Work

When does the orientation of Gibbs-preserving circulation affect route selection? Parity, quasianalytic rigidity, and a Laplace pairing of separatrix germs

Year: 2026
Type: preprint
Abstract: For overdamped diffusions F = −∇V + εℓ with ℓ·∇V = 0 and ∇·ℓ = 0, the nonreversible-sampler class (exactly Gibbs stationary density), we ask when the orientation of the circulation can bias route-sensitive observables: branch ratios between barrier-degenerate escape routes, and winding currents. (i) Parity. In any dimension the characteristic polynomial of ∇F at each critical point is even in ε, so local Eyring–Kramers data are even under circulation reversal at leading prefactor order. For tree-like metastable reductions the orientation-odd branch component between distinct resolved edges vanishes unconditionally at leading Arrhenius order, and to all algebraic orders under a standard expansion hypothesis. (ii) Rigidity. Winding periods are a Laplace pairing at scale D between admissible cycles and separatrix matching germs. In quasianalytic classes the germs vanish: absolute periods vanish, and relative ones reduce to an explicit wall term, nonzero only at reflecting walls of unequal potential. Germ-borne selection is realized in every Gevrey class and graded by the germ's contact order m as Θ(ε Dᵐ⁺¹ exp(−cₑ/D)). (iii) Two mechanisms. In the models examined, the orientation-odd branch logit at linear response is, numerically, the germ-borne term plus a boundary–Green quotient obstruction of the occupation kernel. The latter survives analytic drives and quasistationary initialization, with a barrier-controlled exponent and an anisotropy-dependent prefactor. The Bouchet–Reygner transport law holds numerically at the percent level and a fitting-free closed form at the few-percent level; standard orientation-odd path functionals cannot serve as universal selectors. Rigidity extends to symmetric ring channels in three dimensions, with consequences for sampler activation functions. About this record. Preprint, version v1 (14 pages, two-column; 44 references). Not yet peer reviewed. The record holds two files: Kaneko_2026_Gibbs-preserving-circulation_v1.pdf: the paper. Kaneko_2026_reproduction_package_v1.zip: the reproduction package (88 files). Every number in the paper is produced by one of 80 numbered checks; REPRO.md maps each check to its script, the claim it supports and its runtime tier. Contents: Python scripts for the checks (symbolic parity and inertia checks, sparse Fokker–Planck / committor solves, quasistationary eigenproblems, Monte Carlo, 1D boundary-value reductions in 40-digit and 80-bit long-double arithmetic); verify_outputs.py, which compares the fast checks with the values quoted in the paper within stated tolerances; prefactor tables R(D) to D = 0.010 and the detuning scan (data/); reference logs of the long checks (logs/); the manuscript source; README.md, requirements.txt, CITATION.cff, LICENSE. Quick start. Unzip, change into the folder Kaneko_2026_reproduction_package_v1 and run “pip install -r requirements.txt”, then “python verify_outputs.py” (fast checks, ≈ 3 min, ≈ 1.1 GiB; exit 0 = all pass) and “python verify_outputs.py --long” (adds the long checks, ≈ 25 min, 2.1 GiB peak). Tested with Python 3.13.16, numpy 2.5.3, scipy 1.18.1, sympy 1.14.0, mpmath 1.3.0, matplotlib 3.11.2 on x86-64 Linux. Checks 78–79 require numpy's long double to be 80-bit extended precision (x86-64 Linux, Intel macOS); elsewhere they stop with an error rather than return inaccurate values. Licences. Paper, figures and data: CC BY 4.0. Code (.py and .sh files): MIT. Disclosure. This research was carried out in extended collaboration with Anthropic's Claude (large language model), which contributed derivations, code, and adversarial review under the author's direction. The author takes full responsibility for all statements, has independently re-derived the principal analytical claims, and has verified the numerical outputs through the archived test suite.
Language: English
Open Access status: green
OpenAlex ID: W7226008542
Cites:
Cited by:
Related to: 10