You are on page 1of 40

Lecture Notes on

Computational and Applied Mathematics


(MAT 3310, 3 units)

Academic year 2009/2010: second term

Prepared by
Jun ZOU (2006) and Patrick CIARLET (2010)b
a

We have prepared the lecture notes purely for the convenience of our teaching. We gratefully
acknowledge the contribution of Prof. I-Liang CHERN (National Taiwan University, Taipei, Taiwan),
made during his visit to the Dept of Mathematics of CUHK in 2009.
a
Department of Mathematics, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong
b
Dept of Applied Mathematics, Ecole Nationale Superieure de Techniques Avancees, Paris, France

At some locations, we put the symbol .


This corresponds to paragraphs that may
be omitted in a rst reading.

Equilibrium equations and variational principles

Previously, we formulated the spring-mass system and elastic bar models with the help of
force balance laws. Here, we reformulate those models in terms of the minimization of the
energy of the system we study. These minimization problems are expressed variationally,
with or without constraint. The rst part of this chapter is devoted to the variational
formulation of the discrete model (spring-mass system), and the second part is devoted
to the continuous model (elastic bar). We also propose some physical points of view
along the way.

2.1

Variational formulation for the spring-mass system

In chapter 1, we formulated the spring-mass system at equilibrium as sets of algebraic


equations. We proposed two equivalent formulations, namely (1.6) and (1.5), which we
recall here:
the displacement-force form:
Ku := AT CAu = f.

(2.1)

Above, u , f Rn , m n,26 A Rmn is injective (its kernel, or null-space,


Ker(A) is reduced to {0}), and C Rmm is symmetric positive-denite. As a
result, K Rnn is a symmetric positive-denite matrix27 .
the stress-displacement form:


 

C 1 A
w

0
=
.
AT 0
u
f

(2.2)

Above, w
 Rm .
In this chapter, we add a to the solutions to problems (2.1) and (2.2), to make sure
they are dierent from the other vectors!
26

In a connected system, the number of nodes, n, is always lower or equal to the number of links
minus 1, that is m1. In the special case when n = m1, one can eliminate one unknown ui , 1 i n.
So, the inequality m n always holds.
27
Since C is symmetric positive denite, K := AT CA is automatically symmetric and positive, where
the latter means that (Kv, v) 0 for all v Rn . Check that K is positive denite if, and only if,
Ker(A) = {0}.

27

The rst formulation will be shown to be equivalent (see 2.1.1) to a minimization


problem:
1
MinuRn P (u) := (Ku, u) (f, u).
2
 Physical viewpoint. The physical meaning of P (u) is the total potential energy of the
spring-mass system. The rst term is the potential energy stored in the springs. The
second term is the gravitation potential energy of the masses.
Indeed, if we recall that u0 = un+1 = 0, we nd that
1
1
1
1
(Ku, u) = (AT CAu, u) = (CAu, Au) =
ci (ui ui1 )2
2
2
2
2
i=1
n

which is the sum of the potential energy stored in the spring.


Whereas the term
n

(f, u) =
fi ui
i=1

is the sum of the works done by the masses mi with displacement ui for i = 1, ..., n. In
other words, the term (f, u) is the gravitational potential energy due to the masses mi
with displacements ui .

The second formulation will be shown to be equivalent (see 2.1.2) to a constrained


minimization problem:
1
m Q(w)
MinwR
 w)

 := (C 1 w,

2
 = f.
w
 subject to AT w
Among others, we shall show that the two variational formulations are equivalent. We
will also derive similar variational formulations for the elastic bar at equilibrium.
 Physical viewpoint. The physical meaning of w
 is the stress, or the restoration
force, and it plays the role of the internal force in the spring. Thus, AT w
 = f appears
as the force balance constraint, with f playing the role of the exterior force.
Notice that w
 is a vector with components (w1 , ..., wm ): its component wj is the stress
of the spring sj . The quantity c1
j wj is the elongation of the spring sj . When we let the
stress vary from 0 to wj , the potential energy stored in the spring sj is the integral
 wj
1 1 2
c1
j w dw = cj wj .
2
0
After we sum these potential energies over all springs, we get that the physical meaning
of Q(w)
 is the total potential energy stored in the springs due to the stress w.


28

From now on, let us drop the above vectors in section 2.1...
2.1.1

Minimum principle

Consider the functional dened over Rn ,


1
P (u) := (Ku, u) (f, u),
2

(2.3)

where K is a symmetric positive denite matrix28 . It is made up of two parts, which


play dierent roles in the minimization process: a quadratic part, Pquad (u) := 12 (Ku, u),
and a linear part, Plin (u) := (f, u). In other words, if one multiplies u by a real number
t, the quadratic part is multiplied by t2 , and the linear part is multiplied by t. This
simple remark will be of fundamental importance in the sequel!
Here, we study the minimization problem
MinuRn P (u).

(2.4)

Existence of a minimum point The rst question is: is there a minimum point?
In other words, does there exist umin Rn such that
P (umin ) P (u) u Rn .

(2.5)

The answer to this question is positive!


Let us give an elementary proof of this result:
P is a continuous function of u Rn .
Consider the unit sphere of Rn , denoted by Sn1 . Since Sn1 is a bounded and
closed subset of the nite dimensional vector space Rn , it is actually a compact
set. Now, as Pquad is a continuous function over Sn1 , it takes maximum and
minimum values on Sn1 29 : let smin be such that Pquad (smin ) = MinsSn1 Pquad (s).
In addition, K being positive denite, we know that Pquad (smin ) > 0. Let :=
Pquad (smin ) > 0.
28

We may use the acronyms SP for symmetric positive and SPD for symmetric positive denite.
Unfortunately, the vector space Rn is not bounded, so it is not compact! Therefore, we cannot use
directly this result to prove the existence of a minimum point to P over Rn .
29

29

Consider any u Rn , u = 0: as u/||u|| belongs to Sn1 , we have that Pquad (u/||u||)


, or Pquad (u) ||u||2 . Evidently, this result is also true for u = 0.
With these results, we can now state that P (u) goes to innity when ||u|| goes
to innity. Why is that? According to Cauchy-Schwarz inequality, we know that
(f, u) ||f || ||u||, so it follows that
P (u)

||u||2 ||f || ||u|| u Rn ,


2

which proves the statement, since the right-hand side is a quadratic function of
||u|| with a strictly positive coecient in front of ||u||2 .
Now, consider a given u1 : since P (u) goes to innity when ||u|| goes to innity,
we know that there exists M such that, for all u such that ||u|| > M , we have
P (u) > P (u1 ).
Let us nally consider the closed ball BM := {u Rn | ||u|| M }. With the
previous result, we know that u1 BM , and also that
MinuRn P (u) := MinuBM P (u).
But BM is a closed and bounded subset of Rn , so P takes maximum and minimum
values on BM . This allows us to conclude favorably: there exists umin satisfying
(2.5).
In the above proof, we did not use the fact that the matrix K is symmetric, only that
it is positive denite, so that P (u) goes to innity when ||u|| goes to innity. Note
that, along the way, we have proved the general result below.

Theorem 2.1. Let p : Rn R. Then, if p is continuous over Rn , and if p(u) goes


to innity when ||u|| goes to innity, there exists a unique minimum point umin to p,
solution to
p(umin ) p(u) u Rn .

Characterization of a minimum point Next, let us exhibit some properties of the


minimum point. Instead of writing (2.5), we use the equivalent formulation
P (umin ) P (umin + u) u Rn .
30

(2.6)

What does this condition imply? For all u Rn , there holds


1
(Kumin , umin ) (f, umin )
2

1
(K(umin + u), (umin + u)) (f, (umin + u))
2
K is symmetric 1
1

(Kumin , umin ) + (Kumin , u) + (Ku, u)


2
2
(f, umin ) (f, u)
1
or
0

(Kumin f, u) + (Ku, u).


2
What can one make of this result? Notice that there is a quadratic term in u, and a
linear term in u. So, we use the initial remark we gave at the beginning of the section.
We consider u = t v, with t > 0 and v Rn . It follows that


t
0 t (Kumin f, v) + (Kv, v) .
2

Divide by t, and then let t go to 0+ . One obtains that


0 (Kumin f, v).
In other words, the quadratic term has been removed! Since the previous inequality is
valid for all v Rn , we can use, for any w Rn , v = w and v = w, so we conclude
that
(2.7)
0 = (Kumin f, w) w Rn or Kumin f = 0.
For the time being, we have proved that
umin is a minimum point of P

K is symm.
=
Kumin f = 0.

What of the reciprocal assertion? Consider umin so that (2.7) holds, and let us study
the dierence P (umin + u) P (umin ), for u Rn . We nd easily
P (umin + u) P (umin )

Kis symmetric 1
(Ku, u) + (Kumin f, u)
=
2
1
=
(Ku, u).
2

Now, as K is positive30 , we nd that


P (umin ) P (umin + u) u Rn ,
30

As K is positive denite and not only positive we have the stronger result
P (umin ) < P (umin + u) u Rn ,

so umin is a strict minimum point of P .

31

so umin is a minimum point of P . So, we have proved that


Kumin f = 0

K is SP
=
umin is a minimum point of P .

Finally, we note that since Kumin = f has exactly one solution, the minimum point is
unique (see again30 ). We can write

Theorem 2.2. Let P (u) := 12 (Ku, u)(f, u), for u Rn , where K Rnn is a symmetric
positive denite matrix, and f Rn .
There is a unique minimum point umin to P over Rn .
Furthermore, umin is characterized as the only solution to Kumin = f .
Here, we conclude that u = umin , which means that the set of algebraic equations (2.1)
admits a solution which is identical to the minimum point of the minimization problem
(2.4).
Using derivatives Interestingly, we can recover condition (2.7) by computing the rst
directional derivative of P .
Recall that the directional derivative of a functional p : Rn R at u in the direction
v is dened as
p(u + tv) p(u)
.
p (u).v := lim
t0
t
Remark 2.1. One could write equivalently
p(u + tv) = p(u) + t p (u).v + o(t) t.
Alternately, for a given v, one can consider the real function fv :
fv (t) = p(u + tv).
Then, the directional derivative of p at u in the direction v is simply the derivative of
fv at t = 0:
p (u).v = (fv ) (0).
If p (u) : v
p (u).v is a linear map from Rn to R, then p is dierentiable at u, and
p (u) is called the rst derivative of p at u. Sometimes, it is also denoted by p/u(u).

32

When p (u) = 0, which is equivalent to p (u).v = 0 for all v, one says that u is a
stationary point of P .
As p (u) is a linear map from Rn to R, we may write down its action on v as


p (u).v :=

i=n


gi vi ,

i=1

where (vi )i=1n are the components of the vector v with respect to the orthonormal basis
(ei )i=1n of Rn . The vector of Rn with components (g1 , , gn ) is called the gradient
of p at u: it is denoted by p(u), and there holds p (u).v = (p(u), v) for all v. This
vector is usually written down as

p
(u)
1

x
p

x
(u)
2
,
p(u) =
.

..

p
(u)
xn
p
(u) is called the jth partial derivative of p at u.
and j p(u) := x
j
Why such a name? The reason is simple, if one goes for the coordinates of elements of Rn .

Write u as a combination of elements of the orthonormal basis (ei )i=1n : u = i=n
i=1 xi ei ,
and replace p(u) by the fonction of n variables p(x1 , x2 , xn ). Then, Choosing v = ej ,
the jth vector of the orthonormal basis, one obtains in vector form that

p(u + tej ) = p(u) + t p (u).ej + o(t) = p(u) + t

p
(u) + o(t),
xj

or in coordinates form that


p(x1 , , xj1 , xj +t, xj+1 , , xn ) = p(x1 , , xj1 , xj , xj+1 , , xn )+t

p
(x1 , , xn )+o(t).
xj

In other words, the jth partial derivative of p at u is simply the derivative of x

p(x1 , , xj1 , x, xj+1 , , xn ) at xj .


In the general case, one can prove the assertion:

Theorem 2.3. (Fermats theorem) Let p : Rn R. Then


uext is a local extremum point of p and p (uext ) exists = p (uext ) = 0.

33

Equation p (u) = 0 is sometimes called the Euler equation.


This result means that an extremum point (that is either a minimum point or a maximum
point) of a function dierentiable at this same point is necessarily a stationary point31 .
This notion is local, in the sense that one has to have (case of a maximum point) at
hand > 0 such that p(uext ) p(uext + w), for all w Rn such that ||w|| < . Here,
might be small, be it has to be strictly positive!
The equation p (u) = 0 can be written equivalently p(u) = 0 or, in variational form:
p (u).v = (p(u), v) = 0 v Rn .
The result is easily proven. Indeed, for all v Rn , v = 0, let us dene fv (t) := p(uext +tv):
it is dierentiable at t = 0, and moreover it takes an extremum value at t = 0 over the

, ||v||
[. As a consequence, one gets that (fv ) (0) = 0, or p (uext ).v = 0.
interval ] ||v||
Since this is true for all v, it follows that uext is a stationary point of p.
We can now compute the directional derivative of P given by (2.3). For that, let us
study the dierence:
1
1
(K(u + tv), (u + tv)) (f, u + tv) (Ku, u) + (f, u)
2
2
t2
= t(Ku f, v) + (Kv, v).
2

P (u + tv) P (u) =

Here, we have used K being symmetric. Thus,




t

P (u).v = lim (Ku f, v) + (Kv, v) = (Ku f, v).
t0
2
According to theorem 2.3, if umin is a minimum point of P , we recover condition (2.7).
The vector Ku f is the gradient of P at u: P (u) = Ku f . Under some
assumptions that are fullled here32 , one can then compute the derivative of P at u,
31

The reciprocal assertion is not true! Consider for instance p(x) := x or p(x) := x3 over R.
Basically, the linear map p (u) must not be only dened one direction after the other, but more
appropriately for all directions at once. Namely, p (u) must be a linear map such that
32

p(u + v) = p(u) + p (u).v + o(||v||) v Rn .


This corresponds to the denition of the Frechet derivative, which comes with a uniform estimate in
||v||, independently of the direction. It is thus a stronger denition than the one of the directional
derivative. If in addition p : u
p (u) is continuous over Rn , p is continuously dierentiable. Note
that p can be continuously dierentiable locally, that is in a neighborhood of a given point of Rn .

34

which is the second derivative of P at u:


P  (u).v := lim
t0

P (u + tv) P (u)
= Kv.
t

P  (u) is a linear map from Rn to Rn . Therefore, it can be identied with an n n


matrix, which is called the Hessian of P at u. Evidently, the Hessian of P at u is the
matrix K. Note that it is independent of u in the case of the functional P dened by
(2.3). This is simply a consequence of the fact that P is quadratic in u, so its second
derivative is actually a constant!
2.1.2

Constrained Minimization

Next, let us study the minimization problem with constraint. Consider the functional
dened over Rm ,
1
Q(w) := (C 1 w, w),
(2.8)
2
where C is a symmetric positive denite matrix33 . We recall that K, A and C are related
by K = AT CA.
Here, we study the constrained minimization problem
Minw Q(w)

w subject to AT w = f.

(2.9)

Existence of a minimum point The question is again: is there a minimum point?


We remark that C 1 is symmetric positive denite. Therefore, if the ane set Wconst :=
{w Rm : AT w = f } is not empty, we infer from the previous subsection 2.1.1,
paragraph existence of a minimum point, that there exists a single minimum. Indeed,
we have to solve a unsconstrained minimization problem with a positive denite form,
namely
MinwWconstr Q(w).
But, we already know from chapter 1 that this set is not empty (set w = CAu ), so
existence and uniqueness of minimum point to (2.9) follows.
33

Check that a matrix C is SPD if, and only if, its inverse C 1 is SPD.

35

Characterization of a minimum point Next, let us nd more about the minimum


point. For that, it is helpful to start with the minimum point umin of problem (2.4).
According to theorem 2.2, umin = u is such that Kumin = f , so we dene the candidate
wcand = CAumin : wcand fullls the constraint Awcand = f . Now, let us consider another
w Wconstr . We nd that the dierence w wcand belongs to the kernel of AT : AT (w
wcand ) = 0.
We recall that there holds, according to the rank-nullity theorem34
dim(Ker(AT )) = m dim(Rg(AT )) = m (n dim(Ker(A))) = m n,
since we have assumed from the start that Ker(A) = {0}.
We note that if Ker(AT ) = {0}, then w = wcand so there is exactly one vector of Rm that
fullls the constraint: in this case, Wconstr = {wcand }, and the problem with constraint
is trivial! In this situation, we have n = m, and it follows that A is invertible.
If Ker(AT ) is not reduced to {0}, that is when m > n, the problem is much more
interesting to study!
Since w wcand Ker(AT ), we can write w = wcand + k, for some k Ker(AT ).
Actually, one has the characterization
Wconstr = wcand + Ker(AT ) := {w Rm : k Ker(AT ), w = wcand + k}.
Then, we nd, we all w = wcand + k elements of Wconstr , that
Q(w)

1 1
(C (wcand + k), (wcand + k))
2
1
symm.
=
Q(wcand ) + (C 1 wcand , k) + (C 1 k, k)
2
1 1
=
Q(wcand ) + (C k, k),
2
=

C 1

where for the last equality we used the denition of wcand = CAumin :
(C 1 wcand , k) = (Aumin , k) = (umin , AT k) = 0 as k Ker(AT ).
But, we know that C 1 is positive denite, so for all k = 0, one has (C 1 k, k) > 0, and
it follows that wmin = CAumin is a (strict) minimum point of Q(w), w subject to the
constraint AT w = f .
We can write
34

We use also the result Rg(AT ) = (Ker(A)) , which is a direct consequence of the rank nullity
theorem applied to A and AT , and of the fact that Rg(A) (Ker(AT )) and Rg(AT ) (Ker(A)) .

36

Theorem 2.4. Let Q(w) := 12 (C 1 w, w), for w Rm , where C Rmm is a symmetric


positive denite matrix.
Let A Rmn , with Ker(A) = {0}, and f Rn .
There is a unique minimum point wmin to the constrained problem
Minw Q(w)

w subject to AT w = f.

Furthermore, wmin is characterized as the only solution to C 1 wmin = Aumin , where umin
is the unique solution to AT CAumin = f .
Here, we recall that u = umin , so one has also w = wmin , and we conclude that the set
of algebraic equations (2.2) admits a solution which is identical to the minimum points
of the minimization problems (2.4) and (2.9).
 Physical viewpoint. From the previous physical viewpoints, we know on the one
hand that P (u) is the total potential energy of the spring-mass system, made up of
two parts: the potential energy stored in the spring due to the displacement u, that is
1
(Ku, u), and (f, u) the gravitational energy of the masses (with the same displacement
2
u).
On the other hand, we know that Q(w) is the potential energy stored in the springs
due to the stress w, with constraint AT w = f representing the force balance between
internal and exterior forces.
The actual physical conguration corresponds to the equilibrium of the spring-mass
system. It corresponds to the minimum of P (u), with minimum point u = umin and to
the minimum of Q(w) with constraint, with minimum point w = wmin . We note that
we have wmin = CAumin , which yields
1
1
1
Q(wmin ) = (C 1 wmin , wmin ) = (umin , AT CAumin ) = (umin , Kumin ),
2
2
2
which means that both expressions of the potential energy stored in the springs are
identical at equilibrium, as expected!

Interestingly, we have a remarkable identity for the minimum points of P (unconstrained problem) and Q (constrained problem). One the one hand, we have that
Kumin = f , so
1
1
P (umin ) = (Kumin , umin ) (f, umin ) = (Kumin , umin ).
2
2
37

On the other hand, we have (see previous physical viewpoint): Q(wmin ) = 12 (Kumin , umin ),
so it turns out that
(2.10)
P (umin ) + Q(wmin ) = 0.
2.1.3

Lagrange formulation

To solve the constrained optimization problem, let us turn our attention to a technique of
fundamental importance, which relies on the notion of Lagrangian and Lagrange multipliers. It converts a constrained optimization problem to an unconstrained optimization
problem, that is a problem without constraint. Consider for instance
Min(x,y) f (x, y)

(x, y) subject to g(x, y) = 0 .

(2.11)

This minimization problem with constraint can be recast as a problem without constraint. For that, consider the Lagrangian
L(x, y, ) := f (x, y) + g(x, y).

(2.12)

Above, is called the Lagrange multiplier associated to the constraint g(x, y) = 0. It can
be proven that any local minimum point of (2.11), together with some ad hoc Lagrange
multiplier, is automatically a stationary point of (2.12). In general, the reciprocal assertion is not true: a stationary point of the Lagrangian is not automatically an extremum
point of the constrained problem. Under certain additional assumptions however, the
reciprocal assertion will hold (see theorem 2.7 below).

Theorem 2.5. Consider the constrained minimization problem in R2


Min(x,y) f (x, y)

(x, y) subject to g(x, y) = 0

and dene the associated Lagrangian over R3 :


L(x, y, ) := f (x, y) + g(x, y).
Then, if (x+ , y+ ) R2 is a local minimum point of the constrained problem, and if
g(x+ , y+ ) = 0, there exists + R such that (x+ , y+ , + ) is a stationary point of L:
L
L
L
(x+ , y+ , + ) =
(x+ , y+ , + ) =
(x+ , y+ , + ) = 0.
x
y

38

Let us explain the meaning of this result, before giving its proof.
The last equation L/(x+ , y+ , + ) = 0 simply says that the constraint is fullled at
the minimum point: g(x+ , y+ ) = 0!
On the other hand, rewriting the rst two means that
f
g
(x+ , y+ ) + + (x+ , y+ ) = 0
x
x

g
f
(x+ , y+ ) + + (x+ , y+ ) = 0.
y
y

But the gradient of f and the gradient of g are respectively given by






f
g
(x,
y)
(x,
y)
x
x
, g(x, y) =
,
f (x, y) =
f
g
(x,
y)
(x, y)
y
y
so this can be expressed equivalently as
f (x+ , y+ ) + + g(x+ , y+ ) = 0.
In other words, the gradient of f is parallel to the gradient of g at (x+ , y+ ), with the
Lagrange multiplier playing the role of the multiplicative constant between the two.
Proof. (of theorem 2.5)  As noted above, L/(x+ , y+ , ) = 0 holds for all R.
Let us now consider that both functions f and g are Frechet-dierentiable, and continuously dierentiable (see footnote 32 ), in a neighborhood of the local minimum point
(x+ , y+ ), and for instance that g/y(x+ , y+ ) = 035 .
Then, according to the implicit function theorem, there exists a neighborhood of (x+ , y+ ),
call it V , there exist , R, there exists a function , continuously dierentiable over
], [ such that
{(x, u) V : g(x, y) = 0} = {(x, (x)) : x ], [} .
The function x
g(x, (x)) is thus equal to 0 over ], [. Using the chain rule for
x
g(x, (x)), one nds that the derivative of at x+ is given by:
g
(x+ , y+ )
d
g
g
0=
g(x+ , (x+ )) =
(x+ , y+ ) +
(x+ , y+ )  (x+ ), or  (x+ ) = x
.
g
dx
x
y
(x , y )
y + +
Recall that we assumed that y g(x+ , y+ ) = 0. Above, we replaced (x+ ) by y+ , which
follows from the denition of . With this new expression of the constraint, one can
35

Otherwise, we would have that g/x(x+ , y+ ) = 0, since g(x+ , y+ ) = 0, and we would use the
implicit function theorem replacing x by y.

39

replace the minimization problem with constraint, locally in V , by the unconstrained


local problem below:
Minx],[ f (x, (x)).
Since (x+ , y+ ) is a local minimum point of the original constrained problem, x+ is a
local minimum point of the unconstrained local problem. Furthermore, x
f (x, (x))
is dierentiable at x+ , so thanks to Fermats theorem, we nd that x+ is a stationary
point of this function. According once more to the chain rule:
0=

d
f
f
f (x+ , (x+ )) =
(x+ , y+ ) +
(x+ , y+ )  (x+ ).
dx
x
y

Using the value of  (x+ ) given by the implicit function theorem, we nd that
f
g
f
g
(x+ , y+ ) (x+ , y+ )
(x+ , y+ ) (x+ , y+ ) = 0.
x
y
y
x
This simply means that






f
(x+ , y+ )
x
f
(x+ , y+ )
y

g
(x+ , y+ )
x
g
(x+ , y+ )
y




 = 0.


Or, that f (x+ , y+ ) is parallel to g(x+ , y+ ): but the latter one is not equal to zero,
so there exists a real number, called + (to be coherent!) such that
f (x+ , y+ ) + + g(x+ , y+ ) = 0.
Thus, one concludes that (x+ , y+ , + ) is a stationary point of the Lagrangian.
Remark 2.2. Obviously, one derives the same result for a maximization problem.
The theorem 2.5 can be generalized, and in fact it can be applied to problems like
Minw f (w)

w subject to gi (w) = 0, 1 i n ,

provided that some local inversion property holds at the minimum point.

40

Theorem 2.6. Consider the constrained minimization problem in Rm


Minw f (w)

w subject to gi (w) = 0, 1 i n ,

and dene the associated Lagrangian over Rm Rp :


L(w, ) := f (w) +

i=n


i gi (w).

i=1

Then, if w+ Rm is a local minimum point of the constrained problem, and if the family
(gi (w+ ))i=1, ,n is linearly independent, there exists + Rn such that (w+ , + ) is a
stationary point of L:
L
(w+ , + ) = 0,
w

L
(w+ , + ) = 0.

So, if we consider a function of three variables with two independent constraints, say
Minf (x, y, z)

(x, y, z) subject to g1 (x, y, z) = 0, g2 (x, y, z) = 0,

then we simply introduce two Lagrange multipliers: namely, we dene


L(x, y, z, 1 , 2 ) := f (x, y, z) + 1 g1 (x, y, z) + 2 g2 (x, y, z)
and look for a stationary solution of L:
L
L
L
L
L
=
=
=
=
= 0 at (xst , yst , zst , 1,st , 2,st ).
x
y
z
1
2
The rst three equations mean that f (xst , yst , zst ) is a linear combination of g1 and
g2 , ie. f = 1,st g1 + 2,st g2 , at (xst , yst , zst ). On the other hand, the last two
equations simply mean that the constraints g1 (xst , yst , zst ) = 0 and g2 (xst , yst , zst ) = 0
are fullled.
Remark 2.3. Note that, in the previous example, the original constrained problem is
set in R3 , whereas the Lagrangian is dened over R5 . So, the trade-o for having an
unconstrained problem (nding the stationary points of the Lagrangian) is that it comes
with an increase in the dimension of the problem to be solved!
Going back to the spring-mass system, we expressed it as a minimization problem
with constraint, see (2.9). Since the constraint equation f AT w = 0 is written in Rn ,
41

there should be n Lagrange multipliers, say, 1 , , n . Or, in short, we can consider a


single vector Lagrange multiplier, Rn . In the same spirit as before, we consider the
Lagrangian
1
L(w, ) := Q(w) + (f AT w, )Rn = (C 1 w, w) (, AT w f )Rn .
2

(2.13)

A stationary point for this unconstrained problem is given by36 as


L
(wst , st ) = C 1 wst Ast = 0 ,
w
L
(wst , st ) = AT wst + fst = 0 .

In variational form, the equations are written equivalently


L
(wst , st ).w = (C 1 wst Ast , w ) = 0 w Rm ,
w
L
(wst , st ). = (AT wst + fst ,  ) = 0  Rn .

We note that the stationary point (wst , st ) precisely solves the algebraic set of equations
(2.2), whose only solution is (w , u ) = (wmin , umin ) (see the rst paragraph after theorem 2.4). In other words, the components of u = umin appear as Lagrange multipliers.
So, we conclude that the stationary point (wst , st ) is equal to the minimum point of
the constrained problem (2.9), which is precisely the result we were looking for! Let us
comment this result a bit: the equivalence between minimum point of the constrained
problem and stationary point of the Lagrangian is due to some additional properties
satised by the functional Q (it is quadratic and positive denite), and in the way the
constraint is dened (it is linear)... We summarize all this below.

36

As previously, one can dene the directional derivative of L at (w, ) with respect to w, and with
respect to :
L(w + tw , ) L(w, )
L
(w, ).w := lim
,
t0
w
t

L
L(w, + t ) L(w, )
(w, ). := lim
.
t0

Check the results...

42

Theorem 2.7. Let Q(w) := 12 (C 1 w, w), for w Rm , where C Rmm is a symmetric


positive denite matrix. Let A Rmn , with Ker(A) = {0}, and f Rn .
Then, nding the minimum point wmin to the constrained problem
Minw Q(w)

w subject to AT w = f,

is equivalent to nding the stationary point (wst , st ) of the Lagrangian


L(w, ) := Q(w) + (f AT w, )Rn ,
which satises the equation
L
(wst , st ) = 0
w

L
(wst , st ) = 0 ,

and one has wmin = wst .


 Physical viewpoint. The concept of Lagrangian is also of great physical signicance.
It led to a reformulation of classical mechanics, now known as the Lagrangian mechanics.
However, even though the vocabulary is similar, the mathematical Lagrangian and the
physical Lagrangian are two dierent concepts. As a matter of fact, when all forces
derive from a potential, there is no multiplier in the physical Lagrangian!

2.1.4

Primal and dual problems

Our next goal is to introduce the so-called primal and dual problems for the spring-mass
system, seen as a constrained minimization problem in the variable w, that is (2.9).
Intuitively, the primal problem is concerned with the variable in which the problem is
expressed (w), whereas the dual problem is concerned with the Lagrange multiplier ().
Recall that the functional Q is dened over Rm , Q(w) := 12 (C 1 w, w), where C is an
m m SPD matrix. If we let A Rmn , with Ker(A) = {0}, and f Rn , then the
constraint writes AT w = f .
Next, recall that we saw that we could identify the Lagrange multiplier with u. So, if
we let K := AT CA, an n n (SPD by construction) matrix, the functional associated
to the Lagrange multiplier of Rn will be P () := 12 (K, ) (f, ).

43

In other words, we call the minimization problem (2.9) the primal problem:
Primal Problem:

Minw Q(w)

w subject to AT w = f.

(2.14)

We then introduce the Lagrangian, dened over Rm Rn :


L(w, ) := Q(w) + (f AT w, ).
Finally, the minimization problem (2.4), usually recast as a maximization problem (of
(P )!), is called the dual problem:
Dual Problem:

MaxRn (P )().

(2.15)

We shall establish the following equivalence table:

(1) MaxRn (P ())

(2) Minw subject to AT w=f Q(w)

(3) MaxRn MinwRm L(w, ) (4) MinwRm MaxRn L(w, )

So far, we have established the equivalences (1) (2) (see theorem 2.4). Moreover,
we have the two relations
The maximum point of (P ) (max = umin ) is related to the minimum point wmin
of Q, w subject to AT w = f , by C 1 wmin = Amax .
MaxRn (P ()) = Minw subject to AT w=f Q(w) (see the equality (2.10)).
Next, we show that (2) (4). In fact,
MaxRn L(w, ) = MaxRn


Q(w) (, A w f ) =
T


Q(w)
+

if AT w f = 0
,
if AT w f = 0

Thus, solving (4) is equivalent to solving (2).


To conclude, let us show (1) (3). For a given , we solve the unconstrained
minimization problem with the auxiliary functional d(w) := L(w, ) = 21 (C 1 w, w) +
(f AT w, ):
MinwRm d(w).
44

Since C 1 is SPD, we know from theorem 2.2 that there exists one, and only one,
minimum point w()

to this minimization problem. Furthermore, it is characterized by


the Euler equation
d (w())

= 0,
which writes C 1 w()

A = 0, or w()

= CA. We plug this into the expression of


the Lagrangian to get
1
(CA, A) (, AT CA) + (, f )
2
1
= (AT CA, ) + (, f )
2
= (P )().

L(w(),

) =

Thus,
MinwRm L(w, ) = (P )(),
which shows that (1) (3).
Remark 2.4.  With this last result, it appears that, given the primal problem, we could
have rst introduced the Lagrangian, and second we could have dened the functional
(P ) that appears in the dual problem, with the formula
(P )() = MinwRm L(w, ) Rn .
In this way, the introduction of the dual problem is completely natural, contrary to the
somewhat articial denition given at (2.15), where we had to identify the Lagrange
multiplier to the displacement u...
The equivalence between the primal and dual problems can be nally stated.

Theorem 2.8. Let C Rmm be a symmetric positive denite matrix. Let A Rmn ,
with Ker(A) = {0}, and f Rn .
Then, the primal problem (2.14) and the dual problem (2.15) are equivalent in the following
sense: if wprimal and dual are respectively the solutions of (2.14), (2.15), then they are
related through
C 1 wprimal = Adual .
Moreover, the value of the Lagrangian L at (wprimal , dual ) is equal to
MaxRn MinwRm L(w, ) = MinwRm MaxRn L(w, ).

45

(2.16)

The previous theorem is a special case of the MiniMax theorem of Von Neumann. The
important result (apart from the existence of the minimum and maximum points!) is
the equality (2.16) (see subsection 2.1.5 for more details).
2.1.5

Saddle points

Consider a general Lagrangian  dened over Rm Rn , corresponding to the constrained


minimization of a functional q(w), with n independent constraints, gi (w) = 0, i =
1, , n, over Rm :
(w, ) := q(w) +

i=n


gi (w)i , or (w, ) := q(w) + (g(w), )Rn .

i=1

Then, we always have the inequality


Max Minw (w, ) Minw Max (w, ).

(2.17)

and w,
To prove this simple result, note that we have, for any
Max (w, ).
(w, )
Taking the minimum over w Rm gives
Minw Max (w, ).
Minw (w, )
Rn gives now
The right-hand side is a xed number. Taking the maximum over
Minw Max (w, ).
Max Minw (w, )
So, with the Lagrangian L of the spring-mass system, we have a stronger result (see
theorem 2.8). To explain this, let us introduce the notion of a saddle point.
The couple (wsp , sp ) Rm Rn is said to be a saddle point of the Lagrangian  if
wsp is a minimum point of (, sp ) : w
(w, sp ), over Rm ;
sp is a maximum point of (wsp , ) :
(wsp , ), over Rn .
Or, in other words, (wsp , sp ) is such that
MinwRm (w, sp ) = (wsp , sp ) = MaxRn (wsp , ).
Why is the notion of saddle point so interesting? First, in the case of the spring-mass
system, we have the result below.
46

!#
"#
#
!"#
!!#
!%#
!$#
!
"

!
"

!"
!!

!"
!!

Figure 2.1: Example of a saddle-point.

Theorem 2.9. Let C Rmm be a symmetric positive denite matrix. Let A Rmn ,
with Ker(A) = {0}, and f Rn .
Then, if wprimal and dual are respectively the solutions of (2.14), (2.15), (wprimal , dual )
is a saddle point of the Lagrangian L.
Proof. Introduce the auxiliary functional ddual (w) = L(w, dual ). As we saw in subsection
2.1.4, ddual reaches its minimum at w(
dual ) = CAdual , that is at wprimal . Therefore,
wprimal is a minimum point of L(, dual ) over Rm .
Now, as f AT wprimal = 0, we have that L(wprimal , ) = Q(wprimal ), for all Rn . In
other words, L(wprimal , ) is a constant functional over Rn , so it takes its minimum at
any : it is in particular true at dual !
 Second, going back to the general case, one can check that, if there exists a saddle
point of , then37
Max Minw (w, ) = Minw Max (w, ).
(2.18)
We already know that inequality (2.17) holds, so we just have to show the converse
inequality. For that, we (obviously!) have to use the fact that there is a saddle point of
37

Compare (2.17) to (2.18).

47

, which we call (wsp , sp ):


def. of a minimum over Rm :
def. of the saddle point (wsp , sp ):
def. of a maximum over Rn :

Minw (Max (w, )) Max (wsp , )


= (wsp , sp ) = Minw (w, sp )
Max (Minw (w, )) .

Finally, we conclude by the result below...

Theorem 2.10. If (wsp , sp ) is a saddle point of , then wsp is a minimum point of the
constrained problem
MinwRm q(w) w subject to g(w) = 0.
Proof. By denition, sp is a maximum point of

(wsp , ) = q(wsp ) + (g(wsp ), ).
But, if g(wsp ) = 0, it follows that the maximum is + (take = t g(wsp ), with t +).
Therefore, g(wsp ) = 0, ie. wsp fullls the constraints.
Again by denition, wsp is a minimum point of w
(w, sp ): in other words,
q(wsp ) = (wsp , sp ) (w, sp ) := q(w) + (g(w), sp ) w Rm .
In particular, for all w subject to the constraints g(w) = 0, it follows that q(wsp ) q(w).
This concludes the proof.
2.1.6

Practical corner

Let us now summarize the results we have obtained so far, for the minimization over
nite dimensional vector spaces, such as Rn ... We rewrite the minimization problem in
the following abstract way. Consider K a subset of Rn , and the problem
Find umin K such that q(umin ) q(u) u K .

(2.19)

Maximization? The rst remark is that the results also apply to maximization problems: indeed if one changes the functional to be maximized to minus the same functional,
the maximization problem becomes a minimization problem!

48

First already known result: If K is a closed and bounded subset of Rn and if q


is continuous over K, then q takes minimum and maximum values on K. In particular,
problem (2.19) has a solution.
Unconstrained case: We consider here minimization problems over Rn without constraints. In other words, we have that K = Rn .
Is there a minimum point?
Negative answers: Before trying to nd minimum points, one has to check
whether there is a minimum, such as for instance q1 (u) = u3 in R, for which
limu q(u) = . Or, it can happen that there is a minimum, but no
minimum point. In this case, one speaks of the inmum. See for instance
q2 (u) = (||u||2 +1)1 in Rn , for which q(u) > 0 for all u, but lim||u||+ q(u) =
0. One usually writes that inf uR q(u) = 0...
Positive answers: If q(u) goes to innity when ||u|| goes to innity and if q
is continuous over Rn , then there exists a minimum point (see theorem 2.1).
Consider for instance the case of q3 (u) = exp(||u||) over Rn .
How to nd umin ?
If q is dierentiable: According to theorem 2.3, if q is dierentiable at umin ,
then q  (umin ) = 0 (this is Euler equation).
Otherwise: write down that q(umin ) q(umin + t v), for all small t, and
for all v Rn , and gather information on the possible minimum points umin
(this is what we did several times in chapter 2...)
Check whether the possible choices are minimum points: For instance, consider the interesting case of q4 (u) = |u| over R. We know by applying theorem 2.1 that there exists a minimum point. We also know that, for u < 0,
q  (u) = 1, whereas, for u > 0, q  (u) = +1. So, no point of ] , 0[]0, +[
satises Euler equation. But there its a minimum point, so it can only be
umin = 0!
Constrained case: We consider here minimization problems over Rn with constraints.
In other words, we have that K = Rn . Indeed, we want to nd the minimum of q(u),

49

with u subject to p independent constraints gi (u) = 0, for 1 i p (here, p n). We


just write
K := {u Rn : gi (u) = 0, 1 i p},
to match the two formulations.
Is it possible to describe K? We remark that, if the functionals (gi )i=1, ,p are
continuous, then K is automatically a closed subset of Rn (it is the inverse image
of (0, , 0)).
If one can establish in addition that K is bounded, one recovers existence of a
minimum point (and of a maximum point) provided that q is continuous over the
closed and bounded subset K of Rn .
Or, if q(u) goes to innity when ||u|| goes to innity and if q is continuous over
K, there exists a minimum point.
Is it possible to reformulate as an unconstrained problem?
In some
instances, one can rewrite the minimization problem with constraints as an unconstrained problem. For instance, if one considers the minimization of q5 : R2 R
dened by q5 (u) = 3(u1 )4 + 2(u2 )2 subject to 2u1 + u2 = 4, one notices that the
subset K of R2 made up of elements that fulll the constraint is
K = {u R2 : t R, u1 = t, u2 = 4 2t},
and that the minimization problem with constraint rewrites
MintR q 5 (t)

with q 5 (t) = q5 ((t, 4 2t)).

Then, one can use the results provided in the unconstrained case.
How to nd umin ? If it is not possible to reformulate as an unconstrained
problem, then one introduces the Lagrangian.
Lagrangian: It is equal to
L(u, ) := q(u) +

i=p

i=1

and dened over Rn Rp .

50

i ci (u) = q(u) + (, c(u))Rp ,

If q and (ci )i=1, ,p are dierentiable: According to theorem 2.6, if q and


(ci )i=1, ,p are dierentiable, and if the gradients (ci (umin ))i=1, ,p are linearly
independent, then there exists + such that (umin , + ) is a stationary point
of L:
L
L
(umin , + ) = 0,
(umin , + ) = 0.
u

Otherwise: write down explicitly the condition that umin is a minimum point
of the problem with constraint...
Check whether the possible choices are minimum points.
Remark 2.5. For additional results, see the monograph Introduction to Numerical Linear Algebra and Optimisation, by Philippe Ciarlet, Cambridge Texts in Applied Mathematics (1989).

2.2

Variational formulations for dierential equations

In chapter 1, we formulated the elastic bar model as a dierential equation, supplemented


with boundary conditions. We were able solve it analytically, and even sometimes we derived an explicit formula for simple cases (see 1.2). However, this analytic method does
not work in general: not for higher dimensional problems (starting with two-dimensional
problems), and not even for slightly more general problems in one dimension, such as
the Sturm-Liouville dierential equation:


d
du

c
(x) + q(x)u(x) = f (x), 0 < x < 1
(2.20)
dx
dx
with boundary conditions

u|x=0 = 0,

du
c
dx


= 0.

(2.21)

|x=1

Above, we are looking for the solution function u(x), whereas the functions c(x) and
q(x) are given by the model. Finally, f (x) is the data. In the case of the elastic bar
model:
c(x) appears in Hookes law, whereas q = 0 (these values are prescribed by the
choice of the model: here, the elastic bar model);
f (x) corresponds to the (external) gravitational forces, and it may vary from one
bar to the other, depending on its mass density;
51

and nally u(x) is the resulting displacement of the bar with respect to the reference
conguration.
 Physical viewpoint. The Sturm-Liouville system, with homogeneous or non-homogeneous
boundary conditions have many physical applications. For instance, for some appropriately chosen sets of values of the functions c(x) and q(x):
1. in quantum theory, equation (2.20) is called the one-dimensional Schr
odinger equation;
2. for modeling the oscillations of a drum, or the propagation of electromagnetic
waves, equation (2.20) is called the Bessel equation.

One important method to study the properties of the solution(s) to the boundary value
problem (2.20)-(2.21) is to use an integral form, often called the variational formulation.
We shall discuss how to derive the variational formulation for the second order differential equation (2.20), with boundary conditions (2.21). The same methodology can
be applied to any other second order dierential equations. The derivation is standard
and simple. To do so, we multiply both sides of equation (2.20) by an arbitrary test
function v satisfying v |x=0 = 0 to obtain


du
d
c
v + quv = f v ,

dx
dx
then integrating over ]0, 1[ gives



 1
 1
du
d
c
v + quv dx =

f v dx.
dx
dx
0
0

(2.22)

Now by integration by parts and according to the boundary condition prescribed for the
solution u at x = 1 (see (2.21)), and to the boundary condition on the test function at
x = 0 (v |x=0 = 0), we have

 1
 1
du dv
c
+ quv dx =
f v dx.
dx dx
0
0
This leads to the variational formulation for the boundary value problem (2.20)-(2.21):
Find u with u|x=0 = 0 such that a(u, v) = g(v) v with v |x=0 = 0
52

(2.23)

where a(, ) and g() are forms respectively given by



 1
du dv
+ quv dx ,
a(u, v) =
c
dx dx
0
 1
g(v) =
f v dx .

(2.24)
(2.25)

Remark 2.6. The advantage of this formulation is that it involves only rst order
derivatives of u, not second order derivatives as in the original dierential equation
formulation (2.20). Thus, the regularity requirement on the solution u is weakened.
This is the reason why such variational formulations are also called weak formulations.
Also, one can choose u as the test function (that is v = u) since the solution u fullls
the boundary condition at x = 0. This last property is not always veried, when one
builds variational formulations (see 2.2.3 below).
2.2.1

A study of the variational formulation

Evidently, g() is a linear form. One can check that a(, ) is a bilinear form, that is
linear with respect to each variable, and that it is symmetric, ie. for all u and v,
a(u, v) = a(v, u) .
Furthermore, we can show that a(, ) is also strictly positive, ie.,
a(v, v) > 0 v = 0 with v |x=0 = 0 ,
under the conditions that

c0 > 0 such that c(x) c0
q(x) 0 x ]0, 1[.

x ]0, 1[ ;

Indeed, consider any v (with v |x=0 = 0), then




a(v, v) =
0

dv
dx

2


dx +

2
dv
c
dx

dx
0
 1  2
dv
dx
c0
dx
0
0,


53

qv 2 dx

(2.26)

according to the assumptions on c(x) and q(x). Now, what are the functions v (with
v |x=0 ) and such that a(v, v) = 0? From the above, it follows that
 1  2
dv
dv
= 0.
0=
dx or
dx
dx
0
Therefore, v is a constant function. Because of the homogeneous Dirichlet boundary
condition at x = 0, we nd that v = 0.
Then, there is at most one solution to the variational formulation (2.23)-(2.25) under
the conditions (2.26). Indeed, assuming there are two solutions, say u1 and u2 , one nds
that a(u1 u2 , v) = 0, for all v with v |x=0 = 0. So using v = u1 u2 (which is possible
since (u1 u2 )|x=0 = 0) leads to
a(u1 u2 , u1 u2 ) = 0 ,
and it follows that u1 = u2 , which proves the uniqueness of the solution to the variational formulation.
Concerning the existence of a solution u to the variational formulation (2.23)-(2.25),
the mathematical theory is too complicated to be exposed in every detail here... However,
let us discuss this topic, and give a few pointers below. As we remarked earlier the
solution u, and the test functions v, belong a priori to the same functional space.
Indeed, they satisfy the same boundary condition, and also the forms a(, ) and () both
involve the functions, and their rst derivatives. Let us call it V from now on.
Assume that V is a nite dimensional vector space. Call n its dimension, and let
(vi (x))i=1, ,n be a basis of V . Then, by linearity with the second argument of a(, ),
one can prove that38 that the variational formulation (2.23)-(2.25) is equivalent to
Find u V such that a(u, vi ) = g(vi ) 1 i n ,

(2.27)

with forms a(, ) and g() still dened by (2.24)-(2.25).


j=n
Next, if we express u as u(x) = j=1
Uj vj (x), then (2.27) with (2.24)-(2.25) is
38
equivalent to

 =G
,
 Rn such that AU
Find U
(2.28)
nn 
n
where A R , G R , Aij = a(vj , vi ), Gi = g(vi ) for 1 i, j n,
 Rn is the vector whose components are U1 , , Un .
and U
38

Prove this result, or go to 3.2.

54

To prove existence of a solution, we have now to check that (2.28) has one solution.
Since A is a square matrix, it is enough to verify that the kernel of A is reduced to
{0}. We prove a stronger result below, namely that A is a positive denite matrix.
Given V Rn , one nds that38
(AV , V )Rn = a(v, v),

where v(x) =

j=n


Vj vj (x) .

j=1

Since we know that a(, ) is a strictly positive form, we have that (AV , V )Rn > 0 as
soon as V = 0. So, the matrix A is positive denite, and thus it is invertible. Exis follows, which implies in turn the existence of the solution u to the origtence of U
j=n
inal variational formulation (2.23)-(2.25), with expression u(x) = j=1
Uj vj (x).
If the functional space V is nite dimensional, alls well that ends well! Unfortunately, this is not the case. As a matter of fact, it turns out that V is an
innite dimensional vector space.
 To prove this result, let us give an intuitive proof, by building linearly independent families of elements of V of arbitrary size. So, for any k N , we would like
to dene k independent functions of V : the simplest way to do it is to consider
k functions with disjoint supports39 . Let us introduce the function w(x) dened
over R:

cos(x) 1 for x [0, 2]
w(x) :=
.
0
for x
/ [0, 2]
It is clear that w(x) is continuously dierentiable over R. In terms of regularity,
this is enough for our purpose, since we deal with at most rst order derivatives
in the variational formulation. And, obviously, its support is [0, 2].
We are interested in dening functions that are dened over [0, 1], with 0 value
at x = 0. So, an idea is to localize the support of the functions we build. Let us
introduce, for  = 1, , k, the k functions dened over [0, 1] by


4 1
4 + 1
,
.
v (x) = w((4k + 1)x 4 + 1) with support
4(k + 1) 4(k + 1)
The supports of those k functions are disjoint, and evidently they are all continuously dierentiable, with value 0 at x = 0. As a result, the family (v )1k of
elements of V is linearly independent, for any value of k.
39

The support of a function v(x) is the closure of the complement of the set {x : v(x) = 0}.

55

 In the case of an innite dimensional functional space V , the fact that a(, ) is
positive denite is not enough. In fact, one needs a stronger assumption, namely
that
> 0 such that a(v, v) ||v||2V v V ,
where || ||V is a norm of the space V . This property is called coerciveness of
the form a(, ). It is interesting here to note that we have to provide a measure
for elements of V , with respect to some norm || ||V : this is the main dierence
between the nite dimensional case, where the measure is automatic40 , and the
innite dimensional case... Then, assuming nally continuity of the forms, that is

Ma > 0 such that |a(v, w)| Ma ||v||V ||w||V v, w V
,
Mg > 0 such that |g(v)| Mg ||v||V v V
one concludes that there is a solution u V to the variational formulation (2.23)(2.25). This result is the Lax-Milgram theorem. However, to prove those properties,
one would have to dene precisely the functional space V , which is beyond the
scope of these Lecture Notes...
From now on, we assume existence of the solution u the variational formulation.
Next, we prove an equivalence result between the boundary value problem and the
variational formulation.

Theorem 2.11. The boundary value problem (2.20)-(2.21) is equivalent to the variational
formulation (2.23)-(2.25).
Proof. First, we know already that the solution u(x) to the boundary value problem
(2.20)-(2.21) is also a solution to the variational equation (2.23)-(2.25). Next, we will
verify that any solution u to (2.23)-(2.25) is also a solution to the boundary value problem
(2.20)-(2.21).
In fact, since u satises (2.23)-(2.25), we have

 1
 1
du dv
c
+ quv dx =
f v dx v with v |x=0 = 0 .
dx dx
0
0
 by taking ||V
 ||Rn . So if V is nite dimensional, one takes
In Rn , one can always measure a vector V
 ||Rn , with the bijective correspondance v(x) = j=n Vj vj (x).
simply ||v||V = ||V
j=1
40

56

Using integration by parts, we obtain






x=1  1
 1
du
du
d
c
v + quv dx + c v

=
f v dx v with v |x=0 = 0 . (2.29)
dx
dx
dx x=0
0
0
As the test function v is arbitrary, we can take v to be arbitrary and satisfying the
homogeneous Dirichlet boundary conditions v |x=0 = v |x=1 = 0, then (2.29) becomes



 1
du
d
c
+ qu f v dx = 0 v with v |x=0 = v |x=1 = 0 .

dx
dx
0
Using some powerful result from the theory of Lebesgue integration, one can prove that
this implies41


du
d
c
(x) + q(x)u(x) = f (x), 0 < x < 1.
(2.30)

dx
dx
Plugging this new information into (2.29), it nows reduces to


du
c
|x=1 v |x=1 = 0 v with v |x=0 = 0 ,
dx
which indicates42 that u also satises the condition


du
c
=0.
dx |x=1

(2.31)

The dierential equation (2.30), together with the boundary condition (2.31) tell us that
u is a solution to the boundary value problem (2.20)-(2.21), if one remembers that, by
denition, the solution u to the variational formulation also veries u|x=0 = 0.
2.2.2

Minimum principle for dierential equations

Now we investigate the relation between the boundary value problem (2.20)-(2.21) and
the following (potential energy) functional

   2
 1
1 1
dv
2
+ qv dx
f v dx .
P (v) :=
c
2 0
dx
0
We are going to prove the following results below.
41

In other words, we have enough test functions v at hand, even when we impose the homogeneous
boundary conditions at x = 0 and x = 1... Having enough functions means that the result we refer to
is related to some density property in the functional space V .
42
Propose some test function v...

57

Theorem 2.12. The function u that minimizes P (v) over all v with v |x=0 = 0 is the
solution to the boundary value problem (2.20)-(2.21). In other words, it satises the
dierential equation


du
d
c
(x) + q(x)u(x) = f (x), 0 < x < 1

dx
dx
with the boundary conditions

u|x=0 = 0 and

du
c
dx


|x=1

=0.

The converse implication is also true under the conditions (2.26).


Proof. Note that we can rewrite equivalently P (v) as
1
P (v) = a(v, v) g(v) ,
2
with a(, ) (resp. g()) the symmetric bilinear form (resp. the linear form) of (2.24)(2.25).
Note also that, according to theorem 2.11, we already know that the boundary value
problem (2.20)-(2.21) is equivalent to the variational formulation (2.23)-(2.25). So, we
are now going to prove that solving the minimization problem is equivalent to solving
the variational formulation (under the conditions (2.26) when needed).
To see this, let u be a minimum point of P (v), so we have
P (u) P (v) v with v|x=0 = 0 .

(2.32)

We shall verify that the directional derivative of P at u exists (which is a good exercise
per se!), so all we have to do in this case is to characterize stationary points. To that
aim, we need to compute the directional derivative of P at u in the direction v. So,
given v with v|x=0 = 0, consider the real function
Fv (t) = P (u + tv) .
Recall that we have
P (u + tv) P (u)
Fv (t) Fv (0)
= lim
= (Fv ) (0).
t0
t0
t
t

P  (u).v = lim

58

Now by denition,
1
1
a(u + tv, u + tv) g(u + tv) a(u, u) + g(u)
2
2
t2
= t {a(u, v) g(v)} + a(v, v) ,
2

Fv (t) Fv (0) =

which gives
P  (u).v = (Fv ) (0) = a(u, v) g(v) .
Using (2.32) we know that t = 0 is a minimum point of Fv . It follows that
P  (u).v = (Fv ) (0) = 0.
This is valid for all v with v |x=0 = 0. So, we have
0 = a(u, v) g(v) v with v |x=0 = 0 .
Namely, u is a solution to the variational formulation (2.23)-(2.25).
To see the converse implication under the conditions (2.26) (that imply positive
deniteness of the form a(, )) , for any v such that v |x=0 = 0 we calculate
1
1
a(v, v) g(v) a(u, u) + g(u)
2
2
1
=
a(v u, v u) + a(u, v u) g(v u)
2
1
=
a(v u, v u) .
2

P (v) P (u) =

To reach the last line, we have used the variational formulation (2.23)-(2.25) with v u
as the test function. So it follows that u is a minimum point of P , under the conditions
(2.26) that ensure the positiveness of the bilinear form a(, ). Actually, since a(, ) is
strictly positive, u is a strict minimum point of P .
Remark 2.7. Above, we have seen that the minimum point u is characterized as the
solution to a boundary value problem, or to the equivalent variational formulation. It is
also characterized as the stationary point of P :
P  (u) = 0 ,
which is the Euler equation in the case of the continuous model (with a dierential
equation).
59

2.2.3

Variational formulation for more general boundary conditions

We now consider a bit more general boundary conditions. We want to solve the boundary
value problem:


d
du

c
(x) + q(x)u(x) = f (x), a < x < b
(2.33)
dx
dx
with non-homogeneous boundary conditions


du
c
|x=a = ,
dx

u|x=b = .

(2.34)

As we did previously, we can derive the variational formulation for the boundary value
problem (2.33)-(2.34). To do so, we multiply both sides of equation (2.33) by an arbitrary
test function v with43 v|x=b = 0, then integrate over ]a, b[ to obtain
 b
a

dx

du
c
dx

v + quv dx =

f v dx .

(2.35)

Now using integration by parts and the boundary conditions (2.34) for u, and the homogeneous boundary condition at x = b for v, we deduce

 b
 b
du dv
+ q uv dx =
c
f v dx v|x=a .
dx dx
a
a
This leads to the variational formulation of the boundary value problem (2.33)-(2.34):
Find u with u|x=b = such that a(u, v) = g(v) v with v |x=b = 0
where a(, ) and g() are forms respectively given by

 b
du dv
+ quv dx ,
c
a(u, v) =
dx dx
a
 b
g(v) =
f v dx v|x=a .

(2.36)

(2.37)
(2.38)

Under positivity conditions very similar to (2.26) just replace ]0, 1[ by ]a, b[! we can
prove uniqueness, and assume existence of the solution u the variational formulation.
Next, we prove an equivalence result between the boundary value problem and the
variational formulation.
43

Even if = 0, we are still using test functions with 0 value at x = b, to be able to remove the
boundary term at x = b after integrating by parts.

60

Theorem 2.13. The boundary value problem (2.33)-(2.34) is equivalent to the variational
formulation (2.36)-(2.38).
Proof. First, we know already by the derivation of the variational problem that the
solution u to the boundary value problem (2.33)-(2.34) is also a solution to the variational
formulation (2.36)-(2.38). Next, we will check that any solution u of (2.36)-(2.38) is also
a solution to the boundary value problem (2.33)-(2.34).
In fact, since u satises (2.36)-(2.38), we obtain by integrating by parts that for any
v with v |x=b = 0,
 b
a

dx

du
c
dx


du
v + quv dx c
|x=a v |x=a
dx
 b
=
f vdx v |x=a v with v |x=b = 0. (2.39)
a

Now taking only test functions v which satisfy in addition the homogeneous boundary
condition v |x=a = 0, (2.39) yields
 b
a

dx

du
c
dx


+ qu f

v dx = 0 v with v |x=a = v |x=b = 0 .

This implies (see footnote 41 )




du
d
c
(x) + q(x)u(x) = f (x),

dx
dx
Going back to (2.39), we have the equality


du
c
|x=a v |x=a = v |x=a
dx

a < x < b.

(2.40)

v with v |x=b = 0 ,

which indicates that u also satises the boundary condition




du
c
|x=a = .
dx

(2.41)

Together with the boundary condition u|x=b = 0, (2.40) and (2.41) tell us that u is a
solution to the boundary value problem (2.33)-(2.34).

61

We note that, if = 0, then u cannot be chosen as a test function in the variational


formulation, since one has v |x=b = 0 for all those test functions v. However, one can
introduce
u(x) = u(x) a < x < b ,
which now fullls the appropriate boundary condition u|x=b = 0. The function ulif t (x) =
is called a lifting of the non-homogeneous Dirichlet boundary condition at x = b.
Then, u solves the variational formulation:
Find u with u|x=b = 0 such that a(u, v) = glif t (v) v with v |x=b = 0
where a(, ) is still given by (2.37), and glif t () is now dened by
 b
(f q)v dx v|x=a .
glif t (v) = g(v) a(ulif t , v) =

(2.42)

(2.43)

Remark 2.8. The two variational formulations (2.36) and (2.42) are equivalent, in the
sense that their solutions are unique, and are related by u(x) = u(x) + ulif t (x). On the
other hand, the variational formulation (2.42) exhibits all data (here , and f (x)) in
its right-hand side glif t ()...
Obviously, the lifting function is not unique: choosing44 another lifting ulif t (x), so
that u(x) = u (x) + ulif t (x), one simply ends up modifying the right-hand side of the

variational formulation: nd u such that a(u , v) = glif
t (v) for all v, which writes

 b
dulif t dv



(f qulif t )v c
dx v|x=a .
glif t (v) = g(v) a(ulif t , v) =
dx dx
a
2.2.4

Minimum principle for dierential equations bis

To conclude, we investigate the relation between the boundary value problem (2.33)(2.34) and the minimization of some functional. Note however that we cannot use the
functional originating directly from the natural variational formulation (2.36), because
the solution u and the test functions v do not belong to the same space! So, we have to
rely instead on (2.42), to dene
1
Plif t (v) := a(v, v) glif t (v) v with v |x=b = 0 .
2
In this way, we expect u to be the minimum point of Plif t (v).
44

Propose some alternate choices!

62

Theorem 2.14. The function u that minimizes P (v) over all v with v |x=b = 0 is the
solution to the variational formulation (2.42)-(2.43) and (2.37). Then, u(x) = u(x) +
satises the dierential equation


du
d
c
(x) + q(x)u(x) = f (x), a < x < b

dx
dx
with the boundary conditions


du
c
|x=a =
dx

and u|x=b = .

The converse implication is also true under the conditions (2.26), with u(x) = u(x) .
Proof. The proof of this equivalence is basically the same as we carried out in subsection
2.2.2. So we omit it here.

63

64

You might also like