void
acb_hypgeom_mag_chi(mag_t chi, ulong n)
{
mag_t p, q;
ulong k;
mag_init(p);
mag_init(q);
if (n % 2 == 0)
{
mag_one(p);
mag_one(q);
}
else
{
/* upper bound for pi/2 */
mag_set_ui_2exp_si(p, 843314857, -28);
mag_one(q);
}
for (k = n; k >= 2; k -= 2)
{
mag_mul_ui(p, p, k);
mag_mul_ui_lower(q, q, k - 1);
}
mag_div(chi, p, q);
mag_clear(p);
mag_clear(q);
}
void
mag_exp_tail(mag_t z, const mag_t x, ulong N)
{
if (N == 0 || mag_is_inf(x))
{
mag_exp(z, x);
}
else if (mag_is_zero(x))
{
mag_zero(z);
}
else
{
mag_t t;
mag_init(t);
mag_set_ui_2exp_si(t, N, -1);
/* bound by geometric series when N >= 2*x <=> N/2 >= x */
if (mag_cmp(t, x) >= 0)
{
/* 2 c^N / N! */
mag_pow_ui(t, x, N);
mag_rfac_ui(z, N);
mag_mul(z, z, t);
mag_mul_2exp_si(z, z, 1);
}
else
{
mag_exp(z, x);
}
mag_clear(t);
}
}
/* The error for eta(s) is bounded by 3/(3+sqrt(8))^n */
void
mag_borwein_error(mag_t err, slong n)
{
/* upper bound for 1/(3+sqrt(8)) */
mag_set_ui_2exp_si(err, 736899889, -32);
mag_pow_ui(err, err, n);
mag_mul_ui(err, err, 3);
}
void
arb_sinc(arb_t z, const arb_t x, slong prec)
{
mag_t c, r;
mag_init(c);
mag_init(r);
mag_set_ui_2exp_si(c, 5, -1);
arb_get_mag_lower(r, x);
if (mag_cmp(c, r) < 0)
{
/* x is not near the origin */
_arb_sinc_direct(z, x, prec);
}
else if (mag_cmp_2exp_si(arb_radref(x), 1) < 0)
{
/* determine error magnitude using the derivative bound */
if (arb_is_exact(x))
{
mag_zero(c);
}
else
{
_arb_sinc_derivative_bound(r, x);
mag_mul(c, arb_radref(x), r);
}
/* evaluate sinc at the midpoint of x */
if (arf_is_zero(arb_midref(x)))
{
arb_one(z);
}
else
{
arb_get_mid_arb(z, x);
_arb_sinc_direct(z, z, prec);
}
/* add the error */
mag_add(arb_radref(z), arb_radref(z), c);
}
else
{
/* x has a large radius and includes points near the origin */
arf_zero(arb_midref(z));
mag_one(arb_radref(z));
}
mag_clear(c);
mag_clear(r);
}
void
mag_expinv(mag_t res, const mag_t x)
{
if (mag_is_zero(x))
{
mag_one(res);
}
else if (mag_is_inf(x))
{
mag_zero(res);
}
else if (fmpz_sgn(MAG_EXPREF(x)) <= 0)
{
mag_one(res);
}
else if (fmpz_cmp_ui(MAG_EXPREF(x), 2 * MAG_BITS) > 0)
{
fmpz_t t;
fmpz_init(t);
/* If x > 2^60, exp(-x) < 2^(-2^60 / log(2)) */
/* -1/log(2) < -369/256 */
fmpz_set_si(t, -369);
fmpz_mul_2exp(t, t, 2 * MAG_BITS - 8);
mag_one(res);
mag_mul_2exp_fmpz(res, res, t);
fmpz_clear(t);
}
else
{
fmpz_t t;
slong e = MAG_EXP(x);
fmpz_init(t);
fmpz_set_ui(t, MAG_MAN(x));
if (e >= MAG_BITS)
fmpz_mul_2exp(t, t, e - MAG_BITS);
else
fmpz_tdiv_q_2exp(t, t, MAG_BITS - e);
/* upper bound for 1/e */
mag_set_ui_2exp_si(res, 395007543, -30);
mag_pow_fmpz(res, res, t);
fmpz_clear(t);
}
}
int main(int argc, char *argv[])
{
acb_t s, t, a, b;
mag_t tol;
slong prec, goal;
slong N;
ulong k;
int integral, ifrom, ito;
int i, twice, havegoal, havetol;
acb_calc_integrate_opt_t options;
ifrom = ito = -1;
for (i = 1; i < argc; i++)
{
if (!strcmp(argv[i], "-i"))
{
if (!strcmp(argv[i+1], "all"))
{
ifrom = 0;
ito = NUM_INTEGRALS - 1;
}
else
{
ifrom = ito = atol(argv[i+1]);
if (ito < 0 || ito >= NUM_INTEGRALS)
flint_abort();
}
}
}
if (ifrom == -1)
{
flint_printf("Compute integrals using acb_calc_integrate.\n");
flint_printf("Usage: integrals -i n [-prec p] [-tol eps] [-twice] [...]\n\n");
flint_printf("-i n - compute integral n (0 <= n <= %d), or \"-i all\"\n", NUM_INTEGRALS - 1);
flint_printf("-prec p - precision in bits (default p = 64)\n");
flint_printf("-goal p - approximate relative accuracy goal (default p)\n");
flint_printf("-tol eps - approximate absolute error goal (default 2^-p)\n");
flint_printf("-twice - run twice (to see overhead of computing nodes)\n");
flint_printf("-heap - use heap for subinterval queue\n");
flint_printf("-verbose - show information\n");
flint_printf("-verbose2 - show more information\n");
flint_printf("-deg n - use quadrature degree up to n\n");
flint_printf("-eval n - limit number of function evaluations to n\n");
flint_printf("-depth n - limit subinterval queue size to n\n\n");
flint_printf("Implemented integrals:\n");
for (integral = 0; integral < NUM_INTEGRALS; integral++)
flint_printf("I%d = %s\n", integral, descr[integral]);
flint_printf("\n");
return 1;
}
acb_calc_integrate_opt_init(options);
prec = 64;
twice = 0;
goal = 0;
havetol = havegoal = 0;
acb_init(a);
acb_init(b);
acb_init(s);
acb_init(t);
mag_init(tol);
for (i = 1; i < argc; i++)
{
if (!strcmp(argv[i], "-prec"))
{
prec = atol(argv[i+1]);
}
else if (!strcmp(argv[i], "-twice"))
{
twice = 1;
}
else if (!strcmp(argv[i], "-goal"))
{
goal = atol(argv[i+1]);
if (goal < 0)
{
flint_printf("expected goal >= 0\n");
return 1;
}
havegoal = 1;
}
else if (!strcmp(argv[i], "-tol"))
{
arb_t x;
arb_init(x);
arb_set_str(x, argv[i+1], 10);
arb_get_mag(tol, x);
arb_clear(x);
havetol = 1;
}
else if (!strcmp(argv[i], "-deg"))
{
options->deg_limit = atol(argv[i+1]);
}
else if (!strcmp(argv[i], "-eval"))
{
options->eval_limit = atol(argv[i+1]);
}
else if (!strcmp(argv[i], "-depth"))
{
options->depth_limit = atol(argv[i+1]);
}
else if (!strcmp(argv[i], "-verbose"))
{
options->verbose = 1;
}
else if (!strcmp(argv[i], "-verbose2"))
{
options->verbose = 2;
}
else if (!strcmp(argv[i], "-heap"))
{
options->use_heap = 1;
}
}
if (!havegoal)
goal = prec;
if (!havetol)
mag_set_ui_2exp_si(tol, 1, -prec);
for (integral = ifrom; integral <= ito; integral++)
{
flint_printf("I%d = %s ...\n", integral, descr[integral]);
for (i = 0; i < 1 + twice; i++)
{
TIMEIT_ONCE_START
switch (integral)
{
case 0:
acb_set_d(a, 0);
acb_set_d(b, 100);
acb_calc_integrate(s, f_sin, NULL, a, b, goal, tol, options, prec);
break;
case 1:
acb_set_d(a, 0);
acb_set_d(b, 1);
acb_calc_integrate(s, f_atanderiv, NULL, a, b, goal, tol, options, prec);
acb_mul_2exp_si(s, s, 2);
break;
case 2:
acb_set_d(a, 0);
acb_one(b);
acb_mul_2exp_si(b, b, goal);
acb_calc_integrate(s, f_atanderiv, NULL, a, b, goal, tol, options, prec);
arb_add_error_2exp_si(acb_realref(s), -goal);
acb_mul_2exp_si(s, s, 1);
break;
case 3:
acb_set_d(a, 0);
acb_set_d(b, 1);
acb_calc_integrate(s, f_circle, NULL, a, b, goal, tol, options, prec);
acb_mul_2exp_si(s, s, 2);
break;
case 4:
acb_set_d(a, 0);
acb_set_d(b, 8);
acb_calc_integrate(s, f_rump, NULL, a, b, goal, tol, options, prec);
break;
case 5:
acb_set_d(a, 1);
acb_set_d(b, 101);
acb_calc_integrate(s, f_floor, NULL, a, b, goal, tol, options, prec);
break;
case 6:
acb_set_d(a, 0);
acb_set_d(b, 1);
acb_calc_integrate(s, f_helfgott, NULL, a, b, goal, tol, options, prec);
break;
case 7:
acb_zero(s);
acb_set_d_d(a, -1.0, -1.0);
acb_set_d_d(b, 2.0, -1.0);
acb_calc_integrate(t, f_zeta, NULL, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, 2.0, -1.0);
acb_set_d_d(b, 2.0, 1.0);
acb_calc_integrate(t, f_zeta, NULL, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, 2.0, 1.0);
acb_set_d_d(b, -1.0, 1.0);
acb_calc_integrate(t, f_zeta, NULL, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, -1.0, 1.0);
acb_set_d_d(b, -1.0, -1.0);
acb_calc_integrate(t, f_zeta, NULL, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_const_pi(t, prec);
acb_div(s, s, t, prec);
acb_mul_2exp_si(s, s, -1);
acb_div_onei(s, s);
break;
case 8:
acb_set_d(a, 0);
acb_set_d(b, 1);
acb_calc_integrate(s, f_essing, NULL, a, b, goal, tol, options, prec);
break;
case 9:
acb_set_d(a, 0);
acb_set_d(b, 1);
acb_calc_integrate(s, f_essing2, NULL, a, b, goal, tol, options, prec);
break;
case 10:
acb_set_d(a, 0);
acb_set_d(b, 10000);
acb_calc_integrate(s, f_factorial1000, NULL, a, b, goal, tol, options, prec);
break;
case 11:
acb_set_d_d(a, 1.0, 0.0);
acb_set_d_d(b, 1.0, 1000.0);
acb_calc_integrate(s, f_gamma, NULL, a, b, goal, tol, options, prec);
break;
case 12:
acb_set_d(a, -10.0);
acb_set_d(b, 10.0);
acb_calc_integrate(s, f_sin_plus_small, NULL, a, b, goal, tol, options, prec);
break;
case 13:
acb_set_d(a, -1020.0);
acb_set_d(b, -1010.0);
acb_calc_integrate(s, f_exp, NULL, a, b, goal, tol, options, prec);
break;
case 14:
acb_set_d(a, 0);
acb_set_d(b, ceil(sqrt(goal * 0.693147181) + 1.0));
acb_calc_integrate(s, f_gaussian, NULL, a, b, goal, tol, options, prec);
acb_mul(b, b, b, prec);
acb_neg(b, b);
acb_exp(b, b, prec);
arb_add_error(acb_realref(s), acb_realref(b));
break;
case 15:
acb_set_d(a, 0.0);
acb_set_d(b, 1.0);
acb_calc_integrate(s, f_spike, NULL, a, b, goal, tol, options, prec);
break;
case 16:
acb_set_d(a, 0.0);
acb_set_d(b, 8.0);
acb_calc_integrate(s, f_monster, NULL, a, b, goal, tol, options, prec);
break;
case 17:
acb_set_d(a, 0);
acb_set_d(b, ceil(goal * 0.693147181 + 1.0));
acb_calc_integrate(s, f_sech, NULL, a, b, goal, tol, options, prec);
acb_neg(b, b);
acb_exp(b, b, prec);
acb_mul_2exp_si(b, b, 1);
arb_add_error(acb_realref(s), acb_realref(b));
break;
case 18:
acb_set_d(a, 0);
acb_set_d(b, ceil(goal * 0.693147181 / 3.0 + 2.0));
acb_calc_integrate(s, f_sech3, NULL, a, b, goal, tol, options, prec);
acb_neg(b, b);
acb_mul_ui(b, b, 3, prec);
acb_exp(b, b, prec);
acb_mul_2exp_si(b, b, 3);
acb_div_ui(b, b, 3, prec);
arb_add_error(acb_realref(s), acb_realref(b));
break;
case 19:
if (goal < 0)
abort();
/* error bound 2^-N (1+N) when truncated at 2^-N */
N = goal + FLINT_BIT_COUNT(goal);
acb_one(a);
acb_mul_2exp_si(a, a, -N);
acb_one(b);
acb_calc_integrate(s, f_log_div1p, NULL, a, b, goal, tol, options, prec);
acb_set_ui(b, N + 1);
acb_mul_2exp_si(b, b, -N);
arb_add_error(acb_realref(s), acb_realref(b));
break;
case 20:
if (goal < 0)
abort();
/* error bound (N+1) exp(-N) when truncated at N */
N = goal + FLINT_BIT_COUNT(goal);
acb_zero(a);
acb_set_ui(b, N);
acb_calc_integrate(s, f_log_div1p_transformed, NULL, a, b, goal, tol, options, prec);
acb_neg(b, b);
acb_exp(b, b, prec);
acb_mul_ui(b, b, N + 1, prec);
arb_add_error(acb_realref(s), acb_realref(b));
break;
case 21:
acb_zero(s);
N = 10;
acb_set_d_d(a, 0.5, -0.5);
acb_set_d_d(b, 0.5, 0.5);
acb_calc_integrate(t, f_elliptic_p_laurent_n, &N, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, 0.5, 0.5);
acb_set_d_d(b, -0.5, 0.5);
acb_calc_integrate(t, f_elliptic_p_laurent_n, &N, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, -0.5, 0.5);
acb_set_d_d(b, -0.5, -0.5);
acb_calc_integrate(t, f_elliptic_p_laurent_n, &N, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, -0.5, -0.5);
acb_set_d_d(b, 0.5, -0.5);
acb_calc_integrate(t, f_elliptic_p_laurent_n, &N, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_const_pi(t, prec);
acb_div(s, s, t, prec);
acb_mul_2exp_si(s, s, -1);
acb_div_onei(s, s);
break;
case 22:
acb_zero(s);
N = 1000;
acb_set_d_d(a, 100.0, 0.0);
acb_set_d_d(b, 100.0, N);
acb_calc_integrate(t, f_zeta_frac, NULL, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_set_d_d(a, 100, N);
acb_set_d_d(b, 0.5, N);
acb_calc_integrate(t, f_zeta_frac, NULL, a, b, goal, tol, options, prec);
acb_add(s, s, t, prec);
acb_div_onei(s, s);
arb_zero(acb_imagref(s));
acb_set_ui(t, N);
acb_dirichlet_hardy_theta(t, t, NULL, NULL, 1, prec);
acb_add(s, s, t, prec);
acb_const_pi(t, prec);
acb_div(s, s, t, prec);
acb_add_ui(s, s, 1, prec);
break;
case 23:
acb_set_d(a, 0.0);
acb_set_d(b, 1000.0);
acb_calc_integrate(s, f_lambertw, NULL, a, b, goal, tol, options, prec);
break;
case 24:
acb_set_d(a, 0.0);
acb_const_pi(b, prec);
acb_calc_integrate(s, f_max_sin_cos, NULL, a, b, goal, tol, options, prec);
break;
case 25:
acb_set_si(a, -1);
acb_set_si(b, 1);
acb_calc_integrate(s, f_erf_bent, NULL, a, b, goal, tol, options, prec);
break;
case 26:
acb_set_si(a, -10);
acb_set_si(b, 10);
acb_calc_integrate(s, f_airy_ai, NULL, a, b, goal, tol, options, prec);
break;
case 27:
acb_set_si(a, 0);
acb_set_si(b, 10);
acb_calc_integrate(s, f_horror, NULL, a, b, goal, tol, options, prec);
break;
case 28:
acb_set_d_d(a, -1, -1);
acb_set_d_d(b, -1, 1);
acb_calc_integrate(s, f_sqrt, NULL, a, b, goal, tol, options, prec);
break;
case 29:
acb_set_d(a, 0);
acb_set_d(b, ceil(sqrt(goal * 0.693147181) + 1.0));
acb_calc_integrate(s, f_gaussian_twist, NULL, a, b, goal, tol, options, prec);
acb_mul(b, b, b, prec);
acb_neg(b, b);
acb_exp(b, b, prec);
arb_add_error(acb_realref(s), acb_realref(b));
arb_add_error(acb_imagref(s), acb_realref(b));
break;
case 30:
acb_set_d(a, 0);
acb_set_d(b, ceil(goal * 0.693147181 + 1.0));
acb_calc_integrate(s, f_exp_airy, NULL, a, b, goal, tol, options, prec);
acb_neg(b, b);
acb_exp(b, b, prec);
acb_mul_2exp_si(b, b, 1);
arb_add_error(acb_realref(s), acb_realref(b));
break;
case 31:
acb_zero(a);
acb_const_pi(b, prec);
acb_calc_integrate(s, f_sin_cos_frac, NULL, a, b, goal, tol, options, prec);
break;
case 32:
acb_zero(a);
acb_set_ui(b, 3);
acb_calc_integrate(s, f_sin_near_essing, NULL, a, b, goal, tol, options, prec);
break;
case 33:
acb_zero(a);
acb_zero(b);
k = 3;
scaled_bessel_select_N(acb_realref(b), k, prec);
acb_calc_integrate(s, f_scaled_bessel, &k, a, b, goal, tol, options, prec);
scaled_bessel_tail_bound(acb_realref(a), k, acb_realref(b), prec);
arb_add_error(acb_realref(s), acb_realref(a));
break;
case 34:
acb_zero(a);
acb_zero(b);
k = 15;
scaled_bessel_select_N(acb_realref(b), k, prec);
acb_calc_integrate(s, f_scaled_bessel, &k, a, b, goal, tol, options, prec);
scaled_bessel_tail_bound(acb_realref(a), k, acb_realref(b), prec);
arb_add_error(acb_realref(s), acb_realref(a));
break;
case 35:
acb_set_d_d(a, -1, -1);
acb_set_d_d(b, -1, 1);
acb_calc_integrate(s, f_rsqrt, NULL, a, b, goal, tol, options, prec);
break;
default:
abort();
}
TIMEIT_ONCE_STOP
}
flint_printf("I%d = ", integral);
acb_printn(s, 3.333 * prec, 0);
flint_printf("\n\n");
}
acb_clear(a);
acb_clear(b);
acb_clear(s);
acb_clear(t);
mag_clear(tol);
flint_cleanup();
return 0;
}
/* todo: use log(1-z) when this is better? would also need to
adjust strategy in the main function */
void
acb_hypgeom_dilog_bernoulli(acb_t res, const acb_t z, slong prec)
{
acb_t s, w, w2;
slong n, k;
fmpz_t c, d;
mag_t m, err;
double lm;
int real;
acb_init(s);
acb_init(w);
acb_init(w2);
fmpz_init(c);
fmpz_init(d);
mag_init(m);
mag_init(err);
real = 0;
if (acb_is_real(z))
{
arb_sub_ui(acb_realref(w), acb_realref(z), 1, 30);
real = arb_is_nonpositive(acb_realref(w));
}
acb_log(w, z, prec);
acb_get_mag(m, w);
/* for k >= 4, the terms are bounded by (|w| / (2 pi))^k */
mag_set_ui_2exp_si(err, 2670177, -24); /* upper bound for 1/(2pi) */
mag_mul(err, err, m);
lm = mag_get_d_log2_approx(err);
if (lm < -0.25)
{
n = prec / (-lm) + 1;
n = FLINT_MAX(n, 4);
mag_geom_series(err, err, n);
BERNOULLI_ENSURE_CACHED(n)
acb_mul(w2, w, w, prec);
for (k = n - (n % 2 == 0); k >= 3; k -= 2)
{
fmpz_mul_ui(c, fmpq_denref(bernoulli_cache + k - 1), k - 1);
fmpz_mul_ui(d, c, (k + 1) * (k + 2));
acb_mul(s, s, w2, prec);
acb_mul_fmpz(s, s, c, prec);
fmpz_mul_ui(c, fmpq_numref(bernoulli_cache + k - 1), (k + 1) * (k + 2));
acb_sub_fmpz(s, s, c, prec);
acb_div_fmpz(s, s, d, prec);
}
acb_mul(s, s, w, prec);
acb_mul_2exp_si(s, s, 1);
acb_sub_ui(s, s, 3, prec);
acb_mul(s, s, w2, prec);
acb_mul_2exp_si(s, s, -1);
acb_const_pi(w2, prec);
acb_addmul(s, w2, w2, prec);
acb_div_ui(s, s, 6, prec);
acb_neg(w2, w);
acb_log(w2, w2, prec);
acb_submul(s, w2, w, prec);
acb_add(res, s, w, prec);
acb_add_error_mag(res, err);
if (real)
arb_zero(acb_imagref(res));
}
else
{
acb_indeterminate(res);
}
acb_clear(s);
acb_clear(w);
acb_clear(w2);
fmpz_clear(c);
fmpz_clear(d);
mag_clear(m);
mag_clear(err);
}
slong
hypgeom_bound(mag_t error, int r,
slong A, slong B, slong K, const mag_t TK, const mag_t z, slong tol_2exp)
{
mag_t Tn, t, u, one, tol, num, den;
slong n, m;
mag_init(Tn);
mag_init(t);
mag_init(u);
mag_init(one);
mag_init(tol);
mag_init(num);
mag_init(den);
mag_one(one);
mag_set_ui_2exp_si(tol, UWORD(1), -tol_2exp);
/* approximate number of needed terms */
n = hypgeom_estimate_terms(z, r, tol_2exp);
/* required for 1 + O(1/k) part to be decreasing */
n = FLINT_MAX(n, K + 1);
/* required for z^k / (k!)^r to be decreasing */
m = hypgeom_root_bound(z, r);
n = FLINT_MAX(n, m);
/* We now have |R(k)| <= G(k) where G(k) is monotonically decreasing,
and can bound the tail using a geometric series as soon
as soon as G(k) < 1. */
/* bound T(n-1) */
hypgeom_term_bound(Tn, TK, K, A, B, r, z, n-1);
while (1)
{
/* bound R(n) */
mag_mul_ui(num, z, n);
mag_mul_ui(num, num, n - B);
mag_set_ui_lower(den, n - A);
mag_mul_ui_lower(den, den, n - 2*B);
if (r != 0)
{
mag_set_ui_lower(u, n);
mag_pow_ui_lower(u, u, r);
mag_mul_lower(den, den, u);
}
mag_div(t, num, den);
/* multiply bound for T(n-1) by bound for R(n) to bound T(n) */
mag_mul(Tn, Tn, t);
/* geometric series termination check */
/* u = max(1-t, 0), rounding down [lower bound] */
mag_sub_lower(u, one, t);
if (!mag_is_zero(u))
{
mag_div(u, Tn, u);
if (mag_cmp(u, tol) < 0)
{
mag_set(error, u);
break;
}
}
/* move on to next term */
n++;
}
mag_clear(Tn);
mag_clear(t);
mag_clear(u);
mag_clear(one);
mag_clear(tol);
mag_clear(num);
mag_clear(den);
return n;
}
int main()
{
slong iter;
flint_rand_t state;
flint_printf("get_mpn_fixed_mod_pi4....");
fflush(stdout);
flint_randinit(state);
/* _flint_rand_init_gmp(state); */
for (iter = 0; iter < 100000 * arb_test_multiplier(); iter++)
{
arf_t x;
int octant;
fmpz_t q;
mp_ptr w;
arb_t wb, t, u;
mp_size_t wn;
slong prec, prec2;
int success;
mp_limb_t error;
prec = 2 + n_randint(state, 10000);
wn = 1 + n_randint(state, 200);
prec2 = FLINT_MAX(prec, wn * FLINT_BITS) + 100;
arf_init(x);
arb_init(wb);
arb_init(t);
arb_init(u);
fmpz_init(q);
w = flint_malloc(sizeof(mp_limb_t) * wn);
arf_randtest(x, state, prec, 14);
/* this should generate numbers close to multiples of pi/4 */
if (n_randint(state, 4) == 0)
{
arb_const_pi(t, prec);
arb_mul_2exp_si(t, t, -2);
fmpz_randtest(q, state, 200);
arb_mul_fmpz(t, t, q, prec);
arf_add(x, x, arb_midref(t), prec, ARF_RND_DOWN);
}
arf_abs(x, x);
success = _arb_get_mpn_fixed_mod_pi4(w, q, &octant, &error, x, wn);
if (success)
{
/* could round differently */
if (fmpz_fdiv_ui(q, 8) != octant)
{
flint_printf("bad octant\n");
abort();
}
_arf_set_mpn_fixed(arb_midref(wb), w, wn, wn, 0, FLINT_BITS * wn, ARB_RND);
mag_set_ui_2exp_si(arb_radref(wb), error, -FLINT_BITS * wn);
arb_const_pi(u, prec2);
arb_mul_2exp_si(u, u, -2);
arb_set(t, wb);
if (octant % 2 == 1)
arb_sub(t, u, t, prec2);
arb_addmul_fmpz(t, u, q, prec2);
if (!arb_contains_arf(t, x))
{
flint_printf("FAIL (containment)\n");
flint_printf("x = "); arf_printd(x, 50); flint_printf("\n\n");
flint_printf("q = "); fmpz_print(q); flint_printf("\n\n");
flint_printf("w = "); arb_printd(wb, 50); flint_printf("\n\n");
flint_printf("t = "); arb_printd(t, 50); flint_printf("\n\n");
abort();
}
arb_const_pi(t, prec2);
arb_mul_2exp_si(t, t, -2);
if (arf_sgn(arb_midref(wb)) < 0 ||
arf_cmp(arb_midref(wb), arb_midref(t)) >= 0)
{
flint_printf("FAIL (expected 0 <= w < pi/4)\n");
flint_printf("x = "); arf_printd(x, 50); flint_printf("\n\n");
flint_printf("w = "); arb_printd(wb, 50); flint_printf("\n\n");
abort();
}
}
flint_free(w);
fmpz_clear(q);
arf_clear(x);
arb_clear(wb);
arb_clear(t);
arb_clear(u);
}
flint_randclear(state);
flint_cleanup();
flint_printf("PASS\n");
return EXIT_SUCCESS;
}
void
arb_atan_arf(arb_t z, const arf_t x, slong prec)
{
if (arf_is_special(x))
{
if (arf_is_zero(x))
{
arb_zero(z);
}
else if (arf_is_pos_inf(x))
{
arb_const_pi(z, prec);
arb_mul_2exp_si(z, z, -1);
}
else if (arf_is_neg_inf(x))
{
arb_const_pi(z, prec);
arb_mul_2exp_si(z, z, -1);
arb_neg(z, z);
}
else
{
arb_indeterminate(z);
}
}
else if (COEFF_IS_MPZ(*ARF_EXPREF(x)))
{
if (fmpz_sgn(ARF_EXPREF(x)) < 0)
arb_atan_eps(z, x, prec);
else
arb_atan_inf_eps(z, x, prec);
}
else
{
slong exp, wp, wn, N, r;
mp_srcptr xp;
mp_size_t xn, tn;
mp_ptr tmp, w, t, u;
mp_limb_t p1, q1bits, p2, q2bits, error, error2;
int negative, inexact, reciprocal;
TMP_INIT;
exp = ARF_EXP(x);
negative = ARF_SGNBIT(x);
if (exp < -(prec/2) - 2 || exp > prec + 2)
{
if (exp < 0)
arb_atan_eps(z, x, prec);
else
arb_atan_inf_eps(z, x, prec);
return;
}
ARF_GET_MPN_READONLY(xp, xn, x);
/* Special case: +/- 1 (we require |x| != 1 later on) */
if (exp == 1 && xn == 1 && xp[xn-1] == LIMB_TOP)
{
arb_const_pi(z, prec);
arb_mul_2exp_si(z, z, -2);
if (negative)
arb_neg(z, z);
return;
}
/* Absolute working precision (NOT rounded to a limb multiple) */
wp = prec - FLINT_MIN(0, exp) + 4;
/* Too high precision to use table */
if (wp > ARB_ATAN_TAB2_PREC)
{
arb_atan_arf_bb(z, x, prec);
return;
}
/* Working precision in limbs */
wn = (wp + FLINT_BITS - 1) / FLINT_BITS;
TMP_START;
tmp = TMP_ALLOC_LIMBS(4 * wn + 3);
w = tmp; /* requires wn+1 limbs */
t = w + wn + 1; /* requires wn+1 limbs */
u = t + wn + 1; /* requires 2wn+1 limbs */
/* ----------------------------------------------------------------- */
/* Convert x or 1/x to a fixed-point number |w| < 1 */
/* ----------------------------------------------------------------- */
if (exp <= 0) /* |x| < 1 */
{
reciprocal = 0;
/* todo: just zero top */
flint_mpn_zero(w, wn);
/* w = x as a fixed-point number */
error = _arf_get_integer_mpn(w, xp, xn, exp + wn * FLINT_BITS);
}
else /* |x| > 1 */
{
slong one_exp, one_limbs, one_bits;
mp_ptr one;
reciprocal = 1;
one_exp = xn * FLINT_BITS + wn * FLINT_BITS - exp;
flint_mpn_zero(w, wn);
/* 1/x becomes zero */
if (one_exp >= FLINT_BITS - 1)
{
/* w = 1/x */
one_limbs = one_exp / FLINT_BITS;
one_bits = one_exp % FLINT_BITS;
if (one_limbs + 1 >= xn)
{
one = TMP_ALLOC_LIMBS(one_limbs + 1);
flint_mpn_zero(one, one_limbs);
one[one_limbs] = UWORD(1) << one_bits;
/* todo: only zero necessary part */
flint_mpn_zero(w, wn);
mpn_tdiv_q(w, one, one_limbs + 1, xp, xn);
/* Now w must be < 1 since x > 1 and we rounded down; thus
w[wn] must be zero */
}
}
/* todo: moderate powers of two would be exact... */
error = 1;
}
/* ----------------------------------------------------------------- */
/* Table-based argument reduction */
/* ----------------------------------------------------------------- */
/* choose p such that p/q <= x < (p+1)/q */
if (wp <= ARB_ATAN_TAB1_PREC)
q1bits = ARB_ATAN_TAB1_BITS;
else
q1bits = ARB_ATAN_TAB21_BITS;
p1 = w[wn-1] >> (FLINT_BITS - q1bits);
/* atan(w) = atan(p/q) + atan(w2) */
/* where w2 = (q*w-p)/(q+p*w) */
if (p1 != 0)
{
t[wn] = (UWORD(1) << q1bits) + mpn_mul_1(t, w, wn, p1);
flint_mpn_zero(u, wn);
u[2 * wn] = mpn_lshift(u + wn, w, wn, q1bits) - p1;
mpn_tdiv_q(w, u, 2 * wn + 1, t, wn + 1);
error++; /* w2 is computed with 1 ulp error */
}
/* Do a second round of argument reduction */
if (wp <= ARB_ATAN_TAB1_PREC)
{
p2 = 0;
}
else
{
q2bits = ARB_ATAN_TAB21_BITS + ARB_ATAN_TAB22_BITS;
p2 = w[wn-1] >> (FLINT_BITS - q2bits);
if (p2 != 0)
{
t[wn] = (UWORD(1) << q2bits) + mpn_mul_1(t, w, wn, p2);
flint_mpn_zero(u, wn);
u[2 * wn] = mpn_lshift(u + wn, w, wn, q2bits) - p2;
mpn_tdiv_q(w, u, 2 * wn + 1, t, wn + 1);
error++;
}
}
/* |w| <= 2^-r */
r = _arb_mpn_leading_zeros(w, wn);
/* N >= (wp-r)/(2r) */
N = (wp - r + (2*r-1)) / (2*r);
/* Evaluate Taylor series */
_arb_atan_taylor_rs(t, &error2, w, wn, N, 1);
/* Taylor series evaluation error */
error += error2;
/* Size of output number */
tn = wn;
/* First table lookup */
if (p1 != 0)
{
if (wp <= ARB_ATAN_TAB1_PREC)
mpn_add_n(t, t, arb_atan_tab1[p1] + ARB_ATAN_TAB1_LIMBS - tn, tn);
else
mpn_add_n(t, t, arb_atan_tab21[p1] + ARB_ATAN_TAB2_LIMBS - tn, tn);
error++;
}
/* Second table lookup */
if (p2 != 0)
{
mpn_add_n(t, t, arb_atan_tab22[p2] + ARB_ATAN_TAB2_LIMBS - tn, tn);
error++;
}
/* pi/2 - atan(1/x) */
if (reciprocal)
{
t[tn] = LIMB_ONE - mpn_sub_n(t,
arb_atan_pi2_minus_one + ARB_ATAN_TAB2_LIMBS - tn, t, tn);
/* result can be >= 1 */
tn += (t[tn] != 0);
/* error of pi/2 */
error++;
}
/* The accumulated arithmetic error */
mag_set_ui_2exp_si(arb_radref(z), error, -wn * FLINT_BITS);
/* Truncation error from the Taylor series */
mag_add_ui_2exp_si(arb_radref(z), arb_radref(z), 1, -r*(2*N+1));
/* Set the midpoint */
inexact = _arf_set_mpn_fixed(arb_midref(z), t, tn, wn, negative, prec, ARB_RND);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
TMP_END;
}
}
void
arb_log_arf(arb_t z, const arf_t x, slong prec)
{
if (arf_is_special(x))
{
if (arf_is_pos_inf(x))
arb_pos_inf(z);
else
arb_indeterminate(z);
}
else if (ARF_SGNBIT(x))
{
arb_indeterminate(z);
}
else if (ARF_IS_POW2(x))
{
if (fmpz_is_one(ARF_EXPREF(x)))
{
arb_zero(z);
}
else
{
fmpz_t exp;
fmpz_init(exp);
_fmpz_add_fast(exp, ARF_EXPREF(x), -1);
arb_const_log2(z, prec + 2);
arb_mul_fmpz(z, z, exp, prec);
fmpz_clear(exp);
}
}
else if (COEFF_IS_MPZ(*ARF_EXPREF(x)))
{
arb_log_arf_huge(z, x, prec);
}
else
{
slong exp, wp, wn, N, r, closeness_to_one;
mp_srcptr xp;
mp_size_t xn, tn;
mp_ptr tmp, w, t, u;
mp_limb_t p1, q1bits, p2, q2bits, error, error2, cy;
int negative, inexact, used_taylor_series;
TMP_INIT;
exp = ARF_EXP(x);
negative = 0;
ARF_GET_MPN_READONLY(xp, xn, x);
/* compute a c >= 0 such that |x-1| <= 2^(-c) if c > 0 */
closeness_to_one = 0;
if (exp == 0)
{
slong i;
closeness_to_one = FLINT_BITS - FLINT_BIT_COUNT(~xp[xn - 1]);
if (closeness_to_one == FLINT_BITS)
{
for (i = xn - 2; i > 0 && xp[i] == LIMB_ONES; i--)
closeness_to_one += FLINT_BITS;
closeness_to_one += (FLINT_BITS - FLINT_BIT_COUNT(~xp[i]));
}
}
else if (exp == 1)
{
closeness_to_one = FLINT_BITS - FLINT_BIT_COUNT(xp[xn - 1] & (~LIMB_TOP));
if (closeness_to_one == FLINT_BITS)
{
slong i;
for (i = xn - 2; xp[i] == 0; i--)
closeness_to_one += FLINT_BITS;
closeness_to_one += (FLINT_BITS - FLINT_BIT_COUNT(xp[i]));
}
closeness_to_one--;
}
/* if |t-1| <= 0.5 */
/* |log(1+t) - t| <= t^2 */
/* |log(1+t) - (t-t^2/2)| <= t^3 */
if (closeness_to_one > prec + 1)
{
inexact = arf_sub_ui(arb_midref(z), x, 1, prec, ARB_RND);
mag_set_ui_2exp_si(arb_radref(z), 1, -2 * closeness_to_one);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
return;
}
else if (2 * closeness_to_one > prec + 1)
{
arf_t t, u;
arf_init(t);
arf_init(u);
arf_sub_ui(t, x, 1, ARF_PREC_EXACT, ARF_RND_DOWN);
arf_mul(u, t, t, ARF_PREC_EXACT, ARF_RND_DOWN);
arf_mul_2exp_si(u, u, -1);
inexact = arf_sub(arb_midref(z), t, u, prec, ARB_RND);
mag_set_ui_2exp_si(arb_radref(z), 1, -3 * closeness_to_one);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
arf_clear(t);
arf_clear(u);
return;
}
/* Absolute working precision (NOT rounded to a limb multiple) */
wp = prec + closeness_to_one + 5;
/* Too high precision to use table */
if (wp > ARB_LOG_TAB2_PREC)
{
arf_log_via_mpfr(arb_midref(z), x, prec, ARB_RND);
arf_mag_set_ulp(arb_radref(z), arb_midref(z), prec);
return;
}
/* Working precision in limbs */
wn = (wp + FLINT_BITS - 1) / FLINT_BITS;
TMP_START;
tmp = TMP_ALLOC_LIMBS(4 * wn + 3);
w = tmp; /* requires wn+1 limbs */
t = w + wn + 1; /* requires wn+1 limbs */
u = t + wn + 1; /* requires 2wn+1 limbs */
/* read x-1 */
if (xn <= wn)
{
flint_mpn_zero(w, wn - xn);
mpn_lshift(w + wn - xn, xp, xn, 1);
error = 0;
}
else
{
mpn_lshift(w, xp + xn - wn, wn, 1);
error = 1;
}
/* First table-based argument reduction */
if (wp <= ARB_LOG_TAB1_PREC)
q1bits = ARB_LOG_TAB11_BITS;
else
q1bits = ARB_LOG_TAB21_BITS;
p1 = w[wn-1] >> (FLINT_BITS - q1bits);
/* Special case: covers logarithms of small integers */
if (xn == 1 && (w[wn-1] == (p1 << (FLINT_BITS - q1bits))))
{
p2 = 0;
flint_mpn_zero(t, wn);
used_taylor_series = 0;
N = r = 0; /* silence compiler warning */
}
else
{
/* log(1+w) = log(1+p/q) + log(1 + (qw-p)/(p+q)) */
w[wn] = mpn_mul_1(w, w, wn, UWORD(1) << q1bits) - p1;
mpn_divrem_1(w, 0, w, wn + 1, p1 + (UWORD(1) << q1bits));
error += 1;
/* Second table-based argument reduction (fused with log->atanh
conversion) */
if (wp <= ARB_LOG_TAB1_PREC)
q2bits = ARB_LOG_TAB11_BITS + ARB_LOG_TAB12_BITS;
else
q2bits = ARB_LOG_TAB21_BITS + ARB_LOG_TAB22_BITS;
p2 = w[wn-1] >> (FLINT_BITS - q2bits);
u[2 * wn] = mpn_lshift(u + wn, w, wn, q2bits);
flint_mpn_zero(u, wn);
flint_mpn_copyi(t, u + wn, wn + 1);
t[wn] += p2 + (UWORD(1) << (q2bits + 1));
u[2 * wn] -= p2;
mpn_tdiv_q(w, u, 2 * wn + 1, t, wn + 1);
/* propagated error from 1 ulp error: 2 atanh'(1/3) = 2.25 */
error += 3;
/* |w| <= 2^-r */
r = _arb_mpn_leading_zeros(w, wn);
/* N >= (wp-r)/(2r) */
N = (wp - r + (2*r-1)) / (2*r);
N = FLINT_MAX(N, 0);
/* Evaluate Taylor series */
_arb_atan_taylor_rs(t, &error2, w, wn, N, 0);
/* Multiply by 2 */
mpn_lshift(t, t, wn, 1);
/* Taylor series evaluation error (multiply by 2) */
error += error2 * 2;
used_taylor_series = 1;
}
/* Size of output number */
tn = wn;
/* First table lookup */
if (p1 != 0)
{
if (wp <= ARB_LOG_TAB1_PREC)
mpn_add_n(t, t, arb_log_tab11[p1] + ARB_LOG_TAB1_LIMBS - tn, tn);
else
mpn_add_n(t, t, arb_log_tab21[p1] + ARB_LOG_TAB2_LIMBS - tn, tn);
error++;
}
/* Second table lookup */
if (p2 != 0)
{
if (wp <= ARB_LOG_TAB1_PREC)
mpn_add_n(t, t, arb_log_tab12[p2] + ARB_LOG_TAB1_LIMBS - tn, tn);
else
mpn_add_n(t, t, arb_log_tab22[p2] + ARB_LOG_TAB2_LIMBS - tn, tn);
error++;
}
/* add exp * log(2) */
exp--;
if (exp > 0)
{
cy = mpn_addmul_1(t, arb_log_log2_tab + ARB_LOG_TAB2_LIMBS - tn, tn, exp);
t[tn] = cy;
tn += (cy != 0);
error += exp;
}
else if (exp < 0)
{
t[tn] = 0;
u[tn] = mpn_mul_1(u, arb_log_log2_tab + ARB_LOG_TAB2_LIMBS - tn, tn, -exp);
if (mpn_cmp(t, u, tn + 1) >= 0)
{
mpn_sub_n(t, t, u, tn + 1);
}
else
{
mpn_sub_n(t, u, t, tn + 1);
negative = 1;
}
error += (-exp);
tn += (t[tn] != 0);
}
/* The accumulated arithmetic error */
mag_set_ui_2exp_si(arb_radref(z), error, -wn * FLINT_BITS);
/* Truncation error from the Taylor series */
if (used_taylor_series)
mag_add_ui_2exp_si(arb_radref(z), arb_radref(z), 1, -r*(2*N+1) + 1);
/* Set the midpoint */
inexact = _arf_set_mpn_fixed(arb_midref(z), t, tn, wn, negative, prec);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
TMP_END;
}
}
void
arb_exp_arf_bb(arb_t z, const arf_t x, slong prec, int minus_one)
{
slong k, iter, bits, r, mag, q, wp, N;
slong argred_bits, start_bits;
mp_bitcnt_t Qexp[1];
int inexact;
fmpz_t t, u, T, Q;
arb_t w;
if (arf_is_zero(x))
{
if (minus_one)
arb_zero(z);
else
arb_one(z);
return;
}
if (arf_is_special(x))
{
abort();
}
mag = arf_abs_bound_lt_2exp_si(x);
/* We assume that this function only gets called with something
reasonable as input (huge/tiny input will be handled by
the main exp wrapper). */
if (mag > 200 || mag < -2 * prec - 100)
{
flint_printf("arb_exp_arf_bb: unexpectedly large/small input\n");
abort();
}
if (prec < 100000000)
{
argred_bits = 16;
start_bits = 32;
}
else
{
argred_bits = 32;
start_bits = 64;
}
/* Argument reduction: exp(x) -> exp(x/2^q). This improves efficiency
of the first iteration in the bit-burst algorithm. */
q = FLINT_MAX(0, mag + argred_bits);
/* Determine working precision. */
wp = prec + 10 + 2 * q + 2 * FLINT_BIT_COUNT(prec);
if (minus_one && mag < 0)
wp += (-mag);
fmpz_init(t);
fmpz_init(u);
fmpz_init(Q);
fmpz_init(T);
arb_init(w);
/* Convert x/2^q to a fixed-point number. */
inexact = arf_get_fmpz_fixed_si(t, x, -wp + q);
/* Aliasing of z and x is safe now that only use t. */
/* Start with z = 1. */
arb_one(z);
/* Bit-burst loop. */
for (iter = 0, bits = start_bits; !fmpz_is_zero(t);
iter++, bits *= 2)
{
/* Extract bits. */
r = FLINT_MIN(bits, wp);
fmpz_tdiv_q_2exp(u, t, wp - r);
/* Binary splitting (+1 fixed-point ulp truncation error). */
mag = fmpz_bits(u) - r;
N = bs_num_terms(mag, wp);
_arb_exp_sum_bs_powtab(T, Q, Qexp, u, r, N);
/* T = T / Q (+1 fixed-point ulp error). */
if (*Qexp >= wp)
{
fmpz_tdiv_q_2exp(T, T, *Qexp - wp);
fmpz_tdiv_q(T, T, Q);
}
else
{
fmpz_mul_2exp(T, T, wp - *Qexp);
fmpz_tdiv_q(T, T, Q);
}
/* T = 1 + T */
fmpz_one(Q);
fmpz_mul_2exp(Q, Q, wp);
fmpz_add(T, T, Q);
/* Now T = exp(u) with at most 2 fixed-point ulp error. */
/* Set z = z * T. */
arf_set_fmpz(arb_midref(w), T);
arf_mul_2exp_si(arb_midref(w), arb_midref(w), -wp);
mag_set_ui_2exp_si(arb_radref(w), 2, -wp);
arb_mul(z, z, w, wp);
/* Remove used bits. */
fmpz_mul_2exp(u, u, wp - r);
fmpz_sub(t, t, u);
}
/* We have exp(x + eps) - exp(x) < 2*eps (by assumption that the argument
reduction is large enough). */
if (inexact)
arb_add_error_2exp_si(z, -wp + 1);
fmpz_clear(t);
fmpz_clear(u);
fmpz_clear(Q);
fmpz_clear(T);
arb_clear(w);
/* exp(x) = exp(x/2^q)^(2^q) */
for (k = 0; k < q; k++)
arb_mul(z, z, z, wp);
if (minus_one)
arb_sub_ui(z, z, 1, wp);
arb_set_round(z, z, prec);
}
