# Fast computation/estimation of the nuclear norm of a matrix

The nuclear norm of a matrix is defined as the sum of its singular values, as given by the Singular Value Decomposition of the matrix itself. It is of central importance in Signal Processing and Statistics, where it is used for matrix completion and dimensionality reduction. A question I have is whether it's possible to compute the norm of a given matrix faster than the time needed to compute the svd. Since we don't need all the singular values but only their sum, this seem possible. Alternatively, perhaps it could be possible to approximate it with simulation methods and/or random projections.

-
Where can i find the definition of nuclear form in wikipedia? –  user4140 Feb 7 '13 at 3:15
This is actually a question of active research. I'd say that a more appropriate forum to ask this question is on the SciComp SE site (scicomp.stackexchange.com). In fact I know that one of the more active members there is working on a variety of approaches to solve this problem. In general, though, the nuclear norm requires the computation of all singular values (though not the singular vectors), unless you know in advance that the matrix is low rank, which does happen often in the applications you list. –  Michael C. Grant May 20 '13 at 15:14
If you do choose to post there, make sure to mention that it was already asked here (and include a link), so it isn't flagged as a duplicate. And likewise, do respond here to let us know that you did so. –  Michael C. Grant May 20 '13 at 15:18

Since you asked for an approximation as well, you might find the paper "Some simple estimates for singular values of a matrix" by Liqun Qi useful. There are some nice estimates there.

However, if these are not precise for you, you might consider computing SVD with a low precision, i.e., do one or two or three iterations and then use the above estimates. Depending on the size of your matrices, this might give a nice speedup.

Since the estimates are of the form $\sigma_i \in [ a_i, b_i ]$, you will also have an estimate of error, so you can do iterations of SVD computation until the absolute error $\sum_i (b_i - a_i)$, or some of its relative counterparts, is small enough.

Apart from that, I don't think there is much to be done. SVD exists to avoid the computation of $A^*A$ when you want these eigenvalues, so such tricks would not help, unless your matrices have some nice properties you didn't mention.

Estimates and approximations are really not my area of interest, so maybe there are results newer than the above paper.

-

A tentative answer: the nuclear norm of $A$ is the trace of $\sqrt{A^*A}$ where $A^*$ is the conjugate-transpose of $A$, and $\sqrt{\cdot}$ is the matrix square root. So provided you can calculate matrix square roots faster than singular value decompositions, this might be useful.

-
The problem with this approach, of course, is that one would very much want to avoid computing $\mathbf A^\ast\mathbf A$ in the first place... –  Ｊ. Ｍ. Apr 25 '12 at 13:15