- 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
-
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.
-
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.
-
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
-
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.
-
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.
-
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.
-
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
-
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?
-
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?
-
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?
-
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?
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-finite Brownian reference initialized with Lebesgue measure, R0(dx)=dx. With T=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).