The way I understood it - you take the image of the ground (the object you're projecting shadows onto), divide it into however many pixels you set in the options, and then for each pixel, you take the path from the pixel center to the light source. If the path is obstructed, it's in the shadow, otherwise not.
This procedure seems deterministic, no matter how many times it's done. Therefore, I don't see how there'd be any difference when blending multiple samples. Where am I wrong in my assumptions?
You're totally right - except for when you enable soft shadows. What that does is take that same deterministic approach from different points within the light's radius and averages all the results together. Set the viewport samples to 1 and then 2 and compare to see exactly how it works.
Oh I see, so Blender treats the sphere point light source as if light comes from every point from inside the sphere, uniformly distributed? So for sample size one, it takes one point and treats it as if all light came from there, for 2 it takes 2 points and treats it 50/50 etc.
If so, that makes sense, thanks.