AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems

arXiv:2607.06768 · cs.IT, cs.AI, cs.LG, math.IT · Submitted 2026-07-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems".

Jane: The gist: AirPASS investigates over-the-air federated learning in wireless systems equipped with a multi-waveguide pinching antenna system to jointly optimize device selection, receive beamforming,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're wrapping up our look at "AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems." The authors used this alternating optimization framework with the homotopy-Riemannian margin-consolidation method and the homotopy-assisted geometry optimization method.

Jane: If you think about it, the title itself suggests a major shift because they are tackling that highly nonconvex joint design problem by using these two specific optimization tools.

Lu: The authors show that this approach allows them to increase the number of admissible devices under the aggregation constraint, particularly at low SNR.

Meng: For someone actually running this on a device, what does that mean practically? It means you can potentially use more devices in your learning process without hurting the accuracy of your model too much.

Tom: And they translate that gain into improved learning performance over conventional co-located MIMO baselines, which is something we saw in the simulation results.

Jane: So the implication for wireless systems is that this technique provides a way to leverage physical reconfigurability of antennas to boost communication efficiency in federated learning settings.

Lu: The paper points out that their HRMC block does both the cardinality-oriented search and the margin-consolidation step directly in the original beamforming space, without needing lifted matrix variables.

Meng: That's important because it simplifies things for production deployment because you don't need to manage extra variables that might make training harder to scale up.

Tom: The overall worst-case complexity is O IAO + one)(I1 + I2)KM squared plus IAOI3SMN, which is the complexity they have derived for this method <ref:2607.06768#pg1>.

Jane: It seems like a very concrete contribution to the field of communication-efficient federated learning, showing how physical hardware flexibility can be directly integrated into the learning algorithm design.

Conclusion: Tom: So, we've been looking at "AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems." The main idea here is using this multi-waveguide pinching antenna system to figure out which devices to use and how to steer the signals.

Jane: Right, so it’s about taking a learning objective and coupling it with physical hardware placement—the antenna positions—to get a better outcome.

Lu: The authors tackle that joint optimization problem directly. They set up this alternating framework where they optimize the device selection first, then the antenna geometry second, and repeat.

Meng: It sounds complex to set up in a real deployment because you have to manage two distinct optimization steps that change each other's inputs.

Tom: That’s what they did. They developed two main blocks for this alternating process—one for picking the devices and beamforming, and another for moving the pinching antennas.

Jane: And the results show that this approach consistently beats standard MIMO setups on both datasets we tested.

Lu: What’s interesting is that they use a homotopy method to make sure that when they optimize those antenna positions, it stays feasible and doesn't break anything.

Meng: But what about the cost of running this? Is this just theoretical optimization or something that can actually scale up on a real edge server?

Tom: The complexity analysis shows the worst-case complexity is O IAO plus one times two to k times M squared, which is manageable if you exploit some structure in the signal.

Jane: So, for someone who just wants to know what this means for their app, it means we can potentially use more devices in our learning process without losing too much accuracy.

Lu: And that gain is particularly visible when the signal strength is low, which is a tough condition in real-world wireless communication.

Tom: It’s not just about being more accurate; they say it translates that gain into better learning performance over co-located MIMO baselines.

Jane: So, the paper suggests that having multiple antennas on different waveguides gives us a whole new way to think about how we manage communication efficiency in federated learning.

Lu: It points toward using physical reconfigurability not just for sensing, but as a direct variable in the learning algorithm itself.

Meng: I'm still focused on whether this setup is robust enough for production traffic; it’s one thing to optimize an experiment and another thing entirely when you have real-time data streaming in.

Chalmers University of Technology · Purdue University

cs.IT, cs.AI, cs.LG, math.IT

Submitted: 2026-07-07

Updated: 2026-10-14

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: The gist: AirPASS investigates over-the-air federated learning in wireless systems equipped with a multi-waveguide pinching antenna system to jointly optimize device selection, receive beamforming,

Key concepts

Over-the-Air Federated Learning (AirFL): This is a learning process where devices in a wireless system update a central access point without direct, continuous communication. The goal is to train a shared model using data from multiple devices scattered across the network, relying on over-the-air signal transmission for updates.
Pinching Antenna System (PASS): This is the physical hardware at the access point consisting of multiple dielectric waveguides, each housing several movable pinching antennas. These antennas introduce additional spatial degrees of freedom that can be optimized to improve signal reception and performance in wireless communication systems.
Alternating Optimization Framework: AirPASS uses a two-step iterative process. In one step, it optimizes device selection and receive beamforming assuming the antenna positions are fixed. In the next step, it optimizes the antenna positions assuming the selected devices and beamformer are fixed, repeating until an optimal solution is found.
Homotopy-Riemannian Margin-Consolidation (HRMC): This is a specific technique used to solve device selection and beamforming problems when the antenna configuration is fixed. It replaces a hard counting objective with a smooth surrogate problem, allowing for differentiable optimization while preserving the structure needed to select devices based on their performance margin.

Terminology

AirFL — 2× in this paper · explained in 1 episode
Federated Learning — 2× in this paper · explained in 17 episodes
Aggregation — 1× in this paper · explained in 1 episode

Summary

The gist: AirPASS investigates over-the-air federated learning in wireless systems equipped with a multi-waveguide pinching antenna system to jointly optimize device selection, receive beamforming, and antenna placement under a learning-oriented objective.

System Model

The system considers a wireless FL setup involving an AP equipped with an MIMO PASS architecture consisting of multiple dielectric waveguides, each containing N movable pinching antennas The coordinate of the nth pinching antenna on waveguide m is given by vm,n = [lm,n,(m − 1)d, a]T The effective channel vector between device k and the AP is defined as gk(L) = [gk,1(l1),..., gk,M(lM)]T ∈ C M×1 The AirComp aggregation distortion is measured by the mean-square error (MSE) MSE = E[sˆ− s2].

Optimization Problem Formulation

The joint design problem is formulated as max S, bk⌋,m,η,L S (18a) s.t. MSE ≤ ε, (18b) bk2 ≤ P0, ∀k ∈ S, (18c) 0 ≤ lm,1, ∀m The joint device selection, receive beamforming, and pinching-antenna placement problem can be reformulated as max S,m,L S (25a) s.t. ϕ2k∥m∥2mHgk(L)2 ≤ Γ, ∀k ∈ S.

Device Selection and Receive Beamforming for Fixed PASS Configuration

For a fixed PASS configuration L, the problem is reformulated as max m X K k=1 1/mHQk(L)m ≥ 1 s.t.∥m∥2 = 1. This leads to the development of a homotopy-Riemannian margin-consolidation (HRMC) algorithm for device selection and receive beamforming under fixed PASS configuration. The method uses a smooth surrogate problem Fβ,µ(m) = X K k=1 wk ρµ(˜vk(m)) (35a) s.t.∥m∥2 = 1. This surrogate replaces the discontinuous counting objective by a differentiable function while still preserving the selection structure of the original problem.

PASS Antenna Configuration Optimization

For fixed selected devices and beamformer, the optimization of pinching antenna positions is referred to as homotopy-assisted geometry optimization (HAGO). This block optimizes a smooth worst-device objective fτ (L) = −τ logX k∈S exp −hk(L)/τϕ2k, where hk(L) = mHgk(L)2. The algorithm uses a reparameterization that guarantees feasibility by construction.

AirPASS Alternating Optimization Framework

The proposed AirPASS method alternates between two blocks: (i) device selection and receive beamforming for fixed PASS configuration, and (ii) PASS configuration optimization for fixed selected devices and receive beamformer. This alternating procedure is summarized in Algorithm 3.

Complexity Analysis

The overall worst-case complexity is O IAO (I1 + I2)KM2 + I3SMN + (I1 + I2)KM2, where IAO denotes the number of outer alternating iterations. If the rank-one structure of Qk(L) is exploited, the corresponding bound becomes O((I1 + I2)KM), which yields O((I1 + I2)KM).

Experimental Results

Simulation results showed that AirPASS consistently outperforms both conventional MIMO baselines on the two datasets. The gain over the baseline with only M fixed antennas is particularly pronounced, which highlights the importance of the additional spatial degrees of freedom introduced by the multiple pinching antennas on each waveguide. AirPASS remains close to ideal FedAvg and approaches the performance of SDR-DC while retaining a substantially more favorable complexity profile.

Conclusion

AirPASS increases the number of admissible devices under the aggregation constraint, particularly at low SNR, and translates this gain into improved learning performance over conventional co-located MIMO baselines. The proposed HRMC block performs both the cardinality-oriented search and the margin-consolidation step directly in the original beamforming space, without lifted matrix variables. The HAGO framework addresses these challenges through coupled first-order geometry optimization with a smooth bottleneck objective and a feasible reparameterization. The overall worst-case complexity is O(IAO + 1)(I1 + I2)KM2 + IAOI3SMN.

References

[1] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. Arcas, Communication-efficient learning of deep networks from decentralized data, AISTATS, pp 1273–1282

[2] S. M. Azimi-Abarghouyi, C. Fischione, and K. Huang, Over-the-air federated learning: Rethinking edge AI through signal processing, IEEE Signal Process. Mag., preprint available on arXiv: 2512.03719

[3] M. Goldenbaum, H. Boche, and S. Stanczak, Harnessing interference for analog function computation in wireless sensor networks, IEEE Trans. Signal Proc., vol 61, no 20, pp 4893–4906

[4] S. M. Azimi-Abarghouyi, Out-of-air computation: Enabling structured extraction from wireless superposition, preprint available on arXiv: 2604.04312

[5] M. Mohammadi Amiri and D. Gündüz, Federated learning over wireless fading channels, IEEE Trans. Wireless Commun., vol 19, no 5, pp 3546–3557

[6] K. Yang, T. Jiang, Y. Shi, and Z. Ding, Federated learning via overthe-air computation, IEEE Trans. Wireless Commun., vol 19, no 3, pp 2022–2035

[7] M. Kim, A. L. Swindlehurst, and D. Park, Beamforming vector design and device selection in over-the-air federated learning, IEEE Trans. Wireless Commun., vol 22, no 11, pp 7464–7477

[8] A. Bereyhi, A. Vagollari, S. Asaad, R. R. Muller, W. Gerstacker, and H. V Poor, Device scheduling in over-the-air federated learning via matching pursuit, IEEE Trans Signal Process., vol 71, pp 2188–2203

[9] S. M. Azimi-Abarghouyi and L. Tassiulas, Over-the-air federated learning via weighted aggregation, IEEE Trans. Wireless Commun., vol 23, no 12, pp 18240–18253

[10] M. Krouka, A. Elgabli, C. B. Issaid, and M. Bennis, Communicationefficient federated learning: A second order newton-type method with analog over-the-air aggregation, IEEE Trans Green Commun Netw., vol 6, no 3, pp 1862–1874

[11] S.

Improvements for AI systems

Bold header: Integrated Device Selection and Beamforming via AirPASS

This system can jointly optimize device selection, receive beamforming, and pinching-antenna placement by formulating a problem that jointly couples device selection, receive beamforming, and pinching antenna placement in a multi-waveguide PASS receiver. This allows the AP to actively shape the received signal superposition to substantially improv[e] aggregation accuracy.

Bold header: Robust Device Admission via Homotopy-Reweighted Optimization

The HRMC algorithm can be used for device selection and receive beamforming, which utilizes a homotopy-reweighted Riemannian optimization to solve the problem on the complex unit sphere while avoiding lifted variables required by SDR-DC-type methods. This provides a method to select devices that maximizes participation while ensuring constraints are met.

Bold header: Adaptive Antenna Configuration for Bottleneck Improvement

The HAGO method can be employed to optimize pinching antenna positions under fixed selected devices and beamformer, using a homotopy-assisted geometry optimization with a feasible reparameterization that enforces antenna ordering and minimum spacing by construction. This ensures the configuration maximizes the worst-device margin, addressing the bottleneck-driven configuration design rather than a conventional communication-centric placement problem.

Bold header: Enhanced Communication Efficiency under Low SNR

The AirPASS framework can be deployed to admit more devices under the aggregation constraint, particularly at low SNR, and translate this gain into improved learning performance over conventional baselines, demonstrating effectiveness in the low-to-moderate SNR regime.

Bold header: Complexity-Performance Tradeoff Management

The system offers a substantially more favorable complexity profile than lifted optimization methods by performing updates directly in the original beamforming space, achieving a better balance between performance and computational cost compared to SDR-DC and matching-pursuit scheduling alternatives.