1: 

2: \documentclass[prl,twocolumn,showpacs]{revtex4}

3: \usepackage{graphicx}

4: \usepackage{amsmath}

5: 

6: 

7: \begin{document}

8: 

9: \title{Novel exponents control the  quasi-deterministic limit of the extinction transition}

10: 

11: 

12: \author{David A. Kessler}

13: \affiliation{Department of Physics, Bar-Ilan University, Ramat-Gan IL52900, ISRAEL}

14: \author{Nadav M. Shnerb}

15: \affiliation{Department of Physics, Bar-Ilan University, Ramat-Gan IL52900, ISRAEL}

16: \pacs{05.70.Ln,02.50.Ey,64.60.Ht,87.19.X-}

17: 

18: 

19: 

20: \begin{abstract}

21: The quasi-deterministic limit of the generic extinction transition

22: is considered within the framework of standard epidemiological

23: models. The susceptible-infected-susceptible (SIS) model is known to exhibit

24: a transition from extinction to spreading, as the infectivity is increased, described by the

25: directed percolation equivalence class.  We find that  the distance from the transition point, and the prefactor controlling the divergence of the (perpendicular) correlation length, both

26: scale with the local population size, $N$, with two novel universal exponents. Different

27: exponents characterize the large $N$ behavior of the

28: susceptible-infected-recovered (SIR) model, which belongs to the

29: dynamic percolation class. Extensive numerical studies in a range of

30: systems lead to the conjecture

31: that these characteristics are generic and may be used in order to

32: classify the high density limit of any stochastic process on the

33: edge of extinction.

34: \end{abstract}

35: \maketitle

36: 

37: \section {Introduction}

38: Perhaps the most useful approximation in science is the description

39: of microscopic, stochastic processes via "mean field" deterministic

40: equations. Except for the most fundamental parts of particle

41: physics, almost any other study involves that procedure: classical

42: mechanics is used to describe quantum systems, classical

43: electromagnetism neglects the fluctuations of the photon density and

44: continuum mechanics averages out the microscopic stochasticity

45: involved in the motion of an individual molecule. Chemical reaction

46: kinetics is generally described by rate equations; population

47: dynamics and other ecological processes are depicted using the

48: concept of logistic growth, or by Lotka-Volterra equations. The

49: dynamics of the individual, microscopic, constituents is always

50: stochastic, and is subject to fluctuations; these fluctuations are

51: smeared out when the system is described by deterministic rate

52: equations.

53: 

54: Basically, the underlying assumption beyond all these approximations

55: is that the microscopic fluctuations are averaged out in the ``large

56: density'' (many atoms, animals, quanta) limit. A generic analysis of

57: the deterministic limit, like a $1/N$ expansion where $N$ is the

58: number of elements, is still lacking in many fields. Problems like

59: the quantum classical correspondence in chaotic systems, or the

60: decay of a quasistationary state to an absorbing state are still a

61: subject of intensive studies. The situation becomes even more

62: complicated when spatially extended systems, made of diffusively

63: coupled patches, are considered. What exactly determine the ``large

64: N'' limit in that system? Should the number of microscopic entities

65: be large on a single patch, or within a correlation length? All

66: these questions are still open, as no systematic perturbation theory

67: in $1/N$ exists so far.

68: 

69: In this paper we consider a generic process in population dynamics:

70: the extinction transition with an absorbing state. The framework

71: used is two well-known models for epidemics, namely, the SIS and the

72: SIR models \cite{sa,am}. For a well mixed population of size $N$

73: (say, on a single patch) the stochastic process starts by the

74: introduction of a single infected (``I'') individual into the system.

75: This individual may infect any other, susceptible (``S''), individual

76: with rate $\alpha/N$, where after the infection the susceptible

77: becomes infected and may infect other susceptible members of the

78: community; the only other process is a "recovery" of an infected

79: person; this happens with rate $\beta$. In the SIR model, the

80: recovered individuals are then immune against the disease, while the

81: SIS model describes the case where the recovered became susceptible

82: again. The elementary processes, thus, are

83: \begin{eqnarray} \label{eq1}

84:  S+I \stackrel{\alpha/N}{\rightarrow} 2I  \qquad I \stackrel{\beta}{\rightarrow} \emptyset \qquad (SIR) \nonumber \\

85:   S+I \stackrel{\alpha/N}{\rightarrow} 2I  \qquad I

86: \stackrel{\beta}{\rightarrow} S \qquad (SIS)

87:  \end{eqnarray}

88: Denoting by $I$ the fraction of infected individuals, and by $S$ the

89: fraction of susceptibles,  the mean field equations in the well mixed

90: limit are

91: \begin{eqnarray} \label{eq2}

92: \frac{dI}{dt} = \alpha I S - \beta I  &\qquad& \frac{dS}{dt} = -\alpha I S + \beta I  \quad  (SIS) \nonumber \\

93: \frac{dI}{dt} = \alpha I S - \beta I &\qquad& \frac{dS}{dt} =

94: -\alpha I S \quad  (SIR).

95:  \end{eqnarray}

96: Clearly, the SIR process is self-limiting, as $S$ decreases with

97: time, while the SIS process may support an endemic state with $I =

98: 1- \alpha / \beta$. Another piece of information garnered from the

99: rate equations is the existence of a transition when the infectivity parameter  $R_0 \equiv

100: \alpha  / \beta$ crosses the value $R_c = 1$. Below $R_0=R_c$, both SIS and SIR processes immediately

101: decay; above that value, an outbreak is possible.

102: 

103: The mean field equations (\ref{eq2})  are an approximation for the

104: real stochastic process (\ref{eq1}). Some characteristics of the

105: epidemic, though,  may be calculated exactly for the real stochastic

106: process in the well mixed limit. In particular,

107: 

108: $ \bullet$ Below $R_c$ both

109: processes are the same in the large $N$ limit, since the change of a susceptible becoming reinfected in the SIS model is vanishingly small.

110: 

111:  $ \bullet$ For large $N$, above $R_c$ there are two peaks in the distribution of epidemic sizes.  The first peak is at size 1 (i.e. the initial infected individual, only).  The second peak for the

112:  SIR model is at the size predicted by the deterministic equations~\cite{watson}.  For the SIS model,

113:  the second peak is at an exponentially large value, corresponding to the exponentially

114:  long lifetime of the metastable endemic state\cite{mysis}.

115: 

116: $ \bullet$ At $R_c$, the average size of the epidemic scales for large $N$ as $N^{1/2}$

117: for the SIS model~\cite{dlsis,mysis} and as $N^{1/3}$ for the SIR model~\cite{lof,bn,mysir}.

118: 

119:  Let us consider now  a one dimensional array of L patches, with $N$

120:  susceptible individuals on each patch, where a single infected individual is

121:  introduced at a single patch.

122: 

123: Above $R_c$, and for $N \to \infty$,  the deterministic equations

124: (\ref{eq2}), with an appropriate contact term playing the role of

125:  "diffusion" , give rise to a constant velocity solution.

126:  The mean field dynamics is akin to that  of the FKPP equation \cite{fisher,kpp}, and yield a

127:  velocity proportional to the square root of the distance from

128:  $R_c$.   The solution for the SIS case is that of a front separating the

129:  metastable uninfected state ahead of the front from the endemic state behind.  In the SIR case,

130:  the solution is a pulse, leaving behind a state with a diminished susceptible

131:  population such that the effective infectivity parameter, $R_0 S/N$, is below $R_c$.

132: 

133: What happens in the stochastic models then?  Let us discuss the SIS

134: case first. For $N=1$ the SIS model is equivalent to the contact

135: process~\cite{Harris}, where there is a transition to a propagating state at a

136: finite value of $R_0 > 1$.  This transition is known to be in the

137: directed peculation (DP) equivalence class~\cite{Hinr}.  Above the transition, there

138: is a finite chance of generating a wave of infection whose lifetime

139: is infinite.  The average velocity of the wave is less than that

140: predicted by the deterministic equation.  This has

141: been analyzed by Brunet and Derrida \cite{bd}, who show that as

142: $N\to \infty$, the average velocity approaches the deterministic value, albeit with

143: anomalously large ${\cal{O}}(\ln^{-2}N)$ corrections. Below the

144: directed percolation transition, the typical spatial extent of the

145: epidemic, denoted $\xi_\perp$, is finite, and diverges as the

146: transition point is approached.  Following the Janssen-Grassberger

147: conjecture \cite{jg}, it is widely  believed that \emph{any} generic

148: extinction transition, and, in particular, the transition for the

149: SIS model for \emph{any} N,  falls in the equivalence class of

150: directed percolation.

151: 

152: In order to test these predictions in the large density limit we

153: have simulated the SIS process on a one dimensional array of sites.

154: On each site there are  $N$ individuals, and this number is fixed

155: throughout the process. The epidemics is ignited by the introduction of a

156: single infected individual on a single site. The chance of an $I$ to

157: infect any susceptible individual on the same site is $(1-\chi)

158: \alpha/N$, and its chance to infect an $S$ on one of the two

159: neighboring sites is $\chi \alpha/N$. Thus,  $\chi$ is the

160: inter-site infectivity rate, while the population within a site is

161: considered as well mixed. The  chance of recovery  is independent of

162: the spatial structure and the recovery rate is $\beta$ for any

163: individual. We have used an exact, agent-based simulation in order

164: to study the approach to transition from below, using the divergence

165: of the perpendicular correlation length $\xi_\perp$ as a marker

166: of the transition.

167: 

168: What is the effect of increasing $N$ on the transition? Fixing

169: $\beta$ and $\chi$ one should expect a dependence of  $R_{DP}$, the value of $R_0$ at

170: the DP transition,

171:  on $N$. If $N=1$, for example,

172: the SIS becomes a simple contact process on a line, and the DP

173: transition happens when the infection rate is 3.297 times higher than

174: the recovery rate; accordingly, $R_{DP}(N=1)  =

175: 3.297/\chi$. On the other hand, as $N \to \infty$, $R_{DP}$ should

176: approach the value $R_c \equiv 1$ and becomes  $\chi$ independent,

177: as a single site may support the endemic state. $R_{DP}$ should

178: interpolate between these two limits as a function of $N$.

179: 

180: Another effect has to do with the divergence of $\xi_\perp$ at the

181: transition. According to the stochastic theory, right below the

182: transition $\xi_\perp$ is finite but large. On the other hand, the

183: classical equations predict that $\xi_\perp = 0 $ below the

184: transition. One expects both statements to be true at the large $N$

185: limit. The only way out of that paradox is to understand that the

186: classical description must fail close enough to the transition, but

187: the region in which  it fails must shrink to zero in the $N \to

188: \infty$ limit. Say it another way, at a fixed distance from the

189: transition, and for large enough $N$, the perpendicular correlation

190: length should approach zero.

191: 

192: 

193: Our simulations support all parts of this picture. First, the

194: Jansen-Grassberger conjecture works and in all cases studied  we

195: find that $\xi_\perp$ diverges as the transition point is approached

196: from below with the expected power $-\nu_\perp\approx -1.097$.

197: Second, we find that  $\xi_\perp$  obeys a universal scaling with

198: $N$, independent of the details of the model. Specifically, near the

199: transition

200: \begin{equation}

201: \xi_\perp = A(N) (R_{DP}(N) - R_0)^{-\nu_\perp}

202: \label{trans}

203: \end{equation}

204: where for large $N$

205: \begin{equation}

206: R_{DP} (N) - R_c\sim  N^{-\kappa}, \qquad \kappa \approx 0.66

207: \end{equation}

208: where for large $N$

209: \begin{equation}

210: A(N) \sim N^{-\tau}, \qquad \tau \approx 0.41.

211: \end{equation}

212: The data supporting these findings are presented in Fig. \ref{rdp} for the

213: specific case of  an SIS model with fixed recovery time and $\chi =

214: 0.2$. The applicability of these results, though, is much wider; the

215: same exponents turn out to describe the large $N$ limit of the SIS

216: model with different $\chi$ and with exponentially distributed

217: recovery time, as well as other stochastic  models like

218: branching-annihilation-death ($A \to \emptyset, \  A+A \to

219: \emptyset,  \  A \to 2A$) and so on. We thus conjecture that these

220: exponents describe the large $N$ behavior of \emph{any}

221: one-dimensional extinction transition that belongs to the DP

222: equivalence class.

223: 

224: The data collapse presented in Figure \ref{sisxicol} reveals the

225: existence of much stronger regularity at large $N$'s. Eq. (\ref{trans})

226: implies that $N^{\kappa-\tau/\nu_\perp}\xi_\perp^{-1/\nu_\perp}$ is

227: a linear function of $(R_o - R_c)N^\kappa$ close to the transition. In

228: fact, this result generalizes to a whole scaling regime where the

229: scaled correlation length  is a function of  the scaled distance

230: from the classical transition:

231: \begin{equation}

232: N^{\kappa-\tau/\nu_\perp}\xi_\perp^{-1/\nu_\perp} = F( (R_o - R_c)N^\kappa)

233: \end{equation}

234: Again, we have verified that this scaling behavior is independent of

235: the (nonzero) strength of the intersite coupling.

236: 

237: \begin{figure}

238: \center{\includegraphics[width=0.45\textwidth]{rdp2.eps}}

239: \caption{Directed percolation transition point, $R_{DP}$, of the 1D SIS model as a function of $N$, together with the critical perpendicular correlation length amplitude $A(N)$.  Also shown are power-law fits to the data. $\xi=0.2$}

240: \label{rdp}

241: \end{figure}

242: 

243: \begin{figure}

244: \center{\includegraphics[width=0.45\textwidth]{sisxicollapse3.eps}}

245: \caption{Collapse of the 1D SIS model scaled correlation length, $\xi_\perp^{-1/\nu_\perp} N^{0.284}$

246: versus the scaled infectivity $(R_o-1)N^{0.662}$ for $N=200$, $400$, $800$, $1600$ and $3200$. The contact parameter is again $\chi =0.2$, and the fixed recovery time variant was used. %Also shown for comparison is $N=10$, which is too small to be in the scaling regime.

247: }

248: \label{sisxicol}

249: \end{figure}

250: 

251: We now turn to examine the behavior of the stochastic $SIR$ model at

252: large $N$.  Here, the model is supposed to be in the dynamic

253: percolation universality class~\cite{gras83,grascar85}.  This class does not have a

254: transition in one spatial dimension, and the propagating pulse

255: always dies out in an infinite system.  Still, the correlation

256: length exhibits a scaling behavior with $N$ very similar to that of

257: the SIS model.  We find that the scaled correlation length,

258: $\xi_\perp N^{-0.209}$ is a function of the scaled infectivity $(R_o

259: - R_c)N^{0.454}$.  The data collapse is presented in Fig.

260: \ref{sirxicol}. Here the correlation length is finite for all

261: infectivities, indicating that we are indeed below the percolation

262: transition.  The correlation length does however grow very rapidly with $R_o$.

263: 

264: \begin{figure}

265: \center{\includegraphics[width=0.45\textwidth]{sirxicollapse.eps}}

266: \caption{Collapse of the SIR scaled correlation length, $\xi_\perp N^{-0.209}$

267: versus the scaled infectivity $(R_o-1)N^{0.454}$ for $N=200$, $400$, and $800$. Here too  $\chi =0.2$.}

268: \label{sirxicol}

269: \end{figure}

270: 

271: In the simulations presented above, we have used as our measure of the correlation length the inverse of the

272: exponential falloff rate of the distribution function for infection location.  It is remarkable

273: that in both models, for all $R_o$ studied (below the extinction transition), this distribution function appears to be given {\em{exactly}} (at least to within our

274: statistical errors, over some seven decades) by the functional form

275: $P(x) = C(R_o,N) e^{-|x-x_o|/\xi_\perp(R_o,N)}$, with $x_o$ is the location of the original infected person.  The average {\em total} number of infections is given then roughly by ${\cal{N}}=2C(R_o,N)\xi_\perp(R_o,N)$ (assuming

276: $\xi_\perp\gg 1$, as it is for $N$ large).  We can then deduce the scaling properties of $C(R_o=1,N)$ with $N$.  The number of infections at the central site, $C(1,N)$ are the result of the on-site infections nucleated by the original infected person and by those nucleated by the neighboring sites.  The effective on-site infection rate is $R_o(1-\chi)$,

277: and since this is below $R_c=1$, each initial infection develops essentially independently, giving rise to $1/(1-R_o(1-\chi))$ total infections.  The number of primary infections induced by the neighbors is $\chi C \exp(-1/\xi_\perp) \approx \chi C (1 - 1/\xi_\perp)$.  Thus, we have, for $R_o=1$

278: \begin{eqnarray}

279: C &=& \frac{1}{1-(1-\chi)} \left(1 + \chi C \left(1 - \frac{1}{\xi_\perp}\right)\right)\\

280: &=& C + \frac{1}{\chi} - \frac{C}{\xi_\perp}

281: \end{eqnarray}

282: This implies that

283: \begin{equation}

284: C = \frac{\xi_\perp}{\chi}

285: \end{equation}

286: so $C(1,N)$ scales exactly the same with $N$ as $\xi_\perp(R_o=1)$.  Thus, the total number of infections scales as $\xi_\perp^2(R_o=1)$, in both the SIS and SIR models.

287: For the SIS model, this gives the total number of infections scaling as $N^{2(\kappa\nu_\perp - \tau)} \approx N^{0.623}$, whereas in the SIR model, we have the total number of infections scaling as approximately $N^{0.418}$.  Notice that in both cases, the total

288: number of infections at the critical infectivity scales faster than the 0 dimensional results, $1/2$ and $1/3$, respectively.  We have tested these predictions in our simulations, (data not shown) and found them to be very well satisfied.  Also, preliminary work indicates that these exponents are higher still in two dimensions, and saturate at 1 as the

289: dimension goes to infinity~\cite{long}.

290: 

291: In the SIS model, as in the related contact process, the total number of infections diverges at the directed percolation transition, with the exponent $\gamma=\nu_\parallel + \nu_\perp - 2\beta=2.278$.  However, the total number of infections is equal to $C\xi_\perp$.  Thus, $C$ should diverge at the transition as $\gamma-\nu_\perp = 1.181$.  Combining this with our above result for the scaling of $C$ with $N$ at criticality, we get

292: that

293: \begin{equation}

294: (C N^{-0.311})^{-1/1.181} = {\cal{G}}\left((R-1)N^{0.668}\right)

295: \end{equation}

296: where ${\cal{G}}$ vanishes linearly at the transition.  It is remarkable that even though the directed percolation exponent associated with $C$ differs from that of $\xi_\perp$, the $N$ scaling is the same.

297: 

298: 

299: Our simulations suggest that the quasi-deterministic  region is

300: controlled by the new critical exponents $\kappa$ and $\tau$. At the

301: transition, the  deterministic limit does not exist and the

302: stochastic dynamics of the microscopic constituents determines the

303: system behavior for any time scale. Off transition, on the other

304: hand, the effect of stochasticity vanishes for large enough $N$. The

305: value of $\kappa$ and $\tau$ is parameter independent, as long as

306: the dimensionality of the system and the type of stochastic

307: transition (DP vs. dynamic percolation)  are kept fixed.

308: 

309: It is interesting to point out an exception to our analysis, namely,

310: the branching-annihilation process with no bare death term ($A

311: \stackrel{\sigma}{\rightarrow} 2A, A+A

312: \stackrel{\lambda}{\rightarrow} \emptyset$) for Brownian particles.

313: Here at low densities there is a DP type 

314: transition at finite birth rate, but as $N \to \infty$ (i.e.,

315: $\lambda \to 0$) the transition approaches $\sigma = 0$. The

316: deterministic limit of the transition in that case does \emph{not}

317: belong to the directed percolation equivalence class; in fact, it is

318: known that for $\sigma = 0$ the density decays like $t^{-1/2}$ in

319: one spatial dimension. Indeed, Cardy and Tauber~\cite{cardy} calculated that the

320: distance to the DP transition should scale as $N^{-2}$, (since close to the classical

321: transition the density is inversly proportional to $1/N$). Addition of a

322: spontaneous death process $A \stackrel{\mu}{\rightarrow} 0$ to the

323: model shifts the deterministic transition to $\sigma=\mu$ and

324: provide us with a model that admits a DP transition all the way to $N

325: = \infty$. In fact, our preliminary numerics show that in the Cardy-Tauber case

326: the scaling of the critical $\sigma$ appears to be consistent with  their prediction.

327:  Adding $\mu$ changes the picture

328: dramatically and $\sigma_c \sim \mu + N^{-\kappa}$, where $\kappa$

329: is identical to that measured for the SIS model. It is interesting

330: to note that, if the  Cardy-Tauber perturbative technique is

331: modified to include the spontaneous death process, at large $N$ the

332: distance from the DP transition is predicted to scale as $N^{-1}$ instead of

333: $N^{-2}$.

334: 

335: Many aspects of that problem are still open. In particular a

336: rigorous classification scheme for the quasi-deterministic behavior

337: is still missing. The extinction transition in higher dimensions

338: (where dynamic percolation admits a nontrivial critical point) and the

339: behavior of the system above the transition also need to be investigated. We

340: hope to address these subjects in future work.

341: 

342: \acknowledgments{We thank S. Havlin for useful remarks. The work of NMS is supported in part by the EU 6th framework CO3 pathfinder.}

343: 

344: \begin{thebibliography}{1}

345: \bibitem{sa} N. T. J. Bailey, The Mathematical Theory of Infectious Diseases

346: and its Applications (Hafner Press, New York, 1975).

347: \bibitem{am} R. M. Anderson and R. M. May, \emph{Infectious Diseases of

348: Humans} (Oxford University Press, Oxford, 1991).

349: \bibitem{watson}R. Watson, J. Appl. Prob. {\bf 17}, 912 (1980).

350: \bibitem{mysis}D. A. Kessler,  preprint arXiv:0709.3049.

351: \bibitem{dlsis}R. G. Dolgoarshinnykh and S. P. Lalley, J. Appl. Prob. {\bf 43}, 892 (2006).

352: \bibitem{lof}A. Martin-L\"of, J. Appl. Prob. {\bf 35}, 671 (1998).

353: \bibitem{bn}E. Ben-Naim and P. L. Krapivsky, \pre {\bf 69} 050901 (2004).

354: \bibitem{mysir}D. A. Kessler and N. M. Shnerb, \pre {bf 76} 010901(R) (2007).

355: \bibitem{fisher} R. A. Fisher, Ann. Eugenics {\bf 7}, 353 (1937).

356: \bibitem{kpp} A. N. Kolmogorov, I. G. Petrovskii and N. S. Piskunov, Selected Works of A. N. Kolmogorov.

357: V. M. Tikhomirov (Ed.), Kluwer Academic Publishers, 1991.

358: \bibitem{Harris}T. E. Harris, Ann. Prob. {\bf 2}, 969 (1974).

359: \bibitem{Hinr}For a review of directed percolation, see H. Hinrichsen, Adv. Phys. {\bf 49}, 815 (2000).

360: \bibitem{bd} E. Brunet and B. Derrida, \pre {\bf 56}, 2597 (1997).

361: \bibitem{jg} H. K.

362: Janssen, Z. Phys. B {\bf 42}, 151 (1981);

363: P. Grassberger, Z. Phys. B {\bf 47}, 365 (1982).

364: \bibitem{gras83}P. Grassberger, Math. Biosci. {\bf 62}, 157 (1983).

365: \bibitem{grascar85}J. Cardy and P. Grassberger, J. Phys. A {\bf 18}, L267 (1985).

366: \bibitem{long}D. A. Kessler and N. M. Shnerb, unpublished.

367: \bibitem{cardy} J.L. Cardy and U.C. Tauber, J. Stat. Phys. \textbf{90}, 1

368: (1998).

369: 

370: \end{thebibliography}

371: 

372: \end{document}

373: