Max-Min Algorithms in Optimization

• Max-min algorithms are defined by optimizing the worst-performing component, ensuring fairness and robust performance across distributed and multi-objective systems.
• They are applied in areas like sensor scheduling, grouped bandits, and reinforcement learning, using both local algorithms and global optimization techniques.
• The approach balances efficiency and complexity through methods such as local approximations, SDP relaxations, and Bayesian strategies for decision-making under worst-case scenarios.

Max-min algorithms are optimization methods built around a worst-case or bottleneck objective: they maximize the minimum attained value across objectives, groups, agents, or scenarios. In linear-programming form, this appears as maximizing $\omega$ subject to $A\mathbf{x}\leq \mathbf{1}$, $C\mathbf{x}\geq \omega \mathbf{1}$, and $\mathbf{x}\geq \mathbf{0}$; in grouped bandits, it appears as selecting the group whose worst arm has the highest mean reward; and in multi-objective reinforcement learning it appears as $\max_{\pi}\min_k J_k(\pi)$, explicitly prioritizing fairness among objectives (0809.1489, Wang et al., 2021, Park et al., 2024). The subject spans distributed approximation, fair resource allocation, robust covering, sequential learning, and game-theoretic computation, with a recurring tension between worst-case guarantees, locality or oracle complexity, and global optimality.

1. Formal problem classes and max-min semantics

A canonical max-min linear program is

$$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$

where $V_i=\{v:a_{iv}>0\}$ and $V_k=\{v:c_{kv}>0\}$ are bounded support sets, and each agent $v$ is responsible for choosing $x_v$ in the distributed setting (0710.1499). In the bipartite specialization, the graph is $A\mathbf{x}\leq \mathbf{1}$0, each agent is adjacent to exactly one constraint and exactly one objective, and the degree bounds are captured by $A\mathbf{x}\leq \mathbf{1}$1 and $A\mathbf{x}\leq \mathbf{1}$2 (0804.4815).

In fairness-oriented network optimization, max-min fairness is defined so that no rate $A\mathbf{x}\leq \mathbf{1}$3 can be increased without either losing feasibility or decreasing some other $A\mathbf{x}\leq \mathbf{1}$4, and it is equivalent to lexicographic maximization of the sorted rate vector (Marasevic et al., 2014). In sensor scheduling, the generic allocation problem is

$A\mathbf{x}\leq \mathbf{1}$5

so the worst per-sensor cost is minimized under a shared resource budget (Wu et al., 2019). In grouped bandits, the objective becomes

$A\mathbf{x}\leq \mathbf{1}$6

which operationalizes the idea that a group is only as strong as its weakest arm (Wang et al., 2021). In multi-objective reinforcement learning, the same principle is described as the “egalitarian welfare function,” with the policy chosen to maximize the worst expected return across reward dimensions (Park et al., 2024).

These formulations differ in representation, but they share a common semantic structure: optimize against the weakest supported component of performance. A plausible implication is that max-min algorithms naturally arise whenever feasibility is local but performance is judged globally by a bottleneck criterion.

2. Local approximation and distributed max-min linear programming

A major strand of the literature studies local algorithms, meaning constant-time distributed algorithms in which each agent must choose its variable using only information from a constant-radius neighborhood. In the hypergraph model for max-min LPs, the sets $A\mathbf{x}\leq \mathbf{1}$7 and $A\mathbf{x}\leq \mathbf{1}$8 form hyperedges, and the local horizon $A\mathbf{x}\leq \mathbf{1}$9 is independent of instance size (0710.1499).

The foundational negative result is that in general bounded-degree instances there is no local approximation scheme. More precisely, when $C\mathbf{x}\geq \omega \mathbf{1}$0 and $C\mathbf{x}\geq \omega \mathbf{1}$1, no local algorithm can achieve approximation ratio less than

$C\mathbf{x}\geq \omega \mathbf{1}$2

even when all coefficients are in $C\mathbf{x}\geq \omega \mathbf{1}$3 and every variable appears in exactly one constraint and one objective (0710.1499). The same paper also identifies a positive regime: if the relative neighborhood growth

$C\mathbf{x}\geq \omega \mathbf{1}$4

is slow, then for any $C\mathbf{x}\geq \omega \mathbf{1}$5 there is a local algorithm with horizon $C\mathbf{x}\geq \omega \mathbf{1}$6 achieving approximation ratio $C\mathbf{x}\geq \omega \mathbf{1}$7 (0710.1499). This separates general expanders from grid-like or otherwise slow-growth structures.

For bipartite max-min LPs, the approximability threshold is sharper. There exists, for every $C\mathbf{x}\geq \omega \mathbf{1}$8, a local algorithm achieving approximation ratio

$C\mathbf{x}\geq \omega \mathbf{1}$9

and no local algorithm can achieve the ratio $\mathbf{x}\geq \mathbf{0}$0 in general, even in bipartite 0/1 instances and even with unique node identifiers (0804.4815). That work also introduces graph unfolding, i.e., replacing the original graph by its universal covering tree so that locally indistinguishable neighborhoods can be analyzed on a tree (0804.4815).

A closely related 0/1 line of work studies the case $\mathbf{x}\geq \mathbf{0}$1. There, a local approximation scheme is obtained after reducing constraints to size two and encoding the instance as a colored multigraph. The local decision rule has the explicit form

$\mathbf{x}\geq \mathbf{0}$2

with $\mathbf{x}\geq \mathbf{0}$3 and $\mathbf{x}\geq \mathbf{0}$4 computed from local alternating-walk statistics, yielding approximation ratio $\mathbf{x}\geq \mathbf{0}$5 after lifting back to the original instance (0806.0282). The general sparse case is then resolved by a constant-time distributed algorithm that uses local transformations, layered tree decomposition, a shifting strategy, and averaging over local layer assignments, achieving the optimal ratio $\mathbf{x}\geq \mathbf{0}$6 with a matching unconditional lower bound (0809.1489).

Taken together, these results characterize locality as a genuine algorithmic resource with sharp limits. In particular, max-min structure is harder than standard packing and covering in the fully local model (0809.1489).

3. Fair allocation, routing, and scheduling

In energy harvesting networks, max-min algorithms are used to allocate sensing rates fairly across both nodes and time slots. The setting includes unsplittable, tree, and fractional routing, with time-variable and time-invariable variants. The paper develops a fully combinatorial water-filling method for a given unsplittable routing, introduces problem-specific “fixing-the-rates” rules based on battery evolution and descendant structure, proves that selecting an unsplittable routing or a routing tree that yields lexicographically maximum rate assignment is NP-hard even for a single slot, and gives an FPTAS for the fractional time-variable case, whose feasibility is at least as hard as the 2-commodity feasible flow problem (Marasevic et al., 2014).

In max-min fair sensor scheduling, the allocation problem is reformulated as a zero-sum game between a “judge” choosing weights $\mathbf{x}\geq \mathbf{0}$7 on the simplex and an allocator choosing resources $\mathbf{x}\geq \mathbf{0}$8, with payoff

$\mathbf{x}\geq \mathbf{0}$9

Under the stated assumptions, there exists a unique Nash equilibrium in pure strategy, and the proposed equilibrium-seeking update

$\max_{\pi}\min_k J_k(\pi)$0

converges linearly under mild technical conditions, while diminishing stepsizes can be used when Lipschitz constants are unavailable (Wu et al., 2019). At equilibrium, all active agents satisfy an equalization property: if $\max_{\pi}\min_k J_k(\pi)$1, then $\max_{\pi}\min_k J_k(\pi)$2 equals the maximal cost (Wu et al., 2019).

In multi-objective reinforcement learning, the max-min objective

$\max_{\pi}\min_k J_k(\pi)$3

is treated through an LP-duality perspective based on state-action visitation frequencies, leading to an outer minimization over $\max_{\pi}\min_k J_k(\pi)$4 and an entropy-regularized soft Bellman operator (Park et al., 2024). The resulting model-free algorithm alternates between soft Q-learning and black-box gradient estimation for the weights. The paper reports that the method outperforms Utilitarian and Min-DQN baselines in a four-room environment, traffic light control, and species conservation, while achieving higher minimal return and more balanced allocation across objectives (Park et al., 2024).

Across these domains, max-min algorithms function as fairness mechanisms rather than merely robustness devices. They explicitly suppress solutions that improve aggregate utility by sacrificing a weak component.

4. Robust covering and worst-case decision making

A different tradition studies max-min structure through covering problems and robust optimization. In $\max_{\pi}\min_k J_k(\pi)$5-max-min covering, the task is to identify the $\max_{\pi}\min_k J_k(\pi)$6 demands that are costliest to cover. The thresholded covering template builds an anticipatory solution and repeatedly augments it whenever there exists a single demand whose augmentation cost exceeds a threshold. This yields improved approximation algorithms for $\max_{\pi}\min_k J_k(\pi)$7-robust Steiner tree and set cover, and the first approximation algorithms for $\max_{\pi}\min_k J_k(\pi)$8-robust Steiner forest, minimum-cut, and multicut; as a by-product, it also yields algorithms for max-min problems of the form “given a covering problem instance, which $\max_{\pi}\min_k J_k(\pi)$9 of the elements are costliest to cover?” (0912.1045). The analysis is organized around strongly discriminating algorithms and witness-set constructions via dual rounding (0912.1045).

Large-scale robust optimization is addressed through a max-min-max reformulation of the Lagrangian dual:

$$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$0

where

$$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$1

The resulting Proximal Max-Min-Max algorithm operates directly through subgradient and projection oracles, has oracle complexity $$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$2 in the non-smooth case and $$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$3 when all functions are smooth, and extends to projection-unfriendly uncertainty sets via dualization (Tu et al., 2024). The reported numerical experiments show that it outperforms reformulation, cutting-plane, and two recent first-order baselines on high-dimensional and highly nonlinear robust optimization problems (Tu et al., 2024).

This body of work treats max-min algorithms as adversarial design procedures: the algorithm is optimized not for typical scenarios, but for the most expensive or most restrictive admissible scenario.

5. Learning, bandits, and Bayesian worst-case search

In stochastic learning, max-min grouped bandits formalize a structured identification problem in which arms belong to possibly overlapping groups and the objective is to return the group whose worst arm has the highest mean reward (Wang et al., 2021). Two algorithms are analyzed: a successive elimination procedure that maintains candidate worst arms and candidate optimal groups via confidence bounds, and a robust-optimization-based method adapting StableOpt. The paper derives upper bounds on the number of samples needed for exact or near-exact identification, as well as an algorithm-independent lower bound, while emphasizing that overlapping groups and inter-group information leakage make uniformly tight bounds difficult (Wang et al., 2021).

For expensive black-box objectives, Bayesian optimization has been extended from scalar maximization to min-max robust design. The formulation

$$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$4

is handled by extending Entropy Search and Knowledge Gradient to the min-max setting, producing single-acquisition strategies rather than nested acquisition functions (Weichert et al., 2021). The evaluation reports that Entropy Search and Knowledge Gradient outperform the GP-UCB-based approach on all test problems; Knowledge Gradient converges fastest initially but may stick to local optima, whereas Entropy Search is more reliable on challenging cases such as the modified eggholder benchmark (Weichert et al., 2021).

Although these problems are sequential rather than static, they preserve the central max-min logic: learning effort is directed toward certifying the weakest link, the worst environmental condition, or the most fragile group structure.

6. Semidefinite, parallel, and global methods, and known limitations

For polynomial and rational min-max problems, a semidefinite-programming approach constructs hierarchies of moment relaxations. The framework covers minimizing the supremum of finitely many rational functions over compact basic semi-algebraic sets and two-player zero-sum polynomial games in randomized strategies, with convergence of the SDP hierarchy to the true value and finite convergence detectable through flat-extension rank conditions (0810.3150). Each relaxation can be solved in time polynomial in its input size, and the method provides a unified route to approximate Nash equilibria and min-max strategies for a range of static and dynamic games (0810.3150).

Parallel approximation has also been achieved for a class of min-max problems using the matrix multiplicative weights update method. The algorithm handles transcript-like consistency constraints on lists of semidefinite matrices, gives near-optimal strategies for competitive two-party classical or quantum interactions with adaptive responses, and implies that classes including QRG(2), SQG, DIP, and DQIP collapse to PSPACE; in a special case it yields a direct polynomial-space simulation establishing $$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$5 (Gutoski et al., 2010).

Global methods have recently targeted non-convex or non-concave regimes. EXOTIC reformulates convex-non-concave min-max problems into a non-concave-convex max-min problem over $$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$6, solves the inner minimization approximately by convex optimization, and performs optimistic hierarchical tree search on the outer maximization (Maheshwari et al., 17 Aug 2025). The paper establishes deterministic upper bounds on the optimality gap and reports that EXOTIC outperforms gradient-based methods on both newly proposed and existing benchmarks (Maheshwari et al., 17 Aug 2025).

At the same time, a cautionary dynamical-systems analysis shows that a broad class of first- and zeroth-order min-max algorithms have limit sets contained in internally chain transitive sets of a mean-field dynamical system, that they are attracted to attractors of that system with high probability, and that they avoid unstable invariant sets with probability $$\begin{aligned} \text{maximize}\qquad & w=\min_{k\in K}\sum_{v\in V_k} c_{kv}x_v\ \text{subject to}\qquad & \sum_{v\in V_i} a_{iv}x_v\leq 1\quad \forall i\in I,\ & x_v\geq 0\quad \forall v\in V, \end{aligned}$$7—yet stable spurious attractors may contain no stationary points of the original problem (Hsieh et al., 2020). The paper concludes that existing min-max algorithms may be subject to inescapable convergence failures in non-convex/non-concave settings (Hsieh et al., 2020). This suggests a substantive division inside the field: local and first-order methods deliver scalability, while semidefinite, parallel, or global-search methods target certification against pathological dynamics.