acceptodds
Under review as a conference paper at ICLR 2027

Time Series Constrained Sampling via Refining Trajectory-Level Control

Abstract

Time series generation is central to applications such as finance, energy, and biomedical sensing, where realistic samples must respect partial observations or structural constraints. Existing training-free approaches based on guidance or projection enforce such constraints only locally at each diffusion step, which we show is insufficient to recover the conditional posterior in path space. In this work, we reformulate constrained sampling for time-series diffusion models as a trajectory-level stochastic optimal control problem, and unify guidance- and projection-based methods as instances of local control within this framework. We propose Trajectory-level Time-series Control (TTC), a training-free, inference-time adjoint scheme that parameterizes the control in Fourier and wavelet subspaces aligned with global and localized temporal structure. A single backward pass already moves samples substantially closer to the constraint set at guidance-level cost, additional iterations tighten this further, and a terminal projection supplies exact feasibility. Across four datasets and multiple constraint types, TTC reaches exact constraint satisfaction with the best distributional quality among the compared methods, while running up to an order of magnitude faster than per-step projection.

Primary area:
generative models
Keywords:
Time Series, Optimal Control

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ibBvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a training-free generation framework for time-series modeling. For reward-tilted sampling, the authors employ a stochastic optimal control formulation, where naive reward guidance has pitfalls, and propose an inference-time control method based on the adjoint matching idea. For a control parameterization appropriate for time-series modeling, they restrict the control space to low-frequency Fourier or coarse wavelet subspaces.

Critical strengths
  • For time-series modeling, restricting the control space to Fourier and wavelet subspaces for better control parameterization is interesting.
  • The proposed method shows superior results across multiple datasets and constraint types compared to the baselines.
Critical weaknesses
  • The motivation for functional parameterization in Section 4.1 is interesting, but it is unclear whether the proposed parameterization, which restricts the control space, yields a sufficiently accurate approximation to the optimal control. Specifically, the bound in Theorem 4.1 follows immediately from Girsanov’s theorem and the data-processing inequality. It is expressed entirely through the distance between the optimal control and its projection onto the restricted Fourier or wavelet subspace. The approximation bias therefore depends on how much of the optimal control lies outside this subspace, and that is exactly what the theorem leaves open. The bound vanishes only when the optimal control already lies in the subspace. If a substantial part of it lies outside, the bound becomes large and tells us nothing.

  • In my view, the novelty lies mainly in Section 4.2, since Section 3 largely builds on standard results in the stochastic optimal control literature, and the Fourier and wavelet restrictions in Section 4.1 are fixed linear filters applied to the control. Section 4.2 borrows the adjoint idea from Adjoint Matching. It restricts the control to low-frequency or coarse wavelet components, then updates it iteratively by backpropagating the terminal loss through the sampler. Some methods estimate the h-function in Doob’s h-transform without iterative backpropagation, as described in [1, Section 3.2]. I wonder how the proposed method compares with these methods in terms of performance.

  • It is unclear how the proposed algorithm recovers the optimal tilt. The paper states that the intractable score term (Doob’s h-transform function) is bypassed via PMP (lines 286–287). However, the adjoint used to define the optimal control remains the same intractable object (a conditional expectation over future noise [2,3]). In [2,3], a deterministic adjoint ODE such as Eq. (13) is sufficient because the per-path adjoint serves as a regression target. Its conditional expectation given the intermediate state yields the true adjoint, so the loss gradient is unbiased and the fixed point of the optimal control problem is preserved. However, Algorithm 1 seems to have no such step. The per-path adjoint omits the stochastic term and may not represent the required conditional expectation, yet it is used directly to form the control in the single-pass setting or update it during refinement. It is therefore unclear whether Algorithm 1 preserves the fixed point corresponding to the optimal control.


[1] Potaptchik et al., Meta Flow Maps enable scalable reward alignment
[2] Domingo-Enrich et al., Adjoint Matching through the Lens of the Stochastic Maximum Principle in Optimal Control
[3] Park et al., Functional Adjoint Sampler
Questions for authors
  • How is the terminal loss function L is defined for the experiment? I wonder how the gradient is computed for the reward-guidance or backpropagation.

  • Minor point: The Schrödinger bridge problem targets minimize the KL divergence to the reference process with the marginals prescribed at both ends. However, Eq.(4) fixes only the initial distribution and replaces the terminal-marginal constraint with a soft terminal cost. It is therefore a KL-regularized optimal control problem with a terminal cost, not a bridge between two given marginals.

Reviewer_5mz8OpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper formulates constrained conditional sampling for pretrained diffusion models as a trajectory-level stochastic optimal control (SOC) problem. It provides a unified perspective on guidance- and projection-based methods and characterizes their path-space suboptimality. Building on this formulation, the authors propose Trajectory-level Time-series Control (TTC), which restricts the control to Fourier or wavelet subspaces and optimizes it through an inference-time adjoint scheme, followed by a terminal projection.

Critical strengths

A useful trajectory-level perspective on constrained sampling. Formulating conditional sampling as SOC provides a unified way to analyze guidance and projection, and the paper connects the limitations of these methods to their use of local surrogates rather than trajectory-level control.

A practical method supported by theoretical and empirical analysis. The proposed control parameterization and adjoint-based refinement yield promising results across multiple datasets and constraint types, with substantial computational savings over per-step projection.

Critical weaknesses

The distinction between the general SOC formulation and the time-series-specific contribution could be clearer. Section 3 appears largely applicable to conditional diffusion sampling beyond time series, whereas Section 4 introduces Fourier- and wavelet-based parameterizations motivated by temporal structure. The paper would benefit from a clearer explanation of which theoretical results are new and broadly applicable, and which aspects of TTC specifically depend on time-series structure. This is primarily a concern about the scope and positioning of the contribution, rather than a fundamental flaw in the proposed method.

Questions for authors

Clarification of generality and novelty: Section 3 appears to formulate conditional sampling in a general SOC framework. How does this formulation and its analysis differ technically from existing SOC-based approaches to inference-time diffusion guidance and constrained sampling? Which results in this section are specific to the proposed framework rather than consequences of the standard SOC formulation?

Scope beyond time series: Is the proposed trajectory-level control procedure applicable to other constrained diffusion-sampling problems, with only the control parameterization needing to change? If so, what properties of time-series data are essential to the current method, beyond motivating the Fourier and wavelet subspaces?

Reviewer_awe7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Trajectory-level Time-series Control (TTC), inference-time constrained diffusion sampling as an entropy-regularized stochastic control problem. TTC approximates control through terminal-loss adjoints propagated backward through a sampled diffusion trajectory, restricts updates to low-frequency Fourier or coarse Haar-wavelet subspaces, and iteratively refines the control before a final penalty projection. The reported experimental evaluations use a shared Diffusion-TS backbone and report performance on common benchmark datasets like Stock, ETTh, Energy, and fMRI, obtaining favorable quality ranks and substantially lower runtime than the implemented per-step optimization baseline.

Critical strengths
  1. The trajectory sensitivity and restricted control schemes are plausible, practical ideas, supported meaningfully by the ablations. Backpropagating a terminal requirement through the remaining sampler can capture dependencies missed by a denoised-point surrogate, and the Fourier/wavelet restrictions give a concrete time-series inductive bias. Table 4 shows that better average ranks are attained for these variants than unrestricted TTC-Direct, and no-projection ablations demonstrate that control itself reduces violations.
  2. The paper attempts to separate feasibility from distributional fidelity, and makes clear some limitations of its approximation. In particular, the linear-Gaussian experiment highlights distance to an explicit tilted target rather than relying only on, e.g., difficult to interpret embedding metrics. Appendix B.10 shows the refinement's Gaussian covariance remains uncontracted and Table 12 shows KL to a fixed soft target increasing with iteration, which together are useful interpretations that help disentangle observed reductions in violation from improved posterior sampling.
Critical weaknesses
  1. The main posterior-control theory does not seem to strictly establish posterior correctness, or a controlled approximation to it, for the proposed algorithm. Theorem 4.1 bounds the effect of projecting the unknown optimal feedback control, while TTC instead projects an adjoint estimated along sampled trajectories. The deterministic pathwise sensitivity replacement of the stochastic value-function adjoint (Appendix A.2) can provide a gradient estimator for an open-loop terminal-loss objective, but it is not automatically the optimal adapted feedback control or an approximation with the error claimed for R(u∗)R(u^*). The stationarity proposition is conditional on smoothness and unbiased gradient estimates and concerns the terminal-loss objective, not the original entropy-regularized objective or distance to the posterior.

These theoretical distinctions appear meaningful, and while TTC seems plausibly a good constraint-steering heuristic, its central theoretical contrast with locally guided samplers is not formally an established advantage for conditional posterior sampling. It seems the qualifications made in the appendix should likely be reflected in the main claim (or perhaps additional algorithm-specific argument provided or clarified if I've fundamentally misinterpreted the claims).

Questions for authors
  1. What happens to quality and end-to-end runtime when all methods receive the exact analytical projections for Mean, DeltaMean, and Max with identical final feasibility tolerances? This would be helpful for determining whether the main feasibility/speed advantage remains under more controlled setups.
  2. Relating to novelty claims, how does TTC compare with an inference-time trajectory-control method such as Diffusion Trajectory Matching (cited in Section 2.2) under the same backbone and compute budget?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.