ABSTRACT
This leads us to the method of Sinkhorn balancing. Simply described, the Sinkhorn balancing algorithm derives a doubly stochastic B from a zero-one A by first dividing the entries in each row by the sum of the non-zeros in that row. The result of this is a row stochastic matrix that is not necessarily column stochastic. Next divide by column sums giving a column stochastic matrix that is not necessarily row stochastic. Then divide by row sums again, etc. This will converge quickly [12] if all of the non-zero elements of A are supported and the limit is the Sinkhorn balance of A. (We will return to the
