The discrete yet ubiquitous theorems of Carath\'eodory, Helly, Sperner, Tucker, and Tverberg
Jesus A. De Loera, Xavier Goaoc, Fr\'ed\'eric Meunier, Nabil Mustafa
TL;DR
This paper explores five fundamental discrete theorems from combinatorial topology and geometry, highlighting their connections and broad applications across various fields like game theory and optimization.
Contribution
It provides a unified discussion of five key theorems, emphasizing their interrelations and significance in multiple application domains.
Findings
Identifies connections among Sperner, Tucker, Carath\'eodory, Helly, and Tverberg theorems.
Highlights their broad impact in fields such as game theory, graph theory, and computational geometry.
Emphasizes the importance of these theorems in practical applications.
Abstract
We discuss five discrete results: the lemmas of Sperner and Tucker from combinatorial topology and the theorems of Carath\'eodory, Helly, and Tverberg from combinatorial geometry. We explore their connections and emphasize their broad impact in application areas such as game theory, graph theory, mathematical optimization, computational geometry, etc.
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
The discrete yet ubiquitous theorems of Carathéodory, Helly, Sperner, Tucker, and Tverberg
Jesus A. De Loera111University of California, Department of Mathematics, Davis, CA 95616, USA.
Xavier Goaoc222Université de Lorraine, CNRS, Inria, LORIA, F-54000 Nancy, France.
Frédéric Meunier333Université Paris Est, CERMICS, ENPC, F-77454, Marne-la-Vallée, France.
Nabil Mustafa444Université Paris-Est, Laboratoire d’Informatique Gaspard-Monge, Equipe A3SI, ESIEE Paris, Noisy-le-Grand, France.
Abstract
We discuss five fundamental results of discrete mathematics: the lemmas of Sperner and Tucker from combinatorial topology, and the theorems of Carathéodory, Helly, and Tverberg from combinatorial geometry. We explore their connections and emphasize their broad impact in application areas such as data science, game theory, graph theory, mathematical optimization, computational geometry, etc.
Contents
-
7.1 Equipartitioning: ham sandwich theorem and its relatives
-
7.3 Parametrized partitioning of data via topological methods
1 Introduction
This article surveys the theory and applications of five elementary theorems. Two of them, due to Sperner and Tucker, are from combinatorial topology and are well-known for being the discrete analogues of Brouwer’s fixed point theorem and the Borsuk-Ulam theorem. The other three, due to Carathéodory, Helly, and Tverberg, are the pillars of combinatorial convexity. These theorems are between fifty and one hundred years old, which is not very old as far as mathematics go, but have already produced a closely-knit family of results in combinatorial geometry and topology. They have also found spectacular applications in, among others places, mathematical optimization, equilibrium theorems for games, graph theory, fair-division problems, the theory of geometric algorithms and data analysis.
The first goal of this paper is to introduce some of the many reformulations and variations of our five theorems and explore how these results fit together. It is convenient to split this presentation into two parts. Sections 2 and 3, discuss Sperner-Tucker and Carathéodory-Helly-Tverberg respectively. At a coarse level, the former deals with combinatorial topology and the latter deals with combinatorial geometry. In each case, we include a special section on algorithmic aspects of these results relevant later for applications.
The second goal of this survey is to sample some of the many applications of our five theorems. There, we proceed by broad areas and examine in Sections 4 to 7, examples from game theory and fair division, from graph theory, from optimization, and from geometric data analysis. Some of our illustrations are classical (e.g., Nash equilibria, von Neumann’s min-max theorem, linear programming), others are more specialized (e.g., Dol’nikov’s colorability defect or the polynomial partitioning technique). We aim to show that our five theorems provide simple proofs of them all. This led us to present some new proofs, for instance for Meshulam’s lemma (Section 2) or for the ham sandwich theorem (Section 7).
The research topics that we discuss are vibrant and have already prompted a number of prior surveys [28, 42, 50, 71, 119, 151, 218, 210, 395], but other surveys were focused on a single one of the five theorems or did not cover applications. The important developments that we present here are from the past few years and emphasize both a global view and the value of geometric and topological ideas for modern applied and computational mathematics. This research area abounds with open questions, all the more enticing because they can often be stated without much technical apparatus. We made a particular effort to stress some of them.
1.1 The five theorems at a glance
Let us start with a classical rendition of Brouwer’s fixed point theorem. If you stand in your favorite Parisian boulangerie holding a map of the city in your hands, then crumple and squeeze it (without ripping it apart, mind you) and throw it to the ground, some point on the map must have landed right on top of its precise
location. Brouwer’s theorem follows from the classical Sperner lemma on the labeling of triangulations (see Figure 1). Surprisingly, all that is needed to prove Sperner’s lemma is to understand why a house with an odd number of openings (doors and windows) must have a room with an odd number of openings. This simplicity and its amazing applications attracted the attention of popular newspapers [369] and video sites [213]. Sperner’s lemma is one of five theorems and we present it in detail in Section 2.1.
In the game of Hex, two players take turns coloring, black and white, the hexagonal cells of an diamond-shape board (see picture on the right); the opposite sides of the board have matching colors, and the player that manages to connect the two sides of his/her color wins (here, black wins). Since its invention by Hein in 1942, there has never been a draw in Hex. The fact that there is always a winner happens to have a geometric explanation: for any triangulation of the projective plane and any two-coloring of its vertices, one of the color classes spans a non-contractible cycle [360]. (To see that this implies the impossibility of a draw
in Hex, take the dual of the hexagonal cell decomposition to obtain a triangulation of the diamond, then carefully identify the boundaries to turn that diamond into a projective plane.) This geometric property is equivalent to the two-dimensional case of Tucker’s lemma, whose statement is given in the caption of Figure 2. Tucker’s lemma is discussed in Section 2.1, see in particular the detailed discussion following Proposition 2.2.
As a matter of fact, Gale [175] proved that the game of Hex cannot end in a draw using Brouwer’s fixed point theorem and Nash [301] proved that for boards of arbitrary size, the first player has a winning strategy. Another application of Tucker’s lemma is the ham sandwich theorem, which says that any three finite measures in (such as a piece of bread, a slice of cheese, and a slice of ham for an open-faced sandwich) can be simultaneously bisected by a plane.
Let us now consider finite point sets in the plane. It turns out that any seven points can be partitioned into three parts so that the triangles, segments, and points that they form have a point in common; for example, the seven points on the left admit , , and as such a partition. This is the simplest case of Tverberg’s theorem. Tverberg’s theorem will be discuss at length in Section 3.3. As the number of points grows, so does the number of possible parts in which we can partition the points while assuring all the convex hulls intersect: points allow four parts, points allow five parts, …, and in general points allow parts. A similar phenomenon holds in arbitrary dimension: any set of points can be partitioned into parts whose convex hulls intersect. Coming back to our
(uncrumpled) map of Paris, consider the points that represent the subway stations. By Tverberg’s theorem, they can be partitioned into parts, so that the corresponding 101 triangles and segments all intersect in a common point . Observe that any line passing through must leave at least subway stations on either of its (closed) sides. The median of a list of real numbers separates the list in half by sizes. The properties of point make it an acceptable two-dimensional generalization of the median, but now for the set of subway stations. More generally, the centerpoint theorem, which follows from Tverberg’s theorem, asserts that for any finite measure in the plane there is a point such that for every halfplane that contains . As we will show in Section 7.4 centerpoints are very important objects in applications and have influenced geometry too, e.g., Tverberg himself was motivated to prove his famous theorem (which we discuss at length later), with intention of finding an elegant proof of the centerpoint theorem. See the end of his classic paper [375].
In the (fully supervised) classification problem in machine learning, one is given a data set (e.g., images), each with a tag (e.g., indicating whether the image depicts a cat or a car), and one is presented with new data to be tagged. A natural approach is to map the set of data to a set of points in some geometric space and look for a simple separation in the space (e.g., a line or a circle in the plane) that separates the points with different tags; for instance, perceptron neural networks – some of the basic classifiers – look for a hyperplane best separating the tagged point sets.
It is easy to conclude that a separator exists by producing an explicit hyperplane. Kirchberger’s theorem states that it is also easy to certify when no line separator exists. In the plane, when no line separator exists, then there must exist a point of one of the colors, say blue, contained in a triangle of the opposite color, red, or
a red segment intersecting a blue segment. The set of all lines that define a halfplane containing a given point of defines a convex cone in , so Kirchberger essentially reduces to an intersection property: if a finite family of convex sets in has empty intersection, then some of these sets already have empty intersection. This property is also known as Helly’s theorem and is one of our main theorems. The curious reader may check that the centerpoint theorem, discussed above, also follows easily from Helly’s theorem. It will be carefully studied in Section 3.2.
Let us finally turn our attention to the geometry underlying the popular magic squares from ancient China. A magic square is a square grid of non-negative real numbers such that the entries along any row, column, and diagonals, all add up to the same value. Look at the four examples on the right. It turns out that any magic square can be written as a linear combination, with non-negative coefficients, of only three of these four magic squares! In fact, for any there exists a finite set of magic squares
such that any other magic square can be written using only
[TABLE]
[TABLE]
[TABLE]
[TABLE]
elements of . This last statement follows from Carathéodory’s theorem, which we will study carefully in Section 3.1 : any vector in a cone in is a non-negative linear combination of extreme rays of the cone, and only dimension many are used in the linear combination. Indeed, the set of magic squares forms a polyhedral cone in a vector space of dimension . It may come as a surprise that no one knows what is for all . See [9].
A colorful generalization of Carathéodory’s theorem asserts that if three polygons in the plane, one with red vertices, one with green vertices, and one with blue vertices, all contain a given point , then there exists a colorful triangle, using a vertex of each color, that also contains . This implies that for the centerpoint that we constructed earlier for the Parisian subway stations from Tverberg’s theorem, at least of the triangles spanned by the subway stations contain . In fact, there is a quantitatively stronger statement given by the first selection
lemma. It states that for any set of points in the plane, there exists apoint covered by at least of the triangles they span. We will see more about this topic in Section 7.4.
1.2 Notation and preliminaries
We collect in this subsection notation, terminology, and general basic background on combinatorics, geometry, and topology that will be used in the rest of this survey. The advanced reader may want to skip or move quickly through this section. For a more thorough introduction to the topics listed here, we recommend the classical books and textbooks in combinatorial convexity [52, 193, 194, 256] as well as [344, ]. For topological combinatorics and combinatorial aspects of algebraic topology see [132, 257, 287].
Given , we write to denote the set . If is a set and , we write for the set of -element subsets of . The notation denotes asymptotic notation where we ignore poly-logarithmic factors: if there exists such that .
We denote by the orthonormal frame of . Given two vectors and in , we write to mean that for . We write for the unit ball in and for the unit sphere in .
1.2.1 Polytopes, simplices, polyhedra, cones
Let be a set. The convex hull of , denoted by , is the intersection of all convex sets containing . In other words, is the smallest convex set containing . It is well-known that
[TABLE]
A polytope is the convex hull of a finite set of points in . Here are a few examples. The convex hull of affinely independent points is a simplex; the standard -dimensional simplex is . With is a the -standard unit vector. The convex hull of is the -dimensional cross-polytope. The convex hull of all vectors with entries is the -dimensional hypercube. A face of a polytope is its intersection with a hyperplane that avoids its relative interior. Faces of dimension [math] are vertices and inclusion-maximal faces are facets. A face of a polytope (resp. simplex) is also a polytope (resp. simplex). There is a face of dimension , the empty set.
A polyhedron is the intersection of finitely many halfspaces in . In particular, any polyhedron can be represented as where is an matrix and . A polyhedral cone is a polyhedron that is closed under addition and scaling by a positive constant. In particular, any polyhedral cone writes as .
The polar of a point is the halfspace (and vice-versa) and the dual of is the hyperplane bounding its polar; we speak of polarity or duality relative to to represent the polarity or duality after the coordinate system has been translated to have its origin in .
The Weyl-Minkowski theorem asserts that the polytopes are exactly the bounded polyhedra. (This can be proven using polarity arguments.) Polytopes can thus be represented both as convex hulls of finitely many points, or as intersections of finitely many halfspaces. To see that these viewpoints are complementary, we invite the reader to prove that projections or intersections of polytopes are polytopes. The Weyl-Minkowski theorem similarly implies that any polyhedral cone also writes as the set of convex combinations of a finite set of rays emanating from the origin. By a theorem of Motzkin et al. [283], every polyhedron decomposes into the Minkowski sum of a polytope and a polyhedral cone : .
A polyhedron is pointed if the largest affine subspace it contains is zero-dimensional. Any polyhedron can be decomposed into the Minkowski sum of a pointed polyhedron and a vector space.
1.2.2 Simplicial complexes
A geometric simplicial complex is a family of simplices with two properties: the intersection of any two distinct simplices is a face of both of them; and it contains all the faces of every member of the family. Simplicial complexes may also be considered abstractly by only retaining which sets of vertices span a simplex. Formally, an abstract simplicial complex is a family of finite subsets (the faces) of some ground set (the vertices) that is closed under taking subsets: if and , then . We write for the abstract -dimensional simplex consisting of all subsets of . A facet of a simplicial complex is an inclusion-maximal face.
Let us stress that the meaning of the word face (or facet) depends on the context and can denote a polytope (for polytopes), a simplex (for simplices), or a set of vertices (for abstract simplicial complexes). In particular, we consider to be a polytope, so it has faces of maximal dimension ; in contrast, has a single face of maximal dimension .
Given a geometric simplicial complex , we let denote the underlying topological space, that is . If is the abstract simplicial complex obtained from , we say that is a geometric realization of and put . (The reader can check that all this is well-defined up to homeomorphism.) A triangulation of a topological space is a (geometric or abstract) simplicial complex whose underlying topological space is homeomorphic to .
1.2.3 Homology
We will use some basic notions of homology, mostly simplicial homology over or . To allow readers unacquainted with homology to appreciate at least our simplest examples, we recall here the basic definitions. An important idea in homology theory is that topological spaces can be studied by associating to them some groups, called homology groups. These groups can be defined geometrically (in singular homology) or combinatorially (in simplicial homology) from a triangulation of the space. In the cases that we consider, these approaches produce isomorphic groups, and we mostly work with simplicial homology.
Given a simplicial complex , we denote by the set of finite formal sums of -dimensional faces of , and is the chain complex of . The map that sends every -face of to the formal sum, with coefficients in , of its -faces extends linearly, over , to a map . Notice that has an additive group structure and that the sum of the is a morphism from into itself; this morphism is called the boundary map of . Note that with -coefficients finite formal sums are simply subsets and the boundary operator maps to proper subsets that appear an odd number of times. It turns out that , so we can define the -th homology group of over coefficients as the quotient group .
Intuitively, the rank of relates to the number of independent holes of dimension in ; for example, the rank of counts the number of connected components of . In particular, if is a single vertex then all its homology groups are trivial except the [math]-th one. The reduced homology groups modify slightly so that it is trivial for connected sets; in dimension , homology groups and reduced homology groups coincide and is defined as [math] for nonempty complexes. Going from to other coefficient groups involves only one technical complication: the definition of involves some sign book-keeping, so as to ensure that . See [287] for details.
2 Combinatorial topology
From combinatorial topology, we will focus on two results about labeled or colored triangulations of simplicial complexes: Sperner’s lemma and Tucker’s lemma. The importance of these two lemmas owes much to their special position at the crossroads of topology and combinatorics. As we alluded in the introduction, Sperner’s and Tucker’s lemmas are the combinatorial equivalent versions to the famous topological theorems of Brouwer and Borsuk-Ulam, respectively. Their combinatorial nature makes them particularly well-suited for computations and applications too.
Combinatorial structures have been used in algebraic topology since Poincaré’s foundational analysis situs, so it is not surprising that some topological questions may be studied by combinatorial methods. The lemmas of Sperner and Tucker are well-known for offering an elementary access, via labelings of combinatorial structures, to important results in topology such as the theorems of Brouwer and Borsuk-Ulam.
It is perhaps less obvious that some combinatorial problems may be studied by topological methods. A seminal example of topological methods applied to combinatorics was the use by L. Lovász of the Borsuk-Ulam theorem to settle a conjecture of Kneser on the chromatic number of certain graphs (see Section 5). His paper opened up the application of topological methods in combinatorics that are now common tools. These techniques appear in several books [132, 257] and surveys [62, 224]. In many cases the topological methods hinge on the theorems of Brouwer or Borsuk-Ulam; as we discuss in the application sections, on several occasions the topological machinery can be made implicit, and the combinatorial question settled directly by the lemmas of Sperner or Tucker.
2.1 Sperner and Tucker
A labeling of a simplicial complex by a set is a map from the vertices of to ; the label of a vertex is its image. Sperner’s lemma gives a sufficient condition for the existence of a fully-labeled facet, that is a facet whose vertices have pairwise distinct labels. (Sometimes the labels are called colors and fully-labeled faces are called colorful; we will avoid this terminology in this paper to avoid confusion with the colorful theorems in convex geometry that we discuss in Sections 3 and 5.)
Sperner’s lemma.
Assume that the vertices of a finite triangulation of a simplex are labeled so that any vertex lying in a face of has the same label as one of the vertices of that face. If the vertices of are given pairwise distinct labels, then the number of facets of whose vertices have pairwise distinct labels is odd.
We call a labeling that satisfies the assumptions of Sperner’s lemma a Sperner labeling. A more general version holds for pseudomanifolds, i.e., for pure -dimensional simplicial complexes where every face of dimension is contained in at most two facets. (Recall that a simplicial complex is pure if all its inclusion-maximal faces have the same dimension.) In particular, any triangulation of a -dimensional manifold is a pseudomanifold of dimension . The boundary of a pseudomanifold is the subcomplex generated by its -dimensional simplices that are faces of exactly one -dimensional simplex.
Proposition 2.1**.**
Any labeling by of a -dimensional pseudomanifold without boundary has an even number of fully-labeled facets.
Proposition 2.1 follows from a simple parity argument. Consider the graph where the nodes are the facets and where the edges connect pairs of facets that share a -dimensional face whose vertices use every label in . This graph has only nodes of degree [math], , or , so it consists of vertex-disjoint cycles and paths. The nodes of degree are exactly the fully-labeled facets and there are evenly many of them (twice the number of paths). Coming back to the remark in introduction: this is where it is useful to understand why a house with an odd number of openings has a room with an odd number of openings.
Clearly Sperner’s lemma follows from Proposition 2.1. For this observe that any Sperner labeling of a -dimensional simplex extends into a triangulation of where (i) the outer vertices of form a fully-labeled facet, and (ii) no other added facet is fully-labeled (as illustrated in Figure 3). Knowing a vertex of degree one allows easily, by path following, to find another one; we come back in Section 2.4 on algorithmic applications of this idea. Other arguments can be used to prove Sperner’s lemma [268].
Now we will state the octahedral Tucker lemma. This is a rather streamlined version of Tucker’s lemma that already suffices for all our applications. Given vectors of signs , we write if every nonzero coordinate of is the same as in . We let denote the set of indices such that , and similarly. In particular, if and only if and . Note that each vector of signs uniquely identifies a coordinate (sub)-orthant, and that the order indicates containment. There is an interpretation of as a simplicial complex illustrated in Figure 4. By we mean a choice of one of the two scalars or .
Octahedral Tucker lemma.
Let be such that for all . If for all , then .
The octahedral Tucker lemma was apparently first stated explicitly by Ziegler [394, Lemma 4.1], following its implicit use by Matoušek [258] in his proof of the lower bound on the chromatic number of Kneser graphs from Tucker’s lemma (see Section 5.1). Several classical proofs of Proposition 2.2 can be found in Matoušek’s book [257]. As we explain in Section 2.2, the lemmas of Sperner and Tucker are indeed “topological” in that they essentially state that certain chain maps, namely those induced by the labeling maps, are non-trivial in simplicial homology with coefficients over .
A more common version of Tucker’s lemma deals with triangulations of a ball instead of an octahedron. Tucker’s lemma gives a lower bound on the number of distinct labels used by labelings that avoid certain local patterns. We say that a triangulation of induces a symmetric triangulation of if its boundary forms a centrally-symmetric triangulation of . A labeling of a symmetric triangulation of by integers is antipodal if antipodal vertices have opposite labels.
Proposition 2.2** (Tucker lemma).**
Let be a triangulation of that induces a symmetric triangulation of . Let be a labeling of the vertices of by . If for every vertex of , then there exists an edge in with .
(See Figure 5 for an illustration.) The octahedral version is obtained by applying Proposition 2.2 to the barycentric subdivision of the -dimensional cross-polytope [257, Theorem 2.3.2]. Indeed, consider a map such that for all and for all . Any can be interpreted as the vertex in corresponding to the face of . The edges in connect pairs and such that . Defining , we get a labeling of all vertices of satisfying for every vertex of and having no edge in with . By Proposition 2.2, we must have .
As stated in the introduction, Tucker’s lemma is equivalent to the fact that for any triangulation of the projective plane and any two-coloring of its vertices, one of the color classes spans a non-contractible cycle. Indeed, such a two-coloring can be seen as a two-coloring of a triangulation of the disk, with symmetric vertices of the boundary getting identical colors. If all monochromatic cycles were contractible, we could easily choose a sign for each vertex and get a labeling that would contradict Tucker’s lemma. The reverse implication is also easy and left to the reader.
Consider a triangulation as in Proposition 2.2 and a labeling of its vertices by ; a -dimensional face of is alternating if its vertices can be indexed so that or if . In the first case we call the simplex positively alternating and in the second case negatively alternating. A lemma due to Fan [159] generalizes Tucker’s lemma in terms of a parity counting of alternating simplices.
Theorem 2.3** (Fan’s lemma).**
Let be a triangulation of that induces a symmetric triangulation of . Let be a labeling of the vertices of by such that for every vertex of . If no two adjacent vertices have opposite labels, then has an odd number of alternating facets.
Fan’s lemma readily implies Tucker’s lemma since the existence of an alternating -dimensional face implies . Going in the other direction, it was only recently observed by Alishahi [15] that an existential version of Fan’s lemma is easily derived from Tucker’s lemma.
Let us illustrate that, surprisingly, Fan’s lemma can be easier to prove than Tucker’s lemma: this is one example where a stronger hypothesis facilitates induction. We give an inductive proof for a flag of hemispheres, i.e., a triangulation of such that the restriction of on and on triangulates each of them, where and are the -dimensional hemispheres
[TABLE]
[TABLE]
(Prescott and Su [317] gave another combinatorial constructive proof for this special case.) Consider the graph whose nodes are the facets of and whose edges connect pairs of facets that share a -dimensional face that is positively alternating. We augment this graph with an extra node and add edges connecting to all facets of that have a -dimensional positively alternating face on ; in a sense, represents the “outer facet”. Apart from , all nodes have degree [math], , or . The nodes of degree are exactly the -dimensional alternating facets. The triangulation refines , which is homeomorphic to . So, by induction in dimension , the number of -dimensional alternating faces of in is odd; the same holds for the -dimensional faces of in . The symmetry of , and that of the labeling, imply that the degree of is odd; it follows that there is an odd number of -dimensional alternating facets in .
The above elementary proof requires that the triangulation restricts nicely to lower-dimensional spheres to allow induction. Two proofs of Theorem 2.3 can be found in the literature, both for an equivalent version with a sphere instead of a ball (later we will explain this equivalence for our own proof of Theorem 2.3). On the one hand, Živaljević [396] observed that the labeling is essentially a classifying map that is unique up to -homotopy, so the number of alternating facets (mod ) reformulates as the cap product of a certain cohomology class with a certain homology class. On the other hand, Musin [288] builds a simplicial -map from the triangulation to a -dimensional polytope for which the following holds: a simplex is alternating if and only if its image by this simplicial map contains [math] in its convex hull; a degree argument allows then to conclude. It turns out that Alishahi’s idea to derive Fan’s lemma from Tucker’s lemma leads to a short proof of Theorem 2.3, also based on a degree argument. Before we spell out this (original) proof let us stress that the following question remains:
Open Problem 2.4**.**
Give a direct combinatorial proof of Fan’s lemma (as stated in Theorem 2.3) and of Tucker’s lemma (Proposition 2.2) valid for any centrally symmetric triangulation.
Let us now prove Theorem 2.3 for an arbitrary triangulation of . Let be a labeling of the vertices of satisfying the conditions of Theorem 2.3. We first turn into a triangulation of by gluing two antipodal copies of , each triangulated by ; notice that the number of positively alternating facets of equals the number of alternating facets of (both positive and negative ones), since the negatively alternating facets in one copy of become positively alternating in the other copy. We define next a labeling of the vertices of by : a vertex of corresponds to a simplex of , and we set the absolute value of to be the number of vertices of the largest alternating face of , and its sign according to alternation. This sign is defined uniquely since there cannot be maximal alternating faces of both types in (this can be checked using the fact that no adjacent vertices in can have opposite labels). Now, a crucial observation is that if is an alternating facet (for ), then contains exactly one alternating facet (for ) of the same type; and if is not an alternating facet, then contains no alternating facet. At this point, to establish Theorem 2.3, it suffices thus to prove that contains an odd number of positively alternating facets. This fact follows from basic degree theory. The assumptions on guarantee that induces an antipodal simplicial map from to , the boundary of the -dimensional cross-polytope, whose vertices are identified with the elements in . Any antipodal self-map of is of odd degree [141, Theorem 4.3.32]. Thus, denoting by the formal sum of all facets of and by the formal sum of all facets of , we must have . The only simplices that are mapped to the simplex by are the positively alternating ones, so there are an odd number of them.
2.2 Continuous versions
One of the most famous theorems about fixed points is due to the Dutch mathematician L. E. J. Brouwer and states that any continuous function from a ball into itself has a fixed point. Brouwer’s theorem is often seen as a continuous version of Sperner’s lemma (without the oddness assertion): they can be deduced easily from one another.
Let us sketch how Brouwer’s theorem follows from Sperner’s lemma (we discuss the other direction in Section 2.3). Contrary to Brouwer’s original proof, which says nothing about how to find the fixed point or a good approximation of it, this proof has computational implications (see Section 2.4). Without loss of generality we take the -dimensional standard simplex as a realization of a ball (it is easy to set up a homeomorphism). Then we triangulate the simplex and design a labeling of that triangulation tailored to the continuous function under consideration. Specifically, we associate to a vertex of the triangulation the label if the -th barycentric coordinate of is larger than the -th barycentric coordinate of its image . (So, intuitively, if is labeled , then moves outwards the -th vertex of the standard simplex.) Note that, unless the vertex is a fixed point, there must be at least one such index. If there are several such indices, simply make an arbitrary choice among them. Now, the labeling we provided satisfies the assumptions of Sperner’s lemma, so we can find a fully-labeled simplex of such that the -th barycentric coordinate of the vertex labeled is decreased by . Re-triangulate again and again adding more and more points in such a way that the maximum diameter of the simplices appearing in the triangulation goes to zero in the limit. At each step we find a fully-labeled simplex. The barycenters of all such simplices will produce an infinite sequence of points and, since it is a bounded sequence, it contains a convergent subsequence. Let be the limit of this subsequence. Since the map is continuous, the -th barycentric coordinate of is larger or equal than the -th barycentric coordinate of for every and therefore is a fixed point of the map.
The Knaster-Kuratowski-Mazurkiewicz theorem, also known as the KKM theorem, is a classical consequence of Sperner’s lemma or Brouwer’s theorem. It is used for instance in game theory or for the study of variational inequalities. Consider the -dimensional simplex and closed sets in . This theorem, illustrated in Figure 6, ensures that if for every the face of is covered by , then is nonempty. This statement is somehow reminiscent of Helly’s theorem. A corollary of the KKM theorem can actually be used to prove it, see Section 3.2. This corollary states that if closed sets are such that each of them contains a distinct facet of and such that their union covers , then their intersection is nonempty (this statement is also called a dual KKM theorem in [32]). To see that it is a consequence of the KKM theorem, assign number to the facet covered by ; number the vertices of so that vertex is on facet ; the ’s satisfy then the condition of the KKM theorem.
Several variations of the original KKM theorem exist. Gale [176] proved the following colorful version: given different KKM covers , …, of the -simplex, there exists a permutation of the numbers such that . Clearly choosing all the covers to be the same recovers the classical version of KKM stated above. Gale’s colorful KKM theorem has an intuitive interpretation first stated by Gale himself: “if each of three people paint a triangle red, white and blue according to the KKM rules of covering, then there will be a point which is in the red set of one person, the white set of another, the blue of the third”. A recent strengthening of this colorful theorem [32] states that, given KKM covers of the -simplex , then there exist a point in and bijections for , such that for every . It is interesting to note the proofs of these colorful results combine degree theory with Birkhoff’s theorem on doubly-stochastic matrices. Finally, we note [290] has a common generalizations of Sperner, Tucker, KKM, Fan.
Another fascinating and very useful consequence of Brouwer’s theorem is Kakutani’s 1941 fixed-point theorem. It deals, not with real-valued functions, but with set-valued functions, where points are mapped to subsets. For a suitable notion of continuity, it ensures that for any continuous function mapping points of a ball to convex subsets of it there is an such that . Kakutani’s theorem is especially useful in game theory, its most traditional application being the Nash theorem, see Section 4.
Similar to Sperner’s lemma, Tucker’s lemma has continuous and covering versions. The continuous version is the celebrated Borsuk-Ulam theorem, which has many applications in discrete geometry, combinatorics, and topology. It asserts that there is no continuous function from into that commutes with the central symmetry. Nice proofs of the Borsuk-Ulam theorem from Tucker’s lemma, as well as equivalent formulations and many applications, can be found for instance in the books of Matoušek [257, Chapter 2.3] and de Longueville [132, Chapter 1]. Just as KKM is the covering version of Sperner’s lemma, Tucker’s lemma has a covering version, the Lyusternik-Schnirel’mann theorem [249]. It states that, if the sphere is covered by closed subsets, then one of the sets must contain two antipodal points. This theorem and some of its extensions (e.g., those due to K. Fan) have found many applications in other areas of mathematics, for instance, as for the KKM theorem, in the study of variational inequalities.
2.3 Generalizations and variations
A labeling of a simplicial complex by a set can be interpreted as a map from the vertices of to the vertices of some abstract simplicial complex with vertex set . This viewpoint leads to several interesting developments.
A first idea is to extend into a linear map from into . For a Sperner labeling, maps into itself. Composing with a suitable orthogonal transformation ensures that any fixed point of the resulting map, which must exist by Brouwer’s theorem, lies in a fully-labeled simplex; this is the standard proof of Sperner’s lemma from Brouwer’s theorem [132, Section 1.1]. Using this idea, Sperner’s lemma can be generalized to triangulations of arbitrary polytopes by De Loera, Peterson and Su [130].
Proposition 2.5** (Polytopal Sperner lemma).**
Let be a polytope with vertices, a triangulation of , and a labeling of the vertices of by . If the vertices of have pairwise distinct labels and every vertex of lying on a face of the boundary of has same label as some vertex of , then contains at least fully-labeled facets.
The gist of the proof is that extends into a piece-wise linear map from to (as illustrated in Figure 7). This map can be shown to be surjective, so its image provides a covering of by simplices spanned by its vertices. The number of such simplices required to cover is at least , and each of them is the image (under the extension of ) of a fully-labeled facet of . This approach generalizes to non-convex polyhedral manifolds [272] and broader classes of manifolds [288, 289]. The reader interested in recent progress on lower bounds for the number of fully-labeled facets is refered to the work of Asada et al. [32].
Another idea is to extend into a chain map . Depending on the coefficients ring and the complex , one gets different generalizations of the classical statements. This extension goes as follows. Send every simplex of to if the are pairwise distinct and to [math] otherwise; the linear extension of this map to is the map , and it commutes with (it is a chain map).
For , we obtain a short proof by induction of Sperner’s lemma. Assume that is a Sperner labeling of a triangulation of , so . Let denote the sum of -simplices of and the unique -simplex of . Observe that , where is the number of fully-labeled simplices in mod . A simple induction shows that . Thus, . It comes that , and has an odd number of fully-labeled simplices.
For , the same proof gives a Sperner-type result for oriented simplices. The orientation of a fully-labeled simplex of , where has label , is defined as the sign ( or ) of the determinant
[TABLE]
Specifically, this proves that in any Sperner labeling of a triangulation of by the orientations of the fully-labeled simplices add up to [160, Theorem 2]. This approach also yields a proof of (a signed version of) Fan’s Lemma [132, Theorem 1.10], again for triangulations with flags of hemispheres.
For , this idea leads to generalizations of the lemmas of Tucker and Fan with more than two “signs” [200, 271]. The general insight, following a generalization by Dold [143] of the Borsuk-Ulam theorem, is to replace and the antipodality by a suitable simplicial complex on which acts freely. The resulting -Fan lemmas actually provide combinatorial proofs of Dold’s theorem. The deduction of the -Fan lemmas from their -Tucker versions works as explained above in the case [15].
A variant of the above argument yields a new and elementary proof of a relative of Sperner’s lemma, due to Meshulam [269]. It is a powerful result that has found many applications in graph theory and in discrete geometry, such as the recently-found generalization of the colorful Carathéodory theorem by Holmsen and Karasev [211]. We come back later to some of its applications in discrete geometry (Section 3.1) and in graph theory (Section 5.2). Meshulam’s lemma was first explicitly derived from (a special version of) Leray’s acyclic cover theorem (its first use was implicit; see Kalai and Meshulam [221, Proposition 3.1] for an explicit statement and proof for rational homology).
For the sake of presentation, we consider here homology over but the proof generalizes, mutatis mutandis, to an arbitrary ring. Given a simplicial complex and a subset of its vertices, we denote by the simplicial complex formed by the simplices of whose vertices are in .
Proposition 2.6** (Meshulam lemma).**
Let be a labeling of the vertices of a simplicial complex by . If is trivial for every nonempty , then contains a fully-labeled -dimensional face.
The proof goes as follows. Let denote the chain map induced by and recall that it maps simplices with repeated labels to [math]. We build a chain map such that ; the identity then ensures that contains an odd number of fully-labeled simplices, proving the statement. We build by increasing dimension. We start by setting to be some (arbitrary) vertex in for every ; this is possible because is trivial. Assume that is defined over all chains up to dimension , that it maps any subset of cardinality at most to , and that it commutes with the boundary operator. Now, for any subset of cardinality , we have
[TABLE]
so the -chain is a cycle in . The assumption of the lemma ensures that it is the boundary of some chain , and we set to be . To see that for any , first note that this is straightforward for . For the general case, remark that
[TABLE]
so we can assume by induction that . We conclude by observing that means that is supported only on subsets of , so it must be that .
The parity argument used to prove Proposition 2.1 can also be found, specialized to a certain labeled pseudomanifold, in Scarf’s proofs of the integer Helly theorem [339] and his classical lemma in game theory [338]. There is a related unbounded polar version that will be useful in Section 5.3.
Corollary 2.7** ([231, Theorem 3]).**
Let be a -dimensional pointed polyhedron whose characteristic cone is generated by linearly independent vectors and whose facets are labeled by . If no facet containing the -th extreme direction is labeled , then there exists a extreme point incident to a facet of each label.
Another recent variation of Sperner’s lemma, motivated by applications in approximation algorithms, asks for the minimum possible number of facets in the Sperner labeling that must be non-uniquely-labeled. Mirzakhani and Vondrak [280, 281] settled this question for certain triangulations of the simplex, for which they provided the optimal Sperner labeling. They then proposed two very interesting open questions.
Open Problem 2.8**.**
Is there a theorem that interpolates between the result above (lower bound on the number of simplices with at least two different colors) and the original Sperner’s lemma (lower bound on the number of simplices with vertices of different color) by predicting how many simplices are colored with at least different colors? How does such theorem depend on the structure of the particular triangulation?
It must be mentioned that Tucker-type theorems and Sperner-type theorems are related to each other in interesting ways. For example, it is known that the Borsuk-Ulam theorem implies the Brouwer fixed-point theorem, but at the combinatorial level Nyman and Su [305] proved that Fan’s lemma implies Sperner’s theorem too. Other interconnections can be found in [317, 358].
2.4 Computational considerations
The proof of Sperner’s lemma given for Proposition 2.1 builds a graph where every vertex has degree zero, one, or two, then exhibits a vertex of degree one and argues that any other vertex of degree one must correspond to a fully-labeled simplex. This provides a simple algorithm for finding a fully-labeled simplex: just follow the path! We can combine this simple path-following algorithm for finding fully-labeled simplices with the proof of Brouwer’s theorem, presented at the beginning of Section 2.2, and provide a method for finding an approximate fixed-point of the continuous map . Again assume we are given a continuous map and an . Our goal is to find such that . For this, it suffices to compute a triangulation of with simplices of diameter sufficiently small, depending on and the modulus of continuity of , label it as in the proof of Brouwer’s theorem, and any vertex of a fully-labeled simplex does the job (this fact is more easily formalized by using the norm on the barycentric coordinates). This template of proof was first presented in [337] and is quite flexible, e.g., it applies to non-contracting functions. We left out many details, for instance the choice of the triangulation to speed up the algorithm and the estimation of the modulus of continuity. The interested reader can find more details on methods to compute approximate fixed points based on these ideas in [373].
The theory of computational complexity is a formal way for computer scientists to classify the inherent difficulty of computational problems. Families of problems, called complexity classes, collect problems of equivalent difficulty (a complete introduction can be found in [30]). Famous complexity classes of course include the class P and the class NP, but we briefly discuss here, and in Section 3.4, about the complexity classes that relate to computational versions of our five central theorems.
The path-following algorithm for computing the fully-labeled simplex for Sperner’s lemma is representative of the PPAD class, a complexity class well-suited for studying computational problems whose solution is known to exist, but finding it is not that easy. This class was presented by Papadimitriou [313], see also [303]. The prototypical problem of the class PPAD (which abbreviates Polynomial Parity Argument for Directed graphs) assumes an underlying directed graph where every vertex has in- and out-degrees at most one; the graph may be implicit, and all that is required is the existence of a function that computes the neighborhood of a given vertex in time polynomial in the encoding of that vertex. The PPAD problem asks, given the encoding of an unbalanced vertex (that is, with different in- and out-degrees), to compute the encoding of another unbalanced vertex. (Note that this computational problem does not easily reduce to a meaningful associated decision problem, since the existence of this other unbalanced vertex follows systematically from parity considerations.) A problem is in the PPAD class if it has a polynomial reduction to the PPAD problem, and a problem from the PPAD class is PPAD-complete if every problem from the PPAD-class has a polynomial reduction to that problem. (All reductions here are meant in the usual sense of polynomial reductions [303, ].) PPAD-completeness results imply conditional lower bounds in the following sense: one cannot solve a PPAD-complete problem substantially faster than by path-following, unless there is also a substantially better method than path-following for the PPAD problem (and similarly for every other problem in the PPAD class). As in the case of the P vs NP problem, failure over time to improve on even the most streamlined of these problems supports the conjecture that none of these methods can be substantially improved.
The “Sperner problem” – where one asks, given a Sperner labeling, for a fully-labeled simplex – is well-known to be PPAD-complete in any fixed dimension. (Formalizing this problem properly requires some care, for instance the definition of a canonical sequence of triangulations with simplices of vanishing diameter, which we do not discuss.) Papadimitriou’s seminal paper, which started the theory of PPAD problems [313], settled the three-dimensional case and listed the two-dimensional case as an important open problem; it was settled in the positive a decade later by Chen and Deng [102].
While Tucker’s lemma can also be proved via a path-following argument [166], the computational problem associated to Tucker’s lemma is not known to belong to the PPAD class: contrary to Sperner’s lemma, there is no natural orientation of the edges of the underlying graph. The suitable complexity class to use for the “Tucker problem” is a superclass of the PPAD class, the class PPA. Here PPA abbreviates Polynomial Parity Argument for graphs. This class was introduced at the same time as the PPAD class, its definition is almost the same: instead of working with directed graphs, one works with undirected ones. The underlying graph defining the PPA problem has all its vertices of degree at most two and asks, given the encoding of a vertex of degree one, to output the encoding of another degree one vertex. PPA contains PPAD but it is a famous problem to decide whether the two classes are actually the same, already asked in the paper founding this topic [313]. Experts believe that these two classes are different [190].
Open Problem 2.9**.**
Are the complexity classes PPA and PPAD equal?
As for the Sperner problem, the Tucker problem is PPA-complete already in dimension two (see Aisenberg et al. [11], who corrected an earlier, wrong, assertion of PPADness). Pálvölgyi [311] introduced a clean variation of this problem – the octahedral Tucker problem – together with the open question below: Given a function , computable in time polynomial in and such that and for all , compute such that . Note that contrary to the Tucker problem we just discussed, the dimension is not part of the input. The computational complexity of the algorithmic version of the Octahedral Tucker lemma had been an outstanding challenging problem, but the paper [136] resolved this problem by proving Octahedral Tucker to be PPA-complete.
We will also discuss in our applications, in particular Section 4.1.2, some consequences of the lemmas of Sperner or Tucker whose computational versions may be complete for two related complexity classes. The first class is FIXP, introduced by Etessami and Yannakakis [157]. It consists of the problems whose resolution on an instance reduces to the computation of a fixed point of some function that can be expressed by the operations with rational constants and functions and computed in time polynomial in the size of ; this extends PPAD, which coincides with the case of linear functions. The second class, called [340] (sometimes abbreviated ETR for existence of real solutions, see [180]), studies problems that reduce to deciding the emptiness of a general semi-algebraic set, i.e., the set of real solutions of a system of inequalities with polynomials as constraints. These two complexity classes are relevant in Section 4.
3 Combinatorial convexity
We now focus on three classical combinatorial theorems about convex sets first identified in the early 20th century. These are the theorems of Carathéodory, Helly and Tverberg. The importance of convexity in applications, and hence of these three theorems, owes much to the computational effectiveness of convex optimization algorithms both in practice and in theory [58, 76]. This encourages applied mathematicians to look for convexity, or for ways to approximate complicated sets using convex sets. Surprisingly, convexity appears in unexpected settings.
Extensive surveys were devoted to (subsets and variations on) some of these three theorems by Danzer, Grünbaum, and Klee [119], Eckhoff [151], Holmsen and Wenger [210] and Amenta, De Loera, and Soberón. [28]. An account of early variations of Carathéodory’s theorem is in the memoir by Reay [321].
There is an abundant literature on axiomatic convexity, which studies analogues of the theorems of Carathéodory, Helly, and Tverberg, not over Euclidean spaces as we do here, but over purely combinatorial abstract settings, for instance in the convexity spaces defined by arbitrary graphs, finite geometries, matroids, greedoids, etc. The three theorems play a significant, and interesting role, there too, but we do not cover this topic here. We refer the interested reader to the references [147, 229, 379].
3.1 Carathéodory
We will first consider Carathéodory-type theorems that certify membership of a point in the convex hull of a set via linear non-negative combinations. The original theorem of Carathéodory [89] asserts that any point in the convex hull of a finite point set in is a convex combination of some at most of these points. Equivalently, if a vector belongs to the cone of (i.e., the positive hull of all non-negative real linear combinations of vectors in ), then is a positive combination of at most vectors of . To see this, let and assume that is a solution of
[TABLE]
If the support of has size at least , then has some nontrivial solution with support contained
in the support of . For an adequate value of , the vector is a solution of System (3.1) with smaller support than . A closer examination of this argument yields that, in the plane, any point in the convex hull of four points lies in two of the triangles they span (as illustrated on the right). The following strengthening of Carathéodory’s Theorem will be useful in optimization.
Proposition 3.1**.**
Any point in the convex hull of (at least) points in lies in the convex hull of at least two -element subsets of these points.
A geometric proof of Proposition 3.1 with any one of the points, say , and the point in their convex
hull. Shoot a ray from to and collect the (at most) vertices of the face of a (triangulation of the) convex hull through which this ray exits; these points and contain the point in their convex hull, and any of the points can be used as origin of the ray. The figure on the left illustrates this process in the case of a (blue) point contained in the convex hull of the (white) vertices of a cube in . The green projections to faces help us find the four white points containing .
Several variants of Carathéodory’s theorem have been developed. For instance, Steinitz [363] proved that a point in the interior of the convex hull of a set lies in the interior of the convex hull of some points of . A related classical result is the Krein-Milman theorem [250]: if is a convex set, then every point in is a convex combination of its extreme points, i.e., those that are not convex combinations of others in the set. Allowing more flexible representations, Dobbins [140] proved that any point in an -dimensional polytope is the barycenter of points on its -dimensional skeleton (another proof of Dobbins’ theorem is shown in [67]).
One of the most applicable and powerful variants is the colorful Carathéodory theorem. We saw this in the introduction already.
Colorful Carathéodory theorem.
Let be point sets in . If a point is in the convex hull of every , then there exist , , …, such that lies in the convex hull of .
In other words, if the origin is contained in the convex hull of each of point sets (the color classes), then it is contained in a colorful “simplex”, i.e., one where each of the vertices comes from a different (with the understanding that this “simplex” may be degenerate). This theorem was discovered by Bárány [40] who showed that a colorful simplex that minimizes the distance to the origin must contain it. Indeed, when the distance is still positive, it is attained on a facet and the vertex opposite to that facet may be changed to further decrease the distance to the origin. This approach inspired new proofs and algorithms, by minimization, of the colorful Carathéodory theorem [46] and other results such as Tverberg’s theorem [328, 376, 377] (see more on Tverberg’s theorem later in Section 3.3).
An alternative proof of the colorful Carathéodory theorem applies Meshulam’s lemma (Proposition 2.6) to the join of two abstract simplicial complexes built on top of : one has a simplex for any subset of points with no repeated color, and the other has a simplex for every subset of points not surrounding the origin. (The labeling is given by the identification of the vertices of the join to .) This approach emerged from a colorful Helly theorem of Kalai and Meshulam [221] (see below) and later allowed a purely combinatorial generalization of the colorful Carathéodory theorem by Holmsen [208] where the geometry and the colorfulness are abstracted away into, respectively, an oriented matroid and a matroid. See also the paper [211].
The assumption of the colorful Carathéodory theorem ensures that not only one, but actually many colorful simplices exist; we come back to this question when discussing simplicial depth in Section 7.4. This also underlies its connection to Tverberg’s theorem which we discussed in Section 3.3.
Many variations and strengthenings of the colorful Carathéodory theorem have been explored starting with Bárány’s seminal paper [40] and other collaborations [46]. Recent strengthenings include Deza et al. colorful simplicial depth [138] (discussed in Section 7.4) and Frick and Zerbib’s common generalization of the colorful Carathéodory theorem and the KKM theorem [171]. Another key variation, discovered independently by Arocha et al. [29] and Holmsen et al. [209], is that the assumption that the convex hull of each , , contains the origin can be weakened to only require that the convex hull of each pair , , contains the origin. There are examples showing that it is not sufficient that the convex hulls of triples contain the origin, but weaker relaxations are possible [274]. Arocha et al. [29] also proved another “very colorful Carathéodory theorem”.
Via point-hyperplane duality, one can derive from the colorful Carathéodory theorem, a colorful theorem of Helly. We will not discuss this in detail, but let us explain the basics. Consider families of convex sets inside (researchers think of these as colors). Assume that every colorful selection of of the sets, i.e., one set from each , has a non-empty common intersection. Then, the classical colorful Helly theorem of Lovász (see [40]) says that there is at least one family , whose sets have a non-empty intersection. Here is now the dual of the very colorful Carathéodory theorem of [29]: given a finite family of halfspaces in colored with colors, if every colorful selection of halfspaces has a non-empty common intersection, then there exist two colors classes all of whose members intersect. For more examples and references, see [252] and references therein.
The proof of the colorful Carathéodory theorem also implies that given point sets and a convex set , either one can be separated from by a hyperplane, or there exists a colorful simplex intersecting . Building on this, Mustafa and Ray [294] showed that given sets of points in and a convex object , then either one of the sets can be separated from by a constant number of hyperplanes, or there is a -dimensional colorful simplex intersecting .
The integer Carathéodory problem considers a finite set and in its positive hull and asks whether can be written as a non-negative, integer linear combination of some elements of , and, if true, how many elements are needed. The answer to the first question is negative in general. For instance consider
[TABLE]
is an integral vector of the positive hull of but is not an integer non-negative combination of elements of .
It is thus natural to restrict one’s attention to subsets such that every integral point of the cone of can be written as a non-negative integer combination of elements of ; such sets are called Hilbert generating sets. This restriction is reasonable given that the integer points of any rational polyhedral cone have a finite Hilbert generating set. However, even in this setting, there is no version of an integer Carathéodory theorem with a bound on the size of the representation depending only on the dimension. Take for example
[TABLE]
can be written as an integer combination over , but any such combination requires at least summands. Notice that the coefficients in Example (3.3) grow quickly with . Such growth is necessary to force larger and larger sums because a Carathéodory-type theorem is possible if one wants to bound the number of summands in terms of the dimension and the size of the coordinates. The best upper bound in that direction was recently obtained in [14]. An earlier bound appears in [154].
Theorem 3.2**.**
Let be a finite set, let , let , and let denote the sublattice of of the integer points in the row space of . Any vector representable as non-negative integer combination over can be written as a combination of at most terms.
The proof of Theorem 3.2 starts from some non-negative integer combination and uses some element of the kernel of to eliminate one of the summands. This is very similar to the classical proof of the real-valued Carathéodory theorem, but now the kernel element must be an integral vector with coordinate entries in . That such a vector exists is no longer a rank argument, but follows from Siegel’s lemma (see [351, 378]).
Interestingly, a full-fledged integer Carathéodory theorem, depending only on the dimension, does exist for Hilbert bases of pointed cones. Let us explain. First of all, a Hilbert basis is an inclusion-minimal Hilbert generating set. We say a cone is pointed if it contains no linear subspace other than the nullspace. It is known a pointed cone has a unique Hilbert basis (see e.g., [123, Corollary 2.6.4]). In contrast, when the cone is not pointed, there is no uniqueness, for instance has two Hilbert bases and . As we will see in Sections 6.2.2 and 6.3.2, Hilbert bases play an important role in optimization theory and in the solution of integer optimization problems. Here is the best known upper bound of the number of Hilbert basis elements necessary to write a vector. This is due to Sebő [346]:
Theorem 3.3**.**
If the pointed cone is generated by a Hilbert basis , then any of its integral points can be written as a non-negative integer combination of at most elements of .
A weaker version of Sebő’s theorem, with a constant , was previously obtained by Cook, Fonlupt, and Schrijver [114]. Note that the sets in Example (3.3) do define pointed cones, but are not the Hilbert bases of those cones. Sebö’s theorem gives an upper bound, but the best known lower bound on the size of the linear combination is only for [83], so that leaves an open important problem:
Open Problem 3.4**.**
Determine the best possible constant for the integer Carathéodory theorem on Hilbert bases of pointed cones.
The answer is known to be for [346] and in some special cases such as the cone formed by the bases of any matroid [183].
We conclude with an approximate Carathéodory theorem, recently recovered by Barman [51], which has an interesting application in game theory (see Section 4.1.2). Informally, it says that any point in the convex hull of a point set can be approximated by a convex combination of few elements of . The precise relation between the quality of approximation and the size of the convex combination is quantified as follows:
Theorem 3.5**.**
Let and let . For any point and any , there is a point such that , and can be expressed as a linear combination of at most vectors from .
Observe that the number of points used to represent the approximation to is independent of the ambient dimension . The point can in fact be chosen as the barycenter of points of with non-negative integer weights, the sum of the weights being at most . Barman’s nice probabilistic proof writes as a barycentric combination of points of (by Carathéodory’s theorem) and finds the points adding to by sampling those points using the weights as probabilities (some special care must be taken to ensure a bound independent of the dimension). Theorem 3.5 can also be derived from Maurey’s lemma in functional analysis (see Pisier [315] and Carl [90]). See also [72] for the derivation of a related theorem using the Perceptron algorithm [304]. A very recent new generalization of Theorem 3.5 was presented by Adiprasito, Bárány and Mustafa in [3]. They proved that, given a point set of cardinality , a point , and an integer , , then there exist a subset of elements such that the distance between and is less then . Here the diameter of is the largest distance between a pair of points in .
3.2 Helly
Helly’s theorem asserts that for a finite family of convex subsets of with at least members, if every members intersect, then the whole family intersects. In the contrapositive, the empty intersection of finitely many convex sets in is always witnessed by the empty intersection of some of the sets. See Figure 8
The special case of Helly’s theorem where each subset is a halfspace is of particular interest. Since a family of halfspaces not containing the origin has empty intersection if and only if their inner normals positively span the space, Helly’s theorem for halfspaces is equivalent to Carathéodory’s theorem for their polars. Note that the polar of a set of hyperplanes is a set of ray vectors. The case of halfspaces implies the general case because, given a family of convex sets in , we can replace each set by a polytope that it contains without altering the intersection patterns. It suffices to take a witness point in the intersection of every subfamily, and then replace each set by the convex hull of the witness points that it contains.
Helly’s original proof used the separability of compact convex sets by hyperplanes to set up an induction on the dimension. It is spelled out in the survey of Danzer et al. [119, Section 1] along with references to eight other proofs. The most common proof deduces Helly’s theorem from Radon’s lemma (the case of Tverberg’s theorem) [256, Section 1.3]. Starting with convex sets where any intersect, one picks a witness point . Partition into two subsets with intersecting convex hulls, and observe that this intersection point lies in every convex set. A proof of Helly’s theorem, due to Krasnosselsky [235], fits well with the theme of our survey as it uses the KKM theorem. Lift the to the vertex of the simplex . The map extends into a linear map , and setting produces closed subsets of . Each facet of is covered by a distinct and the ’s cover by Carathéodory’s theorem in . The KKM theorem (see Section 2.2)
yields a point and the point is contained in every . Chakerian showed in [92] that Helly’s theorem also follows from Brouwer’s fixed point theorem, in a similar fashion as the proof of the KKM theorem: the function, instead of moving every point by the vector of its distances to each , moves every point to the barycenter of its projections on each . (See the right-hand picture for an illustration: the projections of to the three convex sets, red, green, and blue, are shown by the white points; the black square is their barycenter).
Helly’s theorem for integral points was first established by Doignon [142, Proposition 4.2], and rediscovered later by Scarf [339] and Bell [56]. Hoffman [207] observed that the techniques apply for more cases than the integer lattice. The proofs of Doignon, Bell, and Hoffman hinge on the following insight: if a polytope in has vertices, each with integer coordinates, and contains no other integral point, then no two vertices may have all their coordinates of the same parity (else their mid-point would yield a contradiction) and thus is at most . Bell’s proof starts with a family of halfspaces in whose intersection contains no integer point and such that removing any halfspace would enlarge the intersection to include some integer point.
Translating each halfspace in the direction of its outer normal until every facet contains a witness point with integer coordinates, one gets witness points that must be distinct and form a polytope as above, so is at most . Hoffman’s proof is more complicated but holds in a more general axiomatic setting. Scarf’s proof is algorithmic and relies on Sperner’s lemma (see also the variation by Todd [372]). Remark that the equivalence via polarity between Helly’s and Carathéodory’s theorems in does not carry over to , as the bounds are respectively and at most (by Theorem 3.3).
Some of our applications will use the following version where some of the coordinates in the intersection may be required to be integers.
Mixed Helly theorem.
Let be a finite family of convex sets in of cardinality at least . If every members of have a common point whose last coordinates are integer numbers, then all members of have an intersection point whose last coordinates are integer numbers.
The mixed Helly theorem was announced by Hoffman [207] as one of the outcomes of his axiomatic setting; he however deferred details of the proof for the mixed analogue of the property of the Doignon theorem to a forthcoming paper, which never appeared. A complete proof came decades later and is due to Averkov and Weismantel [36]. Their proof proceeds in two steps. They start with a family of halfspaces in whose intersection is a nonempty full-dimensional polytope containing no point whose last coordinates are integer (the general case follows). They project onto the last coordinates, obtaining a polytope in with no integer point; Doignon’s theorem then ensures that at most of the halfspaces supporting the facets of already intersect with no integer point. By Carathéodory’s theorem in , each -dimensional halfspace is the projection of the intersection of some at most of the original halfspaces; the bound follows.
Fractional versions of Helly’s theorem play an important role in the study of sampling and hitting geometric set systems. Here, fractional means that one only assumes that some constant fraction of the subfamilies (of a given size) intersect, and concludes that some constant fraction of the whole family intersect.
Theorem** (Fractional Helly theorem).**
Let and be a family of convex sets in . If of the -element subsets of have non-empty intersection, then some elements of intersect.
The first result in this direction was proven by Katchalski and Liu [228]. Starting with a family of convex sets, they assign to any subfamily the lexicographically minimum point in their intersection. The set of points lexicographically larger than a given point is convex, so Helly’s theorem ensures that the minimal point of the intersection of convex sets is also the minimal point in the intersection of some among them. A weak version of the above theorem, where the size of the intersecting subfamily is only guaranteed to be at least , then follows from a pigeonhole argument.
There are few settings in which fractional Helly theorems are known. On the one hand, Matoušek [259] proved, via a general sampling technique due to Clarkson [108, 112], that any set system with bounded VC dimension affords a fractional Helly theorem; his approach holds for other measures of complexity than the VC dimension [314]); we come back to the notions of VC dimension in Section 7.2. On the other hand, Bárány and Matoušek [45] established a fractional Helly theorem for lattices, including over the integers. It is surprising that they only have to check the non-empty integral intersection of a positive fraction of -tuples, instead of the expected -tuples of intersections.
The bound of in Theorem Theorem is sharp and was obtained by Kalai [216] and, independently, by Eckhoff [150], via a study of nerve complexes that led to a more general topological point of view. The nerve of a family of convex sets is the abstract simplicial complex with a vertex for every set in the family, and a simplex for every intersecting sub-family. Helly’s theorem and its fractional version easily translate in terms of nerves: the former states that the nerve cannot contain the boundary of a -dimensional simplex without containing the simplex, and the latter asserts that if the nerve contains a positive fraction of the -dimensional faces, then it must contain a simplex of dimension a positive fraction of . Kalai’s proof uses his technique of algebraic shifting [219] to study how the number of simplices of various dimensions behaves as the nerve is simplified through a sequence of -collapses, a type of filtration available to nerves of convex sets [371]. If a complex is -collapsible, then all its subcomplexes have trivial homology in dimension and above, i.e., it is -Leray. Kalai’s proof of Theorem Theorem extends to -Leray complexes (see [204, ]), and Alon et al. [21] further proved that families of subsets of that are closed under intersection and whose nerve is -Leray also admit weak -nets and -theorems; examples of such families include good covers (this follows from the Nerve theorem [75]) and acyclic covers [134]. Topological versions of Helly theorem have a further application in geometric group theory [161].
Fairly general topological Helly theorems can be derived from non-embeddability results via a construction reminiscent, again, of the setup of the KKM theorem. Let us illustrate the basic idea with five sets in the plane. If any four intersect and any three have path-connected intersection, then we can draw , the complete graph on vertices, inside the family by placing each vertex in the intersections of four sets (different vertices missing different sets) and connecting any two vertices by a path contained in the intersection of the three sets that contain them both. By a classical theorem of Kuratowski for planar graphs [139], there exist two edges that have no common vertex and that intersect. This intersection point must be in all five sets. An induction on the same idea yields that in a family of planar sets, where intersections are empty or path-connected, if every four sets intersect, then they all must intersect. In higher dimension, where all graphs can be drawn without crossing, the same approach can be combined with non-embeddability results derived from the Borsuk-Ulam theorem, e.g., the Van Kampen-Flores theorem which states that does not embed in [380, 164]; cf. [257, Chapter 5]. The discussion we present below on the topological Tverberg theorem is also connected to embeddability of complexes, e.g., the paper [66] proves that the topological Radon theorem implies the Van Kampen-Flores theorem.
The most general result in this direction [187] is that any family of subsets in admits a Helly-type theorem in which the constant that replaces in the case of convex sets is bounded as a function of the dimension and
[TABLE]
where denotes the -th reduced Betti numbers, over , of a space .
Non-embeddability arguments and the study of nerves offer two different pathways to topological Helly theorems. While the former allows more flexible assumptions, the latter offers more powerful conclusions in the form of a sharp fractional Helly theorem. It is not known whether the benefits of both approaches could be combined.
Open Problem 3.6**.**
Given and , is there a fractional Helly theorem for families of subsets in where for any and any ?
This open question relates to a more systematic effort to build a theory of homological VC dimension [220, Conjectures 6 and 7]. There are some recent results for planar sets with connected intersections (the case and ) [212].
Helly-type theorems are too many to list them all in this survey, but we wish to point out at least one more variation. Quantitative Helly’s theorem were introduced by Bárány, Katchalski, and Pach in [43]. In this family of Helly-style theorems one is not content with a non-empty intersection of a family, but the intersections must have measurable or enumerable information in the hypothesis and the conclusion. Typical measurements include the volume, the diameter, or the number of points in a lattice. Motivated by applications in optimization, in the last two years several papers have been published on this subject, both for continuous [80, 126, 302] and for discrete [12, 35, 105, 325] quantitative Helly-type theorems. For other recent Helly-type theorems see [28].
We next discuss the fifth remarkable theorem of our survey.
3.3 Tverberg
Tverberg-type theorems allow for the partition of finite point sets so that the convex hulls of the parts intersect. In its original form we have:
Tverberg theorem.
Any set of at least points in can be partitioned into subsets whose convex hulls all have at least one point in common.
Such a division in parts is often called a Tverberg partition; see Figure 9. The case is known as Radon’s lemma [320] and the case , but general , was proven by Birch [61], before Tverberg proved the general statement.
Tverberg’s first proof of his theorem [375] relies on a deformation argument: start with a configuration with a known Tverberg partition, and move the points continuously to the target configuration. This process is such that, while the number of Tverberg partitions may change, there will always be one present. A simpler proof consists in arguing that a partition of the point set minimizing an adequate function must be a Tverberg partition; this idea, which originates in Bárány’s proof of the colorful Carathéodory theorem, was gradually refined by Tverberg [376], Tverberg and Vrećica [377] and Roudneff [328] (Roudneff minimizes the sum of the squared distances between a point and the convex hulls of the parts). Another proof, due to Sarkaria [333], uses bilinear algebra to deduce Tverberg’s theorem from the colorful Carathéodory theorem; the idea behind Sarkaria’s proof was later made simpler (using explicit tensors instead of number fields) and more algorithmic by Bárány and Onn in [47]. Recently, Bárány and Soberón revisited these ideas and prove a new generalization of Tverberg’s theorem using affine combinations [49].
Just as for Carathéodory’s and Helly’s theorems, there is an integer Tverberg theorem. Its most recent version guarantees that any set of at least integer points in can be partitioned into parts whose convex hulls have a point of in common [129]. The proof of this upper bound goes as follows. From the integer Helly theorem, one can prove that any finite set of integer points has an integer centerpoint: a point such that for every hyperplane containing , one of its half-spaces contains at least points from . Using this centerpoint, it is not difficult to see that points can be partitioned into pairwise disjoint simplices all containing . We will see more about centerpoints in Sections 6.5.5 and 7.4.
The upper bound on the integer Tverberg number is not known to be sharp, and the best lower bound of is due to Doignon (this result was communicated in [152]). Recently in [124] the authors showed that the Tverberg number in is exactly when and improved the upper bounds for the Tverberg numbers of .
The special case of bipartition, i.e., , is called the integral Radon theorem. A sharper upper bound of and a lower bound of were established by Onn [307] (see Figure 10). Even low dimensional cases are hard. The only sharp bound known, also due to Onn, is for : any six integral points in the plane have an integral Radon partition. An upper bound of for the case was proven by Bezdek and Blokhuis [60].
Open Problem 3.7**.**
Determine the exact value of the integer Tverberg numbers. In particular, is the integer Radon number for less than 17? Is it bigger than 11?
More generally, there is the notion of an integer quantitative Tverberg number: any set of at least integer points can be partitioned into parts whose convex hulls have integral points in common [129]; similar results hold more generally for sets that are discrete, i.e., intersect any compact in only finitely many points, for instance the difference between a lattice and one of its sublattices. Recent improvements on the quantitative integer Helly theorem [35, 105] leads to sharper upper bounds for the Tverberg version.
Open Problem 3.8**.**
Determine tighter lower and upper bounds on the integer quantitative Tverberg numbers.
Tverberg’s theorem can be understood as stating that for any linear map from , the -dimensional skeleton of the simplex with vertices, into there must exist disjoint simplices whose images intersect. This reformulation invites the question, going as far back as 1979 (see ending of the important paper [39]), whether the same conclusion holds for all continuous maps. In other words, is there a topological Tverberg theorem? For , this is the question of non-embeddability discussed above in relation to topological Helly theorems.
Positive answers were obtained first for (the topological Radon theorem) by Bajmóczy and Bárány [39], then for prime by Bárány et al. [48], and for a power of a prime by Özaydin [309] and independently, but later, by Volovikov [385] and Sarkaria [334]. Matoušek [257, Chapters 5 and 6] offers an accessible introduction to the techniques behind the topological Tverberg theorem.
For , the proof of the topological Radon theorem uses the notion of deleted product of a geometric simplicial complex with itself, defined as
[TABLE]
Now, for contradiction, suppose there exists a continuous map with the property that points in distinct faces are mapped to distinct points. It induces another continuous map where The map commutes with the central symmetries of and , where the central symmetry of is the exchange of the two components. When is the boundary of the -dimensional simplex, is homotopy equivalent to (this is not trivial). Since the Borsuk-Ulam theorem prevents the existence of an antipodal map from to , the continuous will have two faces intersecting in its image, which gives us a contradiction.
More generally, to prove the topological Tverberg theorem when is a prime, one may start with a map from to with no -wise intersection and use it to build another map from an -fold deleted product , replace the antipodality by the action of the symmetric group, and apply a generalization of the Borsuk-Ulam theorem such as Dold’s theorem [143]. The case of a prime power is technically more involved. (Note that this outline leaves out some issues such as dimension-reduction considerations [341].)
By the late 1990’s, the widespread belief that Bárány’s question had a positive answer for every and was known as the topological Tverberg conjecture. It was only recently refuted by Frick [167], who completed an approach of Mabillard and Wagner [251] building on Özaydin’s work [309]. In a nutshell, Özaydin proved that an equivariant map from the adequate -fold product exists if and only if is not a prime power. Mabillard and Wagner proposed an isotopy-based approach to construct a map with no -wise intersection when such an equivariant map exists, but could only develop it in codimension larger than what the topological Tverberg conjecture allows. Frick overcame this codimension restriction, producing the first series of counter-examples to the topological Tverberg conjecture. The current state of affairs is that a counter-example is known for every that is not a prime power and for every . See the survey [42] for more details.
To conclude our discussion of “all things Tverberg”, let us highlight some natural variants inspired by Tverberg’s theorem for which only very partial results are known. One can find the most recent variants and extensions of Tverberg’s theorem in [50, 125, 316].
A tolerant Tverberg theorem, due to Soberón and Strausz [357], asserts that any set of points can be partitioned into parts such that, after deletion of any points, what remains is a Tverberg partition. This bound was improved to for and to for [284] (bound for is tight). Recently there have been two significant improvements, García-Colín et al. [179] gave an asymptotically tight bound for the tolerant Tverberg theorem when the dimension and the size of the partition are fixed. Later, in [356], Soberón used the probabilistic method to give another asymptotic bound that is polynomial in all three parameters. Still we can ask for precise values.
Open Problem 3.9**.**
What is the smallest number such that any set of points in has a Tverberg partition into parts that tolerates the deletion of points?
A related Carathéodory-type variation of Tverberg’s theorem [29] considers linear maps , assumes that is non-empty for every -dimensional edge of , and concludes the existence of disjoint faces in the simplex , of dimensions summing to and such that is not empty.
A conjectured relaxed version of Tverberg’s theorem, due to Reay, goes as follows. Denote by the minimum positive integer number such that any set of points in (not necessarily distinct) admits a partition into pairwise disjoint sets such that any size subfamily of has a nonempty intersection. Note that Tverberg’s theorem says ; Reay conjectured that the Tverberg constant is tight even for smaller values of :
Conjecture 3.10**.**
for all .
Note that if all intersect then they intersect -wise, so is at most and in particular is finite for any . Moreover, Helly’s theorem ensures that for every . The conjecture is known to hold for or , and some weaker bounds are known in several other cases; we refer the interested reader to [31] and the discussion therein.
Let us take this opportunity to mention another famous conjecture of flavor similar to Reay’s conjecture. A thrackle is a graph that can be drawn in the plane in such a way that any pair of edges intersects precisely once, either at a common vertex or a transverse intersection point.
Conjecture 3.11** (Conway’s Thrackle conjecture).**
For any thrackle, the number of edges is at most the number of vertices.
The conjecture is known to hold if all edges are drawn as straight line segments [155] (it is akin to Reay’s setup for ). We refer the interested reader to the recent progress of Fulek and Pach [174] and the discussion and references therein.
The next big open question was stated in 1979 by Sierksma (he offered an entire Dutch cheese as a prize for whoever can solve this problem). In unpublished mimeographed notes he conjectured about the number of distinct Tverberg partitions of a set of points guaranteed to exist for points in .
Conjecture 3.12** (Sierksma).**
Any set of points in has at least distinct Tverberg partitions into parts.
We do not state the lower bounds here, as they are a bit cumbersome, but lower bounds for the number of Tverberg partitions were first obtained when is prime by Vučić and Živaljević in [387]. They used topological tools to settle this. Hell showed that these bound also hold when is a prime power [205]. Later Hell, without any topology, provided better bounds for the case of the plane [206].
Last, but certainly not least, the colorful Tverberg conjecture was formulated by Bárány and Larman [44] some 25 years ago, but only a few results are known today (see [50, 68, 69, 398] and the references therein).
Conjecture 3.13**.**
Let be sets of points each. There exists a partition of into sets , of points each, such that every contains exactly one point from every and .
Further conjectures along the lines of the Bárány-Larman conjecture, with colorful, discrete and quantitative flavors were formulated by De Loera et al. [129].
3.4 Computational considerations
We continue our discussion of computational issues begun in Section 2.4. We remark that the Carathéodory and Helly theorems, in their classical real-valued versions from Section 3, are dual to each other and, essentially, if one has an algorithm to find the Carathéodory decomposition of a vector in terms of other vectors, one also has an algorithm for finding an intersection point for a family of convex sets. For Helly’s theorem, one wishes to find a point in the intersection of convex sets. The problem of finding such an intersection point can be thought of as a special case of the problem of minimizing a convex function over convex sets. We will see explicit cases for this problem later in Section 6, where the convex sets are explicitly given by constraints (i.e., equations and inequalities), but for now all that we need to know is that: (a) a whole range of different algorithms for solving such problems exist, (b) some of these algorithms are in fact efficient, and (c) depending on the type of input constraints (e.g., convex sets defined linear inequalities versus arbitrary constraints) one can be even more efficient [58, 76]. By the convexity assumption, a local minimum is also a global minimum and, thanks to the Helly and Carathéodory theorems, there are nice necessary and sufficient conditions for when we have found an optimum. We discuss more about this in Section 6.
Convex optimization problems have been classified in levels of increased computational difficulty and different specialized algorithms are available (e.g., Least squares, linear programming, Conic optimization, Semidefinite programming, etc). For example, Carathéodory’s theorem, in the simplest real-valued form presented in Section 3.1, can be formulated as a linear programming problem and Helly’s theorem for halfspaces is reducible to linear programming too. The good news is linear programs have efficient algorithmic solutions, both in theory and in practice [343]. Still, even if we move to the most general version of (real-valued) Helly’s theorem of finite family of arbitrary convex sets, the challenge is to solve a convex programming problem that has many types of algorithms. We cannot cover them here but we recommend [58, 76] for excellent introductions to convexity algorithms and computational methods.
Compare now the good news above to the bad news involving the integer and mixed-integer versions of the Carathéodory and Helly theorems. We are now in the realm of combinatorial, integer, and mixed-integer programming where, for the most part, the problems are not efficiently solvable. Solving combinatorial or mixed-integer optimization problems, that is, finding an optimal solution to such problems, can be a difficult task. Unlike linear programming, whose feasible region is a convex (polyhedral) set, in combinatorial problems, one must search a lattice of feasible points or, in the mixed-integer case, a set of disjoint half-lines or line segments to find an optimal solution. Thus, unlike linear programming, finding a global optimum to the problem requires us to relax, approximate, decompose the solution space, and sometimes we are forced to enumerate all possibilities.
As an example of the higher computational difficulty of discrete versions of the Helly and Carathéodory theorems let us look at the problem of computing a Hilbert basis. This was a particularly simple case for the integer version of Carathéodory’s theorem presented in Section 3.1. Alas, it was proved in [149] that deciding whether a given solution belongs to the Hilbert basis of a given system is coNP-complete. Thus, even in this tame case, the integral Carathéodory property is hard to realize computationally.
We chose to highlight the colorful Carathéodory theorem because it is so general and because it can be used to prove the original Carathéodory theorem and many other existence theorems in high-dimensional discrete geometry, such as Tverberg’s theorem or the centerpoint theorem (see Section 7 for details). While the original Carathéodory’s theorem can be cast as a linear program and thus a solution can be implemented in polynomial time, much less is known about the algorithmic complexity of its colorful version. More precisely, the algorithmic colorful Carathéodory problem is the computational problem of finding such a colorful choice of elements as described in the theorem. Despite several efforts in the past, the computational complexity of the colorful Carathéodory problem in arbitrary dimension is still open. In [275], Meunier et al. showed that the problem lies in the complexity class PPAD.
Open Problem 3.14**.**
What is the complexity of finding a colorful simplex under the hypotheses of the colorful Carathéodory theorem?
This question was formulated for the first time by Bárány and Onn in [47] and there they formulated a general family of related questions that come under the name colorful linear programming.
Meunier and Sarrabezolles [276] have shown that a closely related problem is PPAD-complete: given pairs of points and a colorful choice that contains the origin in its convex hull, find another colorful choice of points that contains the origin in their convex hull.
Since we have no exact combinatorial polynomial-time algorithms for the colorful Carathéodory theorem, approximation iterative algorithms are of interest. This was first considered in [47], but other researchers, e.g., [285], have approached this problem too.
Let us now speak about computational complexity of Tverberg’s theorem. Sarkaria’s proof of Tverberg’s theorem (later simplified by Bárány and Onn [46]) gives a polynomial-time way to compute a Tverberg partitions from a colorful Carathéodory choice with the origin in its convex hull. In this way, the computational issues about Tverberg’s theorem are closely connected to computational issues regarding the colorful Carathéodory theorem. Since one can calculate a Tverberg partition from a colorful selection of Carathéodory, one can show Tverberg’s theorem belongs to the class PPAD. One of the simplest, yet most frustrating, aspects of Tverberg’s result is that it is not clear how to find a Tverberg partition. So it is natural to ask:
Open Problem 3.15**.**
Is there a polynomial-time algorithm to find a Tverberg partition when one exists? That is, given points in , compute, in time polynomial in , a partition into parts with intersecting convex hulls.
Since finding an -Tverberg partition is an open question, approximate versions of Tverberg’s theorem are of interest. Mulzer et al. [285] designed a deterministic algorithm that finds a Tverberg partition into parts in time . This means that for every fixed dimension one can compute an approximate Tverberg point (and hence also an approximate centerpoint) in linear time. Rolnick and Soberón [324] proposed probabilistic algorithms for computing Tverberg partitions into parts with error probability , and with time complexity that is weakly polynomial in .
4 Games and fair division
Mathematics and the social sciences have had rich interactions since Condorcet’s seminal work on the analysis of voting systems. The relevance of (combinatorial) convexity and topology to this interdisciplinary research was first established in the 1940-50’s through the work of Nash, von Neumann, Gale, Shapley, Scarf and many others, and it has been confirmed in the following decades in the development of fair-division algorithms and computational social choice (see e.g., [78, 79, 116, 303] and the many references therein). This section shows how our five discrete theorems appear in these topics too.
4.1 Strategic games
Game theory studies a broad range of “games” that model situations where several agents collaborate or compete. Strategic games model the situations where players interact by each choosing from finitely many “strategies” to play and enjoy a payoff depending on the strategies chosen by all players (his or her choice included). Formally, each player is modeled by the pair where is a finite set of strategies available to him or her and a payoff function . A central theme in the theory of strategic games is the search for equilibria, where each player’s choice is the best response to the other players’ choices.
4.1.1 Nash equilibria
Formally, a Nash equilibrium in pure strategies is a choice of strategy for each player such that for and all
[TABLE]
Let us illustrate pure Nash equilibria with the max-cut game, where an arbitrary graph is fixed and each vertex represents a player. Each player chooses from two strategies , and his/her payoff function is the number of neighbors of in with a different strategy:
[TABLE]
Any bipartition of that maximizes the number of edges between the two parts, also called maximum cut, is a pure Nash equilibrium. Indeed, if a player could strictly increase his or her payoff by switching strategy, then this switch would increase the value of the cut.
Now, unfortunately not every -player game has a pure Nash equilibrium. For example, consider the matching penny game. Players Alice and Bob simultaneously select heads or tails of a coin. If the choice is the same, then Alice wins one penny and Bob loses a penny. If they choose differently, then Bob wins a penny and Alice loses a penny. Each player thus has a choice of two strategies, and the payoffs for each player can be recorded in two matrices ( and , for each player). Note that the pure strategies alone offer no Nash equilibrium as this is a winner-takes-all situation. (In the matching penny game, the payoff matrices can be put together to show ; this is an example of a zero-sum game, a notion to which we come back later.)
It turns out that equilibria always exist when considering a randomized choice of the strategies. We now allow more freedom to the players by making them choose not a single strategy, but a probability distribution over all their strategies. Once choices are made, a random strategy is selected for each player from his or her distribution, and the (random) payoff is determined. Each player then wants to maximize her or his expected payoff. Formally, a mixed strategy for player is a probability distribution on the set of pure strategies . For instance, in the matching penny game, this means that each player decides his or her move according to a biased coin flip and is free to choose the bias. Note that the mixed strategies include all the pure strategies as a special case too.
The set of all possible mixed strategies are the vectors that lie in the convex polytope . We define a product probability measure on by where . Therefore the expected payoff for this probability distribution for the -th player is
[TABLE]
The mixed strategies form a Nash equilibrium in mixed strategies if for each player and for all probability distributions on , modifying to does not increase the expected payoff with respect to the choices of other players, that is
[TABLE]
The literature has plenty of examples of two-player games [261, ]. Thinking about three or more players is more delicate, as illustrated by Nash’s three-man poker game [300, p. 293].
The existence of Nash equilibria for mixed strategies – the theorem, for which John Nash received the Nobel prize – is one of the most celebrated applications of combinatorial topology, following from Brouwer’s theorem. See [299, 300].
Theorem 4.1** (Nash’s theorem).**
Every -player game with continuous payoff functions has at least one Nash equilibrium in mixed strategies.
Nash’s original, very short, proof [299] makes strong use of the combinatorial topology and convexity arguments. It considers the set-valued function that maps each -tuple of mixed strategies to the set of -tuples where is a best response to . Because the probability distributions on are the points of the simplex , Kakutani’s theorem (we saw this in Section 2 after Brouwer) ensures this function has a fixed point, which is the desired equilibrium.
Nash gave a second proof using Brouwer’s fixed point theorem [300]. For this, Nash constructed a continuous function from the polytope , associated to the game above, into itself. For , we define component-wise as follows: , and each entry is equal to where
[TABLE]
where the ’s are the pure strategies available for player . Nash showed this function is a multivariate continuous map from the polytope into itself and thus, by Brouwer’s fixed point theorem, it must have at least one fixed point. Nash then went to show that any fixed point of this function is in fact a Nash equilibrium.
The polytope is actually a Cartesian product of simplices, sometimes called a simplotope. The special structure of simplotopes has been exploited for the computation of fixed-points (see [345, 373] and references there) and in the algebraic solution of equilibrium problems via polynomial equations [267]. There is one more reason why knowing is a polyhedron is useful. Despite its geometric beauty and intricacy, there are several modeling limitations with the notion of Nash equilibria. For example, the assumption behind Nash mixed strategies equilibria is that the choices of each player are independent of those of his/her opponents, but that may not hold. Alternative mathematical models that adjust the definition of Nash mixed strategies to allow correlated equilibria appear in [33]. In other words, the expression for the payoff function does not use anymore the easy product probability structure of a simplotope, but can have a more complicated polyhedral geometry. For example, it is known that all correlated equilibria are described by a finite set of linear inequalities and that it is non-empty, independently of Nash theorem. See [185] and its references.
4.1.2 Two-player games
Nash equilibria have been largely investigated in the area of algorithmic game theory, see for instance the introductory chapter of [303, ]. We only discuss here some of their relations to combinatorial topology and convexity.
The Nash equilibria for two players can be formulated as a linear complementarity problem, the theory of which subsumes both linear programming and two-player game theory (an introduction is in [115]). Let and denote the payoff matrices of the first and second players, respectively. By definition, a pair is a Nash equilibrium if and only if
[TABLE]
Here comes the linear complementarity formulation:
Proposition 4.2**.**
The pair satisfies (4.1) if and only if there exist , , , and such that
[TABLE]
Since all vectors are non-negative, the conditions on inner products imply that the supports of and are disjoint, and similarly for and , hence the aforementioned complementarity. The proof of Proposition 4.2 goes as follows. Start with a Nash equilibrium and let and ; the values of and are forced and the complementarity conditions follow from multiplying the first equation by and the second equation by . Conversely, the complementarity and yield that (and similarly ); the positivity of for any implies that is a best response to ; by a similar argument, is a best response to .
The linear complementarity formulation of Proposition 4.2 yields, after adequate rescaling, the standard method to compute two-player Nash equilibria: find a non-trivial solution (i.e., other than ) to the linear complementarity system
[TABLE]
The standard method to solve a linear complementarity system of this form is the Lemke-Howson pivoting algorithm [241] which operates on feasible bases. A feasible basis of a linear system with non-negativity constraints is a set of indices of columns whose induced sub-matrix has same rank as the system and defines a non-negative solution. The trivial solution gives feasible bases for \left(\begin{array}[]{c}\boldsymbol{v}\\ \boldsymbol{x}\end{array}\right) and for \left(\begin{array}[]{c}\boldsymbol{y}\\ \boldsymbol{u}\end{array}\right), and these bases are complementary, i.e., have disjoint supports. Pick an (arbitrary) element . By Carathéodory’s theorem (in the form of Proposition 3.1), contains another feasible basis for the system . Switching from to is our first pivot step. Remark that and are no longer disjoint, as they share . To remedy this, note that the part of corresponding to lost some element which is also absent from the part of corresponding to . We can thus make enter without further degrading the complementarity; Carathéodory’s theorem ensures that contains another feasible basis , and we pivot from to . The part of corresponding to lost some element . If then and are complementary, and determine our non-trivial solution, otherwise we continue pivoting by making enter , etc. until a pivot makes exit one of the bases; the pair at hands is then complementary.
The Lemke-Howson algorithm is guaranteed to terminate under the non-degeneracy assumption that is not a positive combination of less than columns of or less than columns of . This follows from a non-degenerate Carathéodory theorem (after Proposition 3.1): any point in the convex hull of points of that is non-degenerate, i.e., is not in the convex hull of some of them, lies in the convex hulls of exactly two -element subsets. (Note that the two non-degeneracy assumptions stated above are equivalent via the convex/conic change of viewpoint.) Now, for any pair of non-complementary feasible bases encountered by the algorithm, there is exactly one index not in . A pivot step makes enter either or ; in each case, the non-degenerate Carathéodory theorem yields exactly one other pair of feasible bases. Every pair of non-complementary feasible bases encountered by the algorithm thus has exactly two neighbors through pivot steps. Since the trivial solution has exactly one neighbor, there is no place where the walk can loop back. Remark that this algorithm gives an alternate (constructive) proof of the existence of a Nash equilibrium for two-player games.
The argument that proves termination also reveals that, from a computational complexity point of view, linear complementarity systems of the form of Equation (4.2) are in the PPAD class. Let us point out that the Lemke-Howson algorithm can be understood as a Sperner-type search for a fully-labeled simplex in a pseudomanifold. Let , where positive integers are understood as indices of columns of and negative integers are understood as (minus) indices of columns of . Consider the simplicial complex on whose maximal simplices are the union of the complement of a feasible basis of and the complement of a feasible basis of . The non-degenerate Carathéodory theorem spelled out above ensures that is a pseudomanifold without boundary. If every is labeled by its absolute value , the fully-labeled simplices correspond exactly to the complementary feasible bases.
As a byproduct of the linear complementarity formulation of Proposition 4.2, we also get that the problem of computing a Nash equilibrium for two players is well-posed from the point of view of computational complexity: if the input involves only rational data, there is an equilibrium that involves only rational data and has encoding size polynomial in the input size (see for instance the discussion in the survey of McKelvey and McLennan [266]). This is in sharp contrast with the case of three or more players: Nash’s three-player poker game [300] shows that a three-player game with finitely many strategies and rational payoff arrays may have a (unique) Nash equilibrium with irrational coordinates.
The Lemke-Howson algorithm has exponential complexity in the worst-case [336] and solving general linear complementarity problems is NP-hard [106]. The problem of computing a Nash equilibrium for two players is PPAD-complete [103] (see also [120]). The intractability for games with three or more players is even more stringent, as many decision problems are -complete, i.e., as difficult as deciding the emptiness of a general semi-algebraic set; this includes in particular deciding whether a -player game has a Nash equilibrium within -distance from a given distribution [340, Corollary 5.5] or the existence of more than one equilibrium or of equilibria with payoff or support conditions [180]. Behind this -completeness lurks a more daunting fact: Datta’s universality theorem [121] asserts that arbitrarily complicated semi-algebraic sets can be encoded as sets of Nash equilibria (formally: every real algebraic variety is isomorphic to the set of mixed Nash equilibria of some -player game). Whether the -completeness results stated above could follow from Datta’s proof is an interesting open question [340, Remark 5.6] .
Open Problem 4.3**.**
Can Datta’s universality theorem be improved to yield an efficient polynomial-time, reduction between any semi-algebraic set and the Nash equilibria of a game?
Another open question is whether both few players and few strategies per player already give rise to universality.
Open Problem 4.4**.**
Is there a universality result for the set of Nash equilibria of games with a constant number of players and a constant number of strategies?
4.1.3 Zero-sum games.
The two-player games where what is won by a player is lost by the other are called zero-sum games; this is the case when the payoff matrices satisfy in the formulation (4.1). In this special case, it is customary to consider that one player aims at maximizing the payoff while the other tries to minimize it. Nash’s theorem then asserts that there exist and such that
[TABLE]
This readily implies
[TABLE]
The only inequality that does not follow from Equation (4.3), the central one, holds in fact for arbitrary bivariate functions (see Section 6.4.2). Altogether, this yields the min-max theorem of von Neumann.
Theorem 4.5**.**
For any ,
[TABLE]
In game-theoretic language, the real number is the value of the game. Von Neumann’s theorem has a nice “asynchronous” interpretation: for any choice of a strategy by the minimizing player, the maximizing player can respond so as to ensure a payoff at least the value of the game. Moreover, if the maximizing player cares only about achieving the value of the game, the strategy will work regardless of what the opponent plays. (Of course, symmetric statements hold for the minimizing player.) In zero-sum games, in each of the (possibly many) Nash equilibria, every player gets the same payoff. This is specific to zero-sum games and fails already for broader types of two-player games. A classical result of Dantzig [118] says that the minmax identity of von Neumann’s theorem can be proved, without help of combinatorial topology, via linear programming duality and is thus polynomially solvable. We will discuss more about this in Section 6. More generally, if the rank of is constant then the problem is polynomial [222].
Formulation (4.1) suggests an approximate relaxation and the approximate Carathéodory theorem 3.5 provides a positive complexity result. We say a mixed strategy pair is an -Nash equilibrium if
[TABLE]
Intuitively, a mixed strategy pair is an -Nash equilibrium if no player can benefit more than , in expectation, by a unilateral deviation.
The case when is precisely the case of zero-sum games, for which we know efficient algorithms exist. Barman, using the approximate Carathéodory theorem presented in Theorem 3.5, provided an extension in [51].
Theorem 4.6**.**
Suppose that all entries of the payoff matrices lie in . If the number of non-zero entries in is at most , then an -Nash equilibrium of can be computed in time .
This, in particular, gives a polynomial-time approximation scheme for Nash equilibrium in games with fixed column sparsity . Moreover, for arbitrary bi-matrix games – since can be at most – the running time of this algorithm matches the best-known upper bound, which was obtained by Lipton, Markakis, and Mehta [242].
4.2 Two fair-division problems: cakes and necklaces
In various situations players are eager to divide goods in a “fair way”. There are several examples of such fair-division problems where our five theorems (or their variations) play a key role. We review some famous examples, all with combinatorial-topological proofs. Before we start we remark there are other interesting mathematical challenges arising in distributing resources, some we will not cover here, such as gerrymandering, which is the practice of drawing political maps to gain an advantage. See e.g., [355] for connections to the five theorems in this survey.
Cake cutting
A cake is a metaphor for a heterogeneous, divisible good, such as a piece of land or an inheritance. We consider the problem of dividing a cake between players in such a way that each player prefers his or her part to any other part. We call this envy-free. Let us point out that the literature about fair division, including this and other types of cake-cutting problems, is both old and huge; see e.g., [77, 78, 79, 323, 327, 362]. For example, one of the first envy-free division results was shown by Dubins and Spanier [145].
One setting where pieces are connected is the following: The cake is identified with the unit interval and a division of the cake into pieces is an -tuple , with for all and ; in other words, a division is a point in the -dimensional simplex . Here, represents the size of the -th piece, when ordered from left to right. The preferences of player are modeled by a function mapping each division to a nonempty subset of (indexing the pieces that he or she prefers). A division is envy-free if there exists a choice of pairwise distinct indices, one from each .
It is natural to assume that the set of preferences of player never contains the index of a piece of size zero, (i.e., all players are hungry) and it is common to suppose that the ’s are closed, that is if and for every , then . Stromquist [365] and Woodall [391] proved the following result independently.
Theorem 4.7**.**
Under the assumptions that all players are hungry and the preference functions are closed, there exists an envy-free division with connected pieces.
The original proof relies on the KKM theorem we saw in Section 2. An unpublished proof due to Forest Simmons was improved and adapted by Su [367]. The idea is to refine the usual derivation of Brouwer’s theorem from Sperner’s lemma. Take a sequence of triangulations of whose edge-length goes to [math]. Assign every vertex to a player in a way that every full-dimensional simplex of has a vertex assigned to each player; this may not be possible for any sequence of triangulations, but taking iterated barycentric subdivisions does the job. We label or color every vertex assigned to player by some (arbitrary) element from . The assumption that players are hungry ensures that this is a Sperner labeling. The limit of a converging subsequence of fully-labeled simplices is, by the closedness assumption, an envy-free division.
It may be disappointing in practice that one only gets an iterative infinite process converging to an envy-free division, but it has been shown in [366] that there exists no finite procedure, if you require each person to get a connected piece (i.e., the cake is cut by a minimal number of cuts). In fact, Aziz and Mackenzie [38] showed that there is in fact a bounded finite procedure for -person cake cutting bounded by a huge number of steps, but this would involving breaking the cake into a ridiculous number of pieces. Thus for now one cannot get minimal number of cuts in a finite procedure, but if you allow a lot of cuts, you get a division that is impractical as it destroys the cake. This is why an infinite process converging to an envy-free solution with a minimal number of cuts makes sense. Let us comment that the difficulty of the process is not so surprising perhaps. It is known that, in the polynomial-time function model, where the utility functions are given explicitly by polynomial-time algorithms, the envy-free cake-cutting problem has the same complexity as finding a Brouwer’s fixed point, or, more formally, it is PPAD-complete [137].
The polytopal version of Sperner’s lemma (Proposition 2.5 in Section 3) has many interesting game theoretic applications. Su [368] recently gave a simple elegant proof that Hex does not end in a draw using it. Cloutier, Nyman, and Su [113] applied the polytopal Sperner’s lemma to the multi-cake multi-player fair-division problems. In this type of problems the players have several cakes to choose from, but choices from one cake influence each other, e.g., for a player the amount of vanilla cake may influence how much chocolate cake to order, or after some vanilla cake the player may not want any chocolate cake. Cloutier et al. asked whether there exists an integer , independent of the preferences, such that there exists an envy-free division of the cakes not requiring to divide each cake into more than pieces, some of which are assigned to each of the players (some of the pieces can remain unassigned). Note that Theorem 4.7 for a single cake asserts that . They also used the polytopal version of Sperner’s lemma, and they proved the existence of and and that and . This means that two cakes can be divided into three pieces each in such a way that two players receive the pieces and everyone is satisfied with the fairness of division. Moreover they asked whether . Recently Lebert, Meunier, and Carbonneaux [240] have shown exists for any and its value is at most . They used again the polytopal version of Sperner’s lemma and an inequality between the matching number and the fractional matching number in -partite hypergraphs. Similarly, they showed exists and . Several interesting open questions remain, consider for example:
Open Problem 4.8**.**
Can the bound on be improved? Can one assure the existence of for all values of ?
Finally, there are other surprising variations of the cake-division theorem. Consider one where there are no cuts involved and one divides objects that are not physically divisible. Consider a house with rooms and a total rent amount to be divided among roommates. Assume that for each possible division of the rent amount each roommate can point to one or more rooms as preferred. Then, the theorem proved by Su in [367] states that there exists a division of the rent and an assignment of rooms to each participant, such that each player receives one of his/her preferred rooms. Similar assumptions on the preference function that we made for cake-cutting before Theorem 4.7 must hold again to make this happen, for instance, now it is assumed that each roommate prefers a free room over paying rent. Once more the proof of this theorem is grounded in Sperner’s lemma. In most results on fair-division it is assumed that no player is happy with an empty piece of cake, but imagine a part of the cake is undesirable (burnt cake anyone?) another recent variation in [277, 347] considers the possibility players may prefer an empty piece.
It is well-known that Gale’s colorful KKM theorem (see Section 2.2) has interesting applications in economics, e.g., for the existence of economics equilibria. Now, Asada et al. [32] used Gale’s colorful KKM theorem to prove an extension, by Woodall [391], of Theorem 4.7 and a similar extension of the rental-harmony result of Su. It turns out that there are envy-free cake divisions for any number of players, even if the preferences of one person remain secret! Say the situation is one where one of the cake-cutters (maybe the person celebrating a birthday) is not providing preference information, but still the cutting of the cake can be made without anyone being envious. Similarly, for deciding what the rent should be it suffices to consider the information of all but one of the roommates and still none of them will be jealous. The authors provided a rather nice existence proof of such fair divisions. Recently Frick, Houston-Edwards, and Meunier gave an iterative approximation algorithm for the solution [170].
Necklace splitting
Another fair division problem asks for the fair splitting, between two thieves, of an open necklace with beads threaded along the string. Here, fair means that, for each type of bead, the number of beads assigned to each of the thieves differ by at most one (say because the thieves are unaware of the value of the beads). Perhaps surprisingly, this can be achieved using only a few cuts (which turns out to be convenient should the string material be precious). Contrary to the classical statement, with this notion of fairness we do not need to add any conditions on divisibility.
Theorem 4.9** (Necklace theorem).**
There exists a fair splitting of a necklace that uses no more cuts than there are bead types.
The result is optimal as the number of bead types many cuts are sometimes necessary (e.g., if the beads of the same type are consecutive). Theorem 4.9 was first proved by Goldberg and West [189] and a simpler proof using the Borsuk-Ulam theorem was proposed by Alon and West [24], who also came up with the above popular formulation. More generally the challenge of finding a division of an object into two portions so that each of people believes the portions are equal is called consensus-halving. Note that necklace-splitting is a special case because different preferences are represented by different beads. In [352] Simmons and Su showed how a non-constructive existence result on consensus-halving can be obtained from Borsuk-Ulam. They also showed, by a direct application of Tucker’s lemma, how one can construct an approximate consensus-halving (up to a pre-specified tolerance level).
Later on, a combinatorial proof due to Pálvölgyi [312] used the octahedral Tucker lemma, instead of the Borsuk-Ulam theorem, for necklace-splitting. Here is a sketch of that idea. Let and denote, respectively, the numbers of beads and bead types, let denote the number of beads of type , and assume, ad absurdum, that any fair splitting uses more than cuts. Every vector defines a partial assignment of the beads to the two thieves (identified with and respectively). If every extension of into a complete assignment has at most cuts, then no such extension can be a fair splitting and there is some such that more than beads of type are assigned to the same thief by ; define to be the smallest such index , signed by the thief who gets more than beads of that type. If there is an extension of into a complete assignment with more than cuts, define to be the maximum number of cuts achieved by a completion, signed by the first component of that completion (that this sign is well-defined is straightforward). This map satisfies the condition of the octahedral Tucker lemma with and therefore cannot exist, contradicting the initial assumption.
The necklace splitting problem naturally generalizes from to any number of thieves. Alon [19] showed that cuts suffice (note that they are sometimes necessary). His proof first assumes to be prime and replaces (in a non-trivial way) the stronger form of the Borsuk-Ulam theorem due to Bárány, Shlosman, and Szűcs (see Lemma 4.1 in [19] or Statement B’ in [48]). The case of general follows easily by a recursive argument. (The original proof assumed to be divisible by but this was subsequently relaxed by Alon et al. [23], who obtained “fair roundings” via integrality properties of flows.) It is not known, however, whether Pálvölgyi’s proof can be adapted to the case with more than two thieves; perhaps an ingredient for this could be -generalizations of the octahedral Tucker lemma [273, 394]. Another open question [312] relates to the rounding: For those ’s not divisible by , can one choose the thieves who get the additional beads in the fair splitting? This is easily seen to be true for two thieves and it is also true for three [18]; it is open for .
Open Problem 4.10**.**
Is it possible to choose for each type the thieves who get and those who get in the fair splitting?
For two thieves, a fair splitting can be computed in linear time for and in time for [189]. Let us mention again consensus-halving, the problem of dividing an object into two portions so that each of players believes the portions are equal appears in many contexts. In [352] the Borsuk-Ulam theorem and Tucker’s lemma were used for this purpose. A well-known challenge of computational complexity was to decide whether the computation of a fair splitting is a PPA-complete problem [313]. This was just recently settled by Filos-Ratsikas and Goldberg [162, 163]. The fact that splitting necklace is PPA-complete implies that the algorithmic version of the octahedral Tucker is PPA-complete as well, but there is also a paper directly proving that the algorithmic Octahedral Tucker’s lemma is PPA-complete [136] (see Section 2.4).
The topic of fair-division is very active and once more we can only point to a few additional types of results. One may, for instance, explore necklace splittings with the added constraint that adjacent pieces of the necklace cannot be claimed by certain pairs of thieves; for example, Asada et al. [32] prove that four thieves on a circle can share the beads of the necklace, with the restriction that the two pairs of non-adjacent thieves will not receive adjacent pieces of the necklace. There are also several nice high-dimensional generalizations of (convex) splitting of booty; see [70, 133] and the references therein.
5 Graphs
Graphs are often used to model problems where pairwise interactions are prominent. This includes situations where graph-like structures are apparent, for instance road or train networks, or situations as in Euler’s famous problem on the bridges of Königsberg (although the curious reader may check that Euler’s original article does not use any graph-related notion, but argues purely in terms of words coding paths). In other situations, graphs are not evident, but exist implicitly; for instance, to describe time dependencies between tasks in scheduling problems.
Graph theory developed in many independent directions, driven both by applications (e.g., finding graph matchings to resolve assignments [344]) and deep structural questions (e.g., the graph minor theory [247]). Its interaction with (combinatorial) topology started in the mid-1970s with the proof by Lovász [62, Theorem 6.1] of the conjecture of Frank and Maurer that any -connected graph can be partitioned into subsets that induce connected subgraphs, have prescribed size (summing to ), and each contains a prescribed element. (This result was independently given a non-topological proof by Győry [198].) Lovász [246] followed up shortly after with an astonishing solution to the Kneser conjecture based on the Borsuk-Ulam theorem.
5.1 Chromatic number of graphs
A coloring of a graph by colors is a map from its vertex set into ; a coloring is proper if any adjacent vertices have different colors. The chromatic number of a graph is the smallest integer such that a proper coloring by colors exists. Graph colorings arise in applications such as frequence assignment [1] or scheduling [135].
Proving good lower bounds on chromatic numbers is usually a difficult task, as one needs to show that all colorings with fewer colors are improper. (In contrast, proving an upper bound on the chromatic number of a given graph only requires to exhibit one proper coloring.) In some cases, such as the perfect graphs discussed in Section 5.3, sharp lower bounds can be obtained from the existence of large cliques. Kneser graphs are archetypes where this clique criterion fails dramatically. The Kneser graph , where and are two integers, is the graph with vertex set – the -element subsets of – and where two subsets form an edge if they are disjoint. When , a natural coloring of assigns to every -element subset that intersects its minimal element and to all remaining subsets (see Figure 12: is also known as Petersen’s graph). Kneser conjectured this to be optimal in 1955.
Lovász approached the Kneser conjecture via a topological invariant. Given a graph , consider the simplicial complex that encodes its neighborhoods: has the same vertex set as , and a subset of the vertices forms a simplex whenever they have a common neighbor in . The key invariant is the (homotopy) connectivity of : if is -connected, then is not -colorable. In the case of Kneser graphs, this yields a lower bound that proves the conjecture. The interested reader may refer to the book of Matoušek [257, Section 3.3] and the surveys of Björner [62, Section 6] and Bárány [41, Section 5] for details.
The idea of associating a simplicial complex to a graph and to relate the chromatic number of the latter to topological properties of the former has been especially fruitful and there are now various complexes that can be used to obtain lower bounds for the chromatic number of a graph, see [262] and [233, Chapter 19] for surveys on that approach. Recently Frick, used Tverberg-type results to show bounds for chromatic numbers of (generalized) Kneser graphs and hypergraphs [168, 169].
The octahedral Tucker lemma emerged from a purely combinatorial proof of the Kneser conjecture due to Matoušek [258]. Ziegler [394] then showed that his method can combinatorialize other topological arguments for chromatic numbers. Let us illustrate this method on a bound due to Dol’nikov which deals with hypergraphs.
Recall that a hypergraph is a pair where is a finite set (the vertices) and a set of subsets of (the edges); in particular, hypergraphs whose edges all have size two are graphs. Given a hypergraph and a subset of its vertices, the hypergraph induced on has vertex set and edge set . A hypergraph is -colorable if can be colored so that no edge is monochromatic. To any hypergraph we associate the generalized Kneser graph where
[TABLE]
In particular, the generalized Kneser graph of the -uniform hypergraph with vertex set is the usual Kneser graph . The colorability defect of a hypergraph is the minimum number of vertices to be removed from to ensure that the remaining vertices can be -colored so that no edge of is monochromatic (edges with at least one removed vertex are discarded).
Theorem 5.1** (Dol’nikov [144]).**
For any hypergraph , the chromatic number of is at least .
The combinatorial proof of Theorem 5.1 goes essentially as follows. Consider a hypergraph with vertex set (identified with) . Given a proper coloring of by , we define a map as follows. Consider . If at least one edge of is entirely contained in or in , then we choose such an edge with smallest color and we set
[TABLE]
where records which of or contains the edge. (The sign is unambiguously defined because the coloring is proper.) If neither nor contains an edge of , then we define to be , where the sign is the first nonzero entry of . In the latter case, the edges of contained in induce a subgraph of that is -colorable, so . It then follows that suffices and that in either case, is mapped to disjoint sets of labels, which helps checking that satisfies the condition of the octahedral Tucker lemma. As a consequence, and the announced inequality follows.
Since every graph is isomorphic to some (actually, many) generalized Kneser graph, Theorem 5.1 provides a lower bound on the chromatic number of any graph. In the case of Kneser graphs, this bound is sharp. There exist refinements of the colorability defect that yield better combinatorial bounds [16]. Cases of equality for Theorem 5.1 are remarkable for reasons related to circular chromatic numbers [17] (see [393] for an introduction to circular chromatic numbers). Deciding whether the chromatic number of equals is a natural problem asked in [17]. Very recently this has been proved to be NP-hard by Meunier and Mizrahi (personal communication).
Some generalizations of the Kneser conjecture are still open. A -element subset of is -stable if for any we have (or, equivalently, if and are distance at least apart in ). Let denote the graph with vertices the -stable -element subsets of , and where two vertices span an edge if they are disjoint. The graph , known as Schrijver’s graph [342], has the same chromatic number as the Kneser graph; since is a subgraph of , this strengthens Lovász’s result. As a special case of a conjecture on hypergraphs, Meunier [273] proposed:
Conjecture 5.2**.**
For any and , the chromatic number of is .
Besides the case , the conjecture is known for all even [100] and for and sufficiently large [215]. Some progress has been made by Chen [101]. See also [169] for questions related Kneser hypergraphs.
A more systematic viewpoint recasts graph colorings as special cases of graph homomorphisms [199]. A homomorphism from a graph to a graph is a map such that for every edge the image is an edge of , that is . A proper coloring of with colors corresponds to a homomorphism from to the complete graph with vertices. More generally, associating with every vertex of a -element subset of so that adjacent vertices have disjoint subsets amounts to finding a homomorphism from to . The structure of homomorphisms of Kneser graphs remains to be elucidated, as is shown by the following, broadly open, conjecture.
Conjecture 5.3** (Stahl [359]).**
Let be integers and let and be such that with . There is a homomorphism if and only if .
For more details on partial progress on Stahl’s conjecture, see [199, ].
5.2 Colorful independent sets
A subset of vertices of a graph is independent if there is no edge between any pair of them. Independent sets are sometimes called stable sets. This notion is central in graph theory. For instance, a proper coloring of a graph can be understood as a partition of its vertices into independent sets, namely the sets of vertices with same color. The search for independent sets is also natural in applications such as the design of error-correcting codes [260, ].
Methods from combinatorial topology were particularly effective in finding independent sets with color constraints, in the spirit of the colorful theorems in combinatorial convexity. The following example was first stated explicitly by Aharoni et al. [7], who traced it back to the proof of a result of Haxell [203, Theorem 3] on hypergraph matchings.
Theorem 5.4**.**
Let be a colored graph with maximum degree . There exists an independent set of that intersects every color class of size at least .
(Note that the coloring of is not required to be proper.) It suffices to prove the statement in the case where every color class has size at least because deleting every vertex in a color class of size less than preserves independence and does not augment the maximum degree. The gist of the proof of Aharoni et al. [7] is to apply Meshulam’s lemma (Proposition 2.6) to the independence complex of , the simplicial complex consisting of its independent sets. The connectivity of can be controlled via a variety of domination graph parameters; this principle, which underlies some proofs of Meshulam [270], was made explicit by Aharoni et al. [7, Theorem 2.3] and given a detailed proof by Adamaszek and Barmak [2]. In particular, given a coloring of by , a subset and an integer , if no vertices of dominate the vertices with colors in , then is trivial. (Recall that dominates if every vertex of has a neighbor in .) Here, the condition holds for because dominating vertices requires at least vertices when the maximum degree is . By Meshulam’s lemma (Proposition 2.6), the independence complex therefore contains a colorful simplex; this is an independent set that intersects every color class.
Theorem 5.4 can be improved for graphs with more structure as the following example shows [18].
Theorem 5.5**.**
In any coloring of a path, it is possible to delete a vertex of each color so that the remaining vertices can be partitioned into two independent sets and such that for every color class .
(Again, the coloring does not need to be proper.) We sketch here a direct proof based on the octahedral Tucker lemma in a way reminiscent of Pálvolgyi’s proof of the necklace theorem (Theorem 4.9). We identify the vertex set of the path with and denote the color classes by . The existence of the two disjoint independent sets will be ensured via the notion of alternating subsequences, which has been useful in other similar contexts (e.g.,). A sequence of elements in is alternating if all terms are nonzero and any two consecutive terms are different. Given an , we denote by the maximum length of an alternating subsequence of .
The definition of the map to which we will apply the octahedral Tucker lemma requires the quantity s=\max\big{\{}\operatorname{alt}(\boldsymbol{x})\colon\boldsymbol{x}\in\{+,-,0\}^{n}\;\,\mbox{s.t.}\;I(\boldsymbol{x})=\varnothing\big{\}}, where
[TABLE]
Note that .
Consider a nonzero vector . We distinguish two cases. In the case where , we set , where is the maximum element in and where the sign is defined as follows. When , the sign is if and otherwise. When , the sign is if , and otherwise. In the case where , we set , where the sign is the first nonzero element of . Similarly to the proofs of Theorems 4.9 and 5.5, it can be checked that the map satisfies the condition of the octahedral Tucker lemma with . We have thus and there exists such that and . It implies that there exists such that and . Let and . They are both independent sets. Since , we have for all . The fact that leads then to for all and the conclusion follows.
Many statements about independent sets have analogues in terms of matchings, where a matching in a graph is a set of disjoint edges. This is natural since the matchings of a graph are the independent sets of its line graph, that is the graph in which vertices are the edges of , and where edges with a common endpoint are connected. Matchings are important for theory and applications (see [248] for an excellent book about matchings). For instance, many resource management problems, take after the following example: given a set of tasks, a set of workers, and for each task the list of compatible workers, assign to each task a different worker or report that no such assignment exists. The worker/tasks compatibilities can be modeled by a bipartite graph, so the question is whether there exists a matching that covers the vertex class modeling the tasks.
Colorful matchings still raise many questions, for instance the following well-known conjecture due to Brualdi [82] and Stein [361].
Conjecture 5.6**.**
If the edge set of (complete bipartite graph with vertices in each side) is partitioned into sets of size each, then there exists a matching in consisting of one edge from all but possibly one .
A famous conjecture of Ryser about Latin squares [332] asserts that if is odd, then under the same condition as the Brualdi-Stein conjecture, there exists a perfect matching intersecting each once.
5.3 Kernels in graphs
A kernel in a directed graph is a subset of the vertices that is independent (no two vertices of are joined by an arc) and absorbing (every vertex has an outgoing arc to a vertex ). Kernels naturally arise in certain combinatorial games, where they model the set of winning positions [386], or in stable matchings, as the stable matchings of a graph with preferences are the kernels of the associated line graph [10, ]. Kernels proved effective in revisiting classical questions and are for instance at the heart of the proof by Galvin of Dinitz’s conjecture on list colorings [10, ].
Not every directed graph has a kernel (consider a directed cycle of length three); this is in sharp contrast with the non-directed case, where the independent absorbing sets are the inclusion-maximal independent sets. As shown by a series of works by Richardson, Duchet, Meyniel, Galeana-Sánchez and Neumann-Lara [146, 148, 177, 322], a sufficient condition for the existence of a kernel is that each odd directed cycle has two chords whose heads are two consecutive vertices of the cycle. In particular, any acyclic directed graph has a kernel; this situation is actually what motivated, in the context of combinatorial games [386], the introduction of kernels. In general, however, deciding if a directed graph has a kernel is NP-complete [107].
Sperner’s lemma comes up in the following relation between kernels and perfect graphs. A graph is perfect if for all its induced subgraphs, including itself, the chromatic number is equal to the clique number. These are precisely the graphs for which the trivial lower bound on the chromatic number is sharp for them and all their induced subgraphs. The relation between kernels and perfect graphs is a special case of a conjecture of Berge and Duchet [57] proved by Boros and Gurvich [74].
Theorem 5.7**.**
Any orientation of a perfect graph with no directed cycle of length three has a kernel.
The original proof translates any directed graph into a coalitional game, where the players are the cliques, and the outcomes are the stable sets. Under the theorem’s assumptions, the game has stability properties that ensure, via results from coalitional game theory, the existence of a “non-rejecting” outcome, which translates into the desired kernel. A simpler and much more direct approach based on Scarf’s lemma was proposed by Aharoni and Holzman [8] and further simplified by Király and Pap [231] using Sperner’s lemma.
The proof of Király and Pap goes as follows. Consider an orientation of a perfect graph with no directed cycle of length . Let denote the polyhedron of (possibly negative) vertex weights summing to at most on every clique. This polyhedron has at least one extreme point (assigning to any maximal independent set of and [math] to the remaining vertices does the job), so it is pointed. Moreover, it has exactly independent extreme directions (the where is the unit vector associated to vertex ). Every facet of corresponds to a clique on which the weights sum to exactly . Since there is no directed cycle of length three, every clique has a source, i.e., a vertex that is absorbed by all other vertices of the clique. Label each facet by the source of the corresponding clique. Note that a facet containing is not labeled by . By Corollary 2.7, the polyhedron has an extreme point that is incident to facets of each label. Now, consider the weights on defined by . If we could find an independent set of intersecting every clique of weight , this set would also be absorbing: indeed, every vertex labels a facet incident to , so it is a source of a clique of weight one which intersects . The existence of follows from a classical lemma of Lovász: in a perfect graph, there exists an independent set intersecting every clique of maximum weight. (This commonly used lemma is perhaps difficult to find spelled out in this form; a standard way to prove it is to use perfectness and coloring for cliques of maximum cardinality, then apply the vertex replication lemma of perfect graphs [245], to allow rational weights, then generalize to real weights using linear programming.)
Theorem 5.7 ensures the existence of a kernel, but the proof does not give any efficient method to compute one.
Open Problem 5.8**.**
What is the complexity of computing a kernel in an orientation of a perfect graph with no directed cycle of length three?
6 Optimization
Broadly speaking, mathematical optimization develops mathematical tools for solving optimization problems. We illustrate in this section how the theorems of Carathéodory, Sperner and Helly and their variations provide original viewpoints on different aspects of this field.
6.1 Linear programming
A linear program (LP) asks for the minimum of a linear function under a set of linear constraints and is usually written
[TABLE]
Here, is a matrix, a vector of indeterminates, and vectors in and , respectively, and means that each row of is non-negative. Linear programs may come in different presentations, with in place of or possibly inequalities in place of equalities; these presentations are essentially equivalent [261, ]. Linear programming is by now a central tool in operations research as it allows to model a variety of resource management problems [261, ] and can be solved fairly effectively in practice. The theory of linear programming builds on the study of systems of linear inequalities. While this seems to be just small variation from linear algebra, linear programming was only systematized in the late 1940’s.
6.1.1 The simplex algorithm
Carathéodory’s theorem underlies the simplex algorithm, arguably the standard method to solve linear optimization problems.
On the one hand, Carathéodory’s theorem gives a way to discretize an a priori continuous problem. Indeed, the cone version of Carathéodory’s theorem ensures that if the system with admits a solution, then it admits a solution with support of size the rank of . Such a support, understood as a set of indices of columns of , is a feasible basis. A closer inspection of the proof of Proposition 3.1 reveals that the optimum of (6.1), if one exists, is attained on a solution supported by a feasible basis. Since a feasible basis determines a unique solution of , the optimum can be found in finite (but possibly long) time by enumerating feasible bases which are combinatorially described by their support.
On the other hand, Carathéodory’s theorem, in the form of Proposition 3.1, also explains the pivoting mechanics of the simplex algorithm. Suppose there exists an optimum, and that we have a feasible basis determining a solution . It turns out that if is not optimal, there exists such that increasing improves (i.e., decreases) the objective . The set contains another feasible basis, and it cannot define a worse solution than (again, a consequence of the proof of Proposition 3.1). Switching to that new basis is a pivot step. It is a non-trivial result of Bland that there exist rules for choosing non-cycling sequences of pivot steps, see [343, 11.3] and [261, 5.8]. Broadly speaking, the simplex algorithm starts by computing a feasible basis, and then it performs such pivot steps until no entry outside the basis can be used to improve the objective; the final basis then determines an optimal solution.
6.2 Integer programming
An integer program (IP) adds integrality constraints to linear programs by restricting all of the variables to take their values over rather than , for instance
[TABLE]
This variation arises naturally in the management of undivisible resources or yes/no decision making; the emblematic example is the knapsack problem which asks, given a set of objects with weights and values, for the subset of maximal value whose weight does not exceed a given threshold.
The relaxation of an integer linear program is the linear program obtained by forgetting the integrality conditions, as is (6.1) for (6.2). In general, the solution to the relaxed linear program provides a bound on the solution to the integer program, a lower bound in the case of (6.1) and (6.2). Linear programming and relaxation play a fundamental role in combinatorial algorithms; we refer the reader to the books [384, 390] for more detail.
What we just said applies also on mixed integer programs in which only some of the variables are required to be integer.
6.2.1 Sparsity of integer solutions
Carathéodory’s theorem readily measures the sparsity of optimal solutions. For example, Theorem 3.2 provides the following bound on the support of an optimal solution.
Corollary 6.1**.**
Let , and . The integer point of the polyhedron that minimizes has at most
[TABLE]
non-zero components, where is the largest of the entries of in absolute value.
Similar results have been used, for instance, for solving of bin-packing problems, see e.g., [154, 188]. See [13] for an application to the sparsity of optimal solutions and tighter bounds for special cases such as knapsack problems.
6.2.2 Graver bases
Another example of the influence of Carathéodory’s theorem is the use of Hilbert bases by Graver’s optimization methods. Although we present these ideas for integer programs, they apply more broadly, for instance to convex integer optimization problems, with respect to a convex objective function composed with linear functions, or convex separable functions, see [123, and ] and [308]. Consider the integer program 6.2 and assume, as is usually the case in practice, that . We can decompose the polyhedron into cones
[TABLE]
Each of these cones is pointed and rational, so it has a Hilbert basis [123, Corollary 2.6.4]. The union of these Hilbert bases is called the Graver basis of . Note that Sebő’s integer Carathéodory theorem (Theorem 3.3) ensures that any integer point in the polyhedron can be written as a non-negative integer linear combination of at most vectors from the Graver basis of . Moreover, Graver [191] established the following remarkable augmentation property: any non-optimal feasible solution of the integer program (6.2) can be improved by adding some suitable vector from the Graver basis of . Hence, any integer program of the form (6.2) can be solved by first computing a Graver basis for , then computing a feasible solution, and finally improving this solution by a greedy walk on the set of integer solutions, the candidate steps being provided by the vectors of the Graver basis.
The main obstacle to the practical use of Graver bases is their potentially exponential size. For general matrices, deciding if a given set of vectors is a Hilbert basis is already coNP-complete [149]. The good news, from the last decade of work, are that for highly structured matrices, such as those with regular block decompositions, Graver bases can be computed efficiently and are actually manageable for optimization (see details in Chapters 3 and 4 of [123] and the extensive presentation in [308]).
6.3 LP duality
An important idea for the study of linear programs is the notion of LP duality. This idea naturally arises from the question of certifying the quality of a solution to a linear program. For example, the objective function value of an optimal solution to
[TABLE]
can be seen to be at least by looking at the first constraint and at least by scaling the second constraint by . More generally, the solution of a linear program (P) can be bounded from below by the solution of a linear program (D), where
[TABLE]
and
(D)
s.t.
The linear programs (P) and (D) are said to be dual to one another; the variable of the dual program can be interpreted as the weights of a linear combination of the constraints of the primal program (and conversely). This relation, called weak linear programming duality, can be strengthened.
Theorem 6.2** (Strong Duality).**
Given two dual linear programs, if at least one is feasible, then they have the same optimal value.
The duality theory of linear optimization has many applications such as fast certification of solutions or primal-dual algorithms and in proving combinatorial theorems [343]. But it also plays a role in discrete geometry, for example in the proof of theorems [22]. In the following subsection we are going to prove Theorem 6.2 in an atypical way.
6.3.1 LP duality from the MinMax theorem
LP duality is the classical favorite approach to prove von Neumann’s MinMax theorem (Theorem 4.5) for two-player zero-sum games, as we mentioned in Section 4.1.2. Going in the other direction, Dantzig [118] proposed a deduction of the strong duality theorem from the minimax theorem; as we explain below, Dantzig’s proof required a detour, in some cases, via Farkas’ lemma, another result equivalent to Theorem 6.2. The impression of equivalence between minimax theorem, Nash equilibria for zero-sum two-player games and strong duality theorem nevertheless lingered and became a folklore theorem. It is only recently that Adler [5] filled-in the missing case to give a genuine direct equivalence between these three cornerstones.
Dantzig’s approach proceeds as follows. The weak duality already proves one inequality. The other inequality reduces to finding a solution of the system (6.3.1):
[TABLE]
This system rewrites
[TABLE]
so our task is to find a vector with positive last component and such that . Consider the zero-sum game with payoff matrix and let be a Nash equilibrium. The matrix has dimension , so one may pit each strategy against itself to see that the value of the game satisfies
[TABLE]
Since is antisymmetric, this implies that . Moreover, for any we have , so . Since , writing leads to the desired solution whenever . When , Dantzig concluded by a separate use of Farkas’ lemma (the “incompleteness” in his derivation). Adler was able to complete this missing case, without appealing to Farkas.
6.3.2 Totally Dual Integral polyhedra
Applied optimization models typically involve rational polyhedra, which are expressed as systems of linear inequalities with rational coefficients. An important question for computation is whether a rational polyhedron is integral, that is whether all its vertices have integer coordinates. Indeed, for integral polyhedra, integer optimization (which is typically very hard) becomes linear optimization (which is considered tractable). Let us see how Hilbert bases help when looking for rational polyhedra that are integral. In what follows, we consider a rational polyhedron with and rational.
Checking whether is integral is a finite process, as one can simply list all the vertices. The following structural result allows us to bypass this tedious enumeration in various situations. Observe that if is integral, then for every integral vector , the value is an integer (indeed, it is the inner product of two integral vectors). Surprisingly, a rational polyhedron is integral if and only if it satisfies this condition. This equivalence, due to Edmonds and Giles [153], is still not a practical way to detect integral polyhedra (the set of candidate vectors is infinite) but it suggests to look at things via duality. Indeed, the strong LP duality (Theorem 6.2) states
[TABLE]
so is integral if the vector is integral and the right-hand side minimization problem has an integral optimal value for every integral vector . (Note that in general integrality properties are not preserved through linear programming duality.) A system of inequalities is *totally dual integral *(TDI) if the right-hand side minimization problem above has an integral solution for every integral vector (for which the optimum is finite). A rational polyhedron that can be represented by a TDI system where is integral is thus integral. The converse is true and any integral polyhedron can be represented by a TDI system of inequalities (but which, in practice, may not be easy to find). Let us stress that TDIness is a property of the system of inequalities, not of the underlying polyhedron: Giles and Pulleyblank [186] proved that for every rational system of inequalities , there is a rational number such that is TDI. They also proved, using Carathéodory-style properties, that a system of inequalities is TDI if and only if for every face of , the rows of which are active in form a Hilbert basis for that supporting cone. This makes checking TDIness a finite process, but still not a practical one as checking whether a system of vectors forms a Hilbert bases is not efficient. See [149] and references therein for computational issues.
TDIness and related notions such as box-TDIness often shed new light on results in combinatorial optimization, for instance on the matroid intersection theorem. Consider two matroids and over the same ground set , understood as their sets of independent sets. Any matroid has an associated matroid polytope, obtained by taking the convex hull of its (indicator vectors of) independent sets. It turns out that the convex hull of the independent sets in coincides with the intersection of the matroid polytopes of both matroids. This is remarkable, as in general is different from . See [344, 41] for more on this topic.
A special case of TDIness allows for linear programming proofs of combinatorial results. A matrix is totally unimodular (TU) if every square submatrix has determinant in . For such a matrix, the polyhedron is integral for every integral vector . TU matrices give rise to TDI systems. They are completely characterized and are very important in combinatorics and optimization (see [349, 389]). Since the transpose of a TU matrix is TU again, the strong duality theorem in linear programming (Theorem 6.2) provides alternative proofs of the Kőnig theorem (the maximum cardinality of a matching in a bipartite graph is equal to the minimum cardinality of a set of nodes intersecting each edge), the Kőnig-Rado theorem (the maximum cardinality of a stable set in a bipartite graph without isolated vertices is equal to the minimum number of edges needed to cover all nodes), and the integrality of the Max-Flow-Min-Cut theorem. In all these three cases, the matrix is the node-arc incidence matrix of a directed graph, and it is an easy exercise to check that such a matrix is TU.
6.4 Convex optimization
Using linearization techniques, one may apply ideas from the theory of linear programming, and its duality, to more general optimization problems of the form
[TABLE]
where and ’s are differentiable functions (not just linear as before).
6.4.1 The KKT conditions from LP duality.
A milestone in mathematical programming is the following necessary optimality condition, due to Karush, Kuhn, and Tucker [76].
Theorem 6.3** (KKT condition).**
Let be a feasible solution of the problem (P’) such that the constraints are qualified at . If is a local optimum, then there are nonnegative real numbers such that
[TABLE]
The requirement that “the constraints are qualified” is a regularity condition on the feasible domain near . We do not spell out this rather technical condition but give a sufficient requirement. Call a direction feasible at if contains a segment of positive length with endpoint and direction . (This is where we are approximating: the adequate notion of feasibility is somewhat more flexible.) The constraints are qualified at if the closure of the cone of feasible directions coincides with the tangent cone at , that is . Even this coarser requirement may prove tedious to check, and several simpler sufficient conditions were investigated; the above criterion readily yields that affine constraints are qualified in any feasible point; another important case is that of convex differentiable constraints, which are qualified at any feasible point provided there exists a point satisfying strictly every constraint [76, 5.5].
The factors in Theorem 6.3 are the Lagrange multipliers. The strong duality theorem is classically equivalent to the following lemma of Farkas [343, 7.3].
Lemma 6.4** (Farkas’ lemma).**
Let be a real matrix and let be a vector. There exists such that if and only if for every such that .
Theorem 6.3 can be deduced from Farkas’ lemma via the following linearization argument. Let be a feasible solution of (P’) and write
[TABLE]
Let be a local optimum of (P’). Since is differentiable, for every direction feasible at . Moreover, since the constraints are qualified at , any direction satisfying is in the closure of the cone of directions feasible at , hence
[TABLE]
By Farkas’ lemma, this is equivalent to the existence of a vector in such that . Completing into by zeroes yields Theorem 6.3.
6.4.2 Strong duality in convex programming from the KKT conditions.
The KKT condition (Theorem 6.3) can, in turn, be used to prove a strong duality theorem for convex programming, generalizing Theorem 6.2. Let us introduce the Lagrangian function
[TABLE]
Since
[TABLE]
(P’) is equivalent to
[TABLE]
The same argument as in Equation (4.4) yields
[TABLE]
Finding the right-hand side term consists in solving the following dual program:
[TABLE]
This dual program always asks to maximize a concave function. In the case where Problem (P’) is a linear program, this notion of duality coincides with the LP duality introduced in Section 6.3.
Proposition 6.5** (Strong duality for convex optimization).**
Suppose that in (P’), and are convex functions. Suppose moreover that the constraints are qualified at every feasible solution. If (P’) has an optimal solution, then the dual program has one too and the optimal values of (P’) and (D’) coincide.
The proof goes as follows. Let be an optimal solution of (P’). By Theorem 6.3, there exists such that
[TABLE]
On the one hand, we have
[TABLE]
On the other hand, we have
[TABLE]
The map is convex because , so is a global minimum and . Together with the weak duality of Equation (6.4), this ensures that is an optimal solution for (D’).
6.5 Sampling approaches
Let us now consider optimization problems of the form
[TABLE]
where one minimizes a function over an intersection of subsets , the constraints. Such problems include linear programming (when linear and the are halfspaces), convex programming (when is convex and the are convex sets) or their integral or mixed analogues (via restrictions to ).
6.5.1 Witness sets of constraints
A first use of Helly’s theorem concerns the removal of redundant constraints defining an optimal solution. To begin, consider the linear program
[TABLE]
where (this form differs from those seen so far but is equivalent [261, ]). Assume that the problem is feasible and let denote its solution. The set of feasible solutions is an intersection of halfspaces, and their common intersection with , another halfspace, is empty. By Helly’s theorem, some of these halfspaces must have empty intersection, and must be one of them. It follows that we may drop from (6.6) all but some (carefully chosen) constraints without changing the solution. Recall that Helly’s theorem for halfspaces is dual to Carathéodory’s theorem; in the dual, this argument yields that an optimal solution can be realized by a feasible bases (as introduced in Section 6.1.1).
More generally, given a problem of the form (6.5), a subset is a witness set of constraints if
[TABLE]
and is inclusion-minimal for that property; in other words, a witness set is a non-redundant set of constraints that defines the same optimum as the entire problem. The above argument gives, mutatis mutandis, that the witness sets of any feasible mixed-convex programming over have size at most ; this bound increases by for unfeasible programs.
The relation between Helly’s theorem and witness sets extends beyond convexity, as observed by Amenta [26]. We say that a family admits a Helly-type theorem with constant if the non-empty intersection of every -element subset of implies the non-empty intersection of . Consider a minimization problem of the form (6.5), where is a function from a space to a space . We assume that is equipped with a total order , so that the minimization question makes sense. For any we let .
Proposition 6.6**.**
Let . If every family admits a Helly-type theorem with constant , then any witness set of constraints of (6.5) has size at most (actually, if the problem is feasible).
Note that we make no assumption on (not even continuity!), or . The proof for the feasible case goes as follows. Let and let denote the solution to (6.5). For define as the minimum of over and put
[TABLE]
On the one hand, for every , we have . On the other hand, every elements of intersect (because (6.5) is feasible) and intersects any elements of (by definition of ); thus, every elements in intersect, and the Helly-type theorem on ensures that . By minimality, any witness set thus has size at most .
6.5.2 Combinatorial algorithms for linear programming
Devising algorithms for linear programming with provably good complexity has been a major challenge for the past 70 years. The interior point method of Karmarkar [227] and the analysis of the ellipsoid method by Khachyian [230] only showed that the complexity of LP is polynomial in the number of constraints, the number of variables, and the bit complexity of the entries of the matrix and vectors and . A major question thus remains:
Open Problem 6.7**.**
Is there an algorithm that solves linear programming in time polynomial in the number of constraints and the number of variables, assuming that arithmetic operations on input numbers have unit cost?
(This is problem number nine in Smale’s list of open problems [354].) An algorithm with complexity polynomial in and in the unit cost model is called strongly polynomial. Although the simplex algorithm proves effective in practice, no choice of pivoting rule is known to ensure a number of step polynomial in the number of constraints and the number of variables; in fact, for every pivot rule whose worst-case complexity is established, that complexity is at least exponential in and (see [37, 117, 172, 232] and the references therein). Although no strongly polynomial time algorithm is known, partial progress was made through the 1980’s and 1990’s via combinatorial random sampling algorithms; this approach hinges on the fact that the bounded size of witness sets allows to throw away redundant constraints quickly.
Let us illustrate the basic idea of combinatorial random sampling algorithms in its simplest form, due to Seidel [348] (see also [122, ]). Consider
[TABLE]
where each is a halfspace in . Pick uniformly at random and let denote the solution to the linear program with the constraint removed. The idea is to compute recursively, then deduce from : we check in time whether belongs to . If , then , and if , then must belong to the hyperplane bounding and be the solution to the linear program
[TABLE]
This new linear program has constraints in variables and can be obtained from (6.7) in time . Altogether, the expected time to compute writes
[TABLE]
Observe that if and only if belongs to every witness set for (6.7). Since the size of witness sets is at most , the event occurs with probability at most when is chosen uniformly at random. Altogether, this recursion solves to , which is the running time of Seidel’s algorithm. In fact, Seidel’s algorithm builds on an idea by Clarkson, which we describe next.
The iterated reweighting method of Clarkson [109] consists of assigning a weight , initially set to , to every constraint and iterating a simple process: Sample constraints with probabilities proportional to their current weights and solve the problem on these constraints. If the solution is feasible, we are done; otherwise double the weights of all violated constraints and reiterate. It remains to be shown that, almost surely, the algorithm terminates. This can be seen by comparing the growth rates of the total weight of the system, and the weights of some witness set . On the one hand, every unsuccessful iteration must double the weight of at least one constraint in . On the other hand, as the constraints are chosen in each iteration with probability proportional to their current weight, the expected total weight of the constraints violated at any given iteration is [94]. Thus, after iterations, has weight but the total weight of all the constraints in the system is . Putting these two bounds together implies that the algorithm terminates, with high probability, within iterations. At each iteration one has to solve a linear program on constraints, which can be done in time , say by the simplex method, and compute the set of violating constraints, which takes time . This implies an overall running time of .
Clarkson’s approach was later improved by Matoušek et al. [263] to achieve an expected time complexity of . A similar bound was obtained independently by Kalai [217] via a randomized pivot rule for the simplex algorithm. Clarkson’s algorithm was subsequently derandomized; see [94] and the references therein for the latest developments.
6.5.3 Combinatorial abstractions of LP
Clarkson’s algorithm uses very little structure from linear programming, namely the abilities to solve a small-size problem and to decide if a given solution violates a given constraint. Surprisingly, there are many computational problems for which these two operations can be performed effectively to find a solution and are not “linear”. A simple example is the computation of the smallest enclosing circle of a finite point set . Here, the constraints are the points of , the candidate solutions are the circles, and a circle violates a point if it does not enclose it. Observe that subsets of points that minimally define their enclosing centers have size at most three, so in this case witness sets again have size at most three. It turns out that Clarkson’s algorithm readily applies to this problem. This is not an LP in disguise: a generic instance may have a witness sets of size two or three.
Various combinatorial abstractions of LP were studied in order to understand precisely what class of problems can be solved with the randomized approach we described before. Consider an abstract set of constraints numbered from to , and an objective function that associates to any set the value of the optimum when only the constraints in are considered. A natural black-box model allows to compute when has bounded size (independently of ), or to decide violations asking whether . It turns out that Clarkson’s algorithm can compute effectively in this abstract model under three assumptions [263]: (i) that be decreasing under inclusion, that is whenever , (ii) that be local in the sense that
[TABLE]
and (iii) that witness sets have bounded size, where a witness set is a minimal subset of with . Functions satisfying (i) and (ii) are called LP-type problems or generalized linear programming problems.
Any generic problem of the form (6.5) is LP-type; here by generic we mean that for every subset of constraints, achieves its minimum over the intersection of those constraints in only a bounded number of points. As noted by Proposition 6.6, controlling the size of witness set for such problems, and thus the effectiveness of Clarkson’s approach is a matter of Helly-type theorems. Later a generalization, called violator spaces [181], was shown to give the precise family of problems solved by Clarkson’s approach.
6.5.4 Chance-constrained optimization
Consider the problem of computing, given points in the plane, the smallest disk containing a given proportion of these points (say ). More generally, given a probability measure in the plane, and a positive number , one may consider the optimization problem
[TABLE]
where is a random point chosen from the probability distribution . This quantitative variation of the smallest enclosing circle problem, discussed above, is no longer LP-type but can still be solved effectively. The technique relies, again, on the fact that witness sets have bounded size and applies more generally to chance constrained problem (CCP). A CCP asks to optimize a function of a variable under constraints depending on a parameter , of the form:
[TABLE]
Here, is a convex function, the probability is taken relative to a measure on the space of parameters, is measurable with respect to , is convex for every , and is a convex set. This type of optimization problem naturally arises when modeling with uncertain constraints [350].
An approach to solve CCP, initiated by Calafiore and Campi [87, 88], is to sample from and solve the deterministic convex program
[TABLE]
For any , if , then the solution to is a solution to with probability at least [87, 88].
The proof in [127] goes as follows. For , let so that we are interested in ensuring . Each of the constraints is convex so any witness set has size at most . Now, for every we define
[TABLE]
Note that the decompose according to which are the indices of the witness constraints. Let denote the optimal solution of and the solution to the convex program defined by the constraints alone. The probability of failure is less than or equal to
[TABLE]
The summand corresponding to index set rewrites as
[TABLE]
The first factor is at most and each of the following factors is at most . Altogether, we get and the announced bound follows.
6.5.5 -optimization
There are many situations where one wants to optimize under convex constraints while restricting the solutions to belong to some set ; these are called -optimization problems. This allows to model complicated constraints like mixed-integer constraints (), sparsity constraints (e.g., compressed sensing), or complementarity constraints.
Several of the techniques described above generalize if the intersections of with convex sets of the ambient space admits a Helly-type theorem. We denote its Helly number by , for example . Proposition 6.6 yields that witness sets have size at most , and the reader can check that the analysis of chance constraint programs in Section 6.5.4 applies: the solution to is a solution to with probability at least if . In fact, the proof presented above differs from the initial argument [87, 88] and was found when generalizing CCP to the -optimization setup [127]. Helly-type theorems have been obtained for various sets [34, 128, 178, 318].
Let us now turn our attention to the problem of minimizing a convex function over an arbitrary subset . We assume that is given by a first-order evaluation oracle and that is nonempty and closed. The cutting plane method that we now present allows to approximate the solution. To allow us to control the quality of the approximation, we fix a finite measure supported on . The algorithm starts with a convex set that contains the solution in its interior and is such that . It then builds a sequence where each is also a convex set that contains the solution in its interior. Given , we select a point and compute and a subgradient . We set and define in a way that ensures that
[TABLE]
and that is non-increasing. We stop when is smaller than the desired error and return . This approach leaves many details unspecified, in particular the precise definition of and the way to choose the points . When and is the Lebesgue measure, a possible implementation is the classical ellipsoid method [230]. When and is the counting measure for , we obtain cutting plane algorithms for convex integer optimization problems. Another variant of this method which uses random sampling was explored by Bertsimas and Vempala [59].
The choice of the points in the cutting plane method is important. A particularly good choice are the Tukey centers. Given a vector and a point we let denote the halfspace . Consider the function
[TABLE]
A Tukey center is a point that attains the maximum value of . Lower bounds on , and therefore on the depth of Tukey centers (see Section 7.4), can often be obtained from Helly-type theorems. For instance, if and is the counting measure of a finite subset of , then the centerpoint theorem (7.12) ensures that . The proof of the centerpoint theorem from Helly’s theorem extends to the setting of -convexity: if is nonempty and closed and is finite and supported on , then [55]. For instance, Doignon’s theorem ensures that if and counts integer points inside a compact set , then . It turns out that choosing among the points maximizing , where is the restriction of to , gives the best running times among cutting plane algorithms for convex minimization over [55].
More notions generalize to -optimization. For instance there exists an analogue of the strong duality theorem for -convex optimization under some natural conditions [54]. We will meet Tukey centers again in Section 7.4, and we conclude this section with a challenge:
Open Problem 6.8**.**
What is the complexity of computing Tukey centers for given and ? E.g., can one compute an exact Tukey center of the integer points of a convex polytope in polynomial time in the input size?
7 Data point sets
In this section we consider some computer science results that are either applications of the main theorems or strongly related to them. There are simply far too many results for us to do justice to even a small number of them, thus we restrict ourselves to a few central themes, including classification, geometric shape analysis, and partitioning of points in -dimensional Euclidean spaces. These will involve an interplay between affine geometric and topological techniques, offering the usual mix of advantages and drawbacks of the two: affine tools – Radon’s lemma, Helly’s theorem, linear programming duality, simplicial decompositions – will imply fast algorithms, though apply to a restricted group of geometric objects. Topological tools – the Borsuk-Ulam theorem, Tucker’s lemma—yield broader structural statements, though, at this moment, settling the algorithmic feasibility of these methods remains a major open problem.
7.1 Equipartitioning: ham sandwich theorem and its relatives
We say that a hyperplane bisects a set of points if the two open halfspaces defined by contain at most points of . Note that if is odd, then a point of must necessarily lie on . The famous ham sandwich theorem is the starting point of a large number of results concerning equipartitioning of geometric objects with other geometric objects.
Theorem 7.1**.**
Let be finite point sets in . Then there exists a hyperplane that simultaneously bisects each , .
The theorem holds more generally for finite Borel measures that evaluate to zero on every affine hyperplane. All known proofs of Theorem 7.1 are essentially topological in nature. A classical proof follows from the Borsuk-Ulam theorem by identifying points of with hyperplanes in , and where the function encodes the “unbalance” of the point sets (more generally, measures) with respect to that hyperplane [257]. We now outline another proof of the ham sandwich theorem using Tucker’s lemma. The proof we present was found independently by Holmsen and by the third author (both unpublished).
For simplicity we will assume that the given point sets are in general position, and let . Each pair induces a sign pattern with being the sign of . We will apply Tucker’s lemma (Proposition 2.2) on an abstract simplicial complex induced by these sign patterns to show that there is a pair whose sign pattern is such that and each contain at most half of each . Then the hyperplane bisects each .
Denote by the partially ordered set of all achievable sign patterns endowed with the partial order (see Section 2 before Tucker’s octahedral lemma). It is a well-known result from oriented matroid theory that the order complex of is a triangulation of that is symmetric on its boundary. Suppose, for contradiction, that there is no hyperplane bisecting each . Given an in , we define to be , where is the smallest index such that either or contains more than half of the points of , and where is if it is the first set and is if it is the second. This map is clearly antipodal on the boundary of and labels the vertices of with the elements of . According to the Tucker lemma, there exists an edge of with . Without loss of generality, we assume that and that for some . By definition of , we have then and . Combined with , it implies that , a contradiction.
We will see further applications of the ham sandwich theorem later on, but for now we point out that it gives another proof of Theorem 4.9: given an open necklace with types of beads to be divided equally between two thieves, embed the beads of the necklace along a moment curve in , and use a hyperplane guaranteed by Theorem 7.1 to bisect each type of bead. As any hyperplane intersects a moment curve at points, splits the open necklace into pieces that can then be divided among the two thieves.
Note that from the above discussion, we have the following “computational hierarchy”: computing a solution to this variation of the octahedral Tucker problem is harder than computing a ham sandwich cut (note also that this implies that the latter is in PPA), which is harder than computing a solution to the fair splitting necklace problem. In particular, the Filos-Ratsikas and Goldberg’s paper [163] proves that computing the ham sandwich cut is PPA-complete (see Section 4.2).
As far as the efficiency aspects of Theorem 7.1 are concerned, a line bisecting two given point sets of total points in can be computed in time. In , the best algorithm to computing a ham sandwich cut for point sets in runs in time [244]; in fact the algorithm proposed presents a new proof of Theorem 7.1 that proceeds by induction on the dimension and thus is more amenable to efficient algorithm design.
By now there are dozens of variants of the ham sandwich theorem, and generalizations to other types of bisecting and bisected objects, We now present a few nice examples. For the ham sandwich theorem, see also [257, 397] and references therein.
One variation is called the center transversal theorem: given point sets in where , there exists an -dimensional affine subspace of such that any hyperplane containing has at least -th fraction of the points of on each side, for . In fact, it has been conjectured that the constant can be replaced by , see [86]. Theorem 7.1 is the case , and the case is the important centerpoint theorem that we encountered in the introduction and will visit again later. We refer the reader to the book [257] for many variants of this and related theorems in and higher dimensions. Here are now two famous conjectures.
Conjecture 7.2**.**
Let be a set of points in . Then there exists a set of four hyperplanes such that each of the resulting open regions of contains at most points of .
In the papers [64, 65], the authors showed that in it is indeed possible to find four hyperplanes that divide the set into 16 equal parts. See also [353]. We should mention here a related theorem of Yao and Yao [392] (see also Theorem 7.8): Given a set of points in , one can partition into regions such that the interior of each region contains at most points of , and any hyperplane intersects the interior of at most regions.
Next we consider another conjecture by Tverberg and Vrećica [377].
Conjecture 7.3**.**
Let be a given parameter, and let be finite point sets in . If for , then each can be partitioned into parts such that the sets can be intersected by a -dimensional affine space.
Note that Tverberg’s theorem is the case when in the above statement.
If one wants to partition more than point sets in , then hyperplanes are often insufficient; however the following important variant of the ham sandwich theorem, due to Stone and Tukey [364], shows that then polynomials of a sufficiently high degree can be used to do the partition.
We say that a -variate polynomial bisects a point set if it evaluates to negative on at most points of and likewise evaluates to positive on at most points of . Note that for the case of polynomials of degree one, this coincides with our earlier definition of bisection for hyperplanes.
Theorem 7.4**.**
Let be finite point sets in . Then there exists a -variate polynomial of degree such that bisects each , .
The idea is to reduce the above problem to the usual ham sandwich theorem in a suitably high dimension. As a -variate polynomial of degree has monomials (aside from the constant term), identify each such monomial with a distinct dimension of . Then each -variate polynomial of degree can be identified with a hyperplane in , where the coefficients defining the hyperplane in (i.e., the coordinates of the normal vector of the hyperplane) correspond to the coefficients of the corresponding monomials of . This also gives a mapping – called Veronese mapping – of the points in to , where the -th coordinate of a point is the value of the corresponding monomial on . One can now use Theorem 7.1 on the sets corresponding to the lifted points of to get a hyperplane in that bisects each of the lifted sets. Note that to use the ham sandwich theorem, we require and thus need to satisfy the constraint . The ham sandwich hyperplane corresponds to the required -variate polynomial in , of degree .
7.2 Parametrized partitioning of data via geometric methods
So far the partitioning statements have been of the type where the input geometric configuration precisely fixes the output type – e.g., given point sets in fixes the output of Theorem 7.1 to be a hyperplane. Or given point sets in in Theorem 7.4 fixes the output to be a polynomial of degree . Now we consider statements where, besides the input geometric configuration, one is also given an independent parameter and the complexity of the output is a function of both the geometric configuration as well as the value of . Thus one gets a hierarchy of output structures (varying with ), and one is free to choose the value of depending on the precise problem at hand. This turns out to be very useful for designing hierarchical data structures where one can pick the value of to maximize computational efficiency.
These kinds of partitioning statements – which we call parameterized spatial partitioning – have been a key theme in discrete and computational geometry for both algorithmic and proof purposes. Consider, for example, Hopcroft’s problem studied in the early 1980s: given a set of lines and a set of points , an incidence is a pair , where , , and the point lies on the line . Then given and , is there an efficient method to determine if there exists at least one incidence between them? It is not difficult to see that a spatial partition that either partitions the points or partitions the line (in a suitable sense) is useful for decomposing the original problem into several problems of smaller size; the current best algorithm with a running time of is based on such techniques. We refer the reader to [156] for details on this and other related results on Hopcroft’s problem.
For a more mathematical application in a similar setting, consider the following question posed by Erdős [155]: what is the maximum number of incidences between any points and any lines in the plane? Erdős observed that as the bipartite incidence graph between points and lines is -free, this is upper-bounded by O\big{(}n^{\frac{3}{2}}\big{)}. More generally, the maximum number of incidences between lines and points is . On the other hand, a set of points in a ‘grid-like’ configuration exhibits \Omega\big{(}n^{\frac{4}{3}}\big{)} incidences, and Erdős conjectured this to be, asymptotically, the right bound.
This question was resolved affirmatively by Szemerédi and Trotter by a complicated combinatorial argument [370]. We sketch here a beautiful and simple proof of this theorem by Clarkson et al. [110] that showcases the use of spatial partitioning for proving combinatorial bounds. Given a set of lines in the plane, suppose that for any parameter , there exists a partition of the plane into (possibly unbounded) interior-disjoint triangles such that each triangle , for , intersects at most lines of in its interior. Now one can partition the incidences between and into those for which the points lie on boundary of some triangle, and those for which the points lie in the interior of some triangle of . It is not difficult to see that the former can be only ; on the other hand, a triangle intersecting lines of in its interior and containing points of can contain incidences in its interior (via the graph-theoretic bound). This gives the overall number of incidences lying in the interior of triangles to be . Thus for any , the total number of incidences is bounded by , and setting gives the desired bound!
We now elaborate on this structural partitioning problem and its variations. The key behind the proof is the partition of into triangles, each of which intersects “proportionally few” lines of . More generally, in any dimension one can show the existence of a similar partition of hyperplanes [96].
Theorem 7.5**.**
Given a set of hyperplanes in and a parameter , there exists a partition of into interior-disjoint simplices such that the interior of any simplex intersects at most hyperplanes of .
Such a partition is called a -cutting of . Note that the bound of cannot be improved: each simplex can contain at most vertices induced by -tuples of in its interior, and so there must be simplices to account for all the vertices induced by .
The intuition behind Theorem 7.5 is as follows. Pick each hyperplane of into a random sample independently with probability (to be set later). Then , and so partitions into O\big{(}(|\mathcal{H}|p)^{d}\big{)} induced cells, each of which can then be further partitioned into simplices, to get a partition of into an expected total of O\big{(}(|\mathcal{H}|p)^{d}\big{)} simplices. Furthermore note that each such simplex intersects, in expectation, hyperplanes of in its interior. Setting gives the required statement. This argument was done ‘in expectation’, and it is non-trivial to convert it with the same asymptotic bounds to where each simplex is guaranteed to intersect no more than hyperplanes of .
The proof of Theorem 7.5 is usually presented in the more general abstract framework of the theory of -nets, whose setting we briefly describe now. Given a base set of elements , a set system on , and a parameter , call a set an -net for if contains at least one element from each with . In the case of -cuttings, the set is the set of hyperplanes , and if and only if there exists a simplex with . Then a -cutting can be constructed by taking an -net for with and partitioning into simplices using . As the interior of each simplex induced by does not intersect any hyperplane of , it can only intersect less than hyperplanes of , as desired.
Bounds on -nets have been extensively studied for set systems satisfying a combinatorial condition, called the VC dimension [382]: given a set system , define the projection of onto a set , denoted , as the set system . We say that a subset is shattered by if all subsets of can be realized by intersection with some set of , i.e., if . Then the VC dimension of is defined to be the size of the largest set that is shattered by .
The VC dimension plays an important role in the theory of set systems derived from geometric configurations due to the fact that the VC dimension of such systems is usually quite small. For example, consider the set system where is a finite set of points in , and the subsets in are derived from intersection with halfspaces; here if and only if there exists a halfspace such that . The VC dimension of this set system is ; in other words, given any set of points in , one cannot ‘separate’ all subsets of by intersection with halfspaces. This is an immediate consequence of Radon’s lemma (recall that Radon’s lemma is the case of Tverberg’s theorem for two parts, , namely that any set of points in can be partitioned into two subsets such that ): the convex hulls of the Radon partitions intersect, and thus cannot be separated by a hyperplane. On the other hand, any set of points in general position can be shattered by halfspaces in . By Veronese maps, this implies more generally bounded VC dimension for set systems induced by geometric objects of bounded algebraic complexity (see [256, Chapter 10]).
Returning to -nets, building on the work of Vapnik and Chervonenkis [382], Haussler and Welzl [202] showed the existence of small -nets as a function of the VC dimension of a set system.
Theorem 7.6**.**
Let be a finite set system with VC dimension at most . Then for any real parameter , there exists an -net for of size .
The power of this theorem derives from the fact that the size is independent of the number of elements in and the number of sets in . Combined with the VC dimension bound of on set systems induced on points in by halfspaces, Theorem 7.6 implies the existence of -nets of size , which has been shown to be optimal [238]. The bounds of Theorem 7.6 can be further improved for many geometric set systems, and recent work presents a unified framework for these bounds [95, 291, 383]. We refer the reader to the books [310, Chapter 15], [256, Chapter 10], [97, Chapter 4] and [255, Chapter 5] for a more detailed exposition on -nets and their many applications.
The theory of VC dimension fails to help in the construction of -nets when it is unbounded. A basic case is the set system induced by convex objects in ; namely given a set of points in and a parameter , one would like to show the existence of a small set , called a weak -net for induced by convex objects, such that any convex object containing at least points of must contain at least one point of . Note here that can be any set of points in , and is not just limited to being a subset of – hence the term “weak”.
An initial bound of on the size of was shown by Alon et al. [20] (their proof uses the colorful Carathéodory’s theorem together with Tverberg’s theorem; see the discussion about simplicial depth in Section 7.4), and this was improved to by Chazelle et al. [98]. This was improved even further, by logarithmic factors, by Matoušek and Wagner [264] whose elegant proof we outline now. Partition into equal sized subsets , for a parameter that is chosen optimally, such that any hyperplane intersects the convex hulls of subsets of . Let be a convex object containing points of . When intersects “few” sets of , the proportion of points of contained in some is higher than , hence can be hit by a set constructed inductively for . Otherwise, intersects many sets of . In this case, pick one arbitrary point from each set of intersecting . Let be the centerpoint of those points. Then must be contained in , as otherwise the hyperplane separating from must intersect the convex hulls of many sets in , a contradiction to the definition of .
For the case , Rubin [331] proved the bound of , where is an arbitrarily small constant.
Finding asymptotically optimal bounds for weak -nets induced by convex objects in is a tantalizing open problem. The best known lower bounds of [85] are quite far from the upper bounds. On the other hand, there are partial results that indicate the upper bounds that can be improved; e.g., it is known [292] that one can construct weak -nets from points of in : pick a set of points that form an -net for the set system induced by the intersection of halfspaces. Then adding points lying in Tverberg partitions of carefully chosen subsets of results in a weak -net, though of size . For a different formulation, one can also fix an integer parameter , and then ask for the minimization problem. Find such that for any set of points in , there exists a set of points such that all convex objects containing at least points of are hit by (see [293]).
Open Problem 7.7**.**
What is the asymptotically best bound for the size of the smallest weak -net for the set system induced by convex objects in ?
We next turn to the algorithmic aspects of spatial partitioning. There has been substantial work on improving the constants in the bounds on -nets, as they are directly linked to the approximation ratios of algorithms for the geometric hitting set problem [81, 158]. Given a set of points in and a set of geometric objects, the goal is to compute a minimum subset of that hits all the objects in . This can be written as an integer program, which is then approximated as follows: (a) solve the linear relaxation of the integer program (i.e., the linear program obtained by replacing the integral constraints with real ones), and (b) assign weights to the points of according to the linear program, and finally (c) compute a -net for the set system induced by , where is the value of the linear programming relaxation. As the weight of the points contained in each object of is at least by the linear program, is a hitting set for . The size of is then bounded by the size of a -net; e.g., if an -net exists of size for some constant , then has size at most – i.e., at most times the optimal solution. We refer the reader to the recent surveys [104, 298] for precise bounds on -nets and cuttings for various geometric set systems as well as the current-best algorithms for computing such nets.
Cuttings, together with linear programming duality (or alternatively, Farkas’ lemma), can be used to derive another important partitioning tool we have already encountered in the construction of weak -nets – the simplicial partition theorem [254].
Theorem 7.8**.**
Let be a set of points in , and a given integer parameter. Then there exists a partition of into sets such that for all and any hyperplane intersects the convex-hull of O\big{(}t^{1-\frac{1}{d}}\big{)} sets of .
We outline the general idea of the proof. Given , one can first “discretize” the space of all possible hyperplanes in into a finite set of hyperplanes, so that one only needs to construct a partition such that each hyperplane of intersects the convex hulls of sets of . Construct a -cutting for , consisting of at most simplices, and let be the collection of points of lying in each simplex of the cutting. Then, on average, for each , and furthermore the convex hull of each is intersected by at most hyperplanes of . This is, in a suitable sense, the dual of the statement that we want, where each hyperplane should intersect few cells in the cutting (and thus few convex hulls of sets in ). However we are not far off – the total number of intersecting pairs of hyperplanes of with convex-hulls of sets in is . In other words, the “average” hyperplane in intersects the convex hulls of sets of . Now LP duality [201] (or Farkas’ lemma [182]) shows the existence of the required partition.
7.3 Parametrized partitioning of data via topological methods
Now we move to more recent approaches to parameterized spatial partitioning. These methods use equipartitioning results such as the ham sandwich theorem and Theorem 7.4 to show the existence of polynomials that induce parametrized partitions of . The resulting statements are similar in spirit and use to those we saw earlier for affine objects. The main advantage of these newer approaches is that partitioning space with polynomials circumvents several difficult technical issues that arise when dealing with piece-wise linear objects. On the other hand, new computational challenges arise in the topological approaches, as they often do not lend themselves to efficient algorithm design.
We present now a “polynomial version” of Theorem 7.8, discovered by Guth and Katz [197]. For more on the impact of this theorem see the book [196]. In what follows given a polynomial , we denote by the zero set of .
Theorem 7.9**.**
Let be a set of points in , and let be a given parameter. Then there exists a -variate polynomial of degree such that each connected component of contains at most points.
Here the points lying in the components of play the role of the sets in Theorem 7.8. Observe that any hyperplane in intersects O\big{(}(\deg(f))^{d-1}\big{)}=O(r^{1-\frac{1}{d}}) different components of , quantitatively the same bound as in Theorem 7.8. In contrast with Theorem 7.8, these components are interior disjoint, though they will, in general, not be convex.
We sketch a proof: partition into two equal sized sets by a polynomial, say , of degree (using Theorem 7.1, for example). Then partition these two point sets and into four equal disjoint subsets using a polynomial, say , of degree via Theorem 7.4. Continuing, let be a polynomial of degree at most that equipartitions equal-sized disjoint subsets of , for . Note here that as long as , a hyperplane suffices for . The key observation now is that the polynomial formed by taking the product of all these polynomials – namely – is the required polynomial: as the zero set of is simply the union of the zero sets of all the ’s, each connected region of can contain at most points of . The degree of can be bounded as
[TABLE]
Theorem 7.9 gives another proof of the Szemerédi-Trotter theorem to bound the number of incidences between a set of points and a set of lines in the plane, this time by partitioning the points instead of the lines: apply Theorem 7.9 on with to get a polynomial , with . Note that each line of intersects components of and each of the components of contains at most points of . A simple calculation by summing up the incidences across each component induced by shows that the overall number of incidences is bounded by .
On the computational side, an efficient algorithm to compute the partition guaranteed by Theorem 7.9 was discovered by Agarwal et al. [6], who then used it to construct efficient data-structures for answering range queries with constant-complexity semi-algebraic sets as ranges, in time close to .
We now move on to our last topic in this section, a recent beautiful theorem of Guth [195] that gives a very general theorem that implies many statements that were previously regarded as unconnected.
Theorem 7.10**.**
Let be a set of -dimensional varieties in , each defined by at most polynomial equations of degree at most . For any parameter , there exists a -variate polynomial of degree , so that each component of intersects at most varieties of . The constant in the asymptotic notation depends on and .
Note that this theorem implies Theorem 7.9 by setting and implies Theorem 7.5 (strictly speaking, a polynomial version of it where is partitioned into the components of instead of simplices) by setting . See [63] for an exciting new alternate proof of Theorem 7.10 that extends it to a setting with several algebraic varieties.
We present a brief sketch of the proof of Theorem 7.10, which builds upon the proof of Theorem 7.9 in a natural way. The goal, as before, is to find polynomials , where and for each . Then the required polynomial will be
[TABLE]
as required. To see the idea behind the proof of Theorem 7.10, trace the proof of Theorem 7.9 backwards: the polynomials ’s were constructed using Theorem 7.4, whose proof used the ham sandwich theorem in a suitably high dimension, whose proof identified the coefficients of the polynomial with points of the sphere in a suitable dimension and then applied the Borsuk-Ulam theorem. Note that the ’s were constructed independently via an iterative argument, where first was used to partition the given point set into two sets; these two sets were then equipartitioned with and so on. This approach relied crucially on the fact that a point, outside the zero set of any polynomial , lies in precisely one cell induced by their product. This fact fails for -dimensional varieties when : assume one has constructed polynomials , , such that each component induced by has the same number of incidences with the varieties in . Then, even if the next polynomial equipartitions the incidences of each component of , that does not imply that each component induced by the polynomial will have the same number of incidences with .
The key idea here is to compute simultaneously. Thus we will identify, as before, the coefficients of each with some , but now instead of applying the Borsuk-Ulam theorem separately for each , we will consider the product of these spheres, and apply the following natural variant of the Borsuk-Ulam theorem to get the required polynomials in one step. For an integer , let , and for each , define the functions
[TABLE]
For each , let be a continuous function with the property that . Then there exists a point where for all .
We conclude this section with an open problem – the affine version of Theorem 7.10.
Conjecture 7.11**.**
For any set of -dimensional flats, any set of -dimensional flats in , and any integer , there exists a partition of into simplices such that (a) each simplex intersects O\big{(}\frac{n}{r^{d-k_{1}}}\big{)} flats of , and (b) each flat in intersects O\big{(}r^{k_{2}}\big{)} simplices.
7.4 Depth of point sets
Data science aims to understand the features of data sets. The goal of data depth measures is to generalize the idea of the statistical median of a set of reals to higher dimensions: the data consists of a finite set of points in , and the goal is to compute a point that is a “combinatorial center” of the data . As we will see, there are several natural ways to measure data depth, and they are related to each other in sometimes surprising ways.
Figure 14 shows a set of points in (circles), with “combinatorial centers” under three different measures: halfspace depth (cross), simplicial depth (box) and Oja depth (shaded disk). As the figure shows, the points for these three measures are geometrically close.
Given integers and , let be the point set in of size , with at least points placed at each of the vertices of the standard simplex. Slightly perturb each point so that all points are distinct, and in general position. This point set will be very useful for the remainder of this section, as it essentially captures both the intuition as well as the worst-case behavior with respect to many depth measures.
Halfspace depth.
Given a set of points in , the halfspace depth of a point with respect to is the minimum number of points of in any closed halfspace containing :
[TABLE]
Define the halfspace depth of as the maximum halfspace depth of any point in (this has also been called Tukey depth [374]). The separation theorem implies that any point outside has halfspace depth zero. It is a non-trivial fact, first shown by Rado in 1947, that points of high halfspace depth exist for every point set.
Theorem 7.12** (Centerpoint theorem [319]).**
Any set of points in has halfspace depth at least .
Recall such a point is called a centerpoint of and we saw its importance and generalizations in various places along this survey (e.g., in Section 6.5.5). It turns out the centerpoint theorem is optimal, in the sense that the bound cannot be improved is an example of a point set where it is not possible to do better. By now there are several proofs of the centerpoint theorem: using Brouwer’s fixed-point theorem [91], using Helly’s theorem [256], following from Tverberg’s theorem, and an elementary extremal argument by induction on the dimension [293]. Perhaps the following proof is the simplest: observe that any point hitting all convex objects containing greater than points of is a centerpoint, whose existence now follows from Helly’s theorem.
The centerpoint theorem and its generalizations have found several applications in combinatorial geometry, statistics, geometric algorithms, meshing, and related areas. A beautiful example is by Miller and Thurston [278], who showed that given a set of disjoint disks in the plane, there exists another disk intersecting disks of , and with at least disks of lying completely in the two connected components of induced by . To see this, use an inverse stereographic projection to lift the centers of the disks of to a set of points lying on a carefully chosen sphere in ; then, with high probability, the image of the intersection of a random hyperplane through the centerpoint of with the sphere is the required disk !
A point of highest halfspace depth with respect to is called a Tukey median of . It may not be unique. In general, the set of points of halfspace depth at least , for , forms a convex region called the -deep region of . It is the intersection of all halfspaces containing more than points of . Each facet of this region is supported by a hyperplane that passes through points of . Figure 14 shows that set of all such regions for the earlier point set. Martínez-Sandoval and Tamam, gave a generalization of Tukey depth in [253] which connects depth to fractional Helly theorems.
Algorithms. There has been considerable work on the algorithmic question of computing points of large halfspace depth. We first discuss the case in , which is by now settled. A centerpoint of points in can be computed in time [214], the key tool being the linear-time algorithm for computing ham sandwich partitions of two point sets in the plane. Chan [93] gave an time randomized algorithm for computing a point of the highest halfspace depth, i.e., a Tukey median for a set of points in the plane. The set of all depth contours of points in can be computed in time [279]. A real-time GPU-based algorithm for computing the set of all deep regions of a two-dimensional point set was given in [236]. Turning to , , the current-best algorithms for both computing any centerpoint and the highest depth point take time [93]. Clarkson et al. [111] presented an iterative method to compute approximate centerpoints: the algorithm constructs a -ary tree , where the leaves of are the input points, and each internal node represents the Radon point (namely, the unique point lying in the common intersection of the convex-hulls of the two Tverberg partitions of these points) of its children. This method was improved to the current best algorithm [286] which computes a point of halfspace depth at least in time . In fact, this method computes an approximate Tverberg partition; namely, a partition of into sets whose convex-hulls have a common intersection.
Open Problem 7.13**.**
Can a centerpoint of points in be computed in time?
Simplicial Depth.
A straightforward implication of Proposition 3.1 is that given a set of points in , any point lies in the convex hull of at least tuples of . In fact, any centerpoint must be contained in many more simplices spanned by -tuples of points in . Not surprisingly, the number of -tuples of whose convex hull contains is positively correlated to the halfspace depth of . This leads to the related depth measure of simplicial depth, first defined by Liu [243], which is the number of simplices spanned by points of containing a given point:
[TABLE]
The simplicial depth of is the highest simplicial depth of any point . As mentioned earlier, there is a close relation between halfspace depth and simplicial depth; the current best bound [388] shows that a point of halfspace depth has simplicial depth at least
[TABLE]
Bárány [40] showed that the colorful Carathéodory’s theorem together with Tverberg’s theorem implies that there always exists a point contained in at least simplices spanned by . Let be a constant such that any set of points has simplicial depth at least . The optimal dependency on is a long-standing open problem. Bukh, Matoušek, and Nivash [86] constructed points in so that no point in is contained in, up to lower-order terms, more than simplices defined by . Furthermore, they conjectured that this is the optimal bound.
Conjecture 7.14**.**
Any set of points in has simplicial depth at least .
For , a positive answer to the above conjecture was known already in 1984 by Boros and Füredi [73]. Bukh [84] gave a beautiful short proof: the required point set is the common intersection point of three lines in , having a common intersection and where each of the six induced cones contain at least points of ; the existence of such three lines follows via an elementary topological argument. For , an elementary argument shows that [53]. Using algebraic topology machinery, Gromov [192] improved the bound to the value . This bound for has since been improved even further by Matoušek and Wagner [265] to , and then by Král et al. [234] to . In fact, Gromov proves the bound for general , showing that
[TABLE]
His proof has since been simplified by Karasev [226]. Using the concepts of Gale diagrams and secondary polytopes, one can observe that the concept of simplicial depth is equivalent to a different maximization problem: given a set of points on the sphere , what is the triangulation that uses those points (but possibly not all) as vertices that has the largest number of -dimensional simplices? (see [131, Chapter 5] and the references therein).
We conclude this discussion of simplicial depth with colorful simplicial depth, which was introduced by Deza et al. [138]. Consider a set of points in partitioned into color classes . Suppose that has the property that the origin is in the relative interior of each , for . Recall from Section 3.1 that a colorful simplex is a simplex where each of the vertices comes from a different . While the colorful Carathéodory theorem asserts the existence of at least one colorful simplex containing , one can further ask about the number of distinct colorful simplices containing that must always exist. Define the colorful simplicial depth of , denoted , as the number of colorful simplices in containing . Deza et al. [138] proposed some lower bounds on the colorful simplicial depth, and they conjectured that if for , then . This was proven by Sarrabezolles [335]. The bound is optimal by the work of Deza et al. [138]. They also conjectured the following upper bound which was shown by Adiprasito et al. [4]: let be a point set in with for all . If no colorful simplex spanned by of dimension contains the origin in its convex hull, then .
Algorithms. For the case of the plane, computing the simplicial depth of a query point can be done in time [184], which is optimal. Aloupis et al. [25] presented an algorithm to compute a point of highest simplicial depth in in time . Using the fact that finding the highest simplicial depth is Gale dual to the problem of finding a maximum triangulation of points in the sphere, then one can set up an integer program to find a point of largest simplicial depth for any point configuration (see [131, Chapter 8]). A GPU-based algorithm for computing simplicial depth and colorful simplicial depth of point sets in the plane was given in [237].
Ray-shooting depth.
It turns out that the previous two measures – halfspace depth and simplicial depth – are further related to each other via an even more general depth measure, called ray-shooting depth. Given a set of points in , let be the set of all -simplices spanned by points of . Given a point and a direction , let be the half-infinite ray from in direction . Then the ray-shooting depth of a point is defined as
[TABLE]
The ray-shooting depth of is the maximum ray-shooting depth of any point in . The notion of ray-shooting depth was first introduced in Fox et al. [165], who proved the following using Brouwer’s fixed point theorem.
Theorem 7.15**.**
Any set of points in has ray-shooting depth at least .
Note that any point realizing the maximum ray-shooting depth must have halfspace depth at least and simplicial depth at least : let be a point with ray-shooting depth at least . Then any line through must intersect at least segments in , so both halfspaces defined by it must contain at least points. For simplicial depth, consider, for each point , the ray from in the direction . Then for every edge that intersects this ray, the triangle defined by must contain . Summing up these triangles over all points, each triangle can be counted three times, and so lies in at least distinct triangles spanned by .
The problem of showing the existence of a point with large ray-shooting depth is open in higher dimensions.
Conjecture 7.16**.**
Any set of points in has ray-shooting depth at least \big{(}\lceil\frac{n}{d+1}\rceil\big{)}^{d}.
Other notions of ray-shooting depth for convex sets, instead of points sets, were studied in [173].
Algorithms. The proof in [165] is topological and does not give a method to compute such a point. A combinatorial proof, together with efficient algorithms were obtained in Mustafa et al. [295], who showed how compute a point of ray-shooting depth at least in time .
Oja depth.
It turns out that ray-shooting depth is related to another older depth measure, the Oja depth of a point set, first defined by Oja [306]. Assume, without loss of generality, that . Then define the Oja depth of a point with respect to as
[TABLE]
The Oja depth of is the minimum Oja depth over all . It is easy to see that the Oja depth of is at least . The conjecture [99] is that the lower bound given by is essentially tight.
Conjecture 7.17**.**
The Oja depth of any set of points in is at most \big{(}\frac{n}{d+1}\big{)}^{d}.
The conjecture has been resolved only for the case by Mustafa et al. [297]. For general , it can be shown that the center of mass of has Oja depth at most [99]. This estimate can be improved via ray-shooting depth, as the Oja depth of any point which has ray-shooting depth at least is at least . The reason is the number of triangles spanned by pairs of points in and the point , containing any point , is at most the number of edges spanned by intersecting the ray , which is at most . Integrating over all gives the required bound. A calculation in gives the current-best bound [297]:
Theorem 7.18**.**
Every set of points in , , has Oja depth at most
[TABLE]
Algorithms. For the case , Rousseeuw and Ruts [329] presented a time algorithm for computing the lowest depth point, which was then improved to the current-best algorithm with running time [25]. A point of Oja depth at most can be computed in time [297]. For general , various heuristics for computing points with low Oja depth were given by Ronkainen, Oja and Orponen [326].
Regression depth.
The next depth measure, unlike earlier measures, is a combinatorial analogue of fitting a hyperplane through a set of points. Therefore it will be more convenient to state it in the dual. Given a point , let be its dual hyperplane, and for a set of points , let . Then define the regression depth of a point as
[TABLE]
The regression depth of a set of points in is the maximum regression depth of any point . It was introduced by Rousseeuw and Hubert [330], who showed that any set of points in has regression depth at least . Their proof is elegant: given the set of points, let be a partition of by consecutive -coordinate values, and where for . Then the required line is the ham sandwich cut of the two sets and . The optimal bound for general was discovered later.
Theorem 7.19** ([27, 223, 282]).**
Any set of points in has regression depth at least .
Given a set and a point , the closest point in to (if it exists) is denoted by . The proof in [223] deduces it from the centerpoint theorem: define the function that maps to a centerpoint of the set . This can be done so that is continuous and maps a sufficiently large ball to itself; then observe that the dual of any fixed point of is the required hyperplane.
Algorithms. The method of [330] gives a linear-time algorithm for computing a point of regression depth at least immediately, as it uses only ham sandwich cuts. A point of maximum regression depth in can be computed in time [381], improving upon an earlier time algorithm [239]. For , the best algorithm takes time [381] to compute a point of maximum regression depth.
The -centerpoint conjectures.
It turns out that many of the discussed depth measures are special cases of the following more general conjecture, first proposed by Mustafa et al. [296]. See also the related paper [225].
Conjecture 7.20**.**
For any set of points in and any integer , there exists a point such that any -half flat through intersects at least \big{(}\lceil\frac{n}{d+1}\rceil\big{)}^{k+1} of the -simplices spanned by .
The case corresponds to halfspace depth, to simplicial depth, and to ray-shooting depth.
It is not hard to show that given a set of points in , and an integer , there exists a point such that any -half flat through intersects at least
[TABLE]
-simplices spanned by . For simplicity assume that is a multiple of . The proof follows from the use of Tverberg’s theorem to partition into sets such that there exists a point with for all . Consider any -dimensional half-flat through , where is a -dimensional flat containing . Project onto a -dimensional subspace orthogonal to such that the projection of is a ray in , and and are projected to the point . And let be the projected sets whose convex-hulls now contain the point . Then note that the -dimensional simplex spanned by points intersects the ray if and only if the -dimensional simplex defined by the corresponding set in intersects the flat . Now apply the single-point version (i.e., given any point and sets in such that each contains the origin, there exists a -simplex spanned by and one point from each which also contains the origin) of colorful Carathéodory’s theorem to every -tuple of sets, say , together with the point at infinity in the direction antipodal to the direction of to get a ‘colorful’ simplex, defined by and one point from each , and containing . Then the ray must intersect the -simplex defined by the points of , and so the corresponding points of in span a -simplex intersecting . In total, we get of the -simplices intersecting . Another way is to use the simplicial depth bound of Gromov, that given any set of points in , there exists a point lying in -simplices. Now take any -half flat through . It must intersect at least one -simplex of each -simplex containing it, where each -simplex is counted at most times. This implies the stated bound.
In the plane () the centerpoint theorem, can be re-stated as follows: Given a set P of points in the plane, there exists a point such that if you take any line L passing through , and move it continuously until it arrives outside the convex hull of , then along this motion, the line will intersect at least points of . The following more general statement has also been conjectured in [296]:
Conjecture 7.21**.**
Given a set of points in , and an integer , there exists a point such that the following holds: Let be two -flats, such that and does not intersect the convex-hull of . Then any continuous motion family of -flats, starting at and ending at , must intersect at least -simplices spanned by .
In Conjecture 7.21, the case gives a ‘-’-flat moving to infinity, which can be treated as a stationary point. The validity of these conjectures for the planar case follows from the work of Gromov [192].
Acknowledgments
J. A. De Loera was partially supported by the LabEx Bezout (ANR-10-LABX-58). He is grateful to the labex BEZOUT and the CERMICS research center at École National des Ponts et Chaussées for the support received, and the enjoyable and welcoming environment in which the topics in this paper were discussed and built over several visits. De Loera was also partially supported by NSF grant DMS-1522158. Mustafa was supported by the grant ANR SAGA (JCJC-14-CE25-0016-01). Goaoc was partially supported by Institut Universitaire de France.
The authors are truly grateful to the two anonymous referees who provided a very large and very detailed set of comments, insights, and corrections. Their reading of this work was truly invaluable and it took a lot of effort and time. Very special thanks to Peter Landweber who gave us numerous corrections too. We are also grateful to the following colleagues who gave us support and many comments and references: Ian Agol, Imre Bárány, Pavle Blagojević, Boris Bukh, Sabrina Enriquez, Florian Frick, Alexei Garber, Tommy Hogan, Andreas Holmsen, Wolfgang Mulzer, Oleg Musin, Deborah Oliveros, János Pach, Dőmőtőr Pálvőlgyi, Axel Parmentier, Lily Silverstein, Pablo Soberón, Francis Su, Uli Wagner, and Günter M. Ziegler. We are grateful for their help.
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1[1] Aardal, K. I., van Hoesel, S. P. M., Koster, A. M. C. A., Mannino, C., and Sassano, A. Models and solution techniques for frequency assignment problems. Ann. Oper. Res. 153 , 1 (2007), 79–129.
- 2[2] Adamaszek, M., and Barmak, J. A. On a lower bound for the connectivity of the independence complex of a graph. Discrete Math. 311 , 21 (2011), 2566–2569.
- 3[3] Adiprasito, K., Bárány, I., and Mustafa, N. Theorems of Carathéodory, Helly, and Tverberg without dimension. Ar Xiv:1806.08725 (2018).
- 4[4] Adiprasito, K. A., Brinkmann, P., Padrol, A., Paták, P., Patáková, Z., and Sanyal, R. Colorful simplicial depth, Minkowski sums, and generalized Gale transforms. Int. Math. Res. Not. IMRN (2017), rnx 184.
- 5[5] Adler, I. The equivalence of linear programs and zero-sum games. Internat. J. Game Theory 42 , 1 (2013), 165–177.
- 6[6] Agarwal, P. K., Matoušek, J., and Sharir, M. On range searching with semialgebraic sets. II. SIAM J. Comput. 42 , 6 (2013), 2039–2062.
- 7[7] Aharoni, R., Berger, E., and Ziv, R. Independent systems of representatives in weighted graphs. Combinatorica 27 (2007), 253–267.
- 8[8] Aharoni, R., and Holzman, R. Fractional kernels in digraphs. J. Combin. Theory Ser. B 73 (1998), 1–6.