We construct an $N\times N$ matrix $J$ whose elements are drawn from Gaussian distribution with zero mean and variance $\frac{1}{N}$. Since we want to have different variances for different columns, so that we express the elements as $J_{ij}\sigma_j$ such that the matrix satisfies $\mathbb{Var}[J_{ij}\sigma_j] = \frac{\sigma_j^2}{N}$.

Given a Green function (analogous electrostatic potential) $$\Phi(\omega) = -\frac{1}{N}\mathbb{E}_J\big[ \ln\det((I\omega^* - J^T)(I\omega - J)) \big]$$, where $\omega\in \mathbb{C}$ and $J$ is an $N\times N$ random matrix. Since $J$ is positive semi-definite, by adding diagonal matrix $\epsilon\delta_{ij}$, determinant can be represented by a Gaussian integral over complex variables, we have $$ \Phi(\omega) = \frac{1}{N}\ln\mathbb{E}_J\big[ \int\big( \prod_i\frac{d^2 z_i}{\pi}\big)\exp\big\{ -\epsilon\sum_i|z_i|^2 - \sum_{i, j, k} z_i^*(\omega^*\delta_{ik}-J_{ik}^T)(\omega\delta_{kj}-J_{kj})z_j \big\} \big] $$

Then the potential $\phi$ is determined by

$$ \exp(-N\Phi) = \int\prod_{i = 1}^N \frac{dz_i^* dz_i}{\pi}\exp(-N\hat{Q}) $$ with the average over the distribution of $J$ values given by $$ \exp(-N\hat{Q}) = \int\prod_{i,j = 1}^N \sqrt{\frac{N}{2\pi}}dJ_{ij}\exp(-NQ)$$ where $$ Q = \sum_i\big( \frac{|\omega|^2}{\sigma_i^2} + \epsilon_i \big)\frac{z_i^*z_i}{N} + \frac{1}{2}\sum_{i, j, k} J_{ki}A_{ij}J_{kj} - \sum_{k, j}B_{kj}J_{kj} $$ and $$ A_{ij} = \frac{z_i^*z_j}{N} + \frac{z_j^*z_i}{N} + \delta_{ij}\text{ and } B_{kj} = \frac{\omega^*z_k^*z_j}{\sigma_kN}+ \frac{\omega z_j^*z_k}{\sigma_kN} $$


It confuses me how this later expression expends the previous formula for $\Phi(\omega)$ ? Could someone help me to show some detailed calculations to explain ?

How the formula later with $Q, \hat{Q}, A, B$ etc. to be derived from previous formula of $\Phi(\omega)$ ? They are supposed to be equivalent.

Remark: The question is extracted from two papers for understanding of gap calculation.

Sommers, H. J., et al. "Spectrum of large random asymmetric matrices." Physical Review Letters 60.19 (1988): 1895-1898.

Rajan, Kanaka, and L. F. Abbott. "Eigenvalue spectra of random matrices for neural networks." Physical review letters 97.18 (2006): 188104.


Your first expression for the potential $$ \Phi(\omega)=\frac{1}{N}\log E_J \int d^2z...$$ is equivalent to $$ e^{N\Phi(\omega)}=E_J \int d^2z...$$ Ok? Writing the expectation in $J$ according to its definition gives $$ e^{N\Phi(\omega)}=\int dJ e^{-N\sum_{ij}J_{ij}^2} \int \prod_i\frac{d^2 z_i}{\pi}e^{-\epsilon z^\dagger z}e^{-z^\dagger(\omega^*-(J\sigma)^T)(\omega-J\sigma)z}$$ Notice that the matrix they are interested in is $J\sigma$ (in the second paper you mentioned).

Now redefine the complex integration variables as $z_i\to z_i/\sigma_i$. This will lead to your final expressions.

| cite | improve this answer | |

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.