def _get_best_taylor_coefficients(bs, rs, m):
extrap = _extrapolate(bs, rs, m)
mvec = np.arange(m)
if len(extrap) > 2:
all_coefs, all_errors = dea3(extrap[:-2], extrap[1:-1], extrap[2:])
steps = np.atleast_1d(rs[4:])[:, None] * mvec
coefs, info = _Limit._get_best_estimate(all_coefs, all_errors, steps, (m,))
errors = info.error_estimate
else:
errors = EPS / np.power(rs[2], mvec) * np.maximum(m1, m2)
coefs = extrap[-1]
return coefs, errors
def _get_best_taylor_coefficients(bs, rs, m, max_m1m2):
extrap = _extrapolate(bs, rs, m)
mvec = np.arange(m)
if len(extrap) > 2:
all_coefs, all_errors = dea3(extrap[:-2], extrap[1:-1], extrap[2:])
steps = np.atleast_1d(rs[4:])[:, None] * mvec
# pylint: disable=protected-access
coefs, info = _Limit._get_best_estimate(all_coefs, all_errors, steps, (m,))
errors = info.error_estimate
else:
errors = EPS / np.power(rs[2], mvec) * max_m1m2()
coefs = extrap[-1]
return coefs, errors
def taylor(fun, z0=0, n=1, r=0.0061, num_extrap=3, step_ratio=1.6, **kwds):
"""
Return Taylor coefficients of complex analytic function using FFT
Parameters
----------
fun : callable
function to differentiate
z0 : real or complex scalar at which to evaluate the derivatives
n : scalar integer, default 1
Number of taylor coefficents to compute. Maximum number is 100.
r : real scalar, default 0.0061
Initial radius at which to evaluate. For well-behaved functions,
the computation should be insensitive to the initial radius to within
about four orders of magnitude.
num_extrap : scalar integer, default 3
number of extrapolation steps used in the calculation
step_ratio : real scalar, default 1.6
Initial grow/shrinking factor for finding the best radius.
max_iter : scalar integer, default 30
Maximum number of iterations
min_iter : scalar integer, default max_iter // 2
Minimum number of iterations before the solution may be deemed
degenerate. A larger number allows the algorithm to correct a bad
initial radius.
full_output : bool, optional
If `full_output` is False, only the coefficents is returned (default).
If `full_output` is True, then (coefs, status) is returned
Returns
-------
coefs : ndarray
array of taylor coefficents
status: Optional object into which output information is written:
degenerate: True if the algorithm was unable to bound the error
iterations: Number of iterations executed
function_count: Number of function calls
final_radius: Ending radius of the algorithm
failed: True if the maximum number of iterations was reached
error_estimate: approximate bounds of the rounding error.
This module uses the method of Fornberg to compute the Taylor series
coefficents of a complex analytic function along with error bounds. The
method uses a Fast Fourier Transform to invert function evaluations around
a circle into Taylor series coefficients and uses Richardson Extrapolation
to improve and bound the estimate. Unlike real-valued finite differences,
the method searches for a desirable radius and so is reasonably insensitive
to the initial radius-to within a number of orders of magnitude at least.
For most cases, the default configuration is likely to succeed.
Restrictions
The method uses the coefficients themselves to control the truncation error,
so the error will not be properly bounded for functions like low-order
polynomials whose Taylor series coefficients are nearly zero. If the error
cannot be bounded, degenerate flag will be set to true, and an answer will
still be computed and returned but should be used with caution.
Example
-------
Compute the first 6 taylor coefficients 1 / (1 - z) expanded round z0 = 0:
>>> import numdifftools.fornberg as ndf
>>> import numpy as np
>>> c, info = ndf.taylor(lambda x: 1./(1-x), z0=0, n=6, full_output=True)
>>> np.allclose(c, np.ones(8))
True
>>> np.all(info.error_estimate < 1e-9)
True
>>> (info.function_count, info.iterations, info.failed) == (144, 18, False)
True
References
----------
[1] Fornberg, B. (1981).
Numerical Differentiation of Analytic Functions.
ACM Transactions on Mathematical Software (TOMS),
7(4), 512-526. http://doi.org/10.1145/355972.355979
"""
max_iter = kwds.get('max_iter', 30)
min_iter = kwds.get('min_iter', max_iter // 2)
full_output = kwds.get('full_output', False)
direction_changes = 0
rs = []
bs = []
previous_direction = None
degenerate = failed = False
m = _num_taylor_coefficients(n)
mvec = np.arange(m)
# A factor for testing against the targeted geometric progression of
# FFT coefficients:
crat = m * (np.exp(np.log(1e-4) / (m - 1))) ** mvec
# Start iterating. The goal of this loops is to select a circle radius that
# yields a nice geometric progression of the coefficients (which controls
# the error), and then to accumulate *three* successive approximations as a
# function of the circle radius r so that we can perform Richardson
# Extrapolation and zero out error terms, *greatly* improving the quality
# of the approximation.
num_changes = 0
i = 0
for i in range(max_iter):
# print('r = %g' % (r))
bn = np.fft.fft(fun(_circle(z0, r, m))) / m
bs.append(bn * np.power(r, -mvec))
rs.append(r)
if direction_changes > 1 or degenerate:
num_changes += 1
# if len(rs) >= 3:
if num_changes >= 1 + num_extrap:
break
if not degenerate:
# If not degenerate, check for geometric progression in the fourier
# transform:
bnc = bn / crat
m1 = np.max(np.abs(bnc[:m // 2]))
m2 = np.max(np.abs(bnc[m // 2:]))
# If there's an extreme mismatch, then we can consider the
# geometric progression degenerate, whether one way or the other,
# and just alternate directions instead of trying to target a
# specific error bound (not ideal, but not a good reason to fail
# catastrophically):
#
# Note: only consider it degenerate if we've had a chance to steer
# the radius in the direction at least `min_iter` times:
degenerate = i > min_iter and (m1 / m2 < 1e-8 or m2 / m1 < 1e-8)
if degenerate:
needs_smaller = i % 2 == 0
else:
needs_smaller = (m1 != m1 or m2 != m2 or m1 < m2 or
_poor_convergence(z0, r, fun, bn, mvec))
if (previous_direction is not None and
needs_smaller != previous_direction):
direction_changes += 1
if direction_changes > 0:
# Once we've started changing directions, we've found our range so
# start taking the square root of the growth factor so that
# richardson extrapolation is well-behaved:
step_ratio = np.sqrt(step_ratio)
if needs_smaller:
r /= step_ratio
else:
r *= step_ratio
previous_direction = needs_smaller
else:
failed = True
extrap = _extrapolate(bs, rs, m)
if len(extrap) > 2:
all_coefs, all_errors = dea3(extrap[:-2], extrap[1:-1], extrap[2:])
coefs, info = _Limit._get_best_estimate(all_coefs, all_errors,
np.atleast_1d(rs[4:])[:, None]*mvec, (m,))
errors = info.error_estimate
else:
errors = EPS / np.power(rs[2], mvec) * np.maximum(m1, m2)
coefs = extrap[-1]
if full_output:
info = _INFO(errors, degenerate, final_radius=r,
function_count=i * m, iterations=i, failed=failed)
return coefs, info
return coefs
