Let $X_1,....,X_n$ be some positive random variables.

Is $A_{ij} = \frac{E[X_i*X_j]}{E[X_i]E[X_j]}$ a positive (semi) definite matrix?

(note that $B_{ij} = Cov(X_i,X_j) = E[X_i*X_j] - E[X_i]E[X_j]$ is PSD because it is a covariance matrix.)

  • $\begingroup$ That's an unusual formula; bad things happen when one of the $E[X_i]$ is 0. $\endgroup$ – user856 Oct 14 '10 at 22:38
  • $\begingroup$ thanks for pointing out. $X_i$ are strictly positive, and their expectation is never 0. does that help? $\endgroup$ – normvector Oct 14 '10 at 23:26


Suppose we choose arbitrary $\lambda_1,\ldots,\lambda_n$ and form a new random variable $$Z={{\lambda_1} \over {E[X_1]}}X_1+\cdots+{{\lambda_n} \over {E[X_n]}}X_n$$ Then $$E[Z^2]=\sum_{i=1}^n\sum_{j=1}^n A_{ij} \lambda_i \lambda_j$$ But being the expectation of a non-negative variable, $E[Z^2]$ is non-negative for any choice of $\lambda_1,\ldots,\lambda_n$. Which proves that your matrix $A$ is non-negative semi-definite.

| cite | improve this answer | |
  • $\begingroup$ thanks. that is correct. $\endgroup$ – normvector Oct 15 '10 at 0:34

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.