Suppose that you have a $n\times n$ matrix $A$ that is symmetric and has zero diagonal, such as for example $$ A=\pmatrix{ 0 & 2 & 2\\ 2 & 0 & 1\\ 2 & 1 & 0}, $$ and you want to represent it by a low rank approximation but respecting the symmetry and the zero diagonal in the output of that approximation. Like doing the eigenvalue decomposition by the spectral theorem $$ A=\sum_{i=1}^n \lambda_iv_iv_i^T $$ and then truncate this sum to get an approximation for $k\leq n$ (assuming $|\lambda_1|\geq\ldots\geq|\lambda_n|$). But if you do this for $A$ of course you get something like (with two eigenvalues) $$ A=\pmatrix{ 0 & 2 & 2\\ 2 & 0.5 & 0.5\\ 2 & 0.5 & 0.5 },$$ which is not in the spirit of what I want.

Question: Is there a standard procedure/method for dimension reduction respecting these constraints?

EDIT 1: I realized that since you need to have a zero diagonal output as an approximation matrix, this implies full rank except if some of the entries are zero. So I guess the answer is in choosing which entries are set to zero and how the others are reweighted.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.