ebook img

Spectra of general hypergraphs PDF

0.16 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Spectra of general hypergraphs

Spectra of general hypergraphs 6 1 0 2 Anirban Banerjee1, 2, Arnab Char1, and Bibhash Mondal1 r p A 1Department of Mathematics and Statistics 2Department of Biological Sciences 6 2 Indian Institute of Science Education and Research Kolkata ] Mohanpur-741246, India P S [email protected], [email protected], . h [email protected] t a m [ April 27, 2016 3 v 6 3 Abstract 1 2 Here, we show a method to reconstruct connectivity hypermatri- 0 ces of a general hypergraph (without any self loop or multiple edge) . 1 using tensor. We also study the different spectral properties of these 0 hypermatrices and find that these properties are similar for graphs 6 1 and uniform hypergraphs. The representation of a connectivity hy- : v permatrix that is proposed here can be very useful for the further i development in spectral hypergraph theory. X r a AMS classification: 05C65, 15A18 Keywords: Hypergraph, Adjacency hypermatrix, Spectral theory of hyper- graph, Laplacian Hypermatrix, normalised Laplacian 1 Introduction Spectral graph theory has a long history behind its development. In spectral graph theory, we analyse the eigenvalues of a connectivity matrix which is uniquely defined on a graph. Many researchers have had a great interest to 1 study the eigenvalues of different connectivity matrices, such as, adjacency matrix, Laplacian matrix, signless Laplacian matrix, normalised Laplacian matrix, etc. Now, a recent trend has been developed to explore spectral hypergraph theory. Unlike in a graph, an edge of a hypergraph can be con- structed with more than two vertices, i.e., the edge set of a hypergraph is the subset of the power set of the vertex set of that hypergraph [20]. Now, one of the main challenges is to uniquely represent a hypergraph by a connectivity hypermatrix or by a tensor, and vice versa. It is not trivial for a non-uniform hypergraph, where the cardinalities of the edges are not the same. Recently, the study of the spectrum of uniform hypergraphs becomes popular. In a (m-) uniform hypergraph, each edge contains the same, (m), number of ver- tices. Thus a m-uniform hypergraph of order n can be easily represented by a m order n dimensional connectivity hypermatrix (or tensor). In [7], the results on the spectrum of adjacency matrix of a graph are extended for uni- form hypergraphs by using characteristic polynomial. Spectral properties of adjacency uniform hypermatrix are deduced from matroids in [14]. In 1993, FanChung defined Laplacian of a uniform hypergraph by considering various homological aspects of hypergraphs and studied the eigenvalues of the same [5]. In [9, 10, 16, 17, 11], different spectral properties of Laplacian and sign- less Laplacian of a uniform hypergraph, defined by using tensor, have been studied. In 2015, Hu and Qi introduced the normalised Laplacian of a uni- form hypergraph and analysed its spectral properties [8]. The important tool that has been used in spectral hypergraph theory is tensor. In 2005, Liqun Qi introduced the different eigenvalues of a real supersymmetric tensor [15]. The various properties of the eigenvalues of a tensor have been studied in [3, 13, 4, 21, 22, 12, 18, 19]. But, still the challenge remains to come up with a mathematical frame- work to construct a connectivity hypermatrix for a non-uniform hypergraph, such that, based on this connectivity hypermatrix the spectral graph theory for a general hypergraph can be developed. Here, we propose a unique rep- resentation of a general hypergraph (without any self loop or multiple edge) by connectivity hypermatrices, such as, adjacency hypermatrix, Laplacian hypermatrix, signless Laplacian hypermatrix, normalised Laplacian hyper- matrix and analyse the different spectral properties of these matrices. These properties are very similar with the same for graphs and uniform hyper- graphs. Studying thespectrum ofauniformhypergraphs couldbeconsidered as a special case of the spectral graph theory of general hypergraphs. 2 2 Preliminary Let R be the set of real numbers. We consider an m order n dimensional hypermatrix A having nm elements from R, where A = (a ),a ∈ R and 1 ≤ i ,i ,...,i ≤ n. i1,i2,...,im i1,i2,...,im 1 2 m Let x = (x ,x ,...,x ) ∈ Rn. If we write xm as a m order n dimension 1 2 n hypermatrix with i ,i ,...,i entry x ,x ,...,x , then Axm 1, where the 1 2 m i1 i2 im − multiplication is taken as tensor contraction over all indices, is a n tuple whose i-th component is n a x x ...x . ii2i3...im i2 i3 im i2,i3X,...,im=1 Definition 2.1. Let A be a nonzero hypermatrix. A pair (λ,x) ∈ C×(Cn\ {0}) is called eigenvalue and eigenvector (or simply an eigenpair) if they satisfy the following equation Axm 1 = λx[m 1]. − − Here, x[m] is a vector with i-th entry xm. We call (λ,x) is a H-eigenpair (i.e., i λ and x are called H-eigenvalue and H-eigenvector, respectively) if they are both real. Definition 2.2. Let A be a nonzero hypermatrix. A pair (λ,x) ∈ C×(Cn\ {0}) is called an E-eigenpair (where λ and x are called E-eigenvalue and E-eigenvector, respectively) if they satisfy the following equations Axm 1 = λx, − n x2 = 1. i Xi=1 We call (λ,x) a Z-eigenpair if both of them are real. From the above definitions it is clear that, a constant multiplication of an eigenvector is also an eigenvector corresponding to a H-eigenvalue, but, this is not always true for E-eigenvalue and Z-eigenvalue. Now, we recall some results that are used in the next section. 3 Theorem 2.1 ([15]). The eigenvalues of A lie in the union of n disks in C. These n disks have the diagonal elements of the supersymmetric tensor as their centres, and the sums of the absolute values of the off-diagonal elements as their radii. The above theorem helps us to bound the eigenvalues of a tensor. Lemma 2.1. Let A be an m order and n dimensional tensor and D = diag(d ,...,d ) be a positive diagonal matrix. Define a new tensor 1 n m 1 − B = A.D (m 1).D...D − − z }| { with the entries B = A . i1i2...im i1i2...imdi−1(m−1)di2...dim Then A and B have the same H-eigenvalues. Proof. From the remarks of lemma 3.2 in [21]. Some results of spectral graph theory1 are also hold for general hyper- graphs. If λ is any eigenvalue of an adjacency matrix of a graph G with the maximal degree ∆ then λ 6 ∆. For a k-regular graph k is the maximum eigenvalue with a constant eigenvector of the adjacency matrix of that graph. If λ and µ are the eigenvalues of the adjacency matrices represent the graphs G and H, respectively, then λ+µ is also an eigenvalue of the same for G(cid:3)H, the cartesian product of G and H. All the eigenvalues of a Laplacian matrix ofagrapharenonnegative andavery roughupperboundofthese eigenvalues is 2∆, whereas, the eigenvalue of a normalised Laplacian matrix of a graph lies in the interval [0,2]. Zero is always an eigenvalue for both, Laplacian and normalised Laplacian matrices, of a graph, with a constant eigenvector. If A and L are the normalised adjacency matrix and normalised Laplacian matrix, respectively, of a graph (such that L = 1 − A) then the spectrum of A, σ(A) = 1−σ(L). If M be any connectivity matrix of a graph with r connected components then σ(M) = σ(M )∪σ(M )···∪σ(M ), where M 1 2 r i is the same connectivity matrix corresponding to the component i. 1For different spectral properties of a graph see [2, 6] 4 3 Spectral properties of general hypergraphs Definition 3.1. A (general) hypergraph G is a pair G = (V,E) where V is a set of elements called vertices, and E is a set of non-empty subsets of V called edges. Therefore, E is a subset of P(V) \ {∅}, where P(V) is the power set of V. Example 3.1. Let G = (V,E), whereV = {1,2,3,4,5}andE = {1},{2,3},{1,4,5} . Here, G is a hypergraph of 5 vertices and 3 edges. (cid:8) (cid:9) 3.1 Adjacency hypermatrix and eigenvalues Definition 3.2. Let G = (V,E) be a hypergraph where V = {v ,v ,...,v } 1 2 n and E = {e ,e ,...,e }. Let m = max{|e | : e ∈ E} be the maximum 1 2 k i i cardinality of edges, m.c.e(G), of G. Define the adjacency hypermatrix of G as A = (a ), 1 ≤ i ,i ,...,i ≤ n. G i1i2...im 1 2 m For all edges e = {v ,v ,...,v } ∈ E of cardinality s ≤ m, l1 l2 ls s m! a = , where α = , p1p2...pm α k !k !...k ! X 1 2 s k1,k2,...,ks 1,Pki=m ≥ and p ,p ,...,p are chosen in all possible way from {l ,l ,...,l } with at 1 2 m 1 2 s least once for each element of the set. The other positions of the hypermatrix are zero. Example 3.2. Let G = (V,E) be a hypergraph in example 3.1. Here, the maximum cardinality of edges is 3. The adjacency hypermatrix of G is A = G (a ), where 1 ≤ i ,i ,i ≤ 5. Here, a = 1,a = a = a = a = i1i2i3 1 2 3 111 233 232 223 323 a = a = 1,a = a = a = a = a = a = 1, and the other 332 322 3 145 154 451 415 541 514 2 elements of A are zero. G Definition 3.3. Let G = (V,E) be a hypergraph. The degree, d(v), of a vertex v ∈ V is the number of edges consist of v. Let G = (V,E) be a hypergraph, where V = {v ,v ,...,v } and E = 1 2 n {e ,e ,...,e }. Then, the degree of a vertex v is given by 1 2 k i n d(v ) = a . i ii2i3...im i2,i3X,...,im=1 5 Definition 3.4. A hypergraph is called k-regular if every vertex has the same degree k. Now, we discuss some spectral properties of A of a hypergraph G. Some G of these properties are very similar as in general graph (i.e. for a 2-uniform hypergraph). Theorem 3.1. Let µ be a H-eigenvalue of A . Then |µ| ≤ ∆, where ∆ is G the maximum degree of G. Proof. Let G be a hypergraph of order m and dimension n. Let µ be a H- eigenvalue of A = (a ) with an eigenvector x = (x ,x ,...,x ). Let G i1i2...im 1 2 n x = max{|x |,|x |,...,|x |}. Without loss of any generality we can assume p 1 2 n that x = 1. Now, p n |µ| = |µxm 1| = a x x ...x p − (cid:12) pi2i3...im i2 i3 im(cid:12) (cid:12)(cid:12)i2,i3X,...,im=1 (cid:12)(cid:12) (cid:12) (cid:12) n ≤ |a ||x |m 1 = d(v ) ≤ ∆. pi2i3...im p − p i2,i3X,...,im=1 Thus, for a k-regular hypergraph the theorem (3.1) implies |µ| ≤ k. Theorem 3.2. Let G = (V,E) be a k-regular hypergraph with n vertices. Then, A = (a ) has a H-eigenvalue k. G i1i2...im Proof. Since, G is k-regular, then d(v ) = k for all v ∈ V, i ∈ {1,2,3,...,n}. i i Now, for a vector x = (1,1,1,...,1) ∈ Rn we have n A xm 1 = a = k. G − ii2i3...im i2,i3X,...im=1 Thus the proof. Theorem 3.3. Let G = (V,E) be a k-regular hypergraph with n vertices. Then, A = (a ) has a Z-eigenvalue k( 1 )m 2. G i1i2...im √n − Proof. Thevector x = ( 1 , 1 ,..., 1 ) ∈ Rn satisfies theZ-eigenvalueequa- √n √n √n tions for λ = k( 1 )m 2. √n − 6 Theorem 3.4. Let G be a hypergraph with n vertices andmaximum degree∆. Let x = (x ,x ,...,x ) be a Z-eigenvector of A = (a ) corresponding 1 2 n G i1i2...im to an eigenvalue µ. If x = max |x |,|x |,...,|x | , then |µ| ≤ ∆. p 1 2 n xp (cid:8) (cid:9) Proof. The Z-eigenvalue equations of A for µ and x are Axm 1 = µx, and G − x2 = 1. Therefore, |x | ≤ 1, for all i = 1,2,3,...,n. Now, i i P n |µ||x | = a x x ...x , j (cid:12) ji2i3......im i2 i3 im(cid:12) (cid:12)(cid:12)i2,i3X,...,im=1 (cid:12)(cid:12) (cid:12) (cid:12) which implies |µ||x | ≤ d(j) ≤ ∆,∀j = 1,2,3,...,n. Therefore, |µ| ≤ ∆. j xp Definition 3.5. A hypergraph H = (V ,E ) is said to be a spanning subhy- 1 1 pergraph of a hypergraph G = (V,E), if V = V and E ⊆ E. 1 1 ′ ′ Theorem 3.5. Let G = (V,E) be hypergraph. Let H = (V ,E ) be a subhy- pergraph of G, such that, m.c.e(G) = m.c.e(H) be even. Then, µ (H) ≤ max µ (G), where µ is the highest Z-eigenvalue of the corresponding adja- max max cency hypermatrix. ′ ′ Proof. Let |V| = n, |V | = n (≤ n) and m.c.e(G) = m.c.e(H) = m. Now, µ (H) = max xtA xm 1 (by using lemma (3.1) in [12]) max x m=1 H − || || ′ n = max aH x x ...x ||x||m=1(cid:18) i1i2...im i1 i2 im(cid:19) i1,i2X,...,im=1 n = max aH x x ...x , where aH = x = 0 when i > n′ ||x||m=1(cid:18) i1i2...im i1 i2 im(cid:19) i1..im ir r i1,i2X,...,im=1 n ≤ aG x x ...x (cid:18) i1i2.....im i1 i2 im(cid:19) i1,i2X,...,im=1 ≤ µ (G), max since the each components of x are nonnegative (by Perron-Frobenious theo- rem [3]) and the number of edges of G is greater than or equal to the number edges of H. Hence the proof. Definition 3.6. Let G = (V,E) be a hypergraph with V = {v ,v ,...,v }, 1 2 n E = {e ,e ,...,e }, and m.c.e(G) = m. Let x = (x ,x ,...,x ) be a vector 1 2 k 1 2 n 7 in Rn and p ≥ s−1 be an integer. For an edge e = {v ,v ,...,v } and a l1 l2 ls vertex v , we define li xe/vli := x x ...x , p r1 r2 rp X where the sum is over r ,r ,...,r are chosen in all possible way from 1 2 p {l ,l ,...,l }, such that, all l (j 6= i) occur at least once. Where as, 1 2 s j xe := x x ...x , p r1 r2 rp X where the sum is over r ,r ,...,r are chosen in all possible way from 1 2 p {l ,l ,...,l } with at least once for each element of the set. 1 2 s The symmetric (adjacency) hypermatrix A of order m and dimension n G uniquely defines a homogeneous polynomial of degree m and in n variables by n F (x) = a x x ...x . G i1i2...im i1 i2 im A i1,i2X,,...im=1 We rewrite the above polynomial as: F (x) = aGxe , G e m A X e E ∈ where aG = s, α = m! , and s is the cardinality of the edgeee . α Pk1,k2,...,ks≥1,Pki=m k1!k2!...ks! Definition 3.7. Let G and H be two hypergraphs. The cartesian product, G×H, of G and H is defined by the vertex set V(G×H) = V(G)×V(H) and the edge set E(G×H) = {v}×e : v ∈ V(G),e ∈ E(H) e×{v} : e ∈ E(G),v ∈ V(H) . (cid:8) (cid:9)S(cid:8) (cid:9) Definition3.8. LetGbeahypergraphwiththevertexsetV = {v ,v ,...,v } 1 2 n and m.c.e(G) = m. For an edge e = {v ,v ,...,v } and integer r ≥ m the l1 l2 ls arrangement, (v v ...v ) (where p ,p ,...,p are chosen in all possible p1 p2 pr 1 2 r way from {l ,l ,...,l } with at least once for each element of the set) repre- 1 2 s sents the edge e in order r. Example 3.3. Let G = (V,E) whereV = {1,2,3,4,5}andE = {1,2,3},{2,3,5},{1,3,4,5} , then the arrangement(12233) represents the edge {1,2,3} in orde(cid:8)r 5. (12123) (cid:9) is also a representation of the edge {1,2,3} in order five, whereas, (111123) represents the edge {1,2,3} in 6 order. 8 Let G = (V,E) be a hypergraph with m.c.e(G) = m and E = {e ∈ E : i v ∈ e}. Now, the H-eigenvalue equation for A becomes i G aGxe/vi = λx(m−1), for all i. e m 1 i X − e Ei ∈ Theorem 3.6. Let G and H be two hypergraphs with m.c.e(G) = m.c.e(H). If λ and µ are H-eigenvalue for G and H, respectively, then λ + µ is a H-eigenvalue for G×H. Proof. Let n and n be the number of vertices in G and H, respectively, 1 2 and m.c.e(G) = m.c.e(H) = m. Let (λ,u) and (µ,v) be H-eigenpairs of A G and A , respectively. Let w ∈ Cn1n2 be a vector with the entries indexed H by the pairs (a,b) ∈ [n ]×[n ], such that, w(a,b) = u(a)v(b). Now, we show 1 2 that (λ+µ,w) is an H-eigenpair of A . G H × aGe×Hwme/(a1,b) = aGe×Hwm{a}×1e/(a,b) + aGe×Hwme×{1b}/(a,b) X − X − X − e∈E(a,b) {a}×e∈E(a,b) e×{b}∈E(a,b) with e∈Eb with e∈Ea = aG Hum 1(a)ve/b + aG Hue/a vm 1(b) e× − m 1 e× m 1 − X − X − e∈Hb e∈Ga = um 1(a) aHve/b +vm 1(b) aGue/a − e m 1 − e m 1 X − X − e∈Hb e∈Ea = um 1(a)µvm 1(b)+vm 1(b)λum 1(a) − − − − = (λ+µ)wm 1(a,b). − Hence the proof2. Lemma 3.1. Let A and B be two symmetric hypermatrix of order m and dimension n, where m is even. Then λ (A + B) ≤ λ (A) + λ (B), max max max where λ (A) denotes the largest Z-eigenvalue of A. max Proof. λ (A+B) = max xt(A+B)xm 1 (by using lemma (3.1) in [12]) max x m=1 − || || ≤ max xtAxm 1 +max xtBxm 1 x m=1 − x m=1 − || || || || = λ (A)+λ (B). max max 2For the similar proof on uniform hypergraph see [7]. 9 Let G = (V,E) be a hypergraph with the vertex set V = {v ,v ,...,v } 1 2 n and m = max{|e | : e ∈ E}. We partition the edge set E as, E = E ∪E ∪ i i 1 2 ···∪E , where E contains all the edges of the cardinality i and construct m i a hypergraph G = (V,E ), for a nonempty E . i i i Definition 3.9. Define the adjacency hypermatrix of G in m (> i) -order i by a n dimensional m order hypermatrix Am = (am) , 1 ≤ p ,p ,...,p ≤ n, Gi Gi p1p2...pm 1 2 m (cid:0) (cid:1) such that, for any e = {v ,v ,...,v } ∈ E , l1 l2 li i i m! (am) = , where α = Gi p1p2...pm α k !k !...k ! X 1 2 i k1,k2,...,ki 1,Pkj=m ≥ and p ,p ,...,p are chosen in all possible way from {l ,l ,...,l } with at 1 2 m 1 2 i least once for each element of the set. The other positions of Am are zero. Gi Thus, we can represent a hypergraph G, with m.c.e(G) = s, in higher order m > s by the hypermatrix Am. Clearly, all the eigenvalue equations G show that the eigenvalues of Am1 and Am2 are not equal for m 6= m . G G 1 2 Theorem 3.7. Let G = (V,E) be a hypergraph and m.c.e(G) = m be even. Then λ (A ) ≤ m λ (Am), where λ (A) is the largest Z- max G i=1 max Gi max eigenvalue of A. P Proof. Since A = m Am, the proof follows from the lemma (3.1). G i=1 Gi P Moreover, thetheorem(3.7)impliesλ (A ) ≤ m n λ (Am),where max G i=1 i max i ni is the number of edges of cardinality i and Ami isPthe adjacency hyperma- trix in m-order of a hypergraph contains a single edge of cardinality i. 3.2 Laplacian hypermatrix and eigenvalues Definition 3.10. Let G = (V,E) be a (general) hypergraph without any isolated vertex where V = {v ,v ,...,v } and E = {e ,e ,...,e }. Let 1 2 n 1 2 k m.c.e(G) = m. DefinetheLaplacianhypermatrix, L , ofG = (V,E)asL = G G D −A = (l ) where 1 ≤ i ,i ,...,i ≤ n, where D = (d ) is G G i1i2...im 1 2 m G i1i2...im the m order n dimensional diagonal hypermatrix with d = d(v ) and ii...i i others are zero. The signless laplacian of G is defined as L = D +A . G G G 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.