Sync or Sink: Bounds on Algorithmic Collective Action with Noise and Multiple Groups
Extracted source text · arXiv 2510.18933
\documentclass{article}
\usepackage{graphicx} %
\usepackage{amsfonts}
\usepackage{amsthm}
\usepackage[preprint]{neurips_2025}
\usepackage{amsmath}
\usepackage{cleveref}
\usepackage{url}
\usepackage{enumitem}
\usepackage{booktabs}
\usepackage{subcaption}
\title{Sync or Sink: Bounds on Algorithmic Collective Action with Noise and Multiple Groups}
\author{Aditya Karan}
\author{%
Aditya Karan \\
Siebel School of Computing and Data Science\\
University of Illinois at Urbana-Champaign\\
Urbana, IL 61820 \\
\texttt{karan2@illinois.edu} \\
\And
Prabhat Kalle \\
Siebel School of Computing and Data Science\\
University of Illinois at Urbana-Champaign\\
Urbana, IL 61820 \\
\texttt{pskalle2@illinois.edu} \\
\AND
Nicholas Vincent \\
Simon Fraser University \\
Burnaby, British Columbia \\
\texttt{nvincent@sfu.ca} \\
\And
Hari Sundaram \\
Siebel School of Computing and Data Science \\
Urbana, IL 61820 \\
\texttt{hs1@illinois.edu} \\
}
\newtheorem{theorem}{Theorem}
\newtheorem{corollary}{Corollary}[theorem]
\newtheorem{claim}{Claim}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{arg\,min}
\usepackage{xcolor}
\newcommand\todos[1]{\textcolor{red}{Aditya; #1}}
\DeclareMathOperator*{\E}{\mathbb{E}}
\date{March 2025}
\begin{document}
\maketitle
\begin{abstract}
\input{sections/_0_abstract}
\end{abstract}
\input{sections/_1_introduction}
\input{sections/_2_related_works}
\input{sections/_3_problem_definition}
\input{sections/_4_theorems}
\input{sections/_5_experimental_setup}
\input{sections/_6_results}
\input{sections/_7_discussion}
\input{sections/_8_conclusion}
\bibliographystyle{alpha}
\bibliography{ref}
\appendix
\input{sections/appendix}
\end{document}
\documentclass{article}
\usepackage{neurips_2025}
\usepackage[utf8]{inputenc} %
\usepackage[T1]{fontenc} %
\usepackage{hyperref} %
\usepackage{url} %
\usepackage{booktabs} %
\usepackage{amsfonts} %
\usepackage{nicefrac} %
\usepackage{microtype} %
\usepackage{xcolor} %
\title{Formatting Instructions For NeurIPS 2025}
\author{%
David S.~Hippocampus\thanks{Use footnote for providing further information
about author (webpage, alternative address)---\emph{not} for acknowledging
funding agencies.} \\
Department of Computer Science\\
Cranberry-Lemon University\\
Pittsburgh, PA 15213 \\
\texttt{hippo@cs.cranberry-lemon.edu} \\
}
\begin{document}
\maketitle
\begin{abstract}
The abstract paragraph should be indented \nicefrac{1}{2}~inch (3~picas) on
both the left- and right-hand margins. Use 10~point type, with a vertical
spacing (leading) of 11~points. The word \textbf{Abstract} must be centered,
bold, and in point size 12. Two line spaces precede the abstract. The abstract
must be limited to one paragraph.
\end{abstract}
\section{Submission of papers to NeurIPS 2025}
Please read the instructions below carefully and follow them faithfully.
\subsection{Style}
Papers to be submitted to NeurIPS 2025 must be prepared according to the
instructions presented here. Papers may only be up to {\bf nine} pages long,
including figures.
Additional pages \emph{containing references, checklist, and the optional technical appendices} do not count as content pages.
Papers that exceed the page limit will not be
reviewed, or in any other way considered for presentation at the conference.
The margins in 2025 are the same as those in previous years.
Authors are required to use the NeurIPS \LaTeX{} style files obtainable at the
NeurIPS website as indicated below. Please make sure you use the current files
and not previous versions. Tweaking the style files may be grounds for
rejection.
\subsection{Retrieval of style files}
The style files for NeurIPS and other conference information are available on
the website at
\begin{center}
\url{https://neurips.cc}
\end{center}
The file \verb+neurips_2025.pdf+ contains these instructions and illustrates the
various formatting requirements your NeurIPS paper must satisfy.
The only supported style file for NeurIPS 2025 is \verb+neurips_2025.sty+,
rewritten for \LaTeXe{}. \textbf{Previous style files for \LaTeX{} 2.09,
Microsoft Word, and RTF are no longer supported!}
The \LaTeX{} style file contains three optional arguments: \verb+final+, which
creates a camera-ready copy, \verb+preprint+, which creates a preprint for
submission to, e.g., arXiv, and \verb+nonatbib+, which will not load the
\verb+natbib+ package for you in case of package clash.
\paragraph{Preprint option}
If you wish to post a preprint of your work online, e.g., on arXiv, using the
NeurIPS style, please use the \verb+preprint+ option. This will create a
nonanonymized version of your work with the text ``Preprint. Work in progress.''
in the footer. This version may be distributed as you see fit, as long as you do not say which conference it was submitted to. Please \textbf{do
not} use the \verb+final+ option, which should \textbf{only} be used for
papers accepted to NeurIPS.
At submission time, please omit the \verb+final+ and \verb+preprint+
options. This will anonymize your submission and add line numbers to aid
review. Please do \emph{not} refer to these line numbers in your paper as they
will be removed during generation of camera-ready copies.
The file \verb+neurips_2025.tex+ may be used as a ``shell'' for writing your
paper. All you have to do is replace the author, title, abstract, and text of
the paper with your own.
The formatting instructions contained in these style files are summarized in
Sections \ref{gen_inst}, \ref{headings}, and \ref{others} below.
\section{General formatting instructions}
\label{gen_inst}
The text must be confined within a rectangle 5.5~inches (33~picas) wide and
9~inches (54~picas) long. The left margin is 1.5~inch (9~picas). Use 10~point
type with a vertical spacing (leading) of 11~points. Times New Roman is the
preferred typeface throughout, and will be selected for you by default.
Paragraphs are separated by \nicefrac{1}{2}~line space (5.5 points), with no
indentation.
The paper title should be 17~point, initial caps/lower case, bold, centered
between two horizontal rules. The top rule should be 4~points thick and the
bottom rule should be 1~point thick. Allow \nicefrac{1}{4}~inch space above and
below the title to rules. All pages should start at 1~inch (6~picas) from the
top of the page.
For the final version, authors' names are set in boldface, and each name is
centered above the corresponding address. The lead author's name is to be listed
first (left-most), and the co-authors' names (if different address) are set to
follow. If there is only one co-author, list both author and co-author side by
side.
Please pay special attention to the instructions in Section \ref{others}
regarding figures, tables, acknowledgments, and references.
\section{Headings: first level}
\label{headings}
All headings should be lower case (except for first word and proper nouns),
flush left, and bold.
First-level headings should be in 12-point type.
\subsection{Headings: second level}
Second-level headings should be in 10-point type.
\subsubsection{Headings: third level}
Third-level headings should be in 10-point type.
\paragraph{Paragraphs}
There is also a \verb+\paragraph+ command available, which sets the heading in
bold, flush left, and inline with the text, with the heading followed by 1\,em
of space.
\section{Citations, figures, tables, references}
\label{others}
These instructions apply to everyone.
\subsection{Citations within the text}
The \verb+natbib+ package will be loaded for you by default. Citations may be
author/year or numeric, as long as you maintain internal consistency. As to the
format of the references themselves, any style is acceptable as long as it is
used consistently.
The documentation for \verb+natbib+ may be found at
\begin{center}
\url{http://mirrors.ctan.org/macros/latex/contrib/natbib/natnotes.pdf}
\end{center}
Of note is the command \verb+\citet+, which produces citations appropriate for
use in inline text. For example,
\begin{verbatim}
\citet{hasselmo} investigated\dots
\end{verbatim}
produces
\begin{quote}
Hasselmo, et al.\ (1995) investigated\dots
\end{quote}
If you wish to load the \verb+natbib+ package with options, you may add the
following before loading the \verb+neurips_2025+ package:
\begin{verbatim}
\PassOptionsToPackage{options}{natbib}
\end{verbatim}
If \verb+natbib+ clashes with another package you load, you can add the optional
argument \verb+nonatbib+ when loading the style file:
\begin{verbatim}
\usepackage[nonatbib]{neurips_2025}
\end{verbatim}
As submission is double blind, refer to your own published work in the third
person. That is, use ``In the previous work of Jones et al.\ [4],'' not ``In our
previous work [4].'' If you cite your other papers that are not widely available
(e.g., a journal paper under review), use anonymous author names in the
citation, e.g., an author of the form ``A.\ Anonymous'' and include a copy of the anonymized paper in the supplementary material.
\subsection{Footnotes}
Footnotes should be used sparingly. If you do require a footnote, indicate
footnotes with a number\footnote{Sample of the first footnote.} in the
text. Place the footnotes at the bottom of the page on which they appear.
Precede the footnote with a horizontal rule of 2~inches (12~picas).
Note that footnotes are properly typeset \emph{after} punctuation
marks.\footnote{As in this example.}
\subsection{Figures}
\begin{figure}
\centering
\fbox{\rule[-.5cm]{0cm}{4cm} \rule[-.5cm]{4cm}{0cm}}
\caption{Sample figure caption.}
\end{figure}
All artwork must be neat, clean, and legible. Lines should be dark enough for
purposes of reproduction. The figure number and caption always appear after the
figure. Place one line space before the figure caption and one line space after
the figure. The figure caption should be lower case (except for first word and
proper nouns); figures are numbered consecutively.
You may use color figures. However, it is best for the figure captions and the
paper body to be legible if the paper is printed in either black/white or in
color.
\subsection{Tables}
All tables must be centered, neat, clean and legible. The table number and
title always appear before the table. See Table~\ref{sample-table}.
Place one line space before the table title, one line space after the
table title, and one line space after the table. The table title must
be lower case (except for first word and proper nouns); tables are
numbered consecutively.
Note that publication-quality tables \emph{do not contain vertical rules.} We
strongly suggest the use of the \verb+booktabs+ package, which allows for
typesetting high-quality, professional tables:
\begin{center}
\url{https://www.ctan.org/pkg/booktabs}
\end{center}
This package was used to typeset Table~\ref{sample-table}.
\begin{table}
\caption{Sample table title}
\label{sample-table}
\centering
\begin{tabular}{lll}
\toprule
\multicolumn{2}{c}{Part} \\
\cmidrule(r){1-2}
Name & Description & Size ($\mu$m) \\
\midrule
Dendrite & Input terminal & $\sim$100 \\
Axon & Output terminal & $\sim$10 \\
Soma & Cell body & up to $10^6$ \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Math}
Note that display math in bare TeX commands will not create correct line numbers for submission. Please use LaTeX (or AMSTeX) commands for unnumbered display math. (You really shouldn't be using \$\$ anyway; see \url{https://tex.stackexchange.com/questions/503/why-is-preferable-to} and \url{https://tex.stackexchange.com/questions/40492/what-are-the-differences-between-align-equation-and-displaymath} for more information.)
\subsection{Final instructions}
Do not change any aspects of the formatting parameters in the style files. In
particular, do not modify the width or length of the rectangle the text should
fit into, and do not change font sizes (except perhaps in the
\textbf{References} section; see below). Please note that pages should be
numbered.
\section{Preparing PDF files}
Please prepare submission files with paper size ``US Letter,'' and not, for
example, ``A4.''
Fonts were the main cause of problems in the past years. Your PDF file must only
contain Type 1 or Embedded TrueType fonts. Here are a few instructions to
achieve this.
\begin{itemize}
\item You should directly generate PDF files using \verb+pdflatex+.
\item You can check which fonts a PDF files uses. In Acrobat Reader, select the
menu Files$>$Document Properties$>$Fonts and select Show All Fonts. You can
also use the program \verb+pdffonts+ which comes with \verb+xpdf+ and is
available out-of-the-box on most Linux machines.
\item \verb+xfig+ "patterned" shapes are implemented with bitmap fonts. Use
"solid" shapes instead.
\item The \verb+\bbold+ package almost always uses bitmap fonts. You should use
the equivalent AMS Fonts:
\begin{verbatim}
\usepackage{amsfonts}
\end{verbatim}
followed by, e.g., \verb+\mathbb{R}+, \verb+\mathbb{N}+, or \verb+\mathbb{C}+
for $\mathbb{R}$, $\mathbb{N}$ or $\mathbb{C}$. You can also use the following
workaround for reals, natural and complex:
\begin{verbatim}
\newcommand{\RR}{I\!\!R} %
\newcommand{\Nat}{I\!\!N} %
\newcommand{\CC}{I\!\!\!\!C} %
\end{verbatim}
Note that \verb+amsfonts+ is automatically loaded by the \verb+amssymb+ package.
\end{itemize}
If your file contains type 3 fonts or non embedded TrueType fonts, we will ask
you to fix it.
\subsection{Margins in \LaTeX{}}
Most of the margin problems come from figures positioned by hand using
\verb+\special+ or other commands. We suggest using the command
\verb+\includegraphics+ from the \verb+graphicx+ package. Always specify the
figure width as a multiple of the line width as in the example below:
\begin{verbatim}
\usepackage[pdftex]{graphicx} ...
\includegraphics[width=0.8\linewidth]{myfile.pdf}
\end{verbatim}
See Section 4.4 in the graphics bundle documentation
(\url{http://mirrors.ctan.org/macros/latex/required/graphics/grfguide.pdf})
A number of width problems arise when \LaTeX{} cannot properly hyphenate a
line. Please give LaTeX hyphenation hints using the \verb+\-+ command when
necessary.
\begin{ack}
Use unnumbered first level headings for the acknowledgments. All acknowledgments
go at the end of the paper before the list of references. Moreover, you are required to declare
funding (financial activities supporting the submitted work) and competing interests (related financial activities outside the submitted work).
More information about this disclosure can be found at: \url{https://neurips.cc/Conferences/2025/PaperInformation/FundingDisclosure}.
Do {\bf not} include this section in the anonymized submission, only in the final paper. You can use the \texttt{ack} environment provided in the style file to automatically hide this section in the anonymized submission.
\end{ack}
\section*{References}
References follow the acknowledgments in the camera-ready paper. Use unnumbered first-level heading for
the references. Any choice of citation style is acceptable as long as you are
consistent. It is permissible to reduce the font size to \verb+small+ (9 point)
when listing the references.
Note that the Reference section does not count towards the page limit.
\medskip
{
\small
[1] Alexander, J.A.\ \& Mozer, M.C.\ (1995) Template-based algorithms for
connectionist rule extraction. In G.\ Tesauro, D.S.\ Touretzky and T.K.\ Leen
(eds.), {\it Advances in Neural Information Processing Systems 7},
pp.\ 609--616. Cambridge, MA: MIT Press.
[2] Bower, J.M.\ \& Beeman, D.\ (1995) {\it The Book of GENESIS: Exploring
Realistic Neural Models with the GEneral NEural SImulation System.} New York:
TELOS/Springer--Verlag.
[3] Hasselmo, M.E., Schnell, E.\ \& Barkai, E.\ (1995) Dynamics of learning and
recall at excitatory recurrent synapses and cholinergic modulation in rat
hippocampal region CA3. {\it Journal of Neuroscience} {\bf 15}(7):5249-5262.
}
\appendix
\section{Technical Appendices and Supplementary Material}
Technical appendices with additional results, figures, graphs and proofs may be submitted with the paper submission before the full submission deadline (see above), or as a separate PDF in the ZIP file below before the supplementary material deadline. There is no page limit for the technical appendices.
\newpage
\section*{NeurIPS Paper Checklist}
The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: {\bf The papers not including the checklist will be desk rejected.} The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page
limit.
Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist:
\begin{itemize}
\item You should answer \answerYes{}, \answerNo{}, or \answerNA{}.
\item \answerNA{} means either that the question is Not Applicable for that particular paper or the relevant information is Not Available.
\item Please provide a short (1–2 sentence) justification right after your answer (even for NA).
\end{itemize}
{\bf The checklist answers are an integral part of your paper submission.} They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper.
The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "\answerYes{}" is generally preferable to "\answerNo{}", it is perfectly acceptable to answer "\answerNo{}" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "\answerNo{}" or "\answerNA{}" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer \answerYes{} to a question, in the justification please point to the section(s) where related material for the question can be found.
IMPORTANT, please:
\begin{itemize}
\item {\bf Delete this instruction block, but keep the section heading ``NeurIPS Paper Checklist"},
\item {\bf Keep the checklist subsection headings, questions/answers and guidelines below.}
\item {\bf Do not modify the questions and only use the provided macros for your answers}.
\end{itemize}
\begin{enumerate}
\item {\bf Claims}
\item[] Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the abstract and introduction do not include the claims made in the paper.
\item The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.
\item The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.
\item It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
\end{itemize}
\item {\bf Limitations}
\item[] Question: Does the paper discuss the limitations of the work performed by the authors?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.
\item The authors are encouraged to create a separate "Limitations" section in their paper.
\item The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.
\item The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.
\item The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.
\item The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.
\item If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.
\item While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
\end{itemize}
\item {\bf Theory assumptions and proofs}
\item[] Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not include theoretical results.
\item All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.
\item All assumptions should be clearly stated or referenced in the statement of any theorems.
\item The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.
\item Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.
\item Theorems and Lemmas that the proof relies upon should be properly referenced.
\end{itemize}
\item {\bf Experimental result reproducibility}
\item[] Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not include experiments.
\item If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.
\item If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.
\item Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.
\item While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example
\begin{enumerate}
\item If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.
\item If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.
\item If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).
\item We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
\end{enumerate}
\end{itemize}
\item {\bf Open access to data and code}
\item[] Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that paper does not include experiments requiring code.
\item Please see the NeurIPS code and data submission guidelines (\url{https://nips.cc/public/guides/CodeSubmissionPolicy}) for more details.
\item While we encourage the release of code and data, we understand that this might not be possible, so “No” is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).
\item The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (\url{https://nips.cc/public/guides/CodeSubmissionPolicy}) for more details.
\item The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.
\item The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.
\item At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).
\item Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
\end{itemize}
\item {\bf Experimental setting/details}
\item[] Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not include experiments.
\item The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.
\item The full details can be provided either with the code, in appendix, or as supplemental material.
\end{itemize}
\item {\bf Experiment statistical significance}
\item[] Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not include experiments.
\item The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
\item The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).
\item The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)
\item The assumptions made should be given (e.g., Normally distributed errors).
\item It should be clear whether the error bar is the standard deviation or the standard error of the mean.
\item It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96\% CI, if the hypothesis of Normality of errors is not verified.
\item For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).
\item If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
\end{itemize}
\item {\bf Experiments compute resources}
\item[] Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not include experiments.
\item The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.
\item The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.
\item The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
\end{itemize}
\item {\bf Code of ethics}
\item[] Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics \url{https://neurips.cc/public/EthicsGuidelines}?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.
\item If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.
\item The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
\end{itemize}
\item {\bf Broader impacts}
\item[] Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that there is no societal impact of the work performed.
\item If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.
\item Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.
\item The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.
\item The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.
\item If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
\end{itemize}
\item {\bf Safeguards}
\item[] Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper poses no such risks.
\item Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.
\item Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.
\item We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
\end{itemize}
\item {\bf Licenses for existing assets}
\item[] Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not use existing assets.
\item The authors should cite the original paper that produced the code package or dataset.
\item The authors should state which version of the asset is used and, if possible, include a URL.
\item The name of the license (e.g., CC-BY 4.0) should be included for each asset.
\item For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.
\item If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, \url{paperswithcode.com/datasets} has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.
\item For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.
\item If this information is not available online, the authors are encouraged to reach out to the asset's creators.
\end{itemize}
\item {\bf New assets}
\item[] Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not release new assets.
\item Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.
\item The paper should discuss whether and how consent was obtained from people whose asset is used.
\item At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
\end{itemize}
\item {\bf Crowdsourcing and research with human subjects}
\item[] Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.
\item Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.
\item According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
\end{itemize}
\item {\bf Institutional review board (IRB) approvals or equivalent for research with human subjects}
\item[] Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.
\item Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.
\item We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.
\item For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
\end{itemize}
\item {\bf Declaration of LLM usage}
\item[] Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required.
\item[] Answer: \answerTODO{} %
\item[] Justification: \justificationTODO{}
\item[] Guidelines:
\begin{itemize}
\item The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.
\item Please refer to our LLM policy (\url{https://neurips.cc/Conferences/2025/LLM}) for what should or should not be described.
\end{itemize}
\end{enumerate}
\end{document}
Collective action against algorithmic systems, which enables groups to promote their own interests, is poised to grow. Hence, there will be growth in the size and the number of distinct collectives. Currently, there is no formal analysis of how coordination challenges within a collective can impact downstream outcomes, or how multiple collectives may affect each other's success. In this work, we aim to provide guarantees on the success of collective action in the presence of both coordination noise and multiple groups. Our insight is that data generated by either multiple collectives or by coordination noise can be viewed as originating from multiple data distributions.
Using this framing, we derive bounds on the success of collective action. We conduct experiments to study the effects of noise on collective action. We find that sufficiently high levels of noise can reduce the success of collective action. In certain scenarios, large noise can sink a collective success rate from $100\%$ to just under $60\%$. We identify potential trade-offs between collective size and coordination noise; for example, a collective that is twice as big but with four times more noise experiencing worse outcomes than the smaller, more coordinated one.
This work highlights the importance of understanding nuanced dynamics of strategic behavior in algorithmic systems.
\section{Introduction}
As large platforms (social media, e-commerce) grow, groups within them find it increasingly important and necessary to act strategically (e.g., by changing ratings) in these platforms to get their desired outcome. From delivery drivers coordinating to earn higher wages \cite{sigg_decline_2024}, to fans promoting their favorite artist \cite{xiao_lets_2025, Nugent_2021}, to protesting controversial businesses \cite{payne2024review}, this type of strategic behavior is poised to grow. Strategic behavior ultimately affects the downstream data distributions used to train models. As the incentive to participate increases, we need a more nuanced understanding of the impacts on a system; this includes both understanding what happens with multiple distinct collectives acting on a system as well as considering how internal collective dynamics may impact outcomes. These factors will all impact the downstream data distribution. However, currently, there are no tools or analytical frameworks that incorporate these factors (coordination noise, multiple collectives) to assess their impact on the success of collective action.
In classical collective action problems related to managing common-pool resources, Ostrom notes that smaller, more homogeneous groups are more effective at overcoming barriers to collective action~\cite{ostrom_governing_1990}. She finds that a clear organizational structure and well-defined roles are crucial—one possible interpretation is that smaller groups can more readily and faithfully implement and coordinate their actions. An analogy in the computational world is “noise”: unwanted modifications or deviations from intended behavior (e.g., signal processing). This “noise” can result from poor coordination and may reduce effectiveness.
In algorithmic settings, collective action is mediated through strategic data modifications to influence model behavior. These changes in data distribution can have significant downstream ramifications. \cite{hardt_algorithmic_2023} provided theoretical bounds on the success of collective action in the case of a single, unified collective. The collective modifies their data (e.g., by inserting a watermark), which is then used in training. The learning algorithm observes a distribution $\mathcal{P}$: a mixture of data induced by the collective's action $(\mathcal{P}_1)$ and the underlying data distribution $\mathcal{P}_0$. These bounds help identify the conditions under which collective action may be more successful. However, this analysis assumes perfect coordination within the collective. \cite{karan2025algorithmiccollectiveactioncollectives} developed a framework to identify the factors affecting collective action involving multiple collectives, using simulations to illustrate several outcomes. However, they do not provide any theoretical bounds on success based on the data distributions induced by each collective. A formal understanding of how both noise and multiple collectives interact is key to understanding real-world algorithmic collective action.
In this work, we provide new guarantees on algorithmic collective action in both the presence of noise and multiple distinct collectives. We do this by introducing \textit{multiple distributions} that feed the ultimate observed data distribution $\mathcal{P}$. Instead of considering just a mixture of $\mathcal{P}_0$ and $\mathcal{P}_1$, we study how the addition of $\mathcal{P}_2$ affects a collective's success. Importantly, the source of these distributions can vary: it could occur from multiple distinct collectives or because of coordination challenges within a single collective inducing a new, second data distribution.
A collective that is highly coordinated is one where the size of $\mathcal{P}_2$ is small and similar to $\mathcal{P}_1$, while ones with less cohesion have a larger gap between $\mathcal{P}_1$ and $\mathcal{P}_2$. Multiple collectives can be analyzed as behavior coming from two distinct distributions. Our contributions are as follows:
\begin{description}[style=unboxed,leftmargin= .25cm]
\item [Theoretical bounds with multiple distributions:] We are the first to establish the lower bounds for the success of collective action in the presence of multiple distributions. Prior work has either only analyzed a single distribution coming from a single collective or provided empirical outcomes with multiple collectives \cite{karan2025algorithmiccollectiveactioncollectives, hardt_algorithmic_2023}. We are able to relate this bound to the similarity of the distributions and collective size. This approach can effectively handle many different scenarios: including when there's a fraction of a collective performing actions imperfectly or when multiple distinct collectives are acting upon a system.
\item[Empirical impact of noise:] We study the impact of noise in collective action. Prior work \cite{hardt_algorithmic_2023} has only shown outcomes with perfectly coordinated collectives. We find that different types of noise can impact the success of collective action and find in some situations, smaller, less noisy groups perform better than larger groups with more noise.
\end{description}
We first present a formulation of the algorithmic collective action problem. We establish bounds on the success of collective action with multiple distributions. Our experiments extend \cite{hardt_algorithmic_2023} by considering a text classification task with noise. We find scenarios where noise sinks the success of a task from nearly $100\%$ without noise to under $60\%$. We discuss possible trade-offs in size and noise where small groups with less noise ($0.25\%$ collective size with $10\%$ noise) outperform larger ones with more noise ($0.5\%$ with $40\%$ noise), which can inform organizers of collective action.
\section{Related Work}
Collective action against algorithms has been documented in contexts such as ridesharing~\cite{lei_delivering_2021, wells_just--place_2021, woodside_bottom-up_2021, jeroen_platform_2021, hastie_platform_2020} and in data campaigns aimed at promoting pro-social outcomes~\cite{malchik_data_2022, vincent_can_2021, vincent_data_2021, wu2022reasonable}. \cite{sigg_decline_2024} provides a specific computational model for the delivery driver case, which can be used to analyze incentives and outcomes in these scenarios. \cite{xiao_lets_2025, xiao_it_2025} examine some of the real-world structures involved in promoting online collective action, such as what motivates certain users to act as organizers. They also discuss practical challenges in influencing online marketplaces, including coordinating across different types of devices and teaching users how to engage effectively with the platform to achieve a specific goal. The concept of noise has been used to simulate communication challenges in agent-based modeling~\cite{shirly2013, acquisti_agent-based_2002, beerman_framework_2023} and in modeling bounded rationality~\cite{kahneman2003maps, conlisk1996bounded, RePEc:edn:esedps:169}. However, it has not been studied in the context of online collective action.
Other work has examined collective action in algorithmic settings. \cite{hardt_algorithmic_2023} first examines how different types of strategies can enable small collectives to have a substantial impact on classification outcomes. \cite{baumann2024algorithmic} extends this analysis to examples involving recommender systems. \cite{ben-dov_role_2024} examines how the specifics of the learning algorithm impact collective action. \cite{karan2025algorithmiccollectiveactioncollectives} empirically analyzes the interaction between two collectives. None of these works, however, offer formal guarantees for successful collective action involving multiple data distributions. Our work offers a more comprehensive theoretical understanding of realistic strategic dynamics.
\section{Problem Formulation}
Here we first will describe collective action as described by \cite{hardt_algorithmic_2023}. We will then extend this scenario to include a second distribution. For convenience, \Cref{sec:appendix-notation-table} summarizes the notation.
\subsection{Algorithmic Collective Action with One Distribution}
\label{sec:baseline}
We consider the case where a firm wishes to deploy a classifier $f$ trained on some data. Let $f: \mathcal{X} \rightarrow \mathcal{Y}$. We define a classifier to be $\epsilon_{1}$ suboptimal under a distribution $P_1$ if there exists a $P'$ with $TV(P_1, P') \leq \epsilon_1$ such that $f = \argmax_{y \in \mathcal{Y}} P'(y | x)$ where TV is the total variation distance.
The firm's goal is to minimize the loss with respect to some objective function. The data they train on exists in $\mathcal{Z} = \mathcal{X} \times \mathcal{Y}$ representing the features $\mathcal{X}$ and labels $\mathcal{Y}$. For our purposes, we can think of every data point $z \in \mathcal{Z}$ as belonging to an individual. A collective wishes to work together in order to create a specific outcome on this classifier for a certain set of inputs which we formalize below.
Consider the base distribution $\mathcal{P}_0$ on $\mathcal{Z}$. We define $\mathcal{P}_1$ to be the distribution induced by the collective's intended action. This is operationalized by strategy, $h_1 : \mathcal{Z} \rightarrow \mathcal{Z}$ where $h_1 \in \mathcal{H}$ the set of all available strategies. Examples of potential strategies may include watching a video for a certain amount of time, giving a specific review score, etc. In implementing this strategy, the collective wishes to create an association between the specific inputs to a target label $y^*$.
The collective plants a ``signal", which is done via function $g_1 : \mathcal{X} \rightarrow \mathcal{X}$ which takes some original input $x \sim \mathcal{P}_0$ and modifies it. Examples may include adding certain words to text, adding a watermark to a video, or reviewing an extra product on a marketplace. This planted signal is intended to create the association between the set of inputs generated by $g_1(x)$ and a target label $y^*$.
In some situations, it may also be possible for the collective to act on both the input data $(\mathcal{X})$ and potentially the output label $(\mathcal{Y})$. For the \textit{feature-label} strategy, members can each change their feature and label. For the \textit{feature-only} strategy, they can only change their features. These strategies, using $h_1$ and signal $g_1$, induce the collective's distribution $\mathcal{P}_1$. Let $\alpha_1$ be the fraction of users participating in the collective. We can write the distribution seen by the learning algorithm as a mixture of the distribution of the collective's data and the unmodified data
\[\mathcal{P} = \alpha_1 \mathcal{P}_1 + (1 - \alpha_1) \mathcal{P}_0\]
where $\mathcal{P}_1$ represents the distribution of $h_1(z)$ where $z \sim \mathcal{P}_0$.
The collective's objective is to associate the signal $g_1$ with the target $y^*$. The success rate can be written as
\[S(\alpha_1) = \Pr_{x \sim \mathcal{P}_0}[f(g_1(x)) = y^*] \]
Success also depends on how the signals generated by the collective conflict or compete with data present in the underlying distribution. Let $\mathcal{X}_{1} = \{g_1(x) : x \in \mathcal{X} \}$ the signal set.
We define this as suboptimality gap of $\mathcal{X}_1$ on distribution $\mathcal{P}_0$ for a target $y^*$ as
\[\Delta^0_{1}(y^*) = \max_{x \in \mathcal{X}_1} (\max_{y \in \mathcal{Y}} \mathcal{P}_0(y | x) - \mathcal{P}_0 (y^* | x))\]
We also use $\mathcal{P}_0(\mathcal{X}_1)$ to refer to the background overlap of signal set $\mathcal{X}_1$ - intuitively, this represents how common the modified input $\mathcal{X}_1$ is in the background distribution $\mathcal{P}_0$
Based on these definitions, \cite{hardt_algorithmic_2023} finds a lower bound on the success for this classifier in the single, perfectly coordinated collective for both the \textit{feature-label} strategy and the \textit{feature-only} strategy, restated here.
\begin{theorem}[Feature-label]
\cite{hardt_algorithmic_2023}
The success rate of a collective with size $\alpha_1$ on a $\epsilon_1$ classifier with target $y_1^*$ with signal set $\mathcal{X}_1$ is bounded below by
$$S(\alpha_1) \geq 1 - \frac{1-\alpha_1}{\alpha_1} \mathcal{P}_0(\mathcal{X}_1)\frac{(1 - \epsilon_1)\Delta_1^0(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} $$
\label{thm:feature-label-base}
\end{theorem}
\begin{theorem}[Feature-only]
\cite{hardt_algorithmic_2023}
The success rate of a collective with size $\alpha_1$ on a $\epsilon_1$-suboptimal classifier with target $y_1^*$ with distribution $\mathcal{X}_1$ where there exists a $p$ such that $\mathcal{P}_0(y^* | x) \geq p, \forall x \in \mathcal{X}$ is bounded below by
$$S(\alpha_1) \geq 1 - \frac{1-(1 - \epsilon_1) p - \epsilon_1 \alpha} { (1 - \epsilon_1) p \alpha - \epsilon_1 \alpha} \mathcal{P}_0(\mathcal{X}_1) $$
\label{thm:feature-only-base}
\end{theorem}
However, these theorems do not consider a second distribution. We first develop our multiple distribution setup, which we will use to expand beyond the above bounds.
\subsection{Collective Action with Multiple Distributions}
Here, we extend beyond \cite{hardt_algorithmic_2023} by considering another data distribution. This may arise from internal coordination challenges or a separate collective.
We consider distribution $\mathcal{P}_2$ induced by a strategy $h_2$. $h_2$ can be completely distinct from $h_1$ or can represent a noisy version of $h_1$ (e.g $h_2(z) = h_1(z) + \delta$). $\mathcal{P}_2$ is the distribution of $h_2(z), z \sim \mathcal{P}_0$. For example, suppose a collective aims to create an association between a word $A$ existing at the top of a document and being rated as an important user $y^*$. Members could mistakenly use a similar word $B$ instead of $A$ or place $A$ at the end of the text instead -- this would be represented by $\mathcal{P}_2$.
We can write the final distribution observed by the learning model as
$$\mathcal{P} = \alpha_1 \mathcal{P}_1 + \alpha_2 \mathcal{P}_2 + (1 - \alpha)\mathcal{P}_0$$
where $\alpha_1$ represents the fraction of people who implemented collective one's strategy, $\alpha_2$ represents a separate group's interactions. This can arise because the correct ``action" was not properly shared with all group members, people getting confused about what exactly to do and not executing faithfully, or potential infiltrators within the group.
We let $\alpha = \alpha_1 + \alpha_2$ be the total fraction of people attempting to implement some strategy. We can also equivalently consider $r = \frac{\alpha_1}{\alpha_1 + \alpha_2} = \frac{\alpha_1}{\alpha}$ as the fraction of people implementing the correct strategy.
Suppose a collective that wants to plant a single $g(x)$ where $x \in \mathcal{X}$. Let the signal set $\mathcal{X}_1 = \{g_1(x) : x \in \mathcal{X})$.
Another group implements another signal, which may be a noisy version of $g_1$. Let $g_2$ represent the set of signals produced by the second group.
$\mathcal{X}_2 = \{g_2(x) : x \in \mathcal{X}\}$.
The presence of another distribution requires us to expand upon our definition of suboptimality to consider, across any two distributions, how often the signals from set $\mathcal{X}_i$ are present on the distribution $P_j$. Our expanded definition of suboptimality is:
\[\Delta^j_{i}(y^*) = \max_{x \in \mathcal{X}_i} (\max_{y \in \mathcal{Y}} \mathcal{P}_j(y | x) - \mathcal{P}_j (y^* | x))\]
Intuitively, this measures how ``confusing'' signals from $\mathcal{X}_{i}$ looks to $\mathcal{P}_j$ when targeting $y^*$.
\section{Multiple Distributions and Success of Collective Action}
Based on our expanded definitions, we can quantify the impact that multiple distributions have on the effectiveness of a collective. We consider both the feature-label strategy and the feature-only strategy. Both are important to study, as in different contexts, collectives may be able to alter both their input features and labels, or just their features. We state the theorems for two distributions here and defer the proofs and generalization to $n$ distributions to the \Cref{sec:appendix_n}. While earlier work focused on a single, perfectly coordinated collective, we quantify how either internal disruptions or the presence of a second collective affects the original group's success. Conceptually, the proofs follow by considering the new mixture distribution and the interaction between the new distributions.
For the feature-label strategy, we have the following result (see \Cref{sec:appendix-proof-feature-label} for full proof):
\begin{theorem}[Feature-label with two distributions]
Consider distribution $\mathcal{P}_1$ and $\mathcal{P}_2$ which are distributed according to $h_1(x)$ and $h_2(x)$ respectively, where $x \sim \mathcal{P}_0$. Let $y_1^*$ be the target class. Then success for the first collective against an $\epsilon_1$ classifier to be lower bounded by
\[S(\alpha_1) \geq 1 - \frac{\alpha_2}{\alpha_1} \mathcal{P}_2(\mathcal{X}_1) \frac{(1 - \epsilon_1)\Delta_1^2(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} - \frac{1-\alpha}{\alpha_1} \mathcal{P}_0(\mathcal{X}_1)\frac{(1 - \epsilon_1)\Delta_1^0(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} \]
\label{thm:feature-label-two}
\end{theorem}
Compared with \Cref{thm:feature-label-base}, \Cref{thm:feature-label-two} introduces another term that describes the relationship between $\mathcal{P}_1$ and $\mathcal{P}_2$ independently of the relationship between $\mathcal{P}_0$ and $\mathcal{P}_1$. It consists of the cross-signal overlap between signal set $\mathcal{X}_1$ on distribution $\mathcal{P}_2$ the suboptimality gap of the target $y_1^*$ of signal set $X_1$ and $\mathcal{P}_2$. ($\Delta_1^2(y_1^*)$) and the relative sizes of said groups $\frac{\alpha_2}{\alpha_1}$. \Cref{fig:varying_bounds} illustrates these relationships.
\begin{figure*}[t]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/fraction_vs_p2_suboptimal=0.25.pdf}
\caption{Varying $\mathcal{P}_2(\mathcal{X}_1)$ for fixed $\Delta_2^1(y_1^*) = 0.25$}
\label{fig:varying_suboptimal}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width = \linewidth]{figures/fraction_vs_suboptimal_gaps_p2=0.1.pdf}
\caption{Varying $\Delta_2^1(y_1^*)$ for fixed $\mathcal{P}_2(\mathcal{X}_1) = 0.25$}
\label{fig:varying_p2}
\end{subfigure}
\caption{Relationship between $\Delta_2^1(y_1^*)$, $\mathcal{P}_2(\mathcal{X}_1)$ and success bound. We fix $\alpha_1 + \alpha_2 = 0.3$ and vary the proportion $r$. We fix $\epsilon_1 = 0$, $\Delta_0^1 = 0.01$, and $\mathcal{P}_0(\mathcal{X}_1) = 0.01$}
\label{fig:varying_bounds}
\end{figure*}
If we consider the distributions as coming from two collectives with differing objectives, we can also write down the success bound for the second collective as well.
\begin{corollary}
If we consider $\mathcal{P}_2, \mathcal{X}_2$ as coming from a 2nd collective with (potentially) distinct objective, we can write the probability of success as
\[S(\alpha_2) \geq 1 - \frac{\alpha_1}{\alpha_2} \mathcal{P}_1(\mathcal{X}_2) \cdot \frac{(1 - \epsilon_2)\Delta_2^1(y_2^*) + \epsilon_2}{1 - 2\epsilon_2} - \frac{1-\alpha}{\alpha_2} \mathcal{P}_0(\mathcal{X}_2)\frac{(1 - \epsilon_2)\Delta_2^0(y_2^*) + \epsilon_2}{1 - 2\epsilon_2} \]
\end{corollary}
For the feature-only strategy we have the following result (\Cref{sec:appendix-proof-feature-only} for full proof)
\begin{theorem}[Feature-only with two distributions]
Consider distribution $\mathcal{P}_1$ and $\mathcal{P}_2$ which are distributed by $h_1(x)$ and $h_2(x)$ respectively, where $x \sim \mathcal{P}_0$. Suppose there exist a $p$ such that $\mathcal{P}_0(y^* | x) \geq p, \forall x \in \mathcal{X}$. Then success for the first collective against an $\epsilon_1$ classifier (against $P^*_1)$ is lower bounded by
$$S(\alpha_1) \geq 1 - \frac{\alpha_2}{\alpha_1} \cdot \frac{\mathcal{P}_2(X_1^*) \cdot \Delta_1^2(y^*) (1 - \epsilon_1)}{ p(1 - \epsilon_1) - \epsilon_1} - \frac{1-\alpha}{\alpha} \mathcal{P}_0(\mathcal{X}_1) \cdot \frac{ (1 - p)(1 - \epsilon_1) +\epsilon_1}{ p(1 - \epsilon_1) - \epsilon_1} $$
\label{thm:feature-only-two}
\end{theorem}
In this bound we see the same $\mathcal{P}_2(\mathcal{X}_1) \Delta_2^1(y_1^*) \alpha_2 (1 -\epsilon_1)$ term which captures the cross-signal overlap, suboptimality gap and collective size.
These theorems illustrate how and to what extent the presence of a second distribution can hinder the first collective. If we consider the second distribution to be a noisy variation of the first one, this helps relate the noise characteristics to the success rate. For a second collective, these bounds provide insights into which scenarios collectives may be simultaneously successful or hindering each other. We examine these implications in \Cref{sec:discus_multiple_collectives}.
\section{Experimental Setup}
To demonstrate the empirical implications, we study the case of noisy collective action. We extend \cite{hardt_algorithmic_2023} experiment on resume classification. We use the resume dataset introduced by \cite{jiechieu2021skills} to finetune a BERT-based model for a multilabel prediction task. The goal for the classifier is to predict which set of careers someone is suited for based on the text in the resume. Members of the collective intend to plant a signal in their resume to get the resume classified to some target class $y^*$ inserting a specific character in a certain pattern.
The intended signal in this case is to place this specific character every 20 words (full details in \Cref{sec:appendix-experimental-details}).
We vary $r$, the proportion of users who perform an imperfect collective action by performing a noisy variation. Specifically, we consider the following types of noise variations.
\begin{description}[style=unboxed,leftmargin= .25cm]
\item[Correct Character Usage] The collective attempts to place a specific character into the resume. A noisy variation involves using a different character. We consider variations where the wrong character is sampled across a small subset (Random-Subset), as well sampled randomly across all possible characters (Random-Full).
\item[Modification Location] The collective intended action to place their character every $20$ words. A noisy variation places the character in arbitrary locations (Displaced).
\end{description}
These choices are motivated by considering ``benign'' ways a group of people may misinterpret instructions. If the intended instruction is ``Place character `A' every 20 words'', some users may focus more on the 20 words part and not use the right character, or some users may focus on using ``A'' and not place it every 20 words. We define the full set of variations in \Cref{tab:experiment-variation}.
We also consider the context in which the noise occurs.
\begin{description}[style=unboxed,leftmargin= .25cm]
\item[Noised Input] For feature-label strategy, it is possible that the noise only gets applied in the input feature or gets applied to the feature as well as the label.
\item[Target's Underlying Frequency] The target class underlying frequency (in $\mathcal{P}_0$) has shown to play some impact in algorithmic collective action, especially with the feature only strategy \cite{hardt_algorithmic_2023}. We test a high frequency label (Software Developer) as well as a low frequency label (Database Administrator) as the target class.
\end{description}
Based on these factors, we ask the following questions.
\begin{description}[style=unboxed,leftmargin= .25cm]
\item [RQ1:] Does the frequency of noise (larger $r$'s) affect success more than noise variation?
\item [RQ2:] What class of strategies are more sensitive to noisy behaviors: feature-label or feature-only?
\item [RQ3:] How does the background frequency ($\mathcal{P}_0$) of the target class affect sensitivity to noise?
\end{description}
\begin{table}[!ht]
\begin{center}
\begin{tabular}{p{0.2\linewidth} | p{0.7\linewidth}}
\toprule
\textbf{Variation Name} & \textbf{Variation Description} \\
\hline
\midrule
Baseline & All members use the same characters and place every 20 words \\
\hline
Random-Subset & Some members insert a different character than the signal character, chosen from a small, fixed subset, placed every 20 words. \\
\hline
Random -Full & Some members insert a different character than the signal character, chosen from a large, fixed subset, placed every 20 words. \\
\hline
Displaced-Original &Some members place the collective's signal character at arbitrary places in the text \\
\hline
Displaced-Full & Some members insert a randomly picked character from a large subset at arbitrary places in the text. \\
\bottomrule
\end{tabular}
\end{center}
\caption{List of variations for noise deviations}
\label{tab:experiment-variation}
\end{table}
\section{Results}
We finetune \texttt{distilbert-based-uncased} ~\cite{Sanh2019DistilBERTAD} for five epochs with default hyperparameters using Hugging Face transformer library \cite{wolf2020huggingfaces}.
For the feature-label strategies, we vary $\alpha$, the percent of users in the collective from $0-1\%$. For the feature only strategies, we vary $\alpha$'s from $[0\%, 50\%]$.We consider noise rates $r$ from $[0\%,50\%]$ (hence $\alpha_1 = r\alpha$). Noise variations for input text is detailed in \Cref{tab:experiment-variation}. For noising labels, the label is changed uniformly at random across all possible labels. We measure success by calculating how often the target class ($y_1^*$) is predicted for text that has the desired modification $x^*$ in the test set. All figures are shown with a one standard deviation region shaded.
\subsection{RQ1: Different Noise Variations}
Here we consider the several types of noise variations (as described in \Cref{tab:experiment-variation}). As a reminder, these different noise variations change how the different members of the collective may incorrectly implement a strategy. The $x$-axis represents the collective size while the $y$-axis show the collective's success criteria. \Cref{fig:comparisons} shows how different noise variations result in different efficacies, both for low levels of noise (\Cref{fig:comparisons-0.1}) and higher levels (\Cref{fig:comparisons-0.5}). We find that different levels of noise have more of an impact when the collective size is relatively small; for sufficiently high levels of participation, all variants perform similarly and comparable to the baseline even with noise. At this lower levels, there is a sensitivity to the set of ``wrong'' characters to choose from: the Random-Subset strategy performs better than the Random-Full strategy. We see displacement while keeping the original character performs better than the Random-Subset - which keeps the placement consistent. For this specific model, this may imply that it is more sensitive to character coordination than the specific placement. We also find that across the different noise types, the size of the collective seems can overcome noise.
\begin{figure*}[t]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/"Noise_types_comparisions-noise_level=0.1-strategy=both_change-noise_label=0-target=Software_Developer-char=left.pdf"}
\caption{$r= 10\%$}
\label{fig:comparisons-0.1}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width = \linewidth]{figures/"Noise_types_comparisions-noise_level=0.5-strategy=both_change-noise_label=0-target=Software_Developer-char=left.pdf"}
\caption{$r=50\%$ }
\label{fig:comparisons-0.5}
\end{subfigure}
\caption{Comparison of different noise variations for different collective sizes. The x-axis here is the total collective size, measured by the percentage of the total population is participating. The $y$-axis is the success rate of causing resumes with the ``true'' signal to be classified to the target class. The black dashed line represents success rate without noise. The strategy is feature-label where noise is only applied to the features. We observe that making more characters for mistakes leads to lower success (Random-Subset vs Random-Full and Displaced-Original vs Displaced-Full)}
\label{fig:comparisons}
\end{figure*}
\subsection{RQ2: Feature-Label vs Feature-Only}
Here we examine how noise affects success for the feature-label vs feature-only strategy. We fix a level of participation and examine how success changes as noise increases.
\Cref{fig:rq2-software} shows the decline in success for feature-label and feature-only strategies as a function of the percentage of the collective subject to noise. We examine three cases, a feature-label scenario where noise is applied to both the feature and the label (\Cref{fig:feature-label-feature-label}), a feature-label scenario where noise is applied to just the feature (\Cref{fig:feature-label-feature}) and the feature-only scenario (\Cref{fig:feature-only-software}).
In \Cref{fig:feature-label-feature-label} we see with moderate levels of participation ($0.5\%$) that noise has a major impact (from nearly $100\%$ success rate to below $60\%$). Even at higher levels of participation ($1\%$) we see, at the high end, noise impacting success. However, in \Cref{fig:feature-label-feature}, we see relative robustness to noise at the higher levels of participation ($0.5\%$ and above). The impact of noise in the feature-only strategy is a more gradual decline (comparisons are not apples to apples because the feature-only strategy requires higher participation levels for comparable success). Overall, noise that affects labels has a more significant impact on a collective's success.
\begin{figure*}[t]
\centering
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width = \linewidth]{figures/"efficacy_vs_noise_plots_strategy=both_change-noise_label=1-noise_type=random-target=Software_Developer-char=left.pdf"}
\caption{Feature-label Noise }
\label{fig:feature-label-feature-label}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/"efficacy_vs_noise_plots_strategy=both_change-noise_label=0-noise_type=random-target=Software_Developer-char=left.pdf"}
\caption{Feature-label Partial Noise}
\label{fig:feature-label-feature}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/efficacy_vs_noise_plots_strategy=resume_only-noise_label=0-noise_type=random-target=Software_Developer-char=left.pdf}
\caption{Feature-only}
\label{fig:feature-only-software}
\end{subfigure}
\caption{Impact of noise (Random-Subset) on different strategies. We consider various levels of noise for differing levels of participation. Each curve represents a fixed level of participation. The x-axis is the percentage of that group subject to noise. The left represents the feature-label strategy subject to noise on both the features and the labels. The middle represents the feature-label strategy where only the features are subject to noise. The right figure shows the feature-only strategy. We find that when noise affects both the feature and label, the decline in success is significant - especially at lower levels of participation. When noise just affects the inputs, the declines are more modest.}
\label{fig:rq2-software}
\end{figure*}
\subsection{RQ3: Impact of Baseline Frequency}
Here, we examine how noise impacts target classes that appear more frequently in the $\mathcal{P}_0$ distribution compared to less commonly seen ones. \Cref{fig:rq3-feature-label} shows the impact for feature-label strategies. We find that, counterintuitively, at higher levels of participation (e.g $0.25\%$ and above) low baseline classes are less susceptible to noise. %
The opposite holds true in \Cref{fig:rq3-feature-only}, where the high frequency target class sees some change (some positive/some negative) with the presence of noise, the low frequency target class is mostly negatively affected by noise. This could be since, unlike in the feature-label strategy, the feature-only strategy cannot boost the presence of a target class in the training data ; diluting the signal with noise may more directly impact success.
\begin{figure*}[t]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/"efficacy_vs_noise_plots_strategy=both_change-noise_label=0-noise_type=random-target=Software_Developer-char=left.pdf"}
\caption{High baseline ($\approx 50\%$)}
\label{fig:feature-label-high}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width = \linewidth]{figures/"efficacy_vs_noise_plots_strategy=both_change-noise_label=0-noise_type=random-target=Database_Administrator-char=left"}
\caption{Low baseline ($\approx 10\%$)}
\label{fig:feature-label-low}
\end{subfigure}
\caption{Interaction between underlying frequency and susceptibility to noise (Random-subset) for feature-label strategy. Here we see for moderate levels of participation ($0.25\%$ and above), the low baseline scenario is more robust to noise.}
\label{fig:rq3-feature-label}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{figures/efficacy_vs_noise_plots_strategy=resume_only-noise_label=0-noise_type=random-target=Software_Developer-char=left.pdf}
\caption{High frequency class}
\label{fig:feature-only-high}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width = \linewidth]{figures/efficacy_vs_noise_plots_strategy=resume_only-noise_label=0-noise_type=random-target=Database_Administrator-char=left.pdf}
\caption{Low frequency class}
\label{fig:feature-only-low}
\end{subfigure}
\caption{Impact of noise (Random-subset) on the feature-only strategy. Compared to the feature-label strategy, both the high and low frequencies cases are impacted by noise across differing participation levels. The high frequency case, however, sees more gradual declines in the success rate compared to the low frequency case.}
\label{fig:rq3-feature-only}
\end{figure*}
\section{Discussion}
\subsection{Analysis on Multiple Collectives}
\label{sec:discus_multiple_collectives}
\cite{karan2025algorithmiccollectiveactioncollectives} examined how two distinct collectives with different objectives impact each other's efficacy -- in short, when the same signal set is being used to target different classes, both collective's success rate is greatly reduced. This can be explained by the suboptimality gap ($\Delta_1^2(y_1^*)$) and the cross-signal overlap ($\mathcal{P}_2(\mathcal{X}_1)$). This overlap will be high since the two groups use the same signal. Because they are targeting two different classes, the suboptimality gap may also be large. They also find a case where two collectives, with different target classes and different character usage, still sinks both of their success rates. This can also be explained by the cross-signal overlap - if these character modifications look sufficiently ``close" to each other, this term may be large and cause conflicts.
\subsection{Trade-offs on Size and Noise}
As strategic behavior on algorithmic systems continues to grow, understanding how deviations from a single, unified collective impact on a group's objective is crucial for both organizers of collective action and system developers. \cite{xiao_lets_2025} observes the heterogeneity of members of a fan collective in terms of the actions available (e.g based on device type), as well as the detailed instructions they must give to others, which provides room for misinterpretation. This is also similar to parts of a collective action framework mentioned in \cite{karan2025algorithmiccollectiveactioncollectives} regarding the action availability and collective construction (i.e how the group is organized).
This theoretical analysis can also be used to help define how and where organizers can place resources. In particular, organizers can try to determine whether increasing the size of the collective $(\alpha_1)$ is worth the trade-off in potentially increasing cross signal overlap, $\mathcal{P}_2(\mathcal{X}_1)$, or suboptimality gap, $\Delta_1^2(y_1^*)$. This trade-off has a traditional analogy in managing common pool resources \cite{ostrom_governing_1990}: the importance of smaller, homogeneous groups to overcome barriers in collective action. We see it is possible that a smaller group with less noise may be able to outperform a larger group. In \Cref{fig:feature-label-feature-label} we observe a collective of size $0.25\%$ with $10\%$ noise rate outperforms a collective of size $0.5\%$ at $40\%$ noise rate. Noise here can be used to characterize a group's coordination efficiency. Different types of collectives and algorithmic systems might exhibit trade-offs between size and minimizing noise (if such a trade-off exists). For organizers of collective action knowledge of these characteristics can help to decide whether resources should be allocated to expand the total collective size or to more tightly coordinate within a small group.
\subsection{Broader Impacts} \label{sec:ethical}
Algorithmic collective action can generally be used when there's a difference between the goals of those who generate data and are affected by models and those who train and deploy them. Our aim is to continue the work on promoting socially valuable use cases of collective action (\cite{feng2022has, Abebe_2022}); however, we recognize that there could be malicious use cases. We believe that model developers should be aware of the strategic behavior that can lead to different data distributions.
\subsection{Limitations and Future Work} \label{sec:limitations}
We assumed that we could easily segregate between distributions. In some cases, this may be more natural (multiple collectives) while in others (benign noise) may be difficult to do in practice.
We also considered a limited set of ``noise" variations on a small set of data. Future work could explore more ecologically motivated types of deviations, including non-independent noise and adversarial actions. We also only considered a small classification task, different types of learning paradigms may be affected by noise differently -- this would be an important avenue for future work.
\section{Conclusion}
We investigated the role of multiple distributions on the success of collective action. We derived lower bounds on the success rate in relation to the cross-signal overlap and suboptimality gap. We empirically evaluated noise in collective action. We found that noise variation matters, and that noise that affects labels more severally impacts collective action. We note that, for organizers, understanding the trade-offs between group size and the potential noise is important; different systems and types of actions may push organizers to invest more in effective coordination in a small group vs expansion. As strategic interest on algorithmic systems grows, both developers of algorithms and organizers of collective action must be aware of the potential that differing distributions has on system outcomes.
\section{Notation Table} \label{sec:appendix-notation-table}
We use considerable notation for defining relationship between collectives. Here we provide a concise table for reference as well as some intuition behind these symbols.
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{ll}
\toprule
\textbf{Symbol} & \textbf{Description} \\
\midrule
$TV(\cdot, \cdot)$ & Total variation distance between two distributions \\
$\epsilon_1$ & Classifier error \\
$f$ & Classifier trained on the mixture distribution \\
$\mathcal{X}$ & The entire feature space \\
$\mathcal{Y}$ & The set of labels \\
$\mathcal{Z}$ & The space of all training data $\mathcal{X} \times \mathcal{Y}$ \\
$\mathcal{P}$ & Observed mixture distribution \\
$\mathcal{P}_0$ & Base (non-strategic) data distribution \\
$\mathcal{P}_1$ & Distribution induced by collective one's intended strategy \\
$\mathcal{P}_2$ & Distribution induced by noise or a second collective’s behavior \\
$h_1, h_2$ & Strategies applied by collectives or noisy actors \\
$g_1, g_2$ & Signal planting functions that modify the input $x$ \\
$\alpha_1, \alpha_2$ & Fraction of population following strategy $h_1$, $h_2$ respectively \\
$\alpha$ & Total participating fraction: $\alpha = \alpha_1 + \alpha_2$ \\
$r$ & Proportion of correctly aligned members: $r = \frac{\alpha_1}{\alpha}$ \\
$\mathcal{X}_1, \mathcal{X}_2$ & Signal set induced by $g$ (\textit{e.g} $\mathcal{X}_1 = \{ g_1(x) \mid x \in \mathcal{X} \}$) \\
$\mathcal{P}_i(\mathcal{X}_j)$ & Cross-signal overlap of $P_i$ on signal set $X_j$ \\
$\Delta_i^j(y^*)$ & Suboptimality gap for $X_i$ under $P_j$ \\
$S(\alpha_1)$ & Success probability for collective one \\
$y_1^*$ & Target class label for collective one \\
\bottomrule
\end{tabular}
\caption{Summary of notation used in the paper.}
\end{table}
\section{Experimental Details} \label{sec:appendix-experimental-details}
As described in the main body we finetune a \texttt{distilbert-based-uncased} ~\cite{Sanh2019DistilBERTAD} for five epochs with default hyperparameters using Hugging Face transformer library \cite{wolf2020huggingfaces}.
This experimental setting is the same as \cite{hardt_algorithmic_2023, karan2025algorithmiccollectiveactioncollectives}. The resume dataset from \cite{jiechieu2021skills} was split into 20,000 training points and 5,000 test points. The default or baseline strategy was to place a specific character (in this case the `\{' character) every 20 words. We evaluated on targeting the `Software Developer' class, except for the `low frequency class' where the target was `Database Administrator.'
During training, each training point is select with a probability $r$ (the noise fraction) to be affected by a specific noise condition. If the point was selected to be noised, the noise would be applied to the text, and, for certain contains, the label as well.
\begin{description}
\item[Random-Subset:] Instead of the intended character, a character from the list [U+2E18, `!', `?', `·'] was selected uniformly at random to be used.
\item[Random-Full:] Same as above, but the character chosen was from the set of all single characters in the vocabulary of the tokenizer.
\item[Displaced-Original:] Keep the intended character, but instead of placing every 20 words, choose uniformly at random (5-30 instances) places to insert the character.
\item[Displaced-Full:] Combine the character selection of Random-Full with the placement behavior of Displaced-Original.
\end{description}
To evaluate whether the signal was planted successfully, for each data point in the test set, we applied the ``true'' signal and evaluated whether the trained classifier would predict the target class. We used the top-one accuracy as done in \cite{hardt_algorithmic_2023} as the metric of success.
All experimental conditions were run 15 times. Experiments were run on a ppc64le based cluster with V100 Nvidia GPUs. Each iteration took $30-40$ minutes to complete.
\section{Proof of Theorem 3} \label{sec:appendix-proof-feature-label}
\newtheorem*{remark}{Theorem}
We restate the theorem here:
\begin{remark}
Consider distribution $\mathcal{P}_1$ and $\mathcal{P}_2$ which are distributed according to $h_1(x)$ and $h_2(x)$ respectively, where $x \sim \mathcal{P}_0$. Let $y_1^*$ be the target class. Then success for the first collective against an $\epsilon_1$ classifier to be lower bounded by
$$S(\alpha_1) \geq 1 - \frac{\alpha_2}{\alpha_1} \mathcal{P}_2(\mathcal{X}_1) \frac{(1 - \epsilon_1)\Delta_1^2(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} - \frac{1-\alpha}{\alpha_1} \mathcal{P}_0(\mathcal{X}_1)\frac{(1 - \epsilon_1)\Delta_1^0(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} $$
\end{remark}
\begin{proof}
We follow the same proof strategy as \cite{hardt_algorithmic_2023}.
We consider the multiple distributions present the overall data distribution $\mathcal{P}$
$\mathcal{P}_0$ base distribution;
$\mathcal{P}_1$ group 1's distribution;
$\mathcal{P}_2$ group 2's distribution
We write the mixture distribution as :
\begin{equation}
\mathcal{P}(x, y_1^*) = \alpha_1 \mathcal{P}_1(x, y_1^*) + \alpha_2 \mathcal{P}_2(x, y_1^*) + (1 - \alpha) \mathcal{P}_0(x, y_1^*)
\end{equation}
We also define the suboptimality gap on distribution $i$ for signal set $j$ and target label $y^*$
\begin{equation}
\Delta^j_{i}(y^*) = \max_{x \in X_i^*} (\max_{y \in \mathcal{Y}} \mathcal{P}_j(y | x) - \mathcal{P}_j (y^* | x))
\end{equation}
We define the suboptimality gap for a specific point $x$ as
\begin{equation}
\Delta^j_{i, x}(y^*) = \max_{y \in \mathcal{Y}} \mathcal{P}_j(y | x) - \mathcal{P}_j (y^* | x)
\end{equation}
Our goal is to find, for which value of $\alpha_1$ can we guarantee the model to classify a point $x \in \mathcal{X}_1$ to the target class $y_1^*$
First consider an $\epsilon = 0$ classifier. If the model $f$ classifies any point $x \in \mathcal{X}$ to $y_1^*$ , it must mean that
$\mathcal{P}(y_1^* | x) > \mathcal{P}(y_1 | x)$ or equivalently
$\mathcal{P}(x, y_1^*) - \mathcal{P}(x, y_1 ) > 0$
for every $y_1 \neq y_1^*$.
In this strategy, both the features and labels are changed. Since $\mathcal{P}_1$ is the distribution which correctly performs the intended collective action, every point $x_1^* \in \mathcal{X}_1$ maps to $y_1^*$. Therefore we can simplify and get
$$\mathcal{P}(x, y_1^*) = \alpha_1 \mathcal{P}_1(x) + \alpha_2 \mathcal{P}_2(x, y_1^*) + (1 - \alpha) \mathcal{P}_0(x, y_1^*)$$
Now, for $y \neq y_1^*$ we have
$$\mathcal{P}(x, y) = \alpha_1 \mathcal{P}_1(x, y) + \alpha_2 \mathcal{P}_2(x, y) + (1 - \alpha) \mathcal{P}_0(x, y)$$
Given $\mathcal{P}_1$ maps everything to $y_1^*$ and nothing else this first term is $0$ so we can write simplify to
$$\mathcal{P}(x, y) = \alpha_2 \mathcal{P}_2(x, y) + (1 - \alpha) \mathcal{P}_0(x, y)$$
So we can write $\mathcal{P}(x, y_1^*) - \mathcal{P}(x, y) > 0$ as
$$\alpha_1 \mathcal{P}_1(x) + \alpha_2 \mathcal{P}_2(x, y_1^*) + (1 - \alpha) \mathcal{P}_0(x, y_1^*) - (\alpha_2 \mathcal{P}_2(x, y) + (1 - \alpha) \mathcal{P}_0(x, y)) > 0$$
After rearranging we have
$$\alpha_1 \mathcal{P}_1(x) \geq \alpha_2 \mathcal{P}_2(x)*(\mathcal{P}_2(y|x) -\mathcal{P}_2(y_1 | x)) + (1-\alpha) \mathcal{P}_0(x) (\mathcal{P}_0(y | x) - \mathcal{P}_0(y_1^* | x))$$
This must hold true for any $y$ so we can replace the rhs terms by $\Delta^j_{i, x}(y^*)$, in other words we have a sufficient condition for the size of $\alpha_1$ as
$$\alpha_1 \geq \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)}\alpha_2 \Delta^2_{1, x}(y_1^*) + \frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)}(1-\alpha) \Delta^0_{1, x}(y_1^*)$$
\begin{align*}
S(\alpha)& = \Pr_{x\sim \mathcal{P}_1}\left\{f(x)=y_1^*\right\} \\
& \geq \Pr_{x\sim \mathcal{P}_1}\Biggl\{\alpha_1 \geq \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)}\alpha_2 \Delta^2_{1, x}(y_1^*) + \frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)}(1-\alpha) \Delta^0_{1, x}(y_1^*) \Biggl\} \\
& = \E_{x \sim \mathcal{P}_1}\mathbf{1}\Biggl\{\alpha_1 \geq \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)}\alpha_2 \Delta^2_{1, x}(y_1^*) + \frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)}(1-\alpha) \Delta^0_{1, x}(y_1^*) \Biggl\} \\
& = \E_{x \sim \mathcal{P}_1}\mathbf{1}\Biggl\{1 - \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)}\frac{\alpha_2}{\alpha_1} \Delta^2_{1, x}(y_1^*) - \frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)}\frac{1-\alpha}{\alpha_1} \Delta^0_{1, x}(y_1^*) \geq 0 \Biggl\} \\
& \ge \E_{x \sim \mathcal{P}_1}\left[1 - \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)}\frac{\alpha_2}{\alpha_1} \Delta^2_{1, x}(y_1^*) - \frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)}\frac{1-\alpha}{\alpha_1} \Delta^0_{1, x}(y_1^*) \right]\\
& 1 - \frac{\alpha_2}{\alpha_1} \mathbb{E}_{x \sim \mathcal{P}_1} \left[ \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)} \Delta^2_{1, x}(y_1^*) \right] - \frac{1-\alpha}{\alpha_1} \mathbb{E}_{x \sim \mathcal{P}_1} \left[\frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)} \Delta^0_{1, x}(y_1^*) \right] \\
& \geq 1 - \frac{\alpha_2}{\alpha_1} \mathcal{P}_2(\mathcal{X}_1) \Delta^2_{1}(y_1^*) - \frac{1-\alpha}{\alpha_1} \mathcal{P}_0(\mathcal{X}_1) \Delta^0_{1}(y_1^*)\\
\end{align*}
With the final line using the fact that the delta we max over all $x$'s
Now for distribution where $TV(P, P') \leq \epsilon_1$. By Lemma 11 in \cite{hardt_algorithmic_2023} we have that $\mathcal{P'}(y^* | x ) > \mathcal{P'}(y| x)$ when $\mathcal{P}(y^*|x) > \mathcal{P}(y | x) + \frac{\epsilon_{1}}{1 - \epsilon_{1}}$
We than write this as
$$\mathcal{P}(y_1^* | x)\mathcal{P}(x) > \mathcal{P}(y|x)\mathcal{P}(x) + \frac{\epsilon_1}{1 - \epsilon_1}\mathcal{P}(x)$$
Following the same steps, procedure as before we get
$$\alpha_1 > \alpha_2 \frac{\mathcal{P}_2(x)}{\mathcal{P}_1(x)}\biggl(\frac{(1 -\epsilon_1)\Delta_2^1(y_1^*) + \epsilon_1}{1 + 2\epsilon_1} \biggl) + (1-\alpha)\frac{\mathcal{P}_0(x)}{\mathcal{P}_1(x)} \biggl(\frac{(1 - \epsilon_1)\Delta_0^1(y_1^*) + \epsilon_1}{1+2\epsilon_1}\biggl)$$
Computing the $S(\alpha_1)$ again we get the condition we get
$$S(\alpha_1) \geq 1 - \frac{\alpha_2}{\alpha_1} \mathcal{P}_2(\mathcal{X}_1) \cdot \frac{(1 - \epsilon_1)\Delta_2^1(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} - \frac{1-\alpha}{\alpha_1} \mathcal{P}_0(\mathcal{X}_1)\frac{(1 - \epsilon_1)\Delta_0^1(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} $$
\end{proof}
\section{Proof of Theorem 4} \label{sec:appendix-proof-feature-only}
We restate the theorem:
\begin{remark}[Feature-only with two distributions]
Consider distribution $\mathcal{P}_1$ and $\mathcal{P}_2$ which are distributed by $h_1(x)$ and $h_2(x)$ respectively, where $x \sim \mathcal{P}_0$. Suppose there exist a $p$ such that $\mathcal{P}_0(y^* | x) \geq p, \forall x \in \mathcal{X}$. Then success for the first collective against an $\epsilon_1$ classifier (against $P^*_1)$ is lower bounded by
$$S(\alpha_1) \geq 1 - \frac{\alpha_2}{\alpha_1} \cdot \frac{\mathcal{P}_2(X_1^*) \cdot \Delta_1^2(y^*) (1 - \epsilon_1)}{ p(1 - \epsilon_1) - \epsilon_1} - \frac{1-\alpha}{\alpha} \mathcal{P}_0(\mathcal{X}_1) \cdot \frac{ (1 - p)(1 - \epsilon_1) +\epsilon_1}{ p(1 - \epsilon_1) - \epsilon_1} $$
\end{remark}
\begin{proof}
Similar to Theorem 3, we start with the $\epsilon = 0$. We require that, for any $x^* \in \mathcal{X}_1$ that $\mathcal{P}(y_1^* | x^* ) > \mathcal{P}(y | x^*)$ $\forall y \neq y_1^*$. This is equivalent to $\mathcal{P}(x^*, y_1^*) > \mathcal{P}(x^*, y)$ Our goal is find a sufficient condition for $\alpha_1$
We get a lower bound for $\mathcal{P}(x^*, y^*)$ by writing the mixture distribution as.
$$\mathcal{P}(x^*, y^*) = \alpha_1 \mathcal{P}_1(x^*, y^*) + \alpha_2 \mathcal{P}_2(x^*, y^*) + (1 - \alpha) \mathcal{P}_0(x^*, y^*) \geq \alpha \mathcal{P}_1(x^*, y^*) + \alpha_2 \mathcal{P}_2(x^*, y^*) $$ where $\alpha = \alpha_1 + \alpha_2$
We also have when $y \neq y_1^*$ that $$\mathcal{P}(x^*, y) = \alpha_1 \mathcal{P}_1(x^*, y) + \alpha_2 \mathcal{P}_2(x^*, y) + (1 - \alpha) \mathcal{P}_0(x^*, y) = \alpha_2 \mathcal{P}_2(x^*, y) + (1 - \alpha) \mathcal{P}_0(x^*, y)$$
Hence if
$$\alpha_1 \mathcal{P}_1(x^*, y^*) + \alpha_2 \mathcal{P}_2(x^*, y^*) \geq \alpha_2 \mathcal{P}_2(x^*, y) + (1 - \alpha) \mathcal{P}_0(x^*, y)$$
than $\mathcal{P}(x^*, y^*) > \mathcal{P}(x^*, y)$
Rearranging we have
$$\alpha_1 \mathcal{P}_1(x^*, y^*) \geq \alpha_2 \mathcal{P}_2(x^*, y) - \alpha_2 \mathcal{P}_2(x^*, y^*) + (1 - \alpha) \mathcal{P}_0(x^*, y)$$
$$\alpha_1 \mathcal{P}_1(x^*, y^*) \geq \alpha_2 \mathcal{P}_2(x^*) ( \mathcal{P}_2(y | x^*) - \mathcal{P}_2(y^* | x^*)) + (1 - \alpha) \mathcal{P}_0(y | x^*) \mathcal{P}_0(x^*)$$
$$\alpha_1 \geq \frac{\alpha_2 \mathcal{P}_2(x^*) ( \mathcal{P}_2(y | x^*) - \mathcal{P}_2(y^* | x^*))}{\mathcal{P}_1(x^*, y^*)} + (1 - \alpha) \frac{\mathcal{P}_0(y | x^*) \mathcal{P}_0(x^*)}{\mathcal{P}_1(x^*, y^*)}$$
For our bound on $\alpha_1$ we seek to maximize the rhs. We can do this by upper bounding $ \mathcal{P}_2(y | x^*) - \mathcal{P}_2(y^* | x^*))$ by the suboptimality gap $\Delta_1^2$. By assumption we have that $\mathcal{P}_0(y^* | x) \geq p$ for all $x \in \mathcal{X}$
and hence $\mathcal{P}_0(y | x^*) \leq 1-p$. For the denominator we note that $\mathcal{P}_1(x^*, y*) \geq \mathcal{P}_0(g^{-1}(x^*), y^*) \geq p \mathcal{P}_0(g^{-1}(x^*)$ Using this lowerbound in the denominator we get our required alpha must be at least
$$\alpha_1 \geq \alpha_2 \cdot \frac{ \mathcal{P}_2(x^*) \Delta^2_1}{ p \mathcal{P}_0(g^{-1}(x^*)) } + (1 - \alpha) \frac{(1 - p)\mathcal{P}_0(x^*)}{p \mathcal{P}_0(g^{-1}(x^*))}$$
To compute the success rate we have
$$S(\alpha) = Pr_{x \sim \mathcal{P}_0^*}\{f(g(x)) = y^*\}$$
$$S= \sum_{x^* \in \mathcal{X}_1} \Pr_{x \sim \mathcal{P}_0^*} \{f(g(x)) = y^* |x \in g^{-1} (x^*) \} Pr_{x \in \mathcal{P}_0} \{x \in g^{-1}(x^*)\}$$
$$S= \sum_{x^* \in \mathcal{X}_1} \mathbf{1}\{f(x^*) = y^*\} P_{0}(g^{-1}(x^*))\}$$
For a given fixed $x^*$ we have
$$\mathbf{1}\Biggl\{f(x^*) = y^* \biggl\} \geq \mathbf{1} \biggl\{\alpha_1 \geq \alpha_2 \cdot \frac{ \mathcal{P}_2(x^*) \Delta^2_1}{ p \mathcal{P}_0(g^{-1}(x^*)) } + (1 - \alpha) \frac{(1 - p)\mathcal{P}_0(x^*)}{p \mathcal{P}_0(g^{-1}(x^*))} \Biggl \}$$
$$=\mathbf{1}\Biggl\{1 - \frac{\alpha_2}{\alpha_1} \cdot \frac{ \mathcal{P}_2(x^*) \Delta^2_1}{ p \mathcal{P}_0(g^{-1}(x^*)) } - \frac{(1 - \alpha)}{\alpha_1} \frac{(1 - p)\mathcal{P}_0(x^*)}{p \mathcal{P}_0(g^{-1}(x^*))} > 0 \Biggl \}$$
$$\geq 1 - \frac{\alpha_2}{\alpha_1} \cdot \frac{ \mathcal{P}_2(x^*) \Delta^2_1}{ p \mathcal{P}_0(g^{-1}(x^*)) } - \frac{(1 - \alpha)}{\alpha_1} \frac{(1 - p)\mathcal{P}_0(x^*)}{p \mathcal{P}_0(g^{-1}(x^*))} $$
Computing the summation we get
$$Pr_{x \sim \mathcal{P}_0^*}\{f(g(x)) = y^*\}$$
$$= 1 - \sum_{x^* \in \mathcal{X}_1} \frac{\alpha_2}{\alpha_1} \cdot \frac{ \mathcal{P}_2(x^*) \Delta^2_1}{ p \mathcal{P}_0(g^{-1}(x^*)) } \cdot \mathcal{P}_0(g_1^{-1}(x^*)) - \sum_{x^* \in \mathcal{X}_1} \frac{(1 - \alpha)}{\alpha_1} \frac{(1 - p)\mathcal{P}_0(x^*)}{p \mathcal{P}_0(g^{-1}(x^*))} \cdot \mathcal{P}_0(g_1^{-1}(x^*))$$
$$\geq 1 - \frac{\alpha_2}{p\alpha_1}\mathcal{P}_2(\mathcal{X}_1) \Delta_1^2 - \frac{(1 - \alpha) \cdot (1 - p)}{p\alpha_1} \mathcal{P}_0(X^*) $$
Which gives us the $\epsilon = 0$ case.
When considering $\epsilon > 0$, we again cite Lemma 11 in \cite{hardt_algorithmic_2023}, that: $\mathcal{P}(x^*, y^*) > \mathcal{P}(x^*, y) + \frac{\epsilon}{1-\epsilon} \mathcal{P}(x^*)$
Expanding all terms out with the mixture distribution we have
$$ \mathcal{P}(x^*, y) = \alpha \mathcal{P}_1(x^*, y^*) + \alpha_2 \mathcal{P}_2(x^*, y^*) + (1 - \alpha_1 - \alpha_2) \mathcal{P}_0(x^*, y^*) \geq \alpha \mathcal{P}_1(x^*, y^*) + \alpha_2 \mathcal{P}_2(x^*, y^*)$$
Plugging this into our expression we have
$$ \alpha \mathcal{P}_1(x^*, y^*) + \alpha_2 \mathcal{P}_2(x^*, y^*) > \alpha_2 * \mathcal{P}_2(x^*, y) + (1 - \alpha) \mathcal{P}_0(x^*, y) + \frac{\epsilon}{1-\epsilon} ( \alpha \mathcal{P}_1(x^*) + \alpha_2 \mathcal{P}_2(x^*)+ (1 - \alpha) \mathcal{P}_0(x^*))$$
Using conditionals we can rewrite this as
$$ \alpha_1 \mathcal{P}_1(y^* | x^*) \mathcal{P}_1(x^*) + \alpha_2 \mathcal{P}_2(y^* | x^*) \mathcal{P}_2(x^*) >$$
$$\alpha_2 * \mathcal{P}_2(y | x^*)\mathcal{P}_2(x^*) + (1 - \alpha) \mathcal{P}_0(y | x^*) \mathcal{P}_0(x^*) + \frac{\epsilon}{1-\epsilon} ( \alpha_1 \mathcal{P}_1(x^*) + \alpha_2 \mathcal{P}_2(x^*)+ (1 - \alpha) \mathcal{P}_0(x^*))$$
Rearranging with $\alpha_1$ we get
$$ \alpha_1 > \alpha_2 \frac{\mathcal{P}_2(x^*)}{\mathcal{P}_1(x^*)} \cdot \frac{(\mathcal{P}_2(y | x^*) - \mathcal{P}_2(y^* | x^*))}{\mathcal{P}_1(y^* | x^*) - \frac{\epsilon}{1-\epsilon}} + (1-\alpha)\frac{\mathcal{P}_0(x^*)}{\mathcal{P}_1(x^*)} \cdot \frac{(\mathcal{P}_0(y | x^*) + \frac{\epsilon}{1-\epsilon})}{(\mathcal{P}_1(y^* | x^*) - \frac{\epsilon}{1-\epsilon})}$$
We note that $\mathcal{P}_2(y | x^*) - \mathcal{P}_2(y^*| x^*)$ can be upper bounded by $\Delta_1^2(y^*)$ We can also upper bound and lower bound the numerators and denominators with the following expressions
$\mathcal{P}_1(x^*) \geq \mathcal{P}_0(g^{-1}(x^*))$ as every point $x^*$ for $\mathcal{P}_1$ is mapped from some point in $\mathcal{P}_0$ and likewise
$\mathcal{P}_1(y^* | x^*) \geq \mathcal{P}_0(y^* | g_1^{-1}(x^*)) \geq p$
From the support of $P_0(y_1^* | x) \geq p$ we have
$\mathcal{P}_0(y| x^*) \leq 1 - p$ for $y \neq y_1^*$
All together we get
$$ \alpha_1 > \alpha_2 \frac{\mathcal{P}_2(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{\Delta_1^2(y^*)}{p - \frac{\epsilon}{1-\epsilon}} + (1-\alpha)\frac{\mathcal{P}_0(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{(1-p + \frac{\epsilon}{1-\epsilon})}{(p - \frac{\epsilon}{1-\epsilon})}$$
With this we can compute the expected value as above
For a given fixed $x^*$ we have
$$\mathbf{1}\Biggl\{f(x^*) = y^* \biggl\} \geq \mathbf{1} \biggl\{ \alpha_1 > \alpha_2 \frac{\mathcal{P}_2(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{\Delta_1^2(y^*)}{p - \frac{\epsilon}{1-\epsilon}} + (1-\alpha)\frac{\mathcal{P}_0(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{(1-p + \frac{\epsilon}{1-\epsilon})}{(p - \frac{\epsilon}{1-\epsilon})} \Biggl \}$$
$$=\mathbf{1}\Biggl\{1 - \frac{\alpha_2}{\alpha_1} \frac{\mathcal{P}_2(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{\Delta_1^2(y^*)}{p - \frac{\epsilon}{1-\epsilon}} - \frac{1-\alpha}{\alpha_1}\frac{\mathcal{P}_0(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{(1-p + \frac{\epsilon}{1-\epsilon})}{(p - \frac{\epsilon}{1-\epsilon})} > 0 \Biggl \}$$
$$\geq 1 - \frac{\alpha_2}{\alpha_1} \frac{\mathcal{P}_2(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{\Delta_1^2(y^*)}{p - \frac{\epsilon}{1-\epsilon}} - \frac{1-\alpha}{\alpha_1}\frac{\mathcal{P}_0(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{(1-p + \frac{\epsilon}{1-\epsilon})}{(p - \frac{\epsilon}{1-\epsilon})} $$
Plugging back in we get
$$Pr_{x \sim \mathcal{P}_0^*}\{f(g(x)) = y^*\}$$
$$= 1 - \sum_{x^* \in \mathcal{X}_1}\frac{\alpha_2}{\alpha_1} \frac{\mathcal{P}_2(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{\Delta_1^2(y^*)}{p - \frac{\epsilon}{1-\epsilon}} \cdot \mathcal{P}_0(g_1^{-1}(x^*)) - \sum_{x^* \in \mathcal{X}_1} \frac{1-\alpha}{\alpha_1}\frac{\mathcal{P}_0(x^*)}{\mathcal{P}_0(g_1^{-1}(x^*))} \cdot \frac{(1-p + \frac{\epsilon}{1-\epsilon})}{(p - \frac{\epsilon}{1-\epsilon})} \cdot \mathcal{P}_0(g_1^{-1}(x^*))$$
$$\geq 1 - \frac{\alpha_2}{\alpha_1} \cdot \frac{\mathcal{P}_2(X^*) \cdot \Delta_1^2(y^*)}{ p - \frac{\epsilon}{1 - \epsilon}} - \frac{1-\alpha}{\alpha} \mathcal{P}_0(X^*) \cdot \frac{ 1 - p + \frac{\epsilon}{1-\epsilon}}{ p - \frac{\epsilon}{1 - \epsilon}} $$
$$\geq 1 - \frac{\alpha_2}{\alpha_1} \cdot \frac{\mathcal{P}_2(X^*) \cdot \Delta_1^2(y^*) (1 - \epsilon)}{ p(1 - \epsilon) - \epsilon} - \frac{1-\alpha}{\alpha} \mathcal{P}_0(X^*) \cdot \frac{ (1 - p)(1 - \epsilon) +\epsilon}{p(1 - \epsilon) - \epsilon} $$
\end{proof}
\section{Extension to many collectives} \label{sec:appendix_n}
We can extend the mixture distribution to include an arbitrary number of distributions. This can be written as
$$\mathcal{P} = \sum_{i = 1}^n \alpha_i\mathcal{P}_i + (1 - \alpha) \mathcal{P}_0$$ where $\alpha = \sum_{i=1}^n \alpha_i$
By following the same procedure in \Cref{sec:appendix-proof-feature-label} we provide a claim for success for collective one against an arbitrary number of distributions.
\begin{theorem}
The success of collective action for the feature label strategy against many distributions is lower bounded by
$$S(\alpha_1) \geq 1 - \sum_{i = 2}^n\frac{\alpha_i}{\alpha_1} \mathcal{P}_i(\mathcal{X}_1) \cdot \frac{(1 - \epsilon_1)\Delta_i^1(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} - \frac{1-\alpha}{\alpha_1} \mathcal{P}_0(\mathcal{X}_1)\frac{(1 - \epsilon_1)\Delta_0^1(y_1^*) + \epsilon_1}{1 - 2\epsilon_1} $$
\end{theorem}
\begin{proof}
We repeat the same expansion in \Cref{sec:appendix-proof-feature-label} for an arbitrary number of distributions.
$$\mathcal{P} = \sum_{i = 1}^n \alpha_i\mathcal{P}_i + (1 - \alpha) \mathcal{P}_0$$
The rest follows the same procedure, as each additional mixture term can be handled independently.
\end{proof}
We can do the same for feature-only strategies.
\begin{theorem}
The success of collective action for the feature-only strategy against many distributions is lower bounded by
$$\geq 1 - \sum_{i = 2}^n \frac{\alpha_i}{\alpha_1} \cdot \frac{\mathcal{P}_i(\mathcal{X}_1) \cdot \Delta_1^i(y_1^*) (1 - \epsilon)}{ p(1 - \epsilon) - \epsilon} - \frac{1-\alpha}{\alpha} \mathcal{P}_0(X^*) \cdot \frac{ (1 - p)(1 - \epsilon) +\epsilon}{p(1 - \epsilon) - \epsilon} $$
\end{theorem}
\begin{proof}
We repeat the same expansion in \Cref{sec:appendix-proof-feature-only} for an arbitrary number of distributions with the same expansion of the mixture distribution
and the rest follows the same procedure, as each additional mixture term can be handled independently.
\end{proof}
\newcommand{\etalchar}[1]{$^{#1}$}
\begin{thebibliography}{HMMDZ23}
\bibitem[AHJ{\etalchar{+}}22]{Abebe_2022}
Rediet Abebe, Moritz Hardt, Angela Jin, John Miller, Ludwig Schmidt, and Rebecca Wexler.
\newblock Adversarial scrutiny of evidentiary statistical software.
\newblock In {\em 2022 ACM Conference on Fairness Accountability and Transparency}, FAccT ’22, page 1733–1746. ACM, June 2022.
\bibitem[ASJCB02]{acquisti_agent-based_2002}
Alessandro Acquisti, Maarten Sierhuis, William J.~Clancey, and Jeffrey Bradshaw.
\newblock Agent-{Based} {Modeling} of {Collaboration} and {Work} {Practices} {Onboard} the {International} {Space} {Station}, January 2002.
\bibitem[BBG23]{beerman_framework_2023}
Jack~T. Beerman, Gwendal~G. Beaumont, and Philippe~J. Giabbanelli.
\newblock A framework for the comparison of errors in agent-based models using machine learning.
\newblock {\em Journal of Computational Science}, 72:102119, September 2023.
\bibitem[BDFSS24]{ben-dov_role_2024}
Omri Ben-Dov, Jake Fawkes, Samira Samadi, and Amartya Sanyal.
\newblock The {Role} of {Learning} {Algorithms} in {Collective} {Action}, June 2024.
\newblock arXiv:2405.06582 [cs].
\bibitem[BMD24]{baumann2024algorithmic}
Joachim Baumann and Celestine Mendler-D{\"u}nner.
\newblock Algorithmic collective action in recommender systems: Promoting songs by reordering playlists.
\newblock In {\em The Thirty-eighth Annual Conference on Neural Information Processing Systems}, 2024.
\bibitem[Con96]{conlisk1996bounded}
John Conlisk.
\newblock Why bounded rationality?
\newblock {\em Journal of economic literature}, 34(2):669--700, 1996.
\bibitem[FS22]{feng2022has}
Yunhe Feng and Chirag Shah.
\newblock Has ceo gender bias really been fixed? adversarial attacking and improving gender fairness in image search.
\newblock In {\em Proceedings of the AAAI Conference on Artificial Intelligence}, volume~36, pages 11882--11890, 2022.
\bibitem[Has20]{hastie_platform_2020}
Bethany Hastie.
\newblock Platform {Workers} and {Collective} {Labour} {Action} in the {Modern} {Economy}.
\newblock {\em University of New Brunswick Law Journal}, 71:40, 2020.
\bibitem[HMMDZ23]{hardt_algorithmic_2023}
Moritz Hardt, Eric Mazumdar, Celestine Mendler-Dünner, and Tijana Zrnic.
\newblock Algorithmic {Collective} {Action} in {Machine} {Learning}, June 2023.
\newblock arXiv:2302.04262 [cs, stat].
\bibitem[JGV21]{jeroen_platform_2021}
Meijerink Jeroen, Jansen Giedo, and Daskalova Victoria.
\newblock {\em Platform {Economy} {Puzzles}: {A} {Multidisciplinary} {Perspective} on {Gig} {Work}}.
\newblock Edward Elgar Publishing, August 2021.
\newblock Google-Books-ID: cXg8EAAAQBAJ.
\bibitem[JT21]{jiechieu2021skills}
Kameni Florentin~Flambeau Jiechieu and Norbert Tsopze.
\newblock Skills prediction based on multi-label resume classification using cnn with model predictions explanation.
\newblock {\em Neural Computing and Applications}, 33(10):5069--5087, 2021.
\bibitem[Kah03]{kahneman2003maps}
Daniel Kahneman.
\newblock Maps of bounded rationality: Psychology for behavioral economics.
\newblock {\em American economic review}, 93(5):1449--1475, 2003.
\bibitem[KVKS25]{karan2025algorithmiccollectiveactioncollectives}
Aditya Karan, Nicholas Vincent, Karrie Karahalios, and Hari Sundaram.
\newblock Algorithmic collective action with two collectives, 2025.
\bibitem[Lei21]{lei_delivering_2021}
Ya-Wen Lei.
\newblock Delivering {Solidarity}: {Platform} {Architecture} and {Collective} {Contention} in {China}’s {Platform} {Economy}.
\newblock {\em American Sociological Review}, 86(2):279--309, April 2021.
\newblock Publisher: SAGE Publications Inc.
\bibitem[MF22]{malchik_data_2022}
Caleb Malchik and Joan Feigenbaum.
\newblock From {Data} {Leverage} to {Data} {Co}-{Ops}: {An} {Institutional} {Model} for {User} {Control} over {Information} {Access}, January 2022.
\newblock arXiv:2201.10677 [cs].
\bibitem[Nug21]{Nugent_2021}
Annabel Nugent.
\newblock Taylor swift fans share tips on how to make old ‘fearless’ album disappear to mark release of ‘taylor’s version’.
\newblock {\em Independent}, Apr 2021.
\bibitem[Ost90]{ostrom_governing_1990}
Elinor Ostrom.
\newblock {\em Governing the {Commons}: {The} {Evolution} of {Institutions} for {Collective} {Action}}.
\newblock Cambridge University Press, November 1990.
\newblock Google-Books-ID: 4xg6oUobMz4C.
\bibitem[Pay24]{payne2024review}
Will~B Payne.
\newblock Review bombing the platformed city: Contested political speech in online local reviews.
\newblock {\em Big Data \& Society}, 11(3):20539517241275879, 2024.
\bibitem[Qui07]{RePEc:edn:esedps:169}
Tom Quilter.
\newblock {Noise Matters in Heterogeneous Populations}.
\newblock Edinburgh School of Economics Discussion Paper Series 169, Edinburgh School of Economics, University of Edinburgh, August 2007.
\bibitem[SDCW19]{Sanh2019DistilBERTAD}
Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf.
\newblock Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.
\newblock {\em ArXiv}, abs/1910.01108, 2019.
\bibitem[SHMD24]{sigg_decline_2024}
Dorothee Sigg, Moritz Hardt, and Celestine Mendler-Dünner.
\newblock Decline {Now}: {A} {Combinatorial} {Model} for {Algorithmic} {Collective} {Action}, October 2024.
\newblock arXiv:2410.12633 [cs].
\bibitem[VH21]{vincent_can_2021}
Nicholas Vincent and Brent Hecht.
\newblock Can "{Conscious} {Data} {Contribution}" {Help} {Users} to {Exert} "{Data} {Leverage}" {Against} {Technology} {Companies}?
\newblock {\em Proceedings of the ACM on Human-Computer Interaction}, 5(CSCW1):103:1--103:23, April 2021.
\bibitem[VLT{\etalchar{+}}21]{vincent_data_2021}
Nicholas Vincent, Hanlin Li, Nicole Tilly, Stevie Chancellor, and Brent Hecht.
\newblock Data {Leverage}: {A} {Framework} for {Empowering} the {Public} in its {Relationship} with {Technology} {Companies}.
\newblock In {\em Proceedings of the 2021 {ACM} {Conference} on {Fairness}, {Accountability}, and {Transparency}}, {FAccT} '21, pages 215--227, New York, NY, USA, March 2021. Association for Computing Machinery.
\bibitem[WAC21]{wells_just--place_2021}
Katie~J Wells, Kafui Attoh, and Declan Cullen.
\newblock “{Just}-in-{Place}” labor: {Driver} organizing in the {Uber} workplace.
\newblock {\em Environment and Planning A: Economy and Space}, 53(2):315--331, March 2021.
\newblock Publisher: SAGE Publications Ltd.
\bibitem[WDS{\etalchar{+}}20]{wolf2020huggingfaces}
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven~Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander~M. Rush.
\newblock Huggingface's transformers: State-of-the-art natural language processing, 2020.
\bibitem[WED22]{wu2022reasonable}
Yuxi Wu, W~Keith Edwards, and Sauvik Das.
\newblock “a reasonable thing to ask for”: Towards a unified voice in privacy collective action.
\newblock In {\em Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems}, pages 1--17, 2022.
\bibitem[WSLA13]{shirly2013}
Shir~L Wang, Kamran Shafi, Chris Lokan, and Hussein~A Abbass.
\newblock An agent-based model to simulate and analyse behaviour under noisy and deceptive information.
\newblock {\em Adaptive Behavior}, 21(2):96--117, 2013.
\bibitem[WVM21]{woodside_bottom-up_2021}
Jonathan Woodside, Tara Vinodrai, and Markus Moos.
\newblock Bottom-up strategies, platform worker power and local action: {Learning} from ridehailing drivers.
\newblock {\em Local Economy}, 36(4):325--343, June 2021.
\newblock Publisher: SAGE Publications Ltd.
\bibitem[XFS{\etalchar{+}}25]{xiao_it_2025}
Qing Xiao, Xianzhe Fan, Felix~M. Simon, Bingbing Zhang, and Motahhare Eslami.
\newblock "{It} {Might} be {Technically} {Impressive}, {But} {It}'s {Practically} {Useless} to us": {Motivations}, {Practices}, {Challenges}, and {Opportunities} for {Cross}-{Functional} {Collaboration} around {AI} within the {News} {Industry}, February 2025.
\newblock arXiv:2409.12000 [cs].
\bibitem[XZF{\etalchar{+}}25]{xiao_lets_2025}
Qing Xiao, Yuhang Zheng, Xianzhe Fan, Bingbing Zhang, and Zhicong Lu.
\newblock Let's {Influence} {Algorithms} {Together}: {How} {Millions} of {Fans} {Build} {Collective} {Understanding} of {Algorithms} and {Organize} {Coordinated} {Algorithmic} {Actions}, February 2025.
\newblock arXiv:2409.10670 [cs].
\end{thebibliography}