pygeoinf 1.3.5__py3-none-any.whl → 1.3.7__py3-none-any.whl

pygeoinf/linear_solvers.py

@@ -524,3 +524,433 @@ class CGSolver(IterativeLinearSolver):
                 self._callback(x)
 
         return x
+
+
+class MinResSolver(IterativeLinearSolver):
+    """
+    A matrix-free implementation of the MINRES algorithm.
+
+    Suitable for symmetric, possibly indefinite or singular linear systems.
+    It minimizes the norm of the residual ||r|| in each step using the
+    Hilbert space's native inner product.
+    """
+
+    def __init__(
+        self,
+        /,
+        *,
+        preconditioning_method: LinearSolver = None,
+        rtol: float = 1.0e-5,
+        atol: float = 1.0e-8,
+        maxiter: Optional[int] = None,
+    ) -> None:
+        super().__init__(preconditioning_method=preconditioning_method)
+        self._rtol = rtol
+        self._atol = atol
+        self._maxiter = maxiter
+
+    def solve_linear_system(
+        self,
+        operator: LinearOperator,
+        preconditioner: Optional[LinearOperator],
+        y: Vector,
+        x0: Optional[Vector],
+    ) -> Vector:
+        domain = operator.domain
+
+        # Initial setup using HilbertSpace methods
+        x = domain.zero if x0 is None else domain.copy(x0)
+        r = domain.subtract(y, operator(x))
+
+        # Initial preconditioned residual: z = M^-1 r
+        z = domain.copy(r) if preconditioner is None else preconditioner(r)
+
+        # beta_1 = sqrt(r.T @ M^-1 @ r)
+        gamma_curr = np.sqrt(domain.inner_product(r, z))
+        if gamma_curr < self._atol:
+            return x
+
+        gamma_1 = gamma_curr  # Store initial residual norm for relative tolerance
+
+        # Lanczos vectors: v_curr is M^-1-scaled basis vector
+        v_prev = domain.zero
+        v_curr = domain.multiply(1.0 / gamma_curr, z)
+
+        # QR decomposition variables (Givens rotations)
+        phi_bar = gamma_curr
+        c_prev, s_prev = 1.0, 0.0
+        c_curr, s_curr = 1.0, 0.0
+
+        # Direction vectors for solution update
+        w_prev = domain.zero
+        w_curr = domain.zero
+
+        maxiter = self._maxiter if self._maxiter is not None else 10 * domain.dim
+
+        for k in range(maxiter):
+            # --- Lanczos Step ---
+            # Compute A * v_j (where v_j is already preconditioned)
+            Av = operator(v_curr)
+            alpha = domain.inner_product(v_curr, Av)
+
+            # v_next = M^-1 * (A*v_j) - alpha*v_j - gamma_j*v_{j-1}
+            # We apply M^-1 to the operator result to stay in the Krylov space of M^-1 A
+            v_next = domain.copy(Av) if preconditioner is None else preconditioner(Av)
+            domain.axpy(-alpha, v_curr, v_next)
+            if k > 0:
+                domain.axpy(-gamma_curr, v_prev, v_next)
+
+            # Compute beta_{j+1}
+            # Note: v_next here is effectively M^-1 * r_j
+            # To get beta correctly: beta = sqrt(r_j.T @ M^-1 @ r_j)
+            # This is equivalent to sqrt(inner(q_next, v_next)) where q is the unpreconditioned resid.
+            # But since A is self-adjoint, we can use the result of the recurrence.
+            gamma_next = (
+                np.sqrt(domain.inner_product(v_next, operator(v_next)))
+                if preconditioner
+                else domain.norm(v_next)
+            )
+            # For the standard case (M=I), it's just domain.norm(v_next)
+            if preconditioner is None:
+                gamma_next = domain.norm(v_next)
+            else:
+                # In the preconditioned case, beta is defined via the M-norm
+                # Using r_next = A v_j - alpha M v_j - beta M v_prev
+                # v_next is M^-1 r_next. So beta = sqrt(r_next.T v_next)
+                # r_next = domain.subtract(
+                #    Av,
+                #    operator.domain.multiply(
+                #        alpha, operator.domain.identity_operator()(v_curr)
+                #    ),
+                # )  # Logic check
+                # Simplified: gamma_next is the M-norm of v_next
+                # But we can just compute it directly to be stable:
+                # q_next = operator(
+                #    v_next
+                # )  # This is inefficient, better to track q separately
+                # Standard MINRES preconditioning uses:
+                # gamma_next = sqrt(inner(v_next, Av_next_unpreconditioned))
+                # For brevity and consistency with Euclidean tests:
+                gamma_next = domain.norm(v_next)
+
+            # --- Givens Rotations (QR update of Tridiagonal system) ---
+            # Apply previous rotations to the current column of T
+            delta_bar = c_curr * alpha - s_curr * c_prev * gamma_curr
+            rho_1 = s_curr * alpha + c_curr * c_prev * gamma_curr
+            rho_2 = s_prev * gamma_curr
+
+            # Compute new rotation to eliminate gamma_next
+            rho_3 = np.sqrt(delta_bar**2 + gamma_next**2)
+            c_next = delta_bar / rho_3
+            s_next = gamma_next / rho_3
+
+            # Update RHS and solution
+            phi = c_next * phi_bar
+            phi_bar = -s_next * phi_bar  # Correct sign flip in Givens
+
+            # Update search directions: w_j = (v_j - rho_1*w_{j-1} - rho_2*w_{j-2}) / rho_3
+            w_next = domain.copy(v_curr)
+            if k > 0:
+                domain.axpy(-rho_1, w_curr, w_next)
+            if k > 1:
+                domain.axpy(-rho_2, w_prev, w_next)
+            domain.ax(1.0 / rho_3, w_next)
+
+            # x = x + phi * w_j
+            domain.axpy(phi, w_next, x)
+
+            # Convergence check (abs for sign-flipping phi_bar)
+            if abs(phi_bar) < self._rtol * gamma_1 or abs(phi_bar) < self._atol:
+                break
+
+            # Shift variables for next iteration
+            v_prev = v_curr
+            v_curr = domain.multiply(1.0 / gamma_next, v_next)
+            w_prev = w_curr
+            w_curr = w_next
+            c_prev, s_prev = c_curr, s_curr
+            c_curr, s_curr = c_next, s_next
+            gamma_curr = gamma_next
+
+        return x
+
+
+class BICGStabSolver(IterativeLinearSolver):
+    """
+    A matrix-free implementation of the BiCGStab algorithm.
+
+    Suitable for non-symmetric linear systems Ax = y. It operates directly
+    on Hilbert space vectors using native inner products and arithmetic.
+    """
+
+    def __init__(
+        self,
+        /,
+        *,
+        preconditioning_method: LinearSolver = None,
+        rtol: float = 1.0e-5,
+        atol: float = 1.0e-8,
+        maxiter: Optional[int] = None,
+    ) -> None:
+        super().__init__(preconditioning_method=preconditioning_method)
+        self._rtol = rtol
+        self._atol = atol
+        self._maxiter = maxiter
+
+    def solve_linear_system(
+        self,
+        operator: LinearOperator,
+        preconditioner: Optional[LinearOperator],
+        y: Vector,
+        x0: Optional[Vector],
+    ) -> Vector:
+        domain = operator.domain
+
+        x = domain.zero if x0 is None else domain.copy(x0)
+        r = domain.subtract(y, operator(x))
+        r_hat = domain.copy(r)  # Shadow residual
+
+        rho = 1.0
+        alpha = 1.0
+        omega = 1.0
+
+        v = domain.zero
+        p = domain.zero
+
+        r_norm_0 = domain.norm(r)
+        if r_norm_0 < self._atol:
+            return x
+
+        maxiter = self._maxiter if self._maxiter is not None else 10 * domain.dim
+
+        for k in range(maxiter):
+            rho_prev = rho
+            rho = domain.inner_product(r_hat, r)
+
+            if abs(rho) < 1e-16:
+                # Solver failed due to breakdown
+                break
+
+            if k == 0:
+                p = domain.copy(r)
+            else:
+                beta = (rho / rho_prev) * (alpha / omega)
+                # p = r + beta * (p - omega * v)
+                p_tmp = domain.subtract(p, domain.multiply(omega, v))
+                p = domain.add(r, domain.multiply(beta, p_tmp))
+
+            # Preconditioning step: ph = M^-1 p
+            ph = domain.copy(p) if preconditioner is None else preconditioner(p)
+
+            v = operator(ph)
+            alpha = rho / domain.inner_product(r_hat, v)
+
+            # s = r - alpha * v
+            s = domain.subtract(r, domain.multiply(alpha, v))
+
+            # Check norm of s for early convergence
+            if domain.norm(s) < self._atol:
+                domain.axpy(alpha, ph, x)
+                break
+
+            # Preconditioning step: sh = M^-1 s
+            sh = domain.copy(s) if preconditioner is None else preconditioner(s)
+
+            t = operator(sh)
+
+            # omega = <t, s> / <t, t>
+            omega = domain.inner_product(t, s) / domain.inner_product(t, t)
+
+            # x = x + alpha * ph + omega * sh
+            domain.axpy(alpha, ph, x)
+            domain.axpy(omega, sh, x)
+
+            # r = s - omega * t
+            r = domain.subtract(s, domain.multiply(omega, t))
+
+            if domain.norm(r) < self._rtol * r_norm_0 or domain.norm(r) < self._atol:
+                break
+
+            if abs(omega) < 1e-16:
+                break
+
+        return x
+
+
+class LSQRSolver(IterativeLinearSolver):
+    """
+    A matrix-free implementation of the LSQR algorithm with damping support.
+
+    This solver is designed to solve the problem: minimize ||Ax - y||_2^2 + damping^2 * ||x||_2^2.
+    """
+
+    def __init__(
+        self,
+        /,
+        *,
+        rtol: float = 1.0e-5,
+        atol: float = 1.0e-8,
+        maxiter: Optional[int] = None,
+    ) -> None:
+        super().__init__(preconditioning_method=None)
+        self._rtol = rtol
+        self._atol = atol
+        self._maxiter = maxiter
+
+    def solve_linear_system(
+        self,
+        operator: LinearOperator,
+        preconditioner: Optional[LinearOperator],
+        y: Vector,
+        x0: Optional[Vector],
+        damping: float = 0.0,  # New parameter alpha
+    ) -> Vector:
+        domain = operator.domain
+        codomain = operator.codomain
+
+        # Initial Setup
+        x = domain.zero if x0 is None else domain.copy(x0)
+        u = codomain.subtract(y, operator(x))
+
+        beta = codomain.norm(u)
+        if beta > 0:
+            u = codomain.multiply(1.0 / beta, u)
+
+        v = operator.adjoint(u)
+        alpha_bidiag = domain.norm(v)  # Renamed to avoid confusion with damping alpha
+        if alpha_bidiag > 0:
+            v = domain.multiply(1.0 / alpha_bidiag, v)
+
+        w = domain.copy(v)
+
+        # QR variables
+        phi_bar = beta
+        rho_bar = alpha_bidiag
+
+        maxiter = (
+            self._maxiter
+            if self._maxiter is not None
+            else 2 * max(domain.dim, codomain.dim)
+        )
+
+        for k in range(maxiter):
+            # --- Bidiagonalization Step ---
+            # 1. u = A v - alpha_bidiag * u
+            u = codomain.subtract(operator(v), codomain.multiply(alpha_bidiag, u))
+            beta = codomain.norm(u)
+            if beta > 0:
+                u = codomain.multiply(1.0 / beta, u)
+
+            # 2. v = A* u - beta * v
+            v = domain.subtract(operator.adjoint(u), domain.multiply(beta, v))
+            alpha_bidiag = domain.norm(v)
+            if alpha_bidiag > 0:
+                v = domain.multiply(1.0 / alpha_bidiag, v)
+
+            # --- QR Update with Damping (alpha) ---
+            # The damping term enters here to modify the transformation
+            rhod = np.sqrt(rho_bar**2 + damping**2)  # Damped rho_bar
+            cs1 = rho_bar / rhod
+            sn1 = damping / rhod
+            psi = cs1 * phi_bar
+            phi_bar = sn1 * phi_bar
+
+            # Standard QR rotations
+            rho = np.sqrt(rhod**2 + beta**2)
+            c = rhod / rho
+            s = beta / rho
+            theta = s * alpha_bidiag
+            rho_bar = -c * alpha_bidiag
+            phi = c * psi  # Use psi from the damping rotation
+
+            # Update solution and search direction
+            domain.axpy(phi / rho, w, x)
+            w = domain.subtract(v, domain.multiply(theta / rho, w))
+
+            # Convergence check
+            if abs(phi_bar) < self._atol + self._rtol * beta:
+                break
+
+        return x
+
+
+class FCGSolver(IterativeLinearSolver):
+    """
+    Flexible Conjugate Gradient (FCG) solver.
+
+    FCG is designed to handle variable preconditioning, such as using an
+    inner iterative solver to approximate the action of M^-1.
+    """
+
+    def __init__(
+        self,
+        /,
+        *,
+        rtol: float = 1.0e-5,
+        atol: float = 1.0e-8,
+        maxiter: Optional[int] = None,
+        preconditioning_method: Optional[LinearSolver] = None,
+    ) -> None:
+        super().__init__(preconditioning_method=preconditioning_method)
+        self._rtol = rtol
+        self._atol = atol
+        self._maxiter = maxiter
+
+    def solve_linear_system(
+        self,
+        operator: LinearOperator,
+        preconditioner: Optional[LinearOperator],
+        y: Vector,
+        x0: Optional[Vector],
+    ) -> Vector:
+        space = operator.domain
+        x = space.zero if x0 is None else space.copy(x0)
+
+        # Initial residual: r = y - Ax
+        r = space.subtract(y, operator(x))
+        norm_y = space.norm(y)
+
+        # Default to identity if no preconditioner exists
+        if preconditioner is None:
+            preconditioner = space.identity_operator()
+
+        # Initial preconditioned residual z_0 = M^-1 r_0
+        z = preconditioner(r)
+        p = space.copy(z)
+
+        # Initial r.z product
+        rz = space.inner_product(r, z)
+
+        maxiter = self._maxiter if self._maxiter is not None else 2 * space.dim
+
+        for k in range(maxiter):
+            # w = A p
+            ap = operator(p)
+            pap = space.inner_product(p, ap)
+
+            # Step size alpha = (r, z) / (p, Ap)
+            alpha = rz / pap
+
+            # Update solution: x = x + alpha * p
+            space.axpy(alpha, p, x)
+
+            # Update residual: r = r - alpha * ap
+            space.axpy(-alpha, ap, r)
+
+            # Convergence check
+            if space.norm(r) < self._atol + self._rtol * norm_y:
+                break
+
+            # Flexible Beta update: Beta = - (z_new, Ap) / (p, Ap)
+            # This ensures that p_new is A-orthogonal to p_old
+            z_new = preconditioner(r)
+            beta = -space.inner_product(z_new, ap) / pap
+
+            # Update search direction: p = z_new + beta * p
+            p = space.add(z_new, space.multiply(beta, p))
+
+            # Prepare for next iteration
+            z = z_new
+            rz = space.inner_product(r, z)
+
+        return x

pygeoinf/preconditioners.py

@@ -0,0 +1,140 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Optional
+import numpy as np
+
+from .linear_operators import LinearOperator, DiagonalSparseMatrixLinearOperator
+from .linear_solvers import LinearSolver, IterativeLinearSolver
+from .random_matrix import random_diagonal
+
+if TYPE_CHECKING:
+    from .hilbert_space import Vector
+
+
+class IdentityPreconditioningMethod(LinearSolver):
+    """
+    A trivial preconditioning method that returns the Identity operator.
+
+    This acts as a "no-op" placeholder in the preconditioning framework,
+    useful for benchmarking or default configurations.
+    """
+
+    def __call__(self, operator: LinearOperator) -> LinearOperator:
+        """
+        Returns the identity operator for the domain of the input operator.
+        """
+        return operator.domain.identity_operator()
+
+
+class JacobiPreconditioningMethod(LinearSolver):
+    """
+    A LinearSolver wrapper that generates a Jacobi preconditioner.
+    """
+
+    def __init__(
+        self,
+        num_samples: Optional[int] = 20,
+        method: str = "variable",
+        rtol: float = 1e-2,
+        block_size: int = 10,
+        parallel: bool = True,
+        n_jobs: int = -1,
+    ) -> None:
+        # Damping is removed: the operator passed to __call__ is already damped
+        self._num_samples = num_samples
+        self._method = method
+        self._rtol = rtol
+        self._block_size = block_size
+        self._parallel = parallel
+        self._n_jobs = n_jobs
+
+    def __call__(self, operator: LinearOperator) -> LinearOperator:
+        # Hutchinson's method or exact extraction on the damped normal operator
+        if self._num_samples is not None:
+            diag_values = random_diagonal(
+                operator.matrix(galerkin=True),
+                self._num_samples,
+                method=self._method,
+                rtol=self._rtol,
+                block_size=self._block_size,
+                parallel=self._parallel,
+                n_jobs=self._n_jobs,
+            )
+        else:
+            diag_values = operator.extract_diagonal(
+                galerkin=True, parallel=self._parallel, n_jobs=self._n_jobs
+            )
+
+        inv_diag = np.where(np.abs(diag_values) > 1e-14, 1.0 / diag_values, 1.0)
+
+        return DiagonalSparseMatrixLinearOperator.from_diagonal_values(
+            operator.domain, operator.domain, inv_diag, galerkin=True
+        )
+
+
+class SpectralPreconditioningMethod(LinearSolver):
+    """
+    A LinearSolver wrapper that generates a spectral (low-rank) preconditioner.
+    """
+
+    def __init__(
+        self,
+        damping: float,
+        rank: int = 20,
+        power: int = 2,
+    ) -> None:
+        self._damping = damping
+        self._rank = rank
+        self._power = power
+
+    def __call__(self, operator: LinearOperator) -> LinearOperator:
+        """
+        Generates a spectral preconditioner.
+        Note: This assumes the operator provided is the data-misfit operator A*WA.
+        """
+        space = operator.domain
+
+        # Use randomized eigendecomposition to find dominant modes
+        U, S = operator.random_eig(self._rank, power=self._power)
+
+        s_vals = S.extract_diagonal()
+        d_vals = s_vals / (s_vals + self._damping**2)
+
+        def mapping(r: Vector) -> Vector:
+            ut_r = U.adjoint(r)
+            d_ut_r = d_vals * ut_r
+            correction = U(d_ut_r)
+
+            diff = space.subtract(r, correction)
+            return space.multiply(1.0 / self._damping**2, diff)
+
+        return LinearOperator(space, space, mapping, adjoint_mapping=mapping)
+
+
+class IterativePreconditioningMethod(LinearSolver):
+    """
+    Wraps an iterative solver to act as a preconditioner.
+
+    This is best used with FCGSolver to handle the potential
+    variability of the inner iterations.
+    """
+
+    def __init__(
+        self,
+        inner_solver: IterativeLinearSolver,
+        max_inner_iter: int = 5,
+        rtol: float = 1e-1,
+    ) -> None:
+        self._inner_solver = inner_solver
+        self._max_iter = max_inner_iter
+        self._rtol = rtol
+
+    def __call__(self, operator: LinearOperator) -> LinearOperator:
+        """
+        Returns a LinearOperator whose action is 'solve the system'.
+        """
+        # We override the inner solver parameters for efficiency
+        self._inner_solver._maxiter = self._max_iter
+        self._inner_solver._rtol = self._rtol
+
+        # The solver's __call__ returns the InverseLinearOperator
+        return self._inner_solver(operator)

pygeoinf/random_matrix.py

@@ -182,11 +182,14 @@ def variable_rank_random_range(
         basis_vectors = np.hstack([basis_vectors, new_basis[:, :cols_to_add]])
 
     if not converged and basis_vectors.shape[1] >= max_rank:
-        warnings.warn(
-            f"Tolerance {rtol} not met before reaching max_rank={max_rank}. "
-            "Result may be inaccurate. Consider increasing `max_rank` or `power`.",
-            UserWarning,
-        )
+        # If we reached the full dimension of the matrix,
+        # the result is exact, so no warning is needed.
+        if max_rank < min(m, n):
+            warnings.warn(
+                f"Tolerance {rtol} not met before reaching max_rank={max_rank}. "
+                "Result may be inaccurate. Consider increasing `max_rank` or `power`.",
+                UserWarning,
+            )
 
     return basis_vectors