Submodular Maximization

• Submodular Maximization is the process of optimizing set functions that exhibit a diminishing returns property, ensuring efficient approximations under combinatorial constraints.
• Greedy, threshold-based, and continuous algorithms (using multilinear relaxations) achieve tight approximation guarantees, such as (1-1/e) for monotone functions.
• Recent advances including sharper DR-submodular bounds and bicriteria methods enhance scalability and approximation accuracy in applications like sensor placement and influence maximization.

A submodular function $f:2^V\to\mathbb{R}$ is defined on the subsets of a finite ground set $V$ and satisfies the diminishing returns property: for all $A\subseteq B\subseteq V$ and $x\notin B$, $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$. Submodular maximization focuses on finding a subset $S\subseteq V$ maximizing $f(S)$, possibly subject to combinatorial constraints such as cardinality, matroid, or knapsack constraints, and is a core paradigm in combinatorial and discrete optimization with significant applications in machine learning, economics, and network design.

1. Mathematical Foundations and Problem Settings

Maximizing a submodular function is NP-hard in general, even for basic constraints, but its structure enables approximation algorithms with tight guarantees. The canonical objective is

$\max_{S\in\mathcal{I}} f(S)$

where $\mathcal{I} \subseteq 2^V$ encodes the feasible sets, often a family closed under subset operations (down-closed), e.g., uniform matroid ($|S|\leq k$), partition matroid ($V$0 has at most $V$1 elements in group $V$2), or knapsack (total weight $V$3). In unconstrained or cardinality-constrained settings, the greedy algorithm provides a $V$4-approximation for monotone $V$5 (Li et al., 2020, Fahrbach et al., 2018).

The multilinear extension $V$6, where $V$7 and $V$8 is a random subset where $V$9 is included independently with probability $A\subseteq B\subseteq V$0, is central for fractional relaxations and continuous optimization frameworks (Buchbinder et al., 2023). This extension is DR-submodular, inheriting the diminishing returns structure: $A\subseteq B\subseteq V$1 This permits the use of continuous algorithms for approximate maximization.

2. Key Algorithms and Approximation Guarantees

Greedy and Threshold-Based Methods

The classical greedy method iteratively adds the element with the largest marginal gain, achieving $A\subseteq B\subseteq V$2 approximation for monotone $A\subseteq B\subseteq V$3 under cardinality constraint, tight by known hardness (Li et al., 2020, Fahrbach et al., 2018). For non-monotone objectives, double greedy and random greedy techniques are employed, achieving $A\subseteq B\subseteq V$4 or $A\subseteq B\subseteq V$5-type approximation factors (Bian et al., 2016, Fahrbach et al., 2018). Deterministic threshold-based “clean linear time” algorithms matching this ratio with $A\subseteq B\subseteq V$6 queries have been developed (Li et al., 2020).

Continuous Greedy and Multilinear Relaxations

For matroid and more general constraints, the continuous greedy algorithm, operating on the multilinear extension, combined with swap or pipage rounding, achieves $A\subseteq B\subseteq V$7 for monotone $A\subseteq B\subseteq V$8 (Buchbinder et al., 2023, Bian et al., 2020). For general down-closed constraints and non-monotone $A\subseteq B\subseteq V$9, Buchbinder and Feldman established a $x\notin B$0-approximation via a sophisticated measured continuous greedy (Buchbinder et al., 2023). Recent advances have closed the gap towards the $x\notin B$1 hardness bound (Oveis Gharan & Vondrák), achieving a $x\notin B$2-approximation by leveraging a new full-history DR-submodularity bound (Buchbinder et al., 2023).

Advanced Exact and Branch-and-Bound Approaches

For moderate-scale instances, exact algorithms via branch-and-bound on complex binary integer programming formulations, augmented with batch constraint generation, have been developed to solve cardinality-constrained monotone submodular maximization to optimality, outperforming prior exact methods (Uematsu et al., 2018).

Streaming, Online, and Sliding Window Models

Streaming algorithms address computation over data streams, maintaining $x\notin B$3-approximations in a single pass with memory scaling in $x\notin B$4 (Kazemi et al., 2020). For the sliding window model (focus on most recent $x\notin B$5 items), reductions to streaming algorithms with only constant-factor loss and $x\notin B$6 overhead in time/space enable approximation for both monotone and non-monotone $x\notin B$7 under cardinality and matroid constraints (Chen et al., 2016).

Query Complexity and Adaptivity

Linear query complexity (exactly $x\notin B$8 queries) and nearly optimal adaptivity (parallel rounds) have been achieved by deterministic algorithms guaranteeing constant-factor approximations for monotone and general submodular maximization under matroid constraints (Balkanski et al., 2024). Adaptivity-optimal algorithms for the cardinality-constrained case achieve $x\notin B$9 with $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$0 rounds and $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$1 queries (Fahrbach et al., 2018).

3. Recent Theoretical Advances: New DR-Submodular Bounds

The classic “Feldman $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$2-bound” for the marginal gain loss along continuous trajectories underpins most continuous greedy analysis: $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$3 A new, substantially sharper bound by Buchbinder & Feldman for DR-submodular functions captures the cumulative “damage” across the entire path, yielding bounds of the form

$f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$4

This strengthens both the analysis and practical performance of continuous greedy frameworks: the first $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$5-approximation for maximizing a general (non-monotone) submodular function under down-closed combinatorial constraints is thus established (Buchbinder et al., 2023). This result closes the gap with the $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$6 inapproximability result, and the new bound has implications for other algorithmic frameworks using continuous relaxations.

4. Bicriteria Maximization and Constraint Relaxations

Bicriteria algorithms allow controlled constraint violation, e.g., slightly exceeding budget or cardinality constraints, to obtain improved approximation ratios unattainable under strict feasibility:

• Monotone $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$7: $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$8-bicriteria for cardinality/knapsack constraints via greedy or continuous methods.
• General (non-monotone) $f(A\cup\{x\})-f(A)\geq f(B\cup\{x\})-f(B)$9: $S\subseteq V$0-bicriteria for general down-closed polytopes using repeated continuous greedy and double greedy (Feldman et al., 14 Jul 2025).
• Symmetric $S\subseteq V$1: specialized variants reach $S\subseteq V$2-bicriteria. These relaxations inform both submodular cover and soft-constrained maximization and serve as benchmarks for streaming and parallel algorithms. Hardness results preclude improving the bicriteria factor below certain thresholds for general constraints (Feldman et al., 14 Jul 2025).

5. Streaming and Distributed Submodular Maximization

Efficient streaming and distributed submodular maximization frameworks address the scalability of submodular maximization in modern data-centric applications:

• One-pass streaming algorithms for regularized objectives $S\subseteq V$3 attain $S\subseteq V$4 (Kazemi et al., 2020).
• Distributed MapReduce-style algorithms run in $S\subseteq V$5 rounds and match the classical $S\subseteq V$6 guarantee for monotone $S\subseteq V$7 with $S\subseteq V$8 machine-local memory, improving previous memory bounds (Kazemi et al., 2020).
• Streaming algorithms for sliding windows derive from reductions of infinite-window algorithms, with deterministic $S\subseteq V$9-approximation for monotone $f(S)$0 and $f(S)$1 space (Chen et al., 2016).

6. Submodular Maximization in Continuous Domains and Advanced Extensions

Continuous submodular maximization generalizes set functions to $f(S)$2, with applications spanning influence maximization, DPP MAP inference, and mean-field inference in log-submodular models (Bian et al., 2020, Bian et al., 2016). The Frank-Wolfe variant for DR-submodular $f(S)$3 under convex constraints attains the tight $f(S)$4 bound, while double greedy yields $f(S)$5 for unconstrained non-monotone problems (Bian et al., 2020, Bian et al., 2016). These methods exploit the equivalence between lattice submodularity and weak DR-property, and provide dimension-independent guarantees that extend greedy methods to complex function classes.

7. Applications, Empirical Performance, and Open Problems

Applications of submodular maximization algorithms include data and video summarization, sensor placement, recommendation systems (balancing relevance and diversity), social influence, and robust coverage in networks. Empirical evaluations across real-world datasets demonstrate that threshold-based, continuous, and streaming algorithms achieve close to greedy and sometimes optimal performance with orders-of-magnitude fewer oracle calls (Li et al., 2020, Kazemi et al., 2020, Qin et al., 2023). Recent developments raise further questions:

• Can the gap between the best known and inapproximability factors for general submodular maximization (currently $f(S)$6 vs. $f(S)$7) be closed (Buchbinder et al., 2023)?
• Can adaptivity-optimal, high-approximation algorithms extend uniformly to non-monotone functions and more general constraints?
• Extensions to bicriteria, distributed, and online settings, along with partial monotonicity and compositionality of submodular functions, remain active research directions (Mualem et al., 2022, Feldman et al., 14 Jul 2025).