Interesting observation. This works presumably because the output of a pixel is not just one bit, but some sort of continuous brightness reading (discretised, of course, sooner or later)?
The light is blurred, with a fairly predictable pattern, so if you fit a function to the shape, you can find the peak of the function and that is the most likely center position for the point source.
