Conditional Flow Matching for Transport Between Markov Processes

Syamantak Kumar, Dheeraj Nagaraj, Saptarshi Roy, Purnamrita SarkararXiv 2026 · arXiv

TL;DR

This paper introduces a framework for transporting trajectories between Markov processes using Conditional Flow Matching (CFM). By learning a transition-level mapping that accounts for the current state of both the source and target, the authors provide a method to adapt time-series data across domains (e.g., between different human subjects) while preserving the underlying Markovian structure, backed by rigorous sample complexity bounds and strong empirical results in neural decoding.

What Problem Does This Paper Address?

In many real-world scenarios, we need to map a sequence of data from a "source" process to a "target" process. Think of sensor calibration, where one sensor's readings must be mapped to another's, or brain-computer interfaces, where EEG signals from a new user must be translated into a format a pre-trained model understands.

The challenge is twofold:

  1. Limited Data: We often have very few full trajectories (sometimes just one) from the source and target, even if those trajectories are very long.
  2. Temporal Dependence: Standard machine learning assumes data points are independent. In a time series, each point depends on the previous one. If we treat a whole trajectory as a single data point, we have a sample size of one. If we treat each time step as a data point, we violate the independence assumption.

The authors ask: Can we learn a map from just one source and one target trajectory that successfully transforms a new source trajectory into one that follows the target's dynamics?

Methods

The authors propose a recursive approach based on Conditional Flow Matching (CFM). Instead of trying to map the entire sequence at once, they learn to transport individual transitions.

Figure 1: Each learned flow transports one source transition to one target transition. The index i denotes time in the Markov process, while t ∈[0, 1] denotes flow time within a single transition.

Key components include:

  • Conditional Velocity Learning: They learn a velocity field $v_t$ that describes how to move a source state $x_{i+1}$ toward a target state $y_{i+1}$, conditioned on the current states $(x_i, y_i)$.
  • Recursive Rollout: At test time, given a new source trajectory, they generate the target trajectory step-by-step. The next target state $\hat{y}_{i+1}$ is produced by solving an ODE starting from the source state $x'_{i+1}$, using the learned velocity field conditioned on the previous source and generated target states.
  • Theoretical Guarantees: The paper proves that this method is "population correct"—meaning it perfectly recovers the target distribution given infinite data. More importantly, they derive finite-sample error bounds. They show that the error depends on the mixing time of the Markov processes (how long it takes the process to "forget" its starting state).

Datasets

The researchers evaluated their method on three types of data:

  1. Synthetic Gaussian Transitions: Linear Markov processes in 10 and 100 dimensions to test basic recovery.
  2. Nonlinear Mixture Benchmark: A more complex synthetic task involving bimodal (two-peaked) target noise and nonlinear transition functions.
  3. THINGS-EEG2: A large-scale real-world dataset containing EEG recordings from human subjects viewing 1,854 different images. The goal is image retrieval: identifying which image a person was looking at based on their brain activity.

Key Insights and Ideas

  • Exploiting Time Homogeneity: Because Markov processes are time-homogeneous (the rules don't change over time), a single long trajectory provides many samples of the transition rule. This allows the model to learn effectively even from a single pair of trajectories.
  • The Cost of Dependence: The theoretical analysis reveals that temporal dependence acts as a "tax" on sample size. The effective number of samples is roughly the total number of transitions $m$ divided by the mixing time $K$.
  • Conditioning is Mandatory: The authors show that simply matching the marginal distributions (the overall frequency of states) isn't enough. To preserve the target's transition law, the flow must be conditioned on the history (the current states).
  • Lower Bound Necessity: They prove a lower bound showing that the dependence on mixing time is unavoidable; no estimator can escape the statistical difficulty imposed by slowly mixing chains.

Key Results

  • Synthetic Performance: In both Gaussian and nonlinear tasks, the conditional flow significantly outperformed "marginal flows" (which ignore history) and "ablated flows" (which ignore the source state history). As training data increased, the error for the proposed method dropped much faster than the baselines.
  • EEG Image Retrieval: By adding a conditional flow "adapter" to the state-of-the-art ENIGMA decoder, the researchers improved the mean top-5 retrieval accuracy from 43.87% to 49.05%. This represents an 11.82% relative improvement in decoding accuracy across ten different human subjects.
  • Context Matters: The EEG experiments showed that accuracy improved as the model was given more "context" (previous EEG samples), validating the importance of the Markovian conditioning.

Limitations and Future Work

  • Computational Latency: Generating trajectories requires solving an ODE at every time step. This is sequential and can be slow. The authors suggest exploring parallel methods like Picard iteration to speed this up.
  • Finite Class Assumption: The current statistical proofs assume a finite class of candidate velocity functions. Extending these proofs to the continuous parameter spaces of deep neural networks is a remaining theoretical challenge.
  • Long-Range Memory: The current work focuses on first-order Markov processes. Adapting the framework for processes with longer-term dependencies (higher-order Markov or general non-Markovian time series) is an open direction.

Summary updated

Abstract

Motivated by sequence-to-sequence transport in the context time-series domain adaptation, we study the problem of transportation between trajectories of Markov processes. Given a limited number of trajectories from source distribution and the target distribution, we formulate a flow matching based algorithm which learns a transport map from the source to target trajectory distribution, while preserving the Markov structure. We show that this is consistent in the population limit and derive finite-sample error bounds under mixing time assumptions, following the analysis of classical statistical problems including regression (Nagaraj et al., 2020), principal component analysis (Kumar and Sarkar, 2023), and matrix concentration (Neeman et al., 2024) in the Markov setting. We complement that with a lower-bound construction showing that a mixing-time dependent sample complexity is unavoidable even with regular Gaussian conditional transitions. We evaluate on synthetic and real-world data. For image retrieval from electroencephalography (EEG) on THINGS-EEG2 (Gifford et al., 2022), the task is to identify the viewed image from EEG signals captured from human subjects, which suffers from high inter subject variability. We augment the ENIGMA decoder (Kneeland et al., 2026) with a conditional flow before its subject-specific temporal map. This improves mean top-5 retrieval accuracy from 43.87% to 49.05%, an 11.82% relative improvement.
BibTeX
@misc{kumar2026conditional,
  author = {Syamantak Kumar and Dheeraj Nagaraj and Saptarshi Roy and Purnamrita Sarkar},
  title = {{Conditional Flow Matching for Transport Between Markov Processes}},
  year = {2026},
  eprint = {2610.07229},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.07229},
}
Figure