acceptodds
Under review as a conference paper at ICLR 2027

Finite-Horizon Control Consistency for Rank-Deficient Diffusion Bridges

Abstract

Control-consistency methods for diffusion bridges recover state-space covectors from Brownian-space controls using instantaneous diffusion geometry, which becomes singular under rank-deficient noise even when the drift propagates forced directions into unforced coordinates. We introduce Finite-Horizon Bridge Control (FBC), replacing singular pointwise recovery under rank deficiency with a finite-horizon controllability lift built from dynamically transported noise directions. We show that the self-consistency identity itself does not require full-rank diffusion and connect the finite-horizon Gramian to Malliavin covariance. We prove exact finite-horizon recovery for controllable time-inhomogeneous affine diffusions with deterministic coefficients and characterize the information scales of dynamically reachable directions over short horizons in the local linearized controllability geometry. For nonlinear systems, we give sufficient executable-consistency conditions and an explicit short-window convergence rate for underdamped Langevin dynamics. FBC thus extends control consistency to rank-deficient and hypoelliptic diffusion bridges by exploiting finite-time controllability.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion Bridges, Stochastic Control, Control Consistency, Rank-Deficient Diffusions, Hypoelliptic Diffusions, Finite-Time Controllability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dNPWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends control consistency diffusion bridge (CCDB) from Howard et al., 2026 to the case of a rank-deficient noise (i.e. when the covariance GGTGG^T is not invertible), where the pointwise covector recovery u=(GGT)−1Gvu=\left( GG^T\right)^{-1}Gv breaks down. The method pulls future Brownian columns back to an anchor through the auxiliary flow and solves a ridge least-squares problem over a window to recover the full covector. A short backward recursion propagates these covectors, which then serve as relaxed targets for a residual controller. Experiments include a rank-loss sweep, kinetic OU, full-rank benchmarks, underdamped double-well and Müller–Brown against NGDB.

Critical strengths

The method enables detached control-consistency training in the case of rank-deficient noise, which covers underdamped and constrained molecular dynamics. The previous method itself explicitly leaves this setting open. The extension is cheap as only a small ridge solve in a window space is needed. Notable result beyond small kinetic systems is on constrained-underdamped alanine (problem dimension 90, rank 45).

Critical weaknesses

Some of the theoretical claims that are presented without a clear reference seem like they are part of the contribution (Thm 4.1 and Thm 5.1, Prop 4.2). However, those are already well established results. This limits the novelty of the paper.

Questions for authors

Q1: What is the guide vgv^g, and how well does it perform on its own? The controller is a learned residual around vgv^g (never specified by the authors). This is needed to assess the following:

  • On the linear kinetic OU bridge the Doob control has a closed form so a guide from the linear dynamics could already be exact.
  • On alanine, FBC's terminal evaluation metric (RMSE) varies by only 10−610^{-6}nm which would be consistent with the guide doing most of the work.

Q2: Why not learn the covector directly? Eq. (5) holds without any rank condition on GG, and Appendix B.6 already analyzes a controller defined by a covector field uu, with controlled drift b+GG⊤ub + GG^\top u. For constant GG, the corresponding consistency map is a contraction on short horizons, again with no rank condition (eq. (86) and Thm. B.10). Training this controller on eq. (5) requires no recovery step. Does this approach fail in practice, and if so, why?

3 comments
AuthorsOpenReview ↗

Response to Reviewer dNPW

Response to Reviewer dNPW

We thank the reviewer for identifying the need to distinguish established theory from our contribution, specify the fixed guides, and evaluate direct covector learning. The revised manuscript addresses each point as follows.

1. Theoretical attribution and novelty

We agree that the underlying self-consistency identity, Malliavin–Gramian relation, and Kalman controllability criterion are established results. Section 4.1 now explicitly attributes the self-consistency identity in Theorem 4.1 to CCDB and clarifies that it does not require full-row-rank diffusion. Proposition 4.2 is introduced as a standard stochastic-flow identity with classical references, and Theorem 5.1 is explicitly named and cited as the Kalman criterion.

Our contribution is not these identities, but their use in an executable learning construction under rank deficiency: FBC transports Brownian directions across a finite window, recovers missing covector information by a regularized Gramian solve (Eq. 8), and uses it in detached control-consistency targets (Eqs. 13–14). Theorem 4.3 gives an exact finite-window recovery identity for controllable affine diffusions; Theorem 5.2 bounds the error of the resulting population executable-control estimate, with an explicit underdamped specialization in Corollary 5.3. We have sharpened this distinction in Sections 4–5.

2. Q1: Guide construction and guide-only performance

The learned controller is vθ=vg+gθv_\theta=v^g+g_\theta, with vgv^g fixed. We now specify the guides and report their standalone performance.

Kinetic OU (Appendix C.2.1, Table 7). The guide is obtained by Gaussian terminal conditioning of an auxiliary kinetic system with

\qquad A=\begin{pmatrix}0&1\\-1&-1\end{pmatrix}$$ for the guide and evaluated dynamics, respectively, using the same position-terminal observation $(H,y,R)=((1,0),1,0.04)$. The auxiliary system omits the position-restoring force, so its conditioned guide is not the exact Doob control of the evaluated bridge. | Method | Objective ↓ | Control RMSE ↓ | |---|---:|---:| | Exact Doob | 2.7014 ± 0.0354 | 0 | | Guide-only | 3.9298 ± 0.0318 | 0.5821 | | DCC | 3.2748 ± 0.0306 | 0.5206 | | **FBC** | **2.9047 ± 0.0325** | **0.4551** | FBC lowers the objective by **26.1%** versus Guide-only, and also improves control RMSE. **Constrained-underdamped alanine (Appendix C.6.3, Table 14).** FBC uses a fixed endpoint-only kinetic Gaussian guide conditioned on the target C7ax position, with terminal observation standard deviation $\delta=0.03\,\mathrm{nm}$. Its guiding acceleration is projected into the admissible AllBonds tangent-velocity space. The guide depends on the current position, velocity, time, and terminal conformation. | Method | ETS (kJ/mol) ↓ | Terminal RMSE (nm) ↓ | |---|---:|---:| | Guide-only | 9.26 | 0.0372803 | | NGDB (same guide) | 7.6797 | 0.0322745 | | **FBC** | **6.0692 ± 0.0329** | **0.029273 ± 0.000001** | Relative to Guide-only, FBC reduces ETS by **34.5%** and terminal RMSE by **21.5%**. The same-guide comparison gives additional evidence that improvements are not attributable solely to the fixed guide. The added controls are point estimates; the reported FBC uncertainty is over five seeds. ### 3. Q2: Why not learn the covector directly? We agree this alternative is theoretically valid: Eq. 5 does not require a rank condition, and Appendix B.6 establishes short-horizon contraction of the ideal direct-covector consistency map under its assumptions. We therefore implemented **Direct Covector Consistency (DCC)**, which directly parameterizes the position and momentum covector components, constructs self-consistency targets with the reference-flow Jacobian, and executes $v_\phi=G^\top u_\phi$, without any finite-window Gramian recovery (Appendix C.1). **DCC does not fail in principle or in our tests**: it improves over Guide-only. However, FBC performs better on both rank-deficient benchmarks. On kinetic OU, DCC obtains $3.2748\pm0.0306$ versus FBC's $2.9047\pm0.0325$. On underdamped Double-Well (Appendix C.4, Table 11), Guide-only, DCC, and FBC objectives are $14.7043\pm0.1173$, $12.2162\pm0.1030$, and **$8.2455\pm0.0532$**, respectively; FBC improves on DCC by **32.5%**. These results support the practical value of finite-horizon recovery without claiming that direct covector learning is impossible. We hope these clarifications and additional comparisons resolve the reviewer's concerns and invite reconsideration of the rating.

The novelty of the paper stays very limited even after the rebuttal response. Therefore, I keep my rating.

We thank the reviewer for their time and careful consideration of our response. We appreciate their feedback and will continue to improve the clarity of our contributions and their distinction from established results in the revised manuscript.

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

This paper proposes a control consistency method for diffusion bridges that uses finite-horizon of noise directions collected over a time window to control the diffusion bridge. The method builds upon a prior work on Control Consistency Diffusion Bridges (CCDB), but replaces singular pointwise recovery with a finite-horizon controller. The method is theoretically analyzed. The paper provides an experimental evaluation where the method is compared against the singular pointwise recovery baseline.

Critical strengths

Unfortunately, I find the paper highly inaccessible so I cannot comment on technical contributions or soundness. My concerns are about the clarity and accessibility of the paper, see Weaknesses below.

Critical weaknesses

I find this paper extremely inaccessible to anyone not working on SDE theory or diffusion bridges, and could not follow its technical contributions. Even the sentences in the abstract and the introduction are dense and use a lot of highly technical terminology that is never defined. This does not improve in later sections either, see e.g. the preliminaries where notions such as "endpoint conditioning", "control consistency", "state-space covector" are used without definition. Moreover, the word "geometry" is used frequently and in many different contexts without any definition. All these terms are extensively used throughout the whole paper.

My background is in stochastic control and probabilistic system analysis. Hence, while I am not an expert on SDE theory, I would still expect to be able to follow at least the main arguments and high level picture, which I have not been able to do for this paper. This may partly reflect my lack of expertise, but it is also a sign that this paper is highly inaccessible to much of the ICLR audience.

Questions for authors

See the Weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer qDyC

Response to Reviewer qDyC

We sincerely thank the reviewer for the careful reading and candid assessment. We particularly appreciate the reviewer's transparency regarding their confidence and decision to inform the Area Chair. We fully respect this judgment.

More importantly, we take the concerns about accessibility very seriously. The reviewer's background in stochastic control and probabilistic systems makes this feedback especially valuable: the central ideas of our paper should be understandable to readers beyond specialists in SDE theory. We recognize that the original presentation did not sufficiently achieve this goal.

We have therefore undertaken substantial revisions throughout the manuscript, rather than making only isolated terminological changes.

1. Clarifying terminology and motivation

We have rewritten the Abstract and substantially revised the Introduction to introduce the main concepts more gradually.

In particular, the Introduction now explains:

  • Instantaneous diffusion geometry: the noise directions supplied by the diffusion matrix G(t,x)G(t,x) at a single time.

  • State-space covector: a vector u(t,x)u(t,x) with one component for each state coordinate, related to the executable Brownian-space control through v(t,x)=G(t,x)⊤u(t,x)v(t,x)=G(t,x)^\top u(t,x).

  • Pointwise recovery: reconstructing the covector from the Brownian-space control using the diffusion matrix at a single time.

  • Finite-time controllability: the ability of the dynamics to propagate noise into state directions that are not directly forced.

We also added a concrete kinetic example: in underdamped dynamics, noise acts directly on momentum, while dqt=pt dtdq_t=p_t\,dt propagates its influence into position. This illustrates why information unavailable instantaneously may become accessible over a finite interval.

These explanations are intended to replace unexplained geometric terminology with concrete descriptions of the underlying dynamical mechanisms.

2. Improving the conceptual structure

Section 3 now introduces diffusion bridges from two complementary perspectives: terminal reweighting of path measures and stochastic control. It defines the reference and conditioned path measures, explains endpoint conditioning, and introduces the relationship between state-space covectors and Brownian-space controls before these concepts are used in the method.

Section 4.1 now explicitly introduces the self-consistency principle inherited from CCDB before stating Theorem 4.1. We clarify that this identity remains valid under rank deficiency; the difficulty lies instead in recovering the state-space covector from an instantaneous Brownian-space control.

Section 4.2 then develops the finite-horizon recovery mechanism step by step: transporting noise directions to a common reference time, collecting them in a Gramian, and solving a regularized least-squares problem.

Section 5 explicitly separates two questions: whether the dynamics make missing state directions identifiable over time, and whether the reconstructed executable control approaches the optimal control.

We have also clarified the attribution of established mathematical results, distinguishing classical controllability and stochastic-flow identities from the specific recovery and learning construction proposed in FBC.

3. Making the central contribution easier to understand

The revised manuscript emphasizes a simple distinction:

The control-consistency identity itself does not fail under rank-deficient noise. What fails is the ability to recover the full state-space covector from the control at a single time.

FBC addresses this recovery problem by combining noise information across a finite time window. The dynamics can transport directly forced noise directions into otherwise inaccessible state coordinates, allowing the recovery step to use information unavailable instantaneously.

This is the central methodological contribution, and we have made it explicit in the Introduction and Sections 4–5.

4. Appreciation and further feedback

We recognize that improving accessibility is not simply a matter of adding definitions. A technically correct presentation should also communicate its motivation, main argument, and contribution without requiring readers to reconstruct the conceptual connections themselves.

We have made a substantial effort to address these issues while preserving the mathematical results. We fully respect the reviewer's assessment of their own expertise and the AC's discretion regarding how this review should inform the final decision.

Regardless of the reviewer's preferred role in that decision, we would be genuinely grateful for any further comments on the revised manuscript, including feedback focused solely on clarity, accessibility, or presentation. Such feedback would be valuable for improving the paper for the broader ICLR audience.

We sincerely thank the reviewer again for bringing these issues to our attention.

Reviewer_4Rh2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper extends the method of [1] to the setting where the noise is rank-deficient. This was noted by [1] as an important future direction, as it enables applications in areas such as molecular dynamics. To do so, the authors show that the self-consistency identity from [1] also holds in the rank-deficient case too, and can be used to design an algorithm for learning the diffusion bridge that parameterises the Brownian-space control and regresses onto targets constructed using a Gramian matrix. The authors also provide complementary theoretical results regarding when the missing state-space information can be recovered and how accurately the control can be reconstructed. The method is evaluated on synthetic diffusion bridges and molecular transition path sampling.

[1] Howard et al 2025, Control Consistency Losses for Diffusion Bridges

Critical strengths

Extending the fixed-point style approach of [1] to hold in cases where the ellipticity condition does not hold was noted in as an important question. The proposed approach is sound, as far as I can tell. The results indicate that the method works in settings with rank-deficient noise, which previously was not the case for the method in [1].

Critical weaknesses
  • I find that the presentation of the paper is quite poor. Even as someone familiar with the field, I struggle to follow it. The paper contains a large amount of technical jargon that in my opinion is not introduced sufficiently well, and phrasing and structure that is unnatural and difficult to parse.

    • For example, in only the first few sentences the paper says phrases like 'instantaneous diffusion geometry', 'pointwise recovery', and 'finite-time accessibility'. I do not think it is clear what these phrases mean in this context, despite being familiar with the literature myself. I certainly think they would be confusing to a reader unfamiliar with the topic.

    • Related to this, the authors seem to have written the paper assuming that the reader is already familiar with the 'control consistency' framework from [1] that they are extending. The authors first state the self-consistency identity in Theorem 4.1, with very little introduction explaining its background or motivation. The presentation should proceed along the lines of 'In recent work [1], the authors derived the following self-consistency identity that holds for the optimal control of the conditioned diffusion process', before stating the result and clearly indicating it is an extension of the identity from [1].

  • The main contribution of the paper is to extend the control consistency framework to cases where the noise is rank-deficient. However, I find that the experimental validation in this setting could be more extensive. For example, the paper would benefit from displaying some visualisations of the rank-deficient examples in the paper, rather than the overdamped ones currently displayed in Figure 3. Currently the assessments in the underdamped molecular dynamics examples are quantitative and lack visualisations. However, it can be possible for trajectories to 'tunnel through' the potential and get reasonable looking quantitative results despite not taking realistic paths. I think the results would be more convincing if the obtained paths were visualised and qualitatively compared to the known transition routes, as this would validate that the method is performing reasonably.

  • I have reservations over how much some of the presentation has been lifted directly from [1], with in my opinion not sufficient acknowledgement. For example, all the results in Table 1 (other than for the proposed FBC method) are taken directly from [1], as are the results in Figure 2 and Table 2, but as far as I can see this is not stated anywhere. Moreover, the formatting is often virtually identical - Table 1 is an exact copy of the table in [1] (the formatting of this table is quite unusual, which makes this quite a glaring example), and this is also the case for the three plots in Figure 2 and both plots in Figure 3. It is perfectly ok to use results from previous papers if sufficiently acknowledged, but this is not done so here, and the widespread use of the exact formatting of [1] is in my opinion a bit over the top.

Questions for authors
  • The control consistency procedure in [1] relies on both satisfying the consistency property, and also imposing an additional 'terminal condition' (which in the case of a diffusion bridge is jumping to the endpoint in the final step). I am unsure where in the procedure described in Algorithm 1 the analogous terminal condition is imposed - can the authors please clarify.

  • Could the authors visualise the transition path examples (for Muller-Brown, and alanine dipeptide on a Ramachandran plot) in the underdamped case? This would be valuable as a qualitative assessment of whether the learned paths are behaving as expected (as it is known what routes the paths should roughly take), rather than merely reporting the quantitative values. It can be possible to 'tunnel through' the potential and get reasonable looking results despite not taking realistic paths, so this would help to verify that the proposed method is behaving reasonably.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4Rh2

Response to Reviewer 4Rh2

We thank the reviewer for the careful assessment and constructive suggestions. We have revised the exposition, clarified the terminal condition in Algorithm 1, added underdamped transition-path visualizations, and explicitly acknowledged the provenance of previously reported results.

1. Presentation and relationship to CCDB

We have substantially revised the Introduction and methodological presentation to improve clarity. The Introduction now explains instantaneous diffusion geometry as the noise directions supplied by G(t,x)G(t,x), and pointwise recovery as reconstructing a state-space covector from Brownian-space control at a single time. Section 3 introduces the reference and conditioned path measures and the control representation before developing the method.

Section 4.1 now explicitly introduces the self-consistency identity underlying CCDB [1] before stating Theorem 4.1. The theorem establishes that this identity remains valid without requiring full-row-rank diffusion. We clearly distinguish the inherited identity from FBC's contribution: recovering covector information unavailable through instantaneous noise directions by transporting Brownian directions over a finite horizon and solving a Gramian-based regression problem.

2. Terminal condition in Algorithm 1

We have made the terminal condition explicit in Section 4.3 and Algorithm 1. For full-row-rank hard-endpoint bridges, the prescribed endpoint XT=xTX_T=x_T is imposed at the final discretization step, while the preterminal consistency recursion retains its original initialization.

For bridges with a smooth terminal weight F=e−ΨF=e^{-\Psi}, we use the terminal covector gT=−∇xΨ(XT)g_T=-\nabla_x\Psi(X_T) to construct the terminal-adjoint control target in Eq. (15). This target is combined with the finite-horizon consistency target through Eq. (16), without changing the original backward recursion. Appendix C.4 further describes its use in the underdamped Double-Well experiment.

Thus, both hard-endpoint conditioning and soft-terminal supervision are explicitly incorporated into the training procedure.

3. Underdamped transition-path visualizations

We have added Figure 4, presenting learned transition paths for the rank-deficient underdamped Müller–Brown and constrained-underdamped alanine systems.

For constrained-underdamped alanine, the trajectories exhibit a C5-to-C7ax transition route qualitatively consistent with the Ramachandran pathways presented in Control Consistency Losses for Diffusion Bridges and Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers.

These visualizations provide the requested qualitative assessment of the learned transition routes and complement the quantitative results in Tables 2–3, including terminal accuracy, hit rates, and transition-path objectives.

4. Baseline attribution and presentation

We have revised the captions to explicitly acknowledge previously reported results. Table 1 identifies the non-FBC baseline values as taken from [1], while Figure 2 attributes the CCDB and NGDB baseline curves to [1]. Table 2 distinguishes previously reported full-rank baselines from the constrained-underdamped experiments and acknowledges the original PIPS and TPS–DPS results.

Figure 3 now explicitly acknowledges that its visualization follows the presentation style of [1]. We have also redesigned Table 1 with conventional grouped metric headings and boldfaced best reliable values, removing the original qualitative performance symbols.

These revisions clarify the provenance of the reported comparisons and improve the distinction between our experimental contributions and prior work.

Summary

The revised manuscript introduces the control-consistency framework more clearly, distinguishes FBC's finite-horizon recovery contribution from the identity inherited from CCDB, explicitly specifies terminal conditioning in Algorithm 1, adds the requested underdamped path visualizations, and clarifies the attribution and presentation of prior results.

In light of these revisions, we respectfully ask the reviewer to reconsider and increase the current rating.

Reviewer_4Rh2OpenReview ↗

Response to rebuttal

I thank the authors for their response, and several of their changes have improved the paper. I am a bit surprised by the added underdamped path trajectories visualised in Figure 4 (I would have expected that these should look smooth, see for example [1]).

[1] Seong et al 2024, Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers

Having read the other reviews too, my main concern remains the presentation of the work, which is shared by Reviewer qDyC. The authors' changes go some way to improving this, but in my opinion it is still extremely hard for a reader to follow. Improving this would in my opinion require a full rewrite which is beyond the scope of the rebuttals, hence I unfortunately cannot raise my score.

AuthorsOpenReview ↗

Response to Reviewer 4Rh2

We thank the reviewer for pointing this out. Upon further inspection, we identified an issue in the trajectory visualization and have revised Figure 4 accordingly. We will also continue to improve the overall presentation and readability of the manuscript.

We sincerely appreciate the reviewer's careful attention and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.