acceptodds
Under review as a conference paper at ICLR 2027

RAP: Radial-Angular Random Projection for Hyperbolic Representations

36 views

Abstract

Random projection is a basic compression primitive for Euclidean representation learning, with the Johnson–Lindenstrauss lemma giving a target dimension independent of the ambient dimension. As hyperbolic representations gain increasing attention in modern Transformers, vision-language models, and large language models, the same need for oblivious dimensionality reduction arises, but no analogous primitive yet exists for hyperbolic representations. We study this gap. Two natural attempts fail in dual ways: a tangent-space scheme gives a clean local guarantee but is chart-bound and pays a curvature penalty, while a fully hyperbolic projection in the hyperboloid model is geometrically native but mixes radial and angular error before feeding them into the conditioning of . We propose radial-angular random projection (RAP), which preserves the radial coordinate exactly and applies a Euclidean random projection only to the angular direction, lifting the result back to the target hyperboloid. RAP is hyperbolic-to-hyperbolic by construction, costs one Gaussian matrix-vector product, and reduces hyperbolic compression to a normalized angular JL question on . We prove a finite-set RAP guarantee with explicit conditioning, complemented by a local tangent-space JL theorem and an arcosh-conditioning lemma for the fully hyperbolic comparator. Across an origin-centered broad-radius point cloud, a balanced tree, a clustered Gaussian, and the WordNet noun hierarchy, RAP attains the lowest mean relative distortion in every cell, reducing distortion over the strongest hyperbolic baseline by roughly an order of magnitude on broad-radius data. RAP also scales efficiently to million-point datasets and delivers improved retrieval performance on vision–language representations. Code is available in an anonymous repository: https://anonymous.4open.science/r/RAP-0C10/README.md.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Hyperbolic Geometry, Random Projection

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_d9jJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Radial Angular random Projection (RAP), a hyperbolic-to-hyperbolic dimensionality reduction method. RAP works by converting all coordinates to polar coordinates, preserving the radial component, and applying a Johnson-Lindenstrauss type approach to the angular component. This method was mainly evaluated by measuring the mean relative distortion of low-dimensional embeddings for existing hyperbolic data.

Critical strengths

The proposed Radial Angular Random Projection (RAP) is a novel approach and shows good empirical results against the baselines.

Critical weaknesses

(W1) Relationship to prior literature: [1] uses a similar Johnson-Lindenstrauss idea to do hyperbolic-to-hyperbolic dimensionality reduction, and this paper does not cite this work nor compare to this work’s methods or in experiments. The bounds of [1] appear to be at least as strong as this paper’s. The authors should also be aware of [2], which extends [1], but which may be contemporaneous work.

[1] Benjamini, Itai, and Yury Makarychev. "Dimension reduction for hyperbolic space." Proceedings of the American Mathematical Society (2009): 695-698.

[2] Deng, Chengyuan, et al. "Locality sensitive hashing in hyperbolic space." arXiv preprint arXiv:2603.19724 (2026).

Questions for authors
  1. How does RAP compare to the hyperbolic dimensionality reduction approach of Benjamini and Makarychev [1]?
2 comments
AuthorsOpenReview ↗

Official Response to Reviewer d9jJ

We sincerely thank the reviewer for recognizing RAP's novelty and strong empirical results. We clarify the differences from the cited works and report a direct experimental comparison below. W/Q denotes the concern, and R our response.


W1/Q1. How does RAP relate to the dimensionality reduction methods of Benjamini and Makarychev [1] and Deng et al. [2]?

R: Deng et al. [2] refine the analysis of Benjamini and Makarychev's [1] dimensionality reduction and apply it to hyperbolic locality-sensitive hashing for approximate nearest-neighbor search. Both use the same half-space projection construction.

(1) RAP preserves the radial hierarchy signal exactly, whereas [1,2] do not provide this guarantee. A key advantage of hyperbolic over Euclidean representations is their efficient encoding of hierarchies. When radius encodes depth, exact radial preservation maintains depth rankings, depth-based filtering, and the radial cues used in hierarchical classification and retrieval. Specifically, [1,2] fix height through (z,x)↦(z,Px)(z,\mathbf{x})\mapsto(z,\mathbf{P}\mathbf{x}), but the distance rr to (1,0)(1,\mathbf{0}) at curvature −1-1 satisfies cosh⁡r=(z2+∥x∥2+1)/(2z)\cosh r=(z^2+\|\mathbf{x}\|^2+1)/(2z). Projection-induced norm changes can therefore alter radii, analogous to our vanilla fully hyperbolic baseline. RAP instead fixes r=dH(o,x)r=d_{\mathbb{H}}(o,x) and normalizes only the projected direction: (r,u)↦(r,Pu/∥Pu∥)(r,\mathbf{u})\mapsto(r,\mathbf{P}\mathbf{u}/\|\mathbf{P}\mathbf{u}\|). This preserves radii and distances along every ray from the chosen origin exactly; our radius/separation-dependent bounds characterize the remaining distance distortion.

(2) RAP admits native Lorentz and Poincaré ball implementations, whereas the published formulas in [1,2] require upper-half-space coordinates. Applying those formulas to Lorentz or ball embeddings entails coordinate conversion or an equivalent reformulation. Explicit conversions add arithmetic operations and may introduce floating-point roundoff. RAP's intrinsic polar formulation avoids conversion between hyperbolic coordinate models.

(3) RAP achieves lower mean relative distance distortion than half-space Gaussian JL [1,2] in all 29 evaluated compression settings, with relative reductions of 20.9% to 49.4%. We reran all three methods on identical inputs at equal target intrinsic dimension k, with 20 paired projection seeds, independent data/projection randomness, and all unordered point pairs. The table reports mean distortion ± standard deviation at k = 32. Relative error reduction is measured against half-space JL.

Datasetd → kHalf-space [1,2]Fully hyperbolicRAPRelative error reduction
Broad-radius128 → 320.0770 ± 0.00170.0765 ± 0.00200.0397 ± 0.000848.5%
Balanced tree50 → 320.0911 ± 0.00520.0909 ± 0.00530.0712 ± 0.003821.8%
Clustered50 → 320.0860 ± 0.01030.0853 ± 0.01070.0649 ± 0.004324.5%
WordNet depth 350 → 320.0963 ± 0.00640.0950 ± 0.00550.0682 ± 0.002929.1%
WordNet depth 450 → 320.0963 ± 0.00810.0949 ± 0.00740.0694 ± 0.003828.0%
WordNet depth 550 → 320.0944 ± 0.00840.0930 ± 0.00790.0706 ± 0.005125.2%
WordNet depth 650 → 320.0931 ± 0.00930.0918 ± 0.00900.0725 ± 0.007122.1%

We have cited both papers and include this comparison in the revised discussion.

Esteemed Reviewer,

We sincerely thank you for the time and effort you have devoted to evaluating our work.

As the rebuttal period draws to a close, we would be grateful if you could let us know whether our responses have addressed your concerns. If there are any questions remaining, we would be happy to discuss them and provide further clarification.

Best regards,

The Authors

Reviewer_E2yBOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes RAP, a radia-angular random projection method for a training free dimensionality reduction method for hyperbolic representations. The authors identify that naive attempts at this problem lead to breakdown of the hierarchical geometry and pairwise distance metrics, such as in the tangent space projection and fully hyperbolic projection, while prior work focuses on data dependent dimensionality reduction techniques. To overcome this problem, the authors propose to use the polar corrdinates and keep the radial unchanged while limiting the random projection to only the angular component. The paper provides theoretical guarantees and empirical evidence on both synthetic benchmarks as well as real benchmarks showing that RAP consistently and significantly outperforms other baselines.

Critical strengths
  1. The radial-angular decoupling is a clever and elegant method.
  2. The theoretical guarantees are sound.
  3. Experimental evaluation is comprehensive, and show signitifcant improvements over the non data driven baselines in both synthetic and real world benchmarks.
Critical weaknesses
  1. While the decoupling is elegant, it might put too much burden on the angular component. For example if some dataset has intrinsic complexity in the radial distribution rather than the angular distribution, the compression could be limited compared to methods that compress both axes.
Questions for authors

See weakness above.

3 comments
AuthorsOpenReview ↗

Official Response to Reviewer E2yB

We sincerely thank the reviewer for the careful reading and constructive suggestions, and for recognizing RAP's elegant radial-angular decoupling, sound theoretical guarantees, and significant improvements. Below, W denotes the weakness and R our response.


W. Could complex radial distributions limit RAP's compression?

R: The radius is a single scalar per point, regardless of the complexity of its distribution. RAP preserves this scalar exactly while compressing the angular direction from Sd−1\mathbb{S}^{d-1} to Sk−1\mathbb{S}^{k-1}. Radial information therefore remains intact without placing any additional burden on the angular component. The resulting representation still lies in Hk\mathbb{H}^{k}, with one radial degree of freedom and k−1k-1 angular degrees of freedom.

Moreover, distances between points on the same radial ray are preserved exactly (Proposition 16, Appendix J.3):

=d_c(x,y)=|r_x-r_y|.$$ Thus, when variation is entirely radial, RAP preserves all pairwise distances exactly whenever the projection is defined, regardless of the radial distribution.
Reviewer_E2yBOpenReview ↗

Response to rebuttal

Thanks for the clarification, I understand.

AuthorsOpenReview ↗

Thank you!

Thank you for your response and for the time you have devoted to reviewing our work. We sincerely appreciate your positive assessment.

Reviewer_5xHpOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes RAP, a simple method for reducing the dimensionality of hyperbolic embeddings while trying to preserve their structure and distances. The main idea is to represent each point using a radius (its distance from the origin) and an angle/direction. Instead of projecting the whole point, RAP keeps the radius exactly the same and applies a standard random projection only to the direction. This is motivated by the idea that, in hierarchical hyperbolic embeddings, the radius often represents the depth in the hierarchy. The authors show theoretically how errors in the projected angles affect hyperbolic distances, and experimentally find that RAP produces lower distance distortion than the hyperbolic projection baselines they consider.cr

Critical strengths

The task is novel; the motivation and method is well-motivated. The connection between the method and theory is clear, very well readable. And the consistency in results shows the method is strong.

Critical weaknesses

The theoretical guarantee comes with assumptions that make it weaker than the classical JL analogy may initially suggest. I have asked a question regarding this.

Questions for authors

How restrictive are \R and \tau in realistic hyperbolic representations? And how does the bound deteriorate as \R increases or \tau\rightarrow 0? Could you report the results for binomial tree and Barabási–Albert tree as well? They are more imbalanced trees that shows how your method will do in those edge cases. Do you see a difference across other r-ary trees? Does exact radial preservation necessarily correspond to preservation of useful hierarchy? And could preserving radius exactly actually be undesirable when radial information contains noise or isn't aligned with semantic depth?

1 comment
AuthorsOpenReview ↗

Official Response to Reviewer 5xHp

We sincerely thank the reviewer for the careful reading and constructive suggestions, and for recognizing RAP's motivation, clarity, and consistent empirical results. Below, Q denotes each question and R our response.


Q1. How restrictive are RR and τ\tau, and how does the bound deteriorate?

R: RR and τ\tau characterize the uniform worst-case guarantee; RAP itself requires neither a small-ball restriction nor a prescribed minimum separation. For any finite set of distinct embeddings, a finite RR and positive τ\tau always exist; their magnitudes determine how informative the uniform guarantee is. Under the conditions of Theorem 17 (Appendix J.4), we provide a sharper bound based on each pair's own radii and distance, whose dependence on radius can be much milder. For example, for equal-radius pairs with fixed nonzero angular separation, their distance grows as 2R+O(1)2R+O(1), and the coefficient of ε\varepsilon in the pairwise bound decays as O(1/R)O(1/R) at fixed curvature. Proposition 16 (Appendix J.3) further shows that distances along a common radial ray are preserved exactly whenever RAP is defined, regardless of radius or separation.

In the new controlled experiments (Appendix F.9, Figure 10, Page 27), RAP achieves the lowest mean relative distortion among all four methods in all 50 settings. We test equal and different radii with n=256n=256, d=128d=128, k=32k=32, c=1c=1, and 20 paired runs, evaluating all 32,640 unordered pairs in each run.

(1) At fixed τ=0.1\tau=0.1, increasing RR from 1 to 8 reduces RAP's mean distortion from approximately 0.061 to 0.010 in both constructions.

(2) At fixed R=2R=2, decreasing τ\tau from 0.5 to 10−410^{-4} keeps the equal-radius mean near 0.0409 and slightly reduces the different-radius mean from 0.0430 to 0.0408.

These results further demonstrate empirically that RAP is stable as RR increases and τ\tau decreases over the tested ranges.


Q2. How does RAP perform on binomial, Barabási–Albert, and different r-ary trees?

R: We added the requested tree families (Appendix F.10, Table 11, Page 28) using the paper's embedding procedure in H50\mathbb H^{50} (c=1c=1) and the same four projection methods. Across five families, n∈{128,256,512}n\in\{128,256,512\}, and k∈{16,32}k\in\{16,32\}, RAP achieves the lowest mean distortion in all 30 compression settings, with 20.5%–39.5% reductions relative to Euclidean JL.

Below are results at n=512n=512, k=32k=32, evaluating all 130,816 pairs per run. Entries report mean relative distortion ± sample standard deviation over 20 paired runs. Percentage reductions are relative to Euclidean JL, following the main table.

TreeEuclidean JLLocal tangentFully hyperbolicRAPReduction
Binomial0.1038 ± 0.00770.1000 ± 0.00650.0914 ± 0.00510.0704 ± 0.003832.2%
Barabási–Albert0.1022 ± 0.00550.0994 ± 0.00490.0924 ± 0.00400.0767 ± 0.010025.0%
r = 20.1155 ± 0.01600.0970 ± 0.00860.0854 ± 0.00650.0699 ± 0.004139.4%
r = 40.1001 ± 0.00720.0962 ± 0.00600.0893 ± 0.00530.0710 ± 0.004029.1%
r = 80.1002 ± 0.00500.0986 ± 0.00600.0933 ± 0.00500.0716 ± 0.003228.5%

These results extend RAP's advantage to the requested imbalanced trees. Across r=2,4,8r=2,4,8, its mean distortion remains between 0.06990.0699 and 0.07160.0716, showing little sensitivity to branching factor in these experiments.


Q3. Does exact radial preservation necessarily preserve useful hierarchy?

R: When radius encodes depth, RAP preserves that existing depth signal exactly. Parent and branch relationships also depend on angles, which RAP approximately preserves. We therefore evaluate hierarchy directly: on WordNet at k=32k=32 (Appendix G.1, Table 13), RAP achieves the lowest mean parent rank among all four methods (1.626), together with 98.7% Parent@10. These results complement the exact radial guarantee with direct evidence of preserved parent relationships.


Q4. Could exact radial preservation be undesirable when radii are noisy or misaligned with semantic depth?

R: Like classical Euclidean JL, RAP aims to faithfully compress the geometry supplied by the input representation. Exact radial preservation ensures that compression introduces no additional radial distortion, without requiring radius to encode semantic depth. When an application calls for correcting noisy or semantically misaligned radii, task-specific calibration can be performed during representation learning or preprocessing, after which RAP preserves the resulting radii exactly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Esteemed Area Chair and Reviewers,

We sincerely thank all reviewers for their thoughtful questions, constructive suggestions, and positive comments on RAP. We appreciate their recognition of:

  • A novel, well-motivated method with an elegant radial-angular design (All reviewers).
  • Sound theoretical foundations and clear presentation (Reviewers 5xHp and E2yB).
  • Comprehensive evaluation and strong, consistent empirical improvements (All reviewers).

In response to their questions and suggestions, we provided the following key clarifications and additions:

QuestionReviewerOur Response
How does RAP behave across different geometric regimes and tree structures?5xHpWe clarified the sharper pairwise guarantee and added radius/separation tests and five tree families. RAP consistently achieves the lowest mean distortion, supporting its effectiveness and robustness across diverse datasets and geometric conditions (Appendices F.9–F.10).
Does radial complexity limit angular compression?E2yBWe explained that radius is one scalar regardless of its distribution, so preserving it introduces no additional angular burden. The reviewer responded positively and acknowledged our clarification.
How does RAP relate to prior hyperbolic JL and fitted methods?d9jJ(1) Half-space JL and our fully hyperbolic baseline use different coordinates but both can distort radii. RAP preserves radii exactly and achieves 20.9%–49.4% lower mean distortion (main Tables 1–2). The reviewer confirmed that the related-work question was resolved. (2) We further clarified that PCA and JL serve different objectives. Oblivious projections support distributed training, attribution, and inference, particularly when data distributions are unknown in advance or may shift. PCA requires fitting data and need refitting, whereas RAP provides a shared, training-free projection. We highlighted the difference in Appendix D, page 21.

We thank the reviewers and the AC again for helping strengthen the paper. We hope these clarifications and additional results support favorable consideration of RAP as a principled, training-free compression foundation for hyperbolic representations.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.