NASA Contractor Report 198290

ICASE Report No. 96-13

ICASE
NUMERICAL ALGORITHMS BASED
ON BIORTHOGONAL WAVELETS

Pj. Ponenti t
XLiandrat ~mmm^m^^± i
• Approval fox paWte nlmam j
OUtabtOom DnMmlfd

NASA Contract No. NAS1-19480

February 1996
Institute for Computer Applications in Science and Engineering
NASA Langley Research Center
Hampton, VA 23681-0001
Operated by Universities Space Research Association

National Aeronautics and

19960426 065
Space Administration
Langley Research Center
Hampton, Virginia 23681-0001
 NUMERICAL ALGORITHMS BASED ON
BIORTHOGONAL WAVELETS

Pj. Ponenti
IMA
Institute for Mathematics and Its Applications
University of Minnesota
206 Church St. S.E., 514 Vincent Hall
Minneapolis, MN

/. Liandrat*
IRPHE
12 avenue general Leclerc
13003 Marseille, France
and ESM2, Universite d'Aix Marseille II, IMT
13451 Marseille, France

Abstract

Wavelet bases are used to generate spaces of approximation for the resolution
of bidimensional elliptic and parabolic problems. Under some specific hypothe-
ses relating the properties of the wavelets to the order of the involved operators,
it is shown that an approximate solution can be built. This approximation is
then stable and converges towards the exact solution. It is designed such that
fast algorithms involving biorthogonal multi resolution analyses can be used to
resolve the corresponding numerical problems.
Detailed algorithms are provided as well as the results of numerical tests on
partial differential equations defined on the bidimensional torus.

"This research was supported by the National Aeronautics and Space Administration under
NASA Contract No. NAS1-19480 while the second author was in residence at the Institute
for Computer Applications in Science and Engineering (ICASE), NASA Langley Research
Center, Hampton, VA 23681-0001.
 Introduction

Variational approximation methods for partial differential equations are based

on weak formulations and on suitable spaces of approximation.
Wavelets are known to be unconditional bases for a large variety of spaces
and therefore are good candidates for the generation of approximation spaces
for partial differential equation problems. The goal of this paper is to show that
moreover, wavelet bases may lead to fast and adaptive numerical resolution of
the corresponding approximations.
In this paper, as in previous papers (J. Liandrat and P. Tchamitchian [10],
[11]), the wavelets are used to expand the approximated solution of a partial
differential equation as well as to approximate the kernel of the differential
operator. They are not used only to perform "the linear algebra" (G. Beylkin
[2]) related to more classical methods of resolution.
Starting with an expansion of the form / = Y,\ < /> ** > *A> the solution
of the equation Lu — f where L is a constant coefficient elliptic differential
operator is u{x) = J f(y)K(x,y)dy where K(x,y) = J2\ L'^K^iy)-
Under suitable conditions that will be made precise later, the functions
L~x^*x as well as M\ are pseudo wavelets, very close to wavelets (Y. Meyer
[13]). This turns out to provide a stable approximation of u. However, the effi-
ciency of the corresponding numerical approximation of u relies, at least in this
work, on the hierarchic structure of multiresolution analysis since it provides
fast tree algorithms. We will show that, if the operator satisfies suitable condi-
tions that will be made explicit later, then the above mentioned pseudo wavelets
are directly related to biorthogonal multiresolution and wavelets. Under these
conditions, competitive numerical algorithms involving 0{N) or 0(N log N) op-
erations can then be derived.
This paper provides the analysis of the problem and exhibits the correspond-
ing numerical schemes. It is then organized as follows.
The first part is devoted to the general concept of biorthogonal multiresolu-
tion analysis on L2(lRn). In the second part we focus on the problem of the sta-
bility of the multiresolution framework under the action of constant coefficient
elliptic operators. The cases of homogeneous and non homogeneous operators
are treated separately. The third part deals with the numerical algorithms while
the last section is devoted to numerical tests related to the resolution of elliptic
and parabolic equations in bidimensional spaces.

I Generalities: Biorthogonal Multi-Resolution

Analysis in L2(Mn)
The concept of multiresolution is at the basis of our construction and we there-
fore start with a short description of it :
Definition Ll (Y. Meyer, [12])
A r-regular multiresolution analysis of L2(lRn) is a sequence of increasing
closed subspaces Vj, j G %, Vj C Vj+i, satisfying the following conditions

i) n+~ Vj = {0}, U-~ Vj is dense in L2(Mn);

ii) f(x)eVj^=^f(2x)€Vj+1;
in) f(x) ev0<=* f(x -k)ev0, vke %n;
iv) there exists a function $ in Vo, such that the set of functions {$(x-k), k G
%n}, is a Riesz basis t for Vo;
v) the function $ is regular and localized : $ is Cr~l, $(r_1) is almost
everywhere differentiable, and for almost every x G Mn, for every integer
a <r and for all integer p, it exists Cp such that

\da$(x)\<Cp(l + \x\)-" . (1)

A consequence of ii), in) and iv) is that each Vj is generated by the family
of functions {$,■*(*) = 2'n/2*(2i* - k), k£ Zn}.
For simplicity reasons, we will only consider the case n = 2, but all the
results presented to this article can be generalized in any dimension. We will
always use for vectors a contracted notation: if e is a bidimensional vector then
e = (ei,e2)-

1.1 Orthogonal multiresolution analysis

To build an orthonormal multiresolution analysis, the Riesz basis {$(. — k), k G
Z2} is first orthonormalized in such a way that the resulting orthonormal basis
is still of the form {$(. -*),*€ Z2}.
The wavelets are introduced via the orthogonal complement of VQ in V\: Wo-
More precisely, if E is the set of all vertices in the unit cube [0, l]2, and if
E* = E \ {0}, we have the following theorem:

Theorem 1.1 There are 3 functions *£, s G E*, in Wo, such that the collection
{^(x-k), k G Z2 ,e G E*} is an orthonormal basis of Wo. Moreover, each \P£
satisfies the same property (1) of regularity and localization as $ and, moreover,
satisfies the following cancelation property
3 m G M , such that V k = (kx, k2) &Z2, 0 < h < m,

xkV*(x)dx = 0 .
Jm2
*A collection of vectors {e^, A € A}, in a Hubert space H is a Riesz basis if any vector
x G H can be written in a unique way asi = ^]o> ej where (^ I^AI2)1'2 is finite and
defines a norm equivalent to [|I||H-
 The scaling invariance property it) implies that, for all j, the family{*jfc, k G
2
Z , e G E*} is an orthonormal basis of Wj. We will also use the following
contracted notation: {#A, A G A,} where A;- = {A = 2~' (k + §) , k G ^2,£ G
£"*}. Indeed, there is a straightforward bijection between Aj and the set of pairs
{(e, fc), fc G ^2,£ G -E1*}- We will also use the following sets: A = UjgzAj and

From the inclusion Vo C Vi the following scaling relation can be derived:

$(x) = J2 W* - 0 .
while from Wo C Vi one obtains the following detail relation:

*'(*) = 53 0?$(2a: -/), Ve G £* .

It is very useful to transform these relations using the Fourier transform

which is given by the equality

HO = / f{*)e-*'dx.
Jm.
Indeed, the scaling relations then become

no = Moiawm) (3)
and
$e(£) = M£(£/2)$(£/2) (4)
where M0(£) = £,eZ2 AJC"*'«-') and M£(£) = £,eZ2 fff c"<(€"° are C0
° 2,r
Peri"
odic functions.
This leads to :

$(0 = IIM#)' and^(0 = Me(|)n^o(^r). (5)

3=1 J=l

The following conditions are satisfied

J3 M£(^ + 7re)Me-(C + ire) = 6„> , V (e,e') G £2 (6)

and
M,(7re) = ^1,ei^2,e2, V(£,e)G£2 (7)
and are called, following electrical engineering terms, the quadrature mirror filter
conditions. We will also call the functions Ms quadrature mirror filters.
 Conversely, it is shown in A. Cohen et al. [4] that four 2ir periodic functions
M£(f), £ £ E satisfying the quadrature mirror filter conditions (6) generate
through (5) an orthogonal multiresolution analysis if some specific conditions
are satisfied.
Remark:

• In this paper, we will often use specific multi-resolution analyses oiL2(JR?)

based on a tensorial product of multiresolution analyses of L2(IR). More
precisely, such analyses are defined as follows: if (Vj) is the sequence of
spaces of a ID multiresolution analysis and if Wj,<f, ip, mo and mi are re-
spectively the related wavelet spaces, the scaling function, the associated
wavelet and the quadrature mirror filters, then, the sequence of spaces
(Vj), defined as Vj — Vj ® Vj is a multiresolution analysis in M2. More-
over, $(a:) = <p(xt)ip(x2) is the corresponding scaling function; (Wj), with
Wj = J2e€B. Wj are the wavelet spaces; the three generating wavelets are
tfo,i(z) = <p{xi)^(x2), *i,i(a;) = ip(x1)4>(x2) and *i,0(a;) = VO^iM^);
M£(£) = m£l(^i)m£r2(^2) with s G E are the quadrature mirror filters.

1.2 Biorthogonal Approach

A relaxation of some properties of orthogonal multiresolution analysis can be
performed using the one biorthogonal approach. This approach provides some
flexibility since it allows to distribute the relevant properties of the multires-
olution (number of zero moments, compact support or regularity) to the two
involved multiresolution analyses. Moreover, it will turn out to be that the
biorthogonal framework is "stable" under the action of a large class of opera-
tors while the orthogonal framework is "fragile".

Definition 1.2 We call biorthogonal multiresolution analysis of L2(M2), two

multiresolution analysis (Uj)j^z and (Uj)j^z such that there exists two families
of corresponding scaling functions r and f such that: (T)-*, Tj'k') = bjj'bkk1 for
all j, f e 7Z and, k and k' G 2Z2.

In this case we define the wavelets spaces Xj and Xj as Uj+i = Uj © Xj,

Üj+i = Uj © Xj with UjLXj, UjLXj, and we introduce the functions 6\ =
2j6c(2j.- k) and 0\ = 2j0e(2j.- k), e G E, that generate respectively Xj and
Xj and such that {0\, 6\>) = 6eei6jj'Skk'-
Moreover, following the construction of orthogonal multiresolution analysis
we define 2x4 filters (i.e, C°° 2TT periodic functions), P£ and P£, e G E,
associated with the two biorthogonal multiresolution analyses. These filters
satisfy the biorthogonal quadrature mirror filter relations equivalent to (6) that
are
2
Vs ,e' ande'€E, £ € [0, 2TT] ,

P£(7Te') = <$£ljei<$e2,e2, Pr^e') = ££l,eA2,e2 .

As in the orthonormal case, the generalizations of relations (3) and (4) relate
the scaling functions and the wavelets to the filters as
oo

f(0 = P0(£/2)r(£/2) = Y[Po(2-J0, &(0 = Pe(£/2)?(£/2), (9)

?(0 = P0(£/2)%/2) = lpo(2_j0, *(0 = P(£/2)%2). (10)

As in section I.l,jhe question to know is under which conditions the biorthog-
onal filters Ps and P£ satisfying the quadrature mirror filter conditions (8) de-
fine two biorthogonal multiresolution analyses? Again, a specific condition is
required and has been formulated in A. Cohen et al. [4]. We will use a weaker
version of this formulation adapted to the case of functions with fast decay. It
can be expressed in term of the following theorem for which a complete proof
can be found in Pj. Ponenti [16] and W. Dahmen and A. Michelli [5].
Theorem 1.2 Let a > 0 and let Pe(£) and P£(£)> £ € E, be eight Ca 2TT periodic
functions satisfying the biorthogonal quadrature mirror filters conditions (8).
Defining r, f, 9, and 6 using formula (9) and (10), if
- there exist C and \i > 0 such that for all £ € IR2

\m\ < ca+Ki)-1-", I?(OI < c(n-iei)-1-", (ii)

- V k , etk' e %2 ,
{r{x-k),T(x-k')) = 6kk,, (12)

and if

then
• the sequences of sub spaces (Uj)j£z, and (Uj)j^z generated respectively
by {rjk,k € 2Z2} and {fjk,k € %2} are two biorthogonal multiresolution
analyses;
• the wavelet families { 9\(x) = 2J'*£(2J'a: — k), X £ A } and {0\(x) =
2Jö£(2-7x — k), A 6 A} are two biorthogonal Riesz bases of L2(M2).
II Constant Coefficient Elliptic Differential Op-
erators and Biorthogonal Multiresolution Anal-
ysis
II. 1 General results
The starting point of this section is the following remark. Given (#A) a family
of orthonormal wavelets and knowing / = ]T{/> #A)*A, the solution of the
equation
Lu = f, (14)
where L is an elliptic operator of order s is, at least formally,

«=X;{/.*A>^
_I
[*A] ■ (i5)
When L and L'1 are bounded on L2(R2), the families {L-1$A} and {£*#*}
are two biorthogonal Riesz bases.
The question we address now is related to what happens in the specific case
of wavelets when the operator L is unbounded (as in the case of a differential
operator).
In the following paragraphs, we first see that, assuming some compatibil-
ity conditions between \Pe and L, the two families of functions {Z-1\PA} and
{L*\PA} are wavelets or pseudo wavelets (Y. Meyer [13]) depending on whether
the operator is homogeneous or not. Then, we show that in some cases, a
biorthogonal framework embedding {L-1**} and {L*#A} can be built.
To be more precise, we take (Vj) an r-regular multiresolution of L2{IR?) con-
structed using a tensorial product of two ID multiresolutions and L a constant
coefficient elliptic differential operator. Let us write L - Y,o<a<s aaDa where
D is the operator \-§^. We define in a standard way the symbol of L as

*(0 = £a«*°- (16)

It is a polynomial in £ of degree s and, if / € C°°(iR2), we have the well known
formula
VtelR2,Lf(0 = f{0<r(S). (17)
We note formally
:£*** and#£ = I"1** (18)
and more generally,

e\ = 2->*V*\ and 6\ = V'L-1^. (19)

j c j
Note that when L is homogeneous, 9\(x) = 2 6 (2 x - k) and 6>A(z) =
2^6e(2^x - k) while, in general this is not true.
Then we have
Theorem II. 1 Given a family of r regular wavelets with m + 1 zero moments,
if L is a homogeneous operator or if L is an inhomogeneous operator with a
strictly positive symbol of orderj > 0, and such that r > s and m> s, then
Regularity: 9£ € Hr~s and 6e G Hr+'
Localization: for all multi-indices^ j' and j such that \j'\ < r — s and
\j\ < r + s — 1 and all integers I € IN,

i \dy'd{(x)\ < Cy.yM2> (l + 2>>-A|)"'

1 I^AOOI < CyVWV (1 + 2^|a; - Aj)-3+^-^-IT1

Cancelation: Let xk = (a^1, x22),

/ xk9e(x)dx = 0, 0<\k\<m+a

/ xkee(x)dx = 0, 0<\k\<m-a,

where a = s in the homogeneous case and a — 0 in the inhomogeneous one.

A complete proof of this theorem can be found in Y. Meyer [13] and Ph.
Tchamitchian [17].
Remark:
• Following Y. Meyer [13] and according to Theorem II.1, a factorization
of the operators L* and L~x can be performed as I* = C o T~s and
L~l = C o P where T is a diagonal operator in the wavelet basis denned
as: ip\ *-+ l^tpx and where C and C are bounded on L2 and denned by

C : VA - 0X , C : xl>x^ex. (20)

The operators C and C act just as a transformation between two bases.

The operator T is nothing else but the classical preconditioning operator
for elliptic problems (S. Jaffard [8]) that mimics a diagonal derivation in
the wavelet basis.
In other words, thanks to this factorization, the computation of the image
of a function by an elliptic operator or its inverse can be transformed into
a well conditioned problem using a diagonal operator in a suitable wavelet
basis. This is essential since it provides the numerical stability of the
further developed algorithm.

We can then rewrite (15) as

u = £V'''</, ^x)0x- (21)

t We call a multi-index any couple of integer 7' = (71,72)and 97 & =

axj1 dxl2
 An important issue, as far as numerical applications are concerned, is the
computation of the sum (21). Indeed, even if this sum corresponds to a pseudo
wavelet decomposition, fast algorithms for the computation of the sum are not
available. In the framework of this paper, the fast algorithms are linked to
the concept of multiresolution analysis presented in section I. We prove in
the following sections that, under suitable conditions, the construction of a
multiresolution analysis embedding the function 6\ is possible. Moreover, we
provide explicit expressions of the quadrature mirror filters required for the fast
implementation of (21).
The starting point of our construction is due to P.G. Lemarie [9] who con-
structed in the one dimensional case two biorthogonal multiresolution analyses
from an original orthogonal one and from the derivata operator. We generalize
this approach to any dimensions and for any homogeneous elliptic operator.
In contrast to the classical constructions of multiresolution analysis, this is
an inverse problem. Indeed, knowing the two dual wavelet bases we can define
two sequences of subspaces (Xj) and (Xj):

Xj = span{6>A, A 6 A;, / < j] ,^.

Xj = span{6\,\ € A;, / < j] ,

The open question is the following: how can we construct two sequences of sub-
spaces (Uj) and (Uj) for which (Xj) and (Xj) play the role of two biorthogonal
wavelet spaces?
In other words, the problem is jhe construction of the generalized scaling
functions r and f related to 6 and 6.
In the case of non homogeneous elliptic operators the approach used in the
homogeneous case can not be transposed and we will not be able to define a
multiresolution framework embedding the space sequences (Uj) and (Uj). How-
ever, we will show that the essential property of embedding spaces as well as
the existence of scaling (3) and detail (4) relations can be saved. This will allow
us to derive fast and stable algorithms to sum up (21) even in the case of non
homogeneous operators.

II.2 The case of homogeneous elliptic operator

In that case the natural candidates for r: L~l$\, are not defined in L2 for the
basic reason that $ does not belong to the range of L, or, in other words, suffers
from a lack of zero moments. Using a preconditioning operator, we will adapt
the function $ to the operator L~l (i.e. we will transform $ so that it enters
the range of L) while preserving the two scale relations (3) and (4). We will
then check that the resulting multiresolution analysis is fitted to the functions
6\ and 6\ previously defined.
More precisely, we have the following theorem:
Theorem II.2 Let (Vj) be the family of embedded spaces of an r-regular mul-
tiresolution analysis of L2{IR2)! let L be a homogeneous operator of order s and
of symbol a and let S be a 2TT periodic function, not vanishing on ]0, 2TT[2 and
equivalent to a in zero. Ifm>s+1, r > s+l, and if the eight functions P£(£)
and Pe(£), s E E defined as

P0(O = 2« =H Mo(0, Po(0 = ± ^ Mo(0 , (23)

1 M£(0
Pt(Z) = rS(t) Mt(£), R(0 = --^ (24)

are Ca, a > 0 then they satisfy the quadrature mirror filter conditions (8) and
define two biorthogonal multiresolution analyses. The corresponding wavelets
are the functions 6 and 9 and the scaling functions T and r are derived following
(9) and (10).

Proof:
O By construction all the filters are Ca with some a > 0 and they satisfy
the biorthogonal quadrature mirror filter conditions (8).
The only point to prove is the convergence in L2(IR2) of the infinite prod-
ucts (9) and (10) defining the two scaling functions. We will use the following
lemma:

Lemma II. 1 Let p(x), x € M2, be a homogeneous polynomial of degree s,

and let S and C, be 2w periodic functions; then the following propositions are
equivalent:

i) oo

ficc-'*) = 33.
TTno-iT.\ -
i=i
. . ..
p(x) o»)
S{x) = 2sC{x/2)S{x/2) , (26)
S(x) x^O
~ 'P(x) • (27)

Proof:
O The equation (26) is obtained from (25) written for x and x/2, while (27}
is derived from (25) when x —> 0 since necessarily C(0) = 1.
Conversely, (25) is obtained from (26) . Indeed, since

S{2 NX)
m - C{x/2)
C(x/2)^^- - ~ T\C(2-*x)
P(x) ~ p{x/2) ~ P(2-»x) ÜC(2 X)>
thanks to (27) we obtain (25) when N —► oo. ■

This lemma allows us to calculate the infinite product (9) and (10), and to
get:
CO OO - OO

r(0 = I] fi>(2-'0 = II^I77TII",«'(2"^0"»O(2--''6)

OO oo oo

^l $(0 ' ?(0 = Ü) $(0 '

and finally,
?(0 = (28)
The function $ being r-regular the conditions (11) are trivially satisfied.
Defining the function 9 et 9 by (9) and (10) we get

fr® = w2)?(£/2) = a(om

(29)
~ ~ m
The wavelet admissible condition (13) is immediately satisfied thanks to Theo-
rem II.1. Finally the assumption (12) is satisfied by construction, that completes
the proof. ■

Remarks:

• Note that here, thanks to homogeneity, the subscript \ recovers its classical
"wavelet" meaning since in that case, 9\(x) = 2^29{2^x — k) and 9\(x) =
2'l20{{2>x-k).
• The relation (25) is a generalization in any dimension, of the classical
formula
n coep-'*) = ^
used by P.G. Lemarie for the first order differential operator in one di-
mension.

• The function S can be interpreted as the symbol of a difference operator

that we will call Di. If S is a trigonometric polynomial, then Di is a
finite-difference operator and S is C°°. The fact that S(x) ~ <J{X) is
just the translation that DL is consistent with L. Moreover, S removes
exactly the singularity of f for £ = 0. Conversely for r, the singularity

10
 given by S at points (2wn,2irn), n £ 2Z, will be removed exactly by a in
zero and by the zeros of $ at points {lirn, 2irn), n £ "%*. Notice that this
last point won't be true for inhomogeneous operators.
From a certain point of view, Di can be seen as a preconditioner for L
since r and r are denned by

r=L-1Di$ and f = (D*L)-lL*$. (30)

• In one dimension there is a canonical choice for S and therefore for Di

such that, if the function $ and \? have a compact support, then r, f,
9, and 9 are also compactly supported. Indeed, in that case we have
necessarily cr(£) = a£s,a € (E. Therefore, the canonical choice for S
is S{£) = a(—i)s (l — e~^)s and DL is then a non-centered finite-
difference approximation of L of order 1. Indeed, it is well known (see
I. Daubechies [6]) that the quadrature mirror filters related to orthogonal
compactly supported functions <£ and * are: m0(£) = (l + eJ?) £(£)
and mi(^) = e^ (l - e~if)m£(£ + x) where £ is a finite trigonometric
polynomial. Then we get

P„(0 = 2'(l + e*)m"'£(0.

Pi(0 = 2'äi'c'f (1 - e-!'?)m (1 - c'f)' £(£ + TT),
P0(O = 2"' (1 + e'T (1 + e-'€)'£(Q,
£(£) = 2-'a-1iV« (1 - e-"f)m SC(Z + it) ,

which proves that PQ, PI, PQ and Pi are also finite trigonometric polyno-
mials. Then, using the following lemma borrowed from G. Deslauriers and
S. Dubuc [7] we deduce that the functions f, r, 9, and 9 have compact
support.

Lemma II.2
If nO = En^TnC-"« wUhEnL^yn = 1, then nr=1 r(2">0 15 an
entire function of exponential type. In particular, it is the Fourier trans-
form of a distribution ivith support in [Ni, N2].

Clearly, this canonical form is no longer available if the space dimension

is larger than 1 since the multidimensional quadrature mirror filters can
not be factorized as above.

II.3 The case of inhomogeneous elliptic operator

Here the non homogeneous property of the operator is obviously not adapted
to the scale invariance property of the multiresolution analysis. We will see

11
however, that introducing at each level a new scaling function, an embedded
family of spaces can be constructed which preserves the mathematical properties
relevant for numerical applications.
The natural candidates for T\ are a the functions L~l<b\. They are well
defined in L2(1H?) but suffer now from a lack of localization when j increases.
Indeed, we have
lim || 2-jn'2L-1 [<pjk] (x) - G(x - 2~jk) || = 0
j—*+oo
*2-J—'to

where G is the Greens function of the operator L defined as G(£) = l/c(£). For
example, when L = 1- A, G(£) = 1/(1 +£2) and G(x) = e~W. G decreases fast
but mathematical and numerical difficulties come from the fact that the family
of functions {G(x — &2_J), k £ Z2} is not a set of functions rescaled with j (in
other words, this family is not obtained by rescaling and translation of a single
initial function). This implies that the control of the localization by the index
j is lost. It follows that the family {L-1$A, k £ 2Z2} is not a good basis to
reconstruct our solution.
Let us show now that, however, a process very close to the one used in
the homogeneous case will provide an efficient algorithm for the summation of
formula (21).
We mimic the construction performed in the homogeneous case. Let (Vj)
be an r-regular multiresolution analysis, let L be an elliptic operator of order s
with constant coefficients and <x its symbol (we now suppose <r(£) > <r0 > 0 V£).
Let us also define the homogeneous polynomial of order s, &, as the principal
part of a, and let S(£) be a C°° 2?r periodic function with S(£) ~ £", where
n will be fixed later.
Then, Vj G %, we define a difference operator Dj by its symbol 5(£/2J).
Following the previous section, we define Vj, k £ 2Z
rik = 2>lL-1Dj$jk (31)
and

ejk = v°L-l{yik). (32)

By construction, and thanks to the fact that L is a constant coefficient
operator we have Tjk(x) = TJ(X - ■!-) and Ojk(x) = 8j(x - ■—■) where

%(0 = 2^-VS(U2j)^{p-, (33)

and ^
5 1
f§ (0 = 2^ - )^|p. (34)

12
Remark
• The functions 7j mimic the function f defined in (28). Unfortunately, it is
not possible to define the equivalent of r (28) since DjlL$jk g L2. Note
however that, by chance, (15) involves directly the 6\ functions.

Then, with P0 and P, defined in (23) and 24 we get the following scaling
and detail relations:

*5(0 = 2P0«/2>-+1)5V+1(O

Remark:
• An important point is that the filters Ps are independent of the scale
index j as it is originally the case for standard multiresolution analysis.
Furthermore , since they are defined by (23) and 24, the filters Pc are
directly related to the homogeneous operator L of symbol & if n = s.
This point is essential since it means that if Do is consistent to L, then
the tree algorithms related to the multiresolution spaces (Uj) and used to
sum up (21) are stable even if the functions TJ(X) are not standard scaling
functions. Indeed, S(£)/cr(£) —^—o 0 and therefore fj(0) = 0.
In other words, even if the functions 73 j, (a:) are used as scaling functions on
the range of L, they have zero moments as wavelets have.
Finally, we can prove the following theorem:

Theorem II.3 For 0 < n < s, s < r and s < m, the functions Tjk defined
by (33) and (31) satisfy

\d^Tjk{x)\ < c; 2j^l-5¥ (1 + 2>|* - 2-^|p+s-m"3"171 . (36)

Ifn — s and if Ps(£), £ € E are Ca, then stable tree algorithms are available

J-1
30 < C < C < +00 such that if f — 52A6A ^A#A, then

Cj2\drf <H/II2< C'El^l2- (37)

A A

Proof:

• Since S € C°° the we can apply Theorem II. 1 that proves the localization
inequality (36).

13
 The relations (35) defines directly the tree algorithms required to compute
the coefficients [cjk] such that

and since, when n = s, the involved quadrature mirror filters could be

related to the homogeneous operator L, the stability of the algorithm is
equivalent to the stability of the transform {cjt} >-*• / for / G 0i=o %i-
This transform is stable if and only if the family {r^, k G %2} is a Riesz
basis of 0^ZO Xe.
We have
2

/ / j c
jk T
jk dx
■/ kez2

-i [0,
Yl cik
j
e -i2- (k.£) Hmf
2TT]
k£Z2
$(2~^+2;7r)
a(e+2Z7r) <%.

Since {$jk} is a Riesz basis, since 5(0 is bounded, and since a is bounded
from below, then || / ||2< C"£ \cjk\2, which is the second part of the
inequality (37).
To prove the first part of (37) we use again the fact that the filters PQ and
Ps are related to L.
Indeed, if we define Q\ replacing a by & in (29) and if we define / as
tnen tne trans
/ = Y^x ^A f°rm / *-* {d\} is stable. Moreover thanks to
theorem II. 1, the operator /•-»■/ which can be also defined as *A •-*■ 9\
is also bounded (Y. Meyer [13]). Therefore the operator / ^ {d\} is
bounded that is the first part of (37) and that completes the proof.

Ill Approximation and Numerical Resolution

of Elliptic Problem on the Torus
This section is devoted to the approximation of elliptic problems on a sequence
of embedded Galerkin spaces associated with a multiresolution analysis and to
the corresponding numerical algorithms.

14
 Classically, we will consider the problem with periodic boundary conditions
to avoid the difficulties of general boundary conditions. We will use a r-regular
multiresolution analysis of the torus T[0,ip = {R//Z)2 constructed using a clas-
sical periodization technique (Y. Meyer [12]). We take as granted that, with
minor modifications, the results proved on the whole line can be transposed to
the torus. In that section, homogeneous and inhomogeneous operators will be
treated similarly.

III.l General formulation

The general formulation of the problem is

Find u £ Tjo^p such that

Lu = / (38)
V>) {
with f £ 2]o,i]2 and L a constant coefficient elliptic operator of
order s.

Standard variational approximation (P.A. Raviart and J.M. Thomas [14])

leads us to look for the solution of a weak problem in so called Galerkin approx-
imation spaces V£, where e is a scale related to Ve with Ve —*• Tjo.ip.
A natural choice for VE is Vs = Vf §, where V? belongs to a multiresolution
analysis of T[0)ip of the type described above. Indeed, we then have the following
inequality guaranteed if the involved multiresolution analysis is r-regular,
-JS
V s < r, 3 c> 0 V/ G H' \\f - TLv?fh < C 2 H* (39)

where Uvv, j < 0, stands for the orthogonal projection on Vf.

i
Then a standard Galerkin approximation writes
Find up £ Vpv C 7]o,ip suc
h that

(V) UyvLUypUp — ILyvf- (40)

where II^-p, j < 0, stands for the extension operator from V? to Tfo^p.
This approach leads us to replace L by the approximation Iiv-pLUvv. The
corresponding numerical algorithms are reduced to linear system solvers once a
basis of Vp has been chosen (P.A. Raviart et al. [14]).
§The symbol .v stands for the periodization operator on [0, l]2. We recall that dim Vj
2]
2 = 1/3 dimWj and that Vf = span{*oo = !}•

15
 Numerically, an optimal choice for the expansion basis is the wavelet basis
of Vp: {$o0, $^,Ae A^-1}^, since the corresponding stiffness matrix, is sparse
and can be easily (i.e using diagonal matrices) uniformally preconditioned (S.
Jaffard [8]).

III. 2 A different approximation

The main purpose of this paper is to define a different approximation of L and
the corresponding numerical algorithm. Taking the set of functions {9%, A G
Ag-1} (denned in (31)) and defining U% = span-fL"1^, 9%, A G Ap}, we get
an approximation of u as

up = £ 2-i'(f,nM + (f,*Zo)L-1$Z0 ■ (41)

A6A£
Indeed,
up = L-lKvvf = P~vL-lj (42)

where P~v is the projection on Up orthogonal to Up .

This formulation of up shows that the convergence of up towards u when
p —>■ co is straightforward since the set Up is a family of Galerkin spaces for the
suitable space of definition of u.
Moreover, the stability of the algorithm is a direct consequence from the
classical preconditioning properties of wavelet base expansions (S. Jaffard [8]).

III. 3 General scheme

The numerical algorithm derived from the previous section is now presented
in its collocation version. We call Ij the set of points (Z f][0,2j[)2. Then,
Jj = {2~ik, k £ Ij} is the two dimensional regular grid of scale 2--7 related to
[0,1[2.
For the numerical implementation, we assume that the space Vp is such that
any continuous function f G Vp is unambiguously defined by its values on the set
of points Jp. This assumption (satisfied by the even order spline multiresolution
we will use in the numerical tests) allows us to define the collocation projection,
Cvv, from the set of real sequences (fk)keip to Vpv as

V(fkheif,Cv?{(fk)kei,) = f &fev; and f{2~*k) = /*, V* G Ip.

As soon as we define / G Vp by its coordinates on the basis 5^, the opera-
tor C\-v appears as a discrete convolution operator involving a so-called inter-
polant filter /$ (£). The operator Cvj> is also a convolution operator called the

A Again, A^ - UjinAj with, in the periodized framework, Aj - {\ - 2~l (fc + |) , k €

2Zf}[0,V[)2,eeE*}.

16
point value operator and involves the point value filter PV$p(f) defined from
i$po(Jp)}- Obviously, i$p(£) and PV<s,p(£) are inverse. Let us remark however,
that the point value operator always exists as soon as the functions $pk are
continuous.

For the implementation, we therefore replace TLVV by Cyv in (42) and define
therefore uP as
Up = L~ Cyv f ■
Given the point values of / on the grid points Jp , the algorithm provides the
values of up on the same grid. More precisely, the algorithm can be presented
as follows:

1. The input of the procedure is the set of values (f(JP)) from which the
interpolant function fp G Vpv is constructed using !*„(£):

fp=J2 CPk®P*

2. fp is then decomposed into the wavelet subspaces Wf, 0 < j < p - 1 and
V0V as
f = £ (f,n)*x+coo

where CQO = Evv(f).

3. Up then becomes

up = Y, 2-^(/,^ + c'oo

where 6% = (2jsL~1ip\) . Here, c'o0 = COO/<T-(0) for non-homogeneous

operators. For homogeneous operator c'00 should be given. Note that in
that case cr(0) = 0 and / should have at least s vanishing moments; the
fact that c'00 should be given corresponds to the ill posed property of the
initial problem in L2.
4. up is then expanded in terms of the set of functions {T^, k G Ip} using
the tree algorithms related to 9% and TTJ. as

U
P = 2^i cPkTpk •
k£lp

5. Finally, the grid point values of up on Jp are estimated using the point
value filter PV~ (£).

17
 It appears clearly that various precalculations should be performed. In the
first step, the interpolant filter related to <£Pfc must be known; for the second
step, the orthogonal multiresolution analysis quadrature mirror filters Me should
be used and, for the fourthjstep the corresponding biorthogonal multiresolution
quadrature mirror filters Pt are required; finally, the point value filter related
to fp is used for the last step.
To be more precise, we have to make some remarks that help to reduce the
complexity and storage. For steps one and two, tensorial properties can be used
in a very classical way to reduce the 2D-algorithms involved in 2(dim (V^))1/2 x
ID-algorithms. It is then enough to know the one-dimensional interpolant filter
related to (p*k and the one dimensional quadrature mirror filters m^, £ € E.
For steps four and five, where the filters Pt, and the point value filter PV~ are
involved, we can note apply this simplification and we have a full 2D-problem.
We are now able to summarize all these precalculations in the following step
0:
0. The computation of the following filters is performed (this is presented for
the spline multiresolution analysis case):
-Interpolant filter related to <ppk, Itfip: analytical formulas in one di-
mension are available in V. Perrier and C. Basdevant [15].
-Orthogonal ID multiresolution analysis filters me: analytical formu-
las are also available.
-Filters Pe of the Biorthogonal Multi resolution Analysis: These filters
are constructed from mt and formula (23) and (24). In fact only -P(o,i)
and P{\Xj have to be computed since we have P(i,o)(£i,£2) = -P(o,i)(^2,<^i)-
-Point value filter, PV—Tpk
related to r^\ This filter is computed from
formula (28) and the analytical expression of 5, a and $(£)■ We have
successively

T£(X) = 1/(2T) J2 rpMe2*'™-* , x e M2 ,

w<=Z2

r;{xn) = 1/(2*) £ ( J2 %(w + 2Fr)j e2i

™'Xn - *n€JP. (43)

Practically, 43 is truncated according to a prescribed precision. This is

possible because rp(£) decreases fast.
Remarks:
• One should again emphasize that the entire algorithm is based on con-
volution operators. Thanks to the periodic boundary conditions, the
convolutions can either be performed directly or using a discrete Fourier

18
 transform. The implementation presented in this paper uses the Fourier
transforms since it is optimal for non compactly supported filters on non
adapted spaces of approximation.

III.4 Detailed Algorithm

This section is devoted to the structure of the code. Basic tools, such as Fast
Fourier Transforms, Convolution/Decimation algorithms, or Term by term mul-
tiplications are not described.
As can be seen from the general scheme presented above, the main code
involves only two more elaborate routines that will be called the Precalculus
routine (step 0), and the tree algorithm routines. The tree algorithm routines
may or may not use the tensorial structure. They will be called consequently 2D
Tensorial Tree Algorithm-D (steps 2) and 2D Non Tensorial Tree Algorithm-I
(step 4) where -D and -I stand for direct and inverse. We recall that the steps
1 and 5 are convolutions and the renormalization performed in step 3 is term
by term multiplication.
The tree algorithm routines are becoming very classical and therefore we
will not describe them either. Note however, that since only convolutions are
performed in our algorithm, we only use the discrete Fourier transform of the
wavelet coefficients (and not the corresponding values) at every scale, that re-
duces significantly the complexity.
We now give the detailed description of the main program (ELLIP ) and of
the precalculus program (PRECAL) in pseudo code.
The following example sketches the structure our programs.

[OUTPUTS]=Program(INPUTS)
# Comments
# Body of Program:
■C
[0UTPUTS1]= Subprograml(INPUTS)
INPUTS2 = OUTPUTS1
[OUTPUTS] = Subprogram2(INPUTS2)
>

Our variable descriptors bears some resemblance to the C language as well

as to the MATLAB conventions.

III.4.1 Preliminary computations

The symbols ', *, .*, and ./ used to present this program are borrowed from
MATLAB and mean respectively, the transposition, the matrix product, the
term by term product, and the term by term division. We also use the following
n sub sampling operator a : n : b defined as: If a is a 2D array of size 2P x 2P,

19
b(l : 2n : 2?, 1 : 2n : 2P) is a new array of size 2P~" x 2p~n given by a(i,j) =
a(2ni,2Nj),(i,j)eIp.
Program PRECAL

[QMFBIW.TAUTW] = PRECAL(p,pmax,QMFW,PHIW,SW,CW,SIGMA)
#INPUT:
#p -> index of the approximation space Vp in which the
# elliptic problem is solve.
#pmax -> index of the approximation space Vpmax in which the
# precomputation of TAUTILDE is done (it depends on the
# prescribed precision).
#QMFW -> structure containing the quadrature mirror filters in
# one dimension:
# QMFW.mO -> ID array containing the quadrature mirror filter
# coefficients associated to the scaling functions;
# size(QMFW.mO)~>2~p; QHFW.mO(i) = m0(i/2~p),
# i belong to {0,...,2~p-l}.
# QMFW.ml -> ID array containing the quadrature mirror filter
# coefficients associeted to the wavelet;
# size(QMFW.ml)->2~p;
# QMFW.mi(i) = ml(i/2~p),i belong to {0,...,2~p-l>.
#PHIW -> ID array; size(PHIW)->2~pmax; where pmax is given
# and pmax>p; PHIW(i) = the value of the Fourier transform
# of the ID scaling function at the point i, i belong to
# {0,...,2"pmax-l}. Used to compute the value of tautilde on
# the finer grid.
#SW -> 2D array containing the sampling of the function
# S used for biorthogonal filters;
# Size(SW)->(2"pmax X 2~pmax); SW(i,j)=S(i/2*pmax,j/2"pmax),
# (i,j) belong to {0,...,2~pmax-l}~2.
#
#CW -> 2D array containing the sampling of the function
# S(2w)/(2"s S(w)) Size(CW) -> (2"p X 2'p); CW(i,j) =
# S(2i/2"p,2j/2-p)/ S(i/2-p,j/2-p),
# (i,j) belong to {0,...,2"p-l}"2.
#SIGMA -> 2D array containing the sampling of the symbol
# of the operator Size(SIGMA) -> (2"pmax X 2"pmax);
# SIGMA(i.j) = sigma(i/2~pmax,j/2"p),
# (i,j) belong to {0,...,2~pmax-l}~2.
#
#0UTPUT:
#QMFBIW -> structure containing the biorthogonal quadrature
# mirror filters related to the tautilde functions:
# QMFBIW.PTILDE0-> 2D tab containing the biorthogonal

20
# quadrature mirror filters associated
# to the scaling functions;
# size(QMFBIW.PTILDE0)->(2~p X 2"p):
# QMFBIW.PTILDEl-> 2D tab containing the biorthogonal
# filters associated to the first wav-
# elet;size(QMFBIW.PTILDEl)-> (2~p X 2"p).
# QMFBI¥.PTILDE2-> same as QMFW.PTILDE1 for the second
# wavelet (not computed QMFBIW.PTILDE2 =
# QMFW.PTILDE1 transposed).
# QMFBIW.PTILDE3-> same as QMFW.PTILDE1 for the third
# wavelet.
#

#
# Computation of the filters Ptilde
#
QMFBIW.PTILDEO = ((QMFW.mO)' * (QMFW.mO)) .* CW;
QMFBIW.PTILDEi = ((QMFW.mO)' * (QMFW.ml)) ./
(2~s SW(1:2"(p-pmax):2~pmax,1:2"(p-pmax):2~pmax);
QMFBIW.PTILDE3 = ((QMFW.ml)' * (QMFW.ml)) ./
(2~s SW(1:2"(p-pmax):2"pmax,1:2"(p-pmax):2"pmax);
#
# Computation of the point value filter related to TAUTILDE
#
TAUTW = ( (2"(ps) * PHIW' * PHIW) .* SS)./ SIGMA;
TAUTW = Periodize(TAUTW, p) ;

The subroutine Periodize is not described here, but it is a straight forward

transcription of 43.

III.4.2 Main Program

Main Program Ellip
[UX] = Ellip(FX,QMFW,FIW,QMFBIW,TAUTW)
#INPUT:
#FX -> 2D array containing the sampling of the function
# f; Size(FX) -> (2"p X 2*p); FX(i,j) = f(i/2'p,j/2~p),
# (i,j) belong to {0,...,2~p-i}"2.
#FIW -> ID tab of data containing the interpolation filter
# related to PHI_pO, size(FIW)->2~p;
#QMFW -> see PRECAL
#QMFBIW -> see PRECAL
#TAUTW -> see PRECAL
#

21
#
#OUTPUT:
#UX -> 2D array containing the sampling of the approxi-
# -mation u_p; size(UX)->(2~p X 2~p), UX(i,j) =
# u(i/2-p,j/2-p), (i,j) belong to {0,...,2-p-l}"2.
#
#TAUTW -> 2D array containing the values of the Fourier
# transform of the scaling function TAUTILDE at level p
#
»TEMPORARY DATA:
#FW -> 2D tab containing the fft of FX; Size(FW) ->
# (2*p X 2~p);
#CPW -> 2D tab containing the fft of scaling coefficient of
# FX; Size(CPW) -> (2*p X 2"p);
#DJW -> Structure of 2D array containing the Fourier transform
# of the wavelet coefficients; size(DJW) -> (2~p X 2~p);
#CTILDEPW
# -> same as CPW for UX;
#U¥ -> 2D tab containing the fft of UX;
#
■C
#
# step 0
#
[FW] = Fast Fourier Transform(FX)
#
# step 1
#
[CPW] = FW.*FIW (Term by term multiplication)
#
# step 2
#
[DJW] = 2D Tensorial Tree Algorithm_D (CPW.QMFW)
#
# step 3
#
[DJW] = DJW.*(2~js) (Term by term multiplication)
#
# step 4
#
[CTILDEPW] = 2D Non Tensorial Tree Algorithm_I (DJW.QMFBI)
#
# step 5
#

22
 [UW] = CTILDEPW.*TAUT¥ (Term by term multiplication)
#
[UX] = Inverse Fast Fourier Transform(UW)
}

III.5 Storage and Complexity Analysis

As the computation is clearly separated into precalculations and actual imple-
mentation of the algorithm, we will also present the storage and complexity
analyses separating the two parts. One should remember that the precalcu-
lation is done once and for all while, as it will be the case in section IV, the
algorithm can be applied iteratively.
We will not discuss the complexity related to one-dimensional computations
as well as the storage connected to one-dimensional arrays since both can be ne-
glected in our bidimensional implementation. All the evaluations are performed
forW = dimVrp = 2P x 2".
Storage
Permanant storage (precalculations): The structures QMFBIW and TAUTW rep-
resent four bidimensional arrays of size TV.
Temporary storage (actual algorithm): The storage related to bidimensional
arrays can be reduced to one arrays of size TV.
Finally, the total memory used corresponds to five arrays of size TV.
Complexity analysis
Precalculus: The computation of the four arrays in the structure QMFBIW is
done in C x TV.
The computation of PV~Tp0V is performed in C x TV operations. The value of
. .
C depends on the precision of the calculation.
Main program:
Fast Fourier Transform and Inverse Fast Fourier Transform in-
volve C x TV log(TV) operations.
The complexity of the Term by term multiplication is TV.

Tree algorithm-D and Tree algorithm-I are based on convolution and

decimation operators. These procedures involve therefore C x TV opera-
tions.
Therefore the total complexity is 0(TVlog(TV)).

In the following section we use these programs iteratively,to solve the 2D

Burgers equation after reducing it to a cascade of elliptic problems. We would

23
like to emphasize moreover, that our approach can be also used to solve equa-
tions involving homogeneous pseudo-differential operators. A characteristic ex-
ample is y/—Au = / with periodic boundary conditions on [0, l]2. We have
L = v^Ä and therefore <r(£) = v^i +£>• Tne most natural choice for S is
S(£) - 2-v/sin2(£i/2) + sin2(£2/2) and one easily checks that the hypotheses of
theorem II.2 are then satisfied. The algorithms presented previously can be
used (see Pj. Ponenti [16]).

IV Numerical Application: Resolution of the

2D Burgers equations
In this section, we will use the periodized Battle-Lemarie's multiresolution anal-
ysis of splines of order m (see P.G. Lemarie [9]). The existence of collocation
projectors related to the spline breakpoints requires splines of even order and
the value m = 8 will be used in the applications.
As described in J. Liandrat et al. [11], any parabolic equation of the type
§f + L0u + G{u) = 0

u(0,t) = u(l,t)
(44)
u = uo for i — 0

0 <t <Tmax,0 <x<l

where L0 is a differential operator of even order with positive symbol <r0(w), and
G is generally a nonlinear function of u and its derivatives, can be numerically-
approached using a classical finite difference time discretization scheme followed
by a variational approximation of the resulting elliptic problems. We show now
that the approach developed in the previous sections can be used efficiently to
provide this approximation.
Following J. Liandrat et al. [11] we first introduce a segmentation {tn}n=i
of [0, Tmax] (i.e. a sequence {tn}n=i such tnat 0 =t0 <ti < ... <tM = Tmax)
and now look for a sequence of functions of the x variable {u^Xx)}^ such
that u(n\x) is an approximation of u(x,tn).
With Atn = tn+1 -tn,0<n<M, and considering first (44) as an ordinary-
differential equation in time, a standard finite-difference discretization leads to
the following iterative equation:

£nU("+D = F(ü<n\ ..., «("-!')I A<„,...., Atn-i, G(u<n>),..., <?(«<»-''>)) (45)

where £„ is a step forward operator that together with F is determined by
the choice of the finite-difference approximation of the time-dependent ordinary

24
differential equation. We always assume that this approximation is at least semi
implicit for the linear part LQ and explicit for the nonlinear part. Therefore
3a > 0 such that £„ = (! + aAtnLo) where I stands for the identity operator.
By hypothesis, <r0(w) > 0,Vw and then £n has always a symbol bounded from
below by 1.
Hence, assuming that {«(""'),/ = 0,...,i} and {G(u(n-,),/ = 0, ...j} are
known, the resolution of (45) falls under the scope of paragraph III.3 and the
resolution of (44) can be therefore performed iteratively.
The bidimensional Burgers equation writes, with u — («i, «2):

u(0,t) = u(l,t)
(46)
u= UQ for t = 0

0<t <Tmax,0 < x < 1.

Choosing a constant step segmentation of [0,Tmax] (i.e., a segmentation
such that 3At such that V0 < n < M,tn - nAt), an implicit Crank-Nicholson
time scheme (second order) for the linear term (i/Au), and an explicit second
order Adams-Bashforth scheme for the nonlinear term (Vu.u) we get

I _ V|A) «<n+1> = (i + ^Ay^-öt^Vu^M^-^u^.u^-V)

and the solution can be written as

„(n + l) = Ö(n + X) _ u(n)

with

«("+1) =(l- I/|A) _1 UuW - M^VuH^") - ivW("-1).«("-1))') (47)

To fall completely under the scope of paragraph III.3, one should be able
to evaluate the point values of the nonlinear term of (47). We used the sim-
plest method available that consists, as classically done in spectral methods
(C. Canuto et al. [3]), to "apply the nonlinear operator on the grid points".
More precisely, the approximation of Vu^.u^ we used is PV(Vu(l\u^) =
CZ\{Cvv x (Vu!).Cvv(u')) where x is a term by term multiplication of finite
p p
p
sequences and Cyv is the collocation projection introduced in section III.2.
Then, for each time step nAt, the problem clearly belongs to the class of
elliptic problems studied in section (III) with L = I — v^-A and / = 2u(n) —
8t(%Vu(n\v,(n^ — !Vu(n-1).i((n-1)). The iterative form of the equation induces
some modifications of the general scheme presented in III.3 and we therefore
provide the full scheme for an iteration of the Burgers approximation scheme:

25
 0. The inputs are the values of («<">( J„)), (^(JP)) and (*§£(JP)).

1. F(n\jp) = 2u(-n\jp) - <5tJPV(Vu(n).uw)(Jp) is computed as described

above.
2. Fp"(ar), the function of Vpp interpolating the values Fn(Jp) is constructed

3. F" is then decomposed into the wavelet subspaces Wf, 0 < j < p- 1 and
V0V as

A6AJ-1
where c0o = Uy^i^p)-

4. ul"+1 then becomes

u(n+i) = £ Ti°(F?,Vx)K + coo

AeAp1

where ^ = (2''L-VA)"-

5. t4n+1) is tnen
expanded in terms of the set of functions {T^, k € Ip} as

4n+1) = E c^ •
fc€/p

We also get &fi(n+1) = Efc6/P <**&*£

and^+D = Zkei^pkffik-

6. From the point values of rp{Jp), ^?/(Jp), and j^r^Jp) we compute

(ü<«+1)(Jp)) and its gradient using the corresponding point value filters.

7. Finallv using the values of (un(Jp)) and of its gradient, we get the values
of (u("+1\jp)) = («(n+1)(Jp) - w(n)(./P)) and its gradient.

rV.l Storage and Complexity Analysis

As described above, the numerical code implements the elliptic solver in an it-
erative process. Since the time step At is constant, the characteristics of the
elliptic solver does not depend on the time index n. Then, the solver precalcu-
lations related to the whole parabolic problem are the same as the ones related
to the elementary elliptic solver (see section III.5). This applies to the storage
and to the complexity as well.

26
 As it has been shown in the previous section however, extra work, not con-
nected to the elliptic solver itself but to the computation of the right hand side
term of the iterative equation (47) is required. This extra work is related to the
storage of the fields at the different time steps involved in the three level time
step Adams-Bashforth Crank-Nicholson scheme and to the point value evalu-
ation of the derivatives involved in the nonlinear part. Again, it can be split
into permanent and temporary storage as well as in precalculation and main
program extra work.
Storage (in addition to the elliptic solver storage)
Permanant storage (precalculations): One extra structure containing the
point values -§^f^{Jv) must be stored in one bidimensional array of size N; the
structure containing the point values J^T^(JP) is given by transposition of the
previous one.
Temporary storage (actual algorithm): The two fields u(n\ u("_1) and «(n+1)
can be handled using three arrays of size N.
Complexity analysis (in addition to the elliptic solver complexity)
Precalculation: The computation of PV a ~r v is performed in C x N opera-
s' p
tions where, as in section III.5 the value of C depends on the precision of the
calculation.
Main program: The addition of complexity is related to steps (1), (5), (6)
and (7). Since these steps involve convolutions and term by term products, the
added complexity is again CN log N
Finally, the total memory used is 7 arrays of size N.
The total complexity is 0(Nlog(N)).
Obviously, the total complexity of the whole resolution is M times the com-
plexity of one time step resolution.

IV.2 Numerical Results

Test case on an xi translation invariant problem
The validation of our code has been performed on an X2 translation invari-
ant problem constructed using for the initial condition (u0l,u02)(xi,X2) =
(sin(27rxi),0). Indeed, with such an initial condition, the solution remains x2
translation invariant.
For an easy comparison to the well documented paper of C. Basdevant et al.
[1] we used v = 10~2/7r.
As explained in C. Basdevant et al. [1], the pertinent quantities are
Ölt Ott
ms = SUpt\\-z 0M)lloo = SUPte[o,Tmax]\a (°-M)l
oxi ax\

27
and tms defined such that
du
|^—(0.5,fms)| = ms.
ox i
Table 48 exhibits the numerical results obtain using various values for the
time step At. The maximum time step numerically acceptable was At = 0.0075.
In each case, the values of ms are computed by interpolation and the correspond-
ing values of tms are deduced. The comparison with the expected theoretical
values (first column) shows that our method competes favorably with the ma-
jority of the schemes presented in C. Basdevant et al. [1]. A complete study of
the time step size dependence of the results connected to the stability analysis
of the parabolic algorithm will be published later.

Exact At 0.0005 0.001 0.0025 0.005

-304.0103 ms -304.6308 -305.727 -309.4354 -316.5454 (48)
0.255237 tms 0.253 0.252 0.25 0.245

Test case on a first diagonal translation invariant problem

Our second test case is performed on a first diagonal translation invariant prob-
lem constructed using (u0l,uo3)(xi,x2) = (sin(27r(;ri + x2),sin(27r(a;i + x2)).
Again, the solution can be compared to the reference solution of C. Basdevant
et al. [1] thanks to a 45° rotation and to a time dilation of factor 2. However,
according to our reference axes, it is obviously a fully bidimensional solution.
Figures 1,2 and 3 show the isoline values of the numerical approximations
computed with 6t = .001 at t = 0, t = 0.15 and t = 0.50. The first diagonal
translation invariance is kept and we obtain the values ms = 249.0528 and tms =
0.123. The expected theoretical values are -304.0103 for ms and 0.1276185 for
tms- This is not as good as before but one should note that the resolution in
the direction perpendicular to the front axis is now half the one in our previous
calculations.
Since the ultimate application of all this work is the development of adaptive
algorithms (i.e. the development of algorithms handling approximation spaces
of reduced dimension adapted to the solution regularity (see for instance Pj.
Ponenti [16]),we have estimated, at various times, the wavelet basis adapted
to the approximation and denned as the lowest cardinal m = 8 spline wavelet
basis preserving the I? norm of the approximation with a precision of 10-6.
The columns of table (49) show for each scale 0 < j < 7 the number of wavelets
selected in the adapted basis related to the approximated solution at various
times. It appears that, compared to the full basis of Vs (last column), these
bases have a drastically reduced cardinal (we defined the rate of compression
as cardinal rtjhe^ adapted basis) even if the greLCiiexxts of the solution fill up a large
domain made of two complete lines of the plane (see figures 1, 2 and 3)

28
 Figure 1: Initial condition, U = 0.

Figure 2: Approximated solution, t, = 0.15.

29
 Figure 3: Approximated solution, t{ = 0.50.

Acknowledgments
Authors are very grateful to Philippe Tchamitchian for invaluable assistance
and helpful discussions on the subject of this work.

Conclusion
In this paper, we have proposed an inversion scheme for elliptic problems
based on biorthogonal wavelets. The approximation of elliptic problem solutions
is constructed and leads to stable and fast numerical algorithms.
Numerical tests related to the approximation of the parabolic Burgers equa-
tions transformed into a cascade of elliptic problems are provided.
The approximation scheme is based on convolution operators and can there-
fore be theoretically used in the framework of adapted spaces of approximation.
As mentioned however, the nice tensorial product structure that enforces nu-
merical efficiency is fragile and is generally destroyed when applying the scheme
directly. Other approximations for the step forward operator, should allow
one to use efficiently this approximation in a general context of adapted multi-
dimension spaces of approximation.

30
"o fcT
^
Ö
H CO
v -£
-O JÜ CO CM CO
CM 00 lO co
S 5 CM 00 t- CM T-H
M CO O) (O O M Ol
LO
LO
a ja COHlfr-lhnH'^ CO

O
CD
-*
Tf
to ■* CM -* CO
CO W N CD
II CD tO O CD CO ■* ■^

co
CD t— co o o LO o
-M OCO^Hf-COHH co -*
o ■*
o oo
CO T# T# co
00 CO ^f t~ T-H
II CO CO O CM Ol T-H

CM
O t~- CO O r-H CO Tt*
-» OCO^HNWHH CO LO
^
OS
CD Tf
CD
LO LO
CM CO o CO CM
00 ^tf o l—1 T-H
II CO CO O CO CM co
O N M O H Ol -* CO
•« OOOrfHNCOHH CO LO

O LO
CD t^
CM "* CO CD co
oo o co ■*
CO
II COCOOOIO) co
LO -*
o N n o H o>
-t-i OCO^HNCOHH CO LO

CD oo
LO CO
i—i ■5t< O CM CO
00 O CO LO
II CO ffl O CO ■* CO ■"?
O t— CO O i—I CD co LO
t-T OCO^HNCOHN co LO

o oo
o
T-H oo ■* o LO
COO^f CM i>;
II CO CO O O OS OS
co 1—I
o N coon co
■M
OCOTf<T-Ht^COr^T-H CO LO

CD o
LO CD
o CD o
II CD CD -^
O N » W
CD
CM
CD
HO OCC* r-< t- CM O O CO LO

CD
o T-H
CO
II CM CD
o
-T-o OCCCOOOOOO T}< CD

cu '■*-*

Si o
*» CO
HH CO r-j
O n3 o
Xi CO
13 -o CO
cp
a cu
OHCNCOTflOCON ■d a. CO g-
Car
ad a

II II II II II II II II
•^-> ■?-> '^ *^-s "<-> '^ *^-s '^ * 8

31
References
[1] C. Basdevant, M. Deville, P. Haldenwang, J.M. Lacroix, J. Ouazzani, and
R. Peyret. Spectral and finite difference solution of the Burgers equation.
Computers and Fluids, 14, No. 1, pp. 23-41, 1986.
[2] G. Beylkin. On wavelet based algorithms for solving differential equations.
Wavelets, Mathematics and Applications, pages 449-466, 1995.

[3] C. Canuto, M.Y. Hussaini, A. Quarteroni, and T.A. Zang. Spectral Methods
in Fluid Dynamics. Springer Series in Computational Physics, 1988.
[4] A. Cohen, I. Daubechies, and J.-C. Feauveau. Biorthogonal bases of com-
pactly supported wavelets. CPAM, 45(5):485-560, Jun. 1992.
[5] W. Dahmen and A. Michelli. Dual wavelet expansions for general scalings.
personal communication.
[6] I. Daubechies. Ten lectures on wavelets. SIAM, Philadelphia, 1992.

[7] G. Deslauriers and S. Dubuc. Interpolation dyadique. In Fractals, dimen-

sions non entieres et applications, pages 44-45. Masson, Paris, 1987.
[8] S. Jaffard. Wavelet methods for fast resolution of elliptic problems. SIAM
Journal on Numerical Analysis, 29(4):965-986, 1992.
[9] P.G. Lemarie. Ondelettes ä localisation exponentielles. Journ.de Math.
Pures et Appl, 1989.
[10] J. Liandrat and P. Tchamitchian. Resolution of the ID regularized Burgers
equation using a spatial wavelet approximation: Algorithms and numerical
results. ICASE report 90-83, 1990.
[11] J. Liandrat and P. Tchamitchian. Elliptic operators, adaptivity and
wavelets, preprint IMST, 1993.
[12] Y. Meyer. Ondelettes et Operateurs I: Ondelettes. Hermann, Paris, 1990.
[13] Y. Meyer. Ondelettes et Operateurs II: Operateurs de Calderön-Zygmund.
Hermann, Paris, 1990.
[14] J. M. Thomas P.A. Raviart. Introduction ä l'analyse numerique des
equations aux derivees partielles. Masson, 1983.
[15] V. Perrier and C. Basdevant. La decomposition en ondelettes periodiques,
un outil pour l'analyse de champs inhomogenes, theorie et algorithmes. la
Recherche Aerospatiale, 3:57-67, 1989.

32
[16] Pj. Ponenti. Algorithmes en Ondelettes pour la Resolution d'Equations aux
Derivees Partielles. PhD thesis, Universite d'Aix-Marseille I, Centre de
Physique Theorique de Marseille, October 1994.
[17] P. Tchamitchian. Inversion de certains Operateurs elliptiques a, coefficients
variables. Preprint LAPT, 1993.

33
 Form Approved
REPORT DOCUMENTATION PAGE OMB No. 0704-018S
Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources,
gathering and maintaining the data needed, and completingand reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this
collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operationsand Reports, 1215 Jefferson
Davis Highway, Suite 1204, Arlington, VA 22202-4302, and to the Office of Management and Budget, Paperwork Reduction Project (0704-0188), Washington, DC 20503.
1. AGENCY USE ONLYfieai/e blank) 2. REPORT DATE 3. REPORT TYPE AND DATES COVERED
February 1996 Contractor Report
4. TITLE AND SUBTITLE 5. FUNDING NUMBERS
NUMERICAL ALGORITHMS BASED ON BIORTHOGONAL WAVELETS
C NAS1-19480
WU 505-90-52-01
6. AUTHOR(S)
Pj. Ponenti
J. Liandrat

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION

Institute for Computer Applications in Science and Engineering REPORT NUMBER

Mail Stop 132C, NASA Langley Research Center

ICASE Report No. 96-13
Hampton, VA 23681-0001

9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSORING/MONITORING

National Aeronautics and Space Administration AGENCY REPORT NUMBER

Langley Research Center NASA CR-198290

Hampton, VA 23681-0001 ICASE Report No. 96-13

11. SUPPLEMENTARY NOTES

Langley Technical Monitor: Dennis M. Bushneil
Final Report
Submitted to Numerische Mathematik.
12a. DISTRIBUTION/AVAILABILITY STATEMENT 12b. DISTRIBUTION CODE

Unclassified-Unlimited

Subject Category 64

13. ABSTRACT (Maximum 200 words)

Wavelet bases are used to generate spaces of approximation for the resolution of bidimensional elliptic and parabolic
problems. Under some specific hypotheses relating the properties of the wavelets to the order of the involved op-
erators, it is shown that an approximate solution can be built. This approximation is then stable and converges
towards the exact solution. It is designed such that fast algorithms involving biorthogonal multi resolution analyses
can be used to resolve the corresponding numerical problems.

Detailed algorithms are provided as well as the results of numerical tests on partial differential equations defined on
the bidimensional torus.

14. SUBJECT TERMS 15. NUMBER OF PAGES

wavelets; biorthogonal multiresolution analysis; tree algorithms; elliptic and parabolic 35
numerical problems; two-dimensional Burgers equations 16. PRICE CODE
A03
17. SECURITY CLASSIFICATION 18. SECURITY CLASSIFICATION 19. SECURITY CLASSIFICATION 20. LIMITATION
OF REPORT OF THIS PAGE OF ABSTRACT OF ABSTRACT
Unclassified Unclassified
NSN 7540-01-280-5500 Standard Form 298(Rev. 2-89)
Prescribed by ANSI Std. Z39-18
298-102