Skip to main content
11 events
when toggle format what by license comment
Oct 4, 2024 at 14:47 vote accept biquart
Oct 4, 2024 at 11:11 answer added lightxbulb timeline score: 3
Oct 1, 2024 at 14:59 comment added Kevin Reid Someone (could be you, could be @lightxbulb, could be someone else) should write and post an answer (that can then be marked as accepted, if you like it). Questions should not be left with only comments.
Oct 1, 2024 at 13:20 comment added biquart Yes, I've read that hit and trial is usually the preferred method to get the step size, when you apply some unusual transformation that gives only a distance bound. I'm sort of a first timer on Stack Exchange. Since I think you've addressed nearly everything (thanks!), what would be the appropriate way to close the question?
Sep 30, 2024 at 16:05 comment added lightxbulb Actually that's not necessary with the numerical approximation, it would still work for functions that are almost everywhere differentiable (you can see this from the box sdf for example), and I think those constraints can be relaxed further. I would prefer using the analytical gradient most of the time though, as the step size in the numerical approximation is notoriously hard to get right.
Sep 30, 2024 at 14:42 comment added biquart Yeah, it would also need to be differentiable, at least
Sep 30, 2024 at 10:15 comment added lightxbulb To be precise you probably need a bit more than continuity.
Sep 30, 2024 at 9:07 comment added biquart Hmm, I see. This is quite enlightening, thanks for the comment, that makes sense. I was reasoning this way - if f approximates g, then it doesn't necessarily follow that the derivative of f approximates that of g. Which is true of course, but approximation at one point, and approximation over a surface are quite different.
Sep 29, 2024 at 10:45 comment added lightxbulb My guess is that since $f(p)=0$ must still hold at the points of the surface, and that $f$ is continuous, as you get closer to the boundary the normal must converge to the actual one.
S Sep 28, 2024 at 6:11 review First questions
Sep 28, 2024 at 16:08
S Sep 28, 2024 at 6:11 history asked biquart CC BY-SA 4.0