Pith.

Machine Learning · arXiv:2611.11649 v1 · submitted 8 Nov 2026 · 56 references

A Trust Region Approach for Learning Schrödinger Bridges

Takeshi Koshizuka, Denis Blessing, Max Zimmer, Sebastian Pokutta, Lorenz Richter

The paper says

“We study the numerical approximation of Schrödinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given.”

more of the abstract

In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. This restriction significantly improves numerical stability. Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. We demonstrate the effectiveness of our approach on several challenging sampling tasks.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. A well-executed, transparent algorithmic contribution with strong empirical results, but the convergence guarantee doesn't cover two of the four tasks, so the theoretical narrative is narrower than it first appears. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 56 references say.

The paper says, at §3.2, Appendix A.10

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.88):

“The iterates with the modified anchor are therefore not those of TR-IPF, and our analysis in Section 3.2 does not cover them.”

Pith says: Objection 1 of 3 · major

The convergence analysis (Proposition 3.4) assumes exact updates and an active constraint D_KL = ε, but the practical algorithm on alanine dipeptide and LJ-13 uses a lower-bounded step size, a modified matching anchor, and weight clipping. The paper explicitly states in Appendix A.10 that the iterates with the modified anchor are not those of TR-IPF and that the analysis does not cover them.

more of objection 1

This gap weakens the theoretical support for the method on two of the four main tasks. The authors should discuss this limitation more prominently in the main text and consider providing additional empirical evidence (e.g., monitoring the variance of importance weights under the modifications) to justify that the trust region still controls variance in practice.

The paper says, at §3.2, Eq. (27)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.71):

“We stress, however, that the following argument is local: it relies on a second-order expansion around d⃗P ui+1 d⃗P ui ≈1 and does not provide a global bound on the variance.”

Pith says: Objection 2 of 3 · major

The bound γ_i ≳ √(2ε/M) and the convergence rate rely on a small-ε expansion and the assumption that Var(ℓ_i) ≤ M uniformly. The paper notes that this is a local argument and that the variance of log-weights is approximately 2ε only at leading order. In early iterations when the constraint may not be active, this approximation could break down. The authors should clarify the regime in which the theoretical guarantees are expected to hold and discuss potential failure modes.

The paper says, at §3.1, Eq. (21)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The dual function L_dual(λ) is estimated via Monte Carlo using N_path trajectories. The paper does not discuss the accuracy of this estimate or its sensitivity to N_path and N_state. A poor estimate could lead to a suboptimal step size. The authors should add a remark on the variance of this estimator and practical guidelines for choosing these parameters.

5 minor objections

The paper says, at §1

Pith says: minor objection 1 of 5

The abstract uses the phrase 'simulation-efficient matching structure' but does not define it; clarify that it refers to the adjoint-matching objective from [30].

The paper says, at §3.1, Eq. (25)

Pith says: minor objection 2 of 5

The notation for the importance weight is slightly ambiguous: it should be made explicit that dP^{u_{i+1}}/dP^{u_i} is normalized and given in Proposition 3.1.

show 3 more

The paper says, at §4.1, Figure 4

Pith says: minor objection 3 of 5

The legend in Figure 4 is shared across subplots but the labels in (d) refer to γ, which is defined in the caption as γ = 1/(1+λ). This is clear but the text should remind the reader that λ is the dual variable.

The paper says, at §4.3

Pith says: minor objection 4 of 5

The description of the prior for alanine dipeptide (2% uniform over φ) is important for reproducibility; consider moving it from Appendix B.4 to the main text or at least referencing it more explicitly.

The paper says, at Appendix A.10, Lemma A.13

Pith says: minor objection 5 of 5

The lemma states that the Schrödinger bridge is a fixed point of both iterations, but the proof relies on the assumption that the exact backward half-step is taken. This is fine for the idealized setting, but in practice the backward step is learned. Add a note that this is a theoretical property under exact updates.

What would settle it

Run TR-SBS on alanine dipeptide (the harder benchmark) without the two modifications (i.e., use the original anchor from Eq. (25) and remove the lower bound on γ_i). Compare C7ax population, KL divergences, and training stability against the modified TR-SBS and against ASBS. If the unmodified version fails (e.g., diverges or misses C7ax), then the theoretical analysis does not cover the working algorithm, confirming the gap. If it still succeeds, the analysis is more broadly applicable than stated.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 10 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic partly agrees 27 s · 51,158 tokens read, 3,260 written
  3. Editor wrote the verdict above 12 s · 52,018 tokens read, 978 written
  4. Referee minor revision: 3 major objections, 5 minor 33 s · 53,471 tokens read, 1,218 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Here is my read on the trust-region Schrödinger bridge paper.

What’s new and good: The paper combines a path-space KL trust region with adaptive step-size selection for IPF and importance-weighted matching. That combination is genuinely novel. The adaptive step size is principled, and the importance-weighted objective is a proper divergence, unlike ASBS. On synthetic benchmarks (Many-Well, GMM), the method clearly mitigates mode collapse where ASBS and its variants fail. The alanine dipeptide results are striking: TR-SBS stably samples a rare conformation that ASBS either drains or diverges on. The authors are also unusually transparent about limitations—they explicitly list where and why the theory does not apply.

Soft spots: That transparency cuts both ways. The stress-test note is on point: the convergence proof (Proposition 3.4) assumes exact updates and an active constraint, but on alanine dipeptide and LJ-13 the practical algorithm uses a modified matching anchor and step-size clipping. As the paper states in Appendix A.10, the resulting iterates are not those of TR-IPF, so the theoretical guarantees do not cover the successful variant. This means that for half the tasks, we don’t know whether the trust region per se or the empirical heuristics are responsible for the stability. That gap weakens the paper’s central narrative—the authors claim TR-SBS mitigates mode collapse, but the theoretical rationale is absent for two of the four experiments. On the other hand, the empirical evidence is strong enough to stand on its own, and the authors do not overclaim. Also, the method is several times more expensive than ASBS in target-gradient evaluations, which is acknowledged but could limit adoption.

Who this is for: Researchers in sampling, generative modeling, and Schrödinger bridges. Practitioners who need stable multimodal sampling from unnormalized densities will find the results useful, though the computational cost is nontrivial.

Recommendation: This paper deserves a serious referee. The empirical contributions are solid, the limitations are honestly stated, and the algorithmic idea is clean. A reviewer should ask the authors to clarify the scope of the theory relative to the practical algorithm, and perhaps to include an analysis of the modified anchor (even a heuristic discussion). But the work is coherent, reproducible, and advances the state of the art. I would accept it after minor revisions addressing that narrative gap.

The strongest objection

The central claim that TR-SBS mitigates mode collapse rests on both empirical validation and a theoretical convergence guarantee (Proposition 3.4). However, that guarantee assumes: (i) each TR-IPF step is computed exactly, (ii) the constraint D_KL = ε is active every iteration, (iii) Var(ℓ_i) ≤ M uniformly. On alanine dipeptide and LJ-13, the practical algorithm deviates: it lower-bounds γ_i, replaces the matching anchor ui(X_T,T) by ui(X_t,t), and normalizes/clips importance weights. The paper explicitly states (Appendix A.10) that iterates with the modified anchor are not those of TR-IPF and that the convergence analysis does not cover them. Thus, for two of the four tasks, the method that actually succeeds is not the one whose properties are proven. The empirical results are strong, but the theoretical rationale for mode-collapse mitigation on these tasks is absent; the success could depend on the heuristics rather than the trust region per se. This gap weakens the paper's main narrative and leaves the central claim partly unexplained.

The reader's weakest_assumption already notes that the convergence analysis is idealized and not verified in practice, and that the practical algorithm includes modifications not covered by the analysis. My stress-test sharpens this: the modifications are not minor implementation details—they change the update's fixed point and invalidate the theoretical guarantees for the two most challenging tasks. The paper is transparent about this, but the central claim ('mitigates mode collapse') is less supported theoretically than it appears from Proposition 3.4 alone. The empirical evidence remains strong, so the verdict should stay ACCEPT; however, the paper would benefit from either extending the analysis or acknowledging that on these tasks the trust region's role is heuristic. I choose UNCHANGED because the paper's own limitations discussion sufficiently addresses the concern.

The referee's report

If the results hold, TR-SBS offers a practically useful method for sampling multimodal distributions without mode collapse, addressing a known instability in IPF-based samplers. The theoretical analysis (Proposition 3.4) gives a convergence rate for the idealized iteration, and the empirical results on challenging benchmarks are strong. The paper isolates the contributions of the trust region and importance weighting through ablation studies.

Pith's reading of the paper (a machine's summary, not the authors')

The paper proposes TR-SBS, a variant of iterative proportional fitting (IPF) for solving Schrödinger bridges when only the unnormalized target density is available. It constrains each forward update to lie within a fixed Kullback–Leibler divergence of the previous iterate, which automatically adjusts the mirror-descent step size and stabilizes training. The method also incorporates the importance weight that the forward KL objective requires. Empirical results on synthetic benchmarks and molecular systems show that TR-SBS mitigates mode collapse while retaining the simulation-efficient matching structure of previous approaches.

What carries the argument. The key machinery is the trust-region-constrained forward update of the IPF iteration: instead of a fixed unit step, the step size γ_i = 1/(1+λ_i) is determined adaptively by solving a one-dimensional dual problem that enforces D_KL($P^{{u_{i+1}}$} | $P^{{u_i}}$) = ε. This converts the IPF step into a geometric average of the current and target path measures, and the resulting control is learned via an importance-weighted adjoint matching objective.

Core claim. The central discovery is that combining a trust-region constraint on the path-space KL divergence with importance-weighted matching yields a stable iteration for Schrödinger bridges. The trust region enforces a small, controlled step size at each outer iteration, preventing the mode collapse that plagues fixed-step IPF variants, while the importance weight ensures the matching objective remains a proper divergence that uses the target density itself rather than only its gradient.

The premise it rests on. The convergence analysis (Proposition 3.4) assumes that at every iteration the trust region constraint is exactly active and the conditional variance of the log-ratio ℓ_i is uniformly bounded—conditions not verified in practice, and the practical algorithm deviates from them by clipping weights and bounding step sizes.

If the paper is right (Pith's inferences, not the authors')

TR-SBS achieves lower Earth Mover's Distance and log-partition error than ASBS on the Many-Well potential with up to 2^32 modes.

On high-dimensional Gaussian mixture models, TR-SBS attains lower Sinkhorn divergence and total variation distance of weights than both ASBS and diffusion-based samplers PIS and DIS.

On alanine dipeptide, TR-SBS stably trains for 120,000 epochs and maintains the population of the rare C7ax conformer near the reference, whereas ASBS diverges or drains the conformer.

The trust region radius ε controls a stability–speed trade-off: smaller ε stabilizes updates but requires more outer iterations.

The adaptive step-size selection could be lifted to other iterative proportional fitting settings beyond Schrödinger bridges, such as entropy-regularized optimal transport with a continuous-time Sinkhorn algorithm.

The reliance on an exact backward half-step and the reciprocal structure means that the theoretical convergence guarantee may not hold when the backward step is learned from finite samples, as acknowledged by the authors.

The trust region does not fully resolve the stagnation–variance dilemma when the variance of the log-ratio ℓ_i is very large; the practical fix of bounding γ_i from below gives up variance control, suggesting that additional variance-reduction techniques could be beneficial.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

We study the numerical approximation of Schrödinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given. In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. This restriction significantly improves numerical stability. Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. We demonstrate the effectiveness of our approach on several challenging sampling tasks.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-10

The authors' main claim, as the classifier picked it out 0.56

“This restriction significantly improves numerical stability.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.14 · human participants 0.02 · controlled comparison 0.25 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.76 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
We study the numerical approximation of Schr\"{o}dinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given. background 0.74 0.61 0.04 asserted 0.90
In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. background 0.91 0.13 0.04 asserted 0.85
To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. method 0.81 0.88 0.21 asserted 0.99
This restriction significantly improves numerical stability. result 0.99 0.88 0.30 asserted 1.00
Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. interpretation 0.97 0.92 0.13 asserted 1.00
We demonstrate the effectiveness of our approach on several challenging sampling tasks. result 0.91 0.94 0.32 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.90

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-10: 1 request, 4,889 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (12)
Figure 2 of the paper
Figure 2. Overview of IPF-based samplers. The columns differ in how the mirror descent step size of the forward update is chosen. The rows show the path space update, which reweights the current iterate P⃗ ui by w γ i (Proposition 3.1), its implementation by matching objectives learned from samples of P⃗ ui with this importance weight (IW), and the variants without it. reduces to TR-IPF; we discuss deviations from this setting in Section 5. To this end, let P∗ minimize F(P) := DKL(PT |ptarget) over path measures with P0 = pprior. As stated in Section 2.2, IPF is mirror descent on F with unit s
Figure 3 of the paper
Figure 3. Results on Many-Well potential. We evaluate robustness to increasing modes and robustness to mode distance. (a, b) Performance vs. number of modes. (c, d) Performance vs. mode distance. Our proposed method (TR-SBS, red) achieves the lowest estimation error and marginal distance for large m and δ, handling both combinatorial complexity and high energy barriers. TR-SBS γ=1.0 (ASBS-IW) γ=0.7 γ=0.4 γ=0.1 4 8 16 32 m 0.0 0.4 0.8 1.2 EMD (a) EMD (vs. m) 0 10 20 30 40 50 Iteration i 0.06 0.12 0.18 EMD (b) EMD 0 10 20 30 40 50 Iteration i 50 75 100 Mean energy (c) Mean energy 0 10 20 30 40…
Figure 4 of the paper
Figure 4. Ablation study on the step size γ. Legend is shared across all subplots. (a) Adaptive TR-SBS matches the robustness of conservative γ = 0.1. (b, c) Dynamics reveal the trade-off: Small γ = 0.1 is stable (b) but slow (c). TR-SBS converges efficiently (c) while maintaining stability (b). (d) The adaptive γ starts conservative to prevent collapse and increases to accelerate convergence. To isolate the contributions of the two ingredients of TR-SBS, the trust region and the importance weight, we compare TR-SBS with the original ASBS [30] and two ablations that each add one of them to…
Figure 5 of the paper
Figure 5. Results on GMM (mean ± standard deviation over three seeds). (a, b) Performance vs. dimension d ∈ {10, 25, 50, 100} with K = 10. (c, d) Performance vs. number of components K ∈ {5, 10, 20, 30, 40} with d = 10. TR-SBS (red) attains the lowest SD and TVD in every setting. PIS and DIS use 128 instead of 200 time steps and a different training budget (Appendix B.6). Stability: Figure 4a shows that aggressive fixed step sizes (γ ≥ 0.4) fail to prevent mode collapse as combinatorial complexity increases (m → 32). A conservative step size (γ ≈ 0.1) is required to maintain low error, a…
Figure 6 of the paper
Figure 6. Alanine dipeptide. Left: Ramachandran plots of the reference, of TR-SBS after
Figure 7 of the paper
Figure 7. Additional metrics on GMM (mean ± standard deviation over three seeds). (a)-(c): K = 10, varying d. (d)-(f): d = 10, varying K. Lower is better for W2 (a, d) and sliced W2 (b, e); higher is better for the raw path ESS (c, f). C Additional experimental results C.1 Additional metrics on GMM
Figure 8 of the paper
Figure 8. Samples on GMM with d = 10 and K = 40 (first two coordinates, seed 0). Each panel shows the 4,000 evaluation samples of the checkpoint with the lowest SD during training. Blue points show generated samples; contours show the target marginal log density. All panels use identical contours and coordinate limits and show every sample. PIS and DIS use 128 instead of 200 time steps (Appendix B.6). (rightmost column) are consistently shifted towards higher values compared to the ground truth, confirming that no method can reach the optimal solution. • Success with TR-SBS at Large Noise…
Figure 9 of the paper
Figure 9. Visualization at small noise (σmax = 0.1). Rows (a)–(d) show the four methods; columns show the marginal histograms of dimensions 0–2 and the energy distribution. Due to insufficient noise, all methods (a-d) fail to reach the wells (underfitting). Impact of the maximum noise scale (σmax) A larger σmax yields a smoother optimization land￾scape, making the problem more tractable and significantly accelerating convergence. However, in the context of IPF which approximates the objective using finite samples, an excessively large σmax induces mode collapse. We demonstrate that TR-SBS…
Figure 10 of the paper
Figure 10. Visualization at large noise (σmax = 0.2). Rows and columns as in
Figure 11 of the paper
Figure 11. Ablation study on the maximum noise scale σmax. Legend is shared across subplots (a-c). (a) Large σmax = 0.2 causes mode collapse in ASBS (blue) as barrier distance δ increases, whereas TR-SBS (red) remains robust. (b) Training dynamics confirm that ASBS diverges in EMD while TR-SBS remains stable. (c) A larger σmax leads to faster convergence. (d) Histograms confirm that ASBS collapses to a single mode under large noise, while TR-SBS maintains the global structure. We use d = 16, m = 5, and δ = 6; all other settings are in Appendix B.3. The bistable coordinates induce multiple…
Figure 12 of the paper
Figure 12. Alanine dipeptide: populations of C7ax (left) and

Showing 11 of the paper's 12 figures. Figures are the authors', reproduced from the paper.

References (56)
  1. [1] A. Abdolmaleki, R. Lioutikov, N. Lau, L. P. Reis, J. Peters, and G. Neumann. Model-based relative entropy stochastic search.Advances in Neural Information Processing Systems, 28, 2015 on Pith
  2. [2] S.-i. Amari.Information geometry and its applications. Springer, 2016 on Pith
  3. [3] P.-C. Aubin-Frankowski, A. Korba, and F. Léger. Mirror descent with relative smoothness in measure spaces, with application to Sinkhorn and EM.Advances in Neural Information Processing Systems, 35:17263–17275, 2022 on Pith
  4. [4] J. Berner, L. Richter, M. Sendera, J. Rector-Brooks, and N. Malkin. From discrete-time policies to continuous-time diffusion samplers: Asymptotic equivalences and faster training. Transactions on Machine Learning Research, 2026 on Pith
  5. [5] J. Berner, L. Richter, and K. Ullrich. An optimal control perspective on diffusion-based generative modeling.Transactions on Machine Learning Research, 2024 on Pith
  6. [6] E. Bernton, J. Heng, A. Doucet, and P. E. Jacob. Schrödinger bridge samplers.arXiv preprint arXiv:1912.13170, 2019 on PithPith's reviewarXiv
  7. [7] D. Blessing, J. Berner, L. Richter, C. Domingo-Enrich, Y . Du, A. Vahdat, and G. Neumann. Trust region constrained measure transport in path space for stochastic optimal control and inference. InAdvances in Neural Information Processing Systems, volume 38, pages 165462–165510, 2025 on Pith
  8. [8] D. Blessing, J. Berner, L. Richter, and G. Neumann. Underdamped diffusion bridges with applications to sampling. InInternational Conference on Learning Representations, 2025 on Pith
  9. [9] D. Blessing, L. Richter, J. Berner, E. Malitskiy, and G. Neumann. Bridge matching sampler: Scalable sampling via generalized fixed-point diffusion matching. InInternational Conference on Machine Learning, pages 8496–8541. PMLR, 2026 on Pith
  10. [10] P. G. Bolhuis, C. Dellago, and D. Chandler. Reaction coordinates of biomolecular isomerization. Proceedings of the National Academy of Sciences, 97(11):5877–5882, 2000 on Pith
  11. [11] J. Chen, L. Richter, J. Berner, D. Blessing, G. Neumann, and A. Anandkumar. Sequential controlled Langevin diffusions. InInternational Conference on Learning Representations, 2025 on Pith
  12. [12] T. Chen, G.-H. Liu, and E. A. Theodorou. Likelihood training of Schrödinger bridge using forward-backward SDEs theory. InInternational Conference on Learning Representations, 2022 on Pith
  13. [13] J. Choi, Y . Chen, M. Tao, and G.-H. Liu. Non-equilibrium annealed adjoint sampler. In Advances in Neural Information Processing Systems, volume 38, pages 90956–90988, 2025 on Pith
  14. [14] A. R. Conn, N. I. M. Gould, and P. L. Toint.Trust region methods. SIAM, 2000 on Pith
  15. [15] P. Dai Pra. A stochastic control approach to reciprocal diffusion processes.Applied Mathematics and Optimization, 23(3):313–329, 1991 on Pith
  16. [16] V . De Bortoli, J. Thornton, J. Heng, and A. Doucet. Diffusion Schrödinger bridge with applications to score-based generative modeling.Advances in neural information processing systems, 34:17695–17709, 2021. 11 on Pith
  17. [17] C. Domingo-Enrich, M. Drozdzal, B. Karrer, and R. T. Q. Chen. Adjoint matching: Fine- tuning flow and diffusion generative models with memoryless stochastic optimal control. In International Conference on Learning Representations, 2025 on Pith
  18. [18] C. Domingo-Enrich, J. Han, B. Amos, J. Bruna, and R. T. Q. Chen. Stochastic optimal control matching.Advances in Neural Information Processing Systems, 37:112459–112504, 2024 on Pith
  19. [19] R. Fortet. Résolution d’un système d’équations de M. Schrödinger.Journal de mathématiques pures et appliquées, 19(1-4):83–105, 1940 on Pith
  20. [20] A. Genevay, G. Peyré, and M. Cuturi. Learning generative models with Sinkhorn divergences. InInternational Conference on Artificial Intelligence and Statistics, pages 1608–1617. PMLR, 2018 on Pith
show 36 more
  1. [21] V . L. Ginzburg and L. D. Landau. On the theory of superconductivity. InOn superconductivity and superfluidity: a scientific autobiography, pages 113–137. Springer, 2009 on Pith
  2. [22] G. A. Gottwald and S. Reich. Localized Schrödinger bridge sampler.Journal of Computational Physics, 548:114583, 2026 on Pith
  3. [23] R. Gruhlke, J. Berner, D. Sommer, and L. Richter. Tensor train diffusion: Leveraging low-rank structures for high-dimensional score-based sampling. InInternational Conference on Machine Learning, pages 36905–36941. PMLR, 2026 on Pith
  4. [24] C. Hartmann and L. Richter. Nonasymptotic bounds for suboptimal importance sampling. SIAM/ASA Journal on Uncertainty Quantification, 12(2):309–346, 2024 on Pith
  5. [25] C. Hartmann, L. Richter, C. Schütte, and W. Zhang. Variational characterization of free energy: Theory and algorithms.Entropy, 19(11):626, 2017 on Pith
  6. [26] A. J. Havens, B. K. Miller, B. Yan, C. Domingo-Enrich, A. Sriram, D. S. Levine, B. M. Wood, B. Hu, B. Amos, B. Karrer, X. Fu, G.-H. Liu, and R. T. Q. Chen. Adjoint sampling: Highly scalable diffusion samplers via adjoint matching. InInternational Conference on Machine Learning... on Pith
  7. [27] J. Huang, Y . Jiao, L. Kang, X. Liao, J. Liu, and Y . Liu. Schrödinger-Föllmer sampler.IEEE Transactions on Information Theory, 71(2):1283–1299, 2025 on Pith
  8. [28] M. R. Karimi, Y .-P. Hsieh, and A. Krause. Sinkhorn flow as mirror flow: A continuous-time framework for generalizing the Sinkhorn algorithm. InInternational Conference on Artificial Intelligence and Statistics, pages 4186–4194. PMLR, 2024 on Pith
  9. [29] C. Léonard. A survey of the Schrödinger problem and some of its connections with optimal transport.Discrete and Continuous Dynamical Systems, 34(4):1533–1574, 2014 on Pith
  10. [30] G.-H. Liu, J. Choi, Y . Chen, B. K. Miller, and R. T. Q. Chen. Adjoint Schrödinger bridge sampler. InAdvances in Neural Information Processing Systems, volume 38, pages 15673–15708, 2025 on Pith
  11. [31] G.-H. Liu, Y . Lipman, M. Nickel, B. Karrer, E. A. Theodorou, and R. T. Q. Chen. Generalized Schrödinger bridge matching. InInternational Conference on Learning Representations, 2024 on Pith
  12. [32] L. I. Midgley, V . Stimper, G. N. C. Simm, B. Schölkopf, and J. M. Hernández-Lobato. Flow annealed importance sampling bootstrap. InInternational Conference on Learning Representa- tions, 2023 on Pith
  13. [33] J. Nam, B. Máté, A. P. Toshev, M. Kaniselvan, R. Gómez-Bombarelli, R. T. Q. Chen, B. Wood, G.-H. Liu, and B. K. Miller. Enhancing diffusion-based sampling with molecular collective variables. InInternational Conference on Learning Representations, 2026 on Pith
  14. [34] N. Nüsken and L. Richter. Solving high-dimensional Hamilton-Jacobi-Bellman PDEs using neural networks: perspectives from the theory of controlled diffusions and measures on path space.Partial Differential Equations and Applications, 2(4):48, 2021. 12 on Pith
  15. [35] F. Otto, P. Becker, N. A. Vien, H. C. Ziesche, and G. Neumann. Differentiable trust region layers for deep reinforcement learning. InInternational Conference on Learning Representations, 2021 on Pith
  16. [36] R. Ouyang, L. Grenioux, and J. M. Hernández-Lobato. A diffusive classification loss for learning energy-based generative models. InInternational Conference on Machine Learning, pages 95197–95252. PMLR, 2026 on Pith
  17. [37] S. Peluchetti. Diffusion bridge mixture transports, Schrödinger bridge problems and generative modeling.Journal of Machine Learning Research, 24(374):1–51, 2023 on Pith
  18. [38] S. Peluchetti. Non-denoising forward-time diffusions.arXiv preprint arXiv:2312.14589, 2023 on PithPith's reviewarXiv
  19. [39] J. Peters, K. Mülling, and Y . Altun. Relative entropy policy search. InProceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 1607–1612, 2010 on Pith
  20. [40] L. Richter and J. Berner. Improved sampling via learned diffusions. InInternational Conference on Learning Representations, 2024 on Pith
  21. [41] B. Rosenstein and D. Li.Ginzburg–Landau Theory of Condensates. Cambridge University Press, 2021 on Pith
  22. [42] L. Rüschendorf. Convergence of the iterative proportional fitting procedure.The Annals of Statistics, 23(4):1160–1174, 1995 on Pith
  23. [43] V . G. Satorras, E. Hoogeboom, and M. Welling. E(n) equivariant graph neural networks. In International Conference on Machine Learning, 2021 on Pith
  24. [44] E. Schrödinger. Über die Umkehrung der Naturgesetze.Sitzungsberichte der Preussischen Akademie der Wissenschaften, Physikalisch-mathematische Klasse, pages 144–153, 1931 on Pith
  25. [45] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz. Trust region policy optimization. InInternational conference on machine learning, pages 1889–1897. PMLR, 2015 on Pith
  26. [46] Y . Shi, V . De Bortoli, A. Campbell, and A. Doucet. Diffusion Schrödinger bridge matching. Advances in Neural Information Processing Systems, 36:62183–62223, 2023 on Pith
  27. [47] Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations, 2021 on Pith
  28. [48] J. Sun, J. Berner, L. Richter, M. Zeinhofer, J. Müller, K. Azizzadenesheli, and A. Anand- kumar. Dynamical measure transport and neural PDE solvers for sampling.arXiv preprint arXiv:2407.07873, 2024 on PithPith's reviewarXiv
  29. [49] D. Thalmeier, H. J. Kappen, S. Totaro, and V . Gómez. Adaptive smoothing for path integral control.Journal of Machine Learning Research, 21(191):1–37, 2020 on Pith
  30. [50] F. Vargas. Machine-learning approaches for the empirical Schrödinger bridge problem. Techni- cal report, University of Cambridge, Computer Laboratory, 2021 on Pith
  31. [51] F. Vargas, W. Grathwohl, and A. Doucet. Denoising diffusion samplers. InInternational Conference on Learning Representations, 2023 on Pith
  32. [52] F. Vargas, S. Padhy, D. Blessing, and N. Nüsken. Transport meets variational inference: Controlled Monte Carlo diffusions. InThe Twelfth International Conference on Learning Representations, 2024 on Pith
  33. [53] F. Vargas, P. Thodoroff, A. Lamacraft, and N. Lawrence. Solving Schrödinger bridges via maximum likelihood.Entropy, 23(9):1134, 2021 on Pith
  34. [54] G. Wang, Y . Jiao, Q. Xu, Y . Wang, and C. Yang. Deep generative learning via Schrödinger bridge. InInternational conference on machine learning, pages 10794–10804. PMLR, 2021. 13 on Pith
  35. [55] H. Wu, J. Köhler, and F. Noé. Stochastic normalizing flows.Advances in neural information processing systems, 33:5933–5944, 2020 on Pith
  36. [56] Q. Zhang and Y . Chen. Path integral sampler: A stochastic control approach for sampling. In International Conference on Learning Representations, 2022. 14 Contents of the appendix A Proofs and additional propositions 16 A.1 Notation . . . . . . . . . . . . . . . . . . . . . .... on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment