pygeoinf 1.3.4__py3-none-any.whl → 1.3.6__py3-none-any.whl

pygeoinf/__init__.py

@@ -76,9 +76,14 @@ from .forward_problem import ForwardProblem, LinearForwardProblem
 from .linear_optimisation import (
     LinearLeastSquaresInversion,
     LinearMinimumNormInversion,
+    ConstrainedLinearLeastSquaresInversion,
+    ConstrainedLinearMinimumNormInversion,
 )
 
-from .linear_bayesian import LinearBayesianInversion, LinearBayesianInference
+from .linear_bayesian import (
+    LinearBayesianInversion,
+    ConstrainedLinearBayesianInversion,
+)
 
 from .backus_gilbert import HyperEllipsoid
 
@@ -87,6 +92,8 @@ from .nonlinear_optimisation import (
 )
 
 
+from .subspaces import OrthogonalProjector, AffineSubspace, LinearSubspace
+
 __all__ = [
     # random_matrix
     "fixed_rank_random_range",
@@ -144,11 +151,17 @@ __all__ = [
     # linear_optimisation
     "LinearLeastSquaresInversion",
     "LinearMinimumNormInversion",
+    "ConstrainedLinearLeastSquaresInversion",
+    "ConstrainedLinearMinimumNormInversion",
     # linear_bayesian
     "LinearBayesianInversion",
-    "LinearBayesianInference",
+    "ConstrainedLinearBayesianInversion",
     # backus_gilbert
     "HyperEllipsoid",
     # nonlinear_optimisation
     "ScipyUnconstrainedOptimiser",
+    # Subspaces
+    "OrthogonalProjector",
+    "AffineSubspace",
+    "LinearSubspace",
 ]

pygeoinf/linear_bayesian.py

@@ -5,19 +5,12 @@ This module treats the inverse problem from a statistical perspective, aiming to
 determine the full posterior probability distribution of the unknown model
 parameters, rather than a single best-fit solution.
 
-It assumes that the prior knowledge about the model and the statistics of the
-data errors can be described by Gaussian measures. For a linear forward problem,
-the resulting posterior distribution for the model is also Gaussian, allowing
-for an analytical solution.
-
 Key Classes
 -----------
 - `LinearBayesianInversion`: Computes the posterior Gaussian measure `p(u|d)`
-  for the model `u` given observed data `d`. This provides not only a mean
-  estimate for the model but also its uncertainty (covariance).
-- `LinearBayesianInference`: Extends the framework to compute the posterior
-  distribution for a derived property of the model, `p(B(u)|d)`, where `B` is
-  some linear operator.
+  for the model `u` given observed data `d`.
+- `ConstrainedLinearBayesianInversion`: Solves the inverse problem subject to
+  an affine constraint `u in A`.
 """
 
 from __future__ import annotations
@@ -25,22 +18,19 @@ from typing import Optional
 
 from .inversion import LinearInversion
 from .gaussian_measure import GaussianMeasure
-
-
 from .forward_problem import LinearForwardProblem
-from .linear_operators import LinearOperator
+from .linear_operators import LinearOperator, NormalSumOperator
 from .linear_solvers import LinearSolver, IterativeLinearSolver
-from .hilbert_space import HilbertSpace, Vector
+from .hilbert_space import Vector
+from .subspaces import AffineSubspace
 
 
 class LinearBayesianInversion(LinearInversion):
     """
     Solves a linear inverse problem using Bayesian methods.
 
-    This class applies to problems of the form `d = A(u) + e`, where the prior
-    knowledge of the model `u` and the statistics of the error `e` are described
-    by Gaussian distributions. It computes the full posterior probability
-    distribution `p(u|d)` for the model parameters given an observation `d`.
+    This class applies to problems of the form `d = A(u) + e`. It computes the
+    full posterior probability distribution `p(u|d)`.
     """
 
     def __init__(
@@ -65,45 +55,43 @@ class LinearBayesianInversion(LinearInversion):
     @property
     def normal_operator(self) -> LinearOperator:
         """
-        Returns the covariance of the prior predictive distribution, `p(d)`.
-
-        This operator, `C_d = A @ C_u @ A* + C_e`, represents the total
-        expected covariance in the data space before any data is observed.
-        Its inverse is central to calculating the posterior distribution and is
-        often referred to as the Bayesian normal operator.
+        Returns the Bayesian Normal operator: N = A Q A* + R.
         """
         forward_operator = self.forward_problem.forward_operator
-        prior_model_covariance = self.model_prior_measure.covariance
+        model_prior_covariance = self.model_prior_measure.covariance
 
         if self.forward_problem.data_error_measure_set:
             return (
-                forward_operator @ prior_model_covariance @ forward_operator.adjoint
+                forward_operator @ model_prior_covariance @ forward_operator.adjoint
                 + self.forward_problem.data_error_measure.covariance
             )
         else:
-            return forward_operator @ prior_model_covariance @ forward_operator.adjoint
+            return NormalSumOperator(forward_operator, model_prior_covariance)
 
-    def data_prior_measure(self) -> GaussianMeasure:
+    def kalman_operator(
+        self,
+        solver: LinearSolver,
+        /,
+        *,
+        preconditioner: Optional[LinearOperator] = None,
+    ) -> LinearOperator:
         """
-        Returns the prior predictive distribution on the data, `p(d)`.
-
-        This measure describes the expected distribution of the data before any
-        specific observation is made, combining the uncertainty from the prior
-        model and the data errors.
+        Returns the Kalman gain operator K = Q A* N^-1.
         """
-        if self.forward_problem.data_error_measure_set:
-            # d = A(u) + e  =>  p(d) is convolution of p(A(u)) and p(e)
-            return (
-                self.model_prior_measure.affine_mapping(
-                    operator=self.forward_problem.forward_operator
-                )
-                + self.forward_problem.data_error_measure
+        forward_operator = self.forward_problem.forward_operator
+        model_prior_covariance = self.model_prior_measure.covariance
+        normal_operator = self.normal_operator
+
+        if isinstance(solver, IterativeLinearSolver):
+            inverse_normal_operator = solver(
+                normal_operator, preconditioner=preconditioner
             )
         else:
-            # d = A(u)  => p(d) is just the mapping of the model prior
-            return self.model_prior_measure.affine_mapping(
-                operator=self.forward_problem.forward_operator
-            )
+            inverse_normal_operator = solver(normal_operator)
+
+        return (
+            model_prior_covariance @ forward_operator.adjoint @ inverse_normal_operator
+        )
 
     def model_posterior_measure(
         self,
@@ -114,108 +102,120 @@ class LinearBayesianInversion(LinearInversion):
         preconditioner: Optional[LinearOperator] = None,
     ) -> GaussianMeasure:
         """
-        Returns the posterior Gaussian measure for the model, `p(u|d)`.
-
-        This measure represents our updated state of knowledge about the model
-        `u` after observing the data `d`. Its expectation is the most likely
-        model, and its covariance quantifies the remaining uncertainty.
+        Returns the posterior Gaussian measure p(u|d).
 
         Args:
             data: The observed data vector.
             solver: A linear solver for inverting the normal operator.
-            preconditioner: An optional preconditioner for iterative solvers.
-
-        Returns:
-            The posterior `GaussianMeasure` on the model space.
+            preconditioner: An optional preconditioner.
         """
         data_space = self.data_space
         model_space = self.model_space
         forward_operator = self.forward_problem.forward_operator
-        prior_model_covariance = self.model_prior_measure.covariance
-        normal_operator = self.normal_operator
+        model_prior_covariance = self.model_prior_measure.covariance
 
-        if isinstance(solver, IterativeLinearSolver):
-            inverse_normal_operator = solver(
-                normal_operator, preconditioner=preconditioner
-            )
-        else:
-            inverse_normal_operator = solver(normal_operator)
+        # 1. Compute Kalman Gain
+        kalman_gain = self.kalman_operator(solver, preconditioner=preconditioner)
 
-        # Calculate posterior mean: mu_post = mu_u + C_u*A^T*C_d^-1*(d - A*mu_u - mu_e)
+        # 2. Compute Posterior Mean
+        # Shift data: d - A(mu_u)
         shifted_data = data_space.subtract(
             data, forward_operator(self.model_prior_measure.expectation)
         )
+
+        # Shift for noise mean: d - A(mu_u) - mu_e
         if self.forward_problem.data_error_measure_set:
-            shifted_data = data_space.subtract(
-                shifted_data, self.forward_problem.data_error_measure.expectation
-            )
+            error_expectation = self.forward_problem.data_error_measure.expectation
+            shifted_data = data_space.subtract(shifted_data, error_expectation)
+        else:
+            error_expectation = data_space.zero
 
-        mean_update = (
-            prior_model_covariance @ forward_operator.adjoint @ inverse_normal_operator
-        )(shifted_data)
+        mean_update = kalman_gain(shifted_data)
         expectation = model_space.add(self.model_prior_measure.expectation, mean_update)
 
-        # Calculate posterior covariance: C_post = C_u - C_u*A^T*C_d^-1*A*C_u
-        covariance = prior_model_covariance - (
-            prior_model_covariance
-            @ forward_operator.adjoint
-            @ inverse_normal_operator
-            @ forward_operator
-            @ prior_model_covariance
+        # 3. Compute Posterior Covariance (Implicitly)
+        # C_post = C_u - K A C_u
+        covariance = model_prior_covariance - (
+            kalman_gain @ forward_operator @ model_prior_covariance
+        )
+
+        # 4. Set up Posterior Sampling
+        # Logic: Can sample if prior is samplable AND (noise is absent OR samplable)
+        can_sample_prior = self.model_prior_measure.sample_set
+        can_sample_noise = (
+            not self.forward_problem.data_error_measure_set
+            or self.forward_problem.data_error_measure.sample_set
         )
 
-        return GaussianMeasure(covariance=covariance, expectation=expectation)
+        if can_sample_prior and can_sample_noise:
+
+            def sample():
+                # a. Sample Prior
+                model_sample = self.model_prior_measure.sample()
 
+                # b. Calculate Residual
+                prediction = forward_operator(model_sample)
+                data_residual = data_space.subtract(data, prediction)
 
-class LinearBayesianInference(LinearBayesianInversion):
+                # c. Perturb Residual
+                if self.forward_problem.data_error_measure_set:
+                    noise_raw = self.forward_problem.data_error_measure.sample()
+                    epsilon = data_space.subtract(noise_raw, error_expectation)
+                    data_space.axpy(1.0, epsilon, data_residual)
+
+                # d. Update
+                correction = kalman_gain(data_residual)
+                return model_space.add(model_sample, correction)
+
+            return GaussianMeasure(
+                covariance=covariance, expectation=expectation, sample=sample
+            )
+        else:
+            return GaussianMeasure(covariance=covariance, expectation=expectation)
+
+
+class ConstrainedLinearBayesianInversion(LinearInversion):
     """
-    Performs Bayesian inference on a derived property of the model.
+    Solves a linear inverse problem subject to an affine subspace constraint.
 
-    While `LinearBayesianInversion` solves for the model `u` itself, this class
-    computes the posterior distribution for a property `p = B(u)`, where `B` is a
-    linear operator acting on the model `u`. This is useful for uncertainty
-    quantification of derived quantities (e.g., the average value of a field).
+    This class enforces the constraint `u in A` using either:
+    1. Bayesian Conditioning (Default): p(u | d, u in A).
+       If A is defined geometrically (no explicit equation), an implicit
+       operator (I-P) is used, which requires a robust solver in the subspace.
+    2. Geometric Projection: Projects the unconstrained posterior onto A.
     """
 
     def __init__(
         self,
         forward_problem: LinearForwardProblem,
         model_prior_measure: GaussianMeasure,
-        property_operator: LinearOperator,
+        constraint: AffineSubspace,
         /,
+        *,
+        geometric: bool = False,
     ) -> None:
         """
         Args:
-            forward_problem: The forward problem linking the model to the data.
-            model_prior_measure: The prior Gaussian measure on the model space.
-            property_operator: The linear operator `B` that maps a model `u` to
-                a property `p`.
+            forward_problem: The forward problem.
+            model_prior_measure: The unconstrained prior Gaussian measure.
+            constraint: The affine subspace A.
+            geometric: If True, uses orthogonal projection (Euclidean metric).
+                       If False (default), uses Bayesian conditioning.
         """
-        super().__init__(forward_problem, model_prior_measure)
-        if property_operator.domain != self.forward_problem.model_space:
-            raise ValueError("Property operator domain must match the model space.")
-        self._property_operator: LinearOperator = property_operator
-
-    @property
-    def property_space(self) -> HilbertSpace:
-        """The Hilbert space in which the property `p` resides."""
-        return self._property_operator.codomain
-
-    @property
-    def property_operator(self) -> LinearOperator:
-        """The linear operator `B` that defines the property."""
-        return self._property_operator
+        super().__init__(forward_problem)
+        self._unconstrained_prior = model_prior_measure
+        self._constraint = constraint
+        self._geometric = geometric
 
-    def property_prior_measure(self) -> GaussianMeasure:
+    def conditioned_prior_measure(self) -> GaussianMeasure:
         """
-        Returns the prior measure on the property space, `p(p)`.
-
-        This is computed by propagating the model prior through the property
-        operator.
+        Computes the prior measure conditioned on the constraint.
         """
-        return self.model_prior_measure.affine_mapping(operator=self.property_operator)
+        return self._constraint.condition_gaussian_measure(
+            self._unconstrained_prior, geometric=self._geometric
+        )
 
-    def property_posterior_measure(
+    def model_posterior_measure(
         self,
         data: Vector,
         solver: LinearSolver,
@@ -224,22 +224,22 @@ class LinearBayesianInference(LinearBayesianInversion):
         preconditioner: Optional[LinearOperator] = None,
     ) -> GaussianMeasure:
         """
-        Returns the posterior measure on the property space, `p(p|d)`.
-
-        This is computed by first finding the posterior measure for the model,
-        `p(u|d)`, and then propagating it through the property operator `B`.
+        Returns the posterior Gaussian measure p(u | d, u in A).
 
         Args:
-            data: The observed data vector.
-            solver: A linear solver for the normal equations.
-            preconditioner: An optional preconditioner for iterative solvers.
+            data: Observed data vector.
+            solver: Solver for the data update (inverts A C_cond A* + Ce).
+            preconditioner: Preconditioner for the data update.
 
-        Returns:
-            The posterior `GaussianMeasure` on the property space.
+        Note: The solver for the constraint update is managed internally by
+        the AffineSubspace object passed at initialization.
         """
-        # First, find the posterior distribution for the model u.
-        model_posterior = self.model_posterior_measure(
+        # 1. Condition Prior
+        cond_prior = self.conditioned_prior_measure()
+
+        # 2. Solve Bayesian Inverse Problem with the new prior
+        bayes_inv = LinearBayesianInversion(self.forward_problem, cond_prior)
+
+        return bayes_inv.model_posterior_measure(
             data, solver, preconditioner=preconditioner
         )
-        # Then, map that distribution to the property space.
-        return model_posterior.affine_mapping(operator=self.property_operator)

pygeoinf/linear_optimisation.py

@@ -16,6 +16,8 @@ Key Classes
 - `LinearMinimumNormInversion`: Finds the model with the smallest norm that
   fits the data to a statistically acceptable degree using the discrepancy
   principle.
+- `ConstrainedLinearLeastSquaresInversion`: Solves a linear inverse problem
+  subject to an affine subspace constraint.
 """
 
 from __future__ import annotations
@@ -30,6 +32,7 @@ from .forward_problem import LinearForwardProblem
 from .linear_operators import LinearOperator
 from .linear_solvers import LinearSolver, IterativeLinearSolver
 from .hilbert_space import Vector
+from .subspaces import AffineSubspace
 
 
 class LinearLeastSquaresInversion(LinearInversion):
@@ -160,6 +163,113 @@ class LinearLeastSquaresInversion(LinearInversion):
             return inverse_normal_operator @ forward_operator.adjoint
 
 
+class ConstrainedLinearLeastSquaresInversion(LinearInversion):
+    """
+    Solves a linear inverse problem subject to an affine subspace constraint.
+
+    Problem:
+        Minimize J(u) = || A(u) - d ||_D^2 + alpha * || u ||_M^2
+        Subject to u in A (Affine Subspace)
+
+    Method:
+        The problem is reduced to an unconstrained minimization in the subspace.
+        We decompose the model as u = u_base + w, where u_base is the element
+        of the affine subspace closest to the origin (orthogonal to the tangent space),
+        and w is a perturbation in the tangent space.
+
+        The cost function separates (due to orthogonality) into:
+        J(w) = || A(w) - (d - A(u_base)) ||^2 + alpha * || w ||^2 + (alpha * ||u_base||^2)
+
+        This is solved using the standard LinearLeastSquaresInversion on a
+        reduced forward problem.
+    """
+
+    def __init__(
+        self, forward_problem: LinearForwardProblem, constraint: AffineSubspace
+    ) -> None:
+        """
+        Args:
+            forward_problem: The original unconstrained forward problem.
+            constraint: The affine subspace A where the solution must lie.
+        """
+        super().__init__(forward_problem)
+        self._constraint = constraint
+
+        # 1. Compute the Orthogonal Base Vector (u_base)
+        # u_base = (I - P) * translation
+        # This is the unique vector in the affine space that is orthogonal to the tangent space.
+        # It ensures ||u||^2 = ||u_base||^2 + ||w||^2, decoupling the regularization.
+        self._u_base = constraint.domain.subtract(
+            constraint.translation, constraint.projector(constraint.translation)
+        )
+
+        # 2. Construct Reduced Forward Problem
+        # Operator: A_tilde = A @ P
+        reduced_operator = forward_problem.forward_operator @ constraint.projector
+
+        # The error measure on the data remains valid for the reduced problem
+        # because the noise model is additive and independent of the model parameters.
+        self._reduced_forward_problem = LinearForwardProblem(
+            reduced_operator,
+            data_error_measure=(
+                forward_problem.data_error_measure
+                if forward_problem.data_error_measure_set
+                else None
+            ),
+        )
+
+        # 3. Initialize the internal unconstrained solver
+        self._unconstrained_inversion = LinearLeastSquaresInversion(
+            self._reduced_forward_problem
+        )
+
+    def least_squares_operator(
+        self,
+        damping: float,
+        solver: LinearSolver,
+        /,
+        **kwargs,
+    ) -> NonLinearOperator:
+        """
+        Returns an operator that maps data to the constrained least-squares solution.
+
+        Args:
+            damping: The Tikhonov damping parameter.
+            solver: The linear solver for the reduced normal equations.
+            **kwargs: Additional arguments passed to the solver (e.g., preconditioner).
+
+        Returns:
+            A NonLinearOperator mapping d -> u_constrained.
+        """
+
+        # Get the operator L_tilde such that w = L_tilde(d_tilde)
+        reduced_op = self._unconstrained_inversion.least_squares_operator(
+            damping, solver, **kwargs
+        )
+
+        # Precompute A(u_base) to shift the data efficiently
+        # This represents the data predicted by the "base" model.
+        data_offset = self.forward_problem.forward_operator(self._u_base)
+
+        domain = self.data_space
+        codomain = self.model_space
+
+        def mapping(d: Vector) -> Vector:
+            # 1. Shift Data: d_tilde = d - A(u_base)
+            d_tilde = domain.subtract(d, data_offset)
+
+            # 2. Solve for perturbation w in the tangent space
+            # w = (P A* A P + alpha I)^-1 P A* d_tilde
+            w = reduced_op(d_tilde)
+
+            # 3. Reconstruct full model: u = u_base + w
+            # Note: w is guaranteed to be in the tangent space (Range of P)
+            # because of the structure of the reduced normal equations.
+            return codomain.add(self._u_base, w)
+
+        return NonLinearOperator(domain, codomain, mapping)
+
+
 class LinearMinimumNormInversion(LinearInversion):
     """
     Finds a regularized solution using the discrepancy principle.
@@ -309,3 +419,111 @@ class LinearMinimumNormInversion(LinearInversion):
             normal_operator = forward_operator @ forward_operator.adjoint
             inverse_normal_operator = solver(normal_operator)
             return forward_operator.adjoint @ inverse_normal_operator
+
+
+class ConstrainedLinearMinimumNormInversion(LinearInversion):
+    """
+    Finds the minimum-norm solution subject to an affine subspace constraint
+    using the discrepancy principle.
+
+    Problem:
+        Minimize ||u||
+        Subject to u in A (Affine Subspace)
+        And chi_squared(u, d) <= critical_value
+
+    Method:
+        We decompose the model as u = u_base + w, where u_base is the element
+        of the affine subspace with the smallest norm (orthogonal to the tangent
+        space), and w is a perturbation in the tangent space.
+
+        Because u_base and w are orthogonal, ||u||^2 = ||u_base||^2 + ||w||^2.
+        Minimizing ||u|| is therefore equivalent to minimizing ||w||.
+
+        The problem reduces to finding the minimum norm w such that:
+        || A(w) - (d - A(u_base)) ||_D^2 <= critical_value
+
+        This is solved using the standard LinearMinimumNormInversion on a
+        reduced forward problem.
+    """
+
+    def __init__(
+        self,
+        forward_problem: LinearForwardProblem,
+        constraint: AffineSubspace,
+    ) -> None:
+        """
+        Args:
+            forward_problem: The original unconstrained forward problem.
+            constraint: The affine subspace A where the solution must lie.
+        """
+        super().__init__(forward_problem)
+        if self.forward_problem.data_error_measure_set:
+            self.assert_inverse_data_covariance()
+
+        self._constraint = constraint
+
+        # 1. Compute the Orthogonal Base Vector (u_base)
+        # u_base = (I - P) * translation
+        # This is the vector in the affine space closest to the origin.
+        self._u_base = constraint.domain.subtract(
+            constraint.translation, constraint.projector(constraint.translation)
+        )
+
+        # 2. Construct Reduced Forward Problem
+        # Operator: A_tilde = A @ P
+        reduced_operator = forward_problem.forward_operator @ constraint.projector
+
+        self._reduced_forward_problem = LinearForwardProblem(
+            reduced_operator,
+            data_error_measure=(
+                forward_problem.data_error_measure
+                if forward_problem.data_error_measure_set
+                else None
+            ),
+        )
+
+        # 3. Initialize the internal unconstrained solver
+        self._unconstrained_inversion = LinearMinimumNormInversion(
+            self._reduced_forward_problem
+        )
+
+    def minimum_norm_operator(
+        self,
+        solver: LinearSolver,
+        /,
+        **kwargs,
+    ) -> NonLinearOperator:
+        """
+        Returns an operator that maps data to the constrained minimum-norm solution.
+
+        Args:
+            solver: The linear solver for the reduced normal equations.
+            **kwargs: Arguments passed to LinearMinimumNormInversion (e.g.,
+                      significance_level, rtol, maxiter).
+
+        Returns:
+            A NonLinearOperator mapping d -> u_constrained.
+        """
+
+        # Get the operator L_tilde such that w = L_tilde(d_tilde)
+        reduced_op = self._unconstrained_inversion.minimum_norm_operator(
+            solver, **kwargs
+        )
+
+        # Precompute A(u_base) to shift the data
+        data_offset = self.forward_problem.forward_operator(self._u_base)
+
+        domain = self.data_space
+        codomain = self.model_space
+
+        def mapping(d: Vector) -> Vector:
+            # 1. Shift Data: d_tilde = d - A(u_base)
+            d_tilde = domain.subtract(d, data_offset)
+
+            # 2. Solve for perturbation w in the tangent space
+            w = reduced_op(d_tilde)
+
+            # 3. Reconstruct full model: u = u_base + w
+            return codomain.add(self._u_base, w)
+
+        return NonLinearOperator(domain, codomain, mapping)