What the Gaussian fit predicts

Deep dive 14’s alternating least squares, rank 3, pointed at the crime counts. It knows nothing about counts.

It predicts negative crimes on more than a thousand cells, and three quarters of them are cells whose true count is zero. That is not a numerical accident: with 41% zeros, the cheapest squared-error fit lets the model dip below the floor and average out.

Nothing here is a criticism of ALS. The optimiser found a good minimum of the objective it was given. The objective was the problem, and it was chosen by default.