The SVG standard defines a "bias" property for feConvolveMatrix to offset the result of the filter. It specifies that this value should simply be added to the result for each channel. Implicitly the standard defines that these are premultiplied color values, however, this seems to make little sense, as the bias has very little meaning without taking the alpha channel into account.
The standard also says that one of the applications of a bias is "to have 0.5 gray value be the zero response of the filter", this seems to be directly contradictory with adding bias to the premultiplied values (as then the same bias could suddenly stand for white if alpha was 0.5).
So, what to do? Batik doesn't seem to support bias!=0 at all, so there is also no precedent. I currently let feConvolveMatrix make bias relative to alpha.