Many of our most powerful sampling algorithms—such as Langevin Monte Carlo, Hamiltonian Monte Carlo (HMC), and reverse-time diffusion models—are natively designed as continuous-time dynamical systems. Mathematically, these systems possess elegant properties: if you allow a particle to drift along a vector field dictated by the gradient of a target distribution while subjecting it to continuous thermal noise, its long-term statistical distribution is guaranteed to exactly match your target.
However, digital computers cannot simulate continuous time. They must break continuous trajectories into discrete intervals of a finite step size $\eta$. This transition from continuous differential equations to discrete updates introduces a fundamental pathology: discretization bias. The discrete chain no longer converges to the true target distribution $\pi$, but rather to a perturbed stationary distribution $\pi_\eta$.
Understanding why this bias occurs, why it resists simple fixes in large-scale machine learning, and how modern algorithms neutralize it is essential for designing fast, exact samplers.
The Mechanics of the Shift: Why Mismatches Accumulate
To understand why discretization shifts the stationary distribution, consider the structural decomposition of a single simulation step. As Andre Wibisono demonstrates in Sampling as optimization in the space of measures: The Langevin dynamics as a composite optimization problem, standard discretization schemes combine a discrete forward gradient step for the potential energy with an exact continuous integration of the internal energy (the Brownian motion). This creates a structural mismatch.
When a particle takes a discrete step along a gradient, it implicitly assumes that the geometry of the target distribution is locally flat over the interval $\eta$. If the target distribution is highly curved or ill-conditioned, a straight-line step causes the particle to systematically overshoot or undershoot the true energy contours.
Continuous Trajectory: ───( Smooth Curve )───> Exact Target
Discrete Simulation: ───[ Linear Step ]───> Overshoot / Perturbed Target
Over thousands of iterations, these local geometric errors accumulate. In deterministic settings like Neural Ordinary Differential Equations (Neural ODEs), Aiqing Zhu et al. show that the model ends up learning a solver-specific "perturbed" version of the system rather than the true underlying physical dynamics (On Numerical Integration in Neural Ordinary Differential Equations). In generative diffusion models, evaluating the score function at fixed intervals leads to cumulative trajectory errors that manifest as structural distortions or loss of detail in the final generated images (Faster Diffusion Sampling with Randomized Midpoints: Sequential and Parallel).
The Scalability Dilemma: Why Traditional Fixes Fail
Classically, the standard tool to eliminate discretization bias is the Metropolis-Hastings (MH) correction step. By calculating the probability that a discrete step deviated from the continuous ideal, the algorithm accepts or rejects proposals to ensure the chain preserves the exact target distribution. For example, Raaz Dwivedi et al. demonstrate that adding an MH step to Langevin dynamics (yielding the Metropolis-Adjusted Langevin Algorithm, or MALA) exponentially improves error tolerance, moving from a polynomial dependence on accuracy to a logarithmic one (Log-concave sampling: Metropolis-Hastings algorithms are fast!).
However, the MH step introduces a massive bottleneck for modern, large-scale machine learning. Computing the acceptance probability requires evaluating the likelihood over the entire dataset. When algorithms like Stochastic Gradient Langevin Dynamics (SGLD) or Stochastic Gradient HMC (SGHMC) use mini-batches to scale to large datasets, computing a full MH step defeats the purpose of subsampling (Max Welling et al., Tianqi Chen et al.).
Worse yet, data subsampling introduces its own stochastic noise, which interacts destructively with discretization. As Michael Betancourt proves in The Fundamental Incompatibility of Scalable Hamiltonian Monte Carlo and Naive Data Subsampling, the bias introduced by mini-batch gradient estimation is fundamentally decoupled from the numerical step size. Unlike pure integration errors, which vanish as $\eta \to 0$, subsampling bias causes trajectories to permanently drift away from the target distribution's energy levels regardless of how small the step size is.
Algorithmic Strategies for Neutralizing Bias
Faced with the failure of naive step-size reduction and full MH corrections, researchers have developed several distinct paradigms to mitigate or eliminate discretization bias.
1. Randomized Midpoint Methods
Instead of evaluating the gradient at the beginning of a discrete step, we can evaluate it at a randomized intermediate point. Ruoqi Shen introduced this for log-concave sampling (The Randomized Midpoint Method for Log-Concave Sampling), showing that treating the midpoint stochastically cancels out the first-order geometric errors that cause systematic drift.
By injecting randomness into the discretization interval itself, the overshoots and undershoots symmetrically cancel out on average. Ye He et al. established the asymptotic normality of this method, proving how it stabilizes the stationary distribution without requiring an explicit rejection step (On the Ergodicity, Bias and Asymptotic Normality of Randomized Midpoint Sampling Method).
2. Extrapolation and Amortization
Another approach is to accept the presence of bias during simulation but mathematically subtract it out later. Alain Durmus et al. leverage Richardson-Romberg extrapolation by running parallel simulation chains with different step sizes ($\eta$ and $\eta/2$) and linearly combining their outputs (Stochastic Gradient Richardson-Romberg Markov Chain Monte Carlo). Because the leading-order discretization error is a known function of $\eta$, this combination cancels out the bias term entirely.
Alternatively, Ruqi Zhang et al. propose AMAGOLD, which amortizes the expensive MH correction by applying it only once every $T$ steps (AMAGOLD: Amortized Metropolis Adjustment for Efficient Stochastic Gradient MCMC). This preserves exact convergence to the target distribution while maintaining the computational efficiency of mini-batch updates.
3. Continuous-Time Event-Driven Sampling
The most radical solution is to abandon time-discretization entirely. Piecewise deterministic Markov processes, such as the zigzag sampler analyzed by Jianfeng Lu 0001 et al., simulate continuous trajectories by calculating exact event times (like particle reflections off a potential barrier) rather than stepping forward by a fixed clock tick $\eta$ (Complexity of zigzag sampling algorithm for strongly log-concave distributions). Because the algorithm moves from event to event in continuous time, discretization bias is completely bypassed, allowing for exact sampling on digital hardware.