Lectures on

Semi-group Theory and its

Application to Cauchys Problem
in Partial Dierential Equations
By
K. Yosida
Tata Institute of Fundamental Research, Bombay
1957
Lectures on
Semi-group Theory and its
Application to Cauchys Problem
in Partial Dierential Equations
By
K. Yosida
Notes by
M.S. Narasimhan
Tata Institute of Fundamental Research,
Bombay
1957
Contents
1 Lecture 1 1
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 1
I Survey of some basic concepts and ... 3
2 Normed linear spaces: . . . . . . . . . . . . . . . . . . . 5
3 Pre-Hilbert spaces . . . . . . . . . . . . . . . . . . . . . 6
4 Example of a pre-Hilbert space . . . . . . . . . . . . . . 7
5 Banach spaces . . . . . . . . . . . . . . . . . . . . . . . 8
6 Hilbert space . . . . . . . . . . . . . . . . . . . . . . . 8
7 Example of Banach spaces . . . . . . . . . . . . . . . . 8
8 Example of a Hilbert space . . . . . . . . . . . . . . . . 9
9 Completion of a normed linear space . . . . . . . . . . . 10
10 Additive operators . . . . . . . . . . . . . . . . . . . . . 10
2 Lecture 2 13
1 Linear operators . . . . . . . . . . . . . . . . . . . . . . 13
2 Hahn-Banach lemma . . . . . . . . . . . . . . . . . . . 13
3 Lemma (Hahn-Banach) . . . . . . . . . . . . . . . . . . 14
4 Hahn-Banach extension theorem... . . . . . . . . . . . . 15
5 Hahn-Banach extension theorem for... . . . . . . . . . . 15
6 Existence of non-trivial linear functionals . . . . . . . . 17
7 Orthogonal projection and... . . . . . . . . . . . . . . . 18
iii
iv Contents
3 Lecture 3 21
1 The Conjugate space (dual) of a normed linear space . . 21
2 The Resonance Theorem . . . . . . . . . . . . . . . . . 21
3 Weak convergence . . . . . . . . . . . . . . . . . . . . 23
4 A counter-example . . . . . . . . . . . . . . . . . . . . 24
4 Lecture 4 27
1 Local weak compactness of a Hilbert space . . . . . . . 27
2 Lax-Milgram theorem . . . . . . . . . . . . . . . . . . . 28
II Semi-group Theory 31
5 Lecture 5 35
1 Some examples of semi-groups . . . . . . . . . . . . . . 35
2 The innitesimal generator of a semi-group . . . . . . . 39
6 Lecture 6 45
1 The resolvent set and... . . . . . . . . . . . . . . . . . . 46
2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 47
7 Lecture 7 53
1 The exponential of a linear operator . . . . . . . . . . . 53
2 Representation of semi-groups . . . . . . . . . . . . . . 56
8 Lecture 8 59
1 An application of the representation theorem . . . . . . . 59
2 Characterization of the... . . . . . . . . . . . . . . . . . 59
9 Lecture 9 65
1 Group of operators . . . . . . . . . . . . . . . . . . . . 65
10 Lecture 10 69
1 Supplementary results . . . . . . . . . . . . . . . . . . . 69
Contents v
11 Lecture 11 75
1 Temporally homogeneous Marko... . . . . . . . . . . . 75
2 Brownian motion on a homogeneous Riemannian space . 76
12 Lecture 12 81
1 Brownian motion on a homogeneous... . . . . . . . . . . 81
III Regularity properties of solutions of linear elliptic... 85
13 Lecture 13 87
1 Strong dierentiability . . . . . . . . . . . . . . . . . . 87
2 Weak solutions of linear dierential operators . . . . . . 88
3 Elliptic operators . . . . . . . . . . . . . . . . . . . . . 89
4 Fourier Transforms: . . . . . . . . . . . . . . . . . . . . 89
14 Lecture 14 93
1 Gardings inequality . . . . . . . . . . . . . . . . . . . . 93
15 Lecture 15 99
1 Proof of the Friedrichs - Lax - Nirenberg theorem . . . . 99
16 Lecture 103
1 Proof of Lemma 3 . . . . . . . . . . . . . . . . . . . . . 103
17 Lecture 17 107
1 Proof of Lemma 2 . . . . . . . . . . . . . . . . . . . . . 107
2 Proof of Lemma 1 . . . . . . . . . . . . . . . . . . . . . 108
IV Application of the semi-group theory.... 113
18 Lecture 18 115
1 Cauchy problem for the diusion equation . . . . . . . . 115
2 Gardings inequality . . . . . . . . . . . . . . . . . . . . 118
vi Contents
19 Lecture 19 121
1 The Cauchy problem for the wave equation . . . . . . . 121
20 Lecture 20 127
1 Cauchy problem for the wave equation (continued) . . . 127
21 Lecture 21 133
1 Integration of the Fokker-Planck equation . . . . . . . . 133
22 Lecture 22 139
1 Integration of the Fokker-Planck... . . . . . . . . . . . . 139
23 Lecture 23 145
1 Integration of the Fokker-Planck equation (Contd.) .... . . 145
Lecture 1
1 Introduction
The analytical theory of one-parameter semi-groups deals with the ex- 1
ponential function in innite dimensional function spaces. It is a natural
generalization of the theorem of Stone on one-parameter groups of uni-
tary operators in a Hilbert space.
In these lectures, we shall be concerned with the dierentiability
and the representation of one-parameter semi-groups of bounded lin-
ear operators on a Banach space and with some of their applications to
the initial value problem (Cauchys problem) for dierential equations,
especially for the diusion equation (heat equation) and the wave equa-
tion.
The ordinary exponential function solves the initial value problem:
dy
dt
= y, y(0) = C.
We consider the diusion equation
u
t
= u,
where =
m
_
i=1

2
x
2
i
is the Laplacian in the Euclidean m-space E
m
; we
wish to nd a solution u = u(x, t), t 0, of this equation satisfying the
initial condition u(x, 0) = f (x), where f (x) = f (x
1
, . . . , x
n
) is a given
1
2 1. Lecture 1
function of x. We shall also study the wave equation

2
u
t
2
= u, t
with the initial data
u(x, 0) = f (x) and
_ u
t
_
t=o
= g(x),
f and g being given functions. This may be written in the vector form 2
as follows:

t
_
u
v
_
=
_
0 I
0
__
u
v
_
, v =
u
t
with the initial condition
_
u(0)
v(0)
_
=
_
f (x)
g(x)
_
.
So in a suitable function space the wave equation is of the same form
as the heat equation - dierentiation with respect to the time parameter
on the left and another operator on the right - or again similar to the
equation
dy
dt
= y. Since the solution in the last case is the exponential
function, it is suggested that the heat equation and the wave equation
may be solved by properly dening the exponential functions of the op-
erators and
_
0 I
0
_
in suitable function spaces. This is the motivation
for the application of the semi-group theory to Cauchys problem.
Our method will give an explanation why in the case of the heat
equation the time parameter is restricted to non-negative values, while
in the case of the wave equation it may extend between and .
Before going into the details, we give a survey of some of the ba-
sic concepts and results from the theory of Banach spaces and Hilbert
spaces.
Part I
Survey of some basic
concepts and results from the
theory of Banach spaces
3
2. Normed linear spaces: 5
Denition. A set X is called a linear space over a eld K if the following 3
conditions are satised:
1) X is an abelian group (written additively).
2) There is dened a scalar multiplication: to every element x of X and
each K there is associated an element of X, denoted by x, such
that
( + )x = x + x, , K, x X
(x + y) = x + y, K, x, y X
()x = (x)
1x = x, 1 Kis the unit element of K.
We shall denote by Greek letters the elements of K and by Roman
letters the elements of X. The zero of X and the zero of K will both be
denoted by 0. We have 0.x = 0.
In the sequel we consider linear spaces only over the real number
eld or the complex number eld. A linear space will be said to be real
or complex according as the eld is the real number eld or the complex
number eld. In what follows, by a linear space we always mean a real
or a complex linear space.
Denition. A subset M of a linear space X is called a linear subspace
(or a subspace) if whenever x, y M and , K
2
then x + y M.
2 Normed linear spaces:
4
Denition. A linear space X (real or complex) is called a normed linear
space if, for every x X there is associated a real number, denoted by
||x||, such that
i) ||x|| 0 and ||x|| = 0 if and only if x = 0.
ii) ||x|| = ||||x||, ( is a scalar and || is the modulus of ).
iii) ||x + y|| ||x|| + ||y||, x, y X (triangle inequality). ||x|| is called the
norm of x.
6
A normed linear space becomes a metric space if the distance d(x, y)
between two elements x and y is dened by d(x, y) = ||x y||. We say
that a sequence of elements {x
n
} of X converges strongly to x X, and
write s lim
n
x
n
= x (or simply lim
n
x
n
= x), if lim
n
||x
n
x|| = 0. (This
limit, if it exists, is unique by the triangle inequality).
Proposition. If lim
n

n
= (
n
, K), s lim
n
x
n
= x and s lim
n
y
n
=
y, then s lim
n

n
x
n
= x and s lim
n
(x
n
+ y
n
) = x + y.
Proof.
||(x
n
+ y
n
) (x + y)|| = ||(x
n
x) + (y
n
y)||
||(x
n
x)|| + ||(y
n
y)|| (Triangle inequality)
0.
||
n
x
n
x|| ||x
n
x|| + ||
n
x
n
x
n
||
= |
n
|||x|| + |
n
|||x x
n
||
0.
Proposition . If s lim
n
x
n
= x then lim
n
||x
n
|| = ||x||, i.e., norm is a 5
continuous function.
Proof. We have, from the triangle inequality,

||x|| ||y|| ||x y||;

now take y = x
n
and let n .
3 Pre-Hilbert spaces
A special class of normed linear spaces - pre-Hilbert spaces-will be of
fundamental importance in our later discussion of dierential equations.
These normed linear spaces in which the norm is dened by scalar prod-
uct.
Denition. A linear space X is called a pre-Hilbert space if for every
ordered pair of elements (x, y)(x, y X) there is associated a number
(real number if X is a real linear space and complex number if X is a
complex linear space) such that
4. Example of a pre-Hilbert space 7
i) (x, x) 0 and (x, x) = 0 if and only if x = 0.
ii) (x, y) = (x, y), for every number .
iii) (x, y) = (y, x)[(y, x) denotes the complex conjugate of (y, x).]
iv) (x + y, z) = (x, z) + (y, z) x, y, z X.
(x, y) is called the scalar product between x and y.
If we dene ||x|| =

(x, x), a pre-Hilbert space becomes a normed

linear space, as is veried easily using Schwarzs inequality proved be-
low
Proposition. i) |(x, y)| ||x||||y|| (Schwarzs inequality)
ii) ||x + y||
2
+ ||x y||
2
= 2(||x||
2
+ ||y||
2
) (Euclidean property)
Proof. (ii) is easily veried. To prove (i), we observe that, for every real 6
number ,
0 (x + (x, y)y, x + (x, y)y)
= (x, x) + 2|(x, y)|
2
+
2
|(x, y)|
2
(y, y).
This quadratic form in , being always non-negative should have
non-positive discriminant so that
|(x, y)|
4
||x||
2
||y||
2
|(x, y)|
2
0.
If (x, y) = 0, (i) is obviously satised; if (x, y) 0, Schwarzs in-
equality follows from the above inequality.
4 Example of a pre-Hilbert space
Let R be a domain in Euclidean m-space E
m
. Let D
k
(R) denote the set of
all complex valued functions f (x) = f (x
1
, . . . , x
n
) which are of class C
k
in R(i.e., k times continuously dierentiable) and which have compact
support. These functions form a linear space with the ordinary function
8
sum and scalar multiplication. Dene the scalar product between two
functions f and g by
( f , g)
k
=

|n|k
_
R
D
(n)
f (x)D
(n)
g(x)dx , 0 k < ,
where n = (n
1
, . . . , n
m
) is a system of non-negative integers, |n| = n
1
+
+n
m
and
D
(n)
=

|n|
x
n
1
1
x
n
2
2
x
n
m
m
5 Banach spaces
Denition . A normed linear space is called a Banach spaces if it is
complete in the sense of the metric given by the norm.
(Completeness means that every Cauchy sequence is convergent: if 7
{x
n
} X is any Cauchy sequence, i.e., a sequence {x
n
} for which ||x
m

x
n
|| 0 as m, n independently, then there exists an element x X
such that lim
n
||x
n
x|| = 0.x is unique).
6 Hilbert space
Denition . A pre-Hilbert space which is complete (considered as a
normed linear space) is called a Hilbert space.
The pre-Hilbert space D
K
(R) dened in the last example is not com-
plete
7 Example of Banach spaces
1) C[, ]: Let [, ] be a closed interval < . Let C[, ]
denote the set of all bounded continuous complex-valued functions
x(t) on [, ]. (If the interval is not bounded, we assume further that
x(t) is uniformly continuous). Dene x + y and x by
(x + y)(t) = x(t) + y(t)
8. Example of a Hilbert space 9
(x)(t) = .x(t).
C[, ] is a Banach space with the norm given by
||x|| = sup
t[,]
|x(t)|
Converges in this metric is nothing but uniform convergence on the
whole space.
2) L
p
(, ). (1 p < ). This is the space of all real or complex val-
ued Lebesgue functions f on the open interval (, ) for which | f (t)|
p
is Lebesgue summable over (, ); two functions f and g which are 8
equal almost everywhere are considered to dene the same vector of
L
p
(, ). L
p
(, ) is a Banach space with the norm:
|| f || =
_

| f (t)|
p
dt
_

_
1/p
The fact that || || thus dened is a norm follows from Minkowskis
in-equality; the Riesz-Fischer theorem asserts the completeness of
L
p
.
3) L

(, ): This is the space of all measurable (complex valued) func-

tions f on (, ) which are essentially bounded, i.e., for every f
L

(, ) there exists a > 0 such that | f (t)| almost everywhere.

Dene || f || to be the inmum of such .
(Here also we identify two functions which are equal almost every-
where).
8 Example of a Hilbert space
L
2
(, ) : L
2
(, ) (see example (2) above), is a Hilbert space with the
scalar product
( f , g) =

f (t)g(t)dt.
10
9 Completion of a normed linear space
Just as the completeness of the real number eld plays a fundamen-
tal role in analysis, the completeness of a Banach space will play an
essential role in some of our subsequent discussions. If we have an in-
complete normed linear space we can always complete it; we can imbed
this space in a Banach spaces as an everywhere dense subspace and this
Banach spaces is essentially unique. We have, in fact, the
Theorem. Let X
0
be a normed linear space. Then there exists a com-
plete normed linear space (Banach spaces ) X and a norm preserving
isomorphism T of X
0
onto a subspace X

0
of X which is dense in X in 9
the sense of the norm topology. (That T is a norm preserving isomor-
phism means that T is one-to-one, T(x
0
+y
0
) = T(x
0
) +T(y
0
) and
||x|| = ||T(x)||). Such an X is determined uniquely upto a norm preserv-
ing isomorphism
Sketch of the proof: The proof follows the same idea as that utilized
for dening the real numbers from the rational numbers. Let X be the
totality of all Cauchy sequences {x
n
} X
0
classied according to the
equivalence: {x
n
} {y
n
} if and only if lim
n
||x
n
y
n
|| = 0. Denote by {x
n
}
the class containing {x
n
}.
If x, y X and x = {x
n
}, y = {y
n
}, dene x + y = {x
n
+ y
n
}, x =
{x
n
}, || x|| = lim
n
||x
n
||. These denitions do not depend on the particular
representatives for x, y respectively. Finally if x
0
X
0
denes T(x
0
) =
{x
n
} where each x
n
= x
0
.
10 Additive operators
Denition. Let X and Y be linear spaces over K. An additive operator
from X to Y is a single-valued function T from a subspace M of X into
Y such that
T(x + y) = T x + Ty, x, y M, , K.
M is called the domain of T and is denoted by D(T); the set {z|z Y
10. Additive operators 11
such that z = T x for some x D(T)} is called the range of T and is
denoted by W(T).
If Y is the space of real or complex numbers (according as X is a
real or a complex linear space) and T is an additive operator from X to
Y we say that T is an additive functional.
Denition. Let X and Y be two normed linear spaces. An additive op- 10
erator T is said to be continuous at x
0
D(T) if for every sequence
{x
n
} D(T) with x
n
x
0
we have T x
n
T x
0
. An additive operator
is said to be continuous (on D(T)) if it is continuous at every point of
D(T). It is easy to see that an additive operator T is continuous on
D(T) if it is continuous at one point x
0
D(T).
Proposition . An additive operator T : X Y between two normed
linear spaces is continuous if and only if there exists a real number
> 0 such that
||T x|| ||x|| for every x D(T)
Proof. The suciency of the condition is evident, for if x
n
x
0
||T x
0

T x
n
|| = ||T(x
0
x
n
)|| ||x
0
x
n
|| 0.
Now assume that T is continuous. If there exists no as in the
proposition, then there exists a sequence {x
n
} D(T) such that ||T x
n
|| >
n||x
n
||. Since T(0) = 0, x
n
0. Dene y
n
= x
n
/

n||x
n
||. Then ||y
n
|| =
1

n
0 as n ; as T is continuous T
y
n
must tend to zero as n .
But Ty
n
=
1

n||x
n
||
T x
n
and ||Ty
n
|| =
1

n||x
n
||
||T x
n
|| >

n and so Ty
n
does not tend to zero. This is a contradiction.
Let T be an additive operator from a linear space X into a linear
space Y. T is one-one if and only if T x = 0 implies x = 0. If T is
one-one it has an inverse T
1
, which is an additive operator from Y into
X with domain w(T), dened by
T
1
y = x i f y = T x.
T
1
satises the relations T
1
T x = x for x D(T) and TT
1
y = y 11
12
for y D(T
1
) = W(T). If X and Y are normed linear spaces, T has a
continuous inverse if and only if there exists a > 0 such that ||T x||
||x|| for x D(T).
The sum of two operators T and S , with D(T), D(S ) x and
W(T), W(S ) Y is the operator (T + S ), with domain D(T) D(S ),
dened by:
(T + S )x = T x + S x.
Lecture 2
1 Linear operators
12
Denition. An additive operator T from a normed linear space X into
a normed linear space Y whose domain D(T) is the whole space X and
which is continuous is called a linear operator from X to Y. The norm
||T|| of a linear operator is by denition: ||T|| = ||T||
X
= sup
xX,||x||1
||T x||. If
Y is the real or complex numbers (according as X is a real or a complex
linear space) the linear operator T is called a linear functional on X.
So far we have proved the existence of non-trivial linear functionals.
We shall prove the Hahn-Banach extension theorem which will have as
a consequence the existence of many linear functionals on a normed
linear space.
2 Hahn-Banach lemma
Denition . Let X be a linear space (over real or complex numbers).
A real valued function p on X will be called a semi-group(or a sub-
additive functional) if it satises the following conditions:
i) p(x) = || p(x), for each K and x X.
ii) p(x + y) p(x) + p(y) for all x, y X.
Note that these conditions imply that p(x) 0 for all x X.
13
14 2. Lecture 2
3 Lemma (Hahn-Banach)
Let X be a real linear space and p a semi-norm on X. Let M be a (real)
subspace of X and f a real additive functional on M such that f (x)
p(x) for all x M. Then there exists a real additive functional F on 13
X such that F is an extension of f (i.e., F(x) = f (x) for x M) and
F(x) p(x) for all x X.
Proof. By the application of Zorns lemma or transnite induction, it is
enough to prove the lemma when X is spanned by M and an element
x
0
M, i.e., when
X = {M, x
0
} = {x|x X, x = m + x
0
, m M, real , x
0
M}.
The representation of an element x X in the form x = m+x
0
, (m
M, real) is unique. It follows that if, for any real number c, we dene
F(x) = f (m) + c,
then F(x) is an additive functional on X which is an extension of f (x).
We have now to choose c in such a way that F(x) p(x), x X, i.e.,
f (m) + c p(m + x
0
).
This condition is equivalent to the following two conditions:
_

_
f
_
m

_
+ c p
_
m

+ x
0
_
for > 0
f
_
m

_
c p
_
m

x
0
_
for < 0.
To satisfy these conditions, we shall choose c such that
f (m

) p(m

x
0
) c p(m

+ x
0
) f (m

)
for all m

, m

M. Such a choice of c is possible since

f (m

) + f (m

) = f (m

+ m

)
p(m

+ m

)
= p(m

x
0
+ m

+ x
0
)
4. Hahn-Banach extension theorem... 15
p(m

x
0
) + p(m

+ x
0
).
So f (m

) p(m

x
0
) p(m

+ x
0
) f (m

), m

, m

M.
So 14
sup
m

M
_
f (m

) p(m

x
0
)
_
inf
m

M
_
p(m

+ x
0
) f (m

)
_
and we can choose for c any number in between.
4 Hahn-Banach extension theorem for real normed
linear spaces
Theorem. Let X be a real normed linear space and M a real subspace
of X. Given a (real) linear functional f on M, we can extend f to a (real)
linear functional on the whole space X in such a way that the norm is
preserved:
||F|| = ||F||
X
= || f ||
M
.
Proof. Take p(x) = || f ||
M
||x|| in the Hahn-Banach lemma. We have
f (x) p(x) on M and p(x) is subadditive. We then have an additive
functional F(x) on X which is an extension of f with F(x) || f ||
M
||x||
for all x X. Also F(x) = F(x) || f ||
M
|| x|| = || f ||
M
||x||. Hence
|F(x)| || f ||
M
||x||.
This shows that F is a linear functional on X and ||F||
X
|| f ||
M
. The
reverse inequality, ||F||
X
|| f ||
M
, is trivial as F is an extension of f .
5 Hahn-Banach extension theorem for complex
normed linear spaces (Bohnenblust-Sobczyk)
Theorem. Let X be a complex normed linear space and M a (complex)
subspace. Given a complex linear functional f on M we can extend f to
a complex linear functional F on X in such a way that ||F||
X
= || f ||
M
.
16 2. Lecture 2
Proof. A complex normed linear space becomes a real normed linear 15
space if scalar multiplication is restricted to real numbers and the real
and imaginary parts of a complex linear functional are real linear func-
tionals. If f (x) = g(x) + ih(x) (g(x), h(x) real ), g and h are real linear
functionals on M and ||g||
M
|| f ||
M
, ||h||
M
|| f ||
M
. Since, for each
x M,
g(ix) + ih(ix) = f (ix)
= i f (x)
= i(g(x) + ih(x))
= h(x) + ig(x),
we have h(x) = g(ix), for x M.
By the Hahn-Banach theorem for real linear spaces g can be ex-
tended to a real linear functional G on X with the property ||G||
X
= ||g||
M
.
Now dene
F(x) = G(x) iG(ix).
F is then a complex linear functional on X. (For complex additivity
notice that
F(ix) = G(ix) iG(x) = G(ix) + iG(z) = iF(x)).
F is an extension of f ; for, if x M,
F(x) = G(x) iG(ix) = g(x) ig(ix) = g(x) + ih(x) = f (x).
We have now only to show that the norm is not changed. For this,
writes, for x X, F(x) = re
ie
. Then E
i
F(x) is real. So
|F(x)| = |e
i
F(x)| = |F(e
i
x)|
= |G(e
i
x)| (= since e
i
F(x)is real).
||G|| ||e
i
x||
= ||g||
M
||x||
|| f ||
M
x.
So ||F||
X
|| f ||
M
and the reverse inequality holds since F is an ex- 16
tension of f .
6. Existence of non-trivial linear functionals 17
6 Existence of non-trivial linear functionals
We consider some consequences of the Hahn-Banach extension theo-
rem; we prove the existence of plenty of linear functionals on a normed
linear space.
Proposition. Let X be a normed linear (real or complex) and x
o
0 be
an elements of X. Then there exists a linear functional f
o
on X such that
f
o
on X such that f
o
(x
o
) = ||x
o
|| and || f
o
|| = 1.
Proof. Let M be the subspace spanned by x
o
, i.e., M = {x|x = x
o
for
some number }. Dene f (x) = ||x
o
|| for x = x
o
M. This is a linear
functional on M and || f ||
M
= 1. By the Hahn-Banach extension theorem
there exists a linear functional f
o
on X which extends f in such a way
that || f
o
|| = || f ||
M
= 1; f
o
(x
o
) = f (x
o
) = ||x
o
||.
Remark . For a pre-Hilbert space the existence of such a linear func-
tional is evident; we may take f
o
(x) = (x,
x
o
||x
o
||
). The additivity of f
o
follows from the homogeneity and distributivity of the scalar product.
The continuity of f
o
is a consequence of Schwarzs inequality.
Proposition. Let X be a normed linear space. Let M be a subspace and
x
o
an element X such that d = inf
mM
||x
o
m|| > 0. Then there exists
a linear functional f
o
on X such that f
o
(x) = 0 for every x M and
f
o
(x
o
) = 1.
Proof. Let M

= {x|x = m + x

, m M}. Dene f (x) = for x = m +

x

(m M). f is additive on M

, vanishes on M and f (x

) = 1.
Also f is continuous on M

: if 0, then
x = m + x

0(m M), and 17

| f (x)| = || = ||x||
_
||x||
= ||||x||
_
||m + x
o
||
= ||x||
_
||x
o
(m/)||
d
1
||x||(m/ M);
18 2. Lecture 2
if = 0, f (x) = 0 and the inequality | f (x)| d
1
||x|| is still valid. If f
o
is a linear functional on X which is an extension of f , then f
o
satises
the requirements of the proposition.
7 Orthogonal projection and the Riesz representa-
tion theorem
Denition. Let x and y be two elements of a pre-Hilbert space X; we
say that x is orthogonal to y (written x y) if (x, y) = 0. If x y then
y x; if x x, then x = 0.
Let M be a subset of a pre-Hilbert space; we denote by M

the set
of elements x X such that x y for every y M.
Theorem. Let M be a closed liner subspace of a Hilbert space X. Then
any x
o
X can be decomposed uniquely in the form x
o
= m + n, m
M, n M

. (m is called the orthogonal projection of x

o
on M and is
denoted by P
M
x

).
Proof. The uniqueness of the decomposition is clear from the fact that
an element orthogonal to itself is zero. To prove the existence of the
decomposition we may assume M X and x
o
M (if x
o
M we have
the trivial decomposition with n = 0). Let d = inf
mM
||x
o
m||; since M is
closed and x
o
M, d > 0. Let {m
k
} M be a minimizing sequence, i.e.,
lim
k
||x
o
m
k
|| = d. {m
k
} is a Cauchy sequence; for
||m
k
m
n
||
2
= ||(x
o
m
n
) (x
o
m
k
)||
2
= 2(||x
o
m
n
||
2
+ ||x
o
m
k
||
2
) ||2x
o
m
k
m
n
||
2
((Euclidean property ))
= 2(||x
o
m
n
||
2
+ ||x
o
m
k
||
2
) 4||x
o

m
k
+ m
n
2
||
2
2(||x
o
m
n
||
2
+ ||x
o
m
k
||
2
4d
2
( as
m
k
+ m
n
2
M)
2(d
2
+ d
2
) 4d
2
= as m, n .
By the completeness of the Hilbert space there exists and element 18
7. Orthogonal projection and... 19
m X with lim
k
||m m
k
|| = 0; in fact m M, as M is closed. Also
||x
o
m|| = d. Write x
o
= m + (x
o
m). Putting n = x
0
m we have to
show that n M

. Let m

M. Since, for any real , m + m

M we
have d
2
||x
o
m m

||
2
= ||n m

||
2
= (n m

, n m

)
= ||n||
2
(n, m

) (m

, n) +
2
||m

||
2
.
Since ||n||
2
= d
2
, this gives, for any real ,
0 2R(n, m

) +
2
||m

||
2
.
So R(n, m

) = 0 for every m

M. Replacing m

by im

we have
Im(n, m

) = 0, for every m

M. Thus (n, m

) = 0 for each m

M.
Remark . If x
o
M, then n 0 and f
o
(x) = (x,
n
||n||
2
) satises the
conditions of the last proposition.
Theorem Riesz. Let X be a Hilbert space and f a linear functional on
X. Then there exists a unique element y
f
of X such that
f (x) = (x, y
f
)
for every x X.
Proof. Uniqueness: If (x, y
1
) = (x, y
2
) for every x, (x, y
1
y
2
) = 0 for 19
every x; choosing x = y
1
y
2
we nd y
1
y
2
= 0.
Existence: Let M be the zero manifold of f , i,e,., M = {x| f (x) = 0}.
Since f is additive, M is a linear subspace and since f is continuous M
is closed. The theorem is evident if M = X. i.e., if f (x) = 0 on X; in
this case we need only take y
f
= 0. So suppose M X. Then there
exists, by the last theorem, an element y
0
0 such that y
o
is orthogonal
to every element of M. Dene
y
f
=
f (y
o
)
||y
o
||
2
Y
o
.
20 2. Lecture 2
y
f
meets the condition of the theorem. First, for x M, f (x) = (x, y
f
)
since f (x) = 0 for x M and y
f
M

. For elements x of the form

x = y
0
.
(x, y
f
) = (y
o
, y
f
) =
_

_
,
f (y
0
)
||y
0
||
2
y
0
_

_
= f (y
o
) = f (y
o
)
= f (x).
Since f is linear and (x, y
f
) is linear and (x, y
f
) is linear in x, to show
that f (x) = (x, y
f
) for each x X it is enough to show that X is spanned
by M and y
o
. If x X, write, noting that f (y
f
) 0,
x =
f (x)
f (y
f
)
y
f
+
_
x
f (x)
f (y
f
)
y
f
_
.
f (x)
f (y
f
)
f
y
is of the form y
o
. The second term is an element of M, since
f
_
x
f (x)
f (y
f
)
y
f
_
= f (x)
f (x)
f (y
f
)
y
f
= 0.
Remark.
|| f || = ||y
f
||.
Lecture 3
1 The Conjugate space (dual) of a normed linear
space
Let X be a normed linear space. Let X

be the totality of all linear 20

functionals on X. X

is a linear space with the operations dened by:

( f + g)(x) = f (x) + g(x) f , g X

, x X
(f )(x) = . f (x).
X

is a Banach space with the norm

|| f || = sup
||x||1
| f (x)| ( f X

, x X).
We call the Banach space X

the conjugate space of X.

2 The Resonance Theorem
Lemma Gelfand. Let p(x) be a semi -norm on a Banach space X. Then
there exists a number > o such that
p(x) ||x|| for all x X
if and only if p(x) is lower semi - continuous. (Lower semi - continuity
means this); for any x

and any E > 0, there exists a = (x, E) > 0

such that p(x) p(x
o
) E for ||x x
o
|| .
21
22 3. Lecture 3
Proof. i) Suppose p(x) ||x|| for all x X, > 0; then
p(x
o
) = p(x
o
x + x) p(x
o
x) + p(x)
||x x
o
|| + p(x)
p(x) + , if ||x x
0
|| E/ = .
ii) Conversely assume that p(x) is lower semi - continuous.
To prove that there is a > 0 such that p(x) ||x|| for every x X
it is sucient to show that p(x) is bounded, say by P
1
, in some 21
closed sphere K of positive radius (K =
_
x| ||x x
o
||
_
. For if
x X with ||x|| , then x
o
and x
o
+ x both belong to K and hence
p(x) = p(x
o
+ x
o
+ x) p(x
0
) + p(x
o
+ x)
= p(x
o
) + p(x
o
+ x)
2
1
;
if x is an arbitrary element of X
p(x) = p
_
||x||

x
||x||
_
=
||x||

p
x
||x||

2P
1

||x||
_
as||
x
||x||
|| =
_
and choose = 2
1
/.
Now we assume that p(x) is unbounded in every closed sphere of
positive radius and derive a contradiction. Let
K
o
= {x

||x x
o
|| , > 0};
there exists in interior point x
1
of K
o
such that p(x
1
) > 1. By the lower
semi - continuity of p, there exists a closed sphere K
1
=
_
x; ||x x
1
||

1
< 1,
1
> 0
_
, K
1
K
o
such that p(x) > 1 for each x K
1
. By a
repetition of this argument we may choose a sequence of closed spheres
K
n
=
_
x; ||x x
n
||
n
< 1/n,
n
> 0
_
, n running through all positive
integers, such that K
n
K
n1
and p(x) > n for each x K
n
. For m, m

>
n, Since x
m
, x

m
K
n
, we have ||x
m
x

m
|| ||x
m
x
n
|| + ||x

m
x
n
||
3. Weak convergence 23
2
n
< 2/n; so x
n
is a Cauchy sequence. Since X is complete there exists
an x

X such that s lim

n
x
n
= x

. As ||x
m
x
n
||
n
for m > n, we
have, passing to the limits, ||x

x
n
||
n
. So x

_
n=1
K
n
; this would
mean that p(x

) (which is a real number) is greater than every positive

integer n, which is absurd.
The Resonance theorem: Let X be a Banach space and Y
n
(n = 1, 2, . . .) 22
a sequence of normed linear spaces. Let, for each n, T
n
be a linear
operator from X to Y
n
. Then the boundedness of the sequences
_
||T
n
x||
_
for every x X impels the boundedness of the sequence
_
||T
n
||
_
.
Proof. For each x X, sup
n
||T
n
(x)|| is nite as {||T
n
(x)||} is bounded.
Dene p(x) = sup
n
T
n
(x); p(x) is a semi-norm on X. p(x) is also lower
semi-continuous since it is the supremum of the sequence of continuous
functions
_
||T
n
||
_
. Consequently, by Gelfands lemma, p(x) ||x|| (for
some > o) for such x X; so ||T
n
(x)|| ||x|| for each n and each
x X. Thus ||T
n
|| .
Corollary. Let X be a Banach space Y a normed linear space, and
_
T
n
_
a sequence of linear operators form X to Y. Assume that s lim
n
T
n
(x)
Y exists for each x X. If we dene T
x
= s lim
n
T
n
(x) then T is a
linear operator from X to Y and ||T|| lim
n
||T
n
||.
T is evidently additive. By the Resonance theorem, ||T
n
(x)|| ||x||
( > 0); so||T(x)|| ||x||, i.e., T is continuous. Further, ||T
n
x||
||T
n
||||x||; so ||T x|| lim||T
n
||||x|. Hence ||T|| lim
n
||T
n
||.
3 Weak convergence
Denition . Let X be a normed linear space; we say that a sequence.
{x
n
} X converges weakly to x

X (and write w lim

n
x
n
= x

) if, for
every linear functional f on X, we have lim
n
f (x
n
) = f (x

).
24 3. Lecture 3
Proposition. i) w lim
n
x
n
, if it exists, is unique.
ii) s lim x
n
= x

implies w lim x
n
= x

.
(The converse is not true in general).
iii) if w lim
n
x
n
= x

then lim ||x

n
|| x

. 23
Proof. (i) Let w lim x
n
= x

, w lim x
n
= x

, x

. By the Hahn
-Banach theorem there exists a linear functional f on X such that
f (x

) 0 i.e., f (x

) f (x

). But by the condition of weak

limit we must have f (x

) = lim
n
f (x
n
) = f (x

).
(ii) This follows form the inequality:
| f (x

) f (x
n
)| = f (x

x
n
) || f || ||x

x
n
||,
for each f X

.
(iii) Let f
o
X

with || f || = 1 and f
o
(x

) = ||x

||.
Then
||x

|| = f
o
(x

) lim| f
o
(x
n
)|
lim|| f
o
|| ||x
n
||
= lim
n
||x
n
||.

4 A counter-example
We shall now show by an example that weak convergence does not im-
ply strong convergence in general. Consider the sequence {sin nt} in
L
2
(0, 1) (real). This sequence converges weakly to zero. Since, by
the Riesz theorem, any linear functional is given by the scalar prod-
uct with a function we have to show that
1
_

f (t) sin ntdt 0, for each

4. A counter-example 25
f L
2
(0, 1). But By Bessels inequality,

n=1

_
1
0
f (t) sin ntdt

_
1
0
| f (t)|
2
dt;
so
1
_
0
f (t) sin ntdt 0 as n . But {sin nt} is not strongly conver-
gent, since
|| sin nt sin mt||
2
=
_
1
0
| sin nt sin mt|
2
dt
= 2 for n m.
Lecture 4
1 Local weak compactness of a Hilbert space
24
Theorem. Let {x
n
} be a bounded sequence of elements of a Hilbert space
(i.e., ||x
n
|| C < , n = 1, 2, . . .); then we can choose a subsequence of
{x
n
} which converges weakly to an element of X.
Proof. Let M be the closed linear space spanned by {x
n
}. (M is the clo-
sure in the sense of the norm of the set of all nite linear combinations
_

i
x
i
of the elements{x
i
}). M is separable, there exists a countable set
of elements {y
n
} which is dense in M. We may take for example, the ra-
tional linear combinations of {x
i
} if X is real and if X is complex, linear
combinations of {x
i
} with coecients of the form p+iq, p, q rational.
For each y
k
from {y
n
} the sequence
_
(x
n
, y
k
)
_
is bounded ; |(x
n
.y
k
)|
||x
n
||||y
k
|| C||y
k
||. By the Bolzano - Weierstrass theorem and a diagonal
process we can nd a subsequence {x

n
} of {x
n
} such that
_
(x

n
, y
k
)
_
con-
verges for every k. Actually {(x

n
, z)} converges for each x X. To prove
this, let z = y + where y = P
M
z, M

. Then (x
n
, z) = (x
n
, y) and
we have to prove that {(x
n
, y)}(y M) is convergent. We have
|(x
n
x
m
, y)| = |(x
n
x
m
, y y
k
+ y
k
)|
|(x
n
x
m
, y
k
)| + |(x
n
x
m
., y y
k
)|
|(x
n
x
m
, y
k
)| + ||x
n
x
m
., y y
k
||
|(x
n
x
m
, y
k
)| + 2C||y y
k
||.
Since {(x
n
, y
k
)} is convergent and {y
k
} is dense in M, it follows that 25
27
28 4. Lecture 4
(x
n
, y) is a Cauchy sequence; so {(x
n
, y)} is convergent. Dene g(z) =
lim
n
and | f (z)| = |g(z)| = lim
n

|(x
n
., z)| C||z||, f (z) is continuous. By
the Riesz theorem there exists and element x

X such that f (z) =

(z, x

) for each z X. Since limn

(x
n
, z) = (x

, z) for each z
X, w lim
n
x

n
= x

(by Rieszs Theorem)

We mention without proof that L
p
(, ), 1 < p < is locally
weakly compact. But L(, ), L

(, ) and C[, ] are not locally weak-

ly compact.
We next prove a theorem which will be needed in the study of
Cauchys problem.
2 Lax-Milgram theorem
Let B(u, v) be a bilinear functional on a real Hilbert space X such that
(i) there exists a > 0 such that |B(u, v)| ||u||||v|| for all u, v X,
(ii) there exists a > 0 such that ||u||
2
B(u, u) for each u X.
Then there exists a linear operator S from X to X such that
(u, v) = B(u, S v)
and ||S ||
1
.
Proof. Let V be the set of elements v for which there exists an element
v

such that (u, v) = B(u, v

) for all u X. (V is non-empty; 0 V). 26

v

is uniquely determined by v. For, if w X be such that B(u, w) = 0

for all u, then w = 0 as ||w
2
|| B(w, w) = 0 or ||w|| = 0. V is a linear
subspace. We have an additive operator S with domain V, dened by
S v = v

. S is continuous;
||S v||
2
B(S v, S v) = (S v, v) ||S v||||v||
so that ||S v||
1
||v|| (if ||S v|| = 0 this is trivially true). Moreover V
is closed subspace of X. For, of v
n
V and v
n
v X, then S v
n
is
a Cauchy sequences and so has a limit v

; but (u, v
n
) (u, v) and by
2. Lax-Milgram theorem 29
(i) B(u, S v
n
) B(u, v

) so that (u, v) = B(u, v

) for each u; so v V.
The proof will be complete if we show that V = X. Suppose V X.
Then there exists w X such that w 0 and (w, v) = 0 for each v V.
Consider the functional, as
|F(z)| = |B(z, w)| ||z||||w||.
So by Rieszs theorem, there exists, w

X such that B(z, w) =

(z, w

) for each z X. So w

V and S w

= w. So
||w||
2
B(w, w) = (w, w

)
= 0,
i.e., w = 0
which is a contradiction.
Part II
Semi-group Theory
31
33
Denition. Let {T
t
}
t0
be a one-parameter family of linear operators on 27
a Banach space X into itself satisfying the following conditions:
(1) T
t
T
s
= T
t+s
, T
o
= I, I denoting the identity operator on X (Semi
-group property).
(2) s lim
tt
0
T
t
x = T
t

x 0 and each x X(strong continuity).

(3) there exists a real number 0 such that ||T
t
|| e
t
for t 0.
We call such a family {T
t
} a semi group of linear operators of normal
type on the Banach space X, or simply a semi - group.
Remark. The third condition may look a bit curious but it is nothing but
a restriction of the order of ||T
t
|| near t = 0, because we can prove the
following.
Proposition. The two conditions (1) and (2) imply the following:
(3

) lim
t
t
1
log ||T
t
|| = < ( may be).
(4) ||T
t
|| is bounded in any bounded interval [0, t
o
], o < t
o
< .
Proof. We rst prove (4). Suppose ||T
t
|| is unbounded in some inter-
val [0, t
o
], 0, < t
o
< . Then there would exist a sequence {t
n
} (n =
11, 2, . . .) such that T
t
n
n and o lim
n
t
n
= t

< . Since
_
||T
t
n
||
_
is unbounded, by the resonance theorem,
_
||T
t
n
x||
_
is unbounded
at least for one x X; but by strong continuity, s lim
n
T
t
n
x = T
t

x for
each x X. This is a contradiction.
To prove (3

), let p(t) = log ||T

t
||, p(t) < (may be ). Since 28
||T
t+s
|| = ||T
t
T
s
|| ||T
t
||||T
s
||, we have p(t + s) p(t) + p(s). Let
inf
t>0
t
1
p(t). is either nite or . We shall show that lim
tt
1
p(t)
ex-
ists and is equal to . Assume, rst, is nite. Choose for any E > o,
a number a > o in such a way that p(a) ( + E)a. Let n be an integer
such that na t < (n + 1)a.
Then

p(t)
t

p(na)
t
+
p(t na)
t
34

na
t
p(a)
a
+
p(t na)
t

na
t
( + E) +
p(t na)
t
.
Letting t ,
p(t na)
t
tends to zero since p(t na) is bounded
from above (since, as we have proved above, ||T
s
|| is bounded in any
nite interval of s). Thus lim
t
t
1
p(t) = . The case = can be
treated similarly.
Lecture 5
1 Some examples of semi-groups
29
I In C[o, ] [ the space of bounded uniformly continuous functions
on the closed interval [0, ]] dene
_
T
t
_
t0
by
(T
t
x)(s) = x(t + s) (x C).
_
T
t
_
is a semi-group. Condition (1) is trivially veried. (2) follows
from the uniform continuity of x, as
||T
t
x T
t
o
x|| = sup
s0
|x(t + s) x(t
o
+ s)|.
Finally ||T
t
|| = 1 and so (3) is satised with = 0.
In this example, we could replace C[0, ] by C[, ].
II On the space C[o, ] (or C[, ]), dene
_
T
t
_
t 0
(T
t
x)(s) = e
t
x(s)
where is a xed non-negative number. Again (1) is trivial; for (2)
we have ||T
t
x T
t

x|| =

e
t
e
t
o

sup
s
|x(s)|. Trivially ||T
t
|| = e
t
.
III Consider the space C[, ]. Let
N
t
(u) =
1

2t
e
u
2
/2t
, = < u < , t > 0,
35
36 5. Lecture 5
(the normal probability density). Dene
_
T
t
_
t0
on C[, ] by:
(T
t
x)(s) =
_

N
t
(s u)x(u)du, for t > 0
x(s) for t = 0
Each T
t
is continuous: 30
||T
t
x|| ||x||

N
t
(s u)du = ||x||, as

N
t
(s u)du = 1.
Moreover it follows from this that condition (3) is valid with = 0.
By denition T
o
= I and the semi-group property T
t
T
s
= T
t+s
is a
consequence of the well -known formula concerning the Gaussian
distribution.
1

2(t + t

)
e
u
2
/2(t+t

)
=
1

2t
1

2t

e
(uv)
2
2t
e
v
2
2t

dv.
(Apply Fubinis theorem). To prove the strong continuity, consider
t, t
o
> 0 with t t
0
. (The case t
o
= 0) is treated in a similar
fashion). By denition
(T
t
x)(s) (T
t
o
x)(s) =

_
N
t
(s u)x(u) N
t
o
(s u)x(u)
_
du.
The integral

2t
e
(su)
2
/2t
x(u)du becomes, by the change of
variable
s u

t
= z,
1

e
z
2
/2
x(s

tz)dz. Hence
(T
t
x)(s) (T
t
0
x) (s)) =

N
1
(z)
_
x
_
s

tz
_
x(s

t
o
z)
_
dz x(s)
1. Some examples of semi-groups 37
being uniformly continuous on , , for any > 0 there exists
a number = () > 0 such that |x(s
1
) x(s
2
)| whenever
|s
1
s
2
| . Now, splitting the last integral
|(T
t
x)(s) (T
t

x)(s)|

_
|

tz

t
o
z|
N
1
(z)|x(s

tz) x(s

zdz
+
_
|

tz

z|>
N
1
(z)(. . .)dz
E
_
N
1
(z)dz + 2||x||
_
|

tz

t
o
z>
N
1
(z)dz
= E + 2||x||
_
|z|>|

t
0
|
The second term on the right tends to 0 as |t t
o
| 0, because the 31
integral

N
1
(z)dz converges. Thus
lim
tt
o
sup
<s<
|(T
t
x)(s) (T
t
0
x)(s)| E.
Since E > 0 was arbitrary, we have proved the strong continuity at
t = t
o
of T
t
.
In this example we can also replace C[o, ] by L
p
[o, ] 1 p <
. Consider, for example L
1
[o, ]. In this case, ||T
t
x||

_
_

N
t
(s u)|x(u)|ds
_
du ||x||, applying Fubinis theorem.
As for the strong continuity, we have
(T
t
x)(s) (T
t
o
x)(s)||
38 5. Lecture 5
=

N
1
(z)
_
x(s

tz) x(s

t
o
z)
_
dz|ds

N
1
(z)
_

|x(s

tz) x(s

t
o
z)|ds
_

_
dz
Since N
1
(z)

|x(s

tz) x(s

t
o
z)|ds 2||x||N
1
(z), we may
apply Lebesgues dominated convergence theorem. We then have
lim
tt

||(T
t
x)(S ) (T
t
0
x)(S )||

N
1
(z)
_

_
lim
tt
o

|x(

tz) x(s

t
o
z)|ds
_

_
dz = 0,
by the continuity in mean of the Lebesgue integral.
IV Consider C[, ]. Let > 0, > 0. Dene
_
T
t
_
t0
(T
t
x)(s) = e
t

k=0
(t)
k
k!
x(s k).
{T
t
} is a semi-group. Strong continuity follows from:
| (T
t
X)(s) T
t
0
X)(s) | x|e
t

k=0
(t)
k
k!
e
t
0

k=0
(t
0
)
k
k!
| = 0.
(3) is satised with = 0. To verify (1) 32
(T
w
(t
t
x))(s) = e
w

l=0
(w)
l
l!
_

_
e
t

k=0
(t)
k
k!
f (s k 1)
_

_
= e
(w+t)

p=0
1
p!
_

_
p!
p

1=0
(w)
1
(t)
p1
1! p 1!
f (s p)
_

_
= e
(w+t)

p=0
1
p!
(w + t)
p
f ( + t)
p
f (s p)
2. The innitesimal generator of a semi-group 39
= (T
w+t
x)(s).
2 The innitesimal generator of a semi-group
Denition. The innitesimal generator A of a semi-group T
t
is dened
by:
Ax = s lim
h0
h
1
(T
h
I)x,
i.e., as the additive operator A whose domain is the set
D(A) =
_
x | s lim
h0
h
1
(T
h
I)x exists
_
and for x D(A),
Ax = s lim
ho
h
1
(T
h
I)x.
D(A) is evidently non- empty; it contains at least zero. Actually
D(A) is larger. We prove the
Proposition. D(A) is dense in X ( in the norm topology ).
Proof. Let
n
(s) = ne
ns
. Introduce the linear operator C

n
dened by
C

n
x =

_
0

n
(s)T
s
xds for x X and n > ,
the integral being taken in the sense of Riemann. (The ordinary proce-
dure of dening the Riemann integral of a real or complex valued func- 33
tions can be extended to a function with values in a Banach space, using
the norm instead of absolute value ). The convergence of the integral is
a consequence of the strong continuity of T
s
in s and the inequality,
||
n
(s)T
s
x || ne
(n+)s
| x || .

40 5. Lecture 5
The operator C

n
is a linear operator whose norm satises the in-
equality
||
n
|| n

_
0
e
(n+)s
ds = 1/1 /n.
We shall now show that W(C

n
) D(A) (W(C

n
) denotes the range
of C

n
) for each n > and that for each x X, s lim
n
C

n
x = x; then
_
n>
WC

n
) will be dense in X and a-portion D(A) will be dense in X. We
have
h
1
(T
h
I)C

n
x = h
1

_
0

n
(s)T
h
T
s
xds h
1

_
0

n
(s)T
s
xds
(The change of the order T
h

_
0
=

_
0
T
h
is justied, using the additiv-
ity and the continuity of T
h
, by approximating the integral by Riemann
sums). Then
h
1
(T
h
I)C

n
x = h
1

_
0

n
(s)T
h+s
xds h
1

_
0

n
(s)T
s
xds
= h
1

_
0

n
(s h)T
s
xds h
1

_
0

n
(s)T
s
xds
( by a change of variable in the rst integral ).
= h
1

_
h
{
n
(s h)
n
(s)}T
s
xds
= h
1
h
_
0

n
(s)T
s
xds.
By the strong continuity of
n
(s)T
s
x in s, the second term on the 34
2. The innitesimal generator of a semi-group 41
right converges strongly to
n
(0)T
0
x = nx, as h 0.
h
1

_
h
{
n
(s h)
n
(s)}T
s
xds
=

_
h

n
(s h)T
s
xds (0 < < 1) ( by the mean value theorem )
=

_
0

n
(s)T
s
xds +
h
_
0

n
(s)T
s
xds +

_
h
{

n
(s)

n
(s h)}T
s
xds.
But,
h
_
0

n
(s)T
s
xds 0 as h 0 and
||

_
h
_

n
(s)

n
(s h)
_
T
s
xds ||
n
2

_
h
| e
n(sh)
e
ns
| e
s
|| x || ds
n
2
(e
nh
1)

_
h
e
(n)s
|| x || ds 0 as h 0.( < n).
Thus we have proved that W(C

n
) D(A) and
AC

n
x = n(C

n
I)x
as

n
= n
n
. Next, we show that s lim
n
C

n
(x) = x for each x X.
We observe that
C

n
x x =

_
0
ne
ns
T
s
xds

_
0
ne
ns
xds, ( as

_
0
ne
ns
ds = 1)
42 5. Lecture 5
= n

_
0
e
ns
_
T
s
x x
_
ds.
Approximating the integral by Riemann sums and using the triangle
inequality we have
|| C

n
x x || n

_
0
e
ns
|| T
s
x x || ds
= n

_
0
+ n

, > 0
= I
1
+ I
2
, say .
Given E > 0, by strong continuity, we can choose a > 0 such that 35
|| T
s
x x ||< E for 0 s ; then
I
1
En

_
0
e
ns
ds En

_
0
e
ns
ds = E.
For a xed > 0, using the majorization condition in the denition
of a semi-group,
I
2
n

e
ns
(e
s
+ 1) || x || ds =|| x ||
_
n
e
(n+)s
n
_

|| x ||
_
n
e
ns
n
_

Each of the terms on the right tends to zero as n . So I

2
E,
for n > n
0
. Thus C

n
x x as n .
Remark. That D(A) is dense in X can be proved more easily. But we
need the considerations given in the above proof for later purpose.
Denition. For x X dene D
t
T
t
x by
D
t
T
t
x = s lim
h0
h
1
(T
t+h
T
t
)x
if the limit exists.
2. The innitesimal generator of a semi-group 43
Proposition. If x D(A) then x D(D
t
) and D
t
T
t
x = AT
t
x = T
t
Ax.
Proof. If x D(A), we have, since T
t
is a linear operator,
T
t
Ax = T
t
s lim
h0
h
1
(T
h
I)x
= s lim
h0
h
1
(T
t
T
h
T
t
)x
= s lim
h0
h
1
(T
t+h
T
t
)x
= s lim
h0
h
1
(T
h
I)T
t
x = AT
t
x.
Thus, if x D(A), then T
t
x D(A), and T
t
Ax = AT
t
x = s 36
lim
h0
h
1
(T
t+h
T
t
)x. We have now proved that the strong right derivative
of T
t
x exists for each x D(A). We shall now show that the strong left
derivative exists and is equal to the right derivative. For this, take any
f X

. For xed x, f (T
t
x) is a continuous numerical function (real or
complex - valued ) on t 0. By the above. f (T
t
x) has right derivative
d
+
f (T
t
x)
dt
and
d
+
f (T
t
x)
dt
= f (AT
t
x) = f (T
t
A x).
But f (T
t
A x) is a continuous function. It is well-known that if one
of the Dini-derivatives of a numerical function is ( nite and ) continu-
ous, then the function is dierentiable ( and the derivative, of course, is
continuous ). So f (T
t
x) is dierentiable in t and
f (T
t
x x) = f (T
t
x) f (T
0
x)
=
t
_
0
d
+
f (T
s
x)
ds
ds =
t
_
0
f (T
s
Ax)ds
= f
_

_
t
_
0
A x ds
_

_
.
44 5. Lecture 5
However, if every linear functional vanishes on an element x X,
then x = 0 ( by Hahn - Banach theorem ). Consequently,
T
t
x x =
t
_
0
T
s
Axds.
for each x D(A). Since T
s
is strongly continuous in s, it follows from
this, that T
t
is strongly derivable:
D
t
T
t
x = s lim
h0
h
1
(T
t+h
T
t
)x
= s lim
h0
h
1
_
t+h
t
T
s
Axds
= T
t
Ax.
Lecture 6
Theorem . For n > , the operator (I n
1
A) admits of an inverse 37
J
n
= (I n
1
A)
1
which is linear and satises the relation
J
n
x = n

_
0
e
ns
T
s
xds, for x X (i.e., J
n
= C

n
Also || J
n
|| (1
n
1
)
1
.
Proof. We rst show that (I n
1
A)
1
exists [i.e., (I n
1
A) is one
-one]. If (I n
1
A) is not one-one, there will exist x
0
D(A) such
that || x
0
||= 1 and (I n
1
A)x
0
= 0, i.e., Ax
0
= nx
0
. Let f
0
be
a linear functional on X such that || f
0
||= 1 and f
0
(x
0
) = 1. Dene
(t) = f
0
(T
t
x
0
) = 1. Dene (t) = f
0
(T
t
x
0
). Since x
0
D(A), (t) is
dierentiable and
d(t)
dt
= f

(D
t
T
t
x

) = f

(T
t
Ax

) = f

(T
t
nx

)
= nf

(T
t
x

)
= n(t).

Solving this dierential equation with the initial condition (0) = 1

we get (t) = e
nt
. On the other hand we have
| (t) |=| f
0
(T
t
x
0
) | || f
0
|| || T
t
|| || x
0
||
e
t
;
since (t) = e
nt
and n > this is impossible. So (I n
1
A)
1
exists.
45
46 6. Lecture 6
Since A C

n
x = n(C

n
I)x, we have (I
n
n
1
A)C

n
x = x for all
x X. So (I n
1
A) maps W(C
n
) D(A) on to X; thus (I n
1
A) maps 38
D(A) in a one-one manner onto X. It follows that M(C

n
) D(A) and
(I n
1
A)
1
= C

n
. But C

n
is a linear operator and we have already
proved that || C

n
|| (1 n
1
)
1
.
Corollary.
M(C

n
) = D(A)
AJ
n
x = n(J
n
I)X, x X.
AJ
n
x = J
n
Ax = n(J
n
I)x, x D(A)
s lim
n
J
n
x = x, x X,
D
t
T
t
x = s lim
h0
h
1
(T
t+h
T
t
)x = AT
t
x = T
t
Ax, x D(A).
1 The resolvent set and the spectrum of an additive
operator on a Banach space
We may state our theorem in the terminology of spectral theory.
Let A be an additive operator ( with domain D(A)) from a Banach
space X into X. Let be a complex number ( is assumed to be real if X
is a real space ). Regarding the inverse of the additive operator (I A)
there are various possibilities.
(1) (I A) does not admit of an inverse, i.e., there exists an x 0
such that Ax = x. We then call an eigenvalue of A and x an
eigenvector belonging to the eigenvalue . In this case we also say
that is in the point-spectrum of A.
(2) When (I A)
1
exists there are three possibilities:
(i) D((I A)
1
) is not dense in X. Then is said to be in the
residual spectrum of A.
(ii) D((I A)
1
) is dense in X but (I A)
1
is not continuous. 39
In this case is said to be in the continuous spectrum.
2. Examples 47
(iii) D((I A)
1
) is dense in X and (I A)
1
is continuous in
D((I A)
1
). Then (I A)
1
can be extended uniquely to
a linear operator on the whole space X. In this case is said
to be in the resolvent set; the inverse (I A)
1
is called the
resolvent.
The complement of the resolvent set in the complex plane (or in the
real line if X is real) is called the spectrum of A.
The rst part of the theorem proved above says that if {T
t
} is a semi-
group of normal type (|| T
t
|| e
t
) any number > is in the resolvent
set of the innitesimal generator A.
2 Examples
Using these results we now determine the innitesimal generators of the
semi-groups we considered earlier.
I : C[0, ] : (T
t
x) (s) = x(t + s)
Writing y
n
(s) = (J
n
x)(s) we have
y
n
(s) = n

_
0
e
nt
x(t + s)dt
= n

_
s
e
n(ts)
x(t)dt :
y

n
(s) = ne
n(ss)
x(s) + n
2

_
s
e
n(ts)
x(t)dt
= nx(s) + ny
n
(s)
Comparing this with the general formula
(AJ
n
x)(s) = n((J
n
I)x)(s)
or Ay
n
(s) = ny
n
(s) nx(s)
48 6. Lecture 6
we have Ay
n
(s) = y

n
(s).
For n > , W(J
n
) = D(A). So if y D(A), y

(s) exists and belongs 40

to C[0, ] and
(Ay)(s) = y

(s).
Conversely let y(s) and y

(s) both belong to C[0, ]; we shall show

that y D(A) and (Ay)(s) = y

(s). For dene x(s) by

y

(s) ny(s) = nx(s).

Putting (J
n
x)(s) = y
n
(s), we have, as shown above,
y

n
(s) ny
n
(s) = nx(s).
Writing (s) = y(s) y
n
(s), we obtain

(s) n(s) = 0
or (s) = Ce
ns
. But (s) C[0, ] and this is possible only if C = 0.
Hence y(s) = y
n
(s) D(A) and so (Ay)(s) = y

(s). Thus the domain of

the innitesimal generator A is precisely the set of functions y C[0, ]
and for such a function Ay = y

. We have thus characterized the dier-

ential operator
d
dt
as the innitesimal generator of the semigroup asso-
ciated with the translation by t.
II. In this we give the characterization of the second derivation as the
innitesimal generator of the semi-group associated with the Gaussian
distribution. The space is C[, ] and
(T
t
x)(s) =
_

2t
e
(sv)
2
/2t
x (v) dv if t > 0
x(s) if t = 0.
We have 41
y
n
(s) = (J
n
x)s =

x(v)
_

_
0
n

2t
e
nt(sv)
2
/2t
dt
_

_
dv
2. Examples 49
=

x(v)
_

_
0
2

2
e

2
(sv)
2
n/2
2
d
_

_
dv
(change: t =
2
/n)
Assuming for moment the formula

_
0
e
(
2
+c/
2
)
d =

2
e
2c
, c > 0, with c =

n
| s v |

2
,
we get
y
n
(s) =

x(v)
_

n/2e

2n|sv|
_
dv
=

n/2

x(v)e

2n|sv|
=

2
2
_

_
s
_

_
s

_

_
x(v) being continuous we can dierentiate twice and we then obtain
y

n
(s) = n
_

_
s
x(v)e
2

n(vs)
dv
s
_

x(v)e

2n(sv)
dv
_

_
y

n
(s) = n
_

_
x(s) x(s) + 2

_
s
x(v)e

2n(vs)
dv
+

2n
s
_

x(v)e

2n(sv)
dv
_

_
= 2nx(s) + 2ny
n
(s).
Comparing this with the general formula
(Ay
n
)(s) = (AJ
n
x)(s) = n
_
(J
n
I)x
_
(s)
= n(y
n
(s) x(s))
50 6. Lecture 6
we nd that Ay
n
(s) =
1
2
y

n
(s). For n > , W(J
n
) = D(A). Thus if
y D(A), y

(s) exists and belongs to C[, ] and further (Ay)(s) =

1
2
y

(s). Conversely, let y(s) and y

(s) both belong to C[, ]. Dene

x(s) by
y

(s) 2ny(s) = 2nx(s).

Putting y
n
(s) = (J
n
x)(s), we have, as shown above, 42
y

n
(s) 2ny
n
(s) = 2nx(s).
So, if (s) = y
n
(s) y(s),

(s) 2n(s) = 0.
This (s) = C
1
e

2ns
+ C
2
e

2ns
.
This function cannot be bounded unless both C
1
and C
2
are zero.
Hence y(s) = y
n
(s). So y(s) D(A) and (Ay)(s) =
1
2
y

(s).
Thus the dierential operator
1
2
d
2
dt
2
is the innitesimal generator of
the semi-group associated with the Gaussian process.
We now prove the formula

_
0
e
(
2
+c
2
/
2
)
d =

/
2
e
2c
, c > 0.
We start with the formula

_
0
e
x
2
dx =

/2.
Putting x = c/, we have

2
=

c
e
(c/)
2
(1 + c/
2
)d
2. Examples 51
= e
2c

c
e
(
2
+c
2
/
2
)
(1 + c/
2
)d
= e
2c
_

c
e
(
2
+c
2
/
2
)
d +

c
e
(
2
+c
2
/
2
)
c/
2
d
_

_
Setting = c/t in the last integral

2
= e
2c
_

c
e
(
2
+c
2

2
)
d

c
e
(c
2
/t
2
+t
2
)
dt
= e
2c

_
0
e
(
2
+c
2
/
2
)
d.
Lecture 7
1 The exponential of a linear operator
Example III . In C[, ] consider the semi-group associated with 43
Poison process, viz.,
(T
t
x)(s) = e
t

(t)
k
k!
x(s k) , > 0
Since e
t

_
k=0
(t)
k
k!
= 1, we have
(T
t
x) (s) x(s)
t
=
e
t
t

k=0
(t)
k
k!
(x(s k) x(s))
=
e
t
t
(x(s k) x(s))
+
e
t
t

k=2
(t)
k!
(x(s k) x(s)).
As t 0 the rst term on the right tends uniformly with respect to s
to (x(s ) x(s)); the absolute value of the second term is majorized
by 2 || x ||
e
t
t

_
k=2
(t)
k
k!
which tends to zero as t 0. Thus for any
x C[, ], we have Ax = (x(s ) x(s)). So in this case the
innitesimal generator is the linear operator dened by:
(Ax)(s) = [x(s ) x(s)],
53
54 7. Lecture 7
for x C[, ].
This is the dierence generator.
We now intend to represent the original semi-group {T
t
} by its in-
nitesimal generator. We expect, by analogy with the case of the ordi-
nary exponential function, the result to be given by
T
t
x = exp(tA)x.
But in general A is not dened over the whole space. So if we at-
tempt to dene (exp t A)x by a power series

_
k=0
(tA)
k
k!
x, we encounter
some diculties. First, we have to choose x form

_
k=0
D(A
k
) and we do 44
not know how big this space is. Even if we do this, it will be dicult to
prove the convergence of the series, let alone its convergence to T
t
x. So
we proceed to dene the exponential in another way. As a preparation
to the denition of the exponential function of an additive operator - not
necessarily linear - we consider the exponential of a linear operator.
Proposition. Let B be a linear operator from the Banach space X into
X. Then for each x X, s lim
n

_
k=0
B
k
k!
x exists ; denote this by exp Bx.
Then exp B is a linear operator and || exp B || exp(|| B ||).
Proof. We have || B
k
|| (|| B ||
k
) (k 0).
_
k=0
B
k
k!
x is a Cauchy sequence;
for l > j we have
_
_
_
_
l

k=0
B
k
k!

j

k=0
B
k
k!
_
_
_
_
=
_
_
_
_
1

k=j+1
B
k
k!
_
_
_
_
1

j+1
B
k!
k
and <
_
()
k=0
|| B ||
k
k!
|| x || is convergent. So, by the completeness of the
space, s lim
n

_
k=0
B
k
k!
x exists; and the convergence is uniform in every
sphere || x || M; the above inequality shows that
|| exp B || exp(|| B ||) || x || .

1. The exponential of a linear operator 55

So exp B is a linear operator and
|| exp B || exp(|| B ||).
Remark . In a similar manner one can prove the following: Let a se-
quence of linear operators {S
n
}, on a linear normed space
X with values in a Banach space Y be a Cauchy sequence, i.e., 45
lim
n,m
||S
n
S
m
|| = 0. Then there exists a linear operator S forms X
to Y such that lim
n
||S
n
S || = 0 and S lim
n
S
n
.
Theorem. Let B and C be two linear operators from a Banach space X
into X. Assume that B and C commute, i.e., BC = CB. Then
1) exp B. expC = E exp (B + C)
2) D
t
exp(tB)x = s lim
h
exp(t + h)B exp tB
h
x exists and has the value
B(exp tBx) = (exp tB).Bx.
Proof. i) If and are complex numbers, we have

j=0
(t)
j
j!

l=0
(t)
l
l!
=
t( + )
m
m!
(t > 0);
for, by the absolute convergence of each of the series on the left
and the commutativity of and we may arrange the product on
the left to be equal to the power series on the right. A similar proof
holds when and are replaced by commuting linear operators B
and C on a Banach space.
ii) Since tB and hB commute, we have by 1)
exp(t + h)B = exp(tB). exp(hB) = exp(hB). exp tB.
So,
exp(t + h)B exp tB
h
=
exp tB(exp(hB) I)
h
=
exp(hB) I
h
exp tB.
56 7. Lecture 7
iii) follows since
_
_
_
_
exp(hB) I
h
B
_
_
_
_
=
_
_
_
_

k=2
(hB)
k
k!
_
_
_
_

k=2
B
k
k!
h
k1
0, as h 0.

2 Representation of semi-groups
46
Theorem. Let A be the innitesimal generator of a semi-group {T
t
}.
Then for each y X
T
t
y = s lim
n
exp(tAJ
n
)y
uniformly in any bounded interval of t. (J
n
is the resolvent (I
n
1
A)
1
, n > ).
Proof. (tAJ
n
) = nt(J
n
I) is a linear operator and so exp(tAJ
n
) can be
dened. Since ntI and ntJ
n
commute we have
(exp tAJ
n
) = exp(ntI). exp(ntJ
n
)
= exp(nt). exp(ntJ
n
).

Since ||J
n
|| 1/(1 n
1
) (n > ), we have
|| exp(tAJ
n
)|| exp(nt)|| exp(ntJ
n
)||
exp(nt) exp(ntJ
n
||)
exp(nt) exp(nt/1 n
1
)
= exp(tB/(1 n
1
))
If x D(A), D
t
T
t
x = AT
t
x = T
t
Ax and hence
D
s
_
exp[(t s)AJ
n
)]T
s
x
_
= exp((t s)AJ
n
)T
s
Ax exp((t s)AJ
n
).AJ
n
.T
s
x.
2. Representation of semi-groups 57
Since T
t
T
s
= T
s
T
t
(= T
t+s
),
J
n
= n
_

o
e
nt
T
t
dt
is the limit of Riemannian sums each of which commutes with each T
s
;
so J
n
commutes with each T
s
so that AJ
n
= n(J
n
I) commutes with
each T
s
. Now
T
t
x exp(tAJ
n
)x = [exp((t s)AJ
n
)T
s
x]
t
s=o
Since exp((t s)AJ
n
)T
s
(A AJ
n
)x is strongly continuous in s, we 47
have, for x D(A),
T
t
x exp(tAJ
n
)x =
_
t
o
D
s
_
exp((t s)AJ
n
)T
s
x
_
ds
=
_
t
o
exp((t s)AJ
n
)T
s
(A J
n
Ax) ds
(as AJ
n
x = J
n
Ax, asx D(A))
So
||T
t
x exp(tAs
n
x)||
_
t
o
|| ||ds

_
t
o
|| exp(t s)AJ
n
||||T
s
||||Ax J
n
Ax||ds
||Ax J
n
Ax||
_
t
o
exp
(t s)
1 n
1
exp sds
For each xed t
o
> 0 and n > , the integral is uniformly bounded
for 0 t t
o
as n(> ) ; also we know that for each x X,
s lim
n
J
n
x = x. Thus
T
t
s = s lim
n
exp(tAJ
n
x) uniformly
if xD(A)
in 0 t t
o
,
We now prove the formula for arbitrary y X. Since D(A) is dense
in X, given > 0 we can nd x D(A) such that ||y x|| . Then
58 7. Lecture 7
||T
t
y exp(tAJ
n
y)|| ||T
t
y T
t
x|| + ||T
t
x exp(tAJ
n
x)||
+ || exp(tAJ
n
)x exp(tAJ
n
)y||
exp(t) + ||T
t
x exp(tAJ
n
)x||
+ exp
_
t
1 n
1

_
.
Since x D(A), the middle term on the right tends to zero as n 48
uniformly in any bounded interval of t. So
lim
n
||T
t
y exp(tAJ
n
)y|| 2 exp(t),
and being arbitrary,
T
t
y = s lim
n
(exp tAJ
n
)y, y X,
uniformly in any bounded interval of t
Remark. The above representation of the semi-group was obtained in-
dependently of E. Hille who gave many representations in his book.
One of them reads as follows:
T
t
x = s lim
n
_
I
tA
n
_
1
x
uniformly in any bounded interval of t. It also shows the exponential
character of the representation.
Lecture 8
1 An application of the representation theorem
In C[o, ] consider (T
t
x)(s) = x(t + s). By the representation theorem 49
(T
t
x)(s) = x(t + s) = s lim
n
exp (tAJ
n
x)(s)
= s lim
n

m=o
t
m
m!
(AJ
n
)
m
x(s)
uniformly in any bounded interval. From this we get an operation the-
oretical proof of the Weirstrass approximation theorem. Let z(s) be a
continuous function on the closed interval [0, ], 0 < < . Let
x(s) C[o, ] be such that x(s) = z(s) for s [0, ] (such functions
trivially exist). Put s = 0 in the above formula
(T
t
x)(0) = x(t) = s lim
n

m=o
t
m
[(AJ
n
)
m
x
] (0)
m!
uniformly in [0, ]. Thus shown that z(s) is the uniform limit of polyno-
mials on [0, ].
2 Characterization of the innitesimal general of a
semi-group
We next wish to characterize the innitesimal generator of a semi-group
by some of the properties we have established. First we prove the
59
60 8. Lecture 8
Proposition. Let A be an additive operator on a Banach space X into
itself with the following properties:
(a) D(A) is dense in X;
(b) there exists a 0 such that for n > the inverse J
n
= (I n
1
A)
1
exists as a linear operator satisfying
||J
n
|| (1 n
1
)
1
(n > ).
Then we have 50
i) AJ
n
x = n(J
n
I)x, x X
ii) AJ
n
x = J
n
Ax = n(J
n
I)x, x D(A)
iii) s lim
n
J
n
x = x, for x X.
Proof. i) and ii) are evident. To prove iii) let y D(A).
Then y = J
n
y n
1
J
n
Ay and hence
||y J
n
y|| n
1
||J
n
||||Ay||
n
1
(1 n
1
)
1
||Ay|| 0 as n .
Let x X. Since D(A) is dense in X, given > 0, there exists
y D(A) such that ||y x|| . We then have
||x J
n
x|| ||x y|| + ||y J
n
y|| + ||J
n
y J
n
x||
+ ||y J
n
y|| + (1 n
1
)
1
.
As ||y J
n
y|| as n ,
lim
n
||x J
n
x|| ,
and being arbitrary positive number, iii) is proved.
2. Characterization of the... 61
Theorem. An additive operator A with domain D(A) dense in a Banach
space X and with values in X is the innitesimal generator of a uniquely
determined semi-group {T
t
} with ||T
t
|| e
t
if (and only if ), for n >
, the inverse J
n
= (I n
1
A)
1
exists as a linear operator satisfying
||J
n
|| (1 n
1
)
1
.
Proof. We put T
(n)
t
= (exp tAJ
n
). We have
T
(n)
t
exp(nt) exp(nt||J
n
||)
exp

t
1 n
1

,
D
t
T
(n)
t
x = AJ
n
T
(n)
t
x = T
(n)
t
AJ
n
x, x X,
and T
(n)
t
x x =
_
t
o
T
(n)
s
AJ
n
x ds.
It is easy to see J
n
J
m
= J
m
J
n
; so AJ
n
= n(J
n
I) commutes with 51
T
(m)
t
= exp(tAJ
m
). Thus, as in the proof of the representation theorem,
we have, for any x D(A),
||T
(m)
t
x T
(n)
t
x|| = ||
_
t
o
D
s
_
T
(n)
ts
T
(m)
s
x
_
ds||
=
_
t
o
T
(n)
ts
T
(m)
s
(AJ
m
AJ
n
)x ds|| (asD
s
T
(m)
s
x = T
(m)
s
AJ
m
x)
||(J
m
A J
n
A)x||
_
t
o
exp
(t s)
1 n
1

. exp
s
1 m
1

ds
So lim
m,n
||T
(m)
t
x T
(n)
t
x|| = 0 uniformly in any nite interval of t.
Let y X. Given > 0, there exists x D(A) such that ||y x|| .
Then
T
(m)
t
y T
(n)
t
y = ||T
(m)
t
y T
(n)
t
x|| + T
(m)
t
x T
(n)
t
x
+ ||T
(n)
t
x T
(n)
t
y
exp
_
t
1 m
1

_
+ T
(m)
t
x T
(n)
t
x + exp
t
1 n
1

.
62 8. Lecture 8
So lim
m,n
||T
(m)
t
y T
(n)
t
y|| . 2 exp(t) uniformly in any nite inter-
val of t. Therefore, by the completeness of X, s lim
n
T
(n)
t
y = T
t
y exist
and the convergence is uniform in any bounded interval of t.
By the resonance theorem T
t
is a linear operator; since T
(n)
t
are
strongly continuous in t and the convergence is uniform in any bounded
interval of t, T
t
is strongly continuous in t. Also,
||T
t
|| lim
n
||T
(n)
t
|| (Cor. to response theorem )
exp(t)
We now prove that T
t
T
s
= T
t+s
(T = I, evidently). 52
Since T
(n)
t
T
(n)
t
= T
(m)
t+s
,
||T
t+s
x T
t
T
s
x|| ||T
t+s
x T
(n)
t+s
x|| + ||T
(n)
t+s
x T
(n)
t
T
(n)
s
x||
+ T
(n)
t
T
(n)
s
x T
(n)
t
T
s
x|| + ||T
(n)
t
T
s
x T
t
T
s
x||
||T
t+s
x T
(n)
t+s
|| + exp
t
1 n
1

T
(n)
s
T
s
x
+ ||T
(n)
t
(T
s
x) T
t
(T
s
x)||
0 as n .
Finally let A

be the innitesimal generator of the semi-group T

t
We shall show that A

= A. For this it is enough to prove that A

is
an extension of A (i.e., x D(A) implies x D(A

) and A

x = Ax).
For, (I n
1
A

)(n > ) maps D(A

) onto X in a one-one manner; by

assumption (I n
1
A) maps D(A

) onto X in a one-one manner; but on

D(A), (I n
1
A) = (I n
1
A

) and hence D(A) = D(A

). To prove that
A

is an extension of A, we start with the formula

T
(n)
t
x x =
_
t
o
T
(n)
s
AJ
n
xds, x X.
If x D(A)
||T
s
Ax T
(n)
s
AJ
n
x|| ||T
s
Ax T
(n)
s
Ax|| + ||T
(n)
s
Ax T
(n)
s
AJ
n
x||
||(T
s
T
(n)
s
)Ax|| + exp
s
1n
1
||Ax J
n
Ax||
2. Characterization of the... 63
(AJ
n
x = J
n
Ax, if x D(A)).
As n the rst on the right tends to zero, uniformly in any
bounded interval of s; the second term also tends to zero, uniformly in
any bounded interval of s, as exp
s
1 n
1
stays in such an interval and 53
we know that
s lim
n
J
n
y = y, y X.
Hence
T
t
x x = s lim
n
(T
(n)
t
x x) = s lim
n
_
t
o
T
(n)
s
AJ
n
x ds
=
_
t
o
s lim
n
(T
(n)
s
AJ
n
x)ds
=
_
t
o
T
s
Axds
(using the uniformly of convergence in [o, t]). Therefore
s lim
n
T
t
x x
t
= T
o
Ax = Ax.
i.e., if x D(A) then x D(A

) and A

x = Ax.
The uniqueness of the semi-group {T
t
} with A as the innitesimal
generator follows from the representation theorem for semi-groups pro-
ved earlier.
Lecture 9
1 Group of operators
We add certain remarks which will be useful for the application of semi- 54
group theory to Cauchys problem. The rst of these relates to condi-
tions under which a semi-group becomes a group; this will be useful in
connection with the wave equation.
Denition. A one parameter family T
t <t<
of linear operators T
t
of
a Banach space X is called a group of linear operators of normal type
(or simply a group) if the following conditions are satises:
i) T
t
T
s
= T
t+s
, T
o
= I (group property)
ii) s lim
nt
0
T
t
x = T
t
o
x for each x X and t
o
(, )
iii) there exists a 0 such that for all t
||T
t
|| e
|t|
.
(The innitesimal generator of a group is dened by: Ax = lim
to
T
t
x x
t
).
Theorem. Let A be an additive operator from a Banach space X into X
such that D(A) is dense in X. A necessary and sucient condition that
A be the innitesimal generator of a group T
t
is that there exists a 0
such that for every n with |n| > the inverse J
n
= (I N
1
A)
1
exists as
linear operator with ||J
n
|| /(1 |n|
1
).
65
66 9. Lecture 9
Proof. Necessity. Let {T
t
} be a group. Consider the two semi-groups
{T
t
}
to
,
_

T
t
_
to
where

T
t
= T
t
. The innitesimal generator of the semi-
group {T
t
}
t0
coincides with the innitesimal generator A of the group;
let A

be the innitesimal generator of

_

T
t
_

If we show that A

= A the proof of the necessity part will be 55

complete. Let x D(A

). Then
s lim
n0

T
h
I
h
x = A

x.
Putting x
n
= h
1
(

T
h
I)x, we have
||T
h
x
h
A

x|| ||T
h
x
h
T
h
A

x|| + ||T
h
A

x A

x||
||T
h
||||x
h
A

x|| + ||T
h
A

x A

x||.
(exp h)||x
h
A

x|| + ||T
h
A

x A

x||
0 as h 0.
Thus for x D(A

)
Ax = s lim
n0
h
1
(I T
h
) = s lim
n0
T
h
x
h
= A

x.
Hence x D(A

) implies x D(A) and A

x = Ax. Similarly it is
proved that if x D(A), then x D(A

) and A

x = Ax. So A

= A.
suciency: We can construct two semi -groups {T
t
}
to
and
_

T
t
_
to
as
follows:
T
t
x = s lim
n
T
(n)
t
x = s lim
n
exp (tAJ
n
)x
= s lim
n
exp(nt[(I n
1
A)
1
I]x)

T
t
x = s lim
n
exp(t AJ
n
)x = s lim
n
exp (nt[(I + n
1
A)
1
I]x)
If we show that

T
t
T
t
= T
t

T
t
= I, then

T
t
=
_

_
T
t
for t 0

T
t
for t 0
( < t < )
1. Group of operators 67
will be a group with A as the innitesimal generator.
Since J
n
= (I n
1
A)
1
commutes with J
n
= (I +n
1
A)
1
we have 56
(I n
1
A)
1
+ (I + n
1
A)
1
= [(I + n
1
A) + (I n
1
A)](I n
1
A)
1
(I + n
1
A)
1
= 2(I n
1
A)
1
(I + n
1
A)
1
= 2(I n
2
A
2
)
1
.
Since J
k
maps X onto the dense subspace D(A) of X, J
n
J
n
= (I
n
1
A
2
)
1
maps X onto a dense subspace D(A
2
). Moreover
||(I n
2
A)
1
|| ||J
n
||||J
n
|| (1 /
n
)
1
_
1

n
_
1
= (1
2
/n
2
)
1
.
Therefore A
2
is the innitesimal generator of a semi-group exp(tA
2
).
exp(tA
2
)x = s lim
m
exp(tA
2
(I m
1
A
2
)
1
)x
= s lim
m
exp(m
2
t[(I m
1
A
2
)
1
I])x
the convergence being uniform in t in any nite interval of t.
We have
||T
t

T
t
x T
(n)
t

T
(n)
t
x|| ||T
t

T
t
x T
(n)
t

T
t
x|| + ||T
(n)
t

T
t
x T
(n)
t

T
(n)
t
x||
||
_
T
t
T
(n)
t
_

T
t
X|| + exp
_
t
1 n
1

_
||

T
t
x

T
(n)
t
x||
0 as n ,
uniformly in t in any bounded interval of t.
That the rst on the right tends to zero uniformly in t in any bounded
interval of t may be proved as follows: Let 0 t t
o
< . (t
o
> 0). For
any > 0, we can nd t
1
, . . . , t
k
, 0 t
1
, . . . , t
k
t
o
such that
inf
1ik
||

T
k
x T
t
i
x|| ,
(by the strong continuity of T
t
in t). 57
68 9. Lecture 9
Now
||(T
t
T
(n)
t
)

T
t
i
x|| 0 (i = 1, 2, . . . , k)
uniformly in t for 0 t t
o
, and hence, choosing t
i
properly for given
t, we have
||(T
t
t
(n)
t
)

T
t
X|| ||(T
t
T
(n)
t

T
t
i
x|| + ||(T
t
T
(n)
t
)(

T)t

T
t
i
)x||
||(T
t
T
(n)
t
)

T
t
i
x|| +
_
exp t + exp
t
1 n
1

_
.
So the right side tends to zero uniformly in 0 t t
o
.
Since
T
(n)
t

T
(n)
n
x = exp
_
nt
_
(I n
1
A)
1
+ (I + n
1
A)
1
2I
__
x
= exp
_
st
n
.n
2
_
(I n
2
A
2
)
1
I
_
_
x,
we have
T
t

T
t
x = s lim
n
exp(
2t
n
.n
2
[(I n
2
A
2
)
1
I])x)
the convergence being uniform in any bounded interval of t. Thus
T
t

T
t
x = exp(0.A
2
x) = x.
Similarly

T
t
T
t
x = x.
Remark. For an alternative proof of the above theorem, see E. Hille:
Une g en eralisation du probl` em de Cauchy, Ann. de 1 Institut Fourier,
4 (1952), p.37 (Th eor` eme 4).
Lecture 10
1 Supplementary results
We shall now prove some results which supplement our earlier results; 58
these will be useful in applications.
Theorem. 1. For a semi-group {T
t
} the innitesimal generator A may
be dened by
w lim
ho
T
h
I
h
x.
i.e., if

A is the operator with D(

A) =
_
x|w lim
ho
T
h
I
h
x exists
_
and

Ax = w lim
ho
T
h
I
h
x, then

A = A.
2. If {T
t
}
to
is a family of linear operators on a Banach space X such
that T
t+s
= T
t
T
s
, T
o
= I and ||T
t
|| e
t
, 0 then the following two
conditions are equivalent:
(i) strong continuity of T
t
, i.e., w lim
tt
o
T
t
x = T
t
o
x for each t
o
0
and x X.
(ii) weak right continuity at t = 0, i.e., w lim
ho
T
t
x = x, for x X.
3. The innitesimal generator is a semi-group is a closed operator.
PROOF. It is evident that

A is an extension of A. We shall show that A
is an extension of

A, i.e., if x D(

A), then x D(A) and Ax =

Ax. If
69
70 10. Lecture 10
x D(A),
w lim
ho
T
t+h
T
t
h
x = T
t
_
w lim
ho
T
h
I
h
x
_
= T
t
Ax.
(For, if w lim
ho
x
h
= y, and T is a linear operator, then w
lim
ho
T x
h
= T
y
; in fact, if f X

,

f (y) = f (TY) is a linear functional on X, 59
as |

f (y)| || f || ||T
y
|| || f || ||T|| ||y||, and f (Ty) f (T x
h
) =

f y

f x
h
0
as h 0). So, if x D(

A), f (T
t
x) has right derivative
d
+
dt
f (T
t
x) =
f (T
t

Ax) (t 0), which is continuous for t 0, by the strong continuity
of T
t
. Therefore the derivative
d
dt
f (T
t
x) exists for each t 0 and is
continuous.
So
f (T
t
x x) = f (T
t
x) f (x) =
_
t
o
f (T
s

Ax)ds
= f
__
t
o
T
s

A x ds
_
, for each f X

.
Continuously, by the Hahn-Banach theorem,
T
t
x x =
_
t
o
T
s

Ax ds.
Since T
t
is strongly continuous in t it follows that
s lim
to
T
t
I
t
x = T
o

Ax =

Ax.
Thus if x D(

A), then x D(A) and

Ax = Ax.
PROOF. Evidently (i) implies (ii). T
o
prove that (ii) implies (i), let x
o
be a xed element of X. We shall show that w lim
tt
o
T
t
x
o
= T
t
o
x
o
for
each t 0. Consider the function x(t) = T
t
x
o
. For t
o
0, x(t) is right
continuous at t
o
, as wlim
tt
o
T
t
x
o
= wlim
ho
T
h
T
t
o
x
o
. x(t) has the following
three properties:
1. Supplementary results 71
(a) x(t) is weakly measurable, i.e., for any f X

, f (x(t)) is measurable
(since a right continuous numerical function is measurable).
(b) ||x(t)|| is bounded in any bounded interval of t.
(c) there exists a countable set M = {x
n
} such that x(t) (t 0) is con- 60
tained in the closure of M.
To prove (c), let {t
k
} be the totality of positive rational numbers.
Consider nite linear combinations
_

k
x(t
k
) where
k
are rational num-
bers if X is real and if X is complex
k
= a
k
+ib
k
with a
k
and b
k
rational.
These elements form a countable set M = {x
n
}. The closure of M,

M,
contains x(t), for each t 0.
For, if not, let t
0
0 be a number such that x(t
o
) does not belong
to

M.

M is a closed linear subspace of X. By the Hahn-Banach theorem,
there exists a linear functional f
o
on X such that f
o
(x(t
o
)) 0 and f
o
(x) =
0 for x

M. Take a sequence t

k
t
o
( t

k
positive rational). By the weak
right continuity of x(t) at t
o
,
f
o
(x(t

k
)) f
o
(x(t
o
)).
But f
o
(x(t

k
)) = 0 and f
o
(x(t
o
)) 0. We have thus arrived at a
contradiction.
We next prove a result, due to N. Dunford (On one parameter
group of linear transformations, Ann, of Math., 39(1938), 569 573),
according of which the properties (a), (b) and (c) listed above imply
the strong continuity of x(t). First we show that ||x(t)|| is measurable
in t. Let f
n
X

be such that f
n
(x
n
) = ||x
n
|| and || f
n
|| = 1. Let
f (t) = sup
n1
f
n
(x(t)); since each f
n
(x(t)) is measurable, f (t) is measur-
able in t. But ||x(t)|| = f (t); for
f (t) | f
n
(x(t))| | f
n
(x
n
)| | f
n
(x(t) x
n
)|
||x
n
|| ||x(t) x
n
||
and x(t) is in the closure of the set M so that f (t) ||x(t)||; since 61
| f
n
(x(t)| ||x(t)||, f (t) ||(t)||. Thus f (t) = ||x(t)|| and ||x(t)|| is mea-
surable.
72 10. Lecture 10
By a similar argument, ||x(t) x
n
|| is measurable in t for each n. it
follows, using (c), that the half-line [0 t < ) can be represented, for
each integer m, as a countable union of measurable sets S
m,n
,
[0, ) =

_
n=1
S
m,n
, S
m,n
=
_
t|||x(t) x
n
|| m
1
_
If we dene
S

m,1
= S
m,1
, . . . , S

m,n
= S
m,n

n1
_
k=1
S

m,k
,
we have a decomposition of [0, ) into disjoint measurable sets S

m,n
(n =
1, 2, . . .) such that ||x(t) x
n
|| m
1
in S

m,n
.
Therefore the strongly measurable step-function (i.e., a countably
valued function taking each of its values exactly on a measurable set)
x
m
(t) = x
n
for t S

m,n
converges to x(t) as m uniformly in [0, t), Thus x(t) is a strongly
measurable function, a strongly measurable function being a functional
which is the uniform limit of a sequence of strongly measurable step
functions. We may then dene the Bochner integral of x(t) by:

x(t)dt = s lim
,

x
(m)
(t)dt, 0 < <
(

x
m
(t)dt may be dened, as in the case of the ordinary Lebesgue inte-
gral, as the strong limit of nitely valued functions, each taking each of
its values exactly on a measurable set). We have
||

x(t)dt||

||x(t)||dt.
Let 0 < < < < ( > 0). 62
1. Supplementary results 73
Since
x() = T

x
o
= T

x
o
= T

x( ),
we have
( )x() =

x()d =

( )d,
the integrals being Bochner integrals. So
( ){x( ) x()} =
_

{x( ) x( )}d.
Thus
| | ||x( ) x()|| sup

||T

||

||x( ) x()||d
But the right side tends to zero as 0. (This we see by approximat-
ing x(), in bounded interval, uniformly with bounded. nitely valued
strongly measurable functions. For, then the result is reduced to the case
of numerical measurable step functions.) Thus x() is strongly continu-
ous for > 0.
To prove the strong continuity at = 0 we proceed as follows: For
positive rational t
k
, since
T

x(t
k
) = T

t
t
k
x
o
= T
+t
k
x
o
= x( + t
k
),
we have, using the continuity for > 0 proved above,
s lim
0
T

x(t
k
) = x(t
k
).
It follows that s lim
0
T

x
n
= x
n
for each x
n
; also x(t), t 0, in
particular x(0) = x
o
, belongs to

M (M = {x
n
}). It follows therefore,
from the inequalities,
||x() x
o
|| ||T

x
n
x
n
|| + ||x
n
x
o
|| + ||T

(x
o
x
n
)||
||T

x
n
x
n
|| + ||x
n
x
o
|| + sup
o1
||T

||.||x
o
x
n
||,
that lim
o
x() = x
o
i.e., T

is strongly continuous at = 0. 63
74 10. Lecture 10
PROOF . An additive operator A (with domain D(A)) is said to be
closed if it possesses the following property: if {x
n
} is a sequence of
elements of D(A) such that s lim
n
x
n
= x and s lim
n
Ax
n
= y, then x
belongs to D(A) and Ax = y. Evidently a linear operator is closed.
To prove (3) let k > . Then J
k
=
_
I
A
k
_
1
is a linear operator. Let
{x
n
} be a sequence, x
n
D(A) such that s lim
n
x
n
= x, s lim
n
Ax
n
=
y. Then s lim
n
_
x
n

A
k
x
n
_
= x
y
k
. By the continuity of J
k
, s
lim
n
J
k
_
x
n

A
k
x
n
_
= J
k
_
x
y
k
_
, i.e., x = J
k
_
x
y
k
_
. So x( D(A).
Since
_
I
A
k
_
x =
_
I
A
k
_
J
k
_
x
y
k
_
= x
y
k
,
we have Ax = y.
Remark. It is to be noted that the theory has been extended for {T
t
}
o<t
satisfying
T
t
T
s
= T
t+s
and the strong continuity in t for t > 0.
Lecture 11
1 Temporally homogeneous Marko process on a
locally compact topological space
Let R be a locally compact topological space, countable at innity. We 64
consider in R

a probabilistic movement. Suppose that for each triple
(t, x, E) consisting of a real number t > 0, a point x R and Borel set
E R there is given a real number P(t, x, E) such that the following
conditions are satised.
i) P(t, x, E) 0, P(t, x, R) = 1
ii) for xed t and x, P(t, x, E) is a countably additive set function on
the Borel sets
iii) for xed t and E, P(t, x, E) is a Borel measurable function in x
iv) P(t + s, x, E) =
_
R
P(t, x, dy) P(s, y, E) t, s > 0. (Chapman - Kol-
mogoro relation).
The function P(t, x, E) is called the transition probability; this gives
the probability that, in this process, a point x R is transferred
to the Borel set E after t units of time. We say then that there is
given a temporally homogeneous Marko process on R (temporal
homogeneity means that the motion does not depend on the initial
time but only on the time elapsed).
75
76 11. Lecture 11
2 Brownian motion on a homogeneous Rieman-
nian space
Next, we wish to dene the spatial homogeneity of the process.
We assume that R is an n-dimensional, orientable connected C

65
Riemannian space such that the (full) group of isometries G of R,
which is a Lie group, is transitive on R (i.e., for each pair x, y R
there exists an isometry S

such that S

x = y. The process P(t, x, E)

is called spatially homogeneous if
v) P(t, x, E) = P(t, S

x, S

E) for each S

G, x R, E R. A tem-
porally and spatially homogeneous Marko process on R is called
a Brownian motion on R, if the following condition, known as the
continuity condition of Lindeberg, is satised.
vi) lim
to
t
1
_
dis(x,y)>
P(t, x, dy) = 0, for every > 0 and x R.
Proposition. Let C[R] denote the Banach space of bounded uniformly
continuous real valued functions f (x) on R, with the norm
|| f || = sup
xR
| f (x)|.
Dene
(T
t
f )(x) =
_

_
_
R
P(t, x, dy) f (y), if t > 0.
f (x), if t = 0.
Then T
t
denes a semi -group of normal type in C[R].
Proof. We have by condition (i),
|T
t
f (x)| sup
yR
| f (y)|.

2. Brownian motion on a homogeneous Riemannian space 77

If we dene a linear operator S by (S f )(x) = f (S

x), S

G, we
have T
t
S = S T
t
. For,
(S T
t
f )(x) = (T
t
f )(S

x)
=
_
P(t, S

x, dy) f (y)
=
_
P(t, S

x, d(S

y)) f (S

y)
=
_
P(t, x, dy) f (S

y) = (T
t
S f )(x).
If S

G be such that S

x = x
1
, we have 66
(T
t
f )(x) (T
t
f )(x

) = (T
t
f )(x) (S T
t
f )(x)
= T
t
( f S f )(x).
By the uniform of continuity of f (x) and the above equality, we
see that (T
t
f )(x) is uniformly continuous and bounded. The semi-group
property follows easily from Fubinis theorem and the Chapman-Kolmo-
gor relation (T
o
= I by denition).
To prove the strong continuity, it is enough by and earlier theorem,
to verify weak right continuity at t = 0. Since the conjugate space of
C[R] is the space of measures of nite total variation, it is enough to
show that lim
to
(T
t
f (x) ) = f (x) boundedly in x.
Now
|(T
t
f )(x) f (x)| =

_
R
P(t, x, dy)[ f (y) f (x)]

by(i)
=

R
_
d(x,y)
P(t, x, dy)[ f (y) f (x)]

_
dis(x,y)>
P(t, x, dy)[ f (y) f (x)]

| + 2|| f ||
_
dis(x,y)>
P(t, x, dy)
1.
78 11. Lecture 11
The rst term on the right tends to zero as 0 and, for xed ,
the second term tends to zero boundedly in x as t 0 (by (vi), and the
spatial homogeneity). Thus lim
to
(T
t
f )(x) = f (x) boundedly in x.
Theorem. Let x
o
be a xed point of R. Let us assume that the isotropy
group G
o
=
_
S

|S

G, S

x
o
= x
o
_
is compact. (G
o
, being a closed sub-
group of Lie group, is a Lie group). Let A be the innitesimal generator 67
of T
t
. Then
(i) if f D(A) C
2
(C
2
denoting the set of twice continuously dif-
ferentiable functions), then, for a coordinate system (x
1
x
n
) at
x
o
,
(Af )(x
o
) = a
i
(x
o
)
f
x
i
o
+ b
i j
(x
o
)

2
f
x
i
o
x
j
o
(adapting the summation convention), where
a
i
(x
o
) = lim
to
t
1
_
dis(x
o
,x)
(x
i
x
i
o
)P(t, x
o
, dx)
b
i j
(x
o
) = lim
to
t
1
_
dis(x
o
,x)
(x
i
x
i
o
)(x
j
x
j
o
)P(t, x
o
, dx)
the limits existing independently of suciently small > 0.
(ii) The set D(A) C
2
is big in the sense that, for any C
2
function
with compact support there exists f (x) D(A) C
2
such that
f (x
o
),
f
x
i
o
,

2
f
x
i
o
x
j
o
are arbitrarily near respectively
g(x
o
),
g
x
i
o
,

2
g
x
i
o
x
j
o
.
Proof.
Step 1. Let g(x) be a C

function with compact support.

If f D(A), the convolution
( f g)(x) =
_
G
f (S

y
x)g(S

y
x)dy,
2. Brownian motion on a homogeneous Riemannian space 79
(S

y
denotes a generic element of G and dy a xed right invariant Haar
measure on G) is C

and belongs to D(A). (The integral exists since

the isotropy group is compact and g has compact support). By the uni-
form continuity of f and the compactness of the support of g we can
approximate the integral by Riemann sums
k
_
i=1
f (S

y
i
x)C
i
uniformly in
x : ( f g)(x) = s lim
n
k
_
i=1
f (S

y
i
x)C
i
.
Since T
t
S = S T
t
, S commutes with A, i.e., if f D(A), then S f 68
D(A) and AS f = S Af . Putting h(x) = (Af )(x), (h C[R]).
A
_

_
m

i=1
f (S

y
i
x)C
i
_

_
=
m

i=1
(AS
y
i
f )(x)C
i
=
m

i=1
(S
y
i
Af )(x)C
i
=
m

i=1
h(S

y
i
x)C
i
and the right hand side tends to (h g)(x) = (Af g)(x). Since A is
closed, it follows that f g D(A), and A( f g) = Af g. Since R is
a homogeneous space of the Lie group G (by the closed subgroup G
o
)
we can nd a coordinate neighbourhood U of x
o
and for each x U an
element S

(x) G such that i) S

x = x
o
ii) S

(x)x
o
depends analytically
on the coordinate functions x
1
x
n
. by the right invariance of the Haar
measure,
( f g) (x) =
_
G
f (S

y
S

(x)x
o
)g(S

y
S

(x)x
o
)dy
=
_
G
f (S

y
x
o
)g(S

y
S

(x)x
o
)dy, x U.
The function on the right side is C

in a neighbourhood of x
o
and

q
1
++q
n
(x
1
)
q
1
(x
n
)
q
n
f g(x) =
_
G
f (S

y
x
o
)

q
1
++q
n
g(S

y
S

(x)x
o
(x
1
)
q
1
(x
n
)
q
n
dy

Lecture 12
1 Brownian motion on a homogeneous Riemannian
space (Contd.)
69
Proof.
Step 2. Remarking that D(A) is dense in C[R] and choosing f and g
properly we obtain
(a) there exist C

functions F
1
(x), . . . , F
n
(x) D(A) such that the Ja-
cobian
(F
1
(x), . . . , F
n
(x))
(x
1
, . . . , x
n
)
> 0 at x
o
.
(b) there exists a C

function F
o
(x) D(A) such that
(x
i
x
i
o
)(x
j
x
j
o
)

2
F
x
i
o
x
j
o

i=1
(x
i
x
i
o
)
2
.
We can use F
1
(x), . . . , F
n
(x) as coordinate functions in a neighbour-
hood d(x
o
, x) < ; we denote these newlocal coordinates by (x
1
, . . . , x
n
).
Since F
i
(x) D(A),
s lim
to
T
t
F
i
(x) F
i
(x)
t
exists and = AF
i
(x)
(AF
i
)(x) = lim
to
t
1
_
R
P(t, x
o
, dx)(F
i
(x) F
i
(x
o
))
81
82 12. Lecture 12
= lim
to
t
1
_
d(x,x
o
)
P(t, x, dx)(F
i
(x) F(x
o
))
independent of > 0, by Lindebergs condition. So, for the coordinate
functions x
1
x
n
, (x
i
= F
i
),
lim
to
t
1
_
d(x,x
o
)
(x
i
x
o
)P(t, x
o
, dx) = a
i
(x
o
)
independent of > 0. Since F
o
D(A), we have, using Lindebergs
condition,
(AF
o
)(x
o
) = lim
to
t
1
_
R
P(t, x
o
, dx)(F(x) F
o
(x
o
))
= lim
to
_
d(x,x
o
)
P(t, x
o
, dx)(F(x) F
o
(x
o
))
= lim
to
_

_
t
1
_
d(x,x
o
)
(x
i
x
i
o
)
F
o
x
i
o
P(t, x
o
, dx)
+ t
1
_
d(xx
o
)
(x
i
x
i
o
)(x
j
x
j
o
)
_

2
F
o
x
i
x
j
_
P(t, x
o
, dx)
x = x
o
+ (x x
o
0 < 1.
The rst term on the right has a limit a
i
(x
o
)
F
o
x
i
o
; hence by the posi- 70
tivity of P, and (b),
lim
to
t
1
_
d(x,x
o
)
n

i=1
(x
i
x
i
o
)
2
P(t, x
o
, dx) < (*)
Step 3. Let f D(A) C
2
. Then, expanding f (x) f (x
o
),
T
t
f (x
o
) f (x
o
)
t
= t
1
_
R
f (x) f (x
o
)P(t, x
o
, dx)
1. Brownian motion on a homogeneous... 83
= t
1
_
d(x,x
o
)>
f (x) f (x
o
))P(t, x
o
, dx)
+ t
1
_
d(x,x
o
)
(x
i
x
i
o
)
f
x
o
i
P(t, x
o
, dx)
+ t
1
_
d(x,x
o
)
(x
i
x
i
o
)(x
j
x
j
o
)

2
f
x
i
o
x
j
o
P(t, x
o
, dx)
+ t
1
_
d(x,x
o
)
(x
i
x
i
o
)(x
j
x
j
o
)C
i
j
()P(t, x
o
, dx)
= C
1
(t, ) + C
2
(t, ) + C
3
(t, ) + C
4
(t, ), say ,
where C
i
j
() 0 as 0. We know that lim
to
C
1
(t, ) = 0 for xed 71
> 0 (Condition (vi)) and lim
to
C
2
(t, ) = a
i
(x
o
)
f
x
i
o
, independently of
small . By () and Schwarzs inequality lim
to
C
4
(t, ) = 0, boundedly in
t > 0. Also the left side has a nite limit as t 0. So the dierence
lim
to
C
3
(t, )
lim
t o
C
3
(t, )
can be made arbitrarily small by taking > 0 small. But by (), Sch-
warzs inequality and (vi), the dierence is independent of small > 0.
Thus nite limit lim
to
C
3
(t, ) exists independently of small > 0. Since
we may choose F D(A) C

such that

2
F
x
i

x
i

(i, j = 1, . . . , n)
is arbitrarily near
i j

i j
being constants, it follows, by an argument
similar to the one above that
nite limit
_
d(x,x
o
)
(x
i
x
i
o
) (x
j
x
j
o
)P(t, x
o
, dx) = b
i j
(x
o
)
exists and lim
to
C
3
(t, ) = b
i j
(x
o
)

2
F
x
i
o
x
j
o
.
84 12. Lecture 12
This completes the proof of the theorem.
Remark. i) We have b
i j
(x) = b
i j
(x) and
b
i j
(x
o
)
i

j
0, (
i
real), for ,
(x
i
x
i
o
)(x
j
x
j
o
)
i

j
=
_

(x
i
x
i
o
)
i
_
2
ii) b
i j
(x) is a contravariant tensor: 72

b
i j
= b
kl
x
i
x
k
.
x
j
x
1
(x
1
, . . . , x
n
) ( x
1
, . . . , x
n
)
and a
m
= a
s
x
m
x
s
+ b
kl

2
x
m
x
k
x
l
.
This follows from the equality

b
i j

2
f
x
i
o
x
j
o
+ a
m
f
x
m
= b
k1

2
f
x
k

1
+ a
s
f
x
s
[since each is = (Af )(x
o
)].
Part III
Regularity properties of
solutions of linear elliptic
dierential equations
85
Lecture 13
The results proved in this part will be needed in the application of the 73
semi-group theory to Cauchys problem.
1 Strong dierentiability
Let R be a subdomain of E
m
. We denote by C

(R) the space of in-

denitely dierentiable functions in R and by D

(R) the space of C

functions in R with compact support. We denote by L

2
(R)
loc
the space
of locally square summable functions in R, (i.e., functions in R which
are square summable on every compact subset of R). A function u(x)
L
2
(R)
loc
is said to be k-times strongly dierentiable in R (or of order k
in R) if for every subdomain R
1
of R relatively compact in R there exists
a sequence u
n
(x)(= u
n,R
1
(x) ) of C

functions in R
1
, such that
lim
n
_
R
1
|u u
n
|
2
dx = 0
and lim
n,1
_
R
1
|D
(s)
u
n
D
(s)
u
1
|
2
dx = 0 for |s| k.
Then there exists, for |s| k, functions
u
(s)
(x) = u
(s)
R
1
L
2
(R
1
) such that
lim
n
_
R
1

u
(s)
(x) D
(s)
u
n
(x)

2
dx = 0.
87
88 13. Lecture 13
u
(s)
R
1
(x) is determined independently of the approximating sequence
u
n
; for we have, for each C

function with compact support in R

1
74
_
R
1
(x)u
(s)
(x)dx = lim
n
_
R
1
(x)D
(s)
u
n
(x)dx
= lim
n
(1)
|s|
_
R
1
u
n
(x)D
(s)
(x)dx
= (1)
|s|
_
R
1
u(x)D
(s)
(x)dx
and C

functions with compact support in R

1
are dense in L
2
(R
1
). It
also follows that, for |s| k, there exists a function in L
2
(R)
loc
, de-
noted by

D
(s)
u(x), such that for each subdomain R
1
relatively compact
in R,

D
(s)
u(x) coincides with u
(s)
R
1
(x) almost everywhere in R
1
.

D
(s)
u(x)
is called the strong derivative of u corresponding to the derivation D
(s)
.
2 Weak solutions of linear dierential operators
Let
L =
n

||=||=o
D
()
a

D
()
, a
,
(x) C

(R), a
,
= a
,
for || = || = n,
be a linear dierential operator in R with C

coecients. Let f
L
2
(R)
loc
. A function u L
2
(R)
loc
will be said to be a weak solution
of the equation Lu = f if for every D

(R) we have
_
R
L

udx =
_
R
f dx
where L

is the adjoint of L:
L

=
n

||=||=o
(1)
||+||
D
()
a
,
D
()
.
3. Elliptic operators 89
3 Elliptic operators
Friedrichs - Lax - Nirenberg, theorem: Let L be elliptic in R in the 75
sense that there exists a constant C
o
> 0 such that

||=||=n

1
1

m
m
a

m
;
1

m
(x)

1
1

m
m
C
o
_

_
m

i=1

2
i
_

_
n
for every x R and every real vector (
1
, . . . ,
m
). Then if u
o
is a weak
solution of Lu = f and if f is of order p in R, then u
o
is of order 2n + p
in R.
Sobolevs lemma: If u
o
(x) is of order k in R, then, for k > m/2 +
, h
o
(x) is equal almost everywhere (in R) to a function which is times
continuously dierentiable.
Weyl-Schwartz theorem: Let L be an elliptic operator in R, and u
o
a
weak solution of Lu = f . If f is indenitely dierentiable in R, then u
o
is almost everywhere equal to an indenitely dierentiable function in
R.
This theorem is an immediate consequence of the Friedrichs Lax-
Nirenberg theorem and Sobolevs lemma.
4 Fourier Transforms:
For the proofs we need the following facts about Fourier transforms:
Plancherels theorem: Let f (x) L
2
(E
m
), x = (x
1
, . . . , x
n
). Then the
functions

n
(y) =
_
|x|n
f (x) exp(2ix.y) dx (x.y =

x
i
y
i
)
converge in the L
2
-norm to a function (y
1
, . . . , y
n
) L
2
and the trans- 76
formation F dened by F f = (y) = lim
n
_
|x|n
f (x) exp(2ix.y)dx is
a unitary transformation of L
2
onto itself. (i.e., (F f , Fg) = ( f , g), for
f , g L
2
onto L
2
). The inverse F
1
of F is given by
F
1
(x) = lim
n
_
|y|n
(y) exp(2iy.x)dy
90 13. Lecture 13
F( f ) is called the Fourier transform of f .
As regards the Fourier transform of the derivatives, we have: if f
in L
2
(E
m
) is also in C
k
(E
m
) and D
(s)
f (x) L
2
(E
m
) for |s| k, (D
(s)
=

s
1
++s
n
/x
s
1
1
x
s
m
m
, |s| =
_
n
i=1
s
j
), then
(FD
(s)
f )(y) =
m
_
j=1
(2iy
j
)
s
j.F( f )(y).
Proof of Sobolevs lemma: Let R
1
be any relatively compact subdo-
main of R and (x) aC

function with compact support in R such that

(x) 1 on R
1
. Since u
o
is assumed to be of order k there exists a
sequence {u
n
} of C

functions in R
1
such that
lim
n

|s|k
_
R
1

D
(s)
u
o
D
(s)
u
n

2
dx = 0.
We have, using Leibnitzs formula,
lim
n

|s|k
_

D
(s)
u
o
D
(s)
u
n

2
dx = 0.
Let u
n
(resp. u
o
) denote the function in E
m
dened by:
u
(x)
n
=
_

_
u
n
(x), x Support of
0 x E
m
supp ;
similar denition for u
o
(= u
o
in supp. ). Since the Fourier transform 77
is a unitary transformation, we have
lim
n
||FD
(s)
u
n
F

D
(s)
u
o
||
o,E
m = 0.
But, as remarked earlier,
(FD
(s)
u
n
)(y) = (2i)
s
y
s
1
1
y
s
m
m

U
n
(y)
where

U
n
= F u
n
; also since F is unitary,
lim
n
||

U
n


U
o
||
o,E
m
= 0, where

U
o
= F( u
o
).
4. Fourier Transforms: 91
Therefore there exists a subsequence {n

} of {n} such that for almost

all y E
m
lim
n

U
n
(y) =

U
o
(y) (pointwise limit)
lim
n

U
n
(y)y
s
1
1
y
s
m
m
(2i)
|s|
=

U
|s|
o
=

U
(s)
o
(y)
where

U
(s)
o
= F

D
(s)
u
o
.
Thus for almost all y E
m
,

U
o
(y)y
s
1
1
y
s
m
m
(2i)
|s|
=

U
(s)
o
(y), |s| k.
We shall now show that

U
o
(y) y
q
1
1
y
q
m
m
is integrable on E
m
pro-
vided k >
m
2
+ , where = |q|
m
_
j=1
q
j
. We have

U
o
(y)y
q
1
1
y
q
m
m
=
y
q
1
1
y
q
m
m
1 + |
_
m
i=1
y
2
i
|
k/2

U
o
(y)
_

_
1 + |
m

i=1
y
2
i
|
k/2
_

_
.
Now, in polar coordinates
dy = dy
1
dy
m
= r
m1
drd
m1
(
m1
is the surface of unit sphere in E
m
). So
y
q
1
1
y
q
m
m
1 + |
m
_
i=1
y
2
i
|
k/2
is square
integrable in |z| > (Z E
m
) if 2|q| 2k + m 1 < 1, i.e., if k >
m
2
+. Already we know that U
o
(y)(1+
m
_
i=1
y
2
i
)
k/2
is square integrable in 78
|z| > . So U
o
(y)y
q
m
1
y
qm
m
, begin the product of two square integrable
functions, is integrable in |z| > . We see also that U
o
(y)y
q
1
1
y
q
m
m
is
integrable in |z| .
Thus if k >
m
2
+ |q|, U
o
(y)y
q
1
1
y
q
m
m
is integrable over E
m
.
Suppose k >
m
2
+ , ( > 0 integer). Then

U
o
(y) L
2
L
1
so that (F
1

U
o
)(y) =
_
E
m

U
o
(y) exp(2iy.x)dy, a.e on E
m
; i.e., u
o
(x) =
_
E
m

U
o
(y) exp(2iy.x)dya.e. on E
m
.
92 13. Lecture 13
Let |q| (k >
m
2
+ ); then
D
(q)
x
_

U
o
(y) exp(2i.y.x)
_
=

U
o
(y)
m
_
j=1
(2iy
j
)
q
j
exp 2iy.x
and

U
o
(y)
m
_
j=1
(2iy
j
)
q
j
exp 2iyx

U
o
(y)
m
_
j=1
(2iy
j
)
q
j

and

U
o
(y)
m

j=1
(2iy
j
)
q
j

is a function independent of x and summable (as

a function of y) over E
m
. Therefore D
q
(x) u
o
(x) exists and D
(q)
u
o
(x) =
_
E
m

U
o
(y)
m

j=1
(2iy
j
)
q
j
(exp 2iy.x)dy.
This representation also shows that D
(q)
u
o
(x) is continuous. Thus
u
o
(x) is -times continuously dierentiable; so u
o
(x) is -times contin-
uously dierentiable in R
1
.
Lecture 14
1 Gardings inequality
For the proof of the Friedrichs - Lax - Nirenberg theorem, we need 79
Gardings inequality Let R
1
be a relatively compact subdomain of R
and let L be a linear elliptic dierential operator in R. There exist > 0
and > 0 such that for D

(R
1
),
( + (1)
n
L

, ) ||||
2
n
where ||||
2
n
=
_
R
1

|s|n
|D
(s)
|
2
dx.
Before proving the theorem, we prove a preliminary
Proposition. (i) Dene for D

(R
1
)
||||||
2
j
=

|s|=j
_
R
1

D
(s)

2
dx.
Then for j < n there exists a positive constant e
j,n
such that
||||||
j
e
j,n
||||||
n
(ii) lim
o
sup
D

(R
1
)
_

_
||||
2
n1
||||
2
o
+ ||||
2
n
_

_
93
94 14. Lecture 14
(iii) There exists positive constants and

such that for D

(R
1
)

||=||=n
(D

a
,
D

, ) ||||||
2
n

||||
n1
||||
n
Proof. (i) Let 80
(x) =
_

_
(x) x R
1
0 x E
m
R
1
Then
(x) = (x
1
, . . . , x
m
) =
xs
_

(x
1
, . . . , x
s1
, t, x
s+1
, . . . , x
m
)
t
dt
Hence by Schwarzs inequality
| (x)|
2
L

x
s

2
dx
s
, where L is the diameter of R
1
. So
_
R
1
||
2
dx =
_
R
1

|
2
dx
L.
_
R
1
dx
1
dx
m
_

x
s

2
dx
s
_

_
= L
2
_
R
1

x
s

2
dx
Therefore
||||
2
o
L
2
||

x
s
||
2
o
.
By repeated application of this inequality we get (i).
(ii) Since
FD
(s)
(y) =
m
_
j=1
(2iy
j
)
s
j
(y), ( = F

)
1. Gardings inequality 95
and F is a unitary transformation in L
2
, we obtain
||||||
2
l
=

|s|=1
_
E
m
|FD
(s)
|
2
dx
= (2)
2l

|s|=1
_
E
m
m
_
j=1
y
2s
j
j
|(y)|
2
dy.
Since

_
|s|n1

m
j=1
y
2s
j
j
1 +
_
|t|n

m
j=1
y
2t
j
j
tends to zero uniformly in y as 0. 81
(iii) is proved.
(iv) When a
n
1
,...,n
m
; n

1
,...,n

m
(x) with

n
i
=

i
= n
are constant we have by partial integration and Fourier transform

||=||=n
D
()
a
,
D
()
, ) =

||=||=n
(1)
n
a
,
(D

, D

)
=
_
E
m

||=n
||=n
(2)
2n
y

1
1
y

m
m
a

m
,
1

m
y

1
1
y

m
m
Const
_
E
m

|s|=n
|y
s
1
1
y
s
m
m
|
2
|F(y)|
2
dy|F|
2
dy
(making use of the ellipticity)
= Const .
_
E
m

|s|=n
|D
(s)
|
2
dx
Const ||||||
2
n
.
If a
,
(x), (|| = || = n) are non-constant, put
= sup
,;x

,x

R
1
|a
,
(x

) a
,
(x

)|.
96 14. Lecture 14

Note that may be taken to be arbitrarily small if we choose R

1
suciently small. Let x
o
be a xed point of R
1
. Put a
,
(x
o
) = a
,
o
.
Let D

(R
1
). We have

||=||=n
(1)
n
(a
,
D

, D

)
=

||=||=n
(1)
n
(a
,
o
D

, D

)
0
+

(1)
n
[((a
,
a
,
0
)D

, D

)
o
];
|

(1)
n
(a
,
a
,
o
)D

, D

)
o

||=||=n
||D

||
o
||D

||
0
Const ||||||
2
n
.
So

||=||=n
(1)
n
(a
,
D

, D

) C
1
||||||
2
n
Const ||||||
2
n
C
3
||||||
2
n
(C
3
> 0).
if we choose, R
1
suciently small. This result enables us to deduce 82
(iii) for the general case. For any > 0, R
1
can be covered by a nite
number, say N, of open spheres S
1
, S
2
, . . . , S
N
of radius /2. Let S

i
be
the sphere of radius concentric with S
i
. Let
i
(x) C

(E
m
) satisfy

i
(x) > 0 for x S
i
,
i
(x) = 0 for x S

i
and
i
(x) 0 for x E
m
.
Then
h
i
(x) = (
i
(x)/
N

j=1

j
(x) )
1
2
satises h
i
(x) C

(R
1
), h
i
(x) 0and
N

j=1
h
j
(x) 1 or R.
Thus
(1)
n

||=||=n
(a
,
D

, D

)
o
1. Gardings inequality 97
=
N

j=1
A
j
=
N

j=1
(1)
n

||=||=n
(a
,
h
j
D

, h
j
D
()
)
o
is such that
A
j
= (1)
n
_

||=||=n
(a
,
D

h
j
, D
()
h
j
)
o
R
j
_

_
where, by Leibnitzs formula,
R
j
=

| or |

|<n
(c

, D

)
o
with bounded functions C

. Thus, by Schwarzs inequality,

|R
j
| a
j
||||
n1
||||
n
(a
j
= constant > 0).
For suciently small > 0, we have, by the result obtained already,
(1)
n

||=||=n
(a
,
D

, D

)
o

j=1
(
j
|||h
j
|||
2
n
a
j
||||
n1
||||
n
) (
j
= Const > 0)
Moreover, we have, by the same reasoning as above,
||| h
j
|||
2
n

_
R
h
2
j
(x)

| |=n
| D
()
(x) |
2
dx b
j
|| ||
n1
|| ||
n
with constant b
j
> 0. Therefore, by putting 83
= min(
j
),
N

j=1
(
j
b
j
+ a
j
) =

,
we have
(1)
n

| |=||=n
(a
,
D
()
, D
()
)
o
||| |||
2
n

|| ||
n1
|| ||
n
98 14. Lecture 14
Proof of Gardings inequality: We have, by integration by parts and
from part (iii) of the above proposition, for > 0.
( + (1)
n
L

, )
o
(, )
o
+ ( ||| |||
2
n

|| ||
n1
|| ||
n
)
+

||<n||n
(C
,
D
()
, D
()
)
o
where C
,
are bounded C

functions in R
1
. Then by (i) and Schwarzs
inequality
( + (1)
n
L

, ) ||||
2
o
+ {||||||
2
n
||||
n1
||||
n
}
with some positive constant . Hence for any > 0 we have, remember-
ing
||| ||
2
n
= || ||
2
n

s<n
|| ||
2
s
and using (i),
( + (1)
n
L

, ) || ||
2
0
+
_
|| ||
2
n

|| ||
2
n1

2
_
|| ||
2
n1
+ || ||
2
n

1
_
_
Then by taking
1
> 0 so small that ( /2
1
) > 0 and > 0
suciently small we obtain Gardings inequality by (ii).
Lecture 15
1 Proof of the Friedrichs - Lax - Nirenberg theorem
To prove the Friedrichs - Lax - Nirenberg theorem, we need three lem- 84
mas:
Lemma 1. If u
o
is of order i in R
1
and if

D(s)
u
o
is of order j in R, for
all s with |s| i, then u
o
is of order i + j in R
1
. If u
o
is of order i + j in
R, then

D(s)
u
o
is of order j for |s| i.
Lemma 2. Let R
1
be a relatively compact subdomain of R and let u
o

L
2
(R
1
). Then for any positive integer s
(I + ()
s
)h = u
o
( is the Laplacian)
has weak solution of order 2s in R
1
.
Lemma 3. Let u
o
L
2
(R
1
) be of order n in R
1
and
|(L

, u
o
)| Const ||||
n1
, for all D

(R
1
)
_

_
(, )
o
=
_
R
1

d x; || ||
2
k
=

|s|k
_
R
1
|D
(s)
|
2
dx
_

_
Then u
o
is of order n + 1 in R
1
.
Assuming these lemmas for a moment, we shall give a Proof of the
Friedrichs - Lax - Nirenberg theorem
99
100 15. Lecture 15
First Step 1. If u
o
L
2
(R
1
) is of order n in R
1
and satises |(L

, u
o
)|
Const ||||
nj
for all D

(R
1
), then u
o
is of order n + j in R
1
. This is
proved by induction on j. The result is true for j = 1 (Lemma 3). Let us
assume that j > 1 and that the result is true for j 1 Suppose
|(L

, u
o
)| Const ||||
nj
;
since ||||
nj
||||
n( j1)
, u
o
is of order (n+ j 1) in R
1
by the inductive 85
assumption. For any rst order derivation D, we have |(L

D, u
o
)| Const
||D||
nj
Const ||||
nj+1
. Since u
o
is of order n + 1, we have
(L

D, u
o
) =

||,||n
_
(1)
||+||
D
()
a
,
D
()
D , u
o
_
=

_
(1)
||
D D

, a
,

u
o
_
=
_
(1)
||+1
D

, D(a
,
,

D

u
o
)
_
=

_
(1)
||+1
D

, (D a
,
)

D

u
o
_
+
_
(1)
||+1
D

, a
,

D u
o
_
=

_
(1)
||+1
D

, (D a
,
)

D

u
o
_
(L

,

D u
o
).
Since u
o
is of order (n + j 1) we see by partial integration that
|(L

,

D u
o
)| |(L

D, u
o
)| + Const ||||
2n
(n + j 1)
Const ||||
n( j1)
By Lemma 1,

D u
o
is of order n+ j 1 n+ j 2 n (as j 2).
Hence by the induction assumption

Du
o
is of order n + j 1. So, by
lemma 1 u
o
is of order n + j.
Second Step 1 (Friedrichs theorem ). Let u
o
L
2
(R
1
) be a weak solu-
tion of Lu = f and f be order p in R
1
. If u
o
is of order n in R
1
, then u
o
is of order 2n + p in R
1
.
Proof. This holds for p = 0. For, from (L

, u
o
)
o
= (, f )
o
, we have
|(L

, u
o
)
o
| Const ||||
o
= Const ||||
nn
.
1. Proof of the Friedrichs - Lax - Nirenberg theorem 101
So, by the rst step u
o
is of order n + n = 2n. Suppose p = 1. We have, 86
as above,
(L

,

D u
o
)
o
= (D L

, u
o
) = (1)
||+||+1
(D

D a
,
D

, u
o
)
o
= (1)
||+||+1
(D
()
a
,
D

D , u
o
)
+ (1)
||+||+1
(D

(D a
,
)D

, u
o
)
o
= (L

D, u
o
)
o
+ (,

L

u
o
),
where L

is a dierential operator of degree 2n.

(L

,

Du
o
) = (D, f )
o
+ (,

L

u
o
)
= (,

Df ) + (,

L

u
o
)
(since f is of order 1 at least; the case p = 0 is already proved). Thus
|(L

, D u
o
)
o
| Const ||||
o
= Const ||||
nn
and

D u
o
is of order 2n 1 n. So by the rst step,

D u
o
is of order
n + n = 2n. By Lemma 1, u
o
is of order 2n + 1. For p > 1, we may
repeat the argument.
Third Step 1. Let u
o
L
2
(R
1
) be a weak solution of L u = f and f be
of order p in R
1
. Then u
o
is of order 2n + p in R
1
.
Proof. Let h
o
of order 2n be a weak solution of
(I + (A)
n
)h = u
o
.
h
o
exists by Lemma 2. Then h
o
of order 2n is a weak solution of
L(I + ()
n
)h = f ;
L(I + ()
n
) is an elliptic operator of order 4n. f being of order p, h
o
is 87
of order 4n + p, by the second step. Hence, by Lemma 1,
u
o
= (I + ()
n
)h
o
is of order 4n + p 2n = 2n + p.
Lecture
1 Proof of Lemma 3
Let R be a bounded domain of E
m
. Let u
o
of order n satisfy 88
|(L

, u
0
)
0
| |
n

||=||=0
(D
()
a
,
D
()
, u
0
)
0
Const ||||
n1
for all D

(R)
Let R
2
R
1
R, R
2
, R
1
being subdomains, such that the closure of
R
1
in R is compact. Let D

with (x) = 1 on R
2
. Let
v
h
(x) =
v(x
h
) v(x)
h
, x
h
= (x
1
+ h, x
2
, . . . , x
m
),
h suciently small. Then, as will be proved below,
|| v
h
||
n
Const ( for all suciently small h).
Since the Hilbert space H
n
(R) (completion of D

(R) by the norm

|| ||
n
) is locally weakly compact, there exists a sequence {h
i
} with lim
i
h
i
= 0 such that for |k| n
weak lim
i
v
h
i
= v
weak lim
i

D
k
v
h
i
= v
(k)
exist in L
2
(R
1
). We shall show that
v =

D
1
v (D
1
= /

x
1
)
103
104 16. Lecture
v
(k)
=

D
1

D
(k)
v
=

D
(k)

D
1
v
proving that

D
1
v is of order n in R
1
. Similarly

D
i
v(i = 2, . . . , m) will
be of order n in R
1
. Thus by lemma 1, v is of order n+1 in R
1
and hence
u is of order n + 1 in R. That v =

D
1
v may be proved as follows: For
any D

(R
1
) we have, being a real number such that 0 < < 1,
(, v)
o
= lim
i
(, v
h
i
)
o
= lim
i
(
h
i
, v)
= lim
i
(
x
1
(x

h
i
), v(x))
o
= lim
i
((x
(h
i
)
),

D
1
v(x))
o
= (, D
1
v)
o
.
We have also 89
(

D
k
v)
h
=

D
k
v
h
and thus, in L
2
,
v
k
= weak lim
i

D
k
v
h
i
= w lim
i
(

D
k
v)
h
i
=

D
1
D
(k)
v.
We prove that
||v
h
||
m
Const (for all small h).
We shall make use of Gardings inequality for the 2n order elliptic
dierential operator L

: there exist constants C

1
, C
2
and C
3
such that
C
1
||||
2
n
(L

, )
o
+ C
2
||||
2
o
|(L

, )| C
3
||||
n
||||
n
, , D

(R).
Now,
(L

, v
h
)
o
= (1)
||
(D

, a
,

D
()
( u
o
)
h
)
o
= (1)
||
(D

, a
,
(

D
()
u
o
)
h
)
o
1. Proof of Lemma 3 105
= (1)
||
(D
()
, a
,
(.

D

u
o
)
o
(*)
+ (1)
||
C
,

(D

, a
,
_
D

D
(

)
u
o
_
h
o
(|

| 1)
by applying the Leibnitz formula.
On the other hand, we have, for any function w of order j in R with 90
support completely interior to R
||w
h
||
J1,R
1
||w||
j
, for suciently small |h|, because, for any ap-
proximating functions {u
i
} C

(R)
|| w
h
||
j1,R
1
= lim
i
|| u
h
i
||
j1,R
1
= lim
i
|| u
x
1
(x
(h)
) ||
j1,R
1
|| w ||
j,R
= || w ||
j
.
Thus the absolute value of the second term on the right of () is by
Schwarzs inequality Const || ||
n
|| u ||
n
= Const || ||
n
. Since
(e f )
h
(x) = e
h
(x) f (x
h
) e(x) f
h
(x),
we have
(1)
||
(D

, a
,
(.

D
()
u
o
)
h
)
o
=(1)
||
(D

, [(a
,
.

D

u
o
)
h
(a
,
)
h
(x
h
).

D

u
o
(x
h
)])
o
=(1)
||
((D

)
h
, a
,

u
o
)
o
+(1)
||+1
(D

, (a
,
)
h
(x
h
)

D

u
o
(x
h
))
o
The absolute value of the second term on the right is 91
Const || ||
m
.
We have also
(1)
||
((D

)
h
, a
,

u
o
)
o
=(1)
||
(D

h
, a
,

u
o
)
o
=(1)
||
(a
,
D

h
,

D

u
o
)
o
106 16. Lecture
=(1)
||
(a
,
D

h
,

D

u
o
)
o
=(1)
||
C
,

(a
,
(D

D
(

h
),

D

u
o
)
o
(|

| 1).
The absolute value of the second term on the right is
Const ||
h
||
n1
Const || ||
n1
.
Therefore, by applying the original hypothesis,
|(L

, ( u
o
)
h
)
o
| |(L

h
, u
o
)| + Const ||||
n
Const ||
h
||
n1
+ Const ||||
n
K || ||
n
, K a positive constant .
Thus letting tend, in || ||
n
, to ( u
o
)
h
, we have
C
1
||( u
o
)
h
||
2
n
K||( u
o
)
h
||
n
+ C
2
||( u
o
)
h
||
o
Since ||( u
o
)
h
||
o
Const || u
o
||
1
, the right hand side being indepen-
dent of h, we must have
||( u
o
)
h
||
n
Const ( independent of h).
Lecture 17
1 Proof of Lemma 2
We dene 92
u
o
(x) =
_

_
u
o
(x) if x R
1
0 if x E
m
R
1
.
Let U
o
(y) = (F u
o
)(y). Then
h
o
(x) = F
1
U
o
(Y)
1 + (
_
m
j=1
(2y
j
)
2
)
s
(x)
satises the conditions of the lemma. In the rst place,
h
o
(x) =
_
|y|n
U
o
(y)
1 + (
_
m
j=1
(2y
j
)
2
)
s
exp (2

1 yx)dy
is C

(E
m
). For, since U
o
(y) L
2
(E
m
),
U
o
(y)
1 + (
_
m
j=1
(2y
j
)
2
)
s
m
_
j=1
(2

1 y
j
)
k
j
exp (2

1 yx)
is, for any set of integers k
j
0, integrable over |y| n and majorised
uniformly in x by a summable function (in y). So

k
1
++k
m
x
1
k
1
x
n
k
n
h
n
(x) =
_
|y|n
_

_
U
o
(y)
m

j=1
(2

1 y
j
)
k
j
1 + (
_
j=1
(2 y
i
)
2
)
s
_

_
exp(2

1 yx)dy
107
108 17. Lecture 17
Moreover, for |k| 2s, the function under the curly brackets { }
is in L
2
(E
m
), so that for |k| 2s, D
(k)
h
n
(E
x
). converges in L
2
(E
m
)
Therefore h
o
(x) is of order 2s in E
m
.
Next for any D

(E
m
), we have, by partial integration,
_
E
m
(I + ()
s
) (x) h
o
(x) dx = lim
n
_
E
m
(I + ()
s
)(x) h
n
(x) dx
= lim
n
_
E
m
(x) (I + ()
s
) h
n
(x) dx
=
_
E
m
(x) (F
1
U
o
) (x) dx
This proves that h
o
is a weak solution in E
m
of (I + ()
s
) h = u
o
= 93
F
1
U
o
. Thus h
o
is a weak solution in R
1
of (I + ()
s
)h = u
o
.
2 Proof of Lemma 1
In the proof of Lemma 1 we have to make use of the notion of regular-
isation or molliers. Let j(x) C

(E
m
) such that
i) j(x) 0,
ii) j(x) = 0, for |x| 1
iii)
_
E
m
j(x) dx = 1.
Let for > 0
j

(x) =
n
j(x/)
We have then
i) j

(x) 0,
ii) j

(x) = 0, for |x|

iii)
_
E
m
j

(x) dx = 1.
2. Proof of Lemma 1 109
Let R
1
be a relatively compact subdomain of R E
m
and u(x)
L
2
(R
1
). Let R
2
be a subdomain relatively compact in R
1
. Let d > 0 be
the distance between R
2
and the boundary of R
1
. Let > 0 be such that
< d. For x R
2
, dene
(J

u)(x) =
_
R
1
j

(x y)u

(y) dy.
((J

u)(x) is called regularisation of u(x) and the operators J

are called 94
molliers). Let
||v||
2
o,R
i
=
_
R
i
|v|
2
dx.
We then have
i) || J

u ||
o,R
2
|| u ||
o,R
1
ii) lim
o
|| J

u u ||
o,R
2
= 0
iii) (J

u)(x) is C

in R
2
and if h is of order i in R
1
,
then
D
(s)
(J

u)(x) = (J

D
s
u)(x) for |s| i
in R
2
.
Proof of (iii): We have, for each derivation D
(s)
,
(D
(s)
x
J

u)(x) =
_
R
1
D
(s)
x
j

(x y) u(y) dy.
Suppose u is of order i in R. We have then, for |s| i, by partial
integration,
_
R
1
D
(s)
x
j

(x y) u(y) =
_
R
1
(1)
|s|
{D
(s)
y
j

(x y)} u(y) dy
_
R
1
j

(x y)

D
(s)
u(y) dy
110 17. Lecture 17
since, for each x R
2
, j

(x y) considered as a function of y, has com-

pact support in R
1
.
Proof of (ii): We have, for x R
2
,
_
R
1
j

(x y) dy = 1. Hence
(J

u)(x) u(x) =
_
R
1
j

(x y) (u(y) u(x)) dy.

By Schwarzs inequality
_
R
2
| (J

u)(x) u(x) |
2
dx

_
R
2
dx
_

_
_
R
1
j

(x y) dy
_
R
1
j

(x y)| u(y) u(x)|

2
dy
_

_
=
_
R
2
dx
_
R
1
j

(x y) | u(y) u(x)|
2
dy

_
R
2
dx
_
|z|<
j

(z) |u(x z) u(x)|

2
dz
=
_
|z|<
j

(z)
_

_
_
R
2
| u(x z) u(x)|
2
dx
_

_
dz
Since
_
R
2
| u(x z) u(x)|
2
dx tends to zero as |z| 0, (ii) is proved. 95
Proof of (i): We have, by calculations similar to the above calculations,
|| J

u ||
2
o,R
2
=
_
R
2
dx
_
R
1
j

(x y) | u(y) |
2
dy

_
|z|<
j

(z)
_

_
_
R
2
|u(x z) |
2
dx
_

_
dz

_
|z|<
j

(z)
_

_
_
R
2
| u(x) |
2
dx
_

_
dz
2. Proof of Lemma 1 111
= || u ||
2
o,R
1
.
Proof of Lemma 1. Let u be of order in R
1
and let

D
(s)
u be of order j
in R
1
for each s with |s| i. Then for |t| j,
D
(t)
D
(s)
J

u = D
(t)
J

D
(s)
u = J

D
t

D
(s)
u (|s| i)
by (iii). Hence by (ii), u is of order i + j, in R
1
.
Next let u be of order i + j in R
1
. Since
D
(t)
J

D
(s)
u = J

D
(t)

D
(s)
u (|t| j, |s| i)
we see by (ii) that

D
(s)
u is of order j in R
1
.
Part IV
Application of the
semi-group theory
to the Cauchy problem for
the diusion and wave
equations
113
Lecture 18
1 Cauchy problem for the diusion equation
Let R be a connected n-dimensional oriented Riemannian space with the 96
metric
ds
2
= g
i
j
(x) dx
i
dx
j
.
Let A be a second order linear partial dierential operator in R with
C

coecients:
(Af )(x) = b
i j
(x)

2
f
x
i
x
j
+ a
i
(x)
f
x
i
(x)c(x) f (x);
we assume that b
i j
is a symmetric contravariant tensor and a
i
(x) satises
the transformation rule
a
i
= a
k
x
i
x
k
+ b
kl

2
x
i
x
k
x
l
[(x
1
, . . . , x
n
) ( x
1
, . . . , x
n
)] so that the value (Af )(x) is determined
independent of the choice of the local coordinates. We further assume
that A is elliptic in the strong sense that there exist positive constants
and (0 < < ) such that
g
i j
(x)
i

j
b
i j
(x)
i

j
g
i j
(x)
i

j
for every real vector (
i
, . . . ,
n
) and every x R.
We consider the Cauchy problem in the large on R for the diusion
equation: to nd u(t, x) (x R) such that
_

_
u
t
= A u (t, x), t > 0
u(0, x) = f (x), f (x) being a given function on R.
(**)
115
116 18. Lecture 18
We shall rst give a rough sketch of our method of integration. We 97
wish to integrate the equation in a certain function space L(R) which is a
Banach space (i.e., we want to obtain u(t, x) such that u(t, . . .) L(R) for
each t 0); we assume that L(R) contains D

(R), the space of C

func-
tions with compact support, as a dense subspace. (Examples: L
p
(R), 1
p < ; C(R) if R is compact). We determine an additive operator A
o
such that: (i) C

(R) D(A
o
) D

(R), if f D(A
o
) A
o
f = Af .(ii)
the smallest closed extension

A
o
of A
o
exists and

A
o
is the innitesimal
generator of a semi group T
t
on L(R). We then have
_

_
D
t
T
t
f = s lim
ho
T
t+h
T
t
f
h
=

A
o
T
t
f (= T
t

A
o
f ), t 0
T
o
f = f .
Thus T
t
f is a kind of solution of (). Next, we shall show that,
if the initial function f (x) is prescribed suitably [e.g.,if f D

(R) or
more generally, if A
k
o
f D(A
o
) for all integers k o], there exists
a function u(t, x) denitely dierentiable in t and x such that T
t
f (x) =
u(t, x) almost everywhere in (0, ] R, the measure in R being the one
given by

g d x
1
, . . . , d x
n
, and u(t, x) will be a solution of ().
In carrying out this procedure, we have to solve an equation of the
form
_
u
A
o
m
_
u = f , f is given and u is to be found fromD(A
o
). This is
a kind of boundary value problem connected with the elliptic dierential
operator A.
Theorem. If R is compact, the equation 98
_

_
u
t
= Au = b
i j
(x)

2
u
x
i
x
j
+ a
i
(x)
u
x
i
, t > 0
u(0, x) = f (x) D

(R), ( f (x) given )

admits of a solution C

in (t, x). This solution can be represented in the

form
u(t, x) =
_
R
P(t, x, dy) f (y)
where P(t, x, E) is the transition probability of a Marko process on R.
1. Cauchy problem for the diusion equation 117
The proof will be preceded by two lemmas.
We take for L(R) the Banach space C(R) of continuous functions
with || f || = sup
x
| f (x)|.D

(R) is dense in L(R). The operator A

is dened
as follows:
D(A

) = D

(R) and A

f = Af for f D

(R).
Lemma 1. For any f D

(R) and and any m > 0, we have

max
x
h(x) f (x) min
xR
h(x)
where h(x) = f (x)
(A

f )(x)
m
.
Proof. Let f (x) attain its maximum at x

. We choose a local coordinate

system at x

such that b
i j
(x

) =
i j
(Kronecker delta).
(Such a choice is possible owing to the positive deniteness of
b
i j

j
. Then
h(x

) = f (x

) m
1
(A

f )(x

)
= f (x

) m
1
a
i
(x

)
f
x
i

m
1

n
i=1

2
f
(x
i

)
2
= f (x

).
since we have, at the maximum point x

, 99
f
x
i

= 0 and
n

i=1

2
f
x
i
2

0.
Thus max
x
h(x) f (x). Similarly we have f (x) min
x
h(x).
Corollary. The inverse (Im
1
A

)
1
exists for m > 0 and ||(Im
1
A

)
1
|| 1. Further ((I m
1
A

)
1
h)(x) 0 if h(x) 0. Also
(I m
1
A

)
1
.1 = 1.
Lemma 2. The smallest closed extension

A

of A

exists.
118 18. Lecture 18

f is dened and equal to h if there exists a sequence { f

k
} D

(R)
such that s lim
k
f
k
= f and s lim
k
A

f
k
= h.

f is determined uniquely by f . For if { f

k
} D

(R) be such that

lim
k
f
k
= 0 and lim
k
A

f
k
= h, then we must h = 0.
For by Greens integral theorem, R being compact,
_
R
f
k
A

gdx =
_
R
gAf
k
dx,
for every g D

(R) so that, in the limit,

0 =
_
R
ghdx, for every g D

(R); so h = 0.
To prove that the resolvent (I m
1
A

)
1
exists as a linear operator
in C(R), for m large, it will be sucient to show, in view of the Corollary
to Lemma 1 and the fact that

A

is closed, that the range of (I m

1
A

)
is dense in C(R). We shall show that for any h D

(R) we can nd
f D(R) such that (I m
1
A

) f = h (m large). To this purpose, we

need
2 Gardings inequality
For u, v D

(R), dene 100

(u, v)
0
=
_
R
uvdx (||u||
2

= (u, u)

)
(u, v)
1
= (u, v)
0
+
_
R
g
i j
u
x
i
v
x
j
dx (||u||
2
1
= (u, u)
1
)
Then there exists > 0 and > 0 such that for all suciently large
m > 0,
B

(u, v) =
__
I
A
m
_
u, v
_
2. Gardings inequality 119
satises
|B

(u, v)| ||u||

1
||v||
1
||u||
2
1
B

(u, u) for all u, v D

(R).
This lemma can be proved by partial integration.
Let H

be the Hilbert space of square summable functions in R. We

have D

(R) H

(R). Let A
1
be the operator in H

with domain D

(R)
dened by: A
1
f = A

f , f D

(R). As in Lemma 2, the closure of A

1
in H
0
,

A
1
, exists. We show now that the range of (I
A
1
m
) is dense in
H

, for m large. If (I
A
m
)D

were not dense in H

, there will exists

an element f 0 in H

which will be orthogonal to (I

A
m
)D

. This
mean that f is a weak solution of
_
I
A
m
_
f = 0.
By the Wey1-Schwartz theorem, f may be considered to be in
D

(R). By Gardings inequality, assuming m to be suciently large,

|| f ||
2
1

_
I
A

m
f , f
_
= 0. So f = 0.
Since the range of
_
I
A
1
m
_
is everywhere dense in H

,
_
I

A
1
m
_
1
is
dened everywhere in H

. So for every h D

(R), we can nd f

such that f

is a weak solution of
_
I
A
m
_
f = h.
Again by the Weyl-Schwartz theorem, f will be in D

(R). Thus 101

we see that for large m the resolvent J
m
=
_
I

m
_
1
exists as a linear
operator on L(R) and satises ||J
m
|| 1 (also, (J
m
h)(x) 0 if h(x)
0; J
m
.1 = 1). Consequently, (see Lecture 8)

A

is the innitesimal
generator of a uniquely determined semi-group
T
t
= exp(t

A

) = s limexp(tm(J
m
I)).
We have further
||T
t
|| 1, (T
t
f )(x) 0 if f (x) 0, T
t
.1 = 1.
120 18. Lecture 18
If f D

(R), we have
D
t
T
t
f =

A

T
t
f = T
t

A

f = T
t
A

f
D
2
t
T
t
f =

A

T
t
A

f = T
t
A
2

f
: :
D
k
t
T
t
f = T
t
A
k

f ,
for k 0, since A
k

f D

(R) for integral k 0. By making use

of the strong continuity of T
t
in t we see that (D
2
t
+

A)
k
T
t
f is locally
square summable on the product space (0, ) R. Since (

2
t
2
+ A)
k
is an elliptic operator, it follows the Friedrichs-Lax-Nirenberg theorem
that (T
t
f )(x) is almost everywhere equal to a function u(t, x) indenitely
dierentiable in (t, x) for t 0.
Proof of the latter part of the theorem:
|u(t, x)| = |(T
t
f )x| ||T
t
f || || f ||
Hence u(t, x) is, for xed (t, x) a linear functional of f L(R).
Therefore there exists P(t, x, E) such that
u(t, x) =
_
R
P(t, x, dy) f (y).
The non-negativity of u(t, x) for f (x) 0 implies that P(t, x, E) is 0.
Since T
t
1 = 1, we must have P(t, x, R) = 1.
Lecture 19
1 The Cauchy problem for the wave equation
We consider the Cauchy problem for the wave equation in the m- 102
dimensional Euclidean space E
m
:
_

2
u(t,x)
t
2
= Au(t, x), x E
m
u(0, x) = f (x), u
t
(0, x) g(x), f , g, given ,
where A = a
i j
(x)

2
x
i
x
j
+ b
i
(x)

x
i
+ c(x)
is a second-order elliptic operator. This problem is equivalent to the
matricial equation
_

t
_

_
u(t, x)
v(t, x)
_

_
=
_

_
0 I
A 0
_

_
_

_
u(t, x)
v(t, x)
_

_
(I = identity ).
_

_
u(o, x)
v(o, x)
_

_
=
_

_
f (x)
g(x)
_

_
(1)
We may apply the semi-group theory to integrate (1), by consider-
ing, in a suitable Banach-space the resolvent equation
__
I o
o I
_
n
1
_
o I
A o
__ _
u
v
_
=
_
f
g
_
121
122 19. Lecture 19
for large |n| (n, integral) and obtaining the estimate
||
_
u
v
_
|| (1 + |n
1
|)||
_
f
g
_
||
with a positive independent of u, f and g. As a matter of fact, the es-
timate implies (see Lecture 9) that
_
0 I
A 0
_
is the innitesimal generator
of a group {T
t
}
<t<
and
T
t
_
f (x)
g(x)
_
=
_
u(t, x)
v(t, x)
_
will give a solution of (1) if the initial functions f (x) and g(x) are pre- 103
scribed properly.
We have the
Theorem. Suppose that the coecients a
i j
(x), b
i
(x) and c(x) are C

and
that there exists a positive constant such that
a
i j
(x)
i

2
i
(x E
m
, (
1
, . . . ,
m
) E
m
). Assume further that
= max
_

_
sup
x,i, j
|a
i j
(x)|, sup
x,i, j,k

a
i j
x
k

sup
x;i, j,k,s

2
a
i j
x
k
x
s

, sup
x;i
|b
i
(x)|, sup
x;i,k
b
i
x
k
, sup
x
|c(x)|
_

_
is nite. Then there exists a positive constant such that for suciently
small

, the equation (1) is solvable in the following sense: for any pair
of C

functions
_
f (x), g(x)
_
on E
m
for which A
k
f , A
k
g and their partial
derivatives are square integrable (for each integer k 0) over E
m
, the
equation (1) admits of a C

solution u(t, x) satisfying the energy in-

equality
((u

Au, u)

(u
t
, u
t
)

)
1
2
exp(|t|(( f

Af , f )

(g, g)

)
1
2
1. The Cauchy problem for the wave equation 123
Proof. The proof will be carried out in several steps.
First step: Let H be the space of real-valued C

functions in E
m
for
which
|| f ||
1
=
_

_
_
E
m
| f |
2
dx +

i
_
E
m
| f
x
i
|
2
dx
_

_
1
2
< ,
and let

H
1
(E
m
) be the completion of H with respect to the norm|| ||
1
. The 104
completion of H with respect to || f ||

=
_

_
_
E
m
| f |
2
dx
_

_
1
2
will be denoted
by

H

(E
m
).

H

(E
m
) and

H
1
(E
m
) are Hilbert spaces; actually

H

(E
m
) =
H

(E
m
) = L
2
(E
m
).
One can prove that there exists > 0 and

> 0 such that for

0 < <

there correspond > 0 and

> 0 satisfying

|| f ||
2
1

_

_
( f Af , f )

for f H, Af H

( f A

f , f )

for f H A

f H

.
|( f Af , g)

| (1 + )|| f ||
1
|||g||
1
for f , g H, Af H

.
|( f A

f , g)

| (1 + )|| f ||
1
|||g||
1
for f , g H; A

f H

.
|(af , g)

( f , Ag)

| || f ||
1
||g|||

, for f , g H; Af , Ag H

.
(The proofs of these inequalities will be given in the next lecture).
Thus the bilinear form
B

(u, v) = (u A

u, v)

for u, v H, A

u H

can be extended to a bilinear functional B

(u, v) on H
1
satisfying
_

||u||
2
1
B

(u, u)
|B

(u, v)| (1 + )||u||

1
||v||
1
.
Second step: Let 0 <

. For any f H, the equation u Au = f

admits of a uniquely determined solution u(x) = u
f
(x) H.
124 19. Lecture 19
Proof. The additive functional F(u) = (u, f )

is bounded on H
1
, be-
cause
|F(u)| = |(u, f )

| ||u||

|| f ||

||u||
1
|| f ||

.
So, by Rieszs representation theorem, there exists a uniquely deter- 105
mined v( f )

H
1
such that
(u, f )

= (u, v( f ))
1
.
By the Lax-Milgram theorem (see lecture 4) as applied to the bilin-
ear form B

(u, v) in

H
1
, there corresponds a uniquely determined ele-
ment S v( f ) in

H
1
such that
(u, f )

= (u, v( f ))
1
= B

(u, S v( f )), for u H

1
.
u

= S v( f ) is a weak solution of the equation u Au = f , i.e., for each

u D

(R) we have (u, f )

= (u A

u, S v( f ))

. In fact, let {v
k
} H
be a sequence such that v
k
S v( f ) in

H
1
; then, for
u D

(R), B

(u, S v( f )) = lim
n
B

(u, v
n
)
= lim
n
(u A

u, v
n
)
= (u A

u, S v( f )).
Since f is C

in E
m
and A is elliptic, u

= S v( f ) is almost everywhere
equal to a C

function (Weyl- Schwartz theorem). We thus have a solu-

tion u

H of the equation uAu = f . The uniqueness of the solution

follows from the inequalities given in the rst step.
Third step: If the integer n is such that |n|
1
is suciently small, then
for any pair of functions { f , g} with f , g H and Af H

, the equation
__
I 0
0 I
_
n
1
_
0 I
A 0
__ _
u
v
_
=
_
f
g
_
(2)
or
u n
1
v = f
1. The Cauchy problem for the wave equation 125
v n
1
Au = g
admits of a uniquely determined solution {u, v} u, v H. Moreover, we 106
have
_
B

(u, u) +

(v, v)
_ 1
2
(1 + |n|
1
)(B

( f , f ) +

(g, g)

)
1
2
with a positive constant .
Proof. Let u
1
, v
1
H be such that
u
1
n
2
Au
1
= f v
2
n
2
Av
2
= g.
(See the second step). Then
u = u
1
+ n
1
v
2
v = n
1
Au
1
+ v
2
satises (2).
We have
Au = n(v g) H H

, Av = n(Au Af ) H

.
We may therefore apply the inequalities of the rst step.
Thus by (2),
( f

Af , f )

= (u n
1
v

A(u n
1
v), u n
1
v)

= (u

Au, u)

2n
1
(u, v)

n
1
(Au, v)

n
1
(Av, u)

+ n
2
(v

Av, v)

and

(g, g)

(v n
1
Au, v n
1
Au)

(v, v)

n
1
(v, Au)

n
1
(Au, v)

n
2
(A A)

Hence 107
B

( f , f )

(g, g)
B

(u, u) +

(v, v)

2|n|
1
(u, v)

|n|
1
|(Av, u)

(Au, v)

|
126 19. Lecture 19
B

(u, u) +

(u, v)

2|n
1
| ||u||
1
||v||

|n|
1
||u||
1
||v||

(u, u) +

(v, v)

|n
1
|
_
||u||
2
1
+
1
||v||
2

2
(||u||
2
1
+
1
||v||
2

)
_
( > 0)
Thus, by taking > 0 suciently large and then taking |n| su-
ciently large, we have the desired inequality.
Fourth step: The product space

H
1


H

is a Banach space with the

norm
||(
u
v
)|| = [B

(u, u) +

(v, v)]
1
2
We dene now an operator O in

H
1
H

: the domain of O consists of

the vectors (
u
v
) H
1
H

such that u, v H and A(u n

1
v) H

and
v n
1
Au H and on such elements O(
u
v
) is dened to be
_
0 I
A 0
_ _
u
v
_
.
The third step shows that for suciently large |n|, the range of the
operator
_
I 0
0 I
_
n
1
O coincides with the set pairs (
f
g
) such that f , g
H, Af H

; such vectors (
f
g
) are dense in the Banach space

H
1


H

. It
follows that the smallest closed extension

O of is such that
(I n
1

O), I =
_
I 0
0 I
_
admits, for suciently large |n|, of an inverse I
n
= (I n
1

O)
1
which 108
is linear operator on

H
1
H

satisfying
||I
n
|| (1 + |n
1
|).
So, there exists a uniquely determined group {T
t
}
<t<
with

O as
the innitesimal generator and such that
||T
t
|| exp(t),
strong lim
ho
T
t+h
T
t
h
(
f
g
) =

OT
t
(
f
g
) = T
t

O(
f
g
) if (
f
g
) domain of

O (See
Lecture 9).
Lecture 20
1 Cauchy problem for the wave equation (contin-
ued)
Fifth step: If f and g satisfy the conditions of the theorem, i.e., if 109
A
k
f H, A
k
g H(k = 0, 1, . . .), we have

O
k
_
f
g
_
= O
k
_
f
g
_


H
1
H
o
(k = 0, 1, . . .),
i.e., (
f
g
) is in the domain of every power of

O
k
. So, referring to step 4,
we nd that vectors
_
v(t, x)
v(t, x)
_
= T
t
_
f (x)
g(x)
_
are in the domain of every power of O :

O
k
_
u(t, x)
v(t, x)
_


H
1
H
o
.(k = 0, 1, 2, . . .)
Thus, for integral k 0, u(t, x) is for t xed, a weak solution of the
equation
A
k
u = f
(k)
, with f
k
L
2
(E
m
)
A
k
is an elliptic operator of order 2k and k may be taken arbitrarily
large. We see therefore by the Friidrichs-Lax-Nirenberg theorem and
Sobolevs lemma, that u(t, x) is C

in x (for xed t). And the same

statement holds for v(t, x).
127
128 20. Lecture 20
Since ||T
t
|| exp (|t|) we see that
||u(t, x)||
2
1
+ ||v(t, x)||
2
o
Const . exp(2|t|)
_
|| f ||
2
1
+ ||g||
2
0
_
.
This, combined with the strong continuity of T
t
in t, shows that
u(t, x) and v(t, x) are locally square summable in the product space
( < t < )E
m
. And we have, for the second order strong derivative

2
t
,

2
t
u(t, x) = Au(t, x)
so that (
2
t
+ A)u = 2Au, (
2
t
+ A)
k
u = (2A)
k
u,. 110
Since

2
t
2
+ A is an elliptic operator in ( < t < ) E
m
, we see
that u(t, x) is almost everywhere equal to a function C

in (t, x).
The proof of the rst step is obtained by the
Lemma. Let f , g H and Af H
o
. Then
(Af , g)
o
=
_
a
i j
f
x
i
g
x
j
dx
a
i j
x
j
f
x
i
g d x +
_
b
i
f
x
i
g d x +
_
c f g.
And we can also partially integrate the second and the third terms
on the right, so that the rst order derivatives of
f
x
i
shall be eliminated,
and the integrated terms are nought.
Proof. From Af H
o
and g H we see that a
i j

2
f
x
i
x
j
.g is integrable
over E
m
. Thus, by Fubini theorem,
_
E
m
a
i j

2
f
x
i
x
i
x
j
g d x = lim
+
_

dx
2
dx
m

1
_

1
a
i j

2
f
x
i
x
j
g d x
1

1
_

1
a
i j

2
f
x
i
x
j
gdx
1
=
_
a
i j
f
x
i
g
_
x
1
=
1
x
1
=
1

1
a
i j
f
x
j
g
x
i
dx
1
+

a
i j
x
i
f
x
j
gdx
_

_
+

1
_

i, j1
a
i j

2
f
x
i
x
j
gdx
1
.
1. Cauchy problem for the wave equation (continued) 129
= k
1
(
1
,
1
, x
2
, . . . , x
m
) + k
2
(
1
,
2
, x
2
, . . . , x
m
) + k
3
(
1
,
1
, x
2
, . . . , x
m
)
By Schwarzs inequality, we have

dx
2
. . . dx
m
k
1

dx
2
. . . dx
m

f (
1
, x
2
. . . , x
m
)
2
x
j

dx
2
. . . dx
m
|g(
1
, x
2
.x
m
)
2
)
1
2
+ similar terms pertaining to
1
instead of
1
. 111
Since
_
E
m
g
2
dx =

dx
1
_

|g(x
1
, x
2
, . . . , x
m
)
2
dx
2
, . . . , dx
m
,
_
E
m
|
f
x
j
|
2
dx =
_
dx
1

_

_
||
f
x
j
(x
1
, x
2
, . . . , x
m
)|
2
dx
2
. . . dx
m
,
we see there exists {
(n)
1
} and {
(
1
n)} such that
lim

(n)
1

(n)
1

_
k
1
(
(n)
1
,
(n)
1
, x
2
, . . . , x
m
)dx
2
. . . dx
m
= 0.
On the other hand, since f , g
f
x
j
,
g
x
1
H
o
, we see that
lim

_
k
2
(
1
,
1
, x
2
, . . . , x
m
)dx
2
, . . . , dx
m
=
_
E
m
_
a
i j
f
x
j
g
x
1

i j
x
1
f
x
j
g
_
dx = k
2
is nite, Thus,
_
a
i j

2
f
x
i
x
j
gdx = k
2
+ lim
(n)

(n)
1

_

k
3
(
(n)
1
,
(n)
1
, x
2
, . . . , x)dx
2
. . . dx
m
130 20. Lecture 20
Hence

(n)
1
_

(n)
1

i, j1
a
i j

2
f
x
i
x
j
gdx
1
is integrable over < x
i
< (i = 2, . . . , m).
Hence
k
3
= lim

(n)
1

(n)
1

dx
2
. . . dx
m
k
3
(
(n)
1
,
(n)
1
, x
2
, . . . , x
m
)
= lim

(n)
1

(n)
1

lim

2

_
dx
3
. . . dx
m
_

2
_

2
dx
2

(n)
1
_

(n)
1

i, j1
a
i j

2
f
x
i
x
j
gdx
1
_

_
However 112
_

_
=

(n)
1
_

(n)
1
dx
1

2
_

i, j1
a
i j

2
f
x
i
x
i j
dx
2
=

(n)
1
_

(n)
1
dx
1
_

_
_
a
2j
f
x
j
g
_
x
2
=
2
x
2
=
2
+

2
_

2
a
2j
f
x
j
g
x
2
dx
2

2
_
2
a
2j
x
2
f
x
j
gdx
2
_

_
+

(n)
1
_

(n)
1

2
_

i, j1,2

2
f
x
i
x
j
gdx
1
dx
2
we have

dx
3
. . . dx
m
_

(n)
1

(n)
1
a
2j
f
x
j
gdx
1

1. Cauchy problem for the wave equation (continued) 131

j
_

_
_
dx
1
dx
3
, . . . dx
m

f
x
j

2
_
g
2
dx
1
dx
3
. . . dx
m
_

_
1
2
and so, by the integrability on E
m
of

f
x
j

2
and |g|
2
, there exists
(1)
(2)
,
(1)
(2)
such that
lim

(1)
2

(1)
2

_
dx
3
. . . dx
m

1
_

1
_
a
2j
f
x
j
g
_
x
2
=
1
(1)
x
2
=
2
(1)
dx
1
= 0
uniformly with respect to
1
and
1
.
We have also
lim

(n)
1

(n)
1

lim

_
dx
3
dx
m
_

(n)
1

(n)
1
dx
1
_

2
_
2
_
a
2j
f
x
j
g
x
1

a
2j
x
2
g
x
j
dx
2
_
_

_
=
_
E
m
_
a
2j
f
x
j
g
x
2

a
2j
x
2
g
x
j
g
_
dx.
Therefore 113
_
a
i j

2
f
x
i
x
j
gdx =
_

i or j=1,2
a
i j
f
x
i
g
x
j
dx
_

i or j=1,2
a
i j
x
i
f
x
j
gdx
+ lim

(n)
1

(n)
1

lim

(1)
2

(1)
2

dx
3
dx
m
_

(k)
1

(k)
1
_

(1)
2

(1)
2

i, j1,2
a
i j

2
f
x
i
x
j
gdx
1
dx
2
Repeating the process, we get the Lemma.
Lecture 21
1 Integration of the Fokker-Planck equation
We consider the Fokker-Planck equation 114
u(t, x)
t
= Au(t, x), t 0
(Af )(x) =
1
_
g(x)

2
x
i
x
j
(
_
g(x)a
i j
(x) f (x))
1
_
g(x)

x
i
(g(x)b
i
(x) f (x))
in a relatively compact subdomain R (with a smooth boundary) of an ori-
ented n-dimensional Reimannian space with the metric ds
2
= g
i j
(x)dx
i
dx
j
. As usual the volume element in R is dened by dx =
_
g(x)dx
1
dx
n
. where g(x) = det(g
i j
(x)). We assume that the contravariant ten-
sor a
i j
(x) is such that a
i j

j
> 0 for
m
_
i=1

2
i
> 0,
i
real. The functions
obey, for the coordinate transformation x x, the transformation rule

b
i
( x) =
x
x
k
b
k
+

2
x
x
k
x
s
a
ks
(x).
We assume that g
i j
(x), a
i j
(x) and b
i
(x) are C

function of the local

coordinates x = (x
1
x
m
).
Suggested by the probabilistic interpretation of the Fokker-Planck
equation due to A. Kolmogorov, we shall solve the Cauchy problem in
the space L
1
(R).
133
134 21. Lecture 21
Greens integral formula:
Let A

be the formal adjoint of A;

A

= a
i j
(x)

2
x
i
x
j
+ a
i
(x)

x
i
.
Let G be a subdomain of R with a smooth boundary G. Then we
obtain by partial integration Greens formula:
_
G
{h(x)(Af )(x) f (x)(A

h)(x)}dx
=
_
G
_

_
g(x)a
i j
(x)
x
j

_
g(x)b
i
(x)
_

_
cos(n, x
i
) f (x)h(x)dS
+
_
G
_
g(x)a
i j
(x)
_
h(x)
f
x
j
f (x)
h
x
j
_
cos(n, x
i
)dS
where n denotes the outer normal at the point x of G and ds denotes 115
the hypersurface area of G.
Remark. If a
i j
(x) cos(n, x
i
) cos(n, x
j
) > 0 at x G we may dene the
transversal (or conormal) direction at x by
dx
i
_
g(x)a
i j
(x) cos(n, x
j
)
= d(i = 1, 2, . . . , m)
so that we have
_
g(x)a
i j
(x)(h(x)
f
x
i
f (x)
h
x
j
) cos(n, x
i
)dS
= (h(x)
f
v
f (x)
h
v
)dS .
We consider A to be an additive operator dened on the totality of
D(A) of C

functions f (x) in RUR with compact supports satisfying

the following boundary condition:
_
g(x)a
i j
(x)
f
x
j
cos(n, x
i
) +
_

_
g(x)a
i j
(x)
x
j

_
g(x)a
i
(x)
_

_
.
. cos(n, x
i
) f (x) = 0.
1. Integration of the Fokker-Planck equation 135
(When R is a subdomain of the euclidean space E
m
and A is the
Laplacian the above condition is nothing but the co called reecting
barrier condition)
f
n
= O,
since and n coincide in this case). D(A) is dense in the Banach space
L
1
(R).
To discuss the resolvent of A we begin with 116
Lemma 1. Let f (x) D(A) be positive (or negative) in domain G R
such that f (x) vanishes on G R, (i.e., f (x) vanishes on the part of
G not contained in R). Then we have the inequality
_
G
(Af )(x)dx 0
_

_
resp.
_
G
Af (x)dx 0
_

_
.
Proof. Taking h 1 in Greens formula and remembering the boundary
condition on f (x), we obtain
_
G
(Af )(x)dx =
_
GR
f

ds
0.

Corollary. For f D(A) we have for any > 0 f

1
Af f .
Proof. Let h(x) = 1, 1 or 0 according as f (x) is > 0, < 0 or = 0. Since
the conjugate space of L
1
(R) is the space of essentially bounded function
k(x) with the norm
k

= essential sup
xR
|k(x)|,
we have
f
1
Af
_
R
h(x)
_
f (x)
1
Af (x)
_
dx
136 21. Lecture 21
=
_
| f (x)|dx
1

i
_
P
i
(Af )(x)dx
+
1

i
_
N
j
(Af )(x)dx
where P (resp. N) is connected domain in which f (x) > 0 (resp. < 0)
such that f (x) vanishes on P(resp. N).
Lemma 2. The smallest closed extension

A of A exists and for any > 0
the operator (I
1

A) admits of a bounded inverse, J

= (I
1

A)
1
with norm 1.
Proof. The existence of

A follows from Greens formula. For if { f
k
} 117
D(A) be such that strong lim f
k
= 0, strong lim Af
k
= h, then for
D

(R),
lim
_
R
_
Af
k
f
k
A

_
dx = 0, (or)
_
hdx = 0. So h = 0.
The other part of the lemma follows from the corollary of lemma
1.
Lemma 3.

A is the innitesimal generator of a semi-group T
t
in L
1
(R)
if and only if for suciently large the range {(I
1
A) f , f D(A)} of
the operator (I
1
A) is dense in L
1
(R). Moreover, if this condition is
satised, then J

is a transition operator, i.e., if f (x) 0 and f L

1
(R),
then (J

f )(x) 0 and
_
R
(J

f )(x)dx =
_
R
f (x)dx.
Proof. The rst part is evident. Then latter part may be proved as fol-
lows: For any g(x) 0 of L
1
(R) there exists a sequence { f
k
(x)} D(A)
1. Integration of the Fokker-Planck equation 137
such that slim f
k
= f exists and slim( f
k

1
Af
k
) = f
1

Af = g.
By the boundary condition on f
k
, we have
_
R
( f
k

1
Af
k
)dx =
_
R
f
k
dx,
(Put h(x) 1 in Greens formula). So in the limit we have
_
R
gdx =
_
f dx.
Also, by the Corollary to Lemma 1,
_
| f
k

1
Af
k
|dx
_
| f
k
|dx,
and hence
_
|g|dx
_
| f |dx.
Therefore by the positivity of g(x) 118
_
f (x)dx =
_
g(x)dx =
_
|g(x)|dx
_
| f (x)|dx
proving that J

is a transition operator.
Therefore the semi-group
T
t
u = strong

limexp(t

AJ

)u
= strong

limexp(t(J

I)u)
is a semi-group of transition operators.
Lecture 22
1 Integration of the Fokker-Planck equation
(Continued)
Before going into the proof of the dierentiability of the operator-theo- 119
retical solution u(t, x) = (T
t
u)(x) we shall discuss the question of the
denseness of the range of the set
_
(I
1
A) f , f D(A)
_
.
If the range of (I
1
A) were not dense in L
1
(R), there will exists
h M(R) = L
1
(R)

, h 0 such that
_
R
(I
1
A) f .hdx = 0, f D(A).
h is a weak solution of the equation (I
1
A

)h = 0. Since h L
2
(R)
and A

is elliptic, h is almost everywhere equal to a bounded C

solution
of (I
1
A

)h = 0. Let {R
k
} be a monotone increasing sequence of
domains R with smooth boundary such that R
k
tends to R very
smoothly. Then we have
0 =
_
R
k
h(I
1
A) f dx
_
R
f (I
1
A

)hdx
=
1
_
R
k
(hf f A

h)dx
139
140 22. Lecture 22
=
1
_

_
_
R
k

ga
i j
_
h
f
x
j
f
h
x
j
_
cos(n, x
i
)dS
+
_
R
k
_

ga
i j
x
j

gb
i
_

_
cos(n, x
i
) f (x)h(x)dS
_

_
.
By the boundedness of h and the boundary condition on f , we have
lim
k
_
R
k
_

ga
i j
f
x
j
+
_

ga
i j
x
j

gb
i
_

_
f
_

_
cos(n, x
i
)hdS = 0.
Therefore h must satisfy the boundary condition 120
lim
k
_
R
m

ga
i j
f
h
x
j
cos(n, x
i
)dS = 0 for every f D(A).
Such a bounded solution h of (I
1
A

)h = 0 is identically zero
and hence

A is the innitesimal generator of a semi-group T
t
in L
1
(R) in
either of the following cases:
(i) R is compact (without boundary).
(ii) R is a half-line ir a nite closed interval on the real line and A =
d
2
/dx
2
.
Proof. (i) At a maximum (or minimum) point x

of h(x) we must
have A

h(x

) 0 (resp. 0) so that the continuous solution h(x)

of A

h = h cannot have either a positive maximum or a negative

minimum.
(ii) The boundary condition for h is
h
n
= 0 and the general solution
of A

h = h is

1. Integration of the Fokker-Planck... 141

h = C
1
e

x
C
2
e

x
dh
dx
= C
1

x + C
2

x
so that the vanishing of
dh
dx
at two points implies that C
1
= C
2
= 0. And
the vanishing of
dh
dx
at one point implies either C
1
= C
2
= 0 or C
1
and
C
2
0. The latter contingency contradicts the boundedness of h.
A parametrix for the operator
_

+ A

_
Let (P, Q) = r(P, Q)
2
be the square of the shortest distance be-
tween the points P and Q according to the metric dr
2
= a
i j
dx
i
dx
j
, where
(a
i j
) = (a
i j
)
1
. We have the
Theorem. For any positive k we may construct a parametrix H
k
(P, Q, 121
t ) for
_

+ A

_
of the form
H
k
(P, Q, t ) = (t )
m/2
exp
_

(P, Q)
4(t, )
k

i=0
u
i
(P, Q)(t )
i
_

_
,
where u
i
(P, Q) are C

functions in a vicinity of P and u

i
(P, P) = 1, we
have
_

Q
_
H
k
(P, Q, t ) = (t )
km/2
exp
_

(P, Q)
4(t )
_
C
k
(P, Q)
C
k
(P, Q) being C

functions in a vicinity of P.
Proof. We introduce the normal coordinates* y

of the point Q = (x
1
,
. . ., x
m
) in suitable neighbourhood of P.
y

= {(P, Q)}
1
2
_
dx

dr
_
P=Q
Let dr
2
=
i j
(y)dy
i
dy
j
. We rst show that when we apply the oper-
ator
A

= Ay

=
i j

2
y
i

+
i

y
i
+ e
142 22. Lecture 22
on a function f (, y) ( is function on y) we have,
A

y
= 4

2
f

2
+ 4y

2
f

+ M
f

+ N( f )
N( f ) =
i j

2
f
y
i
y
j
+
i
f
y
i
+ e f

(The dierentiations have to be performed as though and y were

independent variables). To prove this, we need the well-known formu-
lae:
(P, Q) =
i j
(0)y
i
y
j
(1)

i j
(y)y
i
=
i j
(0)y
j
.
Dene 122
d
dy
i
f (y, ) =
f
y
j
+
f

.

y
j
.
Then
d
2
dy
i
dy
j
{ f (y, )} =

y
i
_
f
y
j
+
f

y
j
_
+
f

_
f
y
j
+
f

y
j
_

y
i
=

2
f
y
i
y
j
+

2
f
y
i

y
j
+
f

y
i
y
j
+

2
f
y
j

y
i
+

2
f

2
.

y
j

y
i
So, by (1)
i j
d
2
f
dy
i
dy
j
+
i
d
dy
i
f + e f
=
_

i j

y
i

y
j
_

2
f

2
+ 2
i j

y
j

2
f
y
i

+ M
f

+ N( f )
= 4

2
f

2
+ 4y

2
f
y

+ M
f

+ N( f ).
1. Integration of the Fokker-Planck... 143
Now applying the above formula to H
k
, we have
A

y
H
k
=
k

i=0

4
(t )
i2m/2
exp
_

4(t )
_
u
i
+
k

i=0
(t )
i1m/2
exp
_

4(t )
_ _
y

u
i
y

+
M
4
u
i
N(u
i1
)
_
(t )
km/2
exp
_

4(t )
_
N(u
k
)
where u
1
= 0, N(u
1
) = 0. Since

H
k
=
k

i=0
(t )
i1m/2
exp
_

4(t )
_
u
i
_

m
2
+ i +

4(t )
_
the theorem will be prove d if we can choose u
i
satisfying the relations
y

u
i
y

+
_

m
2
+ i +
M
4
_
u
i
= N(u
i1
),
u
i
(P, Q) being C

in a vicinity of P with u
1
= 0 and u

(P, P) = 1. To 123
see that we may choose such u
i
, put y

s and transform the equation

as an ordinary dierential equation in s containing the parameters :
du
i
ds
+
_

m
2
+ i +
N
4
_
u
i
= N(u
i1
).
Choose
u

= exp
_

s
_

s
1
_

m
2
+
M
4
_
ds
_

_
.
u
i
is C

near s = , because of the order relation M = 2m+o(s). Dene

u
i
successively by the formula
u
i
(P, Q) = u

s
1
s
_

s
i1
u
1

N(u
i1
)ds (i = 1, 2, . . . , k).
Lecture 23
1 Integration of the Fokker-Planck equation
(Contd.) Dierentiability and representation
of the operator-theoretical solution
f (t, x) = (T
t
f )(x), f L
1
(R)
Lemma 1.1. Let h(x, ) be C

in (x, ) x R, t 0, and vanish 124

outside a compact coordinate neighbourhood of P (independent of ).
Then
_
R
f (y, t) h(y, t)dy =
_
R
f (y, o)h(y, o)dy
+
_
t
o
d
_
R
_

f (y, ).h(y, ) + f (y, )

h(y, )

_
dy
where

f (y, ) = strong lim

0
{ f (y, + ) f (y, )}
1
.
Proof. f (y, ) h(y, ) is weakly dierentiable with respect to in L
1
(R)
and the weak derivative is

f (y, ) h(y, ) + f (y, )

h(y, )

145
146 23. Lecture 23
Corollary. We have
_
R
f (y, t) h(y, t)dy =
_
R
f (y, o)h(y, o)dy
+
_
t
o
d
__
R
f (y, )
_
h(y, )

+ A

y
h(y, )
__
dy
Proof. By Lemma 1.1, the right hand side is
=
_
t
o
d
_
f (y, )
_

h(y, )

y
h(y, )
_
h(y, ) (

f (y, )

A
y
f (y, t)
_
dy
=
_
t
o
d
_
R
_

f (y, )h(y, ) + f (y, )

h(y,

_
dy
+
_
t
o
d
_
R
_
f (y, )A

y
h(y, ) h(y, )

A
y
f (y, )
_
dy.

We have, by the denition of the smallest closed extension

A, 125
_
R
_
f (y, )A

y
h(y, ) h(y, )

A
y
f (y, )
_
dy
= lim
k
_
R
_
f
k
(y, )A

y
h(y, ) h(y, )A
y
f
k
(y, )
_
dy
where slim f
k
= f , slim A
y
f
k
=

Af . The integral on the right is zero,
by Greens formula and the fact that h vanishes near the boundary R.
We take for h(y, ) the function
h(y, ) = h(Q, ) = H
k
(P, Q, t + )(P, Q)(P
o
, P);
here P
o
is a point of R, a positive constant and (P, Q) = (r(P, Q))
where (r) is C

function of r such that 0 (r) 1, (r) = 1 for

r 2
1
and = 0 for r . > 0 is chosen so small that the point Q
1. Integration of the Fokker-Planck equation (Contd.) .... 147
satisfying (P
o
, P) (P, q)0 are contained in a compact coordinate neigh-
bourhood of P
o
. We then have
f (Q, t)H
k
(P, Q, ) (P
o
, P) (P, Q)dQ
=f (Q, 0)H
k
(P, Q, t + ) (P
o
, P) (P, Q)dQ (2)

_
t
o
d
_
f (Q, ) K
k
(P, Q, t + )dQ
where
K
k
(P, Q, t + ) =
_

+ A

Q
_
H
k
(P, t + ) (P
o
, P) (P, Q)
If k is chosen such that k
m
2
2, then by lemma 1.1, K
k
(P, Q, t +
) is for r(P
o
, P) 2
1
, devoid of singularity even if t + = 0.
We now show that the left side of (2) tends as 0 to f (P, t) in the
vicinity of P
o
.
_
R
(P
o
, P)dP| f (Q, t)H
k
(P, Q, )(P, Q)dQ
(P, t)
_
H
k
(P, Q, ) (P, Q)dQ|
C
_
(P
o
,Q)2
_

_
_
r(P
o
,P)
| f (Q, t) f (P, t)|dP
_

_
|H
k
(P, Q, )dQ
C
1
_

_
_
_
| f (z +
f rac12
, t) f (z, t)|dzex
_

2
i
4
_

_
d
1
d
n
((z
1
z
m
) and (z
1
+ y
1
, . . . , z
m
+ y
m
) are coordinates of P and C, C
1
are 126
constants). The inner integral on the right converges as 0, to zero
boundedly by Lebesgues theorem.
There exists therefore a sequence {
i
} with
i
0 such that
f (P, t) lim
i
_
H
k
(P, Q,
i
) (P, Q)dQ =
_
R
f (Q, 0)H
k
(P, Q, t)(P, Q)dy

_
t
0
d
_
R
f (Q, )K
k
(P, Q, t )dQ
148 23. Lecture 23
almost everywhere with respect to P in the vicinity of P
o
. Hence f (P, t)
may be considered to be continuously dierentiable once in t > o and
twice in P in vicinity of P
o
if
lim
o
_
R
H
k
(P, Q, )(P, Q)dQ
is positive and twice continuously dierentiable in P in the vicinity of
P
o
. Now,
lim
o
_
H
k
(P, Q, ) (P, Q)dQ = lim
o
_

m/2
exp
r(P,Q)
_
(P, Q)
4
_
dQ
for > 0. Hence, putting
ds
2
=
i j
(y)dy
i
dy
j
, y
i
=
1
2

i
, lim
o
_
R
H
k
(P, Q, )(P, Q)dQ
= lim
o
_

_

1/2
S
exp(
i j
(0)
i

j
) (0)
1
2
d
1
d
n
=
m/2
(((0))
1
2
((0))
1
2
=
m/2
(g(P))
1
2
/((P))
1/2
where g(P) = det(G
i j
(P)) and (P) = det(
i j
(P)). 127
Thus in the vicinity of P
o
, f (P, t) is equivalent to

m/2
g(P)

1
2
(P)
1
2
_
_
R
f (Q, 0) H
k
(P, Q, t) (P, Q)dQ

_
t
o
d
_
R
f (q, )K
k
(P, Q, t )dQ
So it is dierentiable once in t and twice in P. Moreover, we have
| f (P, t)| Const || f (Q, 0)|. Therefore there exists (P, Q, t) bounded in
Q, such that
f (P, t) =
_
(P, Q, t) f (Q, 0)dQ.
Bibliography
[1] F.V. Atkinson: 128
(i) On a theorem of K. Yosida, Proc. Jap. Acad. 28 (1952),327-
329.
[2] N. Dunford:
(i) On one-parameter semi-group of linear transformations,
Ann. of Math. 39 (1938), 569-573.
[3] W. Feller:
(i) Zur Theorie der stochstischen Prozesse, Math. Ann. 113
(1937), 113-160.
(ii) Some recent trends in the mathematical theory of diusions,
Proc. Internat. Congress of Math. Vol. 2. (1950) 322-339.
(iii) Two singular diusion problems, Ann. of Math. 54 (1951),
173-182.
(iv) The parabolic dierential equations and the associated semi-
groups of transformations, Ann. of Math. 55 (1952), 468-
519.
(v) On a generalization of M.Riesz potentials and the semi-
groups generated by them, Comm. Semin. Math. de 1Univ.
de Lund. tome supplement, (1952) dedie a M.Riesz, 73-81.
(vi) On positivity preserving semi-groups of transformations on
C[r
1
, r
1
], Ann. Soc. Polonaise de Math. 25(1952), 85-94.
149
150 BIBLIOGRAPHY
(vii) Semi-groups of transformations in general weak topologies,
Ann. of Math. 57 (1953), 287-308.
(viii) On the generation of unbounded semi-groups of bounded lin-
ear operators, Ann. of Math. 58 (1953), 166-174.
[4] K. 0. Friedrichs:
(i) On the dierentiability of the solutions of linear elliptic dif-
ferential equations, Comm. Pure App. Math. 6 (1953) 299-
326.
[5] M. Fukamiya:
(i) On one-parameter groups of operators, Proc. Imp. Acad.
Tokyo, 16 (1940), 262-265.
[6] I. Gelfand:
(i) On one-parameter groups of operators in a normed space,
C.R.(Doklady) Acad. Sci. URSS, 25 (1939), 713-718.
[7] E. Hille: 129
(i) Functional Analysis and Semi-groups, Amer. Math. Soc.
Colloq. Publ. 31 (1948).
(ii) On the integration problem for Fokker-Plancks equatin, 10th
Congress of Scandin. Math. Trondheim (1949), 183-194.
(iii) Les probabilites continues en chaine, C.R.Acad. Sci. Paris,
230 (1950), 34-35.
(iv) Explosive solutions of Fokker-Plancks equation, Proc. Inter-
nat. Congress of Math. Vol. 2 (1950), 435.
(v) On the generation of semi-groups and the theory of conjugate
functions, Comm. Semin. Math. de 1Univ. de Lund, tome
supplement (1952) dedie a M. Riesz, 1-13.
(vi) Une generalisation du probleme de Cauchy, Ann. LInstitut
Fourier, 4 (1952), 31-48.
BIBLIOGRAPHY 151
(vii) Sur le probleme abstrait de Cauchy, C.R. Acad. Sci. Paris,
236 (1953), 1466-14670.
(viii) Le probleme abstrait de Cauchy, Rendiconti del Seminario
Matermatico di Univ. di Torino, 12 (1952-1953), 95-103.
(ix) A note on Cauchys problem, Ann. Soc. Polonaise de Math.
25 (1952), 56-68.
(x) On the integration of Kolmogoros dierential equations,
Proc. Nat. Acad. Sci. (40 (1954), 20-25.
[8] K. Ito:
(i) On stochastic dierential equations, Memoirs of the Amer.
Math. Soc. No. 4 (1951).
(ii) Brownian motion in a Lie group, Proc. Jap. Acad. 26 (1950),
4-10.
(iii) Stochastic dierential equations in a dierentiable manifold,
Nagoya Math. J.E. (1950), 35-47.
[9] S. Ito:
(i) The fundamental solution of the parabolic equations in a dif-
ferentiable manifold, Osaka Math. J. 5 (1953), 75-92.
[10] T. Kato:
(i) Integration of the equation of evolution in a Banach space,
J.Math. Soc. japan, 5 (1953), 208-234.
(ii) On the semi-group generated by Kolmogoros dierential
equations, forthcoming in the J.Math. Soc. Japan.
[11] A. Kolmogoro: 130
(i) Uber die analytischen Methoden in Nahrscheinlichkeit srech-
nung, Math. Ann. 104 (1931), 415-458.
(ii) Zur-Theorie der stetigen zufalligen Prozesse, Math. Ann. 108
(1933), 149-160.
152 BIBLIOGRAPHY
[12] P.D.Lax:
(i) On Cauchys problem for hyperbolic equations and the dif-
ferentiability of solutions of elliptic equations, Comm. Pure
App. Math., 8 (1955), 615-633.
[13] P. D. Lax and A.N. Milgram:
(i) Parabolic equatins, Contributions to the theory of partial dif-
ferential equations, Princeton, 1954.
[14] S. Minakshisundaram:
(i) Eigenfunctions on Riemannian manifolds, J. Ind. Math. Soc.,
17 (1953 - 1954), 159-165.
[15] S. Minakshisundaram and A. Pleijel:
(i) Some properties of the eigenfunctions of the Laplace opera-
tor on Riemannian manifolds, Canad. J. Math 1 (1949), 242-
256.
[16] L.Nirenberg:
(i) Remarks on strongly elliptic dierential equations, Comm.
Pure App. Math. 8(1955), 649-675.
[17] F. Perrin:
(i) Etude mathematique du mouvement brownien de rotation,
Ann. Ec. norm. sup. 45 (1928), 1-51.
[18] R.S. Phillips:
(i) On one-parameter semi-groups of linear operators, Proc.
Amer. Math. Soc. 2 (1951), 234-237.
(ii) Perturbation theory for semi-groups of linear operators,
Trans. Amer. Math. Soc. 74 (1953), 199-221.
BIBLIOGRAPHY 153
(iii) An inversion formula for Laplace transforms and semi-
groups of linear operators, Ann. of Math. 59 (1954), 325-
356.
[19] G. de Rham: 131
(i) Varietes Dierentiables, Paris, 1955.
[20] F. Riesz and B. Sz. Nagy:
(i) Lecons danalyse fonctionnelle, Budapest, 1952.
[21] M.H.Stone:
(i) Linear transformations in Hilbert space III, Proc. Nat. Acad.
Sci. 16(1930), 172-175.
[22] L. Schwartz:
(i) Theorie des distributions I et II, Paris (1950-1951).
[23] K. Yosida:
(i) On the dierentiability and the representation of one-
parameter semi-group of linear operators, J. Math. Soc.
Japan, 1 (1948), 15-21.
(ii) An operator -theoretical treatment of temporally homoge-
neous Marko process, J. Math. Soc. Japan, 1 (1949) 244-
253.
(iii) Brownian motion on the surface of the 3-sphere, Ann. of
Math. Statist. 20 (1949), 292-296.
(iv) Integration of Fokker-Plancks equation in a compact Rie-
mannian space, Arkiv for Matematik, 1 (1949), 244-253.
(v) Stochastic process built from ows, Proc. Internat. Congress
Math. Vol. 1 (1950).
(vi) Integration of Fokker-Planacks equation with a boundary
condition, J. Math. Soc. Japan, 3 (1951), 29-73.
154 BIBLIOGRAPHY
(vii) A theorem of Liouvilles type for meson equations, Proc. Jap.
Acad. 27 (1951), 214-215.
(viii) Integrability of the backward diusion equation in a compact
Riemannian space, Nagoya Math. J 3(1951), 1-4.
(ix) On Brownian motion in a homogeneous Riemannian space,
Pacic J. of Math. 2 (1952), 263-270.
(x) On the integration of diusion equations In Riemannian
spaces, Proc. Amer. Math. Soc. 3 (1952), 864-873.
(xi) On the fundamental solution of the parabolic equations in a
Riemannian space, Osaka Math. J. 5 (1953), 65-74.
(xii) On the integration of the temporally inhomogeneous diu-
sion equation in a Riemannian space, Proc. Jap. Acad. 30
(1954), 19-23 and 273-275.