Quote - SPPM (just like PPM and plain photon mapping) don't calculate an individual radiance estimate for each point being shaded but do a lookup in a photon map and use use the radiance values of the closest points. That's where the algorithm is biased.

Ah, but that fact is covered. The estimation radius SHRINKS each time around, eventually it will have shrunk too small for your image to resolve the bias at the resolution you set, and the image has converged at that point. (How fast it tightens is controlled by the "alpha" setting in Lux)

 

As I understand it, traditional photon mapping is biased in 2 ways:

1. It takes infinite photons to get everything, which you can't do, it requires infinte memory and you couldn't ever stop to do the eye pass.

2. What Stewer said, there is an estimation radius.

SPPM gets around #1 by repeating the passes over and over, so the effective photon count approaches infinity, and it gets around #2 by tightening the search radius (so as the render runs, it approaches zero)