Four expansion-v2 items (finished after a session-limit interruption): Weak/integral-form discovery (crates/lawsynth-weakform, specs/weak-form) — the top noise-robustness technique from PySINDy: integrate candidates against compactly- supported test functions, NO differentiation of noisy data. ~23x lower coefficient error than finite-difference strong-form under noise. 20 tests, deterministic. Implicit/rational dynamics (crates/lawsynth-implicit, specs/implicit-dynamics) — SINDy-PI alternating-LHS nullspace solve; recovers rational laws (Michaelis-Menten) explicit regression can't express. Honest degree-2 identifiability limit documented (multi-dimensional nullspace); residual-floor + derivative-preferring tie-break for canonical selection. 10 tests. Optimizer breadth (lawsynth-sparse) — FROLS (orthogonal-least-squares, ERR ordering), SSR (stepwise), trapping (stability-biased); --solver frols|ssr|trapping. 24 tests; default path byte-identical. SRBench harness (benchmarks/) — 18 Strogatz ODE systems + scoring + a determinism column (byte-identical on replay — a metric no stochastic competitor can report). 3/3 Strogatz recovered at R²=1.0. srbench_report.py + 5 tests. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Implicit / rational dynamics boundary (v2-A)
This directory specifies implicit and rational dynamics discovery — the SINDy-PI
style method implemented in crates/lawsynth-implicit, which recovers relations
f(x, ẋ) = 0 (including rational laws ẋ = P(x)/Q(x)) that explicit
ẋ = Θ(x)Ξ regression cannot express. It is a boundary specification.
Method
Build an augmented candidate library Θ(x, ẋ) — monomials in the states, plus
those monomials multiplied by the target derivative ẋ. A rational law
ẋ = P(x)/Q(x) is equivalent to the implicit relation Q(x)·ẋ − P(x) = 0, i.e.
a sparse vector ξ with Θ(x, ẋ) · ξ ≈ 0. Discovery finds that sparse nonzero ξ.
Requirements
- Nontrivial solution. ξ = 0 trivially satisfies
Θξ = 0and MUST be excluded. The reference uses the SINDy-PI alternating-LHS scheme: for each candidate column j, normalise its coefficient to 1 (move it to the LHS) and solve a sparse regression for the remainder; a left-hand side whose column is identically zero MUST be skipped. - Selection. Among candidate LHS choices, selection MUST be deterministic.
Near-perfect fits (relative residual below a fixed tolerance) MUST be treated
as tied so that machine-level residual noise does not decide the outcome;
ties break toward fewer active terms, then toward a derivative-bearing LHS
(so the chosen relation is affine in ẋ and reconstructs the canonical
Q(x)·ẋ = P(x)), then toward the earlier library index. - Explicit reconstruction. When the relation is affine in ẋ, the
implementation MUST reconstruct the explicit rational law
ẋ = P(x)/Q(x)and report whether Q(x) is non-vanishing over the observed window (a pole makes the explicit form invalid there). - Determinism. Identical inputs MUST yield bit-identical output.
Honest identifiability limit
Implicit discovery has a genuine identifiability limit: at library degree ≥ 2
the augmented nullspace is multi-dimensional. If r(x, ẋ) = 0 is the true
relation, then m(x)·r(x, ẋ) = 0 for any monomial m(x) also fits to machine
precision. Therefore an implementation is NOT guaranteed to return the minimal,
spurious-free relation at higher degree. The achievable, conforming guarantee is:
- the returned relation is a valid member of the nullspace (fits to tolerance) and, when affine in ẋ, reconstructs an explicit rational law that reproduces the true derivative across the observed window; and
- at the minimal degree (a one-dimensional nullspace) the clean, spurious-free law is recovered.
An implementation MUST NOT claim minimal-form recovery at higher degree; it MUST document this limit (as the reference's degree-2 test does).