I wonder if the same thing can be achieved (or at least the expected result would be the same) by blurring the 1 shadow version somehow?
That's an interesting idea, but I think that approach would fail for sharp shadows (ie long, thin ones), which would render the shadow as intermittent blurry dots rather than a continuous shadow.
At what point would increasing the number of random samples lead to overfitting? And can learning models used to select these random samples?
Seems like the number of random samples is more application based, since games simple want computational speed, while digital renderings want higher quality over speed.