---
title: Covering Multiple Submodular Constraints
url: https://www.emergentmind.com/topics/covering-multiple-submodular-constraints
type: topic
---

# Covering Multiple Submodular Constraints

Covering multiple submodular constraints denotes a family of discrete optimization problems in which a subset must simultaneously satisfy several monotone submodular lower-bound requirements, usually at minimum cost. In a standard formulation, there is a finite ground set \(N\), a cost function \(c:N\to \mathbb{R}_+\), monotone submodular functions \(f_1,\ldots,f_r\), and requirements \(b_1,\ldots,b_r\); the goal is to find a minimum cost subset \(S\subseteq N\) such that \(f_i(S)\ge b_i\) for all \(i\in[r]\). When \(r=1\), this is the well-known Submodular Set Cover problem; when \(r\) is part of the input, approximation factors incur at least an \(\Omega(\log r)\) dependence, whereas for fixed \(r\) one can recover guarantees close to the single-constraint case [2507.09879].

## 1. Core formulations

The basic multiple-constraint model is the problem called **Multiple Submodular Covering Constraints** (MSC):
\[
\min_{S\subseteq N} c(S)\quad\text{s.t.}\quad f_i(S)\ge b_i \qquad \forall i\in[r],
\]
where \(c(S)=\sum_{e\in S} c(e)\) [2507.09879]. This formulation isolates the simultaneous satisfaction of several monotone submodular requirements and is the most direct formalization of “covering multiple submodular constraints.”

Several important variants refine this template rather than replace it. In the fully dynamic setting, there is an active family \(G^{(t)}\) of monotone, nonnegative submodular functions at time \(t\), and the current aggregate covering objective is
\[
f^{(t)}=\sum_{g\in G^{(t)}} g.
\]
The task is to maintain a near-optimal cover for \(f^{(t)}\) while keeping the recourse \(|S_t\triangle S_{t+1}|\) small as functions are inserted or deleted [2009.00800].

A robust constrained formulation appears in **Robust-scsc** and **Robust-scsk**. The former minimizes a worst-case submodular objective subject to multiple submodular lower-bound constraints,
\[
\min_{X\subseteq V}\max_{i=1:l} f_i(X)\ \Big|\ g_i(X)\ge c_i,\ i=1,\dots,k,
\]
while the latter maximizes a worst-case submodular utility subject to multiple submodular upper-bound constraints,
\[
\max_{X\subseteq V}\min_{i=1:k} g_i(X)\ \Big|\ f_i(X)\le b_i,\ i=1,\dots,l.
\]
These forms explicitly separate “cost” and “coverage” roles across multiple submodular functions [1906.06393].

A related single-objective, single-submodular-constraint pair is given by **Submodular Cost Submodular Cover** (SCSC),
\[
\min_{X \subseteq V} f(X) \quad \text{s.t.} \quad g(X) \ge c,
\]
and **Submodular Cost Submodular Knapsack** (SCSK),
\[
\max_{X \subseteq V} g(X) \quad \text{s.t.} \quad f(X) \le b.
\]
These models are central because approximation algorithms for one can be transformed into approximation guarantees for the other [1311.2106].

## 2. Approximation regimes and hardness structure

The approximation picture depends sharply on how many constraints are present and how they are represented. For MSC with arbitrary \(r\), the approximation ratios are at least \(\Omega(\log r)\), and this dependence is unavoidable when \(r\) is part of the input [2507.09879]. This barrier explains why the fixed-\(r\) regime is treated separately in recent work.

In robust multi-constraint formulations, hardness is stronger than a logarithmic loss alone. For the multiple-constraint problems Robust-scsc and Robust-scsk, bicriteria lower bounds depend on curvature and covering complexity, and **Robust-scsk does not admit a single-criterion approximation when \(k>1\) unless \(P=NP\)** [1906.06393]. For SCSC and SCSK, hardness is also curvature-sensitive: for any \(\kappa>0\), there exist submodular functions of curvature \(\kappa\) such that no polynomial-time algorithm can achieve a bi-criterion ratio better than
\[
\frac{\sigma}{\rho} = \frac{n^{1/2-\epsilon}}{1 + (n^{1/2-\epsilon}-1)(1-\kappa)}
\qquad \text{for any }\epsilon>0,
\]
and the known guarantees are tight up to log-factors [1311.2106].

The dynamic setting introduces a different form of impossibility. For fully dynamic Submodular Cover with \(f^{(t)}=\sum_{g\in G^{(t)}} g\), there is a deterministic algorithm with \(O(\log(f_{\max}/f_{\min}))\)-competitiveness, and this competitive ratio is best possible even in the offline setting; the recourse bound is optimal up to the logarithmic factor [2009.00800].

A broader precursor is **Submodular-Cost Covering**, where the objective is a non-decreasing submodular cost function and the feasible region is the intersection of arbitrary covering constraints, each depending on at most \(D\) variables. In that model, a simple greedy procedure is a \(D\)-approximation algorithm, including the online setting where the same greedy rule is \(D\)-competitive [0807.0644]. This earlier line concerns multiple covering constraints with submodular cost rather than multiple submodular covering functions, but it established an important approximation template for structured covering systems.

## 3. Algorithmic paradigms

A common algorithmic template for MSC with fixed \(r\) has four stages: **guess a small set of high-cost elements** from an optimal solution; **solve a fractional relaxation** of the residual instance; **round the fractional solution** using randomized rounding plus a concentration bound; and **greedily fix** any still-unsatisfied constraints [2507.09879]. The multilinear relaxation uses
\[
F_i(x)=\mathbb{E}[f_i(R)],
\]
where \(R\) contains each element independently with probability \(x_e\), together with constraints of the form
\[
\sum_e c(e)x_e \le B,\qquad F_i(x)\ge b_i\ \forall i.
\]
The key technical device is a rounding lemma that combines greedy preprocessing with a Lipschitz reduction and Vondrák-style concentration:
\[
\Pr\big[f(R)\le (1-\delta)F(x)\big] \le \exp\!\left(-\frac{\delta^2F(x)}{2\ell}\right)
\]
for monotone submodular \(\ell\)-Lipschitz functions under independent rounding [2507.09879].

The fully dynamic framework uses a markedly different representation. It maintains a permutation \(\pi\) of the ground set, interprets the solution as the prefix of elements with positive marginal contribution, and repeatedly applies local **swaps** and **\(\gamma\)-moves** until the permutation is stable. The recourse analysis is driven by a potential function inspired by Tsallis entropy,
\[
\Phi_\alpha(f,\pi)=\sum_{i\in N}\left(F_\pi(\pi_i)\right)^\alpha,\qquad \alpha\in(0,1),
\]
together with the notion of **Mutual Coverage**
\[
I_f(A;B)=f(A)+f(B)-f(A\cup B).
\]
For 3-increasing functions, this yields the stronger \(O(\sum_t g_t(N))\) total recourse bound [2009.00800].

For SCSC and SCSK, the main algorithmic paradigm is iterative surrogate optimization. One replaces a submodular cost by modular upper bounds \(m^f_{X^t}\), or by ellipsoidal approximations such as
\[
f^{\mathrm{ea}}(X)= \mathcal f \sqrt{w^{f^\kappa}(X)} + (1-\mathcal f)\sum_{j\in X} f(j),
\]
and then solves a sequence of easier subproblems. This produces algorithms such as **Iterated Submodular Set Cover** and **Iterated Submodular Knapsack**, with guarantees that improve as the cost function curvature decreases [1311.2106].

Robust multi-constraint optimization uses a parallel surrogate philosophy. The main methods are **average approximation**, **majorization-minimization (mmin)**, **ellipsoidal approximation (ea)**, and convex relaxation through the Lovász extension. In mmin, each submodular objective \(f_i\) is replaced by a modular upper bound \(m^{f_i}_{X^t}\), and the next iterate solves
\[
X^{t+1} = \arg\min_{X \in \mathcal C} \max_i m^{f_i}_{X^t}(X).
\]
This converts robust submodular objectives into min-max combinatorial problems with modular costs [1906.06393].

## 4. Fixed-\(r\) covering and exact-feasibility special cases

The fixed-number-of-constraints regime is the clearest positive result for covering multiple submodular constraints. For any fixed integer \(\alpha\ge 1\) and any \(\epsilon>0\), there is a randomized bicriteria approximation for MSC that outputs \(S\subseteq N\) such that
\[
f_i(S)\ \ge\ \left(1-\frac{1}{e^\alpha}-\epsilon\right)b_i \qquad \forall i\in[r],
\]
and
\[
\mathbb{E}[c(S)] \le (1+\epsilon)\alpha\cdot \mathrm{OPT}.
\]
This is essentially the same cost–coverage tradeoff known for the single-constraint case [2507.09879].

For the special case where the \(f_i\) are **weighted coverage functions** from a **deletion-closed set system**, the guarantee strengthens to an exact-feasibility approximation of
\[
(1+\epsilon)\left(\frac{e}{e-1}\right)(1+\beta),
\]
where \(\beta\) is the approximation ratio for the underlying set cover instance obtained via the natural LP relaxation [2507.09879].

| Setting | Guarantee | Notes |
|---|---|---|
| General MSC, fixed \(r\) | \(f_i(S)\ge (1-1/e^\alpha-\epsilon)b_i\), \(\mathbb{E}[c(S)]\le (1+\epsilon)\alpha\cdot \mathrm{OPT}\) | Randomized bicriteria |
| Weighted coverage, deletion-closed set system | \((1+\epsilon)(e/(e-1))(1+\beta)\) | Exact-feasibility approximation |

These results extend directly to several application families. The paper explicitly mentions **Colorful Set Cover**, **Facility Location with Multiple Outliers**, and **Sum of Radii with Multiple Outliers**, and notes that for fixed \(r\) the same framework yields constant-factor approximations in these settings [2507.09879]. The structural point is that the number of submodular constraints, not merely their form, determines whether one can match single-constraint behavior.

## 5. Fairness, partition structure, and robust coverage

Fairness and balance constraints introduce another layer of multiplicity. In **Submodular Cover with Partition Constraints (SCP)**, the ground set is partitioned into disjoint groups
\[
U = U_1 \uplus \cdots \uplus U_N,
\]
and one seeks
\[
\min_{S \subseteq U} \ v \quad \text{s.t.} \quad f(S)\ge \tau,\qquad |S\cap U_j|\le p_j v,\ \forall j\in[N].
\]
The corresponding fairness-constrained variant **SCF** requires both lower and upper balance across groups:
\[
\min_{S\subseteq U} |S| \quad \text{s.t.}\quad p_c|S|\le |S\cap U_c|\le q_c|S|,\ \forall c\in[N],\qquad f(S)\ge \tau.
\]
These models do not introduce multiple submodular functions; instead, they couple submodular cover with multiple partition-based resource constraints [2601.11755].

The main algorithmic strategy is a converting framework: first solve the dual maximization problem approximately, then use it to drive a cover algorithm. For monotone SCKP, the block-greedy subroutine **greedy-knapsack-bi** achieves
\[
f(S)\ge (1-\epsilon)f(OPT),
\]
with per-partition cost blow-up
\[
c(S\cap U_j)\le \frac{2\ln(1/\epsilon)}{\ln 2}\, p_j v,\qquad \forall j,
\]
and after conversion the cover solution satisfies
\[
v_S \le \frac{2(1+\alpha)\ln(1/\epsilon)}{\ln 2}\, v_{OPT}, \qquad f(S)\ge (1-\epsilon)\tau
\]
[2601.11755].

For fairness-constrained cover, **Block-Fair-Bi** achieves
\[
f(S)\ge (1-\epsilon)f(OPT)
\]
with bicriteria factor
\[
\left(1-\epsilon,\ \frac{\ln(1/\epsilon)}{\ln 2}\right),
\]
and after conversion this yields an SCF guarantee of
\[
\left(O(\ln(1/\epsilon)),\ 1-\epsilon\right).
\]
The same work reports a query-complexity improvement from
\[
O\!\left(\frac{n^2(1+\alpha)\log^2(n/\varepsilon)\log n}{\varepsilon^4\alpha}\right)
\]
to
\[
O\!\left(\frac{n\log(n)\kappa\ln(1/\epsilon)}{\epsilon}\right)
\]
for the fair-cover setting [2601.11755].

Robustness provides a different fairness-related perspective. In image collection summarization with multiple queries, the robust formulation
\[
\min_{X \subseteq V} \max_{i=1:l} h(X \mid Q_i) \quad \Big| \quad g(X) \ge c
\]
is exactly a Robust-scsc problem with \(k=1\) and multiple submodular lower-bound constraints indexed by queries. The reported interpretation is that the robust formulation is much fairer across queries than averaging or union-query baselines [1906.06393].

## 6. Applications, related models, and conceptual boundaries

The multiple-submodular-cover viewpoint captures several concrete application domains. The fixed-\(r\) MSC framework is applied to **Colorful Set Cover**, **Facility Location with Multiple Outliers**, and **Sum of Radii with Multiple Outliers** [2507.09879]. Robust constrained formulations are motivated by **robust data subset selection**, **robust co-operative cuts**, **robust co-operative matchings**, and **image collection summarization with multiple queries** [1906.06393]. The dynamic aggregate-cover model subsumes fully dynamic **Set Cover / Hitting Set**, where each arriving or departing hyperedge corresponds to a coverage indicator \(g_t(S)=\mathbbm{1}[S\cap A_t\neq \emptyset]\), yielding the familiar \(O(\log n)\)-competitive and \(O(1)\)-recourse behavior [2009.00800].

Several adjacent topics are often conflated with covering multiple submodular constraints but are technically distinct. **Packing-Covering Submodular Maximization (PCSM)** maximizes a monotone submodular objective subject to linear packing and covering constraints,
\[
\max \left\{ f(S):S\subseteq N,\; P \mathbf{1}_{S}\leq \mathbf{1}_p,\; C \mathbf{1}_{S}\geq \mathbf{1}_c\right\},
\]
and extends to **multiple monotone submodular objectives** through approximate Pareto-set guarantees [1804.10947]. Likewise, bicriteria submodular maximization under cardinality, knapsack, matroid, and convex-set constraints studies a single submodular objective with controlled feasibility violation rather than the simultaneous covering of several submodular lower bounds [2507.10248]. These models are closely related in technique—continuous relaxations, randomized rounding, greedy repair, and bicriteria analysis—but they address different optimization geometries.

A further boundary concerns objective-versus-constraint placement. The SCSC/SCSK framework was motivated precisely by the observation that applications such as sensor placement and data subset selection require maximizing one submodular function while simultaneously minimizing another, and that phrasing the problem as constrained optimization is more natural than minimizing a difference of submodular functions [1311.2106]. In this sense, modern work on multiple submodular covering constraints sits at the intersection of submodular cover, robust optimization, and fairness-aware discrete optimization: the common theme is simultaneous satisfaction of several structured requirements, while the main technical divide is whether multiplicity appears in the lower-bound constraints, in the resource system, in the objective, or across time.

Source: https://www.emergentmind.com/topics/covering-multiple-submodular-constraints