Gyration-Equivariant Latent ODEs for Hierarchies That Evolve in Continuous Time

Many networks that change over time are also hierarchical. Citation graphs, Reddit communities and Wikipedia editing activity all have tree-like structure, and new edges and events arrive at irregular moments. Hyperbolic space is a curved geometry whose volume grows exponentially with radius, like a tree, so it can hold such hierarchies in few dimensions. Neural ordinary differential equations (neural ODEs) handle the timing side: a network defines how a latent state changes, and an ODE solver integrates that rule to any query time. The literature has already combined the two. Manifold neural ODEs and Riemannian continuous normalizing flows integrate learned vector fields on hyperbolic space. Their vector fields, however, are generic networks on coordinates, so the learned law of motion can depend on where in the hierarchy a node sits. Our proposal targets that gap.

We propose latent ODEs whose dynamics are equivariant to the symmetries of hyperbolic space. Equivariance here means that moving a whole neighborhood to a different place in the hierarchy moves its predicted dynamics in the same way. We build this with gyrovector arithmetic, the algebra of Möbius addition $\oplus$ on the Poincaré ball. At each instant the vector field re-centers a node's neighborhood at the origin, computes a velocity there, and translates that velocity back. The construction is only exact if one tracks the gyration. Gyration is the small rotation that hyperbolic translations pick up when they are composed, and it is the reason Möbius addition fails to be associative. The gyration therefore enters as the exact correction that makes "the same rule at every depth" consistent, not as decoration. The claim to test is concrete. On continuous-time dynamic graphs, this weight-sharing across positions should improve generalization from shallow to deep parts of a hierarchy and should keep trajectories stable near the boundary of the ball. Neither generic manifold ODEs nor tangent-space ODEs provide these properties.

Why the obvious idea is not the contribution

The tempting combination is a "gyrovector ODE solver" that steps by $h_{k+1} = h_k \oplus (\Delta t \otimes_c v)$. That step is the Riemannian exponential map written in gyro notation. Intrinsic hyperbolic Euler and Runge–Kutta integrators already exist in Continuous Geometry-Aware Graph Diffusion via Hyperbolic Neural PDE, which builds hyperbolic Euler, fourth-order Runge–Kutta and Adams–Moulton solvers with parallel transport for graph diffusion over network depth. The dynamic chart method of Neural Manifold Ordinary Differential Equations and the projective solver of Riemannian Continuous Normalizing Flows compute the same geodesic flow by other means. A paper whose novelty is the integrator would be a renaming.

We also drop a second tempting idea: that curvature and gyration let a low-dimensional ODE escape the homeomorphism bottleneck documented in Characteristic Neural Ordinary Differential Equation. The flow of any smooth autonomous ODE on a manifold is still a homeomorphism, so trajectories on the ball cannot cross either. The vector field is the only place where a new idea can enter, and equivariance is the property that gyrovector algebra makes cheap and exact.

Core Claim

Let $h_i(t)$ in the Poincaré ball $\mathbb{D}_c^n$ be the latent state of node $i$, and let $\mathcal{N}_i(t)$ be its current neighbors. We define the dynamics as

$$\frac{dh_i}{dt} = dL_{h_i}\big|_0 \; g_\theta\big(\{\ominus h_i \oplus h_j\}_{j \in \mathcal{N}_i(t)},\, t - \tau_i\big).$$

The symbols are as follows. $L_a(x) = a \oplus x$ is left gyrotranslation, the isometry that carries the origin to $a$, and $dL_{h_i}|_0$ is its differential at the origin. $\ominus h_i \oplus h_j$ is neighbor $j$ seen from node $i$'s own frame. $\tau_i$ is the time of node $i$'s last event. $g_\theta$ is a network that outputs a tangent vector at the origin and is equivariant to rotations, meaning it acts through inner products and norms of its inputs, as in standard equivariant message passing.

The gyrotranslation theorem gives the key identity:

$$\ominus(a \oplus h) \oplus (a \oplus z) = \mathrm{gyr}[a, h]\,(\ominus h \oplus z).$$

Here $\mathrm{gyr}[a,h]$ is the gyration rotation. When the whole configuration is translated by $a$, every re-centered neighbor rotates by the same gyration. Translations also compose as $L_a L_h = L_{a \oplus h}\,\mathrm{gyr}[a,h]$, so the push-forward rotates by that same gyration. A rotation-equivariant $g_\theta$ therefore makes the vector field exactly equivariant under all Möbius isometries. If either gyration is dropped, equivariance is lost.

This gives the hypothesis a sharp form. The learned law is shared across positions in the ball, much as convolution shares weights across positions in an image. The network $g_\theta$ also only ever sees re-centered inputs near the origin, where $\tanh$ and $\tanh^{-1}$ are well conditioned. The second property speaks directly to the clamping instabilities reported in Latent Variable Modelling with Hyperbolic Normalizing Flows.

Research Program

The model is a continuous-time dynamic-graph encoder. Between events, every node state evolves under the equation above. When an event (an edge) arrives, a jump update merges the two endpoint states with a Möbius gyromidpoint, the residual construction used in the hyperbolic graph diffusion work cited above. Integration uses a geodesic Euler step, and a two-stage Runge–Kutta step in the origin frame. We are open that the solver is standard. Precision is float64 near the boundary, with no coordinate clamping inside $g_\theta$.

Gradients come from discretize-then-optimize with checkpointing. We do not attempt the gyro-adjoint from Gap 2 unless memory forces it. The Riemannian adjoint in Neural Manifold Ordinary Differential Equations already covers the continuous case, and adjoint drift in reverse time is a known hazard.

Link probabilities are computed from hyperbolic distance. A variant scores links with the Möbius gyroplane logits of Hyperbolic Neural Networks. Stabilization uses gyrogroup batch normalization (Gyrogroup Batch Normalization, code GyroBN), which centers with $\ominus\mu \oplus x$ and is itself equivariant to translation.

The analysis has three parts. The first is an equivariance test, applying random Möbius isometries to whole graphs and measuring how far the outputs deviate. The second is a check that the geodesic Euler step reproduces chart-method trajectories to solver tolerance. The third measures the number of function evaluations against how close node states lie to the boundary.

Baselines

The naive reference point is a Euclidean latent ODE on the same event-driven graph encoder, with the same jump updates and an ordinary dot-product decoder. We run it at matched dimension and also at four times the dimension, because section 5.4 of the file notes that Euclidean models catch up when given spare dimensions. If our method only wins at matched dimension, the claim reduces to compression.

The strongest existing method on the task is Hyperbolic Variational Graph Neural Network for Modeling Dynamic Graphs. That model encodes time with a hyperbolic Bochner kernel on the Lorentz hyperboloid and uses discrete temporal graph network aggregation. It reports a 3% to 3.6% gain in accuracy and average precision over GAT, GraphSAGE and HGCN on Wikipedia. We found no reported absolute result on Lacuna for its Wikipedia, Reddit or DBLP numbers, so we will rerun it ourselves on identical chronological splits with the same negative sampling. Every method gets the same hyperparameter search budget of 30 trials per dataset and the same GPU-hour cap. Beating it means higher average precision with non-overlapping 95% intervals over five seeds on at least two of the three graphs.

The geometric competitor is a manifold ODE whose vector field is a generic multilayer perceptron on ball coordinates. This follows Neural Manifold Ordinary Differential Equations and the geodesic-distance input layer of Riemannian Continuous Normalizing Flows. It has the same solver, geometry, parameter count and event updates as our model. It is also the ablation that isolates equivariance.

A second ablation keeps re-centering but drops the gyration correction. Concretely, $g_\theta$ is a plain multilayer perceptron on re-centered coordinates, which makes it translation-aware but not exactly equivariant. Comparing this ablation with the full model tests whether exact gyration consistency matters or whether approximate re-centering is enough.

As a sanity check that our hyperbolic stack is competitive at all, we will reproduce the static Disease graph result in the file. A hyperbolic network with GyroBN reports 81.2% AUC there, against 75.1% for the base network without GyroBN (Gyrogroup Batch Normalization). Our encoder with frozen time should reach at least that number.

Why not just…?

Why not just run a Euclidean neural ODE in the tangent space at the origin? This is the most serious threat on graphs whose hierarchy stays shallow. Tangent-space dynamics pass through $\log_0$ and $\exp_0$, so they distort distances most strongly far from the origin. Latent Variable Modelling with Hyperbolic Normalizing Flows found that tangent coupling anchored at the origin is consistently beaten by wrapped coupling, which moves transformations across the manifold. That evidence comes from discrete flows on static graphs, not from dynamics, so we include the tangent ODE as a full baseline and report results by node depth. If it matches our method on deep nodes, the hyperbolic dynamics are unnecessary and we will say so.

Why not just use a Riemannian ODE with a generic vector field, as Lou et al. and Mathieu–Nickel do? That baseline has exactly our geometry and solver, and lacks only the equivariance. The depth-transfer experiment below is designed so that this baseline fails if equivariance matters. If it does not fail, our contribution shrinks to a numerical-conditioning trick, and we will state that result.

Why not just discretize, using a hyperbolic temporal graph network with time kernels? Hyperbolic Variational Graph Neural Network for Modeling Dynamic Graphs already handles irregular timing through time-difference kernels, so continuity alone is not obviously needed. Our test is to thin events at random to 50% and 20% density at evaluation time. A continuous state that evolves between events should degrade more gracefully than a kernel-conditioned discrete update. If it does not, the ODE component is not justified.

Why not just add Euclidean dimensions? We run the Euclidean model at four times our dimension. The file suggests hyperbolic advantages shrink as dimension grows (Neural Manifold Ordinary Differential Equations). We therefore report the whole curve of accuracy against dimension rather than a single point.

Proposed Experiments

The first experiment is synthetic and decisive. We generate continuous-time branching processes on a random tree, using a binary diffusion process like the one in Bose et al. extended with Poisson event times. The ground-truth law of how events spread depends only on relative position. We train on events within the top four levels and test on levels five to eight. Because the law is identical everywhere, an equivariant model should transfer, while a generic-field model has never seen states near the boundary. The prediction is that the equivariant model keeps at least 90% of its in-distribution log-likelihood on the deep levels and the generic-field ablation does not. A further check applies a random Möbius isometry to the test graphs. The full model's predictions should change only by solver error, while the version without gyration should change noticeably.

The second experiment is on the real dynamic graphs: Wikipedia, Reddit and DBLP, using chronological splits. We measure transductive and inductive link-prediction average precision and AUC, report them stratified by the hyperbolic norm of each node's state as a proxy for depth, and add the event-thinning test described above.

The third experiment reports cost and stability. We record function evaluations per unit time, wall-clock time, the fraction of runs that produce NaN values, and whether each method needs clamping, all against the norm of the states.

The result that would sink the proposal is that the generic-field Riemannian ODE matches our model on depth transfer and on real-graph average precision. In that case equivariance buys nothing over geometry alone. The result that would establish it is a clear depth-transfer gap on the synthetic task, together with gains concentrated on deep, near-boundary nodes on DBLP and Reddit.