For a matrix $A\in\mathbb R^{n\times m}$, we consider the vector of its singular values $\sigma = [\sigma_1,\dots,\sigma_{\min\{m,n\}}]^T$. We define the $p$-Schatten norm $$ \|A\|_{S,p} := \|\sigma\|_p $$ as the usual $p$-norm of $\sigma$. How do I show that this $p$-Schatten norm satisfies the triangle inequality?

Everywhere, where Schatten norms are used it is implied that it's actually a norm, but I don't have a good idea on how to prove this.

There is this related question on Ky Fan norms, but I actually don't understand the answer.

  • $\begingroup$ Maybe it helps to know that $\DeclareMathOperator{\tr}{tr}$ $$ \|A\|_{S,p} = \Bigl(\tr\bigl((A^TA)^{p/2}\bigr)\Bigr)^{1/p}. $$ But I still don't know where to go from there. $\endgroup$ – Stefan Hante Jun 14 '18 at 9:00
  • $\begingroup$ Pedersen's "Analysis Now" proves this (for arbitrary trace-class operators on a Hilbert space) with polar decomposition; I imagine something similar can be done with singular value decomposition. Bhatia's "Matrix Analysis" has a proof that handles arbitrary unitarily invariant norms. I've been looking at both but I have failed to fined a quick, punchy justification of subadditivity. $\endgroup$ – Ben Grossmann Jun 14 '18 at 15:45
  • $\begingroup$ If you can access Bhatia's Matrix analysis, Theorem IV.2.1 will do the job $\endgroup$ – Ben Grossmann Jun 14 '18 at 15:50

If we adapt the proof of IV.2.1 from Bhatia's Matrix Analysis, it suffices to prove the following facts:

  • The Ky-Fan norms satisfy the triangle inequality
  • For vectors $x, y \in \Bbb R^n_+$: if $$ \sum_{j=1}^k x_k^\downarrow \leq \sum_{j=1}^k y_k^\downarrow, \quad k = 1,\dots,n $$ then $\|x\|_p \leq \|y\|_p$

  • For vectors $x, y \in \Bbb R^n_+$: if $x \leq y$ (entrywise), then $\|x\|_p \leq \|y\|_p$

The result regarding the Ky-Fan norms can be proven as follows:

  • If $A$ is a Hermitian matrix and $\lambda^{\downarrow}_j$ denotes the eigenvalues in decreasing order, then $$ \sum_{j=1}^k \lambda_k^\downarrow = \max \sum_{j=1}^k x_j^TAx_j $$ where the maximum is taken over orthonormal $k$-tuples of vectors $\{x_1,\dots,x_k\} \subset \Bbb R^n$ (known as Ky-Fan's max principle).
  • Consequently: for Hermitian matrices $A,B$, we have $$ \sum_{j=1}^k \lambda_k^\downarrow(A + B) \leq \sum_{j=1}^k \lambda_k^\downarrow(A) + \sum_{j=1}^k \lambda_k^\downarrow(B) $$
  • The same applies to the block matrices $$ \tilde A = \pmatrix{0&A\\A^T & 0}, \quad \tilde B = \pmatrix{0 & B\\B^T & 0} $$ and notably, $\lambda_j^\downarrow(\tilde A) = \sigma_j^\downarrow(A)$.
  • $\begingroup$ That's nice. What do the little downwards arrows mean? $\endgroup$ – Stefan Hante Jun 15 '18 at 11:11
  • $\begingroup$ @Wauzl it means that the relevant quantities are listed from greatest to least. Of course, this is the usual convention for singular values. $\endgroup$ – Ben Grossmann Jun 15 '18 at 13:38
  • $\begingroup$ @BenGrossmann Could you explain why the last equality holds? Can't the eigenvalues of $\tilde A$ even be negative? $\endgroup$ – Florian Ente Dec 13 '20 at 20:19
  • $\begingroup$ @ludoer If $A = U \Sigma V^T$ is a singular value decomposition, then we have $$ \tilde A = \pmatrix{U & 0\\0 & V} \pmatrix{0 & \Sigma\\ \Sigma^T & 0} \pmatrix{U & 0\\ 0 & V}^T. $$ With that in mind, consider the eigenvalues of $$ \pmatrix{0 & \Sigma\\ \Sigma^T & 0}. $$ $\endgroup$ – Ben Grossmann Dec 14 '20 at 1:01
  • $\begingroup$ @ludoer Sorry, I misunderstood your question at first. The last equality holds only for $j = 1,\dots,\min\{m,n\}$. After that, we reach zero-eigenvalues and then the negative eigenvalues $-\sigma_j(A)$. $\endgroup$ – Ben Grossmann Dec 14 '20 at 1:03

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.