While working with Schur polynomials I found what seems like a nice identity, and I wonder if it has a simple proof.

Notation: Suppose $d,n\in\mathbb{N}$, and $\lambda =(\lambda_1,\dots,\lambda_n)$ is an ordered partition of $d$, that is $\lambda_1 \ge \dots \ge \lambda_n \ge 0$, and $\lambda_1 + \dots + \lambda_n = d$.

Let $\Lambda(d,n)$ be the set of all such partitions. For each $\lambda \in \Lambda(d,n)$ define: $$ N(\lambda) = \prod_{1\le i < j \le n} \frac{\lambda_i - \lambda_j + j - i}{j - i}, $$ and $$ W(\lambda) = \prod_{1 \le i \le n} (\lambda_i + n - i)!. $$ Remark: $N(\lambda)=s_\lambda(1,\dots,1)$, where $s_\lambda$ is the Schur polynomial associated with the partition $\lambda$. Now, define $$ A(d,n) = \sum_{\lambda \in \Lambda(d,n)} \frac{N(\lambda)^2}{W(\lambda)}. $$ It seems the following identity holds: $$ A(d,n) = \left( \prod_{k=0}^{n-1} k! \right)^{-1} \frac{n^d}{d!}.$$ Is there a simple proof\explanation?

  • $\begingroup$ It seems that the identity follows from a "Cauchy like identity" (that I guessed): $$\sum_\lambda \frac{s_\lambda(x) s_\lambda(y)}{W(\lambda)} = \prod_{j\ge 1} \frac{e^{x_j y_j}}{(j-1)!}.$$ I don't know if this "identity" (in infinitely many variables) is actually correct\known, or can be made rigorous. $\endgroup$ – sometempname Jul 25 '16 at 22:05

We can use the fact that $N(\lambda)=\left|\text{SSYT}(\lambda)\right|$, the number of semistandard Young tableaux of shape $\lambda$, and that $d!\cdot\left(\frac{\prod_{1\le i < j\le n}(\lambda_i - \lambda_j + j - i)}{\prod_{1 \le i \le n} (\lambda_i + n - i)!}\right)=\left|\text{SYT}(\lambda)\right|$, the number of standard Young tableaux of shape $\lambda$. By some simple rearranging your identity becomes $$\sum_{|\lambda|=d}\left|\text{SSYT}(\lambda)\right|\left|\text{SYT}(\lambda)\right|=n^d.$$ Now, if you track the left hand side through the Robinson-Schensted-Knuth correspondence, you'll find that it corresponds to $d\times n$ matrices with row sums all $1$. So there are exactly $n^d$ of these.

| cite | improve this answer | |
  • $\begingroup$ Where do I miss the $d!$ factor in this argument? $\endgroup$ – sometempname Jul 25 '16 at 23:03
  • $\begingroup$ I had a typo in the hook length formula :) $\endgroup$ – Gjergji Zaimi Jul 25 '16 at 23:04
  • $\begingroup$ Nice proof. Do you expect there's a corresponding "weighted Cauchy identity" for Schur polynomials? (I just guessed the form) $\endgroup$ – sometempname Jul 25 '16 at 23:13
  • 3
    $\begingroup$ If you like representation theory better than RSK, the displayed equation takes the dimension of both sides of Schur-Weyl duality: $\sum_{|\lambda|=d} Sp_{\lambda} \boxtimes S_{\lambda}(V) \cong V^{\otimes d}$ where $V$ is a vector space of dimension $n$, $Sp_{\lambda}$ is the Specht module and $S_{\lambda}$ is the Schur functor. The isomorphism is of representations of $S_d \times GL(V)$. $\endgroup$ – David E Speyer Jul 26 '16 at 13:45

I found a proof which I don't really like, but I'll share it.

For two (real) diagonal matrices $A,B$, the Harish-Chandra-Itzykson-Zuber (HCIZ) integral is $$ I(A,B) = \int_{U(n)} e^{\rm{tr}(U^* A U B)} \, \rm{d} U = c_n \frac{\det\left([e^{a_j b_k}]_{j,k=1}^n\right)}{\Delta(a)\Delta(b)}, $$ where $\Delta(a) = \prod_{j<k} (a_k - a_k)$ is the Vandermonde determinant, and $c_n = \Delta([1,\dots,n]) = \prod_{k=1}^{n-1} k!$ (the integration is with respect to Haar measure on the unitary group).

Using an infinite version of the Cauchy-Binet formula, we can write $$ \det\left(\{e^{a_j b_k}\}_{j,k=1}^n\right) = \sum_\lambda \det\left(\left[\frac{a_j^{\lambda_k+n-k}}{\sqrt{(\lambda_k+n-k)!}}\right]_{j,k=1}^n\right) \det\left(\left[\frac{b_k^{\lambda_j+n-j}}{\sqrt{(\lambda_j+n-j)!}}\right]_{j,k=1}^n\right), $$ where the sum is over all partitions $\lambda$ of size $n$. Using the fact $$ s_\lambda(a) = \frac{\det\left(a_j^{\lambda_k+n-k}\right)}{\Delta(a)},$$ we find $$ I(A,B) = c_n \sum_\lambda \frac{s_\lambda(a)s_\lambda(b)}{\prod_{j=1}^n (\lambda_j+n-j)!} = c_n \sum_\lambda \frac{s_\lambda(a)s_\lambda(b)}{W(\lambda)}. $$ If we substitute $A = B = t \cdot \rm{Id}_n$, then we find $$ e^{t^2 n} = c_n \sum_{d\ge 0} t^{2d} \sum_{\lambda\in\Lambda(d,n)} \frac{\left(s_\lambda(1,\dots,1)\right)^2}{W(\lambda)}. $$ We get the identity by comparing coefficients of $t$.

| cite | improve this answer | |
  • $\begingroup$ What are the vectors $a$ and $b$ in your formulae? I suppose the eigenvalues of $A$ and $B$? $\endgroup$ – Wolfgang Jul 26 '16 at 8:15
  • $\begingroup$ Fixed, thanks. Real diagonal matrices are sufficient for the argument. $\endgroup$ – sometempname Jul 26 '16 at 14:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.