void
arb_poly_exp_series(arb_poly_t f, const arb_poly_t h, slong n, slong prec)
{
slong hlen = h->length;
if (n == 0)
{
arb_poly_zero(f);
return;
}
if (hlen == 0)
{
arb_poly_one(f);
return;
}
if (hlen == 1)
n = 1;
arb_poly_fit_length(f, n);
_arb_poly_exp_series(f->coeffs, h->coeffs, hlen, n, prec);
_arb_poly_set_length(f, n);
_arb_poly_normalise(f);
}
void
_arb_poly_sinh_cosh_series_exponential(arb_ptr s, arb_ptr c,
const arb_srcptr h, slong hlen, slong len, slong prec)
{
arb_ptr t, u, v;
arb_t s0, c0;
hlen = FLINT_MIN(hlen, len);
if (hlen == 1)
{
arb_sinh_cosh(s, c, h, prec);
_arb_vec_zero(s + 1, len - 1);
_arb_vec_zero(c + 1, len - 1);
return;
}
arb_init(s0);
arb_init(c0);
t = _arb_vec_init(3 * len);
u = t + len;
v = u + len;
arb_sinh_cosh(s0, c0, h, prec);
_arb_vec_set(t + 1, h + 1, hlen - 1);
_arb_poly_exp_series(t, t, len, len, prec);
/* todo: part of the inverse could be avoided since exp computes
it internally to half the length */
_arb_poly_inv_series(u, t, len, len, prec);
/* hyperbolic sine */
_arb_vec_sub(s, t, u, len, prec);
_arb_vec_scalar_mul_2exp_si(s, s, len, -1);
/* hyperbolic cosine */
_arb_vec_add(c, t, u, len, prec);
_arb_vec_scalar_mul_2exp_si(c, c, len, -1);
/* sinh(h0 + h1) = cosh(h0) sinh(h1) + sinh(h0) cosh(h1)
cosh(h0 + h1) = cosh(h0) cosh(h1) + sinh(h0) sinh(h1) */
if (!arb_is_zero(s0))
{
_arb_vec_scalar_mul(t, s, len, c0, prec);
_arb_vec_scalar_mul(u, c, len, s0, prec);
_arb_vec_scalar_mul(v, s, len, s0, prec);
_arb_vec_add(s, t, u, len, prec);
_arb_vec_scalar_mul(t, c, len, c0, prec);
_arb_vec_add(c, t, v, len, prec);
}
_arb_vec_clear(t, 3 * len);
arb_clear(s0);
arb_clear(c0);
}
void
_arb_poly_rgamma_series(arb_ptr res, arb_srcptr h, long hlen, long len, long prec)
{
int reflect;
long i, rflen, r, n, wp;
arb_ptr t, u, v;
arb_struct f[2];
hlen = FLINT_MIN(hlen, len);
wp = prec + FLINT_BIT_COUNT(prec);
t = _arb_vec_init(len);
u = _arb_vec_init(len);
v = _arb_vec_init(len);
arb_init(f);
arb_init(f + 1);
/* use zeta values at small integers */
if (arb_is_int(h) && (arf_cmpabs_ui(arb_midref(h), prec / 2) < 0))
{
r = arf_get_si(arb_midref(h), ARF_RND_DOWN);
_arb_poly_lgamma_series_at_one(u, len, wp);
_arb_vec_neg(u, u, len);
_arb_poly_exp_series(t, u, len, len, wp);
if (r == 1)
{
_arb_vec_swap(v, t, len);
}
else if (r <= 0)
{
arb_set(f, h);
arb_one(f + 1);
rflen = FLINT_MIN(len, 2 - r);
_arb_poly_rising_ui_series(u, f, FLINT_MIN(2, len), 1 - r, rflen, wp);
_arb_poly_mullow(v, t, len, u, rflen, len, wp);
}
else
{
arb_one(f);
arb_one(f + 1);
rflen = FLINT_MIN(len, r);
_arb_poly_rising_ui_series(v, f, FLINT_MIN(2, len), r - 1, rflen, wp);
/* TODO: use div_series? */
_arb_poly_inv_series(u, v, rflen, len, wp);
_arb_poly_mullow(v, t, len, u, len, len, wp);
}
}
else
{
/* otherwise use Stirling series */
arb_gamma_stirling_choose_param(&reflect, &r, &n, h, 1, 0, wp);
/* rgamma(h) = (gamma(1-h+r) sin(pi h)) / (rf(1-h, r) * pi), h = h0 + t*/
if (reflect)
{
/* u = gamma(r+1-h) */
arb_sub_ui(f, h, r + 1, wp);
arb_neg(f, f);
_arb_poly_gamma_stirling_eval(t, f, n, len, wp);
_arb_poly_exp_series(u, t, len, len, wp);
for (i = 1; i < len; i += 2)
arb_neg(u + i, u + i);
/* v = sin(pi x) */
arb_const_pi(f + 1, wp);
arb_mul(f, h, f + 1, wp);
_arb_poly_sin_series(v, f, 2, len, wp);
_arb_poly_mullow(t, u, len, v, len, len, wp);
/* rf(1-h,r) * pi */
if (r == 0)
{
arb_const_pi(u, wp);
_arb_vec_scalar_div(v, t, len, u, wp);
}
else
{
arb_sub_ui(f, h, 1, wp);
arb_neg(f, f);
arb_set_si(f + 1, -1);
rflen = FLINT_MIN(len, r + 1);
_arb_poly_rising_ui_series(v, f, FLINT_MIN(2, len), r, rflen, wp);
arb_const_pi(u, wp);
_arb_vec_scalar_mul(v, v, rflen, u, wp);
/* divide by rising factorial */
/* TODO: might better to use div_series, when it has a good basecase */
_arb_poly_inv_series(u, v, rflen, len, wp);
_arb_poly_mullow(v, t, len, u, len, len, wp);
}
}
else
{
/* rgamma(h) = rgamma(h+r) rf(h,r) */
if (r == 0)
{
arb_add_ui(f, h, r, wp);
_arb_poly_gamma_stirling_eval(t, f, n, len, wp);
_arb_vec_neg(t, t, len);
_arb_poly_exp_series(v, t, len, len, wp);
}
else
{
arb_set(f, h);
arb_one(f + 1);
rflen = FLINT_MIN(len, r + 1);
_arb_poly_rising_ui_series(t, f, FLINT_MIN(2, len), r, rflen, wp);
arb_add_ui(f, h, r, wp);
_arb_poly_gamma_stirling_eval(v, f, n, len, wp);
_arb_vec_neg(v, v, len);
_arb_poly_exp_series(u, v, len, len, wp);
_arb_poly_mullow(v, u, len, t, rflen, len, wp);
}
}
}
/* compose with nonconstant part */
arb_zero(t);
_arb_vec_set(t + 1, h + 1, hlen - 1);
_arb_poly_compose_series(res, v, len, t, hlen, len, prec);
arb_clear(f);
arb_clear(f + 1);
_arb_vec_clear(t, len);
_arb_vec_clear(u, len);
_arb_vec_clear(v, len);
}
void
_arb_bell_sum_taylor(arb_t res, const fmpz_t n,
const fmpz_t a, const fmpz_t b, const fmpz_t mmag, long tol)
{
fmpz_t m, r, R, tmp;
mag_t B, C, D, bound;
arb_t t, u;
long wp, k, N;
if (_fmpz_sub_small(b, a) < 5)
{
arb_bell_sum_bsplit(res, n, a, b, mmag, tol);
return;
}
fmpz_init(m);
fmpz_init(r);
fmpz_init(R);
fmpz_init(tmp);
/* r = max(m - a, b - m) */
/* m = a + (b - a) / 2 */
fmpz_sub(r, b, a);
fmpz_cdiv_q_2exp(r, r, 1);
fmpz_add(m, a, r);
fmpz_mul_2exp(R, r, RADIUS_BITS);
mag_init(B);
mag_init(C);
mag_init(D);
mag_init(bound);
arb_init(t);
arb_init(u);
if (fmpz_cmp(R, m) >= 0)
{
mag_inf(C);
mag_inf(D);
}
else
{
/* C = exp(R * |F'(m)| + (1/2) R^2 * (n/(m-R)^2 + 1/(m-R))) */
/* C = exp(R * (|F'(m)| + (1/2) R * (n/(m-R) + 1)/(m-R))) */
/* D = (1/2) R * (n/(m-R) + 1)/(m-R) */
fmpz_sub(tmp, m, R);
mag_set_fmpz(D, n);
mag_div_fmpz(D, D, tmp);
mag_one(C);
mag_add(D, D, C);
mag_div_fmpz(D, D, tmp);
mag_mul_fmpz(D, D, R);
mag_mul_2exp_si(D, D, -1);
/* C = |F'(m)| */
wp = 20 + 1.05 * fmpz_bits(n);
arb_set_fmpz(t, n);
arb_div_fmpz(t, t, m, wp);
fmpz_add_ui(tmp, m, 1);
arb_set_fmpz(u, tmp);
arb_digamma(u, u, wp);
arb_sub(t, t, u, wp);
arb_get_mag(C, t);
/* C = exp(R * (C + D)) */
mag_add(C, C, D);
mag_mul_fmpz(C, C, R);
mag_exp(C, C);
}
if (mag_cmp_2exp_si(C, tol / 4 + 2) > 0)
{
_arb_bell_sum_taylor(res, n, a, m, mmag, tol);
_arb_bell_sum_taylor(t, n, m, b, mmag, tol);
arb_add(res, res, t, 2 * tol);
}
else
{
arb_ptr mx, ser1, ser2, ser3;
/* D = T(m) */
wp = 20 + 1.05 * fmpz_bits(n);
arb_set_fmpz(t, m);
arb_pow_fmpz(t, t, n, wp);
fmpz_add_ui(tmp, m, 1);
arb_gamma_fmpz(u, tmp, wp);
arb_div(t, t, u, wp);
arb_get_mag(D, t);
/* error bound: (b-a) * C * D * B^N / (1 - B), B = r/R */
/* ((b-a) * C * D * 2) * 2^(-N*RADIUS_BITS) */
/* ((b-a) * C * D * 2) */
mag_mul(bound, C, D);
mag_mul_2exp_si(bound, bound, 1);
fmpz_sub(tmp, b, a);
mag_mul_fmpz(bound, bound, tmp);
/* N = (tol + log2((b-a)*C*D*2) - mmag) / RADIUS_BITS */
if (mmag == NULL)
{
/* estimate D ~= 2^mmag */
fmpz_add_ui(tmp, MAG_EXPREF(C), tol);
fmpz_cdiv_q_ui(tmp, tmp, RADIUS_BITS);
}
else
{
fmpz_sub(tmp, MAG_EXPREF(bound), mmag);
fmpz_add_ui(tmp, tmp, tol);
fmpz_cdiv_q_ui(tmp, tmp, RADIUS_BITS);
}
if (fmpz_cmp_ui(tmp, 5 * tol / 4) > 0)
N = 5 * tol / 4;
else if (fmpz_cmp_ui(tmp, 2) < 0)
N = 2;
else
N = fmpz_get_ui(tmp);
/* multiply by 2^(-N*RADIUS_BITS) */
mag_mul_2exp_si(bound, bound, -N * RADIUS_BITS);
mx = _arb_vec_init(2);
ser1 = _arb_vec_init(N);
ser2 = _arb_vec_init(N);
ser3 = _arb_vec_init(N);
/* estimate (this should work for moderate n and tol) */
wp = 1.1 * tol + 1.05 * fmpz_bits(n) + 5;
/* increase precision until convergence */
while (1)
{
/* (m+x)^n / gamma(m+1+x) */
arb_set_fmpz(mx, m);
arb_one(mx + 1);
_arb_poly_log_series(ser1, mx, 2, N, wp);
for (k = 0; k < N; k++)
arb_mul_fmpz(ser1 + k, ser1 + k, n, wp);
arb_add_ui(mx, mx, 1, wp);
_arb_poly_lgamma_series(ser2, mx, 2, N, wp);
_arb_vec_sub(ser1, ser1, ser2, N, wp);
_arb_poly_exp_series(ser3, ser1, N, N, wp);
/* t = a - m, u = b - m */
arb_set_fmpz(t, a);
arb_sub_fmpz(t, t, m, wp);
arb_set_fmpz(u, b);
arb_sub_fmpz(u, u, m, wp);
arb_power_sum_vec(ser1, t, u, N, wp);
arb_zero(res);
for (k = 0; k < N; k++)
arb_addmul(res, ser3 + k, ser1 + k, wp);
if (mmag != NULL)
{
if (_fmpz_sub_small(MAG_EXPREF(arb_radref(res)), mmag) <= -tol)
break;
}
else
{
if (arb_rel_accuracy_bits(res) >= tol)
break;
}
wp = 2 * wp;
}
/* add the series truncation bound */
arb_add_error_mag(res, bound);
_arb_vec_clear(mx, 2);
_arb_vec_clear(ser1, N);
_arb_vec_clear(ser2, N);
_arb_vec_clear(ser3, N);
}
mag_clear(B);
mag_clear(C);
mag_clear(D);
mag_clear(bound);
arb_clear(t);
arb_clear(u);
fmpz_clear(m);
fmpz_clear(r);
fmpz_clear(R);
fmpz_clear(tmp);
}
