← What's New

Tier 1.1 — Multivariate (marked) time-rescaling GOF

2026-05-28 · nstat.fit (core) · branch feat/core-multivariate-time-rescaling

Why

The per-neuron univariate KS test (FitResult.computeKSStats) checks each neuron's marginal conditional intensity in isolation. It can pass for every neuron while the model still gets the coupling between neurons wrong — e.g. treating synchronous cells as independent. Detecting that requires a population-level test.

What shipped

• population_time_rescale(counts_list, lam_per_bin_list, *, n_tau_bins=1) → PopulationTimeRescaleResult — the Tao, Weber, Arai & Eden (2018) marked point-process time-rescaling framework. Pure NumPy/SciPy, core nstat/fit.py.

Two complementary statistics:

• Ground-process KS — pool all spikes into one process with intensity λ• = Σk λk, rescale by Λ•, KS-test the rescaled inter-event intervals vs. the unit exponential. Detects aggregate-temporal and dependency misfit (synchrony / coupling).
• Marked-region Pearson χ² — each neuron-k spike maps to (τ, k) with τ = ∫λk; under a correct model these points uniformly fill R = {(τ, m): 0 ≤ τ ≤ b(m)}. Detects relative-allocation misfit (and, with n_tau_bins > 1, within-neuron timing).

from nstat import population_time_rescale
r = population_time_rescale(counts_per_neuron, lambda_per_bin_per_neuron, n_tau_bins=5)
r.ground_ks_pvalue   # population temporal GOF
r.mark_chi2_pvalue   # relative-allocation / coupling GOF

Validation

Scenario	Result
Correctly-specified population	neither statistic rejects
Wrong overall rate	ground KS rejects (χ² invariant — by design)
Synchronous pair modeled as independent	per-neuron univariate KS both pass; population ground KS rejects (p≈3e-49)
Wrong relative allocation (right total)	mark χ² rejects; ground KS passes
Single neuron	ground process = univariate rescaling

Notes

• The discrete-time (Haslinger, Pipa & Brown 2010) correction already existed in nSTAT (computeKSStats(..., dt_correction=1)); this iteration adds the multivariate (2018) piece.
• Python-only extension — the 2018 method postdates the 2012 MATLAB toolbox, so there is no matlab_gold counterpart (no parity fixture). The existing univariate computeKSStats and its gold fixtures are untouched.

Files changed

• nstat/fit.py — population_time_rescale + PopulationTimeRescaleResult (additive).
• tests/test_population_time_rescale.py — 6 tests (correct/misspecified/coupling/proportion/consistency/validation).
• nstat/__init__.py, tests/test_api_surface.py, AGENT_GUIDE.md, docs/ClassDefinitions.md — exports + helpfiles.