Suppose $\vec x \sim N(\vec \mu, \Sigma)$ is multivariate normal. I want to see that $$\mathrm{Cov}(\vec x^TA\vec x,\vec x^TB\vec x) = 2 \mathrm{Tr}(A \Sigma B \Sigma) + 4 \vec \mu^TA \Sigma B \vec \mu$$ I have been searching the internet for a while, and found multiple sources confirming it, but no proof, for example here: $Var(Q)=2\ tr(A\Sigma A\Sigma)+4\mu^TA\Sigma A\mu$

Do you know how to show it?

  • $\begingroup$ Is this from a self-study question? "A routine question from a textbook, course, or test used for a class or self-study. This community's policy is to "provide helpful hints" for self-study questions." $\endgroup$
    – amoeba
    Sep 16 '17 at 16:45
  • $\begingroup$ I am writing an article, I know it is true, but I do not know why $\endgroup$
    – Mikkel Rev
    Sep 16 '17 at 16:50
  • $\begingroup$ If you provide hints, such that I prove it myself, that is fine $\endgroup$
    – Mikkel Rev
    Sep 16 '17 at 16:51

We have

\begin{align} \operatorname{Var}(x^T(A+B)x)&=\operatorname{Var}(x^TAx+x^TBx) \\&=\operatorname{Var}(x^TAx)+\operatorname{Var}(x^TBx)+2\operatorname{Cov}(x^TAx,x^TBx) \end{align}

So, $$\operatorname{Cov}(x^TAx,x^TBx)=\frac12\left[\operatorname{Var}(x^T(A+B)x)-\operatorname{Var}(x^TAx)-\operatorname{Var}(x^TBx)\right]\tag{1}$$

Note that for the calculation of variance and covariance when $x$ is multivariate normal, it is assumed that $A$ and $B$ are symmetric, which also makes $A+B$ symmetric.

For $y\sim N(0,I)$, it is shown here that $$\operatorname{Var}(y^TAy)=2\operatorname{tr}(A^2)\tag{2}$$

Now suppose $x\sim N(\mu,\Sigma)$ where $\Sigma$ is positive definite.

Then there exists a nonsingular matrix $C$ such that $\Sigma=CC^T$.

And $$x\sim N(\mu,\Sigma)\implies y=C^{-1}(X-\mu)\sim N(0,I)$$

So, $$x^TAx=(\mu+Cy)^TA(\mu+Cy)=\mu^TA\mu+2\mu^TACy+y^T(C^TAC)y$$

Therefore noting that $y$ and $y^T(C^TAC)y$ are uncorrelated,

\begin{align} \operatorname{Var}(x^TAx)&=4\operatorname{Var}(\mu^TACy)+\operatorname{Var}\left(y^T(C^TAC)y\right) \\&=4\mu^T(AC)(AC)^T\mu + 2\operatorname{tr}((C^TAC)(C^TAC))&\small\left[\text{ using }(2)\right] \\&=4\mu^T A\Sigma A\mu + 2\operatorname{tr}(C^TA\Sigma AC) \\&=4\mu^T A\Sigma A\mu + 2\operatorname{tr}(A\Sigma A\Sigma) &\small\left[\because \,\operatorname{tr}(AB)=\operatorname{tr}(BA)\right] \tag{3} \end{align}

Now it follows from $(1)$ and $(3)$ that

$$\operatorname{Cov}(x^TAx,x^TBx)=4\mu^TA\Sigma B\mu + 2\operatorname{tr}(A\Sigma B\Sigma)$$

For a direct calculation of the covariance and hence the variance, one may refer to Graybill's Matrices with Applications in Statistics, 2nd edition (1983).


Just trying to provide a sketch for the proof, the key here is to show the following equations: $$\begin{align}\operatorname{E}[x'Ax]&=\operatorname{tr}(A\Sigma)+\mu'A\mu, \label{1}\tag{1}\\\operatorname{E}[x'Axx'Bx]&=2\operatorname{tr}(A\Sigma B\Sigma)+4\mu'A\Sigma B\mu+(\operatorname{tr}(A\Sigma)+\mu'A\mu)(\operatorname{tr}(B\Sigma)+\mu'B\mu). \label{2}\tag{2}\end{align}$$ The desired result follows immediately since by definition $$\begin{align}\operatorname{Cov}(x'Ax, x'Bx)&=\operatorname{E}[(x'Ax-E[x'Ax])(x'Bx-\operatorname{E}[x'Bx])']\\&=\operatorname{E}[x'Axx'Bx]-\operatorname{E}[x'Ax]\operatorname{E}[x'Bx].\end{align}$$

The proof to equation ($\ref{1}$) is simple and can be found in many introductory text. Equation ($\ref{2}$) is the real deal here, but fortunately a proof can be found in Proofs Section 5 of the Matrix Reference Manual. Check out 5.18 and 5.19 for Isserlis' theorem, and finally 5.28 where they derived an expression for a much more general form: $$\operatorname{E}[(Ax-a)'(Bx-b)(Cx-c)'(Dx-d)].$$

  • $\begingroup$ I think (from memory) you can find this in Seber's linera regression analysis $\endgroup$ Sep 16 '17 at 19:57
  • $\begingroup$ It has some relevant stuff. I've got the book for 4 weeks, are you able to say where in the book it is? $\endgroup$
    – Mikkel Rev
    Sep 23 '17 at 14:29
  • $\begingroup$ I take you mean the book @kjetilbhalvorsen mentioned. $\endgroup$
    – Francis
    Sep 23 '17 at 14:36
  • $\begingroup$ Yes, that's right $\endgroup$
    – Mikkel Rev
    Sep 23 '17 at 14:54

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.