# Bounding Schur symmetric polynomials on the unit circle

Recall the Schur polynomial in $n$ variables, indexed by the partition $\lambda$, with $\ell(\lambda) \leq n$, is given by \begin{equation} s_\lambda(x_1,\ldots, x_n) = a_{\lambda + \delta}(x_1, \ldots, x_n) / a_\delta (x_1, \ldots, x_n), \end{equation} where $\delta = (n-1,n-2,\ldots, 0)$ and $a_\lambda = \det (x_i^{\lambda_j})$ is the Vandermonde determinant of the $n \times n$ matrix whose $(i,j)$ element is $x_i^{\lambda_j}$. My notation here is consistent with Ian Macdonald's Hall polynomial book chapter 1.

Also let $e_j$ be the $j$th elementary symmetric polynomial in $n$ variables, $0 \le j \le n$. These are defined by \begin{equation} \prod_{i=1}^n (1 + x_i t) = \sum_{j=0}^n e_j(x_1,\ldots, x_n) t^j. \end{equation} By the Jacobi-Trudi formula we know that \begin{equation} s_\lambda = \det(e_{\lambda^t_i -i + j}), \end{equation} where $\lambda^t$ is the transpose of the partition $\lambda$, that is, $\lambda^t_i = \mid\{j: \lambda_j \geq i\}\mid$. Thus $s_{1^j} = e_j$, for $j \le n$.

Now we specialize to $x_i \in \mathbb{T}:= \{z \in \mathbb{C}: \mid z \mid = 1\}$. Given that \begin{equation} e_1(x_1, \ldots, x_n) = 0, \end{equation} that is, $\sum x_i = 0$, I am interested in bounding $s_\lambda$.

I have a conjecture for $\lambda = 1^{n/2}$, and $n = 4m$, $m \in \mathbb{N}$, namely, \begin{equation} \mid e_{n/2}(x_1, \ldots, x_n) \mid \le \binom{n/2}{n/4}. \end{equation} This is attained when $x_i = (-1)^i$, that is, when half of them equal $1$ and the other half equal $-1$, since \begin{equation} \prod_i (1 +x_i t) = (1-t)^{n/2} (1+t)^{n/2} = (1-t^2)^{n/2}. \end{equation}

I don't know if this conjecture is true for all qualifying $x_i$'s. Full credit will be given to solve this special case. However, I am also interested in a general conjecture for arbitrary $\lambda$.

• Would an asymptotic bound be useful for you? I can see how to get a bound of $2^{n/2}$ which is along the right lines, and one can probably save the $\sqrt{n}$ too (to almost get your conjecture). Sep 22 '15 at 0:24
• @Lucia: $2^{n/2}$ is more than good enough for me! In fact, in general I am looking for a bound of the form $\mid s_\lambda(x_1, \ldots, x_n) \mid \leq \sqrt{\mid s_\lambda(1,\ldots, 1)\mid }$. Sep 22 '15 at 2:22
• In fact for my actual application, I need to assume $\sum x_i = O(\log n)$ and just need to show $|s_\lambda(x_1,\ldots, x_n) / s_\lambda(1,\ldots, 1)|^4 s_\lambda(1,\ldots, 1) = o(1)$ uniformly over $\lambda$ as $n \to \infty$. Sep 22 '15 at 2:34
• The paper arxiv.org/pdf/1301.0634.pdf derives contour integral formulae for Schur functions that involve ratios $s_\lambda(x)/s_\lambda(1)$. These formulae might help in obtaining some of the bounds that you are after... Sep 23 '15 at 13:30
• @Suvrit: Certainly. This is our paper on mixing time of random walk on $SO(n)$, arxiv.org/pdf/1211.2031v2.pdf. The contour formula for the type B root system case when $x_1 \neq 1$ but all other $x_i =1$ is lemma 2.1. We didn't look at type A, that is, Schur polynomial case. Also appendix B gives the more general case when $x_1, \ldots, x_k \neq 1$. I also found an interesting paper of yours ,suvrit.de/papers/sra_schur_published.pdf. It seems that HCIZ formula may be a good starting point for bounding $|s_\lambda(x_1, \ldots, x_n)|$ in the general case? Sep 27 '15 at 16:52

In the special case mentioned in the problem, I'll show the bound $$|e_{n/2}(x_1,\ldots, x_n)| \le 2^{n/2}.$$ Let $$F(z) = \prod_{j=1}^{n} (1+zx_j) = C \prod_{j=1}^{n} (z + \overline{x_j}),$$ where $C= \prod_{j} x_j$ has magnitude $1$. (The roots of $F$ are $-\overline{x_j}$.)

Now by Cauchy's theorem $$e_{n/2}(x_1,\ldots,x_n) = \frac{1}{2\pi i} \int_{|z|=1} \frac{F(z)}{z^{n/2+1}} dz$$ and so in magnitude this is bounded by $$\le \sup_{|z|=1} |F(z)|.$$

Now we use a very nice Theorem of Carneiro and Vaaler (see Theorem 8.1 there), which gives a bound for the maximum of a polynomial whose roots are on the unit circle in terms of the first few symmetric power sums of the roots. Let me quote their result fully: Suppose $G(z) = \prod_{j=1}^{N}(z-\alpha_j)$ is a polynomial with $|\alpha_j| \le 1$ for all $j$. Then for any natural number $K$ we have $$\sup_{|z|\le 1} \log |G(z)| \le \frac{N}{K+1} \log 2 + \sum_{k=1}^{K} \frac{1}{k} \Big| \sum_{j=1}^{N} \alpha_j^k \Big|.$$ This is Theorem 8.1, display (8.6) of their paper.

Apply this to our polynomial $F$, with $K=1$. Since the sum of the $x_j$ is zero, by assumption, we deduce that $$\max_{|z|=1} \log |F(z)| \le \frac{N}{2} \log 2,$$ which proves the claimed estimate.

Maybe there is an easier proof of this particular bound, rather than appealing to the Carneiro-Vaaler work. I didn't see one immediately, and would be very interested if someone found an alternative approach.

Edit: In the particular case $K=1$, which is what is needed for this question, zeb kindly showed me the following one line proof: assuming $|\alpha_j|\le 1$ and $|z|\le 1$ we have $$|G(z)|^2 =\prod_{j} |z-\alpha_j|^2 \le \prod_{j} (2-2\text{Re }\overline{z}\alpha_j) \le 2^N \prod_j \exp(-\text{Re }\overline{z}\alpha_j)\le 2^N \exp\Big(\Big| \sum_j \alpha_j\Big|\Big).$$

• Thanks a lot for pointing out the nice harmonic analysis work of Carneiro and Vaaler. I wouldn't be able to find it myself in a million years. Sep 22 '15 at 3:01
• I hope it's useful to you in general; and also I found this problem very interesting, so thanks! Sep 22 '15 at 3:08
• Thanks Lucia and Zeb! I tried saddle analysis for the special case of $s_\lambda = e_{(n-1)/2}$ in trying to get the $\sqrt{n}$ factor, but the main difficulty seems to be that there could be $O(n)$ saddles. I may be able get $2^{n/2}$ with some additional small factors, because the saddle points all lie on unit circle. This is not the case for $e_k$, $k \notin \{(n\pm 1)/2\}$. Any idea to chip away a slight generalization of the above problem would be awesome. Oct 18 '15 at 23:25
• This $K = 1$ proof is exactly the same as Carneiro and Vaaler's proof of that case. In Theorem 1.5 there, you have the degree one trigonometric polynomial $u_1(x) = \big(\log{2} - (e(x) + e(-x))/2\big)/2$. This one has Fourier coefficients of $-1/4$ at $n = \pm 1$, which is better than the estimate (1.49); hence the improved coefficient under the exponential in zeb's bound. But it is the unique polynomial $u_1(x)$ for degree $N = 1$ in Carneiro and Vaaler's theorem 1.5, that zeb made explicit. Aug 31 '18 at 19:52