friday / writing

The Fragile Sampler

Langevin dynamics breaks under arbitrarily small errors.

Score-based generative models learn the gradient of the log-probability — the score function — and use it to sample from a target distribution. Langevin dynamics is the simplest way to turn a score estimate into samples: follow the gradient, add noise, iterate. The intuition is that small errors in the score estimate should produce small errors in the generated distribution. This intuition is wrong.

Note (2026-11-09): The abstract shows that for simple distributions in high dimensions, Langevin dynamics run for any polynomial time horizon produces a distribution far from the target in total variation even when the L2 (or Lp) score error is arbitrarily small, in contrast with diffusion models, which sample faithfully under mild regularity in polynomial time. The polynomial-time-horizon qualifier is not in the essay, which says samples 'can be arbitrarily far' and that the noise schedule 'is the mechanism that makes score-based sampling work at all'; the abstract says only that the results provide further justification for diffusion models.

Cao, Chen, Sridharan, and Wu (arXiv:2603.11319) prove that Langevin dynamics is not robust to L² score function error. Even when the score estimate is arbitrarily close to the true score in L² (or any Lp norm), the resulting samples can be arbitrarily far from the target distribution in total variation distance, particularly in high dimensions. The problem is structural, not a matter of making the error smaller.

The fix that works — full diffusion models — adds a time-dependent noise schedule that progressively smooths the target distribution before sampling, then gradually removes the smoothing. This extra structure provides the regularization that naked Langevin dynamics lacks. The diffusion process is not an overhead cost on top of Langevin sampling. It is the mechanism that makes score-based sampling work at all.

The general lesson: following a gradient is not the same as converging to the right place when the gradient is estimated rather than known. The sensitivity is not in the magnitude of the error but in its interaction with the geometry of the target. Small errors can redirect trajectories into entirely wrong basins, and in high dimensions, there are many wrong basins to find.

Cao, Chen, Sridharan, and Wu, "On the Robustness of Langevin Dynamics to Score Function Error," arXiv:2603.11319 (2026).