int
_arb_poly_newton_step(arb_t xnew, arb_srcptr poly, long len,
const arb_t x,
const arb_t convergence_interval,
const arf_t convergence_factor, long prec)
{
arf_t err;
arb_t t, u, v;
int result;
arf_init(err);
arb_init(t);
arb_init(u);
arb_init(v);
arf_set_mag(err, arb_radref(x));
arf_mul(err, err, err, MAG_BITS, ARF_RND_UP);
arf_mul(err, err, convergence_factor, MAG_BITS, ARF_RND_UP);
arf_set(arb_midref(t), arb_midref(x));
mag_zero(arb_radref(t));
_arb_poly_evaluate2(u, v, poly, len, t, prec);
arb_div(u, u, v, prec);
arb_sub(u, t, u, prec);
arb_add_error_arf(u, err);
if (arb_contains(convergence_interval, u) &&
(mag_cmp(arb_radref(u), arb_radref(x)) < 0))
{
arb_swap(xnew, u);
result = 1;
}
else
{
arb_set(xnew, x);
result = 0;
}
arb_clear(t);
arb_clear(u);
arb_clear(v);
arf_clear(err);
return result;
}
void arb_square(arb_t out, const arb_t in, slong prec) {
mag_mul(arb_radref(out), arb_radref(in), arb_radref(in));
int inexact = arf_mul(arb_midref(out), arb_midref(in), arb_midref(in), ARB_RND, prec);
if (inexact)
arf_mag_add_ulp(arb_radref(out), arb_radref(out), arb_midref(out), prec);
}
int main()
{
slong iter;
flint_rand_t state;
flint_printf("get_mag....");
fflush(stdout);
flint_randinit(state);
for (iter = 0; iter < 10000 * arb_test_multiplier(); iter++)
{
acb_t a;
arf_t m2, x, y, s;
mag_t m;
acb_init(a);
mag_init(m);
arf_init(m2);
arf_init(x);
arf_init(y);
arf_init(s);
acb_randtest_special(a, state, 200, 10);
acb_get_mag(m, a);
MAG_CHECK_BITS(m)
/* check m^2 >= x^2 + y^2 */
arf_set_mag(m2, m);
arf_mul(m2, m2, m2, ARF_PREC_EXACT, ARF_RND_DOWN);
arb_get_abs_ubound_arf(x, acb_realref(a), ARF_PREC_EXACT);
arb_get_abs_ubound_arf(y, acb_imagref(a), ARF_PREC_EXACT);
arf_sosq(s, x, y, ARF_PREC_EXACT, ARF_RND_DOWN);
if (arf_cmp(m2, s) < 0)
{
flint_printf("FAIL:\n\n");
flint_printf("a = "); acb_print(a); flint_printf("\n\n");
flint_printf("m = "); mag_print(m); flint_printf("\n\n");
flint_abort();
}
acb_clear(a);
mag_clear(m);
arf_clear(m2);
arf_clear(x);
arf_clear(y);
arf_clear(s);
}
flint_randclear(state);
flint_cleanup();
flint_printf("PASS\n");
return EXIT_SUCCESS;
}
void arf_twobytwo_diag(arf_t u1, arf_t u2, const arf_t a, const arf_t b, const arf_t d, slong prec) {
// Compute the orthogonal matrix that diagonalizes
//
// A = [a b]
// [b d]
//
// This matrix will have the form
//
// U = [cos x , -sin x]
// [sin x, cos x]
//
// where the diagonal matrix is U^t A U.
// We set u1 = cos x, u2 = -sin x.
if(arf_is_zero(b)) {
arf_set_ui(u1, 1);
arf_set_ui(u2, 0);
return;
}
arf_t x; arf_init(x);
arf_mul(u1, b, b, prec, ARF_RND_NEAR); // u1 = b^2
arf_sub(u2, a, d, prec, ARF_RND_NEAR); // u2 = a - d
arf_mul_2exp_si(u2, u2, -1); // u2 = (a - d)/2
arf_mul(u2, u2, u2, prec, ARF_RND_NEAR); // u2 = ( (a - d)/2 )^2
arf_add(u1, u1, u2, prec, ARF_RND_NEAR); // u1 = b^2 + ( (a-d)/2 )^2
arf_sqrt(u1, u1, prec, ARF_RND_NEAR); // u1 = sqrt(above)
arf_mul_2exp_si(u1, u1, 1); // u1 = 2 (sqrt (above) )
arf_add(u1, u1, d, prec, ARF_RND_NEAR); // u1 += d
arf_sub(u1, u1, a, prec, ARF_RND_NEAR); // u1 -= a
arf_mul_2exp_si(u1, u1, -1); // u1 = (d - a)/2 + sqrt(b^2 + ( (a-d)/2 )^2)
arf_mul(x, u1, u1, prec, ARF_RND_NEAR);
arf_addmul(x, b, b, prec, ARF_RND_NEAR); // x = u1^2 + b^2
arf_sqrt(x, x, prec, ARF_RND_NEAR); // x = sqrt(u1^2 + b^2)
arf_div(u2, u1, x, prec, ARF_RND_NEAR);
arf_div(u1, b, x, prec, ARF_RND_NEAR);
arf_neg(u1, u1);
arf_clear(x);
}
int
arf_mul_fmpz_naive(arf_t z, const arf_t x, const fmpz_t y, slong prec, arf_rnd_t rnd)
{
arf_t t;
int r;
arf_init(t);
arf_set_fmpz(t, y);
r = arf_mul(z, x, t, prec, rnd);
arf_clear(t);
return r;
}
int
arf_addmul_naive(arf_t z, const arf_t x, const arf_t y, long prec, arf_rnd_t rnd)
{
arf_t t;
int inexact;
arf_init(t);
arf_mul(t, x, y, ARF_PREC_EXACT, ARF_RND_DOWN);
inexact = arf_add(z, z, t, prec, rnd);
arf_clear(t);
return inexact;
}
void
arb_mul_naive(arb_t z, const arb_t x, const arb_t y, slong prec)
{
arf_t zm_exact, zm_rounded, zr, t, u;
arf_init(zm_exact);
arf_init(zm_rounded);
arf_init(zr);
arf_init(t);
arf_init(u);
arf_mul(zm_exact, arb_midref(x), arb_midref(y), ARF_PREC_EXACT, ARF_RND_DOWN);
arf_set_round(zm_rounded, zm_exact, prec, ARB_RND);
/* rounding error */
if (arf_equal(zm_exact, zm_rounded))
{
arf_zero(zr);
}
else
{
fmpz_t e;
fmpz_init(e);
/* more accurate, but not what we are testing
arf_sub(zr, zm_exact, zm_rounded, MAG_BITS, ARF_RND_UP);
arf_abs(zr, zr); */
fmpz_sub_ui(e, ARF_EXPREF(zm_rounded), prec);
arf_one(zr);
arf_mul_2exp_fmpz(zr, zr, e);
fmpz_clear(e);
}
/* propagated error */
if (!arb_is_exact(x))
{
arf_set_mag(t, arb_radref(x));
arf_abs(u, arb_midref(y));
arf_addmul(zr, t, u, MAG_BITS, ARF_RND_UP);
}
if (!arb_is_exact(y))
{
arf_set_mag(t, arb_radref(y));
arf_abs(u, arb_midref(x));
arf_addmul(zr, t, u, MAG_BITS, ARF_RND_UP);
}
if (!arb_is_exact(x) && !arb_is_exact(y))
{
arf_set_mag(t, arb_radref(x));
arf_set_mag(u, arb_radref(y));
arf_addmul(zr, t, u, MAG_BITS, ARF_RND_UP);
}
arf_set(arb_midref(z), zm_rounded);
arf_get_mag(arb_radref(z), zr);
arf_clear(zm_exact);
arf_clear(zm_rounded);
arf_clear(zr);
arf_clear(t);
arf_clear(u);
}
int
arf_submul(arf_ptr z, arf_srcptr x, arf_srcptr y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn, zn, tn, alloc;
mp_srcptr xptr, yptr, zptr;
mp_ptr tptr, tptr2;
fmpz_t texp;
slong shift;
int tsgnbit, inexact;
ARF_MUL_TMP_DECL
if (arf_is_special(x) || arf_is_special(y) || arf_is_special(z))
{
if (arf_is_zero(z))
{
return arf_neg_mul(z, x, y, prec, rnd);
}
else if (arf_is_finite(x) && arf_is_finite(y))
{
return arf_set_round(z, z, prec, rnd);
}
else
{
/* todo: speed up */
arf_t t;
arf_init(t);
arf_mul(t, x, y, ARF_PREC_EXACT, ARF_RND_DOWN);
inexact = arf_sub(z, z, t, prec, rnd);
arf_clear(t);
return inexact;
}
}
tsgnbit = ARF_SGNBIT(x) ^ ARF_SGNBIT(y) ^ 1;
ARF_GET_MPN_READONLY(xptr, xn, x);
ARF_GET_MPN_READONLY(yptr, yn, y);
ARF_GET_MPN_READONLY(zptr, zn, z);
fmpz_init(texp);
_fmpz_add2_fast(texp, ARF_EXPREF(x), ARF_EXPREF(y), 0);
shift = _fmpz_sub_small(ARF_EXPREF(z), texp);
alloc = tn = xn + yn;
ARF_MUL_TMP_ALLOC(tptr2, alloc)
tptr = tptr2;
ARF_MPN_MUL(tptr, xptr, xn, yptr, yn);
tn -= (tptr[0] == 0);
tptr += (tptr[0] == 0);
if (shift >= 0)
inexact = _arf_add_mpn(z, zptr, zn, ARF_SGNBIT(z), ARF_EXPREF(z),
tptr, tn, tsgnbit, shift, prec, rnd);
else
inexact = _arf_add_mpn(z, tptr, tn, tsgnbit, texp,
zptr, zn, ARF_SGNBIT(z), -shift, prec, rnd);
ARF_MUL_TMP_FREE(tptr2, alloc)
fmpz_clear(texp);
return inexact;
}
int arb_mat_jacobi(arb_mat_t D, arb_mat_t P, const arb_mat_t A, slong prec) {
//
// Given a d x d real symmetric matrix A, compute an orthogonal matrix
// P and a diagonal D such that A = P D P^t = P D P^(-1).
//
// D should have already been initialized as a d x 1 matrix, and Pp
// should have already been initialized as a d x d matrix.
//
// If the eigenvalues can be certified as unique, then a nonzero int is
// returned, and the eigenvectors should have reasonable error bounds. If
// the eigenvalues cannot be certified as unique, then some of the
// eigenvectors will have infinite error radius.
#define B(i,j) arb_mat_entry(B, i, j)
#define D(i) arb_mat_entry(D, i, 0)
#define P(i,j) arb_mat_entry(P, i, j)
int dim = arb_mat_nrows(A);
if(dim == 1) {
arb_mat_set(D, A);
arb_mat_one(P);
return 0;
}
arb_mat_t B;
arb_mat_init(B, dim, dim);
arf_t * B1 = (arf_t*)malloc(dim * sizeof(arf_t));
arf_t * B2 = (arf_t*)malloc(dim * sizeof(arf_t));
arf_t * row_max = (arf_t*)malloc((dim - 1) * sizeof(arf_t));
int * row_max_indices = (int*)malloc((dim - 1) * sizeof(int));
for(int k = 0; k < dim; k++) {
arf_init(B1[k]);
arf_init(B2[k]);
}
for(int k = 0; k < dim - 1; k++) {
arf_init(row_max[k]);
}
arf_t x1, x2;
arf_init(x1);
arf_init(x2);
arf_t Gii, Gij, Gji, Gjj;
arf_init(Gii);
arf_init(Gij);
arf_init(Gji);
arf_init(Gjj);
arb_mat_set(B, A);
arb_mat_one(P);
for(int i = 0; i < dim - 1; i++) {
for(int j = i + 1; j < dim; j++) {
arf_abs(x1, arb_midref(B(i,j)));
if(arf_cmp(row_max[i], x1) < 0) {
arf_set(row_max[i], x1);
row_max_indices[i] = j;
}
}
}
int finished = 0;
while(!finished) {
arf_zero(x1);
int i = 0;
int j = 0;
for(int k = 0; k < dim - 1; k++) {
if(arf_cmp(x1, row_max[k]) < 0) {
arf_set(x1, row_max[k]);
i = k;
}
}
j = row_max_indices[i];
slong bound = arf_abs_bound_lt_2exp_si(x1);
if(bound < -prec * .9) {
finished = 1;
break;
}
else {
//printf("%ld\n", arf_abs_bound_lt_2exp_si(x1));
//arb_mat_printd(B, 10);
//printf("\n");
}
arf_twobytwo_diag(Gii, Gij, arb_midref(B(i,i)), arb_midref(B(i,j)), arb_midref(B(j,j)), 2*prec);
arf_neg(Gji, Gij);
arf_set(Gjj, Gii);
//printf("%d %d\n", i, j);
//arf_printd(Gii, 100);
//printf(" ");
//arf_printd(Gij, 100);
//printf("\n");
if(arf_is_zero(Gij)) { // If this happens, we're
finished = 1; // not going to do any better
break; // without increasing the precision.
}
for(int k = 0; k < dim; k++) {
arf_mul(B1[k], Gii, arb_midref(B(i,k)), prec, ARF_RND_NEAR);
arf_addmul(B1[k], Gji, arb_midref(B(j,k)), prec, ARF_RND_NEAR);
arf_mul(B2[k], Gij, arb_midref(B(i,k)), prec, ARF_RND_NEAR);
arf_addmul(B2[k], Gjj, arb_midref(B(j,k)), prec, ARF_RND_NEAR);
}
for(int k = 0; k < dim; k++) {
arf_set(arb_midref(B(i,k)), B1[k]);
arf_set(arb_midref(B(j,k)), B2[k]);
}
for(int k = 0; k < dim; k++) {
arf_mul(B1[k], Gii, arb_midref(B(k,i)), prec, ARF_RND_NEAR);
arf_addmul(B1[k], Gji, arb_midref(B(k,j)), prec, ARF_RND_NEAR);
arf_mul(B2[k], Gij, arb_midref(B(k,i)), prec, ARF_RND_NEAR);
arf_addmul(B2[k], Gjj, arb_midref(B(k,j)), prec, ARF_RND_NEAR);
}
for(int k = 0; k < dim; k++) {
arf_set(arb_midref(B(k,i)), B1[k]);
arf_set(arb_midref(B(k,j)), B2[k]);
}
for(int k = 0; k < dim; k++) {
arf_mul(B1[k], Gii, arb_midref(P(k,i)), prec, ARF_RND_NEAR);
arf_addmul(B1[k], Gji, arb_midref(P(k,j)), prec, ARF_RND_NEAR);
arf_mul(B2[k], Gij, arb_midref(P(k,i)), prec, ARF_RND_NEAR);
arf_addmul(B2[k], Gjj, arb_midref(P(k,j)), prec, ARF_RND_NEAR);
}
for(int k = 0; k < dim; k++) {
arf_set(arb_midref(P(k,i)), B1[k]);
arf_set(arb_midref(P(k,j)), B2[k]);
}
if(i < dim - 1)
arf_set_ui(row_max[i], 0);
if(j < dim - 1)
arf_set_ui(row_max[j], 0);
// Update the max in any row where the maximum
// was in a column that changed.
for(int k = 0; k < dim - 1; k++) {
if(row_max_indices[k] == j || row_max_indices[k] == i) {
arf_abs(row_max[k], arb_midref(B(k,k+1)));
row_max_indices[k] = k+1;
for(int l = k+2; l < dim; l++) {
arf_abs(x1, arb_midref(B(k,l)));
if(arf_cmp(row_max[k], x1) < 0) {
arf_set(row_max[k], x1);
row_max_indices[k] = l;
}
}
}
}
// Update the max in the ith row.
for(int k = i + 1; k < dim; k++) {
arf_abs(x1, arb_midref(B(i, k)));
if(arf_cmp(row_max[i], x1) < 0) {
arf_set(row_max[i], x1);
row_max_indices[i] = k;
}
}
// Update the max in the jth row.
for(int k = j + 1; k < dim; k++) {
arf_abs(x1, arb_midref(B(j, k)));
if(arf_cmp(row_max[j], x1) < 0) {
arf_set(row_max[j], x1);
row_max_indices[j] = k;
}
}
// Go through column i to see if any of
// the new entries are larger than the
// max of their row.
for(int k = 0; k < i; k++) {
if(k == dim) continue;
arf_abs(x1, arb_midref(B(k, i)));
if(arf_cmp(row_max[k], x1) < 0) {
arf_set(row_max[k], x1);
row_max_indices[k] = i;
}
}
// And then column j.
for(int k = 0; k < j; k++) {
if(k == dim) continue;
arf_abs(x1, arb_midref(B(k, j)));
if(arf_cmp(row_max[k], x1) < 0) {
arf_set(row_max[k], x1);
row_max_indices[k] = j;
}
}
}
for(int k = 0; k < dim; k++) {
arb_set(D(k), B(k,k));
arb_set_exact(D(k));
}
// At this point we've done that diagonalization and all that remains is
// to certify the correctness and compute error bounds.
arb_mat_t e;
arb_t error_norms[dim];
for(int k = 0; k < dim; k++) arb_init(error_norms[k]);
arb_mat_init(e, dim, 1);
arb_t z1, z2;
arb_init(z1);
arb_init(z2);
for(int j = 0; j < dim; j++) {
arb_mat_set(B, A);
for(int k = 0; k < dim; k++) {
arb_sub(B(k, k), B(k, k), D(j), prec);
}
for(int k = 0; k < dim; k++) {
arb_set(arb_mat_entry(e, k, 0), P(k, j));
}
arb_mat_L2norm(z2, e, prec);
arb_mat_mul(e, B, e, prec);
arb_mat_L2norm(error_norms[j], e, prec);
arb_div(z2, error_norms[j], z2, prec); // and now z1 is an upper bound for the
// error in the eigenvalue
arb_add_error(D(j), z2);
}
int unique_eigenvalues = 1;
for(int j = 0; j < dim; j++) {
if(j == 0) {
arb_sub(z1, D(j), D(1), prec);
}
else {
arb_sub(z1, D(j), D(0), prec);
}
arb_get_abs_lbound_arf(x1, z1, prec);
for(int k = 1; k < dim; k++) {
if(k == j) continue;
arb_sub(z1, D(j), D(k), prec);
arb_get_abs_lbound_arf(x2, z1, prec);
if(arf_cmp(x2, x1) < 0) {
arf_set(x1, x2);
}
}
if(arf_is_zero(x1)) {
unique_eigenvalues = 0;
}
arb_div_arf(z1, error_norms[j], x1, prec);
for(int k = 0; k < dim; k++) {
arb_add_error(P(k, j), z1);
}
}
arb_mat_clear(e);
arb_clear(z1);
arb_clear(z2);
for(int k = 0; k < dim; k++) arb_clear(error_norms[k]);
arf_clear(x1);
arf_clear(x2);
arb_mat_clear(B);
for(int k = 0; k < dim; k++) {
arf_clear(B1[k]);
arf_clear(B2[k]);
}
for(int k = 0; k < dim - 1; k++) {
arf_clear(row_max[k]);
}
arf_clear(Gii);
arf_clear(Gij);
arf_clear(Gji);
arf_clear(Gjj);
free(B1);
free(B2);
free(row_max);
free(row_max_indices);
if(unique_eigenvalues) return 0;
else return 1;
#undef B
#undef D
#undef P
}
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;
}
}
int
acb_calc_integrate_taylor(acb_t res,
acb_calc_func_t func, void * param,
const acb_t a, const acb_t b,
const arf_t inner_radius,
const arf_t outer_radius,
long accuracy_goal, long prec)
{
long num_steps, step, N, bp;
int result;
acb_t delta, m, x, y1, y2, sum;
acb_ptr taylor_poly;
arf_t err;
acb_init(delta);
acb_init(m);
acb_init(x);
acb_init(y1);
acb_init(y2);
acb_init(sum);
arf_init(err);
acb_sub(delta, b, a, prec);
/* precision used for bounds calculations */
bp = MAG_BITS;
/* compute the number of steps */
{
arf_t t;
arf_init(t);
acb_get_abs_ubound_arf(t, delta, bp);
arf_div(t, t, inner_radius, bp, ARF_RND_UP);
arf_mul_2exp_si(t, t, -1);
num_steps = (long) (arf_get_d(t, ARF_RND_UP) + 1.0);
/* make sure it's not something absurd */
num_steps = FLINT_MIN(num_steps, 10 * prec);
num_steps = FLINT_MAX(num_steps, 1);
arf_clear(t);
}
result = ARB_CALC_SUCCESS;
acb_zero(sum);
for (step = 0; step < num_steps; step++)
{
/* midpoint of subinterval */
acb_mul_ui(m, delta, 2 * step + 1, prec);
acb_div_ui(m, m, 2 * num_steps, prec);
acb_add(m, m, a, prec);
if (arb_calc_verbose)
{
printf("integration point %ld/%ld: ", 2 * step + 1, 2 * num_steps);
acb_printd(m, 15); printf("\n");
}
/* evaluate at +/- x */
/* TODO: exactify m, and include error in x? */
acb_div_ui(x, delta, 2 * num_steps, prec);
/* compute bounds and number of terms to use */
{
arb_t cbound, xbound, rbound;
arf_t C, D, R, X, T;
double DD, TT, NN;
arb_init(cbound);
arb_init(xbound);
arb_init(rbound);
arf_init(C);
arf_init(D);
arf_init(R);
arf_init(X);
arf_init(T);
/* R is the outer radius */
arf_set(R, outer_radius);
/* X = upper bound for |x| */
acb_get_abs_ubound_arf(X, x, bp);
arb_set_arf(xbound, X);
/* Compute C(m,R). Important subtlety: due to rounding when
computing m, we will in general be farther than R away from
the integration path. But since acb_calc_cauchy_bound
actually integrates over the area traced by a complex
interval, it will catch any extra singularities (giving
an infinite bound). */
arb_set_arf(rbound, outer_radius);
acb_calc_cauchy_bound(cbound, func, param, m, rbound, 8, bp);
arf_set_mag(C, arb_radref(cbound));
arf_add(C, arb_midref(cbound), C, bp, ARF_RND_UP);
/* Sanity check: we need C < inf and R > X */
if (arf_is_finite(C) && arf_cmp(R, X) > 0)
{
/* Compute upper bound for D = C * R * X / (R - X) */
arf_mul(D, C, R, bp, ARF_RND_UP);
arf_mul(D, D, X, bp, ARF_RND_UP);
arf_sub(T, R, X, bp, ARF_RND_DOWN);
arf_div(D, D, T, bp, ARF_RND_UP);
/* Compute upper bound for T = (X / R) */
arf_div(T, X, R, bp, ARF_RND_UP);
/* Choose N */
/* TODO: use arf arithmetic to avoid overflow */
/* TODO: use relative accuracy (look at |f(m)|?) */
DD = arf_get_d(D, ARF_RND_UP);
TT = arf_get_d(T, ARF_RND_UP);
NN = -(accuracy_goal * 0.69314718055994530942 + log(DD)) / log(TT);
N = NN + 0.5;
N = FLINT_MIN(N, 100 * prec);
N = FLINT_MAX(N, 1);
/* Tail bound: D / (N + 1) * T^N */
{
mag_t TT;
mag_init(TT);
arf_get_mag(TT, T);
mag_pow_ui(TT, TT, N);
arf_set_mag(T, TT);
mag_clear(TT);
}
arf_mul(D, D, T, bp, ARF_RND_UP);
arf_div_ui(err, D, N + 1, bp, ARF_RND_UP);
}
else
{
N = 1;
arf_pos_inf(err);
result = ARB_CALC_NO_CONVERGENCE;
}
if (arb_calc_verbose)
{
printf("N = %ld; bound: ", N); arf_printd(err, 15); printf("\n");
printf("R: "); arf_printd(R, 15); printf("\n");
printf("C: "); arf_printd(C, 15); printf("\n");
printf("X: "); arf_printd(X, 15); printf("\n");
}
arb_clear(cbound);
arb_clear(xbound);
arb_clear(rbound);
arf_clear(C);
arf_clear(D);
arf_clear(R);
arf_clear(X);
arf_clear(T);
}
/* evaluate Taylor polynomial */
taylor_poly = _acb_vec_init(N + 1);
func(taylor_poly, m, param, N, prec);
_acb_poly_integral(taylor_poly, taylor_poly, N + 1, prec);
_acb_poly_evaluate(y2, taylor_poly, N + 1, x, prec);
acb_neg(x, x);
_acb_poly_evaluate(y1, taylor_poly, N + 1, x, prec);
acb_neg(x, x);
/* add truncation error */
arb_add_error_arf(acb_realref(y1), err);
arb_add_error_arf(acb_imagref(y1), err);
arb_add_error_arf(acb_realref(y2), err);
arb_add_error_arf(acb_imagref(y2), err);
acb_add(sum, sum, y2, prec);
acb_sub(sum, sum, y1, prec);
if (arb_calc_verbose)
{
printf("values: ");
acb_printd(y1, 15); printf(" ");
acb_printd(y2, 15); printf("\n");
}
_acb_vec_clear(taylor_poly, N + 1);
if (result == ARB_CALC_NO_CONVERGENCE)
break;
}
acb_set(res, sum);
acb_clear(delta);
acb_clear(m);
acb_clear(x);
acb_clear(y1);
acb_clear(y2);
acb_clear(sum);
arf_clear(err);
return result;
}
