Let $g = (g_1, ..., g_n)$ be a random vector distributed uniformly on the sphere $\{ x \in \mathbb{R}^n : \| x \|_2 = 1 \}$. Let $A, B$ be two symmetric $n \times n$ matrices.

I am interested in a simple formula for: $$\mathbb{E}[ g^T A g g^T B g ]$$

In particular, I know that if $g$ is a standard normal $N(0, I)$, then $$ \mathbb{E}[ g^T A g g^T B g ] = 2 \mathrm{Tr}(AB) + \mathrm{Tr}(A)\mathrm{Tr}(B) \:. $$

Does something similar hold in the uniform on a sphere case?


Here is another approach to reduce the problem to the Gaussian case. Let $g \sim N(0,I)$ and let $\theta = g / \|g\|$ and $r = \|g\|$ so that $g = r\theta$. It is not hard to verify that $\theta$ is uniformly distributed on the sphere and is independent of $r$. Then, we have, by independence, \begin{align*} \mathbb E (g^T A g g^T B g) = \mathbb E (\theta^T A \theta \theta^T B \theta) \cdot \mathbb E (r^4) \end{align*} and \begin{align*} \mathbb E (r^4) &= \mathbb E \Big(\sum_i g_i^2\Big)^2 \\ &=\sum_i \mathbb E g_i^4 + \sum_{i \neq j} \mathbb E g_i^2 \mathbb Eg_j^2 \\ &= 3 n + n(n-1) = n(n+2). \end{align*}

  • $\begingroup$ Great! This is much simpler-- thanks $\endgroup$ – steve Dec 22 '18 at 5:40
  • $\begingroup$ @steve, no problem. $\endgroup$ – passerby51 Dec 22 '18 at 5:44

It turns out the answer is:

$$ \mathbb{E}[g^T A g g^T B g] = \frac{1}{n(n+2)} (2 \mathrm{Tr}(AB) + \mathrm{Tr}(A) \mathrm{Tr}(B)) \:. $$

This comes from the Theorem in Section 3 of D. P. Wiens, "On Moments of Quadratic Forms in Non-Spherically Distributed Variables" (https://pdfs.semanticscholar.org/f9e3/9424ff32aec189b441dd189b6f819764de6b.pdf). It is straightforward to check that the conditions (1.1) hold for the uniform distribution over the sphere.

To apply the result, you need to compute $\mathbb{E}[g_1^2 g_2^2]$. It turns out this is equal to $\frac{1}{n(n+2)}$. Here is a simple way to compute this by symmetry. Observe that:

$$ g_1^2 g_2^2 = (1 - g_2^2 - ... - g_n^2) g_2^2 = g_2^2 - g_2^4 - g_3^2 g_2^2 - ... - g_n^2 g_2^2 \:. $$ Let $\mu = \mathbb{E}[g_1^2 g_2^2]$. By symmetry, $\mu = \mathbb{E}[g_i^2 g_j^2]$ for any $i \neq j$. Hence taking expectations, $$ \mu = \mathbb{E}[g_2^2] - \mathbb{E}[g_2^4] - (n-2) \mu \:. $$ Rearrange to obtain: $$ \mu = \frac{1}{n-1}\left( \mathbb{E}[g_2^2] - \mathbb{E}[g_2^4] \right) \:. $$ We can compute $\mathbb{E}[g_2^2] = 1/n$ by a similar symmetry argument. Furthermore, since $g_2^2 \stackrel{d}{=} \mathrm{Beta}(1/2, (n-1)/2)$, we can compute $\mathbb{E}[g_2^4]$ by looking up the formula of the second moment of a Beta distribution, which gives us $\mathbb{E}[g_2^4] = \frac{3}{n(n+2)}$. Hence, $$ \mu = \frac{1}{n-1}\left( \frac{1}{n} - \frac{3}{n(n+2)} \right) = \frac{1}{n(n+2)} \:. $$


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.