Abstract
Stochastic interpolants (SIs) provide a flexible framework for generative modeling by constructing paths between arbitrary distributions. However, their practical implementation relies on discrete-time samplers, creating a gap between continuous-time theory and discrete-time execution. This paper presents the first finite-time convergence analysis for discrete-time SIs. We derive an upper bound on the KL divergence and introduce an exponentially decaying time schedule that achieves approximately a 10x speedup in convergence compared to uniform discretization.
Introduction
Generative modeling has increasingly adopted flow-based and diffusion-based paradigms. Stochastic interpolants (SIs) have emerged as a powerful generalization, enabling the construction of transport paths between any two distributions $\rho_0$ and $\rho_1$. Unlike standard diffusion models, SIs do not require a Markovian structure, allowing for more flexible interpolation functions.
Despite their theoretical appeal, SIs are implemented using discrete-time samplers such as the Euler-Maruyama method. The lack of a Markovian property makes error analysis challenging, leaving a significant gap in our understanding of how discretization affects the final generated distribution.
Theoretical Framework
The authors address the discretization gap through a rigorous analytical framework. The core challenge lies in the fact that SIs use an interpolation function $x_t = I(x_0, x_1, t)$ which, while mathematically elegant, complicates the score and velocity function estimation in discrete steps.
Discretization Error Decomposition
Using Girsanov’s Theorem and Itô’s calculus, the researchers decompose the distribution error into components related to the velocity function (governing the mean drift) and the score function (governing the local density shape). This decomposition allows for a granular view of where errors accumulate during the sampling process.
Exponentially Decaying Schedule
A key finding is that discretization error tends to explode near the boundaries ($t=0$ and $t=1$) because the noise scale $\gamma(t)$ vanishes. To mitigate this, the paper proposes a principled, exponentially decaying time schedule. By setting the step size $h_k$ proportional to $\gamma^2(t)$, the problematic terms in the error bound are neutralized, leading to more stable and faster convergence.
Results and Discussion
The theoretical findings were validated using 2D synthetic datasets (checkerboard and spiral) and high-dimensional Gaussian mixtures.
Convergence and Speedup
The proposed exponentially decaying schedule demonstrated a significant performance boost. In transforming a checkerboard to a spiral, the new schedule reached convergence roughly 10 times faster than a standard uniform step-size approach.
Dimension Scaling
The analysis achieved a KL error bound of $\epsilon^2$ in $\tilde{O}(d^2/\epsilon^2)$ steps. While the theoretical bound suggests a quadratic dependence on dimension $d$, empirical results (Figure 4b) show a near-linear growth. This suggests that the derived bounds are robust, albeit slightly conservative for certain distribution classes.
The Role of Coupling
The study also highlights that the distance between source and target distributions ($E\|x_0 - x_1\|^p$) directly impacts the error bound. This provides theoretical support for the practice of using "closer" base distributions to simplify the generative task.
Conclusion
This work provides the first rigorous finite-time analysis for discrete-time stochastic interpolants, bridging a critical gap in generative modeling theory. The introduction of an optimized time schedule offers a practical 10x speedup for researchers and practitioners. Future work will look toward extending this analysis to ODE-based flow-matching models and narrowing the gap between theoretical $O(d^2)$ bounds and observed linear scaling.
Summary updated