Hi I am trying to calculate the expected value of $$ \mathbb{E}\big[x_i x_j...x_N\big]=\int_{-\infty}^\infty x_ix_jx_k...x_N \exp\bigg({-\sum_{i,j=1}^N\frac{1}{2}x^\top_i A_{ij}x_j}-\sum_{i=1}^Nh_i x_i\bigg)\prod_{i=1}^Ndx_i, $$ note these are higher order correlation functions for a Gaussian generating functional. Also the matrix $A_{ij}=A^\top_{ij}$ (real symmetric) and is also positive definite, thus the eigenvalues of $A_{ij}$ all satisfy $\lambda_i>0$. Note the generating functional is given by $$ \mathcal{F}(h)=\int_{-\infty}^\infty \exp\bigg({-\sum_{i,j=1}^N\frac{1}{2}x^\top_i A_{ij}x_j}-\sum_{i=1}^Nh_i x_i\bigg)\prod_{i=1}^Ndx_i=\frac{(2\pi)^{N/2}}{\sqrt{\det A_{ij}}}\exp\big( \frac{1}{2}\sum_{i,j=1}^N h_i A^{-1}_{ij}h_j\big) $$ where I used $\det(A)=\prod_{i=1}^N \lambda_i$. We calculate this by finding the minimum of the quadratic form $$ \frac{\partial}{\partial x_k}\bigg(\sum_{i,j=1}^{N} \frac{1}{2} x_i A_{ij} x_j-\sum_{i=1}^N h_i x_i \bigg)=\sum_{j=1}^{N} A_{kj}x_j- h_k=0. $$ In order to solve this we need to introduce the inverse matrix of $A$ given by $A^{-1}$. Thus we can write the solution as $$ x_i=\sum_{j=1}^{N} A^{-1}_{ij} h_j. $$ We can now make a change of variables $x_i \mapsto y_i$ to obtain $$ x_i=\sum_{j=1}^{N} K^{-1}_{ij}h_j+y_i. $$ Re-writing $\mathcal{F}(h)$ we obtain $$ \mathcal{F}(h)=\exp\bigg(\sum_{i,j=1}^{N} \frac{1}{2} h_i A^{-1}_{ij} h_j\bigg)\int_{-\infty}^\infty d^Ny \exp\bigg(-\sum_{i,j=1}^{N} \frac{1}{2} y_i A_{ij}y_j \bigg). $$ This integral is now a simple gaussian which we diagonalize by an orthogonal transformation $A=O\lambda_i\delta_{ij}O^\top$ and a linear change of variables $x=Oy$. The Jacobian of the transformation is unity since a rotation leaves the volume invariant. We write the general result as \begin{equation} \mathcal{F}(h)=\big({2\pi}\big)^{N/2} (\det A_{ij})^{-1/2} \exp\left(\sum_{i,j=1}^{N} \frac{1}{2} h_i A^{-1}_{ij} h_j\right). \end{equation} Having calculated this, I now need to calculate the expected value of the higher order moments which is what my question is.

Note in 1 dimension, the expected value I am trying to calculate is similar to $$ \int_{-\infty}^\infty x^{n} e^{-x^2/2-\alpha x}dx,\quad \Re(n)>-1, \alpha\in \mathbb{R}. $$ I have found lower order expected values given by $$ \big<x_i\big>=A^{-1}_{ij}h_j ,\quad \big<x_i x_j\big>=A^{-1}_{ik} h_k A^{-1}_{jl}h_l+A^{-1}_{ij}, $$ but am trying to generalize to higher orders.


You can look up a result of Isserlis, which is proved by him in "On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables". Biometrika 12: 134–139.

Essentially, if $X_1,\ldots,X_n$ follow a zero mean multivariate gaussian distribution, $E[X_1\ldots X_n] = \sum_{\text{Partitions of $\{X_1,\ldots,X_n\}$ into pairs}} \prod_{(X_i,X_j)\text{ is a pair}}E[X_i X_j]$ when $n$ is even (and is 0 when $n$ is odd).

  • $\begingroup$ My case is not the case of zero expected mean. I am familiar with this and Wick's theorem, I am still looking for a general proof of this method though. Also, odd moments of the random variables DO NOT vanish here because $h_i\neq 0$, thus the expected mean is shifted by h. If $h_i=0$, than yes only the even moments have non-vanishing contributions. Specifically in my example, $$ \big<x_i\big>=A^{-1}_{ij}h_j, $$ but I am trying to generalize this... $\endgroup$ – Jeff Faraci Apr 17 '14 at 22:08
  • $\begingroup$ I doubt the case for non-zero mean has a clean solution like this, but you can find the proof in the paper cited. And you can use the solution for the centered case to calculate the non-centered case in principle by linearity of expectation and a ton of algebra. $\endgroup$ – Batman Apr 17 '14 at 22:17
  • $\begingroup$ it has a very nice clean answer actually, my problem is deriving it! $\endgroup$ – Jeff Faraci Apr 17 '14 at 22:18

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.