⚡ electrical engineering

Budget-Aware Federated Dual-Side Channel Estimation for Hybrid mmWave Massive MIMO

This paper proposes BARRNet, a budget-aware neural network architecture that combines a compact residual backbone with lightweight channel-wise recalibration to achieve superior normalized mean squared error and communication efficiency tradeoffs for federated dual-side channel estimation in hybrid mmWave massive MIMO systems under the standard FedAvg framework.

Original authors: Jiawei Chen, Ruining Fan, Mouli Chakraborty, Avishek Nag, Anshu Mukherjee

Published 2026-08-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiawei Chen, Ruining Fan, Mouli Chakraborty, Avishek Nag, Anshu Mukherjee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to send a secret message to a friend, but you are both hiding in a crowded, noisy stadium. You can't shout the whole message at once because the noise would drown it out, and you can't whisper it perfectly because your friend is too far away. This is the daily struggle of modern wireless networks, specifically the super-fast 5G and future 6G systems that use "millimeter-wave" signals. These signals are like high-pitched whistles that carry huge amounts of data but are easily blocked by walls or rain. To make them work, engineers use massive antennas (like giant spiderwebs of wires) and special "hybrid" tricks to squeeze the signal through.

However, there's a catch: to send data correctly, the phone and the tower need to know exactly what the "noise" of the room sounds like. This is called Channel State Information (CSI). Usually, the tower asks every phone to send back a detailed report of the noise. But if thousands of phones do this, the network gets clogged, and privacy becomes a nightmare because everyone is sharing their raw data. Enter "Federated Learning." Think of this as a group project where students (the phones) solve a puzzle on their own and only send their ideas (mathematical models) to the teacher (the tower), not their messy notebooks. The teacher combines these ideas to get a better answer. But here's the problem: if the students' ideas are too big and complex, sending them back and forth takes forever and uses up all the bandwidth. The big question is: How do we design a student's "idea" so it's small enough to send quickly but smart enough to solve the puzzle perfectly?

This paper, titled "Budget-Aware Federated Dual-Side Channel Estimation for Hybrid mmWave Massive MIMO," tackles that exact problem. The authors, a team of researchers from universities in Ireland and China, didn't try to invent a new way to send the messages (the "federated learning" part). Instead, they asked a different question: "If we have a strict limit on how much 'idea space' we can send, how should we build the brain inside that idea?"

They propose a new design called BARRNet (Budget-Aware Recalibrated Refinement Network). To understand how it works, imagine you are trying to clean up a blurry, noisy photo. You have a limited amount of "paint" (computer parameters) to fix it. Most people would just buy a bigger, heavier paintbrush and try to paint over everything uniformly. The authors argue that this is wasteful. Instead, BARRNet uses a small, efficient base brush (a "compact residual backbone") to do the heavy lifting, but adds a tiny, super-smart "adjustment knob" (a channel-wise recalibration module). This knob doesn't paint new details; it just tweaks the colors and brightness of the existing paint to make it pop. It's like having a master chef who uses a standard knife but adds a pinch of a secret spice at the exact right moment to transform a simple dish into a gourmet meal.

The researchers tested this idea in a simulated world of wireless signals. They set up a scenario with a base station and several users, mimicking the real-world challenges of hybrid beamforming. They compared their "secret spice" design (BARRNet) against two other approaches: a standard, heavy-duty neural network (like a giant, clumsy paintbrush) and a "backbone-only" version of their own design (the standard knife without the spice).

The results were clear. In their simulations, BARRNet didn't just work; it was incredibly efficient. When aiming for a specific level of clarity (a target error rate of -13 dB), BARRNet needed 28.9% less communication than the standard "backbone-only" version. To put that in perspective, if the other methods had to send a 100-page report to get the job done, BARRNet only needed to send a 71-page report to get the exact same result. Even compared to the much larger, heavier neural networks, BARRNet was faster and used less data.

Crucially, the paper rules out the idea that "bigger is always better." They showed that simply making the model larger (adding more parameters) didn't help as much as using the same amount of "budget" more wisely. The gain didn't come from having more paint; it came from knowing exactly where to apply the little bit of extra paint they had. The authors emphasize that this is a simulation result, meaning it was proven in a computer model of the wireless world, not yet in a physical lab with real antennas. However, the evidence suggests that for future wireless networks, the key to speed isn't just compressing data, but designing smarter, more budget-conscious brains that know how to make the most of every single bit they send.

Technical Summary: Budget-Aware Federated Dual-Side Channel Estimation for Hybrid mmWave Massive MIMO

1. Problem Statement

The paper addresses the challenge of channel state information (CSI) estimation in hybrid millimeter-wave (mmWave) massive MIMO systems. While hybrid beamforming reduces hardware costs compared to fully digital architectures, it compresses observations into a lower-dimensional RF domain, making accurate CSI recovery difficult due to noise and compression artifacts.

The specific context is Federated Learning (FL), where channel data is distributed across user devices (UEs) to preserve privacy and reduce uplink data transmission. However, standard FL involves repeated exchange of model parameters between clients and the base station (BS). In communication-constrained environments, the overhead of these repeated exchanges is dominated by the number of trainable parameters in the estimator.

Existing communication-efficient FL research primarily focuses on optimizing the aggregation algorithm (e.g., quantization, compression of updates) or reducing the number of rounds. This paper identifies a gap: the estimator-level design under a fixed parameter budget. The core problem is formulated as: How should a limited parameter budget be allocated within a neural network estimator to maximize CSI estimation accuracy (NMSE) while minimizing the cumulative communication cost required to reach a target accuracy under repeated model exchange?

2. Methodology

System Model and Observation Construction

The authors consider a multi-user mmWave system with dual-side hybrid beamforming (digital baseband + analog RF).

Downlink (DL): Structured observations are constructed by projecting received pilots through conjugate DFT bases, resulting in a noisy estimate of the channel matrix $H_k^H$ .
Uplink (UL): Structured observations are formed via pilot correlation and spatial back-projection using equivalent analog combiners, yielding a noisy estimate of $H_k$ .
Unified Input: Both DL and UL observations are converted into real-valued tensors (stacking real and imaginary parts) to allow a single neural architecture to refine both channel directions.

Problem Formulation

The task is framed as a federated supervised learning problem using standard Federated Averaging (FedAvg). The objective is to minimize the global loss (weighted sum of local Mean Squared Errors) subject to a total communication budget constraint ( $B_{max}$ ).

The communication cost per round is modeled as proportional to the number of transmitted parameters ( $P_{model}$ ) and the bitwidth.
The study does not propose a new optimization algorithm but fixes FedAvg to isolate the impact of estimator architecture on the NMSE–communication tradeoff.

Proposed Solution: BARRNet

The authors propose the Budget-Aware Recalibrated Refinement Network (BARRNet). The design philosophy is that under repeated parameter exchange, a small number of parameters should be allocated to selective feature recalibration rather than uniformly scaling a deep backbone.

Architecture Components:

Compact Encoder ( $E_\phi$ ): Extracts latent features from the structured noisy observation.
Channel-wise Recalibration Module ( $G_\psi$ ): Based on the Squeeze-and-Excitation principle. It applies global average pooling to generate channel descriptors, followed by a lightweight Multi-Layer Perceptron (MLP) to generate gating weights. These weights re-scale the encoder features ( $F' = g \odot F$ ).
Residual Refinement Head ( $D_\omega$ ): A compact network that predicts the structured disturbance (noise/error) component $\hat{N}$ .
Reconstruction: The final channel estimate is $\hat{H} = X - \beta \hat{N}$ , where $\beta$ is a learnable scaling factor.

Control Baselines:
To validate the design hypothesis, the authors compare BARRNet against:

Backbone-only: An architecture-matched control where the recalibration module is removed, and the remaining parameters are distributed uniformly in the backbone.
Heavy CNN: A deeper and wider CNN baseline.
DnCNN: A standard denoising CNN baseline.

3. Key Contributions

Structured Observation Pipeline: The construction of unified structured downlink and uplink observations from compressed hybrid measurements, enabling a single learning-based refinement pipeline for dual-side estimation.
Formulation of Budget-Allocation Problem: Framing communication-constrained federated CSI estimation as an estimator-level parameter-budget allocation problem. The work shifts focus from reducing update transmission costs to designing the estimator itself to be efficient under repeated exchange.
BARRNet Architecture: The proposal of a lightweight network that combines a compact residual backbone with a channel-wise recalibration module.
Empirical Evidence of Allocation Efficiency: Demonstration that selective recalibration utilizes a limited parameter budget more effectively than uniform backbone scaling. The paper provides architecture-matched evidence that BARRNet achieves better performance not merely by adding parameters, but by allocating them strategically.

4. Simulation Results

Experiments were conducted in a hybrid mmWave massive MIMO system ( $N_{BS}=64$ , $N_{RF}=16$ , $K=4$ users) using standard FedAvg.

NMSE–Communication Tradeoff: BARRNet achieves the best tradeoff among federated baselines. At 5 dB SNR, to reach a target DL NMSE of -13 dB:
- BARRNet requires 1.93 Gbits of cumulative communication.
- The Backbone-only control requires 2.72 Gbits.
- This represents a 28.9% reduction in communication cost for the same accuracy target compared to the architecture-matched control.
- Compared to the Heavy CNN (which has ~2.5x more parameters), BARRNet requires significantly fewer bits (1.93 Gbits vs. 7.45 Gbits) to reach the same target.
Performance Gains: BARRNet improves the best DL NMSE from -13.89 dB (Backbone-only) to -14.43 dB with only a 2.7% increase in parameter count.
Robustness: The performance advantage is maintained across different user populations ( $K=8$ ), client participation ratios (50%), and SNR regimes (0–10 dB).
Ablation Study: Variants with different parameter allocations (Encoder-heavy, Head-heavy, Recalib-heavy) performed worse than the Balanced allocation used in BARRNet, confirming that simply increasing parameters in one component is less effective than the proposed balanced, recalibration-focused design.
Quantization: Applying 8-bit uplink quantization to BARRNet further reduced communication costs while preserving nearly the same NMSE performance, showing compatibility with classical compression techniques.

5. Significance and Claims

The paper claims that communication-efficient federated CSI estimation relies not only on model compactness or update compression but critically on how the limited model capacity is utilized under the constraint of repeated model exchange.

Design Philosophy: The authors argue that in FL settings where models are exchanged repeatedly, every parameter is a communication cost. Therefore, the design goal should be to allocate a fixed budget to the components that yield the highest performance gain per parameter.
Recalibration vs. Scaling: The results suggest that a lightweight channel-wise recalibration module is a more efficient use of a limited parameter budget than simply scaling up the depth or width of the backbone network.
Scope: The work explicitly states it does not propose a new federated optimization algorithm. Instead, it focuses on the estimator design under standard FedAvg to demonstrate that architectural choices can significantly reduce the communication burden required to achieve a specific estimation accuracy.

The paper concludes that future studies on communication-efficient FL-based CSI estimation should consider budget-aware estimator design alongside communication compression techniques.