Professional Documents
Culture Documents
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance
Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy
is not required from the publisher.
Printed in Singapore
Preface
Calculus of variations first appeared around the time when calculus was in-
vented, over 300 years ago. It used to be a mandatory course for undergraduate
students majoring in mathematics, following calculus and ordinary differential
equations. The main content was to turn variational problems into problems of
solving differential equations. However, since only a few differential equations
have explicit solutions, this limitation hindered the study. Since the second half of
the 20th century, as a course of its own, calculus of variations has been condensed
or merged into other related courses gradually.
Nevertheless calculus of variations has close connections to many other
branches of mathematics. Among the 23 problems proposed by Hilbert, three
of which deal with variational problems, its importance is evident. Variational
problems arise naturally from mechanics, physics, economics, operation research,
and engineering, etc. In particular, since the 1970s, finite element methods and op-
timization techniques have provided numerical solutions to variational problems,
thereby elevating its status in the realm of applied mathematics.
In the past few decades, great development has taken place in both theoretical
and applied aspects of calculus of variations. It is noted by the mathematical com-
munity that calculus of variations is quickly becoming a necessity in undergradu-
ate mathematics education, without it, students would struggle with new demands
of modern science and technology. However, there is no unanimous agreement so
far on how to remedy this shortcoming, it will take some time to explore. This
book attempts to bring the readers up to date in this subject area.
In the academic years between 2006 and 2010, the author taught the course
entitled “Calculus of Variations” to both advanced undergraduate students and
beginning level graduate students in mathematics at the School of Mathematical
Sciences, Peking University. The organization of the course content was based on
the following three principles:
v
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page vi
1. The lectures should not only introduce the classical theory but also the
modern development of calculus of variations. Furthermore, more in-depth stud-
ies were conducted stemming from various research problems.
2. The course should put its main emphasis on the most frequently used theo-
rems as well as techniques.
3. The course should welcome a large audience, including students in the area
of pure mathematics, numerical mathematics, mathematical statistics, information
science, and financial mathematics.
The prerequisites for this course were Mathematical Analysis, Modern Alge-
bra, and Analytic Geometry; in addition, students were expected to be somewhat
familiar with Ordinary Differential Equations, Real Analysis, Functional Analy-
sis, Differential Geometry, and Mathematical Physics.
The entire course was divided into three sections: classical theory of calculus
of variations, the existence and regularity of solutions, and special topics; whereas
the latter played a role in the development of modern day calculus of variation.
This book not only introduces fundamental concepts, basic theorems, and tech-
niques used in calculus of variations, but reinforces them with abundant examples
and counterexamples as well. In particular, it sheds new light on the definitions,
their interplays and compatibilities with existing theorems and methods. Exercises
are given at the end of most lectures.
This book is based on these lectures notes.
Due to the experimental nature of the book, I sincerely welcome any input and
critique on how to improve it.
I am very grateful to Tian Fu Zhao and Shan Nian Lu from the Higher Educa-
tion Publishing Company. Their careful proofreading and comments are greatly
appreciated.
Contents
Preface v
vii
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page viii
Contents ix
Contents xi
Bibliography 305
Index 309
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 1
Lecture 1
1.1 Introduction
The interplay between topology and the calculus of variations has led to
a new brach in mathematics — global analysis. It has vastly accelerated to the
advancement of critical point theory. In particular, Morse theory reveals the
1
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 2
interplay between analysis and topology, it has since become a core topic in dif-
ferential topology. In the same vein, Floer homology has also emerged as their
heir.
Calculus of variations has seeped through to many other areas, such as Rie-
mannian Geometry, Finsler Geometry, Symplectic Geometry, and Conformal Ge-
ometry, etc. Many variational problems with rich geometric background, such as
geodesics, minimal surfaces, harmonic maps, etc. have stimulated research inter-
ests in finding new theory (for example, geometric measure theory), new methods,
and new techniques.
Variational methods play an important role in the study of periodic orbits of
Hamiltonian dynamical systems, Mather set, and chaos.
Malliavin Calculus, the stochastic calculus of variations is an interplay
between differential calculus and probability theory. It has become an essential
part of financial mathematics.
It is clear that variational problems, theory, and methods have profound influ-
ence on various areas of modern mathematics, including pure mathematics, ap-
plied mathematics, numerical mathematics, information science, and mathemati-
cal economics, etc. No doubt, calculus of variations has taken the center stage of
modern mathematics.
Comparing to some classical textbooks (for examples [LL], [Ka], [GF]), this
volume offers the following unique features:
While introducing the classical theory of calculus of variations, emphasis is
put on the first order as well as second order conditions of a minimum.
Since many examples in partial differential equations, differential geometry,
and mathematical physics involve several variables, we give more extensive dis-
cussions on multivariate cases.
In the classical theory of calculus of variations, we strengthen the Hamilton–
Jacobi theory and the conservation law, since they are very useful content in
physics and geometry.
Aside from the classical theory of calculus of variations, we also empha-
size the direct methods and their applications. The direct methods are the main
body of modern calculus of variations; it is the foundation for establishing the
existence of solutions of differential equations and their numerical solutions. This
approach is accessible to students with previous exposure to functional analysis,
and it constitutes for nearly half of the material in this volume.
Eigenvalue problem is one of the central problems in analysis. Our treatment
is to present them as an application of solving constrained optimization problems,
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 3
1.2 Functionals
Calculus of variations examines the extremal values (or more generally, the
critical values) of functionals.
Generally speaking, a functional is a mapping from any set M to the field
of real numbers R or the field of complex numbers C. However, in calculus of
variations, a functional will only take values in R, whose domain M is a set of
functions, i.e. I : M → R.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 4
I2 (u) = u(x0 ),
Z
I3 (u) = [|∇u(x)|2 − F (u(x))]dx
Ω
are all functionals. However, regardless of the choices of M and the single vari-
able function f , the composite function
I4 (u) = f (u(x))
is not a functional.
Given a function L ∈ C 1 (Ω×RN ×RnN ), one mainly considers the following
functional:
Z
I(u) = L(x, u(x), ∇u(x))dx,
Ω
Example 1.1 (The line of steepest descent) Given two points A = (x1 , y1 ) and
B = (x2 , y2 ) in the xy-plane, where x1 < x2 and y1 > y2 . A particle is free
falling along a smooth curve joining A and B. Assuming the initial velocity of
the particle is zero, what trajectory would be the fastest to travel from A to B?
(see Figure 1.1)
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 5
Fig. 1.1
and
s
ds 1 + |u0 (x)|2
dt = = dx.
v 2g(y1 − u(x))
Let
Example 1.2 (Geodesics) Given two points P0 = (x01 , x02 , x03 ) and P1 =
P3
(x11 , x12 , x13 ) on the unit sphere S 2 = {(x1 , x2 , x3 ) ∈ R3 | i=1 x2i = 1}, find
the path joining them whose arclength is the shortest.
We adopt spherical coordinates v = (θ, ϕ) ∈ [− 21 π, 12 π] × [0, 2π) on S 2 such
that
x1 = x1 (v) = cos θ cos ϕ,
x = x2 (v) = cos θ sin ϕ,
2
x3 = x3 (v) = sin θ,
and v i = (θi , ϕi ) corresponds to Pi = (x1 (v i ), x2 (v i ), x3 (v i )) for i = 0, 1.
Let M = {v ∈ C 1 ([0, 1], [− π2 , π2 ] × [0, 2π)) | v(i) = v i , i = 0, 1}, then for all
v ∈ M , u(t) = (x1 (v(t)), x2 (v(t)), x3 (v(t))) for t ∈ [0, 1] is an arc connecting
P0 and P1 .
The square of the line element is
ds2 = dx21 + dx22 + dx23 = dθ2 + cos2 θdϕ2 = (θ0 (t)2 + cos2 θ(t)ϕ0 (t)2 )dt2 .
Hence, the arclength L : M → R is given by
Z 1 Z 1p
L(u) = |ds| = (θ0 (t)2 + cos2 θ(t)ϕ0 (t)2 )dt.
0 0
The arclength L is now a functional of the parametric function v(t) = (θ(t), ϕ(t)),
and we wish to find v ∈ M such that L(v) is a minimum.
Example 1.3 (Minimal surfaces) Given a Jordan curve Γ in R3 , is it possible to
find a surface S bounded by Γ whose area is a minimum?
We give the following parametrization of S : (u, v) 7→ Z = (x, y, z) as a map
from D̄ → R3 , where D ⊂ R2 is the unit circle, u2 + v 2 ≤ 1 such that
x = x(u, v),
y = y(u, v),
z = z(u, v).
We define
and we wish to find a function u1 such that the functional E : M → R attains its
minimum at u1 . Furthermore,
λ1 = min{E(u) | G(u) = 1}
is called the first eigenvalue with the corresponding first eigenfunction u1 . The
eigenpair has great importance in geometry, physics, and engineering.
Often times, inequalities in analysis and geometry can be posed as a variational
problem. For instance, the Sobolev’s inequality asserts
Z Z n−2
N
2N
|∇u(x)|2 dx ≥ SN |u(x)| N −2 dx ,
RN RN
where
−2/N
Γ(N )
SN = N (N − 2)π .
Γ(N/2)
We may turn this into a variational problem as follows. Let
Z
2N
1 N
M = u ∈ H0 (R ) |u(x)| N −2 dx = 1
RN
and
Z
E(u) = |∇u(x)|2 dx,
RN
we now find u ∈ M such that the functional E has a minimum. If the minimal
value is SN , then SN is the best constant for the above inequality.
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 8
Example 1.5 (Vibrations of thin plates) To study the vibrations of a thin elastic
plate (thin means the ratio between the thickness h and the minimal span a of the
plate satisfies h/a 1; the plate is assumed to be homogeneous and isotropic)
under exterior force. In continuum mechanics, Kirchnoff proposed the so-called
“assumption of straight line”, i.e. straight lines normal to the plate remains straight
and no stretch nor strain on the plate occurs during deformation.
Suppose a plane region Ω represents a thin plate, its density function is ρ(x, y).
Let w(x, y) be the displacement of (x, y) ∈ Ω. The stress–strain relation can
then be expressed via the potential energy density, which depends on the Hessian
matrix of w(x, y):
wxx wxy
.
wyx wyy
Since all physical quantities remain independent of the coordinates chosen, the
potential energy density can only depend on the quantities
2
wxx + wyy and wxx wyy − wxy .
Let f (x, y) be the density of the stress put on Ω, the resulting potential energy
density is only considered exterior work done. If we ignore the stress on ∂Ω, the
boundary of Ω and the bending on the boundary, the total potential energy is then
given by
Z
1 2 2
U (w) = [(wxx +wyy ) −2(1−µ)(wxx wyy −wxy )]+f (x, y)w(x, y) dxdy,
Ω 2
where µ is determined by the material of the plate itself, known as the Poisson
ratio.
Suppose now we fix the boundary of the thin plate, i.e. w|∂Ω = ϕ(x) with ϕ
a given function on ∂Ω. Let M = {w ∈ C 2 (Ω̄) | w|∂Ω = ϕ(x)}, then U is a
functional of the displacement function w ∈ M .
From the principles of mechanics, we know that the equilibrium of the plate
must obey variational principle. In other words, the potential energy functional
U of the plate achieves its minimum at the displacement w where the plate is
balanced.
Unlike the above tranditional examples, certain problems may not appear to
be related to problems of finding extremal values of functionals at first, however,
by proper modifications, they can be transformed into variational problems.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 9
Let dVg denote the volume element of M , then the total energy
Z
E(u) = L(u(x), du(x))dVg
M
is a functional of u.
The mapping u itself can also be regarded as a “function”. If we take M =
C 1 (M, N ), then E : M → R is a functional. Those mappings u ∈ M for which
E attains its minimal value are called harmonic mappings. They play an important
role in differential geometry.
Example 1.9 (Hamiltonian system) Given a function H ∈ C 1 (R1 × Rn ×
Rn ). We introduce the following notation: x = (x1 , . . . , xn ), p = (p1 , . . . , pn ),
Pn
(t, x, p) ∈ R1 × Rn × Rn , and (x, p)Rn = i=1 xi pi .
When both x and p are (vector-valued) functions of time t, the system of
ordinary differential equations
(
ẋ = Hp (t, x, p),
ṗ = −Hx (t, x, p)
is called a Hamiltonian system.
August 31, 2016 10:10 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 11
Lecture 2
Before we begin our study on the necessary condition for the extremal values
of a functional, we shall first review the necessary condition for the extremal val-
ues of a real-valued function. Let Ω ⊂ Rn be an open set. Suppose the function
f ∈ C 1 (Ω) attains its (local) minimum at some point x0 ∈ Ω, i.e. there exists an
open neighborhood U ⊂ Ω of x0 such that
f (x) ≥ f (x0 ), ∀ x ∈ U.
Hence, ∀ h ∈ Rn \ {θ}, where θ denotes the zero vector in Rn , ∃ ε(h) > 0 such
that whenever 0 < |ε| < ε(h), x0 + εh ∈ U and
i.e.
1
[f (x0 + εh) − f (x0 )] ≥ 0.
ε
Letting ε → 0, it yields
13
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 14
≥ 0,
∀ ϕ ∈ C01 (J, RN ). We call δI(u∗ , ϕ) the first order variation of I with respect
to ϕ.
If we replace ε > 0 by ε < 0, it is equivalent to replacing ϕ by −ϕ in the
above equation. Hence, ∀ ϕ ∈ C01 (J, RN ), we have:
Z X N Z t
∗ ∗ ∗ ∗
Lui (s, u (s), u̇ (s))ds − Lpi (t, u (t), u̇ (t)) ϕ̇i (t)dt = 0.
J i=1 t0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 15
where C01 (J) = C01 (J, R1 ), which is equal to {u ∈ C 1 (J) | u(t0 ) = u(t1 ) = 0},
then ψ is a constant.
1
Rt
ψ(t)dt and λ(t) = t0 (ψ(s) − c)ds, then λ ∈ C01 (J).
R
Proof Let c = |J| J
Thus,
Z Z Z
(ψ(t) − c)2 dt = ψ(t)(ψ(t) − c)dt = ψ(t) · λ̇(t)dt = 0.
J J J
Note that in M , ∀ δ > 0, U = {v ∈ Lip(J) | |v(t) − u∗ (t)| + |v̇(t) − u̇∗ (t)| <
δ a.e., t ∈ J} is a neighborhood of u∗ .
The E-L equation still holds in the sense of Lebesgue integral for almost all
t ∈ J:
Z t
Lui (s, u(s), u̇(s))ds − Lpi (t, u(t), u̇(t)) = const., 1 ≤ i ≤ N.
t0
To see this, we use the fact that the difference inside the above integral is uniformly
bounded, hence by the Lebesgue’s Dominant Convergence Theorem, we can pass
the limit inside the integral.
Aside from this, in the du Bois–Reymond Lemma, we may replace the re-
quirement ψ ∈ C 1 (J) by ψ ∈ L∞ (J) and λ ∈ C01 (J) by λ ∈ AC0 (J), the space
of all absolutely continuous functions on J which vanish on the boundary of J.
Remark 2.3 A piecewise C 1 continuous function is a Lipschitz function. By
a piecewise C 1 continuous function u, we mean there exists a finite set D =
{a1 , . . . , ak } such that u ∈ C 1 (J \ D) and u̇(ai ± 0) exists for 1 ≤ i ≤ k. In
this case, the integral form of the E-L equation continues to hold for the class of
piecewise C 1 continuous functions.
Many fundamental equations in mechanics and geometry are E-L equations.
We have the following examples:
Example 2.1 (The displacement of a moving particle) A force F is put on a
particle of mass m. Assume the particle’s displacement coordinate is given by
x = (x1 , x2 , x3 ) ∈ R3 with |x|2 = x21 + x22 + x23 , the velocity v = ẋ and kinetic
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 17
where
∂
gjk,i (u) = gjk (u).
∂ui
Consequently,
X X 1
gij (u)üj + (gij,k (u)u̇j u̇k − gkj,i u̇j u̇k ) = 0.
j
2
j,k
Note that
X X d X d X
glj,k (u)u̇k u̇j = glj (u)u̇j = glk u̇k = gkl,j (u)u̇k u̇j .
j
dt dt
j,k k j,k
ik
Let (g ) be the inverse matrix of (gik ), or using the Christoffel symbols of the
second kind
X
Γijk = g il Γjlk ,
l
Fig. 2.1
Next, we examine such local behavior via limits. For simplicity, we shall take
N = 1 and assume L ∈ C 2 , u ∈ C 2 . Note that
Z c+h Z t
EL (u)(s)dsϕ̇(t)dt
I(u + ϕ) − I(u) c−h t0
lim = − lim
∆σ ∆σ
Z c+h
EL (u)(t)ϕ(t)dt
c−h
= lim
∆σ
= EL (u)(c),
h → 0, sup |ϕ̇(t)| → 0.
t∈[c−h,c+h]
Of course, there are many other choices for the set M ; for example, one may
choose to fix only one endpoint and leave the other one free; we may also im-
pose different boundary conditions on the different components of a vector-valued
function.
In some of our latter discussions, we will also encounter other types of bound-
ary conditions, such as periodic and free boundary conditions.
Caution: In all of our previous discussions, we have always assumed all func-
tions u ∈ M are continuously differentiable. If we replace the continuously differ-
entiable functions by piecewise C 1 functions, although the E-L equation remains
the same (locally), it is however necessary to insert corner conditions at those
points where the derivative function has a jump discontinuity (see Exercise 2.4).
For N = 1, we shall consider the following special cases where the E-L equa-
tion can be simplified.
• Case 1. Suppose u is absent from L, then L = L(t, p) and
d
Lp (t, u̇(t)) = 0.
dt
Since Lp (t, u̇(t)) = c is a first order equation without u, assuming we can solve
for u̇ (e.g. Lpp (t, p) 6= 0) to get
u̇(t) = g(t, c),
then by integration, u(t) is also solved.
Example 2.3 Let M = {u ∈ C 1 ([1, 2]) | u(1) = 0, u(2) = 1} and
Z 2p
dt
I(u) = 1 + u̇2 .
1 t
Find u such that it minimizes the functional I.
p
Solution Since L = t−1 1 + p2 ,
p
Lp = p = C.
t 1 + p2
It follows that
Ct
u̇2 (1 − C 2 t2 ) = C 2 t2 or u̇ = ± √ .
1 − C 2 t2
Taking into account of the sign as part of the constant C, we integrate again to
obtain
1p
u= 1 − C 2 t2 + C1 .
C
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 22
q
1
Taking c = 2k , we then have:
ẋ(θ) = k(1 − cos θ).
Thus, we have:
x(θ) = x1 + k(θ − sin θ),
u(θ) = y1 − k(1 − cos θ),
θ ∈ [0, Θ].
Lastly, we use
x(Θ) = x2 ,
u(Θ) = y2 ,
to determine the k and Θ.
Example 2.5 (A minimal surface generated by a surface of revolution) Given
two points P1 = (x1 , y1 ) and P2 = (x2 , y2 ) in the xy-plane, where x1 < x2
and y1 , y2 > 0, we are to find a function u ∈ C 1 ([x1 , x2 ]) whose graph passing
through these points such that the surface generated by revolving the graph of u
about the x-axis has minimal area.
Without loss of generality, we may assume u(xi ) = yi , i = 1, 2 and u(x) > 0.
The area of the surface of revolution is given by
Z x2 p
I(u) = 2π u(x) (1 + u0 (x)2 )dx.
x1
We now find u to minimize I.
Since the Hamiltonian is conservative, we have:
L(u(t), p(t)) − u̇(t)Lp (u(t), p(t)) = C,
hence,
p uu̇2
u 1 + u̇2 − √ = C, ∀ t,
1 + u̇2
or equivalently,
p
u̇ = C −1 u2 − C 2 .
After integration, we arrive at
√
u+ u2 − C 2
C ln = x + C1 ,
C
or the equivalent form of
x + C1
u = C cosh ,
C
which are the standard catenary equations.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 24
whose inverse is
(
t = t(s, v),
u = u(s, v).
The Lagrangian L is now changed to
us + uv q
L̃(s, v, q) = L t(s, v), u(s, v), (ts + tv q).
ts + tv q
Suppose the image of t ∈ [t0 , t1 ] under the transformation is s ∈ [s0 , s1 ], we then
have:
Z t1 Z s1
L(t, u(t), u̇(t))dt = L̃(s, v(s), v̇(s))ds.
t0 s0
Exercises
prove that
H(u(t), u̇(t)) = const.
Given a collection of particles with degree of freedom n, we denote the dis-
placement coordinate by q = (q1 , . . . , qn ), the kinetic energy by
n
1 X
T = aij q̇i q̇j ,
2 i,j=1
Lecture 3
f (x) − f (x0 ) ≥ 0, ∀ x ∈ U.
This means ∃ 0 > 0 such that when 0 < || < 0 , for all h ∈ Rn \ {0},
x0 + h ∈ U and
This implies the one variable function 7→ f (x0 + h) has 0 as its minimum,
hence,
d2
f (x0 + h)|=0 ≥ 0.
d2
Namely,
n
X ∂2f
(d2 f (x0 )h, h) = (x0 )hi hj ≥ 0.
i,j=1
∂xi ∂xj
2
Therefore, the matrix d2 f (x0 ) = ( ∂x∂i ∂x
f
j
)(x0 ) is positive semi-definite.
2
Conversely, suppose the matrix d f (x0 ) is positive definite, then x0 must be a
strict local minimum of f .
29
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 30
d2
= I(u0 + sϕ)|s=0
ds2
d2
Z
= 2 L(t, u0 (t) + sϕ(t), u̇0 (t) + sϕ̇(t))dt|s=0
ds J
XZ
= [Lui uj (t, u0 (t), u̇0 (t))ϕi (t)ϕj (t)
i,j J
We then have:
Z
δ 2 I(u0 , ϕ) = [(Au0 ϕ̇, ϕ̇) + 2(Bu0 ϕ̇, ϕ) + (Cu0 ϕ, ϕ)]dt.
J
Since
Z
g̈(s) = [(Au0 +sϕ ϕ̇, ϕ) + 2(Bu0 +sϕ ϕ̇, ϕ) + (Cu0 +sϕ ϕ̇, ϕ)] dt
J
|Au0 +sϕ − Au0 | + |Bu0 +sϕ − Bu0 | + |Cu0 +sϕ − Cu0 | = o(1),
which yields
Z
g̈(s) − g̈(0) = o(s2 ) (|∇ϕ|2 + |ϕ|2 ) dt.
J
Hence, ∀ ϕ ∈ C01 (J, RN ), as long as |s| is sufficiently small, there exists < λ
such that
Z
I(u0 + ϕ) − I(u0 ) ≥ (λ − ) (|∇ϕ|2 + |ϕ|2 ) dt
J
Although (3.1) and (3.2) give us the necessary and sufficient conditions respec-
tively on u0 such that it minimizes I, its dependence on the arbitrary function ϕ is
nevertheless unsatisfying. We shall continue our study in the subsequent section.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 32
In our previous setting, notice that the roles of the three matrices A0 , B0 , and
C0 in determining whether u0 is a minimum are not all equal.
In fact, ∀ τ ∈ int(J), ∀ ξ ∈ RN , ∀ µ > 0 sufficiently small, one may choose
v ∈ C 1 (R1 ) with v(s) = 0 satisfying for |s| ≥ 1, R1 v̇(s)2 ds = 1. Let
R
t−τ
ϕ(t) = ξµv ,
µ
then
t−τ
ϕ̇(t) = ξ v̇ .
µ
For all µ > 0 sufficiently small, we have:
Z
ϕ̇i ϕ̇j dt = ξi ξj µ,
J
Z Z
2
ϕ̇i ϕj dt = ξi ξj µ v(t)v̇(t)dt,
J R1
Z Z
ϕi ϕj dt = ξi ξj µ3 v(t)2 dt.
J R1
Proof Since
Z t
ϕ(t) = ϕ̇(s)ds,
t0
If we remove the |ϕ|2 term from the integral on the right-hand side of (3.2) and
replacing it by
Z
δ I(u, ϕ) > λ |ϕ̇|2 dt, ∀ ϕ ∈ C01 (J, RN ),
2
(3.2)0
J
On one hand, from the second order variation, it is not difficult to see if the
matrix
Au0 Bu0
Bu0 Cu0
is positive definite, then the solution u0 of the E-L equation must be a minimum.
However, from the next example, we see that the positive definiteness of the matrix
is not necessary for u0 to be a minimum.
Example 3.2 Let I(u) = J (u̇(t)2 − u(t)2 )dt, then for u = 0,
R
AB 1 0
=
BC 0 −1
is not positive semi-definite.
However, when |J| = t1 − t0 is sufficiently small, from Poincaré’s inequality,
we still have;
(t1 − t0 )2
Z Z
δ 2 I(0, ϕ) = (ϕ̇2 − ϕ2 )dt ≥ 1 − ϕ̇2 dt.
J 2 J
Since Qu0 (ϕ) = δ 2 I(u0 , ϕ) ≥ 0, ∀ ϕ ∈ C01 (J, RN ) and Qu0 (θ) = 0, θ must be
a minimum.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 35
Thus, ϕ0 is a minimum of Qu0 . From our previous argument, it must satisfy the
integral form of the E-L equation, hence it also satisfies the differential form of
the E-L equation Ju0 (ϕ0 ) = 0.
Lemma 3.2 Given a sufficiently smooth Lagrangian L, suppose it satisfies the
strict Legendre–Hadamard condition along a solution u0 of the E-L equation,
namely, Au0 is positive definite. Suppose there exists µ > 0 such that
Z
Qu0 (ϕ) ≥ µ |ϕ|2 dt,
J
then there exists λ > 0 such that
Z
Qu0 (ϕ) ≥ λ (|ϕ̇|2 + |ϕ|2 )dt.
J
Consequently, u0 is a strict minimum of the functional
Z
I(u) = L(t, u(t), u̇(t)) dt.
J
R
Proof For any two continuous functions φ, ψ on J, let hφ, ψi = J
φ(t)ψ(t)dt.
Since Au0 is positive definite, there exists α > 0 such that
Z
hAu0 ϕ̇, ϕ̇i ≥ α |ϕ̇|2 dt.
J
From
Qu0 (ϕ) = hAu0 ϕ̇, ϕ̇i + 2hBu0 ϕ̇, ϕi + hCu0 ϕ, ϕi,
we can find two positive constants C1 and C2 such that
Z
α |ϕ̇|2 dt
J
Z 21 Z 21 Z
2 2
≤ Qu0 (ϕ) + C1 |ϕ̇| dt |ϕ| dt + |ϕ|2 dt
Z J Z J J
α 2 2
≤ |ϕ̇| dt + Qu (ϕ) + C2 |ϕ| dt.
2 J J
Using the assumption
Z
|ϕ|2 dt ≤ µ−1 Qu0 (ϕ),
J
we find that
Z
2
(1 + C2 µ−1 )Qu0 (ϕ).
|ϕ̇|2 dt ≤
J α
According to Theorem 3.1 and Poincaré’s inequality, u0 is a strict minimum of I.
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 37
Fig. 3.1
By Theorem 3.2, ξ˜ ∈ C 2 (J, RN ) must satisfy the Jacobi equation Ju0 (ξ) ˜ = 0.
By the uniqueness of the solution of the initial value problem of a second order
ordinary differential equation, ξ˜ ≡ 0, contradictory to ξ 6= 0, hence completes the
proof.
For the special case N = 1, we can also show that the converse of the above
theorem is also true.
We shall henceforth assume u0 is a solution of the E-L equation. Notice that
if u0 has no conjugate points on (t0 , t1 ], then there exists a positive Jacobi field
ψ > 0, ∀ t ∈ J.
To see this, suppose λ is a Jacobi field with the initial conditions λ(t0 ) = 0
and λ̇(t0 ) = 1. By assumption, the next root a satisfies a > t1 . Since the
solution of an ordinary differential equation varies continuously dependent on the
initial values, there exist > 0 and a Jacobi field ψ such that ψ(t0 − ) = 0,
ψ̇(t0 − ) = 1, and ψ(t) > 0, ∀ t ∈ J.
Lemma 3.3 Suppose ψ(t) > 0, ∀ t ∈ J is a Jacobi field along u0 , then for all
ϕ ∈ C01 (J), we have:
Z 0 2
ϕ
Qu0 (ϕ) = Au0 (t)ψ 2 (t) (t) dt.
J ψ
ϕ
Proof Let λ = ψ, then ϕ = λψ and ϕ0 = λ0 ψ + λψ 0 . Hence,
Theorem 3.4 Let N = 1 and assume u0 ∈ C 1 (J) is a solution of the E-L equa-
tion. If ∃ λ > 0 such that Au0 (t) ≥ λ, ∀ t ∈ J and there exists a Jacobi field
ψ > 0 on J, then u0 is a strict minimum.
Proof Denote
α = inf (Au0 (t)ψ 2 (t)) > 0.
J
Exercises
Lecture 4
if there exists > 0 such that for all ϕ ∈ C01 (Ω, RN ) with
kϕkC 0 (J) < (kϕkC 1 (J) < ),
we have:
I(u + ϕ) ≥ I(u).
The C 1 requirement can be replaced by Lip and using the Lip-norm instead of the
C 1 -norm. Without any confusion, we still call this a weak minimum. In Lecture 3,
the notion of a minimum agrees with a weak local minimum (cf. §3.2).
It is clear that a strong minimum is always a weak minimum and a weak min-
imum in the sense of Lipschitz is also a weak minimum under the C 1 -topology.
We have the following example.
Example 4.1 Let
Z 1
I(u) = (u02 + u03 )dx
0
43
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 44
and
M = {u ∈ Lip([0, 1], R1 ) | u(0) = u(1) = 0},
then u = 0 is a weak minimum.
In fact, for kukLip < 12 , we have:
Z 1
1 1 02
Z
02 03
I(u) − I(0) = (u + u )dx ≥ u dx ≥ 0.
0 2 0
Moreover, from
Z 1
δ 2 I(0, ϕ) = ϕ̇2 dt, ∀ ϕ ∈ C01 ([0, 1])
0
and Poincaré’s inequality, we see that u = 0 must be a strict (weak) local
minimum.
On the other hand, we claim u = 0 is not a strong minimum. ∀ 0 < h < 1−h2 ,
let
x
−h,
x ∈ [0, h2 ],
uh (x) =
h(x − 21) , x ∈ [h2 , 1],
1−h
then
1 − 1 ≤− 1 ,
x ∈ [0, h2 ],
h2 h3 2h3
0 0
(uh2 + uh3 )(x) = 2 3
h + h ≤ 2, x ∈ [h2 , 1].
1 − h2 1 − h2
We define
ϕλ (t) = ξψλ (t − τ ),
it is easy to check that kϕλ kC 0 = O(λ2 ), kϕ̇λ kC 0 = kξkRN , and
kϕλ kC 1 ([τ,τ +λ]) = O(λ). In particular, if we choose ϕ = ϕλ , then
I(u + ϕλ ) − I(u) ≥ 0.
It follows from the E-L equation that
Z Z
F (t) dt := {L(t, u(t) + ϕλ (t), u̇(t) + ϕ̇λ (t)) − L(t, u(t), u̇(t))
J J
− ϕλ (t)Lu (t, u(t), u̇(t)) − ϕ̇λ (t)Lp (t, u(t), u̇(t))}dt
≥ 0
and
Z Z τ −λ2 Z τ Z τ +λ Z t1
F (t) dt = + + + F (s) ds.
J t0 τ −λ2 τ τ +λ
Note the first and fourth integrals are both equal to zero, whereas the integrand of
the third integral is o(λ), whence
Z τ +λ
1
lim 2 F (s) ds = 0.
λ→0 λ τ
Lastly, the second integral yields:
Z τ
1
lim F (s) ds
λ→0 λ2 τ −λ2
= L(τ, u(τ ), u̇(τ ) + ξ) − L(τ, u(τ ), u̇(τ )) − ξLp (τ, u(τ ), u̇(τ )).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 46
Fig. 4.1
Fig. 4.2
Fig. 4.3
Fig. 4.4
curves, there are two distinct extremal curves passing through each point of Ω, so
it is not a field of extremals.
In contrast, the family of extremal curves uλ = λ sin t (∀ λ ∈ R1 ) generates
a field of extremals and its directional field where Ω = {(t, λ sin t) | (t, λ) ∈
(, π − ) × R1 } ( ∈ (0, π2 )) and ψ(t, u) = u cot t.
Suppose γ0 = {(t, u0 (t)) | t ∈ J} ⊂ Ω and ψ its directional field, we compare
it with its nearby piecewise C 1 curves γ = {(t, u(t)) | t ∈ J} ⊂ Ω with common
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 49
Thus,
Z
I(u) − I(u0 ) = [L(t, u(t), u̇(t)) − L(t, u(t), ψ(t, u(t)))
J
− (u̇(t) − ψ(t, u(t)))Lp (t, u(t), ψ(t, u(t)))]dt
Z
= E(t, u(t), ψ(t, u(t)), u̇(t))dt.
J
If we further assume
E(t, u, ψ(t, u), p) ≥ 0, ∀ (t, u, p) ∈ Ω × RN ,
then
I(u) ≥ I(u0 ),
which means u0 is a strong minimum.
We next address the following questions:
1. Is it possible to embed the extremal curve of u0 into a field of extremals?
By “embed” we mean there exists an open interval J1 ⊃ J, a continuous
function u : J1 × B → RN such that ∀ α = (α1 , α2 , . . . , αN ) ∈ B ⊂ RN ,
u(t, α) is an extremal curve for which u0 (t) = u(t, 0)|J , ∀ t ∈ J and Ω =
{t, u(t, α) | t ∈ J1 , α ∈ B } is a field of extremals.
2. Why is the above integral independent of path?
We now answer the first question: under what condition can a given extremal
curve γ0 be embedded into a field of extremals Ω?
For simplicity, we shall only present the argument for the case N = 1.
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 50
Fig. 4.5
it yields a family of solution ϕ(t, α), t ∈ (a, b), and |α| < 0 . By the unique-
ness of solutions of ordinary differential equations, ϕ(t, 0) = u0 (t). This fulfills
requirement (4) in Definition 4.2.
3. Define Ω = {(t, ϕ(t, α)) | t ∈ (a, b), |α| < }, < 0 . By Lemma 4.1,
ξ(t) = ∂α ϕ(t, 0)|α=0
is a Jacobi field along u0 which satisfies:
˙
ξ(a) = 0 and ξ(a) = 1.
By our assumption, it has no conjugate point, and Lemma 3.3 shows that there
exists ξ(t) > 0 for all t ∈ [a, t1 ]. According to the continuous dependence of
parameters, for 0 < 1 < 0 ,
∂α ϕ(t, α) > 0, |α| < 1 .
This fulfills requirement (2) in Definition 4.2.
Using the Implicit Function Theorem, we can find 0 < 2 < 1 , ∀ (t, u) ∈
Ω2 , the equation
u = ϕ(t, α)
has a unique continuously differentiable solution α = w(t, u) ∈ B2 (0). This
fulfills requirement (3) in Definition 4.2.
4. Let Ω = Ω2 , it contains γ0 and it is homeomorphic to a quadrilateral,
whence it is simply connected. Furthermore, the directional field ψ(t, u) gener-
ated by ϕ(t, u) is
ψ(t, u) = ∂t ϕ(t, w(t, u)).
It follows that ψ is defined everywhere in Ω and ϕ(t, α) describes a family of
solution curves of the equation
u̇ = ∂t ϕ(t, α) = ∂t ϕ(t, w(t, u)) = ψ(t, u).
This fulfills requirement (1) of Definition 4.2 (see Figure 4.6).
3
Remark 4.1 When N > 1, Lemma 4.2 still holds. It states: let L ∈ C ([t0 , t1 ] ×
RN × RN ) and u0 ∈ C 2 ([t0 , t1 ] × RN ) be a solution of its E-L equation. Suppose
for all t ∈ [t0 , t1 ], the matrix Lpp (t, u0 (t), u̇0 (t)) is invertible. If on (t0 , t1 ],
(t0 , u0 (t0 )) has no conjugate point along the extrema curve γ0 of u0 , then γ0 can
be embedded into a field of extremals Ω, a simply connected region generated by
the family of solution curves.
The proof resembles the proof of Lemma 4.2. We need only replace the scalar
α ∈ R1 in step 2 above by a vector α ∈ RN . The resulting solution ϕ(t, α) of the
E-L equation then satisfies:
∂α ϕj (a, α) = 0, ∂αi ∂t ϕj (a, α) = δij , 1 ≤ i, j ≤ N.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 52
Fig. 4.6
We modify the Jacobi field in the third step above by {wi (t, α) =
∂αi ∂t ϕ(a, α), i = 1, . . . , N }. Since
wi (a, 0) = 0, ∂αi wi (a, 0) = ei , i = 1, . . . , N
and on J1 , it has no conjugate point along γ0 , we have:
det(∂α ∂t ϕj (t, α)) 6= 0, ∀ t ∈ J1 .
The rest of the proof remains the same.
We now examine the second question: under what condition is the integral in
Theorem 4.1 independent of path?
Let
Ri (t, u) = Lpi (t, u, ψ(t, u)),
H(t, u) = hψ(t, u), Lp (t, u, ψ(t, u)i − L(t, u, ψ(t, u)),
where h·, ·i denotes the standard inner product on RN and the 1-form
N
X
ω= Ri dui − H dt.
i=1
We show that
dω = 0.
Definition 4.3 A field of extremals is called a Mayer field if it satisfies the fol-
lowing compatibility condition:
∂ui Lpj (t, u(t), ψ(t, u(t))) = ∂uj Lpi (t, u(t), ψ(t, u(t))), ∀ 1 ≤ i, j ≤ N.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 53
= Dψ L̃pi .
Lemma 4.4 If L ∈ C 2 (J × RN × RN ), then Ω is a Mayer critical field if and
only if dω = 0, i.e.
∂t Ri = −∂ui H, 1 ≤ i ≤ N.
Proof By direct computation, we have:
N
X
j
∂t Ri = ∂t + ψt ∂pj L̃pj .
j=1
Hence,
∂t Ri + ∂ui H
N
X N
X N
X
= ∂t + ψti ∂pi L̃p + j
ψ ∂uj + k j
ψ ∂uk ψ ∂pj L̃pi − L̃ui
i=1 j=1 k=1
= Dψ L̃pi − L̃ui .
Given a Mayer field (Ω, ψ), both R and H are already defined. If γ is any
curve joining the points (t0 , u(t0 )) and (t, u), the line intergal
Z (t,u) N
X
S(t, u) = Ri (t, u) dui − H(t, u) dt
(t0 ,u(t0 )) i=1
Z
= (Lp · du + (L − ψLp )dt)
γ
is independent of path. We shall call this line intergal the Hilbert invariant
integral.
Proof Notice
E(t, u, ψ(t, u), q) = L(t, u, q) − L(t, u, ψ(t, u)) − (q − ψ(t, u))Lp (t, u, ψ(t, u))
where v is in between ψ(t, u) and q. Our assertion follows from Lemmas 4.2 and
4.3, and the argument used in §4.3.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 55
has a solution
cosh t
u0 = .
cosh a
Since the Jacobi equation ϕ̈ − ϕ = 0 has no conjugate points and
has no conjugate points, whence u1 and u2 are both strict weak minima.
Lastly, we take a look at the Weierstrass excess function
have:
N
X
dω = d Ri dui − H dt
i=1
N
X N
X
i i
=d L̃pi du + L̃ − ϕ̇ L̃pi dt
i=1 i=1
N
X
= d L̃ dt + L̃pi (dui − ϕ̇i dt)
i=1
N
X N
X N
X
= L̃uk dϕk ∧ dt + L̃pk dϕ̇k ∧ dt − L̃pi dϕ̇i ∧ dt
k=1 k=1 i=1
N X
X N
+ dL̃pi ϕiαl dαl
i=1 l=1
X N
N X N X
X N
= dL̃uk ϕkαl dαl ∧ dt + ∂t L̃pi ϕiαl dt ∧ dαl
k=1 l=1 i=1 l=1
N X
X N X
N
+ ∂αm L̃pi ϕiαl dαm ∧ dαl
m=1 i=1 l=1
N X
X N N X
X N
N X
= (L̃ui − ∂t L̃pi )ϕiαl dαl ∧ dt + ∂αm L̃pi ϕiαl dαm ∧ dαl .
i=1 l=1 m=1 i=1 l=1
Denote
N
X
[αl , αm ] = (∂αl L̃pi ϕiαm − ∂αm L̃pi ϕiαl )
i=1
and we call it the Lagrange bracket. Using the Lagrange bracket, we can rewrite
the above calculation as the following formula:
N X
X N X
dω = EL (ϕ)i ϕiαl dαl ∧ dt + [αl , αm ] dαl ∧ dαm . (4.2)
i=1 l=1 1≤l<m≤N
Consequently, we have:
Lemma 4.5 Let L ∈ C 3 (J × Rn × RN ). Suppose (Ω, ψ) is an extremal field
determined by a family of extremal curves (ϕ, α), then (Ω, ψ) is a Mayer field if
and only if
EL (ϕ(·, α)) = 0, ∀ α ∈ RN , [αl , αm ] = 0, 1 ≤ l, m ≤ N.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 58
∂ L̃ ∂ L̃ ∂ L̃ ∂ L̃
After differentiating ∂αuli , ∂αumi , ∂αpli , ∂αpmi , and using the fact that L is twice
differentiable, all terms will cancel, which leaves us
∂ l m
[α , α ] = 0.
∂t
In the following, we will strengthen Remark 4.1 for the higher dimensional cases,
embedding a critical curve γ0 defined on J = [t0 , t1 ] into a Mayer field (Ω, ψ).
Theorem 4.5 Suppose the extremal curve γ0 has no conjugate points on J. Sup-
pose also (Lpi pj (t, u(t), u̇(t)) is invertible for all t ∈ J, then γ0 can be embedded
into a family of extremal curves such that this family of curves defines a Mayer
field (Ω, ψ).
Proof Based on Remark 4.1, we obtain a field of extremals (Ω, ψ), we now show
that this is a Mayer field. In fact, using ∂αi ϕj (a, α) = 0, ∀ i, j, we can deduce
d
that [αl , αm ](a, α) = 0, ∀ l, m ∈ [1, N ]. By Lemma 4.6, dt [αl , αm ](t, α) = 0,
l m
hence [α , α ](t, α) = 0. This means (Ω, ψ) is a Mayer field.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 59
Exercises
1. Use the Weierstrass excess function to verify that u0 is not a strong minimum
of I(u) = J (u̇2 + u̇3 ) dt.
R
R2p
2. Verify that for I(u) = 0 u(1 + u̇2 ) and M = {u ∈ C 1 ([0, 2]) | u(0) =
2, u(2) = 5}, there is a two-parameter family of solutions to the E-L equation
2
1 t+β
u(t, α, β) = α + .
α 2
When (α, β) = (1, 2), u(t, 1, 2) ∈ M .
Verify that
2
u0 (t) = − + 3
t
is a strong minimum.
5. Suppose u ∈ C 1 (J × RN , RN ) satisfies El (u(t, α)) = 0, ∀ t ∈ J, ∀ α ∈ RN ,
u(0, α) = θ, ∂t u(0, α) = α, and u0 = u(·, 0). Prove that
(1) ∂αi u(t, α)|α=0 (i = 1, . . . , N ) are Jacobi fields.
(2) If (t, u0 (t)) has no conjugate points on J, then det(∂αi uj (t, α)) 6= 0,
∀ t ∈ J.
Rb
6. Suppose I(u) = a [ 12 u̇2 − V (u)] dt.
(1) When V (u) = cos u, write down the Jacobi operator and the Jacobi field
along the E-L equation solution u = 0.
(2) When V ∈ C 2 (R1 ), V 00 (u) ≤ 0, show that u0 is a strict weak minimum
of I.
May 2, 2013 14:6 BC: 8831 - Probability and Statistical Theory PST˙ws
Lecture 5
we know from the previous lecture that for a field of extremals and its directional
field (Ω, ψ), there exists a 1-form
where γ is any curve connecting (t0 , u0 ) and (t, u). We call this single-valued
function g an eikonal.
Remark 5.1 The eikonal has the following physical meaning: consider the line
integral of ω along an extremal curve γ = (t, u(t)), u̇(t) = ψ(t, u(t)):
61
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 62
Z
g(t2 , u(t2 )) − g(t1 , u(t1 )) = ω
γ
Z
= (hLp (t, u(t), ψ(t, u(t))), dui
J
− [hψ(t, u(t)), Lp (t, u(t), ψ(t, u(t)))i
− L(t, u(t), ψ(t, u(t)))])dt
Z
= hLp (t, u(t), u̇(t)), u̇(t)i
J
− [hu̇(t), Lp (t, u(t), u̇(t))i − L(t, u(t), u̇(t))]dt
Z
= L(t, u(t), u̇(t))dt.
J
This shows that the difference of an eikonal g between any two points along a
given extremal curve is equal to the line integral of the Lagrangian along that
curve.
In optics, the Lagrangian represents the distance of a ray of light traveled in
an instance divided by its speed. The line integral is therefore equal to the time
elapsed when the light ray travels from point one to point two. Consequently, the
wavefront of a light ray from a one point source (or the phase front of waves) can
be expressed via the level sets g(t, u) = const.
The definition shows directly that on the field of extremals (Ω, ψ), the eikonal
g satisfies the following Carathéodory system of equations:
(
∇u g(t, u) = Lp (t, u, ψ(t, u))
(5.2)
∂t g(t, u) = L(t, u, ψ(t, u)) − hψ(t, u), Lp (t, u, ψ(t, u))i.
Fig. 5.1
(3)
−1
∂2
2
∗
∂ f
f (ξ)
= (x) .
∂ξi ∂ξj ξ=∇f (x) ∂xi ∂xj
This is because
x = ψ(∇f (x)) = (∇f ∗ )((∇f )(x)),
differentiating on both sides, we have the identity matrix
∂2 ∂2
∗
I= f (ξ) · f (x).
∂ξi ∂ξj ∂xi ∂xj
ξ=∇f (x)
1
Example 5.1 Let f (x) = xp /p, p > 1, then ξ = f 0 (x) = xp−1 , so x = ξ p−1 .
By definition,
1 0
f ∗ (ξ) = ξ · x − f (x) = 0 ξ p ,
p
1 1
where p0 + p = 1.
Example 5.2 Suppose A = (aij ) is a symmetric invertible N × N matrix. Let
f (x) = 21 hAx, xi, ∀ x ∈ RN , then ξ = ∇f (x) = Ax is invertible and we can
solve for x = A−1 ξ. So the Legendre transformation is
1 1
f ∗ (ξ) = hx, ξi − f (x) = hA−1 ξ, ξi − hξ, A−1 ξi = hA−1 ξ, ξi.
2 2
C2
Given a Lagrangian L : R1× RN × RN −→ R1 .
Suppose det Lpi pj (t, u, p) 6= 0. Let
ξi = Lpi (t, u, p), 1 ≤ i ≤ N.
Using the Implicit Function Theorem, we can solve the system of equations
locally
pi = ϕi (t, u, ξ), ξ = (ξ1 , ξ2 , . . . , ξN ), 1 ≤ i ≤ N.
Fix (t, u), as a function of p, we apply the Legendre transformation on L and let
N
X
H(t, u, ξ) = L∗ (t, u, ξ) = ξi pi − L(t, u, p)|p=ϕ(t,u,ξ) . (5.3)
1
Conversely, given an H-S solution (u(t), ξ(t)), by letting p(t) = u̇(t), and by
ξ(t) = Lp (t, u(t), p(t)), we can deduce (u(t), p(t)) is a solution of the E-L
equation:
d ˙ = −Hu (t, u(t), ξ(t)) = Lu (t, u(t), p(t)).
Lp (t, u(t), u̇(t)) = ξ(t)
dt
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 66
From this, we can establish the following one-to-one correspondence between the
E-L equation and H-S:
(E-L) ←→ (H-S)
(u(t), p(t)) ←→ (u(t), ξ(t)) = (u(t), Lp (t, u(t), p(t))) .
The H-S is also the E-L equation
Z of the functional:
F (u, ξ) = [hu̇(t), ξ(t)i − H(t, u(t), ξ(t))]dt.
J
Its corresponding 1-form is
α = ξdu − Hdt,
which we call the Poincaré–Cartan invariant.
Since H is the Legendre transform of L, L is also the Legendre transform of
H. This implies:
L(t, u(t), u̇(t)) = hu̇(t), ξ(t)i − H(t, u(t), ξ(t)). (5.5)
From this, we see that the integrands in the functional F and the functional I are
in fact the same function expressed via different variables, while the Poincaré–
Cartan invariant is indeed the differential version of the Hilbert invariant integral
from the previous lecture.
It is worth noting that the functional F (u, ξ) corresponding to the H-S is not
bounded below, hence it is impossible to possess any minimal value. The solutions
of the H-S are merely “critical points” of the functional F .
Example 5.3 For a given collection of particles in classical mechanics (cf. Ex-
ample 2.1 in Lecture 2), let q = (q1 , . . . , qN ), p = (p1 , . . . , pN ), T (p) =
1
PN
2 1 aij pi pj , and V = V (q1 , . . . , qN ), its Lagrangian is
L(t, q, p) = T (p) − V (q),
whose corresponding E-L equation is
N
d X
aij q̇j = −∂qi V (q), i = 1, . . . , N.
dt 1
Here the Hamiltonian
N
1 X ij
H(q, ξ) = a ξi ξj + V (q)
2 1
is the energy of the collection of particles, where (aij ) is the inverse matrix of
(aij ). The corresponding H-S is
˙
ξi = −∂qi V (q),
XN
q̇
i
= aij ξj , i = 1, 2, · · · , N.
j=1
When the Hamiltonian H is independent of t, ∀ c ∈ R1 , let H −1 (c) = {(u, ξ) ∈
RN × RN | H(u, ξ) = c} be a level set of the Hamiltonian.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 67
Theorem 5.1 The solution curve {(t, u(t), ξ(t) | ∀ t} of a Hamiltonian system
remains on the same level set.
Proof Suppose (u(t), ξ(t)) is a solution of
˙ = −Hu (u(t), ξ(t)),
ξ(t)
u̇(t) = Hξ (u(t), ξ(t)),
then using
d ˙
H(u(t), ξ(t)) = hHu (u(t), ξ(t)), u̇(t)i + hHξ (u(t), ξ(t)), ξ(t)i = 0,
dt
it is immediate for all t, H(u(t), ξ(t)) = const.
Applying this theorem to a collection of moving particles, it asserts: the total
energy of an isolated system remains constant (the law of conservation of energy).
Given a Hamiltonian H = H(t, u, ξ), we call the first order partial differential
equation
∂t S(t, u) + H(t, u, ∇u S(t, u)) = 0 (5.6)
the Hamilton–Jacobi equation (we shall abbreviate it by H-J equation), where
S = S(t, u) is a function of N + 1 variables.
The H-J equation is a fundamental equation in both classical mechanics and
quantum mechanics.
We know for a Mayer field (Ω, ψ), its eikonal g satisfies the Carathéodory
system of equations (5.2):
∇u g(t, u) = Lp (t, u, ψ(t, u))
∂t g(t, u) = L(t, u, ψ(t, u)) − hψ(t, u), Lp (t, u, ψ(t, u))i.
Substituting the Legendre transform ξ = Lp (t, u, ψ(t, u)) = ∇u g(t, u) into the
Carathéodory system of equations, we see that g satisfies the first order partial
differential equation:
∂t g(t, u) + H(t, u, ∇u g(t, u)) = 0, (5.60 )
where H is the Legendre transform of the Lagrangian L.
Based on the relationship of the Legendre transforms of the Lagrangian and
the Hamiltonian, for a given Hamiltonian H, we can write L and transforming a
solution (u(t), ξ(t)) of the H-S:
ξ˙i (t) = −Hui (u(t), ξ(t)),
into a solution (u(t), p(t)) of the E-L equation. After integrating, we can express
the eikonal g(t, u).
It is evident, if we have a solution (u(t), ξ(t)) of the Hamiltonian system, by
letting
p(t) = Hξ (t, u(t), ξ(t)),
we can then obtain the solution (u(t), p(t)) of the E-L equation.
Moreover, using (5.5), we can obtain the Lagrangian L(t, u(t), p(t)), whence,
Z t
g(t, u(t)) − g(t0 , u(t0 )) = L(t, u(t), p(t))dt
t0
On the other hand, we can also derive the solutions of H-S from the solutions
of the H-J equation.
Definition 5.1 Let g = g(t, u1 , . . . , uN ; λ1 , · · · , λN ) be a family of solutions of
the H-J equation depending on the N independent parameters (λ1 , . . . , λN ) ∈ Λ
(here Λ ⊂ RN is a region). If det(gui λj ) 6= 0, then it is called a complete integral.
Theorem 5.2 (Jacobi) Let the C 2 function g(t, u1 , . . . , uN ; λ1 , . . . , λN ) be
a complete integral of the H-J equation. Given 2N parameters (α, β) =
(α1 , . . . , αN , β1 , . . . , βN ), suppose the function
u = U (t, α, β),
p = P (t, α, β),
satisfies
gαi (t, U (t, α, β), α) = −βi
(5.7)
Pi (t, α, β) = gui (t, U (t, α, β), α) i = 1, 2, . . . , N,
then they form a family of solutions of the Hamiltonian system.
Proof 1. First, differentiating the H-J equation (5.60 ) with respect to αi , we have:
N
X
gt,αi + Hξk (t, u, ∇u g)gαi ,uk = 0, i = 1, 2, . . . , N.
k=1
Remark 5.3 Complete integrals and general solutions have different mean-
ings. This is because the H-J equation is a first order partial differential equa-
tion, and according to the uniqueness of solution to a Cauchy problem, its
general solution should contain an arbitrary function ϕ(u), not just 2N indepen-
dent parameters.
However, the general solutions of the H-S can be determined by a complete
integral g of the H-J equation. In other words, for a given initial value (u0 , ξ0 ), to
solve the initial value problem of the H-S
u̇ = Hξ (t, u, ξ), ξ˙ = −Hu (t, u, ξ), u(0) = u0 , ξ(0) = ξ0 ,
we may proceed as follows:
Once (u0 , ξ0 ) is given, since det(gui αj ) 6= 0, we can first apply the Implicit
Function Theorem to the second equation of the system
∇α g(0, u0 , α) = −β,
ξ0 = ∇u g(0, u0 , α)
to solve for
α0 = α(u0 , ξ0 ).
Setting
β0 = −∇α g(0, u0 , α0 )
and plugging them into (5.12) and (5.13), we will end up with the solution of the
H-S with initial value (u0 , ξ0 ).
Example 5.5 (Harmonic Oscillators) Given the Lagrangian L = 21 (mp2 −ku2 ),
where m and k are positive constants. Then the Hamiltonian is
1 p2
2
H= + ku .
2 m
The corresponding H-S
u̇ = p ,
m
ṗ = −ku.
has solution
r
k
u = C sin (t + t0 ),
m
√
r
p = C mk cos k (t + t0 ),
m
where t0 and C are arbitrary constants.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 72
Exercises
1. For each of the given Lagrangian, find the Hamiltonian and solve the Hamilto-
nian system.
(1) L = (p +pku)2 , k 6= 0.
(2) L = e−u 1 + p2 .
Write the H-J equation and find its complete solutions.
2. Suppose ∀ (t, u), L(t, u, p) is convex in p. Prove that
H(t, u, ξ) = sup {hp, ξi − L(t, u, p)}.
p∈RN
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 73
Lecture 6
u0 ∈ M is said to be a minimum of I on M , if
I(u) ≥ I(u0 ), ∀ u ∈ U,
73
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 74
The assertion follows from a result (Lemma 6.1) similar to the Du Bois Reymond’s
lemma.
We will prove a generalized version of the Du Bois Reymond’s lemma in
higher dimensions. But first, we introduce the 1-variable bump function:
−1
exp |t| < 1
ψ(t) = 1 − |t|2
0 |t| > 1.
where B1 (θ) is the unit ball in Rn centered at the origin. Hence, ∀ δ0 > 0,
Z Z Z
|u(x)−uδ (x)|dx ≤ ϕ(z)dz |u(x)−u(x+δz)|dx → 0, as δ → 0.
Ωδ0 B1 (θ) Ωδ0
It follows from Lemma 6.1 that uδ (x) → u(x) a.e. for x ∈ Ωδ0 , namely u(x) = 0
a.e. for x ∈ Ωδ0 . Since δ0 > 0 is arbitrary, it is immediate u(x) = 0 a.e. for
x ∈ Ω.
Just like in the 1-variable setting, we call EL : u 7→ v = (v1 , . . . , vN ), where
n
X ∂Lpiα (x, u(x), ∇u(x))
vi = − Lui (x, u(x), ∇u(x)), i = 1, . . . , N
α=1
∂xα
Similar to Remarks 2.2 and 2.3 in Lecture 2, when dealing with variational
problems involving multivariate integrals, we can change the domain M of a func-
tional from C 1 (Ω̄, RN ) to Lip(Ω̄, RN ), or more specially, to P W C 1 (Ω̄, RN ), the
class of piecewise C 1 functions. u ∈ P W C 1 (Ω̄, RN ) if there exists a finite col-
lection of n − 1 dimensional piecewise C 1 -hypersurfaces {S1 , . . . , Sk } such that
Sk
the continuous function u ∈ C 1 (Ω̄ \ j=1 Sj , RN ) and u has normal derivatives
on both sides of Sj .
Example 6.1 (Dirichlet integrals) Assume N = 1, L(p) = |p|2 /2, for the
functional
Z
1
D(u) = |∇u(x)|2 dx,
2 Ω
its E-L equation is
n
X ∂2u
∆u = = 0, ∀ x ∈ Ω.
α=1
∂x2α
This is a harmonic equation, also called the Laplace equation.
We can also consider a more generalized variational problem, such as
1 2 a(x) α+1
L(x, u, p) = |p| − |u| , α > 0,
2 α+1
where a ∈ C(Ω̄). Its E-L equation is
−∆u(x) = a(x)|u|α−1 u, ∀ x ∈ Ω.
Example 6.2 (Wave equations) Denote R1 × R3 the time-space continuum, a
given point has coordinate (t, x), where t represents time and x = (x1 , x2 , x3 )
represents its spacial location. We use u = u(t, x1 , x2 , x3 ) to represent the prop-
agation of an elastic wave in the region Ω ⊂ R1 × R3 .
The kinetic energy of the elastic wave is
Z
1
T = |∂t u(t, x)|2 dtdx
2 Ω
and the potential energy is
Z
1
U= |∇x u(t, x)|2 dtdx.
2 Ω
The Lagrangian is
L(u) = T (u) − U (u).
According to the principle of stable action, the propagation of the elastic wave is
a solution of the E-L equation
2u = ∂t2 u − 4u = 0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 77
Comparing (6.2) and (6.3), the mean curvature with mean curvature zero equation
coincides with the minimal surface equation. Therefore, the zero mean curvature
surfaces are usually called minimal surfaces.
We now find a special solution u(x, y) = f (x) + g(y) of the minimal surface
equation. Substituting this into (6.3) with H = 0, it yields
(1 + g 0 (y)2 )f 00 (x) + (1 + f 0 (x)2 )g 00 (y) = 0,
i.e.
f 00 (x) g 00 (y)
= c = − .
1 + f 0 (x)2 1 + g 0 (y)2
From which, we find arctan f 0 (x) = cx or equivalently
1
f (x) = − ln | cos cx|.
c
Likewise,
1
g(y) = ln | cos cy|.
c
Thus,
1 cos cy
u(x, y) = ln .
c cos cx
The minimal surface defined by u is known as the Scherk surface.
Example 6.4 (Maxwell’s equations) Consider a point with space-time coordinate
(x0 , x1 , x2 , x3 ) ∈ R1 × R3 , where x = (x1 , x2 , x3 ) denotes its spacial coordinate
and x0 = ct, t is time and c is the speed of light.
In an electromagnetic field, the electric charge ρ and the electric current j are
both functions of space-time. Let E = (E1 , E2 , E3 ) and B = (B1 , B2 , B3 ) de-
note the electric field and the magnetic field respectively, which are also functions
of space-time.
Maxwell’s equations can be written as
− 1c ∂B = ∇ × E
(Faraday’s law of induction)
∂t
∇·B = 0 (Gauss’s law for magnetism)
∇ × B = 1c ∂E + 4π c j (Ampere’s circuital law)
∂t
∇ · E = 4πρ.
(Gauss’s law for electric charge)
Since ∇ · B = 0, there exists a magnetic potential A = (A1 , A2 , A3 ) such that
∇ × A = B.
From Faraday’s law of induction,
∂A 1 ∂B
∇× E+ =∇×E+ = 0,
∂x0 c ∂t
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 79
E3 B2 −B1 0
We now define the Lagrangian
3 3
1 1X 1 X 2
L=− ji Ai + Fik
c c i=0 16π
i,k=0
to replace (Aj )30 , then the corresponding E, H remain unchanged. The above
transformation is called a gauge transformation. If, in addition, we impose the
Lorentz condition
3
∂A0 X ∂Aj
− + = 0,
∂x0 1
∂xj
it will resolve the non-uniqueness issue.
Bu0 = (bjk
β ) = (Lpj uk (τ )),
β
i.e.
N
X n
X
ajk j k
αβ (x0 )ξ ξ ηα ηβ ≥ 0.
j,k=1 α,β=1
∀ (x, ξ, η) ∈ Ω × RN × Rn . (6.8)
then there exist α > 0 and C0 > 0 such that for all ϕ ∈ C01 (Ω, RN ),
Z X n XN Z Z
jk j k 2
aαβ (x)∂α ϕ ∂β ϕ dx ≥ α |∇ϕ(x)| dx − C0 |ϕ(x)|2 dx.
Ω α,β=1 j,k=1 Ω Ω
then
Z
−2πiξα ϕ̂(ξ) = ∂α ϕ(x) exp [−2πihξ, xiRn ]dx.
Rn
N
X n Z
X
= 4π 2 ajk j k
αβ ξα ξβ ϕ̂ (ξ)ϕ̂ (ξ)dξ
j,k=1 α,β=1 Rn
Z
≥ 4π 2 σ |ξ|2 |ϕ̂(ξ)2
Rn
Z
=σ |∇ϕ(x)|2 dx.
Ω
When (ajk
αβ (x))
are non-constants, we can employ the argument of partition
of unity, treating the coefficients as if they are constants in each small neighbor-
hood, and then piece them together using the estimate given above. The remainder
can be combined into the second integral on the right-hand side be means of the
Schwarz’s inequality. Since the details of this proof is rather complicated and be-
yond the scope of this course, we shall omit it and refer the interested readers to
K. Yosida, Functional Analysis, pp. 175–177.
Lemma 6.3 Let L ∈ C 2 satisfy the strict Legendre–Hadamard condition, namely
∃ σ > 0 such that
Lpi pj (x, u, p)ξ i ξ j ηα ηβ ≥ σ|ξ|2 |η|2 , ∀ (x, u, p) ∈ Ω × RN × RnN .
α β
Suppose u0 is a solution of the E-L equation, and there exists µ > 0 such that
Z
Qu0 (ϕ) ≥ µ |ϕ|2 dx ∀ ϕ ∈ C01 (Ω, RN ),
Ω
From
Z
Qu0 (ϕ) = Au0 (∇ϕ · ∇ϕ) + 2Bu0 (∇ϕ · ϕ) + Cu0 (ϕ · ϕ),
Ω
we then have
Z
2
|∇ϕ(x)|2 dx ≤ (1 + C2 µ−1 )Qu0 (ϕ).
Ω α
Combining the two inequalities, it follows that there exists λ > 0 such that
Z
Qu0 (ϕ) ≥ λ (|∇ϕ|2 + |ϕ|2 )dx.
Ω
the first eigenvalue of the Jacobi operator (details are given in Lecture 12).
Theorem 6.3 Let L ∈ C 2 satisfy the strict Legendre–Hadamard condition. Sup-
pose u0 ∈ M is a weak minimum of I, then λ1 ≥ 0. Furthermore, if λ1 > 0, then
u0 is a strict weak minimum of I.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 87
The idea of the proof follows closely to that of a single variable, however, it is
more complicated, so we shall omit it here.
When n > 1 but N = 1, there is also a similar sufficient condition for strong
minimum (Lichtenstein theorem), we refer to [GH], p. 390.
Exercises
R q
1. Let I(u) = Ω
1 + u2x + u2y dxdy, ∀ ϕ ∈ C0∞ (Ω), find the first variation
δI(u, ϕ) and the second variation δ 2 I(u, ϕ).
2. Suppose g = (gαβ (x))1≤α,β≤n is a continuous positive definite marix defined
on a closed and bounded region Ω. Denote det(g) its determinant and (g αβ (x))
its inverse matrix. Let
Z X n
1
I(u) = g αβ (x)∂α u(x)∂β (x) det(g) dx1 · · · dxn .
2 Ω
α,β=1
Lecture 7
89
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 90
Since the perimeter l is given, our goal is to find a curve (x(θ), y(θ)) such that
functional A is a maximum under the constraint L = l.
We first recall the method used on constrained optimization problems in math-
ematical analysis. Assume f, g ∈ C 1 (Ω, R1 ), where Ω ⊂ Rn is an open subset.
Assume g −1 (1) 6= ∅. If x0 ∈ Ω is such that the function f attains its minimum
under the constraint g(x0 ) = 1 with ∇g(x0 ) 6= 0:
f (x0 ) = min
−1
f (x),
g (1)
then we can apply the Lagrange multiplier to turn this constrained minimization
problem into an unconstrained minimization problem.
To be more explicit, there exists a Lagrange multiplier λ ∈ R1 such that
∇f (x0 ) + λ∇g(x0 ) = 0.
The Lagrange multiplier has a clear geometric meaning: if M = g −1 (1) is a
differentiable manifold, then ∇f (x) is parallel to the outward normal vector at
the point (x, g(x)), whenever x is the constraint minimizer. We restrict f on M
and denote it by f˜ = f |M , whose differential
∇f (x) · ∇g(x)
df˜(x) = ∇f (x) − ∇g(x).
k∇g(x)k2
At an extreme point, it must satisfy
df˜(x) = 0 ⇔ ∇f (x) + λ∇g(x) = 0,
where
∇f (x) · ∇g(x)
λ=−
k∇g(x)k2
is the projection of −∇f (x) onto the unit outward normal ∇g(x) (when
k∇g(x)k = 1).
This indicates by means of the Lagrange multiplier, the solution of the con-
strained extreme value problem becomes a stationary point (critical point) of the
adjusted function f + λg.
Constrained extreme value problem of a functional can also be turned into
an unconstrained extreme value problem of another functional via the Lagrange
multiplier. We have the following:
Theorem 7.1 Given L, G ∈ C 2 (Ω̄ × RN × RnN ), ρ ∈ C 1 (∂Ω, RN ), and M =
{u ∈ C 1 (Ω̄, RN ) | u|∂Ω = ρ}. Define on M the functionals
Z
I(u) = L(x, u(x), ∇u(x))dx
Ω
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 91
and
Z
N (u) = G(x, u(x), ∇u(x))dx.
Ω
Fig. 7.1
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 92
and
Ψ(, τ ) = N (u0 + ϕ + τ ϕ0 ).
Notice
Ψ(0, 0) = N (u0 ) = c, Ψτ (0, 0) = δN (u0 , ϕ0 ) = 1,
for 0 , τ0 > 0 sufficiently small, we now apply the implicit function theorem on
R = (−0 , 0 ) × (−τ0 , τ0 ): ∃ ξ ∈ C 1 (−0 , 0 ) such that (, ξ()) ∈ R is the
unique solution of
Ψ(, τ ) = c
inside R. This means
Ψ(, ξ()) = c, ξ(0) = 0, ξ 0 (0) = −Ψ (0, 0) = −δN (u0 , ϕ).
So inside a small enough R,
N −1 (c) ∩ π = u0 + ϕ + ξ()ϕ0 .
Let
g() = I(u0 + ϕ + ξ()ϕ0 ).
Since v = u0 + ϕ + τ ϕ0 ∈ M,
kv − u0 kC 1 ≤ ||kϕkC 1 + |τ |kϕ0 kC 1 .
However, u0 is a weak minimum of I on M ∩ N −1 (c), so for 0 , τ0 sufficiently
small,
Φ(, ξ()) ≥ Φ(0, 0).
Thus, 0 is a minimum of g.
From this, we deduce that
0 = g 0 (0)
d
= Φ(, ξ(τ ))|=0
d
= Φ (0, 0) + Φτ (0, 0)ξ 0 (0)
= δI(u0 , ϕ) + λδN (u0 , ϕ), (7.2)
where
λ = −Φτ (0, 0)
is independent of ϕ. Letting Q = L + λG, (7.2) is the E-L equation associated to
Q:
div Qp (x, u0 (x), ∇u0 (x)) = Qu (x, u0 (x), ∇u0 (x)).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 93
Proof We will first construct such λ to be continuous locally on each piece, and
then glue them together to a globally defined continuous function on Ω̄.
1. ∀ x0 ∈ Ω, there exists a ball Br (x0 ) ⊂ Ω such that ∇x (M (u0 (x))) = 6 0,
∀ x ∈ Br (x0 ). Since ∇u M (u0 (x)) · ∇u0 (x) = ∇x (M (u0 (x))), ∇u M (u) 6= 0,
∀ u ∈ u0 (Br (x0 )), ∇u0 (x) 6= 0, ∀ x ∈ Br (x0 ). Without loss of generality,
we may assume MuN (u0 (x)) 6= 0, ∀ x ∈ Br (x0 ). If we adopt the notation
ũ = (u1 , . . . , uN −1 ), then we can solve and obtain uN = U (ũ), where U is a C 2
function (see Figure 7.2).
Fig. 7.2
−1
NX
N
p = Uui piα ,
i=1 1≤α≤n
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 96
and
N
X −1
PN = Uui ∇ui .
i=1
Now, if we replace u0 (x) by v(x) on B̄r (x0 ), then we will end up with a new
piecewise differentiable function whose functional value is strictly less than I(u0 ),
which contradicts the fact that u0 is a minimum.
Let
Λ(x, ũ, p̃) = L(x, ũ, U (ũ), p̃, pN ),
since u0 is the minimum of I under the constraint M (u) = 0, ũ0 must be a
minimum of
Z
J(ũ) = Λ(x, ũ(x), ∇ũ(x))dx.
Br (x0 )
(7.4)
However,
N −1
∂pN
α
X ∂2U j ∂Uui
= u =
i ∂uj α
. (7.5)
∂ui j=1
∂u ∂xα
α=1
∂xα ∂xα α pα
∂xα u
n
∂pN
X ∂ ∂ α
= L i + Uui L N + LpN .
α=1
∂xα pα ∂xα pα α
∂ui
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 97
Noting
M (ũ0 , U (ũ0 )) = 0,
upon differentiating, we obtain
Mui + MuN Uui = 0,
i.e.
Mui
Uui = − . (7.7)
MuN
We now define on Br (x0 ) the function
1
λ= (div LpN − LuN ). (7.8)
MuN
Substituting (7.7) and (7.8) into (7.6), it follows that
Lui − div Lpi + λMui = 0, i = 1, . . . , N − 1. (7.9)
Combining (7.8) and (7.9) gives the local E-L equation (7.3) on Br (x0 ).
3. Since Ω̄ is compact, we have a finite open sub-covering of the covering
by {Br (x0 ) | x0 ∈ Ω, r = r(x0 )}, together with the already defined functions
λ|Br (x0 ) . Suppose on Br1 (x1 ), MuN 6= 0 and on Br2 (x2 ), Muj 6= 0, then we
have λ|Br2 (x2 ) = M1 j (div Lpj − Luj ) and λ|Br1 (x1 ) is the λ in (7.8). Thus, when
u
any two such small balls have a nonempty intersection (Br1 (x1 ) ∩ Br2 (x2 ) 6= ∅),
by (7.9), it is immediate that
λ|Br1 (x1 ) = λ|Br2 (x2 ) .
We now glue these λ|Br (x0 ) together to create a globally continuous function λ as
desired.
Remark 7.3 Just like the isoperimetric problem, we can also consider multiple
pointwise constraints. As long as the constraints M1 (u) = M2 (u) = · · · =
Ms (u) = 0 satisfy
∂Mi (u(x))
rank = s, ∀ x ∈ Ω.
∂uj
Now the adjusted Lagrangian is
Xs
Q=L+ λi Mi ,
i=1
where λi ∈ C(Ω), i = 1, . . . , s.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 98
Remark 7.4 If the constraint functions also depend on other variables, then
we call them non-holonomic constraints. As to variational problems with non-
holonomic constraints, is there still a Lagrange multiplier method? The answer to
this question is far more complicated; we refer the interested readers to the book
by Giaquinta and Hidelbrandt [GH]. However, some special cases are also known
to be optimal control problems, which we will address in the third part of this
book.
Example 7.3 (Revisit of geodesics on spheres) In our earlier discussions, we
have studied the geodesics problems from an unconstrained extreme value prob-
lem point of view. We now provide a different viewpoint. Regard the sphere as
the constraint
x2 + y 2 + z 2 = 1.
Find the extrema of the functional
Z b p
I(u) = ẋ2 + ẏ 2 + ż 2 dt
a
given this pointwise constraint.
Let u = (x, y, z) and p = (ξ, η, ζ). Introducing the Lagrange multiplier λ(t),
we have the adjusted Lagrangian
p
Q = ξ 2 + η 2 + ζ 2 + λ(x2 + y 2 + z 2 ).
Its E-L equation is
d u̇
= 2λu;
dt |u̇|
furthermore, it also satisfies |u| = 1. Since
d u̇ d u̇ u̇ du u̇
×u = ×u+ × = 2λu × u + × u̇ = 0,
dt |u̇| dt |u̇| |u̇| dt |u̇|
v = |u̇u̇| × u must be a constant vector. Consequently, u ⊥ v, i.e. the geodesic u
must lie in a plane orthogonal to the constant vector v, hence this curve must be a
piece of a great circle.
Example 7.4 On the slant plane x = y, find the equation of the trajectory of a
moving particle of unit mass by gravity. We denote the coordinate of the particle
by (x, y). The Lagrangian is
1 2
L= (p + q 2 ) − gy,
2
where g is the gravitational constant and the constraint is
M (x, y) = y − x = 0.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 99
∂Q d
−g + λ =
= Qq = ÿ.
∂y dt
Since y = x, λ = g/2, from which we solve to get
1
x = y = − λt2 + ẋ(0)t + x(0),
2
where λ = g/2.
Example 7.5 (Harmonic mappings to spheres) Let Ω be the unit ball in R3 , S 2
be the unit sphere in R3 , and u = (u1 , u2 , u3 ) : Ω → S 2 . If u is a solution of the
constrained extreme value problem
min{I(u) | M (u) = 0},
where
Z Z X 3 X 3
2
I(u) = |∇u| dx = |∂α ui |2 dx,
Ω Ω i=1 α=1
2
M (u) = |u| − 1 = u21 + u22 + u23
− 1,
then we call u a harmonic mapping from the solid unit ball to the unit sphere. Find
the differential equation for which such a harmonic mapping satisfies.
Solution The E-L equation of the adjusted Lagrangian is
−∆ui = λui ,
where ∆ is the Laplace operator and the Lagrange multiplier λ is a continuous
function. From
u21 + u22 + u23 = 1,
by differentiating, it yields
hu, ∇ui = 0,
where h·, ·, i denotes the standard inner product on R3 . Differentiating once more,
it yields
hu, ∆ui + |∇u|2 = 0.
Multiplying by u on both sides of the above E-L equation, it is immediate
λ = −h∆u, ui = |∇u|2 ,
i.e.
−∆u = u|∇u|2 .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 100
i.e. ∀ v ∈ C,
Z
[Lu (x, u(x), ∇u(x))(v(x) − u(x))
Ω
+ Lp (x, u(x), ∇u(x)) · (∇v(x) − ∇u(x))]dx ≥ 0.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 101
Note the only difference between the E-L equation and this expression is that the
former is an equality, whereas the latter is an inequality. This is why we shall call
it a variational inequality.
Returning to the obstacle problem, C = {u ∈ P W C 1 (Ω̄) | u(x) ≤
ψ(x), ∀ x ∈ Ω̄, u|∂Ω = ϕ} is a convex set, so the resulting variational inequality
is
Z
[∇u∇(v − u) − f (v − u)]dx ≥ 0, ∀ v ∈ C.
Ω
Exercises
1. Find
where
Z 1 Z 1
I(u) = u̇2 (t)dt, N (u) = u(t)dt.
0 0
where
Z b Z b
1
I(u) = u(t)dt, N (u) = (1 + u̇(t)2 ) 2 dt.
0 0
Determine the equation of the surface S such that D(X) achieves its minimum
under the constraint V (X) = V0 .
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 103
Lecture 8
Fig. 8.1
103
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 104
Suppose
∂
ηε |ε=0 = X̄(x).
∂ε
Then the deformation
y = ηε (x) = x + εX̄(x) + o(ε).
Moreover, we define v : Rn × (−ε0 , ε0 ) → RN satisfying v(x, 0) = u(x). This
induces a family of functions vε = v(·, ) : Ωε → RN , (y, ε) ∈ Ωε × (−ε0 , ε0 ).
Suppose
∂
vε (x) = ϕ(x).
∂ε ε=0
In our previously discussed variational problems, in the same domain of the func-
tional, all functions are defined over the same underlying region Ω. However, this
requirement is unnecessary, we can indeed allow the functions in M to take on
different domain Ω. To emphasize the functional dependence on the region, we
shall denote
Z
I(u, Ω) = L(x, u(x), ∇u(x))dx.
Ω
i.e.
Z
Φ0 (0) = [EL (u)ϕ + div(LX̄ + Lpi ϕi )]dx. (8.1)
Ω
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 105
Thus,
∂
X(x, u) = ∂ε Y (x, u, ε) ,
ε=0
U (x, u) = ∂ W (x, u, ε)
.
∂ε ε=0
For any u ∈ C 1 (Ω̄, RN ), we want to convert it into the special 1-parameter family
of functions as given above. Let
η(x, ε) = Y (x, u(x), ε),
ω(x, ε) = W (x, u(x), ε),
then
η(x, 0) = x,
ω(x, 0) = u(x).
We also let
∂η(x, ε)
X̄(x) = = X(x, u(x)),
∂ε ε=0
(8.3)
∂ω(x, ε)
Ū (x) = = U (x, u(x)).
∂ε ε=0
That said, for a given u ∈ C 2 (Ω, RN ), we can turn the locally 1-parameter
transformation group φε , whose generating vector field is (8.2), into a special
1-parameter family of functions. Note X̄ and ϕ are determined by (8.3) and
(8.4). Substituting into (8.1), we obtain the Noether’s identity of a general lo-
cal 1-parameter transformation group.
It is worth noting the representation of I(u, Ω) under the transformation {φε }
is Z
I(vε , Ωε ) = L(y, vε (y), ∇vε (y))dy.
Ωε
Definition 8.1 ∀ Ω0 ⊂ Ω̄0 ⊂ Ω. Let (Ω0 )ε = η(Ω0 , ε). If
I(vε , (Ω0 )ε ) = const. (independent of ε), ∀ u ∈ C 1 (Ω̄0 , RN ),
then I is invariant under {φε }.
In fact, if at each point, the following holds:
L(ηε (x), vε (ηε (x)), ∇y vε (ηε (x)))det(ηε (x)) = L(x, u(x), ∇u(x)),
then I is invariant under {φε }. In summary, we arrive at
Theorem 8.1 (Noether) Suppose the local 1-parameter transformation group
{φε } is generated by the vector field (8.2). Let
Z
I(u) = L(x, u(x), ∇u(x))dx.
Ω
Then ∀ u ∈ C 2 (Ω, RN ), the Noether identity (8.1) holds, where
X̄(x) = X(x, u(x)),
Ū (x) = U (x, u(x)),
n
i i
X ∂ui α
ϕ (x) = Ū (x) − X̄ (x).
∂xα
α=1
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 107
is closed, i.e.
dν = 0.
Corollary 8.1 When n = 1, let {φε } be a local 1-parameter transformation group
on RN , whose generating vector field is U (this means X = 0). If I is invariant
under {φε } and u ∈ C 2 is a weak minimum of I, then
N
X
Ū i (u(t))Lpi (t, u(t), u̇(t)) = const.
i=1
Corollary 8.2 When n = 1, if L is autonomous (i.e. L is independent of t), then
for the solution u of the E-L equation of I, we have
XN
i
L− Lpi p = const.
i=1 (u,p)=(u(t),u̇(t))
Proof In fact, for the local 1-parameter transformation group {φε }, its generating
vector fields are X = 1, U = 0, whence ϕ = du dt .
In particular,
1
T00 = (p20 + p21 + p22 + p23 + M 2 u2 ), Tβ0 = −p0 pβ , β = 1, 2, 3.
2
Using the energy–momentum tensor, we can rewrite Noether’s theorem as follows.
Theorem 8.2 Let L ∈ C 2 (Ω × RN × RnN , R1 ). Suppose
Z
I(u) = L(x, u(x), ∇u(x))dx
Ω
is invariant under local 1-parameter transformation group {φε }. If u ∈
C 2 (Ω̄, RN ) is a weak minimum of I, then
Xn n
X Xn
Lpiα Ū i − Tβα X̄ β = 0.
α=1 i=1 β=1
xα
Or simply,
div Lpi Ū i − T · X̄ = 0.
l
X l
X
Pi = mi Ẋi (t) = const.,
i=1 i=1
X = 0, U = (Z1 , Z2 , . . . , Zl ),
Likewise, for rotations in the yz-plane and the xz-plane, we also have similar
identities. This is the conservation of angular momentum:
l
X
mi Xi ∧ Ẋi = const.
i=1
X̄ β = (δγβ )0≤γ≤3 .
where
X
Tγα = gαβ ∂β u∂γ u − δαγ L(x, u, ∇u).
β
Or simply,
div Tγ = 0, γ = 0, 1, 2, 3,
where Tγ = (Tγα )0≤α≤3 .
Choose any [t1 , t2 ] × BR (θ) ⊂ R4 as our domain of integration, then
Z Z Z
x̄
Tγ0 (t2 , x̄)dx̄− Tγ0 (t1 , x̄)dx̄+ Tγ0 (t, x̄)· dtdσ = 0,
BR (θ) BR (θ) [t1 ,t2 ]×∂BR (θ) |x̄|
where dσ is the area element of the 2-sphere and |x̄| is the norm of x̄. If as
R → ∞, Tγ0 (t, x), |x| = R, tend to zero uniformly, and Tγ0 (t, ·) is integrable on
R3 , then
Z
Pγ (t) = Tγ0 (t, x̄)dx̄ = const.
R3
In particular, take γ = 0, then
Z
P0 (t) = T00 (t, x̄)dx̄
R3
Z
= [Lp0 ux0 − L](t, x̄)dx̄
3
ZR
1
= (|∂t u|2 + |∇u|2 + M 2 u2 )dx = const., ∀ t ∈ R1 .
R3 2
This shows the conservation of energy. Next, take γ = 1, 2, 3, then
Z
Pγ (t) = Tγ0 (t, x̄)dx̄
R3
Z
= [Lp0 uxγ ](t, x̄)dx̄
R3
Z
=− ∂t u∂γ u = const., ∀ t ∈ R1 .
R3
This shows the conservation of momentum.
The positive Lorentz transformation group includes the following six rota-
tional generators:
εµν = −ενµ , 0 ≤ µ ≤ ν ≤ 3.
Consider the transformation
3
X
yµ = xµ + gνν µν xν ,
0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 112
Hence,
Z
Mαβ,0 (t)dx̄ = const.,
R3
v (y) = u(x),
i.e.
v = u ◦ ξ = u ◦ η−1 (y).
Hence,
Z
I(v , Ω) = L(y, v (y), ∇v (y))dy
Ω
Z
∂ξ ∂η
= L η (x), u(x), ∇u(x) det dx.
Ω ∂y ∂x
Noting
−1
∂ξ ∂η ∂X
= =I − + o(),
∂y ∂x ∂x
so for u ∈ C 2 , we have
n n X N
∂ui ∂ X̄ α ∂ X̄ α
Z X
d X
I(vε , Ω)|ε=0 = − Lpiβ + Lxα X̄ α + L dx
dε Ω α=1 ∂xα ∂xβ ∂xα
β=1 i=1
Z X n X n
= Lxα − ∂xα (L(x, u(x), ∇u(x))) + ∂xβ Lpiβ uxα X̄ α dx
i
Ω α=1 β=1
Z n
X ∂u α
= EL (u) · − X̄ dx. (8.5)
Ω α=1
∂xα
∂u
EL (u) = 0.
∂xα
8.4∗ Applications
Fig. 8.2
Let vε : Ωε → R1 ,
vε (y) = u((1 + ε)−1 y),
then X̄ = x and
d
u(1 + ε)x)|ε=0 = −x · ∇u.
ϕ(x) =
dε
By Noether’s theorem, we have
d
I(vε , Ωε )|ε=0
dε Z
= div(LX̄ + Lp ϕ)dx
ZΩ
1
= div |∇u|2 − G(u) x − ∇u(x · ∇u) dx. (8.11)
Ω 2
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 117
Z 2
∂u
(x · ν)dσ = 0.
∂Ω ∂ν
Exercises
Prove:
(1)
Z 1
1
I(u) = t2 u̇2 − u6 dt
0 3
is invariant under {ϕε }.
(2) If u is a solution of the E-L equation of I, then
t3 6
u + t3 u̇2 + t2 uu̇ = const.
3
2. Let L = (p + ku)2 , where k is a constant. Let {ϕε } be defined such that
Y (t, u, ε) = t + ε, W (t, u, ε) = u + εαe−kt , α ∈ R1 .
R1
(1) Verify I(u) = 0 (u̇ + ku)2 dt is invariant under {ϕε }.
(2) If u is a solution of the E-L equation of I, find the conservation law for
which u satisfies.
(3) Solve for u by means of Hamiltonian system.
(4) For this u, verify the conclusion of (2).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 119
Lecture 9
Direct methods
Before the 20th century, calculus of variations was largely based on the studies
of E-L equations. Just like in mathematical analysis, finding the extrema of a func-
tion is usually turned into a problem of solving equation of critical points, finding
the extrema of a functional can also be turned into solving the corresponding E-L
equation.
E-L equations are differential equations. When n = 1, these are ordinary
differential equations (or systems of ordinary differential equations); only under
some special circumstances, one can find their analytic solutions. When n > 1,
E-L equations are partial differential equations, the circumstances for which one
will be able to find analytic solutions are extremely rare.
In the 19th century, driven by the studies of electromagnetism and complex
variables, people were seeking solutions of the Laplace equation
4u = 0
in Ω, (9.1)
u|∂Ω = ϕ,
and the Poisson equation
4u = f
in Ω. (9.2)
u|∂Ω = ϕ.
The Riemann (conformal) mapping theorem is a famous example: for a non-
empty simply connected open domain Ω ⊂ C, there exists a biholomorphic map-
ping f (i.e. a bijective holomorphic mapping whose inverse is also holomorphic)
from Ω onto the open unit disk. The fact f is biholomorphic implies that it is
conformal. To establish the existence of the conformal mapping, Riemann turned
this into a boundary value problem of the harmonic equation (9.1). He noticed
119
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 120
that (9.1) is the E-L equation of the Dirichlet integral (regarded as a functional)
Z
D(u) = |∇u(x)|2 dx (9.3)
Ω
1
on the set M = {u ∈ C (Ω̄) | u|∂Ω = ϕ}, the solution of (9.1) can thus be
obtained by finding the minimum of D.
This idea opened up a new path in solving partial differential equations. If
a differential equation is the E-L equation of some functional, then solving the
partial differential equation can be turned into finding the extrema of the corre-
sponding functional.
In order to prove the existence of solutions to the boundary value problem of
the harmonic equation (9.1), it is led to prove the Dirichlet integral achieves its
minimum on M . However, why does such minimum exist? Riemann’s argument
was based on the following Dirichlet’s principle, a widely accepted statement back
in the mid-19th century.
Dirichlet’s principle Since D is bounded below, it “must” attain its infimum.
That is, “there exists” u such that D achieves its minimum at u.
“Rationale”: Choose a sequence {un } ⊂ M such that D(un ) →
inf u∈M D(u). Since {un } is bounded, there exists a convergent subsequence
unk → u0 . This u0 is then the desired solution D(u0 ) = minu∈M D(u).
This argument is clearly flawed in a modern reader’s eyes. Shortly after pub-
lishing the Riemann mapping theorem, a heated debate began about the validity of
the “Dirichlet’s principle”. In 1870, Weierstrass constructed the following coun-
terexample. Consider the following extreme-value problem:
Z 1
I(u) = x2 u02 dx, M = {u ∈ C 1 [−1, 1] | u(−1) = −1, u(1) = 1}.
−1
(1) inf I = 0. In fact, I ≥ 0. Letting
u∈M
arctan x
u = , ∀ > 0,
arctan 1
it follows that
1
2 (x2 + 2 )−1
Z
2
I(u ) < 1 2 dx = → 0, as → 0.
−1 (arctan ) arctan 1
(2) If u0 is a minimum of I on M , then u00 ≡ 0, so u0 = const. This contradicts
the boundary condition!
Any reader with some rigorous background in mathematical analysis under-
stands: it is in general not true that every minimizing sequence contains a subse-
quence converging to the minimal value.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 121
xni → xi , i = 1, 2, 3, . . . .
which implies
∞
X
|ξi0 |2 ≤ M 2 + 1.
i=1
We have thus established the diagonal subsequence xnkk “coordinate-wise con-
verges to” x0 = (ξ10 , ξ20 , . . . , ξk0 , . . .) ∈ l2 .
Weak convergence and weak-∗ convergence both stem from the idea of
“coordinate-wise convergence”.
Definition 9.1 Let X be a normed linear space, the sequence {xn } ⊂ X is
said to converge weakly to x, denoted xn * x, if for any x∗ ∈ X ∗ , we have
hx∗ , xn − xi → 0, where X ∗ is the dual space of X.
Let X ∗ be the dual space of a normed linear space X, a sequence {x∗n } ⊂ X ∗
is said to converge to x∗ in the weak-∗ -topology, denoted x∗n *∗ x∗ , if for any
x ∈ X, we have hx∗n − x∗ , xi → 0.
Remark 9.1 In fact, on X ∗ , we have both the notion of weak convergence and
the notion of weak-∗ convergence. By weak convergence x∗n * x∗ , we mean for
any x∗∗ ∈ X ∗∗ , hx∗∗ , x∗n − x∗ i → 0; by weak-∗ convergence, we mean for any
x ∈ X, hx∗n − x∗ , xi → 0. Since we have the continuous embedding X ,→ X ∗∗ ,
weak convergence implies weak-∗ convergence.
It is evident that norm convergence implies both weak and weak-∗ conver-
gence, but not vice versa.
Example 9.1 In L2 (−∞, ∞), choose any nonzero function ϕ(t) with compact
support, let ϕn (t) = ϕ(t + n), then ϕn * 0, but kϕn k = kϕk 6= 0.
Example 9.2 Note that L2 [0, 2π] is self-dual, its dual space is again L2 [0, 2π].
Consider the sequence {sin (nt)} ⊂ L2 [0, 2π]. According to the Riemann–
Lebesgue lemma, ∀ f ∈ L1 [0, 2π],
Z 2π
f (t) sin (nt)dt → 0,
0
from which we may conclude:
(1) Take X = Lp [0, 2π], 1 ≤ p < ∞ and regard the sequence {sin (nt)} ⊂
0
Lp [0, 2π] = (Lp [0, 2π])∗ , then
sin (nt) *∗ 0, as n → ∞.
(2) Take X = Lp [0, 2π], 1 ≤ p < ∞ and regard the sequence {sin (nt)} ⊂
0
L∞ [0, 2π] ⊂ X, X ∗ = Lp [0, 2π] ⊂ L1 [0, 2π], 1 < p0 ≤ ∞, then
sin (nt) * 0, as n → ∞.
For 1 < p < ∞, Lp [0, 2π] is reflexive, weak and weak-∗ convergence coincide.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 124
then
ϕn * ϕ̄ in Lp (D), 1 ≤ p < ∞,
and
ϕn *∗ ϕ̄ in L∞ (D).
The proof for p = 1 can be modified from the above argument, we omit the proof,
leaving as an exercise.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 125
Consequently, we have the following fundamental result.
Theorem 9.2 Let X be the dual space of a separable Banach space (e.g. a re-
flexive Banach space). Let E ⊂ X be a non-empty weak-∗ sequentially closed
subset. If f : E → R1 is sequentially weak-∗ lower semi-continuous (abbreviated
s.w∗ . l.s.c), and if f is coercive (∀ x ∈ E, when kxk → ∞, f (x) → +∞), then
f attains its minimum on E.
Proof Choose a minimizing sequence {xn } ⊂ E of f ,
lim f (xn ) = inf f (x).
x∈E
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 126
xnk *∗ x0 .
Thus,
We now return to the Dirichlet integral. On the one hand, from the bounded-
ness of the Dirichlet integral D(u), it is impossible to deduce its boundedness in
the C 1 -norm; on the other hand, we do not know whether C 1 (Ω̄) is the dual space
of some normed linear space. Because of this, in order to verify the validity of the
Dirichlet’s Principle, the space C 1 (Ω̄) is not a proper choice.
Closely related to the Dirichlet integral D(u) is the following semi-norm:
Z 21
2
k u k= |∇u| dx ,
Ω
it implies
1
D(u) ≥ D(v) − D(v0 ).
2
This confirms coerciveness.
2. D(u) is sequentially weakly semi-lower continuous.
Let uj = v0 + vj , then
uj * u (H 1 (Ω)) ⇔ vj * v (H01 (Ω)).
Notice
1 1
|D(v0 , ϕ)| ≤ D(v0 ) 2 D(ϕ) 2 ∀ ϕ ∈ H01 (Ω),
ϕ 7→ D(v0 , ϕ) is a continuous linear functional on H01 (Ω). Thus,
D(v0 , vj ) → D(v0 , v).
Likewise,
D(v, vj ) → D(v, v).
By Schwarz’s inequality, we obtain
1 1
D(v) = lim D(v, vj ) ≤ lim inf D(v) 2 D(vj ) 2 ,
i.e.
D(v) ≤ lim inf D(vj ).
That is,
D(u) ≤ lim inf D(uj ).
H01 (Ω) is a Hilbert space, hence is self-dual.
We can also directly verify H01 (Ω) is self-dual, H01 (Ω) is the dual space of the
Banach space H01 (Ω). Furthermore, H01 (Ω) is separable (the detailed proof will
be given in the next lecture). We now apply Theorem 9.2 to deduce the Dirichlet
integral indeed attains its minimum on H01 (Ω), hence affirms Dirichlet’s principle.
Remark 9.2 To verify Dirichlet’s Principle, there is also a more direct method -
orthogonal projection. Geometrically speaking, it is equivalent to minimizing the
distance from a hyperplane to a given point outside the hyperplane (we refer to
Lecture 12). From which, we will derive the Riesz Representation Theorem and
the self-duality of a Hilbert space.
Remark 9.3 In the above example, we obtain the solution u0 ∈ H 1 (Ω). However,
we do not yet know if it is differentiable, nor do we know if it belongs to C 2 . In
our on-going discussions, we must address in what sense will u0 be a solution of
the harmonic equation.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 129
Theorem 9.3 (Banach) Let X be a normed linear space. If the dual space X ∗ is
separable, then so is X itself.
Proof 1. Denote S1∗ the unit sphere in X ∗ , then S∗1∗ is separable. In fact, let
x
{x∗n } ⊂ X ∗ be a countable dense subset. Let yn∗ = kxn∗ k , then {yn∗ } is a countable
n
∗ ∗ ∗
dense subset of S1 . To see this, ∀ x ∈ S1 , ∃ xnj such that kx∗nj − x∗ k → 0,
∗
hx∗∗ ∗ ∗
0 , x0 i = hx0 , xi, ∀ x∗0 ∈ X0∗ . (9.6)
∗
1. Define the mapping T : X → X0∗ ∗ ∗
by T x = x |X0 ,
∗ ∗
hT x , x0 i = hx |X0 , x0 i ∀ x0 ∈ X0 .
T is linear and continuous. Its dual mapping T ∗ ∈ L(X0∗∗ , X ∗∗ ) satisfying
∀ x∗∗ ∗∗ ∗ ∗∗
0 ∈ X0 , T x0 ∈ X .
∗∗
But
hx∗1 , xi = hx∗∗ ∗
0 , x1 |X0 i = 0,
which is a contradiction.
3. We now prove (9.6), i.e. x∗∗ 0 is the image of x ∈ X0 under the natural
mapping. According to the Hanh–Banach theorem, ∀ x∗0 ∈ X0∗ , ∃ x∗ ∈ X ∗ such
that x∗ |X0 = x∗0 . Noting (9.7) implies
hx∗∗ ∗ ∗ ∗
0 , x0 i = hx , xi = hx0 , xi ∀ x∗0 ∈ X0∗ .
and
Z 10
p
p0
kF k(Lp )∗ = |v(x)| dx = kvkp0 .
Ω
p
For all u ∈ L (Ω), 1 < p < ∞,
Z
v 7→ u(x)v(x)dx
Ω
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 132
0
can be viewed as a continuous linear functional G on Lp (Ω). Likewise, there
exists wu ∈ Lp (Ω) such that
Z
0
G(v) = wu (x)v(x)dx ∀ v ∈ Lp (Ω),
Ω
i.e.
Z Z
0
u(x)v(x)dx = G(v) = wu (x)v(x)dx ∀ v ∈ Lp (Ω).
Ω Ω
Exercises
Lecture 10
Sobolev spaces
We have pointed out in our earlier discussions that neither C 1 nor C01 is an ap-
propriate space for verifying the Dirichlet’s principle, instead, one should consider
the spaces H 1 and H01 .
Such scenario frequently occurs when applying direct methods to solve vari-
ational problems. This is because the functionals are usually variational integrals
involving derivatives, and the C-norm associated with the C-space consisting of
the same order derivatives is determined by the maximal value of the pointwise
norm. However, the C-norm cannot be controlled by such variational integrals.
Moreover, in order to possess the weakly sequential compactness, the underly-
ing space must be the dual space of a normed linear space, since such space is at
least complete. The Sobolev spaces introduced in this lecture satisfy the above
requirements.
133
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 134
Definition 10.2 (The space W m,p (Ω)) Suppose p ∈ [1, ∞] and m ∈ N, let
W m,p (Ω) := {u ∈ Lp (Ω) | Dα u ∈ Lp (Ω), ∀ α, |α| ≤ m},
on which, we define the norm by
X Z p1
α p
kukm,p = |D u(x)| dx , 1 ≤ p < ∞,
|α|≤m Ω
X
kukm,∞ = esssupx∈Ω |Dα u(x)|.
|α|≤m
It follows that
hDα uj , ϕi → hgα , ϕi, ∀ α, |α| ≤ m
and
huj , ∂ α ϕi → hg0 , ∂ α ϕi, ∀ α, |α| ≤ m.
That is,
hgα , ϕi = (−1)|α| hg0 , ∂ α ϕi.
From which, we deduce that
gα = Dα g0
and
k Dα uj − gα kp → 0 ∀ α, |α| ≤ m.
m,p
Thus, g0 ∈ W (Ω) and
kuj − g0 km,p → 0, j → ∞.
It is worth mentioning that although the generalized derivatives are dually and
globally defined, they are nevertheless closely related to the locally defined ordi-
nary derivatives. In particular, for functions in the one dimensional Sobolev space,
their generalized derivatives are the original derivatives almost everywhere! As-
sume n = 1 and J = [a, b], we have:
Example 10.3 W 1,1 (J) = AC(J), the space of absolutely continuous functions
defined on J and
Du(x) = u0 (x), a.e. (10.1)
1,1
Proof ∀ u ∈ W (J), we show that
Z x
u(x) − u(a) = Du(t)dt, ∀ x ∈ J.
a
∀ n ∈ N, let
n(t − a) t ∈ [a, a + n1 ],
t ∈ [a + n1 , x − n1 ],
1
ϕn (t) =
−n(x − t) t ∈ [x − n1 , x]
0 t ∈ [x, b].
Since ∃ ξn,k ∈ C0∞ (J), kξn,k kC 1 ≤ 2n such that |ϕn (t) − ξn,k (t)| converges to
R0 uniformly on J andRthat |ϕ0n (t) − ξn,k
0
(t)| → 0 a.e. t ∈ J as k → ∞, from
0
J
u(t)ξn,k (t)dt = − J
Du(t)ξ n,k dt, it follows that
Z Z
u(t)ϕ0n (t)dt = − Du(t)ϕn (t)dt.
J J
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 136
This is (10.1).
Example 10.4 W 1,∞ (J) = Lip(J), the space of Lipschitz functions on J.
Proof “⊃” Assume u ∈ Lip(J), then u is absolutely continuous
Ry 0 on J. Hence, u0
exists almost everywhere and satisfies u(y) − u(x) = x u (t)dt as well as
|u(y) − u(x)|
|u0 (x)| ≤ sup ≤ M.
y∈J |y − x|
From Example 10.3, we see that Du ∈ L∞ (J), i.e. u ∈ W 1,∞ (J) and
kuk1,∞ ≤ kukLip .
Thus,
Definition 10.3 We denote the closure of C0∞ (Ω) in W m,p (Ω) by W0m,p (Ω).
Lemma 10.1 If u ∈ W m,p (Ω), ψ ∈ C0∞ (Ω), then (ψu) ∈ W0m,p (Ω) and
Proof We only prove for m = 1, the rest can be proved by mathematical induc-
tion. In fact, ∀ ϕ ∈ C0∞ (Ω),
Z Z
Dxi (ψu)ϕdx = − ψu∂xi ϕdx
Ω Ω
Z
=− u[∂xi (ψϕ) − ∂xi ψϕ]dx
Z Ω
= [∂xi ψu + ψDxi u]ϕdx.
Ω
Hence,
Dxi (ψu) = ψ(Dxi u) + (∂xi ψ)u.
In order to understand the weak topology on the Banach space W m,p (Ω) (or
W0m,p (Ω)), we need to consider the representations of functionals on these spaces.
We know the dual space of Lp (J) is
0 1 1
(Lp (Ω))∗ = Lp (Ω), 1 ≤ p < ∞, + = 1,
p p0
0
i.e. ∀ f ∈ (Lp (Ω))∗ , ∃ v ∈ Lp (Ω) such that
Z
hf, ui = u(x)v(x)dx ∀ u ∈ Lp (Ω)
Ω
and
Z 10
p
p0
kf k = sup hf, ui = |v| .
kukp ≤1 Ω
space:
i: u 7→ {Dα u, |α| ≤ m},
Z X
α
hf, ui = D u(x)ψα (x) dx.
Ω |α|≤m
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 138
subspace. Since |α|≤m Lp (Ω) is reflexive, by the Pettis theorem, W m,p (Ω) is
Q
also reflexive.
10.4 Modifiers
(3) If u ∈ C0m (Ω), then ∂ α (u ) = (∂ α u) , ∀ α |α| ≤ m for > 0 sufficiently
small.
(4) If u ∈ C0 (Ω), then ku − u kC → 0 as → 0.
(5) If u ∈ Lp (Ω), 1 ≤ p < ∞, then ku − u kLp → 0 as → 0.
(6) C0∞ (Ω) is dense in Lp (Ω).
(7) If u, ∂ α u ∈ Lp (Ω), p ∈ [1, ∞], supp u ⊂ int(Ω), then for sufficiently
small, we have (Dα u) = ∂ α (u) .
In fact,
Z
LHS = Dα u(y)ϕ (x − y)dy
Ω
Z
= (−1)|α| u(y)∂yα ϕ (x − y)dy
Ω
Z
= u(y)∂xα ϕ (x − y)dy
Ω
Z
= ∂xα u(y)ϕ (x − y)dy = RHS.
Ω
(8)
Sobolev spaces are fundamental function spaces, which play an important role
in Harmonic Analysis, Partial Differential Equations, Functional Analysis, and
Calculus of Variations. The important properties of Sobolev spaces are discussed
in many textbooks. In this section, we only list some of the main results and refer
the interested readers for the detailed proofs in the existing literature. However,
in order to assist the readers’ understanding of the significance of these results
as well as the essence of the proofs, we would like to provide proofs for some
particular or simplified cases.
Extension Theorem
Sobolev spaces are function spaces. When the domain of these functions is
an arbitrary region Ω ⊂ Rn , the space is denoted W m,p (Ω); when the domain
of these functions is the whole Rn , it is denoted W m,p (Rn ). A natural question
comes to mind: can we extend each function u ∈ W m,p (Ω) to be a function
ũ ∈ W m,p (Rn ) such that ũ|Ω = u?
The answer is affirmative as long as Ω ⊂ Rn has sufficiently smooth boundary.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 140
ally, W 1,2 (Ω) is a complete metric space containing C 1 (Ω̄), whence H 1 (Ω) ⊂
W 1,2 (Ω). It is natural to ask whether they are the same. We give a positive an-
swer to this question via the following approximation theorem.
Theorem 10.4 (Serrin-Meyers Approximation Theorem) If p ∈ [1, ∞), then
the set S := C ∞ (Ω) ∩ W m,p (Ω) is dense in W m,p (Ω),
Proof Choose a sequence of open subsets {Ωj } satisfying
∅ = Ω−1 = Ω0 ⊂ Ω1 ⊂ Ω̄1 ⊂ Ω2 ⊂ Ω̄2 ⊂ · · · Ωi ⊂ Ω̄i ⊂ Ωi+1 ⊂ · · · ,
∞
[
Ωi = Ω.
i=1
This is possible due to properties (7), (5) in §10.4 and Lemma 10.1. ∀ α, |α| ≤ m,
we have
Poincaré’s Inequality
Poincaré’s inequality in Lecture 9 can be extended to W01,p (Ω), 1 ≤ p < ∞
as follows.
Poincaré’s Inequality If Ω ⊂ Rn is bounded, u ∈ W 1,p (Ω), 1 ≤ p < ∞, then
∃ C = C(p, Ω) such that
Z Z
p
|u| dx ≤ C |∇u|p dx.
Ω Ω
Corollary 10.3 If Ω ⊂ Rn is bounded, then
Z p1
kuk = |∇u|p
Ω
defines an equivalent norm on W01,p (Ω) (1 ≤ p < ∞).
Since W01,p (Ω) (1 < p < ∞) is a closed linear subspace of W 1,p (Ω), it is
itself a reflexive Banach space.
Embedding Theorems
Theorem 10.5 (Sobolev) Both embeddings
1 1 m
W m,q (Rn ) ,→ Lr (Rn ), = − (if mq < n),
r q n
and ∀ j ∈ N,
n
W m+j,q (Rn ) ,→ C j,λ (Rn ), 0 < λ ≤ m − (if mq > n)
q
are continuous.
A proof can be found in [Ad] 5.4–5.10.
Combining the above theorem and the extension theorem, we arrive at
Theorem 10.6 (Sobolev Embedding Theorem) Assume Ω is a bounded region
with uniformly C m boundary, 1 ≤ q < ∞, and m ≥ 0 is an integer, then the
embeddings
1 1 m
W m,q (Ω) ,→ Lr (Ω), = − (if mq < n)
r q n
and ∀ j ∈ N,
n
W m+j,q (Ω) ,→ C j,λ (Ω̄), 0 < λ ≤ m − (if mq > n)
q
are both continuous.
The most frequently used version of this result is for m = 1, and we denote
nq
q ∗ = n−q , then
W 1,q (Ω) ,→ Lr (Ω), r ≤ q ∗ if n > q
and
W 1,q (Ω) ,→ C(Ω̄) if q > n
are both continuous.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 143
Remark 10.1 When n = 1, Ω = (a, b), the conclusion of the embedding theorem
follows easily from Hölder’s inequality and Example 10.3. Since in this case,
the generalized derivatives coincide with the usual derivative functions almost
everywhere, we have
Z x Z b q1
1
u0 (t)dt ≤ |x−y| q0 |u0 (t)|q dt ,
|u(x)−u(y)| = ∀ x, y ∈ (a, b).
y a
Compact Embeddings
Theorem 10.7 (Rellich–Kondrachov compact embedding theorem) Assume Ω
is a bounded region with uniformly C m boundary, 1 ≤ q ≤ ∞, and m ≥ 0 is an
integer, then the embeddings
nq
W m,q (Ω) ,→ Lr (Ω), 1 ≤ r < (if m < n/q)
n − mq
and
W m,q (Ω) ,→ C(Ω̄), (if m > n/q)
are both compact.
The proof can be found in [Ad] Theorem 6.2.
We now give a direct proof for a frequently used special case of the above
result.
Theorem 10.8 (Rellich) If Ω ⊂ Rn is a bounded region, 1 ≤ p < ∞, then a
closed bounded ball in W01,p (Ω) is sequentially compact in Lp (Ω).
Proof It suffices to show the closed unit ball is sequentially compact. Denote B
the closed unit ball in W01,p (Ω) centered at 0. We proceed by showing ∀ > 0,
under the Lp -norm, there exists a finite net of B.
The idea is to find a uniformly bounded and equicontinuous set of functions
in an arbitrary Lp -neighborhood of B.
1. By definition, C0∞ (Ω) is dense in W01,p (Ω). Denote S = C0∞ (Ω) ∩ B. For
any δ > 0, denote Sδ = {vδ | v ∈ S}, where
Z
vδ (x) = v(y)ϕδ (x − y)dy.
Ω
∀ v ∈ S,
kvδ k1,p ≤ kvk1,p ≤ 1.
Since Z
|vδ (x) − v(x)| = ϕ(y)[v(x) − v(x − δy)]dy
|y|≤1
Z Z δ|y|
∂
v x − r y drdy,
≤ ϕ(y)
|y|≤1 0
∂r |y|
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 144
by Hölder’s inequality,
Z Z Z δ|y| p
p
1 1 ∂ y
kvδ − vkp ≤ 0
ϕ(y) ϕ(y)
p p
∂r v x − r |y| drdy dx
Ω |y|≤1 0
Z Z
≤ ϕ(y)δ p |y|p |∇v(x)|p dxdy,
|y|≤1 Ω
i.e.
kvδ − vkp ≤ δk∇vkp .
Thus, ∃ δ0 = δ() such that
kv − vδ kp ≤ , ∀ v ∈ S, ∀ δ ≤ δ0 .
8
Fixing δ = δ0 , we have
Z
C
|vδ (x)| =
v(y)ϕδ (x − y)dy ≤ n kvkp
Ω δ
and
Z
C
|∇vδ (x)| =
v(y)∇ϕδ (x − y)dy ≤ n+1 kvkp .
Ω δ
This shows Sδ0 ⊂ C(Ω̄) is uniformly bounded and equicontinuous. According
to the Arzelà–Ascoli theorem, under the C-norm, Sδ0 has a finite 4mes(Ω) -net
{w1 , w2 , . . . , wl }. That is, ∀ v ∈ Sδ0 , ∃ wi ∈ Sδ0 such that
kvδ − wi kC < .
4mes(Ω)
Thus,
kvδ − wi kp < .
4
2. However, wi may not lie in B, but for each wi , there is a corresponding vδi 0 ,
where v i ∈ S = C0∞ (Ω) ∩ B. If vδi 0 , ∈/ B, its support must lie beyond Ω̄; we
can then take δi ∈ (0, δ0 ] such that the support of wi0 = vδi i is confined in Ω.
Consequently, we have wi0 ∈ B and
kwi − wi0 kp ≤ kwi − v i kp + kwi0 − v i kp ≤ .
4
∀ u ∈ B, ∃ v ∈ S such that ku − vk1,p ≤ 4 , hence
ku − wi0 kp ≤ ku − vkp + kv − vδ0 kp + kvδ0 − wi kp + kwi − wi0 kp < .
{w10 , . . . , wl0 } is the finite -net we are seeking.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 145
on some suitable Sobolev space W 1,q (Ω) and to make sure the E-L equation cor-
responding to the necessary condition of its extremal values actually makes sense,
we must impose some additional assumptions on L as follows:
(1) L, Lu , Lp are continuous;
(2) |L(x, u, p)| ≤ C(1 + |u|q + |p|q );
(3) |Lu (x, u, p)| + |Lp (x, u, p)| ≤ C(1 + |u|q + |p|q ).
In fact, by assumption (2), ∀ u ∈ W 1,p (Ω, RN ),
Z Z
I(u) = |L(x, u(x), ∇u(x))|dx ≤ C (1 + |u(x)|q + |∇u(x)|q )dx < ∞.
Ω Ω
Moreover, by assumption (3),
Z
|Lu (x, u(x), ∇u(x))| + |Lp (x, u(x), ∇u(x))|dx < ∞,
Ω
i.e. λ(x) := Lu (x, u(x), ∇u(x)) and µ(x) := Lp (x, u(x), ∇u(x)) ∈ L1 .
Lemma 10.2 Under the assumptions (1), (2), and (3), ∀ u0 ∈ W 1,q (Ω), ∀ ϕ ∈
C01 (Ω),
d
δI(u0 , ϕ) = I(u0 + sϕ)|s=0
Z dt
= [Lu (x, u(x), ∇u(x))ϕ(x) + Lp (x, u(x), ∇u(x))∇ϕ(x)]dx. (10.3)
Ω
Z 1
µs (x) = Lp (x, u(x) + θsϕ(x), ∇(u(x) + θsϕ(x)))dθ.
0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 146
Thus, the conclusion follows immediately from the Lebesgue Dominated Conver-
gence Theorem.
Remark 10.2 The continuity requirement in all three variables (x, u, p) in Lemma
10.2 assumption (1) can be replaced by the weaker Carathéodory condition:
As a consequence, we have:
Theorem 10.9 Suppose the Carathéodory condition (10.4), (2), and (30 ) hold. If
for any given ρ ∈ W 1,q (Ω, RN ), u0 ∈ M = ρ + W01,q (Ω, RN ) is a minimum of
I in M , then u0 satisfies the following E-L equation:
Z
[Lu (x, u(x), ∇u(x))ϕ(x) + Lp (x, u(x), ∇u(x))∇ϕ(x)]dx = 0,
Ω
∀ ϕ ∈ C0∞ (Ω, RN ). (10.5)
In this sense, we call solutions of (10.5) the generalized solutions of E-L equation.
Remark 10.4 The generalization of the concept of directional derivatives in dif-
ferential calculus on Banach spaces are the Gâteaux derivatives.
Let X be a Banach space, let U ⊂ X be an open subset, and f ∈ C(U, R1 ) be
a function defined on U .
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 147
C(1 + |p|q−1 ),
q > n.
We have the following theorem regarding the Gâteaux derivative of I:
00 1,q N
Theorem 10.10 R Suppose (1), (2) and (3 ) hold. Given ρ ∈ W (Ω, R ).
Let I(u) = Ω L(x, u(x), ∇u(x))dx be a functional defined on M = ρ +
W01,q (Ω, RN ). If I attains its minimum at u0 ∈ M , then u0 is Gâteaux differ-
entiable; furthermore,
Z
dI(u0 , ϕ) = [Lu (x, u(x), ∇u(x))ϕ(x) + Lp (x, u(x), ∇u(x))∇ϕ(x)]dx,
Ω
∀ ϕ ∈ W01,q (Ω, RN ).
Proof Fix any two Lebesgue points a < t0 < t1 < b of f . Let c = f (t0 ). Choose
ε > 0 such that (t0 − ε, t1 + ε) ⊂ J. We construct a piecewise linear function ψ
as follows
0, t∈/ [t0 − ε, t1 + ε]
ψ(t) =
1, t ∈ [t0 , t1 ]
and connecting the rest with straight lines. Clearly, there exists {ϕn } ⊂ C0∞ (J)
such that ϕn → ψ in W 1,1 (J). Hence,
Z Z
f (t)ϕ0n (t)dt → f (t)ψ 0 (t)dt;
J J
consequently,
Z t0 Z t1 +ε
−1 −1
ε f (t)dt − ε f (t)dt = 0.
t0 −ε t1
Exercises
Lecture 11
149
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 150
We also have
Z Z
lim inf c(x)|un (x)|r dx ≥ c(x)|u(x)|r dx.
Ω Ω
Suppose not, there exist > 0 and a subsequence {uj 0 } such that
Z Z
c(x)|uj 0 (x)|r dx < c(x)|u(x)|r dx − .
Ω Ω
a contradiction.
uj → u, Lq (Ω, RN ).
Then by the Riesz theorem, there exists a further subsequence, which is again
denoted by {uj } such that
∀ ε > 0, there exists a compact subset K ⊂ Ω such that mes(Ω\K) < ε and
(1) uj → u uniformly on K (Egorov’s theorem),
R ∇u are continuous on K (Luzin’s theorem),
(2) u and
(3) if Ω L(x, u(x), ∇u(x))dx < +∞, then
Z Z
L(x, u(x), ∇u(x))dx ≥ L(x, u(x), ∇u(x))dx − ε,
K Ω
(the absolute
R continuity of Lebesgue integrals).
If Ω L(x, u(x), ∇u(x))dx = +∞, then we take
Z
1
L(x, u(x), ∇u(x))dx > .
K ε
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 153
= I + II + III.
Remark 11.1 It is worth noting that the theorem does not require the functional
to be bounded from above, this is because we are only concerned with lower semi-
continuity. Without the restriction on the growth of the functional, it is conceivable
that ∃ u ∈ W 1,q (Ω) such that I(u) = +∞. However, this does not interfere with
our discussion on lower semi-continuity.
Utilizing the concept of Carathéodory functions, assumption (1) in Theorem
11.2 can be replaced by
(10 ): for a.e. (x, u), p 7→ L(x, u, p) is differentiable, L and Lp are both
Carathéodory functions.
such that
−∆u = f.
As a matter of fact, by Theorem 11.3, the functional
Z
1 2
I(u) = |∇u| − f u dx
Ω 2
attains its minimum on H01 (Ω).
11.4∗ Quasi-convexity
namely,
|um |1,∞ ≤ kϕ0 kL∞ (J)
and
um → u (in L∞ (J)).
By further applying the facts ϕ̇m (t) = ϕ0 (mt), ϕ ∈ W01,∞ (J, RN ), ϕ̄0 = 0, and
by Example 9.3 in Lecture 9, it follows that
ϕ0m *∗ 0 (in L∞ (J)).
Now ∀ ξ ∈ (W 1,1 (J, RN ))∗ , we have ξ = (ξ0 , ξ1 ) ∈ L1 (J, R2N ). Moreover,
since Dum *∗ Du in L∞ (J), it follows that
Z
hξ, um − ui = [ξ0 (um − u) + ξ1 D(um − u)]dt → 0.
J
Namely, um *∗ u in W 1,∞ (J, RN ).
Notice that
Z
I(u) = L(p0 )dt = L(p0 ),
J
and by the periodicity of ϕ, we have
Z
I(um ) = L(p0 + ϕ0m (t))dt
J
Z
1
= L(p0 + ϕ0 (t))dt
m mJ
Z
= L(p0 + ϕ0 (t))dt.
J
Suppose I is weakly sequentially lower semi-continuous, lim inf I(um ) ≥ I(u),
it yields that
Z
L(p0 + ϕ0 (t))dt ≥ L(p0 ).
J
Inspired by the above argument, we now consider the case n > 1.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 157
Exercises
Lecture 12
we note the first term in the functional I precisely corresponds to the norm induced
by such inner product, whereas the second term can also be expressed via this
163
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 164
which implies the continuity of the functional. Hence, there exists F ∈ (H01 (Ω))∗
such that Z
F (ϕ) = f · ϕ dx.
Ω
By the Riesz representation theorem from functional analysis, this continuous lin-
ear functional can be represented via the inner product as follows: ∃ u0 ∈ H01 (Ω)
such that Z
((u0 , ϕ)) = F (ϕ) = f ϕ dx.
Ω
Consequently, we can rewrite the E-L equation of I in terms of the inner product:
Z
((u, ϕ)) = f · ϕ dx = ((u0 , ϕ)), ∀ ϕ ∈ C0∞ (Ω).
Ω
Since C0∞ (Ω) is dense in H01 (Ω), u = u0 is the solution.
In appearance this approach contains neither the minimizing sequence nor
the weak sequential compactness and the weak sequential lower continuity. The
solution is obtained directly. However, the proof of the Riesz representation the-
orem is itself a variational problem, namely, finding the distance from a point
outside a given hyperplane to the hyperplane in a Hilbert space, along with the
projection of the point onto the hyperplane.
Let us recall this proof. Denote
M = {η ∈ H01 (Ω)|F (η) = 0}.
M is a closed linear subspace of H01 (Ω). For any ϕ ∈ H01 (Ω), we can find the
orthogonal projection η of ϕ onto M , i.e. η ∈ M and ϕ − η = ξ⊥M . If we take
ξ
F (ξ) 2 ξ 6= 0,
v0 = [[ξ]]
0 ξ = 0,
F (ϕ) = F (ξ),
but
ξ
((v0 , ϕ)) = F (ξ) 2 , ϕ = F (ξ),
[[ξ]]
which is exactly what we want.
If we take an equivalent inner product on the Hilbert space, specifically
Z 21
2
[[u]] = |∇u| dx ,
Ω
then the following argument precisely depicts the minimization process used in
solving variational problems.
We now return to examine the existence of orthogonal projections (see
Figure 12.1).
Fig. 12.1
In fact, if there exists η ∈ M such that [[ϕ − η]] = minx∈M [[ϕ − x]], then
ξ = ϕ − η satisfies
((ξ, x − η)) = 0, ∀ x ∈ M.
Fig. 12.2
The question is: does {ηj } converge? ∀ ε > 0, by the parallelogram law and
the fact that M is a linear subspace, we have
2
2 2 2 ηj + ηk
[[ηj − ηk ]] = 2([[ηj − ϕ]] + [[ηk − ϕ]] ) − 4 −ϕ
2
≤ 4(m + ε)2 − 4m2 , for j, k sufficiently large. (12.2)
Hence, {ηj } is a Cauchy sequence. By the completeness of H01 (Ω), {ηj } is con-
vergent, i.e. ηj → η. Since M is closed, η ∈ M and it achieves
[[η − ϕ]] = min [[ϕ − x]] .
x∈M
This argument provides a different angle to verify the Dirichlet’s principle. By the
peculiarity of the problem, it is unnecessary to employ weak convergence, while
the metric completeness of the space plays a key role.
The same method can also be used in handling variational inequalities (cf. Lec-
ture 7). For instance, the obstacle problem of a thin membrane: on the bounded
region Ω ⊂ Rn , given a measurable function ψ(x) as an obstacle and the external
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 167
force function f ∈ L2 (Ω). We seek the equilibrium position of the thin membrane
u : Ω → R1 .
We restate this as a variational problem. Let C = {u ∈ H01 (Ω) | u(x) ≤
ψ(x)
R a.e.} be a closed convex subset of the Hilbert space H01 (Ω). Denote (f, v) =
Ω
f · vdx, we wish to find u ∈ C such that
((u, v − u)) − (f, v − u) ≥ 0, ∀ v ∈ C. (12.3)
According to the above discussion, there exists u0 ∈ H01 (Ω) such that
((u0 , v)) = (f, v), ∀ v ∈ H01 (Ω).
Thus, it becomes
((u − u0 , v − u)) ≥ 0, ∀ v ∈ C. (12.4)
Based on the earlier argument, this can be accomplished by finding
min [[u0 − v]] . (12.5)
v∈C
and
Z
2
N (u) = |u| dx,
Ω
we wish to find
min{I(u)|u ∈ H01 (Ω) ∩ N −1 (1)}. (12.7)
If ϕ1 ∈ C 2 (Ω̄) indeed achieves such minimum, then by the Lagrange multipliers,
for the adjusted E-L equation of the Lagrangian, there exists λ1 ∈ R1 such that
−∆ϕ1 = λ1 ϕ1 in Ω. (12.8)
R 2
Since N (ϕ1 ) = Ω |ϕ1 | dx = 1, ϕ1 is nonzero. Multiplying both sides of (12.8)
by ϕ1 and then integrating, it yields
Z
2
I(ϕ1 ) = |∇ϕ1 | dx = λ1 .
Ω
This means
λ1 = min{I(u)|u ∈ H01 (Ω) ∩ N −1 (1)}. (12.9)
In the following, we shall verify the existence of a solution for the minimization
problem (12.7). Suppose Ω ⊂ Rn is bounded, we already know that I is coercive
and weakly sequentially lower semi-continuous, it remains to show the set
M1 = {u ∈ H01 (Ω) | N (u) = 1}
is weakly sequentially closed. That is, if {uj } ⊂ M1 , uj * u in H01 (Ω), then
u ∈ M1 .
By the Rellich–Kondrachov compact embedding theorem, for the bounded
region Ω, H01 (Ω) ,→ L2 (Ω) is a compact embedding. By assumption, uj * u in
H01 (Ω), it contains a subsequence {u0j } such that
u0j → u in L2 (Ω).
Moreover, since
Z
u2j dx = 1,
Ω
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 169
it follows that
Z
u2 dx = 1,
Ω
we seek to
min{I(u) | u ∈ M2 }.
−∆ϕ2 = λ2 ϕ2 + µ2 ϕ1 . (12.10)
We now prove that µ2 = 0. First, multiplying both sides of (12.8) by ϕ2 and then
integrating, it yields
Z Z Z
∇ϕ2 ∇ϕ1 dx = ∇ϕ1 ∇ϕ2 = λ1 ϕ1 ϕ2 dx = 0.
Ω Ω Ω
Thus, µ2 = 0.
Consequently,
−∆ϕ2 = λ2 ϕ2 in Ω.
R
closed, u ∈ M1 . From Ω
uj ϕ1 = 0, it follows that
Z
uϕ1 = 0.
Ω
and use a similar argument, we can show that each Mn is weakly closed. Hence,
the constrained optimization problem
min{I(u) | u ∈ Mn }
has a solution ϕn 6= 0 satisfying
n−1
X
−∆ϕn = λn ϕn + µj ϕj .
j=1
and
Z Z
((ϕi , ϕj )) = ∇ϕi · ∇ϕj dx = λi ϕi ϕj dx = 0, ∀ i 6= j.
Ω Ω
and we call them generalized Fourier coefficients of u. Consider the partial sum
m
X
sm (x) = cn ϕn (x);
n=1
it follows that
m
X Z
2 2
λn |cn | ≤ |∇u| dx. (12.14)
n=1 Ω
However, we define an equivalent norm and inner product on H01 (J) via I as
follows:
Z 21
kuk = p(t)|u0 (t)|2 + q(t)|u(t)|2 dt ,
J
Z
((u, v)) = p(t)u0 (t)v 0 (t) + q(t)u(t)v(t)dt.
J
Inductively, by introducing the constraints
Z Z
1 2
M1 = u ∈ H0 (J) |u| = 1 , M2 = u ∈ M1
uϕ1 dx = 0 , . . . ,
J
u0 (a) = u0 (b) = 0.
As mentioned in Lectures 1 and 6, for the Neumann problem, we shall use the
space H 1 (Ω) instead of H01 (Ω). The reason is as follows: using integration by
parts:
Z b Z b
(pu0 )0 ϕdx = (pu0 )ϕ|ba − (pu0 )ϕ0 dx,
a a
i.e. replacing H01 (J) by the subspace which is orthogonal to all constant functions.
R b X, Poincaré’s inequality still remains valid. The proof is
On the subspace
identical, since a udx = 0, there exists ξ ∈ (a, b) such that u(ξ) = 0. Thus,
Z x Z b 12
1
0 0 2
|u(x)| =
u (t)dt ≤ (b − a)
2 |u | dt .
ξ a
We then extend Poincaré’s inequality Rto the subspace X; (after inserting an inte-
gral constraint), we minimize I(u) = Ω |∇u|2 dx on X.
Remark 12.4 The same method applies to the eigenvalue problem of the Laplace–
Beltrami operator on a closed (no boundary) compact Riemannian manifold
(M, g):
1 X √
∆g u = √ ∂i (g ij g∂j u),
g ij
and
Z
2
|∇u| dx
µ(En−1 ) = min ZΩ .
⊥
u∈En−1 \{θ} 2
|u| dx
Ω
µ(En−1 ) ≤ λn .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 177
⊥
Let {ϕ1 , . . . , ϕn } be the first n eigenfunctions, then (En−1 \{θ}) ∩
span{ϕ1 , . . . , ϕn } 6= ∅, i.e. there exists a nonzero u such that
Xn
u = ci ϕi ,
Z i=1
j = 1, . . . , n − 1,
uvj dx = 0,
Ω
or
n
X Z
ci ϕi vj dx = 0 j = 1, . . . , n − 1.
i=1 Ω
This system
R of (n − 1) linear equations in the n unknowns c1 , . . . , cn with coeffi-
cients Ω ϕi vj dx must possess a nontrivial solution. From λ1 ≤ λ2 ≤ · · · ≤ λn ,
it follows that
Z n
2 P 2
|∇u| dx λi |ci |
Ω i=1
µ(En−1 ) ≤ Z = P n ≤ λn .
2 2
|u| dx |ci |
Ω i=1
Exercises
prove that
Xn
−
∂i (ai,j (x)∂j u(x)) + c(x)u(x) = λu(x), x ∈ Ω,
i,j=1 x ∈ ∂Ω,
u(x) = 0,
Lecture 13
That is, the minimum is a generalized solution of the corresponding E-L equation.
However, for a functional containing first order derivatives, its E-L equation
is a second order differential equation. A generalized solution is a solution in
the ordinary sense only if it is twice differentiable. From a differential equation
perspective, we need also address whether such a generalized solution would have
enough differentiability to fulfill the E-L differential equation. In other words, is
it possible to deduce u ∈ C 2 from the generalized solution u? Or perhaps more
differentiability, or even analyticity? This is the so-called regularity problem.
We call the problem of finding a generalized solution an “existence” problem,
and call the problem of determining the differentiability of a generalized solution
a “regularity” problem.
That being said, in direct methods, the “existence” and “regularity” are twin
problems.
Among Hilbert’s 23 problems, problem 19 inquires the analyticity of solu-
tions of some regular variational problems (in fact, it is about elliptical equations),
whereas problem 20 is related to the solvability of a regular variational problem
with boundary conditions (namely, the existence).
179
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 180
13.1 Regularity (n = 1)
where c is a constant. Using the integral form of the E-L equation, we have
Lp (t, u∗ (t), u̇∗ (t)) = q(t).
Define ϕ : J¯ × RN → RN via
ϕ(t, p) = Lp (t, u∗ (t), p) − q(t).
We know that ϕ ∈ C 1 (J¯ × RN , RN ) satisfies
det(ϕp (t, u∗ (t))) = det(Lpp (t, u∗ (t), u̇∗ (t))) 6= 0
and
ϕ(t, u̇∗ (t)) = 0.
By the Implicit Function Theorem, the equation
ϕ(t, p) = 0
has a unique local C 1 -solution, i.e. ∀ t0 ∈ J, ¯ there exists a neighborhood U =
U (t0 ) and a unique λ ∈ C 1 (U, RN ) such that in the neighborhood of (t, u̇∗ (t)),
ϕ(t, λ(t)) = 0.
Thus, u̇∗ (t) = λ(t) ∈ C 1 , which implies u∗ ∈ C 2 .
The condition det(Lpi pj (t, u∗ (t), u̇∗ (t)) 6= 0 plays an important role; for oth-
erwise, there exists a functional whose minimum is of C 1 but not of C 2 .
Example 13.1 Let L(t, u, p) = u2 (p − 2t)2 and M = {u ∈
C 1 ([−1, 1]) | u(−1) = 0, u(1) = 1}, then the functional
Z 1
I(u) = u2 (u̇(t) − 2t)2 dt
−1
has a minimum
∗ 0, t < 0,
u (t) =
t2 , t ≥ 0.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 181
It is clear u∗ ∈ C 1 \C 2 ([−1, 1]). Moreover, Lpp (t, u∗ (t), u̇∗ (t)) = 2u∗ (t)2 = 0,
for t < 0.
The above lemma elevated the solution from being C 1 to being C 2 , based on
the Implicit Function Theorem. However, it is not enough to insure the regularity
of the solution. In Lecture 9, we pointed out that the space C 1 is not the suitable
function space for direct methods. Using direct methods, we can only obtain
a solution in some Sobolev space. In order to obtain its regularity, we have to
further establish that the generalized solution is indeed a C 1 solution.
Theorem 13.1 For 1 < r < ∞, assume L satisfies the following growth
condition:
|L(t, u, p)| + |Lu (t, u, p)| + |Lp (t, u, p)| ≤ C(1 + |p|r );
for r = ∞, no growth condition on L is needed.
Furthermore, assume the matrix (Lpi pj (t, u, p)) ∀ (t, u, p) ∈ J¯ × RN × RN
is positive definite.
∗
R If u ∈ W 1,r (J, RN ) is a minimum of the functional I(u) =
Ω
L(t, u(t), u̇(t))dt, then by changing the values of u∗ on a set of measure zero,
∗
u ∈ C 2.
Proof By Lemma 13.1, it suffices to show that u∗ ∈ C 1 .
Define the function ϕ : J × RN × RN × RN → R1 by
ϕ(t, u, p, q) = Lp (t, u, p) − q.
By assumption, det(Lpi pj (t, u, p)) 6= 0, ∀ (t, u, p) ∈ J¯ × RN × RN . To solve
the equation: ϕ(t, u, p, q) = 0, we apply the Implicit Function Theorem to see if
there exists a unique local C 1 -solution
p = λ(t, u, q). (13.1)
On one hand, we show that this solution is globally unique. Suppose p1 , p2 both
satisfy (13.1), then q = Lp (t, u, p1 ) = Lp (t, u, p2 ). Hence,
0 = (Lp (t, u, p1 ) − Lp (t, u, p2 ), p1 − p2 ) = (B(p1 − p2 ), p1 − p2 ),
where
Z 1
B= Lpp (t, u, p1 + τ (p2 − p1 ))dτ.
0
By assumption, B is positive definite, therefore p1 = p2 .
On the other hand, since u∗ ∈ W 1,r (J, RN ), when 1 < r < ∞,
Lu (t, u∗ (t), u̇∗ (t)) ∈ L1 (J) and when r = ∞, Lu (t, u∗ (t), u̇∗ (t)) ∈ L∞ (J).
In any case,
Z t
q(t) = Lu (s, u∗ (s), u̇∗ (s))ds − c is absolutely continuous.
t0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 182
it follows that
q(t) = Lp (t, u∗ (t), u̇∗ (t)), a.e. t ∈ J
and
u̇∗ (t) = λ(t, u∗ (t), q(t)), a.e. t ∈ J.
Since q(t) is absolutely continuous, substituting it into the above expression, by
changing the values of u∗ on a set of measure zero, we have u̇∗ (t) is continuous,
i.e. u∗ ∈ C 1 . Thus, u∗ ∈ C 2 follows readily from Lemma 13.1.
In the above proof, the global positive definiteness of the matrix
(Lpi pj (t, u, p)), ∀ (t, u, p) ∈ J¯ × RN × RN
played a crucial role in establishing both the global uniqueness of the solution as
well as its regularity. This is because the derivatives of functions in the Sobolev
space W 1,r may be discontinuous. Along the graphs of these functions, two
timewise nearby points may fall in different image neighborhoods. Since the Im-
plicit Function Theorem only works for neighborhoods in the image space, it is
no longer applicable.
The following example demonstrates, by removing the global positive defi-
niteness assumption, the solution becomes non-differentiable.
Example 13.2 Let L(p) = (p2 − 1)2 and M = {u ∈ Lip([0, 1]) | u(0) = u(1) =
0}, then the functional
Z 1
I(u) = (u̇2 (t) − 1)2 dt
0
has minimal value 0.
Note that Lpp (t, u, p) = 4(3p2 − 1) is not positive definite.
If u ∈ C 1 is a minimum, then u̇(t) = ±1. However, regardless of u̇(t) = 1 or
u̇(t) = −1, it is impossible to have a solution satisfying the boundary condition
u(0) = u(1) = 0. In other words, there cannot be a C 1 -solution to achieve
the minimal value. On the contrary, there are uncountably many solutions of
this variational problem, which are Lipschitz sawtooth-like functions satisfying
u̇(t) = ±1 (see Figure 13.1).
Lastly, we give an example to show that without the convexity of the La-
grangian, the weak sequential lower semi-continuity of the functional may not
hold.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 183
Fig. 13.1
Example 13.3 (Bolza) Let L(u, p) = u2 + (p2 − 1)2 and M = W01,4 (0, 1), the
the functional Z 1
I(u) = [u2 (t) + (u̇(t)2 − 1)2 ]dt
0
has infimum zero, but I has no minimum in M .
To see this, note on one hand, I(u) ≥ 0; on the other hand, we define a
minimizing sequence of sawtooth-like functions:
k k 2k + 1
t − ,
t∈ , ,
j j 2j
uj (t) =
k+1 2k + 1 k + 1
−t + , t∈
, ,
j 2j j
where j = 2, 3, . . . and 0 ≤ k ≤ j − 1.
1
Since |u̇j (t)| = 1 a.e. and |uj | ≤ 2j , we see that
1
I(uj ) ≤ 2 → 0, j → ∞.
4j
Thus, inf u∈M I = 0.
However, I has no minimum in W01,4 (0, 1). Suppose not, there exists u0 ∈
1,4
W0 (0, 1) such that I(u0 ) = 0, hence u̇0 (t) = 0 a.e. and I(u0 ) = 1, a contra-
diction (see Figure 13.2).
Fig. 13.2
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 184
The regularity problems of partial differential equations are far more compli-
cated than those of ordinary differential equations, they demand special knowl-
edge. In order to give the interested readers a taste of this subject, we provide the
following example.
Theorem 13.2 (Weyl) Suppose u ∈ L1loc (Ω) satisfies
Z
u · 4ϕdx = 0, ∀ ϕ ∈ C0∞ (Ω), (13.2)
Ω
then after changing its value on a set of measure zero, u ∈ C ∞ (Ω).
We call those locally L1 functions which satisfy Eq. (13.2) weakly harmonic
functions. The idea of the proof originates from the mean value property of har-
monic functions.
If u ∈ C 2 (Ω) and ∆u = 0 in Ω, then
Z Z
1 n
u(x) = n−1 u(y)dσ = n u(y)dy, ∀ Br (x) ⊂ Ω, (13.3)
r $n ∂Br (x) r $n Br (x)
where $n denotes the surface area of the unit sphere S ⊂ Rn . We call (13.3) the
mean value property of u.
To verify (13.3), we assume Br (x) ⊂ Ω and choose ρ ∈ (0, r), then
Z Z Z
∂u n−1 ∂
0= ∆u(y)dy = dσ = ρ u(x + ρw)dw.
Bρ (x) ∂Bρ (x) ∂n ∂ρ |w|=1
This implies
Z
u(x + ρw)dw = const. = $n u(x),
|w|=1
hence,
Z Z
1 n
u(x) = u(y)dσ = u(y)dy.
rn−1 $n ∂Br (x) r n $n Br (x)
≤ Cδ0 |x − y|kuk1 .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 186
Proof On [0, T ], we expand the periodic function u as its Fourier series. Since
ū = 0,
∞
X 2πkt 2πkt
u(t) = ak cos + bk sin ,
T T
k=1
hence,
∞
0 2π X 2πkt 2πkt
u (t) = −kak sin + kbk cos .
T T T
k=1
September 1, 2016 9:0 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 188
By Parseval’s identity,
Z T ∞
(2π)2 X 2
|u0 (t)|2 dt = k (|ak |2 + |bk |2 )
0 T
k=1
2 X∞
(2π)
≥ (|ak |2 + |bk |2 )
T
k=1
2 Z T
(2π)
= |u(t)|2 dt.
T2 0
1
∀ u ∈ Hper (0, T ), we decompose
u = ũ + ū,
RT
where ū = T1 0 u(t)dt is a real number. From Wirtinger’s inequality, we see
that ū is not controlled by the values of I. In other words, if we use the H 1 -norm
1
directly on Hper (0, T ), then I is not coercive.
Moreover, since the nonlinear term cos u in I is 2π-periodic, we have
I(u + 2π) = I(u).
1
By which, we need not consider I on the entire Hper (0, T ), but instead, by setting
1
M = {u = ξ + η | ξ ∈ Hper (0, T ), ξ¯ = 0, η ∈ [0, 2π]};
we then restrict I on M . The advantage is that ū now only varies on the bounded
interval [0, 2π].
Noting that M is weakly sequentially closed and
1 ˙ 2 ˙ 2 − |a|;
I(u) ≥ kξk2 − kEk∞ kξk
2
by Wirtinger’s inequality, kξk ˙ 2 is an equivalent norm on H 1 . I is coercive on M .
It is not difficult to verify that I is also weakly lower semi-continuous. According
1
to the existence theorem, there exists a minimum u ∈ M ⊂ Hper . Moreover,
since all conditions in Theorem 13.1 on regularity are met, it follows that u ∈ C 2 .
Example 13.6 Let Ω ⊂ Rn be a bounded region, 1 < r < p < ∞, and f ∈
p
L p−r (Ω), then
Z
1 1
I(u) = |∇u(x)|p − f (x) · |u(x)|r−1 u(x) dx, ∀ u ∈ H01 (Ω)
Ω p r
has a minimum u0 ∈ W01,p (Ω), and it satisfies
Z
[|∇u(x)|p−2 ∇u(x)∇ϕ(x) − |u|r−1 f (x)ϕ(x)]dx = 0, ∀ ϕ ∈ C0∞ (Ω).
Ω
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 189
on W01,p (Ω).
(1) We claim I is coercive.
By Poincaré’s inequality, Hölder’s inequality, and Young’s inequality, ∃ C > 0
such that
Z p
|f ||u|r dx ≤ kf k p−r r r
p kuk ≤ Ckf k p ||u|| ≤ C kf k p
p p−r
p−r
+ rkukp .
p−r
Ω
Thus,
1 C p
where 2∗ = 2n
n−2 (n > 2). Assume
f (x, 0) = 0,
find u ∈ H01 (Ω)\{0} such that
−∆u(x) = λf (x, u(x)),
where λ is a parameter. Furthermore, λ is called an eigenvalue if it corresponds to
a nonzero solution u ∈ H01 (Ω).
Solution Similar to linear equations, we regard this as a constrained variational
problem. On H01 (Ω), we consider the functionals
Z
1
I(u) = |∇u(x)|2 dx,
2 Ω
Z
N (u) = F (x, u(x))dx,
Ω
where
Z t
F (x, t) = f (x, s)ds
0
is an anti-derivative of f (·, t). Since the growth of f is restricted, N (u) is well-
defined on H01 (Ω).
Likewise, we want to find the minimum of I on M = N −1 (1). Since I
is coercive and weakly lower semi-continuous, and by the Rellich–Kondrachov
compact embedding theorem, M is weakly closed, it follows that I attains its
minimum u0 . Moreover, using the Lagrange multipliers, there exists a real number
λ0 such that
−∆u0 (x) = λ0 f (x, u0 (x)).
When f (x, u) = u, this is precisely the eigenvalue problem in Lecture 12. How-
ever, when f (x, u) = ur and 1 ≤ r < 2∗ − 1 with 2∗ = n−2 2n
(n > 2), it
generalizes the linear eigenvalue problem.
Remark 13.4 Recall in Lecture 12, we obtained an increasing sequence of eigen-
values for a linear differential equation. It is natural to ask: for what kind of
nonlinear differential equations would we have similar results. It is not a simple
task to answer this since for nonlinear problems, we no longer have orthogonal-
ity between two eigenvectors with distinct eigenvalues. However, if f is odd in
u, that is, f (x, −u) = −f (x, u); similar results are captured by the Liusternik–
Schnirelmann theory, which requires more in-depth knowledge of topology.
The following gives an example of a functional which is neither bounded
above nor below, so the common variational method does not seem applicable.
However, by a certain technique, we can turn it into a minimal value problem.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 192
1
t = t(v) = p ,
|v|pp−2
which also satisfies
d
I(tv)|t=t(v) = 0.
dt
Substituting into the original functional I, we obtain a functional I˜ on the unit
sphere S by
˜ = 1−1
I(v)
1
2p .
2 p
|v|pp−2
Using the embedding theorem, there exists a constant C > 0 such that
|v|pp ≤ C p kvkp = C p .
Thus,
1 1
≥ p.
|v|pp C
This means I˜ is indeed a continuous functional on S. Using compact embedding,
it is also weakly sequentially continuous. Note that S is itself weakly sequentially
compact, hence I˜ must attain its minimum at some v0 .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 193
d
I(tv)|t=t(v) = (I 0 (t(v)v), v).
dt
Thus,
i.e.
(I 0 (u0 ), ϕ) = 0, ∀ ϕ ∈ X.
At the end of this lecture, we point out, in particular, the solutions of differ-
ential equations are not always obtainable by direct methods. Hadamard gave the
following counterexample: let
∞
X sin m!ϑ
g(ϑ) = ,
m=1
m2
then
∞
X rm! sin m!ϑ
u(r, ϑ) = , ((r, ϑ) ∈ [0, 1] × [0, 2π])
m=1
m2
Exercises
(2)
Z 1
(u̇2 − 1)2 dt, M = W01,4 (0, 1).
0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 195
Lecture 14
The main focus of this lecture includes: the dual least action principle and the
Ekeland variational principle.
The dual least action principle is mainly applied to Hamiltonian systems and
related problems. In general, the functional associated with a Hamiltonian system
is neither bounded above nor below, so variational methods are difficult to apply.
However, if the Hamiltonian in a Hamiltonian system is a convex function, then
by means of convex conjugates, the problem can be transformed into a constrained
variational problem. This is the essence of the dual least action principle.
The Ekeland variational principle is a general minimization result with a broad
variety of applications. It provides a specific method in choosing a minimizing se-
quence; consequently, this minimizing sequence along with some other conditions
give rise to numerous applications.
195
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 196
Its inverse is also a set-valued mapping. If we still denote the inverse (set-valued)
mapping of ∂f (x) by ψ(ξ), then
ψ(ξ) = {x ∈ D(f ) | ξ ∈ ∂f (x)},
i.e.
x ∈ ψ(ξ) ⇐⇒ x is such that f (y) − hξ, yi achieves its minimal value.
Accordingly, we can extend the Legendre transformation to convex functions.
Comparing to (14.1), we introduce the following.
Definition 14.1 (Conjugate function) Let f : Rn → R̄ = R1 ∪ {+∞} be a
proper function, i.e. D(f ) 6= ∅. We call
f ∗ (ξ) = sup {hξ, xi − f (x)}
x∈Rn
whence
hx∗0 , x0 i + λf ∗∗ (x0 ) ≥ α. (14.3)
This contradicts (14.2).
If f ∗∗ (x0 ) < +∞, then λ > 0 still holds, which again yields a contradiction.
It remains to consider the case where f (x0 ) = f ∗∗ (x0 ) = +∞ and λ = 0.
From (14.2), ∃ ε > 0 such that
hx∗0 , x − x0 i ≥ ε, ∀ x ∈ D(f ). (14.4)
∗ ∗ n ∗ ∗
Since f is proper, ∃ x1 ∈ R such that f (x1 ) < +∞ and
hx∗1 , xi − f (x) − f ∗ (x∗1 ) ≤ 0, ∀ x ∈ D(f ). (14.5)
Combining (14.4) and (14.5), ∀ n ∈ N, it yields
hx∗1 − nx∗0 , xi + nhx∗0 , x0 i + nε − f (x) − f ∗ (x∗1 ) ≤ 0, ∀ x ∈ D(f ).
Consequently,
f ∗ (x∗1 − nx∗0 ) + nhx∗0 , x0 i + nε − f ∗ (x∗1 ) ≤ 0
or
nε + hx∗1 , x0 i − f ∗ (x∗1 ) ≤ hx∗1 − nx∗0 , x0 i − f ∗ (x∗1 − nx∗0 ) ≤ f ∗∗ (x0 ).
Letting n → ∞, it follows that f ∗∗ (x0 ) = +∞, a contradiction.
Corollary 14.1 For a proper, convex, lower semi-continuous function f , we have
ξ ∈ ∂f (x) ⇐⇒ x ∈ ∂f ∗ (ξ).
This directly generalizes the Fenchel–Moreau theorem and property (5).
Corollary 14.2 If f : Rn → R̄ is proper, then
f ∗∗ = conv(f ) = sup{ϕ : Rn → R̄ | ϕ(x) ≤ f (x), ∀ x ∈ Rn , ϕ is convex }.
Proof Suppose g ≤ f is convex, then it is proper and convex. By property (3),
f ∗ ≤ g ∗ ; moreover, g ∗∗ ≤ f ∗∗ . By the Fenchel–Moreau theorem, we see that
g = g ∗∗ ≤ f ∗∗ .
Example 14.1 Let f (x) = |x|p /p, 1 < p < ∞, where |x| = (x21 + x22 + · · · +
x2n )1/2 , then
1 0 1 1
f ∗ (ξ) = 0 |ξ|p , where + 0 = 1.
p p p
Example 14.2 Let f (x) = |x|, then
(
∗ 0 if |ξ| ≤ 1,
f (ξ) =
+∞ if |ξ| > 1.
Proof Recall
f ∗ (ξ) = sup{hξ, xi − |x|}.
x
For |ξ| > 1, we let x = tξ and t → +∞, then f ∗ (ξ) = +∞. For |ξ| ≤ 1, since
hξ, xi − |x| ≤ 0, by choosing x = 0, the result follows.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 199
i.e.
(H ∗ )0 (ρ˙0 , −w˙0 ) = λ(w0 , ρ0 ). (14.10)
Since the sub-differential of the conjugate function and the sub-differential of
the original function are inverses of each other, as stated in Corollary 14.1, (14.10)
is equivalent to
ρ˙0 (t) = Hw (t, λw0 (t), λρ0 (t)),
(14.11)
w˙0 (t) = −Hρ (t, λw0 (t), λρ0 (t)).
If we define
η = λρ0 ,
v = λw0 ,
by substituting into (14.11), it gives (14.8).
Lastly, we verify λ > 0. In (14.10), first multiplying both sides by (ρ̇0 , −ẇ0 )
and then integrating, we obtain
h(H ∗ )0 (ρ˙0 , −w˙0 ), (ρ˙0 , −w˙0 )i = −λG(w0 , ρ0 ) = λπ.
It follows from (14.6) that ∇H(θ, θ) = (θ, θ). Using property (5) of a conjugate
function, we see that
H ∗ (θ, θ) = −H(θ, θ) = 0.
Since H ∗ is convex, we have
H ∗ (θ, θ) − H ∗ (ρ˙0 , −w˙0 ) ≥ −h(H ∗ )0 (ρ˙0 , −w˙0 ), (ρ˙0 , −w˙0 )i,
i.e.
h(H ∗ )0 (ρ˙0 , −w˙0 ), (ρ˙0 , −w˙0 )i ≥ H ∗ (ρ˙0 , −w˙0 ) ≥ 0.
Thus, λ ≥ 0.
It remains to show that λ 6= 0. We argue by contradiction. Suppose λ = 0,
then
(H ∗ )0 (ρ˙0 , −w˙0 ) = (θ, θ).
By Corollary 14.1, we see that
ρ˙0 = −Hu (θ, θ),
w˙0 = Hρ (θ, θ).
Using condition (14.6), we conclude that (Hu (θ, θ), Hρ (θ, θ)) = (θ, θ), a contra-
diction to conclusion (2).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 202
so (2) holds.
Lastly, we prove (3) by contradiction. Suppose yε = u∗ does not satisfy (3),
then ∃ w 6= u∗ such that
f (w) ≤ f (u∗ ) − εd(u∗ , w). (14.14)
From (14.12), it follows that
εd(un , u∗ ) ≤ f (un ) − f (u∗ ),
i.e.
f (u∗ ) ≤ f (un ) − εd(un , u∗ ). (14.15)
Combining (14.14) and (14.15), we deduce that
f (w) ≤ f (un ) − εd(un , w), ∀ n.
T∞
Thus, w ∈ n=1 Sn . Using (14.13), we have f (u∗ ) ≤ f (w), which contradicts
(14.14).
Corollary 14.3 Let (X, d) be a complete metric space and f : X → R1 ∪ {+∞}
be a proper function, which is bounded below and lower semi-continuous. Then
∀ ε > 0, ∃ yε ∈ X such that f (x) > f (yε ) − εd(x, yε ), ∀ x 6= yε .
It is worth noting that although the Ekeland variational principle only employs
the metric topology, which involves neither the weak topology nor the various no-
tions of compactness, but the minimum of the functional f is not reached. How-
ever, the significance lies in that by choosing this special minimizing sequence, it
produces a special sequence of approximated minima.
f (xn ) < c + n1 .
The first inequality implies
1
kf 0 (xn )k = sup |df (xn , ϕ)| ≤ .
kϕk=1 n
The second inequality implies
f (xn ) → c.
According to the PSc condition, there exists a convergent subsequence {xnj } such
that xnj → x∗ ∈ X. By continuity, it follows that f (x∗ ) = c.
Since many functionals are neither bounded above nor below, in appearance,
it seems difficult to apply variational methods to find their critical points. How-
ever, for some particular problems, Nehari provided a special technique, which
transforms a critical point problem to an extremal problem.
Let H be a Hilbert space, equipped with inner product (·, ·), and a given func-
tional I ∈ C 2 (H, R1 ). We look for the critical points of I, namely, those points
for which I 0 (u) = 0.
Define G(u) = hI 0 (u), ui. We note that all critical points u of I satisfy:
G(u) = 0.
If the set M = {h ∈ H | G(u) = 0} is a manifold; for example, G0 (u) 6= θ,
˜ Furthermore,
∀ u ∈ M , then we can restrict I on M to obtain a new functional I.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 207
We compute that
Z
0
G(u) = hI (u), ui = [|∇u|2 + µa(x)|u(x)|µ ]dx.
Ω
On G−1 (0),
Z
˜ µ 1 1
I(u) = (µ − 2) a(x)|u(x)| dx = − kuk2
Ω 2 µ
is nonnegative.
1◦ Note that
Z
hG0 (u), vi = [2∇u · ∇v + µ2 a(x)|u(x)|µ−2 u(x)v(x)]dx, ∀ v ∈ H01 (Ω),
Ω
This implies θ is the only isolated point in G−1 (0). Let M = G−1 (0)\{θ}, then
on M ,
Z
hG0 (u), ui = 2kuk2 + µ2 a(x)|u(x)|µ ]dx > 0,
Ω
0
whence G (u) 6= 0, ∀ u ∈ M .
2◦ It is clear I˜ ∈ C 1 (H01 (Ω), R1 ). We now verify the Palais–Smale condition.
Suppose {uj } ⊂ M is a PS-sequence satisfying I(u ˜ j ) → 0 and |I(u
˜ j )| ≤ C.
Since
1 1 ˜ j ) ≤ C,
− kuj k2 = I(u
2 µ
there exists a subsequence uj 0 * u. Moreover, since
I˜0 (u) = I 0 (u) = u + (−∆)−1 µa(x)|u(x)|µ−2 u(x)
and the embedding
H01 (Ω) ,→ Lµ (Ω)
is compact, as I˜0 (uj ) → 0, {uj } has a subsequence which converges in Lµ (Ω).
Composing it with (−∆)−1 , it follows that this subsequence in H01 (Ω) converges
to u0 ∈ M and
I 0 (u0 ) = I˜0 (u0 ) = lim I˜0 (uj ) = 0.
Thus, u0 is the desired nontrivial critical point.
Exercises
√
1. Let R > a > 0 and M = {u ∈ C 1 ([−a, a]) | u(±a) = R2 − a2 },
Z a p
u
I(u) = 1 − u̇2 − dt.
−a R
(1) Compute the first and second variations of I.
(2) Write the Euler–Lagrange
√ equation.
(3) Verify that u0 = R2 − t2 is a weak minimal solution.
(4) Write the Jacobi operator along u0 .
2. Find inf{I(u) | u ∈ M } for each of the following:
(1)
Z 2 p
I(u) = t 1 + u̇2 dt,
1
(2)
Z b
I(u) = π u2 dr,
a
Z b p
1
M = u ∈ C0 ([a, b]) 2π u 1 + u̇2 dt = c .
a
1 1 1 n N
3. Let V = C (R , R ). ∀ (u, p) ∈ R × R , let
N
1X 2 X
L(u, p) = p − V ((ui − uj )2 ).
2 i=1 i
i6=j
Lecture 15
In this lecture, we introduce the theory of finding critical points other than
minima (or maxima). This theory carefully examines the changes of topologi-
cal structure taking place in the level sets of a functional; subsequently provides
criteria to the existence of critical points. Such theory is based on the ideas and
machinery from both algebraic and differential topology. Since the 1970s, critical
point theory has undergone a rapid development; in particular, it has found pro-
found applications in partial differential equations and dynamical systems with
variational structures.
We do not require the reader to possess the needed topological background;
instead, we would like to expose the reader to some of the most fundamental
and commonly used critical points theorems, such as the Mountain Pass Theo-
rem. To make this more accessible, we will take on a more geometrically intuitive
approach.
The standard critical point theory utilizes the gradient flow to accomplish the
deformation between the level sets of a functional. However, this treatment is
beyond the scope of this book and does not fit well with our current content, we
will adopt a more direct approach - introducing these critical point theorems based
on the Ekeland variational principle.
The following intuitive example illustrates the basic idea used in analyzing
saddle points, a special kind of critical point.
Imagine the following scenario: in a valley surrounded by mountains, if a
person starting from a point p1 outside the valley wants to reach a point p0 in
the valley, the optimal path would be along which the highest point is always no
211
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 212
higher than the highest point on a nearby path. The highest point on this path is
likely to be a saddle point - a critical point which is neither the maximum nor the
minimum (see Figure 15.1).
Fig. 15.1
Let
be the set of all paths connecting the two points and let
≤ Cd(l1 , l2 ),
where C is a constant, depending only on f 0 , l1 , and l2 . Using the Ekeland varia-
tional principle, we obtain a sequence {ln } ⊂ Γ such that
1
c ≤ I(ln ) < c + , (15.4)
n
1
I(l) > I(ln ) − d(l, ln ), l 6= ln , n = 1, 2, . . . (15.5)
n
Let
M (l) = {t ∈ [0, 1]| f ◦ l(t) = I(l)}.
Then M is a non-empty compact set and M ⊂ (0, 1). To see this, suppose t0 ∈
M (l) ∩ {0, 1}, then
f ◦ l(t0 ) = max f ◦ l(t) ≥ inf f = α,
t∈[0,1] ∂Ω
but
f ◦ l(t0 ) ≤ max{f (p0 ), f (p1 )} < α,
a contradiction.
Denote Γ0 = {ψ ∈ C([0, 1], X)| ψ(i) = θ, i = 0, 1}, it is a closed linear
subspace of C([0, 1], X) with norm
kψkΓ0 = maxt∈[0,1] kψ(t)k.
By (15.4), ∀ h ∈ Γ0 , khkΓ0 = 1, ∀ λj ↓ 0, ∀ ξj ∈ M (ln + λj h), we have
1
λ−1
j [f ◦ (ln + λj h)(ξj ) − f ◦ ln (ξj )] ≥ − .
n
Since {ξj } ⊂ [0, 1], there exists a subsequence, still denoted ξj , converging to ηn .
The latter depends on ln , λj , and h. Taking the limit, it yields
1
df (ln (ηn ), h(ηn )) ≥ − . (15.6)
n
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 214
Let
m
X
v = v(ξ) = ρi (ξ)vηi .
i=1
Since M (ln ) ⊂ (0, 1), v ∈ Γ0 , kvk ≤ 1. In fact, we can choose a finite covering
and some ξ ∗ ∈ M (ln ) such that there is only one i0 with ξ ∗ ∈ Oηi0 . Hence,
kvkΓ0 = 1, and (15.8) implies
1
df (ln (ξ), v(ξ)) < − , ∀ ξ ∈ M (ln ),
n
contradictory to (15.6). The proof is now complete.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 215
Remark 15.1 The Mountain Pass Theorem was stated in the above version by
A. Ambrosetti and P. Robinowitz in 1974. Its generalizations as well as variations
have since been applied to solving various variational problems. The theorem
originated from the Wall Theorem, discovered by M. Morse while studying the
multiple-solution problem arisen in minimal surfaces.
The proof of the Mountain Pass Theorem by means of Ekeland variational
principle was independently provided by S. Z. Shi (The Chinese Annal of Math-
ematics 1, 1985, 348–355) and J. P. Aubin and I. Ekeland (Applied Nonlinear
Analysis, John Wiley and Sons, 1984).
Remark 15.2 We comment that in Theorem 15.1, the Palais–Smale condition
indeed plays a crucial role. Without it, Brezis and Nirenberg gave the following
counterexample:
In R2 , consider the function
f (x, y) = x2 + (1 − x)3 y 2 .
Let c = inf x2 +y2 = 41 f (x, y) > 0, it actually has a valley Ω = {(x, y) ∈
R2 | f (x, y) ≤ c}, while f (0, 0) = 0 and f (4, 1) = −11, but by direct com-
putation, f has the only critical point (0, 0).
The geometric structure involved in the Mountain Pass Theorem is a special
case of the more general linking structure.
Definition 15.1 Let X be a Banach space, Q ⊂ X be compact manifold with
boundary ∂Q, and S ⊂ X be a closed subset. We say ∂Q and S link, if
(1) ∂Q ∩ S = ∅,
(2) for any continuous ϕ : Q → X satisfying ϕ|∂Q = id|∂Q , we have ϕ(Q) ∩
S 6= ∅.
Link is a property depicting how two sets intersect with one another under contin-
uous deformations, therefore it is a topological property.
Example 15.1 In the Mountain Pass Theorem, Q = {tx0 + (1 − t)x1 | t ∈ [0, 1]},
S = ∂Ω and ∂Q = {x0 , x1 }, hence ∂Ω and S link.
Example 15.2 Let X be a Banach space, X1 be a finite dimensional linear sub-
space, X2 be its complementary space, i.e. X = X1 ⊕ X2 . Let
S = X2 , Q = BR ∩ X1 ,
where BR is the closed ball in X centered at θ with radius R > 0, thus
∂Q = {x ∈ X1 | kxk = R}
(see Figure 15.2).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 216
Fig. 15.2
Fig. 15.3
Following the same proof of the Mountain Pass Theorem, we can also prove
the following.
Theorem 15.2 Let X be a Banach space, Q ⊂ X be compact manifold with
boundary ∂Q, and S ⊂ X be a closed subset linked with ∂Q. Let f ∈ C 1 (X, R1 ).
Suppose there exist α < β such that
sup f (x) ≤ α < β ≤ inf f (x). (15.9)
x∈∂Q x∈S
Let
Γ = {ϕ ∈ C(Q, X) | ϕ|∂Q = id|∂Q } (15.10)
and
c = inf max f (ϕ(ξ)). (15.11)
ϕ∈Γ ξ∈Q
15.2 Applications
Both the Mountain Pass Theorem and the linking theorem have numerous ap-
plications in variational problems. However, we will only exhibit their usefulness
by a few examples to whet the reader’s appetite.
Example 15.4 Given a periodic continuous function a with period T > 0 on the
real line. Define the potential function
1 a(t)
V (t, x) = − |x|2 + |x|p+1 . (15.12)
2 p+1
Suppose p > 1, a(t) ≥ α > 0. Find a non-trivial T -periodic solution x ∈
C 2 ([0, T ], RN ) of the system
ẍ + Vx (t, x) = 0. (15.13)
1
We define, on the space Hper ((0, T ), RN ) := {x ∈ H 1 ((0, T ), RN ) | x(0) =
x(T )}, the functional
Z T
1 2 2 a(t) p+1
I(x) = (|ẋ| + |x| ) − |x| dt. (15.14)
0 2 p+1
It is clear that (15.13) is its E-L equation.
We claim x = θ is a local minimum of I, therefore it is a trivial solution of
(15.13). First, we note
I(θ) = 0.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 220
c
Since 1 − λk+1 > 0, there exist ρ > 0 and β > 0 such that
I(x) ≥ β, for x ∈ ∂Bρ (θ) ∩ X2 . (15.20)
Let S = ∂Bρ (θ) ∩ X2 .
On X1 , we always have
Z T Z T
1 2 2 1 2
I(x) ≤ (|ẋ| − c|x| ) dt ≤ (λk − c)|x| dt ≤ 0.
0 2 0 2
We now take e = cos 2(k+1)πt
T ⊗ e1 , where e1 = (1, 0, . . . , 0) ∈ RN . Since all
norms on a finite dimensional space are equivalent, on the space X1 ⊕ R1 e, as
kxk → ∞, the following holds uniformly
Z T Z T
2 α
I(x) ≤ (λk+1 − c) |x| dt − |x|p+1 dt → −∞.
0 p+1 0
Next, we take
Q = {(x1 , t) ∈ X1 × R1+ | kx1 k2 + t2 = R2 }.
For R > ρ > 0 sufficiently large,
I|∂Q ≤ 0. (15.21)
Additionally, according to Example 15.3, we see that S and ∂Q are linked.
Lastly, we verify the PSc condition. The argument is similar to that of Exam-
ple 15.5. The only difference is that it is less direct in verifying the boundedness
1
of the PS sequence in Hper ((0, T ), RN ). Suppose {xj } ⊂ Hper
1
((0, T ), RN ) sat-
isfies I(xj ) → c and I 0 (xj ) → θ, then
Z T
[|ẋj (t)|2 − c|xj (t)|2 ]dt = C1 + o(kxj k)
0
and
Z T
a(t)|xj (t)|p+1 dt = C2 + o(kxj k).
0
By Hölder’s inequality, we have
Z T Z T 2 Z
p+1 T p−1
p+1
2
2 p+1 − p−1
|xj (t)| dt ≤ a(t)|xj (t)| dt a (t)dt .
0 0 0
Combining the three inequalities above, it follows that
Z T
|ẋj (t)|2 dt = C3 + o(kxj k),
0
whence kxj k ≤ C4 .
According to (15.20), (15.21), and Theorem 15.2, we established that I has a
critical value c ≥ β > 0, which corresponds to a non-trivial solution of (15.13).
Similar methods can also be applied to partial differential equations.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 224
Example 15.6 Let Ω ⊂ Rn be a bounded domain. Let 1 < p < 2∗ −1, a ∈ C(Ω̄),
and a(x) ≥ α > 0, ∀ x ∈ Ω. Find a weak solution of the equation
−∆u(x) = a(x)|u(x)|p−1 u(x), (15.22)
where u ∈ H01 (Ω). On H01 (Ω), define the functional
Z
1 2 a(x) p+1
I(u) = |∇u(x)| − |u(x)| dx; (15.23)
Ω 2 p+1
(15.22) is its E-L equation.
Clearly, u = θ is a trivial solution. We want to find a non-trivial critical point
of I. By Poincaré’s inequality,
Z Z
1 1
I(u) ≥ |∇u(x)|2 dx − o(kuk2 ) ≥ |∇u(x)|2 dx as |uk → 0,
Ω 2 4 Ω
so for r > 0 sufficiently small,
r2
I|∂Br (θ) ≥ .
4
Choose any nonzero function ϕ ∈ H01 (Ω),
Z Z
1 a(x)
I(tϕ) = t2 |∇ϕ(x)|2 dx − tp+1 |ϕ(x)|p+1 dx → −∞.
Ω 2 Ω p + 1
If we take p0 = θ and p1 = tϕ, then for t sufficiently large, we will have the
desired geometric structure in the Mountain Pass Theorem.
It remains to verify the PSc condition. Suppose {uj } satisfies I(uj ) → c and
I 0 (uj ) → θ in H01 (Ω), we want to show that it has a convergent subsequence. In
fact, we have
Z
1 a(x)
|∇uj (x)|2 − |uj (x)|p+1 dx → c (15.24)
Ω 2 p+1
and
Z
[∇uj (x)ϕ(x) − a(x)|uj (x)|p−1 uj (x)ϕ(x)]dx = o(kϕk), ∀ ϕ ∈ H01 (Ω).
Ω
(15.25)
Substituting ϕ = uj into the equation, it yields
Z
[|∇uj (x)|2 − a(x)|uj (x)|p+1 ]dx = o(kuj k). (15.26)
Ω
Lecture 16
We begin this lecture with the motion of a pendulum (see Figure 16.1)
to introduce the concepts of periodic solutions, homoclinic and heteroclinic
orbits.
Fig. 16.1
ϕ̈ + α sin ϕ = 0,
where α = ω02 = gl .
227
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 228
Fig. 16.2
The orbit: (x(t), ẋ(t)) → (±π, 0) as t → ±∞, connecting (−π, 0) and (π, 0)
through the points (0, ±2ω0 ) in the upper or lower half plane is called a hetero-
clinic orbit.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 229
ml2 B 2
If in addition, we assume |B| > 2ω0 , the solution with energy E = 2
passing through the initial points (x, y) = (0, B) is given by
q
y = ± B 2 − 2ω02 (1 − cos x).
These are periodic curves in the upper and lower half planes respectively; if
we let
Z x
ds
t(x) = p
0 B − 2ω02 (1 − cos s)
2
and
τ (B) = t(2π),
then
τ (B)
p(x) = t(x) −
2π
is a 2π-periodic function. We have
(
x(t + τ (B)) = x(t) + 2π,
y(t + τ (B)) = y(t).
In this sense, we call it a periodic solution of the second kind.
In a dynamical system, an orbit x(t) → p as t → ±∞ connecting a saddle
equilibrium point p to itself is called a homoclinic orbit (see Figure 16.3).
Fig. 16.3
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 230
For example, let N = 1 and V (t, u) = −|u|p (1 + ε sin t), where p > 1. Since
−Vuu = −p(p − 1)|u|p−2 (1 + ε sin t) > 0,
I is convex. Since I is weakly lower semi-continuous, it is weakly sequentially
lower semi-continuous. We prove I is bounded below and coercive as follows.
Since u 7→ −V (t, u) is continuous and convex, F is a continuous and convex
function on RN . So ∃ x0 ∈ RN such that F achieves its minimum at x0 . We then
have
Z T
0 = F 0 (x0 ) = − Vu (t, x0 ) dt. (16.3)
0
RT
∀ u ∈ HT1 (0, T ), we have the decomposition u = ũ + ū, where ū = T1 0
u dt.
Combining (16.3) and (16.4), we have
Z T Z T
− V (t, u(t))dt ≥ − (V (t, x0 ) + Vu (t, x0 )(u(t) − x0 ))dt
0 0
Z T
=− (V (t, x0 ) + Vu (t, x0 )(u(t) − ū))dt. (16.5)
0
Let
! 12
Z T Z T
c1 = V (t, x0 ) dt and c2 = |Vu (t, x0 )|2 dt ,
0 0
then
Z T Z T
1 ˙ 2−
I(u) ≥ |ũ| [V (t, x0 ) + Vu (t, x0 )ũ(t)]dt
2 0 0
! 12
Z T Z T
1 ˙ − c1 − c2
2 2
≥ |ũ| |ũ| .
2 0 0
RT RT
By Wirtinger’s inequality, |ũ|2 ≤ c23 ˙ 2 , it follows that
|ũ|
0 0
12
1 T
Z Z
I(u) ≥ ˙ 2 − c1 − c2 c3
|ũ| ˙2
|ũ|
2 0
Z
1 ˙ 2 − c4 .
≥ |ũ|
4
Consequently, the coerciveness of I is determined by whether ū can be bounded
by I(u). By convexity,
u(t) + (−ũ(t))
−V (t, ū/2) = −V t,
2
1
≤ − (V (t, u(t)) + V (t, −ũ(t))),
2
it follows that
1 T ˙2
Z Z T Z T
ū
I(u) ≥ |ũ| − 2 V t, dt + V (t, −ũ(t))dt.
2 0 0 2 0
RT
Since kũk∞ ≤ CkũkH 1 is bounded by I(u), − 0 V (t, −ũ(t))dt is also bounded
by I(u), whence
Z T
ū
− V t, ≤ c5 I(u) + c6 .
0 2
By (16.2), ū is bounded by I(u); this proves that I is coercive.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 232
1
From this, we see that Eq. (16.1) has a solution u ∈ Hper ([0, T ]). By regular-
2
ity, u ∈ C .
Lastly, we examine the periodic condition. By the embedding theorem and
1
u ∈ Hper ([0, T ]), u(0) = u(T ). It remains to show u̇(0) = u̇(T ). In the integral
form of the E-L equation, we choose a period-T function ϕ ∈ C ∞ ([0, T ]), then
we have
Z T
0= [u̇ϕ̇ − Vu (t, u)ϕ]dt
0
Z T
= [−ü − Vu (t, u)]ϕdt + u̇ϕ|T0
0
= u̇(T )ϕ(T ) − u̇(0)ϕ(0).
Since ϕ(0) = ϕ(T ) is arbitrary, we have shown
u̇(T ) = u̇(0).
Theorem 16.1 Under the assumption of (16.2), Eq. (16.1) has a C 2 periodic
solution.
II. V is continuous and periodic. Suppose there exist linearly independent
vectors e1 , . . . , eN ∈ RN such that
V (t, u + ei ) = V (t, u), ∀ (t, u) ∈ [0, T ] × RN . (16.6)
We must again verify
Z T
1
I(u) = |u̇(t)|2 + V (t, u(t)) dt
0 2
is weakly sequentially lower semi-continuous and coercive.
According to Morrey’s Theorem in Lecture 11, I is for certain weakly sequen-
tially lower semi-continuous.
Since V is continuous and satisfies (16.6), there exists a constant C such that
|V (t, u)| ≤ C.
However, we cannot deduce the coerciveness of I directly from
1 T 2
Z
I(u) ≥ |u̇| − CT.
2 0
Thus, we choose the decomposition
u = ũ + ū,
1
RT
where ū = T 0
u(t) dt. Denote
X = {u ∈ HT1 ([0, T ]) | ū = 0},
1
then X is a closed linear subspace of Hper ([0, T ]).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 233
Although {ūj } could be unbounded, but after removing the integer parts, it be-
(j) (j)
comes bounded, i.e. ∃ (λ1 , . . . , λn ) ∈ Zn such that
N
(j)
X X
ūj + λ i ei
≤ kei k , A.
n
i=1 R i=1
P (j)
Let vj = ũj + (ūj + then I(vj ) = I(uj ), whereas {vj } is a bounded
λi ei ),
minimizing sequence. Consequently, it has a weakly convergent subsequence, still
denoted {vj } such that
vj * u∗ .
The same argument can be used to show u∗ is a minimum. Likewise, the same
steps as above can be used to verify u∗ is a periodic solution. Lastly, by regularity,
u∗ ∈ C 2 (R1 ).
Theorem 16.2 Under the assumption of (16.6), Eq. (16.1) has a C 2 periodic
solution.
III. Periodic solutions on the torus
Using the same methods as above, we can study the periodic solutions of the
E-L equation of a functional defined on the torus T N = RN /ZN . Given a La-
grangian on the torus
L(t, u, p) : T × T N × RN → R1
which satisfies
L(t + Z, u + Zn , p) = L(t, u, p)
and the following conditions
−1
c ≤ Lpp ≤ c,
|L | + |Lpu | ≤ c(1 + |p|),
pt
|Lu | ≤ c(1 + |p|2 ),
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 234
min I(u).
u∈ Γ(θ,ξ)
Applying Morrey’s Theorem on any finite interval and by taking limits, it follows
that I is weakly sequentially lower semi-continuous.
Since I ≥ 0, it is bounded below.
1◦ We verify I is coercive. Since
Z
1
|u̇|2 dt ≤ I(u) ≤ C,
2 R1
it suffices to bound |u(0)|.
When u ∈ Γ(θ, ξ), we have
Z b
|u(b) − u(a)| ≤ |u̇(t)| dt
a
!1/2
Z b
≤ (b − a)1/2 |u̇(t)| dt .
a
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 236
As long as
−V (t, u(t)) ≥ M1 , (16.8)
we have
Z b
1 2
I(u) ≥ |u̇| − V (t, u) dt
a 2
1 |u(b) − u(a)|2
≥ + M1 |b − a|
2 |b − a|
p
≥ 2M1 |u(b) − u(a)|. (16.9)
From assumptions (3) and (4), fixing ε > 0, there exists M1 > 0 such that when
u(t) 6∈ Bε (θ) ∪ Bε (ξ),
(16.8) holds.
Noting that V is T -periodic in t, so by letting
(τi u)(t) = u(t − iT ), ∀ i ∈ Z,
we have I(τi u) = I(u).
In order to bound |u(0)|, we use the fact that V is τi invariant. ∀ u ∈ Γ(θ, ξ),
∀ > 0, ∃ i0 such that τ = τi0 satisfying
τ u(t) ∈ Bε (θ), t < 0, τ u(0) ∈ ∂Bε (θ).
Replacing u by ũ = τ uj , it is immediate that kũk = ε. This shows I is coercive.
Moreover, replacing any u ∈ Γ(θ, ξ) by ũ, then I(u) = I(ũ).
2◦ Next, we verify Γ(θ, ξ) is weakly closed with respect to the minimizing
sequence. That is, we want to prove: if {uj } ⊂ Γ(θ, ξ), uj * u and I(uj ) →
inf Γ(θ,ξ) I, then u ∈ Γ(θ, ξ).
Since u is the weak limit of the minimizing sequence and I is weakly lower
semi-continuous, it follows that
I(u) ≤ inf I.
Γ(θ,ξ)
∞ 1 n
By (16.9), u ∈ L (R , R ). Hence, it has a ω-limit point, i.e. ∃ ti → +∞,
∃ α ∈ RN such that u(ti ) → α.
1) We claim the limit point is unique. Suppose ∃ t0i → +∞ such that u(t0i ) →
β. Since I(u) < ∞, it follows from (16.9) that
Z 0
0 1 ti 1 2
|u(ti ) − u(ti )| ≤ √ |u̇| − V (t, u)dt → 0,
2M1 ti 2
hence α = β.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 237
Consequently,
Z ∞ Z ∞
I(u) ≥ − V (t, u(t))dt ≥ M1 = ∞,
t1 t1
a contradiction.
3) We claim α = ξ. Suppose not, then α = θ. ∀ δ > 0, ∃ tδ > 0 such that
when t ≥ tδ , we have u(t) ∈ Bδ (θ).
Since uj * u ∈ Ê, according to the exposition in the last paragraph of 1◦ , we
may choose uj such that uj (t) ∈ Bε (θ) for t ≤ 0 and uj (0) ∈ ∂Bε (θ). Choose
δ < ε/4 and t1 > tδ + 1 such that u(t1 ) ∈ Bδ (θ).
∃ j0 such that for j ≥ j0 , kuj − ukL∞ ([0, t1 ]) < δ, hence uj (t1 ) ∈ B2δ (θ).
It follows from (16.9) that
Z ∞
1 2 εp
I(uj ) ≥ |u̇j | − V (t, uj ) dt + 2M1 .
t1 2 2
We construct the sequence
0, t < t1 − 1,
vj (t) = (t − t1 + 1)uj (t1 ), t ∈ [t1 − 1, t1 ],
uj (t), t > t1 ,
then vj ∈ Γ(θ, ξ) and
Z t1 Z ∞
1 2 1 2
I(vj ) = |uj (t1 )| − V (t, vj (t)) dt + |u̇j | − V (t, uj (t)) dt.
t1 −1 2 t1 2
However,
Z ∞
1 2 εp
|u̇j | − V (t, uj (t)) dt ≤ I(uj ) − 2M1
t1 2 2
and
Z t1
1
|uj (t1 )| − V (t, vj (t)) dt ≤ 2δ 2 + max (−V (t, u)).
2
t1 −1 2 |u|≤2δ
we then obtain
εp
I(vj ) ≤ I(uj ) −
2M1 ,
4
but I(uj ) → inf Γ(θ,ξ) I, which contradicts the fact that {vj } ⊂ Γ(θ, ξ).
We have successfully established Γ(θ, ξ) is weakly closed with respect to the
minimizing sequence. Consequently, the functional I achieves its minimum, say
u∗ . By regularity, u∗ ∈ C 2 (RN ).
3◦ Lastly, we verify u̇∗ (±∞) = 0.
We already know when t > t1 , u∗ (t) ∈ Bε (ξ). Thus, by assumptions (3) and
(1), there exist β1 > 0 and β2 > 0 such that
−V (t, u∗ (t)) ≥ β1 |u∗ (t) − ξ|2 ,
|Vu (t, u∗ (t))| ≤ β2 |u∗ (t) − ξ|.
It folows that
Z ∞ Z ∞
∗
β1 2
|u (t) − ξ| dt ≤ − V (t, u∗ (t)) dt ≤ I(u∗ )
t1 t1
and
Z ∞ Z ∞ Z ∞
|ü∗ |2 dt = |Vu (t, u∗ (t))|2 dt ≤ β22 |u∗ (t) − ξ|2 dt,
t1 t1 t1
R∞
from which, we can deduce t1 |ü | dt < ∞. Together with u̇∗ ∈ L2 (R1 ), it
∗ 2
1
We begin by defining for ∀ k ∈ N, the space Xk = H2kT ([−kT, kT ]) with
norm
Z kT
kxk2k = (|ẋ(t)|2 + |x(t)|2 )dt
−kT
and the functional
Z kT
1
Ik (x) = kxk2k − a(t)|x(t)|µ dt.
2 −kT
According to Example 15.5 in Lecture 15, Ik possesses the geometric structure
described in the Mountain Pass Theorem, that is,
1
Ik (0) = 0, Ik (x) = kxk2k + o(kxk2k ), ∃ ϕ ∈ Xk \{0} such that I(tϕ) → −∞.
2
Furthermore, Ik satisfies the Palais–Smale condition. Thus, there exists a critical
point xk ∈ Xk , i.e.
Z kT
[ẋk ϕ̇ + xk ϕ − µa|xk |µ−2 xk ϕ]dt = 0, ∀ ϕ ∈ Xk , (16.110 )
−kT
with
ck = I(xk ) > 0.
Noting that a is 2T -periodic, while xk (t) is 2kT -periodic, xk (t − jT )|[−kT,kT ]
(j ∈ Z) are also solutions of (16.10) with the same period and the same critical
value. Therefore, we can choose such a xk that
max xk (t) = max xk (t). (16.12)
t∈[0,T ] t∈[−kT,kT ]
and
ϕ1 (t), t ∈ [−T, T ],
zk (t) =
0, t ∈ [−kT, kT ]\[−T, T ].
Then
zk ∈ Xk , Ik (zk ) = I1 (ϕ1 ) ≤ 0.
Hence, we have the estimate
ck ≤ max Ik (tzk ) = max I1 (tϕ1 ) = M.
t∈[0,1] t∈[0,1]
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 240
we obtain that
Z kT
µ
−1 a(t)|xk (t)|µ dt = ck .
2 −kT
Thus,
Z kT
2
kxk2k = 2 Ik (zk ) + µ
a(t)|xk (t)| dt ≤ 2 + ck ≤ M1 .
−kT µ−2
◦
3 We want to prove there exists a solution y ∈ H 1 (R1 ). Notice that for any
1
x ∈ Hloc , we have the embedding inequality
Z t+ 21
2
|x(t)| ≤ 2 (ẋ(r)2 + x(r)2 )dr.
t− 12
This means y ∈ H 1 (R1 ) and it satisfies Eq. (16.11) on the entire real line.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 241
whence
max1 |y(t)| ≥ δ.
t∈R
Lecture 17
17.1 Geodesics
γ : [0, 1] → M, γ̇ 6= θ
Fig. 17.1
243
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 244
L(γ)2 ≤ 2E(γ),
L(γ)2 = 2E(γ).
then
p
inf L(γ) = inf 2E(γ).
γ∈Γ γ∈Γ
Finding the curve with the shortest distance connecting the points P0 and P1 is
equivalent to finding the curve connecting P0 and P1 with minimal energy.
Thus, the curve γ with the shortest arclength also minimizes the energy.
Consequently, γ satisfies the E-L equation of E:
d
Lp (t, u(t), u̇(t)) = Lu (t, u(t), u̇(t)) ,
dt
Pn
where L(t, u, p) = 12 i,j=1 gij (u)pi pj . Namely,
n
X d 1 ∂gkj k j
(gij (u(t))u̇(t)) − (u(t)) u̇ (t)u̇ (t) =0
j=1
dt 2 ∂ui
n n n
X
j
X ∂gij k j X ∂gkj k j
⇐⇒ 2gij ü + 2 u̇ u̇ − u̇ u̇ = 0
j=1
∂uk ∂ui
k=1 k=1
n n n
1 XXX
⇐⇒ üi + g il (2glj,k u̇k u̇j − gkj,l u̇k u̇j ) = 0
2
l=1 k=1 j=1
1 Pn Pn Pn
⇐⇒ üi + g jl (glj,k + gkl,j − gjk,l )u̇k u̇j = 0
2 l=1 k=1 j=1
X n
⇐⇒ üi + Γijk u̇k u̇j = 0.
k,j=1
This is precisely the geodesic equation. So the curves of the shortest arclengths
are themselves geodesics!
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 246
Fig. 17.2
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 248
Recall the square of the area determined by the two vectors Xx and Xy is
Before applying variational methods to study minimal surfaces, let us first insert
certain condition on Γ such that
Xx ∧ Xy dx ∧ dy = Xu ∧ Xv du ∧ dv.
g : R2 → R2 , w 7→ z
|gx |2 − |gy |2 = gx · gy = 0.
A direct calculation shows that although the Diriihlet integral is not diffeomorphi-
cally invariant, it is conformally invariant!
Z Z
|∇w X|2 dx ∧ dy = |∇z X|2 du ∧ dv.
D D
In order to avoid disrupting the proof of existence, we shall postpone the proof
of the above theorem until the end of this lecture.
The advantage of minimizing the Dirichlet integral instead of the area func-
tional is that the conformal group is far smaller than the diffeomorphism group.
The conformal group is the set of all conformal transformations forming a
group. It is generated by three real parameters:
iφ w + a 1
G = g(w) = e |a| < 1, φ ∈ R .
1 + āw
It is worth noting when a = aj → 1, gj (w) is concentrated at one point on the
unit circle. This means ∀ X ∈ C(Γ), the closure of the orbit {X ◦ g | g ∈ G} in
the weak topology on H 1 (D, Rn ) contains constant functions. However, the latter
cannot be a weakly monotone parametrization of Γ, so C(Γ) cannot be a weakly
closed subset in H 1 . Even replacing the functional by the Dirichlet integral, it is
still impossible to get around the weak closedness issue of C(Γ).
To rectify this, we further reduce the influence of the conformal group G: by
imposing additional restriction on C(Γ), we hope to “mod out” the action of G.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 250
2kπ
We call X e 3 i = Pk for 0 6 k 6 2 the three-point condition. Since
we instead find
min{D(X), X ∈ C ∗ (Γ)}.
(2)
Z
d −1
d 1 ∇z (X ◦ gε−1 (z))2 du ∧ dv|ε=0 = 0,
D(X ◦ gε , gε (D))
=
dε ε=0 dε 2 gε (D)
∆X = 0 in D,
ḡε−1 satisfying
However, ḡε (D) is simply connected, so by the Riemann Mapping Theorem, there
exists a conformal mapping hε : D → ḡε (D). Let X̃ε = X̄ε ◦ hε , then X̃ε ∈
C(Γ). Furthermore, by the conformal invariance of the Dirichlet integral, we have
This is a contradiction!
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 251
From (2) and a similar argument used in Example 8.4 in Lecture 8, it follows
that on D,
|Xx |2 = |Xy |2 ,
Xx · Xy = 0,
which is the second equation in (17.1).
Suppose dg dε |ε=0 =P
ε
τ , τ = (τ 1 , τ 2 ). Applying Noether’s formula in
1 n
Lecture 8 to L(p) = 2 ( i=1 (pi1 )2 + (pi2 )2 ), it yields
Z
div(Lτ + Lpi · ϕi ) dxdy = 0,
D
i
where ϕ = −(Xxi τ 1 + Xyi τ 2 ).
Since
Lτ +Lpi · ϕi
1
= (|Xx |2 +|Xy |2 )(τ 1 , τ 2 )−Σi (Xxi τ 1 +Xyi τ 2 )(Xxi , Xyi )
2
1 1
= − (|Xx |2 −|Xy |2 )τ 1 −Xx · Xy τ 2 , −Xx · Xy τ 1 + (|Xx |2 −|Xy |2 )τ 2 ,
2 2
it follows that
1
div(Lτ + Lpi · ϕi ) = (|Xx |2 − |Xy |2 )(τy2 − τx1 ) − Xx · Xy (τy1 + τx2 ).
2
We have thus proved
d
D(X ◦ gε−1 , gε (D))|ε=0
dε Z
d 1 ∇z (X ◦ gε−1 )2 du ∧ dv
=
dε 2 gε (D)
Z
1
=− [(|Xx |2 − |Xy |2 )(τx1 − τy2 ) + 2Xx · Xy (τy1 + τx2 )] dxdy.
2 D
Since λ = τx1 − τy2 and µ = τy1 + τx2 are arbitrary, we conclude
|Xx |2 − |Xy |2 = Xx · Xy = 0 a.e. in D.
We have shown the minimizing function of the energy functional is indeed a
generalized solution of the minimal surface equation.
• Replacing Rn by a more general Riemannian manifold (N n , h), while Γ is
an embedded Jordan curve in N . The corresponding equation becomes
trace(∇dX) = 0,
h(Xx , Xx ) − h(Xy , Xy ) = h(Xx , Xy ) = 0.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 252
where
Z 2π
1 2π
Z
1
a0 = X̂(eiθ )dθ, ak = X̂(eiθ ) cos kθdθ,
2π 0 π 0
1 2π
Z
bk = X̂(eiθ ) sin kθdθ, k = 1, 2, . . . ,
π 0
X∞
k(|ak |2 + |bk |2 ) < ∞,
k=1
then
∞
1 X
X(reiθ ) = a0 + (ak cos kθ + bk sin kθ)rk . (17.3)
2
k=1
as a functional defined on the set C ∗ (Γ) ⊂ X = {X̂ = X|∂D | D(X) < ∞}; on
X , if we define the norm to be
∞ 12
1 2 X 2 2
kX̂k = a + k(|ak | + |bk | ) ,
4 0
k=1
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 253
where ∂s denotes the tangential derivative, Cρ0 is the arc whose center lies on the
unit circle and whose radius is ρ (see Figure 17.3).
Fig. 17.3
Proof By Fubini’s Theorem, for almost all ρ ∈ (0, 1), |∂s X| ∈ L2 (Cρ ) and
Z
2D(X) ≥ |∇X|2 dz
(B√δ (w)\Bδ (w))
√ Z
Z δ
≥ |∂s X|2 dsdρ
δ Cρ
√
Z Z δ
2 dρ
≥ ess inf√ ρ |∂s X| ds .
δ<ρ< δ Cρ δ ρ
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 254
Moreover, from
√
Z δ
dρ 1
= | ln δ|,
δ ρ 2
the assertion now follows.
As a consequence, we have the following.
Lemma 17.2 ∀ M > 0, the set
n o
CM = X̂ : ∂D −→ Γ X̂ = X|∂D , X ∈ C ∗ (Γ), D(X) ≤ M
is equicontinuous.
Proof Since Γ is a Jordan curve, ∀ ε > 0, ∃ d > 0, ∀ P 6= P 0 ,
|P − P 0 | < d
=⇒ Among the two components of Γ\{P, P 0 }, at least one has diam < ε.
1◦ . ∀ ε > 0, assume ε < mini6=j |Pi − Pj |. Choose d > 0 as above, choose
δ > 0 such that |8πM 2
ln δ| < d . ∀ X ∈ CM , ∀ w ∈ ∂D, set Cρ = D ∩ ∂Bρ (w). By
the Courant–Lebesgue lemma,
2
|X(w1 ) − X(w2 )|2 ≤ [diam X(Cρ )]
Z 2
≤ |∂s X|ds
Cρ
Z
≤ L(Cρ ) |∂s X|2 ds
Cρ
8πM
≤ , ∀ w1 , w2 ∈ Cρ ,
| ln δ|
where L(Cρ ) is the arclength of Cρ . Hence, {w1 , w2 } = ∂D ∩ ∂Bρ (w) =⇒
among the two components of Γ\{X(w1 ), X(w2 )}, at least one has diameter less
than ε.
2◦ Choose δ > 0 such that ∀ z ∈ ∂D, at least two values of k satisfy
√
2πk
z − e 3 i ≥ δ.
Suppose ∂D\{w1 , w2 } = C1 ∪ C2 , C1 ∩ C = {w
2 2kπ 1 , w2 }, then at least one of C1 ,
C2 , say C1 , contains at most one point of e 3 i k = 1, 2, 3 . Hence,
diam X̂(C1 ) < ε.
∀ z, z 0 ∈ ∂D, when |z − z 0 | < δ, choose w0 ∈ ∂D and ρ > 0 such that z, z 0 ∈ C1 .
Thus,
|X̂(z) − X̂(z 0 )| ≤ diam X̂(C1 ) < ε =⇒ CM is equicontinuous.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 255
From which, we conclude ψ ∈ W 1,1 (∂D). This means the line integral on the
right-hand side of the equation can also be extended to g ∈ S̄. Namely, if {g j } ⊂
S, g ∈ S̄ such that g j * g, then
Z
2 |Xu × Xv |w=(g(z)) det(∇g) dxdy
D
Z
= lim [− [Gg (g1j ∂y g2j − g2j ∂y g1j ) − Hg (g1j ∂x g2j − g2j ∂x g1j )] dxdy
j→∞ D
Z
+ |Xu × Xv |w=(g(eiθ )) ψj0 (θ)dθ]
∂D
Z
= lim 2 |Xu × Xv |w=(g(z)) det(∇g j ) dxdy,
j→∞ D
0
where ψ (θ) = g1j ∂θ g2j − g2j ∂θ g1j .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 258
it follows that
inf D(X) ≤ inf A(X).
X∈C(Γ) X∈C(Γ)∩C 2 (D̄,Rn )
Lecture 18
The direct method for solving a variational problem consists of two main parts:
(1) the construction of a minimizing sequence,
(2) taking the limit of the minimizing sequence in order to obtain a solution.
L. Rayleigh and W. Ritz proposed a numerical method for finding a mini-
mizing sequence, known as the Rayleigh–Ritz method. Let M0 be a linear sub-
space
R of the functions space X over some domain Ω, I is the integral functional
Ω
L(x, u(x), ∇u(x)) dx on Ω, whose domain M = ϕ0 + M0 , where ϕ0 is a
function on Ω (determined by its inhomogeneous boundary value).
The idea of the Rayleigh–Ritz method is to choose a complete basis
{e1 , e2 , . . .} for M0 and let
Mn = ϕ0 + span{e1 , e2 , . . . , en } ⊂ M.
We then search, on Mn , for the minimum ϕn ∈ Mn of the functional I. Since
any function u in Mn can be expressed as
n
X
u = ϕ0 + ξi ei ξ = (ξ1 , . . . , ξn ) ∈ Rn (18.1)
i=1
Consequently, we can either directly apply optimization technique to find the min-
imum ϕn of I|Mn , or by solving the system of n homogeneous equations:
n
∂ X
I ϕ0 + ξi ei = 0, j = 1, 2, . . . , n. (18.3)
∂ξj i=1
259
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 260
Closely related to this idea is the Ritz–Galerkin method. However, this method is
not targeted at the functional I, rather at its E-L equation
Z
δI(u, ϕ) = (Lu (τ )ϕ(x) + Lp (τ )∇ϕ(x)) dx = 0, ∀ ϕ ∈ M0 ,
Ω
where
n
X n
X
τn (x) = x, ϕ0 (x) + ξi ei (x), ∇ϕ0 (x) + ξi ∇ei (x) .
i=1 i=1
namely,
n
X
aij ξi − bj = 0, j = 1, . . . , n. (18.7)
i=1
Since aij = aji for all i and j, (18.6) and (18.7) are identical to each other.
Remark 18.1 The Ritz–Galerkin method is not only applicable to the E-L equa-
tion of a functional, but also applicable to all differential equations whose weak
solutions are expressed in terms of integrals. For example, suppose aij , bj , and c
are continuous functions, we define a weak solution u ∈ H01 (Ω) of the equation
X n X n
∂ ∂ ∂
aij (x) u(x) + bj (x) u(x)+c(x)u(x) = f (x), x ∈ Ω
i,j=1
∂x j ∂x j j=1
∂x j
u ∂Ω = 0
(18.50 )
1
to be such that ∀ ϕ ∈ H0 (Ω), it satisfies the integral equation
Z n Xn
X ∂u(x) ∂ϕ(x) ∂u(x)
− aij (x) + bj (x) +c(x)u(x)−f (x) ϕ(x) dx = 0.
Ω i,j=1
∂xj ∂xi j=1
∂xj
Pn
By substituting u = k=1 ξk ek (x) and ϕ(x) = el (x), l = 1, 2, . . . , n, we can
again obtain a system of n linear equations in n variables.
{Pi − P0 | 1 ≤ i ≤ n}
ei (Nj ) = δij i, j = 1, 2, . . . , M.
In fact, for any simplex K ∈ J, the value of the basis function ei on K is com-
pletely determined by its value at each vertex (if the vertex lies on ∂Ω, then set the
value be zero for homogeneous Dirichlet boundary condition and set the value be
a fixed constant for inhomogeneous Dirichlet boundary condition).
In particular, taking K = (P0 , . . . , Pn ), Pi = (p1i , . . . , pni ) for i = 0, 1, . . . , n,
we define n + 1 functions via
(
vj (x) = xj j = 1, 2, . . . , n
v0 (x) = 1,
Fig. 18.1
where K = {P0 , . . . , Pn }.
Consequently, Π : C(Ω̄) → H01 (Ω) is a bounded linear operator with
kΠukH 1 ≤ CkukC(Ω̄) ,
.
where C = n sup1≤j≤M kej kH 1 .
Lemma 18.1 Let u ∈ C 2 (Ω̄), then
n2 (n + 1) h2J
k∇u − ∇(Πu)k1,∞ ≤ k∇2 uk∞ , (18.9)
2 ρJ
n2 (n + 1) 2 2
ku − Πuk∞ ≤ hJ k∇ uk∞ , (18.10)
2
where k · k1,∞ is the norm on W 1,∞ , k · k∞ is the norm on L∞ ,
hJ = sup hK , hK = diam(K),
K∈J
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 264
and
ρJ = inf ρK , ρK = sup{2R | BR (x) ⊂ K, x ∈ K}.
K∈J
Proof It suffices to consider the difference of ∇u(x) and ∇(ΠK u) on each sim-
plex K = {P0 , P1 , . . . , Pn }. We have the Taylor expansion of ∇(ΠK u):
n
X
∇(ΠK u)(x) = u(Pj )∇ej (x)
j=0
n
X 1 2 2
= u(x) + ∇u(x)(Pj − x) + ∇ u(ξx )(Pj − x) ∇ej (x),
j=0
2
(18.11)
where ξx ∈ k. Differentiating (18.8), we obtain
n
j
X ∂ei
δ = pji , j, k = 1, 2, . . . , n
k
∂xk
i=0
n
X ∂ei
0 = .
i=0
∂xk
where
1 2
Rj (x) = ∇ u(ξx )(Pj − x)2 .
2
Since
∂ei 1 1
kPj − xk ≤ diamK ≤ hJ , ∂xk ≤ ρK ≤ ρJ ,
and
|u − Πu|H 1 ≤ Cβ hJ |u|H 2 ,
and
h2K
|u − Πu|H 1 (K) ≤ C |u|H 2 (K) .
ρK
Thus,
X
ku − Πuk2L2 (Ω) = ku − ΠK uk2L2 (K)
K∈J
X
≤ C 2 h4K |u|2H 2 (K)
K∈J
≤ C 2 h4J |u|2H 2
and
X
|u − Πu|2H 1 (Ω) = |u − ΠK u|2H 1 (K)
K∈J
X h4K 2
≤ C2 |u| 2
ρ2K H (K)
K∈J
1
≤ C 2 h2J |u|2 2 .
β 2 H (K)
Remark 18.2 The geometric meaning of condition (18.12) is that in a simplicial
subdivision, the angle of each simplex cannot be arbitrarily small.
Remark 18.3 The advantage of adopting the interpolation method in construct-
ing approximating functions is evident not only in theoretical research but also in
practical computations. The reason is that the interpolation method only requires
the solution to be continuous, i.e. u ∈ C(Ω̄), however, the resulting approxima-
tion is in the sense of H 1 . In particular, for linear second order elliptic equations,
u ∈ H 2 (Ω). When n ≤ 3, by the embedding theorem, H 2 (Ω) ,→ C(Ω̄).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 266
it follows that
a(u − uh , v) = 0, ∀ v ∈ Vh .
In particular,
αku − uh k2 ≤ a(u − uh , u − uh )
= a(u − uh , u − v) + a(u − uh , v − vh )
≤ Cku − uh k ku − vk,
where v ∈ Vh . This means
C
ku − uh k ≤ ku − vk, ∀ v ∈ Vh ,
α
i.e.
C
ku − uh k ≤ min ku − vk.
α v∈Vh
We now return to address the linear second order elliptic equation (18.50 ) men-
tioned in Remark 18.1. Suppose α, β, γ > 0 satisfy β 2 < αγ and
Pn 2 n
(1) i,j=1 aij (x)ξi ξj ≥ α|ξ| , ∀ ξ = (ξ1 , . . . , ξn ) ∈ R , |aij (x)| ≤ C for
1 ≤ i, j ≤ n, ∀ x ∈ Ω,
(2) c(x) ≥ γ, ∀ x ∈ Ω,
(3) max1≤i≤n kbi k∞ < 2β.
Let δ = α − β 2 /γ, then δ > 0. We also let
Z Xn n
X
a(u, v) = aij (x)∂j u ∂i v + bi (x)∂i u · v + c(x)u · v dx,
Ω i,j i=1
Fig. 18.2
then it gives the steepest descent. The step length λi can then be computed by
minimizing the single variable function
φ(t) = f (xi + tσ i )
in the direction of σ i . This iterative algorithm is known as the gradient method,
which is outlined as follows.
1. Given an initial point x0 .
2. Starting from xi , use (18.14) to compute σ i ; then use
0 i
φ (λ ) = 0,
φ(λi ) < φ(λ), ∀ λ ∈ [0, λi ),
to compute λi > 0, where φ(t) is defined as in (18.15).
3. Use (18.13) to compute xi+1 based on xi , σ i , and λi .
If σ i+1 = 0, then stop and xi+1 is the desired point. Otherwise, continue.
It is not difficult to show, if f is coercive with only one minimum, then starting
from any initial point x0 , the gradient method will either reach the minimum in
finite steps, or it will produce a minimizing sequence whose limit is the minimum.
However, the gradient method is not ideal, since the rate of convergence could
be painfully slow. The following example clearly portraits such weakness. Let
1
f (x1 , x2 ) = x21 + x22 , ε > 0.
ε
This is a quadratic function whose level curves are all ellipses and whose mini-
mum is at (0, 0). Starting from xi = (xi1 , xi2 ), by direct calculation, we have
λi
i+1 i i
x = (1 + λ )x1 , 1 + xi2 ,
ε
where
− (xi1 )2 + ε12 (xi2 )2
i
λ = .
(xi1 )2 + ε13 (xi2 )2
So for ε > 0 very small,
(xi1 )2
xi+1 ≈ (1 − ε)xi1 , −ε2 .
xi2
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 270
Fig. 18.3
This is why the descending path zigzags slowly approaching (0, 0) (see Fig-
ure 18.3).
The conjugate gradient method emerged to compensate for this weakness. Its
main idea is in a small neighborhood of the minimum, the function f can be
approximated by its second order Taylor polynomial:
1
f (x) = f (x0 )+(∇f (x0 ), x−x0 )+ ∇2 f (x0 )(x − x0 ), x − x0 +o(kx−x0 k2 ).
2
For the quadratic function
1
c + (b, x) + (Ax, x), c ∈ R1 , b ∈ Rn , A ∈ R,
2
where A is positive definite, it has a unique minimum. When n = 2, its level
curves are ellipses. We will take advantage of the special feature of this quadratic
function to avoid the zigzagging descending path.
From analytic geometry, we know a given vector w = (w1 , w2 ) determines an
axis of a family of ellipses, while its conjugate axis is the trace of the midpoints
lying on the chords parallel to this vector. We call the direction u of the conjugate
axis the conjugate direction of w (see Figure 18.4).
We are inspired by a geometric fact: for a positive definite quadratic function
on the plane, regardless of the initial point x0 , given any direction w, the minimum
along w must be the midpoint of a chord in the ellipse parallel to w. The next step
is, if we can find the minimum in the conjugate direction u of w, then we can reach
the center of the family of ellipses in merely two steps, which coincides with the
minimum of the original function.
In order to extend this geometric fact to higher dimensions, we must be able to
analytically express the concept of “conjugate” direction. For simplicity, we take
b = 0.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 271
Fig. 18.4
(Awi , wj ) = 0, 1 ≤ i < j ≤ n.
Set
xi+1 = xi + λi σ i .
3. If ri+1 = 0, then stop and xi+1 is the desired minimum; otherwise, con-
tinue. For a quadratic function, the above iteration will terminate in at most n
steps.
In fact, it is not difficult to verify by mathematical induction: for m ≤ n, we
have
(1) if φ(t) = f (xi + tσ i ), then φ0 (λi ) = 0;
(2) span{σ 0 , . . . , σ m } = span{r0 , . . . , rm } = span{r0 , Ar0 , . . . , Am r0 };
(3) [σ i , σ j ]A = 0 for i 6= j and 1 ≤ i < j ≤ m.
Remark 18.4 Since the optimization problem of a quadratic function can be
reduced to solving a system of linear equations, the conjugate gradient method is
also a very effective iterative algorithm in solving systems of linear equations.
The systems of linear equations generated by the finite element method in
general correspond to sparse matrices, where the conjugate gradient method is
proven to be particularly effective.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 273
Remark 18.5 The conjugate gradient method is not only applicable to optimiza-
tion problems of quadratic functions, but is also applicable to convex functions.
Despite the fact that it may not reach the minimum in only n steps, the rate of
convergence is nevertheless notably improved.
Remark 18.6 Of particular note, in calculations of quantum mechanics and math-
ematical physics, it is often required to compute the spectrum of a positive definite
self-adjoint operator. Under certain compactness assumption, the spectrum con-
sists only of eigenvalues. In which case, based on the minimax description of the
eigenvalues (see Lecture 12), we can apply the Ritz method to obtain approxi-
mated solutions. Both the finite element method as well as the conjugate gradient
method are numerical implementations of the Ritz method.
May 2, 2013 14:6 BC: 8831 - Probability and Statistical Theory PST˙ws
Lecture 19
Example 19.1 (A rolling cart problem) A small rolling cart with mass m = 1 is
placed on the x-axis. Starting from the stationary point x = 0, the cart is to reach
the point x = L at exactly time T . Without friction and air resistance, we seek
how to apply a nonnegative external force u ≥ 0 on the cart such that the total
work done W is minimum.
Denote the position of the cart at time t by x(t). Let v(t) be the velocity of
the cart at time t and u(t) be the external force put on the cart at time t. Their
275
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 276
It follows that
Z t0
L
L= v(t) dt + (T − t0 ),
0 T
i.e.
Z t0
L
t0 = v(t) dt.
T 0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 277
Thus,
L
v(t) ≡ .
T
This contradicts (19.2).
3◦ We consider a discontinuous u. Let
(
a, t ∈ [0, T0 ],
u(t) =
0, t ∈ (T0 , T ],
where a and T0 are parameters. The solution is
(
at, t ∈ [0, T0 ],
v(t) =
aT0 , t ∈ (T0 , T ].
Now we have
T0
a2 T02
Z
W = a2 t dt = ,
0 2
T
aT02 aT 2
Z
L= v dt = + aT0 (T − T0 ) = aT T0 − 0 .
0 2 2
2
L L
When aT0 = T, as T0 → 0, W → 2T 2 , the controlled external force u(t) =
L
T δ(t), where δ(t) is a pulse function.
Therefore, the pulse force u(t) = TL δ(t) produces constant velocity v = TL ,
L2
which realizes the minimal work W = 2T 2 . However, pulse force is not physical.
◦
4 If u is allowed to change signs, then the control set becomes U =
L∞ ([0, T ]).
If we set u = a( T2 − t), then
t2
T
v=a t− , v(T ) = 0,
2 2
hence W = 0. However, there are many solutions for W = 0. For example, if we
take
a,
t ∈ [0, T0 ],
u(t) = 0, t ∈ (T0 , T − T0 ),
−a,
t ∈ [T − T , T ], 0
then
at,
t ∈ [0, T0 ],
v(t) = aT0 , t ∈ (T0 , T − T0 ),
a(T − t),
t ∈ [T − T0 , T ],
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 278
Example 19.1 (continued) In the above example, the state equations are
(
ẋ = v,
t ∈ [0, T ],
v̇ = u,
the initial state and the terminal set are respectively given by
x(0) 0 L
= , B= .
v(0) 0 ∗
The set of permissible control functions is
U = {u ∈ L∞ ([0, T ], R1 ) | u(t) ≥ 0}.
The Lagrangian L = u · v, which defines the price functional
Z T Z T
J(u) = W = u · v dt = L(v, u) dt.
0 0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 279
Example 19.2 We find how to reinvest the products in order to maximize the
overall production of the goods.
Suppose the production rate of a certain goods (the unit time production) is
q = q(t). Suppose its rate of increase q̇ is proportional to the percentage of
reinvestment u,
q̇ = αuq, t ∈ [0, T ],
where α > 0 is given. The initial state is also given,
q(0) = q0 .
However, the terminal state has no restriction whatsoever, i.e. B = R1 .
The overall production is given by
Z T
I= (1 − u)q dt.
0
The controlled reinvestment percentage
u ∈ U = {v ∈ L∞ (0, T ) | v(t) ∈ [0, 100]}.
The goal is to maximize the overall production, i.e.
max{I | u ∈ U }.
Example 19.3 (The Ramsey economic growth model) In 1928, Ramsey pro-
posed the following economic growth model.
Suppose the investment per worker is i, the consumption per worker is c, and
the output per worker is x = i + c.
Suppose x is a function f (k) of the capital intensity (capital per worker) k.
Suppose the population growth rate is α, then the labor L, as a function of time t,
follows the exponential growth L(t) = L0 eαt .
Since the total capital is K = kL, the total investment I = dK dt . They are
related via
I 1 dK dk k dL dk
i= = = + = + αk.
L L dt dt L dt dt
The target function measuring the consumption of the society is
Z ∞
W (c) = e−pt u(c(t))dt,
0
where p > 0 is called the depreciation rate of capital, while u(c) is called
the consumer’s utility function. Ramsey proposed to maximize W (c) under the
constraint
dk
= f (k) − αk − c, k(0) = k0 .
dt
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 280
λ̇ = −Hx (x∗ , λ∗ , u∗ ),
where
Furthermore, along the optimal orbit {(x∗ (t), λ(t), u∗ (t)) | t ∈ (0, T )}, H is the
constant
The function H(x, λ, u) is also called the Hamiltonian, whereas λ is called the
conjugate variable.
Proof (The proof is rather lengthy and it is beyond the scope of this book. We
hereby adopt the outline of the proof given by Macki and Struass [MS], we also
refer to MacCluer [Mac] for details.)
1◦ Suppose for any initial state x0 , ∃ u = u(x0 , t) ∈ U minimizes the func-
tional. This means there exists an optimal path, y = x(x0 , t) for t ∈ [0, T0 ] such
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 281
According to the local minimizing principle, (u(x0 , t), x(x0 , t)) is also optimal in
(t, T0 ], hence
Z T0
0
I(x(x , t)) = L(x(x0 , s), u(x0 , s)) ds.
t
◦
2 Suppose I is differentiable and u is piecewise continuous. Then except for
finitely many points, for all t, we have
0 = L(x(x0 , t), u(x0 , t)) + ∇I(x(x0 , t)) · ẋ(x0 , t)
= L(x(x0 , t), u(x0 , t)) + ∇I(x(x0 , t))F (x(x0 , t), u(x0 , t)).
When x0 = x0 , by setting u = u∗ (x0 , t), we have x = x∗ (x0 , t) and
0 = L(x∗ (x0 , t), u∗ (x0 , t)) − λF (x∗ (x0 , t), u∗ (x0 , t))
= −H(x∗ (x0 , t), λ(t), u∗ (x0 , t)),
where λ(t) = −∇I(x∗ (x0 , t)). Since x∗ (x0 , t) is piecewise differentiable, λ is
piecewise continuous and along the optimal orbit (x∗ , λ, u∗ ),
H(x∗ (x0 , t), λ(t), u∗ (x0 , t)) = 0.
3◦ We now prove
H(x∗ (t), λ(t), v) ≤ 0. ∀ v ∈ C, ∀ t ∈ (0, T ).
In fact, ∀ v ∈ C, the system
(
ẋ = F (x, v),
x(0) = x0 , ∀ t ∈ [0, T0 )
has a local solution x = x̃(t). Generally speaking, ∀ v ∈ C, x̃ is not the optimal
orbit from x0 to x̃(t), we only have
Z t
I(x0 ) ≤ L(x̃, v) ds + I(x̃(t)).
0
So for t > 0 sufficiently small,
t
I(x̃(t)) − I(x0 )
Z
1
− ≤ L(x̃, v) ds.
t t 0
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 282
Remark 19.2 In this theorem, if the terminal vector x(T ) = x1 has some com-
ponents which are not pre-determined, then the corresponding components in the
conjugate variable λ(T ) must be zero.
We now revisit the previous two examples using the maximal principle.
Example 19.1 (further continuation) Suppose the control variable has a con-
straint 0 ≤ u(t) ≤ a, where a > 0 is a prescribed constant. By introducing the
conjugate variable λ(t) = (α(t), β(t)), the Hamiltonian is
v
H = −L + λF = −uv + (α, β) = αv + u(β − v).
u
Suppose (x∗ (t), v ∗ (t)) is the optimal orbit with conjugate orbit (α, β) such that it
achieves the maximum of the Hamiltonian. Since v ∗ (t) is fixed, u(β − v ∗ ) must
attain its maximum. Thus,
(
a, β > v∗ ,
u=
0, β < v∗ .
Since v̇ = u and v(0) = 0, it follows that there exists a constant v0 such that
(
∗ at, β > v∗ ,
v =
v0 , β < v∗ .
In addition, the conjugate variable satisfies the equation
λ̇ = (α̇, β̇) = −H(x,v) = (0, u − α).
From λ(t) = (α(t), β(t)), we deduce α = α0 must be a constant, and
(
(a − α0 )t + c, u = a,
β=
−α0 t + d, u = 0,
where c and d are constants. Since both v ∗ and β are continuous, it follows that
β − v ∗ = −α0 t + c, d = c + v0 .
Noting the terminal v(T ) is indeterminate, according to Remark 19.2, β(T ) = 0.
However, since
−α0 t + c = 0
has only one real root T0 = c/α0 , this implies c ≤ α0 T .
at, t ∈ [0, T0 ),
v ∗ (t) =
aT ,
0 t ∈ [T0 , T ].
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 284
(
a, t ∈ [0, T0 ),
u(t) =
0, t ∈ [T0 , T ].
Moreover, by the terminal condition x(T ) = L, we also have
Z T
aT 2
L= vdt = aT T0 − 0 ,
0 2
r
2 2L
T − T 2 − L, a≥ ,
a T2
T0 =
no solution, 2L
. a<
T2
Example 19.2 (continued) We return to analyze how to maximize the overall
production.
Introducing the conjugate variable λ, the Hamiltonian is
H = −L + λF = (1 − u)q + λ(αuq) = (1 − u + λαu)q.
1◦ Since q ≥ 0, maximizing H is equivalent to maximizing −u + λαu =
(λα − 1)u. Hence, the optimal control variable should be
(
1, αλ(t) > 1,
u(t) =
0, αλ(t) < 1.
2◦ The conjugate variable λ satisfies the equation
λ̇ = −Hq = −(1 − u + λαu),
namely, λ̇ + λαu = u − 1, therefore,
(
λ̇ + αλ = 0, u = 1,
λ̇ = −1, u = 0.
It yields the solution
(
ce−αt , u=1
λ(t) =
−t + d, u = 0,
where c and d are constants.
3◦ The terminal condition. Since q(T ) is indeterminate, λ(T ) = 0. This
implies u(T ) 6= 1, which is only possible if
u = 0, d = T.
This conclusion is reasonable, since at the end of a production cycle, there is no
need to reinvest.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 285
4◦ The change of the control variable. From 1◦ , we already know, the values
of the control variable u should change at αλ(t) = 1, i.e. when ts = T − α1 .
Thus, the final solution is given by
(
1, t ∈ (0, ts ),
u(t) =
0, t ∈ (ts , T ),
(
q0 eαt , 0 < t < ts
q(t) =
q0 eαts , ts < t < T,
Z T
q0 α(T − 1 )
I= q0 eαts dt = q0 eαts (T − ts ) = e α .
ts α
Example 19.3 (continued) Returning to the Ramsey’s economic growth model,
to maximize W (c) is equivalent to minimizing −W (c). By introducing the con-
jugate variable λ and the Hamiltonian
H(k, λ, c) = e−pt u(c) + λ(f (k) − αk − c),
we obtain the optimal conditions
∂H
= u0 (c)e−pt − λ = 0,
∂c
and
k 0 (t) = f (k) − αk − c,
∂H
λ0 (t) = −
= −λ(f 0 (k) − α)
∂k
together with the initial condition k(0) = k0 . Furthermore, since the problem is
proposed on the positive half line, we also have the terminal condition
lim λk(t) = 0.
t→∞
Theorem 19.2 In the linear system (19.5), if there exists a control variable u0 ∈
L∞ (0, T ) to move from the initial state x(0) = x0 to the terminal state x(T ) =
x1 , then it must be realized by a bang-bang control u.
This theorem is in fact a corollary of the Krein–Milman theorem in functional
analysis. The Krein–Milman theorem concerns the extreme points of a compact
convex subset in a locally convex topological vector space. It is an equivalent
statement to the Hahn–Banach theorem.
Let X be a locally convex topological vector space and E ⊂ X be a convex
subset. A point x ∈ E is called an extreme point, if there are no y, z ∈ E such
that x = (y + z)/2. The set of all extreme points is called the extremal set of E,
denoted by D(E).
The Krein–Milman theorem Let X be a locally convex topological vector space
and E ⊂ X be a compact convex subset, then E is the closure of the extremal set
D(E).
Proof of Theorem 19.2 Without loss of generality, we may assume ku0 (t)k∞ ≤ 1.
Let
Z t
U = v ∈ L∞ (0, T ) | kvk∞ ≤ 1 x(T ) = eA x0 + eA(t−s) Bv(s)ds .
0
Lecture 20
289
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 290
for every partition π = {a < t1 < · · · < tn+1 < b} of (a, b). We call
then µu ({α}) = u(α + 0) − u(α). Hence, for any Borel set E ⊂ J, we have
!
[
1
µu (E) = L [u(x), u(x + 0)] ,
x∈E
Let Ω ⊂ Rn be a bounded open subset. Denote C01 (Ω, Rn ) the function space
of all continuously differentiable functions on Ω with compact support in Rn .
Analogous to the single variable situation, we define
Defintion 20.3 A function u ∈ L1 (Ω) is said to be of bounded variation, if
Z
udivϕ dx ϕ ∈ C01 (Ω, Rn ), kϕk∞ ≤ 1 < ∞.
kDuk(Ω) = sup
Ω
it satisfies
|L(ϕ)| ≤ Ckϕk∞ , ∀ ϕ ∈ C01 (Ω, Rn ).
By letting
L(ϕ0 ) = lim L(ϕk ),
k→∞
we see that not only the limit L(ϕ0 ) exists, but it is independent of the particular
choices of ϕk as well. This means L has a unique linear continuous extension
L : C0 (Ω, Rn ) → R1 .
Namely, L is a linear continuous functional on C0 (Ω, Rn ). By the Riesz repre-
sentation theorem, there exists a vector-valued Radon measure µ = (µ1 , . . . , µn )
such that
Z Z X n Z
u div ϕ dx = − ϕi dµi = − ϕ dµ, ∀ ϕ = (ϕ1 , . . . , ϕn ) ∈ C01 (Ω, Rn ).
Ω Ω i=1 Ω
This means
Du = dµ.
We henceforth denote the space of all Radon measures on Ω taking values in Rn
by M(Ω, Rn+1 ), on which, by defining the norm
Z
kµkM = sup ϕ dµ kϕkC0 (Ω,Rn+1 ) ≤ 1 ,
Ω
it becomes a Banach space.
Example 20.1 Let u ∈ W 1,1 (Ω), then u ∈ BV(Ω) and kukW 1,1 = kukBV . To
prove this, note that on one hand,
Z Z
Du · φ ≤ kDuk1 , ∀ φ ∈ C01 (Ω, Rn ), kφk∞ ≤ 1.
u div φ =
Ω Ω
On the other hand, ∀ ε > 0, ∃ ϕε ∈ C01 (Ω, Rn ) with kϕε k∞ ≤ 1, such that
Z Z Z
|Du| ≤ Du · ϕε dx + ε = u div ϕε dx + ε ≤ kDuk(Ω) + ε.
Ω Ω Ω
Consequently, kDuk(Ω) = kDuk1 .
Example 20.2 Let S ⊂ Rn be a C ∞ compact closed hypersurface of dimension
n − 1. We equip S the induce metric from Rn and denote its (n − 1) dimensional
Hausdorff measure by Hn−1 (S).
Assume Ω is the region bounded by S with characteristic function χΩ . Then
χΩ ∈ BV(Rn ) and
kDχΩ k(Rn ) = Hn−1 (S).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 295
then
hµi , ϕi = −hµ0 , ∂i ϕi, i = 1, 2, . . . , n, ∀ ϕ ∈ C01 (Ω).
Using techniques from measure theory, it is not difficult to verify that µ0 is indeed
absolutely continuous. Therefore, there exists u ∈ L1 (Ω) such that µ0 = uLn .
Since
Z
kT ukM(Ω,Rn+1 ) = sup [u · ϕ0 + Du · ϕ̂] dx kφkC0 (Ω,Rn+1 ) ≤ 1 ,
Ω
it follows that
kukBV = kukL1 + kDuk(Ω) ≤ 2kT uk ≤ 2kukBV .
As a consequence, we can regard BV(Ω) as the dual space of the separable Banach
space (X/Y ). By the Banach–Alaoglu theorem, every bounded set is weak-∗
sequentially compact.
We now analyze the weak-∗ topology on BV(Ω).
Theorem 20.1
uj → u, in L1 ,
BV weak∗
uj −−−−−−→ u ⇐⇒
uj is bounded in BV.
Fig. 20.1
They are defined as follows. Denote Br (x) the ball of radius r > 0 with center
x in Rn . Define two functions (see Figure 20.1)
Ln {y ∈ Br (x) | u(y) > t}
+
u (x) = inf t ∈ [−∞, +∞] lim =0 ,
r→0 rn
Ln {y ∈ Br (x) | u(y) < t}
−
u (x) = sup t ∈ [−∞, +∞] lim =0 .
r→0 rn
A point x ∈ Ω is called a Lebesgue point if
Z
1
lim |u(y) − u(x)| dy = 0.
r→0 |Br (x)| B (x)
r
and
u(x) = u+ (x) = u− (x).
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 299
Under certain growth restriction, we know the weak sequential lower semi-
continuity of a functional I is equivalent to the quasi-convexity of the Lagrangian
in p. In Lecture 13, we mentioned that Bolza gave the counterexample (Example
13.3) of L(u, p) = u2 + (p2 − 1)2 . It is not convex in p, I(u) has a minimizing
sequence {un } in W 1,4 (0, 1) which converges weakly to 0, and the corresponding
functional values I(un ) → 0, but I(0) = 1.
However, from the viewpoint of extremal values of a functional, in this exam-
ple, the value of I(0) is not of particular importance, but instead, the minimizing
sequence itself reflects how small the values of I may become.
In order to showcase the characteristic of the functional in this regard, we
introduce the concept of a relaxation function.
Definition 20.4 Let (X, τ ) be a topological space and f : X → R̄ = R1 ∪{+∞}.
We call Rf (x) = sup{ϕ(x) | ϕ : X → R̄ is lower semi-continuous, ϕ(y) ≤
f (y), ∀ y ∈ X} the relaxation function of f .
The following are immediate from the definition.
(1) If f is lower semi-continuous, then Rf = f .
Proof Choosing ϕ = f , then f ≤ Rf . Conversely, by definition, Rf ≤ f .
(2) Rf (x) is lower semi-continuous. This is because the supremum of a family
of lower semi-continuous functions is itself lower semi-continuous.
From this, we see that the relaxation function Rf of f is the largest among all
lower semi-continuous functions which are not greater than f .
(3) Let f : X → R̄ be bounded below. Suppose f has a minimizing se-
quence {xj } and x0 is a limit point of {xj }, then x0 is the minimum of Rf and
Rf (x0 ) = inf f .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 300
Proof Without loss of generality, we may assume c = inf f = lim f (xj ). On one
hand,
min Rf = inf f.
X X
Proof Choose a minimizing sequence {xj } of f such that lim f (xj ) = inf X f .
*w
Since {xj } is bounded in X, it contains a weak-∗ convergent subsequence xnj −→
x0 . By definition, the relaxation function Rf with respect to the weak-∗ topology
is weak -∗ lower semi-continuous, by property (3), x0 is the minimum of Rf , i.e.
Noting that
thus, we have
Z
RI(u) = conv(L)(∇u(x)) dx,
Ω
Fig. 20.2
Example 20.3 (The double well) Let L(t, u, p) = (1 − p2 )2 (see Figure 20.2),
Z 1
I(u) = L(t, u(t), u̇(t))dt.
−1
Define
(1 − p2 )2 , |p| > 1,
L̃(t, u, p) =
0, |p| ≤ 1,
which is the convex hull of L, then
Z 1
RI(u) = L̃(t, u(t), u̇(t))dt.
−1
where K is the convolution operator derived from the point source dispersion
function of the optical instrument itself. It can be viewed as a bounded linear oper-
ator on L2 (Ω). Moreover, n(x) denotes the noise. The so-called image restoration
is the process of deducing u from ud .
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 302
The most natural approach is to employ the least square method in finding
Z
inf |ud − Ku|2 dx, (20.3)
Ω
whose E-L equation is
K ∗ ud − K ∗ Ku = 0, (20.4)
where K ∗ is the adjoint operator of K. Since K ∗ K does not have bounded in-
verse, (20.2) is in general ill-posed.
Tikhonov and Arsenin proposed a regularization method in 1977, by adding
an energy term
Z Z
Tλ (u) = |ud − Ku|2 dx + λ |∇u|2 dx, where λ is a parameter
Ω Ω
to the target functional in (20.3).
The advantage of this method can not only overcome the ill-posedness of
(20.2), but also smooth out the solution u.
If we seek the minimum of the functional on H 1 (Ω), then the E-L equation is
K ∗ ud − K ∗ Ku + λ∆u = 0,
with the Neumann boundary condition
∂u
= 0.
∂n ∂Ω
The regularity of solution of an elliptic equation insures the smoothness of the
solution u. By doing so, some noise can also be filtered out. However, this ho-
mogenized smoothing process has its flaw, it inevitably blurs the boundary of the
image.
Several attempts were made to remedy this issue. For example, replacing the
L2 norm of the gradient ∇u by the Lp norm, and by decreasing the value of p, one
can hopefully retain the image on the
R boundary. Rudin–Osher–Fatemi suggested
to use the L1 norm, namely, using Ω |∇u| dx instead of Ω |∇u|2 dx.
R
Unfortunately, W 1,1 (Ω) is neither reflexive, nor is it the dual space of any
Banach space. Consequently, it is very difficult to determine the existence of
minimum of a functional on it.
We have stated that BV(Ω) is the dual space of a separable Banach space.
However, a BV solution u may have jump-type discontinuity, while the set of
jump-type discontinuity is a rectifiable curve. If we solve the problem in BV(Ω),
then not only the existence of the solution is relatively easy to establish, but the
edge of the body is also easier to meet. The model they proposed is to find the
minimum u of the functionalZ Z
1 2
E(u) = |ud − Ku| dx + λ φ(|∇u|) dx, (20.5)
2 Ω Ω
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 303
Bibliography
The main references for Lectures 1–8: [BGH], [LL], [GF], [JJ], [Ka], [Mac],
[Mos].
The main references for Lectures 9–14: [Ad], [AE], [CL], [Da], [Ek], [Gi],
[JJ], [Mos], [Mor], [Ne], [Po].
[Ad] Adams, R. A. (1975) Sobolev Spaces, Acad. Press. (Chinese translation by Ye,
Q. X., Wang, Y. D., Ying, L. A., Han, H. D., and Wu, L. C., Beijing, People’s
Education Press, 1983.
[AFP] Ambrosio, L., Fusco, N., and Pallara, D. (2000). Functions of Bounded Varia-
tion and Free Discontinuous Problems, Clarendon Press, Oxford.
[AE] Aubert G. and Ekeland I. (1984). Applied Nonlinear Analysis, A John Wiley-
Interscience Publication.
[AK] Aubert, G. and Kornprobst, P. (2002). Mathematical Problems in Image Pro-
cessing, Partial Differential Equations and the Calculus of Variations, Applied
Math Sciences 147, Springer.
[BM] Brechtken, U. and Manderschild, U. (1991). Introduction to the Calculus of
Variations, Chapman Hill.
[Bu] Buttazzo, G. (1989). Semicontinuity, relaxation and integral representation in
the Calculus of Variations, Pitman Research Notes in Mathematics 207, Long-
man Scientific and Technical.
[BGH] Buttazzo, G., Giaquinta, M., and Hildebrandt, S. (1998). One-dimensional
Variational Problems, An Introduction, Clarendon Press. Oxford.
[CL] Chang, K. C. and Lin, Y. Q. (1987). Lecture notes on Functional Analysis I
(Chinese), Beijing, Peking University Press.
305
August 23, 2016 13:50 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 306
[Ch] Chang, K. C. (1993). Infinite Dimensional Morse Theory and Multiple Solution
Problems, Birkhäuser.
[Co] Courant, R. (1950). Dirichlet’s Principle, Conformal Mapping, and Minimal
Surfaces, New York, Interscience.
[Da] Dacorogna, B. (1989). Direct Methods in the Calculus of Variations, Applied
Math. Sciences, 78, Springer Verlag.
[Ek] Ekeland, I. (1979). Non-convex Minimization Problems, Bull. Amer. Math.
Soc., pp. 443–474.
[Fe] Federer, H. (1969). Geometric Measure Theory, Springer-Verlag.
[GF] Gelfand, I. M. and Fomin, S. V. Calculus of Variations, (English translation by
Silverman, R. A., Prentice Hall, 1964).
[Gi] Giaquinta, M. (1983). The regularity problem of extremals of variational inte-
grals, Proc. NATO/LMS Advance Study Inst. on “Systems of nonlinear partial
differential equations”, Reidel Publ. Co.
[GH] Giaquinta, M. and Hildebrandt, S. (1996). Calculus of Variations I The La-
grangian formalism, Grundlehren der mathematischen Wissenschaften, 310,
Springer.
[GT] Gilbarg, D. and Trudinger, N. (1983). Elliptic Partial Differential Equations of
Second Order, 2nd ed. Grundlehren der mathematischen Wissenschaften, 224
Springer..
[Ha] Hardt, R. (Ed. 2004). Six Theorems in Variations, Student Math. Lib. 26 AMS.
[Hi1] Hilbert, D. (1900). Uber das Dirichletsche Prinzip, Jber Deutsch Math. Vere.,
8, pp.184–188.
[Hi] Hildebrandt, S. (1969). Boundary behavior of Minimal Surfaces, Arch. Rat.
Mech. Anal., 35, pp. 47–82.
[HT] Hofer, H. and Toland, J. F. (1984). Homoclinic, heteroclinic and periodic so-
lutions for indefinite Hamiltonian systems, Math. Ann., 268, pp. 387–403.
[Hu] Hughes, T. (2000). Finite Element method, Linear Static and Dynamic Finite
Element Analysis, Dover Publications.
[JJ] Jost, J. and Jost, X. Li. (1988). Calculus of Variations, Cambridge University
Press.
[Jo] Jost, J. (1990). Two Dimensional Geometric Variational Problems, John Wiley
and Sons.
[Joh] Johnson, C. (1987). Numerical Solution of Partial Differential Equations by
the Finite Element Method, Studentlitteratur, Lund.
[Ka] Kato, T. Calculus of Variations and its Applications (Chinese Translation by
Zhou, H. S., Shanghai Science and Technology Press, 1961).
[LL] Lavrentiev, M. A. and Liusternik, L. A. Lecture Notes on the Calculus of Vari-
ations (Chinese Translation by Zeng, D. H., Deng, H. Y. and Wang, Z. K., High
Education Press, Beijing, 1955).
[Le] Leitmann, G. (1981). The Calculus of Variations and Optimal Control, An In-
troduction, Plenum Press, 1981.
[Mac] MacCluer, C. R. (2005). Calculus of Variations, Mechanics, Control, and
Other Applications, Pearson Education, Ltd.
[MS] Macki, J. and Strauss, A. (1982). Introduction to Optimal Control Theory,
Springer-Verlag, 1982.
August 19, 2016 10:39 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 307
Bibliography 307
Index
309
August 30, 2016 14:14 ws-book9x6 BC: 10157 – Lecture Notes on Calculus of Variations book page 310
Index 311
There is a section that makes use of the Johnson solids (called the Johnson-
Zalgeller solids in the book). These are the 92 convex polyhedra that are more or less
irregular, but whose faces are regular polygons. Edge models of all 92, made from
coloured plastic rods, are shown in a single picture about 12 cm by 10 cm. It is
impossible to make out much detail: they are unlabelled, some overlap others, and
the smallest is a triangular blob about 4 mm by 2 mm. In the text some are referred
to only by number, e.g. J3, J7, J65, with no illustration or other indication of their
shape. Although the numbering is standard and the book claims to be self-contained,
we have to look elsewhere to find what solids these represent.
The English slips occasionally (‘... every repeated pattern has one of two
dimensional symmetry groups ... ’), and there are a few misprints. I was quietly
amused to see a British newspaper that was once renowned for its misprints referred
to as ‘The Gardian’. The mathematician Hallard Croft might not be so amused to see
his name given consistently in the references as H. T. Craft.
There is a lot of jargon, probably unavoidable, but wading through uni-trunk
holders, canonical parallelohedra, (P,Q)-chimera superimpositions,
equicomplementability and equidecomposability is perhaps not for the faint-hearted.
There are chapters that stand alone, and can easily be dipped into. But sections such
as the work leading to Hilbert's third problem and Dehn's theorem are serious
mathematics that has to be worked at. You have only to look at the given definition
of the Dehn invariant to see that it makes nonsense of the claim on the back cover
that ‘It does not require more than a high school level of knowledge ...’.
Despite my strictures this is an interesting work, with much that cannot be found
elsewhere. I shall certainly try to fathom how the 858-box works. This is a book,
perhaps, for the library, with a notice ‘Handle with care’.
10.1017/mag.2017.164 MICHAEL FOX
2 Leam Road, Leamington Spa CV31 3PA
e-mail: vulpius1@gmail.com
classical theory includes sufficient conditions for extremal solutions, the Legendre-
Hadamard condition, the isoperimetric problem, Mayer field, Hilbert's invariant
integral, and the Hamilton-Jacobi theory. There is also a lecture on Noether's
theorem, which gives a unified approach to various conservation laws in physics and
mechanics from the invariant properties of the Lagrangian under certain group
actions. Dirichlet's principle, the analysis of the existence of extreme values of
functionals, and the need to introduce notions such as weak sequential lower semi-
continuity are considered in the lectures on foundation and development, which also
include lectures on Sobolov spaces and the Ekeland variational principle. The
lectures on applications include the Mountain Pass theorem, geodesics and minimal
surfaces, numerical methods (Rayleigh-Ritz, finite element, optimisation, conjugate
gradient), optimal control, and image processing.
I learned much from studying the lectures on applications. Take, as an example,
Lecture 16, with the title ‘Periodic solutions, homoclinic and heteroclinic orbits’,
which opens with the consideration of the simple pendulum. The dynamical equation
concerned is ϕ ¨ + (g / l) sin ϕ = 0, where g and l have their usual meanings; readers
will be familiar with the simple harmonic solution if sin ϕ is replaced by its first
order approximation ϕ. Here we are given a proper treatment for the problem, and
the topics in the title are considered in some depth.
The prerequisite is perhaps too high for most undergraduates—readers are
expected to be thoroughly familiar with concepts from measure theory, functional
analysis, differential geometry and topology, as well as various techniques in applied
mathematics. Sometimes esoteric theorems and methods are used with neither
comment nor reference, for example, the embedding theorems of Rellich-
Kondrachov, and of Nash. The presentation can be somewhat economical in places,
where supplementary work is required for a proper understanding. However, there
are suitably chosen examples throughout the text to illustrate new notions and
concepts, and also exercises at the end of most of the lectures. The book should be
useful to advanced undergraduates and graduate students, and it will also serve as a
reference text.
10.1017/mag.2017.165 PETER SHIU
353 Fulwood Road, Sheffield S10 3BQ
e-mail: p.shiu@yahoo.co.uk