NaN and Infs popping in my matrices after a few iterations of EM algorithm

Hello ;

So, I'm implementing the EM algorithm in Matlab, but my matrices quickly end up contaminated by NaN and Inf values. I think it might be caused by matrix inversions, but I'm not sure it's the only reason.

Here is the code : em_algo.m http://pastebin.com/vJGqQPw7 iterates until convergence (which never happens due to my issue) and calls smoother.m http://pastebin.com/yD9Je37i at each iteration. The NaNs and Infs start popping around the ~8th iteration, usually.

I guess in there somewhere I'm doing something unholy with my matrices, but I really have no clue about what's wrong. I trust your expertise.

Thanks in advance for the help!

Answer 1

[José-Luis]

It sounds like you are running into numerical stability issues. Look at the following example, that might help you understand what is happening:

 val = 1;
 while (val ~= Inf)
     val = val / 10^-100
 end

The NaN's that pop up can be a result of

 0/0

Answer 2

[Jérôme]

Thanks José-Luis. Indeed, something like that might be the source of my problem. But I have a hard time finding out WHERE it happens in the code. :/

Another issue (or most likely another aspect of the same one) is the fact that some values seem to grow out of bonds which probably leads to Inf values. But again, I have no idea why this is happening.

[Walter Roberson]

dbstop if naninf

[José-Luis]

If your matrices are close to rank deficient, and you invert them, then numerical problems can pop up.

Answer 3

[Matt J]

Set a conditional breakpoint at line 45 (making it conditional on i==8 or whenever the NaNs start to pop up). When the code stops there, see whether inv(A) has sensible values.