Struct CollapsedStat 

pub struct CollapsedStat {
    pub observed_sum_ds: DMatrix<f32>,
    pub imputed_sum_ds: DMatrix<f32>,
    pub size_s: DVector<f32>,
    pub observed_sum_db: DMatrix<f32>,
    pub n_bs: DMatrix<f32>,
    pub matched_bs: DMatrix<f32>,
    pub anchor_batches: Vec<usize>,
    pub size_ds: Option<DMatrix<f32>>,
    pub obs_mask_db: Option<DMatrix<f32>>,
    pub exclude_unmatched_from_delta: bool,
}

a struct to hold the sufficient statistics for the model

Fields

observed_sum_ds: DMatrix<f32>

imputed_sum_ds: DMatrix<f32>

size_s: DVector<f32>

observed_sum_db: DMatrix<f32>

n_bs: DMatrix<f32>

matched_bs: DMatrix<f32>

Counterfactual SOURCE mass per (batch, sample): how much of sample s’s imputed profile was drawn from batch b’s cells (each matched cell’s weights sum to 1, so a column sums to the matched cell count). With n_bs this is what identifies δ: E[imputed_gs] = μ_gs Σ_b δ_gb w_bs against E[observed_gs] = μ_gs Σ_b δ_gb n_bs.

anchor_batches: Vec<usize>

Batches whose frame IS the frame: δ is normalised so their per-gene geometric mean is 1. Empty = pooled, where every batch’s mass weighs in.

size_ds: Option<DMatrix<f32>>

Per-(gene, sample) effective size: the mass of sample s whose SOURCE actually measures gene g. None means fully observed — every gene sees size_s[s], and optimize takes the exact historical code path.

This is what distinguishes “unmeasured” from “measured as zero”: a gene absent from one cohort’s panel reads 0 in all of its cells, and without this denominator the estimate is silently dragged toward zero by cells that never looked. Set by [attach_observability]; see the plan note there for the factorized construction.

obs_mask_db: Option<DMatrix<f32>>

Per-(gene, batch) observability of δ’s evidence: 1.0 when any source in batch b measures gene g, else 0.0. Zeroing δ’s denominator there sends δ to its prior (≈ 1, “no adjustment”) instead of to an extreme driven by structurally-absent counts. None = all observed.

exclude_unmatched_from_delta: bool

When true (CNV strata present), samples with no counterfactual mass (matched_bs column sum 0) are excluded from the δ numerator and denominator so private-clone observed mass cannot pull δ. update_mu still uses the full masses. Default false — bit-identical to before.

Implementations

impl CollapsedStat

pub fn new(ngene: usize, nsample: usize, nbatch: usize) -> Self

pub fn num_genes(&self) -> usize

pub fn num_samples(&self) -> usize

pub fn num_batches(&self) -> usize

pub fn frame_weights(&self) -> DVector<f32>

Per-batch weights that define the frame δ is pinned to: the anchors when there are any (each weighted 1), else every batch by its own cell mass.

pub fn clear(&mut self)

pub fn select_columns(&self, indices: &[usize]) -> Self

Select a subset of sample columns (groups) by index.

pub fn select_rows(&self, r0: usize, nrows: usize) -> Self

Select a contiguous block of feature rows (r0..r0+nrows). Per-gene stats (observed/imputed/observed_db) are sliced; the per-sample size_s and per-batch n_bs/matched_bs are shared, so they’re copied whole. Used by the gene-blocked optimize.

Trait Implementations

impl Clone for CollapsedStat

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CollapsedStat

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

Formats the value using the given formatter.