def pod(A, modes=None, product=None, tol=None, symmetrize=None, orthonormalize=False,
check=None, check_tol=None):
assert isinstance(A, VectorArrayInterface)
assert len(A) > 0
assert modes is None or modes <= len(A)
assert product is None or isinstance(product, OperatorInterface)
tol = defaults.pod_tol if tol is None else tol
symmetrize = defaults.pod_symmetrize if symmetrize is None else symmetrize
orthonormalize = defaults.pod_orthonormalize if orthonormalize is None else orthonormalize
check = defaults.pod_check if check is None else check
check_tol = defaults.pod_check_tol if check_tol is None else check_tol
B = A.gramian() if product is None else product.apply2(A, A, pairwise=False)
if symmetrize: # according to rbmatlab this is necessary due to rounding
B = B + B.T
B *= 0.5
eigvals = None if modes is None else (len(B) - modes, len(B) - 1)
EVALS, EVECS = eigh(B, overwrite_a=True, turbo=True, eigvals=eigvals)
EVALS = EVALS[::-1]
EVECS = EVECS.T[::-1, :] # is this a view?
last_above_tol = np.where(EVALS >= tol)[0][-1]
EVALS = EVALS[:last_above_tol + 1]
EVECS = EVECS[:last_above_tol + 1]
if len(EVALS) == 0:
return type(A).empty(A.dim)
POD = A.lincomb(EVECS / np.sqrt(EVALS[:, np.newaxis]))
if orthonormalize:
POD = gram_schmidt(POD, product=product)
if check:
if not product and not float_cmp_all(POD.dot(POD, pairwise=False), np.eye(len(POD)), check_tol):
err = np.max(np.abs(POD.dot(POD, pairwise=False) - np.eye(len(POD))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
elif product and not float_cmp_all(product.apply2(POD, POD, pairwise=False), np.eye(len(POD)), check_tol):
err = np.max(np.abs(product.apply2(POD, POD, pairwise=False) - np.eye(len(POD))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
return POD
def allclose(self, mu):
assert isinstance(mu, Parameter)
if set(self.keys()) != set(mu.keys()):
return False
if not all(float_cmp_all(v, mu[k]) for k, v in self.iteritems()):
return False
else:
return True
def allclose(self, mu):
'''Compare to |Parameters| using :meth:`~pymor.tools.floatcmp.float_cmp_all`.
Parameters
----------
mu
The |Parameter| with which to compare.
Returns
-------
`True` if both |Parameters| have the same |ParameterType| and all parameter
components are almost equal, else `False`.
'''
assert isinstance(mu, Parameter)
if self.viewkeys() != mu.viewkeys():
return False
elif not all(float_cmp_all(v, mu[k]) for k, v in self.iteritems()):
return False
else:
return True
def numpy_trivial_basis_extension(basis, U):
'''Trivially extend basis by just adding the new vector.
We check that the new vector is not already contained in the basis, but we do
not check for linear independence.
Parameters
----------
basis
The basis to extend.
U
The new basis vector.
Returns
-------
The new basis.
Raises
------
ExtensionError
Is raised if U is already contained in basis.
'''
assert isinstance(U, NumpyVectorArray)
if basis is None:
return U
assert isinstance(basis, NumpyVectorArray)
basis = basis.data
U = U.data
if not all(not float_cmp_all(B, U) for B in basis):
raise ExtensionError
new_basis = np.empty((basis.shape[0] + 1, basis.shape[1]))
new_basis[:-1, :] = basis
new_basis[-1, :] = U
return NumpyVectorArray(new_basis)
def almost_equal(self, other, rtol=None, atol=None):
assert self.dim == other.dim
return float_cmp_all(self._array, other._array, rtol=rtol, atol=atol)
def gram_schmidt(A, product=None, tol=1e-14, offset=0, find_duplicates=True,
reiterate=True, reiteration_threshold=1e-1, check=True, check_tol=1e-3,
copy=False):
"""Orthonormalize a |VectorArray| using the Gram-Schmidt algorithm.
Parameters
----------
A
The |VectorArray| which is to be orthonormalized.
product
The scalar product w.r.t. which to orthonormalize, given as a linear
|Operator|. If `None` the Euclidean product is used.
tol
Tolerance to determine a linear dependent row.
offset
Assume that the first `offset` vectors are already orthogonal and start the
algorithm at the `offset + 1`-th vector.
find_duplicates
If `True`, eliminate duplicate vectors before the main loop.
reiterate
If `True`, orthonormalize again if the norm of the orthogonalized vector is
much smaller than the norm of the original vector.
reiteration_threshold
If `reiterate` is `True`, re-orthonormalize if the ratio between the norms of
the orthogonalized vector and the original vector is smaller than this value.
check
If `True`, check if the resulting VectorArray is really orthonormal.
check_tol
Tolerance for the check.
copy
If `True`, create a copy of `A` instead of working directly on `A`.
Returns
-------
The orthonormalized |VectorArray|.
"""
logger = getLogger('pymor.la.gram_schmidt.gram_schmidt')
if copy:
A = A.copy()
# find duplicate vectors since in some circumstances these cannot be detected in the main loop
# (is this really needed or is in this cases the tolerance poorly chosen anyhow)
if find_duplicates:
i = 0
while i < len(A):
duplicates = A.almost_equal(A, ind=i, o_ind=np.arange(max(offset, i + 1), len(A)))
if np.any(duplicates):
A.remove(np.where(duplicates)[0])
logger.info("Removing duplicate vectors")
i += 1
# main loop
remove = []
norm = None
for i in xrange(offset, len(A)):
# first calculate norm
if product is None:
oldnorm = A.l2_norm(ind=i)[0]
else:
oldnorm = np.sqrt(product.apply2(A, A, V_ind=i, U_ind=i, pairwise=True))[0]
if float_cmp_all(oldnorm, 0):
logger.info("Removing null vector {}".format(i))
remove.append(i)
continue
if i == 0:
A.scal(1/oldnorm, ind=0)
else:
first_iteration = True
# If reiterate is True, reiterate as long as the norm of the vector changes
# strongly during orthonormalization (due to Andreas Buhr).
while first_iteration or reiterate and norm < reiteration_threshold:
# this loop assumes that oldnorm is the norm of the ith vector when entering
if first_iteration:
first_iteration = False
else:
logger.info('Orthonormalizing vector {} again'.format(i))
# orthogonalize to all vectors left
for j in xrange(i):
if j in remove:
continue
if product is None:
p = A.dot(A, ind=i, o_ind=j, pairwise=True)[0]
else:
p = product.apply2(A, A, V_ind=i, U_ind=j, pairwise=True)[0]
A.axpy(-p, A, ind=i, x_ind=j)
# calculate new norm
if product is None:
norm = A.l2_norm(ind=i)[0]
else:
norm = np.sqrt(product.apply2(A, A, V_ind=i, U_ind=i, pairwise=True))[0]
# remove vector if it got too small:
if norm / oldnorm < tol:
logger.info("Removing linear dependent vector {}".format(i))
remove.append(i)
break
A.scal(1 / norm, ind=i)
oldnorm = 1.
if remove:
A.remove(remove)
if check:
if not product and not float_cmp_all(A.dot(A, pairwise=False), np.eye(len(A)),
atol=check_tol, rtol=0.):
err = np.max(np.abs(A.dot(A, pairwise=False) - np.eye(len(A))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
elif product and not float_cmp_all(product.apply2(A, A, pairwise=False), np.eye(len(A)),
atol=check_tol, rtol=0.):
err = np.max(np.abs(product.apply2(A, A, pairwise=False) - np.eye(len(A))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
return A
def pod(A, modes=None, product=None, tol=None, symmetrize=None, orthonormalize=None,
check=None, check_tol=None):
'''Proper orthogonal decomposition of `A`.
If the |VectorArray| `A` is viewed as a map ::
A: R^(len(A)) ---> R^(dim(A))
then the return value of this method is simply the array of left-singular
vectors of the singular value decomposition of `A`, where the scalar product
on R^(dim(A) is given by `product` and the scalar product on R^(len(A)) is
the Euclidean product.
Parameters
----------
A
The |VectorArray| for which the POD is to be computed.
modes
If not `None` only the first `modes` POD modes (singular vectors) are
returned.
products
Scalar product given as a linear |Operator| w.r.t. compute the POD.
tol
Squared singular values smaller than this value are ignored. If `None`,
the `pod_tol` |default| value is used.
symmetrize
If `True` symmetrize the gramian again before proceeding. If `None`,
the `pod_symmetrize` |default| value is chosen.
orthonormalize
If `True`, orthonormalize the computed POD modes again using
:func:`la.gram_schmidt.gram_schmidt`. If `None`, the `pod_orthonormalize`
|default| value is used.
check
If `True`, check the computed POD modes for orthonormality. If `None`,
the `pod_check` |default| value is used.
check_tol
Tolerance for the orthonormality check. If `None`, the `pod_check_tol`
|default| value is used.
Returns
-------
|VectorArray| of POD modes.
'''
assert isinstance(A, VectorArrayInterface)
assert len(A) > 0
assert modes is None or modes <= len(A)
assert product is None or isinstance(product, OperatorInterface)
tol = defaults.pod_tol if tol is None else tol
symmetrize = defaults.pod_symmetrize if symmetrize is None else symmetrize
orthonormalize = defaults.pod_orthonormalize if orthonormalize is None else orthonormalize
check = defaults.pod_check if check is None else check
check_tol = defaults.pod_check_tol if check_tol is None else check_tol
B = A.gramian() if product is None else product.apply2(A, A, pairwise=False)
if symmetrize: # according to rbmatlab this is necessary due to rounding
B = B + B.T
B *= 0.5
eigvals = None if modes is None else (len(B) - modes, len(B) - 1)
EVALS, EVECS = eigh(B, overwrite_a=True, turbo=True, eigvals=eigvals)
EVALS = EVALS[::-1]
EVECS = EVECS.T[::-1, :] # is this a view? yes it is!
above_tol = np.where(EVALS >= tol)[0]
if len(above_tol) == 0:
return type(A).empty(A.dim)
last_above_tol = above_tol[-1]
EVALS = EVALS[:last_above_tol + 1]
EVECS = EVECS[:last_above_tol + 1]
POD = A.lincomb(EVECS / np.sqrt(EVALS[:, np.newaxis]))
if orthonormalize:
POD = gram_schmidt(POD, product=product, copy=False)
if check:
if not product and not float_cmp_all(POD.dot(POD, pairwise=False), np.eye(len(POD)), check_tol):
err = np.max(np.abs(POD.dot(POD, pairwise=False) - np.eye(len(POD))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
elif product and not float_cmp_all(product.apply2(POD, POD, pairwise=False), np.eye(len(POD)), check_tol):
err = np.max(np.abs(product.apply2(POD, POD, pairwise=False) - np.eye(len(POD))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
return POD
def numpy_gram_schmidt(A, product=None, tol=None, row_offset=0, find_row_duplicates=True, find_col_duplicates=False,
check=None, check_tol=None):
'''Orthonormnalize a matrix using the Gram-Schmidt algorithm.
Parameters
----------
A
The matrix which is to be orthonormalized.
product
The scalar product w.r.t. which to orthonormalize. Either a `DiscreteLinearOperator`
or a square matrix.
tol
Tolerance to determine a linear dependent row.
row_offset
Assume that the first `row_offset` rows are already orthogonal and start the
algorithm at the `row_offset + 1`-th row.
find_row_duplicates
If `True`, eliminate duplicate rows before the main loop.
find_col_duplicates
If `True`, eliminate duplicate columns before the main loop.
check
If `True`, check if the resulting matrix is really orthonormal. If `None`, use
`defaults.gram_schmidt_check`.
check_tol
Tolerance for the check. If `None`, `defaults.gram_schmidt_check_tol` is used.
Returns
-------
The orthonormalized matrix.
'''
A = A.copy()
tol = defaults.gram_schmidt_tol if tol is None else tol
check = defaults.gram_schmidt_tol if check is None else check
check_tol = check_tol or defaults.gram_schmidt_check_tol
# find duplicate rows since in some circumstances these cannot be detected in the main loop
# (is this really needed or is in this cases the tolerance poorly chosen anyhow)
if find_row_duplicates:
for i in xrange(A.shape[0]):
for j in xrange(max(row_offset, i + 1), A.shape[0]):
if float_cmp_all(A[i], A[j]):
A[j] = 0
# find duplicate columns
if find_col_duplicates:
for i in xrange(A.shape[1]):
for j in xrange(i, A.shape[1]):
if float_cmp_all(A[:, i], A[:, j]):
A[:, j] = 0
# main loop
for i in xrange(A.shape[0]):
if i >= row_offset:
if product is None:
norm = np.sqrt(np.sum(A[i] ** 2))
else:
norm = np.sqrt(product.apply2(A[i], A[i]))
if norm < tol:
A[i] = 0
else:
A[i] = A[i] / norm
j = max(row_offset, i + 1)
if product is None:
p = np.sum(A[j:] * A[i], axis=-1)
else:
p = product.apply2(A[j:], A[i], pairwise=False)
A[j:] -= p[..., np.newaxis] * A[i]
rows = np.logical_not(np.all(A == 0, axis=-1))
A = A[rows]
if check:
if not float_cmp_all(A.dot(A.T), np.eye(A.shape[0]), check_tol):
err = np.max(np.abs(A.dot(A.T) - np.eye(A.shape[0])))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
return A
def gram_schmidt(A, product=None, tol=None, offset=0, find_duplicates=True,
check=None, check_tol=None):
'''Orthonormnalize a matrix using the Gram-Schmidt algorithm.
Parameters
----------
A
The VectorArray which is to be orthonormalized.
product
The scalar product w.r.t. which to orthonormalize.
tol
Tolerance to determine a linear dependent row.
offset
Assume that the first `offset` vectors are already orthogonal and start the
algorithm at the `offset + 1`-th vector.
find_duplicates
If `True`, eliminate duplicate vectors before the main loop.
check
If `True`, check if the resulting VectorArray is really orthonormal. If `None`, use
`defaults.gram_schmidt_check`.
check_tol
Tolerance for the check. If `None`, `defaults.gram_schmidt_check_tol` is used.
Returns
-------
The orthonormalized matrix.
'''
tol = defaults.gram_schmidt_tol if tol is None else tol
check = defaults.gram_schmidt_tol if check is None else check
check_tol = check_tol or defaults.gram_schmidt_check_tol
# find duplicate vectors since in some circumstances these cannot be detected in the main loop
# (is this really needed or is in this cases the tolerance poorly chosen anyhow)
if find_duplicates:
for i in xrange(len(A)):
duplicates = A.almost_equal(A, ind=[i], o_ind=xrange(max(offset, i + 1), len(A)))
if np.any(duplicates):
A.remove(np.where(duplicates))
# main loop
i = 0
remove = []
while i < len(A):
if i >= offset:
if product is None:
norm = A.l2_norm(ind=[i])[0]
else:
norm = np.sqrt(product.apply2(A, A, ind=[i], o_ind=[i], pairwise=True))[0]
if norm < tol:
remove.append(i)
continue
else:
A.iadd_mult(None, factor=1/norm, o_factor=0, ind=[i])
for j in xrange(max(offset, i + 1), len(A)):
if product is None:
p = A.prod(A, ind=[j], o_ind=[i], pairwise=True)[0]
else:
p = product.apply2(A, A, V_ind=[j], U_ind=[i], pairwise=True)[0]
A.iadd_mult(A, o_factor=-p, ind=[j], o_ind=[i])
i += 1
if remove:
A.remove(remove)
if check:
if not float_cmp_all(A.prod(A, pairwise=False), np.eye(len(A)), check_tol):
err = np.max(np.abs(A.prod(A, pairwise=False) - np.eye(len(A))))
raise AccuracyError('result not orthogonal (max err={})'.format(err))
return A
