acceptodds
Under review as a conference paper at ICLR 2027

Backward Kolmogorov Transport: Sampling Invariant Laws of Reversible Diffusions from Trajectory Data

Abstract

We consider a reversible diffusion with unknown drift, observed through trajectories, and construct samplers of its invariant law. The trajectories contain samples of this law, but a sampler built from them by density or score estimation inherits the errors of these estimates, in particular on the barriers between metastable states. The trajectories also contain the dynamics, and for a reversible diffusion the ratio between the law of the process and its invariant law solves the backward Kolmogorov equation, so that the eigenpairs of the generator propagate it in closed form. Backward Kolmogorov transport (BKT) evaluates this ratio with eigenpairs estimated from the trajectories and coefficients computed once from the initial particles, and moves the particles along the Wasserstein gradient flow of the Kullback-Leibler divergence. The particles do not interact, BKT defines a transport map, and no density, score or drift is estimated. With exact eigenpairs BKT transports the spectral projection of the initial law exactly when this projection is positive. For estimated eigenpairs we prove a transport identity in which truncation enters only through the initial law and estimation only through the residual of the eigenpairs, and a local stability bound in the 2-Wasserstein distance. With coefficients computed from the transported particles, the sampling error of the retained modes cancels, and a central limit theorem identifies the asymptotic variance. Experiments on Ornstein-Uhlenbeck processes, multi-well potentials, separable products and alanine dipeptide, including comparisons with particle methods based on density estimates, agree with these results, and with an accurate spectrum the errors of BKT reach or fall below those of independent samples.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
particle-based sampling, Wasserstein gradient flow, Koopman operator, probability flow ODE, Langevin dynamics

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_phtdOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript proposes a method for an interesting setup of the sampling problem, that assuming the landscape (e.g., a potential function) is not known but given trajectory data derived from the overdamped Langevin equation driven by the potential function, how one could leverage the trajectory data to perform additional sampling task. The theoretical core involves (1) the identification of the weight function with respect to the invariant measure (i.e., density ratio) evolves the Kolmogorov Backward Equation (KBE), cleverly leveraging the reversibility condition of the process, (2) use the standard Koopman spectral analysis for performing eigen-analysis for making prediction, and (3) to generate new samples, using the property that a Kullback-Leibler divergence-driven Wasserstein gradient flow and the forward diffusion share the same time-dependent marginal distribution to enable a relaxation and flow-based model to transport samples from an a priori defined prior distribution.

Critical strengths
  • The setup of the problem is interesting and definitely relevant to more physical sciences-driven applications, such as molecular dynamics.

    • The setup carefully dissociates the generation of the "data" (trajectories) from the sampling algorithm. This is relevant in practice. Most generative modeling-based sampling methods require in-loop evaluation of the potential function; for complex systems, such an intrusive approach is time-consuming and could waste compute in repetitively sampling the same region in the state space.
  • The theoretical core linking standard results - the connection between KBE and the density ratio evolution under the reversibility condition and approximate Koodman analysis - together is clever and insightful from the perspective of stochastic processes. Turning this connection to explicit flow-based particle transport with error analysis is useful.

Critical weaknesses
  • Despite the theoretical clarity, the reviewer remains skeptical about the practical aspects of the proposed methods, especially in well-separated multimodal settings.
    • The basic idea is to estimate the velocity field of the effective KL-Wasserstein flow from the trajectory data. For multimodal distributions, obtaining representative stationary data from trajectories can require simulation times comparable to or longer than the characteristic inter-basin transition times. Even with stationary sampling, the probability of observing the transition regions can become exponentially small as mode separation or barrier height increases. It is hard for the reviewer to believe that this method can accurately resolve rare-event-driven relaxation under practical data and computational budgets, and such fundamental limitations were acknowledged but not adequately resolved or quantified in the manuscript.

      • The reviewer tried to use the provided codebase to perform the following, idealized problems, supplying independent exact target samples for spectral estimation to give the method favorable training coverage. The reviewer invites the authors to consider these analytically tractable problems to address the limitation of their methods.

        • [Mode discovery.] Consider R1\mathbb{R}^1, Gaussian mixture target with density ∼12N(0.9,σ2=0.012)+12N(−0.9,σ2=0.012)\sim \frac{1}{2} \mathcal{N}(0.9, \sigma^2 = 0.012) + \frac{1}{2}\mathcal{N}(-0.9, \sigma^2 =0.012) and assume initial μ0=N(0.9,σ2=0.012)\mu_0 = \mathcal{N}(0.9, \sigma^2 = 0.012). This is to mimic that the Langevin dynamics has already discovered some local mode at x=0.9x=0.9, and the question we ask is if the method can reasonably populate the other mode at −0.9-0.9 by transporting particles initially concentrated around 0.90.9.
        • [Weight equilibration.] Again consider the Gaussian mixture target with density ∼1011N(−0.9,σ2=0.012)+111N(0.9,σ2=0.012)\sim \frac{10}{11} \mathcal{N}(-0.9, \sigma^2 = 0.012) + \frac{1}{11}\mathcal{N}(0.9, \sigma^2 =0.012), but now with an initial mixture target with density ∼12N(0.9,σ2=0.012)+12N(−0.9,σ2=0.012)\sim \frac{1}{2} \mathcal{N}(0.9, \sigma^2 = 0.012) + \frac{1}{2}\mathcal{N}(-0.9, \sigma^2 =0.012). The question we ask is if the method can reasonably recover the target population weights of these two modes, which encode their free-energy difference in physical sciences.

        With the limited review time, the reviewer could run the code, but the results were substantially inaccurate in the tested configurations. For example, in the discovery problem, the estimated spectral gap was approximately 6.8×1066.8\times10^6 times the reference value, and no inter-basin crossings were detected even at ten times the reference mode-to-mode mean first-passage time (a recent paper provided the analytical formula based on mean first-passage time analysis.) In the equilibration problem, a few crossings occurred, but the left-basin fraction remained approximately 49%, rather than the target 90.9%, even at ten reference relaxation times. Particle collapse was also observed. Enforcing the constant eigenfunction exactly did not resolve these failures. These tests used the S regime; long-horizon integration used a stiff solver with tolerance checks and an early-time comparison against the supplied RK4 implementation. The reviewer cannot claim that the tested dictionary and configuration were optimally tuned.

    • Theoretically, such a rare event-driven transition can require extremely long time.

      • Because the exact KL-Wasserstein gradient flow tracks the time-dependent marginal distribution of the diffusion process, the resulting particle dynamics must transport samples at the corresponding timescale. The reviewer questions whether accurate numerical particle transport across fast intra-basin relaxation and slow inter-basin redistribution can be achieved at practical computational cost without problem-specific adaptation. The long physical timescale alone does not establish that numerical integration must be prohibitively expensive.
      • An additional problem is that the estimated spectral gap λ^1\widehat{\lambda}_1 could be way off for these systems.
      • Numerically, accurately estimating the eigenpairs by projecting onto a finite set of a priori selected basis functions is notoriously difficult and sensitive in high dimensions. On the one hand there is finite-sample spectral estimation error. On the other hand, for well-separated metastable basins, the first nonconstant eigenfunction is approximately constant within each basin, connected by a narrow transition layer near the separating barrier. It is difficult for the reviewer to believe that the chosen radial basis functions - furthermore, not even with fully data-driven basis functions (meaning, solving the inverse of the full N×NN\times N Gram matrix, NN: number of samples), since their centers are placed on fixed spatial grids - can reliably resolve φ1\varphi_1 and its derivatives in this regime without a demonstrated choice of grid space, bandwidth, and regularization.
Questions for authors
  • Could the authors discuss the limitations of the method, particularly for the rare-event-driven processes described above? Clarifying the regime in which accurate spectral estimation and particle transport are practically achievable, and diagnostics that can signal unreliable spectral estimates or transport results, would materially affect the reviewer’s assessment.

  • What basis functions and parameter-selection strategies would the authors recommend for high-dimensional problems (where approximating eigenfunctions and their derivatives is known to be challenging)? What practical diagnostics can detect inaccurate eigenpairs or incorrect transported basin weights, even when the spectral coefficient diagnostic indicates convergence?

  • The unweighted Dirichlet-form estimator requires representative equilibrium sampling. What trajectory-design or coverage assumptions are required to recover the relative equilibrium weights of different basins? In particular, if the available trajectories remain confined to one basin, what information allows the method to determine the population of an unvisited basin? Can the proposed diagnostics detect this lack of information? How does the unweighted Dirichlet-form estimator accommodate trajectories initialized out of equilibrium?

  • The main spectral estimation appears to depend only on the empirical distribution of trajectory frames, with temporal information used separately for lag-consistency checks. Could the authors clarify whether the central task is learning dynamics from trajectories or estimating the generator of a reversible diffusion with prescribed diffusivity from stationary samples? What role does temporal information play beyond these checks?

4 comments

Part 1/2. (Weakness, Q1)

We thank the reviewer for the most demanding review we received. Running our code on two tests is the necessary validation the paper needed. In the revised PDF, the changes we made for this review are in magenta and the changes we made on our own are in blue. The line numbers below are those in the margin of the PDF.

First we want to clarify our position in brief. The review is right on three points, and the paper now reflects them. First, the submission named the limitation at high barriers without quantifying it. Second, the diagnostic χ^2\hat\chi^2 does not detect wrong eigenpairs. Third, the Dirichlet-form estimator uses the data through the stationary samples, so the submission overstated the role of the dynamics (Q4). On one point we reach a different conclusion. The two tests fail in the estimation of λ1\lambda_1 and φ1\varphi_1, which no method can obtain from data without samples in the transition layer, whereas the transport itself works across these time scales at a small cost.

Regarding to the two tests, for s2=0.012s^2=0.012 the barriers are 33.133.1 and 34.234.2 in units of the temperature, and the two-state formula gives λ1=7.5⋅10−13\lambda_1=7.5\cdot10^{-13} and 1.3⋅10−121.3\cdot10^{-12}. About four of 2⋅1052\cdot10^5 exact samples fall in ∣x∣<0.45|x|<0.45, and none near the saddle. We reproduced both tests (lines 1559–1565 and 1586–1590).

  • With eigenpairs estimated from the samples, λ^1\hat\lambda_1 is too large by factors between 6⋅1036\cdot10^3 and 1.1⋅1051.1\cdot10^5. No mass reaches the other mode in (i), and the left-well mass stays at 0.500.50 in (ii). This is the outcome of the reviewer's runs, including the collapse of the particles, which pile up at sinks of the estimated field.
  • With the exact two-state eigenpair, BKT carries half of the mass across the barrier in (i) (0.5020.502 for a target of 0.50.5) and recovers the weight in (ii) (0.9080.908 for 0.9090.909), with about 100100 accepted steps per particle over a horizon of order 101310^{13} and without collapse.
  • With the exact eigenfunction and the eigenvalue 106λ110^6\lambda_1 the failure reappears, with left-well masses 0.0000.000 in (i) and 0.490.49 in (ii).

The transport therefore works at 3333 units, at a cost of about a hundred steps per particle, and the failure comes from the estimate of the slow eigenpair.

Now we want to explain why a longer integration cannot help. By the new Remark 4.3 (lines 294–299), wrong eigenvalues do not simply rescale time. If the eigenvalues are too large by a factor κ\kappa, the particles follow the field vκtv_{\kappa t} and move κ\kappa times slower than the exact flow, so the field has decayed before the mass has crossed. In the reviewer's discovery test the estimated gap was about 6.8⋅1066.8\cdot10^6 times the reference value. The estimated field was therefore negligible after about a millionth of the reference time scale, and integrating up to ten mean first-passage times could not move any mass.

We also want to clarify the limit is in the data. The estimators are functions of the sampled states. If no sample falls in the transition layer, the barrier height is not determined by the data, for our estimator and for any other (lines 1538–1539). This is the limitation the review describes, and no change of the transport can remove it. The scan below quantifies it, which the submission had not done.

Q1. The regime in which estimation and transport work. For V=a(x2−1)2V=a(x^2-1)^2, a∈{1,2,3,4,6,8,10}a\in\{1,2,3,4,6,8,10\}, we used the data protocol of the paper (2⋅1052\cdot10^5 pairs from 500500 chains, lag 0.10.1) over ten realizations (Table 9, lines 1541–1557).

  • With FD eigenpairs BKT reaches the reference level in 139139 of 140140 runs up to a=10a=10, where λ1=7.8⋅10−4\lambda_1=7.8\cdot10^{-4}.
  • With estimated eigenpairs it reaches the reference level for a≤4a\le4 (at least 16501650 observed transitions), in six of ten realizations for a=6a=6 (about 330330 transitions) and in none for a≥8a\ge8 (at most 6767 transitions).

The practical range is therefore set by the number of observed transitions, a few hundred here, rather than by the relaxation time. This agrees with the reviewer's analysis of the data requirement and with McBride et al. (2026), which Section 5 now cites (lines 501–509).

Regarding to the cost across time scales, the concern that the particles must be integrated at the cost of the slow time scale does not apply to BKT. The field is explicit in tt and the equations of the particles are decoupled, so each particle is integrated with its own adaptive step in physical time (lines 1547–1549). A particle takes small steps while it crosses a barrier and large steps otherwise. This means 2020 to 4848 accepted steps on average in the scan and about 100100 for the exact pair at 3333 units, where the relaxation time is 10910^9 times longer than at a=10a=10.

Part 2/2. (Q1 to Q4)

About diagnostics (Q1, Q2). The reviewer is right that the diagnostic χ^2(t)\hat\chi^2(t) of the submission can indicate convergence while the eigenpairs are wrong. It measures only the decay of the retained coefficients, and the paper now says so. The revision adds four checks, computed from the data and the output (lines 1592–1604).

  1. Transitions. The trajectory contains at least 16501650 transitions between the wells where the estimated eigenpairs work in the scan, and at most 6767 where they fail with stationary data.
  2. Stationarity. The drift of the basin occupations between the two halves of the trajectory is at most 0.0050.005 on average for chains started in both wells, and +0.087+0.087 and +0.039+0.039 at a=6a=6 and 88 for chains started in one well.
  3. Generator and lag eigenvalues. Their relative difference for a retained mode, computed from the same pairs, vanishes for exact eigenpairs. Its median is at most 0.030.03 where the estimated eigenpairs work and above 0.10.1 where no run reaches the reference level.
  4. Output and data. The basin masses of the output differ from those of the stationary data by at most 0.0120.012 in the runs that reach the reference level and by 0.210.21 to 0.510.51 in the runs at a≥8a\ge8.

Q2. We agree that a fixed grid of basis functions does not carry to high dimension. In three coupled dimensions a tensor grid with the per-axis resolution of the two-dimensional experiments reaches the reference level, and a coarser one does not (Appendix D.5). In higher dimension the eigenpairs should come from kernel estimators, with the full Gram matrix or a Nyström approximation, or from neural estimators, and the bounds of Section 4 transfer their errors without change. As for the choice of functions, bounded ones matter, since unbounded polynomials are unstable in the tails, and check 3 selects the retained modes from the data, as in the ten-dimensional example (Appendix D.2).

Q3. The data must sample the basins with their equilibrium weights and the transition layers between them. If the trajectories stay in one basin, no information determines the population of the other, for BKT or for any other method, and BKT then samples the law of the data. The scan with all chains started in the right well shows this (lines 1551–1557). The left-well fractions of the data are 0.310.31, 0.090.09 and 0.020.02 at a=6a=6, 88 and 1010, and BKT returns 0.300.30, 0.070.07 and 0.020.02. Checks 1 and 2 detect this case, and for trajectories started out of equilibrium the reweighting of Wu et al. (2017) and Nüske et al. (2017) applies. Section 2 now states the data assumption (lines 155–160).

Q4. The reviewer's reading is correct. For a reversible diffusion with known diffusion matrix the generator is determined by π\pi, and the Dirichlet-form estimator of the paper uses the data only through the sampled states. The time order enters through checks 2 and 3, through estimators built on pairs (X0,Xτ)(X_0,X_\tau), which BKT accepts without change, and through the reweighting of non-stationary data. The central task is to sample π\pi through the spectrum of the generator, estimated in weak form and without a density estimate. The abstract, the introduction, Section 2 and the conclusion have been reworded accordingly (lines 16, 58–65, 155–160 and 515).

For your convenience, we list the changes (magenta) made:

  • Abstract, introduction, Section 2 and conclusion, the role of the dynamics and the data assumption (lines as above).
  • Remark 4.3, wrong eigenvalues (lines 294–299).
  • Section 5, the paragraph "High barriers" (lines 501–509).
  • Appendix D.6 with Table 9 (lines 1536–1604).
  • Appendix D.7, the limitations (lines 1606–1614).

We thank the reviewer again. The two tests led to Remark 4.3 and to the four checks, and we are glad to run any further test the reviewer finds useful.

The reviewer thanks the authors for their substantial revisions, additional experiments, and candid discussion of the observed failures. The reviewer will increase the score in recognition of these improvements and the clearer specification of the method’s limitations. The reviewer continues to appreciate the idea of using data from reversible processes to estimate the backward Kolmogorov generator and construct an explicit sampling flow.

The additional controls help distinguish spectral-estimation errors from numerical integration costs. In particular, the reviewer accepts that an extremely long physical relaxation time does not necessarily imply prohibitively expensive adaptive integration when an accurate velocity field is already available. The clarification that the principal Dirichlet-form estimator uses stationary sample locations and known diffusivity is also helpful.

The following points would further strengthen the manuscript. The reviewer is putting them here as a public record because everything will be publicly available regardless of the acceptance/rejection/withdrawn decision of the manuscript; this is to provide different aspects for future reviewers, practitioners, and even LLM's.

  1. Relationship to McBride et al. (2026). The reviewer appreciates the added citation. A more detailed discussion would help distinguish the metastable relaxation timescale inherited from the diffusion, the computational cost of integrating an explicitly available spectral field, and the cost of acquiring data and estimating that field accurately. The new adaptive-integration results address the second but not the first and third.

  2. Data acquisition and practical applicability. The revised coverage assumptions and barrier-height scan substantially clarify the requirements. However, the Gaussian-mixture tests show that even independent exact equilibrium samples, with correct basin coverage and weights, can leave the transition region insufficiently resolved and lead to unsuccessful transport. Could the authors discuss how practitioners should proceed in this situation, and more importantly, how does the practitioner know that they are in this situation?

  3. Spatial eigenfunction approximation near the saddle. This remains the most important technical clarification. One has very few samples near the saddle, so constraining the eigenfunction there may be difficult. The failure with independent exact equilibrium samples in the Gaussian-mixture tests shows that equilibrium sampling alone does not ensure an accurate spectral estimate, particularly of the transition-layer profile and its derivatives. Accurate eigen analysis in the basins alone need not imply an accurate velocity field.

  4. Interpretation of the diagnostics. The four additional checks are useful empirical warning signs, but their limitations should also be stated. The reviewer remains skeptical. Stable occupations can occur in a trajectory confined to one basin, and agreement between output and training occupations can reproduce biased training data. Generator–lag agreement on the same finite dataset likewise does not establish coverage of unvisited regions. Please distinguish detecting failures in the reported experiments from certifying reliable sampling more generally. The observation that a few hundred transitions suffice should also be qualified as specific to the tested configurations.

  5. Basis selection beyond the demonstrated examples. The coupled three-dimensional experiment is helpful. The remaining question is how to select basis resolution, bandwidth, regularization, and retained rank when reference eigenpairs are unavailable, particularly to resolve transition-layer derivatives. Kernel and neural estimators are plausible alternatives, but their availability alone does not establish that they resolve this difficulty.

We thank the reviewer for deciding to increase the score and for stating the remaining points so precisely. We agree with all five, and the revised PDF states them in magenta.

1. We agree that the three quantities must be kept apart and that our adaptive-integration results address only the second. BKT follows the Fokker-Planck path and inherits its time to equilibrium, of order 1/λ11/\lambda_1. With an explicit field this time is a parametrization of the path. In the variable e−λ1te^{-\lambda_1t} a particle in a well moves along the gradient of the mean first-passage time to the other well, the measure of slowness of McBride et al., at a velocity of order one. The number of adaptive steps therefore grows only linearly with the barrier height. The exponential time returns in the data. The diffusion that they analyze as a sampler is the process that generates our trajectories, and in the scan these had to be several hundred times longer than 1/λ11/\lambda_1. We therefore reach their conclusion. The field of BKT contains nonlocal information, the slow eigenpairs, and BKT moves the cost of metastability from the transport to the data without removing it (Remark 5.1, lines 510–521, and Appendix E).

2. This case can be recognized by counting the samples between the wells along a coordinate that separates them, or the transitions in a trajectory. In the scan the estimated eigenpairs work with about 30003000 or more samples in ∣x∣<0.3|x|<0.3 and fail with about 120120 or fewer, and the mixtures have about four. Exact samples still determine the weights of the wells, but not λ1\lambda_1 or the profile of φ1\varphi_1 across the layer, which BKT needs, and no choice of the dictionary supplies them (lines 1637–1648). Samples in the layer must then be added. Appendix E reports a test with reweighted data from a simulation with a lowered barrier (lines 1749–1770). With these data λ^1\hat\lambda_1 is within 6%6\% of λ1\lambda_1, and 99, 66 and 33 of ten runs reach the reference level at a=6a=6, 88 and 1010, compared with 55, 11 and 00 for unbiased trajectories.

3. We agree that this is the central point (lines 1623–1636). A function with the right constants in the wells is close to φ1\varphi_1 in L2(π)L^2(\pi) whatever its profile in the layer, so accuracy in the basins says nothing about the field. The energy norm of our bounds does see the layer, where the energy of φ1\varphi_1 is concentrated, but the estimators replace it by an average over the samples, which sees the layer only through the samples that fall into it. Theorem 4.5 requires more, since the paths of the particles through the layer must stay in the validity region, whose conditions are pointwise. The basins matter as well, because the particles move with ∇φ^1/λ^1\nabla\hat\varphi_1/\hat\lambda_1, a ratio of two quantities of order λ1\lambda_1 there. In the biased test an eigenvalue that is 22 to 6%6\% too large alone lowers the left-well mass from 0.500.50 to 0.490.49, and at a=10a=10 the estimated eigenfunctions limit the result although the layer contains about 1.7⋅1041.7\cdot10^4 samples.

4. It's correct and the paper now calls the checks warning signs that do not certify the output (lines 1650–1654). For the chains started in the right well at a=10a=10 the occupations in Table 9 are stable, with a drift of +0.008+0.008, and the output reproduces the weights of such data, with 0.300.30 for 0.310.31 at a=6a=6. Agreement of the generator and lag eigenvalues says nothing about regions without data. The checks also miss failures within the scan, since at a=6a=6 they do not separate the four realizations that miss the reference level from the six that reach it. The few hundred transitions are specific to the tested configurations, as the paper now says.

5. We acknowledge that the paper gives no general rule and it now says so (lines 1656–1666). Ranking dictionaries by their estimated eigenvalues is not sufficient. In our three-dimensional example the slow eigenvalues of the coarse dictionary are within 3%3\% of the reference values, and yet the output keeps 0.260.26 of the mass in the initial well where the finer dictionary gives 0.210.21. Refining the dictionary until the output no longer changes shows this without a reference, but it sees a transition layer only through its samples. The same holds for kernel and neural estimators, which avoid the exponential growth of a tensor dictionary with the dimension but are fitted to the same samples (line 498).

Changes made (in magenta).

  • Section 2, the data must also sample the transition regions (lines 155–160).
  • Section 5, the sentence on kernel and neural estimators (line 498) and Remark 5.1 (lines 510–521).
  • Appendix D.7, rewritten (lines 1612–1666).
  • Appendix E, new, with Table 10 (lines 1668–1777).

We thank the reviewer again. These comments have made the limits of the method clearer and new paper is updated.

Reviewer_W8mQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

: The paper proposes a method for sampling the invariant distribution of a reversible diffusion. The main idea is to use the generator's eigenpairs to predict how an initial distribution evolves toward equilibrium, and then realize this evolution through a deterministic particle transport. The authors analyze errors arising from the spectral approximation and show how reusing the initial particles to construct the transport can reduce sampling fluctuations. The method is evaluated on synthetic examples and alanine dipeptide.

Critical strengths

The proposed methodology is interesting, particularly the idea of using the spectral representation to construct a transport without continuously updating the particle distribution.

The theoretical analysis is sound, particularly the error analysis and the intuition of how reusing the initial particles can reduce sampling fluctuations. The latter is also supported by the numerical experiments.

Critical weaknesses

The comparison with existing spectral sampling methods is somewhat limited. Although BKT reaches comparable accuracy earlier than LAWGD on alanine dipeptide, LAWGD is not included in the synthetic experiments. It is therefore unclear whether the observed advantage is specific to this application or extends to other distributions.

The higher-dimensional experiments rely on rather specific problem structures. In the 10D non-Gaussian example, only one coordinate is initially out of equilibrium, while the 50D example consists of independent double wells treated separately. The coupled Gaussian example uses analytical eigenfunctions. I therefore find it difficult to assess how well the approach would perform on a higher-dimensional, coupled non-Gaussian problem where the eigenfunctions must be learned jointly.

Questions for authors
  1. Could the authors include LAWGD in some of the synthetic experiments, particularly the coupled double-well examples, using the same estimated eigenpairs? This would help assess whether the faster convergence observed on alanine extends to other distributions.

  2. Could the authors elaborate on the applicability of BKT to higher-dimensional, coupled non-Gaussian distributions? In particular, how does the need to estimate a joint spectrum affect the computational cost and accuracy of the approach?

4 comments

Part 1/2. Related to LAWGD (W1, Q1)

We thank the reviewer for the careful reading and for two requests that were concrete enough to carry out as stated. We ran both experiments. In the revised PDF, the changes we made for this review are in green and the changes we made on our own are in blue. The line numbers below are those in the margin of the PDF.

Both weaknesses are well founded. On W1 the new experiment shows that the earlier convergence of BKT on alanine dipeptide does not extend to the synthetic systems, and the revision says so. On W2 the new coupled example shows that BKT works with a jointly estimated spectrum, and it also shows where BKT is more sensitive than LAWGD.

W1 and Q1. LAWGD on the synthetic systems. We ran LAWGD on the two-dimensional double wells, β∈{0,0.25,0.5,1}\beta\in\{0,0.25,0.5,1\}, and on the four-well product in exactly the setting of Table 1. In each of ten realizations the two methods share everything that can be shared, the Koopman (RBF) eigenpairs (J=145J=145, n=2⋅105n=2\cdot10^5, r=64r=64), the 20002000 initial particles, the target sample and the sliced distance. LAWGD moves the particles by x˙i=−∑k≤rλ^k−1mk(s)∇φ^k(xi)\dot x_i=-\sum_{k\le r}\hat\lambda_k^{-1}m_k(s)\nabla\hat\varphi_k(x_i) with the moments mk(s)m_k(s) of the current particles, and its normalized time ss is compared with s=λ1ts=\lambda_1t for BKT. Both methods use the same RK4 steps, velocity bound and box (Appendix D.4, lines 1455–1493, Figure 8). The table gives the sliced W2W_2 error, mean ±\pm sd over ten realizations.

systemBKT, s=0.5s=0.5LAWGD, s=0.5s=0.5BKT, s=8s=8LAWGD, s=8s=8reference
β=0\beta=00.470±0.0040.470\pm0.0040.493±0.0050.493\pm0.0050.028±0.0030.028\pm0.0030.026±0.0010.026\pm0.0010.040±0.0080.040\pm0.008
β=0.25\beta=0.250.461±0.0040.461\pm0.0040.489±0.0050.489\pm0.0050.029±0.0030.029\pm0.0030.028±0.0010.028\pm0.0010.041±0.0080.041\pm0.008
β=0.5\beta=0.50.442±0.0050.442\pm0.0050.473±0.0050.473\pm0.0050.031±0.0040.031\pm0.0040.028±0.0010.028\pm0.0010.042±0.0100.042\pm0.010
β=1\beta=10.440±0.0050.440\pm0.0050.479±0.0040.479\pm0.0040.027±0.0030.027\pm0.0030.027±0.0010.027\pm0.0010.042±0.0120.042\pm0.012
four wells0.712±0.0070.712\pm0.0070.601±0.0080.601\pm0.0080.028±0.0020.028\pm0.0020.023±0.0020.023\pm0.0020.043±0.0080.043\pm0.008

Both methods reach the reference level by s=4s=4. BKT is ahead for s≤0.5s\le0.5 on the double wells and LAWGD for s≤2s\le2 on the four wells. At s=8s=8 LAWGD has a smaller spread and a slightly smaller error, within the spread of BKT on the double wells and lower by 0.0050.005 on the four wells. A velocity evaluation costs the same for both methods (Remark 4.8, lines 414–416), and so does a run, 33 to 44 s on the double wells and 1010 s on the four wells on one core.

The reviewer's doubt is therefore confirmed. The advantage seen on alanine dipeptide is specific to that system, and we no longer present it as a general property. Section 5 says that both methods reach the reference level (lines 468–469), the full comparison is in Appendix D.4, and Appendix D.7 (lines 1607–1611) lists it among the limitations.

Here is how we distinguish the two methods. The final error does not separate the two methods, because both reach the same cancellation of the sampling error. LAWGD imposes the moment conditions at the final positions of the particles and BKT at the initial ones (Remark 4.8), which is consistent with the smaller final spread of LAWGD. The two methods differ in structure. The field of BKT is fixed after t=0t=0, so BKT is a transport map whose particles move independently of each other once the coefficients are computed, and the error analysis of Section 4 applies to it. LAWGD reaches the cancellation through the interaction of the particles.

Part 2/2.

W2 and Q2. Regarding to the coupled non-Gaussian problem with a jointly estimated spectrum, I think that the reviewer is right that the 10D and 50D examples isolate the transport from the estimation. The revision adds the missing case, a coupled three-dimensional double well V=∑i=13(xi2−1)2+β2∑i=12(xi+1−xi)2V=\sum_{i=1}^3(x_i^2-1)^2+\frac\beta2\sum_{i=1}^2(x_{i+1}-x_i)^2 with β=0.5\beta=0.5 (Table 1, lines 492–494, Appendix D.5, lines 1495–1534). It has eight wells and seven slow modes with non-separable eigenfunctions. The source sits in one well, so all three coordinates are out of equilibrium and the transport has to populate all eight wells. The eigenpairs are estimated jointly from 2⋅1052\cdot10^5 trajectory pairs with tensor Gaussian dictionaries, with FD eigenpairs on a 41341^3 grid as the control.

The table gives the results at r=64r=64 over ten realizations. The reference is 0.046±0.0160.046\pm0.016, and the total variation between the masses of the eight wells of an independent target sample and of the FD density is 0.0280.028.

eigenpairsBKT, sliced W2W_2LAWGD, sliced W2W_2TV of the well masses, BKT / LAWGD
FD-41 (control)0.029±0.0020.029\pm0.002--0.0230.023 / --
12 Gaussians per coordinate, J=1729J=17290.037±0.0050.037\pm0.0050.037±0.0040.037\pm0.0040.0270.027 / 0.0240.024
8 Gaussians per coordinate, J=513J=5130.083±0.0060.083\pm0.0060.052±0.0030.052\pm0.0030.0720.072 / 0.0270.027

With the per-axis resolution of the two-dimensional experiments (1212 Gaussians per coordinate) both methods reach the reference level, BKT for every rr and LAWGD for r≥32r\ge32. With 88 per coordinate BKT leaves 0.260.26 of the mass in the initial well instead of 0.190.19. The slow eigenvalues of this dictionary are within 3%3\% of the FD values, so this is an error of the coarse eigenfunctions. LAWGD, whose moments are re-estimated from the particles, is less sensitive to it. Appendix D.7 lists this among the limitations (lines 1609–1611).

Now we show the cost and accuracy with the joint spectrum. With a joint spectrum the cost lies in the estimation, which takes O(nJ2+J3)O(nJ^2+J^3) operations for nn pairs and JJ functions, whereas one velocity evaluation takes O(MJd)O(MJd) for MM particles. In the three-dimensional example, fitting takes 1515 s for J=513J=513 and 107107 s for J=1729J=1729 on one core, and a BKT run of about 270270 RK4 steps 11 s and 33 s when the dictionary is evaluated through its tensor structure. The accuracy is governed by the slow eigenfunctions inside the wells, as the experiment with the coarse dictionary shows. How many modes are needed depends on how many slow modes carry weight of the source. A product of dd double wells has 2d−12^d-1 slow modes, and a source in one well has nonzero coefficients on all of them. A tensor dictionary grows exponentially with dd, which is the cost of a joint estimation with a fixed dictionary. The way to higher dimension is therefore through kernel or neural estimators of the joint spectrum. BKT accepts them without change, because the bounds of Corollary B.1 transfer the errors of any eigenpairs in energy norm and eigenvalue, and it adds neither density estimation nor interaction between the particles on top of them.

For your convenience, we list the changes (green) made:

  • Abstract, the three-dimensional example and the comparison with LAWGD (lines 31–33).
  • Table 1, the LAWGD column, the three-dimensional row and the caption (lines 432–444).
  • Remark 4.8, the cost of a velocity evaluation (lines 414–416).
  • Section 5, LAWGD on the synthetic systems (lines 468–469) and the three-dimensional example (lines 492–494).
  • Appendix D.4 with Figure 8 (lines 1455–1493).
  • Appendix D.5 with Table 8 (lines 1495–1534).
  • Appendix D.7, the comparison with LAWGD and the sensitivity to the eigenfunctions (lines 1607–1611).

We thank the reviewer again for the two requests, which improved the paper, and we are glad to run any further comparison.

I thank the authors for their detailed rebuttal and additional experiments. As noted by the authors, the new results highlight limitations in the approach. The empirical advantage over LAWGD is not consistent and there seems to be a high sensitivity to eigenfunction estimation errors. Scalability also seems to be a concern in high-dimensional coupled systems.

I will maintain my original recommendation.

We thank the reviewer for engaging with our rebuttal and additional experiments. To ensure an accurate record of what the new results show, we briefly clarify three points:

  1. Our primary contribution is not claiming uniform empirical dominance over LAWGD, but constructing a transport map with a pre-fixed velocity field covered by the error analysis in Section 4. Given the same eigenpairs, BKT matches both direct-sampling accuracy and LAWGD's computational cost (Table 1) while providing this theoretical guarantee.

  2. Sensitivity to eigenfunction errors appeared only in a single "stress test" where the dictionary of the 3D example was deliberately coarsened from 1212 to 88 Gaussians per coordinate. Under standard resolution (the 2D experiments), BKT is in fact more robust at low ranks, reaching reference accuracy for every rank rr, whereas LAWGD requires r≥32r \ge 32 (lines 1525–1531).

  3. High-dimensional spectral estimation is a bottleneck shared by all spectrum-based samplers, including LAWGD, rather than a limitation specific to BKT. Because BKT accepts eigenpairs from any upstream estimator, advances in spectral estimation transfer directly to our framework (lines 520–523).

We thank the reviewer again for the discussion; these clarification was incorporated into the manuscript.

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

This paper proposes Backward Kolmogorov Transport to study systems that have stochastic nature and partial measurement observations. They consider non-interacting particles transported via Fokker-Planck given trajectory information. The paper provides a detailed theoretical study with support of numerical experiments which include OU process, 2d double well potential, multi-well potential and alanine dipeptide.

Critical strengths

Theoretical study: The paper provides a detailed theoretical study of the proposed method including formulation of transport identity, stability of Wasserstein distance and same-sample coefficients.

Support by empirical evidence: The claims are validated and supported by small-scale numerical experiments for which there is available codebase in the reproducibility section.

Critical weaknesses

Presentation: The paper does not seem to have a Related work section which would help understand the positioning. It would also help moving some of the theoretical background into the appendix and providing further empirical validations in the main text.

Applicability and scalability: It is difficult to assess which areas the method could benefit. The paper mentions domains that have partially observed measurements such as molecular dynamics and climate applications as well as experiment on alanine dipeptide. But given that it requires precomputing eigenpairs it is unclear how it could scale to larger application domains (climate or other larger molecular systems).

Questions for authors
  • Could you explain which systems this could be more broadly applied to given the constraint that the method considers non-interacting particles? It would be great to know since most systems in nature have interacting particles?

  • Would it be possible to explain whether Backward Kolmogorov could help theoretically validate experiments that include interacting particles?

4 comments

Part 1/2. Why it is not restricted to non-interacting systems (Q1, Q2)

We thank the reviewer for the time spent on our paper and for the questions, which showed us where the submission was easy to misread. In the revised PDF, the changes we made for this review are in orange and the changes we made on our own are in blue. The line numbers below are those in the margin of the PDF.

The review reads the paper as a study of non-interacting particles under partial observation. Neither is the case, and both questions and the concern about applicability follow from this reading. We therefore address it first.

The data are one long trajectory of a system in equilibrium, for example the two backbone angles of a small molecule recorded along a molecular dynamics simulation. The goal is to draw new samples from the equilibrium distribution of the recorded coordinates, starting from any cloud of points we choose. From the trajectory we estimate the slow eigenfunctions and eigenvalues of the generator of the dynamics. These give how any initial distribution relaxes to equilibrium in closed form, and BKT moves each point of the cloud along this relaxation until the cloud follows the equilibrium distribution. The trajectory is the only input, and the paper makes no assumption of partial observation. Whatever coordinates the trajectory records, the method samples their equilibrium law.

Q1. The method has no constraint of non-interacting particles. The "particles" of the paper are the points of this cloud, not the atoms or molecules of the physical system. Each point is one complete state of the system, for alanine dipeptide one pair of backbone angles, that is, one conformation of the molecule. The physical system may have any interactions, since they shape the dynamics that produced the trajectory and BKT sees them through it. Our main example is of this kind. The alanine dipeptide trajectory (Section 5, line 450, and Appendix D.3, line 1344) comes from a simulation of the molecule in explicit solvent, with all the interactions between the atoms and the water. BKT therefore applies to interacting systems, and more generally to any system whose recorded coordinates follow a reversible diffusion, or are modeled as one.

In the submission, "non-interacting" was a statement about the sampler. Each point of the cloud moves on its own, along a velocity field that depends only on its position and on time, so BKT is a transport map and any number of points can be moved independently and in parallel. The situation is the same as with independent Monte Carlo runs of a molecule, where the runs do not interact with each other while the atoms of the molecule do. This independence is an advantage of the method. Spectral samplers such as LAWGD (Remark 4.8) do not have it, because the velocity of each of their points depends on the positions of all the others.

The reviewer is right that the phrase "non-interacting particles" invites the physical reading, and we no longer use it. The abstract and the first contribution now say that each sample point follows its own ordinary differential equation and that BKT is a transport map (lines 23–25 and 100–105). The introduction defines a particle where the word first appears, as a sample point of a particle method (lines 50–51), and says that the physical system may have interacting degrees of freedom (lines 89–99). Section 3 says that the particles are the sample points that represent the transported law, as in particle methods for transport equations, and that their equations are decoupled (lines 173–175 and 191–194).

Q2. The theory of Section 4 holds for any reversible diffusion, whatever the interactions inside the physical system. It bounds the error of the transported points by the errors of the estimated eigenfunctions and eigenvalues (Corollary 4.2 and Theorem 4.5). For an interacting system it therefore gives a sampler of the equilibrium law of the recorded coordinates whose error is traced back to the estimated spectrum. This is a statement about the sampler, and the paper does not claim to validate a physical experiment.

Part 2/2. Related work

Related work. The paper does discuss the related work, in three paragraphs of the introduction instead of a separate section, and each paragraph says how BKT differs. Lines 47–58 treat particle methods built on kernel density or score estimates. These need the estimates at every step, and their errors do not decrease along the flow, whereas BKT estimates no density. Lines 71–84 treat density forecasting, which uses the same eigenfunction expansion to predict distributions but does not move samples, and the heat-flow map, density-equalizing maps and diffusion models, where the operator is known. Lines 86–98 treat the spectral samplers of the LAWGD type, which use an estimated spectrum as BKT does but whose particles interact, and probability-flow samplers and flow matching, which move particles by a single flow map as BKT does but learn the velocity field. In BKT the field comes in closed form from the estimated spectrum.

We followed the suggestion to move theory to the appendix and to bring experiments into the main text. The technical construction of the validity regions moved from Section 4.2 to Appendix B.3 (lines 974–987), and Section 4 now points to it (lines 354–357). Each main result of Section 4 is followed by an explanation of its terms in words (lines 277–278, 290–293, 345–353 and 396–404). In the main text, Table 1 has a new LAWGD column and a new coupled three-dimensional example (lines 432–444, 468–469 and 492–494), and Section 5 ends with a new paragraph on high barriers (lines 501–509). BKT is for systems whose recorded coordinates follow a reversible diffusion, or are modeled as one, and for which an equilibrium trajectory is available. In practice these are a few slow coordinates of a molecular system, as for alanine dipeptide.

Precomputing the eigenpairs is not a cost specific to BKT, since it is the estimation step of Koopman and transfer operator methods, which are in routine use for molecular simulation data, and BKT accepts the eigenpairs of any such estimator. This cost grows with the number of functions used to represent the eigenfunctions. Moving the points is cheap by comparison, a few evaluations of the eigenfunctions per point and time step, and the points do not interact. The error bounds hold for any estimated eigenpairs, so kernel and neural estimators of the spectrum, which scale to larger systems, can be used without any change to the method. The new three-dimensional example (Table 1, lines 492–494, Appendix D.5) estimates the eigenfunctions jointly in a coupled system with eight wells. With 1212 Gaussians per coordinate (J=1729J=1729), fitted in about two minutes on one core, BKT reaches the accuracy of direct sampling (sliced Wasserstein error 0.037±0.0050.037\pm0.005, compared with 0.046±0.0160.046\pm0.016 between two independent samples of the target). For climate models or large molecules in their full coordinates the slow eigenpairs would come from such estimators, and we report no experiment at that scale. The paper establishes that the transport and its error bounds stay the same whatever the estimator. The limitations of the method are collected in Appendix D.7 (lines 1606–1614).

Changes (orange).

  • Abstract, the sample points and their own equations (lines 23–25).
  • Introduction, the definition of a particle (lines 50–51), the comparison with interacting particle systems and the interacting physical system (lines 89–99), and the first contribution (lines 100–105).
  • Section 3, what the particles are and why their equations are decoupled (lines 173–175 and 191–194).
  • Conclusion (lines 513–515).
  • Appendix B.3, the construction of the validity regions moved from Section 4.2 (lines 974–987).

We thank the reviewer again. The score rests on the reading of "non-interacting particles" that Part 1 corrects, so we kindly ask the reviewer to reassess the paper in this light. We are glad to answer any further question.

Thank you for the clarifications provided in your responses and for incorporating these changes in the revised manuscript. The set-up and the assumptions regarding particle interactions are clearer now. Therefore, I will be raising the score.

Thanks for reading our revision and raising your score. We really appreciate your time and feedback.

Comments on the paper

1 comment

We thank the three reviewers for their time and for reviews that were specific enough to act on. One review ran our code on two hard tests, one asked for two concrete experiments, and one showed us that a phrase of the submission was misleading. We ran every experiment that was asked for, and the revised PDF reports the results as they came out, including those that are less favorable to our method.

In the revised PDF, blue marks the changes we made on our own. The other three colors mark what we changed in response to the reviews, one color for each review. The line numbers in our responses are those in the margin of the revised PDF. Remark 4.3 is new, so Theorem 4.7 is Theorem 4.6 of the submission. The experimental details are now Appendix D, with the new material in Appendices D.4 to D.7.

Clarification in original submission.

  1. BKT does not assume non-interacting physical particles. In the paper a particle is a sample point that the sampler moves, as in particle methods for transport equations. Each sample point follows its own ordinary differential equation, so BKT is a transport map. The degrees of freedom of the physical system may interact in any way, and they enter only through the trajectory. We agree that the phrase "non-interacting particles" of the submission invites the physical reading, so we dropped it and now define the term where it first appears (lines 23–25, 50–51 and 173–175).

  2. At high barriers the transport is not what fails. With the exact slow eigenpair BKT carries the correct mass across a barrier of 3333 times the temperature in about 100100 steps per particle. The failure in such tests comes from the estimation of λ1\lambda_1 and φ1\varphi_1 from data that have no samples in the transition layer, and every method that estimates rates from trajectories has this limit. The submission named this limit without quantifying it. The revision shows that with trajectory data a few hundred observed transitions are needed, and it gives four checks, computed from the data and the output, that detect wrong basin masses (lines 501–509, Appendix D.6).

  3. The earlier convergence of BKT on alanine dipeptide is specific to that system. On the synthetic systems, with the same eigenpairs and particles, BKT and LAWGD reach the reference level by the same flow time and at the same cost, and the revision says so (Table 1, lines 432–444 and 468–469, Appendix D.4). BKT differs from LAWGD in that its field is fixed after the initial time, so its particles are decoupled and the error analysis of Section 4 applies.

What is new in the paper.

  1. We added LAWGD with the same estimated eigenpairs, initial particles and steps as BKT on the two-dimensional double wells and the four-well product (Table 1, Appendix D.4).
  2. We added a coupled three-dimensional double well with eight wells and jointly estimated eigenpairs, with an FD control and LAWGD (Table 1, lines 492–494, Appendix D.5).
  3. We showed a barrier-height scan with trajectory data and two tests on Gaussian mixtures with barriers of 3333 and 3434 times the temperature, with estimated eigenpairs and with the exact eigenpair (Appendix D.6).
  4. We provide a more precise statement of the role of the dynamics. For a reversible diffusion with known diffusion matrix, the estimators of the paper obtain the spectrum of the generator from the sampled states through its Dirichlet form. The time order of the data enters only through estimators built on pairs and through the consistency checks. The abstract, introduction, Section 2 and conclusion now say so (lines 16, 58–65, 155–160 and 515).
  5. The limitations of the method, is also discussed in Appendix D.7 (lines 1606–1614).

Each review is answered point by point in its own thread, with different color.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.