Variations 2015
Variations 2015
Lecture Notes
Erich Miersemann
Department of Mathematics
Leipzig University
1 Introduction 9
1.1 Problems in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.2 Nash equilibrium . . . . . . . . . . . . . . . . . . . . . 10
1.1.3 Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Ordinary differential equations . . . . . . . . . . . . . . . . . 11
1.2.1 Rotationally symmetric minimal surface . . . . . . . . 12
1.2.2 Brachistochrone . . . . . . . . . . . . . . . . . . . . . 13
1.2.3 Geodesic curves . . . . . . . . . . . . . . . . . . . . . . 14
1.2.4 Critical load . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.5 Euler’s polygonal method . . . . . . . . . . . . . . . . 20
1.2.6 Optimal control . . . . . . . . . . . . . . . . . . . . . . 21
1.3 Partial differential equations . . . . . . . . . . . . . . . . . . . 22
1.3.1 Dirichlet integral . . . . . . . . . . . . . . . . . . . . . 22
1.3.2 Minimal surface equation . . . . . . . . . . . . . . . . 23
1.3.3 Capillary equation . . . . . . . . . . . . . . . . . . . . 26
1.3.4 Liquid layers . . . . . . . . . . . . . . . . . . . . . . . 29
1.3.5 Extremal property of an eigenvalue . . . . . . . . . . . 31
1.3.6 Isoperimetric problems . . . . . . . . . . . . . . . . . . 31
1.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2 Functions of n variables 39
2.1 Optima, tangent cones . . . . . . . . . . . . . . . . . . . . . . 39
2.1.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2 Necessary conditions . . . . . . . . . . . . . . . . . . . . . . . 47
2.2.1 Equality constraints . . . . . . . . . . . . . . . . . . . 49
2.2.2 Inequality constraints . . . . . . . . . . . . . . . . . . 52
2.2.3 Supplement . . . . . . . . . . . . . . . . . . . . . . . . 56
2.2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 58
3
4 CONTENTS
7
8 CONTENTS
Chapter 1
Introduction
1.1 Problems in Rn
1.1.1 Calculus
Let f : V 7→ R, where V ⊂ Rn is a nonempty set. Consider the problem
9
10 CHAPTER 1. INTRODUCTION
1.1.3 Eigenvalues
Consider the eigenvalue problem
Ax = λBx,
where A and B are real and symmetric matrices with n rows (and n columns).
Suppose that hBy, yi > 0 for all y ∈ Rn \ {0}, then the lowest eigenvalue λ1
is given by
hAy, yi
λ1 = min .
y∈Rn \{0} hBy, yi
where −∞ < a < b < ∞ and f is sufficiently regular. One of the basic
problems in the calculus of variation is
(P ) minv∈V E(v).
d
fu′ (x, u(x), u′ (x)) = fu (x, u(x), u′ (x))
dx
in (a, b).
Proof. Exercise. Hints: For fixed φ ∈ C 2 [a, b] with φ(a) = φ(b) = 0 and
real ǫ, |ǫ| < ǫ0 , set g(ǫ) = E(u + ǫφ). Since g(0) ≤ g(ǫ) it follows g ′ (0) = 0.
Integration by parts in the formula for g ′ (0) and the following basic lemma
in the calculus of variations imply Euler’s equation. ✷
12 CHAPTER 1. INTRODUCTION
ub
ua
a b x
Proof. Assume h(x0 ) > 0 for an x0 ∈ (a, b), then there is a δ > 0 such that
(x0 − δ, x0 + δ) ⊂ (a, b) and h(x) ≥ h(x0 )/2 on (x0 − δ, x0 + δ). Set
2
δ 2 − |x − x0 |2 if x ∈ (x0 − δ, x0 + δ)
φ(x) = .
0 if x ∈ (a, b) \ [x0 − δ, x0 + δ]
S
a
l x
be the surface defined by rotating the curve around the x-axis. The area of
this surface is Z l p
|S(v)| = 2π v(x) 1 + (v ′ (x))2 dx.
0
Set
V = {v ∈ C 1 [0, l] : v(0) = a, v(l) = b, v(x) > 0 on (a, b)}.
Then the variational problem which we have to consider is
min |S(v)|.
v∈V
1.2.2 Brachistochrone
In 1696 Johann Bernoulli studied the problem of a brachistochrone to find
a curve connecting two points P1 and P2 such that a mass point moves from
P1 to P2 as fast as possible in a downward directed constant gravitational
field, see Figure 1.3. The associated variational problem is here
Z t2 p ′ 2
x (t) + y ′ (t)2
min p dt ,
(x,y)∈V t1 y(t) − y1 + k
x
P
1
m g
y P
2
and Chapter 3. These functions are solutions of the system of the Euler
differential equations associated to the above variational problem.
One arrives at the above functional which we have to minimize since
q p
v = 2g(y − y1 ) + v12 , v = ds/dt, ds = x1 (t)2 + y ′ (t)2 dt
and Z Z
t2 t2
ds
T = dt = ,
t1 t1 v
where T is the time which the mass point needs to move from P1 to P2 .1
Set E = xv1 · xv1 , F = xv1 · xv2 , G = xv2 · xv2 . The functions E, F and G
are called coefficients of the first fundamental form of Gauss. Then we get
11
2
mv 2 = mgy, principle of conversation of energy
1.2. ORDINARY DIFFERENTIAL EQUATIONS 15
x3
S
P
2
P1
x2
x
1
min L(v).
v∈V
where
Z l
a(u, v) = EI u′′ (x)v ′′ (x) dx
0
Z 2
b(u, v) = u′ (x)v ′ (x) dx
0
and
E modulus of elasticity,
16 CHAPTER 1. INTRODUCTION
P P
v v
l l
(a) (b)
which leads to the critical value P = EIπ 2 /l2 . If the beam is clamped at
the lower end and free (no condition is prescribed) at the upper end, see
Figure 1.5(b), then
Unilateral buckling
If there are obstacles on both sides, see Figure 1.7, then we have in the case
of a beam simply supported at both ends
V = {v ∈ C 2 [0, l] : v(0) = v(l) = 0 and φ1 (x) ≤ v(x) ≤ φ2 (x) on (0, l)}.
The critical load is here
a(v, v)
P = inf .
V \{0} b(v, v)
18 CHAPTER 1. INTRODUCTION
l
x
Figure 1.7: Unilateral beam
It can be shown, see [37, 38], that this number P is the lowest point of
bifurcation of the eigenvalue variational inequality
r(x)
Suppose that ρ/E is sufficiently small to avoid that the column is unstable
without any load P . If the column is clamped at the lower end and free at
the upper end, then we set
where r0 , r1 are given positive constants and M is the given volume of the
column. That is, we consider the saddle point problem
max min q(r, v) .
r∈U v∈V \{0}
20 CHAPTER 1. INTRODUCTION
where v ∈ V with
be a subdivision of the interval [a, b]. Then we replace the graph defined
by v(x) by the polygon defined by (x0 , A), (x1 , v1 ), ... , (xn , vn ), (xn+1 , B),
where vi = v(xi ), see Figure 1.9. Set hi = xi − xi−1 and v = (v1 , . . . , vn ),
a b x
f : [t0 , t1 ] × Rn × Rm 7→ Rn .
max E(v).
v∈Uad
The governing necessary condition for this type of problems is the Pon-
tryagin maximum principle, see [48] and Section 3.5.
V = {v ∈ C 1 (Ω) : v = h on ∂Ω}.
in Ω.
Proof. Exercise. Hint: Extend the above fundamental lemma of the calculus
of variations to the case of multiple integrals. The interval (x0 − δ, x0 + δ) in
the definition of φ must be replaced by a ball with center at x0 and radius
δ. ✷
△u = 0 in Ω, u = h on ∂Ω
1.3. PARTIAL DIFFERENTIAL EQUATIONS 23
min D(v).
v∈V
But these problems are not equivalent in general. It can happen that the
boundary value problem has a solution but the variational problem has no
solution. For an example see Courant and Hilbert [9], Vol. 1, p. 155, where
h is a continuous function and the associated solution u of the boundary
value problem has no finite Dirichlet integral.
The problems are equivalent, provided the given boundary value function
h is in the class H 1/2 (∂Ω), see Lions and Magenes [35].
V = {v ∈ C 1 (Ω) : v = h on ∂Ω}.
where E = yv1 · yv1 , F = yv1 · yv2 , G = yv2 · yv2 are the coefficients of the
first fundamental form of Gauss. Then an associated variational problem is
min |S(y)|,
y∈V
2
An experiment from Beutelspacher’s Mathematikum, Wissenschaftsjahr 2008, Leipzig
1.3. PARTIAL DIFFERENTIAL EQUATIONS 25
where ν = (ν1 , . . . , νn ) is the exterior unit normal at the boundary ∂Ω. This
follows after integration by parts from the basic lemma of the calculus of
variations.
26 CHAPTER 1. INTRODUCTION
x3
x2
x
1
Set Z Z
1
E(v) = |∇v|2 dx − h(x)v ds,
2 Ω ∂Ω
then the associated boundary value problem is
△u = 0 in Ω
∂u
= h on ∂Ω.
∂ν
div (T u) = κu in Ω
ν · T u = cos γ on ∂Ω,
1.3. PARTIAL DIFFERENTIAL EQUATIONS 27
div (T u) is the left hand side of the minimal surface equation and it is twice
the mean curvature of the surface defined by z = u(x1 , x2 ), see an exercise.
The above problem describes the ascent of a liquid, water for example,
in a vertical cylinder with constant cross section Ω. It is assumed that
the gravity is directed downwards in the direction of the negative x3 axis.
Figure 1.13 shows that liquid can rise along a vertical wedge. This is a
consequence of the strong non-linearity of the underlying equations, see
Finn [16]. This photo was taken from [42].
where
Y potential energy per unit mass, for example Y = gx3 , g = const. ≥ 0,
ρ local density,
28 CHAPTER 1. INTRODUCTION
NS
x Ωv
3
( vapor)
NΓ
S γ
Ωs Ωl
(solid)
( liquid )
|Ωl (S)| = V.
It turns out that a minimizer S0 of the energy functional under the volume
constraint satisfies, see [16],
2σH = λ + gρx3 on S0
cos γ = β on ∂S0 ,
where H is the mean curvature of S0 and γ is the angle between the surface
S0 and the container wall at ∂S0 .
Remark. The term −σβ|W | in the above energy functional is called wetting
energy.
Liquid can hang on the edge of a glass, see Figure 1.15. This picture was
taken from [42]. Here the capillary surface S satisfies a variational inequality
at ∂S where S meets the container wall along an edge, see [41].
1.3. PARTIAL DIFFERENTIAL EQUATIONS 29
solid Ωs
Ωl
S
liquid N
Ωv
vapor
Let
E(S) = σ|S| + w(S) − µ|Dl (S)| (1.1)
be the energy (grand canonical potential) of the problem, where
σ surface tension, |S|, |Ωl (S)| denote the area resp. volume of S, Ωl (S),
30 CHAPTER 1. INTRODUCTION
Z
w(S) = − F (x) dx , (1.2)
Ωv (S)
is the disjoining pressure potential, where
Z
dy
F (x) = c . (1.3)
Ωs |x − y|p
Here is c a negative constant, p > 4 a positive constant (p = 6 for nitrogen)
and x ∈ R3 \ Ωs , where Ωs denotes the closure od Ωs , that is, the union of
Ωs with its boundary ∂Ωs . Finally, set
µ = ρkT ln(X) ,
where
ρ density of the liquid,
k Boltzmann constant,
T absolute temperature,
X reduced (constant) vapor pressure, 0 < X < 1.
−△u = λu in Ω
u = 0 on ∂Ω.
for all u ∈ C01 (R2 ). After an appropriate definition of the integral on the
right hand side this inequality holds for functions from the Sobolev space
H01 (Ω), see [1], or from the class BV (Ω), which are the functions of bounded
variation, see [15]. The set of characteristic functions for sufficiently regular
32 CHAPTER 1. INTRODUCTION
domains is contained in BV (Ω) and the square root of the integral of the
right hand defines the perimeter of Ω. Set
1 : x∈Ω
u = χΩ =
0 : x 6∈ Ω
then
1
|Ω| ≤ |∂Ω|2 .
4π
max |Ω|,
Ω∈V
where V is the set of all sufficiently regular bounded and connected domains
Ω ⊂ R3 with p prescribed perimeter |∂Ω|. The solution of this problem is a
ball BR , R = |∂Ω|/(4π), and the solution is uniquely determined, see [5],
for instance, where it is shown that the isoperimetric inequality
|∂Ω|3 − 36π|Ω|2 ≥ 0
holds for all sufficiently regular Ω, and equality holds only if Ω is a ball.
1.4. EXERCISES 33
1.4 Exercises
1. Let V ⊂ Rn be nonempty, closed and bounded and f : V 7→ R lower
semicontinuous on V . Show that there exists an x ∈ V such that
f (x) ≤ f (y) for all y ∈ V .
Hint: f : V 7→ Rn is called lower semicontinuous on V if for every
sequence xk → x, xk , x ∈ V , it follows that
where 0 < ǫ < ǫ0 , ǫ0 sufficiently small, and use g(0) ≤ g(ǫ) which
implies that g ′ (0) ≥ 0.
Hint: 2x, x + y ∈ C if x, y ∈ C.
where a 6= b.
Hint:
a + b b − a arctan(x/ǫ)
v(x; ǫ) = +
2 2 arctan(1/ǫ)
defines a minimal sequence, that is, limǫ→0 E(v(ǫ)) = inf v∈V E(v).
8. Set Z b
g(ǫ) := f (x, u(x) + ǫφ(x), u′ (x) + ǫφ′ (x)) dx,
a
d
fu ′ = fu ,
dx
p
if f = 1 + (u′ )2 .
10. Set Z 1
E(v) = v 2 (x) + xv ′ (x) dx
0
and
V = {v ∈ C 1 [0, 1] : v(0) = 0, v(1) = 1}.
Show that minv∈V E(v) has no solution.
y(1 + y ′2 ) = α2 , y = y(x),
with a constant α.
(ii) Substitution
c
y = (1 − cos u), u = u(x),
2
implies that x = x(u), y = y(u) define cycloids (in German: Rollkur-
ven).
Functions of n variables
In this chapter, with only few exceptions, the normed space will be the
n-dimensional Euclidean space Rn . Let f be a real function defined on a
nonempty subset X ⊆ Rn . In the conditions below where derivatives occur
we assume that f ∈ C 1 or f ∈ C 2 on an open set X ⊆ Rn .
where ||y − x|| denotes the Euclidean norm of the vector x − y. We always
assume that x ∈ V is not isolated, that is, we assume that V ∩ Bρ (x) 6= {x}
for all ρ > 0.
39
40 CHAPTER 2. FUNCTIONS OF N VARIABLES
and we say that x ∈ V defines a strict local minimum if the strict inequality
holds for all y ∈ V ∩ Bρ (x), y 6= x.
For each given ρ > 0 there exists a λ = λ(ρ) such that y(λ) ∈ Bρ (x) and
f (y(λ)) < f (x). This is a contradiction to the assumption. ✷
Proof. Exercise.
In the following we use for (fx1 (x), . . . , fxn (x)) the abbreviations f ′ (x),
∇f (x) or Df (x).
x1 := (1 − λ)y + λx and h := y − x1 .
Then
1−λ
x = x1 − h.
λ
Since we suppose that the inequality of the theorem holds, we have
After multiplying the first inequality with (1−λ)/λ we add both inequalities
and get
1−λ 1−λ
f (y) − f (x1 ) + f (x) − f (x1 ) ≥ 0
λ λ
(1 − λ)f (y) + λf (x) − (1 − λ)f (x1 ) − λf (x1 ) ≥ 0.
Thus
(1 − λ)f (y) + λf (x) ≥ f (x1 ) ≡ f ((1 − λ)y − λx)
Remark. The inequality of the theorem says that the surface S defined by
z = f (y) is above of the tangent plane Tx defined by z = hf ′ (x), y−xi+f (x),
see Figure 2.1 for the case n = 1.
Tx
x y
x
T(V,x)
In the following the Hesse matrix (fxi xj (x))ni,j=1 is also denoted by f ′′ (x),
fxx (x) or D2 f (x).
hf ′′ (x)w, wi ≥ 0
44 CHAPTER 2. FUNCTIONS OF N VARIABLES
wk := tk (xk − x) → w.
Then
hf ′′ (x)wk , wk i + ||wk ||2 η(||xk − x||) ≥ 0,
which implies that
hf ′′ (x)w, wi ≥ 0.
(ii) Since
1
f (y) − f (x) − hf ′ (x), y − xi = hf ′′ (x + δ(y − x))(y − x), y − xi,
2
where 0 < δ < 1, and the right hand side is nonnegative, it follows from
Theorem 2.1.4 that f is convex on V . ✷
2.1. OPTIMA, TANGENT CONES 45
2.1.1 Exercises
1. Assume V ⊂ Rn is a convex set. Show that Y = V − V := {x − y :
x, y ∈ V } is a convex set in Rn , 0 ∈ Y and if y ∈ Y then −y ∈ Y .
Φ(τ ) = f (x + τ (y − x))
and
x + τ (y − x) = x + (λs1 + (1 − λ)s2 ) (y − x)
= λ (x + s1 (y − x)) + (1 − λ) (x + s2 (y − x)) .
2.2. NECESSARY CONDITIONS 47
We recall that
For a proof of Rademacher’s Theorem see [15], pp. 81, for instance.
Proof. From the definition of the tangent cone it follows that the corollary
implies Theorem 2.2.1. On the other hand, fix y ∈ V and define xk :=
(1−λk )y+λk x, λk ∈ (0, 1), λk → 1. Then xk ∈ V , (1−λk )−1 (xk −x) = y−x.
That is, y − x ∈ T (V, x). ✷
x = pV (x − qF (x)) ,
Proof. From
f (x) ≤ f (xk )
1
= f (x) + hf ′ (x), xk − xi + hf ′′ (x)(xk − x), xk − xi
2
k 2 k
+||x − x|| η(||x − x||),
where limt→0 η(t) = 0, we obtain
1
0 ≤ tk hf ′ (x), wk i + hf ′′ (x)wk , wk i + ||wk ||2 η(||xk − x||).
2
By taking lim inf the assertion follows. ✷
In the next sections we will exploit the explicit nature of the subset V .
When the side conditions which define V are equations, then we obtain
under an additional assumption from the necessary condition of first order
the classical Lagrange multiplier rule.
and from the definition of the local tangent cone it follows for each j that
From this lemma follows the simplest version of the Lagrange multiplier rule
as a necessary condition of first order.
Proof. From the necessary condition hf ′ (x), wi ≥ 0 for all w ∈ T (V, x) and
from Lemma 2.2.1 it follows that hf ′ (x), wi = 0 for all w ∈ Y ⊥ since Y ⊥ is
a linear space. This equation implies that
′
f ′ (x) ∈ Y ≡ span {g1′ (x), . . . , gm (x)}.
2.2. NECESSARY CONDITIONS 51
f (x + ǫz + y) = L(x + ǫz + y, λ)
= L(x, λ) + hL′ (x, λ), ǫz + yi
1
+ hL′′ (x, λ)(ǫz + y), ǫz + yi + o(ǫ2 )
2
1
= f (x) + hL′′ (x, λ)(ǫz + y), ǫz + yi + o(ǫ2 ),
2
52 CHAPTER 2. FUNCTIONS OF N VARIABLES
since L′ (x, λ) = 0 and x satisfies the side conditions. Hence, since f (x) ≤
f (x + ǫz + y), it follows that hL′′ (x, λ)z, zi ≥ 0.
1
f (xk ) − f (x) = hL′′ (x, λ)(xk − x), xk − xi + o(||xk − x||2 ).
2
On the other hand,
1
f (xk ) − f (x) = hf ′ (x), xk − xi + hf ′′ (x)(xk − x), xk − xi + o(||xk − x||2 ).
2
Consequently
m
′ k 1X
hf (x), x − xi = λj hgj′′ (x)(xk − x, xk − xi + o(||xk − x||2 ),
2
j=1
If j ∈ E, we obtain from
and if j ∈ E, then
(1)
is satisfied, there exist non-negative real numbers µj if j ∈ I0 , µj if j ∈ E
(2)
and µj if j ∈ E such that
X X (1) ′ X (2)
f ′ (x) = µj −gj′ (x) + µj gj (x) + µj −gj′ (x)
j∈I0 j∈E j∈E
X X (1) (2)
= − µj gj′ (x) + µj − µj gj′ (x).
j∈I0 j∈E
This follows from the Minkowski–Farkas Lemma, see Section 2.6: let A be
a real matrix with m rows and n columns and let b ∈ Rn , then hb, yi ≥ 0
∀y ∈ Rn with Ay ≥ 0 if and only if there exists an x ∈ Rm , such that x ≥ 0
and AT x = b. ✷
The following corollary says that we can avoid the consideration whether
the inequality constraints are active or not.
2.2. NECESSARY CONDITIONS 55
(i) T (V0 , x) = Z,
(ii) hL′′ (x, λ)z, zi ≥ 0 for all z ∈ T (V0 , x) (≡ Z),
1 X
(iii) lim tk hf ′ (x), wk i = λj hgj′′ (x)w, wi, wk := tk (xk − x),
k→∞ 2
j∈E∪I0
Proof. Assertion (i) follows from the maximal rank condition and the im-
plicit function theorem. Since f (y) = L(y, λ) for all y ∈ V0 , we obtain
1
f (y) − f (x) = hL′ (x, λ), y − xi + hL′′ (x, λ)(y − x), y − xi + o(||x − y||2 )
2
1 ′′
= hL (x, λ)(y − x), y − xi + o(||x − y||2 ).
2
56 CHAPTER 2. FUNCTIONS OF N VARIABLES
2.2.3 Supplement
In the above considerations we have focused our attention on the necessary
condition that x ∈ V is a local minimizer: hf ′ (x), wi ≥ 0 for all w ∈
C(V, x). Then we asked what follows from this inequality when V is defined
by equations or inequalities. Under an additional assumption (maximal rank
condition) we derived Lagrange multiplier rules.
In the case that V is defined by equations or inequalities there is a
more general Lagrange multiplier rule where no maximal rank condition is
assumed. Let
V = {y ∈ Rn : gj (y) ≤ 0 for each j ∈ I and gj (y) = 0 for each j ∈ E} .
The case where the side conditions are only equations is included, here I is
empty.
Proof. We will show by contradiction that the vectors f ′ (x), gj′ (x), j ∈
E ∪ I0 , must be linearly dependent if x defines a local minimum.
By assumption we have gj (x) = 0 if j ∈ E ∪ I0 and gj (x) < 0 if j ∈ I \ I0 .
Assume that the vectors f ′ (x), g1′ (x), . . . , gm
′ (x), g ′ ′
m+l1 (x), . . . , gm+lk (x) are
linearly independent, where I0 = {m + l1 , . . . , m + lk }. Then there exists a
regular quadratic submatrix of N = 1 + m + k rows (and columns) of the
associated matrix to the above (column) vectors. One can assume, after
renaming of the variables, that this matrix is
fx1 (x) g1,x1 (x) · · · gN −1,x1 (x)
fx2 (x) g1,x2 (x) · · · gN −1,x2 (x)
................................... .
fxN (x) g1,xN (x) · · · gN −1,xN (x)
2.2. NECESSARY CONDITIONS 57
Here gi,xj denote the partial derivative of gi with respect to xj . Set h = f (x),
then we consider the following system of equations:
f (y1 , . . . , yN , xN +1 , . . . , xn ) = h + u
gj (y1 , . . . , yN , xN +1 , . . . , xn ) = 0 , j ∈ E ∪ I0 ,
2.2.4 Exercises
1. Suppose that f ∈ C 1 (Bρ ). Show that
(ii) Assume the necessary condition (⋆) is satisfied and V is convex, then
assumption A holds with M = 0.
Definition. Let Tf ′ (V, x) be the set of all w ∈ T (V, x) such that, if xk and
tk = ||xk −x||−1 are associated sequences to w, then lim supk→∞ t2k hf ′ (u), xk −
xi < ∞.
Set
f ′ (x)⊥ = {y ∈ Rn ; hf ′ (x), yi = 0} .
convex), then
Tf ′ (V, x) ⊆ T (V, x) ∩ f ′ (x)⊥ .
Proof. Assume that w ∈ Tf ′ (V, x) and let tk and xk are associated sequences.
If w 6= 0, then tk → ∞ and assumption A implies that hf ′ (x), wi ≥ 0, see
Remark 2.2. On the other hand, the inequality lim inf k→∞ t2k hf ′ (x), xk −xi <
∞ yields hf ′ (x), wi ≤ 0. ✷
Proof. If x does not define a strict local minimum, then there exists a
sequence xk ∈ V , xk → x, xk 6= x, such that
0 ≥ f (xk ) − f (x)
1
= hf ′ (x), xk − xi + hf ′′ (x)(xk − x), xk − xi + o(||xk − x||2 ) .
2
Set tk = ||xk − x||−1 , then
1
0 ≥ t2k hf ′ (x), xk − xi + hf ′′ (x)(tk (xk − x)), tk (xk − x)i
2
o(||x k − x||2 )
+t2k ||xk − x||2 .
||xk − x||2
The following example shows that Tf ′ (V, x) can be a proper subset of C(V, x)∩
f ′ (x)⊥ .
where λ = (λ1 , . . . , λm ), λj ∈ R.
1 o(||xk − x||2 )
0 ≥ hL′′ (x, λ)tk (xk − x), tk (xk − x)i + ,
2 ||xk − x||2
We recall (see Lemma 2.2.2) that the tangent cone T (V, x) is a hyperplane
given by {y ∈ Rn ; hgj′ (x), yi = 0, for every j} if the maximal rank condition
of Lemma 2.2.2 is satisfied.
Proof. Set
X
G(y, λ) := − λj gj (y) ,
j∈I0+
then G(x, λ) = 0, G(y, λ) ≥ 0 for all y ∈ V and f (y) ≡ L(y, λ) + G(y, λ).
If x is no strict local minimizer, then there exists a sequence xk ∈ V, xk →
2.3. SUFFICIENT CONDITIONS 63
x, xk 6= x, such that
0 ≥ f (xk ) − f (x)
= L(xk , λ) − L(x, λ) + G(xk , λ) − G(x, λ)
1
= hL′ (x, λ), xk − xi + hL′′ (x, λ)(xk − x), xk − xi + G(xk , λ)
2
+o(||xk − x||2 )
1 ′′
= hL (x, λ)(xk − x), xk − xi + G(xk , λ) + o(||xk − x||2 ) ,
2
Set tk = ||xk − x||−1 , then
1 o(||xk − x||2 )
0 ≥ hL′′ (x, λ)(tk (xk − x)), tk (xk − x)i + t2k G(xk , λ) + . (2.4)
2 ||xk − x||2
This inequality implies that t2k G(xk , λ) is bounded from above. Since G(y, λ) ≥
0 for all y ∈ V , it follows that tk G(xk , λ) → 0. On the other hand
o(||xk − x||)
tk G(xk , λ) = hG′ (x, λ), tk (xk − x)i + ,
||xk − x||
which follows since G(xk , λ) − G(x, λ) = G(xk , λ) holds. Thus we find that
hG′ (x, λ), wi = 0, where w is the limit of a subsequence of tk (xk − x), tk ≡
||xk − x||−1 . Since w ∈ C(V, x) we have hgj′ (x), wi ≤ 0 if j ∈ I0+ . Hence,
since per definition λj > 0 if j ∈ I0+ , we obtain from the definition of G(y, λ)
that
hgj′ (x), wi = 0 for each j ∈ I0+ (2.5)
From G(xk , λ) ≥ 0 it follows from inequality (2.4) that hL′′ (x, λ)w, wi ≤ 0.
This inequality and equations (2.5) contradict the assumption of the theo-
rem. Since the proof shows that T0 6= {0} if x is no strict local minimizer,
it follows that x defines a strict local minimum if T0 = {0}. ✷
Remark. The above proof is mainly based on the observation that the
sequence t2k G(xk , λ) remains bounded from above. In the general case of a set
V which defines the side condition we have that the sequence t2k hf ′ (x), xk −xi
remains bounded from above. In the infinitely dimensional case we must
exploit this fact much more then in the above finitely dimensional case
where it was enough to use the conclusion that hf ′ (x), wi = 0.
64 CHAPTER 2. FUNCTIONS OF N VARIABLES
2.3.3 Exercises
1. Show that assumption A implies that the necessary condition of first
order hf ′ (x), wi ≥ 0 holds for all w ∈ T (V, x).
2. Show that Tf ′ (V, x) = {0} in the above example if 1 < α < 2 holds.
V = {y ∈ X : g j (y) ≥ 0, j = 1, . . . , m},
(P ) maxy∈V f (y).
for all x ∈ X and for all λ ≥ 0, see Figure 2.3 for an illustration of a saddle
point. The relationship between saddle points and the problem (P) is the
x
λ
g j (x0 ) ≥ 0, j = 1, . . . , m,
m
X
λ0j g j (x0 ) = 0.
j=1
Proof. Since
m
X
f (x) + λ0j g j (x) ≤ f (x0 )
j=1
k
X
λj g j (x) ≤ 0
j=1
for all x ∈ X.
2.4. KUHN-TUCKER THEORY 67
Zx = {z ∈ Rk : z < g(x)}
hp0 , zi ≥ hp0 , 0i
Thus,
hp, g(x) − ǫi ≤ 0
for all x ∈ X and for all ǫ > 0. Consequently hp, g(x)i ≤ 0 for all x ∈ X. ✷
Theorem 2.4.5. Suppose that the assumptions of the previous theorem are
fulfilled and additionally that the Slater condition holds. Then there are non-
negative constants λ0j , j = 1, . . . , m such that (x0 , λ0 ), λ0 = (λ01 , . . . , λ0m ), is
P
a saddle point of the Lagrange function L(x, λ) = f (x) + m j
j=1 λj g (x).
P
Proof. If p0 = 0, then m j
j=1 pj g (x) ≤ 0 for all x ∈ X, and, in particular,
Pm j 1
j=1 pj g (x ) ≤ 0 which is a contradiction to the Slater condition. Set
λ0j = pj /p0 , j = 1, . . . , m, then
Lemma. Suppose that (x0 , λ0 ) is a saddle point of L(x, λ), X is convex and
f , g ∈ C 1 . Then
hL′ (x0 , λ0 ), x − x0 i ≤ 0
for all x ∈ X.
From the above Theorem 2.4.4, Theorem 2.4.5 and the previous lemma it
follows
for all x ∈ X. The second inequality is the first of the Kuhn-Tucker condi-
tions. On the other hand we have
Thus we have shown that (x0 , λ0 ) is a saddle point. Then the assertion of
the theorem follows from Theorem 2.4.2. ✷
70 CHAPTER 2. FUNCTIONS OF N VARIABLES
This example was taken from [3], p. 65. Suppose that a firm produces a
good and let q be the real number of the produced goods (output). Assume
there are n different goods which the production needs (input) and let xj
the real number of the good j which are used. Set x = (x1 , . . . , xn ) and
let r = (r1 , . . . , rn ) be the associated given price vector. We make the
assumption q ≤ f (x), where f denotes the given production function. Then
the problem of profit maximizing is
X = {(q, x) ∈ Rn+1 : q ≥ 0, x ≥ 0}
and let
L(x, λ) = p q − hr, xi + λ(f (x) − q)
be the associated Lagrange function. We suppose that f is concave in x ≥ 0,
f (0) = 0, f (x) > 0 if x ≥ 0 and not x = 0, and fxj (0) > 0 for at least one j,
which implies that the Slater condition is satisfied. From above we have that
the Kuhn-Tucker conditions are necessary and sufficient that (q 0 , x0 ) ∈ X is
a global maximizer of the profit pq − hr, xi under the side conditions. That
is, we have to consider the following inequalities and one equation for q 0 ≥ 0
and x ≥ 0: P
(i) (p − λ0 )(q − q 0 ) + nj=1 −rj + λ0 fxj (x0 ) (xj − x0j ) ≤ 0 for all (q, x) ∈ X,
(ii) f (x0 ) − q 0 ≥ 0,
(iii) λ0 (f (x0 ) − q 0 ) = 0,
(iv) λ0 ≥ 0.
2.4.1 Exercises
1. Set Zx = {z ∈ Rk : z < f (x)} and Z = ∪x∈X Zx , see the proof of
Theorem 2.4.3. Show that Z is convex if X is convex and f is concave
on X.
2.5 Examples
2.5.1 Maximizing of utility
This example was taken from [3], p. 44. Set X := {x ∈ Rn : xi >
0, i, . . . , n}, which is called the consumption set. By writing x >> 0 we
mean that xi > 0 for each i. Let p ∈ Rn , p >> 0, the vector of prices for
the n commodities xi and m denotes the available income of the consumer.
Concerning the utility function U (x) we assume that U ∈ C 2 (X) is strictly
concave and U ′ >> 0 ∀x ∈ X. The assumption U ′ >> 0 reflects the
microeconomic principle “more is better”. Set V = {x ∈ X : hp, xi ≤ m}
and consider the problem of maximizing of the utility, that is, maxx∈V U (x)
under the budget restriction hp, xi ≤ m. Assume x0 ∈ V is a solution, then
hp, x0 i = m, which follows from the assumption Uxi > for each i. Thus,
one can replace the above problem by maxx∈V ′ U (x), where V ′ = {x ∈ X :
hp, xi = m}. From assumption on U it follows that a local maximum is also
a global one. The associated Lagrange function is here
The previous equation follows from the necessary condition of first order.
Consider the system U ′ (x0 ) − λ0 p = 0, hp, x0 i − m = 0 of n + 1 equations,
then it follows from the necessary condition of first order and the above
sufficient condition that the matrix
′′ 0
U (x ) p
,
pT 0
2.5.2 V is a polyhedron
Suppose that f ∈ C 2 and that V is given by
V = {y ∈ Rn : hli , yi ≤ ai , i = 1, . . . , m} ,
P
where hli , yi := nk=1 lik yk , lik ∈ R, are given linear functionals. Assume
x ∈ V is a solution of the variational inequality hf ′ (x), y − xi > 0 for all
y ∈ V , (or equivalently of the corresponding Lagrange multiplier equation).
Define the cone
hf ′′ (x)y, yi
λ1 = min ,
y∈K\{0} hy, yi
is positive.
where the minimum is taken over y 6= 0 which satisfy hBx(l) , yi = 0 for all
l, 1 ≤ l ≤ k − 1.
74 CHAPTER 2. FUNCTIONS OF N VARIABLES
Proof. Set
1 1
f (y) := hAy, yi , g(y) := hBy, yi .
2 2
Step 1. We consider the problem to minimize f (y) under the side condition
g(y) = 1. There exists a minimizer x(1) . The vector x(1) is a regular point
since Bx(1) 6= 0 which follows since B is positive definite. Then Lagrange
multiplier rule implies that there exists an eigenvalue λ1 and that x(1) is an
associated eigenvector. Since hAx(1) , x(1) i = λ1 hBx(1) , x(1) i it follows that
λ1 = min(1/2)hAy, yi under the side condition (1/2)hBy, yi = 1 .
Step 2. We consider the problem to minimize f (y) under the side conditions
g(y) = 1 and hg ′ (x(1) ), yi = 0. We recall that g ′ (y) ≡ By. By the same
reasoning as above we find a minimizer x(2) which is a regular vector since
Bx(1) and Bx(2) are linearly independent (Exercise). Then there exists
λ2 , µ ∈ Rn such that
From the side conditions it follows that the multipliers µl are zero. Moreover,
λk = min(1/2)hAy, yi, where the minimum is taken over
1
{y ∈ Rn ; hBy, yi = 1, hBx(l) , yi = 0, l = 1, . . . , k − 1}.
2
multiplier rule. Let x(k) be a minimizer of the problem min f (y) under the
side conditions g(y) = 1, and hg ′ (x(l) ), yi = 0, 1 ≤ l ≤ k − 1, where x(l)
are mutually B-orthogonal, that is, hBx(l) , x(k) i = 2δlk , and suppose that
Ax(l) = λl Bx(l) , 1 ≤ l ≤ k − 1. Equivalently, x(k) is a solution of
hAy, yi
min =: λk ,
hBy, yi
where the minimum is taken over all y 6= 0 which satisfy the side conditions
hg ′ (x(l) ), yi = 0, 1 ≤ l ≤ k − 1. We will show that x(k) is an eigenvector to
the eigenvalue λ = λk . Set for ǫ, |ǫ| < ǫ0 , ǫ0 sufficiently small,
hg ′ (x(l) ), yi = 0 , 1 ≤ l ≤ k − 1 . (2.6)
for all y which satisfy the above side conditions (2.6). It remains to show
that (2.7) is true for all y ∈ Rn . Set Z = span {x(1) , . . . , x(k−1) }. Then
Rn = Z ⊕ Z ⊥ , where the orthogonal decomposition is taken with respect
to
Pk−1the scalar product hBx, yi. For y ∈ Rn we have the decomposition y =
(l) + w, c ∈ R, w ∈ Z ⊥ . We must show that
l=1 cl x l
k−1
X k−1
X
hAx(k) , cl x(l) + wi = λk hBx(k) , cl x(l) + wi
l=1 l=1
If 1 ≤ l ≤ k − 1, then
since x(k) satisfies the side conditions (2.6) and since B is symmetric. It
remains to show that hAx(k) , x(l) i = 0. On the other hand, we have for
1 ≤ l ≤ k − 1 the equations λl hBx(l) , yi = hAx(l) , yi are satisfied for all
76 CHAPTER 2. FUNCTIONS OF N VARIABLES
Proof. Set z (1) = x(1) , . . . , z (k−1) = x(k−1) , where x(l) denotes the above
eigenvector to the eigenvalue λl . Then
1
min hAy, yi = λk ,
2
where the minimum is taken over V (x(1) , . . . , x(k−1) ). That is,
λk ≤ sup Λk (z (1) , . . . , z (k−1) ) .
z (1) ,...,z (k−1)
P
On the other hand, let x b := kl=1 cl x(l) , where we choose coefficients cl such
P
that 21 hBb bi = 1, that is, kl=1 c2l = 1 and hBz (l) , x
x, x bi = 0, l = 1, . . . , k − 1,
for fixed vectors z (1) , . . . , z (k−1) . Then
X kX k
(1) (k−1) 1
Λk (z ,...,z ) ≤ hAb bi =
x, x c2l λl ≤ λk c2l = λk .
2
l=1 l=1
Consequently
sup Λk (z (1) , . . . , z (k−1) ) ≤ λk .
z (1) ,...,z (k−1)
the variational inequality (2.9) and the Lagrange multiplier equation (2.11)
are equivalent.
T (V, x) = {y ∈ Rn : yj ≤ 0 if xj = bj and yj ≥ 0 if xj = aj }.
2.5. EXAMPLES 79
T0 = {z ∈ T (V, x) : zj = 0 if j ∈ I0± }.
It follows from the sufficient criterion Theorem 2.3.3 that (x, λ0 ) defines a
strict local minimum of F (y, λ0 ) if T0 = {0} or if A − λ0 B is positive on
T0 \ {0}, i. e., if h(A − λ0 B)z, zi > 0 for all z ∈ T0 \ {0}.
(iii) A set of payoff functions fi : S → R. The value fi (s) is the payoff for
the i-th player if the players choose the strategy s ∈ S.
We will denote such a game by {Si , fi }i∈N . To formulate the concept
of a Nash equilibrium we need some notations. Set S−i = S1 × S2 × · · · ×
Si−1 × Si+1 × · · · × Sn and for a given strategy vector s = (s1 , s2 , . . . , sn ) ∈ S
we define s−i ∈ S−i by s−i = (s1 , s2 , . . . , si−1 , si+1 , . . . , sn ). Finally, set
s−i \ t = (s1 , s2 , . . . , si−1 , t, si+1 , . . . , sn ), t ∈ Si .
Example: Oligopoly
Each concave function is also quasiconcave, but there are quasiconcave func-
tions which are not concave, see for example f (x) = −x2 + 1 on 0 ≤ x ≤ 1
and −x + 1 on [1, ∞) which is quasiconcave on [0, ∞].
The existence of such a fixed point follows from the fixed point theorem of
Kakutani [27, 22]. ✷
Suppose that fi ∈ C 1 and that s∗ defines a local equilibrium. Then for every
i ∈ N = {1, . . . , n}
Sufficient conditions follow from the results of Section 2.3. To simplify the
considerations assume that each Si ⊂ Rmi is a parallelepiped Si = {x ∈
Rmi ; aki ≤ xki ≤ bki , k = 1, . . . , mi }. Define
⊥
fi,s i
(s∗ ) = {y ∈ Rmi : hfi,si (s∗ ), yi = 0}
and
λi = maxhfi,si ,si (s∗ )y, yi ,
y
where the maximum is taken over y ∈ T (Si , s∗i ) ∩ fi,s ⊥ (s∗ ) which satisfy
i
⊥ (s∗ ) = {0} we set λ = −∞.
hy, yi = 1. In the case that T (Si , s∗i ) ∩ fi,s i i
Assume the payoff functions fi ∈ C 2 and that s∗ satisfies the necessary con-
ditions. If λi < 0 for every i, then s∗ defines a local equilibrium. If λi > 0
for at least one i, then s∗ defines no equilibrium.
2.5.6 Exercises
1. Show that the matrix, see Section 2.4.1,
U ′′ (x0 ) p
pT 0
is regular.
2. Set
V = {y ∈ Rn : aj ≤ yj ≤ bj , j = 1, . . . , n},
Show that λ0 > 0 holds, provided that the real matrices A and B are
symmetric and positive, and that aj ≤ 0 ≤ bj for all j.
Hint: The variational inequality is equivalent to the Lagrange rule (2.11).
3. Let
2 −1 0 1 0 0
A = −1 2 −1 , B= 0 1 0 .
0 −1 2 0 0 1
and V = {y ∈ R3 ; yi ≤ 1, i = 1, 2, 3}.
6. Let
8. Suppose that
Pn the oligopoly has a linear demand function, that is,
p(r), r = i qi is given by
a − br if 0 ≤ r ≤ a/b
p(r) = ,
0 if r > a/b
where a and b are given positive constants. Assume that the cost
functions are linear, then the payoff functions are given by
n
X
πi (q) = p(r)qi − ci qi , r := qk .
k=1
Show that these payoff functions are continuous and individually quasi-
concave. Consequently there exists a Nash equilibrium of this oligopoly.
Let the strategy set of the i-th firm be 0 ≤ xi ≤ 1, the capacity bound
of the i-th firm is given by a positive constant yi , and a, b are positive
constants. Set
gi (x, y) = −fi,xi (x, y)
and V = [0, 1]n . Then a necessary condition that x⋆ defines a local
Nash equilibrium is
12. Consider the oligopoly of the previous exercise but with the additional
side condition
X n
a
yk x k ≤ .
b
k=1
Remark. A problem where the strategy set of the i-th player depends
on the strategy of the other players is called a social system, see [11].
2.6. APPENDIX: CONVEX SETS 87
(a) (b)
H(p, α) = {y ∈ Rn : hp, yi = α}
is called hyperplane.
(i) hp, yi < α for all y ∈ A and hp, yi > α for all y ∈ B,
(ii) hp, yi > α for all y ∈ A and hp, yi < α for all y ∈ B.
x .
.z X
implies that
x ∈ X : hx − z, y − xi ≥ 0 for all y ∈ X.
Set p = x − z and α = hx − z, xi, then hp, yi ≥ α for all y ∈ X.
Inequality hp, zi < α holds since
hp, zi = hx − z, zi
= −hx − z, x − zi + hx − z, xi
= −||x − z||2 + α
< α.
Then the hyperplane H(p, α⋆ ), where hp, zi < α∗ < α, separates X and z
strictly. ✷
or
hp, yi ≤ α for all y ∈ X and hp, xi = α.
pk αk pk
h , yi ≥ ≥ h , xk i
||pk || ||pk || ||pk ||
for all y ∈ X. Thus H(pk⋆ , αk⋆ ), where pk⋆ = pk /||pk || and αk⋆ = αk /||pk ||,
separate X and xk . Choose a subsequence of xk such that the associated
subsequences pk⋆ and αk⋆ converge, say to p and α, respectively. It follows
that
hp, yi ≥ α ≥ hp, xi for all y ∈ X.
These inequalities imply that α = hp, xi since x ∈ X. ✷
hp, xi ≥ hp, yi
hp, xi ≥ α ≥ hp, yi
(j) hy, x1 i
lim λ1 = ,
j→∞ hx1 , x1 i
provided that x1 6= 0.
(ii) Suppose the lemma is shown for all k satisfying 1 ≤ k ≤ s − 1. Then we
will show the lemma if k = s. In the case that the cone C contains all of the
vectors −x1 , . . . , −xs , then C is a subspace of Rn . Then the lemma is shown
since a subspace is closed. Assume at least one of the vectors −x1 , . . . , −xs
is not in C, say −xs . Then the cone
s−1
X
′ n
C := {x ∈ R : x = λl xl , λl ≥ 0}
l=1
S = {s ∈ Rn : s = AT x, x ≥ 0}
and T = {b}. The above lemma implies that the convex cone S is closed.
Since S and T are disjoint, there exists a hyperplane which separates these
sets strictly. Thus there are p ∈ Rn , p 6= 0, and α ∈ R such that
for all s ∈ S. Thus hp, AT xi > α for all x ≥ 0. Set x = 0, then we see that
α < 0. Let x = xj ej , where xj ∈ R and xj > 0, and ej denotes the standard
basis vectors in Rm . Then
α
hp, AT ej i >
xj
since x0 ≥ 0, z ≥ 0, y 0 ≥ 0 and Ay 0 ≥ 0.
(ii) Suppose that there is no non-negative solution of AT x ≤ b. Then there
are no non-negative x ∈ Rm , z ∈ Rn such that AT x + z = b. Set w = (x, z)
and B T = AT En , where
a11 · · · am1 1 · · · 0
AT En = . . . . . . . . . . . . . . . . . . . . . . . . . .
a1n · · · amn 0 · · · 1
see an exercise.
Proof. Exercise.
In the case that V is a closed convex cone K with vertex at the origin, then
there is an interesting decomposition result due to Moreau [45].
u = u1 + u2 .
(1 − t)hu1 , u2 i ≤ 0,
hu − u1 , v − u1 i = hu2 , v − u1 i = hu2 , vi ≤ 0,
which implies that u1 = pK (u), see (2.13). By the same reasoning we con-
clude that u2 = pK ∗ (u) since for v ′ ∈ K ∗ we have
hu − u2 , v ′ − u2 i = hu1 , v ′ − u2 i = hu1 , v ′ i ≤ 0.
K*
f (w) + λ0 g(w) = 0
for all w ∈ V .
Proof. Following the proof of of the previous lemma, we find that hF, hi ≥ 0
for all h ∈ cl K ∩ cl Z. Thus −F ∈ (cl K ∩ cl Z)∗ . Then the proof is based
on the formula, see the lemma below,
(cl K ∩ cl Z)∗ = cl (K ∗ + Z ∗ ) .
−F ∈ cl (K ∗ + span {G}) .
2.6. APPENDIX: CONVEX SETS 97
hwn , hi − λn hG, hi ≤ 0
(K1 ∩ · · · ∩ Km )∗ = cl (K1∗ · · · + Km
∗
)
(K1∗ · · · + Km
∗
) ⊂ (K1 ∩ · · · ∩ Km )∗
follows since we have for given vi ∈ Ki∗ that hvi , hi ≤ 0 for all h ∈ Ki .
Consequently hv1 + · · · + vm , hi ≤ 0 for all h ∈ K1 ∩ · · · ∩ Km . Thus
v1 + · · · + vm ∈ (K1 ∩ · · · ∩ Km )∗ .
(ii) Set C = cl (K1∗ · · · + Km ∗ ). Let w ∈ (K ∩ · · · ∩ K )∗ be given and
1 m
suppose that w ∈ / C. From a separation theorem, see one of the following
exercises, it follows that there is a p ∈ H such that hp, wi > 0 and hp, yi ≤ 0
for all y ∈ C. We have hw, vi ≤ 0 for all v ∈ K1 ∩ · · · ∩ Km and hp, yi ≤ 0
for all y ∈ K1∗ · · · + Km
∗ . The previous inequality shows that p ∈ K for all
i
i. Then hw, pi ≤ 0 in contrast to a separation theorem. ✷
98 CHAPTER 2. FUNCTIONS OF N VARIABLES
2.6.5 Exercises
1. Prove the related Theorem 2.6.1 where X is a closed and convex subset
of a real Hilbert space.
and divide y j by aj .
is equivalent to
w ∈ V : hw − z, y − wi ≥ 0 for all y ∈ V.
7. Show that the polar cone K ∗ is a convex and closed cone with vertex
at the origin.
8. Let K be a closed convex cone with the vertex at the origin. Show
that (K ∗ )∗ = K.
for all w ∈ V .
for all w ∈ K.
Hint: There are Gj ∈ H such that Z = {h ∈ H : hGj , hi = 0, j =
P
1, . . . , N }. Set M = {G ∈ H : G = Nj=1 λj Gj , λj ∈ R} and show
that Z = M ∗ .
100 CHAPTER 2. FUNCTIONS OF N VARIABLES
2.7 References
The main part of this material is quite standard. The first order neces-
sary condition and the concept of a local tangent cone was adopted from
Lions [34].
The presentation of the main results and proofs concerning Lagrange
multiplier rules was adopted from Hestenes [23], see also Luenberger [36].
Concerning applications, the study of eigenvalue equations is quite stan-
dard, see for instance Courant and Hilbert [9]. The case of unilateral eigen-
value problems is a finite dimensional version of problems due to Mierse-
mann [37]. The first part concerning non-cooperative games is adopted from
Luenberger [36]. The equilibrium concept for non-cooperative games is due
to Cournot [10]. A first existence proof was given by Nash [46]. The con-
cept of a local equilibrium is a straightforward generalization. References
for non-cooperative games with applications to economy are Debreu [11],
Friedman [18] and Luenberger [36], for instance. For other applications of
finite dimensional variational calculus to economics see [58].
A source for variational calculus in Rn is Rockafellar and Wets [49] and
Rockafellar [50] for convex sets.
Chapter 3
Ordinary differential
equations
The most part of this chapter concerns classical results of the calculus of
variations.
A weak local minimizer is said to be a strict weak local minimizer if E(u) <
E(v) for all v ∈ V , v 6= u, ||v − u||B < ρ.
1
More precisely, we assume that there is an injective embedding j : B 7→ H, that is, j
is linear and bounded and ||v||B 6= 0 implies ||j(v)||H 6= 0.
101
102 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Proof. Exercise.
Corollaries.
Proof. Exercise.
Corollary. Suppose the Gâteaux derivative exists then also the Gâteaux
variation and δE(u)(h) = hE ′ (u), hi.
104 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
3.1.1 Exercises
1. Suppose that V ⊂ H is not empty, where H is a Hilbert space. Show
that T (V, x) is weakly closed in H.
4. Suppose that the Gâteaux derivative exists. Show that the Gâteaux
variation exists and (δE)(h) = hE ′ (u), hi.
5. Set for y ∈ R2 (
y1 y22
y12 +y22
: y 6= (0, 0)
f (y) =
0 : y = (0, 0)
Show that there exists the first variation at (0, 0), and that the Gâteaux
derivative at (0, 0) does not exist.
δE(u)(λh) = λδE(u)(h)
for all λ ∈ R.
(ii) Show that the right variation is positive homogeneous of degree
one.
R1 R1
Example. E(v) = 0 v ′ (x)2 dx, hE ′ (u), hi = 2 0 u′ (x)v ′ (x) dx, B =
C 1 [0, 1], H = H 1 (0, 1).
thus
ub
ua
a b x
is
(P ) minv∈V E(v).
Proof. Set g(ǫ) = E(u + ǫφ) for fixed φ ∈ Φ and |ǫ| < ǫ0 . Since g(0) ≤ g(ǫ)
it follows g ′ (0) = 0 which is equation (3.2). ✷
d
fu′ (x, u(x), u′ (x)) = fu (x, u(x), u′ (x))
dx
in (a, b), provided that u ∈ C 2 (a, b). We will see that the assumption u ∈ C 2
is superfluous if fu′ u′ 6= 0 on (a, b).
Proof. Set
Z x Z b
x−a
φ0 (x) = h(ζ) dζ − h(ζ) dζ.
a b−a a
Then Z b
1
φ′0 (x) = h(x) − h(ζ) dζ.
b−a a
108 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Thus Z b
1
h(x) = h(ζ) dζ.
b−a a
✷
Proof. Set Z x
P (x) = fu (ζ, u(ζ), u′ (ζ)) dζ.
a
for all φ ∈ Φ. The above lemma implies that fu′ − P = const = c on [a, b].
Set Z x
F (x, p) = fu′ (x, u(x), p) − fu (ζ, u(ζ), u′ (ζ)) dζ − c.
a
Proof. Exercise.
max E(K).
K∈V
Set
F (t, K, K ′ ) = U (f (K) − K ′ )e−ρt ,
then the associated Euler equation is
d
FK ′ = FK
dt
on 0 < t < T . We have
FK ′ = −U ′ (f (K) − K ′ )e−ρt
FK = U ′ (f (K) − K ′ )f ′ (K)e−ρt
FK ′ K ′ = U ′′ (f (K) − K ′ )e−ρt .
110 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Thus
minimizer, then
hE ′′ (u)φ, φi ≥ 0
for all φ ∈ V − V .
Proof. Set g(ǫ) = E(u + ǫφ) for |ǫ| < ǫ0 and fixed φ ∈ Φ, then
1
g(0) ≤ g(ǫ) = g(0) + g ′ (0)ǫ + g ′′ (0)ǫ2 + o(ǫ2 )
2
as ǫ → 0. Since g ′ (0) = 0 it follows g ′′ (0) ≥ 0, which is the inequality of the
theorem. ✷
on [a, b].
Proof. (i) Since the inequality of Theorem 3.2.6 holds for φ in the Sobolev
space H01 (a, b) the following function φh is admissible. Let φh (x) be contin-
uous on [a, b], zero on |x − x0 | ≥ h, φh (x0 ) = h and linear on x0 − h < x < x0
and x0 < x < x0 + h. Set φ = φh in the necessary condition, then
Z x0 +h Z x0 +h Z x0 +h
2
0≤ fu′ u′ dx + 2h |fuu′ | dx + h |fuu | dx,
x0 −h x0 −h x0 −h
112 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
which implies
0 ≤ 2hfu′ u′ (x1 , u(x1 ), u′ (x1 )) + 4h2 max |fuu′ | + 2h3 max |fuu |,
[x0 −h,x0 +h] [x0 −h,x0 +h]
(ii) The inequality of the theorem follows also by inserting the admissible
function
1 2
h3
h2 − |x − x0 |2 if |x − x0 | ≤ h
φh (x) = .
0 if |x − x0 | > h
✷
d
Lv ≡ (Rv ′ ) − Sv = 0
dx
Lv = 0 in (a, b) (3.3)
′
v(a) = 0, v (a) = 1.
We suppose that the strict Legendre condition fu′ u′ > 0 is satisfied on [a, b]
and that there exists C 1 -extensions of R and S onto C 1 [a − δ, b + δ] for a
(small) δ > 0.
Proof. If not, then a < ζ < b. We construct a w ∈ H01 (a, b) such that
hE ′′ (u)φ, φi < 0. We choose a fixed h ∈ C 2 [a, b] such that h(a) = h(b) = 0,
h(ζ) > 0, for example h(x) = (x − a)(b − x) and define
v(x) + κh(x) if a≤x≤ζ
w(x) = ,
κh(x) if ζ<x≤b
where v is the solution of the above initial value problem (3.3). The positive
114 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
for all 0 < κ < κ0 , κ0 sufficiently small. We recall that R(ζ) > 0, v ′ (ζ) < 0
and h(ζ) > 0. ✷
Lemma. Assume that the strict Jacobi condition is satisfied. Then there
6 0 on [a, b].
exists a solution v of the Jacobi equation such that v =
Proof. Consider the initial value problem Lv = 0 on (a, b), v(a) = α, v ′ (a) =
1, where α is a small positive constant. Let v(α; x) be the solution and ζ(α)
the lowest zero of v(α; x). Then ζ(α) → ζ(0) as α → 0, which is a result
in the theory of ordinary differential equations (continuous dependence of
solutions on data). ✷
3.2. NECESSARY CONDITIONS 115
Consequently
Z b
hE ′′ (u)φ, φi = (S + z ′ )φ2 + 2zφφ′ + Rφ′2 dx.
a
P
The integrand of the right hand side is a quadratic form aij ζi ζj , where
ζ1 = φ′ , ζ2 = φ and a11P= R, a12 = z, a22 = S + z ′ . Set ζ = U (x)η, where
U is orthogonal, then aij ζi ζj = λ1 η12 + λ2 η22 . The requirement that one
of the eigenvalues of the matrix (aij ) is zero leads to
z 2 = R(S + z ′ ), (3.4)
which is a Riccati equation for z. Let V ∈ C 1 [a, b], V 6= 0 on [a, b], then the
substitution
V′
z = −R (3.5)
V
transforms the Riccati equation into the Jacobi equation LV = 0 for V . On
the other hand, let V 6= 0 on [a, b], then (3.5) is a solution of the Riccati
equation (3.4). The transformation (3.5) is called Legendre transformation.
Thus the second variation is
Z b
′′ z 2
hE (u)φ, φi = R φ′ + φ dx, (3.6)
a R
since S + z ′ = z 2 /R.
Theorem 3.2.9. Suppose the strict Legendre condition R > 0 on [a, b] and
the strict Jacobi condition ζ > b are satisfied. Then hE ′′ (u)φ, φi > 0 for all
φ ∈ Φ which are not identically zero.
Proof. Exercise.
Remark. For systems we have some related definitions and results as for
scalar equations. A weak extremal is in C 2 [a, b] if
m
det fu′i u′k (x, u(x), u′ (x)) 6= 0
i,j=1
is called Legendre condition, and is called strict Legendre condition if the left
hand side is positive for all Rm \ {0}. As in the scalar case it follows from
E ′′ (u)(φ, φ) ≥ 0 for all φ ∈ V − V that the Legendre condition is satisfied,
see an exercise.
3.2. NECESSARY CONDITIONS 117
Example: Brachistochrone
x′
fx ′ = p √ = a, (3.7)
x′2 + y ′2 y − y1 + k
with a constant a. Suppose that P1 and P2 are not on a straight line parallel
to the y-axis, then a 6= 0. Let t = t(τ ) be the map defined by
x′ (t)
p = cos τ. (3.8)
x′ (t)2 + y ′ (t)2
1
y0 (τ ) − y1 + k = cos2 τ
a2
1
= (1 + cos(2τ )). (3.9)
2a2
Equation (3.9) implies that
x − x1 + β = ±α(u − sin u)
y − y1 + k = α(1 − cos u),
Consider the case where v1 = 0, P1 = (0, 0), and that P2 = (x2 , y2 ) satisfies
x2 > 0 and y2 > 0. Then
x = α(u − sin u)
y = α(1 − cos u),
x2 = α(u − sin u)
y2 = α(1 − cos u),
see an exercise.
Here is V = C 1 [a, b]
and h(α, β) is a sufficiently regular function. Let u be
a local minimizer, then for fixed φ ∈ V
E(u) ≤ E(u + ǫφ)
for all ǫ, |ǫ| < ǫ0 , ǫ0 sufficiently small,
Z b
fu (x, u, u′ )φ + fu′ (x, u, u′ )φ′ dx
a
+hα (u(a)), u(b))φ(a) + hβ (u(a), u(b))φ(b) = 0
for all φ ∈ V . Assume that u ∈ C 2 (a, b), then
Z b
d
fu − fu′ φ dx + [fu′ φ]ba (3.10)
a dx
+hα (u(a)), u(b))φ(a) + hβ (u(a), u(b))φ(b) = 0.
Since C01 (a, b) ⊂ V , it follows
Z b
d
fu − fu′ φ dx = 0
a dx
for all φ ∈ C01 (a, b), which implies that
d
fu − fu ′ = 0
dx
on (a, b). Then, from (3.10) we obtain
(fu′ φ)(b) − fu′ φ)(a) + hα (u(a)), u(b))φ(a) + hβ (u(a), u(b))φ(b) = 0
for all φ ∈ C 1 [a, b]. Choose a φ such that φ(b) = 0 and φ(a) = 1, it follows
fu′ = hα at x = a, and take then a φ such that φ(b) = 1, φ(a) = 0, we
obtain fu′ = −hβ at x = b.
These boundary conditions are called free boundary conditions. These
conditions are not prescribed, they result from the property that u is a
minimizer of he associated energy functional.
120 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Proof. Exercise.
and let V = C m [a, b]be the set of the admissible comparison functions.
That is, no boundary conditions are prescribed. From u ∈ V : E(u) ≤ E(v)
for all v ∈ V , ||v − u||C[a,b] < ǫ for an ǫ > 0, it follows the weak Euler
equation
Z bXm
fu(k) (x, u(x), ...u(m) (x))φ(k) dx = 0
a k=0
for all φ ∈ C m [a, b]. Assume that u ∈ C 2m [a, b], which is a regularity
assumption on u, it follows by integration by parts the differential equation
m
X
(−1)k (fu(k) )(k) = 0
k=0
on (a, b) and the free boundary conditions (ql )(a) = 0, (ql )(b) = 0, l =
0, ..., m − 1, where
m−l
X
ql = (−1)k−1 (fu(k+l) )(k) .
k=1
Proof. Exercise.
Consider the energy functional, see [33] for the related physics,
Z l Z l
1 ′′ 2
J(v) = EI (v (x)) dx − f (x)v(x) dx,
2 0 0
3.2. NECESSARY CONDITIONS 121
and the prescribed and free boundary conditions depend on how the beam
is supported, see the related figures.
(i) (ii)
(iii) (iv)
(iii) Clamped at one end and simply supported at the other end. Prescribed
conditions: u(0) = u′ (0) = 0, u(l) = 0,
free boundary condition: u′′ (l) = 0.
(iv) Clamped at one end, no prescribed conditions at the other end. Pre-
scribed conditions: u(0) = u′ (0) = 0,
free boundary conditions: u′′ (l) = 0, u′′′ (l) = 0.
Proof. Exercise.
f given and sufficiently regular. Set fv = (fv1 , fv2 ) and fv′ = (fv1′ , fv2′ ).
Proof. Let τ0 such that u(1) = γ(τ0 ). Since E(u) ≤ E(v) for all v ∈ V0 ,
where
V0 = {v ∈ V : v(0) = P, v(1) = u(1)},
it follows the system of Euler equations
d
fu − fu ′ = 0
dt
in (0, 1). The transversality condition is a consequence of variations along
the target curve γ, see Figure 3.2. There is a family v(t, τ ) of curves such
that v ∈ C 1 (D), where D = (0, 1) × (τ0 − ǫ0 , τ0 + ǫ0 ) for an ǫ0 > 0, and
x2 γ
. γ(τ)
. γ(τ )
v
0
u
x
1
Set g(τ ) = E(v). Since g(τ0 ) ≤ g(τ ), |τ − τ0 | < ǫ0 , it follows that g ′ (τ0 ) = 0.
Consequently
Z 1
fu · vτ (t, τ0 ) + fu′ · vτ′ (t, τ0 ) dt = 0,
0
Since the system of Euler differential equations is satisfied and since v(0, τ ) =
P , |τ − τ0 | < ǫ, it follows
x2 γ
2
γ
1
x
1
x3
S
.
P x2
x
1
Proof. Exercise.
Remark 2. The result of the theorem and of the above remark hold in Rn .
Proof. Exercise.
Remark 3. Consider the case R3 and let the target be a sufficiently regular
surface S, see Figure 3.4, then the transversality condition is fu′ ⊥ S.
Proof. Exercise.
3.2. NECESSARY CONDITIONS 125
1
2n
...
1 1
2n
are smooth, that is they are at least in C 2 . In general, it can happen that
extremals have corners or edges, respectively, even if the integrands are
analytically in their arguments.
v2
.u(t )
0
v1
Proof. Let η > 0 small enough such that there is no further corner of the
extremal in (t0 − η, t0 + η). Then for all φ ∈ C01 (t0 − η, t0 + η) we have, where
a = t0 − η and b = t0 + η,
Z b
0 = fu · φ + fu′ · φ′ dt
a
Z t0 Z b
= fu · φ + fu′ · φ′ dt + fu · φ + fu′ · φ′ dt
a t0
Z t0 t 0
d
= fu − fu′ · φ dt + fu′ · φ
a dt a
Z b b
d
+ fu − fu′ · φ dt + fu′ · φ
t dt t0
0
= − fu′ (t0 ) · φ(t0 )
3.2. NECESSARY CONDITIONS 127
x
a c b
then [Fx′ ] (c) = 0 and Fy′ (c) = 0, which are the equations of the corollary.
The following consideration is a justification of that argument. Let u be
a minimizer of E(v) in V . For fixed φ1 , φ2 ∈ C01 (a, b) set
where t ∈ [a, b], |ǫ| < ǫ0 , ǫ0 sufficiently small. Then x defines a C 1 diffeo-
morphism from [a, b] onto [a, b] and x′ 6= 0 for each ǫ, |ǫ| < ǫ0 . Here we set
x′ = xt (ǫ; t). Let t = t(ǫ; x) be the inverse of the first of the two equations
above, and set
Y (ǫ; x) = y(ǫ; t(ǫ; x)).
Then Y (ǫ; x) defines a C 1 [a, b] graph, i. e., Y ∈ V , and
Z b Z b
′
f (x, u(x), u (x)) dx ≤ f (x, Y (ǫ; x), Y ′ (ǫ; x)) dx
a a
Z b
y ′ (ǫ; t)
= f x(ǫ; t), y(ǫ; t), ′ x′ (ǫ; t) dt
a x (ǫ, t)
= : g(ǫ).
Since g(0) ≤ g(ǫ), |ǫ| < ǫ0 , it follows g ′ (0) = 0 which implies the conditions
of the corollary. ✷
Remark. The first condition of the corollary follows also by a direct applica-
tion of the argument of the proof of Theorem 3.2.12. The second condition
is a consequence of using a family of diffeomorphism of the fixed interval
[a, b], which are called sometimes “inner variations”.
ξ ξ ξ
1 2
Set
ξ1 = (u′ )− ≡ u′ (c − 0), ξ2 = (u′ )+ ≡ u′ (c + 0)
η1 = f − ≡ f (c, u(c)u′ (c − 0)), η2 = f − ≡ f (c, u(c)u′ (c + 0))
and
fu−′ = fu′ (c, u(c), (u′ )− ), fu+′ = fu′ (c, u(c), (u′ )+ ).
Then the two tangent lines are given by
η − f − = fu−′ (ξ − (u′ )− )
η − f + = fu+′ (ξ − (u′ )+ ).
From the first condition of the corollary we see that the tangent lines must
be parallel, then the second condition implies that the lines coincides, see
Figure 3.8.
Suppose that h(ξ) = f (x, u, ξ) is strongly convex or strongly concave for all
(x, u) ∈ [a, b] × R, then there are no corners of extremals.
Proof. If not, then there are ξ1 6= ξ2 which implies the situation shown in
Figure 3.8. ✷
all v ′ , then corners can occur as the example f = (v ′2 − 1)2 shows, see
Figure 3.5.
v = (v1 , . . . , vn ). Let
Set Z b
E(v) = v(x) dx
a
and
V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub , g(v) = L},
where Z b p
g(v) = 1 + v ′2 (x) dx
a
is the given length L of the curve defined by v. We assume that
p
c > (b − a)2 + (ub − ua )2 .
Then we consider the problem maxv∈V E(v) of maximizing the area |Ω|
between the x-axis and the curve defined by v ∈ V , see Figure 3.9.
3.2. NECESSARY CONDITIONS 131
.
. Ω
a b x
This problem is a special case of a more general problem, see Section 1.3.3.
It is also a problem which is governed by a partial differential equation, but
it is covered by the Lagrange multiplier rule below. Consider a capillary tube
with a bottom and filled partially with a liquid. The gravity g is directed
downwards in direction of the negative x3 -axis. The interface S, which
separates the liquid from the vapor, is defined by x3 = v(x), x = (x1 , x2 ),
see Figure 3.10. Set
Z
1
V = v ∈ C (Ω) : v dx = const. ,
Ω
A large class of problems fit into the following framework. Suppose that
E : B 7→ R and gj : B 7→ R, j = 1, . . . , m. We recall that B is a real Banach
space and H a real Hilbert space such that B ⊂ H is continuously embedded:
||v||H ≤ c||v||B for all v ∈ B. Moreover, we suppose that ||v||B 6= 0 implies
||v||H 6= 0 for v ∈ B, that is, B ⊂ H is injectively embedded.
x3
g
vapor
NΓ
S
NS
liquid
x2 Ω
x1
m
X
F (c) = E(u + c j φj )
j=1
m
X
Gi (c) = gi (u + c j φj )
j=1
are in C 1 in a neighborhood of c = 0, c ∈ Rm .
Set
V = {v ∈ B : gi (v) = 0, i = 1, . . . , m}.
there exist m + 1 real numbers, not all of them are zero, such that
m
X
′
λ0 E (u) + λi gi′ (u) = 0B ∗ .
i=1
Set A(c, η) = (F (c)−M −η, G1 (c), . . . , Gm (c))T , then we can write the above
system as A(c, η) = 0m+1 . We have A(0, 0) = 0m+1 , and, if the functionals
l0 , . . . , lm are linearly independent, that the m×m-matrix Ac (0, 0) is regular.
From the implicit function theorem we obtain that there exists an η0 > 0
and a C 1 (−η0 , η0 ) function c(η) such that c(0) = 0 and A(c(η), η) ≡ 0 on
−η0 < η < η0 . Then we take an η < 0 from this interval and obtain a
contradiction to the assumption that u is local minimizer of E in V , if u is
a maximizer, then we choose a positive η. ✷
Set
m
X
F (x, v, v ′ , λ) = f (x, v, v ′ ) + λk lk (x, v).
k=1
d
Fu′ = Fu
dx
on (a, b).
where |ǫ| < ǫ0 , ǫ0 sufficiently small, and consider on [a, b] the system
∂wl
ηl (x) = .
∂ǫ ǫ=0
3.2. NECESSARY CONDITIONS 135
Set Z b
h(ǫ) = f (x, w(x, ǫ), w′ (x, ǫ)) dx.
a
Since h′ (0) = 0, we see that
Z n
!
b X
ful ηl + fu′l ηl′ dx = 0. (3.14)
a l=1
P2
.
.
P1
1 x
since v − u ∈ T (V, u) if v ∈ V .
Let
where φ ∈ C 1 [0, 1] is given and satisfies φ(0) ≤ 0 and φ(1) ≤ 0, see Fig-
ure 3.12. Set Z 1
2
E(v) = v ′ (x) dx,
0
138 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
The following problem was studied by Link [32].2 Consider a simply sup-
ported beam compressed by a force P along the negative x-axis, where the
deflections are restricted by, say, a parallel line to the x-axis, see Figure 3.13.
It turns out that u(k; x) defines a local minimizer of the associated energy
functional Z l Z
1 ′′ 2 P l ′ 2
J(v) = EI v (x) dx − v (x) dx,
2 0 2 0
where EI is a positive constant (bending stiffness), in the set
Remark. Related problems for the circular plate and the rectangular plate
were studied in [43, 44], where explicit stability bounds were calculated.
2
I would like to thank Rolf Klötzler for showing me this problem.
3.2. NECESSARY CONDITIONS 139
u(k;x)
d
P
k l−k l x
We suppose that p ∈ C 1 [a, b], q, ρ ∈ C[a, b], and that p, q and ρ are positive
on the finite interval [a, b]. Set
Z b
a(u, v) = p(x)u′ (x)v ′ (x) + q(x)u(x)v(x) dx
a
Z b
b(u, v) = ρ(x)u(x)v(x) dx.
a
where H = H01 (a, b). Then we ask whether or not the associated eigenfunc-
tion does not change sign in (a, b). In our case of this second order problem
for an ordinary differential equation it can be easily shown that each eigen-
value is simple. Instead of looking for minimizers of the above Rayleigh
quotient in H, we consider the problem directly in the set of non-negative
functions. Define the closed convex cone with vertex at the origin
Let
a(v, v)
λK = min .
v∈K\{0} b(v, v)
140 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Proposition. λH = λK .
Remark. One can apply this idea to more general eigenvalue problems.
In particular, it can be shown that the first eigenvalue of a convex simply
supported plate is simple and the associated eigenfunction has no zero inside
of the convex domain Ω ⊂ R2 , see [39]. Plate problems are governed by forth
order elliptic equations.
3.2. NECESSARY CONDITIONS 141
3.2.9 Exercises
1. Consider the example ”How much should a nation save?” Show that
Z T
hE ′′ (K)ζ, ζi ≡ (FKK ζ 2 + 2FKK ′ ζζ ′ + FK ′ K ′ ζ ′2 ) dt
0
≤ 0
7. Show that a solution of the weak Euler equation of the vector valued
variational problem, see Section 3.2.2, is in C 2 [a, b] if
m
det fu′i u′k (x, u(x), u′ (x)) 6= 0
i,k=1
on [a, b].
8. Consider the case of a system, see Section 3.2.2, and show that
hE ′′ (u)φ, φi ≥ 0
11. Find the shortest distance between the x-axis and the straight line
defined by x + y + z = 1 and x − y + z = 2.
12. Find the the shortest distance between the origin and the surface (ro-
tational paraboloid) defined by z = 1 − x2 − y 2 .
13. Let Z t2 p
E(v) = g(v(t)) v ′ (t) · v ′ (t) dt,
t1
Find λ.
P1 , P2 ∈ E ∩ {x ∈ R3 : x1 = 0}.
18. Set
vǫ = v + ǫζδh2 v,
where
Z b
a(u)(φ, φ) = Rφ′2 dx,
a
Z b
b(u)(φ, φ) = − 2Qφφ′ + P φ2 dx.
a
with
d
S=P−
Q.
dx
If the strict Legendre condition is satisfied on [a, b], then u ∈ C 2 [a, b] and
the quadratic form a(u)(φ, φ) is equivalent to a norm on H = H01 (a, b) and
b(u)(φ, φ) is a completely continuous form on H.
Proof. Assumption (ii) implies that hE ′′ (u)φ, φi ≥ 0 for all φ ∈ H01 (a, b). If
u is no strict local minimizer, then we will show that there is a φ0 ∈ H01 (a, b),
φ0 6= 0 such that hE ′′ (u)φ0 , φ0 i = 0. Thus φ0 is a solution of Jacobi equation
hE ′′ (u)φ0 , ψi = 0 for all ψ ∈ H01 (a, b). A regularity argument, we omit the
proof here since this problem is addressed in another course, shows that
φ ∈ V − V . The idea of proof is a purely variational argument. We insert
for ψ the admissible function ζ(x)ψ −h (x), where ζ is a sufficiently regular
cut off function and ψ −h is the backward difference quotient. After some
calculation one can show that φ0 ∈ H 2 (a, b) which implies that φ0 ∈ C 1 [a, b].
Set B = C 1 [a, b] and H = H01 (a, b). If u is no strict local minimizer, then
there is a sequence un → u in B, un 6= u in B, such that
0 ≥ a(u)(un − u, un − u) − b(u)(un − u, un − u)
+η(||un − u||B )||un − u||2H .
0 ≥ a(u)(wn , wn ) − b(u)(wn , wn )
+η(||un − u||B )||wn ||2H .
−1 b(u)(v, v)
λ+
1 = max .
v∈H\{0} a(u)(v, v)
V = {v ∈ H : a(u)(v, v) ≤ 1},
(ii) Suppose that there is no positive eigenvalue or that the lowest positive
eigenvalue satisfies λ+
1 > 1.
a(u)(v, v) − b(u)(v, v) ≥ 0
for all v ∈ H \ {0}. If not, then we have for a w ∈ H \ {0} that a(u)(w, w) =
b(u)(w, w). Thus a(u)(w, w) ≤ 0, which implies that w = 0.
(iib) Suppose that there is a w ∈ H such that b(u)(w, w) > 0. Then there is
at least one positive eigenvalue and the lowest positive eigenvalue satisfies,
see the lemma above,
−1 b(u)(v, v)
λ+
1 ≥
a(u)(v, v)
for all v ∈ H \ {0}. According to the assumption there is a positive ǫ such
that −1
1 − ǫ = λ+1 .
It follows that
a(u)(v, v) − b(u)(v, v) ≥ ǫa(u)(v, v)
for all v ∈ H. ✷
Remark. In general, the lowest positive eigenvalue λ+1 is not known explic-
itly. Thus, the above theorem leads to an important problem in the calculus
of variations: find lower bounds of λ+
1.
suppose that hE ′′ (u)v, hi and hgi′′ (u)v, hi are bilinear forms on B × B which
has continuous extensions to symmetric, bounded bilinear forms on H × H.
Rb
EXAMPLE: B = C 1 [a, b], H = H 1 (a, b) and E(v) = a (v ′ (x))2 dx, then
Z b Z b
′ ′ 1
E(u + h) = E(u) + u (x)h (x) dx + h′ (x)h′ (x) dx.
a 2 a
Let
V = {v ∈ B : gi (v) = 0, i = 1, . . . , m}
and assume
where a(u, λ0 )(v, h) and b(u, λ0 )(v, h) are bounded bilinear symmetric forms
on H × H, a(u, λ0 )(v, v) is non-negative on H and
1/2
(a) a(u, λ0 )(v, v) is equivalent to a norm on H,
Set −1/2
tn = a(u, λ0 )(un − u, un − u)
and wn = tn (un − u). Then
0 ≥ 1 − b(u, λ0 )(w, w)
and
a(u, λ0 )(w, w) ≤ 1.
Summarizing, we arrive at
for a w 6= 0 satisfying
hgj′ (u), wi = 0, j = 1, . . . , m.
Proof. Exercise.
see Theorem 3.2.1. For the definition of T (V, u) see Section 3.1. We recall
that we always suppose that u is not isolated in V .
For given u ∈ B we assume
1
E(u + h) = E(u) + hE ′ (u), hi + hE ′′ (u)h, hi + η(||h||B )||h||2H ,
2
where hE ′ (u), hi is a bounded linear functional on B which admits an ex-
tension to a bounded linear functional on H, and hE ′′ (u)v, hi is a bilinear
form on B × B which has a continuous extensions to a symmetric, bounded
bilinear form on H × H. Moreover, we suppose that
Definition. Let TE ′ (V, u) be the set of all w ∈ T (V, u) such that, if un and
tn = ||un − u||−1
H are associated sequences to w, then
Proof. Exercise.
E(u) ≥ E(u + un − u)
= E(u) + hE ′ (u), un − ui
1
+ [(a(u)(un − u, un − u) − b(u)(un − u, un − u)]
2
+η(||un − u||B )||un − u||2H .
Set
tn = (a(u)(un − u, un − u))−1/2
and wn = tn (un − u). Then
1
0 ≥ tn hE ′ (u), wn i + [1 − b(u)(wn , wn )] + η(||un − u||B )||wn ||2H , (3.21)
2
which implies that
which implies that w0 6= 0. Since the square of the norm on a real Hilbert
space defines a weakly lower semicontinuous functional, we have
a(u)(w0 , w0 ) ≤ 1.
Remark. Assumption
It follows immediately
Consider the example ”A unilateral problem for the beam” of Section 3.2.8,
see Figure 3.13. Set
Z l
a(u, v) = u′′ (x)v ′′ (x) dx
0
Z l
b(u, v) = u′ (x)v ′ (x) dx
0
and
1 λ
E(v, λ) = a(v, v) − b(v, v),
2 2
where u, v ∈ V and λ = P/(EI). The family of functions
√ √
d
λx + sin( λx : 0≤x<k
π
u = u(k; x) = √ d
: k ≤x≤l−k ,
√
d λ(1 − x) + sin( λ(1 − x) : l−k ≤x<l
π
Proposition. Suppose that l/4 < k ≤ l/2, then u = u(k; x) is a strict local
minimizer of E in V , that is, there is a ρ > 0 such that E(u, λ) < E(v, λ)
for all v ∈ V satisfying 0 < ||u − v||H 2 (0,l) < ρ, and u is no local minimizer
if k < l/4.
see the following lemma. We show, see Theorem 3.3.5 and the second Corol-
lary, that a(v, v) − λb(v, v) > 0 for all v ∈ L(k) \ {0} if l/4 < k ≤ l/2. We
recall that λ = (π/k)2 . Consider the eigenvalue problem
In fact, this problem splits into three problems if 0 < k < l/2. The first one
is to find the lowest eigenvalue of a compressed beam of length l − 2k which
is clamped at both ends, and the other two consist in finding the lowest
eigenvalue of the compressed beam of length k which is simply supported
at one end and clamped at the other end. Thus the lowest eigenvalue µ1
of (3.22) is ( )
τ 2 2π 2
1
µ1 = min , ,
k l − 2k
where τ1 = 4.4934... is the lowest positive zero of tan x = x. Then u is a
strict weak local minimizer if
π 2
λ= < µ1
k
is satisfied. This inequality is satisfied if and only if l/4 < k ≤ l/2.
If 0 < k < l/4, then u(k; x) is no local minimizer since E(u+w; λ) < E(u; λ),
where w(x), k ≤ x ≤ l−k, is the negative eigenfunction to the first eigenvalue
of the compressed beam of length l−2k which is clamped at both ends x = k
and x = l − k. On 0 ≤ x ≤ k and on l − k ≤ x ≤ l we set w(x) = 0. ✷
It remains to show
We have
wn (k) ≤ 0 and wn (l − k) ≤ 0,
where wn = tn (un − u), and
with
A1 = u′′′ (k − 1), A2 = u′′′ (l − k + 0)
156 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
ǫn = t−1
n ≡ ||un − u||H ,
then
−A1 wn (k) − A2 wn (l − k) ≤ c1 ǫn . (3.24)
It follows that wn (k), wn (l − k) → 0 as n → ∞, which implies
w′ (k) = 0 and w′ (l − k) = 0.
wn (k + h) = wn (k) + wn′ (k)h + O |h|3/2 ,
1
u(k + h) = u(k) + u′ (k)h + u′′ (k)h2 + O |h|5/2
2
5/2
= d + O |h|
3
Let x ± h ∈ (0, l) and v ∈ H m (0, l). Then
1
v(x + h) = v(x) + v ′ (x)h + . . . v (m−1) (x)hm−1 + Rm ,
(m − 1)!
where Z h
1
Rm = (h − t)m−1 v (m) (x + t) dt
(m − 1)! 0
which satisfies 2m−1
|Rm | ≤ c||v (m) ||H m (0,l) |h| 2 .
3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 157
which implies
3/2 ! 5/2
2αn 2αn
ǫn αn − c 2 ≤ c3 .
β β
Consequently
ǫn ≤ c4 αn3/2 (3.27)
for all n ≥ n0 , n0 sufficiently large. Combining this inequality with inequal-
ity (3.24), we find
A1 αn ≤ c1 c4 αn3/2
which is a contradiction to A1 > 0. Consequently wn′ (k) → 0. Thus w′ (k) =
0. We repeat the above consideration at x = l−k and obtain that w′ (l−k) =
0. ✷
158 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
3.3.4 Exercises
1. Show that the square of the norm on a Hilbert space defines a weakly
lower semicontinuous functional.
1
E(u + h) = E(u) + hE ′ (u), hi + hE ′′ (u)h, hi + η(||h||B )||h||2H ,
2
where
Z b
′
hE (u), hi = [fu (x, u, u′ )h + fu′ (x, u, u′ )h′ ] dx
a
Z b
hE ′′ (u)h, hi = [fuu (x, u, u′ )h2 + 2fuu′ (x, u, u′ )hh′
a
+fu′ u′ (x, u, u′ )h′2 ] dx,
Rb
Hint: Set g(ǫ) = a f (x, u + ǫh, u′ + ǫh′ ) dx. Then
1 1
g(1) = g(0) + g ′ (0) + g ′′ (0) + [g ′′ (δ) − g ′′ (0)],
2 2
where 0 < δ < 1.
3. Set Z b
a(u)(h, h) = fu′ u′ (x, u, u′ )h′ (x)2 dx,
a
if φn ⇁ φ and ψl ⇁ ψ in H.
Hint: (i) The sequences {φn }, {ψl } are bounded in H 1 (a, b).
(ii) The sequences {φn }, {ψl } are equicontinuous sequences.
(iii) Use the Arzela-Ascoli Theorem.
and for v ∈ V Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a
where f (x, y, p) is a given and sufficiently regular function f : [a, b]×R×R 7→
R.
We recall that u ∈ V is called a weak local minimizer of E in V if there
exists a ρ > 0 such that E(u) ≤ E(v) for all v ∈ V satisfying ||v −u||C1 [a,b] <
ρ. And u ∈ V is said to be a strong local minimizer of E in V if there exists
a ρ > 0 such that E(u) ≤ E(v) for all v ∈ V satisfying ||v − u||C[a,b] < ρ.
If the strict Legendre condition is satisfied on [a, b], then u ∈ C 2 [a, b],that
is, u is a solution of the Euler equation. Assume
d
fv′ (x, v, v ′ ) = fv (x, v, v ′ ) on (a, b) (3.28)
dx
v(a) = ua + µ (3.29)
v(b) = ub + µ, (3.30)
by
d
− dx fv′ (x, v, v ′ ) + fv (x, v, v ′ )
M (v, µ) = v(a) − ua − µ .
v(b) − ub − µ
d
− [R(x, u, u′ ) v ′ ] + S(x, u, u′ ) v = 0 on (a, b)
dx
v(a) = 1
v(b) = 1.
162 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
The solution v1 is positive on [a, b]. To show this, we set ζ(x) = max{−v1 (x), 0},
then
Z b
0 = [−(Rv ′ )′ ζ + Svζ] dx
a
Z b
= (Rv ′ ζ ′ + Svζ) dx
a
Z
= − (Rζ ′2 + Sζ 2 ) dx
{−v1 (x)>0}
Z b
= − (Rζ + Sζ 2 ) dx.
′2
a
It follows that ζ = 0, that is, v1 (x) ≥ 0 on [a, b]. Assume there is a zero
x0 ∈ (a, b) of v1 (x), then v ′ (x0 ) = 0. Consequently v(x) ≡ 0 on [a, b], which
contradicts the boundary conditions.
Finally, for given (x, y) ∈ Dδ (u), δ > 0 sufficiently small, there exists a
unique solution µ = µ(x, y) of v(x, µ) − y = 0 since vµ (x, 0) = v1 (x) > 0 on
[a, b]. ✷
Let v(x, µ) be the solution of (3.28)–(3.31). Set F (x, µ) = v ′ (x, µ). From
the previous lemma we have that for given (x, y) ∈ Dδ (u) there exists a
unique µ = µ(x, y) which defines the curve of the foliation which contains
(x, y). Set
Φ(x, y) = F (x, µ(x, y)),
and consider the vector field A = (Q, −P ), where
P (x, y) = f (x, y, Φ(x, y)) − Φ(x, y)fp (x, y, Φ(x, y))
Q(x, y) = fp (x, y, Φ(x, y)).
Thus
Py − Qx = fy − Φ(fpy + fpp Φy ) − fpx − fpp Φx .
The right hand side is zero in Dδ (u). To show this, let (x0 , y0 ) ∈ Dδ (u) be
given, and consider the curve of the foliation defined by
We recall that
if y(x) defines a curve of the foliation, since c(x, y) =const. along this curve.
Then
y ′ (x0 ) = Φ(x0 , y0 )
y ′′ (x0 ) = Φx (x0 , y0 ) + Φy (x0 , y0 )y ′ (x0 )
= Φx (x0 , y0 ) + Φy (x0 , y0 )Φ(x0 , y0 ).
Inserting y ′ (x0 ) and y ′′ (x0 ) from above into the Euler equation, which is
satisfied along every curve of the foliation,
since
Z b
U (u) = [f (x, u(x), Φ(x, u(x)) − Φ(x, u(x))fp (x, u(x), φ(x, u(x))
a
+fp (x, u(x), u′ (x))u′ (x)] dx
Z b
= f (x, u(x), u′ (x)) dx.
a
Corollary. Suppose that fpp (x, y, p) ≥ 0 for all (x, y) ∈ Dδ (u) and for all
p ∈ R, then E ≥ 0 in Dδ (u) × R × R.
Proof. Let v ∈ V ∩ Dδ (u). From equation (3.32) and Lemma 3.4.2 we see
that
E(v) − E(u) = E(v) − U (u)
= E(v) − U (v)
Z b
= [f (x, v, v ′ ) − f (x, v, Φ(x, v)
a
+(Φ(x, v) − v ′ )fp (x, v, Φ(x, v))] dx
Z b
= E(x, v(x), Φ(x, v), v ′ ) dx
a
≥ 0.
✷
3.4.1 Exercises
Rb
1. Suppose that a (Rφ′2 + Sφ2 ) dx > 0 for all φ ∈ (V − V ) \ {0}.
Show that there exists a unique solution of the Jacobi boundary value
problem
d
− [R(x, u, u′ ) v ′ ] + S(x, u, u′ ) v = g(x) on (a, b)
dx
v(a) = va
v(b) = vb ,
2. Show that the solution u of the above example defines a global strong
minimizer of E in V .
166 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
f : [t0 , t1 ] × Rn × Rm 7→ Rn .
where
f 0 : [t0 , t1 ] × Rn × Rm 7→ R,
v ∈ Uad and y(t) is the solution of the above boundary value problem with
the control v.
(iii) p0 = 1 or p0 = 0.
Remarks. (i) In the case that we do not prescribe the endpoint x1 , which
is called the free endpoint case, then we have to add in (i) the additional
endpoint condition p(t1 ) = 0. For this case of a free endpoint there is an
elementary proof of the Pontryagin maximum principle, see below.
(ii) If the endpoint condition is y(t1 ) ≥ x1 , then p(t1 ) ≥ 0, and if the optimal
trajectory satisfies x(t1 ) > 0, then p(t1 ) = 0.
168 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
3.5.2 Examples
This example was taken from [55], pp. 78. Suppose that the admissible set
of controls is V = [0, 1], the cost functional is given by
Z T
E(v) = U (1 − v(t)) dt,
0
and the differential equation and the boundary conditions which defines the
trajectory y : [0, T ] 7→ R are
H = p0 U (1 − v) + pv,
p′ (t) = −Hx .
Since H does not depend on y we find that p(t) = c = const. and c ≥ 0, see
the second remark above. Thus, if u(t) is an optimal control then we have
in regular points t the inequality
This example was taken from [55], pp. 222. Suppose that we have two
admissible sets of controls u(t) ∈ U = [a, b], 0 < a < b < 1, and v(t) ∈ V =
170 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
We are looking for a Nash equilibrium, that is, for piecewise continuous
controls u∗ (t) ∈ U , v ∗ (t) ∈ V such that
W (u∗ , v ∗ ) ≥ W (u, v ∗ ) for all u ∈ U
C(u∗ , v ∗ ) ≥ C(u∗ , v) for all v ∈ V.
Suppose there exists a Nash equilibrium and p0 = 1, then the associated
Hamilton functions are given by:
HW = uK + pv ∗ (t)(1 − u)K
HC = (1 − v)(1 − u∗ (t))K + qv(1 − u∗ (t))K,
where p and q are the associate adjoint functions.
then
where ǫ > 0 is sufficiently small and v ∈ U . Let xǫ (t), see Figure 3.15, be
z=v
z=u(t)
t0 τ−ε τ t1 t
xε (t)
x(t)
t0 τ−ε τ t1 t
Since
xǫ (τ − ǫ) − xǫ (τ ) = −x′ǫ (τ )ǫ + o(ǫ)
x(τ − ǫ) − x(τ ) = −x′ (τ )ǫ + o(ǫ)
as ǫ → 0, it follows that
then the changed trajectory xǫ (t) on τ < t < t1 is given by, see an exercise,
where
fx11 · · · fx1n
fx = . . . . . . . . . . . . . .
fxn1 · · · fxnn
174 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
We recall that Θ(t, t) = I, I the identity matrix, and w(t) = Θ(t, τ ) a is the
solution of the initial value problem w′ (t) = fx w(t), w(τ ) = a. We have
Z τ
E(uǫ ) − E(u) = [f 0 (t, xǫ (t), v) − f 0 (t, x(t), u(t))] dt
τ −ǫ
Z t1
+ [f 0 (t, xǫ (t), u(t)) − f 0 (t, x(t), u(t))] dt
τ
= ǫ[f 0 (t∗ , xǫ (t∗ ), v)) − f 0 (t∗ , x(t∗ ), u(t∗ ))]
Z t1
+ hfx0 (t, , x(t), u(t)), xǫ (t) − x(t)i
τ
+o(|xǫ (t) − x(t)|) dt
= ǫ[f 0 (t∗ , xǫ (t∗ ), v)) − f 0 (t∗ , x(t∗ ), u(t∗ ))]
Z t1
+ǫ hfx0 (t, x(t), u(t)), y(t)i dt + o(ǫ),
τ
t∗
where ∈ [τ − ǫ, τ ] and y(t) := Θ(t, τ )w(τ, v).
From the assumption E(uǫ ) ≤ E(u), ǫ > 0, it follows
E(uǫ ) − E(u)
lim ≤ 0.
ǫ→0 ǫ
Combining this inequality with the previous expansion of E(uǫ ) − E(u), we
obtain
Z t1
0 0
f (τ, x(τ ), v)) − f (τ, x(τ ), u(τ )) + hfx0 (t, x(t), u(t)), y(t)i dt ≤ 0
τ
for every regular τ ∈ (t0 , t1 ). Then the theorem follows from the formula
Z t1
hfx0 (t, x(t), u(t)), y(t)i dt = hf (τ, x(τ ), u(τ )), p(τ )i−hf (τ, x(τ ), v), p(τ )i,
τ
(3.41)
where p(t) is the solution of the initial value problem
p′ (t) = −fxT p − fx0 , p(t1 ) = 0.
Formula (3.41) is a consequence of
d
hp(t), y(t)i = hp′ (t), y(t)i + hp(t), y ′ (t)i
dt
= −hfxT p, yi − hfx0 , yi + hp, fx yi
= −hfx0 , yi.
3.5. OPTIMAL CONTROL 175
which implies that yn+1 (t) = t. The second new coordinate is defined by
where y = (y1 , . . . , yn ).
Let
P = (p0 , p, pn+1 ), p = (p1 , . . . , pn ).
Define the Hamilton function by
H(x(t), xn+1 (t), u(t), P (t)) ≥ H(x(t), xn+1 (t), v, P (t)) (3.46)
v1
v2
t0 τ 1−ε a 1 τ1 τ 2 −εa 2 τ2 t1
system
W ′ (t) = A(t) W (t), (3.48)
3.5. OPTIMAL CONTROL 177
where
fx00 · · · fx0n+1
A = ................. .
fxn+1
0
· · · fxn+1
n+1
Define for an s-tuple z = (τ1 , . . . , τs ), t0 < τi < t1 , τi are different from each
other, and for v = (v1 , . . . , vs ), where vl ∈ U , the set
s
X
C(z, v) = {Y ∈ Rn+2 : Y = ai Θ(t1 , τi )W (τi , vi ), ai > 0}.
i=1
This set is a convex cone with vertex at the origin. Denote by Z(s) the set
of all s-tuples z from above and let V (s) be the set of all s-tuples v such
that the coordinates are in U . Define the set
C = ∪∞
s=1 ∪z∈Z(s),v∈V (s) C(z, v).
This set is a convex cone in Rn+2 with vertex at the origin, see an exercise.
Consider the ray L = r e0 , r > 0, where e0 = (1, 0, . . . , 0) ∈ Rn+2 . If
L is not in the interior of C, then we will prove the maximum principle by
using separation results for convex sets. If L is in the interior of C, then
we are led to a contradiction to the assumption that u is optimal. In this
case we will show by using Brouwer’s fixed point theorem that there is an
admissible needle variation which produces an associated trajectory Xǫ (t)
such that the first coordinate satisfies x0,ǫ (t1 ) > x0 (t1 ), where x0 (t) is the
first coordinate of the trajectory X(t) associated to the optimal control u(t).
Case (i). L is not in the interior of C. From Theorem 2.6.1 and Theorem 2.6.2
it follows, see two of the following exercises, that there exists a vector P1 ∈
Rn+2 , P1 6= 0, such that
e0 Π
P1
where v ∈ U . Then
ǫΘ(t1 , t)W (t, v) ∈ C,
where ǫ > 0. Then, see (3.50),
see Figure 3.18. Let λi (z) are the (uniquely determined) barycentric coor-
L
3
A2
A
z0 . A1
dinates4 of z ∈ S, then
n+2
X
z= λi (z)Ai .
i=1
4
Barycentric coordinates
Pm λ0 , . . . , λm of z ∈ Rm are called the real numbers in the
l Pm 0 m
representation z = l=0 λl x , l=0 λl = 1, where x , . . . , x ∈ Rm are given and the
l 0
m vectors x − x , l = 1, . . . , m are linearly independent.
P The m-dimensionalPmsimplex S
defined by suchPvectors is the set of all z =P m l
l=0 λl x , where λl ≥ 1 and l=0 λl = 1.
Thus z − x0 = m l
l=1 λl x , where λl ≥ 0 and
m
l=1 λl ≤ 1.
180 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Since Ai are continuous with respect to τli , which are all regular by assump-
tion, we can assume that all τli are different from each other. Then the
associated perturbed trajectory at t = t1 is given by
n+2
X
Xǫ (t1 ) = X(t1 ) + ǫ [λi + bi (λ, ǫ)]Ai ,
i=1
3.5.5 Exercises
1. Consider the ”Two sector model”, see [55], p. 89. Suppose that the
admissible set of controls is V = [0, 1], the cost functional is given by
Z T
E(v) = y2 (t) dt,
0
and the differential equations and the side conditions which define the
trajectory y : [0, T ] 7→ R2 are
where a is positive constant and the initial data y10 and y20 are given.
2. Consider a model for ”Growth that pollutes”, see [55], pp. 92. Suppose
that the admissible set of controls is V = [0, 1], the cost functional is
given by
Z T
E(v) = [(1 − v(t))y1 (t) − by2 (t)] dt,
0
where the data y10 and y20 and y2T are given.
4. Show that the solution xǫ of the initial value problem, see the proof of
Pontryagin’s maximum principle in the case of a free endpoint,
satisfies
xǫ (t) = x(t) + ǫΘ(t, τ )w(τ, v).
Hint: z := (∂xǫ /∂ǫ) ǫ=0 is the solution of the initial value problem
z ′ (t) = fx z(t), z(τ ) = w(τ, v).
5. Show that the mapping Mv (u, 0), see the proof of Lemma 3.4.1, is
regular.
defined by
y ′ (t) − f (t, y(t)
M (y, µ) = ,
y(τ ) − a − µ
and apply an implicit function theorem, see for example [28].
9. Show that the set C defined in Section 3.5.4 is a convex cone with
vertex at the origin.
11. Define the fundamental matrix for the linear system Y ′ (t) = A(t) Y (t),
where A(t) is a piecewise continuous quadratic matrix.
12. See [29], pp. 10. Let K ⊂ Rn be compact and convex and let F :
K 7→ K be continuous. Show that F admits a fixed point by assuming
that there is a fixed point if K is a ball Σ or an n-dimensional simplex
Σ.
Hint: Consider for x ∈ Σ the mapping F (pK (x)), where pK is the
orthogonal projection onto K.
184 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Bibliography
[3] B. Beavis and I. M. Dobbs, Optimization and Stability Theory for Eco-
nomic Analysis. Cambridge University Press 1990.
[7] E. Bombieri, E. De Giorgi and E. Giusti, Minimal cones and the Bern-
stein problem. Inv. Math. 7 (1969), 243–268.
185
186 BIBLIOGRAPHY
[35] J.-L. Lions and E. Magenes, Problémes aux limites non homogénes et
applications. Dunod, Paris, 1968.
[53] L. Scheeffer, Über die Bedeutung der Begriffe ”Maximum und Mini-
mum” in der Variationsrechnung. Math. Annalen 26 (1886), 197–208.
[60] E. Zeidler, Nonlinear Functional Analysis and its Applications. III Vari-
ational Methods and Optimizations. Springer, 1985.