acceptodds
Under review as a conference paper at ICLR 2027

Reactive Flux Matching: Mechanism Discovery and Adaptive Sampling of Rare Events

Abstract

Path sampling methods generate ensembles of reactive trajectories connecting metastable states, but extracting mechanistic insight from these data remains nontrivial. We introduce Reactive Flux Matching, a framework that learns two complementary objects directly from reactive trajectory data: a current velocity whose streamlines trace the dominant reaction pathways, and a scalar potential , obtained from a weighted Helmholtz–Hodge decomposition of the reactive current, that serves as a data-driven reaction coordinate. Both minimize quadratic functionals over the reactive path ensemble, analogous to the flow matching loss in generative modeling, and require no knowledge of the underlying dynamics or stationary distribution. Unlike committor-based methods, and remain well-defined under projection onto non-Markovian collective variables, and their level sets in turn provide adaptive interfaces for improved sampling with enhanced sampling methods. Flux Matching is validated through the generation of current velocity trajectories, adaptive sampling, and rate constant calculations on molecular systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, generative modeling for science, learned vector fields, AI for molecular dynamics, rare event sampling, representation learning, collective variables, reaction coordinates, Helmholtz–Hodge decomposition, non-Markovian dynamics, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mYHUOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduce Reactive Flux Matching, which learns two objects from an ensemble of reactive trajectories by quadratic regression, without access to the dynamics or stationary distribution. First is the current velocity, the ratio of the reactive current to the reactive density, whose streamlines are used to visualize transition mechanisms. Second is the scalar potential obtained from a weighted Helmholtz–Hodge decomposition of the reactive current, which reduces to the logit of the committor for Markovian dynamics with detailed balance. The authors argue that, unlike the committor, both objects remain well-defined under projection onto non-Markovian collective variables. They use the potential as a collective variable for shooting point selection in TPS and for binning in weighted ensemble. Validation covers the Müller–Brown potential, alanine dipeptide and AIB9.

Critical strengths

Recasting the analysis of reactive trajectories as a flow matching problem is novel and well executed. Defining the reactive density and current through averages over the path measure yields a regression target that needs no dynamics, no stationary distribution and no committor, works with data from any path sampler, and does not depend on Markovianity. The observation that the loss for the potential is exactly the loss for the current velocity ties the two learned objects together cleanly, and the recovery of transition path theory in the detailed balance limit gives the framework a solid anchor. This is consequential because it opens rare event analysis to standard, scalable flow matching, and I expect it to be adopted and extended by both the ML and the molecular simulation communities.

Critical weaknesses

The advantages over committor-based methods are not tested. These advantages are the paper's main selling points over existing reaction coordinate methods, so this gap is central to my assessment.

(a) There is no committor baseline. A projected committor E[q∣ϕ(x)=z]\mathbb{E}[q \mid \phi(x) = z], the minimiser of the likelihood loss on shooting outcomes (Peters & Trout, 2006; Jung et al., 2023), could be an baseline with equally well-defined regression target under projection.

(b) The claim that uu and hh remain meaningful under non-Markovian projection is not checked against a ground truth.

(c) h=log[q/(1−q)]h = {\rm log}[q/(1-q)] is shown only for Markovian dynamics with detailed balance and DD equal to the true diffusion tensor. The molecular experiments (underdamped dynamics, D=ID = I) satisfy none of these conditions, and there is no committor histogram test, so the claim that the transition state is correctly identified in Figure 7–8 would be unsupported.

Questions for authors

(a) How does hh compare with a projected committor trained by likelihood maximization on the same TPS data, with the same features and architecture and a budget, both (i) as a reaction coordinate and (ii) when used for shooting point selection or weighted ensemble binning?

(b) Does the learned uu match the exact current velocity in a non-Markovian setting such as position-only underdamped Müller–Brown?

(c) What is the committor distribution of configurations on the hh level set for alanine dipeptide?

5 comments

We thank the reviewer for their constructive critique.

(a) Projected-committor baseline. The likelihood-based projected committor qˉ(z)=E[q∣z]\bar q(z)=\mathbb{E}[q\mid z] is well defined but is a different object: the hidden variables are averaged over equilibrium, it carries no current, and ρR≠ρ qˉ(1−qˉ)\rho_R\neq\rho\,\bar q(1-\bar q) under projection (Thm 2.7). The two also use different data. qˉ\bar q is fitted to the outcomes of all shots, successful or not, launched from configurations whose hidden variables are drawn from equilibrium. hh is learned from the reactive trajectories alone, and its loss is an average over the reactive ensemble rather than a Dirichlet form over equilibrium, which concentrates on the rarely visited transition region (Remark 2.6). In (c) below, σ(a(h−c))\sigma(a(h-c)) is a Peters–Trout model with hh as the reaction coordinate, in which the shooting data set only aa and cc. We are comparing its held-out likelihood on shooting outcomes with that of qˉ\bar q trained on the same outcomes, features and architecture, and the two as coordinates for shooting-point selection, and will report the results here.

(b) uu under non-Markovian projection. We ran the proposed test: 5,000 transition paths of underdamped Langevin dynamics on Müller–Brown (γ=10\gamma=10), with RFM trained on positions only. The reference is the position-only current velocity u(x)=ER[v∣x]u(x)=\mathbb{E}_R[v\mid x], obtained by binning the velocities along the reactive trajectories, with no Markov assumption (App. J.4, Table 3, Fig. 10).

Rel. L2L_2 (well-sampled bins)Cosine
learned ∇h\nabla h vs reference0.3090.980

Here the solenoidal part of the current is small, so uu and ∇h\nabla h nearly coincide; the residual error comes mostly from sparsely sampled bins near the states. The reference must be built from reactive paths: averaging the full-space objects over Maxwellian velocities gives zero current, because conditioning on reactivity changes the velocity distribution at fixed position. This holds at any friction, so the comparison with the equilibrium projection is discriminating even here. The new Thm 2.7 makes this precise.

(c) hh and the committor on alanine dipeptide. hh is defined up to an additive constant, and its relation to the committor depends on the tensor DD in the decomposition (8). Under detailed balance with D=κDtrueD=\kappa D_{\text{true}}, h=κ−1logit q+consth=\kappa^{-1}\mathrm{logit}\,q+\text{const} exactly, i.e. q=σ(a(h−c))q=\sigma(a(h-c)) with a=κa=\kappa (Eq. 16). On overdamped Müller–Brown, where κ=β=0.08\kappa=\beta=0.08, the fitted value is a=0.083a=0.083. For alanine dipeptide (underdamped dynamics, dihedral coordinates, D=ID=I) no exact relation holds, and we test the two-parameter form empirically: we obtained empirical committors of 368,617 conformations by 1000 shots each with Maxwellian velocities, fitted a,ca,c on half of them, and tested on the other half. The mean committor in each hh-bin follows σ(a(h−c))\sigma(a(h-c)) closely, with a maximum deviation of 0.046 (App. J.6, Fig. 11). We replaced the transition-state claims for Figs. 7–8 by this test.

Dear Reviewer, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns. In particular, we ran the two tests you proposed: the learned field matches a model-free ground truth for position-only underdamped Müller–Brown (cosine 0.98, App. J.4), and on alanine dipeptide the shooting committor of 368,617 conformations follows σ(a(h−c))\sigma(a(h-c)) to within 0.046 (App. J.6). We would be glad to clarify anything further before then. Thank you again for your review.

Following up on your point (a): we have now compared hh with the neural committor trained by likelihood maximization on shooting outcomes, with the same features and architecture, on alanine dipeptide. Both coordinates drive the same active-learning loop (two-way TPS, 300 shots per round, 30 rounds, retrained each round) at identical MD cost: the committor is trained on the outcomes of the loop's own shots, hh on its reactive paths.

(ii) Shooting-point selection. MD frames per decorrelated path, pooled (TPS acceptance), one run each:

Rounds 1–101–30vs uniform (1–30)
Uniform selection10046 (0.02)10046 (0.02)1×
hh2577 (0.10)1676 (0.13)6.0×
Committor (MLE)1238 (0.16)1055 (0.19)9.5×

(i) Reaction coordinate. Binomial NLL per shot of the round-30 models, both evaluated on the same data (hh through σ(a(h−c))\sigma(a(h-c)) with two fitted parameters; constant predictor 0.693):

Held-out shots (2,694)Empirical committors, 368,617 configurations
Committor (MLE)0.5270.173
hh0.6350.196
Floor (empirical committor itself)—0.126

Both learned coordinates improve on uniform selection by large factors. On these two tasks the committor, trained directly on shooting outcomes, is the stronger of the two: it samples 1.6× more efficiently than hh over 30 rounds and predicts shooting outcomes better. hh, learned from the reactive paths alone, comes reasonably close on both, and in addition provides the current velocity and the reaction mechanism, which the committor does not. The comparison is in the revised PDF (App. J.8).

As the discussion period closes on Nov 18, we would be glad to know whether our replies to your points (a)–(c) address your concerns.

Reviewer_uPWkOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

WriteThe paper introduces a method called “Reactive Flux Matching”, which learns two objects: a committor-like control and a velocity field they call “current velocity”, both regressed from existing transition paths. This addresses an issue the previous committor learning methods have swept aside, namely, that working in the full space of Cartesian coordinates of the system (which, for systems in water and scientifically relevant systems of study, can reach millions) is not feasible; a reduced set of coordinates is used. However, in a reduced coordinate space, the dynamics become non-Markovian because the missing degrees of freedom appear as a memory term, which violates the assumptions of these committor learning methods. They demonstrate their method on the Müller-Brown potential (overdamped and underdamped), alanine dipeptide and AIB9, including a classic TPS loop setup.

Critical strengths

WriteThe main strength is the aforementioned addressing of a major theoretical pain point in the literature of committor learning (and here committor-like quantity learning), and the general theoretical developments that lead up to this. However, some of these theoretical advances also appear in Lu et al. [1], and should be cited, namely Proposition 1.12 in Lu et al. [1] is very close, but still a few algrebraic operations away. [1] “REACTIVE TRAJECTORIES AND THE TRANSITION PATH PROCESS” JIANFENG LU AND JAMES NOLEN

Critical weaknesses

The main weakness is the substitution of the boundary terms in the loss, which moves the elegant variational principle to a substitution which is pragmatic but less sound and critically not fully explored in the paper, neither mathematically nor experimentally.

Second, the experiments ablating the number of paths to start the loop from ablate down to 300 paths, which for such a trivial system amounts to basically full sampling, in all the senses that matter.

Questions for authors

How does the cross-entropy surrogate change the learned hh? This should be checkable using by looking at how ∇h\nabla h compares to the FEM reference in scale and flux on the 2d MB potential. How many transitions does the method need to get a loop of the ground? Is the uu-loss weighted by path duration in the actual code? Or am I wrong in thinking it needs it? If not, please fix and quantify it.

6 comments

We thank the reviewer for their careful reading.

Lu and Nolen. Thank you for the reference, which we now cite. For Markov diffusions, Lu and Nolen give a rigorous pathwise construction of transition path theory: reactive trajectories follow a Doob hh-transform of the dynamics, and the reactive density and current are recovered from empirical averages over reactive segments. Our Def. 2.2 adopts this pathwise viewpoint, and our Prop. D.4 is their construction, now attributed to them. What is new here is that the pathwise averages become the definition rather than a consequence of the SDE. This has three effects: (i) we yield regression losses that need only path data; (ii) we define the potential hh through a weighted Helmholtz–Hodge decomposition; (iii) we apply to any path ensemble, including projected, non-Markovian dynamics, where the committor-based expressions no longer hold (new Thm 2.7).

Effect of the surrogate on hh. We ran the suggested check on Müller–Brown: same paths, ∇h\nabla h compared with the FEM reference V=β−1∇ logit qV=\beta^{-1}\nabla\,\mathrm{logit}\,q (App. J.5, Table 4).

Cross-entropy surrogateExact endpoint lossTheory
Scale aa in q=σ(a(h−c))q = \sigma(a(h - c))0.08300.0418β=0.08\beta = 0.08
a/β\quad a/\beta1.040.521
Flux ratio Φ[ρR∇h]/Φ[ρRV]\Phi[\rho_R \nabla h] / \Phi[\rho_R V], q=0.25q = 0.251.0351.6611
q=0.50\quad q = 0.501.0271.8681
q=0.75\quad q = 0.751.0721.9961
Rel. L2L^2 error of ∇h\nabla h vs. VV, 0.05<q<0.950.05 < q < 0.950.1570.8740
Mean cosine (∇h, V), 0.05 < q < 0.950.9900.9961

The target logit q\mathrm{logit}\,q diverges on ∂A\partial A and ∂B\partial B, so any loss with this minimizer asks the network to represent a singular function there. The surrogate caps the boundary gradients. With either loss the direction of ∇h\nabla h matches FEM, and the magnitude error is concentrated in a thin layer around the states. For use as a reaction coordinate or to place interfaces only the level sets of hh matter, and these are what the cosine measures. Magnitude enters through the flux, reported in the last column.

Paths needed to start the loop. Correctness does not depend on hh: the Metropolis–Hastings acceptance includes the ratio of shooting-point selection probabilities, so the sampled ensemble is exact whatever hh is. Efficiency does depend on hh. We reran the loop on alanine dipeptide from only 30 seed paths, with 30 rounds of 30 one-way shooting moves and hh retrained each round, and compared with uniform selection (3,000 moves). Values are pooled within blocks of rounds (± SEM for acceptance, Poisson for decorrelation counts).

MovesAcceptanceMoves / decorr.MD frames / decorr.
hh-flat, rounds 1–103000.62 ± 0.0217.6 ± 4.32377 ± 576
hh-flat, rounds 11–203000.64 ± 0.0312.0 ± 2.41523 ± 305
hh-flat, rounds 21–303000.73 ± 0.037.5 ± 1.2698 ± 110
Uniform30000.56 ± 0.0116.3 ± 1.22596 ± 191

In the first ten rounds, with hh learned from 30 paths, hh-flat selection is no worse than uniform; the gain then grows, reaching 3.7× fewer MD frames per decorrelated path in rounds 21–30 and 2.0× pooled over all 900 moves. The loop thus starts from a few tens of paths and becomes efficient after a few hundred moves.

Duration weighting of the uu-loss. In the code, minibatches are drawn uniformly over all frame triplets of all paths, so each path contributes in proportion to its duration. This is the correct weighting, since u=jR/ρRu=j_R/\rho_R is an average under ρR\rho_R, the density of time spent on reactive paths. Algorithm 1 wrongly showed per-path sampling of tt; it is corrected.

Dear Reviewer, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns. In particular, we have now run the sampling loop on alanine dipeptide from only 30 seed paths: the gain appears after about ten rounds and reaches 3.7× fewer MD frames per decorrelated path in rounds 21–30 (2.0× pooled). We have also compared the cross-entropy surrogate and the exact loss with the FEM reference, and clarified our contribution relative to Lu and Nolen. We would be glad to clarify anything further before then. Thank you again for your review.

Reviewer_uPWkOpenReview ↗

thanks for the answer

Thank you for the detailed response and the new experiments! The attribution, the duration-weighting fix, and the level-set comparison (cosine 0.99) resolve those points, and thus the main issues, for me.

Regarding the surrogate magnitude, Table 4 is hard for me to reconcile with the loss as written. Varying the bulk plus cross-entropy terms gives Robin conditions, n·ρR D∇h = −½σ(h)ρ∂A on ∂A (and the mirror condition on ∂B). The minimizer's flux would then be ½E[σ(h(z_0))] < ½ of the reference, not ≈1. Softplus is also not scale-invariant, so the surrogate's magnitude depends on the units of time and on D, while the exact minimizer simply rescales. I would therefore read a/β ≈ 1 as incidental rather than as a property of the method. Could you rerun the surrogate with time rescaled by 10 (or D = 10I) and report a/β, and state how the exact-endpoint run was trained?

Seed paths. Thanks for the 30-seed run. As I read it, the gain appears once the pool reaches ~330 paths, so the message is that the seed count matters little because a poor h falls back to roughly uniform selection. I find that convincing and would frame it that way in the paper.

Reviewer_VD22OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of sampling rare events in transition path sampling (TPS), where the goal is to sample lower-probability states that have a chance of appearing in transition paths between two distinct low-energy regions. The paper's approach is learning-based, and they proposed to learn the following quantities which they show sufficiently reproduce path distributions: the current velocity (current divided by reactive density), and the potential (defined through the Helmholtz-Hodge decomposition). The resulting TPS method is evaluated on MB, alanine dipeptide, and AIb9.

Critical strengths
  • The theory and math, to my reading, is all correct and well-formulated.
  • The approach to my knowledge of having learning focused on specifically current velocity to approach TPS is, to my understanding, novel.
Critical weaknesses
  • There are a number of other learning-based methods for transition path sampling that need to be compared against, eg [1][2][3][4], some of which only need to train on uncorrelated states rather than full trajectories. These need to be compared against fairly in their ability to find diverse, viable transition paths.
  • The scale of the molecules tested is, to my understanding, small relative to the claim of section 3.3 (since it just uses backbone dihedral angles of AIB9?).
Questions for authors

What is the main benefit of this particular learning-based TPS method compared to already published work [1][2][3][4]?

6 comments

We thank the reviewer for their careful reading.

Benefit compared with [1]–[4]. These methods and RFM solve different problems. [1]–[4] generate transition paths in the full state space: by stochastic optimal control (Holdijk et al., 2023), a variational Doob transform (Du et al., 2024; Lee et al., 2025), or minimization of the Onsager–Machlup action (Raja et al., 2025). To do so they need the force field and a Markovian model of the dynamics in the sampled variables, overdamped or underdamped Langevin with explicit transition kernels, and they rely on further approximations: Gaussian path families, added noise to make the diffusion invertible, or most-probable rather than typical paths. RFM does not generate paths. It extends transition path theory beyond the Markovian setting: from reactive paths alone, with no model of the dynamics, it estimates the current velocity uu, whose streamlines show the mechanism, and a potential hh that serves as a reaction coordinate. The two are complementary: paths produced by [1]–[4] are valid input for RFM, as are paths from TPS, TIS, FFS or weighted ensemble, which make no assumption on the dynamics. A comparison on the diversity of generated paths would therefore measure the sampler, not RFM. We now cite and discuss [1]–[4] in the related work.

Methods trained on uncorrelated states. The reactive current is a dynamical quantity and cannot be estimated from configurations alone, which is why RFM uses reactive paths. Methods that avoid trajectory data, such as [2]–[4], replace it with an explicit model of the dynamics; RFM uses the paths in place of that model.

RFM as an add-on. Applied to the output of an existing sampler, RFM also makes it more efficient: interface-based methods (TIS, FFS, milestoning) and weighted ensemble need a reaction coordinate to place interfaces or bins, and the level sets of hh provide one from data. On alanine dipeptide, selecting TPS shooting points uniformly in hh reduces the MD cost per decorrelated path by 2.6× (Sec. 3.2.2), and binning weighted ensemble in hh gives earlier first-passage events and faster convergence of the rate estimate than the standard dihedral coordinate (App. J.3).

Scale. For AIB9 we model the 29 backbone dihedrals, a projection of the 129-atom system; for alanine dipeptide we also train on full Cartesian coordinates. Modeling a subset of coordinates, where the projected dynamics is non-Markovian, is the setting the method is designed for (Thm 2.7), and one that samplers requiring a Markovian model cannot address. We agree Sec. 3.3 should state the claim in these terms and have revised it.

Dear Reviewer, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns. In particular, the revised paper now cites and discusses the learned transition path samplers [1]–[4], explains why RFM is complementary to them as an add-on to established samplers, and clarifies that the models are trained on 7 to 66 coordinates for alanine dipeptide and 29 for AIB9, the 2D maps being projections for display only. We would be glad to clarify anything further before then. Thank you again for your review.

Thank you for the response, and for including the mentioned references and discussion in related work. I also appreciate the difference between TPS and RFM, and appreciate the novelty (as far as I'm aware of) for a learning-based approach on these particular mathematical objects for RFM, and the flexibility it buys. Int his lens from this discussion, I do agree that [1,2,4] are less important as baselines, as your framework does not require explicit dynamics models.

But this does not exclude e.g. [3], which to my understanding only requires a diffusion model trained over states (not needing trajectory ordering). Even if in spirit this needs an underlying model of forces, and TPS and RFM are indeed different problems, a good learning-based TPS method can still be used for RFM.

Likewise, all tested datasets/domain have good force field models available, so if the downstream applicability is to use this method for RFM on similar environments, it should still be compared against other learning-based methods on overlapping tasks, even if they are TPS-focused. I agree that good RFM without relying on a force field model would be useful, but this method is not tested on any environments where the force field is hard to estimate.

I think this is an interesting method, my main point of contention right now is the lack of need to compare to any other learning-based MD methods.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful and constructive reports. Each reviewer receives a self-contained reply; this note summarizes the discussion.

What the reviewers valued.

  • Recasting the analysis of reactive trajectories as a flow matching problem, with a regression target that needs no dynamics, no stationary distribution and no Markov property, and works with data from any path sampler (mYHU, VD22).
  • Addressing a major theoretical pain point of committor learning: reduced coordinates make the dynamics non-Markovian (uPWk).
  • The link between the two losses, and the recovery of transition path theory in the Markovian limit (mYHU); a correct and well-formulated theory (VD22).

What the paper does. We define the reactive density and current directly as averages over reactive paths, and learn the current velocity uu and the potential hh from these paths by regression. This has three consequences.

  • Beyond transition path theory. TPT describes reactive trajectories through the committor, which is defined only when the dynamics is Markovian in the observed variables. Our objects exist for any path ensemble: for a Markov diffusion TPT is recovered exactly, and for projected or otherwise non-Markovian dynamics uu and hh remain exact where committor-based expressions do not apply (Thm 2.7).
  • Reactive rather than equilibrium averages. The standard variational principle for the committor is a Dirichlet form, an equilibrium average whose integrand is concentrated in the transition region, where the equilibrium density is smallest. Our losses are averages over the reactive ensemble, so every sampled frame is informative, and they use the reactive trajectories alone, with no record of the trials the sampler rejected (Remark 2.6).
  • An add-on to existing samplers. RFM is not a path sampler. It analyzes the output of TPS, TIS, FFS, milestoning, weighted ensemble or learned samplers a posteriori, and the level sets of hh provide interfaces, bins or shooting points that make these samplers more efficient.

What they questioned, and what we did.

ConcernReviewerResolution
Comparison with learned TPS samplersVD22These generate paths and need a Markovian model of the dynamics; RFM needs neither and takes paths from any sampler, including these, and accelerates it (2.6× on TPS, faster WE rates). Now cited and discussed; scope stated in the introduction.
Scale of AIB9VD22Training on 29 dihedrals of a 129-atom system is a projection, the setting the method targets (Thm 2.7). §3.3 reworded.
Overlap with Lu & NolenuPWkCited; their Doob-transform construction credited (Prop. D.4). New: regression losses, the potential hh, and the non-Markovian extension.
Cross-entropy surrogateuPWkCompared with the FEM reference: the direction of ∇h\nabla h, hence its level sets, is unchanged (cosine 0.99; App. J.5).
Duration weighting of the uu-lossuPWkThe code weights by duration, as required; Algorithm 1 corrected.
Paths needed to start the loopuPWkFrom 30 seed paths, the gain appears after ~10 rounds and reaches 3.7× fewer MD frames per decorrelated path in rounds 21–30 (2.0× pooled).
No ground truth under projectionmYHUPosition-only underdamped Müller–Brown against binned reactive velocities: cosine 0.97–0.98 (App. J.4, Fig. 10).
hh vs. committor; transition-state claimsmYHURelation q=σ(a(h−c))q=\sigma(a(h-c)) derived and its conditions stated (Eq. 16); tested by shooting on ADP (App. J.6, Fig. 11). Claims for Figs. 7–8 replaced by this test.
Projected-committor baselinemYHUE[q∣z]\mathbb{E}[q\mid z] averages over equilibrium and carries no current (Thm 2.7); comparison as a CV in progress.

New theoretical result. If z=ϕ(x)z=\phi(x) and the states are defined in zz, the reactive density and current in zz are pushforwards of those in xx, with the unobserved variables averaged over the reactive ensemble, not the equilibrium one (Thm 2.7). Projecting transition path theory objects with equilibrium averages is therefore biased; for underdamped dynamics observed in position it gives zero current. RFM learns the projected objects directly from the projected paths, so this conditioning is built in. The diffusion formula of the submission becomes Cor. E.1, in which we corrected a missing term.

Revised paper. The revised PDF (10 pages) is uploaded. It contains Thm 2.7, Remark 2.6 on reactive vs. equilibrium averages, the rewritten introduction, the Lu–Nolen attribution, the corrected Algorithm 1, and the three new experiments (App. J.4–J.6).

Still in progress. The projected-committor baseline on alanine dipeptide and the low-seed sampling loop, to be posted as follow-up comments before November 18.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.