Hermite normal form

The Hermite normal form is the canonical basis of a lattice in \mathbb{Z}^n given by generators, and the integer analogue of the reduced row echelon form. Two integer matrices have the same Hermite normal form exactly when their rows generate the same lattice. FLINT computes it by several algorithms, from classical elimination with gcds to a randomised algorithm of Pernet and Stein that reduces most of the work to linear system solving; fmpz_mat_hnf chooses among them.

Definition#

An m \times n integer matrix H is in (row-style) Hermite normal form if

• it is in row echelon form, with its r nonzero rows first: the first nonzero entry (pivot) of row i is in column j_i, with j_0 < j_1 < \cdots < j_{r-1};
• the pivots are positive;
• the entries above each pivot are reduced: 0 \le H_{k j_i} < H_{i j_i} for k < i.

For every integer matrix A there is a unique H in Hermite normal form with H = UA for some unimodular U (\det U = \pm 1). Its nonzero rows are a basis of the lattice spanned by the rows of A, the product of its pivots is the determinant (covolume) of that lattice when A has full column rank, and the pivot columns are the column rank profile of A. fmpz_mat_is_in_hnf checks the definition.

Column-style Hermite normal forms, H = AV with the columns spanning the lattice, are obtained by transposing: the column form of A is the transpose of fmpz_mat_hnf of A^T.

Usage#

• H must have the dimensions of A and may be A.
• fmpz_mat_hnf_minors and the modular functions require A to have full column rank (in particular m \ge n), which is not checked; the other functions accept any matrix.
• fmpz_mat_hnf_modular(H, A, D) requires D to be a positive multiple of the determinant of the lattice, for example the absolute value of a nonzero n \times n minor of A when A has full column rank, or |\det A| for a nonsingular square A. fmpz_mat_hnf_modular_eldiv(A, D) transforms A in place and requires D to be a positive multiple of the largest elementary divisor of A.
• fmpz_mat_hnf_transform(H, U, A) also sets the m \times m unimodular U with UA = H; U is not unique when the rank is less than m.

Choice of algorithm#

fmpz_mat_hnf uses fmpz_mat_hnf_classical when the number of rows m is below a cutoff that decreases with the bit size b of the largest entry:

Otherwise it calls fmpz_mat_hnf_pernet_stein. That function in turn uses fmpz_mat_hnf_minors for matrices of full column rank with fewer than about 35 to 52 columns (depending on b), and falls back to fmpz_mat_hnf_minors or fmpz_mat_hnf_classical when the rank is too small for its decomposition.

Algorithms#

• fmpz_mat_hnf_classical eliminates column by column with the entry of smallest absolute value as pivot, by integer division with remainder, until the pivot divides the other entries of its column [Coh1996, Algorithm 2.4.4]. Entries can grow exponentially in the worst case.
• fmpz_mat_hnf_xgcd combines pairs of rows with the cofactors of an extended gcd, which clears an entry in one step [Coh1996, Algorithm 2.4.5].
• fmpz_mat_hnf_minors is the algorithm of Kannan and Bachem [KanBac1979]: it brings the leading k \times k principal submatrix into Hermite normal form for k = 1, 2, \ldots, reducing the entries above the diagonal at each step, which keeps all entries polynomially bounded. It requires full column rank.
• fmpz_mat_hnf_modular is the algorithm of Domich, Kannan and Trotter [DomKanTro1987; Coh1996, Algorithm 2.4.8]: since D\mathbb{Z}^n is contained in the lattice, all computations can be done modulo D, which bounds the entries by D.
• fmpz_mat_hnf_modular_eldiv computes the strong echelon form of A modulo D and puts D on zero diagonal entries; the lattice contains D\mathbb{Z}^n when D is a multiple of the largest elementary divisor [FieHof2014].
• fmpz_mat_hnf_pernet_stein implements [PernetStein2010]. It chooses r - 1 rows forming, together with one of two further rows, a nonsingular submatrix modulo a random prime, computes the two determinants obtained with each of the two further rows by solving one linear system over \mathbb{Q}, and combines the two rows with the cofactors of the gcd g of these determinants. The Hermite normal form of the resulting (r - 1) \times (r - 1) block is computed modulo g, which is small for random matrices; then the remaining columns are added by solving a linear system, and the remaining rows are added one at a time by an incremental update. A random prime giving wrong pivots is detected and the computation restarted.

See Hermite and Smith normal forms for the mathematics and the complexities.

Transformation matrix#

fmpz_mat_hnf_transform uses fmpz_mat_hnf_minors_transform when m \ge n and A has full column rank modulo a random prime; that function performs the Kannan–Bachem elimination on A and applies the same operations to the identity matrix. Otherwise it computes the Hermite normal form of the m \times (n + m) matrix (A \mid I_m), which has full row rank, and reads off H from the first n columns and U from the last m.

Example#

The lattice spanned by three vectors in \mathbb{Z}^3, its Hermite normal form computed by three algorithms, the determinant of the lattice as the product of the pivots, and a transformation matrix:

#include <flint/flint.h>
#include <flint/fmpz.h>
#include <flint/fmpz_mat.h>

int main(void)
{
    fmpz_mat_t A, H, H2, U, UA;
    fmpz_t d;
    slong i, j;
    slong rows[3][3] = { { 4, 6, 10 }, { 2, -2, 8 }, { 6, 14, -6 } };

    fmpz_init(d);
    fmpz_mat_init(A, 3, 3);
    fmpz_mat_init(H, 3, 3);
    fmpz_mat_init(H2, 3, 3);
    fmpz_mat_init(U, 3, 3);
    fmpz_mat_init(UA, 3, 3);

    for (i = 0; i < 3; i++)
        for (j = 0; j < 3; j++)
            fmpz_set_si(fmpz_mat_entry(A, i, j), rows[i][j]);

    fmpz_mat_hnf(H, A);
    fmpz_mat_print_pretty(H);
    flint_printf("\nin HNF: %d\n", fmpz_mat_is_in_hnf(H));

    fmpz_mat_hnf_classical(H2, A);
    flint_printf("classical agrees: %d\n", fmpz_mat_equal(H, H2));

    /* the lattice contains |det A| Z^3 */
    fmpz_mat_det(d, A);
    fmpz_abs(d, d);
    flint_printf("|det A| = %{fmpz}\n", d);
    fmpz_mat_hnf_modular(H2, A, d);
    flint_printf("modular agrees: %d\n", fmpz_mat_equal(H, H2));

    fmpz_mat_hnf_transform(H2, U, A);
    fmpz_mat_mul(UA, U, A);
    flint_printf("U A = H: %d\n", fmpz_mat_equal(UA, H));
    fmpz_mat_det(d, U);
    flint_printf("det U = %{fmpz}\n", d);

    fmpz_mat_clear(A);
    fmpz_mat_clear(H);
    fmpz_mat_clear(H2);
    fmpz_mat_clear(U);
    fmpz_mat_clear(UA);
    fmpz_clear(d);
    return 0;
}

[[2 8 2]
[0 10 12]
[0 0 18]
]
in HNF: 1
classical agrees: 1
|det A| = 360
modular agrees: 1
U A = H: 1
det U = 1

The product of the pivots, 2 \cdot 10 \cdot 18 = 360, is the index of the lattice in \mathbb{Z}^3, which equals |\det A|.