add acb_mat_eig_multiple_rump, handling overlapping eigenvalues

This commit is contained in:
fredrik 2018-12-02 12:20:13 +01:00
parent eecc160028
commit bbf6860121
4 changed files with 365 additions and 0 deletions

View file

@ -423,6 +423,8 @@ int acb_mat_eig_simple_vdhoeven_mourrain(acb_ptr E, acb_mat_t L, acb_mat_t R,
int acb_mat_eig_simple(acb_ptr E, acb_mat_t L, acb_mat_t R,
const acb_mat_t A, acb_srcptr E_approx, const acb_mat_t R_approx, slong prec);
int acb_mat_eig_multiple_rump(acb_ptr E, const acb_mat_t A, acb_srcptr E_approx, const acb_mat_t R_approx, slong prec);
/* Special functions */
void acb_mat_exp_taylor_sum(acb_mat_t S, const acb_mat_t A, slong N, slong prec);

181
acb_mat/eig_multiple_rump.c Normal file
View file

@ -0,0 +1,181 @@
/*
Copyright (C) 2018 Fredrik Johansson
This file is part of Arb.
Arb is free software: you can redistribute it and/or modify it under
the terms of the GNU Lesser General Public License (LGPL) as published
by the Free Software Foundation; either version 2.1 of the License, or
(at your option) any later version. See <http://www.gnu.org/licenses/>.
*/
#include "acb_mat.h"
static void
acb_approx_mag(mag_t res, const acb_t x)
{
mag_t t;
mag_init(t);
arf_get_mag(res, arb_midref(acb_realref(x)));
arf_get_mag(t, arb_midref(acb_imagref(x)));
mag_hypot(res, res, t);
mag_clear(t);
}
static int
close(const acb_t x, const acb_t y, const mag_t eps)
{
arf_t t;
mag_t a, b;
int result;
mag_init(a);
mag_init(b);
arf_init(t);
arf_sub(t, arb_midref(acb_realref(x)), arb_midref(acb_realref(y)), MAG_BITS, ARF_RND_UP);
arf_get_mag(a, t);
arf_sub(t, arb_midref(acb_imagref(x)), arb_midref(acb_imagref(y)), MAG_BITS, ARF_RND_UP);
arf_get_mag(b, t);
mag_hypot(a, a, b);
result = (mag_cmp(a, eps) <= 0);
mag_clear(a);
mag_clear(b);
arf_clear(t);
return result;
}
int
acb_mat_eig_multiple_rump(acb_ptr E, const acb_mat_t A, acb_srcptr E_approx, const acb_mat_t R_approx, slong prec)
{
slong c, i, j, k, n;
acb_mat_t X;
acb_ptr F;
int result;
slong iter;
mag_t escale, eps, tm, um;
slong ** cluster;
slong * cluster_size;
slong num_clusters;
n = acb_mat_nrows(A);
if (n == 0)
return 1;
cluster = flint_malloc(sizeof(slong *) * n);
for (i = 0; i < n; i++)
cluster[i] = flint_malloc(sizeof(slong) * n);
cluster_size = flint_malloc(sizeof(slong) * n);
mag_init(eps);
mag_init(escale);
mag_init(tm);
mag_init(um);
mag_zero(escale);
for (i = 0; i < n; i++)
{
acb_approx_mag(tm, E_approx + i);
mag_max(escale, escale, tm);
}
/* todo: when num_clusters = 1, could use fallback global enclosure */
/* todo: initial clustering could be allowed to be zero */
/* M 2^(-0.75p), M 2^(-0.5p), M 2^(-0.25p), ... */
for (iter = 0; iter < 2; iter++)
{
mag_mul_2exp_si(eps, escale, -prec + (iter + 1) * prec/4);
/* Group the eigenvalue approximations. */
num_clusters = 0;
for (i = 0; i < n; i++)
{
int new_cluster = 1;
for (j = 0; j < num_clusters && new_cluster; j++)
{
if (close(E_approx + i, E_approx + cluster[j][0], eps))
{
cluster[j][cluster_size[j]] = i;
cluster_size[j]++;
new_cluster = 0;
}
}
if (new_cluster)
{
cluster[num_clusters][0] = i;
cluster_size[num_clusters] = 1;
num_clusters++;
}
}
result = 1;
F = _acb_vec_init(num_clusters);
for (c = 0; c < num_clusters && result; c++)
{
k = cluster_size[c];
acb_mat_init(X, n, k);
for (i = 0; i < n; i++)
for (j = 0; j < k; j++)
acb_set(acb_mat_entry(X, i, j), acb_mat_entry(R_approx, i, cluster[c][j]));
acb_mat_eig_enclosure_rump(F + c, NULL, X, A, E_approx + cluster[c][0], X, prec);
if (!acb_is_finite(F + c))
result = 0;
acb_mat_clear(X);
}
for (i = 0; i < num_clusters; i++)
{
for (j = i + 1; j < num_clusters; j++)
{
if (acb_overlaps(F + i, F + j))
result = 0;
}
}
if (result)
{
i = 0;
for (c = 0; c < num_clusters; c++)
{
for (j = 0; j < cluster_size[c]; j++)
{
acb_set(E + i, F + c);
i++;
}
}
}
_acb_vec_clear(F, num_clusters);
if (result)
break;
}
if (!result)
_acb_vec_indeterminate(E, n);
for (i = 0; i < n; i++)
flint_free(cluster[i]);
flint_free(cluster);
flint_free(cluster_size);
mag_clear(eps);
mag_clear(escale);
mag_clear(tm);
mag_clear(um);
return result;
}

View file

@ -0,0 +1,153 @@
/*
Copyright (C) 2018 Fredrik Johansson
This file is part of Arb.
Arb is free software: you can redistribute it and/or modify it under
the terms of the GNU Lesser General Public License (LGPL) as published
by the Free Software Foundation; either version 2.1 of the License, or
(at your option) any later version. See <http://www.gnu.org/licenses/>.
*/
#include "acb_mat.h"
static void
acb_get_mid(acb_t res, const acb_t x)
{
arb_get_mid_arb(acb_realref(res), acb_realref(x));
arb_get_mid_arb(acb_imagref(res), acb_imagref(x));
}
int main()
{
slong iter;
flint_rand_t state;
flint_printf("eig_multiple_rump....");
fflush(stdout);
flint_randinit(state);
for (iter = 0; iter < 2000 * arb_test_multiplier(); iter++)
{
acb_mat_t A, R;
acb_ptr E, F;
acb_t b;
slong i, j, n, prec, count, count2;
int result;
n = n_randint(state, 8);
prec = 2 + n_randint(state, 200);
acb_init(b);
acb_mat_init(A, n, n);
acb_mat_init(R, n, n);
E = _acb_vec_init(n);
F = _acb_vec_init(n);
if (n_randint(state, 10) != 0)
{
for (i = 0; i < n; i++)
acb_randtest(E + i, state, prec, 2);
}
else
{
/* Randomly repeat eigenvalues. */
for (i = 0; i < n; i++)
{
if (i == 0 || n_randint(state, 2))
acb_randtest(E + i, state, prec, 2);
else
acb_set(E + i, E + n_randint(state, i));
}
}
if (n_randint(state, 2))
{
for (i = 0; i < n; i++)
acb_get_mid(E + i, E + i);
}
acb_mat_randtest_eig(A, state, E, prec);
acb_mat_approx_eig_qr(F, NULL, R, A, NULL, 0, prec);
/* Perturb F further. */
if (n_randint(state, 10) == 0)
{
for (i = 0; i < n; i++)
{
acb_randtest(b, state, prec, 1);
acb_mul_2exp_si(b, b, -n_randint(state, prec));
acb_add(F + i, F + i, b, prec);
}
}
/* Perturb R further. */
if (n_randint(state, 10) == 0)
{
j = n_randint(state, n);
for (i = 0; i < n; i++)
{
acb_randtest(b, state, prec, 1);
acb_mul_2exp_si(b, b, -10 - n_randint(state, prec));
acb_add(acb_mat_entry(R, i, j), acb_mat_entry(R, i, j), b, prec);
}
}
result = acb_mat_eig_multiple_rump(F, A, E, R, prec);
if (result)
{
count = 0;
count2 = 0;
for (i = 0; i < n; i++)
{
for (j = 0; j < n; j++)
{
if (j == 0 || !acb_equal(F + j, F + j - 1))
count += acb_contains(F + j, E + i);
}
for (j = 0; j < n; j++)
{
if (j == 0 || !acb_equal(F + j, F + j - 1))
count2 += acb_overlaps(F + j, E + i);
}
}
if (count != n || count2 != n)
{
flint_printf("FAIL: count\n\n");
flint_printf("A = \n"); acb_mat_printd(A, 20); flint_printf("\n\n");
flint_printf("R = \n"); acb_mat_printd(R, 20); flint_printf("\n\n");
flint_printf("count = %wd, count2 = %wd\n\n", count, count2);
flint_printf("E = \n");
for (j = 0; j < n; j++)
{
acb_printd(E + j, 20);
flint_printf("\n");
}
flint_printf("F = \n");
for (j = 0; j < n; j++)
{
acb_printd(F + j, 20);
flint_printf("\n");
}
flint_abort();
}
}
acb_mat_clear(A);
acb_mat_clear(R);
_acb_vec_clear(E, n);
_acb_vec_clear(F, n);
acb_clear(b);
}
flint_randclear(state);
flint_cleanup();
flint_printf("PASS\n");
return EXIT_SUCCESS;
}

View file

@ -692,3 +692,32 @@ than necessary. Manually balancing badly scaled matrices may help.
complexity `O(n^3)`.
The default version currently uses *vdhoeven_mourrain*.
By design, these functions terminate instead of attempting to
compute eigenvalue clusters if some eigenvalues cannot be isolated.
To compute all eigenvalues of a matrix allowing for overlap,
:func:`acb_mat_eig_multiple_rump` may be used as a fallback.
.. function:: int acb_mat_eig_multiple_rump(acb_ptr E, const acb_mat_t A, acb_srcptr E_approx, const acb_mat_t R_approx, slong prec)
Computes all the eigenvalues of the given *n* by *n* matrix *A*.
On success, the output vector *E* contains *n* complex intervals,
each representing one eigenvalue of *A* with the correct
multiplicities in case of overlap.
The output intervals are either disjoint or identical, and
identical intervals are guaranteed to be grouped consecutively.
Each complete run of *k* identical intervals thus represents a cluster of
exactly *k* eigenvalues which could not be separated from each
other at the current precision, but which could be isolated
from the other `n - k` eigenvalues of the matrix.
The user supplies approximations *E_approx* and *R_approx*
of the eigenvalues and the right eigenvectors.
The initial approximations can, for instance, be computed using
:func:`acb_mat_approx_eig_qr`.
No assumptions are made about the structure of *A* or the
quality of the given approximations.
This function groups approximate eigenvalues that are close and
calls :func:`acb_mat_eig_enclosure_rump` repeatedly to validate
each cluster. The complexity is `O(m n^3)` for *m* clusters.