Trait BlockEffectiveJacobian

Source

pub trait BlockEffectiveJacobian: Send + Sync {
    // Required method
    fn effective_jacobian_rows(
        &self,
        state: &FamilyLinearizationState<'_>,
        rows: Range<usize>,
    ) -> Result<Array2<f64>, String>;

    // Provided methods
    fn effective_jacobian_at(
        &self,
        state: &FamilyLinearizationState<'_>,
    ) -> Result<Array2<f64>, String> { ... }
    fn n_outputs(&self) -> usize { ... }
    fn eta_row_scaling_for_skewness(&self) -> Option<Arc<[f64]>> { ... }
}

Expand description

β-dependent Jacobian callback for a parameter block.

Principled long-term contract for expressing how a block contributes to the stacked linear predictor at a given β:

J(β) ∈ ℝ^{n_rows · n_outputs × p_block}

Single-output linear block: returns design.clone().
Row-scaled block (RowScaledJacobian): returns diag(eta_scaling) · design (still linear in β).
Multi-output block (e.g. survival marginal-slope with η0, η1, ad1): stacks ∂eta_r/∂β_k for r ∈ 0..n_outputs, row-major ordering.

The default impl on ParameterBlockSpec::effective_jacobian_at is:

jacobian_callback = None → design.clone().
jacobian_callback = Some(cb) → delegates to cb.effective_jacobian_at.

Required Methods§

Source

fn effective_jacobian_rows( &self, state: &FamilyLinearizationState<'_>, rows: Range<usize>, ) -> Result<Array2<f64>, String>

Stacked multi-output Jacobian for a contiguous observation row range.

Shape: (rows.len() * n_outputs, p_block), with the same channel-major layout as Self::effective_jacobian_at: row channel * rows.len() + local_row is rows.start + local_row in that output channel. Implementations should keep this as the single source of row math so large construction-time audits can stream chunks instead of materialising all n * p * K entries at once.

Provided Methods§

Source

fn effective_jacobian_at( &self, state: &FamilyLinearizationState<'_>, ) -> Result<Array2<f64>, String>

Stacked multi-output Jacobian at the current β.

Shape: (n_rows * n_outputs, p_block), channel-major: rows r * n_rows .. (r + 1) * n_rows carry output channel r’s row Jacobian, so stacked[r * n_rows + i, j] is observation i’s row at output r and coefficient column j. Every consumer that destacks this matrix (audit, canonicaliser, fit) relies on this layout — see BlockJacobianAsRowOp::from_callback for the destacking transpose. For n_outputs = 1 this is identical to the (n_rows, p_block) effective design used by the flat identifiability audit.

Source

fn n_outputs(&self) -> usize

Number of stacked output channels. 1 for most blocks.

Source

fn eta_row_scaling_for_skewness(&self) -> Option<Arc<[f64]>>

Returns the per-row scaling vector when this callback is a simple diagonal-scaling block (RowScaledJacobian). Used by the identifiability audit’s skewness-aware bias correction (T25).

Returns None for all blocks except RowScaledJacobian.