How to compute entropy more efficiently?

Dear All

I want to computer the entropy for the distribution over hundreds outcomes for many times. I did it as follows: (copied from profiler, first column for time consumed , and second column for call numbers).

Do you know any more efficient way to do this?

Since it has been vectorised, I guess 'mex' may not help too much. Is it right?

Thanks!!!

==================

 6.31  993683   16 p1 = Alpha1./sum(Alpha1); 

5.06  993683   17 p2 = Alpha2./sum(Alpha2); 

 16.56  993683   18 h1 = sum(p1.*log(p1)); 

 13.05  993683   19 h2 = sum(p2.*log(p2));

===================

Answer 1

[Bjorn Gustavsson]

On my version of matlab I get a big reduction in run-time by doing:

h1 = p1*log(p1.');

or

h1 = p1.'*log(p1);

depending on which version is the inner product for vector p1.

HTH

[Yingke]

I found a package which provide better (C and ASM) implementation of log. Anyway, your suggestion is still good. It really helped a bit. Thanks!

[Bjorn Gustavsson]

Ops, I was wrong. Seems most time gain in my version was that I forgot to call the log -> the log-ing is what takes time.