12
$\begingroup$

If one generates an $n\times n$ Haar random unitary $U$, then clearly $\Pr(U=I)=0$. However, for every $\epsilon>0$, the probability $$\Pr(\|U-I\|_{\rm op}<\varepsilon)$$ should be positive. How can this quantity be computed?

$\endgroup$
10
  • $\begingroup$ Exatly? Approximately? Upper/lower bounds? Analytically? Numerically? $\endgroup$ – Norbert Schuch Aug 14 '20 at 11:20
  • 3
    $\begingroup$ arxiv.org/abs/1506.07259 could contain some useful information or techniques (they compute said quantity, but for a different distance measure). $\endgroup$ – Norbert Schuch Aug 14 '20 at 11:27
  • $\begingroup$ Let's say we want an analytical lower-bound? $\endgroup$ – Calvin Liu Aug 15 '20 at 15:08
  • 1
    $\begingroup$ @CalvinLiu I'll leave this as a comment having not carefully worked out the details, but you can think of this probability as computing the ratio of the volume of an $\epsilon$-ball around $I$ to the volume of the unitary group (with respect to the operator norm), equivalently you can think about this as the 1/size of an $\epsilon$-net for $U(n)$. Very roughly, this is going to be ${\rm Pr}(\|U-I\|_\infty \leq \epsilon) \sim (n/\epsilon^2)^{-n^2}$. $\endgroup$ – 4xion Aug 25 '20 at 15:36
  • 1
    $\begingroup$ If you need a rigorous lower bound you should be able to compute this more precisely (look up refs related to volumes of balls in the unitary group and $\epsilon$-nets) $\endgroup$ – 4xion Aug 25 '20 at 15:38
2
$\begingroup$

$U-I$ is a normal matrix so $||U-I||_{op}$ is its eigenvalue with the largest magnitude. The eigenvalue equation for this matrix is $$(U-I)|\psi\rangle=\lambda|\psi\rangle,$$ so $$|\lambda|^2=(\cos\phi-1)^2+\sin^2\phi=4\sin^2\frac{\phi}{2}\Rightarrow |\lambda|=2\left|\sin\frac{\phi}{2}\right|,$$ where $e^{i\phi}$ is some eigenvalue of $U$.

Now, the distribution of eigenvalues for a random unitary matrix is known: there are $n$ eigenvalues $e^{i\phi_n}$ each distributed over $-\pi\leq \phi_n\leq \pi$ with probability density $$p(\phi_1,\cdots,\phi_n)=\frac{1}{n!(2\pi)^n}\prod_{1\leq j< k\leq n}\left|e^{i\phi_j}-e^{i\phi_k}\right|^2=\frac{2^n}{n!\pi^n}\prod_{1\leq j< k\leq n}\sin^2\frac{\phi_j-\phi_k}{2}.$$ So now we can express the desired quantity as $$\mathrm{Pr}(||U-I||_{op}\leq\epsilon)=\frac{2^n}{n!\pi^n}\int_{-\pi}^{\pi} d\phi_1\cdots d\phi_n \prod_{1\leq j< k\leq n}\sin^2\frac{\phi_j-\phi_k}{2}\left(2\max\{\left|\sin \frac{\phi_1}{2}\right|,\cdots,\left|\sin \frac{\phi_n}{2}\right|\}\leq \epsilon\right).$$

For the maximum to be less than or equal to $\epsilon/2$, all of the quantities must be less than or equal to $\epsilon/2$. This happens when $|\phi_i|\leq 2\arcsin{\frac{\epsilon}{2}}$, so the result is $$\mathrm{Pr}(||U-I||_{op}\leq\epsilon)=\frac{2^n}{n!\pi^n}\int_{-2\arcsin\frac{\epsilon}{2}}^{2\arcsin\frac{\epsilon}{2}} d\phi_1\cdots d\phi_n \prod_{1\leq j< k\leq n}\sin^2\frac{\phi_j-\phi_k}{2} .$$

Right now I can't evaluate that for arbitrary $n$... the integrals can all be done analytically but they are tedious. For $n=2$ the answer is $$-\frac{4\epsilon^2}{\pi^2}+\frac{\epsilon^4}{\pi^2}+\frac{16}{\pi^2}\arcsin^2\frac{\epsilon}{2}\approx \frac{4\epsilon ^4}{3\pi^2}+O\left(\epsilon ^{6}\right),$$ for $n=3$ the answer is $$\frac{1}{8\pi^3} \left(\left(\epsilon^6-8 \epsilon^4+28 \epsilon^2-48\right) \epsilon^2 \sin ^{-1}\left(\frac{\epsilon}{2}\right)+\left(\epsilon^4-6 \epsilon^2+8\right) \sqrt{4-\epsilon^2} \epsilon^3+64 \sin ^{-1}\left(\frac{\epsilon}{2}\right)^3\right)\approx \frac{4 \epsilon^9}{135 \pi ^3}+O\left(\epsilon^{11}\right),$$ and for $n=$ the answer is (I haven't simplified the trig terms here but they're doable) $$\frac{96 \sin ^3\left(2 \sin ^{-1}\left(\frac{\epsilon}{2}\right)\right) \sin ^{-1}\left(\frac{\epsilon}{2}\right) \left(-5 \epsilon^2+2 \cos \left(6 \sin ^{-1}\left(\frac{\epsilon}{2}\right)\right)+10\right)+4 \epsilon^2 \left(\epsilon^2-4\right) \left(2 \epsilon^8-16 \epsilon^6+53 \epsilon^4-84 \epsilon^2+108\right) \sin ^{-1}\left(\frac{\epsilon}{2}\right)^2+1152 \sin ^{-1}\left(\frac{\epsilon}{2}\right)^4+\sin ^4\left(2 \sin ^{-1}\left(\frac{\epsilon}{2}\right)\right) \left(-172 \cos \left(4 \sin ^{-1}\left(\frac{\epsilon}{2}\right)\right)+\cos \left(8 \sin ^{-1}\left(\frac{\epsilon}{2}\right)\right)-45\right)}{1152 \pi ^4}\approx \frac{\epsilon^{16}}{23625 \pi ^4}+O(\epsilon^18).$$

$\endgroup$
2
  • 1
    $\begingroup$ Why do you think that for a Haar random unitary the distribution of its eigenvalues is a simple product of uniform distributions? The paper arxiv.org/abs/1506.07259 has the exact formula (eq. 3) for that distribution. Those angles are not independent. Also check en.wikipedia.org/wiki/Circular_ensemble $\endgroup$ – Danylo Y Jun 3 at 22:56
  • $\begingroup$ Oh right of course.. also I forgot to take square roots... will see what I can salvage, otherwise will remove $\endgroup$ – Quantum Mechanic Jun 4 at 2:21

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.