Enum BlockWorkingSet 

pub enum BlockWorkingSet {
    Diagonal {
        working_response: Array1<f64>,
        working_weights: Array1<f64>,
    },
    ExactNewton {
        gradient: Array1<f64>,
        hessian: SymmetricMatrix,
    },
}

Working quantities supplied by a custom family for one block.

Observed vs expected information (see response.md Section 3)

For the outer REML/LAML criterion, the Hessian used in log|H| and trace terms must be the observed (actual) Hessian at the mode, not the expected Fisher.

• ExactNewton: provides -nabla^2 log L directly, which is the observed Hessian by construction. This is always correct.

• Diagonal: provides IRLS working weights W such that the per-block Hessian is X’WX. For canonical links (logit-Binomial, log-Poisson), W_obs = W_Fisher. For supported non-canonical diagonal links, W must be the observed weight W_obs = W_Fisher - (y-mu)*B so the outer REML uses the exact Laplace Hessian. The matching CustomFamily::diagonalworking_weights_directional_derivative callback must differentiate the same observed W surface; silently using Fisher weights or zero dW would change the criterion into a PQL-type surrogate.

Variants

Diagonal

Standard IRLS/GLM-style diagonal working set for eta-space updates.

Fields

working_response: Array1<f64>

IRLS pseudo-response for this block’s linear predictor.

working_weights: Array1<f64>

IRLS working weights for this block (non-negative, length n).

For the inner solver, Fisher or observed weights both find the same mode. For the outer REML/LAML log|H| term, observed weights are the correct Laplace choice (see response.md Section 3). Canonical-link families need no correction since observed = Fisher.

ExactNewton

Exact Newton block update in coefficient space.

gradient is nabla log L wrt block coefficients. hessian is -nabla^2 log L wrt block coefficients (positive semidefinite near optimum).

This is the observed Hessian by construction (actual second derivative of the log-likelihood), which is the correct quantity for the outer REML Laplace approximation.

Fields

gradient: Array1<f64>

hessian: SymmetricMatrix

Implementations

impl BlockWorkingSet

pub fn diagonal_checked( working_response: Array1<f64>, working_weights: Array1<f64>, ) -> Result<Self, String>

Construct a Diagonal working set with the length invariant (working_response.len() == working_weights.len()) enforced at the type boundary. Use this from any new code path that produces a diagonal IRLS block; the legacy struct-literal form is preserved for existing call sites pending a full migration.

Trait Implementations

impl Clone for BlockWorkingSet

fn clone(&self) -> BlockWorkingSet

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BlockWorkingSet

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

Formats the value using the given formatter.