Octavian-Eugen Ganea, Gary Bécigneul and Thomas Hofmann (2018), Advances in Neural Information Processing Systems 31 (NeurIPS 2018) — ARXIV-1805.09112
Key takeaways
- The toolkit. The Poincaré ball of curvature −c, Dⁿ_c, with Ungar’s
Möbius addition x ⊕_c y and Möbius scalar multiplication r ⊗_c x, is
tied to its Riemannian geometry. The paper gives:
- the exponential and logarithmic maps at any point in closed form (Lemma 2), so exp_x(v) = x ⊕_c tanh(√c λ_x‖v‖/2) v/(√c‖v‖);
- Möbius scalar multiplication as exp₀(r log₀(x)) (Lemma 3);
- parallel transport from the origin to x as the scaling λ₀/λ_x (Theorem 4), proved from the Christoffel symbols;
- the geodesic hyperplane through p normal to a, exp_p({a}^⊥), and the distance from a point to it in closed form (Theorem 5). Every operation recovers its Euclidean counterpart as c → 0. I checked Lemma 2, Lemma 3, Eq. 15, Theorem 4 and Theorem 5 numerically (agreement to 10⁻¹¹ or better, Theorem 5 to the resolution of a grid search in D²), and Theorem 4 analytically as well.
- Where the curvature enters. A map f between Euclidean spaces is lifted as its “Möbius version” exp₀ ∘ f ∘ log₀, which covers matrix multiplication and pointwise nonlinearities. The paper notes the consequence itself: a stack of such layers is exp₀ ∘ (Euclidean network) ∘ log₀. Only the bias translations x ⊕_c b between layers and the distance-based output layer make the network differ from a Euclidean one written in log₀ coordinates.
- Hyperbolic logistic regression. Euclidean softmax regression is rewritten as signed distances to class hyperplanes, and the hyperplanes replaced by geodesic ones, so each logit is (λ_p‖a‖/√c) sinh⁻¹(2√c⟨−p ⊕_c x, a⟩/((1 − c‖−p ⊕_c x‖²)‖a‖)). The normal a is a tangent vector at p, made trainable by transporting it from the origin. Its decision regions are hyperbolic half-spaces.
- A hyperbolic GRU whose update gate is derived, after Tallec and Ollivier, from invariance to time warping, using a gyro-derivative and a chain rule for it that the paper proves (Appendix E).
- The evidence is thin and low-dimensional. All embeddings have
dimension 5, and each Table 1 cell is the test score of the best of
three runs by validation.
- On SNLI (entailment against the rest), hyperbolic and Euclidean encoders score 78.2–81.5%, with the Euclidean GRU best.
- On a synthetic task, deciding whether one sentence is a noisy prefix of another, hyperbolic models win most at 10% noise (error 3.1% against 10.4% for the RNN) and least at 50%, where the fully hyperbolic RNN is worst (62.9% against 72.1%).
- In a separate test, logistic regression separates four WordNet subtrees from the rest in pre-trained Poincaré embeddings. Hyperbolic hyperplanes beat Euclidean ones in 15 of 16 settings, by most at 2–5 dimensions; at 10 dimensions the two are close for the larger subtrees. Those embeddings were trained on the whole hierarchy, test nodes included, so this measures separability in a fitted embedding, not generalisation.
Standing in the record
Filed at the owner’s request on 2026-10-09. The owner asked for the works cited by the batch on hierarchy and hyperbolic geometry (LIT-865 to LIT-877) that neither record held, and this was one of them. Sharpee 2019 (LIT-875) cites it as its ref. 9, its one machine-learning example of hyperbolic representation. The batch’s curation entry named it among the works its papers build on. It is the companion of the same authors’ entailment-cones paper (LIT-866), which it cites for the closed-form exponential map and the hyperbolic Riemannian SGD it uses. Read on its own merits (NOTE-686). The reading filed no THEORY; the backfill of 2026-10-16 filed its structural finding as THEORY-566 (Proposed).
LIT-866 encodes implication as nested cones. This paper’s geometric object for a class is a hyperbolic half-space bounded by a geodesic hyperplane, the analogue of a linear attribute direction with a threshold. That bears on QUESTION-025, which asks whether attributes that imply one another still get linear directions. The paper’s WordNet test shows that, in an embedding fitted to the hierarchy, a subtree is close to a hyperbolic half-space, and closer than to a Euclidean one in low dimensions. It does not answer the question. There is no co-occurrence, the embedding was trained on the hierarchy itself, and no claim is made about nested subtrees giving nested half-spaces.
It is chiefly about machine-learning practice, building layers and
training them, and an anthology topic could hold it. The anthology does
not hold it as far as its clone shows, so it carries
anthology-candidate.