q-bio0404035/SIRS.tex
1: \documentclass[aps,pre,showpacs,preprint]{revtex4}
2: \usepackage{graphicx,amsmath,bm}
3: %\renewcommand{\baselinestretch}{2}
4: 
5: \begin{document}
6:  
7: \title{Pair Approximation of the stochastic susceptible-infected-recovered-susceptible epidemic model 
8: on the hypercubic lattice}
9: 
10: \author{Jaewook Joo}
11: \affiliation{Department of Physics, Rutgers University, New Brunswick, New 
12: Jersey 08854, USA}
13: \author{Joel L. Lebowitz} 
14: \affiliation{Department of Mathematics and Physics, Rutgers University, New Brunswick, New 
15: Jersey 08854, USA}
16: \date{\today}
17: 
18: \begin{abstract}
19: We investigate the time-evolution and steady states of the stochastic 
20: susceptible-infected-recovered-susceptible~(SIRS) epidemic model
21: on one- and two- dimensional lattices. 
22: We compare the behavior of this system, obtained from computer simulations, 
23: with those obtained from the mean-field approximation~(MFA) 
24: and pair-approximation~(PA). The former~(latter) approximates higher order moments 
25: in terms of first~(second) order ones. 
26: We find that the PA gives consistently better results than the MFA. 
27: In one dimension the improvement is even qualitative.
28: \end{abstract}
29: \pacs{87.23.Ge,05.70.Ln}
30: \maketitle
31: 
32: \section{\label{sec:intro}Introduction}
33: 
34: The mathematical modeling of the spread of epidemics is a subject of continuing theoretical 
35: and practical interest~\cite{anderson:1992,diekmann:2000}. 
36: This is enhanced by the fact that the same or similar models are used for 
37: describing other phenomena such as plant and animal dispersal, and successional dynamics 
38: in ecology~\cite{neuhauser:2000,Jeger:1990}. 
39: 
40: The level of description provided by a model 
41: can be purely macroscopic and deterministic or individual and stochastic~\cite{durrett:1994a}. 
42: In the first case one uses (partial-) differential equations 
43: to describe the time evolution of different subpopulations; e.g., susceptible, 
44: infectious and recovered. In the second case one typically uses stochastic dynamics 
45: on a lattice (or more general graphs) where the variables at each node represent the state 
46: of an individual or a small spatial region. 
47: The time evolution of these variables is stochastic, e.g., an infected individual 
48: at site $i$ has a certain probability per unit time (rate) 
49: $\lambda$ to infect a susceptible individual at a neighboring site $j$. 
50: These systems fall into the category of what mathematicians call interacting particle 
51: systems~\cite{liggett:1985,durrett:1988b} and physicists call stochastic 
52: lattice gases~\cite{marro:1999} - systems of great interest also 
53: in the study of equilibrium phase transitions, phase segregation kinetics, etc., 
54: fields very different from epidemiology and ecology.
55: 
56: 
57: The connection between these modes of description and various intermediate ones 
58: has been investigated extensively in recent years, e.g., see ~\cite{lebowitz:1985-86a,
59: lebowitz:1986b,lebowitz:1988,durrett:1994,durrett:1994a,bramson:1997}. 
60: Mathematically this involves the use of the so called hydro-dynamical scaling limit. 
61: This uses a rigorous separation of space and time scales 
62: to derive deterministic macroscopic equations 
63: from the microscopic dynamics of stochastic lattice systems. 
64: Other approaches are based on more heuristic methods 
65: such as the mean field approximation (MFA) and improvement thereof~\cite{dickman:1986,
66: matsuda:1992,benavraham:1992,petermann:2004,levin:1995,levin:1996,levin:1997,durrett:1998,keeling:1999,
67: levin:2000,schnaiz:2002,geometry:2002}
68: 
69: 
70: The present work falls in the latter category. We apply a pair approximation (PA) 
71: scheme to a microscopic stochastic epidemic model in which individuals recovered 
72: from an infection 
73: enjoy a period of immunity before again becoming susceptible at a rate $\gamma$: 
74: the SIRS model. 
75: The PA approximation was used by Durrett and Levin~\cite{levin:1996} for the simpler 
76: susceptible-infected-susceptible~(SIS) model where 
77: recovered individuals immediately become susceptible again. 
78: They compared the results of the PA and MFA with those of the stochastic 
79: SIS model and found that the 
80: PA gave a quantitative improvement over the MFA. 
81: Here we consider the general SIRS model. 
82: We obtain the behavior of the stochastic model from extensive computer simulations. 
83: We then solve the PA and MFA models analytically for the stationary state and numerically 
84: for the time dependent case. We find that 
85: the PA gives considerably better agreement with the simulations than the MFA both for the 
86: time evolution and for the 
87: steady state. For the latter the PA reproduces the qualitative difference between 
88: the one and higher dimensional phase diagram of this model found in 
89: Ref.~\cite{kuulasmaa:1982, durrett:1991,andjel:1996,berg:1998}. This is reminiscent of 
90: the relation between the MFA and the Bethe-Peierls approximation
91: ~(which the PA closely resembles) for equilibrium lattice systems~\cite{huang}. 
92: 
93: 
94: %[similar models]
95: %For stochastic lattice models which are closely related to the SIRS case, one could 
96: %naturally expect that these models have a similar phase diagram with that of the SIRS 
97: %case and that the PA will reproduce it qualitatively.   
98: %These models include forest fire models~\cite{bak:1990,drossel:1993}, Prey-predator 
99: %models~\cite{antal:2001} and Immunization models~\cite{dammer:2003}.
100: 
101: \section{\label{sec:model}The stochastic SIRS model}
102: 
103: We first recall the stochastic lattice model of the SIRS epidemic process~\cite{murray:1980}.
104: A site $x$ of a $d$-dimensional lattice can be occupied by an individual in a state of  
105: $S$~(healthy and susceptible), $I$~(infected), or $R$~(recovered, i.e., healthy and immune).
106: The system evolves according to the following transition rates, 
107: \begin{eqnarray} 
108: S \rightarrow I &\text{ at rate }& \lambda n(x), \label{transition}
109: \\ \nonumber
110: I \rightarrow R &\text{ at rate }& \delta,
111: \\ \nonumber
112: R \rightarrow S &\text{ at rate }& \gamma, 
113: \end{eqnarray}
114: where $n(x)$ is the number of infected (nearest) neighbors of $x$, 
115: $\lambda$ is the infection rate, $\delta$ is the recovery rate 
116: and $\gamma$ is the rate at which immunization ceases.
117: The limit $\gamma \rightarrow \infty$ corresponds to the case where 
118: a recovered site passes instantaneously through the state $R$; 
119: this is the SIS model, also known as the contact process.
120: We shall choose time units in which $\delta$=1.
121: 
122: 
123: %[Summary of Previous work, sketch of phase diagram]
124: One can obtain some rigorous qualitative information about this and 
125: related models via probabilistic approaches such as those used in 
126: interacting particle systems~\cite{kuulasmaa:1982, durrett:1991, andjel:1996,berg:1998}. 
127: Of particular interest is the behavior of the 
128: stationary state on an infinite lattice which is a good approximation for 
129: the quasi-steady state behavior of large systems: see Appendix~\ref{MCsimulation}.
130: . This information is encoded in the phase 
131: diagram of the stationary state which depends on the infection rate $\lambda$, 
132: the recovery rate $\gamma$ and the topology of the lattice. For small $\lambda$, 
133: the only stationary 
134: state is one in which all sites are in the susceptible (disease-free) state
135: while for large $\lambda$ there is (for the infinite system) 
136: also a stationary state containing non zero fraction of $I$ and $R$ individuals. 
137: 
138: 
139: The critical infection rate $\lambda_{c}(\gamma)$ is defined as the smallest value 
140: of $\lambda$, for a given $\gamma$, above which the infection can persist forever. 
141: For the SIS or contact process~($\gamma=\infty$), the critical infection value is 
142: known with high accuracy,
143: $\lambda_{c}(\infty) \simeq 1.6489$ in $d=1$ and 
144: $\lambda_{c}(\infty)\simeq 0.4122$ in $d=2$~\cite{liggett:1985,marro:1999}.
145: Considerably less is known about the phase diagram of the SIRS model.
146: Interestingly there is a qualitative difference in the behavior of 
147: $\lambda_{c}(\gamma)$ in one and in higher dimension when 
148: $\gamma \rightarrow 0$. 
149: It has been shown that $lim_{\gamma \rightarrow 0} \lambda_{c}(\gamma)=\lambda_{c}(0)$ 
150: is finite when $d \geq 2$ while $\lambda_{c}(0)=\infty$ when $d=1$
151: ~\cite{kuulasmaa:1982, durrett:1991,andjel:1996,berg:1998}.
152: 
153: 
154: To go beyond qualitative results we need to carry out simulation or 
155: make some approximations. This is the subject of the rest of the paper.
156: 
157: 
158: \section{\label{sec:PA}The pair approximation}
159: 
160: %[master equation] 
161: The time evolution of the single site probabilities in the stochastic SIRS 
162: epidemic process can be written in the following form.
163: \begin{subequations}
164: \label{allmaster}
165: \begin{eqnarray}
166: \frac{dP_{t}(S_{x})}{dt}&=&-\lambda \sum_{y \in {\cal N}(x)} P_{t}(S_{x},I_{y})+\gamma 
167: P_{t}(R_{x}), 
168: \label{mastera}
169: \\ 
170: \frac{dP_{t}(I_{x})}{dt}&=&\lambda \sum_{y \in {\cal N}(x)} P_{t}(S_{x},I_{y})-P_{t}(I_{x}), 
171: \label{masterb}
172: \\ 
173: \frac{dP_{t}(R_{x})}{dt}&=&P_{t}(I_{x})-\gamma P_{t}(R_{x}),
174: \label{masterc}
175: \end{eqnarray} 
176: \end{subequations}
177: Here ${\cal N}(x)$ is the neighborhood~(nearest neighbor sites) of a site $x$, 
178: $P_{t}(\alpha_{x})$ is the probability of having a state 
179: $\alpha$ at site $x$ at time $t$ and $P_{t}(\alpha_{x},\beta_{y})$ is 
180: the joint probability to have state $\alpha$ at site $x$ and state 
181: $\beta$ at site $y$, at time $t$. We always have 
182: $P_{t}(S_{x})+P_{t}(I_{x})+P_{t}(R_{x})=1$.
183: 
184: 
185: %[Intro to PA]
186: Eqs.~(\ref{mastera})-~(\ref{masterc}) are, as is usual for moment equations, 
187: not a closed system. One can extend them by including 
188: equations for the time evolution of $P_{t}(S_{x},I_{y})$ which 
189: in turn involve higher moments of 
190: the spatial correlations. This leads to an infinite hierarchy. 
191: To solve such a hierarchy one 
192: usually resorts to some approximation scheme which expresses 
193: the higher order moments in terms of 
194: the lower order ones and truncates the equations at some point 
195: ; this is referred to as the moment closure 
196: method~\cite{dickman:1986,matsuda:1992,benavraham:1992,petermann:2004,
197: levin:1995,levin:1996,levin:1997,levin:2000,
198: keeling:1999,geometry:2002,durrett:1998,schnaiz:2002}.
199: Both the MFA and PA are such schemes. In the MFA Eqs.~(\ref{mastera})-(\ref{masterc}) 
200: are closed by assuming that $P_{t}(S_{x},I_{y})=P_{t}(S_{x})P_{t}(I_{y})$, 
201: i.e., it neglects correlations between different sites. 
202: This leads to a pair of coupled equations which have been studied in \cite{murray:1980}.
203: In the PA scheme 
204: $P_{t}(\alpha_{x})$ and $P_{t}(\alpha_{x}, \beta_{y})$ are kept as unknowns while the 
205: higher-order moments are expressed, via some appropriate approximation, in terms of 
206: these quantities. 
207: 
208: 
209: To carry out the PA we complement Eq.~(\ref{allmaster}) by equations 
210: for the second moments $P_{t}(\alpha_{x},\beta_{y})$
211: for nearest neighbor sites $x$ and $y$ based on the transition rule that we have 
212: described in Eq.~(\ref{transition}). These are 
213: \begin{subequations}
214: \label{alltwomoment}
215: \begin{eqnarray}
216: \frac{dP_{t}(S_{x},I_{y})}{dt} &=& \gamma P_{t}(R_{x},I_{y})- ( \lambda+1 
217: )P_{t}(S_{x},I_{y})+ 
218: \sum_{w \in {\cal N}^{x}(y)} \lambda P_{t}(S_{x},S_{y},I_{w}) 
219: \label{twomomenta}
220: \\ 
221: &-& \sum_{w \in {\cal N}^{y}(x)} \lambda P_{t}(I_{w},S_{x},I_{y}), 
222: \nonumber
223: \\ 
224: \frac{dP_{t}(S_{x},R_{y})}{dt} &=& P_{t}(S_{x},I_{y})+\gamma 
225: P_{t}(R_{x},R_{y})-\gamma 
226: P_{t}(S_{x},R_{y})-
227: \sum_{w \in {\cal N}^{y}(x)} \lambda P_{t}(I_{w},S_{x},R_{y}), 
228: \label{twomomentb}
229: \\ 
230: \frac{dP_{t}(R_{x},I_{y})}{dt} &=& 
231: -(\gamma+1)P_{t}(R_{x},I_{y})+P_{t}(I_{x},I_{y})+
232: \sum_{w \in {\cal N}^{x}(y)} \lambda P_{t}(R_{x},S_{y},I_{w})
233: \label{twomomentc}
234: \end{eqnarray}
235: \end{subequations}
236: where ${\cal N}^{x}(y)$ is the set of nearest neighbor sites of $y$ excluding the site $x$. 
237: $P_{t}(\alpha_{x},\beta_{y},\chi_{w})$ is the joint probability to have state $\alpha$ 
238: at site $x$, state $\beta$ at site $y$ and state $\chi$ at site $w$ at time $t$.
239: For a derivation of Eq.~(\ref{alltwomoment})
240: see Appendix~\ref{derivation_twomoment}.
241:  
242: 
243: To close the system~(\ref{allmaster}) and~(\ref{alltwomoment}) and derive a set of autonomous 
244: equations for $P_{t}(\alpha_{x})$ and 
245: $P_{t}(\alpha_{x},\beta_{y})$ we approximate the triad joint probability 
246: $P_{t}(\alpha_{x},\beta_{y},\chi_{w})$ for $x$ and $w$ nearest neighbors 
247: of $y$, by the product of two pair probabilities 
248: $P_{t}(\alpha_{x},\beta_{y})$ and $P_{t}(\beta_{y},\chi_{w})$ divided
249: by the probability $P_{t}(\beta_{y})$
250: ~\cite{dickman:1986,matsuda:1992,levin:1996,benavraham:1992,petermann:2004}, i.e., we set
251: \begin{equation}
252: P_{t}(\alpha_{x}, \beta_{y}, \chi_{w}) = \frac{ P_{t}(\alpha_{x}, \beta_{y}) 
253: P_{t}(\beta_{y}, \chi_{w})} {P_{t}(\beta_{y})}
254: \label{myPA}
255: \end{equation}
256: Note that we have made use here of the structure of the hypercubic 
257: lattice. In such lattices, three adjacent sites, $x,y,w$ can not form a triangle 
258: but form only linear chains. This is not so in other lattices, 
259: e.g., the triangular lattice, where other configurations need also be considered.
260: 
261: 
262:  
263: While there are other choices for a PA, the approximation in Eq.~(\ref{myPA}) 
264: allows one to get the steady state solutions analytically. With other 
265: pair approximations~\cite{geometry:2002}, 
266: one has to solve the resulting differential equations numerically, making it 
267: impossible to obtain analytic expressions for the critical curve. 
268: 
269: 
270: To actually carry out computations with the PA 
271: we will assume from now on that our system is spatially uniform.  
272: The site $x$, in Eqs.~(\ref{allmaster}) 
273: and~(\ref{alltwomoment}), can now be chosen to be the origin. 
274: We also define: $P_{t}(S,I)=\frac{1}{z}\sum_{y \in {\cal N}(x)} P_{t}(S_{x},I_{y})$, 
275: $P_{t}(\alpha,\beta,\chi)=\frac{1}{z-1}\sum_{w \in {\cal N}^{x}(y)}
276: P_{t}(\alpha_{x},\beta_{y},\chi_{w})$ where $z=2d$ is the number 
277: of nearest neighbors of a site in the $d$-dimensional cubic lattice. 
278: The truncated equations for the PA-SIRS can now be written, by using the exact 
279: Eqs.~(\ref{allmaster})-~(\ref{alltwomoment}) and the approximate Eq.~(\ref{myPA}), 
280: as a closed set of five coupled equations, 
281: \begin{subequations} 
282: \label{allPASIRS}
283: \begin{eqnarray}
284: \frac{d P_{t}(I)}{dt}&=& z \lambda P_{t}(S,I) -P_{t}(I) 
285: \label{PASIRSa}
286: \\
287: \frac{d P_{t}(R)}{dt}&=& P_{t}(I)-\gamma P_{t}(R) 
288: \label{PASIRSb}
289: \\
290: \frac{d P_{t}(S,R)}{dt}&=& P_{t}(S,I)+\gamma (P_{t}(R)-P_{t}(R,I)-2P_{t}(S,R))
291: -\frac{(z-1) \lambda P_{t}(S,I)P_{t}(S,R)}{1-P_{t}(R)-P_{t}(I)} 
292: \label{PASIRSc}
293: \\ 
294: \frac{d P_{t}(R,I)}{dt}&=& -(2+\gamma )P_{t}(R,I)+P_{t}(I)-P_{t}(S,I)
295: +\frac{(z-1) \lambda P_{t}(S,I) P_{t}(S,R)}{1-P_{t}(R)-P_{t}(I)} 
296: \label{PASIRSd}
297: \\
298: \frac{d P_{t}(S,I)}{dt}&=& \gamma P_{t}(R,I) -(\lambda+1) P_{t}(S,I)
299: \label{PASIRSe} 
300: \\
301: &+&\frac{(z-1) \lambda P_{t}(S,I)}
302: {1-P_{t}(I)-P_{t}(R)}(1-P_{t}(R)-P_{t}(I)-P_{t}(S,R)-2P_{t}(S,I)). \nonumber 
303: \end{eqnarray}
304: \end{subequations}
305: Note that we always have $P_{t}(\alpha)=P_{t}(\alpha,S)+P_{t}(\alpha,I)+P_{t}(\alpha,R)$ 
306: which determines 
307: $P_{t}(I,I)$ and $P_{t}(S,S)$.
308: 
309: 
310: %[Check with PA-SIS]
311: In the limit $\gamma \rightarrow \infty$, 
312: $P_{t}(R)$ and $P_{t}(R,\alpha)$ as well as their time derivatives 
313: will go to zero. This yields $\gamma P_{t}(R)=P_{t}(I)$ and $\gamma P_{t}(R,I)=P_{t}(I)-P_{t}(S,I)$
314: ~\cite{murray:1980}. 
315: In this limit Eq.~(\ref{allPASIRS}) reduces to the PA 
316: equations of the SIS~considered in~\cite{levin:1996},
317: \begin{subequations}
318: \label{PASIS}
319: \begin{eqnarray}
320: \frac{d P_{t}(I)}{dt}&=& z \lambda P_{t}(S,I) -P_{t}(I), 
321: \label{PASISa}
322: \\ 
323: \frac{d P_{t}(S,I)}{dt}&=& P_{t}(I)-(\lambda+2)P_{t}(S,I)
324: +\frac{(z-1)\lambda P_{t}(S,I)}{1-P_{t}(I)}(1-P_{t}(I)-2P_{t}(S,I)).
325: \label{PASISb}
326: \end{eqnarray}
327: \end{subequations}
328: 
329: As already noted the MFA approximates the joint probability 
330: $P_{t}(S,I)$ in Eq.~(\ref{PASIRSa}) by the product, $P_{t}(S,I)=P_{t}(S)P_{t}(I)$. 
331: This leads to the closed set of MFA of equations
332: for the SIRS~\cite{murray:1980},
333: \begin{subequations}
334: \label{MFPASIRS}
335: \begin{eqnarray}
336: \frac{dP_{t}(S)}{dt}&=&-z \lambda P_{t}(S)P_{t}(I)+\gamma 
337: P_{t}(R)
338: \label{MFPASIRSa} 
339: \\ 
340: \frac{dP_{t}(I)}{dt}&=&z \lambda P_{t}(S)P_{t}(I)-P_{t}(I)
341: \label{MFPASIRSb}
342: \\
343: \frac{dP_{t}(R)}{dt}&=&P_{t}(I)-\gamma P_{t}(R)
344: \label{MFPASIRSc}
345: \end{eqnarray}
346: \end{subequations}
347: For $\gamma \rightarrow \infty$, $\gamma P_{t}(R) \rightarrow P_{t}(I)$ and 
348: $P_{t}(S) \rightarrow 1-P_{t}(I)$. Eq.~(\ref{MFPASIRS}) then reduces to the MFA 
349: for the SIS.
350: 
351: \section{\label{stationary}Stationary solutions of the PA-SIRS model}
352: 
353: Let us first consider the steady state solutions of 
354: the PA-SIS obtained  by setting the l.h.s of Eq.~(\ref{PASIS}) 
355: equal to zero~\cite{levin:1996}. This gives for the critical value of the PA-SIS 
356: epidemic process $\lambda_{c}(\infty)=1/(z-1)$.
357: For $\lambda \leq \lambda_{c}(\infty)$, 
358: both $P_{t}(I)$ and $P_{t}(S,I)$ $\rightarrow 0$ as $t \rightarrow \infty$ 
359: for all initial states.
360: When $\lambda > \lambda_{c}(\infty)$ there is, in addition to the disease-free 
361: state corresponding 
362: to $P(I)=0$, also a stationary state consisting of a finite fraction of infected 
363: individuals:
364: \begin{subequations}
365: \label{stationarySIS}
366: \begin{eqnarray}
367: \bar{P}(S,I)&=&\bar{P}(I)/(z\lambda)
368: \label{StationarySISa}
369: \\ 
370: \bar{P}(I)&=&\frac{z[(z-1)\lambda-1]}{z(z-1)\lambda-1}.
371: \label{stationarySISb}  
372: \end{eqnarray}
373: \end{subequations}
374: It is these non-zero steady states which are approached as $t \rightarrow \infty$ 
375: when starting from any initial state with $P_{0}(I)>0$.
376: 
377: 
378: The steady state solutions of the PA-SIRS system 
379: is obtained by setting the l.h.s. of Eq.~(\ref{allPASIRS}) equal to zero. 
380: Setting $x=\bar{P}(I)$ this yields, 
381: \begin{subequations}
382: \label{steadyPASIRS}   
383: \begin{eqnarray}
384: \bar{P}(R) &=& x/\gamma  
385: \label{steadyPASIRSa}
386: \\
387: \bar{P}(S,I) &=& x/(z \lambda)
388: \label{steadyPASIRSb}
389: \\ 
390: \bar{P}(S,R) &=& \frac{x(\frac{1}{z
391: \lambda}+\frac{1}{\gamma+1})}
392: {\gamma(1+\frac{1}{\gamma+1}+\frac{(z-1)x}{z(\gamma-(1+\gamma)x)})}
393: \label{steadyPASIRSc}
394: \\
395: \bar{P}(R,I) &=& \frac{x-\gamma 
396: \bar{P}(S,R)}{\gamma+1} \\ \nonumber 
397:                &=& \frac{x}{\gamma+1}
398: \Bigl (
399: 1-\frac
400: {
401: \frac{1}{z \lambda}+\frac{1}{\gamma+1}
402: }
403: {
404: 1+\frac{1}{\gamma+1}+\frac{(z-1)x}{z(\gamma-(\gamma+1)x)}
405: }
406: \Bigr )
407: \label{steadyPASIRSd}
408: \end{eqnarray} 
409: \end{subequations}
410: where $\bar{P}(\alpha,\beta)$ are the approximate probabilities for having states 
411: $\alpha$ and $\beta$ on neighboring sites. After further simplifications, 
412: we find that $x$ has to satisfy the cubic equation,
413: \begin{equation}
414: x(a_{1} x^{2}+a_{2} x+a_{3} )=0
415: \label{Pequation}
416: \end{equation}
417: Both the derivation of Eq.~(\ref{Pequation}) and the explicit  expressions for 
418: $a_{1}$, $a_{2}$ and $a_{3}$ as functions of $\lambda$ and $\gamma$ are given in 
419: Appendix~\ref{Pcoefficients}.
420: 
421: 
422: The root $x=0$ corresponds to the all healthy steady state, which is always 
423: a solution. The critical curve $\lambda_{c}(\gamma)$ is determined by 
424: the existence of a root of Eq.~(\ref{Pequation}) such that $x$ and all other 
425: stationary probabilities, are strictly positive.
426: It turns out that this strictly positive root is unique.
427: Thus when $\lambda \leq \lambda_{c}(\gamma)$, 
428: $x=0$ is the only steady state solution. For $\lambda> \lambda_{c}(\gamma)$, 
429: there is also a steady state in which the infection is endemic:
430: $\bar{P}(I)=\gamma \bar{P}(R)=x$ and $\bar{P}(S)=1-(1+1/\gamma)x$, see Appendix
431: ~\ref{Pcoefficients}.
432:   
433: 
434: The critical curve $\lambda_{c}(\gamma)$ is obtained in Appendix~\ref{Pcoefficients}. 
435: It is given by the equation,  
436: \begin{equation}
437: \lambda_{c}(\gamma)=\frac{\gamma+1}{2d-2+(2d-1)\gamma}, \textrm{  d=1,2,3,...}
438: \end{equation}
439: As $\gamma \rightarrow \infty$, $\lambda_{c}(\infty)=(2d-1)^{-1}$, the critical 
440: point of the PA-SIS epidemic 
441: process. On the other hand, as $\gamma$ approaches zero,
442: the critical curve shows different behavior depending on the 
443: dimension of the lattice: $\lambda_{c}(0)$ diverges to infinity for $d=1$, 
444: while $\lambda_{c}(0)$ is finite for $d\geq2$.
445: The PA thus reproduces the qualitative difference between the one and higher 
446: dimensional phase diagram of the SIRS model found in Ref.
447: ~\cite{kuulasmaa:1982, durrett:1991,andjel:1996,berg:1998}. 
448: 
449: 
450: The MFA, Eq.~(\ref{MFPASIRS}), yields the mean field critical value, $\lambda^{MF}_{c}=1/z$ 
451: independent of $\gamma$. 
452: In the coexistence region $\lambda>\lambda^{MF}_{c}$ the mean field stationary states are  
453: $\bar{P}(I)=\gamma \bar{P}(R)=\frac{\gamma (\lambda z-1)}{\lambda z(\gamma+1)}$ and 
454: $\bar{P}(S)=\frac{1}{z \lambda}$.
455: 
456: Both the steady state and critical value of the MFA and PA fail to correctly represent 
457: the results of the stochastic SIRS process for small $\gamma$: 
458: see Figs.~\ref{fig1} and \ref{fig2}. 
459: Note in particular that $\bar{P}(S)$ of the stochastic SIRS process is
460: considerably larger than that of the MFA or PA for large $\lambda$ and small $\gamma$. 
461: This is due to the fact that the susceptible sites can be surrounded by recovered ones 
462: and thus protected from contacting infected ones in the stochastic case.
463: 
464: \section{Comparison of the stochastic, the  PA and MFA steady states.}
465: 
466: %[Steady state of second moments]
467: We compare in Figs.~\ref{fig3}~-~\ref{fig6} the steady state values of 
468: $\bar{P}(\alpha)$ and $\bar{P}(\alpha,\beta)$ obtained from the MFA and PA 
469: with the results from the stochastic SIRS process
470: as a function of $\lambda$ at fixed values of $\gamma$.
471: Clearly the PA gives results closer to those obtained from the stochastic model.
472: For the methods used to obtain the steady state results from the numerical simulation, 
473: see appendix.~\ref{MCsimulation}. 
474: 
475: 
476: Figs.~\ref{fig3} and~\ref{fig5} show that both the MFA and PA overestimate  
477: $\bar{P}(I)$ as well as $\bar{P}(\alpha,I)$, $\alpha=S,R$.
478: This is due to the strong tendency of infected sites in the stochastic model 
479: to cluster into localized islands, reducing the contacts between $S$ and $I$.
480: This is partially taken into account by the PA as seen by the behavior of $\bar{P}(S,I)$ and 
481: $\bar{P}(I,I)$ 
482: in Figs.~\ref{fig3} and~\ref{fig5}. This clustering effect is also observed in the stochastic 
483: SIS process~\cite{levin:1996}. It is more pronounced in one dimension. 
484: 
485: 
486: Note that $\bar{P}(S,I)$ becomes zero both at 
487: $\lambda<\lambda_{c}(\gamma)$, when $\bar{P}(I)=0$, and 
488: at $\lambda=\infty$ when $\bar{P}(S)=0$, reaching 
489: a peak at a positive value of $\lambda$ which depends on $\gamma$. 
490: For large values of $\gamma$, the steady state values of $\bar{P}(\alpha)$ and 
491: $\bar{P}(\alpha,\beta)$ obtained from the PA, or the MFA agree well with the 
492: numerical simulation, away from the critical $\lambda_{c}(\gamma)$. 
493: Moreover the PA yields steady state curves remarkably similar to those from 
494: the numerical simulation, see Figs.~\ref{fig4} and~\ref{fig6}.
495: 
496: 
497: \section{Linear stability analysis of the Pair approximation}
498: 
499: To study the stability of the stationary PA state, 
500: Eq.~(\ref{allPASIRS}) is linearized about the steady state values
501: ~\cite{murray:1980}, see Appendix~\ref{PAjacobian}. 
502: This leads to the study of the roots of the characteristic fifth order polynomial $P_{5}(\xi)$, 
503: obtained from $|A-\xi I |=0$ where $A$ is the Jacobian of the linearized PA-SIRS system.
504: If $Re \xi <0$, the solution of the linearized equation is stable, i.e., 
505: a small perturbation around the steady state will decay back 
506: to the steady state. 
507: We used the Routh-Hurwitz conditions~\cite{murray:1980} to obtain the sign of 
508: the real part of eigenvalues of the Jacobian.  
509: As expected, the positive steady state solution is stable for $\lambda>\lambda_{c}(\gamma)$. 
510: The zero steady state solution is stable for $\lambda \leq \lambda_{c}(\gamma)$ 
511: and unstable for $\lambda > \lambda_{c}(\gamma)$. 
512: 
513: 
514: The eigenvalues of $P_{5}(\xi)$ have non-zero 
515: imaginary parts in some regions of the parameter space. 
516: In such regions the PA-SIRS system in 
517: Eq.~(\ref{allPASIRS}) will converge to the steady state in a damped 
518: oscillatory manner. Such oscillations are seen in 
519: Fig.~\ref{fig7} and \ref{fig8}.
520: 
521: 
522: \section{Time dependent behavior}
523: 
524: To study the time evolution of an epidemic following an initial infection of 
525: a healthy population we performed dynamical 
526: Monte Carlo simulations~\cite{grassberger:1979} as well as solutions of 
527: Eqs.~(\ref{allPASIRS}) and (\ref{MFPASIRS}). For the stochastic evolution we started 
528: with infected sites placed either randomly or in a cluster and followed 
529: the time evolution averaged over $10^{3}$ realizations of the SIRS process.
530: To obtain the time evolution of the MFA and PA we solved Eq.~(\ref{MFPASIRS}) 
531: and Eq.~(\ref{allPASIRS}) numerically by using a 4th-order 
532: Runge-Kutta method. We plot the results in Figs.~\ref{fig7} and~\ref{fig8}. 
533: 
534: 
535: To set the unit of time of the simulation 
536: we started with a fully infected state, $P_{0}(I)=1$ and $\lambda=0$ and obtained 
537: the exponentially decaying pattern of $P_{t}(I)$. We then set the slope (death rate) 
538: of the graph, 
539: $log P_{t}(I)$ vs $t$, from the numerical simulation equal to those from the MFA and PA. 
540:  
541: 
542: Starting with a small value of $P_{0}(I)$, $P_{t}(I)$ displays  
543: an initial "exponential" growth in both the MFA and PA. 
544: Similar growth patterns are observed in all $P_{t}(\alpha,I)$, $\alpha=S,I,R$.
545: This is explained by the initially abundantly available susceptible population. 
546: Once the susceptible 
547: population is reduced, the infected population reaches a maximum and then 
548: decreases to the steady state 
549: endemic level. Note the damped oscillatory pattern in Figs.~\ref{fig7} and \ref{fig8} 
550: for this choice of the parameters ($\lambda$,$\gamma$).
551: 
552: 
553: The numerical simulation of the stochastic time evolution does not show the pronounced 
554: growth patterns of the PA and MFA when the initial fraction of infected sites is small, 
555: as seen in Fig.~\ref{fig8}. The formation of clusters of infected sites makes the 
556: infected population grow more slowly in the stochastic model. 
557: When the initial fraction of infected population 
558: increases to more than one percent the stochastic model shows significant change 
559: in its growth pattern, becoming similar to the PA and MFA. 
560: If however the same fraction of infected sites are initially placed in a single cluster 
561: the stochastic epidemic process exhibits slower growth patterns, similar to those starting with 
562: a small fraction of initially infected sites. These studies confirm that the clustering 
563: of infected sites in the stochastic model reduces both the speed of growth 
564: and the maximum fraction of infected sites. In realistic situations the population 
565: is not well mixed so we would expect 
566: growth patterns more similar to that of the stochastic epidemic model, starting with 
567: a fraction of infected sites initially placed in a single cluster.
568: 
569: 
570: \section{Summary}
571: 
572: 
573: We investigated the stochastic SIRS epidemic process and compared the results with those
574: obtained from the deterministic MFA and PA. These approximations close the hierarchy of  
575: dynamical equations by expressing the higher order moments in terms of the lower order ones.
576: The PA is found to improve over the MFA both for the stationary and 
577: for the time dependent states. The time evolution 
578: of the system shows damped oscillatory 
579: behavior in some parameter ranges.
580: 
581: 
582: \begin{acknowledgments}
583: Work supported by NSF DMR-01-279-26 and by AFOSR AF 49620-01-1-0154 and  
584: by DIMACS grants NSF DBI 99-82983 and NSF EIA 02-05116.
585: \end{acknowledgments}
586: 
587: 
588: \appendix
589: 
590: \section{\label{derivation_twomoment}Derivation of differential 
591: equation for $P_{t}(S_{x},I_{y})$}
592: 
593: Eq.~(\ref{twomomenta}) is derived by considering all transitions leaving or entering the pair 
594: configuration $(S_{x},I_{y})$. We list them as follows: A pair $(R_{x},I_{y})$ 
595: changes to a pair $(S_{x},I_{y})$ with a rate $\gamma$. A pair $(S_{x},I_{y})$ changes to 
596: a pair $(I_{x},I_{y})$ with a rate $\lambda$ and also changes to a pair $(S_{x},R_{y})$ 
597: with a rate 1. 
598: A triad configuration $(S_{x},S_{y},I_{w})$ 
599: transits to a triad $(S_{x},I_{y},I_{w})$ with 
600: a rate $\lambda$ such that a pair configuration $(S_{x},S_{y})$ is changed to $(S_{x},I_{y})$. 
601: A triad $(I_{w},S_{x},I_{y})$ 
602: changes to a triad $(I_{w},I_{x},I_{y})$ with a rate $\lambda$. 
603: The equations for $P_{t}(S_{x},R_{y})$ and $P_{t}(R_{x},I_{y})$ 
604: in Eq.~(\ref{alltwomoment}) can be obtained in a similar way.
605: The relation $P_{t}(\alpha_{x})=P_{t}(\alpha_{x},\alpha_{y})
606: +P_{t}(\alpha_{x},\beta_{y})+P_{t}(\alpha_{x},\chi_{y})$ can be used to obtain the 
607: other joint probabilities $P_{t}(\alpha_{x}, \beta_{y})$ 
608: which are not shown in Eq.~(\ref{alltwomoment}).
609: 
610: 
611: \section{\label{Pcoefficients}Derivation of Eq.~(\ref{Pequation})}
612: The steady states in Eq.~(\ref{steadyPASIRS}) are obtained by setting the l.h.s. of 
613: Eq.~(\ref{PASIRSa})-(\ref{PASIRSd}) equal to zero. In addition we set Eq.~(\ref{PASIRSe}) 
614: equal to zero
615:  and replace a single site and joint probabilities with the steady states in 
616: Eq.~(\ref{steadyPASIRS}). After simplifications, we obtain Eq.~(\ref{Pequation})
617: with the coefficients, 
618: \begin{eqnarray} 
619: a_{1}&=&
620: \gamma^{3} \{ z^{2}(z-1)\lambda-z \}
621: +\gamma^{2} \{ z(2z^{2}-2z-1)\lambda -2z-1\} \label{Pcoefficient}
622: \\ \nonumber 
623: &+&\gamma \{ 2z(z^{2}-z-1) \lambda -2z-1 \} 
624: +z \{(z^{2}-z-1)\lambda-1\}\\ \nonumber 
625: a_{2}&=&
626: z\gamma \Bigl \{
627: \gamma^{2} \{ z+1-2z(z-1)\lambda \}
628: +\gamma \{z+3-(3z^{2}-4z-1)\lambda \}
629: +z+1-(2z^{2}-3z-1)\lambda
630: \Bigr \}\\ \nonumber
631: a_{3}&=&
632: z^{2}\gamma^{2} \Bigl \{
633: \gamma\{-1+\lambda(z-1)\}-1+\lambda(z-2)
634: \Bigl \}.
635: \end{eqnarray}
636: 
637: The critical curve $\lambda_{c}(\gamma)$ is given by setting 
638: $a_{3}=0$. Only for $\lambda>\lambda_{c}(\gamma)$ does the quadratic factor of 
639: Eq.~(\ref{Pequation}) have a positive root. 
640: 
641: \section{\label{PAjacobian} The Jacobian of the linearized PA-SIRS}
642: The Jacobian of the linearized PA-SIRS is written,
643: \[
644: \mathbf{A}=
645: \left(
646: \begin{array}{ccccc}
647: -\gamma & 1 & 0 & 0 & 0 \\
648: 0 & -1 & z\lambda & 0 & 0 \\
649: -K_{2}K_{0} 
650: & -K_{2}K_{0}
651: & K_{3}
652: & -\frac{K_{1}}{\bar{P}(SR)} & \gamma \\
653: \gamma-K_{2} & -K_{2} & 1-\frac{K_{1}}{\bar{P}(IS)} 
654: & -2\gamma-\frac{K_{1}}{\bar{P}(SR)} & -\gamma \\
655: K_{2} &  1+K_{2}  & -1+\frac{K_{1}}{\bar{P}(IS)} 
656: &  \frac{K_{1}}{\bar{P}(SR)}  &  -\gamma-2
657: \end{array} \right)
658: \]
659: where 
660: $K_{0}=1+2\frac{\bar{P}(IS)}{\bar{P}(SR)}$, 
661: $K_{1}=\frac{(z-1)\lambda 
662: \bar{P}(IS)\bar{P}(SR)}{1-\bar{P}(R)-\bar{P}(I)}$, 
663: $K_{2}=\frac{(z-1)\lambda
664: \bar{P}(IS)\bar{P}(SR)}{(1-\bar{P}(R)-\bar{P}(I))^{2}}$, and  
665: $K_{3}=(z-2)\lambda-1-K_{1}(\frac{1}{\bar{P}(IS)}+\frac{4}{\bar{P}(SR)})
666: $.
667: 
668: 
669: \section{\label{sec:stabilityMFSIRS}Linear Stability Analysis of the MF-SIRS}
670: The Jacobian matrix B of linearized MF-SIRS is given by~\cite{murray:1980}
671: \[
672: \mathbf{B}=
673: \left( \begin{array}{cc}
674: -\lambda z \bar{P}(I)-\gamma & -\lambda z \bar{P}(S)-\gamma \\
675: \lambda z \bar{P}(I) & \lambda z \bar{P}(S)-1 
676: \end{array}
677: \right)
678: \]
679: The characteristic polynomial of the second order, $P_{2}(\xi)=\xi^{2}+a_{1}\xi+a_{2}=0$, 
680: is obtained from $|B-\xi I|=0$. 
681: 
682: 
683: The necessary and sufficient (Routh-Hurwitz) conditions~\cite{murray:1980} 
684: for $Re\xi<0$ is $a_{2}>0$ and $a_{1}>0$.
685: In the coexistence region where $z\lambda> 1$, 
686: $a_{2}=\gamma (z \lambda -1)>0$ and 
687: $a_{1}=\frac{\gamma}{\gamma+1}(\gamma+z\lambda)
688: >0$ for all $\gamma>0$.
689: In the no-coexistence region where 
690: $z\lambda<1$, 
691: $a_{2}=\gamma(1-z\lambda)>0$ and $a_{1}=\gamma +(1-z \lambda)>0$ 
692: for all $\gamma>0$. Both in the coexistence and 
693: no-coexistence region, the real part of the eigenvalues is negative 
694: and thus the mean field steady states are stable. 
695: 
696: 
697: Now we turn our attention to the oscillatory behavior.
698: The eigenvalues of the characteristic polynomial $P_{2}(\xi)$ is given by, 
699: \begin{equation}
700: \xi_{\pm}=\frac{-\gamma^{2}-z \gamma \lambda \pm 
701: \sqrt{ (2\gamma z^{2}\lambda^{2}-2z\lambda (\gamma^{2}+2z\gamma \lambda 
702: +2)+\gamma^{3}+4\gamma^{2}+8\gamma+4) \gamma }}
703: {2(\gamma+1)}
704: \end{equation}
705: In the range of $\lambda_{-}(\gamma)<\lambda(\gamma)<\lambda_{+}(\gamma)$
706: the imaginary part of the eigenvalues is non-zero: 
707: $\lambda_{\pm}(\gamma)=\frac{2+4\gamma+\gamma^{2}\pm 
708: 2(1+\gamma)^{3/2}}{z\gamma}$. In this range of $\lambda$, the steady states 
709: correspond to the stable spiral and the system converges to the steady state in 
710: damped oscillatory pattern. Even in the damped oscillatory region, any 
711: oscillation is hardly visible in the large $\gamma$ limit and becomes noticeable 
712: only in small $\gamma$ limit.
713: 
714: 
715: \section{\label{MCsimulation}Monte Carlo simulation}
716: 
717: The numerical simulations described here used lattices with 
718: periodic boundary conditions. In one dimension, rings of 
719: $5000 \leq N \leq 15000$ sites were used. 
720: In two dimensions, torii of $50^{2} \leq N \leq 200^{2}$ sites were employed. 
721: 
722: 
723: To obtain the steady state of the SIRS process a random initial 
724: configuration of susceptible and infected sites is evolved according to 
725: the transition rates in Eq.~(\ref{transition}). 
726: In practice a site is randomly chosen and a random number~($\in [0,1]$) is 
727: also chosen: if it is greater than the given transition probability 
728: for that site, which is equal 
729: to the rate$\times \Delta t$, its state is updated: 
730: $\Delta t$ is chosen to be so small that transition
731: probability is not greater than 1 for a range of ($\lambda, \gamma$)
732: ~\cite{marro:1999, durrett:1994b}. Otherwise its state remains the same. 
733: 
734: 
735: For a finite system the only true stationary state of the SIRS process is 
736: the absorbing state corresponding to $P(S)=1$, $P(I)=P(R)=0$. To learn about 
737: the active state from simulations of 
738: a finite system we study the quasi-stationary state. These are determined 
739: from averages over the surviving representatives of $10^{3}$-$10^{4}$ 
740: independent realizations of the SIRS process with the same parameter $(\lambda,\gamma)$, 
741: beginning with random initial distribution of the $I$'s. 
742: Surviving sample averages converge to stationary values as $N \rightarrow \infty$. 
743: To obtain the steady states and critical curve 
744: we extrapolated quasi-stationary values of finite systems to 
745: those of the infinite system.
746: 
747:  
748: The finite size scaling theory~\cite{marro:1999} can be used to 
749: obtain the critical curve $\lambda^{z}_{c}(\gamma)$. We can 
750: assume a scaling function of the surviving probability: 
751: $P_{t}(I)\sim t^{-\beta/\nu_{\|}}
752: f((\lambda-\lambda_{c})t^{1/\nu_{\|}})$. 
753: At criticality, $\lambda=\lambda_{c}(\gamma)$, 
754: the survival probability of the infection, starting from a single infected 
755: site, has a power law behavior in time. In the subcritical region, 
756: it decays exponentially while in the supercritical region it 
757: reaches non-zero steady state in a short time. The power law behavior 
758: of the survival probability at criticality enables one to extract 
759: the critical curve $\lambda^{z}_{c}(\gamma)$ from the time evolution data 
760: of the SIRS process. This dynamical Monte Carlo simulation 
761: is reliable when the system size is sufficiently large so that 
762: the evolution of the system is approximately confined, for the duration of the 
763: simulation to a region smaller than the size of the system~\cite{grassberger:1979}. 
764: However we found that this surviving probability oscillates 
765: wildly when $\gamma$ is small. Because of this 
766: the dynamical Monte Carlo method is not used to determine the critical curve near $\gamma=0$.  
767: 
768: 
769: \begin{thebibliography}{99}
770: 
771: \bibitem{anderson:1992} R. M. Anderson and R. M. May, 
772: {\sl Infectious Diseases of Humans: Dynamics and Control}, 
773: Oxford university Press, Oxford, 1992.
774: 
775: \bibitem{diekmann:2000} O. Diekmann and J. A. P. Heesterbeek,
776: {\sl Mathematical Epidemiology of Infectious Diseases: model building, 
777: analysis and interpretation}, Willey, New York, 2000
778: 
779: \bibitem{neuhauser:2000} S. M. Krone and C. Neuhauser, {\sl J. Appl. Prob.} 
780: {\bf 37}, 1044(2000). 
781: 
782: \bibitem{Jeger:1990} J. Kranz, ed., {\sl Comparative Epidemiology of Plant Diseases}, 
783: New York, 2002. 
784: 
785: \bibitem{durrett:1994a} R. Durrett and S. Levin,
786: {\sl Theor. Popul. Biol.}, {\bf 28}, 263(1994)
787: 
788: \bibitem{liggett:1985} T. M. Liggett,
789: {\sl Interacting Particle Systems}, Springer-Verlag, New York(1985).
790: 
791: \bibitem{durrett:1988b} R. Durrett, {\sl Lecture Notes on Particle Systems and 
792: Percolation}, Wadsworth Publishing, 1988.
793:                                                                                
794: \bibitem{marro:1999} J. Marro and R. Dickman,
795: {\sl Nonequilibrium Phase Tranistions in Lattice Models}, Cambridge 
796: University Press, Cambridge, 1999.
797: 
798: \bibitem{lebowitz:1985-86a} A. De Masi, P. A. Ferrari and J. L. Lebowitz, 
799: {\sl Phys. Rev. Lett.} {\bf 55}, 1947(1985). 
800: {\sl J. Stat. Phys.} {\bf 44}, 589(1986). 
801: 
802: \bibitem{lebowitz:1986b} J. L. Lebowitz, Physica, {\bf 140A}, 232(1986).
803: 
804: \bibitem{lebowitz:1988} J. L. Lebowitz, E. Presutti and H. Spohn, 
805: {\sl J. Stat. Phys.} {\bf 51}, 841(1988). 
806: 
807: \bibitem{durrett:1994} R. Durrett and C. Neuhauser, {\sl Ann. Prob.}, {\bf 22}, 289(1994). 
808:  
809: \bibitem{bramson:1997} M. Bramson, {\sl Ann. Appl. Prob.} {\bf 7}, 565(1997). 
810: 
811: \bibitem{dickman:1986} R. Dickman, 
812: {\sl Phys. Rev. A} {\bf 34} 4246(1986) and {\sl Phys. Rev.A} {\bf 38}, 2588(1988).
813: 
814: \bibitem{matsuda:1992} H. Matsuda, N. Ogata, A. Sasaki, and K. Sato,
815: {\sl Prog. theor. Phys} {\bf 88}, 1035(1992).
816: 
817: \bibitem{levin:1996} S. Levin and R. Durrett, 
818: {\sl Phil. Trans. R. Soc. Lond.}, {\bf 351}, 1615(1996).
819: 
820: \bibitem{benavraham:1992} D. ben-Avraham and J. Kohler, {Phys. Rev. A} {\bf 45}, 
821: 8358(1992).
822: 
823: \bibitem{petermann:2004} T. Petermann and Paolo De Los Rios, Los Alamos archive, 
824: q-bio.PE/0401028. 
825: 
826: \bibitem{levin:1995} D. Mollison and S. Levin, in 
827: {\sl Ecology of Infectious Diseases in Natural Populations}, 384, B. T. Grenfell 
828: and A. P. Dobson, eds., Cambridge University, Cambridge, 1995.
829: 
830: \bibitem{levin:1997} D. Tilman and P. Kareiva, eds., 
831: {\sl Spatial Ecology: The Role of Space in Population Dynamics and 
832: Interspecific Interactions}, Princeton university, Princeton, 1997.
833: 
834: \bibitem{durrett:1998} R. Durrett and S. Levin, {\sl Theor. Popul. Biol.} {\bf 53}, 30(1998). 
835: 
836: \bibitem{keeling:1999} M. J. Keeling, {\sl Proc. R. Soc. Lond. B}  {\bf 266}, 859(1999), 
837: {\sl Proc. R. Soc. Lond. B}  {\bf 266}, 953(1999) and {\sl Ecol. Lett.} {\bf 5}, 20(2002). 
838: 
839: \bibitem{levin:2000} A. Gandhi, S. Levin and S. Orszag, {\sl Bulletin of 
840: Mathematical Biology} {\bf 62}, 595(2000).
841: 
842: \bibitem{schnaiz:2002} R. B. Schnazi, {\sl Theor. Popul. Biol.} {\bf 61}, 163(2002). 
843: 
844: \bibitem{geometry:2002} U. Diekmann, R. Law and J. A. Metz, eds., {\sl The Geometry 
845: of Ecological Interactions: Simplifying Spatial Complexity}, Cambridge University Press, 
846: Cambridge, 2000.
847: 
848: \bibitem{berg:1998} J. Van Den Berg, G. Grimmett and R. Schinazi, 
849: {\sl Ann. Appl. Prob.}, {\bf 8}, 317(1998). 
850: 
851: \bibitem{andjel:1996} E. Andjel and R. Schinazi, 
852: {\sl J. Appl. Prob.}, {\bf 33}, 741(1996). 
853: 
854: \bibitem{durrett:1991} R. Durrett and C. Neuhauser,
855: {\sl Ann. Appl. Prob.}, {\bf 1}, 189(1991). 
856: 
857: \bibitem{kuulasmaa:1982} K. Kuulasmaa,
858: {\sl J. Appl. Prob.}, {\bf 25}, 745(1982).
859: 
860: \bibitem{huang} K. Huang, {\sl Statistical Mechanics}, Wiley, New York(1987).
861: 
862: \bibitem{murray:1980} J. D. Murray,
863: {\sl Mathematical Biology}, Springer-Verlag, New York(1980).
864: 
865: \bibitem{durrett:1994b} R. Durrett and S. Levin,
866: {\sl Phil. Trans. R. Soc. Lond. B} {\bf 343}, 329(1994). 
867: 
868: \bibitem{grassberger:1979} P. Grassberger and A. de la Torre, {\sl Ann. Phys.}
869: {\bf 122}, 373(1979).
870: 
871: \end{thebibliography}{99}
872: 
873: \begin{figure}
874: \begin{center}
875: \includegraphics[height=10cm,width=10cm,angle=0]{Fig1.eps}
876: %{Fig.091703-1.eps}
877: \caption{\label{fig1}Phase diagram of the SIRS process in two dimensions.
878: The coexistence phase of $S$-$I$-$R$ and the no-coexistence phase are 
879: separated by the critical curve from the simulation~(open circles with dotted line for 
880: eye-guidance), the PA (thick solid line) and the MFA (long dashed line). 
881: The critical curve is obtained on periodic square lattice of different sizes $N$ 
882: from simulations extrapolated to infinite system : $N=50^{2},70^{2},100^{2},150^{2},200^{2}$. 
883: }
884: \end{center}
885: \end{figure}
886: 
887: \newpage
888: 
889: \begin{figure}
890: \begin{center}
891: \includegraphics[height=10cm,width=10cm,angle=0]{Fig2.eps}
892: %{Fig.091903-1.eps}
893: \caption{\label{fig2}Phase diagram of the SIRS process in one dimension. 
894: The critical curve from numerical simulations of ring lattice of different sizes $N$ 
895: is extrapolated to infinite system: $N=5000,7000,10000,15000$. The same symbols are used as 
896: in the Fig.~\ref{fig1}. 
897: }
898: \end{center}
899: \end{figure}
900: 
901: \newpage
902: 
903: \begin{figure}
904: \begin{center}
905: \includegraphics[height=10cm,width=10cm,angle=0]{Fig3.eps}
906: %{Fig.100903-1.eps}
907: \caption{\label{fig3}First and second order moments of the steady state SIRS 
908: in two dimensions at $\gamma=0.2$. 
909: The steady-state values of the density of infection in Fig.~\ref{fig3}(a) 
910: and the second moments in Fig.~\ref{fig3}(b)-(f) 
911: are drawn from the numerical simulation~(open circle with dotted line 
912: for eye-guidance), the PA~(thick solid line), and the MFA~(long-dashed line). 
913: For the numerical simulation we used a system of size $N=100^{2}$.}
914: \end{center}
915: \end{figure}
916: 
917: \newpage
918: 
919: \begin{figure}
920: \begin{center}
921: \includegraphics[height=10cm,width=10cm,angle=0]{Fig4.eps}
922: %{Fig.101503-3.eps}
923: \caption{\label{fig4}First and second order moments of the steady state SIRS 
924: in two dimension at $\gamma=2$. The same symbols are 
925: used as in the Fig.~\ref{fig3}.}
926: \end{center}
927: \end{figure}
928: 
929: \newpage
930: 
931: \begin{figure}
932: \begin{center}
933: \includegraphics[height=10cm,width=10cm,angle=0]{Fig5.eps}
934: %{Fig.101503-1.eps}
935: \caption{\label{fig5}First and second order moments of the steady state SIRS 
936: process in one dimension at $\gamma=1$. 
937: The same symbols are used as in the Fig.~\ref{fig3}.}
938: \end{center}
939: \end{figure}
940: 
941: \newpage
942: 
943: \begin{figure}
944: \begin{center}
945: \includegraphics[height=10cm,width=10cm,angle=0]{Fig6.eps}
946: %{Fig.101503-2.eps}
947: \caption{\label{fig6}First and second order moments of the steady state SIRS 
948: process in one dimension at $\gamma=4$. The same 
949: symbols are used as in the Fig.~\ref{fig3}.}
950: \end{center}
951: \end{figure}
952:             
953: 
954: \newpage
955: 
956: \begin{figure}
957: \begin{center}
958: \includegraphics[height=10cm,width=10cm,angle=0]{Fig7.eps}
959: %{Fig.031204-4.eps}
960: \caption{\label{fig7}Time-evolution of the first and the second order moments of 
961: the SIRS process in two dimensions. All sub-graphs are from numerical 
962: simulations (open circles), the PA~(solid line), and the MFA~(dashed line) 
963: at $\gamma=0.2$ and $\lambda=2$. A periodic square lattice of $N=10^{4}$ sites is used 
964: in the numerical simulations averaged over $10^{3}-10^{4}$ realizations starting with 
965: random initial distribution with 1 percent of infected sites. }
966: \end{center}
967: \end{figure}
968:             
969: 
970: \newpage
971: 
972: \begin{figure}
973: \begin{center}
974: \includegraphics[height=10cm,width=10cm,angle=0]{Fig8.eps}
975: %{Fig.031204-1.eps}
976: \caption{\label{fig8}Time-evolution of fraction of infected sites of the 
977: SIRS process in two dimensions at $\gamma=0.2$ and $\lambda=2$. 
978: A periodic square lattice of $N=100^{2}$ is used in numerical 
979: simulation averaged over $10^{3}-10^{4}$ realizations.
980: Main: Simulation starts with 1$\%$ of infected sites placed either randomly
981: ~(filled circles) or in a single cluster~(open circles) on a lattice.
982: Both the PA and MFA takes an initial value 0.01 for 
983: $P_{0}(I)$. 
984: Inset: Simulation starts with different fractions of infected sites 
985: randomly placed in a lattice: 0.1, 1 and 5 percents of the system.}
986: \end{center}
987: \end{figure}
988: 
989: \end{document}
990: 
991: 
992: 
993: 
994: 
995: 
996: 
997: 
998: 
999: 
1000: 
1001: 
1002: 
1003: 
1004: