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);
}
}
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;
}
}
void
mag_root(mag_t y, const mag_t x, ulong n)
{
if (n == 0)
{
mag_inf(y);
}
else if (n == 1 || mag_is_special(x))
{
mag_set(y, x);
}
else if (n == 2)
{
mag_sqrt(y, x);
}
else if (n == 4)
{
mag_sqrt(y, x);
mag_sqrt(y, y);
}
else
{
fmpz_t e, f;
fmpz_init_set_ui(e, MAG_BITS);
fmpz_init(f);
/* We evaluate exp(log(1+2^(kn)x)/n) 2^-k where k is chosen
so that 2^(kn) x ~= 2^30. TODO: this rewriting is probably
unnecessary with the new exp/log functions. */
fmpz_sub(e, e, MAG_EXPREF(x));
fmpz_cdiv_q_ui(e, e, n);
fmpz_mul_ui(f, e, n);
mag_mul_2exp_fmpz(y, x, f);
mag_log1p(y, y);
mag_div_ui(y, y, n);
mag_exp(y, y);
fmpz_neg(e, e);
mag_mul_2exp_fmpz(y, y, e);
fmpz_clear(e);
fmpz_clear(f);
}
}
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
arb_const_log2(arb_t res, slong prec)
{
if (prec < ARB_LOG_TAB2_LIMBS * FLINT_BITS - 16)
{
slong exp;
/* just reading the table is known to give the correct rounding */
_arf_set_round_mpn(arb_midref(res), &exp, arb_log_log2_tab,
ARB_LOG_TAB2_LIMBS, 0, prec, ARF_RND_NEAR);
_fmpz_set_si_small(ARF_EXPREF(arb_midref(res)), exp);
/* 1/2 ulp error */
_fmpz_set_si_small(MAG_EXPREF(arb_radref(res)), exp - prec);
MAG_MAN(arb_radref(res)) = MAG_ONE_HALF;
}
else
{
arb_const_log2_hypgeom(res, prec);
}
}
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__ void
_mag_vec_get_fmpz_2exp_blocks(fmpz * coeffs,
double * dblcoeffs, fmpz * exps, slong * blocks, const fmpz_t scale,
arb_srcptr x, mag_srcptr xm, slong len)
{
fmpz_t top, bot, t, b, v, block_top, block_bot;
slong i, j, s, block, bits, maxheight;
int in_zero;
mag_srcptr cur;
fmpz_init(top);
fmpz_init(bot);
fmpz_init(t);
fmpz_init(b);
fmpz_init(v);
fmpz_init(block_top);
fmpz_init(block_bot);
blocks[0] = 0;
block = 0;
in_zero = 1;
maxheight = ALPHA * MAG_BITS + BETA;
if (maxheight > DOUBLE_BLOCK_MAX_HEIGHT)
abort();
for (i = 0; i < len; i++)
{
cur = (x == NULL) ? (xm + i) : arb_radref(x + i);
/* Skip (must be zero, since we assume there are no Infs/NaNs). */
if (mag_is_special(cur))
continue;
/* Bottom and top exponent of current number */
bits = MAG_BITS;
fmpz_set(top, MAG_EXPREF(cur));
fmpz_submul_ui(top, scale, i);
fmpz_sub_ui(bot, top, bits);
/* Extend current block. */
if (in_zero)
{
fmpz_swap(block_top, top);
fmpz_swap(block_bot, bot);
}
else
{
fmpz_max(t, top, block_top);
fmpz_min(b, bot, block_bot);
fmpz_sub(v, t, b);
/* extend current block */
if (fmpz_cmp_ui(v, maxheight) < 0)
{
fmpz_swap(block_top, t);
fmpz_swap(block_bot, b);
}
else /* start new block */
{
/* write exponent for previous block */
fmpz_set(exps + block, block_bot);
block++;
blocks[block] = i;
fmpz_swap(block_top, top);
fmpz_swap(block_bot, bot);
}
}
in_zero = 0;
}
/* write exponent for last block */
fmpz_set(exps + block, block_bot);
/* end marker */
blocks[block + 1] = len;
/* write the block data */
for (i = 0; blocks[i] != len; i++)
{
for (j = blocks[i]; j < blocks[i + 1]; j++)
{
cur = (x == NULL) ? (xm + j) : arb_radref(x + j);
if (mag_is_special(cur))
{
fmpz_zero(coeffs + j);
dblcoeffs[j] = 0.0;
}
else
{
mp_limb_t man;
double c;
man = MAG_MAN(cur);
/* TODO: only write and use doubles when block is short? */
/* Divide by 2^(scale * j) */
fmpz_mul_ui(t, scale, j);
fmpz_sub(t, MAG_EXPREF(cur), t);
fmpz_sub_ui(t, t, MAG_BITS); /* bottom exponent */
s = _fmpz_sub_small(t, exps + i);
if (s < 0) abort(); /* Bug catcher */
fmpz_set_ui(coeffs + j, man);
fmpz_mul_2exp(coeffs + j, coeffs + j, s);
c = man;
c = ldexp(c, s - DOUBLE_BLOCK_SHIFT);
if (c < 1e-150 || c > 1e150) /* Bug catcher */
abort();
dblcoeffs[j] = c;
}
}
}
fmpz_clear(top);
fmpz_clear(bot);
fmpz_clear(t);
fmpz_clear(b);
fmpz_clear(v);
fmpz_clear(block_top);
fmpz_clear(block_bot);
}
int
arb_get_unique_fmpz(fmpz_t z, const arb_t x)
{
if (!arb_is_finite(x))
{
return 0;
}
else if (arb_is_exact(x))
{
/* x = b*2^e, e >= 0 */
if (arf_is_int(arb_midref(x)))
{
/* arf_get_fmpz aborts on overflow */
arf_get_fmpz(z, arb_midref(x), ARF_RND_DOWN);
return 1;
}
else
{
return 0;
}
}
/* if the radius is >= 1, there are at least two integers */
else if (mag_cmp_2exp_si(arb_radref(x), 0) >= 0)
{
return 0;
}
/* there are 0 or 1 integers if the radius is < 1 */
else
{
fmpz_t a, b, exp;
int res;
/* if the midpoint is exactly an integer, it is what we want */
if (arf_is_int(arb_midref(x)))
{
/* arf_get_fmpz aborts on overflow */
arf_get_fmpz(z, arb_midref(x), ARF_RND_DOWN);
return 1;
}
fmpz_init(a);
fmpz_init(b);
fmpz_init(exp);
/* if the radius is tiny, it can't be an integer */
arf_bot(a, arb_midref(x));
if (fmpz_cmp(a, MAG_EXPREF(arb_radref(x))) > 0)
{
res = 0;
}
else
{
arb_get_interval_fmpz_2exp(a, b, exp, x);
if (COEFF_IS_MPZ(*exp))
{
flint_printf("arb_get_unique_fmpz: input too large\n");
abort();
}
if (*exp >= 0)
{
res = fmpz_equal(a, b);
if (res)
{
fmpz_mul_2exp(a, a, *exp);
fmpz_mul_2exp(b, b, *exp);
}
}
else
{
fmpz_cdiv_q_2exp(a, a, -(*exp));
fmpz_fdiv_q_2exp(b, b, -(*exp));
res = fmpz_equal(a, b);
}
if (res)
fmpz_set(z, a);
}
fmpz_clear(a);
fmpz_clear(b);
fmpz_clear(exp);
return res;
}
}
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);
}
/* 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);
}
}
}
