⚡ electrical engineering

Scalable Cross-Attention Transformer for Cooperative Multi-AP OFDM Uplink Reception

The paper proposes a scalable cross-attention Transformer that enables joint decoding of uplink OFDM signals across multiple coordinated access points by learning time-frequency structures and fusing receiver data without explicit channel estimates, achieving robust performance that rivals perfect channel state information baselines while remaining efficient on commodity hardware.

Original authors: Xavier Tardy, Grégoire Lefebvre, Apostolos Kountouris, Haïfa Fares, Amor Nafkha

Published 2026-04-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xavier Tardy, Grégoire Lefebvre, Apostolos Kountouris, Haïfa Fares, Amor Nafkha

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

The Big Picture: The "Group Chat" for Wi-Fi

Imagine you are trying to hear a friend (the User) who is shouting a message across a noisy, crowded room. The room has bad acoustics, echoes, and people talking over each other.

In a traditional Wi-Fi setup, you might have just one person (a single Access Point or "AP") trying to listen to your friend. If that person is standing behind a pillar or in a noisy corner, they might miss parts of the message or hear it garbled.

This paper proposes a smarter way: The Group Chat. Instead of relying on one listener, we have a team of three (or more) listeners standing in different corners of the room. They all hear the same shout, but from different angles and with different levels of background noise.

The goal of this research is to build a "super-brain" that listens to all these people simultaneously, combines their stories, and figures out exactly what the friend said—even if some listeners are in a bad spot.

The Problem: Why Current Wi-Fi Struggles

Currently, Wi-Fi systems handle this in two ways, both of which have flaws:

The "Solo Listener" (Classical Methods): Each listener tries to figure out the message alone, then they just average their answers.
- The Flaw: If one listener is deaf (bad signal), they drag the whole group down. Also, they don't share how they heard the message, just the final guess.
The "Super-Computer" (Old AI Models): Some new AI models try to look at all the data from all listeners at once to find patterns.
- The Flaw: This is like trying to organize a meeting where everyone talks to everyone else at the same time. As you add more listeners, the chaos (computational cost) grows exponentially. It becomes too slow and expensive to run on real hardware.

The Solution: The "Cross-Attention" Transformer

The authors built a new AI model called a Cross-Attention Transformer. Here is how it works, broken down into three simple steps:

1. The Individual Reporters (The Shared Encoder)

First, every listener (AP) writes a quick report about what they heard.

The Analogy: Imagine each listener is a journalist. They don't just write "I heard 'Hello'." They write a detailed note: "I heard a 'Hello', but it sounded a bit muffled, and there was a lot of static on the high notes."
The Tech: The AI uses a "Shared Encoder." This means all journalists use the same training manual. They learn to spot patterns (like echoes or noise) independently, but they all speak the same "language" so they can be compared later.

2. The Smart Editor (Token-Wise Cross-Attention)

This is the magic part. Instead of just averaging the reports, the AI acts like a Smart Editor sitting at a desk with all the reports.

The Analogy: The Editor looks at a specific word in the message (e.g., the word "Hello").
- Listener A says: "I heard 'Hello' clearly."
- Listener B says: "I heard 'H...llo' but it was very quiet."
- Listener C says: "I heard 'Hello' but it sounded like 'Gello'."
- The Decision: The Editor realizes Listener A is the most reliable for this specific word. They give Listener A's report a high weight (trust it 90%) and Listener B's a low weight (trust it 20%).
The Innovation: The Editor does this for every single word (or data point) in the message. If Listener A is good at the beginning of the sentence but bad at the end, the Editor switches to trusting Listener B for the end.
Why it's fast: The Editor doesn't make everyone talk to everyone. They just look at the specific word and pick the best sources. This keeps the process linear (fast) even if you add 10 or 20 listeners.

3. The Final Verdict (The Decoder)

Once the Editor has combined the best parts of all the reports, they produce a "confidence score" for every bit of the message. This score is sent to a standard error-checking system (like a spell-checker) to fix any remaining mistakes and output the final, perfect message.

Why This Paper Matters (The Results)

The researchers tested this system using realistic Wi-Fi scenarios (simulating a busy city environment). Here is what they found:

It's a Master at Teamwork: When they added more listeners (APs), the system got much better at hearing the message. It didn't just get slightly better; it got significantly better, outperforming old methods by a wide margin.
It Works with Few Clues: Wi-Fi usually sends "pilot signals" (like test tones) to help the receiver understand the channel. The new system works great even when these test tones are very sparse (like having only one clue instead of two). It's like a detective who can solve a crime with very little evidence.
It's Fast and Cheap: Unlike the "Super-Computer" AI models that get slow as you add more people, this new system stays fast. It can run on standard computer chips (CPUs) without needing expensive graphics cards.
It Beats "Perfect" Knowledge: In some cases, this AI performed better than a theoretical system that already knew exactly what the noise was. It learned to adapt so well that it didn't need that perfect knowledge.

The Bottom Line

This paper introduces a smart, scalable way for Wi-Fi networks to use multiple antennas together.

Think of it as upgrading from a single person shouting back and forth to a well-organized team of translators who listen to a speaker from different angles, compare notes in real-time, and produce a perfect transcript. This technology will be crucial for the next generation of Wi-Fi (Wi-Fi 8), making connections faster, more reliable, and capable of handling crowded environments like stadiums or airports.

1. Problem Statement

The paper addresses the challenge of cooperative uplink reception in next-generation OFDM systems (e.g., Wi-Fi 7/8, 5G/6G), where multiple Access Points (APs) coordinate to decode signals from a single User Equipment (UE).

Limitations of Classical Pipelines: Traditional receivers rely on a three-stage process: pilot-based channel estimation (e.g., LS, LMMSE), equalization, and demapping. These methods often fail in non-stationary environments due to:
- Sensitivity to noise and interference (LS).
- Requirement for accurate second-order channel statistics which are often unavailable (LMMSE).
- Independent processing at each AP, ignoring spatial correlations between APs.
- Inability to adaptively fuse signals based on the varying reliability of individual links.
Limitations of Existing Deep Learning Approaches: While neural receivers exist, many operate on flat-fading models rather than full 2D time-frequency grids. Furthermore, Transformer-based approaches using full self-attention incur quadratic complexity ( $O(N^2)$ ) relative to the number of time-frequency elements and APs, making them computationally prohibitive for large-scale, dense multi-AP deployments.

2. Methodology

The authors propose a Cross-Attention Transformer architecture designed specifically for joint decoding of multi-AP OFDM signals without requiring explicit Channel State Information (CSI).

System Model

Scenario: A single-antenna UE transmits to $N_R$ coordinated APs over an OFDM grid ( $N_c$ subcarriers $\times$ $N_s$ symbols).
Input: Raw received complex grids $\{Y^{(r)}\}$ from each AP and noise variance estimates.
Output: Soft Log-Likelihood Ratios (LLRs) for a standard channel decoder (LDPC).
Training Objective: The model is trained to maximize the Bit-Metric Decoding (BMD) rate, which serves as a differentiable surrogate for minimizing Bit Error Rate (BER).

Architecture Design

The proposed network consists of three distinct stages:

Per-AP Shared Encoder (Self-Attention):
- A shared Transformer encoder processes the time-frequency grid of each AP independently.
- Input tokens are formed from the real/imaginary parts of received symbols and noise variance, augmented with 2D positional encodings.
- Multi-Head Self-Attention (MHSA) captures local and global dependencies within the 2D grid of a single receiver.
- Key Benefit: Parameter sharing ensures the model scales linearly with the number of APs.
Token-Wise Cross-Attention Fusion:
- Instead of concatenating all AP data and applying full self-attention (which is $O(N_{AP}^2)$ ), the model performs fusion per time-frequency resource element (RE).
- For a specific grid position $(f, t)$ , the latent representations from all $N_R$ APs are treated as a sequence of tokens.
- Mechanism: One AP (e.g., AP 1) acts as an "anchor" to generate the Query ( $Q$ ), while all APs contribute to the Keys ( $K$ ) and Values ( $V$ ).
- The attention mechanism learns to dynamically weight the contributions of each AP based on reliability, effectively down-weighting noisy or faded links and up-weighting strong ones.
- Complexity: This reduces the fusion complexity to linear with respect to the number of APs ( $O(N_R)$ ).
Prediction Head:
- A lightweight Multi-Layer Perceptron (MLP) maps the fused representation of each RE to the bit-level LLRs required by the channel decoder.

3. Key Contributions

Scalable Architecture: A novel Transformer design that achieves linear complexity with respect to the number of APs, overcoming the quadratic bottleneck of full self-attention models.
Adaptive Fusion: A token-wise cross-attention mechanism that learns to fuse multi-view observations based on per-receiver reliability, adapting to heterogeneous link qualities and noise levels without explicit CSI.
End-to-End Soft Output: The model outputs soft LLRs directly, acting as a drop-in replacement for the entire classical chain (estimation + equalization + demapping) while remaining compatible with standard LDPC decoders.
Robustness: The system demonstrates high resilience to sparse pilot patterns and strong frequency selectivity.

4. Performance Results

The model was evaluated using 3GPP TR 38.901 Urban Microcell (UMi) channels with varying numbers of APs ( $N_R = 1, 2, 3$ ) and pilot densities.

BER Performance:
- Cooperation Gain: Increasing $N_R$ from 1 to 3 provided significant spatial diversity gains. For a target BER of $10^{-6}$ , the proposed model reduced the required $E_b/N_0$ by approximately 9.5 dB when moving from 1 to 3 APs.
- Comparison: The model consistently outperformed classical LS, LMMSE, and CNN-based baselines.
- Perfect-CSI Benchmark: In multi-AP scenarios ( $N_R=3$ ), the model approached or even surpassed the performance of a local Perfect-CSI reference with SNR-based fusion, indicating it effectively exploits inter-AP correlations beyond simple weighting.
- Pilot Sparsity: The model remained highly robust when pilot density was reduced (from 2 columns to 1), outperforming CNNs by ~3.5 dB at $10^{-6}$ BER in sparse conditions.
Computational Efficiency:
- Complexity: The proposed model has 0.15M parameters and scales linearly with $N_R$ .
- Latency: On commodity CPU hardware, the proposed model was 3x faster than a full self-attention Transformer baseline at $N_R=10$ (2120ms vs. 7700ms).
- Comparison: It offers a better trade-off between performance and latency compared to CNNs (which have high FLOPs) and full self-attention Transformers (which scale poorly).

5. Significance and Impact

This work bridges the gap between theoretical cooperative reception gains and practical implementation constraints in distributed Wi-Fi and cellular networks.

Feasibility: By reducing complexity to linear scaling, the architecture makes real-time, joint multi-AP decoding feasible on edge hardware (CPUs) without requiring massive fronthaul bandwidth or expensive GPUs.
Next-Gen Wi-Fi: The solution is particularly relevant for Wi-Fi 8 (802.11bn) and cell-free architectures, where dense AP coordination is essential for coverage and interference mitigation.
Paradigm Shift: It demonstrates that deep learning can replace the entire physical layer receiver chain (estimation to demapping) with a single, robust, and adaptive neural module that learns to handle channel uncertainty and spatial diversity implicitly.