Problem

The evolution towards 6G necessitates increasingly high mobile data traffic, extreme spectral efficiency, and adaptability across diverse service scenarios. Existing 5G feedback-based Multiple-Input Multiple-Output (MIMO) transmission faces significant limitations, including substantial feedback overhead, increased latency, and susceptibility to pilot contamination, leading to performance degradation, particularly in high-mobility environments. Furthermore, the shift to higher frequency bands in 5G results in reduced coverage areas and higher power consumption. Current cellular architectures with coupled uplink (UL) and downlink (DL) transmissions further restrict base station (BS) coverage. While Fully-Decoupled Radio Access Networks (FD-RAN) are a promising architectural candidate for 6G due to their ability to physically separate UL and DL functionalities, this decoupling renders conventional feedback-based transmission methods infeasible. This creates a critical challenge in achieving feedback-free MIMO transmission and efficient multi-dimensional resource allocation in the expanded coverage scenarios of FD-RAN.

Method

The proposed framework, CaFTRA (Frequency-domain Correlation-aware Feedback-free MIMO Transmission and Resource Allocation), is specifically designed for FD-RAN. At the physical (PHY) layer, CaFTRA introduces a Learnable Queries-driven Transformer Network (LQTN) for Channel State Information (CSI) prediction. This network leverages user geolocation information (BS and UE locations) and employs multi-head attention with learnable query embeddings to accurately capture frequency-domain correlations among Resource Blocks (RBs). This enables the prediction of CSI parameters (Rank Indicator, Channel Quality Indicator, Precoding Matrix Indicator) for all RBs without requiring real-time uplink feedback from User Equipments (UEs). The feedback-free MIMO transmission process involves offline training of the LQTN using historical CSI data, online prediction of CSI from real-time user geolocation by the edge cloud, and subsequent feedback-free MIMO transmission by DL BSs using these predicted CSI parameters. At the Medium Access Control (MAC) layer, CaFTRA addresses resource scheduling challenges by proposing a low-complexity many-to-one matching theory-based algorithm, termed M3-MAMA. This algorithm efficiently handles multi-BS association and multi-RB allocation in the extensive-coverage scenarios of FD-RAN, ensuring each RB serves only one UE and each UE receives a minimum number of RBs (QoS). The M3-MAMA algorithm is designed to achieve pairwise stability and is proven to converge within a limited number of iterations, offering a computationally feasible solution compared to traditional exhaustive search methods.

Results

Simulations demonstrate that CaFTRA consistently achieves superior performance compared to existing 5G methods. In terms of CSI prediction, CaFTRA exhibits significantly lower Normalized Mean Absolute Error (MAE) across all CSI components (RI, CQI1, CQI2, PMI) compared to baseline independent Transformer networks, validating its frequency-domain correlation-aware design. For static user scenarios, CaFTRA achieves comparable throughput to 5G closed-loop spatial multiplexing (CLSM) while eliminating CSI feedback. Crucially, in high-mobility scenarios, CaFTRA dramatically outperforms 5G CLSM; at a user velocity of 10 km/h, CaFTRA shows a 20% increase in per-user throughput, and at 15 km/h, it surpasses 5G CLSM by 93%, showcasing its robustness against feedback delays. Furthermore, CaFTRA achieves substantial gains in spectral efficiency, showing a 60% improvement over 5G Round-Robin and approximately 15% over 5G Best CQI scheduling algorithms. It also ensures significantly higher user fairness, as indicated by Jain's Fairness Index, across various user densities and QoS levels. Both average per-user and per-RB throughput are higher with CaFTRA. The M3-MAMA matching algorithm is proven to converge to a stable solution within a polynomial time complexity, making it practical for real-world deployment. The LQTN model itself also significantly reduces space complexity by approximately 84% compared to independent Transformer networks.

Implications

The CaFTRA framework offers a compelling solution for the design of AI-native 6G and beyond wireless systems. By effectively eliminating real-time CSI feedback, CaFTRA significantly reduces spectral overhead and enhances transmission reliability, particularly in dynamic and high-mobility environments where traditional feedback-based systems struggle. The framework leverages the architectural advantages of FD-RAN, such as decoupled UL/DL functionalities and extended downlink coverage, to enable more flexible and efficient multi-BS cooperation and resource scheduling. This leads to substantial improvements in spectral efficiency, user fairness, and overall throughput. These findings underscore CaFTRA's potential as a foundational technology for future wireless communication networks, addressing critical challenges related to CSI feedback and resource allocation scalability, and contributing valuable insights toward 6G standardization efforts in areas like power control and spectrum sharing under feedback-free MIMO.