acceptodds
Under review as a conference paper at ICLR 2027

Reusable Schrödinger Bridges from Coarse Endpoint Interactions

Abstract

Schrödinger bridges connect prescribed distributions through reference dynamics and provide a common framework for control, generation, and scientific inference. When many endpoint pairs share the same dynamics, can their bridges also share a smaller optimization problem? We show that fine endpoint information and the interaction governing endpoint pairing can be represented separately. Because the marginals are fixed, additive single-endpoint terms leave the optimizer unchanged. Approximating the remaining interaction on coarse states therefore gives a reusable solver that preserves the original fine marginals exactly. For positive finite references, a uniform log-interaction residual bounds each Kullback-Leibler (KL) direction by , with a sharp quadratic coefficient. A sparse-query converse then leads to a reference-only partition rule with an equal-budget approximation guarantee. We further decompose query error and quantify the dependence recovered by nested refinement, while fast mixing explains why coarse interactions arise dynamically. Our experiments verify the predicted error structure and demonstrate that reusing coarse endpoint interactions provides a valid and practical way to solve repeated bridge queries in molecular and stochastic-dynamics settings.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Schrödinger bridges, entropic optimal transport, multiscale stochastic dynamics, coarse-graining

Reviews

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

paper studies how to reuse a compact representation of a fixed reference process when solving Schrödinger bridge problems with different endpoint distributions. The main observation is that, with fixed marginals, terms depending on only one endpoint do not affect the optimal coupling. The authors therefore approximate the remaining endpoint interaction using coarse states, and reconstruct the fine coupling using the conditional distributions of each query. The resulting construction preserves the original endpoint marginals, including sparse and asymmetric distributions. The paper provides KL and TV error bounds, a sparse-query converse, a representation-selection guarantee, and an error decomposition that separates coarse-kernel mismatch from information lost through the partition. It also connects coarse interactions to fast-slow dynamics. The theory is evaluated on controlled stochastic systems and molecular examples, including alanine dipeptide, Chignolin, and WLALL. The experiments examine approximation quality, storage requirements, and computational savings for repeated coupling and full-path queries.

Critical strengths
  1. I think the strongest part of the paper is the theoretical treatment of endpoint interaction. The idea of removing single-endpoint terms is simple, but the authors use it to characterize which information must be retained for repeated bridge queries. Theorem 4.3 gives bounds on both KL directions in terms of the uniform log-interaction residual, including the sharp quadratic coefficient. I particularly like the converse based on sparse two-point queries. It shows that certain interactions lost through coarse-graining cannot be recovered simply by choosing a different coarse kernel. This makes the approximation criterion more meaningful than a sufficient error bound alone. The connection to Hilbert projective distance and the fixed-budget guarantee in Theorem 6.1 is also interesting. Although farthest-first clustering is not new, relating its covering radius to worst-case bridge-query error is a useful contribution.

  2. The decomposition (F=C+J) in Theorem 4.4 is another result I found valuable. It separates the error caused by the chosen coarse kernel from the error that cannot be removed without refining the partition. The nested-refinement identity makes this distinction more precise. The experiments are also fairly thorough. The authors compare several partition and kernel choices, NMF baselines, conditional lifting, and repeated-query computation. I appreciate that the same reference and endpoint distributions are used across comparisons, and that offline processing and conditional-path sampling are accounted for separately. The treatment of the observed molecular data is reasonably careful. In particular, the WLALL evaluation distinguishes the five independent test trajectories from the 30 queries constructed from them. This is an important distinction when interpreting the experimental evidence.

Critical weaknesses
  1. My main reservation is the gap between the representation-selection theory and the empirical results. Theorem 6.1 proposes farthest-first clustering under the projective distance (d_K), with a representative coarse kernel. Appendix D.4 describes a reference-distance partition based on log-ratio distances, but I could not tell whether the reported implementation exactly follows the construction in the theorem. If the "Reference distance / R" results in Table 10 correspond to this construction, their performance is somewhat disappointing. The median symmetric KL is essentially identical to independent endpoint pairing on three query families: 0.0345 for alanine, 0.100 for observed Chignolin, and 0.355 for WLALL. For synthetic Chignolin queries, it is slightly worse than independent pairing (0.171 versus 0.167). These results do not mean that the resulting couplings are independent, and I understand that a worst-case guarantee does not necessarily imply good average performance. Still, I would like to understand why the reference-distance representation captures so little useful dependence in these cases. Spectral aggregation performs substantially better across these benchmarks. NMF also achieves lower errors at higher ranks, although at a greater storage cost. For example, on WLALL, rank-8 NMF uses 62.88 KiB compared with 4.43 KiB for the spectral representation (Section 7.2). The main empirical advantage therefore seems to be the accuracy-storage tradeoff of spectral aggregation, rather than the theoretically guaranteed partition construction. I think the paper would benefit from a more direct discussion of this difference. In particular, under what conditions should the guaranteed construction work well for the endpoint queries encountered in practice?

  2. The practical benefits are less convincing beyond the tested settings. The computational results support the proposed approach, but I am not sure how broadly the reported savings will carry over. The largest experiment, with 12,288 states, uses an artificially constructed reference with an exact coarse factorization. This is a useful test of computational scaling, but it is also a particularly favorable setting for the method. The molecular references are much smaller, with state spaces containing only a few hundred states. There is also a limitation for continuous problems. As discussed in Appendix B.1, Brownian interactions on unbounded Euclidean support cannot have a uniformly bounded interaction residual under a finite partition. The Gaussian experiments use finite endpoint banks, which avoids this issue but does not establish the same guarantee for unrestricted continuous domains. Finally, the end-to-end improvements become relatively modest once complete path sampling is included. The reported speedups are 1.70× and 1.24× for the Brownian experiments, and 1.16× and 1.32× for Chignolin and WLALL (Section 7.3). This is understandable because the conditional sampler is shared, but it also means that large savings in coupling computation do not necessarily translate into similarly large savings for complete path generation. I do not consider this a fundamental weakness of the method. However, the practical conclusions should be framed more carefully. The paper demonstrates a useful computational mechanism, while its advantage on larger, less structured problems remains to be established.

Questions for authors
  1. Does the "Reference distance / R" baseline in Table 10 implement the exact farthest-first construction under (d_K) from Theorem 6.1? If so, could the authors explain why its performance is so close to independent endpoint pairing on several molecular benchmarks? Are these couplings actually close to independence, or is the similarity limited to the median KL values? I think this would help clarify the practical relevance of the representation-selection guarantee.

  2. Relationship to previous work. Could the authors distinguish their contributions more explicitly from the fixed-marginal stability results of Eckstein and Nutz (2022, particularly Proposition 3.12) and the conditional reconstruction approach in QDSB (Fuchs et al., 2026)? My understanding is that the sharp constants, sparse-query converse, and fixed-budget representation guarantee are the main advances here, rather than quadratic stability or conditional lifting by themselves. A clearer comparison would help establish the novelty.

  3. Reproducibility of the molecular experiments. Could the authors point to the specific supplementary files containing the final WLALL shrinkage parameters, partition configurations, and frozen inputs? I would also appreciate clarification on how test frames falling outside the training-defined active support are handled, and whether this affects a meaningful fraction of the data.

6 comments

Thank you very much for taking the time to review and for your support. We appreciate your careful discussion of the theory and experiments so much, and we will try our best to address your concerns and questions below.

1. Concern: The reference-distance construction captures little dependence on the molecular benchmarks. When should its guarantee translate into practical accuracy?

A1. The reference-distance construction seeks a coarse interaction that approximates fine-state pairing preferences uniformly across endpoint pairs. For a positive reference kernel KK and a partition qq, the relevant decomposition is

log⁡Kij=ai+bj+cq(i),q(j)+eij,\log K_{ij}=a_i+b_j+c_{q(i),q(j)}+e_{ij},

where ai,bja_i,b_j are single-endpoint terms, cc is the interaction between coarse states, and ee is the residual. The construction compares incoming and outgoing pairing preferences over the entire reference. Thus its structural bias is toward the strongest distinctions that a future query could expose, as required by the worst-query guarantee. After conditional lifting, states in the same cell share one coarse interaction, and any remaining within-cell pairing dependence contributes to the partition error JJ in Theorem 4.4.

Spectral aggregation chooses the cells using stationary-mass-weighted dynamical coordinates. Consequently, it can allocate more resolution to regions carrying substantial molecular occupancy and preserve more of the dependence used by those queries. It still uses a hard partition and the same conditional lift, so the improvement comes from the placement of the cells and the aggregation of their interactions. The residual theory applies to this representation as well.

Nonnegative matrix factorization (NMF) takes a different approach by retaining state-specific positive factors. At rank rr on NN fine states, these factors contain 2Nr2Nr entries, allowing more detailed interactions at a higher storage cost. By comparison, our representation stores the state labels and a B×BB\times B coarse kernel, where BB is the number of cells. On WLALL, rank-eight NMF uses 62.88 KiB, while the eight-cell spectral representation uses 4.43 KiB. This explains the accuracy-storage tradeoff highlighted in your review.

The molecular occupancies make this distinction particularly concrete. At the reported budgets, one reference-distance cell contains 94.99% of Chignolin's reference mass (Lazzeri, 2025) and 99.36% of WLALL's (Scherer et al., 2015). As a result, eight of ten observed Chignolin queries and 16 of 30 WLALL queries have one endpoint supported on a single coarse cell. The coarse marginals then uniquely determine a product coupling. Thus the interaction-covering objective, together with the concentrated molecular occupancies, explains the independence-like behavior.

We further examined this explanation by crossing the two partitions with both representative and aggregated kernels on the ten observed Chignolin queries at B=16B=16. In the exact decomposition F=C+JF=C+J, forward KL error FF separates into coarse-kernel mismatch CC and the best error JJ attainable with that partition and conditional lift. The rounded means are:

Partition and kernelForward error FKernel term CPartition term J
Reference distance, representative0.1416170.0045260.137092
Reference distance, aggregation0.1383120.0012210.137092
Spectral, representative0.0552290.0183360.036894
Spectral, aggregation0.0377210.0008280.036894

The decomposition identifies the source of the gain directly: spectral partitioning retains more query-relevant dependence, reducing JJ, and aggregation improves the coarse fit, reducing CC. This also suggests a natural direction for future work: combine interaction coverage with stationary-mass information when selecting the partition, bringing the uniform guarantee and the intended query population into closer alignment.

While not strictly necessary, the following favorable conditions allow the guaranteed construction to achieve its best practical performance:

  • Fast conditional mixing: Under the assumptions of Section 5, rapid equilibration within fibers makes endpoint interactions compressible and yields the mixing-dependent residual rate.
  • A dynamically aligned partition: The cells distinguish the important pairing preferences. For the reference-distance construction, a small covering radius rBr_B gives worst-query KL error at most min⁡2rB,rB2/2\min\\{2r_B,r_B^2/2\\}.
  • Query mass on accurately represented interactions: Small query-weighted residual moments give bounds adapted to the actual endpoints, as developed in Theorem 4.4.
  • Many queries sharing one reference: Reuse spreads the offline representation cost across requests and realizes the online computational advantage.

Together, these conditions explain how the dynamics, partition, and query population determine accuracy, and how the workload turns that accuracy into a computational benefit. The uniform theorem controls the most demanding endpoint pair, while the query-weighted analysis describes the interactions used in a given application.

 

2. Concern: The practical advantage on larger, less structured problems remains unclear.

A2. We have added a larger experiment with a geometrically defined reference to examine this question directly. The data are self-generated from seven Gaussian components in 12 dimensions, with unequal coordinate scales and a nonlinear deformation. This produces multimodal point clouds of 4,096 and 8,192 states, from which we compute the Gaussian endpoint kernel before constructing the coarse representation.

We apply the exact reference-distance/representative construction from Theorem 6.1 and compare it with exact Sinkhorn and the endpoint-quantization implementation of QDSB (Fuchs et al., 2026). Each of ten queries uses half of the points at each endpoint. Every method solves the same endpoint problem and uses the same exact Brownian conditional sampler, returning 2,048 paths at five times. We tested budgets B=8,16,32,64,128B=8,16,32,64,128, with the following results at B=128B=128:

StatesMethodSymmetric KLStored representationComplete request
4,096Exact Sinkhorn0.000000128 MiB65.091 ms
4,096QDSB quantization0.039831160 KiB21.655 ms
4,096Ours0.036948160 KiB3.122 ms
8,192Exact Sinkhorn0.000000512 MiB300.052 ms
8,192QDSB quantization0.041290192 KiB36.263 ms
8,192Ours0.039627192 KiB3.388 ms

Entries are medians over ten queries. Symmetric KL averages both directions against exact Sinkhorn. Complete requests include coupling, sampler preparation, and conditional-path simulation, with five timing repetitions on one CPU thread. Storage counts the dense kernel, our labels and coarse kernel, or QDSB's query labels and coarse plan.

At equal representation storage, our construction achieves lower median error and faster complete requests, with median paired speedups of 6.91× and 10.60× over QDSB. Furthermore, including the one-time kernel construction and compilation costs of 11.60 s and 51.79 s, the projected totals for 10,000 requests are 43.31 s versus 217.38 s and 86.06 s versus 364.71 s. These projections use mean service costs across the ten measured queries. The results therefore demonstrate practical gains for the guaranteed construction on geometrically defined kernels beyond the exact-factorization experiment.

 

3. Question: Does Reference distance/R implement Theorem 6.1 exactly, and are its molecular couplings actually close to independence?

A3. Yes. Reference distance/R implements the farthest-first construction under dKd_K in Theorem 6.1. Thank you for asking us to clarify this connection. It computes the incoming and outgoing log-ratio distances over all reference states, selects centers by exact farthest-first traversal, and assigns states to their nearest centers. The representative kernel is Wab=Kca,cbW_{ab}=K_{c_a,c_b}, where ca,cbc_a,c_b are the selected centers. Each center belongs to its own cell, and the initial center is the state with the largest stationary mass, consistent with the theorem's arbitrary initialization.

We also measured total variation (TV) distance to independent endpoint pairing directly. For alanine dipeptide (Nüske et al., 2017) and the Chignolin and WLALL queries discussed above, the results at the reported budgets are:

Reference distance/R settingMedian TV to independenceMaximum TVQueries with TV below 10−810^{-8}
Alanine, B=28.10×10−138.10\times10^{-13}1.77×10−91.77\times10^{-9}30/30
Chignolin observed, B=165.97×10−175.97\times10^{-17}0.1796148/10
WLALL, B=81.62×10−151.62\times10^{-15}0.00518116/30

These measurements confirm that the similarity is visible in the couplings themselves. For Chignolin and WLALL, the numerically independent cases are precisely those with one endpoint supported on a single coarse cell, which forces the product coupling described in A1. Alanine combines a weak representative interaction, whose maximum absolute log cross-ratio is 0.003946, with concentrated endpoints, whose median largest-cell masses are 0.999988 and 0.999985. Increasing the WLALL budget to B=32B=32 raises median TV to independence to 0.043239, with every query occupying multiple cells at both endpoints. These observations can help clarify how query mass and partition resolution determine the dependence retained in practice.

Reviewer_iv6KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper investigates whether schrodinger bridges between between various end points with similar dynamics can be reused. They investigate settings where coarse dynamics on the endpoints can be optimized to create schrodinger bridges which can be reused exactly to recreate accurate (or "fine") dynamics under certain conditions. They do experiments on molecular and stochastic dynamics in settings where these conditions apply, but the numerical results are not incredible.

Critical strengths
  • The theory is strong to my eye, although I must admit that I found it difficult to follow in some places. The notion of being able to separate the bridge from one endpoint would obviously have applications in (coarse-grained) molecular dynamics and other common use cases. It was not obvious to me whether the settings in which their conditions hold would be useful for many practical use cases.
  • The numerical results are fairly extensive for a paper with so much theory.
  • I had a positive impression about the paper and think it belongs in the conference.
Critical weaknesses
  • Despite strong theoretical guarantees, the speedups are fairly modest on practical systems such as molecules. The limited applicability doesn't count toward rejecting the paper but slightly downweighs it. Perhaps a stronger quantitative comparison with Quantized diffusion schrodinger bridges by Tobias Fuchs, et. al. 2026 would be appealing.
  • It's cannot be denied that the paper is extremely dense. A significant amount of time spent organizing the information would be appealing.
Questions for authors
  • Can you justify that the realistic dynamics are actually related to the coarse dynamics by this log-interaction residual (egn 9)? Under what circumstances?
  • Why not do some work at the beginning of the paper to give an overview of your goals and how to achieve them theoretically rather than diving directly into quite strong jargon?
  • There are other works that factorize coarse to fine dynamics. One example is Making transport more robust and interpretable by moving data through a small number of anchor points by Chi-Heng Lin, et. al. 2021. You don't have many baselines to compare to, perhaps this would be a starting point?
4 comments

Thank you so much for taking the time to review and for your support to help this work better. We will try our best to address your concerns and questions about the practical gains, presentation, and relationship to previous work below.

 

1. Concern: The speedups on practical systems are modest, and a comparison with QDSB would be helpful.

A1. Actually, the speedup varies across practical systems because the reference size and the cost of conditional path sampling determine how much of each request benefits from the coarse solve.

For example, the Chignolin reference (Lazzeri, 2025) has 128 states, and dense Sinkhorn already takes 0.471 ms. At this scale, fixed solver overheads account for an appreciable part of the 0.401 ms coarse solve. This follows the computational structure in Section 6: a query costs O(N+IB2)O(N+IB^2) for NN fine states, BB coarse states, and II iterations, compared with O(IN2)O(IN^2) for dense iterations. Furthermore, complete path requests share the conditional-path simulation cost HH. Their speedup is therefore (Cdense+H)/(Ccoarse+H)(C_{\mathrm{dense}}+H)/(C_{\mathrm{coarse}}+H), where CC includes coupling and endpoint-pair sampling. As the shared simulation cost increases, the total speedup naturally approaches one.

Meanwhile, on the 503-state WLALL pentapeptide reference (Scherer et al., 2015), eight coarse states reduce coupling time from 2.820 ms to 0.408 ms, giving a 6.91× speedup with the same marginal tolerance of 10−1010^{-10}. The recorded 1,000-request replay takes 0.540 s including representation construction, compared with 2.834 s for dense scaling. This illustrates the practical value of serving many endpoint requests through a compact dynamical representation. Effective reaction coordinates and compact transfer-operator descriptions are well established in molecular modeling (Noé and Clementi, 2017, Schütte et al., 2023), providing a natural setting for this computational mechanism.

Following your suggestion, we also compared against the endpoint-quantization implementation of QDSB (Fuchs et al., 2026). We generated 12-dimensional data from seven Gaussian components, with unequal coordinate scales and a nonlinear deformation, producing point clouds of 4,096 and 8,192 states. The Gaussian endpoint kernel is computed from these points, and every method uses the same exact Brownian conditional sampler. We evaluated ten endpoint queries per size across budgets B=8,16,32,64,128B=8,16,32,64,128. At B=128B=128, the results are:

StatesMethodSymmetric KLStored representationComplete request
4,096Exact Sinkhorn0.000000128 MiB65.091 ms
4,096QDSB quantization0.039831160 KiB21.655 ms
4,096Ours0.036948160 KiB3.122 ms
8,192Exact Sinkhorn0.000000512 MiB300.052 ms
8,192QDSB quantization0.041290192 KiB36.263 ms
8,192Ours0.039627192 KiB3.388 ms

Entries are medians over ten queries. Symmetric Kullback–Leibler (KL) divergence averages both directions against exact Sinkhorn. A complete request includes coupling, sampler preparation, and 2,048 paths evaluated at five times. Each query is timed five times on one CPU thread. Storage counts the dense kernel, our labels and coarse kernel, or QDSB's query labels and coarse plan.

Thus, at the same representation storage, our method achieves lower median error and faster complete requests, placing these configurations on the measured accuracy-time frontier. The median paired speedups over QDSB are 6.91× and 10.60×. Moreover, reuse pays off after including the one-time kernel construction and compilation costs of 11.60 s and 51.79 s. Using the measured query costs, the projected totals for 10,000 requests are 43.31 s versus 217.38 s and 86.06 s versus 364.71 s, respectively. We will add this comparison to make the practical benefit of a reusable representation clearer.

 

2. Concern: The paper is dense and would benefit from better organization.

A2. We are so sorry for the reading burden caused by the density of the presentation. The nine-page limit of ICLR led us to compress several connected results, which is sad for for both us and all those who care about our work and want to read our papers more easily. We will make a substantial effort to improve their presentation, because our goal is to help readers recognize our main ideas intuitively and quickly, and to inspire researchers in broader communities for all the time. In particular, we will try our best to add a short takeaway after each important result, explain the central formulas in as plain language as possible.

3. Question: Under what circumstances do realistic dynamics admit the log-interaction residual in Equation (9)?

A3. Equation (9) quantifies this relation by decomposing the reference log-kernel into single-endpoint terms, a coarse interaction, and a residual. For a positive reference kernel KK and a partition qq of the fine states, we write

log⁡Kij=ai+bj+cq(i),q(j)+eij.\log K_{ij}=a_i+b_j+c_{q(i),q(j)}+e_{ij}.

Here aia_i and bjb_j depend on individual endpoints, cc represents the interaction between coarse states, and ee is the approximation residual. With fixed endpoint marginals μ,ν\mu,\nu, the single-endpoint contribution is the constant ∑iμiai+∑jνjbj\sum_i\mu_i a_i+\sum_j\nu_j b_j. Consequently, the residual measures precisely how well the coarse interaction preserves the pairing preferences that determine the fine bridge. Fast conditional mixing provides one concrete sufficient mechanism. Under the assumptions of Section 5, the decomposition can be chosen with η≤A(ε+e−cmixT/ε)\eta\le A(\varepsilon+e^{-c_{\mathrm{mix}}T/\varepsilon}), where η=max⁡ij∣eij∣\eta=\max_{ij}|e_{ij}|, ε\varepsilon is the scale ratio, TT is the horizon, cmix>0c_{\mathrm{mix}}>0 is the mixing rate, and AA is independent of ε\varepsilon.

Importantly, the paper further relaxes this uniform requirement for individual queries. Theorem 4.4 bounds the error using query-weighted exponential moments of the residual and gives the exact decomposition F=C+JF=C+J. Here FF is forward KL error, CC measures coarse-kernel mismatch, and JJ measures the dependence lost through the partition. The analysis therefore accounts for which interactions the actual endpoints use. On the 30 synthetic Chignolin queries, for example, the median forward KL is 0.0831 and the refined query-weighted bound is 0.0838, directly connecting this relaxation to practical accuracy.

There is also substantial support for these mechanisms in realistic dynamics, and effective coarse dynamics have a substantial foundation in molecular modeling. Legoll and Lelièvre (2010) derive effective dynamics using conditional expectations. Importantly, Bittracher et al. (2018) show that, under a suitable reducibility condition, low-dimensional reaction coordinates can preserve dominant timescales even in metastable systems lacking local fast–slow separation. More broadly, Schütte et al. (2023) review transfer-operator methods for describing long-timescale molecular dynamics. Together, these results support the use of effective coarse coordinates beyond explicit local fast–slow decompositions. For a given representation, our log-interaction residual provides the additional bridge-specific accuracy criterion, while the query-weighted extension adapts the bound to the endpoint distributions of interest.

 

4. Question: Could the introduction first explain the goal and the theoretical approach intuitively?

A4. Thank you very much for this helpful suggestion. We fully agree that readers should first understand what we want to achieve and how the main idea leads to the theory. We will add the following explanation to the introduction:

“Schrödinger bridges connect two observed distributions through reference dynamics, supporting applications in control, generation, and molecular modeling. In many applications, the dynamics remain fixed while the endpoint observations change. Then why should each new pair require solving the same large interaction problem again? Each query already supplies its detailed endpoint distributions. We can therefore precompute a smaller description of how groups of states prefer to pair, solve the pairing problem between these groups, and restore the original endpoints using their within-group distributions. The theory then tells us how accurately this reusable representation reproduces the fine bridge and which interactions it must retain.”

We hope this overview can lead naturally to the approximation bound, the converse, and the reusable solver, with a short takeaway explaining the role of each result.

5. Question: Could anchor-based transport provide a starting point for comparison?

A5. Both approaches use an intermediate transport representation to organize fine-scale interactions, while the objectives and reuse settings differ. In particular, the work of Lin et al. (2021) and our approach actually address quite orthogonal aspects of this construction.

  • Goal: Latent Optimal Transport uses a three-stage anchor objective to promote robust and interpretable alignment. Our goal is to approximate the Schrödinger bridge associated with prescribed reference dynamics across many endpoint requests.
  • Representation selection: Their source and target anchors are learned for the supplied distributions. Our partition and coarse kernel are selected from the reference before the future endpoint distributions arrive.
  • Theory: Their analysis studies the structured transport objective and its rank properties. Our analysis connects endpoint-uniform KL accuracy, a sparse-query converse, and a fixed-budget representation-selection guarantee.

Accordingly, using Latent Optimal Transport as a direct baseline for our problem would require adapting its endpoint-specific anchor construction to a fixed reference while preserving the original Schrödinger objective. Thus, this may not be a sufficiently appropriate baseline. The QDSB comparison above is abel to provide a potentially effective experimental comparison for coarse endpoint pairing followed by conditional-path sampling. We will clarify these complementary goals in the related-work discussion.

Dear Reviewer iv6K,

 

Thank you for your time and efforts in reviewing our paper.

Based on your suggestions and feedback, we provide clarifying responses and additional experiments to answer your questions. We sincerely hope that this addresses your concerns.

We would like to kindly remind that the discussion period will end in a few days. We wonder whether there are additional clarifications that we could further provide to address your concerns (if there are any). It would be very much appreciated if you could once again help review our responses and let us know if these address your concerns and if our explanations are heading in the right direction. Please also let us know if there are any further questions or comments about this paper.

We strive to consistently improve the paper and it would be our pleasure to have your precious feedback!

Thank you very much!

 

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.