T

Chapter 4

4.1

AF

VARIATIONAL METHODS

Introduction to variational principles

Certain classes of partial differential equations possess a variational structure. This means
that their solutions u can be interpreted as extremal points over a properly defined function
space U, with reference to given functionals I[u]. By way of introduction to variational
methods, consider a functional I[u] defined as
I[u] =

k
2

u
x1

2

k
+
2

u
x2

2


+ f u d ,

(4.1)

where k = k(x1 , x2 ) > 0 and f = f (x1 , x2 ) are continuous functions in . In addition,

DR

assume that the domain possesses a smooth boundary with uniquely defined outward

unit normal n.

The functional I[u] attains an extremum if, and only if, its first variation vanishes, namely

I[u] =

k
x1

u
x1




u
u
+ k
+ f u d

x2
x2
Z
 u u

u u
=
k
+
k
+ f u d = 0 , (4.2)
x1
x2 x2
x1

where u is assumed continuously differentiable. Following the developments of Section 3.2,

integration by parts and application of the divergence theorem on (4.2) yields


Z
Z




u
u
u

u
I[u] =
k
n1 +
n2 u d
(k
)+
(k
) f u d = 0 .
x1
x2
x1
x2 x2

x1
(4.3)
63

64

Variational methods

I[u] =

u
n
x1 1

u
n
x2 2

u
u d
n

u
,
n

write



u

u
(k
) +
(k
) f u d = 0 .
x1
x2 x2
x1

Recalling that

(4.4)

Owing to the arbitrariness of u, the localization theorem of Section 3.1 implies that

u
(k
) +
(k
) = f
x1 x1
x2 x2
and

in

(4.5)

u
u = 0 on ,
(4.6)
n
conditional upon sufficient smoothness of the respective fields. The first of the above two

AF

equations is identical to the Laplace-Poisson equation (3.5)1 , while the second equation
presents three distinct alternatives:
(i) Set

u = 0 on .

(4.7)

This condition implies that the dependent variable u is prescribed throughout the
boundary . The space of admissible fields u is defined as


U = u H 1 () | u = u on ,

(4.8)

where u is prescribed independently of the functional I[u], in the sense that the functional contains no information regarding the actual value of u on . Boundary conditions such as u = u, which appear in the space of admissible fields, are referred to as

DR

essential (or geometrical).

(ii) Set

u
= 0 on .
(4.9)
n
In this case, the boundary condition applies on the extremal function u, and is exactly
k

derivable from the functional. Boundary conditions that directly apply to the extremal
function (and its derivatives) are referred to as natural (or suppressible). No boundary
restrictions are imposed on U in the present case.

(iii) Admit a decomposition of boundary into parts u and q , such that = u q .

Subsequently, set

u = 0 on u ,
u
= 0 on q .
k
n

ME280A

(4.10)

Version: September 22, 2010, 22:11

Introduction to variational principles

65

Here, essential and natural boundary conditions are enforced on mutually disjoint

portions of the boundary. In this case, the problem is said to involve mixed boundary
conditions, and the space of admissible fields is defined as
U =


u H 1 ()

u = u on u

(4.11)

It can be concluded from the above, with reference to (4.4) that essential boundary conditions
appear on variations of u and, possibly, its derivatives (and therefore place restrictions on the

AF

space of admissible fields), while natural boundary conditions appear directly on derivatives
of the extremal function u. Equation (4.4) reveals that extremization of the functional
in (4.1) yields a function u which satisfies the differential equation (3.5)1 and boundary
conditions selected in conjunction to the space of admissible fields U.

Following case (iii), note that the space of admissible variations U0 is defined as
U0 =


u H 1 ()

u = 0 on u

= H01 () .

(4.12)

It can be easily seen that the option of non-homogeneous natural boundary conditions of
the form

u
= q on q
n
can be accommodated, if the original functional is amended so that it reads



2
Z
Z
 k u 2

k u

I[u] =
qu d ,
+
+ f u d +
x1
2 x2
2
q

DR

(4.13)

(4.14)

where q = q(x1 , x2 ) is a continuous function on q .

implies that
Vanishing of the first variation of I[u]
Z


 u u
u u
k
+
k
+ f u d +
x1
x2 x2
x1

qu d = 0 .

(4.15)

Equation (4.15) is termed the weak (variational) form of boundary-value problem (3.5).
Comparing the above equation to (3.14), it is obvious that they are identical provided that
the space of admissible field W for w in (3.14) is identical to that of u in (4.15).
The nature of the extremum point u (i.e., whether it renders I[u] minimum, maximum or

merely stationary) can be determined by means of the second variation of I[u]. Specifically,
write

I[u] = I[u]
=

Version: September 22, 2010, 22:11

u
x1

2

+ k

u
x2

2

d ,

(4.16)
ME280A

66

Variational methods

> 0, for all u 6= 0, provided u 6= . This is true because, if u is

and note that 2 I[u]

assumed to be constant throughout the domain, it has to vanish everywhere, by definition

= 0 and 2 I[u]
> 0 are sufficient for any I[u] to
of U0 . It turns out that the conditions I[u]
is also bounded from below at u, namely
attain a local minimum at u, provided that 2 I[u]
that

2 I[u]
ckuk2 ,

AF

where c is a positive constant.

(4.17)

The weak (variational) form of problem (3.5) can be stated operationally as follows: find
u U, such that for all u U0

B(u, u) + (u, f ) + (u, q)q = 0 ,

(4.18)

where the bilinear form B(u, u) is defined as

B(u, u) =

Z 

u u
u u
k
+
k
x1 x1
x2 x2

d ,

(4.19)

and the linear forms (u, f ) and (u, q)q are defined, respectively, as
(u, f ) =

uf d

(4.20)

DR

and

(u, q)q =

u
q d .

(4.21)

The correspondence of the above operational form with that of Section 3.2 is noted for the
purpose of the forthcoming comparison between the Galerkin method and the variational
method, when applied to problem (3.5).

In addition to the above weak (variational) form, there exists a variational principle

associated with the solution u of problem (3.5). This can be stated as follows: find u U,
such that

,
I[u]
I[v]

(4.22)

is defined in (4.14).
for all v U, where I[u]
ME280A

Version: September 22, 2010, 22:11

Weak (variational) forms and variational principles

67

Remark:
Directional derivatives can be used in deriving the weak (variational) equation (4.15)

from functional I[u].

Indeed, write
h d
i
+ v]
I[u
= 0
d
=0
Z
Z

 u v
u v
qv d = 0 , (4.23)
k
+
k
+ f v d +
x2 x2
q
x1 x1

AF

Dv I[u]
= 0

where v U0 .

4.2

Weak (variational) forms and variational principles

The analysis in Section 4.1, as applied to the model problem (3.5), provides an attractive
perspective to the solution of certain partial differential equations: the solution is identified
with a point, which minimizes an appropriately constructed functional over an admissible function space. Weak (variational) forms can be made fully equivalent to respective
strong forms, as evidenced in the discussion of the weighted residual methods, under certain
smoothness assumptions. However, the equivalence between weak (variational) forms and
variational principles is not guaranteed: indeed, there exists no general method of constructing functionals I[u], whose extremization recovers a desired weak (variational) form. In this

DR

sense, only certain partial differential equations are amenable to analysis and solution by
variational methods.

Vainbergs theorem provides the necessary and sufficient condition for the equivalence

of a weak (variational) form to a functional extremization problem. If such an equivalence

holds, the functional is referred to as a potential.
Theorem (Vainberg)

Consider a weak (variational) form

G(u, u) = B(u, u) + (f, u) + (

q , u)q = 0 ,

(4.24)

where u U, u U0 , and f and q are independent of u. Assume that G possesses a Gateaux derivative in a neighborhood N of u, and the Gateaux differen-

tial Du1 B(u, u2) is continuous in u at every point of N . Then, the necessary and

Version: September 22, 2010, 22:11

ME280A

68

Variational methods

that

sufficient condition for the above weak form to be derivable from a potential in N is
Du1 G(u, u2) = Du2 G(u, u1) ,

(4.25)

namely that Du1 G(u, u2) be symmetric for all u1 , u2 U0 and all u N .
Preliminary to proving the above theorem, introduce the following two lemmas:
Lemma 1
Show that

I[u + v] I[u]
.

AF

Dv I[u] = lim

(4.26)

To prove that Lemma 1 holds, use the definition of the directional derivative of I in the
direction v, so that

 d

I[u + v] =0
d

I[u + v + v] I[u + v]
=
lim
=0
0

 I[u + v + v] I[u + v]
= lim
=0
0

I[u + v] I[u]
= lim
.
0

Dv I[u] =

In the above derivation, note that operations

d
d

(4.27)

and |=0 are not interchangeable (as they

both refer to the same variable ), while lim0 and |=0 are interchangeable, conditional
upon sufficient smoothness of I[u].

DR

Lemma 2 (Lagranges formula)

Let I[u] be a functional with Gateaux derivatives everywhere, and u, u + u be any

points of U. Then,

I[u + u] I[u] = Du I[u + u]

0<<1.

(4.28)

To prove Lemma 2, fix u and u + u in U, and define function f on R as

f () = I[u + u] .

It follows that
df
f ( + ) f ()
f =
= lim
0
d

I[u + u + u] I[u + u]
= Du I[u + u] ,
= lim
0

ME280A

(4.29)

(4.30)

Version: September 22, 2010, 22:11

Weak (variational) forms and variational principles

69

where Lemma 1 was invoked. Then, using the standard mean-value theorem of calculus,

write

I[u + u] I[u] = f (1) f (0)

=

(4.31)

f (1) f (0)
= f () = Du I[u + u] ,
1 0

where 0 < < 1.

Given Lemma 2, one may proceed directly to the proof of Vainbergs theorem. First,

AF

prove the necessity of (4.25), namely that if an appropriate I[u] exists, then (4.25) holds.
To this end, fix u N and define the scalar quantity as

= I[u + a u1 + b u2 ] I[u + a u1 ] I[u + b u2 ] + I[u] ,

(4.32)

where a, b are non-zero scalars, such that u + u1 + u2 N , for all 0 a and

0 b. Also, define functional J[u] as

so that

J[u] = I[u + a u1 ] I[u] ,

(4.33)

= J[u + b u2 ] J[u] .

(4.34)

Using Lemma 2 and the above definitions of J[u] and , write

= J[u + b u2 ] J[u] = Dbu2 J[u + 1 b u2 ]

= Dbu2 I[u + 1 b u2 + a u1] Dbu2 I[u + 1 b u2 ]

DR

= B(u + 1 b u2 + a u1 , bu2 ) + (f, b u2 ) + (

q , b u2 )q
B(u + 1 b u2 , b u2) (f, b u2 ) (
q, b u2 )q

= b B(au1 , u2) = ab B(u1 , u2) .

(4.35)

Alternatively, let functional K[u] be defined as

K[u] = I[u + b u2 ] I[u] ,

(4.36)

= K[u + a u1 ] K[u] .

(4.37)

so that is also written as

Using the steps followed in the derivation of (4.35), it can be readily concluded that
= ab B(u2 , u1 ) ,

Version: September 22, 2010, 22:11

(4.38)
ME280A

70

Variational methods

so that (4.35) and (4.38) lead to

B(u1 , u2) = B(u2 , u1 ) .
Noting that, due to the linearity of B,

 d

B(u + u1 , u2) =0
d
 d 

=
B(u, u2 ) + B(u1, u2 ) =0 = B(u1 , u2 ) ,
d

and, similarly,

AF

Du1 B(u, u2) =

Du2 B(u, u1) = B(u2 , u1 ) ,

(4.39)

(4.40)

(4.41)

it follows that condition (4.25) holds.

In order to show the sufficiency of (4.25), namely prove that (4.25) implies the existence
of an appropriate functional I[u], define
Z 1
I[u] =
B(tu, u) dt + (f, u) + (
q , u)q .

(4.42)

Since

and

d
B(tu, u + u) = B(tu, u) ,
d

DR

d
B( u, u + u) = B(u, u) + 2B(u, u) ,
d
the directional derivative of I[u] in the direction u is written as
Z 1


Du I[u] = Du
B(tu, u) dt + (f, u) + (
q, u)q .

(4.43)

(4.44)

(4.45)

With the aid of (4.43) and (4.44), the first term on the right-hand side of the above equation
may be rewritten as
Z 1
Z 1


Du
B(tu, u) dt =
Du B(tu, u) dt
0
0
Z 1
 d

=
B(tu + t u, u + u) =0 dt
d
Z0 1

 d
d
=
B(tu, u + u) +
B(t u, u + u) =0 dt
d
d
0
Z 1


=
B(tu, u) + B(t u, u) dt .
(4.46)
0

ME280A

Version: September 22, 2010, 22:11

Weak (variational) forms and variational principles

71

Du

Exploiting the assumed symmetry of B, it follows from the above that

1

B(tu, u) dt = 2

tB(u, u) dt

 t2 1
= B(u, u) ,
= 2B(u, u)
2 0

(4.47)

which proves that I[u], as defined in (4.42), is indeed an appropriate functional.

AF

Remarks:

Apart from some technicalities, Vainbergs theorem can be proved following the above
general procedure, even when B is non-linear in u.

Checking condition (4.25) is typically an easy task.

Vainbergs theorem not only establishes a condition for potentiality of a weak from,
but also provides a direct definition of the potential in the form of (4.42).

Example:

Recall the weak (variational) form of Section 3.2, which is associated with boundary-value problem (3.5). In this case,
Z 

B(u, u) =

DR

u u
u u
k
+
k
x1 x1
x2 x2

(f, u) =

and

(
q , u)q =

d ,

f u d ,

q u d .

Using Vainbergs theorem, it can be immediately concluded that, since B is symmetric, there exists
a potential I[u], which, according to (4.42), is given by
1
B(u, u) + (f, u) + (
q , u)q
2




Z
Z

 k u 2
k u 2
u
q d ,
+
+ f u d +
=
x1
2 x2
q
2

I[u] =

whose extremization yield the above weak (variational) form.

Version: September 22, 2010, 22:11

ME280A

72

Rayleigh-Ritz method

4.3

Variational methods

The Rayleigh-Ritz method provides approximate solutions to partial differential equations,

whose weak (variational) form is derivable from a functional I[u]. The central idea of the
Rayleigh-Ritz method is to extremize I[u] over a properly constructed subspace Uh of the
space of admissible fields U. To this end, write

N
X
.
u = uh =
I I + 0 ,

(4.48)

AF

I=1

where I , I = 1, . . . , N, is a specified family of interpolation functions that vanish where

essential boundary conditions are enforced. In addition, function 0 is defined so that uh
satisfy identically the essential boundary conditions. Consequently, a proper N-dimensional
subspace Uh is completely defined by (4.48). Extremization of I[u] over Uh yields
I[uh ] = I[

N
X

I I + 0 ] = 0 .

(4.49)

I=1

Instead of directly obtaining the weak (variational) form of the problem by determining the
explicit form of I[uh ] as a function of uh , one may rewrite the extremization statement as
a function of parameters I , I = 1, . . . , N, namely

I(1 , . . . , N ) = 0 .

(4.50)

It follows from (4.50) that I(1 , . . . , N ) is an integral scalar function which attains an

DR

extremum over Uh if, and only if,

I
I
I
1 +
2 + . . . +
N = 0 .
1
2
N

(4.51)

Since the variations I , I = 1, . . . , N, are arbitrary, it may be immediately concluded that

I
= 0
I

I = 1, . . . , N .

(4.52)

Equations (4.52) may be solved for parameters I , so that an approximate solution to the

variational problem is expressed by means of (4.48).

Example: Consider the functional I[u] defined in the domain (0, 1) as
I[u] =

ME280A

1

1
2

du
dx

2



+ u dx + 2ux=1 ,

Version: September 22, 2010, 22:11

Rayleigh-Ritz method

73

and the associated essential boundary condition u(0) = 0. The above functional is associated with
the one-dimensional version of the Laplace-Poisson equation discussed in Section 3.2. In particular,
it can be readily established that extremization of I[u] recovers the solution to a boundary-value
problem of the form
d2 u
= 1
dx2
du

= 2
dx
u = 0

in = (0, 1) ,
on q = {1} ,

AF

on u = {0} .

In order to obtain a Rayleigh-Ritz approximation to the solution of the preceding boundary-value

problem, write uh as
N
X
I I (x) + 0 (x) ,
uh (x) = uN (x) =
I=1

and set, for simplicity, 0 = 0, so that the homogeneous essential boundary condition at x = 0 be
satisfied. A one-parameter Rayleigh-Ritz approximation can be determined by choosing 1 (x) = x.
Then,
1


1 2
1 + 1 x dx + 21
0 2
5
1 2
= 1 + 1 .
2
2

I[u1 ] =

Setting the first variation of I[u1 ] to zero, it follows that

1 +

5
= 0,
2

DR

from where it is concluded that 1 = 52 , and

5
u1 (x) = x .
2

Similarly, one may consider a two-parameter polynomial Rayleigh-Ritz approximation by choosing

1 (x) = x and 2 (x) = x2 . In this case, I[u] takes the form
1


1
(1 + 22 x)2 + 1 x + 2 x2 dx + 2(1 + 2 )
0 2
1
2
5
7
= 21 + 1 2 + 22 + 1 + 2 .
2
3
2
3

I[u2 ] =

Setting the first variation of I[u2 ] to zero, results in the system of equations
5
,
2
7
= ,
3

1 + 2 =

1 +

Version: September 22, 2010, 22:11

4
2
3

ME280A

74

Variational methods

whose solution gives 1 = 3 and 2 = 12 , hence

1 2
x .
2

u2 (x) = 3x +

AF

The approximate solution u2 (x) coincides with the exact solution of the boundary-value problem.
Furthermore, u1 (x) and u2 (x) coincide with the respective solutions obtained in Section 3.2 using
the Bubnov-Galerkin method with the same interpolation functions.
A different approximate solution u
2 can be obtained using the Rayleigh-Ritz method in connection with piece-wise linear polynomial interpolation functions of the form

2x
if 0 x 0.5
1 (x) =
2(1 x) if 0.5 < x 1
and

2 (x) =

0
2(x

if 0 x 0.5
if 0.5 < x 1

1
2)

where functions 1 and 2 are depicted in Figure 4.1. Then, I[

u2 ] is written as

DR

Figure 4.1: Piecewise linear interpolations functions in one dimension

0.5 


1
(21 )2 + 21 x dx
2
0
Z 1
1
1 
+
(21 + 22 )2 + 21 (1 x) + 22 (x ) dx + 22
2
0.5 2
1
9
= 221 21 2 + 22 + 1 + 2 .
2
4

I[
u2 ] =

Again, setting the variation of I[

u2 ] to zero yields

1
,
2
9
= ,
4

41 22 =

21 + 22

5
so that 1 = 11
8 and 2 = 2 , and

u
2 (x) =

ME280A

if 0 x 0.5
11
4 x
14 (1 + 9x) if 0.5 < x 1

Version: September 22, 2010, 22:11

Exercises

75

1.0
u2

u2

u1

1.0

0.5

AF

Figure 4.2: Comparison of exact and approximate solutions

Solutions u1 , u2 and u
2 are plotted in Figure 4.2

The Rayleigh-Ritz method is related to the Bubnov-Galerkin method, in the sense that,
whenever the former is applicable, it yields identical approximate solutions with the latter,
when using the same interpolation functions. However, it should be understood that, even
in these cases, the methods are fundamentally different in that the former is a variational
method, whereas the latter is not.

4.4

Exercises

DR

Problem 1
Consider the fourth-order ordinary differential equation
d4 u
= f
dx4

in = (0, 1) ,

where f is

a function of x. No boundary conditions are prescribed at this stage on =
{0}, {1} .

(a) Multiply the differential equation by a function v and subsequently integrate over the
domain (note that other than the standard integrability requirement, no restrictions
are placed on v, since no boundary conditions have been specified).

(b) Perform two successive integrations by parts on the above integral to obtain
Z

nZ  1


d4 u

f
v
dx
=
D
v
dx4
2
3
3
d u
d u
(1) v(1) 3 (0) v(0)
3
dx
dx

Version: September 22, 2010, 22:11

 o
d2 u
2

f
u
dx +
dx2
d2 u
dv
d2 u
dv
(1)
(1)
+
(0) (0) .
2
2
dx
dx
dx
dx
ME280A

76

Variational methods

(c) Conclude from part (b) that stationarity of the functional I[u], defined as
Z 

1 d2 u
2
I[u] =

f
u
dx ,
2
2 dx

implies that the given differential equation is satisfied in and, moreover,

= 0,
= 0,

AF

d3 u
(1) v(1)
dx3
d3 u
(0) v(0)
dx3
d2 u
dv
(1) (1)
2
dx
dx
dv
d2 u
(0) (0)
2
dx
dx

= 0,
= 0.

(d) Identify all possible essential and natural boundary conditions on . Note that essential boundary conditions appear on the variations v (and therefore restrict the admissible
fields), while natural boundary conditions appear directly on derivatives of the extremal
function u.
(e) Consider an expanded functional I1 [u] given by
Z 

du
1 d2 u
2

f
u
dx + q1 (0)u(0) + q1 (1) (1) ,
I1 [u] =
2
2
dx
dx

DR

where q1 is defined on , and derive the boundary equations associated with stationarity of I1 [u]. Again, identify all possible essential and natural boundary conditions on
. Can the functional be further amended so that it read
Z 

1 d2 u
2
du
d2 u

f
u
dx
+
q
(0)u(0)
+
q
(1)
(1)
+
q
(1)
(1) ,
I2 [u] =
1
1
2
2
dx
dx2
2 dx
where q2 is defined at x = 1? Clearly explain your answer.

Problem 2
Consider the initial-value problem

u

u

in (0, T ) ,
k
f = l
x x
t
u = u
(x, t) on (0, T ) ,
u(x, 0) = u0 (x) in ,

for the determination of u = u(x, t), where R, and k, l and f are given non-vanishing
functions of x. Starting from the above strong form, obtain the weak (variational) form of
the problem and use Vainbergs theorem to show that there exists no variational theorem
associated with the weak form.

ME280A

Version: September 22, 2010, 22:11

Problem 3
Consider the boundary-value problem

77

Exercises

u

d
(1 + x)
= 0 in (0, 1) ,
dx
x
u(0) = 0 ,
u(1) = 1 .

AF

Construct the weak (variational) form of this problem and use Vainbergs theorem to ascertain
that there exists a variational principle associated with the problem. Also, construct the
relevant functional I[u] and specify the space over which it attains a minimum. Obtain a
Rayleigh-Ritz approximation to the solution of the above problem, assuming a two-parameter
polynomial approximation. Submit a plot of the exact and the approximate solution.

Problem 4
Consider the boundary-value problem
 2

u 2u
+
= f
in = {(x1 , x2 ) | 0 < x1 , x2 < 1} ,
x21 x22
u
= 0
on q = {(x1 , x2 ) | x1 = 0 or x2 = 0} ,
n
u = 0
on u = {(x1 , x2 ) | x1 = 1 or x2 = 1} ,
where f = f (x1 , x2 ). In particular, minimize an appropriate functional I[u] over a properly
defined functional space U (treat boundary conditions on x1 = 0 and x2 = 0 as natural)
using a one-parameter polynomial approximation
u1 (x1 , x2 ) = 1 1 ,

DR

where 1 (x1 , x2 ) = (1x1 )(1x2 ). Also, suggest a two-parameter polynomial approximation

u2 (x, y) = 1 1 + 2 2

by specifying 2 to be the next to 1 in the hierarchy of admissible polynomials in U. In this

part, do not solve for the coefficients 1 and 2 .

Problem 5
Consider the homogeneous boundary-value problem
4u
4u
4u
+
2
+
= f
x41
x21 x22
x42
u = 0 ,

u
= 0
n

in R2 ,

on ,

where f = f (x1 , x2 ).

Version: September 22, 2010, 22:11

ME280A

78

Variational methods

x2
x1

AF

Show that the solution u of the above problem extremizes the functional I[u] defined as
Z h n 2
i
2u 2u
2 u
2 o
1 u
2
+
2

f
u
d ,
+
I[u] =
x21
x21 x22
x22
2

over the set of all admissible functions U. Verify that the same conclusion can be reached for
the functional I1 [u] defined as
Z h n 2
i
2 u
2
2 u
2 o
1 u
2

f
u
d .
+
2
+
I1 [u] =
x1 x2
x21
x22
2

4.5

Suggestions for further reading

Sections 4.1

[1] K. Washizu. Variational Methods in Elasticity & Plasticity. Pergamon Press, Oxford, 1982. [This is a classic book on variational methods with emphasis on structural
and solid mechanics].

DR

[2] H. Sagan. Introduction to the Calculus of Variations. Dover, New York, 1992.
[This book contains a complete discussion of the theory of first and second variation].

Section 4.2

[1] M.M. Vainberg.

Variational Methods for the Study of Nonlinear Operators.

Holden-Day, San Francisco, 1964. [This book contains many important mathematical
results, including a non-linear version of Vainbergs theorem].

Section 4.3

[1] B.A. Finlayson and L.E. Scriven. The method of weighted residuals a review.
Appl. Mech. Rev., 19:735748, 1966. [This review article contains on page 741 an
exceptionally clear discussion of the relationship between Rayleigh-Ritz and Galerkin
methods].

ME280A

Version: September 22, 2010, 22:11