From Caustics to Groups — experiments

Code for the research program scoped in the blog post _posts/2026-07-15-caustics-to-groups-research-program.md: given an observed field of caustics (the singular set of a sub-Riemannian exponential map), infer the underlying Lie-group structure — the reverse map. Built around the ADE group-blindness obstruction and the three structure-specific observables that escape it (tangent-cone growth vector; conjugate-locus symmetry and moduli at the pole; abnormal-geodesic stratum), matched by the nilpotent-deviation statistic. Full derivation in docs/METHODS.md.

Layout

• src/ — dependency-light math core (numpy only), one module per group / concern:
  • heisenberg.py — Heisenberg H³ forward model: closed-form geodesics, exponential map, and the known conjugate locus (t_c = 2π/|w|, z-axis). The leakage-free Phase 0 ground truth.
  • caustics.py — generic first-conjugate-time / caustic detection for any geodesic map, via the zero of the exponential map’s Jacobian determinant. Validated against the Heisenberg closed form; reused for later groups.
  • liegroup.py — generic vectorized Lie–Poisson normal-geodesic engine; candidate-group specs (Heisenberg, SE(2), Engel, Cartan; SE(3) upcoming) as frame + structure constants + horizontal set.
  • growth.py — tangent-cone growth-vector estimator (metric M1) via geodesic reach scaling with a noise-floor-aware weight fit.
  • caustics.py — first-conjugate-time detection: a generic finite-difference routine and a fast batched one for 3D contact groups.
  • fingerprint.py — the three-component fingerprint and the group classifier (M1 growth vector + M2 nilpotent-deviation δ that splits the (2,3) alias).
  • (upcoming) a data-driven abnormal-stratum test (M4); SE(3) forward model.
• scripts/run_e0.py — E0 growth-vector recovery sweeps (noise, sample size).
• scripts/run_e1.py — E1 fingerprint classifier + confusion matrix.
• scripts/smoke_test.py — assert-based self-checks + golden-file regression.
• tests/golden/ — committed reference outputs (e.g. the Heisenberg conjugate locus); regenerate with smoke_test.py --update-golden.
• docs/METHODS.md — living derivation of the fingerprint + deviation statistic.
• docs/adr/ — architecture/method decision records (ADR-0001: candidate list and identifiability target).
• artifacts/ — result JSONs (gitignored, regenerable).

Status

• Phase 0 (scaffold + ground truth): done — Heisenberg model, generic engine, golden test green.
• E0 (growth-vector recovery, tests H1): done — H1 confirmed. All three vectors recovered exactly on clean data; graceful noise degradation; Engel (step 3) is the hardest (needs more samples, tolerates less noise); Heisenberg and SE(2) alias at (2,3) as pre-registered. See docs/E0-growth-vector.md.
• E1 (fingerprint + confusion matrix, tests H2): done — H2 confirmed. Perfect 4-way separation on clean data (1.00 vs 0.25 chance); δ breaks the Heisenberg/SE(2) alias (δ ≈ 0.002 vs 0.140); noise degrades step-3 groups first and into “unknown” (abstention), never wrong-group confusion. See docs/E1-confusion.md.
• E2 (aliasing map + abnormal leg, tests H3): done — H3 confirmed. Honest aliasing map with two rigidity points (Heisenberg/SE(2) rides on δ alone; Engel/Cartan on the growth vector alone). Standout finding: the coarse abnormal bit (M4) is far more noise-robust than the full growth vector — perfect for Cartan at σ=3e-2 where M1 recovery is 0.00. See docs/E2-aliasing.md.
• Classifier upgrade (M4 fallback): done — tagging the coarse class from the abnormal bit when the growth vector is unresolved lifts class accuracy to 1.00 at σ=1e-2 and 0.96 at σ=3e-2 (from 0.91 / 0.57 strict), with no new wrong-group errors. Degradation is now graceful (falls to correct coarse class, not “unknown”).
• Blog Parts 2 & 3 drafted from these results: Part 2 “The Forward Map” and Part 3 “The Inverse Map” (confusion matrix, aliasing map, calibrated silence).
• E4 + SE(3) (calibrated silence + H4 both halves): done — confirmed. Added the SE(3) fibre-tracking tangent cone (growth vector (3,6), M1-verified with the frame↔C guard). Confident on genuine homogeneous groups including SE(3) (the H4 yes-half mechanism, 97%); correctly abstains on an effective flow (growth vector varies across base points → “field of varying tangent cones”). See docs/E4-calibrated-silence.md.
• Blog Part 4 (“In the Wild”): the four-part series is complete — Part 4.
• Remaining (future): inference on a real (credential-gated) DW-MRI acquisition; optional theory appendices C1–C5; the curved SE(3) exact geodesics (Euler-angle frame) for a full conjugate-locus deviation on SE(3).

Results at a glance

Experiment	Hypothesis	Verdict
E0 growth-vector recovery	H1 identifiability	confirmed
E1 fingerprint + confusion matrix	H2 discrimination	confirmed
E2 aliasing map + abnormal leg	H3 rigidity	confirmed
E4 calibrated silence	H4 (no half)	confirmed (right verdict, wrong reason — see E5)
SE(3) synthetic field	H4 (yes half, mechanism)	confirmed
real DW-MRI acquisition	H4 (yes half, real data)	pending credentialed data
E5 gravity has no SR structure	H-A (astrophysics Phase A)	confirmed; refuted our own Q>n criterion

Astrophysics phase

Plan: docs/PLAN-astrophysics-local-groups.md. E5 done (docs/E5-no-sr-structure-in-gravity.md): gravity is a deterministic flow, not a control system — no distribution, no growth vector. The R³×S² lift’s constraint is a tautology (residual 5e-16 for any smooth flow), so the growth vector measures the instrument. And a Zel’dovich fold fakes Heisenberg’s Q=4 exactly, so Q>n pointwise is not sufficient; the correct criterion is Q>n on a set of full measure (Zel’dovich 0%, Heisenberg 100%). Next: Phase B — the deformation-tensor isotropy stratification, validated against the published D₄-umbilic criterion (first external validation in this repo).

E6 done (docs/E6-local-symmetry-doroshkevich.md) — the programme’s first external validation. The local group of the cosmic web is the isotropy group of the deformation tensor, stratified by eigenvalue degeneracy. Measured against Doroshkevich (1970): the Hessian covariance matches theory (0.2025 vs 1/5; 0.0664 vs 1/15), the eigenvalue law matches to ~0.1% (KS D ≈ 0.002), and the Vandermonde-induced eigenvalue repulsion is confirmed (α = 0.973 vs 1.0). Hence the SO(2)-symmetry stratum has codimension 2 (curves, β = 1.944 vs 2.0), and D₄ umbilics — degeneracy ∩ fold, corank 2 — are isolated points. Local groups are detectable; they are found with the deformation tensor, never a growth vector.

E8 done (docs/E8-magnetized-plasma.md) — the technique’s astrophysical home. Suppressed cross-field transport is anisotropic Riemannian (Q=n even at D⊥/D∥ = 1e-6), and D⊥→0 gives an integrable rank-1 foliation. But on the extended (position, flux) space, [X₁,X₂] = B ∂_z, so the structure is contact wherever B≠0: Q=4>n=3 on full measure (30/30 base points), and for uniform B the frame is literally Heisenberg. At a magnetic null the weight jumps 2→3 (Martinet), Q: 4→5 — so the growth vector is a magnetic-null / reconnection-site detector.

The astrophysics phase is complete. Gravity: no SR structure (E5). Cosmic-web local groups: isotropy groups of the deformation tensor, externally validated (E6). SR technique’s home: magnetized flux geometry (E8).

Reproduce

cd research/caustics-to-groups
python3.13 -m venv .venv && .venv/bin/pip install numpy scipy
.venv/bin/python scripts/smoke_test.py     # model + detector + M1 + classifier + golden
.venv/bin/python scripts/run_e0.py         # E0 growth-vector recovery sweeps (~30s)
.venv/bin/python scripts/run_e1.py         # E1 fingerprint + confusion matrix (~20s)