Vous êtes sur la page 1sur 42

Contents

1 Introduction 3
1.1 Particle and free-Lagrange methods : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 3
1.1.1 Particle methods : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 3
1.1.2 Free-Lagrange methods : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 4
1.1.3 Moving-grid methods : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 4
1.2 The semi-Lagrangian method : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 4
1.2.1 Trajectory tracing : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 5
1.2.2 Interpolation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 5
1.2.3 Other interpolations : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6
A spectral Lagrange-Galerkin method for convection-dominated di usion 1.3 Finite element methods : : : : : : : : : : : : : : : : :
1.3.1 The characteristic Galerkin method : : : : : : :
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
: 6
: 6
problems 1.3.2 Lagrange-Galerkin methods : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 7
1.3.3 The weak Lagrange-Galerkin method : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 8
1.4 Spectral methods : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 8
1.4.1 Spectral approximation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 8
A. F. Ware 1.4.2 The discrete Fourier transform : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 8
1.4.3 Spectral methods for time-dependent problems : : : : : : : : : : : : : : : : : : : : : : : : : : : 9
October 1991 1.5 Summary of contents : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 9
1.6 Notations : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 10
2 A model hyperbolic problem 13
2.1 Discussion of the model problem; the evolution operator E : : : : : : : : : : : : : : : : : : : : : : : : : 13
2.2 Construction of the numerical method : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 20
2.2.1 Stability : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 20
2.2.2 Convergence : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 21
2.3 The introduction of numerical quadrature : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 22
2.3.1 Accuracy : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 23
2.3.2 Stability analysis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 24
2.3.3 Convergence : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 28
3 An ecient implementation 30
3.1 Local interpolation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 30
3.2 The algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 32
3.3 Error analysis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 32
3.4 Numerical experiments : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 35
3.4.1 Experiment 1 : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 36
3.4.2 Experiment 2 : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 36
3.4.3 Experiment 3 : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 37
3.4.4 Experiment 4 : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 40
4 Advection{di usion equations 44
4.1 Introduction : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 44
4.1.1 Mathematical setting of the problem : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 44
4.1.2 The material derivative : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 45
4.1.3 First-order timestepping|an example : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 45
4.2 A Lagrangian form of the equation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 47
4.2.1 Further properties of E (t; s) : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 47
4.2.2 The operator A(t; s) : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 48

1
4.3 Time discretisation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 50
4.3.1 Stability : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 52
4.3.2 Convergence : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 54
4.4 Space discretisation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 55
4.4.1 Stability : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 55
4.4.2 Convergence : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 56
4.4.3 The initialisation procedure :
4.4.4 An example revisited : : : :
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
57
57
Chapter 1
4.5 A nonlinear example : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 59
4.5.1 A modi ed Newton method :
4.5.2 Numerical results : : : : : :
5 The Navier-Stokes equations
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
60
61
62
Introduction
5.1 Introduction : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 62
5.1.1 Notation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 62
5.2 The equations and their approximation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 63 The modelling of pure convection or convection-dominated processes is a central problem in elds such as meteorology, oil
5.2.1 Statement of the problem : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 63 reservoir simulation, modelling of aerodynamic or geophysical ows, magnetohydrodynamics and many others. For many
5.2.2 Lagrangian form : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 63 standard numerical methods there is a trade-o between excessive numerical di usion on the one hand, and unphysical
5.2.3 Spectral Lagrange-Galerkin approximation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 64 oscillations on the scale of the mesh on the other. However, at least as far back as the work of Courant, Isaacson and
5.3 Error analysis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 64 Rees [17] and Ansorge [3], it was realised that exploitation of the characteristics could give rise to e ective numerical
5.3.1 L2 error estimates : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 66 schemes for the solution of hyperbolic problems. When other terms, such as di usion, are present in small amounts,
5.3.2 Error estimates in V : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 69 the equation may be regarded as `nearly hyperbolic', and various options for the incorporation of the extra terms into
5.4 Stability : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 70 a scheme designed for the purely hyperbolic case are possible, giving rise to methods that are capable of avoiding the
5.4.1 De nition of stability : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 70 undesirable trade-o mentioned above.
5.4.2 Reformulation of the problem : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 71 A closely related idea is the use of Lagrangian coordinates, which may be used to improve the treatment of convective
5.4.3 Stability analysis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 71 terms in uid dynamical applications. As opposed to Eulerian methods, where an observer associated with a grid point
5.5 Numerical experiments : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 72 watches the world evolve around him from a xed vantage point, in a scheme based on Lagrangian coordinates a similar
5.5.1 The algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 72 observer nds himself travelling with a uid particle. From this perspective the time evolution of the ow may well be
5.5.2 The experiments : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 73 smoother, and thus easier to model numerically.
5.6 Conclusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 74 Such concepts have formed the basis of a wide variety of numerical schemes since their initial application, and are the
5.6.1 Future developments : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 75 central theme of this Thesis. We shall begin by giving a brief review of a selection of the range of methods that have been
devised for convection-dominated problems, giving particular emphasis to those based on the characteristics/Lagrangian
ideas discussed above.

1.1 Particle and free-Lagrange methods


The common feature of the methods to be described in this section is the presence of a set of points or particles that
are transported under the action of some convective eld. Consider a time-dependent convection-di usion equation
ut + r  f (u) = r  (bru) in Rn  (0; T ]; (1.1a)
u(; 0) = u0 ; (1.1b)
where r  f (u) = a  ru; a, b are functions of x and t, and a may depend on u. Let x(t) satisfy the equation
dx(t) = a(x(t); t): (1:2)
dt
Then, writing u = u(x(t); t), (1.1a) may be rewritten in terms of coordinates satisying (1.2), giving
du = r  (bru); (1:3)
dt
which is in the form of a heat equation.

1.1.1 Particle methods


Particle methods (see [63]) model the solution of (1.1) by calculating the evolution of a set of particles whose coordinates
satisfy (1.2). In the absence of di usion (b = 0) (1.1) thus reduces to the problem of solving the set of ordinary di erential
equations (1.2). In the uid dynamics arena, particle methods were rst applied to the solution of Euler's equations in
vorticity-streamfunction form. Based on the principle that, for non-viscous ows, the vorticity in the eld is transported
by the ow, the particles are taken to be vortex `blobs', and particle methods in this context are known as vortex methods.

2 3
The earliest examples of vortex methods involved the use of point vortices, so that the solution was represented by t
a linear combination of Dirac distributions. However, when two vortices approach one another, the velocities that they
induce upon one another may become arbitrarily large. Thus the Dirac distributions are molli ed [14] to give `blob'
vortices. Calculations based on these have been very successful and are surveyed in [45].
The convergence of vortex methods was rst proved in [32], [33], and these results were generalised and extended
to three dimensions by Beale and Majda [6], where they demonstrated that with a judicious choice of the molli er used tm+1 P
for the Dirac distributions, in nite order convergence analogous to that of spectral methods may be obtained. In [2],
Anderson and Greengard give convergence results for a fully discrete scheme relying on stability and consistency results
obtained in [6]. A common feature of the above convergence results is that if  is a measure of the support of the
molli ed Dirac distributions and h the initial spacing between the particles, then the convergence rate depends on the
way in which =h ! 0. Typically, a relationship of the form  = h1+ is enforced, with 12   < 1.
The extension of the vortex method to viscous ows introduces issues which have still to be fully resolved. Two tm 
main approaches have been advocated. Chorin [14] used a random walk method to simulate the di usion process. This C
approach has been successfully applied in a variety of situations, but su ers from low accuracy. An alternative is a
deterministic vortex method: one possibility is to introducing a splitting of the equations as studied by Beale and Majda
[7], involving a redistribution of the vorticity at each time level. More recently various generalisations that involve no
splitting have been proposed [51], [25]. One problem here is that, since particle methods are automatically adaptive, the
particles are likely to congregate in areas of high spatial activity, and to be missing from others, and the accurate treatment A
tm 1
of di usion will be dicult in this case. Such considerations lead to the proposal [15] of a particle-grid superposition B
scheme, whereby a grid of xed particles is maintained along with the moving grid to help with the di usion.

1.1.2 Free-Lagrange methods


Most of the algorithms mentioned in the previous section are truly Lagrangian, in the sense that the particles are allowed
to move freely throughout the computation according to the ow eld. Another type of method that ts this description x
is the free-Lagrange method [27]. Here there is a grid which develops according to the ow eld, thus incorporating xj 3 xj 2 xj 1 xj
the convective terms, and the remaining terms may be dealt with in a nite-di erence, nite-element or nite-volume Figure 1.1: The trajectory tracing of the semi-Lagrangian method.
manner. For each point a record is kept of which are the neighbouring grid points, which must be updated regularly
to avoid problems such as mesh tangling. A nite element/volume formulation is developed in [64], [65], based on a
locally-Delaunay triangulation with an ecient algorithm for the update. Such methods seek to avoid the problems The recent review by Staniforth and C^ote [73] gives an excellent overview of work in this area. A major problem
associated with points absenting themselves from areas where they might be required for the approximation of di usion faced by numerical methods in meteorology is that the maximum timestep is governed by stability rather than accuracy
etc. by allowing for the addition and removal of grid points as and when required. An alternative strategy, a moving considerations. The semi-Lagrangian treatment introduced by Robert [69] lead to a reported factor of six increase in
point method as described by Farmer [23], is to interpolate on to a xed grid at each time step. Such an approach the size of the allowable timestep. The size of the timestep is here dictated primarily by accuracy, since the scheme is
combines advantages of both the Eulerian and Lagrangian frameworks, but the use of interpolation requires care to be unconditionally stable. Morevover, the phase error is greatly reduced as compared to an Eulerian method.
taken to avoid numerical di usion.
1.2.1 Trajectory tracing
1.1.3 Moving-grid methods The tracing of the trajectories is found to be a decisive factor in the accuracy of the schemes. Consider equation (1.1)
We mention here a further class of methods, generically termed moving-grid methods. These methods are adaptive, in the absence of di usion (b = 0) and with r  a = 0. The three time-level scheme rst utilised by Robert is given by
in that the grid points are required to move around according to some criterion, but they are more general than those um+1 (x) um 1 (x 2 m (x)) = 0;
described above, since this criterion need not be the satisfaction of (1.2). A commonly applied criterion is the principle (1.4a)
of equidistribution, enforced by the use of an arc-length monitor function. A comparison of two nite di erence moving 2t
grid methods and a moving nite element method is carried out in [28], where their comparison of the methods takes m (x) = tam (x m (x)): (1.4b)
into account their robustness and also the time-smoothness of the resulting system of ordinary di erential equations. Equation (1.4b) represents the use of the midpoint rule for the integration of (1.2), and is solved by a Picard iteration.
In the moving nite element method [53, 54], [4], [90], the movement of the nodes is controlled by a minimisation Although (1.4a) places no stability-related timestep restriction, the convergence of this iteration does introduce a restric-
procedure, in line with the nite element approach. The method has encountered problems with points drifting close tion on t, but this remains an order of magnitude larger than the stability restriction to which an Eulerian timestepping
together, and with parallelism, causing the mass matrix to become singular, and has needed problem-dependent tuning scheme is subject.
to avoid these diculties. Controlling the mesh in a systematic way remains an area of active research. Figure 1.1 illustrates the solution of (1.2) by (1.4b). The solid line represents the exact trajectory passing through
the point (xj ; tm+1 ), landing at A, and the dashed line is the approximation given by (1.4b), which passes through C
1.2 The semi-Lagrangian method (coordinates (xj m (xj ); tm )) and lands at B (coordinates (xj 2 m (xj ); tm 1 )).
The three level scheme (1.4) is almost a set of two decoupled two level integrations, apart from the fact that in
Purely Lagrangian methods of the type discussed in the previous section are susceptible to the fact that an initially applications the velocity eld a depends on u either explicitly or implicitly through other equations to which (1.1) is
regularly-spaced set of points will generally evolve into a highly irregular set. In meteorological circles semi-Lagrangian coupled. Temperton and Staniforth [88], by extrapolating a to tm from tm 1 and tm 3 , managed to e ectively decouple
methods have achieved great popularity. Here a di erent set of points is used at each time step. These are chosen so these processes and arrived at a two time level scheme, reducing the amount of work involved.
that they arrive exactly at points of a regular mesh at the top time level; in general then the feet of the trajectories of
these points will not be points on the regular mesh, and a characteristic feature of the semi-Lagrangian methods used in 1.2.2 Interpolation
meteorological circles has been the use of interpolation from values at nearby mesh points. We shall return to this point
below. A second crucial factor in the success of these methods is the interpolation that is used at the feet of the trajectories.
Too low an interpolation introduces excessive damping|too high a formula is expensive. For most applications, cubic
interpolation is found to be an e ective compromise. For some meteorological applications, the avoidance of overshoots

4 5
and undershoots is of prime importance; with this in mind Williamson and Rasch [91] make use of shape-preserving and
monotone interpolation. Another alternative is the non-interpolating semi-Lagrangian method of Ritchie [66]. Here the
and (1.5) is then discretised by Z
velocity eld a is split up into two parts. The rst part a is such that each trajectory through a grid point at the top
hU m+1; i i = U m (x)i(y)dy; (1:7)
time level lands exactly on the nearest grid point to the base of the original trajectory through that grid point, thus where here h; i represents the L2 inner product over Rn .
requiring no interpolation. The residual velocity eld a+ = a a is dealt with in a standard Eulerian fashion. The To obtain the characteristic Galerkin method, (1.7) is then rewritten in an equivalent form,
stability advantages of the semi-Lagrangian method are maintained, since this residual velocity always satis es a unit CFL
hU m+1 U m ; i i + thr  f (U m ); i i = 0; (1:8a)
condition, and the damping associated with the interpolation is avoided. However, the dispersion properties of Eulerian where Z x+a(U m(x))t
methods have returned, albeit with smaller coecients associated with them (since they are associated with the residual
velocity). These ideas have been generalised and extended by Smolarkiewicz and Rasch [72], who develop a formalism i (x) = ja(U m1(x))jt i(z )dz: (1:8b)
x
The scheme (1.8) is unconditionally stable, conservative, and highly accurate (splines of order s give accuracy of order
which may be used to convert any advection algorithm into a semi-Lagrangian format.
Ritchie [66] also argues that the non-interpolating scheme opens the way for the incorporation of a spectral space- 2s 1). Moreover, the accuracy can be enhanced through recovery procedures, and (1.8) is still valid when shocks form.
discretisation into the semi-Lagrangian method. Spectral methods have long been used in meteorological models, but However, there is a diculty in the application of the method to systems of conservation laws, which has inhibited the
their advantages would have been lost with the low order grid point interpolation of the semi-Lagrangian method. Ritchie use of the method.
[67, 68] has incorporated a spectral discretisation and compared the interpolating and non-interpolating versions; he nds
the interpolation does indeed lead to a deterioration of performance at short scales. 1.3.2 Lagrange-Galerkin methods
Let us return to (1.1) and (1.2) and rewrite (1.2) in the more explicit form
1.2.3 Other interpolations d
Similar ideas have appeared in the engineering literature, with applications in such areas as subsurface hydrology. In ds X (x; t; s) = a(X (x; t; s); s) (1.9a)
the Holly-Preissman scheme [38], two functions are evolved in time, representing the solution and its rst derivative at X (x; t; t) = x: (1.9b)
the grid points. These are found at the new time level by tracing back along characteristics through grid points and For a function w on Rn we de ne [E (t; s)w](x) = w(X (x; t; s)). Then the basic Lagrange-Galerkin method for approx-
using a cubic Hermite polynomial interpolation of the solution and its rst derivatives at the grid points adjacent to the imating (1.1) takes the form
feet of the characteristics to nd the function values there. Di usion and other terms are incorporated using a splitting
technique. Recent extensions include the inclusion of second derivatives into the above scheme, resulting in a fth-order hU m+1 ; i i + thbm+1rU m+1 ; ri i = hE (tm+1 ; tm )U m; i i: (1:10)
accurate method [92], and allowing the characteristics to reach back more than one time level, with a view to reducing When a depends on u, the discretised form of (1.9) must usually be solved by some iterative means. Often a forward
the numerical dispersion associated with the interpolation [93]. Euler discretisation of (1.9) will suce, although the midpoint rule of the semi-Lagrangian method is also an option. The
method is, of course, similar to the semi-Lagrangian methods already described, although it incorporates the accuracy
1.3 Finite element methods and exibility of the nite element method, and, at least in its pure form, it avoids the disadvantages of the use of
interpolation at the feet of the characteristics.
It is well known that nite element approximations enjoy optimal approximation properties in appropriate integral norms, Such methods have been developed in various settings ([21], [8], [9], [36], [49], [60], [75] and [76]), with each group
most straightforwardly when applied to self-adjoint elliptic problems. Similar results for non self-adjoint problems can of authors using a di erent name. The modi ed method of characteristics [21] was developed with a view to applications
be obtained by the Petrov-Galerkin approach, whereby the test functions are selected with a view to symmetrising the in petroleum engineering. Generalisations to higher order in time versions were studied in [22]; we shall return to this
bilinear form associated with the method [5]. subject in Chapter 4. The problem of boundary conditions is one that has received scanty treatment in the literature|this
Semidiscrete Galerkin nite element approximations of convection-di usion problems also possess many attractive issue was addressed by Russell and his co-workers in [71], where they use space-time elements. The transport-di usion
approximation and conservation properties. However, when time discretisations are introduced, these properties dete- algorithm was applied to the Navier-Stokes equations and analysed in [60] (obtaining slightly sub-optimal convergence
riorate noticeably, and unless severe restrictions are placed on the scheme, unphysical oscillations occur on the scale rates), and the Lagrange-Galerkin method for the same equations was analysed in [76] (obtaining optimal convergence
of the mesh. Attempts to avoid such phenomena include the adapting of the Petrov-Galerkin approach, choosing the rates and demonstrating non-linear stability). The engineering literature has preferred the name Eulerian-Lagrangian
test functions to satisfy a unit CFL condition [56], and the Taylor-Galerkin approach of Donea [19], which enjoyed methods for these and related ideas (e.g. [58]), and this nomenclature was adopted by Russell in [71].
similar accuracy properties yet was easier to extend to more complicated problems. Another approach, which has been The convergence and stability properties of the method (1.10) are impressive (the scheme is unconditionally stable).
extremely successful, is the streamline-di usion method, rst introduced by Hughes and Brookes [39]. The method can However, much of the above mentioned analysis was carried out under the assumption that the inner products involved
be viewed as a Petrov-Galerkin method with a test function modi ed by adding a multiple of a linearised form of the in evaluating the right hand side are carried out exactly. In practice some form of quadrature is necessary. Unfortunately
hyperbolic operator applied to the test function. This gives added stability without sacri cing accuracy, and in the scalar this causes the unconditional stability to be lost. Instability was rst observed by Priestley [61] and the analysis in [57]
one-dimensional case can be viewed as adding arti cial di usion along the streamlines. References to recent work can and [77] showed that for most commonly used quadrature schemes, the method is only conditionally stable, and that for
be found in [42]. some the stability condition cannot be met by reducing the size of the timestep. The presence of di usion does have a
stabilising in uence on the method, however, and Jack [40], [41] has investigated the impact of these theoretical results
in practice.
1.3.1 The characteristic Galerkin method In [57], an alternative technique was presented, using ideas developed for particle-in-cell methods. This is termed
Here we consider a method primarily designed for the treatment of hyperbolic conservation laws, although it is intimately area weighting and involves transporting the support of each nite element basis function without rotation according
related to more generally applicable schemes to be discussed below. The characteristic Galerkin method, as the name to the path taken by the centroid of the element, before the integrals in (1.10) are carried out. This has the e ect
suggests, combines the method of characteristics with a Galerkin approach, o ering distinct improvements over the use of restoring the stability properties of the exactly-integrated scheme without signi cantly deteriorating the accuracy
of interpolation. For an extensive account of the method we refer to [13]; here we give a brief description of some of the [57]. Its e ectiveness, however, is limited to cartesian meshes, and this led Jack [41] to investigate another alternative,
main features. called subdivision. However, in practice, he concluded that this was less e ective than the careful use of quadrature. It is
Consider (1.1), with b = 0. Since u is constant along the characteristics, they are straight lines. Thus we may write worth commenting at this point that the area-weighted Lagrange-Galerkin method is very similar to a particle-in-cell nite
element method which Bermejo [10] has shown, on rectangular cartesian meshes, to be equivalent to the semi-Lagrangian
um+1 (y) = um (x) where y = x + a(um(x))t: (1:5) method with cubic spline interpolation.
The deterioration of the stability properties of the Lagrange-Galerkin method when quadrature is introduced can be
A nite element representation of um in terms of the basis functions i may be written attributed to the lack of smoothness in the basis functions. For Eulerian schemes this is not a problem, but here the
X basis functions are shifted, and the discontinuity becomes apparent to the integration formula. Such considerations were
Um = Uim i ; (1:6) a major factor in the original proposal of the spectral Lagrange-Galerkin method [78], which will be investigated in detail
i below.

6 7
1.3.3 The weak Lagrange-Galerkin method interpolant of u at the quadrature nodes, so that this is the suitable transform to use in a collocation method for, say, the
evaluation of a derivative at the collocation/quadrature points. For the most common systems (Fourier and Chebyshev
We bring this brief survey to a close by describing an alternative nite element method for (1.1), which may be formulated polynomials), the discrete transform may be calculated eciently by using the FFT in 25 N log N operations, where N is
by shifting the test functions instead of the trial functions. Let us integrate (1.1) in space and time against time-dependent the number of polynomials in the expansion, compared with 2N 2 operations without the use of the FFT.
test functions. We have Z m+1
t
h@tu + r  f r  (bru); i idt = 0; 1.4.3 Spectral methods for time-dependent problems
tm
which, after some integration by parts, gives Most standard spectral methods for time-dependent problems fall within the framework of the method of lines (m.o.l.)
Z tm+1 Z tm+1 approach, where the partial di erential equation is discretised rst in space, giving rise to a system of ordinary di erential
hum+1 ; m+1 hum ; m equations which may be solved by the method of choice. Analysis of such schemes thus centres rst on the semidiscrete
i i i i+ m hbru; r i idt = [hu; @t i i + hf; r i i]dt: (1:11) equations formed by the spatial discretisation. The review by Tadmor [81] is interesting for our present purpose since
t tm
he deals with the periodic, hyperbolic case exclusively. As expected, the stability of the Fourier-Galerkin method exactly
Let us set c = f=u. Then if we replace a by c in (1.9) and set i (x; t) = i (X (x; tm+1; t)), the right hand side of (1.11) mirrors the well-posedness of the original equation. However, once the discrete Fourier transform, and along with it
vanishes, and on approximating the integral on the left hand side in some way we arrive at the weak Lagrange-Galerkin aliasing, is introduced, stability is less straightforward, at least in theory. In fact, the theoretical stability of the method
method of Benque et. al. [8]. This formulation has the advantage of being conservative in the case b = 0, and for this remains an open problem, although in practical computations the introduction of aliasing has not been observed to cause
reason is favoured by Priestley [62] in his implementation of the spectral Lagrange-Galerkin method on a meteorological any problems. (For a discussion of this point see also [12]).
test problem. It is also the formulation of an independent development of spectral Lagrange-Galerkin methods by Ho et. There are at present two ways around this diculty [44]. One path one might take is the skew-symmetrisation of the
al. [37]. They use a quadrature formula on the term hum ; im i which uses points which have been shifted according spatial di erentiation operator. Pasciak [59] analyses a problem already in this form. The second way to obtain stability is
to (1.9). To evaluate um at those points they use an ecient `regridding' technique which is based on solving a pure the introduction of ltering. This is discussed in detail by Kreiss and Oliger in [44], where they propose a delicate ltering
advection subproblem for which they use an Adams-Bashforth time integrator. All this is combined with a spectral procedure that enforces a minimal decay rate in the higher region of the spectrum, and leaves unchanged polynomials
element method to enable the solution of problems on non-rectangular domains. which are suciently smooth, in the sense that they satisfy an inequality in which some high order Sobolev norm is
bounded by the L2 -norm. As in Tadmor [81], stability is more straightforward to demonstrate with the (less accurate)
application of the crudest of lters|merely cutting o a xed proportion of the highest modes. This can be seen as
1.4 Spectral methods corresponding to the use of quadrature with the number of quadrature points proportional to and strictly greater than
Having discussed various implementations of Lagrangian approaches for convection-di usion problems, we now outline the number of modes in the solution. For the spectral Lagrange-Galerkin method, the situation is analogous, although
by way of introduction some aspects of spectral methods, and mention some work that is of relevance to us, before we of necessity consider the fully discrete case, since the time discretisation is carried out rst (see Chapter 4 for a
setting out the programme for the remainder of this Thesis. full discussion of this point). An exactly-integrated Fourier-Galerkin spatial approximation gives rise to a scheme whose
Spectral methods have developed into a widely used technique for the solution of di erential equations. Their increase stability properties mirror the well-posedness of the original equation, whereas the use of quadrature introduces aliasing
in popularity is due in part to the fact that the approximations underlying spectral methods have very high accuracy errors which must be controlled in some way. This is achieved by ensuring that the number of quadrature points exceeds
and that, with the availability of the Fast Fourier Transform (FFT), they may be implemented eciently. More recently, that of the number of coecients in the solution. The resulting scheme can also be viewed as a ltered collocation
much work has been put into overcoming the limitation of standard spectral methods to simple domains, and into method. What is in some ways the central result of this Thesis is that the ltering that is used by Tadmor [81] in his
extending their remarkable e ectiveness to problems with non-smooth solutions. Spectral element, and spectral domain- proof is sucient to guarantee unconditional stability for the spectral Lagrange-Galerkin method. A nal point to make
decomposition techniques, are areas of particularly active development in this respect. The standard reference on spectral in this comparison is that in practical computations the need for such ltering has not been observed.
methods is [12], which contains a uni ed theory of their mathematical analysis, together with a fairly comprehensive For time-dependent problems, the ordinary di erential equations resulting from spectral discretisations in space
description of applications of spectral methods in uid dynamics. are often solved by standard methods, such as a linear multistep, or a Runge-Kutta method. In contrast with the
spectral Lagrange-Galerkin method, such calculations are usually constrained by a CFL-type condition, which for Jacobi
polynomials takes the form t  CN 2 (In the Fourier case this is replaced by the less restrictive t  CN 1 ). This is
1.4.1 Spectral approximation sometimes heuristically viewed as being a relation between the size of the timestep and the minimum gridsize. However,
The expansion of a function in terms of an in nite sequence of orthogonal functions underlies many numerical methods Gottlieb and Tadmor [30] show that a more rigorous explanation is that the N 2 -term corresponds to the growth of the
of approximation, including spectral methods. The usefulness of these methods depends upon their accuracy and upon eigenvalues of the Sturm-Liouville problem associated with the Jacobi polynomials.
the eciency of their implementation. With regard to the accuracy of spectral methods, the most familiar results With this procedure, a high order discretisation in space is coupled with a compara tively low order discretisation
concern the approximation of periodic functions by Fourier series. Here the kth coecient of the expansion decays in time. For problems in which the solution changes over a time scale which is comparable to the spatial scale, as in
faster than any inverse power of k when the function is in nitely smooth. (For holomorphic functions, this decay is many hyperbolic problems and advection-dominated di usion problems, the time errors will dominate the spatial errors,
actually exponential [80]). A particular characteristic of this decay is that it is not observed until the expansion contains necessitating the use of very small timesteps. Gottlieb and Turkel [31] are led by this observation to introduce a coupling
enough terms to e ectively resolve the salient features of the function. However, once this point is reached the series will of the space and time discretisations. They obtain schemes which are unconditionally stable, but still are subject to
provide a very good approximation with the inclusion of a very few additional terms. This behaviour is referred to as the accuracy restrictions on the size of the timestep.
spectral accuracy of the Fourier method. For smooth, non-periodic functions, such spectral accuracy may be attained Tal-Ezer [83] has proposed a spectral method in time for hyperbolic equations (and for parabolic equations [84]),
by expanding in terms of Jacobi polynomials. also with a view to eliminating the discrepancy between the temporal and spatial discretisations. His method applies to
linear problems with known bounds on the spectrum of the spatial operator L and is based on an orthogonal polynomial
expansion of the evolution operator generated by L. His method is relatively ecient; in Chapter 3 we shall compare its
1.4.2 The discrete Fourier transform performance with that of the spectral method of characteristics for a linear hyperbolic problem.
This expansion in terms of an orthogonal system introduces a linear transformation between u and the sequence of its
expansion coecients fu^(k)g. If the system is complete in a suitable Hilbert space, this transform can be inverted. Hence 1.5 Summary of contents
functions can be described both through their values in physical space and through their coecients in transform space.
These coecients can rarely be calculated exactly, since they are integrals depending on all the values of u in physical We shall begin, in the next chapter, by considering a simple advection equation|a scalar hyperbolic problem. Our
space. If a high precision quadrature formula is used, approximate values for a nite number of expansion coecients purpose here will be to investigate the properties of the numerical treatment of the evolution operator associated with
may be calculated. Such a process describes a discrete transform between the values of u at the nodes of the quadrature evolving the solution of the hyperbolic equation through one time step. This is the spectral-Lagrange treatment of the
formula and the set of discrete coecients. The quadrature points are chosen so that an approximation de ned by such convective terms, and it is shown to be unconditionally stable, both in the L2-norm and in higher order Sobolev norms.
discrete coecients has the same accuracy properties as the exactly integrated expansion, and it may be used instead
in practical computations. If the quadrature formula is chosen appropriately, the approximating function is actually the

8 9
The central result here is that these results continue to hold even with the introduction of quadrature, which amounts we denote the norm on L2# (
) by kk. We note that the set fk gk2Z n forms a Hilbert basis for L2# (
), so that any
to turning the Fourier-Galerkin method into a Fourier pseudospectral method with truncation of higher modes. v 2 L2# (
) may be written X
The method as it is described in Chapter 2, although it has very desirable stability and convergence properties, is very v= v^(k)k ; (1:13)
inecient to implement, especially in higher dimensions. In Chapter 3 we describe an additional approximation which, k 2Z n
if included in the method, serves to maintain the spectral accuracy of the original scheme and yet may be implemented
eciently. After providing stability and convergence results for this we illustrate these results and those of the previous where
chapter by means of a series of numerical experiments. v^(k) = (v; k ); k 2 Z n;
Chapter 4 describes the application of the method to convection-di usion problems. The equation is rewritten in are the Fourier coecients of v.
Lagrangian coordinates and is discretised in this form rstly in time and then in space, and backward di erence methods For 2-periodic functions f , g we de ne the convolution by
up to order 6 are analysed within a general framework provided by Le Roux [46]. Again the theoretical results are Z
illustrated by numerical experiments. We then begin to turn our attention to non-linear problems, and we consider rst (f  g)(x) = (2) n f (x y)g(y)dy: (1:14)
a discretisation of Burgers equation. Here we break with the pattern of previous investigations and content ourselves

with numerical results. They provide a lead in to the nal chapter, where we consider the Navier-Stokes equations. This enjoys all the usual properties of convolution in Rn .
A full analysis for the spectral Lagrange-Galerkin method with a backward-Euler timestep is given, followed by some For k 2 N0 [ f1g, C#k (
) is the space of functions from
to R which are the restrictions to
of functions from
results obtained by a numerical study of the evolution of a perturbed double shear layer. Scope for future investigation Rn to R which are k-times continuously di erentiable and 2-periodic in each coordinate direction. (For k = 0 we write
is discussed, as well as conclusions drawn from the current work. C# (
)). We note that for v 2 C#k (
) the Fourier coecients v^(p) of v decay like jpj k .
The norm on C#(
) will be denoted by j  j1 , and for m  1, the norm of u = (u1 ; : : : ; um ) 2 C# (
)m will be
given by
1.6 Notations X
m ! 12
Here we introduce much of the notation that will be used below, and describe some of the main function spaces in which
juj1 = jui j21 :
i=1
we will be operating.
Z denotes the set of integers, N is the set of positive integers and N0 the set of nonnegative integers. R denotes the For i = 1; : : : ; n, let Di = @=@xi ; the gradient operator r : C#k+1(
)m 7! (C#k (
)m )n may then be written r =
set of real numbers and C is the set of complex numbers. The integer part of a nonnegative real number  is denoted (D1 ; : : : ; Dn ). If = ( 1; : : : ; k ), with each i 2 f1; : : : ; ng, we write n( ) = k and D = D 1 : : : D k . (We shall
by [], and we further de ne [] = minfn 2 N0 : n  g. For x = (x1 ; : : : ; xn), y = (y1 ; : : : ; yn ) 2 Rn , we de ne make use of this non-standard multiindex notation throughout this Thesis.) Then the norm of u 2 C#k (
)m is given by
0k  1 12
x  y := xi yi ; X k X
jujk;1 = @ j n( )=j jD uj21 A ;
where we have made use of the summation convention, and j=0
X
n which may also be written
0k  11
jxj1 := jxi j;
i=1 @ X k j 2 A2
jujk;1 = jr uj1 : j
jxj1 := 1max
jn
jxj j: j=0
We introduce the dual D#0 (
) of C#1 (
) (consisting of conjugate-linear functionals), with duality pairing h; i; this
jxj will denote the Euclidean norm of the vector x. is the space of periodic distributions. For further details to the description below, we refer to [52]. For f 2 D0 (
) we
#
For s 2 N , let s0 = f0; 1gs , and s = f0; 1gs nf(0; : : : ; 0)g. Then the binomial theorem generalises to can de ne the Fourier coecients of f by
Ys X Ys f^(p) = hf; p i:
(ar + br ) = ar r b1r r : (1:12) f may be expanded in a Fourier series as in (1.13), but with convergence in D#0 (
) instead of in L2# (
). An important
r=1 2s0 r=1 result in this respect is that f 2 D#0 (
) if and only if 9 > 0 : (1 + jpj2 )  f^(p) ! 1.
Suppose that K 2 N0 and  = (1 ; : : : ; n), 0  j  K , then Periodic distributional derivatives are de ned by
hD( ) f; i = ( 1) hf; D( ) i; 8 2 C#1 (
): (1:15)
X X
K X
K
:= ::: : We note that derivatives in the above sense do not in general coincide with those in the sense of D0(
);when a function
 1 =0 n =0 is regular but not periodic, the non-periodicity on the boundary of
will induce the same behaviour in the (periodic
distributional) derivative of that function as would a corresponding discontinuity in the interior of
. For  2 N0 , we
Let
= [0; 2)n. For any N 2 N0 , we denote by SN the space of trigonometric polynomials on
with degree de ne the Sobolev space H# (
) by
 N , i.e. 8 9
< X = H# (
) = fu : D u 2 L2# (
); 0   g;
SN = :v = v^k k ; ;
jkj1N (where the derivatives are in the sense of periodic distributions) with norm
where 0  1 21
k (x) := eikx : X  X
kuk = @ k n( )=k kD uk2 A : (1:16)
Let L2# (
) be the space of square-integrable functions on
, periodically extended to the whole of Rn . Then N will k=0
denote the orthogonal projector in L2# (
) onto SN , and, for u, v 2 L2#(
), (u; v) will denote the L2# (
) inner product We also de ne the semi-norm j  j on H# (
) by
Z 0 1 12
(2) n u(x)v(x)dx; X

juj = @ kD uk2 A :
n( )=

10 11
De ning the norm on H# (
)m equivalently to that on C#k (
)m , we may rewrite
juj = kr uk ;
and kuk may be rewritten
X ! 12
kuk =
p
(1 + jpj2 ) ju^(p)j2 : (1:17) Chapter 2
This allows us to extend the de nition of H# (
) to non-integer :
H# (
) = fu 2 D#0 (
) : kuk < 1g: (1:18) A model hyperbolic problem
For  < 0, H# (
) is de ned to be the dual space of H#  (
), and may also be characterised by (1.18) and (1.17).
Let T > 0 and 0 = t0 < t1 < : : : < tM = T . For a Banach space X ,
  As indicated in the previous chapter, the Lagrange-Galerkin method is designed for problems in which convection is the
l1 (X ) := v : ft0 ; : : : ; tM g ! X j kvkl1 (X ) = 0max
mM
kv(tm )kX < 1 : dominant term in the governing equation. It falls into a particular class of schemes, each of which is based on a particular
strategy for dealing with the convective term. In this chapter we focus attention on the strategy that is characteristic of
Further, C k (X ) denotes the space of k-times continuously di erentiable mappings from [0; T ] into X ; for 1  p  1, the spectral Lagrange-Galerkin method, and so we consider the application of the scheme to a linear scalar hyperbolic
Lp(X ) will often be used as shorthand for Lp(0; T ; X ), and W k;p(X ) for W k;p (0; T ; X ), for k  0. For any positive problem.
integer m, when there is no risk of confusion with the norm on C#k (
)m , the norm on Lp (C#k (
))m will be denoted We begin by posing the problem in a weak form. We construct solutions by means of the method of characteristics,
j  jk;p . For s; t 2 [0; T ] the norm on Lp (s; t; C#k (
))m will be denoted j  jk;p;[s;t]. For any Banach spaces Y and Z , and discuss their regularity properties. This involves results concerning the well-posedness of the problem in various
L(Y ; Z ) denotes the Banach space of continuous linear operators from Y to Z equipped with the induced operator norm. norms, and leads naturally to the formation of a numerical method which has analogous stability properties, and which
The letter C will denote a generic positive constant, and C0 , C1 , etc. will stand for speci c positive constants. In most forms the basis of the spectral Lagrange-Galerkin method. In subsequent sections we consider the e ects of further
places, the nature of the dependence of these constants will be made explicit. approximations that must be made in order to create a scheme that may be implemented eciently. These are: the use
of quadrature in the evaluation of the integrals that form an essential part of the method, and the introduction of a local
interpolation (in order to improve eciency). In both cases we show that the original stability and accuracy properties
can be maintained.

2.1 Discussion of the model problem; the evolution operator E


For u0 2 L2# (
) and a 2 L1 (C#1 (
)n), consider the following initial boundary value problem: nd u in C (L2#(
)) \
W 1;1 (H# 1 (
)) such that
@u + a  ru = 0 in
 (0; T ], (2.1a)
@t
u(; 0) = u0 in
. (2.1b)
We shall establish the existence and uniqueness of solutions to (2.1), and discuss their regularity, using the method of
characteristics as our main tool. This approach yields results which are central to much of the analysis in this Thesis. We
want to construct curves along which solutions of (2.1) remain constant. With this end in view we proceed as follows:
for x 2 Rn and t0 2 [0; T ], we de ne the mapping X (x; t0; ) : [0; T ] ! Rn as the solution of the initial value problem
d
dt X (x; t0 ; t) = a(X (x; t0; t); t); a:e: t 2 [0; T ]nft0 g; (2.2a)
X (x; t0 ; t0 ) = x; (2.2b)
where we have periodically extended a to the whole of Rn . As we shall show below, the solutions of (2.2) are continuous
curves in space-time passing through the points (x; t0). Given sucient smoothness of u and a, we can apply the chain
rule and obtain
d dX
dt u(X (x; t0; t); t) = ut(X (x; t0; t); t) + dt (x; t0 ; t)  ru(X (x; t0 ; t); t)
= [ut + a  ru](X (x; t0 ; t); t):
Thus smooth solutions of (2.1) will satisfy, for each x 2 Rn ,
u(X (x; t0; t); t) = constant,
and then (2.2) de nes the curves we seek.

12 13
For each x 2 Rn and t0 2 [0; T ], (2.2) is mathematically equivalent to the following problem: Corollary 2.1 For x 2 Rn and t0 ; t;  2 [0; T ],
d X (x; t0 ; t) = X (X (x; t0;  );  ; t): (2:8)
dt Y (x; t0; t) = a(x + Y (x; t0 ; t); t); a:e: t 2 [0; T ]nft0 g; (2.3a)
Y (x; t0 ; t0 ) = 0; (2.3b) Proof Since both the left and the right hand sides of (2.8) (regarded as functions of t) are solutions to
(2.2), they satisfy
where Y (x; t0 ; t) = X (x; t0; t) x, and we note that solutions to (2.3) will be 2-periodic in their spatial variables.
d
Theorem 2.1 Let a 2 L1 (C#1 (
)n), x 2 Rn and t0 2 [0; T ]. Then (2.3) has a unique solution Y (x; t0 ; ) 2 W 1;1 (0; T ). dt (t) = a((t); t); a:e: t 2 [0; T ];
Moreover, if a 2 L1 (C#r (
)n) for integer r  1, then Y (x; t0 ; t) and @Y=@t0 (x; t0 ; t) are continuous in each of their ( ) = X (x; t0;  );
variables, and for each t0 2 [0; T ], the map  : (x; t) 7! Y (x; t0; t) satis es  2 W 1;1 (C#r (
)n) and the map : (x; t) 7! the result follows since such solutions are unique.
@Y=@t0 (x; t0 ; t) satis es 2 W 1;1 (C#r 1(
)n)). Setting
We now de ne the evolution operator E.
X (x; t0 ; t) = x + Y (x; t0 ; t); x 2 Rn ; t0 ; t 2 [0; T ]; (2:4)
De nition 2.1 Let w be a function from a subset A of Rn to R. For t0 ; t 2 [0; T ] and x; X (x; t0; t) 2 A de ne
then provides the unique solution to (2.2), with corresponding smoothness properties.
E (t0 ; t)w(x) = w(X (x; t0; t)): (2:9)
Remark. This is only a slight generalisation on standard results for such systems of ordinary di erential equations (as
found in [35], for example). We give a sketch of the proof.
Proof Integrating (2.3) formally, we obtain Motivated by the discussion following (2.2), our solution to (2.1) will be de ned by
Zt u(x; t) = u0 (X (x; t; 0)) = E (t; 0)u0(x): (2:10)
Y (x; t0 ; t) = a(x + Y (x; t0 ;  );  )d: (2:5)
t0 In order to ascertain whether (2.10) will de ne a solution to (2.1) under the weaker assumptions we have made on u0 ,
De ning, iteratively, and to be able to discuss its regularity properties, we have to answer some questions about E .
The rst question is whether E (t0 ; t) : L2#(
) ! L2# (
). This is answered in the following lemma.
Y0 (x; t0; t) = 0
Zt Lemma 2.1 Let t0 ; t 2 [0; T ]; then E (t0 ; t) is a quasi-isometry of L2#(
) onto itself, with
Yn+1 (x; t0; t) = a(x + Yn(x; t0 ;  );  )d;
t0 kE (t0 ; t)kL(L2# (
))  e 21 jraj0;1;[t0;t] : (2:11)
we obtain a sequence converging uniformly for t in Bt0 = ft 2 [0; T ] : jraj0;1;[t0 ;t] < 1g. Since each
Yn(x; t0 ; t) is continuous for t 2 Bt0 , this holds true for the limit Y (x; t0; t). It also follows readily that
Y (x; t0 ; ) solves (2.3) on Bt0 and is unique there. To see how to extend this construction of the solution Proof De ne  : Rn ! Rn by (x) = X (x; t0 ; t); x 2 Rn . By virtue of Liouville's formula applied to
to the whole of [0; T ], we take some t1 2 Bt0 (t1  t0 ) such that jraj0;1;[t0 ;t1]  12 . We know that such (2.6) (see [35]), the Jacobian J (x; t0 ; t) of the mapping x ! X (x; t0 ; t) is given by
a point exists, except perhaps if T 2 Bt0 , because of the continuity of the map t 7! jraj0;1;[t0 ;t] . Now we R t (ra)(X(x;t0; ); )d
let Bt1 be de ned analogously to Bt0 and de ne for t 2 Bt1 the iteration J (x; t0 ; t) = e t0 :
Y00 (x; t0; t) = Y (x; t0 ; t1 )
Zt
Thus Z Z
jv((x))j2dx = jv(y)j2 J (y; t; t0 )dy;
Yn0 +1(x; t0; t) = Y (x; t0 ; t1 ) + a(x + Yn0 (x; t0 ;  );  )d:
(
)
t1 where, (making use of (2.8)), we have replaced x by y = (x) = X (x; t0; t) (so that x = X (y; t; t0)). We
The limit Y 0 (x; t0 ; t) solves (2.5) on Bt1 , by uniqueness, and agrees with Y (x; t0 ; t) on Bt0 \Bt1 . Continuing want to replace the integral over (
) by one over
. For j 2 Z n de ne the set
j by
in this way we can extend the solution on to a sequence of intervals increasing to the right. It is impossible
that these intervals should remain within [0; T ] inde nitely, since the way we constructed the points t0 ; t1 ; : : :
j := [0; 2)n + 2j:
would then imply that jraj0;1;[0;T ] is unbounded, contradicting a 2 L1 (C#1 (
)n). Thus we can extend the Then, since (x) x is 2-periodic,
solution to [t0 ; T ], and by changing directions in the above argument, to the whole of [0; T ]. We obtain
a unique continuous map t 2 [0; t] 7! Y (x; t0 ; t); since a 2 L1(C#1 (
)n) this map is di erentiable almost (
k ) \
l = ((
k l) \
0 ) + 2l 8k; l 2 Z n:
everywhere and satis es (2.3). Moreover, Y (x; t0; ) 2 W 1;1 (0; T ) as required, and continuity in both x The sets
j (and (
j )) are mutually disjoint and cover Rn , and so it follows that, if f is a locally integrable
and t0 is also assured. 2-periodic function on Rn ,
The rst step towards obtaining the further regularity results is to di erentiate formally (2.3). Making use Z XZ
of the summation convention, we obtain f (x)dx = f (x)dx
@ (
0 ) l2Z n Z(
0 )\
l
@t Di Yj (x; t0 ; t) = (ik + Di Yk (x; t0 ; t))Dk aj (x + Y (x; t0 ; t); t); (2.6a) X
= f (x)dx
Di Yj (x; t0 ; t0 ) = 0; (2.6b) l2Z n Z((
l )\
0 )+2l
X
and = f (x)dx
@ @Yi @Yj lZ2Z n (
l )\
0
@t @t0 (x; t0; t) = @t0 (x; t0 ; t)Dj ai (x + Y (x; t0; t); t); (2.7a)
= f (x)dx:
@Yi (x; t ; t ) = a (x; t ); (2.7b)
0
@t0 0 0 i 0
Thus Z Z
with (2.6a) and (2.7a) holding a.e. t 2 [0; T ]nft0 g. jv(y)j2 J (y; t; t0 )dy = jv(y)j2 J (y; t; t0 )dy;
Analogous arguments to those applied to (2.3) may be applied to (2.6) and to (2.7), and the conclusions (
)

of the theorem for r = 1 follow. For higher values of r an inductive argument is used. and the result follows.

14 15
The next question concerns the action of E on the spaces H# (
),   0, and is addressed by the following lemma. The application of D 0 to the right hand side of (2.16) will necessitate the use of Leibnitz' rule. We de ne,
 for  2 s0 1 , D 0; = D 11 : : : D ss 11 . Then, combining (2.15) and (2.16), we have
Lemma 2.2 Let   0, and suppose that a 2 L1 (C#[] (
)n). Let t0 ; t 2 [0; T ]; then E (t0 ; t) is a quasi-isometry of
H# (
) onto itself, with D E (t)u = D 0 E (t)D s u + D 0  sk (t)E (t)Dku  
kE (t0 ; t)kL(H# (
))  e 21 jraj0;1;[t0;t] +C1jaj[];1;[t0 ;t] ; (2:12) 0 X 0;
= D E (t)D s u + D  s k (t) D 0 ;1  E (t)Dk u :
where C1 = C1 (n; ; jaj[] ;1;[t0 ;t]), and if  = 0, C1 = 0. 20s 1
In addition, for integer , E (t0 ; t) is also a quasi-isometry of C# (
) to itself, with (2.17)
kE (t0 ; t)kL(C# (
))  eC1 jaj[] ;1;[t0;t] : (2:13) By density, (2.17) holds, in the sense of distributions, for u 2 H# (
), and henceforth we only assume that
u comes from H# (
).
Finally, an estimate for the size of C1 may be provided. If we de ne, inductively, We multiply both sides of (2.17) by D E (t)u, integrate over
, sum over all such that n( ) = s, and
Pk obtain, by the Cauchy-Schwarz inequality,
1
c0 = 1; c k = ek j =0 cj jajj +1;1;[t0;t] ; (2.14) 0 n 1 21
P  X @X
where k = ( kj=1 kj ) 21 , then C1 is such that jE (t)ujs  jE (t)rujs 1 + jjk (t)j2jj1;1 A jE (t)rujs 1 j j 1
2s0 1 j;k=1
eC1 jaj[] ;1;[t0;t]  c[] : s 1 s 1
X
= jE (t)rujs 1 + r r (t)jE (t)rujs 1 r ; (2.18)
r=0
Proof We prove the theorem rst for  integer. For non-integer , the result may be deduced by inter- Pn  21
where r (t) = j;k=1 jjk (t)jr;1 . Again using the Cauchy-Schwarz inequality, and writing Es (t) =
2
polation. We use induction on , and so we suppose that the statement of the theorem holds when  is
replaced by 0; : : : ;  1, and deduce that it holds for . Ps s 2 (t) 12 , we nd
r=0 r r
Since C#1 (
) is dense in H# (
), we rst consider the action of E (t0 ; t) on u 2 C#1(
). We denote E (t0 ; t)
by E (t). We shall also denote X (; t0 ; t) by X (t). jE (t)ujs  jE (t)rujs 1 + Es 1 (t) kE (t)ruks 1 : (2:19)
P
For v = p v^(p)p 2 H# (
), But then, repeating the above calculation, and writing jvjs = krs vk,
X ! 21 X
s 1
kv k  = (1 + jpj2 ) jv^(p)j2 jE (t)ujs  kE (t)rsuk + Er (t) E (t)rs r u r : (2:20)
p r=0
  X
X ! 12
If we denote e 12 jraj0;1;[t0;t] by c, we can bound kE (t)rs r ukr (for r  s 1; s  ) by ccr kuks , and this
= s jpj2s jv^(p)j2 in turn by ccr kuk . Thus, multiplying (2.20) by s jE (t)ujs and summing, we obtain
s=0 p
  ! 21
X 0 s1 !2  1 21
X X  A
s jvjs : kE (t)uk  c kuk + kuk @
= 2
s=0
Er (t)ccr s
s=1 r=0
Now, if = ( 1; : : : ; s ) 2 f1; : : : ; ng , and we de ne D = D
s " X1 #
1 : : : D s , we can write
0 11  c kuk 1 +  Er (t)cr : (2.21)
r=0
@ X 2A 2
jvjs = kD v k : It remains to bound Er (t), 0  r   1. Integrating (2.6) from t0 to t, we nd
n( )=s
Zt
Applying D to E (t)u, we obtain ij (t) = [ik ( )E ( )Dkaj ( ) + E ( )Di aj ( )]d: (2:22)
t0
D E (t)u = D 1 : : : D s E (t)u
Di erentiating (2.22),
= D 1 : : : D s 1 E (t)D s u + D 1 : : : D s 1 [D s ; E (t)]u
0 0
= D E (t)D s u + D [D s ; E (t)]u; (2.15) XZt  
D ij (t) = D ;ik ( ) D ;1  E ( )Dk aj ( ) d
0
where D = D 1 : : : D s 1 . Now, for j = 1; : : : ; n, 2s0 t0
Zt
[Dj ; E (t)]u = Dj Xk (t)E (t)Dku E (t)Dj u + D E ( )Di aj ( )d: (2.23)
t0
= (Dj Xk (t) jk )E (t)Dk u
= jk (t)E (t)Dk u; (2.16) Again, we may multiply both sides of (2.23) by D ij (t), take the maximum norm in space, and sum over
j j1 = s, i; j = 1; : : : ; n, to obtain
with jk (t) = Dj Xk (t) jk , j; k = 1; : : : ; n. (Here and elsewhere in this proof we use the summation XZt
convention.) s (t)  jj1 ( )jE ( )ra( )js jj1;1 d
2s0 t0

16 17
Zt Now, rewriting (2.6) and (2.7) in terms of X instead of Y and comparing them, we nd that
+ jE ( )ra( )js;1d
t0
s sZ t
@Xk (x; t; 0) = a (x; t)D X (x; t; 0);
X @t i i k
= r r ( )jE ( )ra( )js r;1d
r=0 t0 and so we obtain
Zt u(tj) (x; t) = ai (x; t)DiXk (x; t; 0)E (t; 0)Dku(0j) (x)
+ jE ( )ra( )js;1d
Zt t0
Zt = ai (x; t)DiE (t; 0)u(0j) (x): (2.31)
 Es( )As( )d + jE ( )ra( )js;1d; (2.24) Letting j ! 1, we see that, in L1 (H# 1 (
)),
t0 t0
ut = a  ru; (2:32)
Ps s 2 2
1
where As(t) = r=0 r jE (t)ra(t)jr;1 . Now (2.24) holds for 0  s   1. For these values of s we and so u 2 W 1;1 (H# 1 (
)), and is a solution of (2.1).
deduce
! 12 To show uniqueness we prove that any solution of (2.1) must satisfy (2.10). Let  be a periodic Friedrichs'
s s
Zt X Zt molli er, constructed as follows. Let  be a real-valued function in C 1 (
0 ) (where
0 = ( ; )n) satisfying
Es (t)  r Er ( ) Ar ( )
2 2 d + As( )d Z
t0 r=0
Zt Zt
t0   0; supp  fx : jxj1  1g; dx = 1:

0
  Es( )As( ) + As ( )d: (2.25) For example, (
t0 t0 1
Gronwall's Lemma now gives (x) = Ce (1 jxj2 ) jxj < 1;
0 otherwise:
Zt R
  As ( 0 )d 0 Now de ne  by
Es (t)  As( )e t0 d  1 n  x 
t0  R t As ( )d   (x) =    ;
= 1 e t0 1 : (2.26) and periodically extend  to the whole of Rn . Then  2 C#1 (
) and satis es

But, by the inductive hypothesis,
Z
Zt   0; supp \
 fx : maxi jxi jmod2  g;  dx = 1:
As ( )d  cs jajs+1;1;[t0;t]: (2:27)

t0 Suppose that v 2 L2# (


), and write v =   v. Then v 2 C#1 (
), and v ! v in L2#(
) as  ! 0+. For
Substituting (2.27) into (2.21), we obtain Z 2
" X1 h i# jv (x) v(x)j2 =  (y)(v(x y) v(x))dy
kE (t)uk  c kuk 1 +  cr jajr+1;1;[t0;t]
cr e 1 Z

Pr=0  jv(x y) v(x)j2  (y)dy;


= c kuk e r=0   1 cr jajr+1;1;[t ;t]
0 supp
= cc kuk : so that
Z Z
The establishment of the bound on kE (t0 ; t)kL(C# (
)) follows an analogous argument to that used in the kv vk2  jv(x y) v(x)j2  (y)dydx
rst part of this proof.

supp
Z
 sup jv(x y) v(x)j2 dx;
We are now in a position to give the following existence and uniqueness theorem for (2.1). y2supp

Theorem 2.2 Let u0 2 L2# (


) and a 2 L1 (C#1 (
)n). Then there is a unique solution u of (2.1) satisfying u 2 and this tends to zero as  ! 0+ by Lebesgue's Lemma (see [55]).

C (L2#(
)) \ W 1;1 (H# 1 (
)), and given by (2.10). Moreover, if u0 2 H# (
),   0, and a 2 L1(C#[] (
)n), then Now let u be a solution of (2.1), and let u =   u. Then
1;1 
u 2 C (H# (
)) \ W (H# (
)), and satis es
 1 Ztd
E (0; t)u (x; t) u;0 (x) = E (0; s)u (x; s)ds
ku(; t)k  eC2 (0;t) ku0 k ; t 2 [0; T ]; (2:28) Z0t ds @
1 jr  aj
= E (0; s)[ @t u + a  ru ](x; s)ds: (2.33)
where C2 = C2 (0; t) = 2 0;1;[0;t] + C1 jaj[] ;1;[0;t] ,
and C1 is as in Lemma 2.2. 0

Proof Let u be de ned by (2.10). Then by Lemma 2.1, u 2 C (L2#(


)). By density, we may take a Now, the left hand side of (2.33) converges in C (L2#(
)) to
sequence of functions u(0j) in C#1 (
) converging to u0 in L2# (
). If, analogously to the de nition of u in E (0; t)u(x; t) u0 (x):
terms of u0 by (2.10), we de ne u(j) by For the right hand side we have
u(j) (x; t) = u(0j) (X (x; t; 0)) = E (t; 0)u(0j) (x); (2:29) @
@t u + a  ru =   (ut + a  ru) + [a  r;  ]u:
then we nd that u(j) ! u in C (L2#(
)) by Lemma 2.1. By Theorem 2.2, @X @t (x; t; 0) is continuous in x Since u satis es (2.1) a.e. t 2 [0; T ], the rst term on the right hand side is zero. The second term converges
and t, and so we may di erentiate (2.29) to obtain to zero in L2# (
) by Lemma 2.3 below, and so u satis es (2.10).
To complete the proof, we notice that, because of Lemma 2.2, if u0 2 H# (
) then u 2 H# (
). The
u(tj) (x; t) = @X k (j )
@t (x; t; 0)E (t; 0)Dku0 (x): (2:30) remaining assertions of the theorem now follow from (2.1) and from Lemma 2.2.

18 19
We quote the following lemma, which was used in the proof of Theorem 2.2. For the proof in the non-periodic case 2.2.2 Convergence
we refer to [55].
Here we will be mainly concerned to establish the convergence (as t ! 0 and N ! 1) of U to u in the l1 (L2#(
))-
Lemma 2.3 (Friedrichs) Let b 2 C#1 (
), u 2 L2# (
). For a commutator C which is de ned as norm, although, as above, extensions of the results to higher order norms will be indicated. At each time step, as the
projection is carried out, an error is incurred, and these errors accumulate as the calculation proceeds. Our concern,
C u = [a  r;  ]u then, will be to estimate the size of these errors, and how they are a ected by the size of the timestep. The principal
= a    ru   a  ru; result of this subsection is the following.

we have: Theorem 2.4 Let u0 2 H# (
), a 2 L1(C#[] (
)n),   0. Then
1. kC uk  C kuk, where C depends only on  ;  
2. C u ! 0 in L2# (
) as  ! 0+ . ku U kl1(L2# (
))  eC2 N  1 + min( T 12
t ; n N jaj0;1 ) ku0 k : (2:38)
We have completed our discussion of the model problem, and are now ready to describe the spectral Lagrange-
Galerkin method as it applies the problem we have been considering, and to discuss how well the approximations it
produces conform to the exact solutions of the problem. Remark. We note that the term min( Tt ; n 21 N jaj0;1 ) which appears in (2.38) is bounded by Tt min(1; n 21 N tjaj0;1 ),
thus indicating the role played by the Courant number in this estimate.
2.2 Construction of the numerical method Proof For m = 0; : : : ; M , denote u(; tm ) by um , and de ne m = (I N )um and  m = N um U m .
Then, for each m = 1; : : : ; M ,
As will become a familiar procedure in later chapters, we construct our solution by rst discretising in time (although we  m = N E (tm ; tm 1 )( m 1 + m 1 ); (2:39)
note that in this instance our time discretisation involves no approximation). For m = 0; : : : ; M , let tm := mt, where
t = T=M . Then, because of (2.10), the solution u of (2.1) satis es, for each  2 L2# (
)
1 ;tm ] (
m 1 + m 1 ):
so that
k m k  e 12 jraj0;1;[tm (2:40)
(u(; tm ); ) = (E (tm; tm 1 )u(; tm 1 ); ); m = 1; : : : ; M (2.34a)
(u(; 0); ) = (u0 ; ): (2.34b) Alternatively, we may expand N E (tm ; tm 1 )m 1 to give

This leads us to introduce the following numerical method for approximating (2.1): nd U = fU m gM N E (tm ; tm 1 )m 1 = N (E (tm ; tm 1 ) I )m 1
m=1  SN such Z tm 1 d
that = N E (tm ; t)m 1 dt
tm dt
(U m ; ) = (E (tm ; tm 1 )U m 1 ; ); 8 2 SN ; m = 1; : : : ; M (2.35a) Z tm 1
(U 0 ; ) = (u0 ; ); 8 2 SN : (2.35b) = N E (tm ; t)(a  rm 1 )dt;
tm
By taking  = p , jpj1  N , then we obtain expressions for the coecients Ud m (p) of U m , in terms of an integral
so that
involving E (tm ; tm 1 )U m 1 . Unfortunately, calculating these integrals exactly entails at least as much work as nding
the exact solution of the original di erential equation. Thus some form of numerical quadrature must be used. We  E(tm; tm 1 )m 1  Z tm E(tm; t)(a  rm 1 ) dt
will return to this subject in the next section, after having discussed the stability and convergence properties of the N
tm 1
exactly-integrated scheme (2.35). Z tm
 e 21 jraj0;1;[tm 1;tm ] m 1 a  rm 1 dt;
t
2.2.1 Stability
The only approximation involved in (2.35) is the introduction of the projections carried out in obtaining the initial datum
thus we obtain
k m k  e 21 jraj0;1;[tm 1;tm ] (  m 1 + jaj0;1;[tm 1 ;tm] rm 1 ): (2:41)
and at each time step. Since the projection operator N is non-expansive in H# (
),   0, the stability properties of
Since  0 = 0, it follows by induction (and by combining (2.40) and (2.41)) that, for m = 1; : : : ; M ,
(2.35) exactly mirror those of (2.1). This result is expressed in the following theorem, which is an immediate consequence
of Lemma 2.1. mX1 1
k m k 

e 2 jraj0;1;[tk 1 ;tk ] min( k ; jaj0;1;[tk 1 ;tk ] rk )
Theorem 2.3 Let u0 2 L2#(
) and a 2 L1 (C#1 (
)n). Then the method (2.35) is unconditionally stable in l1 (L2#(
)),
k=0
and
kU kl1(L2# (
))  e 12 jraj0;1 U 0 : (2:36) mX1
 N  e 21 jraj0;1 min(1; n 21 N jaj0;1;[tk 1 ;tk ] ) uk  ; (2.42)

Moreover, if a 2 L1 (C#[] (
)n), for some   0, then unconditional stability in H# (
) applies, and we have k=0

kU kl1 (H# (
))  eC2 U 0 H# (
) ; (2:37)
where we have made use of the estimate
k(I N )vks  N s  kvk ;   s; v 2 H# (
) (2:43)
where C2 is as in Theorem 2.2. (see, for example, [52]). The theorem now follows k  m k to both sides of (2.42), again applying
by adding
(2.43), and applying (2.2) to bound each uk  by ku0 k .
Remark. If kr  akL1 (0;1;C# (
)) < 1, then the method is unconditionally asymptotically stable in L2# (
).
An alternative analysis of the error terms may be carried out by writing, for m = 0; : : : ; M ,  m =  m + m . Then
we nd that
 m = E (tm ; tm 1 ) m 1 + (I N )E (tm ; tm 1 )U m 1 : (2:44)
The mechanism for the generation and transmission of the error terms as the calculation proceeds is perhaps clearer from
this perspective than in (2.39). We see from (2.44) that, in the mth time step, the error as it stands at tm 1 is evolved

20 21
according to the exact equation, but an additional error is incurred in projecting E (tm ; tm 1 )U m 1 . The analysis may so that PL is a projection in the sense that PL2 = PL .
be continued in a similar way to the proof of Theorem 2.4 by making use of Lemma 2.2. We may now rewrite (2.46) in a more compact operator notation as
The accumulation of errors committed at each time step (corresponding to the term Tt ) indicates that the total
error is minimised when only one time step is used. This is of course true in this instance. However, when other terms U m = N PL E (tm ; tm 1 )U m 1 ; m = 1; : : : ; M (2.47a)
(such as di usion) are included, or when nonlinear problems are considered, other error terms will appear, which can only U 0 = N PL u0 : (2.47b)
be reduced in size when the timestep t is reduced.
Here, as t is reduced, the analysis predicts that the error will increase, up to a maximum value which will be
bounded from above by 2.3.1 Accuracy
eC2 N  [1 + N jaj0;1 ] ku0 k : Later in this section our concern will be to discuss the e ect of the introduction of quadrature on the stability and
As regards the behaviour as N ! 1, we see from (2.38) that, if t ! 0 simultaneously, the convergence rate is N 1  , convergence properties established in the previous section for the exactly-integrated scheme (2.35). Here we discuss the
whereas for each xed t, the error will decrease as N  . This suboptimal convergence in the former case is consistent extra error incurred, which may be seen to be a result of the aliasing phenomenon (2.45). The e ect of aliasing has
with standard results in the semidiscrete case for the Fourier pseudospectral method for advection problems [52], [59]. been the subject of much discussion among researchers and users of spectral methods. Several ingeneous approaches
Pasciak [59] comments that since in the constant coecient case an optimal estimate holds, this result is possibly not towards controlling (or eliminating) its e ect have been employed. Now the consensus appears to be that its e ect
sharp; in practice he observes the faster convergence rate. Our numerical experiments (Chapter 3) also point to this on the accuracy of the spectral approximation is not signi cant, and that its presence will not be detrimental to the
conclusion, and we shall discuss this point further in that chapter. success of any particular computation. Theoretical justi cation for this con dence is the following approximation result,
which demonstrates that the interpolation error decays at the same rate as the projection error when the degree of the
approximating polynomial is increased.
2.3 The introduction of numerical quadrature
Lemma 2.4 Let u 2 H# (0; 2),  > 21 , and let 0    . Then
As already mentioned, the integrals appearing on the right hand side of (2.35) can be evaluated exactly only in exeptional
cases (such as when a is constant); in general it is necessary to employ some form of numerical quadrature. Here, a k(PL I )uk  C L  kuk ; (2:48)
compound trapezium rule is used with nodes x = =L,  = (1; : : : ; n), 0  j  2L 1, and then the inner product
on L2# (
) is approximated by the discrete inner product de ned, for u, v 2 C# (
), by where C = (2 (2)) 12 , and  () is Riemann's zeta function.
X
(u; v)L := (2L) n u(x )v(x ): The proof of this standard result may be found in [52]. The extension of the result to general dimensions is
 straightforward. We quote it below, and run through its proof, since the method of proof, and some of the notation
introduced in the course of the proof, will be used again in later results.
If u, v 2 SN and N < L, then
(u; v)L = (u; v): Lemma 2.5 Let u 2 H# (
),  > n2 , and let 0    . Then
By analogy with the continuous case, we de ne the discrete Fourier coecients of v 2 C# (
) by k(PL I )uk  Cn; L  kuk ; (2:49)
v^p = (v; p )L; jpj1  L; n
where Cn; = [(1 + C=n ) 1].
then we have, according to the Poisson summation formula, the aliasing relation
X Proof We must rst introduce some notation. For x 2 Rn , write
v^p = v^(p + 2kL); jpj1  L: (2:45) X
n
k 2Z n x= xi e(i) ;
i=1
We demonstrate that the results of the previous section hold when all inner products are replaced by their discrete
counterparts, as long as L is large enough in comparison to N . Thus the numerical method (2.35) is replaced by: nd where e(i) is a unit vector in the ith coordinate direction. Further, write
U = fU m gMm=1  SN such that
x?(i) = x xi e(i) :
(U m ; )L = (E (tm ; tm 1 )U m 1 ; )L; 8 2 SN ; m = 1; : : : ; M (2.46a)
(i)
(U 0 ; )L = (u0 ; )L; 8 2 SN : (2.46b) We can now de ne PL , the interpolation operator in the ith coordinate direction, by
!
By taking  = p , jpj1  N , then we obtain expressions for the coecients Ud m
pm of U , in terms of a summation X 1 1 2X
L 1
involving values of E (tm ; tm 1 )U m 1 at the quadrature points. Once fX (x ; tm ; tm 1 )g is known, these values may PL(i) u(x) =  (p ) 2 L u(x?(i) + y e(i) )pi (y ) pi (xi );
themselves be calculated using the coecients of U m 1 . jpijL i  =0
The use of the discrete inner product above is equivalent to the application of an interpolation operator to E (tm ; tm 1 )U m 1 , where y = =2L,  = 0; : : : ; 2L 1. Then the following operator identity holds:
with the interpolation points being the same as the quadrature points. (When N = L, the resulting scheme is actually
a collocation method.) Let PL denote the interpolation operator from C#(
) into SL de ned by Y
n
PL = PL(i) :
X 1 i=1
PL v = v^p p ;
jpj1 L c(p) We should like to expand PL I as a sum of products of terms such as PL(i) I , since we can apply Lemma
Q
where c(p) = ni=1 (pi ), with 2.4 to such terms. Making use of (1.12), the above operator identity may be rewritten
1
if jkj 6= L
(k) = Y
n Y
n
2 if jkj = L : PL = PL(i) = ((PL(i) I ) + I )
i=1 i=1
It is readily checked that PL interpolates at the quadrature points x (this is ensured by the term c(1p) ), and that, for XY n
v 2 C# (
), = (PL(i) I )i ;
(PL v; p )L = v^p ; 2n0 i=1

22 23
on the understanding that, if F is an operator, and (2.51) follows since, by function space interpolation,
F if i = 1
F i = ka  rV kL1 (t0;t;H# (
))  C jaj[] ;1;[t0 ;t] krV k :
I if i = 0:
Thus
XY
n
Lemma 2.6 works because of the smoothness we have imposed on a, and therefore on E (t0 ; t). PL is bounded on
PL I = (PL(i) I )i ; (2:50)
2n i=1 SN , and this smoothness ensures that the application of E (t0; t) does not take V too far away from SN , so that the
and we have, by (2.48), e ect of PL on E (t0 ; t)V is controllable. In fact we are able to bound the di erence between PL E (t0; t)V and E (t0 ; t)V
X Y by O(jt0 tj), albeit at the cost of one derivative. Stability follows from a simple corollary of Lemma 2.6 and Lemma
n 2.2, which we now state.
k(PL I )uk = (PL(i) I )i u
2n i=1 Corollary 1 Suppose that the conditions of Lemma 2.6 hold, and that L > N . Then there is a constant C4 =
X Y n
(PL(i)
 C4 (n; ; t0; jaj[] ;1 ) such that, for jt0 tj  t0 , and for V 2 SN ,
 I )i u   1+   
2n i=1  kN PL E (t0 ; t)V k  eC2 1 + C4 NL  jaj[] ;1;[t0;t] kV k ; (2:52)
X jj1  
 C=jj1 L kuk
2n where C2(t0 ; t) is as given in Theorem 2.2.
X jj1  
 C=n L kuk To ensure the stability of (2.47) via (2.52), the loss of one derivative must be countered by making L > N in such
2n
a way that as N and L are increased, 
and (2.49) then follows by applying (1.12) in reverse. N 1+  =L   C: (2:53)
[We note that the more smoothness we have on a, the closer N and L can remain.] When (2.53) holds, and taking
2.3.2 Stability analysis (t0 ; t) = (tm ; tm 1 ) at successive time steps, the stability of (2.47) is guaranteed independently of t, so that the
The above results indicate that the introduction of quadrature (and therefore aliasing) is not detrimental to the accuracy unconditional stability of the exactly-integrated scheme (2.35) is carried over to the method (2.47).
of a spectral approximation|at least not to the asymptotic rate of convergence. Its e ect on the stability (and therefore The relationship (2.53) is one between the number of quadrature points and the smallest scale being represented by
convergence) of the method as applied to a time-dependent p.d.e. is perhaps more delicate. In the exactly integrated the discrete solution. A similar relationship, between the number of particles and the smallest scale represented by the
scheme, stability was an automatic consequence of the well-posedness of the exact problem and the fact that the operator solution, is often required in the context of vortex methods (e.g. [16]).
N is non-expansive. Unfortunately, as a result of (2.45), far from being non-expansive, PL is not even bounded in Stability was proved in [79] assuming the slightly more restrictive relationship
L2#(
). (This is indicated by the restriction  > n=2 in Lemma 2.5.) Thus in this case stability will not be so
straightforward to establish. L  (1 + N n2= )N: (2:54)
Our results will be based on regarding the term N PL E (tm ; tm 1 )U m 1 as a perturbation of N E (tm ; tm 1 )U m 1 , Tadmor [81] gives a relationship of the form
and using the aliasing relation (2.45) to bound the di erence. We begin with the following lemma, which makes use of L  (1 + )N (2:55)
Lemma 2.51 .

Lemma 2.6 Let a 2 L1(C#[] (
)n),  > n2 , and let 0    . Let t0 > 0. Then there is a constant C3 = as a sucient condition for the stability of a standard semidiscrete Fourier-Galerkin approximation. In [78], (2.55) was
C3(n; ; t0 ; jaj[] ;1 ) such that for t0 ; t 2 [0; T ], with jt0 tj  t0 , and for V 2 N PL SN with L  N , used, together with the additional assumption N t  C , to establish the stability of the method (2.47). This latter
assumption could be viewed as a CFL-type restriction on the method, which is something we are trying to avoid by using
k(PL I )E (t0; t)V k  eC2 (t0;t) C3jaj[] ;1;[t0 ;t]L  krV k ; (2:51) the Lagrange-Galerkin approach. The approach we shall now employ is similar to that in [79], but with some re nement,
and we are able to prove stability subject to a relationship of the form (2.55), with no timestep limitation.
where C2 (t0 ; t) is as given in Theorem 2.2. 
Proof Since PL N PL = N PL when N  L, (PL I )V = 0, so that Lemma 2.7 (c.f. Lemma 2.6) Let a 2 L1 (C#[] (
)n), with  > n+1 2 when n > 1, and with   1 when n = 1, and
let 0    . Let t0 > 0, and let L and N be positive integers related by (2.55), for some  > 0. Then there is a
(PL I )E (t0; t)V = (PL I )(E (t0; t) I )V: constant C5 = C5(n; ; t0 ; ; jaj[] ;1 ) such that for t0 ; t 2 [0; T ], with jt0 tj  t0 , and for V 2 SN ,
Now kN (PL I )E (t0; t)V k  eC2 (t0;t) C5jaj[] ;1;[t0 ;t]N   kV k ; (2:56)
Zt d
(E (t0; t) I )V = d (E (t0 ;  )V )d where C2(t0 ; t) is as given in Theorem 2.2.
Zt0t Proof In order to communicate e ectively the main ideas in the proof, we demonstrate the result rst in
= E (t0 ;  )(a  rV )d (by the chain rule): one dimension, for the case  = 1,  = 0. The proof involves temporarily expanding the left hand side in
t0
terms of the Fourier coecients of V , and exploiting the nature of the Fourier basis functions p .
Application of Lemma 2.5 and Lemma 2.2 now gives From the aliasing relation (2.45) we can write
k(PL I )E (t0; t)V k X 1 X
Z t N (PL I )E (t0; t)V = (E (t0; t)V; q+2kL )q :
 Cn; L  E (t0 ;  )(a  rV )d jqjN c(q) k6=0
t0 
 Cn; L  eC2 (t0 ;t) ka  rV kL1(t0 ;t;H# (
)) ; The right hand side can be expanded further to give
X 1 X ^ X
1
We note that (2.45) plays a key role in the proof of Lemma 2.4, and therefore also in that of Lemma 2.5, of which the rst N (PL I )E (t0; t)V = V (p) (E (t0; t)p ; q+2kL )q :
stability result is an almost direct application. jqjN c(q) jpjN k6=0

24 25
Now With L and N related by (2.55),
E (t0 ; t)p(x) = p (X (x; t0; t)) = p (X (x; t0; t) x)p(x): X X
1
max jq + 2kLj 2(r+1)  2 j2kL Lj 2(r+1)
Thus if we drop the explicit dependence on t0 and t and de ne q
k6=0 k=1
E p (x) = p (X (x; t0; t) x); 2(r+1)
X
1
 2[(1 + )N ] (2k 1) 2(r+1)
then E (t0 ; t)p = (E p)p, and, taking the bounds jpj; jqj  N and jqj  LL to be understood, k=1
2(r+1)  (2);
XX 1 ^ X  2[(1 + )N ] (2.62)
N (PL I )E (t0; t)V =
pc(q) V (p) (E p; q
q k6=0
p+2kL )q : (2:57)
where  is Riemann's zeta function. Moreover,
EDr+1V 2  e2C2(t0;t) X jqj2(r+1) jV^ (q)j2
Now, we may integrate by parts the term (E p; q p+2kL ), giving
q
(E p; q p+2kL ) = (D(i(Eqp)p; +q 2kL
p+2kL ) :
) (2:58)  e2C2 (t0;t) N 2r kDV k2 : (2.63)
Together, (2.59), (2.61), (2.62) and (2.63) give that
The term appearing in the integrand on the right hand side of (2.58) may be expanded in the following way:
d  kAr k  (2 (2)) 21 (ejaj1;1;[t0;t] 1)eC2 (t0;t)(1 + ) (r+1) N 1 kV k
1;
D(E p(x)) = dx (X (x; t0; t) x) (ip)E p(x): so that 1
kN (PL I )E (t0; t)V k  (2 (2))
2 jaj1;1;[t ;t]
In order to ease presentation, we write dxd (X (x; t0; t) x) = B (x); then in notation drawn from the proof  (e
0 1)eC2 (t0;t) N 1 kV k1 : (2:64)
of Lemma 2.2 we have that kB kC#1 (
) = E0(t), and (2.26) provides the bound In this way the proof for the one-dimensional case with  = 1 and  = 0 is completed. The case  > 0 (in
general dimensions) may be reduced to the case  = 0 by noting that the left hand side of (2.56) is a norm
E0 (t)  ejaj1;1;[t0;t] 1: (2:59) of a function in SN , and for v 2 SN we have the inverse inequality
The denominator of (2.58) may be rewritten in a binomial expansion kvk  (1 + nN 2 )=2 kvk :
1 X1 (ip)r The extension to higher dimensions is somewhat of a notational headache, although there is otherwise no
i(q p + 2kL) = r=0 [i(q + 2kL)]r+1 ; particular technical diculty. The basic idea is the following. We expand PL I by the formula (2.50),
and note that, since  = 0 is excluded, each term in the summation involves the application of PL(i) I for
so that (2.58) becomes at least one i. The result is then obtained by applying the one-dimensional version of Lemma 2.7 in that
direction, and by applying Lemma 2.5 in the other directions for which i = 1.
(E p ; q p+2kL ) = (D(i(Eqp)p; +q 2kL
p+2kL )
)
By rearranging (2.50) we can write, for n > 1,
2 3
X1 X
n
4X Y
j 1
(Ni) (PL(i) I )i 5 (Nj) (PL(j) I );
(ip )r+1
N (PL I ) = (1) (1)
N (PL I) +
= r+1 (B E p ; q p+2kL )
r=0 [i(q + 2kL)] j=2 2j0 1 i=1
X (ip)r+1
1
= r+1 (BE p ; q+2kL ): where (Ni) is projection with respect to the ith coordinate direction only, and use is made of the fact that
r=0 [i(q + 2kL)] (Ni) commutes with PL(j) for i 6= j . We set Wj = (Nj) (PL(j) I )E (t0; t)V , j = 1; : : : ; n. Then Lemma 2.5
Thus, substituting this back into (2.57) we have implies that, so long as  1 > (n 1)=2,
X
1 X X 1 (BEDr+1V;  X
n
N (PL I )E (t0; t)V = q+2kL )  :
q (2:60) kN (PL I )E (t0; t)V k  kW1 k + Cj 1; L ( 1) kW
j k 1 : (2:65)
r=0 q k6=0 c(q) [i(q + 2kL)]r+1 j=2
P1 A , and proceed to bound the terms A in turn. We
We shall write the right hand side of (2.60) as
Thus we seek to bound terms of the form kWj k , for  = 0 and  =  1. We shall investigate the case
r=0 r r  integer|the corresponding result for non-integer  being obtained thence by interpolation.
have
0 1 We may expand Wj to give
X @ 1 X j(BEDr+1V; q+2kL)j A2 Wj = (Nj) (PL(j) I )E (t0; t)V
kAr k2  jq + 2kLjr+1 X 1 X ^ X
q c(q) k6=0
0 10 1 = (q ) V (p) (E p; qj +2kj L )j qj ; (2.66)
X@ 1 X r+1V; q+2kL )j2 A @ 1 X jq + 2kLj 2(r+1) A
jqj jN j jpj1N kj 6=0
 j(BED
q c(q) k6=0 c(q) k6=0 where (; )j denotes an integral carried out in the j th coordinate direction only, and where qj is a function
0 1 of xj only. The term involving the inner product may be rewritten
X 00 1 1
 kB k2C# (
) EDr+1V 2 max
q
@ jq + 2 kLj 2(r+1) A : (2.61) Y A
k6=0 (E p ; qj +2kj L)j = @@ Epi E pj ; qj pj +2kj L A
i6=j j

26 27
 Q  
Dj i6=j Epi E pj ; qj pj +2kj L j Theorem 2.6 (Convergence) Suppose that the conditions of Lemma 2.7 hold, and that u0 2 H# (
). Suppose also
= that t  t0 . Then
i(qj pj + 2kj L)
X
n ( E  ;  kj L)j ku U kl1 (L2#(
)) 
= ipl ijl(q pp +qj +2 ;  
j 2kj L)
l=1 j
eC2 +C5 jaj[];1 N  ku0 k 1 + Cn; + min( Tt ; n 21 N jaj0;1 ) + C5jaj[] ;1 ; (2.70)
where jl denotes Dj Xl jl as in the proof of Lemma 2.2. Now let be a multiindex of order  , and
let be another multiindex of order  1, which is such that i = i when i 6= j . After some integration where C2 and C5 are as in Lemma 2.7.
by parts in the inner product, and expanding 1=(qj pj + 2kj L) as in the one-dimensional case, we obtain
Proof For m = 0; : : : ; M , we denote u(; tm ) by um , and we de ne m = (I N )um and  m =
1 XX
X n X 1 N um U m . Then, for each m = 1; : : : ; M ,
D Wj = r
(qj ) (D (jl EDj Dl V ); qj +2kj L)j 
r=0 qj l=1 kj 6=0  m = N PL E (tm ; tm 1 ) m 1 + N E (tm ; tm 1 )m 1
j N (PL I )E (tm; tm 1 )N um 1 :
 [i(q + (2iqkj )L)] j +r+1 qj : (2.67) (2.71)
j j
As in Theorem 2.5,
The remaining steps are now very similar to the one-dimensional case. We write the right hand side as a  P E(tm; tm 1)m 1
sum of terms Ar , and bound each one as follows. N L
0n  e 12 jraj0;1;[tm 1 ;tm ] +C5 jaj[] ;1;[tm 1 ;tm ]
m 1 ; (2.72)
X @X X 1
kAr k 
2
j(D (jl EDjr Dl V ); qj +2kj L )j j and as in Theorem 2.4,
qj l=1 kj 6=0 (qj )
j 2  E(tm; tm 1)m 1
 (q +q2j k L) jqj + 2kj Lj  r N
um 1 :
j j  N  e 21 jraj0;1;[tm 1 ;tm ] min(1; n 21 N jaj
0;1;[tm 1 ;tm ] )  (2.73)
Xn
D (jlEDjr DlV ) 2 max X
 qj
jqj + 2kj Lj2(  r) It remains to bound the third term on the right hand side of (2.71). Lemma 2.7 provides
l=1

kj 6=0
 (P I )E(tm; tm 1 ) um 1
 2 (2)L2(  r) D ( EDr D V ) 2 N L N
1 ;tm ] C5 jaj  m 1 m N 
jl
j l
 e 21 jraj0;1;[tm [] ;1;[t ;t ] um 1  : (2.74)
 2 (2)L2(  r) E2 1 (t)e2C2 (t0;t) N 2r kV k2 ; (2.68)
where E 1 (t) is as in the proof of Lemma 2.2, and may be bounded in terms of jaj[] ;1;[t0 ;t] by use of The inequalities (2.72), (2.73) and (2.74) may now be combined with (2.71) to give
(2.26) and (2.27). The proof may now be completed by summing over r and over all multiindices of k m k  e 12 jraj0;1;[tm 1 ;tm ] +C5 jaj[] ;1;[tm 1 ;tm ]
m 1
order  . 1 
+N  um 1  e 2 jraj0;1;[tm 1;tm ] C5 jaj[] ;1;[tm 1 ;tm]
We now end our discussion of stability with the following theorem. 12 i
+ min(1; n N jaj0;1;[tm 1 ;tm] )
Theorem 2.5 (Stability) Suppose that the conditions of Lemma 2.7 hold, and that t  t0 . Then the method
(2.47) is unconditionally stable, and for any 0    , and this may be used to show inductively that

kU kl1(H# (
))  eC2 (0;T )+C5 jaj[];1 U 0  
(2:69) k m k  eC2 +C5 jaj[];1   0

+N  ku0 k min( Tt ; n 21 N jaj0;1 ) + C5jaj[] ;1 :
Proof The proof is by induction. The inductive step is provided by
The result now follows by adding km k to sides, applying (2.43), applying (2.2) to bound kumk by
both
kU m k  N E (tm 1 ; tm )U m 1  + N (PL I )E (tm 1; tm )U m 1  : ku0 k , and by including an estimate for  0 provided by Lemma 2.5.
Use of Lemma 2.2 and Lemma 2.7 now gives The theoretical estimates we have given in this chapter indicate that the spectral Lagrange-Galerkin method is
unconditionally stable and retains the accuracy of the underlying spectral approximation, even when quadrature is
kU m k C2 (tm 1 ;tm)+C5 jaj[];1;[tm 1 ;tm ] U m 1 introduced. In the next chapter we concern ourselves with the implementation of the method, rst describing an
e ; additional approximation that must be made in order for the method to be implemented eciently, and then illustrating
as is required. the theory with various numerical experiments.

2.3.3 Convergence
Our goal here is to give an analogous convergence result to Theorem 2.4, showing that the method (2.47) enjoys similar
spectral accuracy to (2.35), and that this is irrespective of the size of the timestep under the condition on L and N
given by (2.55). Our result is the following.

28 29
(0; 2) (2; 2)

Chapter 3

An ecient implementation
As is common when discussing spectral methods, we are concerned with the eciency of the scheme, because the
global nature of the basis functions can mean that the evaluation of derivatives or of inner products of functions is very
expensive. Such operations usually involve the use of a discrete Fourier transform. Using a Fast Fourier Transform (FFT),
this may be performed eciently (by which is usually meant in O(N n log N ) operations, where n is the dimension of (0,0) (2; 0)

the problem). Without the use of the FFT, the operations count is still only O(N n+1), because of the tensor-product
nature of the basis functions.
The timestepping procedure central to the spectral Lagrange-Galerkin method (e.g. (2.47)) involves, at each time Figure 3.1: A typical distribution of points X (marked ) in
, together with the grid points x (marked by
step, the evaluation of E (tm ; tm 1 )U m 1 (x ) = U m 1 (X ) at the (2L)n quadrature points x , where U m 1 2 SN ) and the sets B (denoted by ) surrounding the grid points.
and X = X (x ; tm ; tm 1 ). Unfortunately, the points X are not in general evenly spaced, and so the FFT cannot be
used in these evaluations. Also, when the evolution operator E (tm ; tm 1 ) is applied to the basis functions, they lose
their tensor-product properties, so that in general, the operation count for the evaluation of E (tm ; tm 1 )U m 1 (x ) is
in fact O(N 2n). This is a severe disadvantage, especially in more than one dimension. We de ne, for  2 2L 1 ,
B = fx 2 Rn jd(x; x )  bg;
3.1 Local interpolation and
B = [ 2 2L 1 B ;
In this section we describe and analyse an interpolation process which is designed to enable the ecient and accurate
evaluation of a trigonometric polynomial at a randomly distributed set of points. The method involves rst setting so that X  B. Thus, for our interpolation procedure to yield an accurate approximation to fV (X )g 2 2L 1 , we are
down an equally-spaced grid of points and, in regions surrounding those points, replacing the trigonometric polynomial mainly concerned that it should be accurate in B. Our strategy is to replace V in each component B by a nite sum of
by a local Chebyshev polynomial interpolant, with the degree of the interpolating polynomial identical in each region. (tensor product) Chebyshev polynomials. This is achieved by laying down an auxiliary grid of interpolation points on B,
The error incurred by the use of this local interpolation is shown to decay faster than factorially with the degree of identical in each component B . Once the coecients of the polynomials in each component are known, the value of V
the interpolating polynomial. Thus, when the process is applied to U m 1 at each time step, it gives rise to a method at any point in X may be calculated from the polynomial in the component of B in which it lies. In practice we exploit
which retains the stability and convergence properties described in the previous section. We also show that, especially the fact that the sets B are centered around the points of G (which are equally-spaced) to calculate the coecients
for problems where a high degree of resolution is required, this modi cation leads to substantial savings in time with of the polynomials in each of the sets B simultaneously. This computation takes advantage of the FFT and may be
e ectively no loss of accuracy. performed in O(Ln log L) operations.
Let V 2 SN . Then V may be expanded in the form In order to meaningfully speak of a polynomial de ned on B we shall have to identify B with a particular subset
of Rn which shall be the canonical representation of B . As indicated in Figure 3.1, this shall be the set of points in Rn
X^ whose distance in the ordinary l1 -norm from the grid point x 2
is not greater than b.
V= V (p)p;
p2 Z n Let K be a positive integer (K will be the degree of the local interpolating polynomials). Let yj = cos j K,
j = 0; : : : ; K . We de ne, for p = N; : : : ; N and 0 < k  K ,
where V^ (p) = 0 if jpj1 > N . Let 2L 1 = f0; : : : ; 2L 1gn , and let G = fx g 2 2L 1 be an equally-spaced grid of X
K
points indexed by 2L 1 , where x = (x11 ; : : : ; xnn ), and xii = 2iL , i = 1; : : : ; n. Let X = fX g 2 2L 1 be some a^p;k = K2 00 p (byj )Tk (yj );
other set of points (also indexed by 2L 1 ) which we may take to be a subset of
by the periodicity of V . A typical j=0
distribution of these two sets of points in two dimensions is given in Figure 3.1.
Let Rn2 be the vector space of equivalence classes of points in Rn , where two such points are regarded as equivalent where Tk (y) = cos(k cos 1 y) (y 2 [ 1; 1]) is the kth Chebyshev polynomial on [ 1; 1], and the double prime denotes
if they are separated by an integer multiple of 2 in each coordinate direction. In the sequel we shall speak freely of the fact that the rst and last terms in the summation are halved; for k = 0, a^p;0 is de ned by
points in Rn (and often
), but we shall regard them as representatives of their particular equivalence classes. We can
de ne a metric d on Rn2 by X
K
d(x; y) = max jxi yi jmod2 : (3:1) a^p;0 = K1 00 p(byj );
i j=0
Then, for x 2
we denote by bxc the nearest grid point to x with respect to d. Let
and for k = K , a^p;K is de ned by
b =  2max d(bX c; X ): (3:2) X
K
2L 1 a^p;K = K1 00 ( 1)j p(byj ):
j=0
We note that (3.2) implies that b  2L .

30 31
With these de nitions,
PK a^ T ( y ) interpolates  (y) at the points fby gK in [ b; b]. For x 2
de ne interpolation has been carried out in some coordinate directions but not in others. We shall describe the one-dimensional
k=0 p;k k b p j j=0
( PK case rst.
a^ T ( x i bxci
) (bxc ) x 2 B
i For N  p  N , let p = ~p p. Let  2 2L 1 = f0; : : : ; 2L 1g, and consider p (x) for x 2 B . We can
~p(x ) =
i k=0 p;k k
p (xi )
b p
otherwise; write such an x as x + by for some y 2 [ 1; 1]. Then

~p(xi )Qagrees with p(xi ) when xP


i = xi + byj , for any  2 2L 1 and any j = 0; : : : ; K . Then, for p 2 Z n we put X
K
 p (x) = p (x )( a^p;k Tk (y) p (by)):
~ p = ni=1 ~pi and we nd that p2Z n V^ (p)~ p interpolates V on the set [ 2 2L 1 [2 K fx + (by1 ; : : : ; byn )g k=0
(the auxiliary grid). We de ne the local interpolation operator PK;L (acting on functions in SN ) by
When x is not in B, p (x) = 0, so that, from the standard error estimate for Lagrange interpolation,
X^ ~
PK;L V = V (p)p (3:3) X
K
a^p;k Tk (y) p(by)j  2Kjpb j K +1
p2 Z n jp j1 = y2max
[ 1;1]
j (K + 1)! : (3:6)
k=0
3.2 The algorithm Now for V 2 SN we can write X^
We want to evaluate PK;L V at the points X of X . Expanding PK;L V (X ) we have (PK;L I )V = V (p)p ;
p2 Z n
PK;L V (X ) so that (3.6) implies that
X^ XY n   i i 
a^pi;ki Tki X bbX c p (bX c) K +1 X
=
p2 Z n
V (p)
k2 K i=1
k(PK;L I )V k1  2K (bK + 1)! jV^ (p)jjpjK +1
X "Y n  i #2
i X Yn
3 p2 Z n
K +1 p N )
= Tki X bX c 4 b V^ (p) a^pi;ki p (bX c)5 : (3.4)  jV jK +1 b2K (K +(21)! : (3.7)
k2 K i=1 p2 Z n i=1
So far this concerns the local interpolation, without reference to the set of points at which V is to be evaluated. For
(3.4) indicates how the calculation may be performed inexpensively. For each k 2 K we have a term of the form any particular set of points fX g, we shall be able to say more about b. For instance, we note that here b satis es
X^
V (p)fk (X bX c)gk (p)p(bX c); b  minf 2L ; jaj0;1;[tm 1 ;tm] g (3.8)
p
and the algorithm is as follows:
 minf 2L ; jaj0;1 tg; (3.9)
 multiply the V^ (p) by the gk (p), so that we have
 evaluate W (x ) = Pp V^ (p)g(p)p(x ) by the use of a FFT, P E(tm; tm 1 )(P
L

K;L I )V
 multiply W (bX c) by fk (X bX c). ! 21
= 1 2XL 1
j(PK;L I )V (X )j2
The total number of operations for one loop is O(Ln log L) (assuming that L  N ), so that the operation count for 2L  =0
evaluating fPK;L V (X )g 2 2L 1 is now O(K nLn log L). We shall show, both by means of theoretical estimates and
numerical experiments, that this represents a signi cant saving over the cost of a full evaluation (O(L2n)).  max j(P
 2 2L 1 K;L
I )V (X )j
We note here that the local interpolation described above is not the only means of obtaining an approximation to  K p
fV (X )g 2 2L 1 that may
P be implemented by the use of this algorithm. The relevant property of the approximation  C6 jaj0;1;[tm 1 ;tm] jV jK +1 4L (K(2+N1)!) (3.10)
of p (X bX c) by k fk (X bX c)gk (p) is that the dependences on p and on  have been decoupled. Any
approximation with this property will provide an ecient procedure, but one based on the use of Chebyshev polynomials In the one-dimensional case, the constant C6 appearing in (3.10) may be taken to be unity. In general dimensions,
will in general be amongst the most accurate [34]. however, it depends on C5 , jaj[] ;1 , n and t0 , where C5 is as in Lemma 2.7 and t0 is an upper bound on t. In
The incorporation of this local interpolation into the timestepping procedure of the previous chapter results in the order to obtain a stability result analogous to Theorem 2.5 we let 0    min(K + 1; ) and we have
following numerical method. Find U = fU m gM m=1  SN such that  P E(tm; tm 1 )(P
N L K;L I )V   (1 + nN 2 )=2 N PL E (tm ; tm 1 )(PK;L I )V
U m = N PL E (tm ; tm 1 )PK;L U m 1 ; m = 1; : : : ; M (3.5a)  N K N 32
U 0 = N PL u0 : (3.5b)  C7jaj0;1;[tm 1 ;tm] kV k 4L (K + 1)! ; (3.11)
Here, the points X in X used to de ne PK;L are given by X = X (x ; tm ; tm 1 ), with  2 2L 1 . Calculation of U m where C7 is just a multiple of C6 .
thus involves evaluating PK;L U m 1 on the set fX g 2 2L 1 . These X are not randomly distributed and in fact, for We seek to bound the right hand sides of (3.11) independently of K or N , and so we require that K and N are
small Courant numbers (jaj0;1Lt), they will be grouped around their respective grid points. Also, when a is smooth, related by, for example,
there will be a corresponding smooth variation between neighbouring points in X . In the next section we shall give an  4L K
error analysis for the local interpolation as it used in (3.5) which exploits these observations. N 23  N (K + 1)!: (3:12)
In order to illustrate the relationship between N and K that (3.12) implies, we consider sample values of K and N
3.3 Error analysis satisfying (3.12), shown in Table 3.1. Since the saving in time resulting from the use of the procedures described in this
section is of the order of N=K , we see that this can be quite substantial. The numerical experiments in the next section
The analysis rests on the standard error estimate for Lagrange interpolation of a smooth function. Most of the subsequent will show how substantial these savings are in practice. It is perhaps worth remarking that a relationship of the form
complication arises when (in more than one dimension) (3.4) is expanded further, resulting in terms where the local (3.12) implies that K = O(log N ) so that the asymptotic operations count is O(Ln(log L)n+1).

32 33
then, for V 2 SN , (PK;L I )V may be expanded (c.w. (2.50))
K N 0 1
1 1 X X@ X ^
2 4 (PK;L I )V = V (p + q)p A q : (3:17)
3 13 2n q2Zn p2Z1n 
4 46
5 179 We further de ne X ^
6 772 V;q = V (p + q)p;
7 3630 p2Z1n 
8 18451 and note that V;q 2 SN . Then
9 100608 P E(tm; tm 1 )(P
10 584571 L K;L I )V
Table 3.1: Sample values of K and N . X X
m
= PL E (t ; t )m 1 V;q q
n
X X 2m mq2Z1
n
 PL E (t ; t )V;q q


Making use of (3.11) we can ensure the stability of (3.5) (under the conditions of Theorem 2.5, together with (3.12))
2n q2Zn
as follows: X X

kU m k = N PL E (tm ; tm 1 )PK;L U m 1 
 kq kC# (
)n PL E (tm ; tm 1 )V;q : (3.18)
2n q2Zn
 N PL E (tm ; tm 1 )U m 1  + N PL E (tm ; tm 1 )(PK;L I )U m 1 

 eC2 (tm 1 ;tm)+C5 jaj[];1;[tm 1 ;tm ] U m 1 
Under the conditions of the theorem,
P E(tm; tm 1)V  (1 + C jaj  m 1 ;tm] ) kV;q k : (3:19)
+C7 jaj0;1;[tm 1 ;tm] U  :
m 1 L ;q 5 [] ;1;[t

As in the proof of Theorem 2.5 we deduce that


Moreover X ^
kV;q k =
2
jV (p + q)j ;
2 (3:20)

kU m k  e(C2 +C5 jaj[];1 +C7 jaj0;1) U 0  ; 0  m  M: (3:13) p2Z1n 
and, for q 2 Zn ,
Convergence also follows, analogously to Theorem 2.6. We nd that (2.71) is replaced by n  jq bjK +1 i
Y
kq kC# (
)n  i
 m = N PL E (tm ; tm 1 )PK;L  m 1 + N E (tm ; tm 1 )m 1 i=1 2K (K + 1)! : (3:21)
N (PL I )E (tm; tm 1 )N um 1 By the Cauchy-Schwarz inequality, (3.20) and (3.21) imply that
N PL E (tm ; tm 1 )(PK;L I )N um 1 : (3.14) X
Making use of the above stability result, the rst three terms are dealt with in the same way as the corresponding terms
kq kC# (
)n kV;q k
q2Zn
in (2.71). By (3.10) and (3.12) we may bound the nal term by
 bK+1p(2N ) jj1 0 X X Y n
1 12
C6 jaj0;1;[tm 1;tm] N (K +1) jum 1 jK +1 :  @ jq i j2(K +1)i jV^ (p + q )j2 A
2K (K + 1)! p2Z1n  q2Zn i=1
As in the proof of Theorem 2.6, this, when included in with the other terms, yields the following convergence result.
 bK+1p(2N ) jj1
ku U kl1(L2# (
))   N (K +1)(jj1 1) jV jK +1
2K (K + 1)!
eC2 +C5 jaj[];1 +C7jaj0;1 N 0 ku0 k
  (Nb)K+1p(2N ) jj1 jV j
  K +1
 1 + Cn; + min( Tt ; n 21 N jaj0;1 ) + C5jaj[] ;1 + C6 jaj0;1 ; (3.15)
= 2K (K + 1)! N K +1 :
Then, from (3.18) and (3.19),
where 0 = minf1 + K; g.  P E(tm; tm 1)(P
We have now proved, for the one-dimensional case, the following stability and convergence result for the method N L K;L I )V
(3.5).
jV j  p  
)K +1 (2N ) n 1
Theorem 3.1 Suppose that the conditions of Theorem 2.5 and Theorem 2.6 hold, together with the condition (3.12).  N KK+1
+1
(1 + C5jaj[] ;1 ) 1 + (NbK
2 (K + 1)!
Then (3.13) and (3.15) hold.
and (3.10) follows, with (3.11) a straightforward consequence.
Proof All that remains to be proved is that (3.10) and (3.11) hold in general dimensions. We de ne, for
p 2 Z n , x 2 Rn , Y
p (x) = pi (xi ): 3.4 Numerical experiments
i:pi 6=0
Let  2 n . We note that Z n = Zn  Z1n  where Zn is as de ned by In this section we perform various numerical experiments in order to illustrate the theoretical predictions of the previous
two chapters. We begin by considering the performance of the method on two one-dimensional advection equations.
Zn = fk 2 Z n jki = 0 , i = 0; i = 1; : : : ; ng; (3:16)

34 35
K N = 3, L = 4 N = 7, L = 8 N = 15, L = 16 N = 31, L = 32
4 1:880  10 1 4:457  10 5 8:560  10 7 4:597  10 8
10
-1 -1
10 N=8 6 1:882  10 1 3:209  10 5 8:620  10 10 8:672  10 12
N=16 8 1:882  10 1 3:209  10 5 4:277  10 13 1:629  10 15
N=32 10 1:882  10 1 3:209  10 5 1:100  10 15 3:953  10 16
0.03 0.03
N=64
Table 3.2: l1 (L2 ) errors for Experiment 2.
Error

-2 -2
N=128
10 10

0.003 0.003

-3 -3
10 10

1 2 3 4
10 10 10 10
No. of timesteps

Figure 3.2: Results from Experiment 1.

3.4.1 Experiment 1
In this section our model problem is the following.
ut (sin x)ux = 0; x 2 [0; 2); t 2 (0; 1:571]; Z
Y X
with initial condition
u(x; 0) = f (x);
for which the exact solution is  
u(x; t) = f 2 tan 1 (et tan x2 ) : Figure 3.3: Initial data for the rotating cone problem (p = 1).
The advection is `pushing' the solution out towards the boundaries of the domain, and so in general the solution tends
to develop sharp layers at x = 0 and x = 2. where SN is the space in which the problem is being approximated. In order to demonstrate his theoretical predictions,
We consider the case  he integrates the above problem up to time 50:27. To achieve machine accuracy ( 10 13 ) he requires 840 polynomials,
sin x 0  x  
f (x) = with N =16.
0   x  2: For comparison we also integrated the above problem up to time 50:27 ( 16). Because of the fact that the exact
This enables us to illustrate the behaviour of the method (2.35) as t is reduced when the initial data and the velocity solution is 2-periodic in space and in time, integration to this nal time in just one time step using the spectral method
eld a have only nite smoothness. Figure 3.2 shows the L2 error for various values of N and t when (2.35) was used of characteristics (with the additional interpolation) is almost a trivial problem. We therefore used three time steps. The
to integrate the equation up to time t = 3:14. (In fact we had to use the scheme (2.47), but we took L = 2048 to feet of the characteristics were found to machine accuracy by repeated bisection. The results obtained for various values
e ectively eliminate the quadrature error). of K , N and L are shown in Table 3.2. We note that for N = 15 and L = 16, the l1 (L2 ) error in the computed
As predicted by Theorem 2.4, for each xed N the error increases up to a maximum value as more time steps are solution was down to  10 15 with the use of the additional interpolation PK;L with K =10, so that the amount of work
taken. In Figure 3.2 this increase does not show itself until N t  1:5. Until that point it is overshadowed by the required was a fraction of that for the time-integration scheme of Tal-Ezer. The results also illustrate the high accuracy
contribution to the error of the projection of the initial data. Although this agrees with Theorem 2.4 in a qualitative sense, of the spectral approximation for such smooth problems.
the predicted reduction in the order of convergence for the time-converged solution is not evident in this experiment. A
more careful analysis might be able to reveal the balance between the various contributions to the error more clearly.

3.4.2 Experiment 2
Here we consider the scalar hyperbolic equation
ut (2 + cos x) 1 ux = 0; x 2 [0; 2); t 2 (0; 50:27];
subject to periodic boundary conditions and the initial condition
3.4.3 Experiment 3
This test problem, known as the rotating cone problem, is the one in which Priestley (see [76]) rst observed the instability
u(x; 0) = sin(2x + sin x): to which the nite-element version of the method is prey when quadrature is introduced. For our purposes it also serves
to illustrate the theoretically predicted convergence rates for initial datum that is not in nitely smooth, and to illustrate
The exact solution is the savings arising from use of the approximate evolution operator.
u(x; t) = sin(2x + sin x + t): The governing equation is
Tal-Ezer has also studied this problem [83]. He uses a spectral method in time, based on a Chebyshev polynomial ut + a(x)  ru = 0; x 2
0 ; t 2 (0; 20];
expansion of the evolution operator. His theory predicts that in order to e ectively resolve the behavior of the solution
in time, the degree of the polynomial needs to be proportional to both the time over which he is integrating, and to N ,

36 37
p N=L Umax Umin l1 (L2) error order
7=8 0:9996 2:1  10 2 5:112  10 3
1 15=16 1:0008 4:9  10 3 8:172  10 4 2:65
31=32 1:0002 1:4  10 3 1:448  10 4 2:50
63=64 1:00005 4:4  10 4 2:660  10 5 2:44
7=8 0:8676 6:0  10 2 1:432  10 2
2 15=16 0:9998 7:6  10 4 1:385  10 4 6:69
31=32 :99999 3:7  10 5 5:185  10 6 4:74
Y
Z
X
63=64 :999999 2:5  10 6 2:301  10 7 4:49
Table 3.4: Rotating Cone Problem

Figure 3.4: Initial data for the rotating cone problem (p = 2).
method N=L Umax Umin l1 (L2) error
exactly-integrated 0:9917 4:112  10 2 6:840  10 3
quadrature-based 7=8 0:9847 2:802  10 2 6:827  10 3
approx. (K = 5) 0:9853 2:829  10 2 6:842  10 3
exactly-integrated 1:0001 6:544  10 3 9:598  10 4
quadrature-based 15=16 0:9991 4:138  10 3 7:468  10 4
approx. (K = 5) 0:9992 5:145  10 3 9:081  10 4
Table 3.3: Rotating Cone Problem (p = 1)

subject to periodic boundary conditions and the initial condition


8 cos2p(2jx x0j); jx x0j  =4;
< 10
-2 t=1.E0 10
-2

u(x; 0) = : 10
-3
t=1.25E-2 -3
10
0; otherwise, 10
-4
t=1.25E-3 -4
10

where p 2 N , a(x) = 2(x2 ; x1 ), x0 = (0; =2) and


0 = ( ; )2 . The initial datum is shown, for p = 1 and p = 2,
-5 -5
10 10
-6
in Figs. 3.3 and 3.4. Again the characteristics may be calculated explicitly. Moreover, because of the linearity of the
-6 10
10
-7
velocity eld it is possible to perform the inner products involved in the scheme (2.35) exactly. Thus we can observe -7

Error
10 10
the e ect of quadrature. The solutions obtained were visibly indistinguishable from the initial data, so we illustrate the -8 -8
10 10
performance of the schemes by tabulating the errors. In Table 3.3, the results show the l1 (L2 ) error after ve revolutions 10
-9 -9
10
(t = 5) with 250 time steps, for the exactly integrated scheme (2.35), the quadrature-based scheme (2.46), and for the -10
10
-10
10
scheme (3.5) involving the additional interpolation. The value of K given here is the minimum required to reproduce -11
10
-11
10
the results obtained by the scheme (2.46). -12 -12
10
The parameter p in the description of the problem enables us to control the smoothness of the initial datum. We
10
-13

nd that u0 2 H#2p+1=2  (
0 ) for any  > 0. Thus we can check the theoretically predicted orders of convergence for
-13 10
10
-14 -14

the method. Results for p = 1 and p = 2 are given in Table 3.4 for the method (2.46), together with the orders of 10 10
2 3 4 5 6 7 8 9 10
convergence represented by the results. 25 time steps were performed to reach a nal time of t = 0:5. The results for K
N=L = 63=64 were obtained with the approximate method (3.5) with K = 5. Figure 3.5: Results from Experiment 4.
These calculations were performed on a Convex C1/XP, which is a vector machine. On this machine the scheme
(3.5) was approximately seven times as fast per time step as the scheme (2.46) for N = 63, L = 64 and K = 5. On
the VAX 11/785, a scalar machine, this di erence is increased from a factor of seven to a factor of fty. The reason for
this disparity is that the FFT routine used in the approximate trajectories scheme does not take full advantage of the
vectorizing capabilities of the Convex machine.

38 39
K c.p.u. time speedup values of k that managed to sneak through the rst test. This then is the procedure we have used in practice: having
1 2.0s 630 assigned a tolerance at the beginning of the run, we rst discard those values of k for which (3.23) exceeds that tolerance,
2 3.4s 370 and then check the remaining values according to (3.22).
3 5.2s 242 In Figures 3.6 and 3.7 the shaded areas represent those values of k which were actually used in each particular
4 7.5s 168 calculation. In each table the tolerance parameter is shown, together with the resulting l1 error. The results indicate
5 10.3s 122 that the tolerance parameter is a fairly reliable guide to the size of the resulting error. Table 3.5 shows how the c.p.u.
6 13.4s 94 time decreases with K ; correspondingly, the shaded areas in the diagrams in Figs. 3.6 and 3.7 directly relate to the c.p.u.
7 17.1s 73 time taken by the respective calculations.
8 21.1s 59
9 25.7s 49
10 30.0s 42
Table 3.5: Example c.p.u. times for Experiment 4.

3.4.4 Experiment 4
Our nal experiment in this chapter is a more thorough investigation of the behaviour of the additional interpolation
than was a orded by the rotating cone problem. We shall consider the problem of the evaluation of a two-dimensional
trigonometric polynomial at set of points randomly distributed in (0; 2)2. The polynomial U will be a function from SN ,
with the value of N being 32, and we take 642 points. We choose U to be the projection of the vorticity eld obtained
at the nal time level from the Navier-Stokes calculation performed at the end of the nal chapter. We shall actually
perform two experiments here. In the rst, we shall rst lay down an equally-spaced grid of 642 points on (0; 2)2, and
then, to obtain the points at which we are to evaluate U , we allow each grid point to move to a random position within
a ball of radius t (in the l1 -norm) of its original position. Then we allow the degree (K ) of the local interpolation to
increase, and measure the maximum error. The results are illustrated, for various values of t, in Figure 3.5. Note that the
spacing between the points of the original equally-spaced grid is  9:8  10 2 , so that setting t = 1 means that the nal
set of points is randomly distributed throughout the domain. In Table 3.5 we illustrate the c.p.u. times corresponding
to various values of K , and compare these to the c.p.u. time (1260s) taken by a full evaluation of U at the random
points. The savings are substantial, and are even more dramatic at higher spatial resolutions (i.e. higher values of N and
L). The exponential accuracy of the interpolation is evident in these results. Moreover, in the context of the spectral
Lagrange-Galerkin method, Fig. 3.5 indicates that the smaller the value of t, the less work necessary for each time
step.
Our second experiment examines a method (implemented in the Navier-Stokes calculations of Chapter 5) for avoiding
unnecessary work in the evaluation of U . We notice that in the expansion (3.4) the Chebyshev coecients of pi (by)
appear in a product. For high values of ki , the coecient a^pi;ki will be small. When two or more coecients each
corresponding to high values of the respective component of k are multiplied together, the result will be negligibly small.
Yet the calculation of the term corresponding to this outlying value of k takes just as much work as the calculation
of the term corresponding to k = (0; : : : ; 0) for example, which will be of the same order of magnitude as the entire
function. In order to avoid wasting time on such calculations, we set a tolerance and employ an upper bound of the size
of the contribution arising from a particular value of k, and exclude those values of k from our calculation for which this
measure falls below the set tolerance.
The simplest bound on the size of the contribution arising from the k-term on the right hand side of (3.4) is given
by "
Y  i #2 3
n i X Yn
max

Tki X bbX c 4 V^ (p) a^pi;ki p (bX c)5
i=1 p2 Z n i=1
X ^ Y n
 V (p) a^pi;ki (3.22)
p2 Z n i=1
0 n 11
X Y 2 2
 kV k @ a^pi;ki A
p2Z n i=1
0 1 21
Y Xn
= kV k @ j^api;ki j2 A : (3.23)
i=1 pi2Z
The expression (3.23) is easy to calculate, and provides a rst check for which values of k 2 K to ignore. A tighter
bound is given by (3.22). This expression is much more expensive to calculate, and yet need only be checked for those

40 41
Figure 3.6: Results from Experiment 4, t = 1=80. Figure 3.7: Results from Experiment 4, t = 1=800.

42 43
Problem 1 Let a 2 L1 (C#1 (
)n) \ L1(C# (
)n ) with r  a = 0. Given u0 2 H , nd u : [0; T ] ! V satisfying (in the
sense of distributions, with values in V 0 )
du + (a  r)u + Au = 0; (4.2a)
dt
u(0) = u0 : (4.2b)

Chapter 4
The following theorem expresses a standard existence and uniqueness result for Problem 1. (See for example [87],
Theorem 3.1.)
Advection{di usion equations Theorem 4.1 There exists a unique solution u of Problem 1 such that
u 2 L2 (V ) \ C (H ); (4.3a)
u0 2 L2 (V 0 ): (4.3b)

4.1 Introduction Regularity results concerning the solution of Problem 1 may also be found in [87].
In this chapter we consider the model advection{di usion problem
@u + a  ru  r2 u = 0 in
 (0; T ]; 4.1.2 The material derivative
@t (4:1a) In line with the discretisation of the hyperbolic problem in the previous chapter, our approach is to rewrite the advection
subject to periodic boundary conditions and the initial condition terms in (4.1a) as a directional derivative along the particle paths de ned by the ow eld a. This is seen to be
equivalent to rewriting the equation in a Lagrangian form with the Lagrangian variables being given by the particle paths.
u(x; 0) = u0 (x) in
; (4:1b) In the hyperbolic case, this gives the solution explicitly. Here the process results in a new di erential equation for which
where
= (0; 2)n, T a xed nal time and  a non-negative di usion coecient. The initial data u0 and the velocity appropriate higher-order discretisations present themselves fairly readily.
eld a are assumed to be periodic in the spatial variables, and we assume that r  a = 0. We recall for completeness the de nition of the Lagrangian variables. For x 2
, t 2 [0; T ], X (x; t; ) is de ned to
The combination of the advection and di usion phenomena modelled in this problem is characteristic of a wide be the solution of the initial value problem
variety of physical processes. In isolation the two phenomena merit quite distinct approaches towards the numerical dX (x; t; s) = a(X (x; t; s); s); s 2 [0; T ]nftg;
approximation of equations in which they occur. Our interest is in cases where the advection is the dominant term, and ds (4.4a)
so our numerical approximation of this problem will be based on the treatment of the advective term presented in the X (x; t; t) = x; (4.4b)
preceding chapters. We shall nd that our discrete equations will reduce to those of Chapter 2 in the limit as  ! 0.
We shall begin, in the remainder of this section, by describing the mathematical setting of the problem upon which various properties of solutions to (4.4) are given in Theorem 2.1. The operator E (t; s) is de ned by its action on a
the analysis of this chapter is founded, and giving a preliminary discussion of a backward-Euler discretisation of this function w on Rn by
problem, based on the approach introduced in Chapter 2. In the second section a Lagrangian form of the equation is E (t; s)w() = w(X (; t; s)):
introduced, and shown to be equivalent to the original form. The third section introduces and analyses time-discrete
schemes derived from the Lagrangian form, and in the nal section the fully discrete schemes are discussed. We recall that when r  a = 0, E (t; s) is an isometry on H . We also have, for t; s;  2 [0; T ], by Corollary 2.1,
E (t; s) = E (t;  )E ( ; s):
4.1.1 Mathematical setting of the problem
The material derivative Dtw of a function w on Rn  [0; T ] is de ned, where it exists, by
Suppose that u Ris a classical solution of (4.1). It is a straightforward consequence of the divergence theorem that if
r  a = 0 then u is constant, so that we can assume that u has zero mean. With this assumption, we de ne two

Dtw(; t) = dtd E (s; t)w(; t) ;
function spaces that will be central to our discussion of the problem. Let s=t
H := H#0 (
)=R and, for smooth w, we have, by the chain rule,
with norm denoted by kk, and
V := H#1 (
)=R; Dtw = @w @t + a  rw: (4:5)
with the norm on V given by
kukV = kruk : Thus for smooth solutions u, (4.1a) may be replaced by
We de ne the bilinear form A(; ) on V  V by Dt u + Au = 0: (4:6)
A(u; v) =  (ru; rv); u; v 2 V: We note that Dt is a directional derivative, the direction at (x; t) being that of the tangent to the curve (X (x; t; s); s).
By the Riesz Representation Theorem we may associate with A a continuous linear operator A from V to V 0 . For u 2 V ,
Au is that element of V 0 which satis es 4.1.3 First-order timestepping|an example
hAu; vi = A(u; v) 8v 2 V:
A is an isomorphism, and may be associated with  r2 . We de ne A straightforward backward-Euler discretisation of (4.6) between times tm 1 and tm results in
D(A) = fv 2 H jAv 2 H g; um + tAum = E (tm ; tm 1 )um 1 : (4:7)
which, with the graph norm, is a Hilbert space, isomorphic to H#2 (
) \ H . When um 1 2 H , E (tm ; tm 1 )um 1 2 H , and so, by the Lax-Milgram Theorem, (4.7) has a unique solution u 2 H .
m
A weak form of (4.1) is We shall sketch the derivation of an error estimate for the above method.

44 45
schemes that are higher-order in time. These can be obtained in a most natural way from the Lagrangian form of the
equations to be derived below.
-2

4.2 A Lagrangian form of the equation


10

-3
10

-4 Writing formally A(t; s) = E (t; s)AE (s; t), and v(t; s) = E (t; s)u(; s), (4.6) takes the form
d 
10


ds + A(t; s) v(t; s) = 0: (4:10)
-5
10

-6 s=t
10
Our initial aim in this section is to make the de nition of the operator A(t; s) precise and to describe some of its
properties. We shall then be in a position to give a weak form of (4.10) (removing the restriction s = t) and to apply
-7
10

-8 standard existence and uniqueness theorems, on which the analysis of the numerical schemes will be based. Our rst
10
nu=1.e-2
task is derive some preliminary results concerning the operator E (t; s).
-9
10 nu=1.e-4

10 20 30
nu=1.e-6
40 50 60
4.2.1 Further properties of E (t; s)
Figure 4.1: Results for the advection-di usion test problem The rst property we require concerns the commutator [E (t; s); r] of E (t; s) with the gradient operator r and is summed
up in the following lemma
Suppose that u is a smooth solution of Problem 1. Then Lemma 4.1 Let a 2 L1 (C#1 (
)n). Then for each t; s 2 [0; T ] the commutator [E (t; s); r] belongs to L(V; H n), and
we have
Z tm d k[E (t; s); r]kL(V;H n )  ejaj1;1;[t;s] 1: (4:11)
u(tm ) + tAu(tm ) E (tm ; tm 1 )u(tm 1 ) = tAu(tm ) + m 1 ds
E (tm ; s)u(s)ds
t
Z tm R
= Au(tm ) E (tm ; s)Au(s)ds ProofR The rst statement of the lemma is established by Lemma 2.2, once we have shown that
u =
Z tm 1
Z tm d 0 )
E (t; s)u = 0, which is a consequence of r  a = 0. To establish (4.11) we use the notation of
tm Lemma 2.2 to write
m
=
m 1 d E (t ;  )Au( )dds [Dj ; E (t)]u = jk (t)E (t)Dku;
Zt t m s
where we have taken s = 0 without loss of generality. It follows that
= (s tm 1 )E (tm ; s)DtAu(s)ds;
tm 1 k[r; E (t)]uk  0 (t) kruk :
and so u(tm) + tAu(tm) E(tm; tm 1)u(tm 1)  t kD Auk Now, by (2.26), Rt
t L1 (tm 1 ;tm ;H ) :
0 (t)  e 0 A0( )d 1  ejaj1;1;[0;t] 1;
It follows that
max u(tk ) uk   t D r2 u : (4:8) and the result follows.
k=0;:::;m t L1(H )
It will be shown below that (4.8) still holds if u is the solution to Problem 1 furnished by Theorem 4.1. We see that Next, we present a result which, like Lemma 4.1, is almost a corollary of Lemma 2.2, and which will be useful in the
the time-truncation properties of the scheme are good: when  ! 0 the discretisation is exact, recovering the situation analysis of time-discrete schemes in the next section.
already observed in the hyperbolic case. This is in contrast
2
to the errors committed by standard Eulerian timestepping Lemma 4.2 Let a 2 L1 (C#2 (
)n ). Then for each t; s 2 [0; T ] the commutator [E (t; s); A] is in L(D(A); H ), and
schemes, where the truncation error is of the form dtd 2 u 1 , and will not vanish as  ! 0.
L (H )
Example 1. In order to illustrate the meaning of (4.8), let us consider the application of (4.7), when coupled with k[E (t; s); A]kL(D(A);H )  e2jaj1;1;[t;s] +jaj2;1;[t;s] 1: (4:12)
the spectral discretisation in space described in Chapter 2, to the test problem
@u (sin x) @u  @ 2 u = 0 in (0; 2)  (0; T ] (4.9a) Proof Again following the notation of Lemma 2.2 (and setting s = 0), we have
@t @x @x2
u(x; 0) = sin x in (0; 2): (4.9b) Di Di E (t)u = Di (E (t)Di u + ij (t)E (t)Dj u)
= E (t)Di Di u + 2ij (t)E (t)Dj Di u
The exact solution to this problem develops sharp layers at x = 0 and x = 2, which eventually decay. Although
the ow eld here is not divergence-free, an error estimate similar to (4.8) may still be obtained, since E (t; s) maps +(Di ij (t))E (t)Dj u + ij (t)ik (t)E (t)Dk Dj u:
H  ! H  (and then also V  ! V  via Lemma 2.2), where H  (respectively V  ) is the restriction of H (respectively It follows that
V ) to the subspace consisting of odd functions.
A standard Fourier Galerkin method with a fourth order Runge-Kutta time integrator at high resolution was employed k[E (t); ]uk  20 (t) kuk + 1 (t) kruk + 02 (t) kuk
to obtain an `exact' solution for the purpose of comparison at time t = 1:57. In Figure 4.1 the error measured in the  (1 (t) + (0 (t) + 1)2 1) kuk :
L -norm is plotted on a log-scale against L (with N = L 1), for various values of the di usion  . 320 time steps of
2
the scheme (4.7) were taken to reach the nal time. The exponential accuracy of the spectral approximation is evident, Now, (2.23) translates to
although eventually it is obscured by the error due to the time discretisation. This nal error is proportional to  , as Zt
predicted by the theory, but only for very small values of  will it be comparable to the spatial discretisation error. Di ij (t) = [(Di ik (s))E (s)Dk aj (s) + 2ik (s)E (s)DiDk aj (s)
When approximating problems with smooth solutions, for which the spectral method is evidently capable of producing 0
extremely accurate results, we would like to obtain a more balanced global error. Thus we shall introduce discretisation +ik (s)il (s)E (s)Dk Dl aj (s) + E (s)Di Di aj (s)]ds;

46 47
so that Proof The properties (4.14b), (4.14c) follow immediately from Lemma 2.2. Let u; v 2 V and t 2 [0; T ].
Zt Consider (s) (s) for s; s 2 (0; T ). We have, denoting E (s; t) by E and E (s ; s) by E  ,
1 (t)  [1 (s)jra(s)j1 + 20 (s)ja(s)j1
0 (s) (s) =  (rE  Eu; rE  Ev)  (rEu; rEv)
+02 (s)ja(s)j1 + ja(s)j1 ]ds =  (rE  Eu; rE  Ev)  (E rEu; E  rEv)
Zt =  ([r; E  ]Eu; rE  Ev) +  (E rEu; [r; E  ]Ev):
 (1 (s) + e2jaj1;1;[0;s] )ja(s)j1 ds;
0
Thus, by Lemma 4.1,
using the fact that 0 (t)  ejaj1;1;[0;s] 1. Then, by Gronwall's lemma,
Zt R t ja( )j1d j(s) (s)j   (e2jraj0;1;[s;s] 1)e2jraj0;1;[t;s] kukV kvkV ; (4:15)
1 (t)  e2jaj1;1;[0;t] ja(s)j1 e s ds so that  is Lipschitz, which implies its absolute continuity. Hence  is di erentiable a.e. on [0; T ]. (4.14a)
0 and (4.14d) follow. Suppose s 2 (0; T ) is a point at which  is di erentiable. Then (4.14e) is obtained
= e2jaj1;1;[0;t] (ejaj2;1;[0;t] 1); from (4.15) by dividing through by js sj and taking the limit as s ! s.
and the result follows. Since the bilinear form A(s; tj; ) is coercive, the Lax-Milgram theorem allows us to deduce that A(t; s) is an
Let u 2 H and v 2 V . Lemma 2.1 implies that isomorphism from V onto V 0 . We can also de ne the domain of A(t; s) in H by

(E (t; s)u; v) = (u; E (s; t)v); D(A(t; s)) = fu 2 H j A(t; s)u 2 H g:


where E (s; t)v 2 V . This motivates the following extension of E (t; s) to V 0 , the dual space of V . The form A(s; tj; ) is symmetric so that the operator A(t; s) is self-adjoint in H . Its inverse A 1 (t; s) is also self-adjoint,
in H .
De nition 4.1 Let v0 2 V 0 ; then E (t; s)v0 is that element of V 0 such that We are now ready to state a Lagrangian form of Problem 1 as an abstract Cauchy problem in V . More precisely, we
state:
hE (t; s)v0; vi = hv0 ; E (s; t)vi 8v 2 V; Problem 2 Given t 2 [0; T ], ut ;0 2 H , nd ut : [0; T ] ! V satisfying (in the sense of distributions, with values in
where h; i is the duality pairing between V and V 0 . V 0)
Then, for each s; t 2 [0; T ], E (t; s) is an isometry on V 0 and bounds similar to (2.12) hold. d  
dt ut (t) + A(t ; t)ut (t) = 0 on (0; T ];
 (4.16a)

4.2.2 The operator A(t; s) ut (0) = ut ;0 : (4.16b)

Our rst step towards a precise de nition of the operator A(t; s) is now open to us. For s; t 2 [0; T ] we de ne the bilinear
form A(s; tj; ) on V  V by The following results for Problem 2 are standard (see Theorems 3.4{3.6 of [87] for a general statement).
A(s; tju; v) =  (rE (s; t)u; rE (s; t)v); 8u; v 2 V: (4:13) Theorem 4.2 Suppose that (4.14a{4.14c) hold. Then there exists a unique solution ut of Problem 2 such that
The continuity of A(s; tj; ) is assured by Lemma 4.1. By the Riesz Representation Theorem we may associate with ut 2 L2 (V ) \ C (H );
A(s; tj; ) a continuous linear operator A(t; s) from V to V 0 as follows: for every u 2 V , A(t; s)u is de ned to be that u0t 2 L2 (V 0 ):
(4.17)

element of V 0 which satis es (4.18)


hA(t; s)u; vi = A(s; tju; v) 8v 2 V: If, moreover, (4.14d) and( 4.14e) hold, and we assume that
We note that for all s; t 2 [0; T ],
A(t; s) = E (t; s)AE (s; t) in L(V; V 0 ): ut ;0 2 D(A(t; 0)); (4:19)
The following lemma encapsulates further properties of the bilinear form A(s; tj; ) (and so of A(t; s)). then
Lemma 4.3 Let a 2 L1(C#1 (
)n) \ L1 (C#(
)n). Then ut (t) 2 D(A(t; t)), 8t 2 [0; T ] and
For every u; v 2 V , t 2 [0; T ], s ! (s) = A(s; tju; v) is measurable; (4.14a)
the map t ! A(t; t)ut (t) is continuous from [0; T ] into H , (4.20)

A(s; tju; v)j  e2jraj0;1;[t;s] kuk V kvkV ; 8u; v 2 V; 8s; t 2 [0; T ]; (4.14b) and
A(s; tju; u)  e 2jraj0;1;[t;s] kuk2V ; 8u 2 V; 8s; t 2 [0; T ]; (4.14c)
For every u; v 2 V ,
u0t 2 L2 (V ) \ C (H ); (4.21)

(s) is absolutely continuous with derivative s ! 0 (s) = A0 (s; tju; v); (4.14d)
u00t 2 L2 (V 0 ): (4.22)

For a.e. s; t 2 [0; T ] , A0 (s; tj; ) is a bilinear continous form on V with Suppose now that a 2 L1(C#2 (
)n), so that (see Lemma 4.2) D(A(t; t)) = D(A) = H#2 (
)\H . Replacing assumptions
jA0 (s; tju; v)j  2jraj0;1 e2jraj0;1;[t;s] kukV kvkV ; 8u; v 2 V: (4.14e) (4.19) by
ut ;0 2 V; (4:23)
we have
ut 2 L2 (D(A)) \ C (V ): (4:24)

48 49
Suppose that we discretise Problem 2 in time by a q-step linear multistep method. We obtain
q 0 1 2 3 4 5 6 0 X
q
1 1 -1 1 ( i + t i A(t; tm i ))utm i = 0; (4:27)
2 1 - 43 1 2
i=0
3 1 - 18 93 - 112 63
4 1
11
- 48
11
36 - 16 3 11
12 where m ranges from q to M , and t = T=M . We shall be interested primarily in the backward di erence formulae
5 1
25
- 300
25
300 25
- 200
25
75 12
- 137
25
60 (see for example Gear [29]), which are strongly A0 -stable for q  6. These have i = 0 for i > 0. Full details of
6 1
137
- 360
137
450 137
- 400
137
225 72
- 147 10 137
60 the coecients of these formulae are given in Table 4.1. Under stronger assumptions on the velocity eld a than are
147 147 147 147 147 147 necessary for the backward di erence formulae, we shall also consider extensions to the case i  0 for i > 0. We shall
Table 4.1: Coecients for the backward di erentiation formulae. in either case assume that the method used is strongly A0-stable. Le Roux's results may be applied to these schemes
to give stability and optimal-order convergence to the solution of Problem 2. (4.27) represents a time-discretisation
of Problem 1 in Lagrangian coordinates which originate at time t . We shall derive from (4.27) an approximation to
Problem 1 involving a di erent set of Lagrangian coordinates at each time level (with the coordinates used to obtain
Remark. It is perhaps worth noting that of the results in Theorem 4.2 only (4.24) depends on the symmetry of A(s; tj; ). um from (4.27) originating at time tm ) and present stability and convergence results for this. We shall only sketch the
It would seem at rst sight that Problem 2 is actually a one-parameter family of problems fP2(t)gt 2[0;T ] . However, proofs, since they are at all points analogous to the proofs to be found in [46].
it is readily seen that, if s ; t 2 [0; T ] and ut is the solution of P2(t ) with initial datum ut ;0 , then E (s ; t )ut satis es Setting um;i = E (tm+i ; t )um m+i m m
t = E (t ; t )u , i = 0; : : : ; q, we nd, when i = 0 for i > 1, that the method
P2(s) with initial datum E (s ; t)ut ;0 . Thus, since E (s ; t ) is an isomorphism on V , the problems P2(t) and P2(s) represented by (4.27) is equivalent to
are equivalent.
The central question of whether Problem 2 is equivalent to Problem 1 is more delicate. The answer is contained in X
q
the following lemma. um;0 = um = ( 0 + t 0 A) 1 E (tm ; tm 1 )( i + t i A)um i;i 1 (4.28a)
i=1
Lemma 4.4 Problem 1 and Problem 2 are equivalent, in the sense that the unique solution to Problem 1 (with initial um i;i = E (tm ; tm 1 )um i;i 1 ; i = 1; : : : ; q 1; (4.28b)
datum u0 ) provided by Theorem 4.1 may be identi ed with the unique solution to Problem 2 (with initial datum which is the form used in practice.
E (t; 0)u0 ) provided by Theorem 4.2. Fig. 4.2 illustrates (4.28) for a second-order backward di erence time step with q = 2.
Proof Let u be the solution to Problem 1. We shall show that u also satis es Problem 2 and so it may be
identi ed with the unique solution provided by Theorem 4.2. um;0 = (1 + 32 tA) 1 ( 43 um 1;1 13 um 2;2 ) (4.29a)
Our rst step is to show that um i;i = E (tm ; tm 1 )um i;i 1 ; i = 1; 2: (4.29b)
d  d
  The meaning of the labelling is as follows. For example, the positioning of the label um 1;1 corresponds to the point

dt E (t ; t)u(t) = E (t ; t) dt + a  r u(t) (4:25) Xj , the foot of the trajectory through (xj ; tm ). We emphasise that the m 1 in the superscript indicates that um 1;1
represents the solution at time tm 1 , and the 1 corresponds to the fact that xj ! Xj represents evolution through one
in D0 (0; T ; V 0 ). Since u 2 H 1 (V 0 ) \ L2 (V ), there is (by density) a sequence fuj g  C 1 ((0; T ); C#1(
)) time step (E (tm; tm 1 )). One can think of the procedure given by (4.29) as being
which converges to u in the norm kkH 1(V 0 )\L2 (V ) = kkH 1 (V 0) + kkL2 (V ) . For each j , (4.25) holds in the (um 1;0 ; um 2;1 ) ! (um 1;1 ; um 2;2 ) ! (um;0 ; um 1;1 ).
classical sense with u replaced by uj . Let  2 D(0; T ) and v 2 V . Integrating (4.25) against v over
and (4.29b) (4.29a)
against  over [0; T ] we obtain
For the purposes of our analysis we shall de ne the q  q matrices of operators
ZT Z T    0 ( + tA) 1
hE (t; t)uj (t); vi 0 (t)dt = E (t; t) dtd + a  r uj (t); v (t)dt; (4:26) BB 0 0
1
CC
0 0 I
 = B
B  CC ;
where h; i is the duality pairing between V 0 and V . Since, for each t 2 [0; T ], E (t ; t) is an isomorphism BB  CC
on V 0 , uniformly bounded in t, and since dtd uj ! dtd u in L2(V 0 ) and a  ruj ! a  ru in L2 (H ), we @  A
may take the limit through (4.26) as j ! 1 and replace u with uj , so that (4.25) holds in the sense of
distributions (with values in V 0 ). 0 ( + tA)     ( +I tA) 1
q q
In this sense, we may write BB 1 1 I 0 0 CC
d  d   B
B    CC ;
   = B CC
dt + A(t ; t) ut (t) = E (t ; t) dt + a  r + A u(t); B@   
A
 0 
where ut = E (t; )u, and we deduce that ut satis es (4.16a). I 0
0 E(tm; tm 1) 1
The equivalence of Problems 1 and 2 means that we can derive discrete approximations to Problem 1 from schemes BB  CC
designed to approximate Problem 2. A natural way to construct such approximations is to consider the time-discrete Em = B B@  CC ;
case rst, introducing discretisations in the space variables as a second step. This is the pattern our analysis will follow.  A
E (tm ; tm 1 )
4.3 Time discretisation 0 1 I 0   0 0 1
BB 0 I  C C
Several authors (see, for example, [18, 46, 47, 48, 89]) have considered the problem of the time discretisation of abstract
1 = B
B   C C;
Cauchy problems such as Problem 1 and Problem 2. The results of Le Roux in [46] concerning higher-order time BB   C CA
discretisations are especially relevant here, and we shall draw heavily on her work in this section. @  0 
I 0

50 51
kX1 X k 1
t = (j j1 )[; Ep ](k j k1 j ) + j1 [; Ep ](k j k1 j )
j=1 j=0
X
k Xk
+ (j j k j
1 )[; Ep ]1 + 1 [; Ep ]k1 j :
j (4.34)
tm um;0 j=1 j=0
From [46], p.133, we have
1=2 j j
A ( 1) L(Hq ) = (j j1)A1=2 L(Hq )  Ce
jt
(j t)1=2
; (4:35)

tm um 1;1 um 1;0
1 and from [46], Lemma 3, we have j  Ce j : (4:36)
1 L(H q )
We thus seek suitable bounds on kA 1 [; Ep ]A 2 k
L(H q ) for f1 ; 2 g = f 21 ; 12 g, f 21 ; 0g, f0; 12 g, and f0; 0g. We have
A 1 [; E ]A 2
p L(H q )
um 2;1
tm 2 Xq
1
um 2;2  A [( 0 + 0 tA) 1 ( i + i tA); E (tp ; tp 1 )]A 2 : (4.37)
i=1
Now, denoting E (tp ; tp 1 ) by E ,
A 1 [( 0 + 0 tA) 1 ( i + i tA); E ]A 2 (4.38)
= tA 1 ( 0 + 0 tA) 1( i 0 0 i )[A; E ]A 2 ( 0 + 0 tA) 1 ; (4.39)
xj 3 xj 2 xj 1 xj x
Figure 4.2: Illustration of a second order time step and we have (see [46], p.132) A 1 ( + tA) 1 1 1
0 0 L(H;V )  C t 2 ; (4:40)
and we shall denote and ( + tA) 1A 2
m =  Em  and  =  : 0 0  C t2 12 :
L(V 0 ;H ) (4:41)
Writing um = (um;0 ; : : : ; um q+1;q 1 )T , the method (4.28) may then be written A bound on k[A; E ]kL(V;V 0 ) may be obtained as follows. Let u; v 2 V : then
um =  m um 1 : (4:30)
h[A; E ]u; Evi =  (rEu; rEv)  (ru; rv)
4.3.1 Stability =  ([r; E ]u; rEv) +  (E ru; [r; E ]v);
Our rst concern will be to demonstrate that the scheme is stable. By (4.30) this is equivalent to showing that there is so that [A; E(tp; tp 1)] 2jraj0;1;[tp
L(V;V 0)   (e
a constant C such that 1 ;tp ] 1): (4:42)
kk : : : q kL(H q )  C; Combining (4.37{4.42) we obtain
uniformly in k. The positive constants C and  will, in the following discussion, be generic constants depending only on A 1 [; E ]A 2  C (e2jraj0;1;[tp
the particular multistep formula used (i.e. on 0 ; : : : ; q and 0; : : : ; q ), and sometimes on t0 , an upper bound on p L(H q )
1 ;tp ] 1)t1 +2 : (4:43)
the size of the timestep.  will also be proportional to minf; 1t0 g. Any other dependence will be explicitly stated. We
have the following operator identities. (4.43), together with (4.33){(4.36), may be used as in [46] pp.131{134 to give
k : : : q = Ek : : : Eq k q+1 [k+1; E ]  C (e2jraj0;1;[tp 1 ;tp ] 1)e (k+1)t : (4:44)
p L(H q )
kX
+1
+ Ek : : : Ep (k+1 p p 1 Ep 1 k+2 p )p 2 : : : q ; (4.31) The second term appearing on the right hand side of (4.32), k (p Ep ), vanishes when the multistep scheme being
p=q+1 used is a backward di erence formula ( i = 0; i > 0). When for example 1 6= 0, we must assume that a 2 L1 (C#2 (
)n )
and so that Lemma 4.2 applies. Then we have
k p Epk+1 = [k+1 ; Ep] + k (p Ep ): (4:32) k ( E )  C jaj p 1 ;tp ] e (k+1)t : (4:45)
p p L(H q ) 2;1;[t
As in [46], Proposition 4, we have, for k  1,
k  Ce kt: (4:33) We can now deduce from (4.31), (4.32), (4.44) and (4.45) that
L(H q )
kk : : : q kL(H q )  Ce (k q+1)t
It remains to bound the two terms appearing on the right hand side of (4.32). For the rst, the following operator
identity holds. kX
+1
+C jaj1;1;[tp 1 ;tp ] e (k+2 p)t kp 2 : : : q kL(H q ) ; (4.46)
X
k
p=q+1
[k+1 ; Ep] = j [; Ep ]k j
j=0

52 53
with jaj1;1;[tp 1 ;tp ] replaced by jaj2;1;[tp 1 ;tp] if 1 6= 0. If we write bq = 1, ak = jaj1;1;[tk ;tk+1] and bk = e(k q)t kk : : : q kL(H q ) , Then we have
for k  q + 1, (4.46) becomes  m = m  m 1 + t (T m ; 0; : : : ; 0)T
X
k
mXq
bk+1  C (1 + ap bp ):
= m : : : q  q 1 + t

m : : : m k+1  T m k ; 0; : : : ; 0 T
p=q
k=1
The discrete Gronwall lemma now allows us to deduce that
Pk +t (T m ; 0; : : : ; 0)T ;
bk+1  CeC p=q ap ; so that, making use of Theorem 4.3,

k m kH q  km : : : q kL(H q )  q 1 H q
which when translated back into the original terms gives us the following result.
mXq
Theorem 4.3 (Stability) Suppose that the multistep method (4.27) is strongly A0 -stable, and is such that i = 0 for
i > 0, or, if i geq for i  1, that a 2 L1 (C#2 (
)n). Then there is a positive constant C depending only on 0 ; : : : ; q , +t km : : : m k+1 kL(H q ) T m k + t kT m k
k=1
0 ; : : : ; q and t0 , and a constant  depending on t0 and  , such that, for m = q; : : : ; M ,
 CeC jaj1;1 (m q+1)t  0
Hq
km : : : q kL(H q )  CeC jaj1;1 (m q+1)t ; (4:47) mXq
+CeC jaj1;1 tp e kt Dtp+1 u L1 (tm q k ;tm k ;H )
with jaj1;1 replaced by jaj2;1 when i(1) 6= 0. k=1
and (4.48) follows.
4.3.2 Convergence These results concern the performance of the Lagrange-Galerkin method's approach to timestepping, which may be
Let u be the solution to Problem 1 with initial datum u0 . Then, by Theorem 4.1, u 2 C (H ) and so we can sensibly implemented in combination with a variety of spatial discretisation methods, such as nite di erence methods, nite
refer to u(t), for t 2 [0; T ]. Suppose that u0;q 1 ; : : : ; uq 1;0 are given as starting values for the method (4.28), and that element methods, or spectral methods. In the next section we shall examine the e ect of the incorporation of the spectral
they correspond to approximations of the terms E (tq 1 ; 0)u0 ; E (tq 1 ; t1)u(t1 ); : : : ; u(tq 1 ). Then we have the following projection and interpolation operators introduced in Chapter 2.
convergence result.
Theorem 4.4 (Convergence) We suppose that the conditions of Theorem 4.3 are satis ed, and that Dtp+1 u 2 L1 (H ), 4.4 Space discretisation
where the multistep scheme (4.27) has order of accuracy p. Then, for m = q 1; : : : ; M ,
In this section we turn our attention to fully-discrete practicable methods, and we restrict ourselves to those arising from
u(tm) um;0  CeCjaj1;1 e (m q+1)t X E (tq 1 ; ti )u(ti )
q 1 backward di erence time discretisations (4.27) with i = 0; i > 0. Let K; L and N be positive integers, with L and N
ui;q 1 i
related by (2.53). Corresponding to (4.28) we have, for m = q; : : : ; M ,
i=0 !
ZT X
q
e (T t) Dtp+1u(t) dt : U m;0 = ( 0 + t 0 A) i N PL E (tm ; tm 1 )PK;L U m i;i 1
1 (4.51a)
+ t p (4.48)
0 i=1
U m i;i m m
= N PL E (t ; t )PK;L U
1 m i;i 1 ; i = 1; : : : ; q 1; (4.51b)
Proof We begin by de ning the truncation error T m at time tm by and we assume that U 0;q 1 ; : : : ; U q 1;0 (each in SN ) are either given or have been calculated by some other means to
approximate utq 1 (0); : : :; utq 1 (tq 1 ), where u is the solution to Problem 1, and, for s; t 2 [0; T ], us (t) = E (s; t)u(t).
X
q De ning N PL and PK;L on H q in the obvious way, we further de ne, for m = q; : : : ; M ,
T m = 1t E (tm ; tm i )( i + t i A)u(tm i ): (4:49)
i=0 e m =  N PL Em PK;L  ;
Now, let ut (t) = E (t ; t)u(t), t; t 2 [0; T ]. Lemma 4.4 established that ut is then a solution of Problem and, for m = q 1; : : : ; M ,  ;
2, and it is readily seen that, writing Ttm = E (t ; tm )T m , U m = U m;0 ; : : : ; U m q+1;q 1 T

1 X
q so that (4.51) may be rewritten
Ttm =  m i  m i
t i=0 ( i + t i A(t ; t ))ut (t ) U m = e m U m 1 : (4:52)
1 X q
 m i 0 m i 4.4.1 Stability
= t i=0 i ut (t ) t i ut (t ) ;
Our concern is to obtain a bound on e m : : : e q in the L(SNq )-norm. We have, for k = q; : : : ; m,
and standard Taylor-series expansion then yields, changing back to our original terms, e k : : : e q = N k : : : N q
Z tm X
k
kT m k  C tp 1 m q
Dtp+1u dt: (4:50) + N k : : : N j+1 (e j N j )e j 1 : : : e q : (4.53)
t j=q
Now let us de ne, for m = q 1; : : : ; M , The stability of the rst term follows exactly as in Theorem 4.3 since we have, for j = q; : : : ; k,
m = utm (tm ) um;0 ; : : : ; utm (tm q+1 )

um q+1;q 1 T : N k : : : N j = N Ek : : : N Ej k+1 j (4.54)
X
k
+ N Ek : : : N Ep+1k p N (p Ep )N p 1 : : : N j ; (4.55)
p=j

54 55
and it is readily seen that, with the positive constants C and  exactly as in Theorem 4.3, Then, by expanding the right hand side of (4.61),

kN k : : : N j kL(Sq )  CeC jaj1;1 (k+1 j)t :
N
(4:56) k m kH q  e m : : : e q L(Sq )  q 1 H q
N
m
Our attention thus centres on the term
+
X e m : : : e j+1 (t T j + Tej ): (4.62)
L(Sq ) Hq
(e j N j ) =  (N PL Ej PK;L N Ej ) j=q N

appearing in (4.53). The central term may be expanded We must now obtain a bound on Tem q . To begin with, we bound the terms on the right hand side of (4.59).
H
Combining the results of (2.73), (2.74) and (3.11), we obtain
N PL Ej PK;L N Ej = N PL Ej (PK;L I ) + N (PL I )Ej ; m;i
so that, under the conditions of Theorem 3.1, there is a positive constant C depending on C5 and on C6 such that Te  CN 0 ku0k0 [min(1; n 21 N jaj0;1;[tm 1 ;tm] ) + jaj[] ;1;[tm 1 ;tm ] ]; (4:63)

kN PL Ej PK;L N Ej kL(SNq )  C jaj[] ;1;[tj 1 ;tj ] ; (4:57) where 0 = minfK + 1; g.
Continuing as in previous convergence proofs, we obtain the following result.
so that, from (4.53) and (4.56), Theorem 4.6 Suppose that the multistep method (4.27) is strongly A0 -stable, and is such that i = 0 for i > 0, and
that Dtp+1 u 2 L1 (H ), where the multistep scheme (4.27) has order of accuracy p. Suppose also that (3.10) and the
e k : : : e q L(Sq )  CeCjaj1;1 (k q+1)t conditions of Lemma 2.6 hold. Then, for m = q 1; : : : ; M ,
N
Xk u(tm) um;0  CeCjaj[];1 e (m q+1)t X
q 1
e j 1 : : : e q q :
1 ;tj ] eC jaj1;1 (k j )t  E(tq 1 ; ti)u(ti) ui;q 1 i
+C jaj[] ;1;[tj L(S )
j=q N
0
 1 i=0

+N  [min  T 21
In the same way that Theorem 4.3 was deduced from (4.46), we obtain the following stability result for the fully discrete t (1 e ); n N!jaj0;1 + jaj[] ;1 ]
scheme (4.52). ZT
Theorem 4.5 Suppose that the multistep method (4.27) is strongly A0 -stable, and is such that i = 0 for i > 0. Suppose + tp e (T t) Dtp+1u(t) dt : (4.64)
0
also that the conditions of Theorem 3.1 hold. Then there is a positive constant C , depending only on 0 ; : : : ; q , 0 ,
t0 , C5 and C6 , and another constant  proportional to minf; 1t0 g, such that, for m = q; : : : ; M ,
4.4.3 The initialisation procedure
e m : : : e q L(Sq )  CeCjaj[];1 (m q+1)t : (4:58)
N In all of the above analysis it has been assumed that suitable starting values have been obtained for the quantities
E (tq 1 ; ti )u(ti ), i = 0; : : : ; q 1. In this section we describe one procedure for obtaining approximations to these values
with appropriate accuracy, which is used in each of the calculations presented in the remainder of this Thesis.
4.4.2 Convergence The initial data for our numerical procedure consists of the projection of the initial data for the original problem,
N PL u0 . We set t0 = 2 s t, where s is an integer which is such that t0  tq 1 (e.g. s = [(q 2)j log tj= log 2]+
The strategy employed in the proof of Theorem 4.5 (i.e. that of regarding N PL Ej PK;L as a perturbation of N Ej ) is 1), and perform q 1 rst-order time steps with timestep t0 . This enables us to nd approximate values for E ((q
exactly the same as that employed in the proofs of the stability results in Chapter 2. In the same way, our demonstration 1)t0 ; it0 )u(it0 ), i = 0; : : : ; q 1; this is enough information to be able to perform a qth-order time step with a
of the convergence of the fully-discrete scheme here will, in its essential ideas, resemble that of the convergence results timestep of t0 . Enough of these are then performed to provide enough information to carry out a qth-order time step
in that chapter. but with double the size of the timestep, and this process is repeated until a timestep of size t may be used.
For m = q 1; : : : ; M and i = 0; : : : ; q 1 we de ne The algorithm is illustrated in Fig. 4.3, for a fourth-order time step with s = 3 (so that t0 = 8t ). Time is
advancing to the right, and each successive vertical level represents a step in the calculation. On any particular level,
 m;i = N utm+i (tm ) U m;i and m;i = (I N )utm+i (tm): the circles represent times at which a value for the approximate solution is known at that step; solid circles indicate
If we denote E (tm ; tm 1 ) by E , and write those values which will be used in subsequent steps. For example, at the 6th stage, the solution is calculated at times
3 8t ; : : : ; 6 8t . Values at times 0 and 4t are carried over from previous calculations. Of these values, only those at times
Tem;i = N Em i;i 1 N (PL I )E N utm 1 (tm i ) 0, 4t , 2t and 3 4t are kept, and are used for the next stage of the calculation.
N PL E (PK;L I )N utm 1 (tm i ); i = 1; : : : ; q: (4.59)
4.4.4 An example revisited
we nd, from (4.49) and (4.51), that In Figure 4.4 the method (4.51) is used, for q = 1; : : : ; 4, on Problem 1 above. The nal time is (as before) 1.57, and
X
q ! the error is plotted against the number of time steps taken; the value of  is 10 2 , and the values of N and L are 63
 m;0 = ( 0 + t 0 A) 1 i (N PL E PK;L  m 1;i 1 + Tem;i ) + tN T m and 64 respectively, so that spatial resolution has been obtained. For the higher values of q, the imbalance between
i=1 spatial and temporal error has to some extent been redressed. Figure 4.5 plots results from the same set of experiments
(4.60a) as Figure 4.4, but this time the error is plotted against the c.p.u. time used (on a Sun SPARC Station). The numbers
 m i;i = N PL E PK;L  m i;i 1 + Tem;i ; i = 1; : : : ; q 1; (4.60b) of time steps used to obtain these results were as shown in the following table.
P
(compare (4.59) and (4.60) with (3.14)). De ning Tem = ( qi=1 Tem;i ; Tem;1 ; : : : ; Tem;q 1 )T and  m = ( m;0 ; : : : ;  m q+1;q 1 )T q range
we can rewrite (4.60) as 1 10{1280
2 10{640
 m = e m  m 1 + t (T m; 0; : : : ; 0)T + Tem : (4:61) 3 10{400
4 10{240

56 57
12

11 10
-3

-4
10
-5
10
-6
10 10
-7
10
-8
10
q=1
-9 q=2
10
9 q=3
-10
10 q=4

8 5 10 20 50 100 200

Figure 4.5: L2 error against c.p.u. time (secs.) for the advection-di usion test problem.
7
q 1 2 3 4 5 6
6 1 1
2 3 - 12
2
23
5
3 12 - 16
12
5
12
4 4 55 - 59 37 - 249
24
1901 24 24
3 5 9982 - 2774
9982
2616
9982 - 1274
9982
251
9982
6 4277 - 7923 9982 - 7298 2877 475
- 1440
1440 1440 1440 1440 1440
2
1 Table 4.2: Coecients for the Adams-Bashforth formulae.

The results are only illustrative, since the measure of c.p.u. time used can vary slightly from run to run. They indicate
that for maximum eciency the order of the method should be selected with a view to the accuracy required: for high
accuracy the 4th order time step is to be preferred, whereas for low to moderate accuracy the 2nd or 3rd order schemes
are quite adequate.

4.5 A nonlinear example


Figure 4.3: Initialisation procedure for fourth order time step. Thus far our discussion has been restricted to linear problems. However, our eventual aim is the application of these ideas
to the solution of (for example) the Navier-Stokes equations, which are, of course, non-linear. A suitable model problem
for the development of techniques for the solution of the Navier-Stokes equations is the viscous Burgers equation
-3 ut + uux = uxx in (0; 2)  (0; T ]; (4.65a)
u(t = 0) = u0 in (0; 2):
10
-4
(4.65b)
10
-5 This generalises (4.1) in that it is nonlinear, and in that the velocity eld is not divergence-free. The analysis of this
chapter will not apply straightforwardly to this problem. For the time being we content ourselves with experiments,
10

10
-6
although the theory does promise to be interesting and will be pursued elsewhere.
10
-7 The main additional problem to be overcome here by the numerical method is the solution of the trajectory equations
d  
dt X (x; t ; t) = u(X (x; t ; t); t) in (0; 2)  (0; T ]; (4.66a)
-8
10

X (x; t ; t ) = x in (0; 2): (4.66b)


-9
10
q=1
-10

In order to be able to implement (4.51), we must know how to apply the operator E (tm ; tm 1 ), and to do this we must
10 q=2
q=3
10
-11
q=4 know the solution to the trajectory equation (4.66), which itself depends on the solution to (4.65). The approach taken
is to discretise (4.66) by an Adams-Bashforth scheme of order q, and to solve this in tandem with the discretisation of
(4.65) given by (4.51). Since we seek a di erent set of coordinates at each time level the initial conditions are set at tm ,
20 50 10 200 500

Figure 4.4: L2 error against number of time steps for the advection-di usion test problem.

58 59
and so the fully-discrete equations are
X
q -2

x X (x ; tm; tm 1 ) = t i U m i;i 1 (X (x ; tm ; tm 1 )); 0    2L 1; (4:67)


10

i=1 10
-3

(where i are given in Table 4.2) which must be solved by some iterative procedure, which we shall discuss below. Thus,
to obtain fU m i;i g0iq 1 from fU m 1 i;i g0iq 1 we rst solve (4.67), then use the values of X (x ; tm ; tm 1 ) to 10
-4

de ne the operator N PL E (tm ; tm 1 ), and nally we can solve (4.51) just as in the linear case.
-5
10

4.5.1 A modi ed Newton method 10


-6

In this section we describe the iterative method that is used to solve (4.67), which we shall write as the set of equations
y(x ) + f (x + y(x )) = 0; 0    2L 1: (4:68)
-7
10 q=1
P
Here y(x ) = X (x ; tm; tm 1 ) x and f = t qi=1 i U m i;i 1 . Now, in our case, the function f can be evaluated 10
-8
q=2
q=3
rapidly (by means of a FFT) at the points fx g, but the further the arguments fx + y(x )g deviate from the set
q=4

fx g, the more expensive this evaluation becomes. (This is graphically illustrated by the results from Experiment 4 in 50 100 200 400 800
Chapter 3). With an iterative scheme for solving (4.68) we obtain sequences of points fym (x )g tending towards the
points fy(x )g that solve (4.68). Each iteration, however, involves an evaluation of f whose cost is of the same order Figure 4.6: L2 error against number of time steps for Burgers equation
as the evaluation of f at the points fy(x )g. This is precisely because, for each  , ym (x ) ! y(x ): yet it ought to be
possible to use this convergence to reduce the cost of the later iterations, since the change between subsequent iterates When Newton's method is to be used, all that changes is that (4.72b) and (4.72e) are replaced by
is reducing to zero. This hope is what underlies the development of the modi ed iteration described below.
We shall rst describe our method in the context of a xed-point iteration, and then show how the approach can be 0 (x) = f (x)=(1 + f 0 (x));
applied to Newton's method, which is what we use in practice. Consider the functional equation n+1(x) = fn (x)=(1 + fn0 (x))
y(x) + f (x + y(x)) = 0: (4:69)
respectively.
f will be small, so y will be small and we take our rst iteration to be The analysis of this procedure is awkward and is incomplete at the present time. The method is designed so that
y0 (x) = f (x): yn (x) ! y(x) for each x as n ! 1, and our experience with the method suggests that this occurs (rapidly), at least
when f is small.
This will give We have described the procedure as applied to a continuous functional equation. Of course in (4.68) we are interested
y0 (x) + f (x + y0 (x))  0; in the solution at only a nite number of points x . We thus start by nding f0 (x )g and fy0 (x )g. These are used
we can then use y0 to de ne a new function smaller than f (if y0 is a good guess), to de ne
f0 = PL (y0 + Ey0 PK;L f ); (4:73)
f0 (x) = y0 (x) + f (x + y0 (x)): where for functions g and h, Eg h(x) = h(x + g(x)), and we proceed inductively by using fn+1 (x )g to de ne
Let us seek now a solution  to the equation
fn+1 = PL (n+1 + En+1 PK;L fn ): (4:74)
 (x) + f0 (x +  (x)) = 0: (4:70)
This process describes 2L semi-independent iterations (one for each value of  )|they are coupled together only in as
If we can nd such a solution then we can construct a solution to (4.69) by much as all of the 2L points are used in the de nitions of the new functions fn . (4.68) is thus solved to spectral accuracy.
y(x) =  (x) + y0 (x +  (x)): (4:71) We note that the value of K in the right hand side of (4.74) will be very much smaller than that in the right hand side
of (4.73), an observation which forms the justi cation for the introduction of this procedure.
However, from our point of view, (4.70) has an advantage over (4.69) since, usually,  will be much smaller than y, and
so f0 (x +  (x)) will be much cheaper to evaluate than f (x + y(x)). We have carried out, in going from (4.69) to (4.70)
and (4.71), one loop in our modi ed iteration procedure. The full method is outlined below. 4.5.2 Numerical results
The method (4.51), coupled with (4.67) solved by the Newton form of (4.72), was used to solve (4.65) in several test
Equation to be solved runs, with the same initial data as Example 1.
In Fig. 4.6 the behaviour of the error as the timestep is decreased is illustrated. The values of N and L are 63 and 64
y(x) + f (x + y(x)) = 0: (4.72a) respectively, the nal time is t = 1:57, and the value of  is 10 2 . Thus the calculation is analogous to that illustrated in
Starting step Fig. 4.4. Here we also see that the high order timestepping schemes are producing the required convergence. Moreover,
0 (x) = f (x); (4.72b) for the fourth order time step we see that we have indeed achieved a balance between spatial and temporal error. The
y0 (x) = 0 (x); (4.72c) spatial error here is higher than in Fig. 4.4 because the gradients in the solution are higher in this case.
f0 (x) = y0 (x) + f (x + y0 (x)): (4.72d)
Recursive step
n+1 (x) = fn (x); (4.72e)
yn+1 (x) = n+1 (x) + yn (x + n+1 (x)); (4.72f)
fn+1 (x) = n+1 (x) + fn (x + n+1 (x)); (4.72g)
and it is shown by induction that
fn+1 (x) = yn+1 (x) + f (x + yn+1 (x)): (4.72h)

60 61
5.2 The equations and their approximation
5.2.1 Statement of the problem
The Navier-Stokes equations for a viscous incompressible uid in two or three space dimensions take the form
ut + (u  r)u  u + rp = f on
 (0; T ); (5.1a)
Chapter 5 r  u = 0 on
 (0; T ); (5.1b)
u(x; 0) = u0 (x) on
; (5.1c)

The Navier-Stokes equations where u(x; t) is the velocity of the uid, p(x; t) the kinematic pressure,  the kinematic viscosity, f (x; t) the density
of body force per unit mass, and u0 the initial velocity. The problem is completed in this case by periodic boundary
conditions
u(x + 2ei ; t) = u(x; t) 8x 2 Rn ; 8t 2 [0; T ]; (5:1d)
where e1 ; : : : ; en is the canonical basis for Rn . We assume that the mean ow is zero.
5.1 Introduction The following functional-analytic formulation of the problem is due to J. Leray. Details may be found in Temam's
works [85, 86].
As mentioned in Chapter 1, the nite element Lagrange-Galerkin approximation of the Navier-Stokes equations has been For f and u0 given,
analysed by Pironneau [60] and subsequently by Suli in [76], where optimal-order error estimates were obtained in both f 2 L2(H ); (5:2a)
the L2 and H 1 norms and the non-linear stability of the scheme was demonstrated subject to certain constraints on the
size of the timestep in terms of the mesh spacing. Our initial aim in this chapter is to describe the spectral Lagrange- u0 2 V; (5:2b)
Galerkin approximation of the Navier-Stokes equations, and to derive error estimates that correspond to those in [76]. nd the strong solution u satisfying
The stability and convergence results obtained are subject to weaker constraints on the timestep than for the nite u 2 L2(H#2 (
)n) \ L1 (V ); (5:2c)
element version of the method. The e ect of quadrature in the nonlinear case will be the subject of future investigation.
d
dt (u; v) + ((u  r)u; v) +  (ru; rv) = (f; v) 8v 2 V;
We begin by introducing some of the notation that will be employed in the remainder of the chapter. Section 2 (5:2d)
describes the incompressible Navier-Stokes equations and recalls some well known existence and uniqueness results. The
concept of Lagrangian coordinates is also recalled and the Lagrange-Galerkin time discretisation is described. Finally the and
fully discrete method is presented, based on a spectral spatial discretisation. Section 3 is devoted to the convergence u(0) = u0 : (5:2e)
analysis of the spectral Lagrange-Galerkin method, beginning with some technical lemmata followed by statements and The following theorem encapsulates some standard existence and uniqueness results for this problem. Again we refer
proofs of the error estimates. Convergence is shown directly; no particular concept of stability is explicitly referred to. to [85, 86] for further details.
However, the stability of the method is still important, and so is dealt with in the next section. There we brie y discuss
the concept of non-linear stablity employed, and state and prove the relevant stability result. The penultimate section Theorem 5.1 For n = 2 there exists a unique solution to problem (5.2) satisfying
describes the application of the method to the two-dimensional Navier-Stokes equations. The particular setting is a study
of a perturbed double shear layer. We end this chapter by setting out the conclusions of the Thesis and indicating areas u 2 L2 (H#2 (
)n) \ C (V ) \ H 1 (H ): (5:3)
of future development for the work.
For n = 3, given f 2 L1 (H ) instead of (5.2a), the same result holds for T small enough.
5.1.1 Notation Under the assumptions of Theorem 5.1, Dt u = @u=@t +(u r)u, the material derivative of u, belongs to L2 (L2(
)n),
We begin with some de nitions speci cR to this chapter. We de ne L_ 2 (
) (resp. H_ #s (
), s  0) as the set of all functions so that (5.2d) may be rewritten as
u in L2(
) (resp. H#s (
)) such that
udx = 0. (Dtu; v) +  (ru; rv) = (f; v) 8v 2 V: (5:4)
Let  = (1; : : : ; n) denote the unit outward normal to @
and let i (resp. i+n ) be the face of
with i = 1 The crucial aspect of the Lagrange-Galerkin approach is the discretisation of the material derivative along particle
(resp. i = 1). Two spaces central to our discussion will be trajectories. To this end we seek to cast the equations into a Lagrangian form.
H := fu 2 L_ 2(
)njr  u = 0; uij i = uij i+n ; i = 1; : : : ; ng; 5.2.2 Lagrangian form
where the latter condition is that u   , with  the outward normal, is periodic (the trace of u   on @
exists when We assume here that the solution u of (5.2) satis es, in addition to (5.3),
u 2 L2 (
)n and r  u 2 L2(
), [86]); and
u 2 C (C 0;1(
)n): (5:5)
V := fu 2 H_ #1 (
)njr  u = 0g:
Using von Neuman's measurable selection theorem, it is possible to give a description of a Lagrangian representation of
For s  1 we set Vs = H#s (
)n \ V . We de ne a norm on Vs equivalent to that induced by H#s (
)n by the ow under weaker assumptions on u (see for example [87]). Since, however, we shall need u to satisfy (5.5) for most
X !1=2 of the results in this chapter, we are happy to assume it here.
For x 2 Rn , t 2 [0; T ], let Xu (x; t; ) denote the trajectory of the particle of uid whose motion is governed by the
kukVs = jpj2s ju^(p)j2 : velocity eld u and which is at position x at time t. Then Xu (x; t; ) is the solution of the initial value problem
p
d
Finally we denote by C 0;1 = C 0;1 (
) the space of Lipshitz-continuous functions on the closure of
. ds Xu (x; t; s) = u(Xu (x; t; s); s); s 2 [0; T ]nftg; (5.6a)
Xu (x; t; t) = x: (5.6b)

62 63
The map x ! Xu (x; t; s) is, for each s; t 2 [0; T ], an isometric homeomorphism from Rn onto itself, and by virtue Our rst error estimate will be obtained from (5.9), taking v =  k . First, however, we present some preliminary results.
of Theorem 2.1 it is di erentiable almost everywhere. Moreover, since u is divergence-free, the map has the volume-
preserving property, i.e. its Jacobian is equal to 1 almost everywhere. Lemma 5.1 Let v 2 S_ N = N L_ 2 (
)n, n = 1; 2; 3. Then
Associated with Xu (x; t; s), for each t; s 2 [0; T ], we can de ne the linear operator Eu (t; s) on H_ #0 (
)n (which kvk1  Dn (N ) kvk1 ;
coincides exactly with L_ 2 (
)n) as follows. Each v 2 H_ #0 (
)n may be periodically extended to the whole of Rn . We then
de ne (Eu (t; s)v)(x) := v(Xu (x; t; s)), x 2 Rn . The restriction of Eu (t; s)v to
is in H_ #0 (
)n. Moreover Eu (t; s) is where
invertible (with inverse Eu (s; t)) and, because of the volume preserving property, D1 (N )  1=2 ;
kEu (t; s)vk = kvk : D2 (N )  [(2 + 4 ln(1 + N 2 ))]1=2;

Under the assumptions of Theorem 5.1, together with the additional assumption (5.5), we may replace the term Dt u D3 (N )  [(9 + 12 ln(1 + N 2 ) + 32p3N )]1=2 :
in (5.4) by dsd u(Xu (; t; s); s) s=t. The rst step in obtaining the discrete equations described in the next subsection
is to replace this time derivative by a rst-order backward di erence formula. Higher order backward di erence formulae
may also be used; their properties in the non-linear case will be the subject of future investigations. Proof . We note that
X
5.2.3 Spectral Lagrange-Galerkin approximation kvk1  jv^(p)j
6 jpj1N
0=
Having described how we will discretise (5.4) in time, we form our fully discrete equations by employing a spectral 2 31=2
Galerkin method for our spatial discretisation. Thus we seek U = (U 0 ; : : : ; U M )T 2 (N V )M +1 satisfying
4 X
1 k 1  (1 + jpj ) 5 kvk1
2 1
k k k 1 k 1 k
t (U ; v) +  (rU ; rv) = t (EU k 1 (t ; t )U ; v) + (f ; v)
6 jpj1N
0=
8v 2 VN = N V  V; k = 1; : : : ; M; (5.7) =: Dn (N ) kvk1 ;
with U 0 = N u0 , where, for x 2
and t 2 [0; T ], we de ne XU k 1 (x; t; s) by and the result follows readily.
d k 1
Lemma 5.2 For k = 1; : : : ; M , let ck = uk C 0;1 : then
ds XU k 1 (x; t; s) = U (XU k 1 (x; t; s)); s 2 [0; T ]nftg; (5.8a)
XU k 1 (x; t; t) = x; (5.8b)
kXu XU k 1 k  c 1 (eck 1 t 1)(  k
+ du
dt L1 (tk 1 ;tk ;L2 (
)n ) ): (5:10)
1
and then EU k 1 (t; s) is de ned by k 1
(EU k 1 (t; s)w)(x) = w(XU k 1 (x; t; s)); x 2
:
We note that, similarly to Eu (t; s), EU k 1 (t; s) is volume-preserving and invertible (with inverse EU k 1 (s; t)). These Proof Integrating (5.6) and the corresponding equation for XU k 1 between tk t and tk we obtain
properties will be used repeatedly in the subsequent analysis.
Given the solution U k 1 at time tk 1 , XU k 1 (x; tk ; tk 1 ) is found by solving (5.8) to whatever accuracy is required. (Xu XU k 1 )(; tk ; tk t)
The right hand side of (5.7) is then well de ned, giving a diagonal system to solve for U k . In practice, the solution of Z tk t
(5.8) is carried out by using a forward Euler discretisation to reduce it to a functional equation, which is solved at equally = Eu (tk ; s)u(s) EU k 1 (tk ; s)U k 1 ds
tkZ
spaced values of x by Newton iteration. These values of x are the quadrature points (for the trapezium rule) which t
are then used in the evaluation of the integrals implicit in the right hand side of (5.7). An additional approximation = fEU k 1 (tk ; tk s)(uk 1 U k 1)
is employed, taking the form of a piecewise Chebyshev interpolant of U k 1 , in order to speed up the evaluation of 0
EU k 1 (tk ; tk 1 )U k 1 . The e ect of these approximations is analysed in the linear case in [79] and shown to be minimal. + (Eu EU k 1 )(tk ; tk s)uk 1
Their e ect in the nonlinear case will be the subject of future investigation. + Eu (tk ; tk s)(u(tk s) uk 1 )gds;
so that
5.3 Error analysis (X X k 1 )(; tk; tk t)
u
In what follows, C1; C2 ; : : : will stand for speci c constants whose values will often be given explicitly. They will be Z t U
di erent from those constants appearing in previous chapters. For simplicity of presentation, we shall sometimes write  f EU k 1 (tk ; tk s)(uk 1 U k 1 )
u(; t) as u(t), and we de ne without ambiguity, for k = 1; : : : ; M , 0
+ (Eu EU k 1 )(tk ; tk s)uk 1

uk := u(tk );
+ Eu (tk ; tk s)(u(tk s) uk 1 ) gds

Eu := Eu (tk ; tk 1 );
EU k 1 := EU k 1 (tk ; tk 1 );  t(  k 1 + du
dt L1 (tk 1 ;tk ;L2 (
)n ) )
Xu := Xu (x; tk ; tk 1 ); Z t
XU k 1 := XU k 1 (x; tk ; tk 1 ): + c k 1 (X X k 1 )(; tk; tk s) ds:
u U
0
We set  k = N uk U k , k = (I N )uk and  k =  k + k , and derive, from (5.4) and (5.7),
1 k A straightforward application of Gronwall's lemma yields the required result.
k
t ( ; v) +  (r ; rv) Lemma 5.3 For k = 1; : : : ; M , v 2 V , and t; s 2 [0; T ],
= t (EU k 1  k 1 ; v) + 1t ((Eu EU k 1 )uk 1 ; v)
1
k(EU k 1 (t; s) I )vk  jt sj U k 1 krvk : (5:11)
1
+ 1t (EU k 1 k 1 ; v) + ( 1t (uk Eu uk 1 ) Dtuk ; v) 8v 2 VN : (5.9)

64 65
Proof We have Proof We choose v =  k in (5.9) to obtain
Zs d 1 k 2 k 2
t  +  r
(EU k 1 (t; s) I )v = (EU k 1 (t;  )v) d
Zt s d 
= EU k 1 (t;  ) U k 1  rv d:  1t  k  k 1 + 1t (Eu EU k 1 )uk 1  k
t
k Eu uk 1 D uk k
The result then follows by taking the norm of both sides. + 1t j(EU k 1 k 1 ;  k )j + u  t t
The nal lemma deals with the time truncation error in the backward Euler approximation of the material derivative = A1 + : : : + A4 : (5.16)
along particle trajectories. We shall bound the terms A2 to A4 in turn. We have, for A2,
Lemma 5.4 For k = 1; : : : ; M , (E E k 1 )uk 1  c kX X k 1 k ;
u U k 1 u U

(Dtu)k uk Euuk 1
 Dt2u L1(tk 1 ;tk ;L2 (
)n ) : (5:12) and so, from Lemma 5.2,
t
 k 1 + du !
A2  1t (eck 1 t 1) dt L1(tk 1;tk;L2(
)n) k
Proof 
  k 1t (eck 1 t 1)  k 1
uk Eu uk 1 (D u)k
t t
!)
Z tk
d E (tk ; s)u(s) d E (tk ; t)u(t) ds +ck 1 eck 1 t k 1 + du dt L1 (tk 1 ;tk ;L2 (
)n ) ;
= 1t u u t=tk
tk 1 ds dt
Z t k Z tk 2
d E (tk ; t)u(t)dt ds where we have written  k 1 as  k 1 + k 1 .
= 1t u In order to bound A3 we make use of the fact that Eu and EU k 1 are invertible and volume preserving.
tk 1 s dt2
Z t k Thus we have, applying Lemma 5.1 and Lemma 5.3,
= 1t (t tk 1 ) dtd 2 Eu (tk ; t)u(t)dt;
2
t k 1 A3  1t j(k 1 ; (EU k1 1 I ) k)j
and the result follows.
  k 1 U k 1 1 r k

 k 1 r k N uk 1 1 + Dn (N ) r k 1 :

5.3.1 L2 error estimates
Now, by Sobolev's imbedding theorem and the contractivity of N in Vr for any real r, it follows that there is
We assume that a constant C1 , independent of N but dependent on s > n=2, such that kN ukl1 (L1)  C1 kukL1 (Vs ) =:
f 2 C (H ) and u0 2 C 0;1 \ Vs ; s > n=2; (5:13) C2 . Including the bound on A4 which follows from Lemma 5.4 we nd, repeatedly making use of the
and that the corresponding solution u of (5.2) satis es inequality ab  2 a2 + 21 b2 (valid for all real a, b and for  > 0) and gathering like terms, that
u 2 C (C 0;1 \ Vs ); s > n=2; (5.14a) 1 k 2 k 2
du=dt 2 L2 (H ) and (5.14b) t  +  r
ck 1 t 
Dt2 u 2 L2 (H ): (5.14c)  e t  k  k 1 + r k k 1 C2 + Dn(N ) r k 1
( k 1 !
Then we have the following +  k ck 1 eck 1 t du
dt L1 (tk 1 ;tk ;L2(
)n ) + 
Theorem 5.2 Suppose that u satis es (5.2), U satis es (5.7), and that (5.13) and (5.14) hold. Then there exist two o
positive constants t0 and N0 such that, for all t  t0 and all N  N0, + Dt2 u L1(tk 1 ;tk;L2 (
)n )
 
ku U kl1 (H )  C3(C4N s + C5t) (5:15)  21 t  k 2 + e2ck 1 t  k 1 2 + c2k 1  k 2 + 2 r k 2
where 2ck 1 t du 2 !
C3 = expf2 juj2l2 (C 0;1 ) + jujl1 (C 0;1 ) )g; + e 2 dt +  k 1 2
!1=2
2( k 1 D (N ))L12(t k 1 ;tk ;L2 (
)n )

4C12 kuk2L1(Vs ) + n r k 1 2
C4 = kukL1 (Vs ) +  +1 kukl2 (Vs ) ; 
2
+ 1 Dt2 u 2L1(tk 1 ;tk ;L2(
)n ) + 2C 2 k 1 2 ;
 2 1=2 !
C5 = du
dt L2 (H ) + 
Dt2u L2(H) ;
where we have also made use of the fact that kvk  krvk for v 2 V .
and C1 is a positive constant which depends only on s. Because of the smoothness of u, there is an N0 such that for N  N0 ,
 2 2
The remainder of this subsection will be taken up with the proof of this theorem. 2 kkC (H ) Dn(N )  4 :

66 67
We assume that we are dealing with such an N , so that multiplying through by 2t we obtain 5.3.2 Error estimates in V
k 2 k 2 2ck 1 t k 1 2 Theorem 5.3 Suppose that the conditions of Theorem 5.2 hold. Suppose also that N and t are related by
(1
k 1 )  +4C2t r 2 e  t  2
2tc2
 ((ln N ) 1=2); n = 2
+t e2ck 1 t +  2 k 1 + 2 r k 1 t = O (5:24)
2 ! O(N 1=2 ); n = 3;
+t2 e2ck 1 t du 2 2 2 and that t0 , N0 , C1 , C3 , C4 and C5 are as in Theorem 5.2. Then, for t  t0 and N  N0 ,
dt L2(tk 1 ;tk ;L2(
)n ) +  Dt u L2 (tk 1 ;tk ;L2 (
)n ) :
(5.17) ku U kl1(V )  C6(C7N 1 s + C8t) (5:25)
We proceed under the assumption that t  t0 where where 1  
! C6 = exp  C12 kuk2l2 (Vs ) + C9 2 ;
t0 = min T; 1 : (5:18)
4 jujl1 (C 0;1 )
2
!
 
1=2  
For such t, we can bound (1 2tc2k 1 ) from below by e 4tc2k 1 . Then we deduce from (5.17) that C7 = kukl1(Vs ) 1 + 5 C1 kukl2 (Vs ) + C9
k 2 + e4tc2k 1  t rk 2  1=2
 + 5 jujl2(C 0;1 ) C3 C4et0 jujl1(C0;1 ) ;
 e4tc2k 1 +2tck 1  k 1 2 +  2 t e4tc2k 2 r k 1 2
 2  5 1=2 !
+ t 1 + 4C 2 k 1 2 C8 =

et0 jujl1(C0;1 ) jujl1 (C 0;1 ) du + C3 C5 jujl2 (C 0;1 )
2 )  dt L2(L2 (
)n )
+ t2 du + 2t2 D2 u 2 :  1=2
+ 5 Dt2u L2(L2(
)n)
dt L2(tk 1 ;tk ;L2(
)n )  t L 2 (t k 1 ;t k ;L 2 (
)n )

(5.19)
This inequality is valid for all k  1 if we set c 1 = 0. It is slightly less sharp than (5.17), since we have and C9 is an upper bound on Dn(N ) r k l2 (H ) .
multiplied some of the terms on the right hand side of (5.17) by terms greater than unity. Our reason for
doing this is to enable us to make use of the following lemma, whose proof (by induction) is straightforward. Remark. The constant C9 is guaranteed to exist and to be independent of N and t by Lemma 5.6 and (5.24).
k k 1
Lemma 5.5 Let fak g, fbk g, fdk g and fk g be sequences of non-negative real numbers satisfying, for Proof We begin by considering (5.9), adding 1 k 1; v)
t ( to both sides, and taking v =  t , giving
k = 1; : : : ; M ,   k k 1 2  n 2
a k + bk  e  k ak 1 + 21 bk 1 + dk : (5:20) k rk 1 2o
t + 2t r
1 1
Then the following inequality holds: k k 1 
MY1 ! MX1 !   t 1t (EU k 1 I ) k 1 + 1t (Eu EU k 1 )uk 1
X
M

aM + 12 bk  ek a0 + b20 + dk : (5:21) k Eu uk 1 D uk  :
k=0 k=0 k=0 + 1t (EU k 1 I )k 1 + u  t t (5.26)
P 1 2 
If we denote 2t Mk=0 2ck + ck by , then from (5.19){(5.21) we deduce that We deal with each of the terms on the right hand side of (5.26) in turn to obtain
M 2 +  t X
M 1 k 1
e4tc2k 1 r k 2 t (E U k 1 I ) 
2 k=1  U k 1 1 r k 1
  2  k 1 
X
M
e t 1 + 4C 2
2
 k 1  C1 u Vs + Dn (N ) r k 1 r k 1 ; (5.27a)
k=1
du 2 ! 1 k 1
2 2 2 t (Eu EU k 1 )u !
+t2 e dt L2 (L2(
)n ) +  Dt u L2 (L2(
)n ) : (5.22)
 ck 1 eck 1 t  k 1 + du
dt L1 (tk 1 ;tk ;L2 (
)n ) ;
(5.27b)
Since the same bound holds for  k 2 , for any k  M , the conclusions of the theorem follow.
1 k 1
Remark. The result is such that, given sucient smoothness and decay of u, it is valid for a semi-in nite time interval. t (E U k 1 I ) 
Remark. (5.22) also provides us with a bound on the l2 norm of r , which we encapsulate in the following lemma, and  U k 1 1 r k 1
will make use of in the next subsection.  k 1 
Lemma 5.6 Under the conditions of Theorem 5.2, and with C4 and C5 as de ned there, we have  C1 u Vs + Dn (N ) r k 1 rk 1 ; (5.27c)
 2 1=2 uk E uk 1
rk  
C3 C4N s + C5t : (5:23)
u Dt uk
l2(H )   t

 Dt2 u L1 (tk 1 ;tk ;L2 (
)n ) : (5.27d)

68 69
Substituting (5.27a){(5.27d) into (5.26) yields 5.4.2 Reformulation of the problem
 n k 2 r k 1 2 o In the previous section the discrete equations were parameterised by small positive numbers h. In our case, however, there
2t r are two parameters, t and N , associated with the discrete equations. We thus denote by h the pair (t; 1=N ) 2 (0; 1)2.

 r k 1 2 + 52 rk 1 2  C 2 uk 1 2 + D (N )2 rk 1 2 In fact we invoke the limitations on t and N involved in Theorem 5.2, so that h 2 (0; t0 ]  (0; 1=N0].
1 Vs n The spaces Xh and Yh will be distinguished only by their norms. They will both consist of vectors
 k 1 2 + t du 2 !  M +1
+ 25 c2k 1 e2ck 1 t dt L2(tk 1;tk;L2(
)n) U = (U 0 ; : : : ; U M )T 2 S_ N :

+ 52 t Dt2u 2L2 (tk 1 ;tk ;L2(
)n ) : (5.28) For U 2 Xh we de ne
kUkXh := kUkl1 (H ) +  1=2 kUkl2 (V ) ;
and for U 2 Yh we de ne
t  2 k + D (N )2 rk 1 2 :
Let us write
k 1 = 2 1 2
M !1=2
X
 C1 u Vs n
kUkYh := U 0 + t U k 2 :
Then, multiplying (5.28) by 2t= , we obtain k=1
rk 2  ek 1  rk 1 2 + 5k 1 rk 1 2 We suppose that u satis es (5.2), (5.13) and (5.14), for some s > (n + 2)=2. Then we set our target element
2 uh 2 Xh to be N u. We are now in a position to de ne the operator h : Xh ! Yh . Let V = (V 0 ; : : : ; V M )T 2 Xh .
! For each k = 1; : : : ; M , we have V k 1 ; V k 2 S_ N , and we construct F k 2 S_ N as follows. For (x; t) 2
 [0; T ] we
5 t 
+  e ck 1 e1 2 2c 1 t  k 1 2 + t du 2 de ne XV k 1 (x; t; ) to be the solution of the initial value problem
k k
dt L2(tk 1 ;tk ;L2 (
)n )
d k 1

+ 5t ek 1 D2 u 2
2
t : ds XV k 1 (x; t; s) = V (XV k 1 (x; t; s)); s 2 [0; T ]nftg;
 L2(tk 1 ;tk;L2 (
)n ) XV k 1 (x; t; t) = x:
A simple application of Lemma 5.5 completes the proof of Theorem 5.3. We then de ne, for s; t 2 [0; T ], EV k 1 (t; s) : H ! H analogously to EU k 1 (t; s) in the discussion of convergence, and
similarly denote EV k 1 (tk ; tk 1 ) by EV k 1 . Furthermore we de ne F k 2 S_ N to be the solution of
5.4 Stability (F k ; v) = 1t ((V k EV k 1 V k 1 ); v) +  (rV k ; rv) (f k ; v) 8v 2 VN ;
Although in the previous section we have demonstrated the convergence of the spectral Lagrange-Galerkin method for
the Navier-Stokes equations, we have not yet addressed the pertinent question of the stability properties of the scheme. and F 0 to be the solution of
These properties become important, for example, in the presence of rounding errors. Here we carry out a stability 
analysis within the framework introduced by Lopez-Marcos and Sanz-Serna [50], and we begin with a brief discussion of (F 0 ; v) = (V 0 N u0 ; v) + 21  t r(V 0 N u0 ); rv 8v 2 VN :
this framework.
In this way we construct F = (F 0 ; : : : ; F M )T from V , and we de ne h : Xh ! Yh by
5.4.1 De nition of stability h (V ) = F :
Let u be a solution of the equation (u) = 0, where  is a mapping from a Banach space X into a Banach space
Y . (The Navier-Stokes equations as described earlier may be set in this form). Let H be a set of positive numbers (or The spectral Lagrange-Galerkin method for the Navier-Stokes equations may then be succinctly described as follows:
vectors in R2 with positive entries) with zero in mum. For each h 2 H, let Uh be a numerical approximation to u, nd U 2 Xh such that
obtained by solving h (U ) = 0: (5:31)
h (Uh ) = 0; (5:29)
where h is a xed mapping with domain Dh 2 Xh and taking values in Yh , with both Xh and Yh being nite-dimensional 5.4.3 Stability analysis
linear spaces. We choose a norm kkXh in Xh , a norm kkYh in Yh and a `target element' uh in the interior of Dh , Theorem 5.4 There is a pair of positive numbers (t0 ; 1=N0) such that for N  N0 and t  t0 the spectral
which should be some discrete representation of the analytical solution u in Xh . Then we have the following de nition Lagrange-Galerkin method (5.31) is unconditionally non-linearly stable with stability threshold Rh = (NDn(N )) 1.
of stability for (5.29) (c.f. [50]).
De nition 5.1 For h 2 H, let Rh 2 (0; 1]. The discretisation (5.29) is said to be stable, restricted to the thresholds Proof Choose V ; W 2 Xh \ B (uh ; Rh ), and write F = h (V ) and G = h (W ). For k = 0; : : : ; M , we
write  k = V k W k . Thus, for k = 1; : : : ; M and for all v 2 VN , we have
Rh , if there exist positive constants h0 and S such that for h 2 H, jhj  h0 , the open ball B (uh ; Rh ) is contained in
Dh and for any Vh ; Wh in that ball, 1 k k 1 E k 1 W k 1 ); v +  (r k ; rv) = (F k Gk ; v);
t  (EV k 1 V W
kVh Wh kXh  S kh (Vh ) h (Wh )kYh : (5:30)
so that, taking v =  k , we nd
The motivation for this de nition was in part to introduce a concept of stability that could be used in a general \stability 
1  k 2 +  r k 2   k 1 E k 1 V k 1 E k 1 W k 1
+ consistency ) convergence" type of Theorem. Knowing this, it would have been natural to have sought to prove the V W
t
convergence of the method by this well-trodden route. However, although such a strategy would have indeed produced
a convergence result, the assumptions of the smoothness of the exact solution u would have had to be more severe. t
+ F k Gk : (5.32)
The di erence between this and the problem-speci c proof given above is the introduction of the `target element' in the
above general de nition of stability. The key part of the proof is the bound on EV k 1 V k 1 EW k 1 W k 1 . We write
In order to make use of this de nition we have to cast the spectral Lagrange-Galerkin method into an appropriate
form. This is the aim of the next subsection.
E k 1 V k 1 E k 1 W k 1   k 1 + (E k 1 E k 1 )W k 1 ;
V W V W

70 71

where we have bounded EV k 1  k 1 by  k 1 . The second term may be bounded by kXV k 1 XW k 1 k W k
1 ,
C 0;1
where j  jC 0;1 is the Lipschitz seminorm, and it is relatively straightforward to show, by use of Gronwall's No. of time steps L2-error Order of convergence
lemma, that 10 2:65  10 2
XW k 1 k  t  k etjW jC0;1 :
k 1
kXV k 1 1 15 7:04  10 3 3.26
25 1:15  10 3 3.55
Under the assumptions that we have on t and W , and making use of Sobolev's embedding theorem, we 50 8:48  10 5 3.76
have 100 5:69  10 6 3.90

t W k
k
N uk 1 V2 + t N uk

C 0;1  tDn (N ) W
1 1 1
C 0;1 Table 5.1: Results for the rst Navier-Stokes experiment
1=
 t NDn (N )Rh + tC10 kukl1(Vs ) ;
2 (5.33)

etj W j C0;1 may be bounded by a constant C11. We


k 1
and so according to the assumptions of the theorem where
deduce from (5.32), making use of the relation  k  r k , that F m = N PL f (tm ); (5:37d)
1 k 2 +  r k 2  1   k 2 +  k 1 2  +  r k 2 and where PL E (tm ; tm 1 ) may be implemented by solving
t   2 k 2
+ 1 C11
t 2

W 1 2C 0;1  k 1 2 + F k Gk 2 : (5.34) x X = t
X
q
i PK;L U m i;i 1 (X );  2 2L 1 : (5:37e)
i=1
Multiplying through by 2t we obtain The coecients i , i and i are as in (4.27) and (4.67). The calculation is commenced by means of the initialisation
procedure described in Chapter 4. Once this is complete the solution is advanced in time by means of the qth-order
 k 2 +  t r k 2   k 1 2 1 + C112 t W k 1 2  algorithm with constant timestep. Each particular time step is performed as follows.
P
 C 0;1 The rst task is the solution of (5.37e). The Fourier coecients of the sum qi=1 i U m i;i 1 are formed and used
2 t
+  F k Gk :
2
(5.35)
to form the right hand side of (5.37e), which is then solved by means of the modi ed Newton iteration described in
the context of Burgers equation in Chapter 4. The calculation is performed to a predetermined tolerance (TOL1). In
2 t k 1 2 2 practice, no more than three iterations have ever been needed to achieve machine accuracy in this calculation.
Bounding 1 + C11
 W C 0;1 by expf C11t W k 1 2C 0;1 g and applying Lemma 5.6 gives that Once this is known, the coecients of the term
"X #
 1=2 C112 2 N PL E (tm ; tm 1 )PK;L
q
i U m i;i t 1 (rP m Fm  U m 1;0 )
WkXh  2 1 +  e 2 jWjl2 (C0;1) kh (V ) h (W )kYh :
1 1 1
kV (5:36) i=1

We bound W k 1 l2 (C 0;1 ) by employing an argument identical to (5.33), and the stability result follows. may be calculated and stored. It is then a simple matter to calculate U m;0 by including the forcing term, applying R,
and inverting the left hand side of (5.37a), and to update the other terms U m i;i and calculate the pressure according
to (5.37b) and (5.37c) respectively.
5.5 Numerical experiments The operator PK;L is used repeatedly in this process. The time-saving devices described in Chapter 4 are used, thus
introducing a second tolerance parameter (TOL2) which is again chosen at the beginning of the calculation.
In order to be able to complete the analysis of this chapter, we have had to restrict ourselves to considering the
exactly-integrated version of the method with a backward-Euler time step. In practice quadrature must be introduced,
the additional local interpolation may be used to improve eciency, and higher-order timestepping procedures may be
5.5.2 The experiments
incorporated. All of these elements are included in the calculations presented below, and we shall now describe the The experiments we have conducted thus far have been primarily for the purpose of veri cation of the code. More
algorithm in detail, giving the formulae which serve to de ne the method, and discussing some of the details of the extensive experimentation will be carried out and presented elsewhere.
computational procedure. The rst results show that the time-accuracy of the method is as required. It is easy to design a problem with a
predetermined exact solution by choosing the forcing function appropriately. In this experiment the exact solution is
5.5.1 The algorithm
 cos ye t+sin x+sin y 
u(x; y; t) = cos xe t+sin x+sin y ; (5.38a)
Let R be the L2-projection from S_ N to the subset VN consisting of functions with zero divergence and zero mean. R
may be identi ed with the operator (I r 1 r), and so may be applied straightforwardly. The canonical basis for p(x; y; t) = e t+sin x+sin y ; (5.38b)
VN is fRp g0=6 jpj1N . Functions are represented by the coecients of their expansions in these basis functions. The and this is generated by the forcing function
approximation is obtained by rst discretising in time the Lagrangian form of (5.1) and then projecting onto divergence-  (cos x(1 + sin y) cos y  cos y(cos2 x sin x sin2 y 3 sin y)) 
free nite-dimensional subspaces. The full set of formulae for the qth order method is: f (x; y; t) =
"X
q
(cos x + cos y(1 + sin x) +  cos x(cos2 y sin y sin2 x 3 sin x)) ; (5:39)
(I  t 0 )U m;0 = RN PL E (tm ; tm 1 )PK;L i U m i;i 1 where = e t+sin x+sin y .
i=1  The calculations were carried out using a fourth-order backward di erence scheme with the following parameters:
t 1 (rP m 1 F m 1  U m 1;0 ) N = 31
+t 0 RN PL F m ; (5.37a) L = 32
U m i;i = N PL E (tm ; tm 1 )PK;L U m i;i 1 ; i = 1; : : : ; q 1; (5.37b)
"X
q
K = 5
t 0 rP m = (I R)N PL E (tm ; tm 1 )PK;L i U m i;i 1 T = 1
i=1   = 10 4
t 1 (rP m 1 F m 1  U m 1;0 ) TOL1 = 10 8
+t 0 (I R)N PL F m ; (5.37c) TOL2 = 10 8 ;

72 73
and the results for various values of t are given in Table 5.1. 5.6.1 Future developments
The second experiment is somewhat more realistic. We consider the evolution of a perturbed double shear layer,
generated by an initial vorticity distribution The most obvious extension of this work is the consideration of non-periodic problems. The major diculty here will be
the incorporation of boundary conditions, a long-standing thorn in the side for some Lagrangian methods. It is hoped
!(x; y) = sech2 (10(y )) sech2 (10(y )) + 10 3 cos(2x) + 10 4 cos(4x); (5:40) that the detailed exploration of the properties of the spectral Lagrange-Galerkin method in the absence of boundaries
contained herein will have shed light on e orts being made to apply the method to such problems.
with = (1 1=8) and = (1 + 1=8). This problem is considered by Kreiss et. al. [43], where they use a As already mentioned, the algorithm is highly amenable to implementation on shared-memory parallel machines, and
spectral discretisation of the equations in vorticity-velocity form combined with a fourth-order Runge-Kutta method for in such a setting would provide a highly competitive method for the solution of the Navier-Stokes equations. There is a
the discretisation in time. Our calculation was carried out with the following parameters: long history of the use of Fourier methods in the modelling of two-dimensional turbulence. The spectral Lagrange-Galerkin
method is a fresh approach and could make a signi cant contribution to the area.
N = 63 A nal area of interest is in the analysis of the method as applied to non-linear problems with non-trivial dynamics.
L = 64 There is an increasing awareness of the need to ensure that a numerical scheme mirrors the dynamics of the original
K = 5 system accurately, and the approach of timestepping along particle trajectories as opposed to timestepping along straight
lines may signi cantly alter the dynamics of a numerical method. We feel that there is room for further investigation in
T = 150 this area; the stability results of this chapter may be a rst step in this direction.
t = 5  10 2
 = 10 4
TOL1 = 10 8
TOL2 = 10 8 :
The results, shown in the form of contour plots of the vorticity in Figs. 5.1{5.4, are visibly indistinguishable from those
obtained in [43]. They show a wake ow unstable to the perturbations imposed upon it. The cos(4x) perturbation
governs the initial formation of a von Karman vortex street, but eventually the cos(2x) perturbation grows and these
vortices pair o into a nal stable state.

5.6 Conclusions
We have in this Thesis presented and analysed a new numerical method for the solution of convection-dominated di usion
problems. The scheme is based on a Lagrangian-Eulerian approach that rewrites the equations in Lagrangian coordinates
that take care of the convective terms, and avoids the distortion that might cause diculties in dealing with the di usion
and other terms by using a di erent set of Lagrangian coordinates at each time step.
This approach has been widely and successfully implemented in combination with a variety of spatial discretisation
methods. The use of spectral methods is relatively new. In our case it was motivated by the stability problems faced
by the nite-element version of the method when quadrature was introduced. Our analysis has shown that the spectral
version remains unconditionally stable, even with the introduction of quadrature.
The combination of spectral methods with Lagrangian approaches su ers from the drawback of unreasonable expense
due to the distortion of the coordinates. We have introduced a new procedure by which this can be circumvented with
negligible loss in accuracy. The algorithm is such that it could be readily implemented on a shared memory parallel
machine, and would be able to make full use of a modestly parallel computer, with 10{20 processors. On such a machine
the method would be highly competitive. This new procedure is based on the use of a local Chebyshev interpolation. Its
application is more general than just the spectral Lagrange-Galerkin method. It will be of use whenever it is required to
evaluate a trigonometric (or a Chebyshev) polynomial at an unstructured set of points, as might be the case if one were
to seek to track a set of particles released into a velocity eld which was represented in terms of such a polynomial.
In order to capitalise upon the high accuracy of spectral methods for space discretisation, a highly accurate time
discretisation is required. We have described and analysed multistep methods (speci cally backward di erence methods)
for the time discretisation of the equations in Lagrangian form. These give rise to fully discrete schemes with well
balanced error terms without the need to use excessively small timesteps. Their implementation in non-linear problems
involves tracking the `particle trajectories' with corresponding accuracy. It was found that the most convenient way to
do this was by means of an Adams-Bashforth formula used to extrapolate along the trajectory itself. This then gives
rise to set of implicit problems which must be solved by some iterative scheme. In the search for greater eciency we
have introduced a modi ed Newton iteration, which leads to savings in time of at least 50 percent because of the nature
of the local interpolation used in evaluating the functions involved. We are not aware of any other applications for this
iterative procedure.
The culmination of the development of the method in this Thesis is its application to the Navier-Stokes equations.
The analysis, applied to a slightly idealised form of the method, has shown it to be convergent (with optimal order of
convergence) and non-linearly stable. Moreover, with suitable decay of the exact solution, these results hold for in nite
time. Alongside the theory, the method has been made into a computer code, which has been successfully implemented
on a simulation of two-dimensional wake instability.

74 75
Figure 5.1: Contour plot of initial data. Figure 5.2: Contour plot at time t = 50.

76 77
Figure 5.3: Contour plot at time t = 100. Figure 5.4: Contour plot at time t = 150.

78 79
[22] Ewing, E. R. and Russell, T. F. (1981). Multistep Galerkin methods along characteristics for convection-di usion
problems. In Advances in Computer Methods for Partial Di erential Equations, Vichnevetsky, R. and Stepleman, R. S.,
Eds., pp 28{36. IMACS, Rutgers University.
[23] Farmer, C. L. (1985). A moving point method for arbitrary Peclet number multi-dimensional convection-di usion
problems. IMA J. Numer. Anal., 5:465{480.
Bibliography [24] Federer, H. (1969). Geometric Measure Theory. Berlin Heidelberg New York: Springer.
[25] Fishelov, D. (1990). A new vortex scheme for viscous ows. J. Comp. Phys., 86:211{224.
[26] Foias, C., Guillope, C. and Temam, R. (1985). Lagrangian representation of a ow. J. Di . Eq. 57:440{449.
[27] Fritts, M.J., Crowley, W.P. and Trease, H.E. eds., (1985). Free-Lagrange methods for compressible hydrodynamics
[1] Adams, R.S. (1975). Sobolev Spaces. Academic Press, New York. in two space dimensions. Lecture Notes in Physics 238, The Free-Lagrange Method, Springer-Verlag.
[2] Anderson, C. and Greengard, C. (1985). On vortex methods. SIAM J. Numer. Anal., 22:413{440. [28] Furzeland, R.M., Verwer, J.G. and Zegeling, P.A. (1990). A numerical study of three moving-grid methods for
[3] Ansorge, R. (1963). Die Adams-Verfahren als Characteristikenverfahren Ordnung zur Losung von hyperbolischen one-dimensional partial di erential equations which are based on the method of lines. J. Comp. Phys., 89:349{388.
Systemen halblinearer Di erentialgleichungen. Numer. Math., 5: 443{460. [29] Gear, C.W. (1971). Numerical Initial-Value Problems in Ordinary Di erential Equations. Prentice-Hall.
[4] Baines, M.J. (1991). Analysis of the moving nite element procedure. SIAM J. Numer. Anal., To appear. [30] Gottlieb, D. and Tadmor, E. (1991). The CFL condition for spectral approximations to hyperbolic initial-boundary
[5] Barrett, J. W. and Morton, K. W. (1984). Approximate symmetrisation and Petrov-Galerkin methods for di usion- value problems. Math. Comp., 56:565{588.
convection problems. Comp. Meth. Appl. Mech. Engrg., 45:97{122. [31] Gottlieb, D. and Turkel, E. (1980). On time discretisation for spectral methods. Stud. Appl. Math., 63:67{86.
[6] Beale, J. T. and Majda, A. (1982). Vortex methods. I: Convergence in three dimensions. Math. Comp., 39:1{27. [32] Hald, O. and Del Prete, V.M., (1978). Convergence of vortex methods for Euler's equations. Math. Comp.
[7] Beale, J. T. and Majda, A. (1981). Rates of convergence for viscous splitting of the Navier-Stokes equations. Math. 32:791{809.
Comp., 37:243{259. [33] Hald, O. (1979). Convergence of vortex methods for Euler's equations. II SIAM J. Numer. Anal., 16:726{755.
[8] Benque, J. P., Labadie, G., and Ronat, J. (1982). A new nite element method for Navier-Stokes equations coupled [34] Handscomb, D.C. (1989). Private communication.
with a temperature equation. In Proc. 4th Int. Symp. on Finite Element Methods in Flow Problems, Kawai, T., Ed.,
pp 295{301. North Holland. [35] Hartman, P., (1964). Ordinary Di erential Equations. New York London Sydney: Wiley.
[9] Bercovier, M. and Pironneau, O. (1982). Characteristics and the nite element method. In Proc. 4th Int. Symp. on [36] Hasbani, Y., Livne, E., and Bercovier, M. (1983). Finite elements and characteristics applied to advection-di usion
Finite Element Methods in Flow Problems, Kawai, T., Ed., pp 67{73. North Holland. equations. Comput. & Fluids, 11:71{83.
[10] Bermejo, R. (1990). On the equivalence of semi-Lagrangian schemes and particle-in-cell nite element methods. [37] Ho, L-W., Maday, Y., Patera, A.T. and Rnquist, E.M. (1989). A high-order Lagrangian-decoupling method for
Mon. Wea. Rev., 118:979{987. the incompressible Navier-Stokes equations. ICASE Report No. 89-57.
[11] Cai, W., Gottlieb, D. and Shu, C-W. (1989). Essentially nonoscillatory spectral Fourier methods for shock wave [38] Holly Jr., F.M. and Preissmann, A. (1977). Accurate evaluation of transport in two dimensions. J. Hydraul. Div.
calculations. Math. Comp., 52:389{410. ASCE, 98:1259{1277.
[12] Canuto, C., Hussaini, M. Y., Quarteroni, A. and Zang, T. A. (1988). Spectral Methods in Fluid Dynamics. Springer- [39] Hughes, T. J. R. and Brooks, A. (1979). A multi-dimensional upwind scheme with no crosswind di usion, In Finite
Verlag. element methods for convection-dominated ows, Hughes, T. J. R., Ed., AMD Vol. 34, ASME, New York., (1979),
pp 19{35.
[13] Childs, P.N. and Morton, K.W. (1990). Characteristic Galerkin methods for scalar conservation laws in one dimen-
sion. SIAM J. Numer. Anal., 27:553{594. [40] Jack, R. O. (1987). Convergence properties of the Lagrange-Galerkin method with and without exact integration.
OUCL Report No. 87/10, Oxford.
[14] Chorin, A.J. (1973). A numerical study of slightly viscous ow. J. Fluid Mech. 57:785{796.
[41] Jack, R. O. (1988). Stability of the Lagrange-Galerkin method: the performance of quadrature in theory and
[15] Cottet, G-H. (1989). A particle-grid superposition method for the Navier-Stokes equations. J. Comp. Phys., practice. OUCL Report No. 88/15, Oxford.
89:301{318.
[42] Johnson, C., Szepessy, A. and Hansbo, P. (1990). On the convergence of shock-capturing streamline-di usion nite
[16] Cottet, G-H. (1991). Large-time behaviour of deterministic particle approximations to the Navier-Stokes equations. element methods for hyperbolic conservation laws. Math. Comp., Vol. 54, No. 189, pp. 107{129.
Math. Comp., Vol. 56, No. 193, pp. 45{59.
[43] Kreiss, H.-O., Henshaw, W.D. and Reyna, L.G. (1988). On the smallest scale for the incompressible Navier-Stokes
[17] Courant, R., Isaacson and E, Rees, M. (1952). On the solution of non-linear hyperbolic di erential equations by equations in two dimensions. ICASE Report No. 88-8.
nite di erences. Comm. Pure Appl. Math., 5: 243{264.
[44] Kreiss, H.-O. and Oliger, J. (1979). Stability of the Fourier method. SIAM J. Numer. Anal., 16:421{433.
[18] Crouzeix, M. (1980). Une methode multipas implicite-explicite pour l'approximation des equations d'evolution
paraboliques. Numer. Math., 35: 257{276. [45] Leonard, A. (1980). Vortex methods for ow simulations. J. of Comp. Phys., 37:289{335.
[19] Donea, J. (1984). A Taylor-Galerkin method for convective-transport problems. Int. J. Numer. Meth. Engrg., [46] Le Roux, M-N. (1979). Semi-discretisation en temps pour les equations d'evolution paraboliques lorsque l'operateur
20:101{119. depend du temps. R.A.I.R.O. Numerical Analysis, Vol. 13, No. 2, pp. 119{137.
[20] Douglas, Jr., J. (1983). Finite di erence methods for two phase incompressible ow in porous media. SIAM J. [47] Le Roux, M-N. (1979). Semidiscretisation in time for parabolic problems. Math. Comp., Vol. 33, No. 147, pp.
Numer. Anal., 20:681{696. 919{931.
[21] Douglas, Jr., J. and Russell, T. F. (1982). Numerical methods for convection-dominated di usion problems based [48] Le Roux, M-N. (1980). Methodes multipas pour des equations paraboliques non lineaires. Numer. Math., 35:143{
on combining the method of characteristics with nite element or nite di erence procedures. SIAM J. Numer. Anal., 162.
19:871{885.

80 81
[49] Lesaint, P. (1977). Numerical solution of the equation of continuity. In Topics in Numerical Analysis III, J.J.H. [76] Suli, E. (1988). Convergence and nonlinear stability of the Lagrange-Galerkin method for the Navier-Stokes equa-
Miller, Ed., Academic Press, London, New York, San Francsico, pp. 199-222. tions. Numer. Math., 53:459{483.
[50] Lopez-Marcos, J. and Sanz-Serna, J. (1987). Stability and convergence in numerical analysis III: Linear investigation [77] Suli, E. (1988). Stability and convergence of the Lagrange-Galerkin method with non-exact integration. In The
of nonlinear stability. IMA J. Numer. Anal. 8: 71{84. Mathematics of Finite Elements and Applications VI, Whiteman, J. R., editor, pages 435{442. Academic Press.
[51] Mas-Gallic, S. (1990). Deterministic particle methods: di usion and boundary conditions. Internal report 90029, [78] Suli, E. and Ware, A. (1988). A spectral method of characteristics for rst-order hyperbolic equations. OUCL
Analyse Numerique, Universite Pierre et Marie Curie. Report No. 88/6, Oxford.
[52] Mercier, B. (1989). An Introduction to the Numerical Analysis of Spectral Methods. Lecture Notes in Physics 318. [79] Suli, E. and Ware, A. (1991). A spectral method of characteristics for rst-order hyperbolic equations. SIAM J.
Springer-Verlag. Numer. Anal. Vol. 28, No. 2, pp. 423{445.
[53] Miller, K. and Miller, R.N. (1981). Moving nite elements part I. SIAM J. Numer. Anal., 18:1019{1032. [80] Tadmor, E. (1986). The exponential accuracy of Fourier and Chebyshev di erencing methods. SIAM J. Numer.
[54] Miller, K. (1981). Moving nite elements part II. SIAM J. Numer. Anal., 18:1033{1057. Anal., Vol. 23, pp. 1{10.
[55] Mizohata, S. (1973). The Theory of Partial Di erential Equations. Cambridge University Press. [81] Tadmor, E. (1987). Stability analysis of nite di erence, pseudospectral and Fourier-Galerkin approximations for
time-dependent problems. SIAM Review, Vol. 29, pp. 525{555.
[56] Morton, K. W. and Parrott, A.K. (1980). Generalised Galerkin methods for hyperbolic problems. J. Comput. Phys.,
36:247{270. [82] Tadmor, E. (1989). Convergence of spectral methods for nonlinear conservation laws. SIAM J. Numer. Anal., Vol.
26, pp. 30{44.
[57] Morton, K. W., Priestley, A., and Suli, E. (1988). Stability of the Lagrange-Galerkin method with non-exact
integration. RAIRO M2 AN, vol. 22, no. 4, pp. 123{151. [83] Tal-Ezer, H. (1986). Spectral methods in time for hyperbolic equations. SIAM J. Numer. Anal., Vol. 23, pp. 11{26.
[58] Neuman, S.P. (1981). A Eulerian-Lagrangian numerical scheme for the dispersion-convection equation using con- [84] Tal-Ezer, H. (1989). Spectral methods in time for parabolic equations. SIAM J. Numer. Anal., Vol. 26, pp. 1{11.
jugate space-time grids. J. Comp. Phys., 41:270{294. [85] Temam, R. (1983). Navier-Stokes Equations and Nonlinear Functional Analysis. CBMS-NSF Regional Conference
[59] Pasciak, J.E. (1980). Spectral and pseudo-spectral methods for advection equations. Math. Comp., Vol. 35, No. Series in Applied Mathematics, SIAM, Philadelphia.
152, pp. 1081{1092. [86] Temam, R. (1984). The Navier-Stokes Equations, Theory and Numerical Analysis. 3rd revised edition, North-
[60] Pironneau, O. (1982). On the transport-di usion algorithm and its applications to the Navier-Stokes equations. Holland.
Numer. Math., 38:309{332. [87] Temam, R. (1989). In nite-Dimensional Dynamical Systems in Mechanics and Physics. Applied Mathematical
[61] Priestley, A. (1986). Lagrange and characteristic Galerkin methods for evolutionary problems. D. Phil. Thesis, Sciences 68, Springer-Verlag.
University of Oxford. [88] Temperton, C. and Staniforth, A. (1987). An ecient two time-level semi-Lagrangian implicit integration scheme.
[62] Priestley, A. (1989). The spectral Lagrange-Galerkin method for the atmospheric transportation of pollutants. Q. J. Roy. Met. Soc., 113:1025{1039.
Numerical Analysis Report 2/89, University of Reading. [89] Tomarelli, F. (1984). Regularity theorems and optimal error estimates for linear parabolic Cauchy problems. Numer.
[63] Raviart, P.-A. (1985). An analysis of particle methods. Lecture Notes in Mathematics 1127, Springer. Math., 45:23{50.
[64] Rees, M.D. (1988). Moving point, particle and free-Lagrange methods. D.Phil Thesis, Oxford University. [90] Tourigny, Y. and Suli, E. (1991). The nite-element method with nodes moving along the characteristics for
[65] Rees, M.D. and Morton, K. W. (1991). Moving point, particle and free-Lagrange methods for convection-di usion convection-di usion equations. Numer. Math., 59:399{412.
equations. SIAM J. Sci. Statist. Comput., 12:547{572. [91] Williamson, D. and Rasch, P. (1989). Two-dimensional semi-Lagrangian transport with shape-preserving interpola-
[66] Ritchie, H. (1986). Eliminiating the interpolation associated with the semi-Lagrangian method. Mon. Wea. Rev., tion. Mon. Wea. Rev., 117:102{129.
114:135{146. [92] Yang, G., Belleudy, P. and Temperville, A. (1991). A higher-orderEulerian scheme for coupled advection-di usion
[67] Ritchie, H. (1987). Semi-Lagrangian advection on a Gaussian grid. Mon. Wea. Rev., 115:608{619. transport. Int. J. Num. Meth. Fl., 12:43{58.
[68] Ritchie, H. (1988). Application of the semi-Lagrangian method to a spectral model of the shallow water equations. [93] Yang, J-C. and Hsu, E-L. (1991). On the use of the reach-back characteristics method for calcuation of dispersion.
Mon. Wea. Rev., 116:1587{1598. Int. J. Num. Meth. Fl., 12:225{235.
[69] Robert, A. (1981). A stable numerical integration scheme for the primitive meteorological equations. Atmos. Ocean,
19:35{46.
[70] Russell, T. F. (1985). Timestepping along characteristics with incomplete iteration for a Galerkin approximation of
miscible displacement in porous media. SIAM J. Numer. Anal., 22:970{1013.
[71] Russell, T. F. (1989). Eulerian-Lagrangian localized adjoint methods for advection-dominated problems. In Numerical
Analysis 1989, D.F. Griths and G.A. Watson Eds., Pitman Research Notes in Mathematics, Vol.228. Longman/Wiley.
pp. 206{228.
[72] Smolarkiewicz, P.K. and Rasch, P.J. (1991). Monotone advection on a sphere: an Eulerian versus semi-Lagrangian
approach. J. Atmos. Sci., 48:793{810.
[73] Staniforth, A. and C^ote, J. (1990). Semi-Lagrangian integration schemes for atmospheric models|a review. Sub-
mitted to Monthly Weather Review.
[74] Stuart, A. (1989). Private communication.
[75] Suli, E. (1985). Lagrange-Galerkin mixed nite element approximation of the Navier-Stokes equations. In Numerical
Methods for Fluid Dynamics, Morton, K. W. and Baines, M. J., Eds. pp 439{448. Oxford University Press.

82 83