Mercurial > hg > camir-aes2014
annotate toolboxes/FullBNT-1.0.7/KPMtools/logsumexpv.m @ 0:e9a9cd732c1e tip
first hg version after svn
| author | wolffd |
|---|---|
| date | Tue, 10 Feb 2015 15:05:51 +0000 |
| parents | |
| children |
| rev | line source |
|---|---|
| wolffd@0 | 1 function s = logsumexpv(a) |
| wolffd@0 | 2 % Returns log(sum(exp(a))) while avoiding numerical underflow. |
| wolffd@0 | 3 % |
| wolffd@0 | 4 % e.g., log(e^a1 + e^a2) = a1 + log(1 + e^(a2-a1)) if a1>a2 |
| wolffd@0 | 5 % If a1 ~ a2, and a1>a2, then e^(a2-a1) is exp(small negative number), |
| wolffd@0 | 6 % which can be computed without underflow. |
| wolffd@0 | 7 |
| wolffd@0 | 8 % Same as logsumexp, except we assume a is a vector. |
| wolffd@0 | 9 % This avoids a call to repmat, which takes 50% of the time! |
| wolffd@0 | 10 |
| wolffd@0 | 11 a = a(:)'; % make row vector |
| wolffd@0 | 12 m = max(a); |
| wolffd@0 | 13 b = a - m*ones(1,length(a)); |
| wolffd@0 | 14 s = m + log(sum(exp(b))); |
| wolffd@0 | 15 |