Struct AtomFunctionalReport 

pub struct AtomFunctionalReport {
    pub peak_contrast: Option<AtomFunctionalEstimate>,
    pub average_value: Option<AtomFunctionalEstimate>,
    pub decoder_variation_norm: Option<AtomFunctionalEstimate>,
}

Descriptive penalty-debiased POINT summaries of one fitted atom’s decoder curve (#1097, narrowed under #1115). Each field is a scalar functional of the atom’s inner smooth g_k(t), reported as a plug-in value and a one-step penalty-debiased value (the regularization bias relative to the conditional target is removed through the atom fit’s penalized Hessian). No standard error and no confidence interval are reported — by design (see below).

Why these carry NO coverage claim (#1115)

Conditional on the fitted latent coordinates t̂ and assignment â, each functional is an ordinary linear functional of the penalized-WLS coefficients β with a well-defined conditional population value, and one-step debiasing validly removes the penalty bias for that conditional target. The point estimates are therefore meaningful. A standard error, however, would only be honest if t̂ and â were fixed/known. They are not: they are generated regressors estimated from the very activations that also form the response Z, so Z enters both the design (via t̂(Z), â(Z)) and the response. An influence-function SE built from the β-only Hessian and row scores carries no ∂t̂/∂Z / ∂â/∂Z channel — exactly the generated-regressor correction the marginal-slope family (#461 Stage 2) is defined by — so it omits a first-order variance term and is generally anti-conservative. Rather than ship an SE/CI that silently under-covers, this report exposes only the debiased point summaries; a coverage-valid interval would require either freezing the dictionary on a held-out split or propagating the generated-regressor Jacobian, neither of which the fixed inner-fit snapshot supports.

Fields

peak_contrast: Option<AtomFunctionalEstimate>

g(t_peak) − g(t_mode): the peak-vs-baseline contrast of the fitted decoder, penalty-debiased through the inner-fit Hessian. Point summary only (no coverage claim — see the type doc).

average_value: Option<AtomFunctionalEstimate>

E_data[g(t_i)]: the data-averaged decoder value over the atom’s active rows, penalty-debiased. Point summary only.

decoder_variation_norm: Option<AtomFunctionalEstimate>

E_data[∂g/∂t] along the atom’s leading latent axis: how much the fitted decoder curve varies across the data distribution, conditional on the fit. A descriptive variation measure of the fitted curve, NOT a population “marginal slope” (the latent coordinate is itself a fitted, generated regressor). Point summary only.

Despite the historical _norm suffix this is the signed mass-weighted mean derivative E_data[∂g/∂t] over the single leading axis, not a magnitude — it can be negative, and a value near 0 means the average slope cancels (a symmetric bump), not that the curve is flat. Use AtomSmoothSignificance::log_e_nonconstant for an honest non-constancy test; this field only describes the average local slope.

Trait Implementations

impl Clone for AtomFunctionalReport

fn clone(&self) -> AtomFunctionalReport

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AtomFunctionalReport

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

Formats the value using the given formatter.