puremp

Pure-Rust, MIT-licensed, arbitrary-precision arithmetic — integers, rationals, MPFR-class floating point, and base-10 decimals, plus derived modular integers, complex numbers, polynomials, matrices, interval & ball arithmetic, p-adic numbers, finite fields, elliptic curves, and exact real algebraic numbers — with no foreign-code dependencies. Usable as a Rust crate, a C library, and a command-line calculator.

Why

A GMP + MPFR-class toolkit that is:

• Pure, safe Rust — no C, no inline assembly, no intrinsics. The only unsafe in the crate is the opt-in C ABI module.
• Clean-room & MIT-licensed — algorithms come from the open literature (Knuth; Brent & Zimmermann's Modern Computer Arithmetic; the HAC), never from GMP/MPFR source. Use it anywhere, including closed-source projects.
• no_std + alloc — runs on bare metal with an allocator; no OS assumptions in the core. Verified on 32-bit thumbv7em-none-eabi in CI.

Quick start (Rust)

[dependencies]
puremp = "0"

use puremp::{Int, Rational};

let big = Int::from_i64(2).pow(100);
assert_eq!(big.to_string(), "1267650600228229401496703205376");

let sum = Rational::new(Int::from_i64(1), Int::from_i64(2))?   // 1/2
    .add(&Rational::new(Int::from_i64(1), Int::from_i64(3))?); // + 1/3
assert_eq!(sum.to_string(), "5/6");
# Ok::<(), puremp::Error>(())

Quick start (CLI)

$ cargo run --bin puremp
puremp> 2 ** 100
1267650600228229401496703205376
puremp> x = 1000
puremp> x * x - 1
999999
puremp> (2**64) * (2**64)
340282366920938463463374607431768211456
puremp> :quit

Supports + - * / % **, parentheses, unary minus, decimal literals, and name = expr variables (/ and % are truncated integer division).

Quick start (C)

Build the static and/or shared library and link against the header in include/puremp.h:

$ cargo rustc --lib --release --features ffi --crate-type staticlib
$ cargo rustc --lib --release --features ffi --crate-type cdylib
$ cc myprog.c -I include target/release/libpuremp.a -lpthread -ldl -lm -o myprog

#include "puremp.h"
#include <stdio.h>

int main(void) {
    PurempInt *two = puremp_int_from_i64(2);
    PurempInt *big = puremp_int_pow(two, 100);
    char *s = puremp_int_to_string(big);
    printf("2^100 = %s\n", s);
    puremp_string_free(s);
    puremp_int_free(big);
    puremp_int_free(two);
    return 0;
}

Feature flags

Beyond the base types, Int/Rational provide a number-theory toolkit — factorize (trial division → Pollard rho → Lenstra ECM → quadratic sieve), sqrt_mod (Tonelli–Shanks), jacobi/legendre, crt, random_prime, factorial/binomial/fibonacci, and continued-fraction approximate — plus ModInt for modular arithmetic.

The exact-algebra layers stack on top of these:

• Poly::factor factors a rational polynomial into irreducible factors over ℚ (Berlekamp–Zassenhaus with van Hoeij LLL recombination).
• Algebraic is a real root of a rational polynomial, with exact + − × ÷, comparison, and sqrt — and Algebraic::from_float recovers one exactly from a floating-point approximation.
• lattice::lll_reduce is an exact LLL lattice reduction; on top of it, find_integer_relation recognizes constants (e.g. is a number a·√2 + b?) and minimal_polynomial recovers a real algebraic number's defining polynomial.
• Poly<T> and Matrix<T> are generic over any Ring — including the context-carrying ModInt (ℤ/nℤ) and GfElement (GF(pᵏ)), so polynomials and matrices over finite fields and modular integers work, not just over Int/Rational. Over a Field, FieldMatrix gives determinant/inverse/solve/rank by Gaussian elimination, FactorOverField factors polynomials over a finite field (Cantor–Zassenhaus), and RingMatrix gives a division-free determinant and characteristic polynomial (Samuelson–Berkowitz) over any commutative ring — including non-fields like ℤ/nℤ (composite n) or Matrix<Poly<Int>>.

For a bare no_std build: --no-default-features (add --features int for the integer types).

Design & provenance

Bottom-up layers, each building only on the ones below: machine-word carry primitives (adc/sbb/mac) → unsigned magnitudes (Nat, home of the hard algorithms) → tagged signed Int → Rational, with the optional Float and the derived types layered on top. Signed integers inline single-limb magnitudes (no heap allocation until a value exceeds 64 bits).

The implementation is clean-room: GMP and MPFR are LGPL and their source is never consulted. Algorithms come from the open literature —

• Knuth, TAOCP Vol. 2 §4.3 (schoolbook arithmetic; Algorithm D for division);
• Brent & Zimmermann, Modern Computer Arithmetic (sub-quadratic multiply/ divide, GCD, base conversion);
• Menezes, van Oorschot & Vanstone, Handbook of Applied Cryptography;
• primary papers: Karatsuba; Toom–Cook; Burnikel–Ziegler; Möller–Granlund; Faddeev–LeVerrier (algebraic numbers); Sturm sequences (real-root isolation); Lenstra–Lenstra–Lovász (LLL); Cantor–Zassenhaus and van Hoeij (polynomial factorization).

Correctness is checked against published values and, in the dev-only test harness, a trusted reference — never a runtime dependency.

Non-goals: constant-time / side-channel resistance across the general API (for constant-time crypto see the sibling purecrypto crate); drop-in GMP/MPFR C-header compatibility (puremp ships its own cleaner C ABI).

Run cargo run --release --example bench for a throughput harness across the core operations and the derived types.

Roadmap

Candidate directions not yet implemented, all specifiable from open literature (so they preserve the clean-room provenance). Ordering is rough interest, not commitment. Brent & Zimmermann's Modern Computer Arithmetic (MCA; freely available drafts) is the umbrella reference for much of this list.

Faster algorithms (existing operations, correct today, just not maximally fast):

• Deeper multi-prime argument reduction — the current exp multi-prime fast path finds the reducing prime-log combination with an f64 Babai nearest-plane, which caps its winning range; a higher-precision Babai over more primes (and an extension to sin/cos) would widen it. Johansson, arXiv:2207.02501.
• AGM-based transcendentals — π and log/exp via the arithmetic–geometric mean (Brent–Salamin / Gauss–Legendre), O(M(n)·log n) in ~2·lg n quadratically converging steps; a large implicit constant means it complements binary splitting, winning only at very high precision. Brent (1976); Borwein & Borwein, Pi and the AGM; MCA §4.8.

Candidate new capabilities (new operations / types):

• Primality proving for arbitrary inputs — n∓1 certificate proofs (Pocklington + BLS n^{1/3}) already prove any number with a sufficiently factorable n∓1; the general case (a large prime whose n∓1 is hard) needs ECPP (Goldwasser–Kilian → Atkin → Morain; heuristic Õ((log N)⁴⁻⁵)) or the deterministic APR-CL.
• Factorization past ~50 digits — SIQS handles balanced semiprimes into the ~50-digit range; a large-prime variation, a sparse GF(2) solver (block Lanczos / Wiedemann in place of dense Gaussian elimination), and ultimately the number-field sieve (GNFS) are the path toward ~80–100+ digits.

License

MIT — see LICENSE.