Source code for probnum.utils.linalg._cholesky_updates

"""Cholesky updates."""


import typing

import numpy as np

__all__ = ["cholesky_update", "tril_to_positive_tril"]


[docs]def cholesky_update( S1: np.ndarray, S2: typing.Optional[np.ndarray] = None ) -> np.ndarray: r"""Compute Cholesky update/factorization :math:`L` such that :math:`L L^\top = S_1 S_1^\top + S_2 S_2^\top` holds. This can be used in various ways. For example, :math:`S_1` and :math:`S_2` do not need to be Cholesky factors; any matrix square-root is sufficient. As long as :math:`L L^\top = S_1 S_1^\top + S_2 S_2^\top` is well-defined (and admits a Cholesky-decomposition), :math:`S_1` and :math:`S_2` do not even have to be square. Parameters ---------- S1 : First matrix square-root. Not necessarily a Cholesky factor, any (possibly even non-square) matrix :math:`S` such that :math:`C = S S^\top` holds, is sufficient. S2 : Second matrix square-root. Not necessarily a Cholesky factor, any (possibly even non-square) matrix :math:`S` such that :math:`C = S S^\top` holds, is sufficient. Optional. Default is None. Returns ------- Lower Cholesky factor :math:`L` of :math:`L L^\top =S1 S1^\top + S2 S2^\top`, if ``S2`` was not None. Otherwise, lower Cholesky factor of :math:`L L^\top =S1 S1^\top`. Examples -------- >>> from probnum.utils.linalg import cholesky_update >>> from probnum.problems.zoo.linalg import random_spd_matrix >>> import numpy as np Compute the Cholesky-factor of a sum of SPD matrices. >>> rng = np.random.default_rng(seed=3) >>> C1 = random_spd_matrix(rng, dim=5) >>> S1 = np.linalg.cholesky(C1) >>> C2 = random_spd_matrix(rng, dim=5) >>> S2 = np.linalg.cholesky(C2) >>> C = C1 + C2 >>> S = cholesky_update(S1, S2) >>> np.allclose(np.linalg.cholesky(C), S) True Turn a (potentially non-square) matrix square-root into a Cholesky factor >>> A = np.random.rand(3, 5) >>> S = cholesky_update(A @ S1) >>> np.allclose(np.linalg.cholesky(A @ C1 @ A.T), S) True """ if S2 is not None: stacked_up = np.vstack((S1.T, S2.T)) else: stacked_up = np.vstack(S1.T) upper_sqrtm = np.linalg.qr(stacked_up, mode="r") if S1.ndim == 1: lower_sqrtm = upper_sqrtm.T elif S1.shape[0] <= S1.shape[1]: lower_sqrtm = upper_sqrtm.T else: lower_sqrtm = np.zeros((S1.shape[0], S1.shape[0])) lower_sqrtm[:, : -(S1.shape[0] - S1.shape[1])] = upper_sqrtm.T return tril_to_positive_tril(lower_sqrtm)
[docs]def tril_to_positive_tril(tril_mat: np.ndarray) -> np.ndarray: r"""Orthogonally transform a lower-triangular matrix into a lower-triangular matrix with positive diagonal. In other words, make it a valid lower Cholesky factor. The name of the function is based on `np.tril`. Parameters ---------- tril_mat: A lower-triangular matrix. """ d = np.sign(np.diag(tril_mat)) # Numpy assigns sign 0 to 0.0, which eliminate entire rows in the operation below. d[d == 0] = 1.0 # Fast(er) multiplication with a diagonal matrix from the right via broadcasting. with_pos_diag = tril_mat * d[None, :] return with_pos_diag