0
$\begingroup$

Given a multivariate normal random $n\times 1$ vector $X \sim N(\mu,\Sigma)$, what is the expectation $$\mathbb{E}[exp(X^TAX+b^TX)]$$ where $A$ is a $n\times n$ matrix and $b$ is a n-dimensional vector? I know that $\mathbb{E}[X^TAX] = \mathbb{E}[X^T]A\mathbb{E}[X]+tr(A\Sigma)$, but how should I solve this problem?

$\endgroup$
1
$\begingroup$

$$ X\sim N(\mu,\Sigma)\quad\Rightarrow\quad \text{pdf}=f_X(X)=\frac{1}{\sqrt{(2\pi)^n|\Sigma|}}\exp(-\frac{1}{2}(X-\mu)^T\Sigma^{-1}(X-\mu)) $$ Therefore: $$ \begin{split} E[\exp(X^TAX+b^TX)]&=\int_{\mathbb R^n}\frac{1}{\sqrt{(2\pi)^n|\Sigma|}}\exp(-\frac{1}{2}(X-\mu)^T\Sigma^{-1}(X-\mu)+X^TAX+b^TX)dX\\ &=\int_{\mathbb R^n}\frac{1}{\sqrt{(2\pi)^n|\Sigma|}}\exp(-\frac{1}{2}(X-\Delta)^T\Sigma'^{-1}(X-\Delta)+\Lambda)dX \end{split} $$ where $$ \Sigma'=(\Sigma^{-1}-2A)^{-1}, \Delta=\Sigma'(b+\Sigma^{-1}\mu), \Lambda=\frac{1}{2}(\Delta^T\Sigma'^{-1}\Delta-\mu^T\Sigma^{-1}\mu). $$ Therefore: $$ \begin{split} E[\exp(X^TAX+b^TX)]&=\sqrt{\frac{|\Sigma'|}{|\Sigma|}}\exp(\Lambda)\int_{\mathbb R^n}\frac{1}{\sqrt{(2\pi)^n|\Sigma'|}}\exp(-\frac{1}{2}(X-\Delta)^T\Sigma'^{-1}(X-\Delta))dX\\ &=\sqrt{\frac{|\Sigma'|}{|\Sigma|}}\exp(\Lambda) \end{split} $$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks for replying. I have two questions to ask you if you don't mind: 1. What is the explanation for the first line of equation you have written after "Therefore:"? I found a mention of a similar problem at this link goo.gl/fNqvoA and I can see that your answer can be simplified to the form given in this link, but I am having problem understanding the first step of your solution itself. 2. Can you suggest me a book to study this topic in detail so that I can solve the problems on my own? $\endgroup$ – AbbyJ Aug 24 '14 at 14:59
  • $\begingroup$ @abbyj The link you gave refer to a completely different problem. Do you know what a closed form means? $\endgroup$ – Troy Woo Aug 24 '14 at 15:10
  • $\begingroup$ yes and I guess you mean that the closed form expression here shouldn't contain any integrals over infinite space, right? $\endgroup$ – AbbyJ Aug 24 '14 at 15:16
  • $\begingroup$ @abbyj multivariate normal distribution is much like the single variable one if you use diagonalization. What I did is basically completion of squares, which is high school math. The rest is utilizing the fact that probability density function integrates to $1$ over the entire domain $\mathbb R^n$. There are plenty of book on this subject, I didn't read a single one so I cant recommend you any reference. $\endgroup$ – Troy Woo Aug 24 '14 at 15:23
  • $\begingroup$ Got your explanation. Thanks for helping. $\endgroup$ – AbbyJ Aug 24 '14 at 15:37

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.