I am trying to compute the following gaussian integral over all possible real matrices $J$:

$$I=\int \exp\left\{-\frac{N}{2}\text{Tr}\left[\mathbf{J}\mathbf{A}\;\mathbf{J}^T+2\mathbf{BJ}-\gamma \mathbf{JJ} \right]\right\}\mathrm{d}\mathbf{J}$$

Where $\mathbf{A}$ and $\mathbf{B}$ are Hermitian matrices.

When $\gamma=0$ I can complete the square and integrate this Gaussian integral without any problem (assuming I know the eigenvalues and determinant of $\mathbf{A}$):


However for general $\gamma\in \mathbb{R}$ I cannot seem to know how to evaluate this integral by completing the square: $\mathbf{J}^T\mathbf{A}\;\mathbf{J}+2\mathbf{BJ}-\gamma \mathbf{JJ}$

$\mathbf{J}$ is real but not symmetric. when $\gamma=0$ this integral converges so I do not see any reason why it would not be generalised to general $\gamma$ with an appropriate $\mathbf{A}$.

Any remark or advice is always appreciated. Thank you.

Edit : A different way to express the integral $I$ is the following:

$$I=\int \left(\prod_{ij}\mathrm{d}J_{ij}\right)\exp\left\{-\frac{N}{2} \sum_{i, j, k} J_{k i} A_{i j} J_{k j}+N\sum_{k, j} B_{k j} J_{k j}+\frac{N\gamma}{2}\sum_{ij}J_{ij}J_{ji}\right\}$$

Assuming I already know the eigenvalues of $\mathbf{A}$ and thus $\det(\mathbf{A})$, how can I compute the integral $I$?

  • $\begingroup$ Should that be $- \gamma \,J^TJ$ instead? $\endgroup$ Feb 25 '20 at 9:32
  • $\begingroup$ Unfortunately not, otherwise I could have included it in the expression of $\mathbf{A}$. Here it is meant to be $\text{Tr}\mathbf{JJ}=\sum_{ij}J_{ij}J_{ji}$. $\endgroup$
    – Matt
    Feb 25 '20 at 9:35
  • 1
    $\begingroup$ In that case, there is no way to "complete the square" here. Do you have a way of computing the integral in a simple case such as $B = 0$ and $A = I$? I suspect that the integral does not converge for $A = 0$. $\endgroup$ Feb 25 '20 at 9:38
  • 1
    $\begingroup$ Ah okay, I was hoping that there might have been a typo at some point (: $\endgroup$ Feb 25 '20 at 9:42
  • 1
    $\begingroup$ Note, by the way, that the map $J \mapsto PJP^{-1}$ has determinant $1$, so there is no factor to multiply by here. $\endgroup$ Feb 25 '20 at 10:04

In principle you could write the $J$ matrix as a $N\times N$ long "super"vector and then you would have a "simple" quadratic form $J_{ij} \Gamma^{ijkl} J_{kl}$ where all the transposes etc are encoded in the $\Gamma$ super matrix. By redefining an index pair ${ij}=\alpha$ you could put the $\Gamma$ tensor in a "super" matrix form and find the relevant determinant etc. It looks daunting, but maybe there are some shortcuts, e.g. there is a super matrix $\mathcal T$ that transforms any "super"vector into it's transpose.


Following the idea of writing the $J$ matrix as a $N^2$ long "super" vector: my starting point is the following result: $$\int_{\mathbb{R}^{n^{2}}} \exp\left\{-\frac{1}{2} \mathbf{x}^{T} \mathbf{\Sigma} \mathbf{x}\right\}\mathrm{d}\mathbf{x}=\frac{(2 \pi)^{n^{2} / 2}}{\sqrt{\operatorname{det}(\Sigma)}}$$ Where $\mathbf{\Sigma}\in \mathbb{R}^{N^2\times N^2}$

Integrating over all $\mathrm{d}J_{ij}$ is equivalent to integrating over all $\mathrm{d}x_i$ if we find an appropriate transformation. The main obstacle is this $\sum_{ij}J_{ij}J_{ji}$ sum. So I will consider here the following integral:

$$I=\int\left(\prod_{i j} \mathrm{d} J_{i j}\right) \exp \left\{-\frac{1}{2} \sum_{i, j, k} J_{k i} A_{i j} J_{k j}-\frac{b}{2} \sum_{i j} J_{i j} J_{j i}\right\}$$

First, we can write: $$b\sum_{ij}J_{ij}J_{ji}=\sum_{i=1}^{n} \sum_{j=1}^{n} \sum_{k=1}^{n} \sum_{l=1}^{n} J_{i j} J_{k l} \;\delta_{i l} \delta_{j k}b$$ Similarly: $$\sum_{i, j, k} J_{k i} A_{i j} J_{k j}=\sum_{i, j, k} J_{i j} A_{jk} J_{i k}=\sum_{i, j, k,l} J_{i j} A_{jk} J_{l k}\delta_{il}=\sum_{i, j, k,l} J_{i j} A_{jl} J_{k l}\delta_{ik}$$ Thus our integral is now: $$I=\int\left(\prod_{i j} \mathrm{d} J_{i j}\right) \exp \left\{-\frac{1}{2} \sum_{i, j, k,l} J_{i j} A_{jl} J_{k l}\delta_{ik}-\frac{1}{2}\sum_{i,j,k,l}^{n}J_{i j} J_{k l} \;\delta_{i l} \delta_{j k}b\right\}$$

$$\implies I=\int\left(\prod_{i j} \mathrm{d} J_{i j}\right) \exp \left\{-\frac{1}{2} \sum_{i, j, k,l} J_{i j} \left(A_{jl} \delta_{ik} +\delta_{i l} \delta_{j k}b\right)J_{k l}\right\}$$ Now, we can define $x$ such that : $$x:=\left(\begin{array}{c} J_{11} \\ J_{12} \\ \vdots \\ J_{1 n} \\ J_{21} \\ J_{22} \\ \vdots \\ J_{n n} \end{array}\right) \in \mathbb{R}^{n^{2}}$$

Therefore: $$J_{i j}= x_{n(i-1)+j} \quad \forall i, j \in \mathbb{N} \cap[1, n]$$ \begin{equation} \Sigma_{n(i-1)+j, n(k-1)+l}= A_{jl} \delta_{ik} +\delta_{i l} \delta_{j k}b\quad \forall i, j, k, l \in \mathbb{N} \cap[1, n] \end{equation}

For N=2 we have: $$\Sigma=\left( \begin{array}{cccc} A_{1,1}+b & A_{1,2} & 0 & 0 \\ A_{2,1} & A_{2,2} & b & 0 \\ 0 & b & A_{1,1} & A_{1,2} \\ 0 & 0 & A_{2,1} & A_{2,2}+b \\ \end{array} \right)$$ For general $N\in \mathbb{N}$ we recover : $\Rightarrow x^{T} \Sigma x=\sum_{i,j,k,l}^{n} J_{i j}\left(A_{jl} \delta_{ik} +\delta_{i l} \delta_{j k}b\right) J_{k l}$

Using our initial result we see that $$I=\frac{(2 \pi)^{n^{2} / 2}}{\sqrt{\operatorname{det}(\Sigma)}}$$

Thus our problem is reduced to finding the determinant of the following $N^2\times N^2$ matrix: $\Sigma_{n(i-1)+j, n(k-1)+l}= A_{jl} \delta_{ik} +b \delta_{i l} \delta_{j k}$. This is where my answer ends. It would be interesting to derive a formula to compute the determinant of $\Sigma$.

After tinkering I find that:

In the case $N=2$:


In the case $N=3$:


Where $\operatorname{Tr}\left(\Lambda^2A\right)$ is the trace of the $k$th exterior power of A, taken from this formula: $$ |A-bI|=-\sum_{k=0}^{n} b^{n-k}(-1)^{k} \operatorname{tr}\left(\Lambda^{k} A\right) $$ It would be interesting to solve for general $N$. I might create a new post on this question.

  • 1
    $\begingroup$ So as I understand the trick was to used a delta twice and not once, what I don't get is why appears a factor of $|A-Ib|$ in the cases you wrote $\endgroup$
    – Dabed
    Mar 9 '20 at 5:16
  • 1
    $\begingroup$ Thank you for your comment. Indeed, I made a mistake: the correct term is $|A+bI|$ and not $|A-Ib|$. I edited my answer. $\endgroup$
    – Matt
    Mar 9 '20 at 8:55
  • $\begingroup$ Actually now I have another problem I read that we have $\Sigma=A_{jl} \delta_{ik} +b \delta_{i l}\delta_{j k}=A+bI$ but is $I=\delta_{i l}\delta_{j k}$ shouldn't $\delta_{i l}\delta_{j k}$ be identified to a real value (trace) instead of a matrix or I'm just reading things wrong $\endgroup$
    – Dabed
    Mar 9 '20 at 11:21
  • $\begingroup$ I am not sure I understand correctly: indeed $A_{j l} \delta_{i k}+b \delta_{i l} \delta_{j k}\neq A+b I$. I edited my answer and added a representation of $\Sigma$ in the case $N=2$. Did I misunderstood your problem? $\endgroup$
    – Matt
    Mar 9 '20 at 11:25
  • $\begingroup$ I thought I did but now with your example I think I get it $\endgroup$
    – Dabed
    Mar 9 '20 at 12:16

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.