Struct WorkingState 

pub struct WorkingState {
    pub eta: LinearPredictor,
    pub gradient: Array1<f64>,
    pub hessian: SymmetricMatrix,
    pub log_likelihood: f64,
    pub deviance: f64,
    pub penalty_term: f64,
    pub firth: FirthDiagnostics,
    pub ridge_used: f64,
    pub hessian_curvature: HessianCurvatureKind,
    pub gradient_natural_scale: f64,
}

Working state at a PIRLS iterate: gradient, Hessian, deviance, etc.

Fields

eta: LinearPredictor

gradient: Array1<f64>

hessian: SymmetricMatrix

log_likelihood: f64

deviance: f64

penalty_term: f64

firth: FirthDiagnostics

ridge_used: f64

hessian_curvature: HessianCurvatureKind

gradient_natural_scale: f64

Implementations

impl WorkingState

pub fn jeffreys_logdet(&self) -> Option<f64>

pub fn relative_gradient_norm(&self, g_norm: f64) -> f64

Scale-invariant relative gradient residual.

Returns ||g||_2 / (1 + ||score||_2 + ||Sbeta||_2 + ridge||beta||_2). g_norm is the projected/constrained stationarity residual in the current PIRLS basis; the denominator is the natural magnitude of the penalized gradient and is invariant under uniform rescaling of the objective.

pub fn certifies_kkt(&self, g_norm: f64, tol: f64) -> bool

Strict KKT acceptance: g_norm certifies stationarity under EITHER scale-invariant criterion (dimension-based or data-driven natural-scale).

Both certificates are invariant under uniform rescaling of the objective F → c·F (in the limit where the natural scale dominates the additive 1 floor). Acceptance under either is sufficient because:

• the natural-scale bound is tighter when the data are well-scaled (it tracks actual gradient component magnitudes);
• the dimension bound is tighter when the design matrix has unusual scaling (so the natural scale is dominated by a single component).

pub fn near_stationary_kkt(&self, g_norm: f64, tol: f64) -> bool

Near-stationary band (10× the strict KKT tolerance) under EITHER scale-invariant criterion. Used as a “good-enough” plateau check that classifies a fit as StalledAtValidMinimum rather than as a hard non-convergence. The band is 10 · tol without a floor — a caller asking for tol = 1e-12 gets a 1e-11 band, not the 1e-5 the old tol.max(1e-6) * 10 formula silently widened it to. The 1e-6 floor was masking real convergence regressions (e.g. constant_prior_mean_centers_penalty’s LM-ridge induced 2.5e-8 bias visible only when the user asked for sub-1e-6 precision).

Trait Implementations

impl Clone for WorkingState

fn clone(&self) -> WorkingState

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for WorkingState

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

Formats the value using the given formatter.