Consider first general matrices in $\mathbb{C}^{n \times m}$. Using norm duality, the nuclear norm (sum of singular values) can be expressed as

$$\|A\|_* = \max \{ | \langle A, B \rangle | : \|B\|_2 \leq 1 \}$$

where $\|\|_2$ denotes the spectral norm (largest singular value) and $\langle A,B \rangle = \mathrm{Tr} AB^*$.

How can I show that, if $A$ is self-adjoint, it is sufficient to consider self-adjoint $B$ in the above?

  • $\begingroup$ Do you mean $\langle A, B \rangle = \text{Tr}(AB^*)$? $\endgroup$
    – angryavian
    Sep 3 '19 at 16:01
  • $\begingroup$ @angryavian Yes of course! Corrected. $\endgroup$
    – YLee
    Sep 3 '19 at 16:21

When $A$ is self-adjoint, it is diagonalizable, and the nuclear norm becomes the sum of the absolute values of the eigenvalues, say, $|\lambda_1| + \cdots + |\lambda_n|$ where $A = UDU^*$ and the diagonal entries of $D$ are $\lambda_1, \ldots, \lambda_n$.

Let $\tilde{D}$ be diagonal with diagonal entries $\lambda_1 / |\lambda_1|, \ldots, \lambda_n / |\lambda_n|$ and let $B = U\tilde{D} U^*$. You can check that $\|A\|_* = |\langle A, B \rangle|$ and that $\|B\|_2 \le 1$.

Thus, if $A$ is self-adjoint, there exists a self-adjoint $B$ with $\|B\|_2 \le 1$ such that $\|A\|_* = |\langle A, B \rangle|$, so the maximization can be restricted to self-adjoint $B$.


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.