Represent Schatten-p norm in CVX


#1

I know that the CVX function

norm_nuc(X)

is the nuclear norm of X, which is also Schatten-1 norm. For Frobenius Norm (or Schatten-2 norm), it is represented by

norm(X,'fro')

But how can I represent the general Schatten-p norm for 1<p<2 in CVX?


(Michael C. Grant) #2

I do not believe it is possible.


#3

@mcg Really? Well, such a bad news.


(Michael C. Grant) #4

If CVX could solve everything (or even just everything convex), what would people write journal articles about? :slight_smile:


#5

@mcg Yeah, you are right! Thanks!


(Madeleine Udell) #6

Actually, any symmetric monotone function of the singular values of a matrix is DCP-representable. See Ben-Tal and Nemirovski “Lectures on Modern Convex Optimization”, Proposition 4.2.2. But the best representation (presented there, at least) is huge!

Let’s see how this works for a PSD matrix A. (The case of rectangular matrices is similar.)

  • the sum S_k(A) of the top k eigenvalues of A is the solution to the problem
minimize sk + Tr(Z)
subject to
       Z ⪰ 0
       A ⪰ 0
       Z + sI ⪰ A
  • g(\lambda(A)) is the solution to the problem
minimize g(s)
subject to
    s[1] >= ... >= s[n]
    S_k(A) <= s[1:k],  k = 1, ..., n
    trace(A) == sum(s)

The only difficulty is that this representation introduces a new n x n PSD variable for each eigenvalue, resulting in a problem with n n x n SDP constraints — which is not exactly fast to solve.


Computing or bounding a modified Ky Fan norm
Using svds(A,k)
(Michael C. Grant) #7

This is great, Madeleine. Of course, we do have the sum of the top k eigenvalues in A. I’d say that even though this corrects my claim that it is not possible, it is certainly not practical for all but the smallest matrices… If someone were to implement this, it would be interesting to see!


(Mark L. Stone) #8

What kind of run times for some representative problems?


(Mark L. Stone) #9

Madeline,

Per my answer in Maximize the minimum singular value , the minimum singular value of a matrix of dimension >= 2 is neither convex nor concave in general. This is not inconsistent with your statement though, because min is not a monotone function.

Truth in advertising: I had to look up the definition of monotone function.


#11

This is fantastic – I have been looking for something like this for ages! I went ahead and implemented this for the (k,p)-norms (i.e., the Schatten p-norms of the k largest singular values), and the code is available here as part of my “QETLAB” package (but you don’t need the whole package: just this one .m file).

It actually runs much better than I thought it would. On my laptop, it computes the Schatten 3-norm of a random 100-by-100 matrix in about 3.5 seconds, and the Schatten 5-norm of a random 200-by-200 matrix in about 11.5 seconds.


(Michael C. Grant) #12

Wonderful! Thanks again Madeleine!