- 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
-
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.
-
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
-
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?
-
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
-
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.
-
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.
-
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
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 and a partition , the relevant decomposition is
where are single-endpoint terms, is the interaction between coarse states, and 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 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 on fine states, these factors contain entries, allowing more detailed interactions at a higher storage cost. By comparison, our representation stores the state labels and a coarse kernel, where 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 . In the exact decomposition , forward KL error separates into coarse-kernel mismatch and the best error attainable with that partition and conditional lift. The rounded means are:
| Partition and kernel | Forward error F | Kernel term C | Partition term J |
|---|---|---|---|
| Reference distance, representative | 0.141617 | 0.004526 | 0.137092 |
| Reference distance, aggregation | 0.138312 | 0.001221 | 0.137092 |
| Spectral, representative | 0.055229 | 0.018336 | 0.036894 |
| Spectral, aggregation | 0.037721 | 0.000828 | 0.036894 |
The decomposition identifies the source of the gain directly: spectral partitioning retains more query-relevant dependence, reducing , and aggregation improves the coarse fit, reducing . 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 gives worst-query KL error at most .
- 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 , with the following results at :
| States | Method | Symmetric KL | Stored representation | Complete request |
|---|---|---|---|---|
| 4,096 | Exact Sinkhorn | 0.000000 | 128 MiB | 65.091 ms |
| 4,096 | QDSB quantization | 0.039831 | 160 KiB | 21.655 ms |
| 4,096 | Ours | 0.036948 | 160 KiB | 3.122 ms |
| 8,192 | Exact Sinkhorn | 0.000000 | 512 MiB | 300.052 ms |
| 8,192 | QDSB quantization | 0.041290 | 192 KiB | 36.263 ms |
| 8,192 | Ours | 0.039627 | 192 KiB | 3.388 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 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 , where 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 setting | Median TV to independence | Maximum TV | Queries with TV below |
|---|---|---|---|
| Alanine, B=2 | 30/30 | ||
| Chignolin observed, B=16 | 0.179614 | 8/10 | |
| WLALL, B=8 | 0.005181 | 16/30 |
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 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.
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.