"""Module exposing the `Matrices` and `MatricesMetric` class."""
import logging
import math
from functools import reduce
import geomstats.backend as gs
import geomstats.errors
from geomstats.algebra_utils import flip_determinant, from_vector_to_diagonal_matrix
from geomstats.geometry.base import MatrixVectorSpace
from geomstats.geometry.diffeo import VectorSpaceDiffeo
from geomstats.geometry.euclidean import EuclideanMetric
from geomstats.vectorization import repeat_out
[docs]
def matrix_matrix_transpose(point):
r"""Matrix multiplication with transpose.
.. math::
f(A) = A A^{\top}
Parameters
----------
point : array-like, shape=[n, n]
Matrix.
Returns
-------
mat : array-like, shape=[n, n]
Matrix resulting from operation.
"""
return Matrices.mul(point, Matrices.transpose(point))
[docs]
def tangent_matrix_matrix_transpose(tangent_vec, base_point):
r"""Tangent matrix multiplication with transpose.
.. math::
d_A f (X) = X A^\top + A X^\top
Parameters
----------
tangent_vec : array-like, shape=[..., n, n]
Tangent vector.
base_point : array-like, shape=[..., n, n]
Base point.
Returns
-------
image_tangent_vec : array-like, shape=[n, n]
Matrix resulting from operation.
"""
return Matrices.mul(tangent_vec, Matrices.transpose(base_point)) + Matrices.mul(
base_point, Matrices.transpose(tangent_vec)
)
[docs]
class FlattenDiffeo(VectorSpaceDiffeo):
"""A diffeo from matrices to Euclidean by flattening.
Parameters
----------
m, n : int
Integers representing the shapes of the matrices: m x n.
"""
def __init__(self, m, n=None):
super().__init__(space_ndim=2, image_space_ndim=1)
self.m, self.n = m, n or m
def __call__(self, base_point=None):
"""Diffeomorphism at base point.
Parameters
----------
base_point : array-like, shape=[..., m, n]
Base point.
Returns
-------
image_point : array-like, shape=[..., m*n]
Image point.
"""
if base_point is None:
return None
return Matrices.flatten(base_point)
[docs]
def inverse(self, image_point=None):
r"""Inverse diffeomorphism at image point.
:math:`f^{-1}: N \rightarrow M`
Parameters
----------
image_point : array-like, shape=[..., m*n]
Image point.
Returns
-------
base_point : array-like, shape=[..., m, n]
Base point.
"""
if image_point is None:
return None
return gs.reshape(image_point, image_point.shape[:-1] + (self.m, self.n))
[docs]
class BasisRepresentationDiffeo(VectorSpaceDiffeo):
"""A diffeo from matrices to Euclidean through basis representation.
Parameters
----------
space : MatrixVectorSpace
"""
def __init__(self, space):
super().__init__(space_ndim=2, image_space_ndim=1)
self._space = space
def __call__(self, base_point=None):
"""Diffeomorphism at base point.
Parameters
----------
base_point : array-like, shape=[..., m, n]
Base point.
Returns
-------
image_point : array-like, shape=[..., dim]
Image point.
"""
if base_point is None:
return None
return self._space.basis_representation(base_point)
[docs]
def inverse(self, image_point=None):
r"""Inverse diffeomorphism at image point.
:math:`f^{-1}: N \rightarrow M`
Parameters
----------
image_point : array-like, shape=[..., dim]
Image point.
Returns
-------
base_point : array-like, shape=[..., m, n]
Base point.
"""
if image_point is None:
return None
return self._space.matrix_representation(image_point)
[docs]
class Matrices(MatrixVectorSpace):
"""Class for the space of matrices (m, n).
Parameters
----------
m, n : int
Integers representing the shapes of the matrices: m x n.
"""
def __init__(self, m, n, equip=True):
geomstats.errors.check_integer(n, "n")
geomstats.errors.check_integer(m, "m")
super().__init__(shape=(m, n), equip=equip)
self.m = m
self.n = n
def _create_basis(self):
"""Create the canonical basis."""
m, n = self.m, self.n
return gs.reshape(gs.eye(n * m), (n * m, m, n))
[docs]
@staticmethod
def default_metric():
"""Metric to equip the space with if equip is True."""
return MatricesMetric
[docs]
def basis_representation(self, matrix_representation):
"""Compute the coefficient in the usual matrix basis.
This simply flattens the input.
Parameters
----------
matrix_representation : array-like, shape=[..., n, n]
Matrix.
Returns
-------
basis_representation : array-like, shape=[..., dim]
Representation in the basis.
"""
return self.flatten(matrix_representation)
[docs]
@staticmethod
def equal(mat_a, mat_b, atol=gs.atol):
"""Test if matrices a and b are close.
Parameters
----------
mat_a : array-like, shape=[..., dim1, dim2]
Matrix.
mat_b : array-like, shape=[..., dim2, dim3]
Matrix.
atol : float
Tolerance.
Optional, default: backend atol.
Returns
-------
eq : array-like, shape=[...,]
Boolean evaluating if the matrices are close.
"""
return gs.all(gs.isclose(mat_a, mat_b, atol=atol), (-2, -1))
[docs]
@staticmethod
def mul(*args):
"""Compute the product of matrices a1, ..., an.
Parameters
----------
a1 : array-like, shape=[..., dim_1, dim_2]
Matrix.
a2 : array-like, shape=[..., dim_2, dim_3]
Matrix.
...
an : array-like, shape=[..., dim_n-1, dim_n]
Matrix.
Returns
-------
mul : array-like, shape=[..., dim_1, dim_n]
Result of the product of matrices.
"""
return reduce(gs.matmul, args)
[docs]
@classmethod
def bracket(cls, mat_a, mat_b):
"""Compute the commutator of a and b, i.e. `[a, b] = ab - ba`.
Parameters
----------
mat_a : array-like, shape=[..., n, n]
Matrix.
mat_b : array-like, shape=[..., n, n]
Matrix.
Returns
-------
mat_c : array-like, shape=[..., n, n]
Commutator.
"""
return cls.mul(mat_a, mat_b) - cls.mul(mat_b, mat_a)
[docs]
@staticmethod
def transpose(mat):
"""Return the transpose of matrices.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
transpose : array-like, shape=[..., n, n]
Transposed matrix.
"""
ndim = gs.ndim(mat)
axes = list(range(0, ndim))
axes[-1] = ndim - 2
axes[-2] = ndim - 1
return gs.transpose(mat, axes)
[docs]
@staticmethod
def diagonal(mat):
"""Return the diagonal of a matrix as a vector.
Parameters
----------
mat : array-like, shape=[..., m, n]
Matrix.
Returns
-------
diagonal : array-like, shape=[..., min(m, n)]
Vector of diagonal coefficients.
"""
return gs.diagonal(mat, axis1=-2, axis2=-1)
[docs]
@staticmethod
def is_square(mat):
"""Check if a matrix is square.
Parameters
----------
mat : array-like, shape=[..., m, n]
Matrix.
Returns
-------
is_square : array-like, shape=[...,]
Boolean evaluating if the matrix is square.
"""
shape = mat.shape[:-2]
if mat.shape[-1] == mat.shape[-2]:
return gs.ones(shape, dtype=bool)
return gs.zeros(shape, dtype=bool)
[docs]
@classmethod
def is_diagonal(cls, mat, atol=gs.atol):
"""Check if a square matrix is diagonal.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default: backend atol.
Returns
-------
is_diagonal : array-like, shape=[...,]
Boolean evaluating if the matrix is square and diagonal.
"""
diagonal_mat = from_vector_to_diagonal_matrix(cls.diagonal(mat))
is_diagonal = gs.all(gs.isclose(mat, diagonal_mat, atol=atol), axis=(-2, -1))
return is_diagonal
[docs]
@classmethod
def is_lower_triangular(cls, mat, atol=gs.atol):
"""Check if a square matrix is lower triangular.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default : backend atol.
Returns
-------
is_tril : array-like, shape=[...,]
Boolean evaluating if the matrix is lower triangular
"""
return cls.equal(mat, gs.tril(mat), atol)
[docs]
@classmethod
def is_upper_triangular(cls, mat, atol=gs.atol):
"""Check if a square matrix is upper triangular.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default : backend atol.
Returns
-------
is_triu : array-like, shape=[...,]
Boolean evaluating if the matrix is upper triangular.
"""
return cls.equal(mat, gs.triu(mat), atol)
[docs]
@classmethod
def is_strictly_lower_triangular(cls, mat, atol=gs.atol):
"""Check if a square matrix is strictly lower triangular.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default : backend atol.
Returns
-------
is_strictly_tril : array-like, shape=[...,]
Boolean evaluating if the matrix is strictly lower triangular
"""
return cls.equal(mat, gs.tril(mat, k=-1), atol)
[docs]
@classmethod
def is_strictly_upper_triangular(cls, mat, atol=gs.atol):
"""Check if a square matrix is strictly upper triangular.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default : backend atol.
Returns
-------
is_strictly_triu : array-like, shape=[...,]
Boolean evaluating if the matrix is strictly upper triangular
"""
return cls.equal(mat, gs.triu(mat, k=1))
[docs]
@classmethod
def is_symmetric(cls, mat, atol=gs.atol):
"""Check if a square matrix is symmetric.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default: backend atol.
Returns
-------
is_sym : array-like, shape=[...,]
Boolean evaluating if the matrix is symmetric.
"""
return cls.equal(mat, cls.transpose(mat), atol)
[docs]
@classmethod
def is_pd(cls, mat):
"""Check if a matrix is positive definite.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
is_pd : array-like, shape=[...,]
Boolean evaluating if the matrix is positive definite.
"""
if mat.ndim == 2:
return gs.array(gs.linalg.is_single_matrix_pd(mat))
shape = mat.shape
if mat.ndim > 3:
mat = gs.reshape(mat, (-1,) + shape[-2:])
is_pd = gs.array([gs.linalg.is_single_matrix_pd(m) for m in mat])
return gs.reshape(is_pd, shape[:-2])
[docs]
@classmethod
def is_spd(cls, mat, atol=gs.atol):
"""Check if a matrix is symmetric positive definite.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default: backend atol.
Returns
-------
is_spd : array-like, shape=[...,]
Boolean evaluating if the matrix is symmetric positive definite.
"""
return gs.logical_and(cls.is_symmetric(mat, atol), cls.is_pd(mat))
[docs]
@classmethod
def is_skew_symmetric(cls, mat, atol=gs.atol):
"""Check if a square matrix is skew symmetric.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
atol : float
Absolute tolerance.
Optional, default: backend atol.
Returns
-------
is_skew_sym : array-like, shape=[...,]
Boolean evaluating if the matrix is skew-symmetric.
"""
return cls.equal(mat, -cls.transpose(mat), atol)
[docs]
@classmethod
def to_diagonal(cls, mat):
"""Make a matrix diagonal.
Make a matrix diagonal by zeroing out non
diagonal elements.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
diag : array-like, shape=[..., n, n]
"""
return cls.to_upper_triangular(cls.to_lower_triangular(mat))
[docs]
@classmethod
def to_lower_triangular(cls, mat):
"""Make a matrix lower triangular.
Make a matrix lower triangular by zeroing
out upper elements.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
tril : array-like, shape=[..., n, n]
Lower triangular matrix.
"""
return gs.tril(mat)
[docs]
@classmethod
def to_upper_triangular(cls, mat):
"""Make a matrix upper triangular.
Make a matrix upper triangular by zeroing
out lower elements.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
triu : array-like, shape=[..., n, n]
"""
return gs.triu(mat)
[docs]
@classmethod
def to_strictly_lower_triangular(cls, mat):
"""Make a matrix strictly lower triangular.
Make a matrix stricly lower triangular by zeroing
out upper and diagonal elements.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
tril : array-like, shape=[..., n, n]
Lower triangular matrix.
"""
return gs.tril(mat, k=-1)
[docs]
@classmethod
def to_strictly_upper_triangular(cls, mat):
"""Make a matrix stritcly upper triangular.
Make a matrix strictly upper triangular by zeroing
out lower and diagonal elements.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
triu : array-like, shape=[..., n, n]
"""
return gs.triu(mat, k=1)
[docs]
@classmethod
def to_symmetric(cls, mat):
"""Make a matrix symmetric.
Make a matrix symmetric by averaging it
with its transpose.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
sym : array-like, shape=[..., n, n]
Symmetric matrix.
"""
return 1 / 2 * (mat + cls.transpose(mat))
[docs]
@classmethod
def to_skew_symmetric(cls, mat):
"""Make a matrix skew-symmetric.
Make matrix skew-symmetric by averaging it
with minus its transpose.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
skew_sym : array-like, shape=[..., n, n]
Skew-symmetric matrix.
"""
return 1 / 2 * (mat - cls.transpose(mat))
[docs]
@classmethod
def to_lower_triangular_diagonal_scaled(cls, mat, K=2.0):
"""Make a matrix lower triangular.
Make matrix lower triangular by zeroing out
upper elements and divide diagonal by factor K.
Parameters
----------
mat : array-like, shape=[..., n, n]
Matrix.
Returns
-------
tril : array-like, shape=[..., n, n]
Lower triangular matrix.
"""
slt = cls.to_strictly_lower_triangular(mat)
diag = cls.to_diagonal(mat) / K
return slt + diag
[docs]
def random_point(self, n_samples=1, bound=1.0):
"""Sample from a uniform distribution in a cube.
Parameters
----------
n_samples : int
Number of samples.
Optional, default: 1.
bound : float
Bound of the interval in which to sample each entry.
Optional, default: 1.
Returns
-------
point : array-like, shape=[..., m, n]
Sample.
"""
m, n = self.m, self.n
size = (n_samples, m, n) if n_samples != 1 else (m, n)
point = bound * (gs.random.rand(*size) - 0.5)
return point
[docs]
@classmethod
def congruent(cls, mat_1, mat_2):
r"""Compute the congruent action of mat_2 on mat_1.
This is :math:`mat\_2 \ mat\_1 \ mat\_2^T`.
Parameters
----------
mat_1 : array-like, shape=[..., n, n]
Matrix.
mat_2 : array-like, shape=[..., n, n]
Matrix.
Returns
-------
cong : array-like, shape=[..., n, n]
Result of the congruent action.
"""
return cls.mul(mat_2, mat_1, cls.transpose(mat_2))
[docs]
@staticmethod
def frobenius_product(mat_1, mat_2):
"""Compute Frobenius inner-product of two matrices.
The `einsum` function is used to avoid computing a matrix product. It
is also faster than using a sum an element-wise product.
Parameters
----------
mat_1 : array-like, shape=[..., m, n]
Matrix.
mat_2 : array-like, shape=[..., m, n]
Matrix.
Returns
-------
product : array-like, shape=[...,]
Frobenius inner-product of mat_1 and mat_2
"""
return gs.einsum("...ij,...ij->...", mat_1, mat_2)
[docs]
@staticmethod
def trace_product(mat_1, mat_2):
"""Compute trace of the product of two matrices.
The `einsum` function is used to avoid computing a matrix product. It
is also faster than using a sum an element-wise product.
Parameters
----------
mat_1 : array-like, shape=[..., m, n]
Matrix.
mat_2 : array-like, shape=[..., m, n]
Matrix.
Returns
-------
product : array-like, shape=[...,]
Frobenius inner-product of mat_1 and mat_2
"""
return gs.einsum("...ij,...ji->...", mat_1, mat_2)
[docs]
@staticmethod
def flatten(mat):
"""Return a flattened form of the matrix.
Flatten a matrix (compatible with vectorization on data axis 0).
The reverse operation is reshape. These operations are often called
matrix vectorization / matricization in mathematics
(https://en.wikipedia.org/wiki/Tensor_reshaping).
The names flatten / reshape were chosen to avoid confusion with
vectorization on data axis 0.
Parameters
----------
mat : array-like, shape=[..., m, n]
Matrix.
Returns
-------
vec : array-like, shape=[..., m * n]
Flatten copy of mat.
"""
batch_shape = mat.shape[:-2]
mat_shape = mat.shape[-2:]
return gs.reshape(mat, batch_shape + (math.prod(mat_shape),))
[docs]
def reshape(self, vec):
"""Return a matricized form of the vector.
Matricize a vector (compatible with vectorization on data axis 0).
The reverse operation is matrices.flatten. These operations are often
called matrix vectorization / matricization in mathematics
(https://en.wikipedia.org/wiki/Tensor_reshaping).
The names flatten / reshape were chosen to avoid confusion with
vectorization on data axis 0.
Parameters
----------
vec : array-like, shape=[..., m * n]
Vector.
Returns
-------
mat : array-like, shape=[..., m, n]
Matricized copy of vec.
"""
is_data_vectorized_on_axis_0 = gs.ndim(gs.array(vec)) == 2
if is_data_vectorized_on_axis_0:
vector_size = vec.shape[1]
shape = (vec.shape[0], self.m, self.n)
else:
vector_size = vec.shape[0]
shape = (
self.m,
self.n,
)
if vector_size != self.m * self.n:
raise ValueError("Incompatible vector and matrix sizes")
return gs.reshape(vec, shape)
[docs]
@classmethod
def align_matrices(cls, point, base_point):
"""Align matrices.
Find the optimal rotation R in SO(m) such that the base point and
R.point are well positioned.
Parameters
----------
point : array-like, shape=[..., m, n]
Point on the manifold.
base_point : array-like, shape=[..., m, n]
Point on the manifold.
Returns
-------
aligned : array-like, shape=[..., m, n]
R.point.
"""
mat = gs.matmul(cls.transpose(point), base_point)
left, singular_values, right = gs.linalg.svd(mat, full_matrices=False)
det = gs.linalg.det(mat)
conditioning = (
singular_values[..., -2] + gs.sign(det) * singular_values[..., -1]
) / singular_values[..., 0]
if gs.any(conditioning < gs.atol):
logging.warning(
f"Singularity close, ill-conditioned matrix "
f"encountered: "
f"{conditioning[conditioning < 1e-10]}"
)
if gs.any(gs.isclose(conditioning, 0.0)):
logging.warning("Alignment matrix is not unique.")
flipped = flip_determinant(cls.transpose(right), det)
return Matrices.mul(point, left, cls.transpose(flipped))
[docs]
class MatricesMetric(EuclideanMetric):
"""Euclidean metric on matrices given by Frobenius inner-product."""
[docs]
def inner_product(self, tangent_vec_a, tangent_vec_b, base_point=None):
"""Compute Frobenius inner-product of two tangent vectors.
Parameters
----------
tangent_vec_a : array-like, shape=[..., m, n]
Tangent vector.
tangent_vec_b : array-like, shape=[..., m, n]
Tangent vector.
base_point : array-like, shape=[..., m, n]
Base point.
Optional, default: None.
Returns
-------
inner_prod : array-like, shape=[...,]
Frobenius inner-product of tangent_vec_a and tangent_vec_b.
"""
inner_prod = Matrices.frobenius_product(tangent_vec_a, tangent_vec_b)
return repeat_out(
self._space.point_ndim, inner_prod, tangent_vec_a, tangent_vec_b, base_point
)
[docs]
def squared_norm(self, vector, base_point=None):
"""Compute the square of the norm of a vector.
Squared norm of a vector associated to the inner product
at the tangent space at a base point.
Parameters
----------
vector : array-like, shape=[..., dim]
Vector.
base_point : array-like, shape=[..., dim]
Base point.
Optional, default: None.
Returns
-------
sq_norm : array-like, shape=[...,]
Squared norm.
"""
init_axis = len(vector.shape[:-2])
axis = (init_axis, init_axis + 1)
sq_norm = gs.linalg.norm(vector, axis=axis) ** 2
return repeat_out(self._space.point_ndim, sq_norm, vector, base_point)
[docs]
def norm(self, vector, base_point=None):
"""Compute norm of a matrix.
Norm of a matrix associated to the Frobenius inner product.
Parameters
----------
vector : array-like, shape=[..., dim]
Vector.
base_point : array-like, shape=[..., dim]
Base point.
Optional, default: None.
Returns
-------
norm : array-like, shape=[...,]
Norm.
"""
norm = gs.linalg.norm(vector, axis=(-2, -1))
return repeat_out(self._space.point_ndim, norm, vector, base_point)
[docs]
class MatricesDiagMetric(EuclideanMetric):
"""Flat metric on matrices given by a diagonal metric matrix.
Parameters
----------
space : MatrixVectorSpace
metric_coeffs: array-like, shape=[m,n]
"""
def __init__(self, space, metric_coeffs=None):
super().__init__(space)
self._check_metric_coeffs(space, metric_coeffs)
if metric_coeffs is None:
n = space.n
m = space.m if hasattr(space, "m") else n
metric_coeffs = gs.ones((m, n))
self.metric_coeffs = metric_coeffs
@staticmethod
def _check_metric_coeffs(space, metric_coeffs):
"""Check shape of metric coefficients."""
if metric_coeffs is None:
return
n = space.n
m = space.m if hasattr(space, "m") else n
expected_shape = (m, n)
if metric_coeffs.shape != expected_shape:
raise ValueError(
f"metric_coeffs shape is {metric_coeffs.shape};"
f"expected: {expected_shape}"
)
[docs]
def inner_product(self, tangent_vec_a, tangent_vec_b, base_point=None):
"""Compute Frobenius inner-product of two tangent vectors.
Parameters
----------
tangent_vec_a : array-like, shape=[..., m, n]
Tangent vector.
tangent_vec_b : array-like, shape=[..., m, n]
Tangent vector.
base_point : array-like, shape=[..., m, n]
Base point.
Optional, default: None.
Returns
-------
inner_prod : array-like, shape=[...,]
Frobenius inner-product of tangent_vec_a and tangent_vec_b.
Notes
-----
- Implementation relies on fact that the metric_matrix is diagonal, i.e.
we just need to multiply each element of the tangent vector by the corresponding
diagonal coefficient. (Avoids unnecessary reshapings.)
"""
inner_prod = Matrices.frobenius_product(
tangent_vec_a, self.metric_coeffs * tangent_vec_b
)
return repeat_out(2, inner_prod, tangent_vec_a, tangent_vec_b, base_point)