PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS
NEGARSADAT MIRSATTARI
Contents Dedication 2 1. Introduction 3 1.1. Basic Definitions 4 1.2. The Characterization Problem for Hypergraphic Sequences 4 1.3. ThePolytopeofHypergraphicSequences 6 2. CharacterizationsofDegreeSequences 8 2.1. Characterizations of Graphic Sequences 8 2.2. DegreeSequencesofOtherClassesofGraphs 10 2.3. PropertiesofHypergraphicSequences 12 3. ASetofNecessaryConditionsonHypergraphicSequences 14 3.1. AnOrderonClassesofHyperedges 15 3.2. Upperbounds and Proofs 18 3.3. More on Upperbounds 21 4. Polytope of Degree Sequences 24 4.1. Properties for Dn 24 4.2. PolytopeofHypergraphicSequences 28 4.3. Generalizing Koren Properties for Dn(k) 31 5. Reduction to the Problem of Lattice Points 35 5.1. The Lift Operation and Hidden Points 35 5.2. A Bijection and Reduction to Lattice points 37 5.3. Some Interesting Properties of the Threshold Hypergraphs 40 Appendix A. Preliminary Steps for Theorem 3.8 44 References 52
1 2 February 9, 2011
Dedication. To Navid, my brother, who proves me ∄ l ∈{limits} one can not rise beyond. PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS 3
1. Introduction Problems involving graphs first appeared in the mathematical folklore as puz- zles [1]. Since then, graphs have been useful tools in other fields of mathematics, such as algebra, number theory, geometry, topology, and geometry [2], as well as in other fields of computer science, such as scheduling, optimization, networks, and bioinformatics. Graphs have been also extensively used in other fields of science, such as electrical engineering, chemistry, psychology, and economics [1].
Using graphs is a powerful way to model systems in which relations involve a pair of elements. However, for more complicated systems, where relations involve sets of elements of size larger than two, we need concepts more general than the tra- ditional graphs. A generalized form of graphs which can model relations involving sets of elements, are called hypergraphs. In particular, a special case of hypergraphs, where all of the edges are distinct sets and of the same size k, are called k-families or k-(uniform) hypergraphs. k-hypergraphs have been the main focus of studies in the hypergraph theory.
Hypergraphs can be used to simplify the already existing graph models, as well as to model more complicated systems. Hypergraphs also facilitate generalizing concepts or systems modeled by graphs. Modeling a system, usually we are inter- ested in more information than just individual states of the system.
The degree sequence of a graph or hypergraph is the integer sequence which records the number of collections a given element is contained in. Degree sequences of (hyper)graphs and convex hull polytopes of the degree sequences provide us with extremely helpful tools to model and gather information about possible states of complicated systems, boundary or extreme possible states, transitions between dif- ferent states, and a wide range of statistics about the states of the systems.
In this paper, we investigate properties of degree sequences of k-hypergraphs as well as the properties of their convex hull polytopes. No characterization of these sequences is known. In Section 3, we introduce a set of necessary conditions on a n n sequence ϕ = (d1, d2, .., dn) ∈ N , where i=1 di ≡ 0 (mod k)tobe k-hypergraphic, as follows (Theorem 3.8): P z z dj ≤ αiMi + dj − γiMi /γz+1 αz+1 j∈S i=0 j∈T i=0 X X X X for all S,T such that S ∪ T ⊂ [n] and S ∩ T = ∅.
The Mi are parameters measuring the maximum possible number of hyperedges in the class of hyperedges containing αi vertices from S and γi vertices from T . We also introduce conditions similar or equivalent to this result that we use in proving properties of the convex hull polytopes of k-hypergraphic sequences.
For properties of the convex hull polytopes of k-hypergraphic sequences, first in Section 4.3, we generalize some known properties of the graphical case to hyper- graphic case, using our result for properties of degree sequences of k-hypergraphs. Then in Section 5, we reduce an open problem concerning holes in the polytope 4 February 9, 2011 of degree sequences (Problem 1.12) to a problem on lattice points (Problem 5.9.) Also we use this reduction to better understand these holes if they exist.
In the rest of this section, we will explain the basic definitions and the prob- lems under study. In Section 2, we review known characterizations or properties of several classes of graphs. In Section 3, we explain our results about the proper- ties of the degree sequences of hypergraphs. Appendix A contains the preliminary steps to this result. The known properties and our result about polytopes of degree sequences of hypergraphs form Section 4. In Section 5, we reduce a problem con- cerning the polytope of degree sequences to a problem on lattice points and use this reduction to gain some interesting properties of the polytope of degree sequences. 1.1. Basic Definitions. A hypergraph is a generalization of a graph, where an edge can connect any number of vertices. Formally, a hypergraph H is a pair H = (V, E), where V is a set of elements, called vertices or nodes, and E = {S1,S2, ..., Sm}, where each Si is a non-empty subset of V , called a hyperedge, block or link. A special case of hypergraphs, where all of the Si are distinct and of the same size k, are called k-families or (simple) k-(uniform) hypergraphs. A simple graph (without loops and multiple edges) is a 2-family.
Definition 1.1. (k-families or k-uniform hypergraph:) A “k-uniform hypergraph” K is a pair K = (V, E), where V is a set of elements and E = {S1,S2, ..., Sm} is a collection of distinct k-subsets of V .
In this text, for the purpose of simplicity we consider V = {1, 2, ..., n} = [n] and [n] therefore, for all i ∈ [m], Si ⊂ [n]. In particular, for k-families Si ∈ k . The degree of a vertex v, denoted by deg(v), is equal to the number of (hy- per)edges incident with the vertex v. The (ordered) degree sequence of a (hy- per)graph with vertices v1, v2, ..., vn is defined as d(H) = (deg(v1), deg(v2), ..., deg(vn)). Note that in this definition vertices are ordered and there is no non- increasing condition on the sequence. Also note that the domain of a degree is the natural numbers, i.e. N = {0, 1, 2, 3,...}. One can also look at the degree sequences as a n dimensional point or vector in Nn, where the value of the ith dimension is equal to deg(vi).
Example 1.2. H = ([6], E), where E = {{v1, v2, v3}, {v1, v2, v5}, {v1, v3, v6}, {v4, v5, v6}}, is a 3-family on [6]. Here, we will simply write E = {123, 125, 136, 456}. d(H) = (3, 2, 2, 1, 2, 2) is the degree sequence of H. Figure 1 shows a graphic rep- resentation of H.
1.2. The Characterization Problem for Hypergraphic Sequences. A non- negative integer sequence d = (d1, d2, .., dn) is called graphic if there exists a (sim- ple) graph G, for which d is the degree sequence. In this case, d is called realizable by G and G realizes or is a realization of d. A graphic sequence might be realizable by more than one graph. When a sequence is realized only by one graph (or its isomorphisms) it is called uniquely realizable. Similarly, PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS 5
Figure 1. Graphic representation of the 3-family from Example 1.2
Definition 1.3. (k-hypergraphic sequence:) A non-negative integer sequence d = (d1, d2, .., dn) is called “k-hypergraphic” if there exists a k-family H, for which d is the degree sequence.
Similarly, the terms realizable, to realize, realization, and uniquely realizable could be used for hypergraphic sequences.
Example 1.4. The sequence d = (3, 2, 2, 1, 2, 2) from Example 1.2 is the degree sequence for several 3-families , but it is not 6-graphic.
How could we recognize if a given integer sequence d = (d1, d2, ..., dn) is k- n hypergraphic? It is easy to see that i=1 dn ≡ 0 (mod k), but this is not sufficient. P Open problem 1.5. Give a characterization of k-hypergraphic sequences, d = (d1, d2, .., dn).
This problem has been studied extensively. In the literature, characterizations for graphic sequences have been proven in different formats. Also the characteriza- tions of some specific classes of graphs including bipartite graphs and r-graphs are know. However, in the general case of hypergraphs, the answer to this problem is not known.
In Section 2, we review some known characterizations and properties of several classes of graphs, including simple graphs, r-graphs and hyper graphs. In Section n 3, we introduce a set of necessary conditions on a sequence ϕ = (d1, d2, .., dn) ∈ N , n where i=1 di ≡ 0 (mod k) to be k-hypergraphic, as follows (Theorem 3.8): z z P dj ≤ αiMi + dj − γiMi /γz+1 αz+1 j∈S i=0 j∈T i=0 X X X X for all S,T such that S ∪ T ⊂ [n] and S ∩ T = ∅.
The Mi are parameters measuring the maximum possible number of hyperedges in the class of hyperedges containing αi vertices from S and γi vertices from T . We 6 February 9, 2011 also introduce propositions containing conditions similar or equivalent to this result that we use in proving properties of the convex hull polytopes of k-hypergraphic sequences.
1.3. The Polytope of Hypergraphic Sequences. Most well-known characteri- zations of graphic sequences (see Section 2.1) are systems of (redundant) inequal- ities. Each system could be used to determine a polytope in Rn. These polytopes and their properties, such as extreme points and adjacency of points, have been the subject of several studies. The construction of such polytopes has also been extended to k-hypergraphic sequences.
Definition 1.6. Define the set of all k-families on a given set [n] as Fn(k)= {G : G is a k-family on [n]}.
Definition 1.7. Define the set of integer points of all K-families in Nn as DSn(k)= {(d1, d2, ..., dn)|(d1, d2, ..., dn) is k-hypergraphic }, i.e. DSn(k)= { degree sequences of members of Fn(k)}.
n Definition 1.8. Define the polytope of k-hypergraphic sequences Dn(k) ⊂ R as Dn(k)= Convex hull of {DSn(k)}.
For the special case k = 2, the K is dropped and we write, Fn, DSn, and Dn.
Definition 1.9. (integer point:) Call each point d = (d1, d2, ..., dn) ∈ Dn(k) an “integer point” of Dn(k) if
(1) for all i ∈ [n], di ∈ N and n (2) i=1 di ≡ 0 (mod k). P Each K ∈ Fn(k) maps to an integer point d(K) in Dn(k), which is the vector representation of K’s degree sequence. We know that it is possible for an integer point in Dn(k) to be the degree sequence for more than one hypergraph in Fn(k). But is each integer point in Dn(k) the degree sequence for at least one hypergraph in Fn(k)?
Example 1.10. The degree sequence of the 3-family K from Example 1.2, d(k)= (3, 2, 2, 1, 2, 2), is an integer point in D6(3). This sequence is the degree sequence for several other 3-families on [6], including K′ = ([6], E′), where E′ = {126, 125, 136, 345}. All these 3-families on [6] are members of F6(3).
Definition 1.11. We call an integer point d ∈ Dn(k) a “hole” if there does not exist a k-family on [n], for which d is the degree sequence. i.e. n n d = (d1, d2, ..., dn) ∈ Dn(k) is a “hole” if d ∈ N and i=1 di ≡ 0 (mod k) , but d∈ / DSn(k). P PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS 7
Open problem 1.12. Are there any holes in Dn(k)?
For the case k = 2, it is known that there is no hole in Dn. However this prob- lem is still open for cases k > 2. We will review the known properties about Dn and Dn(k) in Section 4. In particular, in Section 4.3, we generalize some known properties of Dn to Dn(k), using our result for properties of degree sequences of k-hypergraphs. Then in Section 5, we reduce an open problem concerning holes in the polytope of degree sequences (Problem 1.12) to a problem on lattice points (Problem 5.9.) Also we use this reduction to prove some interesting properties of the polytopes of degree sequences. 8 February 9, 2011
2. Characterizations of Degree Sequences The problem of characterizing degree sequences of k-uniform hypergraphs, i.e. Problem 1.5, is a long standing open question in hypergraph theory [2]. In the literature, characterizations for graphic sequences have been proved in different formats. Also the characterizations of some specific classes of graphs including bi- partite graphs and r-graphs are known. However, in the general case, the answer to this problem is not known.
In the rest of this section, first we mention some of the well-known characteriza- tions of graphic sequence in Section 2.1 and some of the known characterizations or properties of other classes of graphs in Section 2.2. Then we review the partial results known for hypergraphic sequences in Section 2.3. We introduce new neces- sary conditions on hypergraphic sequences in Section 3.
2.1. Characterizations of Graphic Sequences. A nonincreasing sequence ϕ = n (d1, d2,...,dn), where i=1 di is even, is graphic if and only if the following equiv- alent characterizations hold: P
(1) The Erd¨os-Gallai characterization [6] : This is perhaps the most famous and the most cited characterization of graphic sequences.
k n di ≤ k(k − 1) + min{k, di} for all k ∈ [n], (2.1) i=1 X i=Xk+1 which is equivalent to the following system of linear inequalities [21]:
k n di − k(k − 1) ≤ (l − k)k + di for all k,l ∈ [n],l ≥ k. (2.2) i=1 X i=Xk+1 (2) The Ryser characterization [6, 25] : This is one of the oldest character- izations known for graphic sequences. Ryser defines f = max{i|di ≥ i} and
di +1, if i ∈ [f] d˜i = (di, otherwise.
ϕ = (d1, d2,...,dn) is graphic if and only if (d˜1, d˜2, ..., d˜n : d˜1, d˜2, ..., d˜n) is bigraphic. We will explain the definition of bigraphic sequences and their characterizations in Section 2.2 under item (2).
(3) The Berge characterization [2] : Berge defines the sequence d¯= (d¯1, d¯2,..., th d¯n) such that each d¯i is the sum of the i column of the n×n (0, 1)-matrix, th th which has the di leading terms of the i row equal to 1 except for the (i,i) term which is 0. Also the remaining entries are 0. The characterization is
k k di ≤ d¯i for all k ∈ [n]. (2.3) i=1 i=1 X X PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS 9
Example 2.1. If d = (3, 2, 2, 2, 1), then the (0, 1)-matrix is equal to 01110 10100 11000 11000 10000 and d¯= (4, 3, 2, 1, 0).
(4) The H¨asselbarth characterization [19, 26] : Define f = max{i|di ≥ i} and th the sequence d´ = (d´1, d´2,..., d´n) such that each d´i is the sum of the i th column of the n × n (0, 1)-matrix, which has the di leading terms of the i row equal to 1 and the remaining entries are 0. The characterization is k k di ≤ (d´i − 1) for all k ∈ [f]. (2.4) i=1 i=1 X X (5) The Bollob´as characterization [5] : k n k di ≤ di + min{k − 1, di} for all k ∈ [n]. (2.5) i=1 i=1 X i=Xk+1 X (6) The Gr¨unbaum characterization [15] : k n max{k − 1, di}≤ k(k − 1) + di for all k ∈ [n]. (2.6) i=1 X i=Xk+1 (7) The Fulkerson-Hoffman-McAndrew characterization [11, 13] : k n di ≤ k(n − m − 1) + di (2.7) i=1 i=n−m+1 X X for all k ∈ [n],m ≥ 0, and k + m ≤ n. (8) The Havel-Hakimi Algorithm [18, 17] : Havel and Hakimi, independently introduced an algorithm to recognize graphic sequences, which in known as the Havel-Hakimi algorithm. This algorithm inspired many algorithms for recognition of other classes of graphs, some of which we will see in Section 2.2. Havel-Hakimi algorithm: for n ≥ 1, the nonnegative integer sequence ϕ on size n is graphic if and only if ϕ′ is graphic, where ϕ′ is the list of integers obtained from ϕ by deleting its largest element ∆ and subtracting one from each of the next ∆ largest elements.
(9) The Koren characterization [21] : The Koren characterization is similar to the Fulkerson-Hoffman-McAndrew characterization, item (7), only the nonincreasing condition on the sequence has been relaxed.
di − dj ≤ |S|(n − 1 − |T |) (2.8) Xi∈S jX∈T for all S,T such that ∅ 6= S ∪ T ⊂ [n] and S ∩ T = ∅. 10 February 9, 2011
However, it is easy to see that if the inequality 2.8 holds for some S and T , where S consists of vertices with the largest degrees and T consists of vertices with the lowest degrees, then it would hold for any S′ and T ′, where |S′| = |S| and |T ′| = |T |. Also if the inequality does not hold for some S′ and T ′, then it would not hold for S and T , where S consists of the |S′| vertices with the largest degrees and T consists of the |T ′| vertices with the lowest degrees. Therefore, the Koren characterization is equivalent to the Fulkerson-Hoffman-McAndrew characterization. Example 2.2. Consider the sequence d = (3, 2, 2, 2, 1) from Example 2.1. Let us check the Koren characterization for S = {v1, v4} and T = {v3, v5}:
di − dj =5 − 3 ≤ |S|(n − 1 − |T |) = 2(5 − 1 − 2)=4, Xi∈S jX∈T which holds. Proposition 3.8, a set of conditions for degree sequences of hypergraphs, in the special case of (simple) graphs is the same as the Koren character- ization (see Example 3.11). Therefore, our result can be considered as a generalization of the Koren characterization.
2.2. Degree Sequences of Other Classes of Graphs. In addition to char- acterizations of degree sequences of (simple) graphs, characterizations of degree sequences of other special classes of graphs are also known. Here we restate some of this information. There have also been some studies on the properties of more complicated classes of graphs that we do not describe here. See [22] for an example.
(1) The Characterization of Degree Sequences of General Graphs: A loop increases the degree of a vertex by two. If we do not restrict our- selves to simple graphs, specifically allow multiple loops, it is easy to see that: Any nonnegative sequence of integers ϕ = (d1, d2,...,dn) is the degree se- n quence for some graph if and only if i=1 di is even.
(2) Condition on Bigraphic Sequences: P Let P = (p1,...pn) and Q = (q1,...,qm) be sequences of nonnegative integers. The pair (P : Q) is called bigraphic if it is the degree sequence of a bipartite graph. Conditions on sequences to be bigraphic are also known. For example, the Havel-Hakimi algorithm for bigraphic sequences [22, 29] is: The pair of nonnegative sequences of integers (P : Q) is bigraphic if and only if (P ′ : Q′) is bigraphic, where (P ′ : Q′) is obtained from (P : Q) by deleting the largest element pi of P and subtracting one from the pi largest elements of Q.
(3) Condition on Digraphic Sequences: Let P = (p1,...pn) and Q = (q1,...,qn) be sequences of nonnegative in- tegers. The pair (P : Q) is called digraphic if there exists a directed graph G~ for which P is the out-degree sequence and Q is the in-degree sequence. PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS 11
Knowing the Havel-Hakimi algorithm for (simple) graphs it is easy to rec- ognize digraphic sequences: The pair of nonnegative sequences of integers (P : Q) is digraphic if and only if (P ′ : Q′) is bigraphic, where (P ′ : Q′) is obtained from (P : Q) by deleting the largest element pi of P and subtracting one from the pi largest elements of Q other than qi.
(4) The Characterization of r-Graphic Sequences: A r-graph is a loopless undirected graph with at most r edges joining a pair of vertices. A sequence is called r-graphic if it is the degree sequence for a r-graph. Based on this definition (simple) graphs are 1-graphs and graphic sequences could also be called 1-graphic. The characterization of r-graphic sequences has been studied in several articles, including [9]. However, we only state the characterization of a generalization of r-graphic sequences, i.e. f-graphic sequences under item 5.
(5) The Characterization of f-Graphic Sequences: f-graphs are a generalization of r-graphs where the maximum allowed multiplicity of each edge e is defined by f(e). A sequence is called f- graphic if it is the degree sequence for a f-graph. In [13], Fulkerson et. al. introduced a simple characterization, but it holds only for graphs that satisfy a condition on odd-cycles. Tutte [28], presents an answer to the problem of characterization of f-graphic sequences while he is explaining more general results. His conditions are more complicated but they hold for all f-graphic sequences. Here we first give the definitions. Then we will write the Tuttes characterization in a format suitable for our problem. Let G be the graph with all the allowed multiple edges. Let T be any set of vertices of G. Tutte denotes by S(T ) the set of all vertices c of G having the following properties: (a) c is not an element of T . (b) Each edge of G having c as an end has its other end in T . If T does not include every vertex of G denote by k(T ) the number of com- ponents H of the induced subgraph G(T ) having the following properties: (a) H has more than one vertex. (b) The sum of the numbers di, taken over all vertices of H, is odd. If T is the set of vertices of G write k(T )=0. Any nonnegative sequence of integers ϕ = (d1, d2,...,dn) is f-graphic if and only if for any subset T of vertices the following inequality holds:
di + k(T ) ≤ di. (2.9) i∈XS(T ) Xi∈T (6) The Characterization of Threshold Sequences: A graph with vertex set {v1, v2,...,vn} is a threshold graph if there exist real weights c1,c2,...,cn n and t such that the 0 − 1 solutions of the inequality i=1 cixi ≤ t are precisely the characteristic vectors of the edges of the graph. The degree sequence of a threshold graph is called a threshold sequenceP . Golumbic [14] and Hammer et. al. [16] showed that: A nonnegative integer sequence d = (d1, d2,...,dn) is a threshold sequence if and only if d = d¯, where the (threshold) sequence d¯ = (d¯1, d¯2,..., d¯n) 12 February 9, 2011
such that each d¯i = n − 1 − di. Threshold graphs have been the subject of several studies. This class of graphs has been used to study the polytope of degree sequences of simple graphs as well as hyper graphs [24, 20]. We discuss hypergraphs and this work in more detail in Section 4.1.2. We should point out here that, the same characterization is valid for the general case of threshold ¯ n−1 k-hypergraphic sequences, where each di = k−1 − di. 2.3. Properties of Hypergraphic Sequences. The characterization of k-hypergraphic sequences is not known [3, 8, 12]. Even for the case k = 3 the problem seems to be difficult [8]. Currently, there is neither a polynomial time algorithm to test 3- hypergraphicness nor a proof that it is a NP-complete problem. Here we gathered some of the known conditions on 3-hypergraphic sequences.
(1) The Choudum Property [9] : Choudum gives a set of sufficient conditions for a sequence to be 3-hypergraphic. This set of conditions is a generalization of the Erd¨os-Gallai characterization, item (1) in Section 2.1. Let - Mk(ϕ) = max{|E(G)| : G is a k-graph on n vertices, such that degG(vi) ≤ di, for every i, 1 ≤ i ≤ n}, - x+ = max(0, x), and + + + - (x1,...,xn) = (x1 ,...,xn ). Any nonnegative sequence of integers ϕ = (d1, d2,...,dn) is 3-hypergraphic n if i=1 di ≡ 0 (mod 3) , and for all k ∈ [n], k n P k − 1 k − 1 d ≤ k + 2 min d , i 2 j 2 i=1 X j=Xk+1 k k + + M d − ,...,d − . (2.10) k k+1 2 n 2 Choudum has also introduced a formula to calculate Mk(ϕ) in polynomial time. Therefore, one can check this set of conditions in of polynomial time. However, it is easy to find 3-hypergraphs whose degree sequence does not satisfy 2.10.
(2) The Billington Properties [4] : Billington has introduced seven necessary conditions and one sufficient condition (algorithm) for sequences to be 3-hypergraphic. Here the notation of the original paper is changed to be consistent with the notation of other references and our paper: (a) Necessary Conditions: If a sequence ϕ = (d1, d2,...,dn) is 3-hypergraphic then the following seven conditions all hold n (i) i=1 di = 3q for some non-negative integer q. Moreover, q is the number of hyperedges in any realization of ϕ. P n−1 (ii) d1 ≤ 2 . (iii) d − d ≤ n−2 . 1 n 2 (iv) For each k ∈ [n], k d =2q + k . i=1 i 3 (v) For each k ∈ [n], k d = q + (n − 1) k − k+1 . Pi=1 i 2 3 P PROPERTIES OF DEGREE SEQUENCES OF k-UNIFORM HYPERGRAPHS 13
(vi) For each k ∈ [n], k k k − 1 n k d ≤ 3 − (d − ) +2 min d , i 3 k 2 j 2 i=1 X h i j=Xk+1 n + min dj , k(k − j − 1) . (2.11) j=Xk+1 (vii) For a hyperedge e = {vi, vj , vk}, let the weight of e be e = i+ j + k. For a sequence ϕ = (d1, d2,...,dn), let the weight of ϕ be n P wt(ϕ)= i=1 idi, and ϕmin(n, q) be the degree sequence of the 3-uniform hypergraph with the q least weighted hyperedges. If a P sequence ϕ = (d1, d2,...,dn) is 3-hypergraphic and 3q = mn+r where 0 ≤ r