acceptodds
Under review as a conference paper at ICLR 2027

Taming the Boundary: Truncated Poincaré Quasi-Metric for Stable Hyperbolic Retrieval

Abstract

Poincar\'e-ball representation learning provides a powerful inductive bias for hierarchical and multi-scale data, but its practical use is often limited by severe numerical instability. As embeddings approach the boundary, distances become ill-conditioned, increasing sensitivity to floating-point errors and destabilizing optimization. In this paper, we propose a truncated Poincar\'e geometry that preserves the standard Poincar\'e geometry within a radius while replacing the near-boundary region with scaled Euclidean geometry. We theoretically show that, unlike the standard Poincar\'e distance, whose gradient can grow as when a point is away from the boundary, the proposed truncated quasi-metric distance has constant, boundary-independent gradient magnitude. We evaluate the method on twelve retrieval datasets spanning image, text, video, and cross-modal retrieval. Experimental results show that truncated Poincar\'e geometry substantially improves numerical stability while achieving competitive or improved retrieval performance across diverse settings. These findings suggest that quasi-metric truncation offers a practical stability patch for Poincar\'e retrieval models without discarding the useful inductive bias of Poincar\'e representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperbolic, Multimodal, Retrieval

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_dJLrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies the numerical instability of Poincaré-ball representations near the boundary and proposes a truncated Poincaré quasi-metric. It retains standard Poincaré geometry within radius r0r_0, switches to scaled Euclidean geometry outside, and uses a Ray Intersection Approximation (RIA) to approximate cross-boundary distances. The paper further analyzes near-boundary gradient stability and evaluates the method on 12 datasets spanning image, text, video, and cross-modal retrieval. The overall goal is to preserve the advantages of Poincaré representations while mitigating boundary-induced numerical instability.

Critical strengths
  1. The method is simple and directly addresses a practical problem. Near-boundary numerical instability is an important issue for Poincaré representations. By truncating the outer region and using a closed-form RIA, the method provides a direct stabilization mechanism with low additional computational complexity and can be readily integrated into existing retrieval pipelines.
  2. The experimental coverage is broad. The paper evaluates image, text, video, and cross-modal retrieval across 12 datasets, and further includes analyses of hyperparameter sensitivity, model scale, and embedding-norm dynamics.
Critical weaknesses
  1. The proof of Theorem 1 is inconsistent with the RIA actually defined in the paper. In the main text and Appendix B, the cross-boundary intersection point is defined as p=x+t∗(y−x)p=x+t^*(y-x), i.e., the intersection of the line segment from xx to yy with the truncation sphere. However, Appendix C instead uses p=r0y∥y∥p=r_0\frac{y}{\|y\|} when proving the general-xx case, which is a radial projection from the origin. Thus, the current proof analyzes a different construction from the general RIA defined in the method. In addition, Eq. (25) only establishes a constant upper bound on the gradient, which directly supports O(1)O(1) rather than the claimed Θ(1)\Theta(1). This issue directly affects the paper’s main theoretical conclusion.
  2. The experimental setup contains inconsistencies that weaken the reliability of the performance and stability conclusions. For image retrieval, the main text states that LoRA is applied to all transformer blocks, whereas Appendix E.1 states that only the last two transformer blocks are fine-tuned, with different epoch and batch-size settings. For text retrieval, the baseline compare-all runs and the tuned truncated-Poincaré runs use different batch sizes; for InfoNCE, this changes the number of in-batch negatives. In addition, the definitions and measurement protocols of the core stability metrics in Figure 3 are not sufficiently clear. These issues make it difficult to determine whether the reported gains arise from the proposed geometry or from differences in training and tuning.
Questions for authors
  1. Regarding Theorem 1, could the authors provide a complete gradient-boundedness proof for the RIA actually used in the paper, p=x+t∗(y−x)p=x+t^*(y-x)? Please also clarify why the conclusion should be Θ(1)\Theta(1) rather than only O(1)O(1).
  2. Regarding experimental fairness and reproducibility, please specify the exact training configurations corresponding to the main-table results and re-evaluate the major baselines under the same batch size, optimizer, training schedule, and tuning budget. Please also provide precise mathematical definitions and measurement protocols for MeanRelErr, MaxRelErr, Max distortion, and Diameter delta in Figure 3.
Reviewer_gybnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a truncated Poincaré quasi-metric for contrastive retrieval, retaining hyperbolic distances within an interior radius, using scaled Euclidean distances outside it, and approximating cross-boundary distances through a closed-form ray intersection. Its main contributions are this efficient hybrid similarity and a theoretical argument that near-boundary gradients remain bounded, unlike standard Poincaré distance gradients. Evaluation spans 12 image, text, video, and cross-modal retrieval datasets, with geometric hyperparameter and backbone ablations. Results show competitive performance and generally modest, dataset-dependent gains. The claimed continuity and gradient guarantees warrant closer scrutiny, while retrieval accuracy alone does not establish numerical stability.

Critical strengths

The paper addresses a practical limitation of Poincaré representation learning through a simple, closed-form similarity that can replace standard hyperbolic distance without changing the backbone or contrastive objective. Its combination of interior hyperbolic geometry and exterior Euclidean scaling offers an interpretable stabilization mechanism. Evaluation across 12 datasets and multiple retrieval modalities, together with geometric hyperparameter and backbone ablations, demonstrates broad applicability and competitive, though generally modest, performance gains.

Critical weaknesses

The central theoretical claims require stronger justification: the piecewise definition may be discontinuous when transitioning between cross-boundary and outside–outside cases, and the stated distance bounds do not themselves establish gradient boundedness or well-conditioned optimization. Moreover, bounded asymptotic gradient growth should not be described as constant gradient magnitude. Empirically, small and setting-dependent gains need multi-seed uncertainty estimates, matched tuning budgets, and direct numerical-stability measurements. Comparisons with simpler Euclidean mixtures and analysis of branch usage are also needed to establish whether benefits arise from the proposed geometry rather than increased Euclidean behavior.

Questions for authors
  1. How does the proposed distance remain continuous when one point crosses the truncation boundary while the other remains outside, particularly when their connecting segment passes through the inner ball?
  2. Are the performance gains consistent across random seeds and matched hyperparameter-search budgets? How do they compare with a simple hyperbolic–Euclidean distance mixture?
  3. What proportion of embedding pairs uses each distance branch, and do the gains persist when controlling for increased Euclidean behavior?
  4. Why do several Poincaré baselines produce identical reported results, and what is the proposed method’s computational overhead relative to these baselines?
Reviewer_5nRKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses a well-known problem with Poincaré embeddings: their numerical instability near the disk boundary. To address this, the authors propose a hybrid approach that combines the hyperbolic and Euclidean metrics, along with a method for computing distances between points in different regions. The authors evaluate their proposal against a variety of baselines on multiple datasets spanning text, images, video, and cross-modal retrieval.

Critical strengths

S1. Exact distance computation across hyperbolic and Euclidean metric spaces is expensive and has no closed form. The authors therefore propose a closed-form quasi-metric, and Theorem 1 shows theoretically that it substantially improves numerical stability compared to the hyperbolic metric.

S2. The method is not sensitive to hyperparameters, so its performance is stable. In Figure 4, the heatmap coloring exaggerates the differences, but the actual values barely vary: the largest difference within the matrix is only 0.0084. This low parameter sensitivity indicates that performance is stable.

Critical weaknesses

W1. The performance gain of the proposal is quite small, so its contribution is not significant.

  • Theorem 1 proves that, compared with the hyperbolic metric, the RIA-based quasi-metric improves numerical stability by several orders of magnitude. However, the retrieval accuracy in the experiments shows only very small differences between the hyperbolic and Euclidean metrics. Specifically, across the text, image, video, and cross-modal experiments, the performance difference from the baselines is at most 1%, and only about 0.1% on average.
  • Moreover, the experiments for each modality lack discussion. The paper mostly restates the results without analyzing them. For example, Section 3.3 states that “its gains are dataset- and tuning-dependent rather than uniform,” but it offers no particular findings. The authors should analyze which methods are suitable for data with which characteristics, and which design principles of the methods are effective.
Questions for authors

No questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.