acb_inv: tighter error bounds (in particular, ensure nonzero input -> finite output)

This commit is contained in:
Fredrik Johansson 2017-02-18 17:27:17 +01:00
parent 109afaab0f
commit f0570449f9
3 changed files with 320 additions and 25 deletions

View file

@ -79,7 +79,7 @@ acb_div(acb_t z, const acb_t x, const acb_t y, slong prec)
}
else
{
if (prec > 256 && acb_bits(y) <= prec / 2)
if (prec > 256 && acb_bits(y) <= prec / 2 && acb_is_exact(y))
{
arb_t t, u, v;

193
acb/inv.c
View file

@ -1,5 +1,5 @@
/*
Copyright (C) 2013 Fredrik Johansson
Copyright (C) 2017 Fredrik Johansson
This file is part of Arb.
@ -11,44 +11,189 @@
#include "acb.h"
void
acb_inv(acb_t z, const acb_t x, slong prec)
static void
_arb_arf_div_rounded_den(arb_t res, const arf_t x, const arf_t y, int yinexact, slong prec)
{
#define a acb_realref(x)
#define b acb_imagref(x)
#define c acb_realref(z)
#define d acb_imagref(z)
int inexact = arf_div(arb_midref(res), x, y, prec, ARB_RND);
if (arb_is_zero(b))
if (yinexact && !arf_is_special(arb_midref(res)))
arf_mag_set_ulp(arb_radref(res), arb_midref(res), prec - 1);
else if (inexact)
arf_mag_set_ulp(arb_radref(res), arb_midref(res), prec);
else
mag_zero(arb_radref(res));
}
static void
_arb_arf_div_rounded_den_add_err(arb_t res, const arf_t x, const arf_t y, int yinexact, slong prec)
{
int inexact = arf_div(arb_midref(res), x, y, prec, ARB_RND);
if (yinexact && !arf_is_special(arb_midref(res)))
arf_mag_add_ulp(arb_radref(res), arb_radref(res), arb_midref(res), prec - 1);
else if (inexact)
arf_mag_add_ulp(arb_radref(res), arb_radref(res), arb_midref(res), prec);
}
void
acb_inv(acb_t res, const acb_t z, slong prec)
{
mag_t am, bm;
slong hprec;
#define a arb_midref(acb_realref(z))
#define b arb_midref(acb_imagref(z))
#define x arb_radref(acb_realref(z))
#define y arb_radref(acb_imagref(z))
/* choose precision for the floating-point approximation of a^2+b^2 so
that the double rounding result in less than
2 ulp error; also use at least MAG_BITS bits since the
value will be recycled for error bounds */
hprec = FLINT_MAX(prec + 3, MAG_BITS);
if (arb_is_zero(acb_imagref(z)))
{
arb_inv(c, a, prec);
arb_zero(d);
arb_inv(acb_realref(res), acb_realref(z), prec);
arb_zero(acb_imagref(res));
return;
}
else if (arb_is_zero(a))
if (arb_is_zero(acb_realref(z)))
{
arb_inv(d, b, prec);
arb_neg(d, d);
arb_zero(c);
arb_inv(acb_imagref(res), acb_imagref(z), prec);
arb_neg(acb_imagref(res), acb_imagref(res));
arb_zero(acb_realref(res));
return;
}
if (!acb_is_finite(z))
{
acb_indeterminate(res);
return;
}
if (mag_is_zero(x) && mag_is_zero(y))
{
int inexact;
arf_t a2b2;
arf_init(a2b2);
inexact = arf_sosq(a2b2, a, b, hprec, ARF_RND_DOWN);
if (arf_is_special(a2b2))
{
acb_indeterminate(res);
}
else
{
_arb_arf_div_rounded_den(acb_realref(res), a, a2b2, inexact, prec);
_arb_arf_div_rounded_den(acb_imagref(res), b, a2b2, inexact, prec);
arf_neg(arb_midref(acb_imagref(res)), arb_midref(acb_imagref(res)));
}
arf_clear(a2b2);
return;
}
mag_init(am);
mag_init(bm);
/* first bound |a|-x, |b|-y */
arb_get_mag_lower(am, acb_realref(z));
arb_get_mag_lower(bm, acb_imagref(z));
if ((mag_is_zero(am) && mag_is_zero(bm)))
{
acb_indeterminate(res);
}
else
{
arb_t t;
arb_init(t);
/*
The propagated error in the real part is given exactly by
arb_mul(t, a, a, prec);
arb_addmul(t, b, b, prec);
(a+x')/((a+x')^2+(b+y'))^2 - a/(a^2+b^2) = P / Q,
arb_div(c, a, t, prec);
arb_div(d, b, t, prec);
P = [(b^2-a^2) x' - a (x'^2+y'^2 + 2y'b)]
Q = [(a^2+b^2)((a+x')^2+(b+y')^2)]
arb_neg(d, d);
where |x'| <= x and |y'| <= y, and analogously for the imaginary part.
*/
mag_t t, u, v, w;
arf_t a2b2;
int inexact;
arb_clear(t);
mag_init(t);
mag_init(u);
mag_init(v);
mag_init(w);
arf_init(a2b2);
inexact = arf_sosq(a2b2, a, b, hprec, ARF_RND_DOWN);
/* compute denominator */
/* t = (|a|-x)^2 + (|b|-x)^2 (lower bound) */
mag_mul_lower(t, am, am);
mag_mul_lower(u, bm, bm);
mag_add_lower(t, t, u);
/* u = a^2 + b^2 (lower bound) */
arf_get_mag_lower(u, a2b2);
/* t = ((|a|-x)^2 + (|b|-x)^2)(a^2 + b^2) (lower bound) */
mag_mul_lower(t, t, u);
/* compute numerator */
/* real: |a^2-b^2| x + |a| ((x^2 + y^2) + 2 |b| y)) */
/* imag: |a^2-b^2| y + |b| ((x^2 + y^2) + 2 |a| x)) */
/* am, bm = upper bounds for a, b */
arf_get_mag(am, a);
arf_get_mag(bm, b);
/* v = x^2 + y^2 */
mag_mul(v, x, x);
mag_addmul(v, y, y);
/* u = |a| ((x^2 + y^2) + 2 |b| y) */
mag_mul_2exp_si(u, bm, 1);
mag_mul(u, u, y);
mag_add(u, u, v);
mag_mul(u, u, am);
/* v = |b| ((x^2 + y^2) + 2 |a| x) */
mag_mul_2exp_si(w, am, 1);
mag_addmul(v, w, x);
mag_mul(v, v, bm);
/* w = |b^2 - a^2| (upper bound) */
if (arf_cmpabs(a, b) >= 0)
mag_mul(w, am, am);
else
mag_mul(w, bm, bm);
mag_addmul(u, w, x);
mag_addmul(v, w, y);
mag_div(arb_radref(acb_realref(res)), u, t);
mag_div(arb_radref(acb_imagref(res)), v, t);
_arb_arf_div_rounded_den_add_err(acb_realref(res), a, a2b2, inexact, prec);
_arb_arf_div_rounded_den_add_err(acb_imagref(res), b, a2b2, inexact, prec);
arf_neg(arb_midref(acb_imagref(res)), arb_midref(acb_imagref(res)));
mag_clear(t);
mag_clear(u);
mag_clear(v);
mag_clear(w);
arf_clear(a2b2);
}
mag_clear(am);
mag_clear(bm);
#undef a
#undef b
#undef c
#undef d
#undef x
#undef y
}

150
acb/test/t-inv.c Normal file
View file

@ -0,0 +1,150 @@
/*
Copyright (C) 2012 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.h"
static void
acb_inv_naive(acb_t z, const acb_t x, slong prec)
{
#define a acb_realref(x)
#define b acb_imagref(x)
#define c acb_realref(z)
#define d acb_imagref(z)
if (arb_is_zero(b))
{
arb_inv(c, a, prec);
arb_zero(d);
}
else if (arb_is_zero(a))
{
arb_inv(d, b, prec);
arb_neg(d, d);
arb_zero(c);
}
else
{
arb_t t;
arb_init(t);
arb_mul(t, a, a, prec);
arb_addmul(t, b, b, prec);
arb_div(c, a, t, prec);
arb_div(d, b, t, prec);
arb_neg(d, d);
arb_clear(t);
}
#undef a
#undef b
#undef c
#undef d
}
int main()
{
slong iter;
flint_rand_t state;
flint_printf("inv....");
fflush(stdout);
flint_randinit(state);
for (iter = 0; iter < 100000 * arb_test_multiplier(); iter++)
{
acb_t a, b, c, d, e, f;
arf_t t;
slong prec;
acb_init(a);
acb_init(b);
acb_init(c);
acb_init(d);
acb_init(e);
acb_init(f);
arf_init(t);
prec = 2 + n_randint(state, 1000);
acb_randtest_special(a, state, 1 + n_randint(state, 1000), 100);
acb_randtest_special(b, state, 1 + n_randint(state, 1000), 100);
acb_inv(b, a, prec);
acb_inv_naive(c, a, prec);
if (!acb_overlaps(b, c))
{
flint_printf("FAIL: overlap\n\n");
flint_printf("a = "); acb_printd(a, 30); flint_printf("\n\n");
flint_printf("b = "); acb_printd(b, 30); flint_printf("\n\n");
flint_printf("c = "); acb_printd(c, 30); flint_printf("\n\n");
abort();
}
acb_set(c, a);
acb_inv(c, c, prec);
if (!acb_equal(b, c))
{
flint_printf("FAIL: aliasing\n\n");
flint_printf("a = "); acb_printd(a, 30); flint_printf("\n\n");
flint_printf("b = "); acb_printd(b, 30); flint_printf("\n\n");
flint_printf("c = "); acb_printd(c, 30); flint_printf("\n\n");
abort();
}
acb_randtest(a, state, 1 + n_randint(state, 1000), 10);
acb_randtest(b, state, 1 + n_randint(state, 1000), 10);
acb_zero(d);
arf_set_mag(t, arb_radref(acb_realref(a)));
if (n_randint(state, 2))
arf_neg(t, t);
arf_add(arb_midref(acb_realref(d)),
arb_midref(acb_realref(a)), t, ARF_PREC_EXACT, ARF_RND_DOWN);
arf_set_mag(t, arb_radref(acb_imagref(a)));
if (n_randint(state, 2))
arf_neg(t, t);
arf_add(arb_midref(acb_imagref(d)),
arb_midref(acb_imagref(a)), t, ARF_PREC_EXACT, ARF_RND_DOWN);
acb_inv(b, a, 2 + n_randint(state, 1000));
acb_inv(d, d, 2 + n_randint(state, 1000));
if (!acb_overlaps(b, d))
{
flint_printf("FAIL: corner test\n\n");
flint_printf("a = "); acb_printd(a, 30); flint_printf("\n\n");
flint_printf("b = "); acb_printd(b, 30); flint_printf("\n\n");
flint_printf("d = "); acb_printd(d, 30); flint_printf("\n\n");
abort();
}
acb_clear(a);
acb_clear(b);
acb_clear(c);
acb_clear(d);
acb_clear(e);
acb_clear(f);
arf_clear(t);
}
flint_randclear(state);
flint_cleanup();
flint_printf("PASS\n");
return EXIT_SUCCESS;
}