Bicriteria Algorithms for Submodular Cover with Partition and Fairness Constraints
Extracted source text · arXiv 2601.11755
\documentclass{article}
\usepackage[margin=1in]{geometry}
% if you need to pass options to natbib, use, e.g.:
% \PassOptionsToPackage{numbers, compress}{natbib}
% before loading neurips_2025
% ready for submission
%\usepackage{iclr2026_conference,times}
% to avoid loading the natbib package, add option nonatbib:
% \usepackage[nonatbib]{neurips_2025}
%\usepackage[utf8]{inputenc} % allow utf-8 input
%\usepackage[T1]{fontenc} % use 8-bit T1 fonts
\usepackage{hyperref} % hyperlinks
\usepackage{url} % simple URL typesetting
\usepackage{booktabs} % professional-quality tables
\usepackage{amsfonts} % blackboard math symbols
\usepackage{nicefrac} % compact symbols for 1/2, etc.
%\usepackage{microtype} % microtypography
\usepackage{xcolor} % colors
\usepackage{natbib}
\title{Bicriteria Algorithms for Submodular Cover with Partition and Fairness Constraints}
\author{Wenjing Chen\thanks{Texas A\&M University, \texttt{jj9754@tamu.edu}, \texttt{shuoxing@tamu.edu}, \texttt{samsonzhou@gmail.com}, \texttt{vcrawford@tamu.edu}} \And Shuo Xing\samethanks \And Samson Zhou\samethanks \And Victoria G. Crawford\samethanks}
\author{%
Wenjing Chen \\
\texttt{jj9754@tamu.edu} \\
Computer Science \& Engineering\\
Texas A\&M University\\
\and
Yixin Chen \\
\texttt{chen777@tamu.edu}\\
Computer Science \& Engineering\\
Texas A\&M University\\
\and
Victoria G. Crawford \\
\texttt{vcrawford@tamu.edu} \\
Computer Science \& Engineering\\
Texas A\&M University\\
% examples of more authors
%\And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \AND
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
}
\date{}
% If accepted, instead use the following line for the camera-ready submission:
% \usepackage[accepted]{icml2025}
% For theorems and such
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{mathtools}
\usepackage{amsthm}
% if you use cleveref..
%\usepackage[capitalize,noabbrev]{cleveref}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% THEOREMS
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage{xcolor}
\usepackage{subfiles}
% \usepackage{authblk}
% % \usepackage[margin=1.25in]{geometry}
\usepackage{times}
\usepackage{helvet}
\usepackage{courier}
% \usepackage[hyphens]{url}
\usepackage{graphicx}
\usepackage{natbib}
\usepackage{enumerate}
\bibliographystyle{unsrtnat}
\usepackage{caption}
\usepackage{soul}
%\usepackage[utf8]{inputenc}
% \usepackage{courier}
% \usepackage[hyphens]{url}
% \theoremstyle{plain}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{assumption}[theorem]{Assumption}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{example}{Example}
\usepackage{booktabs}
\usepackage{algorithm}
% \usepackage{algorithmic}
\usepackage[noend]{algpseudocode}
\usepackage{thm-restate}
\usepackage{enumitem}
\usepackage{subfigure}
\usepackage{amsfonts}
\usepackage{nicefrac}
%\usepackage{microtype}
\usepackage{xspace}
% Todonotes is useful during development; simply uncomment the next line
% and comment out the line below the next line to turn off comments
%\usepackage[disable,textsize=tiny]{todonotes}
\usepackage[textsize=tiny]{todonotes}
\newcommand{\scpklong}{Submodular Cover under Partition Knapsack constraint\xspace}
\newcommand{\scpk}{SCPK\xspace}
\newcommand{\smpk}{SMPK\xspace}
\newcommand{\convk}{\texttt{convert}\xspace}
\newcommand{\problong}{Submodular Maximization over Partition matroid\xspace}
\newcommand{\prob}{SMP\xspace}
\newcommand{\argmin}{\text{argmin}}
\newcommand{\argmax}{\text{argmax}}
\newcommand{\pargreedy}{\texttt{partition-greedy}\xspace}
\newcommand{\chunkgreedy}{\texttt{Block-Greedy}\xspace}
\newcommand{\blockmono}{\texttt{Greedy-Subroutine-Mono}\xspace}
\newcommand{\blocknonmono}{\texttt{Greedy-Subroutine-Nonmono}\xspace}
\newcommand{\nonmonosc}{\texttt{nonmono-bi}\xspace}
\newcommand{\algksm}{\texttt{Alg-SM}\xspace}
\newcommand{\blockksc}{\texttt{greedy-knapsack-bi}\xspace}
\newcommand{\greedy}{\texttt{Greedy}\xspace}
\newcommand{\chunkgreedynonmono}{\texttt{chunk-greedy-nonmono}\xspace}
\newcommand{\pargreedynew}{\texttt{partition-greedy-nonuniform}\xspace}
\newcommand{\fairgreedy}{\texttt{Block-Fair-Bi}\xspace}
\newcommand{\algsub}{\texttt{Greedy-Subroutine}\xspace}
\newcommand{\proofsketch}{\vspace*{-1ex} \noindent {\bf Proof Sketch: }}
\newcommand{\ksml}{The Knapsack Constrained Submodular Maximization Problem\xspace}
\newcommand{\ksm}{KSM\xspace}
\newcommand{\csml}{The Cardinality Constrained Submodular Maximization Problem\xspace}
\newcommand{\csm}{CSM\xspace}
\newcommand{\sckl}{The Knapsack Cost Submodular Cover Problem\xspace}
\newcommand{\ksc}{KSC\xspace}
\newcommand{\algksc}{Block-G-knapsack\xspace}
% \newcommand{\prob}{SMP\xspace}
\newcommand{\mE}{\mathbb{E}}
\newcommand{\vect}{\textbf}
\newcommand{\aksc}{\texttt{KSC-Sto-Alg}}
\newcommand{\cmin}{c_{\text{min}} }
\newcommand{\cmax}{c_{\text{max}} }
\newcommand{\filter}{\texttt{filter}\xspace}
\newcommand{\stocover}{\texttt{Stochastic Greedy for Submodular Cover (SGSC)}\xspace}
\newcommand{\tg}{\texttt{threshold-bi}\xspace}
\newcommand{\inprod}[2]{\langle #1, #2\rangle}
\newcommand{\convr}{\texttt{convert-rand}\xspace}
\newcommand{\wenjing}[1]{{\color{blue}#1}}
\begin{document}
\maketitle
\begin{abstract}
In many submodular optimization applications, datasets are naturally partitioned into disjoint subsets. These scenarios give rise to submodular optimization problems with partition-based constraints, where the desired solution set should be in some sense balanced, fair, or resource-constrained across these partitions. While existing work on submodular cover largely overlooks this structure, we initiate a comprehensive study of the problem of Submodular Cover with Partition Constraints (SCP) and its key variants. Our main contributions are the development and analysis of scalable bicriteria approximation algorithms for these NP-hard optimization problems for both monotone and nonmonotone objectives. Notably, the algorithms proposed for the monotone case achieve optimal approximation guarantees while significantly reducing query complexity compared to existing methods.
Finally, empirical evaluations on
real-world and synthetic datasets further validate the efficiency and effectiveness
of the proposed algorithms.
\end{abstract}
\input{Sections/Introduction}
\input{Sections/Preliminaries}
\input{Sections/RelatedWork.tex}
\input{Sections/problemsetup}
\input{Sections/Algorithms.tex}
% \input{Sections/hardnessResult}
% \input{Sections/nonmonotone}
% \input{Sections/FairnessCover}
% \input{Sections/knapsack}
\input{Sections/experiments}
\section{Reproducibility statement}
All theoretical results in this paper are supported by complete proofs, which are provided in the main text and the appendix. Approximation guarantees and detailed proofs are described to ensure that the theoretical contributions can be independently verified.
\bibliography{paper}
% \bibliographystyle{icml2025}
% \input{Sections/CheckList}
\input{Sections/appendix}
% \include{Sections/appendix}
\end{document}
\section{Introduction}
\label{sec:intro}
Submodular optimization algorithms have emerged as a cornerstone of modern machine learning, driving advancements across a range of impactful applications. From curating high-quality pretraining and fine-tuning datasets for large language models \citep{ji2024submodular, kumari2024end, agarwal2024delift} to powering diversified online recommendation systems \citep{hiranandani2020cascading, chen2024linear}, multi-agent optimization in robotics \citep{zhou2022risk, xu2024performance}, and enabling precise image attribution in computer vision \citep{chen2024less}.
%From the application of selecting high-quality and relevant pretraining data in Large Language Model training tasks \cite{ji2024submodular, kumari2024end} to other high-impact applications such as diversified online-recommendation system \cite{yue2011linear, hiranandani2020cascading, chen2024linear} and computer vision \cite{chen2024less}, submodular optimization algorithms are becoming increasingly prevalent in technologies and decision-making processes in the modern world.
Submodular functions informally satisfy a diminishing returns property that is exhibited by many objective functions for fundamental optimization problems in machine learning. Formally, let $f:2^U\to\mathbb{R}$ be defined over subsets of a ground set $U$ of size $n$. Then the function $f$ is \textit{submodular} if for all $A\subseteq B\subseteq U$ and $x\notin B$, $f(A\cup\{x\})-f(A) \geq f(B\cup\{x\})-f(B)$. Further, $f$ is \textit{monotone} if $f(Y)\ge f(X)$ for every $X\subseteq Y\subseteq U$.
The submodular cover (SC) problem is an important optimization problem with a variety of applications \citep{iyer2013submodular,chen2024bicriteria,crawford2019submodular,mirzasoleiman2015distributed}.
%In nearly all of the current literature,
In the classical form, the goal of submodular cover is to find a subset $S\subseteq U$ of minimum cost such that $f(S)\geq\tau$, where the cost function is typically cardinality or some additive cost. Existing results on SC take advantage of its relationship with submodular maximization \citep{iyer2013submodular,chen2024bicriteria}, which is to find $\arg\max\{f(S): c(S)\leq\kappa\}$. For example, \citep{chen2024bicriteria, iyer2013submodular} proposed converting algorithms that could convert any bicriteria algorithm for submodular maximization to an algorithm for SC. In particular, an $(\alpha,\beta)$-bicriteria approximation algorithm for the SC problem returns a solution $X$ such that $|X|\leq\alpha |OPT|$ and $f(X)\geq\beta\tau$.
However, a significant limitation of these classical formulations is their inability to model critical applications where the ground set $U$ is partitioned into disjoint groups $U_1,\ldots, U_N$, and the objective is to find a subset that has a budget within each partition, or alternatively is balanced or fair across the partitions. We further illustrate the submodular cover with partition constraints setting with several applications.
In video summarization \citep{mirzasoleiman2018streaming}, the elements of $U$ are frames that are each associated with one of $N$ consecutive regions of time in the video. A submodular function $f$ is formulated to measure how effectively a subset of frames $X$ summarizes the entire video $U$ \citep{tschiatschek2014learning}. The goal is to find a subset of frames that is a sufficiently good summary, i.e. $f(X)\geq\tau$, while limiting the proportion of frames from each time region in the solution, i.e., $c(X\cap U_i)\leq p_i v$ where $p_j\in[0,1]$ and $v$ is the budget on the cost.
As a second example, consider influence maximization \citep{tschiatschek2014learning}, where the ground set of users $U$ may be divided into $N$ partitions, reflecting demographics information such as language. In order to choose a subset with balanced distribution across different partitions, we enforce the fairness constraint where $p_j|S|\leq|S\cap U_j|\leq q_j|S|$. Then the objective is to find a fair solution with minimum cardinality such that $f(S)\geq\tau$.
% A final example application is deep neural network pruning or quantization where $U$ corresponds to the neurons in the network that belong to $N$ types that depend on the structure of the neural network.
Many further applications exist in the literature, including neural network pruning \citep{chen2024fair}, high-quality data selection for learning \citep{killamsetty2021glister}, and data summarization \citep{el2020fairness}.
Despite the importance and widespread applications of this problem, prior work remains limited. \cite{chen2024fair} studied a special case of the submodular cover with constraints defined on partitions, which is the fairness constraints. However, their proposed discrete method attains only a suboptimal approximation ratio, while the continuous approach incurs prohibitively high query complexity. In contrast, our approach achieves optimal bicriteria approximation ratios with significantly lower query complexity on the problem of SC with fairness constraint.
In this work, we study several distinct submodular cover problems with constraints defined on partitions of the universe $U$, including but not limited to the submodular cover with fairness constraints.
Our approach follows the general converting framework by developing converting algorithms that can convert submodular maximization algorithms into submodular cover algorithms. In particular, to construct solutions with objective values closer to the target threshold $\tau$, we propose bicriteria algorithms for submodular maximization with partition constraints. Notably, unlike traditional greedy algorithms that add one feasible element at a time based on marginal gain, our method incrementally selects blocks of elements in each round, where each block respects the cost distribution across different partitions to ensure that elements are selected proportionally to the budget cost.
This block-greedy strategy is particularly effective in the submodular cover setting, where achieving values close to the threshold $\tau$ may require selecting sets that exceed the feasibility limits of standard submodular maximization algorithms.
In particular, our contributions are summarized as follows.
\begin{enumerate}
\item In Section \ref{sec:nonmono_sc}, we study the Submodular Cover with Partition Constraint (SCP) problem of $\arg\min_{S\subseteq U}\{v: f(S)\geq\tau, |S\cap U_j|\leq p_jv, \forall j\in[N]\}$ in the case where $f$ is nonmonotone. We first propose a general converting algorithm to convert any randomized algorithms for the dual problem of Submodular Maximization with Partition
constraint (SMP), into an algorithm for SCP.
By proposing a bicriteria algorithm for SMP, we can obtain an algorithm for nonmonotone SCP with a bicriteria approximation ratio of $(O(\frac{(1+\alpha)}{\epsilon}), 1/e-\epsilon)$.
\item Section \ref{sec:knapsack} addresses the problem of Monotone Submodular Cover with Knapsack Partition Constraints
(SCKP), which is to find $\arg\min_{S\subseteq U}\{v: f(S)\geq\tau, c(S\cap U_j)\leq p_jv, \forall j\in[N]\}$.
We first develop an algorithm for the dual optimization problem of Submodular Maximization under Knapsack Partition Constraints
(SMKP), which adopts the block-greedy structure. By utilizing a converting procedure, we achieve a $(\frac{(1+\alpha)\ln1/\epsilon}{\ln2},1-\epsilon)$ bicriteria-approximation ratio for SCKP.
\item Section \ref{sec:fair} considers the monotone Submodular Cover problem with Fairness Constraint (SCF), which was recently introduced by \cite{chen2024fair}. The proposed algorithm achieves the nearly optimal approximation ratio of $(\mathcal{O}(\ln(1/\epsilon)),1-\epsilon)$. This matches the approximation ratio for the algorithm of \citet{chen2024fair}, but their method is continuous and requires $\mathcal{O}(\frac{n^2(1+\alpha)\log^2(\frac{n}{\varepsilon})\log n}{\varepsilon^4\alpha})$ queries of $f$ while our method only requires a query complexity of $\mathcal{O}(\frac{n\log(n)\kappa\ln(1/\epsilon)}{\epsilon})$.
\end{enumerate}
Finally, we conduct an experimental evaluation of our algorithms for nonmonotone SCP, monotone SCKP, and monotone SCF. Our results demonstrate that our proposed algorithm for nonmonotone SCP achieves a higher function value compared to the baseline algorithms, and SCKP achieves an improvement in the budget of the cost. Additionally, our algorithm for SCF outperforms the other algorithms proposed in \cite{chen2024fair} in terms of the solution set size and fairness difference.
%\begin{enumerate}
%\item First of all, we propose an algorithm for monotone \problong (monotone \prob), achieving an approximation ratio of $1-e^{-1}-\frac{e^{-1}}{\phi+1}$, where $\phi=\sqrt{\min_{i\in[N]}k_i}-1$. Notably, as $\phi$ grows, the approximation ratio asymptotically approaches the best-known result of $1-e^{-1}$, which is especially relevant in large-scale problems. Additionally, to underscore the significance of our result, we provide a class of tight examples demonstrating that the standard greedy algorithm cannot exceed an approximation ratio of $1/2+\epsilon$ even when $\phi$ is large, thereby validating the non-triviality of our approach.
%\item Next, we propose an algorithm that follows the \chunkgreedy framework with the subroutine algorithm \blocknonmono for the problem of nonmonotone \prob, which achieves an approximation ratio of $1/e(1-\frac{3}{\phi+1})$ in expectation. In particular, when $\phi\geq18$, the approximation ratio surpasses the best-known result of $0.305-\epsilon$ for discrete algorithms, as established in \cite{chen2024discretely}.
%\end{enumerate}
%Apart from the submodular maximization (SM) problems, we can also show the potential of our algorithmic framework for solving submodular cover problems.
%\begin{enumerate}
%\item We consider the {\scpklong (\scpk)} problem, which is defined to find $\arg\min_{S\subseteq U}\{v: f(S)\geq\tau, c(S\cap U_j)\leq p_jv, \forall j\in[N]\}$. (see more description in Section \ref{sec:knapsack})
% Here $c$ is a modular cost penalty function. The objective is to find a solution set of minimum cost that satisfies $f(S)\geq\tau$ while balancing the cost across subgroups according to specified allocation of budgets.
%We start by proposing the first bicriteria algorithm for \smpk, which follows the \chunkgreedy framework and achieves the bicriteria approximation ratio of $(1-\epsilon, \frac{2\ln1/\epsilon}{\ln2})$. By developing a converting algorithm that can convert algorithms for the \smpk into ones for \scpk, we obtain an algorithm that achieves a $(\frac{2(1+\alpha)\ln1/\epsilon}{\ln2},1-\epsilon)$ bicriteria-approximation ratio.[TODO:update later.]
%\item Furthermore, we extend the idea of \chunkgreedy and propose an algorithm for the problem of Fair Submodular Cover, which is first introduced in \cite{chen2024fair}. The proposed algorithm achieves the optimal approximation ratio of $(\mathcal{O}(\ln(1/\epsilon)),1-\epsilon)$ but requires much fewer queries compared with the continuous method proposed in \cite{chen2024fair}.
%\end{enumerate}
% In this paper, we propose new algorithms for a series of Submodular Maximization problems over Partition Matroid. The main idea is to split the partition matroid into chunks. We refer to this kind of approach as "chunk greedy". In particular, suppose the ground set $U$ is divided into disjoint groups $U_1,\ldots,U_N$ such that $U=\cup_{i\in[N]}U_i$ and $U_i\cap U_j=\emptyset$ for any $i\neq j$. The partition matroid is defined as $\mathcal{M}:=\{S\subseteq U:|S\cap U_i|\leq k_i\}$.
% Finally, we apply the idea of "chunk greedy" to the problem of Fair Submodular Cover and demonstrate that the proposed algorithm achieves the optimal approximation ratio but requires much fewer queries compared with the continuous method proposed in \cite{chen2024fair}.
% \subsection{Preliminaries}
% \label{section:prelim}
% \textcolor{red}{Not sure all of these that are used, or if they are the right version for this paper, copied this from some of our other papers}
% We first provide a number of preliminary definitions that will be used throughout the paper:
% % An $(\alpha,\beta)$-bicriteria approximation algorithm for a submodular maximization returns a solution $X$ such that $f(X)\geq \alpha f(OPT)$ and $|X|\leq\beta\kappa$. Notice that in the $(\alpha,\beta)$ notation, the approximation on the objective is first, and the approximation on the constraint is second. \textcolor{red}{Need to mention or add in partition constraint.}
% (vii) The marginal gain of adding an element $u\in U$ to a set $S\subseteq U$ is denoted as $\Delta f(S,u)=f(S\cup u)-f(S)$;
% % (viii) The function $f_{\tau}=\min\{f,\tau\}$.
% (i) \textcolor{red}{Definition of submodular max with partition}
% (iv) $OPT$ is used to refer to the optimal solution to the instance of submodular cover that should be clear from the context;
% (v) $OPT_{SM}$ is used to refer to the optimal solution to the instance of submodular maximization that should be clear from the context;
\subsection{Related Work}
% In particular, algorithms for submodular maximization with a matroid constraint are studied for these problems \cite{lee2010maximizing,ene2019submodular}, since many partition constraints either are a special case of a matroid (e.g. a partition matroid) or can be converted into one (e.g. a fairness matroid \cite{el2020fairness,chen2024fair}).
%submodular maximization with partition structure has received attention \cite{lee2010maximizing,ene2019submodular}.
% \textcolor{red}{Should emphasize somewhere how partition constraints themselves are still important.}
In the context of submodular maximization, matroid constraints represent a fundamental and well-studied class of feasibility constraints, with partition constraints serving as a class of particularly prominent special case with widespread applications \citep{el2020fairness,chen2024fair}.
Therefore, algorithms for submodular maximization with a general matroid constraint, which has been extensively studied, can be employed \citep{nemhauser1978analysis,fisher1978analysis,calinescu2011maximizing,badanidiyuru2014fast,chekuri2019submodular,buchbinder2024deterministic}.
%Foundational work by \cite{nemhauser1978analysis, fisher1978analysis} analyzed greedy and local search algorithms, deriving key results such as a $1/(p + 1)$-approximation algorithm for monotone submodular maximization with $p$ matroid constraints.
The best known approximation ratio for monotone submodular maximization with a matroid constraint is $1-1/e$ \citep{calinescu2011maximizing,buchbinder2018deterministic}. For the more general maximization of a non-monotone submodular function with a matroid constraint, the best-known hardness result is 0.478~\citep{gharan2011submodular,qi2024maximizing}. The algorithm with the current best approximation ratio is a continuous one that achieves $0.401$ \citep{buchbinder2024constrained}.
The combinatorial algorithm with the best approximation ratio is that of \citet{chen2024discretely}, which achieves a $0.305-\epsilon$ approximation guarantee in $O\left(k^5\log(k)n/\epsilon\right)$ queries of $f$.
Partition type of constraints are widely found in submodular optimization applications, but despite this there has been little attention towards algorithms specifically designed for them. An exception is that fairness constraints have recently been of interest \citep{el2020fairness,el2023fairness,chen2024fair}.
%\citeauthor{calinescu2011maximizing} proposed a continuous method using the multi-linear extension, achieving the best-known approximation ratio of $1-1/e$ in expectation.
%However, continuous methods often suffer from high query complexity. More recently, \citeauthor{buchbinder2024deterministic} proposed deterministic algorithms for general matroids, achieving $1-1/e$ approximation guarantees in $\mathcal{O}(n\kappa)$ query complexity where $\kappa$ is the rank of the partition matroid, but their query complexity has a high dependence on the error parameter $\epsilon$.
%Submodular optimization involving fairness constraints, which are also defined over partitions of $U$, has recently been of interest \cite{el2020fairness,el2023fairness,chen2024fair}.
%Another line of work studied in this paper is the fair submodular optimization problem \cite{el2020fairness, el2023fairness}. It has been noted in
\citeauthor{el2020fairness} showed that maximization of a monotone submodular function under a fairness constraint can be converted into monotone SM under a matroid constraint.
The Submodular Maximization under Knapsack Partition Constraints (SMKP) problem, which is the dual problem of the SCKP, is defined as $\arg\max \{f(S): \sum_{s \in X \cap U_j} c(s) \leq p_j v, \forall j\in[N]\}$. This formulation can be regarded as the generalization of submodular maximization subject to a single knapsack constraint \citep{amanatidis2020fast,cui2025linear} in the case where there is only one partition in the universe. Recent work has also studied submodular maximization under both knapsack and partition constraints \citep{cui2024fairness, li2025fair}. Specifically, these papers address fairness-aware submodular maximization subject to: (i) a knapsack constraint on the total cost of the selected subset across all partitions, and (ii) cardinality constraints within each partition to ensure fair representation. In contrast, SMKP enforces per-partition knapsack constraints (as opposed to a single global knapsack constraint), without imposing cardinality requirements.
In the classical submodular cover problem with integral-valued objective functions, the standard greedy algorithm---which repeatedly selects the element with the highest marginal gain until the objective reaches a threshold $\tau$---achieves an approximation ratio of $O(\log \max_{e \in U} f(e))$~\citep{Wolsey82}. For real-valued submodular functions, a common modification is to stop once the function value reaches $(1 - \epsilon)\tau$, yielding algorithms with a $(\ln(1/\epsilon),\, 1 - \varepsilon)$-bicriteria approximation ratio~\citep{krause2008robust,chen2024bicriteria}. For the Fair Submodular Cover (FSC) \citep{chen2024fair} problem, the discrete algorithm of \citeauthor{chen2024fair} achieves a bicriteria approximation ratio of $(\mathcal{O}(1/{\epsilon}), 1-\epsilon)$ while our algorithm achieves an improved approximation ratio of $(\mathcal{O}(\ln1/{\epsilon}), 1-\epsilon)$, which matches the approximation ratio of the continuous method proposed in \citeauthor{chen2024fair} but requires much fewer queries.
% In contrast, our proposed dicrete algorithm achieves an approximation guarantee that matches the continuous approach in \cite{chen2024fair}.
% In particular, \cite{iyer2013submodular} proposed a method of converting algorithms for monotone submodular maximization with a cardinality constraint to algorithms for monotone submodular cover with uniform cost: Given a deterministic $(\gamma,\beta)$-bicriteria approximation algorithm for the maximization version, their algorithm produces a deterministic $((1+\alpha)\beta,\gamma)$-bicriteria approximation algorithm for the cover version.
%The algorithm works by making $\log_{1+\alpha}(n)$ guesses for $|OPT|$ (which is unknown in \scp), running the \smp algorithm with the budget set to each guess, and returning the smallest solution with $f$ value above $\gamma\tau$. However, this approach is limited by the approximation guarantees of existing algorithms for \smp. The best $\gamma$ for monotone \smp is $1-1/e$, and the best for general \smp where $f$ is not assumed to be monotone is significantly lower \citep{gharan2011submodular}.
%\citeauthor{chen2024fair} initiated the study of Fair Submodular Cover (FSC) and proposed a general converting theorems that can be used to convert algorithms for SM under fairness constraint into fair submodular cover respectively. The discrete algorithm achieves a bicriteria approximation ratio of $(\mathcal{O}(1/{\epsilon}), 1-\epsilon)$ while our algorithm achieves an improved approximation ratio of $(\mathcal{O}(\ln1/{\epsilon}), 1-\epsilon)$, which matches the approximation ratio of the continuous method proposed in \citeauthor{chen2024fair} but requires much fewer queries.
%For the more general maximization of a non-monotone submodular function with a matroid constraint, the best-known hardness result is 0.478~\citep{gharan2011submodular,qi2024maximizing}. The algorithm with the current best approximation ratio is a continuous one that achieves $0.401$ \cite{buchbinder2024constrained}.
%\citet{feldman2011unified} proposed a continuous algorithm with a $1/e$ approximation ratio. This was later improved to $0.385$ by \citet{buchbinder2019constrained}, and further to $0.401$ by \citet{buchbinder2024constrained}, which is the current best approximation ratio.
%But as previously mentioned, continuous algorithms often require a substantial number of queries.
% to estimate the continuous oracle from the discrete oracle, making them impractical in most scenarios
%For combinatorial algorithms, \citet{lee2009non} introduced the first constant-factor ($\frac{1}{4+\epsilon}$) approximation algorithms based on local search, requiring $O\left(k^5 \log(k)n/\epsilon\right)$ queries. Most recently, \citet{chen2024discretely} improved the approximation ratio to $0.305-\epsilon$ while maintaining the same query complexity.
% \textcolor{red}{Do you think we can delete this? Or shorten it significantly.}
% A related line of work is the $k$-submodular optimization problem \cite{ohsaka2015monotone,nguyen2020streaming}. \citeauthor{iwata2016improved} proved that the problem of $k$-submodular maximization problem is equivalent to submodular maximization problem under a specific type of partition matroid constraint where the size capacity of each group is upper bounded by $1$. In contrast, we address the problem of SM under general partition matroid constraint and especially focus on the case where the cardinality constraint within each partition is large.
%In this section, we present preliminary definitions and notations that will be used throughout the paper. First, we define a set function $f:2^U\rightarrow\mathbb{R}$ to be monotone i.f.f $f(Y)\ge f(X)$ for every $X\subseteq Y\subseteq U$. Throughout the paper, we define $[m]:=\{1,2,...m\}$.
% \textcolor{red}{Aren't all of these problems defined in their sections? I think we should just define them in their sections and remove the below definitions.}
% Next,
% we define the related problems studied in this paper. The first problem is Submodular Maximization under Partition matroid (SMP), which is defined to find the set $\arg\max_{S\in\mathcal{P}}f(S)$, where the constraint $\mathcal{P}$ is defined to be $\mathcal{P}=\{S\subseteq U:|S\cap U_j|\leq k_j, \forall j\in[N]\}$ and is also referred to as partition matroid constraint. We study the problem both in the monotone and nonmonotone setting. Another line of problems studied in this paper are the submodular cover problems. The first studied problem is \scpklong (\scpk). Let $c:U\rightarrow \mathbb{R}_+$ be the cost function defined on the ground, \scpk is defined to as
% of $\arg\min\{v:f(X)\geq\tau, \sum_{s\in X\cap U_j}c(s)\leq p_jv, \forall j\in[N]\}$
% \begin{align}
% &\min v\nonumber\\
% &\text{s.t. } f(X)\geq\tau\nonumber\\
% &~~\sum_{s\in X\cap U_j}c(s)\leq p_jv, \qquad\forall j\in[N].
% \end{align}
% The second submodular cover problem in this paper is the Fair Submodular Cover problem. Given input threshold $\tau$, and bounds on the proportion of the elements in each group $u_c$ and $l_c$, FSC is to find a set
% \begin{align*}
% &\min_{S\in U} |S|\nonumber\\
% s.t.\qquad& p_j|S|\leq|S\cap U_j|\leq q_j|S|\qquad \forall j\in[N]\nonumber\\
% &f(S)\geq\tau.
% \end{align*}
% The dual problem of FSC is the problem of Submodular Maximization under Fairness constraint, which is defined to be $$
\section{Algorithms and Theoretical Analyses}
\label{sec:main_results}
% \textcolor{red}{Should mention somewhere above how these problems are related to each other (which ones are special cases of others). I did that below but double check if correct.}
We now present the main results of our paper\footnote{We summarized our results in a table in the appendix. Please refer to Table \ref{tab:submodular-comparison}. }. We first address the general case of not necessarily monotone, submodular cover with a partition constraint in Section \ref{sec:nonmono_sc}. Next, we consider monotone submodular cover with a partition constraint, and our results apply even for the more general knapsack cost, in Section \ref{sec:knapsack}. Finally, we consider the more restricted, but with many interesting applications, setting of fair submodular cover in Section \ref{sec:fair}.
Central to all of our results is the novel algorithmic framework proposed for submodular maximization problems that achieves a bicriteria approximation ratio by running greedy in blocks, where each block is a feasible subset. This block-wise greedy strategy departs from prior approaches that focus on the matroid structure of partition constraints. In contrast, our method exploits the intrinsic relationship between partition constraints and cardinality constraints, leading to improved query complexity and approximation ratio. Throughout the paper, we define the marginal gain of adding an element $u\in U$ to a set $S\subseteq U$ as $\Delta f(S,u)=f(S\cup u)-f(S)$. Besides, $OPT$ is used to refer to the optimal solution to the instance of submodular optimization that should be clear from the context.
% Central to all of our results is the novel algorithmic framework proposed for submodular maximization problems with bicriteria approximation ratio that runs a greedy procedure in blocks. This allows us to explore
%which are organized in order of the problem that we consider. In particular, we address the fair submodular cover problem (FSC) in Section \ref{sec:fair}, which is a generalization of the SCP problem. In Section \ref{sec:knapsack}, we study the submodular cover with knapsack partition constraints (SCKP). In Section \ref{sec:nonmono_sc}, we address the submodular cover problem with partition matroid constraint. Throughout the paper, we define $[m]:=\{1,2,...m\}$.
\subsection{Non-Monotone Submodular Cover with Partition Constraints}
\label{sec:nonmono_sc}
% \textcolor{red}{I wonder if its best to present a "there exists an algorithm for SCP with guarantees ..." type of theorem, which states the main contibution, right at the beginning before presenting the algorithms. That way, they will see the result quickly rather than being first presented with the converting algorithm which might be confusing for some.}
In this section, we consider the general nonmonotone
%case of submodular cover subject to a partition constraint, where the submodular function $f$ is not necessarily monotone. In particular, we consider the
Submodular Cover with Partition Constraint (SCP) problem, which is to find a set $S\subseteq U$ that solves
\begin{align*}
\min_{S \subseteq U} \quad & v \\
\text{s.t.} \quad & |S \cap U_j| \leq p_j v, \quad \forall j \in [N], \\
& f(S) \geq \tau.
\end{align*}
The $v$ represents a budget to allocate over the partitioned sets, which our goal is to minimize while ensuring $f$ is sufficiently high. The $p_j$ represents the desired portion of the budget to allocate to the $j$-th partition. Without loss of generality, we assume $\sum_{j\in[N]}p_j=1$. If there is only one single partition in the universe, i.e., $N=1$, then the optimal value of $v$ is $|S|$, and we recover the classic submodular cover problem. To further illustrate the problem definition, consider the example application of video summarization described in Section \ref{sec:intro}, where frames are grouped by scene or content type and costs are uniform. Then this would mean the objective is to find a solution with a minimum total budget, while maintaining a balanced allocation across different partitions and ensuring that the summary achieves sufficiently high quality.
%upper bound on the size of $S$. We define SCP this way with this additional variable $v$, as opposed to $\text{argmin}\{|S|: |S\cap U_j|\leq p_j|S|, \forall j\in[N], f(S)\geq\tau\}$, in order .
%\begin{align*}
% &\min v\\
% s.t.\qquad&|S\cap U_j|\leq p_jv, \forall j\in[N]\}\\
% &f(S)\geq \tau
%\end{align*}
% Notice that unlike the monotone SCP problem, where we can propose algorithms that outputs solution set with function values being arbitrarily close to being feasible.
In our first result, taking advantage of the relationship between submodular cover and submodular maximization, we introduce a converting algorithm, \convr, that can convert any randomized bicriteria algorithm for nonmonotone Submodular Maximization with Partition matroid constraint (SMP) into a bicriteria algorithm for nonmonotone SCP. In particular, the SMP with an input budget $v$ is defined as $\max\{f(S):|S\cap U_i|\leq p_iv, \forall i\in[N]\}$. We formally define the notion of bicriteria approximation for both SMP and SCP in the following.
\begin{definition}
An $(\alpha,\beta)$-approximation algorithm for SMP with input budget $v$ returns a solution $X$ that satisfies
\begin{align*}
&f(X)\geq\alpha f(OPT),\\
& |X\cap U_j|\leq \beta p_jv,
\end{align*}
where $OPT$ is the optimal solution of SMP, i.e., $OPT:=\arg\max\{f(S):|S\cap U_i|\leq p_iv, \forall i\in[N]\}$.
\end{definition}
\begin{definition}
An $(\alpha,\beta)$-approximation algorithm for SCP returns a solution $X$ with objective value $v_X$ that satisfies
\begin{align*}
&v_X\leq\alpha v_{OPT},\\
& |X\cap U_j|\leq p_jv_X\\
&f(X)\geq\beta\tau.
\end{align*}
Here $OPT$ is the optimal solution of SCP, i.e., $OPT:=\arg\min\{v: |S\cap U_j|\leq p_jv, \forall j\in[N], f(S)\geq\tau\}$. $v_{OPT}$ is the optimal value of SCP.
\end{definition}
Notice that in \cite{chen2024bicriteria}, they proposed an algorithm to convert randomized submodular maximization algorithms into submodular cover algorithms in the case of monotone submodular objectives. In fact, a key challenge in this setting arises from the need to ensure a high-probability guarantee on the function value $f$ by repeatedly invoking the submodular maximization subroutine and applying concentration inequalities. To reduce the number of oracle queries, the algorithm in \cite{chen2024bicriteria} applies Markov's inequality and operates on a truncated objective function $f_\tau:=\min\{\tau, f\}$ throughout the converting algorithm. However, the assumption that $f_\tau$ is submodular only holds when $f$ is monotone. In contrast, our analysis extends to the non-monotone setting by avoiding the truncated objective and instead employing a more delicate analysis when applying the concentration inequality. Specifically, we analyze the deviation of the random variable $\beta f(OPT_g)-f(S_i)$ where $S_i$ is the output solution set of the randomized subroutine algorithm for SMP.
% This approach yields a $(1+\alpha,0.305-\epsilon)$-bicriteria approximation algorithm for SCP with high probability in $O(n\log_{1+\alpha}(|OPT|)\ln(1/\delta)\ln(1/\epsilon)/\ln(\frac{1-\gamma+\epsilon}{1-\gamma}))$ queries of $f$.
We now present \convr and its theoretical guarantees. The pseudocode for \convr is described in Algorithm \ref{alg:conver} in Section \ref{appdx:nonmono_sc} in the supplementary material. \convr runs by iteratively guessing the value of the optimal budget $v$. For each guess, \convr runs the corresponding dual submodular maximization algorithms over multiple independent trials. The theoretical guarantee of \convr is provided in Theorem \ref{thm:convert}. We defer the analysis to Section \ref{appdx:nonmono_sc} in the supplementary material.
\begin{theorem}
\label{thm:convert}
Any randomized $(\gamma,\beta)$-bicriteria approximation algorithm for nonmonotone SMP that runs in time $\mathcal{T}(n)$ where $\gamma$ holds only in expectation can be converted into an approximation algorithm for nonmonotone SCP that with probability at least $1-\delta$ is a $((1+\alpha)\beta,\gamma-\epsilon)$-bicriteria approximation algorithm that runs in time $O(\log_{1+\alpha}(|OPT|)\ln(1/\delta)\mathcal{T}(n)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma}))$.
\end{theorem}
%$O(\ln(|OPT|)\ln(1/\delta)\mathcal{T}_{sg}(n))$ queries of $f$ where $\mathcal{T}_{sg}(n)$ is the query complexity of the stochastic greedy algorithm for SM.
Since the best-known result of algorithms for nonmonotone submodular maximization over the partition matroid is the one proposed in \cite{chen2024discretely}, which achieves an approximation ratio of $0.305-\epsilon$. Therefore by applying Theorem \ref{thm:convert} to the randomized algorithm in \cite{chen2024discretely}, we have a $(1+\alpha,0.305-\epsilon)$-bicriteria approximation algorithm for SCP with high probability in $O(n|OPT|\log_{1+\alpha}(|OPT|)\ln(1/\delta)/\ln({1+\epsilon}))$ queries of $f$.
However, a factor of $0.305-\epsilon$ of $\tau$ is not very close to a feasible solution, and a natural question arises whether an algorithm that achieves a better feasibility guarantee exists. An important relevant result \citep{crawford2023scalable} is that it has been shown that a feasibility factor better than $1/2$ is impossible for the nonmonotone submodular cover problem. Since this problem is a special case of SCP, this result holds for SCP as well. Still, this leaves us with uncertainty of whether there exist scalable algorithms with approximation ratios in the gap between $0.305$ to $0.5$.
In the rest of this section, we present a scalable algorithm, \nonmonosc, that can output a solution set arbitrarily close to $\tau/e$. The pseudocode of \nonmonosc is provided in Algorithm \ref{alg:nonmono_sc}. \nonmonosc uses the idea of gradually developing a solution in blocks greedy algorithm, and achieves the bicriteria-approximation ratio as below.
% The main result of \nonmonosc for the problem of nonmonotone SMP is presented in Theorem \ref{thm:nonmonobi} below.
%On the other hand, unlike the monotone SCP problem, where we can propose algorithms that outputs solution set with function values being arbitrarily close to being feasible, when $f$ is nonmonotone, it is hard to obtain a solution arbitrarily close to being feasible. The traditional submodular cover problem defined as
% $\arg\min_{f(S)\geq\tau} |S|$ can be viewed as a special case of SCP where $N=1$, and $p_1=1$, the problem is reduced to the traditional nonmonotone SC problem. In this case, there doesn't exist any polynomial algorithms that can return a solution set $S$ that achieves
% $f(S)\geq \alpha \tau$ for any $\alpha>0.5$. Therefore, we can't propose an algorithm that runs on polynomial time and outputs a solution arbitrarily close to being feasible. In the rest of this section, we focus the rest of this section on developing an algorithm, \nonmonosc, that can output a solution set arbitrarily close to $\tau/e$. The pseudocode of \nonmonosc is provided in Algorithm \ref{alg:nonmono_sc}. \nonmonosc uses the idea of gradually developing a solution in blocks greedy algorithm and achieves the bicriteria-approximation ratio as below.
\begin{algorithm}[t] \caption{\nonmonosc}\label{alg:nonmono_sc}
\begin{algorithmic}[1]
% \State \textbf{Output: }$S\in U$
\State \textbf{Input: }$\epsilon$, partition constraint parameters, total budget $v$
\State \textbf{Output: } $S\subseteq U$
\For {$i=1$ \textbf{to} $\frac{2}{\epsilon} $}
\For{$j=1$ \textbf{to} $N $}
\For{$l=1$ \textbf{to} $p_jv $}
\State Let $M\subseteq U_j/S$ be a set of size $\frac{2p_jv}{\epsilon}$ maximizing $\sum_{x\in M}\Delta f(S,x)$.\label{line:greedy_nonmono_sc}
\State $u\gets$ uniformly sample an element from $M$
\State $S\gets S\cup \{u\}$
\EndFor
\EndFor
\EndFor
% \textbf{return }$S$
\end{algorithmic}
\end{algorithm}
\begin{theorem}\label{thm:nonmonobi}
Suppose that \nonmonosc is run for an instance of nonmonotone SMP with budget $v$,
then \nonmonosc
outputs a solution $S$ that satisfies a bicriteria approximation ratio of
$(1/e-\epsilon,\frac{2}{\epsilon})$ in expectation in at most $O(\frac{nv}{\epsilon})$ number of queries.
\end{theorem}
The analysis and the proof are deferred to Section \ref{appdx:nonmono_sc} in the supplementary material. From the results, we can get that using \nonmonosc as a subroutine in \convr yields a $(\frac{2(1+\alpha)}{\epsilon}, 1/e-2\epsilon)$-bicriteria approximation algorithm for nonmonotone SCP. The algorithm runs in $O(\frac{n|OPT|\log_{1+\alpha}(|OPT|)\ln(1/\delta)}{\epsilon\ln({1+\epsilon^2})})$ number of queries.
\subsection{Monotone Submodular Cover with Knapsack Partition Constraints}
\label{sec:knapsack}
\input{Sections/coverknapsack}
\begin{algorithm}[t]
\caption{\blockksc}
\label{procd:ksm}
\begin{algorithmic}[1]
\State \textbf{Input: } $\epsilon$, an instance of SMKP
\State \textbf{Output: } solution set $S\subseteq U$
% \State \textbf{Output: } $A$
% \Procedure{Greedy Block}{}
\For{$i=1$ \textbf{to} $\frac{\ln\frac{1}{\epsilon}}{\ln 2}$}
\For{$j=1$ \textbf{to} $N$} \label{line:knapsack_block_starts}
\State $A\gets\emptyset$, $ B_j\gets p_jv$.
% \Comment{$B_j$ is the budget of Partition $j$}
\While{\textbf{true}}
\State $s\gets\argmax_{x\in U_j/S,c(x)\leq B_j}\frac{\Delta f(S\cup A,x)}{c(x)}$\label{line:ksm_greedy}
% \State $A\gets S_{j}^i\cup\{s\}$
\State $A\gets A\cup \{s\}$
\If{$c(A)\geq B_j$}
\State $S\gets S\cup A$
\State \textbf{break}
\EndIf
\EndWhile
\EndFor\label{line:knapsack_block_ends}
\EndFor
\State \textbf{return} $S$
% \EndProcedure
\end{algorithmic}
\end{algorithm}
\subsubsection{Converting Theorem for SCKP}
In order to convert any bicriteria algorithms for SMKP into bicriteria algorithms for SCKP, we propose and analyze the converting algorithm, denoted as \convk.
The pseudocode for \convk is in Algorithm \ref{alg:convk}, and its theoretical guarantee is outlined in Theorem \ref{thm:convk}, both in Section \ref{appdx:convk} of the appendix. By leveraging the result in the converting theorem, we can obtain the following corollary.
\begin{corollary}
By using the \blockksc as a subroutine for the converting algorithm \convk, we can obtain an algorithm for SCKP that returns a solution set $S$ and $v_S$ that satisfies
\begin{align*}
&v_S\leq \frac{2(1+\alpha)\ln(1/\epsilon)}{\ln 2}v_{OPT}\\
&f(S)\geq(1-\epsilon)\tau\\
&c(S\cap U_j)\leq p_jv_S
\end{align*}
where $OPT$ and $v_{OPT}$ are the optimal solution set and the optimal value of the problem SCKP defined in (\ref{eqn:opt_prob_knapsack}) respectively. The total runtime of the algorithm is upper bounded by $\mathcal{O}(n^2\log_{1+\alpha}(\frac{c_{\max}n}{c_{\min}}))$. Here $c_{\max}$ and $c_{\min}$ are the maximum and minimum values of the cost of a single element respectively.
\end{corollary}
\input{Sections/FairnessCover}
% \wenjing{check later}
We now consider the problem of Monotone Submodular Cover with Knapsack Partition Constraints (SCKP).
SCKP models the setting where we want to balance the cost across different partitions, and the costs of different elements are nonuniform for different elements in the ground set. We illustrate an example of SCKP. Consider a neural network training task in deep learning, we want to select pretraining data points from $N$ different predefined groups, and the goal is to select a subset of data points with minimal budget of the cost, where the cost of each data point may reflect computational, labeling, or storage expenses. Simultaneously, we would want a solution with balanced cost allocation across predefined groups in the dataset while ensuring that the submodular utility function (e.g., coverage of diverse features) meets a specified threshold $\tau$.
More formally, the definition of SCKP is as follows. Define a cost function $c:U\to\mathbb{R}_{\geq 0}$, and let $c(S) = \sum_{x \in S} c(x)$ for any subset $S \subseteq U$. Then SCKP is:
%Let $c:2^U\rightarrow \mathbb{R}$ be a non-negative, modular cost function defined over subsets of a ground set$U$, satisfying $c(s) \geq 0$ for each element $s \in U$ and $c(S) = \sum_{x \in S} c(x)$ for any subset $S \subseteq U$. The problem of SCKP is defined as follows.
% of $\arg\min\{v:f(X)\geq\tau, \sum_{s\in X\cap U_j}c(s)\leq p_jv, \forall j\in[N]\}$
\begin{align}
\label{eqn:opt_prob_knapsack}
&\min v\nonumber\\
&\text{s.t. } f(S)\geq\tau\nonumber\\
&~~c(S\cap U_j)\leq p_jv, \qquad\forall j\in[N].
\end{align}
In the definition of SCKP, $v$ represents the budget for the total cost. More specifically, $v$ is the upper bound on the total cost. The second constraint ensures that the cost of the solution set $S$ within each partition $U_j$ does not exceed a specified fraction, $p_j$, of $v$. Without loss of generality, we assume that $\sum_{j\in[N]}p_j=1$.
This formulation naturally arises in various real-world applications, including influence maximization in social network analysis, where activating different nodes incurs varying costs, pretraining data selection for deep learning where different data points might require different computational or memory costs, and task allocation in multi-agent systems. Please see the Appendix \ref{appdx:knapsack_application} for a detailed discussion on the motivating examples. In the following part, we discuss the pretraining data selection as a motivating example of the SCKP problem.
% Unlike traditional submodular cover problems \cite{iyer2013submodular, chen2024bicriteria}, SCKP introduces an additional variable $v$. The benefit of introducing the $v$ provides two key advantages: (i). It ensures the feasibility of the problem (as long as $f(U)\leq\tau$, which is easy to verify) while preserving the essential requirement of obtaining a solution with balanced costs across partitions, (i.i). {\color{red}The solution of SCKP has a similar quality to the traditional submodular cover problems (see a detailed discussion in Appendix \ref{appdx:knapsack_prob_explain}),} (i.i.i). The presence of $v$ preserves the dual relationship with Submodular Maximization under Knapsack Partition Constraints (SMKP), enabling conversion between the two problems (see a detailed discussion in Appendix \ref{appdx:knapsack_prob_explain}).
To address SCKP, we first propose an algorithm for the dual problem of Submodular Maximization with Knapsack Partition Constraint (SMKP), which is defined as $\arg\max \{f(S): \sum_{s \in X \cap U_j} c(s) \leq p_j v, \forall j\in[N]\}$. Notice that when the cost is uniform, i.e., $c(s)=c,\forall s\in U$. SMKP is a monotone submodular maximization problem with a partition matroid constraint. Therefore, we can apply any submodular maximization algorithm with a matroid constraint. However, the output of standard algorithms can't achieve an objective value arbitrarily close to the optimal. Therefore, we propose the \blockksc algorithm, which proceeds in $\phi:=\frac{\ln\frac{1}{\epsilon}}{\ln 2}$ blocks. The pseudocode of \blockksc is described in Algorithm \ref{procd:ksm}. Within each block from Line \ref{line:knapsack_block_starts} to Line \ref{line:knapsack_block_ends}, the algorithm visits each partition $U_j$ in the ground set $U$ and adds elements greedily with the highest density of marginal gain. Below we present the theoretical guarantee of \blockksc for SMKP.
\begin{restatable}{theorem}{SMKthm}
\label{thm:SM_knapsack}
Suppose that \blockksc described in Algorithm \ref{procd:ksm} is run for an instance of SMKP, then \blockksc outputs a solution set that satisfies
\begin{align*}
&f(S)\geq (1-\epsilon)f(OPT)\\
&c(S\cap U_j)\leq \frac{2\ln\frac{1}{\epsilon}}{\ln 2}p_jv,\qquad \forall j\in[N],
\end{align*}
where $OPT$ is the optimal solution of SMKP.
\end{restatable}
This theorem guarantees that the solution set returned by \blockksc achieves an objective value arbitrarily close to the optimal while the cost constraints are satisfied up to a violation factor of $\frac{2\ln\frac{1}{\epsilon}}{\ln 2}$. In the special case of uniform costs, the theorem implies that \blockksc achieves a $(1 - \epsilon, O(\ln \tfrac{1}{\epsilon}))$ bicriteria approximation guarantee. This matches the best-known approximation ratio for monotone submodular maximization under a cardinality constraint. The improved performance stems from the blockwise structure of \blockksc, which effectively leverages the intrinsic similarity between the submodular maximization problem under the cardinality constraint and the partition matroid constraint. The proof and analysis of Theorem \ref{thm:SM_knapsack} are deferred to Appendix \ref{appdx:proof_SMKP}.
\subsection{Submodular Cover with Fairness Constraint}
\label{sec:fair}
In this section, we consider the monotone Submodular Cover problem with Fairness Constraint (SCF), recently proposed by \citet{chen2024fair} (also referred to as Fair Submodular Cover in \cite{chen2024fair}). SCF is defined as follows: Given a monotone and submodular function $f$, a threshold $\tau$, and bounds $p_c$ and $q_c$ on the proportion limits of the elements in each group, SCF aims to find
\begin{align*}
&\argmin_{S\in U} |S|\nonumber\\
s.t.\qquad& p_c|S|\leq|S\cap U_c|\leq q_c|S|,\qquad \forall c\in[N]\nonumber\\
&f(S)\geq\tau.
\end{align*}
\citeauthor{chen2024fair} proposed a converting approach that takes bicriteria algorithms for the dual problem of Submodular Maximization with Fairness constraint (SMF) \citep{el2020fairness} and converts them into algorithms for SCF. Formally, SMF seeks to maximize $f(S)$ subject to the constraint that $S\in \mathcal{M}_{fair}$. $\mathcal{M}_{fair}$ represents the fairness matroid and is defined as
$\mathcal{M}_{fair}=\{S \subseteq U : |S \cap U_c| \leq u_c, \forall c \in [N],
\sum_{c\in [N]}\max\{|S\cap U_c|, l_c\}\leq k\}.$
If we set $l_c=0$ for all $c\in[N]$ and we set $k=\sum_{c\in[N]}u_c$, then $\mathcal{M}_{fair}$ is equivalent to $\mathcal{M}_{fair}=\{S \subseteq U : |S \cap U_c| \leq u_c, \forall c \in [N]\}.$ Therefore, SMF can be viewed as a generalized form of submodular maximization with a partition matroid constraint.
A bicriteria approximation guarantee for SMF is defined as follows.
\begin{definition}
\label{def:bicri-SM}
A discrete algorithm for SMF with an $(\alpha,\beta)$-bicriteria approximation ratio returns a solution $X$ such that
\begin{align*}
&f(X)\geq \alpha f(OPT),\\
&|X\cap U_c|\leq \beta u_c\qquad\forall c\in[N], \\
&\sum_{c\in [N]}\max\{|X\cap U_c|,\beta l_c\}\leq \beta k.
\end{align*}
%\begin{align*}
%&f(X)\geq \alpha f(OPT)\\
%&|X\cap U_c|\leq \phi u_c\qquad\Forall c\in[N]\\
%&\sum_{c\in [N]}\max\{|X\cap U_c|,\phi l_c\}\leq \phi k.
%\end{align*}
Here $OPT$ is the optimal solution of the problem SMF, i.e., $OPT=\arg\max_{S\in \mathcal{M}_{fair}}f(S)$.
\end{definition}
Therefore, in order to leverage the conversion approach of \citeauthor{chen2024fair}, we propose an algorithm for SMF called \fairgreedy that uses a greedy block formation technique. \fairgreedy is an improvement over the \texttt{greedy-fairness-bi} in Algorithm 5 in \cite{chen2024fair} . This leads to a bicriteria approximation guarantee for SCF of $(\frac{1+\ln\frac{1}{\epsilon}}{\ln 2}, 1-\epsilon)$, which is a significant improvement compared to the best-known results for discrete algorithms of $(\frac{1}{\varepsilon}+1,1-\mathcal{O}(\varepsilon))$ in \citeauthor{chen2024fair}.
We now describe \fairgreedy, pseudocode for which is presented in Algorithm \ref{alg:fair}. \fairgreedy adopts a similar approach of running greedy algorithms in blocks. By definition, the $\beta$-extension of the fairness matroid constraint $\mathcal{M}_{fair}$ is given by $\mathcal{M}_\beta:=\{S \subseteq U : |S \cap U_c| \leq \beta u_c, \forall c \in [N], \sum_{c\in [N]}\max\{|S\cap U_c|, \beta l_c\}\leq \beta \kappa\}$ as introduced in \cite{chen2024fair}. Therefore, an algorithm with a bicriteria approximation ratio of $(\alpha,\beta)$ for the SMF problem outputs a solution set that belongs to $\mathcal{M}_\beta$. The key intuition behind \fairgreedy is that any set in the $\mathcal{M}_\beta$ can be expressed as the union of $\beta$ subsets, each belonging to $\mathcal{M}_{fair}$. This motivates our algorithm of dividing the capacity of the solution set into $\beta$ blocks. Here $\beta:=\frac{\ln(1/\epsilon)}{\ln 2}$. Within each block from Line \ref{line:fair_outer_for_loop_starts} to Line \ref{line:fair_outer_for_loop_ends} in Algorithm \ref{alg:fair}, the algorithm operates in a greedy fashion: it iteratively adds the element with the highest marginal gain while maintaining $B\in\mathcal{M}_{fair}$. The main theoretical result is stated in Theorem \ref{thm:fair_greedy} below. The proof of the theorem is deferred to Appendix \ref{appdx:fair}.
\begin{algorithm}[t]
\caption{\fairgreedy}\label{alg:fair}
\begin{algorithmic}[1]
\State \textbf{Input: }fairness parameters
\State \textbf{Output: }$S\in U$
\State $S\gets\emptyset$
\For{$i=1$ \textbf{to} $\frac{\ln(1/\epsilon)}{\ln 2}$}
\State $B\gets\emptyset$\label{line:fair_outer_for_loop_starts}
\While{$\exists s$ \textbf{ s.t. } $B\cup\{s\}\in\mathcal{M}_{fair}$}
\State $x\gets \arg\max_{s\in U,B\cup\{s\}\in\mathcal{M}_{fair}}\Delta f(S,s)$
\State $B\gets B\cup\{x\}$
\State $S\gets S\cup\{x\}$\label{line:fair_outer_for_loop_ends}
\EndWhile
\EndFor
\textbf{return }$S$
\end{algorithmic}
\end{algorithm}
\begin{restatable}{theorem}{fairthm}
\label{thm:fair_greedy}
Suppose that \fairgreedy is run for an instance of SMF,
then \fairgreedy outputs a solution $S$ that satisfies a $(1-\varepsilon, \frac{\ln\frac{1}{\varepsilon}}{\ln 2})$-bicriteria approximation guarantee in at most $O\left(n\kappa \ln(1/\varepsilon)\right)$ queries of $f$.
\end{restatable}
\section{Experiments}
\label{sec:exp}
\begin{figure*}[t!]
\centering
\hspace{-0.5em}
\subfigure[euall $f$]
{\label{fig:euall_f}\includegraphics[width=0.24\textwidth]{graphs/euallblock_nonmono_n265214-f.pdf}}
\hspace{-0.5em}
\subfigure[euall budget]
{\label{fig:euall_budget}\includegraphics[width=0.24\textwidth]{graphs/euallblock_nonmono_n265214-budget.pdf}}
\hspace{-0.5em}
\subfigure[twitch $f$]
{\label{fig:twitch_f}\includegraphics[width=0.24\textwidth]{graphs/large_twitch_threshold_n2000-f.pdf}}
\hspace{-0.5em}
\subfigure[twitch budget]
{\label{fig:twitch_budget}\includegraphics[width=0.24\textwidth]{graphs/large_twitch_threshold_n2000-budget.pdf}}
\hspace{-0.5em}
\subfigure[corel cost]
{\label{fig:cover5000_fair_c}\includegraphics[width=0.24\textwidth]{graphs/threshold_n4500-cost.pdf}}
\hspace{-0.5em}
\subfigure[corel fairness difference]
{\label{fig:cover5000_fair_difference}\includegraphics[width=0.24\textwidth]{graphs/threshold_n4500-fair-difference.pdf}}
\hspace{-0.5em}
\subfigure[synthetic cost]
{\label{fig:synthetic_c}\includegraphics[width=0.24\textwidth]{graphs/synthetic_set_cover_threshold_n1400-cost.pdf}}
\hspace{-0.5em}
\subfigure[synthetic fairness difference]
{\label{fig:synthetic_fair_difference}\includegraphics[width=0.24\textwidth]{graphs/synthetic_set_cover_threshold_n1400-fair-difference.pdf}}
\caption{The experimental results of running the algorithms on the euall dataset, the twitch dataset, the Corel5k dataset, and the synthetic dataset. Budget: $\max_{i\in[N]}\frac{c(S\cap U_i)}{p_i}$. Cost: the size of the solution. Fairness difference: $(\max_c |S \cap U_c| - \min_c |S \cap U_c|) / |S|$. }
\label{fig:exp_result}
\end{figure*}
In this section, we present an empirical evaluation of our proposed algorithms. In particular, we evaluate our \nonmonosc algorithm on instances of graph cut in Section \ref{sec:nonmono_exp}. Next, we evaluate \blockksc on set cover instances and \fairgreedy on both max cover and image summarization tasks in Sections~\ref{sec:knapsack_exp} and~\ref{sec:fair_exp}, respectively. Additional details about the applications, setup, and results can be found in Section \ref{appdx:exp} in the supplementary material.
\subsection{Experiments on Nonmonotone SCP}
\label{sec:nonmono_exp}
We evaluate the performance of our algorithms on several instances of graph cut over social
network data. The dataset used in the main paper is the email-EuAll dataset (n = 265214,
420045 edges) from the SNAP large network collection \citep{leskovec2016snap}. We compare the solutions returned by the \convr algorithm with four subroutines including: (i). our \nonmonosc algorithm ("BLOCK-G") (ii). the streaming algorithm from \citet{feldman2018less} ("STREAM"); (iii). the randomized algorithm of~\citet{chen2024discretely}, initialized with the twin-greedy solution proposed in~\citet{han2020deterministic} ("GUIDED-RG"). (iv). The random-greedy algorithm for the submodular maximization problem with the cardinality constraint being the input guess of budget $\kappa$ \citep{buchbinder2014submodular} ("RG"). The algorithms are evaluated in terms of the function value $f(S)$ returned by the solution $S$, the query complexity, and the minimum value of budget $v$ that satisfies the partition constraints, i.e., $\max_{i\in[N]}\frac{|S\cap U_i|}{p_i}$, and the execution time. The results are compared for different values of the threshold $\tau$.
% For all of the algorithms, the parameter $\alpha$ in the converting algorithm \convr is $0.2$. $\epsilon$ is set to be $0.05$.
The results in terms of $f$ and the minimum budget are described in Figure \ref{fig:euall_f} and \ref{fig:euall_budget}. From the results, one can see that RG and BLOCK-G consistently achieve higher objective values $f$ than the other methods. This is consistent with our theoretical results as the approximation ratio on the function value of these two algorithms satisfies $f(S)\geq (1/e-\epsilon)\tau$ and $f(S)\geq (1/e-2\epsilon)\tau$ respectively while the other two algorithms STREAM, GUIDED-RG achieve worse approximation ratios on the function value of $1/0.583-\epsilon$, $0.305-2\epsilon$ respectively. However, in terms of budget, we can see that the RG algorithm performs poorly, since it does not account for partition constraints, resulting in imbalanced budget allocations. The STREAM and GUIDED-RG algorithm returns solutions with smaller budgets since both these two algorithms achieve a bicriteria approximation ratio such that $v_S\leq (1+\alpha)v_{OPT}$. While BLOCK-G has a higher budget due to its weaker bicriteria guarantee of $v_S \leq \frac{2(1+\alpha)}{\epsilon}v_{\text{OPT}}$, it does achieve a significantly higher function value.
\subsection{Experiments on SCKP}
\label{sec:knapsack_exp}
We evaluate three different algorithms on the instance of max cover: the converting algorithm \convk with two different subroutines: \blockksc ("BLOCK-G") and a standard greedy algorithm without the block-wise structure ("GREEDY"), and the greedy algorithm for submodular cover without the partition-knapsack constraint ("GREEDY-Knapsack"). Further details regarding the GREEDY and the GREEDY-Knapsack algorithms, and the experimental setup are provided in Appendix \ref{appdx:setup}.
The results in terms of the minimum budget, which can be calculated by $\max_{i\in[N]}\frac{c(S\cap U_i)}{p_i}$, and the function value $f$ are plotted in Figure \ref{fig:twitch_f}
and Figure \ref{fig:twitch_budget}. From the results, one can see that the $f$ values of solutions returned by BLOCK-G,
GREEDY-Knapsack, and GREEDY are nearly the same. This is because the theoretical guarantees on $f$ are about
the same for the different algorithms. However, the budget of the solution returned by our algorithm BLOCK-G is smaller than the other two algorithms, which demonstrates the effectiveness of our approach of running greedy in blocks.
\subsection{Experiments on SCF}
\label{sec:fair_exp}
For the SCF problem, we evaluate algorithms using the conversion framework from \cite{chen2024fair} with different subroutines: our \fairgreedy algorithm ("BLOCK-G-Fair"), the standard greedy algorithm ("GREEDY-Fair") and the threshold greedy algorithm ("THRES-Fair"). These algorithms are compared in terms of solution cost (cardinality), fairness difference, objective function value, query complexity, and execution time for varying values of the threshold $\tau$. Here we set $\alpha=0.2$ for the converting algorithms in Algorithm 1 in \cite{chen2024fair}. The parameters in the fairness constraint are set to be $ u_c = 1.1 / N, l_c = 0.9 / N$. (where $N$ is the number of groups). Additional details about the applications, setup, and results can be found in Section \ref{appdx:exp} in the appendix.
% \subsection{Experimental Results}
% We first analyze the results for the monotone SMP problem. The results in terms of the $f$ values and the number of queries are presented in Figure \ref{fig:cover5000_f}, \ref{fig:cover5000_q}, \ref{fig:synthetic_f} and \ref{fig:synthetic_q}. From the plots Figure \ref{fig:cover5000_f} and \ref{fig:synthetic_f}, one can see that the $f$ values returned by GREEDY and THRES, are nearly the same, and are smaller than the ones returned by BLOCK-G. This is unsurprising, because the theoretical guarantees on $f$ are about the same for the GREEDY and THRES algorithms ($1/2$ and $1/2-\mathcal{O}(\epsilon)$), and is worse than that of BLOCK-G algorithm ($1-1/e-\frac{1}{\phi+1}$). From the plots Figure \ref{fig:cover5000_q} and \ref{fig:synthetic_q}, we can also see that the number of queries required by BLOCK-G is smaller than the GREEDY algorithm, which aligns with our theoretical results as the BLOCK-G algorithm restricts its greedy selection process to only one partition at each step.
Figures \ref{fig:cover5000_fair_c} and \ref{fig:synthetic_c} illustrate the cost (cardinality) of the solution sets, while Figures \ref{fig:cover5000_fair_difference} and \ref{fig:synthetic_fair_difference} show the fairness differences across varying $\tau$ values. In most cases, BLOCK-G-Fair achieves a lower cost than THRES-Fair and GREEDY-Fair, aligning with our theoretical results. In the case where $\tau$ is large on the corel dataset, the cost of the THRES-Fair is smaller than the BLOCK-G-Fair. However, Figures \ref{fig:cover5000_fair_difference} and \ref{fig:synthetic_fair_difference} reveal that fairness differences in THRES-Fair and GREEDY-Fair are significantly larger than in BLOCK-G-Fair, demonstrating that BLOCK-G-Fair produces more balanced solutions. This is expected given that the fairness constraint from \cite{chen2024fair} ensures that $\beta\lfloor \frac{p_c|S|}{\beta}\rfloor \leq |S \cap U_c| \leq \beta\lceil \frac{q_c|S|}{\beta}\rceil$, which means the solution set might break the fairness constraint by an additive factor of $\beta$. Notably, $\beta=\mathcal{O}(1/\epsilon)$ for the THRES-Fair and GREEDY-Fair and $\beta =\frac{\ln(\frac{1}{\epsilon})}{\ln2}$ for BLOCK-G-Fair, which means our method achieves an enhanced fairness guarantee.
\appendix
\onecolumn
\section*{Appendix}
\section{The Use of Large Language Models (LLMs)}
In this paper, we use Large Language Models solely to polish the writing to improve the clarity and presentation. All theoretical results and technical contributions in this paper were not developed by LLMs.
\section{Appendix for Section \ref{sec:nonmono_sc}}
\label{appdx:nonmono_sc}
In this section, we present missing content from Section \ref{sec:nonmono_sc}, where we considered non-monotone submodular cover with partition constraints. First, pseudocode for the converting algorithm \convr, which we only informally described in Section \ref{sec:nonmono_sc}, is given in Algorithm \ref{alg:conver}. Next, we present the omitted proofs of Theorems \ref{thm:convert} and Theorem \ref{thm:nonmonobi}.
%First of all, we provide the omitted content about the converting algorithm \convr and its theoretical guarantee. In particular, we prove the theoretical guarantee of \convr as stated in Theorem \ref{thm:convert}. The pseudocode of the converting algorithm \convr is presented in Algorithm \ref{alg:conver}. Next, we provide the omitted proof of Theorem \ref{thm:nonmonobi} on the theoretical guarantee of the algorithm \nonmonosc. We now provide the proof of Theorem \ref{thm:convert} as below.
\begin{algorithm}[t!]
\caption{\convr}\label{alg:conver}
\textbf{Input}: An SCP instance with threshold $\tau$, a $(\gamma,\beta)$-bicriteria approximation algorithm for SMP, $\alpha>0$\\
\textbf{Output}: $S\subseteq U$
\begin{algorithmic}[1]
\State $S_i\gets\emptyset$, $\forall i\in\{1,..., \ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})\}$
\State $g\gets(1+\alpha)$
\While{$f(S_i)<(\gamma-\epsilon)\tau$ $\forall i$}\label{line:convrtest}
\For{$i\in\{1,..., \ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})\}$}
\State $S_i\gets(\gamma,\beta)$-bicriteria approximation for SMP with objective function $f$ and budget $g$ \label{line:convertmax}
\EndFor
\State $g\gets(1+\alpha)g$
\EndWhile
\State \textbf{return} $S$
\end{algorithmic}
\end{algorithm}
\textbf{Theorem \ref{thm:convert}. }\textit{
Any randomized $(\gamma,\beta)$-bicriteria approximation algorithm for SMP that runs in time $\mathcal{T}(n)$ where $\gamma$ holds only in expectation can be converted into a $((1+\alpha)\beta,\gamma-\epsilon)$-bicriteria approximation algorithm for SCP that runs in time $O(\log_{1+\alpha}(|OPT|)\ln(1/\delta)\mathcal{T}(n))/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma}))$ where $\gamma$ holds with probability at least $1-\delta$.}
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|p{2.5cm}|p{4cm}|p{3.5cm}|p{5.5cm}|}
\hline
\textbf{Problem} & \textbf{Algorithm} & \textbf{Approximation Ratio} & \textbf{Query Complexity} \\
\hline
Nonmonotone SCP & \convr + \nonmonosc &
\(\left(O\left(\frac{1+\alpha}{\epsilon}\right), \frac{1}{e}-\epsilon\right)\) &
\(\mathcal{O}\left(n^2 \log_{1+\alpha} n \ln(1/\delta)/\epsilon/\ln(1+\epsilon)\right)\) \\
% \hline
% Nonmonotone SCP & \convr + GUIDED-RG \cite{chen2024discretely} &
% \(\left(1+\alpha, 0.305 - 2\epsilon\right)\) &
% \(\mathcal{O}\left(n^2 \log_{1+\alpha} n \ln(1/\delta)/\epsilon/\ln(1+\epsilon)\right)\) \\
\hline
Monotone SCKP & \blockksc + convert &
\(\left(\frac{(1+\alpha)\ln(1/\epsilon)}{\ln 2}, 1 - \epsilon\right)\) &
\(\mathcal{O}\left(n^2 \log_{1+\alpha}\left(\frac{c_{\max}n}{c_{\min}}\right)\right)\) \\
\hline
Monotone SCF & \fairgreedy + convert-fair \citep{chen2024fair} &
\(\left(\mathcal{O}(\ln(1/\epsilon)), 1 - \epsilon\right)\) &
\(\mathcal{O}\left(\frac{n \log n \kappa \ln(1/\epsilon)}{\epsilon}\right)\) \\
% \hline
% Monotone SCF & greedy-fairness-bi \cite{chen2024fair} + convert-fair \cite{chen2024fair} &
% \(\left(\mathcal{O}\left(\frac{1}{\varepsilon}\right), 1 - \varepsilon\right)\) &
% \(\mathcal{O}\left(n \log(n) \kappa / \varepsilon^2\right)\) \\
% \hline
% Monotone SCF & cont-thresh-greedy-bi \cite{chen2024fair} + convert-continuous \cite{chen2024fair} &
% \(\left(O\left(\ln\left(\frac{1}{\varepsilon}\right)\right), 1 - \varepsilon\right)\) &
% \(\mathcal{O}\left(\frac{n^2 (1+\alpha) \log^2\left(\frac{n}{\varepsilon}\right) \log n}{\varepsilon^4 \alpha}\right)\) \\
\hline
\end{tabular}
\caption{Summarization of approximation algorithms for Submodular Cover Problems in this paper.}
\label{tab:submodular-comparison}
\end{table}
\begin{proof}
Consider the run of the algorithm for SMP on Line \ref{line:convertmax}
of Algorithm \ref{alg:conver} when the guess of optimal value $g$ falls into the region
\begin{align*}
v_{OPT}\leq g\leq(1+\alpha)v_{OPT}.
\end{align*}
Let us denote the partition matroid with budget $g$ as $\mathcal{M}$, i.e., $\mathcal{M}:=\{S\subseteq U:|S\cap U_j|\leq p_jv, \forall j\in[N]\}$. The SMP problem is then defined to find $\arg\max\{f(S):S\in\mathcal{M}\}$. We denote the optimal solution of the SMP problem with budget $g$ as $OPT_g$, i.e., $$OPT_g:=\arg\max\{f(S):|S\cap U_j|\leq p_jv, \forall j\in[N]\}.$$ Besides, we define the optimal solution for SCP as $OPT$. By the fact that the optimal solution $OPT$ is feasible for SCP, we have that
\begin{align}
|OPT\cap U_j|\leq p_j v_{OPT}\leq p_jv
\end{align}
which means that $OPT\in\mathcal{M}$, and therefore $f(OPT)\leq\max_{S\in\mathcal{M}}f(S)= f(OPT_g)$. Since $ f(OPT)\geq \tau$, we have $f(OPT_g)\geq \tau$. It then follows that for each $i\in\{1,..., \ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})\}$,
\begin{align*}
P\Big(f(S_i)\leq(\gamma-\epsilon)\tau\Big)&\leq P\Big(f(S_i)\leq(\gamma-\epsilon)f(OPT_g)\Big)\\
&\leq P\Big(\beta f(OPT_g)-f(S_i)\geq(\beta-\gamma+\epsilon)f(OPT_g)\Big)
\end{align*}
By the theoretical guarantees of the algorithm for SMP, we have that for all $i\in\{1,..., \ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})\}$, we have that $\mE f(S_i)\geq \gamma f(OPT_g)$ and $|S_i\cap U_j|\leq p_j\beta g$ for each $j\in[N]$. It then follows that
\begin{align*}
P\Big(f(S_i)\leq(\gamma-\epsilon)\tau\Big)
&\leq P\Big(\beta f(OPT_g)-f(S_i)\geq\frac{\beta-\gamma+\epsilon}{\beta-\gamma}(\beta f(OPT_g)-\mE f(S_i))\Big)
\end{align*}
Let us denote $\mathcal{M}_\beta:=\{S\subseteq U:|S\cap U_j|\leq p_j\beta g, \forall j\in[N]\}$, i.e., $\mathcal{M}_\beta$ is the $\beta$-extension of the matroid $\mathcal{M}$ as is defined in \cite{chen2024fair}. Since $S_i$ satisfies $|S_i\cap U_j|\leq p_j\beta g$ for any $j\in[N]$, we have that $S_i\in \mathcal{M}_\beta$ for each $i\in\{1,..., \ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})\}$. Notice that for any set $A\in\mathcal{M}_\beta$, we can express $A$ as the union of $\beta$ disjoint subsets in $\mathcal{M}$. Let us denote them as $A_1, A_2,..A_m$. Thus we have that
\begin{align*}
f(A)&=f(\cup_{i\in[\beta]}A_i)\\
&\leq \sum_{i\in[\beta]}f(A_i)\leq \beta f(OPT_g),
\end{align*}
where the first inequality follows from the submodularity of $f$. It then follows that $\max_{S\in\mathcal{M}_\beta}f(S)\leq \beta f(OPT_g)$. Notice that $S_i\in\mathcal{M}_\beta$, we have that $ \beta f(OPT_g)-f(S_i)\geq 0$. Thus, we can apply Markov's inequality on the random variable $\beta f(OPT_g)-f(S_i)$. Therefore, we can get that for each $i\in\{1,..., \ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})\}$
\begin{align*}
P\Big(f(S_i)\leq(\gamma-\epsilon)\tau\Big)
&\leq P\Big(\beta f(OPT_g)-f(S_i)\geq\frac{\beta-\gamma+\epsilon}{\beta-\gamma}(\beta f(OPT_g)-\mE f(S_i))\Big)\\
&\leq \frac{\beta-\gamma}{\beta-\gamma+\epsilon}
\end{align*}
Then the probability that none of the subsets $S_i$ can reach the stopping condition can be bounded by
\begin{align*}
P(f(S_i)\leq(\gamma-\epsilon)\tau, \forall i)&=P(f(S_i)\leq(\gamma-\epsilon)\tau, \forall i)\notag\\
&=\prod_{i=1}^{\ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})}P(f(S_i)\leq(\gamma-\epsilon)\tau)\\
&\leq(\frac{\beta-\gamma}{1+\epsilon-\gamma})^{\ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})}=\delta.
\end{align*}
This means with probability at least $1-\delta$, \convr stops when $g$ reaches the region where $v_{OPT}\leq g\leq(1+\alpha)v_{OPT}$ since the condition of the \textbf{while} loop is not satisfied. Therefore, by the assumption that the subroutine algorithm is a $(\gamma,\beta)$-bicriteria approximation algorithm, we have that the output solution $S$ satisfies that $|S\cap U_j|\leq p_j\beta g\leq p_j\beta(1+\alpha)v_{OPT}$. Then the objective value of the optimal solution $S$ can be set to be $v_S=\beta(1+\alpha)v_{OPT}$. It also implies that there are at most $O(\log_{1+\alpha}v_{OPT})$ number of guesses of the cardinality of the optimal solution. Since for each guess, we run the SMP for $\ln(1/\delta)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma})$ times, the algorithm runs in time $O(\log_{1+\alpha}(v_{OPT})\ln(1/\delta)\mathcal{T}(n)/\ln(\frac{\beta-\gamma+\epsilon}{\beta-\gamma}))$.
\end{proof}
Next, we present the proof for the result in Theorem \ref{thm:nonmonobi}.
\textbf{Theorem \ref{thm:nonmonobi}.}\textit{
Suppose that \nonmonosc is run for an instance of nonmonotone SMP with budget $v$,
then \nonmonosc
outputs a solution $S$ that satisfies a bicriteria approximation ratio of
$(1/e-\epsilon,\frac{2}{\epsilon})$ in expectation in at most $O(\frac{nv}{\epsilon})$ number of queries.
}
\begin{proof}
Let us denote the solution set after adding the $l$-th element in $j$-th subgroup during the $i$-th round of the outer for loop in Line \ref{line:greedy_nonmono_sc} in Algorithm \ref{alg:nonmono_sc} as $S_{i,j,l}$, and we define the solution set after completing adding all the elements in the $j$-th subgroup during the $i$-th round in Algorithm \ref{alg:nonmono_sc} as $S_{i,j}$. For notation simplicity, we also define $\phi:=\frac{2}{\epsilon}$. From the greedy selection strategy, we have that
\begin{align*}
\mE [f(S_{i,j,l})-f(S_{i,j,l-1})]\geq\frac{\sum_{a\in OPT\cap U_j}\Delta f(S_{i,j,l-1},a)}{p_jv\phi}.
\end{align*}
By submodularity, we would have that
\begin{align*}
\mE [f(S_{i,j,l})-f(S_{i,j,l-1})]\geq\frac{\sum_{a\in OPT\cap U_j}\Delta f(S_{i,j},a)}{p_jv\phi}.
\end{align*}
By summing over all $l\in[p_jv]$, it then follows that
\begin{align*}
\mE [f(S_{i,j})-f(S_{i,j-1}]&\geq\frac{\sum_{a\in OPT\cap U_j}\Delta f(S_{i,j},a)}{\phi}.
\end{align*}
Let us denote the solution after completing the entire $i$-th round as $S_i$. By submodularity, it then follows that
\begin{align*}
\mE[f(S_{i,j})-f(S_{i,j-1})]&\geq\frac{\mE[\Delta f(S_{i,j},OPT_j)]}{\phi}\\
&\geq\frac{\mE[\Delta f(S_{i},OPT_j)]}{\phi},
\end{align*}
By summing over all $j\in[N]$, we have
\begin{align*}
\mE [f(S_{i})-f(S_{i-1})]&\geq\frac{\sum_{j=1}^N\mE[\Delta f(S_{i},OPT_j)]}{\phi}\\
&\geq\frac{\mE[\Delta f(S_{i},OPT)]}{\phi}.
\end{align*}
Then it follows that
\begin{align*}
\mE [f(S_{i})-f(S_{i-1})]\geq\mE[\frac{f(S_{i}\cup OPT)-f(S_i)}{\phi}].
\end{align*}
Notice that by the greedy selection step, for each group $j$ and each element $a\in OPT\cap U_j$ appears in $S_i$ with probability at most $1-(1-\frac{1}{p_jv\phi })^{p_jvi}$. Since $(1-\frac{1}{x})^x$ increases with $x$ in the range of $[1,+\infty)$, we have that $(1-\frac{1}{p_jv\phi })^{p_jv\phi }\geq (1-\frac{1}{\phi})^\phi$. Therefore, we would get $1-(1-\frac{1}{p_jv\phi })^{p_jvi}\leq 1-(1-\frac{1}{\phi})^i$. From Lemma 2.2 in \cite{buchbinder2014submodular}, we can conclude that
\begin{align*}
\mE [f(S_{i}\cup OPT)]\geq(1-\frac{1}{\phi})^if(OPT).
\end{align*}
By rearranging the above inequality, we can get that
\begin{align*}
\mE [f(S_{i})]\geq\frac{\phi}{\phi+1}\mE [f(S_{i-1})]+\frac{1}{\phi+1}(1-\frac{1}{\phi})^if(OPT).
\end{align*}
By induction, we have that the output solution set satisfies that
\begin{align}
\label{eqn:nonmono_induction}
\mE [f(S)]&=\mE [f(S_{\phi})]\nonumber\\
&\geq\frac{1}{\phi+1}\sum_{i=1}^\phi(\frac{\phi}{\phi+1})^{\phi-i}(1-\frac{1}{\phi})^if(OPT)\nonumber\\
&\geq\frac{1}{\phi+1}\sum_{i=1}^\phi(\frac{\phi-1}{\phi})^{\phi-i}(1-\frac{1}{\phi})^if(OPT)\nonumber\\
&\geq\frac{\phi}{\phi+1}(1-\frac{1}{\phi})^\phi f(OPT)\geq\frac{1}{e}(1-\epsilon)f(OPT).
\end{align}
where the last inequality follows from the fact that $(1-\frac{1}{\phi})^{\phi-1}\geq e^{-1}$ for any $\phi>1$, and that $\phi=\frac{2}{\epsilon}$.
\end{proof}
\section{Appendix for Section \ref{sec:knapsack}}
\label{appdx:scpk}
We now present omitted content from Section \ref{sec:knapsack}, where we studied monotone submodular cover with knapsack partition constraints. Our first goal is to provide additional detail to motivate and explain our proposed optimization problem formulation. In particular, several detailed motivating examples of SCKP are presented in Section \ref{appdx:knapsack_application}, and then further we provide detailed discussion on the formulation of SCKP in Section \ref{appdx:knapsack_prob_explain}. Then in Section \ref{sec:knapsack}, we provide the omitted proofs from Section \ref{sec:knapsack} in the main paper. Namely, we present the missing proofs of the theoretical guarantee for the \chunkgreedy algorithm with \algksm as the subroutine, stated in Theorem \ref{thm:SM_knapsack}, and we present the converting algorithm \convk for transforming an algorithm for SMKP to an algorithm for SCKP, stated in Theorem \ref{thm:convk}.
\subsection{Motivating Applications of SCKP}
\label{appdx:knapsack_application}
In this portion of the appendix, we provide a series of examples to motivate our study of the SCKP problem, where the objective is to find a solution set $S$ which minimizes the total cost while maintaining a certain level of utility ($f(S)\geq\tau$) and a balanced cost constraint across different partitions ($c(S\cap U_j)\leq p_jv$). The motivating examples of this problem include
\begin{itemize}
\item \textbf{Influence Maximization}: In this application, we might want to select a set of nodes with minimum cost (e.g., limited budget funds to be allocated) while ensuring a certain level of influence spread. The cost should also be balanced among each partition of the universe, which is splitted by the demographic or geographic attributes. Different nodes (e.g., influential users or groups) may require different costs to be activated (e.g., through targeted ads or promotions), and thus the cost is non-uniform among different nodes.
\item \textbf{Pretraining Data Selection}: In pretraining data selection, the goal is to select a subset of data points with minimal cost, where costs may reflect computational, labeling, or storage expenses. The problem involves balancing costs across predefined groups in the dataset while ensuring that the utility function (e.g., coverage of diverse features) meets a specified threshold.
\item \textbf{Multi-Agent Task Allocation:} The objective is to find a set of tasks that minimizes the total cost of the assigned tasks while achieving an overall utility or performance of a certain level and a balanced cost across different types of tasks (e.g., delivery, inspection, or cleaning). Tasks have different execution costs depending on complexity, duration, or required resources and thus the cost is nonuniform.
\end{itemize}
\subsection{Clarification of the Problem Definition of SCKP}
\label{appdx:knapsack_prob_explain}
In this section, we provide some illustrations of the problem formulation of SCKP defined in Section \ref{sec:knapsack} in the main paper. First of all, recall that the classical Minimum Cost Submodular Cover (MCSC) studied in previous work \citep{iyer2013submodular,crawford2019efficient} is defined as $\arg\min\{c(S): f(S)\geq\tau\}$ where $c:2^U\rightarrow \mathbb{R}$ is a modular, positive cost function. In our setting, we also want to ensure a balanced budget allocation across different partitions. Therefore, one of the definition of our problem should be $\arg\min\{c(S): f(S)\geq\tau, c(S\cap U_j)\leq p_j c(S), \forall j\in[N]\}$.
However, the problem defined above can have feasibility issues in many cases. In particular, the constraint of $c(S\cap U_j)\leq p_jc(S)$ for each $j\in[N]$ can be really hard to satisfy, and can even render the problem infeasible. For example, if we set $p_j=1/N$ for each $j\in[N]$, and that for each $s\in U_{j_1}$ $c(s)=\pi$, and each $s\in U_{j_2}$, $c(s)=1$. From the definition of $(P1)$, we can see that there is no subset $S\subseteq U$ that satisfies $c(S\cap U_j)\leq p_jc(S)$ for each $j\in[N]$.
To solve this feasibility issue, we can relax the constraint on the balanced solution such that it can be slightly broken by the cost of a single element. Let us define $c_j=\max\{c(s):s\in U_j\}$ to be the maximum singleton cost within the partition $U_j$. It then follows that the definition of the relaxed problem should be $\arg\min\{c(S): f(S)\geq\tau, c(S\cap U_j)\leq p_j c(S)+c_j, \forall j\in[N]\}$
For notation simplicity, we use $(P1)$ to denote this problem, i.e.,
\begin{align}
\label{eqn:problem_p1}
(P1): \qquad &\min_{S\subseteq U}c(S)\nonumber\\
& f(S)\geq\tau\nonumber\\
&c(S\cap U_j)\leq p_jc(S)+c_j,\qquad \forall j\in[N].
\end{align}
To solve this problem, we can slightly relax the constraint on the cost by introducing another variable $\mu$ to the constraint, i.e., $c(S\cap U_j)\leq p_j\mu c(S)$ for each $j\in[N]$. Notice that here we also want to minimize the level of breaking the constraint, to do that, we replace the objective function from minimizing $c(S)$ to $\mu c(S)$ in the optimization problem, Next, by replacing the term $\mu c(S)$ with $v$, we obtain the definition of the SCKP problem. Additionally, compared with the optimization problem defined in $(P1)$, the problem defined in SCKP in (\ref{eqn:opt_prob_knapsack}) preserves the feasibility as long as the threshold $\tau$ satisfies $f(U)\geq\tau$.
\begin{align}
\label{eqn:problem_p2}
(P2): \qquad &\min_{S\subseteq U}v(S)\nonumber\\
& f(S)\geq\tau\nonumber\\
&c(S\cap U_j)\leq p_jv,\qquad \forall j\in[N],
\end{align}
Let us define the optimal solution and optimal value of $(P1)$ as $OPT_{P1}$ and $v_{OPT}$ respectively, and we denote the optimal solution of SCKP defined as $OPT$. It is worth noting that the optimal solution in the optimization problem $(P1)$ has a similar quality to our SCKP problem $(P2)$. In particular, we have that the optimal value of $P1$ and $(P2)$ satisfies the following lemma:
\begin{lemma}
\label{lem:SCKP_prob_equivalence} The optimal value of $(P1)$ and $(P2)$ satisfies
\begin{enumerate}
\item $ v_{OPT}\leq c(OPT_1)+\max_{i\in[N]}\frac{c_i}{p_i}$.
\item When the optimal value of $(P2)$ satisfies that $ p_jv_{OPT}\leq c(U_j)$, we have that $c(OPT_1)\leq \alpha v_{OPT}+\sum_{i\in[N]}c_j$, where $\alpha=\sum_{j\in[N]}p_j$.
\end{enumerate}
\end{lemma}
\begin{proof}
We first prove the first part of the lemma. Since $OPT_1$ is feasible for problem $(P1)$, it must satisfy all constraints of $(P1)$. In particular, for each $j \in [N]$, we have:
\begin{align}
c(OPT_1\cap U_j)&\leq p_jc(OPT_1)+c_j\notag\\
&=p_j\Big(c(OPT_1)+\frac{c_j}{p_j}\Big)\notag\\
&\leq p_j\Big(c(OPT_1)+\max_{i\in[N]}\frac{c_i}{p_i}\Big)
\end{align}
Setting $v = c(OPT_1) + \max_{i\in[N]}\frac{c_i}{p_i}$, we observe that $S = OPT_1$ satisfies the constraints of the SCKP problem defined in $(P2)$. Therefore, $v_{OPT} \leq v$, proving the first result.
We prove the second result by constructing a set $A$ by the following procedure.
\begin{enumerate}
\item Initialize $A \leftarrow \text{OPT}_2$.
\item For $j = 1$ to $N$ do:
\begin{enumerate}
\item While $c(A \cap U_j) \leq p_j v_{\text{OPT}}$:
\begin{enumerate}
\item $x \leftarrow \arg\min_{x' \in U_j \setminus A} c(x')$
\item $A \leftarrow A \cup \{x\}$
\end{enumerate}
\end{enumerate}
\end{enumerate}
Notice that for each $j\in[N]$, $c(OPT_2\cap U_j) \leq p_jv_{OPT}$, and that $c(U_j)\geq p_jv_{OPT}$. Therefore, upon the termination of the above procedure, set $A$ satisfies that
\begin{align}
p_jv_{OPT}\leq c(A\cap U_j)\leq p_jv_{OPT}+c_j
\end{align}
It then follows that $\sum_{j\in[N]}p_jv_{OPT}\leq\sum_{j\in[N]}c(A\cap U_j)$, and thus $\alpha v_{OPT}\leq c(A)$. Therefore, for each $j\in[N]$, we have that
\begin{align}
c(A\cap U_j)\leq p_jv_{OPT}+c_j\leq p_jc(A)+c_j
\end{align}
It implies that $A$ is feasible for problem $(P1)$, therefore, we can conclude
$c(OPT_1)\leq c(A)\leq\sum_{j\in[N]}p_jv_{OPT}+c_j\leq \alpha v_{OPT}+\sum_{j\in[N]}c_j$.
\end{proof}
Besides, we want to point out that another benefit of the definition of our problem is that it preserves the dual relationship between the SCKP problem and the SMKP problem, which is defined as $\arg\max {f(S): \sum_{s \in X \cap U_j} c(s) \leq p_j v}$. In particular, here the variable $v$ in SMKP also serves as the budget of the cost constraint. This property facilitates our application of converting theorems, which is used to convert bicriteria algorithms for SMF to algorithms for SCF.
\subsection{Proof of Theorem \ref{thm:SM_knapsack}}
\label{appdx:proof_SMKP}
In this portion of the appendix, we present the missing proofs of the theoretical guarantee for the \chunkgreedy algorithm with \algksm as the subroutine. The theorem statement is provided in Theorem \ref{thm:SM_knapsack}.
\SMKthm*
Let us denote the solution set after adding the $l$-th element in $j$-th subgroup during the $i$-th round of the outer for loop from Line \ref{line:knapsack_block_starts} to Line \ref{line:knapsack_block_ends} in Algorithm \ref{procd:ksm} as $S_{i,j,l}$, and we define the solution set after completing adding all the elements in the $j$-th subgroup during the $i$-th round in in Algorithm \ref{procd:ksm} as $S_{i,j}$. Before we prove Theorem \ref{thm:SM_knapsack}, we prove the result in the following lemma.
\begin{lemma}
\label{lem:knapsack}
Let $S_{i,j}$ be the solution set of the algorithm \blockksc in Algorithm \ref{procd:ksm} after completing adding all the elements in the $j$-th subgroup during the $i$-th round in in Algorithm \ref{procd:ksm}, then we would get that
\begin{align*}
f(S_{i,j})-f(S_{i,j-1})\geq \Delta f(S_{i,j},OPT_j)
\end{align*}
where $OPT_j:=OPT\cap U_j$ is the intersection of the optimal solution set $OPT$ and the $j$-th partition $U_j$, and that
\begin{align*}
B_j\leq c(S_{i,j}/S_{i,j-1})\leq 2B_j.
\end{align*}
Here $B_j:=p_jv$.
\end{lemma}
\begin{proof}
Let $A_{i,j,l}$ be the set $A$ after adding the $l$-th element to the subgroup $j$ in the iteration $i$, and let $s_{i,j,l}$ be the $l$-th element $s$ added to the set $A$ during the $i$-th outer loop in the subgroup $j$ in Algorithm \ref{procd:ksm}. It then follows that for any element $o\in OPT_j$,
\begin{align*}
\frac{\Delta f(S_{i,j-1}\cup A_{i,j,l-1},s_{i,j,l})}{c(s_{i,j,l})}\geq\frac{\Delta f(S_{i,j-1}\cup A_{i,j,l-1},o)}{c(o)}.
\end{align*}
By rearranging the above inequality, we can get
\begin{align*}
c(o)\Delta f(S_{i,j-1}\cup A_{i,j,l-1},s_{i,j,l})\geq c(s_{i,j,l})\Delta f(S_{i,j-1}\cup A_{i,j,l-1},o).
\end{align*}
Summing over all $o\in OPT_j$ and by submodularity, we can get
\begin{align*}
c(OPT_j)\Delta f(S_{i,j-1}\cup A_{i,j,l-1},s_{i,j,l})\geq c(s_{i,j,l})\Delta f(S_{i,j-1}\cup A_{i,j,l-1},OPT_j)
\end{align*}
Let us denote the total number of iterations in Algorithm \ref{procd:ksm} as $T$. By submodularity, it then follows that
\begin{align*}
c(OPT_j)\Delta f(S_{i,j-1}\cup A_{i,j,l-1},s_{i,j,l})\geq c(s_{i,j,l})\Delta f(S_{i,j-1}\cup A_{i,j,T},OPT_j).
\end{align*}
Since
\begin{align*}
\Delta f(S_{i,j-1}\cup A_{i,j,l-1},s_{i,j,l})=f(S_{i,j-1} \cup A_{i,j,l})-f(S_{i,j-1} \cup A_{i,j,l-1}),
\end{align*}
we can sum over all $l\in[T]$ and get
\begin{align*}
c(OPT_j)\{f(S_{i,j-1} \cup A_{i,j,T})-f(S)\}\geq c(A_{i,j,T})\Delta f(S_{i,j-1} \cup A_{i,j,T},OPT_j).
\end{align*}
Since $S_{i,j-1} \cup A_{i,j,T} = S_{i,j}$, we have
\begin{align*}
c(OPT_j)\{f(S_{i,j})-f(S)\}\geq c(A_{i,j,T})\Delta f(S_{i,j},OPT_j).
\end{align*}
By the stopping condition of Algorithm \ref{procd:ksm}, we have that $c(A_{i,j,T-1})\leq B_j$, therefore,
\begin{align*}
B_j\leq c(A_{i,j,T})\leq 2B_j
\end{align*}
Since $c(OPT_j)\leq B_j$, it then follows that
\begin{align*}
f(S_{i,j})-f(S)\geq\Delta f(S_{i,j},OPT_j).
\end{align*}
We can then conclude the proof by the fact that $A_{i,j,T}= S_{i,j}/S_{i,j-1}$.
\end{proof}
With Lemma \ref{lem:knapsack}, we can prove the result of Theorem \ref{thm:SM_knapsack} as follows.
\begin{proof}
Let us denote the solution set after completing the $i$-th round in Algorithm \ref{procd:ksm} as $S_{i}$, i.e., $S_i=S_{i,N}$. By the result in Lemma \ref{lem:knapsack}, it then follows that
\begin{align*}
f(S_{i,j})-f(S_{i,j-1})\geq\Delta f(S_{i,j},OPT_j)\geq \Delta f(S_{i},OPT_j),
\end{align*}
where the second inequality follows from submodularity. Summing over all $j\in[N]$, we would get
\begin{align*}
f(S_i)-f(S_{i-1})\geq \sum_{j\in[N]}\Delta f(S_{i},OPT_j)\geq \Delta f(S_{i},OPT)
\end{align*}
Therefore, $f(S_i)\geq\frac{f(S_{i-1})+f(OPT)}{2}$. By induction, we have that the final output solution set $S$ satisfies
\begin{align*}
f(S)=f(S_\phi)\geq(1-\epsilon)f(OPT).
\end{align*}
Notice that $S=\cup_{i=1}^\phi S_i/S_{i-1}$. From Lemma \ref{lem:knapsack}, we can get $c((S_i/S_{i-1})\cap U_j)\leq 2p_jv$. Therefore
$c(S\cap U_j)=\sum_{i=1}^\phi c((S_i/S_{i-1})\cap U_j)\leq\frac{2\ln\frac{1}{\epsilon}}{\ln 2}p_jv$.
\end{proof}
\subsection{Theoretical Analysis of Algorithm \ref{alg:convk}}
\label{appdx:convk}
In this portion of the appendix, we present the converting algorithm \convk for transforming an algorithm for SMKP to an algorithm for SCKP. The pseudocode is described in Algorithm \ref{alg:convk}. The theoretical guarantee of \convk is provided in Theorem \ref{thm:convk}.
\begin{theorem}
\label{thm:convk}
Suppose that we have an algorithm \algksm for SMKP, and given budget $v$, \algksm is guaranteed to return a set $S$ such that
$f(S)\geq \gamma f(OPT_{SM})$ and $c(S\cap U_j)\leq\beta p_jv$, in time $T(n)$, where $OPT_{SM}$ is the optimal solution of SMKP.
Then the algorithm \convk using \algksm as a subroutine returns a set $S$ and a value $v_S$ in time $\mathcal{O}(\log_{1+\alpha}(\frac{c_{\max}n}{c_{\min}})T(n))$ such that $v_S\leq \beta(1+\alpha) v_{OPT}$,
$c(S\cap U_j)\leq p_jv_S$ and
$f(S)\geq \gamma\tau.$ Here $v_{OPT}$ is the optimal value of SCKP. $c_{\max}$ and $c_{\min}$ are the maximum and minimum values of the cost of a single element respectively.
\end{theorem}
\begin{algorithm}[t]
\caption{\convk}
\label{alg:convk}
\textbf{Input}: $\alpha$, $\epsilon$\\
\textbf{Output}: $S\subseteq U$
\begin{algorithmic}[1]
\State $v_g\gets (1+\alpha)c_{\min}$, $S\gets\emptyset$
\While{$f(S)<\gamma\tau$}
\State $S\gets$ Algorithm for SMKP run with budget parameter $v=v_g$
\State $v_g\gets (1+\alpha)v_g$
\EndWhile
\State $v_S=\beta v_{g}$
\State \textbf{return} $S$, $v_S$
\end{algorithmic}
\end{algorithm}
\begin{proof}
Let $OPT$ be the optimal solution to the instance of SCKP.
Consider the iteration of \convk where $v_g$ has just increased above $v_{OPT}$, i.e., $v_{OPT}\leq v_{g}\leq (1+\alpha)v_{OPT}$. Then we run \algksm with budget $v_{OPT}\leq v_{g}\leq (1+\alpha)v_{OPT}$. Then by the assumptions on \algksm we have that
\begin{align}
\label{eqn:knapsack_converting}
f(S)\geq \gamma f(OPT_{SM}).
\end{align}
Notice that the optimal solution $OPT$ for SCKP satisfies that $c(OPT\cap U_j)\leq p_jv_{OPT}\leq p_jv_{g}$. It then follows that $OPT$ is feasible for the SMKP problem with input $v_g$. Let us denote the optimal solution of SMKP as $OPT_SM$. Then we have that $$f(OPT_{SM})\geq f(OPT).$$
Since $OPT$ is the optimal solution for SCKP, then
$$f(OPT_{SM})\geq f(OPT)\geq\tau.$$
Combining the above inequality with the result in (\ref{eqn:knapsack_converting}), we can get that $f(S)\geq \gamma\tau$. Therefore, the algorithm stops before $v_g$ reaches $(1+\alpha)v_{OPT}$.
The cost of each partition would satisfy
\begin{align*}
c(S\cap U_j)\leq\beta p_j v_g\leq(1+\alpha)\beta p_j v_{OPT}
\end{align*}
The proof is completed by setting $v_S=\beta v_{g}$.
\end{proof}
\section{Appendix for Section \ref{sec:fair}}
\label{appdx:fair}
In this section, we present the missing proofs of the theoretical results of \fairgreedy from Section \ref{sec:fair}. The theorem statement is provided in Theorem \ref{thm:fair_greedy}.
\fairthm*
\begin{proof}
Denote the solution set after the $i$-th chunk as $S_i$, and we denote the subset $B$ after the $i$-th chunk as $ B_i$, then it follows that $S_i=S_{i-1}\cup\{B_i\}$. We can prove the following lemma.
\begin{restatable}{lemma}{lemfair}
\label{lem:fair}
For any $i\leq\frac{\ln1/\epsilon}{\ln 2}$, the solution set $S_i$ satisfies that
\begin{align*}
f(S_i)-f(S_{i-1})\geq f(OPT)-f(S_{i})
\end{align*}
and that
\begin{align*}
&|S_i\cap U_c|\leq u_ci,\\
&\sum_{c\in [N]}\max\{|S_i\cap U_c|,l_ci\}\leq ki.
\end{align*}
\end{restatable}
\begin{proof}
Let us denote the solution set after adding the $j$-th element to the solution set $S$ during the $i$-th chunk as $S_{i,j}$. In addition, we denote the $j$-th element adding to $B_i$ as $b_j$, and that $B_{i,j}=(b_1,\ldots,b_{j-1})$. By the definition of matroid, there exists a mapping from the set $B$ to the optimal solution $OPT=\{o_1,...,o_\kappa\}$ s.t. $B_{i,j}\cup\{o_j\}\in\mathcal{P}_{fair}$.
\begin{align*}
f(S_{i,j})-f(S_{i,j-1})\geq f(S_{i,j-1}\cup\{o_i\})-f(S_{i,j-1})\geq \Delta f(S_{i,\kappa},o_i)
\end{align*}
Summing over all $j$, it follows that
\begin{align*}
f(S_{i,\kappa})-f(S_{i,0})\geq \sum_{i=1}^\kappa\Delta f(S_{i,\kappa},o_i)\geq f(OPT)-f(S_{i,0}).
\end{align*}
Since $f(S_{i,\kappa})=f(S_{i+1,0})=f(S_{i+1})$,
\begin{align*}
f(S_{i+1})-f(S_i)\geq f(OPT)-f(S_{i+1})
\end{align*}
Next, we prove the result on the size of the solution set $S$. When $i=1$, $S_1=B_1$. By the fact that $B_1\in\mathcal{M}_{fair}$, we have that the result in the lemma holds. Let us assume that the result in the lemma holds for $i$. Then for $i+1$, we have that $|S_{i+1}\cap U_c|=|S_{i}\cup B_i\cap U_c|\leq|S_{i}\cap U_c|+|B_i\cap U_c|\leq u_c(i+1)$. For the total cardinality constraint,
\begin{align*}
\max\{|S_{i+1}\cap U_c|,l_c(i+1)\}&\leq\max\{|S_{i}\cap U_c|+|B_{i}\cap U_c|,l_c(i+1)\}\\
&\leq\max\{|S_{i}\cap U_c|,l_ci\}+\max\{|B_{i}\cap U_c|,l_c\}
\end{align*}
where the first inequality comes from the fact that $S_{i+1}=S_i\cup B_i$. The second inequality is due to the inequality of $\max\{a+b,c+d\}\leq\max\{a,c\}+\max\{b,d\}$.
It then follows that
\begin{align*}
\sum_{c\in[N]}\max\{|S_{i+1}\cap U_c|,l_c(i+1)\}\leq\sum_{c\in[N]}\max\{|S_{i}\cap U_c|,l_ci\}+\sum_{c\in[N]}\max\{|B_{i}\cap U_c|,l_c\}\leq \kappa(i+1)
\end{align*}
\end{proof}
Next, by leveraging this Lemma \ref{lem:fair}, we can prove the results in Theorem \ref{thm:fair_greedy}.
Denote $\phi=\frac{\ln1/\epsilon}{\ln 2}$, then by the Lemma, we have that
\begin{align*}
&|S_\phi\cap U_c|\leq u_c\phi,\\
&\sum_{c\in [N]}\max\{|S_\phi\cap U_c|,l_c\phi\}\leq k\phi.
\end{align*}
Since $f(S_{i})\geq\frac{f(OPT)+f(S_{i-1})}{2}$, by induction, it follows that
\begin{align*}
f(S_\phi)\geq (1-\frac{1}{2^{\phi}})f(OPT)=(1-\epsilon)f(OPT).
\end{align*}
\end{proof}
% Next, we prove the following lemma regarding the result in the \blockmono algorithm.
\section{Submodular Maximization under Partition Matroid Constraint}
% \section{Appendix for monotone and nonmonotone case being moved from the main paper}
In the previous sections and in the main paper, we demonstrated that block-greedy algorithms can be effective for solving submodular cover problems under partition-based constraints. Interestingly, this block-greedy approach also proves to be valuable in designing algorithms for submodular maximization problems. In this section, we introduce \chunkgreedy, a novel algorithmic framework tailored for submodular maximization subject to a partition matroid constraint.
\chunkgreedy proceeds by greedily adding blocks—i.e., sets of elements—to the solution. Our algorithms improve upon existing methods in both solution quality and query complexity.
This section is structured as follows. We first present our main results in Sections \ref{section:framework}, \ref{sec:mono}, and \ref{sec:nonmono}. Section \ref{section:framework} introduces the block-greedy framework that underpins the algorithms discussed throughout. Then, we address two specific settings: monotone submodular maximization with a partition matroid constraint (monotone SMP) in Section \ref{sec:mono}, and nonmonotone submodular maximization with a partition matroid constraint (nonmonotone SMP) in Section \ref{sec:nonmono}.
Finally, we include additional content and discussions in Section \ref{appdx:hardness} and Section \ref{appdx:mono_result_discuss}. Section \ref{appdx:hardness} provides the missing discussion and proof of Theorem \ref{thm:chunkgreedy} from Section \ref{sec:tight}, while Section \ref{appdx:mono_result_discuss} elaborates on omitted content from Section \ref{sec:mono_alg_descript}.
\subsection{Block Greedy Framework}
\label{section:framework}
\input{Sections/blockgreedy}
\subsection{Monotone \prob}
\label{sec:mono}
\input{Sections/classicpartition}
\subsection{Nonmonotone \prob}
\label{sec:nonmono}
In this section, we propose the algorithm for the problem of nonmonotone \problong (SMP). The proposed algorithm follows the framework in Algorithm \ref{alg:partition_greedy_general} with $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$, and the subroutine algorithm \blocknonmono is described in Algorithm \ref{alg:blocknonmono}. Here the parameter $r_j:=\lfloor k_j/\phi\rfloor$. The algorithm uniformly selects an element from the set of elements with the top $r_j\phi$ marginal gain to add to the solution set. The intuition behind the \blocknonmono algorithm is similar to that of the Random Greedy algorithm proposed in \cite{buchbinder2014submodular}. However, in \blocknonmono, the size of the candidate set considered for inclusion in the solution is adjusted to $r_j\phi$ to ensure an important result that $\mE [f(S_{i}\cup OPT')]\geq(1-\frac{1}{\phi})^if(OPT')$ where $\mathcal{P}'=\{S\subseteq U:|S\cap U_i|\leq r_i\phi, \forall i\in[N]\}$ and $OPT'=\arg\max_{S\in\mathcal{P}'}f(S)$.
\begin{algorithm}[t] \caption{\blocknonmono($S$, $i$, $j$)}\label{alg:blocknonmono}
\begin{algorithmic}[1]
\State \textbf{Input: }$S,i,j$
% \State \textbf{Output: }$S\in U$
\For{$l=1$ \textbf{to} $r_j $}
\State Let $M\subseteq U/S$ be a set of size $r_j\phi$ maximizing $\sum_{x\in M}\Delta f(S,x)$.
\State $u\gets$ uniformly sample an element from $M$
\State $S\gets S\cup \arg\max_{x\in U_j}\Delta f(S,x)$\label{line:greedy_nonmono}
\EndFor
% \textbf{return }$S$
\end{algorithmic}
\end{algorithm}
Below we present the main result of \chunkgreedy for the problem of nonmonotone \prob.
\begin{restatable}{theorem}{thmchunkgreedynonmono}
\label{thm:chunkgreedynonmono}
Suppose that \chunkgreedy is run for an instance of nonmonotone \prob with the subroutine algorithm \blocknonmono as described in Algorithm \ref{alg:blockmono},
then \chunkgreedy
outputs a solution $S$ that satisfies an approximation ratio of
$\frac{1}{e}-\frac{3}{e(\phi+1)}$ where $\phi=\sqrt{\min_{i\in[N]}k_i}-1$ in expectation.
\end{restatable}
Notice that the approximation ratio is close to $1/e$, which matches the bound of the random greedy algorithm for submodular maximization under cardinality constraint, with the difference bounded by $\mathcal{O}(\frac{1}{\sqrt{k_{\min}}})$, where $k_{\min}=\min_{i\in[N]}k_i$. In this sense, the proposed algorithm achieves an approximation ratio for Nonmonotone \prob that bridges the gap between submodular maximization over cardinality constraint and partition matroid constraint. The proof of Theorem \ref{thm:chunkgreedynonmono} is provided below.
% \label{appdx:nonmono}
% In this section, we present the detailed proofs of the theorem for the nonmonotone SMP problem, which is provided in Theorem \ref{thm:chunkgreedynonmono}.
% \thmchunkgreedynonmono*
Let us define $\mathcal{P}'=\{S\subseteq U:|S\cap U_i|\leq r_i\phi, \forall i\in[N]\}$ and we denote the optimal solution set of $OPT'=\arg\max_{S\in\mathcal{P}'}f(S)$. First of all, we prove the following lemma for the subroutine algorithm \blocknonmono.
\begin{lemma}
\label{lem:nonmono}
Let us denote the input and output of Algorithm \ref{alg:blocknonmono} as $S$ and $S'$ respectively, then it follows that
\begin{align*}
\mE[f(S')-f(S)]\geq\frac{\mE[\Delta f(S',OPT_j)]}{\phi},
\end{align*}
where $OPT_j=OPT'\cap U_j$.
\end{lemma}
\begin{proof}
Let us denote the solution set after adding the $l$-th element as $S^{(l)}$. Similar to the proof of Theorem \ref{thm:chunkgreedy} for the monotone \prob, we have that
\begin{align*}
\mE [f(S^{(l)})-f(S^{(l-1)})]\geq\frac{\sum_{a\in OPT'\cap U_j}\Delta f(S^{(l-1)},a)}{r_j\phi}.
\end{align*}
By submodularity, we would have that
\begin{align*}
\mE [f(S^{(l)})-f(S^{(l-1)})]&\geq\mE [\frac{\sum_{a\in OPT'\cap U_j}\Delta f(S^{(l-1)},a)}{r_j\phi }]\\
&\geq\mE[\frac{\sum_{a\in OPT'\cap U_j}\Delta f(S',a)}{r_j\phi }].
\end{align*}
By summing over all $l\in[r_j]$, it follows that
\begin{align*}
\mE [f(S')-f(S)]&\geq\mE [\frac{\sum_{a\in OPT'\cap U_j}\Delta f(S',a)}{\phi}]\\
&\geq\mE [\frac{\Delta f(S',OPT_j)}{\phi}].
\end{align*}
\end{proof}
Next, leveraging the result in Lemma \ref{lem:nonmono}, we prove the approximation ratio in Theorem \ref{thm:chunkgreedynonmono}.
\begin{proof}
Similar to the proof of Theorem \ref{thm:chunkgreedy}, we denote the solution set after running the subroutine algorithm in $j$-th subgroup during the $i$-th round of the outer for loop in Line \ref{line:outer_for_loop_starts} in Algorithm \ref{alg:partition_greedy_general} as $S_{i,j}$, and we define the solution set after completing the $i$-th round in Algorithm \ref{alg:partition_greedy_general} as $S_{i}$. From Lemma \ref{lem:nonmono}, we have that
\begin{align*}
\mE[f(S_{i,j})-f(S_{i,j-1})]&\geq\frac{\mE[\Delta f(S_{i,j},OPT_j)]}{\phi}\\
&\geq\frac{\mE[\Delta f(S_{i},OPT_j)]}{\phi},
\end{align*}
By summing over all $j\in[N]$, we have
\begin{align*}
\mE [f(S_{i})-f(S_{i-1})]&\geq\frac{\sum_{j=1}^N\mE[\Delta f(S_{i},OPT_j)]}{\phi}\\
&\geq\frac{\mE[\Delta f(S_{i},OPT')]}{\phi}.
\end{align*}
Then it follows that
\begin{align*}
\mE [f(S_{i})-f(S_{i-1})]\geq\mE[\frac{f(S_{i}\cup OPT')-f(S_i)}{\phi}].
\end{align*}
Notice that by the greedy selection step, for each group $j$ and each element $a\in OPT'\cap U_j$ appears in $S_i$ with probability at most $1-(1-\frac{1}{r_j\phi })^{r_ji}$. Since $(1-\frac{1}{x})^x$ increases with $x$ in the range of $[1,+\infty)$, we have that $(1-\frac{1}{r_j\phi })^{r_j\phi }\geq (1-\frac{1}{\phi})^\phi$. Therefore, we would get $1-(1-\frac{1}{r_j\phi })^{r_ji}\leq 1-(1-\frac{1}{\phi})^i$. From Lemma 2.2 in \cite{buchbinder2014submodular}, we can conclude that
\begin{align*}
\mE [f(S_{i}\cup OPT')]\geq(1-\frac{1}{\phi})^if(OPT').
\end{align*}
By rearranging the above inequality, we can get that
\begin{align*}
\mE [f(S_{i})]\geq\frac{\phi}{\phi+1}\mE [f(S_{i-1})]+\frac{1}{\phi+1}(1-\frac{1}{\phi})^if(OPT').
\end{align*}
By induction, we have that the output solution set satisfies that
\begin{align*}
\mE [f(S)]&=\mE [f(S_{\phi})]\nonumber\\
&\geq\frac{1}{\phi+1}\sum_{i=1}^\phi(\frac{\phi}{\phi+1})^{\phi-i}(1-\frac{1}{\phi})^if(OPT')\nonumber\\
&\geq\frac{1}{\phi+1}\sum_{i=1}^\phi(\frac{\phi-1}{\phi})^{\phi-i}(1-\frac{1}{\phi})^if(OPT')\nonumber\\
&\geq\frac{\phi}{\phi+1}(1-\frac{1}{\phi})^\phi f(OPT')\geq\frac{1}{e}(1-\frac{2}{\phi+1})f(OPT').
\end{align*}
where the last inequality follows from the fact that $(1-\frac{1}{\phi})^{\phi-1}\geq e^{-1}$ for any $\phi>1$. From the definition that $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$, it then follows that $k_i- \phi\lfloor k_i/\phi\rfloor\leq \lfloor k_i/\phi\rfloor$ for any $i\in[N]$. Therefore, from the result in Lemma \ref{lem:partial_OPT}, we get that
\begin{align}
\label{eqn:nonmono_partial_OPT}
f(OPT')\geq\frac{\phi}{\phi+1}f(OPT)
\end{align}
where $OPT$ is the optimal solution to the submodular maximization problem $\arg\max_{S\in\mathcal{P}}f(S)$. By combining (\ref{eqn:nonmono_induction}) and (\ref{eqn:nonmono_partial_OPT}) together, we can prove the result in the Lemma.
\end{proof}
\begin{lemma}
\label{lem:partial_OPT}
Suppose the ground set $U$ is divided into $N$ disjoint subgroups $U_1$, $U_2$,..., $U_N$, then for any partition matroid $\mathcal{P}=\{S\subseteq U:|S\cap U_j|\leq k_j, \forall j\in[N]\}$, let us define the matroid $\mathcal{P}'=\{S\subseteq U:|S\cap U_j|\leq \lfloor \frac{k_j}{c}\rfloor c, \forall j\in[N]\}$ for some positive integer $c$. If for any $j\in[N]$, it satisfies that $\lfloor \frac{k_j}{c}\rfloor\geq k_j-c\lfloor \frac{k_j}{c }\rfloor$, then it follows that for any submodular function $f$, we have
\begin{align*}
\max_{S\in\mathcal{P}'}f(S)\geq\frac{c }{c +1}\max_{S\in\mathcal{P}}f(S).
\end{align*}
\end{lemma}
\begin{proof}
For notation simplicity, we also define $r_j=\lfloor \frac{k_j}{c }\rfloor$ where $j\in[N]$, and we define two matroids $\mathcal{P}_0=\{S\subseteq U:|S\cap U_j|\leq r_j, \forall j\in[N]\}$ and $\mathcal{P}_1=\{S\subseteq U:|S\cap U_j|\leq r_j(c+1), \forall j\in[N]\}$. Let us denote the optimal solution of $\max_{S\in\mathcal{P}_1}f(S)$ as $OPT''$. Then by definition, we can see that $OPT''$ can be divided into $(c+1)$ disjoint subsets $O_1,...,O_{c +1}$ such that each $O_i\in\mathcal{P}_0$. Without loss of generality, we assume that the subsets are chosen greedily such that the index satisfies $\Delta f(\cup_{j\in[i-1]}O_j,O_i)\geq\Delta f(\cup_{j\in[i-1]}O_j,O_l)$ for any $1\leq i\leq c$ and $l>i$. It then follows that by submodularity, $\Delta f(\cup_{j\in[i-1]}O_j,O_i)\geq\Delta f(\cup_{j\in[i-1]}O_j,O_l)\geq\Delta f(\cup_{j\in[l-1]}O_j,O_l)$ for any $l>i$. Therefore,
\begin{align*}
f(OPT'')-f(\cup_{j\in[c]}O_j)&=\Delta f(\cup_{j\in[c]}O_j,O_{c +1})\\
&\leq\frac{\sum_{i\in[c]}\Delta f(\cup_{j\in[i-1]}O_j,O_{i})}{c }\\
&=\frac{f(\cup_{j\in[c]}O_j)-f(\emptyset)}{c }.
\end{align*}
By rearranging the above inequality, we would get that $f(\cup_{j\in[c]}O_j)\geq\frac{c }{c +1}f(OPT'')$. Since $\cup_{j\in[c]}O_j\in\mathcal{P}'$, then we have that $\max_{S\in\mathcal{P}'}f(S)\geq\frac{c }{c +1}f(OPT'')$. Notice that $\lfloor \frac{k_j}{c }\rfloor(c+1)\geq k_j$ implies that for any subset $S\in\mathcal{P}$, it also satisfies that $S\in\mathcal{P}_1$. Therefore, $\max_{S\in\mathcal{P}_1} f(S)\geq \max_{S\in\mathcal{P}} f(S)$ and we can conclude the proof.
% Define a series of set $O_0,O_1,...,O_c$ such that $O_0=\emptyset$ and $O_i=O_{i-1}\cup\arg\max_{S\in\mathcal{P}_0, S\subseteq OPT/O_{i-1}}\Delta f(O_{i-1},S)$. By the definition, we have that for any $i<j$, $O_j/O_{j-1}\in\mathcal{P}_0$ and that $O_j/O_{j-1}\subseteq OPT/O_{i-1}$. Therefore, it follows that $\Delta f(O_{i-1},O_j/O_{j-1})\leq\Delta f(O_{i-1},O_i/O_{i-1})$. By submodularity, we have that $\Delta f(O_{j-1},O_j/O_{j-1})\leq\Delta f(O_{i-1},O_i/O_{i-1})$. Next, we prove that the set $OPT/O_c$ satisfies that $|OPT/O_c\cap U_j|\leq k_j-r_j\phi $. In the case where $f$ is monotone, $|O_c\cap U_j|=r_j\phi $, by the result that $r_j\geq k_j-r_j\phi $, we have that $|OPT/O_c\cap U_j|\leq r_j$ for any $j\in[c]$. If $f$ is nonmonotone, and if there exists some group $j_0$ such that $|OPT/O_c\cap U_{j_0}|>k_{j_0}-r_{j_0}c$, then there exist at least one element $x\in OPT/O_c\cap U_{j_0}$ such that $|O_c\cup\{x\}\cap U_{j_0}|\leq r_{j_0}c$. Since $x\in OPT$ and that $\Delta f(O_c,x)\geq 0$, $x$ should be added in $O_c$ during the construction of $O_1,...,O_c$, which results in a contradiction.
\end{proof}
% \subsection{Appendix for Section \ref{sec:mono}}
% \label{appdx:mono}
\subsection{Proof of Theorem \ref{thm:hardness}}
\label{appdx:hardness}
In this section, we present the omitted proof of Theorem \ref{thm:hardness}. To prove the theorem, we construct a class of instances to demonstrate that the standard greedy algorithm can't achieve an approximation ratio better than $1/2$. We begin by presenting relevant definitions of the set functions and constraints, followed by a formal description of the hardness instance.
Suppose the ground set $U$ is partitioned into $N$ disjoint subsets $U_1, U_2, \dots, U_N$, with each subset $U_i$ containing $2k_i$ elements. Define a set function $t: U \to 2^M$, where $M$ is a finite set, and let $c: 2^M \to \mathbb{R}_+$ be a non-negative, monotone, modular function. We define the submodular function $f: 2^U \to \mathbb{R}_+$ as follows:
\[
f(S) = c\left(\bigcup_{s \in S} t(s)\right)=\sum_{x\in\bigcup_{s \in S} t(s)}c(x).
\]
The partition matroid is denoted as $\mathcal{P}=\{S\subseteq U:|S\cap U_i|\leq k_i, \forall i\in[N]\}$. Without loss of generality, we assume the partition matroid constraint parameters satisfy that $k_1\leq k_2\leq,...\leq k_N$. For notation simplicity, let us denote the $j$-th element in group $U_i$ as $s_j^{(i)}$. The hardness example is defined as follows. Suppose $\epsilon$ is a constant such that $\epsilon\in(0,1/2)$, then
\begin{enumerate}
\item \textbf{Set Assignments in $U_1 $}: In the first partition $U_1$, assume that the sets $t(s_{j_1}^{(1)})$ and $t(s_{j_2}^{(1)})$ are disjoint for any distinct $j_1, j_2$, i.e., $t(s_{j_1}^{(1)}) \cap t(s_{j_2}^{(1)}) = \emptyset $.
\item \textbf{Function Values in $U_1 $}: Set the modular function values for the elements in $U_1 $ as follows:
\begin{itemize}
\item For $j \leq k_1 $, $c(t(s_j^{(1)})) = \frac{1}{2} + \epsilon $.
\item For $k_1 < j \leq 2k_1 $, $c(t(s_j^{(1)})) = \frac{1}{2} $.
\end{itemize}
\item \textbf{Set Assignments and Function Values for $i > 1 $}: For each $i_1, i_2 $ such that $1 < i_1, i_2 \leq N $, define:
\begin{itemize}
\item If $j \leq k_1 $, then $t(s_{j}^{(i_1)}) = t(s_{j}^{(i_2)}) \subseteq t(s_{j}^{(1)}) $.
\item If $j > k_1 $, then $t(s_{j}^{(i_1)}) = t(s_{j}^{(i_2)}) = t(s_{1}^{(i_1)}) $.
\item Set $c(t(s_{j}^{(i)})) = \frac{1}{2}$ for any $i > 1 $ and $j \in [2k_i] $.
\end{itemize}
\end{enumerate}
Given this construction, the standard greedy algorithm proceeds as follows: for the first $k_1$ steps, the algorithm would add the elements $s_1^{(1)},s_2^{(1)},...,s_{k_1}^{(1)}$ from $U_1 $, yielding a marginal gain of $\frac{1}{2} + \epsilon$ per step. Thus, the total value after these steps is $\frac{k_1}{2} + k_1 \epsilon $. After the first $k_1$ steps, the algorithm can only select elements from partitions $U_i$ where $i > 1$, with a marginal gain of $0$ at each step due to the structure of $f$ under the current set assignments. Therefore, the value of the submodular objective returned by the standard greedy algorithm is $\frac{k_1}{2}+k_1\epsilon$.
Next, we consider the optimal solution of the constructed example. Notice that for any of the partitions $U_i$ such that $i>1$, $f(U_i)=k_1/2$. Besides, by the construction, we have that $\bigcup_{s\in U_1}t(s)=\bigcup_{s\in U_2}t(s)$ for any $i_1,i_2>1$. Therefore, $f(\bigcup_{i>1}U_i)=k_1/2$. It then follows that for any subset $S\subseteq U$, \begin{align}\label{eqn:hardness_exp_proof}
f(S\cap\bigcup_{i>1}U_i )\leq k_1/2.
\end{align}
Next, we claim that the optimal solution should satisfy that $f(OPT)\leq k_1 $. We prove the claim by considering the following cases. For any $S\in\mathcal{P}$, then
\begin{enumerate}
\item If $f(S\cap\bigcup_{i>1}U_i )= k_1/2$, which means $\bigcup_{s\in S\cap\bigcup_{i>1}U_i} t(s) =\bigcup_{j\leq k_1} t(s_{j}^{(2)})$, then for each $j\leq k_1$, the marginal gain of adding the $j$-th element in the first partition to the the set $S\cap\bigcup_{i>1}U_i$ satisfies that $\Delta f(S\cap\bigcup_{i>1}U_i ,s_{j}^{(1)}) = c(t(s_{j}^{(1)}))-c(t(s_{j}^{(2)}))=\epsilon$ , and for each $j> k_1$, $\Delta f(S\cap\bigcup_{i>1}U_i ,s_{j}^{(1)})=f(s_{j}^{(1)})=1/2$ . It then follows that $\Delta f(S\cap\bigcup_{i>1}U_i ,S\cap U_1)\leq\sum_{s\in S\cap U_1}\Delta f(S\cap\bigcup_{i>1}U_i ,s)\leq k_1/2$. Therefore, $f(S)=\Delta f(S\cap\bigcup_{i>1}U_i ,S\cap U_1)+ f(S\cap\bigcup_{i>1}U_i )\leq k_1$.
\item If $f(S\cap\bigcup_{i>1}U_i )< k_1/2$, then we have that there exists at least one element $s_j^{(2)}$ for $j\leq k_1$ such that the set $t(s_{j}^{(2)})\notin \bigcup_{s\in S\cap\bigcup_{i>1}U_i} t(s) $ . Let us denote $E=\{j\leq k_1:t(s_{j}^{(2)})\notin S\cap\bigcup_{i>1}U_i \}$. It then follows that $f(S\cap\bigcup_{i>1}U_i )=\frac{k_1-|E|}{2}$. For each $j\leq k_1$, if $j\in E$, $\Delta f(S\cap\bigcup_{i>1}U_i ,s_{j}^{(1)}) = c(t(s_{j}^{(1)}))=1/2+\epsilon$, if $j\notin E$, then $\Delta f(S\cap\bigcup_{i>1}U_i ,s_{j}^{(1)}) = c(t(s_{j}^{(1)}))-c(t(s_{j}^{(2)}))=\epsilon$. For each $j> k_1$, $\Delta f(S\cap\bigcup_{i>1}U_i ,s_{j}^{(1)})=f(s_{j}^{(1)})=1/2$. Similarly, we have that $\Delta f(S\cap\bigcup_{i>1}U_i ,S\cap U_1)\leq\sum_{s_{j}^{(1)}\in S\cap U_1}\Delta f(S\cap\bigcup_{i>1}U_i ,s_{j}^{(1)})\leq |E|(1/2+\epsilon)+(k_1-|E|)/2=|E|\epsilon+k_1/2$. Therefore, we can conclude that $f(S)=\Delta f(S\cap\bigcup_{i>1}U_i ,S\cap U_1)+ f(S\cap\bigcup_{i>1}U_i )\leq k_1 -|E|(1/2-\epsilon)\leq k_1$.
\end{enumerate}
It then follows that the $f(\text{OPT}) \leq k_1$. Notice that the set $O=\{s_{k_1+1}^{(1)},...s_{2k_1}^{(1)}, s_{1}^{(2)},...,s_{k_1}^{(2)}\}$ achieves an objective value of: $f(O) = k_1$. Therefore, $f(\text{OPT}) = k_1$.
Consequently, the approximation ratio of the standard greedy algorithm should be $\frac{k_1/2+k_1\epsilon}{k_1}=1/2+\epsilon$. When $\epsilon$ approaches $0$, then the approximation ratio goes to $1/2$.
\subsection{Discussion on Theorem \ref{thm:chunkgreedy}}
\label{appdx:mono_result_discuss}
In this portion of the appendix, we illustrate the results of Theorem \ref{thm:chunkgreedy}. First of all, we discuss the difference between the approximation ratio of our proposed algorithm \chunkgreedy and the optimal approximation ratio $1-1/e$ achieved by the previous continuous method. In particular, the difference is $\mathcal{O}(\frac{1}{\sqrt{k_{\min}}})$ with $k_{\min}=\min_{i\in[N]}k_i$.
In fact, we notice that this difference results from the fact that it scales in the order of $\mathcal{O}(\frac{1}{\phi})$. In the algorithm \chunkgreedy with \blockmono as the subroutine in Section \ref{sec:mono}, $\phi$ is set to be $\phi = \lfloor\sqrt{k_{\min}}\rfloor-1$, which is designed to bound the difference between $k_i$ and $\lfloor\frac{k_i}{\phi}\rfloor\phi$ to ensure that the optimal value of the monotone $\max_{S\in\mathcal{P}'}f(S)$ approximates the optimal value of $\max_{S\in\mathcal{P}}f(S)$ where $\mathcal{P}':=\{S\subseteq U:|S\cap U_i|\leq \lfloor\frac{k_i}{\phi}\rfloor\phi,\forall i\in[N]\}$ and $\mathcal{P}:=\{S\subseteq U:|S\cap U_i|\leq k_i,\forall i\in[N]\}$.
This motivates the following result: in some cases, if we can design the parameter $\phi$ such that $\lfloor\frac{k_i}{\phi}\rfloor=\frac{k_i}{\phi}$ for any $i\in[N]$, then the partition matroid $\mathcal{P}=\mathcal{P}'$ and we don't need to bound the difference of $\max_{S\in\mathcal{P}'}f(S)$ and $\max_{S\in\mathcal{P}}f(S)$. Therefore, we can further refine the difference between the approximation ratio of \chunkgreedy and the optimal result of $1-1/e$. The result is stated as follows.
\begin{restatable}{theorem}{thmgreedy-nonuniform}
\label{thm:greedy-gcd}
Suppose that $gcd(k_1,k_2,\ldots,k_N)=c$, and that \chunkgreedy with \blockmono as a subroutine and $\phi=c$ and $r_j=k_j/c$ for each $j\in[N]$ is run for an instance of monotone SMP over partition matroid $\mathcal{P}:=\{S\subseteq U:|S\cap U_i|\leq k_i, \forall i\in[N]\}$,
then \chunkgreedy
outputs a solution set $S$ that satisfies an approximation ratio of $1-1/e-1/c$.
\end{restatable}
\begin{proof}
Following the similar proof of Theorem \ref{thm:chunkgreedy}, we can get that the output solution set $S$ satisfies
\begin{align*}
f(S)-f(\emptyset)\geq (1-(\frac{\phi}{\phi+1})^\phi)f(OPT'),
\end{align*}
where $OPT'=\arg\max_{S\in\mathcal{P}'}f(S)$ and $\mathcal{P}':=\{S\subseteq U:|S\cap U_i|\leq r_i\phi, \forall i\in[N]\}$. By the assignment of $r_i$ and $\phi$ in this case, we can get that $r_i\phi=k_i$. It then follows that $\mathcal{P}'=\mathcal{P}$ and that $OPT'$ is also the optimal solution to our problem. Therefore,
\begin{align*}
f(S)&\geq(1-(\frac{\phi}{\phi+1})^\phi)f(OPT)\\
&\geq(1-1/e-1/\phi)f(OPT)=(1-1/e-1/c)f(OPT).
\end{align*}
\end{proof}
In particular, if $c=\mathcal{O}(k_{\min})$ such as in the case where $k_1=k_2=,...=k_N=k$, we have that the approximation ratio is $1-1/e-1/k$. Therefore, the difference between the approximation ratio and the optimal one is decreased to $\mathcal{O}(\frac{1}{k_{min}})$. The result is stated in Corollary \ref{coro:mono}
Next, we prove that we can improve the approximation ratio of \chunkgreedy algorithm by adding more elements to the solution set. First, we notice that there are two drawbacks of the proposed algorithm \chunkgreedy compared with the standard greedy algorithm. First of all, the approximation ratio of $1-1/e-\frac{1}{\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor}$ is only better than the approximation ratio of the standard greedy algorithm, which is $1/2$, when the capacity $k_i$ within partition $U_i$ satisfies that $k_i\geq 64$ for each $i\in[N]$.
Second, the output solution satisfies that $|S\cap U_i|\leq
r_i\phi$ for each $i\in[N]$. Notice that $r_i\phi\leq k_i$. If $r_i\phi< k_i$, we can add more elements to the solution set $S$ until it reaches the full rank of the partition matroid. Since the objective function is monotone, we can see that adding more elements would not incur a decrease in the marginal gain. In the following part, we claim that if the standard greedy procedure (Algorithm \ref{alg:greedy}) is applied to the output of \chunkgreedy with \blockmono as the subroutine, the resulting solution set achieves an approximation ratio of $\max\{1/2,1-1/e-\frac{1}{\phi+1}\}$.
\begin{algorithm}
\caption{\greedy}\label{alg:greedy}
\begin{algorithmic}[1]
\State \textbf{Input: }the output solution set $S$ obtained by running \chunkgreedy with \blockmono as the subroutine and $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$ and $r_j:=\lfloor k_j/\phi\rfloor$
\State \textbf{Output: }$A\in U$
\State $A\gets S$
\While{$\exists x$ such that $A\cup\{x\}\in\mathcal{P}$}
\State $A\gets A\cup \arg\max_{x\in U, A\cup\{x\}\in\mathcal{P}}\Delta f(A,x)$
\EndWhile
\textbf{return }$A$
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\label{thm:standard_greedy}
Suppose we run the standard greedy algorithm in Algorithm \ref{alg:greedy} with input being the output solution set of the \chunkgreedy algorithm, then the output solution set achieves an approximation ratio of $\max\{1/2,1-1/e-\frac{1}{\phi+1}\}$ where $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$.
\end{theorem}
\begin{proof}
First of all, notice that $S\subseteq A$, by the result of Theorem \ref{thm:chunkgreedy}, we can see that $f(S)\geq(1-1/e-\frac{1}{\phi+1})f(OPT)$. Since $f$ is monotone, $f(A)\geq f(S)\geq(1-1/e-\frac{1}{\phi+1})f(OPT)$. Then to prove the result in the Theorem \ref{thm:standard_greedy}, it suffices to prove that $f(A)\geq f(OPT)/2$. Here we use the same notations as in the proof of Theorem \ref{thm:chunkgreedy}, which means that we define the partition matroid of $\{S\subseteq U:|S\cap U_j|\leq
r_j\phi \}$ as $\mathcal{P}'$, and we define the optimal solution of the problem $\max_{S\in\mathcal{P}'}f(S)$ as $OPT'$.
Denote the solution set after completing the $i$-th round of the outer for loop in Line \ref{line:outer_for_loop_starts} in Algorithm \ref{alg:partition_greedy_general} as $S_{i}$. Following the similar idea in the proof of Theorem \ref{thm:chunkgreedy}, we can see that for any $O\in\mathcal{P}'$, it holds that
\begin{align*}
f(S_{i})-f(S_{i-1})
&\geq \frac{\Delta f(S_{i},O)}{\phi}\\
&\geq \frac{\Delta f(S,O)}{\phi},
\end{align*}
where the last inequality follows from submodularity and the fact that $S_i\subseteq S_{\phi}=S$. Summing over all $i$, then we get
\begin{align}
\label{eqn:result_of_block_g}
f(S)-f(\emptyset)&\geq\Delta f(S,O).
\end{align}
Let us define the partition matroid $\mathcal{P}'':=\{S\subseteq U:|S\cap U_i|\leq k_i-r_i\phi, \forall i\in[N]\}$. Let us define the solution set $A$ before the $i$-th round in Algorithm \ref{alg:greedy} as $A_i$, and the element added in the $i$-th round as $a_i$. Since $\mathcal{P}''$ is a matroid, we have that for any $O'\in \mathcal{P}''$, there exists an ordering of $O'=\{o_1',o_2',...,o_t'\}$ such that for each $i\in[t]$, $A_i/S\cup \{o_i'\}\in\mathcal{P''}$. Therefore, for each $i\in[t]$, $A_i\cup\{o_i'\}\in\mathcal{P}$. By the greedy selection rule in Algorithm \ref{alg:greedy}, we have that
\begin{align*}
\Delta f(A_i,a_i)\geq \Delta f(A_i,o_i')\geq \Delta f(A,o_i'),
\end{align*}
where the second inequality follows from the fact that $A$ is the output of Algorithm \ref{alg:greedy} and that $A_i\subseteq A$.
Summing over all $i$, we can get that
\begin{align}
\label{eqn:standard_greedy}
f(A)-f(S)\geq\sum_{i}\Delta f(A,o_i')\geq \Delta f(A,O').
\end{align}
Summing over (\ref{eqn:result_of_block_g}) and (\ref{eqn:standard_greedy}), we can get that
\begin{align*}
f(A)-f(\emptyset)&\geq \Delta f(A,O')+\Delta f(S,O)\\
&\geq\Delta f(A,O')+\Delta f(A,O)\\
&\geq\Delta f(A, O'\cup O).
\end{align*}
Since the above inequality holds for any $O\in\mathcal{P}'$ and $O'\in\mathcal{P}''$. Therefore,
\begin{align*}
f(A)-f(\emptyset)&\geq\max_{O\in\mathcal{P}',O'\in\mathcal{P}''}\Delta f(A, O'\cup O).
\end{align*}
Notice that any set in $\mathcal{P}$ can be decomposed into the union of a set in $\mathcal{P}'$ and a set in $\mathcal {P}''$. It then follows that $\max_{O\in\mathcal{P}',O'\in\mathcal{P}''}\Delta f(A, O'\cup O)\geq\Delta f(A, OPT)$. Therefore, $f(A)\geq f(OPT)/2$.
\end{proof}
\section{Appendix for Section \ref{sec:exp}}
\label{appdx:exp}
In this section, we present the additional experimental setup and results omitted in Section \ref{sec:exp}. In particular, we present additional details about the experimental setup in Section \ref{appdx:setup}, and additional experimental results in Section \ref{appdx:results}.
\subsection{Experimental setup}
\label{appdx:setup}
In this section, we provide additional details about the applications used to evaluate our algorithms, which include set cover, max cover, and graph cut. Below, we define each application and describe the associated setup in detail.
In the application of set cover, the function $f$ is defined to be the number of tags covered by the elements in a subset. The problem is defined as follows.
\begin{definition}{(\textbf{Set Cover})}
Suppose there are a total of $n$ elements denoted as $U$. Let $T$ be a set of tags. Each element in $U$ is tagged with a set of elements from $T$ via a function $t: U\rightarrow 2^T$. The function $f$ is defined as
\begin{align*}
f(S)=|\cup_{s\in S}t(s)|,\qquad\forall S\in U.
\end{align*}
\end{definition}
Next, we introduce the definition of max cover, which is a monotone submodular function defined on graphs.
\begin{definition}{(\textbf{Max Cut})}
Let $G = (V, E)$ be a graph, and $w:E\rightarrow \mathbb{R}_{\geq 0}$ be a function that assigns a weight to every edge in the graph. The function $f:2^V\rightarrow \mathbb{R}_{\geq 0}$ maps a subset of vertices $X\subseteq V$ to the total weight of edges between $X$ and $V\backslash X$. More specifically,
\begin{align*}
f(X)=\sum_{x\in X \text{ or } y\in X}w(x,y).
\end{align*}
\end{definition}
We also evaluate our experiments on the instance of image summarization. For this task, we use a subset of the ImageNet dataset (ImageNet\_50).
\begin{definition}{(\textbf{Image Summarization})}
Let \( N \subseteq \mathbb{R}^d \) denote the ground set, where each item \( x \in N \) (e.g., an image) is represented by a feature vector. The objective is to maximize the \textit{Determinantal Point Process (DPP)} function \citep{iyer2015submodular}, which is a monotone submodular function defined as:
\[
f(S) = \log \det (I + K_S),
\]
where \( I \) is the identity matrix, \( K \in \mathbb{R}^{|N| \times |N|} \) is a positive semidefinite kernel matrix, and \( K_S \) denotes the principal submatrix of \( K \) indexed by the subset \( S \subseteq N \).
\end{definition}
% The additional results are presented in the tables below.
For general SCP, where $f$ can be nonmonotone, the application we consider is where $f$ is a graph cut function, which is a submodular but not necessarily monotone function.
\begin{definition}{(\textbf{Graph Cut})}
Let $G = (V, E)$ be a graph, and $w:E\rightarrow \mathbb{R}_{\geq 0}$ be a function that assigns a weight to every edge in the graph. The function $f:2^V\rightarrow \mathbb{R}_{\geq 0}$ maps a subset of vertices $X\subseteq V$ to the total weight of edges between $X$ and $V\backslash X$. More specifically,
\begin{align*}
f(X)=\sum_{x\in X, y\in V\backslash X}w(x,y).
\end{align*}
\end{definition}
Next, we present more details about the experimental setup in the order of the problems we consider. For the experiments on nonmonotone SCP, the dataset is the email-Euall dataset, where the dataset is partitioned into $5$ different subgroups based on the synthetic labels of the dataset. The group proportions are uniform, i.e., the parameter $p_j$ used in this experiment satisfies $p_1=p_2=\cdots=p_5=1/5$. To speed up the experiments, the conversion algorithm's subroutine is parallelized across 10 threads. Additional details about the values of the parameters in the experiments are presented as follows. The parameter $\alpha=0.2$, $\epsilon = 0.05$ and $\delta = 0.1$.
For the experiments on monotone SCKP, we run the experiments on two instances, which include max cover and set cover. For the max cover instance, we use a subset of the Twitch Gamers dataset~\citep{twitch}, selecting 2,000 users speaking six major languages which include English, German, French, Spanish, Russian, and Chinese. For the set cover instance, we use two datasets here. The first one is the core dataset, which is the Corel5k set of
images in \cite{duygulu2002object} ($n=4500$). We assign a label to each element in the dataset uniformly selected from $\{0,1,2,3,4\}$. Another dataset we use here is the synthetic dataset. The synthetic dataset is generated with $5$ partitions with $40*i+200$ number of elements in partition $i$ for each $i\in[4]$. The synthetic dataset has a similar structure as the tightness example in Section \ref{appdx:hardness} in the appendix. In the first partition, each element is mapped to a disjoint set of tags. For the elements partition $i$ where $i>1$, the mapped sets of tags of $100$ elements are the same, with the other elements mapped to disjoint sets of $25$ tags. The cost of each element in the synthetic dataset and in the twitch dataset is generated randomly in the range of $[0.001,10]$. The other parameters in the experiments comparing different values of $\tau$ include: $\alpha = 0.2$, $\epsilon=0.05$. The parameters for the experiments comparing different values of $\alpha$ include: threshold value $\tau=700$, $\epsilon=0.05$. The parameter $p_j$ used in this experiment satisfies $p_1=p_2=\cdots=p_5=1/5$. Next, we illustrate the two algorithms used in the experiments. The GREEDY algorithm uses the converting theorem in Algorithm \ref{alg:conver} with the subroutine being a greedy algorithm. In particular, the subroutine greedy algorithm adds the element $s=\arg\max_{x:S\cup x\in \mathcal{P}}\frac{\Delta f(S,x)}{c(x)}$ to the solution set $S$, where $\mathcal{P}=\{S\subseteq U:c(S\cap U_j)\leq p_jv,\forall j\in[N]\}$. Here the subroutine algorithm of GREEDY is not guaranteed with any approximation ratio. In this sense, this algorithm can be regarded as a heuristic algorithm. The GREEDY-Knapsack algorithm runs by iteratively adding the element with the highest density of marginal gain, i.e., $s=\arg\max_{x\in U}\frac{\Delta f(S,x)}{c(x)}$ until $f(S)\geq(1-\epsilon)\tau$.
For the SCF experiments, we consider the same synthetic dataset and the corel dataset used in the experiment for SCKP. Apart from these datasets, we also consider the image summarization task, where the goal is to select a diverse and representative image subset across all classes. The dataset used here is ImageNet \citep{deng2009imagenet}, consisting of 50 classes and 26,112 images (ImageNet\_50). Each image is represented by a feature vector extracted using ResNet-18. Additionally, in our experiment, we set $K$ as a Gaussian kernel matrix such that $K_{ij} = e^{-||x_i - x_j||^2 / \sigma^2}$.
To ensure a fair comparison among the used algorithms, we keep the approximation ratio on the function value $f$ the same by setting $\varepsilon = 0.05$ for THRES-Fair and $\varepsilon = 0.1$ for GREEDY-Fair and BLOCK-G-Fair while keeping the other parameters the same.
\subsection{Additional experimental results}
\label{appdx:results}
The additional experimental results comparing different algorithms for different problems are presented in Figure \ref{fig:exp_result_appdx_knapsack}, \ref{fig:exp_result_appdx_knapsack_alpha} and \ref{fig:exp_result_appdx}. The additional experimental results for SCKP algorithms on the corel dataset and the synthetic dataset are presented in Figure \ref{fig:exp_result_appdx_knapsack} and \ref{fig:exp_result_appdx_knapsack_alpha}. The results demonstrate that our algorithm, BLOCK-G, achieves a slightly lower cost compared to GREEDY and significantly outperforms GREEDY-Knapsack in this regard. Additionally, BLOCK-G requires substantially fewer function queries and has a much faster runtime than GREEDY, highlighting its practical efficiency. The function values $f$ for different algorithms are similar, which is because all algorithms (including baselines) in the experiments are designed to terminate once the function value reaches $(1-\epsilon)\tau$. For the experiments on different values of $\alpha$, we can see that increasing
$\alpha$ significantly reduces the number of evaluations, leading to dramatic improvements in runtime. This also corresponds to the results of query complexity of our converting algorithm (Algorithm \ref{alg:convk}) as proved in Theorem \ref{thm:convk}.
Further results on the query complexity and runtime for the non-monotone SCP problem are provided in Figures~\ref{fig:euall_q} and~\ref{fig:euall_time}. From the results, we can see that the BLOCK-G algorithm runs faster than the STREAM algorithm and the GUIDED-RG algorithm, which demonstrates the efficiency of our algorithm.
The additional results of comparing different algorithms in terms of the query complexity for the experiments on the SCF problem are presented in Figure \ref{fig:cover5000_fair_q} and \ref{fig:synthetic_fair_q}. From the results, we can see that the query complexity of the BLOCK-G-Fair algorithm is better than that of the GREEDY-Fair algorithm, and is worse than THRES-Fair. This is because these three algorithms differ in the subroutine algorithm used in the converting algorithm in Algorithm 1 in \cite{chen2024fair} developed to convert an algorithm for SMF to SCF. Specifically, THRES-Fair used the threshold greedy algorithm, which runs in time complexity of $\mathcal{O}(\frac{n}{\epsilon}\log\frac{n}{\epsilon})$ while the subroutine algorithms for BLOCK-G-Fair and GREEDY-Fair both run in time $\mathcal{O}(nk_g\beta)$ where $k_g$ is the guess of $|OPT|$ and the parameter $\beta$ refers to the approximation ratio. Therefore, the query complexity of BLOCK-G-Fair is lower than GREEDY-Fair because the parameter $\beta$ for BLOCK-G-Fair is $\frac{\ln\frac{1}{\epsilon}}{\ln 2}$, which is smaller than the GREEDY-Fair, which is $\mathcal{O}(\frac{1}{\epsilon})$.
The results of the function values for different assignments of $\tau$ on the experiments of SCF are presented in Figure \ref{fig:synthetic_fair_f} and \ref{fig:cover5000_fair_f}.
From the plots, we can see that the function values of the returned solutions of different algorithms are almost the same, and are linear in the threshold value $\tau$. This aligns with our theoretical guarantee of different algorithms, which requires that $f(S)\geq0.9\tau$ for all of the algorithms. Finally, we also provide the results of the execution time of running different algorithms in Figure \ref{fig:cover5000_fair_time}, and \ref{fig:synthetic_fair_t}.
The additional experimental results on the ImageNet\_50 dataset are presented in Figure \ref{fig:exp_result_imagebet}. From these results, we observe that block-greedy consistently achieves significantly better fairness performance and, in many cases, returns solutions with lower or comparable cost to baselines. This demonstrates its practical effectiveness, especially in fairness-sensitive applications.
Finally, we also plot the distribution of different labels in the solutions produced by these algorithms on the corel dataset with $\tau = 300$, as is presented in Figure \ref{fig:threshold_fair_diff}, \ref{fig:greedy_fair_differ}, and \ref{fig:blockG_fair_differ}. From the plots, we can see that over $30 \%$ of the elements in the solution returned by GREEDY-Fair and THRES-Fair have the label $1$, which indicates a lack of fairness in the output distribution. While the solutions produced by our algorithm BLOCK-G-Fair exhibit significantly fairer distributions across different labels, demonstrating the effectiveness of our proposed algorithms.
\begin{figure*}[t!]
\centering
\hspace{-0.5em}
\subfigure[corel time (SCKP)]
{\label{fig:cover5000_time_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/corel_threshold_n4500-time.pdf}}
\hspace{-0.5em}
\subfigure[corel $f$ (SCKP)]
{\label{fig:cover5000_f_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/corel_threshold_n4500-f.pdf}}
\hspace{-0.5em}
\subfigure[corel queries (SCKP)]
{\label{fig:cover5000_q_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/corel_threshold_n4500-q.pdf}}
\hspace{-0.5em}
\subfigure[corel budget (SCKP)]
{\label{fig:cover5000_c_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/corel_threshold_n4500-budget.pdf}}
\hspace{-0.5em}
\subfigure[synthetic time (SCKP)]
{\label{fig:synthetic_time_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/synthetic_set_cover_knapsack-time.pdf}}
\hspace{-0.5em}
\subfigure[synthetic $f$ (SCKP)]
{\label{fig:synthetic_f_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/synthetic_set_cover_knapsack-f.pdf}}
\hspace{-0.5em}
\subfigure[synthetic queries (SCKP)]
{\label{fig:synthetic_q_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/synthetic_set_cover_knapsack-q.pdf}}
\hspace{-0.5em}
\subfigure[synthetic budget (SCKP)]
{\label{fig:synthetic_c_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/synthetic_set_cover_knapsack-budget.pdf}}
\caption{The experimental results of running the algorithms on the Corel5k dataset and the synthetic dataset. Samples: the number of queries. Budget: $\max_{i\in[N]}\frac{c(S\cap U_i)}{p_i}$.}
\label{fig:exp_result_appdx_knapsack}
\end{figure*}
\begin{figure*}[t!]
\centering
\hspace{-0.5em}
\subfigure[twitch time (SCKP)]
{\label{fig:twitch_alpha_time_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/large_twitch_edges.csv_alpha_n2000-time.pdf}}
\hspace{-0.5em}
\subfigure[twitch $f$ (SCKP)]
{\label{fig:twitch_alpha_f_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/large_twitch_edges.csv_alpha_n2000-f.pdf}}
\hspace{-0.5em}
\subfigure[twitch queries (SCKP)]
{\label{fig:twitch_alpha_q_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/large_twitch_edges.csv_alpha_n2000-q.pdf}}
\hspace{-0.5em}
\subfigure[twitch budget (SCKP)]
{\label{fig:twitch_alpha_c_knapsack}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/large_twitch_edges.csv_alpha_n2000-budget.pdf}}
\hspace{-0.5em}
\caption{The experimental results for the SCKP problem on the Twitch dataset across different $\alpha$ values. Samples: the number of queries. Budget: $\max_{i\in[N]}\frac{c(S\cap U_i)}{p_i}$.}
\label{fig:exp_result_appdx_knapsack_alpha}
\end{figure*}
\begin{figure*}[t!]
\centering
\hspace{-0.5em}
\subfigure[twitch $q$ (SCKP)]
{\label{fig:twich_q}\includegraphics[width=0.24\textwidth]{graphs/graphs_knapsack/large_twitch_threshold_n2000-q.pdf}}
\hspace{-0.5em}
\subfigure[corel $f$ (SCF)]
{\label{fig:cover5000_fair_f}\includegraphics[width=0.24\textwidth]{graphs/threshold_n4500-f.pdf}}
\hspace{-0.5em}
\subfigure[corel queries (SCF)]
{\label{fig:cover5000_fair_q}\includegraphics[width=0.24\textwidth]{graphs/threshold_n4500-q.pdf}}
\hspace{-0.5em}
\subfigure[corel time (SCF)]
{\label{fig:cover5000_fair_time}\includegraphics[width=0.24\textwidth]{graphs/threshold_n4500-time.pdf}}
\hspace{-0.5em}
\subfigure[twitch time (SCKP)]
{\label{fig:twitch_time}\includegraphics[width=0.2\textwidth]{graphs/graphs_knapsack/large_twitch_threshold_n2000-time.pdf}}
\hspace{-0.5em}
\subfigure[synthetic $f$ (SCF)]
{\label{fig:synthetic_fair_f}\includegraphics[width=0.24\textwidth]{graphs/synthetic_set_cover_threshold_n1400-f.pdf}}
\hspace{-0.5em}
\subfigure[synthetic queries (SCF)]
{\label{fig:synthetic_fair_q}\includegraphics[width=0.24\textwidth]{graphs/synthetic_set_cover_threshold_n1400-q.pdf}}
\hspace{-0.5em}
\subfigure[synthetic time (SCF)]
{\label{fig:synthetic_fair_t}\includegraphics[width=0.24\textwidth]{graphs/synthetic_set_cover_threshold_n1400-time.pdf}}
\hspace{-0.5em}
\subfigure[euall queries (nonmonotone SCP)]
{\label{fig:euall_q}\includegraphics[width=0.24\textwidth]{graphs/euallblock_nonmono_n265214-q.pdf}}
\hspace{-0.5em}
\subfigure[euall time (nonmonotone SCP)]
{\label{fig:euall_time}\includegraphics[width=0.24\textwidth]{graphs/euallblock_nonmono_n265214-time.pdf}}
\caption{The experimental results of running the algorithms on the Corel5k dataset and the synthetic dataset. Samples: the number of queries. Cost: the size of the returned solution. Budget: $\max_{i\in[N]}\frac{c(S\cap U_i)}{p_i}$. Fairness difference: $(\max_c |S \cap U_c| - \min_c |S \cap U_c|) / |S|$.}
\label{fig:exp_result_appdx}
\end{figure*}
\begin{figure*}[t!]
\centering
\hspace{-0.5em}
\subfigure[ImageNet\_50 $f$ (SCF)]
{\label{fig:imagenet_fair_f}\includegraphics[width=0.24\textwidth]{graphs/image-net_50_classes-f.pdf}}
\hspace{-0.5em}
\subfigure[ImageNet\_50 queries (SCF)]
{\label{fig:imagenet_fair_q}\includegraphics[width=0.24\textwidth]{graphs/image-net_50_classes-q.pdf}}
\hspace{-0.5em}
\subfigure[ImageNet\_50 fairness difference (SCF)]
{\label{fig:imagenet_fair_difference}\includegraphics[width=0.24\textwidth]{graphs/image-net_50_classes-fair-difference.pdf}}
\hspace{-0.5em}
\subfigure[ImageNet\_50 cost (SCF)]
{\label{fig:imagenet_fair_cost}\includegraphics[width=0.24\textwidth]{graphs/image-net_50_classes-cost.pdf}}
\caption{The experimental results of running the algorithms on the ImageNet\_50 dataset on the SCF problem. Samples: the number of queries. Cost: the size of the returned solution. Fairness difference: $(\max_c |S \cap U_c| - \min_c |S \cap U_c|) / |S|$.}
\label{fig:exp_result_imagebet}
\end{figure*}
\begin{figure*}[t!]
\centering
\hspace{-0.5em}
\subfigure[THRES-Fair]
{\label{fig:threshold_fair_diff}\includegraphics[width=0.3\textwidth]{graphs/threshold_threshold_300.0_fairness.pdf}}
\hspace{-0.5em}
\subfigure[GREEDY-Fair]
{\label{fig:greedy_fair_differ}\includegraphics[width=0.3\textwidth]{graphs/greedy_threshold_300.0_fairness.pdf}}
\hspace{-0.5em}
\subfigure[BLOCK-G-Fair]
{\label{fig:blockG_fair_differ}\includegraphics[width=0.3\textwidth]{graphs/block-greedy_threshold_300.0_fairness.pdf}}
\caption{Radar plots of the label distributions for the experiments on the instance of SCF on the corel dataset with $\tau=300$.}
\label{fig:exp_result_appdx2}
\end{figure*}
\section{Broader Impact}\label{apdx:braoder-impact}
This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.
% \textcolor{red}{If we end up having time, we could consider putting in an image for chunk greedy.}
The \chunkgreedy algorithm serves as the core framework for most of our proposed algorithms, except for the \fairgreedy algorithm used in the Fair Submodular Cover problem. \chunkgreedy repeatedly runs a greedy subroutine. In each of the subroutines, a ``block'' of elements is added into the final solution from each of the partitions of the universe $U$. The value of the parameter $\phi$ and the subroutine \algsub are problem-specific and vary depending on the submodular optimization problem being solved. The pseudocode for \chunkgreedy is in Algorithm \ref{alg:partition_greedy_general}.
In the following part, we introduce the subroutine algorithm for different problems and present the analysis for these proposed algorithms.
%\chunkgreedy algorithm operates by running greedy algorithms in multiple blocks. Specifically, for each block, only elements in a particular partition $U_j$ are considered to be added to the solution set. Here $\phi$ is a parameter depending on the problem type, and \algsub denotes a subroutine algorithm, which may vary depending on the submodular optimization problem used. We will illustrate the advantage of running a greedy algorithm in blocks in Section \ref{sec:mono} by discussing the hardness result for the problem of monotone \prob.
\begin{algorithm}[t] \caption{\chunkgreedy}
\begin{algorithmic}[1]
\State \textbf{Input: } Partitions of the ground set $U_1, U_2,...,U_N$, problem definition and parameters
\State \textbf{Output: }$S\subseteq U$
\State $S\gets\emptyset$
\For{$i=1$ \textbf{to} $\phi$ \label{line:outer_for_loop_starts}}
\For{$j=1$ \textbf{to} $N$}
\State \algsub($S$, $i$, $j$)
\EndFor
\EndFor\label{line:outer_for_loop_ends2}
\State\textbf{return }$S$
\end{algorithmic}
\label{alg:partition_greedy_general}
\end{algorithm}
\vspace{-1em}
%\textcolor{red}{Not sure if this is somewhere, but might want to also put bicriteria statement in.}
%\textcolor{red}{Another algorithm where it might be nice to have a figure.}
We first consider the classic problem of Monotone Submodular Maximization with a Partition Matroid Constraint (SMP).
Given positive integers $k_1,\cdots k_N$ such that $k_j\leq|U_j|$ for any $j\in[N]$, the partition matroid constraint is defined as $\mathcal{P}=\{S\subseteq U:|S\cap U_j|\leq k_j, \forall j\in[N]\}$. The monotone SMP is defined to find the set $\arg\max_{S\in\mathcal{P}}f(S)$ for a monotone, submodular objective function $f$. Before presenting our algorithm, we illustrate the intuitions and benefits of our proposed algorithm in contrast to the standard greedy algorithm through a tight hardness result.
\subsubsection{Tight Examples}
\label{sec:tight}
The standard greedy algorithm iteratively selects the element with the highest marginal gain while maintaining feasibility. It is well-known that this algorithm achieves a $1/2$-approximation ratio for monotone submodular maximization with general matroid constraint.
Despite partition matroids being a simpler special case, in the theorem below, we prove this ratio is tight by constructing a class of instances where the standard greedy algorithm cannot achieve an approximation ratio better than $1/2$.
\begin{theorem}
\label{thm:hardness}
For any given positive integers $k_1,...,k_N$, there exists an instance of monotone \prob with size constraints $k_1,...,k_N$, i.e., $$\max_{S\in\mathcal{P}}f(S)$$ where $\mathcal{P}:=\{S\subseteq U:|S\cap U_i|\leq k_i\}$, such that the best approximation ratio achievable by the standard greedy algorithm is $1/2$.
\end{theorem}
We defer the detailed proof of Theorem \ref{thm:hardness} to the supplementary material in Section \ref{appdx:hardness}. We give a brief illustration of the proof by constructing a toy example. Suppose $U=[8]$ which is split into two groups $U_1=\{1,2,3,4\}$, $U_2=\{5,6,7,8\}$. Let $t:U\rightarrow M$ be a function that assigns tags to each element in the universe: $t(1) = t(5) = t(7) = t(8) = "a"$, $t(2) = t(6) = "b"$, $t(3) = "c"$, and $t(4) = "d"$. Define a set cover function $f$ that maps a subset to the number of unique tags covered, $f(S) = |\cup_{s \in S} t(s)|$. Here, the partition matroid is defined by $k_1 = k_2 = 2$. In this case, the standard greedy algorithm might first select elements $1$ and $2$. Subsequently, all remaining elements either become infeasible or have zero marginal gain, yielding a solution set $S$ with $f(S) = 2$, whereas the optimal solution set $OPT = {3,4,5,6}$ achieves $f(OPT) = 4$. Therefore, $f(S) = f(OPT)/2$. Thus we can conclude the proof.
This example highlights a key limitation of the standard greedy algorithm: it greedily adds the element with the highest marginal gain by searching over all feasible elements in each step, which can lead certain partitions to quickly reach their cardinality limits, making remaining elements in those subgroups infeasible for later selections. This strategy prevents the standard greedy algorithm from achieving a better approximation ratio.
In fact, in most of the continuous methods developed in existing works \citep{badanidiyuru2014fast,calinescu2011maximizing}, the key idea for achieving the optimal approximation ratio of $1-1/e$ is by incrementally increasing some coordinates by small fractions in each step. This technique ensures that all of the elements in the ground set $U$ remain feasible throughout most of the algorithm's execution. Inspired by this insight, the \chunkgreedy algorithm carefully balances the number of elements being added to the solution set across different partitions during each round. In particular, the number of elements selected by \chunkgreedy at each round within each partition $i$ is proportional to the budget capacity $k_i$ of the partition to ensure the feasibility of the elements for the majority of the algorithm's runtime, thereby effectively addressing this limitation. We provide a detailed description of our proposed algorithm for monotone SMP in the next section.
\subsubsection{Subroutine Algorithm for Monotone SMP}
\label{sec:mono_alg_descript}
We propose the subroutine algorithm \blockmono (Algorithm \ref{alg:blockmono}), to be used in \chunkgreedy along with the parameter $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$, and show that it can be used to achieve an approximate solution that is at least as good as the standard greedy and often strictly better. Further, \chunkgreedy makes fewer queries to $f$, depending on the structure of the partition matroid constraint. Here the parameter $r_j$ is defined as $r_j:=\lfloor k_j/\phi\rfloor$ for each $j\in[N]$.
\begin{algorithm} \caption{\blockmono($S$, $i$, $j$)}\label{alg:blockmono}
\begin{algorithmic}[1]
\State \textbf{Input: }$S,i,j$
% \State \textbf{Output: }$S\in U$
\For{$l=1$ \textbf{to} $r_j $}
\State $S\gets S\cup \arg\max_{x\in U_j}\Delta f(S,x)$\label{line:greedy_mono}
\EndFor
% \textbf{return }$S$
\end{algorithmic}
\end{algorithm}
\begin{restatable}{theorem}{thmchunkgreedy}
\label{thm:chunkgreedy}
Suppose that \chunkgreedy is run for an instance of monotone \prob with the subroutine algorithm \blockmono as described in Algorithm \ref{alg:blockmono},
then \chunkgreedy
outputs a solution set $S$ that satisfies an approximation ratio of $1-1/e-\frac{1}{\phi+1}$ where $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$.
\end{restatable}
Intuitively, \chunkgreedy achieves an approximation close to $1-1/e$ when the parameters $k_i$ are large. The reason that the term involving $\phi$ arises in Theorem \ref{thm:chunkgreedy} is because there are a total of $\phi$ rounds in the outer loop of \chunkgreedy. In particular, if $\phi\geq7$, then the approximation guarantee described in Theorem \ref{thm:chunkgreedy} is strictly better than 1/2. To further ensure the bound is better than 1/2, we can greedily add new elements with maximum marginal gain to the returned solution by \chunkgreedy algorithm until the cardinality of the solution set reaches the rank of the partition matroid, in which case the approximation ratio of \chunkgreedy is $\max\{1/2,1-1/e-\frac{1}{\phi+1}\}$ (see Appendix \ref{appdx:mono_result_discuss} for proof).
Notably, the difference between the approximation ratio for \chunkgreedy and the optimal result of $1-1/e$ is bounded by $\mathcal{O}(\frac{1}{\sqrt{k_{\min}}})$ where $k_{\min}=\min_{i\in[N]}k_i$. In particular, in some cases where $k_1=k_2=\cdots=k_N$, the bound can be improved further to $\mathcal{O}(\frac{1}{k_1})$, as shown in the following corollary:
\begin{corollary}
\label{coro:mono}
Suppose \chunkgreedy with the \blockmono subroutine is run for instance of monotone \prob with $k_1=k_2=...=k_N$. If we set $\phi=k_1$ and $r_j=1$ for each $j$, then \chunkgreedy outputs a solution set $S$ with an approximation ratio of $1-1/e-1/k_1$.
\end{corollary}
The corollary can be proved by applying Theorem \ref{thm:greedy-gcd}, which is presented and analyzed in Appendix \ref{appdx:mono_result_discuss}.
%Besides, notice that the output solution set when the algorithm terminates satisfies that $|S\cap U_j|\leq r_j\phi\leq k_j$. Therefore, we can greedily add new elements with maximum marginal gain until the solution is full. By this method, the output solution set achieves an approximation ratio of $\max\{1/2,1-1/e-\frac{1}{\phi+1}\}$. For detailed proof and discussion, please refer to the appendix.
An additional important benefit to \chunkgreedy compared to the standard greedy algorithm is that its query complexity is potentially much better. This improvement arises because \chunkgreedy selects elements with maximum marginal gain within one partition rather than over the entire universe $U$. In particular, the query complexity of \chunkgreedy is upper bounded by $\sum_{i\in [N]}|U_i|k_i\leq n(\sum_{i\in [N]}k_i)$.
Next, we present the proof of the theorem. First of all, we prove the result in Lemma \ref{lem:mono}.
\begin{lemma}
\label{lem:mono}
Let us define the partition matroid of $\{S\subseteq U:|S\cap U_j|\leq
r_j\phi \}$ as $\mathcal{P}'$, and we define the optimal solution of the problem $\max_{S\in\mathcal{P}'}f(S)$ as $OPT'$. Let us denote the input and output of Algorithm \ref{alg:blockmono} as $S$ and $S'$ respectively, then it follows that
\begin{align*}
f(S')-f(S)\geq\frac{\Delta f(S',OPT_j)}{\phi},
\end{align*}
where $OPT_j=OPT'\cap U_j$.
\end{lemma}
\begin{proof}
For notation simplicity, we define the solution set after the $l$-th step in the for loop of Algorithm \ref{alg:blockmono} as $S^{(l)}$. By the greedy selection step in Line \ref{line:greedy_mono}, we have that for any $o\in OPT_j:=OPT'\cap U_j$,
\begin{align*}
f(S^{(l)})-f(S^{(l-1)})\geq\Delta f(S^{(l-1)}, o).
\end{align*}
Therefore,
\begin{align*}
f(S^{(l)})-f(S^{(l-1)})&\geq\frac{\sum_{o\in OPT_j}\Delta f(S^{(l-1)}, o)}{|OPT_j|}\\
&\geq\frac{\sum_{o\in OPT_j}\Delta f(S^{(l-1)}, o)}{r_j\phi}\\
&\geq\frac{\sum_{o\in OPT_j}\Delta f(S^{(r_j)}, o)}{r_j\phi}\\
&\geq\frac{\Delta f(S^{(r_j)}, OPT_j)}{r_j\phi},
\end{align*}
where the second inequality follows from the fact that $OPT'\in\mathcal{P}'$, and therefore $|OPT_j|\leq r_j\phi$. Summing over all $l\in[r_j]$, it follows that
\begin{align*}
f(S^{(r_j)})-f(S^{(0)})\geq\frac{\Delta f(S^{(r_j)}, OPT_j)}{\phi}.
\end{align*}
Notice that $S^{(0)}$ is the input of the algorithm and $S^{(l)}$ is the output of the algorithm, so we can prove the result.
\end{proof}
With the result in Lemma \ref{lem:mono}, we can prove the result in Theorem \ref{thm:chunkgreedy}.
\begin{proof}
Let $S_{i,j}$ represent the solution set after executing the subroutine algorithm \blockmono on the
$j$-th subgroup during the $i$-th iteration of the outer for loop in Line~\ref{line:outer_for_loop_starts} in Algorithm \ref{alg:partition_greedy_general}, and we define $S_{i}$ as the solution set after completing the $i$-th round of the outer for loop in Algorithm \ref{alg:partition_greedy_general}, i.e., $S_i=S_{i,N}$. Then by the result in Lemma \ref{lem:mono}, we have that we have that
\begin{align*}
f(S_{i,j})-f(S_{i,j-1})\geq\frac{\Delta f(S_{i,j},OPT_j)}{\phi}
\end{align*}
Since $S_{i,j}\subseteq S_{i,N}$ for any $j\in[N]$, by submodularity, we have that $\Delta f(S_{i},OPT_j)=\Delta f(S_{i,N},OPT_j)\leq \Delta f(S_{i,j},OPT_j)$. Then
\begin{align*}
f(S_{i,j})-f(S_{i,j-1})\geq\frac{\Delta f(S_{i},OPT_j)}{\phi}
\end{align*}
Summing over all $j$, it then follows that
\begin{align*}
f(S_{i,N})-f(S_{i,0})&\geq\frac{\sum_{j\in[N]}\Delta f(S_{i}, OPT_j)}{\phi}\\
&\geq \frac{\Delta f(S_{i},OPT')}{\phi},
\end{align*}
where the last inequality follows from submodularity and the fact that $OPT'=\cup_{j\in[N]}OPT_j$.
Notice that here $S_{i,N}$ is equivalent to $S_{i}$, and that $S_{i,0}$ is equivalent to $S_{i-1}$. Then we get
\begin{align*}
f(S_{i})-f(S_{i-1})&\geq\frac{f(OPT'\cup S_i)-f(S_i)}{\phi}\\
&\geq\frac{f(OPT')-f(S_i)}{\phi}.
\end{align*}
By rearranging the inequality and by induction, we have that
\begin{align*}
f(S_\phi)-f(\emptyset)\geq (1-(\frac{\phi}{\phi+1})^\phi)f(OPT').
\end{align*}
% Let us sequentially denote the $OPT^{(i)}=\arg\max_{S\in \mathcal{P}_0,S\subseteq OPT}\Delta f(\cup_{j\in [i-1]}OPT^{(j)},S)$. Then it follows that $OPT_2\in\mathcal{M}'$. By the definition of $OPT'$, we have that $f(OPT')\geq f(\cup_{i\in[c]}OPT^{(i)})=\sum_{i\in[c]}\Delta f(\cup_{j\in [i-1]}OPT^{(j)},OPT^{(i)})$. Let us denote $OPT_2=\cup_{i\in[c]}OPT^{(i)}$ and $OPT_3=OPT/OPT_2$. Since $c=\sqrt{\min_{i\in[N]}k_i}-1$, it then follows that $k_i- c\lfloor k_i/c\rfloor\leq \lfloor k_i/c\rfloor$ for any $i\in[N]$. Therefore, $OPT_3\in\mathcal{P}_0$. From the definition of $OPT^{(i)}$, we have that $\Delta f(\cup_{j\in[i-1]}OPT^{(j)},OPT_3)\leq \Delta f(\cup_{j\in [i-1]}OPT^{(j)},OPT^{(i)})$. By submodularity, we have that $\Delta f(\cup_{j\in[c]}OPT^{(j)},OPT_3)\leq \Delta f(\cup_{j\in [i-1]}OPT^{(j)},OPT_3)$ for any $i\in[c]$. It then follows that $\Delta f(OPT_2,OPT_3)\leq \Delta f(\cup_{j\in [i-1]}OPT^{(j)},OPT^{(i)})$ for any $i\in[c]$. Thus we can conclude that
% \begin{align*}
% \Delta f(OPT_2,OPT_3)\leq \frac{\sum_{i\in[c]}\Delta f(\cup_{j\in [i-1]}OPT^{(j)},OPT^{(i)})}{\phi}=\frac{f(OPT_2)}{\phi}
% \end{align*}
By the definition of $\phi$ that $\phi=\lfloor\sqrt{\min_{i\in[N]}k_i}\rfloor-1$, we have that $k_i- \phi\lfloor k_i/\phi\rfloor\leq \lfloor k_i/\phi\rfloor$ for any $i\in[N]$. By Lemma \ref{lem:partial_OPT}, it follows that
\begin{align*}
\max_{S\in\mathcal{P}'}f(S)\geq\frac{\phi}{\phi+1}\max_{S\in\mathcal{P}}f(S).
\end{align*}
Therefore,
\begin{align*}
f(S_\phi)&\geq (1-(\frac{\phi}{\phi+1})^\phi)f(OPT')\\
&\geq (1-(\frac{\phi}{\phi+1})^\phi)(\frac{\phi}{\phi+1})f(OPT)\\
&\geq(1-e^{-1}-\frac{1}{\phi+1})f(OPT).
\end{align*}
% Since the $S_\phi$ is a subset of the output solution set $S$, and that $f$ is monotone, we have that the output solution set satisfies $f(S)\geq (1-e^{-1}-\frac{1}{\phi+1})f(OPT)$.
\end{proof}