You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

837 lines
32 KiB
Python

#
# Created by: Pearu Peterson, March 2002
#
""" Test functions for linalg.matfuncs module
"""
from __future__ import division, print_function, absolute_import
import random
import functools
import numpy as np
from numpy import array, matrix, identity, dot, sqrt, double
from numpy.testing import (
assert_array_equal, assert_array_less, assert_equal,
assert_array_almost_equal, assert_array_almost_equal_nulp,
assert_allclose, assert_)
import pytest
from scipy._lib._numpy_compat import _assert_warns, suppress_warnings
import scipy.linalg
from scipy.linalg import (funm, signm, logm, sqrtm, fractional_matrix_power,
expm, expm_frechet, expm_cond, norm)
from scipy.linalg import _matfuncs_inv_ssq
import scipy.linalg._expm_frechet
from scipy.optimize import minimize
def _get_al_mohy_higham_2012_experiment_1():
"""
Return the test matrix from Experiment (1) of [1]_.
References
----------
.. [1] Awad H. Al-Mohy and Nicholas J. Higham (2012)
"Improved Inverse Scaling and Squaring Algorithms
for the Matrix Logarithm."
SIAM Journal on Scientific Computing, 34 (4). C152-C169.
ISSN 1095-7197
"""
A = np.array([
[3.2346e-1, 3e4, 3e4, 3e4],
[0, 3.0089e-1, 3e4, 3e4],
[0, 0, 3.2210e-1, 3e4],
[0, 0, 0, 3.0744e-1]], dtype=float)
return A
class TestSignM(object):
def test_nils(self):
a = array([[29.2, -24.2, 69.5, 49.8, 7.],
[-9.2, 5.2, -18., -16.8, -2.],
[-10., 6., -20., -18., -2.],
[-9.6, 9.6, -25.5, -15.4, -2.],
[9.8, -4.8, 18., 18.2, 2.]])
cr = array([[11.94933333,-2.24533333,15.31733333,21.65333333,-2.24533333],
[-3.84266667,0.49866667,-4.59066667,-7.18666667,0.49866667],
[-4.08,0.56,-4.92,-7.6,0.56],
[-4.03466667,1.04266667,-5.59866667,-7.02666667,1.04266667],
[4.15733333,-0.50133333,4.90933333,7.81333333,-0.50133333]])
r = signm(a)
assert_array_almost_equal(r,cr)
def test_defective1(self):
a = array([[0.0,1,0,0],[1,0,1,0],[0,0,0,1],[0,0,1,0]])
r = signm(a, disp=False)
#XXX: what would be the correct result?
def test_defective2(self):
a = array((
[29.2,-24.2,69.5,49.8,7.0],
[-9.2,5.2,-18.0,-16.8,-2.0],
[-10.0,6.0,-20.0,-18.0,-2.0],
[-9.6,9.6,-25.5,-15.4,-2.0],
[9.8,-4.8,18.0,18.2,2.0]))
r = signm(a, disp=False)
#XXX: what would be the correct result?
def test_defective3(self):
a = array([[-2., 25., 0., 0., 0., 0., 0.],
[0., -3., 10., 3., 3., 3., 0.],
[0., 0., 2., 15., 3., 3., 0.],
[0., 0., 0., 0., 15., 3., 0.],
[0., 0., 0., 0., 3., 10., 0.],
[0., 0., 0., 0., 0., -2., 25.],
[0., 0., 0., 0., 0., 0., -3.]])
r = signm(a, disp=False)
#XXX: what would be the correct result?
class TestLogM(object):
def test_nils(self):
a = array([[-2., 25., 0., 0., 0., 0., 0.],
[0., -3., 10., 3., 3., 3., 0.],
[0., 0., 2., 15., 3., 3., 0.],
[0., 0., 0., 0., 15., 3., 0.],
[0., 0., 0., 0., 3., 10., 0.],
[0., 0., 0., 0., 0., -2., 25.],
[0., 0., 0., 0., 0., 0., -3.]])
m = (identity(7)*3.1+0j)-a
logm(m, disp=False)
#XXX: what would be the correct result?
def test_al_mohy_higham_2012_experiment_1_logm(self):
# The logm completes the round trip successfully.
# Note that the expm leg of the round trip is badly conditioned.
A = _get_al_mohy_higham_2012_experiment_1()
A_logm, info = logm(A, disp=False)
A_round_trip = expm(A_logm)
assert_allclose(A_round_trip, A, rtol=1e-5, atol=1e-14)
def test_al_mohy_higham_2012_experiment_1_funm_log(self):
# The raw funm with np.log does not complete the round trip.
# Note that the expm leg of the round trip is badly conditioned.
A = _get_al_mohy_higham_2012_experiment_1()
A_funm_log, info = funm(A, np.log, disp=False)
A_round_trip = expm(A_funm_log)
assert_(not np.allclose(A_round_trip, A, rtol=1e-5, atol=1e-14))
def test_round_trip_random_float(self):
np.random.seed(1234)
for n in range(1, 6):
M_unscaled = np.random.randn(n, n)
for scale in np.logspace(-4, 4, 9):
M = M_unscaled * scale
# Eigenvalues are related to the branch cut.
W = np.linalg.eigvals(M)
err_msg = 'M:{0} eivals:{1}'.format(M, W)
# Check sqrtm round trip because it is used within logm.
M_sqrtm, info = sqrtm(M, disp=False)
M_sqrtm_round_trip = M_sqrtm.dot(M_sqrtm)
assert_allclose(M_sqrtm_round_trip, M)
# Check logm round trip.
M_logm, info = logm(M, disp=False)
M_logm_round_trip = expm(M_logm)
assert_allclose(M_logm_round_trip, M, err_msg=err_msg)
def test_round_trip_random_complex(self):
np.random.seed(1234)
for n in range(1, 6):
M_unscaled = np.random.randn(n, n) + 1j * np.random.randn(n, n)
for scale in np.logspace(-4, 4, 9):
M = M_unscaled * scale
M_logm, info = logm(M, disp=False)
M_round_trip = expm(M_logm)
assert_allclose(M_round_trip, M)
def test_logm_type_preservation_and_conversion(self):
# The logm matrix function should preserve the type of a matrix
# whose eigenvalues are positive with zero imaginary part.
# Test this preservation for variously structured matrices.
complex_dtype_chars = ('F', 'D', 'G')
for matrix_as_list in (
[[1, 0], [0, 1]],
[[1, 0], [1, 1]],
[[2, 1], [1, 1]],
[[2, 3], [1, 2]]):
# check that the spectrum has the expected properties
W = scipy.linalg.eigvals(matrix_as_list)
assert_(not any(w.imag or w.real < 0 for w in W))
# check float type preservation
A = np.array(matrix_as_list, dtype=float)
A_logm, info = logm(A, disp=False)
assert_(A_logm.dtype.char not in complex_dtype_chars)
# check complex type preservation
A = np.array(matrix_as_list, dtype=complex)
A_logm, info = logm(A, disp=False)
assert_(A_logm.dtype.char in complex_dtype_chars)
# check float->complex type conversion for the matrix negation
A = -np.array(matrix_as_list, dtype=float)
A_logm, info = logm(A, disp=False)
assert_(A_logm.dtype.char in complex_dtype_chars)
def test_complex_spectrum_real_logm(self):
# This matrix has complex eigenvalues and real logm.
# Its output dtype depends on its input dtype.
M = [[1, 1, 2], [2, 1, 1], [1, 2, 1]]
for dt in float, complex:
X = np.array(M, dtype=dt)
w = scipy.linalg.eigvals(X)
assert_(1e-2 < np.absolute(w.imag).sum())
Y, info = logm(X, disp=False)
assert_(np.issubdtype(Y.dtype, np.inexact))
assert_allclose(expm(Y), X)
def test_real_mixed_sign_spectrum(self):
# These matrices have real eigenvalues with mixed signs.
# The output logm dtype is complex, regardless of input dtype.
for M in (
[[1, 0], [0, -1]],
[[0, 1], [1, 0]]):
for dt in float, complex:
A = np.array(M, dtype=dt)
A_logm, info = logm(A, disp=False)
assert_(np.issubdtype(A_logm.dtype, np.complexfloating))
def test_exactly_singular(self):
A = np.array([[0, 0], [1j, 1j]])
B = np.asarray([[1, 1], [0, 0]])
for M in A, A.T, B, B.T:
expected_warning = _matfuncs_inv_ssq.LogmExactlySingularWarning
L, info = _assert_warns(expected_warning, logm, M, disp=False)
E = expm(L)
assert_allclose(E, M, atol=1e-14)
def test_nearly_singular(self):
M = np.array([[1e-100]])
expected_warning = _matfuncs_inv_ssq.LogmNearlySingularWarning
L, info = _assert_warns(expected_warning, logm, M, disp=False)
E = expm(L)
assert_allclose(E, M, atol=1e-14)
def test_opposite_sign_complex_eigenvalues(self):
# See gh-6113
E = [[0, 1], [-1, 0]]
L = [[0, np.pi*0.5], [-np.pi*0.5, 0]]
assert_allclose(expm(L), E, atol=1e-14)
assert_allclose(logm(E), L, atol=1e-14)
E = [[1j, 4], [0, -1j]]
L = [[1j*np.pi*0.5, 2*np.pi], [0, -1j*np.pi*0.5]]
assert_allclose(expm(L), E, atol=1e-14)
assert_allclose(logm(E), L, atol=1e-14)
E = [[1j, 0], [0, -1j]]
L = [[1j*np.pi*0.5, 0], [0, -1j*np.pi*0.5]]
assert_allclose(expm(L), E, atol=1e-14)
assert_allclose(logm(E), L, atol=1e-14)
class TestSqrtM(object):
def test_round_trip_random_float(self):
np.random.seed(1234)
for n in range(1, 6):
M_unscaled = np.random.randn(n, n)
for scale in np.logspace(-4, 4, 9):
M = M_unscaled * scale
M_sqrtm, info = sqrtm(M, disp=False)
M_sqrtm_round_trip = M_sqrtm.dot(M_sqrtm)
assert_allclose(M_sqrtm_round_trip, M)
def test_round_trip_random_complex(self):
np.random.seed(1234)
for n in range(1, 6):
M_unscaled = np.random.randn(n, n) + 1j * np.random.randn(n, n)
for scale in np.logspace(-4, 4, 9):
M = M_unscaled * scale
M_sqrtm, info = sqrtm(M, disp=False)
M_sqrtm_round_trip = M_sqrtm.dot(M_sqrtm)
assert_allclose(M_sqrtm_round_trip, M)
def test_bad(self):
# See https://web.archive.org/web/20051220232650/http://www.maths.man.ac.uk/~nareports/narep336.ps.gz
e = 2**-5
se = sqrt(e)
a = array([[1.0,0,0,1],
[0,e,0,0],
[0,0,e,0],
[0,0,0,1]])
sa = array([[1,0,0,0.5],
[0,se,0,0],
[0,0,se,0],
[0,0,0,1]])
n = a.shape[0]
assert_array_almost_equal(dot(sa,sa),a)
# Check default sqrtm.
esa = sqrtm(a, disp=False, blocksize=n)[0]
assert_array_almost_equal(dot(esa,esa),a)
# Check sqrtm with 2x2 blocks.
esa = sqrtm(a, disp=False, blocksize=2)[0]
assert_array_almost_equal(dot(esa,esa),a)
def test_sqrtm_type_preservation_and_conversion(self):
# The sqrtm matrix function should preserve the type of a matrix
# whose eigenvalues are nonnegative with zero imaginary part.
# Test this preservation for variously structured matrices.
complex_dtype_chars = ('F', 'D', 'G')
for matrix_as_list in (
[[1, 0], [0, 1]],
[[1, 0], [1, 1]],
[[2, 1], [1, 1]],
[[2, 3], [1, 2]],
[[1, 1], [1, 1]]):
# check that the spectrum has the expected properties
W = scipy.linalg.eigvals(matrix_as_list)
assert_(not any(w.imag or w.real < 0 for w in W))
# check float type preservation
A = np.array(matrix_as_list, dtype=float)
A_sqrtm, info = sqrtm(A, disp=False)
assert_(A_sqrtm.dtype.char not in complex_dtype_chars)
# check complex type preservation
A = np.array(matrix_as_list, dtype=complex)
A_sqrtm, info = sqrtm(A, disp=False)
assert_(A_sqrtm.dtype.char in complex_dtype_chars)
# check float->complex type conversion for the matrix negation
A = -np.array(matrix_as_list, dtype=float)
A_sqrtm, info = sqrtm(A, disp=False)
assert_(A_sqrtm.dtype.char in complex_dtype_chars)
def test_sqrtm_type_conversion_mixed_sign_or_complex_spectrum(self):
complex_dtype_chars = ('F', 'D', 'G')
for matrix_as_list in (
[[1, 0], [0, -1]],
[[0, 1], [1, 0]],
[[0, 1, 0], [0, 0, 1], [1, 0, 0]]):
# check that the spectrum has the expected properties
W = scipy.linalg.eigvals(matrix_as_list)
assert_(any(w.imag or w.real < 0 for w in W))
# check complex->complex
A = np.array(matrix_as_list, dtype=complex)
A_sqrtm, info = sqrtm(A, disp=False)
assert_(A_sqrtm.dtype.char in complex_dtype_chars)
# check float->complex
A = np.array(matrix_as_list, dtype=float)
A_sqrtm, info = sqrtm(A, disp=False)
assert_(A_sqrtm.dtype.char in complex_dtype_chars)
def test_blocksizes(self):
# Make sure I do not goof up the blocksizes when they do not divide n.
np.random.seed(1234)
for n in range(1, 8):
A = np.random.rand(n, n) + 1j*np.random.randn(n, n)
A_sqrtm_default, info = sqrtm(A, disp=False, blocksize=n)
assert_allclose(A, np.linalg.matrix_power(A_sqrtm_default, 2))
for blocksize in range(1, 10):
A_sqrtm_new, info = sqrtm(A, disp=False, blocksize=blocksize)
assert_allclose(A_sqrtm_default, A_sqrtm_new)
def test_al_mohy_higham_2012_experiment_1(self):
# Matrix square root of a tricky upper triangular matrix.
A = _get_al_mohy_higham_2012_experiment_1()
A_sqrtm, info = sqrtm(A, disp=False)
A_round_trip = A_sqrtm.dot(A_sqrtm)
assert_allclose(A_round_trip, A, rtol=1e-5)
assert_allclose(np.tril(A_round_trip), np.tril(A))
def test_strict_upper_triangular(self):
# This matrix has no square root.
for dt in int, float:
A = np.array([
[0, 3, 0, 0],
[0, 0, 3, 0],
[0, 0, 0, 3],
[0, 0, 0, 0]], dtype=dt)
A_sqrtm, info = sqrtm(A, disp=False)
assert_(np.isnan(A_sqrtm).all())
def test_weird_matrix(self):
# The square root of matrix B exists.
for dt in int, float:
A = np.array([
[0, 0, 1],
[0, 0, 0],
[0, 1, 0]], dtype=dt)
B = np.array([
[0, 1, 0],
[0, 0, 0],
[0, 0, 0]], dtype=dt)
assert_array_equal(B, A.dot(A))
# But scipy sqrtm is not clever enough to find it.
B_sqrtm, info = sqrtm(B, disp=False)
assert_(np.isnan(B_sqrtm).all())
def test_disp(self):
from io import StringIO
np.random.seed(1234)
A = np.random.rand(3, 3)
B = sqrtm(A, disp=True)
assert_allclose(B.dot(B), A)
def test_opposite_sign_complex_eigenvalues(self):
M = [[2j, 4], [0, -2j]]
R = [[1+1j, 2], [0, 1-1j]]
assert_allclose(np.dot(R, R), M, atol=1e-14)
assert_allclose(sqrtm(M), R, atol=1e-14)
def test_gh4866(self):
M = np.array([[1, 0, 0, 1],
[0, 0, 0, 0],
[0, 0, 0, 0],
[1, 0, 0, 1]])
R = np.array([[sqrt(0.5), 0, 0, sqrt(0.5)],
[0, 0, 0, 0],
[0, 0, 0, 0],
[sqrt(0.5), 0, 0, sqrt(0.5)]])
assert_allclose(np.dot(R, R), M, atol=1e-14)
assert_allclose(sqrtm(M), R, atol=1e-14)
def test_gh5336(self):
M = np.diag([2, 1, 0])
R = np.diag([sqrt(2), 1, 0])
assert_allclose(np.dot(R, R), M, atol=1e-14)
assert_allclose(sqrtm(M), R, atol=1e-14)
def test_gh7839(self):
M = np.zeros((2, 2))
R = np.zeros((2, 2))
assert_allclose(np.dot(R, R), M, atol=1e-14)
assert_allclose(sqrtm(M), R, atol=1e-14)
class TestFractionalMatrixPower(object):
def test_round_trip_random_complex(self):
np.random.seed(1234)
for p in range(1, 5):
for n in range(1, 5):
M_unscaled = np.random.randn(n, n) + 1j * np.random.randn(n, n)
for scale in np.logspace(-4, 4, 9):
M = M_unscaled * scale
M_root = fractional_matrix_power(M, 1/p)
M_round_trip = np.linalg.matrix_power(M_root, p)
assert_allclose(M_round_trip, M)
def test_round_trip_random_float(self):
# This test is more annoying because it can hit the branch cut;
# this happens when the matrix has an eigenvalue
# with no imaginary component and with a real negative component,
# and it means that the principal branch does not exist.
np.random.seed(1234)
for p in range(1, 5):
for n in range(1, 5):
M_unscaled = np.random.randn(n, n)
for scale in np.logspace(-4, 4, 9):
M = M_unscaled * scale
M_root = fractional_matrix_power(M, 1/p)
M_round_trip = np.linalg.matrix_power(M_root, p)
assert_allclose(M_round_trip, M)
def test_larger_abs_fractional_matrix_powers(self):
np.random.seed(1234)
for n in (2, 3, 5):
for i in range(10):
M = np.random.randn(n, n) + 1j * np.random.randn(n, n)
M_one_fifth = fractional_matrix_power(M, 0.2)
# Test the round trip.
M_round_trip = np.linalg.matrix_power(M_one_fifth, 5)
assert_allclose(M, M_round_trip)
# Test a large abs fractional power.
X = fractional_matrix_power(M, -5.4)
Y = np.linalg.matrix_power(M_one_fifth, -27)
assert_allclose(X, Y)
# Test another large abs fractional power.
X = fractional_matrix_power(M, 3.8)
Y = np.linalg.matrix_power(M_one_fifth, 19)
assert_allclose(X, Y)
def test_random_matrices_and_powers(self):
# Each independent iteration of this fuzz test picks random parameters.
# It tries to hit some edge cases.
np.random.seed(1234)
nsamples = 20
for i in range(nsamples):
# Sample a matrix size and a random real power.
n = random.randrange(1, 5)
p = np.random.randn()
# Sample a random real or complex matrix.
matrix_scale = np.exp(random.randrange(-4, 5))
A = np.random.randn(n, n)
if random.choice((True, False)):
A = A + 1j * np.random.randn(n, n)
A = A * matrix_scale
# Check a couple of analytically equivalent ways
# to compute the fractional matrix power.
# These can be compared because they both use the principal branch.
A_power = fractional_matrix_power(A, p)
A_logm, info = logm(A, disp=False)
A_power_expm_logm = expm(A_logm * p)
assert_allclose(A_power, A_power_expm_logm)
def test_al_mohy_higham_2012_experiment_1(self):
# Fractional powers of a tricky upper triangular matrix.
A = _get_al_mohy_higham_2012_experiment_1()
# Test remainder matrix power.
A_funm_sqrt, info = funm(A, np.sqrt, disp=False)
A_sqrtm, info = sqrtm(A, disp=False)
A_rem_power = _matfuncs_inv_ssq._remainder_matrix_power(A, 0.5)
A_power = fractional_matrix_power(A, 0.5)
assert_array_equal(A_rem_power, A_power)
assert_allclose(A_sqrtm, A_power)
assert_allclose(A_sqrtm, A_funm_sqrt)
# Test more fractional powers.
for p in (1/2, 5/3):
A_power = fractional_matrix_power(A, p)
A_round_trip = fractional_matrix_power(A_power, 1/p)
assert_allclose(A_round_trip, A, rtol=1e-2)
assert_allclose(np.tril(A_round_trip, 1), np.tril(A, 1))
def test_briggs_helper_function(self):
np.random.seed(1234)
for a in np.random.randn(10) + 1j * np.random.randn(10):
for k in range(5):
x_observed = _matfuncs_inv_ssq._briggs_helper_function(a, k)
x_expected = a ** np.exp2(-k) - 1
assert_allclose(x_observed, x_expected)
def test_type_preservation_and_conversion(self):
# The fractional_matrix_power matrix function should preserve
# the type of a matrix whose eigenvalues
# are positive with zero imaginary part.
# Test this preservation for variously structured matrices.
complex_dtype_chars = ('F', 'D', 'G')
for matrix_as_list in (
[[1, 0], [0, 1]],
[[1, 0], [1, 1]],
[[2, 1], [1, 1]],
[[2, 3], [1, 2]]):
# check that the spectrum has the expected properties
W = scipy.linalg.eigvals(matrix_as_list)
assert_(not any(w.imag or w.real < 0 for w in W))
# Check various positive and negative powers
# with absolute values bigger and smaller than 1.
for p in (-2.4, -0.9, 0.2, 3.3):
# check float type preservation
A = np.array(matrix_as_list, dtype=float)
A_power = fractional_matrix_power(A, p)
assert_(A_power.dtype.char not in complex_dtype_chars)
# check complex type preservation
A = np.array(matrix_as_list, dtype=complex)
A_power = fractional_matrix_power(A, p)
assert_(A_power.dtype.char in complex_dtype_chars)
# check float->complex for the matrix negation
A = -np.array(matrix_as_list, dtype=float)
A_power = fractional_matrix_power(A, p)
assert_(A_power.dtype.char in complex_dtype_chars)
def test_type_conversion_mixed_sign_or_complex_spectrum(self):
complex_dtype_chars = ('F', 'D', 'G')
for matrix_as_list in (
[[1, 0], [0, -1]],
[[0, 1], [1, 0]],
[[0, 1, 0], [0, 0, 1], [1, 0, 0]]):
# check that the spectrum has the expected properties
W = scipy.linalg.eigvals(matrix_as_list)
assert_(any(w.imag or w.real < 0 for w in W))
# Check various positive and negative powers
# with absolute values bigger and smaller than 1.
for p in (-2.4, -0.9, 0.2, 3.3):
# check complex->complex
A = np.array(matrix_as_list, dtype=complex)
A_power = fractional_matrix_power(A, p)
assert_(A_power.dtype.char in complex_dtype_chars)
# check float->complex
A = np.array(matrix_as_list, dtype=float)
A_power = fractional_matrix_power(A, p)
assert_(A_power.dtype.char in complex_dtype_chars)
@pytest.mark.xfail(reason='Too unstable across LAPACKs.')
def test_singular(self):
# Negative fractional powers do not work with singular matrices.
for matrix_as_list in (
[[0, 0], [0, 0]],
[[1, 1], [1, 1]],
[[1, 2], [3, 6]],
[[0, 0, 0], [0, 1, 1], [0, -1, 1]]):
# Check fractional powers both for float and for complex types.
for newtype in (float, complex):
A = np.array(matrix_as_list, dtype=newtype)
for p in (-0.7, -0.9, -2.4, -1.3):
A_power = fractional_matrix_power(A, p)
assert_(np.isnan(A_power).all())
for p in (0.2, 1.43):
A_power = fractional_matrix_power(A, p)
A_round_trip = fractional_matrix_power(A_power, 1/p)
assert_allclose(A_round_trip, A)
def test_opposite_sign_complex_eigenvalues(self):
M = [[2j, 4], [0, -2j]]
R = [[1+1j, 2], [0, 1-1j]]
assert_allclose(np.dot(R, R), M, atol=1e-14)
assert_allclose(fractional_matrix_power(M, 0.5), R, atol=1e-14)
class TestExpM(object):
def test_zero(self):
a = array([[0.,0],[0,0]])
assert_array_almost_equal(expm(a),[[1,0],[0,1]])
def test_single_elt(self):
# See gh-5853
from scipy.sparse import csc_matrix
vOne = -2.02683397006j
vTwo = -2.12817566856j
mOne = csc_matrix([[vOne]], dtype='complex')
mTwo = csc_matrix([[vTwo]], dtype='complex')
outOne = expm(mOne)
outTwo = expm(mTwo)
assert_equal(type(outOne), type(mOne))
assert_equal(type(outTwo), type(mTwo))
assert_allclose(outOne[0, 0], complex(-0.44039415155949196,
-0.8978045395698304))
assert_allclose(outTwo[0, 0], complex(-0.52896401032626006,
-0.84864425749518878))
class TestExpmFrechet(object):
def test_expm_frechet(self):
# a test of the basic functionality
M = np.array([
[1, 2, 3, 4],
[5, 6, 7, 8],
[0, 0, 1, 2],
[0, 0, 5, 6],
], dtype=float)
A = np.array([
[1, 2],
[5, 6],
], dtype=float)
E = np.array([
[3, 4],
[7, 8],
], dtype=float)
expected_expm = scipy.linalg.expm(A)
expected_frechet = scipy.linalg.expm(M)[:2, 2:]
for kwargs in ({}, {'method':'SPS'}, {'method':'blockEnlarge'}):
observed_expm, observed_frechet = expm_frechet(A, E, **kwargs)
assert_allclose(expected_expm, observed_expm)
assert_allclose(expected_frechet, observed_frechet)
def test_small_norm_expm_frechet(self):
# methodically test matrices with a range of norms, for better coverage
M_original = np.array([
[1, 2, 3, 4],
[5, 6, 7, 8],
[0, 0, 1, 2],
[0, 0, 5, 6],
], dtype=float)
A_original = np.array([
[1, 2],
[5, 6],
], dtype=float)
E_original = np.array([
[3, 4],
[7, 8],
], dtype=float)
A_original_norm_1 = scipy.linalg.norm(A_original, 1)
selected_m_list = [1, 3, 5, 7, 9, 11, 13, 15]
m_neighbor_pairs = zip(selected_m_list[:-1], selected_m_list[1:])
for ma, mb in m_neighbor_pairs:
ell_a = scipy.linalg._expm_frechet.ell_table_61[ma]
ell_b = scipy.linalg._expm_frechet.ell_table_61[mb]
target_norm_1 = 0.5 * (ell_a + ell_b)
scale = target_norm_1 / A_original_norm_1
M = scale * M_original
A = scale * A_original
E = scale * E_original
expected_expm = scipy.linalg.expm(A)
expected_frechet = scipy.linalg.expm(M)[:2, 2:]
observed_expm, observed_frechet = expm_frechet(A, E)
assert_allclose(expected_expm, observed_expm)
assert_allclose(expected_frechet, observed_frechet)
def test_fuzz(self):
# try a bunch of crazy inputs
rfuncs = (
np.random.uniform,
np.random.normal,
np.random.standard_cauchy,
np.random.exponential)
ntests = 100
for i in range(ntests):
rfunc = random.choice(rfuncs)
target_norm_1 = random.expovariate(1.0)
n = random.randrange(2, 16)
A_original = rfunc(size=(n,n))
E_original = rfunc(size=(n,n))
A_original_norm_1 = scipy.linalg.norm(A_original, 1)
scale = target_norm_1 / A_original_norm_1
A = scale * A_original
E = scale * E_original
M = np.vstack([
np.hstack([A, E]),
np.hstack([np.zeros_like(A), A])])
expected_expm = scipy.linalg.expm(A)
expected_frechet = scipy.linalg.expm(M)[:n, n:]
observed_expm, observed_frechet = expm_frechet(A, E)
assert_allclose(expected_expm, observed_expm)
assert_allclose(expected_frechet, observed_frechet)
def test_problematic_matrix(self):
# this test case uncovered a bug which has since been fixed
A = np.array([
[1.50591997, 1.93537998],
[0.41203263, 0.23443516],
], dtype=float)
E = np.array([
[1.87864034, 2.07055038],
[1.34102727, 0.67341123],
], dtype=float)
A_norm_1 = scipy.linalg.norm(A, 1)
sps_expm, sps_frechet = expm_frechet(
A, E, method='SPS')
blockEnlarge_expm, blockEnlarge_frechet = expm_frechet(
A, E, method='blockEnlarge')
assert_allclose(sps_expm, blockEnlarge_expm)
assert_allclose(sps_frechet, blockEnlarge_frechet)
@pytest.mark.slow
@pytest.mark.skip(reason='this test is deliberately slow')
def test_medium_matrix(self):
# profile this to see the speed difference
n = 1000
A = np.random.exponential(size=(n, n))
E = np.random.exponential(size=(n, n))
sps_expm, sps_frechet = expm_frechet(
A, E, method='SPS')
blockEnlarge_expm, blockEnlarge_frechet = expm_frechet(
A, E, method='blockEnlarge')
assert_allclose(sps_expm, blockEnlarge_expm)
assert_allclose(sps_frechet, blockEnlarge_frechet)
def _help_expm_cond_search(A, A_norm, X, X_norm, eps, p):
p = np.reshape(p, A.shape)
p_norm = norm(p)
perturbation = eps * p * (A_norm / p_norm)
X_prime = expm(A + perturbation)
scaled_relative_error = norm(X_prime - X) / (X_norm * eps)
return -scaled_relative_error
def _normalized_like(A, B):
return A * (scipy.linalg.norm(B) / scipy.linalg.norm(A))
def _relative_error(f, A, perturbation):
X = f(A)
X_prime = f(A + perturbation)
return norm(X_prime - X) / norm(X)
class TestExpmConditionNumber(object):
def test_expm_cond_smoke(self):
np.random.seed(1234)
for n in range(1, 4):
A = np.random.randn(n, n)
kappa = expm_cond(A)
assert_array_less(0, kappa)
def test_expm_bad_condition_number(self):
A = np.array([
[-1.128679820, 9.614183771e4, -4.524855739e9, 2.924969411e14],
[0, -1.201010529, 9.634696872e4, -4.681048289e9],
[0, 0, -1.132893222, 9.532491830e4],
[0, 0, 0, -1.179475332],
])
kappa = expm_cond(A)
assert_array_less(1e36, kappa)
def test_univariate(self):
np.random.seed(12345)
for x in np.linspace(-5, 5, num=11):
A = np.array([[x]])
assert_allclose(expm_cond(A), abs(x))
for x in np.logspace(-2, 2, num=11):
A = np.array([[x]])
assert_allclose(expm_cond(A), abs(x))
for i in range(10):
A = np.random.randn(1, 1)
assert_allclose(expm_cond(A), np.absolute(A)[0, 0])
@pytest.mark.slow
def test_expm_cond_fuzz(self):
np.random.seed(12345)
eps = 1e-5
nsamples = 10
for i in range(nsamples):
n = np.random.randint(2, 5)
A = np.random.randn(n, n)
A_norm = scipy.linalg.norm(A)
X = expm(A)
X_norm = scipy.linalg.norm(X)
kappa = expm_cond(A)
# Look for the small perturbation that gives the greatest
# relative error.
f = functools.partial(_help_expm_cond_search,
A, A_norm, X, X_norm, eps)
guess = np.ones(n*n)
out = minimize(f, guess, method='L-BFGS-B')
xopt = out.x
yopt = f(xopt)
p_best = eps * _normalized_like(np.reshape(xopt, A.shape), A)
p_best_relerr = _relative_error(expm, A, p_best)
assert_allclose(p_best_relerr, -yopt * eps)
# Check that the identified perturbation indeed gives greater
# relative error than random perturbations with similar norms.
for j in range(5):
p_rand = eps * _normalized_like(np.random.randn(*A.shape), A)
assert_allclose(norm(p_best), norm(p_rand))
p_rand_relerr = _relative_error(expm, A, p_rand)
assert_array_less(p_rand_relerr, p_best_relerr)
# The greatest relative error should not be much greater than
# eps times the condition number kappa.
# In the limit as eps approaches zero it should never be greater.
assert_array_less(p_best_relerr, (1 + 2*eps) * eps * kappa)