2
$\begingroup$

Consider arbitrary unit vectors $w,x,y,z \in \mathbb{C}^d$. Is there an explicit formula for what this average is? $$ \int \mathrm{Tr}( \psi \psi^* \, \, w x^* \,\, \psi \psi^* \,\, y z^*) d\psi $$ where the average is over a Haar-random unit vector $\psi \in \mathbb{C}^d$. Here, $\psi \psi^*$ is the rank-one matrix formed by the outer product of $\psi$ with itself. I'm looking for a formula that relates this average to inner products between $w, x, y, ,z$, for example.

A related question I'm wondering about is what the expected value of the following inner product is: suppose you pick a Haar-random vector $\psi$, followed by a Haar-random vector $\psi^\perp$ that's orthogonal to $\psi$, and then look at $\langle \psi, x \rangle \langle y, \psi^\perp \rangle$.

Thanks!

$\endgroup$

1 Answer 1

5
$\begingroup$

$\psi_\alpha$, $\alpha=1,2,\ldots d$, is a column vector of a $d\times d$ unitary matrix $U$; averaging over the Haar measure gives $$\int d\psi\, \psi_{\alpha} \psi^\ast_\beta \psi_{\alpha'}\psi^\ast_{\beta'}=\frac{1}{d+d^2}\left(\delta_{\alpha\beta}\delta_{\alpha'\beta'}+\delta_{\alpha\beta'}\delta_{\alpha'\beta}\right)$$

so your integral equals

$$\int \mathrm{Tr}( \psi \psi^* \, \, w x^* \,\, \psi \psi^* \,\, y z^*) \,d\psi= \langle z|\psi\rangle\langle \psi|w\rangle\langle x|\psi\rangle\langle \psi|y\rangle$$ $$\qquad\qquad = \frac{1}{d+d^2}\bigl(\langle z|w\rangle\langle x|y\rangle+ \langle z|y\rangle\langle x|w\rangle\bigr)$$

in answer to your second question: the average of $\langle\psi|x\rangle\langle y|\psi'\rangle$ with $\psi,\psi'$ two different column vectors of $U$ is zero (obviously, because $\psi,\psi'$ and $-\psi,\psi'$ are equally probable)

$\endgroup$
3
  • $\begingroup$ Great, thanks! Where did you get the first formula, though? Also, as a followup: would you know what kind of measure of concentration is satisfied by these inner products? That is, what is the probability that $\mathrm{Tr}( \psi \psi^* \, \, w x^* \,\, \psi \psi^* \,\, y z^*)$ deviates from the average by more than some amount $\epsilon$? $\endgroup$
    – Henry Yuen
    Mar 30, 2016 at 14:22
  • $\begingroup$ for the first formula, see equation B5 of arxiv.org/abs/cond-mat/9612179 ; measure of concentration: for large $d$ the elements of $\psi$ are independent Gaussians (zero mean, variance $1/d$) $\endgroup$ Mar 30, 2016 at 14:28
  • $\begingroup$ Do you have a more precise estimate of the error (as a function of $d$) when you try to treat the elements of $\psi$ as independent Gaussians? $\endgroup$
    – Henry Yuen
    Mar 30, 2016 at 14:37

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.