The question I have is quite specific. So in the hope that this post might help others in the future, my problem boils down to solving the following integral:

$$I_\tau=\int \prod_{i, j=1}^{N} d J_{i j} \exp \left\{-\frac{N}{2(1-\tau^2)}\left( \sum_{i, j, k} J_{k i} A_{i j} J_{k j}-\tau\sum_{ij}J_{ij}J_{ji}\right)\right\}$$ Assuming we can compute the determinant of $A$, is there a general identity to compute $I_\tau$?

Simplifying their argument, in the following paper [1], the authors compute the following Gaussian integral over all $N\times N$ real asymmetric matrices $J_{ij}$:

$$I=\int \prod_{i, j=1}^{N}d J_{i j} \exp \left\{-\frac{N}{2} \sum_{i, j, k} J_{k i} A_{i j} J_{k j}\right\}$$

Where $A_{i j}=\frac{z_{i}^{*} z_{j}}{N}+\frac{z_{j}^{*} z_{i}}{N}+\delta_{i j}$ and $z_i\in \mathbb{C}$.

If I define the vector $\mathbf{x}$ such that: $$\mathbf{x}:=\left(\begin{array}{c} J_{11} \\ J_{12} \\ \vdots \\ J_{1 n} \\ J_{21} \\ J_{22} \\ \vdots \\ J_{n n} \end{array}\right) \in \mathbb{R}^{n^{2}}$$ $$\implies x_{N(i-1))+j}=J_{ij}$$ Then we can represent $I$ as the following integral: $$I=\int \left(\prod_{i, j=1}^{N} \mathrm{d}x_{N(i-1))+j} \right)\exp \left(-\frac{N}{2} \mathbf{x}^T\Sigma \mathbf{x}\right)$$ Where $\Sigma$ simply is: $$\Sigma=\left(\begin{array}{cccc} \mathbf{A} & 0 & \cdots & 0 \\ 0 & \mathbf{A} & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & \mathbf{A} \end{array}\right)$$

Thus $I\propto \frac{1}{\sqrt{|\Sigma|}}=|A|^\frac{-N}{2}$.

Defining $r\equiv\frac{1}{N}\sum_i |z_i|^2$, the authors observe that to order $1/\sqrt{N}$, the matrix $A$ has all but two of its eigenvalues equal to one. The two exceptions are the eigenvectors $z_i$ and $z^*_i$ , both with eigenvalue $(1+r)$. This observation allows them to compute the determinant of $\Sigma$ and thus the Gaussian integral $I$.

Now, considering the integral $I_\tau$ using the same method: $$I_\tau=\int \left(\prod_{i j} \mathrm{d} J_{i j}\right)\exp \left\{-\frac{N}{2(1-\tau^2)}\left( \sum_{i, j, k} J_{k i} A_{i j} J_{k j}-\tau\sum_{ij}J_{ij}J_{ji}\right)\right\}$$ $$I_\tau=\int\left(\prod_{i j} \mathrm{d} J_{i j}\right) \exp \left\{-\frac{N}{2(1-\tau^2)} \sum_{i j, k, l} J_{i j}\left(A_{j l} \delta_{i k}-\tau\delta_{i l} \delta_{j k}\right) J_{k l}\right\}$$ $$\implies I_\tau=\int\left(\prod_{i, j=1}^{N} \mathrm{d} x_{N(i-1))+j}\right) \exp \left(-\frac{N}{2} \mathbf{x}^{T} \Sigma \mathbf{x}\right)$$ Where this time: $$\Sigma_{n(i-1)+j,n(k-1)+l}=A_{j l} \delta_{i k}-\tau\delta_{i l} \delta_{j k} \quad \forall i, j, k, l \in \mathbb{N} \cap[1, N]$$

For large $N$, how can we compute the determinant of $\Sigma$ this time? It seems to be a non-trivial question.

Any remark or comment is always appreciated, thank you.

[1]: Rajan, K., & Abbott, L. F. (2006). Eigenvalue spectra of random matrices for neural networks. Physical review letters, 97(18), 188104.

Link: https://www.researchgate.net/profile/Kanaka_Rajan/publication/6643146_Eigenvalue_Spectra_of_Random_Matrices_for_Neural_Networks/links/0912f50f7347f4863b000000.pdf


Only the symmetric part of $A$ contributes to the integrand, so we may assume $A$ is symmetric and diagonalize it as $A=O\Lambda O^T$ with $O$ orthogonal and $\Lambda={\rm diag}\,(\lambda_1,\lambda_2,\ldots\lambda_N)$. The orthogonal matrix may be removed by a change of variables, so without loss of generality we can take a diagonal $A$. So we seek the integral $$I=\int \prod_{i, j=1}^{N} d J_{i j} \exp \left\{-\frac{N}{2(1-\tau^2)}\left( \sum_{i, j} \lambda_j J_{ij}^2-\tau\sum_{ij}J_{ij}J_{ji}\right)\right\}.$$ I now proceed similarly to here. Decompose the sum over $i,j$ as $$\sum_{ij}\lambda_j J_{ij}^2-\tau\sum_{ij}J_{ij}J_{ji}=$$ $$\qquad\qquad=\sum_{i}(\lambda_i-\tau)J_{ii}^2+\sum_{i<j}\left(\lambda_j J_{ij}^2+\lambda_i J_{ji}^2-2\tau J_{ij}J_{ji}\right)$$ $$\qquad\qquad\equiv\sum_{i} A_i+\sum_{i<j} B_{ij}.$$ Then perform the Gaussian integrals separately for each term in the sum, $$I=\left(\prod_{i=1}^N\int e^{-\beta A_i}dJ_{ii}\right)\left(\prod_{i<j=1}^{N}\int\int e^{-\beta B_{ij}}dJ_{ij}dJ_{ji}\right)$$ $$\qquad\qquad=(\pi/\beta)^{N^2/2}\left(\prod_{i=1}^N(\lambda_i-\tau)^{-1/2}\right)\left(\prod_{i<j=1}^N(\lambda_i\lambda_j-\tau^2)^{-1/2}\right),$$ where I have defined $\beta=\frac{1}{2}N(1-\tau^2)^{-1}$, and assumed that $\beta>0$, $\lambda_i>\tau$ for all $i$.

So knowledge of only the determinant of $A$ is not sufficient to evaluate the integral, you need to know the individual eigenvalues -- not just their product. In particular, if $\tau$ happens to be close to one particular eigenvalue of $A$, then that eigenvalue will dominate the integral.


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.