Is Flow Matching Just Trajectory Replay for Sequential Data?

Flow matching (FM) is increasingly used for time-series generation, but it is not well understood whether it learns a general dynamical structure or simply performs an effective “trajectory replay”. We study this question by deriving the velocity field targeted by the empirical FM objective on sequential data, in the limit of perfect function approximation. For the Gaussian conditional paths commonly used in practice, we show that the implied sampler is an ODE whose dynamics constitutes a nonparametric, memory-augmented continuous-time dynamical system. The optimal field admits a closed-form expression as a similarity-weighted mixture of instantaneous velocities induced by past transitions, making the dataset dependence explicit and interpretable. This perspective positions neural FM models trained by stochastic optimization as parametric surrogates of an ideal nonparametric solution. Using the structure of the optimal field, we study sampling and approximation schemes that improve the efficiency and numerical robustness of ODE-based generation. On nonlinear dynamical system benchmarks, the resulting closed-form sampler yields strong probabilistic forecasts directly from historical transitions, without training.

💡 Research Summary

The paper investigates a fundamental question that has emerged alongside the rapid adoption of flow‑matching (FM) methods for time‑series generation: does FM learn a genuine continuous‑time dynamical model, or does it simply act as a sophisticated “trajectory replay” mechanism that stitches together observed one‑step transitions? To answer this, the authors adopt the idealised setting of perfect function approximation and derive the exact velocity field that minimizes the empirical FM loss on a finite set of sequential transitions.

First, the authors formalise the data. They assume N independent trajectories of a d‑dimensional state X_t, each sampled at uniform intervals, and extract every one‑step transition (x_τ, x_{τ+1}). Collecting all such pairs across all trajectories yields a transition dataset D_M = {X^{(j)} = (X^{(j)}1, X^{(j)}2)}{j=1}^M, which serves as a memory bank. The underlying (unknown) dynamics may be deterministic (X{τ+1}=F(X_τ)) or a discretisation of a continuous‑time ODE dX_t = f(X_t)dt; in either case the empirical distribution of transitions is taken as a Monte‑Carlo approximation of the true transition law.

Flow‑matching constructs a conditional probability path p_t(z|X) that transports a simple base distribution (typically a Gaussian) to the data distribution at t=1. The authors specialise to the Gaussian “bridge” path that is widely used in FM literature: for a given transition X^{(j)} they define

 Z_t^{(j)} = (1−t) X^{(j)}_1 + t X^{(j)}_2 + c_t ξ, ξ∼N(0,I),

with c_t² = σ_min² + σ² t(1−t). This path linearly interpolates between the two observed states while adding isotropic Gaussian noise whose variance varies smoothly with t. The associated velocity field is affine in z:

 v(t,z|X^{(j)}) = a_t(X^{(j)}) z + b_t(X^{(j)}),

where a_t and b_t are explicit functions of σ, σ_min and the transition endpoints.

The empirical FM objective is

 \hat L_{CFM}