Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

I read this sentence.

Suppose that the matrix $A_{ij}$ of dimension $n_i \times n_j$ has rank $k$ to precision $\epsilon$, then there exists a factorization of $A_{ij}$ of the form: $A_{ij} = L_i S_{ij} R_j + \text{O}(\epsilon)$.

I wonder what does matrix rank $k$ to precision $\epsilon$ mean?

Thank you.

share|improve this question
    
Where did you read this? –  Chris Eagle Oct 25 '11 at 14:17
1  
It must mean that there's a rank-$k$ matrix within a distance of $\epsilon$ from $A$, for some appropriate (but unidentified) norm on the space of $n_i\times n_j$ matrices. I wonder what $k$ has to do with the conclusion of the claim, though. –  Henning Makholm Oct 25 '11 at 14:19
1  
@Henning: the "unidentified" norm is often the 2-norm, especially in the case of diagnosing badly-behaved least-squares problems and other problems that necessitate the use of orthogonal matrices for decompositions. –  J. M. Oct 25 '11 at 14:37
    
@ChrisEagle This is a restatement of Theorem 3 in ON THE COMPRESSION OF LOW RANK MATRICES by H.Cheng. You could get it through Google Scholar search. –  Yao Jin Oct 25 '11 at 16:59
add comment

1 Answer

The rank of a matrix is the number of its nonzero singular values.

Rank to precision $\epsilon$ means that in computing the rank of the matrix, we consider every singular value of the matrix that is less than $\epsilon$ as zero.

This is also known as "numerical rank": the number of singular values greater than $\epsilon$.

share|improve this answer
    
I see. Thank you. –  Yao Jin Oct 25 '11 at 14:28
3  
Put another way: the SVD is a useful tool for diagnosing the numerical rank of a matrix. If the matrix has any singular values less than machine epsilon (or more often, the norm of the matrix times machine epsilon), you have what is called an ill-conditioned matrix. –  J. M. Oct 25 '11 at 14:34
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.