Crate spectrafit_levenberg_marquardt 

spectrafit-levenberg-marquardt — the Levenberg–Marquardt family of solvers.

This crate is one method on the graph-agnostic spectrafit-trust-region framework: it owns the classic LM outer loop (Nielsen λ/ν damping, gain-ratio acceptance, Moré column scaling) and the regime-adaptive damped Gauss–Newton step (normal-equations + Cholesky, or thin-SVD secular), and layers two opt-in variants on the same loop:

• TRF — Coleman–Li bound scaling folded into the damping (StrategyConfig::bound_scaling).
• geodesic — Transtrum/Sethna second-order acceleration (StrategyConfig::geodesic).

The shared problem contract (TrustRegionProblem) and outcome types (Report, Termination) live in the framework crate and are re-exported here, so a consumer needs only a spectrafit-levenberg-marquardt dependency for the full LM API surface.

Structs

Report

Outcome of a solve. The optimised parameters live in the problem (read via TrustRegionProblem::params); this only carries diagnostics.

StepOutput

A computed trial step and the cost reduction the linear model predicts for it.

StrategyConfig

Tuning for a single Levenberg–Marquardt solve (covers LM, TRF and geodesic).

Enums

StepError

Why a step could not be computed for the current λ.

StepFactor

A once-per-outer-iteration factorization of the (column-scaled) Jacobian, reused across every λ trial and by geodesic acceleration. This is the key to keeping the inner λ search cheap: the O(m·p²) work (forming JᵀJ, or the thin SVD of J) happens once; each λ trial is then only O(p³) (Cholesky) or O(p²) (closed form).

StepKind

Which linear-algebra path the step uses. Chosen per-fit by select_regime.

Termination

Why the solve stopped.

Traits

TrustRegionProblem

A weighted nonlinear least-squares problem driven by the trust-region core.

Functions

factorize

Factor the step operator once for the current Jacobian and column scaling. diag (D) is the per-iteration damping scale; for the SVD path it is baked into the factorization (J̃ = J/D), for the NE path it is applied per λ.

minimize

Minimise ½‖r(p)‖² over the free parameters of problem with Levenberg–Marquardt.

select_regime

Choose the step factorization path from the problem shape.