"""Belief update in a matrix-based inference view where the information is given by
matrix-vector multiplication."""
import numpy as np
import probnum # pylint: disable="unused-import"
from probnum import linops, randvars
from probnum.linalg.solvers.beliefs import LinearSystemBelief
from .._linear_solver_belief_update import LinearSolverBeliefUpdate
class MatrixBasedLinearBeliefUpdate(LinearSolverBeliefUpdate):
r"""Gaussian belief update in a matrix-based inference framework assuming linear information.
Updates the belief over the quantities of interest of a linear system :math:`Ax=b` given matrix-variate Gaussian beliefs with Kronecker covariance structure and linear observations :math:`y=As`. The belief update computes :math:`p(M \mid y) = \mathcal{N}(M; M_{i+1}, V \otimes W_{i+1})`, [1]_ [2]_ such that
.. math ::
\begin{align}
M_{i+1} &= M_i + (y - M_i s) (s^\top W_i s)^\dagger s^\top W_i,\\
W_{i+1} &= W_i - W_i s (s^\top W_i s)^\dagger s^\top W_i.
\end{align}
References
----------
.. [1] Hennig, P., Probabilistic Interpretation of Linear Solvers, *SIAM Journal on
Optimization*, 2015, 25, 234-260
.. [2] Wenger, J. and Hennig, P., Probabilistic Linear Solvers for Machine Learning,
*Advances in Neural Information Processing Systems (NeurIPS)*, 2020
"""
[docs] def __call__(
self, solver_state: "probnum.linalg.solvers.LinearSolverState"
) -> LinearSystemBelief:
# Inference for A
A = self._matrix_based_update(
matrix=solver_state.belief.A,
action=solver_state.action,
observ=solver_state.observation,
)
# Inference for Ainv (interpret action and observation as swapped)
Ainv = self._matrix_based_update(
matrix=solver_state.belief.Ainv,
action=solver_state.observation,
observ=solver_state.action,
)
if solver_state.belief.b is None:
b = randvars.Constant(solver_state.problem.b)
else:
b = solver_state.belief.b
return LinearSystemBelief(A=A, Ainv=Ainv, x=None, b=b)
def _matrix_based_update(
self, matrix: randvars.Normal, action: np.ndarray, observ: np.ndarray
) -> randvars.Normal:
"""Matrix-based inference update for linear information."""
if not isinstance(matrix.cov, linops.Kronecker):
raise ValueError(
f"Covariance must have Kronecker structure, but is '{type(matrix.cov).__name__}'."
)
pred = matrix.mean @ action
resid = observ - pred
covfactor_Ms = matrix.cov.B @ action
gram = action.T @ covfactor_Ms
gram_pinv = 1.0 / gram if gram > 0.0 else 0.0
gain = covfactor_Ms * gram_pinv
covfactor_update = linops.aslinop(gain[:, None]) @ linops.aslinop(
covfactor_Ms[None, :]
)
resid_gain = linops.aslinop(resid[:, None]) @ linops.aslinop(
gain[None, :]
) # residual and gain are flipped due to matrix vectorization
return randvars.Normal(
mean=matrix.mean + resid_gain,
cov=linops.Kronecker(A=matrix.cov.A, B=matrix.cov.B - covfactor_update),
)