hyperbolix: Hyperbolic Deep Learning in JAX
Timo Klein, Thomas Lang, Yllka Velaj, Sebastian Tschiatschek
cs.LG · cs.MS
Quick verdict
Claims the first general hyperbolic deep learning library for JAX, with verified numerics; a credible, honestly benchmarked systems contribution that still lacks end-to-end training benchmarks.
Summary
The paper introduces hyperbolix, an open-source MIT-licensed hyperbolic deep learning library for JAX and Flax NNX, positioned as the first comprehensive, general-purpose offering in that ecosystem. It provides six manifolds sharing a 19-operation interface (Poincaré ball, hyperboloid, proper velocity, signed-curvature κ-stereographic, Euclidean, and mixed-curvature products), 76 layer exports covering linear, convolution, attention, normalization, positional-encoding, regression, activation, and vector-quantization families, optax-compatible Riemannian Adam/SGD, wrapped distributions, and HoroPCA/CO-SNE. Correctness rests on 1,038 tests checked against independent oracles (hand-derived closed forms, NumPy/SciPy transcriptions, finite differences) rather than other libraries; this caught a digamma argument swap also present in the Intrinsic Lorentz NN reference. The central numerical result is a cancellation-free hyperboloid distance formula that keeps float32 error at rounding level to geodesic radius 10, where geoopt already returns NaN from radius 9. On the Poincaré ball accuracy matches geoopt/HypLL (within 2.2× per cell), and speed versus geoopt is mixed and hardware-dependent. The authors candidly report no training-quality benchmarks, single-seed speed sweeps, and uneven manifold coverage (attention hyperboloid-only, VQ Poincaré-only).
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 2
● low 1
6 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- hyperbolix is, to the authors' knowledge, the first comprehensive, general-purpose hyperbolic deep learning library for JAX (Section 3.4, Table 2, Appendix C).
- Every checked operation is verified against an independently derived oracle (closed form, paper transcription, or finite difference) rather than another library, which catches bugs cross-library tests would inherit (Section 5).
- Cancellation-free hyperboloid formulas keep float32 distance error within 8× of the input-rounding error up to geodesic radius ~10, where geoopt already returns NaN from radius ~9 (Section 4, Figures 2–3, Section H.7).
- The three hyperbolic models are connected by exact isometries, so learned representations transfer between Poincaré ball, hyperboloid, and proper-velocity coordinates without retraining (Section 3.1, Table 16).
- The digamma argument swap the oracle tests caught is also present in the Intrinsic Lorentz Neural Network paper and its reference implementation (Shi et al., 2026) (Section 5).
Strengths
Fills a demonstrated ecosystem gap
Appendix C (Table 14, verified against default branches on 2026-08-03) shows no JAX library offers hyperbolic NN layers: Rieoptax has 9 undocumented classes and is inactive since Oct 2023, RiemannAX's hyperbolic manifolds are unreleased, GeoJAX has no layers; geoopt's released version exposes zero NN layers (Table 2).
Independent-oracle testing discipline
Section 5: 1,038 test functions across 55 files (4,446 parametrized items) checked against hand-derived closed forms, NumPy/SciPy transcriptions, or finite differences — never another library. Concrete payoff: catching the digamma argument swap in Hyperboloid.log_radius_concat that the ILNN paper (Shi et al., 2026) reference also contains; Section G.3 pins wrapped-normal log-prob from four independent directions including integration to unity and a SciPy flat-curvature limit.
Principled cancellation fix with a validating check
Unusually honest comparative and self-reporting
Benchmark provenance is clean
Actionable numerical guidance
Weaknesses
No end-to-end training validation
Single-seed, single-machine performance evidence
Insular development provenance
Not uniformly faster than prior art
Documented sharp edges in its own API guarantees
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a software/library paper: hyperbolix v1.3.0, an MIT-licensed hyperbolic deep learning library for JAX/Flax NNX, offered as the first comprehensive general-purpose library in that ecosystem. Scope is threefold: (1) API design (Section 2: pure-function, unbatched manifold ops with call-time curvature; manifolds as plain classes outside the NNX graph so they survive `nnx.scan`; LearnableCurvature with log/softplus/identity parameterizations; `riemannian_adam/sgd` as optax transformations keyed on a `ManifoldParam` subclass); (2) breadth (76 `nn_layers` exports across six manifolds with a 19-operation protocol and exact pairwise isometries, Appendices A–B, D); (3) numerical superiority on the hyperboloid plus a verification methodology (Sections 4–5). The claims are scoped carefully — "to our knowledge" for priority, "checked operations" for the test suite — and the abstract's headline (float32 accurate where geoopt returns NaN) is backed by the constructed-experiment data in Section H.7.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 6 strengths & 5 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is hyperbolix?
An open-source (MIT license) hyperbolic deep learning library for JAX and Flax NNX from University of Vienna researchers. Per the paper, it is the first comprehensive, general-purpose hyperbolic DL library in JAX: six manifold models with a common 19-operation interface, 76 layer exports (linear, convolution, attention, normalization, positional encoding, regression, activations, vector quantization), optax-compatible Riemannian optimizers, wrapped distributions, and HoroPCA/CO-SNE dimensionality reduction.
How does hyperbolix differ from geoopt?
geoopt is a PyTorch Riemannian-optimization library whose released version ships no neural network layers; hyperbolix is JAX-native with 44 layer classes plus 31 functions, jit/vmap-compatible pure functions, call-time curvature, and cancellation-free hyperboloid arithmetic. The paper's benchmarks show comparable accuracy on the Poincaré ball and mixed, hardware-dependent speed results.
Why do hyperboloid distance computations return NaN at large radii in other libraries?
The Minkowski inner product subtracts two large, nearly equal terms that both grow like e^{2a} with geodesic radius a, erasing shared significant digits ('ambient cancellation'). In float32 the mantissa is exhausted near a≈9, so arcosh receives an invalid argument and geoopt returns NaN. hyperbolix's haversine-style formula (Eq. 1) sums non-negative radial and angular terms instead, staying at float32 rounding level to radius ~10.
What are the main limitations of the hyperbolix paper?
The authors state in Section 7 that the paper reports no training-quality benchmarks ('nothing here shows hyperbolix reaching a published accuracy on a published task'), all speed sweeps are single-seed on shared hardware, library coverage is uneven (attention only on the hyperboloid, vector quantization only on the Poincaré ball), and float32 still fails beyond a finite radius that the new formulas push outward without removing.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2–6 were assessed from their captions, axes, and the in-text numbers; the rendered images themselves were not analyzed.
- The extraction appears complete through Appendix H (all tables present), but claimed missing sections cannot be fully ruled out.
- External-survey claims (Table 14: library activity, licenses, layer counts) are verifiable only as of the authors' stated snapshot date 2026-08-03.
- The GitHub repository (github.com/timoklein/hyperbolix) was not inspected; reproducibility statements rest on the paper's description of the released v1.3.0 wheel and measurement scripts.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.