Many physical and financial time series are modeled as a hidden process that follows a stochastic differential equation with a strong restoring force. The double-well and Ginzburg–Landau potentials are standard examples, and in both the pull back toward the center grows faster than linearly with distance. A particle filter tracks the hidden state by simulating many candidate trajectories and reweighting them against the data. Recursive maximum likelihood fits the model's parameters online by following the gradient of the filter with respect to those parameters. Both must simulate the dynamics with a numerical step. A plain Euler step overshoots in the tails and sends particles to infinity. "Taming", which shrinks the drift where it is large, fixes the overshoot. This note argues that the usual taming also quietly removes the property that makes long-horizon filtering provably stable: the model forgets where it started, at a rate that does not depend on how far out it started.
The core observation fits in two sentences. A superlinear restoring force is exactly what lets a diffusion on an unbounded space forget its initial condition uniformly, because from any starting point, however distant, it returns to a bounded region within a fixed amount of time. Standard tamed steps cap how far a particle can move per step, so a far-away particle takes arbitrarily long to come back. The discretized model therefore loses the uniform forgetting that time-uniform error proofs require. We propose tamed steps that are "infinity-compressing": every state is mapped into a bounded region in one step. We then show that this is the missing ingredient for time-uniform particle estimates of filter gradients on unbounded state spaces, a setting where existing guarantees require a compact state space.
Core Claim
Time-uniform guarantees for particle filters and their parameter derivatives rest on a strong mixing assumption. Uniform Stability of a Particle Approximation of the Optimal Filter Derivative proves that an $O(N^2)$ backward-kernel estimator of the filter derivative has error $C/\sqrt{N}$ uniformly in time. The proof assumes a compact state space and transition densities bounded away from zero and infinity (their Assumption A). The authors state that Gaussian dynamics on $\mathbb{R}^d$ fall outside this assumption. Additive smoothing error in backward variational inference for general state-space models and Variational excess risk bound for general state space models make the same assumption. These papers motivate the gap. The taming literature supplies the numerical method. Tamed Langevin sampling under weaker conditions rescales the drift by $1+\gamma|\nabla U|$, and Nonasymptotic convergence analysis for the tamed unadjusted stochastic Langevin algorithm uses $1+\sqrt{\lambda}|\theta|^r$. Both prove uniform-in-time moment bounds and convergence to a fixed target, but neither considers filtering.
A uniform moment bound is weaker than uniform forgetting. With a tamed step $T_h$, the per-step displacement $|T_h(x)-x|$ is bounded by a constant $K_h$. A particle started at radius $r$ therefore needs at least $(r-R)/K_h$ steps to reach a ball of radius $R$. The Dobrushin contraction coefficient of any fixed block of steps is then 1, taken over all starting points. The coefficient measures how much two different starting distributions still differ after the block, and a value of 1 means forgetting fails uniformly. We call a tamed map infinity-compressing if
$$\sup_{x \in \mathbb{R}^d} |T_h(x)| \le R_h < \infty ,$$
where $T_h$ is the deterministic part of one numerical step with step size $h$ and $R_h$ is a radius. Our central conjecture, to be proven and tested, is as follows. For latent dynamics with superlinear dissipativity, a Gaussian-noise scheme built on an infinity-compressing $T_h$ satisfies a minorization over any block of physical time $\Delta \ge \Delta_0$. A minorization means the block transition density is bounded below by a fixed fraction $\varepsilon$ of a common reference measure. The constant $\varepsilon$ should depend on $d$ and $\Delta$ but not on $h$ or on the starting state. Given this condition, the stability argument of Del Moral et al. carries over to the block kernel, and the backward-kernel filter-derivative estimator becomes uniform in time on $\mathbb{R}^d$. Standard bounded-increment taming does not satisfy this condition.
The error propagation follows the familiar "leaky bucket" recursion $E_{k+1} \le (1-\varepsilon)E_k + \tau_h$. Here $E_k$ is the filter error after block $k$ and $\tau_h$ is the local error that taming injects per block. The recursion caps the long-run error at $\tau_h/\varepsilon$. The whole point is that $\varepsilon$ stays bounded away from zero.
Research Program
The first strand is construction. For potentials that split as $U(x) = u(|x|) + V(x)$, with $u$ a power law carrying the superlinear growth and $V$ of lower order, one infinity-compressing step applies the exact radial flow of $-\nabla u$ over time $h$ and then an ordinary tamed Euler step for $\nabla V$. The radial flow is available in closed form for power laws. For $u(r) = c r^{2m+2}$, its image over time $h$ is contained in a ball of radius of order $(ch)^{-1/(2m)}$, whatever the starting point. This is the discrete analogue of a diffusion "coming down from infinity." For non-separable potentials we will study a smooth saturation map composed with a standard tamed step. Unlike the projection of Projected Langevin Monte Carlo algorithms in non-convex and super-linear setting, this map has no boundary discontinuity and its transition density is differentiable in the parameters. That differentiability matters because recursive maximum likelihood needs the gradient of the log transition density.
Each fine substep has a Gaussian transition density centered at $T_h(x)$. The backward kernel of Del Moral et al. can therefore be run on the fine grid, with observations every $M$ substeps, at cost $O(N^2 M)$ per observation.
The second strand is analysis. It has three parts. The positive result is the block minorization with $h$-independent $\varepsilon$ described above, combined with an $O(h)$ local weak error. Together these give time-uniform bias and $L_p$ error for the filter and its derivative. The matching negative result is a lower bound for bounded-increment taming. The filter started from a point mass at $x_0$ should forget $x_0$ only after a time that grows linearly in $|x_0|$. Consequently, derivative error after an outlier or a far initialization should scale with $|x_0|$. The third part addresses the tension flagged by Diffusion of Credit in Markovian Models: fast forgetting stabilizes the filter but erases long-lag gradient signal. Compression intentionally speeds forgetting in the tails. It could therefore bias the estimate of exactly the parameter that governs the tails, the stiffness $c$. We will quantify this identifiability cost rather than assume it away. We will also report whether a decaying compression radius removes it.
Baselines
The naive reference point is an untamed Euler discretization inside a bootstrap particle filter, with the $O(N^2)$ backward-kernel derivative. The resampling step is left to discard runaway particles. This measures how often the problem actually bites, and it is the configuration most practitioners would try first.
The strongest competitor is the same filter and derivative estimator with a standard bounded-increment tamed step. We will use the TUSLA form from Nonasymptotic convergence analysis for the tamed unadjusted stochastic Langevin algorithm and the wd-TULA form from Tamed Langevin sampling under weaker conditions. These are the best-analyzed stabilizers for superlinear drift, and they already come with uniform moment bounds. Their authors report that in a 100-dimensional double-well, TUSLA reaches first-moment errors as low as $3.5 \times 10^{-5}$, and that vanilla ULA "failed across all tested step sizes" while wd-TULA stayed stable. These are static-sampling numbers. On filtering and recursive maximum likelihood with tamed dynamics, no reported result was found on Lacuna, so this proposal must run that comparison itself. We will also run our compressing scheme as a pure sampler on the same double-well. The check is that compression does not cost stationary accuracy relative to the TUSLA figure.
The third baseline is projection onto a ball, following PLMC. It is the other scheme that bounds the image, so it is the closest mechanism to ours. The fourth is implicit (backward) Euler.
Two ablations isolate the proposal's ingredients. The first swaps only the tamed map, compressing versus bounded-increment, while holding the local error $\tau_h$ matched at the stationary distribution. The second runs the compressing scheme with the $O(N)$ path-space derivative instead of the backward kernel. This separates the contribution of taming from that of the estimator. In the stochastic-volatility benchmark of Del Moral et al., the path-space variance grew linearly while the backward kernel's "remained flat."
For fairness, all methods share the same particle count, the same number of fine substeps per observation, and matched wall-clock budgets. Implicit Euler's inner nonlinear solve counts against its budget. Step sizes and taming exponents are tuned on a held-out horizon with an equal grid budget for each method. We count a method as beaten if its derivative-variance or parameter-RMSE curves grow with time, or scale with the initialization radius, while ours stay flat, at equal or lower cost.
Why not just…?
Why not just let resampling kill runaway particles? When observations are frequent and informative, exploded particles receive negligible weight and vanish. This is a real threat in that regime, and we expect untamed filters to look acceptable there. The failure should appear with sparse, missing, or low signal-to-noise observations. It should also appear in the derivative estimator, where a single non-finite particle contaminates the backward sums. More fundamentally, the untamed discretized chain is itself transient, so the filter is tracking a model that does not mix. The observation-density sweep below is designed to locate exactly where this baseline breaks.
Why not just use standard taming, which already has uniform moment bounds? Bounded moments under the stationary regime do not bound the time needed to forget a far-away state. The lower bound in our analysis predicts a recovery time linear in the initial radius for bounded-increment taming. The outlier and far-initialization experiments test this directly. If standard taming recovers in time independent of $|x_0|$ in practice, our negative result is wrong and the claim shrinks to a proof technique.
Why not just project onto a ball, or use implicit Euler? Projection trivially makes the state space compact, so Assumption A holds. However, the minorization constant then degrades with the ball radius, and the parameter gradient is discontinuous at the boundary. Implicit Euler compresses $|x|$ to roughly $(|x|/h)^{1/3}$ in one step, which is strong but not uniform. It needs a nonlinear solve, and its transition density needs a Jacobian term. Implicit Euler is the most serious alternative. If it matches our scheme empirically at equal cost, the contribution becomes a theory result that explains why it works, and we will report it that way.
Proposed Experiments
We use three benchmarks. The first is the stochastic-volatility model of Del Moral et al., run for 20,000 to 100,000 steps with $N=500$, as a sanity check that nothing regresses on a benign model. The second is a latent double-well state-space model observed through Gaussian noise, in dimensions 1 to 100. The third is a stochastic Ginzburg–Landau field of up to 1,000 sites observed at a subset of sites, reusing the potentials of Huang et al. as latent dynamics rather than as static targets.
The metrics are four. We track the variance of the filter-derivative estimate over time. We track the RMSE of the recursive-maximum-likelihood parameters, including the stiffness $c$. We measure filter bias against a reference filter with large $N$ and small $h$. Finally, we estimate the empirical forgetting rate as the distance between two filters started at the origin and at radius $r$, for $r$ up to $10^3$.
The decisive checks are three. First, is recovery time after a far initialization, or after an injected outlier observation, flat in $r$ for compressing taming and growing in $r$ for bounded-increment taming? Second, does derivative variance stay flat over 100,000 steps as the observation interval grows, while baselines drift? Third, does compression bias the estimate of $c$? The theory predicts the first two and leaves the third open.
References
- Uniform Stability of a Particle Approximation of the Optimal Filter Derivative
- Additive smoothing error in backward variational inference for general state-space models
- Variational excess risk bound for general state space models
- Tamed Langevin sampling under weaker conditions
- Nonasymptotic convergence analysis for the tamed unadjusted stochastic Langevin algorithm
- Projected Langevin Monte Carlo algorithms in non-convex and super-linear setting
- Diffusion of Credit in Markovian Models