arxiv: 2604.02615 · v1 · submitted 2026-04-03 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Complex-Valued GNNs for Distributed Basis-Invariant Control of Planar Systems

Samuel Honor , Mohamed Abdelnaby , Kevin Leahy

Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:52 UTC · model grok-4.3

classification 💻 cs.LG

keywords graph neural networkscomplex-valued networksbasis invariancedistributed controlplanar systemsimitation learningflocking control

0 comments

The pith

A complex-valued graph neural network learns distributed control policies that stay unchanged no matter which local coordinate frame each agent uses.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops a graph neural network for controlling groups of agents moving in a plane, where each agent may use its own rotated coordinate system without a shared compass or GPS. By representing positions and velocities as complex numbers and using special activation functions that respect phase shifts, the network ensures that the overall policy looks the same from a global view regardless of local rotations. A reader should care because many real-world robot teams or sensor networks operate without reliable global references, and current methods break when bases mismatch. The approach is tested on an imitation learning task where agents learn to flock together, showing it needs less data and performs better than standard real-valued networks.

Core claim

The architecture expresses all 2D geometric features and basis transformations in the complex domain. Inside each GNN layer, it applies complex-valued linear transformations followed by phase-equivariant activations. From any fixed global frame, every learned policy is strictly invariant to the choice of local frames at each node.

What carries the argument

Complex-valued linear layers paired with phase-equivariant activation functions, which together enforce invariance to local basis rotations when 2D data is encoded as complex numbers.

If this is right

Any control policy produced by the network will produce identical actions when viewed globally, even if individual agents rotate their local frames arbitrarily.
The method requires fewer training examples to reach good performance on imitation tasks.
Tracking accuracy and ability to generalize to new situations improve over real-valued GNN baselines in the flocking example.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar techniques might extend invariance to other geometric groups beyond planar rotations, such as translations or 3D orientations.
Deployment in compass-denied environments like indoors or underwater could become more reliable without needing to align sensors manually.
Testing on tasks beyond flocking, such as formation control or obstacle avoidance, would reveal how broadly the invariance helps.

Load-bearing premise

The specific form of complex linear layers and phase-equivariant activations is sufficient to guarantee strict global invariance no matter what the underlying planar dynamics are.

What would settle it

Run the trained policy on a system where local frames are rotated differently for each agent and check if the output actions match what a global-frame policy would produce; any mismatch would disprove the invariance claim.

Figures

Figures reproduced from arXiv: 2604.02615 by Kevin Leahy, Mohamed Abdelnaby, Samuel Honor.

**Figure 2.** Figure 2: As seen in the left-hand plot, the smallest and largest equivariant models tested both track the nominal controller very well. The seven intermediate [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Shown above are the results of the extended 5 second rollouts of the basis-invariant GNN, nominal controller, and baseline GNN on an environment [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: The extended run velocity variance of the best-performing basis-invariant and baseline models indicate different levels of generalization to situations [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Lowering the communication radius increased the overall velocity variance of both the basis-invariant and baseline GNNs. However, the basis [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Reducing the communication radius had a far greater impact on the baseline GNN’s performance than on the basis-invariant’s. The swarm [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

read the original abstract

Graph neural networks (GNNs) are a well-regarded tool for learned control of networked dynamical systems due to their ability to be deployed in a distributed manner. However, current distributed GNN architectures assume that all nodes in the network collect geometric observations in compatible bases, which limits the usefulness of such controllers in GPS-denied and compass-denied environments. This paper presents a GNN parametrization that is globally invariant to choice of local basis. 2D geometric features and transformations between bases are expressed in the complex domain. Inside each GNN layer, complex-valued linear layers with phase-equivariant activation functions are used. When viewed from a fixed global frame, all policies learned by this architecture are strictly invariant to choice of local frames. This architecture is shown to increase the data efficiency, tracking performance, and generalization of learned control when compared to a real-valued baseline on an imitation learning flocking task.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The complex GNN parametrization for basis-invariant planar control is a reasonable idea but lacks the formal derivation needed to support the strict invariance claim.

read the letter

The paper's main contribution is a GNN architecture that represents planar geometry in the complex domain and uses complex linear layers plus phase-equivariant activations to produce control policies that stay the same when viewed from a fixed global frame, regardless of each node's local basis choice. This directly targets the limitation in prior distributed GNN control work that assumed aligned frames, which is relevant for GPS-denied robotics settings. The approach is new in this specific combination for control policies, and the abstract reports clear gains in data efficiency, tracking performance, and generalization versus a real-valued baseline on an imitation-learning flocking task. That empirical result is the strongest part of what is shown so far. The algebraic motivation makes sense on its face because complex multiplication encodes rotations cleanly, and the chosen activations should preserve the necessary equivariance property through the layers. The soft spot is the missing general argument. The stress-test note is right that there is no inductive derivation confirming the invariance survives arbitrary graph topologies and arbitrary geometric interaction rules in planar systems. Without that step, the performance improvements cannot be confidently attributed to the invariance rather than other factors in the specific flocking setup. Experiments also lack reported error bars or protocol details, which makes the gains harder to evaluate. This work is aimed at researchers in learned multi-agent control and geometric GNNs. A reader already working on orientation-agnostic policies would get practical value from the parametrization idea and the flocking results, even if the theory needs tightening. I would send it to peer review. The core technical move is worth referee time and the authors can address the derivation gap in revision.

Referee Report

2 major / 2 minor

Summary. The paper introduces a complex-valued GNN parametrization for distributed control of planar dynamical systems. Geometric features and basis transformations are represented in the complex domain, with complex-valued linear layers and phase-equivariant activations inside each GNN layer. The central claim is that, when viewed from a fixed global frame, all learned policies are strictly invariant to independent choices of local frames at each node. Empirical results on an imitation-learning flocking task show gains in data efficiency, tracking performance, and generalization relative to a real-valued GNN baseline.

Significance. If the strict invariance property holds for arbitrary planar systems and graph topologies, the architecture would enable reliable distributed controllers in GPS- and compass-denied settings without requiring a shared reference frame. The reported improvements in sample efficiency and generalization on the flocking task suggest practical utility for multi-agent control, provided the invariance can be rigorously established rather than assumed from the layer definitions.

major comments (2)

[Abstract, Section 3] Abstract and Section 3 (layer definitions): the claim that complex-valued linear layers with phase-equivariant activations produce policies that are strictly invariant to local frame choice for arbitrary planar dynamical systems is stated without a general inductive argument or derivation showing that the full message-passing composition preserves global invariance under independent local basis rotations for arbitrary graph topologies and interaction rules.
[Section 4] Section 4 (experiments): the reported gains in data efficiency, tracking performance, and generalization on the imitation-learning flocking task are presented without error bars, details of the experimental protocol, or ablation isolating the contribution of the invariance property versus other architectural choices.

minor comments (2)

[Section 3] Notation for complex multiplication and phase-equivariant activations should be defined explicitly with an equation reference rather than left implicit.
[Introduction] The manuscript should include a clear statement of the precise class of planar systems and graph structures for which the invariance is claimed to hold.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will incorporate revisions to strengthen the rigor and experimental reporting.

read point-by-point responses

Referee: [Abstract, Section 3] Abstract and Section 3 (layer definitions): the claim that complex-valued linear layers with phase-equivariant activations produce policies that are strictly invariant to local frame choice for arbitrary planar dynamical systems is stated without a general inductive argument or derivation showing that the full message-passing composition preserves global invariance under independent local basis rotations for arbitrary graph topologies and interaction rules.

Authors: We agree that an explicit inductive argument for the full GNN would improve clarity. The layer definitions in Section 3 establish invariance for individual operations, but we will add a formal proof by induction on depth in the revised manuscript. This will show that the complete message-passing composition preserves strict global invariance under independent local basis rotations for arbitrary connected graph topologies and interaction rules. revision: yes
Referee: [Section 4] Section 4 (experiments): the reported gains in data efficiency, tracking performance, and generalization on the imitation-learning flocking task are presented without error bars, details of the experimental protocol, or ablation isolating the contribution of the invariance property versus other architectural choices.

Authors: We acknowledge the need for greater experimental transparency. In the revision we will report mean and standard deviation over 5 random seeds, provide a complete description of the training protocol (including optimizer, learning rate schedule, and data generation), and add an ablation comparing the full complex-valued model against a real-valued GNN with identical capacity but without phase-equivariant activations. revision: yes

Circularity Check

0 steps flagged

No circularity: invariance derived from complex algebra

full rationale

The manuscript defines a GNN using complex-valued linear layers and phase-equivariant activations, then states that the resulting policies are strictly invariant to local frame choice when observed in a global frame. This property is presented as following directly from the algebraic rules of complex multiplication and the activation design rather than from any fitted quantity, self-referential definition, or load-bearing self-citation. The empirical evaluation on the flocking imitation task is reported separately and does not serve as the justification for the invariance claim. No step in the provided derivation chain reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard properties of complex numbers for 2D geometry and the design choice of phase-equivariant activations; no free parameters, new axioms, or invented entities are introduced in the abstract.

axioms (2)

standard math Complex multiplication represents 2D rotations and basis transformations invariantly.
Standard algebraic property of the complex field used to express geometric features.
domain assumption Phase-equivariant activation functions preserve invariance under local frame changes.
Design assumption required for the layers to maintain the claimed global invariance.

pith-pipeline@v0.9.0 · 5456 in / 1318 out tokens · 55799 ms · 2026-05-13T20:52:57.478720+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel (J uniqueness via reciprocal symmetry) echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

Theorem 4.1 (SO(2) Equivariance of complex matrix multiplication): For any complex weight matrix W_C ... ρ_C(R) (W_C x) = W_C (ρ_C(R) x)
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking (D=3 rotational linking) echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

split phase-amplitude AF that applies tanh to the amplitude and fixes the phase

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 4 internal anchors

[1]

Translational and Rotational Invariance in Networked Dynamical Systems,

C.-I. Vasile, M. Schwager, and C. Belta, “Translational and Rotational Invariance in Networked Dynamical Systems,” IEEE Transactions on Control of Network Systems, vol. 5, no. 3, pp. 822–832, Sep. 2018, conference Name: IEEE Transactions on Control of Network Systems. [Online]. Available: https://ieeexplore.ieee.org/document/7805256/?arnumber=7805256

work page arXiv 2018
[2]

Coordination of groups of mobile autonomous agents using nearest neighbor rules,

A. Jadbabaie, Jie Lin, and A. Morse, “Coordination of groups of mobile autonomous agents using nearest neighbor rules,”IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003. [Online]. Available: http://ieeexplore.ieee.org/document/1205192/

work page arXiv 2003
[3]

Formation control using range-only measurements,

M. Cao, C. Yu, and B. D. Anderson, “Formation control using range-only measurements,”Automatica, vol. 47, no. 4, pp. 776–781, 2011. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0005109811000823

work page 2011
[4]

Winkelmann, Dynamical Simulation of Electron Backscatter Diffraction Patterns, in: A.J

K. Meyer, G. Hall, and D. Offin,Introduction to Hamiltonian Dynamical Systems and the N-Body Problem, ser. Applied Mathematical Sciences, J. Hale, P. Holmes, J. Keener, J. Keller, R. Laubenbacher, B. Matkowsky, A. Mielke, C. Peskin, K. Sreenivasan, A. Stevens, S. Antman, J. E. Marsden, and L. Sirovich, Eds. New York, NY: Springer New York, 2009, vol. 90. ...

work page doi:10.1007/978-0- 2009
[5]

How Powerful are Graph Neural Networks?

K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How Powerful are Graph Neural Networks?” Feb. 2019, arXiv:1810.00826 [cs]. [Online]. Available: http://arxiv.org/abs/1810.00826

work page internal anchor Pith review Pith/arXiv arXiv 2019
[6]

Learning Decentralized Controllers for Robot Swarms with Graph Neural Networks,

E. Tolstaya, F. Gama, J. Paulos, G. Pappas, V . Kumar, and A. Ribeiro, “Learning Decentralized Controllers for Robot Swarms with Graph Neural Networks,” inProceedings of the Conference on Robot Learning. PMLR, May 2020, pp. 671–682, iSSN: 2640-3498. [Online]. Available: https://proceedings.mlr.press/v100/tolstaya20a.html

work page 2020
[7]

Learning Decentralized Flocking Controllers with Spatio-Temporal Graph Neural Network,

S. Chen, Y . Sun, P. Li, L. Zhou, and C.-T. Lu, “Learning Decentralized Flocking Controllers with Spatio-Temporal Graph Neural Network,” in2024 IEEE International Conference on Robotics and Automation (ICRA), May 2024, pp. 2596–2602. [Online]. Available: https://ieeexplore.ieee.org/document/10610627/

work page arXiv 2024
[8]

Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

M. M. Bronstein, J. Bruna, T. Cohen, and P. Veli ˇckovi´c, “Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges,” May 2021, arXiv:2104.13478 [cs]. [Online]. Available: http://arxiv.org/abs/2104.13478

work page internal anchor Pith review Pith/arXiv arXiv 2021
[9]

Inductive Representation Learning on Large Graphs

W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive Representation Learning on Large Graphs,” Sep. 2018, arXiv:1706.02216 [cs]. [Online]. Available: http://arxiv.org/abs/1706.02216

work page Pith review arXiv 2018
[10]

Semi-Supervised Classification with Graph Convolutional Networks

T. N. Kipf and M. Welling, “Semi-Supervised Classification with Graph Convolutional Networks,” Feb. 2017, arXiv:1609.02907 [cs]. [Online]. Available: http://arxiv.org/abs/1609.02907

work page internal anchor Pith review Pith/arXiv arXiv 2017
[11]

Threat-Adaptive Cooperative Control for Heterogeneous UA V Swarms via Dynamic Graph Neural Networks,

L. Zhang, X. Feng, B. Yang, and Z. Lei, “Threat-Adaptive Cooperative Control for Heterogeneous UA V Swarms via Dynamic Graph Neural Networks,” inProceedings of the 2025 9th International Conference on Computer Science and Artificial Intelligence, ser. CSAI ’25. New York, NY , USA: Association for Computing Machinery, Mar. 2026, pp. 690–699. [Online]. Avai...

work page doi:10.1145/3788149.3788175 2025
[12]

Comprehensive Survey of Complex-Valued Neural Networks: Insights into Backpropagation and Activation Functions,

M. M. Hammad, “Comprehensive Survey of Complex-Valued Neural Networks: Insights into Backpropagation and Activation Functions,” 2024, version Number: 1. [Online]. Available: https://arxiv.org/abs/2407.19258

work page arXiv 2024
[13]

Robot control using high dimensional neural networks,

Y . Maeda, T. Fujiwara, and H. Ito, “Robot control using high dimensional neural networks,” in2014 Proceedings of the SICE Annual Conference (SICE), Sep. 2014, pp. 738–743. [Online]. Available: https://ieeexplore.ieee.org/document/6935220/

work page arXiv 2014
[14]

Continuous complex-valued back-propagation learning,

A. Hirose, “Continuous complex-valued back-propagation learning,”Electronics Letters, vol. 28, no. 20, pp. 1854–1855, Sep. 1992. [Online]. Available: http://digital- library.theiet.org/doi/10.1049/el%3A19921186

work page doi:10.1049/el 1992
[15]

J. E. Grabowski,Representation Theory: A Categorical Approach. Open Book Publishers, Sep. 2025. [Online]. Available: https://www.openbookpublishers.com/books/10.11647/obp.0492

work page doi:10.11647/obp.0492 2025
[16]

A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning,

S. Ross, G. Gordon, and D. Bagnell, “A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics. JMLR Workshop and Conference Proceedings, Jun. 2011, pp. 627–635, iSSN: 1938-7228. [Online]. Available: https://proceedings.mlr....

work page 2011
[17]

Stable Flocking of Mobile Agents Part II: Dynamic Topology,

H. Tanner, A. Jadbabaie, and G. Pappas, “Stable Flocking of Mobile Agents Part II: Dynamic Topology,” in42nd IEEE International Conference on Decision and Control, vol. 2, Dec. 2003, pp. 2016–2021 V ol.2, iSSN: 0191-2216. [Online]. Available: https://ieeexplore.ieee.org/document/1272911/

work page arXiv 2003
[18]

Adam: A Method for Stochastic Optimization

D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” Jan. 2017, arXiv:1412.6980 [cs]. [Online]. Available: http://arxiv.org/abs/1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2017