Exploiting the Superposition Property of Wireless Communication for Max-Consensus Problems in Multi-Agent Systems
Extracted source text · arXiv 1805.06724
\documentclass[12pt]{article}
\usepackage{natbib}
\usepackage[latin1]{inputenc}
\usepackage{amsthm}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amstext}
%\usepackage{amssymb}
\usepackage{algorithm}
\usepackage[noend]{algpseudocode}
\usepackage{graphicx}
\usepackage{natbib}
%\usepackage{cleveref}
%\usepackage{caption}
\usepackage{subcaption}
%\usepackage{epstopdf}
%\usepackage{epsfig}
%\usepackage{mathtools}
\usepackage[dvipsnames]{xcolor}
\usepackage{soul}
%%% ONLY FOR REVIEWING WITH HORG
\usepackage{setspace}
% Yellow highlight
\newcommand{\hlc}[2][yellow]{{%
\colorlet{foo}{#1}%
\sethlcolor{foo}\hl{#2}}%
}
% Green highlight
\newcommand{\hlcg}[2][GreenYellow]{{%
\colorlet{foo}{#1}%
\sethlcolor{foo}\hl{#2}}%
}
% Red highlight
\newcommand{\hlcr}[2][RedOrange]{{%
\colorlet{foo}{#1}%
\sethlcolor{foo}\hl{#2}}%
}
\newtheorem{definition}{Definition}
\newtheorem{proposition}{Proposition}
\newtheorem{corollary}{Corollary}
%\newtheorem{lemma}{Lemma}
\newtheorem{remark}{Remark}
%\newtheorem{assumption}{Assumption}
\newtheorem{example}{Example}[section]
%\DeclarePairedDelimiter{\norm}{\lVert}{\rVert}
\newcommand{\sgn}{\text{sgn}}
\newcommand\numberthis{\addtocounter{equation}{1}\tag{\theequation}}
\begin{document}
\title{Exploiting the Superposition Property of Wireless Communication for Max-Consensus Problems in Multi-Agent Systems}
\author{
Fabio Molinari\thanks{F. Molinari is with the Control Systems Group - Technische Universit\"at Berlin, Germany.}
\and S\l awomir Sta\'nczak\thanks{S. Sta\'nczak is with the Network Information Theory Group - Technische Universit\"at Berlin, Germany \& Fraunhofer Heinrich Hertz Institute, Germany.}
\and J\"org Raisch\thanks{J. Raisch is with the Control Systems Group - Technische Universit\"at Berlin, Germany \& Max-Planck-Institut f\"ur Dynamik komplexer technischer Systeme, Germany.}
}
\maketitle
\begin{abstract}
\footnote{This work was funded by the German Research Foundation (DFG) within their priority programme SPP 1914 "Cyber-Physical
Networking (CPN)".
}
This paper presents a consensus protocol that achieves max-consensus in multi-agent systems over wireless channels.
Interference, a feature of the wireless channel,
is exploited:
each agent receives a superposition of broadcast data, rather than individual values.
With this information, the system endowed with the proposed consensus protocol
reaches max-consensus in a finite number of steps.
A comparison with traditional approaches shows that the proposed consensus protocol achieves a faster convergence.
\end{abstract}
%===============================================================================
\section{Introduction}
%\doublespacing
Consensus is a useful concept
in cases where several agents need to achieve agreement over a variable of common interest, \cite{ren2007information}.
Each agent retains a local guess of this variable, which is referred to as its \textit{information state}.
The ingredients of a consensus strategy are, first, an exchange of information between agents (\textit{communication}),
then, an update of information states according to a suitable algorithm (\textit{computation}).
This paper focuses on a specific consensus called max-consensus (\cite{olfati2004consensus}).
%In this context, any max-consensus protocol drives all agents to an agreement,
%In this context, agreement is achieved when information states are .
%whose value will be the maximum initial information state of the system.
% In literature, among all examples, this problem has been analyzed for the minimum-time rendezvous problems \cite{lin2005multi,lin2007multi} or
% for consensus-based decentralized auctions, in the context of robust task allocation \cite{choi2009consensus} or of automation of road intersections \cite{molinari2018automation}.
The convergence of max-consensus algorithms has been studied, e.g., via max-plus algebra techniques
in \cite{nejad2009max}, showing that an agreement is achieved
in a finite number of steps.
Convergence rate is an important property.
Many applications, as \cite{molinari2018automation}, show that getting to an agreement as fast as possible is a main issue,
also for safety reasons.
%In average-consensus problems, many strategies were adopted. For example, in
%\cite{mirali2017distributed}, the communication topology is dynamically adapted for increasing the convergence rate.
%;
%on the other hand, \cite{olfati2005ultrafast} achieved considerably faster agreement by randomly rewiring existing communication links between agents.
Traditionally, communication and computation have been treated as two distinct aspects.
However, \cite{goldenbaum2013harnessing} claim that, in a wireless communication framework, a considerable increase of convergence rate can be achieved by merging these two aspects.
Following this path, \cite{molinari2018exploiting} propose an average-consensus protocol which harnesses the interference of the wireless channel,
thus achieving a much faster agreement.
In the proposed scheme, all agents broadcast their values simultaneously, instead of creating orthogonal channels.
Via an appropriate consensus protocol,
the received interfered signals can lead to an agreement.
The broadcast property of the wireless channel has been harnessed for achieving
max-consensus via randomized protocols in \cite{iutzeler2012analysis}.
To the best of our knowledge, no deterministic broadcast protocols for max-consensus have been proposed so far.
% Authors propose a randomized broadcast approach, which is claimed to achieve max-consensus in a finite number of steps, with probability one.
% At the best of our knowledge, broadcast max-consensus algorithms which deterministically achieve consensus in a finite number of steps have never been presented.
In the remainder of the paper, a general consensus problem over wireless channel is described in Section \ref{sec:problemdesc}.
%and a possible usage of the interference is analyzed.
A broadcast max-consensus protocol is presented in Section \ref{sec:asympt} and proven to converge asymptotically to an agreement.
Achieving consensus in a finite number of steps is possible by using the consensus algorithm presented in Section \ref{section:finite}.
Simulations are then analyzed in Section \ref{sec:simul}. Finally, in Section \ref{sec:concl}, concluding remarks are provided.
% This last
% Concerned of speed
% Also in CPS
% Therefore communication merge comput
% An example for average
% Maximum not simple nomographic
% Max paper but randomized
\subsection{Notation}
{We use $\mathbb{N}$, $\mathbb{R}$, $\mathbb{R}_{>0}$, and $\mathbb{R}_{\geq0}$ to denote the set of positive integers, the set of real numbers, the set of positive real numbers, and the set of nonnegative real numbers, respectively. }
%The set of natural numbers is denoted by $\mathbb{N}$, while the set of real numbers is denoted by $\mathbb{R}$. The sets of positive and nonnegative real numbers are, respectively, $\mathbb{R}_{>0}$ and $\mathbb{R}_{\geq0}$.
The vector $\mathbf{1}_n\in\mathbb{R}^n$ denotes the all-ones vector of dimension $n\in\mathbb{N}$.
The transpose of a matrix $A\in\mathbb{R}^{n\times m}$ is denoted by $A'$.
%while the element of matrix $A$ at position $(i-j)$ is $[A]_{ij}$.
%Given a vector $\mathbf{v}$ its infinity norm is denoted by $||\mathbf{v}||_{\infty}$.
The cardinality of a {finite} set $\mathcal{S}$ is {denoted} by $|\mathcal{S}|$.
The indicator function $I_A(x):X\rightarrow\{0,1\}$ on a set $X$ is defined to be
$$
%X\rightarrow \{0,1\} :\
I_A(x):=
\begin{cases}
1\ \ \ \text{if } x\in A,\\
0\ \ \ \text{else,}
\end{cases}
\ A\subseteq X.
$$
{An undirected} graph is a pair $(\mathcal{N},\mathcal{A})$, where $\mathcal{N}$ is a set of nodes, while $\mathcal{A}$ is the corresponding set of arcs.
If an arc connects nodes $i,j\in\mathcal{N}$, then $(i,j)\in\mathcal{A}$.
Given a graph, we define a \textit{path} to be a sequence of nodes, in which each adjacent pair is connected by an arc.
A graph is said to be \textit{connected} if there exists a path between any distinct pair of nodes.
Given a node $i\in\mathcal{N}$, its set of neighbors is denoted by $N_i$ and {we have}
$$N_i=\{ j\in\mathcal{N}\mid (i,j)\in\mathcal{A} \} .$$
% Graph
\section{Problem Description}
\label{sec:problemdesc}
{We consider a discrete-time multi-agent system with $n\geq1$ agents communicating over a wireless network modeled by the graph
$(\mathcal{N},\mathcal{A})$ where $\mathcal{N}=\{1,\dots,n\}$. In what follows, let $k\in\mathbb{N}$ be the time index.}
%A multi-agent system is composed of a set of agents, $\mathcal{N}=\{1,\dots,n\}$, communicating over a wireless network modeled by the graph $\left(\mathcal{N},\mathcal{A}\right)$.
%Given any agent $i\in\mathcal{N}$, we denote by $N_i$ the set of its neighbors.
%Throughout the paper, the following assumption will be hold.
%\begin{assumption}
% The undirected graph $\left(\mathcal{N},\mathcal{A}\right)$ is connected.
%\end{assumption}
{The system has a variable of common interest which the agents have to agree on.
Each agent $i\in\mathcal{N}$ has its initial estimate of this variable}, i.e., $x_{i_0}\in\mathbb{R}_{\geq0}$, which we refer to as its initial \textit{information state}.
%In a discrete-time framework where $k\in\mathbb{N}$ is the time index,
The objective of the consensus protocol
is to enable all agents to reach an agreement over their information states.
To this end,
each agent, say agent $i$, will dynamically update its information state, i.e. $x_i:\mathbb{N}\rightarrow\mathbb{R}_{\geq0}$, according to the consensus protocol and to the information received from its neighbouring agents. Let $\forall i\in\mathcal{N}$, $\mathbf{x}_{N_i}:\mathbb{N}\rightarrow\mathbb{R}_{\geq0}^{|N_i|+1}$ be a vector containing the information states of agents in the set $N_i\cup\{i\}$ at instant $k\in\mathbb{N}$, i.e.
\begin{equation}
\mathbf{x}_{N_i}(k)=[x_i(k), x_{j1}(k),\dots,x_{j_{m_i}}(k)]',
\end{equation}
where $j_1,\dots,j_{m_{i}}\in N_i$ and $m_i=|N_i|$.
%$\mathbf{x}_{N_i}(k)=[x_{i_1},\dots,x_{i_n}]'$, where $i_1,\dots,i_n\in N_i$.
%$\mathbf{x}_{N_i}(k)=\{ x_j(k) \mid j\in N_i \}$.
% Questo è un insieme, e non un vettore. Come esprimere un vettore in questo caso?
Accordingly, a general discrete-time consensus protocol is
\begin{equation}
\label{eq:generalConsProt}
x_i(k+1)=f_i(\mathbf{x}_{N_i}(k)),
\end{equation}
where $f_i:\mathbb{R}_{\geq0}^{|N_i|+1}\rightarrow\mathbb{R}_{\geq0}$ %, $k\in\mathbb{N}$ is the discrete-time index,
and, $\forall i\in\mathcal{N},\ x_i(1)=x_{i_0}$.
Let $\mathbf{x}(k)$ be the vector of all information states, i.e. $\forall i\in\mathcal{N},\ \left[\mathbf{x}(k)\right]_i=x_i(k)$.
Then, we say that the multi-agent system converges to max-consensus if
\begin{equation}
\forall i\in\mathcal{N},\ \lim\limits_{k\rightarrow\infty} x_i(k) = x^* = \max(\mathbf{x}(0)).
\end{equation}
Moreover, if the system converges to max-consensus in a finite number of steps, then {the strategy is referred to as}
\textit{finite-time max-consensus}. This is formally achieved if $\exists \bar{k}\in\mathbb{N}$, such that $\forall k>\bar{k}$,
\begin{equation}
\forall i\in\mathcal{N},\ x_i({k})= x^* = \max(\mathbf{x}(0)).
\end{equation}
%The following definition will be used later in this paper.
\begin{definition}
Given $k\in\mathbb{N}$, an agent $i\in\mathcal{N}$ {is said to be maximal at $k\in\mathbb{N}$ if} $x_i(k)=\max(\mathbf{x}(k))$.
\end{definition}
\subsection{Traditional Max-Consensus Protocols}
{Widely-considered max-consensus protocols are of the form}
%Traditionally, a widely used consensus protocol is
% nelle sources anche il mio paper che usa max-consensus
\begin{equation}
\label{eq:maxcons}
x_i(k+1)=\max(\mathbf{x}_{N_i}(k)).
\end{equation}
\cite{nejad2009max}, under the assumption of a time-invariant and connected network topology, show that the consensus protocol (\ref{eq:maxcons}) ensures max-consensus in a finite-number of steps.
The wireless communication is usually based on orthogonal channel access methods, which establish interference-free transmission links between neighbours and provide each agent with the knowledge of individual information states of the neighbors.
However, {by the data processing inequality} \cite[p. 32]{cover2012elements}, the amount of information contained in $\max(\mathbf{x}_{N_i}(k))$ is in general less than the amount carried by the vector $\mathbf{x}_{N_i}(k)$ itself. Therefore, reconstructing neighbors' information states appears a suboptimal strategy if the goal is to reconstruct only the maximum value of the information states.
%Moreover, as claimed in \cite{molinari2018exploiting,goldenbaum2012nomographic}, channel access methods are bad performing in terms of delays and in scenarios with scarce wireless resources.
\subsection{Interference model}
Let $N_i=\{{j_1},\dots,{j_{m_i}}\}\in\mathcal{N}$ be a set of agents {that} transmit their respective information states to agent $i\in\mathcal{N}$ at discrete-time step $k\in\mathbb{N}$.
If orthogonal channel access methods are not used, then interference occurs.
Accordingly, agent $i\in\mathcal{N}$ receives the signal $\zeta_i(k)$ which is a superposition of $x_{j_1}(k),\dots,x_{j_{m_i}}(k)$.
This is often modeled by an affine model of the wireless multiple access channel (MAC), as in \cite{molinari2018exploiting}, i.e.
\begin{equation}
\label{eq:MAC}
\zeta_i(k)
=
\sum_{j\in N_i} h_{ij}(k)x_j(k) + v_i(k),
\end{equation}
where
$\forall k\in\mathbb{N},\ h_{ij}(k)\in(0,1]\subset\mathbb{R}$
are referred to as channel coefficients, and $\forall i\in\mathcal{N},\ \forall k\in\mathbb{N},\ v_i(k)$ is the receiver noise.
In the following, we ignore both channel coefficients and receiver noise, as in \cite{goldenbaum2012nomographic}; accordingly, (\ref{eq:MAC}) reduces to an ideal MAC
\begin{equation}
\label{eq:MAC_ideal}
\zeta_i(k)
=
\sum_{j\in N_i} x_j(k).
\end{equation}
{This shows that the nature of interference is \textit{superposition}}.
\subsection{Nomographic Representation}
By the \textit{superposition theorem} of \cite{kolmogorov1963representation}, any multivariate function $f:\mathbb{R}^n\rightarrow\mathbb{R}$ has a \textit{nomographic representation}
\begin{equation}
\label{eq:nomRepr}
f_i(x_1,\dots,x_n)=\psi_i(\sum_{j=1}^n \phi_j(x_j)),
\end{equation}
where the univariate functions $\psi_i:\mathbb{R}\rightarrow\mathbb{R}$ and $\phi_j:\mathbb{R}\rightarrow\mathbb{R}$, $j=1\dots n$, are referred to as post-processing function and pre-processing functions, respectively.
%Moreover, pre-processing functions can be chosen to be independent of $f$ \cite{buck1976approximate}.
By (\ref{eq:nomRepr}), and according to the interference model in (\ref{eq:MAC_ideal}), each function can be computed over the wireless channel by harnessing its interference property.
In the context of a consensus problem,
if each agent $i\in\mathcal{N}$ has to compute (\ref{eq:generalConsProt}),
%for reaching an agreement
a procedure, which aims to merge communication and computation over the channel, is presented in Algorithm \ref{algo:nomo_algo} and has to be run $\forall k\in\mathbb{N}$.
\begin{algorithm}
\caption{}
\label{algo:nomo_algo}
\begin{algorithmic}
\State $\forall j\in\mathcal{N}$, agent $j$ broadcasts the pre-processed value of its current information state, $\phi_j(x_j(k))$;
\State $\forall i\in\mathcal{N}$, agent $i$ receives the superposed signals from neighbors, $z_i(k)=\sum_{j\in N_i} \phi_j(x_j(k))$;
\State $\forall i\in\mathcal{N}$, agent $i$ computes $\psi_i(\phi_i(x_i(k))+z_i(k))$, thus getting the desired $f_i(\mathbf{x}_{N_i}(k))$.
\end{algorithmic}
\end{algorithm}
Computing a function over the wireless channel, by using its nomographic representation and the interference, results in a much faster and more efficient solution. A quantitative analysis can be found in \cite{goldenbaum2013harnessing}.
\subsection{{Approximated Nomographic Representation}}
\label{subsec:approx_nom_repr}
However, \cite{buck1982nomographic} states that, in general,
functions do not have a continuous real-valued nomographic representation.
%This can lead to the impossibility of using the interference with the proposed method.
In \cite{limmer2015simple} the idea of a nomographic approximation is suggested.
However, as in the suggested nomographic approximations for the max-function the error is always positive,
it accumulates over time in a consensus algorithm and will therefore lead to divergence.
%However, two possible solutions which do not make use of approximation are presented in the next sections.
In the following, we suggest how to circumvent this problem, i.e., how to use the superposition principle without the need of a nomographic representation of the max-function.
% :
% with regards to the infinity norm, continuous pre- and post-processing functions form a nomographic approximation of $f(x_1,\dots,x_n)$ if $\exists\epsilon \geq0$ such that
% \begin{equation}
%% \norm{ f(x_1,\dots,x_n) - \psi\left( \sum_{j=1}^n \phi_j(x_j) \right) }_{\infty} \leq \epsilon.
% \end{equation}
% Yet, the impact of the approximation on the problem has to be studied for every single case.
%
% In the case agents have to compute (\ref{eq:maxcons}), the following protocol can be used, $\forall i\in\mathcal{N}$,
% \begin{equation}
% \label{eq:approx_nom_max_prot}
% x_i(k+1)=\frac{1}{N}\log\left( \sum_{j\in N_i\cup\{i\}} e^{Nx_j(k)} \right),
% \end{equation}
% where $N>0$ is a chosen parameter.
% %where the parameter $N(\epsilon)$ is function of the approximation error $\epsilon\geq0$.
% It is straightforward to show that, $\forall k\in\mathbb{N}$,
% \begin{equation}
% \label{eq:approx_err}
% \frac{1}{N}\log\left( \sum\limits_{j\in N_i\cup\{i\}} e^{Nx_j(k)} \right)-\max(\mathbf{x}_{N_i}(k))=\epsilon_m,
% \end{equation}
% where $\epsilon_m>0$ is the nomographic approximation error.
% Randomized experiments have been performed, in order to analyze the value of the nomographic approximation error $\epsilon_m$ for different realizations of $\mathbf{x}_{N_i}(k)$. The entries of $\mathbf{x}_{N_i}(k)$ are randomly generated out of an uniform distribution between $0$ and $10$.
% The dependency of $\epsilon_m$ on $N$ is shown in Figure \ref{fig:approx_err}.
% \begin{figure}[h]
% \includegraphics[width=\columnwidth]{figures/approx_err.eps}
% \caption{Nomographic approximation error for the max-function, by varying the parameter $N$.}
% \label{fig:approx_err}
% \end{figure}
% Intuitively, implementability issues arise. First, a small approximation error $\epsilon_m$ is achieved by taking a large $N$. This makes any broadcast signal $e^{Nx_j}$ enormous with regards to $x_j$, thus giving evident implementability issues.
% Secondarily, the approximation error gives birth to a diverging behaviour. In fact, by (\ref{eq:approx_err}), $\forall k\in\mathbb{N}$,
% \begin{equation}
% \max(\mathbf{x}(k+1))>\max(\mathbf{x}(k))>\max(\mathbf{x}(0)).
% \end{equation}
% Neither an asymptotic nor a finite-time max-consensus can be achieved just by using a nomographic approximation.
% In the remainder, other viable solutions are presented.
\section{Asymptotic Max-Consensus Protocol}
\label{sec:asympt}
Computing the average function over {an ideal MAC of form} (\ref{eq:MAC_ideal}) by exploiting the interference is straightforward, since the function is trivially nomographic (\cite{molinari2018exploiting}). In the following, we present a max-consensus protocol which makes use of the average function to achieve an agreement.
The strategy is inspired by the following observation.
\begin{proposition}
\label{core_idea_expl}
%Given a connected communication topology $(\mathcal{N},\mathcal{A})$,
Given a set of agents $\mathcal{N}$ and a non-empty subset $\mathcal{M}\subseteq\mathcal{N}$,
$\forall k\in\mathbb{N}$, $\forall i\in\mathcal{N}$,
$$x_i(k)<\frac{\sum_{j\in \mathcal{M}}x_j(k)}{|\mathcal{M}|} \implies x_i(k)<\max_{j\in\mathcal{N}}({x}_j(k)).$$
%At every time-step $k\in\mathbb{N}$, for each agent $i\in\mathcal{N}$, if the received neighborhood average $u_i(k)$ is larger than the current state $x_i(k)$, then $x_i(k)<\max(\mathbf{x}(0))$.
\begin{proof}
$\forall k\in\mathbb{N}$, $\forall i\in\mathcal{N}$,
%$(\mathcal{N},\mathcal{A})$ connected implies that $N_i\not=\emptyset$. Then
$x_i(k)<\frac{\sum_{j\in \mathcal{M}}x_j(k)}{|\mathcal{M}|} \implies \exists j\in \mathcal{M},\ j\not=i: x_j(k)>x_i(k)$,
which immediately implies $x_i(k)<\max_{j\in\mathcal{M}}({x}_j(k))\leq\max_{j\in\mathcal{N}}({x}_j(k))$.
\qed
\end{proof}
\end{proposition}
% Correction by Joerg
% The proposition suggests that a non maximal agent does not need to broadcast its state.
% {The problem is however that agents do not have the global knowledge of the network, therefore they do not know whether they are maximal or not.
% Any agent can have a guess of it.
% Therefore, each agent needs to estimate whether it is maximal or not at some time point $k$.
% In other words, an agent $i\in\mathcal{N}$ needs to estimate a boolean state $y_i(k)\in\{0,1\}$ that is associated with it.}
%% An additional (boolean) state variable is then added to each agent $i\in\mathcal{N}$ and will be denoted by $y_i(k)$.
% \begin{definition}
% Given $k\in\mathbb{N}$, an agent $i\in\mathcal{N}$ for which $y_i(k)=1$ is said to be a \textbf{maximal candidate} at instant $k\in\mathbb{N}$.
% \end{definition}
Clearly, for max-consensus, any non-maximal agent does not need to broadcast its information state.
In general, however, agents do not know whether they are maximal.
We therefore settle for necessary conditions that can be locally evaluated.
The result of this local evaluation for agent $i$ at time $k$ is stored in the boolean variable $\tilde{y}_i(k)$.
If this variable is $1$, this means that agent $i$ satisfies the respective necessary condition at time $k$
and it is said to be a \textbf{maximal candidate} at time $k$.
If (and only if) this is true, the agent will be allowed to broadcast at the next time instant.
This will be expressed by an \textit{authorization variable} $y_i$, where $y_i(k)=\tilde{y}_i(k-1)$.
%State variable $y_i(k)$ will be computed as follows.
%If $y_i(k)=1$, then $i\in\mathcal{N}$ is a max-candidate at time-step $k\in\mathbb{N}$.
Let
%$\forall i\in\mathcal{N}$, $N_i^m(k)\subseteq N_i$ be
$
%\label{eq:def_N_i^m}
\forall i\in\mathcal{N},\ N_i^m(k) = \{j\in N_i\mid y_j(k)=1\}\subseteq N_i
$
be the set of neighboring agents of $i$ authorized to broadcast at time $k$.
%$i$'s neighboring agents, candidate to be maximal nodes at time-step $k\in\mathbb{N}$.
If $|N_i^m(k)|>0$, by Proposition \ref{core_idea_expl}, at time-step $k\in\mathbb{N}$, each agent $i\in\mathcal{N}$ for which
\begin{equation}
\label{eq:isMax}
x_i(k)<\frac{\sum_{j\in N_i^m(k)}x_j(k)}{|N_i^m(k)|},
\end{equation}
%where $N_i^m(k)$ is assumed to be non-empty,
cannot be a maximal candidate at time-step $k\in\mathbb{N}$.
{Therefore $y_i$ can be updated as follows:}
\begin{multline}
\forall i \in\mathcal{N},\
y_i(k+1)=\tilde{y}_i(k)=\\
=I_{\mathbb{R}_{\geq0}}\left( x_i(k)-\frac{\sum_{j\in N_i^m(k)}x_j(k)}{|N_i^m(k)|} \right).
\end{multline}
%An additional (boolean) state variable will be added to each agent $i\in\mathcal{N}$ and will be denoted by $y_i(k)$. If $y_i(k)=1$, then $i\in\mathcal{N}$ is a max-candidate at time-step $k\in\mathbb{N}$.
% This strategy creates a state-dependent communication topology.
% In fact, given a transmitting node $j\in\mathcal{N}$ and a receiving node $i\in\mathcal{N}$, the communication at time-step $k\in\mathbb{N}$ is established if $a_{ij}y_j(k)=1$.
% However, unlike \cite{mesbahi2003state}, $a_{ij}y_j(k)$ does not only depend on $x_i(k)$ and $x_j(k)$, but on each $x_l(k)$, where $l\in N_i^m(k)\cup\{i\}$.
\subsection{Protocol Design}
\label{subsec:comm_prot}
Under the assumption of a noiseless and non-fading channel, the following communication protocol, which makes use of an ideal MAC of order 2 (see \cite{goldenbaum2013harnessing}), is employed.
At every time-step $k\in\mathbb{N}$, each agent $j\in\mathcal{N}$ broadcasts two orthogonal signals, $\tau_j(k)=y_j(k)x_j(k)$ and $\tau_j'(k)=y_j(k)$.
Each agent $i\in\mathcal{N}$ receives two mutually orthogonal signals from its neighbors,
\begin{align}
\label{eq:zeta}
&z_i(k) = \sum_{j\in N_i}y_j(k)x_j(k)=\sum_{j\in N_i^m(k)}x_j(k)\\
\text{and}\nonumber&\\
&z_i'(k) = \sum_{j\in N_i}y_j(k)=\sum_{j\in N_i^m(k)}1=|N_i^m(k)|
\label{eq:zetaprime}
.
\end{align}
By making use of these two signals, each agent $i\in\mathcal{N}$, at every time $k\in\mathbb{N}$, can compute the average of information states of agents in $N_i^m(k)$, i.e.
\begin{align}
\label{eq:u_signal}
u_i(k)=
\begin{cases}
\frac{z_i(k)}{z_i'(k)}=\frac{\sum_{j\in N_i^m(k)}x_j(k)}{|N_i^m(k)|} &\text{if }N_i^m(k)\not=\emptyset\\
0 &\text{else}.
\end{cases}
\end{align}
Notice that interference has been exploited for computing $u_i(k)$.
% \begin{remark}
% If, given an agent $i\in\mathcal{N}$, at time $k\in\mathbb{N}$, all agents in $N_i$ are not max-candidates, $z_i(k)=z_i'(k)=0$. In this case, we assume $u_i(k)=0$.
% \end{remark}
\subsection{Controller Design}
Each agent $i\in\mathcal{N}$ is endowed with the following consensus dynamics:
%, which is computed at every discrete-time step $k\in\mathbb{N}$:
\begin{equation}
\label{eq:asympt_cons_prot}
\forall k\in\mathbb{N}:\ \ \
%\mathcal{S}_i:
\begin{cases}
x_i(k+1)=\max(x_i(k), u_i(k))\\
y_i(k+1)=I_{\mathbb{R}_{\geq0}}(x_i(k)-u_i(k))
\end{cases},
\end{equation}
where $y_i(1)=1$, $x_i(1)=x_{i_0}$, {and $u_i(k)$ computed as in} (\ref{eq:u_signal}).
Note that $u_i(k)$ depends on both $x_i(k)$ and $y_i(k)$.
In vector-form, (\ref{eq:u_signal})-(\ref{eq:asympt_cons_prot}) become
\begin{equation}
\label{eq:asympt_cons_prot_matForm}
\mathbf{w}(k+1)=
% \begin{bmatrix}
% \mathbf{x}(k+1)\\
% \mathbf{y}(k+1)
% \end{bmatrix}
% =
% g
% \left(
% \begin{bmatrix}
% \mathbf{x}(k)\\
% \mathbf{y}(k)
% \end{bmatrix}
% \right)
% =
g(\mathbf{w}(k))
,
\end{equation}
where
\begin{equation}
\mathbf{w}(k)=
\begin{bmatrix}
\mathbf{x}(k)\\
\mathbf{y}(k)
\end{bmatrix},
\end{equation}
and, $\forall i\in\mathcal{N},\ [\mathbf{y}(k)]_i=y_i(k)$ and $g:\mathbb{R}^{n}\times\{0,1\}^n\rightarrow\mathbb{R}^{n}\times\{0,1\}^n$ the corresponding nonlinear map.
{Using Lyapunov theory}, e.g. \cite[p. 87]{aastrom2013computer}, the convergence of (\ref{eq:asympt_cons_prot_matForm}) to max-consensus can be formalized.
To this end, we need to study some properties of the system.
\begin{proposition}
\label{prop:xincreas}
Given a network topology $(\mathcal{N},\mathcal{A})$ and
a consensus dynamics (\ref{eq:asympt_cons_prot_matForm}),
$
\forall
\mathbf{x}(1)
\in
\mathbb{R}_{\geq0}^{n}
,\ \forall i\in\mathcal{N},\
\forall k\in\mathbb{N},\
x_i(k)\leq x_i(k+1)\leq \max(\mathbf{x}(1)).
$
% \begin{multline*}
% \forall
% \mathbf{x}(0)
% \in
% \mathbb{R}_{\geq0}^{n}
% ,
% \forall k\in\mathbb{N},\ \forall i\in\mathcal{N},\\
% x_i(k)\leq x_i(k+1)\leq \max(\mathbf{x}(0)).
% \end{multline*}
\begin{proof}
By (\ref{eq:u_signal}), $\forall i\in\mathcal{N}$, $\forall k\in\mathbb{N}$, $u_i(k)\leq\max(\mathbf{x}(k))$.
Hence with (\ref{eq:asympt_cons_prot}),
the iteration (\ref{eq:asympt_cons_prot_matForm}) {generates a non-decreasing bounded sequence}
$x_i(0)\leq x_i(1)\leq\dots\leq x_i(k)\leq x_i(k+1) \leq \max(\mathbf{x}(k+1))$, $1\leq i\leq n$.
%$x_i(k)\leq x_i(k+1)$ and $x_i(k+1)\leq \max(\mathbf{x}(k))$,
This implies $x_i(k)\leq x_i(k+1)\leq \max(\mathbf{x}(k+1))=\max(\mathbf{x}(1))$.
% By (\ref{eq:asympt_cons_prot}), $\forall k\in\mathbb{N}$, $\max(\mathbf{x}(0))=\max(\mathbf{x}(k))$.
% Then, by (\ref{eq:u_signal}), $\forall i\in\mathcal{N},\ \forall k\in\mathbb{N}$, $u_i(k)\leq\max(\mathbf{x}(0))$.
% This implies that $\forall i\in\mathcal{N},\ \forall k\in\mathbb{N}$, $x_i(k+1)\leq\max(x_i(k),\max(\mathbf{x}(0)))\leq\max(\mathbf{x}(0))$ and $x_i(k)\leq x_i(k+1)$.
\qed
\end{proof}
\end{proposition}
\begin{proposition}
\label{prop:equil}
If $(\mathcal{N},\mathcal{A})$ is a connected graph,
% $
% \forall
% \mathbf{x}(0)
% \in
% \mathbb{R}_{\geq0}^{n}
% $
% ,
$
\mathbf{w}^*=
\begin{bmatrix}
\mathbf{x}^*\\
\mathbf{1}_n
\end{bmatrix}
$,
where
$
\mathbf{x}^*
%=
%\lim\limits_{k\rightarrow\infty}\mathbf{x}(k)
=
x^*\mathbf{1}_n
,\
x^*=\max(\mathbf{x}(0))
,
$
is an equilibrium point for (\ref{eq:asympt_cons_prot_matForm}),
%and is unique.
\begin{proof}
% If $(\mathcal{N},\mathcal{A})$ is connected, then $\forall i\in\mathcal{N}$, $N_i\not=\emptyset$.
% Then,
% by (\ref{eq:u_signal}) and with $\mathbf{w}(k)=\mathbf{w}^*$, $\forall i\in\mathcal{N}$, $u_i(k)=x^*$, thus yielding $y_i(k+1)=1$ and $x_i(k+1)=\max(x^*,x^*)=x^*$.
% Therefore,
% $\mathbf{w}^*$ is an equilibrium of (\ref{eq:asympt_cons_prot_matForm})
% since $\mathbf{w}^*=g(\mathbf{w}^*)$, as in \cite[p. 21]{lalo2014advanced}.
% % which is immediately proven.
{Clearly},
$
\mathbf{w}^*=
\begin{bmatrix}
{\mathbf{x}^*}',
{\mathbf{y}^*}'
\end{bmatrix}'
$
is an equilibrium point if and only if, $\forall i\in\mathcal{N}$,
\begin{align}
%\begin{cases}
\label{eq:1_proof_eq}
\mathbf{x}_i^*&=\max(\mathbf{x}_i^*,\mathbf{u}_i^*)\\
\label{eq:2_proof_eq}
\mathbf{y}_i^*&=I_{\mathbb{R}_{\geq0}}(\mathbf{x}_i^*-\mathbf{u}_i^*)
%\end{cases},
\end{align}
where
\begin{align}
\mathbf{u}_i^*=
\begin{cases}
\frac{\sum_{j\in N_i^m}\mathbf{x}_j^*}{|N_i^m|} &\text{if }|N_i^m|\not=0\\
0 &\text{else.}
\end{cases}
\end{align}
For $\mathbf{x}^*=x^*\mathbf{1}_n$ and $\mathbf{y}^*=\mathbf{1}_n$, $N_i^m=N_i$.
As $(\mathcal{N},\mathcal{A})$ is connected, $|N_i|\geq1$.
Therefore, $\mathbf{u}_i^*={x}^*$ and (\ref{eq:1_proof_eq})-(\ref{eq:2_proof_eq}) are satisfied.
\qed
\end{proof}
\end{proposition}
\begin{corollary}
\label{prop:equil_unique}
Given a connected network topology $(\mathcal{N},\mathcal{A})$, then
$
\forall
\mathbf{x}(1)
\in
\mathbb{R}_{\geq0}^{n}
$
,
$
\mathbf{w}^*
$
is the unique equilibrium point of (\ref{eq:asympt_cons_prot_matForm}).
\begin{proof}
% %Let's proceed with a \textit{reductio ad absurdum}. Let
% {The proposed proof is by contradiction. Hence, let}
% $
% \tilde{\mathbf{w}}^*=
% \left[
% \tilde{\mathbf{x}}^{*'},\
% \mathbf{1}_n'
% \right]'
%% \tilde{\mathbf{w}}^*=
%% \begin{bmatrix}
%% \tilde{\mathbf{x}}^*\\
%% \mathbf{1}_n
%% \end{bmatrix}
% $
% be an equilibrium point for (\ref{eq:asympt_cons_prot_matForm}), such that $\tilde{\mathbf{x}}^*\not={\mathbf{x}}^*$.
% By Proposition \ref{prop:xincreas}, each maximal agent at $k=0$, say agent $\bar{i}\in\mathcal{N}$, remains maximal for all $k$, i.e. $\forall k\in\mathbb{N}$, $x_{\bar{i}}(k)=x^*$.
% Therefore, since $(\mathcal{N},\mathcal{A})$ is connected, $\exists i,j\in\mathcal{N}$ such that $[\mathcal{A}]_{ji}=1$ and $[\tilde{\mathbf{x}}^*]_i>[\tilde{\mathbf{x}}^*]_j$, otherwise $\tilde{\mathbf{x}}^*={\mathbf{x}}^*$.
% Let $\underline{i}\in\mathcal{N}$ be such that $[\tilde{\mathbf{x}}^*]_{\underline{i}}=\min(\tilde{\mathbf{x}}^*)$.
% Since the network is connected and given the considerations above, there {are} always an agent $\underline{i}$ and an agent $j\in N_{\underline{i}}$, such that $[\tilde{\mathbf{x}}^*]_j>[\tilde{\mathbf{x}}^*]_{\underline{i}}$.
% Therefore, by (\ref{eq:u_signal}), $u_{\underline{i}}>[\tilde{\mathbf{x}}^*]_{\underline{i}}$, which contradicts the assumption that $\tilde{\mathbf{w}}^*$ is an equilibrium for the system (\ref{eq:asympt_cons_prot_matForm}).
% This shows that ${\mathbf{w}}^*$ is the unique equilibrium for (\ref{eq:asympt_cons_prot_matForm}) and completes the proof.
The proof is by contradiction.
Hence, assume that $\mathbf{w}^*=[{\mathbf{x}^*}',{\mathbf{y}^*}']'$
with $\mathbf{x^*}=x^*\mathbf{1}_n$
and ${\mathbf{y}^*}\not=\mathbf{1}_n$ is an equilibrium point.
The latter implies that $\mathbf{y}_i^*=0$ for at least one $i\in\mathcal{N}$.
Hence, from (\ref{eq:2_proof_eq}),
$\mathbf{x}_i^*<\mathbf{u}_i^*$
and (\ref{eq:1_proof_eq}) is violated.
This establish that $\mathbf{w}^*$ with ${\mathbf{y}^*}\not=\mathbf{1}_n$ is not an equilibrium point.
Now consider the case $\mathbf{x^*}\not=x^*\mathbf{1}_n$ and ${\mathbf{y}^*}=\mathbf{1}_n$.
Choose $i$ such that $\mathbf{x}_i^*=\min(\mathbf{x}^*)$.
As $\mathbf{x^*}\not=x^*\mathbf{1}_n$, $\mathbf{x}_i^*<x^*$.
On the other hand, by Proposition \ref{prop:xincreas}, an agent that is maximal at $k=1$
remains so $\forall k\in\mathbb{N}$.
Hence, there exists $l\in\mathcal{N}$ such that $\mathbf{x}_l^*=x^*>\mathbf{x}_i^*$.
Then, as $(\mathcal{N},\mathcal{A})$ is connected, there exists $(i,j)\in\mathcal{A}$ such that $\mathbf{x}_j^*>\mathbf{x}_i^*$.
As ${\mathbf{y}}^*=\mathbf{1}_n$, $N_i^m=N_i$ and therefore $\mathbf{u}_i^*>\mathbf{x}_i^*$.
This violates (\ref{eq:1_proof_eq}) and therefore contradicts the assumption.
\qed
\end{proof}
\end{corollary}
% COROLLARY FOR X^*_>W^*
% \begin{corollary}
% \label{prop:x_w}
% Given a multi-agent system with dynamics (\ref{eq:asympt_cons_prot_matForm}) and a connected communication topology $(\mathcal{N},\mathcal{A})$,
% the following implication holds
% $\forall k\in\mathbb{N}$,
% $$\mathbf{x}(k)=\mathbf{x}^*\implies\mathbf{w}(k+1)=\mathbf{w}^*.$$
% \begin{proof}
% We present a proof only for the non-trivial case when $\exists k\in\mathbb{N}$ such that $\mathbf{x}(k)=\mathbf{x}^*$ and $\mathbf{w}(k)\not=\mathbf{w}^*$, which is equivalent to $\mathbf{y}(k)\not=\mathbf{1}_n'$. The condition $\mathbf{x}(k)=\mathbf{x}^*$ implies, $\forall i\in\mathcal{N}$, $u_i(k)=x^*=x_i(k)$ (or $u_i(k)=0$ in case $N_i^m(k)$ is empty). Consequently, $\forall i\in\mathcal{N}$, $y_i(k+1)=1$, which verifies $\mathbf{w}(k+1)=\mathbf{w}^*$.
% \qed
% \end{proof}
% \end{corollary}
\begin{proposition}
\label{prop:contraction}
A connected network topology $(\mathcal{N},\mathcal{A})$ is given. Then,
$
\forall
\mathbf{x}(1)
\in
\mathbb{R}_{\geq0}^{n}
$
,
$\forall k\in\mathbb{N}$,
\begin{equation}
\label{eq:prop_diff2_0}
\sum_{i\in\mathcal{N}} \left( x_i(k+2)-x_i(k) \right) =0
\implies
\mathbf{x}(k)=\mathbf{x}^*
.
\end{equation}
\begin{proof}
%\begin{itemize}
%\item[$\Leftarrow$] It comes as a consequence of Proposition \ref{prop:equil}.
%\item[$\Rightarrow$]
By Proposition \ref{prop:xincreas}, $\{ x_i(k) \}_{k\in\mathbb{N}} $ is a non-decreasing sequence of nonnegative entries.
Therefore, $\sum_{i\in\mathcal{N}} (x_i(k+2)-x_i(k))=0$ if and only if $\mathbf{x}(k)=\mathbf{x}(k+1)=\mathbf{x}(k+2)$,
which implies (because of (\ref{eq:asympt_cons_prot})), $\forall i\in\mathcal{N},\ x_i(k)\geq u_i(k)$ and $x_i(k+1)\geq u_i(k+1)$ and therefore $\mathbf{y}(k+1)=\mathbf{y}(k+2)=\mathbf{1}_n$.
Therefore, $\mathbf{w}(k+1)=\mathbf{w}(k+2)$, which is possible if and only if $\mathbf{w}(k+1)=\mathbf{w}^*$ due to Corollary \ref{prop:equil_unique}.
$\mathbf{w}(k+1)=\mathbf{w}^*$ is equivalent to $\mathbf{x}(k+1)=\mathbf{x}^*$ and $\mathbf{y}(k+1)=\mathbf{1}_n$.
The latter implies that $\mathbf{x}(k)=\mathbf{u}(k)$ and therefore $\mathbf{x}(k)=\mathbf{x}(k+1)=\mathbf{x}^*$.
% which corresponds to the case $\mathbf{y}(k+1)=\mathbf{y}(k+2)=\mathbf{1}_n$.
% Therefore, $\mathbf{w}(k+1)=\mathbf{w}(k+2)$, which is possible if and only if $\mathbf{w}(k+1)=\mathbf{w}^*$ due to Corollary \ref{prop:equil_unique}.
% The statement $\mathbf{w}(k+1)=\mathbf{w}^*$ {holds if and only if} $\mathbf{x}(k)=\mathbf{x}(k+1)=\mathbf{x}^* $.
% From this, (\ref{eq:prop_diff2_0}) immediately follows.
\qed
%\end{itemize}
\end{proof}
\end{proposition}
By Proposition \ref{prop:contraction} and Proposition \ref{prop:xincreas},
$$\mathbf{x}(k)\not=\mathbf{x}^*\implies \sum_{i\in\mathcal{N}}( x_i(k+2)-x_i(k))>0.$$
% \begin{proposition}
% \label{prop:lyap}
% Given a connected communication topology $(\mathcal{N},\mathcal{A})$,
% $
% \forall k\in\mathbb{N},
% $
% $$
% V
% \left(
% \begin{bmatrix}
% \mathbf{x}(k)\\
% \mathbf{x}(k-1)
% \end{bmatrix}
% \right)
% =2n\max(\mathbf{x}(k)) - \sum_{i\in\mathcal{N}}x_i(k)+x_i(k-1).
% $$
% is a \textit{candidate Lyapunov function}
% for
% (\ref{eq:asympt_cons_prot_matForm}).
%
% \begin{proof}
% Due to the dynamics of $\mathbf{y}(k)$,
% it is possible to perform a change of variables in (\ref{eq:asympt_cons_prot_matForm}), such that the system dynamics can be described by the state ${\mathbf{v}}(k)$,
% where
% $$
% {\mathbf{v}}(k)
% =
% \begin{bmatrix}
% \mathbf{x}(k)\\
% \mathbf{x}(k-1)
% \end{bmatrix},
% $$
% and by a nonlinear map $\tilde{g}$, so that
% \begin{equation}
% \label{eq:equil_state_change}
% {\mathbf{v}}(k+1)
% =
% \tilde{g}(
% {\mathbf{v}}(k)
% ).
% \end{equation}
% The equilibrium $\mathbf{w}^*$ of (\ref{eq:asympt_cons_prot_matForm}) corresponds to the equilibrium
% $$\mathbf{v}^*
% =
% \begin{bmatrix}
% \mathbf{x}^*\\
% \mathbf{x}^*
% \end{bmatrix}
% =
% x^*\mathbf{1}_{2n}
% $$
% of (\ref{eq:equil_state_change}).
% The following analysis is based on \cite[p. 88]{aastrom2013computer} and \cite[p. 22]{lalo2014advanced}, which formalize the Lyapunov method for discrete-time systems.
%
% \begin{itemize}
% \item[a)] $V(\mathbf{v})$ is continuous in $\mathbb{R}_{\geq0}^n$, since difference of continuous functions in $\mathbb{R}_{\geq0}^n$;
% \item[b)] $V(\mathbf{v}^*)=0$;
% \item[c)] $V(\mathbf{v})$ is positive definite $\forall \mathbf{v}\in\mathbb{R}_{\geq0}^{2n}$, $\mathbf{v}\not=\mathbf{v}^*$, by Proposition \ref{prop:xincreas} and \ref{prop:equil}.
% \item[d)] $\Delta V(\mathbf{v}(k))=V(\mathbf{v}(k+1))-V(\mathbf{v}(k))=\sum_{i\in\mathcal{N}}x_i(k-1)-x_i(k+1)$.
%
% By Proposition \ref{prop:xincreas}, $\sum_{i\in\mathcal{N}}x_i(k-1)-x_i(k+1)<0$.
%
% \end{itemize}
% Since $V(\mathbf{x}(k))$ is a generalized energy function for (\ref{eq:equil_state_change}) which strictly decreases along the trajectories of the system,
% by \cite[p. 87]{aastrom2013computer},
% the state $\mathbf{v}(k)$ will asymptotically converge to $\mathbf{v}^*$.
%
%% If this Lyapunov function strictly decreases along the trajectories of the system. By Proposition \ref{prop:xincreas} and Proposition \ref{prop:contraction},
%% $$V(\mathbf{x}(k+2))-V(\mathbf{x}(k))=\sum_{i\in\mathcal{N}}x_i(k)-x_i(k+2)<0,$$
%% until consensus is reached.
%% $V(\mathbf{x}(k))$ is then a candidate Lyapunov function for the system (\ref{eq:asympt_cons_prot_matForm}) modeled by a connected communication topology $(\mathcal{N},\mathcal{A})$.
% \qed
% \end{proof}
% \end{proposition}
\begin{proposition}
\label{prop:lyap}
Given a connected network topology $(\mathcal{N},\mathcal{A})$. For every initial state $\mathbf{x}(1)\in\mathbb{R}_{\geq0}$,
the consensus protocol (\ref{eq:u_signal})-(\ref{eq:asympt_cons_prot}) converges asymptotically to max-consensus.
\begin{proof}
By (\ref{eq:asympt_cons_prot}),
$ y_i(k+1)=1 \iff x_i(k)\geq u_i(k)$
and
$ x_i(k+1)=x_i(k) \iff x_i(k)\geq u_i(k).$
Therefore, $y_i(k+1)\iff(x_i(k+1)=x_i(k))$, which can lead to rewriting $y_i(k)$ as
$y_i(k)=I_{\mathbb{R}_{\geq0}}(x_i(k-1)-x_i(k))$.
% Due to the dynamics of $\mathbf{y}(k)$,
Hence, we can rewrite (\ref{eq:asympt_cons_prot_matForm})
such that the system dynamics is described by the state
${\mathbf{v}}(k)$
% \footnote{
% This is possible, since the protocol can be rewritten as $x_i(k+1)=\max(x_i(k),u_i(k))$
% where
% $
% u_i(k)
% =
% \frac{
% \sum_{j\in N_i}
% I_{\mathbb{R}_{\geq 0}}
% (x_j(k-1)-x_j(k))
% x_j(k)
% }
% {
% \sum_{j\in N_i}
% I_{\mathbb{R}_{\geq 0}}
% (x_j(k-1)-x_j(k))
% }
% $ ($u_i(k)=0$ if $\sum_{j\in N_i}
% I_{\mathbb{R}_{\geq 0}}
% (x_j(k-1)-x_j(k))=0$).
% },
given by
$
{\mathbf{v}}(k)
=
\left[
\mathbf{x}(k)',\
\mathbf{x}(k-1)'
\right]',
$
and by {the nonlinear map $\tilde{g}$},
\begin{equation}
\label{eq:equil_state_change}
{\mathbf{v}}(k+1)
=
\tilde{g}(
{\mathbf{v}}(k)
).
\end{equation}
The equilibrium $\mathbf{w}^*$ of (\ref{eq:asympt_cons_prot_matForm}) corresponds in (\ref{eq:equil_state_change}) to the equilibrium
$\mathbf{v}^*
=
\left[
\mathbf{x}^{*'},\
\mathbf{x}^{*'}
\right]'
=
x^*\mathbf{1}_{2n}.
$
The following analysis is based on \cite[p. 88]{aastrom2013computer} and \cite[p. 22]{lalo2014advanced}, which formalize the Lyapunov method for discrete-time systems.
Let
\begin{align*}
V(\mathbf{v(k)})&=2n\max(\mathbf{v}(k))-\mathbf{1}_{2n}'\mathbf{v}(k)\\
&=2nx^* - \sum_{i\in\mathcal{N}}x_i(k)+x_i(k-1)\numberthis
\end{align*}
be a candidate Lyapunov function for (\ref{eq:equil_state_change}).
The following properties hold:
\begin{itemize}
\item[a)] $V(\mathbf{v})$ is continuous in $\mathbb{R}_{\geq0}^{2n}$;%, since difference of continuous functions in $\mathbb{R}_{\geq0}^{2n}$;
\item[b)] $V(\mathbf{v}^*)=0$;
%\item[c)] $V(\mathbf{v})$ is positive definite $\forall \mathbf{v}\in\mathbb{R}_{\geq0}^{2n}\setminus\{\mathbf{v}^*\}$, by Proposition \ref{prop:xincreas} and \ref{prop:equil}.
\item[c)] By Proposition \ref{prop:xincreas}, % and \ref{prop:equil},
$V(\mathbf{v})$ is positive definite on any trajectory of $\mathbf{v}$ satisfying (\ref{eq:equil_state_change}) unless $\mathbf{v}=\mathbf{v}^*$;
\item[d)] By Proposition \ref{prop:contraction}, $\forall \mathbf{v}(k)\not=\mathbf{v}^*$,
\begin{align*}
\Delta V(\mathbf{v}(k))&=V(\mathbf{v}(k+1))-V(\mathbf{v}(k))\\
&=\sum_{i\in\mathcal{N}}(x_i(k-1)-x_i(k+1))<0.
\end{align*}
\end{itemize}
Since $V(\mathbf{v}(k))$ is a generalized energy function for (\ref{eq:equil_state_change})
and it strictly decreases along the trajectories of the system,
by \cite[p. 87]{aastrom2013computer},
the state $\mathbf{v}(k)$ will asymptotically converge to $\mathbf{v}^*$.
\qed
\end{proof}
\end{proposition}
\begin{corollary}
\label{prop:soonerOrLater}
Given a connected network topology $(\mathcal{N},\mathcal{A})$, the following holds
\begin{multline}
\forall
\mathbf{x}(1)
\in
\mathbb{R}_{\geq0}^{n},\
\forall i\in\mathcal{N},
\\
x_i(1)<\max(\mathbf{x}(1))
\implies
\exists k_i\in\mathbb{N}:\ y_i(k_i)=0.
\end{multline}
\begin{proof}
By Proposition \ref{prop:lyap}, $\forall i\in\mathcal{N}$,
$$
\lim\limits_{k\rightarrow\infty}x_i(k)=x^*=\max(\mathbf{x}(1)).
$$
As a consequence, $\forall i\in\mathcal{N}$,
$$
x_i(1)<\max(\mathbf{x}(1))
\implies
\exists k_i\in\mathbb{N}:\ x_i(k_i)>x_i(k_i-1),
$$
which implies $y_i(k_i)=0$ (see proof of Proposition \ref{prop:lyap}).
\qed
\end{proof}
\end{corollary}
\begin{example}
\label{example:asympt}
Let a multi-agent system be modeled by a connected communication topology $(\tilde{\mathcal{N}},\tilde{\mathcal{A}})$ which is represented in Figure \ref{fig:graph_as_conv}. The vector of initial information states is $\mathbf{x}_{(1)}(1)=[4,3,3,3]'$. A simulation of (\ref{eq:u_signal})-(\ref{eq:asympt_cons_prot}) is shown in Figure \ref{fig:as_conv_plot}.
\end{example}
%% Asymptotic
\begin{figure}[t]
\centering
\begin{subfigure}{.49\columnwidth}
\centering
\includegraphics[width=\columnwidth]{graph_as_conv.png}
\caption{$(\tilde{\mathcal{N}},\tilde{\mathcal{A}})$ and $\mathbf{x}_{(1)}(0)$.}
\label{fig:graph_as_conv}
\end{subfigure}
\begin{subfigure}{.49\columnwidth}
\centering
\includegraphics[width=\columnwidth]{graph_as_conv_diffx0.png}
\caption{$(\tilde{\mathcal{N}},\tilde{\mathcal{A}})$ and $\mathbf{x}_{(2)}(0)$.}
\label{fig:graph_as_conv_diffx0}
\end{subfigure}
% \begin{subfigure}[b]{.3\columnwidth}
% \centering
% \includegraphics[width=\columnwidth]{figures/graph_as_conv_diffTop}
% \caption{$(\tilde{\mathcal{N}},\tilde{\mathcal{A}}^*)$ and $\mathbf{x}_{(1)}(0)$.}
% \label{fig:graph_as_conv_diffTop}
% \end{subfigure}
\caption{Topology and initial conditions for the examples.}
\end{figure}
%% Finite time x0
\begin{figure}[h]
\centering
\begin{subfigure}{\columnwidth}
\includegraphics[width=\columnwidth]{as_conv_plot-eps-converted-to}
\caption{Asymptotic convergence of $x_i(k)$.}
\label{fig:as_conv_plot}
\end{subfigure}
\begin{subfigure}{\columnwidth}
\includegraphics[width=\columnwidth]{as_conv_plot_diffx0-eps-converted-to}
\caption{Finite-time agreement in $6$ steps.}
\label{fig:as_conv_plot_diffx0}
\end{subfigure}
\caption{Evolution of information states for the Examples \ref{example:asympt} (top) and \ref{example:finite-time} (bottom).}
\end{figure}
% %% Finite time top
% \begin{figure}[h]
% \centering
% \begin{subfigure}[b]{\columnwidth}
% \includegraphics[width=\columnwidth]{figures/as_conv_plot_diffTop}
% \caption{Finite-time convergence of $x_i(k)$.}
% \label{fig:as_conv_plot_diffTop}
% \end{subfigure}
% \caption{$(\mathcal{N},\mathcal{A})$ and $\mathbf{x}(0)$ which give finite-time convergence in $7$ steps.}
% \end{figure}
% {Every connected multi-agent system will asymptotically converge to the max-consensus.}
% Yet, experimentally, protocol (\ref{eq:asympt_cons_prot}) allows in most cases for reaching consensus in a finite-number of steps.
% In fact, by slightly changing the problem set-up, as in Examples \ref{example:asympt1}, a finite-time convergence is achieved;
% however it should be evaluated for every single case.
We have shown that the suggested consensus protocol achieves asymptotic convergence to the max consensus
for arbitrary initial information states if the network topology is represented by a connected graph.
Moreover, extensively numerical experiments have indeed verified that in most cases finite time convergence is achieved.
This is demonstrated in the following example.
\begin{example}
\label{example:finite-time}
A multi-agent system with communication topology $(\tilde{\mathcal{N}},\tilde{\mathcal{A}})$, as in Example \ref{example:asympt}, is given. However, the vector of initial information states is now $\mathbf{x}_{(2)}(1)=[4,3,1,3]'$.
The system shows finite-time convergence, as shown in Figure \ref{fig:as_conv_plot_diffx0}.
\end{example}
% \begin{exmp}
% \label{example:asympt2}
% The communication topology of Example \ref{example:asympt} is slightly changed, as in Figure \ref{fig:as_conv_plot_diffTop},
% and is denoted by $(\tilde{\mathcal{N}},\tilde{\mathcal{A}}^*)$.
% Initial vector state is maintained as in Example \ref{example:asympt}, i.e. $\mathbf{x}_{(1)}(0)$.
% The system shows finite-time agreement, as in Figure \ref{fig:as_conv_plot_diffTop}.
% \end{exmp}
\section{Finite-Time Max-Consensus Protocol }
\label{section:finite}
By using orthogonal channel access methods and the standard max-consensus protocol (\ref{eq:maxcons}),
max-consensus is always achieved in a finite number of steps if the underlying graph is connected (\cite{nejad2009max}).
With the consensus protocol (\ref{eq:asympt_cons_prot}), max-consensus is achieved by exploiting interference.
However, in general, consensus will be reached asymptotically.
Based on the consensus protocol in Section \ref{subsec:comm_prot},
we can suggest a switching consensus dynamics that will exploit the superposition property of the wireless channel
and achieve finite-time max-consensus for arbitrary vectors of initial states.
% Via the communication protocol already presented in Section \ref{subsec:comm_prot},
% it is possible to design a switching consensus dynamics
% which allows for a fast finite-time agreement with any connected topology and any initial state vector.
% {The strategy is as follows}:
\begin{subequations}\label{eq:switching_top_dyn}
\begin{align*}
\intertext{\textbullet\ if $k = 2T_i(k):$}
&\begin{cases}
x_i(k+1)=\max(x_i(k), u_i(k))\\
y_i(k+1)=\Pi_{t=T_i(k)}^{k}y_i(t)\\
T_i(k+1)=k
\end{cases}
\label{eq:switching_top_dyn1} \numberthis&,
\intertext{\textbullet\ else:}
%\mathcal{S}_i:
&\begin{cases}
x_i(k+1)=\max(x_i(k), u_i(k))\\
y_i(k+1)=I_{\mathbb{R}_{\geq0}}(x_i(k)-u_i(k))\\
T_i(k+1)=T_i(k)
\end{cases}
\label{eq:switching_top_dyn2} \numberthis&.
\end{align*}
\end{subequations}
An additional state variable, $T_i:\mathbb{N}\rightarrow\mathbb{N}$, has been added to the system and has initial conditions $T_i(1)=1$.
The initial conditions $x_i(1)$ and $y_i(1)$ are the same as in (\ref{eq:asympt_cons_prot}).
The controller switches between two dynamics: it keeps (\ref{eq:switching_top_dyn2})
(which has the same dynamics as (\ref{eq:asympt_cons_prot})) in every time step with the exception of $k=2^{p},\ p\in\mathbb{N}$,
when the protocol has dynamics (\ref{eq:switching_top_dyn1}).
In this case, $T_i$ is updated by $T_i(k+1)=2T_i(k)$.
Moreover, $y_i(k+1)$ is computed so that any agent that has not been a
\textit{maximal candidate} in all of its last $T_i(k)$ time steps
will not be a maximal candidate at time $k+1$.
%Remark \ref{prop:sameDyn} and Corollary \ref{prop:soonerOrLater} will be of use for the finite-time convergence proof.
% \begin{remark}
% \label{prop:sameDyn}
% %Given the same initial conditions and the same topology $(\mathcal{N},\mathcal{A})$,
% (\ref{eq:asympt_cons_prot}) and (\ref{eq:switching_top_dyn2}) exhibit the same dynamics for $x_i(k)$ and $y_i(k)$.
% \end{remark}
% Proposition \ref{prop:Tidiverge} analyzes the behaviour of the new state variable $T_i(k)$.
% \begin{proposition}
% \label{prop:Tidiverge}
% Given the switching dynamics (\ref{eq:switching_top_dyn}),
% \begin{equation}
% \label{eq:Tieverywhere}
% \forall k\in\mathbb{N},\ \forall i,j\in\mathcal{N},\ T_i(k)=T_j(k)
% \end{equation}
% and
% \begin{equation}
% \label{eq:Tidiverge}
% \forall i\in\mathcal{N},\ \lim\limits_{k\rightarrow\infty}T_i(k)=\infty.
% \end{equation}
%
% \begin{pf}
% {$T_i(k)$ does only depend on the discrete-time index $k\in\mathbb{N}$, therefore each agent retains the same value (given that they all have the same initial conditions), as in }(\ref{eq:Tieverywhere}).
% It is straightforward to show that the sequence $$\{T_i(k-1)+T_i(k) \}_{k\in\mathbb{N}}=\{1,1,2,3,5,8,13,\dots \}$$ is a \textit{Fibonacci series}, which is a diverging integer series. This proves (\ref{eq:Tidiverge}).
% \qed
% \end{pf}
% \end{proposition}
\begin{proposition}
\label{prop:finite_time}
Let $(\mathcal{N},\mathcal{A})$ be a connected graph.
Then, the switching consensus protocol (\ref{eq:switching_top_dyn}) achieves finite-time max-consensus
for any $\mathbf{x}(1)\in\mathbb{R}_{\geq0}$.
\begin{proof}
System (\ref{eq:switching_top_dyn}) has the same dynamics as (\ref{eq:asympt_cons_prot}) in all time steps except for $k=2^{p},\ p\in\mathbb{N}$, in which the system dynamics is (\ref{eq:switching_top_dyn1}).
%Let us assume {the case when the protocol} (\ref{eq:asympt_cons_prot}) {does not allow in general the system for finite-time max-consensus} (e.g. Example \ref{example:asympt}).
If, at time $k\in\mathbb{N}$, all non-maximal agents are not
%maximal candidates
authorized to broadcast
(i.e. $\forall i\in\mathcal{N}, x_i(k)<x^*\implies y_i(k)=0$),
all agents in set
$$
\mathcal{L}_1(k)=\{ i\in N_j\mid x_j(k)=x^* \}
$$
will get to max-consensus at step $k+1$.
By Corollary \ref{prop:soonerOrLater},
%$$
%x_i(0)<x^*
%\implies
%\exists k_i\in\mathbb{N}:\ y_i(k_i)=0.
%$$
%Therefore, $\exists k\in\mathbb{N}$ large enough, such that, $\forall i\in\mathcal{N}$,
$\exists \tilde{k}\in\mathbb{N}$ large enough, such that, $\forall i\in\mathcal{N}$,
\begin{multline*}
x_i(T_i(\tilde{k}))<x^*%\max(\mathbf{x}(T_i(k))=x^*\\
\implies
\exists k_i\in[T_i(\tilde{k}),\tilde{k}]:\ y_i(k_i)=0.
\end{multline*}
As a consequence and according to (\ref{eq:switching_top_dyn1}),
%$\exists k\in\mathbb{N}$, such that
$\forall i\in\mathcal{N}$,
$$
x_i(T_i(\tilde{k}))<x^*%\max(\mathbf{x}(T_i(k))
\implies
y_i(2T_i(\tilde{k})+1)=0,
$$
which implies that
$$
\forall i\in\mathcal{L}_1(2T_i(\tilde{k})),\ x_i(2T_i(\tilde{k})+2)=x^*.
$$
%This happens recursively until $\exists \bar{k}\in\mathbb{N}:\ \forall i\in\mathcal{N},\ x_i(\bar{k})=x^*$.
As $|\mathcal{N}|$ is finite, within a finite number of {recursions}, max-consensus is achieved.
\qed
\end{proof}
\end{proposition}
\section{Simulation}
\label{sec:simul}
% In the following,
% multi-agent systems endowed with
% the switching consensus dynamics (\ref{eq:switching_top_dyn})
% are simulated.
% First, we propose again the scenario in Example \ref{example:asympt},
% which is shown to achieve finite-time consensus with protocol (\ref{eq:switching_top_dyn}).
% Then, randomized simulations are performed,
% thus showing what happens with arbitrarily large networks.
% Finally, a comparison with the traditional solution
% is presented.
\subsection{Scenario}
The multi-agent system with communication graph $(\tilde{\mathcal{N}},\tilde{\mathcal{A}})$ and with initial state $\mathbf{x}_{(1)}(1)$ as
in Figure \ref{fig:graph_as_conv} converges asymptotically to the agreement.
%Finite-time consensus was not achievable with dynamics (\ref{eq:asympt_cons_prot}).
As claimed in Section \ref{section:finite}, by using (\ref{eq:switching_top_dyn}),
any system with a connected communication topology
achieves finite-time consensus.
The agents in $(\tilde{\mathcal{N}},\tilde{\mathcal{A}})$ are therefore endowed with the switching consensus dynamics (\ref{eq:switching_top_dyn}).
In Figure \ref{fig:fin_conv_plot}, the agents' information states are plotted
and shown to achieve agreement in $\bar{k}=9$ steps.\
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{fin_conv_plot-eps-converted-to}
\caption{System in Figure \ref{fig:graph_as_conv} endowed with (\ref{eq:switching_top_dyn}) achieves finite-time agreement.}
\label{fig:fin_conv_plot}
\end{figure}
\subsection{Randomized Scenarios}
A multi-agent system endowed with consensus protocol (\ref{eq:switching_top_dyn}) and
with network topology $(\mathcal{N}_l,\mathcal{A}_l)$, with initial state vector $\mathbf{x}_l(1)$,
as in Figure \ref{fig:random_graph},
is simulated.
The network size is $|\mathcal{N}_l|=20$, and its communication topology is connected.
Each entry of the initial information state vector $\mathbf{x}_l(1)$ is drawn from a uniform distribution $\mathcal{U}(0,2\pi)$.
Figure \ref{fig:random_plot} shows that agents get to consensus in a finite number of steps, i.e.
$\forall i\in\mathcal{N}_l,\ x_{i}(\bar{k})=\max(\mathbf{x}_l(0))$. In this example, $\bar{k}=21$.
% In the following, a possible relation between the communication topology and $\bar{k}$ is presented.
%
%
% The algebraic connectivity is a measure of how well the graph is connected \cite{de2007old}. It is also referred to as \textit{Fiedler's eigenvalue}, since it is the second smallest eigenvalue of the graph Laplacian. Intuitively, if a graph is better connected, the convergence to a consensus should be faster. This has been empirically analyzed in case consensus dynamics is (\ref{eq:switching_top_dyn}).%, by using randomization.
% \begin{figure}[h]
% \includegraphics[width=\columnwidth]{figures/randomized/k_algCon}
% \caption{Finite-time convergence as function of the graph connectivity. The randomized analysis consists of $323068$ different scenarios. Networks having between $10$ and $100$ nodes have been analyzed.}
% \label{fig:k_algCon}
% \end{figure}
% In fact, a large number of different scenarios have been randomly simulated by a calculator.
% Each one consisted of a multi-agent system with a randomized connected communication topology
% and an initial state vector, whose entries were drawn out of a uniform distribution $\mathcal{U}(0,2\pi)$.
% In Figure \ref{fig:k_algCon}, a possible relation between $\bar{k}$ and the algebraic connectivity $\lambda_2$ can be identified.
% As intuitively stated, systems whose topologies have smaller connectivity measures need more steps to achieve consensus.
%
% Another interesting result emerges from Figure \ref{fig:k_algCon}.
% The value of $\bar{k}$ does not increase with $|\mathcal{N}|$.
% This will be of high interest in the next section, since traditional protocols normally exhibit an increase of delays and complexity for larger networks.
\begin{figure}
\centering
\begin{subfigure}{\columnwidth}
\centering
\includegraphics[width=\columnwidth]{graph}
\caption{$({\mathcal{N}}_l,{\mathcal{A}}_l)$ and $\mathbf{x}_{l}(0)$.}
\label{fig:random_graph}
\end{subfigure}
\begin{subfigure}{\columnwidth}
\includegraphics[width=\columnwidth]{plot-eps-converted-to}
\caption{Finite-time convergence of $x_i(k)$.}
\label{fig:random_plot}
\end{subfigure}
\caption{System with dynamics (\ref{eq:switching_top_dyn}).}
\end{figure}
\subsection{Comparison with traditional solution}
\begin{figure}[h]
\includegraphics[width=\columnwidth]{_TDMA-eps-converted-to}
\caption{Ratio of steps required for converging with traditional ($\bar{k}_{t}$) and broadcast protocols ($\bar{k}_{b}$). }
\label{fig:_TDMA}
\end{figure}
In the following, by \textit{traditional protocols}, we denote
those strategies which use orthogonal channel access methods such as TDMA (Time Division Multiple Access) for communication and a consensus dynamics like (\ref{eq:maxcons}).
With TDMA, the transmission is multiplexed in time-domain.
Each agent will then be allowed to transmit its information state in a pre-assigned time slot (this will avoid interference between signals coming from different agents).
Accordingly, for a network modeled by topology $(\mathcal{N},\mathcal{A})$, $|\mathcal{N}|$ time slots should be assigned one-to-one to all nodes within one sampling interval.
Intuitively, each run of (\ref{eq:maxcons})
corresponds to $\frac{|\mathcal{N}|}{2}$ runs of (\ref{eq:switching_top_dyn}),
since the latter uses a communication protocol based on the broadcast of only two orthogonal signals,
as presented in Section \ref{subsec:comm_prot}.
A comparison (via randomized analysis) between traditional protocols and the one proposed in this paper in terms of steps required for converging
%always by using randomization analysis,
is presented in Figure \ref{fig:_TDMA}.
Given a topology and an initial state vector, $\bar{k}_{t}$ denotes the number of steps required to achieve consensus with traditional protocols;
$\bar{k}_{b}$ represents, for the same problem, the number of steps required to achieve consensus with broadcast protocol (\ref{eq:switching_top_dyn}).
Their ratio is $r=\frac{\bar{k}_{t}}{\bar{k}_{b}}$.
For networks with more than circa $20$ agents, adopting broadcast solutions gives faster convergence.
In case $|\mathcal{N}|=100$, broadcast algorithm (\ref{eq:switching_top_dyn}) achieves consensus between $5$ and $20$ times faster than traditional approaches.
\section{Conclusion}
\label{sec:concl}
In this paper,
a consensus protocol for reaching max-consensus in a finite number of steps
is suggested.
Its main characteristic is that it employs the broadcast and superposition properties of the wireless channel.
This can drastically reduce the number of required messages.
Indeed
simulations show that this algorithm exhibits considerably faster convergence
than traditional approaches.
The wireless channel has been modeled by an ideal MAC.
Future work will analyze the impact of nonidealities, such as channel coefficients and receiver noise.
%, on the problem.
%Finally, approaches which will allow for identifying the initial max agent will be in future proposed.
\bibliography{biblio}
\bibliographystyle{IEEEtran}
\end{document}
\begin{thebibliography}{14}
\providecommand{\natexlab}[1]{#1}
\providecommand{\url}[1]{\texttt{#1}}
\providecommand{\urlprefix}{URL }
\expandafter\ifx\csname urlstyle\endcsname\relax
\providecommand{\doi}[1]{doi:\discretionary{}{}{}#1}\else
\providecommand{\doi}{doi:\discretionary{}{}{}\begingroup
\urlstyle{rm}\Url}\fi
\bibitem[{{\AA}str{\"o}m and Wittenmark(2013)}]{aastrom2013computer}
{\AA}str{\"o}m, K.J. and Wittenmark, B. (2013).
\newblock \emph{Computer-controlled systems: theory and design}.
\newblock Courier Corporation.
\bibitem[{Buck(1982)}]{buck1982nomographic}
Buck, R.C. (1982).
\newblock Nomographic functions are nowhere dense.
\newblock \emph{Proceedings of the American Mathematical Society}, 195--199.
\bibitem[{Cover and Thomas(2012)}]{cover2012elements}
Cover, T.M. and Thomas, J.A. (2012).
\newblock \emph{Elements of information theory}.
\newblock John Wiley \& Sons.
\bibitem[{Goldenbaum et~al.(2012)Goldenbaum, Boche, and
Sta{\'n}czak}]{goldenbaum2012nomographic}
Goldenbaum, M., Boche, H., and Sta{\'n}czak, S. (2012).
\newblock Nomographic gossiping for f-consensus.
\newblock In \emph{Modeling and Optimization in Mobile, Ad Hoc and Wireless
Networks (WiOpt), 2012 10th International Symposium on}, 130--137. IEEE.
\bibitem[{Goldenbaum et~al.(2013)Goldenbaum, Boche, and
Sta{\'n}czak}]{goldenbaum2013harnessing}
Goldenbaum, M., Boche, H., and Sta{\'n}czak, S. (2013).
\newblock Harnessing interference for analog function computation in wireless
sensor networks.
\newblock \emph{IEEE Transactions on Signal Processing}, 61(20), 4893--4906.
\bibitem[{Iutzeler et~al.(2012)Iutzeler, Ciblat, and
Jakubowicz}]{iutzeler2012analysis}
Iutzeler, F., Ciblat, P., and Jakubowicz, J. (2012).
\newblock Analysis of max-consensus algorithms in wireless channels.
\newblock \emph{IEEE Transactions on Signal Processing}, 60(11), 6103--6107.
\bibitem[{Kolmogorov(1963)}]{kolmogorov1963representation}
Kolmogorov, A.N. (1963).
\newblock On the representation of continuous functions of many variables by
superposition of continuous functions of one variable and addition.
\newblock \emph{Translations American Mathematical Society}, 2(28), 55--59.
\bibitem[{Limmer et~al.(2015)Limmer, Mohammadi, and
Sta{\'n}czak}]{limmer2015simple}
Limmer, S., Mohammadi, J., and Sta{\'n}czak, S. (2015).
\newblock A simple algorithm for approximation by nomographic functions.
\newblock In \emph{Communication, Control, and Computing (Allerton), 2015 53rd
Annual Allerton Conference on}, 453--458. IEEE.
\bibitem[{Magni and Scattolini(2014)}]{lalo2014advanced}
Magni, L. and Scattolini, R. (2014).
\newblock \emph{Advanced and multivariable control}.
\newblock Pitagora edirice Bologna.
\bibitem[{Molinari and Raisch(2018)}]{molinari2018automation}
Molinari, F. and Raisch, J. (2018).
\newblock Automation of road intersections using consensus-based auction
algorithms.
\newblock In \emph{American Control Conference (ACC)}. IEEE.
\bibitem[{Molinari et~al.(2018)Molinari, Sta{\'n}czak, and
Raisch}]{molinari2018exploiting}
Molinari, F., Sta{\'n}czak, S., and Raisch, J. (2018).
\newblock Exploiting the superposition property of wireless communication for
average consensus problems in multi-agent systems.
\newblock In \emph{European Control Conference (ECC)}.
\bibitem[{Nejad et~al.(2009)Nejad, Attia, and Raisch}]{nejad2009max}
Nejad, B.M., Attia, S.A., and Raisch, J. (2009).
\newblock Max-consensus in a max-plus algebraic setting: The case of fixed
communication topologies.
\newblock In \emph{Information, Communication and Automation Technologies,
2009. ICAT 2009. XXII International Symposium on}, 1--7. IEEE.
\bibitem[{Olfati-Saber and Murray(2004)}]{olfati2004consensus}
Olfati-Saber, R. and Murray, R.M. (2004).
\newblock Consensus problems in networks of agents with switching topology and
time-delays.
\newblock \emph{IEEE Transactions on automatic control}, 49(9), 1520--1533.
\bibitem[{Ren et~al.(2007)Ren, Beard, and Atkins}]{ren2007information}
Ren, W., Beard, R.W., and Atkins, E.M. (2007).
\newblock Information consensus in multivehicle cooperative control.
\newblock \emph{IEEE Control Systems}, 27(2), 71--82.
\end{thebibliography}