Struct WorkingModelPirlsOptions 

pub struct WorkingModelPirlsOptions {
    pub max_iterations: usize,
    pub convergence_tolerance: f64,
    pub adaptive_kkt_tolerance: Option<AdaptiveKktTolerance>,
    pub max_step_halving: usize,
    pub min_step_size: f64,
    pub firth_bias_reduction: bool,
    pub coefficient_lower_bounds: Option<Array1<f64>>,
    pub linear_constraints: Option<LinearInequalityConstraints>,
    pub initial_lm_lambda: Option<f64>,
    pub geodesic_acceleration: bool,
    pub arrow_schur: Option<ArrowSchurInnerConfig>,
}

Fields

max_iterations: usize

convergence_tolerance: f64

adaptive_kkt_tolerance: Option<AdaptiveKktTolerance>

max_step_halving: usize

min_step_size: f64

firth_bias_reduction: bool

coefficient_lower_bounds: Option<Array1<f64>>

Optional lower bounds on coefficients (same coordinate system as beta). Use -inf for unconstrained entries.

linear_constraints: Option<LinearInequalityConstraints>

Optional linear inequality constraints in current coefficient coordinates: A * beta >= b.

initial_lm_lambda: Option<f64>

Optional warm-start hint for the Levenberg-Marquardt damping coefficient. When set, the inner solver seeds λ_LM to this value instead of the default 1e-6. Clamped on consumption to [1e-6, 1e-3] so a stale or pathological hint cannot poison the solve: the upper bound costs at most three damping halvings versus the cold default, which is dwarfed by the savings when the hint is informative.

Used by execute_pirls_if_needed (in solver::reml::outer_eval) to persist the converged λ across consecutive PIRLS calls in a single REML outer optimization, so the inner Newton does not have to rediscover problem-specific damping at every accepted outer iterate.

geodesic_acceleration: bool

Enable the Transtrum-Sethna geodesic-acceleration second-order correction on each accepted Levenberg-Marquardt step. When true, after the standard LM direction δp = −(H + λ_lm·diag(H))⁻¹ g is computed and accepted by the LM gain test, the solver computes a finite-difference estimate of the directional second derivative of the gradient along δp, solves a second linear system with the same (already-factored) Hessian, and adds the correction δp₂ to the step only if ‖δp₂‖ ≤ α‖δp‖ (the Transtrum-Sethna 2011 acceptance criterion, α = 0.75 here). The correction costs two extra full WorkingModel::update calls per accepted step (for the FD evaluations); it is most useful for fits whose penalized Hessian is near-singular (latent-coordinate fits, near-collinear bases). Default false; opt-in until validated across the broader family of likelihoods and penalties.

arrow_schur: Option<ArrowSchurInnerConfig>

Optional arrow-Schur structured-inner-solve descriptor.

When Some, every accepted LM Newton step inside the inner loop is computed by the per-observation arrow-Schur path (crate::arrow_schur::ArrowSchurSystem) instead of the β-only solve_newton_direction_dense. When None, the existing β-only path is used unchanged (back-compat: every existing call site that does not opt in is unaffected).

Scope note. This wires the inner Gauss–Newton step. The REML outer-loop gradient w.r.t. t (which carries a shared Schur⁻¹ factor) is a separate plumbing change owned by the REML driver and is not handled here.

Trait Implementations

impl Clone for WorkingModelPirlsOptions

fn clone(&self) -> WorkingModelPirlsOptions

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for WorkingModelPirlsOptions

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.