acceptodds
Under review as a conference paper at ICLR 2027

SPIM: Solving Schrödinger Bridges Via Alternating Stochastic Optimal Control

Abstract

Schrödinger bridges (SBs) and stochastic optimal control (SOC) are closely related, but the two-endpoint distribution constraints of SBs make terminal-cost SOC solvers difficult to apply directly. We propose Schrödinger Potential Iteration Matching (SPIM), which solves SBs through alternating SOC that realize the fixed-point iteration of Schrödinger potentials. In each direction, the terminal cost is given by the corresponding relative endpoint energy minus the initial-time value returned by the opposite direction SOC. Under the Cole–Hopf transform, each exact SOC performs one Schrödinger-potential propagation, so a forward–backward pair completes one fixed-point iteration. Under compactness and uniform kernel positivity, exact alternating SOC inherits the Hilbert-metric contraction of the classical potential iteration, yielding geometric convergence of the SOC value functions up to additive constants. The limiting value functions determine the unique optimal generalized Schrödinger bridge. The framework supports energy-to-energy transport from relative endpoint energies without endpoint samples or normalizing constants, extends naturally to generalized Schrödinger bridges with running state costs, and allows different SOC solvers in the two directions. Experiments show that SPIM achieves competitive endpoint and interior transport accuracy across classical and generalized SB benchmarks, scales to higher-dimensional problems, and supports both running-cost and two-endpoint reward fine-tuning.

Primary area:
generative models
Keywords:
Schrödinger Bridge Problem, Stochastic Optimal Control

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

This article introduces the "Schrödinger Potential Iteration Matching" (SPIM) method, which solves generalized Schrödinger bridge problems by alternating between forward and backward stochastic optimal control (SOC) problems. In this approach, each terminal cost function combines the associated relative terminal energy with the negative initial value of the corresponding opposing SOC problem. Through the Cole-Hopf transformation, these update steps implement the classical Schrödinger potential iteration.

Critical strengths
  1. Appendix B establishes a contraction on a complete space of anchored log-potentials and subsequently demonstrates, via a mass balance argument, that the projective fixed point possesses an eigenvalue of one. The final identity $ J(P)-J(P^\star)=\mathrm{KL}(P\Vert P^\star) $ establishes global optimality within the set of admissible path distributions.

  2. The exchange of scalar boundary values ​​creates a clear interface between direction-based solution methods. Incorporating running costs while simultaneously allowing for the subsequent weighting of endpoints—all within a single, unified formulation—is particularly advantageous in practice.

  3. PI-VM achieves the lowest endpoint error documented to date across all four classical benchmarks, while TR-AM yields strong results for specific problem settings that include running costs.

Critical weaknesses
  1. The theory assumes a single reference law and its corresponding time-reversal. Appendix E.3 states that Brownian reference processes are used for synthetic experiments; however, details regarding their initial distribution, marginal densities, backward drift, and the resulting relative endpoint energies are missing.

  2. The assumptions concerning compact kernels that are uniformly positive do not directly cover the Euclidean experiments. Appendix C requires uniform accuracy of the marginal values ​​(up to constants), rather than merely a low empirical training loss. For the neural solvers, however, this accuracy is neither demonstrated nor measured.

  3. The primary SPIM benchmarks utilize endpoint samples augmented via Gaussian noise for training. The "Two-Dynamic-Bank" ablation study, which starts from Gaussian-distributed particles, offers an interesting indication but is limited to tracking sinusoidal trajectories. Consequently, this does not substantiate robust energy-to-energy training for more complex, multimodal, or higher-dimensional examples.

  4. Table 8 indicates that SPIM employs 180,000 updates for the classical tasks, whereas DSBM and D2E are allocated 90,000 updates each. Furthermore, SPIM uses a batch size of 3,200, compared to 512 and 128, respectively. During fine-tuning for images, 60,000 SPIM updates are used, versus 12,000 AM updates. A homotopy is employed by SPIM only in the "Spherical-Meridian" benchmark. Given these discrepancies, it is difficult to definitively attribute the reported performance gains to the marginal value coupling. Comparisons based on equivalent computational effort are required, as are performance-over-time plots and a more robust AM baseline trained with the same budget.

Other unresolved issues include the broken reference “Table ??” in Section 4.2, as well as inconsistent notation for PI-VM/PIVM and TR-PI-VM/TR-PIVM. Figure 3 lacks numerical axis labels, which precludes a quantitative interpretation of the cost improvements achieved. The implementation description should specify the number of seeds used, the variance of the Gaussian augmentation, the values ​​for the “control caps,” and the schedule for updating the “trust regions.” Furthermore, providing explicit pseudocode for the neural Jacobi implementation would help distinguish it from the sequential algorithm analyzed in the main text.

Questions for authors
  1. What specific forward and backward reference dynamics, reference endpoint densities, and relative terminal energies were used in each of the synthetic experiments? If both reference drifts are zero, what reference measurement convention justifies this?

  2. Can you provide the explicit scalar-valued training objectives for AM, TR-AM, and TR-PI-VM, and measure their boundary-value errors—modulo constants—using an analytically solvable problem?

  3. Do the significant improvements persist under comparable computational costs, particularly when AM undergoes 60,000 image fine-tuning updates and competing spherical meridian methods employ similar continuation strategies?

  4. Can you describe a complete energy-to-energy (E2E) experiment involving multimodal or higher-dimensional endpoints, in which no endpoint samples were used for training or database initialization, and where the information access for D2E is comparable?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Rykg

We thank the reviewer for the suggestions. We clarify the reference, training objectives, theoretical scope, and implementation, and provide error measurements and additional experiments.

Reference processes and relative endpoint energies (W1, Q1)

Response: The synthetic experiments use a σ\sigma-finite Brownian reference initialized with Lebesgue measure, R0(dx)=dxR_0(dx)=dx. With T=1T=1, the forward and time-reversed reference dynamics are

d\bar X_s=\sigma\,d\bar W_s,\qquad s=1-t.$$ The reference marginal density is $r_t(x)=1$ at every time, so both reference drifts are zero. The task dimensions and diffusion coefficients are: - Gaussian transport (5D, 20D, and 50D), Ring-mixture transport (2D), Double-well transport (16D), and Banana transport (2D): $\sigma=\sqrt{2}$. - Sinusoidal tracking, Bimodal splitting, and Moving double well (all 2D), and Spherical-meridian transport (3D): $\sigma=1$. - In the additional energy-to-energy comparisons, Ring-mixture transport uses $\sigma=0.3$ and Double-well transport uses $\sigma=\sqrt{2}$. Both use the time interval $[0,1]$. The target endpoint energies are $E_i(x)=-\log\mu_i(x)+\mathrm{const}$ for each task’s prescribed endpoint densities. The relative endpoint energies are $\widetilde E_i=E_i+\log r_i=E_i$, up to additive constants. The Lebesgue reference measure is distinct from the finite particle banks used during training. ### **Training objectives, boundary-value errors, and scope of the theory (W2, Q2)** **Response:** The exact theorem requires compactness and uniformly positive kernels, so it does not directly cover the Euclidean experiments. Appendix C assumes accurate boundary values, which low empirical loss alone cannot guarantee. We will clarify this scope in the revision (Sec. 3.3, Appendices A and C). All three solvers use a scalar network $V_\theta$ to parameterize $u_\theta=-\sigma^\top\nabla V_\theta$. Their training objectives are $$\begin{aligned} \mathcal L_{\mathrm{AM}}&=\mathbb E_{\mathcal D}\|u_\theta-y_{\mathrm{AM}}\|^2, \end{aligned}$$ $$\begin{aligned} \mathcal L_{\mathrm{TR\text{-}AM}}&=\tfrac12\mathbb E_{\mathcal D} [w_\alpha\|u_\theta-y_\alpha\|^2], \end{aligned}$$ $$\begin{aligned} \mathcal L_{\mathrm{TR\text{-}PI\text{-}VM}}&=\mathbb E_{\mathcal D} [(V_\theta-\operatorname{sg}[Y_\alpha])^2/s_{\mathrm{out}}^2]. \end{aligned}$$ Here, $y_{\mathrm{AM}}=-\sigma^\top A$ is the adjoint control label, and $y_\alpha=(1-\alpha)a-\sigma^\top A^\alpha$ includes a correction for the frozen anchor control $a$. The conditional weights $w_\alpha$ are normalized within each group sharing the same starting point, $Y_\alpha$ is the path-integral value target with anchor and likelihood-ratio corrections, and $s_{\mathrm{out}}$ is the output normalization scale. **Boundary-value errors.** Using saved analytical evaluations, we measure forward RMSE at $t=0$ and backward RMSE at local $s=0$ (physical $t=1$), removing a separate optimal additive constant in each direction: | Task | Method | Forward boundary RMSE ↓ | Backward boundary RMSE ↓ | | --- | --- | --- | --- | | Bimodal splitting | SPIM (TR-PI-VM) | 0.015514 ± 0.006215 | 0.023286 ± 0.009575 | | Bimodal splitting | DeepGSB | 0.022057 ± 0.006800 | 0.055967 ± 0.011440 | | Moving double well | SPIM (TR-PI-VM) | 0.071434 ± 0.026987 | 0.074099 ± 0.040877 | | Moving double well | DeepGSB | 0.080196 ± 0.023070 | 0.079350 ± 0.07976 | | | | | | ### **Experimental validation of energy-to-energy capability (W3, Q4)** **Response:** See “Experimental validation of energy-to-energy capability” in the General response for the shared reply and results. ### **Additional comparisons with matched training budgets (W4, Q3)** **Response:** See “Additional comparisons with matched training budgets” in the General response for the shared reply and results. ### **Implementation details and presentation (other issues)** **Response:** The missing reference is fixed (line 418, Sec. 4.2, Table 10). PI-VM/PIVM and TR-PI-VM/TR-PIVM denote the base and trust-region solvers, which are different algorithms. We will add numerical ticks and units or dimensionless conventions to the time, objective, and ground-MSE axes in Figure 3 (Sec. 4.3). The revision specifies seed counts (Appendix E.3.1), Gaussian augmentation variance and control caps (Appendix E.2.3), trust-region schedules (Appendix E.2.2), and neural Jacobi pseudocode (Appendix E.2.1, Algorithm 2).
Reviewer_5bVSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an approach for solving Schrödinger Bridge (SB) and Generalized SB (GSB) problems by iteratively solving half-bridge stochastic optimal control (SOC) problems. Specifically, the method alternates between forward and backward SOC problems, each SOC optimization updates a boundary value function, which is subsequently used to define the terminal cost of the opposite problem. This alternating procedure updates the forward and backward potentials, which jointly characterize a single bridge satisfying both marginal constraints. The authors further establish uniqueness of the solution and provide convergence guarantees. Finally, they evaluate their approach on synthetic SB and GSB problems, LiDAR terrain, and MNIST-EMNIST setups, additionally ablating particle-bank types and SOC solvers.

Critical strengths

[S1.] The paper is mostly clearly written and easy to read. It introduces all required information in the background, explicitly positions the method in the related literature (Table 1), formally states all used assumptions (Appendix A).

[S2.] The method itself is simple yet elegant. It has strong connections with Iteration Proportional Fitting (IPF) literature, it is based on the known relation between SOC and Generalized Schr\”odinger Bridges (GSB). Thus, this fact makes it more believable in the correctness of the method.

Critical weaknesses

Below I list my two main concerns:

[W.1] Lack of compelling motivation for SPIM. The authors motivate the proposed method primarily as a new way of finding the SB (see first paragraph of Introduction). However, the paper does not sufficiently establish the relevance of pursuing this alternative approach, as it remains unclear which limitations of existing SB/SOC methods it aims to address or what practical advantages it could offer. This concern is further amplified by the limited discussion of the experimental results (see [W.2]) and the lack of applications that could particularly benefit from SPIM, as the experimental section mostly relies on standard datasets.

Additionally, the following two experimental setups do not seem to provide much motivation from an application perspective: (1) distributions specified by energy functions and (2) running state costs. For the first, the paper provides insufficient explanation of the motivation and potential applications of this setting, making its practical relevance difficult to assess, particularly for readers less familiar with this literature (as me). For the second, while running state costs are relevant to applications such as navigation, the comparisons with GSBM and TSBM are mostly limited to synthetic problems, and the LiDAR experiments only consider GSBM.

[W.2] As a consequence of [W.1], I don’t understand the purpose of the experiments conducted in this paper (except for the ablation studies). This issue is further worsened by the limited discussion of obtained results. Specifically, Sections 4.2, 4.3 just report the results without sufficiently analyzing or interpreting them. As a result, the discussion is mostly about the sanity of checking SPIM or comparing to the baselines, rather than what experiments actually demonstrate about the method and its possible benefits. Moreover, the metrics on the GSB (sinusoidal tracking, bimodal splitting, moving double well, and spherical-meridian transport) and High-Dimensional Benchmarks mostly are comparable or lower than baselines for both endpoint and interior distributions. Next, while I understand the low number of SOC solvers in main experiments, the ablation study could test more of them. Also, missing one table (line 418).

In summary, I found the paper well organized, with its main ideas clearly presented. However, the proposed method is not sufficiently motivated, as it remains unclear which specific limitations of existing SB/SOC methods it aims to address. This concern is further amplified by the experimental evaluation, which does not convincingly demonstrate the practical advantages of SPIM over existing approaches. Therefore, I recommend weak rejection for the current state of the paper.

Questions for authors

None, refer to weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5bVS

We thank the reviewer for the suggestions. We clarify SPIM’s motivation and applications, expand the experimental discussion, and add SOC solver ablations.

Motivation and applications (W1)

Response: A terminal-cost SOC solver fixes one endpoint and optimizes a given terminal cost, without enforcing a prescribed distribution at the other endpoint. SPIM exchanges boundary values between two SOC problems to update both terminal costs and enforce two endpoint constraints through an outer Schrödinger-potential iteration.

This reuses and combines existing SOC solvers for relative energies, running costs, and rewards. The DSBM/GSBM implementations compared here require endpoint samples, while D2E’s energy-to-energy formulation handles zero running cost. SPIM offers:

  1. Reduced reliance on endpoint samples. Relative energies specify both distributions without endpoint samples or normalizing constants, supporting targets that are difficult to sample.
  2. Path optimization under endpoint constraints. A reference transport can be refined while preserving both endpoint constraints by optimizing reference deviation plus running costs.
  3. Coordinating rewards at both endpoints. Boundary exchange coordinates both directions of a pretrained reference to connect high-reward regions at both endpoints.

The revision clarifies these limitations, applications, and links to the experiments (Sec. 3.4, Appendix E.1).

Rationale for the experimental settings (W1)

Response: Energy-defined endpoints can model transitions between conformations or modes in biology and materials. For a pretrained reference, rewards define relative energies, μi(dx)∝eλri(x)νi(dx)\mu_i(dx)\propto e^{\lambda r_i(x)}\nu_i(dx) and E~i=−λri\widetilde E_i=-\lambda r_i (up to constants), enabling two-endpoint reward fine-tuning without endpoint density models. MNIST–EMNIST tests this setting.

LiDAR tests terrain-aware path optimization under endpoint constraints. TSBM produced no valid results in our tests and was excluded from the original quantitative comparison.

Discussion of experimental results (W2)

Response: The revised discussion (Secs. 4.2–4.4) highlights:

  1. Classical bridges. PI-VM has the lowest mean endpoint error in all four tasks, showing consistent endpoint matching across high-dimensional and multimodal targets.
  2. Generalized bridges. SPIM has the lowest mean interior errors on tracking and the moving well, and second-lowest on splitting and spherical transport, showing flexible path control across geometries with running costs.
  3. LiDAR. SPIM improves forward path cost and ground MSE over GSBM on Mount St. Helens and Sunset Crater, demonstrating practical terrain-aware path control.
  4. Two-endpoint rewards. On MNIST–EMNIST, SPIM improves over independent AM on reward in 8/9 settings and attribute W1W_1 in 5/9, supporting coupled reward coordination at both endpoints.

SOC solver ablation (W2)

Response: We added REACT-VM and tested all nine directional solver combinations (Sec. 4.5, Appendix E.7, Table 12), using 30 bidirectional iterations, 2,000 updates per direction per iteration, and three seeds. AM → AM has the lowest mean endpoint error, while AM → REACT-VM has the lowest mean interior-path error. These results show that solver choice affects endpoint and path accuracy within SPIM.

Forward → backwardEndpoint W22↓W_2^2\downarrowInterior path W22↓W_2^2\downarrow
PI-VM → PI-VM †0.023 ± 0.0170.015 ± 0.007
PI-VM → AM †0.017 ± 0.0140.008 ± 0.003
PI-VM → REACT-VM0.036 ± 0.0320.018 ± 0.018
AM → PI-VM0.019 ± 0.0130.007 ± 0.002
AM → AM †0.013 ± 0.0050.006 ± 0.001
AM → REACT-VM0.019 ± 0.0140.005 ± 0.001
REACT-VM → PI-VM0.024 ± 0.0170.007 ± 0.001
REACT-VM → AM0.022 ± 0.0110.006 ± 0.001
REACT-VM → REACT-VM0.030 ± 0.0040.005 ± 0.000

Missing table reference and writing issues (W2)

Response: We fixed the missing reference (line 418, Sec. 4.2, Table 10) and corrected writing issues throughout.

I thank the authors for their detailed response, which has addressed my concerns. I have also reviewed the discussions with the other reviewers and the updated manuscript. I believe the paper has improved following the revisions, and I have therefore updated my recommendation to weak accept. Overall, I lean more toward acceptance than rejection.

We sincerely thank the reviewer for the encouraging feedback and for taking the time to carefully review our responses. We are delighted that our revisions have addressed your concerns. We greatly appreciate your constructive suggestions and positive reassessment, which have helped us further improve the quality and clarity of our work.

Comments on the paper

5 comments
AuthorsOpenReview ↗

General response and additional experiments

We thank all three reviewers for their careful review and suggestions. We have added experiments and theoretical derivations and revised the manuscript. Below we address shared concerns:

Experimental validation of energy-to-energy capability

This response addresses W2 and Q2 from R2 (WmBS), and W3 and Q4 from R3 (Rykg).

Response: We test Ring-mixture transport (2D multimodal) and Double-well transport (16D), each with three training seeds. Both methods initialize each endpoint bank with 10,000 prescribed Gaussian particles. No true endpoint samples enter initialization or training, including SPIM’s normalization and gauge. They are used only for independent evaluation.

SPIM fully replaces both banks each round with bidirectional transport outputs. Final evaluation uses model-generated endpoints. D2E accesses the same energies, gradients, physical diffusion coefficient, and proposal scale, using its native Langevin bank updates:

TaskMethodCoordinate-wise DcoordD_{\mathrm{coord}} ↓Full-dimensional W22W_2^2 ↓
Ring-mixture transportSPIM (PI-VM)0.00400 ± 0.001250.01470 ± 0.00538
Ring-mixture transportD2E0.04331 ± 0.002590.09749 ± 0.00538
Double-well transportSPIM (PI-VM)0.00601 ± 0.0014111.94401 ± 0.12258
Double-well transportD2E0.04977 ± 0.0055212.09918 ± 0.08782

Under the reported configurations, SPIM has lower coordinate-wise and full-dimensional errors on both tasks, supporting energy-only generation of multimodal and high-dimensional endpoints from Gaussian initialization. The comparison is included in Appendix E.6 (Table 11).

Additional comparisons with matched training budgets

This response addresses W1 and Q1 from R2 (WmBS), and W4 and Q3 from R3 (Rykg).

Response: We reran baselines with mismatched budgets and retained SPIM runs already satisfying the matched budgets for classical and generalized bridges, high-dimensional Gaussian transport, and image fine-tuning. Complete budgets and results appear in Appendix E.3 (Table 8) and the tables below.

  • Classical bridges. PI-VM achieves the lowest mean endpoint error on Gaussian transport, Ring-mixture transport, Double-well transport, and Banana transport (Sec. 4.2, Table 2).
MethodGaussian transportRing-mixture transportDouble-well transportBanana transport
DSBM0.60434 ± 0.007560.07908 ± 0.0009212.11999 ± 0.079700.03189 ± 0.00263
D2E0.59990 ± 0.005070.08569 ± 0.0060711.97078 ± 0.100340.19265 ± 0.01599
ASBS0.59995 ± 0.004841.71019 ± 0.0217113.80886 ± 1.252490.04093 ± 0.02848
SPIM (TR-AM)0.61100 ± 0.005730.43941 ± 0.2215412.00319 ± 0.029260.02780 ± 0.00410
SPIM (PI-VM)0.59982 ± 0.000400.02020 ± 0.0058111.90751 ± 0.053220.01667 ± 0.00100
  • Generalized Schrödinger bridge tasks. TR-AM and TR-PI-VM achieve the lowest mean interior-path errors on Sinusoidal tracking and Moving double well, respectively. DeepGSB achieves the lowest errors on both metrics for Bimodal splitting, with TR-PI-VM ranking second on both. On Spherical-meridian transport, TR-AM achieves the lowest mean endpoint error, whereas TSBM achieves the lowest mean interior-path error (Sec. 4.2, Appendix E.4, Tables 3 and 9).
MethodSinusoidal trackingBimodal splittingMoving double wellSpherical-meridian transport
GSBM0.01185 ± 0.000620.02784 ± 0.014400.02487 ± 0.002120.066524 ± 0.011680
DeepGSB0.00678 ± 0.000680.01328 ± 0.000590.01204 ± 0.003734.86566×10104.86566\times 10^{10}
TSBM0.01212 ± 0.001100.02024 ± 0.005820.03293 ± 0.003190.046241 ± 0.015290
SPIM (TR-AM)0.00559 ± 0.000240.02458 ± 0.008940.01526 ± 0.002670.051869 ± 0.016917
SPIM (TR-PI-VM)0.00771 ± 0.001670.01389 ± 0.001870.01051 ± 0.000341.22095 ± 1.08585
  • High-dimensional Gaussian transport. PI-VM achieves the lowest mean endpoint errors in both 20 and 50 dimensions (Sec. 4.2, Table 4).
  • Image reward fine-tuning. Across the nine settings in the original manuscript, SPIM achieves higher rewards in 8/9 settings and lower attribute W1W_1 in 5/9. The number of settings in which SPIM leads on attribute W1W_1 decreases from 8/9 in the original short-budget comparison to 5/9 (Sec. 4.4, Table 7).
AuthorsOpenReview ↗

Response to Reviewer WmBS

We thank the reviewer for the suggestions on fair comparisons and validation of energy-to-energy capability. We rerun the main experiments with matched training budgets, add energy-to-energy experiments without target endpoint samples.

Additional comparisons with matched training budgets (W1, Q1)

Response: See “Additional comparisons with matched training budgets” in the General response for the shared reply and results.

Experimental validation of energy-to-energy capability (W2, Q2)

Response: See “Experimental validation of energy-to-energy capability” in the General response for the shared reply and results.

Reviewer_WmBSOpenReview ↗

Remaining Questions on Optimality and Empirical Validation (Reject)

  1. How do the endpoint-only results in Tables 2 and 11 establish minimization of the SB path-space objective, rather than merely marginal matching? Where are the path-KL or control-cost comparisons at comparable endpoint accuracy?

  2. With only 2,048 evaluation samples in 20D/50D (Appendix E.3.1), how have you ruled out empirical Wasserstein bias favoring underdispersed outputs? Do the Table 4 rankings survive real–real calibration and checks of generated means and covariances?

  3. Given the mixed starting laws (μi+νi)/2(\mu_i+\nu_i)/2 at λ=10,100\lambda=10,100 (Appendix E.3.6), what evidence establishes both full reward-weighted endpoint distributions? How should this claim be reconciled with worse attribute W1W_1 than AM on all three shape tasks at λ=100\lambda=100?

  4. How do the reported boundary RMSEs validate the per-round oscillation errors relative to the current exact SOC update required by Eq. (49)? What corresponding evidence is available for AM/TR-AM?

  5. Since D2E already supports energy-to-energy transport, why is SPIM's distinctive combination of sample-free endpoints and nonzero running costs not directly tested? What measured benefit establishes the value of the SOC interface beyond implementing classical potential iteration?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.