Greedy Maximum Coverage Algorithm

Updated 19 January 2026

Greedy maximum coverage algorithm is a combinatorial strategy that iteratively selects sets to maximize union coverage while ensuring polynomial-time efficiency.
It exploits submodularity and monotonicity to guarantee a 1-1/e approximation, with improved performance under specific structural conditions.
Extensions like Big Step Greedy and curvature-refined analysis broaden its applicability to fields such as active learning, computational geometry, and multi-agent systems.

The greedy maximum coverage algorithm is a fundamental combinatorial optimization strategy for the maximum coverage problem, which seeks to select a fixed number of sets from a collection to maximize the cardinality of their union. This algorithm is characterized by its iterative selection of the set(s) that cover the largest number of still-uncovered elements at each step—a process grounded in the principles of monotonicity and submodularity. The greedy algorithm has become the standard practical approach due to its polynomial-time complexity and its proven approximation guarantee, key results in submodular maximization, and broad applicability from computational geometry to multi-agent systems, discrete geometry, and machine learning.

1. Maximum Coverage Problem: Definitions and Complexity

Formally, the maximum $k$ -coverage problem is defined as follows. Given a finite universe %%%%1%%%% and a family of subsets $S = \{S_1, S_2, \dots, S_n\}$ , the goal is to identify a subfamily $C \subset S$ with $|C| = k$ such that the union $\cup_{S\in C} S$ is maximized, i.e.,

$C^* = \arg\max_{C \subseteq S, |C| = k} |\cup_{S\in C} S|.$

This objective is NP-hard; Feige demonstrated that, unless $\mathrm{P}=\mathrm{NP}$ , no polynomial-time algorithm can achieve a better approximation factor than $1-1/e$ in the worst case for arbitrary set systems (Badanidiyuru et al., 2011).

2. Classical Greedy Algorithm and Its Analysis

The classical greedy algorithm for maximum coverage proceeds in $k$ iterations. At each step, it selects the set covering the greatest number of currently uncovered elements. Denote the previously selected sets by $G_t$ at iteration $t$ . The decision rule is to pick

$S^* = \arg\max_{S_i \in S \setminus G_t} |S_i \setminus (\cup_{S \in G_t} S)|.$

This process exploits the monotonicity and submodularity of the coverage function $f(C) = |\cup_{S \in C} S|$ , which ensures diminishing returns as $C$ grows. The optimality analysis—originating with Nemhauser, Wolsey, and Fisher—yields an approximation ratio: $\frac{|G_k|}{|\mathrm{OPT}|} \geq 1 - \left(1 - \frac{1}{k}\right)^k \xrightarrow{k \rightarrow \infty} 1-\frac{1}{e}.$ This $(1-1/e)$ guarantee is tight for general instances (Badanidiyuru et al., 2011, Sun et al., 2017, Welikala et al., 2024).

3. Extensions: Big Step Greedy and Generalizations

A notable extension is the "Big Step Greedy" heuristic (Chandu, 2015). Rather than adding a single set at each step, it selects $p$ sets simultaneously (where $1 \leq p \leq k$ ), choosing the $p$ -subset whose union yields maximal incremental coverage. The pseudocode is as follows:

Input: S = {S₁, ..., Sₙ}, k, step size p
C ← ∅, Covered ← ∅
While |C| < k:
    q ← min(p, k–|C|)
    For each q-combination I ⊆ S\C:
        Evaluate union size |Covered ∪ (⋃_{S∈I} S)|
    Select I* with largest union
    C ← C ∪ I*
    Covered ← Covered ∪ (⋃_{S∈I*} S)
Output C

For $p=1$ , this reduces to the classical greedy algorithm; for $p=k$ , it tests all $k$ -subsets, behaving as a brute-force optimum. The Big Step variant interpolates between speed and solution quality, with empirical results indicating increased $p$ can yield significant average-case improvements, though worst-case guarantees remain at $1-1/e$ (Chandu, 2015).

4. Structural Conditions and Improved Approximation Bounds

The standard $1-1/e$ ratio can be improved if the set system exhibits additional structure. For instance, if every set has cardinality at most $r$ , or more generally, if the instance has covering multiplicity $r$ (every greedy choice can be "explained" by $r$ optimal sets), the greedy approximation ratio becomes

$1 - \left(1 - \frac{1}{r}\right)^r$

which can be significantly larger than $1-1/e$ for small $r$ (Badanidiyuru et al., 2011). In the specific case of sets defined by planar halfspaces ( $\mathbb{R}^2$ ), the multiplicity is $2$, and thus greedy achieves a tight $3/4$-approximation. However, in dimension four or higher, the lower bound reverts to $1-1/e$, and surpassing this is APX-hard (Badanidiyuru et al., 2011).

5. Curvature-Refined Performance and Submodularity

Recent studies in multi-agent coverage and active learning establish that submodularity implies greedy's worst-case $1-1/e$ bound, but tighter analysis exploits curvature metrics. Several curvature definitions (total, greedy, elemental, partial, and extended greedy curvature) allow for refined, instance-dependent performance bounds, sometimes approaching unity as curvature decreases (Sun et al., 2017, Welikala et al., 2024). The coverage function’s diminishing returns ensure monotonicity and submodularity, underpinning these guarantees.

Curvature Type	Definition (compact)	Approximation Guarantee
Total ( $\alpha_t$ )	$1-\frac{\Delta J(e\|X\setminus\{e\})}{\Delta J(e\|\emptyset)}$	$\beta_t = \frac{1}{\alpha_t}[1-(1-\frac{\alpha_t}{N})^N]$
Greedy ( $\alpha_g$ )	$1-\frac{\Delta J(e\|S^i)}{\Delta J(e\|\emptyset)}$	$\beta_g=1-\alpha_g (1-\frac{1}{N})$
Elemental ( $\alpha_e$ )	See data (Welikala et al., 2024)	Complex closed forms (see table)
Partial ( $\alpha_p$ )	$1-\frac{\Delta J(e\|S\setminus\{e\})}{\Delta J(e\|\emptyset)}$	$\beta_p$ similar to $\beta_t$
Extended ( $\alpha_u$ )	See greedy partitioning method (Welikala et al., 2024)	$\beta_u=J(S^{G})/\alpha_u$

Empirically, these refined bounds can reach $0.90$–$1.00$ for “weakly submodular” instances, far exceeding the general $1-1/e$ lower limit (Welikala et al., 2024).

6. Algorithmic Complexity and Implementational Aspects

The classical greedy algorithm computes, at each of $k$ steps, the marginal gain for $O(n)$ remaining sets, with each gain evaluated in $O(m)$ time, for $O(knm)$ total. The Big Step Greedy with step size $p$ evaluates up to $\binom{n}{p}$ combinations per step—rendering it practical only for small $p$ and moderate $n$ . For $p=k$ this becomes brute-force optimal enumeration (Chandu, 2015). In active learning with kernel-based objectives, maintaining and updating coverage arrays enables $O(kN)$ time per selection after an $O(N^2)$ kernel computation (Bae et al., 2024).

7. Applications and Empirical Performance

The greedy maximum coverage algorithm and extensions are central to many fields. Key applications include:

Active learning: Greedy selection of samples (“ProbCover,” “MaxHerding”) maximizes a surrogate coverage criterion directly connected to downstream classification error. MaxHerding generalizes the standard coverage algorithm via soft kernels, retaining the classical $(1-1/e)$ guarantee for monotone submodular objectives (Bae et al., 2024).
Geometric modeling: Multi-sphere particle approximation converts the clump construction problem in DEM into a greedy maximum coverage instance, leveraging the greedy guarantee for minimum set cover and ensuring mechanical fidelity through post-selection linear programming (Yuan, 2018).
Multi-agent systems: Agent placement for joint event detection admits a submodular greedy solution, with rigorous theoretical and empirical validation demonstrating substantial improvement using curvature-refined bounds and hybrid greedy-gradient approaches (Sun et al., 2017, Welikala et al., 2024).
Computational geometry: In set systems of low VC-dimension or bounded set cardinality, greedy can outperform its generic bound, showing tightness for particular geometric classes (Badanidiyuru et al., 2011).

Empirical findings indicate that modest increases in the step size $p$ for Big Step Greedy heuristics (e.g., $p=2,3,4$ ) often result in increased average coverage, with the hybrid approach (“best of $p=1,2,3,4$ ”) frequently outperforming both the standard greedy and randomized variants in practice, albeit at greater computational cost (Chandu, 2015).

In summary, the greedy maximum coverage algorithm occupies a central place in submodular optimization, offering both robust theoretical guarantees and considerable empirical efficacy. Its structural extensions, curvature-based analyses, and wide-ranging applications illustrate the continuing evolution of greedy methods in combinatorial optimization (Chandu, 2015, Badanidiyuru et al., 2011, Welikala et al., 2024, Bae et al., 2024, Yuan, 2018, Sun et al., 2017).