1: \documentclass[twocolumn,showpacs,aps]{revtex4}
2: %\documentclass[preprint,showpacs,aps]{revtex4}
3: \usepackage{epsf}
4: \begin{document}
5: \title{Size of Outbreaks Near the Epidemic Threshold}
6: \author{E.~Ben-Naim}\email{ebn@lanl.gov} \affiliation{Theoretical
7: Division and Center for Nonlinear Studies, Los Alamos National
8: Laboratory, Los Alamos, NM, 87545}
9: \author{P.~L.~Krapivsky}\email{paulk@bu.edu} \affiliation{Center for
10: BioDynamics and Department of Physics, Boston University, Boston, MA,
11: 02215}
12: \begin{abstract}
13:
14: The spread of infectious diseases near the epidemic threshold is
15: investigated. Scaling laws for the size and the duration of
16: outbreaks originating from a single infected individual in a large
17: susceptible population are obtained. The maximal size of an
18: outbreak $n_*$ scales as $N^{2/3}$ with $N$ the population size.
19: This scaling law implies that the average outbreak size $\langle
20: n\rangle$ scales as $N^{1/3}$. Moreover, the maximal and the
21: average duration of an outbreak grow as $t_*\sim N^{1/3}$ and
22: $\langle t\rangle\sim \ln N$, respectively.
23:
24: \end{abstract}
25:
26: \pacs{02.50.-r, 05.40.-a, 87.23.Cc, 87.19.Xx}
27:
28: \maketitle
29:
30: Infection processes typically involve a threshold
31: \cite{ntjb,ma,hwh,wd,hed,lp,bc,ad,jg}. Below the epidemic threshold,
32: outbreaks quickly die out, while above the threshold, outbreaks may
33: take off. We study epidemic outbreaks near the threshold. Such
34: outbreaks arise naturally. On the one hand, human efforts at disease
35: prevention reduce the infection rate thereby crossing the epidemic
36: threshold \cite{ma}. On the other hand, evolution may increase the
37: infection rate of diseases hovering just below the threshold,
38: enhancing the likelihood of near-threshold outbreaks
39: \cite{antia}. Typically, detection, modeling, and eradication of
40: infectious diseases are subtle for outbreaks near the epidemic
41: threshold.
42:
43: The total number of infected individuals is a basic measure of the
44: severity of an epidemic outbreak. We study outbreaks originating from
45: a single infected individual in a large susceptible population. Our
46: main result is that near the epidemic threshold, the maximal outbreak
47: size $n_*$ grows as a power-law of the population size $N$,
48: \begin{eqnarray}
49: \label{size-max}
50: n_*\sim N^{2/3}\,.
51: \end{eqnarray}
52: In contrast, below the epidemic threshold, endemic outbreaks involve a
53: small number of infected individuals, while above the epidemic
54: threshold, pandemic outbreaks involve a fraction of the population
55: $n_*\sim N$. Therefore, outbreaks near the epidemic threshold have a
56: distinct intermediate size between a pandemic and an endemic outbreak
57: \cite{ntjb}. Loosely speaking, epidemics come in three sizes: large,
58: medium, and small.
59:
60: The scaling law (\ref{size-max}) has several important implications
61: concerning the statistics of both the size and the duration of the
62: outbreaks. It implies that the average size of outbreaks $\langle
63: n\rangle$ and the maximal duration of outbreaks $t_*$ both scale as
64: $\langle n\rangle\sim t_*\sim N^{1/3}$ near the epidemic
65: threshold. Furthermore, the average duration of the outbreaks $\langle
66: t\rangle$ scales logarithmically, $\langle t\rangle\sim \ln N$. These
67: behaviors hold in a sizable range of infection rates, namely in a
68: window of the order ${\cal O}(N^{-1/3})$ around the epidemic
69: threshold.
70:
71: These scaling laws are demonstrated for the classic
72: Susceptible-Infected-Recovered (SIR) infection process
73: \cite{ntjb,ma,hwh}. In this model, the population consists of $s$
74: susceptible, $i$ infected, and $r$ recovered individuals with
75: $N=s+i+r$. These sub-populations change due to two competing
76: processes: infection and recovery. The disease is transmitted from an
77: infected individual to a susceptible one with rate $\alpha/N$, where
78: $\alpha$ is the infection rate:
79: \begin{equation}
80: (s,i,r)\buildrel {\alpha si}/N \over \longrightarrow (s-1,i+1,r).
81: \end{equation}
82: Infected individuals recover with a unit rate:
83: \begin{equation}
84: (s,i,r)\buildrel i \over \longrightarrow (s,i-1,r+1).
85: \end{equation}
86: The infection process starts with a single infected individual,
87: $(s,i,r)=(N-1,1,0)$, and it ends with none $(s,i,r)=(N-n,0,n)$.
88:
89: \begin{figure}[t]
90: \centerline{\epsfxsize=8cm\epsfbox{fig1.eps}}
91: \caption{The average outbreak size versus the population size for the
92: SIR infection process at the epidemic threshold ($\alpha=1$). Shown
93: are Monte Carlo simulation results representing an average over
94: $10^9$ independent realizations of the infection process (circles). A
95: line of slope $1/3$ is also shown as a reference. A least-square-fit
96: to $\langle n\rangle \sim N^\gamma$ in the range $10^3<N<10^9$ yields
97: $\gamma=0.334\pm 0.001$.}
98: \end{figure}
99:
100: The total size of the outbreak $n$ and the duration of the outbreak
101: $t$ are the outcomes of a stochastic process. We study statistical
102: properties of these random variables, particularly, their average and
103: maximal size, as a function of the population size (We implicitly
104: consider an average over infinitely many realizations of the infection
105: process.)
106:
107: In the infinite population limit, the epidemic threshold is
108: $\alpha=1$. Since infection occurs with probability
109: $\frac{\alpha}{1+\alpha}$ and recovery with probability
110: $\frac{1}{1+\alpha}$, the average outbreak size satisfies
111: \hbox{$\langle n\rangle
112: =\frac{1}{1+\alpha}+2\frac{\alpha}{1+\alpha}\langle n\rangle$}. Thus,
113: below the threshold ($\alpha<1$), a finite number of individuals is
114: infected, $\langle n\rangle=(1-\alpha)^{-1}$. Above the threshold
115: ($\alpha>1$), there is a pandemic outbreak with a finite fraction of
116: the population infected: $\langle n\rangle =rN$ \cite{ntjb,ar}. At the
117: threshold ($\alpha=1$), the probability that the outbreak size equals
118: $n$, $G_n$, is found recursively:
119: \hbox{$G_n=\frac{1}{2}\sum_{m=1}^{n-1} G_mG_{n-m}$} starting with
120: $G_1=1/2$. This recursion reflects the fact that the first infection
121: event results in two independent infection processes \cite{teh}. The
122: generating function is \hbox{$\sum_{n\geq 1} G_nz^n =1-\sqrt{1-z}$},
123: from which the size distribution is a power-law,
124: \begin{equation}
125: \label{size-dist}
126: G_n\sim n^{-3/2},
127: \end{equation}
128: for sufficiently large outbreaks $n\gg 1$.
129:
130: For a finite, yet large population, the outbreak size distribution
131: (\ref{size-dist}) holds, but only up to the maximal outbreak size:
132: $1\ll n \ll n_*$. Outbreaks beyond the maximal size are practically
133: impossible. Therefore, the average outbreak size grows according to
134: $\langle n\rangle =\sum_{n=1}^{n_*} n G_n\sim n^{1/2}_*$. Naively
135: assuming that a finite fraction of the population may become infected,
136: $n_*\sim N$, would lead to \hbox{$\langle n\rangle \sim
137: N^{1/2}$}. While consistent with the generic statistical
138: uncertainties, this law is in fact {\it erroneous}. Instead, the
139: outbreak size is much smaller because the epidemic outbreak weakens as
140: more individuals become infected, and it finally dies out when the
141: number of infected individuals becomes of the order $n_*$. When there
142: are \hbox{$N-n_*$} susceptible individuals, the total infection rate
143: $\alpha (N-n_*)i/N$ shows that the infection rate is effectively
144: reduced, \hbox{$\alpha_{\rm eff}=1-n_*/N$}. Therefore, the epidemic
145: becomes essentially endemic. (This is clearly a finite population
146: effect: the susceptible population ``reservoir'' is never affected in
147: the infinite population limit.) Equating the outbreak size in the
148: endemic phase \hbox{$\langle n\rangle\sim (1-\alpha_{\rm
149: eff})^{-1}\sim N/n_*$} with that estimated from the size distribution,
150: $\langle n\rangle\sim n_*^{1/2}$, gives the scaling law
151: (\ref{size-max}) governing the maximal outbreak size. Hence, in the
152: worse case scenario, only a fraction of the order of $N^{-1/3}$ of the
153: entire population can ever be infected.
154:
155: As a byproduct we obtain the scaling law for the average outbreak size
156: \begin{eqnarray}
157: \label{size-ave}
158: \langle n\rangle\sim N^{1/3}\,.
159: \end{eqnarray}
160: Large scale Monte Carlo simulations confirm this behavior (Fig.~1).
161: The simulations are a straightforward realization of the infection
162: process. When there are $s$ susceptible individuals, with probability
163: $1/(1+\alpha s/N)$ a recovery event occurs, and otherwise, an
164: infection event occurs. The simulation results represent an average
165: over a remarkably large number of independent realizations.
166:
167: \begin{figure}[t]
168: \centerline{\epsfxsize=8cm\epsfbox{fig2.eps}} \caption{The normalized
169: cumulative distribution $U_n(N)/U_n(\infty)$ versus the normalized
170: outbreak size $n/N^{2/3}$. The data corresponds to an average over
171: $10^6$ independent realizations.}
172: \end{figure}
173:
174: Statistical properties of the outbreak size are self-similar as they
175: follow a universal, population-size independent law. Once the
176: outbreak size distribution and the outbreak size are properly
177: normalized by the infinite population distribution and the maximal
178: outbreak size respectively, a universal behavior emerges:
179: \hbox{$G_n(N)/G_n(\infty)\to {\cal G}(n/N^{2/3})$}. This universality,
180: reminiscent of finite-size scaling in critical phenomena \cite{pf},
181: was confirmed numerically by studying the cumulative distribution
182: $U_n(N)=\sum_{m\geq n}G_n(N)$ (Fig.~2). This provides further
183: verification of the scaling law (\ref{size-max}).
184:
185: The scaling laws characterizing the outbreak size hold not only at the
186: threshold but also in a window around the threshold. Equating the
187: average outbreak size (\ref{size-ave}) with the behavior in the
188: endemic phase, $\langle n\rangle=(1-\alpha)^{-1}$, we find that the
189: threshold window (i.e., the range of infection rates for which the
190: intermediate behavior holds) diminishes with the population size as
191: \begin{equation}
192: \label{window}
193: |1-\alpha|\sim N^{-1/3}.
194: \end{equation}
195: This parameter range can be sizable for moderate populations --- for
196: example, when $N=10^3$, the threshold window is roughly
197: $0.9<\alpha<1.1$ and the maximal outbreak size is smaller than the
198: population size by a factor of $10$.
199:
200: \begin{figure}[t]
201: \centerline{\epsfxsize=8cm\epsfbox{fig3.eps}}
202: \caption{The near threshold behavior. Shown is the normalized outbreak
203: size $\langle n\rangle/N^{1/3}$ versus the normalized distance from
204: the threshold $(\alpha-1)N^{1/3}$. The data corresponds to an average
205: over $10^6$ independent realizations.}
206: \end{figure}
207:
208: The behavior of $\langle n\rangle$ near the epidemic threshold
209: provides another manifestation of the scaling law
210: (\ref{window}). Indeed, plotting the average outbreak size versus the
211: infection rate normalized according to (\ref{size-ave}) and
212: (\ref{window}), respectively, shows a universal behavior:
213: \hbox{$\langle n\rangle/N^{1/3}\to {\cal
214: Q}\left[(1-\alpha)N^{1/3}\right]$} (Fig.~3).
215:
216: The threshold window is larger than the canonical $N^{-1/2}$ estimate
217: arising either from the standard large-population analysis
218: \cite{ngv,dss} or from the widely-used deterministic SIR ordinary
219: differential equations \cite{jdm}, describing the evolution of the
220: average susceptible and infected populations \cite{bk}. Moreover, the
221: related SI (sometimes also termed SIS) model, where a recovered
222: individual immediately becomes susceptible, is characterized by the
223: simpler behavior $n_*\sim N$ and $\langle n\rangle \sim N^{1/2}$;
224: finite size effects are not as pronounced because there is no
225: depletion of the susceptible reservoir.
226:
227: The scaling laws for the outbreak size have direct implications
228: concerning the dynamics and in particular, the duration of infection
229: processes near the epidemic threshold. To obtain these scaling laws,
230: we again consider first the infinite population limit. At the epidemic
231: threshold, $\alpha=1$, infection and recovery occur with equal
232: probabilities and therefore, the average number of infected
233: individuals is conserved, $I(t)=1$. The probability $P_i(t)$ that
234: there are $i$ infected individuals at time $t$ satisfies
235: \begin{equation}
236: \frac{d}{dt}P_i=(i+1)P_{i+1}+(i-1)P_{i-1}-2iP_i
237: \end{equation}
238: together with the initial condition $P_i(0)=\delta_{i,1}$. The
239: distribution is geometric, $P_i(t)=t^{i-1}(1+t)^{-(i+1)}$ \cite{bk,edg}
240: for $i\geq 1$, and $P_0(t)=t(1+t)^{-1}$ for $i=0$. Therefore, the
241: survival probability of the outbreak, i.e., the probability that the
242: outbreak is still active at time $t$ is simply
243: \begin{equation}
244: \label{Pt}
245: P(t)=(1+t)^{-1}
246: \end{equation}
247: since $P(t)=1-P_0(t)$. Restricting attention to active outbreaks, the
248: average number of infected individuals grows linearly with time
249: $\langle i\rangle=I(t)/P(t)=1+t$. Consequently, the typical number of
250: recovered individuals $r\sim \int_0^t dt'\,(1+t')$ grows quadratically
251: with time: $r\sim t^2$.
252:
253: \begin{figure}[t]
254: \centerline{\epsfxsize=8cm\epsfbox{fig4.eps}}
255: \caption{The average outbreak duration at the epidemic threshold
256: versus the population size. Simulation results, obtained from an
257: average over $10^9$ realizations are consistent with the theoretical
258: prediction (\ref{duration-ave}). A best fit to $\langle t\rangle =
259: \beta\ln N$ yields $\beta=0.32\pm 0.01$.}
260: \label{tav}
261: \end{figure}
262:
263: For finite populations, the probability that the outbreak is still
264: alive at time $t$ decays as $P(t,N)\sim t^{-1}$ up to the maximal time
265: scale $t\ll t_*$. The survival probability is sharply suppressed for
266: times larger than the maximal time. The maximal duration of outbreaks
267: is estimated by equating the time dependent outbreak size $n\sim r
268: \sim t^2$ with the maximal outbreak size $n_*\sim N^{2/3}$. Therefore,
269: \begin{equation}
270: \label{duration-max}
271: t_*\sim N^{1/3}.
272: \end{equation}
273:
274: The maximal duration of outbreaks greatly exceeds both the typical
275: duration that is of the order of one and the average duration of an
276: outbreak $\langle t\rangle$ which exhibits an interesting logarithmic
277: growth. To derive the logarithmic law, we first note that, by
278: definition, the average duration of an outbreak is $\langle
279: t\rangle=\int_0^{\infty}\, dt\, t\, \frac{d}{dt}P(t,N)$. Using the
280: infinite population result (\ref{Pt}) and integrating up to $t_*$ that
281: plays the role of a cutoff, we get
282: \begin{equation}
283: \label{duration-ave}
284: \langle t\rangle\simeq \frac{1}{3}\ln N.
285: \end{equation}
286: Numerical simulations confirm this behavior (Fig.~4). The probability
287: distribution for the duration of outbreaks also follows a
288: population-size independent law: \hbox{$P(t,N)/P(t)\to {\cal
289: P}\left(t/N^{1/3}\right)$} as shown in Fig.~5. However, the
290: convergence to this law is not uniform: it is slow for short durations
291: but fast at large durations.
292:
293: \begin{figure}[t]
294: \centerline{\epsfxsize=8cm\epsfbox{fig5.eps}}
295: \caption{The survival probability at the epidemic threshold. Shown is
296: the normalized survival probability $P(t,N)/P(t)$ versus the
297: normalized duration time $t/N^{1/3}$. The data corresponds to an
298: average over $10^8$ realizations.}
299: \label{st}
300: \end{figure}
301:
302: In summary, we found that outbreaks in the vicinity of the epidemic
303: threshold have a distinct size, characterized by a distinct power-law
304: dependence of the population size. This behavior describes a range of
305: infection processes in the vicinity of the epidemic threshold. The
306: size of this threshold window is larger than expected from the
307: traditional large system size analysis techniques or from the
308: deterministic description. We conclude that statistical fluctuations
309: and finite population effects are most pronounced and may be quite
310: subtle near the epidemic threshold.
311:
312: The scaling laws have concrete implications regarding the
313: computational complexity of near-threshold infection
314: processes. Typically, one has to compute $P_{i,r}$, the probability
315: that there are $i$ infected individuals and $r$ removed individuals
316: from the master equations. Although there are $N^2$ such coupled
317: ordinary differential equations, the scaling laws $i\sim N^{1/3}$ and
318: $r \sim N^{2/3}$ imply that the number of relevant equations is much
319: smaller and scales only linearly with the population size.
320:
321: Several questions arise, e.g., what is the shape of the scaling
322: functions ${\cal G}\left(n/N^{2/3}\right)$ and ${\cal
323: P}\left(t/N^{1/3}\right)$ characterizing the size and duration of
324: outbreaks near the epidemic threshold? Analytical determination of
325: these functions is very challenging as it requires treatment of the
326: full master equations describing the stochastic infection process
327: \cite{ntjb}, that is, the distribution $P_{i,r}(t,N)$ is needed
328: \cite{bk}.
329:
330: Further related problems include the corresponding near-threshold
331: scaling laws for spatial epidemic models, where the geometry and the
332: spatial structure of the infected domain play a role
333: \cite{sa,pg,ad1,wss}, and infection processes on networks
334: \cite{pv,mejn}.
335:
336: \acknowledgments We are grateful to Aric Hagberg for initial
337: collaboration on this work. We also thank Gary Doolen, Hans
338: Frauenfelder, and Sergei Rudchenko for careful reading of the manuscript.
339: This research was supported in part by DOE(W-7405-ENG-36).
340:
341: \begin{thebibliography}{99}
342:
343: \bibitem{ntjb}
344: N.~T.~J.~Bailey,
345: {\em The Mathematical Theory of Infectious Diseases}
346: (Charles Griffin, London, 1975).
347:
348: \bibitem{ma}
349: R.~Anderson and R.~May,
350: {\em Infectious Diseases: Dynamics and Control}
351: (Oxford Univ. Press, Oxford, 1991).
352:
353: \bibitem{hwh}
354: H.~W.~Hethcote,
355: SIAM Rev. {\bf 42}, 599 (2000).
356:
357: \bibitem{wd}
358: G.~H.~Weiss and M.~Dishon,
359: Math. Biosci. {\bf 11}, 261 (1971).
360:
361: \bibitem{hed}
362: H.~E.~Daniels,
363: Biometrika {\bf 59}, 211 (1972).
364:
365: \bibitem{lp}
366: C.~Lefevre and C.~Picard,
367: Adv. Appl. Prob. {\bf 22}, 25 (1990).
368:
369: \bibitem{bc}
370: F.~Ball and D.~Clancy,
371: Adv. Appl. Prob. {\bf 25}, 721 (1993).
372:
373: \bibitem{ad}
374: H.~Andersson and B.~Djehiche,
375: J. Appl. Prob. {\bf 35}, 662 (1998).
376:
377: \bibitem{jg}
378: J.~Grasman, Math. Biosci. {\bf 152}, 13 (1998).
379:
380: \bibitem{antia} J.~Bull and D.~Dykhuizen, Nature {\bf 426}, 609
381: (2003); R.~Antia, R.~R.~Regoes, J.~C.~Koella, and C.~T.~Bergstrom,
382: Nature {\bf 426}, 658 (2003).
383:
384: \bibitem{ar}
385: The fraction $r$ satisfies $r+\exp(-\alpha r)=1$.
386:
387: \bibitem{teh}
388: T.~E.~Harris,
389: {\em The Theory of Branching Processes} (Dover, New York, 1989).
390:
391: \bibitem{pf}
392: V.~Privman and M.~E.~Fisher,
393: Phys.\ Rev.\ B {\bf 30}, 322 (1984).
394:
395: \bibitem{ngv}
396: N.~G.~Van Kampen, {\em Stochastic Processes in Physics and Chemistry}
397: (North Holland, Amsterdam, 2003).
398:
399: \bibitem{dss}
400: C.~R.~Doering, K.~V.~Sargysyan, and L.~M.~Sander,
401: {\it q-bio}/0401016.
402:
403: \bibitem{jdm}
404: J.~D.~Murray, {\em Mathematical Biology}
405: (Springer-Verlag, New York, 1989).
406:
407: \bibitem{bk}
408: E.~Ben-Naim and P.~L.~Krapivsky, in preparation.
409:
410: \bibitem{edg}
411: E.~Ben-Naim and P.~L.~Krapivsky,
412: Phys.\ Rev.\ E {\bf 68}, 031104 (2003).
413:
414: \bibitem{sa}
415: D.~Stauffer and A.~Aharony, {\em Introduction to
416: Percolation Theory} (Taylor \& Francis, Briston PA, 1994).
417:
418: \bibitem{pg}
419: P.~Grassberger,
420: Math.\ Biosci. {\bf 63}, 157 (1983).
421:
422: \bibitem{ad1}
423: H.~Andersson and B.~Djehiche,
424: J. Appl. Prob. {\bf 34}, 698 (1998).
425:
426: \bibitem{wss}
427: C.~P.~Warren, L.~M.~Sander, and I.~M.~Sokolov,
428: Phys.\ Rev.\ E {\bf 66}, 056105 (2002).
429:
430: \bibitem{pv}
431: R.~Pastor-Satorras and A.~Vespignani,
432: Phys.\ Rev.\ Lett. {\bf 86}, 3200 (2001).
433:
434: \bibitem{mejn}
435: M. E. J. Newman,
436: Phys.\ Rev.\ E {\bf 66}, 016128 (2002).
437:
438: \end{thebibliography}
439:
440: \end{document}
441: