void
mag_sqrt(mag_t y, const mag_t x)
{
if (mag_is_special(x))
{
mag_set(y, x);
}
else
{
double t;
fmpz e;
t = MAG_MAN(x) * ldexp(1.0, -MAG_BITS);
e = MAG_EXP(x);
if (!COEFF_IS_MPZ(e))
{
if (e % 2 != 0)
{
e = (e - 1) >> 1;
t *= 2.0;
}
else
{
e >>= 1;
}
t = sqrt(t) * (1 + 1e-13);
mag_set_d_2exp_fmpz(y, t, &e);
}
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);
}
}
double
mag_get_d(const mag_t z)
{
if (mag_is_zero(z))
{
return 0.0;
}
else if (mag_is_inf(z))
{
return D_INF;
}
else if (MAG_EXP(z) < -1000 || MAG_EXP(z) > 1000)
{
if (fmpz_sgn(MAG_EXPREF(z)) < 0)
return ldexp(1.0, -1000);
else
return D_INF;
}
else
{
return ldexp(MAG_MAN(z), MAG_EXP(z) - MAG_BITS);
}
}
static __inline__ long
rec_fac_bound_2exp_si(slong n)
{
if (n < TABSIZE)
{
return rec_fac_bound_2exp_si_tab[n];
}
else
{
mag_t t;
mag_init(t);
mag_rfac_ui(t, n); /* todo: check for overflow */
return MAG_EXP(t);
}
}
void
mag_set_d_lower(mag_t z, double c)
{
if (c < 0.0)
c = -c;
if (c == 0.0 || (c != c))
{
mag_zero(z);
}
else if (c == D_INF)
{
mag_inf(z);
}
else
{
_fmpz_demote(MAG_EXPREF(z));
MAG_SET_D_2EXP_LOWER(MAG_MAN(z), MAG_EXP(z), c, 0);
}
}
void
mag_rfac_ui(mag_t z, ulong n)
{
if (n < MAG_FAC_TABLE_NUM)
{
_fmpz_demote(MAG_EXPREF(z));
MAG_EXP(z) = mag_rfac_tab[n * 2];
MAG_MAN(z) = mag_rfac_tab[n * 2 + 1];
}
else
{
double x = n;
x = ceil((((x+0.5)*mag_d_log_lower_bound(x) - x) * 1.4426950408889634074) * -0.9999999);
/* x + 1 could round down for huge x, but this doesn't matter
as long as the value was perturbed up above */
fmpz_set_d(MAG_EXPREF(z), x + 1);
MAG_MAN(z) = MAG_ONE_HALF;
}
}
slong
_arb_mat_exp_choose_N(const mag_t norm, slong prec)
{
if (mag_is_special(norm) || mag_cmp_2exp_si(norm, 30) > 0 ||
mag_cmp_2exp_si(norm, -prec) < 0)
{
return 1;
}
else if (mag_cmp_2exp_si(norm, -300) < 0)
{
slong N = -MAG_EXP(norm);
return (prec + N - 1) / N;
}
else
{
double c, t;
c = mag_get_d(norm);
t = d_lambertw(prec * LOG2_OVER_E / c);
t = c * exp(t + 1.0);
return FLINT_MIN((slong) (t + 1.0), 2 * prec);
}
}
void
_arb_sin_cos_generic(arb_t s, arb_t c, const arf_t x, const mag_t xrad, slong prec)
{
int want_sin, want_cos;
slong maglim;
want_sin = (s != NULL);
want_cos = (c != NULL);
if (arf_is_zero(x) && mag_is_zero(xrad))
{
if (want_sin) arb_zero(s);
if (want_cos) arb_one(c);
return;
}
if (!arf_is_finite(x) || !mag_is_finite(xrad))
{
if (arf_is_nan(x))
{
if (want_sin) arb_indeterminate(s);
if (want_cos) arb_indeterminate(c);
}
else
{
if (want_sin) arb_zero_pm_one(s);
if (want_cos) arb_zero_pm_one(c);
}
return;
}
maglim = FLINT_MAX(65536, 4 * prec);
if (mag_cmp_2exp_si(xrad, -16) > 0 || arf_cmpabs_2exp_si(x, maglim) > 0)
{
_arb_sin_cos_wide(s, c, x, xrad, prec);
return;
}
if (arf_cmpabs_2exp_si(x, -(prec/2) - 2) <= 0)
{
mag_t t, u, v;
mag_init(t);
mag_init(u);
mag_init(v);
arf_get_mag(t, x);
mag_add(t, t, xrad);
mag_mul(u, t, t);
/* |sin(z)-z| <= z^3/6 */
if (want_sin)
{
arf_set(arb_midref(s), x);
mag_set(arb_radref(s), xrad);
arb_set_round(s, s, prec);
mag_mul(v, u, t);
mag_div_ui(v, v, 6);
arb_add_error_mag(s, v);
}
/* |cos(z)-1| <= z^2/2 */
if (want_cos)
{
arf_one(arb_midref(c));
mag_mul_2exp_si(arb_radref(c), u, -1);
}
mag_clear(t);
mag_clear(u);
mag_clear(v);
return;
}
if (mag_is_zero(xrad))
{
arb_sin_cos_arf_generic(s, c, x, prec);
}
else
{
mag_t t;
slong exp, radexp;
mag_init_set(t, xrad);
exp = arf_abs_bound_lt_2exp_si(x);
radexp = MAG_EXP(xrad);
if (radexp < MAG_MIN_LAGOM_EXP || radexp > MAG_MAX_LAGOM_EXP)
radexp = MAG_MIN_LAGOM_EXP;
if (want_cos && exp < -2)
prec = FLINT_MIN(prec, 20 - FLINT_MAX(exp, radexp) - radexp);
else
prec = FLINT_MIN(prec, 20 - radexp);
arb_sin_cos_arf_generic(s, c, x, prec);
/* todo: could use quadratic bound */
if (want_sin) mag_add(arb_radref(s), arb_radref(s), t);
if (want_cos) mag_add(arb_radref(c), arb_radref(c), t);
mag_clear(t);
}
}
void
arb_mat_exp(arb_mat_t B, const arb_mat_t A, slong prec)
{
slong i, j, dim, wp, N, q, r;
mag_t norm, err;
arb_mat_t T;
dim = arb_mat_nrows(A);
if (dim != arb_mat_ncols(A))
{
flint_printf("arb_mat_exp: a square matrix is required!\n");
abort();
}
if (dim == 0)
{
return;
}
else if (dim == 1)
{
arb_exp(arb_mat_entry(B, 0, 0), arb_mat_entry(A, 0, 0), prec);
return;
}
wp = prec + 3 * FLINT_BIT_COUNT(prec);
mag_init(norm);
mag_init(err);
arb_mat_init(T, dim, dim);
arb_mat_bound_inf_norm(norm, A);
if (mag_is_zero(norm))
{
arb_mat_one(B);
}
else
{
q = pow(wp, 0.25); /* wanted magnitude */
if (mag_cmp_2exp_si(norm, 2 * wp) > 0) /* too big */
r = 2 * wp;
else if (mag_cmp_2exp_si(norm, -q) < 0) /* tiny, no need to reduce */
r = 0;
else
r = FLINT_MAX(0, q + MAG_EXP(norm)); /* reduce to magnitude 2^(-r) */
arb_mat_scalar_mul_2exp_si(T, A, -r);
mag_mul_2exp_si(norm, norm, -r);
N = _arb_mat_exp_choose_N(norm, wp);
mag_exp_tail(err, norm, N);
_arb_mat_exp_taylor(B, T, N, wp);
for (i = 0; i < dim; i++)
for (j = 0; j < dim; j++)
arb_add_error_mag(arb_mat_entry(B, i, j), err);
for (i = 0; i < r; i++)
{
arb_mat_mul(T, B, B, wp);
arb_mat_swap(T, B);
}
for (i = 0; i < dim; i++)
for (j = 0; j < dim; j++)
arb_set_round(arb_mat_entry(B, i, j),
arb_mat_entry(B, i, j), prec);
}
mag_clear(norm);
mag_clear(err);
arb_mat_clear(T);
}
void acb_hypgeom_u_asymp(acb_t res, const acb_t a, const acb_t b,
const acb_t z, slong n, slong prec)
{
acb_struct aa[3];
acb_t s, t, w, winv;
int R, p, q, is_real, is_terminating;
slong n_terminating;
if (!acb_is_finite(a) || !acb_is_finite(b) || !acb_is_finite(z))
{
acb_indeterminate(res);
return;
}
acb_init(aa);
acb_init(aa + 1);
acb_init(aa + 2);
acb_init(s);
acb_init(t);
acb_init(w);
acb_init(winv);
is_terminating = 0;
n_terminating = WORD_MAX;
/* special case, for incomplete gamma
[todo: also when they happen to be exact and with difference 1...] */
if (a == b)
{
acb_set(aa, a);
p = 1;
q = 0;
}
else
{
acb_set(aa, a);
acb_sub(aa + 1, a, b, prec);
acb_add_ui(aa + 1, aa + 1, 1, prec);
acb_one(aa + 2);
p = 2;
q = 1;
}
if (acb_is_nonpositive_int(aa))
{
is_terminating = 1;
if (arf_cmpabs_ui(arb_midref(acb_realref(aa)), prec) < 0)
n_terminating = 1 - arf_get_si(arb_midref(acb_realref(aa)), ARF_RND_DOWN);
}
if (p == 2 && acb_is_nonpositive_int(aa + 1))
{
is_terminating = 1;
if (arf_cmpabs_ui(arb_midref(acb_realref(aa + 1)), n_terminating) < 0)
n_terminating = 1 - arf_get_si(arb_midref(acb_realref(aa + 1)), ARF_RND_DOWN);
}
acb_neg(w, z);
acb_inv(w, w, prec);
acb_neg(winv, z);
/* low degree polynomial -- no need to try to terminate sooner */
if (is_terminating && n_terminating < 8)
{
acb_hypgeom_pfq_sum_invz(s, t, aa, p, aa + p, q, w, winv,
n_terminating, prec);
acb_set(res, s);
}
else
{
mag_t C1, Cn, alpha, nu, sigma, rho, zinv, tmp, err;
mag_init(C1);
mag_init(Cn);
mag_init(alpha);
mag_init(nu);
mag_init(sigma);
mag_init(rho);
mag_init(zinv);
mag_init(tmp);
mag_init(err);
acb_hypgeom_u_asymp_bound_factors(&R, alpha, nu,
sigma, rho, zinv, a, b, z);
is_real = acb_is_real(a) && acb_is_real(b) && acb_is_real(z) &&
(is_terminating || arb_is_positive(acb_realref(z)));
if (R == 0)
{
/* if R == 0, the error bound is infinite unless terminating */
if (is_terminating && n_terminating < prec)
{
acb_hypgeom_pfq_sum_invz(s, t, aa, p, aa + p, q, w, winv,
n_terminating, prec);
acb_set(res, s);
}
else
{
acb_indeterminate(res);
}
}
else
{
/* C1 */
acb_hypgeom_mag_Cn(C1, R, nu, sigma, 1);
/* err = 2 * alpha * exp(...) */
mag_mul(tmp, C1, rho);
mag_mul(tmp, tmp, alpha);
mag_mul(tmp, tmp, zinv);
mag_mul_2exp_si(tmp, tmp, 1);
mag_exp(err, tmp);
mag_mul(err, err, alpha);
mag_mul_2exp_si(err, err, 1);
/* choose n automatically */
if (n < 0)
{
slong moreprec;
/* take err into account when finding truncation point */
/* we should take Cn into account as well, but this depends
on n which is to be determined; it's easier to look
only at exp(...) which should be larger anyway */
if (mag_cmp_2exp_si(err, 10 * prec) > 0)
moreprec = 10 * prec;
else if (mag_cmp_2exp_si(err, 0) < 0)
moreprec = 0;
else
moreprec = MAG_EXP(err);
n = acb_hypgeom_pfq_choose_n_max(aa, p, aa + p, q, w,
prec + moreprec, FLINT_MIN(WORD_MAX / 2, 50 + 10.0 * prec));
}
acb_hypgeom_pfq_sum_invz(s, t, aa, p, aa + p, q, w, winv, n, prec);
/* add error bound, if not terminating */
if (!(is_terminating && n == n_terminating))
{
acb_hypgeom_mag_Cn(Cn, R, nu, sigma, n);
mag_mul(err, err, Cn);
/* nth term * factor */
acb_get_mag(tmp, t);
mag_mul(err, err, tmp);
if (is_real)
arb_add_error_mag(acb_realref(s), err);
else
acb_add_error_mag(s, err);
}
acb_set(res, s);
}
mag_clear(C1);
mag_clear(Cn);
mag_clear(alpha);
mag_clear(nu);
mag_clear(sigma);
mag_clear(rho);
mag_clear(zinv);
mag_clear(tmp);
mag_clear(err);
}
acb_clear(aa);
acb_clear(aa + 1);
acb_clear(aa + 2);
acb_clear(s);
acb_clear(t);
acb_clear(w);
acb_clear(winv);
}
/* note: z should be exact here */
void acb_lambertw_main(acb_t res, const acb_t z,
const acb_t ez1, const fmpz_t k, int flags, slong prec)
{
acb_t w, t, oldw, ew;
mag_t err;
slong i, wp, accuracy, ebits, kbits, mbits, wp_initial, extraprec;
int have_ew;
acb_init(t);
acb_init(w);
acb_init(oldw);
acb_init(ew);
mag_init(err);
/* We need higher precision for large k, large exponents, or very close
to the branch point at -1/e. todo: we should be recomputing
ez1 to higher precision when close... */
acb_get_mag(err, z);
if (fmpz_is_zero(k) && mag_cmp_2exp_si(err, 0) < 0)
ebits = 0;
else
ebits = fmpz_bits(MAG_EXPREF(err));
if (fmpz_is_zero(k) || (fmpz_is_one(k) && arb_is_negative(acb_imagref(z)))
|| (fmpz_equal_si(k, -1) && arb_is_nonnegative(acb_imagref(z))))
{
acb_get_mag(err, ez1);
mbits = -MAG_EXP(err);
mbits = FLINT_MAX(mbits, 0);
mbits = FLINT_MIN(mbits, prec);
}
else
{
mbits = 0;
}
kbits = fmpz_bits(k);
extraprec = FLINT_MAX(ebits, kbits);
extraprec = FLINT_MAX(extraprec, mbits);
wp = wp_initial = 40 + extraprec;
accuracy = acb_lambertw_initial(w, z, ez1, k, wp_initial);
mag_zero(arb_radref(acb_realref(w)));
mag_zero(arb_radref(acb_imagref(w)));
/* We should be able to compute e^w for the final certification
during the Halley iteration. */
have_ew = 0;
for (i = 0; i < 5 + FLINT_BIT_COUNT(prec + extraprec); i++)
{
/* todo: should we restart? */
if (!acb_is_finite(w))
break;
wp = FLINT_MIN(3 * accuracy, 1.1 * prec + 10);
wp = FLINT_MAX(wp, 40);
wp += extraprec;
acb_set(oldw, w);
acb_lambertw_halley_step(t, ew, z, w, wp);
/* estimate the error (conservatively) */
acb_sub(w, w, t, wp);
acb_get_mag(err, w);
acb_set(w, t);
acb_add_error_mag(t, err);
accuracy = acb_rel_accuracy_bits(t);
if (accuracy > 2 * extraprec)
accuracy *= 2.9; /* less conservatively */
accuracy = FLINT_MIN(accuracy, wp);
accuracy = FLINT_MAX(accuracy, 0);
if (accuracy > prec + extraprec)
{
/* e^w = e^oldw * e^(w-oldw) */
acb_sub(t, w, oldw, wp);
acb_exp(t, t, wp);
acb_mul(ew, ew, t, wp);
have_ew = 1;
break;
}
mag_zero(arb_radref(acb_realref(w)));
mag_zero(arb_radref(acb_imagref(w)));
}
wp = FLINT_MIN(3 * accuracy, 1.1 * prec + 10);
wp = FLINT_MAX(wp, 40);
wp += extraprec;
if (acb_lambertw_check_branch(w, k, wp))
{
acb_t u, r, eu1;
mag_t err, rad;
acb_init(u);
acb_init(r);
acb_init(eu1);
mag_init(err);
mag_init(rad);
if (have_ew)
acb_set(t, ew);
else
acb_exp(t, w, wp);
/* t = w e^w */
acb_mul(t, t, w, wp);
acb_sub(r, t, z, wp);
/* Bound W' on the straight line path between t and z */
acb_union(u, t, z, wp);
arb_const_e(acb_realref(eu1), wp);
arb_zero(acb_imagref(eu1));
acb_mul(eu1, eu1, u, wp);
acb_add_ui(eu1, eu1, 1, wp);
if (acb_lambertw_branch_crossing(u, eu1, k))
{
mag_inf(err);
}
else
{
acb_lambertw_bound_deriv(err, u, eu1, k);
acb_get_mag(rad, r);
mag_mul(err, err, rad);
}
acb_add_error_mag(w, err);
acb_set(res, w);
acb_clear(u);
acb_clear(r);
acb_clear(eu1);
mag_clear(err);
mag_clear(rad);
}
else
{
acb_indeterminate(res);
}
acb_clear(t);
acb_clear(w);
acb_clear(oldw);
acb_clear(ew);
mag_clear(err);
}
void
mag_log1p(mag_t z, const mag_t x)
{
if (mag_is_special(x))
{
if (mag_is_zero(x))
mag_zero(z);
else
mag_inf(z);
}
else
{
fmpz exp = MAG_EXP(x);
if (!COEFF_IS_MPZ(exp))
{
/* Quick bound by x */
if (exp < -10)
{
mag_set(z, x);
return;
}
else if (exp < 1000)
{
double t;
t = ldexp(MAG_MAN(x), exp - MAG_BITS);
t = (1.0 + t) * (1 + 1e-14);
t = mag_d_log_upper_bound(t);
mag_set_d(z, t);
return;
}
}
else if (fmpz_sgn(MAG_EXPREF(x)) < 0)
{
/* Quick bound by x */
mag_set(z, x);
return;
}
/* Now we must have x >= 2^1000 */
/* Use log(2^(exp-1) * (2*v)) = exp*log(2) + log(2*v) */
{
double t;
fmpz_t b;
mag_t u;
mag_init(u);
fmpz_init(b);
/* incrementing the mantissa gives an upper bound for x+1 */
t = ldexp(MAG_MAN(x) + 1, 1 - MAG_BITS);
t = mag_d_log_upper_bound(t);
mag_set_d(u, t);
/* log(2) < 744261118/2^30 */
_fmpz_add_fast(b, MAG_EXPREF(x), -1);
fmpz_mul_ui(b, b, 744261118);
mag_set_fmpz(z, b);
_fmpz_add_fast(MAG_EXPREF(z), MAG_EXPREF(z), -30);
mag_add(z, z, u);
mag_clear(u);
fmpz_clear(b);
}
}
}
