Randomization of Approximate Bilinear Computation for Matrix Multiplication
We present a method for randomizing a formula for bilinear computation of matrix products. We consider the implications of such randomization when the formula itself is approximate, and when the formula is exact but its computation is plagued by numerical error due to finite precision arithmetic. Our theoretical results and numerical experiments indicate that our method can improve performance in both settings for a negligible increase in computational complexity.
READ FULL TEXT