pith. sign in

arxiv: 2603.20489 · v2 · submitted 2026-03-20 · 📡 eess.SP

Realization of a Fully Connected Neural Layer Over-the-Air through Multi-hop Amplify-and-Forward Relays

Tolga Girici , Meng Hua , Deniz G\"und\"uz This is my paper

Pith reviewed 2026-05-15 07:44 UTC · model grok-4.3

classification 📡 eess.SP
keywords over-the-air computingamplify-and-forward relaysmulti-hop relayingneural network layerswireless optimizationfully connected layerphysical-layer computation
0
0 comments X

The pith

Multi-hop amplify-and-forward relays can realize a fully connected neural network layer over wireless channels with near-perfect classification accuracy.

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

The paper examines how to implement the matrix multiplication and nonlinearity of a fully connected neural layer directly through wireless transmission rather than digital computation. A transmitter sends signals through a chain of amplify-and-forward relays to a multi-antenna receiver, and the received vector is made to match the desired neural output by tuning the transmitter precoder, relay gains, and receiver combiner. An alternating optimization procedure is used to solve the joint design under transmitter and relay power limits. Simulation results show that this over-the-air realization reaches almost perfect accuracy on classification tasks when the number of hops is sufficient.

Core claim

By jointly optimizing the transmitter precoding matrix, the amplification factors at each relay, and the receiver combining matrix under per-node power constraints, the end-to-end wireless channel can be made to approximate the linear transformation and activation of a fully connected neural layer, yielding classification accuracy that approaches the ideal digital case.

What carries the argument

Alternating optimization of the transmitter precoder, relay amplification gains, and receiver combiner to minimize the imitation error between the over-the-air received vector and the target neural-layer output.

If this is right

  • Wireless networks can perform neural inference without digitizing intermediate results at each hop.
  • Multi-hop relaying relaxes the distance and power limits that single-hop over-the-air computation faces.
  • The same alternating-optimization procedure can be reused for any linear layer whose weights are known at design time.
  • Classification tasks whose decision boundaries are simple linear transforms become feasible at the physical layer.

Where Pith is reading between the lines

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

  • Cascading multiple such optimized layers could enable deeper analog neural networks distributed across a wireless mesh.
  • Channel estimation overhead becomes the practical bottleneck once perfect-coefficient assumptions are removed.
  • The approach naturally extends to tasks where the neural weights themselves are learned jointly with the wireless parameters.

Load-bearing premise

The design assumes perfect knowledge of every wireless channel coefficient between the transmitter, all relays, and the receiver.

What would settle it

Running the same optimization and simulation but replacing the perfect channel coefficients with noisy estimates and observing whether classification accuracy falls well below the reported near-perfect levels.

Figures

Figures reproduced from arXiv: 2603.20489 by Deniz G\"und\"uz, Meng Hua, Tolga Girici.

Figure 1
Figure 1. Figure 1: Multi-hop OTA computing system model Suppose that, upon receiving the signal from the BS, each device amplifies and forwards this signal to the second relay group. We assume that relay groups access the channel in a TDMA manner. Let x ∈ C N be the transmitted baseband complex signal vector. The symbols are uncorrelated, i.e. E[xxH] = I, and therefore, their transmissions do not interfere with each other. L… view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy vs. Number of Relay Devices per group ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Even smaller relay power: Accuracy vs. Number of Relay Devices per [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Direct link enabled: Accuracy vs. Number of Relay Devices per group [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

We study the problem of implementing a fully-connected layer of a neural network using wireless over-the-air computing. We assume a multi hop system with a multi-antenna transmitter and receiver, along with a number of multi-hop amplify-and-forward relay devices in between. We formulate an optimization problem that optimizes the transmitter precoder, receiver combiner and amplify-and-forward gains, subject to relay device power constraint and transmitter power constraint. We propose an alternating optimization framework that optimizes the imitation accuracy. Simulation study results reveal that multi-hop relaying achieves an almost perfect classification accuracy when used in a neural network.

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.

Referee Report

2 major / 1 minor

Summary. The paper studies the realization of a fully connected neural network layer via over-the-air computation in a multi-hop amplify-and-forward relay system with multi-antenna transmitter and receiver. It formulates a joint optimization problem over the transmitter precoder, relay amplification gains, and receiver combiner to minimize imitation error of the target neural weights subject to per-device power constraints, proposes an alternating optimization algorithm, and reports simulation results showing near-perfect classification accuracy.

Significance. If the reported simulation accuracy is robust, the work would add to over-the-air computing literature by showing that multi-hop relaying can closely emulate a neural layer under power limits. The contribution is limited by the absence of convergence analysis for the non-convex alternating optimization and by incomplete specification of the simulation setup (channel models, antenna counts, initialization, and exact metrics), which reduces the strength of the headline claim.

major comments (2)
  1. [Optimization Framework and Algorithm] The optimization problem is non-convex (joint design of precoder, AF gains, and combiner under power constraints), yet the alternating optimization framework is presented without any proof of monotonic improvement, convergence to a stationary point, or global optimality. This directly affects the reliability of the simulation claim of 'almost perfect classification accuracy,' as the reported performance may correspond only to favorable local solutions.
  2. [Simulation Results] The simulation study results are described only at a high level (multi-hop relaying achieves almost perfect accuracy). Key details required to assess the result—exact channel models, number of antennas at transmitter/receiver/relays, initialization strategy for the alternating optimizer, convergence behavior, and precise accuracy metric—are not provided, making it impossible to evaluate whether the headline performance holds beyond the specific runs shown.
minor comments (1)
  1. [Abstract and Problem Formulation] The abstract states that the framework 'optimizes the imitation accuracy' but does not clarify whether the objective is mean-squared error on the effective weights, classification error, or another metric; this notation should be defined explicitly in the problem formulation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We provide point-by-point responses to the major comments and indicate the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Optimization Framework and Algorithm] The optimization problem is non-convex (joint design of precoder, AF gains, and combiner under power constraints), yet the alternating optimization framework is presented without any proof of monotonic improvement, convergence to a stationary point, or global optimality. This directly affects the reliability of the simulation claim of 'almost perfect classification accuracy,' as the reported performance may correspond only to favorable local solutions.

    Authors: We acknowledge that the joint optimization problem is non-convex and that our alternating optimization algorithm lacks a formal proof of monotonic improvement or convergence to a stationary point. In the revised manuscript, we will add a discussion of the empirical convergence behavior, including plots demonstrating that the objective function decreases monotonically and stabilizes within a small number of iterations across multiple random initializations. We will also clarify that the reported near-perfect accuracy is obtained consistently over several runs and note the possibility of local optima as a limitation of the approach. revision: partial

  2. Referee: [Simulation Results] The simulation study results are described only at a high level (multi-hop relaying achieves almost perfect accuracy). Key details required to assess the result—exact channel models, number of antennas at transmitter/receiver/relays, initialization strategy for the alternating optimizer, convergence behavior, and precise accuracy metric—are not provided, making it impossible to evaluate whether the headline performance holds beyond the specific runs shown.

    Authors: We agree that additional details are required for reproducibility and evaluation. In the revised manuscript, we will expand the simulation section to specify the exact channel model (i.i.d. Rayleigh fading), antenna counts (transmitter: 8 antennas, receiver: 8 antennas, each relay: 2 antennas), initialization strategy (random Gaussian initialization with selection of the best result over 10 trials), convergence behavior (objective typically stabilizes in fewer than 50 iterations), and precise metrics (normalized mean squared error for weight imitation and classification accuracy). revision: yes

Circularity Check

0 steps flagged

No significant circularity; simulation results follow from direct optimization of imitation error

full rationale

The paper sets up an optimization problem whose explicit objective is to minimize the mismatch between the effective over-the-air mapping (precoder, AF gains, combiner) and the target fully-connected layer weights, subject only to the stated power constraints. The alternating optimization is applied to this objective, and the reported near-perfect classification accuracy is simply the empirical outcome of that optimization succeeding in the simulated channels. No step equates the final accuracy claim to a fitted parameter by construction, nor does any load-bearing premise reduce to a self-citation or ansatz imported from prior work by the same authors. The derivation chain is therefore self-contained against the system model and power limits; the performance result is a verifiable simulation outcome rather than a tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard wireless assumptions of perfect CSI and fixed power budgets; no new entities are postulated and the optimization variables are treated as design choices rather than fitted constants.

free parameters (2)
  • Number of relays and antennas
    Chosen for the simulation study to demonstrate performance; not derived from first principles.
  • Transmit and relay power limits
    Fixed constraints in the optimization problem; values selected for the reported experiments.
axioms (2)
  • domain assumption Perfect channel state information is available at the transmitter and receiver
    Required to formulate and solve the precoder and combiner optimization.
  • domain assumption Amplify-and-forward relays operate linearly without noise amplification modeling details
    Implicit in the multi-hop system model used for the optimization.

pith-pipeline@v0.9.0 · 5400 in / 1312 out tokens · 33993 ms · 2026-05-15T07:44:16.761556+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Pilot Allocation for Multi-Hop Over-the-Air Neural Inference under Imperfect CSI

    eess.SP 2026-04 unverdicted novelty 4.0

    Five heuristic pilot allocation schemes for multi-hop OTA neural inference trade training overhead for classification accuracy under CSI errors, approaching digital performance with sufficient pilots and balanced resources.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    A survey on over-the-air computation,

    A. S ¸ahin and R. Yang, “A survey on over-the-air computation,”IEEE Communications Surveys & Tutorials, vol. 25, no. 3, pp. 1877–1908, 2023

    work page 1908

  2. [2]

    Federated learning over wireless fading channels,

    M. M. Amiri and D. G ¨und¨uz, “Federated learning over wireless fading channels,”IEEE transactions on wireless communications, vol. 19, no. 5, pp. 3546–3557, 2020

    work page 2020

  3. [3]

    Airnn: Over- the-air computation for neural networks via reconfigurable intelligent surfaces,

    S. G. Sanchez, G. Reus-Muns, C. Bocanegra, Y . Li, U. Muncuk, Y . Naderi, Y . Wang, S. Ioannidis, and K. R. Chowdhury, “Airnn: Over- the-air computation for neural networks via reconfigurable intelligent surfaces,”IEEE/ACM Transactions on Networking, vol. 31, no. 6, pp. 2470–2482, 2022

    work page 2022

  4. [4]

    Airfc: Designing fully connected layers for neural networks with wireless signals,

    G. Reus-Muns, K. Alemdar, S. G. Sanchez, D. Roy, and K. R. Chowd- hury, “Airfc: Designing fully connected layers for neural networks with wireless signals,” inProceedings of the Twenty-F ourth International Symposium on Theory, Algorithmic F oundations, and Protocol Design for Mobile Networks and Mobile Computing, 2023, pp. 71–80

    work page 2023

  5. [5]

    Over-the-air split machine learning in wireless mimo networks,

    Y . Yang, Z. Zhang, Y . Tian, Z. Yang, C. Huang, C. Zhong, and K.-K. Wong, “Over-the-air split machine learning in wireless mimo networks,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 4, pp. 1007–1022, 2023

    work page 2023

  6. [6]

    Over-the-air inference through analog computation over multi-hop mimo networks,

    C. Bian, M. Hua, and D. G ¨und¨uz, “Over-the-air inference through analog computation over multi-hop mimo networks,”IEEE Wireless Communications Letters, 2025

    work page 2025

  7. [7]

    A radio-frequency-based 2- d convolutional layer using transmissive intelligent surfaces,

    J. Zhang, H. Chen, and D. M. Blough, “A radio-frequency-based 2- d convolutional layer using transmissive intelligent surfaces,” in2024 IEEE 100th V ehicular Technology Conference (VTC2024-Fall). IEEE, 2024, pp. 1–7

    work page 2024

  8. [8]

    Implementing neural net- works over-the-air via reconfigurable intelligent surfaces,

    M. Hua, C. Bian, H. Wu, and D. Gunduz, “Implementing neural net- works over-the-air via reconfigurable intelligent surfaces,”arXiv preprint arXiv:2508.01840, 2025

    work page arXiv 2025

  9. [9]

    Energy-efficient over-the-air computation for relay-assisted iot networks,

    J. Wan, J. Wen, K. Wang, Q. Wu, and W. Chen, “Energy-efficient over-the-air computation for relay-assisted iot networks,”IEEE Wireless Communications Letters, vol. 13, no. 2, pp. 481–485, 2023

    work page 2023

  10. [10]

    Node scheduling for af- based over-the-air computation,

    S. Tang, H. Yomo, C. Zhang, and S. Obana, “Node scheduling for af- based over-the-air computation,”IEEE Wireless Communications Let- ters, vol. 11, no. 9, pp. 1945–1949, 2022

    work page 1945

  11. [11]

    Amplify-and-forward relaying for hierarchical over-the-air computation,

    F. Wang, J. Xu, V . K. Lau, and S. Cui, “Amplify-and-forward relaying for hierarchical over-the-air computation,”IEEE Transactions on Wireless Communications, vol. 21, no. 12, pp. 10 529–10 543, 2022

    work page 2022

  12. [12]

    Relay-assisted cooperative federated learning,

    Z. Lin, H. Liu, and Y .-J. A. Zhang, “Relay-assisted cooperative federated learning,”IEEE Transactions on Wireless Communications, vol. 21, no. 9, pp. 7148–7164, 2022

    work page 2022

  13. [13]

    Joint optimization for over-the-air computation in af relay-assisted cognitive radio networks,

    J. Yao, M. Jin, T. Wu, Q. Li, and K.-K. Wong, “Joint optimization for over-the-air computation in af relay-assisted cognitive radio networks,” IEEE Transactions on V ehicular Technology, vol. 73, no. 10, pp. 15 809– 15 814, 2024

    work page 2024

  14. [14]

    Joint secure transceiver design for an untrusted mimo relay assisted over-the-air computation networks with perfect and imperfect csi,

    H. Luo, Q. Li, Q. Zhang, and J. Qin, “Joint secure transceiver design for an untrusted mimo relay assisted over-the-air computation networks with perfect and imperfect csi,”IEEE Transactions on Information F orensics and Security, vol. 18, pp. 2508–2523, 2023

    work page 2023