but they will not take over too soon imho.. a bit too artifacts prone.
Numeric stability problems are gone on the G80 with fp32 (and maybe even better with fx32 - haven't got DX10 up and running yet). I suspect R600 will be able to do the same, or something similar. Note that hardware filtering isn't even required to get really nice results, but more on that in the near future
You can also reduce/eliminate light bleeding by just lopping off the tail of the distribution (Mint suggested this at one point I think). Of course no matter how much one cuts off, a degenerate case can be constructed. Still this solution has proven to be very effective in my testing, and it's an artist-editable one-liner.
Furthermore anything short of brute-force PCF of the entire filter region (which is prohibitively expensive) will have *some* incorrect case... i.e. there is no silver bullet to visibility.
I'm becoming pretty certain that VSM is the best you can do with two pieces of data. It's certainly the best upper bound that one can get with that amount of information, but I think one would be hard pressed to find a more suitable approximation even. In particular by modifying the falloff function as mentioned above, light bleeding can be eliminated at the cost of some over-darkening. The key however is that the shadow edge will still be (projectively) anti-aliased even if the whole of the distribution function is removed (i.e. converted to a step function).
Anyways getting back on topic: shadows and visibility are just very hard problems, particularly when multiple overlapping occluders are involved. Almost all shadow implementations have similar problems due to this complexity.
I tended to find that, because statistics assumes a bell-shaped curve, a lot of transitions in the shadow map simply didn't match that model very well. <shrug>
I'm not sure what you mean - there's nothing inherent about using statistics that assumes "bell-shaped" distributions. In particular, Chebyshev's Inequality is an upper bound for *all* distributions with a given mean and variance, which is precisely why it is useful.
PS: Simon F, I'd love to see some of those pretty pictures if you're willing to share