Abstract
We consider the multi-cell joint power control and scheduling problem in cellular wireless networks as a weighted sum-rate maximization problem. This formulation is very general and applies to a wide range of applications and QoS requirements. The problem is inherently hard due to objective's non-convexity and the knapsack-like constraints. Moreover, practical system requires a distributed operation. We applied an existing algorithm proposed by Scutari et al. in distributed optimization literature to our problem. The algorithm performs local optimization followed by consensus update repeatedly. However, it is not fully applicable to our problem, as it requires all decision variables to be maintained at every base station (BS), which is impractical for large-scale networks; also, it relies on the Lipschitz continuity of the objective function's gradient, which does not hold here. We exploited the nature of our objective function, and proposed a localized version of the algorithm. Furthermore, we relaxed the requirements of Lipschitz continuity with the proximal approximation. Convergence to local optimal solutions was proved under some conditions. Future work includes proving the above results from a stochastic approximation perspective, and investigating non-linear consensus schemes to speed up the convergence.
Introduction
Distributed optimization serves as the backbone for modern large-scale infrastructure, ranging from massive AI model training to the management of 5G wireless networks. In these environments, multiple "agents"—such as base stations or sensors—must collaborate to solve a global optimization problem using only local data and neighbor-to-neighbor communication. A prominent framework for this is NEXT (In-network nonconvex optimization), which utilizes local convex approximations combined with consensus steps to reach a solution.
However, standard implementations of NEXT face two significant "scalability killers." First, they typically require Global Knowledge, where every node must store and update the entire vector of decision variables for the whole network. In a large-scale cellular network, it is impractical for a single base station to track the power levels of every other tower in the system. Second, they rely on Smoothness Assumptions, specifically that the gradients of objective functions are Lipschitz continuous. In wireless resource allocation, gradients can often become unbounded, leading to numerical instability or failure in standard algorithms.
Methodology: Localization and Proximal Approximations
To address these challenges, the authors propose two primary generalizations to the NEXT framework: Localization and Proximal Approximations.
Exploiting Sparsity through Localization
The authors observe that most physical systems have a "sparse" dependency structure. By partitioning decision variables into blocks, they developed a scheme where nodes only maintain and communicate variables that directly affect them or their immediate neighbors. This reduces memory requirements and communication bandwidth significantly.
As shown in Figure 1, the communication graph (a) defines the physical links, while the local dependency sets (b and c) define which nodes share information about specific variables. By aligning the algorithm's communication pattern with these physical dependencies, the system achieves high efficiency without sacrificing global convergence.
Handling Non-Smoothness with Proximal Approximations
To manage non-smooth or unbounded gradients, the researchers introduced a sequence of smooth, Lipschitz-continuous functions—such as the Lasry-Lions double envelope—that gradually converge to the original objective. This "Proximal NEXT" approach acts as a mathematical guide, starting with a very smooth version of a difficult problem and slowly increasing complexity. This prevents the oscillations and "jumps" that typically plague gradient-descent methods in non-convex territory.
These innovations are synthesized into Algorithm 1: Localized Proximal Inexact NEXT, which alternates between a local optimization step using the smooth approximation and a consensus step restricted to local variables.
Evaluation and Key Results
The proposed framework was evaluated using both synthetic mathematical examples and a high-fidelity 19-Cell Wrap-Around Model, a standard industry benchmark for cellular network simulations.
The results demonstrate significant improvements over traditional methods:
- 4x Throughput Boost: In wireless simulations, the localized methods achieved approximately a four-fold average throughput gain compared to traditional single-cell heuristic methods.
- Communication Efficiency: The localized version reached optimal solutions while communicating only a fraction of the total network state, proving its scalability.
- Numerical Robustness: In scenarios where objective functions had unbounded gradients, the Proximal NEXT algorithm converged smoothly, whereas standard NEXT failed.
- Faster Convergence: The localized "LXLP-PL" variant showed significantly faster convergence than globalized versions in multi-cell scenarios.
Conclusion and Future Work
This work provides a bridge between theoretical distributed optimization and practical wireless network implementation. By leveraging the natural sparsity of network dependencies and employing proximal smoothing, the authors have created a robust, scalable framework for resource allocation.
Future research will focus on addressing integer constraints in scheduling through formal rounding schemes, exploring non-linear consensus schemes to further accelerate convergence, and investigating the theoretical challenges of unbounded gradients at the boundaries of constraint sets.
Summary updated