$\DeclareMathOperator\tr{\mathrm{tr}}$Let $x$ be a standard normal $p$-variate random variable, which is independent of a symmetric positive definite random matrix $Y$. I would like to compute the mean and variance of the quadratic form $x^TYx$.

If $Y$ were not a random variable, then we could argue that the fact that quadratic forms are diagonalizable means that without the loss of generality we can assume that $Y$ is a diagonal matrix. And an entry-by-entry computation shows that the mean will be $\tr(Y)$ with variance $2 \tr(Y)$.

But if $Y$ is a random variable, I'm not sure it's so easy since though the mean is still $\tr(\mathbb{E}[Y])$, the variance of each entry now depends on $Y$ in a more complicated way, and since diagonalizing changes the diagonal entries in non-trivial ways, is it no longer possible to relate the variance to the trace of $Y$ or a related expression?


1 Answer 1


$\DeclareMathOperator\E{\mathbb E} \DeclareMathOperator\tr{\mathrm{tr}} \DeclareMathOperator\Cov{\mathrm{Cov}} \DeclareMathOperator\Var{\mathrm{Var}}$ Using trace rotation and the commutivity of linear operators: \begin{align} \E[x^T Y x] &= \E[\tr(x^T Y x)] \\&= \E[\tr(Y x x^T)] \\&= \tr(\E[Y x x^T]) \\&= \tr(\E[Y] \E[x x^T]) \\&= \tr(\E[Y]) .\end{align}

This Wikipedia page or this math.SE answer tell us, slightly disagreeing with what you said, that for normal $x$ and symmetric $Y$ $$\Var[x^T Y x \mid Y] = 2 \tr(Y Y).$$ Then via the law of total variance, for random $Y$ we have \begin{align} \Var(x^T Y x) &= \E_Y \Var_x(x^T Y x \mid Y) + \Var_Y( \E_x[x^T Y x \mid Y] ) \\&= \E[ 2 \tr( Y^2 ) ] + \Var( \tr(Y \E[x x^T])) \\&= 2 \tr(\E[ Y^2 ]) + \Var(\tr(Y)) .\end{align}


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.