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

2: 

3: \usepackage{url}

4: \usepackage{defns}

5: % \usepackage[colour]{ajr}

6: % \newcommand{\cA}{{\cal A}}

7: \newcommand{\cJ}{{\cal J}}

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

9: \newcommand{\uj}{u_j}

10: \newcommand{\sfrac}[2]{{\textstyle\frac{#1}{#2}}}

11: \newcommand{\ibc}{\textsc{ibc}}

12: \newcommand{\pde}{\textsc{pde}}

13: 

14: % \renewcommand{\baselinestretch}{1.5}

15: 

16: \begin{document}

17: 

18: % \title{Holistic finite differences rationalise projecting 

19: % initial conditions onto a numerical discretisation}

20: \title{Holistic projection of initial conditions onto a finite 

21: difference approximation}

22: 

23: \author{A. J. Roberts 

24: \thanks{\protect\url{mailto:aroberts@usq.edu.au}}\\

25: \textit{Dept Maths \& Comput., University of Southern 

26: Queensland,}\\ \textit{Toowoomba, Queensland~4352, Australia}.

27: }

28: \maketitle

29: 

30: \begin{abstract}

31: Modern dynamical systems theory has previously had little to say about 

32: finite difference and finite element approximations of partial 

33: differential equations (\pde{}s) \cite{Archilla98}.

34: However, recently I have shown one way that centre manifold theory may 

35: be used to create and support the spatial discretisation of \pde{}s 

36: such as Burgers' equation \cite{Roberts98a} and the 

37: Kuramoto-Sivashinsky equation \cite{MacKenzie00a}.

38: In this paper the geometric view of a centre manifold is used to 

39: provide correct initial conditions for numerical discretisations 

40: \cite{Roberts97b}.

41: The derived projection of initial conditions follows from the physical 

42: processes expressed in the \pde{}s and so is appropriately 

43: conservative.

44: This rational approach increases the accuracy of forecasts made with 

45: finite difference models.

46: \end{abstract}

47: 

48: \paragraph{PACS:} 02.60.Lj, 02.70.Bf, 05.45.-a

49: 

50: \paragraph{Keywords:} Burgers' equation, initial condition, holistic finite 

51: differences.

52: 

53: 

54: \section{Introduction}

55: 

56: Consider the equations for some physical field $u(x,t)$ evolving in 

57: space-time that we wish to model numerically.

58: Imagine a given initial field $u_0(x)$ and a finite difference model 

59: written in terms of $\uj(t)=u(x_j,t)$ for equi-spaced grid points 

60: $x_j=jh$ say; for example, in \S\ref{Scmth} for Burgers' 

61: equation~(\ref{eq:burg}) we find

62: \begin{equation}

63:      \frac{d\uj}{dt}+\frac{a}{2h}\mu\delta\, \uj^2\approx

64:     \frac{1}{h^2}\delta^2\uj 

65:     +\frac{a^2}{16}(\delta^2\uj^3-\uj^2\delta^2\uj) \,,

66:     \label{eq:dmod}

67: \end{equation}

68: in terms of the central difference operator $\delta\,\uj 

69: =u_{j+1/2}-u_{j-1/2}$ and central mean operator $\mu\,\uj 

70: =(u_{j+1/2}+u_{j-1/2})/2$\,.

71: One might expect that the correct initial condition for this 

72: discretisation is simply to project the initial field $u_0(x)$ onto 

73: the finite dimensional space of the model by setting the initial 

74: discretisation values to the value of the initial field at the grid: 

75: $\uj(0)=u_0(x_j)$.

76: But if the initial field is localised away from any grid point then 

77: physically we know to distribute the initial field among nearby grid 

78: points.

79: I use dynamical arguments to show that the correct initial condition 

80: is, to leading order, the correctly conservative element average

81: \begin{equation}

82:     \uj(0)\approx\frac{1}{h}\int_{x_j-h/2}^{x_j+h/2} u_0(x) \,dx\,.

83:     \label{eq:leadzj}

84: \end{equation}

85: This formula, and higher order corrections that involve neighbouring 

86: elements, are derived systematically herein.

87: For a numerical model, this is the first time a dynamical rationale 

88: has been used to provide initial conditions.

89: 

90: Such projection of initial fields onto the discretisation is supported 

91: by centre manifold theory \cite[e.g.]{Carr81}: the Relevance Theorem 

92: asserts that each of the nearby solutions of the governing \pde{} 

93: exponentially quickly in time approach a solution of the numerical 

94: model; this holds even for finite grid spacing~$h$.

95: The algebraic techniques developed by Roberts~\cite{Roberts97b}, based 

96: upon analysing with the aid of computer algebra the adjoint of a 

97: linearisation of the \pde{}, determines the initial condition for the 

98: discretisation so that we ensure the finite difference model 

99: faithfully tracks the correct particular solution of the \pde.

100: 

101: 

102: \section{Burgers' equation is discretised with centre manifold theory}

103: \label{Scmth}

104: 

105: Consider the dynamics of Burgers' equation

106: \begin{equation}

107:     u_t+auu_x=u_{xx}

108:     \label{eq:burg}

109: \end{equation}

110: as a prototype advection-diffusion equation.

111: Roberts~\cite{Roberts98a} first constructed finite difference 

112: approximations to the spatial derivatives using centre manifold theory 

113: to ensure nonlinear, subgrid-scale processes were systematically 

114: modelled.

115: We summarise the approach in this section.

116: 

117: Divide the spatial domain~$I$ into a number, say $m$, of 

118: elements of equi-size~$h$.

119: We analyse the dynamics of the elements away from any physical 

120: boundary to derive a discretisation for the interior of the domain.

121: Artificially crafted internal boundary conditions (\ibc's) between 

122: the elements are introduced:

123: \begin{equation}

124:     \left[u_x-\half au^2\right]=0\,,\quad

125:     (1-\gamma)h\overline{(u_x-\half au^2)}=\gamma\left[u\right]\,,

126:     \label{eq:ibc}

127: \end{equation}

128: where $[\phantom{u}]$ denote the jump across each internal boundary, 

129: $\overline{\phantom{u}}$ denotes the average value from the two sides 

130: of the boundary, and distinct from earlier work~\cite{Roberts98a} 

131: these \ibc's are expressed in terms of the flux $q=-u_x+\half au^2$\,.

132: See that when $\gamma=0$ the right-hand side of the second \ibc{} 

133: disappears so that the two conditions then completely insulate an 

134: element from its neighbours.

135: Whereas when $\gamma=1$, the left-hand side disappears and the two 

136: conditions ensure sufficient continuity of the physical field to 

137: recover Burgers' dynamics throughout the domain.

138: 

139: The centre manifold and the evolution thereon is straightforwardly 

140: constructed using the computer algebra algorithm described in 

141: \cite{Roberts98a,Roberts96a}.

142: Here we find the subgrid field in the $j$th~element is

143: \begin{eqnarray}

144:     && u(x,t) = \uj +\half ah\xi\uj^2 +\quarter a^2h^2\xi^2\uj^3

145:     +\gamma\left[\xi\mu\delta\,\uj 

146:     +\half\xi^2\delta^2\uj\right] \nonumber \\

147:      &  & {} +ah\gamma\left[ -{\textstyle\frac18}\xi( 

148:      \uj\delta^2\uj +\delta^2\uj^2 +4\uj^2) +{\textstyle\frac18}\xi^2( 

149:      2\uj\mu\delta\,\uj-\mu\delta\,\uj^2) 

150:      +\third\xi^2\uj\delta^2\uj  \right]  \nonumber \\

151:      &  & {} +a^2h^2\gamma\left[ 

152:      {\textstyle\frac{1}{16}}\xi( \uj\mu\delta\,\uj^2 +\mu\delta\uj^3) 

153:      -{\textstyle\frac{3}{32}}\xi^2( 3\uj^2\delta^2\uj +2\uj\delta^2\uj 

154:      -\delta^2\uj^3 +8\uj^3) 

155:      \right. \nonumber \\ &  & \left.\quad{}

156:      +{\textstyle\frac{1}{6}}\xi^3( 2\uj^2\mu\delta\,\uj 

157:      -\uj\mu\delta\,\uj^2) +{\textstyle\frac{5}{24}}\xi^4\uj^2\delta^2\uj 

158:      \right]  +\Ord{\gamma^2,a^3}\,,

159:      \label{eq:cm}

160: \end{eqnarray}

161: where $\xi=(x-x_j)/h$ ranges over $[-1/2,1/2]$\,.

162: The evolution on this centre manifold, when evaluated at $\gamma=1$ to 

163: restore continuity, forms the finite difference model~(\ref{eq:dmod}) 

164: for Burgers' equation: see that the first three terms 

165: in~(\ref{eq:dmod}) form a standard discretisation of each term but now 

166: appearing automatically from the discretisation when mediated by the 

167: flux form~(\ref{eq:ibc}) of the \ibc's; whereas the last term gives 

168: $\Ord{a^2}$ corrections to account for interactions between the 

169: nonlinear advection and the diffusive dissipation.

170: Such nonlinear modifications of standard discretisations can be 

171: extremely effective~\cite{Roberts98a}.

172: 

173: To find the correct initial condition, $\uj(0)$, for numerical models 

174: such as~(\ref{eq:dmod}) corresponding to any given field $u_0(x)$, we 

175: follow the procedure described in~\cite{Roberts97b}.

176: The aim is to determine projection vectors $z_j(x)$, such as those 

177: shown in Figure~\ref{fig:zj}, so that

178: \begin{equation}

179:     \left<z_j,u_0(x)-v(\vec u(0),x)\right>=0

180:     \quad\mbox{using}\quad

181:     \left<z,u\right>=\frac{1}{h}\int_I zu\,dx

182:     \label{eq:proj}

183: \end{equation}

184: as the inner product.

185: \begin{figure}[tbp]

186:     \centering

187:     \includegraphics[width=\textwidth]{zjplot.eps}

188:     \caption{leading two orders of approximation to the projection 

189:     vectors~$z_j(x)$ for purely diffusive dynamics, errors: 

190:     $\Ord{\gamma}$, dashed; $\Ord{\gamma^2}$,  solid.  }

191:     \label{fig:zj}

192: \end{figure}%

193: Now the dynamics linearised about the nonlinear centre manifold, 

194: $u=v(\vec u,x)$, is governed by the operator

195: \begin{displaymath}

196:     \cJ =\partial_x^2-av_x -av\partial_x\,,

197: \end{displaymath}

198: with  \ibc's linearised about~(\ref{eq:ibc}) of

199: \begin{equation}

200:     \left[u_x\right]=0\,,\quad

201:     (1-\gamma)h\overline{ (u_x- avu)}=\gamma\left[u\right]\,,

202:     \label{eq:libc}

203: \end{equation}

204: Then in the above inner product the adjoint of $\cJ$ is

205: \begin{equation}

206:     \cJ^{\dagger} z=\partial_x^2z+av\partial_xz\,,

207:     \quad\mbox{such that}\quad

208:          \left[z_x\right]=0\,,\quad

209:      (1-\gamma)h\overline{z_x}=\gamma\left[z\right]\,.

210:     \label{eq:adj}

211: \end{equation}

212: To find the projection vectors~$z_j(x)$ we start with the leading 

213: approximation $z_j(x)\approx\chi_j(x)$ corresponding 

214: to~(\ref{eq:leadzj}) and plotted in Figure~\ref{fig:zj}, where 

215: $\chi_j(x)$ denotes the characteristic function that is~$1$ in the 

216: $j$th~element and otherwise is~$0$.

217: Then successive corrections are sought by iteration to ultimately 

218: satisfy the appropriate version of the equations derived 

219: in~\cite{Roberts97b}: defining the dual operator $\cD z=\D 

220: tz+\cJ^{\dagger}z$ we must solve

221: \begin{equation}

222:     \cD z_j-\sum_i\left<\cD z_j,e_i\right>z_i=0\,,

223:     \label{eq:projdz}

224: \end{equation}

225: subject to the \ibc's in~(\ref{eq:adj}) and the normalisation 

226: condition

227: \begin{equation}

228:     \left<z_j,e_i\right>=\delta_{i,j}\,,

229:     \label{eq:norm}

230: \end{equation}

231: where $e_j=\partial v/\partial \uj$ is the tangent vector of the 

232: centre manifold.

233: We seek solutions in a power series in $\gamma$ to errors 

234: $\Ord{\gamma^{\ell}}$ corresponding to the finite difference 

235: approximation of stencil width $2\ell-1$.

236: A computer algebra program available from the author does all the 

237: necessary algebra.

238: 

239: 

240: 

241: \section{Project onto Burgers' discretisation}

242: 

243: In this section we solve to quantities with errors 

244: $\Ord{a^3,\gamma^2}$: the finite difference model for Burgers' 

245: equation is then~(\ref{eq:dmod}); and the corresponding centre 

246: manifold over the whole domain is given by~(\ref{eq:cm}).

247: Calculating to errors $\Ord{a^3,\gamma^2}$ the projection onto the 

248: numerical model must be orthogonal to

249: \begin{eqnarray}

250:     z_j&\!\approx\!&\left[1-\frac{h^2a^2}{16}\uj^2\right]\chi_j

251:     \nonumber\\&&{}

252:     +\gamma\left[ \left(\sfrac16-\xi^2\right)\chi_j

253:     +\left(-\sfrac1{12}+\half\xi+\half\xi^2\right)\chi_{j-1}

254:     +\left(-\sfrac1{12}-\half\xi+\half\xi^2\right)\chi_{j+1} \right]

255:     \nonumber\\&&{}

256:     +\frac{ha\gamma}{48}\left[ \left( -(12\xi-16\xi^3)\uj 

257:     +u_{j+1}-u_{j-1} \right)\chi_j

258:     \right.\nonumber\\&&\quad\left.{}

259:     +\left( +\uj-(3-6\xi+8\xi^3)u_{j-1} \right)\chi_{j-1}

260:     \right.\nonumber\\&&\quad\left.{}

261:     +\left( -\uj+(3+6\xi-8\xi^3)u_{j+1} \right)\chi_{j+1} \right]

262:     \nonumber\\&&{}

263:     +\frac{h^2a^2\gamma}{384}\left[ \left( 8(1+3\xi^2)\uj^2 

264:     +4\uj(u_{j+1}+u_{j-1}) +2(u_{j+1}^2+u_{j-1}^2) \right)\chi_j 

265:     \right.\nonumber\\&&\quad\left.{}

266:     +\left( 3\uj^2+4\uj u_{j-1}-(5+12\xi^2+16\xi^3)u_{j-1}^2 

267:     \right)\chi_{j-1}

268:     \right.\nonumber\\&&\quad\left.{}

269:     +\left( 3\uj^2+4\uj u_{j+1}-(5+12\xi^2-16\xi^3)u_{j+1}^2 

270:     \right)\chi_{j+1}

271:     \right]\,.

272:     \label{eq:zj1ad}

273: \end{eqnarray}

274: Higher order expressions may be straightforwardly computed by computer 

275: algebra.

276: I conclude by further interpreting the physical effects incorporated 

277: in the projection defined by the above~$z_j$.

278: 

279: \subsection{Linear diffusion}

280: 

281: Set $a=0$ in this subsection to analyse the linear diffusion equation 

282: $u_t=u_{xx}$\,.

283: Then the projection vector~(\ref{eq:zj1ad}), evaluated at $\gamma=1$ 

284: to recover the physically relevant case as plotted in 

285: Figure~\ref{fig:zj}, is

286: \begin{equation}

287:     z_j\approx\left(\sfrac76-\xi^2\right)\chi_j

288:     +\left(-\sfrac1{12}+\half\xi+\half\xi^2\right)\chi_{j-1}

289:     +\left(-\sfrac1{12}-\half\xi+\half\xi^2\right)\chi_{j+1}\,.

290:     \label{eq:zj1}

291: \end{equation}

292: To find the correct initial condition using this in~(\ref{eq:proj}) 

293: note that in these linear diffusion dynamics $\left<z_j,v(\vec 

294: u,x)\right>=\uj$ by the normalisation~(\ref{eq:norm}); thus here 

295: $\uj(0) =\left< z_j,u_0(x) \right>$\,.

296: For example, see that a point release in the $k$th element, 

297: $u_0(x)=\delta(x-x_k-h\eta)$, requires the slightly distributed 

298: initial condition

299: \begin{equation}

300:     h\uj(0)=\left(\sfrac{7}{6}-\eta^2\right)\delta_{k,j}

301:     +\left(-\sfrac{1}{12}-\half\eta+\half\eta^2\right)\delta_{k-1,j}

302:     +\left(-\sfrac{1}{12}+\half\eta+\half\eta^2\right)\delta_{k+1,j} 

303:     \,.

304:     \label{eq:diffptg}

305: \end{equation}

306: Such a specific initial condition corresponds via~(\ref{eq:cm}) to a 

307: field on the centre manifold as shown in Figure~\ref{fig:ifld} for the 

308: three cases

309: $\eta=0$, $1/4$ and~$1/2$.

310: \begin{figure}[tbp]

311:     \centering

312:     \includegraphics{v0plot.eps}

313:     \caption{initial fields $u=v(\vec u(0),x)$ corresponding to a 

314:     unit-mass point release at: $\xi=0$, dot-dash; $\xi=1/4$, dashed; 

315:     $\xi=1/2$, solid.}

316:     \label{fig:ifld}

317: \end{figure}%

318: 

319: See that these initial conditions ensure that the first moment of the 

320: numerical solution is correct for all time: in the numerical 

321: model~(\ref{eq:dmod}) the first moment is constant in time so it is 

322: enough to check that the first moment is correct in the initial 

323: conditions.

324: Define $\left<u\right>=\left<1,u\right>$\,, then for all time 

325: $\left<u\right>=1$ both in the model and in the exact solutions.

326: The first moment in the exact solution is its initial value 

327: $m_1=\left<(x-x_k)u_0(x)\right>=h\eta$\,; from~(\ref{eq:diffptg}) the 

328: first moment in the numerical model is the same

329: \begin{displaymath}

330:     m_1=\left<(x-x_k)v(\vec u(0),x)\right>=h\eta\,.

331: \end{displaymath}

332: However, the second moment $m_2=\left<(x-x_k)^2u\right>$ has 

333: $\Ord{h^2}$ errors: it evolves in time at the correct rate 

334: $dm_2/dt=2$, but the initial value is $h^2(\eta^2-1/6)$ instead of~$0$.

335: Determining the projection of initial conditions to higher orders in 

336: the coupling parameter~$\gamma$ obtains such higher order moments 

337: correctly.

338: Note that the rational approach adopted here does better than the 

339: usually chosen initial conditions which incur $\Ord{h}$ errors.

340: 

341: 

342: \subsection{Nonlinear dynamics}

343: 

344: Consider the $\Ord{a}$ terms from~(\ref{eq:zj1ad}) that 

345: modify~(\ref{eq:zj1}), namely

346: \begin{eqnarray*}

347: z_j'&=&{}

348:     +\frac{ha}{48}\left[ \left( (-12\xi+16\xi^3)\uj 

349:     +u_{j+1}-u_{j-1} \right)\chi_j

350:     \right.\nonumber\\&&\quad\left.{}

351:     +\left( +\uj+(-3+6\xi-8\xi^3)u_{j-1} \right)\chi_{j-1}

352:     \right.\nonumber\\&&\quad\left.{}

353:     +\left( -\uj+(+3+6\xi-8\xi^3)u_{j+1} \right)\chi_{j+1} \right]\,.

354: \end{eqnarray*}

355: Realise that the leading order effect of including these terms is to 

356: \emph{modify} the initial condition by $\left<z_j',u_0(x)\right>$\,.  

357: For example, if the initial field is approximately constant, 

358: $u_0(x)\approx U$, then 

359: \begin{displaymath}

360:     z_j'=\frac{Uha}{48}\left[ 

361:     (6\xi-8\xi^3)(\chi_{j+1}-2\chi_j+\chi_{j-1}) 

362:     +2(\chi_{j+1}-\chi_{j-1}) \right]\,;

363: \end{displaymath}

364: that the coefficients of the characteristic functions~$\chi_k$ sum to 

365: zero reflects that the the projection conserves the field~$u$.

366: More specifically, if $u_0(x)$ is~$U$ except for a symmetric bump in 

367: the $k$th element, then as well as the direct symmetric distribution 

368: identified for linear diffusion, the component in 

369: $\chi_{j+1}-\chi_{j-1}$ causes $u_{k-1}(0)$ to increase and 

370: $u_{k+1}(0)$ to decrease by an amount proportional to~$Ua$ reflecting 

371: that the self advection of the bump is not as great as that induced by 

372: assigning the mass of the bump solely to~$u_k(0)$.

373: Conversely, for an antisymmetric perturbation in the $k$th element, 

374: positive to the left of $x_k$, the component in 

375: $\chi_{j+1}-2\chi_j+\chi_{j-1}$ increases $u_k(0)$ and decreases 

376: $u_{k\pm 1}(0)$ in proportion to~$Ua$ to reflect the increased delay 

377: in~$u$ advecting out of the $k$th element because more of it is 

378: further to the left initially.

379: The $\Ord{a^2}$ terms in~(\ref{eq:zj1ad}) reflect more subtle physical 

380: processes.

381: 

382: 

383: \section{Conclusion}

384: 

385: Based upon the method of analysis and the discussion in the previous 

386: sections, we deduce that this centre manifold approach to finding 

387: correct initial conditions for finite difference models accounts for 

388: subgrid scale processes that occur as initial transients decay.

389: No other method does this.

390: 

391: Extensions of this approach to higher spatial dimensions is 

392: straightforward.

393: For example, consider the class of diffusive \pde's

394: \begin{displaymath}

395: 	 \D tu=\delsq u+f(u,\grad u)\,,

396: \end{displaymath}

397: where $f$ represents nonlinear reaction or advection effects.

398: After tessellating space into finite elements---using \ibc's of the 

399: form (cf (\ref{eq:ibc}))

400: \begin{displaymath}

401:     \left[q_n\right]=0

402:     \quad\mbox{and}\quad

403:     (1-\gamma)h\overline{q_n}=\gamma[u]

404: \end{displaymath}

405: where $q_n$ is the flux of~$u$ normal to the internal boundary and $h$ 

406: is a size of the element---the fundamental problem in constructing a 

407: model is simply to solve Poisson's equation with forced Neumann 

408: boundary conditions on each element.

409: The adjoint of this problem lies at the heart of the 

410: dual~(\ref{eq:projdz}) for determining initial conditions of the 

411: approximation.

412: Although these sub-grid problem may itself need to be done 

413: numerically, in the simplest case of a regular tessellation it need 

414: only be done once for each term in the model, just like the 

415: computation of the interaction terms in a traditional finite element 

416: approximation.

417: 

418: In the case where there are variations in the size or shape of the 

419: elements of the discretisation, one would build formulae for the 

420: approximation parametrised by the shapes of the element and those 

421: neighbouring elements to which it is coupled by the \ibc's.

422: The algebraic detail becomes more complicated but the principles are 

423: the same.

424: 

425: 

426: \small

427: \bibliography{ajr,new,bib}

428: \bibliographystyle{unsrt}

429: 

430: \end{document}

431: