Hybrid Downlink Beamforming with Outage Constraints under Imperfect CSI using Model-Driven Deep Learning

Technical Summary: Hybrid Downlink Beamforming with Outage Constraints under Imperfect CSI using Model-Driven Deep Learning

Problem Statement
The paper addresses the challenge of energy-efficient multi-user hybrid downlink beamforming (BF) and power allocation in the presence of imperfect Channel State Information (CSI). In practical systems, particularly those operating in millimeter-wave or future 6G bands, CSI degradation arises from estimation errors, feedback quantization, and time-varying fading. The objective is to minimize the total transmit power while satisfying probabilistic Quality of Service (QoS) constraints (specifically, outage probability constraints) for multiple users.

Classical optimization approaches for this problem typically rely on computationally expensive conic or semidefinite programming, often requiring knowledge of the CSI error distribution which is rarely available in practice. Conversely, generic deep learning (DL) architectures lack interpretability, require massive training datasets, and struggle to generalize across varying network configurations (e.g., different numbers of users or CSI qualities) without retraining. Furthermore, standard DL methods often fail to effectively handle non-convex or probabilistic QoS constraints.

Methodology
The authors propose a lightweight, model-aided deep learning (DL) architecture that integrates domain knowledge with learnable components. The core methodology involves the following elements:

1. Model-Driven Architecture: The proposed network, denoted as U-G-HBF-GCN, is based on a greedy selection algorithm for analog beam codewords. It unrolls the iterative process of selecting analog beams and solving the digital beamforming subproblem.
2. Instance-Adaptive Error Modeling: Instead of assuming a fixed error distribution, the architecture employs a Graph Convolutional Network (GCN) to estimate the impact of CSI errors. The GCN takes instance-specific features (such as channel estimates, QoS requirements, and pilot power) and outputs adaptive coefficients that modify the virtual channel covariance matrices. This allows the model to implicitly learn the error distribution from data without prior statistical knowledge.
3. Implicit Differentiation for Gradients: A significant technical hurdle is computing gradients for the nested optimization problems (digital BF and power allocation) which lack closed-form solutions. The authors derive a novel, efficient implicit representation of the underlying uplink BF problem. By formulating the Karush-Kuhn-Tucker (KKT) conditions as a system of nonlinear equations and applying the implicit function theorem, they prove sufficient conditions for the existence of the gradient. This approach avoids the computational cost of differentiating through eigendecompositions or semidefinite relaxations.
4. Annealing-Based Loss Function: To handle probabilistic outage constraints, the paper introduces an adaptive annealing-based approximation in the loss function. Unlike fixed quantile-based penalties, this method uses an annealing coefficient that adapts based on the current level of QoS violation. This allows the model to learn effectively during early training stages (when constraints are violated) and converge to the exact probabilistic constraint as training progresses.

Key Contributions

• Novel Architecture: The paper presents a model-aided DL framework for hybrid BF under outage constraints that does not predict power allocation directly but relies on an adaptive mapping of CSI using GCNs. This avoids the need for second-order cone or semidefinite programming during inference.
• Generalization: The proposed method demonstrates the ability to generalize to varying numbers of users, different QoS requirements, and fluctuating CSI qualities using a single trained model. It supports both instantaneous and statistical CSI.
• Theoretical Gradient Derivation: The authors provide a rigorous derivation of the gradient for the constrained uplink/downlink BF problem, proving sufficient conditions for its existence via implicit differentiation, thereby enabling end-to-end training.
• Adaptive Loss Function: The introduction of an annealing-based approach for integrating probabilistic constraints is shown to accelerate training and yield a better power-outage trade-off compared to conventional quantile-based methods.

Experimental Results
Simulations were conducted using a 4x4 and 8x8 antenna array with varying numbers of users (1 to 4) and CSI degradation levels (simulated via noisy MMSE estimation and DFT feedback quantization).

• Performance: The proposed U-G-HBF-GCN significantly outperforms benchmark greedy algorithms (G-HBF-marg) that use imperfect CSI with margin adjustments. It achieves the nominal outage level (e.g., 10%) while allocating less power than benchmarks.
• Comparison to Benchmarks: The model achieves performance close to fully digital beamforming (U-FDBF-GCN) when the number of RF chains is sufficient, and it consistently outperforms fully connected network variants (U-G-HBF-FCN), highlighting the efficacy of the GCN in capturing user interference correlations.
• Generalization: The model successfully adapts to unseen numbers of users (extrapolation) and varying CSI qualities without retraining.
• Training Efficiency: The adaptive annealing loss function resulted in faster convergence (fewer training steps) and a smoother loss landscape compared to quantile-based approaches.

Significance and Claims
The paper claims that its primary significance lies in bridging the gap between the computational efficiency of deep learning and the robustness required for practical wireless systems with imperfect CSI. By moving away from black-box architectures and rigid optimization assumptions, the proposed method offers a scalable solution for energy-efficient hybrid beamforming. The authors emphasize that their approach does not require knowledge of the analytical CSI error distribution, making it suitable for real-world deployment where such distributions are unknown. Additionally, the work provides a theoretical foundation for differentiating through constrained optimization problems in wireless communications, enabling the training of complex, unrolled architectures with probabilistic constraints. The paper concludes that with sufficient RF chains, hybrid beamforming can attain performance comparable to fully digital systems while maintaining lower hardware complexity and energy consumption.