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

Suppose that I have two positive matrices, $A$, and $B$, and I know their Frobenius-Perron eigenvalues ($\lambda_A$, $\lambda_B$) and eigenvectors ($v_A$, $v_B$). I'm interested in what I can say about the Frobenius-Perron eigenvalue and eigenvector of $C=A+B$. I'm sure that this question must have been considered before, but I haven't found a good source.

Here are two particular questions:

(1) One clear bound on $\lambda_C$ is that $\max(\lambda_A,\lambda_B) \le \lambda_C$. Is there a nice upper bound? Also, if $v_A=v_B$, then $\lambda_C = \lambda_A + \lambda_B$. But can we do better than this if we know $v_A$ and $v_B$?

(2) What can we say about $v_C$? I expect it to be "somewhere between" $v_A$ and $v_B$, but that's just vague intuition. (Assume all the eigenvectors are normalized.) Is there a simple bound on the elements of $v_C$ given knowledge of $\lambda_A$, $\lambda_B$, $v_A$, and $v_B$?

I'd prefer not to assume that $A$ and $B$ are symmetric, but if the question has only been solved for the symmetric case, that's fine.

Edit: I realized that I should clarify that all of the $v$'s are right eigenvectors. If it helps to know the left Frobenius-Perron eigenvectors $u_A$ and $u_B$, then feel free to use this too.

Edit 2: In light of David's answer, then I've corrected my "intuitive" bound on $\lambda_C$ which was completely wrong.

share|improve this question

1 Answer 1

First, the definition of $positive$ should be clarified. It could mean all entries strictly positive, or merely nonnegative, or that the matrix is primitive (all entries nonnegative and some power strictly positive).

I will use $\rho(M)$ to denote the spectral radius of a matrix $M$, which in case $M$ is nonnegative, is the Perron eigenvalue.

Note a weird example: if $M$ is strictly upper triangular with all entries above the diagonal strictly positive, then the spectral radius is $0$, and the same for $M^T$. However, $M+M^T$ is primitive, so its spectral radius exceeds zero. We can modify this to obtain strictly positive $M$ with $\rho(M)$ small, but $\rho(M+M^T)$ big, simply by adding a tiny amount to all the zero entries. In this case, $\rho(C)$ can be made much larger than $\rho(A) + \rho(B) $.

A well-known result ($well-known$ means I can't give a reference off the top of my head) is that if $C \geq A$ entrywise, $C \neq A$, and $A$ is strictly positive (or merely primitive), then $\rho(C) > \rho (A)$. (But the argument is easy: use an H-transform to convert $A$ to one with all column sums equal; adding anything will increase the spectral radius.) So if $A$ and $B$ are primitive, then $\rho(A+B) > \max \{\rho(A), \rho(B) \}$.

I doubt there's much of anything beside the obvious that can be said about the Perron eigenvectors.

Perhaps you have matrices in special forms in mind?

share|improve this answer
    
Thanks. I corrected the bound above. I don't have special restrictions in mind right now except that $A_{ij}$ and $B_{ij}$ are in $(0,1]$. But if the problem can't be solved in general, then maybe it can be solved for random matrices with entries in $(0,1]$, and the result can be used as a heuristic (although not a rigorous result). A good heuristic is really all I need. –  sasquires Nov 13 '13 at 13:49

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.