Given $Y = [Y_1, Y_2, \ldots, Y_n]^T \sim N(0, \Sigma)$. That is $f_{Y}(y_1, y_2, \ldots, y_n)$ is a multivariate gaussian with mean $0$ and covariance matrix $\Sigma$.

I'm asked to compute the conditional correlation of $y_i, y_j$ given $y_k$, that is $\rho(y_i, y_j | y_k)$. I know that in order to do this, I need to compute $$E[y_i \cdot y_j|y_k] = \int_{-\infty}^{\infty} y_i \,y_k \,\, f(y_i, y_j | y_k) \,d$$.

How can I compute $f(y_i, y_j | y_k)$? I know that this expression can be computed by $$f(y_i, y_j | y_k) = \frac{f(y_i, y_j)}{f(y_k)}$$

But I'm not sure how to find these expressions in closed form.

  • $\begingroup$ I would take a look at the Schur Complement, in particular the section labeled "Applications to probability theory and statistics." $\endgroup$ – Jonathan Lisic Nov 11 '16 at 1:00

The Schur Complement, is the goto for this kind of computation. Let $Y \sim \mathcal{N}(0,\Sigma)$, where $Y$ and $\Sigma$ are organized such that the last element and row/column belong to the variable you wish to condition on. Then, $Y_{-k}|Y_k$ is distributed, $\mathcal{N}\left(\mu_{-k}, \Sigma_{-k}\right)$, where

  • $Y_{-k}$ consists of the elements of $Y$ not including $k$;
  • $\Sigma = \left( \begin{array}{rr} A & B\\ B^{T} & C\\ \end{array} \right)$ in slightly R'ish parlance $A = \Sigma[-k,-k]$, $B=\Sigma[-k,k]$, and $C = \Sigma[k,k]$;

  • $\Sigma_{-k} = A - BC^{-1}B^T$;

  • and $\mu_{-k} = BC^{-1}Y$.

The covariance of $Y_i$ and $Y_j$ given $Y_k$ can then be retrieved from $\Sigma_{-k}$. A fast way to calculate this numerically is through the sweep function (Dempster 1969).

Dempster, A.P. (1969). Elements of continuous multivariate analysis. Reading, MA: Addison- Wesley.

| 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.