I am wondering why some large exponentials yield wrong values, specifically always 1 or 0 julia> 123^(456^789) 1 julia> (0.1)^(456^789) 1.0 julia> 456^789 0 This results in wrong values inside functions julia> atan((0.1)^(456^789)) 0.7853981633974483 julia> atan(1) 0.7853981633974483 I’d expe…

[image] Discussion about integer overflow Internals & Design Integer overflow has been thought about since at least 2012 , is documented in the FAQ , has been discussed repeatedly, has been implemented in a package , and finally Julia is not at that stage of development anymo…

To summarize that very long discussion: No, this is expected behavior. Julia integers are machine integers, ^ on integers does integer exponentiation. Checking for overflow & throwing an error/giving a warning would prevent a lot of optimizations like vectorization to take place. If you want to do …

So this is where the difference lies: julia> 456^789 0 julia> 456.0^789 Inf

And Inf is not defined for integers anyways… Only for floating point numbers…

[image] nil: julia> 456^789 0 This results in wrong values inside functions This is right(!) in some technical (modular) sense. Modular arithmetic is the default for machine integers, and they can’t be Inf. So Int^Int, CAN’T result in Inf, if, given that it results in Int. I would argue it s…

There has been some serious discussion of doing an overflow check for integer^integer non-literal powers, though I can’t find it at the moment (is there an issue)? Rationale: integer^integer powers rarely seem to be performance sensitive in real applications, and are expensive enough that an ove…

Yes, I was searching for it in the very long discussion, but gave up. I believe Karpinski or some such was open to it, but I just opened an issue on rather the more general Float64 result.

Sorry for reviving this issue, I should have checked more thoroughly. I think I now get where the problem’s coming from, just have to be more careful with the modular nature of integer exponentials on a finite memory computer.

I guess an error message would do it for starters because I could only imagine hunting for such a problem in a larger codebase, that would certainly not be pleasant. Admittedly, I don’t know how easy such an implementation would be in a general case or possible at all, where values aren’t given befo…

[image] nil: what I just find weird and confusing (especially for starters coming from python) is the fact that integer exponentiation can overflow to 0 That’s one of the possible design choices for integer exponentiation and frequently chosen when speed gets the highest priority. (Strictly s…

Should large exponentials not yield Inf?

General Usage

nilshg May 16, 2022, 10:26am 2

Topic		Replies	Views
Weird floating point under/overflow behavior General Usage	3	300	March 31, 2024
Integer power of large Integer returning very wrong results General Usage base	7	581	May 15, 2023
Exponent operator failing for large exponents General Usage	11	471	October 5, 2022
Julia messes up integer exponents? Performance integer-overflow	106	7408	February 21, 2019
Long Math Operation result is 0, In Python isn't New to Julia question , integer-overflow	7	1193	January 2, 2022

Should large exponentials not yield Inf?

Related topics