def test_jdqz_smallest_magnitude_eigenvectors(dtype):
numpy.random.seed(1234)
tol = numpy.finfo(dtype).eps * 1e3
atol = tol * 10
n = 20
k = 5
a = generate_test_matrix([n, n], dtype)
b = generate_test_matrix([n, n], dtype)
alpha, beta, v = jdqz.jdqz(a, b, num=k, tol=tol, return_eigenvectors=True)
jdqz_eigs = numpy.array(sorted(alpha / beta, key=lambda x: abs(x)))
jdqz_eigs = jdqz_eigs[:k]
eigs = scipy.linalg.eigvals(a, b)
eigs = numpy.array(sorted(eigs, key=lambda x: abs(x)))
eigs = eigs[:k]
assert_allclose(jdqz_eigs.real, eigs.real, rtol=0, atol=atol)
assert_allclose(abs(jdqz_eigs.imag), abs(eigs.imag), rtol=0, atol=atol)
i = 0
while i < k:
ctype = numpy.dtype(numpy.dtype(dtype).char.upper())
if dtype != ctype and alpha[i].imag:
assert norm(beta[i].real * a @ v[:, i] -
alpha[i].real * b @ v[:, i] +
alpha[i].imag * b @ v[:, i + 1]) < atol
assert norm(beta[i].real * a @ v[:, i + 1] -
alpha[i].imag * b @ v[:, i] -
alpha[i].real * b @ v[:, i + 1]) < atol
i += 2
else:
assert norm(beta[i] * a @ v[:, i] - alpha[i] * b @ v[:, i]) < atol
i += 1
def test_normalization(dtype):
atol = numpy.finfo(dtype).eps * 100
n = 20
x = generate_random_dtype_array([n], dtype)
assert norm(x) > 1
orthogonalization.normalize(x)
assert_allclose(norm(x), 1, rtol=0, atol=atol)
def repeated_mgs(V, w, W=None, M=None, MV=None, MW=None):
prev_nrm = norm(w, M)
modified_gs(V, w, W, M, MV, MW)
nrm = norm(w, M)
eta = 1 / sqrt(2)
while nrm < eta * prev_nrm:
modified_gs(V, w, W, M, MV, MW)
prev_nrm = nrm
nrm = norm(w, M)
return nrm
def DGKS(V, w, W=None, M=None, MV=None, MW=None):
prev_nrm = norm(w, M)
gram_schmidt(V, w, W, M, MV, MW)
nrm = norm(w, M)
eta = 1 / sqrt(2)
while nrm < eta * prev_nrm:
gram_schmidt(V, w, W, M, MV, MW)
prev_nrm = nrm
nrm = norm(w, M)
return nrm
def check_eigenvalues(A_op, B_op, eigs, v, num_evs, tol):
from jadapy.utils import norm
idx = range(len(eigs))
idx = numpy.array(sorted(idx, key=lambda i: abs(eigs[i])))
for i in range(num_evs):
j = idx[i]
assert norm(A_op @ v[:, j]) > tol
assert_allclose(norm(A_op @ v[:, j] - B_op @ v[:, j] * eigs[j]),
0,
rtol=0,
atol=tol)
def test_Epetra_eigenvectors():
try:
from PyTrilinos import Epetra
from jadapy import EpetraInterface
except ImportError:
pytest.skip("Trilinos not found")
dtype = numpy.float64
numpy.random.seed(1234)
tol = numpy.finfo(dtype).eps * 1e3
atol = tol * 10
n = 20
k = 5
comm = Epetra.PyComm()
map = Epetra.Map(n, 0, comm)
a1, a2 = generate_Epetra_test_matrix(map, [n, n], dtype)
b1, b2 = generate_Epetra_test_matrix(map, [n, n], dtype)
interface = EpetraInterface.EpetraInterface(map)
alpha, beta, v = jdqz.jdqz(a2,
b2,
num=k,
tol=tol,
return_eigenvectors=True,
interface=interface)
jdqz_eigs = numpy.array(sorted(alpha / beta, key=lambda x: abs(x)))
jdqz_eigs = jdqz_eigs[:k]
eigs = scipy.linalg.eigvals(a1, b1)
eigs = numpy.array(sorted(eigs, key=lambda x: abs(x)))
eigs = eigs[:k]
assert_allclose(jdqz_eigs.real, eigs.real, rtol=0, atol=atol)
assert_allclose(abs(jdqz_eigs.imag), abs(eigs.imag), rtol=0, atol=atol)
i = 0
while i < k:
if alpha[i].imag:
assert norm(a2 @ v[:, i] * beta[i].real -
b2 @ v[:, i] * alpha[i].real +
b2 @ v[:, i + 1] * alpha[i].imag) < atol
assert norm(a2 @ v[:, i + 1] * beta[i].real -
b2 @ v[:, i] * alpha[i].imag -
b2 @ v[:, i + 1] * alpha[i].real) < atol
i += 2
else:
assert norm(a2 @ v[:, i] * beta[i].real -
b2 @ v[:, i] * alpha[i].real) < atol
i += 1
def DGKS(V, w, W=None, M=None, MV=None, MW=None):
prev_nrm = norm(w, M)
_proj(V, w, MV)
_proj(W, w, MW)
nrm = norm(w, M)
eta = 1 / sqrt(2)
while nrm < eta * prev_nrm:
_proj(V, w, MV)
_proj(W, w, MW)
prev_nrm = nrm
nrm = norm(w, M)
return nrm
def test_orthogonalization_no_vectors(dtype, otype):
n = 20
x = numpy.ndarray(())
y = generate_random_dtype_array([n], dtype)
orthogonalization.orthogonalize(x, y, method=otype)
assert norm(y) > 1
def test_orthonormalization(dtype, otype):
atol = numpy.finfo(dtype).eps * 100
n = 20
x = generate_random_dtype_array([n], dtype)
orthogonalization.normalize(x)
y = generate_random_dtype_array([n], dtype)
orthogonalization.orthonormalize(x, y, method=otype)
assert_allclose(dot(x, y), 0, rtol=0, atol=atol)
assert_allclose(norm(y), 1, rtol=0, atol=atol)
def normalize(w, nrm=None, M=None, verbose=True):
if nrm is None:
nrm = norm(w, M)
if verbose and nrm < eps(w):
# print('Warning: norm during normalization is nearly zero: %e' % nrm)
raise Exception('Norm during normalization is nearly zero: %e' % nrm)
w /= nrm
return nrm
def test_orthonormalization_multiple_vectors(dtype, otype):
atol = numpy.finfo(dtype).eps * 100
n = 20
k = 5
x = generate_random_dtype_array([n, k], dtype)
orthogonalization.orthonormalize(x, method=otype)
for i in range(k):
for j in range(k):
if i == j:
continue
assert_allclose(dot(x[:, i], x[:, j]), 0, rtol=0, atol=atol)
assert_allclose(norm(x[:, i]), 1, rtol=0, atol=atol)
def normalize(w, nrm=None, M=None, verbose=True, interface=None):
if nrm is None:
nrm = norm(w, M)
if verbose and nrm < eps(w):
if not interface:
raise Exception('Norm during normalization is nearly zero: %e' %
nrm)
else:
warnings.warn(
'Warning: norm during normalization is nearly zero: %e. Taking a random vector.'
% nrm)
if len(w.shape) > 1:
w[:, 0] = interface.random()
else:
w[:] = interface.random()
w /= norm(w, M)
return w
w /= nrm
return nrm
def test_orthonormalization_multiple_vectors_twice_with_mass(dtype, otype):
atol = numpy.finfo(dtype).eps * 100
n = 20
k = 5
x = generate_random_dtype_array([n, k], dtype)
M = generate_random_mass_matrix([n, n], dtype)
orthogonalization.orthonormalize(x, method=otype, M=M)
y = generate_random_dtype_array([n, k], dtype)
orthogonalization.orthonormalize(x, y, method=otype, M=M)
for i in range(k):
for j in range(k):
assert_allclose(dot(x[:, i], M @ y[:, j]), 0, rtol=0, atol=atol)
if i == j:
continue
assert_allclose(dot(y[:, i], M @ y[:, j]), 0, rtol=0, atol=atol)
assert_allclose(norm(y[:, i], M), 1, rtol=0, atol=atol)
def test_orthonormalization_multiple_vectors_epetra(otype):
try:
from PyTrilinos import Epetra
from jadapy import EpetraInterface
except ImportError:
pytest.skip("Trilinos not found")
dtype = numpy.float64
atol = numpy.finfo(dtype).eps * 100
n = 20
k = 5
comm = Epetra.PyComm()
map = Epetra.Map(n, 0, comm)
x = EpetraInterface.Vector(map, k)
x.Random()
orthogonalization.orthonormalize(x, method=otype)
for i in range(k):
for j in range(k):
if i == j:
continue
assert_allclose(x[:, i].dot(x[:, j]), 0, rtol=0, atol=atol)
assert_allclose(norm(x[:, i]), 1, rtol=0, atol=atol)
def jdqr(A,
num=5,
target=Target.SmallestMagnitude,
tol=1e-8,
M=None,
prec=None,
maxit=1000,
subspace_dimensions=(20, 40),
initial_subspace=None,
arithmetic='real',
return_eigenvectors=False,
return_subspace=False,
interface=None):
if arithmetic not in ['real', 'complex', 'r', 'c']:
raise ValueError("argument must be 'real', or 'complex'")
if not prec:
prec = _prec
solver_tolerance = 1.0
n = A.shape[0]
subspace_dimensions = (min(subspace_dimensions[0],
n // 2), min(subspace_dimensions[1], n))
it = 1
k = 0 # Number of eigenvalues found
m = 0 # Size of the search subspace
nev = 1 # Amount of eigenvalues currently converging
alpha = None
evs = None
dtype = A.dtype
if interface:
dtype = interface.dtype
ctype = numpy.dtype(dtype.char.upper())
if arithmetic in ['complex', 'c']:
dtype = ctype
if not interface:
interface = NumPyInterface(n, dtype)
extra = 0
if dtype != ctype:
# Allocate extra space in case a complex eigenpair may exist for a real matrix
extra = 1
# Eigenvalues
aconv = numpy.zeros(num + extra, ctype)
# Schur matrices
R = numpy.zeros((num + extra, num + extra), dtype)
# Schur vectors
Q = interface.vector(num + extra)
# Preconditioned Q
Y = interface.vector(num + extra)
H = numpy.zeros((num + extra, num + extra), dtype)
MQ = Q
if M is not None:
MQ = interface.vector(num + extra)
# Orthonormal search subspace
V = interface.vector(subspace_dimensions[1])
# AV = A*V without orthogonalization
AV = interface.vector(subspace_dimensions[1])
# MV = M*V without orthogonalization
MV = None
if M is not None:
MV = interface.vector(subspace_dimensions[1])
# Residual vector
r = interface.vector(1 + extra)
# Low-dimensional projection: VAV = V'*A*V
VAV = numpy.zeros((subspace_dimensions[1], subspace_dimensions[1]), dtype)
while k < num and it <= maxit:
if it == 1:
if initial_subspace is not None:
nev = min(initial_subspace.shape[1], subspace_dimensions[1])
V[:, 0:nev] = initial_subspace[:, 0:nev]
else:
V[:, 0] = interface.random()
normalize(V[:, 0], M=M)
else:
solver_maxit = 100
sigma = evs[0]
# Build an initial search subspace in an inexpensive way
# and as close to the target as possible
if m < subspace_dimensions[0]:
solver_tolerance = 0.5
solver_maxit = 1
if target != 0.0:
sigma = target
if M is not None:
V[:, m:m + nev] = solve_generalized_correction_equation(
A, M, prec, MQ[:, 0:k + nev], Q[:, 0:k + nev],
Y[:, 0:k + nev], H[0:k + nev, 0:k + nev], sigma, 1.0,
r[:, 0:nev], solver_tolerance, solver_maxit, interface)
else:
V[:, m:m + nev] = solve_correction_equation(
A, prec, Q[:, 0:k + nev], Y[:, 0:k + nev],
H[0:k + nev, 0:k + nev], sigma, r[:, 0:nev],
solver_tolerance, solver_maxit, interface)
orthonormalize(V[:, 0:m],
V[:, m:m + nev],
M=M,
MV=None if MV is None else MV[:, 0:m])
AV[:, m:m + nev] = A @ V[:, m:m + nev]
if M is not None:
MV[:, m:m + nev] = M @ V[:, m:m + nev]
# Update VAV = V' * A * V
for i in range(m):
VAV[i, m:m + nev] = dot(V[:, i], AV[:, m:m + nev])
VAV[m:m + nev, i] = dot(V[:, m:m + nev], AV[:, i])
VAV[m:m + nev, m:m + nev] = dot(V[:, m:m + nev], AV[:, m:m + nev])
m += nev
[S, U] = schur(VAV[0:m, 0:m])
found = True
while found:
[S, U] = schur_sort(S, U, target)
nev = 1
if dtype != ctype and S.shape[0] > 1 and abs(S[1, 0]) > 0.0:
# Complex eigenvalue in real arithmetic
nev = 2
alpha = S[0:nev, 0:nev]
evcond = norm(alpha)
Q[:, k:k + nev] = V[:, 0:m] @ U[:, 0:nev]
Y[:, k:k + nev] = prec(Q[:, k:k + nev], alpha)
if M is not None:
MQ[:, k:k + nev] = MV[:, 0:m] @ U[:, 0:nev]
for i in range(k):
H[i, k:k + nev] = dot(MQ[:, i], Y[:, k:k + nev])
H[k:k + nev, i] = dot(MQ[:, k:k + nev], Y[:, i])
H[k:k + nev, k:k + nev] = dot(MQ[:, k:k + nev], Y[:, k:k + nev])
r[:, 0:nev] = A @ Q[:, k:k + nev] - MQ[:, k:k + nev] @ alpha
orthogonalize(MQ[:, 0:k + nev],
r[:, 0:nev],
M=None,
MV=Q[:, 0:k + nev])
rnorm = norm(r[:, 0:nev]) / evcond
evs = scipy.linalg.eigvals(alpha)
ev_est = evs[0]
print(
"Step: %4d, subspace dimension: %3d, eigenvalue estimate: %13.6e + %13.6ei, residual norm: %e"
% (it, m, ev_est.real, ev_est.imag, rnorm))
sys.stdout.flush()
# Store converged Ritz pairs
if rnorm <= tol:
# Compute R so we can compute the eigenvectors
if return_eigenvectors:
if k > 0:
AQ = AV[:, 0:m] @ U[:, 0:nev]
for i in range(k):
R[i, k:k + nev] = dot(Q[:, i], AQ)
R[k:k + nev, k:k + nev] = alpha
# Store the converged eigenvalues
for i in range(nev):
print("Found an eigenvalue:", evs[i])
sys.stdout.flush()
aconv[k] = evs[i]
k += 1
if k >= num:
break
# Reset the iterative solver tolerance
solver_tolerance = 1.0
# Remove the eigenvalue from the search space
V[:, 0:m - nev] = V[:, 0:m] @ U[:, nev:m]
AV[:, 0:m - nev] = AV[:, 0:m] @ U[:, nev:m]
if M is not None:
MV[:, 0:m - nev] = MV[:, 0:m] @ U[:, nev:m]
VAV[0:m - nev, 0:m - nev] = S[nev:m, nev:m]
S = VAV[0:m - nev, 0:m - nev]
U = numpy.identity(m - nev, dtype)
m -= nev
else:
found = False
solver_tolerance = max(solver_tolerance / 2, tol)
if m >= min(subspace_dimensions[1], n - k) and k < num:
# Maximum search space dimension has been reached.
new_m = min(subspace_dimensions[0], n - k)
print("Shrinking the search space from %d to %d" % (m, new_m))
sys.stdout.flush()
V[:, 0:new_m] = V[:, 0:m] @ U[:, 0:new_m]
AV[:, 0:new_m] = AV[:, 0:m] @ U[:, 0:new_m]
if M is not None:
MV[:, 0:new_m] = MV[:, 0:m] @ U[:, 0:new_m]
VAV[0:new_m, 0:new_m] = S[0:new_m, 0:new_m]
m = new_m
elif m + nev - 1 >= min(subspace_dimensions[1], n - k):
# Only one extra vector fits in the search space.
nev = 1
it += 1
if return_eigenvectors:
evs, v = scipy.linalg.eig(R[0:k, 0:k], left=False, right=True)
if ctype == dtype:
if return_subspace:
return evs, Q[:, 0:k] @ v, Q[:, 0:k]
return evs, Q[:, 0:k] @ v
i = 0
while i < k:
Y[:, i] = Q[:, 0:k] @ v[:, i].real
if evs[i].imag:
Y[:, i + 1] = Q[:, 0:k] @ v[:, i].imag
i += 1
i += 1
if return_subspace:
return evs, Y[:, 0:k], Q[:, 0:k]
return evs, Y[:, 0:k]
if return_subspace:
return aconv[0:num], Q[:, 0:k]
return aconv[0:num]
def jdqz(A,
B,
num=5,
target=Target.SmallestMagnitude,
tol=1e-8,
lock_tol=None,
prec=None,
maxit=1000,
subspace_dimensions=(20, 40),
initial_subspaces=None,
arithmetic='real',
testspace='Harmonic Petrov',
return_eigenvectors=False,
return_subspaces=False,
interface=None):
if arithmetic not in ['real', 'complex', 'r', 'c']:
raise ValueError("argument must be 'real', or 'complex'")
if not prec:
prec = _prec
if not lock_tol:
lock_tol = tol * 1e2
solver_tolerance = 1.0
n = A.shape[0]
subspace_dimensions = (min(subspace_dimensions[0],
n // 2), min(subspace_dimensions[1], n))
it = 1
k = 0 # Number of eigenvalues found
m = 0 # Size of the search subspace
nev = 1 # Amount of eigenvalues currently converging
alpha = None
beta = None
evs = None
sort_target = target
dtype = A.dtype
if interface:
dtype = interface.dtype
ctype = numpy.dtype(dtype.char.upper())
if arithmetic in ['complex', 'c']:
dtype = ctype
if not interface:
interface = NumPyInterface(n, dtype)
extra = 0
if dtype != ctype:
# Allocate extra space in case a complex eigenpair may exist for a real matrix
extra = 1
# Generalized eigenvalues
aconv = numpy.zeros(num + extra, ctype)
bconv = numpy.zeros(num + extra, dtype)
# Generalized Schur matrices
RA = numpy.zeros((num + extra, num + extra), dtype)
RB = numpy.zeros((num + extra, num + extra), dtype)
# Generalized Schur vectors
Q = interface.vector(num + extra)
Z = interface.vector(num + extra)
# Preconditioned Z
Y = interface.vector(num + extra)
QZ = numpy.zeros((num + extra, num + extra), dtype)
# Orthonormal search subspace
V = interface.vector(subspace_dimensions[1])
# Orthonormal test subspace
W = interface.vector(subspace_dimensions[1])
# AV = A*V without orthogonalization
AV = interface.vector(subspace_dimensions[1])
# BV = B*V without orthogonalization
BV = interface.vector(subspace_dimensions[1])
# Residual vector
r = interface.vector(1 + extra)
# Low-dimensional projections: WAV = W'*A*V, WBV = W'*B*V
WAV = numpy.zeros((subspace_dimensions[1], subspace_dimensions[1]), dtype)
WBV = numpy.zeros((subspace_dimensions[1], subspace_dimensions[1]), dtype)
while k < num and it <= maxit:
if it == 1:
if initial_subspaces is not None:
nev = min(initial_subspaces[0].shape[1],
subspace_dimensions[1])
V[:, 0:nev] = initial_subspaces[0][:, 0:nev]
if len(initial_subspaces) > 1:
W[:, 0:nev] = initial_subspaces[1][:, 0:nev]
else:
V[:, 0] = interface.random()
normalize(V[:, 0])
else:
solver_maxit = 100
sigma_a = evs[0, 0]
sigma_b = evs[1, 0]
# Build an initial search subspace in an inexpensive way
# and as close to the target as possible
if m < subspace_dimensions[0]:
solver_tolerance = 0.5
solver_maxit = 1
sigma_a = target
sigma_b = 1.0
V[:, m:m + nev] = solve_generalized_correction_equation(
A, B, prec, Q[:, 0:k + nev], Z[:, 0:k + nev], Y[:, 0:k + nev],
QZ[0:k + nev, 0:k + nev], sigma_a, sigma_b, r[:, 0:nev],
solver_tolerance, solver_maxit, interface)
orthonormalize(V[:, 0:m], V[:, m:m + nev], interface=interface)
AV[:, m:m + nev] = A @ V[:, m:m + nev]
BV[:, m:m + nev] = B @ V[:, m:m + nev]
if it > 1 or initial_subspaces is None or len(initial_subspaces) < 2:
nu, mu = _set_testspace(testspace, target, alpha, beta, dtype,
ctype)
if nu.shape[0] < nev:
# Repeat nu and mu in case only an initial V was passed
nu = numpy.diag(numpy.repeat(nu[0, 0], nev))
mu = numpy.diag(numpy.repeat(mu[0, 0], nev))
W[:, m:m + nev] = AV[:, m:m + nev] @ nu[
0:nev, 0:nev] + BV[:, m:m + nev] @ mu[0:nev, 0:nev]
orthogonalize(Z[:, 0:k], W[:, m:m + nev])
orthonormalize(W[:, 0:m], W[:, m:m + nev], interface=interface)
# Update WAV = W' * A * V
for i in range(m):
WAV[i, m:m + nev] = dot(W[:, i], AV[:, m:m + nev])
WAV[m:m + nev, i] = dot(W[:, m:m + nev], AV[:, i])
WAV[m:m + nev, m:m + nev] = dot(W[:, m:m + nev], AV[:, m:m + nev])
# Update WBV = W' * B * V
for i in range(m):
WBV[i, m:m + nev] = dot(W[:, i], BV[:, m:m + nev])
WBV[m:m + nev, i] = dot(W[:, m:m + nev], BV[:, i])
WBV[m:m + nev, m:m + nev] = dot(W[:, m:m + nev], BV[:, m:m + nev])
m += nev
[S, T, UL, UR] = generalized_schur(WAV[0:m, 0:m], WBV[0:m, 0:m])
found = True
while found:
[S, T, UL, UR] = generalized_schur_sort(S, T, UL, UR, sort_target)
nev = 1
if dtype != ctype and S.shape[0] > 1 and abs(S[1, 0]) > 0.0:
# Complex eigenvalue in real arithmetic
nev = 2
alpha = S[0:nev, 0:nev]
beta = T[0:nev, 0:nev]
evcond = sqrt(norm(alpha)**2 + norm(beta)**2)
Q[:, k:k + nev] = V[:, 0:m] @ UR[:, 0:nev]
Z[:, k:k + nev] = W[:, 0:m] @ UL[:, 0:nev]
Y[:, k:k + nev] = prec(Z[:, k:k + nev], alpha, beta)
for i in range(k):
QZ[i, k:k + nev] = dot(Q[:, i], Y[:, k:k + nev])
QZ[k:k + nev, i] = dot(Q[:, k:k + nev], Y[:, i])
QZ[k:k + nev, k:k + nev] = dot(Q[:, k:k + nev], Y[:, k:k + nev])
r[:,
0:nev] = A @ Q[:, k:k + nev] @ beta - B @ Q[:, k:k + nev] @ alpha
orthogonalize(Z[:, 0:k + nev], r[:, 0:nev])
rnorm = norm(r[:, 0:nev]) / evcond
evs = scipy.linalg.eigvals(alpha, beta, homogeneous_eigvals=True)
ev_est = evs[0, 0] / evs[1, 0]
print(
"Step: %4d, subspace dimension: %3d, eigenvalue estimate: %13.6e + %13.6ei, residual norm: %e"
% (it, m, ev_est.real, ev_est.imag, rnorm))
sys.stdout.flush()
if rnorm <= lock_tol:
sort_target = ev_est
# Store converged Petrov pairs
if rnorm <= tol and m > nev:
# Compute RA and RB so we can compute the eigenvectors
if return_eigenvectors:
if k > 0:
AQ = AV[:, 0:m] @ UR[:, 0:nev]
BQ = BV[:, 0:m] @ UR[:, 0:nev]
for i in range(k):
RA[i, k:k + nev] = dot(Z[:, i], AQ)
RB[i, k:k + nev] = dot(Z[:, i], BQ)
RA[k:k + nev, k:k + nev] = alpha
RB[k:k + nev, k:k + nev] = beta
# Store the converged eigenvalues
for i in range(nev):
print("Found an eigenvalue:", evs[0, i] / evs[1, i])
sys.stdout.flush()
aconv[k] = evs[0, i]
bconv[k] = evs[1, i].real
k += 1
if k >= num:
break
# Reset the iterative solver tolerance
solver_tolerance = 1.0
# Unlock the target
sort_target = target
# Remove the eigenvalue from the search space
V[:, 0:m - nev] = V[:, 0:m] @ UR[:, nev:m]
AV[:, 0:m - nev] = AV[:, 0:m] @ UR[:, nev:m]
BV[:, 0:m - nev] = BV[:, 0:m] @ UR[:, nev:m]
W[:, 0:m - nev] = W[:, 0:m] @ UL[:, nev:m]
WAV[0:m - nev, 0:m - nev] = S[nev:m, nev:m]
WBV[0:m - nev, 0:m - nev] = T[nev:m, nev:m]
S = WAV[0:m - nev, 0:m - nev]
T = WBV[0:m - nev, 0:m - nev]
UL = numpy.identity(m - nev, dtype)
UR = numpy.identity(m - nev, dtype)
m -= nev
else:
found = False
solver_tolerance = max(solver_tolerance / 2, tol / 100)
if m >= min(subspace_dimensions[1], n - k) and k < num:
# Maximum search space dimension has been reached.
new_m = min(subspace_dimensions[0], n - k)
print("Shrinking the search space from %d to %d" % (m, new_m))
sys.stdout.flush()
V[:, 0:new_m] = V[:, 0:m] @ UR[:, 0:new_m]
AV[:, 0:new_m] = AV[:, 0:m] @ UR[:, 0:new_m]
BV[:, 0:new_m] = BV[:, 0:m] @ UR[:, 0:new_m]
W[:, 0:new_m] = W[:, 0:m] @ UL[:, 0:new_m]
WAV[0:new_m, 0:new_m] = S[0:new_m, 0:new_m]
WBV[0:new_m, 0:new_m] = T[0:new_m, 0:new_m]
m = new_m
elif m + nev - 1 >= min(subspace_dimensions[1], n - k):
# Only one extra vector fits in the search space.
nev = 1
it += 1
if return_eigenvectors:
evs, v = scipy.linalg.eig(RA[0:k, 0:k],
RB[0:k, 0:k],
left=False,
right=True,
homogeneous_eigvals=True)
if ctype == dtype:
if return_subspaces:
return evs[0], evs[1], Q[:, 0:k] @ v, Q[:, 0:k], Z[:, 0:k]
return evs[0], evs[1], Q[:, 0:k] @ v
i = 0
while i < k:
Y[:, i] = Q[:, 0:k] @ v[:, i].real
if evs[0][i].imag:
Y[:, i + 1] = Q[:, 0:k] @ v[:, i].imag
i += 1
i += 1
if return_subspaces:
return evs[0], evs[1], Y[:, 0:k], Q[:, 0:k], Z[:, 0:k]
return evs[0], evs[1], Y[:, 0:k]
if return_subspaces:
return aconv[0:num], bconv[0:num], Q[:, 0:k], Z[:, 0:k]
return aconv[0:num], bconv[0:num]
