Sublinear-time Collision Detection with a Polynomial Number of States in Population Protocols
Extracted source text · arXiv 2411.09957
\documentclass[11pt,a4paper]{article}
\usepackage[T1]{fontenc}
\usepackage[linesnumbered,ruled,vlined,noresetcount]{algorithm2e}
\DontPrintSemicolon
\SetArgSty{textnormal}
\SetKwProg{Fn}{function}{:}{end}
\SetKwProg{pFor}{for each}{ do in parallel}{end}
\SetKwProg{When}{when}{ do}{end}
% \SetKwInput{VarAgent}{Variables in Agent Memory}
% \SetKwInput{VarNode}{Variables in $v$'s Whiteboard}
\SetKwInput{Variables}{Variables}
\SetKwInput{Notation}{Notation}
\SetKwInput{Note}{Note}
\def\protocol#1{{\normalfont\textsc{#1}}\xspace}
\newcommand{\approximate}{\protocol{Approximate}}
\newcommand{\aspace}{\vspace{10pt}}
\newcommand{\varspace}{\hspace{10pt}}
\usepackage{graphicx}
\usepackage{multirow}
\usepackage{color}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\usepackage{authblk}
\usepackage{geometry}
\geometry{left=2.5cm,right=2.5cm,top=2.5cm,bottom=2.5cm}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{graphicx}
\usepackage{hyperref}
\newtheorem{lemma}{Lemma}
\newtheorem{theorem}{Theorem}
\newtheorem{remark}{Remark}
\newtheorem{proposition}{Proposition}
\theoremstyle{definition}
\newtheorem*{psketch}{Proof Sketch}
\newtheorem*{note}{Note}
\newtheorem{definition}{Definition}
%カリグラフ体
\newcommand{\calp}{\mathcal{P}}
\newcommand{\calm}{\mathcal{M}}
\newcommand{\calw}{\mathcal{W}}
\newcommand{\calg}{\mathcal{G}}
\newcommand{\calr}{\mathcal{R}}
\newcommand{\cald}{\mathcal{D}}
\newcommand{\cale}{\mathcal{E}}
%アルゴリズム名称
\newcommand{\slow}{\protocol{SlowCollisionDetection}}
%\newcommand{\slow}{\protocol{SlowCollisionDetection}}
\newcommand{\cold}{\protocol{CollisionDetection}}
%\newcommand{\cold}{\protocol{CD}}
\newcommand{\coldbfull}{\protocol{CollisionDetection\allowbreak{}WithBounds}}
\newcommand{\coldb}{\protocol{CDWB}}
% \newcommand{\cold}{\protocol{CollisionDetection}}
% \newcommand{\coldb}{\protocol{CollisionDetection\allowbreak{}WithBounds}}
\newcommand{\phaseclock}{\protocol{PhaseClock}}
\newcommand{\epochupdate}{\protocol{EpochUpdate}}
\newcommand{\epidemic}{\protocol{Epidemic}}
%関数（サブルーチン）
\newcommand{\main}{\mathtt{MAIN}}
\newcommand{\reset}{\mathtt{RESET}}
\newcommand{\spread}{\mathtt{SPREAD\_STATE}}
\renewcommand{\check}{\mathtt{CHECK\_COLLISION}}
\newcommand{\up}{\mathtt{EPOCH\_UP}}
\newcommand{\init}{\mathtt{Initialize}}
%変数
\newcommand{\collision}{\mathtt{collision}}
\newcommand{\rank}{\mathtt{rank}}
\newcommand{\col}{\mathtt{collision}}
\newcommand{\epoch}{\mathtt{epoch}}
\newcommand{\timer}{\mathtt{timer}}
%\newcommand{\name}{\mathtt{name}}
\newcommand{\grid}{\mathtt{gid}}
\newcommand{\tick}{\mathtt{tick}}
\newcommand{\children}{\mathtt{infectivity}}
\newcommand{\countfin}{\mathtt{countFin}}
\newcommand{\level}{\mathtt{level}}
\newcommand{\lognum}{\mathtt{logNum}}
\newcommand{\var}{\mathtt{var}}
\newcommand{\leader}{\mathtt{leader}}
%定数
\newcommand{\tr}{\mathit{True}}
\newcommand{\fl}{\mathit{False}}
%\newcommand{\nul}{\mathit{null}}
%呼びかけ側、応答側
\newcommand{\aaa}{a}
\newcommand{\bbb}{b}
%記号
\newcommand{\nlower}{n_{L}}
\newcommand{\nupper}{n_{U}}
%関数（数学）
\newcommand{\inputs}{f_{\mathrm{in}}}
\newcommand{\outputs}{f_{\mathrm{out}}}
\newcommand{\sinit}{s_{\mathrm{init}}}
\newcommand{\tdone}{t_{\mathrm{done}}}
\newcommand{\rs}{\mathbf{\Gamma}}
%状況の集合
\newcommand{\call}{\mathcal{C}_{\mathrm{all}}}
\newcommand{\cinit}{\mathcal{C}_{\mathrm{init}}}
%ラテン文字
\newcommand{\ie}{i.e.}
\newcommand{\eg}{e.g.}
\newcommand{\etal}{et~al.}
%確率表記
\newcommand{\ex}{\mathbf{E}}
%その他
\newcommand{\poly}{\mathit{poly}}
\newcommand{\suchthat}{:}
%特定の値
%\newcommand{\childmax}{\left \lfloor (\log \nlower - \log \log \nlower)/2 \right \rfloor -2}
\newcommand{\childmax}{\left \lfloor \log \nlower - \log \ell \right \rfloor-2}
\bibliographystyle{plainurl}% the mandatory bibstyle
\title{
Sublinear-time Collision Detection with a Polynomial Number of States in Population Protocols
} %TODO Please add
\date{}
%\titlerunning{Dummy short title} %TODO optional, please use if title is longer than one line
%\author{\red{Anonymous Authors}}
\author{Takumi Araya}
\author{Yuichi Sudo}
% % \author[1]{Yuichi Sudo\thanks{Corresponding Author: sudo@hosei.ac.jp}}
% % \author[2]{Masahiro Shibata}
% % \author[3]{Junya Nakamura}
% % \author[4]{Yonghwan Kim}
% % \author[5]{Toshimitsu~Masuzawa}
\affil{Hosei University, Tokyo, Japan}
% \affil[1]{Hosei University, Japan}
% \affil[2]{Kyusyu Institute of Technology, Japan}
% \affil[3]{Toyohashi University of Technology, Japan}
% \affil[4]{Nagoya Institute of Technology, Japan}
% \affil[5]{Osaka University, Japan}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
\maketitle
%TODO mandatory: add short abstract of the document
\begin{abstract}
This paper addresses the collision detection problem in population protocols. The network consists of state machines called agents. At each time step, exactly one pair of agents is chosen uniformly at random to have an interaction, changing the states of the two agents. The collision detection problem involves each agent starting with an input integer between $1$ and
$n$, where $n$ is the number of agents, and requires those agents to determine whether there are any duplicate input values among all agents. Specifically, the goal is for all agents to output false if all input values are distinct, and true otherwise.
In this paper, we present an algorithm that requires a polynomial number of states per agent and solves the collision detection problem with probability one in sub-linear parallel time, both with high probability and in expectation. To the best of our knowledge, this algorithm is the first to solve the collision detection problem using a polynomial number of states within sublinear parallel time, affirmatively answering the question raised by Burman, Chen, Chen, Doty, Nowak, Severson, and Xu [PODC 2021] for the first time.
\end{abstract}
\section{Introduction}
\label{sec:introduction}
In this paper, we explore the \emph{population protocol} model, introduced in 2004 and studied extensively since then \cite{AAD+06,AAE08,AG15,AAG18,GS18,GSU18,SOK+20,SOI+20,BGK20,SEI+21}. The model consists of a network, or \emph{population}, of $n$ state-machines, referred to as \emph{agents}. At each time step, a pair of agents is selected uniformly at random to engage in an \emph{interaction} (\ie, pairwise communication), during which they update their states.
Agents are anonymous, \ie, they lack unique identifiers.
% In the population protocol model, time complexity is often measured in \emph{parallel time}, defined as the number of time steps divided by $n$ (the number of agents).
% In practice, many interactions typically occur simultaneously across a population: parallel time approximately captures that.
% Throughout the remainder of this section (\ie Section \ref{sec:introduction}), we will discuss time complexity in terms of parallel time.
In the population protocol model, time complexity is commonly measured in \emph{parallel time}, defined as the number of time steps divided by the population size $n$ (the number of agents). In practice, interactions often occur simultaneously across the population; parallel time approximately captures this behavior. Throughout the remainder of this section (\ie, Section \ref{sec:introduction}), we assume parallel time when discussing time complexity.
In this paper, we study the collision detection problem in the population protocol model.
Before presenting our main results, we briefly explain prior studies on problems closely related to collision detection, namely self-stabilizing leader election and self-stabilizing ranking.
Leader election has been extensively studied in the population protocol model. Leader election can be solved by a simple two-state protocol \cite{AAD+06}, where initially, all agents are leaders. The protocol employs only one transition rule: when two leaders meet, one of them becomes a follower (\ie, a non-leader). This simple protocol elects a unique leader in $O(n)$ parallel time.
Furthermore, this protocol is time-optimal: Doty and Soloveichik \cite{DS18} demonstrated that any constant-space protocol requires linear time to elect a unique leader. Later, in 2015, Alistarh and Gelashvili \cite{AG15} developed a leader election protocol that converges in $O(\log^3 n)$ parallel time and uses $O(\log^3 n)$ states per agent.
Subsequently, numerous papers have focused on fast leader election, including \cite{AAG18,GS18,GSU18,SOI+20,BGK20}. G{\k{a}}sieniec, Staehowiak, and Uznanski \cite{GSU18} developed an algorithm that converges in $O(\log n \log \log n)$ time and uses a surprisingly small number of states: only $O(\log \log n)$ per agent. This is considered space-optimal because it is established that every leader election protocol requiring $O(n/\mathrm{polylog}(n))$ time also requires $\Omega(\log \log n)$ states \cite{AAE+17}.
Sudo \etal~\cite{SOI+20} presented a simple protocol that elects a unique leader within $O(\log n)$ time and utilizes $O(\log n)$ states per agent. This is time-optimal, as any leader election protocol requires $\Omega(\log n)$ time, even if it uses an arbitrarily large number of states and the agents know the exact size of the population \cite{SM20}. (This lower bound may appear obvious, yet it does not directly follow from a simple coupon collector argument because we can specify an initial configuration where all agents are followers.)
Finally, in 2020, Berenbrink \etal~\cite{BGK20} provided a time and space-optimal protocol, \ie, an $O(\log n)$-time and $O(\log \log n)$-states leader election protocol.
Self-stabilizing leader election (SS-LE) has garnered significant attention within this model. This variant of leader election stipulates that (i) starting from any configuration, the population must reach a safe configuration where exactly one leader exists; and (ii) once a safe configuration is reached, the unique leader must be maintained indefinitely. These conditions ensure tolerance against finitely many transient faults, which is critical since many protocols (both self-stabilizing and non-self-stabilizing) assume the presence of a unique leader. Consequently, SS-LE is essential for enhancing the fault tolerance of the population protocol model itself.
However, no protocol can solve SS-LE unless each agent in the population knows the \emph{exact size} $n$ of the population \cite{FJ06,CIW12}\footnotemark{}.
\footnotetext{
Strictly speaking, the cited works proves a slightly weaker impossibility. Nevertheless, this impossibility can be proved using a similar technique: a simple partitioning argument. See \cite{SOK+20} for details (page 618, footnote).
}
Numerous studies have focused on overcoming this impossibility by employing various strategies, including assuming oracles \cite{BBB13,FJ06,CP07}, assuming that agents precisely know the population size $n$ \cite{CIW12,BCC+21}, restricting the topology \cite{AAF+08,CC19,CC20,YSM21,YSO+23}, or slightly relaxing the requirements of SS-LE~\cite{Izu15,SOK+14,SMD+16,SNY+12,SOK+20det,SOK+18,SOK+20,SSN+21,SEI+21}. Among these, all algorithms that adopt the $n$-knowledge approach \cite{CIW12,BCC+21} elect the unique leader by solving a more general problem, called \emph{self-stabilizing ranking}. This problem stipulates that: (i) each agent $v$ maintains an output variable $v.\rank \in \{1,2,\dots,n\}$,
(ii) starting from any configuration, the population must reach a safe configuration where no two agents share the same $\rank$ value; and (iii) once a safe configuration is reached, no agent updates its $\rank$. SS-LE can be straightforwardly reduced to self-stabilizing ranking because, once ranking is achieved, there is exactly one agent with $\rank=1$. This agent can thus be regarded as the unique leader.
Cai, Izumi, and Wada~\cite{CIW12} present an algorithm that solves the self-stabilizing ranking problem within $O(n^2)$ parallel time, using $n$ states per agent. In contrast, Burman, Chen, Chen, Doty, Nowak, Severson, and Xu~\cite{BCC+21} introduce significantly faster algorithms that, however, require more states. Specifically, they offer an $O(n)$ parallel time algorithm with $O(n)$ states per agent and an $O(\log n)$ parallel time algorithm with a super-exponential number of states. This leads to a natural question: Can self-stabilizing ranking be solved in sublinear parallel time using only a polynomial number of states per agent?
To solve self-stabilizing ranking, agents must detect whether any distinct agents share the same rank. Consequently, Burman \etal\ raise an open question: Is there a sublinear parallel time algorithm with $\mathit{poly}(n)$ states that can solve the following problem, which we refer to as the \emph{collision detection problem} in this paper (See Section \ref{sec:prodef} for the formal definition
of this problem):
\begin{quote}
Each agent $v$ is given an input $\rank \in \{1,2,\dots,n\}$. The goal for each agent is to decide whether at least one pair of agents have the same input value in $\rank$.
\end{quote}
\subsection*{Our Contribution}
In this paper, we affirmatively answer the open question raised by Burman \etal~\cite{BCC+21}. Specifically, we introduce a collision detection algorithm that stabilizes within $O(\sqrt{n \log n})$ parallel time in expectation and $O(n^{1/2} \cdot \log^{3/2} n)$ parallel time with high probability. This algorithm uses $\tilde{O}(n)$ states per agent with high probability, excluding the input variable $\rank$, which requires $O(n)$ states.
The proposed algorithm is always correct; that is, it eventually reaches a stable configuration where all agents output the correct answer with probability 1.
% It remains an open question whether self-stabilizing ranking or weaker variants of this problem, such as loosely-stabilizing ranking, can be solved within sublinear parallel time using a polynomial number of states per agent. This still remains open because our collision detection protocol is not self-stabilizing.
It remains an open question whether self-stabilizing ranking, or weaker variants such as loosely-stabilizing ranking, can be solved within sublinear parallel time using a polynomial number of states per agent. This question persists primarily because our collision detection protocol is not self-stabilizing.
\subsection*{Organization of This Paper}
Section 2 introduces the preliminaries, including key terminologies and the definition of the model. Section 3 describes the basic submodules that are instrumental in designing the proposed algorithm $\cold$. Section 4 presents $\cold$, proves its correctness, and bounds its time and space complexities.
In the remainder of this paper, we will not use parallel time; instead, we will discuss stabilization time in terms of the number of time steps (or interactions).
\section{Preliminaries}
\label{sec:preliminaries}
%\subsection{Notation}
Throughout this paper,
we denote the set of \emph{non-negative} integers by $\mathbb{N}$
and represent the set
$\{z \in \mathbb{N} \mid x \leq z \leq y\}$
by $[x,y]$.
When the base of a logarithm is omitted, it is assumed to be 2.
For any positive integer $i$, we denote the $i$-th projection map by $\pi_i$,
\ie, for any element $\mathbf{x} = (x_1, x_2, \dots, x_k) \in X_1 \times X_2 \times \dots \times X_k$,
we define $\pi_i(\mathbf{x}) = x_i$.
The $\tilde{O}$-notation hides a poly-logarithmic factor; that is, for any function $f(x)$,
$\tilde{O}(f(x))$ denotes $O(f(x) \cdot \log^c x)$ for some constant $c$.
\subsection{Model}
\label{sec:model}
A \emph{population} is a network consisting of {\em agents}.
We denote the set of all the agents by $V$ and let $n = |V|$.
We assume that a population is a complete graph,
thus every pair of agents $(u,v)$ can interact,
where $u$ serves as the \emph{initiator}
and $v$ serves as the \emph{responder} of the interaction.
% Throughout this paper, we use the phrase
% ``with high probability'', abbriviated w.h.p.~to denote
% probability $1-O(n^{-1})$.
A \emph{protocol} $P(Q,\sinit,X,Y,T,\outputs)$ consists of
a set $Q$ of states,
the initial state $\sinit$,
a set $X$ of input symbols,
a set $Y$ of output symbols,
a transition function
$T: (Q\times X)^2 \to Q^2$,
%an input function $\inputs : X \to Q$, and
and an output function $\outputs : Q \times X \to Y$.
% when an execution of protocol $P$ starts,
% every agent $v \in V$ is given an input symbol $x \in X$,
% deciding its initial state $\inputs(x)$.
% Some protocol does not require any input
% and let all agents be in the common initial state $\sinit$ at time step $0$.
% In this case, we suppose that $X=\{\bot\}$
% and $\inputs(\bot)=\sinit$.
Initially, all agents are in state $\sinit \in Q$.
When two agents interact,
$T$ determines their next states
according to their current states and inputs.
The \emph{output} of an agent is determined by $\outputs$:
the output of an agent in state $q \in Q$ and input $x \in X$
is $\outputs(q,x)$.
% The state of each agent can be represented by a set of local variables.
% In this paper, we use the notation $v.\var$ to denote the value of a variable $\var$ maintained by agent $v$,
% and by abuse of notation, for any state $q \in Q$, we use $q.\var$ to denote $v.\var$ when $v$ is in state $q$.
A protocol may be given design parameters, such as a lower bound $\nlower$ and an upper bound $\nupper$ on the population size $n$.
In this case, each component of the protocol, namely $Q$, $\sinit$, $X$, $Y$, $T$, and $\outputs$, may depend on these parameters.
A \emph{configuration} is a mapping $C : V \to Q\times X$ that specifies the states and the inputs
of all agents in the population.
% Given a protocol $P$, the set of all possible configurations for $P$ on the population of size $n$
% is denoted by $\call(P,n)$. We omit $n$ and simply use $\call(P)$ when there is no risk of ambiguity.
We say that a configuration $C$ \emph{changes} to another configuration $C'$ via an interaction $e = (u,v)$,
denoted by $C \stackrel{P,e}{\to} C'$,
if $(\pi_1(C'(u)), \pi_1(C'(v))) = T(C(u), C(v))$, $\pi_2(C'(u))=\pi_2(C(u))$, $\pi_2(C'(v))=\pi_2(C(v))$,
and $C'(w) = C(w)$ for all $w \in V \setminus \{u, v\}$.\footnote{
In this paper, we assume that the inputs provided to the agents do not change.
}
% In this context, $u$ and $v$ are referred to as the \emph{initiator}
% and the \emph{responder} of the interaction $e$, respectively.
We say that a configuration $C'$ is reachable from a configuration $C$
if there is a sequence of configurations $C = C_0, C_1, \dots, C_k$ such that
$C_i$ changes to $C_{i+1}$ via some interaction for each $i \in [0, k-1]$.
A configuration $C$ is \emph{stable} if, for every configuration $C'$ reachable from $C$,
the output of each agent remains the same in both $C$ and $C'$,
\ie, $\outputs(C(v)) = \outputs(C'(v))$.
% Furthermore, we say that an agent $v \in V$
% \emph{participates} in an interaction $e$ if $v$ is either the initiator or the responder of $e$.
%A \emph{schedule} $\gamma = \gamma_0,\gamma_1,\dots =(u_0,v_0),(u_1,v_1),\dots~$
%is a (possibly infinite) sequence of interactions happening at each time step $t$.
%under schedule $\gamma$.
We assume the \emph{uniformly random scheduler} $\rs$, which
selects two agents to interact at each step uniformly at random from all ordered pairs of agents.
Specifically, $\rs=\Gamma_0, \Gamma_1,\dots$
where each $\Gamma_t$ is a random variable such
that $\Pr(\Gamma_t = (u,v)) =\frac{1}{n(n-1)}$
for any $t \ge 0$ and any distinct $u,v \in
V$.
%We use capital letter $\rs$ (e.g., $P$) to represent the random schedule not fixed yet,
%in contrast to using lower cases to represent a realized schedule.
%Given a configuration $C_0 \in \call(P)$,
The \emph{execution} of protocol $P$ starting from a configuration $C_0$
under the uniformly random scheduler $\rs$
is defined as the sequence of configurations $\Xi_{P}(C_0,\rs) = C_0,C_1,\dots$ such that
$C_t \stackrel{P,\Gamma_t}{\to} C_{t+1}$ holds for all $t \ge 0$.
Note that each $C_i$ is also a random variable.
In this paper, we use the notation $v.\var$ to denote the value of a variable $\var$ maintained by an agent $v$. %Each state $q \in Q$ can be regarded as a function that specifies the value of each variable.
By abuse of notation,
we use $C(v).\var$ to represent the value of $\var$ in a configuration $C$.
%when an agent $v$ is in state $q \in Q$.
\subsection{With High Probability}
\label{sec:whp}
In this paper, we frequently use the term ``with high probability,'' abbreviated as w.h.p.
Various definitions for this term exist, such as with probability $1-n^{-1}$ or with probability $1-n^{\Omega(1)}$.
Following \cite{BEF21}, we adopt the below definition.
\begin{definition}[with high probability]
\label{def:whp}
A property of a protocol $P$ holds \emph{with high probability}
if, for any constant $\eta = \Theta(1)$, we can ensure
the property is satisfied with a probability of $1-O(n^{-\eta})$
by adjusting the constant parameters of $P$, if necessary.
\end{definition}
% \begin{definition}[with high probability]
% \label{def:whp}
% An event $A$ in an execution of a protocol $P$ occurs within $O(f(n))$ time steps \emph{with high probability}
% if, for any constant $\eta$, the parameters of $P$ can be adjusted (if necessary) so that
% there exists a constant $c$, ensuring that
% $A$ occurs within $c \cdot f(n)$ time steps with a probability of at least $1-n^{-\eta}$
% for each sufficiently large population size $n$.
% \end{definition}
\noindent
For example, if a protocol $P_{\tau}$ with a design parameter $\tau=\Theta(1)$
elects a leader in $\tau \cdot n \log n = O(n \log n)$ time steps with probability $1-n^{-\tau}$,
we say that $P_{\tau}$ elects a leader in $O(n\log n)$ time steps w.h.p.
In this case, we can also explicitly state the constant factor of the running time and say that
$P_{\tau}$ elects a leader in $\tau \cdot n \log n$ time steps w.h.p.
Note that the requirement for the success probability (i.e., $1-O(n^{-\eta})$) is defined asymptotically;
therefore, we only need to consider sufficiently large population sizes $n$.
% In this case, we can also explicitly write the constant factor of the running time and say that
% $P_{\tau}$ elects a leader in $\tau \cdot n \log n$ time steps w.h.p.
%Note that $\tau \cdot n \log n = O(n \log n)$ because $\tau$ is a constant parameter.
%A property of a protocol $P$ may hold w.h.p.\ even if $P$ does not have a design parameter.
%For example,
% For simplicity of presentation, we also introduce the term ``with very high probability'',
% abbreviated as w.v.h.p., as follows.
% \begin{definition}[with very high probability]
% An event $A$ occurs within $O(x)$ time steps \emph{with very high probability}
% if for any constant $\eta$, there exists a constant $c$ such that
% $A$ occurs within $c \cdot x$ time steps with a probability of at least $1-n^{-\eta}$.
% \end{definition}
\subsection{Collision Detection Problem}
\label{sec:prodef}
% A problem $\mathcal{P}$ specifies (i) a set of \emph{valid} input assignment functions $\mathcal{I}$,
% where each $\iota$ assigns each agent in $V$ an input symbol,
% (ii) a set of feasible output
% In the collision detection problem, each agent $v$ is given an input symbol in $[1,n]$, denoted by $v.\rank$, where $n=|V|$ is the population size.
The collision detection problem requires a protocol $P = (Q, \sinit, X, Y, T, \outputs)$ to accept the input symbol $[1,n]$, \ie, $[1,n] \subseteq X$, where $n = |V|$ is the population size. Each agent $v$ is assigned an input symbol in $[1,n]$, denoted by $v.\rank$.
A configuration $C : V \to Q \times X$ is termed \emph{initialized} if, for every $v \in V$, $v$ is in the initial state $\sinit$ and its input $v.\rank$ is within $[1,n]$ in $C$.
More formally, $C$ is initialized if $\pi_1(C(v)) = \sinit$ and $\pi_2(C(v)) \in [1,n]$ for all $v \in V$.
We define $\cinit(P)$ as the set of these initialized configurations.
The goal of this problem is to detect whether or not a \emph{collision} exists, \ie, $u.\rank = v.\rank$ for some distinct agents $u, v \in V$. Each agent must output either $1$ or $0$, representing ``there is at least one collision among the ranks of the agents'' and ``there is no such collision,'' respectively. For simplicity, we assume every protocol $P$ includes a variable $\collision \in \{0,1\}$, and an agent $v$ outputs $1$ if and only if $v.\collision = 1$.
That is, $\outputs(C(v)) = v.\collision$ for any configuration $C$ and any agent $v \in V$.
\begin{definition}[Collision Detection Problem]
Let $P=(Q,\sinit, X,Y,T,\outputs)$ be a protocol, where $[1,n] \subseteq X$. A protocol $P$ \emph{solves} the collision detection problem if for any initialized configuration $C_0 \in \cinit(P)$, both of the following conditions are satisfied:
\begin{itemize}
\item
%Let $C$ be any stable configuration reachable from $C_0$.
For any stable configuration $C$ reachable from $C_0$,
$C(v).\collision = 1$ for all $v \in V$ if $C_0(u).\rank = C_0(w).\rank$ for some distinct agents $u,w \in V$; otherwise, $C(v).\collision = 0$ holds for all $v \in V$.
\item The execution $\Xi_{P}(C_0,\rs)$ eventually reaches a stable configuration with probability 1.
\end{itemize}
\end{definition}
For any execution $\Xi = C_0, C_1, \dots$, we define the \emph{stabilization time} of $\Xi$
as the minimum $t$ such that $C_t$ is a stable configuration.
We say that a protocol $P$ \emph{stabilizes} within $t$ time steps with high probability (respectively, in expectation)
if, for any initialized configuration $C_0 \in \cinit(P)$,
the stabilization time of $\Xi_{P}(C_0,\rs)$ is at most $t$ with high probability (respectively, in expectation).
\subsection{Uniform Protocols}
\label{sec:uniform}
A protocol is \emph{uniform} if it does not depend on the population size at all,
that is, if it does use any knowledge on the population size.
In the field of population protocols, originally,
the number of states of a protocol $\mathcal{P}=(Q,\rho,Y,\pi,\delta)$ is simply defined as $|Q|$.
Thus, the number of states for uniform protocols must be either $\Theta(1)$ or infinite.
Some problems inherently require memory space of a non-constant size.\footnote{
For example, Doty and Soloveichik \cite{DS18} proved that $\omega(1)$ states are necessary to solve the leader election problem within $o(n^2)$ expected interactions under the uniformly random scheduler.
}
This means that uniform protocols devoted for those problems always require an infinite number of states by definition,
while each execution of these protocols may use only small number of states depending on population size.
(A uniform protocol must be defined independently from the population size, but of course, its execution may depend on
the population size.)
Thus, we require another way to evaluate the number of states for uniform protocols as several works do
\cite{CMN+11,BKR19,DE19,BEF21}. This paper adopts the following simple definition of \emph{the number of states} of
a protocol $P=(Q,\sinit, X,Y,T,\outputs)$. %, introduced by \cite{BKR19}:
%The generalized model, which accommodates both uniform and super-constant protocols, can be described roughly as follows:
\begin{itemize}
\item An agent maintains a constant number of variables $x_1, x_2, \dots, x_s$, and the combination of their values constitutes the state of the agent. The first element $Q$ of protocol $P=(Q,\sinit, X,Y,T,\outputs)$ can be regarded as the set of all such states.
The domain of each variable may not be bounded, thus $Q$ may be an infinite set.
\item For any agent $v \in V$, define $v$'s amount of information at time step $t$ as $f(v,t)=\sum_{i=1}^s \iota_i$, where each $\iota_i$ is the number of bits required to encode the value of variable $x_i$. The number of states of an execution $\Xi$ is defined as $\#(\Xi) = \max\{2^{f(v,t)} \mid v \in V, t=0,1,\dots\}$ in the execution.
We say that the number of states of a protocol $P$
is at most $z$ with high probability (respectively, in expectation) if
for any initialized configuration $C_0 \in \cinit(P)$,
$\#(\Xi_{P}(C_0,\rs)) \le z$ holds with high probability (respectively, in expectation).
% The number of states of a protocol $P$ starting from a configuration $C_0$ is defined as
% $\#(\Xi_{P}(C_0,\rs))$.
% %We evaluate the number of states of $P$ as the maximum
% %$\#(\Xi_{P}(C_0,\rs))$ over $C_0 \in \sinit(P)$, \eg,
% For example,
% we write that a protocol $P$ uses $\tilde{O}(\sqrt{n})$ states with high probability
% starting from any configuration $C_0 \in \sinit(P)$
% if $\#(\Xi_{P}(C_0,\rs))=\tilde{O}(\sqrt{n})$ holds with high probability for any $C_0 \in \sinit(P)$.
% %Therefore, the number of states of $\mathcal{P}$'s execution may depend on the population size $n$, even if $\mathcal{P}$ itself does not depend on $n$.
\end{itemize}
%Since we assume the uniformly random scheduler,
%Refer to Section 2 in \cite{DE19} for the formal definition of the generalized model, which is based on a Turing Machine.
\section{Tools}
\label{sec:tools}
\begin{algorithm}[t]
\caption{$\epidemic(\var)$
at an interaction where
$\aaa$ and $\bbb$ are an initiator and a responder, respectively.
}
\label{al:epidemic}
$\bbb.\var \gets \max(\aaa.\var,\bbb.\var)$
\end{algorithm}
\subsection{One-way Epidemic}
The one-way epidemic protocol was introduced and analyzed by Angluin, Aspnes, and Eisenstat \cite{AAE08} and has been widely used thereafter.
The goal of this protocol is to propagate the maximum value of a given variable $\var$ to all agents.
The strategy is simple: when an initiator $a$ and a responder $b$ meet, and if $a.\var > b.\var$ holds,
$b.\var$ is updated to $a.\var$. (See Algorithm \ref{al:epidemic}.)
Angluin \etal~\cite{AAE08} prove the following useful lemma,
which guarantees that
%the one-way epidemic completes
the maximum value is propagated to the entire population
in $O(n \log n)$ steps w.h.p.
% (Note that Definition \ref{def:whp} does not
% necessarily require the existence and adjustment of a design parameter of a protocol.)
\begin{lemma}
\label{lem:epidemic}
Suppose that
an execution of the protocol $\epidemic(\var)$
starts from a configuration $C$ where $M = \max_{v \in V} v.\var$.
Then, for any fixed $\eta > 0$,
there exists a constant $d$ such that
for sufficiently large population size $n$,
with probability $1-n^{-\eta}$,
every agent $v \in V$ satisfies $v.\var \ge M$
within $d n \log n$ time steps.
\end{lemma}
\begin{algorithm}[t]
\caption{$\phaseclock(F)$
at an interaction where
$\aaa$ and $\bbb$ are an initiator and a responder, respectively.
}
\label{al:phaseclock}
\Variables{$\timer \in [0,m-1], \epoch\in [0,F], \leader \in \{0,1\}$}
\Note{$m$ is a sufficiently large constant. (See Lemma \ref{lem:phase_clock}.)}
\Notation{
$\max_m(x,y)=
\begin{cases}
x& x \in [0,y-\lceil m/2 \rceil] \cup [y+1,y+\lfloor m/2 \rfloor]\\
y& \textbf{otherwise}
\end{cases}$
}
%$\bbb.\var \gets \max(\aaa.\var,\bbb.\var)$
\uIf{$\bbb.\leader = 1 \land \aaa.\timer = \bbb.\timer$}{
$\bbb.\timer \gets (\bbb.\timer + 1) \bmod m$\;
\If{$\bbb.\timer = 0$}{$\bbb.\epoch = \min(\bbb.\epoch+1,F)$\;}
}\ElseIf{
$\bbb.\leader=0$
}{
$\bbb.\timer \gets \max_m(\aaa.\timer,\bbb.\timer)$\;
}
Execute $\epidemic(\epoch)$\;
% \uIf{$\aaa.\timer$ decreased above}{
% $\aaa.\tick \gets 1$\;
% }\Else{
% $\aaa.\tick \gets 0$\;
% }
% $\bbb.\tick\gets
% \begin{cases}
% 1& \bbb.\timer\text{ has decreased above}\\
% 0 & \textbf{otherwise}
% \end{cases}$
\end{algorithm}
\subsection{Phase Clock with a Leader}
Together with the epidemic protocol, Angluin \etal~\cite{AAE08}
introduce the phase clock protocol.
This protocol requires the presence of a unique leader.
Specifically, it assumes that each agent maintains a variable $\leader \in \{0,1\}$,
and there exists an agent $a_L \in V$ such that
$a_L.\leader = 1$ and $b.\leader = 0$ for all $b \in V \setminus \{a_L\}$ consistently.
The goal of this protocol %, as shown in Algorithm \ref{al:phaseclock},
is allowing the leader $a_L$ to determine whether $\Theta(n \log n)$ time steps has passed or not.
In this paper, we modify this protocol slightly so that
all agents are synchronized with a variable $\epoch \in [0,F]$,
where $F$ is a given (possibly non-constant) integer.
In addition to $\epoch$, this protocol maintains a variable $\timer \in [0, m-1]$, where $m$ is a constant parameter.
Initially, $a_L.\timer = 0$ and $v.\timer = m-1$ for all $v \in V \setminus \{a_L\}$,
while the $\epoch$ value of all agents are $0$.
In this protocol, only the responder updates its variables at each interaction.
Lines 2--8 in Algorithm \ref{al:phaseclock} specify how the responder updates $\timer$ and $\epoch$.
The leader $a_L$ increments its $\timer$ by one modulo $m$ when it encounters an initiator
with the same timer value.
The timer of a non-leader $b$ is overwritten when
it encounters an initiator with a $\timer$ value from $b.\timer+1$
to $b.\timer + \lfloor m/2 \rfloor$ modulo $m$,
\ie, a $\timer$ value in $[0, y-\lceil m/2 \rceil] \cup [y+1, y+\lfloor m/2 \rfloor]$,
where $y=b.\timer$.
The leader $a_L$ increments its $\epoch$ by one when it increments its $\timer$ from $m-1$ to $0$
unless $a_L.\epoch$ has already reached the given maximum value $F$.
The increased $\epoch$ value is propagated to all agents via the epidemic protocol.
The following lemma directly follows from Lemma \ref{al:epidemic}
and the analysis of the phase clock protocol presented by Angluin \etal~\cite{AAE08}.
% For completeness, the proof of this lemma is presented in the appendix.
%(See the proof for Theorem 1 in \cite{AAE08}.)
%Angluin \etal~\cite{AAE08} prove that the $\timer$ value of the leader $a_L$
% While we do not explain how the agents maintain the $\timer$ variable,
% it is worthwhile mentioning that during each interaction, only the responder updates its $\timer$ value.
% The update rule for the $\tick$ variable is simple: the responder $b$ of an interaction sets the $\tick$ flag to 1 if $b.\timer$ decreases during the interaction. Otherwise, $b.\tick$ is reset to 0.
% but the typical behavior of this protocol can be described as follows:
% (i) The timer values of all agents are always close in modulo $m$,
%Let $v \in V$ be an agent.
% We divide time steps into \emph{epochs} according to the $\timer$ value of the leader $a_L$.
% Let $\Xi=C_0, C_1,\dots$ be an execution of $\phaseclock$.
% Define $f_{\Xi}:\mathbb{N} \to \mathbb{N}$ as follows:
% (i) $f_\Xi(0) = 0$, (ii) for any $t \ge 1$, $f_\Xi(t) = f_\Xi(t-1)+1$ if $C_t(a_L).\timer = 0$ and $C_{t-1}(a_L).\timer = m-1$,
% \ie, the leader change its $\timer$ value from $m-1$ to $0$ at time step $t-1$;
% otherwise $f_\Xi(t)=f_\Xi(t-1)$. The $i$-th epoch of this execution $\Xi$
% is defined as $\cale(\Xi,i)= \{t \in \mathbb{N} \mid f(t)=i\}$.
% Angluin \etal~\cite{AAE08} prove the following theorem,
% which guarantees the size of each epoch is $\Theta(n \log n)$ w.h.p.
% \begin{theorem}[\cite{AAE08}
% \footnote{
% In the original statement of the theorem in \cite{AAE08}, the constants $c$ and $\eta$ are equal. However, according to their proof,
% the theorem is also valid for cases where $c \neq \eta$, and we have thus modified the statement accordingly.
% }
% ]
% \label{theorem:phaseclock}
% % Let $C_0$ be a configuration where
% % (i) there exists exactly one agent $a_L$ such that $a_L.\leader=1$,
% % (ii) $a_L.\timer = 0$ holds,
% % and (iii) $v.\timer = m-1$ for all $v \in V \setminus\{a_L\}$.
% Let $C_0$ be a configuration with exactly one leader $a_L$, where
% the $\timer$ values of all agents are initialized as specified above.
% Let $\Xi_m = \Xi_{\phaseclock}(C_0,\rs)$,
% where $\phaseclock$ is the protocol specified in Algorithm \ref{al:phaseclock}
% with parameterized $m$.
% For any fixed $c, \eta, d_1 \ge 0$,
% there exists constants $m$ and $d_2$ such that,
% for all sufficiently large population size $n$,
% with probability $1-n^{-\eta}$,
% $d_1 n \log n \le |\cale(\Xi_m,i)| \le d_2 n \log n$ holds for all $i \in [0,n^c]$.
% \end{theorem}
% % the following inequality holds for all $i \in [0,n^c]$:
% % $$d_1 n \log n \le |\cale_\Xi(i)| \le d_2 n \log n.$$
% Let $\Xi=C_0, C_1,\dots$ be an execution of $\phaseclock$.
% We define the \emph{epoch} of an agent $v \in V$ at time step $t \ge 0$ in $\Xi$,
% denoted by $\cale_\Xi(v,t)$, as the count of times the agent raises its tick flag. Specifically,
% (i) $\cale_\Xi(v,0) = 0$, and
% (ii) for any $t \ge 1$, $\cale_\Xi(v,t) = \cale_\Xi(v,t-1) + 1$
% if $C_{t-1}(v).\tick = 0$ and $C_{t}(v).\tick = 1$;
% otherwise, $\cale_\Xi(v,t) = \cale_\Xi(v,t-1)$.
% Angluin \etal~\cite{AAE08} prove the following lemma,
% which guarantees that roughly speaking,
% with high probability,
% all agents simultaneously stay in each of the first polynomial number of epochs
% with $\Theta(n \log n)$ interactions.
% \begin{lemma}
% \label{lem:phaseclock}
% Let $C_0$ be a configuration where $v.\timer = v.\tick = 0$ holds for all $v \in V$
% and there exists exactly one agent $a_L$ such that $a_L.\leader=1$.
% Let $\Xi = \Xi_{\phaseclock}(C_0,\rs) = C_0,C_1,\dots$.
% Then, for any constant $c$ and any constant $\eta$,
% all of the following statements hold w.h.p.
% \begin{itemize}
% \item $|\{t \ge 0 \mid \forall v \in V: \cale_\Xi(v,t)=j \}|=\Omega(n \log n)$ holds for all $j \in [1,n^c]$
% \item There exists $M = \Theta(n \log n)$ such that
% $\cale_\Xi(v,t) < \cale_\Xi(v,t+M)$ holds for all $v \in V$ and all $t \ge 0$.
% \end{itemize}
% \end{lemma}
% Combining $\epidemic$ and $\phaseclock$, we can easily synchronize the population.
% The resulting protocol, $\epochupdate$, is depicted in Algorithm \ref{al:epochupdate},
% where all agents recognize the current epoch index of the execution and store it in a variable $\epoch$.
% Initially, $\epoch$ for all agents is set to zero.
% A leader increases its $\epoch$ value by one each time it resets its timer to zero (from $m-1$),
% unless the epoch has already reached the given maximum value $F$.
% The incremented $\epoch$ value is then propagated to the entire population via $\epidemic(\epoch)$.
% The following corollary directly follows from Lemma \ref{lem:epidemic} and Theorem \ref{theorem:phaseclock}.
% (For completeness, the proof is presented in the appendix.)
\begin{lemma}
\label{lem:phase_clock}
Let $C_0$ be a configuration with exactly one leader $a_L$, where
the variables of all agents are initialized as specified above.
Let $\Xi_m = \Xi_{\phaseclock(F)}(C_0,\rs)$,
where $\phaseclock$ is the protocol specified in Algorithm \ref{al:phaseclock}
with the parameter $m$.
For any fixed constants $c, d_1, \eta \ge 0$,
there exist constants $m$ and $d_2$ such that,
for all sufficiently large population sizes $n$,
the execution $\Xi_m$ satisfies the following properties with probability $1-n^{-\eta}$:
\begin{itemize}
\item For any $i \in [0,\min(n^c, F)]$, there are at least $d_1 \log n$ time steps where $v.\epoch=i$ holds for all agents $v$, and
\item For any $i \in [0,\min(n^c, F)-1]$, each agent $v$ maintains $v.\epoch = i$ for at most $d_2n \log n$ steps.
\end{itemize}
\end{lemma}
By this lemma, as long as $F=\poly(n)$,
we can ensure that with high probability, all agents simultaneously experience each of the epochs $0, 1, \dots, F$ during an arbitrarily large $\Omega(n \log n)$ time steps with a controllable constant factor $d_1$,
while each epoch $0,1,\dots,F-1$ finishes in $O(n \log n)$ time steps.
% The first item guarantees that no epoch finishes ``too early''
% and we can control the lower bound $d \log n$ of the duration of each epoch (by choosing $m$).
% The second item guarantees that no epoch continues ``too long''
% \begin{algorithm}[t]
% \caption{$\slow$ at an interaction where
% $\aaa$ and $\bbb$ are an initiator and a responder, respectively.
% }
% \label{al:slow}
% \If{$\aaa.\rank = \bbb.\rank$}{$\bbb.\collision \gets 1$\;}
% %$\aaa.\collision \gets \max(\aaa.\collision,\bbb.\collision)$
% Execute $\epidemic(\collision)$
% \end{algorithm}
% \paragraph*{Slow Collision Detection}
\section{Collision Detection}
In this section, we present a collision detection protocol $\cold$, which
stabilizes within $O(n^{3/2} \cdot \sqrt{\log n})$ time steps (respectively, $O(n^{3/2} \cdot \log^{3/2} n)$ time steps) in expectation
(respectively, with high probability) and uses $\tilde{O}(n)$ states per agent with high probability, excluding the input variable $\rank$, which requires $O(n)$ states.
This protocol is always correct,
\ie, it eventually reaches a stable configuration where all agents output the correct answer
with probability $1$.
In Section \ref{sec:without}, we present a collision detection protocol under the assumption that (i) there is a pre-elected leader, and (ii) all agents are aware of a common lower bound and a common upper bound on $n$, both of which are asymptotically tight. In Section \ref{sec:with}, we eliminate these assumptions.
\subsection{Collision Detection with a Leader and Rough knowledge of $n$}
\label{sec:without}
In this section, we present a protocol $\coldbfull(\nlower, \nupper)$,
abbreviated as $\coldb(\nlower, \nupper)$.
This protocol requires that there is always a unique leader $a_L \in V$
and that
the two arguments $\nlower$ and $\nupper$ are asymptotically tight lower and upper bounds on $n$.
Specifically, $c_1 n \le \nlower \leq n \leq \nupper \le c_2 n$ must hold
for some constants $c_1$ and $c_2$.
We assume that the agents are aware of these constants $c_1$ and $c_2$, allowing us to incorporate them into the design of this protocol.
\begin{algorithm}[t]
\caption{$\coldb(\nlower,\nupper)$ at an interaction where
$\aaa$ and $\bbb$ are an initiator and a responder, respectively.
}
\label{al:coldb}
% Execute the counting protocol \approximate \cite{BKR19}\\
% \tcp*{This protocol runs the phase clock protocol given by \cite{GS18}}
%\If{$\aaa.\countfin=\bbb.\countfin=1$}{
%Let $N=2^{\aaa.\ell}$\;
%Let $r = \eta \log n/\log \chi(\nupper)$\;
\If{$\aaa.\rank = \bbb.\rank$}{$\bbb.\collision \gets 1$\;}
Let $r = \lceil 3 \eta \log \nupper \rceil$, $\ell = \left \lceil \sqrt{\nlower \log \nlower} \right \rceil$, and $z = \lceil \nupper / \ell \rceil$\;
Execute $\phaseclock(r\cdot z +1)$\;
\If{$\bbb.\epoch$ has increased in $\phaseclock$}{
%$\init(a_i)$
%$\bbb.\epoch = \min(\bbb.\epoch+1,M)$\;
Let $k=(\bbb.\epoch-1)\bmod z$\;
\uIf{$\bbb.\rank \in [k\cdot \ell+1,(k+1)\cdot \ell]$}{
%$\bbb.\name \gets (\text{a number chosen uniformly at random from }[1,\chi(\nupper)])$\;
Choose $\chi$ uniformly at random from $\{0,1\}$\;
$\bbb.\grid \gets (\bbb.\rank - k\cdot \ell,\chi)$\;
$\bbb.\children \gets \childmax$\;
}\Else{
$\bbb.\grid \gets \bot$\;
$\bbb.\children \gets 0$\;
}
}
%}
\If{
$\aaa.\epoch = \bbb.\epoch > 0$
}{
\If{$\aaa.\grid \neq \bot \land \aaa.\children > 0 \land \bbb.\grid = \bot$}{
$\bbb.\grid \gets \aaa.\grid$\;
$\aaa.\children \gets \bbb.\children \gets \aaa.\children-1$
}
\If{$\exists x,y,z: \aaa.\grid=(x,y) \land \bbb.\grid=(x,z) \land y\neq z$}{
$\bbb.\collision \gets 1$\;
}
}
%Execute $\slow$\;
Execute $\epidemic(\collision)$
% \If{$\aaa.\rank = \bbb.\rank$}{$\aaa.\collision \gets 1$\;}
% $\aaa.\collision \gets \bbb.\collision \gets \max(\aaa.\collision,\bbb.\collision)$
\end{algorithm}
The pseudocode of this protocol is presented in Algorithm \ref{al:coldb}.
The algorithm consists of three parts.
The first part, referred to as \emph{the backup protocol}, detects collisions through direct interactions (Lines 9--10):
when two agents with the same rank meet, the responder $\bbb$ raises a collision flag,
\ie, $\bbb.\collision$ is set to $1$.
If only one pair of agents $u, v \in V$ shares a common rank,
it can be easily observed that this process requires $\binom{n}{2} = \Theta(n^2)$ time steps in expectation.
The second part detects collisions much more quickly, requiring $O(n^{3/2} \sqrt{\log n})$ time steps in expectation
and $O(n^{3/2} \log^{3/2} n)$ time steps with high probability, using a square root decomposition technique (Lines 11-27).
Both the first and second parts allow \emph{some} agent to raise a collision flag if a collision is detected.
The third part propagates the raised flag to the entire population via the epidemic protocol (Lines 28).
Although the backup protocol is much slower,
it is still necessary because the second part may fail to detect a collision with non-zero probability.
Moreover, the backup protocol assists the second part in quickly detecting collisions.
%Specifically, if there are too many agents whose rank are in a common segment $S_k$,
Specifically, we present the following lemma:
\begin{lemma}
\label{lem:not_so_many_colliding_pairs}
Let $C_0$ be a configuration where there are $x = \Omega(\sqrt{n\log n})$ pairs of colliding agents,
\ie, $|\{(u,v) \in V^2 \mid u \neq v, u.\rank = v.\rank\}| = \Omega(\sqrt{n\log n})$ holds in $C_0$.
Then, in an execution of the backup protocol starting from $C_0$,
at least one agent raises a collision flag within $O(n^{3/2})$ time steps, both in expectation and with high probability.
\end{lemma}
\begin{proof}
At each time step, one of the $x$ colliding pairs is selected to interact
with a probability of $x / \binom{n}{2} = \Omega(\log n / n^{3/2})$, resulting in the responder of the pair raising a collision flag.
Therefore, a collision flag is raised within $O(n^{3/2} / \log n)$ steps in expectation,
and within $O(n^{3/2})$ steps with high probability.
\end{proof}
\noindent
This lemma implies that, regardless of the performance of the second part,
%$\coldb(\nlower, \nupper)$
the proposed protocol solves the collision detection problem within $O(n^{3/2})$ time steps, both in expectation and with high probability, if the number of colliding pairs is $\Omega(\sqrt{n \log n})$.
Therefore, for the remainder of this section, we will assume that the number of colliding pairs is $o(\sqrt{n \log n})$.
\begin{figure}[t]
\centering
\includegraphics[width=0.85\textwidth]{segments.pdf}
\caption{Segments of the second part of $\coldb(\nlower,\nupper)$}
\label{fig:segments}
\end{figure}
\begin{table}
\centering
\caption{Notations used in the second part of $\coldb(\nlower,\nupper)$}
\label{table:notations}
\begin{tabular}{c c c}
\hline
Notation & Definition & Explanation \\
\hline
$\ell$ & $\lceil \sqrt{\nlower \log \nlower} \rceil$ & Length of a segment \\
$z$ & $\lceil \nupper/\ell \rceil$ & Number of segments \\
$S_k$ & $[k\cdot \ell + 1, (k+1)\cdot \ell] \cap [1, \nupper]$ & $k$-th segment
($0 \le k < z$)\\
$r$ & $\lceil 3 \eta \cdot \log \nupper \rceil$ & Number of epochs dedicated to each segment \\
\hline
\end{tabular}
\end{table}
We will now explain how the agents behave in the second part, which is the main focus of the proposed protocol.
Utilizing the unique leader assumption, the agents selected for each interaction
execute the phase clock protocol presented in Section \ref{sec:tools} (Line 12).
Let $\ell=\left \lceil \sqrt{\nlower \log \nlower} \right \rceil$
and $z = \lceil \nupper/\ell \rceil$.
In addition, let $r = \lceil 3 \eta \cdot \log \nupper \rceil$, where $\eta$ is a hidden design parameter ensuring that the targeted time complexity, \ie, $O(n^{3/2} \cdot \log^{3/2} n)$ time steps, are achieved with probability $1-O(n^{-\eta})$.
The argument $F$ of the phase clock protocol is set to
$r \cdot z + 1$. By Lemma \ref{lem:phase_clock}, it is guaranteed with high probability that
for each $i = 1, 2, \dots, r\cdot z$,
all agents are simultaneously in epoch $i$
for at least $d_1 n \log n$ steps with any sufficiently large constant $d_1$,
and all agents enter the final epoch $r \cdot \ell + 1$ within
$\Theta(r z n \log n)=\Theta(n^{3/2} \log^{3/2} n)$ steps.
We divide the set of integers $[1,\nupper]$ into $z$ segments with at most size $\ell$,
\ie, for any $k=0,1,2,\dots,z-1$,
the $k$-th segment is $S_k = [k\cdot \ell+1, (k+1)\cdot \ell] \cap [1,\nupper]$
(Figure \ref{fig:segments}).
We assign $r$ epochs $k-1, k-1 + z, k-1 + 2z, \dots, k-1 + (r-1)z$ to
detect collisions of ranks in $S_k$.
Conversely, each epoch $i \in [1, r\cdot z]$ is dedicated to collision detection in $S_{(i-1) \bmod z}$.
% We omit ``$\bmod z$'' when it is clear from the context, \eg, we write $S_{i-1}$ instead of $S_{(i-1) \bmod z}$,
% in the rest of this section.
The notations introduced here are summarized in Table \ref{table:notations}.
%In what follows, we explain how the agents detect the collision in $S_{\lceil i/r \rceil}$ in each epoch $i \in [1,r\cdot z]$.
The goal of epoch $i \in [1,r\cdot z]$ is to detect whether or not there exists $x \in S_{(i-1) \bmod z}$ such that
$|\{v \in V \mid v.\rank = x\}| \ge 2$,
in $O(n\log n)$ time steps with probability at least $1/3$ for a sufficiently large $n$, while
ensuring that no agent mistakenly detect a collision if no such $x$ exists.
By the achievement of this goal, all $r$ epochs devoted to each segment results in
the correct decision with probability $1-(1/3)^r \ge 1-n^{-\eta}$.
To achieve this goal, each agent $v \in V$ maintains a variable $v.\grid \in ([1,\ell]\times \{0,1\}) \cup \{\bot\}$ and $v.\children \in [0,\childmax]$.
Suppose that an agent $v$ enters epoch $i$ now.
If $v.\rank \in S_{k}$, where $k=(i-1) \bmod z$, it generates a group identifier
$(v.\rank-k\cdot\ell, \chi)$, where $\chi$ is a random number (or \emph{nonce})
chosen uniformly at random in $[0,\lceil {\nupper}^{\eta/r}\rceil-1]$,
and stores it in $v.\grid$ (Lines 15--18). In addition, $v.\children$ is reset to the maximum value $\childmax$.
Otherwise, $v.\grid$ and $v.\children$ are set to $\bot$ and $0$, respectively
(Lines 19--21).
We say that an agent with $\grid=\bot$ a \emph{null} agent.
A non-null agent with $\children = x$ has ability to make $2^x$ copies of non-null agent:
whenever a non-null agent $a$ with $a.\children > 0$ and a null agent $b$
have an interaction as the initiator and the responder, respectively,
$a.\grid$ is copied to $b.\grid$ and both $a.\children$ and $b.\children$
are set to $a.\children - 1$ (Lines 23--25).
If two agents have an interaction at which
their group identifier share a common rank,
but have different nonce, they notice the existence of a collision.
Then, the responder of the interaction raises a collision flag (Line 27).
%$\pi_1(u.\grid)=\pi_1(v,\grid)$
% This proliferation of the group identifier is similar to the epidemic protocol \cite{AAE08}.
% However, unlike the epidemic protocol, this process bounds the maximum number of non-null agents.
% As mentioned later, this limitation contributes to quick proliferation of all group identifiers equally so that
% they complete in $O(n \log n)$ time steps with high probability.
% Consider the case that there is a colliding pair $u,v \in V$ such that $u.\rank = v.\rank \in S_k$,
% $u$ and $v$ will generate different nonce with probability $1/2$, thus their group identifiers will be different
% with probability $1/2$.
The proliferation of group identifiers resembles the epidemic protocol \cite{AAE08}. However, unlike the epidemic protocol, this process limits the maximum number of non-null agents. This constraint
enhances the speed of proliferation, ensuring that proliferation completes in every epoch with high probability. This is formalized in the following lemma:
\begin{lemma}
\label{lem:proliferation}
Let $P = \coldb(\nlower, \nupper)$,
$i \in [1, r \cdot z]$, and $k = (i-1) \bmod z$.
Let $C_0 \in \cinit(P)$ be any initialized configuration
where there are at most $2 \cdot \ell$ agents whose rank is in $S_k$,
and let $\Xi_{P}(C_0, \rs) = C_0, C_1, \dots$.
Let $t_b$ be the minimum integer such that all agents are in epoch $i$ in the $t_b$-th configuration $C_{t_b}$. Then, with high probability,
there exists a non-negative integer $\tdone = t_b + O(n \log n)$ such that
for all $v \in V$, $v.\epoch = i$ and $v.\children = 0$ in $C_{\tdone}$.
\end{lemma}
\begin{proof}
Let $t_e$ be the maximum integer such that all agents are in epoch $i$ in configuration $C_{t_e}$,
and let $V_k$ be the set of agents whose rank is in $S_k$ in configuration $C_0$.
Since $|V_k| \leq 2\ell$, during epoch $i$,
\ie, in configurations $C_{t_b}, C_{t_b+1}, \dots, C_{t_e}$,
there are always at most $2 \ell \cdot 2^{\childmax} \leq n / 2$ non-null agents,
thus there are at least $n / 2$ null agents.
We will prove the lemma using the concept of \emph{virtual agents} introduced by \cite{SNY+12} with slight modifications.
Let $U(t)$ be the set of null agents at time step $t$. From the above discussion,
$U(t) \geq n/2$ holds for all $t \in [t_b, t_e]$.
Define $\iota(v) = C_{t_b}(v).\children$. %for any $v \in U(t_b)$.
If proliferation completes correctly in epoch $i$,
each $v \in V\setminus U(t_b)$ must have generated $2^{\iota(v)}$ non-null agents, including itself.
In this proof, we map those agents to distinct integers from $0$ to $2^{\iota(v)}-1$.
For any $v \in V \setminus U(t_b)$, $\alpha \ge 0$, and $j \ge 1$, we denote
the $j$-th bit from the end of $\alpha$ by $b(v,\alpha,j)$, \ie, $b(v,\alpha,j)=\lfloor \alpha / 2^{j-1} \rfloor \bmod 2$.
By definition, $b(v,\alpha,j) = 0$ for any $j \ge \iota(v)+1$.
Then, for any $t \geq t_b$, we define the \emph{virtual agent} $\nu(v,\alpha,t)$ and its number of divisions $\kappa(v,\alpha,t)$ as follows:
$$ \nu(v,\alpha,t_b)=v, \quad \kappa(v,\alpha,t_b)=0,$$
and for any $t \ge t_b$,
\begin{align*}
\nu(v,\alpha,t+1)
&=
\begin{cases}
u & \text{if } b(v,\alpha,\kappa(v,\alpha,t)+1)=1 \land \exists u \in U(t): \Gamma_{t} = (\nu(v,\alpha,t),u)\\
\nu(v,\alpha,t) & \text{otherwise},
\end{cases}
\\
\kappa(v,\alpha,t+1)
&=
\begin{cases}
\min(\kappa(v,\alpha,t)+1,\iota(v)) & \text{if } \exists u \in U(t): \Gamma_{t} =(\nu(v,\alpha,t),u)\\
\kappa(v,\alpha,t) & \text{otherwise}.
\end{cases}
\end{align*}
\noindent
Clearly, proliferation completes before time step $t$ if $\kappa(v,\alpha,t) = \iota(v)$ for all $v \in V \setminus U(t_b)$ and $\alpha \in [0, 2^{\iota(v)}-1]$. Note that a virtual agent $\nu(v,\alpha,t)$ is determined solely by $C_{t_b}$ and $\Gamma_{t_b}, \Gamma_{t_b+1}, \dots, \Gamma_{t-1}$. Therefore, at any time step $t \in [t_b, t_e]$, the virtual agent $\nu(v,\alpha,t)$ interacts with a non-null agent as the initiator with a probability of $\frac{1}{n} \cdot \frac{U(t)}{n-1} \geq \frac{1}{2(n-1)}$. Consequently, by the Chernoff Bound, $\kappa(v,\alpha,t) = \iota(v)$ holds with high probability for some $\tdone = t_b + O(n \log n)$. By applying the union bound, with high probability, this condition is met for all $v \in V \setminus U(t_b)$ and $\alpha \in [0, 2^{\iota(v)}-1]$, indicating that the proliferation completes within $O(n \log n)$ steps with high probability. Note that we can disregard the probability that $t_e < \tdone$, as we can choose arbitrarily large constant $d_1$ in Lemma \ref{lem:phase_clock}.
\end{proof}
\begin{lemma}
\label{lem:coldb}
Let $P=\coldb(\nlower,\nupper)$.
%Let $C_0 \in \cinit(P)$ be any initialized configuration.
Starting from any initialized configuration $C_0 \in \cinit(P)$,
execution $\Xi_{P}(C_0,\rs)$ solves the collision detection.
It stabilizes within
$O(n^{3/2} \cdot \sqrt{\log n})$ time steps (respectively, $O(n^{3/2} \cdot \log^{3/2} n)$ time steps) in expectation
(respectively, with high probability).
This protocol uses $\tilde{O}(n)$ states per agent.
%the number of agents whose rank is in $S_k$ is at most $2\cdot \ell$.
\end{lemma}
\begin{proof}
We can easily verify that the number of states is $\tilde{O}(n)$ by checking the domain of the variables, excluding the input variable $\rank$: $\grid$ uses $O(\sqrt{n})$ states and $\epoch$ uses $O(r\cdot z)=\tilde{O}(\sqrt{n})$ states, while all other variables use only a poly-logarithmic number of states.
Proving the correctness is also easy. The first part of the protocol (\ie, the backup protocol)
eventually makes an agent raise a flag if and only if there is at least one colliding pair of agents.
The raised flag is propagated to the entire population via the epidemic protocol.
False-positive errors are not permitted also in the second part of $\coldb$.
In each epoch $i \in [1, r \cdot z]$, the group identifiers are initialized. When no agents share a common rank, all non-null agents $v$ with $\pi_1(v.\grid) = \alpha$ originate from a common agent with a rank $k \cdot \ell + \alpha$, where $k = (i-1) \bmod z$. Therefore, all such agents have the same nonce, thus preventing any of them from raising a collision flag.
Thus, starting from any initialized configuration $C_0 \in \cinit(P)$,
execution $\Xi_{P}(C_0,\rs)$ eventually reaches a stable configuration with probability $1$,
where all agents output the correct answer.
Next, we bound the stabilization time of $P$. From the above discussion, the stabilization time is zero when there is no rank collision in $C_0$. %Therefore, without loss of generality,
Thus, assume the presence of a colliding pair $u, v \in V$, \ie, $u.\rank = v.\rank$, in $C_0$.
Let $k \in \{0,1,\dots,z-1\}$ be the integer such that $C_0(u).\rank \in S_k$.
By Lemma \ref{lem:not_so_many_colliding_pairs}, we assume that there are only $o(\sqrt{n\log n})$ colliding pairs of agents without loss of generality.
Under this assumption, $|V_k| \le 2\ell$,
where $V_k$ is the set of agents whose ranks are in $S_k$ in configuration $C_0$.
Otherwise, there are $\sum_{j=1}^{\ell} \binom{x_j}{2} \ge \sum_{j=1}^{\ell} (x_j-1)>\ell = \Omega(\sqrt{n\log n})$ colliding pairs, yielding a contradiction,
where $x_j$ is the number of agents whose rank is $k\cdot \ell+j$.
%Let $i \in [1,r\cdot z]$ be any integer such that $k = (i-1) \bmod z$.
With probability $1/2$, $u$ and $v$ generate different nonces and thus obtain distinct group identifiers when they enter epoch $i$.
By Lemma \ref{lem:proliferation}, with high probability,
each of their group identifiers is copied to $2^{\childmax}\ge \nlower/(8\ell)$ %$\Omega(\sqrt{n/\log n})$
agents, respectively, in the first half period of epoch $i$, at least $\frac{d_1}{2} n \log n$ time steps, where
$d_1$ is the constant that appears in Lemma \ref{lem:phase_clock}.
Thereafter, during the next $\frac{d_1}{2} \cdot n \log n$ time steps, there is at least one time step at which
one agent in $u$'s group and one agent in $v$'s group have an interaction
with probability at least
$$1-\left(1- \frac{\binom{\nlower/8\ell}{2}}{\binom{n}{2}}\right)^{\frac{d_1}{2} \cdot n \log n}
= 1 - \left (1-\Omega \left ( \frac{1}{n \log n}\right ) \right)^{\frac{d_1}{2} \cdot n \log n} \ge 1-\epsilon$$
for any constant $\epsilon$ because we can choose an arbitrarily large constant for $d_1$.
To conclude, some agent raises a collision flag in epoch $i$ with probability $1/2 -\epsilon \ge 1/3$.
There are $r$ such epochs dedicated to $S_k$, and those epochs appear in every $z$ epochs.
Thus, $\Xi(C_0,\rs)$ stabilizes within $O(r\cdot z \cdot n \log n)= O(n^{3/2} \log^{3/2} n)$ steps with probability
at least $1-(1-1/3)^{r} \ge 1-2^{-r} \ge 1-n^{-\eta}$, \ie, with high probability.
%Therefore, by Lemma \ref{lem:proliferation},
%$\Xi_{P}(C_0,\rs)$
% Therefore, the number of non-null agents in epoch $i$ is bounded by $2\ell \cdot 2^{\childmax} \le \frac{\nlower}{2}$.
The expected stabilization time is bounded by
$$\sum_{j=1}^{r} \left (1-\frac{1}{3}\right )^{j-1} \cdot \frac{j z}{3} \cdot O(n \log n)
+ \left(1-\frac{1}{3} \right ) ^r \cdot O(n^2) = O(z n \log n)=O\left (n^{\frac{3}{2}} \sqrt{\log n} \right ).$$
\end{proof}
% Since $v.\epoch$ is monotonically non-decreasing for any $v \in V$, the following proposition holds:
\begin{remark}[No false-positive error even without the assumptions]
\label{remark:coldb}
The epoch $v.\epoch$ of each agent $v \in V$ is monotonically non-decreasing.
Therefore, no false-positive errors occur in the execution of $\coldb(\nlower, \nupper)$ even without the previously stated assumption that there is always a unique leader $a_L$ and the parameters satisfy $\nlower \leq n \leq \nupper$, $\nlower = \Theta(n)$, and $\nupper = \Theta(n)$.
\end{remark}
\begin{remark}[Derandomization]
\label{remark:derandomization}
One might consider $\coldb$ a randomized algorithm because it generates a binary nonce at Line 16, seemingly deviating from the model defined in Section \ref{sec:model}. However, we can easily derandomize our protocol by exploiting the randomness from the uniformly random scheduler $\rs$. The derandomization can proceed as follows:
(i) When an agent increases its epoch, it enters the \emph{waiting mode} instead of initializing $\grid$ and $\children$;
(ii) When an agent in the waiting mode has an interaction,
the agent can generate a 1-bit nonce $\chi$, setting $\chi = 0$ if it is the initiator of the interaction, and $\chi = 1$ otherwise.,
thus it initializes $\grid$ and $\children$ with the generated nonce,
and returns to the \emph{normal mode}.
Let $u$ and $v$ be any two agents with the same rank. Note that the nonce generated by $u$ and $v$ may not be independent, which could potentially complicate the analysis. However, their nonces are dependent only if they have previously met, at which point they would have raised a collision flag via the backup protocol. Therefore, this dependency does not pose a problem.
\end{remark}
% As mentioned later, this limitation of proliferation contributes to guarantee that
% with high probability,
% every agent with a rank in $S_k$
% forms a group with the same identifier whose size is at least $2^{\childmax}\ge \frac{\nlower}{8\ell}=\Omega(\sqrt{\nlower/\log \nlower})$.
% As mentioned previously, by Lemma \ref{lem:not_so_many_colliding_pairs}, we assume that there are only $o(\sqrt{n\log n})$ colliding pairs
% of agents without loss of generality.
% Under this assumption, there are at most $2\ell$ agents whose rank is in $S_{k}$;
% Otherwise, there are $\sum_{j=1}^{\ell} \binom{x_j}{2} \ge \sum_{j=1}^{\ell} (x_j-1)>\ell = \Omega(\sqrt{n\log n})$ colliding pairs,
% where $x_j$ is the number of agents whose rank is $k\cdot \ell+j$, a contradiction.
% Therefore, the number of non-null agents in epoch $i$ is bounded by $2\ell \cdot 2^{\childmax} \le \frac{\nlower}{2}$.
% Specifically, if the number of agents with a rank in $S_{i-1}$ is at most $2\ell$,
% the number of non-null agents would be at most
% $2\ell \cdot 2^{\childmax} \le \frac{\ell}{2} \cdot \sqrt{\nlower/\log \nlower} \le \frac{\nlower}{2}$.
% Thus, at least half of the agents are null, ensuring the proliferation completes in $O(n \log n)$ time steps
% with high probability (See Lemma \ref{lem:proliferation}).
% Otherwise, the number of \emph{colliding pairs} $u$ and $v$ such that $u.\rank = v.\rank \in S_{i-1}$
% is at least $n$, which ensures that some agent detects a collision in $O(n \log n)$ steps.
%there are $\Theta(n \log n)$ steps with any large hidden constant $d_1$
%where all agents $v$ satisfy $v.\epoch = i$,
\begin{algorithm}[t]
\caption{%Collision Detection Algorithm
%The transition rule of
$\cold$ at an interaction where
$\aaa$ and $\bbb$ are an initiator and a responder, respectively.
%an initiator $\aaa$ and a responder $\bbb$ meet
}
\label{al:cold}
Execute the counting protocol \approximate \cite{BKR19}\;
%\tcp*{This protocol runs the phase clock protocol given by \cite{GS18}}
%\For{$i \in \{0,1\}$}{
\If(\tcp*[f]{$\aaa$ never updates its $\level$}){$\bbb.\level$ increased at this interaction}{
Reset all variables of $\bbb$ maintained by $\coldb$ and its submodules
%, including $\collision$,
to their initial values.
}
%}
\If{$
(\aaa.\level, \aaa.\lognum)=(\bbb.\level, \bbb.\lognum)
\land
\aaa.\countfin = \bbb.\countfin = 1
% \aaa.\level = \bbb.\level
% \land
% \aaa.\lognum = \bbb.\lognum
% \land
% \aaa.\countfin=\bbb.\countfin=1
$
}
{
Execute $\coldb(r,2^{\aaa.\lognum-1},2^{\aaa.\lognum+1})$
}
%Execute $\slow$\;
\If{$\aaa.\rank = \bbb.\rank$}{$\bbb.\collision \gets 1$\;}
Execute $\epidemic(\collision)$\;
% $\aaa.\collision \gets \bbb.\collision \gets \max(\aaa.\collision,\bbb.\collision)$
%}
\end{algorithm}
\subsection{Collision Detection without Assumptions}
\label{sec:with}
To run $\coldb(\nlower,\nupper)$, we require a unique leader $a_L \in V$ and two parameters, $\nlower$ and $\nupper$, that satisfy $\nlower \leq n \leq \nupper$, $\nlower = \Theta(n)$, and $\nupper = \Theta(n)$. We address the collision detection problem without this assumption
by running this protocol and the counting protocol $\approximate$, given by Berenbrink, Kaaser, and Radzik~\cite{BKR19}, in parallel. The protocol $\approximate$ elects a leader using the leader election protocol presented by G{\k{a}}sieniec and Stachowiak~\cite{GS18}, and computes $\lfloor \log n \rfloor$ or $\lceil \log n \rceil$ with high probability.
This protocol uses $O(\log n \cdot \log \log n)$ states with high probability.
%The guarantees provided by $\approximate$ are complex, and we explain these in detail.
The protocol $\approximate$ maintains four variables: $\lognum \in \mathbb{N}$, $\level \in \mathbb{N}$, $\countfin \in \{0,1\}$, and $\leader \in \{0,1\}$~\footnote{
We have changed the names of the variables for consistency with the notations of this paper.
}.
The following proposition holds by the definition of $\approximate$, which enables us to integrate these two protocols. (See \cite{BKR19} for the definition of $\approximate$.)
%In this paper, we use $\approximate$ as the black box, but
\begin{proposition} %[\cite{BKR19}]
\label{prop:approximate}
An execution of the protocol $\approximate$ has the following properties:
\begin{itemize}
\item For all $v \in V$, $v.\level$ is monotonically non-decreasing
and increases only when $v$ joins an interaction as the responder,
%$v.\countfin$ is monotonically non-decreasing while $v.\level$ remains the same,
\item For all $v \in V$, $v.\lognum$ never changes while $v.\countfin = 1$,
\item For all $v \in V$,
$v.\countfin$ reverts from $1$ to $0$ only if $v.\level$ increases, and
\item When two agents with the same level meet, the $\countfin$ value is updated according to the one-way epidemic from the initiator to the responder. Only an agent with $\leader = 1$ can set $\countfin$ to $1$, in addition to this rule,
\end{itemize}
\end{proposition}
For any variable $\var$, we say that a configuration $C$ is $\var$-stable if, for every configuration $C'$ reachable from $C$ and every agent $v \in V$, the value of $\var$ for $v$ is the same in both $C$ and $C'$, \ie, $C(v).\var = C'(v).\var$.
For any set of variables $\chi$, we say that a configuration $C$ is $\chi$-stable if $C$ is $\var$-stable for every $\var$ in $\chi$.
Due to the analysis in \cite{BKR19}, we have the following lemma:
\begin{lemma}[\cite{BKR19}]
\label{lem:approximate}
Let $C_0 \in \cinit(P)$ be any initialized configuration and let $\Xi_{P}(C_0, \rs) = C_0, C_1, \dots$,
where $P=\approximate$.
With high probability, there is an integer $t = O(n\log^2 n)$ such that:
\begin{itemize}
\item $C_{t}$ is a $\{\lognum, \leader, \level\}$-stable configuration,
\item All agents share the same value in $\lognum$, which is either $\lfloor \log n \rfloor$ or $\lceil \log n \rceil$, in $C_t$,
\item There exists exactly one agent $a_L$ with $\leader = 1$ in $C_t$,
\item All agents satisfy $\countfin = 0$ in $C_{t-1}$,
\item Exactly two agents satisfy $\countfin = 1$, one of whom is $a_L$, in $C_{t}$.
\end{itemize}
The protocol $\approximate$ uses $O(\log n \cdot \log \log n)$ states with high probability.
\end{lemma}
% If agents can determine whether they have already reached a configuration $C_t$, integrating $\coldb$ and $\approximate$ becomes straightforward: when two agents $\aaa$ and $\bbb$ meet, they execute $\approximate$ if either $\aaa.\countfin = 0$ or $\bbb.\countfin = 0$; otherwise, they execute $\coldb(\nlower, \nupper)$, setting $\nlower = 2^{\aaa.\lognum - 1}$ and $\nupper = 2^{\aaa.\lognum + 1}$. However, such straightforward integration is not feasible as agents cannot detect this condition. In particular, the value of the variable $\countfin$ in $\approximate$ may revert from $1$ back to $0$.
% Fortunately, this issue can be addressed by the following rule:
% each time an agent's level increases, it terminates the current execution of $\coldb$ and restarts it.
% We call the resulting protocol $\cold$.
If agents could determine whether they have already reached a configuration $C_t$, integrating $\coldb$ and $\approximate$ would be straightforward: when two agents $\aaa$ and $\bbb$ meet, they execute $\approximate$ if either $\aaa.\countfin = 0$ or $\bbb.\countfin = 0$; otherwise, they execute $\coldb(\nlower,\nupper)$, setting $\nlower = 2^{\aaa.\lognum - 1}$ and $\nupper = 2^{\aaa.\lognum + 1}$. However, such straightforward integration is not feasible, as agents cannot detect this condition. In particular, the value of the variable $\countfin$ in $\approximate$ may revert from $1$ back to $0$.
% Fortunately, this issue can be resolved, roughly speaking, by the following modification: each time an agent's $\level$ increases, it terminates the current execution of $\coldb$ and resets its variables maintained by $\coldb$ to their initial values.
% We call the resulting protocol $\cold$.
%(See the arXiv version \cite{arxiv} for the details.)
%Nonetheless, The properties of Proposition \ref{prop:approximate} provide the means to integrate these two protocols.
The pseudocode of $\cold$ is presented in Algorithm \ref{al:cold}, which utilizes the properties outlined in Proposition \ref{prop:approximate} and \ref{lem:approximate} to integrate $\coldb$ and $\approximate$.
Suppose that two agents $\aaa$ and $\bbb$ have an interaction where they serve as the initiator and the responder, respectively. Initially, they execute $\approximate$ (Line 29). As specified in Proposition \ref{prop:approximate}, $\aaa.\level$ does not increase during $\approximate$. If $\bbb.\level$ does increase, all variables maintained by $\coldb$ for $\bbb$ are reset to their initial values (Lines 30, 31). In essence, each time an agent's level increases, it terminates the current execution of $\coldb$ and restarts it.
Subsequently, $\aaa$ and $\bbb$ execute $\coldb(\nlower, \nupper)$ with $\nlower = 2^{\aaa.\lognum - 1}$ and $\nupper = 2^{\aaa.\lognum + 1}$, but only if they have matching values in both $\level$ and $\lognum$, and both observe $\countfin = 1$. According to Lemma \ref{lem:approximate}, these conditions will eventually be met with high probability.
%allowing the agents to accurately compute either $\lfloor \log n \rfloor$ or $\lceil \log n \rceil$.
However, there remains a small but non-zero probability that $\approximate$ fails to meet these conditions, which would prevent $\coldb$ from executing. Thus, we also implement a backup protocol outside of $\coldb$ (Lines 34–36), eliminating the risk of false-negative errors.
% We have the following theorem. See the arXiv version \cite{arxiv} for the details of $\cold$,
% including its pseudocode, and the proof of Theorem \ref{theorem:main}.
\begin{theorem}
\label{theorem:main}
Protocol $\cold$ solves the collision detection problem with probability $1$.
The execution of $\cold$ under the uniformly random scheduler
stabilizes within $O(n^{3/2} \log^{3/2} n)$ time steps and uses $\tilde{O}(n)$ states per agent (excluding the space needed to store the input variable $\rank$) with high probability. It stabilizes within $O(n^{3/2} \sqrt{\log n})$ time steps in expectation.
%The stabilization time of $\coldb$ is $O(n^{3/2})$ time steps
\end{theorem}
\begin{proof}
The protocol $\cold$ is always correct. Specifically, the execution of $\cold$ eventually stabilizes to correctly output:
$1$ if any agents share the same rank, and $0$ otherwise.
This correctness immediately follows from the fact that (i) a backup protocol (Lines 34--36) eventually computes the correct answer, and (ii) $\coldb$ (Line 33) does not cause false-positive errors (see Remark \ref{remark:coldb}).
Next, we analyze the stabilization time. Let $C_0 \in \cinit(P)$ be any initialized configuration, and let $\Xi = \Xi_{P}(C_0, \rs) = C_0, C_1, \dots$, where $P = \cold$. According to Lemma \ref{lem:approximate}, with high probability, there exists $t = O(n \log^2 n)$ that meets the conditions specified by Lemma \ref{lem:approximate}. Since $C_t$ is $\{\lognum, \leader, \level\}$-stable, in the suffix $C_t, C_{t+1}, C_{t+2}, \dots$ of the execution, no agent changes the values of $\lognum$, $\leader$, or $\level$. Therefore, the assumptions made for $\coldb$ are always satisfied in this suffix. At interaction $\Gamma_{t-1}$, which transitions $C_{t-1}$ to $C_t$, the unique leader $a_L$ and one non-leader agent, say $u \in V$, raise the $\countfin$ flag and execute $\coldb$ for the first time since their level reaches $C_t(a_L).\level$. Thereafter, as stated in Proposition \ref{prop:approximate}, other agents also raise the $\countfin$ flag according to the one-way epidemic and never revert the flag from $1$ to $0$.
Moreover, in $C_t$, all agents except $a_L$ and $u$ retain the initial values for all variables maintained by $\coldb$, while during interaction $\Gamma_{t-1}$, $a_L$ and $u$ simulate one interaction of $\coldb$ with those initialized variables. Therefore, the suffix $C_t, C_{t+1}, C_{t+2}, \dots$ exactly corresponds to the execution of $\coldb$ because in an execution of $\coldb$, no agent changes its state before the unique leader $a_L$ has its first interaction. Specifically, an agent $v$ updates its states by interaction $\Gamma_{t'}$ only if there is an increasing sequence of integers $t_1, t_2, \dots, t_{k-1}$ and a sequence of agents $v_1, v_2, \dots, v_{k}$ such that:
(i) $v_1 = v_L$, (ii) $v_k = v$, (iii) $t_{k-1} \le t'$,
and (iv) $\Gamma_{t_i} = (v_i, v_{i+1})$ for each $i \in [1, k-1]$.
Therefore, by Lemmas \ref{lem:coldb} and \ref{lem:approximate}, $\cold$ stabilizes within $O(n^{3/2} \log^{3/2} n)$ time steps with high probability and $O(n^{3/2} \sqrt{\log n})$ time steps in expectation.
The claim for the number of states immediately follows from Lemmas \ref{lem:coldb} and \ref{lem:approximate}.
\end{proof}
\paragraph*{Acknowledgments}
This work is supported by JST FOREST Program JPMJFR226U and
JSPS KAKENHI Grant Numbers JP19K11826, JP20H04140, and JP20KK0232.
%%
%% The next two lines define the bibliography style to be used, and
%% the bibliography file.
% \bibliographystyle{ACM-Reference-Format}
\bibliography{population}
%%
%% If your work has an appendix, this is the place to put it.
%\appendix
% \section{Pseudocode of \texorpdfstring{$\rootopt$}{RootOptimal}}
% \label{sec:pseudocode_root_optimal}
% The pseudocode of $\rootopt$ is shown in Algorithm \ref{al:root_optimal}.
\end{document}
\begin{thebibliography}{10}
\bibitem{AAE+17}
Dan Alistarh, James Aspnes, David Eisenstat, Rati Gelashvili, and Ronald~L Rivest.
\newblock Time-space trade-offs in population protocols.
\newblock In {\em Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms}, pages 2560--2579. SIAM, 2017.
\bibitem{AAG18}
Dan Alistarh, James Aspnes, and Rati Gelashvili.
\newblock Space-optimal majority in population protocols.
\newblock In {\em Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms}, pages 2221--2239. SIAM, 2018.
\bibitem{AG15}
Dan Alistarh and Rati Gelashvili.
\newblock Polylogarithmic-time leader election in population protocols.
\newblock In {\em Proceedings of the 42nd International Colloquium on Automata, Languages, and Programming}, pages 479--491, 2015.
\bibitem{AAF+08}
Dana. Angluin, J.~Aspnes, Michael.~J Fischer, and Hong. Jiang.
\newblock Self-stabilizing population protocols.
\newblock {\em ACM Transactions on Autonomous and Adaptive Systems}, 3(4):1--28, 2008.
\bibitem{AAD+06}
Dana Angluin, James Aspnes, Zo\"{e} Diamadi, Michael~J. Fischer, and Ren\'{e} Peralta.
\newblock Computation in networks of passively mobile finite-state sensors.
\newblock {\em Distributed Computing}, 18(4):235--253, 2006.
\bibitem{AAE08}
Dana Angluin, James Aspnes, and David Eisenstat.
\newblock Fast computation by population protocols with a leader.
\newblock {\em Distributed Computing}, 21(3):183--199, 2008.
\bibitem{BBB13}
Joffroy Beauquier, Peva Blanchard, and Janna Burman.
\newblock Self-stabilizing leader election in population protocols over arbitrary communication graphs.
\newblock In {\em International Conference on Principles of Distributed Systems}, pages 38--52, 2013.
\bibitem{BEF21}
Petra Berenbrink, Robert Els{\"a}sser, Tom Friedetzky, Dominik Kaaser, Peter Kling, and Tomasz Radzik.
\newblock Time-space trade-offs in population protocols for the majority problem.
\newblock {\em Distributed Computing}, 34:91--111, 2021.
\bibitem{BGK20}
Petra Berenbrink, George Giakkoupis, and Peter Kling.
\newblock Optimal time and space leader election in population protocols.
\newblock In {\em Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing}, pages 119--129, 2020.
\bibitem{BKR19}
Petra Berenbrink, Dominik Kaaser, and Tomasz Radzik.
\newblock On counting the population size.
\newblock In {\em Proceedings of the 2019 ACM Symposium on Principles of Distributed Computing}, pages 43--52, 2019.
\bibitem{BCC+21}
Janna Burman, Ho-Lin Chen, Hsueh-Ping Chen, David Doty, Thomas Nowak, Eric Severson, and Chuan Xu.
\newblock Time-optimal self-stabilizing leader election in population protocols.
\newblock In {\em Proceedings of the 2021 ACM Symposium on Principles of Distributed Computing}, pages 33--44, 2021.
\bibitem{CIW12}
Shukai Cai, Taisuke Izumi, and Koichi Wada.
\newblock How to prove impossibility under global fairness: On space complexity of self-stabilizing leader election on a population protocol model.
\newblock {\em Theory of Computing Systems}, 50(3):433--445, 2012.
\bibitem{CP07}
Davide Canepa and Maria~G. Potop-Butucaru.
\newblock Stabilizing leader election in population protocols.
\newblock {\em http://hal.inria.fr/inria-00166632}, 2007.
\bibitem{CMN+11}
Ioannis Chatzigiannakis, Othon Michail, Stavros Nikolaou, Andreas Pavlogiannis, and Paul~G Spirakis.
\newblock Passively mobile communicating machines that use restricted space.
\newblock In {\em Proceedings of the 7th ACM ACM SIGACT/SIGMOBILE International Workshop on Foundations of Mobile Computing}, pages 6--15, 2011.
\bibitem{CC19}
Hsueh-Ping Chen and Ho-Lin Chen.
\newblock Self-stabilizing leader election.
\newblock In {\em Proceedings of the 38th ACM Symposium on Principles of Distributed Computing}, pages 53--59, 2019.
\bibitem{CC20}
Hsueh-Ping Chen and Ho-Lin Chen.
\newblock Self-stabilizing leader election in regular graphs.
\newblock In {\em Proceedings of the 39th ACM Symposium on Principles of Distributed Computing}, pages 210--217, 2020.
\bibitem{DE19}
David Doty and Mahsa Eftekhari.
\newblock Efficient size estimation and impossibility of termination in uniform dense population protocols.
\newblock In {\em Proceedings of the 2019 ACM Symposium on Principles of Distributed Computing}, page 34–42, 2019.
\bibitem{DS18}
David Doty and David Soloveichik.
\newblock Stable leader election in population protocols requires linear time.
\newblock {\em Distributed Computing}, 31(4):257--271, 2018.
\bibitem{FJ06}
Michael~J. Fischer and Hong Jiang.
\newblock Self-stabilizing leader election in networks of finite-state anonymous agents.
\newblock In {\em International Conference on Principles of Distributed Systems}, pages 395--409, 2006.
\bibitem{GS18}
Leszek G{\k{a}}sieniec and Grzegorz Stachowiak.
\newblock Fast space optimal leader election in population protocols.
\newblock In {\em Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms}, pages 2653--2667. SIAM, 2018.
\bibitem{GSU18}
Leszek G{\k{a}}sieniec, Grzegorz Stachowiak, and Przemyslaw Uznanski.
\newblock Almost logarithmic-time space optimal leader election in population protocols.
\newblock In {\em The 31st ACM on Symposium on Parallelism in Algorithms and Architectures}, pages 93--102. ACM, 2019.
\bibitem{Izu15}
Taisuke Izumi.
\newblock On space and time complexity of loosely-stabilizing leader election.
\newblock In {\em International Colloquium on Structural Information and Communication Complexity}, pages 299--312, 2015.
\bibitem{SEI+21}
Yuichi Sudo, Ryota Eguchi, Taisuke Izumi, and Toshimitsu Masuzawa.
\newblock Time-optimal loosely-stabilizing leader election in population protocols.
\newblock In {\em 35nd International Symposium on Distributed Computing (DISC 2021)}, pages 40:1--40:17, 2021.
\bibitem{SM20}
Yuichi Sudo and Toshimitsu Masuzawa.
\newblock Leader election requires logarithmic time in population protocols.
\newblock {\em Parallel Processing Letters}, 30(01):2050005, 2020.
\bibitem{SMD+16}
Yuichi Sudo, Toshimitsu Masuzawa, Ajoy~K Datta, and Lawrence~L Larmore.
\newblock The same speed timer in population protocols.
\newblock In {\em the 36th IEEE International Conference on Distributed Computing Systems}, pages 252--261, 2016.
\bibitem{SNY+12}
Yuichi Sudo, Junya Nakamura, Yukiko Yamauchi, Fukuhito Ooshita, Hirotsugu. Kakugawa, and Toshimitsu Masuzawa.
\newblock Loosely-stabilizing leader election in a population protocol model.
\newblock {\em Theoretical Computer Science}, 444:100--112, 2012.
\bibitem{SOI+20}
Yuichi Sudo, Fukuhito Ooshita, Taisuke Izumi, Hirotsugu Kakugawa, and Toshimitsu Masuzawa.
\newblock Time-optimal leader election in population protocols.
\newblock {\em IEEE Transactions on Parallel and Distributed Systems}, 31(11):2620--2632, 2020.
\bibitem{SOK+14}
Yuichi Sudo, Fukuhito Ooshita, Hirotsugu Kakugawa, and Toshimitsu Masuzawa.
\newblock Loosely-stabilizing leader election on arbitrary graphs in population protocols.
\newblock In {\em International Conference on Principles of Distributed Systems}, pages 339--354, 2014.
\bibitem{SOK+20det}
Yuichi Sudo, Fukuhito Ooshita, Hirotsugu Kakugawa, and Toshimitsu Masuzawa.
\newblock Loosely stabilizing leader election on arbitrary graphs in population protocols without identifiers or random numbers.
\newblock {\em IEICE Transactions on Information and Systems}, 103(3):489--499, 2020.
\bibitem{SOK+18}
Yuichi Sudo, Fukuhito Ooshita, Hirotsugu Kakugawa, Toshimitsu Masuzawa, Ajoy~K Datta, and Lawrence~L Larmore.
\newblock Loosely-stabilizing leader election for arbitrary graphs in population protocol model.
\newblock {\em IEEE Transactions on Parallel and Distributed Systems}, 30(6):1359--1373, 2018.
\bibitem{SOK+20}
Yuichi Sudo, Fukuhito Ooshita, Hirotsugu Kakugawa, Toshimitsu Masuzawa, Ajoy~K Datta, and Lawrence~L Larmore.
\newblock Loosely-stabilizing leader election with polylogarithmic convergence time.
\newblock {\em Theoretical Computer Science}, 806:617--631, 2020.
\bibitem{SSN+21}
Yuichi Sudo, Masahiro Shibata, Junya Nakamura, Yonghwan Kim, and Toshimitsu Masuzawa.
\newblock Self-stabilizing population protocols with global knowledge.
\newblock {\em IEEE Transactions on Parallel and Distributed Systems}, 32(12):3011--3023, 2021.
\bibitem{YSM21}
Daisuke Yokota, Yuichi Sudo, and Toshimitsu Masuzawa.
\newblock Time-optimal self-stabilizing leader election on rings in population protocols.
\newblock {\em IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences}, 104(12):1675--1684, 2021.
\bibitem{YSO+23}
Daisuke Yokota, Yuichi Sudo, Fukuhito Ooshita, and Toshimitsu Masuzawa.
\newblock A near time-optimal population protocol for self-stabilizing leader election on rings with a poly-logarithmic number of states.
\newblock In {\em Proceedings of the 2023 ACM Symposium on Principles of Distributed Computing}, pages 2--12, 2023.
\end{thebibliography}