1: %% This document created by Scientific Word (R) Version 2.0
2: %% Starting shell: mathart1
3:
4:
5: \documentclass[12pt,thmsa]{article}
6: \usepackage{amssymb}
7: %%%%%%
8: \usepackage{sw20bams}
9:
10: %TCIDATA{TCIstyle=Article/art1.lat,bams,mathart1}
11:
12: \input tcilatex
13: \begin{document}
14:
15: \author{Steven R. Finch}
16: \title{CLT\ Variance Associated with Baxendale's SDE}
17: \date{September 30, 2008}
18: \maketitle
19:
20: \begin{abstract}
21: Simple analysis of the leftmost eigenvalue of Ince's equation (a boundary
22: value problem with periodicity) resolves an open issue surrounding a
23: stochastic Lyapunov exponent. Numerical verification is also provided.
24: \end{abstract}
25:
26: \footnotetext{
27: Copyright \copyright\ 2008 by Steven R. Finch. All rights reserved.}Let $a>b$
28: and $\sigma >0$. Consider the stochastic differential equation (SDE)
29: \[
30: \begin{array}{ccc}
31: dX_t=\left(
32: \begin{array}{cc}
33: a & 0 \\
34: 0 & b
35: \end{array}
36: \right) X_t\,dt+\sigma \left(
37: \begin{array}{cc}
38: 0 & -1 \\
39: 1 & 0
40: \end{array}
41: \right) X_t\circ dW_t, & & X_t\in \Bbb{R}^2
42: \end{array}
43: \]
44: where $W_t$ is scalar Brownian motion with unit variance and $\circ $
45: denotes the use of Stratonovich calculus. A measure of the stability of such
46: a system is provided by the (almost-sure) Lyapunov exponent
47: \[
48: \ell =\lim_{t\rightarrow \infty }\frac 1t\ln |X_t|
49: \]
50: for $X_0\neq 0$. Baxendale \cite{Bx, KP} computed a formula for $\ell $ and,
51: further, proved an associated central limit theorem (CLT)
52: \[
53: \lim_{t\rightarrow \infty }\limfunc{P}\left( \frac{\dfrac 1t\ln |X_t|-\ell }{%
54: \dfrac s{\sqrt{t}}}\leq v\right) =\frac 1{\sqrt{2\pi }}\dint\limits_{-\infty
55: }^v\exp \left( -\frac{u^2}2\right) du.
56: \]
57: No formula has been known for the variance $s^2$ until now. Our calculation
58: of $s^2$ is based on the boundary value problem (BVP)
59: \[
60: y^{\prime \prime }(x)+c\sin (2x)y^{\prime }(x)+\left( \lambda -\mu \,c\cos
61: (2x)\right) y(x)=0,
62: \]
63: \[
64: y^{\prime }(0)=y^{\prime }(\pi )=0
65: \]
66: due to Ince \cite{Ic}, where $c=-(a-b)/\sigma ^2$ and $\mu \approx 0$. More
67: precisely, if $\lambda (\mu )$ is the leftmost eigenvalue of the Ince BVP,
68: given $\mu $, then
69: \[
70: \begin{array}{ccc}
71: \ell =\dfrac{a+b}2-\dfrac{\sigma ^2}2\lambda ^{\prime }(0), & & s^2=-\dfrac{%
72: \sigma ^2}2\lambda ^{\prime \prime }(0).
73: \end{array}
74: \]
75: In the next three sections, we discuss how Ince's equation arises and the
76: details of computing $\lambda ^{\prime }(0)$ and $\lambda ^{\prime \prime
77: }(0)$. The final section is devoted to numerical verification of the
78: preceding.
79:
80: \section{Equivalence}
81:
82: Let $\gamma /\sigma ^2$ denote the rightmost eigenvalue of the differential
83: operator \cite{Bx}
84: \[
85: \frac 12\frac d{dx^2}-\frac{a-b}{2\sigma ^2}\sin (2x)\frac d{dx}+\frac \mu
86: {\sigma ^2}(a\cos ^2x+b\sin ^2x)
87: \]
88: (which is obtained, in part, by projecting the solution $X_t$ of Baxendale's
89: SDE onto the unit circle). By the double angle formula for cosine, the BVP\
90: we wish to study is
91: \[
92: y^{\prime \prime }(x)-\frac{a-b}{\sigma ^2}\sin (2x)y^{\prime }(x)+\left[ -%
93: \frac{2\gamma }{\sigma ^2}+\frac \mu {\sigma ^2}\left( (a+b)+(a-b)\cos
94: (2x)\right) \right] y(x)=0
95: \]
96: hence $c=-(a-b)/\sigma ^2$ follows, as does
97: \[
98: \lambda =-\frac{2\gamma }{\sigma ^2}+\frac{a+b}{\sigma ^2}\mu ;
99: \]
100: hence
101: \[
102: \begin{array}{ccccc}
103: \lambda ^{\prime }=-\dfrac{2\gamma ^{\prime }}{\sigma ^2}+\dfrac{a+b}{\sigma
104: ^2} & & \therefore & & \ell =\dfrac{a+b}2-\dfrac{\sigma ^2}2\lambda
105: ^{\prime }(0),
106: \end{array}
107: \]
108: \[
109: \begin{array}{ccccc}
110: \lambda ^{\prime \prime }=-\dfrac{2\gamma ^{\prime \prime }}{\sigma ^2} & &
111: \therefore & & s^2=-\dfrac{\sigma ^2}2\lambda ^{\prime \prime }(0).
112: \end{array}
113: \]
114: This argument demonstrates the equivalence of Baxendale's setting (in terms
115: of $\gamma ^{\prime }(0)$ and $\gamma ^{\prime \prime }(0)$) and our setting
116: (in terms of $\lambda ^{\prime }(0)$ and $\lambda ^{\prime \prime }(0)$).
117:
118: \section{First Derivative}
119:
120: Let $y(x,\mu )$ denote the solution of the Ince BVP\ with $\lambda =\lambda
121: (\mu )$ that satisfies the initial conditions $y(0)=1$, $y^{\prime }(0)=0$.
122: Recall that $\lambda (\mu )$ is the leftmost such eigenvalue. Define
123: \[
124: z(x,\mu )=\frac \partial {\partial \mu }y(x,\mu ).
125: \]
126: Differentiate Ince's equation with respect to $\mu $, yielding
127: \[
128: z^{\prime \prime }(x,\mu )+c\sin (2x)z^{\prime }(x,\mu )+\left( \lambda
129: ^{\prime }(\mu )-c\cos (2x)\right) y(x,\mu )+\left( \lambda (\mu )-\mu
130: \,c\cos (2x)\right) z(x,\mu )=0.
131: \]
132: Call this $(*)$. Set $\mu =0$, yielding
133: \[
134: z^{\prime \prime }(x,0)+c\sin (2x)z^{\prime }(x,0)+\lambda ^{\prime
135: }(0)-c\cos (2x)=0
136: \]
137: because $\lambda (0)=0$ and $y(x,0)=1$. Multiply both sides of $(*)$ by $%
138: \exp (-\frac c2\cos (2x))$, yielding
139: \[
140: \left[ \exp (-\tfrac c2\cos (2x))z^{\prime }(x,0)\right] ^{\prime }=\left(
141: c\cos (2x)-\lambda ^{\prime }(0)\right) \exp (-\tfrac c2\cos (2x)).
142: \]
143: Since $z^{\prime }(0,0)=0$,
144: \[
145: \exp (-\tfrac c2\cos (2x))z^{\prime }(x,0)=\dint\limits_0^x\left( c\cos
146: (2\theta )-\lambda ^{\prime }(0)\right) \exp (-\tfrac c2\cos (2\theta
147: ))d\theta .
148: \]
149: Now $z^{\prime }(x,0)$ has period $\pi $, hence $z^{\prime }(\pi ,0)=0$ and
150: therefore
151: \[
152: \dint\limits_0^\pi \left( c\cos (2\theta )-\lambda ^{\prime }(0)\right) \exp
153: (-\tfrac c2\cos (2\theta ))d\theta =0.
154: \]
155: It follows that
156: \[
157: \lambda ^{\prime }(0)=\frac{c\dint\limits_0^\pi \cos (2\theta )\exp (-\tfrac
158: c2\cos (2\theta ))d\theta }{\dint\limits_0^\pi \exp (-\tfrac c2\cos (2\theta
159: ))d\theta }=c\frac{I_1(-\frac c2)}{I_0(-\frac c2)}
160: \]
161: where $I_0$, $I_1$ are modified Bessel functions \cite{AS}. This reproduces
162: Baxendale's formula for $\ell $.
163:
164: \section{Eigenfunction}
165:
166: In the next section, we will need a formula for $z(x,0)$. Note that
167: \[
168: y(x,\mu )\approx 1+\mu \,z(x,0)
169: \]
170: when $\mu \approx 0$, thus a byproduct of our work is an approximation of
171: the leftmost eigenfunction of Ince's equation to first order.
172:
173: From \cite{AS}, the following expressions hold:
174: \[
175: \exp (-\tfrac c2\cos (2\theta ))=I_0(-\tfrac c2)+2\dsum\limits_{k=1}^\infty
176: I_k(-\tfrac c2)\cos (2k\,\theta ),
177: \]
178: \[
179: \exp (-\tfrac c2\cos (2\theta ))\cos (2\theta )=I_1(-\tfrac
180: c2)+\dsum\limits_{k=1}^\infty \left( I_{k-1}(-\tfrac c2)+I_{k+1}(-\tfrac
181: c2)\right) \cos (2k\,\theta )
182: \]
183: hence the integrand $\left( c\cos (2\theta )-\lambda ^{\prime }(0)\right)
184: \exp (-\tfrac c2\cos (2\theta ))$ is equal to
185: \begin{eqnarray*}
186: &&c\dsum\limits_{k=1}^\infty \left( I_{k-1}(-\tfrac c2)+I_{k+1}(-\tfrac
187: c2)\right) \cos (2k\,\theta )-2c\frac{I_1(-\tfrac c2)}{I_0(-\tfrac c2)}%
188: \dsum\limits_{k=1}^\infty I_k(-\tfrac c2)\cos (2k\,\theta ) \\
189: &=&c\dsum\limits_{k=1}^\infty \stackunder{j_k}{\underbrace{\left(
190: I_{k-1}(-\tfrac c2)-2\frac{I_1(-\tfrac c2)}{I_0(-\tfrac c2)}I_k(-\tfrac
191: c2)+I_{k+1}(-\tfrac c2)\right) }}\cos (2k\,\theta ).
192: \end{eqnarray*}
193: Integrating once gives
194: \[
195: \exp (-\tfrac c2\cos (2x))z^{\prime }(x,0)=\frac c2\dsum\limits_{k=1}^\infty
196: \frac{j_k}k\sin (2k\,x).
197: \]
198: Integrating twice gives
199: \[
200: z(x,0)=\frac c2\dsum\limits_{k=1}^\infty \frac{j_k}k\dint\limits_0^x\exp
201: (\tfrac c2\cos (2\theta ))\sin (2k\,\theta )d\theta .
202: \]
203: Although we do not explicitly write out the integrals here, they are
204: elementary and can be computed symbolically for arbitrary $k$.
205:
206: \section{Second Derivative}
207:
208: Define
209: \[
210: w(x,\mu )=\frac{\partial ^2}{\partial ^2\mu }y(x,\mu )=\frac \partial
211: {\partial \mu }z(x,\mu ).
212: \]
213: Differentiate equation $(*)$ with respect to $\mu $, yielding
214: \[
215: \begin{array}{c}
216: w^{\prime \prime }(x,\mu )+c\sin (2x)w^{\prime }(x,\mu )+\lambda ^{\prime
217: \prime }(\mu )y(x,\mu )+2\left( \lambda ^{\prime }(\mu )-c\cos (2x)\right)
218: z(x,\mu ) \\
219: +\left( \lambda (\mu )-\mu \,c\cos (2x)\right) w(x,\mu )=0.
220: \end{array}
221: \]
222: Set $\mu =0$, yielding
223: \[
224: w^{\prime \prime }(x,0)+c\sin (2x)w^{\prime }(x,0)+\lambda ^{\prime \prime
225: }(0)+2\left( \lambda ^{\prime }(0)-c\cos (2x)\right) z(x,0)=0
226: \]
227: because $\lambda (0)=0$ and $y(x,0)=1$. We again multiply both sides by $%
228: \exp (-\frac c2\cos (2x))$:
229: \[
230: \left[ \exp (-\tfrac c2\cos (2x))w^{\prime }(x,0)\right] ^{\prime }=\left[
231: 2\left( c\cos (2x)-\lambda ^{\prime }(0)\right) z(x,0)-\lambda ^{\prime
232: \prime }(0)\right] \exp (-\tfrac c2\cos (2x)).
233: \]
234: Since $w^{\prime }(0,0)=0$,
235: \[
236: \exp (-\tfrac c2\cos (2x))w^{\prime }(x,0)=\dint\limits_0^x\left[ 2\left(
237: c\cos (2\theta )-\lambda ^{\prime }(0)\right) z(\theta ,0)-\lambda ^{\prime
238: \prime }(0)\right] \exp (-\tfrac c2\cos (2\theta ))d\theta .
239: \]
240: Now $w^{\prime }(x,0)$ has period $\pi $, hence $w^{\prime }(\pi ,0)=0$ and
241: therefore
242: \[
243: \dint\limits_0^\pi \left[ 2\left( c\cos (2\theta )-\lambda ^{\prime
244: }(0)\right) z(\theta ,0)-\lambda ^{\prime \prime }(0)\right] \exp (-\tfrac
245: c2\cos (2\theta ))d\theta =0.
246: \]
247: It follows that
248: \[
249: \lambda ^{\prime \prime }(0)=\frac 2{\pi \,I_0(-\frac c2)}\dint\limits_0^\pi
250: \left( c\cos (2\theta )-\lambda ^{\prime }(0)\right) z(\theta ,0)\exp
251: (-\tfrac c2\cos (2\theta ))d\theta
252: \]
253: where $z(\theta ,0)$ is defined via an infinite series in the preceding
254: section.
255:
256: \section{Numerical Verification}
257:
258: One way to compute $\lambda (\mu )$ is to construct the infinite tridiagonal
259: matrix \cite{Vm}
260: \[
261: M=\left(
262: \begin{array}{cccccc}
263: r_0 & \sqrt{2}q_{-1} & 0 & 0 & 0 & \\
264: \sqrt{2}q_0 & r_1 & q_{-2} & 0 & 0 & \\
265: 0 & q_1 & r_2 & q_{-3} & 0 & \\
266: 0 & 0 & q_2 & r_3 & q_{-4} & \\
267: 0 & 0 & & q_3 & r_4 & \ddots \\
268: & & & & \ddots & \ddots
269: \end{array}
270: \right)
271: \]
272: where $r_n=4n^2$ and $q_n=(-n+\mu /2)c$. The leftmost eigenvalue of $M$ is $%
273: \lambda (\mu )$.
274:
275: Another way to compute $\lambda (\mu )$ is to solve the continued fraction
276: equation \cite{Vm}
277: \[
278: -\frac \lambda 2=\dfrac{\left. p_0\;\;\;\;\;\;\;\;\;\;\right| }{\left|
279: 4\cdot 1^2-\lambda \right. }-\dfrac{\left. p_1\;\;\;\;\;\;\;\;\;\;\right| }{%
280: \left| 4\cdot 2^2-\lambda \right. }-\dfrac{\left.
281: p_2\;\;\;\;\;\;\;\;\;\;\right| }{\left| 4\cdot 3^2-\lambda \right. }-\dfrac{%
282: \left. p_3\;\;\;\;\;\;\;\;\;\;\right| }{\left| 4\cdot 4^2-\lambda \right. }-%
283: \dfrac{\left. p_4\;\;\;\;\;\;\;\;\;\;\right| }{\left| 4\cdot 5^2-\lambda
284: \right. }-\cdots
285: \]
286: where $p_n=(-n+\mu /2)(n+1+\mu /2)c^2$.
287:
288: High precision estimates of the derivatives of $\lambda (\mu )$ at zero are
289: found via
290: \[
291: \begin{array}{ccc}
292: \lambda ^{\prime }(0)\approx \dfrac{\lambda (\mu )}\mu , & & \lambda
293: ^{\prime \prime }(0)\approx \dfrac{\dfrac{\lambda (\mu )}\mu -c\dfrac{%
294: I_1(-\frac c2)}{I_0(-\frac c2)}}{\dfrac \mu 2}
295: \end{array}
296: \]
297: for $\mu \approx 0$. For example, when $a=1$, $b=-2$ and $\sigma =10$, we
298: obtain
299: \[
300: \ell =-0.4887503163943852244580286...,
301: \]
302: \[
303: s^2=0.0112485762885419873084837...
304: \]
305: as the CLT\ parameter values. As another example,
306: \[
307: \ell =0.3941998582469360577816389...,
308: \]
309: \[
310: s^2=0.3841476218435126147382099...
311: \]
312: when $\sigma =1$ (and $a$, $b$ remain unchanged). The same results were
313: obtained using exact formulation in sections 2 and 4, confirming our work.
314:
315: \section{Acknowledgements}
316:
317: Peter Baxendale suggested that I work on Ince's equation, Hans Volkmer
318: sketched out the derivation of $\lambda ^{\prime }(0)$ and $\lambda ^{\prime
319: \prime }(0)$, and Joe Keane found the generating function in \cite{AS} for $%
320: \exp (r\cos (\theta ))$. My grateful thanks go to all three!
321:
322: \begin{thebibliography}{9}
323: \bibitem{Bx} P. H. Baxendale, Moment stability and large deviations for
324: linear stochastic differential equations, \textit{Probabilistic Methods in
325: Mathematical Physics}, Proc. 1985 Katata/Kyoto conf., ed. K. It\^o and N.
326: Ikeda, Academic Press, 1987, pp. 31--54; MR0933817 (89c:60068).
327:
328: \bibitem{KP} P. E. Kloeden and E. Platen, \textit{Numerical Solution of
329: Stochastic Differential Equations}, Springer-Verlag, 1992, pp. 545--548;
330: MR1214374 (94b:60069).
331:
332: \bibitem{Ic} E. L. Ince, A linear differential equation with periodic
333: coefficients, \textit{Proc. London Math. Soc.} 23 (1925) 56--74.
334:
335: \bibitem{AS} M. Abramowitz and I. A. Stegun, \textit{Handbook of
336: Mathematical Functions}, Dover, 1972, pp. 375--376, eqns. (9.6.10) and
337: (9.6.34); MR1225604 (94b:00012).
338:
339: \bibitem{Vm} H. Volkmer, Coexistence of periodic solutions of Ince's
340: equation, \textit{Analysis} 23 (2003) 97--105; MR1983977 (2004c:34074).\\
341:
342: \begin{tabular}{lll}
343: & Steven R. Finch & \\
344: & Department of Statistics & \\
345: & Harvard University & \\
346: & Science Center & \\
347: & 1 Oxford Street & \\
348: & Cambridge, MA 02138 & \\
349: & \textit{Steven.Finch@inria.fr} &
350: \end{tabular}
351: \end{thebibliography}
352:
353: \end{document}
354: