When we use numerical methods to simulate a differential equation, our standard guarantees are built around a finite time horizon $[0, T]$. Classical numerical analysis ensures that as the step size $h$ approaches zero, the numerical trajectory stays close to the true trajectory up to time $T$.
But what happens when $T \to \infty$?
In many scientific fields—from molecular dynamics to financial engineering—we do not care about a single finite trajectory. Instead, we want to know the system's long-term statistical behavior: its steady states, its energy profiles, or its invariant measure (the stationary probability distribution it settles into after running for an indefinite period).
If you take a standard numerical scheme and run it forever, finite-time error bounds break down. Tiny local errors accumulate, often causing the simulation to diverge catastrophically from the true long-term physics. This article explains why standard methods fail over infinite horizons and how modern numerical frameworks achieve long-time stability and convergence.
The Finite-Horizon Illusion
To understand why long-time simulation is uniquely challenging, consider a linear stochastic oscillator (a system modeling noisy, repetitive motion like a pendulum in a windstorm).
If you simulate this oscillator using the standard Euler-Maruyama method—the stochastic equivalent of the classic forward Euler scheme—it behaves reasonably well over short intervals. However, as demonstrated by Aslaug H.Strømmen Melbø in [Numerical simulation of a linear stochastic oscillator with additive noise], the long-term energy of the true continuous system grows linearly over time, but the Euler-Maruyama simulation exhibits exponential energy growth. The short-term accuracy of the method creates an illusion of stability that vanishes over infinite horizons. Similar discrepancies between short-term convergence and long-term qualitative features appear across deterministic iterative solvers, a relationship formalized by Andrew M. Stuart et al. [Waveform relaxation as a dynamical system].
The root of the problem lies in the error bounds. Standard proofs rely on Grönwall's inequality, which yields error bounds that scale as $O(e^{CT}h^\gamma)$, where $h$ is the step size and $\gamma$ is the convergence order. If $T$ grows large, the error bound explodes exponentially, rendering the guarantee useless for infinite-horizon simulations.
Capturing the Invariant Measure
For a well-behaved, dissipative stochastic differential equation (SDE), trajectories are pulled toward a global equilibrium. Even if two trajectories start far apart, the randomness and damping cause them to forget their initial states and settle into a shared statistical distribution: the invariant measure ($\mu$).
Continuous SDE: Initial State ───(Long Time T → ∞)───> True Invariant Measure (μ)
▲
│ (Does this converge?)
▼
Numerical Scheme: Initial State ───(Long Time T → ∞)───> Numerical Invariant Measure (μ_h)
As Raz Kupferman et al. [LONG-TERM BEHAVIOUR OF LARGE MECHANICAL SYSTEMS WITH RANDOM INITIAL DATA] show, complex high-dimensional systems (like large mechanical networks) can often be simplified into Langevin SDEs that exhibit this kind of ergodic behavior.
To evaluate a system's long-term behavior numerically, we need our discretization scheme to possess its own numerical invariant measure ($\mu_h$) that closely approximates the true invariant measure ($\mu$). We look for two types of convergence:
- Strong Long-Time Convergence: The numerical trajectory remains close to the true trajectory in an average sense across the entire infinite time horizon.
- Weak Convergence of the Invariant Measure: The statistical averages (expectations of target functions) calculated over the numerical steady state match those of the true steady state: $$\left| \int \phi(x) d\mu_h(x) - \int \phi(x) d\mu(x) \right| \le C h^p$$ Crucially, the constant $C$ here must be time-independent.
The Catastrophe of Non-Globally Lipschitz Coefficients
Designing schemes with time-independent bounds is difficult because real-world physical systems rarely feature perfectly linear or bounded forces. Instead, they are governed by potentials that grow superlinearly—such as a cubic restoring force in a molecular bond. Mathematically, these systems lack globally Lipschitz continuous coefficients.
When a drift coefficient grows superlinearly, an explicit Euler step becomes dangerous. If a large random fluctuation pushes the simulation state far from the origin, the explicit step calculates a massive restoring force based entirely on that extreme position. Instead of gently pulling the particle back, this oversized force overcorrects, launching the particle even further out in the opposite direction. On the next step, the force is even larger, causing the simulation to explode to infinity.
To prevent these non-physical explosions, researchers modify how explicit steps handle steep gradients. Martin Hutzenthaler et al. [Strong convergence of an explicit numerical method for SDEs with nonglobally Lipschitz continuous coefficients] introduced a "tamed" explicit Euler method. If the drift coefficient grows too large over a step, the method dampens (tames) the step size dynamically, preserving an optimal strong convergence rate of $1/2$ without requiring expensive implicit equations.
Designing Time-Independent Schemes
For true infinite-horizon reliability, we must prove that our numerical errors do not compound over time. This requires specialized mathematical structures.
Implicit and Projected Methods
Instead of taming explicit steps, one can use implicit methods, which naturally handle stiff, rapidly growing forces by evaluating the drift at the next time step.
Xiaojie Wang [Strong convergence rates for long-time approximations of SDEs with non-globally Lipschitz continuous coefficients] established a general theoretical framework proving that the backward Euler method maintains a strong convergence rate of $h^{1/2}$ uniformly across infinite time horizons for non-globally Lipschitz systems.
For weak convergence, Xiaojie Wang [Linear implicit approximations of invariant measures of semi-linear SDEs with non-globally Lipschitz coefficients] developed the Linear-Theta-Projected Euler (LTPE) scheme. This method achieves a time-independent weak convergence rate of order one, ensuring that long-term statistical averages remain accurate over indefinite simulation periods. To achieve higher-order accuracy without prohibitive computational costs, Assyr Abdulle et al. [High Order Numerical Approximation of the Invariant Measure of Ergodic SDEs] introduced a framework that decouples the weak order of the underlying trajectory generator from the high-order accuracy achieved for the invariant measure itself.
Advanced Error Analysis: The Poisson Equation
How do we mathematically prove that a local numerical error made at step 100 doesn't ruin the simulation at step 1,000,000?
The breakthrough mechanism relies on the Poisson equation. As formalized by Jonathan C. Mattingly et al. [Convergence of Numerical Time-Averaging and Stationary Measures via Poisson Equations], the Poisson equation acts as a bridge. It maps the accumulated long-term statistical error back to a sum of local, finite-time errors. If the underlying SDE is geometrically ergodic—meaning it pulls states toward its equilibrium exponentially fast—the long-term impact of an early numerical error decays exponentially. This allows researchers to establish sharp, uniform error bounds over long durations, a technique extended to non-local fractional systems by Junqing Jia et al. [Improved uniform error estimates for the two-dimensional nonlinear space fractional Dirac equation with small potentials over long-time dynamics].
Local Numerical Error ───(Poisson Equation)───> Global Invariant Measure Error
│ ▲
└───────────(Geometric Ergodicity)───────────────────┘
(Past errors decay exponentially)
Scaling Up: Acceleration and Data-Driven Schemes
With stable, long-time convergent schemes in hand, the focus shifts to computational efficiency and model discovery.
Computational Acceleration
Computing expectations over an invariant measure requires simulating trajectories for a very long time. To accelerate this, Michael B. Giles et al. [Multi-level Monte Carlo methods for the approximation of invariant measures of stochastic differential equations] adapted Multi-level Monte Carlo (MLMC) methods to invariant measures. By combining many computationally cheap trajectories evaluated with large step sizes with a few expensive trajectories evaluated with fine step sizes, they achieve optimal computational complexity even for non-globally Lipschitz drifts.
Data-Driven Flow Maps
When the underlying SDE coefficients are unknown or too complex to evaluate efficiently, we can turn to data-driven methods. Xingjie Helen Li et al. [ISALT: Inference-based schemes adaptive to large time-stepping for locally Lipschitz ergodic systems] developed the ISALT framework. Instead of running a costly implicit simulation at every step, ISALT uses data to learn an explicit approximation of the long-term "flow map." This allows scientists to use large time steps while mathematically preserving the true ergodic measures of locally Lipschitz systems.
Complementing this, sparse inference methods like those proposed by Youssef Mabrouk et al. [Sparse inference and active learning of stochastic differential equations from data] allow the direct discovery of governing Langevin equations from noisy experimental data, ensuring that the inferred models themselves are structured for long-term stability.
Summary
Simulating SDEs over infinite horizons requires a fundamental shift from tracking individual trajectories to preserving geometric and statistical properties. By moving beyond standard finite-time Lipschitz assumptions and leveraging tools like taming, projection, and Poisson error analysis, modern numerical analysis ensures that our simulations remain anchored to physical reality, no matter how long they run.
Key Foundational Work
LONG-TERM BEHAVIOUR OF LARGE MECHANICAL SYSTEMS WITH RANDOM INITIAL DATA (2002) — contributed 1 concept elements to this direction.
Waveform relaxation as a dynamical system (1997) — contributed 1 concept elements to this direction.