Calculus of Variations

Lecture Notes

Erich Miersemann
Department of Mathematics
Leipzig University

Version October, 2015

2
Contents

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

2.3 Sufficient conditions . . . . . . . . . . . . . . . . . . . . . . . 59

2.3.1 Equality constraints . . . . . . . . . . . . . . . . . . . 61
2.3.2 Inequality constraints . . . . . . . . . . . . . . . . . . 62
2.3.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.4 Kuhn-Tucker theory . . . . . . . . . . . . . . . . . . . . . . . 65
2.4.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.5.1 Maximizing of utility . . . . . . . . . . . . . . . . . . . 72
2.5.2 V is a polyhedron . . . . . . . . . . . . . . . . . . . . 73
2.5.3 Eigenvalue equations . . . . . . . . . . . . . . . . . . . 73
2.5.4 Unilateral eigenvalue problems . . . . . . . . . . . . . 77
2.5.5 Non-cooperative games . . . . . . . . . . . . . . . . . 79
2.5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.6 Appendix: Convex sets . . . . . . . . . . . . . . . . . . . . . . 87
2.6.1 Separation of convex sets . . . . . . . . . . . . . . . . 87
2.6.2 Linear inequalities . . . . . . . . . . . . . . . . . . . . 91
2.6.3 Projection on convex sets . . . . . . . . . . . . . . . . 93
2.6.4 Lagrange multiplier rules . . . . . . . . . . . . . . . . 95
2.6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 98
2.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

3 Ordinary differential equations 101

3.1 Optima, tangent cones, derivatives . . . . . . . . . . . . . . . 101
3.1.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.2 Necessary conditions . . . . . . . . . . . . . . . . . . . . . . . 105
3.2.1 Free problems . . . . . . . . . . . . . . . . . . . . . . . 105
3.2.2 Systems of equations . . . . . . . . . . . . . . . . . . . 116
3.2.3 Free boundary conditions . . . . . . . . . . . . . . . . 119
3.2.4 Transversality conditions . . . . . . . . . . . . . . . . 121
3.2.5 Nonsmooth solutions . . . . . . . . . . . . . . . . . . . 125
3.2.6 Equality constraints; functionals . . . . . . . . . . . . 130
3.2.7 Equality constraints; functions . . . . . . . . . . . . . 133
3.2.8 Unilateral constraints . . . . . . . . . . . . . . . . . . 137
3.2.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 141
3.3 Sufficient conditions; weak minimizers . . . . . . . . . . . . . 145
3.3.1 Free problems . . . . . . . . . . . . . . . . . . . . . . . 145
3.3.2 Equality constraints . . . . . . . . . . . . . . . . . . . 148
3.3.3 Unilateral constraints . . . . . . . . . . . . . . . . . . 151
3.3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 158
3.4 Sufficient condition; strong minimizers . . . . . . . . . . . . . 160
CONTENTS 5

3.4.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 165

3.5 Optimal control . . . . . . . . . . . . . . . . . . . . . . . . . . 166
3.5.1 Pontryagin’s maximum principle . . . . . . . . . . . . 167
3.5.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 168
3.5.3 Proof of Pontryagin’s maximum principle; free endpoint172
3.5.4 Proof of Pontryagin’s maximum principle; fixed end-
point . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
3.5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 181
6 CONTENTS
Preface

These lecture notes are a straightforward introduction to the calculus of

variations which can serve as a textbook for undergraduate and beginning
graduate students.
The main body of Chapter 2 consists of well known results concerning
necessary or sufficient criteria for local minimizers, including Lagrange mul-
tiplier rules, of real functions defined on a Euclidean n-space. Chapter 3
concerns problems governed by ordinary differential equations.
The content of these notes is not encyclopedic at all. For additional
reading we recommend following books: Luenberger [36], Rockafellar [50]
and Rockafellar and Wets [49] for Chapter 2 and Bolza [6], Courant and
Hilbert [9], Giaquinta and Hildebrandt [19], Jost and Li-Jost [26], Sagan [52],
Troutman [59] and Zeidler [60] for Chapter 3. Concerning variational prob-
lems governed by partial differential equations see Jost and Li-Jost [26] and
Struwe [57], for example.

7
8 CONTENTS
Chapter 1

Introduction

A huge amount of problems in the calculus of variations have their origin

in physics where one has to minimize the energy associated to the problem
under consideration. Nowadays many problems come from economics. Here
the main point is that the resources are restricted. There is no economy
without restricted resources.

Some basic problems in the calculus of variations are:

(i) find minimizers,

(ii) necessary conditions which have to satisfy minimizers,
(iii) find solutions (extremals) which satisfy the necessary condition,
(iv) sufficient conditions which guarantee that such solutions are minimizers,
(v) qualitative properties of minimizers, like regularity properties,
(vi) how depend minimizers on parameters?,
(vii) stability of extremals depending on parameters.

In the following we consider some examples.

1.1 Problems in Rn
1.1.1 Calculus
Let f : V 7→ R, where V ⊂ Rn is a nonempty set. Consider the problem

x ∈ V : f (x) ≤ f (y) for all y ∈ V.

If there exists a solution then it follows further characterizations of the

solution which allow in many cases to calculate this solution. The main tool

9
10 CHAPTER 1. INTRODUCTION

for obtaining further properties is to insert for y admissible variations of x.

For example let V be a convex set. Then for given y ∈ V
f (x) ≤ f (x + ǫ(y − x))
for all real 0 ≤ ǫ ≤ 1. From this inequality one derives the inequality
h∇f (x), y − xi ≥ 0 for all y ∈ V,
provided that f ∈ C 1 (Rn ).

1.1.2 Nash equilibrium

In generalization to the above problem we consider two real functions fi (x, y),
i = 1, 2, defined on S1 × S2 , where Si ⊂ Rmi . An (x∗ , y ∗ ) ∈ S1 × S2 is called
a Nash equilibrium if
f1 (x, y ∗ ) ≤ f1 (x∗ , y ∗ ) for all x ∈ S1
f2 (x∗ , y) ≤ f2 (x∗ , y ∗ ) for all y ∈ S2 .
The functions f1 , f2 are called payoff functions of two players and the sets
S1 and S2 are the strategy sets of the players. Under additional assumptions
on fi and Si there exists a Nash equilibrium, see Nash [46]. In Section 2.4.5
we consider more general problems of noncooperative games which play an
important role in economics, for example.

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

The higher eigenvalues can be characterized by the maximum-minimum

principle of Courant, see Section 2.5.

In generalization, let C ⊂ Rn be a nonempty closed convex cone with vertex

at the origin. Assume C 6= {0}. Then, see [37],
hAy, yi
λ1 = min
y∈C\{0} hBy, yi
1.2. ORDINARY DIFFERENTIAL EQUATIONS 11

is the lowest eigenvalue of the variational inequality

x ∈ C : hAx, y − xi ≥ λhBx, y − xi for all y ∈ C.

Remark. A set C ⊂ Rn is said to be a cone with vertex at x if for any

y ∈ C it follows that x + t(y − x) ∈ C for all t > 0.

1.2 Ordinary differential equations

Set Z b
E(v) = f (x, v(x), v ′ (x)) dx
a
and for given ua , ub ∈ R

V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub },

where −∞ < a < b < ∞ and f is sufficiently regular. One of the basic
problems in the calculus of variation is

(P ) minv∈V E(v).

That is, we seek a

u ∈ V : E(u) ≤ E(v) for all v ∈ V.

Euler equation. Let u ∈ V be a solution of (P) and assume additionally

u ∈ C 2 (a, b), then

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

Figure 1.1: Admissible variations

Basic lemma in the calculus of variations. Let h ∈ C(a, b) and

Z b
h(x)φ(x) dx = 0
a

for all φ ∈ C01 (a, b). Then h(x) ≡ 0 on (a, b).

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]

Thus φ ∈ C01 (a, b) and

Z b Z x0 +δ
h(x0 )
h(x)φ(x) dx ≥ φ(x) dx > 0,
a 2 x0 −δ

which is a contradiction to the assumption of the lemma. ✷

1.2.1 Rotationally symmetric minimal surface

Consider a curve defined by v(x), 0 ≤ x ≤ l, which satisfies v(x) > 0 on [0, l]
and v(0) = a, v(l) = b for given positive a and b, see Figure 1.2. Let S(v)
1.2. ORDINARY DIFFERENTIAL EQUATIONS 13

S
a

l x

Figure 1.2: Rotationally symmetric surface

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

Solutions of the associated Euler equation are catenoids (= chain curves),

see an exercise.

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

where V is the set of C 1 [t1 , t2 ] curves defined by (x(t), y(t)), t1 ≤ t ≤ t2 , with

x′ (t)2 + y ′ (t)2 6= 0, (x(t1 ), y(t1 )) = P1 , (x(t2 ), y(t2 )) = P2 and k := v12 /2g,
where v1 is the absolute value of the initial velocity of the mass point, and
y1 := y(t1 ). Solutions are cycloids (German: Rollkurven), see Bolza [6]
14 CHAPTER 1. INTRODUCTION

x
P
1

m g

y P
2

Figure 1.3: Problem of a brachistochrone

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

1.2.3 Geodesic curves

Consider a surface S in R3 , two points P1 , P2 on S and a curve on S
connecting these points, see Figure 1.4. Suppose that the surface S is defined
by x = x(v), where x = (x1 , x2 , x3 ) and v = (v1 , v2 ) and v ∈ U ⊂ R2 .
Consider curves v(t), t1 ≤ t ≤ t2 , in U such that v ∈ C 1 [t1 , t2 ] and v1′ (t)2 +
v2′ (t)2 6= 0 on [t1 , t2 ], and define

V = {v ∈ C 1 [t1 , t2 ] : x(v(t1 )) = P1 , x(v(t2 )) = P2 }.

The length of a curve x(v(t)) for v ∈ V is given by

Z t2 r
dx(v(t)) dx(v(t))
L(v) = · dt.
t1 dt dt

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

Figure 1.4: Geodesic curves

for the length of the curve under consideration

Z t2 q
L(v) = E(v(t))v1′ (t)2 + 2F (v(t))v1′ (t)v2′ (t) + G(v(t))v2′ (t)2 dt
t1

and the associated variational problem to study is here

min L(v).
v∈V

For examples of surfaces (sphere, ellipsoid) see [9], Part II.

1.2.4 Critical load

Consider the problem of the critical Euler load P for a beam. This value is
given by
a(v, v)
P = min ,
V \{0} b(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

I surface moment of inertia, EI is called bending stiffness,

V is the set of admissible deflections defined by the prescribed conditions at
the ends of the beam. In the case of a beam simply supported at both ends,
see Figure 1.5(a), we have

P P

v v

l l

(a) (b)

Figure 1.5: Euler load of a beam

V = {v ∈ C 2 [0, l] : v(0) = v(l) = 0}

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

V = {v ∈ C 2 [0, l] : v(0) = v ′ (0) = 0},

and the critical load is here P = EIπ 2 /(4l2 ).

Remark. The quotient a(v, v)/b(v, v) is called Rayleigh quotient (Lord

Rayleigh, 1842-1919).

Example: Summer house

As an example we consider a summer house based on columns, see Fig-

ure 1.6:
9 columns of pine wood, clamped at the lower end, free at the upper end,
9 cm × 9 cm is the cross section of each column,
2,5 m length of a column,
1.2. ORDINARY DIFFERENTIAL EQUATIONS 17

9 - 16 · 109 N m−2 modulus of elasticity, parallel fiber,

0.6 - 1 · 109 N m−2 modulus of elasticity, perpendicular fiber,
Z Z
I= x2 dxdy, Ω = (−4.5, 4.5) × (−4.5, 4.5),
Ω
I = 546.75 ·10−8 m4 ,
E := 5 × 109 N m−2 ,

P=10792 N, m=1100 kg (g:=9.80665 ms−2 ),

9 columns: 9900 kg,

18 m2 area of the flat roof,
10 cm wet snow: 1800 kg.

Figure 1.6: Summer house construction

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

u ∈ V : a(u, v − u) ≥ λb(u, v − u) for all v ∈ V.

A real λ0 is said to be a point of bifurcation of the above inequality if there

exists a sequence un , un 6≡ 0, of solutions with associated eigenvalues λn
such that un → 0 uniformly on [0, l] and λn → λ0 .

Optimal design of a column

Consider a rotationally symmetric column, see Figure 1.8. Let
l be the length of the column,
r(x) radius of the cross section,
I(x) = π(r(x))4 /4 surface moment of inertia,
ρ constant density of the material,
E modulus of elasticity.
Set
Z l Z Z l 
4 ′′ ′′ 4ρ l
a(r)(u, v) = r(x) u (x)v (x) dx − r(t) dt u′ (x)v ′ (x) dx
2
0 E 0 x
Z l
b(r)(v, v) = u′ (x)v ′ (x) dx.
0
1.2. ORDINARY DIFFERENTIAL EQUATIONS 19

r(x)

Figure 1.8: Optimal design of a column

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

V = {v ∈ C 2 [0, l] : v(0) = v ′ (0) = 0}

and consider the Rayleigh quotient

a(r)(v, v)
q(r, v) = .
b(r)(v, v)
We seek an r such that the critical load P (r) = Eπλ(r)/4, where

λ(r) = min q(r, v),

v∈V \{0}

approaches its infimum in a given set U of functions, for example

Z l
U = {r ∈ C[a, b] : r0 ≤ r(x) ≤ r1 , π r(x)2 dx = M },
0

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

Let (r0 , v0 ) be a solution, then

q(r, v0 ) ≤ q(r0 , v0 ) ≤ q(r0 , v)

for all r ∈ U and for all v ∈ V \ {0}.

1.2.5 Euler’s polygonal method

Consider the functional
Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a

where v ∈ V with

V = {v ∈ C 1 [a, b] : v(a) = A, v(b) = B}

with given A, B. Let

a = x0 < x1 < . . . < xn < xn+1 = b

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

Figure 1.9: Polygonal method

and replace the above integral by

n+1
X  
vi − vi−1
e(v) = f xi , v i , hi .
hi
i=1

The problem minv∈Rn e(v) is an associated finite dimensional problem to

minv∈V E(v). Then one shows, under additional assumptions, that the finite
1.2. ORDINARY DIFFERENTIAL EQUATIONS 21

dimensional problem has a solution which converges to a solution to the

original problem if n → ∞.

Remark. The historical notation ”problems with infinitely many variables”

for the above problem for the functional E(v) has its origin in Euler’s polyg-
onal method.

1.2.6 Optimal control

As an example for problems in optimal control theory we mention here a
problem governed by ordinary differential equations. For a given function
v(t) ∈ U ⊂ Rm , t0 ≤ t ≤ t1 , we consider the boundary value problem

y ′ (t) = f (t, y(t), v(t)), y(t0 ) = x0 , y(t1 ) = x1 ,

where y ∈ Rn , x0 , x1 are given, and

f : [t0 , t1 ] × Rn × Rm 7→ Rn .

In general, there is no solution of such a problem. Therefore we consider

the set of admissible controls Uad defined by the set of piecewise continuous
functions v on [t0 , t1 ] such that there exists a solution of the boundary value
problem. We suppose that this set is not empty. Assume a cost functional
is given by Zt1
E(v) = f 0 (t, y(t)), v(t)) dt,
t0
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.

The functions f, f 0 are assumed to be continuous in (t, y, v) and contin-

uously differentiable in (t, y). It is not required that these functions are
differentiable with respect to v.

Then the problem of optimal control is

max E(v).
v∈Uad

A piecewise continuous solution u is called optimal control and the solution

x of the associated system of boundary value problems is said to be optimal
trajectory.
22 CHAPTER 1. INTRODUCTION

The governing necessary condition for this type of problems is the Pon-
tryagin maximum principle, see [48] and Section 3.5.

1.3 Partial differential equations

The same procedure as above applied to the following multiple integral leads
to a second-order quasilinear partial differential equation. Set
Z
E(v) = F (x, v, ∇v) dx,
Ω

where Ω ⊂ Rn is a domain, x = (x1 , . . . , xn ), v = v(x) : Ω 7→ R, and

∇v = (vx1 , . . . , vxn ). It is assumed that the function F is sufficiently regular
in its arguments. For a given function h, defined on ∂Ω, set

V = {v ∈ C 1 (Ω) : v = h on ∂Ω}.

Euler equation. Let u ∈ V be a solution of (P), and additionally u ∈

C 2 (Ω), then
Xn
∂
Fu = Fu
∂xi xi
i=1

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. ✷

1.3.1 Dirichlet integral

In two dimensions the Dirichlet integral is given by
Z


D(v) = vx2 + vy2 dxdy
Ω

and the associated Euler equation is the Laplace equation △u = 0 in Ω.

Thus, there is natural relationship between the boundary value problem

△u = 0 in Ω, u = h on ∂Ω
1.3. PARTIAL DIFFERENTIAL EQUATIONS 23

and the variational problem

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].

1.3.2 Minimal surface equation

The non-parametric minimal surface problem in two dimensions is to find a
minimizer u = u(x1 , x2 ) of the problem
Z q
min 1 + vx21 + vx22 dx,
v∈V Ω

where for a given function h defined on the boundary of the domain Ω

V = {v ∈ C 1 (Ω) : v = h on ∂Ω}.

Suppose that the minimizer satisfies the regularity assumption u ∈ C 2 (Ω),

then u is a solution of the minimal surface equation (Euler equation) in Ω
! !
∂ u x1 ∂ u x2
p + p = 0.
∂x1 1 + |∇u|2 ∂x2 1 + |∇u|2

In fact, the additional assumption u ∈ C 2 (Ω) is superfluous since it follows

from regularity considerations for quasilinear elliptic equations of second
order, see for example Gilbarg and Trudinger [20].
Let Ω = R2 . Each linear function is a solution of the minimal surface
equation. It was shown by Bernstein [4] that there are no other solutions of
the minimal surface equation. This is true also for higher dimensions n ≤ 7,
see Simons [56]. If n ≥ 8, then there exists also other solutions which define
cones, see Bombieri, De Giorgi and Giusti [7].
The linearized minimal surface equation over u ≡ 0 is the Laplace equa-
tion △u = 0. In R2 linear functions are solutions but also many other
functions in contrast to the minimal surface equation. This striking differ-
ence is caused by the strong non-linearity of the minimal surface equation.
24 CHAPTER 1. INTRODUCTION

Figure 1.10: Comparison surface

More general minimal surfaces are described by using parametric rep-

resentations. An example is shown in Figure 1.112 . See [52], pp. 62, for
example, for rotationally symmetric minimal surfaces, and [47, 12, 13] for
more general surfaces. Suppose that the surface S is defined by y = y(v),
where y = (y1 , y2 , y3 ) and v = (v1 , v2 ) and v ∈ U ⊂ R2 . The area of the
surface S is given by
Z p
|S(y)| = EG − F 2 dv,
U

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

where V is a given set of comparison surfaces which is defined, for example,

by the condition that y(∂U ) ⊂ Γ, where Γ is a given curve in R3 , see
Figure 1.12. Set V = C 1 (Ω) and
Z Z
E(v) = F (x, v, ∇v) dx − g(x, v) ds,
Ω ∂Ω

2
An experiment from Beutelspacher’s Mathematikum, Wissenschaftsjahr 2008, Leipzig
1.3. PARTIAL DIFFERENTIAL EQUATIONS 25

Figure 1.11: Rotationally symmetric minimal surface

where F and g are given sufficiently regular functions and Ω ⊂ Rn is a

bounded and sufficiently regular domain. Assume u is a minimizer of E(v)
in V , that is,
u ∈ V : E(u) ≤ E(v) for all v ∈ V,
then
Z n
X

Fuxi (x, u, ∇u)φxi + Fu (x, u, ∇u)φ dx
Ω i=1
Z
− gu (x, u)φ ds = 0
∂Ω

for all φ ∈ C 1 (Ω). Assume additionally that u ∈ C 2 (Ω), then u is a solution

of the Neumann type boundary value problem
n
X ∂
Fu = Fu in Ω
∂xi xi
i=1
n
X
Fuxi νi = gu on ∂Ω,
i=1

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

Figure 1.12: Minimal surface spanned between two rings

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

1.3.3 Capillary equation

Let Ω ⊂ R2 and set
Z p Z Z
2
κ 2
E(v) = 1 + |∇v| dx + v dx − cos γ v ds.
Ω 2 Ω ∂Ω

Here is κ a positive constant (capillarity constant) and γ is the (constant)

boundary contact angle, that is, the angle between the container wall and
the capillary surface, defined by v = v(x1 , x2 ), at the boundary. Then the
related boundary value problem is

div (T u) = κu in Ω
ν · T u = cos γ on ∂Ω,
1.3. PARTIAL DIFFERENTIAL EQUATIONS 27

where we use the abbreviation

∇u
Tu = p ,
1 + |∇u|2

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].

Figure 1.13: Ascent of liquid in a wedge

The above problem is a special case (graph solution) of the following

problem. Consider a container partially filled with a liquid, see Figure 1.14.
Suppose that the associate energy functional is given by
Z
E(S) = σ|S| − σβ|W (S)| + Y ρ dx,
Ωl (S)

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 )

Figure 1.14: Liquid in a container

σ surface tension, σ = const. > 0,

β (relative) adhesion coefficient between the fluid and the container wall,
W wetted part of the container wall,
Ωl domain occupied by the liquid.
Additionally we have for given volume V of the liquid the constraint

|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

Figure 1.15: Liquid hangs on an edge

1.3.4 Liquid layers

Porous materials have a large amount of cavities different in size and geom-
etry. Such materials swell and shrink in dependence on air humidity. Here
we consider an isolated cavity, see [54] for some cavities of special geometry.
Let Ωs ∈ R3 be a domain occupied by homogeneous solid material. The
question is whether or not liquid layers Ωl on Ωs are stable, where Ωv is the
domain filled with vapor and S is the capillary surface which is the interface
between liquid and vapor, see Figure 1.16.

solid Ωs

Ωl
S
liquid N
Ωv

vapor

Figure 1.16: Liquid layer in a pore

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.

More precisely, ρ is the difference between the number densities of the

liquid and the vapor phase. However, since in most practical cases the vapor
density is rather small, ρ can be replaced by the density of the liquid phase.
The above negative constant is given by c = H/π 2 , where H is the
Hamaker constant, see [25], p. 177. For a liquid nitrogen film on quartz one
has about H = −10−20 N m.
Suppose that S0 defines a local minimum of the energy functional, then
−2σH + F − µ = 0 on S0 , (1.4)
where H is the mean curvature of S0 .
A surface S0 which satisfies (1.4) is said to be an equilibrium state. An
existing equilibrium state S0 is said to be stable by definition if
 2 
d
E(S(ǫ)) >0
dǫ2 ǫ=0
for all ζ not identically zero, where S(ǫ) is an appropriate one-parameter
family of comparison surfaces.
This inequality implies that
1 ∂F
−2(2H 2 − K) + > 0 on S0 , (1.5)
σ ∂N
where K is the Gauss curvature of the capillary surface S0 , see Blaschke [5], p. 58,
for the definition of K.
1.3. PARTIAL DIFFERENTIAL EQUATIONS 31

1.3.5 Extremal property of an eigenvalue

Let Ω ⊂ R2 be a bounded and connected domain. Consider the eigenvalue
problem

−△u = λu in Ω
u = 0 on ∂Ω.

It is known that the lowest eigenvalue λ1 (Ω) is positive, it is a simple eigen-

value and the associated eigenfunction has no zero in Ω. Let V be a set of
sufficiently regular domains Ω with prescribed area |Ω|. Then we consider
the problem
min λ1 (Ω).
Ω∈V
p
The solution of this problem is a disk BR , R = |Ω|/π, and the solution is
uniquely determined.

1.3.6 Isoperimetric problems

Let V be a set of all sufficiently regular bounded and connected domains
Ω ⊂ R2 with prescribed length |∂Ω| of the boundary. Then we consider the
problem
max |Ω|.
Ω∈V

The solution of this problem is a disk BR , R = |∂Ω|/(2π), and the solution

is uniquely determined. This result follows by Steiner’s symmetrization,
see [5], for instance. From this method it follows that

|∂Ω|2 − 4π|Ω| > 0

if Ω is a domain different from a disk.

Remark. Such an isoperimetric inequality follows also by using the in-

equality Z Z
1
|u| dx ≤ |∇u|2 dx
R 2 4π R 2

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

The associated problem in R3 is

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

lim inf f (xk ) ≥ f (x).

k→∞

2. Let V ⊂ Rn be the closure of a convex domain and assume f : V → 7 R

is in C 1 (Rn ). Suppose that x ∈ V satisfies f (x) ≤ f (y) for all y ∈ V .
Prove
(i) h∇f (x), y − xi ≥ 0 for all y ∈ V ,
(ii) ∇f (x) = 0 if x is an interior point of V .
3. Let A and B be real and symmetric matrices with n rows (and n
columns). Suppose that B is positive, i. e., hBy, yi > 0 for all y ∈
Rn \ {0}.
(i) Show that there exists a solution x of the problem
hAy, yi
min .
y∈Rn \{0} hBy, yi
(ii) Show that Ax = λBx, where λ = hAx, xi/hBx, xi.
Hint: (a) Show that there is a positive constant such that hBy, yi ≥
chy, yi for all y ∈ Rn .
(b) Show that there exists a solution x of the problem miny hAy, yi,
where hBy, yi = 1.
(c) Consider the function
hA(x + ǫy), x + ǫyi
g(ǫ) = ,
hB(x + ǫy, x + ǫyi
where |ǫ| < ǫ0 , ǫ0 sufficiently small, and use that g(0) ≤ g(ǫ).
4. Let A and B satisfy the assumption of the previous exercise. Let C
be a closed convex nonempty cone in Rn with vertex at the origin.
Assume C 6= {0}.
(i) Show that there exists a solution x of the problem
hAy, yi
min .
y∈C\{0} hBy, yi
34 CHAPTER 1. INTRODUCTION

(ii) Show that x is a solution of

x ∈ C : hAx, y − xi ≥ λhx, y − xi for all y ∈ C,

where λ = hAx, xi/hBx, xi.

Hint: To show (ii) consider for x, y ∈ C the function

hA(x + ǫ(y − x)), x + ǫ(y − x)i

g(ǫ) = ,
hB(x + ǫ(y − x)), x + ǫ(y − x)i

where 0 < ǫ < ǫ0 , ǫ0 sufficiently small, and use g(0) ≤ g(ǫ) which
implies that g ′ (0) ≥ 0.

5. Let A be real matrix with n rows and n columns, and let C ⊂ Rn be

a nonempty closed and convex cone with vertex at the origin. Show
that
x ∈ C : hAx, y − xi ≥ 0 for all y ∈ C
is equivalent to

hAx, xi = 0 and hAx, yi ≥ 0 for all y ∈ C.

Hint: 2x, x + y ∈ C if x, y ∈ C.

6. R. Courant. Show that

Z 1
1/4
E(v) := 1 + (v ′ (x))2 dx
0

does not achieve its infimum in the class of functions

V = {v ∈ C[0, 1] : v piecewise C 1 , v(0) = 1, v(1) = 0},

that is, there is no u ∈ V such that E(u) ≤ E(v) for all v ∈ V .

Hint: Consider the family of functions

(ǫ − x)/ǫ : 0 ≤ x ≤ ǫ < 1
v(ǫ; x) =
0 : x>ǫ

7. K. Weierstraß, 1895. Show that

Z 1
E(v) = x2 (v ′ (x))2 dx
−1
1.4. EXERCISES 35

does not achieve its infimum in the class of functions


V = v ∈ C 1 [−1, 1] : v(−1) = a, v(1) = b ,

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

where ǫ, |ǫ| < ǫ0 , is a real parameter, f (x, z, p) in C 2 in his arguments

and u, φ ∈ C 1 [a, b]. Calculate g ′ (0) and g ′′ (0).

9. Find all C 2 -solutions u = u(x) of

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.

11. Is there a solution of minv∈V E(v), where V = C[0, 1] and

Z Z !
1 v(x)
E(v) = (1 + ζ 2 ) dζ dx ?
0 0

12. Let u ∈ C 2 (a, b) be a solution of Euler’s differential equation. Show

that u′ fu′ − f ≡ const., provided that f = f (u, u′ ), that is, f depends
not explicitly on x.
36 CHAPTER 1. INTRODUCTION

13. Consider the problem of rotationally symmetric surfaces minv∈V |S(v)|,

where Z l p
|S(v)| = 2π v(x) 1 + (v ′ (x))2 dx
0
and

V = {v ∈ C 1 [0, l] : v(0) = a, v(l) = b, v(x) > 0 on (a, b)}.

Find C 2 (0, l)-solutions of the associated Euler equation.

Hint: Solutions are catenoids (chain curves, in German: Kettenlinien).

14. Find solutions of Euler’s differential equation to the Brachistochrone

problem minv∈V E(v), where

V = {v ∈ C[0, a]∩C 2 (0, a] : v(0) = 0, v(a) = A, v(x) > 0 if x ∈ (0, a]},

that is, we consider here as comparison functions graphs over the x-

axis, and
Z a √
1 + v ′2
E(v) = √ dx .
0 v

Hint: (i) Euler’s equation implies that

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).

15. Prove the basic lemma in the calculus of variations: Let Ω ⊂ Rn be a

domain and f ∈ C(Ω) such that
Z
f (x)h(x) dx = 0
Ω

for all h ∈ C01 (Ω). Then f ≡ 0 in Ω.

16. Write the minimal surface equation as a quasilinear equation of second

order.
1.4. EXERCISES 37

17. Prove that a minimizer in C 1 (Ω) of

Z Z
E(v) = F (x, v, ∇v) dx − g(v, v) ds,
Ω ∂Ω

is a solution of the boundary value problem, provided that additionally

u ∈ C 2 (Ω),
n
X ∂
Fu = Fu in Ω
∂xi xi
i=1
n
X
Fuxi νi = gu on ∂Ω,
i=1

where ν = (ν1 , . . . , νn ) is the exterior unit normal at the boundary ∂Ω.

38 CHAPTER 1. INTRODUCTION
Chapter 2

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 .

2.1 Optima, tangent cones

Let f be a real-valued functional defined on a nonempty subset V ⊆ X.

Definition. We say that an element x ∈ V defines a global minimum of f

in V if
f (x) ≤ f (y) for all y ∈ V,
and we say that x ∈ V defines a strict global minimum if the strict inequality
holds for all y ∈ V, y 6= x.

For a ρ > 0 we define a ball Bρ (x) with radius ρ and center x:

Bρ (x) = {y ∈ Rn ; ||y − x|| < ρ},

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.

Definition. We say that an element x ∈ V defines a local minimum of f in

V if there exists a ρ > 0 such that

f (x) ≤ f (y) for all y ∈ V ∩ Bρ (x),

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.

That is, a global minimum is a local minimum. By reversing the direc-

tions of the inequality signs in the above definitions, we obtain definitions
of global maximum, strict global maximum, local maximum and strict local
maximum. An optimum is a minimum or a maximum and a local optimum
is a local minimum or a local maximum. If x defines a local maximum etc.
of f , then x defines a local minimum etc. of −f , that is, we can restrict
ourselves to the consideration of minima.
In the important case that V and f are convex, then each local minimum
defines also a global minimum. These assumptions are satisfied in many
applications to problems in microeconomy.

Definition. A subset V ⊆ X is said to be convex if for any two vectors

x, y ∈ V the inclusion λx + (1 − λ)y ∈ V holds for all 0 ≤ λ ≤ 1.

Definition. We say that a functional f defined on a convex subset V ⊆ X

is convex if
f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y)
for all x, y ∈ V and for all 0 ≤ λ ≤ 1, and f is strictly convex if the strict
inequality holds for all x, y ∈ V , x 6= y, and for all λ, 0 < λ < 1.

Theorem 2.1.1. If f is a convex functional on a convex set V ⊆ X, then

any local minimum of f in V is a global minimum of f in V .

Proof. Suppose that x is no global minimum, then there exists an x1 ∈ V

such that f (x1 ) < f (x). Set y(λ) = λx1 + (1 − λ)x, 0 < λ < 1, then

f (y(λ)) ≤ λf (x1 ) + (1 − λ)f (x) < λf (x) + (1 − λ)f (x) = f (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. ✷

Concerning the uniqueness of minimizers one has the following result.

Theorem 2.1.2. If f is a strictly convex functional on a convex set V ⊆ X,

then a minimum (local or global) is unique.

Proof. Suppose that x1 , x2 ∈ V define minima of f , then f (x1 ) = f (x2 ),

2.1. OPTIMA, TANGENT CONES 41

see Theorem 2.1.1. Assume x1 6= x2 , then for 0 < λ < 1

f (λx1 + (1 − λ)x2 ) < λf (x1 ) + (1 − λ)f (x2 ) = f (x1 ) = f (x2 ).

This is a contradiction to the assumption that x1 , x2 define global minima.

✷

Theorem 2.1.3. a) If f is a convex function and V ⊂ X a convex set, then

the set of minimizers is convex.
b) If f is concave, V ⊂ X convex, then the set of maximizers is convex.

Proof. Exercise.

In the following we use for (fx1 (x), . . . , fxn (x)) the abbreviations f ′ (x),
∇f (x) or Df (x).

Theorem 2.1.4. Suppose that V ⊂ X is convex. Then f is convex on V if

and only if

f (y) − f (x) ≥ hf ′ (x), y − xi for all x, y ∈ V.

Proof. (i) Assume f is convex. Then for 0 ≤ λ ≤ 1 we have

f (λy + (1 − λ)x) ≤ λf (y) + (1 − λ)f (x)

f (x + λ(y − x)) ≤ f (x) + λ(f (y) − f (x))
′
f (x) + λhf (x), y − xi + o(λ) ≤ f (x) + λ(f (y) − f (x)),

which implies that

hf ′ (x), y − xi ≤ f (y) − f (x).
(ii) Set for x, y ∈ V and 0 < λ < 1

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

f (y) − f (x1 ) ≥ hf ′ (x1 ), y − x1 i = hf ′ (x1 ), hi

1−λ ′ 1
f (x) − f (x1 ) ≥ hf ′ (x1 ), x − x1 i = − hf (x ), hi.
λ
42 CHAPTER 2. FUNCTIONS OF N VARIABLES

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)

for all 0 < λ < 1. ✷

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

Figure 2.1: Figure to Theorem 2.1.4

The following definition of a local tangent cone is fundamental for our

considerations, in particular for the study of necessary conditions in con-
strained optimization. The tangent cone plays the role of the tangent plane
in unconstrained optimization.

Definition. A nonempty subset C ⊆ Rn is said to be a cone with vertex at

z ∈ Rn , if y ∈ C implies that z + t(y − z) ∈ C for each t > 0.

Let V be a nonempty subset of X.

2.1. OPTIMA, TANGENT CONES 43

Definition. For given x ∈ V we define the local tangent cone of V at x by

T (V, x) = {w ∈ Rn : there exist sequences xk ∈ V, tk ∈ R, tk > 0,

such that xk → x and tk (xk − x) → w as k → ∞}.

This definition implies immediately

x
T(V,x)

Figure 2.2: Tangent cone

Corollaries. (i) The set T (V, x) is a cone with vertex at zero.

(ii) A vector x ∈ V is not isolated if and only if T (V, x) 6= {0}.

(iii) Suppose that w 6= 0, then tk → ∞.

(iv) T (V, x) is closed.

(v) T (V, x) is convex if V is convex.

In the following the Hesse matrix (fxi xj (x))ni,j=1 is also denoted by f ′′ (x),
fxx (x) or D2 f (x).

Theorem 2.1.5. Suppose that V ⊂ Rn is nonempty and convex. Then

(i) If f is convex on V , then the Hesse matrix f ′′ (x), x ∈ V , is positive

semidefinite on T (V, x). That is,

hf ′′ (x)w, wi ≥ 0
44 CHAPTER 2. FUNCTIONS OF N VARIABLES

for all x ∈ V and for all w ∈ T (V, x).

(ii) Assume the Hesse matrix f ′′ (x), x ∈ V , is positive semidefinite on

Y = V − V , the set of all x − y where x, y ∈ V . Then f is convex on V .

Proof. (i) Assume f is convex on V . Then for all x, y ∈ V , see Theo-

rem 2.1.4,
f (y) − f (x) ≥ hf ′ (x), y − xi.
Thus
1
hf ′ (x), y − xi + hf ′′ (x)(y − x), y − xi + o(||y − x||2 ) ≥ hf ′ (x), y − xi
2
hf ′′ (x)(y − x), y − xi + ||y − x||2 η(||y − x||) ≥ 0,

where limt→0 η(t) = 0. Suppose that w ∈ T (V, x) and that tk , xk → x are

associated sequences, i. e., xk ∈ V , tk > 0 and

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 .

2. Prove Theorem 2.1.3.

3. Show that T (V, x) is a cone with vertex at zero.

4. Assume V ⊆ Rn . Show that T (V, x) 6= {0} if and only if x ∈ V is not

isolated.

5. Show that T (V, x) is closed.

6. Show that T (V, x) is convex if V is a convex set.

7. Suppose that V is convex. Show that

T (V, x) = {w ∈ Rn ; there exist sequences xk ∈ V, tk ∈ R, tk > 0,

such that tk (xk − x) → w as k → ∞}.

8. Assume if w ∈ T (V, x), w 6= 0. Then tk → ∞, where tk are the reals

from the definition of the local tangent cone.

9. Let p ∈ V ⊂ Rn and |p|2 = p21 + . . . + p2n . Prove that

p
f (p) = 1 + |p|2

is convex on convex sets V .

Hint: Show that the Hesse matrix f ′′ (p) is nonnegative by calculating
hf ′′ (p)ζ, ζi, where ζ ∈ Rn .

10. Suppose that f ′′ (x), x ∈ V , is positive on (V − V ) \ {0}. Show that

f (x) is strictly convex on V .
Hint: (i) Show that f (y)−f (x) > hf ′ (x), y −xi for all x, y ∈ V , x 6= y.
(ii) Then show that f is strictly convex by adapting part (ii) of the
proof of Theorem 2.1.4.

11. Let V ⊂ X be a nonempty convex subset of a linear space X and

f : V 7→ R. Show that f is convex on V if and only if Φ(t) :=
f (x + t(y − x)) is convex on t ∈ [0, 1] for all (fixed) x, y ∈ V .
Hint: To see that Φ is convex if f is convex we have to show

Φ(λs1 + (1 − λ)s2 )) ≤ λΦ(s1 ) + (1 − λ)Φ(s2 ),

46 CHAPTER 2. FUNCTIONS OF N VARIABLES

0 ≤ λ ≤ 1, si ∈ [0, 1]. Set τ = λs1 + (1 − λ)s2 ), then

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

2.2 Necessary conditions

The proof of the following necessary condition follows from assumption f ∈
C 1 (X) and the definition of the tangent cone.
We recall that f is said to be differentiable at x ∈ X ⊂ Rn , X open, if
all partial derivatives exist at x. In contrast to the case n = 1 it does not
follow that f is continuous at x if f is differentiable at that point.

Definition. f is called totally differentiable at x if there exists an a ∈ Rn

such that
f (y) = f (x) + ha, y − xi + o(||x − y||)
as y → x.

We recall that

(1) If f is totally differentiable at x, then f is differentiable at x and a =

f ′ (x).

(2) If f ∈ C 1 (Bρ ), then f is totally differentiable at every x ∈ Bρ .

(3) Rademacher’s Theorem. If f is locally Lipschitz continuous in Bρ , then

f is totally differentiable almost everywhere in Bρ , that is, up to a set of
Lebesgue measure zero.

For a proof of Rademacher’s Theorem see [15], pp. 81, for instance.

Theorem 2.2.1 (Necessary condition of first order). Suppose that f ∈

C 1 (X) and that x defines a local minimum of f in V . Then

hf ′ (x), wi ≥ 0 for all w ∈ T (V, x).

Proof. Let tk , xk are associated sequences to w ∈ T (V, x). Then, since x

defines a local minimum, it follows

0 ≤ f (xk ) − f (x) = hf ′ (x), xk − xi + o(||xk − x||),

which implies that

0 ≤ hf ′ (x), tk (xk − x)i + ||tk (xk − x)||η(||xk − x||),

48 CHAPTER 2. FUNCTIONS OF N VARIABLES

where limt→0 η(t) = 0. Letting n → ∞ we obtain the necessary condition.

✷

Corollary. Suppose that V is convex, then the necessary condition of The-

orem 1.1.1 is equivalent to

hf ′ (x), y − xi ≥ 0 for all y ∈ V.

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). ✷

The variational inequality above is equivalent to a fixed point equation, see

Theorem 2.2.2 below.

Let pV (z) be the projection of z ∈ H, where H is a real Hilbert space, onto

a nonempty closed convex subset V ⊆ H, see Section 2.6.3 of the appendix.
We have w = pV (z) if and only if

hpV (z) − z, y − pV (z)i ≥ 0 for all y ∈ V. (2.1)

Theorem 2.2.2 (Equivalence of a variational inequality to an equation).

Suppose that V is a closed convex and nonempty subset of a real Hilbert
space H and F a mapping from V into H. Then the variational inequality

x ∈ V : hF (x), y − xi ≥ 0 for all y ∈ V.

is equivalent to the fixed point equation

x = pV (x − qF (x)) ,

where 0 < q < ∞ is an arbitrary fixed constant.

Proof. Set z = x − qF (x) in (2.1). If x = pV (x − qF (x)) then the variational

inequality follows. On the other hand, the variational inequality

x ∈ V : hx − (x − qF (x)), y − xi ≥ 0 for all y ∈ V

implies that the fixed point equation holds. ✷

2.2. NECESSARY CONDITIONS 49

Remark. This equivalence of a variational inequality with a fixed point

equation suggests a simple numerical procedure to calculate solutions of
variational inequalities: xk+1 := pV (xk − qF (xk )). Then the hope is that
the sequence xk converges if 0 < q < 1 is chosen appropriately. In these
notes we do not consider the problem of convergence of this or of related
numerical procedures. This projection-iteration method runs quite well in
some examples, see [51], and exercises in Section 2.5.

In generalization to the necessary condition of second order in the un-

constrained case, hf ′′ (x)h, hi ≥ 0 for all h ∈ Rn , we have a corresponding
result in the constrained case.

Theorem 2.2.3 (Necessary condition of second order). Suppose that f ∈

C 2 (X) and that x defines a local minimum of f in V . Then for each w ∈
T (V, x) and every associated sequences tk , xk the inequality
1
0 ≤ lim inf tk hf ′ (x), wk i + hf ′′ (x)w, wi
k→∞ 2
holds, where wk := tk (xk − 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.

2.2.1 Equality constraints

Here we suppose that the subset V is defined by
V = {y ∈ Rn ; gj (y) = 0, j = 1, . . . , m}.
50 CHAPTER 2. FUNCTIONS OF N VARIABLES

Let gj ∈ C 1 (Rn ), w ∈ T (V, x) and tk , xk associated sequences to w. Then

0 = gj (xk ) − gj (x) = hgj′ (x), xk − xi + o(||xk − x||),

and from the definition of the local tangent cone it follows for each j that

hgj′ (x), wi = 0 for all w ∈ T (V, x). (2.2)

′ (x)} and let Rn = Y ⊕Y ⊥ be the orthogonal

Set Y = span {g1′ (x), . . . , gm
decomposition with respect to the standard Euclidean scalar product ha, bi.
We recall that dim Y ⊥ = n − k if dim Y = k.
Equations (2.2) imply immediately that T (V, x) ⊆ Y ⊥ . Under an addi-
tional assumption we have equality.

Lemma 2.2.1. Suppose that dim Y = m (maximal rank condition), then

T (V, x) = Y ⊥ .

Proof. It remains to show that Y ⊥ ⊆ T (V, x). Suppose that z ∈ Y ⊥ , 0 <

ǫ ≤ ǫ0 , ǫ0 sufficiently small. Then we look for solutions y = o(ǫ), depending
on the fixed z, of the system gj (x + ǫz + y) = 0, j = 1, . . . , m. Since z ∈ Y ⊥
and the maximal rank condition is satisfied, we obtain the existence of a
y = o(ǫ) as ǫ → 0 from the implicit function theorem. That is, we have
x(ǫ) := x + ǫz + y ∈ V , where y = o(ǫ). This implies that z ∈ T (V, x) since
x(ǫ) → x, x(ǫ), x ∈ V and ǫ−1 (x(ǫ) − x) → z as ǫ → 0. ✷

From this lemma follows the simplest version of the Lagrange multiplier rule
as a necessary condition of first order.

Theorem 2.2.4 (Lagrange multiplier rule, equality constraints). Suppose

that x defines a local minimum of f in V and that the maximal rank con-
dition of Lemma 2.1.1 is satisfied. Then there exist uniquely determined
λj ∈ R, such that
m
X
f ′ (x) + λj gj′ (x) = 0.
j=1

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

Uniqueness of the multipliers follows from the maximal rank condition. ✷

There is a necessary condition of second order linked with a Lagrange func-

tion L(x, λ) defined by
m
X
L(x, λ) = f (x) + λj gj (x).
j=1

Moreover, under an additional assumption the sequence tk hf ′ (x), wk i, where

wk := tk (xk −x), is convergent, and the limit is independent of the sequences
xk , tk associated to a given w ∈ T (V, x).
We will use the following abbreviations:
m
X m
X
′ ′
L (x, λ) := f (x) + λj gj′ (x) ′′
and L (x, λ) := f (x) + ′′
λj gj′′ (x).
j=1 j=1

Theorem 2.2.5 (Necessary condition of second order, equality constraints).

Suppose that f, gj ∈ C 2 (Rn ) and that the maximal rank condition of Lemma 2.2.1
is satisfied. Let x be a local minimizer of f in V and let λ = (λ1 , . . . , λm )
are the (uniquely determined) Lagrange multipliers of Theorem 2.2.4 Then

(i) hL′′ (x, λ)z, zi ≥ 0 for all z ∈ Y ⊥ (≡ T (V, x)),

m
′ k 1X
(ii) lim tk hf (x), w i = λj hgj′′ (x)w, wi, wk := tk (xk − x),
k→∞ 2
j=1

for all w ∈ T (V, x) and for all associated sequences xk , tk

to w ∈ T (V, x) (≡ Y ⊥ ).

Proof. (i) For given z ∈ Y ⊥ , ||z|| = 1, and 0 < ǫ ≤ ǫ0 , ǫ0 > 0 sufficiently

small, there is a y = o(ǫ) such that gj (x + ǫz + y) = 0, j = 1, . . . , m, which
follows from the maximal rank condition and the implicit function theorem.
Then

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.

(ii) Suppose that xk ∈ V, tk > 0, such that wk := tk (xk − x) → w. Then

m
X
L(xk , λ) ≡ f (xk ) + λj gj (xk )
j=1
1
= L(x, λ) + hL′′ (x, λ)(xk − x), xk − xi + o(||xk − x||2 ).
2
That is,

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

which implies (ii) of the theorem. ✷

2.2.2 Inequality constraints

We define two index sets by I = {1, . . . , m} and E = {m + 1, . . . , m + p} for
integers m ≥ 1 and p ≥ 0. If p = 0 then we set E = ∅, the empty set. In
this section we assume that the subset V is given by

V = {y ∈ Rn ; gj (y) ≤ 0 for each j ∈ I and gj (y) = 0 for each j ∈ E}

and that gj ∈ C 1 (Rn ) for each j ∈ I ∪ E. Let x ∈ V be a local minimizer

of f in V and let I0 ⊆ I be the subset of I where the inequality constraints
are active, that is, I0 = {j ∈ I; gj (x) = 0}. Let w ∈ T (V, x) and xk , tk are
associated sequences to w, then for k ≥ k0 , k0 sufficiently large, we have for
each j ∈ I0

0 ≥ gj (xk ) − gj (x) = hgj′ (x), xk − xi + o(||xk − x||).

2.2. NECESSARY CONDITIONS 53

It follows that for each j ∈ I0

hgj′ (x), wi ≤ 0 for all w ∈ T (V, x).

If j ∈ E, we obtain from

0 = gj (xk ) − gj (x) = hgj′ (x), xk − xi + o(||xk − x||)

and if j ∈ E, then

hgj′ (x), wi = 0 for all w ∈ T (V, x).

That is, the tangent cone T (V, x) is a subset of the cone

K = {z ∈ Rn : hgj′ (x), zi = 0, j ∈ E, and hgj′ (x), zi ≤ 0, j ∈ I0 }.

Under a maximal rank condition we have equality. By |M | we denote the

number of elements of a finite set M , we set |M | = 0 if M = ∅ .

Definition. A vector x ∈ V is said to be a regular point if

dim span {gj′ (x)}j∈E∪I0 = |E| + |I0 |.

It means that in a regular point the gradients of functions which de-

fine the active constraints (equality constraints are included) are linearly
independent.

Lemma 2.2.2. Suppose that x ∈ V is a regular point, then T (V, x) = K.

Proof. It remains to show that K ⊆ T (V, x). Suppose that z ∈ K, 0 < ǫ ≤

ǫ0 , ǫ0 sufficiently small. Then we look for solutions y ∈ Rn of the system
gj (x + ǫz + y) = 0, j ∈ E and gj (x + ǫz + y) ≤ 0, j ∈ I0 . Once one has
established such a y = o(ǫ), depending on the fixed z ∈ K, then it follows
that z ∈ T (V, x) since x(ǫ) := x + ǫz + y ∈ V, x(ǫ) → x, x(ǫ), x ∈ V and
ǫ−1 (x(ǫ) − x) → z as ǫ → 0.
Consider the subset I0′ ⊆ I0 defined by I0′ = {j ∈ I0 ; hgj′ (x), zi = 0}.
Then, the existence of a solution y = o(ǫ) of the system gj (x + ǫz + y) =
0, j ∈ E and gj (x + ǫz + y) = 0, j ∈ I0′ follows from the implicit function
theorem since  
dim span {gj′ (x)}j∈E∪I0′ = |E| + |I0′ |
54 CHAPTER 2. FUNCTIONS OF N VARIABLES

holds. The remaining inequalities gj (x + ǫz + y) ≤ 0, j ∈ I0 \ I0′ , are satisfied

for sufficiently small ǫ > 0 since hgj (x), zi < 0 if j ∈ I0 \ I0′ . ✷

Thus the necessary condition of first order of Theorem 2.1.1 is here

hf ′ (x), wi ≥ 0 for all w ∈ K (≡ T (V, x)), (2.3)

if the maximal rank condition of Lemma 2.2.2 is satisfied, that is, if x ∈ V

is a regular point.
In generalization of the case of equality constraints the variational in-
equality (2.3) implies the following Lagrange multiplier rule.

Theorem 2.2.6 (Lagrange multiplier rule, inequality constraints). Suppose

that x is a local minimizer of f in V and that x is a regular point. Then
there exist λj ∈ R, λj ≥ 0 if j ∈ I0 , such that
X
f ′ (x) + λj gj′ (x) = 0.
j∈E∪I0

Proof. Since the variational inequality (2.3) with

K = {z ∈ Rn : hgj′ (x), zi ≥ 0 and h−gj′ (x), zi ≥ 0 for each j ∈ E,

and h−gj′ (x), zi ≥ 0 for each j ∈ I0 }

(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

Kuhn–Tucker conditions. Let x be a local minimizer of f in V and

suppose that x is a regular point. Then there exist λj ∈ R, λj ≥ 0 if j ∈ I,
such that
X
f ′ (x) + λj gj′ (x) = 0 ,
j∈E∪I
X
λj gj (x) = 0 .
j∈I

As in the case of equality constraints there is a necessary condition of

second order linked with a Lagrange function.

Theorem 2.2.7 (Necessary condition of second order, inequality constraints).

Suppose that f, gj ∈ C 2 (Rn ). Let x be a local minimizer of f in V which is
regular and λj denote Lagrange multipliers such that
X
f ′ (x) + λj gj′ (x) = 0,
j∈E∪I0

where λj ≥ 0 if j ∈ I0 . Let I0+ = {j ∈ I0 ; λj > 0}, V0 = {y ∈ V ; gj (y) =

}, Z = {y ∈ Rn : hgj′ (x), yi = 0 for each j ∈ E ∪ I0+ } and
0 for each j ∈ I0+P
L(y, λ) ≡ f (y) + j∈E∪I0 λj gj (y). Then

(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

for all w ∈ T (V0 , x) and for all associated sequences xk , tk

to w ∈ T (V0 , x) (≡ Z).

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

On the other hand

1
f (y) − f (x) = hf ′ (x), y − xi + hf ′′ (x)(y − x), y − xi + o(||y − x||2 ).
2
Since f (y) ≥ f (x) if y is close enough to x, we obtain (ii) and (iii) by the
same reasoning as in the proof of Theorem 2.2.5. ✷

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.

Theorem 2.2.8 (General Lagrange multiplier rule). Suppose that x defines

a local minimum or a local maximum of f in V and that |E| + |I0 | < n.
Then there exist λj ∈ R, not all are zero, such that
X
λ0 f ′ (x) + λj gj′ (x) = 0 .
j∈E∪I0

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 ,

where y1 , . . . , yN are unknowns. The real number u is a given parameter in

a neighborhood of zero, say |u| < u0 for a sufficiently small u0 . From the
implicit function theorem it follows that there exist solutions yi = φi (u), i =
1, . . . , N , where φi (0) = xi . Set x∗ = (φ1 (u), . . . , φN (u), xN +1 , . . . , xn ).
Then f (x∗ ) > f (x) if u > 0 and f (x∗ ) < f (x) if u < 0, i. e., x defines
no local optimum. ✷

From this multiplier rule it follows immediately:

1. If x is a regular point, then λ0 6= 0. After dividing by λ0 the new coef-

ficients λ′j = λj /λ0 , j ∈ E ∪ I0 , coincide with the Lagrange multipliers
of Theorem 2.7.

2. If λ0 = 0, then x is no regular point.

58 CHAPTER 2. FUNCTIONS OF N VARIABLES

2.2.4 Exercises
1. Suppose that f ∈ C 1 (Bρ ). Show that

f (y) = f (x) + hf ′ (x), y − xi + o(||y − x||)

for every x ∈ Bρ , that is, f is totally differentiable in Bρ .

2. Assume g maps a ball Bρ (x) ⊂ Rn in Rm and let g ∈ C 1 (Bρ ). Sup-

pose that g(x) = 0 and dim Y = m (maximal rank condition), where
Y = span{g1′ (x), . . . , gm
′ }. Prove that for fixed z ∈ Y ⊥ there ex-

ists y(ǫ) which maps (−ǫ0 , ǫ0 ), ǫ0 > 0 sufficiently small, into Rm ,

y ∈ C 1 (−ǫ0 , ǫ0 ) and y(ǫ) = o(ǫ) such that g(x + ǫz + y(ǫ)) ≡ 0 in
(−ǫ0 , ǫ0 ).
Hint: After renaming the variables we can assume that the matrix
gi,xj , i, j = 1, . . . , m is regular. Set y := (y1 , . . . , ym , 0, . . . , 0) and
f (ǫ, y) := (g1 (x + ǫz + y), . . . , gm (x + ǫz + y)). From the implicit
function theorem it follows that there exists a C 1 function y = y(ǫ),
|ǫ| < ǫ1 , ǫ1 > 0 sufficiently small, such that f (ǫ, y(ǫ)) ≡ 0 in |ǫ| < ǫ1 .
Since gj (x + ǫz + y(ǫ)) = 0, j = 1, . . . , m and y(ǫ) = ǫa + o(ǫ), where
a = (a1 , . . . , am , 0, . . . , 0) it follows that a = 0 holds.

3. Find the smallest eigenvalue of the variational inequality

x∈C: hAx, y − xi ≥ λhx, y − xi for all y∈C ,

where C = {x ∈ R3 ; x1 ≥ 0 and x3 ≤ 0} and A is the matrix

 
2 −1 0
A =  −1 2 −1  .
0 −1 2
2.3. SUFFICIENT CONDITIONS 59

2.3 Sufficient conditions

As in the unconstrained case we will see that sufficient conditions are close
to necessary conditions of second order. Let x be a local minimizer of f ∈ C 2
in the unconstrained case V ≡ Rn , then f ′ (x) = 0 and hf ′′ (x)z, zi ≥ 0 for
all z ∈ Rn . That means, the first eigenvalue λ1 of the Hessian matrix f ′′ (x)
is nonnegative. If λ1 > 0, then from expansion (1.2) follows that x defines
a strict local minimum of f in V .
Let V ⊆ Rn be a nonempty subset and suppose that x ∈ V satisfies the
necessary condition of first order, see Theorem 2.2.1:

(⋆) hf ′ (x), wi ≥ 0 for all w ∈ T (V, x).

We are interested in additional assumptions under which x defines a local

minimum of f in V . For the following reasoning we will need an assumption
which is stronger then the necessary condition (⋆).

Assumption A. Let w ∈ T (V, x) and xk and tk associated sequences. Then

there exists an M > −∞ such that

lim inf t2k hf ′ (x), xk − xi ≥ M .

k→∞

Remarks. (i) Assumption A implies that the necessary condition (⋆) of

first order holds.

(ii) Assume the necessary condition (⋆) is satisfied and V is convex, then
assumption A holds with M = 0.

The following subcone of T (V, x) plays an important role in the study

of sufficient conditions, see also Chapter 3, where the infinitely dimensional
case is considered.

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} .

Corollary. Suppose that assumption A is satisfied (this is the case if V is

60 CHAPTER 2. FUNCTIONS OF N VARIABLES

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. ✷

From an indirect argument follows a sufficient criterion.

Theorem 2.3.1. Suppose that f ∈ C 2 (Rn ). Then a nonisolated x ∈ V

defines a strict local minimum of f in V if Tf ′ (V, x) = {0} or if assumption
A is satisfied for each w ∈ Tf ′ (V, x) with an M ≥ 0 and hf ′′ (x)w, wi > 0
holds for all w ∈ Tf ′ (V, x) \ {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

For a subsequence we have tk (xk − x) → w, ||w|| = 1. The above inequality

implies that w ∈ Tf ′ (V, x). Since assumption (A) is satisfied with M ≥ 0
it follows that 0 ≥ hf ′′ (x)w, wi, a contradiction to the assumption of the
theorem. Since w ∈ Tf ′ (V, x) 6= {0} if x is no strict local minimizer, it
follows that x defines a strict local minimum if Tf ′ (V, x) = {0}. ✷

The following example shows that Tf ′ (V, x) can be a proper subset of C(V, x)∩
f ′ (x)⊥ .

Example. Let f (x) = x2 − c(x21 + x22 ), c > 0, and V = {x ∈ R2 : 0 ≤ x1 <

∞ and xα1 ≤ x2 < ∞}, where 1 < α < ∞. Since f ′ (0) = (0, 1), the vector
2.3. SUFFICIENT CONDITIONS 61

x = (0, 0) satisfies the necessary condition hf ′ (0), y − 0i ≥ 0 for all y ∈ V .

We ask whether x = (0, 0) defines a local minimum of f in V . A corollary
from above implies that Tf ′ (V, x) = {0} or that Tf ′ (V, x) = {y ∈ R2 ; y2 =
0 and y1 ≥ 0}. If 1 < α < 2, then Tf ′ (V, x) = {0}, see an exercise. In this
case (0, 0) defines a strict local minimum of f in V , see Theorem 2.3.1. In
the case 2 ≤ α < ∞ we find that the vector (1, 0) is in Tf ′ (V, x). Thus the
assumption of Theorem 2.3.1 is not satisfied since
 
′′ 1 0
f (0) = −2c .
0 1
By taking a sequence y = (x1 , xα1 ), x1 → 0, it follows that (0, 0) defines no
local minimum if 2 < α < ∞. In the borderline case α = 2 it depends on c
whether or not (0, 0) defines a local minimum.

2.3.1 Equality constraints

We assume here that the subset V is given by
V = {y ∈ Rn : gj (y) = 0, j = 1, . . . , m} ,
where f and gj are C 2 functions. Set
m
X
L(y, λ) = f (y) + λj gj (y) ,
j=1

where λ = (λ1 , . . . , λm ), λj ∈ R.

Theorem 2.3.2 (Sufficient condition, equality constraints). Assume a non-

isolated x ∈ V satisfies L′ (x, λ) = 0 and hL′′ (x, λ)z, zi > 0 for all z ∈
T (V, x) \ {0}. Then x defines a strict local minimum of f in V .

Proof. Since f (y) = L(y, λ) for all y ∈ V we obtain for x, xk ∈ V

1
f (xk ) − f (x) = hL′ (x, λ), xk − xi + hL′′ (x, λ)(xk − x), xk − xi
2
+o(||xk − x||2 )
1 ′′
= hL (x, λ)(xk − x), xk − xi + o(||xk − x||2 ) .
2
If x is no strict local minimizer, then there exists a sequence xk ∈ V , xk → x
and xk 6= x such that
1
0 ≥ hL′′ (x, λ)(xk − x), xk − xi + o(||xk − x||2 )
2
62 CHAPTER 2. FUNCTIONS OF N VARIABLES

holds, which implies that

1 o(||xk − x||2 )
0 ≥ hL′′ (x, λ)tk (xk − x), tk (xk − x)i + ,
2 ||xk − x||2

where tk := ||xk −x||−1 . For a subsequence we have tk (xk −x) → w, ||w|| = 1

and w ∈ T (V, x), which is a contradiction to the assumption of the theorem.
✷

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.

2.3.2 Inequality constraints

Here we use the notations of previous sections (inequality constraints, nec-

essary conditions).

Theorem 2.3.3 (Sufficient condition, inequality constraints). Suppose that

x ∈ V is non-isolated and that there exist multipliers λj such that L′ (x, λ) =
0, where
X
L(y, λ) = f (y) + λj gj (y) ,
j∈E∪I0

λj ≥ 0 if j ∈ I0 . Let I0+ be the subset of I0 defined by I0+ = {j ∈ I0 ; λj > 0}.

Set T0 ≡ {z ∈ T (V, x) : hgj′ (x), zi = 0 for each j ∈ I0+ }. Then x is
strict local minimizer of f in V if T0 = {0} or if hL′′ (x, λ)z, zi > 0 for all
z ∈ T0 \ {0}.

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.

3. Assume f ∈ C 1 (Rn ) and that V is given by V = {y ∈ Rn : ai ≤ yi ≤

bi }. Show that the variational inequality x ∈ V : hf ′ (x), y − xi ≥ 0
for all y ∈ V is equivalent
P to the corresponding Lagrange multiplier
equation f ′ (x) = − j∈I0 λj ej , where λj ≥ 0 if xj = bj and λj ≤ 0 if
xj = aj . The index set I0 denotes the set of active indices.
2.4. KUHN-TUCKER THEORY 65

2.4 Kuhn-Tucker theory

Here we consider again the problem of maximizing a real function under side
conditions given by inequalities. Set

V = {y ∈ X : g j (y) ≥ 0, j = 1, . . . , m},

where X ⊂ Rn is a given nonempty convex set, and consider the maximum

problem

(P ) maxy∈V f (y).

In contrast to the previous considerations, we do not assume that f or gj

are differentiable. But the most theorems in this section require that f and
g j are concave functions. Define, as in the previous section, the Lagrange
function
m
X
L(x, λ) = f (x) + λj g j (x).
j=1

Definition. A couple (x0 , λ0 ), where x0 ∈ X and λ0 ≥ 0 is called saddle

point of L(x, λ) if
L(x, λ0 ) ≤ L(x0 , λ0 ) ≤ L(x0 , λ)

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

Figure 2.3: Saddle point

content of the Kuhn-Tucker theory.

66 CHAPTER 2. FUNCTIONS OF N VARIABLES

Theorem 2.4.1. Suppose that (x0 , λ0 ) is a saddle point. Then

g j (x0 ) ≥ 0, j = 1, . . . , m,
m
X
λ0j g j (x0 ) = 0.
j=1

Proof. The assumption says that

m
X m
X m
X
f (x) + λ0j g j (x) 0
≤ f (x ) + λ0j g j (x0 ) 0
≤ f (x ) + λj g j (x0 )
j=1 j=1 j=1

for all x ∈ X and for all λ ≥ 0. Set λj = 0 if j 6= l, divide by λl > 0

and
Pm letting λl → ∞, we get g l (x0 ) ≥ 0 for every l. Set λ = 0. Then
0 j 0 0 j 0
j=1 λj g (x ) ≤ 0. Since λj ≥ 0 and g (x ) ≥ 0, the equation of the
theorem follows. ✷

Theorem 2.4.2. Suppose that (x0 , λ0 ) is a saddle point. Then x0 is a global

maximizer of f in V .

Proof. Since
m
X
f (x) + λ0j g j (x) ≤ f (x0 )
j=1

it follows that f (x) ≤ f (x0 ) for all x ∈ X satisfying g j (x) ≥ 0. ✷

The following is a basic result on inequalities in convex optimization. We

write w > 0 or sometimes w >> 0, if all coordinates of the vector w are
positive.

Theorem 2.4.3. Suppose that X ⊂ Rn is nonempty and convex and g j :

X ∈ R, j = 1, . . . , k, are concave. Assume there is no solution x ∈ X of the
system of inequalities g j (x) > 0, j = 1, . . . , k. Then there are λj ≥ 0, not
all of them are zero, such that

k
X
λj g j (x) ≤ 0
j=1

for all x ∈ X.
2.4. KUHN-TUCKER THEORY 67

Proof. Set g(x) = (g 1 (x), . . . , g k (x)) and define

Zx = {z ∈ Rk : z < g(x)}

and Z = ∪x∈X Zx . We have 0k 6∈ Z, otherwise 0 < g j (x) for an x ∈ X

and for all j, a contradiction to the above assumption. Since Z is convex,
see an exercise, it follows from a separation theorem for convex sets, see
Section 2.6, that there is a p0 6= 0 such that

hp0 , zi ≥ hp0 , 0i

for all z ∈ Z. We have p0 ≤ 0 since with z = (z1 , . . . , zl , . . . , zk ) ∈ Z

also z ′ = (z1 , . . . , t, . . . , zk ) ∈ Z for all t ≤ zl . Dividing by negative t and
let t → −∞ we find that every coordinate of p0 must be nonpositive. Set
p = −p0 , then hp, zi ≤ 0 for all z ∈ Z. Another representation of Z is

Z = {g(x) − ǫ : x ∈ X, ǫ > 0}.

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. ✷

Replacing g by −g, we get

Corollary. Suppose X ⊂ Rn is convex and all g j : X 7→ R are convex.

Then either the system g j (x) < 0, j =, . . . , k has a solution x ∈ X or there
is a p ≥ 0, not all coordinates zero, such that hp, g(x)i ≥ 0 for all x ∈ X.

The main theorem of the Kuhn-Tucker theory, is

Theorem 2.4.4 (Kuhn and Tucker [31]). Suppose that X ⊂ Rn is nonempty

and convex, and let f, g j : X 7→ R are concave. If x0 is a solution of problem
(P), then there exist non-negative constants p0 , p1 , . . . , pm , not all of them
are zero, such that
m
X
p0 f (x) + pj g j (x) ≤ p0 f (x0 ) for all x ∈ X and
j=1
m
X
pj g j (x0 ) = 0.
j=1
68 CHAPTER 2. FUNCTIONS OF N VARIABLES

Proof. By assumption there is no solution x ∈ X of the system g(x) ≥ 0 and

f (x)−f (x0 ) > 0. Then there is no solution of g(x) > 0 and f (x)−f (x0 ) > 0
too. Then there are non-negative constants p0 , p1 , . . . , pm , not all of them
are zero, such that
m
X
p0 (f (x) − f (x0 )) + pj g j (x) ≤ 0
j=1

for all x ∈ X, see Theorem 2.4.3. Set x = x0 , then it follows that

m
X
pj g j (x0 ) ≤ 0.
j=1

In fact, we have equality since pj ≥ 0 and g j (x0 ) ≥ 0. ✷

Under an additional assumption (Slater condition) we have p0 > 0.

Definition. We say that the system of inequalities g(x) ≥ 0 satisfies the

Slater condition if there exists an x1 ∈ X such that g(x1 ) > 0.

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

f (x) + hλ0 , g(x)i ≤ f (x0 ).

Since hλ0 , g(x0 )i = 0, we obtain that L(x, λ0 ) ≤ L(x0 , λ0 ), and L(x0 , λ0 ) ≤

L(x0 , λ) follows since λ ≥ 0 and g(x0 ) ≥ 0. ✷

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.

Proof. The lemma is a consequence of L(x0 , λ0 ) ≥ L(x, λ0 ) for all x ∈ X.

2.4. KUHN-TUCKER THEORY 69

Definition The following equations and inequalities are called Kuhn-Tucker

conditions for (x0 , λ0 ):
(i) hL′ (x0 , λ0 ), x − x0 i ≤ 0 for all x ∈ X,
(ii) g(x0 ) ≥ 0,
(iii) hλ0 , g(x)i = 0,
(iv) λ0 ≥ 0.

From the above Theorem 2.4.4, Theorem 2.4.5 and the previous lemma it
follows

Theorem 2.4.6 (A necessary condition). Assume X ⊂ Rn is convex, f , g j

are in C 1 and concave on X and that the Slater condition holds. If x0 is
a solution of the maximum problem (P), then there exists a vector λ0 such
that the Kuhn-Tucker conditions are satisfied.

Theorem 2.4.7 (A sufficient condition). Suppose that X is convex, f , g j

are concave and in C 1 . If (x0 , λ0 ) satisfies the Kuhn-Tucker conditions then
x0 is a global maximizer in X of f under the side conditions g(x) ≥ 0, that
is of the problem (P).

Proof. The function

L(x, λ0 ) = f (x) + hλ0 , g(x)i

is concave in x since λ0 ≥ 0. It follows that

L(x, λ0 ) − L(x0 , λ0 ) ≤ hL′ (x0 , λ0 , x − x0 i

≤ 0

for all x ∈ X. The second inequality is the first of the Kuhn-Tucker condi-
tions. On the other hand we have

L(x0 , λ) − L(x0 , λ0 ) = hLλ (x0 , λ0 ), λ − λ0 i

= hg(x0 ), λ − λ0 i
= hg(x0 ), λi
≥ 0.

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

Example: Profit maximizing

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

max (p q − hr, xi)

under the side conditions q ≤ f (x), x ≥ 0 and q ≥ 0.

Define the convex set

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.

Corollaries. (a) q 0 = 0 implies that x0 = 0.

(b) If q 0 > 0, then λ0 = p, f (x0 ) = q 0 and rj = pfxj (x0 ) if xj > 0.

Remark. If q 0 > 0 and x0 > 0, then x0 is a solution of the nonlinear

system pf ′ (x) = r. If the Hesse matrix f ′′ (x) is negative or positive definite
on x > 0, which implies that f is strictly concave, then solutions are uniquely
determined, see an exercise.
2.4. KUHN-TUCKER THEORY 71

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. Prove that solutions x ∈ Rn , x > 0, of f ′ (x) = b, where b ∈ Rn , are

uniquely determined if the Hesse matrix f ′′ (x) is negative or positive
definite for all x > 0.
72 CHAPTER 2. FUNCTIONS OF N VARIABLES

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

L(x, λ) = U (x) + λ(m − hp, xi) .

The necessary condition of first order is Uxj − λ0 pj = 0 for each j = 1, . . . , n,

λ0 ∈ R. Hence λ0 > 0 since Uxj > 0. The vector x0 is a strict local
maximizer if
hL′′ (x0 , λ0 )z, zi < 0
for all z ∈ Rn \ {0} satisfying hp, zi = 0. This equation is hg ′ (x), zi = 0,
where the side condition is given by g(x) ≡ hp, xi − m = 0. Or, equivalently,

hU ′′ (x0 )z, zi < 0 ∀z ∈ Rn , z 6= 0 and hU ′ (x0 ), zi = 0.

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

where p is a column vector and pT it’s transposed, is regular (Exercise).

From the implicit function theorem it follows that there exist continuously
differentiable demand functions xi = f i (p, m) and a continuously function
λ0 = f (p, m), |p − p0 | < δ, |m − m0 | < δ, δ > 0 sufficiently small, where
(x0 , λ0 ) is a solution of U ′ (x) − λp0 = 0 and hp0 , xi = m0 .
2.5. EXAMPLES 73

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

K = {y ∈ Rn ; hli , yi ≤ 0 for each i = 1, . . . , m and hf ′ (x), yi = 0} .

Suppose that K = {0} or, if K 6= {0}, that λ1 > 0, where

hf ′′ (x)y, yi
λ1 = min ,
y∈K\{0} hy, yi

then x is a strict local minimizer of f in V . If λ1 < 0, then x is no local

minimizer. Equivalently, x is a strict local minimizer if K = {0} or if the
lowest eigenvalue of the variational inequality

w ∈ K : hf ′′ (x)w, z − wi ≥ λhw, z − wi for all z ∈ K

is positive.

2.5.3 Eigenvalue equations

Consider the eigenvalue equation Ax = λBx, where A and B are real and
symmetric matrices with n rows (and n columns).

To illustrate the Lagrange multiplier method in the case of equations as

side conditions, we will prove the following well known result.

Assume B is positive definite, then there exist n eigenvalues λ1 ≤ λ2 ≤

. . . ≤ λn such that the associated eigenvectors x(k) are B-orthogonal, that is
hBx(k) , x(l) i = 0 if k 6= l. The k-th eigenvalue is given by
hAy, yi
λk = min ,
hBy, yi

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

Ax(2) = λ2 Bx(2) + µBx(1) .

By (scalar) multiplying with x(1) we obtain that µ = 0, and by multiply-

ing with x(2) we see that λ2 = min(1/2)hAy, yi under the side conditions
(1/2)hBy, yi = 1, hBx(1) , yi = 0, which implies that λ1 ≤ λ2 is satisfied.
Step 3. Assume x(k) , k ≤ n, is a minimizer of the problem min f (y) under
the side conditions g(y) = 1, hg ′ (x(1) ), yi = 0, . . . , hg ′ (x(k−1) , yi = 0, where
Bx(1) , . . . , Bx(k−1) are linearly independent and hBx(l) , x(m) i = 0 if l 6= m.
Then there exist λk , µ1 , . . . , µk−1 such that
k−1
X
(k) (k)
Ax = λk Bx + µl Bx(l) .
l=1

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

Thus we obtain n linearly independent eigenvectors x(l) , l = 1, . . . , n, which

satisfy hBx(k) , x(l) i = 2δkl , where δkl denotes the Kronecker symbol defined
by δkl = 1 if k = l and δkl = 0 if k 6= l. The associated eigenvalues satisfy
the inequalities λ1 ≤ λ2 ≤ . . . ≤ λn .

Another proof. Exploiting the special structure of the problem, we can

prove the above proposition more directly without relying on the Lagrange
2.5. EXAMPLES 75

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,

hA(x(k) + ǫy), x(k) + ǫyi

h(ǫ) = ,
hB(x(k) + ǫy), x(k) + ǫyi
where y is a fixed vector satisfying the side conditions

hg ′ (x(l) ), yi = 0 , 1 ≤ l ≤ k − 1 . (2.6)

Then h(0) ≤ h(ǫ), which implies that h′ (0) = 0 or

hAx(k) , yi = λk hBx(k) , yi (2.7)

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

holds. Since w satisfies the side condition (2.6) we have

hAx(k) , wi = λk hBx(k) , wi.

If 1 ≤ l ≤ k − 1, then

hBx(k) , x(l) i = hBx(l) , x(k) i = 0 (2.8)

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

y ∈ Rn . It follows, in particular, that λl hBx(l) , x(k) i = hAx(l) , x(k) i which

implies that λl hBx(k) , x(l) i = hAx(k) , x(l) i = 0, because of equation (2.8) and
since the matrices A, B are symmetric.

As a corollary to the above theorem we obtain the maximum-minimum

principle of Courant. The advantage of this principle is that we can define
the eigenvalue λk without the knowledge of the eigenvectors x(1) , . . . , x(k−1) .
For given k − 1 vectors z (l) ∈ Rn set
1
Vk−1 ≡ V (z (1) , . . . , z (k−1) ) : = {y ∈ Rn : hBy, yi = 1,
2
(l)
hBz , yi = 0, l = 1, . . . , k − 1}
and
1
Λk (z (1) , · · · , z (k−1) ) := min hAy, yi .
Vk−1 2
✷

Maximum-minimum principle of Courant. The k-th eigenvalue λk is

given by
λk = max Λk (z (1) , . . . , z (k−1) ) ,
where the maximum is taken over all (k−1)-tuples of vectors z (1) , . . . , z (k−1) .

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)

Since λk = Λk (x(1) , . . . , x(k−1) ), we can replace sup by min. ✷

2.5. EXAMPLES 77

2.5.4 Unilateral eigenvalue problems

Discretization of some obstacle problems in mechanics lead to the following
type of problems. Let A and B be real and symmetric matrices with n rows
(and n columns). Set as above
1 1
f (y) = hAy, yi, g(y) = hBy, yi, y ∈ Rn ,
2 2
and assume that the matrix B is positive definite and that the set of admis-
sible vectors is given by
V = {y ∈ Rn : ai ≤ yi ≤ bi , i = 1, . . . , n} ,
where ai ∈ [−∞, ∞) and bi ∈ (−∞, ∞] are given and satisfy ai < bi for
each i. If ak = −∞, then we suppose that yi satisfies the inequality −∞ <
yk , if bk = ∞, then yk < ∞, respectively. The set V is a closed convex
subset of Rn . Then we consider the eigenvalue problem
x∈V : hAx, y − xi ≥ λhBx, y − xi for all y∈V , (2.9)
that is, we seek a λ ∈ R such that (2.9) has a solution x 6= 0.
The constrained minimum problem
min f (y), (2.10)
y∈Ms

where Ms = {y ∈ V ; g(y) = s} for a given s > 0, is closely related to

the variational inequality (2.9) and vice versa. If x is a regular point with
respect to the side condition g(y) = s and the side conditions which define
V , then there exist λ0 , λj ∈ R such that
X
Ax = λ0 Bx − λj ej , (2.11)
j∈I0

where ej = (0, . . . , 0, 1, 0, . . . , 0) denotes the vectors of the standard basis in

Rn , and I0 denotes the set of indices where the constraints which define V
are active. One has λj ≥ 0 if xj = bj and λj ≤ 0 if xj = aj .
One finds easily that x is a regular point if and only if at least one
coordinate of Bx with an index k 6∈ I0 is not zero (exercises).

Thus we have shown

Assume that a solution x of the minimum problem (2.10) satisfies hBx, ek i = 6

0 for a k 6∈ I0 , then there exist λ0 , λj ∈ R, j ∈ I0 , such that the Lagrange
multiplier rule holds, where λj ≥ 0 if xj = bj and λj ≤ 0 if xj = aj .
78 CHAPTER 2. FUNCTIONS OF N VARIABLES

Another useful observation is that

the variational inequality (2.9) and the Lagrange multiplier equation (2.11)
are equivalent.

Proof. (i) Assume x ∈ V : hAx, y − xi ≥ λ0 hBx, y − xi for all y ∈ V .

Set I = {1, . . . , n} and I0a = {i ∈ I : xi = ai }, I0b = {i ∈ I : xi = bi }.
The variational inequality implies that (Ax)i = λ0 (Bx)i if i ∈ I \ (I0a ∪
I0b ), (Ax)i ≥ λ0 (Bx)i if i ∈ I0a , (Ax)i ≤ λ0 (Bx)i if i ∈ I0b . One can write
these inequalities as equation (2.11) with appropriate λj .
(ii) Multiplying the Lagrange multiplier equation (2.11) with y−x, we obtain
X X
hAx, y − xi − λ0 hBx, y − xi = − λj hej , y − xi − λj hej , y − xi ≥ 0
j∈I0a j∈I0b

since λj ≤ 0, hej , y − xi ≥ 0 if j ∈ I0a and λj ≥ 0, hej , y − xi ≤ 0 if j ∈ I0b .

✷

Now we consider the question whether a solution (x, λ0 ) ∈ V × R, x 6= 0, of

the variational inequality (2.9) or, equivalently, of the Lagrange multiplier
equation (2.11), defines a strict local minimum of the functional
1 λ0
F (y, λ0 ) := f (y) − λ0 g(y) ≡ hAy, yi − hBy, yi
2 2
in V = {y ∈ Rn ; ai ≤ yi ≤ bi }.

The phenomenon that an eigenvector defines a strict local minimum of the

associated functional F (y, λ0 ) is due to the side conditions. There is no such
behavior in the unconstrained case. In this case we have
1
F (x + y, λ0 ) = F (x, λ0 ) + hF ′ (x, λ), yi + hF ′′ (x, λ0 )y, yi
2
1 ′′
= F (x, λ0 ) + hF (x, λ0 )y, yi
2
= F (x, λ0 ) + hAy − λ0 By, yi .

Set y = ǫx, ǫ 6= 0, then we obtain F (x + ǫx, λ0 ) = F (x, λ0 ). Thus x is no

strict local minimizer of F (y, λ0 ).

In our example, the tangent cone T (V, x) is given by

T (V, x) = {y ∈ Rn : yj ≤ 0 if xj = bj and yj ≥ 0 if xj = aj }.
2.5. EXAMPLES 79

Let I0± = {j ∈ I0 : λj 6= 0}, where the λj are the multipliers in for-

mula (2.11), and set

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}.

2.5.5 Non-cooperative games

Non-cooperative games are games without binding agreements between the
players. A non-cooperative game consists of
(i) A set of n players N = {1, 2, . . . , n}.

(ii) A collection of nonempty strategy sets Si , where Si is the strategy set

of the i-th player and a subset of a Euclidean space, say of Rmi . The
set S = S1 ×S2 ×· · ·×Sn is the strategy set of the game and an element
si ∈ Si is a strategy for the player i and a point s = (s1 , s2 , . . . , sn ) ∈ S
is a strategy of the game.

(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

An oligopoly consists of n firms producing identical common products.

Firm i has a cost function ci (qi ) where qi is the amount of the product
which this firm produces. The inverse demand is given by p(r), where
r = q1 + q2 + . . . + qn . A inverse demand function is the inverse of the
demand function which gives the amount of this product that consumers
will buy as a function of its price. The players are the n firms. The player i
chooses the strategy of his strategy set Si = [0, Mi ], where Mi is the capacity
of the i-th firm. The payoff function of the i-th firm is its profit

πi (q) = p(q1 + . . . + qn )qi − ci (qi ).

80 CHAPTER 2. FUNCTIONS OF N VARIABLES

A generalization of an oligopoly is a monopolistic competition. Let pi =

pi (q1 , q2 , . . . , qn ) be the inverse demand function for the i-th firm. Then the
payoff functions of a monopolistic competition are given by

πi (q) = p(q1 , . . . , qn )qi − ci (qi ).

Definition. A point s∗ ∈ S is a Nash equilibrium of the game {Si , fi }i∈N if

for every i ∈ N

fi (s∗−i \ si ) ≤ fi (s∗ ) for all si ∈ Si .

To formulate the general existence result of Nash we define what indi-

vidually quasiconcave means.

Definition. A function f : V → R, where V is a convex subset of Rn , is

said to be quasiconcave on V if x1 , x2 ∈ V and f (x1 ) ≥ c, f (x2 ) ≥ c implies
that f (λx1 + (1 − λ)x2 ) ≥ c for all λ, 0 < λ < 1.

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, ∞].

Definition. The payoff function fi is said to be individually quasiconcave if

for each value s−i ∈ S−i the function fi (s−i \ t) is quasiconcave with respect
to t ∈ Si .

We recall that a function f is concave if −f is convex.

Theorem (Nash [46]). Let {Si , fi }i∈N be a non-cooperative game. Suppose

that the strategy sets Si ⊂ Rmi are convex and compact and that the payoff
functions fi are continuous and individually quasiconcave. Then a Nash
equilibrium exists.

Sketch of proof. Define multivalued mappings µi : S 7→ Si by

µi (s) = {x ∈ Si : fi (s−i \ x) = max fi (s−i \ y)}

y∈Si

n µ (s). Then s⋆ ∈ S is a Nash equilibrium if and only if

and set µ(s) = Xi=1 i
s is a fixed point of the multivalued mapping µ, that is, if s⋆ ∈ µ(s⋆ ) holds.
⋆
2.5. EXAMPLES 81

The existence of such a fixed point follows from the fixed point theorem of
Kakutani [27, 22]. ✷

From this theorem it follows the existence of an equilibrium of an oligopoly

or an monopolistic competition if the involved functions satisfy additional
assumptions.

In generalization to local minimizers we define a local Nash equilibrium.

Definition. A point s∗ ∈ S is a local Nash equilibrium for the game

{Si , fi }i∈N if there exists a ρ > 0 such that for every i ∈ N

fi (s∗−i \ si ) ≤ fi (s∗ ) for all si ∈ Si ∩ Bρ (s∗i ) ,

where Bρ (s∗i ) is a ball with center s∗i ∈ Si and radius ρ > 0.

From the above definition of a local equilibrium we obtain immediately

a necessary condition for a local equilibrium. Set fi , si (s) = ∇si fi (s), si has
mi coordinates si = (s1i , . . . , sm
i ). Then:
i

Suppose that fi ∈ C 1 and that s∗ defines a local equilibrium. Then for every
i ∈ N = {1, . . . , n}

hfi,si (s∗ ), wi ≤ 0 for all w ∈ T (Si , s∗i ) .

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

From Section 2.3 we obtain the following sufficient condition:

82 CHAPTER 2. FUNCTIONS OF N VARIABLES

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.

Let Si be the interval ai ≤ y ≤ bi . Then


⊥ ∗ ∗ R if fi,si (s∗ ) = 0
fi,si (s ) = {y ∈ R : fi,si (s )y = 0} = .
{0} if fi,si (s∗ ) 6= 0

The necessary conditions of first order are

fi,si (s∗ )(y − s∗i ) ≤ 0 for all y ∈ Si .

Let N1 , N2 ⊆ N be defined as follows: i ∈ N1 if fi,si (s∗ ) 6= 0 and i ∈ N2 if

fi,si (s∗ ) = 0. Then s∗ defines a local equilibrium if fi,si ,si (s∗ ) < 0 for every
i ∈ N2 , and s∗ is no equilibrium if fi,si ,si (s∗ ) > 0 for at least one i ∈ N2 .
2.5. EXAMPLES 83

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},

where aj < bj . Suppose that λ0 is an eigenvalue of the variational

inequality

x ∈ V : hAx, y − xi ≥ λhBx, y − xi for all y ∈ V.

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}.

(a) Show that x = (1, 1, 1) is an eigenvector to each eigenvalue λ0 ∈

[1, ∞) of the variational inequality

x ∈ V : hAx, y − xi ≥ λhBx, y − xi for all y ∈ V.

(b) Show that x = (a, 1, a), 0 < a <√1, is an eigenvector to the

eigenvalue λ = 2 − (1/a) if a satisfies 1/ 2 ≤ a < 1.

(c) Show that x = (1, a, 1), 0 < a < 1, is no eigenvector.

Hint: Use that the inequality is equivalent to a Lagrange multiplier
rule.
84 CHAPTER 2. FUNCTIONS OF N VARIABLES

4. Under which conditions on a define the eigenvectors and associated

eigenvalues of the previous exercise a strict local minimum of f (y) =
hAy, yi − λhBy, yi in V ?
Hint: Use Theorem 2.3.3 (sufficient criterion).
Pn
5. Show that a local equilibrium s∗ satisfies a system of i=1 mi equa-
tions if s∗ is an interior point of S.

6. Let

f1 (s) = 2 720 000 s1 − 33 600 s1 s2 − s41 ,

f2 (s) = 2 720 000 s2 − 33 600 s1 s2 − s42

and S1 = S2 = [−100, 90]. Show that s∗ = (−42.3582, 90) defines a

local equilibrium.

7. Consider the case n = 2 for an oligopoly, that is, a duopoly. Find

conditions under which Nash equilibria are no interior points.

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.

9. Consider the previous example. Find conditions under which Nash

equilibria are no interior points.

10. Let V = {y ∈ Rn : yi ≤ 1, i = 1, . . . , n} and set

 
2 −1 0 0 ···
 −1 2 −1 0 ··· 
A=  0 −1
 .
2 −1 · · · 
......................
2.5. EXAMPLES 85

Apply the projection-iteration method

xk+1 = pV (xk − q(Axk − λxk ))

of Section 2.2 to find eigenvectors of the variational inequality

x ∈ V : hAx − λx, y − xi ≥ 0 for all y ∈ V

which are no eigenvectors of the associated eigenvalue equation.

Remark. The above problem comes from a difference approximation

of the unilateral eigenvalue problem
Z b Z b
′′ ′′
u∈V : u (x)(v(x) − u(x)) dx ≥ λ u′ (x)(v(x) − u(x))′ dx
a a

for all v ∈ V , where V = {v ∈ H 2 (a, b)∩H01 (a, b) : v(x) ≤ 1 on (a, b)}.

11. Consider an oligopol with payoff functions

n
X
fi (x, y) = yi (a − b yk xx )xi − ci yi xi .
k=1

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

x⋆ : hg(x⋆ , y), x − x⋆ i ≥ 0 for all x ∈ V.

Apply the projection-iteration method xk+1 = pV (xk − qg(xk , y)), 0 <

q < ∞, of Section 2.2 to find local Nash equilibria of an example of
the above oligopol, i. e., for given data a, b, ci and yi .

Remark. According to the above sufficient criterion, x⋆ defines a local

Nash equilibrium if fi,xi ,xi (x⋆ , y) < 0 for all i where fi,xi (x⋆ , y) = 0.
In this example we have fi,xi ,xi = −2byi2 . Thus x⋆ is a local Nash
equilibrium since yi > 0.
86 CHAPTER 2. FUNCTIONS OF N VARIABLES

12. Consider the oligopoly of the previous exercise but with the additional
side condition
X n
a
yk x k ≤ .
b
k=1

Then the strategy set of the i-th firm is

( n
!)
1 a X
Si (x) = 0 ≤ xi ≤ 1 : 0 ≤ xi ≤ − yk x k + yi x i ,
yi b
k=1

i. e., in fact Si (x) does not depend on xi . Set

V (x) = S1 (x) × . . . × Sn (x).

Then we seek solutions of

x⋆ ∈ V (x⋆ ) : hg(x⋆ , y), x − x⋆ i ≥ 0 for all V (x⋆ ).

This variational inequality is equivalent to the fixed point equation

x = pV (x) (x − qg(x, y))

with a given positive constant q.

Find solutions x⋆ , for given data a, b, ci and yi , by using the iteration
procedure
xk+1 = pV (xk ) (xk − qg(xk , y)).

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

2.6 Appendix: Convex sets

Some existence results for systems of linear inequalities as well as Lagrange
multiplier rules for variational inequalities follow from separation theorems.

2.6.1 Separation of convex sets

Here we consider separations by hyperplanes. There is not always a separa-
tion by a hyperplane of two given sets, see Figure 2.4.

(a) (b)

Figure 2.4: Separation of sets

Definition. For given p ∈ Rn , p 6= 0, and real α the set

H(p, α) = {y ∈ Rn : hp, yi = α}

is called hyperplane.

Definition. A hyperplane H(p, α) separates two nonempty sets A, B ⊂ Rn

if one of the two conditions is satisfied for a p ∈ Rn , p 6= 0, and a real α:

(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.

A hyperplane H(p, α) separates strictly two nonempty sets A, B ⊂ Rn if

one of the two conditions is satisfied for a p ∈ Rn , p 6= 0, and a real α:
88 CHAPTER 2. FUNCTIONS OF N VARIABLES

(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.

Theorem 2.6.1 (Separation of a closed convex set and a point). Let X ⊂ Rn

be nonempty, closed and convex, and z 6∈ X. Then there exists a hyperplane
which separates X and z strictly.

Proof. There exists a solution of

x ∈ X : ||z − x||2 ≤ ||z − y||2 for all y ∈ X,

see Figure 2.5 for an illustration. Replacing y by x + λ(y − x), 0 ≤ λ ≤ 1,

x .
.z X

Figure 2.5: Projection of z onto 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. ✷

Definition. A hyperplane H(p, α) is called supporting plane of X at x if

hp, yi ≥ α for all y ∈ X and hp, xi = α

2.6. APPENDIX: CONVEX SETS 89

or
hp, yi ≤ α for all y ∈ X and hp, xi = α.

Theorem 2.6.2 (Supporting plane of closed convex sets). Suppose that X ⊂

Rn is nonempty, closed, convex, and that the boundary ∂X is nonempty. Let
x ∈ ∂X, then there exist a supporting plane of X at x.

Proof. Let x ∈ ∂X. Then there exists a sequence xk 6∈ X such that xk → x

as k → ∞. Without restriction of generality, we can assume, see Theo-
rem 2.6.1, that there exists hyperplanes H(pk , αk ) such that

hpk , yi ≥ αk ≥ hpk , xk i for all y ∈ X.

Moreover we can assume that ||pk || = 1 since

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. ✷

Remark. A supporting plane can be considered as a generalization of a

tangent plane in the case that this plane does not exist, see Figure 2.6.

Figure 2.6: Supporting planes

Theorem 2.6.3 (Separation of a point and a not necessarily closed convex

set). Suppose that X ⊂ Rn , not necessarily closed, is nonempty, convex and
that z 6∈ X. Then there exists a hyperplane which separates X and z.
90 CHAPTER 2. FUNCTIONS OF N VARIABLES

Proof. Assume z 6∈ cl X, where cl X denotes the closure of X. Then the

assertion follows from Theorem 2.2.9. In the case that z ∈ cl X the theorem
is a consequence of Theorem 2.2.10. ✷

This theorem implies the following more general result.

Theorem 2.6.4 (Minkowski). Suppose that X, Y ⊂ Rn , not necessarily

closed, are nonempty, convex and that X ∩ Y = ∅. Then there exists a
separating hyperplane.

Proof. Set S = X − Y . Since 0 6∈ X, there exists a hyperplane which

separates S and 0. That is, there is a p ∈ Rn , p 6= 0, such that hp, si ≥ hp, 0i
for all s ∈ S, or equivalently

hp, xi ≥ hp, yi

for all x ∈ X and for all y ∈ Y . Thus

inf hp, xi ≥ suphp, yi,

x∈X y∈Y

which implies that there exists an α such that

hp, xi ≥ α ≥ hp, yi

for all x ∈ X and for all y ∈ Y . ✷

Figure 2.7: Separation of convex sets

2.6. APPENDIX: CONVEX SETS 91

2.6.2 Linear inequalities

Important consequences from the previous separation results are theorems
about systems of linear inequalities.

Lemma. Let xl ∈ Rn , l = 1, . . . , k are given, and set

k
X
C := {x ∈ Rn : x = λl xl , λl ≥ 0},
l=1

then the cone C is closed.

Proof. The proof is by induction with respect to k.

(j)
(i) Let k = 1. Suppose that yj := λ1 x1 → y if j → ∞, then

(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

is closed by assumption. Consider a sequence y j → y as j → ∞. Then

′ ′
y j = xj + λ(j) xs , xj ∈ C ′ , λ(j) ≥ 0. (2.12)
′
Suppose first that the sequence λ(j) is unbounded. Let λ(j ) → ∞ for a
′
subsequence λ(j ) , then it follows from the above decomposition (2.12) that
′
x′j
lim = −xs .
j →∞ λj ′
′

That is, −xs ∈ C ′ since C ′ is closed. This is a contradiction to −xs 6∈ C ′ .

′
If the sequence λ(j) is bounded, then also the sequence xj , see the de-
composition (2.12). Then it follows from (2.12) that y = x′ + λ0 xs , where
x′ ∈ C ′ and λ0 ≥ 0. ✷
92 CHAPTER 2. FUNCTIONS OF N VARIABLES

Theorem 2.6.5. Let A = A(m, n) be a real matrix with m rows and n

columns and let b ∈ Rn . Then there exists a solution of Ay ≥ 0 and hb, yi < 0
if and only if there is no solution x ∈ Rm of AT x = b and x ≥ 0.

Proof. (i) Suppose that there is no solution of AT x = b and x ≥ 0. Set

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

hp, bi < α < hp, si

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

for all positive xj . It follows that hp, AT ej i ≥ 0 for every j = 1, . . . , m. Thus

p is a solution of Ay ≥ 0 and hb, yi < 0.
(ii) Suppose that there is a solution y 0 of Ay ≥ 0 and hb, yi < 0. Then there
is no solution of of AT x = b and x ≥ 0, x ∈ Rm . If not, then

hb, y 0 i = hAT x, y 0 i = hx, Ay 0 i ≥ 0.

The next theorem is a consequence of the previous result.

Theorem 2.6.6 (Minkowski-Farkas Lemma). Let A = A(m, n) be a real

matrix with m rows and n columns and let b ∈ Rn . Then hb, yi ≥ 0 for all
y ∈ Rn satisfying Ay ≥ 0 if and only if there exists an x ∈ Rm , x ≥ 0, such
that AT x = b.

Proof. (i) Suppose that hb, yi ≥ 0 for all y ∈ Rn satisfying Ay ≥ 0. If there

is no solution of AT x = b, x ∈ Rm , x ≥ 0, then the above Theorem 2.6.5
says that there is a solution of Ay ≥ 0 and hb, yi < 0, a contradiction to the
assumption.
(ii) Assume there exists an x0 ∈ Rm , x0 ≥ 0, such that AT x0 = b. If there
is a y ∈ Rn such that Ay ≥ 0 and hb, yi < 0, then there is no solution of
2.6. APPENDIX: CONVEX SETS 93

AT x = b and x ≥ 0, see Theorem 2.6.5, which is a contradiction to the

assumption. ✷

Another consequence of Theorem 2.6.5 is

Theorem 2.6.7 (Alternative Theorem). Either there exists a nonnegative

solution of AT x ≤ b or there is a non-negative solution of Ay ≥ 0 and
hb, yi < 0.

Proof. (i) Suppose that there is a non-negative solution x0 of AT x ≤ b. Set

z = b − AT x0 , then there exists a nonnegative solution of AT x + z = b.
Assume there is a non-negative solution y 0 of Ay ≥ 0 and hb, yi < 0, then

0 > hb, y 0 i = hAT x0 + z, y 0 i = hx0 , Ay 0 i + hz, y 0 i ≥ 0

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

Since there is no non-negative solution of B T w = b, we have a solution y 0

of By ≥ 0 and hb, yi < 0, see Theorem 2.2.13. Thus Ay ≥ 0 and y ≥ 0 since
these inequalities are equivalent to By ≥ 0. ✷

2.6.3 Projection on convex sets

Let pV (z) be the projection of z ∈ H, where H is a real Hilbert space, onto

a nonempty subset V ⊆ H defined by

||pV (z) − z|| = min ||y − z|| .

y∈V

This projection exists if H = Rn and if V is closed or in the case of a general

real Hilbert space if V is closed and convex.
Thus we have w = pV (z) if and only if w ∈ V solves

w ∈ V : ||w − z||2 ≤ ||y − z||2 for all y ∈ V.

94 CHAPTER 2. FUNCTIONS OF N VARIABLES

Set y = w + ǫ(y − x), 0 ≤ ǫ ≤ 1, then we observe that this inequality is

equivalent to

hpV (z) − z, y − pV (z)i ≥ 0 for all y ∈ V, (2.13)

see an exercise.

Corollary. The projection pV of a real Hilbert space H onto a closed

nonempty convex subset V is non-expansive, i. e.,

||pV (x) − pV (y)|| ≤ ||x − y||.

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].

Definition. The cone

K ∗ = {v ∈ H : hv, ui ≤ 0 for all u ∈ K}

is called polar cone to K.

Moreau’s decomposition lemma (Moreau [45]). For given u ∈ H there

are uniquely determined u1 ∈ K, u2 ∈ K ∗ satisfying hu1 , u2 i = 0, such that

u = u1 + u2 .

Moreover, u1 = pK (u) and u2 = pK ∗ (u).

Proof. (i) Existence of the decomposition. Set u1 = pK u, u2 = u − u1 .

Then, see (2.13), hu − u1 , v − u1 i ≤ 0 for all v ∈ K. Thus

hu2 , v − u1 i ≤ 0 for all v ∈ K. (2.14)

Replacing in (2.14) v by the admissible element

 
1 1
v + u1 ≡ 2 v + u1 ,
2 2
then
hu2 , vi ≤ 0 for all v ∈ K. (2.15)
2.6. APPENDIX: CONVEX SETS 95

Thus u2 ∈ K ∗ . Replacing v in (2.14) by tu1 , t > 0, we get

(1 − t)hu1 , u2 i ≤ 0,

which implies that hu1 , u2 i = 0.

(ii) Uniqueness and u1 = pK (u), u2 = pK ∗ (u). Suppose that u = u1 + u2 ,
where u1 ∈ K, u2 ∈ K ∗ and hu1 , u2 i = 0. Let v ∈ K, then

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*

Figure 2.8: Moreau’s decomposition lemma

2.6.4 Lagrange multiplier rules

There is a large variety of Lagrange multiplier rules for equations and in-
equalities, see for example [60]. We will present two Lagrange multiplier
rules. The following lemmas can easily extended to more than one side
conditions.
Let H be a real Hilbert space with the inner product hu, vi, u, v ∈ H.
Suppose that f (h) = 0 for all h ∈ V ∩ Z, where f is a bounded linear
functional on H, V ⊂ H a nonempty subspace, Z = {h ∈ H : g(h) = 0},
and g is another bounded linear functional defined on H. Then we have
96 CHAPTER 2. FUNCTIONS OF N VARIABLES

Lagrange multiplier rule (equation). There exists a real λ0 such that

f (w) + λ0 g(w) = 0

for all w ∈ V .

Proof. There are F, G ∈ cl V , where cl V denotes the closure of V with

respect to the Hilbert space norm, such that

f (h) = hF, hi, g(h) = hG, hi

for all h ∈ cl V . Set Y = span G, then cl V = Y ⊕ Y ⊥ . Then F = F1 + F2 ,

where F1 ∈ Y and F2 ∈ Y ⊥ Since hF, F2 i = 0, we get F2 = 0. Consequently
F + λ0 G = 0, or
hF, hi + λ0 hG, hi = 0
for all h ∈ cl V . ✷

Assume f (h) ≥ 0 for all h ∈ K ∩ Z, where K ⊂ V is a nonempty convex

cone with vertex at zero, Z = {h ∈ H : g(h) = 0} and f , g are bounded
linear functionals defined on H. We recall that K is said to be a cone
with vertex at zero if h ∈ K implies that t h ∈ K for all t > 0. By C ∗
we denote the polar cone of a cone with vertex at the origin. The polar
cone of a cone C ⊂ cl V with the vertex at zero is defined to be the cone
C ∗ = {v ∈ cl V : hv, wi ≤ 0 for all w ∈ C}.

Lagrange multiplier rule (variational inequality). Suppose that there is

an h0 ∈ K such that −h0 ∈ K and g(h0 ) 6= 0. Then there exists a real λ0
such that
f (w) + λ0 g(w) ≥ 0
for all w ∈ K.

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 ∗ ) .

Thus, since Z ∗ = span {G}, it follows

−F ∈ cl (K ∗ + span {G}) .
2.6. APPENDIX: CONVEX SETS 97

Then there are sequences zn ∈ K ∗ , yn ∈ span {G} such that zn + yn → −F

in cl V . If the sequence yn remains bounded, then there is a convergent
subsequence yn′ → y. Consequently zn′ → z ∈ K ∗ which implies that
−F ∈ K ∗ + y. Thus there is a real λ0 satisfying −F − λ0 G ∈ K ∗ , or
equivalently, hF + λ0 G, hi ≥ 0 for all h ∈ cl K.
Suppose that the sequence yn ∈ span {G} is unbounded. Set wn =
zn + yn , then wn − yn = zn ∈ K ∗ . Thus hwn − yn , hi ≤ 0 for all h ∈ cl K, or

hwn , hi − λn hG, hi ≤ 0

for all h ∈ cl K. Since |λn | → ∞, we get hG, hi ≤ 0 for all h ∈ cl K or

hG, hi ≥ 0 for all h ∈ cl K, which is a contradiction to the assumption of
the lemma. ✷

Extending [49], Corollary 11.25(b), p. 495, or [50], Corollary 16.4.2, p. 146,

to a real Hilbert space we get the following lemma.

Lemma. Let H be a real Hilbert space. Suppose that K1 , . . . , Km ⊂ H are

nonempty, closed and convex cones with vertex at the origin. Then

(K1 ∩ · · · ∩ Km )∗ = cl (K1∗ · · · + Km
∗
)

Proof. (i) The inclusion

(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.

2. Show that the closure of X is convex if X is convex.

3. Let xl ∈ Rn , l = 1, . . . , k, are given linearly independent vectors, and

set
X k
C := {x ∈ Rn : x = λl xl , λl ≥ 0}.
l=1
Show that C is closed by using the following hints.
P (j) j)
Hint: Let y j ∈ C, i. e., y j = kl=1 λl xl , λl ≥ 0, where λl ≥ 0, and
y j → y as j → ∞. Then consider two cases
(j)
(a) all sequences λl are bounded,
(b) not all of these sequences are bounded. Then set
(j) (j)
aj = max{λ1 , . . . , λk }

and divide y j by aj .

4. Suppose that V ⊂ H is a nonempty, convex and closed subset of a real

Hilbert space. Show that

w ∈ V : ||w − z||2 ≤ ||y − z||2 for all y ∈ V

is equivalent to

w ∈ V : hw − z, y − wi ≥ 0 for all y ∈ V.

5. Suppose that V ⊂ H is nonempty, convex and closed. Show that for

given z ∈ H there exists a solution of

min ||v − z||2 ,

v∈V

and this solution is uniquely determined.

Hint: Theorem of Banach-Saks: let V ⊂ H be closed and convex, then
V is weakly closed.

6. Show that the projection of a real Hilbert space on a nonempty closed

convex set is a nonexpansive mapping.
Hint: Use formula (2.13).
2.6. APPENDIX: CONVEX SETS 99

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.

9. Separation theorem. Let H be a real Hilbert space and V a nonempty,

closed and convex subset. Let w ∈ H and w ∈ / V . Show that there is
a real λ such that hp, yi ≤ λ < hp, wi for all y ∈ V .
Hint: Consider the minimum problem miny∈V ||y − v||2 and use the
Banach-Saks theorem that a closed convex subset is weakly closed.

10. Separation theorem. Let V in the previous exercise be a closed convex

cone C with vertex at zero. Then hp, yi ≤ 0 < hp, wi for all y ∈ C.

11. Generalization of the Lagrange multiplier rule for equations. Suppose

that f (h) = 0 for all h ∈ V ∩ Z, where Z = {h ∈ H : gj (h) = 0, j =
1, . . . N } and gj are bounded linear functionals on H. Then there exist
real λj such that
N
X
f (w) + λj gj (w) = 0
j=1

for all w ∈ V .

12. Generalization of the Lagrange rule for variational inequalities. Let

K ⊂ V be a convex cone with vertex at the origin. Suppose that
f (h) ≥ 0 for all h ∈ K ∩ Z, where Z = {h ∈ H : gj (h) = 0, j =
1, . . . N } and gj are bounded linear functionals on H. Assume there
are hl ∈ K, l = 1, . . . , N such that −hl ∈ K and gj (hl ) = δjl . Then
there exist real λj such that
N
X
f (w) + λj gj (w) ≥ 0
j=1

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.

3.1 Optima, tangent cones, derivatives

Let B be a real Banach space and H a real Hilbert space such that B ⊆ H
is continuously embedded, that is, ||v||H ≤ c||v||B for all v ∈ B. Moreover,
we assume that ||v||B 6= 0 implies ||v||H 6= 0 for v ∈ B.1
In most applications of this chapter we have B = C 1 [a, b] and H =
1
H (a, b), which is the Sobolev space of all functions v which have generalized
derivatives of first order which are together with the functions itselve in
L2 (a, b).
Let V ⊆ B be a nonempty subset and suppose that E : V 7→ R.

Definition. A u ∈ V is said to be a weak local minimizer of E in V if there

is a ρ > 0 such that

E(u) ≤ E(v) for all v ∈ V, ||v − u||B < ρ.

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

Remark. A local minimizer is said to be a strong local minimizer with

respect to a given norm which allows a larger class of comparison elements
as above. In the classical calculus of variations this norm is the C[a, b]-norm.

Definition. The local tangent cone T (V, u) of V at u ∈ V is the set of all

w ∈ H such that there exist sequences un ∈ V , tn ∈ R, tn > 0, such that
un → u in B and tn (un − u) ⇁ w in H.

Corollaries. (i) The set T (V, u) is a cone with vertex at zero.

(ii) If T (V, u) 6= {0} then u is not isolated.

(iii) Suppose that w 6= 0, then tn → ∞.

(iv) T (V, u) is weakly closed in H.

(v) T (V, u) is convex if V is convex.

(vi) Assume V is convex. Then

T (V, u) = {w ∈ H : there exist sequences un ∈ V, tn ∈ R, tn > 0,

such that tn (un − u) ⇁ w as n → ∞}.

Proof. Exercise.

Definition (Fréchet derivative). The functional E is said to be Fréchet

differentiable at u ∈ B if there exists a bounded linear functional l on B
such that
E(u + h) = E(u) + l(h) + o(||h||B ),
as ||h||B → 0.

Notation: l = DE(u) Fréchet derivative of E at u.

Definition (Gâteaux derivative). For t ∈ R and fixed h ∈ B set Φ(t) =

E(u + th). The functional E is said to be Gâteaux differentiable at u ∈ B
if there is a bounded linear functional l on B such that Φ′ (0) exists and
Φ′ (0) = l(h).
3.1. OPTIMA, TANGENT CONES, DERIVATIVES 103

Notation: l = E ′ (u) Gâteaux derivative of E at u.

Corollaries.

(i) If f is Fréchet differentiable at u then f is Gâteaux differentiable at u.

(ii) If E ′ exists and is continuous in a neighborhood of u, then E ′ (u) =

DE(u).

Proof. Exercise.

Definition (First and second variation). The derivative

 
d
δE(u)(h) := E(u + ǫh) ,
dǫ ǫ=0

if it exists, is said to be the first variation (or first Gâteaux variation) of E

at u in direction h.
The derivative  2 
2 d
δ E(u)(h) := E(u + ǫh) ,
dǫ2 ǫ=0
if it exists, is said to be the second variation (or second Gâteaux variation)
of E at u in direction h.
The limits, if they exist,

E(u + th) − E(u)

δ + E(u)(h) = lim
t→0,t>0 t
and
E(u + th) − E(u)
δ − E(u)(h) = lim
t→0,t<0 t
are called right variation and left variation, respectively.

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.

2. Show that E ′ (x) = Df (u) if E ′ (v) exists and is continuous in a neigh-

borhood of u.

3. Show that, in general, the existence of the Gâteaux derivative does

not imply the existence of the Fréchet derivative.
Hint: Consider X = R2 and the derivatives of f at (0, 0), where
(  
y1 y22 2
2 : (y1 , y2 ) 6= (0, 0)
f (y) = y1 +y2
0 : (y1 , y2 ) = (0, 0)

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.

6. (i) Show that δE(u)(h) is homogeneous of degree one, i. e.,

δE(u)(λh) = λδE(u)(h)

for all λ ∈ R.
(ii) Show that the right variation is positive homogeneous of degree
one.

7. Show that δ 2 E(u)(h) is homogeneous of degree two.

8. Set Φ(t) = E(u + th) and suppose that Φ ∈ C 2 in a neighborhood of

t = 0 . Show that
t2 2
E(u + th) = E(u) + tδE(u)(h) + δ E(u)(h) + ǫ2 (th),
2
where limt→0 ǫ2 (th)/t2 = 0 for fixed h.
3.2. NECESSARY CONDITIONS 105

3.2 Necessary conditions

Let u, h ∈ B, and assume that the expansion

E(u + h) = E(u) + hE ′ (u), hi + η(||h||B )||h||H (3.1)

holds as ||h||B → 0, where limt→0 η(t) = 0, and hE ′ (u), hi is a bounded

linear functional on B which admits an extension to a bounded linear func-
tional on H.

This assumption implies that E is Fréchet differentiable at u.

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).

Theorem 3.2.1 (Necessary condition). Let V ⊂ B be a nonempty subset

and suppose that u ∈ V is a weak local minimizer of E in V , then

hE ′ (u), wi ≥ 0 for all w ∈ T (V, u).

Proof. Let tn , un are associated sequences to w ∈ T (V, u). Then, if n is

sufficiently large,

E(u) ≤ E(un ) = E(u + (un − u))

= E(u) + hE ′ (u), un − ui + η(||un − u||B )||un − u||H ,

thus

0 ≤ hE ′ (u), un − ui + η(||un − u||B )||un − u||H ,

0 ≤ hE ′ (u), tn (un − u)i + η(||un − u||B )||tn (un − u)||H .

Letting n → ∞, the theorem is shown. ✷

3.2.1 Free problems

Set Z b
E(v) = f (x, v(x), v ′ (x)) dx
a
106 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

and for given ua , ub ∈ R

V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub },
where −∞ < a < b < ∞ and f is sufficiently regular. See Figure 3.1 for
admissible variations. One of the basic problems in the calculus of variation
y

ub

ua

a b x

Figure 3.1: Admissible variations

is

(P ) minv∈V E(v).

It follows from the necessary condition (Theorem 3.2.1) that

Z b
 
fu (x, u(x), u′ (x))φ(x) + fu′ (x, u(x), u′ (x))φ′ (x) dx = 0 (3.2)
a
for all φ ∈ V − V , since the left hand side of (3.2) is equal to hE ′ (u), φi and
since V − V ⊂ T (V, u). The previous inclusion follows from Corollary (v)
of Section 3.1, or directly since for given v ∈ V we have n(un − u) = v − u,
where un := u + (v − u)/n, n an integer.

In our case of admissible comparison functions we can derive this equa-

tion under weaker assumptions.

Definition. A u ∈ V is said to be a weak local minimizer of E in V if there

exists an ǫ0 > 0 such that
E(u) ≤ E(v) for all v ∈ V : ||v − u||C 1 [a,b] < ǫ.
3.2. NECESSARY CONDITIONS 107

A u ∈ V is called a strong local minimizer of E in V if there exists an ǫ0 > 0

such that
E(u) ≤ E(v) for all v ∈ V : ||v − u||C[a,b] < ǫ.

We say that u ∈ V is a local minimizer if u is a weak or a strong local

minimizer.

Corollary. A strong local minimizer is a weak local minimizer.

Theorem 3.2.3. Let u ∈ V be a local minimizer of (P). Assume the first

variation of E at u in direction φ ∈ V − V exists, then equation (3.2) holds.

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). ✷

Definition. A solution u ∈ V of equation (3.2) is said to be a weak extremal.

From the basic lemma in the calculus of variations, see Chapter 1, it

follows that a weak extremal satisfies the Euler equation

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).

Lemma (Du Bois-Reymond). Let h ∈ C[a, b] and

Z b
h(x)φ′ (x)dx = 0
a

for all φ ∈ Φ, then h =const. 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

Since φ0 ∈ Φ, in particular φ0 (a) = φ0 (b) = 0, it follows

Z b Z b Z b Z b
1
0= hφ′0 dx = φ′0 (x)2 dx + h(ζ) dζ φ′0 (x) dx
a a b−a a a
Z b
= φ′0 (x)2 dx.
a

Thus Z b
1
h(x) = h(ζ) dζ.
b−a a
✷

Theorem 3.2.5 (Regularity). Suppose that f ∈ C 2 and that u ∈ C 1 [a, b] is

a weak extremal. Assume

fu′ u′ (x, u(x), u(x)) 6= 0

on [a, b]. Then u ∈ C 2 [a, b].

Proof. Set Z x
P (x) = fu (ζ, u(ζ), u′ (ζ)) dζ.
a

Then (3.2) is equivalent to

Z b
(−P + fu′ )φ′ dx = 0
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

Let x0 ∈ [a, b] and p0 = u′ (x

0 ). Since F (x0 , p0 ) = 0 and Fp (x0 , p0 ) =
fu′ u′ (x0 , u(x0 ), u′ (x0 )), it follows from the implicit function theorem that
there is a unique p = p(x), p ∈ C 1 in a neighborhood of x0 such that
p(x0 ) = p0 and F (x, p(x)) ≡ 0 in a neighborhood of x0 . The uniqueness
implies that p(x) = u′ (x) in a neighborhood of x0 . ✷

Corollary. Suppose f ∈ C m in its arguments, m ≥ 2, then u ∈ C m [a, b].

3.2. NECESSARY CONDITIONS 109

Proof. Exercise.

Example: How much should a nation save?

This example was taken from [55], pp. 13. Let

K = K(t) be the capital stock of the nation at time t,
C(t) consumption,
Y = Y (t) net national product.
We assume that Y = f (K), where f is sufficiently regular and satisfies
f ′ (K) > 0 and f ′′ (K) ≤ 0. Then the national product is a strictly increasing
concave function of the capital stock. Further we assume that

C(t) = f (K(t)) − K ′ (t),

which means that ”consumption=net production - investment”.

U (C) denotes the utility function of the nation. We suppose that U ′ (C) > 0
and U ′′ (C) < 0,
ρ denotes the discount factor.
Set
V = {K ∈ C 1 [0, T ] : K(0) = K0 , K(T ) = KT },
where T > 0, K0 and KT are given, and let
Z T


E(K) = U f (K(t)) − K ′ (t) e−ρt dt.
0

Then we consider the maximum problem

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

It follows from the above assumption that a weak extremal K0 satisfies

FK ′ < 0, FK ′ K ′ > 0 on [0, T ]. Consequently a weak extremal is in C 2 [0, T ]
if the involved functions are sufficiently regular.

The above assumptions imply that

Z T
′′
hE (K)ζ, ζi ≡ (FKK ζ 2 + 2FKK ′ ζζ ′ + FK ′ K ′ ζ ′2 ) dt
0
≤ 0

for all K ∈ V and for all ζ ∈ V − V . If additionally f ′′ < 0, then

Z T
hE ′′ (K)ζ, ζi ≤ −c(K, T ) ζ 2 dt
0

for all K ∈ V and for all ζ ∈ V − V , c(K, T ) is a positive constant, see an

exercise. As a consequence we get the following result.

A weak extremal K0 ∈ V is a global maximizer of E(K) in V . If additionally

f ′′ < 0, then weak extremals are uniquely determined.

Proof. Set h(t) = E(K0 + t(K − K0 )). Then

Z t
′
h(t) − h(0) = h (0)t + (t − s)h′′ (s) ds.
0

Thus

E(K) − E(K0 ) = hE ′ (K0 ), K − K0 i

Z 1
+ (1 − s)hE ′′ (K0 + s(K − K0 ))(K − K0 ), K − K0 i ds.
0

Consider again the general functional

Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a

where v ∈ V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub }. We will see in the

next section that the following necessary condition of second order is close
3.2. NECESSARY CONDITIONS 111

to a sufficient condition for a weak local minimizer. Set

Z b
hE ′′ (u)φ, φi = fu′ u′ (x, u(x), u′ (x))φ′ (x)2
a
+2fuu′ (x, u(x), u′ (u))φ(x)φ′ (x)


+fuu (x, u(x), u′ (x))φ(x)2 dx.

Theorem 3.2.6 (Necessary condition of second order). Let u ∈ V be a local

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. ✷

From this necessary condition it follows a condition which is close to the

assumption from which regularity of a weak extremal follows.

Theorem 3.2.7 (Legendre condition). Assume u ∈ V satisfies the necessary

condition of the previous theorem. Then

fu′ u′ (x, u(x), u′ (x)) ≥ 0

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]

where x1 = x1 (h) ∈ [x0 − h, x0 + h]. Then divide by h and letting h to zero.

(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
✷

Definition. A weak extremal is said to be satisfying the Legendre condition

if fu′ u′ (x, u(x), u(x)) ≥ 0 on [a, b] and it satisfies the strict Legendre condi-
tion if fu′ u′ > 0 on [a, b].

From the regularity theorem (Theorem 3.2.5) it follows immediately

Corollary. If f ∈ C 2 and an extremal u satisfies the strict Legendre condi-

tion, then u ∈ C 2 [a, b].

In the following we will derive a further necessary condition which follows

from hE ′′ (u)φ, φi ≥ 0 for all φ ∈ Φ. From the strict inequality for all
φ ∈ Φ \ {0} it follows that u defines a strict weak local minimizer provided
the strict Legendre condition is satisfied. Set

R = fu′ u′ (x, u(x), u′ (x)),

P = fuu (x, u(x), u′ (x)),
Q = fuu′ (x, u(x), u′ (x)).

Suppose that u ∈ C 2 [a, b]. This assumption is satisfied if u is a weak ex-

tremal and if R 6= 0 on [a, b], see Theorem 3.2.5 (regularity). Set
d
S=P− Q,
dx
then the second variation is
Z b
′′
hE (u)φ, φi = (Rφ′2 + Sφ2 ) dx.
a
3.2. NECESSARY CONDITIONS 113

We recall that φ(a) = φ(b) = 0.

Definition. The Euler equation

d
Lv ≡ (Rv ′ ) − Sv = 0
dx

associated to the second variation is called Jacobi equation.

Consider the initial value problem for the Jacobi equation

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.

Definition. The lowest zero ζ, a < ζ, of the solution of (3.3) is said to be

conjugate point of a with respect to L.

Theorem 3.2.8 (Necessary condition of Jacobi). Assume hE ′′ (u)φ, φi ≥ 0

for all φ ∈ Φ and fu′ u′ (x, u(x), u′ (x)) > 0 on [a, b]. Then ζ ≥ b.

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

constant κ will be determined later. Then

Z ζ Z b
′′ ′2 2
hE (u)w, wi = (Rw + Sw ) dx + (Rw′2 + Sw2 ) dx
a ζ
Z ζ
= − wLw dx + (Rw′ w)(ζ − 0)
a
Z b
− wLw dx − (Rw′ w)(ζ + 0)
ζ
Z ζ
= − κ(v + κh)Lw dx + κR(ζ)(v ′ (ζ) + κh′ (ζ))h(ζ)
a
Z b
2
−κ hLh dx − κ2 R(ζ)h′ (ζ)h(ζ)
ζ
Z b Z ζ
′ 2
= κR(ζ)v (ζ)h(ζ) − κ hLh dx − κ vLh dx
a a
 Z b 
′
= κ 2R(ζ)v (ζ)h(ζ) − κ hLh dx
a
< 0

for all 0 < κ < κ0 , κ0 sufficiently small. We recall that R(ζ) > 0, v ′ (ζ) < 0
and h(ζ) > 0. ✷

Definition. The inequality ζ > b is called strict Jacobi condition.

If the strict Jacobi condition is satisfied, then there is a solution of the

Jacobi equation which is positive on the closed interval [a, b]. Once one has
such a positive solution then we can rewrite the second variation from which
it follows immediately that this form is positive if φ 6≡ 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

Let z ∈ C 1 [a, b] be an arbitrary function. Since

d
(zφ2 ) = 2zφφ′ + z ′ φ2
dx
it follows for φ ∈ Φ that
Z
(2zφφ′ + z ′ φ2 ) dx = 0.

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. From (3.6) it follows hE ′′ (u)φ, φi ≥ 0 and ” = ” if and only if

φ′ + (z/R)φ = 0 on [a, b]. Since φ(a) = 0, this differential equation implies
that φ is identically zero on [a, b]. ✷
116 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

3.2.2 Systems of equations

Set Z b
E(v) = f (x, v(x), v ′ (x) dx,
a
and v(x) = (v1 (x), . . . , vm (x)), v ′ (x) = (v1′ (x), . . . , vm
′ (x)). Let

V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub },

where ua , ub ∈ Rm are given.

Theorem 3.2.10. Suppose that u ∈ V is a C 2 (a, b) local minimizer of E(v)

in V , then u satisfies the system of Euler differential equations
d
f ′ = fu j
dx uj
for j = 1, . . . , m.

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

on [a, b], see an exercise. A u ∈ V is said to be a weak extremal if

Z b m 
X 
fu′k φ′k + fuk φk dx = 0
a k=1

for all φ ∈ V − V . The condition

X
fu′i u′k ζi ζk ≥ 0 for all ζ ∈ Rm
i,k=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

Consider the problem of a Brachistochrone, see Section 1.2.2, to find a reg-

ular curve from

V = {(x(t), y(t)) ∈ C 1 [t1 , t2 ] : x′2 + y ′2 6= 0,

(x(t1 ), y(t1 )) = P1 , (x(t2 ), y(t2 )) = P2 }

which minimizes the functional

Z t2
E(x, y) = f (t, y, x′ , y ′ ) dt
t1

in the class V , where p

x′2 + y ′2
f=√ .
y − y1 + k
For notations see Section 1.2.2. Since fx = 0, it follows from an equation of
the system of Euler’s equations that (fx′ )′ = 0. Thus

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

Set x0 (τ ) = x(t(τ )) and y0 (τ ) = y(t(τ )). From (3.7) we get

1
y0 (τ ) − y1 + k = cos2 τ
a2
1
= (1 + cos(2τ )). (3.9)
2a2
Equation (3.9) implies that

y0′ (τ ) = −2α sin(2τ ), α := 1/(2a2 ),

and from (3.8) we see that

x′0 (τ ) = ±4α cos2 τ.

118 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Set 2τ = u − π, then it follows that

x − x1 + β = ±α(u − sin u)
y − y1 + k = α(1 − cos u),

where (x1 , y1 ) = P1 and β is a constant, and x(u) := x0 (τ ), y(u) := y0 (τ ).

Thus extremals are cycloids.

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),

where 0 ≤ u ≤ u1 . For given P2 = (x2 , y2 ) one finds u1 and α from the

nonlinear system

x2 = α(u − sin u)
y2 = α(1 − cos u),

see an exercise.

Example: N-body problem

(i) (i) (i)

Consider N mass points with mass mi located at x(i) = (x1 , x2 , x3 ) ∈ R3 .
Set
X mi mj
U =−
i6=j
|x(i) − x(j) |

and consider the variational integral

   
Z t2 N
X  dx(i) 2
 1  
E(x) = mi   − U (x) dt,
t1 2  dt 
i=1

where x = (x(1) , . . . , x(N ) ). The associated system of the 3N Euler equations

is
d2 x(i)
mi = −∇x(i) U.
dt2
3.2. NECESSARY CONDITIONS 119

3.2.3 Free boundary conditions

In previous sections there are ”enough” boundary conditions prescribed. In
many problems some further conditions follow from variational considera-
tions. A typical example is minv∈V E(v), where
Z b
E(v) = f (x, v(x), v ′ (x)) dx + h(v(a), v(b)).
a

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

Mixed boundary conditions

If we choose V = {v ∈ C 1 [a, b] : v(a) = ua }, where ua is prescribed, as the
admissible comparison set instead of C 1 [a, b], then a local minimizer of E
in V satisfies the weak Euler equation and the additional (free) boundary
condition fu′ = −hβ at x = b.

Proof. Exercise.

Higher order problems

Set Z b
E(v) = f (x, v(x), ..., v (m) (x)) dx
a

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.

Example: Bending of a beam.

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

where v ∈ C 2 [0, l], EI is a positive constant (bending stiffness), and f

denotes the force per unit length, see Figure (i). The Euler equation is here

EIu(4) = f on (0, l).

and the prescribed and free boundary conditions depend on how the beam
is supported, see the related figures.

(i) (ii)

(iii) (iv)

(i) Simply supported at both ends. Prescribed conditions: u(0) = 0, u(l) = 0,

free boundary conditions: u′′ (0) = 0, u′′ (l) = 0.

(ii) Clamped at both ends. Prescribed conditions: u(0) = u′ (0) = 0, u(l) =

u′ (l) = 0,
free boundary conditions: none.

(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.

3.2.4 Transversality conditions

The condition which we will derive here is a generalization of the previous
case (iv), where the right end of the curve can move freely on the target line
which is parallel to the y-axis.
122 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Definition. A curve γ in R2 is said to be a simple C 1 -curve if there is a

parameter representation v(t) = (v1 (t), v2 (t)), ta ≤ t ≤ tb , ta < tb , such that
vi ∈ C 1 [ta , tb ], v1′ (t)2 + v2′ (t) 6= 0 and v(t1 ) 6= v(t2 ) for all t1 , t2 ∈ [ta , tb ]
satisfying t1 6= t2 .

Remark. A regular parameter transformation t = t(τ ), i. e., a mapping

t ∈ C 1 [τa , τb ] satisfying t(τa ) = ta , t(τb ) = tb and t′ (τ ) 6= 0 on τa ≤ τ ≤ τb ,
τa < τb , maps a simply C 1 -curve onto a simple C 1 -curve.

Proof. Exercise.

Let γ = γ(τ ) be a given simple C 1 -curve and consider the set

V = {v : v = v(t), 0 ≤ t ≤ 1, simple C 1 − curve, v(0) = P, v(1) ∈ γ},

where P 6∈ γ is given. Let v ∈ V , then we consider the functional

Z 1
E(v) = f (t, v(t), v ′ (t)) dt,
0

f given and sufficiently regular. Set fv = (fv1 , fv2 ) and fv′ = (fv1′ , fv2′ ).

Theorem 3.2.11. Suppose that u ∈ V ∩ C 2 [0, 1] is a local minimizer of E

in V , then
d
(fu′ ) = fu on (0, 1)
dt
fu′ (1, u(1), u′ (1)) ⊥ γ.

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

v(t, τ0 ) = u(t), v(0, τ ) = P, v(1, τ ) = γ(τ ).

3.2. NECESSARY CONDITIONS 123

x2 γ

. γ(τ)
. γ(τ )
v
0
u

x
1

Figure 3.2: Admissible variations

For instance, such a family is given by

v(t, τ ) = u(t) + (γ(τ ) − γ(τ0 )) η(t),

where η(t), 0 ≤ t ≤ 1, is a fixed C 1 -function such that η(0) = 0 and η(1) = 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

where v ′ = vt . Integration by parts yields

Z 1   h it=1
d
fu − fu′ · vτ (t, τ0 ) dt + fu′ · vτ (t, τ0 ) = 0.
0 dt t=0

Since the system of Euler differential equations is satisfied and since v(0, τ ) =
P , |τ − τ0 | < ǫ, it follows

fu′ (1, u(1), u′ (1)) · vτ (1, τ0 ) = 0.

Finally, we arrive at the result of the theorem since v(1, τ ) = γ(τ ). ✷

Remark 1. If both ends move on curves γ1 , γ2 , respectively, see Figure 3.3,

then
fu′ ⊥ γ1 at t = 0, and fu′ ⊥ γ2 at t = 1,
if u(0) ∈ γ1 and u(1) ∈ γ2 .
124 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

x2 γ
2

γ
1

x
1

Figure 3.3: Both ends move on curves

x3
S

.
P x2

x
1

Figure 3.4: Target is a surface

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

Figure 3.5: Nonsmooth solutions

3.2.5 Nonsmooth solutions

Under additionally assumptions extremals of variational problems associated
to integrals of the type
Z b Z
′
f (x, v, v ) dx or F (x, v, ∇v) dx
a Ω

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.

Example. Consider the class

V = {v ∈ C[0, 1] : v piecewise C 1 , v(0) = v(1) = 0}.

A u ∈ C[a, b] is called piecewise in C s if there are at most finitely many

points 0 < t1 < t2 . . . < tm < 0 such that u ∈ C s [tk , tk+1 ], k = 0, . . . , m. Set
t0 = 0 and tm+1 = 1. For v ∈ V let
Z 1

2
E(v) = v ′ (x)2 − 1 dx.
0

There is a countable set of nonsmooth solutions, see Figure 3.5.

Let V be the class of functions v : [t1 , t2 ] 7→ Rn in C[t1 , t2 ], piecewise in C 1

and v(t1 ) = u1 , v(t2 ) = u2 , where u1 , u2 are given. Consider the functional
Z t2
E(v) = f (t, v(t), v ′ (t)))dt,
t1

where v ∈ V and f is given and sufficiently regular.

126 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

v2

.u(t )
0

v1

Figure 3.6: Corner of the extremal

Let u ∈ V be a weak extremal, that is,

Z t2


fu · φ + fu′ · φ′ dt = 0
t1

for all φ ∈ C01 (t1 , t2 ).

Theorem 3.2.12 (Weierstrass-Erdmann corner condition). Suppose that

u ∈ V and in C 2 on the closed subintervals where u is in C 1 , and that u′ is
possibly discontinuous at t0 ∈ (t1 , t2 ), see Figure 3.6, then
   
fu′ (t0 ) ≡ fu′ (t, u(t), u′ (t))t0 +0 − fu′ (t, u(t), u′ (t))t0 −0 = 0.

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

Figure 3.7: Corner of a graph extremal

for all φ(t0 ) ∈ Rn . ✷

As a corollary we derive a related condition for nonparametric integrands.

Set Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a
where v : [a, b] 7→ R, v ∈ V and V is defined by

V = {v ∈ C[a, b] : v piecewise in C 1 , v(a) = ua , v(b) = ub }.

Corollary. Suppose that u ∈ V satisfies u ∈ C 2 [a, c] and u ∈ C 2 [c, b], where

a < c < b, is a local minimizer of E in V , see Figure 3.7. Then
h  h 
fu′ (c) = 0 and f − u′ fu′ (c) = 0.

Proof. The formal proof is to replace x through x = x(t), a ≤ t ≤ b, where

x is a C 1 -bijective mapping from [a, b] onto [a, b] such that x′ 6= 0 on [a, b].
Then
Z b Z b  
′ y ′ (t)
f (x, v(x), v (x)) dx = f x(t), y(t), ′ x′ (t) dt,
a a x (t)
where y(t) = v(x(t). Set
 
y′
′ ′
F (x, y, x , y ) = f x, y, ′ x′ ,
x
128 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

 
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

x(ǫ; t) = t + ǫφ1 (t)

y(ǫ; t) = u(t) + ǫφ2 (t),

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”.

There is an interesting geometric interpretation of the conditions of the

corollary. Let u be an extremal and a < x0 < b.

Definition. The function

η = f (x0 , u(x0 ), ξ) =: h(ξ)

is called characteristic of f at (x0 , u(x0 )).

Let (ξi , ηi ), i = 1, 2, two points on the characteristic curve of f at (c, u(c)),

a < c < b, and let Ti tangent lines of the characteristic curve at (ξi , ηi ),
which are given by

η − ηi = fu′ (c, u(c), ξi )(ξ − ξi ).

3.2. NECESSARY CONDITIONS 129

ξ ξ ξ
1 2

Figure 3.8: Geometric meaning of the corollary

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.

As a consequence of this consideration we have:

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. ✷

Thus, √extremals of variational problems to the integrands f = v ′2 or f =

a(x, y) 1 + v ′2 , a > 0, have no corners. If the integrand is not convex for
130 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

all v ′ , then corners can occur as the example f = (v ′2 − 1)2 shows, see
Figure 3.5.

3.2.6 Equality constraints; functionals

In 1744 Euler considered the variational problem minv∈V E(v), where
Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a

v = (v1 , . . . , vn ). Let

V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub , gk (v) = 0, k = 1, . . . , m}

for given ua , ub ∈ Rn , and define gk by

Z b
gk (v) = lk (x, v(x), v ′ (x)) dx.
a

The functions f and lk are given and sufficiently regular.

Example: Area maximizing

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

Figure 3.9: Area maximizing

Example: Capillary tube

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

that is we prescribe the volume of the liquid. Let

Z p Z
κ 
E(v) = 1 + |∇v|2 + v 2 dx − cos γ v ds,
Ω 2 ∂Ω

where κ is a positive constant (capillary constant) and γ is the angle be-

tween the normals on the cylinder wall and on the capillary surface S at the
boundary of S. Then the variational problem is minv∈V E(v).

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.

Assumptions: (i) The functionals E and gj are Frechéchet differentiable at

u ∈ B.
132 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

x3

g
vapor

NΓ
S

NS

liquid

x2 Ω

x1

Figure 3.10: Capillary tube

(ii) For fixed u ∈ B and given φ1 , . . . , φm ∈ B the functions

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}.

Definition. A u ∈ V is said to be a local minimizer with respect to m-

dimensional variations of E in V if for given φ1 , . . . , φm ∈ B there ex-
ists an ǫ > 0 such that E(u) ≤ E(v) for all v ∈ V satisfying v − u ∈
span {φ1 , . . . , φm } and ||u − v||B < ǫ.

Theorem 3.2.13 (Lagrange multiplier rule). Let u ∈ V be a local minimizer

or maximizer with respect to m-dimensional variations of E in V . Then
3.2. NECESSARY CONDITIONS 133

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

Proof. We will show by contradiction that the functionals l0 = E ′ (u), l1 =

′ (u) are linearly dependent in B. Suppose that these
g1′ (u), . . . , lm = gm
functionals are linearly independent, then there are φj ∈ B, j = 0, 1, . . . , m,
such that li (vj ) = δij , see for example [28]. Set M = E(u) and consider for
small η ∈ R and c ∈ Rm the system of m + 1 equations
m
X
F (c) : = E(u + c j φj ) = M + η
j=0
Xm
Gi (c) : = gi (u + cj φj ) = 0.
j=0

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. ✷

Corollary. If g1′ (u), . . . , gm

′ (u) are linearly independent, then λ 6= 0.
0

3.2.7 Equality constraints; functions

Set Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a
where v = (v1 , . . . , vn ), and

V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub ,

lk (x, v(x)) = 0 on [a, b], k = 1, . . . , m}

ua , ub ∈ Rn , and lk and f are given sufficiently regular functions. We

assume m < n. The problem minv∈V E(v) is called Lagrange problem.
134 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Set
m
X
F (x, v, v ′ , λ) = f (x, v, v ′ ) + λk lk (x, v).
k=1

Theorem 3.2.14 (Lagrange multiplier rule). Let u be a local minimizer or

maximizer of E in V . Suppose that a fixed (m × m)-submatrix of lv (x, u(x))
is regular for all x ∈ [a, b]. Then there are functions λl ∈ C 1 [a, b] such that

d
Fu′ = Fu
dx
on (a, b).

Proof. Suppose that

∂(l1 , . . . , lm ) 

∂(v1 , . . . , vm ) v=u(x)
is regular for all x ∈ [a, b]. Choose n − m functions ηm+r ∈ C 1 , r =
1, . . . , n − m, satisfying ηm+r (a) = 0, ηm+r (b) = 0. Set

wm+r (x, ǫ) = um+r (x) + ǫηm+r (x),

where |ǫ| < ǫ0 , ǫ0 sufficiently small, and consider on [a, b] the system

lk (x, v1 , . . . , vm , wm+1 (x, ǫ), . . . , wn (x, ǫ)) = 0,

k = 1, . . . , m, for the unknowns v1 , . . . , vm . From the implicit function

theorem we get solutions vl = wl (x, ǫ), l = 1, . . . , m, vl ∈ C 1 on [a, b] ×
(−ǫ0 , ǫ0 ) satisfying vl (x, 0) = ul (x) on [a, b]. These solutions are uniquely
determined in a C-neighbourhood of u(x). Thus lk (x, w(x, ǫ)) = 0 on [a, b]
for every k = 1, . . . , m. We have

wm+r (a, ǫ) = um+r (a), wm+r (b, ǫ) = um+r (b). (3.11)

Hence, since the above solution is unique, we obtain for k = 1, . . . , m that

wk (a, ǫ) = uk (a), wk (b, ǫ) = uk (b). (3.12)

Thus w(x, ǫ) is an admissible family of comparison functions. Set

∂wl 
ηl (x) =  .
∂ǫ ǫ=0
3.2. NECESSARY CONDITIONS 135

From (3.11) and (3.12) we get for l = 1, . . . , n that

ηl (a) = 0, ηl (b) = 0. (3.13)

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

From lk (x, w(x, ǫ)) = 0 on [a, b] × (−ǫ0 , ǫ0 ) we obtain

n
X ∂lk
ηj = 0, k = 1, . . . , m.
∂vj
j=1

Multiplying these equations with functions λk ∈ C[a, b], we get

Xn Z b
∂lk
λk (x) ηj dx = 0, (3.15)
a ∂vj
j=1

k = 1, . . . , m. We add equations (3.14) and (3.14) and arrive at

Z b Xn  
Fuj ηj + Fu′j ηj′ dx = 0,
a j=1
P
where F = f + m ′
k=1 λk lk . We recall that lk are independent of u . Suppose
1
for a moment that λk ∈ C [a, b], then
Z b X n  
d
Fuj − F ′ ηj dx = 0 (3.16)
a dx uj
j=1

Since we can not choose the functions ηj arbitrarily, we determine the m

functions λk from the system
d
Fui − F ′ = 0, i = 1, . . . , m.
dx ui
That is, from the system
X m
d
fu i − fu′i + λk lk,ui = 0,
dx
k=1
136 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

P2
.

.
P1

Figure 3.11: Geodesic curve

i = 1, . . . , m. It follows that λk ∈ C 1 [a, b]. Then (3.16) reduces to

X
Z b n−m
d

Fum+r − Fu′m+r ηm+r dx = 0.
a dx
r=1

Since we can choose ηm+r arbitrarily, the theorem is shown. ✷

Example: Geodesic curves

Consider a surface S defined by φ(v) = 0, where φ : R3 7→ R is a C 1 -function

satisfying ∇φ 6= 0. Set
V = {v ∈ C 1 [t1 , t2 ] : v(t1 ) = P1 , v(t2 ) = P2 , φ(v) = 0}.
Then we are looking for the shortest curve v ∈ V which connects two given
points P1 and P2 on S, see Figure 3.11. The associated variational integral
which is to minimize in V is
Z t2 p
E(v) = v ′ (t) · v ′ (t) dt.
t1

A regular extremal satisfies

 
d u′
√ = λ∇φ.
dt u′ · u′
Choose the arc length s instead of t in the parameter representation of u,
then
u′′ (s) = λ(s)(∇φ)(u(s)),
which means that the principal normal is perpendicular on the surface S,
provided the curvature is not zero.
We recall that ∇φ ⊥ S, and that the principal curvature is defined by
u′′ (s)/κ(s), where κ(s) = ||u′′ (s)|| (Euclidean norm of u′′ ) is the curvature.
3.2. NECESSARY CONDITIONS 137

1 x

Figure 3.12: String above an obstacle

3.2.8 Unilateral constraints

Let u, h ∈ B, E : B 7→ R, and assume that the expansion

E(u + h) = E(u) + hE ′ (u), hi + η(||h||B )||h||H (3.17)

holds as ||h||B → 0, where limt→0 η(t) = 0 and hE ′ (u), hi is a bounded linear

functional on B which admits an extension to a bounded linear functional
on H.
Let V ⊂ B nonempty and suppose that u ∈ V is a weak local minimizer of
E in V . Then, see Theorem 3.2.1.

hE ′ (u), wi ≥ 0 for all w ∈ T (V, u).

If V is a convex set, then

hE ′ (u), v − ui ≥ 0 for all v ∈ V.

since v − u ∈ T (V, u) if v ∈ V .

Example: String above an obstacle

Let

V = {v ∈ C 1 [0, 1] : v(0) = v(1) = 0 and v(x) ≥ ψ(x) on (0, 1)},

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

and consider the variational problem minv∈V E(v). Suppose that u is a

solution, then u satisfies the variational inequality
Z 1
u∈V : u′ (x)(v(x) − u(x))′ dx ≥ 0 for all v ∈ V.
0

Remark. The existence follows when we consider the above problem in

the associated convex set in the Sobolev space H01 (0, 1). Then we find a
weak solution u ∈ H01 (0, 1) satisfying u(x) ≥ ψ(x) on (0, 1). Then from a
regularity result due to Frehse [17] we find that u ∈ C 1 [0, 1], provided ψ
is sufficiently regular. Such kind of results are hold also for more general
problems, in particular, for obstacle problems for the beam, the membrane,
the minimal surface or for plates and shells.

Example: A unilateral problem for the beam

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

V = {v ∈ H01 (0, l) ∩ H 2 (0, l) : v(x) ≤ d on (0, l)}

of admissible deflections if l/4 < k < l/2 and it is no local minimizer if

0 < k < l/4, see Section 3.3.3 or [40].

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

Figure 3.13: A unilateral beam

Example: Positive solutions of eigenvalue problems

Consider the eigenvalue problem

d

− p(x)u′ (x) + q(x)u(x) = λρ(x)u(x) in (a, b)
dx
u(a) = u(b) = 0.

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

Then the lowest eigenvalue λH , which is positive, is given by

a(v, v)
λH = min ,
v∈H\{0} b(v, v)

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

K = {v ∈ H01 (a, b) : v(x) ≥ 0 on (a, b)}.

Let
a(v, v)
λK = min .
v∈K\{0} b(v, v)
140 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

As in Chapter 2 we find that λK is the lowest eigenvalue of the variational

inequality

u ∈ H \ {0} : a(u, v − u) ≥ λb(u, v − u) for all v ∈ H.

Under the above assumptions we have

Proposition. λH = λK .

Proof. It remains to show that λH ≥ λK . Let uH be an eigenfunction to

λH , then
a(uH , v) = λH b(uH , v) for all v ∈ H. (3.18)
Moreau’s decomposition lemma, see Section 2.6.3, says that uH = u1 + u2 ,
where u1 ∈ K, u2 ∈ K ∗ and a(u1 , u2 ) = 0. We recall that K ∗ denotes the
polar cone associated to K. Inserting v = u1 into (3.18), we get

a(u1 , u1 ) = λH b(u1 , u1 ) + λH b(u2 , u1 )

≤ λH b(u1 , u1 )

since b(u2 , u1 ) ≤ 0, see an exercise. If u1 6= 0, then it follows that λK ≤ λH .

If u1 = 0, then uH ∈ K ∗ , which implies that −uH ∈ K, see an exercise. ✷

Remark. The associated eigenfunction has no zero in (a, b).

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

for all K ∈ V and for all ζ ∈ V − V . If additionally f ′′ < 0 is satisfied,

then Z T
hE ′′ (K)ζ, ζi ≤ −c(K, T ) ζ 2 dt
0
for all K ∈ V and for all ζ ∈ V − V , c(K, T ) is a positive constant.
2. Consider the example ”How much should a nation save?”. Find all
extremals if
1
U (C) = C 1−v and f (K) = bK,
1−v
where v ∈ (0, 1) and b are constants. Suppose that b 6= (b − ρ)/v.
3. Suppose that l1 , . . . , lN are linearly independent functionals on a Hilbert
space H. Show that there are v1 , . . . , vN ∈ H such that li (vj ) = δij .
4. Consider the example of area maximizing of Section 3.2.6. Show that
λ0 6= 0, where λ0 and λ1 are the Lagrange multipliers according to the
Lagrange multiplier rule of Theorem 3.2.1.
5. Consider the example of the capillary tube of Section 3.2.6. Show that
λ0 6= 0, where λ0 and λ1 are the Lagrange multipliers according to the
Lagrange multiplier rule of Theorem 3.2.1.
6. Weierstraß. Show that the integral
Z b
f (x(t), x′ (t)) dt,
a

where x(t) = (x1 (t), . . . , xn (t)), is invariant with respect to a regular

parameter transformation t = t(τ ), that is, t ∈ C 1 and t′ > 0, if and
only if f (x, p) is positive homogeneous in p, i. e., f (x, λp) = λf (x, p)
for all positive real λ.
Hint: To show that it follows that f (x, p) is positive homogeneous,
differentiate the integral with respect to the upper bound τb and then
consider the mapping t = τ /λ, where λ is a positive constant.
142 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

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

for all φ ∈ V − V implies the Legendre condition

X
fu′i u′k ζi ζk ≥ 0 for all ζ ∈ Rm .
i,k=1

Hint: Set φl = ζl φh (x), where ζ ∈ Rm and φh is the function defined

in the proof of Theorem 3.2.7.

9. Find the Brachistochrone if P1 = (0, 0), P2 = (1, 1) and v1 = 0.

10. Determine the shortest distance between two straight lines in R3 .

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

v = (v1 , v2 ) and g is continuous. Show that the corner points of

extremals are contained in the set {(x1 , x2 ) ∈ R2 : g(x1 , x2 ) = 0}.

14. Let u be a solution of the isoperimetric problem

Z 1
max v(x) dx,
v∈V 0

where RV = {v ∈ C 1 [a, b] : v(0) = v(1) = 0, l(v) = π/2} with

1 p
l(v) = 0 1 + (v ′ (x))2 dx.
Find u.
3.2. NECESSARY CONDITIONS 143

15. Let u be a solution of the isoperimetric problem

Z 1 p
min 1 + (v ′ (x))2 dx,
v∈V 0

where RV = {v ∈ C 1 [a, b] : v(0) = v(1) = 0, l(v) = π/8} with

1
l(v) = 0 v(x) dx.
Find u.

16. A geodesic on a surface defined by φ(x) = 0, x ∈ R3 and ∇φ 6= 0

satisfies, see Section 3.2.7,

u′′ (s) = λ(s)(∇φ)(u(s)).

Find λ.

17. Find geodesics u(s), 0 ≤ s ≤ L, of length 0 < L < πR on a sphere

with radius R.
Consider geodesics x(s) = (x1 (s), x2 (s), x3 (s)) on an ellipsoid E, de-
fined by
 x 2  x 2  x 2
1 2 3
+ + = 1,
a b c
where a, b, c are positive constants. Let

P1 , P2 ∈ E ∩ {x ∈ R3 : x1 = 0}.

Show that a geodesic connecting P1 and P2 , P1 6= P2 , satisfies x1 (s) ≡

0.

18. Set

V = {v ∈ C 1 [−1, −1] : v(−1) = v(1) = 0, v(x) ≥ −x2 +1/4 on (−1, 1)}.

Find the solution of Z 1

min (v ′ (x))2 dx.
v∈V −1

Is the solution in the class C 2 (0, 1)?

19. Set V = {v ∈ C[a, b] : ψ1 (x) ≤ v(x) ≤ ψ(x)}, where ψ1 and ψ2 are in

C[a, b], ψ1 (x) ≤ ψ2 (x) on (a, b), ψ1 is convex and ψ2 concave on [a, b].
Let
1
δh2 v = 2 (v(x + h) − 2v(x) + v(x − h))
h
144 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

be the central difference quotient of second order. For fixed ζ ∈

C0 (a, b), 0 ≤ ζ ≤ 1, define

vǫ = v + ǫζδh2 v,

where ǫ is a positive constant. We suppose that h is a sufficiently small

positive constant such that vǫ is defined on [a, b].
Show that vǫ ∈ V , provided that 0 ≤ ǫ ≤ h2 /2 is satisfied.

Remark. Such type of admissible comparison functions were used by

Frehse [17] to prove regularity properties of solutions of elliptic varia-
tional inequalities.

20. Consider the example ”Positive solutions of eigenvalue equations” of

Section 3.2.8. Show that u ≤ 0 on (a, b) if u ∈ K ∗ .

Hint: If u ∈ K ∗ , then u ∈ H01 (a, b) and

Z b
(pu′ v ′ + quv) dx ≤ 0 for all v ∈ K.
a

Inserting v(x) = max{u(x), 0}.

21. Consider the example ”Positive solutions of eigenvalue equations” of

Section 3.2.8. Show that a non-negative eigenfunction is positive on
(a, b).
Prove the related result for the eigenvalue problem −△u = λu in Ω,
u = 0 on ∂Ω. Here is Ω ∈ Rn a bounded and sufficiently regular
domain.
3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 145

3.3 Sufficient conditions; weak minimizers

3.3.1 Free problems
Consider again the problem minv∈V E(v), where
Z b
E(v) = f (x, v(x), v ′ (x)) dx
a

and V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub }. The next theorem shows

that an extremal u is a strict weak local minimizer if the assumptions of
Theorem 3.2.9 are satisfied. In contrast to the n-dimensional case, the as-
sumption hE ′′ (u)φ, φi > 0 for all φ ∈ (V − V ) \ {0} alone is not sufficient
such that u is a weak local minimizer. A counterexample is, see [53],

f = (x − a)2 y ′2 + (y − a)y ′3 , a < x < b.

The second variation admits the decomposition

hE ′′ (u)φ, φi = a(u)(φ, φ) − b(u)(φ, φ),

where
Z b
a(u)(φ, φ) = Rφ′2 dx,
a
Z b

b(u)(φ, φ) = − 2Qφφ′ + P φ2 dx.
a

If u ∈ C 2 [a, b], then

Z b Z b
′ 2


2Qφφ + P φ dx = Sφ2 dx,
a 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.

Theorem 3.3.1. Suppose that

(i) u ∈ V is a solution of the weak Euler equation,
(ii) hE ′′ (u)φ, φi > 0 for all φ ∈ (V − V ) \ {0},
146 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

(iii) fu′ u′ (x, u(x), u′ (x)) > 0 on [a, b].

Then u is a strict weak local minimizer of E in V .

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

E(u) ≥ E(un ) = E(u + (un − u))

= E(u) + hE ′ (u), un − ui
1
+ hE ′′ (u)(un − u), un − ui + η(||un − u||B )||un − u||2H .
2
Then we can write the above inequality as

0 ≥ a(u)(un − u, un − u) − b(u)(un − u, un − u)
+η(||un − u||B )||un − u||2H .

Set tn = (a(u)(un − u, un − u))−1/2 and wn = tn (un − u). Then

0 ≥ a(u)(wn , wn ) − b(u)(wn , wn )
+η(||un − u||B )||wn ||2H .

Since a(u)(wn , wn ) = 1 it follows for a subsequence wn ⇁ w that b(u)(w, w) ≥

1, in particular w 6= 0, and a(u)(w, w) ≤ 1 since a(u)(v, v) is lower semicon-
tinuous on H. It follows a(u)(w, w) − b(u)(w, w) ≤ 0. Since by assumption
a(u)(v, v) − b(u)(v, v) ≥ 0 for all v ∈ H it follows that hE ′′ (u)w, wi = 0. ✷

There is an interesting relationship between hE ′′ (u)φ, φi > 0 for all φ ∈

(V − V ) \ {0} and an associated eigenvalue problem. Again we suppose that
the strict Legendre condition fu′ u′ (x, u(x), u′ (x)) > 0 on [a, b] is satisfied.
Set H = H01 (a, b) and consider the eigenvalue problem

w ∈ H \ {0} : a(u)(w, ψ) = λb(u)(w, ψ) for all ψ ∈ H. (3.19)

3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 147

Lemma. Suppose that there is a w ∈ H such that b(u)(w, w) > 0. Then

there exists at least one positive eigenvalue of (3.19), and the lowest positive
eigenvalue λ+ 1 is given by

−1 b(u)(v, v)
λ+
1 = max .
v∈H\{0} a(u)(v, v)

Proof. The idea of proof is taken from Beckert [2]. Set

V = {v ∈ H : a(u)(v, v) ≤ 1},

and consider the maximum problem

max b(u)(v, v).

v∈V

There is a solution v1 which satisfies a(u)(v1 , v1 ) ≤ 1. From the assumption

we see that a(u)(v1 , v1 ) = 1. Then

max b(u)(v, v) = max b(u)(v, v),

v∈V v∈V1

where V1 = {v ∈ H : a(u)(v, v) = 1}. The assertion of the lemma follows

since for all v ∈ H \ {0} we have
b(u)(v, v) b(u)(sv, sv)
= ,
a(u)(v, v) a(u)(sv, sv)

where s = (a(u)(v, v))−1/2 . ✷

Theorem 3.3.2. The second variation

hE ′′ (u)φ, φi = a(u)(φ, φ) − b(u)(φ, φ)

is positive for all φ ∈ H \ {0} if and only if there is no positive eigenvalue

of (3.19) or if the lowest positive eigenvalue satisfies λ+
1 > 1.

Proof. (i) Suppose that the second variation is positive, then

b(u)(v, v)
<1
a(u)(v, v)
for all v ∈ H \ {0}. If b(u)(v, v) ≤ 0 for all H, then there is no positive
eigenvalue of (3.19). Assume b(u)(v, v) > 0 for a w ∈ H, then we obtain
from the above lemma that the lowest positive eigenvalue satisfies λ+
1 > 1.
148 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

(ii) Suppose that there is no positive eigenvalue or that the lowest positive
eigenvalue satisfies λ+
1 > 1.

(iia) Consider the subcase that b(u)(v, v) ≤ 0 for all v ∈ H, then

a(u)(v, v) − b(u)(v, v) ≥ 0

for all v ∈ H. It follows that

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.

3.3.2 Equality constraints

Suppose that E : B 7→ R and gi : B 7→ R, i = 1, . . . , m, and for u, h ∈ B
1
E(u + h) = E(u) + hE ′ (u), hi + hE ′′ (u)h, hi + η(||h||B )||h||2H ,
2
1
gi (u + h) = gi (u) + hgi′ (u), hi + hgi′′ (u)h, hi + η(||h||B )||h||2H ,
2
where limt→0 η(t) = 0, hE ′ (u), hi, hgi′ (u), hi are bounded linear functionals
on B which admit a extensions to bounded linear functionals on H. We
3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 149

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

Set for (v, λ) ∈ B × Rm

m
X
L(v, λ) = E(v) + λj gj (v)
j=1
m
X
L′ (v, λ) = E ′ (v) + λj gj′ (v)
j=1
Xm
L′′ (v, λ) = E ′′ (v) + λj gj′′ (v).
j=1

Let
V = {v ∈ B : gi (v) = 0, i = 1, . . . , m}

and assume

hL′′ (u, λ0 )h, hi = a(u, λ0 )(h, h) − b(u, λ0 )(h, h),

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,

(b) b(u, λ0 )(v, h) is a completely continuous form on H × H.

Theorem 3.3.3. Suppose that (u, λ0 ) ∈ V × Rm satisfies L′ (u, λ0 ) = 0 and

a(u, λ0 )(h, h) − b(u, λ0 )(h, h) > 0

for all h ∈ H \ {0} satisfying hgj′ (u), hi = 0 for every j = 1, . . . , m. Then u

is a strict weak local minimizer of E in V .
150 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Proof. The proof is close to the proof of Theorem 3.3.1. If u is no strict

local weak minimizer, then there exists a sequence un ∈ V , ||un − u||B 6= 0,
un → u in B such that

0 ≥ E(un ) − E(u) = L(un , λ0 ) − L(u, λ0 )

1
= hL′ (u, λ0 ), un − ui + hL′′ (u, λ0 )(un − u), un − ui
2
+η(||un − u||B )||un − u||2H
1 1
= a(u, λ0 )(un − u, un − u) − b(u, λ0 )(un − u, un − u)
2 2
+η(||un − u||B )||un − u||2H .

Set
−1/2
tn = a(u, λ0 )(un − u, un − u)
and wn = tn (un − u). Then

0 ≥ 1 − b(u, λ0 )(wn , wn ) + 2η(||un − u||B )||wn ||2H .

Let wn ⇁ w in H for a subsequence, then

0 ≥ 1 − b(u, λ0 )(w, w)

and
a(u, λ0 )(w, w) ≤ 1.
Summarizing, we arrive at

a(u, λ0 )(w, w) − b(u, λ0 )(w, w) ≤ 0

for a w 6= 0 satisfying

hgj′ (u), wi = 0, j = 1, . . . , m.

The previous equations follow from the above expansion of gj (u + h). ✷

There is a related result to Theorem 3.3.2 for constraint problems considered

here. Set
W = {h ∈ H : hgj′ (u), hi = 0, j = 1, . . . , m}
and consider the eigenvalue problem

w ∈ W \ {0} : a(u, λ0 )(w, ψ) = λb(u, λ0 )(w, ψ) for all ψ ∈ W. (3.20)

3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 151

Theorem 3.3.4. Suppose that (u, λ0 ) ∈ V × Rm satisfies L′ (u, λ0 ) = 0.

Then u defines a strict local weak minimizer of E in V if there is no positive
eigenvalue of (3.20) or if the lowest positive eigenvalue satisfies λ+
1 > 1.

Proof. Exercise.

3.3.3 Unilateral constraints

Assume E : B 7→ R. Let V ⊂ B be a nonempty subset and suppose that
u ∈ V is a weak local minimizer of E in V , then

hE ′ (u), wi ≥ 0 for all w ∈ T (V, u),

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

hE ′′ (u)h, hi = a(u)(h, h) − b(u)(h, h),

where a(u)(v, h) and b(u)(v, h) are bounded bilinear symmetric forms on

H × H, a(u)(v, v) is non-negative on H and

(i) (a(u)(v, v))1/2 is equivalent to a norm on H,

(ii) b(u)(v, h) is a completely continuous form on H × H.

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

lim sup t2n hE ′ (u), un − ui < ∞.

n→∞

Corollary. If u ∈ V satisfies the necessary condition hE ′ (u), wi ≥ 0 for all

w ∈ T (V, u), then hE ′ (u), wi = 0 for all w ∈ TE ′ (V, u).
152 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Proof. Exercise.

Theorem 3.3.5. Suppose that u ∈ V satisfies the necessary condition of

Theorem 3.2.1 and that

lim inf t2n hE ′ (u), un − ui ≥ 0

n→∞

for all associated sequences un , tn to w ∈ TE ′ (V, u). Then u is a strict weak

local minimizer of E in V if TE ′ (V, u) = {0} or if

a(u)(w, w) − b(u)(w, w) > 0 for all w ∈ TE ′ (V, u) \ {0}.

Proof. If u is no strict local weak minimizer of E in V , then there exists a

sequence un ∈ V satisfying ||un − u||B 6= 0, un → u in B, such that

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

lim sup t2n hE ′ (u), un − ui < ∞.

n→∞

It follows, if w0 is a weak limit of a subsequence wn′ of wn , that w0 ∈

TE ′ (V, u), and inequality (3.21) yields
1
0 ≥ lim inf hE ′ (u), tn′ wn′ i + [1 − b(u)(w0 , w0 )].
n→∞ 2
Since the first term on the right hand side is non-negative by assumption,
we get
0 ≥ 1 − b(u)(w0 , w0 ),
3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 153

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.

Combining the two previous inequalities, we obtain finally that

a(u)(w0 , w0 ) − b(u)(w0 , w0 ) ≤ 1 − b(u)(w0 , w0 ) ≤ 0,

which is a contradiction to the assumptions of the theorem. ✷

Remark. Assumption

lim inf t2n hE ′ (u), un − ui ≥ 0

n→∞

is satisfied if V is convex since hE ′ (u), v − ui ≥ 0 for all v ∈ V .

Corollary. A u ∈ V satisfying the necessary condition is a strict weak local

minimizer of E in V if
sup b(u)(v, v) < 1,
where the supremum is taken for all v ∈ TE ′ (V, u) satisfying a(u)(v, v) ≤ 1.

Proof. Inequality a(u)(v, v) − b(u)(v, v) > 0 for v ∈ TE ′ (V, u) is equivalent

to 1 − b(u)(v, v) > 0 for v ∈ TE ′ (V, u) satisfying a(u)(v, v) = 1. We recall
that TE ′ (V, u) is a cone with vertex at zero. ✷

It follows immediately

Corollary. Let K be a closed cone with vertex at zero satisfying TE ′ (V, u) ⊂

K. Suppose that u ∈ V satisfies the necessary condition. Then u is a strict
weak local minimizer of E in V if

µ := max b(u)(v, v) < 1,

where the maximum is taken over v ∈ K satisfying a(u)(v, v) ≤ 1.

Remark. If K is convex and if there exists a w ∈ K such that b(u)(w, w) >

0, then µ−1 is the lowest positive eigenvalue of the variational inequality,
see [37],

w∈K: a(u)(w, v − w) ≥ λb(u)(w, v − w) for all v ∈ K.

154 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Example: Stability of a unilateral beam

Consider the example ”A unilateral problem for the beam” of Section 3.2.8,
see Figure 3.13. Set

V = {v ∈ H01 (0, l) ∩ H 2 (0, l) : v(x) ≤ d on (0, l)},

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

where 0 < k ≤ l/2 and λ = (π/k)2 , define solutions of the variational

inequality
u ∈ V : hE ′ (u, λ), v − ui ≥ 0 for all v ∈ V.

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.

Proof. The cone TE ′ (V, u) is a subset of the linear space

L(k) = {v ∈ H01 (0, l) ∩ H 2 (0, l) : v(k) = v(l − k) = 0, v ′ (k) = v ′ (l − k) = 0},

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

w ∈ L(k) : a(w, v) = µb(w, v) for all v ∈ L(k). (3.22)

3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 155

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

Lemma. The cone TE ′ (V, u) is a subset of the linear space

L(k) = {v ∈ H01 (0, l) ∩ H 2 (0, l) : v(k) = v(l − k) = 0, v ′ (l) = v ′ (l − k) = 0}.

Proof. Let w ∈ TE ′ (V, u), that is, there is a sequence un ∈ V , un → u in H

such that tn (un − u) ⇁ w, where tn = ||un − u||−1 H and

lim sup t2n hE ′ (u, λ), un − ui < ∞.

n→∞

We have
wn (k) ≤ 0 and wn (l − k) ≤ 0,
where wn = tn (un − u), and

hE ′ (u, λ), wn i = −A1 wn (k) − A2 wn (l − k), (3.23)

with
A1 = u′′′ (k − 1), A2 = u′′′ (l − k + 0)
156 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

are positive constants. From the definition of TE ′ (V, u) it follows that

hE ′ (u, λ), wn i ≤ c1 t−1

n .

By ci we denote positive constants. Set

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.

Now we prove that

w′ (k) = 0 and w′ (l − k) = 0.

We have u + ǫn wn = un , that is u + ǫn wn ≤ d on [0, l], or

ǫn wn ≤ d − u on [0, l]. (3.25)

Since u ∈ H 3 (0, l), which follows directly by calculation or from a general

regularity result due to Frehse [17], and wn ∈ H 2 (0, l), we have the Taylor
expansions3

 
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

since u′ (k) = u′′ (k) = 0. Then we obtain from (3.25) that

 
ǫn wn (k) + wn′ (k)h − c2 |h|3/2 ≤ c3 |h|5/2 . (3.26)

We consider two cases.

Case 1. Suppose that wn (k) = 0 for a subsequence, then we get from (3.26)
that wn′ (k) = 0 for this sequence, which implies w′ (k) = 0.
Case 2. Suppose that wn (k) < 0 for all n > n0 . Set

αn = −wn (k), βn = |wn′ (k)|.

From above we have that αn → 0 as n → ∞. Assume βn ≥ β with a positive

constant β. Set
2αn
h= sign(wn′ (k)),
β
then we obtain from inequality (3.26) that
 3/2 !  5/2
βn 2αn 2αn
ǫn −αn + 2αn − c2 ≤ c3 ,
β β β

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.

2. Set B = C 1 [a, b], H = H 1 (a, b) and

Z b
E(v) = f (x, v(x), v ′ (x)) dx,
a

where v ∈ B, and f is assumed to be sufficiently regular.

Show, if u, h ∈ B, then

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,

and limt→0 η(t) = 0.

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

where u, h ∈ C 1 [a, b]. Show that (a(u)(h, h))1/2 is equivalent to a

norm on H01 (a, b), provided that the strict Legendre condition

fu′ u′ (x, u(x), u′ (x)) > 0

is satisfied on [a, b].

3.3. SUFFICIENT CONDITIONS; WEAK MINIMIZERS 159

Hint: Show that there exists a positive constant c such that

Z b Z b
′ 2
h (x) dx ≥ c h(x)2 dx
a a

for all h ∈ C01 (a, b).

4. Show that the bilinear form defined on H × H, where H = H 1 (a, b),

Z b
b(u)(φ, ψ) = − [fuu′ (φψ ′ + φ′ ψ) + fuu φψ] dx,
a

where fuu′ , fuu ∈ C[a, b], is completely continuous on H × H, that is,

lim b(u)(φn , ψl ) = b(u)(φ, ψ)

n,l→∞

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.

5. Prove Theorem 3.3.4.

160 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

3.4 Sufficient condition; strong minimizers

The following consideration concerns a class of free problems. There are
related results for isoperimetric problems, see [6, 52, 19], for instance. Set

V = {v ∈ C 1 [a, b] : v(a) = ua , v(b) = ub }

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.

Let u ∈ V be a weak solution of the Euler equation, that is

u ∈ V : hE ′ (u), φi = 0 for all φ ∈ V − V.

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

hE ′′ (u)φ, φi > 0 for all φ ∈ (V − V ) \ {0},

then u is a weak local minimizer of E in V , see Theorem 3.3.1. If additionally

fpp (x, y, p) ≥ 0 for all (x, y) ∈ Dδ (u) and p ∈ R, where for a δ > 0

Dδ (u) = {(x, y) : a ≤ x ≤ b, u(x) − δ ≤ y ≤ u(x) + δ},

then we will prove that u is a strong local minimizer of E in V .

Definition. A set D ⊂ R2 is said to be simply covered by a family of curves

defined by y = g(x, c), c ∈ (c1 , c2 ), if each point (x, y) ∈ D is contained in
exactly one of these curves. Such a family is called a foliation of D. If a given
curve defined by y = u(x) is contained in this family, that is u(x) = g(x, c0 ),
c1 < c0 < c1 , then u is called embedded in this family.

Lemma 3.4.1 (Existence of an embedding family). Let u ∈ V ∩ C 2 [a, b] be

a solution of the Euler equation. Suppose that the strict Legendre condition
fpp (x, u(x), u′ (x)) > 0 is satisfied on [a, b] and that the second variation
3.4. SUFFICIENT CONDITION; STRONG MINIMIZERS 161

hE ′′ (u)φ, φi is positive on (V − V ) \ {0}. Then there exists a foliation v(µ),

|µ| < ǫ, of Dδ (u), provided δ and ǫ are sufficiently small. Every element of
this foliation solves the Euler equation, and v(0) = u.

Proof. Consider the family of boundary value problems

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)

where µ is a constant, µ ∈ (−ǫ, ǫ), ǫ > 0. Define the mapping

M (v, µ) : C 2 [a, b] × (−ǫ, ǫ) 7→ C[a, b] × R × R

by
 d

− dx fv′ (x, v, v ′ ) + fv (x, v, v ′ )
M (v, µ) =  v(a) − ua − µ .
v(b) − ub − µ

We seek a solution v(µ) of M (v, µ) = 0. Since M (v(0), 0) = 0, where

v(0) = u, and Mv (u, 0) defined by
 d

− dx (Rh′ )′ + Sh
Mv (u, 0)h =  h(a) 
h(b)

is a regular mapping from C 2 [a, b] →

7 C[a, b] × R × R. It follows from an
implicit function theorem, see [28], for instance, that there exists a unique
solution
v(x, µ) = u + µv1 (x) + r(x, µ) (3.31)

of (3.28), (3.29), (3.29), where r ∈ C 1 ([a, b] × (−ǫ, ǫ)), r(x, µ) = o(µ),

limµ→0 rµ (x, µ) = 0, uniformly on [a, b], and v1 is the solution of the Ja-
cobi boundary value problem.

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)).

Lemma 3.4.2 (Hilbert’s invariant integral). Suppose that Cv is a curve in

Dδ (u) defined by y = v(x), v ∈ C 1 [a, b], v(a) = ua and v(b) = ub . Then the
integral Z
U (v) = P (x, y)dx + Q(x, y)dy,
Cv
is independent of v.

Proof. We show that Py = Qx in Ω. This follows by a straightforward

calculation. We have
Py = fy + fp Φy − Φy fp − Φ(fpy + fpp Φy )
Qx = fpx + fpp Φx .
3.4. SUFFICIENT CONDITION; STRONG MINIMIZERS 163

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

y ′ (x) = Φ(x, y(x))

y(x0 ) = y0

We recall that

Φ(x, y) = F (x, c(x, y))

≡ F (x, c(x, y(x)))
≡ Φ(x, y(x)),

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,

fy − fpx − fpy y ′ − fpp y ′′ = 0,

we obtain that Py − Qx = 0 at (x0 , y0 ) ∈ Dδ (u). ✷

On the weak local minimizer u in consideration we have

E(u) = U (u) (3.32)

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

We recall that u′ (x) = Φ(x, u(x)).

164 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Definition. The function

E(x, y, p, q) = f (x, y, q) − f (x, y, p) + (p − q)fp (x, y, p)
is called Weierstrass excess function.

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.

Theorem 3.4.1. Let u ∈ V be a solution of the weak Euler equation, that

is of hE ′ (u), φi = 0 for all φ ∈ V − V . Suppose that
(i) fpp (x, u(x), u′ (x)) > 0 on [a, b],
(ii) hE ′′ (u)φ, φi > 0 for all φ ∈ (V − V ) \ {0},
(iii) E(x, y, p, q) ≥ 0 for all (x, y) ∈ Dδ (u) and for all p, q ∈ R.
Then u is a strong local minimizer of E in V .

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.
✷

EXAMPLE: Set V = {v ∈ C 1 [0, l] : v(0) = 0, v(l) = 1} and for v ∈ V

Z l
 ′ 
E(v) = (v (x))2 − (v(x))2 dx.
0

The solution of the Euler equation is u = sin x/ sin l, provided l 6= kπ,

k = ±1, ±2, . . . . Assumption (iii) of the above theorem is satisfied since
Rl
E(x, y, p, q) = (p − q)2 , and assumption (ii), that is 0 (φ′2 − φ2 ) dx > 0 for
all φ ∈ (V − V ) \ {0} holds if the lowest eigenvalue λ1 = (π/l)2 of −φ′′ = λφ
on (0, l), φ(0) = φ(l) = 0 satisfies λ1 > 1, see Theorem 3.3.2. Thus u is a
strong local minimizer of E in V if 0 < l < π.
In fact, u is a strong global minimizer of E in V if 0 < l < π, see an exercise.
3.4. SUFFICIENT CONDITION; STRONG MINIMIZERS 165

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 ,

where g ∈ C[a, b] and va , vb ∈ R are given.

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

3.5 Optimal control

For a given function v(t) ∈ U ⊂ Rm , t0 ≤ t ≤ t1 , we consider the boundary
value problem

y ′ (t) = f (t, y(t), v(t)), y(t0 ) = x0 , y(t1 ) = x1 , (3.33)

where y ∈ Rn , x0 and x1 are given, and

f : [t0 , t1 ] × Rn × Rm 7→ Rn .

In general, there is no solution of such a problem. Therefore we consider

the set of admissible controls Uad defined by the set of piecewise continu-
ous functions v on [t0 , t1 ] such that there exists a continuous and piecewise
continuously differentiable solution of the boundary value problem. Such a
solution is continuously differentiable at all regular points of the control v.
A point t ∈ (t0 , t1 ) where v(t) is continuous is called a regular point of the
control v. We suppose that this set is not empty. Assume a cost functional
is given by
Z t1
E(v) = f 0 (t, y(t)), v(t)) dt,
t0

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.

The functions f, f 0 are assumed to be continuous in (t, y, v) and contin-

uously differentiable in (t, y). It is not required that these functions are
differentiable with respect to v.

Then the problem of optimal control is

max E(v). (3.34)

v∈Uad

A piecewise continuous solution u is called optimal control, and the contin-

uous and piecewise continuously differentiable solution x of the associated
system of boundary value problems is said to be an optimal trajectory.
3.5. OPTIMAL CONTROL 167

3.5.1 Pontryagin’s maximum principle

The governing necessary condition for this type of problems is the following
maximum principle, see Pontryagin et al. [48]. Set

H(t, y, v, p0 , p) = p0 f 0 (t, y, v) + hp, f (t, y, v)i,

where p0 ∈ R and p ∈ Rn . This function H is called Hamilton function

associated to the above optimal control problem.

Theorem 3.5.1 (Pontryagin’s maximum principle). Let u(t) be a piecewise

continuous solution of the maximum problem, and let x(t) be the associated
continuous and piecewise continuously differentiable trajectory. Then there
exists a constant p0 and a continuous and piecewise continuously differen-
tiable vector function p(t), not both are zero, such that

(i) p(t) is a solution of the linear system

p′ (t) = −Hx (t, x(t), u(t), p0 , p(t)), (3.35)

in all regular points.

(ii) In all regular points t of the optimal control u(t) we have

H(t, x(t), u(t), p0 , p(t)) ≥ H(t, x(t), v, p0 , p(t)) for all v ∈ U.

(iii) p0 = 1 or p0 = 0.

Definition. The vector function p(t) is called adjoint function.

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

Example: Consumption versus investment

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

y ′ (t) = v(t), y(0) = x0 , y(T ) ≥ x1 .

We suppose additionally that

x0 < x1 < x0 + T. (3.36)

For the utility function U (s), 0 ≤ s < ∞, we assume U ∈ C 2 , U ′ > 0 and

U ′′ < 0.
Economic interpretation: x(t) level of infrastructure at time t,
u(t) level of investment in infrastructure at t,
1 − u(t) level of consumption at t,
[0, T ] planning period.

The Hamilton function is here

H = p0 U (1 − v) + pv,

then equation (3.35) is given by

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

H = p0 U (1 − u(t)) + cu(t) ≥ p0 U (1 − v) + cv for all v ∈ [0, 1]. (3.37)

We recall that the non-negative constants p0 and c are not simultaneously

zero. Since Huu = p0 U ′′ (1 − u), where p0 ≥ 0 and U ′′ < 0, we find from the
3.5. OPTIMAL CONTROL 169

maximum property (3.37) three cases:

(i) u = 0 is a maximizer of H, then Hu ≤ 0 at u = 0,
(ii) u = a, where a, 0 < a < 1, is a constant maximizes H, then Hu = 0 at
u = a,
(iii) u = 1 is a maximizer of H, then Hu ≥ 0 at u = 1.
Since Hu = −p0 U ′ (1 − u) + c, we have for regular t: if

u(t) = 0, then c ≤ p0 U ′ (1), (3.38)

′
0 < u(t) < 1, then p0 U (1 − u(t)) = c, (3.39)
u(t) = 1, then p0 U ′ (0) ≤ c. (3.40)

We show that p0 = 1. If not, then p0 = 0. Then u = 1 is a maximizer

of H for all t. It follows from the differential equation x′ (t) = u(t) ≡ 1
that x(t) = t + x0 , thus x(T ) = T + x0 . The assumption (3.36) implies
that the optimal trajectory satisfies x(T ) > x1 . This case is covered by
Section 3.5.3 (free endpoint) below. In this case we have c = p(T ) = 0.
Since p0 = 1, the Hamilton function H = U (1 − v) + cv is strictly concave,
which implies that there exists a unique maximizer of H in [0, 1] which does
not depend of t since H is independently of t. Then the optimal control is
u(t) = u∗ = const., u∗ ∈ [0, 1].
We have u∗ > 0. If not, then we get from x′ (t) = u(t), x(0) = x0 , that
x(T ) = x0 , a contradiction to the assumption (3.36).
The inequality u∗ > 0 implies that c > 0, see (3.38)-(3.40). Then x(T ) = x1 .
If x(T ) > x1 , then p(T ) = c = 0, see the remark above.
If u∗ = 1, then there is no consumption, which contradicts the side con-
dition (3.36) since in this case x(t) = t + x0 . Thus x(T ) = T + x0 , a
contradiction to x(T ) ≥ x1 and x1 < T + x0 .
We get finally that u∗ ∈ (0, 1), which implies that x(t) = x0 + u∗ t. Thus we
have
x1 − x0
u(t) = u∗ =
T
since x(T ) = x1 , and the associated optimal trajectory is given by
x1 − x0
x(t) = x0 + t.
T

Example: Workers versus capitalists

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

[0, 1], and two cost functionals W and C given by

Z T
W (u, v) = u(t)K(t) dt,
0
Z T
C(u, v) = (1 − v(t))(1 − u(t))K(t) dt.
0
and the differential equation and the boundary conditions which define the
trajectory K(t) : [0, T ] 7→ R are
K ′ (t) = v(t)(1 − u(t))K(t), K(0) = K0 > 0, K(T ) free.
That is, no assumption on the final state is posed. We suppose additionally
that
1 1
T > and T > .
b 1−b

Economic interpretation: K(t) capital stock of the firm. Rate of production

is proportional to K,
u(t)K(t) share of profit of the workers,
(1 − u(t))K(t) profit of the firm,
v(t)(1 − u(t))K(t) is investment of the company and the rest (1 − v(t))(1 −
u(t))K(t) remains for consumption of the firm.

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.

A discussion similar to the previous example leads to the following result,

exercise or [55], pp 224,
Case b ≥ 1/2. Set t′ = T − 1/(1 − b), then
u∗ (t) = a, v ∗ (t) = 1 if t ∈ [0, t′ ],
u∗ (t) = b, v ∗ (t) = 0 if t ∈ (t′ , T ].
3.5. OPTIMAL CONTROL 171

Case b < 1/2. Set

 
′ 1 1 1
t =T− , t′′ = t′ − ln ,
1−b 1−b 2b

then

u∗ (t) = a if t ∈ [0, t′′ ], u∗ (t) = b if t ∈ (t′′ , T ],

v ∗ (t) = 1 if t ∈ [0, t′ ], v ∗ (t) = 0 if t ∈ (t′ , T ].
172 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

3.5.3 Proof of Pontryagin’s maximum principle; free end-

point
Here we prove Theorem 3.5.1, where p0 := 1, and the boundary conditions
in (3.33) are replaced by y(t0 ) = x0 , no condition at t1 . The proof is close
to a proof given in [24].
Let u(t) be an optimal control and let τ ∈ (t0 , t1 ) be a regular point. We
define a new admissible control uǫ (t), see Figure 3.14 by a needle variation

u(t) : t 6∈ [τ − ǫ, τ ]
uǫ (t) = ,
v : t ∈ [τ − ǫ, τ ]

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

Figure 3.14: Needle variation

xε (t)

x(t)

t0 τ−ε τ t1 t

Figure 3.15: Variation of the trajectory

3.5. OPTIMAL CONTROL 173

the perturbed trajectory associated to uǫ (t) is defined by


x(t) : t0 < t < τ − ǫ
xǫ (t) = ,
x∗ (t) : τ − ǫ < t < t1

where x∗ denotes the solution of the initial value problem

y ′ (t) = f (t, y(t), uǫ (t)) τ − ǫ < t < t1 ,

y(τ − ǫ) = x(τ − ǫ).

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

xǫ (τ ) = x(τ ) + ǫ[x′ǫ (τ ) − x′ (τ )] + o(ǫ)

= x(τ ) + ǫ[f (τ, xǫ (τ ), v) − f (τ, x(τ ), u(τ ))] + o(ǫ)
= x(τ ) + ǫ[f (τ, x(τ ), v) − f (τ, x(τ ), u(τ ))] + o(ǫ).

The previous equation is a consequence of xǫ (τ ) = x(τ ) + O(ǫ). We recall

that τ is a regular point. Set

w(τ, v) = f (τ, x(τ ), v) − f (τ, x(τ ), u(τ )),

then the changed trajectory xǫ (t) on τ < t < t1 is given by, see an exercise,

xǫ (t) = x(t) + ǫΘ(t, τ )w(τ, v) + o(ǫ),

where Θ(t, τ ) is the fundamental matrix associated to the linear system

w′ (t) = fx (t, x(t), u(t)) w(t),

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

3.5.4 Proof of Pontryagin’s maximum principle; fixed end-

point
Here we will prove Theorem 3.5.1, see [48], pp. 84. The following proof is
close to [48] and to [3], where an important part of the proof (see Case (ii)
below) is sketched. See also [59] for a sketch of the proof. We will give an
easy proof of Pontryagin’s maximum principle. See also [21], pp. 75, for a
proof based on Brouwer’s fixed point theorem.
The idea is to use more than one needle variation of the given optimal
control in order to achieve the fixed target x1 with a perturbed trajectory. At
first, we transform problem (3.34) into an autonomous problem of Mayer’s
type, i. e., we maximize a single coordinate of a new trajectory vector.
Define the first new coordinate by the initial value problem
′
yn+1 (t) = 1, yn+1 (t0 ) = t0 ,

which implies that yn+1 (t) = t. The second new coordinate is defined by

y0′ (t) = f 0 (yn+1 (t), y(t), v(t)), y0 (t0 ) = 0.

Then the maximum problem (3.34) is equivalent to

max y0 (t1 , v), (3.42)

v∈Uad

where y0 (t) ≡ y0 (t, v). Set

Y = (y0 , y1 , . . . , yn , yn+1 ), F = (f 0 , f 1 , . . . , f n , 1),

then the new trajectory satisfies the differential equation

Y ′ = F (y, yn+1 , v) (3.43)

and the boundary and initial conditions are

y0 (t0 ) = 0, yn+1 (t0 ) = t0 , y(t0 ) = x0 , y(t1 ) = x1 , (3.44)

where y = (y1 , . . . , yn ).

Let
P = (p0 , p, pn+1 ), p = (p1 , . . . , pn ).
Define the Hamilton function by

H(y, yn+1 , v, P ) = hP, F (y, yn+1 , v)i.

176 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Suppose that u is an optimal control, that is, a solution of problem (3.42),

and X is the associated trajectory satisfying

X ′ = F (x, xn+1 , u) ≡ HP (x, xn+1 , u), t0 < t < t1 ,

where x = (x1 , . . . , xn ) and X = (x0 , x, xn+1 ).

We will show that there is a continuous and piecewise continuously differ-

entiable vector function P (t) 6= 0 which solves

P ′ (t) = −HX (x, xn+1 , u, P ), t0 < t < t1 , (3.45)

and at all regular points t ∈ (t0 , t1 )

H(x(t), xn+1 (t), u(t), P (t)) ≥ H(x(t), xn+1 (t), v, P (t)) (3.46)

for all v ∈ U , and

p0 (t) = const. ≥ 0. (3.47)

We consider a finite set of needle variations at regular points τi of the given

optimal control u(t) defined by replacing u(t) by a constant vi ∈ U on the
intervals [τi − ǫai , τi ], where t0 < τi < t1 , τi different from each other, ai > 0
are fixed and ǫ > 0 is sufficiently small, see Figure 3.16. Consider the linear

v1

v2

t0 τ 1−ε a 1 τ1 τ 2 −εa 2 τ2 t1

Figure 3.16: Needle variations

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

The matrix A(t) is piecewise continuous on (t0 , t1 ). As in the previous proof

we see that the perturbed trajectory is given by
s
X
Xǫ (t) = X(t) + ǫ ai Θ(t, τi )W (τi , vi ) + o(ǫ), (3.49)
i=1

where Θ(t, τ ) is the fundamental matrix to the system (3.48) and

W (τi , vi ) := F (X(τi ), vi ) − F (X(τi ), u(τi )).

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

hP1 , yi ≤ 0 for all y ∈ C and hP1 , re0 i ≥ 0. (3.50)

178 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Let Π be the plane defined by, see Figure 3.17,

Π = {z ∈ Rn+2 : hP1 , zi = 0}.

Consider the initial value problem

e0 Π
P1

Figure 3.17: Separation of C and L = re0

P ′ (t) = −FX (x(t), xn+1 , u(t)) P (t), P (t1 ) = P1 .

Let Ψ(t, τ ) be the fundamental matrix of this system, then

P (t) = Ψ(t, t0 ) P (t0 ), P (t0 ) = Ψ−1 (t1 , t0 ) P (t1 ).

Let t ∈ (t0 , t1 ) be a regular point of the optimal control u(t). Set

W (t, v) = F (x(t), xn+1 (t), v) − F (x(t), xn+1 (t), u(t)), (3.51)

where v ∈ U . Then
ǫΘ(t1 , t)W (t, v) ∈ C,
where ǫ > 0. Then, see (3.50),

hP (t1 ), Θ(t1 , t)W (t, v)i ≤ 0.

Since ΨT (t, τ )Θ(t, τ ) = I, where I is the identity matrix, see an exercise,

and from P (t1 ) = Ψ(t1 , t)P (t) we obtain

hP (t), W (t, v)i ≤ 0.

Consequently, see the definition (3.51) of W (t, v),

H(x(t), xn+1 (t), u(t), P (t)) ≥ H(x(t), xn+1 (t), v, P (t))

3.5. OPTIMAL CONTROL 179

for all v ∈ U and for all regular points t ∈ (t0 , t1 ).

Finally, from Hx0 = 0 it follows that p0 (t) = const., and from the second
inequality of (3.50) we get that p0 ≥ 0. Thus we can assume that p0 = 0 or
p0 = 1. This follows since we can replace P1 by P1 /p0 in the considerations
above if p0 > 0.

Case (ii). Suppose that L is in the interior of C. Then z 0 = r0 e0 , r0 > 0,

is in the interior of C, and there are n + 2 linearly independent Ai ∈ C such
that z 0 is an interior point of the closed simplex S defined by
n+2
X n+2
X
λi Ai , λi ≥ 0, λi ≤ 1,
i=1 i=1

see Figure 3.18. Let λi (z) are the (uniquely determined) barycentric coor-
L

3
A2
A

z0 . A1

Figure 3.18: Simplex in consideration

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

We recall that λi (z) are continuous

Pn+2in z, and z is in the interior of S if and
only if λi (z) > 0 for every i and i=1 λi (z) < 1. From the definition of C
we see that
si
X
i
A = ail Θ(t1 , τli )W (τli , vli ).
l=1
Consider needle variations of the optimal control at τli with associated λi ail ,
vli , where τli , ail and vli are fixed, and
n+2
X
λi ≤ 1, λi > 0.
i=1

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

where bi (λ, ǫ) are continuous in λ, for each fixed ǫ, 0 < ǫ ≤ ǫ0 , ǫ0 sufficiently

small, and bi (λ, ǫ) → 0 as ǫ → 0, uniformly on {λ ∈ Rn+2 : 0 ≤ λi ≤ 1, i =
1, . . . , n + 2}.
Let z be in the interior of S and let λi (z) are the associated barycentric
coordinates. Set
n+2
X
q(z, ǫ) = [λi (z) + bi (λ(z), ǫ)]Ai
i=1
n+2
X
≡ z+ bi (λ(z), ǫ)Ai ,
i=1

and consider for fixed ǫ, 0 < ǫ ≤ ǫ0 , the mapping T : S 7→ Rn+2 , defined by

T (z, ǫ) := z − q(z, ǫ) + z 0 .
This mapping is continuous in z and maps the closed ball Bρ (z 0 ) ⊂ Rn+2 ,
which is in the interior of S, into this ball, provided that ǫ0 is sufficiently
small. Brouwer’s fixed point theorem, see [8, 30, 22], says that there is a
z ∗ ∈ Bρ (z 0 ), z ∗ = z ∗ (ǫ), such that T (z ∗ , ǫ) = z ∗ . Set λi = λi (z ∗ ) in the
needle variation above, we get finally
Xǫ (t1 ) = X(t1 ) + ǫr0 e0 .
✷
3.5. OPTIMAL CONTROL 181

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

y1′ (t) = av(t)y1 (t), y1 (0) = y10

y2′ (t) = a(1 − v(t))y1 (t) y2 (0) = y20 ,

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

b is a positive constant, v(t) ∈ V piecewise continuous, and the dif-

ferential equations and the side conditions which define the trajectory
y : [0, T ] 7→ R2 are

y1′ (t) = v(t)y1 (t), y1 (0) = y10 , y1 (T ) free,

y2′ (t) = y1 (t), y2 (0) = y20 , y2 (T ) ≤ y2T ,

where the data y10 and y20 and y2T are given.

3. Consider a model for ”Consumption versus investment”, see [55], pp.

113. Suppose that the admissible set of controls is V = [0, 1], the cost
functional is given by
Z T  
E(v) = 1 − e−(1−v(t))y(t) dt,
0

where v(t) ∈ V is piecewise continuous, and the differential equation

and the side conditions which defines the trajectory y : [0, T ] 7→ R are

y ′ (t) = v(t) y(t), y(0) = y 0 > 0, y(T ) free.

182 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

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,

z ′ (t) = f (t, z(t), u(t)), τ < t < t1

z(τ ) = x(τ ) + ǫw(τ, v) + o(ǫ)

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.

6. Let x : [t0 , t1 ] 7→ R be a C 1 [t0 , t1 ]-solution of the initial value problem

x′ (t) = f (t, x(t)) in (t0 , t1 ),

x(τ ) = a,

where a is given and f is sufficiently regular with respect to (t, x).

Show that there exists a solution y(t) of

y ′ (t) = f (t, y(t)) in (t0 , t1 ),

y(τ ) = a + µ,

where µ ∈ (−µ0 , µ0 ), µ > 0 sufficiently small.

Hint: Consider the mapping

M (y, µ) : C 1 [t0 , t1 ] × R 7→ C[t0 , t1 ] × R

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].

7. Let K ⊂ Rn be a nonempty convex cone with vertex at the origin and

assume x 6∈ cl K. Then there is a p ∈ Rn \ {0} such that hp, xi > 0
and hp, yi ≤ 0 for all y ∈ cl K.
Hint: Apply Theorem 2.6.1.
3.5. OPTIMAL CONTROL 183

8. Let K ⊂ Rn be a nonempty convex cone with vertex at the origin, and

assume x ∈ ∂K. Then there is a p ∈ Rn \ {0} such that hp, xi ≥ 0 and
hp, yi ≤ 0 for all y ∈ K.
Hint: Theorem 2.6.2 says that there are p ∈ Rn \ {0} and α ∈ R such
that hp, xi = α and hp, yi ≤ α for all y ∈ cl K.

9. Show that the set C defined in Section 3.5.4 is a convex cone with
vertex at the origin.

10. Let A(t) be a continuous N × N -matrix, t0 < t < t1 . Consider

the initial value problems w′ (t) = A(t)w(t), w(τ ) = w0 and v ′ (t) =
−AT (t)v(t), v(τ ) = v0 , where τ ∈ (t0 , t1 ). Denote by Θ(t, τ ) and
Ψ(t, τ ) the associated fundamental matrix, respectively. Show that
ψ T (t, τ )Θ(t, τ ) = I, where I denotes the identity matrix.
Hint: If t = τ , then Θ(τ, τ ) = Ψ(τ, τ ) = I. Let ξ ∈ RN , set y(t) =
ΨT (t, τ )Θ(t, τ )ξ and show that y ′ (t) = 0. Thus ξ = ΨT (t, τ )Θ(t, τ )ξ
for all ξ ∈ RN .

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

[1] R. A. Adams, Sobolev Spaces. Academic Press, New York, 1975.

[2] H. Beckert, Über eine bemerkenswerte Klasse gemischtfreier Variation-

sprobleme höherer Ordnung. Math. Zeitschr. 103 (1968), 18–29.

[3] B. Beavis and I. M. Dobbs, Optimization and Stability Theory for Eco-
nomic Analysis. Cambridge University Press 1990.

[4] S. Bernstein, Sur un théorème de géométrie et son application aux

dérivées partielles du type elliptique. Comm. Soc. Math. de Kharkov
(2)15,(1915–1917), 38–45. German translation: Math. Z. 26 (1927),
551–558.

[5] W. Blaschke, Vorlesungen über Differentialgeometrie I. Springer,

Grundlehren 1 (1921).

[6] O. Bolza, Vorlesungen über Variationsrechnung. Koehler und Amelang,

Leipzig 1949.

[7] E. Bombieri, E. De Giorgi and E. Giusti, Minimal cones and the Bern-
stein problem. Inv. Math. 7 (1969), 243–268.

[8] L. E. J. Brouwer, Über Abbildung von Mannigfaltigkeiten. Math. Ann.

71 (1912), 97–115.

[9] R. Courant und D. Hilbert, Methoden der Mathematischen Physik.

Band 1 und Band 2. Springer-Verlag, Berlin, 1968. English translation:
Methods of Mathematical Physics. Vol. 1 and Vol. 2, Wiley-Interscience,
1962.

[10] A. Cournot, Researches into the Mathematical Principles of the Theory

of Wealth. Translation from French, New York, Kelley, 1960.

185
186 BIBLIOGRAPHY

[11] G. Debreu, Existence of competitive equilibrium. Handbook of Math-

ematical Economics, Vol. II North-Holland Publishing Company, pp.
697-743 (1982).

[12] U. Dierkes, S. Hildebrandt, A. Küster and O. Wohlrab, Minimal Sur-

faces. I, II. Springer, Grundlehren 295, 296 (1992).

[13] U. Dierkes, S. Hildebrandt and F. Sauvigny, Minimal Surfaces.

Springer, Grundlehren 339 (2010).

[14] L. C. Evans, Partial Differential Equations. Graduate Studies in Math-

ematics , Vol. 19, AMS, Providence, 1991.

[15] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of

Functions. Studies in Advanced Mathematics, CRC Press, Boca Raton,
1992.

[16] R. Finn, Equilibrium Capillary Surfaces. Springer, Grundlehren 284

(1986).

[17] J. Frehse, Zum Differenzierbarkeitsproblem bei Variationsungleichun-

gen höherer Ordnung. Abhandl. Math. Seminar Hamburg 36 (1971),
140–149.

[18] J. W. Friedman, Game Theory with Applications to Economics. Oxford

University Press, New York, 1986.

[19] M. Giaquinta and S. Hildebrandt, Calculus of Variations. I, II, Springer,

Grundlehren 310, 311 (2004).

[20] D. Gilbarg and N. S. Trudinger, Elliptic Partial Differential Equations

of Second Order. Springer, Grundlehren 224 (1983).

[21] H. Halkin, On the necessary condition for optimal control of nonlinear

systems. J. Analyse Math. 12 (1964),1–82.

[22] H. Heuser, Lehrbuch der Analysis II. Teubner-Verlag, Stuttgart, 1993.

[23] M. R. Hestenes, Calculus of Variations and Optimal Control Theory.

John Wiley & Sons, Inc. New York, 1966.

[24] A. D. Ioffe und V. M. Tichomirov, Theorie der Extremalaufgaben. VEB

Deutscher Verlag der Wissenschaften, Berlin, 1979. Translation from
Russian.
BIBLIOGRAPHY 187

[25] J. Israelachvili, Intermolecular and Surface Forces. Academic Press,

London, 1995.

[26] J. Jost and X. Li-Jost, Calculus of Variations. Cambridge University

Press, 1998.

[27] S. Kakutani, A generalisation of Brouwer’s fixed point theorem. Duke

Math. J. 8 (1941), 457–459.

[28] L. W. Kantorowitsch and G. P. Akilow, Funktionalanalysis in

normierten Räumen. Translation from Russian. Akademie-Verlag,
Berlin, 1964.

[29] D. Kinderlehrer and G. Stampacchia, An Introduction to Variational

Inequalities and Their Applications. Academic Press, New York, 1980.

[30] B. Knaster, K. Kuratowski und S. Mazurkiewicz, Ein Beweis des Fix-

punktsatzes für n-dimensionale Simplexe. Fund Mat. 14 (1929), 132–
137.

[31] H. W. Kuhn and A. W. Tucker, Nonlinear programming. Proc. of the

Second Berkeley Symp. on Math. Stat. and Prob., Berkeley, Calif., 1951,
481–492.

[32] H. Link, Über den geraden Knickstab mit begrenzter Durchbiegung.

Ing.-Arch. 22 (1954), 237–250.

[33] L. D. Landau and E. M. Lifschitz, Lehrbuch der Theoretischen Physik.

Vol. 1., Akademie-Verlag, Berlin, 1964. German translation from Rus-
sian. English translation: Course of Theoretical Physics. Vol. 1, Perga-
mon Press, Oxford, 1976.

[34] J. L. Lions, Optimal Control Governed by Partial Differential Equa-

tions. Translation from French, Springer, Berlin, 1971.

[35] J.-L. Lions and E. Magenes, Problémes aux limites non homogénes et
applications. Dunod, Paris, 1968.

[36] D. G. Luenberger, Microeconomic Theory. MacGraw-Hill Company,

New York, 1995.

[37] E. Miersemann, Verzweigungsprobleme für Variationsungleichungen.

Math. Nachr. 65 (1975), 187 – 209.
188 BIBLIOGRAPHY

[38] E. Miersemann, Über nichtlineare Eigenwertaufgaben in konvexen Men-

gen. Math. Nachr. 88 (1979), 191 – 205.

[39] E. Miersemann, Über positive Lösungen von Eigenwertgleichungen mit

Anwendungen auf elliptische Gleichungen zweiter Ordnung und auf ein
Beulproblem der Platte. ZAMM 59 (1979), 189–194.

[40] E. Miersemann, Zur Lösungsverzweigung bei Variationsungleichungen

mit einer Anwendung auf den geraden Knickstab mit begrenzter Durch-
biegung. Math. Nachr. 102 (1981), 7 – 15.

[41] E. Miersemann, Wetting barriers. Calc. Var. 15 (2002), 237 – 256.

[42] E. Miersemann, Kapillarflächen. Ber. Verh. Sächs. Akad. Wiss. Leipzig,

Math.-Natur. Kl. 130 (2008), Heft 4, S. Hirzel, Leipzig, 2008.

[43] E. Miersemann and H. D. Mittelmann, A free boundary problem and

stability for the circular plate. Math. Meth. in the Appl. Sci. 9 (1987),
240–250.

[44] E. Miersemann and H. D. Mittelmann, A free boundary problem and

stability for the rectangular plate. Math. Meth. in the Appl. Sci. 12
(1990), 129–138.

[45] J. J. Moreau, Décomposition orthogonale dans un espace HILBERTien

selon deux cônes mutuellement polaires. C. R. Acad. Sci., Paris 255
(1962), 238–240.

[46] J. F. Nash, Equilibrium points in N-person games. Proceedings Nat.

Acad. Sciences, U.S.A., 36 (1950), 48-49.

[47] R. Osserman, A Survey of Minimal Surfaces. Dover Publications

(1986).

[48] L. S. Pontryagin, V. G. Boltjanskij, R. V. Gamkrelidze und E. F.

Miščenko, The Mathematical Theory of Optimal Processes. 2nd Edi-
tion, Moscow, 1969 (in Russian). Translations into German (1964) and
English (1962), for example, exist from the first edition.

[49] R. T. Rockafellar and R. J.-B. Wets, Variational Analysis. Springer,

Grundlehren 317 (1998)

[50] R. T. Rockafellar, Convex Analysis. Princeton University Press, 1970.

BIBLIOGRAPHY 189

[51] M. Roscher, Fortsetzung lokaler Nash-Gleichgewichte. Diplomarbeit.

Department of Mathematics, Leipzig University, 1999.

[52] H. Sagan, Introduction to the Calculus of Variations. Dover, New York,

1992.

[53] L. Scheeffer, Über die Bedeutung der Begriffe ”Maximum und Mini-
mum” in der Variationsrechnung. Math. Annalen 26 (1886), 197–208.

[54] P. Schiller, H.-J. Mögel and E. Miersemann, Stability of liquid layers

on solid materials. J. math. fluid mech. 12 (2) (2010), 293–305.

[55] A. Seierstad and K. Sydsæter, Optimal Control Theory with Economic

Applications. North-Holland, Amsterdam, 1987.

[56] J. Simons, Minimal varieties in riemannian manifolds. Ann. of Math(2)

88 (1968), 62–105.

[57] M. Struwe, Variational Methods. Springer, 1980.

[58] A. Takayama, Mathematical Economics. Cambridge University Press

1985.

[59] J. L. Troutman, Variational Calculus and Optimal Control; Optimiza-

tion with Elementary Convexity. Springer, 1996.

[60] E. Zeidler, Nonlinear Functional Analysis and its Applications. III Vari-
ational Methods and Optimizations. Springer, 1985.