2
$\begingroup$

Given a $d$-dimensional standard Gaussian variable $\varepsilon$, a variable $X$ independent of $\varepsilon$ and a continuous function $F : \mathbb{R}^d \to \mathbb{R}^d$ , I am trying to compute $$\mathbb{E}[ \mathrm{exp} \, (\langle F(X), \epsilon \rangle + \tfrac{1}{4} \Vert \varepsilon \Vert^2) | X]$$

which "amounts" to computing

$$\mathbb{E}[ \mathrm{exp} \,(\Vert F(X) + \tfrac{1}{2} \varepsilon \Vert^2) | X]$$ as the first term if we develop the square is $X$-measurable. But I don't really know where to go further, as the exponetial prevents me from splitting the sum of the scalar product.

Edit: After discussion, it is related to the MGF of a non central chi-square distribution: for a fixed $x$, $N = \Vert 2 F(x) + \varepsilon \Vert^2$ follows a chi-square distribution with $d$ degrees of freedom and $\lambda = 4 \Vert F(x) \Vert^2 $. It remains to apply the formula for the MGF of a non central chi-square, with $t = 1/4$: $$\mathbb{E}[e^{t N}] = \frac{e^{\lambda /2}}{2^{d/2}}$$

$\endgroup$
3
  • $\begingroup$ How far are you trying to get? $\endgroup$
    – user515599
    Mar 12, 2019 at 8:35
  • $\begingroup$ @orange is it hopeless to try to get a closed form formula in X ? $\endgroup$ Mar 12, 2019 at 8:36
  • $\begingroup$ It turns into a product not asum, my bad $\endgroup$
    – user515599
    Mar 12, 2019 at 9:53

1 Answer 1

2
$\begingroup$

First we can use the formula $$ \mathbb E\left[h\left(X,Y\right)\mid X\right]=g(X) $$ where $X$ is a vector independent of $Y$ and $g(x)=\mathbb E\left[h\left(x,Y\right)\right]$. The wanted conditional expectation is thus $g(X)$, where $$ g(x)=\mathbb E\left[\exp\left(\sum_{k=1}^d\left(F_k(x)\varepsilon_k+\frac 14\varepsilon_k^2\right)\right)\right]. $$ Using independence of the sequence $\left(\varepsilon_k\right)$ gives $$ g(x)=\prod_{k=1}^d\mathbb E\left[\exp\left( F_k(x)N+\frac 14N^2 \right)\right], $$ where $N$ has a standard normal distribution. Rewriting the later expectation as an integral and reducing the square leads to a nice formula for $g$.

$\endgroup$
1
  • $\begingroup$ Thaks, I had forgotten about the substitution formula to fix $x$. I have also been given a proof involving chi-square distributions, which I've included in an edit. $\endgroup$ Mar 12, 2019 at 10:43

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.