Emergent structure and dynamics of tropical forest-grassland landscapes
Extracted source text · arXiv 2207.14118
\documentclass[reprint,english,twocolumn,amsfonts,amssymb,amsmath]{revtex4-1}
% Use the lineno option to display guide line numbers if required.
% 9pt,twocolumn,twoside
% note: avoid eqnarray (has issues with referencing and numbering)
%\templatetype{pnasresearcharticle} % Choose template
\usepackage[T1]{fontenc}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{stackrel}
\usepackage{graphicx}
\usepackage{natbib}
\usepackage{svg}
%\usepackage[notcite,notref]{showkeys}
\usepackage{bbm}
\usepackage{float}
\usepackage{mathtools}
\usepackage[unicode=true,pdfusetitle,
bookmarks=true,bookmarksnumbered=false,bookmarksopen=false,
breaklinks=false,backref=false,colorlinks=true]
{hyperref}
\hypersetup{linkcolor=blue, urlcolor=blue, citecolor=blue}
\usepackage[nameinlink,capitalise]{cleveref}
\usepackage{multirow}
%full reference to sections
\makeatletter
\renewcommand{\p@subsection}{\p@section\thesection} % for cross-refs to subsections
\renewcommand{\p@subsubsection}{\p@subsection\thesubsection} % for cross-refs to subsubsections
\makeatother
\creflabelformat{equation}{#2\textup{#1}#3}%remove brackets from eq refs
%commenting:
%% JS:
\newcommand{\jsq}[1]{\textbf{\color{red}[JS: #1]}}
\newcommand{\js}[1]{{\color{blue}#1}}
%\newcommand{\jsc}[2]{{\color{blue}#1}\textbf{\color{red}[previous: #2]}}
\newcommand{\jsc}[2]{{\color{blue}#1}{}}%\textbf{\color{red}[previous: #2]}}
\newcommand{\sla}{$\leftarrow$}
\newcommand{\sra}{$\rightarrow$}
\newcommand{\fgg}{\langle[FG]\rangle_{\mathrm{cg}}}
%% BW:
\newcommand{\bwq}[1]{\textbf{\color{purple}[BW: #1]}}
\newcommand{\bw}[1]{\textbf{\color{teal}#1}}
\newcommand{\todo}{{\textbf{\color{purple}[...]}}}
% Please add a significance statement to explain the relevance of your work (<=120w)
%\significancestatement{
%It has been hypothesised that an amplifying feedback between forest loss and fire spread
%on grass patches can cause abrupt dieback of tropical forest, but this has been
%challenging to test unequivocally. To obtain tractable
%models, most theoretical analysis so far has
%disregarded the spatial nature of fire and forest spread by assuming that vegetation patches are well mixed. We analyse a recently
%proposed microscale model, using a method from
%control theory for tractability. Taking into account
%that spreading processes occur near the forest perimeter instead of in
%a well-mixed fashion, we find emergent relations between forest structure and dynamics. These relations can be used to test if and where fire-vegetation feedbacks are strong enough to cause abrupt shifts.
%}
% Please include corresponding author, author contribution and author declaration information
%\authorcontributions{BW performed research; JS supervised research; BW \& JS wrote the paper.}
%\authordeclaration{The authors declare no conflict of interest.}
%\correspondingauthor{\textsuperscript{1}To whom correspondence should be addressed. E-mail: b.wuyts@ex.ac.uk}
% At least three keywords are required at submission. Please provide three to five keywords, separated by the pipe symbol.
\keywords{tropical forest $|$ bistability $|$ cellular automata $|$ percolation $|$ nonlinear dynamics $|$ mean field $|$ feedback control}
%\dates{This manuscript was compiled on \today}
%\doi{\url{www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX}}
\begin{document}
\title{Emergence of nonlinear dynamics from spatial structure \\in tropical forest-grassland landscapes}
\author{Bert Wuyts}
\affiliation{College of Engineering, Mathematics and Physical Sciences,
University of Exeter, EX4 4QF, UK}
\author{Jan Sieber}
\affiliation{College of Engineering, Mathematics and Physical Sciences,
University of Exeter, EX4 4QF, UK}
\email{b.wuyts@ex.ac.uk}
\begin{abstract}
It is thought that tropical forests can exist as an alternative stable state to savanna \cite{Staver2011b,Hirota2011}. Therefore,
the cumulative effects of perturbation by climate change or human impact may
lead to crossing of a tipping point beyond which there is rapid large-scale forest dieback
that is not easily reversed \cite{Lenton2008,Scheffer2015}. Empirical evidence for
bistability due to fire-vegetation feedbacks relies on tree cover bimodality in
satellite-observed data \cite{Staver2011b,Hirota2011}, but this may also be
explained by spatial heterogeneity \cite{Wuyts2017},
or by biases in the data \cite{Hanan2014,Gerard2017}. Most modelling studies of alternative stable tree cover states have so far either relied on mean-field assumptions
\cite{Staver2012,Schertzer2015,Touboul2018,Goel2020,Wuyts2017,Wuyts2019}
or not included the spatiotemporal dynamics of fire \cite{Patterson2021}, making it
hard to compare model results to spatial data.
In this work, we analyse a microscopic model of tropical forest dynamics and fire spread,
proposed by \cite{Hebert-Dufresne2018}, to show how dynamics of forest
area emerge from spatial structure.
We find that the relation between forest
perimeter and area determines the nonlinearity in forest growth while
forest perimeter weighted by adjacent grassland area
determines the nonlinearity in forest loss. Together with the linear changes,
which are independent of spatial structure, these two effects lead to
an emergent relation between forest area change and forest area,
defining a single-variable ordinary differential equation.
Such a relation between pattern and dynamics offers falsifiable theoretical
predictions of the bistability hypothesis
that are more closely linked to the underlying mechanism
than bimodality and provides a criterion for forest vulnerability.
\end{abstract}
\maketitle
\section{Introduction}
%In tropical ecosystems, tree cover suppresses local
%fire spread \cite[e.g.][]{Hennenberg2006}
%and enhances regional rainfall
%\cite{Spracklen2018}, such that losing forest promotes fire spread and
%decreases regional rainfall, leading to further forest loss. It is thought that
%these feedbacks make parts of the ecosystem prone to tipping \cite{Lenton2008,Scheffer2015}
%from a high tree cover forest state to a degraded low tree cover state,
%via climate drying in combination with shocks, such as
%drought or deforestation. Due to the hysteresis associated
%with these \emph{alternative stable states}, restoration to the high tree cover state
%would only be possible after sufficiently large climate wettening.
Satellite observations of tropical tree cover show that in a wide
range of mean annual rainfall \cite{Staver2011b,Hirota2011},
both savanna (low tree cover) and forest (high tree cover) are possible, making the
distribution of tree cover \emph{bimodal}.
On the one hand, fire exclusion experiments show that fire can maintain low tree cover
\cite{Bond2008}. On the other hand, fire
occurs almost exclusively below a tree cover threshold of about 40\%
\cite{Staver2011b,Hoffmann2012,Archibald2009,Wuyts2017,VanNes2018}, which is
consistent with fire being a \emph{percolation} process \cite{Stauffer1994,Christensen2005} on grass patches \cite{Schertzer2015,Cardoso2022}. Such a highly nonlinear response of fire to grass together with an empirically consistent
response of vegetation to fire was shown to be sufficient for inducing \emph{bistability} in simple
models \cite{Staver2012}. Taken together, the bimodality, the two-way interaction between fire and
vegetation, and the availability of a plausible underlying mechanism provide compelling
%\jsq{strong}\bwq{(creating contrast with the below by implying here that it is tempting to consider it as conclusive.)}
evidence that forest and savanna are \emph{alternative
stable states}, maintained by a feedback between vegetation and fire \cite{Staver2011b},
and between which transitions would neither be gradual nor easily reversed
\cite{Lenton2008,Scheffer2015}. This local-scale fire-vegetation feedback may exacerbate the effects of the regional-scale feedback between vegetation and rainfall \cite{Spracklen2018}.
%Existence of such alternative stable states implies that parts of the ecosystem are prone to
%\emph{tipping} \cite{Lenton2008,Scheffer2015}
%from a high tree cover forest state to a degraded low tree cover state,
%via climate drying in combination with shocks, such as
%drought or deforestation. Due to hysteresis, restoration to the high tree cover state
%would only be possible after sufficiently large climate wettening.
Nevertheless, subsequent analysis of the same data
showed that at least part of the bimodality can be explained by independent spatial heterogeneity
of parameters other than mean annual rainfall \citep{Wuyts2017}. Yet another possibility is
that the bimodality results from
artificial discontinuities in the data, introduced by its generative machine-learning algorithm
\cite{Hanan2014,Gerard2017}.
Therefore, the fire-mediated alternative stable state hypothesis is
still contested, with some left unconvinced and in favour
of other causal mechanisms that distinguish between savanna and forest
\cite{Veenendaal2018}. Taking into consideration the possible alternative
explanations for bimodality, further tests of bistablity using remote-sensed
data would have to rely on more specific indicators than bimodality.
For their specificity, they should ideally
be closely tied to the assumed underlying mechanism, while for comparability with spatial data, they should relate to space.
Indicators are provided by theoretical models. Despite the fact that tropical vegetation
dynamics is a high-dimensional spatial stochastic process, its conceptual analysis has mostly been concerned with models consisting of
a small number of ordinary differential equations (ODE) representing macroscale dynamics \cite[e.g.][]{VanLangevelde2003,Accatino2010,Staver2012}, possibly including a noise
\cite{DOdorico2006,Touboul2018} or impulsive \cite[e.g.][]{Baudena2010,Beckage2011,Hoyer-Leitzel2021} term. Such an approach is justified
under validity of \emph{mean-field assumptions}: (\emph{i}) well mixedness, allowing the probability of
interaction between two species to be written as the product of their overall occurrence
probability, (\emph{ii}) large system size and validity of the law of large numbers in the deterministic case or an appropriate limit theorem for the noise term in the stochastic case, (\emph{iii}) spatial homogeneity.
%These permit
%writing the macroscopic or mesoscopic dynamics of a high-dimensional stochastic spatial system in the form of a small number of deterministic differential equations.
E.g., the Staver-Levin
model of tropical tree cover bistability \cite{Staver2012} is an ODE in which the variables
represent the system-scale fractions of area covered by grass, forest trees,
savanna trees or savanna saplings. Species interact through competition for
space or through the mediating effect of grass cover on fire occurrence.
By the well-mixedness
assumption, competition terms involve products of these macroscopic cover fractions and
fire effects on tree cover are functions of macroscale grass cover multiplied by
macroscale tree cover. Taking account of the
% also mention homogeneity of pars; large size->no noise?
threshold dependence of fire on grass cover, with fire
affecting forest trees, but not the fire-adapted savanna trees, except for their recruitment from saplings, the Staver-Levin model
exhibits tree cover bistability but also oscillatory
behaviour \cite{Staver2012,Touboul2018}.
Assuming well-mixedness allows a greatly simplified description,
but it comes at the expense of disregard for spatial structure.
Therefore, mean-field models cannot provide predictions or indicators that relate to spatial structure.
Spatial mean-field models of tree cover bistability, such as
\emph{reaction-diffusion} \cite{Wuyts2017,Wuyts2019,Goel2020} and \emph{integro-partial-differential} \cite{Patterson2021} equations,
partially alleviate this problem by
allowing for spatial variation on the mesoscale. They show that, due to stationary fronts in heterogeneous environments, bistability does not necessarily lead to bimodality, unless interfacial effects are large
\cite{Goel2020} or the forest and savanna components are sufficiently decoupled \cite{Wuyts2019}. However, they still rely on
mean-field assumptions between the meso- and microscale, thereby still neglecting the effects of microscopic spatial structure. As trees or forest patches are typically highly clustered in space, this neglect of structure by the well-mixedness assumption is usually unjustified.
%In tropical forests, mean-field assumptions are usually
%violated: forest patches are highly clustered, tree or patch population sizes are finite,
%spreading processes such as plant dispersal or fire can generate large
%fluctuations \cite{Henkel2008}, particularly near tipping points, and environmental
%conditions are often highly heterogeneous.
Contrary to mean-field models, \emph{microscopic spatial simulation models},
such as interacting particle systems or cellular automata,
capture detailed spatial structure, thereby also allowing them to provide
indicators that are testable by spatial data.
The \emph{interacting particle system} of Patterson et al. \cite{Patterson2021} reproduces many of the behaviours of the Staver-Levin model, to which it was proven to converge under mean-field assumptions, but it
also shows front solutions, as in the spatial mean-field models,
and finite-size effects, such
as extinction after a finite simulation time. However, it did not include
fire as a percolation process,
even though percolation was assumed to play the key role in inducing
nonlinearity and bistability \cite{Staver2012,Schertzer2015,Archibald2012a,Cardoso2022}.
%Fire percolation models on a fixed landscape
%consisting of a mix of flammable grass patches and non-flammable tree patches
%show the appearance of a giant connected component of grass cells
%above a grass cover of about 60\% \citep{Schertzer2015,Archibald2012a}, leading to
%system-scale fires.
The cellular automaton of H\'ebert-Dufresne et
al. \cite{Hebert-Dufresne2018}, which can be considered as a simplified
version of \cite{Favier2004}, captures the joint dynamics of forest cover and fire
percolation. In their model,
fires spread rapidly and nearly perfectly on grass but can only burn forests at their edge. Besides bistability
of forest cover, they also found that isolated forest patches obtain a steady state shape
characterised by a perimeter-area power law relation consistent with fits based on remote-sensed observations in 86 protected areas of the Brazilian Cerrado.
Stochastic spatial simulation models are, however, far
harder to analyse than mean-field models. In particular, simulations only give access to the stable parts of phase space, unless one is
prepared to wait prohibitively long.
In this work, we analyse the nonlinear dynamics of the cellular automaton
proposed by \cite{Hebert-Dufresne2018} without relying on mean-field assumptions.
Instead, we start from the observation that fire spreads on grass much faster than forest and assume that it spreads perfectly, reaching the entire perimeter of the ignited
grass patch, and damaging the forest at this perimeter. This assumption makes two structural quantities relevant for the long-term evolution of the
forest-grass landscape: the \emph{forest perimeter} for gain from forest spread, and, the \emph{grassland-weighted forest perimeter}, which counts each part of the forest perimeter weighted by the size of its adjacent grass patch, for loss induced by fire. Via simulations,
we test how these perimeter quantities affect the dynamics.
We are able to explore the unstable parts of phase space by first obtaining unstable
solutions from feedback-controlled simulations.
This will show that both perimeter quantities are approximate functions
of forest area throughout the dynamics, allowing the evolution of forest area to be
written as a single-variable ODE. In this equation,
the nonlinear terms emerge from competition of forest and grass patches at their perimeter,
while the linear terms arise from spatially independent spontaneous conversions.
\section{Results}
\subsection{The FGBA probabilistic cellular automaton}
\begin{figure*}[t]
\centering
\includesvg[inkscapelatex=false,width=1.7\columnwidth]{./figures/FGBA0}
\caption{The FGBA stochastic cellular automaton: (a) state transition
diagram (coloured rates: spread to neighbouring cell, black rates: spontaneous conversion within cell), (b) example time series of a simulation starting at zero
tree cover, (c,d) $10$x and $10^7$x zoom of (b), (e-h) snapshots of
a simulation at indicated times for low fire ignition rate ($\phi N=0.075$). The fire in (f) spreads throughout grassland in the whole domain whereas that in (g) went extinct locally because forest splits grassland in clusters (notice the area of ash near the top). Remaining parameters are shown in \cref{tab:rates}. Domain size: 200x200 cells.}
\label{fig:FGBAplots}
\end{figure*}
\begin{figure}[t]
\centering
\includesvg[inkscapelatex=false,width=1\columnwidth]{./figures/bifbistab2}
\caption{Steady states and bistability of forest area in the FGBA cellular automaton. (a) Bifurcation diagram of forest area fraction $[F]$ versus fire ignition rate $\phi$ (shade: two-standard deviation confidence interval of the mean). (b) Simulations initiated at two different points on the saddle ($\phi N{=}0.257$ and $\phi N{=}1.32$). Remaining parameters are shown in \cref{tab:rates}. Domain size: 100x100 cells.}
\label{fig:intro}
\end{figure}
The FGBA probabilistic cellular automaton
(adapted from \citep{Hebert-Dufresne2018} -- see Methods) models the
stochastic dynamics of
tropical vegetation and fire on a square lattice and in continuous time.
The key empirical facts of tropical forest and fire dynamics captured by the FGBA automaton are: (\emph{i}) fires
only naturally ignite in grasslands but they can spread into forest, (\emph{ii}) fires
spread faster in grassland than in forest, such that forests suppress
fires, albeit imperfectly, (\emph{iii}) forest dynamics occurs on a \emph{strongly separated
time scale} from fire spread and grass regrowth.
This results in the following reaction
rules in the cellular automaton. At any time, each lattice cell
can be in one of four states: $F$ -- forest, $G$ -- grass, $B$ -- burning, $A$ -- ash. Conversions between these
states can occur spontaneously or due to spread to neighbouring cells (\cref{fig:FGBAplots}a and \cref{tab:rates}). The spontaneous conversions are: forest recruitment on grass cells due to long-distance seed dispersal ($G\to F$ at rate $\beta$), forest mortality ($F\to G$ at rate $\gamma$),
fire ignition on grass cells ($G\to B$ at rate $\phi$), and grass regrowth on ash cells
($A\to G$ at rate $\lambda$). The conversions due to spread to neighbours are: forest recruitment due to short-distance seed dispersal on grass or ash cells ($GF\to FF,\;AF\to FF$ at rate $\alpha$), fire spread on grass ($GB\to BB$ at rate $\rho_g$) or on tree cells ($FB\to BB$ at rate $\rho_f$). Chosen parameters are in the ranges empirically justified by
\cite{Hebert-Dufresne2018} for a square domain of $100$x$100$ cells, with cell size $\Delta x=\Delta y=30$m. The time scale separation between fire and forest dynamics implies
that the rates satisfy $\rho_g,\rho_f,\mu,\lambda\gg\alpha,\beta,\gamma$. In
particular, we choose
\begin{align}
\rho_g,\mu\sim10^6>\rho_f\sim10^5&\gg1\text{y}^{-1},\label{eq:fast-timescale1}\\
\alpha,\beta,\gamma\sim10^{[-4,-2]}&\ll\lambda\sim1\text{y}^{-1}.\label{eq:fast-timescale2}
\end{align}
So, fire spreading and extinction $\rho_g$, $\rho_f$, $\mu$ occurs on the scale of hours, while grass regrows on ash over months ($\lambda$) and forest spread, growth and mortality $\alpha$, $\beta$, $\gamma$ occurs over decades.
We take fire ignition rate $\phi\sim1/N$ such that fires spontaneously occur about once per year in the modelled area.
\Cref{fig:FGBAplots}b--d shows a typical time profile
for fractions of cells in each state. The initial state is all grass and the ignition rate $\phi$ is low, such that nearly full forest cover is the only steady state (reached after $300$ years, \cref{fig:FGBAplots}h). Brief events of rapidly spreading fire counteract a gradual spread of forest. Time scale separation of forest dynamics (\cref{fig:FGBAplots}b), grass regrowth
(\cref{fig:FGBAplots}c), and fire spread (\cref{fig:FGBAplots}d)
shows clearly.
%The domain fractions of cells in each state will further
%be written as $[F]$, $[G]$, $[B]$ and $[A]$ (see Methods).
\Cref{fig:intro}a shows a bifurcation diagram of steady state forest cover in the FGBA cellular automaton, denoted by $[F]$ (see \cref{eq:def:xxy}),
versus fire ignition rate $\phi$. Unstable steady states (saddles)
were obtained by applying feedback control to the simulations (see Methods).
\emph{Bistability} occurs
above a critical ignition rate $\phi$. Simulations initiated at the saddle will tip up
or down depending on initial stochasticity (\cref{fig:intro}b). Near the lower end
of the bistability
range, the saddle solution is fairly homogeneous, but for higher $\phi$
values, a single hole of grass in forest arises (insets in \cref{fig:intro}a).
\subsection{Fast and slow subprocesses}
The time scale separation (\cref{eq:fast-timescale1,eq:fast-timescale2}) permits treatment of the joint vegetation and fire dynamics
as a fast-slow system. Fire spread occurs on the fast time scale, where
the vegetation landscape is treated as constant. Forest dynamics occurs on the slow
time scale, where the effects of fire are a steady state function of the vegetation
landscape.
\paragraph{Fast process: fires spreading in a given landscape}
On the time scale of a single fire event, forest dynamics are negligible
($\alpha,\beta,\gamma\ll1/$day)
such that we can consider the total landscape
of forest patches as fixed. For each ignition event, this results in the following
dynamics. A fire ignites on a grass cell, then spreads
across its grassland cluster at a rate $\rho_g$ per $BG$ link, after which it reaches
the interface with adjacent forest, where it starts intruding the forest at a rate
$\rho_f$ per $BT$ link. At any time, burning cells can spontaneously extinguish,
converting to ash at a rate $\mu$. The probabilities of fire spreading into a neighbouring grass or forest cell
before the originating cell extinguishes are given by
\begin{equation}\label{eq:occupation:prob}
p_{g}:=\frac{\rho_g}{\rho_g+\mu}=0.9, \qquad p_{f}:=\frac{\rho_f}{\rho_f+\mu}=0.1,
\end{equation}
where we have shown the chosen values in our simulations (adopted from
\cite{Hebert-Dufresne2018}). Since new fires arise and grass regrows only with much lower rates than the fire spreading rate ($\phi N\lesssim\lambda\ll\rho_f,\mu,\rho_g$),
we observe repeated fire waves well separated in time (\cref{fig:FGBAplots}b).
When a fire in grassland cluster with index $j$ reaches its interface with adjacent forest,
the resulting forest loss due to this single fire event can be approximated as a fraction
$p_f$ of the forest cells at this interface. This leads to the approximate loss of forest per fire event
\begin{equation}\label{eq:forest:loss}
\Delta_{F,j}^{\mathrm{loss}}=p_f[FG]_{j},
\end{equation}
where $[FG]_{j}$ counts the number of forest cells adjacent to grassland cluster $j$
(with both sides of the equation optionally normalised by $N$). This approximation relies on the assumptions that the fire reaches the
whole interface with forest (i.e. $p_g\to1$) and only once per fire (i.e.
$\rho_g\gg\lambda\gg\phi N$), and that burning forest does not
pass on fires further to adjacent forest (i.e. $p_f$ small).
\paragraph{Slow processes: forest demography and fire damage}
Forest demography and loss due to repeated fires occur on the slow time scale.
%The forest-grass landscapes in which fires spread in the GBA process have been shaped by the competition between the (slow) forest spreading process (called GF below) and the cumulative damage due to fire exposure.
Writing the number of neighbour pairs with $FG$ as $[FG]$ (divided by $N$,
equivalently the domain perimeter of forest or grass patches, see \cref{eq:def:xxy}),
%\jsq{It might be clearer to just state ``Writing the number of $FG$ lattice links (scaled by $N$) as $[FG]$'', same with $[FG]_j$ above}
the dynamics for tree recruitment
and mortality result in an expected rate of change for $[F]$
\begin{equation}
\label{eq:slow:gain}
\Delta_{F}^{\mathrm{gain}}:=\beta[G]-\gamma[F]+\alpha[FG].
\end{equation}
In \cref{eq:slow:gain}, the rates of change are $\beta[G]$ for spontaneous forest growth on grass, $\gamma[F]$ for forest mortality, and $\alpha[FG]$ for spread of forest into grass at its perimeter.
%This process (without fire events) is equivalent to the susceptible-infected-susceptible
%(SIS) epidemic spreading \citep[e.g.][]{Kiss2017} (or the contact process \citep{Harris1974,Henkel2008}) with a spontaneous infection term. This process
%models disease spread through a susceptible
%population that cannot acquire immunity. The steady-state landscapes it generates without interacting with fire are different from the landscapes shaped by fire spread, see \cref{fig:GF} for a systematic study without fire spread.
The rate of forest erosion at its perimeter due to fire damage over many fire events
is the weighted sum over all grass clusters $j=1,...,n_c$, i.e.
\begin{align}
\Delta_{F}^{\mathrm{loss}}&:=\sum_{j=1}^{n_c}\phi N[G]_j\Delta_{F,j}^{\mathrm{loss}}\nonumber\\
&=\phi N p_f\sum_{j=1}^{n_c}[G]_j[FG]_{j},\label{eq:slow:loss}
\end{align}
where $[G]_j$ is the fraction of $G$ cells in grass cluster
$j$ (so, $[G]=\sum_{j=1}^{n_c}[G]_j$), $\phi N[G]_j$ is the rate at which fires spontaneously ignite in grass cluster $j$ ($\phi$ is the rate per cell and $N[G]_j$ is the area of the cluster), and $\Delta_{F,j}^{\mathrm{loss}}=p_f[FG]_j$ is the conversion of forest to ash caused by each fire event (see \cref{eq:forest:loss}) (note also that $[FG]=\sum_{j=1}^{n_c}[FG]_j$). By defining the \emph{grassland-weighted forest perimeter} as
\begin{equation}\label{eq:defFGavcg}
\fgg:=\sum_{j}^{n_c}\frac{[G]_j}{[G]} [FG]_j,
\end{equation}
\cref{eq:slow:loss} becomes
\begin{equation}
\label{eq:slow:loss2}
\Delta_{F}^{\mathrm{loss}}:=\phi p_f N[G]\fgg.
\end{equation}
The grassland-weighted forest perimeter $\fgg$ is the average perimeter
of forest clusters weighted by the relative size of their adjacent grass cluster.
\subsection{Emergent slow dynamics}
\begin{figure}[t]
\centering
\includesvg[inkscapelatex=false,width=1\columnwidth]{./figures/expectedchanges}
\caption{Rate of change according to \cref{eq:dFdtclust}. (a,b) Spatial snapshots at indicated times, (c,d) time series of $[F]$, (e,f) time series of $\Delta_F^{\mathrm{gain}}$, $\Delta_F^{\mathrm{loss}}$, (g,h) time series of the right-hand side of \cref{eq:dFdtclust} (gain minus loss). At $t{=}0$, the simulation is started on the saddle (on the left, $[F](0)\approx0.2$ and on the right, $[F](0)\approx0.7$, see blue and red circles in \cref{fig:intro}). Towards the left (along $t^+$), a simulation that tips up and towards the right (along $t^-$) a simulation that tips down is shown. Parameters are shown in \cref{tab:rates}. Columns correspond to leftmost and rightmost vertical dashed lines in \cref{fig:intro} ($\phi N=0.257$ and $\phi N=1.32$). Domain size: 100x100 cells.}
\label{fig:gainlossterms}
\end{figure}
We now form the balance between the slow processes discussed above,
assuming fire converts trees immediately to grass (i.e. $\lambda\gg\phi N$).
%neglecting interaction of forest spread into ash at rate $\alpha$ (as grass regrows much faster on ash $\lambda\gg\alpha$).
The resulting expected rate of forest cover change during a short time interval is
\begin{align}
\langle\frac{\mathrm{d}[F]}{\mathrm{d} t}\rangle =& \Delta_{F}^{\mathrm{gain}}-\Delta_{F}^{\mathrm{loss}},\nonumber\\
\frac{\mathrm{d}[F]}{\mathrm{d}t}=&\beta [G] - \gamma [F] + \alpha [FG] - \phi p_f N[G]\fgg,\label{eq:dFdtclust}
\end{align}
where we used \cref{eq:slow:gain,eq:slow:loss2}, and assumed on the left-hand side that $N$
is sufficiently large, such that, via the law of large numbers,
$\langle \mathrm{d}[F]/\mathrm{d}t\rangle\approx \mathrm{d}[F]/\mathrm{d}t$.
\Cref{eq:dFdtclust} can be understood intuitively as
forest and grass competing for space within patches (spontaneous terms)
and at their interface (interaction terms). A larger interface $[FG]$ leads
simultaneously to faster forest spread (proportional to its perimeter $[FG]$), and to increased exposure to fires (proportional to its grassland-weighted perimeter $\fgg$). Fires are most
damaging to forest when $[G]$ forms a single cluster, i.e. $\fgg=[FG]$,
such that each fire reaches the whole interface. Conversely, when forest patches break $[G]$
into several clusters $\fgg$ is smaller than $[FG]$, such that several ignitions are required to have
the same effect, slowing forest erosion down. Additionally, the total amount of grass $N[G]$ determines the number of ignitions
and hence the rate at which grass spreads into forest. The parameters determine the relative weight
of each of the discussed effects.
\Cref{fig:gainlossterms}
shows example simulations along trajectories starting from the saddle equilibria of \cref{fig:intro}, showing
forest area $[F]$ in space and time (a--d), the gain/loss terms $\Delta_F^{\mathrm{gain}}$ and $\Delta_F^{\mathrm{loss}}$ defined in \cref{eq:slow:gain,eq:slow:loss2} (e--f), and the right-hand side of \cref{eq:dFdtclust}
(gain minus loss, g--h). The left column of \cref{fig:gainlossterms} shows simulations for
low fire ignition rate $\phi$ and low $[F](0)$, and the right column for high $\phi$ and high $[F](0)$. Each
column shows two realisations, both starting from the same saddle steady state. One
realisation evolves toward high forest cover,
shown on axis $t^+$ (increasing to the left from $t=0$), the other realisation evolves toward low forest cover,
shown on axis $t^-$ (increasing to the right from $t=0$). In the stable steady states, gain (green) and loss (red)
terms vary around the same mean. On the saddle (at time $0$)
gain and loss functions cross. This shows that the steady states
and changes are accurately captured by \cref{eq:dFdtclust}. The largest changes in forest cover $[F]$ occur when there are large changes in forest loss due to fire.
The snapshots in
\cref{fig:gainlossterms}a,b show that the high-cover state changes as
an expanding/contracting hole in the forest, whereas the low-cover state
is more homogeneous.
%Invoking potential wells here is probably unnecessary and possibly confusing
%The snapshots in
%\cref{fig:gainlossterms}a,b of the landscape show that \js{for large fire ignition rate $\phi$} a noise-induced ``climb out of the potential well'' with high forest fraction $[F]$ is
%an expanding hole in the forest. Correspondingly a drop into the potential well is a shrinking hole in the forest, and the saddle corresponds to a hole of a critical size. For smaller $\phi$ the landscape is more homogeneous and has less forest cover at the saddle.
\subsection{Emergent nonlinear relations}
\begin{figure}[t]
\centering
\includesvg[inkscapelatex=false,width=1.\columnwidth]{./figures/percanalysis13scatter3phi}
\caption{Emergent relations between key quantities and forest area $[F]$: (a--c) forest perimeter $[FG]$ and grassland-weighted forest perimeter $\fgg$, (d--f) forest gain terms and loss terms in \cref{eq:slow:gain,eq:slow:loss}, (g--i) forest area rate of change $(\mathrm{d}/\mathrm{d}t)[F]$ from \cref{eq:dFdtclust}. Columns correspond to vertical dashed lines in \cref{fig:intro} ($\phi N{=}0.257,\phi N{=}0.38,\phi N{=}1.32$). Domain size: 100x100 cells.}
\label{fig:scatterplots_changes}
\end{figure}
\Cref{eq:dFdtclust} explains how the rate of change of $[F]$ depends on the perimeter quantities $[FG]$ and $\fgg$.
%, both measuring spatial clustering and correlations of the forest-grass interface.
\Cref{fig:scatterplots_changes}a--c
shows a scatterplot of $[FG](t)$ and $\fgg(t)$ versus $[F](t)$ for three
different values of $\phi$, and for an ensemble of simulations starting from the saddle in \cref{fig:intro}a, with each point being a value
observed at a discrete observation time. Remarkably, we observe that
$[FG]$ and $\fgg$ lie on a narrow band around some steady state
functions $[FG]^*$ and $\fgg^*$ of $[F]$ (and $\phi$), which
implies that $[FG],\langle[FG]\rangle_{c_g}$ are changing on
a much faster time scale, making them slaved to $[F]$.
%I didn't find this clearer
%that the $FG$-interface and the grass cluster distribution as measured
%by $[FG]$ and $\fgg$ are on the slow time scale determined by the forest area fraction $[F]$ (also known as $[FG]$ and $\fgg$ being ``slaved'' by $[F]$).
\Cref{fig:scatterplots_changes}d--f shows the terms on the right-hand side of
\cref{eq:dFdtclust} depending on $[F]$, splitting between gain and loss
terms $\Delta_F^\mathrm{gain},\Delta_F^\mathrm{loss}$, as defined in \cref{eq:slow:gain,eq:slow:loss2}. Steady states occur when
gain equals loss ($\Delta_F^\mathrm{gain}{=}\Delta_F^\mathrm{loss}$). The resulting graph of $\mathrm{d}[F]/\mathrm{d}t$
versus $[F]$ in \cref{fig:scatterplots_changes}g--i clearly shows the bistability of
$[F]$.
Replacing the quantities $[FG]$ and $\fgg$ by their steady-state functions $[FG]^*$ and $\fgg^*$ results in the single-variable ODE for $[F]$,
\begin{equation}\label{eq:dFdtclust2ss}
\frac{\mathrm{d}[F]}{\mathrm{d}t} = \beta[G] - \gamma [F] + \alpha [FG]^* - \phi p_f N [G] \fgg^*,
\end{equation}
where $[FG]^*,\fgg^*$ are functions of $[F]$ and $\phi$ (as shown in
\cref{fig:scatterplots_changes}a-c), and $[G]=1-[F]$. With these functions $[FG]^*$ and
$\fgg^*$, the observed bistability is caused by a
classic scalar double-well potential of the gradient system
\cref{eq:dFdtclust2ss}.
In this ODE, nonlinearities emerge due to the
equilibrium dependence of the interface on forest area (affecting $[FG]^*$
and $\fgg^*$),
due to the segmentation of grass cells near and below the percolation threshold
(affecting $\fgg^*$) and due to dependence of the ignition
rate on grass patch size (multiplying $\fgg^*$ with $[G]$).
In \cref{fig:singlemultipleclust_vs_ctrl},
we show the roots of \cref{eq:dFdtclust2ss}
using a nonparameteric fit of $[FG]^*([F];\phi)$
and $\langle[FG]\rangle_{c_g}^*([F];\phi)$. These match well with
the steady states obtained via control (dot-dashed red).
% \paragraph{Steady state interface and cluster distribution} So far, we only
% substituted values obtained from the simulations into the right-hand side of
% \cref{eq:dFdtclust}. To gain general insight in the typical behaviour
% and how it emerges from the right-hand-side terms, we would
% like to solve \cref{eq:dFdtclust}. Therefore, it needs to be determined how
% $[FG],\langle[FG]\rangle_{c_g}$ or their change in time depend on existing variables,
% such that a closed system of equations can be formed.
% However, we show here that $[FG],\langle[FG]\rangle_{c_g}$ are changing on
% a much faster time scale, which makes them slaved to $[F]$, such that
% all we need to know is how their steady states $[FG]^*$ and
% $\langle[FG]\rangle_{c_g}^*$ depend on $[F]$. \Cref{fig:scatterplots_changes}a-c
% shows a scatterplot of $[FG](t),\langle[FG]\rangle_{c_g}(t)$ versus $[F](t)$ for three
% different values of $\phi$,
% obtained from the simulations starting at the saddle, with each point being a value
% observed at a discrete observation time. That
% $[FG]$ and $\langle[FG]\rangle_{c_g}$ lie on a narrow band around some steady state
% functions $[FG]^*$ and $\langle[FG]\rangle_{c_g}^*$ of $[F]$ (and $\phi$) confirms
% that their time-scale separation with $[F]$ is large, such that the FG-interface and the
% grass cluster distribution (via \cref{eq:defFGavcg}) are approximately determined by
% $[F]$.
% \Cref{fig:scatterplots_changes}d-i shows the right-hand
% side of \cref{eq:dFdtclust} as a function of $[F]$,
% with panels d-f splitting between gain
% loss terms (as defined above). Steady states occur when gain equals loss.
% The resulting plot of $d[F]/dt$ versus $[F]$
% clearly shows the bistability of $[F]$.
If there is only one connected
component of grass cells, we have $\fgg=[FG]$,
such that \cref{eq:dFdtclust2ss} simplifies to
\begin{equation}\label{eq:dFdtclust1ss}
\frac{\mathrm{d}[F]}{\mathrm{d}t} = \beta[G] - \gamma [F] + (\alpha - \phi p_f N [G]) [FG]^*.
\end{equation}
For homogeneous initial conditions (i.e. $[F]$ is about the same in different
large subsections of the domain), this approximation is expected to be valid for small $[F]$, where
most grass cells belong to the giant component.
\Cref{fig:singlemultipleclust_vs_ctrl} shows
the resulting steady states of \cref{eq:dFdtclust1ss} as a function of
$\phi$ and $[F]$ when only using the fit of $[FG]^*([F];\phi)$ (dashed blue).
The approximation is good for landscapes with low forest cover ($[F]\lesssim0.2$). Above $[F]\approx0.2$, it fails because the
grassland breaks up into multiple clusters and fires are smaller than in case of a single
cluster. \Cref{fig:scatterplots_changes}a--c already indicated that the single-cluster approximation is accurate for low forest cover since
$\fgg\approx[FG]$ for low $[F]$ in the scatterplots.
\section{Discussion}
Using as high-dimensional reference model a
continuous-time stochastic cellular automaton, adapted from
ref. \cite{Hebert-Dufresne2018}, that includes reaction rules for both forest dynamics
and fire percolation, we have shown how low-dimensional nonlinear dynamics of tropical forest emerge from the spatial configuration of forest patches.
Taking the assumption on the fast time scale that fire spreads perfectly on grass,
reaching the whole perimeter of adjacent forest,
makes two macroscopic structural quantities relevant for the
evolution of forest area on the slow time scale (\cref{eq:dFdtclust}): the forest perimeter $[FG]$ for gain from forest spread, and, the grassland-weighted forest perimeter $\fgg$ for loss induced by fire. Simulations initiated at the unstable steady states allowed us to explore the whole
phase space and revealed that these quantities are evolving on
a much faster time scale than forest area, such that on the slow time scale and for
given parameters, they are equilibrium functions of forest area. Hence,
replacing the perimeter quantities with their
steady state functions of forest area reduces the evolution
of forest area on the slow time scale to a one-variable ODE,
\cref{eq:dFdtclust2ss}.
In this emergent relation between forest change and forest area,
nonlinear terms arise due to spatially structured interactions at the forest-grass
interface whereas linear terms arise from spontaneous transitions, which are independent of spatial structure.
In the rest of our discussion, we will compare our result to what is obtainable by mean-field models, illustrate the role of finite-size effects in bistability, and consider the implications for empirical tests.
\paragraph{Inadequacy of mean-field models}
Mean-field models fail to capture the role played by forest perimeter
and grassland clusters because they neglect spatial
structure (by the well-mixedness assumption, point (\emph{i}) in the introduction). The mean-field approximation for the FGBA process is given in \cref{sec:sup:MF} (Supporting Information).
Mean-field model \cref{eq:MFeqs} assumes absence of pairwise correlations, which permits substitutions such as
$[FG]=4[F][G]$ in \cref{eq:MFeqs1} and closure of the system of equations. Such a substitution of
counts of larger units by an algebraic expression in
terms of counts of smaller units in order to close a system of equations is
known as a \emph{closure formula} \cite{Wuyts2021}.
Neglect of structure leads to a well-known
bias in the location of steady states compared to the simulations \cite[e.g.][]{Keeling1999}. This may be mitigated by mean-field models
of higher order, which consider units of larger size than single cells,
but for dynamics on the square lattice in regimes with long-range dependence
(such as induced by fire or forest spreading) this would require an impractically
large number of equations \cite{Wuyts2021}, defeating its purpose of low-dimensional
approximation.
%\js{Revisiting our analysis of the fast subprocess (fire spreading) makes clear why this bias will persist also for closure formulas and mean-field models of higher order. Since fire spreading on grass is nearly perfect, all cells in a grass cluster contribute to the fire damage rate at its forest perimeter on the slow time scale. This long-range spatial interaction leads to long-range correlations in the resulting forest-grass landscape, which is fundamentally beyond the reach of mean-field models.}
By exploiting time scale separation of fire and forest
dynamics, we arrived at \cref{eq:dFdtclust2ss}, which respects the spatially structured nature of the dynamics, in contrast to mean-field approximations.
In our description, the steady state functions $[FG]^*$ and $\fgg^*$ of $[F]$
are closure formulas to obtain \cref{eq:dFdtclust2ss} from \cref{eq:dFdtclust}.
Here the closure is observed emerging in simulations instead of derived from an a priori
imposed assumption of
statistical independence (see also \cite{Gross2008}). Closure at low dimension is
feasible due to time-scale separation between forest perimeter quantities and
forest area.
\paragraph{Finite sizes and bistability} In simple bistable chemical systems, it is known
that bistability converges to an abrupt phase transition
in the thermodynamic limit ($N{\to}\infty$) \cite{Ge2009,Thiele2019}, at a value known as the Maxwell point \cite[e.g.][]{Goldenfeld2018}, making the macroscopic state of the system
deterministically dependent on the parameters (e.g. pressure or temperature). With only
forest and grass or provided that savanna and forest components are
sufficiently decoupled, the same behaviour occurs in spatial mean-field
models of tropical forest, with a front between forest and nonforest that is
deterministically dependent on environmental drivers \cite{Wuyts2019}.
We do not expect such determinism as $N{\to}\infty$ to arise in the FGBA cellular
automaton.
The infinite FGBA cellular automaton possesses grass clusters of arbitrary size,
such that, even when assuming that fire spreads instantly on grass, there will
always be some parts of the forest perimeter shielded from intruding fires by
adjacent ash cells. Were it not for this shielding effect, then there would be
a deterministic dependence of the dynamics on fire ignition rate away from the absorbing states: $\phi{=}0$ would lead to forest spread while $\phi{>}0$ would lead
to forest extinction (see \cref{sec:supp:fronts}, for the case $\beta{=}\gamma{=}0$).
In reality, finite fire spreading rates and, in particular,
the effects of heterogeneity in time
(e.g. the influence of weather on spreading rates) or space
impose stronger limits on the reach of fires.
%We can
%illustrate that in case of $\beta=\gamma=0$, such determinism arises as well in the
%infinite-size FGBA cellular automaton.
%As shown in \cref{sec:supp:fronts}, for $\beta=\gamma=0$, a domain initially separated
%into two contiguous grass and forest patches will have its dynamics determined entirely by how a forest-grass interface moves depending on the size of the adjacent grass cluster.
%Namely, for a sufficiently large grass cluster, the forest will recede and the grass cluster will expand, while the reverse occurs if the grass cluster is small.
%Now we consider an infinite domain with many grass clusters of which the size is a random
%variable (with support $[0,\infty)$). Here, there will be initial grass clusters of arbitrarily large size, which will expand and eventually drive forest extinct.
%Thus, we expect a deterministic dependence of the dynamics on fire ignition rate away from the absorbing states for $N\to\infty$: $\phi=0$ leads to spread of forest while $\phi>0$ leads to extinction of forest.
Hence, while, due to the
large size of a typical system of molecules, the thermodynamic
limit is usually of practical relevance in context of materials,
%where the concept of Maxwell point first emerged \cite[e.g.][]{Goldenfeld2018},
in case of ecosystems,
\emph{finite sizes} and their effects are the norm. In finite
domains,
%(with fire return rate $\phi N$ for the area of order $1$)
both the FGBA system and scalar reaction-diffusion equations show bistability
due to critical patch
sizes and dependence on interfacial characteristics \cite[e.g.][]{Liehr2013,Goel2020}, but this correspondence requires further scrutiny.
In realistic scenarios, the amount of bistability then depends (besides the parameters) on a `characteristic system
size', which in turn depends on e.g. fire spreading and/or plant dispersal ranges.
%(and fire return rate $\phi N$) <- already included in 'besides parameters'
How this is affected by environmental heterogeneity is subject of future work.
\paragraph{Implications for empirical tests}
By taking the simplest possible model that
was able to include fire percolation and forest spread, we were able to obtain
structural indicators of bistable forests prone to fire percolation effects. These
indicators are theoretical predictions additional to bimodality that can be tested
against spatial data.
%(for the first simple test, disregarding spatial structure and heterogeneity, see
%\cite{VanNes2018}).
We emphasise here that the perimeter quantities $[FG]$ and $\fgg$ do not have to be in equilibrium
for \cref{eq:dFdtclust} to be valid, such that it can also be tested in places
where human impact keeps them away from their natural equilibrium.
Furthermore, they are in principle easy to determine from observations. In the original treatment of the FGBA
cellular automaton \cite{Hebert-Dufresne2018}, a climate-dependent equilibrium
perimeter-area relation for \emph{isolated} forest patches
in low-cover simulations was found to be consistent with data of protected areas
in the Brazilian Cerrado. The perimeter-area relation at the patch level for
given parameters determines the perimeter-area relation at domain level
$[FG]^*$ versus $[F]$. Therefore, we
expect this consistency with data to hold at the domain level, at
least for low cover areas.
Inclusion of additional realism, such as
environmental heterogeneity, larger neighbourhoods, non-lattice geometry,
presence and dynamics
of other tree types and life stages, will lead to modified perimeter quantities, but these can be derived and tested in a similar manner. We hope that the additional opportunities that our theoretical predictions
offer for falsification
lead to a faster rate of progress in the field. If this progress leads to theory
with sufficient reliability and validity, it may eventually function as a tool
for conservation of existing or restoration of lost tropical forests.
\section{Methods}
\renewcommand{\theequation}{M\arabic{equation}}
\renewcommand{\thefigure}{M\arabic{figure}}
\renewcommand{\thetable}{M\arabic{table}}
\setcounter{equation}{0} % reset counter
\setcounter{section}{0} % reset counter
\setcounter{figure}{0} % reset counter
\paragraph{Details of the FGBA probabilistic cellular automaton}
The FGBA probabilistic cellular automaton is a minimal spatial
stochastic process that models the joint dynamics of tropical vegetation
and fire. It is an adapted
version of the BGT(A) model of ref. \citep{Hebert-Dufresne2018}.
The modifications compared to ref. \citep{Hebert-Dufresne2018} are:
(\textit{i}) it runs in continuous
time, (\textit{ii}) it includes a spontaneous forest mortality rate $\gamma$, (\textit{iii}) it optionally includes rainfall feedback (for \cref{fig:FGBA_phik})
(\textit{iv}) species $T$ is labelled as $F$,
consistent with other models of tropical vegetation dynamics \citep{Staver2012,Wuyts2019,Patterson2021}. Note that according to some definitions,
probabilistic continuous-time cellular automata are considered as interacting particle
systems. In general, when studying
the stochastic dynamics of a number $n$ of interacting species on
a square lattice with $N$ cells, the state of the system can be represented
as
\[
X:=(X_{1},X_{2},...,X_{N}),
\]
where $X_{i}$ is the label of the species that occupies cell $i$.
Each cell is occupied by exactly one of four possible species: grass, forest, burning and ash, with labels
$G$, $F$, $B$ and $A$. Transitions between states (species) occur in continuous time, resulting in a continuous-time Markov chain with $n^{N}$ possible states. The reaction rules for transitions between states are shown in \cref{tab:rates},
where spontaneous conversions are shown on the left and conversions
due to nearest-neighbour interactions on the right (see also
\cref{fig:FGBAplots}a for the state transition diagram).
The latter type of
interaction occurs over each four nearest neighbour connections
of the indicated type. E.g., fire will spread into a given grass cell with
a rate $\rho_g$ for each burning neighbour. For realistic time scales,
our parameters satisfy \cref{eq:fast-timescale1,eq:fast-timescale2},
which were empirically justified in \cite{Hebert-Dufresne2018}.
This results in separated time scales for forest spread (slow) and
fire spread (fast). We borrow our notation from the moment closure
literature \cite[e.g.][]{Keeling1999,Kiss2017,Wuyts2021}, writing
the global fraction of species with label $x$ and the interface
between species with label $x$ with label $y$ respectively as
\begin{equation}\label{eq:def:xxy}
[x]:=\frac{1}{N}\sum_{i}^{N}\delta_x(X_{i}),
\quad
[xy]:=\frac{1}{N}\sum_{i,j}^{N}\mathsf{A}_{ij}\delta_x(X_{i})\delta_y(X_{j}),
\end{equation}
where both are normalised by $N$, $\delta$ is the Kronecker delta function ($\delta_x(y){=}1$ if $x{=}y$ and $0$ otherwise) and $\mathsf{A}\in\{0,1\}^{N\times N}$
the adjacency matrix. We simulated the cellular automaton via a Gillespie algorithm
\citep{Gillespie2007} and used a domain
of $N{=}100{\times}100$ ($N{=}200{\times}200$ in \cref{fig:FGBAplots}) cells with periodic boundary conditions.
\begin{table}
\setlength{\tabcolsep}{3.5pt}
\centering
\caption{Reactions and rates (y$^{-1}$) in the FGBA cellular automaton.\label{tab:rates}}
\begin{tabular}{|cl|cl|}
\hline
\multicolumn{2}{|c|}{Spontaneous}&\multicolumn{2}{c|}{Spread}\\
\hline
$G\overset{\beta}{\to}F,$ & $\beta=2\cdot10^{-4}$ &
$GF\overset{\alpha}{\to}FF,$ & $\alpha_{\;}=3\cdot10^{-2}$\\
$F\overset{\gamma}{\to}G,$&$\gamma=2\cdot10^{-2}$&$AF\overset{\alpha}{\rightarrow}FF,$ & $\alpha_{\;}=3\cdot10^{-2}$\\
$G\overset{\phi}{\to}B,$ & $\phi=[0,2]\cdot10^{-4}$ &
$GB\overset{\rho_g}{\to}BB,$ & $\rho_g=9\cdot10^6$\\
$B\overset{\mu}{\to}A,$ & $\mu=10^6$ & $FB\overset{\rho_f}{\rightarrow}BB,$&$\rho_f=1.11\cdot10^5$\\
$A\overset{\lambda}{\to}G,$ & $\lambda=5$& &\\
\hline
\end{tabular}
\end{table}
\paragraph{Noninvasive feedback control}
To study steady states regardless of their stability in a simulation, we apply noninvasive feedback control, similar to what was used
in \citep{SGNWK08,schilder2015experimental,barton2017control,neville2018shape} for
tracking unstable
vibrations and buckling states in mechanical experiments.
To study the dependence of equilibria of $[x]$ on one of the conversion rates $r$,
we treat the simulation like an ideal physical
experiment and introduce an artificial stabilizing feedback loop of the form
\begin{equation}
r(t)=r_{0}+g([x](t)-[x]_{\mathrm{ref}})\mbox{.}\label{eq:control}
\end{equation}
In \cref{eq:control}, $r$ is one of the conversion rates of the lattice model.
The artificial feedback makes this rate time dependent by coupling it to the species
occurrence fraction $[x](t)$ through the relation (\ref{eq:control}).
The factor $g$ is called the feedback control gain and is problem
specific. The property of noninvasiveness means that the controlled
simulations have the same equilibria as
regular simulations \citep{BS13,SOW14,renson2017experimental}. This implies
that if one extracts the equilibrium values of the controlled simulation $(r^*,[x]^*)$,
one can use them to plot a 1-parameter bifurcation diagram of the stochastic
lattice model.
%If each reaction is labelled by the elements
%of a vector $\boldsymbol{e}$, $E(t)$ is a random variable returning the
%event labels at the time of their occurrence, and $\mathbf{N}$ is a
%vector with the size of the population on which the corresponding
%events in $\boldsymbol{e}$ act, then the effective rate of conversion
%according to the reaction with label $e_k$ is
%\begin{equation}\label{eq:effrate}
% \bar{r}_k=\frac{\int_{t_0}^{t_0+T}\delta_D(t-s)\delta_K[E(s),e_k]ds}{ \int_{t_0}^{t_0+T}N_k(t)dt},
%\end{equation}
%where $t_0$ is the time after which the dynamics have settled to a steady
%state, $T$ the averaging time, $\delta_D,\delta_K$ are the Dirac and
%Kronecker delta functions.
\begin{figure}[t]
\centering
\includesvg[inkscapelatex=false,width=1.\columnwidth]{./figures/ctrlrelease40case1}
%\includegraphics[width=\columnwidth]{./figures/placeholder2}
\caption{Feedback control applied to the cellular automaton without spontaneous conversions ($\beta{=}\gamma{=}0$): (a) the unstable steady state of the bifurcation diagram (dashed) was derived via feedback control by letting $\phi$ be a function of $[F]$ (blue line) such that it is stabilised, then obtaining $(\phi,[F])$ by averaging and repeating this for many $[F]_\mathrm{ref}$ (with appropriate $g$), (b) a regular simulation with the same $\phi$ value (solid grey in (a)) and starting from the final state of the controlled simulation tips up or down depending on initial perturbations, (c) snapshot of the domain at the saddle for the control indicated in (a) (black: forest, white: grass). Other parameter values are shown in \cref{tab:rates}.}
\label{fig:ctrlFGBA}
\end{figure}
\Cref{fig:ctrlFGBA} shows
an example for which the controlled rate is the spontaneous conversion
rate $\phi$ for reaction $G\to B$ and the output variable $[x]$ is total forest fraction $[F]$. The
feedback \cref{eq:control}, indicated in blue in \cref{fig:ctrlFGBA}a,
stabilises a steady state that is unstable in a regular simulation. This
can be seen in \cref{fig:ctrlFGBA}b, where the unstable steady state is first
stabilised via control, after which the control is removed and a regular
simulation is started with the effective rate (and initial conditions)
obtained from the controlled
simulation. Depending on initial perturbations, the regular simulation gets
either attracted to the 100\% forest state or to the low tree cover state.
When the controlled simulation is in equilibrium (steady part of the
blue curve in \cref{fig:ctrlFGBA}b), the steady state
values of $[x]$ for $x\in\{F,G,B,A\}$ are obtained via taking the time average, i.e.
\begin{equation}\label{eq:effout}
\overline{[x]}=\frac{1}{T}\int_{t_0}^{t_0+T}[x](t)dt,
\end{equation}
where $t_0$ is the time after which the dynamics have settled to a steady
state and $T$ the averaging time.
If $n_{G\to B}$ is the number of ignition events between $t=t_0$ and $t=t_0+T$,
the steady state of $r$ is obtained by calculating the mean ignition
rate as $n_{G\to B}/T$ and dividing this by the mean number of grass
cells, such that
\begin{equation}
\bar{\phi}=\frac{n_{G\to B}/T}{\overline{[G]}},
\end{equation}
where $\overline{[G]}$ is obtained as in \cref{eq:effout}.
When repeating this exercise for many $[F]_\mathrm{ref}$ values (with
for each an
appropriately chosen gain $g$), one can get multiple points on the
unstable branch. Points on the stable branches can be obtained with
regular simulations. On the final selection of points, we applied
Gaussian process regression to obtain smooth curves and used
moving block bootstrapping \citep{Kreiss2012} to obtain confidence
intervals.
One of the advantages of applying control is that one can obtain
states for which one would have to wait prohibitively long in a
regular simulation due to their instability. The inset in \cref{fig:ctrlFGBA}b
shows for instance a snapshot of the unstable steady state
obtained with the control indicated in \cref{fig:ctrlFGBA}a.
\section*{Acknowledgements}
This work was supported by the UK Engineering and Physical Sciences Research Council (EPSRC) grants EP/N023544/1 and EP/V04687X/1. This project has received funding from the European Union’s Horizon 2020
research and innovation program under grant agreement 820970.
\bibliography{./lib}
%\appendix
\onecolumngrid
\newpage
\renewcommand{\theequation}{S\arabic{equation}}
\renewcommand{\thefigure}{S\arabic{figure}}
\renewcommand{\thesection}{S\arabic{section}}
\setcounter{equation}{0} % reset counter
\setcounter{section}{0} % reset counter
\setcounter{figure}{0} % reset counter
\begin{center}
\fontsize{20}{24}\bfseries\scshape{Supplement}
\end{center}
%\twocolumn
\section{Mean-field equations of the FGBA process on the square lattice}\label{sec:sup:MF}
Here we show how to obtain
the mean-field equations. Note that ref. \citep{Hebert-Dufresne2018}
already analysed the dynamics of the mean-field model (\cref{eq:MFeqs}) with $\gamma=k=0$. Writing $x_i$ as shorthand for $\delta_x(X_{i})$ (equalling $1$ if $X_i=x$ and $0$ otherwise) and taking expectations in each cell $i$,
we obtain equations for the rate of change of the expectation that cell $i$ is occupied by species $x\in\{G,F,B,A\}$,
\begin{align}\label{eq:MFeqs0}
\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle G_{i}\bigr\rangle &= \lambda\bigl\langle A_{i}\bigr\rangle-\bigl\langle G_{i}\Bigl[\beta+\phi+\sum_{j\in{\cal N}(i)}(\rho_gB_{j}+\alpha F_{j})\Bigr]\bigr\rangle+\gamma\bigl\langle F_{i}\bigr\rangle,\nonumber \\
\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle F_{i}\bigr\rangle &= \beta\bigl\langle G_{i}\bigr\rangle+\alpha\bigl\langle(A_{i}+G_{i})\sum_{j\in{\cal N}(i)}F_{j}\bigr\rangle-\rho_f\bigl\langle F_{i}\sum_{j\in{\cal N}(i)}B_{j}\bigr\rangle-\gamma\bigl\langle F_{i}\bigr\rangle,\\
\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle B_{i}\bigr\rangle &= \phi\bigl\langle G_{i}\bigr\rangle-\mu\bigl\langle B_{i}\bigr\rangle+\bigl\langle(\rho_fF_{i}+\rho_gG_{i})\sum_{j\in{\cal N}(i)}B_{j}\bigr\rangle,\nonumber \\
\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle A_{i}\bigr\rangle &= -\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle G_{i}\bigr\rangle-\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle F_{i}\bigr\rangle-\frac{\mathrm{d}}{\mathrm{d}t}\bigl\langle B_{i}\bigr\rangle=\mu\bigl\langle B_{i}\bigr\rangle-\lambda\bigl\langle A_{i}\bigr\rangle-\alpha\bigl\langle A_{i}\sum_{j\in{\cal N}(i)}F_{j}\bigr\rangle.\nonumber
\end{align}
One may think of $\langle\cdot\rangle$ as ensemble averages here. This set of
equations can be derived rigorously from the master equation \cite[e.g.][]{Tome2015}.
Summing
over $i$, dividing by $N$ and using \cref{eq:def:xxy}, we obtain equations for total
frequencies of all four species in the domain (relative to $N$):
\begin{align}\label{eq:MFeqs1}
\frac{\mathrm{d}}{\mathrm{d}t}\langle [G]\rangle &= \lambda\langle [A]\rangle-(\beta+\phi)\langle[G]\rangle- \rho_g \langle[GB]\rangle-\alpha \langle[FG]\rangle+\gamma\langle [F]\rangle,\nonumber \\
\frac{\mathrm{d}}{\mathrm{d}t}\langle [F]\rangle &= \beta\langle [G]\rangle+\alpha\langle[FA]\rangle+\alpha\langle[FG]\rangle-\rho_f\langle [FB]\rangle-\gamma\langle [F]\rangle,\\
\frac{\mathrm{d}}{\mathrm{d}t}\langle [B]\rangle &= \phi\langle [G]\rangle-\mu\langle [B]\rangle+\rho_f \langle[FB]\rangle+\rho_g \langle[GB]\rangle,\nonumber \\
\frac{\mathrm{d}}{\mathrm{d}t}\langle [A]\rangle &= \mu\langle [B]\rangle-\lambda\langle [A]\rangle-\alpha\langle [FA]\rangle,\nonumber
\end{align}
where $[xy]$ is the total number of neighbouring $xy$ pairs divided by $N$, which will
later be referred to as the $xy$ interface or $xy$ perimeter. \Cref{eq:MFeqs1} is not a
closed system. To close the system, we need to determine
all undetermined terms $[xy]$ on the right-hand side without creating
new unknowns. The simplest way to do this is to assume absence of
pairwise correlations, i.e. $\langle [xy]\rangle=4\langle[x]\rangle\langle [y]\rangle$.
We take the additional assumption of $N\to\infty$,
such that the law of large numbers applies and
$[x]\to\langle[ x]\rangle$. Taking these assumptions leads to
the \emph{simple mean-field approximation}
of the cellular automaton model:
\begin{align}
\frac{\mathrm{d}}{\mathrm{d}t}[ G] &= \lambda[A]-(\beta+\phi)[G] -4\rho_g[G][B]-4\alpha[F][G]+\gamma[F],\nonumber \\
\frac{\mathrm{d}}{\mathrm{d}t}[ F] &= \beta[ G]+4\alpha[F][A]+4\alpha[F][G]-4\rho_f[F][B]-\gamma[ F],\label{eq:MFeqs}\\
\frac{\mathrm{d}}{\mathrm{d}t}[ B] &= \phi[ G]-\mu[ B]+4\rho_f[F][B]+4\rho_g[G][B],\nonumber \\
\frac{\mathrm{d}}{\mathrm{d}t}[ A] &=\mu[ B]-\lambda[ A]-4\alpha[F][ A],\nonumber
\end{align}
In the main text we focus exclusively on the fire-vegetation feedback due to
percolation effects but here we also test the effect of a rainfall feedback
that works globally, by taking a linear dependence of forest spreading
rate on total forest area, i.e.
\begin{equation}\label{eq:rainfb}
\alpha([F])=\alpha_{0}+k[F].
\end{equation}
The mean field assumes spatial homogeneity,
well-mixedness and large system size. These assumptions are valid when the
dynamics occur on a large complete graph and interactions occur
at contact rates of order $1/N$.
%More realistic spatial settings have
%more localised structure, e.g. as in case of the square
%lattice, which leads to higher percolation thresholds \citep{Keeling1999}.
%This means that the steady states of \cref{eq:MFeqs} can
%have considerable bias. \jsq{we should demonstrate that in pictures.} Nonetheless, usually the mean-field
%approximation gives a qualitatively
%correct picture of the steady states and bifurcations \citep{Henkel2008,Tauber2012}. Higher-order
%mean-field approximations \citep{Wuyts2021} are more
%accurate but often require considerably more effort to analyse due to
%combinatorial growth of the number of equations
%and slow convergence with approximation order. Mean-field approximations
%designed for spatial problems, either by being spatially explicit
%\citep{Patterson2021} or by use of spatial perturbation
%methods \citep{Cornell2019} have better convergence and allow spatial
%heterogeneity but
%still require assumptions that do not need to be taken in equation-free
%methods, such as feedback control, explained in the next section.
\begin{figure*}[h]
\centering
% \includegraphics[width=1.6\columnwidth]{./figures/FGBA_phik}
\includegraphics[width=0.8\columnwidth]{./figures/FGBA_phik}
\caption{FGBA process (\cref{tab:rates} for simulations and \cref{eq:MFeqs} for mean field): (a) mean-field steady states of forest cover $[F]$ as a function of fire ignition rate $\phi$ and rainfall feedback strength $k$ (blue: steady state manifold, green: contours at fixed $k$ values, red: saddle-node bifurcations), (b) steady states of $[F]$ in the simulations versus $\phi$ (see legend for parameter values, shading: two-standard deviation confidence interval of the mean). The curves in (b) were obtained via a Gaussian process regression on the output of a finite number of simulations (regular for the stable and controlled for the unstable branches).}
\label{fig:FGBA_phik}
\end{figure*}
The FGBA cellular automaton exhibits bistability of tree cover (first shown by
ref. \citep{Hebert-Dufresne2018}) in ranges of all
parameters that are expected to show considerable spatial
heterogeneity in a given ecosystem: $\alpha,\beta,\gamma,\phi$.
The bistability range is widened by the rainfall feedback, as shown
in \cref{fig:FGBA_phik}b (red versus blue curve). Yet, for the mean-field
model at low rainfall feedback gain $k$, bistability can also be decreased (\cref{fig:FGBA_phik}a).
Hence, the effect may not be monotonic. That bistability of tree cover occurs
even without rainfall feedback shows that both fire and
rainfall feedbacks can cause it independently. Note that
the mean field shows significant bias.
\section{Evolution of fronts --- heterogeneous states}\label{sec:supp:fronts}
\begin{figure*}[ht]
\centering
% \includesvg[inkscapelatex=false,width=1.4\columnwidth]{./figures/case1bet0gam0sepic}
%\jsq{Do we have how this looks for $\beta,\gamma$ as in the main simulation? possibly single column?}\bwq{not yet}
\includesvg[inkscapelatex=false,width=0.7\columnwidth]{./figures/case1bet0gam0sepic}
\caption{Perimeter quantities and change rates according to \cref{eq:dFdtclust} when initial conditions are heterogeneous and there are no spontaneous transitions ($\beta{=}\gamma{=}0$) for $\phi{=}6.98\cdot10^{-5}$, displayed as in \cref{fig:gainlossterms,fig:scatterplots_changes}. (a) Snapshots of the domain at indicated times. (b,d,f) Cover, interface and rate of change from \cref{eq:dFdtclust} versus time. (c,e,g) Interface, grass-cluster weighted average of the interface, and rate of change from \cref{eq:dFdtclust} versus cover. Remaining parameters are $\alpha{=}0.03,\rho_{g}{=}10^{10},\rho_{f}{=}1.11\cdot10^{5},\mu{=}10^6,\lambda{=}5$.}
\label{fig:linearinterface}
\end{figure*}
% The steady states we obtained from the simulations
% arise spontaneously when the environment and the initial conditions are homogeneous.
% Under spatial heterogeneity, different
% relations between interface and cluster distribution versus cover will prevail,
% leading in turn to different dynamics and steady states.
Here, we illustrate the case where grass and forest are initially separated into
two contiguous areas with their interface extending along a straight line
%\jsq{why this?$\rightarrow$ and not changing with the dynamics}.% \js{}, without spontaneous conversions between grass and forest ($\beta=\gamma=0$)
Because for this type of initial conditions the single-cluster approximation
(\cref{eq:dFdtclust1ss}) is valid, we can focus on the evolution of the interface.
As spontaneous conversion between forest and grass (with rates $\beta$ and $\gamma$) increases mixing and promotes homogeneity at large scale, we expect the effects of heterogeneous initial conditions to be most persistent when the spontaneous conversion rates $\beta$ and $\gamma$ are small.
Therefore, we will set $\beta=\gamma=0$, for which
\cref{eq:dFdtclust} becomes
\begin{equation}\label{eq:dFdt:nospont}
\frac{\mathrm{d}[F]}{\mathrm{d}t} = (\alpha - \phi p_f N [G]) [FG].
\end{equation}
Hence, the precise shape of the interface $[FG]$ does not affect the location of the steady states, only
the rate at which they are approached or receded from.
The trivial steady states of \cref{eq:dFdt:nospont} (for $[FG]=0$) are $[F]=0$ and $[F]=1$,
which are stable, and between them, there is the (unstable) saddle
\begin{equation}\label{eq:dFdt:nospont:saddle}
[F]^*=1-\frac{\alpha}{\phi N p_f}.
\end{equation}
As seen in \cref{fig:linearinterface_analysis}, this analytical prediction (solid black)
matches the
controlled simulations with $p_g=0.9999$ (shaded blue). For $p_g=0.9$ (shaded red), which
we used before, there is a small bias.
In the limit of $N\to\infty$, \cref{eq:dFdt:nospont:saddle} converges to $[F]^*_\infty=1$,
implying that in an infinite domain, any positive fire rate leads to extinction of forest
below $[F]^*=1$. When ignoring the effect of ash, this would also occur for heterogeneous
initial conditions. That is, considering an infinite domain with many grass clusters of which the size is a random
variable (with support $[0,\infty)$), there will be initial grass clusters of arbitrarily large size, which will expand and eventually drive forest extinct.
However, such determinism does not occur in
the simulations because at high fire rates, patches with ash start to block fires,
and the rate of exposure of the forest interface to fire
becomes limited by the rate at which ash is converted back into grass.
As a first correction for this, one can multiply $p_f$ with the average proportion of
grass sites that are in the ash state after the expected waiting time between fires
$1-\exp(-\lambda/\phi N)$, such that
%\begin{equation}\label{eq:dFdt:nospont:saddle:corr}
%[F]^*=1-\frac{\alpha}{\phi N p_f[1-\exp(-\frac{\lambda}{\phi N})]},
%\end{equation}
%and
$[F]^*_\infty=1-\alpha/\lambda p_f$.
Keeping in mind that we are focusing on heterogeneous states,
the analysis here implies that for $\gamma=\beta=0$, there is a critical
patch size above which the forest patch expands
and below which it contracts. The intuition is that above the critical forest patch size,
there is not enough grass area to reach the minimum number of ignitions
required to erode the forest.
In \cref{fig:linearinterface}, we show how the dynamics and steady states
arise from $[FG]$, as we did in \cref{fig:gainlossterms} and \cref{fig:scatterplots_changes}, but now starting with separated patches of
grass and forest that interface on a line on both sides (showing for $[F](0)=0.57$).
From $[FG](0)=2L/N=2N^{-1/2}=0.02$, the interface
quickly gains roughness due to the dynamics, until it reaches a
steady state of $[FG]^*\approx6L/N=6N^{-1/2}=0.06$ (see \cref{fig:linearinterface}a--c), which appears
independent of $\phi$ (not shown).
\Cref{fig:linearinterface}b
confirms that $\fgg\approx[FG]$ (except when
forest cover approaches $[F]=0$ or $[F]=1$), confirming that the single-cluster approximation is valid.
Away from $[F]=0$ and $[F]=1$, $[FG]$ stays about constant as the forest interface moves
(\cref{fig:linearinterface}b--c).
%\jsq{this implies that the slope of $[FG]$ in \cref{fig:intro}c is $\gtrsim\beta[F]$ and suggests that for $\beta,\gamma>0$ \cref{fig:intro}c is valid even from heterogeneous initial conditions?}\bwq{To be checked.}
Therefore the gain and loss terms (as defined in \cref{eq:slow:gain,eq:slow:loss2}) are now, respectively,
constant and linearly decreasing with $[F]$ (\cref{fig:linearinterface}e),
such that $\mathrm{d}[F]/\mathrm{d}t$ increases linearly with $[F]$ (\cref{fig:linearinterface}g),
except near $[F]=0$ and $[F]=1$, where it connects to 0, as here $[FG]=0$.
\newpage
\section{Supplementary figures}
\begin{figure}[H]
\begin{center}
\includesvg[inkscapelatex=false,width=.33\columnwidth]{./figures/singlemulticluster_approx_vs_ctrl}
\caption{Steady states of \cref{eq:dFdtclust2ss} (multicluster -- dot-dashed red) and of \cref{eq:dFdtclust1ss} (single cluster -- dashed blue) compared to controlled simulations (solid black with shading indicating two-standard deviation confidence interval of the mean). Parameters in the cellular automaton: $\gamma{=}0.02,\alpha{=}0.03,\beta{=}2\cdot10^{-4},\rho_{g}{=}9\cdot 10^{6},\rho_{f}{=}1.11\cdot10^{5},\mu{=}10^6,\lambda{=}5$.}
\label{fig:singlemultipleclust_vs_ctrl}
\end{center}
\end{figure}
\begin{figure}[H]
\begin{center}
\includesvg[inkscapelatex=false,width=.33\columnwidth]{./figures/sepicpb09999pb09analysis}
\caption{Saddle of \cref{eq:dFdt:nospont} via \cref{eq:dFdt:nospont:saddle} (single cluster -- solid black) compared to controlled simulations (shaded red: $p_g=\rho_{g}/(\rho_g+\mu)=0.9$, shaded blue: $p_g=0.9999$) in case of heterogeneous initial conditions and without spontaneous interactions ($\beta{=}\gamma{=}0$). Other parameters: $\alpha{=}0.03,\rho_{f}{=}1.11\cdot10^{5},\mu{=}10^6,\lambda{=}5$.}
\label{fig:linearinterface_analysis}
\end{center}
\end{figure}
\end{document}
%merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked
%Control: key (0)
%Control: author (8) initials jnrlst
%Control: editor formatted (1) identically to author
%Control: production of article title (-1) disabled
%Control: page (0) single
%Control: year (1) truncated
%Control: production of eprint (0) enabled
\begin{thebibliography}{49}%
\makeatletter
\providecommand \@ifxundefined [1]{%
\@ifx{#1\undefined}
}%
\providecommand \@ifnum [1]{%
\ifnum #1\expandafter \@firstoftwo
\else \expandafter \@secondoftwo
\fi
}%
\providecommand \@ifx [1]{%
\ifx #1\expandafter \@firstoftwo
\else \expandafter \@secondoftwo
\fi
}%
\providecommand \natexlab [1]{#1}%
\providecommand \enquote [1]{``#1''}%
\providecommand \bibnamefont [1]{#1}%
\providecommand \bibfnamefont [1]{#1}%
\providecommand \citenamefont [1]{#1}%
\providecommand \href@noop [0]{\@secondoftwo}%
\providecommand \href [0]{\begingroup \@sanitize@url \@href}%
\providecommand \@href[1]{\@@startlink{#1}\@@href}%
\providecommand \@@href[1]{\endgroup#1\@@endlink}%
\providecommand \@sanitize@url [0]{\catcode `\\12\catcode `\$12\catcode
`\&12\catcode `\#12\catcode `\^12\catcode `\_12\catcode `\%12\relax}%
\providecommand \@@startlink[1]{}%
\providecommand \@@endlink[0]{}%
\providecommand \url [0]{\begingroup\@sanitize@url \@url }%
\providecommand \@url [1]{\endgroup\@href {#1}{\urlprefix }}%
\providecommand \urlprefix [0]{URL }%
\providecommand \Eprint [0]{\href }%
\providecommand \doibase [0]{http://dx.doi.org/}%
\providecommand \selectlanguage [0]{\@gobble}%
\providecommand \bibinfo [0]{\@secondoftwo}%
\providecommand \bibfield [0]{\@secondoftwo}%
\providecommand \translation [1]{[#1]}%
\providecommand \BibitemOpen [0]{}%
\providecommand \bibitemStop [0]{}%
\providecommand \bibitemNoStop [0]{.\EOS\space}%
\providecommand \EOS [0]{\spacefactor3000\relax}%
\providecommand \BibitemShut [1]{\csname bibitem#1\endcsname}%
\let\auto@bib@innerbib\@empty
%</preamble>
\bibitem [{\citenamefont {Staver}\ \emph {et~al.}(2011)\citenamefont {Staver},
\citenamefont {Archibald},\ and\ \citenamefont {Levin}}]{Staver2011b}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont
{Staver}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Archibald}}, \
and\ \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont {Levin}},\ }\href
{\doibase 10.1126/science.1210465} {\bibfield {journal} {\bibinfo {journal}
{Science}\ }\textbf {\bibinfo {volume} {334}},\ \bibinfo {pages} {230}
(\bibinfo {year} {2011})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Hirota}\ \emph {et~al.}(2011)\citenamefont {Hirota},
\citenamefont {Holmgren}, \citenamefont {{Van Nes}},\ and\ \citenamefont
{Scheffer}}]{Hirota2011}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Hirota}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Holmgren}},
\bibinfo {author} {\bibfnamefont {E.~H.}\ \bibnamefont {{Van Nes}}}, \ and\
\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Scheffer}},\ }\href
{\doibase 10.1126/science.1210657} {\bibfield {journal} {\bibinfo {journal}
{Science}\ }\textbf {\bibinfo {volume} {334}},\ \bibinfo {pages} {232}
(\bibinfo {year} {2011})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Lenton}\ \emph {et~al.}(2008)\citenamefont {Lenton},
\citenamefont {Held}, \citenamefont {Kriegler}, \citenamefont {Hall},
\citenamefont {Lucht}, \citenamefont {Rahmstorf},\ and\ \citenamefont
{Schellnhuber}}]{Lenton2008}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {T.~M.}\ \bibnamefont
{Lenton}}, \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Held}},
\bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Kriegler}}, \bibinfo
{author} {\bibfnamefont {J.~W.}\ \bibnamefont {Hall}}, \bibinfo {author}
{\bibfnamefont {W.}~\bibnamefont {Lucht}}, \bibinfo {author} {\bibfnamefont
{S.}~\bibnamefont {Rahmstorf}}, \ and\ \bibinfo {author} {\bibfnamefont
{H.~J.}\ \bibnamefont {Schellnhuber}},\ }\href {\doibase
10.1073/pnas.0705414105} {\bibfield {journal} {\bibinfo {journal}
{Proceedings of the National Academy of Sciences}\ }\textbf {\bibinfo
{volume} {105}},\ \bibinfo {pages} {1786} (\bibinfo {year}
{2008})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Scheffer}\ \emph {et~al.}(2015)\citenamefont
{Scheffer}, \citenamefont {Carpenter}, \citenamefont {Dakos},\ and\
\citenamefont {{Van Nes}}}]{Scheffer2015}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Scheffer}}, \bibinfo {author} {\bibfnamefont {S.~R.}\ \bibnamefont
{Carpenter}}, \bibinfo {author} {\bibfnamefont {V.}~\bibnamefont {Dakos}}, \
and\ \bibinfo {author} {\bibfnamefont {E.~H.}\ \bibnamefont {{Van Nes}}},\
}\href {\doibase 10.1146/annurev-ecolsys-112414-054242} {\bibfield {journal}
{\bibinfo {journal} {Annual Review of Ecology, Evolution, and Systematics}\
}\textbf {\bibinfo {volume} {46}},\ \bibinfo {pages} {145} (\bibinfo {year}
{2015})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Wuyts}\ \emph {et~al.}(2017)\citenamefont {Wuyts},
\citenamefont {Champneys},\ and\ \citenamefont {House}}]{Wuyts2017}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Wuyts}}, \bibinfo {author} {\bibfnamefont {A.~R.}\ \bibnamefont
{Champneys}}, \ and\ \bibinfo {author} {\bibfnamefont {J.~I.}\ \bibnamefont
{House}},\ }\href {\doibase 10.1038/ncomms15519} {\bibfield {journal}
{\bibinfo {journal} {Nature Communications}\ }\textbf {\bibinfo {volume}
{8}},\ \bibinfo {pages} {15519} (\bibinfo {year} {2017})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Hanan}\ \emph {et~al.}(2014)\citenamefont {Hanan},
\citenamefont {Tredennick}, \citenamefont {Prihodko}, \citenamefont
{Bucini},\ and\ \citenamefont {Dohn}}]{Hanan2014}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {N.~P.}\ \bibnamefont
{Hanan}}, \bibinfo {author} {\bibfnamefont {A.~T.}\ \bibnamefont
{Tredennick}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont
{Prihodko}}, \bibinfo {author} {\bibfnamefont {G.}~\bibnamefont {Bucini}}, \
and\ \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Dohn}},\ }\href
{\doibase 10.1111/geb.12122} {\bibfield {journal} {\bibinfo {journal}
{Global Ecology and Biogeography}\ }\textbf {\bibinfo {volume} {23}},\
\bibinfo {pages} {259} (\bibinfo {year} {2014})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Gerard}\ \emph {et~al.}(2017)\citenamefont {Gerard},
\citenamefont {Hooftman}, \citenamefont {van Langevelde}, \citenamefont
{Veenendaal}, \citenamefont {White},\ and\ \citenamefont
{Lloyd}}]{Gerard2017}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont
{Gerard}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Hooftman}},
\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {van Langevelde}},
\bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Veenendaal}}, \bibinfo
{author} {\bibfnamefont {S.~M.}\ \bibnamefont {White}}, \ and\ \bibinfo
{author} {\bibfnamefont {J.}~\bibnamefont {Lloyd}},\ }\href {\doibase
10.1111/geb.12592} {\bibfield {journal} {\bibinfo {journal} {Global Ecology
and Biogeography}\ }\textbf {\bibinfo {volume} {26}},\ \bibinfo {pages} {854}
(\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Staver}\ and\ \citenamefont
{Levin}(2012)}]{Staver2012}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont
{Staver}}\ and\ \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont
{Levin}},\ }\href {https://www.ncbi.nlm.nih.gov/pubmed/22766932} {\bibfield
{journal} {\bibinfo {journal} {The American Naturalist}\ }\textbf {\bibinfo
{volume} {180}},\ \bibinfo {pages} {211} (\bibinfo {year}
{2012})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Schertzer}\ \emph {et~al.}(2015)\citenamefont
{Schertzer}, \citenamefont {Staver},\ and\ \citenamefont
{Levin}}]{Schertzer2015}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {E.}~\bibnamefont
{Schertzer}}, \bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont
{Staver}}, \ and\ \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont
{Levin}},\ }\href {\doibase 10.1007/s00285-014-0757-z} {\bibfield {journal}
{\bibinfo {journal} {Journal of Mathematical Biology}\ }\textbf {\bibinfo
{volume} {70}},\ \bibinfo {pages} {329} (\bibinfo {year} {2015})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Touboul}\ \emph {et~al.}(2018)\citenamefont
{Touboul}, \citenamefont {Staver},\ and\ \citenamefont
{Levin}}]{Touboul2018}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {J.~D.}\ \bibnamefont
{Touboul}}, \bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont {Staver}},
\ and\ \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont {Levin}},\
}\href {\doibase 10.1073/pnas.1712356115} {\bibfield {journal} {\bibinfo
{journal} {Proceedings of the National Academy of Sciences}\ ,\ \bibinfo
{pages} {201712356}} (\bibinfo {year} {2018})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Goel}\ \emph {et~al.}(2020)\citenamefont {Goel},
\citenamefont {Guttal}, \citenamefont {Levin},\ and\ \citenamefont
{Staver}}]{Goel2020}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {N.}~\bibnamefont
{Goel}}, \bibinfo {author} {\bibfnamefont {V.}~\bibnamefont {Guttal}},
\bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont {Levin}}, \ and\
\bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont {Staver}},\ }\href
{\doibase 10.1086/708270} {\bibfield {journal} {\bibinfo {journal}
{American Naturalist}\ }\textbf {\bibinfo {volume} {195}},\ \bibinfo {pages}
{833} (\bibinfo {year} {2020})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Wuyts}\ \emph {et~al.}(2019)\citenamefont {Wuyts},
\citenamefont {Champneys}, \citenamefont {Verschueren},\ and\ \citenamefont
{House}}]{Wuyts2019}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Wuyts}}, \bibinfo {author} {\bibfnamefont {A.~R.}\ \bibnamefont
{Champneys}}, \bibinfo {author} {\bibfnamefont {N.}~\bibnamefont
{Verschueren}}, \ and\ \bibinfo {author} {\bibfnamefont {J.~I.}\ \bibnamefont
{House}},\ }\href {\doibase 10.1371/journal.pone.0218151} {\bibfield
{journal} {\bibinfo {journal} {PLoS ONE}\ }\textbf {\bibinfo {volume}
{14}},\ \bibinfo {pages} {e0218151} (\bibinfo {year} {2019})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Patterson}\ \emph {et~al.}(2021)\citenamefont
{Patterson}, \citenamefont {Levin}, \citenamefont {Staver},\ and\
\citenamefont {Touboul}}]{Patterson2021}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {D.~D.}\ \bibnamefont
{Patterson}}, \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont
{Levin}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {Staver}}, \
and\ \bibinfo {author} {\bibfnamefont {J.~D.}\ \bibnamefont {Touboul}},\
}\href {\doibase 10.1137/19M1298329} {\bibfield {journal} {\bibinfo
{journal} {SIAM Journal on Applied Dynamical Systems}\ }\textbf {\bibinfo
{volume} {19}},\ \bibinfo {pages} {2682} (\bibinfo {year}
{2021})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {H{\'{e}}bert-Dufresne}\ \emph
{et~al.}(2018)\citenamefont {H{\'{e}}bert-Dufresne}, \citenamefont
{Pellegrini}, \citenamefont {Bhat}, \citenamefont {Redner}, \citenamefont
{Pacala},\ and\ \citenamefont {Berdahl}}]{Hebert-Dufresne2018}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {L.}~\bibnamefont
{H{\'{e}}bert-Dufresne}}, \bibinfo {author} {\bibfnamefont {A.~F.}\
\bibnamefont {Pellegrini}}, \bibinfo {author} {\bibfnamefont
{U.}~\bibnamefont {Bhat}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Redner}}, \bibinfo {author} {\bibfnamefont {S.~W.}\ \bibnamefont {Pacala}},
\ and\ \bibinfo {author} {\bibfnamefont {A.~M.}\ \bibnamefont {Berdahl}},\
}\href {\doibase 10.1111/ele.12942} {\bibfield {journal} {\bibinfo
{journal} {Ecology Letters}\ }\textbf {\bibinfo {volume} {21}},\ \bibinfo
{pages} {794} (\bibinfo {year} {2018})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Bond}(2008)}]{Bond2008}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {W.~J.}\ \bibnamefont
{Bond}},\ }\href {\doibase 10.1146/annurev.ecolsys.39.110707.173411}
{\bibfield {journal} {\bibinfo {journal} {Annual Review of Ecology,
Evolution, and Systematics}\ }\textbf {\bibinfo {volume} {39}},\ \bibinfo
{pages} {641} (\bibinfo {year} {2008})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Hoffmann}\ \emph {et~al.}(2012)\citenamefont
{Hoffmann}, \citenamefont {Geiger}, \citenamefont {Gotsch}, \citenamefont
{Rossatto}, \citenamefont {Silva}, \citenamefont {Lau}, \citenamefont
{Haridasan},\ and\ \citenamefont {Franco}}]{Hoffmann2012}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {W.~A.}\ \bibnamefont
{Hoffmann}}, \bibinfo {author} {\bibfnamefont {E.~L.}\ \bibnamefont
{Geiger}}, \bibinfo {author} {\bibfnamefont {S.~G.}\ \bibnamefont {Gotsch}},
\bibinfo {author} {\bibfnamefont {D.~R.}\ \bibnamefont {Rossatto}}, \bibinfo
{author} {\bibfnamefont {L.~C.}\ \bibnamefont {Silva}}, \bibinfo {author}
{\bibfnamefont {O.~L.}\ \bibnamefont {Lau}}, \bibinfo {author} {\bibfnamefont
{M.}~\bibnamefont {Haridasan}}, \ and\ \bibinfo {author} {\bibfnamefont
{A.~C.}\ \bibnamefont {Franco}},\ }\href {\doibase
10.1111/j.1461-0248.2012.01789.x} {\bibfield {journal} {\bibinfo {journal}
{Ecology Letters}\ }\textbf {\bibinfo {volume} {15}},\ \bibinfo {pages} {759}
(\bibinfo {year} {2012})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Archibald}\ \emph {et~al.}(2009)\citenamefont
{Archibald}, \citenamefont {Roy}, \citenamefont {van Wilgen},\ and\
\citenamefont {Scholes}}]{Archibald2009}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Archibald}}, \bibinfo {author} {\bibfnamefont {D.~P.}\ \bibnamefont {Roy}},
\bibinfo {author} {\bibfnamefont {B.~W.}\ \bibnamefont {van Wilgen}}, \ and\
\bibinfo {author} {\bibfnamefont {R.~J.}\ \bibnamefont {Scholes}},\ }\href
{\doibase 10.1111/j.1365-2486.2008.01754.x} {\bibfield {journal} {\bibinfo
{journal} {Global Change Biology}\ }\textbf {\bibinfo {volume} {15}},\
\bibinfo {pages} {613} (\bibinfo {year} {2009})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {{Van Nes}}\ \emph {et~al.}(2018)\citenamefont {{Van
Nes}}, \citenamefont {Staal}, \citenamefont {Hantson}, \citenamefont
{Holmgren}, \citenamefont {Pueyo}, \citenamefont {Bernardi}, \citenamefont
{Flores}, \citenamefont {Xu},\ and\ \citenamefont {Scheffer}}]{VanNes2018}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {E.~H.}\ \bibnamefont
{{Van Nes}}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Staal}},
\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Hantson}}, \bibinfo
{author} {\bibfnamefont {M.}~\bibnamefont {Holmgren}}, \bibinfo {author}
{\bibfnamefont {S.}~\bibnamefont {Pueyo}}, \bibinfo {author} {\bibfnamefont
{R.~E.}\ \bibnamefont {Bernardi}}, \bibinfo {author} {\bibfnamefont {B.~M.}\
\bibnamefont {Flores}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont
{Xu}}, \ and\ \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Scheffer}},\ }\href {\doibase 10.1371/journal.pone.0191027} {\bibfield
{journal} {\bibinfo {journal} {PloS One}\ }\textbf {\bibinfo {volume}
{13}},\ \bibinfo {pages} {e0191027} (\bibinfo {year} {2018})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Stauffer}\ and\ \citenamefont
{Aharony}(1994)}]{Stauffer1994}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {D.}~\bibnamefont
{Stauffer}}\ and\ \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont
{Aharony}},\ }\href@noop {} {\emph {\bibinfo {title} {{Introduction to
percolation theory}}}}\ (\bibinfo {publisher} {CRC press},\ \bibinfo {year}
{1994})\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Christensen}\ and\ \citenamefont
{Moloney}(2005)}]{Christensen2005}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {K.}~\bibnamefont
{Christensen}}\ and\ \bibinfo {author} {\bibfnamefont {N.~R.}\ \bibnamefont
{Moloney}},\ }\href {\doibase 10.1142/P365} {\emph {\bibinfo {title}
{{Complexity and Criticality}}}},\ \bibinfo {series} {Imperial College Press
Advanced Physics Texts}, Vol.~\bibinfo {volume} {1}\ (\bibinfo {publisher}
{Imperial College Press and Distributed},\ \bibinfo {year}
{2005})\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Cardoso}\ \emph {et~al.}(2022)\citenamefont
{Cardoso}, \citenamefont {Archibald}, \citenamefont {Bond}, \citenamefont
{Coetsee}, \citenamefont {Forrest}, \citenamefont {Govender}, \citenamefont
{Lehmann}, \citenamefont {Makaga}, \citenamefont {Mpanza}, \citenamefont
{Ndong}, \citenamefont {{Koumba Pambo}}, \citenamefont {Strydom},
\citenamefont {Tilman}, \citenamefont {Wragg},\ and\ \citenamefont
{Staver}}]{Cardoso2022}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {A.~W.}\ \bibnamefont
{Cardoso}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Archibald}},
\bibinfo {author} {\bibfnamefont {W.~J.}\ \bibnamefont {Bond}}, \bibinfo
{author} {\bibfnamefont {C.}~\bibnamefont {Coetsee}}, \bibinfo {author}
{\bibfnamefont {M.}~\bibnamefont {Forrest}}, \bibinfo {author} {\bibfnamefont
{N.}~\bibnamefont {Govender}}, \bibinfo {author} {\bibfnamefont
{D.}~\bibnamefont {Lehmann}}, \bibinfo {author} {\bibfnamefont
{L.}~\bibnamefont {Makaga}}, \bibinfo {author} {\bibfnamefont
{N.}~\bibnamefont {Mpanza}}, \bibinfo {author} {\bibfnamefont {J.~E.}\
\bibnamefont {Ndong}}, \bibinfo {author} {\bibfnamefont {A.~F.}\ \bibnamefont
{{Koumba Pambo}}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont
{Strydom}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Tilman}},
\bibinfo {author} {\bibfnamefont {P.~D.}\ \bibnamefont {Wragg}}, \ and\
\bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont {Staver}},\ }\href
{\doibase 10.1073/pnas.2110364119} {\bibfield {journal} {\bibinfo {journal}
{Proceedings of the National Academy of Sciences of the United States of
America}\ }\textbf {\bibinfo {volume} {119}} (\bibinfo {year} {2022}),\
10.1073/pnas.2110364119}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Spracklen}\ \emph {et~al.}(2018)\citenamefont
{Spracklen}, \citenamefont {Baker}, \citenamefont {Garcia-Carreras},\ and\
\citenamefont {Marsham}}]{Spracklen2018}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {D.~V.}\ \bibnamefont
{Spracklen}}, \bibinfo {author} {\bibfnamefont {J.~C.}\ \bibnamefont
{Baker}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont
{Garcia-Carreras}}, \ and\ \bibinfo {author} {\bibfnamefont {J.~H.}\
\bibnamefont {Marsham}},\ }\href {\doibase
10.1146/annurev-environ-102017-030136} {\enquote {\bibinfo {title} {{The
effects of tropical vegetation on rainfall}},}\ } (\bibinfo {year}
{2018})\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Veenendaal}\ \emph {et~al.}(2018)\citenamefont
{Veenendaal}, \citenamefont {Torello-Raventos}, \citenamefont {Miranda},
\citenamefont {Sato}, \citenamefont {Oliveras}, \citenamefont {van
Langevelde}, \citenamefont {Asner},\ and\ \citenamefont
{Lloyd}}]{Veenendaal2018}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {E.~M.}\ \bibnamefont
{Veenendaal}}, \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Torello-Raventos}}, \bibinfo {author} {\bibfnamefont {H.~S.}\ \bibnamefont
{Miranda}}, \bibinfo {author} {\bibfnamefont {N.~M.}\ \bibnamefont {Sato}},
\bibinfo {author} {\bibfnamefont {I.}~\bibnamefont {Oliveras}}, \bibinfo
{author} {\bibfnamefont {F.}~\bibnamefont {van Langevelde}}, \bibinfo
{author} {\bibfnamefont {G.~P.}\ \bibnamefont {Asner}}, \ and\ \bibinfo
{author} {\bibfnamefont {J.}~\bibnamefont {Lloyd}},\ }\href {\doibase
10.1111/nph.14940} {\bibfield {journal} {\bibinfo {journal} {New
Phytologist}\ }\textbf {\bibinfo {volume} {218}},\ \bibinfo {pages} {153}
(\bibinfo {year} {2018})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {{Van Langevelde}}\ \emph {et~al.}(2003)\citenamefont
{{Van Langevelde}}, \citenamefont {{Van De Vijver}}, \citenamefont {Kumar},
\citenamefont {{Van De Koppel}}, \citenamefont {{De Ridder}}, \citenamefont
{{Van Andel}}, \citenamefont {Skidmore}, \citenamefont {Hearne},
\citenamefont {Stroosnijder}, \citenamefont {Bond}, \citenamefont {Prins},\
and\ \citenamefont {Rietkerk}}]{VanLangevelde2003}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {{Van
Langevelde}}}, \bibinfo {author} {\bibfnamefont {C.~A.}\ \bibnamefont {{Van
De Vijver}}}, \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Kumar}},
\bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {{Van De Koppel}}},
\bibinfo {author} {\bibfnamefont {N.}~\bibnamefont {{De Ridder}}}, \bibinfo
{author} {\bibfnamefont {J.}~\bibnamefont {{Van Andel}}}, \bibinfo {author}
{\bibfnamefont {A.~K.}\ \bibnamefont {Skidmore}}, \bibinfo {author}
{\bibfnamefont {J.~W.}\ \bibnamefont {Hearne}}, \bibinfo {author}
{\bibfnamefont {L.}~\bibnamefont {Stroosnijder}}, \bibinfo {author}
{\bibfnamefont {W.~J.}\ \bibnamefont {Bond}}, \bibinfo {author}
{\bibfnamefont {H.~H.}\ \bibnamefont {Prins}}, \ and\ \bibinfo {author}
{\bibfnamefont {M.}~\bibnamefont {Rietkerk}},\ }\href {\doibase
10.1890/0012-9658(2003)084[0337:EOFAHO]2.0.CO;2} {\bibfield {journal}
{\bibinfo {journal} {Ecology}\ }\textbf {\bibinfo {volume} {84}},\ \bibinfo
{pages} {337} (\bibinfo {year} {2003})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Accatino}\ \emph {et~al.}(2010)\citenamefont
{Accatino}, \citenamefont {{De Michele}}, \citenamefont {Vezzoli},
\citenamefont {Donzelli},\ and\ \citenamefont {Scholes}}]{Accatino2010}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont
{Accatino}}, \bibinfo {author} {\bibfnamefont {C.}~\bibnamefont {{De
Michele}}}, \bibinfo {author} {\bibfnamefont {R.}~\bibnamefont {Vezzoli}},
\bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Donzelli}}, \ and\
\bibinfo {author} {\bibfnamefont {R.~J.}\ \bibnamefont {Scholes}},\ }\href
{\doibase 10.1016/j.jtbi.2010.08.012} {\bibfield {journal} {\bibinfo
{journal} {Journal of Theoretical Biology}\ }\textbf {\bibinfo {volume}
{267}},\ \bibinfo {pages} {235} (\bibinfo {year} {2010})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {D'Odorico}\ \emph {et~al.}(2006)\citenamefont
{D'Odorico}, \citenamefont {Laio},\ and\ \citenamefont
{Ridolfi}}]{DOdorico2006}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {P.}~\bibnamefont
{D'Odorico}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {Laio}}, \
and\ \bibinfo {author} {\bibfnamefont {L.}~\bibnamefont {Ridolfi}},\ }\href
{\doibase 10.1086/500617} {\bibfield {journal} {\bibinfo {journal} {The
American naturalist}\ }\textbf {\bibinfo {volume} {167}} (\bibinfo {year}
{2006}),\ 10.1086/500617}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Baudena}\ \emph {et~al.}(2010)\citenamefont
{Baudena}, \citenamefont {D'Andrea},\ and\ \citenamefont
{Provenzale}}]{Baudena2010}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Baudena}}, \bibinfo {author} {\bibfnamefont {F.}~\bibnamefont {D'Andrea}}, \
and\ \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Provenzale}},\
}\href {\doibase 10.1111/j.1365-2745.2009.01588.x} {\bibfield {journal}
{\bibinfo {journal} {Journal of Ecology}\ }\textbf {\bibinfo {volume}
{98}},\ \bibinfo {pages} {74} (\bibinfo {year} {2010})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Beckage}\ \emph {et~al.}(2011)\citenamefont
{Beckage}, \citenamefont {Gross},\ and\ \citenamefont {Platt}}]{Beckage2011}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Beckage}}, \bibinfo {author} {\bibfnamefont {L.~J.}\ \bibnamefont {Gross}},
\ and\ \bibinfo {author} {\bibfnamefont {W.~J.}\ \bibnamefont {Platt}},\
}\href {\doibase 10.1016/j.ecolmodel.2011.01.015} {\bibfield {journal}
{\bibinfo {journal} {Ecological Modelling}\ }\textbf {\bibinfo {volume}
{222}},\ \bibinfo {pages} {2227} (\bibinfo {year} {2011})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Hoyer-Leitzel}\ and\ \citenamefont
{Iams}(2021)}]{Hoyer-Leitzel2021}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {A.}~\bibnamefont
{Hoyer-Leitzel}}\ and\ \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Iams}},\ }\href {\doibase 10.1007/s11538-021-00944-x} {\bibfield {journal}
{\bibinfo {journal} {Bulletin of Mathematical Biology}\ }\textbf {\bibinfo
{volume} {83}} (\bibinfo {year} {2021}),\
10.1007/s11538-021-00944-x}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Archibald}\ \emph {et~al.}(2012)\citenamefont
{Archibald}, \citenamefont {Staver},\ and\ \citenamefont
{Levin}}]{Archibald2012a}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Archibald}}, \bibinfo {author} {\bibfnamefont {A.~C.}\ \bibnamefont
{Staver}}, \ and\ \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont
{Levin}},\ }\href {\doibase 10.1073/pnas.1118648109} {\bibfield {journal}
{\bibinfo {journal} {Proceedings of the National Academy of Sciences}\
}\textbf {\bibinfo {volume} {109}},\ \bibinfo {pages} {847} (\bibinfo {year}
{2012})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Favier}\ \emph {et~al.}(2004)\citenamefont {Favier},
\citenamefont {Chave}, \citenamefont {Fabing}, \citenamefont {Schwartz},\
and\ \citenamefont {Dubois}}]{Favier2004}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {C.}~\bibnamefont
{Favier}}, \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Chave}},
\bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Fabing}}, \bibinfo
{author} {\bibfnamefont {D.}~\bibnamefont {Schwartz}}, \ and\ \bibinfo
{author} {\bibfnamefont {M.~A.}\ \bibnamefont {Dubois}},\ }\href {\doibase
10.1016/j.ecolmodel.2003.07.003} {\bibfield {journal} {\bibinfo {journal}
{Ecological Modelling}\ }\textbf {\bibinfo {volume} {171}},\ \bibinfo {pages}
{85} (\bibinfo {year} {2004})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Wuyts}\ and\ \citenamefont
{Sieber}(2021)}]{Wuyts2021}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {B.}~\bibnamefont
{Wuyts}}\ and\ \bibinfo {author} {\bibfnamefont {J.}~\bibnamefont {Sieber}},\
}\href {https://arxiv.org/abs/2111.07643v1} {\enquote {\bibinfo {title}
{{Automated derivation and closure of moment equations for multistate
dynamics on networks}},}\ } (\bibinfo {year} {2021}),\ \Eprint
{http://arxiv.org/abs/2111.07643} {arXiv:2111.07643} \BibitemShut {NoStop}%
\bibitem [{\citenamefont {Keeling}(1999)}]{Keeling1999}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {M.~J.}\ \bibnamefont
{Keeling}},\ }\href {\doibase 10.1098/rspb.1999.0716} {\bibfield {journal}
{\bibinfo {journal} {Proceedings of the Royal Society B: Biological
Sciences}\ }\textbf {\bibinfo {volume} {266}},\ \bibinfo {pages} {859}
(\bibinfo {year} {1999})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Gross}\ and\ \citenamefont
{Kevrekidis}(2008)}]{Gross2008}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {T.}~\bibnamefont
{Gross}}\ and\ \bibinfo {author} {\bibfnamefont {I.~G.}\ \bibnamefont
{Kevrekidis}},\ }\href {\doibase 10.1209/0295-5075/82/38004} {\bibfield
{journal} {\bibinfo {journal} {Epl}\ }\textbf {\bibinfo {volume} {82}},\
\bibinfo {pages} {38004} (\bibinfo {year} {2008})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Ge}\ and\ \citenamefont {Qian}(2009)}]{Ge2009}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {H.}~\bibnamefont
{Ge}}\ and\ \bibinfo {author} {\bibfnamefont {H.}~\bibnamefont {Qian}},\
}\href {\doibase 10.1103/PhysRevLett.103.148103} {\bibfield {journal}
{\bibinfo {journal} {Physical Review Letters}\ }\textbf {\bibinfo {volume}
{103}},\ \bibinfo {pages} {148103} (\bibinfo {year} {2009})}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Thiele}\ \emph {et~al.}(2019)\citenamefont {Thiele},
\citenamefont {Frohoff-H{\"{u}}lsmann}, \citenamefont {Engelnkemper},
\citenamefont {Knobloch},\ and\ \citenamefont {Archer}}]{Thiele2019}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {U.}~\bibnamefont
{Thiele}}, \bibinfo {author} {\bibfnamefont {T.}~\bibnamefont
{Frohoff-H{\"{u}}lsmann}}, \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont
{Engelnkemper}}, \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont
{Knobloch}}, \ and\ \bibinfo {author} {\bibfnamefont {A.~J.}\ \bibnamefont
{Archer}},\ }\href {\doibase 10.1088/1367-2630/ab5caf} {\bibfield {journal}
{\bibinfo {journal} {New Journal of Physics}\ }\textbf {\bibinfo {volume}
{21}},\ \bibinfo {pages} {123021} (\bibinfo {year} {2019})},\ \Eprint
{http://arxiv.org/abs/1908.11304} {arXiv:1908.11304} \BibitemShut {NoStop}%
\bibitem [{\citenamefont {Goldenfeld}(2018)}]{Goldenfeld2018}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {N.}~\bibnamefont
{Goldenfeld}},\ }\href {\doibase 10.1201/9780429493492} {\emph {\bibinfo
{title} {Lectures on Phase Transitions and the Renormalization Group}}}\
(\bibinfo {publisher} {CRC Press},\ \bibinfo {year} {2018})\ pp.\ \bibinfo
{pages} {1--394}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Liehr}(2013)}]{Liehr2013}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {A.}~\bibnamefont
{Liehr}},\ }\href@noop {} {\emph {\bibinfo {title} {{Dissipative solitons in
reaction diffusion systems}}}},\ Vol.~\bibinfo {volume} {70}\ (\bibinfo
{publisher} {Springer},\ \bibinfo {year} {2013})\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Kiss}\ \emph {et~al.}(2017)\citenamefont {Kiss},
\citenamefont {Miller},\ and\ \citenamefont {Simon}}]{Kiss2017}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {I.~Z.}\ \bibnamefont
{Kiss}}, \bibinfo {author} {\bibfnamefont {J.~C.}\ \bibnamefont {Miller}}, \
and\ \bibinfo {author} {\bibfnamefont {P.~L.}\ \bibnamefont {Simon}},\ }\href
{\doibase 10.1007/978-3-319-50806-1} {\emph {\bibinfo {title}
{Interdisciplinary Applied Mathematics}}},\ \bibinfo {series}
{Interdisciplinary Applied Mathematics}, Vol.~\bibinfo {volume} {46}\
(\bibinfo {publisher} {Springer International Publishing},\ \bibinfo
{address} {Cham},\ \bibinfo {year} {2017})\ pp.\ \bibinfo {pages}
{1--413}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Gillespie}(2007)}]{Gillespie2007}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {D.~T.}\ \bibnamefont
{Gillespie}},\ }\href {\doibase 10.1146/annurev.physchem.58.032806.104637}
{\bibfield {journal} {\bibinfo {journal} {Annual Review of Physical
Chemistry}\ }\textbf {\bibinfo {volume} {58}},\ \bibinfo {pages} {35}
(\bibinfo {year} {2007})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Sieber}\ \emph {et~al.}(2008)\citenamefont {Sieber},
\citenamefont {Gonzalez-Buelga}, \citenamefont {Neild}, \citenamefont
{Wagg},\ and\ \citenamefont {Krauskopf}}]{SGNWK08}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {J.}~\bibnamefont
{Sieber}}, \bibinfo {author} {\bibfnamefont {A.}~\bibnamefont
{Gonzalez-Buelga}}, \bibinfo {author} {\bibfnamefont {S.~A.}\ \bibnamefont
{Neild}}, \bibinfo {author} {\bibfnamefont {D.~J.}\ \bibnamefont {Wagg}}, \
and\ \bibinfo {author} {\bibfnamefont {B.}~\bibnamefont {Krauskopf}},\
}\href@noop {} {\bibfield {journal} {\bibinfo {journal} {Phys. Rev. Lett.}\
}\textbf {\bibinfo {volume} {100}},\ \bibinfo {pages} {244101} (\bibinfo
{year} {2008})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Schilder}\ \emph {et~al.}(2015)\citenamefont
{Schilder}, \citenamefont {Bureau}, \citenamefont {Santos}, \citenamefont
{Thomsen},\ and\ \citenamefont {Starke}}]{schilder2015experimental}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {F.}~\bibnamefont
{Schilder}}, \bibinfo {author} {\bibfnamefont {E.}~\bibnamefont {Bureau}},
\bibinfo {author} {\bibfnamefont {I.~F.}\ \bibnamefont {Santos}}, \bibinfo
{author} {\bibfnamefont {J.~J.}\ \bibnamefont {Thomsen}}, \ and\ \bibinfo
{author} {\bibfnamefont {J.}~\bibnamefont {Starke}},\ }\href@noop {}
{\bibfield {journal} {\bibinfo {journal} {Journal of Sound and Vibration}\
}\textbf {\bibinfo {volume} {358}},\ \bibinfo {pages} {251} (\bibinfo {year}
{2015})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Barton}(2017)}]{barton2017control}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {D.}~\bibnamefont
{Barton}},\ }\href@noop {} {\bibfield {journal} {\bibinfo {journal}
{Mechanical Systems and Signal Processing}\ }\textbf {\bibinfo {volume}
{84}},\ \bibinfo {pages} {54} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Neville}\ \emph {et~al.}(2018)\citenamefont
{Neville}, \citenamefont {Groh}, \citenamefont {Pirrera},\ and\ \citenamefont
{Schenk}}]{neville2018shape}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {R.~M.}\ \bibnamefont
{Neville}}, \bibinfo {author} {\bibfnamefont {R.~M.}\ \bibnamefont {Groh}},
\bibinfo {author} {\bibfnamefont {A.}~\bibnamefont {Pirrera}}, \ and\
\bibinfo {author} {\bibfnamefont {M.}~\bibnamefont {Schenk}},\ }\href@noop {}
{\bibfield {journal} {\bibinfo {journal} {Physical review letters}\
}\textbf {\bibinfo {volume} {120}},\ \bibinfo {pages} {254101} (\bibinfo
{year} {2018})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Barton}\ and\ \citenamefont {Sieber}(2013)}]{BS13}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {D.~A.~W.}\
\bibnamefont {Barton}}\ and\ \bibinfo {author} {\bibfnamefont
{J.}~\bibnamefont {Sieber}},\ }\href@noop {} {\bibfield {journal} {\bibinfo
{journal} {Phys. Rev. E}\ }\textbf {\bibinfo {volume} {87}},\ \bibinfo
{pages} {052916} (\bibinfo {year} {2013})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Sieber}\ \emph {et~al.}(2014)\citenamefont {Sieber},
\citenamefont {Omel'chenko},\ and\ \citenamefont {Wolfrum}}]{SOW14}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {J.}~\bibnamefont
{Sieber}}, \bibinfo {author} {\bibfnamefont {O.~E.}\ \bibnamefont
{Omel'chenko}}, \ and\ \bibinfo {author} {\bibfnamefont {M.}~\bibnamefont
{Wolfrum}},\ }\href@noop {} {\bibfield {journal} {\bibinfo {journal} {Phys.
Rev. Lett.}\ }\textbf {\bibinfo {volume} {112}},\ \bibinfo {pages} {054102}
(\bibinfo {year} {2014})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Renson}\ \emph {et~al.}(2017)\citenamefont {Renson},
\citenamefont {Barton},\ and\ \citenamefont
{Neild}}]{renson2017experimental}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {L.}~\bibnamefont
{Renson}}, \bibinfo {author} {\bibfnamefont {D.}~\bibnamefont {Barton}}, \
and\ \bibinfo {author} {\bibfnamefont {S.}~\bibnamefont {Neild}},\
}\href@noop {} {\bibfield {journal} {\bibinfo {journal} {International
Journal of Bifurcation and Chaos}\ }\textbf {\bibinfo {volume} {27}},\
\bibinfo {pages} {1730002} (\bibinfo {year} {2017})}\BibitemShut {NoStop}%
\bibitem [{\citenamefont {Kreiss}\ and\ \citenamefont
{Lahiri}(2012)}]{Kreiss2012}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {J.~P.}\ \bibnamefont
{Kreiss}}\ and\ \bibinfo {author} {\bibfnamefont {S.~N.}\ \bibnamefont
{Lahiri}},\ }in\ \href {\doibase 10.1016/B978-0-444-53858-1.00001-6} {\emph
{\bibinfo {booktitle} {Handbook of Statistics}}},\ Vol.~\bibinfo {volume}
{30}\ (\bibinfo {year} {2012})\ pp.\ \bibinfo {pages} {3--26}\BibitemShut
{NoStop}%
\bibitem [{\citenamefont {Tom{\'{e}}}\ and\ \citenamefont {{De
Oliveira}}(2015)}]{Tome2015}%
\BibitemOpen
\bibfield {author} {\bibinfo {author} {\bibfnamefont {T.}~\bibnamefont
{Tom{\'{e}}}}\ and\ \bibinfo {author} {\bibfnamefont {M.~J.}\ \bibnamefont
{{De Oliveira}}},\ }\href@noop {} {\emph {\bibinfo {title} {{Stochastic
dynamics and irreversibility}}}}\ (\bibinfo {publisher} {Springer},\
\bibinfo {year} {2015})\ p.\ \bibinfo {pages} {394}\BibitemShut {NoStop}%
\end{thebibliography}%