HP-50g: help remembering library with Interval Arithmetic

[johnb]

(02-05-2024 02:33 PM)Gjermund Skailand Wrote:  As far as I remember, the main problem with the (simple implementation) is that the uncertainty tends to blow up if the intervals are ...too... big, resulting in something that is practically unusable.

@Gjermund - thank you for your efforts in producing this, however long ago it might have been!

I've found similar behavior in my own (mostly dumb, brute force) implementation as well, but some of my blowups are actually a good warning that a chained calculation has gone wrong.

(02-05-2024 02:33 PM)Gjermund Skailand Wrote:  I spent a lot of time trying to get border cases calculated correctly.

That effort is much appreciated: it represents your code's chief value to me. Also, I think you're very likely a better mathematician or numerics programmer than I am. So it's great to have your implementation to compare mine with.

My 48g code is all in User RPL so I might have to learn System RPL to make it worthwhile to others.

Of note: My experimental code differs from everything else that I've seen in that I store triplets not tuples: not just the endpoints of the range but also the "most likely value," which permits it to track asymmetrical error bands. I've searched the web (also for C/C++ implementations, as well as other languages) and I'm quite honestly surprised no one else has tried to do this.

I understand that the serious Numerical Analysis community is exploring hardware implementations. I would love to see a CPU that directly supports my triplet approach, along with a standard library that properly uses this for transcendentals and other functions!