Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

I have a problem that necessitates solving a large non-negative least-squares problem. My matrix A is large, sparse, highly rectangular (num rows >> num cols) and nearly binary. However, A is not necessarily of full column-rank, causing my non-negative least-squares solver (http://www.jasoncantarella.com/webpage/index.php?title=Tsnnls) to fail. Is there an efficient algorithm that will allow me to select a maximum cardinality set of linearly-independent columns from A so that I can solve my least-squares problem?

share|improve this question
add comment

1 Answer

up vote 0 down vote accepted

By nearly binary do you mean most entries are $0$ or $1$ but a few are not? I will assume that all entries are $0,1.$ My main comments involve merely noting zero and non-zero.

To be sure of having maximum rank you will need to look at each column. You can start with any independent set (such as the set consisting of a single column) and then examine the rest of the columns in some order. If the next column is independent of the current set adjoin it as a new member otherwise discard it. This will result in a maximum size set (a basis of the column space). Depending on how sparse it is, it could be pretty easy to decide when the next column can be added, at least for a while.

Here are a few simple ideas to boost efficiency:

  • If the matrix is very sparse then use efficient sparse matrix algorithms and representations, these might involve storing it as a set of ordered tuples $(i,j)$ or $(i,j,a_{ij})$ showing the location (and value) of the non-negative entries. There is bound to be highly efficent and well validated existing code which will provide you with a basis for the column space of a (sparse) matrix.

  • Let the depth of a column be the first row with a non-zero entry. Sort the columns according to depth (don't bother to break ties by looking further.) So far this might not require you to look at all the entries, just those down to the first non-zero entry. If the matrix is quite sparse then one might expect many different heights.

  • Pick one column from each depth class, This will give you a starting independent set which might include a relatively large number of columns. Assume that this has been done.

  • Go through the current set of independent columns, look for any rows which are all zero (for the selected column vectors) Now look through the undiscarded columns and try to find one which is non-zero in that position. If one is found then add it to the set. This will enlarge the set of columns.

share|improve this answer
    
Thanks for the response Aaron! By binary, I do indeed mean that most entries (but not necessarily all) are 0 or 1. If the final method necessitates a binary matrix, then the non-zero entries can be set to 1 without too much cause for concern. To solve the general problem, my current approach is to consider the procedure you outlined below to obtain an initial set of independent columns, then to proceed via a Gram-Schmidt procedure. Orthogonalize this set of columns, and for each remaining column, try to add it to the basis, dropping it if it's orthogonalized norm is 0. I think this works. –  Rob Oct 17 '12 at 18:07
    
Using Gram-Schmidt means using floating point computations. Especially if the entries are all integers or even $0,1$ just do column reduction. But again, there are sure to be optimized algorithms for this exact problem in general and in the case of sparse matrices. –  Aaron Meyerowitz Oct 18 '12 at 7:42
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.