Struct JointPoissonFitConfig 

pub struct JointPoissonFitConfig {
    pub max_iter: usize,
    pub lambda_init: f64,
    pub lambda_up: f64,
    pub lambda_down: f64,
    pub armijo_c: f64,
    pub backtrack: f64,
    pub tol_d: f64,
    pub tol_param: f64,
    pub fd_step: f64,
    pub enable_polish: bool,
    pub polish: NelderMeadConfig,
    pub compute_covariance: bool,
}

Configuration for joint_poisson_fit.

Fields

max_iter: usize

Maximum number of damped-Fisher iterations in stage 1.

lambda_init: f64

Initial damping factor (Marquardt λ) on the Fisher matrix diagonal.

lambda_up: f64

Multiplicative factor to increase λ on a rejected step.

lambda_down: f64

Multiplicative factor to decrease λ on an accepted step.

armijo_c: f64

Armijo sufficient-decrease coefficient.

backtrack: f64

Backtracking factor during line search.

tol_d: f64

Convergence tolerance on relative deviance change.

tol_param: f64

Convergence tolerance on normalized parameter step.

fd_step: f64

Finite-difference step for gradient fallback.

enable_polish: bool

Enable Nelder-Mead polish after stage 1.

Default false as of #486. The polish tolerances (xatol = 1e-9, fatol = 1e-10) were originally matched to the EG5 synthetic benchmark (memo 35 §P2.1) where D stays O(1), so fatol is physically meaningful. On real-data regimes where D saturates at 10⁴–10⁵ (un-modelled upstream physics — memo 35 §P3/§P4), fatol / D drops below f64 ULP and polish cannot self-terminate — it burns its full max_iter = 5000 every fit at 70–260× wall cost, and the three-scenario ablation on real VENUS Hf 120-min data (issue #486) showed the resulting parameter shift is ≤ 0.35 Fisher σ on every parameter in every scenario — i.e. below the solver’s own reported uncertainty floor.

The polish mechanism itself is sound (self-terminates cleanly on synthetic D≈1 data per ablation S3); only the absolute tolerance defaults are mis-calibrated for real counts data. A future scale-aware rescale (fatol_rel vs D_stage1) can re-enable polish as a useful opt-in refinement.

Set this to true (via with_counts_enable_polish(Some(true)) at the pipeline level) when you specifically want the polish stage on a synthetic / clean-data scenario where the absolute tolerance defaults are physically meaningful.

polish: NelderMeadConfig

Polish (Nelder-Mead) configuration. Used only when enable_polish == true. Default xatol = 1e-9, fatol = 1e-10 match the EG5 synthetic benchmark tolerances from memo 35 §P2.1 — physically meaningful when D ≈ 1 (clean data) but sub-f64- ULP on real counts where D ≈ 10⁴–10⁵, which is why enable_polish defaults to false. See #486.

compute_covariance: bool

Compute and return the Fisher covariance and parameter uncertainties.

Trait Implementations

impl Clone for JointPoissonFitConfig

fn clone(&self) -> JointPoissonFitConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for JointPoissonFitConfig

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

Formats the value using the given formatter.

impl Default for JointPoissonFitConfig

fn default() -> Self

Returns the “default value” for a type.