JPALB: Joint Power Allocation & AP Load Balancing

• JPALB is a resource optimization method that jointly addresses discrete AP activation and continuous power allocation to minimize total network energy consumption while meeting strict QoS targets.
• It employs advanced algorithmic frameworks such as MISOCP, group-sparsity relaxation, and successive convex approximation to solve its complex, non-convex mixed-integer formulations.
• Empirical results indicate that JPALB significantly reduces power consumption and AP load variance, offering scalable solutions in cell-free, heterogeneous, and integrated sensing-communication systems.

Joint Power Allocation and AP Load Balancing (JPALB) algorithms form a fundamental class of resource optimization methodologies for wireless networks that jointly address the discrete problem of AP (access point) activation/deactivation and the continuous allocation of transmit powers. The primary goal is to minimize total network power consumption—including transmit, hardware, and sometimes fronthaul or computation-associated costs—while satisfying stringent user-centric QoS constraints, such as SINR/SE/latency targets, fronthaul capacity, and fairness/load balancing. These algorithms have been realized in cell-free massive MIMO, heterogeneous networks, cooperative LTE, cognitive radio, and integrated sensing-communication systems, and are mathematically formulated as highly structured, often non-convex, mixed-integer optimization problems. The following sections provide an in-depth account of system models, mathematical problem formulations, algorithmic approaches, complexity and trade-offs, and empirical performance of state-of-the-art JPALB methods.

1. System Models and Key Variables

JPALB arises in diverse multi-AP environments, including cell-free Massive MIMO (Chien et al., 2019, Chien et al., 2020, Abbas et al., 19 Aug 2025), heterogenous networks (Yang et al., 2017), edge computing (Chen et al., 2020), and ISAC systems (Singh et al., 23 Jan 2026). The unified structure involves:

• Multiple distributed APs: $M$ APs, potentially each with $N$ antennas, serve $K$ single-antenna users.
• AP on/off variables: Binary indicators $s_m \in \{0,1\}$ (or $z_m$, $\alpha_k$), encoding AP activation.
• Continuous power variables: $\rho_{mk} \ge 0$ for downlink transmission from AP $m$ to user $k$, or per-user transmit powers in other frameworks.
• QoS variables and constraints: SINR/Spectral Efficiency (SE) constraints ($\mathrm{SINR}_k \ge \nu_k$ or $R_k \ge \xi_k$), per-AP total power budgets ($\sum_k \rho_{mk} \le P_{\max,m} s_m$), fronthaul, and possibly heterogeneous user priorities or task-latency targets.
• Augmentations: ISAC settings introduce additional sensing SINR constraints and fronthaul traffic-dependent power expenditures (Singh et al., 23 Jan 2026); edge scenarios further couple to offloading, computing, and harvesting variables (Chen et al., 2020).

Channel models are typically block-fading, with per-user large-scale coefficients $\beta_{m k}$; TDD operation, MRT or ZF precoding, and MMSE channel estimation are common in the massive MIMO context.

2. Mathematical Formulation and Problem Structure

The canonical JPALB problem in cell-free MIMO is formulated as

$$\min_{s_m \in \{0,1\}, \rho_{mk} \ge 0} \sum_{m=1}^M \left[ \Delta \sum_{k=1}^K \rho_{mk} + P_{\text{act},m} s_m \right]$$

subject to

$$\mathrm{SINR}_k(\{\rho_{mk}\}, \{s_m\}) \ge \nu_k, \quad \sum_{k=1}^K \rho_{mk} \le P_{\max} s_m, \quad \forall m,k$$

where $\Delta \ge 1$ is the PA inefficiency, $P_{\text{act},m}$ the static power per AP, and all terms above track active transmit and hardware power (Chien et al., 2019, Chien et al., 2020).

These problems are inherently non-convex mixed-integer programs (MIPs), due to discrete AP activation variables and SINR coupling. Prominent variants reformulate the objective with group-sparsity surrogates on per-AP power vectors to induce sparse AP selection, effectively relaxing the integer constraint (Chien et al., 2019, Chien et al., 2020). In integrated sensing-and-communication and edge-computing setups, additional constraints enforce sensing performance and computation-capacity (Singh et al., 23 Jan 2026, Chen et al., 2020).

In generalized HetNet JPALB, resource allocation couples to user association indicators $x_{ij}$, fractional scheduling $y_{ij}$, BS load $d_i = \sum_j y_{ij}$, and power $p_i$, with nonlinear interference-coupled SINR expressions:

$$\eta_{ij} = \frac{p_i g_{ij}}{\sum_{k \neq i} d_k p_k g_{kj} + \sigma^2}$$

and corresponding log-utility maximization objectives (Yang et al., 2017).

3. Solution Algorithms and Theoretical Approaches

Several algorithmic paradigms are employed to resolve the coupled discrete-continuous optimization:

3.1 Mixed-Integer Second-Order Cone Programming (MISOCP)

• Reformulates all constraints (including non-convex SINR, per-AP power, and objective) into SOC form, introduces binary activations as integer variables, and solves via branch-and-bound. This yields global optima but incurs exponential worst-case complexity in $M$ and is tractable only for moderate-scale systems (Chien et al., 2019, Chien et al., 2020).
• Example: For $M=20$, $K=20$, typical total power is reduced by $\approx$49% compared to a baseline with all APs active (Chien et al., 2019, Chien et al., 2020).

3.2 Group-Sparsity Relaxation and IRLS

• Relaxes AP binary indicators to a group-sparsity regularizer on $\rho_m = [\sqrt{\rho_{m1}},...,\sqrt{\rho_{mK}}]^T$.
• Utilizes an iteratively reweighted $\ell_2$ minimization (IRLS) under SOC constraints, converging rapidly (5–10 iterations) to stationary points, and followed by discrete AP subset selection from ranked per-AP power norms (Chien et al., 2019, Chien et al., 2020).
• Yields near-optimal performance (within 17–27% of the global optimum) at polynomial complexity $O(MK^{3.5})$.

3.3 Successive Convex Approximation and Difference-of-Convex Methods

• In ISAC (Singh et al., 23 Jan 2026), the mixed-integer program is relaxed via difference-of-convex penalties and successive linearization for binary indicators, transforming the problem into a sequence of convex SOCPs over (relaxed) AP activity and power. This approach enables efficient solution for high-dimensional settings with joint communication and sensing constraints.

3.4 Fixed-Point and Decomposition Methods

• Monotone and strictly subhomogeneous (MSS) fixed-point iterations decompose the utility-balancing JPALB into cluster-based BS assignment/power steps and per-BS antenna tilt/power updates. Global convergence is ensured via the Nuzman–Yates theorem (Liao et al., 2016).
• Alternating optimization of user association, BS load distribution, and power with convexification steps, using dual variables or exponential variable transforms, is prevalent in HetNet JPALB (Yang et al., 2017, You et al., 2016).

3.5 Accelerated Projected Gradient and Nonconvex Relaxation

• Large-scale cell-free MIMO with multicast-unicast support employs an accelerated projected gradient (APG) method after surrogate relaxation of binary association, yielding sublinear convergence to stationary points and allowing practical scalability (Abbas et al., 19 Aug 2025).

3.6 Distributed Best-Response and Potential Games

• In cognitive and multi-channel environments, JPALB can be cast as a noncooperative game with discrete-continuous actions, solved via distributed, convergent best-response updates exploiting the potential game structure (Hong et al., 2011).

4. Complexity, Scalability, and Trade-offs

Method	Complexity Scaling	Optimality	Typical Use/Scale
MISOCP (branch & bound)	$O(2^M \cdot (MK)^{a})$	Global optimal	Moderate M, K
Group-sparsity/IRLS	$O(N_{\text{iter}}\,MK^{3.5})$	Stationary/local	Large $M,K$
APG (relaxed penalty)	$O(N(U+K_M)^2)$ per iter	Stationary (non-convex)	Large $N,U,K_M$
MSS fixed-point	$O(\text{ITER}\,[C N + N|\Theta| + CK])$	Global (MSS)	Cellular SON
Distributed BR (potential)	$O(N M K)$ per round	Nash eq. (potential game)	Cognitive networks
Successive convexification (ISAC)	Per iter $O((KU\!+\!K)^3)$	Stationary (non-convex)	ISAC/edge systems

MISOCP ensures benchmarking accuracy, but group-sparsity, APG, and fixed-point methods offer practical scaling and over 40% total power reduction versus naïve approaches (Chien et al., 2019, Chien et al., 2020, Abbas et al., 19 Aug 2025). Trade-offs are dictated by the network size, required optimality, and real-time application constraints.

5. Empirical Performance and Insights

Empirical results across representative JPALB applications demonstrate:

• Total Power Reduction: In cell-free MIMO ($M=20$, $K=20$), global JPALB delivers $\sim$49–50% total power savings over all-AP-on baselines ($\sim$51–52 W vs. 102 W), with group-sparsity achieving within 17–27% of the optimum (Chien et al., 2019, Chien et al., 2020). ISAC extensions maintain $~$33% reduction for joint communication and sensing QoS (Singh et al., 23 Jan 2026).
• AP/Site Sparsity: At optimality or near-optimality, typically half or more APs can be deactivated while respecting all user constraints (Chien et al., 2019, Chien et al., 2020, Singh et al., 23 Jan 2026).
• Performance–Complexity Trade-off: Group-sparsity, APG, and IRLS heuristics practically match global benchmarks at a tiny fraction of computational cost. APG outperforms SCA-based benchmarks by 10–60$\times$ in runtime, with negligible loss in spectral efficiency (Abbas et al., 19 Aug 2025).
• Load Balancing and Fairness: Effective JPALB drastically reduces AP load variance (up to 50%), improves worst-case user SINR by up to 4 dB, and achieves near-optimal sum-utility compared to reference heuristics (Liao et al., 2016, Yang et al., 2017).
• Robustness and Flexibility: JPALB maintains all SINR, SE, sensing, and per-AP power constraints under dynamic topologies and supports a variety of precoding choices (MRT, ZF, F-ZF).

6. Extensions and Thematic Variants

JPALB methodologies extend to:

• Joint task offloading, cooling, and power control for MEC: Incorporates WPT, edge/local computation, and cooling-aware models. Alternating semi-closed-form updates achieve up to 90% energy savings over fixed/computation-only offloading strategies (Chen et al., 2020).
• Cognitive Radio and Distributed Networks: Potential-game based JPALB (JASPA) ensures distributed Nash equilibrium with rapid convergence, load-aware user self-association, and near-centralized throughput (Hong et al., 2011).
• HetNet User Prioritization: Accommodates user weights, proportional fairness, and exploits exponential transformations to address non-convexity in network utility maximization, leading to closed-form resource allocations and improved network fairness (Yang et al., 2017).
• ISAC Load-Balanced Architectures: Integrates communication and sensing QoS via mixed-integer convex-approximation, supporting URLLC with fronthaul and static power savings (Singh et al., 23 Jan 2026).

7. Practical Recommendations and Open Challenges

JPALB algorithms are central for energy-efficient resource management in next-generation wireless networks, especially as densification, heterogeneous topologies, and joint multi-service paradigms proliferate. For large-scale operation, group-sparsity-type IRLS and APG methods are recommended for balancing optimality, complexity, and practicability, especially as they naturally yield sparse AP activation patterns. MISOCP remains critical for benchmarking and moderate-scale deployments. Dynamic environments and distributed networks benefit from potential-game and fixed-point based approaches.

A continuing challenge is the extension of JPALB to scenarios with highly dynamic user mobility, stochastic service demands, and integrated non-linear cross-layer constraints (e.g., MEC and ISAC architectures with real-time adaptation). The development and theoretical guarantees of scalable, low-complexity algorithms that can seamlessly integrate discrete AP activation, continuous power allocation, and additional systems-level objectives constitute a current frontier in wireless systems optimization.

References: (Chien et al., 2019, Chien et al., 2020, Liao et al., 2016, Yang et al., 2017, Singh et al., 23 Jan 2026, Abbas et al., 19 Aug 2025, Chen et al., 2020, Hong et al., 2011, You et al., 2016)