Struct AdaptiveTraceEvidence 

pub struct AdaptiveTraceEvidence {
    pub logdet_hessian: f64,
    pub traces: Array1<f64>,
    pub stderrs: Array1<f64>,
    pub probe_count: usize,
    pub converged: bool,
}

Adaptive-K Hutchinson trace schedule with common random numbers (CRN).

Repeatedly invokes evidence_derivatives_hutchinson_gpu with probe counts K = 16, 32, 64, 128, stopping at the first K that satisfies the per-coordinate relative-SE criterion

max_j  s_j / (sqrt(K) · max(|t_j|, τ))  ≤  ε

where s_j is the sample standard deviation across the K probes (the raw quadratic-form sample, without the (1/2) REML logdet scaling) and t_j is the running mean. Because the SplitMix probe RNG is stateless ((seed, k_index, i) → ±1), the first K_prev probes of a K = 2·K_prev re-run are bit-identical to the previous batch, so each step extends the prior estimate rather than starting fresh in expectation. The implementation re-runs from scratch at each K for simplicity; CRN is preserved by the stateless RNG seed.

Returns the raw traces t_j = tr(H⁻¹ H_j) = mean_k q_{j,k} (length D), the log|H| from the cached Cholesky, and the final probe count K actually used. The raw traces (not the (1/2) REML logdet gradient) are what the outer evaluator wants — it applies the logdet-gradient half-factor itself.

Fields

logdet_hessian: f64

traces: Array1<f64>

stderrs: Array1<f64>

Per-probe sample standard deviation of the raw trace estimator q_{j,·} (NOT divided by sqrt(K)); divide by sqrt(probe_count) to obtain the standard error of the running mean traces[j].

probe_count: usize

converged: bool