2
$\begingroup$

How can one maximize the following function:

$ f(V) = || V \otimes V - U_1 \otimes U_2 ||$

where $U_1, U_2 \in SU(n)$ are given and we seek to maximize over $V \in SU(n)$.

Both the maximum value of $f$ and the $V$ which maximizes it are interesting.

$\endgroup$

1 Answer 1

6
$\begingroup$

$\newcommand{\diag}{\operatorname{diag}}% $The maximum distance achieved is 2, except possibly when $n=2$. By left multiplying by the unitary $U_1^{-1} \otimes U_1^{-1}$, we are reduced to $U_1 = I$ and $U_2 = U$, a unitary.

Rewrite $V\otimes V - I \otimes U = (V \otimes VU^{-1} - I)(I \otimes U)$. Since the latter factor is unitary, it suffices to find $V$ such that $-1$ is an eigenvalue of $V \otimes VU^{-1}$ [this will show that the norm is at least 2, but of course it cannot exceed 2 in any case], that is, there is an eigenvalue $a$ of $V$ and an eigenvalue $b$ of $VU^{-1}$ such that $ab = -1$.

Let $v$ be an eigenvector for $U$, with corresponding eigenvalue $\lambda$ (obviously of modulus one), and let $w$ be orthogonal to $v$. By Gram-Schmidt, we can find $V$ in SU($n$) such that $Vv = \lambda v$ and $Vw = -w$ (here is where we use $n \geq 3$; for $n = 2$, all we can get is a unitary, not necessarily of determinant one). Then $v$ is an eigenvector for $VU^{-1}$, with eigenvalue $1$, and $-1$ is an eigenvalue of $V$, so $V \otimes VU^{-1}$ has $-1$ as an eigenvalue, and we are done.

For the case that $n=2$, I imagine there is a brute force argument; we can diagonalize $U$, and try to find the corresponding $V$ so that the set of products of the eigenvalues includes $-1$. If the result fails, it shouldn't be too difficult to optimize.

Edit: Actually, the $n=2$ case is easier than I thought. Reduce to $U = \diag(a,1/a)$, find $d$ in the unit circle such that $d^2 = -a$, and set $V = \diag (d, 1/d)$. Then the eigenvalues of $VU^{-1}$ are $d/a$ and $a/d$, so among the products are $d^2/a = -1$.

So in all cases, we can find optimal $V$ that commutes with $U = U_1^{-1}U_2$.

$\endgroup$

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.