Let $x\sim \mathcal{N(\mu_x,\Sigma_x \succ 0)}$ and $y\sim \mathcal{N(\mu_y,\Sigma_y \succ 0)}$ be multivariate Gaussian distributions, with $\mu_x \in \mathbb{R}^n$ and $\mu_y \in \mathbb{R}^m$ being the expected value vectors of $x$ and $y$ respectively. Let $M\in \mathbb{R}^{n \times m}$ be any matrix. Suppose that $x$ and $y$ are independent, that is $$\begin{bmatrix}x\\y\end{bmatrix}\sim \mathcal{N}\left(\begin{bmatrix}\mu_x\\\mu_y\end{bmatrix},\begin{bmatrix}\Sigma_x&0_{n \times m}\\0_{m \times n}&\Sigma_y \end{bmatrix}\right)\,.$$

What is the expression of $\mathbb{E}(x^T M y)$ in terms of $\mu_x,\mu_y,\Sigma_x,\Sigma_y$?

My guess would be $\mathbb{E}(x^T M y)=\mu_x^T M \mu_y$, but I'm confused how to prove this.


$x$ and $y$ are independent so $x$ and $My$ are independent. Therefore, $$ \mathbb E[x^TMy]=\mathbb E[x^T]\mathbb E[My]=\mathbb E[x]^TM\mathbb E[y]=\mu_x^TM\mu_y. $$

$x$ and $y$ do not need to be Gaussian.


Consider the general case

$$\begin{bmatrix}x\\y\end{bmatrix}\sim \mathcal{N}\left(\begin{bmatrix}\mu_x\\\mu_y\end{bmatrix},\begin{bmatrix}\Sigma_{xx}&\Sigma_{xy}\\\Sigma_{yx}&\Sigma_{yy} \end{bmatrix}\right)\,.$$

Recall that the Cross-covariance_matrix between $x$ and $y$ is

$$\Sigma_{xy} = \Sigma_{yx}^T = \text{Cov}(x,y) = E[(x-\mu_x)(y-\mu_y)^T] = E[xy^T] - \mu_x \mu_y^T $$

Consequently $E[yx^T] = \Sigma_{yx} + \mu_y\mu_x^T$. Using the linearity of the expectated value and the trace trick we obtain

\begin{align} \def\tr{\operatorname{tr}} E[xMy^T] &= E[\tr(x^T M y)]\\ &=E[\tr(Myx^T)]\\ &= \tr(E[Myx^T])\\ &= \tr(ME[yx^T]) \\ &= \tr(M(\Sigma_{yx}+\mu_y\mu_x^T)) \\ &= \mu_x^T M\mu_y + \tr(M\Sigma_{yx}) \end{align}

In the special case when $\Sigma_{yx}=0$, the formula simplifies to $E[xMy^T]=\mu_x^T M\mu_y$


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.