Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Introduction
Throughout this paper, we will use graph to mean an undirected graph without parallel edges, but possibly with loops. Let G, H be two graphs; we write Hom(G, H) for the set of homomorphisms from G to H. When G, H are unweighted, we dene hom(G, H) = |Hom(G, H)|. A denition of hom(G, H) for weighted graphs will be given later in this paper. We will write G H for the tensor product of the graphs G and H. Let i(G) denote the number of independent sets in G. It is easily seen that i(G) = hom(G, I2 ) for the two-vertex graph I2 which has the two vertices adjacent and one vertex having a loop. Alon examined the extremal behavior of i(G) in d-regular graphs, and conjectured an upper bound of i(Kd,d )N/2d when G has N vertices. Some time later, Kahn, using entropy methods, proved this conjecture under the additional assumption that G is bipartite. Theorem 1. (Kahn). Let G be a bipartite, d-regular graph with |V (G)| = N . Then i(G) i(Kd,d )N/2d . Galvin and Tetali extended Kahns methods to prove similar bounds on hom(G, H) when G is d-regular and bipartite. Theorem 2. (GT). Let G be a bipartite, d-regular graph with |V (G)| = N . Then hom(G, Y ) hom(Kd,d , Y )N/2d for any graph Y . 1
Recently, Zhao used a clever swapping trick to extend Kahns result to all d-regular graphs, thus proving Alons conjecture. Theorem 3. (Zhao). Let G be a graph, not necessarily regular or bipartite. Then i(G)2 i(G K2 ) . Corollary 1. Let G be a d-regular graph with |V (G)| = N . Then i(G) i(Kd,d )N/2d . Proof. For G bipartite, this follows from the theorem of Kahn. Suppose G is not bipartite; then G K2 is, and is d-regular on 2N vertices. We know that: i(G)2 i(G K2 ) i(Kd,d )N/d , from which the claim follows immediately. In this paper, we show that Theorem 3 is equivalent to an inequality of the form hom(X, G) hom(X, H) for certain 4-vertex graphs G, H, and use this fact to give a novel proof of the result. In addition, we explore new approaches to other inequalities of this form, and formulate a conjecture which would imply an analogue of Alons conjecture for 3-colorings.
2
2.1
Preliminaries
Graph exponentiation
In order to have any hope of interpreting Theorem 3 in terms of functions of the form hom(X, G), we need a way to relate expressions of the form hom(GH, K) to simpler expressions. Fortunately, there is a tool that allows us to do this, at least in the case where H = K2 . Denition 1. (following Lovasz). Let G be a graph. The exponentiation of G to the power K2 , written GK2 , is a graph dened as follows: V (GK2 ) is the set of ordered pairs of vertices of G (not necessarily distinct); the vertices (u0 , v0 ) and (u1 , v1 ) are adjacent i (u0 , v1 ) and (u1 , v0 ) are edges of G. Note. It is possible in general to exponentiate to the power of any graph, or indeed even a digraph. However, the denition becomes more subtle, and in addition a graph to the power of another graph is not even a graph in general, but a digraph. (In fact, Lovaszs denition would strictly speaking even have GK2 be a digraph!) However, in this paper we only consider products of the form G K2 , and the above denition, though modied, retains the desired properties of Lovaszs construction. 2
Lemma 1. (adapted from Lovasz). hom(G K2 , H) = hom(G, H K2 ). Proof. G K2 has vertex set V (G) {0, 1}, with (x0 , y0 ) and (x1 , y1 ) adjacent if x0 , x1 are adjacent in G and y0 = y1 . Let : G K2 H be a homomorphism. We dene a homomorphism f () : G H K2 by setting f ()(v) = (((v, 0)), ((v, 1)). It is easy to check that f () is indeed a homomorphism. Conversely, given a homomorphism : G H K2 we may dene a homomorphism f 1 () : G K2 H as follows: If (v) = (u, w), then f 1 ()((v, 0)) = u, f 1 ()((v, 1)) = w. Once more it is straightforward to check that f 1 () is a homomorphism and that f and f 1 are inverses. So we have a bijection between Hom(G K2 , H) and Hom(G, H K2 ).
2.2
Swappability
Call an (ordered) pair of graphs (G, H) swappable if hom(X, G) hom(X, H) for all (nite, simple) graphs X. Then, recalling that hom(G, H K) = hom(G, H) hom(G, K), we immediately obtain the following
K Corollary 2. Theorem 3 is equivalent to the statement that (I2 I2 , I2 2 ) is swappable.
In the next two lemmas, we give sucient conditions for a pair (G, H) to be swappable. Lemma 2. Suppose that there exists a graph homomorphism : G H which is injective on vertices. Then (G, H) is swappable. Proof. Consider a homomorphism : X G. Then the composition is a graph homomorphism from X to H. Furthermore, suppose that 0 , 1 : X G are such that 0 = 1 . Consider these as maps from V (X) to V (H); since a graph homomorphism with xed source and target is determined by its action on the vertices, this doesnt lose any information. As is injective on vertices, there is a map : V (H) V (G) such that is the identity map on V (G). Now 0 = 0 = 1 = 1 . Therefore, composition with gives an injective map between the Hom-sets Hom(X, G) Hom(X, H), from which the claim follows. Lemma 3. Suppose that (G0 , H0 ) and (G1 , H1 ) are swappable. Then: 1. (G0 G1 , H0 H1 ), where is the tensor product, is swappable; 2. (G0 G1 , H0 H1 ), where is the disjoint union, is swappable; 3. (G0 + G1 , H0 + H1 ), where + is the join of two graphs, is swappable. Proof. 1. Follows immediately from the fact that hom(X, GH) = hom(X, G)hom(X, H).
2. Follows immediately from the fact that hom(X, G H) = hom(X, G) + hom(X, H).
3. The full proof is omitted for space; we provide a brief outline. Note that by assumption there exist families of injections fX : Hom(X, G0 ) Hom(X, H0 ) and gX : Hom(X, G1 ) Hom(X, H1 ). Write V (G0 + G1 ) = V (G0 ) V (G1 ). Let : X G0 + G1 be a homomorphism. Set V (X0 ) = ( 1)(V (G0 )), V (X1 ) = 1 (V (G1 )), and let X0 , X1 be the induced subgraphs of X on these vertex sets. We can consider as essentially made up of two distinct homomorphisms 0 : X0 G0 and 1 : X1 G1 . By applying fX0 , gX1 , respectively, to these homomorphisms, we get a new mapping from X to H0 + H1 . Checking that this construction is injective on the Hom-sets and gives a bona de homomorphism is tedious but straightforward.
In this section, we propose a general approach to inequalities of the form hom(X, G0 ) hom(X, G1 ). The high-level view is as follows: We dene a (weighted) graph-valued function Gt which deforms G0 into G1 . Then for xed X, hom(X, Gt ) is a function of t; we wish to nd some nice property of hom(X, Gt ), independent of X, from which the inequality follows. For the remainder of the paper, we will concentrate on a specic such weighted graph-valued function, namely a convex sum of G0 , G1 . First we need to dene some basic notions for weighted graphs. Fix a commutative ring R. An R-weighted graph (also weighted graph for R generic; we may also say weighted graph when the choice of R is clear) is dened as a vertex set V along with a symmetric weight function W : V V R. (Here, R will be either the real numbers R, or the real polynomial ring R[t]; it is conceivable that other rings may be of interest.) Intuitively, one thinks of the edges of G as having (nonzero) weights in R. One may view an ordinary unweighted graph as a weighted graph by setting W (u, v) = 1 if u, v V (G) are connected, and W (u, v) = 0 otherwise. We can extend the denition of hom(X, Y ) to cases where the target graph Y is a weighted graph, by dening hom(X, Y ) =
:V (X)V (Y ) uvE(X)
W ((u), (v))
. When Y is unweighted, the above denition agrees with the rst one (after applying the unique morphism from Z to the base ring R.) Note. It is possible in some cases (e.g., when all weights are real and nonnegative) to dene hom(X, Y ) where X, Y are both weighted graphs. However, this is not necessary for this paper, and it does not extend to graphs with weights in a general ring; I have elected to omit it for simplicity of presentation.
For a xed vertex set V , the set of all R-weighted graphs on V forms a free R-module (of dimension |V |(|V |+1)/2) in the obvious way. We will take implicit advantage of this module structure by speaking of things such as G0 + G1 , or rG (for r R), but deeper properties of modules play no signicant role in this paper. We will consider weighted graphs over the polynomial ring R[t]. Let G0 , G1 be graphs on the same vertex set V ; dene (G0 , G1 ) = (1/2+t)G0 +(1/2t)G1 . In addition, dene PG0 ,G1 (X; t) = hom(X, (G0 , G1 )) . When G0 , G1 are understood, we will often abbreviate this to P (X; t). Note that PG0 ,G1 (X; t) is a polynomial for any xed X, and furthermore that P (X; 1/2) = hom(X, G1 ) and P (X; 1/2) = hom(X, G0 ). The special case where G0 = H H and G1 = H K2 is particularly important, so we give it its own notation: for a xed (unweighted) graph G, dene QG (X; t) = hom(X, (G G, GK2 )) .
Proof of Theorem 3
Let V = {a, b}; call the graph on V with only a single edge between a and b K2 , and call the graph on V with only loops at a and b L2 . Lemma 4. (K2 , L2 ) is swappable. Note. We give two proofs; the rst is a short Book proof, and is morally similar to Zhaos approach. The second, though longer and less elegant, is meant to provide a concrete demonstration of the weighted-graph machinery. Proof. 1. We wish to show that hom(X, K2 ) hom(X, L2 ). Assume that X is connected; this can be done without loss of generality since hom(G H, K) = hom(G, K)hom(H, K). There are two cases: (i) If X is not bipartite, then hom(X, K2 ) = 0, and the inequality in this case is trivial. (ii) If X is bipartite, then a homomorphism : X K2 is determined by its value at any given vertex; thus hom(X, K2 ) = 2. But there are two homomorphisms from X to L2 as well, namely the map that sends every vertex to a and the map that sends every vertex to b. So here hom(X, K2 ) = hom(X, L2 ) = 2. Proof. 2. We will show that, for any simple graph X with m edges and odd integer n, the coecient of tn in PK2 ,L2 (X; t) is nonpositive. From this the desired result will follow immediately; to see this, write P (X; t) = E(X; t) + O(X; t) where E(X; t) is even and O(X; t) is odd. Then P (X; 1/2) P (X; 1/2) i
O(X; 1/2) O(X; 1/2). But when all odd-power coecients are negative, O(X; t) has the opposite sign from t, so the desired inequality automatically follows. For convenience we will denote by KL2 the weighted graph (K2 , L2 ). For Hom(X, KL2 ), we will denote by L() (respectively K()) the set of edges of X that map to edges from L2 (respectively K2 ). Write () = |L()|, k() = |K()|. Then: hom(X, LK2 ) =
:XKL2
()
. Let an denote the coecient of tn in this polynomial. By the binomial theorem, we can write an =
:XKL2 r+s=n
(1)r (1/2)mn
() r
k() s (1)r
() k() r s r+s=n can be interpreted as follows: Each subset of edges with r edges mapping to L2 and s edges mapping to K2 contributes a term of (1)r to the sum. This interpretation allows us to rewrite the sum as: Now, pulling out the factor of (1/2)mn , the sum an = (1/2)mn
:XKL2 SE(X),|S|=n
(1)|L()S| (1)|L()S|
= (1/2)mn
SE(X),|S|=n :XKL2
. From here we proceed by contradiction; let G be a simple graph with m edges, and let n be an odd integer such that: (i) The coecient of tn in PK2 ,L2 (G; t) is positive. (ii) For any graph Y on fewer than m edges, and any odd integer n , the coecient of tn in PK2 ,L2 (Y ; n ) is nonpositive. Fixing S E(G), we see that (1)|L()S|
:GKL2
where G[S] is the induced subgraph of G on S. But this latter sum is the coecient of tn in P (X[S]; t), and by the minimality assumption this is negative unless S = E(G) and n = m. 6
Suppose that G has some vertex v with odd degree. Then there is an involution of Hom(G, KL2 ) given by changing the image of v. It is easily checked that (), (()) have opposite parity. It follows that if G has a vertex of odd degree, am = 0. Suppose that G has all vertices of even degree, and an odd number of edges. Then (1) () = 2|V (G)|
:GKL2
. Indeed, for any : G KL2 , k() is even, and therefore () is odd. To see this, write V (KL2 ) = {0, 1}, and note that when maps every vertex to 0, k() = 0. Now suppose we change a homomorphism to by changing the image of just one vertex w from 0 to 1. Further say that w is adjacent to r vertices which map to 1 under . Then k( ) = k() r + ((w) r) = k() 2r + (w) . Since w has even degree by assumption, we conclude that k( ) has the same parity as k(); it follows as a corollary that every homomorphism has k() even. Thus we conclude that the coecient of tn in PK2 ,L2 (G; t) is negative; but this is a contradiction, and we are done. Proof. (Theorem 3). By Lemma 4, (K2 , L2 ) is swappable. Then, by two applications of Lemma 3, ((K2 K1 ) + L1 , (L2 K1 ) + L1 ) is swappable, where K1 is the simple graph on one vertex and L1 is a loop on one vertex. Now I2 I2 K is isomorphic to (K2 K1 ) + L1 , and I2 2 isomorphic to (L2 K1 ) + L1 ; so, by Lemma 1 and some basic identities, we obtain the statement of Theorem 3.
Future directions
This research was motivated by an attempt to extend Zhaos result on independent sets to 3-colorings, i.e., to prove that hom(X, K3 ) hom(Kd,d , K3 )N/2d when X is a d-regular graph on N vertices. One way to do this would be to show (as in the second proof of Lemma 4 above) that QK3 (X; t) has nonpositive coecients of tn when n is odd. However, there are some diculties in extending the proof of the analogous result for I2 to this case. Most crucially, the analogous sum for an works out to be: (1/2)bn
nbm ,|E()|=b SE(),|S|=n
(1)|L()S|
, where E() is the set of edges of X mapped to edges of weight 1/2 t. In particular, there are choices of X, E() such that, with E() xed, the sum is positive. (For instance, take X to be a triangular prism and E() to be 7
the edges of one of the triangles.) Thus, its not possible to directly extend the passing-to-subgraphs argument of Lemma 4 to the 3-coloring case. However, its possible that a more complex argument would work and computer searches have shown that QK3 (X; t) has nonpositive odd coecients for all X with at most 5 vertices.