A rooted variant of Stanley’s chromatic symmetric function

The left side X_0(T_*) is the weighted sum of all colorings in \operatorname{\textsc{col}}_0(T_*). We build each such coloring uniquely by making the following choices. First, color the root vertex r with color 0; the weight monomial for this step is x_0. Next, color the subtree T_{1*} using any coloring in \operatorname{\textsc{col}}_{\neq 0}(T_{1*}); color the subtree T_{2*} using any coloring in \operatorname{\textsc{col}}_{\neq 0}(T_{2*}); and so on. The colorings of the subtrees can be chosen independently since the subtrees do not connect with each other except through the root r. The recursion (6) follows by the Product Rule for Weighted Sets [loehr-comb, § 5.8].

By reducibility of f, there is a proper factorization f=gh where g,h\in\mathbb{Q}[x_0,\ldots,x_N] are not initially known to be homogeneous. Write g=g^{(d)}+\text{(lower terms)}, where d>0 is the maximum degree of any monomial in g, and g^{(d)} is the sum of all monomials in g having degree d. Write h=h^{(e)}+\text{(lower terms)}, where e>0 is the maximum degree of any monomial in h, and h^{(e)} is the sum of all monomials in h having degree e. Then gh=g^{(d)}h^{(e)}+\text{(lower terms)}. Since \mathbb{Q}[x_0,\ldots,x_N] is an integral domain, the product g^{(d)}h^{(e)} (which is homogeneous of degree d+e) cannot be zero. Since f is homogeneous, this product must equal f. So f=g^{(d)}h^{(e)}, where both factors are homogeneous and non-constant.

Let f'=f|_{x_i=0}. To get a contradiction, assume f' is irreducible (hence non-constant and nonzero) and f is reducible. By the previous lemma, there is a proper factorization f=gh where g and h are both homogeneous of positive degree. Letting g'=g|_{x_i=0} and h'=h|_{x_i=0}, we have f'=g'h'. Since f'\neq 0, g' and h' are nonzero. It follows at once that g' and h' are homogeneous of positive degree. Thus f'=g'h' is a proper factorization of f', which is a contradiction.

All polynomials mentioned in the lemma are homogeneous of degree |V(G_*)| and are nonzero, since the hypothesis N\geq M\geq k ensures there are enough colors to produce at least one coloring satisfying all needed conditions. Part (1) follows from Lemma 10, since setting x_0=0 in X_{\neq 0}(G_*;x_0,\ldots,x_M) produces X_G(x_1,\ldots,x_M). Parts (2) and (3) follow from Lemma 10 by setting x_N=0,\ldots,x_{M+1}=0 one at a time.

Let n=|V(G)|; we may assume n>1. Note that f=\sum_{j=0}^n a_jq^j, where a_j\in\mathbb{Z} is the number of proper colorings \kappa:V(G)\rightarrow\{1,2,\ldots,M\} such that the number of v with \kappa(v)\in\{1,2,\ldots,k\} is j. This holds since specializing X_G to f replaces the weight monomial x_1^{e_1}x_2^{e_2}\cdots x_M^{e_M} of each proper coloring \kappa by q^{e_1+e_2+\cdots+e_k}, where e_1+\cdots+e_k is the number of vertices that \kappa colors using 1,2,\ldots,k.

The following counting argument leads to a formula for a_j, where 0\leq j\leq n is fixed. To build a proper coloring counted by a_j, first choose a j-element subset S of V(G), and let S^c=\{v\in V(G): v\not\in S\}. Let G|S be the graph with vertex set S and edge set consisting of all edges in G joining two vertices in S, and define G|S^c similarly. Choose a proper coloring \kappa_1 of G|S using color set \{1,2,\ldots,k\}, and choose a proper coloring \kappa_2 of G|S^c using color set \{k+1,k+2,\ldots,k+p\}. The union of \kappa_1 and \kappa_2 is a proper coloring of G counted by a_j, and all such colorings arise uniquely by this construction process. In summary,

In particular, a_n=\chi_G(k) and a_0=\chi_G(p).

We now confirm that f satisfies the hypotheses of Eisenstein’s Criterion (Theorem 8). By choice of k and p, a_n is nonzero and not divisible by p. We know \chi_G is c_1x plus higher terms all divisible by x^2, and p does not divide c_1. Replacing x by p, we see that \chi_G(p) is c_1p plus a multiple of p^2, so p^2 does not divide a_0=\chi_G(p). For 0\leq j<n, each subgraph S^c used in the sum for a_j is nonempty with n-j>0 vertices. So each \chi_{G|S^c} is divisible by x, each \chi_{G|S^c}(p) is divisible by p, and a_j itself is therefore divisible by p. By Eisenstein’s Criterion, f is irreducible in \mathbb{Q}[q].

We prove the contrapositive. Given F and f as in the lemma setup, assume there is a proper factorization F=GH, where G,H\in\mathbb{Q}[x_1,\ldots,x_M] are not constant. Choose i,j so that x_i appears in G and x_j appears in H. Since F is symmetric, we can assume i=1 and j\in\{1,2\} by applying a permutation of the variables to the factorization F=GH.

Let R=\mathbb{Q}[x_{k+1},\ldots,x_M] and view F,G,H as elements of the ring R[x_1,\ldots,x_k]. This ring is graded by total degree in the variables x_1,\ldots,x_k. So we can write

where F^{(d)} is the sum of all the highest degree monomials in F (say of degree d), and similarly for G^{(d_1)} and H^{(d_2)}. Since F=GH and there are no zero divisors, we must have F^{(d)}=G^{(d_1)}H^{(d_2)} and d=d_1+d_2. Moreover, d_1>0 since x_1 appears in G, and d_2>0 since x_1 or x_2 appears in H and k\geq 2.

We now apply the specialization that replaces each variable x_1,\ldots,x_k by the new formal variable q. For each P\in R[x_1,\ldots,x_k], let P'\in R[q] be the image of P under this specialization. We have F'=G'H' where F', G', H' are in R[q]. Write

where u,r,s\in R are nonzero. Since R[q] is graded by degree in q and has no zero divisors, u=rs and e=e_1+e_2. On one hand, e_1\leq d_1 and e_2\leq d_2 since the degree in q of each monomial after the specialization matches the original total degree in x_1,\ldots,x_k, but some terms might perhaps cancel. On the other hand, e=d since all coefficients of F are nonnegative. But now e=e_1+e_2\leq d_1+d_2=d=e, so we must actually have e_1=d_1>0 and e_2=d_2>0. In other words, G' and H' are non-constant polynomials in q.

For all p\in R=\mathbb{Q}[x_{k+1},\ldots,x_M], write \overline{p} for the specialization of p where x_{k+1}=\cdots=x_M=1. Now u has all nonnegative coefficients (since F does), so \overline{u}>0. Then \overline{r}\cdot\overline{s}=\overline{rs}=\overline{u}>0, and hence \overline{r}\neq 0\neq \overline{s}. So when we further specialize the factorization F'=G'H' by setting x_{k+1}=\cdots=x_M=1, we get a factorization f=gh in \mathbb{Q}[q] where g still has degree d_1>0 in q and h still has degree d_2>0 in q. Thus, f is reducible in \mathbb{Q}[q].

Write X_{\neq 0}^{2,p}(G_*)=\sum_{j\geq 0} a_jq^j. Adapting the argument leading to (?), we can show

where we sum over all lists (S_0,S_1,U) satisfying these conditions: V(G_*) is the disjoint union of S_0, S_1, and U; |S_0\cup S_1|=j; the root of G_* is not in S_0; no two vertices in S_0 are joined by an edge; and no two vertices in S_1 are joined by an edge. We get a proper coloring contributing to the coefficient of q^j in X_{\neq 0}^{2,p}(G_*) by coloring every vertex in S_0 with color 0, every vertex in S_1 with color 1, and choosing a proper coloring of G|U using the p available colors \{2,3,\ldots,p+1\}. Formula (Proof) follows from the Sum Rule, the Product Rule, and the definition of \chi_{G|U}.

Since G_* is connected and bipartite, there is exactly one proper coloring of G_* using color set \{0,1\} where the root r of G_* is not colored 0. In more detail, for each vertex s of G, there exists a path in G from r to s (since G is connected), and the lengths of all paths from r to s have the same parity (since G is bipartite). To obtain a coloring \kappa with the specified properties, we must set \kappa(s)=1 for all s that are an even number of edges away from r, and \kappa(s)=0 for all s that are an odd number of edges away from r. So the coefficient of q^n in X_{\neq 0}^{2,p}(G_*) is 1, and hence this polynomial is monic of degree n=|V(G)|. (This conclusion also holds for X_0^{2,p}(G_*), by interchanging the roles of colors 0 and 1.)

For 0\leq j<n, p divides a_j since p divides each summand \chi_{G|U}(p) on the right side of (Proof). (We need j<n to ensure that G|U is a nonempty graph.) Since p does not divide the coefficient of x in \chi_G, a_0=\chi_G(p) is not divisible by p^2. Eisenstein’s Criterion now shows that X_{\neq 0}^{2,p}(G_*) is irreducible in \mathbb{Q}[q]. The conclusion for rooted trees follows since every rooted tree T_* is non-empty, connected, and bipartite, and the coefficient of x in \chi_T=x(x-1)^{n-1} is (-1)^{n-1}. When p=2, the conclusion for any connected bipartite rooted graph follows from the next lemma.

Let L(G) be the coefficient of x in \chi_G. We use induction on the number of edges in G. The result is immediate if G has no edges, or if G is not connected (L(G) is zero in the latter case). For the induction step, assume G is connected with at least one edge. It follows from the deletion-contraction recursion [west-graph, Thm. 5.3.6] that for any edge e of G, L(G)=L(G-e)-L(G_e), where G-e is obtained from G by deleting edge e, and G_e is obtained from G by contracting edge e. Now consider several cases.

Case 1. There is an edge e of G whose removal disconnects G, so L(G-e)=0. Now G_e is connected, and G_e is bipartite if and only if G is bipartite, since contracting e does not create or destroy any cycles. So L(G) and L(G_e) have the same parity, and the result follows by induction.

Case 2. G is connected and bipartite, but Case 1 does not apply to G. Let e be any edge of G. Then G-e is connected and bipartite, while G_e is connected and not bipartite (as e must be part of an even-length cycle of G, and contraction makes this an odd cycle). By induction, L(G-e) is odd and L(G_e) is even, so L(G) is odd.

Case 3. G is connected with an odd cycle, and G has an edge e not used in that odd cycle. We can assume Case 1 does not apply to G. Then G-e is connected with an odd cycle, and (as is readily checked) G_e also has an odd cycle. By induction, L(G-e) and L(G_e) are even, so L(G) is even.

Case 4. None of the first three cases apply to G. Then G itself must be an odd cycle. For any edge e on this cycle, G-e is a path and G_e is an even cycle (or a single edge), so G-e and G_e are connected and bipartite. By induction, L(G-e) and L(G_e) are odd, so L(G) is even.

For a permutation \sigma of \{0,1,2,3\} and a coloring \kappa:V(G_*)\rightarrow \{0,1,2,3\}, \sigma\circ\kappa is the coloring obtained from \kappa by replacing each color c by color \sigma(c). Note \kappa is a proper coloring of G if and only if \sigma\circ\kappa is a proper coloring of G. A coloring \kappa contributes to the coefficient of q^j in f_0(G_*) if and only if (0,1)\circ \kappa contributes to the coefficient of q^j in f_1(G_*). So f_0(G_*)=f_1(G_*), and f_2(G_*)=f_3(G_*) follows similarly by comparing \kappa to (2,3)\circ\kappa. On the other hand, a coloring \kappa contributes to the coefficient of q^j in f_0(G_*) iff \kappa'=(0,2)(1,3)\circ\kappa contributes to the coefficient of q^{n-j} in f_2(G_*). So f_2(G_*)=\operatorname{rev}_n(f_0(G_*)). This proves (a), and we prove (b) by the same symmetry arguments. For example, if \kappa colors exactly j vertices using \{0,1\} and does not color the root with color 0, then (0,2)(1,3)\circ\kappa colors exactly n-j vertices using \{0,1\} and does not color the root with color 2 (and conversely). Since f_G=f_0(G_*)+f_1(G_*)+f_2(G_*)+f_3(G_*), part (c) follows from part (a). Similarly, part (d) follows from part (b) and the observation \sum_{i=0}^3 f_{\neq i}(G_*)=\sum_{j=0}^3 3f_j(G_*)= 3f_G. The first equality in part (e) holds by definition, and the second equality follows from part (d).

It is routine to check that isomorphic rooted graphs have the same X_0 polynomial, so (a) implies (b). Since N\geq 3, (b) implies (c) by setting x_0=x_1=q, x_2=x_3=1, and x_k=0 for 3<k\leq N. To prove (c) implies (a), we use induction on n=\deg(f_0(T_*)). Let T_* and U_* be rooted trees with f_0(T_*)=f_0(U_*). Both T_* and U_* have n=\deg(f_0(T_*)) vertices. In the base case n=1, we have T_*\cong U_* since both trees consist of a single root vertex. Assume n>1 from now on. Let the root r of T_* have principal subtrees T_{1*},\ldots,T_{c*}, and let the root s of U_* have principal subtrees U_{1*},\ldots,U_{d*}. Applying Proposition 6 to T_* and to U_* and specializing at (q,q,1,1), the assumption f_0(T_*)=f_0(U_*) becomes

Theorem 14 shows that each side of (Proof) is a factorization of the polynomial f_0(T_*)=f_0(U_*) into monic irreducible factors. Because \mathbb{Q}[q] is a unique factorization domain, we conclude that c=d and (after reordering factors appropriately) f_{\neq 0}(T_{j*})=f_{\neq 0}(U_{j*}) for 1\leq j\leq c. For each j, we can apply Proposition 16(e), taking n there to be \deg(f_{\neq 0}(T_{j*}))=|V(T_{j*})|=|V(U_{j*})|, to deduce that f_0(T_{j*})=f_0(U_{j*}). Each principal subtree T_{j*} has fewer than n vertices. By the induction hypothesis, T_{j*}\cong U_{j*} for 1\leq j\leq c. Finally, T_*\cong U_* follows from (?).

Let S=\operatorname{\textsc{col}}_0(G_*), T=\operatorname{\textsc{col}}_0(G_*-e), and U=\operatorname{\textsc{col}}_0(G_{e*}). The set T is the disjoint union of the two subsets

There is a bijection from U' to U sending \kappa to \overline{\kappa}, where the two colorings agree on all vertices other than r and s, and \overline{\kappa} colors the new root vertex 0. It follows at once that \operatorname{wt}(\kappa)=x_0\operatorname{wt}(\overline{\kappa}). Since T is the disjoint union of S and U', the Weighted Sum Rule gives X_0(G_*-e)=X_0(G_*)+x_0X_0(G_{e*}), as needed. The initial condition is immediate from the Weighted Product Rule.

We build the proper colorings \kappa\in\operatorname{\textsc{col}}_0(G_*) contributing to the coefficient of x_0^k in X_0(G_*) as follows. First, choose a k-element independent set A\in I(G_*,k) and color the vertices of A (including the root r) with color 0. Second, choose any proper coloring of the graph G-A using color set \{1,2,\ldots,N\}. The formula in (a) follows from the Sum and Product Rules for Weighted Sets. Part (b) is proved in the same way, but now we choose A\in I'(G_*,k) to ensure the root does not receive color 0.

Assume the condition on multisets is true; we prove Conjecture 1. Let T and U be any n-vertex trees with X_T=X_U. By the multiset condition, [X_0(T_*^r): r \in V(T)]=[X_0(U_*^s): s \in V(U)]. So there exist r\in V(T) and s\in V(U) with X_0(T_*^r)=X_0(U_*^s). By Theorem 17, T_*^r and U_*^s are isomorphic as rooted graphs. So T and U are isomorphic as graphs.

We prove the converse implication by contradiction. Assume Conjecture 1 is true, but the condition on multisets is false. Then there exist trees T and U such that X_T=X_U, but [X_0(T_*^r):r\in V(T)]\neq [X_0(U_*^s):s\in V(U)]. Since X_T=X_U, Conjecture 1 says T\cong U, so there is a graph isomorphism f:V(T)\rightarrow V(U). For each r\in V(T), f is a rooted graph isomorphism from T_*^r to U_*^{f(r)}. But then X_0(T_*^r)=X_0(U_*^{f(r)}) for all r\in V(T), which means the two multisets are equal. This gives the required contradiction.

The proof is an application of the pointing construction to the generating function X_G (see [flajolet, Theorem I.4]). Each coloring \kappa\in\operatorname{\textsc{col}}(G) contributes a monomial x_0^{e_0}\cdots x_N^{e_N} to X_G. Applying the operator x_0\frac{\partial}{\partial x_0} multiplies this monomial by e_0, which is the number of vertices colored 0 by \kappa. The coloring \kappa belongs to exactly e_0 of the sets \operatorname{\textsc{col}}_0(G_*^r) as r varies through V(G), since the root r must be colored 0. Thus, each coloring \kappa makes the same contribution (namely e_0x_0^{e_0}\cdots x_N^{e_N}) to both sides of (25).

Let A^v be the set of all (not necessarily proper) colorings \kappa:V(G)\rightarrow\{0,1,\ldots,N\} where \kappa(v)=0. For each edge e, let A_e^v be the set of \kappa\in A^v that assign the same color to the two endpoints of edge e. For each S\subseteq E(G), we can count the weighted set \bigcap_{e\in S} A_e^v as follows. Write \lambda_v^+(G_s)=k, where k is the size of the component C^v of G_S containing the root v. Write \lambda_v^-(G_S)=(\lambda_1,\ldots,\lambda_m), where each \lambda_i is the size of another component C_i of G_S. We build a coloring in \bigcap_{e\in S} A_e^v as follows. Color all k vertices in C^v with color 0 (as we must), giving a weight contribution of x_0^k. For each remaining component C_i, choose the common color for all \lambda_i vertices in that component. The weight contribution for that choice is x_0^{\lambda_i}+x_1^{\lambda_i}+\cdots+x_N^{\lambda_i} =p_{\lambda_i}(x_0,x_1,\ldots,x_N). By the Product Rule, the generating function for \bigcap_{e\in S} A_e^v is p_{\lambda_v^-(G_S)}(x_0,x_1,\ldots,x_N)x_0^{\lambda_v^+(G_S)}. Equation (29) follows at once from the Inclusion–Exclusion Formula, since X_0(G_*^v) counts colorings in A^v outside \bigcup_{e\in E(G)} A_e^v.

Let Z=\{S\subseteq E(H): H_S\text{ is connected}\}, and let the sign of S be (-1)^{|S|-k} for each S\in Z. It suffices to define a sign-reversing involution I:Z\rightarrow Z whose fixed points are positive and correspond to the internal spanning trees of H. Given S\in Z, compute I(S) by the following algorithm. Process the edges e_1,e_2,\ldots,e_m in the given order. When edge e_j is reached, do the following. If e_j\in S and e_j is the largest edge in some cycle of H_S, then return I(S)=S\setminus\{e_j\}. If e_j\not\in S and adding e_j to S creates a cycle with largest edge e_j, then return I(S)=S\cup\{e_j\}. Otherwise, proceed to the next edge. If no output is returned after processing all edges, then return I(S)=S.

Since I acts by deleting or adding one edge from S, I reverses signs when acting on a non-fixed point. To see that I is an involution, suppose we compute I(S)=S' by returning an answer when edge e_j is processed. The decisions made when processing edges e_1,\ldots,e_{j-1} do not depend on whether e_j is or is not in S. So, when we compute I(S'), the same decisions will be made for edges e_1,\ldots,e_{j-1} (namely, continue without returning an output). When the algorithm processes edge e_j, it returns I(S')=S, as needed.

Now consider the fixed points of I. On one hand, let S\in Z correspond to a spanning subgraph of H that is not a tree. This subgraph has a cycle, so there exists an edge e_j that is the largest edge in some cycle of H_S. Then I matches S to some S' of opposite sign when processing e_j (or perhaps earlier). On the other hand, by comparing the definition of internal spanning tree to the definition of I, we see that the edge set S of a spanning tree of H is a fixed point of I iff this spanning tree is internal. All such trees have k edges and therefore have positive sign.

Fix k, and let \mathcal{H} be the set of all graphs H satisfying the conditions in the theorem statement. Let \mathcal{S} be the set of all S\subseteq E(G) such that \lambda_v^+(G_S)-1=k. By the defining formula (?), the coefficient of (-z)^k in P_{G,v} is

Define a function h:\mathcal{S}\rightarrow\mathcal{H} by letting h(S)=H, where V(H) is the connected component of v in G_S. For each H\in\mathcal{H}, we can build all S\in\mathcal{S} with h(S)=H as follows. Choose S_1\subseteq E(H) such that H_{S_1} is connected, choose S_2\subseteq E(G\setminus H) arbitrarily, and let S=S_1\cup S_2. By definition, p_{\lambda_v^-(G_S)}=p_{\lambda((G\setminus H)_{S_2})}. Therefore, equation (Proof) can be rewritten as

By Lemma 32, the sum over S_1 equals f(H). By (?), the sum over S_2 equals X_{G\setminus H}.