Source
Ganea, Bécigneul & Hofmann (2018), LIT-879 (NeurIPS 2018), read in full in NOTE-686. The reader checked Lemmas 2–3, Eq. 15 and Theorems 4–5 numerically and re-derived Theorem 4. NOTE-686 filed no THEORY: it judged the closed forms established geometry and the experiments too thin. This account states the structural point the reading found (its Key insight and claim C5, rated strong), which is what the record lacks, and files it Proposed because the part that would make it matter is an inference.
What was actually shown
The lift. A map f between Euclidean spaces is lifted to the Poincaré ball of curvature −c as its “Möbius version” f^{⊗c} = exp₀ ∘ f ∘ log₀ (Definition 3.2). That covers matrix multiplication and pointwise nonlinearities. Möbius scalar multiplication has this form (Lemma 3), which motivates the definition; nothing forces it. The paper notes the consequence itself (§3.2): a composition of lifted maps is exp₀ ∘ (the Euclidean composition) ∘ log₀.
Where the chart cannot absorb the geometry. The bias x ⊕_c b equals exp_x((λ₀/λ_x) log₀(b)), parallel transport of log₀(b) from the origin to x followed by the exponential map at x (Theorem 4 gives the transport as the scaling λ₀/λ_x). The output layer’s logits are hyperbolic distances to geodesic hyperplanes exp_p({a}^⊥), in closed form (Theorem 5). Every operation recovers its Euclidean counterpart as c → 0.
What the experiments show. All at dimension 5, each cell a best of three runs. On SNLI, hyperbolic and Euclidean encoders score 78.2–81.5%, with a Euclidean GRU best. On a synthetic noisy-prefix task, hyperbolic models win most at 10% noise (error 3.1% against 10.4% for the RNN), and at 50% noise the fully hyperbolic RNN is worst (62.9% against 72.1%). In pre-trained Poincaré embeddings of WordNet, hyperbolic logistic regression separates subtrees better than Euclidean in 15 of 16 settings, most at 2–5 dimensions (Table 2).
What this does not say
- That the networks are Euclidean. With biases and the hyperbolic output layer they are not, and those are the parts the account names.
- That the gains come from the two operations. No ablation removes either; that is this record’s inference, and the promote_when.
- That the lift is canonical. It privileges the origin: translating the inputs by an isometry does not commute with it. Lifting at a learned point, or intrinsic operations such as Fréchet means, are not compared.
- That hyperbolic networks help on tree-like data in general. The gains are on one synthetic task whose tree-likeness is not measured, and on embeddings fitted to the hierarchy being classified, test nodes included, which measures separability, not generalisation.
- That nested subtrees get nested half-spaces. Each subtree’s classifier is trained alone.
Connections
- QUESTION-025 lists encodings of hierarchy. This paper adds one, hyperbolic half-spaces, one per subtree, the analogue of a Euclidean attribute direction with a threshold (THEORY-185). It does not answer the question (NOTE-686).
- THEORY-189 is the same authors’ entailment cones (LIT-866), which make an order relation into nested regions. This paper uses the cones paper’s exponential map and optimiser, not its cones.
- THEORY-421 (Yang et al., LIT-871) names this paper among those that assumed training would place general concepts near the origin.
Note of 2026-10-19: the composition argument
Carried out for the promotion audit’s batch 027. On the Poincaré ball of curvature −c, exp₀(v) = tanh(√c‖v‖) v/(√c‖v‖) maps ℝⁿ onto the open ball, and log₀(y) = artanh(√c‖y‖) y/(√c‖y‖) is its inverse. A lifted map is f^{⊗c} = exp₀ ∘ f ∘ log₀, so g^{⊗c} ∘ f^{⊗c} = exp₀ ∘ g ∘ f ∘ log₀, and by induction a stack of lifted layers is exp₀ ∘ (the Euclidean composition) ∘ log₀. The Euclidean network inside does not depend on c. In the chart the bias is F_b(v) = log₀(exp₀(v) ⊕_c b), which by Theorem 4 is a c-dependent map, and reduces to v + b only as c → 0. The output layer’s logits are distances to geodesic hyperplanes, in closed form with c (Theorem 5).
“Conjugated” is meant in this sense. Any map on the ball can be written as exp₀ ∘ F ∘ log₀ for some F. The content is that, for the lifted layers, F is the standard c-free layer, while for the bias and the output layer it is a c-dependent map. Whether the networks’ gains come from those two operations is QUESTION-162.