acceptodds
Under review as a conference paper at ICLR 2027

The price of regularizing squared distance in hyperbolic optimal transport

Abstract

Changing a transport cost to obtain regularity also changes the task being solved. We quantify this price for squared distance on hyperbolic space, whose negative curvature violates the Ma–Trudinger–Wang (MTW) condition. Among normalized radial costs in dimension , critical log-cosh minimizes uniform deviation from squared distance among full-A3w costs, uniquely on each bounded distance interval. A prescribed transverse MTW floor admits an explicit sharp envelope with interval rigidity. Its weakest full margin and mixed-Hessian condition number obey , exposing the sensitivity cost of this optimal repair. Positive full margins require additional distortion: finite directional candidates and a sixth-order obstruction sharpen analytic bounds, while calibrated log-cosh supplies a feasible family-wise frontier. The critical repair's distortion and conditioning penalties vanish in the flat limit. In a fixed three-target OT problem, new-seed experiments show that repair changes the optimal transport regions at a measurable original-objective cost. Separate inversion experiments recover the predicted local sensitivity. Together, the theory and experiments distinguish the geometric benefit of regularization from its task and numerical prices.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Optimal transport, hyperbolic geometry, MTW regularity, numerical conditioning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_v3QcOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper seems to consider changing the OT ground cost to improve the regularity while assessing how much the underlying geometry is affected. Unfortunately I could not get the paper's idea beyond this point since I was unable to decipher the paper.

Critical strengths

I was not able to assess the strengths of this paper due to apparently unclear writing (at least to me).

Critical weaknesses

This paper is extremely hard to understand with many concepts and notions undefined. I believe this is not due to my lack of expertises but rather the fundamental issue existing in the submission. The following points are what tripped me up.

  • L036: What is Ma-Trudinger-Wang condition? While it may be introduced later, readers cannot continue to read if we do not even have its informal interpretation.
  • L039: What is log-cosh? What is A3w? The A3w condition seems to be a rather known set of regularity conditions in Riemannian OT, which cannot be assumed standard for ML audiences in ICLR.
  • L040: What is the minimum repair price? The "repair" and "price" are used repeatedly without a solid description, which complicates the concept of this paper.
  • L041: What is the curvature obstruction?
  • L049: What is ℓ\ell? While this is defined later in Section 3 and rather a minor issue, please do not use math symbols out of nowhere.
  • L049: Ditto; what is rr?
  • Figure 1: While the figure seems to illustrate "the critical repair" (mentioned in the caption), I really do not have idea what corresponds to it in the figure. Also, what corresponds to the transport plan in this figure? How much do (a)-(c) differ with each other?
  • L117: Hκd\mathbb H_\kappa^d is not defined.
  • L117: The sectional curvature is rather a standard notion in Riemannian geometry, and skipping its definition might be okay depending on the audience, but we should describe it in front of the ML audiences. I believe a very few proportion of the audiences are sufficiently familiar with it.
  • L118: What is cut locus?
  • L120: It is unclear what it means by "profiles with an even C4C^4 extension."
  • L127: What is the Lee-Li convention?
  • L129: Is this the definition of MTW? Or an implication of the Lee-Li convention?
  • L131: Please explain DxD_x briefly. Should cc be cℓc_\ell instead? Also, can we always ensure the invertibility of y↦−Dxc(x,y)y\mapsto -D_x c(x,y)?
  • L132: I think this is the first place where A3w (and A3s) is introduced formally. We should emphasize that this is the definition of A3w.
  • L157: What is the purpose of this design problem? Why do we need to minimize the fidelity? What is the meaning of the constraint on Λ⊥\Lambda_\perp? And what does Λ⊥\Lambda_\perp mean at the first place?
  • L163: What are u0u_0 and w0w_0?
  • L181: Again, the notion "repair" has not been defined yet.

Let me stop here. Some are minor and fixable, but others are really significant to understand the paper.

Questions for authors

See "Critical Weaknesses"

1 comment
AuthorsOpenReview ↗

Response to Reviewer v3Qc

Thank you for identifying the points at which the original presentation becomes difficult for readers unfamiliar with MTW regularity. We agree that the research question and key definitions should be introduced before the technical machinery. We have substantially rewritten the Introduction, Figure 1 explanation, and Setup (revised pp. 1–3), and added an actual-source allocation visualization in Appendix K.

Clarification of the central problem.

In hyperbolic optimal transport, the ground cost determines which assignments are preferred. Squared geodesic distance is a natural choice, but it fails the weak MTW condition relevant to transport-map regularity. Replacing this ground cost can remove that geometric obstruction while changing optimal assignments and increasing numerical sensitivity.

Our central question is: What is the smallest unavoidable distortion of squared distance required to satisfy MTW regularity, and what numerical sensitivity accompanies this modification?

This concerns modifying the ground cost, not entropic regularization of the coupling. We do not claim unconditional map regularity or downstream accuracy improvements.

Point-by-point revisions addressing the review:

  • L036, L039–L041, L181 (MTW, A3w/A3s, log-cosh, repair, price, and curvature obstruction): Revised p. 1 introduces these concepts informally before the technical results and clarifies their relevance to the transport task.

  • L049 (normalization and distance-range notation): Revised pp. 1 and 3 define ℓ′′(0)=1\ell''(0)=1, the distance variable rr, the hard distance cap DD, and the distortion measure Δ\Delta.

  • Figure 1 (interpretation of panels and transport plans): Revised p. 2 explains the three cost settings and the discrete dual-potential cells. The new Figure K.1 separately illustrates actual finite source-point allocations. Neither figure is presented as evidence of smooth Monge-map regularity.

  • L117–L120 (hyperbolic geometry, cut locus, and even extension): Revised p. 3 introduces Hκd\mathbb{H}_\kappa^d, sectional curvature, the absence of a cut locus, and the meaning of an even C4C^4 extension.

  • L127–L132 (MTW definition, Lee–Li convention, and c-exponential invertibility): Revised p. 3 explicitly distinguishes the MTW definition from its sign and scale convention, introduces A3w/A3s, and specifies that the c-exponential is inverted on its natural momentum image, not necessarily the entire cotangent space.

  • L157–L163 (design objective, margin constraints, and coefficients): Revised p. 3 explains why worst-case cost distortion is minimized, distinguishes the transverse margin Λ\Lambda from the full margin Γ\Gamma, and defines α,β,γ,δ\alpha,\beta,\gamma,\delta before the directional analysis.

The revised presentation now follows a clearer sequence: original transport task → MTW obstruction → ground-cost repair → fidelity objective → numerical sensitivity → formal definitions.

We also distinguish the solved zero-full-margin problem from the open positive-full-margin problem and explicitly retain the limitations concerning continuous-map regularity and downstream performance.

We appreciate the reviewer's detailed feedback and hope these concrete revisions make the mathematical contributions and their assumptions clearer and easier to assess.

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

The paper studies how much the squared-distance cost on hyperbolic space must change to satisfy MTW regularity. It restricts the search to normalized radial costs and measures fidelity by the maximum deficit from squared distance over a hard distance cap. After deriving a fourth-order obstruction near the diagonal, it proves a sharp envelope for a prescribed transverse MTW floor in dimensions d≥3d\geq 3. At zero floor, critical log-cosh is the unique least-distorting full-A3w profile on the chosen interval.

The authors also compute the envelope's mixed-Hessian spectrum and establish m(ℓa⋆;D)K(ℓa⋆;D)2=Λm(\ell_a^\star;D)K(\ell_a^\star;D)^2=\Lambda. The positive full-margin optimization problem is not solved exactly; the paper gives bounds, a sixth-order restriction under additional assumptions, and an optimum within the calibrated log-cosh family. The experiments test discrete assignment changes, squared-cost regret, and local inversion. They do not establish a downstream matching improvement.

Critical strengths
  1. The sharp fidelity result is a genuine addition to the known MTW examples. Lee and Li (2012) had already established the admissibility of log-cosh. Theorem 7 and Corollary 8 go further: within normalized radial costs in d≥3d\geq 3, the paper proves an exact fidelity optimum and shows that any other optimizer agrees with the critical profile on [0,D][0,D]. The interval-rigidity statement makes this more than another construction of an admissible cost.

  2. The paper identifies a specific conditioning cost of the repair. For the sharp transverse-envelope profile, Theorem 10 relates the weakest full MTW margin to mixed-Hessian conditioning through mK2=ΛmK^2=\Lambda. I also appreciate that the numerical section treats assignment changes, objective regret, and inverse sensitivity as different measurements rather than presenting them as interchangeable evidence.

Critical weaknesses
  1. The fidelity criterion is not yet well connected to preservation of the transport solution. The main optimum concerns uniform ground-cost deficit. Proposition 19 converts this into a distribution-free bound on the original objective, but a bound on the distance between plans requires an additional growth condition. Figure 1 illustrates why the distinction matters: the assignment changes despite a small squared-cost regret. The exact-OT tests also have a maximum reported R0/ΔR_0/\Delta of only 0.003940.00394, so the bound is quite conservative on these instances. None of this contradicts the theorem, and the authors acknowledge the limitation, but it makes the practical meaning of "least-distorting" less clear to me.

  2. The strongest optimality statement has narrow scope, and the experiments do little to extend it. The exact full-A3w fidelity optimum is for normalized radial costs with d≥3d\geq 3. The positive full-margin finite-cap problem remains open; calibrated log-cosh gives a feasible upper bound and is optimal only within its own family. I would not demand downstream accuracy gains from a theory paper, but the application results are inconclusive, and the numerical checks mostly test mechanisms predicted by the analysis. I am left uncertain how far the proposed design criterion carries beyond the setting of the theorem.

Questions for authors

Q1. Can the growth condition in Proposition 19 be instantiated with an explicit modulus for a nontrivial class of hyperbolic OT problems? A concrete example where the fidelity bound gives meaningful control of the optimal plan would materially change my assessment of the paper's task-preservation motivation.

Q2. Is there any analytical or numerical evidence that the calibrated log-cosh upper bound for a positive full MTW margin is close to the best normalized radial cost on a finite cap, or that a different feasible radial profile improves it? This would help clarify how much of the positive-margin design problem is captured by the current results.

3 comments
AuthorsOpenReview ↗

Response to Reviewer MaTe

We thank the reviewer for recognizing the sharp fidelity result and for identifying the missing link between ground-cost fidelity and transport-plan preservation. We agree that an objective-regret bound alone does not control the distance between optimal plans. We address the two questions separately.

Q1. An explicit growth modulus and a nontrivial smooth hyperbolic example.

The revised Section 8.1 (Proposition 23; full proof in Appendix J) makes the growth condition in Proposition 19 explicit.

Let X∼μX\sim\mu, and suppose the squared-cost optimal coupling is given by π0=Law⁡(X,T0(X))\pi_0=\operatorname{Law}(X,T_0(X)), where T0T_0 is the squared-cost optimal transport map. Assume that an optimal target-side Kantorovich potential vv admits a global C2C^2 extension satisfying Hess⁡v⪯(1−η)g\operatorname{Hess}v\preceq(1-\eta)g for some η>0\eta>0.

By the Hessian comparison theorem on hyperbolic space, the function y↦12dκ(x,y)2−v(y)y\mapsto \frac12 d_\kappa(x,y)^2-v(y) is η\eta-strongly geodesically convex. Dual slack then gives, for any coupling π\pi with the same marginals,

\geq\frac{\eta}{2}W_{2,\mathrm{prod}}(\pi,\pi_0)^2.$$ Here $W_{2,\mathrm{prod}}$ is the quadratic Wasserstein distance between couplings on the source–target product space, not between the two marginals. Thus $\omega(t)=\eta t^2/2$, and Proposition 19 yields $$W_{2,\mathrm{prod}}(\pi_c,\pi_0)\leq\sqrt{2\Delta(c;D)/\eta},$$ provided $D$ bounds every source–target pair in $\operatorname{supp}(\mu)\times\operatorname{supp}(\nu)$. The dual-Hessian condition is a sufficient assumption, not a claim about arbitrary OT data. **The example is genuinely nontrivial.** Proposition 24 constructs smooth, compactly supported source and target densities in $\mathbb{H}_\kappa^3$ from an anisotropic perturbation of a radial target-side dual potential. The source momentum one-form of the squared-cost optimal map, after the log-cosh radial reweighting, has a nonzero exterior derivative on an open set. Hence the same map cannot be optimal for the repaired cost. Strict radial convexity yields the twist condition, and absolute continuity of the source ensures uniqueness of the repaired optimal Monge coupling. The two couplings therefore differ while the conditional stability bound applies. This is an analytic existence-and-stability example: we do not claim that its actual continuum plan distance was measured or that the bound is sharp. The local numerical audit only checks the nonclosedness expansion. **Q2. Tightness of the positive full-margin cost-design bound.** From Corollary 14, let $L=\Delta(\ell^\star_{\Gamma/3};D)$ and $U=\Delta(\ell_{q_\Gamma};D)$. Since $L\leq V_{\mathrm{full}}(\Gamma,D)\leq U$, we obtain $$0\leq\frac{U-V_{\mathrm{full}}}{V_{\mathrm{full}}}\leq\frac{U}{L}-1.$$ The new 16-setting analytic table in Section 8.2, with $\kappa=1$, $D\in\{0.5,1,2,3\}$, and $\Gamma\in\{0.1,0.25,0.5,1\}$, gives a maximum relative suboptimality upper bound of **6.57088%**, attained on this grid at $(D,\Gamma)=(2,1)$. Therefore, the calibrated candidate's worst-case cost distortion is provably within this relative excess of the unknown optimum over all admissible normalized radial costs at these tested settings. This does not determine the exact optimizer or exclude improvements by other radial profiles. Nor is it a uniform guarantee over other parameters, a plan-distance bound, or a downstream performance claim. **In summary**, we have made explicit when cost fidelity also controls a transport solution and how close the candidate cost design is to the unrestricted normalized-radial optimum on the specified parameter grid, while keeping these two claims logically distinct.

Thank you for the thoughtful response and substantial revisions. I appreciate the effort to address both of my questions.

The new plan-stability result and the 6.57% near-optimality certificate meaningfully strengthen the theoretical contribution and alleviate my earlier concerns.

I still have some reservations about the practical significance of the fidelity criterion, given the additional assumptions required for plan stability.

I also think the presentation could be improved. The paper remains quite dense, and the main contributions are sometimes buried under technical details and qualifications. Moving Table 6 and Figure 3 into the main text would help readers better understand the scope and significance of the results.

Overall, while I appreciate the improvements and am more positive about the paper, I will keep my rating at lean reject, mainly due to the remaining concerns about practical significance and presentation.

Thank you for your careful follow-up and for recognizing the additional theoretical results. We understand your remaining concerns about practical significance and presentation.

In the latest revision, we have brought Figure 3 and Table 6 into the main text, while retaining their appendix copies for continuity. We have also clarified how the fidelity, conditioning, and inverse-gain bounds can be used to screen cost repairs against application-specific budgets.

We agree that this is a cost-design guarantee rather than evidence of improved downstream accuracy, and that the plan-stability result remains conditional on the stated global dual-Hessian assumption.

The mathematical results, experimental data, and negative application findings remain unchanged. We appreciate your constructive suggestions.

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

The paper studies the cost of modifying squared-distance optimal transport on hyperbolic space to satisfy the MTW regularity condition. Among normalized radial costs in dimensions d >=3, it proves that critical log-cosh uniquely minimizes the uniform deviation from squared distance on any bounded distance interval. It also derives a sharp envelope under a prescribed transverse MTW lower bound and an exact relationship between this envelope’s full MTW margin and mixed-Hessian conditioning. For positive full MTW margins, the paper provides analytic bounds, a sixth-order local obstruction, and an optimal choice within the calibrated log-cosh family; the unrestricted optimum remains open. Validation includes numerical checks of the theoretical formulas, local inversion experiments confirming predicted sensitivity, and a controlled three-target OT experiment demonstrating changed transport assignments and measurable regret under the original squared-distance objective. Additional synthetic and real-data matching experiments do not establish a robust downstream benefit. Overall, the contribution is primarily theoretical: it quantifies the fidelity and local numerical sensitivity costs of MTW repair, while the experiments illustrate these costs rather than directly establish improved regularity of continuous transport maps.

Critical strengths
  1. The paper proves that critical log-cosh uniquely minimizes uniform deviation from squared distance among normalized radial A3w costs in dimensions d >=3, with rigidity on each bounded distance interval. This strengthens the previously known admissibility of log-cosh into a principled optimality result. It is the strongest argument for acceptance because it answers a precise cost-design question over a function class rather than merely comparing selected examples.
  2. For the optimal transverse envelope, the paper derives the exact identity alongside explicit mixed-Hessian singular values and inverse amplification. These results explain how a geometrically admissible repair can incur substantial numerical sensitivity, with inversion experiments supporting the predicted local behavior. This adds a consequential computational perspective to the theoretical characterization.
Critical weaknesses
  1. The sharp envelope optimizes a transverse constraint, whereas a prescribed full MTW margin requires control over all orthogonal directions. For the latter problem, the paper provides bounds and a family-restricted log-cosh optimum, but neither identifies the unrestricted optimum nor establishes a positive gap above the unattainable lower bound. The sixth-order obstruction additionally assumes fourth-order saturation. This limits the completeness of the cost-design contribution when a strictly positive uniform margin is required.
  2. The three-target experiment demonstrates changed assignments and objective regret, but its atomic targets cannot validate continuous-map regularity. The inversion experiments validate local sensitivity rather than end-to-end solver performance, while the application experiments show no robust downstream improvement and include near-random recovery or ineffective registration. Moreover, the original M0–M3 pipelines are not fully reproducible from the supplied exports, as acknowledged in Appendix H.8. Consequently, the paper quantifies several costs of repair convincingly, but provides limited evidence of when those costs yield a useful practical benefit.
Questions for authors

How tight is the bracket in Corollary 14 for the parameter ranges used in the experiments? Does it provide a meaningful quantitative guarantee for calibrated log-cosh relative to the unrestricted optimum? This would clarify how much practical guidance the theory provides beyond the solved zero-margin case. Lee–Li’s Theorem already implies B'(z)≤0 for radial A3w costs in dimensions d >=3. Under the paper’s normalization, integrating this inequality appears to yield the critical log-cosh envelope directly; the positive transverse-floor extension follows similarly from a quadratic bound on B. Could the authors clarify which results require substantive new arguments beyond these consequences of the existing criterion, and where they locate the main technical novelty?

Do any existing results demonstrate improved regularity of a continuous transport map under the required density and domain assumptions? The three-target experiment establishes assignment changes, while the inversion experiments establish local sensitivity. Please clarify whether the empirical contribution is limited to illustrating the costs of repair, or also provides evidence of its benefits.

Can you provide the complete M0–M3 implementations and input-generation procedures, together with run identifiers that reconcile the 33 aggregate solver failures with the 31 failures reported in the diagnostic panels? This would materially strengthen confidence in results currently supported primarily by exported records.
1 comment
AuthorsOpenReview ↗

Response to Reviewer jfsB

We appreciate the technically precise review and its distinction between the solved zero-margin problem and the open positive-full-margin problem. The revision clarifies the novelty relative to Lee--Li, quantifies the design bracket, and strengthens independent reproducibility evidence without presenting the negative application results as benefits.

  1. What is technically new beyond Lee--Li (2012)? We agree that the radial A3w criterion already implies B′(z)≤0B'(z)\le0 in d≥3d\ge3, and that integrating it gives the basic transverse envelope comparison. We do not claim that integration itself is technically novel. The additional results and proof obligations are: (i) an extremal cost-design formulation with its equality case and interval rigidity, plus attainment of the full A3w zero-margin optimum (Theorem 7/Corollary 8); (ii) optimization over all orthogonal MTW directions for the envelope extremizer, identifying the radial--transverse weakest direction and deriving the radial and transverse branches of the mixed-Hessian spectrum, thereby proving m(ℓa⋆;D)K(ℓa⋆;D)2=3am(\ell_a^\star;D)K(\ell_a^\star;D)^2=3a for this extremizer only (Theorem 10); (iii) a positive-coefficient representation that verifies the full-directional uniform MTW margin of calibrated log-cosh (Theorem 12); and (iv) a radial--transverse sixth-order restriction for profiles saturating the fourth-order price (Theorem 15). The exact unrestricted positive-full-margin optimizer remains open. These proof-level distinctions are now emphasized in the Introduction and Discussion.
  2. How tight is Corollary 14? Write L≤Vfull≤UL\le V_{\mathrm{full}}\le U for its analytic bracket. Then (U−Vfull)/Vfull≤U/L−1(U-V_{\mathrm{full}})/V_{\mathrm{full}}\le U/L-1. Evaluating the exact formulas at κ=1\kappa=1, D∈0.5,1,2,3D\in{0.5,1,2,3}, and full-margin budgets Γ∈0.1,0.25,0.5,1\Gamma\in{0.1,0.25,0.5,1} yields a maximum certified relative cost-distortion excess of 6.57088% over these 16 configurations (at D=2,Γ=1D=2,\Gamma=1, L≈0.873237L\approx0.873237, U≈0.930616U\approx0.930616). This certificate compares the calibrated profile with the unknown global optimum over the stated normalized radial class; it is not an assertion of exact optimality or of a uniform bound beyond the grid. We also emphasize that near-optimality need not mean a small absolute price: at this configuration, U/(D2/2)≈46.5U/(D^2/2)\approx46.5%, K≈13.95K\approx13.95, and maximal inverse gain G≈72.06G\approx72.06.
  3. Does the paper demonstrate smoother continuous OT maps? No, and the revised manuscript explicitly avoids that inference. The three-target LP tests discrete assignment changes and original-objective regret; the local inversion experiment tests sensitivity. Neither establishes improved regularity of continuous Monge maps or robust downstream accuracy. Existing theory (e.g., Figalli--Kim--McCann, 2013) gives conditional regularity under A3w together with appropriate twist/nondegeneracy, density bounds and strong cc-convexity assumptions. Our empirical setups do not verify all those hypotheses. The null and inconclusive WordNet/OAEI outcomes are retained.
  4. M0--M3 source and 33 versus 31 failures. We have newly implemented 270 independent SciPy/HiGHS LP solves covering every declared three-target configuration, with a separately implemented sorting solver. Maximum objective difference is 1.22×10−151.22\times10^{-15}, and maximal marginal residual is 2.49×10−142.49\times10^{-14}. Additional independent sampled checks cover 24 M0 directional cases and 48 M1 inverse-gain cases (Appendix K and the anonymous reproducibility archive). These checks do not restore the original M0--M3 pipelines. The historical input clouds, checkpoints, complete logs, and run identifiers are unavailable; therefore the 33 aggregate failures cannot be reconciled at run level with 31 diagnostic-panel flags. We do not infer an unsupported failure rate. Revised Appendix H.7--H.8 and the supplement README explicitly disclose that the 14 historical plotting exports are absent from the new independent-audit package, while retaining the original negative application results. We hope the distinction between new mathematical structure, reproduced numerical mechanisms, and unresolved experimental limitations supports reassessing the submission as a carefully scoped theory paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.