Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Error Analysis of Discrete Flow with Generator Matching

Zhengyan Wan, Yidong Ouyang, Qiang Yao, Liyan Xie, Fang Fang, Hongyuan Zha, Guang Cheng

TL;DR

This work addresses the lack of theoretical understanding for discrete flow models by developing a unified generator-matching framework anchored in CTMC theory. It derives a Girsanov-type bound on path measures to bound the KL divergence between CTMC marginals and provides non-asymptotic error bounds that decompose into stochastic error, approximation error, and early-stopping error, with and without a bounded-rate assumption. A key novelty is the use of uniformization for exact sampling, which yields zero initialization and discretization errors, and the explicit handling of transition-rate estimation errors. The results include fast-rate bounds under boundedness, nearly-optimal rates in terms of sample size, and practical guidance validated by simulations, illustrating how to balance estimation and early stopping. Overall, the paper delivers the first comprehensive non-asymptotic error analysis for discrete flow models and clarifies how discrete-flow methods can be rigorously controlled and deployed in practice.

Abstract

Discrete flow models offer a powerful framework for learning distributions over discrete state spaces and have demonstrated superior performance compared to the discrete diffusion models. However, their convergence properties and error analysis remain largely unexplored. In this work, we develop a unified framework grounded in stochastic calculus theory to systematically investigate the theoretical properties of discrete flow models. Specifically, by leveraging a Girsanov-type theorem for the path measures of two continuous-time Markov chains (CTMCs), we present a comprehensive error analysis that accounts for both transition rate estimation error and early stopping error. In fact, the estimation error of transition rates has received little attention in existing works. Unlike discrete diffusion models, discrete flow incurs no initialization error caused by truncating the time horizon in the noising process. Building on generator matching and uniformization, we establish non-asymptotic error bounds for distribution estimation without the boundedness condition on oracle transition rates. Furthermore, we derive a faster rate of total variation convergence for the estimated distribution with the boundedness condition, yielding a nearly optimal rate in terms of sample size. Our results provide the first error analysis for discrete flow models. We also investigate model performance under different settings based on simulation results.

Error Analysis of Discrete Flow with Generator Matching

TL;DR

This work addresses the lack of theoretical understanding for discrete flow models by developing a unified generator-matching framework anchored in CTMC theory. It derives a Girsanov-type bound on path measures to bound the KL divergence between CTMC marginals and provides non-asymptotic error bounds that decompose into stochastic error, approximation error, and early-stopping error, with and without a bounded-rate assumption. A key novelty is the use of uniformization for exact sampling, which yields zero initialization and discretization errors, and the explicit handling of transition-rate estimation errors. The results include fast-rate bounds under boundedness, nearly-optimal rates in terms of sample size, and practical guidance validated by simulations, illustrating how to balance estimation and early stopping. Overall, the paper delivers the first comprehensive non-asymptotic error analysis for discrete flow models and clarifies how discrete-flow methods can be rigorously controlled and deployed in practice.

Abstract

Discrete flow models offer a powerful framework for learning distributions over discrete state spaces and have demonstrated superior performance compared to the discrete diffusion models. However, their convergence properties and error analysis remain largely unexplored. In this work, we develop a unified framework grounded in stochastic calculus theory to systematically investigate the theoretical properties of discrete flow models. Specifically, by leveraging a Girsanov-type theorem for the path measures of two continuous-time Markov chains (CTMCs), we present a comprehensive error analysis that accounts for both transition rate estimation error and early stopping error. In fact, the estimation error of transition rates has received little attention in existing works. Unlike discrete diffusion models, discrete flow incurs no initialization error caused by truncating the time horizon in the noising process. Building on generator matching and uniformization, we establish non-asymptotic error bounds for distribution estimation without the boundedness condition on oracle transition rates. Furthermore, we derive a faster rate of total variation convergence for the estimated distribution with the boundedness condition, yielding a nearly optimal rate in terms of sample size. Our results provide the first error analysis for discrete flow models. We also investigate model performance under different settings based on simulation results.

Paper Structure

This paper contains 40 sections, 20 theorems, 142 equations, 4 figures, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Theoretical Background on Discrete Flow-Based Models
  3. Continuous-Time Markov Chain as Stochastic Integral
  4. Discrete Flow-Based Models
  5. Main Results
  6. Additional Notations
  7. Bound for KL Divergence
  8. Error analysis without boundedness condition
  9. Error analysis with boundedness condition
  10. Simulations
  11. Related Works
  12. Conclusion and Future Works
  13. Uniformization Algorithm
  14. Additional Results for CTMCs
  15. Compensated Random Measure
  16. ...and 25 more sections

Key Result

Proposition 1

Assume that $u_t(z,x)_{z,x\in\mathcal{S}^\mathcal{D}}$ is a rate matrix satisfying that (a) $-u_t(x,x)\leq M$ for any $x\in\mathcal{S}^\mathcal{D}$; (b) $u_t(z,x)$ is $L$-Lipschitz continuous in $t$ for any pair $(z,x)\in\mathcal{S}^\mathcal{D}\times\mathcal{S}^\mathcal{D}$. Suppose that $T_1,T_2,\d where $R_t\leq (M^2+L)h^2=O(h^2)$. Thus, $X(t)$ is a CTMC with the rate $u_t$.

Figures (4)

  • Figure 1: Prediction risk v.s. early stopping parameter.
  • Figure 2: Total variation v.s. early stopping parameter with uniformization algorithm.
  • Figure 3: Prediction risk v.s. dimension.
  • Figure 4: Prediction risk v.s. sample size.

Theorems & Definitions (43)

  • Definition 1: CTMC
  • Proposition 1: Uniformization
  • Definition 2: Random Measure Associated with CTMC
  • Remark 1: Comparison to Lévy-Itô Decomposition
  • Proposition 2: Kolmogorov Forward Equation
  • Theorem 1: Bound for KL Divergence
  • Proposition 3: Estimation Error Decomposition 1
  • Definition 3: Uniform Covering Number, jiao2023deep
  • Theorem 2: Stochastic Error
  • Theorem 3: Approximation Error
  • ...and 33 more