Optimal Stochastic Coordinated Beamforming with Compressive CSI Acquisition for Cloud-RAN

Yuanming Shi, Jun Zhang, Khaled Ben Letaief

Abstract

This paper introduces a generic framework for coordinated beamforming in wireless networks that accounts for the inevitable uncertainty in channel state information (CSI). By reformulating a highly intractable joint chance-constrained problem into a Difference-of-Convex (DC) program, we provide a stochastic algorithm that guarantees optimality—a significant leap over previous heuristic or overly conservative methods. Simulations demonstrate that the proposed approach achieves superior power efficiency and stability compared to traditional scenario-based methods.

Introduction

In modern wireless architectures like Cloud-RAN, multiple base stations or Radio Access Units (RAUs) coordinate to beamform signals to users, maximizing efficiency and minimizing interference. This coordination requires exact Channel State Information (CSI). However, in practical deployments, CSI is imperfect due to estimation errors, feedback delays, and hardware deficiencies.

Researchers typically address this using "chance constraints" to ensure Quality of Service (QoS) is met with high probability (e.g., 90% of the time). Mathematically, this leads to a Joint Chance Constrained Program (JCCP). JCCPs are notoriously difficult to solve because probability constraints often lack closed-form expressions and are frequently non-convex. Existing solutions, such as the "scenario approach" or "Bernstein approximation," are often overly conservative, leading to wasted transmit power.

Proposed Methodology

The authors propose a Stochastic DC (Difference-of-Convex) Programming algorithm to solve the JCCP. The approach consists of three primary components:

  1. DC Reformulation: The intractable probability constraint is rewritten as a DC constraint (the difference between two convex functions). This transforms the JCCP into a manageable, though still non-convex, DC program.
  2. Joint Approximation: An approximation parameter ($\kappa$) is introduced to handle non-smooth indicator functions. Unlike previous methods that keep $\kappa$ fixed, this framework jointly optimizes $\kappa$ alongside beamforming vectors to accelerate convergence.
  3. Sample Average Approximation (SAA): To ensure practicality, Monte Carlo sampling (SAA) is used to estimate gradients and functions at each iteration. This makes the framework "distribution-blind," requiring only samples of channel uncertainty rather than a specific mathematical model.

Experimental Evaluation and Results

The algorithm was evaluated using Monte Carlo simulations in a Cloud-RAN setup with 5 single-antenna RAUs and 3 single-antenna users. The model incorporated path-loss, shadowing, and small-scale fading, with a mixed CSI model where only 40% of channel coefficients were obtained noisily.

Power Efficiency and Constraint Tightness

The proposed algorithm achieved an average total transmit power of 10.52 dBm, significantly lower than the 11.10 dBm required by the scenario approach. A key reason for this efficiency is the "tightness" of the constraints. While the scenario approach is overly conservative, the proposed algorithm hits the target probability constraint (0.9) exactly, avoiding unnecessary power expenditure.

Figure 1: Comparison of probability constraint tightness and stability between the proposed Stochastic DC Programming and the Scenario Approach.

Stability and Convergence

A major finding was the stability of the Stochastic DC Programming algorithm. The variance in the optimal value across 50 random replications was significantly smaller ($\Delta_1 = 0.1190$ dBm) compared to the scenario approach ($\Delta_2 = 0.9566$ dBm). Furthermore, jointly optimizing the approximation parameter $\kappa$ allowed the algorithm to converge in roughly half the iterations compared to fixed-parameter methods.

Figure 2: Total transmit power across 50 replications, highlighting the lower power requirements and higher stability of the proposed method.

Conclusion

The proposed Stochastic DC Programming framework provides a robust and efficient solution for coordinated beamforming under CSI uncertainty. It is the first to guarantee local optimality for non-convex JCCPs and global optimality for convex ones. While the computational cost of large-scale SAA is a current limitation, future work will focus on more efficient solvers like ADMM to scale the approach for very large networks. Additionally, extending the framework to handle "ambiguous" distributions through distributionally robust methods remains a promising research direction.

Summary updated

BibTeX
@article{shi2015optimal,
  author = {Yuanming Shi and Jun Zhang and Khaled Ben Letaief},
  title = {{Optimal Stochastic Coordinated Beamforming with Compressive CSI Acquisition for Cloud-RAN}},
  journal = {IEEE Trans. Signal Process.},
  year = {2015},
  doi = {10.1109/tsp.2014.2385669},
}
Figure