10

Recently, I am interested in the graph polynomial of the product of cycles. Let $G = (V , E)$ be an undirected multi-graph with vertex set $\{1,\cdots,n\}$. The graph polynomial of $G$ is defined by $$f_G(x_1,x_2,\cdots,x_n)=\prod_{1\leq i<j\leq n, (i,j)\in E}(x_i-x_j).$$

Conjecture: Let $G $ be the Cartesian product graph $C_{2n+1}\Box C_{2m}$, then the coefficient of $x_1^2x_2^2\cdots x_{(2n+1)(2m)}^2$ in the the graph polynomial $f_G(x_1,x_2,\cdots, x_{(2n+1)(2m)})$ is nonzero.

For $C_3\Box C_{2n}$, the conjecture is true. See the coefficient of a special term in the expansion of the graph polynomial

This conjecture generalized the result about $C_3\Box C_{2n}$. I think it may be true. But I have no idea about the proof on the general cases.

I hope someone could give suggestions about the conjecture. I will appreciate it even if given some special cases for the conjecture such as $n=2,3$,etc.

Jacob.Z.Lee
  • 767
  • 3
  • 11

1 Answers1

9

The following proof is obtained jointly with Alexey Gordeev.

We start with a general

Lemma. Let $H=(X,E)$ be a $2d$-regular graph on the vertex set $X$, $G=H\square C_k$ be the product of $H$ and a cycle of length $k$. Fix a field $\mathbb{F}$ and a subset $A\subset \mathbb{F}$, $|A|=d+2$. Let $\mathcal{U}$ denote the set of all proper $d$-colorings $u:X\to A$ of the vertices of $H$ with colors taken from $A$. Consider the square matrix $M$ with rows and columns indexed by the elements of $\mathcal{U}$: $$ M_{u,v}=f_H(u)\cdot \prod_{x\in X} \frac{u(x)-v(x)}{\prod_{b\in A\setminus u(x)} (u(x)-b)} $$ for two proper colorings $u,v\in \mathcal{U}$. Here $f_H(u)=\prod_{(x,y)\in E} (u(y)-u(x))$ is a graph polynomial of $H$ (with somehow fixed sign.) Then $$ \left[\prod_i x_i^{d+1}\right] f_G=\mathrm{tr}\,M^k, $$ where the variables $x_i$ correspond to all $k\cdot |X|$ vertices of $G$.

Proof. Use the formula for the coefficient of $f_G$: $$ \left[\prod_i x_i^{d+1}\right] f_G=\sum_{a_i\in A} \frac{f_G(a_1,\ldots)}{\prod_i \prod_{b\in A\setminus a_i} (a_i-b)} \quad \quad \quad \quad (1) $$ The non-zero summands in the RHS of (1) correspond to proper colorings of $G$. Any proper coloring of $G$ corresponds to a sequence $u_1,\ldots, u_k\in \mathcal{U}$ of the proper colorings of $H$. For such a sequence, the corresponding summand in RHS of (1) reads as $$ M_{u_1,u_2}\cdot M_{u_2,u_3}\cdot \ldots \cdot M_{u_k,u_1}. $$ The sum of such products is nothing but $\mathrm{tr}\, M^k$. In our situation $H=C_{2n+1}$, $k=2m$ is even. We choose the field $\mathbb{F}=\mathbb{C}$ and the set $A=\{1,w,w^2\}$, where $w=e^{2\pi i/3}$. The reason why $\mathrm{tr}\, M^{2m}\ne 0$ is that $M$ is a non-zero antihermitian matrix. That implies that the eigenvalues of $M$ are purely imaginary and not all of them are equal to 0, that yields $(-1)^m\mathrm{tr}\, M^{2m}> 0$. It remains to prove that $M$ is antihermitian ($M$ is certainly non-zero). That is the point why we have chosen such a set of colors. Let $u=[u_1,\ldots,u_{2n+1}]\in \mathcal{U}$ be a proper 3-coloring of $C_{2n+1}$ with 3 colors $1,w,w^2$. Denote by $u_i^*$ the unique element of the set $\{1,w,w^2\}\setminus \{u_i,u_{i-1}\}$. Assuming that the sign of $f_{H}$ is chosen so that $f_H(u)=\prod_i (u_i-u_{i-1})$ (the indices are cyclic modulo $2n+1$), applying the obvious relation $$\frac1{u_i-u_{i}^*}=\frac{u_i-u_{i-1}} {\prod_{b\in A\setminus u_i} (u_i-b)}$$ we get $$ M_{u,v}=\prod_{i=1}^{2n+1}\frac{u_i-v_i}{u_i-u_i^*}.\quad \quad (2) $$ So the relation to check is $$ M_{v,u}=-\overline{M_{u,v}}.\quad \quad \quad \quad (3) $$ Applying (2) and substituting $\bar{z}=1/z$ for roots of unity $z=u_i,v_i,u_i^*$ we simplify (3) to the following: $$ \prod_i \frac{u_i-u_i^*}{u_i}=-\prod_i \frac{v_i^*-v_i}{v_i^*} \quad \quad (4) $$ provided that $u_i\ne v_i$ for all $i$ (if $u_i=v_i$ for some $i$, then $M_{u,v}=M_{v,u}=0$). Denote $\varepsilon_i=u_i/u_{i-1}$, then $\varepsilon_i\in \{w,w^2\}$ and $u_i^*=u_i \varepsilon_i$. Analogously write $v_i^*=v_i \delta_i$ so we rewrite (4) as $$ \prod_i(1-\varepsilon_i)=-\prod_i (1-\overline{\delta_i})=(-1)^{1+(2n+1)}\prod \delta_i^{-1}\prod(1-\delta_i). $$ Obviously $\prod \varepsilon_i=\prod \delta_i=1$ so this in turn rewrites as $$ \prod_i(1-\varepsilon_i)=\prod_i (1-\delta_i). \quad \quad (5) $$ Now we have $1-\varepsilon=\pm i\sqrt{3}\varepsilon^2$ for $\varepsilon\in \{w,w^2\}$ (the signs are distinct for $w$ and $w^2$). Substituting this for $\varepsilon_i$'s and $\delta_i$'s and using $\prod \delta_i=\prod \varepsilon_i=1$, we reduce (5) to the following fact: the total number of $\varepsilon_i$'s and $\delta_i$'s which are equal to $w$ is even. Call the index $i$ white if $u_i=w\cdot v_i$ and black if $u_i=w^2\cdot v_i$. Then $\varepsilon_i=\delta_i$ if $i-1,i$ have the same color and $\varepsilon_i\ne \delta_i$ if $i-1,i$ have different color. Obviously there are even number of indices $i$ of the second type, thus the result.

Fedor Petrov
  • 102,548
  • @Petrov Exellent work! Thanks a lot. There maybe a mistake in (4), I think it should be $$\prod_i \frac{u_i-u_i^}{u_i^}=-\prod_i \frac{v_i^*-v_i}{v_i} \quad \quad (4).$$ In the lemma , if $ k=2m+1$ is the matrix $M$ still antihermitian matrix? Furthermore, I conjeture the corresponding cofficient of the graph polynomial of $C_{2n+1}\Box C_{2m+1}$ is zero. – Jacob.Z.Lee Jun 18 '19 at 12:54
  • @Jacob.Z.Lee I think, there is no mistake, and what you write is just the same equation upto interchange of $u$ and $v$. – Fedor Petrov Jun 18 '19 at 17:21
  • @Jacob.Z.Lee the matrix $M$ does not depend on $k$, it depends only on $H$. So yes, for the product of two odd cycles the coefficient is 0. But this is clear a priori: if we reverse the cyclic order in each $C_{2n+1}$, we get the same polynomial with the sign changed, thus corresponding coefficent $A$ equals to $-A$. – Fedor Petrov Jun 18 '19 at 17:44
  • @Petrov More specifically, if reverse the cyclic order in $C_{2n+1}$ and $C_{2m+1}$, we get the same graph polynomial and the sigh of each $M_{u,v}$ changed. thus we get $$tr (M^{2m+1})=-tr (M^{2m+1}).$$ Hence we get the $$ tr (M^{2m+1})=0.$$ Is it right? – Jacob.Z.Lee Jun 19 '19 at 01:29
  • @Jacob.Z.Lee You do not even need to use that the coefficient is the trace here. It vanishes by its own right. – Fedor Petrov Jun 19 '19 at 07:09
  • @Petrov ~ I do have difficulties to understand it. Let us take for an example, $C_3\Box C_3.$ Its graph polynomial $$f_G=(x_1- x_2)(x_2- x_3)(x_3- x_1)(x_4- x_5)(x_5- x_6)(x_6- x_4)(x_7- x_8)(x_8- x_9)(x_9- x_7)(x_1- x_4)(x_4- x_7)(x_7- x_1)(x_2- x_5)(x_5- x_8)(x_8- x_2)(x_3- x_6)(x_6- x_9)(x_9- x_3).$$ If we reverse the cyclic order in each horizontal and vertical $C_3$, we get $$f'_G=(x_2- x_1)(x_3- x_2)(x_1- x_3)(x_5- x_4)(x_6- x_5)(x_4- x_6)(x_8- x_7)(x_9- x_8)(x_7- x_9)(x_4- x_1)(x_7- x_4)(x_1- x_7)(x_5- x_2)(x_8- x_5)(x_2- x_8)(x_6- x_3)(x_9- x_6)(x_3- x_9).$$ – Jacob.Z.Lee Jun 19 '19 at 12:52
  • Obviously, $f_G=f'G$. But how do we know the sign of the coefficient $[\prod{i=1}^9 x_i^2]f_G$ is changed? would you please give me an explaination about it? Thanks. – Jacob.Z.Lee Jun 19 '19 at 12:52
  • No, $f_G=-f_G'$. You may think about it also as follows: imagine that you take something from any bracket $x_i-x_j$ and get $\pm \prod x_i^2$. Now take the other term from each bracket. You get $\mp \prod x_i^2$. Thus the summands do cancel. – Fedor Petrov Jun 19 '19 at 14:49
  • @Petrov There are 18 factors in $f_G$. I take out $-1$ from each factor to get the polynomial $f'_G$. Since $(-1)^{18}=1$, we get $f_G=f'_G.$ Why $f_G=-f_G'$? Unless my $f'_G$ is not your $f'_G.$ I donnot catch your method yet. would you please take the $C_3\Box C_3$ for example to illustrate your method? I will appreciate it .The example will be good to show the spirit. Thanks! – Jacob.Z.Lee Jun 20 '19 at 10:22
  • Oh, indeed 18, sorry. Then this obvious argument does not work at all, while the eigenvalues argument works: the trace of $M^{2m+1}$ is purely imaginary, on the other hand this coefficient is real, therefore it is 0. – Fedor Petrov Jun 20 '19 at 12:06
  • @Petrov Yes. it need to rely on the trace method. – Jacob.Z.Lee Jun 20 '19 at 12:14