1: \tolerance=10000
2: \documentclass{appolb}
3: \usepackage{amsmath}
4: \usepackage{amssymb,amsthm}
5: \usepackage{graphicx}
6: \begin{document}
7:
8: \author{P.\ F.\ G\'ora\thanks{e-mail:gora@if.uj.edu.pl}
9: \address{M.\ Smoluchowski Institute of Physics and Complex Systems Research Center\\
10: Jagellonian University, Reymonta~4, 30-059 Krak\'ow, Poland}}
11:
12: \title{Stationary distributions of a noisy logistic process}
13:
14: \date{}
15:
16: \maketitle
17:
18: \begin{abstract}
19: Stationary solutions to a Fokker-Planck equation corresponding to a~noisy
20: logistic equation with correlated Gaussian white noises are constructed. Stationary
21: distributions exist even if the corresponding deterministic system
22: displays an unlimited growth. Positive correlations between the noises
23: can lead to a minimum of the variance of the process and to the stochastic
24: resonance if the system is additionally driven by a periodic signal.
25: \end{abstract}
26:
27: \PACS{05.40.Ca, 02.50.Ey, 87.23.Cc}
28:
29: \section{Introduction}
30:
31: The logistic equation
32:
33: \begin{equation}
34: \dot x = ax(1-x)\,,\ a>0\,,\ x\geqslant0\,,
35: \end{equation}
36:
37: \noindent is one of the best-known and most popular models in population
38: dynamics. Perturbing this equation by a multiplicative noise is an obvious
39: generalization of the deterministic theory, aiming at describing populations
40: that live in an ever-changing environment. The logistic equation with
41: a~fluctuating growth rate
42:
43: \begin{equation}\label{fp:logistic-noise-basic}
44: \dot x = (a+p\,\xi(t))x(1-x)\,,
45: \end{equation}
46:
47: \noindent has been first discussed
48: by Leung in Ref.~\cite{Leung} and later by many other authors. Recently in
49: Ref.~\cite{appb} we have discussed a further generalization of
50: \eqref{fp:logistic-noise-basic} in which both the growth rate and the limiting population
51: level fluctuate, and these fluctuations are correlated in time:
52:
53: \begin{equation}\label{fp:logistic-noise-general2}
54: \dot x = (a+p\,\xi_{\text{m}}(t))x - (b + q\,\xi_{\text{a}}(t))x^2\,.
55: \end{equation}
56:
57: \noindent Here $\xi_{\text{m},\text{a}}$ are two Gaussian white noises (GWNs) that satisfy
58: $\left\langle\xi_i(t)\right\rangle=0$, $\left\langle\xi_i(t_1)\xi_i(t_2)\right\rangle=
59: \delta(t_1-t_2)$, $i=\text{m},\text{a}$, $\left\langle\xi_{\text{m}}(t_1)\xi_{\text{a}}(t_2)\right\rangle=
60: c\,\delta(t_1-t_2)$, $p$, $q$ are the amplitudes of the two noises
61: and the correlation coefficient $c\in[-1,1]$. For the sake of terminology, we will sometimes
62: call the noise $\xi_{\text{m}}(t)$ ``multiplicative'' and the noise $\xi_{\text{a}}(t)$ ``additive'';
63: see Eq.~\eqref{fp:linear} below for a~rationale behind these names. Please note,
64: though, that on the level of Eq.~\eqref{fp:logistic-noise-general2} both these noises
65: are coupled multiplicatively to the process $x(t)$. Note also that
66: if $b>0$, the corresponding deterministic equation converges to a stable fixed point;
67: accordingly, we will call a system with a positive $b$ ``convergent''. If $b\leqslant0$,
68: the corresponding deterministic system displays unlimited growth. We will call a system with
69: a~negative $b$ ``exploding''. Recently Mao \etal have shown in Ref.~\cite{marion} that in
70: the absence of $\xi_m$, the system \eqref{fp:logistic-noise-basic} remains positive
71: and bounded even in the ``exploding'' case. In Ref.~\cite{njp} we have discussed certain
72: difficulties that may arise in numerical simulations of such a~system. This
73: paper generalizes the result of Mao \etal to the case of both noises present.
74:
75: In Ref.~\cite{appb} we have shown how the dynamics \eqref{fp:logistic-noise-general2}
76: is related to the problem of
77: a linear stochastic resonance. We have mapped the nonlinear equation
78: \eqref{fp:logistic-noise-general2} into a linear Langevin equation with two correlated
79: noises and used the solutions of the latter to heuristically explain the behaviour
80: of the noisy logistic equation. Specifically, the substitution
81:
82: \begin{equation}\label{fp:substitution}
83: y=\frac{1}{x}
84: \end{equation}
85:
86: \noindent converts Eq.~\eqref{fp:logistic-noise-general2} into a linear equation
87:
88: \begin{equation}\label{fp:linear}
89: \dot y = -(a+p\,\xi_{\text{m}}(t))y + b + q\,\xi_{\text{a}}(t)
90: \end{equation}
91:
92: \noindent which can be solved exactly for realizations of the process $y(t)$.
93: The process \eqref{fp:linear} has a convergent mean if
94: \begin{subequations}
95: \begin{equation}\label{fp:condition-mean}
96: a-\frac{1}{2}p^2>0
97: \end{equation}
98: and a convergent variance if a stronger condition
99: \begin{equation}\label{fp:condition-variance}
100: a-p^2>0
101: \end{equation}
102: \end{subequations}
103: holds. Using the properties of the process $y(t)$, useful prediction can be made
104: about the noisy logistic process. Since in the presence of correlations, for certain
105: values of parameters the variance of $y(t)$ first shrinks and then grows as a function
106: of the ``multiplicative'' noise strength, $q$, one expects a~similar behaviour for
107: the logistic process $x(t)$ as well. These predictions have been corroborated
108: numerically in Ref.~\cite{appb}.
109: In particular, if $c=\pm1$, $bp\mp aq=0$ and the condition
110: \eqref{fp:condition-variance} is satisfied, the variance of the process $y(t)$ vanishes.
111: The relation between the processes $y(t)$ and $x(t) = 1/y(t)$ \textit{intuitively}
112: means that if almost all realizations of the former asymptotically reach the same
113: constant value, so do almost all realizations of the latter. However, as
114: a~\textit{formal} relation between moments of these processes is not trivial,
115: the predictions based on the properties of $y(t)$ have only a heuristic value.
116:
117: In the following we will
118: construct mathematically exact stationary solutions to the Fokker-Planck
119: equation corresponding to Eq.~\eqref{fp:logistic-noise-general2} and re-examine the above
120: results from the point of view of these stationary solutions. Furthermore, we
121: will show numerically that if the parameters undergo
122: periodic (for example, circaannual or seasonal) oscillations,
123: a positive correlation between the noises leads to a stochastic resonance.
124:
125: In the Appendix we extend to the case of two correlated noises
126: the proof originally proposed by Mao \etal in Ref.~\cite{marion}
127: that solutions to Eq.~\eqref{fp:logistic-noise-general2}, when started from
128: a positive initial condition, never become negative almost surely.
129:
130: \section{The Fokker-Planck equation}
131:
132:
133: The problem of constructing a Fokker-Planck equation corresponding to
134: a~process driven by two correlated Gaussian white noises has been first
135: discussed in Ref.~\cite{Telejko}, where the two noises have been decomposed
136: into two independent processes. The same result has been later re-derived
137: in \cite{cao94}, where the authors have attempted to avoid an explicit
138: decomposition of the noises but eventually resorted to a disguised form of
139: the decomposition. The general Langevin equation
140:
141: \begin{equation}\label{fp:general}
142: \dot x = h(x) + g_1(x)\xi_{\text{m}}(t) + g_2(x)\xi_{\text{a}}(t)\,,
143: \end{equation}
144:
145: \noindent where $x(t)$ is a one-dimensional process and $\xi_{\text{m},\text{a}}$ are
146: as in Eq.~\eqref{fp:logistic-noise-general2}, leads to the following Fokker-Planck
147: equation in the Ito interpretation:
148:
149: \begin{subequations}\label{fp:general-FP}
150: \begin{equation}
151: \frac{\partial P(x,t)}{\partial t} =
152: -\frac{\partial}{\partial x}h(x)P(x,t) +
153: \frac{1}{2}\frac{\partial^2}{\partial x^2}B(x)P(x,t)\,,
154: \end{equation}
155: \noindent where
156: \begin{equation}
157: B(x) = [g_1(x)]^2 + c\,g_1(x)g_2(x) + [g_2(x)]^2\,.
158: \end{equation}
159: \end{subequations}
160:
161: In the case of Eq.~\eqref{fp:logistic-noise-general2} the corresponding Fokker-Planck
162: equation therefore reads
163:
164: \begin{equation}\label{fp:FP}
165: \frac{\partial P(x,t)}{\partial t} =
166: -\frac{\partial}{\partial x}
167: \left[(a - bx)xP(x,t)\right]
168: +\frac{1}{2}\frac{\partial^2}{\partial x^2}
169: \left[x^2(p^2 - 2cpqx + q^2x^2)P(x,t)\right].
170: \end{equation}
171:
172: \noindent It is apparent that the absolute signs of the two noise amplitudes
173: do not influence the solutions to the above equations, only their relative
174: sign, $\text{sgn}(pq)$, does. In the following we will assume that $\text{sgn}(pq)=+1$.
175: This comes at no loss to the generality as Eq.~\eqref{fp:FP} is invariant
176: under a simultaneous change of signs of $pq$ and the correlation coefficient,~$c$.
177:
178: Stationary solutions to Eq.~\eqref{fp:FP} are the normalizable
179: solutions to \cite{Risken,Gardiner}
180:
181: \begin{equation}\label{fp:stationary}
182: x^2(p^2 - 2cpqx + q^2x^2)\frac{dP_{\text{st}}(x)}{dx}
183: +2x\left(2q^2x^2+(b{-}3cpq)x-(a{-}p^2)\right) P_{\text{st}}(x)
184: =0\,.
185: \end{equation}
186:
187: A slight modification of the argument presented originally in Ref.~\cite{marion}
188: shows that all solutions to Eq.~\eqref{fp:logistic-noise-general2} that start
189: from a~positive initial condition remain positive almost surely; see the Appendix
190: for a~proof. Physically speaking, this results from a presence of an absorbing barrier
191: at $x=0$ in Eq.~\eqref{fp:logistic-noise-general2}: should the population suddenly
192: drop to zero, it would stay there forever. The formal result of Ref.~\cite{marion},
193: extended here to the case of two noises present, ensures that the
194: population never actually becomes nonpositive although in certain cases (see below)
195: it may dynamically cluster in a close proximity of $x=0^+$.
196: Therefore, we can divide both sides of \eqref{fp:stationary}
197: by $x$ and obtain
198:
199: \begin{equation}\label{fp:stationary-x}
200: x(p^2 - 2cpqx + q^2x^2)\frac{d P_{\text{st}}(x)}{dx}
201: +2\left(2q^2x^2+(b-3cpq)x-(a-p^2)\right) P_{\text{st}}(x)
202: =0\,,
203: \end{equation}
204:
205: \noindent provided $P_{\text{st}}(x)$ is normalizable over the $x>0$ semiaxix.
206:
207: One may be tempted to try to immediately solve Eq.~\eqref{fp:stationary-x}
208: by standard methods, but a word of caution is needed here:
209: should the coefficient
210: at $dP_{\text{st}}/dt$ vanish, special care must be taken.
211:
212:
213: \subsection{The case of only one noise present}
214:
215: Before proceeding to the general case, we will discuss the special cases where
216: only one of the amplitudes $p$, $q$ does not vanish.
217:
218: \textit{Purely ``multiplicative'' noise.}
219: If $q=0$, the Langevin equation \eqref{fp:logistic-noise-general2} reduces to
220:
221: \begin{equation}\label{fp:logistic-noise-reduced1}
222: \dot x = (a+p\,\xi_{\text{m}}(t))x - bx^2
223: \end{equation}
224:
225: \noindent and we find for $P_{\text{st}}(x)$
226:
227: \begin{equation}\label{fp:special1}
228: P_{\text{st}}(x)
229: ={\cal N} x^{2(a-p^2)/p^2} \exp\left(-\frac{2bx}{p^2}\right),
230: \end{equation}
231:
232: \noindent where $\cal N$ is a normalization constant. This function is normalizable
233: for $x>0$ if $b>0$, or if the system is convergent, and
234:
235: \begin{subequations}
236: \begin{equation}\label{fp:condition1}
237: a-\frac{1}{2}p^2>0\,,
238: \end{equation}
239: which determines the behaviour of the distribution around $x=0$. Moreover, if
240:
241: \begin{equation}\label{fp:condition2}
242: a-p^2>0\,,
243: \end{equation}
244: \end{subequations}
245:
246: \noindent
247: $P_{\text{st}}(x)$ goes to zero as $x\to0^+$ and has a maximum at $x=(a-p^2)/b$.
248: Note that the conditions \eqref{fp:condition1}, \eqref{fp:condition2} coincide
249: with the conditions \eqref{fp:condition-mean}, \eqref{fp:condition-variance}
250: determining the properties of the linear process \eqref{fp:linear}. For
251: $p^2> a>\frac{1}{2}p^2$, the distribution is mildly divergent at $x=0$ and decreases
252: monotically with an increasing $x$.
253:
254: \textit{Purely ``additive'' noise.}
255: If $p=0$, the Langevin equation takes the form
256:
257: \begin{equation}\label{fp:logistic-noise-reduced2}
258: \dot x = (a - bx)x + qx^2\xi_{\text{a}}(t)
259: \end{equation}
260:
261: \noindent and we obtain for the stationary distribution
262:
263: \begin{equation}\label{fp:special2}
264: P_{\text{st}}(x) = \frac{\cal N}{x^4}\exp\left(-\frac{a-2bx}{q^2x^2}\right)
265: \end{equation}
266:
267: \noindent which is normalizable whenever $a>0$. Note that no bounds on $b$ are
268: imposed: The stationary distribution exists for both convergent and exploding systems.
269: This special case
270: falls into a broad category discussed recently by Mao \etal in Ref.~\cite{marion},
271: without further generalizations provided by the present paper.
272: The distribution \eqref{fp:special2} has a maximum at the positive root of
273:
274: \begin{equation}\label{fp:special2-minimum}
275: 2q^2x^2 + bx -a=0\,.
276: \end{equation}
277:
278: \subsection{The general case}
279:
280: If both noises are present and are not maximally correlated, $|c|\not=1$,
281: the solution to \eqref{fp:stationary-x} reads
282:
283: \begin{equation}\label{fp:solution-general}
284: P_{\text{st}}(x) =
285: \frac{{\cal N}x^{2(a-p^2)/p^2}}{(p^2{-}2cpqx {+} q^2x^2)^{(a+p^2)/p^2}}
286: \exp\left[
287: -\frac{2(bp{-}acq)\arctan\left(\frac{qx-cp}{\sqrt{1 - c^2}p}\right)}{\sqrt{1-c^2}\,p^2q}
288: \right].
289: \end{equation}
290:
291: \begin{figure}
292: \begin{center}
293: \includegraphics{fig1.eps}
294: \end{center}
295: \caption{Stationary distributions \eqref{fp:solution-general} in a strongly
296: correlated case, $c=0.99$. Clockwise, from top-left $q=0.1$, $q=0.5$, $q=0.6$,
297: and $q=5.0$. Other parameters, common for all panels, are $p=0.5$, $a=b=1$.}
298: \label{fp:fig1}
299: \end{figure}
300:
301: \noindent Since the exponential term is limited, the convergence (normalization)
302: properties of \eqref{fp:solution-general} are determined by those of the fractional
303: term. The denominator is always strictly positive.
304: For $x\to\infty$, $P_{\text{st}}(x) \sim x^{-4}$
305: for all possible values of parameters. $P_{\text{st}}(x)$ is therefore
306: normalizable if it does not diverge too rapidly at $x\to0^+$, or again if the
307: condition \eqref{fp:condition1} holds. Either in this case, no bounds on $b$ are imposed.
308: If the condition \eqref{fp:condition2} holds
309: as well, the distribution \eqref{fp:solution-general} approaches zero as $x\to0^+$,
310: but this condition is no longer associated with the presence of a maximum.
311: The maximum of $P_{\text{st}}(x)$ coincides with the
312: positive root of
313:
314: \begin{equation}\label{fp:root}
315: 2q^2x^2 + (b-3cpq)x -(a-p^2) =0\,,
316: \end{equation}
317:
318: \noindent cf.~Eq.~\eqref{fp:stationary-x}, provided such a root exists.
319: It certainly does for $a-p^2>0$, but it can appear also for
320: $p^2>a>\frac{1}{2}p^2$, where the distribution is mildly divergent.
321: Example stationary distributions in a strongly correlated case are presented
322: on Fig.~\ref{fp:fig1}. For large values of the additive noise strength, $q$,
323: the distributions are highly skewed and squeezed against the $x=0$ axis.
324: For comparison, on Fig.~\ref{fp:fig2} we show example stationary distributions
325: for the uncorrelated and a strongly anticorrelated cases. The distributions
326: presented are skewed and much wider than the distribution from Fig.~\ref{fp:fig1}
327: with the same value of $q=0.5$. These distributions
328: also get squeezed as the additive noise strength becomes large.
329: Note that the distribution corresponding to the anticorrelated noises is already
330: more squeezed than the distribution for the uncorrelated case.
331:
332: \begin{figure}
333: \begin{center}
334: \includegraphics{fig2.eps}
335: \end{center}
336: \caption{Stationary distributions for the uncorrelated ($c=0$, left panel) and
337: a~strongly anticorrelated ($c=-0.99$, right panel) cases. Other parameters are
338: $a=b=1$, $p=q=0.5$.}
339: \label{fp:fig2}
340: \end{figure}
341:
342: If the distribution \eqref{fp:solution-general} is normalizable, it has a~convergent
343: mean and a~variance. Its higher moments are divergent.
344:
345:
346: \subsection{The maximally correlated case}
347:
348: If the two noises are maximally (anti)correlated, $c=\pm1$,
349: Eq.~\eqref{fp:stationary-x} takes the form
350:
351: \begin{equation}\label{fp:stationary-resonant}
352: x(p\mp qx)^2\frac{dP_{\text{st}}}{dx}
353: + 2\left[(p\mp qx)^2 + q^2x^2 +(b\mp pq)x -a\right]P_{\text{st}} = 0\,,
354: \end{equation}
355:
356: \noindent leading to the following candidate solution:
357:
358: \begin{equation}\label{fp:stationary-resonant-solution}
359: P_{\text{st}}(x) = {\cal N}
360: \frac{x^{2(a-p^2)/p^2}}{(p\mp qx)^{2(a+p^2)/p^2}}
361: \exp\left[\mp\frac{2(bp\mp aq)}{pq(p\mp qx)}\right].
362: \end{equation}
363:
364: With our sign convention adopted, $\text{sgn}(pq)=+1$, this solution
365: is normalizable if $c=-1$ and the condition
366: \eqref{fp:condition1} holds. The distribution \eqref{fp:stationary-resonant-solution}
367: with the ``$+$'' sign can be obtained from
368: Eq.~\eqref{fp:solution-general} by taking the limit $c\to-1$.
369: This distribution decreases as $x^{-4}$ with $x\to\infty$.
370: The maximum
371: of \eqref{fp:stationary-resonant-solution}, if it exists, coincides with the
372: positive root of Eq.~\eqref{fp:root} with $c=-1$.
373:
374: The case of $c=+1$ is more challenging. First, if the resonant condition
375:
376: \begin{equation}\label{fp:resonantcondition}
377: bp-aq=0
378: \end{equation}
379:
380: \noindent holds, Eq.~\eqref{fp:stationary-resonant} is solved by
381:
382: \begin{equation}\label{fp:delta}
383: P_{\text{st}}(x) = \delta\left(x-\frac{p}{q}\right)
384: \end{equation}
385:
386: \noindent regardless of the value of $p$.
387: This result is stronger than that reported in Ref.~\cite{appb}
388: where we could predict a $\delta$-shaped distribution only in the $a\nobreak-\nobreak{}p^2>0$ case,
389: or when the variance of the corresponding linear system was convergent, as otherwise
390: any predictions based on the linear system failed. Note that with the condition
391: \eqref{fp:resonantcondition} statisfied, Eq.~\eqref{fp:logistic-noise-general2}
392: reduces to a rescaled form of Eq.~\eqref{fp:logistic-noise-basic}.
393:
394: If $c=+1$ and the condition \eqref{fp:resonantcondition} does not hold,
395: Eq.~\eqref{fp:stationary-resonant} does not have a~normalizable solution.
396: This observation is slightly surprising, but formally speaking, it results
397: from the fact that the double limit
398:
399: \begin{equation}\label{fp:doublelimit}
400: \lim\limits_{\begin{array}{c}\scriptstyle x\to p/q\\\scriptstyle c\to1\end{array}}\!\!
401: \exp\left[
402: -\frac{2(bp-acq)\arctan\left(\frac{qx-cp}{\sqrt{1 - c^2}p}\right)}{\sqrt{1-c^2}\,p^2q}
403: \right]
404: \end{equation}
405:
406: \noindent does not exits: Its value depends on which route the singularity is approached.
407: The nonexistence of stationary solutions in the fully correlated, non-resonant
408: case is, therefore, related to the essential singularity of the complex exponential
409: at infinity. The fact that with $c=+1$, $bp-aq\not=0$, the drift and diffusive terms in
410: Eq.~\eqref{fp:stationary-resonant} both vanish, but at different points, is the physical
411: reason for this apparent oddity: If the population gets located around the point of the
412: vanishing diffusion, it is washed away by the drift, and if it gets located around
413: the point of the vanishing drift, it diffusively leaks from there.
414: Nevertheless, if $a-\frac{1}{2}p^2>0$, in numerical simulations the cases of
415: $c=1$ and $c=1-\varepsilon$ with $0<\varepsilon\ll1$ are undistinguishable. In the
416: latter case, the distribution \eqref{fp:solution-general} is perfectly
417: normalizable.
418:
419: \section{Resonant effects and the shape of the stationary distribution}
420:
421: Perhaps the most important prediction based on the analysis of the linear equation
422: \eqref{fp:linear} and discussed in Ref.~\cite{appb} is that, for certain values of parameters,
423: the variance of the process $x(t)$ should, in the asymptotic regime, first shrink,
424: reach a minimum, and then grow as a function of the additive noise strength, $q$.
425: As we have mentioned before, these are heuristic, intuitive conclusions based on
426: the behaviour of the linear system associated with the logistic process, but because
427: of the complicated relation between the moments of these two processes, they do not
428: amount to a formal proof. In Ref.~\cite{appb} we have confirmed these predictions
429: numerically for a certain range of the additive noise strengths. As we have seen
430: above, in the fully correlated and resonant case,
431: the stationary distribution becomes $\delta$-shaped and its variance indeed vanishes,
432: much as predicted by the linear system. Since we now know the mathematically exact
433: stationary distributions, we can test the behaviour of the variance in the general case
434: directly.
435:
436: \begin{figure}
437: \begin{center}
438: \includegraphics{fig3.eps}
439: \end{center}
440: \caption{The variance $\left\langle x^2\right\rangle-\left\langle x\right\rangle^2$
441: determined form the distribution \eqref{fp:solution-general} as a function of the
442: additive noise strength, $q$. Main panel: $p=0.5$, the curves, from bottom to top,
443: correspond to $c=0.99$, $c=0.90$, $c=0.75$, $c=0.50$, $c=0.25$, $c=0$, and $c=-0.25$,
444: respectively. Inset: $p=1.1$, the curves correspond, from bottom to top,
445: to $c=0.99$, $c=0.98$, $c=0.97$, $c=0.96$, $c=0.95$, and $c=0.94$, respectively.
446: Other parameters, common for all curves presented, are $a=b=1$.}
447: \label{fp:fig3}
448: \end{figure}
449:
450:
451: Recall that the distribution \eqref{fp:solution-general} has the two first moments
452: convergent whenever it is normalizable. Unfortunately, analytical expressions for
453: these moments cannot be obtained, mainly due to the presence of the complicated
454: exponential term. Therefore, we have calculated the moments by numerically integrating
455: over the distribution \eqref{fp:solution-general}. Results are presented on
456: Fig.~\ref{fp:fig3}. If the distribution approaches zero as $x\to0^+$, or when the
457: condition \eqref{fp:condition2} is satisfied, and if the two noises are positively
458: correlated, $0<c<1$, the variance $\left\langle x^2\right\rangle-
459: \left\langle x\right\rangle^2$ displays a clear minimum as a function of the
460: additive noise strength, $q$. The minimum becomes shallower as the correlations
461: decrease towards zero, where it eventually disappears. It is not present for the
462: negative correlations or when one of the amplitudes vanishes. These effects agree
463: with predictions based on the linear system \eqref{fp:linear}. The presence of
464: the minimum of the variance is a clear and beneficial effect of positive correlations
465: between the two noises. However, for larger values of $q$ a new phenomenon appears:
466: The variance starts decreasing again. This is because for large values of the additive noise,
467: the stationary distribution
468: gets squeezed against the $x=0$ axis, cf.~Fig.~\ref{fp:fig1} above. This effect
469: cannot be predicted within the linear approach --- note that the process
470: described by Eq.~\eqref{fp:linear} has a support that formally spreads
471: over the entire real axis and, moreover, is Gaussian whenever the condition
472: \eqref{fp:condition2} holds, while the noisy logistic process is restricted to the
473: positive semiaxix.
474:
475: If the stationary distribution mildly diverges at zero, or if $p^2>a>\frac{1}{2}p^2$,
476: a distinct minimum in the variance of $x$ also appears but it is present only for
477: fairly large (and positive) values of the correlation coefficient,
478: cf.\ the inset on Fig.~\ref{fp:fig3}.
479: Note that this effect cannot possibly be predicted by analysing the linear system
480: \eqref{fp:linear} as in this regime the variance of the linear process diverges
481: and any predictions break.
482:
483: \begin{figure}
484: \begin{center}
485: \includegraphics{fig4a.eps}
486: \includegraphics{fig4b.eps}
487: \end{center}
488: \caption{Stochastic resonance in the system \eqref{fp:logistic-noise-oscillations}.
489: The upper panel --- the condition \eqref{fp:condition2} is satisfied,
490: $p=0.5$. The lower panel --- the condition \eqref{fp:condition2} is not satisfied,
491: $p=1.1$. Other parameters, common for the two panels, are $a=b=1$, $A=0.5$,
492: $\Omega=2\pi$. Curves presented correspond, back to front, to $c=1.0$, $0.99$, $0.9$ (lower panel only),
493: $0.75$, $0.5$, $0.25$, $0.0$, and $-0.25$, respectively.}\label{fp:fig4}
494: \end{figure}
495:
496:
497:
498: \section{Stochastic resonance}
499:
500: We now assume that parameters of the logistic process are not only subjected to noise,
501: but also to periodic, deterministic perturbations, resulting for example from seasonal
502: changes in the environment. Specifically, we consider
503:
504: \begin{equation}\label{fp:logistic-noise-oscillations}
505: \dot x = (a+p\,\xi_{\text{m}}(t))x - (b+A\sin(\Omega t +\varphi) + q\,\xi_{\text{a}}(t))x^2\,.
506: \end{equation}
507:
508: We have shown analytically in Ref.~\cite{appb} that the linear system associated
509: with Eq.~\eqref{fp:logistic-noise-oscillations} displays a stochastic resonance (SR) if the
510: noises are positively correlated. SR is one of the most spectacular examples of
511: a~constructive role of noise --- see Ref.~\cite{SRreview} for a review.
512: Because we do not know exact solutions of
513: a~time-dependent Fokker-Planck equation corresponding to
514: Eq.~\eqref{fp:logistic-noise-oscillations}, we will demonstrate the SR phenomenon
515: numerically. We will use the Signal-To-Noise Ratio (SNR) as a~measure of the SR:
516:
517: \begin{equation}\label{appb:SNR-def}
518: \mathrm{SNR} =
519: 10 \log_{10}\frac{P_{\text{signal}}}{P_{\text{noise}}(\omega=\Omega)}\,,
520: \end{equation}
521: %
522: \noindent where $P_{\text{signal}}$ is the height of the peak in the
523: power spectrum at the driving frequency and $P_{\text{noise}}$ is the noise-induced
524: background.
525:
526: We have solved the equation \eqref{fp:logistic-noise-oscillations} numerically
527: with the Euler-Maryuama algorithm and a timestep equal $2^{-16}$. The GWNs have
528: been generated by the Marsaglia algorithm \cite{Marsaglia} and the famous
529: Mersenne Twister \cite{Mersenne} has been used as the underlying uniform generator.
530: We have let the system to equilibrate, run the simulations
531: for $2^{25}$ steps and collected the results of every $2^9$-th step,
532: calculated the power spectrum, calculated the
533: SNR and averaged the results over 128 realizations of the stochastic
534: processes and the initial phases,~$\varphi$. Selected results are presented on
535: Fig.~\ref{fp:fig4}.
536:
537: The upper panel corresponds to the situation when the condition \eqref{fp:condition2}
538: holds, or when the noisy logistic process without the periodic signal has a~convergent
539: variance. We can clearly see that the system \eqref{fp:logistic-noise-oscillations}
540: displays a~SR for positive correlations between the two noises and disappears
541: for $c\leqslant0$: For positive correlations between the noises, there is a certain
542: level of the ``additive'' noise that maximizes the impact that seasonal
543: changes in the environment have on the population. As we have
544: shown in the preceding Section, this range of parameters corresponds to the presence
545: of the minimum in the variance of \eqref{fp:solution-general}. One may be tempted
546: to conclude that the SR and the minimum of the variance are two facets of the
547: same phenomenon, much as in the linear case. However, the lower panel of Fig.~\ref{fp:fig4},
548: corresponding to the situation when the condition \eqref{fp:condition2} is not
549: satisfied, shows that this is not the case. The SR, albeit much weaker than in the
550: previous case, is clearly present even when the signal-free system no longer displays
551: a minimum of the variance. A minimum of the variance and the stochastic resonance
552: are two different constructive effects of positive correlations between the noises.
553:
554:
555: \section{Conclusions}
556:
557: In this paper we have constructed stationary distributions corresponding to a noisy
558: logistic process driven by two correlated GWNs. These distributions are restricted
559: to the positive semiaxis and if they are normalizable, they have two (and only two)
560: convergent moments. In particular, if the noises are maximally correlated and a certain
561: resonant condition holds, the stationary distribution is $\delta$-shaped, which
562: has been reported previously as a result of many numerical simulations. Surprisingly,
563: if the noises are maximally correlated but the resonant condition does not hold,
564: the process does not have a stationary distribution,
565: even though it can numerically manifest itself as if it had one.
566:
567: Positive correlations between the noises lead to a minimum of the variance of the
568: noisy logistic process and to a stochastic resonance if the parameters of the
569: system undergo additional periodic changes. As we have numerically demonstrated,
570: these are two different effects. By constructing the exact stationary distributions,
571: we have extended our previous analysis of the system performed mainly by formally
572: converting the system into a~linear one. Several features of the system,
573: and ``squeezing'' of the stationary distribution in case of
574: a strong ``additive'' noise in particular, cannot be described by analysing the
575: linear process. This
576: is because the linear process \eqref{fp:linear} is Gaussian if it has two convergent
577: moments but the nonlinear logistic process is not.
578: Nevertheless, there are nice parallels between the
579: properties of the logistic process and its formal linearization:
580: If $a-\frac{1}{2}p^2>0$, the linear process has a convergent mean and the
581: logistic process has a normalizable stationary distribution that decreases for large
582: $x$ as $x^{-4}$. If $a-p^2>0$, the stationary distribution of the logistic process approaches
583: zero as $x\to0^+$ and the linear process has a convergent variance.
584: If $\frac{1}{2}p^2<a<p^2$, the stationary distribution of the noisy logistic
585: process mildly diverges at $x=0^+$ and the population dynamically clusters around that
586: point. Note that for $a>p^2$, the stationary distribution has a maximum and the population
587: is actually pushed away from $x=0^+$. Thus, the level of the multiplicative noise
588: $p^2=a$ marks a qualitative change in the population described by the noisy logistic
589: equation.
590:
591: It is, perhaps, surprising that a stationary distribution of the noisy logistic
592: process may exist even if the corresponding deterministic process is clearly
593: divergent. The fact that noise can prevent a population from
594: exploding has been recently reported by Mao \etal in Ref.~\cite{marion} for
595: a~more restricted, in a sense, class of systems. The present work is
596: an extension of this research to a~class that includes two correlated
597: sources of the noise.
598:
599: I am very grateful to Prof.\ Andrzej Fuli\'nski for his thorough criticism of
600: a~draft version of this paper.
601:
602: \appendix
603:
604: \section{}
605:
606: To show that solutions to Eq.~\eqref{fp:logistic-noise-general2} remain positive
607: almost surely when started from a positive initial condition, we first
608: decompose the correlated noises $\xi_{\text{m}}$, $\xi_{\text{a}}$ into
609: two independent processes:
610:
611: \begin{subequations}\label{fp:decomposition}
612: \begin{eqnarray}
613: \xi_{\text{m}}(t) &=& \xi(t)\,,\\
614: \xi_{\text{a}}(t) &=& c\xi(t) + \sqrt{1-c^2}\,\eta(t)
615: \end{eqnarray}
616: \end{subequations}
617:
618: \noindent where $\eta(t)$, $\xi(t)$ are two identical, uncorrelated GWNs.
619: The decomposition \eqref{fp:decomposition} is a variant of the method originally
620: used in Ref.~\cite{Telejko}.
621: We now cast Eq.~\eqref{fp:logistic-noise-general2} in a form customarily used by
622: mathematicians:
623:
624: \begin{equation}\label{fp:wiener}
625: dx = (ax-bx^2)\,dt + (px-qcx^2)\,du -q\sqrt{1-c^2}\,x^2\,dw\,,
626: \end{equation}
627:
628: \noindent where $du$, $dw$ are differentials over two identical, independent
629: Wiener processes. Incidentally, observe that the Fokker-Planck equation \eqref{fp:FP}
630: follows immediately from Eq.~\eqref{fp:wiener}.
631:
632:
633: \newtheorem{positive}{Theorem}
634:
635: \begin{positive}\label{fp:theorem}
636: If the initial condition $x_0>0$, for any $q\not=0$ the solution to
637: Eq.~\eqref{fp:wiener} remains positive for all $t>0$ almost surely.
638: \end{positive}
639:
640: \begin{proof}
641:
642: The proof of this theorem follows closely that of Theorem~2.1 from the work of Mao
643: \etal \cite{marion} and we encourage readers interested in mathematical details to
644: familiarize with that proof first; to save the space, we will show only this
645: part in which the proof of Theorem~\ref{fp:theorem} differs from that of Mao \etal
646:
647: First, the authors of Ref.~\cite{marion} consider a multispecies (multidimensional)
648: system, while we restrict ourselves to a simpler single-species case.
649:
650: Second, the proof is based on properties of the function
651:
652: \begin{equation}\label{fp:V}
653: V(s) = \sqrt{s} - 1 - {\textstyle\frac{1}{2}}\ln s\,.
654: \end{equation}
655:
656: \noindent This function is nonnegative for any $s>0$. We calculate $V(x(t))$ along
657: the trajectory generated by Eq.~\eqref{fp:wiener} with an initial condition $x_0>0$
658: and calculate the stochastic differential of $V(x(t))$ using Ito formula.
659: Because Eq.~\eqref{fp:wiener} differs from that considered by Mao \etal, we obtain
660: a slightly different expression. Specifically, if $x(t)>0$,
661:
662: \begin{gather}
663: dV(x(t)) = {\textstyle\frac{1}{2}}\left(x^{-1/2}-x^{-1}\right)
664: \!
665: \left[(ax{-}bx^2)\,dt + (px{-}qcx^2)\,du -q\sqrt{1{-}c^2}x^2\,dw\right]
666: \nonumber\\
667: \label{fp:dV}
668: {}+
669: {\textstyle\frac{1}{4}}\left(x^{-2}-{\textstyle\frac{1}{2}}x^{-3/2}\right)
670: \left[(px-qcx^2)^2+q^2(1-c^2)x^4\right]\,dt
671: \end{gather}
672:
673: \noindent The second term in \eqref{fp:dV} would be absent if the noises were
674: interpreted in the Stratonovich sense. After a simple algebra,
675:
676: \begin{eqnarray}
677: dV(x(t))&=&\left[{\textstyle\frac{1}{2}}\left(x^{1/2}-1\right)(a{-}bx)
678: +{\textstyle\frac{1}{4}}\left(1-{\textstyle\frac{1}{2}}x^{1/2}\right)(p^2{-}2pqcx{+}q^2x^2)\right]
679: dt
680: \nonumber\\\label{fp:dV2}
681: &-&{\textstyle\frac{1}{2}}\left(x^{1/2}-1\right)(p-qcx)\,du
682: -{\textstyle\frac{1}{2}}\left(x^{1/2}-1\right)q\sqrt{1-c^2}\,x\,dw\,.
683: \nonumber\\
684: \end{eqnarray}
685:
686: If $q=0$ and $b<0$, or when the corresponding deterministic system explodes, the
687: coefficient at $dt$ in \eqref{fp:dV2} may assume arbitrarily large values.
688: On the contrary, for any $q\not=0$ and regardless of the sign of $b$,
689: this coefficient is bounded from above by a certain positive number $K$.
690: Thus\footnote{Some subtleties of the notation are omitted here, see~\cite{marion}
691: for a fully rigorous treatment.}
692:
693: \begin{eqnarray}\label{fp:dV-calka}
694: \int_0^T dV(x(t)) &\leqslant& KT
695: -\int_0^T {\textstyle\frac{1}{2}}\left(x^{1/2}-1\right)(p-qcx)\,du
696: \nonumber\\
697: &&\hphantom{KT}
698: -\int_0^T {\textstyle\frac{1}{2}}\left(x^{1/2}-1\right)q\sqrt{1-c^2}\,x\,dw\,,
699: \end{eqnarray}
700:
701: \noindent where $T$ is a time such that $x(t)$ is positive for
702: $0\leqslant t<T$ almost surely. By taking the expectation values, we obtain
703:
704: \begin{equation}\label{fp:V(t)}
705: \left\langle V(x(T))\right\rangle \leqslant V(x_0) + KT\,.
706: \end{equation}
707:
708: The rest of the proof now proceeds exactly as in Ref.~\cite{marion} to
709: show that $T=\infty$.
710: \end{proof}
711:
712: \begin{thebibliography}{10}
713:
714:
715: \bibitem{Leung} H. K. Leung, Phys. Rev. A {\bf37}, 1341 (1988).
716:
717: \bibitem{appb} P. F. G\'ora, Acta Phys. Pol. B {\bf35}, 1583 (2004) .
718:
719: \bibitem{marion} X. Mao, G. Marion, and E. Renshaw, Stochastic Process. Appl.
720: {\bf 97}, 95 (2002).
721:
722: \bibitem{njp} P. F. G\'ora, New J. Phys. {\bf7}, 36 (2005).
723:
724: \bibitem{Telejko} A. Fuli\'nski and T. Telejko, Phys. Lett. A {\bf 152} (1991) 11.
725:
726: \bibitem{cao94} Wu Da-jin, Cao Li, and Ke Sheng-zi, Phys. Rev. E {\bf50}, 2496 (1994).
727:
728: \bibitem{Risken} H. Risken, \textit{The Fokker-Planck Equation}
729: (Springer, Berlin, 1984).
730:
731: \bibitem{Gardiner} C. W. Gardiner, \textit{Handbook of Stochastic Methods}
732: (Springer, Berlin, 1993).
733:
734: \bibitem{SRreview} L. Gammaitoni, P. H\"anggi, P. Jung, and F. Marchesoni,
735: Rev. Mod. Phys. {\bf70}, 223 (1998).
736:
737: \bibitem{Marsaglia} G. Marsaglia and T. A. Bray, SIAM Review {\bf6}, 260 (1964);
738: A. J. Kinderman and J. G. Ramage, J. Amer. Statist. Assoc. {\bf 71}, 893 (1976);
739: R. Wieczorkowski and R. Zieli\'nski, \textit{Komputerowe generatory liczb losowych}
740: (WNT, Warszawa, 1997) (in Polish).
741:
742: \bibitem{Mersenne} M. Matsumoto and T. Nishimura, ACM Trans. on Modeling and
743: Computer Simulation, {\bf 8}, 3 (1998).
744:
745: \end{thebibliography}
746:
747: \end{document}
748: