GROUPOID

Pre-alpha research prototype. This is an early-stage exploration of groupoid-based aggregation for federated learning on Riemannian manifolds. It is not a production federated learning system. See STATUS.md, LIMITATIONS.md, and CORRECTION_NOTICE.md.

Research authorship: Santiago Maniches (ORCID 0009-0005-6480-1987) is the sole scientific author and independent maintainer of this research prototype. AI-assisted development and internal review are acknowledged as tooling, not independent scientific validation. Responsibility for its design, implementation decisions, interpretation, and published claims remains his.

Overview

GROUPOID explores using transport groupoids, cellular sheaves, and Riemannian geometry to aggregate model parameters across heterogeneous federated clients. The core idea: instead of naive Euclidean averaging (FedAvg), transport client parameters to a common frame via explicit point-action morphisms, evaluate cycle-holonomy consistency, and compute the intrinsic Karcher mean on the parameter manifold.

Implemented and tested

These components have working implementations with property-based and integration tests:

• Karcher mean on Riemannian manifolds via geomstats (groupoid.manifold)
• Transport groupoid: morphism composition, inverse, composition associativity verified by Hypothesis (groupoid.groupoid)
• Cycle-basis holonomy defect: cycle_basis_holonomy_defect computes max_{gamma in B} ||Hol(gamma) - I||_F over the NetworkX cycle basis. Exact vanishing has a flatness interpretation under explicit assumptions, while the nonzero magnitude is basis- and representation-dependent. The historical compute_h1 name is retained as a deprecated compatibility alias; the scalar is not a canonical H^1 norm. An incompletely specified basis cycle raises IncompleteCocycleError; conflicting explicitly supplied opposite arrows raise NonReciprocalTransportError (groupoid.cohomology)
• Cellular sheaf: restriction maps with functoriality tested (groupoid.sheaf)
• Sheaf Laplacian: connection Laplacian L = delta^T delta; PSD and delta^T-delta equality verified on non-orthogonal restriction maps, plus spectral analysis, algebraic connectivity, and diffusion convergence tested (groupoid.laplacian)
• Federated aggregation pipeline: point-valued aggregation through explicit invertible point actions, with a cycle-basis holonomy-defect threshold diagnostic and multi-round convergence tested. The current S^2 evidence validates explicit SO(3) point actions (groupoid.aggregation)
• Parallel transport: Schild's ladder and pole ladder (groupoid.transport) are tangent-vector transport utilities. Pole ladder is validated against geomstats' analytic tangent-vector parallel transport on S^2 -- it matches to round-off for every rung count (S^2 is a symmetric space, where the pole ladder is exact). Schild's ladder is the first-order geodesic-parallelogram scheme: its error halves each time n_rungs doubles, and the tests pin that convergence order. The former register_transport_from_points point-valued integration is withdrawn because tangent parallel transport does not by itself define the invertible point action required by the aggregator. The ambient tangent matrix helper is restricted to 1D coordinate-array point representations. See CORRECTION_NOTICE.md and LIMITATIONS.md
• Persistent homology: Vietoris-Rips filtration for divergence tracking (groupoid.persistence), wired into the pipeline via the aggregator's opt-in track_divergence flag (per-round H0-vs-H0 bottleneck distance on the transported parameters, exposed as FederatedRound.divergence). Unit-tested against point clouds of known topology: a circle's dominant 1-cycle (via maximum persistence), two-cluster component counting (betti_0 == 2 at a finite filtration), and a translation-invariant bottleneck distance. The persistence diagram retains a homology-dimension label, and track_divergence compares H0 against H0 only (it does not pool features across dimensions); this is verified against an independent minimum-spanning-tree reconstruction of the H0 diagram. A zero value is a geometric diagnostic, not a certificate of client-model convergence: translations and H1-only changes can leave it unchanged. A finite H0 diagram compared with an empty one incurs the nonzero bottleneck-to-diagonal cost. The Betti numbers are degenerate under the default thresh=inf filtration; see LIMITATIONS.md.

Implemented and validated, not yet integrated

This module is validated against known-correct references but not yet wired into the main aggregation pipeline:

• Riemannian optimizers: SGD and Adam with exponential map retraction; the momentum velocity and Adam first moment are parallel-transported between iterates (with a projection fallback for metrics without parallel transport). Validated: descent to a known target on S^2 (geodesic-distance objective) for SGD, momentum SGD, and Adam; transported moments preserve norm exactly where projection would annihilate them; the curvature-adaptive learning rate is covered for both its damping and fallback branches. No general convergence-rate guarantees are established (groupoid.optimizer)

Status

Pre-alpha. See STATUS.md for details.

Related work / why not just use X?

GROUPOID sits at the intersection of three existing toolchains and is not a replacement for any of them. It is an exploratory prototype of one specific idea -- transport-groupoid aggregation with a cycle-holonomy consistency diagnostic -- not a federated learning framework.

• Flower / FedML / TensorFlow Federated -- mature federated learning frameworks providing the client/server communication, orchestration, and real training loops that GROUPOID deliberately does not implement (see LIMITATIONS.md: "Not a federated learning framework"). GROUPOID is about the aggregation operator, not the FL plumbing; in principle a transport-aware aggregator like this one would be dropped into such a framework, not used instead of it.
• geomstats / pymanopt -- Riemannian-geometry libraries. GROUPOID uses geomstats for the manifold primitives (the Karcher mean delegates to geomstats FrechetMean). What GROUPOID adds on top is the transport groupoid, the cycle-basis holonomy diagnostic, and the cellular-sheaf Laplacian wiring -- not the manifold geometry itself.
• Cellular-sheaf spectral methods (the sheaf-Laplacian line of work, e.g. Hansen and Ghrist's spectral theory of cellular sheaves, and sheaf neural networks) -- GROUPOID's sheaf Laplacian follows this line and is the geometric machinery for detecting inconsistency across clients. The contribution here is applying it to the federated-aggregation setting, not the sheaf-Laplacian construction in the abstract.

In short: use Flower/FedML/TFF for the FL system, use geomstats/pymanopt for manifold math; GROUPOID is a research prototype testing a transport-groupoid aggregation operator against Euclidean FedAvg. A preregistered synthetic benchmark (experiments/) supports the transport benefit under frame misalignment -- an effect largely built into the synthetic setup -- and shows that the fixed NetworkX cycle-basis holonomy defect is associated with aggregation error when the two corruption levels are pooled (Spearman rho = 0.587, permutation p = 1e-4). The within-level correlations are near zero, so the statistic does not rank runs by error within a fixed corruption level. The hypothesis remains unvalidated on real federated learning tasks (see STATUS.md).

Installation

Requires Python 3.11 or 3.12. Python 3.10 reached end of life on October 1, 2026 and is no longer supported. Python 3.13+ is not supported: the numpy<2.0 / scipy<1.16 pins (needed for geomstats compatibility, see LIMITATIONS.md) have no wheels there, so pip will refuse with a Requires-Python message rather than attempt a source build.

From PyPI:

pip install groupoid

From source (for development):

git clone https://github.com/smaniches/GROUPOID.git
cd GROUPOID
pip install -e ".[dev]"

Quick example

import networkx as nx
import numpy as np
from geomstats.geometry.hypersphere import Hypersphere
from groupoid import TransportGroupoidAggregator

manifold = Hypersphere(dim=2)
graph = nx.DiGraph([("A", "B"), ("A", "C")])

aggregator = TransportGroupoidAggregator(
    manifold=manifold, graph=graph, base_node="A"
)

# Register SO(3) point actions.
theta = np.pi / 6
R = np.array([
    [np.cos(theta), -np.sin(theta), 0],
    [np.sin(theta),  np.cos(theta), 0],
    [0, 0, 1],
])
aggregator.register_transport("A", "B", R)
aggregator.register_transport("A", "C", R.T)

client_params = {
    "A": np.array([0.0, 0.0, 1.0]),
    "B": np.array([0.1, 0.0, 0.995]),
    "C": np.array([-0.1, 0.0, 0.995]),
}
client_params = {k: v / np.linalg.norm(v) for k, v in client_params.items()}

result = aggregator.aggregate(client_params)
print(
    f"cycle-basis defect = {result.cycle_basis_holonomy_defect:.2e} "
    f"(below configured threshold: {result.passes_consistency_threshold})"
)

Running tests

pytest tests/ -v

The suite reaches 100% line and branch coverage of the groupoid package on Python 3.11-3.12, enforced in CI:

pytest tests/ --cov=groupoid --cov-branch --cov-fail-under=100

Coverage measures which lines run, not whether behavior is correct. See STATUS.md for the per-component validation depth, which coverage alone does not capture.

Documentation

smaniches.github.io/GROUPOID

License

Copyright 2026 TOPOLOGICA LLC. Licensed under the Apache License, Version 2.0.