#sum #accurate #f64 #float #exact

fsum

The library to calculate accurate sum of floats

3 releases

0.1.2 Feb 25, 2022
0.1.1 Feb 22, 2022
0.1.0 Feb 21, 2022

#951 in Math

Download history 275/week @ 2024-11-17 305/week @ 2024-11-24 286/week @ 2024-12-01 360/week @ 2024-12-08 1007/week @ 2024-12-15 210/week @ 2024-12-22 192/week @ 2024-12-29 387/week @ 2025-01-05 334/week @ 2025-01-12 290/week @ 2025-01-19 439/week @ 2025-01-26 396/week @ 2025-02-02 391/week @ 2025-02-09 410/week @ 2025-02-16 554/week @ 2025-02-23 529/week @ 2025-03-02

1,897 downloads per month
Used in 11 crates (3 directly)

MIT/Apache

11KB
132 lines

fsum is the Rust library by Piotr Beling to calculate accurate sum of floats.

Example

use fsum::FSum;

assert_eq!(FSum::new().add(1e100).add(1.0).add(-1e100).value(), 1.0);
assert_eq!(FSum::with_all((0..10).map(|_| 0.1)).value(), 1.0);
assert_eq!(FSum::with_all(&[1e100, 1.0, -1e100]).value(), 1.0);

let mut s = FSum::new();
assert_eq!(s.value(), 0.0);
s += 3.0;
assert_eq!(s.value(), 3.0);
s -= 1.0;
assert_eq!(s.value(), 2.0);

Complexity

The complexities of summing n numbers are:

  • time: from O(n) (optimistic) to O(n²) (pessimistic)
  • memory: from O(1) (optimistic) to O(n) (pessimistic)

Usually the complexities are close to optimistic.

References

Calculation code bases on (is mostly copied from) sum method of test::stats::Stats implementation for f64 (which probably reimplements math.fsum from Python's library) and source of CPython. See also:

The method sacrifices performance at the altar of accuracy Depends on IEEE-754 arithmetic guarantees. See proof of the correctness in Adaptive Precision Floating-Point Arithmetic and Fast Robust Geometric Predicates

No runtime deps