PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Recent Advances in Optimization Methods for Wireless Communications

Ya-Feng Liu, Tsung-Hui Chang, Mingyi Hong, Zheyu Wu, Anthony Man-Cho So, Eduard A. Jorswieck, Wei Yu

arXiv:2401.12025v3cs.ITeess.SPmath.OC

TL;DR

Wireless communication design produces structured optimization problems spanning nonconvex, global, integer, distributed, and learning-based settings. This paper surveys recent methods and applications, emphasizing algorithms and neural architectures that exploit such structure, while also identifying open challenges and practical scope boundaries.

  • Problem

    Wireless-system optimization has become increasingly structured and difficult, motivating methods tailored to diverse problem forms and system scenarios.

  • Method

    The paper surveys optimization theory, algorithms, and applications across nonconvex optimization, global optimization, integer programming, distributed optimization, federated learning, and learning-based optimization.

  • Results

    The survey organizes recent advances and successful wireless-communication applications around selecting or developing algorithms and neural architectures that exploit underlying problem structure.

  • Takeaways & Limitations

    Suitable structure-aware optimization methods can provide efficient, provable, and interpretable approaches for wireless communication system design.

  • Takeaways & Limitations

    Learning-based wireless optimization can involve highly nonlinear training objectives that are difficult to optimize, creating a tradeoff between alternative formulations.

Abstract

from arXiv · show

Mathematical optimization is now widely regarded as an indispensable modeling and solution tool for the design of wireless communications systems. While optimization has played a significant role in the revolutionary progress in wireless communication and networking technologies from 1G to 5G and onto the future 6G, the innovations in wireless technologies have also substantially transformed the nature of the underlying mathematical optimization problems upon which the system designs are based and have sparked significant innovations in the development of methodologies to understand, to analyze, and to solve those problems. In this paper, we provide a comprehensive survey of recent advances in mathematical optimization theory and algorithms for wireless communication system design. We begin by illustrating common features of mathematical optimization problems arising in wireless communication system design. We discuss various scenarios and use cases and their associated mathematical structures from an optimization perspective. We then provide an overview of recently developed optimization techniques in areas ranging from nonconvex optimization, global optimization, and integer programming, to distributed optimization and learning-based optimization. The key to successful solution of mathematical optimization problems is in carefully choosing or developing suitable algorithms (or neural network architectures) that can exploit the underlying problem structure. We conclude the paper by identifying several open research challenges and outlining future research directions.

I. INTRODUCTION

Wireless advances from 3G through 6G have made optimization problems more structurally challenging, while optimization remains central to wireless-system design. This survey reviews recent methods and guides algorithm selection by exploiting problem structure.

  • Wireless evolution: Emerging wireless requirements and 6G usage scenarios continue to drive technological development and new optimization needs.These requirements include data rate, latency, efficiency, connectivity density, and related KPIs.
  • Role of optimization: Optimization is a powerful and indispensable tool for modeling, analyzing, and solving wireless communication system-design problems.It supports formulation, structural analysis, algorithm development, and convergence analysis.
  • Optimization challenges: Compared with 3G, 4G–6G systems produce problems that are often nonconvex, nonsmooth, non-Lipschitz, nonseparable, nondeterministic, or mixed-variable.These features substantially complicate understanding, analysis, and solution methods.
  • Survey scope: The survey reviews nonconvex nonsmooth, global, distributed, and learning-based optimization, emphasizing theoretical properties and wireless applications.Its coverage extends beyond earlier surveys centered primarily on convex optimization.
  • Paper goals: The paper guides algorithm and neural-network design by relating suitable methods to the special structures and features of wireless optimization problems.It also seeks cross-fertilization between optimization research and wireless communications.
  • Illustrative problems: Hybrid beamforming illustrates the structural difficulty of jointly designing unit-modulus analog beamformers and digital beamformers.The analog matrix is typically implemented with phase shifters, while digital beamformers have reduced dimension.

2) Optimization Problems with Integer Variables:

Integer and discrete variables arise naturally in massive-MIMO problems such as MIMO detection and symbol-level precoding. These formulations introduce discrete optimization challenges, including large problem sizes that can limit existing algorithms.

  • Discrete formulations: Constellation-symbol optimization produces discrete optimization problems in massive MIMO.The section introduces integer or discrete variables through massive-MIMO examples.
  • MIMO detection: MIMO detection recovers transmitted symbols from received signals using knowledge of the channel matrix.The transmitted symbols come from discrete constellations such as QAM or PSK.
  • MIMO detection: Massive-MIMO detection has a large problem size that can prevent the use of algorithms efficient only for small- to medium-sized problems.Semidefinite-programming-relaxation methods are cited as an example.
  • Symbol-level precoding: Symbol-level precoding designs a downlink transmit signal so received signals align closely with desired constellation points.The formulation assumes channel-state information is available at the transmitter.

3) Optimization Problems with Mixed Variables:

Wireless design problems increasingly combine continuous and discrete decisions, producing mixed-variable optimization challenges. These challenges also reflect larger, more nonlinear, and structurally difficult systems in 5G and 6G.

  • Mixed-Variable Structure: Mixed-integer variables arise in admission control, user scheduling, and BS-user association, and are generally harder to solve than continuous counterparts.Binary on-and-off variables are a common form of the discrete component.
  • Joint Admission Control and Multicast Beamforming: Admission control selects users whose SNR constraints can be met, while alternative formulations maximize admitted users or minimize power for a target cardinality.The selected-user formulation uses binary variables alongside a continuous multicast beamforming vector.
  • Joint Uplink Scheduling and Power Control: Joint uplink scheduling and power control couples discrete user-selection variables with continuous transmit powers.The scheduling variable identifies the user served in each cell, while power control determines its transmission power.
  • Emerging System Complexity: Wireless optimization problems have grown more challenging because system dimensions and nonlinear couplings increase with antennas, users, subcarriers, devices, and RIS elements.RIS systems can multiplicatively couple reflective and transmit beamforming variables within fractional and logarithmic expressions.
  • Emerging System Complexity: Many formulations also lack favorable properties such as convexity, smoothness, Lipschitz continuity, separability, or determinism.The supplied discussion attributes these properties partly to regularizers that promote sparsity, low-rankness, and fairness.

C. Structural Properties of Optimization Problems

Wireless optimization problems should be analyzed through their structural properties before algorithms are selected. The survey connects these properties to tailored methods for fractional, sparse, nonconvex, and mixed-structure problems.

  • Structural Properties: Recognizing function properties, variable types, coupling, hidden convexity, duality gaps, projections, bounds, and solution conditions guides algorithm selection.The survey emphasizes that structural recognition is central to both analysis and solution design.
  • Structural Properties: MIMO detection can use gradient projection because its variables are decoupled in the constraints and its quadratic objective has a Lipschitz-continuous gradient.The feasible set therefore has an easy-projection property.
  • Structural Properties: Riemannian conjugate gradient exploits the unit-modulus manifold in RIS and hybrid beamforming, improving on Euclidean gradient projection for faster convergence.The method projects gradients onto the tangent space of the complex circle manifold.
  • Structural Properties: Hidden convexity can convert seemingly nonconvex beamforming formulations into equivalent convex reformulations.The survey defines hidden convexity as the existence of such an equivalent convex formulation.
  • Computational Complexity: Complexity analysis helps determine whether to pursue exact algorithms, efficient special cases, fast heuristics, or relaxations.For hard problems, the survey advises lowering the priority of efficient exact algorithms.
  • Fractional Programming: Fractional programming addresses ratio-based wireless metrics such as SINR, using transforms that can preserve global optimality for concave-convex single-ratio problems.The Charnes-Copper and Dinkelbach transforms are identified as classic single-ratio techniques.
  • Fractional Programming: Quadratic-transform alternating optimization can converge to a stationary point when the transformed problem is convex in one variable block at a time.The transform also extends to broader sum-of-functions-of-ratio and matrix settings.
  • Fractional Programming: The Lagrangian dual transform moves SINRs outside logarithms, while the quadratic transform decouples numerators and denominators for alternating optimization.Under suitable convexity conditions, alternating optimization converges to a stationary point.

2) Application Examples:

The survey applies structural optimization tools to wireless beamforming, scheduling, power control, and sparse recovery. These transformations can yield efficient alternating or distributed procedures while generally targeting stationary or high-quality solutions.

  • Downlink Beamforming: FP reformulations solve downlink sum-rate beamforming by alternating updates of auxiliary variables and beamformers, with convergence to a stationary point.The beamformer update can use a Lagrange multiplier found by bisection under the total power constraint.
  • Downlink Beamforming: The FP algorithm is equivalent to WMMSE for sum-rate maximization, although different treatments of terms can produce different update rules in complex scenarios.The survey describes the FP form as practically preferable when it yields distributed optimization.
  • Uplink Scheduling and Power Control: For joint uplink scheduling and power control, FP transforms produce an alternating formulation whose scheduling and power variables decouple across cells when the auxiliary variable is fixed.This permits independent per-cell scheduling and power optimization.
  • Optimization Transforms: Quadratic and Lagrangian dual transforms lift difficult fractional problems into equivalent higher-dimensional forms that are easier to optimize blockwise.They can also enable distributed optimization in settings with discrete scheduling or multiplicatively coupled variables.
  • Sparse Optimization and Recovery: Sparse optimization exploits solutions with few nonzero entries and supports compressed-sensing recovery from underdetermined measurements.The ℓ0 formulation is generally NP-hard, motivating computationally efficient alternatives such as ℓ1 minimization.
  • Sparse Optimization and Recovery: Under the stated RIP condition, ℓ1 minimization recovers the same k-sparse signal as ℓ0 minimization, while noisy recovery error is O(ϵ).The supplied passage states the recovery equivalence and the noise-error order.

2) Application Examples:

Sparse optimization and compressed sensing support wireless channel modeling and massive-access detection by exploiting limited angular or user activity structure. The surveyed methods provide models and detection guarantees under stated measurement and asymptotic conditions.

  • Localized Statistical Channel Modeling: Localized statistical channel modeling uses sparse angular power spectra inferred from beam-wise RSRP measurements.The RSRP vector and channel APS are linked linearly through a sensing matrix, with high angular resolution producing a large ambient dimension.
  • Localized Statistical Channel Modeling: Limited scattering produces a sparse channel APS, enabling optimization-based construction of localized models from underdetermined measurements.The sensing matrix depends on beam waveforms and antenna gains.
  • Localized Statistical Channel Modeling: Localized statistical channel models are statistically indistinguishable from the true propagation environment and facilitate offline network-performance evaluation.The passage connects the model construction to simulation for offline network optimization.
  • Device Activity Detection: Grant-free mMTC random access detects active devices from signature sequences when only a small subset of users is active.The received pilot signals can be written as Y = SX + Z, where sporadic activity makes most rows of X zero.
  • Device Activity Detection: Group-sparse formulations and AMP exploit the row sparsity of the channel-activity matrix, recovering activities and channel estimates.The ℓ2,1-norm promotes group sparsity, and AMP performance is analyzed through state evolution.
  • Device Activity Detection: As M tends to infinity, AMP can make missed-detection and false-alarm probabilities tend to zero under the surveyed formulation.This result is stated for the device-activity detection approach that exploits sparse structure.
  • Device Activity Detection: As M tends to infinity, covariance-based detection can identify active devices with probability at least 1 −exp(−c2L), and detectable activity scales quadratically with signature length L.The stated guarantee assumes the signature and activity conditions described in the passage.
  • Device Activity Detection: Covariance-based formulations use the sample covariance of received signals and have been extended to joint detection, multicell, asynchronous, low-resolution-ADC, and unsourced-access settings.The passage lists these extensions as applications of the covariance-based approach.

3) Remarks:

The PG and GP algorithms exploit problem structure to solve challenging wireless communication designs, with general convergence guarantees and stronger problem-specific results. Applications include massive MIMO detection and joint base-station clustering with beamformer design.

  • PG algorithm: PG minimizes a first-order quadratic upper approximation while retaining the nonsmooth term, and its efficiency depends on computing the proximal operator.For suitable step sizes, the update fits the MM framework; many practical proximal operators have closed-form or efficient solutions.
  • PG algorithm: In the nonconvex case, PG iterates converge to a critical point under mild closedness and Kurdyka-Łojasiewicz conditions, including semi-algebraic functions.Inexact proximal calculations can also be accommodated in established convergence analyses.
  • Massive MIMO detection: For massive MIMO detection, GP is computationally efficient because projections onto the decoupled PSK or QAM constellation set reduce to simple componentwise projections.Each iteration mainly requires two matrix-vector multiplications and one easily computed projection.
  • Massive MIMO detection: Under mild conditions, massive-MIMO GP iterates converge to the true symbol vector within finitely many iterations, exploiting constellation and channel structure.The conditions are roughly small noise variance and a large M/K ratio, yielding a guarantee stronger than generic critical-point convergence.
  • Joint BS clustering and beamformer design: Joint base-station clustering and beamformer design uses reformulation, block coordinate descent, and a dual approach when quadratic constraints make direct PG inefficient.The dual inner problems are separable convex subproblems solvable by PG, while the outer problem is one-dimensional and convex.

3) Remarks:

Penalty methods transform structured discrete or difficult constrained problems into more tractable continuous formulations while controlling the relaxation gap. Their exactness can support algorithm design, especially when the penalty exploits problem-specific structure.

  • Penalty methods: Penalty methods replace difficult constrained problems with easier subproblems and rely on exactness to preserve solutions when the penalty parameter is sufficiently large.Exactness means the penalized problem eventually shares a solution with the original constrained problem.
  • Structured constraints: The penalty formulation relaxes one-hot binary assignment constraints to a simplex and then adds a negative-square term to drive solutions toward the discrete boundary.The relaxed feasible set is the convex hull of the original feasible set.
  • Structured constraints: For PSK-based MIMO detection, a negative-square penalty tightens a convex relaxation by penalizing the relaxation gap and can recover the original discrete feasible set.When λ exceeds the largest eigenvalue of Q, the penalized objective is strictly concave and attains its solution on the boundary corresponding to the original feasible set.

2) Remarks:

The survey highlights duality and KKT structure as routes to efficient globally optimal algorithms for wireless optimization. These methods depend on nontrivial convex reformulations and careful exploitation of solution structure, while newer communication settings introduce additional discrete and combinatorial difficulty.

  • Duality-based algorithms: Lagrangian and uplink-downlink duality reveal intrinsic structure in wireless power-control and beamforming problems.Duality-based algorithms often use simple fixed-point iterations and can achieve global optimality.
  • Duality-based algorithms: A joint beamforming and compression problem is first reformulated as an equivalent convex SDP, then its enhanced KKT conditions are separated into dual and primal fixed-point systems.Each system can be solved through fixed-point iteration after exploiting the problem’s structure.
  • Remarks: Globally efficient duality-based algorithms require an equivalent convex reformulation and careful use of the underlying solution structure.The first step can be highly nontrivial, while the second is essential for computational efficiency.
  • Remarks: Modern multiple-access systems such as NOMA add discrete decoding-order variables to already NP-hard interference-channel sum-rate optimization.This creates additional combinatorial challenges across several multicarrier NOMA variants.

IV. PROBLEM-SPECIFIC GLOBAL OPTIMIZATION

The survey presents branch-and-bound and branch-and-cut as foundational frameworks for problem-specific global optimization in wireless communications. Their effectiveness depends on bounding, pruning, cuts, and problem-tailored design, while global optimality can require exponential worst-case complexity.

  • Global optimization seeks global rather than local or suboptimal solutions, but generally requires more difficult algorithmic design.
  • Branch-and-bound: Branch-and-bound partitions feasible regions recursively, uses bounds to prune unpromising regions, and returns a global solution after tree exploration.
  • Branch-and-bound: Branch-and-bound design choices include subdivision, node selection, bounding, reduction, feasibility checks, and feasible-point generation.
  • Branch-and-bound guarantees global optimality, but its worst-case complexity is generally exponential and depends on tree depth, branching, and subproblem-solving cost.
  • Branch-and-cut: Branch-and-cut strengthens relaxations with valid cutting planes before or during branch-and-bound, preserving integer solutions while tightening the feasible region.
  • Problem-specific techniques surveyed include tight semidefinite-relaxation bounds for complex quadratic problems, monotonicity-based bounds, and valid cuts for mixed-integer formulations.

C. Mixed Monotonic Programming

Mixed monotonic programming provides lower bounds over rectangular regions and integrates with branch-and-bound to globally solve wireless optimization problems. Bound quality is central: tighter bounds can accelerate global search, while representation choices affect tightness.

  • Mixed monotonic programming represents objectives through mixed monotonic functions and obtains lower bounds over rectangular feasible-region subsets.
  • The framework globally solves the K-user interference-channel sum-rate problem by combining an objective representation with standard branch-and-bound components.
  • MMP bounds are always better than the compared difference-of-monotonic bounds for the sum-rate maximization problem, explaining faster branch-and-bound convergence.
  • The MMP framework includes difference-of-monotonic programs as a special case and has implementations in C++ and Python.
  • Different mixed monotonic representations produce different bounds, so selecting a representation that yields tight bounds is a key design issue.

D. Valid Cuts for Mixed-Integer Problems

The survey describes valid cuts and mixed-integer relaxations as mechanisms for strengthening global optimization of mixed-integer wireless problems. A branch-and-cut procedure alternates relaxation, cut generation, and optimality checks, with finite termination for the considered binary formulation.

  • Valid inequalities tighten mixed-integer relaxations without excluding feasible integer solutions, but deriving them generally requires exploiting special problem structure.
  • Keeping binary variables intact while relaxing nonconvex quadratic constraints yields a mixed-integer semidefinite relaxation for the JABF problem.
  • Gaussian randomization converts the mixed-integer relaxation solution into a feasible JABF solution with a provable guarantee.
  • The branch-and-cut design first constructs an outer relaxation whose constraints can be expressed as second-order cone constraints for finite selected sets.
  • Solving the inner semidefinite problem either verifies global optimality or generates a valid inequality that eliminates the current candidate and tightens the relaxation.
  • Because the feasible binary solution set is finite and each iteration eliminates one binary solution, the branch-and-cut algorithm terminates with an optimal solution in finitely many iterations.

V. DISTRIBUTED OPTIMIZATION AND FEDERATED LEARNING

Distributed optimization decomposes wireless system design across entities that exchange localized information, supporting scalability and collaborative interference management. The survey contrasts dual decomposition with ADMM, emphasizing ADMM’s broader assumptions and convergence advantages alongside parallelization trade-offs.

  • Distributed optimization enables multiple entities to solve global wireless optimization problems collectively through localized computations.
  • In coordinated multi-cell and cell-free MIMO systems, collaborating base stations can mitigate inter-cell interference and improve cell-edge-user QoS.
  • Distributed optimization is positioned as important for edge intelligence, where computation at network edges supports real-time decision-making and latency reduction.
  • Dual decomposition: Dual decomposition exploits separable convex structure by solving local subproblems in parallel and coordinating them through a shared constraint.
  • Dual decomposition: Dual decomposition may converge slowly when strict convexity is absent, and its output is not generally guaranteed to be feasible.
  • ADMM: ADMM splits convex problems into simpler decoupled or parallel subproblems using an augmented Lagrangian and converges under mild assumptions.
  • ADMM: The augmented-Lagrangian update gives ADMM a faster convergence rate than dual decomposition, while proximal ADMM enables parallel subproblem updates under a suitable positive-definite matrix choice.

2) Application Example:

Distributed optimization reformulates coupled multi-cell beamforming constraints so base stations can solve separable local subproblems using ADMM and related methods.

  • ADMM is applied to multi-cell coordinated beamforming in distributed wireless system design.
  • Introducing local interference copies transfers coupling from SINR constraints to equality constraints that can be decoupled.
  • The reformulated objective and constraints become separable across base stations, enabling local ADMM updates.
  • Each base station shares relevant local variables, computes the global interference vector, and updates its multiplier until convergence.
  • ADMM can track solution variation in time-varying channels and, with semidefinite relaxation, address robust beamforming under imperfect CSI.

B. Federated Learning in Wireless Edge Networks

Federated learning coordinates edge clients through a server that aggregates local model updates, with FEDAVG supporting full or partial participation under communication constraints.

  • The wireless edge network uses an edge server to orchestrate clients collaboratively solving a distributed learning problem via federated learning.
  • Each client’s local objective is an expected loss over its local dataset, while the global objective aggregates the clients’ data distributions.
  • FEDAVG extends consensus-based distributed SGD to a star network through broadcasting, local SGD updates, and server aggregation.
  • Full participation aggregates all clients, whereas partial participation selects a smaller client subset in each communication round.
  • Full participation can be difficult when uplink bandwidth is limited and many clients must transmit model updates.
  • The selected-client averaging scheme provides an unbiased estimate of the global model average.

2) Performance Analysis:

FEDAVG performance depends on client participation, local updating, and data heterogeneity, while wireless optimization and learning methods address difficult interference-limited design problems.

  • FEDAVG performance is influenced by selected-client count, local update steps, and data heterogeneity, whose effects interact.
  • Under stated smoothness, unbiased-gradient, bounded-variance, and heterogeneity assumptions, increasing selected clients alleviates stochastic variance and heterogeneity effects.
  • Partial participation yields convergence rate O(1/T^1/4), while full participation removes its associated term and improves the rate.
  • Wireless resource allocation studies jointly address client selection, transmission reliability, quantization, and convergence in non-ideal FL environments.
  • Power control and spectrum allocation are generally NP-hard in interference-limited networks, motivating computationally efficient and data-driven methods.

A. Black-Box Based Approaches

Learning-based optimization replaces or approximates iterative wireless solvers with neural networks, while also enabling optimization from richer representations without explicit channel estimation.

  • A. Black-Box Based Approaches: Learning-to-optimize treats an iterative optimization algorithm as a nonlinear mapping from problem specifications to decision variables.
  • A. Black-Box Based Approaches: A DNN can approximate WMMSE using a relatively simple architecture, achieving 25 to 250 times faster execution than the best C implementation.
  • A. Black-Box Based Approaches: Unsupervised learning can optimize system utilities directly without requiring existing algorithms to generate training labels, but its WSR objective is difficult to optimize.
  • A. Black-Box Based Approaches: Deep unfolding builds multi-stage networks that imitate finite iterations of known algorithms, reducing the number of learned parameters relative to black-box DNNs.
  • A. Black-Box Based Approaches: Channel estimation becomes increasingly difficult as antenna, reflector, and network sizes grow, making accurate network-wide CSI a potential optimization bottleneck.
  • A. Black-Box Based Approaches: Learning-based optimization can combine CSI with locations, images, or sensing data and potentially reduce or eliminate explicit channel estimation.

2) Beamforming and RIS Reconfiguration with Implicit Channel Estimation:

The paper contrasts traditional channel-estimation-based optimization with model-free neural networks that directly map received pilots to optimized beamforming and RIS configurations. These approaches can reduce pilot requirements, improve rates under estimation error, and extend optimization to sensing and tracking, while leaving theoretical, communication, and deployment challenges.

  • Implicit channel estimation: Traditional optimization estimates the channel from pilots before optimizing, but channel-model and estimation choices involve tradeoffs.
  • Implicit channel estimation: Neural networks can directly map received pilots to optimized solutions using channel information implicitly contained in the observations.
  • RIS reconfiguration: Model-free RIS optimization can use fewer pilots than traditional channel-estimation approaches and achieve higher overall rate than manifold optimization and block coordinate descent when estimation error is considered.
  • Sensing and beam alignment: The same learning-based framework supports sensing and localization, including localization and mmWave massive MIMO initial beam alignment, where the beamformer aligns with an incoming ray during the pilot stage.
  • Sensing and beam alignment: Deep learning can perform sequential sensing and incorporate visual imaging data for beam tracking and alignment without explicitly estimating the channel matrix.
  • Open problems: Open challenges include theoretical guarantees, neural-network architecture and sample selection, constraint handling, general duality-based algorithms, communication-efficient distributed optimization, and practical interconnection costs.
Loading 2401.12025v3…