autolens.potential_correction.IterFitDpsiSrcInterferometer

class IterFitDpsiSrcInterferometer

Bases: object

Iteratively solves for the pixelized source s and potential corrections dpsi of an interferometer dataset, minimizing the penalized visibility objective with a Levenberg-Marquardt loop on the combined state x = [s | dpsi]. Each accepted step re-ray-traces the real-space grid through the corrected lens (the macro model plus an InputPotential built from the current dpsi), so the corrections feed back into the source mapping.

Requires the dataset’s sparse operator (dataset.apply_sparse_operator()) — the loop runs through the w-tilde normal-equation space and never materializes visibility-space matrices.

Parameters:

• dataset – The al.Interferometer dataset carrying a sparse operator.

• lens_start (Galaxy) – The macro lens galaxy the corrections perturb.

• dpsi_pixelization (DpsiPixelization) – The dpsi mesh + regularization model, on the real-space mask.

• src_pixelization (Pixelization) – The source pixelization.

• gauge_constraints (bool) – Whether to impose <dpsi, 1> = <dpsi, x> = <dpsi, y> = 0 via an equality-constrained KKT step.

• src_image_mesh – An image mesh whose image-plane mesh grid is preloaded into the source inversion.

• dpsi_mask – An optional 2D bool mask restricting the dpsi mesh to a sub-region of the real-space mask (typically an arc-tracing mask from al.pc.util.arc_mask_from): the corrections are only constrained where the lensed arcs are, and restricting the mesh there stabilises the inversion. Defaults to the full real-space mask.

• settings_inversion (Optional[Settings]) – The inversion settings; defaults to the border relocator without the positive-only solver (the LM state is signed).

• preloads (Optional[dict]) – Precomputed attributes set directly onto the fit.

• n_iter (int) – The maximum number of outer LM iterations.

• tol (float) – The step-norm convergence tolerance.

• reg_optimize_every (Optional[int]) – When set, every N accepted outer iterations (and once at the start) the regularization strength multipliers (a_src, a_dpsi) are re-optimized by maximising the Laplace evidence at the current normal equations — the per-iteration objective control of Koopmans 2005 / Vegetti & Koopmans 2009 that prevents the source block absorbing potential perturbations. Each candidate costs one Cholesky solve at fixed F, D.

• reg_optimize_grid (int) – The number of log-spaced multipliers per axis of the re-optimization grid (spanning 1e-2..1e2 around the current scales).

• verbose (bool) – Whether to log per-iteration costs.

Methods

log_evidence	The Laplace-approximation log evidence at (s, dpsi) (defaulting to the optimized state): 0.5 [ logdet R_s + logdet R_dpsi - logdet(F + R) - chi^2 - x^T R x ] (+ the complex-noise normalization), with chi^2 through the normal-equation identity.
solve_joint_optimization	Runs the Levenberg-Marquardt loop on the combined state x = [s | dpsi] in the real-space normal-equation space, accepting steps that decrease the penalized cost.

Attributes

data_weighted_norm	The scalar d^H C^-1 d over real and imaginary components.
dpsi_gradient_matrix
dpsi_points
dpsi_regularization_matrix
pair_dpsi_data_obj

property pair_dpsi_data_obj

property dpsi_gradient_matrix

property dpsi_points

property dpsi_regularization_matrix

property data_weighted_norm: float

The scalar d^H C^-1 d over real and imaginary components.

solve_joint_optimization(x0=None)

Runs the Levenberg-Marquardt loop on the combined state x = [s | dpsi] in the real-space normal-equation space, accepting steps that decrease the penalized cost. Returns the optimized (s, dpsi); also stored as s_opt / dpsi_opt.

log_evidence(s=None, dpsi=None, include_noise_normalization=True)

The Laplace-approximation log evidence at (s, dpsi) (defaulting to the optimized state): 0.5 [ logdet R_s + logdet R_dpsi - logdet(F + R) - chi^2 - x^T R x ] (+ the complex-noise normalization), with chi^2 through the normal-equation identity.