Consider the complex domain ℂ. If U and V are 2 unitary random matrices and A is a deterministic matrix.

What is the distribution of $u^HAv$ ( or $||u^HAv||^2$)

where : u is a column vector of U. v is a column vector of V.

Simulations give that $||u^HAv||^2$ is exponential; is it true ? if yes, how to prove it ?

Have a nice day!


a simple and exact answer follows if $A$ is a unitary $n\times n$ matrix; then $\tilde{v}=Av$ is a unit vector and since $u$ is a randomly oriented unit vector the distribution of the scalar product $u^HAv=(u,\tilde{v})$ does not depend on the orientation of $\tilde{v}$, so we may take it in the $x_1$ direction, hence

$$P(u^H Av)=P(u_1).$$

the distribution of a single element $u_1$ of the random unit vector $u$ follows upon integration over the other $n-1$ elements,

$$P(u_1)= \int du_2\int du_3\cdots\int du_n\;\delta\left(1-\sum_{i=1}^{n}|u_i|^2\right)={\rm constant}\times(1-|u_1|^2)^{n-2},$$

hence $\xi=|u^H Av|^2$ has the distribution


this is exact for any $n\geq 2$; for $n\gg 1$ the distribution becomes exponential to leading order in $1/n$,

$$P(\xi)\approx ne^{-n\xi}.$$

for non unitary $A$, the answer will depend on its singular values $a_1,a_2,\ldots a_n$; I doubt that you can find a simple closed-form expression for any $n$, but for $n\gg 1$ again an exponential distribution will appear.

To see this, define the unit vector $v'=|Av|^{-1}\,Av$ and first consider the distribution of $\xi'=|u^H v'|^2$. Here the same argument as above applies, so for $n\gg 1$ we have $P(\xi')\propto\exp(-n\xi')$. The quantity $\xi$ we want is related to $\xi'$ by $\xi=|Av|^2\xi'$. The factor $|Av|^2$ has a normal distribution in the large-$n$ limit, with a mean $$E[|Av|^2]=\frac{1}{n}{\rm Tr}\,AA^H$$ and a variance that is an order $1/n$ smaller; to leading order in $1/n$ we may therefore just equate

$$\xi\approx E[Av|^2]\xi'\Rightarrow P(\xi)\approx\frac{n^2}{{\rm Tr}\,AA^H}\exp\left(-\frac{n^2}{{\rm Tr}\,AA^H}\xi\right).$$

  • $\begingroup$ Could you please give me some references for the cases you have explained. thank you very much $\endgroup$ – tam Jun 1 '14 at 15:28
  • $\begingroup$ hmm, I don't really have references, if you tell me for which step you would like to have further explanations, I can try to fill you in. $\endgroup$ – Carlo Beenakker Jun 1 '14 at 15:42
  • $\begingroup$ in the mean time, you might check your numerics that for unitary $A$ you find $P(\xi)\propto(1-\xi)^{n-2}$, exactly for small $n$ (the exponential distribution only follows in the large-$n$ limit) $\endgroup$ – Carlo Beenakker Jun 1 '14 at 15:44
  • $\begingroup$ "for non unitary A and n>>1 I would think the same exponential distribution will appear.." if you don't have the answer, I will be grateful if you give me the explanation for the case where A is unitary (and n>>1). $\endgroup$ – tam Jun 1 '14 at 15:52
  • $\begingroup$ added details at various steps of the derivation, and generalized it to the general case of nonunitary $A$. $\endgroup$ – Carlo Beenakker Jun 1 '14 at 18:12

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.