I

ELEMENTS OF FUNCTIONAL
ANALYSIS

In this chapter we will recall some fundamental results of functional anal-

ysis. We firstly present the notion of a Hilbert space and discuss some
basic properties of the orthogonal projection operator. We then introduce
the concepts of closeness and completeness of a system of elements which
belong to a Hilbert space. The completeness of the system of elementary
sources is a necessary condition for the solution of scattering problems in
the framework of the discrete sources method. After this discussion, we will
briefly present the notions of Schauder and Riesz bases. We will use these
concepts when we will analyze the convergence of the null-field method.
We then consider projection methods for the operator equation

Au = / ,

where A is a linear bounded and bounded invertible operator from a Hilbert

space H onto itself. We will consider the equivalent variational problem

B{u, x) = J^*{x) for all x e H,

where B is a bounded and strictly coercive sesquilinear form and J" is a

linear and continuous functional. Convergent projection schemes will be
2 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

constructed by appealing on the fundamental theorem of discrete approx-

imation. Later, we will particularize these results for the space of square
integrable tangential vector functions. We conclude this chapter by analyz-
ing projection methods for a linear operator A acting from a Hilbert space
H onto a Hilbert space G, and for the operator equation

where JB is a compact operator.

1 HILBERT SPACES. ORTHOGONAL PROJECTION OPERATOR

Let if be a complex vector space (linear space). The function.(, )j^ :
H X H -^ C is called a Hermitian form if

(a) {au 4- /3v, w)j^ = a {u,w) fj -\- f3 {v, w)jj , (linearity)

(b) (w, v)fj = {v, u)]^ , (symmetry)

for all u,v,w e H and all a,/3 E C. Here, a* denotes the complex conju-
gate of a. A Hermitian form with the properties

(a) (w,ix)^>0, (positivity)

(b) (u, u)ff=Oii and only ii u = 9HI (definiteness)

where ^H stands for the zero element of H, is called a scalar product or an
inner product. The vector space with a scalar product specified is called a
inner product space or a pre-Hilbert space.
In terms of the scalar product in /if, a norm

\HH = \fM~H (1-1)

can be introduced, after which H becomes a normed space. The follow-

ing important inequality is the basis for the statement that inner product
spaces contain all the elements of Euclidean geometry, while normed spaces
have length, but nothing corresponding to angle. It is the Cauchy-Schwarz
inequality and is given by

|(u,i;)^| < \\u\\tj \\V\\H for all u.veH. (L2)

The Cauchy-Schwarz inequality and the definition of scalar product imply

that the norm properties:
 1. HILBERT SPACES. ORTHOGONAL PROJECTION OPERATOR 3

(a) \\u\\fj > 0, (positivity)

(t>) ll^ll// = 0 if and only \i u = 0//, (definiteness)

(c) ||aix||^ = \a\ \\u\\jj , (homogeneity)

(d) ||u + v\\f^ < \\u\\ff -f ||i;||^ , (triangle inequality)

for all u,v e H and all a € C are satisfied. Therefore any scalar prod-
uct induces a norm, but in general, a norm || ||^ is generated by a scalar
product if and only if the parallelogram identity

ll« + v\\l + \\u - v\f„ = 2 {\\u\\l + Ml) (1.3)

holds.
Given a sequence {un) of elements of a normed space X, we say that
Un converges to an element u of H if \\un — u\\^ -+ 0 as n —• oo. A se-
quence (un) of elements in a normed space X is called a Cauchy sequence
if \\un — UmWx —* 0 as n , m ^ oc.
A subset M of a normed space X is called complete if every Cauchy
sequence of elements in M converges to an element in M. A normed space
is called a Banach space if it is complete. An inner product space is called
a Hilbert space if it is complete.
A sequence (un) in a Hilbert space H converges weakly to u £ H if for
any v E H, {un,v)fj —> (u,i;)^ as n —>> oo. Ordinary (norm) convergence is
often called strong convergence, to distinguish it from weak convergence.
The terms 'strong' and 'weak' convergence are justified by the fact that
strong convergence implies weak convergence, and, in general, the converse
implication does not hold. If a sequence is contained in a compact set,
then weak convergence implies strong convergence. Note that every weakly
convergent sequence in a Hilbert space is bounded and every bounded
sequence in a Hilbert space has a weakly convergent subsequence.
Two elements u and v of an inner product space H are called orthogonal
if {u,v)fj = 0; we then write u±v. If an element u is orthogonal to each
element of a set M, we call it orthogonal to the set M and write u±M.
Similarly, if each element of a set M is orthogonal to each element of the set ,
K, we call these sets orthogonal, and write M±K. The Pytagora theorem
states that

\\u±v\\l^\\u\\l + \\v\\l (1.4)

for any orthogonal elements u and v.

A set in a Hilbert space is called orthogonal if any two elements of the
set are orthogonal. If, moreover, the norm of any element is one, the set is
called orthonormal.
4 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

A subset M of a normed space is said to be closed if it contains all

its limit points. For any set M in a normed space, the closure of M is
the union of M with the set of all limit points of M, The closure of M is
written M. Obviously, M is contained in M, and M = M if M is closed.
Note the following properties of the closure:
(a) For any set M, M is closed.
(b) If_A/ C K, then M C F .
(c) M is the smallest closed set containing A/; that is, '\{ M <Z K and
K is closed, then JI C K.
Complete sets are closed and each closed subset of a complete set is
complete.
Next, we define the orthogonal projection operator. Let J/ be a Hilbert
space and M a subspace of H (i.e. a complete vector subspace of H).
Let u E H. Since for any v € M we have ||^/ — ^||// > 0, we see that
the set {\\u — vW^^ / t ' G A/} posses an infimum. Let d = mi^^^M ||^ ~" ^IIH
and let {vn) be a minimizing sequence, i.e. (f^,) C M and \\u - VUWH ~^
dasn —V oo. Since M is a vector subspace, ^{vn + v^n) G Af, whence
11 !L___!Ii|| > d^ Using this and the parallelogram identity
H

Vn 4-1;,,
,,||5, = 2(||7i-i;,||^, + | | u - i ; , n | | ^ ) - 4 (1-5)
H

gives

\\V„ - Vr,,\\l < 2 (||U - VnWl + II" " ^mll«) ' ^d^\ (1-6)

whence, by letting n, m —• oo, ||i;„ — I'mll// —* 0 follows. Thus, {vn) is

a Cauchy sequence and since M is complete, there exists w £ M such
that \\vn — 'w\\f^ —> 0 as n —> oo; moreover \\u — i^n||// —^ ||^ - ?^||// = rfas
n —• 00. Suppose now that there exists another element w' for which the
function \\u — u||^ attains its minimum; then d = \\u - if ||^ = ||w — vj'^n •
Clearly, \(w •\- w') E M and we have

d — inf \U - l^llrr < w+ w

H
(1.7)

w -\- w'
Thus, u - — d, and by the parallelogram identity
H

w ^-w'
\\W - w'W^j = 2 (||n - xot„ + \\u - w't^^) - 4 I L = 0, (1.8)
 1. HILBERT SPACES. ORTHOGONAL PROJECTION OPERATOR 5

we find w = w'.
The vector w gives the best approximation of u among all the vectors
of M. Note that d is called the distance from u to M and is also noted by
p(u, M). The operator P : H —^ M mapping u onto its best approximation,
i.e.

Pu=:us (1.9)

where ||t/ — t/;||j^ = d — miy^M 11^ "" ^11// ' ^^ ^ bounded Hnear operator
with the properties: P^ = P and {Pu.v)jj — {u,Pv)ff for any u,v € H.
It is called the orthogonal projection operator from H onto M, and w is
called the projection of u onto M.
The following statements characterizing the projection are equivalent:

(a) \\U-W\\H < \\u-v\\ff,

(b) Re{u- w,v - w)fj < 0,

(c) Re{u-v,w- v)fj > 0,

toT ue H, w ^ Pu e M and any v € M,
Let M be a subset of a Hilbert space //. The set of all elements or-
thogonal to M is called the orthogonal complement of M,
M^ = {ueH/u±M}.

Clearly, M-^ is a subspace of H, To show this we firstly observe that A/-^

is a vector subspace, since for any scalars a and /? and any u.v £ A/-^,
{au-\- (3v,(fi)^ = 0 for all (^ G A/; whence au -{- f3v e A/-^ follows. To
prove that Af-*- is complete, let us choose a Cauchy sequence {n„) C A/-^;
it converges to some u £ H because H is complete. We must show that
u E M^. Since for any v e H, and in particular for any v e A/, we have
{uny v)fj —> (w, v)fj as n —> oc and (un? ^')H = 0, n = 1,2,..., it follows that
(w, i^)// = 0 for any v G A/. Hence, tz G A/-'- and so Af-^ is complete.
Now, let if be a Hilbert space, M a subspace of if, and P the or-
thogonal projection operator of H onto M. Let u e H. From the prop-
erties of the projection we see that Re{u - Pu.v — Pu)fj < 0 for any
t; G Af. Choose v = Pu ± (p with (p being an arbitrary element of A/. Then
Re {u — Pu, i ^ ) / f < 0, whence Re {u — Pu, ^)fj = 0 . Replacing in the last
relation (f by J V (J^ = 1) we get

Re(w ~ Pu,j(f)ff ~ R e [ - j ( w - Pu,(f)fj] = Im(w - Pu,ip)ff = 0. (LIO)

Thus, for a given u e H the projection u; = Pu satisfies u — w 1 M.

Therefore, any element u £ H can be uniquely decomposed as
u = w-^w^, (Lll)
6 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

where w e M and w^ G M-^. This result is known as the theorem of

orthogonal projection.
The operator Q : H -^ M-^ given by

Qu = u-Pu (1.12)

is the orthogonal projection operator from H onto M^.

2 CLOSED AND COMPLETE SYSTEMS IN HILBERT SPACES.

BASES
Let X be a normed space and M a subset of X. M is dense in X if for
any u e X and any £ > 0 there exist u^ E M such that \\u — u^W^ < e.
Equivalently, M is.dense in X if and only if for any u e X there exists a
sequence {un) C M such that \\un — w||x —• 0 as n —> oo.
Every set is dense in its closure, i.e. M is dense in M. M is the largest
set in which M is dense; that is, if M is dense in K, then K C M.li M is
dense in a Hilbert space if, then M — H. Conversely, if M = i / , then M
is dense in H.
Let H he Si Hilbert space. If M is dense in H and u is orthogonal to
M, then u = 6H- Indeed, let uJLM and choose an arbitrary v E H. Since
M = H there exists a sequence (vn) C M such that \\vn — v||^ —> 0 as
n —> GO. Consequently, {u^Vn)^ —• (^?^)// as n —> oo. From {u,Vn)ff =
0, n = 1,2,..., it follows that (w, v)^ = 0 for any v e H. Thus, u ± H. The
element w is orthogonal to any element of H and in particular is orthogonal
to itself, i.e. {u,u)ff = ||u||^ = 0. Hence, u = OH-
Elements V^i,^2' "">'^N ^f ^ vector space X are called linearly depen-
dent if there exists a linear combination Yli^i ^i'^i = 0 in which the co-
efficients do not vanish, i.e. Yli=i l^^l > ^- ^ ^ ^ vectors are called linearly
independent if they are not linearly dependent, or equivalently, if there
exists no non-trivial vanishing linear combination. If any finite number of
elements of an infinite set {t/^J^i is linearly independent, the set {V'Ji^i
is called linearly independent.
A system of elements {V^jj^i is called closed in H if there are no
elements in H orthogonal to any element of the set except the zero element,
that means

(w,V^,)^ = 0 , z = 1,2,..., implies IX = 0H-

A system of elements {ipi}^i is called complete in H if the linear span

CX)
of {ipi}^i or the set of all finite linear combinations of {ipi}^
1=1

S p { ^ i , ^2. •••} = lu = J ^ a . ^ ^ Q i G C,7V = 1,2,... i

2. CLOSED AND COMPLETE SYSTEMS IN HILBERT SPACES. BASES 7

is dense in H, i.e. Bp {-^j, ip2^ •••} = H. Equivalently, if {t/^J^i is complete

in H then for any u e H and any e > 0 there exist an integer N = N{e)
and a set {ttr}^_i such that \\u — X)i=i ^f^^i "^ ^•
Let us observe that the closure of the linear span of any set {T/^^}^! is
a subspace of H. It is a vector subspace by its very definition and it is also
complete as a closed subset of a complete set.
Obviously, if the system {V^J^i is complete in H, then the only ele-
ment orthogonal to {ipi}^i is the zero element of H] thus the set {0i}i^i
is closed in H. The converse result is also true. To show this let { ^ J ^ i be
a closed system in H. Let us denote by W the linear span of {ipi}^i - Then
any element u£ H can be uniquely represented as w = Pu 4- Qu, where P
is the orthogonal projection operator from H onto W^ and Q is the orthog-
onal projection operator from H onto W . Since Qu G W and \l)i € W,
2 = 1,2,..., we get {Qu, il^^)u = 0, i = 1,2,.... The closeness of {V^J^i in H
implies Qu = 6H-> and therefore for any element u € H we have u = Pu €
W. Thus, H C W, and since W C H we get W = H; therefore W is dense
in H. We summarize this result in the following theorem.
T H E O R E M 2.1: Let H be a Hilbert space. A systetn of elements {'4^i}^i
is complete in H if and only if it is closed in H.
A set {ipi }?=i is called a finite basis for the vector space X if it is linearly
independent and it spans X. A vector space is said to be n-dimensional if
it has a finite basis consisting of n elements. A vector space with no finite
basis is said to be infinite-dimensional.
Let HN be a finite-dimensional vector subspace of a Hilbert space H
with orthogonal basis {<f>i}^-i. Then the orthogonal projection operator
from H onto HN is given by
N
PNU = Y2 (^' ^t)H ^i^ ueH.
t=l

For the time being we note a simple but important result characterizing
the convergence of the projections. Let {ipi}^i be a complete and linear
independent system in a Hilbert space /f, let H^ stand for the linear span
of {ipi}i^i, and let us denote by PN the orthogonal projection operator
from H onto HN. We have

\\u - P N + I ^ I I H = inf \\u - t;||^

Vfc/lN + l
(1.13)
< inf ||u - v\\fj = ||u - PNU\\H
v€riN

for any u e H; thus the sequence ||n - PN'^WH ^^ convergent. Since { ^ J ^ i

is complete in H we find a subsequence (uiv„) C HN^ such that ||u - UNn \\H
8 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

—> 0 as n —* oo. Then, from 0 < \\u — PN^'^^'WH ^ N — ^iVnll// we get

||u — PNa'^Wn —^ 0 as n —>• oc; thus the convergent sequence \\u — PNU\\H
possesses a subsequence which converge to zero. Therefore, for any xi £ H
we have

\\u - PNU\\U -^Oas N -^OO. (1.14)

A map i4 of a vector space X into a vector space Y is called linear if A

transforms linear combinations of elements into the same linear combina-
tions of their images, i.e. if ^ ( a i ^ i -]-a2U2 +...) = aiA{ui)-ha2A{u2)-{-.."
Linear maps are also called linear operators. In the linear algebra one usu-
ally writes arguments without brackets, A{u) = Au. Linearity of a map, is
for normed spaces, a very strong condition which is shown by the following
equivalent statements:
(a) A transforms sequences converging to zero into bounded sequences,
(b) A is continuous at one point (for instance at tx = 0),
(c) A satisfies the Lipschitz condition ||i4u||y < c||u||;^ for all u e X
and c independent on tx,
(d) A is continuous at every point.
Each number c for which the inequality (c) holds is called a bound for
the operator A.
Let C{X, Y) be the linear space of all linear continuous maps of a
normed space X into a normed space Y. The norm of an operator

uex,uy^0x \m\x l|u|lx=i

satisfies all the axioms of the norm in a normed space, whence the linear
space £(X, Y) is a normed space. Note that the number \\A\\ is the smallest
bound for yl. It is not difficult to prove that the space C{X, Y) is complete
if the space Y is such.
A map of a vector space into the space C of scalars is called a functional.
The above statements are valid for linear functional. The space £{X^ C)
is called the conjugate space of X and is denoted by X*. It is always a
Banach space.
A system {xpj}^^ is called minimal if no elements of this system belongs
to the closure of the linear span of the remaining elements. In order that
the system {V^l^i be minimal in a Banach space X, it is necessary and
sufficient that a system of linear and continuous functional defined on X
exist forming with the given system a biorthogonal system; that is, a system
of Hnear and continuous functionals { ^ j j ^ j such that ^j ( ^ J = 6ij^ where
6ij is the Kronecker symbol. If the system {V^^ j ^ j is complete and minimal,
then the system of functionals {^j}Jli is defined in a unique manner. In
2. CLOSED AND COMPLETE SYSTEMS LN HILBERT SPACES. BASES 9

a Hilbert space H, by Riesz theorem (see section 1.3), there exists ipj such
that J'j (u) = (^,^j) for any u € //; therefore (tj^^j)^
= 6ij. In this
case the system ^^^^^ _ is called biorthogonal to the system {t''j}^_|.
A system { ^ J ^ j is called a Schauder basis of a Banach space X if
any element u e X can be uniquel}^ represented as u = X]^^l ^?^^n where
the convergence of the series is in the norm of X. Every basis is a complete
minimal system. However, a complete minimal system may not be a basis in
the space. For example, the trigonometric system I/JQ (t) = 1/2, 02n-i(^) =
sin(n;^), ^2M (^) — cos(?if),n = 1,2, ...,is a complete minimal system in
the space C([—TT, TT]) but it does not form a basis in it. In an arbitrarily
separable Hilbert space if, every complete orthogonal systems of elements
forms a basis. Thus, the trigonometric system of functions forms a basis in
L2([-;r,7r]).
The system { 0 j ^ j is called an unconditional basis in the Banach
space A' if it remains a basis for an arbitrary rearrangement of its elements.
Let T : X -^ X hea bounded linear operator with a bounded inverse. If the
system {ipi}^i is a basis, then the system { T ^ j j ^ j is a basis. If {u%}^^
is an unconditional basis, then {Tu'i}^i is an unconditional basis. In a
Hilbert space, every orthogonal basis is unconditional. It can be shown that
an arbitrary unconditional basis in a Hilbert space is representable in the
form { T 0 ^ } ^ j , where {0^}J^i is an orthonormal basis oi H. Such bases are
called Riesz bases. If { ' 0 j ^ i is a Riesz basis then the biorthogonal system
\pi > is also a Riesz basis. A complete system {i^i}^i forms a Riesz
basis of H if the Gramm matrix G = [Gij], Gjj = {^i^'^'j)ff > generates
an isomorphism on /^. The system {t'^jj^i forms a Riesz basis of H if the
inequalities

N N

ci^\o^if < l]«^i^^ <C2^K|' (1.15)

i=l H *=1

hold for any constants QJ and for any iV, where the positive constants cj
and C2 should not depend on A^ and a^. Equivalently, { ^ J ^ i forms a Riesz
basis of H if there exist the positive constants ci and C2 such that
oo oo

\'<\\u\\l<C2'£\{u,i>,)i/ (1.16)
?=1 1=1

for arbitrary u E H. Note that if {t\}^i is a Riesz basis, then sup^ 11^/11// <
C2, infj ll^^llj:^ > ci and similar inequalities hold for the biorthogonal system

{*'}:,•
10 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

3 PROJECTION METHODS
One of the most important result of the theory of Hilbert spaces is the Riesz
theorem. This theorem states that if J^ is a hnear and continuos functional
on a Hilbert space i/, then there exists an unique element u/ e H such
that

J^{v) = {v,Uf)fj for all v e H, (1.17)

and IIJ^II^.
— 11^/11// • Conversely, any element u e H define a linear and
continuous functional on H by the relation
J^u{v) = {v,u)fj for all v e H, (1.18)

and we have ||^W|IH* ~ \W\\H-

Let ^ be a linear and continuous functional defined on a Hilbert space
H and {un) a weak convergent sequence with limit u e H. Then by the
Riesz theorem, J-^{un) —• ^{u) as n —• oo.
Let H he a, Hilbert space. The function B : H x H -^ C is called a
sesquilinear form on H if it is linear in the first argument and antilinear in
the second one, i.e.
B{ax-\-py,z) = aB(x,^) +/?B(y,^),
(1.19)
B{x,ay-\-f3z) = a'B{x,y)-h0'B(x,z),
for all x^y^z G H and all a,/? G C. A sesquilinear form B is bounded, if
there exists a constant M > 0 such that

\B{x,y)\ <M\\x\\ff \\y\\f, for all x,y e H, (1.20)

and strictly coercive, if there exists a constant c > 0 such that

ReB{x,x) > c||x||^ for all x e H. (1.21)

Let H he a. Hilbert space and B a bounded sesquilinear form on H. With

X being an element of H we define the functional T: i / —• C by J^{y) =
B*{x^ y). Then f E H^. According to Riesz theorem there exists an unique
element x/ G i/, such that J-^{y) = {yi^f)f{ for all y £ H. We define the
operator A : H -^ H hy Ax — x/ .Then B*{x^y) = {y,Ax)ff , and further
B{x^y) = {Ax^y)ff for all x^y e H. Let us now prove that A G C{H,H).
The linearity of -<4 is a consequence of the linearity in the first argument of
B. The boundedness of A follows from the boundedness of B, i.e. from

\\Ax\\l = {Ax, Ax)f, = B{x, Ax) < M ||a:||^ \\Ax\\^ (1.22)

we obtain ma;||;^ < M||a:||^ for all x e H. The operator A is uniquely

determined. To show this we suppose that there exist two operators Ai and
 3. PROJECTION METHODS 11

A2 such that B{x^y) = {Aix,y)jj = {A2X,y)fj for all x,y 6 H. Then, we

have {Aix — A2X, y)fj = 0 for all x^y e H, which implies Aix = A2X for all
X E H. Therefore, for any bounded sesquilinear form B : H x H -^ C there
exists an uniquely determined linear and bounded operator A : H —^ H
such that

B{x, y) = {Ax, y)fj for all x,y e H. (1.23)

In a similar manner we can prove the existence of a linear and bounded

operator A^ : H -^ H such that

B{x, y) = (x, A \ \ for all x,y £ H. (1.24)

The operator A^ is called the adjoint operator of A. Note that if B is

strictly coercive then A is strictly coercive, that is Re{AXyX)ff > c||rr||;^
for sll X € H and c > 0.
The Lax-Milgram lemma states that if B is a bounded and strictly
coercive sesquilinear form on a Hilbert space H, then the strictly coercive
bounded operator A : H —^ H generated by B has a bounded inverse
A-^ :H -^ H.
As a consequence of Riesz theorem and Lax-Milgram lemma if 6 is a
bounded and strictly coercive sesquilinear form and / " a bounded linear
functional on a Hilbert space H then the variational problem

B{%x) = T*{x) for all x e H, (1.25)

is unique solvable and the solution solves the operator equation

Au = / , (1.26)

where A is the operator generated by B and / is the uniquely determined

element corresponding to J^.
We are now well prepared to present the main result of this chapter,
namely the fundamental theorem of discrete approximation. This theo-
rem is frequently used in the finite-element method for solution of various
boundary-value problems by discrete schemes.
T H E O R E M 3.2: (fundamental theorem of discrete approxima-
tion) Let H be a Hilbert space and B a bounded sesquilinear form on H
satisfying

\B{x, x)\>c \\x\\]j for all xeH, (1.27)

Let T be a linear and continuous functional on H and {^i}^x ^ complete

and linearly independent system in H. Then
12 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

(a) the algebraic system of equations

N
Y,B{^,,iPj)a^=T*{i^j), j = l,...,iV, (1.28)

possess a unique solution^

(b) the sequence
N
UN = Y^ a^il^i (1.29)

is convergent; if \\UM — u\\fj —> 0 as N -^ oo, then u is the unique

solution to the variational problem
^*(x) = B{u,x) for all x G H, (1.30)

Proof: Before we present the proof we note that condition (1.27) is

weaker than the coerciveness condition (1.21). Coming now to the proof of
(a) we define the matrix B = [Bij] by Bij — B{'tl)^^ ipj)^ z, j = 1,2,..., N. Let
HN = Sp{^i,...,i/^;v} ^^d let PN be the orthogonal projection operator
from H onto Hjsf. With A standing for the operator generated by the
sesquilinear form B, i.e. B{x,y) = {Ax,y)fj , we have

Bij = 5 ( ^ i , ^ , ) = ( M , ^ i > H = (A^i.PN'ipj),, = (PNAiPi.tlj^)^ .

(1.31)
For any x € HN we use
c\M]t < \Bix,x)\ = \{Ax,x)ff\ = \{Ax,PNx)f,\ = \{PNAx,x)jf\

< WPNAXWJ, \\X\\H

(1.32)
to obtain ||P/v>la:||^ > c||x||^ . Consequently, the operator PjsfA : H^ —^
HN is invertible. Since {t/^jl^^^x form a basis of H^ we see that the vectors
(fi = PisfAtp^, i = . l,...,iV, form a basis of f//^. Let us denote by T =
[Tij] , z, j = 1,2,..., A/', the nonsingular transition matrix passing from the
basis {t/^jjli to the basis {iPi}^^i, i.e. (p^ = Ylk^i^ik^k^^^ ^ = 1,...,A^.
Then, we have
N N

k=l k=l

where $ = [*ij] , *ij = {'^ii'^j)fj ^h j =" 1? 2,..., AT, is the Gramm matrix
of the linearly independent system {ipi}^^i - The matrix B is expressed as
a product of two nonsingular matrices. Hence, B is nonsingular.
 3. PROJECTION METHODS 13

For proving (b) we rewrite (1.28) as

B{uN,i^j)^r{i^j), j-l,...,Ar. (1.34)

Multiplying the above relations by a^* and summing over j we obtain

B{UNJUN) = T*{UN). Then from

c||«^||^ < \B{UN,UN)\ = \r{uN)\ < WTWH' I K I I H (1.35)

we deduce that {UN) is bounded; thus we can pick up a weak convergent

subsequence {uNk) • Let u be the weak limit of this subsequence. Prom
(1.34) we get B{uNk^'^j) = T*{tl)j), j = l,,..,Nk. Since for any fixed j
the mapping x G i / , x «-* B{x^ ipj) is a linear and continuous functional on
H^ and since u^^ ~> u weakly asfc-^ oo, we obtain B{u,il)j) = J^*{tpj)
for any j = 1,2,.... Next, the completeness of the system { ^ J ^ i gives
B{u,x) = J^*{x) for any a: e H, Let us prove that u is unique. Assume
that there exists u' ^ u such that B{u\x) — ^*{x) for any x E H. Then
B{u — w', a:) = 0 for any x E H and from

0 = \B{u -u'.U'- u')\ >c\\u- u'tfj > 0 (1.36)

the conclusion readily follows. Thus, all weak convergent subsequences have
the same weak limit; whence UN -^ u weakly as AT —> oo (cf. Dinca [43]).
Let us now prove the more stronger result, namely that \\UN — u\^ —• 0
as iV —> oo. Using

(1.37)
< 15(1*AT, WAT) - S(tiAr, u) - B{u, UN) -f B(u, u)\

and the identities B{U^UN) = T*{UN) and B{UN,UN) = J^*{UN) we get

c \\UN - u\\]j < \B[uN.u) - B{u, u)\. (1.38)

Since UN -^ u weakly as AT -> oo and the mapping x 6 if, a; H-> B{X, U)

defines a linear and continuous functional on H we obtain B{UN^U) —•
B{u^ u) as N —* oo, and the conclusion readily follows.
Evidently, the fundamental theorem of discrete approximation is also
valid for a strictly coercive sesquilinear form B. In this context, the unique
solution to the the variational problem (1.30) coincides with the unique
solution to the operator equation (1.26). The projection relations (1.28)
may be written as

< ^ w j v ~ / , ^ ^ ) ^ = 0 , i = l,...,iV, (1.39)

14 CHAPTER I ELEMENTS OF FUNCTIONAL ANALYSIS

or equivalently as

PNAPNUN = PN/^ (1.40)

where P/v is the orthogonal projection operator from H onto Hjsf and
HN = Sp{V^i,..,V^^}. The above projection method is also called the
Galerkin method. The strongest condition which guarantee the conver-
gence of the projection scheme is the strictly coercivity of the sesquilinear
form B. According to (1.32) we see that this condition implies

WPNAPNUW^^ > c \\PNU\\H for all ueH. (1.41)

Let us generalize the above results when A is a. linear bounded and

boundedly invertible operator from a Hilbert space H onto a Hilbert space
G. Let HN C HN^I with dimHN = AT be a sequence of subsets limit
dense in H, i.e. for any u £ H, P{U^HN) —• 0 as iV —^ OO, and let PN
stands for the orthogonal projection operator onto HN. Analogously, let
GN C Giv+i with dim GAT = AT be a sequence of subsets limit dense in G
and let QN stands for the orthogonal projection operator onto GN. The
projection method giving the approximate solution u^ of (1.26) is

QNAPNUN = QN/- (1.42)

Then we can formulate the following result (cf. Ramm [128]).

T H E O R E M 3.3: Let A : H -^ G be a linear bounded and. boundedly
invertible operator. Equation (1.4^) is unique solvable for all sufficiently
large N, and

\W - '^NWH - • 0 as AT - • 00, (1.43)

if and only if

WQNAPNUW^ > c \\PNu\\fj for all w € ^ and AT > A^o, (1-44)

where NQ is some integer and c > 0 does not depend on N and u.

Proof: Let us prove the necessity. Assume that (1.43) holds and
(1.42) is unique solvable. Then for / G G we have \\UN — u\\ff —> 0 as
N -^ oo, where UN = {QNAPN)~^ QNf and u = A~^f. Thus

sup {QNAPNr'QN\\ <c<oo. (1.45)

N

Since QNAPNU = Q^QNAP^U and P^PN'^ = PNU we have

QNQNAPMU = QNAPNPNU. (1.46)

 3. PROJECTION METHODS 15

Thus,

\\PNU\\H = WiQNAPNr^ QNQNAPNU\\ <c\\QNAPNu\\a (1.47)

II IIH
and (1.44) is proved. For proving the sufficiency we assume (1.44). Con-
sequently, the operator Q^APN \ P^H -^ QNG is an injective mapping
between two A/'-dimensional spaces and therefore this mapping is surjective.
Hence, (1.42) is unique solvable for N > No. From

QNA [PNU -f (/ - PN) U] = Qiv/,

(1.48)
QNAPNUN = QN/I

we obtain

QNAPN {UN - PNU) = QNA (/ - PN) U. (1.49)

Since HN is limit dense in il, it follows that \\u — P/v^||// —> 0, iV -* oo.
Then, from
\\UN-PNU\\U = \\PN{UN -PNy)\\u

< '\\QNAPN{UN-PNU)\\G^-\\QNA{I^PN)U\\C
c c

< l\\A{I-PN)u\\a<^\\u-PMu\\j,
(1.50)
we see that \\UN — -Pivw||^ —^ 0 as A^ —^ oo; whence, by the triangle in-
equality, we find that ||t* — i^ivll// —* 0 as iV —> oo. This finishes the proof
of the theorem.
The following theorem will also be used many times in the sequel.
T H E O R E M 3.4: Let A : H —^ G he a linear hounded and houndedly
invertible operator satisfying (1.44)- Let B : H —^ G be a compact operator
and A-\- B he hounded invertihle. Then,

^QN{A + B)PNu\\ci>c\\PNu\\jj foralUG^andiV>iVo, (1.51)

where NQ is some integer and c > 0 does not depend on N and u.
Proof: For the proof we refer to Ramm [128].
Theorems 3.3 and 3.4 show that the equation

QN{A-i-B)PNUN = QNf (1.52)

is unique solvable for all sufficiently large AT and \\u — UNWH ~^ 0 as iV —>
oo, where u is the exact solution to the operator equation [A -f B)u = / .