numerically stable way to multiply log probability matrices in numpy

Question

I need to take the matrix product of two NumPy matrices (or other 2d arrays) containing log probabilities. The naive way np.log(np.dot(np.exp(a), np.exp(b))) is not

Answer 1

You are accessing columns of res and b, which has poor locality of reference. One thing to try is to store these in column-major order.