1: \documentclass[12pt,a4paper,twoside]{article}

2: % \usepackage{ajr}

3: % defines commonly used commands, headings.

4: 

5: \newcommand{\D}[2]{\frac{\partial #2}{\partial #1}}

6: \newcommand{\DD}[2]{\frac{\partial^2 #2}{\partial #1^2}}

7: \newcommand{\Dn}[3]{\frac{\partial^#2 #3}{\partial #1^#2}}

8: \newcommand{\DDD}[2]{\frac{\partial^3 #2}{\partial #1^3}}

9: \newcommand{\DDDD}[2]{\frac{\partial^4 #2}{\partial #1^4}}

10: \newcommand{\cm}{\mbox{cm}}

11: \newcommand{\secs}{\mbox{s}}

12: \newcommand{\gm}{\mbox{gm}}

13: \renewcommand{\vec}[1]{\mbox{\boldmath$#1$}}

14: \newcommand{\Ord}[1]{{\cal O}\left(#1\right)}

15: \newcommand{\divv}{\vec\nabla\cdot}

16: \newcommand{\grad}{\vec\nabla}

17: \newcommand{\curl}{\vec\nabla\times}

18: \newcommand{\marginal}[1]{\marginpar{\raggedright\footnotesize\sf #1}}

19: 

20: % pagestyle, very naughtily crunching \MakeUppercase

21: \pagestyle{headings}

22: \renewcommand{\MakeUppercase}[1]{\textsf{#1}}

23: 

24: \usepackage{url}

25: \usepackage{graphicx}

26: 

27: \newcommand{\pde}{{\textsc{pde}}}

28: \newcommand{\cM}{{\cal M}}

29: \newcommand{\ur}{u^+}

30: \newcommand{\ul}{u^-}

31: \newcommand{\up}{u_{j+1}}

32: \newcommand{\um}{u_{j-1}}

33: \newcommand{\umm}{u_{j-2}}

34: \newcommand{\upp}{u_{j+2}}

35: \newcommand{\uj}{u_{j}}

36: \newcommand{\xjmh}{\xi_{j-1/2}}

37: \newcommand{\xjph}{\xi_{j+1/2}}

38: \newcommand{\xjpmh}{\xi_{j\pm1/2}}

39: \newcommand{\cA}{{\cal A}}

40: \newcommand{\cD}{{\cal D}}

41: \newcommand{\KS}{Ku\-ra\-mo\-to-Siva\-shin\-sky}

42: \newcommand{\reduce}{\textsc{reduce}}

43: 

44: %\usepackage{showkeys}

45: %\usepackage{hyperref}

46: 

47: \begin{document}

48: 

49: \title{\sf Holistic finite differences accurately model the dynamics of the

50: Kuramoto-Sivashinsky equation }

51: \author{T. MacKenzie   \and

52: A.J. Roberts \thanks{Dept Maths \& Comp, University of Southern

53: Qld, Toowoomba, Qld 4350, Australia.

54: \protect\url{mailto:mackenzi@usq.edu.au} and

55: \protect\url{mailto:aroberts@usq.edu.au} respectively.}}

56: \maketitle

57: 

58: \begin{abstract}

59: We analyse the nonlinear Kuramoto-Sivashinsky equation to develop

60: an accurate finite difference approximation to its dynamics. The

61: analysis is based upon centre manifold theory so we are assured

62: that the finite difference model accurately models the dynamics

63: and may be constructed systematically. The theory is applied after

64: dividing the physical domain into small elements by introducing

65: insulating internal boundaries which are later removed. The

66: Kuramoto-Sivashinsky equation is used as an example to show how

67: holistic finite differences may be applied to fourth order,

68: nonlinear, spatio-temporal dynamical systems. This novel centre

69: manifold approach is holistic in the sense that it treats the

70: dynamical equations as a whole, not just as the sum of separate

71: terms.

72: \end{abstract}

73: 

74: %\tableofcontents

75: 

76: 

77: \section{Introduction}

78: 

79: We continue exploring the new approach to finite difference

80: approximation introduced by Roberts \cite{Roberts98a} by

81: approximating the dynamics of solutions to the \KS{}

82: equation \cite{Kuramoto78,Hyman86b,Pomeau84}. In some

83: non-dimensional form we take the following partial differential

84: equation (\pde) to govern the evolution of $u(x,t)$:

85: \begin{equation}

86:     \D tu+u\D xu+R\DD xu+\DDDD xu=0\,.

87:     \label{Epde}

88: \end{equation}

89: This model equation includes the mechanisms of linear growth

90: $u_{xx}$ controlled by the parameter $R$, high-order dissipation,

91: $u_{xxxx}$,

92: and nonlinear advection/steepening, $uu_x$.  Consider implementing

93: the method of lines by discretising in $x$ and integrating in time

94: as a set of ordinary differential equations.  A finite difference

95: approximation to~(\ref{Epde}) on a regular grid in $x$ is

96: straightforward, say $x_j=jh$ for some grid spacing $h$.  For

97: example, the linear term

98: \begin{displaymath}

99:     \DD xu = \frac{\up-2\uj+\um}{h^2}+\Ord{h^2}\,.

100: \end{displaymath}

101: However, there are differing valid alternatives for the nonlinear

102: term $uu_x$: two possibilities are

103: \begin{equation}

104:     u\D xu= \frac{\uj(\up-\um)}{2h}+\Ord{h^2}

105:     =\frac{\up^2-\um^2}{4h}+\Ord{h^2}\,.

106:     \label{Ealt}

107: \end{equation}

108: Which is better?  The answer depends upon how the discretisation

109: of the nonlinearity interacts with the dynamics of other terms.

110: The conventional approach of considering the discretisation of

111: each term separately does not tell us.  Instead, in order to find

112: the best discretisation we consider the influence of all

113: terms in the equation in a holistic approach.

114: 

115: As introduced in \cite{Roberts98a} and discussed in \S\ref{Scm},

116: centre manifold theory \cite[e.g.]{Carr81,Roberts97a} has

117: appropriate characteristics to do this.  It addresses the

118: evolution of a dynamical system in a neighbourhood of a marginally

119: stable fixed point; based upon the linear dynamics the theory

120: guarantees that an accurate low-dimensional description of the

121: nonlinear dynamics may be deduced. For example the analysis herein

122: supports the first approximation in~(\ref{Ealt}) but with higher-order and

123: nonlinear corrections.  The analysis of the \KS{}

124: equation~(\ref{Epde}) in \S~\ref{Sbe} favours the

125: discretisation

126: \begin{eqnarray}

127:     &&

128:     \frac{d\uj}{dt} +\left[

129:     \frac{\uj(-\upp+9\up-9\um+\umm)}{16h}\right.

130:     \nonumber \\ &&

131:     \quad{}+\left.\left(\frac{\up^2-\um^2}{16h}\right)-\left(\frac{

132:     \upp\up-\umm\um}{48h}\right)

133:     \right]

134:     \nonumber \\ &&

135:     \quad{}+R\left(\frac{-\upp+16\up-30\uj+16\um-\umm}{12h^2}\right)

136:     \nonumber \\ &&

137:     \quad{}+\frac{\upp-4\up+6\uj-4\um+\umm}{h^4} = 0\,,

138:     \label{Efd}

139: \end{eqnarray}

140: as a low-order approximation.  Provided the initial conditions are

141: not too extreme, centre manifold theory assures us that such a

142: discretisation models the dynamics of~(\ref{Epde}) to errors

143: $\Ord{\|u\|^3,h^2}$.  Such accuracy on a relatively coarse grid is

144: extremely useful for such stiff \pde{}s.  Further, because the

145: centre manifold is composed of actual solutions to the dynamical

146: system, we are assured that equation~(\ref{Efd}) models the whole

147: of the \KS{} equation, \emph{independent} of its algebraic form.

148: 

149: The discretisation~(\ref{Efd}) is a low-order approximation,

150: centre manifold theory also provides systematic corrections.  Analysis

151: to higher orders in the nonlinearity, discussed in

152: \S~\ref{Sbe}, shows higher order corrections to the

153: discretisation of the nonlinear terms.  The specific finite

154: difference models presented here were derived by a computer

155: algebra program.  Computer algebra is an effective tool because of

156: the systematic nature of centre manifold theory \cite{Roberts96a}.

157: 

158: In this preliminary exploration of the approximation of the

159: \KS{} equation~(\ref{Epde}) we only consider an

160: infinite domain or strictly periodic solutions in finite domains.

161: Then all elements of the discretisation are identical by symmetry and

162: the analysis of all elements is simultaneous.  However, if physical

163: boundaries to the domain of the \pde{} are present, then those

164: elements near a physical boundary will need special treatment.

165: Further research is needed on this and other issues.

166: 

167: 

168: 

169: 

170: 

171: \section{Centre manifold theory underpins the fidelity}

172: \label{Scm}

173: 

174: Here we describe in detail one way to place the discretisation of the

175: \KS{} equation~(\ref{Epde}) within the purview of

176: centre manifold theory.

177: 

178: The discretisation is established via an equi-spaced grid of

179: collocation points, $x_j=jh$ say, for some small spacing $h$. Here

180: we scale the \KS{} to the scale of the grid spacing $h$. Thus our

181: view of the dynamics shrinks with $h$.  This is different to the

182: analysis in \cite{Roberts98a} and allows the linear dynamics to be

183: dominated by simply the highest order spatial derivative term.  We

184: work on the scale of the grid by transforming~(\ref{Epde}) to the

185: following space and time scales: $\xi=x/h$ and $\tau=t/h^4$,

186: giving

187: \begin{equation}

188:      \D \tau u+h^3u\D\xi u+h^2R\DD \xi u+\DDDD \xi u=0\,.

189:      \label{Eks}

190: \end{equation}

191: Then the crucial step: at midpoints

192: $\xi_{j+1/2}=(\xi_j+\xi_{j+1})/2$ artificial boundaries are

193: introduced:

194: \begin{eqnarray}

195:     \left[\begin{array}{c}\D \xi\ur

196:     -\D \xi\ul \\

197:     \frac{\partial^3 \ur}{\partial \xi^3}

198:     -\frac{\partial^3 \ul}{\partial \xi^3}

199:     \end{array}\right]

200:     &=& \left[\begin{array}{c}0\\0\end{array}\right]\,,

201:     \label{Ebcsd}\\

202:     \left(\frac{1-\gamma}{2}\right)

203:     \left[\begin{array}{c}\D \xi\ur

204:     +\D \xi\ul \\

205:     \frac{\partial^3 \ur}{\partial \xi^3}

206:     +\frac{\partial^3 \ul}{\partial \xi^3}

207:     \end{array}\right]

208:     &=&\gamma \cA

209:     \left[\begin{array}{c}

210:     \ur-\ul\\

211:     \frac{\partial^2 \ur}{\partial \xi^2}

212:     -\frac{\partial^2 \ul}{\partial \xi^2}

213:     \end{array}\right]\,,

214:     \label{Esbci}

215: \end{eqnarray}

216: where $\ur$ is just to the right of a midpoint and $\ul$ to the

217: left. The introduction of the near identity operator

218: \begin{equation}

219:     \cA

220:     =1+\frac{\partial^2_\xi}{12}-\frac{\partial^4_\xi}{720}+

221:     \frac{\partial^6_\xi}{30240}+\cdots

222:     =\frac{\partial_\xi}{2}\coth\left(\frac{\partial_\xi}{2}\right)\,,

223:     \label{Efudge}

224: \end{equation}

225: ensures that high-order approximations to linear terms are

226: obtained exactly as discussed in \cite[\S4]{Roberts98a}:  it is

227: remarkable that the exactly equivalent operator works for both

228: Burgers' equation and the \KS{} equation.  These boundaries divide

229: the domain into a set of elements, the $j$th element centred upon

230: $\xi_j$ and of width $\Delta \xi=1$.  A non-zero value of the

231: parameter $\gamma$ couples these elements together so that when

232: $\gamma=1$ the grid scaled \KS{} \pde{}~(\ref{Eks}) is effectively

233: restored over the whole domain by ensuring sufficient continuity

234: between adjacent elements.

235: 

236: The application of centre manifold theory is based upon a linear

237: picture of the dynamics.  Adjoin the dynamically trivial equations

238: \begin{equation}

239:     \D \tau \gamma=\D \tau h=0\,,

240:     \label{Etriv}

241: \end{equation}

242: and consider the dynamics in the extended state space

243: $(u(\xi),\gamma,h)$. This is a standard trick used to unfold

244: bifurcations \cite[\S1.5]{Carr81} or to justify long-wave

245: approximations \cite{Roberts88a}.  Within each element

246: $u=\gamma=0$ is a fixed point. Linearized about each fixed point,

247: that is to an error $\Ord{\|u\|^2+\gamma^2+h^2}$, the \pde{} is

248: \begin{displaymath}\quad

249:     \frac{\partial u}{\partial \tau}=\frac{\partial^4 u}{\partial \xi^4}\,,

250:     \quad\mbox{s.t.}\quad

251:     \left.\D \xi u\right|_{\xi=\pm 1/2}=

252:     \left.\DDD \xi u\right|_{\xi=\pm 1/2}=0\,,

253: \end{displaymath}

254: namely the hyperdiffusion equation with essentially insulating boundary

255: conditions.  There are thus linear eigenmodes associated with each

256: element:

257: \begin{equation}\quad

258: \gamma=0\,,\quad

259:     u\propto \left\{

260:     \begin{array}{ll}

261:         e^{\lambda_n \tau}\cos[n\pi(\xi-\xjmh)]\,, & \xjmh<\xi<\xjph\,,  \\

262:         0\,, & \mbox{otherwise}\,,

263:     \end{array}\right.

264:     \label{Emode}

265: \end{equation}

266: for $n=0,1,\ldots$, where the decay rate of each mode is  $

267: \lambda_n=-{n^4\pi^4}\,;$

268: %\begin{equation}

269: %    \lambda_n=-{n^4\pi^4}\,;

270: %    \label{Eeigen}

271: %\end{equation}

272: together with the trivial modes $\gamma=\mbox{const}$,

273: $h=\mbox{const}$ and $u=0$.  In a domain with $m$ elements,

274: evidentally all eigenvalues are negative, $-\pi^4$ or less, except

275: for $m+2$ zero eigenvalues: $1$ associated with each of the $m$

276: elements and $2$ from the trivial~(\ref{Etriv}).  Thus, provided

277: the nonlinear terms in~(\ref{Eks}) are sufficiently well behaved,

278: the existence theorem (\cite[p281]{Carr83b}

279: or~\cite[p96]{Vanderbauwhede89}) guarantees that a

280: $m+2$~dimensional centre manifold $\cM$ exists

281: for~(\ref{Eks}--\ref{Etriv}).  The centre manifold $\cM$ is

282: parameterized by $\gamma$, $h$ and a measure of $u$ in each

283: element, say $\uj$: using $\vec u$ to denote the collection of

284: such parameters, $\cM$ is written as

285: \begin{equation}

286:     u(\xi,\tau)=v(\xi;\vec u,\gamma,h)\,.

287:     \label{Ecmv}

288: \end{equation}

289: In this the analysis has a very similar appearance to that of finite

290: elements.  The theorem also asserts that on the centre manifold the

291: parameters $\uj$ evolve deterministically

292: \begin{equation}

293:     \dot\uj=g_j(\vec u,\gamma,h)\,,

294:     \label{Ecmg}

295: \end{equation}

296: where $\dot\uj$ denotes $d\uj/d\tau$, and $g_j$ is the restriction

297: of~(\ref{Eks}--\ref{Etriv}) to $\cM$.  In this approach the

298: parameters of the description of the centre manifold may be

299: anything that sensibly measures the size of $u$ in each

300: element---we simply choose the value of $u$ at the grid points,

301: $\uj(\tau)=u(\xi_j,\tau)$. This provides the necessary amplitude

302: conditions, namely that $\uj=v(\xi_j;\vec u,\gamma,h)$.\par

303: %\begin{equation}

304: %    \uj=v|_{\xi_j}\,.

305: %    \label{Eamp}

306: %\end{equation}

307: The above application of the theorem establishes that in principle we

308: may find the dynamics~(\ref{Ecmg}) of the interacting elements of the

309: discretisation.  A low order approximation written in unscaled

310: variables is given in~(\ref{Efd}).

311: 

312: The next outstanding question to answer is: how can we be sure

313: that such a description of the interacting elements does actually

314: \emph{model} the dynamics of the original

315: system~(\ref{Eks}--\ref{Etriv})?   Here, the relevance theorem of

316: centre manifolds, \cite[p282]{Carr83b}

317: or~\cite[p128]{Vanderbauwhede89}, guarantees that all solutions

318: of~(\ref{Eks}--\ref{Etriv}) which remain in the neighbourhood of

319: the origin in $(u(\xi),\gamma,h)$ space are exponentially quickly

320: attracted to a solution of the $m$ finite difference

321: equations~(\ref{Ecmg}).  For practical purposes the rate of

322: attraction is estimated by the leading negative eigenvalue, here

323: $-\pi^4$.  Centre manifold theory also guarantees that the

324: stability near the origin is the same in both the model and the

325: original.  Thus the finite difference model will be stable if the

326: original dynamics are stable.  After exponentially quick

327: transients have died out, the finite difference

328: equation~(\ref{Ecmg}) on the centre manifold accurately models the

329: complete system~(\ref{Eks}--\ref{Etriv}).

330: 

331: The last piece of theoretical support tells us how to approximate

332: the shape of the centre manifold and the evolution thereon.

333: Approximation theorems such as that by Carr \& Muncaster

334: \cite[p283]{Carr83b} assure us that upon substituting the

335: ansatz~(\ref{Ecmv}--\ref{Ecmg}) into the

336: original~(\ref{Eks}--\ref{Etriv}) and solving to some order of

337: error in $\|u\|$, $\gamma$ and $h$, then $\cM$ and the evolution

338: thereon will be approximated to the same order.  The catch with

339: this application is that we need to evaluate the approximations at

340: $\gamma=1$ because it is only then that the artificial internal

341: boundaries are removed.  In some applications of such an

342: artificial homotopy good convergence in the parameter $\gamma$

343: \cite{Roberts94c} has been found.  Thus although the order of

344: error estimates do provide assurance, the actual error due to the

345: evaluation at $\gamma=1$ should be also assessed otherwise.  Here

346: we have crafted the interaction~(\ref{Esbci}) between elements so

347: that low order terms in $\gamma$ recover the exact finite

348: difference formula for linear terms. Note that although centre

349: manifold theory ``guarantees'' useful properties near the origin

350: in $(u(\xi),\gamma,h)$ space, because of the need to evaluate

351: asymptotic expressions at $\gamma=1$, we have used a weaker term

352: elsewhere, namely ``assures''.

353: 

354: 

355: 

356: \section{Numerical comparisons show the effectiveness}

357: \label{Sbe}

358: 

359: We now turn to a detailed description of the centre manifold model for

360: the \KS{} equation~(\ref{Epde}).

361: 

362: The algebraic details of the derivation of the centre manifold

363: model~(\ref{Ecmv}--\ref{Ecmg}) are handled by computer

364: algebra.\footnote{The \reduce{} computer algebra source code is

365: available from the authors upon request.}  In an algorithm

366: introduced in~\cite{Roberts96a}, the program iterates to drive to

367: zero the residuals of the governing differential

368: equation~(\ref{Eks}) and its boundary

369: conditions~(\ref{Ebcsd}--\ref{Efudge}).  Hence by the

370: Approximation theorems we assuradly construct appropriate

371: approximations to the centre manifold model.

372: 

373: %%\begin{figure}[tbp]

374: %%    \centering

375: %%?? % \includegraphics[width=0.9\textwidth]{sinelem.eps}

376: %%\caption{\narrower low order

377: %%approximations~(\ref{Ecm23}) to the centre manifold field $u(x)$ within

378: %%each element for a fixed set of data points (circles) corresponding to

379: %%$m=8$ intervals on $[0,2\pi)$ with parameters $\alpha=6$, $\beta=0$ and

380: %%$\gamma=1$: $\cdots$, the normal field associated with finite

381: %%differences; $-\cdot-\cdot-$, includes

382: %%the first effects of the nonlinear advection term, $\alpha uu_x$;

383: %%$-~-~-$, includes the second-order effects in $\alpha$.  }

384: %%    \label{Fsine}

385: %%\end{figure}

386: %%The shape of the centre manifold gives the field $u$ as a function of

387: %%the parameters $\uj$ and the coupling $\gamma$.  To low-order in $u$

388: %%and $\gamma$, the solution field in the $j$th element is

389: %%\begin{eqnarray} ??

390: %%    u & = & \uj

391: %    +\gamma\left[ \frac{\up-\um}{2}\xi

392: %    +\frac{\up-2\uj+\um}{2}\xi^2 \right]

393: %    \nonumber\\&&{}

394: %    +h\gamma\left[\alpha \uj(\up-2\uj+\um)\left(

395: %    \frac{1}{6}\xi^3-\frac{1}{8}\xi \right)\right]

396: %    \nonumber\\&&{}

397: %    +h^2\gamma\left[ \alpha^2 \uj^2(\up-2\uj+\um)\left(

398: %    \frac{1}{24}\xi^4-\frac{1}{48}\xi^2 \right)

399: %    \right.\nonumber\\&&\quad\left.{}

400: %    +\beta (\up^3-3\up\uj^2+4\uj^3-3\um\uj^2+\um^3)

401: %    \left(-\frac{1}{24}\xi^4-\frac{1}{48}\xi^2\right)

402: %    \right.\nonumber\\&&\quad\left.{}

403: %    +\beta (\up^3-3\up\uj^2+3\um\uj^2-\um^3)

404: %    \left(-\frac{1}{12}\xi^3+\frac{1}{48}\xi\right)

405: %    \right.\nonumber\\&&\quad\left.{}

406: %    -\beta\frac{\uj^2}{8}\left( (\up-\um)\xi+(\up-2\uj+\um)\xi^2 \right)

407: %    \right]

408: %    \nonumber\\&&{}

409: %    +\Ord{\|u\|^4,\gamma^2}\,.

410: %    \label{Ecm23}

411: %\end{eqnarray}

412: %\marginal{?? revise this accordingly---although it is expendable}

413: %Observe the first line, when evaluated at $\gamma=1$, is simply the

414: %quadratic interpolant based upon second order accurate centred

415: %differences.  This is normal finite differences.  As displayed in

416: %Figure~\ref{Fsine}, the second line shows that this field should be

417: %modified because of the nonlinear advection term $\alpha uu_x$; the

418: %modification in proportion to the second difference at $x_j$ is

419: %reasonable because when $u$ has a local maximum the field must be

420: %decreased/increased to the left/right due to the advection to the

421: %right of the lower/higher levels of $u$ respectively.  Note that the

422: %same discretisation for this nonlinear term is obtained here whether

423: %we code it as $\alpha uu_x$ or $\alpha(u^2)_x/2$---the centre manifold

424: %is independent of any valid change in the algebraic description of the

425: %dynamics.  The third line gives the next order correction due to the

426: %nonlinear advection and is also displayed in Figure~\ref{Fsine}.  It might

427: %appear odd to introduce the dysfunctional behaviour between elements

428: %shown in Figure~\ref{Fsine}, but centre manifold theory reasonably

429: %assures us that it is appropriate for the nonlinear dynamics we wish

430: %to model.  In short, the description of the centre manifold is based

431: %upon standard differencing formulae, but includes effects upon $u$

432: %within each element due to the nonlinear processes that act in the

433: %continuum dynamics.

434: 

435: The finite difference model is given by the evolution on the

436: centre manifold.  In order to represent the spatial fourth

437: derivative in the \KS{} equation we need to determine the

438: interactions between next-nearest neighbouring elements.  Thus the

439: first approximation we can consider involves quadratic terms in

440: $\gamma$.  After returning to $x$ and $t$ variables it is

441: \begin{eqnarray}

442:     \frac{d \uj}{d t} &=&

443:     -\frac{\gamma R}{h^2}\left(\up-2\uj+\um\right)

444:     -\frac{\gamma}{2h}\uj(\up-\um)\nonumber\\&&{}

445:     -\frac{\gamma^2}{h^4}\left(\upp-4\up+6\uj-4\um+\umm\right)\nonumber \\&&{}

446:     +\frac{\gamma^2 R}{12h^2}\left(\upp-4\up+6\uj-4\um+\umm\right)

447:     \nonumber \\&&{}

448:     +\frac{\gamma^2}{48h}\left(\upp\up+3\upp\uj-3\up-3\up\uj \right.

449:     \nonumber \\&&\quad{}\left.

450:     +3\um\uj+3\um-3\umm\uj-\umm\um\right)

451:     \nonumber \\&&{}

452:     +\frac{\gamma h^2 u_j^2}{120}\left(\up-2\uj+\um\right)\nonumber\\&&{}

453:     +\frac{\gamma^2

454:     h^2}{60480}\left[\uj^2\left(-30\upp-170\up+256\uj-170\um-30\umm\right)\right.

455:     \nonumber \\&& \quad{}

456:     +\uj\left(-126\upp\up-54\up\um-126\umm\um\right)

457:     \nonumber \\&& \quad{}

458:     +\up^2\left(10\upp-20\up+235\uj\right)

459:     \nonumber \\&& \quad{}

460:     \left.+\um^2\left(10\um-20\um+235\uj\right)\right]

461:     \nonumber \\&&{}

462:     % ??

463:     %\dot\uj & = &

464:     %\frac{\gamma}{h^2}\left( \up-2\uj+\um \right)

465:     %-\frac{\gamma\alpha}{2h}\uj(\up-\um)

466:     %-\beta\uj^3

467:     %\nonumber\\&&{}

468:     %+\gamma\left[ \frac{\alpha^2}{12}\uj^2(\up-2\uj+\um)

469:     %\right.\nonumber\\&&\left.\quad{}

470:     %+\frac{\beta}{24}\left(-\up^3-3\up\uj^2+8\uj^3-3\um\uj^2-\um^3\right) \right]

471: %\nonumber\\&&{}

472: %    +h\gamma\left[ \frac{\alpha\beta}{8}\uj^3

473: %    (\up-\um) \right]

474: %    \nonumber\\&&{}

475: %    +h^2\gamma\left[

476: %    -\frac{\alpha^4}{720}\uj^4(\up-2\uj+\um)

477: %    \right.\nonumber\\&&\left.\quad{}

478: %    +\frac{\alpha^2\beta}{5760}\uj^2\left( 9\up^3 -113\up\uj^2 +208\uj^3

479: %    -113\um\uj^2 +9\um^3 \right)

480: %    \right.\nonumber\\&&\left.\quad{}

481: %    +\frac{\beta^2}{1920}\left( -21\up^5 +6\up^3\uj^2 +7\up\uj^4 +16\uj^5

482: %    \right.\right.\nonumber\\&&\left.\left.\qquad{}

483: %    +7\um\uj^4 +6\um^3\uj^2 -21\um^5 \right)

484: %    \phantom{\frac11}\right]

485: %    \nonumber\\&&{}

486:     +\Ord{\|u\|^4,\gamma^3,h^4}\,.

487:      \label{Ecmg24}

488: \end{eqnarray}

489: %\marginal{?? revise here as well}

490: The first two lines recorded here, when evaluated for $\gamma=1$,

491: form the conventional second-order finite difference equation for

492: the \KS{} equation~(\ref{Epde}). The third line when

493: evaluated for $\gamma=1$ gives the fourth order accurate

494: corrections to the $Ru_{xx}$ term. The fourth and further lines

495: above start accounting systematically for the variations in the

496: field $u$ within each element and how they affect the evolution

497: through the nonlinear term.

498: %The approximation formed by the first three lines is that reported

499: %in the Introduction as~(\ref{Efd}).  The fourth and further lines

500: %above, through $\alpha\beta$ and $\alpha^2\beta$ effects, show

501: %that this approach also accounts for interaction between the

502: %nonlinear terms in the \pde{} in the finite difference

503: %approximation.

504: Finite difference equations derived via this approach holistically

505: model all the interacting dynamics of the entire \pde.

506: 

507: \begin{figure}[tbp]

508:     \centering

509:     \includegraphics[width=0.9\textwidth]{a1r2con8pt3.eps}

510:     \caption{Contours

511:     of an accurate solution $u(x,t)$, -----, to compare with numerical

512:     approximations~(\ref{Ecmg24}): $\cdots$, the conventional

513:     approximation, errors ${\cal O}\left(h^2\right)$; $-~-~-$, the first

514:     correction, errors ${\cal O}\left(\|u\|^3\right)$; $-\cdot-\cdot-$

515:     , the second correction,errors ${\cal O}\left(\|u\|^4\right)$.

516:     \KS{} equation~(\ref{Epde}) with

517:     parameter $R=2$ is discretised on just $m=8$

518:     elements in $[0,2\pi)$ and drawn with contour interval $\Delta

519:     u=3$.}

520:     \label{F286}

521: \end{figure}

522: To show the effectiveness of the approach we compare the finite

523: difference model obtained from various truncations

524: of~(\ref{Ecmg24}) to accurate numerical solutions obtained on a

525: much finer grid. Choosing $m$ intervals on $[0,2\pi)$ gives an

526: element length $h=2\pi/m$ and grid points $x_j=jh$ for

527: $j=0,\ldots,m-1$. Because of the antisymmetry in $u(x,t)$ about

528: $x=k\pi$, when starting from the initial condition $u(x,0)=10\sin x$,

529: we only display the interval $[0,\pi]$. There are three different

530: approximations from~(\ref{Ecmg24}), with $\gamma=1$, depending

531: upon where the expansion is truncated.

532: % (or equivalently in $\alpha$):

533: %\begin{eqnarray}

534: %    \dot\uj & \approx &

535: %    \frac{1}{h^2}\left( \up-2\uj+\um \right)

536: %    -\frac{\alpha}{2h}\uj(\up-\um)

537: %    \nonumber\\&&{}

538: %    +\frac{\alpha^2}{12}\uj^2(\up-2\uj+\um)

539: %    \nonumber\\&&{}

540: %    -h^2\frac{\alpha^4}{720}\uj^4(\up-2\uj+\um)\,.

541: %    \label{Eburg2}

542: %\end{eqnarray}

543: The first two lines form a model with ${\cal O}\left(h^2\right)$

544: errors (a conventional finite difference approximation), the first

545: five lines provide the first correction with ${\cal O}

546: \left(\|u\|^3\right)$ errors, and all shown terms form the model

547: with ${\cal O}\left(\|u\|^4\right)$ errors. The solutions of these

548: models over $0<t<1$ with $m=8$ and $R=2$ are shown in

549: Figure~\ref{F286}. Observe that the leading approximation (dotted)

550: is significantly in error whereas the next two refinements

551: (dot-dashed and dashed) are overall more accurate, especially near

552: the peak. Such accuracy is remarkable considering the

553: nonlinearity, and the few points in the discretisation, $m=8$.

554: 

555: % \begin{table}[tbp]

556: %   \centering

557: % \caption{Comparison of the errors $\varepsilon$, defined

558: % in~(\ref{Eerr}), of the spatially second order model~(\ref{Eburg2}) at

559: % the two first truncations in the nonlinear parameter $\alpha$ for

560: % various grid subintervals $m$.  Observe that the higher order

561: % model is significantly less sensitive to the nonlinearity.}

562: % \begin{tabular}{|r|rr|rr|rr|}

563: %   \hline

564: %    & \multicolumn{2}{c|}{$m=8$}  & \multicolumn{2}{c|}{$m=16$}  &

565: %    \multicolumn{2}{c|}{$m=32$}   \\

566: %   $\alpha$ & $o(\alpha)$ & $o(\alpha^2)$ & $o(\alpha)$ & $o(\alpha^2)$

567: %   & $o(\alpha)$ & $o(\alpha^2)$ \\

568: %   \hline\hline

569: %   1 & .0116 & .0095 & .0031 & .0025 & .0008 & .0006  \\

570: %   \hline

571: %   3 & .0356 & .0109 & .0102 & .0040 & .0026 & .0011  \\

572: %   \hline

573: %   6 & .0694 & .0115 & .0215 & .0059 & .0054 & .0018  \\

574: %   \hline

575: %   10 & .0971 & .0186 & .0318 & .0055 & .0081 & .0026  \\

576: %   \hline

577: % \end{tabular}

578: %   \label{Ta2}

579: % \end{table}

580: % For an assessment of errors over a range of parameters we record in

581: % Table~\ref{Ta2} the error

582: % \begin{equation}

583: %   \varepsilon=\max_t\left(\frac{1}{m}\sum_j|\epsilon_j(t)|\right)\,,

584: %   \label{Eerr}

585: % \end{equation}

586: % where $\epsilon_j(t)$ is the difference between an approximation

587: % and the exact solution at the $j$th grid point.  Observe the usual

588: % properties that the errors decrease with increasing spatial

589: % resolution $m$, and that the errors increase with increasing

590: % nonlinearity $\alpha$.  Our interest lies more in the comparison

591: % between the $o(\alpha)$ and $o(\alpha^2)$ models (here the

592: % $o(\alpha^4)$ model is virtually indistinguishable from the

593: % $o(\alpha^2)$ model).  For small nonlinearity, $\alpha=1$, there is

594: % very little difference; presumably the errors are dominated by the

595: % second-order approximation of spatial derivatives.  As the

596: % nonlinearity $\alpha$ increases, the error of the $o(\alpha)$

597: % approximation increases roughly in proportion to $\alpha$, but the

598: % error of the $o(\alpha^2)$ approximation grows much more slowly.

599: % We conclude, as centre manifold theory assures us, that the

600: % nonlinearly corrected $o(\alpha^2)$ model more accurately captures

601: % the nonlinear dynamics of \KS{} equation.

602: 

603: 

604: 

605: 

606: \section{Conclusion}

607: 

608: Centre manifold theory is a powerful new approach to deriving finite

609: difference models of dynamical systems.  Many details need to

610: researched for a general application of the theory.  However, there

611: are many promising features of this application to the \KS{}

612: equation~(\ref{Epde}) and the earlier example of Burgers' equation

613: \cite{Roberts98a}.

614: 

615: \bibliographystyle{plain}

616: \bibliography{more2}

617: %,bib,new}

618: 

619: 

620: \end{document}

621: