🔢 mathematics

Joint Channel Estimation and Dynamics-Aware Grouping for Time-Varying RIS-Assisted OTA Federated Learning

This paper proposes a unified framework for RIS-assisted over-the-air federated learning that integrates GRU-based temporal channel estimation, dynamics-aware user grouping, and physical-layer optimization to mitigate the performance degradation caused by time-varying channels, imperfect CSI, and user heterogeneity.

Original authors: Ziqi Li, Shuangzhi Li, Uchechukwu Awada, Jiankang Zhang

Published 2026-07-21

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Ziqi Li, Shuangzhi Li, Uchechukwu Awada, Jiankang Zhang

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 a world where your smartphone, your smartwatch, and your neighbor's tablet all work together to teach a super-smart AI without ever showing each other their private photos or messages. This is the promise of "Federated Learning," a way for devices to learn collectively while keeping data safe. But there's a catch: to learn together, these devices need to talk to a central tower. In a perfect world, they would send their lessons one by one, but that takes forever. So, scientists invented "Over-the-Air" learning, where everyone shouts their lesson at the same time, and the tower listens to the combined noise to figure out the answer. It's like a choir where everyone sings a different note, but the conductor hears the perfect chord instantly.

However, real life is messy. The airwaves are full of interference, and people are always moving, causing the "channel" (the invisible path the signal travels) to wobble and change shape constantly. If the tower doesn't know exactly how the signal is wobbling, the choir sounds like a disaster, and the AI learns the wrong things. To fix this, engineers use "Reconfigurable Intelligent Surfaces" (RIS)—giant, smart mirrors on walls that can bounce signals around obstacles and steer them perfectly to the tower. But here's the tricky part: these mirrors need to know exactly where everyone is and how they are moving to work, and figuring that out in a moving, noisy world is incredibly hard.

This paper tackles that exact headache. The researchers propose a clever new system that acts like a team of detectives, a choir director, and a traffic controller all rolled into one. They built a framework that uses a special type of artificial intelligence called a Gated Recurrent Unit (GRU) to predict how the wireless signals will move and change over time, almost like a weather forecaster predicting a storm before it hits. Instead of trying to force every single device to sing in perfect unison (which is impossible when some are far away and some are zooming by), they group the devices based on how "dynamic" their connection is. Devices with stable connections form one choir, while those with wobbly, fast-changing connections form another. By letting these groups practice separately before joining the big concert, the system ensures that the final lesson the AI learns is accurate, even when the wind is blowing and the devices are racing down the street. The authors tested this in computer simulations with 10 users and found that their method significantly reduced errors and helped the AI learn faster, especially when there wasn't much time to send "practice" signals or when users were moving very quickly.

The Big Idea: A Smart Choir in a Windy City

Imagine you are trying to teach a robot to recognize cats by having 10 different friends send you photos. In a perfect world, you'd just ask them to send the photos one by one. But in this paper's world, the friends are in a busy city with tall buildings and strong winds, and they have to shout their photos to you all at once. If they shout at the same time, their voices mix up. If the wind blows one friend's voice away, you miss their photo. If a friend is running fast, their voice sounds different (like a siren passing by).

The researchers in this paper are trying to solve the problem of how to get a clear message when everyone is shouting at once in a chaotic, moving environment. They use three main tricks:

1. The Time-Traveling Detective (GRU)
The first problem is that the "wind" (the wireless channel) changes every second. To fix this, the team uses a tool called a GRU (Gated Recurrent Unit). Think of this as a detective who doesn't just look at the current weather report but remembers the last few days of weather to predict what the wind will do next. Instead of just guessing the signal's path right now, the GRU looks at the history of how the signal has moved to predict where it will be a split-second later. This helps the central tower (the base station) know exactly how to tune its ears to hear the friends clearly, even if they are moving.

2. The Smart Mirror (RIS)
To help the signals get through the tall buildings, the system uses a Reconfigurable Intelligent Surface (RIS). Imagine a giant wall covered in thousands of tiny, adjustable mirrors. If a friend is behind a building, the tower can tell the mirror to tilt just right to bounce the signal around the corner. But the mirror needs to know exactly where the friend is and how fast they are moving to aim correctly. The paper's system uses the "Time-Traveling Detective" to tell the mirror exactly how to tilt, ensuring the signal bounces perfectly to the tower.

3. The Grouping Strategy (Dynamics-Aware Grouping)
Here is the most creative part. In a normal choir, everyone tries to sing the same song at the same volume. But in this wireless world, some friends are standing still in a quiet park (stable connection), while others are sprinting down a highway (unstable, fast-changing connection). If they all try to sing together, the sprinters mess up the quiet singers.

The authors realized that instead of forcing everyone to sing together, they should group the singers.

Group A: The "Steady" singers (people with stable, slow-moving connections).
Group B: The "Wobbly" singers (people moving fast or far away).

The system first lets everyone try to sing together to get started quickly. But once the system realizes who is who, it splits them into two separate groups. Each group practices its own song with its own set of instructions for the smart mirror. This prevents the fast-moving friends from drowning out the slow ones. The paper shows that this "grouping" strategy makes the final result much more accurate, especially when there are "tail users"—those difficult friends who are far away or moving very fast and usually ruin the signal for everyone else.

What They Found (and What They Didn't)

The researchers ran many computer simulations to test their idea. They set up a virtual world with a base station, a smart mirror with 64 tiny elements, and 10 users. They tested different scenarios: some where everyone was walking slowly, and others where some users were zooming around at high speeds.

The Results:

Better Prediction: Their "Time-Traveling Detective" (the GRU) was much better at guessing the signal path than older methods. In simulations, it reduced the "noise" (errors) in the signal significantly, even when they didn't have much time to send "practice" signals.
Stronger Grouping: When they used the grouping strategy, the system handled the "wobbly" users much better. The stable users got clearer signals, and the fast-moving users didn't get kicked out of the system; they just got their own group to practice in.
Robustness: The system worked well even when the "wind" was very strong (high mobility) or when the "practice" signals were very short (low pilot length).

What They Didn't Claim:
It is important to note that these results come from computer simulations, not a real-world test with actual people running around with phones. The authors are very clear that they simulated the physics of the signals and the movement of the users. They did not claim to have built a physical prototype or tested this on a live 6G network. They also didn't claim that their method is perfect in every single possible situation, but rather that it is a significant improvement over current methods in the specific, chaotic scenarios they simulated.

Why This Matters

You might wonder, "Why do we care about a choir in a windy city?" The answer is the future of our connected world. As we move toward 6G networks and "edge intelligence" (where your phone and your car learn together), we will have millions of devices moving at high speeds. If we can't figure out how to get them to talk to each other clearly without wasting battery or time, our smart cities and autonomous cars won't work well.

This paper suggests that by combining smart prediction (the GRU), smart mirrors (RIS), and smart grouping, we can make these systems much more reliable. It's a step toward a future where your devices can learn together, no matter how fast you are driving or how many buildings are in the way, without you ever having to worry about the signal dropping. The authors show that by treating the wireless channel as a dynamic, changing thing that needs to be predicted and grouped, rather than a static road, we can build a much smarter, more resilient internet for the future.

Technical Summary: Joint Channel Estimation and Dynamics-Aware Grouping for Time-Varying RIS-Assisted OTA Federated Learning

Problem Statement
Over-the-air federated learning (OTA-FL) offers communication efficiency by exploiting the superposition property of wireless channels for simultaneous model aggregation. However, in time-varying wireless environments, OTA-FL suffers from severe performance degradation due to imperfect Channel State Information (CSI), time-varying channel dynamics, and strong user heterogeneity. These factors induce aggregation distortion and model update cancellation. While Reconfigurable Intelligent Surfaces (RIS) can reshape propagation environments to improve link quality, RIS-assisted OTA-FL introduces tightly coupled challenges: cascaded channel estimation, pilot overhead limitations, and the dynamic nature of user mobility. Existing approaches often decouple channel estimation, RIS configuration, and federated optimization, or rely on device selection strategies that exclude "weak" users, potentially compromising data diversity and fairness. Furthermore, standard personalized FL methods often assume static or ideal channels, failing to address the coupling between time-varying channel dynamics and federated optimization.

Methodology
The paper proposes a unified, closed-loop framework that jointly optimizes channel estimation, RIS phase control, and federated aggregation under imperfect CSI. The system consists of a multi-antenna Base Station (BS), an RIS, and $K$ single-antenna edge users. The framework integrates three core modules:

GRU-Based Cascaded Channel Estimation:
To address time-varying channels, the authors employ a Gated Recurrent Unit (GRU) with a Conv1D backbone for temporal modeling. The estimator decomposes the channel into a slowly varying large-scale path-loss component and a fast-varying small-scale fading component. This structure prevents the entanglement of spatial geometry variations with Doppler-driven temporal fluctuations. The model is trained offline using Linear MMSE (LMMSE) reference labels and operates online to predict instantaneous CSI for over-the-air aggregation and RIS optimization.
Personalized Federated Aggregation (Reptile-based):
The learning architecture adopts a shared-backbone design where a global backbone ( $\phi$ ) captures transferable temporal representations, and user-specific heads ( $\psi_k$ ) absorb local statistical heterogeneity. The system uses a Reptile-style meta-update for the shared backbone, which is compatible with the communication-constrained OTA aggregation scenario where full gradient exchange is infeasible. Only the backbone updates are transmitted via AirComp.
Dynamics-Aware User Grouping:
To mitigate aggregation cancellation caused by heterogeneity, a dynamic grouping strategy is introduced. This strategy utilizes "dynamics proxies" derived from the estimated path-loss and short-term channel dynamics (mobility).
- Mechanism: Users are initially aggregated globally. Once the population-level dynamic proxies stabilize (indicating the system has learned stable representations), the system switches to a grouped mode.
- Optimization: A robust grouping objective minimizes intra-group risk while maximizing inter-group separation, incorporating RIS-induced correlation signatures. This is solved via Successive Convex Approximation (SCA).
- Philosophy: Unlike device selection methods that drop weak users, this approach retains them in a "high-risk" group, ensuring full participation and fairness while allowing the "low-risk" group to converge faster.

Key Contributions

Unified Closed-Loop Framework: The paper proposes a system-level design that jointly optimizes channel estimation, RIS phase control, and federated aggregation, explicitly incorporating OTA-induced distortion into the learning process.
Temporal Modeling for CSI: A personalized federated learning architecture is designed using GRUs as a temporal backbone, enabling accurate prediction of time-varying cascaded CSI while supporting heterogeneous local adaptation.
Dynamics-Aware Grouping: A novel grouping strategy is developed that exploits long-term path-loss and short-term channel dynamics to improve aggregation stability under strong user heterogeneity without excluding weak users.
Theoretical Analysis: The paper provides convergence analysis for the Reptile-based shared backbone under OTA distortion and establishes the stationary-point convergence of the SCA-based grouping solver.

Results
Numerical simulations were conducted in RIS-assisted uplink scenarios with varying mobility settings (mild, tail-user, and high mobility) and user distributions (homogeneous and near/far heterogeneous).

CSI Estimation: The proposed GRU-based estimator consistently achieved lower Normalized Mean Square Error (NMSE) compared to baselines (CNN-based estimators and LMMSE), particularly in low-pilot regimes and high-mobility scenarios.
Aggregation Performance: The framework demonstrated substantial gains in OTA aggregation performance (lower aggregation NMSE) and reduced proxy distortion.
Grouping Efficacy: The dynamic grouping strategy effectively separated users with different channel dynamics. Low-risk users achieved lower error floors, while high-risk users maintained meaningful partial convergence rather than being discarded.
Robustness: The system showed consistent improvements in both mild and highly heterogeneous scenarios, with significant gains observed when pilot overhead was limited.

Significance and Claims
The authors claim that the proposed framework bridges the critical gap between communication reliability and learning stability in RIS-assisted OTA-FL systems. By treating channel estimation, physical-layer optimization, and federated learning as a coupled problem rather than separate sequential tasks, the system achieves reliable learning under imperfect CSI and heterogeneous dynamics. The paper emphasizes that the joint optimization of temporal channel learning, personalized federated optimization, and dynamics-aware grouping is essential for distributed intelligence in dynamic wireless environments. The results suggest that retaining weak users through grouping, rather than excluding them, preserves data diversity and fairness while maintaining system robustness.

The Big Idea: A Smart Choir in a Windy City

What They Found (and What They Didn't)

Why This Matters

Technical Summary: Joint Channel Estimation and Dynamics-Aware Grouping for Time-Varying RIS-Assisted OTA Federated Learning

More like this