> I think I agree with Thorsten and Christoph that 3.6 behaviour is
> mathematically correct and that 3.5 had a bug/feature.

I admit that a part of me agrees that the clipping is kind of an ugly,
non-mathematically-correct thing to do. But the fact is that it must be done
at some point when the image formats that we're outputting to require it. I
do like the idea of not clipping until just before output because in a way
it *feels* like the right thing to do - not introducing the limitation until
it's actually necessary.

Nonetheless, this "feeling" that clipping before sampling is kind of ugly is
not entirely relevant. I've "felt" a lot in the past that my mathematical
results were right or wrong, and I've often been incorrect.

Clipping *at all* is mathematically incorrect. If a sphere has color rgb 10,
then values of 10 should be written to the output. But obviously that can't
be done if the output format doesn't support it.

So we have to live with the clipping one way or another. We have, at the
point of admitting this, already introduced a certain level of mathematical
incorrectness. Our goal should be to do it in the most useful way.

I have seen no evidence that clipping after averaging samples is any more
"mathematically correct" than clipping beforehand. As I just said, both ways
are slightly incorrect, but we have no choice, we must clip at some point.
What I'm saying now is that there is no evidence that I've seen that one way
is *more correct* than the other.

However, I have seen evidence that clipping after averaging is much less
useful.

If anyone has a concrete reason why clipping after averaging is more
mathematically correct, I'm open to hearing it. If one has already been
mentioned, please point me to it. But I don't believe I have seen a concrete
reason.

In summary, my argument here is that neither method is more mathematically
correct than the other, but the 3.5 method was most certainly more useful.

 - Slime
 [ http://www.slimeland.com/ ]

Post a reply to this message