acceptodds
Under review as a conference paper at ICLR 2027

HyRef-SB: Hyperbolic Schrödinger Bridges Guided by Geometric References

39 views

Abstract

Endpoint distributions specify where stochastic transport starts and ends, while reference dynamics provide a geometric inductive bias for the resulting stochastic paths between them. We introduce Hyperbolic Reference-guided Schrödinger Bridges (HyRef-SB), combining reference diffusion, stochastic control, and endpoint distribution constraints on a hyperbolic manifold. An invertible calibration preserves the input representation, while a prescribed base metric defines diffusion and control costs. An endpoint-preserving path teacher calibrates the reference drift; differentiable rollouts then jointly minimize intrinsic residual control energy and a soft endpoint penalty under the fixed reference. A shared backbone and conditional adapters model multiple terminal distributions. Our analysis characterizes the radial and transverse costs induced by calibration and quantifies, for regular teacher marginals, how reference regression error and density regularity bound reference marginal error and a feasible residual-control budget. Continuous and finite-step entropy identities identify the regularized path laws, and a population penalty limit connects soft endpoint fitting to hard marginal constraints. Experiments yield low-collision navigation and recover tree-structure statistics, while improving cellular distribution matching over the evaluated baselines on most terminal and held-out intermediate metrics. Hyperbolic geometry reduces terminal errors relative to scale-matched Euclidean controls on three cellular tasks, and the Trametinib ablation shows further gains from the learned reference over a Brownian reference. The complete code will be made publicly available.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Schrödinger Bridges, Hyperbolic Geometry, Stochastic Control, Reference Dynamics, Conditional Distribution Transport, Manifold Generative Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Vha2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HyRef-SB, a framework for conditional stochastic transport on a hyperbolic–Euclidean product manifold. An invertible calibration maps preprocessed observations into geometric coordinates. An endpoint-preserving teacher supplies trajectories whose velocities train a reference drift. This reference is then frozen while a residual controller is optimized through stochastic rollouts, using intrinsic control energy, soft terminal distribution discrepancies, and optional support costs.

Critical strengths
  1. The same metric defines diffusion, control energy, and exponential-map updates. The factors associated with the Poincaré convention g0=4Ig_0=4I are handled consistently: the radial coefficient, transverse amplification, and origin-matched Euclidean comparison agree with this convention. Preserving all preprocessed coordinates also avoids conflating geometric calibration with dimensionality reduction.

  2. The comparison drift $ b^\dagger=b_\ast+\frac{\sigma^2}{2}\nabla_g\log p_t $ correctly compensates for diffusion while preserving the teacher marginals. The resulting bound separates reference regression error from the density correction required by stochastic dynamics. The paper appropriately retains the Fisher-information term even when regression error vanishes.

Critical weaknesses
  1. Theorem 5.4 concerns approximate minimizers over population path measures as the endpoint penalty tends to infinity. The implementation uses a finite penalty, restricted neural controls, empirical discrepancies, and a fixed discretization. Corollary C.14 correctly identifies approximation, optimization, and estimation errors, but none of these errors is controlled experimentally.

  2. All main benchmarks and geometry/module ablations use a single training seed. The five-seed diagnostic is useful but does not establish robustness of the navigation, tree, or cellular rankings. This matters especially for small differences, such as Mouse Hematopoiesis intermediate Wasserstein errors.

  3. Appendix F.5.1 specifies that Clonidine and Trametinib use all 50 PCs for MMD but only the first two PCs for W1W_1 and W2W_2. This distinction is not prominent in Tables 3 and 7 or their interpretation. The paper should report full-dimensional Wasserstein or sliced-Wasserstein diagnostics alongside the two-PC results and identify evaluation coordinates directly in the table captions.

  4. Equation (25) combines generated-endpoint readouts with fixed source-conditioned weights q(k∣x)q(k\mid x). However, the construction of these weights for unseen source cells is insufficiently specified.

  5. Collision detection examines only interior grid points. This is a valid discrete evaluation, but it can miss obstacle crossings between observations. The reported near-perfect safety should therefore be supported by grid-refinement or more conservative crossing checks, with the total number of evaluated paths stated.

  6. The existing geometry and Brownian-reference ablations are helpful, but they do not establish how much performance comes from residual stochastic optimization beyond the teacher and fitted reference. Cellular experiments also use a small total noise scale of 0.050.05. Comparisons against the frozen reference, a deterministic counterpart with matched architecture and action penalty, and several noise/control-weight settings would help establish when the full stochastic bridge formulation is necessary.

Notation is generally consistent, but η\eta is reused for teacher parameters, the endpoint-penalty multiplier, and the calibration coefficient in Appendix F.3. Distinct symbols would improve readability. Figure 2 is too small for reliable inspection of its axes and trajectories. A compact notation table and a single implementation specification would be more useful than some of the repeated explanations of standard identities.

Questions for authors
  1. What are the exact numerical safeguard maps and thresholds? Are they identical for controlled and reference transitions as required by Proposition C.19, and how often are they activated?

  2. Can you evaluate endpoint discrepancy and control energy across increasing penalty strengths on a problem with a known bridge solution? Which parts of the population penalty theorem are empirically supported by the trained model?

  3. Do the main rankings persist across independent training seeds and comparable total computational budgets? How do the frozen-reference and deterministic counterparts perform?

  4. Can you provide full-dimensional drug-task Wasserstein or sliced-Wasserstein results and explicitly label the coordinate space used by every reported metric?

  5. How is q(k∣x)q(k\mid x) constructed and extended to held-out LARRY source cells? What accuracy is obtained using these weights and terminal-class readouts without learned transport?

Reviewer_zkk3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an approach for solving Schr\”odinger Bridge (SB) problem on hyperbolic geometry with a pretrained reference process.

Critical strengths

Unfortunately unable to find any

Critical weaknesses

[W.1] Poor clarity and presentation. The paper is poorly written (and at times appears possibly AI generated). Given the complexity of the topic, this makes the paper extremely difficult to read, and I do not think the presentation meets the standards expected at an A* conference. I would like to emphasize that this is not simply a matter of occasional awkward wording, missing explanations, or minor presentation issues. Rather, the writing is consistently unclear, making the paper unnecessarily difficult to follow. Here are some examples:

  1. Literally the first sentence of the abstract starts by describing some dynamics without ever introducing what these dynamics are.

  2. The whole abstract and introduction read like a combination of sophisticated terminology, much of which is not common knowledge, without sufficiently explaining the underlying ideas. The abstract almost requires reading the entire paper to understand what is being said.

  3. In the Introduction (lines 41–49), the paragraph reads more like a collection of separate sentences than a connected and meaningful discussion. Specifically, what is the transition between the sentence about the reference process (2nd sentence) and the one about geometry (3rd sentence)? Why are some properties of hyperbolic geometry discussed before hyperbolic geometry is even introduced in the next sentence? How does learning geometry relate to the previous discussion at all? Why exactly is learning the geometry presented as the solution, rather than simply changing the geometry from Euclidean to hyperbolic? I find it very difficult to follow the logic of this paragraph.

  4. The Related Work section is another example. The whole section mostly lists the core ideas of related methods without providing specific details that would clearly distinguish them from the proposed approach. For example, the “Bridges and references” paragraph explains that the proposed method learns a reference process without clearly stating that other methods typically use a fixed one. Similar issues appear in the other paragraphs of this section.

  5. In Sections 3 and 4, much of the information is introduced rather abruptly, without sufficient context or motivation. For example, how is the information about parallel transport (line 129) related to the preceding discussion in the preliminaries? Why is a special case in which hyperbolic geometry becomes Euclidean mentioned at all? This information is almost never used later in the main text. These details make the presentation harder to follow rather than helping the reader understand the method. Why do you introduce two categorical distributions χk\chi_k and πk\pi_k? What is the difference between them?

  6. Some terms are not clearly defined or introduced with sufficient information. For example, “invertible calibration” appears three times (according to a text search), and the authors discuss some of its properties (line 137) without first clearly explaining what it actually means. What is the “support cost” VUV_\mathcal{U}? How is it defined? Why does its notation appear without any introduction in Equation (3)? What is F(z)F(z)? In line 171 (first equation) it is discussed where the terms sjs_j appears. Similarly, line 164 introduces several additional terms without defining them. I believe such concepts should be clearly explained before they are used.

  7. Some mathematical constructions lack sufficient motivation. For example, why do the authors construct the manifold as a product of two spaces? Why can we not simply work directly in hyperbolic space? The paper should explain the motivation behind such design choices rather than simply introducing them.

As a consequence of such poor writing, it is extremely difficult (and some parts completely impossible by point 6) to fully understand and properly evaluate the paper. Therefore, all of my comments focus on this issue.

[W.2] It's possible that the template is broken. Specifically, the formatting of the reference list looks wrong.

Questions for authors

See Weaknesses

Reviewer_mCqhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

I am unsure what the paper actually aims to do. The paper is concerned with Schrödinger bridges in hyperbolic manifolds (or at least with reference processes on such manifolds). It was not entirely clear what the paper contributed or what the authors hoped to achieve with the work.

Critical strengths

I found the work too difficult to read to isolate any particular strengths.

Critical weaknesses

The paper's main weakness is that I found it nearly incomprehensible. I work with generative models and differential geometry, yet I found myself unable to determine what problem the paper aims to solve.

The experiments appear a bit "all over the place". I was unable to determine the nature of the tackled problem by inspecting the experiments (since I didn't understand the main body of the paper, I tried to determine the aim of the paper through the experiments, but I did not succeed).

Remark on Soundness score: I was simply unable to assess the presented work and therefore unable to determine if it is sound.

I understand that this review is not particularly helpful to the authors, and for this I am sorry. Yet, I do not consider the manuscript to be in a form suitable for review.

Questions for authors

What problem are you trying to solve in this paper?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.