Does the Solver Move the Equilibrium? Time-Uniform Stationary Bias for Hypoelliptic Neuron Models

Many noisy physical and biological systems inject randomness into only some of their variables. A classic case is a neuron model where noise drives a slow "recovery" variable while membrane voltage evolves deterministically under a cubic restoring force. Scientists fit such models to data by simulating them for a long time and matching long-run statistics: the histogram of voltage values and the frequency content of spikes. That procedure quietly assumes the simulator's long-run behavior matches the true model's long-run behavior. Nobody has proved this for systems that combine partial noise with forces growing faster than linearly. It matters in practice. Two recent inference pipelines fit the same stochastic FitzHugh–Nagumo neuron model to the same rat recordings, and they disagree on the parameters. The authors could not tell whether the gap comes from the simulator, from parameters that the data cannot distinguish, or from a wrong model.

This note proposes to answer "does the solver move the equilibrium?" for this class of models. There are two parts. The first is a proof that the splitting integrator both pipelines use has a long-run bias that does not grow with simulation length and scales as $h^2$ in the step size $h$. The second is an experiment that converts that bias into an equivalent shift in fitted parameters and compares it against the observed disagreement between the two pipelines.

Core Claim

The setting has three ingredients, each introduced in plain terms here. A hypoelliptic diffusion is a stochastic differential equation (SDE) whose noise reaches only some coordinates but still spreads to all of them through the drift. Its density evolves by a degenerate parabolic equation, meaning a diffusion equation that is missing diffusion in some directions. A superlinear drift is a force that grows faster than linearly, such as $-V^3$. It makes naive explicit simulation explode.

Consider the stochastic FitzHugh–Nagumo system, where $V$ is voltage, $W$ is recovery, $B$ is a Brownian motion, and $\epsilon,\gamma,\beta,\sigma$ are parameters:

$$dV_t=\tfrac{1}{\epsilon}(V_t-V_t^3-W_t)\,dt,\qquad dW_t=(\gamma V_t-W_t+\beta)\,dt+\sigma\,dB_t .$$

More generally, the claim covers semi-linear hypoelliptic SDEs whose nonlinearity is a dissipative flow that can be solved exactly. For this class we claim that the Strang splitting scheme has a unique invariant measure $\mu_h$. The scheme alternates an exactly solved linear Gaussian step with an exactly solved cubic ODE step. Its invariant measure $\mu_h$ is the long-run distribution of the simulated chain, and it satisfies

$$\Big|\int\phi\,d\mu_h-\int\phi\,d\mu\Big|\le C_\phi\,h^2,$$

with $C_\phi$ independent of the simulation horizon. Here $\mu$ is the true stationary distribution and $\phi$ is any smooth test function of polynomial growth. The exponent 2 comes from the symmetric composition cancelling the first-order term in the step-size expansion of the long-run bias (a Talay–Tubaro expansion). For the Lie–Trotter variant, which does one sub-step and then the other without symmetrizing, the same argument gives order 1. We then use this bound, together with measured constants, to test whether solver bias can explain the posterior disagreement between the two published analyses of rat dorsal rootlet data.

Why Existing Work Stops Short

The research file shows that each piece exists, but the three pieces do not appear together. Three results supply partial methods but break exactly at this intersection.

Convergence of Numerical Time-Averaging and Stationary Measures via Poisson Equations turns finite-time weak error orders into long-run stationary bias. It does this by solving the Poisson equation $L\psi=\phi-\bar\phi$, where $L$ is the SDE's generator and $\bar\phi$ is the stationary average. It explicitly allows hypoelliptic generators, but it works on compact spaces or with globally Lipschitz coefficients. Cubic drift on $\mathbb{R}^2$ violates both. We reuse its structure and replace compactness with Lyapunov-weighted bounds on the derivatives of $\psi$.

Linear implicit approximations of invariant measures of semi-linear SDEs with non-globally Lipschitz coefficients proves a time-independent $O(h)$ bias under cubic nonlinearities. Its proof needs a negative-definite linear part. The FitzHugh–Nagumo linear part is not negative definite: it couples $V$ and $W$ through off-diagonal terms, and noise sits only in $W$.

An explicit splitting SAV scheme for the kinetic Langevin dynamics handles hypoelliptic superlinear systems with an explicit energy-stable splitting. It introduces an auxiliary energy variable, and its weak error bound grows polynomially in time rather than staying uniform.

Two further lines constrain the design. Contractive kinetic Langevin samplers beyond global Lipschitz continuity shows that ordinary taming, meaning capping the drift's magnitude per step, destroys the monotonicity needed for contraction. Its fix, a monotonicity-preserving taming, relies on a strongly convex potential. FitzHugh–Nagumo is not a gradient system at all and has a limit cycle, so the fix does not transfer. The tamed Euler analysis in Geometric Ergodicity and Strong Error Estimates for Tamed Schemes of Super-linear SODEs and the overdamped results in The tamed unadjusted Langevin algorithm both assume noise in every coordinate.

On the regularity side, Loss of regularity for Kolmogorov equations is a warning. When partial noise fails Hörmander's bracket condition, solutions can lose Hölder continuity and Euler schemes converge slower than any power of the step count. We address this head on. FitzHugh–Nagumo satisfies Hörmander's condition, because the $-W/\epsilon$ term carries noise into $V$. Our proof therefore uses hypoelliptic smoothing, and we check numerically that the predicted power law actually appears.

The application is motivated by a concrete disagreement. Inference for the stochastic FitzHugh-Nagumo model from real action potential data via approximate Bayesian computation fits the model with approximate Bayesian computation. It uses summary statistics built from the invariant density and the spectral density, computed by a structure-preserving splitting. Inference for Diffusion Processes via Controlled Sequential Monte Carlo and Splitting Schemes fits the same model on 12,500 recordings with Strang splitting, diffusion bridges and controlled sequential Monte Carlo. It reports estimates that "differed from previous studies using Approximate Bayesian Computation (ABC)". Neither paper bounds $\mu_h$ against $\mu$.

Research Program

The theoretical work has three steps.

The first step builds a Lyapunov function $\mathcal{V}(v,w)$ that is uniform in $h$ for the discrete Strang chain. It uses the fact that the exact cubic flow contracts $|V|$ outside a compact set whatever the step length. This is the property that gives splitting its reported 0% explosion rate. The step yields geometric ergodicity of the chain and moment bounds uniform in $h$.

The second step is the technical crux, and it may fail. We must show that the Poisson solution $\psi$, or equivalently the semigroup $P_t\phi$, has derivatives bounded by a polynomial in $\mathcal{V}$ that decays exponentially in $t$. The difficulty is that the noise is degenerate and the state space is non-compact. We plan to obtain this from hypoelliptic interior estimates on unit time intervals, glued together by the exponential convergence from the first step. We will state a partial result, a bias constant growing like $\log T$, if full uniformity resists.

The third step is the local error expansion of the Strang step against the generator, which gives order 2. We also derive the corresponding order-1 statement for Lie–Trotter.

The computational work measures $C_\phi$ and the bias exponent. It then pushes the bias through the actual inference summaries: histograms of the voltage values and spectral densities at the step sizes the two published pipelines used.

Baselines

The naive reference point is explicit Euler–Maruyama at the same step sizes. Inference for Diffusion Processes via Controlled Sequential Monte Carlo and Splitting Schemes reports that it explodes in up to 40% of simulated intervals under high noise, against 0% for splitting. We will report its explosion rate and, among runs that do not explode, its stationary bias. It anchors the scale but is not the real competitor.

The strongest existing competitor for long-run accuracy on hypoelliptic superlinear systems is the explicit splitting SAV scheme of An explicit splitting SAV scheme for the kinetic Langevin dynamics. It has order-one strong convergence and a weak error that grows polynomially in time. We adapt it to FitzHugh–Nagumo by placing the auxiliary energy variable on the cubic voltage potential. We give it the same number of drift evaluations and a grid search over its tuning parameters $\alpha$ and $C_H$ with the same budget we spend on any tuning of our own. The file gives no reported stationary-bias number for it: no reported result found on Lacuna for invariant-measure bias of SSAV or of splitting schemes on FitzHugh–Nagumo. Beating it means two things: a bias slope near 2 against its expected slope near 1, and a bias that is flat in $T$ where its bias drifts upward.

We also run the dual-tamed Euler scheme of Geometric Ergodicity and Strong Error Estimates for Tamed Schemes of Super-linear SODEs. On its own elliptic double-well benchmark it reports strong slopes of exactly 0.5 and 1.0. Here we test whether it stays ergodic when the noise is degenerate.

The key ablation is Lie–Trotter against Strang splitting, with identical sub-flows. If the $h^2$ rate is real, only the symmetric composition shows it. This isolates the claimed mechanism, which is cancellation of first-order bias, from the separate benefit of exact nonlinear sub-steps.

The ground-truth reference for $\int\phi\,d\mu$ is Strang splitting at very small step size with Richardson extrapolation. As an independent check that does not involve any solver, we use moment bounds from the semidefinite programs of Bounding Stationary Averages of Polynomial Diffusions via Semidefinite Programming. These apply because FitzHugh–Nagumo is polynomial. They are guaranteed valid but not guaranteed tight for non-elliptic systems, so we use them only where the upper and lower bounds nearly meet.

Why not just…?

Why not just run the simulator with a tiny step and call that the truth? For a two-dimensional model this genuinely works as a numerical reference, and we adopt it as our primary ground truth rather than argue against it. What it cannot do is tell a practitioner, before running ABC with hundreds of thousands of simulations, which step size keeps the bias below the posterior's width. It also cannot explain why the bias stays flat as simulations get longer. The proof and the measured constant $C_\phi$ supply exactly that. The fine-step reference is also only trustworthy if the power law holds, and the roughening results of Loss of regularity for Kolmogorov equations show that under partial noise the power law can fail. We cross-check it against the semidefinite bounds for that reason.

Why not just use an implicit scheme, which is stable by construction? Implicit Euler also removes explosion. It pays for this with a nonlinear solve at every step across the 200,000 to 300,000 simulations the ABC pipeline reported needing. It is also only first order, so it does not address the bias question any better. We include it as a secondary accuracy reference at equal wall-clock time.

Why not just use monotonicity-preserving taming, which already has contraction guarantees? The guarantees in Contractive kinetic Langevin samplers beyond global Lipschitz continuity require a strongly convex potential. FitzHugh–Nagumo has an oscillating limit cycle, and a double-well potential is non-convex. We run the tamed OBABO sampler from that paper as a stress test on a kinetic double-well target. If it matches Strang splitting's bias there anyway, our argument for splitting over taming weakens to the solvability of the sub-flows alone. We will say so.

Why not just attribute the ABC–cSMC disagreement to parameters the data cannot distinguish? This is the most likely answer, and our experiment is designed to confirm it or rule it out rather than assume it. If the solver-induced shift in parameters turns out much smaller than the observed disagreement, the solver is exonerated. That outcome would direct attention to identifiability and misspecification. It is still a useful result for these papers.

Proposed Experiments

The first check tests the rate. It uses the one-dimensional cubic SDE $dX=(X-X^3)dt+\sigma dW$ as a sanity case where $\mu$ is known in closed form. It then uses FitzHugh–Nagumo at the parameters of both published studies, in excitable and oscillatory regimes. For polynomial and indicator-smoothed test functions we plot the stationary bias against $h$ on log–log axes. We expect a slope near 2 for Strang and near 1 for Lie–Trotter and SSAV. For comparison, the file reports a slope of 1.0 for the projected Euler scheme of Pang, Wang and Wu on its own elliptic benchmarks.

The second check tests time uniformity. It plots the bias of time averages against horizon $T$ from $10^2$ to $10^6$ at fixed $h$. A flat curve for Strang, against an upward drift for SSAV, supports the uniform bound.

The third check is the decisive application test. It uses the two summaries from the ABC pipeline: the voltage invariant density and the spectral density. For each, we compute the shift caused by step size, $\Delta s(h)=s_h-s_0$, at the step sizes used by the two published pipelines. We convert this shift to a parameter shift $\Delta\theta\approx J^{-1}\Delta s$, where $J$ is the sensitivity of the summaries to the four parameters, estimated by finite differences at the reference solver. We then compare $\Delta\theta$ with the reported gap between the ABC and controlled sequential Monte Carlo posteriors on the rat dorsal rootlet recordings, and with the posterior widths. The file gives this gap only qualitatively, so we will recompute it by rerunning both published procedures with a shared solver at the published step sizes and at $h/4$. Suppose the gap shrinks substantially as $h$ decreases while $\Delta\theta$ matches the shrinkage. Then solver bias is a real contributor. If the gap persists unchanged, the solver is ruled out as a cause.

A final check guards against failure of the smoothing assumption. On the non-Hörmander counterexample of Hairer, Hutzenthaler and Jentzen, the proof's assumptions fail. There we expect, and will report, the absence of any clean power law. This confirms that the observed rates come from hypoelliptic smoothing rather than from the scheme alone.