Let $M$ be a square matrix, $\|\cdot\|_*$ be the nuclear (trace) norm, and $\|\cdot\|_F$ be the Frobenius norm. The following inequality holds between the norms:

$$\|M\|^2_* \leq \text{rank}(M) \|M\|^2_F.$$

This is pretty easy to show by using the definitions of the norms in terms of the singular values $\sigma_i$, since $\|M\|_* = \sum_i \sigma_i$ and $\|M\|_F = \sqrt{\sum_i \sigma_i^2}$ and the result follows by Cauchy-Schwarz.

However, out of curiosity I have been trying to prove this using the definitions of the norms as $\|M\|_* = \text{trace} (\sqrt{M^* M})$ and $\|M\|_F = \sqrt{\text{trace}(M^* M)}$. Can the above inequality be shown using these definitions and without invoking the singular values explicitly?


1 Answer 1


For convenience, let $|M| = \sqrt{M^*M}$.

It suffices to note that $$ \langle A,B \rangle = \operatorname{Tr}(AB^*) = \operatorname{Tr}(B^*A) $$ is an inner product on the space of $m \times n$ matrices. From there, the properties of an inner product are enough to prove the Cauchy-Schwarz inequality: $$ |\operatorname{Tr}(AB^*)| = |\langle A,B \rangle| \leq \|A\| \cdot \|B\| \ = \sqrt{\operatorname{Tr}(A^*A)\operatorname{Tr}(B^*B)} = \sqrt{\operatorname{Tr}(|A|^2) \operatorname{Tr}(|B|^2)} $$ Now, if $M$ has rank $r$, then the polar decomposition tells us that there exists a partial isometry $U$ (i.e. $U^*U$ is an orthogonal projection) with $\operatorname{rank}(U) = M$ such that $M = U|M|$ and $|M| = U^*M$. From there, $$ \operatorname{Tr}(|M|)^2 = \langle M,U \rangle^2 \leq \langle M, M \rangle \langle U,U \rangle = \operatorname{Tr}(M^*M) \cdot \operatorname{rank}(M) $$

  • $\begingroup$ Thanks, this makes sense, although I was under the impression that the polar decomposition (for matrices) is always defined with $U$ being unitary. Is this a common generalization? $\endgroup$
    – egrr
    Aug 18, 2017 at 0:24
  • $\begingroup$ @egrr This "generalization" is more common in functional analysis, where the "matrices" at play act on infinite dimensional vector spaces. $\endgroup$ Aug 18, 2017 at 0:34

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.