A note on connections between the Föllmer process and the denoising diffusion probabilistic model

Abstract: The Föllmer process is a Brownian motion conditioned to have a pre-specified distribution at time 1. This process can be interpreted as an "augmented" time-compressed version of the reverse stochastic differential equation (SDE) for the denoising diffusion probabilistic model (DDPM). While this fact has been indirectly used to analyze DDPM sampling errors via discretization of the reverse SDE, connections between direct discretization of the Föllmer process and the DDPM sampler have not yet been fully explored. This note aims to clarify this point while surveying relevant results from existing work. We show that discretized Föllmer processes give natural hyper-parameter settings of the DDPM sampler. Moreover, this allows us to systematically recover state-of-the-art results on DDPM sampling error bounds with slight improvements.

• The paper establishes that discretized Föllmer processes are mathematically equivalent to canonical DDPM sampler frameworks.
• It demonstrates that the Euler-Maruyama scheme yields controlled KL divergence error bounds through precise hyper-parameter selection.
• The findings provide a unifying framework for adapting diffusion models to intrinsic low-dimensional data, enhancing both theoretical insights and practical implementations.

Connections Between the Föllmer Process and the Denoising Diffusion Probabilistic Model

Introduction

This paper investigates the under-explored connections between the Föllmer process—a Brownian motion conditioned to meet a target distribution at terminal time—and denoising diffusion probabilistic models (DDPMs), focusing on their implications for the theory and practice of diffusion-based generative modeling. While the analytic relationship between the Föllmer process and the time-reversed SDE underpinning DDPMs is known, the systematic study of how discretization approaches interrelate, and how this informs hyper-parameter selection and error analysis in DDPMs, has been lacking. This work rigorously clarifies these aspects, providing a framework that subsumes state-of-the-art error analyses for DDPMs and offers several improvements and theoretical insights.

The Föllmer Process: Markovian Structure, Entropy, and Drift Representation

A Föllmer process $X = (X_t)_{t \in [0,1]}$ associated with a target probability measure $\mu$ on $\mathbb{R}^d$ is a Brownian bridge process, initiating at the origin and conditioned such that $X_1 \sim \mu$. Its Markov property holds unconditionally, without entropy regularity assumptions on $\mu$. This process provides a weakest solution to the SDE:

$$dX_t = \left( \frac{X_t}{t} + \nabla \log p_t(X_t) \right) dt + dW_t, \quad X_0 = 0,$$

where $p_t$ is the marginal law of $X_t$ and $\nabla \log p_t$ is its score function.

A core analytic connection is the identity:

$$H(\mu \mid \gamma_d) = \frac{1}{2} \int_0^1 \mathbb{E}[|v_t|^2] dt,$$

where $\mu$0, $\mu$1 is relative entropy, and $\mu$2 is the Föllmer drift. This identity links the sample-path structure to the entropy profile along the process. Notably, the drift can be represented in two analytically distinct forms: as a function of the instantaneous score or as the scaled difference between the (conditional) mean at $\mu$3 and the path value, establishing a bridge to stochastic localization frameworks and the de Bruijn identity.

Stochastic Differential Equations in DDPMs and the Time-Compressed Föllmer Process

The reverse-time SDE formulation of DDPMs takes the form:

$\mu$4

with the reverse process realizing an SDE similar in structure to the Föllmer process, but under time rescaling and augmentation. Explicitly, the work exhibits that the Föllmer process can be mapped into the reverse SDE of DDPMs by the transformation $\mu$5. This formalizes the heuristic that DDPMs, as discretizations of reverse SDEs, are also discretizations of a time-compressed Föllmer process path.

Discretization, Score Estimation, and Sampling Error Bounds

The core technical contribution is the rigorous analysis of the effects of discretizing the Föllmer SDE and relating this to practical DDPM simulation algorithms. The Euler-Maruyama scheme is shown to correspond precisely to the standard DDPM iteration, including the canonical hyper-parameter setting for the update coefficients. The paper proves that under mild regularity (e.g., bounded second moment and consistent score estimation), the KL divergence between the terminal law of the discretized Föllmer process and the target can be tightly controlled. The explicit upper bounds provided improve prior work by eliminating dependence on ambient dimension under suitable initialization, and by sidestepping an explicit dependence on the duration parameter $\mu$6 (often used as a variance schedule proxy in DDPM literature) in regimes where such dependence is provably unnecessary.

Further, the error analysis is shown to adapt to both the Fisher information scenario (when $\mu$7 with finite Fisher information), thereby allowing early stopping to be avoided, and to settings where data are effectively low-dimensional (intrinsic dimension $\mu$8). For the latter, bounds depend on entropy and Gaussian width quantities related to the support of $\mu$9, not the ambient dimension.

Intrinsic Dimensionality and Adaptation

Recent theory shows that standard DDPM discretizations adapt to data concentrated near a low-dimensional manifold provided specific parameter and noise scheduling conditions are met. This work elucidates that the “mean-only” discretization—where only the conditional mean in the Föllmer drift is discretized—yields schemes whose error bounds scale with the intrinsic data complexity, not ambient dimension. The hyper-parameterization developed here, originally realized in the DDPM context via empirical tuning, is deduced formally from the properties of the Föllmer process and its drift representations. Under additional covering/entropy conditions on data support, dimension-free KL bounds are rigorously established, significantly sharpening the theoretical understanding of generative sampling on manifolds.

Implications and Future Directions

These results have clear implications for both theoretical investigation and practical deployment of diffusion models. The Föllmer process provides a unifying framework for analyzing score-based generative samplers, clarifying when reverse SDE discretizations suffice and how hyper-parameters should be set for optimal convergence. The intrinsic adaptivity of DDPMs to low-dimensional structure is put on rigorous foundations, informing both architectural choices—such as network classes for score estimation—and variance schedule design.

From a theoretical perspective, the deep connections to entropy, Fisher information, and stochastic localization underpin possibilities for further synthesis between information theory, functional inequalities, and generative modeling. Notably, the connections drawn here invite further analysis of Schrödinger bridges and optimal transport analogies in discrete-time generative modeling. There are open questions regarding minimax optimality rates, robustness to model misspecification in the score, and convergence in other metrics (e.g., Wasserstein, total variation). Finally, as recent works explore non-Euclidean targets, extension to manifold-valued Föllmer processes and corresponding diffusion models is anticipated.

Conclusion

This paper rigorously establishes that discretized Föllmer processes are mathematically equivalent to canonical DDPM sampler architectures, yielding natural hyper-parameter regimes and directly recovering (and in some settings improving upon) recent KL-based error bounds for diffusion models. These insights clarify the roles of discretization, score estimation, initialization, and intrinsic data dimension in generative model sample quality. Thus, this study both unifies the mathematical landscape underlying modern score-based generation and points to structured pathways for improved theoretical and algorithmic design of diffusion samplers (2605.18040).