Max-Min Fairness Optimization Strategy

• Max-min fairness is an optimization strategy that maximizes the minimum utility among agents, ensuring robust protection for the worst-off user.
• It employs methods such as closed-form water-filling, convexification via minorization-maximization, and alternating optimization to achieve Pareto optimal fairness.
• Widely applied in wireless communications, machine learning, and public resource allocation, the strategy balances efficiency with equitable outcomes.

Max-Min Fairness Optimization Strategy

Max-min fairness (MMF) is a foundational optimization paradigm for resource allocation in multi-agent and multi-user systems. It seeks to maximize the minimum utility (rate, allocation, reward) across all participants, ensuring robust protection of the worst-off agent or group. In contrast to sum-utility (efficiency-oriented) objectives, the MMF criterion provides strong egalitarian guarantees and is widely deployed in communication networks, machine learning, public goods allocation, and digital services. The formal definition of the MMF objective is

$$\max_{\text{allocations}}\;\min_{k}\;\text{utility}_k\qquad \text{subject to feasibility constraints}.$$

Below, key principles, canonical formulations, algorithmic methodologies, analytical insights, and domain-specific applications of max-min fairness optimization are detailed with technical precision.

1. Mathematical Formulation and Principle

The max-min fairness objective seeks an allocation that maximizes the utility achieved by the worst-off user. In vector form, for resource allocation vector $\mathbf{x}=(x_1,\dots,x_N)$:

$$\max_{\mathbf{x}\in\mathcal{F}}\;\min_{k=1,\dots,N}\;U_k(\mathbf{x}),$$

where $\mathcal{F}$ encodes arbitrary feasibility constraints (e.g., convex hulls, combinatorial/network polytopes), and $U_k$ are continuous (often concave) per-user utility/rate/loss functions. When direct expression as a convex program is feasible, an epigraphic reformulation is standard:

$$\begin{aligned} &\max_{t,\mathbf{x}\in\mathcal{F}}\; t\ &\text{s.t.}\; U_k(\mathbf{x})\ge t,\;\forall k. \end{aligned}$$

This formulation guarantees Pareto optimality with respect to the worst-case outcome and is immune to efficiency–fairness trade-offs commonly observed in sum-utility maximization.

2. Algorithmic Techniques: Closed-Form, Convexification, Alternating Optimization

Approaches for max-min fairness optimization depend critically on the structure of $U_k$ and $\mathcal{F}$:

• Closed-Form and Water-Filling: In specialized cases (notably two-user RSMA downlink), the MMF allocation admits a full closed-form, leveraging fixed precoder directions, water-filling for power allocation, and a one-dimensional search over corner points. For example, in the two-user MISO broadcast RSMA problem (Luo et al., 2023), the optimal minimum rate is realized at one of a small set of analytically characterized points, each corresponding to an operational regime (pure multicast, SDMA, NOMA, or RSMA). The entire algorithm involves $O(1)$ closed-form evaluations and $O(N_t)$ operations per realization, bypassing iterative convex solvers.
• Convex Surrogate and Minorization-Maximization (MM): When $U_k(\mathbf{x})$ is non-concave or non-smooth, convexification procedures such as MM are used. MM constructs an affine (or convex) surrogate that minorizes the original objective at the current iterate, leading to monotonic convergence toward stationarity. At each step, a convex program (e.g., SOCP) is solved, often in the epigraph form above (Naghsh et al., 2019).
• Alternating Optimization with Block Updates: For multi-block (e.g., beamforming, power, rate-splitting) variables, alternating procedures—where each block is optimized while others are fixed—yield practical solutions. Typical subproblems include (i) power allocation (often convexified as geometric programming), (ii) beamformer optimization (quadratic programming, Riemannian manifold methods), and (iii) rate splitting via water-filling or closed-form splits (Quran et al., 13 May 2025, Ginige et al., 20 Apr 2025). Convergence to a KKT point is ensured if each block update is globally optimal.
• Fractional Programming and Extragradient-VI: Recent advances handle non-convex, non-smooth max-min RSMA formulations via fractional programming transformations that make log-SINR objectives block-concave and subsequently apply extragradient methods to solve the associated variational inequalities (VIs) efficiently and scalably, exploiting optimal beamforming structures (span of user channels) (Luo et al., 6 Jul 2025).

Examples of these algorithmic methods, together with tight pseudocode, are present in the detailed algorithm sections of the referenced works.

3. Application Domains

A. Wireless Communication Networks

Max-min fairness is the principal fairness notion in multiuser transmission. Formulations and algorithms span:

• Rate-Splitting Multiple Access (RSMA) Systems: Closed-form optimizers (Luo et al., 2023) and advanced extragradient-based methods (Luo et al., 6 Jul 2025) for downlink MISO channels ensure symmetric rates or bounded rate-drops for the weakest user, with direct computational benefits compared to conventional WMMSE or SCA algorithms.
• Interference and Broadcast Channels: MMF is applied to precoder design via MM–SOCP iterations that handle nonconvex rate (log-det) expressions even in robust regimes with channel or noise uncertainty (Naghsh et al., 2019).
• Stacked Intelligent Metasurface (SIM) and ISAC Systems: Hybrid applications optimize both beamforming and power coefficients under fairness, using alternating optimization with geometric programming, gradient ascent-descent, and Riemannian conjugate gradient updates (Ginige et al., 20 Apr 2025, Quran et al., 13 May 2025, Kanani et al., 24 Jul 2025).
• Cell-Free Massive MIMO: MMF is achieved via alternating generalized eigenvalue solvers and geometric programming or inferential transformer-based predictors for scalable, configuration-agnostic power control (Arachchi et al., 2019, Chafaa et al., 5 Mar 2025).
• Full-Duplex OFDMA: In multi-level user assignment and resource scheduling, MMF emerges as a computationally demanding combinatorial-programming problem; relaxation, nonconvex regularization, and greedy rounding yield near-optimal solutions with substantial empirical gains (Zhang et al., 2018).

B. Networks and Public Resource Allocation

• 802.11 Mesh Networks: Log-convex rate regions enable convex optimization for max-min throughput allocations, coupled with distributed algorithms exploiting per-flow bottleneck characterizations and local AIMD-style adjustments (Leith et al., 2010).
• Millimeter-Wave Backhaul: Progressive-filling airtime allocation with per-clique constraints assures end-to-end flow fairness in multihop environments subject to spatial conflict graphs (Li et al., 2017).
• Dynamic Allocation in Public Goods: Stochastic allocation of indivisible resources among agents is explained via threshold policies, with robust guarantees and data-driven policies shown to yield improved welfare at symmetric equilibria (Onyeze et al., 24 Jan 2025).

C. Machine Learning and Recommender Systems

• Group/Subgroup Fairness in Prediction: MMF constraints take the form $\min_{g\in G}U_g(\theta)$, with tailored primal-dual algorithms (FairDual, M$^2$FGB) that reweight group losses in each mini-batch, adjust group exposure probabilities, and provide theoretical convergence and Jensen gap bounds in stochastic settings (Xu et al., 13 Feb 2025, Pereira et al., 16 Apr 2025).
• Provider-Fair Recommender Systems: MMF-regularized online re-ranking routines (P-MMF) model exposure allocation as a resource LP, leveraging dual reduction and momentum gradient descent in high-dimensional dual space, with provable sublinear regret and efficient dynamical adaption (Xu et al., 2023).
• Active Sampling for Min-Max Fairness: The simple strategy of always updating the model on the currently worst-off group provably drives worst-group error to the min-max fair optimum at O($1/\sqrt{T}$) or O($1/T$) rates in convex loss settings (Abernethy et al., 2020).

4. Analytical Properties and Performance Guarantees

• Existence and Uniqueness: On compact convex rate or allocation regions with free disposal, MMF solutions exist and are unique (Leith et al., 2010, Li et al., 2017).
• Closed-form Solutions and Structural Insights: In several two-user or symmetric settings, MMF split/power-allocation can be given in closed form, facilitating symbol-period implementation (Luo et al., 2023).
• Complexity: MMF is NP-hard for general nonconvex, combinatorial, or interference-coupled systems (Naghsh et al., 2019, Zhang et al., 2018). Nevertheless, specialized convexification/alternating algorithms achieve practical runtimes (polynomial in relevant system dimensions), and with closed-form structures or learning-based methods, complexity can be decoupled from system size (Luo et al., 6 Jul 2025, Chafaa et al., 5 Mar 2025).
• Scalability and Implementation: Many MMF algorithms are fully distributed or require limited information sharing (e.g., local idle-probability, or a small set of positive scalars between clusters), enabling implementation in practical mesh/backhaul, wireless, and recommendation systems (Leith et al., 2010, Li et al., 2017, Huang et al., 2012).
• Robustness and Generalization: MMF solutions generalize to robust optimization with imperfect CSI/noise by substituting lower bounds on rates into the same architecture (Naghsh et al., 2019, Luo et al., 6 Jul 2025). In learning, primal-dual reweighting ensures subgroup improvement at modest overall utility cost, with monotonic tradeoff in fairness parameter $\lambda$ (Pereira et al., 16 Apr 2025).

5. Domain-Specific Example Algorithms and Performance

The following table gives an overview of MMF optimization strategies deployed in core application domains:

Domain	MMF Formulation	Core Algorithmic Approach
RSMA Downlink (2-user MISO)	Maximize $\min_k\{R_k + C_k\}$	Closed-form water-filling, 1D search, fixed beam directions (Luo et al., 2023)
General MIMO-IC	Maximize $\min_k R_k$	Minorization-Maximization + SOCP at each iteration (Naghsh et al., 2019)
Metasurface-Aided Multi-User MISO	Maximize min-rate (users)	Alternating GP (power) and GDA/RCG (phase), quantization (Ginige et al., 20 Apr 2025)
Cell-Free Massive MIMO	Maximize $\min_k$ SINR	Alternating generalized eigensolver (combining) and GP (power) (Arachchi et al., 2019)
ML Group Fairness	Minimize $\max_g L_g(\theta)$	Dual (reweighting) SGD/mirror descent, minibatch gap control (Xu et al., 13 Feb 2025)
Recommender (Provider MMF)	Maximize min-exposure per provider	Dual reduction, online momentum gradient in dual space (Xu et al., 2023)
Digital FM Resource Allocation	Maximize $\min_i x_i$, $\leq d_i$, $\sum x_i \leq B$	Sort-and-fill, $O(N\log N)$ for embedded devices (Martínez et al., 2024)

Empirical results across domains show that MMF-based methods reliably yield significantly improved minimum-rate or minimum-exposure outcomes relative to baseline policies (SDMA, NOMA, sum-rate maximization, proportional fairness). For example, in 2-user RSMA, closed-form MMF achieves 92-93% of the full WMMSE/SCA with $10^3$ speedup (Luo et al., 2023); in full-duplex OFDMA, MMF-aware scheduling and power allocation achieve up to $25\times$ improvement in the 80th-percentile MMF rate over heuristics (Zhang et al., 2018).

6. Generalizations, Extensions, and Practical Considerations

• Multiuser Generalization: Closed-form insights gained in the two-user MMF RSMA setting extend conceptually to $K$-user systems via multi-dimensional water-filling and critical point analysis, albeit with combinatorially more candidate solutions (Luo et al., 2023, Luo et al., 6 Jul 2025).
• Robust MMF: Extensions to settings with imperfect channel state information, noise variance, or demand uncertainty are enabled by lower-bound substitutions in the MMF objective (Naghsh et al., 2019, Luo et al., 6 Jul 2025).
• Distributed/Federated MMF: In mesh/backhaul and multicell wireless, distributed protocols achieving MMF exchange only small numbers of scalars or easily measurable local statistics, and leverage iterative interference functions or uplink-downlink duality (Leith et al., 2010, Li et al., 2017, Huang et al., 2012).
• Fairness-Efficiency Trade-Off: Theory and empirical analysis establish the inherently monotonic trade-off between maximizing the worst-case utility (fairness) and overall efficiency (sum-utility), with max-min designs often achieving Pareto boundary points for the feasible region (Huang et al., 2012).
• Numerical Stability and Implementation: Several algorithms exploit invariance principles (beamformer structure, dual variable magnitude) and closed-form Jacobians to ensure robust fast convergence. Embedded and real-time contexts benefit from the inherently low complexity and finite evaluation structure of many MMF optimizers (Luo et al., 2023, Martínez et al., 2024).

7. Theoretical and Empirical Performance Highlights

• Optimality and Convergence: MMF-centric algorithms—where block updates are globally or locally optimal and convexification is used judiciously—feature provable convergence to saddle points or KKT conditions, respecting the non-convexity of resource coupling.
• Scalability: Reducing algorithmic dependence on network size (antennas, users) is achieved both via structural problem reductions (channel-span beamforming) and machine-learning-based predictors for large-scale real-time regimes (Luo et al., 6 Jul 2025, Chafaa et al., 5 Mar 2025).
• Empirical Utility: Across wireless, networking, and ML recommender domains, the implementation-ready MMF strategies robustly achieve fairness targets, protect the worst-off agents/groups/flows, and—especially in complex interference and combinatorial regimes—outperform heuristic or legacy designs by wide margins.
• General Applicability: The principles and algorithmic frameworks of MMF optimization are generic and extensible to diverse constrained allocation problems in digital and physical networks, providing strong theoretical guarantees and practical efficacy throughout modern multi-agent system design.