Function projected 

pub fn projected(
    learner: impl StreamingLearner + 'static,
    d_in: usize,
    rank: usize,
    lambda: f64,
) -> ProjectedLearner

Wrap any streaming learner with online projection learning (PAST algorithm).

Applies Welford normalization, then projects d_in-dimensional inputs to rank dimensions via the PAST subspace tracker (Yang, 1995 — “Projection approximation subspace tracking”). The forgetting factor lambda controls the PAST half-life: lambda = 0.9999 ≈ 6 931-sample half-life.

When the inner model exposes RLS readout weights (HasReadout), the projection updates toward prediction-relevant directions (supervised mode); otherwise it tracks variance-maximising directions (PCA mode).

Panics

Panics if rank > d_in (delegated to SubspaceTracker::new).

use irithyll::{projected, rls, StreamingLearner};

let mut model = projected(rls(0.99), 10, 4, 0.9999);
for i in 0..20 {
    model.train(&[i as f64; 10], i as f64);
}
let pred = model.predict(&[1.0; 10]);
assert!(pred.is_finite());