Greatest common divisors

Over a field $\mathbb{F}_p$, every pair of polynomials has a greatest common divisor, unique up to a constant factor, and the non-underscore functions return the monic one. nmod_poly_gcd computes $G = \gcd(A, B)$, nmod_poly_xgcd also the cofactors $S, T$ with $SA + TB = G$, and nmod_poly_gcdinv only the cofactor $S$ of $A$, which is what is needed for inverses modulo $B$. The modulus must be prime; over $\mathbb{Z}/n\mathbb{Z}$ with composite $n$ the Euclidean algorithm can meet a non-invertible leading coefficient, which throws an exception.

Algorithms#

The Euclidean algorithm replaces $(A, B)$ by $(B, A \bmod B)$ until the remainder is zero. In the typical case each step lowers the degree by one and costs $O(\deg B)$ operations, for $O(\deg A \deg B)$ in total. FLINT's implementation calls _nmod_poly_rem, which has a dedicated routine for quotients of length 2, and for moduli with at least half the bits of a word (but at most $\mathtt{FLINT\_BITS} - 2$ bits) and $\deg B \ge 31$ it runs in Montgomery arithmetic (_nmod_poly_gcd_euclidean_redc_fast), where the long chains of dependent multiplications are cheaper.

The half-GCD algorithm [ThullYap1990] uses the fact that the quotients of the first half of the Euclidean algorithm depend only on the top half of the coefficients. It computes them recursively from the top coefficients and applies them to the full polynomials as a $2 \times 2$ matrix of polynomials, so that the whole remainder sequence costs $O(M(n) \log n)$ operations. nmod_poly_hgcd exposes one such step: it returns the matrix $M$ and the two consecutive remainders $A, B$ of the sequence that straddle half the degree of the input.

nmod_poly_gcd uses the half-GCD algorithm for $\mathit{lenB} \ge$ nmod_poly_gcd_hgcd_cutoff(mod), between 470 and 2199 depending on the bit size of $n$, and nmod_poly_xgcd for $\mathit{lenB} \ge$ nmod_poly_xgcd_hgcd_cutoff(mod), between 72 and 388; the cofactors make the Euclidean algorithm more expensive, so the extended version switches earlier.

Examples#

Two random polynomials of degree 2500 modulo a 50-bit prime, multiplied by a common factor of degree 500; both algorithms find the factor, made monic, and the cofactors of nmod_poly_xgcd satisfy Bézout's identity:

#include <flint/flint.h>
#include <flint/ulong_extras.h>
#include <flint/nmod_poly.h>

int main(void)
{
    flint_rand_t state;
    nmod_poly_t a, b, c, g, h, s, t, u;
    ulong p = n_nextprime(UWORD(1) << 49, 1);

    flint_rand_init(state);
    nmod_poly_init(a, p); nmod_poly_init(b, p); nmod_poly_init(c, p);
    nmod_poly_init(g, p); nmod_poly_init(h, p);
    nmod_poly_init(s, p); nmod_poly_init(t, p); nmod_poly_init(u, p);

    nmod_poly_rand_monic(c, state, 501);
    nmod_poly_rand(a, state, 2501);
    nmod_poly_rand(b, state, 2501);
    nmod_poly_mul(a, a, c);
    nmod_poly_mul(b, b, c);

    nmod_poly_gcd(g, a, b);               /* half-GCD at this length */
    nmod_poly_gcd_euclidean(h, a, b);
    flint_printf("deg gcd = %wd\n", nmod_poly_degree(g));
    flint_printf("hgcd == euclidean: %d\n", nmod_poly_equal(g, h));
    flint_printf("gcd == c: %d\n", nmod_poly_equal(g, c));

    nmod_poly_xgcd(h, s, t, a, b);
    nmod_poly_mul(s, s, a);
    nmod_poly_mul(t, t, b);
    nmod_poly_add(u, s, t);
    flint_printf("s a + t b == g: %d\n", nmod_poly_equal(u, h));

    nmod_poly_clear(a); nmod_poly_clear(b); nmod_poly_clear(c);
    nmod_poly_clear(g); nmod_poly_clear(h);
    nmod_poly_clear(s); nmod_poly_clear(t); nmod_poly_clear(u);
    flint_rand_clear(state);
    return 0;
}

deg gcd = 500
hgcd == euclidean: 1
gcd == c: 1
s a + t b == g: 1

The common factor is found exactly because the random cofactors are coprime, which happens with probability $1 - 1/p$.

nmod_poly_gcdinv returns the inverse of $A$ modulo $B$ together with the GCD, so that one call both tests and inverts. Modulo 7, $x^2 + 1$ is invertible modulo $x^3 + x + 1$ but not modulo $x^3 + x^2 + x + 1 = (x + 1)(x^2 + 1)$:

#include <flint/flint.h>
#include <flint/nmod_poly.h>

int main(void)
{
    nmod_poly_t a, b, g, s, t;

    nmod_poly_init(a, 7);
    nmod_poly_init(b, 7);
    nmod_poly_init(g, 7);
    nmod_poly_init(s, 7);
    nmod_poly_init(t, 7);

    nmod_poly_set_str(a, "3 7  1 0 1");          /* x^2 + 1 */
    nmod_poly_set_str(b, "4 7  1 1 0 1");        /* x^3 + x + 1 */
    nmod_poly_gcdinv(g, s, a, b);
    flint_printf("g = "); nmod_poly_print_pretty(g, "x");
    flint_printf(", s = "); nmod_poly_print_pretty(s, "x");
    nmod_poly_mulmod(t, s, a, b);
    flint_printf(", s a mod b = "); nmod_poly_print_pretty(t, "x");
    flint_printf("\n");

    nmod_poly_set_str(b, "4 7  1 1 1 1");        /* (x + 1)(x^2 + 1) */
    nmod_poly_gcdinv(g, s, a, b);
    flint_printf("g = "); nmod_poly_print_pretty(g, "x");
    flint_printf(", s = "); nmod_poly_print_pretty(s, "x");
    flint_printf("\n");

    nmod_poly_clear(a);
    nmod_poly_clear(b);
    nmod_poly_clear(g);
    nmod_poly_clear(s);
    nmod_poly_clear(t);
    return 0;
}

g = 1, s = 6*x, s a mod b = 1
g = x^2+1, s = 1