KarlNaumann / MacroStat / 26446002830

91%

push

github

refactor(KirmansAnts): Lamperti-only SDE with torch.Generator (#74)

* refactor(KirmansAnts): drop rejection/reflect; Lamperti only

Strip the three-integrator dispatch table introduced under Phase-1
boundary-handling exploration. The Lamperti-transformed integrator
(Moran et al. 2020) is the only one that recovers O(dt) weak
convergence under the non-Lipschitz x-space diffusion; the rejection
and reflect integrators carried bias that this paper documents but
no longer needs to run.

- Delete KirmansAntsDiagnostics, _INTEGRATORS dispatch, _rng_aux,
  _compute_drift_diffusion, _milstein_substep, _advance_rejection,
  _advance_reflect, _unpack_rho_mu.
- Rename _advance_lamperti -> advance_lamperti (public) so autodocs
  harvests the Equations block; private names are filtered out by
  parse_behavior_docstrings.
- Add the forward Lamperti map phi = arcsin(2x - 1) to the Equations
  align block for a self-contained derivation.
- Drop boundary_method and max_attempts hyperparameters. Fix the
  wrong Lamperti formula in the parameters docstring (was
  sqrt(2/mu) arcsin(sqrt(x)); should be arcsin(2x - 1)).
- Module docstring now leads with the method, not the choice.

* test(KirmansAnts): simplify for Lamperti-only; add KS gates

- Drop ALL_METHODS/NEW_METHODS parametrisation, dispatch tests,
  Milstein and drift-diffusion identity tests, diagnostics tests,
  rejection xfail.
- Demote the xfail-marked bimodal KS test to a plain passing test
  under Lamperti; add explicit unimodal and uniform analogues.
- Add @pytest.mark.slow test_stationary_paper_budget -- three
  scenarios at T=1000, KS-1samp against Beta(a, a) CDF. Gives a
  measurable gate for the fig_ant_dynamics figure at paper budget
  rather than a visual handwave.

* refactor(KirmansAnts): use torch.Generator for noise; inline Lamperti drift

Replace self.numpy_rng.standard_normal with a per-instance torch.Generator
seeded from hyper["seed"] in initialize(). Each macro-step draws its
substeps Gaussi... (continued)

465 of 475 branches covered (97.89%)

Branch coverage included in aggregate %.

3154 of 3301 relevant lines covered (95.55%)

0.96 hits per line