Sie sind auf Seite 1von 29

Numer Algor (2006) 43:355–383

DOI 10.1007/s11075-007-9066-6
ORIGINAL PAPER

A fully discrete H 1 -Galerkin method with quadrature


for nonlinear advection–diffusion–reaction equations

M. Ganesh · K. Mustapha

Received: 9 May 2006 / Accepted: 16 January 2007 /


Published online: 15 February 2007
© Springer Science + Business Media B.V. 2007

Abstract We propose and analyze a fully discrete H 1 -Galerkin method with quadra-
ture for nonlinear parabolic advection–diffusion–reaction equations that requires
only linear algebraic solvers. Our scheme applied to the special case heat equation
is a fully discrete quadrature version of the least-squares method. We prove second
order convergence in time and optimal H 1 convergence in space for the computer
implementable method. The results of numerical computations demonstrate optimal
order convergence of scheme in H k for k = 0, 1, 2.

Keywords advection–diffusion–reaction equations · parabolic equations ·


Galerkin method · quadrature

Mathematics Subject Classifications (2000) Primary 65M12 · 65M60

Support of the Australian Research Council is gratefully acknowledged.


M. Ganesh (B)
Department of Mathematical and Computer Sciences, Colorado School of Mines,
Golden, CO 80401, USA
e-mail: mganesh@mines.edu

K. Mustapha
Department of Mathematical Sciences, King Fahd University of Petroleum and Minerals,
Dharhan 31261, Saudi Arabia
e-mail: kassem@kfupm.edu.sa
356 Numer Algor (2006) 43:355–383

1 Introduction

We consider the nonlinear parabolic advection–diffusion–reaction equation


∂u
(x, t) = a(x, u) "u + f (x, t, u, ∇u), (x, t) ∈ # × (0, T], (1.1)
∂t
u(x, 0) = g(x), x ∈ #, (1.2)
u(x, t) = 0, (x, t) ∈ ∂# × (0, T], (1.3)
where # ⊂ R2 is a bounded domain. The functions f and g are defined respectively
on # × [0, T] × R3 and #, and for some positive constant amin , the nonlinear
diffusion coefficient a, defined on # × R, satisfies
a(x, α) ≥ amin , x ∈ #, α ∈ R. (1.4)
The algorithm and analysis in this paper are applicable for a large class of linear
and nonlinear functions (including polynomials and exponentials) in the unknown
variables. Throughout the paper, we assume the following mild Lipschitz continuity
conditions on a and f : there exist positive constants C and s such that for x ∈ #, t ∈
[0, T], and α1 , α2 , β, γ ∈ R,
! "s
|a(x, α1 ) − a(x, α2 )| ≤ C e|α1 | + e|α2 | ) |α1 − α2 | , (1.5)
! |α1 | "s
| f (x, t, α1 , β, γ ) − f (x, t, α2 , β, γ )| ≤ C e + e|α2 | + e|β| + e|γ | ) |α1 − α2 | , (1.6)
! "s
| f (x, t, β, α1 , γ ) − f (x, t, β, α2 , γ )| ≤ C e|α1 | + e|α2 | + e|β| + e|γ | ) |α1 − α2 | , (1.7)
! "s
| f (x, t, β, γ , α1 ) − f (x, t, β, γ , α2 )| ≤ C e|α1 | + e|α2 | + e|β| + e|γ | ) |α1 − α2 | . (1.8)
A brief, but stronger, version of the above assumptions is that we require each
entry in the Jacobian of the nonlinear terms in (1.1) has at most exponential
growth. The model problem (1.1), (1.2) and (1.3) includes the non-divergence
diffusion processes in biological applications, and the standard divergence form
∇ · [a(x, u(x, t))∇u(x, t)]. In the next section, we describe further regularity and
domain assumptions required for analysis, after introducing appropriate function
spaces.
The main aim of the paper is to propose and analyze a generalized fully discrete
quadrature version of the least-squares H 1 -Galerkin method for the nonlinear
problem. The least-squares method and analysis have been investigated by many
authors for linear and semi-linear problems, following the seminal work [4, 5], on
smooth domains. The smoothness (C ∞ ) assumption is essential for the convergence
analysis of the standard least-squares method [3, Theorem 2.1]. In this work we do
not require such smoothness assumptions on the domain and our algorithm and
analysis are applicable even for domains with corners (satisfying the assumptions
described in the next section).
The finite element test functions in the standard least-squares method are ob-
tained by applying the given strongly elliptic operator on the trial functions [3–5].
In particular, for a model second-order semi-linear elliptic problem Lw = ψ(w),
with linear elliptic operator L and homogeneous Dirichlet boundary condition the
standard least-squares method is to find wh ∈ Sh such that
(Lwh , L() = (ψ(wh ), L() , ( ∈ Sh , (1.9)
Numer Algor (2006) 43:355–383 357

where for v, z ∈ H 0 (!) = L2 (!),


!
(v, z) = v(x)z(x) dx, ∥v∥20 = (v, v), (1.10)
!

and Sh ⊂ H 2 (!) ∩ H01 (!) denotes the space of all splines that are polynomials of
degree at most r ≥ 3 on each element of a quasi-uniform triangulation Th of ! (with
largest mesh size h) such that for φ ∈ H k (!) ∩ H01 (!), 2 ≤ k ≤ r + 1

inf ∥φ − χ ∥ℓ ≤ Chk−ℓ ∥φ∥k , ℓ = 0, 1, 2. (1.11)


χ∈Sh

Such finite elements are usually constructed by starting with a reference element ρ "
and taking the boundary vertices of the triangulation to be on ∂!. (If the interior !h
of the union of the finite elements is not equal to !, then the functions in Sh vanish
on ! \ !h . Details of construction of C 1 splines on arbitrary topology are in [15, 16]).
Throughout the paper, for a nonnegative integer k, the standard norm in the
Sobolev space H k (!) is denoted by ∥ · ∥k , L2 (!) = H 0 (!); H01 (!) denotes the space
of all functions φ ∈ H 1 (!) with φ = 0 on ∂!; C denotes a generic positive constant
which may depend on r, but which is independent of h, the time-discretization
parameter τ and the exact solution of the partial differential equation.
We remark that the H 1 -Galerkin method for second-order parabolic problems
requires an H 2 trial space (spanned by C 1 basis functions). Hence this method is
considered to be less practical compared to the standard Galerkin method based
on an H 1 trial space, for linear problems described in divergence form that are
suitable for weak formulations, allowing reduction in the smoothness requirement.
However, for our nonlinear model problem (1.1) in non-divergence form with
diffusion coefficient depending linearly or nonlinearly on the unknown solution, the
choice of H 2 trial space is natural. Construction of C 1 splines (and hence H 2 trial
spaces) is no longer considered to be difficult even on arbitrary topology, in view of
the recent work [15, 16] and references therein. In comparison with C 0 finite element
Galerkin methods, C 1 smoothness of the H 1 -Galerkin approximate solutions leads
to significantly smaller linear systems.
Trial spaces spanned by C 1 splines are essential for collocation finite element
methods, see the survey paper [7] and the recent work [19] and references therein.
The finite element Galerkin method for biharmonic problems based on a weak
formulation requires an H 2 trial space, see the recent work [1, 2] and references
therein. A byproduct of the analysis in this work in Section 3 yields optimal
order convergence of a quadrature finite element Galerkin solution for a nonlinear
biharmonic problem, extending the analysis in [2]. Further, the scheme in this paper
has the advantage of allowing a wider class of nonlinear processes in the advection–
diffusion–reaction model (1.1), (1.2) and (1.3).
In (1.1), (1.2) and (1.3), the variable coefficients of the second order elliptic
operator depend on the unknown solution u, and hence the standard H 1 -Galerkin
method (with test space obtained by applying the given elliptic operator) is not
appropriate for the model problem. Further, the integral in (1.10) with a non-
polynomial integrand cannot in general be evaluated exactly and hence a computer
implementable version of (1.9) requires a quadrature approximation for the inner
product (·, ·).
Since the integrand in (1.9) for the stiffness matrix is a polynomial of degree at
most 2r − 2 on each triangular finite element, we consider a quadrature rule with
358 Numer Algor (2006) 43:355–383

degree of precision 2r − 2 over the reference element ρ ! with positive weights and
quadrature points in the interior of ρ
!. This rule induces a quadrature formula (·, ·)ρ,h
on each element ρ ∈ Th so that
"
v(x)z(x) dx = (v, z)ρ ≈ (v, z)ρ,h , ("1 , "2 )ρ = ("1 , "2 )ρ,h , "1 "2 ∈ P2r−2 ,
ρ

(1.12)

where P j is the space of all polynomials of degree at most j. Let


#
(v, z)h = (v, z)ρ,h , ∥v∥2h = (v, v)h . (1.13)
ρ∈Th

Replacing the continuous inner product (·, ·) in (1.9) by (·, ·)h , we get a fully
discrete computer implementable quadrature approximation version of the standard
least–square scheme in (1.9). However, a similar approach for the time–dependent
problem, using stable implicit rules for discretization of the time derivative operator,
will lead to the expensive requirement of solving a nonlinear algebraic system at each
time step.
Now we are ready to describe a new fully discrete H 1 -Galerkin method with
quadrature to solve the parabolic problem (1.1), (1.2) and (1.3), that requires solving
only linear system at each time step. In our new scheme, the finite element trial and
test spaces for (1.1), (1.2) and (1.3) are chosen to be the same as that used in the
least-squares method for the Poisson equation.
Nt
For a positive integer N t , let #t = {tn }n=0 be a uniform partition of the time in-
terval [0, T] such that tn = nτ, where τ = T/N t , and let tn+ 1 = tn + τ/2. Throughout
2
the paper, we use the following notation for a function φ.
φ n+1 − φ n 1 φ n+1 + φ n 3φ n − φ n−1
φ n = φ(tn ), ∂t φ n = , φ n+ 2 = , !n+ 12 =
φ .
τ 2 2
(1.14)

Our fully discrete quadrature scheme to solve (1.1), (1.2) and (1.3) is: find U :
#t → Sh , such that
!n+ 12 'U n+ 12 , '")h +(F (tn+ 1 )U
(∂t U n ,'")h =(AU !n+ 12 , '")h ,"∈ Sh, n=1,. . ., N t−1,
2

(1.15)

where A, F (t), t ∈ [0, T] are defined for ψ ∈ C 1 ()) by

[Aψ](x) = a(x, ψ(x)), [F (t)ψ](x) = f (x, t, ψ(x), ∇ψ(x)), x ∈ ). (1.16)

The linear system in (1.15) requires selection of U 0 , U 1 ∈ Sh . Given U 0 ∈ Sh ,


depending on the initial data u0 in (1.2), we select U 1 ∈ Sh by solving the following
predictor-corrector linear systems
1
(∂t V 0 , '")h = (AV 0 'V 2 , '")h + (F (t 1 )V 0 , '")h , " ∈ Sh , (1.17)
2

1 1 1
(∂t U 0 , '")h = (AV 2 'U 2 , '")h + (F (t 1 )V 2 , '")h , " ∈ Sh , (1.18)
2

where V 0 = U 0 and V 1 ∈ Sh .
Numer Algor (2006) 43:355–383 359

The purpose of rest of the paper is to establish second order in time and optimal
order H 1 norm error bounds in space for the computer implementable scheme (1.15),
(1.16), (1.17) and (1.18), and to demonstrate convergence of the algorithm in H k
norms for all k = 0, 1, 2 with numerical experiments.
The outline of this paper is as follows. In the next section, we discuss preliminary
results including the derivation of a non-standard approximation property in Sh . The
convergence analysis is based on analyzing the error in two stages, through an elliptic
projection of the exact solution, for each fixed time. In Section 3, for each fixed time,
we investigate the stability and convergence of a finite element elliptic projection
with quadrature. In Section 4, using the elliptic projections as comparison functions,
we prove the uniform boundedness, second order in time and optimal order H 1 norm
convergence of the approximate solutions satisfying (1.15), (1.16), (1.17) and (1.18).
Numerical experiments in Section 5 confirm the theoretical results, and demonstrate
optimal order convergence of the algorithm for a nonlinear parabolic problem in H k
norm for all k = 0, 1, 2.

2 Preliminaries

For a nonnegative integer k, in addition to the Sobolev space H k (!) (with norm
∥ · ∥k ), we use the standard norm ∥ · ∥k,∞ in the Banach space C k (!). (Note that
H 0 (!) = L2 (!), (., .) = ∥ · ∥20 = ∥ · ∥2L2 (!) , C 0 (!) = C (!), ∥ · ∥0,∞ = ∥ · ∥∞ .) We de-
fine the “broken” H m and C m norms ∥v∥m,Th and ∥v∥m,∞,Th on the quasi-uniform
triangulation Th by
!
∥v∥2m,Th = ∥v∥2m,ρ , ∥v∥2m,∞,Th = max ∥v∥2m,∞,ρ ,
ρ∈Th
ρ∈Th

where ∥ · ∥m,ρ = ∥ · ∥ Hm (ρ) and ∥ · ∥m,∞,ρ = ∥ · ∥C m (ρ) .


We assume throughout the paper that u is a solution of (1.1), (1.2) and (1.3) with

u ∈ C 5 (! × [0, T]), u, ut ∈ C (Hr+3 (!), [0, T]),

(where ut = ∂u
∂t
), and that the coefficient a in (1.1) is such that

a ∈ C 5 (! × [−δ, δ]), δ = max ∥u∥∞ . (2.1)


0≤t≤T

Throughout the paper, for φ, φt ∈ C (!), C(φ) denotes a generic positive constant
depending only on r, φ, φt , and Ci (φ) is a specific C(φ), for i = 1, 2, 3.
For each fixed t ∈ [0, T], optimal regularity requirement of the exact solution in
the standard H 1 -Galerkin method analysis (without quadrature), using splines of
degree r, is Hr+1 (!). It is common in quadrature finite element analysis to assume
extra regularity [1, 2, 6, 7, 10, 19]. We require Hr+3 (!) regularity, mainly due to
technical details involved in analysis of the quadrature error in Lemma 3.1. Such
extra regularity assumption on the exact solution is essential for analysis of C 1 spline
collocation methods that do not involve integrals, but require evaluation of functions
at certain quadrature points [7, 10, 19]. The analysis in [1, 2] for finite element
Galerkin method with quadrature for a linear biharmonic problem, restricted to
the C 1 cubic spline (r = 3) case, requires Hr+5 (!) regularity. As demonstrated in [6,
Section 5], the extra regularity assumptions on the solution and coefficients may be
360 Numer Algor (2006) 43:355–383

required only in quadrature finite element analysis, and may not affect convergence
rates in practical computations. It is useful to recall that implementation of the
scheme (1.15), (1.16), (1.17) and (1.18) does not require such regularity conditions.
Using the approximation property (1.11), we obtain the following result.

Lemma 2.1 If φ ∈ H k (") ∩ H01 ("), 2 ≤ k ≤ r + 1, then there exists # ∈ Sh such that

∥φ − #∥ℓ ≤ Chk−ℓ ∥φ∥k , ℓ = 0, 1, 2.

The result of the next lemma is well known in the approximation theory [9].

Lemma 2.2 If φ ∈ H k ("), k ≥ 2, then there exists a spline % of degree k − 1 defined


on Th such that

∥φ − %∥ j ≤ Chk− j∥φ∥k , j = 0, · · · , k − 1.

For each # ∈ Sh , &# is a polynomial of degree r − 2 on ρ ∈ Th , and hence the


degree precision 2r − 2 of the quadrature (see (1.12)), integration by parts, the
Poincaré inequality, and results in [9, Chapter 4], yield properties in the following two
lemmas. (The constant C = 1 in the first two inequalities in Lemma 2.3 for triangular
elements.)

Lemma 2.3 For any #, % ∈ Sh ,

∥&#∥h ≤ C∥&#∥0 ≤ C∥&#∥h , ∥#∥21 ≤ C(#, −&#)h , and (#, &%)h = (&#, %)h .

Lemma 2.4 If X is a piecewise polynomial of degree r on Th and g ∈ H 2 (ρ), ρ ∈ Th ,


then
! !
!(X , g)ρ − (X , g)ρ,h ! ≤ Ch2 ∥g∥2,ρ ∥X ∥0,ρ

Remark 2.5 Depending on the shape of the domain ", it may be convenient to use
a rectangular partition Th of ", to define the finite dimensional approximation space
Sh spanned by splines that are polynomials of degree at most r (with respect to each
variable) on each rectangular element ρ.
In case of rectangular elements, if % is a polynomial of degree r on ρ, &% is also a
polynomial of degree r on ρ. Hence, unlike in the triangular elements case, choosing
a quadrature formula (·, ·)ρ,h on each element ρ with degree of precision 2r − 2 and
(1.13) need not directly imply properties stated in Lemmas 2.3 and 2.4. We need to
be more precise about the choice of quadrature rule and proof of Lemmas 2.3 and 2.4
in this case. We discuss the rectangular elements and quadrature in Appendix (see
Section 6.1).

We use the following bounds throughout the paper. Using the H 2 regularity of the
unique solution of the homogeneous Dirichlet problem for the Poisson equation on
a convex or a C 2 domain " [14, 18], we get the equivalence relation

∥&φ∥0 ≤ ∥φ∥2 ≤ C∥&φ∥0 , φ ∈ H 2 (") ∩ H01 ("). (2.2)


Numer Algor (2006) 43:355–383 361

The Cauchy–Schwarz inequality applied to the quadrature sum yields


|(φ, ψ)h | ≤ C∥φ∥h ∥ψ∥h , for any φ, ψ defined on all quadrature points in #.
(2.3)
Next we derive a non-standard approximation and stability property in Sh using
the H 4 regularity of the classical biharmonic problem:
If ψ ∈ L2 (#), there exists χ ∈ H 4 (#) such that %2 χ = ψ, and ∥χ ∥4 ≤ C∥ψ∥0 .
(A1)
The assumption (A1) holds for # ⊂ R2 with a piecewise smooth boundary satisfying
a maximum interior angle condition at the corners [8, Theorem 2] (for example
rectangles), and also for C 4 domains [17, 18]. A direct application of the assumption
(A1) is required only in the proof of Lemma 2.6. An important relation connecting
the quadrature finite element analysis with a variable coefficient biharmonic problem
is given in the next section.
Thus, we assume the following sufficient domain condition for (2.2) and (A1) to
hold: The domain # is either C 4 or it is convex and Lipschitz continuous with interior
angles at the corner points do not exceed 126 degrees.

! ∈ L2 (#). Then there exist φ, φ u ∈ H 2 (#) ∩


Lemma 2.6 Let (A1) hold and let φ
1
H0 (#), and & ∈ Sh such that
!,
%φ = φ φ u = φ/Au, ∥φ u − %&∥0 ≤ C(u)h2 ∥φ
!∥0 , !∥0 .
∥%&∥2,Th ≤ C(u)∥φ
(2.4)

Proof The existence of solution φ ∈ H 2 (#) ∩ H01 (#) of the Poisson equation satisfy-
!∥0 is well known [18]. Hence φ u in (2.4) exists and ∥φ u ∥2 ≤ C(u)∥φ
ing ∥φ∥2 ≤ C∥φ !∥0 .
u 2 4
Since %φ ∈ L (#), using (A1) there exists ψ ∈ H (#) such that
%2 ψ = %φ u , ∥ψ∥4 ≤ C∥%φ u ∥0 ≤ C∥φ u ∥2 ≤ C(u)∥φ
!∥0 . (2.5)
Hence using r ≥ 3, Lemma 2.1, and (2.2), there exists & ∈ Sh , such that
∥ψ − &∥ℓ ≤ Ch4−ℓ ∥ψ∥4 ≤ C(u)h4−ℓ ∥φ
!∥0 ℓ = 0, 1, 2. (2.6)
Using (2.5) and (2.6),
∥φ u − %&∥0 = ∥%(ψ − &)∥0 ≤ C∥ψ − &∥2 ≤ C(u)h2 ∥φ
!∥0 . (2.7)
Lemma 2.1 and the last equality in (2.5) yield a spline ( ∈ Sh such that
∥φ u − (∥ℓ ≤ C(u)h2−ℓ ∥φ
!∥0 , ℓ = 0, 1, 2. (2.8)
The triangle and inverse inequalities, (2.8), and (2.7) yield
" #
∥%&∥2,Th ≤ ∥%& − (∥2,Th + ∥( − φ u ∥2 + ∥φ u ∥2 ≤ C(u) h−2 ∥%& − (∥0 + ∥φ u ∥2
" #
≤ C(u) h−2 ∥%& − φ u ∥0 + h−2 ∥φ u − (∥0 + ∥φ u ∥2 ≤ C(u)∥φ
!∥0 . )

Remark 2.7 Our main analysis in the rest of the paper is based only on the properties
stated in Lemmas 2.1 to 2.6. Accordingly, the algorithm and analysis in this paper can
be extended for (1.1), (1.2) and (1.3) defined on any domain # ⊂ Rd , d = 1, 2, 3, · · ·
362 Numer Algor (2006) 43:355–383

with corresponding approximation space and quadrature rule satisfying results in this
section. For notational convenience, throughout the paper we restrict to the case
d = 2.

3 Properties of fully discrete elliptic projections

Following a traditional approach, our convergence analysis is based on splitting the


error between the exact solution u and the computable U as u − U = (u − W) +
(W − U), where W is a comparison function obtained using a fully discrete elliptic
projection of the exact solution u in Sh , for each fixed time. In this section we
analyze the first stage error and stability of the comparison function, by choosing
W : [0, T] → Sh to be the solution of

(Au !W, !")h = (Au !u, !")h , " ∈ Sh , t ∈ [0, T]. (3.1)

For each fixed t ∈ [0, T], the comparison function W(t) ∈ Sh in (3.1) may also be
considered as a quadrature finite element Galerkin solution of the biharmonic-type
problem !(α(x)!w) = χ (x) for the unknown solution w subject to the boundary
conditions w = Bw = 0 on ∂&, where B = ! or ∂n , the outer normal derivative.
Hence convergence analysis in this section generalizes that in the recent work [2] (for
the constant coefficient biharmonic operator) to the variable coefficient biharmonic
problem with less regularity requirement on the solution.
We bound the comparison function error, denoted throughout the paper by η =
u − W, after studying the effect of quadrature approximations of various integrals
that will occur in our main analysis.

Lemma 3.1 Let D be the linear differential operator, defined for ψ ∈ H 2 (&) by

Dψ =(D2 +D1 +D0 )ψ, D2 ψ =α!ψ, D1 ψ =β.∇ψ, D0 ψ =γ ψ, α, γ ∈ R, β ∈ R2 .


(3.2)

Let * ∈ Sh and let φ !, for some φ ∈ Hr+3 (&) and *


!= φ − * ! ∈ Sh .

!∥0 ≤ C(φ)hr−1 , then for any v ∈ C 2 (&),


(a) If ∥Dφ
"# $ # $ "
" !*, v Dφ! − !*, v Dφ ! " ≤ C(v)C(φ)hr+1 ∥!*∥2,Th , (3.3)
h

!∥0 ≤ C(φ)hr−1 , and wi ∈ C 2 (&), i = 1, 2 then with w = (w1 , w2 ),


(b) If ∥∇xi φ
"# $ # $ "
" !*, w.∇ φ
! − !*, w.∇ φ ! " ≤ C(w)C(φ)hr+1 ∥!*∥2,Th . (3.4)
h

!∥0 ≤ C(φ)hk− j, for some 3 ≤ k ≤ r + 1, then


(c) For j = 0, 1, 2, if ∥D jφ

!∥h ≤ C(φ)hk− j,
∥D j φ j = 0, 1, 2. (3.5)

!∥0 ≤ C(φ)hr−1 , then for any v ∈ C 2 (&)


(d) If ∥D*
"# $ # $ "
" !*, v D*! − !*, v D* ! " ≤ C(v)C(φ)hr+1 ∥!*∥2,Th . (3.6)
h
Numer Algor (2006) 43:355–383 363

Proof Since Dφ ∈ Hr+1 ("), using Lemma 2.1, there exists a spline # ∈ Sh of degree
r such that

∥Dφ − #∥ℓ ≤ Chr+1−ℓ ∥φ∥r+3 , ℓ = 0, 1, 2. (3.7)

Using the triangle inequality,


! # $ !
!(%&, v Dφ " !
") − %&, v Dφ
h
! ! !# $
≤ !(%&, v[Dφ − #]) − (%&, v[Dφ − #])h ! + ! %&, v[# − D&
"]
# % &$ !
− %&, v # − D& " ! ≡ I1 + I2 . (3.8)
h

Using Lemma 2.4 (with g = v[Dφ − #], X = %&), Leibniz’s rule, the Cauchy–
Schwarz inequality, and (3.7), we have
' '
I1 ≤ Ch2 ∥v[Dφ − #]∥2,ρ ∥%&∥0,ρ ≤ C(v)h2 ∥Dφ − #∥2,ρ ∥%&∥0,ρ
ρ∈Th ρ∈Th

≤ C(v)h2 ∥Dφ − #∥2 ∥%&∥0 ≤ C(v)hr+1 ∥φ∥r+3 ∥%&∥0 . (3.9)

To bound I2 , we use Lemma 2.4 (with g = v%&, X = [# − D& "]), Leibniz’s rule, the
" in (a), and
Cauchy–Schwarz inequality and triangle inequalities, assumption on φ
(3.7) to obtain
'
I2 ≤ Ch2 " − #∥0,ρ ≤ C(v)h2 ∥%&∥2,Th ∥D&
∥v%&∥2,ρ ∥D& " − #∥0
ρ∈Th
# $
"∥0 + ∥Dφ − #∥0 ≤ C(v)C(φ)hr+1 ∥φ∥r+3 ∥%&∥2,Th .
≤ C(v)h2 ∥%&∥2,Th ∥Dφ
(3.10)

Hence, (3.3) follows from (3.8), (3.9) and (3.10). The bound (3.4) can be proved
similarly.
For j = 0, 1, 2, since D jφ ∈ Hr+3− j(") with r ≥ 3, using Lemma 2.2, there exists a
spline # j of degree r + 2 − j defined on Th such that

∥D jφ − # j∥ℓ ≤ Chr+3− j−ℓ ∥φ∥r+3 , j = 0, 1, 2 ℓ = 0, 1, 2, 3. (3.11)

The triangle inequality yields


( )
"∥2h ≤ C J1j + J2j + ∥D jφ − # j∥20 + ∥# j − D j&
∥D j φ "∥20 , j = 0, 1, 2, (3.12)

where
j j
J1 = ∥D jφ − # j∥2h − ∥D jφ − # j∥20 , "∥2h − ∥# j − D j&
J2 = ∥# j − D j& "∥20 . (3.13)

Using the triangle inequality, (3.11), and assumption on φ " in (b), we obtain for j =
0, 1, 2, and 3 ≤ k ≤ r + 1,
# $
∥D jφ − # j∥20 + ∥# j − D j& "∥20 ≤ C(φ)h2k−2 j∥φ∥r+3
"∥20 ≤ C ∥D jφ − # j∥20 + ∥D jφ 2
.
(3.14)
364 Numer Algor (2006) 43:355–383

Using the Bramble–Hilbert lemma (see [9]) for the quadrature formula, Leibniz’s
rule, Cauchy–Schwarz inequalities, and (3.11), for j = 0, 1, 2, and 3 ≤ k ≤ r + 1,
we get
# #
j
! ! " # ∂ α1 +α2 #
3 # 2 #
J1 ≤ Ch # ∂ xα1 ∂ yα2 [D jφ − % j] (x, y)# dxdy
α +α =3 ρ
ρ∈Th 1 2

! $% % % % % % % % &
≤ Ch3 %D jφ − % j% %D jφ − % j% + %D jφ − % j% %D jφ − % j%
0,ρ 3,ρ 1,ρ 2,ρ
ρ∈Th
'% % % % % % % % (
≤ Ch3 %D jφ − % j%0 %D jφ − % j%3 + %D jφ − % j%1 %D jφ − % j%2

≤ Ch2r+6−2 j∥φ∥r+3
2
≤ Ch2k−2 j∥φ∥r+3
2
. (3.15)

Similarly, we obtain

j
! $% % % % % % % % &
J2 ≤ Ch3 ) % %% j − D j &
%% j − D j&
0,ρ 3,ρ
)% %% j − D j&
)% + %% j − D j&
1,ρ
)%
2,ρ
.
ρ∈Th

Hence, the inverse inequality and (3.14), for j = 0, 1, 2, and 3 ≤ k ≤ r + 1, yield


j
!
J2 ≤ C )∥20,ρ ≤ Ch2k−2 j∥φ∥r+3
∥% j − D j& 2
. (3.16)
ρ∈Th

The bound (3.5) now follows from using (3.14), (3.15) and (3.16) in (3.12).
Finally, using Lemma 2.4 (with g = v'&, X = '& )) and the assumption on &
),
we get

|('&, v D& )∥0 ≤ C(v)C(φ)hr+1 ∥'&∥2,Th .


))h | ≤ Ch2 ∥v'&∥2,Th ∥D&
)) − ('&, v D& &

For each t ∈ [0, T], since u, ut ∈ Hr+3 ((), using Lemma 2.1, there exist W,
) W* ∈ Sh
such that

) ℓ ≤ Chr+1−ℓ ∥u∥r+1 ,
∥u − W∥ * ℓ ≤ Chr+1−ℓ ∥ut ∥r+1 ,
∥ut − W∥ ℓ = 0, 1, 2.
(3.17)

Hence from Lemma 3.1(c)

) h ≤ C(u)hr−1 ,
∥'(u − W)∥ * h ≤ C(u)hr−1 .
∥'(ut − W)∥ (3.18)

Next we bound the comparison function error η in H j norm, and D jη in discrete


norm ∥ · ∥h , for j = 0, 1, 2, where D j is as defined in (3.2) with α = γ = β1 = β2 = 1.
Using (3.1), the comparison function error η = u − W is the solution of

(Au 'η, '%)h = 0, % ∈ Sh , t ∈ [0, T].

Theorem 3.2 For each t ∈ [0, T], (3.1) has a unique solution and
% %
∥η∥ j ≤ C(u)hr+1− j, %D jη% ≤ C(u)hr+1− j,
h
j = 0, 1, 2. (3.19)
Numer Algor (2006) 43:355–383 365

Proof Since the linear system corresponding to (3.1) is square, it is enough to show
that the solution of (3.1) is unique. If there exist two solutions W and Z ∈ Sh of (3.1),
then
(Au !(W − Z ), !")h = 0, " ∈ Sh . (3.20)
Using (2.2), Lemma 2.3, (1.4), and " = W − Z in (3.20), we get

∥W − Z ∥22 ≤ C∥!(W − Z )∥2h ≤ C∥ Au !(W − Z )∥2h
= C(Au !(W − Z ), !(W − Z ))h = 0,
and hence W = Z .
Since η = u − W, with u ∈ Hr+3 ($) and W ∈ Sh , using Lemma 3.1(c), it is suffi-
cient to prove the first inequality in (3.19). The triangle inequality and (3.17) yield

∥η∥2 ≤ ∥u − W∥ ! − W∥2 ≤ Chr−1 ∥u∥r+1 + ∥W


! 2 + ∥W ! − W∥2 . (3.21)

Using (2.2), Lemma 2.3, (1.4), (3.1), (2.3), (3.18), and Lemma 2.3,
! − W∥22 ≤ C∥!(W
∥W ! − W)∥2h

! − W)∥2h
≤ C∥ Au !(W
" #
! − W), !(W
= C Au !(W ! − W)
h
" #
! − u), !(W
= C Au !(W ! − W)
h

! − u)∥h ∥!(W
≤ C∥Au∥∞ ∥!(W ! − W)∥h

≤ Chr−1 ∥Au∥∞ ∥u∥r+3 ∥W


! −W∥2 , t ∈ [0, T]. (3.22)
Hence from (3.21) and (3.22), we obtain ∥η∥2 ≤ C(u)hr−1 .
Next we bound η in L2 and H 1 norms. For each t ∈ [0, T], using Lemma 2.6, with
φ = η, and φ, φ u as in (2.4), there exists & ∈ Sh and such that
!

∥φ u − !&∥0 ≤ C(u)h2 ∥η∥0 , ∥!&∥2,Th ≤ C(u)∥η∥0 . (3.23)


Using φ, η = 0 on ∂$, integration by parts and (3.1), we obtain
∥η∥20 = (!φ, η) = (!(Auφ u ), η) = (Auφ u , !η)
= (φ u − !&, Au !η) + {(!&, Au !η) − (!&, Au !η)h } ≡ J1 + J2 . (3.24)
The Cauchy–Schwarz inequality, (3.23), and (3.19) for ℓ = 2 yield
|J1 | ≤ C∥φ u − !&∥0 ∥Au !η∥0 ≤ C(u)h2 ∥η∥0 ∥η∥2 ≤ C(u)hr+1 ∥η∥0 . (3.25)
Since the first inequality in (3.19) holds for j = 2, using Lemma 3.1(a) (with α = 1
being the only non-zero coefficient in (3.2)), (3.3) and (3.23), we obtain
|J2 | ≤ C(u)hr+1 ∥η∥0 . (3.26)
Using (3.25) and (3.26) in (3.24) we get the first inequality in (3.19) for j = 0.
Further, the Poincaré inequality, integration by parts, the Cauchy–Schwarz
inequality, and the first inequality in (3.19) for j = 0, 2 yield
∥η∥21 ≤ −C(!η, η) ≤ C ∥η∥2 ∥η∥0 ≤ C(u)h2r ,
366 Numer Algor (2006) 43:355–383

and hence, the first desired result in (3.19) is obtained for j = 0, 1, 2. Using (3.2),
the triangle inequality and the first bound in (3.19), we obtain ∥D jη∥0 ≤ C(u)hr+1− j.
!, we obtain the second bound in (3.19).
Using this and (3.5) with η in place of φ %

Next we show that ηt has optimal error bounds in H 1 and H 2 norms, and hence
we obtain a sub-optimal error bound of ηt in discrete norm ∥·∥h .

Theorem 3.3 For each t ∈ [0, T],

∥ηt ∥2 ≤ C(u)hr−1 , ∥ηt ∥1 ≤ C(u)hr , ∥ηt ∥h ≤ C(u)hr . (3.27)

Proof Differentiating both sides of (3.1) with respect to t, we get that W, Wt ∈ Sh


satisfy
(At u #W + Au #Wt , #$)h = (At u #u + Au #ut , #$)h , v ∈ Sh , t ∈ [0, T],
(3.28)
where [At u(t)] = ∂t∂ [Au(t)].
Following (3.1), let W ∗ : [0, T] → Sh be such that
" #
Au #W ∗ , #$ h = (Au #ut , #$)h , $ ∈ Sh , t ∈ [0, T]. (3.29)
! and (3.1)
Using the triangle inequality, (3.17), arguments similar to (3.22) (with W, W
∗ $
replaced by W , W and (3.29) respectively), we obtain
∥ut − W ∗ ∥2 ≤ ∥ut − W∥ $ − W ∗ ∥2 ≤ C(u)hr−1 .
$ 2 + ∥W (3.30)
The triangle inequality and (3.30) yield
∥ηt ∥2 ≤ ∥ut − W ∗ ∥2 + ∥W ∗ − Wt ∥2 ≤ C(u)hr−1 + ∥W ∗ − Wt ∥2 . (3.31)
Using (2.2), Lemma 2.3, (1.4), (3.28), (3.29), (2.3), Theorem 3.2 and Lemma 2.3,
we get

∥W ∗ − Wt ∥22 ≤ C∥#(W ∗ − Wt )∥2h ≤ C∥ Au #(W ∗ − Wt )∥2h
" # " #
= C Au #(Wt − W ∗ ), #(Wt − W ∗ ) h = C At u #η, #(Wt − W ∗ ) h

≤ C∥At u∥∞ ∥#η∥h ∥#(W ∗ − Wt )∥h ≤ C(u)hr−1 ∥W ∗ − Wt ∥2 . (3.32)


Using (3.32) in (3.30), we obtain ∥ηt ∥2 ≤ C(u)hr−1 . Next, to bound ηt in H 1 norm, we
define a function p ∈ H 2 (&) and obtain its bound using the triangle inequality, (3.27)
and the first inequality in (3.19) with j = 2:
At u
p = ηt + η, ∥ p∥2 ≤ C(∥ηt ∥2 + C(u)∥η∥2 ) ≤ C(u)hr−1 . (3.33)
Au
Since
% & % &
At u At u At u
#p = #ηt + q + #η, q=# η + 2∇ .∇η, (3.34)
Au Au Au
using (3.28), it is easy to check that
(Au #p, #$)h = (Au q, #$)h , $ ∈ Sh , t ∈ [0, T]. (3.35)
Numer Algor (2006) 43:355–383 367

We approximate p by P, where P ∈ Sh is the solution of


(Au !P, !")h = (Au q, !")h , " ∈ Sh , t ∈ [0, T]. (3.36)
Using (2.2), Lemma 2.3, (3.36), the Cauchy–Schwarz inequality, (3.34), (1.4), and
Theorem 3.2, we obtain
∥P∥22 ≤ C∥!P∥20 ≤ C∥!P∥2h ≤ C(Au!P, !P)h ≤ C(u)hr ∥!P∥0 , (3.37)
Further, using (2.2), the triangle inequality, (3.33), and (3.37), we obtain
∥ p − P∥2 ≤ C∥!( p − P)∥0 ≤ C (∥!p∥0 + ∥!P∥0 ) ≤ C(u)hr−1 (3.38)
! = p − P, and φ, φ u as in (2.4), there exists $ ∈ Sh such
Using Lemma 2.6, with φ
that
∥φ u − !$∥0 ≤ C(u)h2 ∥ p − P∥0 , ∥!$∥2,Th ≤ C(u)∥ p − P∥0 , (3.39)
Using (3.35) and (3.36),
(Au!( p − P), !$)h = 0. (3.40)
Hence !(Au φ u ) = p − P and (3.40) yield
∥ p − P∥20 = (φ u −!$, Au !( p − P))
J1 + !
+ {(!$, Au !( p− P)) − (!$, Au !( p − P))h } ≡ ! J2 . (3.41)
Using arguments similar to the ones used to bound J1 in (3.25), with η replaced by
p − P, we have
J1 ≤ C(u)hr+1 ∥ p − P∥0 .
! (3.42)
"5
J2 , we use (3.34), the triangle inequality, and write !
To bound ! J2 ≤ i=1 J2i , where,
!
# #
J 21 = #(!$, Au !ηt ) − (!$, Au !ηt )h #
!
#$ $ % % $ $ % % #
2
#
# At u At u #
!
J 2 = # !$, Au ! η − !$, Au ! η ##
Au Au h
#$ $ % % $ $ % % #
3
#
# At u At u #
!
J 2 = 2 # !$, Au ∇ .∇η − !$, Au ∇ .∇η ##
Au Au h
# #
J24 = #(!$, At u !η) − (!$, At u !η)h #
!
# #
J 52 = # (!$, Au !P) − (!$, Au !P)h #.
!

Since the first inequalities in (3.19) and (3.27) hold for j = 2, ∥!P∥0 ≤ Chr , and using
Lemma 3.1 (with coefficients in (3.2) chosen appropriately to be zero or one), (3.3),
(3.4) and (3.39), we obtain
5
&
!
J2 ≤ J 2i ≤ C(u)hr+1 ∥ p − P∥0 .
! (3.43)
i=1

Therefore, using (3.42) and (3.43) in (3.41), we get


∥ p − P∥0 ≤ C(u)hr+1 . (3.44)
368 Numer Algor (2006) 43:355–383

The Poincaré inequality, integration by parts, the Cauchy–Schwarz inequality, (3.38),


and (3.44) yield

∥ p − P∥21 ≤ −C(!( p − P), p − P) ≤ C ∥!( p − P)∥0 ∥( p − P)∥0 ≤ C(u)h2r . (3.45)

Now we are ready to bound ∥ηt ∥1 . Using (3.33), the triangle inequality, (3.19), (3.45),
and (3.37), we obtain
At u
∥ηt ∥1 = ∥ p − η∥1 ≤ C(u)[∥ p∥1 + ∥η∥1 ] ≤ C(u)[∥ p − P∥1 + ∥P∥1 + hr ] ≤ C(u)hr .
Au

Finally, since ut ∈ Hr+3 (#), Wt ∈ Sh , the last inequality in (3.27) follows from the
bound ∥ηt ∥0 ≤ ∥ηt ∥1 ≤ C(u)hr , and Lemma 3.1 (b) (with j= 0, γ = 1, k =r, φ != ηt ).
&

We complete this section with uniform boundedness of the comparison functions


W, Wt in various norms.

Lemma 3.4 For each t ∈ [0, T],

∥W∥ ≤ C(u), ∥Wt ∥ ≤ C(u), ∥ · ∥ = ∥ · ∥1,∞ , or ∥ · ∥ = ∥ · ∥2 , or ∥ · ∥ = ∥ · ∥2,∞,Th .

Proof Since u ∈ Hr+3 (#), using Lemma 2.2, there exists a spline & of degree r + 2
such
∥u − &∥k ≤ Chr+3−k ∥u∥r+3 k = 0, · · · , r + 2.

Hence using the triangle and inverse inequalities, the Sobolev embedding theorem,
and (3.19), we obtain

∥W∥ ≤ ∥W − &∥ + ∥u − &∥ + ∥u∥


≤ Ch−1 ∥W − &∥2 + C∥u − &∥r+2 + C∥u∥r+2
≤ Ch−1 [∥η∥2 + ∥u − &∥2 ] + C(u) ≤ C(u),

and hence, the first desired inequality is obtained. Similarly (using (3.27) instead of
(3.19)), we obtain the second inequality. &

4 Convergence analysis

In this section, we prove optimal order convergence of our scheme by analyzing the
Nt
error un − U n , for n = 0, . . . , N t in H 1 norm, where U 1 and {U n }n=2 are defined by
(1.18) and (1.15) respectively, and U 0 = W 0 (with W as in (3.1)).
Since un − U n = ηn − ξ n , with ηn = un − W n , ξ n = U n − W n , n = 0, . . . , N t , first
we derive bounds on ξ between any two consecutive time levels in Theorem 4.1. Then
we bounded the term ξ 1 in Lemma 4.1 using the fact that ξ 0 = 0 and assuming that
the time discretization parameter τ and the spatial mesh size h satisfy the relation
τ ≤ Ch2/3 .
We prove the following two results in Appendix (see Section 6.2).
Numer Algor (2006) 43:355–383 369

Theorem 4.1 If h and τ are sufficiently small, then for n = 1, · · · , N t − 1, we have


! n+1
#
"
−(ξ n+1 , #ξ n+1 )h + (ξ n , #ξ n )h ≤ C(u)τ [d(n, u, U)]2 τ 4 + h2r + (ξ i , −#ξ i )h ,
i=n−1

(4.1)
where
$ % & (s+1
'n+ 12 ∥1,∞ )
d(n, u, U) = exp max ∥u∥1,∞ + exp(∥U . (4.2)
0≤t≤T

Lemma 4.2 Assume that h and τ are sufficiently small, τ ≤ Ch2/3 , V 0 = U 0 = W 0 and
V 1 , U 1 ∈ Sh are defined by (1.17) and (1.18). Then
) *
−(ξ 1 , #ξ 1 )h ≤ C(u) τ 4 + h2r . (4.3)

Now we are ready to prove uniform boundedness and optimal order convergence
of the approximate solutions to the exact solution u of (1.1), (1.2) and (1.3) in H 1
norm. We need to restrict our analysis to H 1 , mainly due to obtaining only h2r error
bound in (4.1). For convergence results in L2 norm, we need h2r+2 instead of h2r in
(4.1).
In addition to obtaining optimal order error bounds in H 1 norm for the algorithm
(1.15), (1.16), (1.17) and (1.18), an important contribution in this paper is to prove
uniform boundedness of the approximate solutions in supremum norm (see (4.8))
with mild nonlinear conditions (1.5), (1.6), (1.7) and (1.8) for the model problem
(1.1), (1.2) and (1.3).

Theorem 4.3 Assume that h and τ are sufficiently small, τ ≤ Ch2/3 , V 0 = U 0 = W 0


and V 1 , U 1 ∈ Sh are defined by (1.17) and (1.18). Then
) *
max t ∥un − U n ∥1 ≤ C(u) τ 2 + hr . (4.4)
0≤n≤N

Proof Using Lemmas 3.4 and 4.2, there exists a constant C1 (u) independent of h and
τ such that

max ∥W∥1,∞ ≤ C1 (u)/4, (ξ 1 , −#ξ 1 )h ≤ C1 (u)(τ 4 + h2r ). (4.5)


0≤t≤T
) *
The relation U'1+ 12 = 3(W 1 + ξ 1 ) − W 0 /2, the triangle and inverse inequalities,
Lemma 2.3, (4.5), τ 3 ≤ Ch2 , and h sufficiently small yield a constant C2 (u) indepen-
dent of h and τ such that
+ ,
∥U'1+ 12 ∥1,∞ ≤ C ∥W 0 ∥1,∞ + ∥W 1 ∥1,∞ + h−1 ∥ξ 1 ∥1 ≤ C2 (u). (4.6)

Let
C3 (u) = max {C1 (u), C2 (u)} . (4.7)

Next, we claim using mathematical induction that


1
'n+ 2 ∥1,∞ ≤ C3 (u),
∥U n = 1, . . . , N t . (4.8)
370 Numer Algor (2006) 43:355–383

Using (4.6) and (4.7), clearly (4.8) holds for n = 1. Assuming that ∥U !n+ 12 ∥1,∞ ≤
C3 (u), n = 1, . . . , ℓ, for some 1 ≤ ℓ ≤ N t − 1, and using (4.2), we have d(n, u, U) ≤
C(u), n = 1 . . . , ℓ. Hence Theorem 4.1 gives

" n+1
$
#
n+1 n+1 n n 4 2r i i
−(ξ , #ξ )h +(ξ , #ξ )h ≤C(u)τ τ +h + (ξ , −#ξ )h , n=1, . . . , ℓ.
i=n−1

(4.9)

Summing both sides of (4.9) for n = 1, . . . , k − 1, where 2 ≤ k ≤ ℓ + 1, we obtain

" k
$
#
−(ξ k , #ξ k )h ≤ C(u) −(ξ 1 , #ξ 1 )h + τ 4 + h2r + τ (ξ n , −#ξ n )h . (4.10)
n=0

Clearly (4.10) holds for k = 1, and since ξ 0 = 0, it also holds for k = 0. Hence, for τ
sufficiently small, the discrete analogue of Gronwall’s inequality (see Lemma 4.7 in
[12]) yields
% &
−(ξ n , #ξ n )h ≤ C(u) −(ξ 1 , #ξ 1 )h + τ 4 + h2r , n = 0, . . . , ℓ + 1. (4.11)

8
Using the inverse inequality, Lemma 2.3, (4.11), (4.5), τ 4 ≤ Ch 3 , and taking h
sufficiently small, we get

' ( [C3 (u)]2


∥ξ n ∥21,∞ ≤ Ch−2 ∥ξ n ∥21 ≤ Ch−2 (ξ n , −#ξ n )h ≤ C(u)h−2 τ 4 + h2r ≤ ,
16
n = ℓ, ℓ + 1. (4.12)

!ℓ+1+ 12 = W
Using the relation U ! ℓ+1+ 12 + ! 1
ξ ℓ+1+ 2 , the triangle inequality, (4.5), and
(4.12), we obtain

1
!ℓ+1+ 2 ∥1,∞ ≤ 3 1 3 1
∥U ∥W ℓ+1 ∥1,∞ + ∥W ℓ ∥1,∞ + ∥ξ ℓ+1 ∥1,∞ + ∥ξ ℓ ∥1,∞ ≤ C3 (u),
2 2 2 2

which completes the proof of (4.8) by induction.


To show (4.4), we use (4.8) and follow the derivation of (4.11) to obtain
% &
(ξ n , −#ξ n )h ≤ C(u) (ξ 1 , −#ξ 1 )h + τ 4 + h2r , n = 0, . . . , N t .

Hence Lemmas 2.3 and 4.2 imply that


' (
∥ξ n ∥21 ≤ C(u) τ 4 + h2r , n = 0, . . . , N t . (4.13)

Since un − U n = ηn − ξ n , (4.4) follows from the triangle inequality, Theorem 3.2,


and (4.13). &

Numer Algor (2006) 43:355–383 371

5 Numerical experiments

We consider the nonlinear parabolic test problem


∂u ! "
(x, t) = ∇ · u2 (x, t)∇u(x, t) + t3 exp(u(x, t)) + #
f (x, t), (x, t) ∈ " × (0, T], (5.1)
∂t
u(x, 0) = g(x), x ∈ ", (5.2)
u(x, t) = 0, (x, t) ∈ ∂" × (0, T], (5.3)

where " is the unit square, the forcing term # f (x, t) and initial condition g(x) are
chosen so that the exact solution of (5.1), (5.2) and (5.3) is

u(x, t) = [sin(t) + cos(t)] sin(π x) sin(π y), x = (x, y) ∈ ", t ∈ [0, T].

The model problem (5.1), (5.2) and (5.3) with (Frank–Kamenetskii type) expo-
nential reaction, and polynomial advection and diffusion terms can be written in the
form (1.1), (1.2) and (1.3), with a(x, u(x, t)) = u2 (x, t) (not satisfying the positivity
requirement (1.4) in analysis),
$ %2 $ %2
3 # ∂u ∂u
f (x, t, u, ∇u) = t exp(u(x, t)) + f (x, t) + 2u(x, t) (x, t) + 2u(x, t) (x, t) .
∂x ∂y
We computed the numerical solutions U n ∈ Sh , n = 1, · · · , N t , of (5.1), (5.2) and
(5.3) by solving the linear systems (1.15), (1.17) and (1.18) in two finite element
spaces Sh , spanned by the cubic (r = 3) and quartic (r = 4) spline bases. The basis
functions were constructed on various triangulations of " using rectangular elements
with a uniform mesh size parameter h. The uniform time partition parameter
τ = T/N t was chosen so that τ 2 ≤ hr+1 to measure the rate of convergence of the
computed solutions.
The initial approximate solution U 0 ∈ Sh , required for (1.17) and (1.18), was
computed by solving the linear system (3.1) with t = 0, using the initial condition
(5.2). We used the quadrature described in Section 6.1. Our analysis and the number
of quadrature points prescribed in Section 6.1 is optimal in the sense that reducing
the precision of quadrature by even one degree did not yield optimal experimental
convergence rates observed in the following tables. We calculated the errors

ϵk,∞ = max t ∥un − U n ∥k , k = 0, 1, 2,


0≤n≤N

using 25600 Gauss quadrature points in ", and then calculated the rate of conver-
gence (R(H k )) in H k norm, for k = 0, 1, 2. Results in Tables 1 and 2 confirm the

Table 1 Errors and optimal order convergence rates for the case r = 3

h ϵ0,∞ R(H 0 ) ϵ1,∞ R(H 1 ) ϵ2,∞ R(H 2 )

0.25000 5.5713e-03 6.0496e-02 1.0520e+00


0.12500 2.3454e-04 4.5701 4.1345e-03 3.8710 1.4383e-01 2.8707
0.06250 1.0904e-05 4.4269 3.0451e-04 3.7631 2.0137e-02 2.8364
0.03125 6.7513e-07 4.0135 3.1648e-05 3.2663 4.3848e-03 2.1993
372 Numer Algor (2006) 43:355–383

Table 2 Errors and optimal order convergence rates for the case r = 4

h ϵ0,∞ R(H 0 ) ϵ1,∞ R(H 1 ) ϵ2,∞ R(H 2 )

0.25000 1.1427e-03 1.7491e-02 4.4372e-01


0.12500 1.4639e-05 6.2865 2.0892e-04 6.3876 1.0837e-02 5.3555
0.06250 3.0518e-07 5.5840 8.9139e-06 4.5507 7.2817e-04 3.8956
0.03125 9.0067e-09 5.0825 4.0232e-07 4.4696 6.6759e-05 3.4472

optimal convergence rates (r + 1 − k) of our scheme in H k norm, proved in Theorem


4.3 for k = 1, and expected for k = 0, 2.

6 Appendix

In Section 6.1, we describe a special quadrature treatment for rectangular elements


mentioned in Remark 2.5, and in Section 6.2 we prove Theorem 4.1 and Lemma 4.2.

6.1 Rectangular elements and quadrature

As described in Remark 2.5, in case of rectangular elements, we need to be precise


about the choice of quadrature rule to obtain results in Section 2. We give details
below by assuming in this subsection that " is the unit square.
For a given small parameter h, it is convenient to use a quasi-uniform col-
lection Th of rectangular elements ρ to partition ", with h = max(hx , h y ), where
ρ = (xk−1 , xk ) × (yl , yl−1 ), hx,k = xk − xk−1 k = 1, · · · , N x , l = 1, · · · , N y , hx =
max x hx,k and h y defined similarly.
1≤k≤N
In this case, Sh ⊂ H 2 (") ∩ H01 (") is a tensor-product space consisting of splines
that are polynomials of degree at most r (with respect to each variable x, y) on each
rectangular element ρ. The approximation properties stated in Section 2 (Lemmas
2.1 and 2.2) hold for this finite dimensional space. (See for example Theorem 4 in
[11] and Theorem 3.3 in [12]). Next we consider a concrete quadrature rule.
We start with a J = r − 1-point Gauss quadrature rule on (0, 1) (for r ≥ 4) with
weights and nodes {w j} Jj=1 and {ξ j} Jj=1 . This rule induces a quadrature rule with degree
of precision 2J − 1, on ρ = (xk−1 , xk ) × (yl , yl−1 ), k = 1, · · · , N x , l = 1, · · · , N y ,
defined by
! J
"
v(x)z(x) dx = (v, z)ρ ≈ (v, z)ρ,h = hx,k h y,l wm wn (vz)(xk,m , yl,m ), (6.4)
ρ m,n=1

with xk,m = xk−1 + hx,k ξm and yl,n = yl−1 + h y,l ξn . Lemma 2.4 for this rectangular
element discrete inner product follows from application of [6, Lemma 2.6], provided
that J ≥ (r + 2)/2. Hence we need J = r for the cubic spline r = 3 case. (The proof
of [6, Lemma 2.6] is based on arguments used in [9, Chapter 4]).
Below we give a proof of Lemma 2.3 to complete all results in Section 2 for
rectangular elements. Let %, & ∈ Sh . First we prove that

(%xx , % yy )h ≥ 0 . (6.5)
Numer Algor (2006) 43:355–383 373

For r=3, since !∈ Sh , the chosen quadrature rule has degree of precision 2r − 1,
! yy (α, y) = !xx (x, α) = 0 for α = 0, 1, x ∈ [0, 1], using integration by parts we
obtain
(!xx , ! yy )h = (!xx , ! yy ) = −(!x , !xyy ) = (!xy , !xy ) = ∥!xy ∥20 ≥ 0. (6.6)

For r ≥ 4, using Lemma 3.1 in [10] and !xx (x, α) = 0 for α = 0, 1, x ∈ [0, 1], with
yl− 1 = yl−12+yl , we obtain
2

x y
N
! r−1
! N
! r−1
!
x,k y,l
(!xx , ! yy )h = h wm h wn (!xx ! yy )(xk,m , yl,n )
k=1 m=1 l=1 n=1

N
!
x
r−1
! "# 1
=− hx,k wm (!xxy ! y )(xk,m , y) dy
k=1 m=1 0

N y $
! ∂ 2r
y,l 2r−1
+C (h ) (!xx !)(xk,m , yl− 1 )
∂ y2r 2
l=1
# x
1 N! ,r−1
=− hx,k wm (!xxy ! y )(xk,m , y) dy
0 k,m=1
y x
N
! N! ,r−1
y,l 2r−1 ∂ 2r
−C (h ) hx,k wm (!xx !)(xk,m , yl− 1 ). (6.7)
∂ y2r 2
l=1 k,m=1

Similarly and using ! y (α, y) = 0 for α = 0, 1, y ∈ [0, 1], we get


x
N
! r−1
!
− hx,k wm (!xxy ! y )(xk,m , y)
k=1 m=1
# 1
= (!xy !xy )(x, y) dx
0
N
!
x " %2
x,k 2r−1 ∂r
+C (h ) !y (xk− 1 , y) ≥ C∥!xy ∥2 , (6.8)
∂ xr 2
k=1

and for l = 1, · · · , N y ,
x
N
! r−1
! ∂ 2r
− hx,k wm (!xx !)(xk,m , yl− 1 )
m=1
∂ y2r 2
k=1
# 1" %2
∂r
= !x (x, yl− 1 ) dx
0 ∂ yr 2

N
!
x " %2
x,k 2r−1 ∂r k
+C (h ) ! (x, yl− 1 ) ≥ 0. (6.9)
∂ xr 2
k=1

Therefore, using (6.7), (6.8) and (6.9) we obtain (6.5).


374 Numer Algor (2006) 43:355–383

Using (1.6) in [20] and (17) in [13], we obtain

∥!xx ∥h ≤ C∥!xx ∥0 ≤ C∥!xx ∥h , ∥! yy ∥h ≤ C∥! yy ∥0 ≤ C∥!xx ∥h . (6.10)

Using (2.3), the Cauchy’s inequality, (6.10), and [18, p.180, (8.24)], we obtain

∥"!∥2h = ("!, "!)h = ∥!xx ∥2h + ∥! yy ∥2h + 2(!xx , ! yy )h


! "
≤ C ∥!xx ∥2h + ∥!xx ∥2h ≤ C∥"!∥20 . (6.11)

The triangle inequality, (6.10), and the inequality (!xx , ! yy )h ≥ 0 yield


! " ! "
∥"!∥20 ≤ C ∥!xx ∥20 + ∥!xx ∥20 ≤ C ∥!xx ∥2h + ∥!xx ∥2h ≤ C∥"!∥2h , (6.12)

and hence, the first result in Lemma 2.3 follows from (6.11) and (6.12).
Using (2.6) in [20] and Lemma 3.3 in [10], we get

# 1 $# 1 % N
&
y
J
& # 1
y,l
∥!x ∥20 = !2x (x, y)dy dx ≤ C h wn !2x (x, yl,n )dx
0 0 l=1 n=1 0

N y
J
' Nx J
(
& & & &
≤ −C h y,l wn hx,k wm (!!xx )(xk,m , yl,n )
l=1 n=1 k=1 m=1

= −C(!, !xx )h . (6.13)

Similarly, we obtain

∥! y ∥20 ≤ − C(!, ! yy )h (6.14)

Since ! = 0 on ∂$, the second desired result follows from the Poincaré inequality,
(6.13) and (6.14). The last inequality in Lemma 2.3, for r = 3, is obtained by using the
fact that the degree of precision is 2r − 2 and using integration by parts . For r ≥ 4,
we use (1.13) and Lemma 3.1 in [10], to obtain

x y
N
& r−1
& N
& r−1
&
x,k
(!, "%)h = h wm h y,l wn (![%xx + % yy ])(xk,m , yl,n )
k=1 m=1 l=1 n=1
x y
N
& r−1
& N
& r−1
&
= hx,k wm h y,l wn (!xx % + ! yy %)(xk,m , yl,n ) = ("!, %)h .
k=1 m=1 l=1 n=1

6.2 Proofs of Theorem 4.1 and Lemma 4.2

The first part of a proof of Theorem 4.1 is based on the idea of expanding and splitting
first two terms of the method in (1.15) into eight terms, by choosing appropriate test
functions and U in (1.15) replaced by the error ξ . The major part of the proof is then
to bound each of the eight terms.
Numer Algor (2006) 43:355–383 375

Proof of Theorem 4.1 We first consider the first two terms of the method in (1.15)
for n = 1, . . . , N t − 1, with ! ∈ Sh and U replaced by ξ = U − W. Using η = u − W,
and (3.1),

!n+ 12 %ξ n+ 12 , %!)h
− (∂t ξ n , %!)h + (AU
1 1
= −(∂t U n , %!)h + (AU
!n+ 2 %U n+ 2 , %!)h
1 1
+ (∂t W n , %!)h − (AU
!n+ 2 %W n+ 2 , %!)h
1
!n+ 2 , %!)h
= −(F (tn+ 1 )U
2

+ (∂t (un − ηn ), %!)h


1 1 1 1
un+ 2 %W n+ 2 , %!)h − (AU
+ [(A! !n+ 2 %W n+ 2 , %!)h ]
1 1 1 1 1 1
un+ 2 %W n+ 2 , %!)h − (Aun+ 2 %W n+ 2 , %!)h ] − (Aun+ 2 %W n+ 2 , %!)h .
− [(A!
(6.15)

Further using (1.14) we get

1 1
− (Aun+ 2 %W n+ 2 , %!)h
1
= (1/2)[(Aun+1 %(W n+1 − W n ), %!)h − (Aun+ 2 %(W n+1 − W n ), %!)h ]
1 1 1
− ([Aun+ 2 − (Au)n+ 2 ]%W n , %!)h − ((Au%W)n+ 2 , %!)h . (6.16)

Adding and subtracting F (tn+ 1 )u(tn+ 1 ) in (6.15) and using (6.16), (3.1), and (1.1), we
2 2
get the following identity for n = 1, . . . , N t − 1

!n+ 12 %ξ n+ 12 , %!)h = (∂t un − ut (tn+ 1 ), %!)h − (∂t ηn , %!)h


− (∂t ξ n , %!)h + (AU
2

n+ 12 n+ 12 ! n+ 12 n+ 12
u
+ [(A! %W , %!)h − (AU %W , %!)h ]
n+ 12 n+ 12 n+ 12 n+ 12
u
− [(A! %W , %!)h − (Au %W , %!)h ]
1
+ (1/2)[(Aun+1 %(W n+1 − W n ), %!)h − (Aun+ 2 %(W n+1 − W n ), %!)h ]
1 1 1
− ([Aun+ 2 − (Au)n+ 2 ]%W n , %!)h + ((Au%u)(tn+ 1 ) − (Au%u)n+ 2 , %!)h
2

!n+ 12 , %!)h .
+ (F (tn+ 1 )u(tn+ 1 ) − F (tn+ 1 )U (6.17)
2 2 2

1
With ! = ξ n+ 2 in (6.17), we obtain

8
"
1 1 1 1
−(∂t ξ n , %ξ n+ 2 )h + (AU
!n+ 2 %ξ n+ 2 , %ξ n+ 2 )h = Iin , n = 1, . . . , N t − 1, (6.18)
i=1
376 Numer Algor (2006) 43:355–383

where

1 1
I1n = (∂t un − ut (tn+ 1 ), "ξ n+ 2 )h , I2n = −(∂t ηn , "ξ n+ 2 )h , (6.19)
2

1 1 1 1 1 1
I3n = (A!
un+ 2 "W n+ 2 , "ξ n+ 2 )h − (AU
!n+ 2 "W n+ 2 , "ξ n+ 2 )h , (6.20)
1 1 1 1 1 1
I4n = (Aun+ 2 "W n+ 2 , "ξ n+ 2 )h − (A!
un+ 2 "W n+ 2 , "ξ n+ 2 )h , (6.21)
1
I5n = (1/2)[(Aun+1 "(W n+1 − W n ), "ξ n+ 2 )h
1 1
− (Aun+ 2 "(W n+1 − W n ), "ξ n+ 2 )h ], (6.22)
1 1 1
I6n = −([Aun+ 2 − (Au)n+ 2 ]"W n , "ξ n+ 2 )h , (6.23)
n+ 12 n+ 12
I7n = ((Au"u)(tn+ 1 ) − (Au"u) , "ξ )h , (6.24)
2

1 1
I8n = (F (tn+ 1 )u(tn+ 1 ) − F (tn+ 1 )U
!n+ 2 , "ξ n+ 2 )h . (6.25)
2 2 2

For the first term on the left hand side of (6.18), we use Lemma 2.3 to obtain

1 1 n+1
−(∂t ξ n , "ξ n+ 2 )h = − (ξ − ξ n , "ξ n+1 + "ξ n )h

1
= − {(ξ n+1 , "ξ n+1 )h − (ξ n , "ξ n )h }. (6.26)

For the second term on the left hand side of (6.18), we use (1.16), a ≥ amin > 0 and
Lemma 2.4 to obtain

!n+ 12 "ξ n+ 12 , "ξ n+ 12 )h ≥ C∥"ξ n+ 12 ∥2h ≥ C∥"ξ n+ 12 ∥20 .


(AU (6.27)

We bound terms on the right hand side of (6.18) starting with I1n . Since

⎡ ⎤
$ t 1 $ tn+1
1 ⎣ n+
∂t un − ut (tn+ 1 ) = (s − tn )2 uttt ds + (s − tn+1 )2 uttt ds⎦ ,
2

2 2τ tn tn+ 1
2

using (2.3) and Lemma 2.3, the triangle and ϵ inequalities, we obtain

1 1
|I1n | ≤ C∥ut (tn+ 1 ) − ∂t un ∥h ∥"ξ n+ 2 ∥h ≤ C∥ut (tn+ 1 ) − ∂t un ∥∞ ∥"ξ n+ 2 ∥0
2 2
$ tn+1
1 1
≤ Cτ ∥uttt ∥∞ ds∥"ξ n+ 2 ∥0 ≤ ϵ1 ∥"ξ n+ 2 ∥20 + C(u)ϵ1−1 τ 4 . (6.28)
tn
Numer Algor (2006) 43:355–383 377

Next we bound I2n . Since for any φ ∈ C 1 [0, T] and any t′ , t′′ ∈ [0, T],
! t′′
′′ ′
φ(t ) − φ(t ) = φt ds, (6.29)
t′

using (3.27), we have


! tn+1
" n"
"∂t η " ≤ Cτ −1
h
∥ηt ∥h ds ≤ C(u)hr . (6.30)
tn

Hence, (2.3), (6.30), Lemma 2.3, and the ϵ inequality yield


1 1 1
|I2n | ≤ C∥∂t ηn ∥h ∥&ξ n+ 2 ∥h ≤ C(u)hr ∥&ξ n+ 2 ∥0 ≤ ϵ2 ∥&ξ n+ 2 ∥20 + C(u)ϵ2−1 h2r . (6.31)
Using (6.28) and (6.31), we get
1 # $# $
|I1n | + |I2n | ≤ (ϵ1 + ϵ2 )∥&ξ n+ 2 ∥20 + C(u) ϵ1−1 + ϵ2−1 h2r + τ 4 . (6.32)
Using (6.20), (2.3), Lemma 3.4, (1.5), Lemma 2.3, the triangle inequality, U n − un =
ξ n − ηn , and Theorem 3.2, we obtain
1 1
|I3n | ≤ C∥&W n+ 2 ∥∞,Th ∥A% %n+ 12 ∥h ∥&ξ n+ 12 ∥h
u n+ 2 − AU
1
%n+ 12 ∥h ∥&ξ n+ 12 ∥0
u n+ 2 − U
≤ C(u)d(n, u, U)∥%
n
& 1
≤ C(u)d(n, u, U) (∥ξ i ∥h + ∥ηi ∥h )∥&ξ n+ 2 ∥0
i=n−1
' n
(
& 1
≤ C(u)d(n, u, U) ∥ξ i ∥0 + hr+1 ∥&ξ n+ 2 ∥0 ,
i=n−1

and thus, the ϵ inequality, and Lemma 2.3 yield


' n
(
1
&
|I3n | ≤ ϵ3 ∥&ξ n+ 2 ∥20 + C(u)ϵ3−1 [d(n, u, U)]2 h 2r+2
+ i
(ξ , −&ξ )h . i
(6.33)
i=n−1

Using (6.21), (2.3), Lemma 3.4, (1.5), and Lemma 2.3, we get
1 1 1 1
|I4n | ≤ C∥&W n+ 2 ∥∞,Th ∥A%
u n+ 2 − Aun+ 2 ∥h ∥&ξ n+ 2 ∥h
1 1 1
u n+ 2 − un+ 2 ∥h ∥&ξ n+ 2 ∥0 .
≤ C(u)∥% (6.34)
′ ′′
Since, for any φ ∈ C 2 [0, T], and t′ , t′′ , t∗ ∈ [0, T], where t∗ = t +t 2
,
)! ∗ ! t′′ *
t
φ(t′ ) + φ(t′′ ) ∗ 1 ′ ′′
− φ(t ) = (s − t )φtt ds − (s − t )φtt ds , (6.35)
2 2 t′ t∗

using (1.14),
" n+1 "
n+ 12 n+ 12
"u
" + un−1 "
n"
u
∥% −u ∥h = " −u "
2 h
"! tn ! tn+1 "
1""
" 2
= " (s − tn−1 )utt ds − (s − tn+1 )utt ds"
" ≤ C(u)τ . (6.36)
2 tn−1 tn h
378 Numer Algor (2006) 43:355–383

Similarly using (6.35), we obtain


1
∥un+ 2 − u(tn+ 1 )∥h ≤ C(u)τ 2 . (6.37)
2

Using (6.34), (6.36), and the ϵ inequality, we get


1
|I4n | ≤ ϵ4 ∥#ξ n+ 2 ∥20 + C(u)ϵ4−1 τ 4 . (6.38)

Using (6.22), (2.3), (1.5), (6.29) (with φ replaced by u and W respectively), Lemmas
2.3 and 3.4, and the ϵ inequality, we have
1 1
|I5n | ≤ |([Aun+1 − Aun+ 2 ]#(W n+1 − W n ), #ξ n+ 2 )h |
1 1
≤ ∥Aun+1 − Aun+ 2 ∥∞ ∥#W n+1 − #W n ∥h ∥#ξ n+ 2 ∥h
1 1
≤ C(u)∥un+1 − un+ 2 ∥∞ ∥#W n+1 − #W n ∥h ∥#ξ n+ 2 ∥0
! tn+1 ! tn+1
1 1
≤ C(u) ∥ut ∥∞ ds ∥#Wt ∥h ds∥#ξ n+ 2 ∥0 ≤ C(u)τ 2 ∥#ξ n+ 2 ∥0
tn tn

n+ 12
≤ ϵ5 ∥#ξ ∥20 + C(u)ϵ5−1 τ 4 . (6.39)

Next (6.23), (2.3), Lemmas 2.3 and 3.4, and the triangle inequality, yield
1 1 1
|I6n | ≤ ∥Aun+ 2 − (Au)n+ 2 ∥∞ ∥#W n ∥h ∥#ξ n+ 2 ∥h
1 1 1
≤ C(u)∥Aun+ 2 − (Au)n+ 2 ∥∞ ∥#ξ n+ 2 ∥0
1 1 1
≤ C(u)[∥Aun+ 2 − Au(tn+ 1 )∥∞ + ∥Au(tn+ 1 ) − (Au)n+ 2 ∥∞ ]∥#ξ n+ 2 ∥0 . (6.40)
2 2

Using (1.5) and (6.37), we obtain


1 1
∥Aun+ 2 − Au(tn+ 1 )∥∞ ≤ C(u)∥un+ 2 − u(tn+ 1 )∥∞ ≤ C(u)τ 2 . (6.41)
2 2

Using (6.35) with Au in place of φ,


⎡ ⎤
! ! tn+1
n+ 12 1 ⎣ tn+ 21
Au(tn+ 12 ) − (Au) =− (s − tn )Att u ds − (s − tn+1 )Att u ds⎦ ,
2 tn tn+ 1
2

∂2
where [Att u(t)] = ∂t2
[Au(t)], and hence the triangle inequality and (6.29) give
! tn+1
n+ 12
∥Au(tn+ 1 ) − (Au) ∥∞ ≤ Cτ ∥Att u∥∞ ds ≤ C(u)τ 2 . (6.42)
2
tn

Using (6.40), (6.41) and (6.42), and the ϵ inequality, we obtain


1 1
|I6n | ≤ CC4 (u)τ 2 ∥#ξ n+ 2 ∥0 ≤ ϵ6 ∥#ξ n+ 2 ∥0 + C(u)ϵ6−1 τ 4 . (6.43)
Numer Algor (2006) 43:355–383 379

Using (2.3), Lemma 2.3, (6.35) with Au!u in place of φ, and the ϵ inequality we get
1 1
|I7n | ≤ C∥(Au!u)(tn+ 1 ) − (Au!u)n+ 2 ∥∞ ∥!ξ n+ 2 ∥0
2
! tn+1
1 1
≤ Cτ ∥(Au!u)tt ∥∞ ds∥!ξ n+ 2 ∥0 ≤ C(u)τ 2 ∥!ξ n+ 2 ∥0 .
tn
1
≤ ϵ7 ∥!ξ n+ 2 ∥20 + C(u)ϵ7−1 τ 4 . (6.44)

Using (6.39), (6.43), and (6.44), we obtain


7
# 7 $ # 7 $
" " "
n+ 12 2
|Iin | ≤ ϵi ∥!ξ ∥0 + C(u) −1
ϵi τ 4. (6.45)
i=5 i=5 i=5

To bound I8n in (6.25). we use (1.16), (2.3), the triangle inequality, and (1.6), (1.7) and
(1.8),
% & % & % & n+ 1
|I8n | ≤ C∥F tn+ 1 u tn+ 1 − F tn+ 1 U' 2 ∥h ∥!ξ n+ 12 ∥h
2 2 2
(%) % & % & % % & % & n+ 12 &)
≤ C )F tn+ 1 u tn+ 1 − f ., tn+ 1 , u tn+ 1 , ux tn+ 1 , U
'y )
h
2 2 2 2 2

) % % & % & n+ 12 & % % & n+ 12 1 &)


+ ) f ., tn+ 1 , u tn+ 1 , ux tn+ 1 , U
'y 'x , U
− f ., tn+ 1 , u tn+ 1 , U 'yn+ 2 )
2 2 2 2 2 h
) % % & n+ 12 1& % *
& n+ 1 ) ) n+ 1 )
+ ) f ., tn+ 1 , u tn+ 1 , U
'x , U 'yn+ 2 − F tn+ 1 U ' 2 ) )!ξ 2 )
2 2 h 2 h
+
) % & ) ) % &
'n+ 12 ) + )ux tn+ 1
≤ Cd(n, u, U) )u tn+ 1 − U h 2 2

,
n+ 12 ) ) % & n+ 12 ) ) )
'
−U x ) ) '
+ u y tn+ 1 − U y ) )!ξ n+ 12 ) . (6.46)
h 2 h h

The relation u(tn+ 1 ) − U 'n+ 12 = [u(tn+ 1 ) − un+ 12 ] + [un+ 12 − ' 1


un+ 2 ] + '
1
ηn+ 2 − '
1
ξ n+ 2 , the
2 2
triangle inequality, Lemma 2.3, (6.36), (6.37), and Theorem 3.2, give
( *
'n+ 12 ∥h ≤ C ∥u(tn+ 1 ) − un+ 12 ∥h + ∥un+ 12 − '
∥u(tn+ 1 ) − U
1
un+ 2 ∥h + ∥'
1
ηn+ 2 ∥h + ∥'
1
ξ n+ 2 ∥h
2 2
- 1
.
≤ C C(u)[τ 2 + hr+1 ] + ∥' ξ n+ 2 ∥0 . (6.47)

Similarly with ux (or u y ) and U x (or U y ) in place of u and U respectively, we obtain


/ 1 1 0 - .
'xn+ 2 ∥h , ∥u y (tn+ 1 ) − U
max ∥ux (tn+ 1 ) − U 'yn+ 2 ∥h ≤ C C(u)[τ 2 + hr ] + ∥' 1
ξ n+ 2 ∥1 .
2 2

(6.48)

Using (6.46), (6.47) and (6.48), the ϵ inequality, and Lemma 2.3, we get
# n
$
1
"
|I8n | ≤ ϵ8 ∥!ξ n+ 2 ∥20 + C(u)ϵ8−1 [d(n, u, U)]2 τ 4 + h2r + (ξ i , −!ξ i )h . (6.49)
i=n−1
380 Numer Algor (2006) 43:355–383

Using (6.18), (6.26), (6.27), (6.32), (6.33), (6.38), (6.45), (6.49), and the triangle
inequality, we obtain for n = 1, . . . , N t − 1,
! 8 #
1 1
" 1
− [(ξ n+1 , #ξ n+1 )h +(ξ n , #ξ n )h ] + C∥#ξ n+ 2 ∥20 ≤ ϵi ∥#ξ n+ 2 ∥20
2τ i=1
! 8
# $ n+1
%
" "
2 −1 4 2r i i
+ C(u) [d(n, u, U)] + ϵi τ +h + (ξ , −#ξ )h ,
i=1 i=n−1

and hence, taking ϵi , i = 1, · · · , 8, sufficiently small, and multiplying through by 2τ ,


we obtain for n = 1, . . . , N t − 1,
$ n+1
%
"
n+1 n+1 n n 2 4 2r i i
−(ξ , #ξ )h + (ξ , #ξ )h ≤ C(u)[d(n, u, U)] τ τ + h + (ξ , −#ξ )h . ⊓
%
i=n−1

We follow arguments used in proving Theorem 4.1, to prove Lemma 4.2 for the
predictor-corrector scheme (1.17) and (1.18). Main changes are in obtaining concrete
error bounds for all terms involved in the splitting first two terms in (1.17) and (1.18)
with appropriate test functions and error terms. Below we show that the predictor
scheme (1.17) yields only the sub-optimal τ 3/2 error in time; this is then corrected by
using (1.18) to obtain optimal second-order in time error bound.

Proof of Lemma 4.2 Let & ξ n = V n − W n , n = 0, 1. Then, as in the first part of proof
of Theorem 4.1, using (1.17) and (1.18), we have
' ( ' 1
( ' (
− ∂t&ξ 0 , #& + AV 0 #& ξ 2 , #& = ∂t u0 − ut (t 1 ), #&
h h 2 h
1 1
− (∂t η0 , #&)h + [(Au0 #W , #&)h − (AV 0 #W , #&)h ]
2 2

)' 1
( ' 1 1
( * )' ' ( (
− Au0 #W 2 , #& − Au 2 #W 2 , #& +(1/2) Au1 # W 1 − W 0 , #&
h h h
' 1
' ( ( * ') 1 1
* ( ' ' (
− Au 2 # W 1 − W 0 , #& − Au 2 − (Au) 2 #W 0 , #& + (Au#u) t 1
h h 2

' ( 12 ( ' ' ( ' ( ' ( (


− Au#u , #& + F t 1 u t 1 − F t 1 V 0 , #& , & ∈ Sh , (6.50)
h 2 2 2 h

and
' ( ' 1 1
( ' ' ( ( ' (
− ∂t ξ 0 , #& + AV 2 #ξ 2 , #& = ∂t u0 − ut t 1 , #& − ∂t η0 , #&
h h 2 h h
)' 1 1
( ' 1 1
( *
+ Au 2 #W 2 , #& − AV 2 #W 2 , #&
h h
)' ( ' 1
( *
+ (1/2) Au1 #(W 1 − W 0 ), #& − Au 2 #(W 1 − W 0 ), #&
h h
') 1 1
* ' ' ( 1
(
− Au 2 − (Au) 2 #W 0 , #&)h + (Au#u) t 1 − (Au#u) 2 , #&
2 h
' ' ( ' ( ' ( 1 (
+ F t 1 u t 1 − F t 1 V 2 , #& , & ∈ Sh . (6.51)
2 2 2 h
Numer Algor (2006) 43:355–383 381

1
Taking ! = !
ξ 2 in (6.50), we obtain
8
"
1 1 1
ξ 0 , $!
−(∂t! ξ 2 )h + (AV 0 $!
ξ 2 , $!
ξ 2 )h = Ii0 , (6.52)
i=1

where Ii0 , i = 1, · · · , 8 are defined as in (6.19), (6.20), (6.21), (6.22), (6.23), (6.24)
1
and (6.25), with n = 0, ξ , # # 12 replaced by !
u 2 and U ξ , u0 and V 0 respectively. Using
0 0
V = W and Lemma 3.4,

∥V 0 ∥1,∞ = ∥W 0 ∥1,∞ ≤ C(u). (6.53)


1
Following the derivations of (6.34) and (6.49) with n = 0, and u0 , V 0 , and $ξ̃ 2 , in
1
place of # u2 , U# 12 , and $ξ 12 , respectively, and using the relation V 0 − u0 = −η0 , and
(6.53), we obtain
$ % 1
|I40 | + |I80 | ≤ C(u) ∥u1 − u0 ∥1,∞ + ∥u0 − u(t 1 )∥1,∞ + hr ∥$! ξ 2 ∥0 . (6.54)
2

Using (6.29),
& t1
∥u1 − u0 ∥1,∞ + ∥u0 − u(t 1 )∥1,∞ ≤ C(u) ∥ut ∥1,∞ ds ≤ C(u)τ. (6.55)
2
0

Hence, using (6.54), (6.55), and the ϵ inequality, we have


1 ' (
ξ 2 ∥20 + C(u)ϵ4−1 h2r + τ 2 .
|I40 | + |I80 | ≤ ϵ4 ∥$! (6.56)
1
Using (6.52), and (6.26), (6.27), (6.33), (6.32), (6.45), with n = 0 and u0 , ξ̃ 2 , and V 0
1 1
in place of #
u 2 , ξ 2 , and U# 12 , respectively, we get
) 3 7
*
1 1 1
" " 1
− (! ξ , $ξ̃ 1 )h + C∥$! ξ 2 ∥20 ≤ ϵi + ϵi ∥$! ξ 2 ∥20 + |I40 | + |I80 |
2τ i=1 i=5
+ 3 7
,
" "
−1 −1
+C C(u) + ϵi + ϵi
i=1 i=5
- 2 2r
.
× τ + h + (ξ , −$ξ )h , !1 !1

and hence, (6.56), taking ϵi , i = 1, · · · , 7, sufficiently small, and multiplying through


by 2τ yield
- .
−(ξ̃ 1 , $ξ̃ 1 )h ≤ C(u) τ 3 + τ h2r + τ (! ξ 1 , −$!
ξ 1 )h ,

and consequently, for τ sufficiently small, we get


' (
− (!ξ 1 , $!
ξ 1 )h ≤ C(u) τ 3 + h2r . (6.57)
1
The relation V 2 = (W 1 + W 0 + ! ξ 1 )/2, the triangle and inverse inequalities, Lemmas
2.3 and 3.4, (6.57), and τ 3 ≤ Ch2 we get
1
/ 0
∥V 2 ∥1,∞ ≤ C ∥W 0 ∥1,∞ + ∥W 1 ∥1,∞ + h−1 ∥ξ̃ 1 ∥1 ≤ C(u). (6.58)
382 Numer Algor (2006) 43:355–383

1
Next, we use (6.57) and (6.58) to bound −(ξ 1 , "ξ 1 )h . Taking v = ξ 2 in (6.51), we
obtain
8
!
1 1 1 1
0
−(∂t ξ , "ξ )h + (AV "ξ , "ξ )h =
2 2 2 2 Ii0 , (6.59)
i=1

where Ii0 , i = 1, · · · , 8 are defined in (6.19), (6.20), (6.21), (6.22), (6.23), (6.24) and
1
(6.25), with "u 2 and U " 12 replaced by u 12 and V 12 respectively. Following (6.31) (6.32)
1
and (6.33), (6.49) with n = 0, V 2 in place of U " 12 , u 12 in place of " 1
u 2 , and using the
1 1 1 1
relation V 2 − u 2 = ξ̃ 2 − η 2 , ξ̃ 0 = 0, and (6.58), we obtain
1 # %
|I30 | + |I80 | ≤ ϵ3 ∥"ξ 2 ∥20 + C(u)ϵ3−1 h2r + τ 4 + ($ ξ 1 , −"$
ξ 1 )h . (6.60)

Using (6.59), ξ 0 = 0, (6.26), (6.27), (6.32), (6.38), (6.45), with n = 0, (6.58), (6.60),
taking ϵi , i = 1, · · · , 7, sufficiently small, and multiplying through by 2τ , we obtain
& '
−(ξ 1 , "ξ 1 )h ≤ C(u) τ 5 + τ h2r + τ (ξ 1 , −"ξ 1 )h + τ (ξ̃ 1 , −"ξ̃ 1 )h ,

and hence, for τ sufficiently small, the desired result follows from (6.57). %

References

1. Aitbayev, R.: Multilevel preconditioners for a quadrature Galerkin solution of a biharmonic


problem. Numer. Methods Partial Differ. Equ. 22, 847–866 (2006)
2. Aitbayev, R.: Convergence analysis of a quadrature finite element Galerkin scheme for a bihar-
monic problem. IMA J. Numer. Anal. (to appear)
3. Baker, G.: Simplified proofs of error estimates for the least squares method for Dirichlet’s
problem. Math. Comput. 327, 229–235 (1973)
4. Bramble, J., Schatz, A.: Rayleigh–Ritz–Galerkin methods for Dirichlet’s problem using sub-
spaces without boundary conditions. Commun. Pure Appl. Math. 23, 653–675 (1970)
5. Bramble, J., Schatz, A.: Least-squares methods for 2mth order elliptic boundary value problems.
Math. Comput. 25, 1–32 (1971)
6. Bialecki, B., Ganesh, M., Mustapha, K.: A Petrov-Galerkin method with quadrature for elliptic
boundary value problems. IMA J. Numer. Anal. 24, 157–177 (2004)
7. Bialecki, B., Fairweather, G.: Orthogonal spline collocation methods for partial differential
equations. J. Comput. Appl. Math. 128, 55–85 (2001)
8. Blum, H., Rannacher, R.:On the boundary value problem of the biharmonic operator on domains
with angular corners. Math. Methods Appl. Sci. 2, 556–581 (1980)
9. Ciarlet, P.G.: The Finite Element Method for Elliptic Problems. North-Holland, Amsterdam
(1978)
10. Douglas, J., Dupont, T.: Collocation method for parabolic equations in a single variable. In:
Lecture Notes in Math, vol. 385. Springer, Berlin Heidelberg New York (1974)
11. Douglas, J., Dupont, T., Rachford, H.H., Wheeler, M.F.: Local H −1 Galerkin and adjoint local
H −1 Galerkin procedures for elliptic equations. RAIRO Anal. Numer. 11, 3–12 (1977)
12. Fairweather, G.: Finite element Galerkin methods for differential equations. In: Lecture Notes
in Pure and Applied Mathematics, vol. 34. Marcel Dekker, New York (1978)
13. Grigorieff, R., Sloan, I.H., Brandts, J.H.: Superapproximation and commutator properties of
discrete orthogonal projections for continuous splines. J. Approx. Theory 107, 244–267 (2000)
14. Gilbarg, D., Trudinger, N.S.: Elliptic Partial Differential Equations of Second Order. Springer,
Berlin Heidelberg New York (1977)
15. He, Y., Jin, M., Gu, X., Qin, H.: A C 1 globally interpolatory spline of arbitrary topology. Lect.
Notes Comput. Sci. 3752, 295–306 (2005)
16. He, Y., Gu, X., Qin, H.: Automatic shape control of triangular B-splines of arbitrary topology.
J. Comput. Sci. Technol. 21, 232–237 (2006)
Numer Algor (2006) 43:355–383 383

17. Hackbusch, W.: Elliptic Differential Equations, Theory and Numerical Treatment. Springer,
Berlin Heidelberg New York (1992)
18. Ladyzhenskaya, O., Uraltseva, N.: Linear and Quasilinear Elliptic Equations. Academic, New
York (1968)
19. Ma, N., Lu, T., Yang, D.: Analysis of incompressible miscible displacement in porous media by
characteristics collocation method. Numer. Methods Partial Differ. Equ. 22, 797–814 (2006)
20. Percell, P., Wheeler, M.F.: A C 1 finite element collocation method for elliptic equations. SIAM
J. Numer. Anal. 17, 605–622 (1980)