Question

[Edit: The whole thing has a very simple solution: the matrix used the single datatype instead of the default double]

I have just noticed a somewhat peculiar (I think) behaviour in matlab and wonder what’s causing it. I have a 10000×500 matrix M with values ranging from

min(min(M)) = -226.9723 to 
max(max(M)) =  92.8173

and

exp(-227) =  2.6011e-99
exp(93) = 2.4512e+40

but if I exp the entire matrix, this matrix has inf values:

ii = isinf(exp(M));
sum(sum(ii))
ans =
     2

How does Matlab store the values in the matrix so that operations on individual elements can give a different result than when doing the same operation on the matrix itself?

I.e.

expM = exp(M);
exp(M(1)) == expM(1) ; %can be false, which I find surprising

I know I have to change the algorithm anyway as the high exponents will give inexact results even if I can avoid inf values. It happens in a formula for a artificial neural network calculation like:

 sum(log(1+exp(ones(numcases,1)*b_h + data*w_vh)),2);

so my plan is to split this up into two cases, first where the exponent is small I do the calculation as above, for high values I approximate

log(1+exp(ones(numcases,1)*b_h + data*w_vh)

with

ones(numcases,1)*b_h + data*w_vh

Does that sound reasonable? My reasoning of course is that

log(1+exp(x)) ≈ log(exp(x)) ≈ x, for large x

btw: is there a better way to get the maximum element of a matrix other than doing max twice as in max(max(M))?

Answer 1

Ok, I found the error: my matrix was of type single, but when I copied the value in a new variable that would be a double with of course a different max value. I answer this myself here so the question won’t stay unanswered. Thanks for the tips, I found the cause when trying to build a repro-case 🙂