Let $X \in \mathbb{R}^d \sim N(0, \Sigma)$ be a multivariate Gaussian random vector. For any $\beta\in \mathbb{R}^d$, consider the regression model $$ Y = X^\top \beta + \epsilon, $$ where $\epsilon \sim N(0,\sigma^2)$ is the random noise. How to compute expectation $ \mathbb{E} [ Y^2 XX^\top ] $?


The vector $(X,Y)$ is Gaussian, because any sum is normal: $$\sum a_iX_i + bY= \sum c_i X_i + b\epsilon\sim\mathcal N(\mu,\sigma^2)$$ for some mean and variance. (assuming $\epsilon$ is independent of all other variables.)

If $(X,Y)$ is a zero-mean multivariate Gaussian, then we can use Isserlis' theorem: For some $i,j$ let us calculate the scalar $$ E[Y^2X_i X_j]=E[Y\cdot Y\cdot X_i \cdot X_j]$$ $$E[Y\cdot Y\cdot X_i \cdot X_j]=E[YY]E[X_iX_j]+E[YX_i]E[YX_j]+E[YX_i]E[YX_j]$$

Everything here should be known: $E[X_iX_j]=(\Sigma)_{i,j}$, $$E[YX_i]=E[X_iX^T\beta ] + E[\epsilon X_i]=\sum E[X_iX_k]\beta_k+0,$$ and $$E[Y^2]=E[(X^T\beta+\epsilon)^2]=E[\epsilon^2]+0+\beta^T E[X X^T]\beta.$$

| cite | improve this answer | |

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.