2
$\begingroup$

I am concerned with the asymptotical behavior of integrals like this for large $n$ $$\frac{1}{n!}\intop_{\Omega}\prod_{1\leq i<j\leq n}(x_{j}-x_{i})^{2}\,\prod_{j=1}^{n}e^{-x_{j}^{2}}dx_{j},$$ where $\Omega$ is one of these infinite rectangular regions in $\mathbb{R}^n$: $$\Omega_1^{(n)}=\{(x_1,\ldots,x_n):x_1>\lambda,\,x_2<\lambda,\,x_3<\lambda\ldots,\,x_n<\lambda\},$$ $$\Omega_2^{(n)}=\{(x_1,\ldots,x_n):x_1>\lambda,\,x_2>\lambda,\,x_3<\lambda\ldots,\,x_n<\lambda\},$$ $$\ldots$$ $$\Omega_n^{(n)}=\{(x_1,\ldots,x_n):x_1>\lambda,\,x_2>\lambda,\,x_3>\lambda\ldots,\,x_n>\lambda\},$$ and $\lambda$ is a large positive number. Some information about such integrals has already been obtained in https://mathoverflow.net/a/200622/22773

Originally, I am trying to estimate a polynomial sum of integrals over $\Omega_k$. Namely, let $a$ be any number such that $a>1$. For convenience, define

$$\Omega_0^{(n)}=\{(x_1,\ldots,x_n):x_1<\lambda,\,x_2<\lambda\ldots,\,x_n<\lambda\},$$

$$I_k^{(n)}(\lambda)=\frac{1}{n!}\intop_{\Omega_k^{(n)}}\prod_{1\leq i<j\leq n}(x_{j}-x_{i})^{2}\,\prod_{j=1}^{n}e^{-x_{j}^{2}}dx_{j},\;k=0,1,\ldots,n$$

and

$$H^{(n)}=\frac{1}{n!}\intop_{\mathbb{R}^n}\prod_{1\leq i<j\leq n}(x_{j}-x_{i})^{2}\,\prod_{j=1}^{n}e^{-x_{j}^{2}}dx_{j}=(2\pi)^{n/2}2^{-n^2/2}\prod_{k=1}^{n-1}k!$$

(the last is the classical Selberg integral). Now consider the sum

$$S^{(n)}(a,\lambda)=\sum_{k=1}^n a^k\binom{n}{k}I_k^{(n)}(\lambda),$$

so that

$$H^{(n)}=\sum_{k=0}^n \binom{n}{k}I_k=I_0^{(n)}(\lambda)+S^{(n)}(1,\lambda).$$

The conjecture to be proven is that for any $a>1$ (the case $0<a\leq 1$ is trivial) there exist positive numbers $\lambda_0$ and $C$ (which possibly depend on $a$) such that

$$\boxed{\forall n\in\mathbb{N} \;\forall \lambda>\lambda_0 \;\; S^{(n)}(a,\lambda)\leq C(H^{(n)}-I^{(n)}_0(\lambda))= C S^{(n)}(1,\lambda).}$$

It might be useful that the following asymptotics are known for the Selberg integral:

$$H^{(n)}=e^{\zeta'(-1)} \pi^n e^{-3n^2/4} 2^{n-n^2/2} n^{n^2/2-1/12}(1+o(1)),$$

where $\zeta'(-1)=-0.165...$ is the derivative of the Riemann zeta-function at $-1$.

One guess is that if there is an estimate like

$$I_{k+1}^{(n)}(\lambda)\leq\frac{f(\lambda_0)}{n}I_{k}^{(n)}(\lambda),$$

where $f(\lambda)=o(1)$ for large $\lambda$, then the previous statement follows immediately (with, say, $C=2a$).

Edit: Another important fact that I remembered is that $I_k^{(n)}(\lambda)$ is, up to a constant, the probability that a $n\times n$ random matrix from the Gaussian Unitary Ensemble has exactly $k$ eigenvalues on $(\lambda,+\infty)$. When $n\rightarrow\infty$, the analogue of this is also known as $E(k,\lambda)$ and the asymptotics of these functions as $\lambda\rightarrow -\infty$ have been examined in Tracy&Widom "Level-Spacing Distributions and the Airy Kernel". Let's call $J=(\lambda,\infty)$ and define $E_k^{(n)}(\lambda)$ as the probability of a random $n\times n$ matrix in the GUE having $k$ eigenvalues in $J$. Then the following expression is known:

$$E_k^{(n)}(\lambda)=\binom{n}{k}\frac{I_k^{(n)}}{H^{(n)}}=\frac{1}{k!}\left(\frac{\mathrm{d}}{\mathrm{d}z}\right)^k R(z,J)\mid_{z=-1},$$

where

$$R^{(n)}(z,J)\equiv R^{(n)}(z,\lambda)=\frac{1}{n!H^{(n)}}\intop_{\mathbb{R}^n}\prod_{1\leq i<j\leq n}(x_{j}-x_{i})^{2}\,\prod_{j=1}^{n}e^{-x_{j}^{2}}\left[1+z\chi(x_j)\right]\mathrm{d}x_{j}$$

and $\chi$ is the characteristic function of $J$. It is easy to see that

$$R^{(n)}(-1,\lambda)=\frac{I_0^{(n)}(\lambda)}{H^{(n)}}.$$

Using these relations, it is straightforward to check that

$$S^{(n)}(\lambda)=H^{(n)}(R^{(n)}(a-1,\lambda)-R^{(n)}(-1,\lambda)),$$

therefore the estimate that we seek is equivalent to

$$\boxed{R^{(n)}(a-1,\lambda)-R^{(n)}(-1,\lambda)\leq C(1-R^{(n)}(-1,\lambda)).}$$

So, $R^{(n)}$ is a very common object when studying the Gaussian Unitary Ensemble, but typically one is interested in its values or derivatives at $z=-1$. In my problem, however, values at arbitrary $z$ come up.

It would probably suffice to prove that $R^{(n)}(z,\lambda),\, n=1,2,3,\ldots$ is a convergent sequence of analytic functions in $z$ and it converges uniformly for all large $\lambda$.

It is also worth noting that $R^{(n)}(z,\lambda)$ is the generating function of integrals over $J^k \times \mathbb{R}^{n-k}$ (they are, in fact, integrals over $J^k$ of the $k$-point correlation functions):

$$R^{(n)}(z,\lambda)=\sum_{i=0}^n \frac{\rho_i}{i!}z^i,$$

where $\rho_0=1$ and

$$\rho_k=\frac{1}{(n-k)!H^{(n)}}\intop_{\mathbb{R}^n}\prod_{i=1}^k\chi(x_i)\prod_{1\leq i<j\leq n}(x_{j}-x_{i})^{2}\,\prod_{j=1}^{n}e^{-x_{j}^{2}}\mathrm{d}x_{j}.$$

Moreover, $R^{(n)}(z,\lambda)$ is equal to the determinant of the following $n\times n$ matrix:

$$R^{(n)}(z,\lambda)=\det[G_{ij}]_{i,j=0,1,\ldots,n-1},$$

$$G_{ij}=\frac{1}{h_i}\int_\mathbb{R} p_i(x)p_j(x)e^{-x^2}[1+z\chi(x)]\mathrm{d}x,$$

where $p_i$ is the $i$-th Hermite polynomial orthogonal with the weight $e^{-x^2}$ and $\int p_i^2(x)e^{-x^2}\mathrm{d}x=h_i$.

$\endgroup$
2
$\begingroup$

Calculations for $\boldsymbol n$ up to $\boldsymbol 6$ indicate that, for $1\leq k\leq n$, $0\leq r\leq k-1$, there are polynomials $P_{n,k,r}\in\mathbb Q[\lambda]$ of certain degrees $d_{n,k-r}$ - which are even if $k-r$ is even and odd if $k-r$ is odd - and constants $Q_{n,k}\in\mathbb Q$ such that the following holds true: $$ \boxed{\sum_{l=k}^n \binom{n-k}{l-k}\,I_l^{(n)}(\lambda) = \sum_{r=0}^{k-1}e^{-\left(k-r\right)\lambda^2}\pi^{\frac{n-k+r}2} P_{n,k,r}(\lambda)\,E^r(\lambda) + Q_{n,k}\,\pi^{\frac{n}2}E^k(\lambda)} $$ (For the left-hand side: The integral is extended over the region $\lambda<x_i<\infty$ for $1\leq i\leq k$ and $-\infty<x_i<\infty$ for $k+1\leq i\leq n$.) Here, $E(\lambda)=\frac2{\sqrt{\pi}}\int_\lambda^\infty e^{-y^2}dy$ is the complementary error function. Notice that $E(\lambda)=\frac{e^{-\lambda^2}}{\sqrt{\pi}\,\lambda}\left(1+O\left(\frac1{\lambda^2}\right)\right)$ as $\lambda\to\infty$. In particular, one then has $$ I_k^{(n)}(\lambda) \sim c_{n,k}\,\pi^{\frac{n-k}2}\,e^{-k\lambda^2} \,\lambda^{N_{n,k}} \tag{1} $$ as $\lambda\to\infty$ for some constant $c_{n,k}\in\mathbb Q$, where $N_{n,k}= \max \{d_{n,k-r}-r\,\mid\,0\leq r\leq k-1\}$. For $k=1$, this was already observed in this previous post.

Remark. The $P_{n,k,r}$ appear to depend only on $(n,k-r)$, up to a multiplicative constant.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Although this is interesting, I am constantly returning to the thought that the actual solution might be much simpler. I added some information to the post which might help in searching for a proof. The quantity that you calculated here is $\rho_l$ in my post. It is possible that some estimates for $R^{(n)}(z,\lambda)$ are already known, since it is such a common object in random matrix theory, but I haven't been able to find anything yet. $\endgroup$ – level1807 Mar 23 '15 at 10:37
  • $\begingroup$ Your additional info is certainly helpful. In order to get what is calculated above you need to replace $\chi_{J^n}$ with $\chi_{J^k\times{\mathbb R}^{n-k}}$. $\endgroup$ – ifw Mar 23 '15 at 12:38
1
$\begingroup$

... I am constantly returning to the thought that the actual solution might be much simpler.

I do not think that you are right. To obtain estimates for fixed $n$ is not hard, the big issue is uniformity with respect to $n$. Of course, it is possible that the specific context allows to play some algebraic tricks which facilitate the computations. But in the end it comes down to getting these integrals estimated, and I am convinced that any argument that does not take into account the fine structure of the problem is doomed to fail.

Here is a heuristic argument to support my point of view: You are seeking to prove an estimate like $$ I_{k+1}^{(n)}(\lambda) \leq \frac{f(\lambda)}{n}\,I_k^{(n)}(\lambda) $$ with $f(\lambda)= o(1)$ as $\lambda\to\infty$. Inserting the asymptotics from (1), so neglecting remainder terms, one gets $$ \frac{c_{n,k+1}}{c_{n,k}}\,e^{-\lambda^2}\lambda^{N_{n,k+1}-N_{n,k}} \leq \frac{C\, f(\lambda)}{n}. $$ The calculations performed before suggest that $d_{n,1}=2n-3$, $d_{n,2}\geq 2n-2$, and $d_{n,k}=d_{n,2}-k+2$ for $k\geq2$. If so, then one has $N_{n,1}=2n-3$ and $N_{n,k}=d_{n,2}-k+2$ for $k\geq2$.

In case $k=1$, this yields $$ \frac{c_{n,2}}{c_{n,1}}\,e^{-\lambda^2}\lambda^{d_{n,2}-2n+3} \leq \frac{C\, f(\lambda)}{n}. $$ Now, if $d_{n,2}\geq 2n + \gamma_0\log n +\gamma_1$ for some constants $\gamma_0>0$ and $\gamma_1\in\mathbb R$ (which is not unlikely to hold), then one would need $$ \frac{c_{n,2}}{c_{n,1}} = O\left(n^{-1-\gamma_0}\right) $$ as $n\to \infty$.

In case $k\geq2$, one obtains $$ \frac{c_{n,k+1}}{c_{n,k}}\,\frac{e^{-\lambda^2}}{\lambda} \leq \frac{C\, f(\lambda)}{n}, $$ so here one would need $$ \frac{c_{n,k+1}}{c_{n,k}} = O\left(n^{-1}\right) $$ as $n\to\infty$, but uniformly so in $2\leq k< n$.

Admittedly, these arguments are not rigorous, but in my opinion they demonstrate that one has to proceed with extreme care when doing the estimates.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Yes, I noticed that too. Combinatorics in expanding the polynomial is pretty awful. I am also hoping that the traditional GUE theory might provide us with some magical properties of $R^{(n)}$. Specialists might know. $\endgroup$ – level1807 Mar 23 '15 at 20:22

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.