NOTES ON

CLASSICAL
POTENTIAL THEORY

M. Papadimitrakis
Department of Mathematics
University of Crete

January 2004
2

Forword

During the fall semester of the academic year 1990-1991 I gave a course on
Classical Potential Theory attended by an excellent class of graduate students
of the Department of Mathematics of Washington University. That was my first
time to teach such a course and, I have to say, besides sporadic knowledge of a
few facts directly related to complex analysis, I had no serious knowledge of the
subject. The result was: many sleepless nights reading books, trying to choose
the material to be presented and preparing hand-written notes for the students.
The books I found very useful and which determined the choice of material
were the superb El ements de la Theorie Classique du Potentiel by M. Brelot
and the Selected Problems on Exceptional Sets by L. Carleson. Other sources
were: Some Topics in the Theory of Functions of One Complex Variable by
W. Fuchs, unpublished notes on Harmonic Measures by J. Garnett, Subhar-
monic Functions by W. Hayman and P. Kennedy, Introduction to Potential
Theory by L. Helms, Foundations of Modern Potential Theory by N. Land-
kof, Subharmonic Functions by T. Rado and Potential Theory in Modern
Function Theory by M. Tsuji.
This is a slightly expanded version of the original notes with very few
changes. The principle has remained the same, namely to present an overview
of the classical theory at the level of a graduate course. The part called Pre-
liminaries is new and its contents were silently taken for granted during the
original course. The main material is the Divergence Theorem and Greens For-
mula, a short course on holomorphic functions (, since their real parts are the
main examples of harmonic functions in the plane and, also, since one of the
central results is the proof of the Riemann Mapping Theorem through potential
theory), some basic facts about semi-continuous functions and very few ele-
mentary results about distributions and the Fourier transform. Except for the
Divergence Theorem, the Arzela-Ascoli Theorem, the Radon-Riesz Representa-
tion Theorem and, of course, the basic facts of measure theory and functional
analysis, all of which are used but not proved here, all other material contained
in these notes is proved with sufficient detail.
Material which was not included in the original notes: the section on har-
monic conjugates in the first chapter (it, actually, contains a new proof of the
existence of a harmonic conjugate in a simply-connected subset of the plane);
the section on the differentiability of potentials in the second chapter; the sec-
tions on superharmonic functions at and on Poisson integrals at in the
fourth chapter; an additional proof of the result about the direct connection
between Greens function and harmonic measure in the fifth chapter (indicating
the role of the normal derivative of Greens function as an approximation to
the identity); the subadditivity of capacity in the eigth chapter; the sections on
polar sets and thin sets in the ninth chapter. The definition of the notion of
quasi-almost everywhere in the eigth chapter has been changed. The proof of
the Riemann Mapping Theorem in the ninth chapter is corrected and given in
 3

full detail, not relying on obvious topological facts any more. In the original
course the proof (taken from the notes of J. Garnett) of Wieners Theorem was
presented only in dimension 2. Now, the proof is given in all dimensions.
A short and very classical application of potential theory in dimension 1 on
the convergence of trigonometric series is missing from this set of notes, since
it is quite specialized. What is, also, missing is a short chapter on the metrical
properties of capacity and an example of a Cantor-like set. But this will be
included very soon, after it is expanded as a chapter on Capacity, capacitability
and Hausdorff measures.
Besides the new material, there is a re-organization which results, I hope, to
better exposition.
Here, I would like to thank the Department of Mathematics of Washington
University for giving me the opportunity to teach the original course and the
graduate class which attended it with great care and enthusiasm.
4
Contents

I Preliminaries 9
0.1 Euclidean Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
0.2 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
0.3 Holomorphic Functions . . . . . . . . . . . . . . . . . . . . . . . . 18
0.4 Equicontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
0.5 Semi-continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
0.6 Borel Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
0.7 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
0.8 Concavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
0.9 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . 51

II Main Theory 65
1 Harmonic Functions 67
1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
1.2 Maximum-minimum principle . . . . . . . . . . . . . . . . . . . . 69
1.3 Differentiability of harmonic functions . . . . . . . . . . . . . . . 70
1.4 Holomorphy and harmonic conjugates . . . . . . . . . . . . . . . 71
1.5 Fundamental solution . . . . . . . . . . . . . . . . . . . . . . . . 75
1.6 Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
1.7 Flux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
1.8 The representation formula . . . . . . . . . . . . . . . . . . . . . 80
1.9 Poisson integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
1.10 Consequences of the Poisson formula . . . . . . . . . . . . . . . . 85
1.11 Monotone sequences . . . . . . . . . . . . . . . . . . . . . . . . . 90
1.12 Normal families of harmonic functions . . . . . . . . . . . . . . . 91
1.13 Harmonic distributions . . . . . . . . . . . . . . . . . . . . . . . . 93

2 Superharmonic Functions 97
2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
2.2 Minimum principle . . . . . . . . . . . . . . . . . . . . . . . . . . 98
2.3 Blaschke-Privaloff parameters . . . . . . . . . . . . . . . . . . . . 100
2.4 Poisson modification . . . . . . . . . . . . . . . . . . . . . . . . . 105
2.5 Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5
6 CONTENTS

2.6 Differentiability of potentials . . . . . . . . . . . . . . . . . . . . 109

2.7 Approximation, properties of means . . . . . . . . . . . . . . . . 112
2.8 The Perron process . . . . . . . . . . . . . . . . . . . . . . . . . . 116
2.9 The largest harmonic minorant . . . . . . . . . . . . . . . . . . . 117
2.10 Superharmonic distributions . . . . . . . . . . . . . . . . . . . . . 119
2.11 The theorem of F. Riesz . . . . . . . . . . . . . . . . . . . . . . . 122
2.12 Derivatives of superharmonic functions . . . . . . . . . . . . . . . 123

3 The Problem of Dirichlet 125

3.1 The generalized solution . . . . . . . . . . . . . . . . . . . . . . . 125
3.2 Properties of the generalized solution . . . . . . . . . . . . . . . . 128
3.3 Wieners Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 132
3.4 Harmonic measure . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3.5 Sets of zero harmonic measure . . . . . . . . . . . . . . . . . . . 138
3.6 Barriers and regularity . . . . . . . . . . . . . . . . . . . . . . . . 140
3.7 Regularity and the problem of Dirichlet . . . . . . . . . . . . . . 143
3.8 Criteria for regularity . . . . . . . . . . . . . . . . . . . . . . . . 145

4 The Kelvin Transform 149

4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.2 Harmonic functions at . . . . . . . . . . . . . . . . . . . . . . 150
4.3 Superharmonic functions at . . . . . . . . . . . . . . . . . . . 154
4.4 Poisson integrals at . . . . . . . . . . . . . . . . . . . . . . . . 155
4.5 The effect of the dimension . . . . . . . . . . . . . . . . . . . . . 156
4.6 Dimension 2, in particular . . . . . . . . . . . . . . . . . . . . . . 157

5 Greens Function 161

5.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.2 Greens function, the problem of Dirichlet and harmonic measure 162
5.3 A few examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.4 Monotonicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.5 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.6 Greens function and regularity . . . . . . . . . . . . . . . . . . . 166
5.7 Extensions of Greens Function . . . . . . . . . . . . . . . . . . . 167
5.8 Greens potentials . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.9 The Decomposition Theorem of F. Riesz . . . . . . . . . . . . . . 172
5.10 Greens function and harmonic measure . . . . . . . . . . . . . . 173
5.11 as interior point. Mainly, n = 2 . . . . . . . . . . . . . . . . . 183

6 Potentials 187
6.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
6.2 Potentials of non-negative Borel measures . . . . . . . . . . . . . 189
6.3 The maximum principle for potentials . . . . . . . . . . . . . . . 192
6.4 The continuity principle for potentials . . . . . . . . . . . . . . . 195
CONTENTS 7

7 Energy 197
7.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
7.2 Representation of energy: Greens kernel . . . . . . . . . . . . . . 199
7.3 Measures of finite energy: Greens kernel . . . . . . . . . . . . . . 205
7.4 Representation of energy: kernels of first type . . . . . . . . . . . 206
7.5 Measures of finite energy: kernels of first type . . . . . . . . . . . 212

8 Capacity 217
8.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
8.2 Equilibrium measures . . . . . . . . . . . . . . . . . . . . . . . . 221
8.3 Transfinite diameter . . . . . . . . . . . . . . . . . . . . . . . . . 229
8.4 The Theorem of Evans . . . . . . . . . . . . . . . . . . . . . . . . 232
8.5 Kernels of variable sign . . . . . . . . . . . . . . . . . . . . . . . 234

9 The Classical Kernels 237

9.1 Extension through sets of zero capacity . . . . . . . . . . . . . . 237
9.2 Sets of zero harmonic measure . . . . . . . . . . . . . . . . . . . 238
9.3 The set of irregular boundary points . . . . . . . . . . . . . . . . 239
9.4 The support of the equilibrium measure . . . . . . . . . . . . . . 243
9.5 Capacity and conformal mapping . . . . . . . . . . . . . . . . . . 244
9.6 Capacity and Greens function in R2 . . . . . . . . . . . . . . . . 249
9.7 Polar sets and the Theorem of Evans . . . . . . . . . . . . . . . . 251
9.8 The theorem of Wiener . . . . . . . . . . . . . . . . . . . . . . . 255
9.9 Thin Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
8 CONTENTS
 Part I

Preliminaries

9
0.1. EUCLIDEAN SPACES 11

0.1 Euclidean Spaces

1. We work in the Euclidean space Rn and denote the Euclidean norm of x Rn
by |x| and the Euclidean inner product of x, y Rn by x y .
B(x; r) is the open ball with center x and radius r, B(x; r) is the closed ball
and the sphere S(x; r) is the boundary of B(x; r).
d(x, B) = inf yB |x y| denotes the Euclidean distance of the point x Rn
from the non-empty subset B of Rn and d(A, B) = inf xA,yB |x y| denotes
the Euclidean distance between the non-empty sets A and B. If B is closed and
x / B, then d(x, B) > 0 and, if A is compact, B is closed and the two sets are
disjoint, then d(A, B) > 0.
2. Suppose is an open subset of Rn and consider the open sets
{ 1 }
(m) = x : d(x, ) > , |x| < m , m N.
m
It is easy to check the following four properties:

1. every (m) is a compact subset of ,

2. (m) (m+1) for all m,

3. +
m=1 (m) = and

4. every compact subset of is contained in (m) for a sufficiently large m.

This increasing sequence {(m) } of open sets, or any other with the same
four properties, is called an open exhaustion of .
The increasing sequence {K(m) }, with K(m) = (m) , where {(m) } is an
open exhaustion of , is called a compact exhaustion of .
n
3. Vn = (n 2+1) is the volume of Bn = B(0; 1). Hence,
2

m 2(2)m
V2m = , V2m+1 = .
m! 1 3 5 (2m + 1)

Also, n1 = nVn is the standard surface area of S n1 = S(0; 1) .

4. If dm is the Lebesgue measure in Rn and d is the standard surface measure
in S n1 , then we have the formula
R
f (y) dm(y) = f (x + rt) d(t) rn1 dr
B(x;R) 0 S n1

5. We define the surface-mean-value of f over S(x; r) by

1 1
Mrf (x) = f (x + rt) d(t) = f (y) dS(y)
n1 S n1 n1 rn1 S(x;r)

for all f integrable with respect to dS, the surface measure in S(x; r).
12

We also define the space-mean-value of f over B(x; r) by

1 1
Arf (x) = f (x + ry) dm(y) = f (y) dm(y)
Vn Bn Vn rn B(x;r)

for all f integrable with respect to dm in B(x; r).

By the formula in paragraph 4,
R
n
Af (x) = n
R
Mrf (x)rn1 dr .
R 0

6. Define Rn , the one-point compactification of Rn , by adjoining the point

at to Rn :
Rn = Rn {} .
The -neighborhoods of points x Rn are the usual balls B(x; ), while the
-neighborhood of is defined to be the set {x Rn : |x| > 1 } {} .
We define open sets in Rn through these neighborhoods, in the usual way,
and we, also, define closed sets (complements of open sets) and the notion of
convergent sequence: a sequence in Rn converges to some point of Rn , if the
sequence is, eventually, contained in every -neighborhood of this point.
Hence, if the limit point is in Rn , then the new notion of convergence coin-
cides with the usual one, while xm is equivalent to |xm | +.
Whenever we write A and A, for any A Rn , we mean the closure and
the boundary, respectively, of A with respect to Rn .
Hence, if A Rn is bounded, then these two sets coincide with the usual
closure and boundary in Rn , while if A Rn is unbounded, then the two sets
are the usual closure and boundary in Rn with the point adjoined to them.
Any A Rn is open in Rn in the usual sense if and only if it is open in Rn .
It is easy to see that Rn and, hence, every closed subset of it, is compact.
If A is a bounded subset of Rn , then A is closed in Rn if and only if it is
closed in Rn . But, if A is an unbounded subset of Rn , then A is closed in Rn
if and only if A {} is closed in Rn .
The spherical metric in Rn is defined by

2|xy|
1+|x|2 1+|y|2 , if x, y R
n

dS (x, y) = 2 , if x Rn and y =

1+|x|2
0, if x = y = .
The spherical metric induces exactly the open sets, closed sets and conver-
gent sequences in Rn which were described above. In fact, one can, easily,
prove that if x Rn , then, for every -neighborhood BS (x; ) with respect to
the spherical metric, there is some 0 -neighborhood of x, as this was defined
above, contained in BS (x; ) and conversely.
7. A subset E of a metric space (X, d) is called connected, if, whenever we
write E = A B with A B = with A, B being both open relative to E, it is
implied that one of A, B is empty.
0.2. DERIVATIVES 13

A subset F of E is called a connected component of E, if it is a maximal

connected subset of E. This means that F is connected and that there is no
connected subset of E strictly containing F .
Every E can be uniquely decomposed in connected components: there exists
a unique (perhaps uncountable) family F such that
1. every F F is a connected component of E,
2. the elements of F are pairwise disjoint,

3. F F F = E.

Especially if Rn is open, then is connected if and only if every two

of its points can be connected by a polygonal path which is contained in .
The connected components of an open are all open sets and there are at
most countably many of them.
Therefore, for every open Rn there exist (at most countably many) sets
Um such that
1. every Um is connected,
2. the sets Um are pairwise disjoint and

3. m Um = .
A subset E of a metric space (X, d) is called a continuum if it is connected,
compact and contains at least two points.

0.2 Derivatives
1. If is an open subset of Rn , then C() = C 0 () is the space of all complex-
valued functions which are continuous in and C k (), 1 k +, is the
space of all functions which are k times continuously differentiable in .
Similarly, we denote by C k () the space of all functions whose derivatives
of all orders up to k are continuous in and can be continuously extended in
. Note that a complex-valued function continuous in can be continuously
extended in if and only if it is uniformly continuous in with respect to
the spherical metric dS (or, equivalently, with respect to the usual Euclidean
distance in case the set is bounded).
For all multi-indices = (1 , . . . , n ), we denote the -derivative of f by

1 ++n f || f
D f = n = .
x1 xn
1
x1 x
1
n
n

The order of this derivative is || = 1 + + n .

We denote, as usual, the gradient of a real- or complex-valued function by
( f f )

grad f = ,...,
x1 xn
14

and is the Laplace operator or Laplacian

2f 2f
f = + + .
x21 x2n


If V = (V1 , V2 , . . . , Vn ) is a vector-function, we denote its divergence by

V1 Vn
div V = + + .
x1 xn

2. We say that the open set is C 1 at its boundary point y Rn , if

there is an open neighborhood V of y and a : V R which is in C 1 (V ) with

grad (y) 6= 0, so that V = {x V : (x) = 0}, V = {x V : (x) > 0}

and V \ = {x V : (x) < 0} .
Such a function is called a defining function for in the neighborhood
V of the boundary point y. A defining function is not unique; for example any
multiple of it by a positive constant is, also, a defining function.
In case there exists a defining function C k (V ) in a neighborhood V of
the boundary point y, we say that is C k at its boundary point y.
If is C 1 at its boundary point y Rn , we denote by
(y) the unit vector
at y which is normal to and has the direction towards the exterior of . In
terms of any defining function for in a neighborhood of y,


grad (y)
(y) = .
|grad (y)|

The directional derivative of an f in the direction of the unit normal

(y)
is
f
(y) = grad f (y)

(y) .

If the bounded open set is C k at all of its boundary points, we say that
has C k -boundary.
If the bounded open set has C 1 -boundary, then there is the standard
surface measure defined in , denoted by dS.
Theorem 0.1 (The Divergence Theorem and Greens Formulas) Let be a
bounded open set with C 1 -boundary.
1.



div V (x) dm(x) = V (y)

(y) dS(y)



for all vector-functions V whose components are in C() C 1 ().
2. If f is in C() C 1 () and g is in C 1 () C 2 (), then

( ) g
f (x)g(x) + grad f (x) grad g(x) dm(x) = f (y) (y) dS(y) .

0.2. DERIVATIVES 15

3. If g is in C 1 () C 2 (), then

g
g(x) dm(x) = (y) dS(y) .

4. If f and g are in C 1 () C 2 (), then

( )
( ) g f
f (x)g(x)g(x)f (x) dm(x) = f (y) (y)g(y) (y) dS(y).

Proof:
The proof of 1 is considered known.
g g
Applying 1 to V1 = f x 1
, . . . , Vn = f x n
, we prove 2.
We prove 3 from 2, using the constant function f = 1 in .
Finally, we prove 4, changing places of f and g in 2 and subtracting.

In these notes whenever we refer to the Greens Formula, we understand any

one of the above four formulas.
3. Since open sets with C k -boundary are widely used, (in particular, to apply
Greens Formula), we shall, briefly, describe a standard way to produce such
sets arbitrarily close to other given sets.
Suppose that K , where K is compact and is an open subset of Rn .
In practice, we are, usually, given K and choose = {x : d(x, K) < } with
arbitrarily small or we are given and choose K = {x : d(x, )
, |x| 1 } .
We shall construct a bounded open set O with C k -boundary so that

K O O .

Consider 0 < 1n d(K, ) and observe that, if a cube Q of sidelenghth 0

intersects K, then Q is contained in .
Now, consider the n coordinate-hyperplanes together with all other hyper-
planes which are parallel to them and at distances which are integer multiples
of 0 . The space Rn is, thus, divided into a mesh of cubes

Qk = {x = (x1 , . . . , xn ) : kj 0 xj (kj + 1)0 , 1 j n}

of sidelenght 0 , where k = (k1 , . . . , kn ) is an arbitrary multi-index with integer

coordinates. The set of all these cubes we denote by Q0 .
Now, K is intersected by finitely many qubes in Q0 and we define the set

F = {Q Q0 : Q K 6= } .

F is a compact subset of with K F and it is clear that the boundary of

F consists of certain of the (n 1)-dimensional faces of the cubes that are used
to construct F . If a face of one of the cubes contained in F intersects K, then
the adjacent cube is, also, one of the cubes contained in F and, hence, this face
16

is not contained in the boundary of F . Therefore, the boundary of F consists

of faces which do not intersect K and, thus, K is contained in the interior U of
F.
We have produced an open set U so that K U U whose boundary
consists of faces of cubes contained in U . The points at which this boundary is
not C k belong to the (n 2)-dimensional edges of these faces.
We, now, modify slightly the boundary to make it smooth at its edges and
we, thus, produce an open set O slightly different from U , containing K and
with C k -boundary. The smoothening process is geometrically clear and it is
not worth seeing the actual technical details.
4. A subset of Rn is said to be C 1 at its point y, if there exists some open
neighborhood V of y and a real-valued : V R which is in C 1 (V ) with

grad (y) 6= 0, so that V = {x V : (x) = 0} .

Any such is called a defining function for in the neighborhood of its
point y. Again, is not unique; for example, any multiple of it by a non-zero
real constant is, also, a defining function.
If the defining function can be chosen to be in C k (V ), then we say that
is C k at y.
is called a C k -hypersurface, if it is C k at all of its points.
A C 1 -hypersurface has a standard surface measure dS naturally defined
on it.
If is C k at y and C k (V ) is a defining function for in a neighborhood
V of y, then, because of continuity, we may choose V to be small enough so

that grad 6= 0 everywhere in V . Now, if M = supV and m = inf V , then
m < 0 < M and the sets
t = {x V : (x) = t} , m<t<M ,
are pairwise disjoint C k -hypersurfaces constituting a continuous partition of
V . In fact, itself is a defining function for each t at every point of it.
One can, easily, prove that, for every f C(V ),

f (x) dS(x)
t

(where dS is the surface measure in t ) is a continuous function of t (m, M ).

At every point y where is C 1 there are exactly two unit vectors normal to
. These, in terms of any defining function in a neighborhood of y, are


grad (y)
(y) = .
|grad (y)|
The C 1 -hypersurface is called orientable, if, for every y , we can
choose one of the two possible (y) so that the resulting
: Rn is
continuous. Then, we call
a continuous unit vector field normal to .
In such a case, the function
is, also, a continuous unit vector field
normal to and, if is connected, these are the only continuous unit vector
fields normal to .
0.2. DERIVATIVES 17

If is a bounded open set with C 1 -boundary, then = is an orientable

1
C -hypersurface and there are two continuous unit vector fields normal to ;
one of them has the direction towards the exterior of and the other has the
opposite direction.
In this special case, we keep the notation only for the vector field with
the direction towards the exterior of , in agreement with the discussion in
paragraph 2.
5. A function f which is in C k (B(x0 ; R)) has a Taylor-expansion of order k
at the point x0 . Using the notation ! = 1 ! n ! and y = y11 ynn , this
means that
1
f (x) = D f (x0 )(x x0 ) + Rk (x; x0 ) , x B(x0 ; R) ,
!
||k

|Rk (x;x0 )|
where |xx0 |k
0 as x x0 . In fact,
1 ( )
Rk (x; x0 ) = D f (x0 ) D f (x0 ) (x x0 )
!
||=k

for some x0 = x0 (x) contained in the linear segment [x0 , x] and, thus,

|Rk (x; x0 )| Ck,n |x x0 |k sup |D f (x0 ) D f (x0 )| ,

||=k,|x0 x0 ||xx0 |

where Ck,n is a constant depending only on k and n.

A function f defined in an open neighborhood of x0 Rn is called real-
analytic at x0 , if there is some R > 0 and constants a for all multi-indices ,
so that
f (x) = a (x x0 )

for every x B(x0 ; R).

This series expansion is unique and is called the Taylor-series of f at x0 .
Then, f is real-analytic at every other point of B(x0 ; R), it is infinitely many
times differentiable in B(x0 ; R) and we have the formulas
1
a = D f (x0 ) .
!
6. Suppose that is C 1 at its point y and let C 1 (V ) be a defining function
for in the neighborhood V = B(y; r) of y.
We write the Taylor-expansion of order 1 at y

(x) = grad (y) (x y) + R1 (x; y) , x B(y; r) ,

we choose t with 0 < t < |grad (y)| and we take r small enough so that
|R1 (x; y)| t|x y| for all x B(y; r).
It is, then, a matter of simple calculations to see that the two sets

F = {x : t|x y| < grad (y) (x y) < tr}
18

are two open truncated cones with common vertex y and contained in the sets

{x : (x) > 0} B(y; r)

respectively.
As a special case we have that, if the open set is C 1 at its boundary point
y, then there are two open truncated cones with common vertex y so that one
of them is contained in and the other is contained in Rn \ .
Now, suppose that is C 2 at its point y and let C 2 (V ) be a defining
function for in the neighborhood V = B(y; r) of y.
We, now, write the Taylor-expansion of order 2 at y in the simplified form

(x) = grad (y) (x y) + R2 (x; y) , x B(y; r) ,

where
|R2 (x; y)| M |x y|2
for all x B(y; r). In fact, M = Cn sup||=2,xB(y;r) |D (x)|, where Cn is a
constant depending only on( n. )
If we choose = min 2M 1
, 
r
 , then it is, again, a matter of
2grad (y)
calculations to see that the two open balls
{  ( )  }
B = x : x y grad (y)  < grad (y)

are mutually tangent at the point y and they are contained in the sets

{x : (x) > 0} B(y; r)

respectively.
Hence, if the open set is C 2 at its boundary point y, then there are two
open balls mutually tangent at the point y so that one of them is contained in
and the other is contained in Rn \ .

0.3 Holomorphic Functions

If is an open subset of C = R2 , a function f = <f + i=f : C is
called holomorphic in , if <f, =f C 1 () and they satisfy the system of
Cauchy-Riemann equations

(<f ) (=f ) (<f ) (=f )

= , =
x1 x2 x2 x1
f f
everywhere in or, equivalently, x 1
+ i x 2
= 0 everywhere in .
It is trivial to see that the property of holomorphy is preserved under ad-
dition, multiplication, division and composition of functions and that the usual
formulas (f + g)0 = f 0 + g 0 , (f g)0 = f 0 g + f g 0 , (f g)0 = (f 0 g)g 0 hold.
0.3. HOLOMORPHIC FUNCTIONS 19

If f is holomorphic in , writing the Taylor-expansions of order 1 at every

x for <f and =f and using the Cauchy-Riemann equations, we easily find
that
f f
f (y) = f (x) + (x)(y x) + R(y; x) = f (x) i (x)(y x) + R(y; x)
x1 x2

where |R(y;x)| 0
|yx| 0 as y x. Therefore, the limit f (x) = limyx
f (y)f (x)
yx
exists at every x . This limit is called the (complex) derivative of f at x
and
f f
f 0 (x) = (x) = i (x) , x.
x1 x2
In elementary courses on the theory of holomorphic functions it is, actually,
proved that the converse is, also, true: if the complex derivative of f exists at
every point of , then <f and =f are in C 1 () and satisfy the Cauchy-Riemann
equations. We shall not need this result.
In this section, our aim is to develop only a very small part of the theory of
holomorphic functions leading, through Cauchys Theorem, up to the Argument
Principle and one of its consequences. These results will not be stated in the
generality (requiring homological considerations) in which they are presented
in the standard courses on the theory of holomorphic functions. We shall be
restricted to the study of curvilinear integrals only over boundaries of open sets
with C 1 -boundary. Our main tool, therefore, is the Divergence Theorem.

Theorem 0.2 (Cauchys Theorem) Let f be holomorphic in and the open

set 1 with C 1 -boundary be such that 1 . Then,

f (y) dy = 0 .
1

Remark This integral is called the curvilinear integral of f over 1 in

the positive direction of 1 with respect to 1 . Here, we define

dy = i(y) dS(y) , y 1 ,

where = 1 +i2 is the continuous unit vector field normal to 1 and directed
towards the exterior of 1 .
Proof:
From the Cauchy-Riemann equations and the Divergence Theorem,
( )
f f
0 = i (x) + i (x) dm(x)
x1 x2
1
= i f (y)1 (y) dS(y) f (y)2 (y) dS(y)
1 1

= f (y)i(y) dS(y) = f (y) dy .
1 1
20

Theorem 0.3 (Cauchys Formula) Let f be holomorphic in and the open set
1 with C 1 -boundary be such that 1 . Then,

1 f (y)
f (x) = dy
2i 1 y x
for every x 1 .
Proof:
Let r > 0 be small so that B(x; r) 1 and let 2 = 1 \ B(x; r). We
apply Cauchys Theorem to the function g(z) = fzx
(z)
, z \ {x}, which is
holomorphic in \ {x} and get

f (y) f (y) f (y)
0 = dy = dy dy .
2 y x 1 y x B(x;r) y x

Therefore,

1 f (y) 1 f (y)
dy = dy
2i 1 yx 2i B(x;r) y x

1 f (y) y x
= dS(y)
2 B(x;r) y x r

1 ( )
= f (x) + f (y) f (x) dS(y) .
2r B(x;r)
The continuity of f at x implies that the last term tends to 0 as r 0+
and the proof is complete.

Example The first case below is implied by Cauchys Formula and the second
by Cauchys Theorem:
{
1 1 1 , if x B(x0 ; r)
dy =
2i B(x0 ;r) y x 0 , if x
/ B(x0 ; r) .

Theorem 0.4 Let f be holomorphic in . Then,

1. f 0 is holomorphic in and
2. for every B(x; r) , f can be expanded in B(x; r) in a unique way as an
absolutely convergent power series

+
f (z) = an (z x)n , z B(x; r) .
n=0

Proof: f (y)
1. From Cauchys Formula f (z) = 2i 1
B(x;r) yz
dy, z B(x; r), it is trivial
to prove, using difference quotients for both sides and interchanging limit and
integration, that

1 f (y)
f 0 (z) = dy , z B(x; r) .
2i B(x;r) (y z)2
0.3. HOLOMORPHIC FUNCTIONS 21

Interchanging, again, partial derivatives and integration, we easily see that

f 0 f 0 f 0 f 0
x1 and x 2
are continuous in and x 1
+ i x 2
= 0 everywhere in .
2. For every z B(x; r), the geometric series

1 1 1
+
(z x)n
= =
yz y x 1 yx
zx
n=0
(y x)n+1
 
 zx  |zx|
converges absolutely and uniformly in B(x; r), since  yx  = r < 1. This
permits us to interchange integration and summation in Cauchys Formula to
get

+
f (z) = an (z x)n
n=0
1
f (y)
with an = 2i B(x;r) (yx)n+1
dy.
If

+
+
f (z) = an (z x)n = a0n (z x)n , z B(x; r) ,
n=0 n=0

then, using z = x, we get a0 = a00 . Cancelling a0 and simplifying, we find

+
+
an+1 (z x)n = a0n+1 (z x)n , z B(x; r) \ {x} .
n=0 n=0

By the continuity of both power series at x, we get a1 = a01 . We continue

inductively to conclude that an = a0n for all n.

We may, inductively, see that, if f is holomorphic in , then it has (complex)

derivatives of all orders and that, for every B(x; r) ,

n! f (y)
(n)
f (z) = dy , z B(x; r) .
2i B(x;r) (y z)n+1
+
Therefore, the coefficients in the power series f (z) = n=0 an (z x)n are given
(n)
by an = f n!(x) .
Suppose, now, that f is holomorphic in and consider its expansion as a
power series in any B(x; r) with B(x; r) .
If all coefficients an are equal to 0, then f = 0 everywhere in B(x; r) and we
say that x is a zero of f of infinite multiplicity.
If an 6= 0 for at least one n, and N is the smallest such n, then

+
f (z) = (z x)N an+N (z x)n , z B(x; r) .
n=0

We, now, define { f (z)

(zx)N
, if z \ {x}
g(z) =
aN 6= 0 , if z = x
22

and it is trivial to prove that g is holomorphic in . Therefore,

f (z) = (z x)N g(z) , z,

where g is holomorphic in and g(x) 6= 0. In this case we say that x is a zero

of f of finite multiplicity and the number N is called the multiplicity of x
as a zero of f and it is denoted by m(x; f ).
Of course, f (x) 6= 0 is equivalent to m(x; f ) = 0. Whenever we say that x is a
zero of f , we shall understand that f (x) = 0 or, equivalently, that m(x; f ) 1.
Using the series expansion, we, easily, see that the set of the zeros of finite
multiplicity and the set of zeros of infinite multiplicity are both open sets and,
hence, we get
Proposition 0.1 If f is holomorphic in the open connected , then either all
points of are zeros of f of finite multiplicity or f = 0 everywhere in .
We, also, see that if f is holomorphic in the open connected and is not
= 0 identically in , then every zero of f is isolated. Therefore,
Theorem 0.5 (Analytic Continuation Principle) If f is holomorphic in the
open connected and is not = 0 identically in , then every compact subset of
contains at most finitely many zeros of f .

Theorem 0.6 (Argument Principle) Let f be holomorphic in and the open

set 1 with C 1 -boundary be such that 1 . If 1 contains no zeros of f ,
then
1 f 0 (y)
dy = m(x; f ) .
2i 1 f (y)
x1

Proof:
By the Analytic Continuation Principle, 1 contains at most finitely many
zeros of f (which are all contained in 1 ), x1 , . . . , xN , and let mj = m(xj ; f )
be the corresponding multiplicities. We may, then, write

f (x) = (x x1 )m1 (x xN )mN g(x) , x,

where g is holomorphic in and has no zeros in 1 . Therefore,

f 0 (x)
N
mk g 0 (x)
= + , x \ {x1 , . . . , xN } \ {x : g(x) = 0} .
f (x) x xk g(x)
k=1

We consider small closed discs B(xj ; rj ) which are pairwise disjoint and are
all contained in 1 and the open set 2 = 1 \ N j=1 B(xj ; rj ).
f0
Then, f is holomorphic in an open set containing 2 , and, by Cauchys
Theorem,
N
1 f 0 (y) 1 f 0 (y) 1 f 0 (y)
0 = dy = dy dy .
2i 2 f (y) 2i 1 f (y) j=1
2i B(xj ;rj ) f (y)
0.3. HOLOMORPHIC FUNCTIONS 23

Thus,
N N
1 f 0 (y) 1 mk
dy = dy
2i 1 f (y) j=1
2i B(xj ;rj ) y xk
k=1

N
1 g 0 (y)
+ dy
j=1
2i B(xj ;rj ) g(y)

N
N
= mj = m(xj ; f ) = m(x; f ) ,
j=1 j=1 x1

from the example after Cauchys Formula and from Cauchys Theorem applied
0
to gg which is holomorphic in 1 .

If f (x) = w, then x is a zero of f w and its multiplicity is denoted by

m(x; f, w).
Theorem 0.7 Let f be holomorphic in and the open set 1 with C 1 -boundary
be such that 1 . If w0 , w00 C are contained in the same component of the
complement of the compact set f (1 ), then

m(x; f, w0 ) = m(x; f, w00 ) .
x1 x1

Proof:
Applying the Argument Principle to the function f w, we get

1 f 0 (y)
m(x; f, w) = dy , w/ f (1 ) .
2i 1 f (y) w
x1

The integral in the right side is a continuous function of w in the complement of

f (1 ) and it is integer-valued, as the left side shows. Therefore, this function
is constant in each component of the complement of f (1 ).

An important example of holomorphic function is the exponential func-

tion exp : C C \ {0} given by

exp(x) = ex1 (cos x2 + i sin x2 ) , x = x1 + ix2 C .

(Of course, there is no contradiction to use the notation ex , instead of exp(x),

and we shall very often do so.)
This satisfies the identity exp(x0 + x00 ) = exp(x) exp(x00 ) and it is periodic
with period i2. In fact, the only periods of exp are the numbers ik2, k Z.
A, perhaps, even more important function is the logarithmic function,
denoted by log, which is the many-valued inverse of exp. For each x C \ {0},
we have that log(x) = log |x| + i + ik2, where is any real number so that
x
ei = |x| and k takes all integer values.
24

To be strict, one has to talk about branches of the logarithmic function, as

follows. A function f in any A C \ {0} is called a branch of the logarithm
in A, if it is continuous in A and

exp(f (x)) = x , xA.

More generally, if g : A C \ {0} is continuous in A, then f : A C is

called a branch of the logarithm of g in A if it is continuous in A and

exp(f (x)) = g(x) , xA.

If is open and g is holomorphic in , then it is trivial to show that every

branch f of the logarithm of g in is holomorphic in and that
g 0 (x)
f 0 (x) = , x.
g(x)
It is, also, trivial to show, by continuity of the branches, that, if A is con-
nected, then any two branches of the logarithm of g in A differ by a constant of
the form ik2, k Z.
As the simplest example, if 0 = C \ {x = x1 + ix2 : x2 = 0, x1 0} then
the principal branch of the logarithm in 0 is log0 : 0 C given by

log0 (x) = log |x| + i ,

x
where is the unique real number so that cos + i sin = |x| and < < .
Therefore, the totality of branches of the logarithm in 0 are all functions of
the form log0 (x) + ik2, x 0 , where k runs in Z.
If = C \ {x = rei : r 0}, then the totality of branches of the
logarithm in are the functions of the form log0 (ei x) + i + ik2, x ,
where k runs in Z.
Here is a negative result.
Lemma 0.1 Let A and A+ be connected with A and A+ + .
0
If these two sets have a common point x0 = |x0 |ei with < 0 < + and
00 00 i 00
another common point x = |x |e with < 00 < , then there is no
branch of the logarithm in A = A A+ .
Proof:
Let f be a branch of the logarithm in A = A A+ .
Since A is connected, f (x) = log0 (ei x) + i + ik2 for all x A and
 0
for some k Z. Hence, f (x0 ) f (x00 ) = log xx00  + 0 00 .
Since A+ is connected, f (x) = log0 (ei(+) x) + i( + ) + il2 for all
 0
x A+ and for some l Z. Hence, f (x0 ) f (x00 ) = log xx00  + 0 00 2.
We, thus, get a contradiction.

For instance, there is no branch of the logarithm in C \ {0} or, even, in any
circle S(0; r).
Lemma 0.1 is used to prove a purely topological result.
0.4. EQUICONTINUITY 25

Theorem 0.8 Let the compact sets A and A+ be connected with A

and A+ + . Suppose, also, that these two sets have a common point
0 00
x0 = |x0 |ei with < 0 < + and another common point x00 = |x00 |ei with
00
< < .
Then, 0 (which is not in( A = A A) + ) does not belong to the unbounded
component of C \ A = C \ A A+ .
Proof:
Consider the functions ga : C \ {a} given by
ga (x) = x a , x C \ {a} .
If |a0 a| < d(a, A), then the existence of a branch of the logarithm of ga in
A implies the existence of a branch of the logarithm of ga0 in A. Indeed, if fa is
a branch of the logarithm of ga in A, then, for all x A,
aa0
( a a0 )
efa (x)+log0 (1+ xa ) = ga (x) 1 + = ga0 (x) .
xa
It is obvious now that, if |a0 a| < d(a0 , A), then the non-existence of a
branch of the logarithm of ga in A implies the non-existence of a branch of the
logarithm of ga0 in A.
Now, take any R so that A B(0; R) and any point x0 / B(0; R). Then,
x0 belongs to the unbounded component O of C \ A.
The set of points a O such that there exists a branch of the logarithm of
ga in A and the set of points a O such that there does not exist a branch of
the logarithm of ga in A are, by the previous discussion, both open sets and x0
is in the first set. By the connectedness of O, the second set is empty. Lemma
0.1, finally, implies that 0
/ O.

0.4 Equicontinuity
Let F be a family of complex-valued functions defined in a subset E of a metric
space (X, d).
The family is called bounded at x E, if supf F |f (x)| < + .
The family is called equicontinuous at x E, if for every  > 0 there is
= () > 0, so that for all f F it holds |f (y) f (x)| <  whenever y E
and d(y, x) < .
Observe that , in this definition, does not depend on f F .
The following is one of several versions of the Ascoli-Arzela Theorem.
Theorem 0.9 (Arzela and Ascoli) Suppose F is a family of functions defined
in some compact subset K of a metric space and let F be bounded and equicon-
tinuous at every point of K.
Then, from every sequence in F we can extract a subsequence which con-
verges uniformly in K to some function (not necessarily belonging to F).
In these notes we shall apply this result when the metric space is either Rn
with the Euclidean metric or Rn with the spherical metric.
26

0.5 Semi-continuity
Suppose that E is a subset of a metric space (X, d).
Definition 0.1 A function f is called lower-semicontinuous in E, if
1. < f (x) + for all x E and
2. f (x) lim inf E3yx f (y) for all x E
or, equivalently, if
1. < f (x) + for all x E and
2. {y E : < f (x)} is open relative to E for every real .
We call f upper-semicontinuous in E, if f is lower-semicontinuous
in E: in all relations above we just reverse the inequalities, replace by
and replace lim inf by lim sup.
Properties of semicontinuous functions
(1) f is continuous in E if and only if it is simultaneously lower- and upper-
semicontinuous in E.
(2) Linear combinations of lower-semicontinuous functions with non-negative
coefficients are lower-semicontinuous. This is true for upper-semicontinuous
functions, also.
The proofs are easy.
(3) The supremum of any family of lower-semicontinuous functions is lower-
semicontinuous. There is a dual statement for upper-semicontinuous functions.
Suppose that each f F is lower-semicontinuous in E and let F (x) =
supf F f (x) for every x E. Then, it is obvious that < F (x) for all x E
and, since f (y) F (y) for all y E and all f F , we find

f (x) lim inf f (y) lim inf F (y)

E3yx E3yx

for x E. Taking the supremum over all f F , we conclude

F (x) lim inf F (y) .

E3yx

(4) The minimum of finitely many lower-semicontinuous functions is lower-

semicontinuous. There is a dual statement for upper-semicontinuous functions.
It is enough to consider two lower-semicontinuous functions f1 , f2 and let
f = min(f1 , f2 ). Then, it is obvious that f (x) > for all x E.
Take < f (x) and, hence, < f1 (x) and < f2 (x). There exist 1 > 0 and
2 > 0 so that < f1 (y) for all y E with d(y, x) < 1 and < f2 (y) for all
y E with d(y, x) < 2 .
If = min(1 , 2 ), then < f (y) for all y E with d(y, x) < . Therefore,
f (x) lim inf E3yx f (y).
0.5. SEMI-CONTINUITY 27

The following four propositions state properties of semicontinuous functions

which we shall make constant use of in later chapters.
Proposition 0.2 If f is lower-semicontinuous in a compact set K, then f is
bounded from below in K and takes a minimum value in K.
There is an obvious dual statement for upper-semicontinuous functions.
Proof:
Let m = inf xK f (x) and take {xk } in K so that f (xk ) m. Replacing
{xk } by some subsequence, if necessary, we may assume that xk x for some
x K. But then,

m f (x) lim inf f (y) lim f (xk ) = m

E3yx k+

and we get that < m = f (x).

Another proof of the boundedness from below runs as follows. Since f does
not take the value ,

+
K {x K : k < f (x)} .
k=1

The terms of the union are open and increasing with k and, since K is
compact, K is contained in one of them.

The following is a partial converse of property (3) of lower-semicontinuous

functions.
Proposition 0.3 If f is lower-semicontinuous in a compact set K, then there
exists an increasing sequence of continuous functions {fk } in K which converges
to f pointwise in K.
There is the usual dual statement for upper-semicontinuous functions.
Proof:
If f = + identically in K, we consider fk = k identically in K.
Otherwise, we define

fk (x) = inf (f (y) + kd(x, y))

yK

for all x K. Since, by Proposition 0.2, both f and d(x, ) are bounded from
below in K, we have that fk (x) is a real number.
From the inequality
 
(f (y) + kd(x, y)) (f (y) + kd(x0 , y)) kd(x, x0 ) ,

we, easily, prove that

|fk (x) fk (x0 )| kd(x, x0 ) ,

implying that fk is continuous in K.

28

It is, also, clear that {fk } is increasing and that fk (x) f (x) for all k and
all x: just take y = x in the definition of fk (x).
Now, fix x K and < f (x).
By the lower-semicontinuity of f , there exists > 0 so that f (y) for all
y K with d(x, y) < .
If we take k large enough, then f (y) + kd(x, y) for all y K with
d(x, y) . In fact, by Proposition 0.2, minK f is finite and, then it is enough
to take k 1 ( minK f ).
Therefore, if k is large enough,

inf (f (y) + kd(x, y)) = fk (x) .

yK

This, together with fk (x) f (x), implies

lim fk (x) = f (x)

k+

for all x K.

Proposition 0.4 (First Minimum Principle) Let f be lower-semicontinuous

in a connected subset E of a metric space with the property that, if it has a
minimum value at some point, then it is constant in some open (relative to E)
neighborhood of the same point.
Then, if f takes a minimum value in E, it is constant in E.
There is a dual Maximum Principle for upper-semicontinuous functions.

Proof:
The assumptions imply that both sets {x E : inf E f = f (x)} and {x E :
inf E f < f (x)} are open relative to E. Therefore, since E is connected, one of
them is empty and the other is all of E.

Proposition 0.5 (Second Minimum Principle) Suppose that O is an open con-

nected subset of a metric space with compact closure O and with non-empty
boundary. Let f be lower-semicontinuous in O with the property that, if it has
a minimum value at some point, then it is constant in some open neighborhood
of the same point.

1. If f takes a minimum value in O, then it is constant in O.

( )
2. Let m = inf yO lim inf O3xy f (x) . Then m f (x) for all x in O.
If m = f (x) for some x O, then f = m identically in O.

There is a dual Maximum Principle for upper-semicontinuous functions.

Proof:
The first part is straightforward from the First Minimum Principle.
0.6. BOREL MEASURES 29

As for the second part, extend f in O, defining

f (y) = lim inf f (x)

O3xy

for all y O.
Then, it is easy to see that f becomes lower-semicontinuous in O and, from
Proposition 0.2, this extended f takes a minimum value in O, say m .
If m < m, then m is taken at a point of O and, by the first part, f must
be constant, f = m , in O. Therefore, taking any boundary point y,

m f (y) = lim inf f (x) = m

O3xy

and we get a contradiction.

Hence, m m .
If for some x O we have m = f (x), then, by the first part, f = m in O.

0.6 Borel Measures

1. The Borel -algebra B(Rn ) is the smallest -algebra of subsets of Rn
which contains all open sets. Its elements are called Borel sets.
If A is any Borel set, then we define the Borel -algebra of A by

B(A) = {B B(Rn ) : B A} .

A complex measure d on B(A) is called a complex Borel measure in A.

A non-negative measure d on B(A) is called a non-negative Borel mea-
sure in A if, additionally, d(K) < + for every compact K A.
A signed Borel measure in A is any difference of two non-negative Borel
measures in A at least one of which is finite.
2. A Borel measure d of any kind (complex, signed, non-negative) in Rn is
said to be supported in a Borel set A, if d(B) = 0 for all Borel sets B such
that B A = .
For any kind of Borel measure d in Rn and any Borel set A, we define the
restriction of d in A by

dA (B) = d(B A)

for all Borel sets B. This is a Borel measure supported in A.

One should observe and keep in mind the difference between a Borel measure
in a Borel set A and a Borel measure in Rn supported in the Borel set A.
3. Every signed Borel measure d has a non-negative variation d+ and
a non-positive variation d , which are both non-negative Borel measures,
they are supported in disjoint Borel sets, at least one of them is finite and satisfy
d = d+ d . Then, the non-negative Borel measure |d| = d+ + d is
called the absolute variation of d.
30

It is true that,
m for every signed Borel measure d and every Borel set A,
|d|(A) = sup k=1 |d(Bk )| over all m N and all partitions A = m
k=1 Bk of
A into pairwise disjoint Borel subsets of it.
We extend to the case of complex Borel measures and, for every complex
Borel measure d and every Borel set A, we define

m
|d|(A) = sup |d(Bk )|
k=1

over all m N and all partitions A = mk=1 Bk of A into pairwise disjoint Borel
subsets of it.
Then, |d| is a finite non-negative Borel measure and the finite number

kdk = |d|(Rn )

is called the total variation of d.

4. If f is integrable in Rn , then a complex Borel measure df is defined by

df (A) = f (x) dm(x)
A

for all Borel sets A. This measure is denoted, also, by

f dm = df .

The integrable function f is called the density function or the Radon-

Nikodym derivative with respect to dm of the complex Borel measure
f dm.
If f + = max(f, 0) and f = max(f, 0), whence f = f + f and |f | =
f + f , then
+

(f dm)+ = f + dm , (f dm) = f dm , |f dm| = |f | dm

and
kf dmk = |f (x)| dm(x) .
Rn

We, also, have that f dm is supported in the Borel set A if and only if f = 0
almost everywhere in Rn \ A.
The complex Borel measures of the form d = f dm are called absolutely
continuous and are characterized by the property:

d(A) = 0 for all Borel sets A with dm(A) = 0 .

For every x Rn we define the Dirac mass at x as the measure

{
1 , if x A
dx (A) =
0 , if x
/A
0.6. BOREL MEASURES 31

for all Borel sets A. This is not absolutely continuous and is supported in {x}.
5. The space M(Rn ) of all complex Borel measures in Rn is a linear space and
k k is a norm on it. Under this norm, M(Rn ) becomes a Banach space.
If the complex Borel measure d is supported in the Borel set A, then,
obviously, |d| is also supported in A and kdk = |d|(A). We, then, denote by
M(A) the space of all complex Borel measures supported in A. This space is
just a closed linear subspace of M(Rn ) and, hence, a Banach space itself.
6. If A is a subset of a metric space, then CB (A) is the Banach space of all
complex-valued functions continuous and bounded in A, with the norm

kf k = sup |f (x)| , f CB (A) .

xA

If A is compact, then CB (A) = C(A).

We shall need the following simple version of an important theorem.
Theorem 0.10 (Representation Theorem of J. Radon and F. Riesz) Let L be a
bounded linear functional on the Banach space C(K), where K is any compact
subset of Rn . I.e.
L(af + bg) = aL(f ) + bL(g)
for all a, b C and all f, g C(K) and

|L(f )| kf k

for all f C(K) and some not depending on f .

Then, there exists a unique complex Borel measure d supported in K so
that
L(f ) = f (x) d(x)
K
for all f C(K). We, also, have that

kdk = kLk = sup{|L(f )| : f C(K), kf k 1} .

If L is, also, non-negative, i.e.

L(f ) 0

for all f C(K) with f 0 everywhere in K, then d is non-negative.

If for every d M(K) we define the function

Ld (f ) = f (x) d(x) , f C(K) ,
K

then Ld is a continuous linear functional on the Banach space C(K), i.e. an

element of C(K) .
The content of the Representation Theorem of J. Radon and F. Riesz is that
the (obviously, linear) mapping

M(K) 3 d 7 Ld C(K)
32

is a bijective isometry.
7. If {dm } is a sequence of Borel measures, then we say that it converges
weakly on the compact set K to the Borel measure d, if for all f C(K),

f (x) dm (x) f (x) d(x) .
K K

In this case, by the Uniform Boundedness Theorem, it is true that there is

a constant M so that |d|(K) M and |dm |(K) M for all m.
As an application of the Banach-Alaoglou Theorem in the Banach space
C(K), we get that for any sequence {dm } of complex Borel measures supported
in K with kdm k M for some M < + not depending on m, there is some
subsequence converging weakly on K to some complex Borel measure supported
in K.
If the sequences {d1m } and {d2m } converge weakly on the compact subsets
K1 and K2 of Rn to d1 and d2 , respectively, then the product measures
d1m d2m converges weakly on K1 K2 to d1 d2 . To show this, consider
an arbitrary product f1 (x)f2 (y) of continuous functions in K1 and K2 . Then,
by the Theorem of Fubini, it is clear that

f1 (x)f2 (y) d1m d2m (x, y) f1 (x)f2 (y) d1 d2 (x, y) .
K1 K2 K1 K2

K2 , we use the Stone-Weierstrass

For the arbitrary continuous f (x, y) in K1
Theorem to approximate f by a polynomial x y uniformly in K1 K2 and
that kdm dm k kdm kkdm k M1 M2 < +, to prove
1 2 1 2

f (x, y) dm dm (x, y)
1 2
f (x, y) d1 d2 (x, y) .
K1 K2 K1 K2

8. All kinds of Borel measures d are regular. This means that, for every Borel
set B with finite d(B) and every  > 0, there is an open set U and a compact
set K so that K B U and |d|(U \ K) <  .
9. Let d be a non-negative Borel measure and let the extended-real-valued f
be defined in the Borel set A. We, then, define the upper integral of f in A
by

f (x) d(x) = inf (x) d(x) ,
A A

where the infimum is taken over all bounded from below lower-semicontinuous
in A with f everywhere in A. We, also, define the lower integral

f (x) d(x) = sup (x) d(x) ,
A A

where the supremum is taken over all bounded from above upper-semicontinuous
in A with f everywhere in A.
0.7. DISTRIBUTIONS 33

It is, then, true that

f (x) d(x) f (x) d(x)
A A

It is, also, true that f is d-integrable in A if and only if

< f (x) d(x) = f (x) d(x) < + ,
A A

and,
in this case, the common value of the upper and the lower integral is equal
to A f (x) d(x).

0.7 Distributions
The following is only a short exposition of very few elementery facts about
distributions with brief proofs of only those of them which are considered new
or not easy enough to be proved by the inexperienced reader.
If f is a measurable function in Rn , then x Rn is said to be a support-
point for f , if f is almost everywhere 0 in no neighborhood of x.
The support of a measurable function f : Rn C is the smallest closed
set in Rn outside of which f is almost everywhere 0 and it is defined by

supp(f ) = {x Rn : x is a support-point for f } .

Saying that the support of a measurable function is bounded is equivalent to

saying that it is a compact subset of Rn and equivalent to saying that the func-
tion vanishes almost everywhere outside a compact subset of Rn . In this case,
we say that the function is compactly supported or that it has compact
support.
In case f is continous in Rn , then it is easy to see that its support is the
smallest closed set in Rn outside of which f is everywhere 0.
If d is a Borel measure (of any kind), we say that the point x Rn is a
support-point of d, if d is the zero measure in no neighborhood of x. The
support or, sometimes called, closed-support of d is defined by

supp(d) = {x Rn : x is a support-point of d} .

It is easy to see, using the regularity of d, that supp(d) is the smallest

closed set in Rn outside of which d is the zero measure.

Example +
The finite non-negative Borel measure d = k=1 2k d k1 is supported in
the set { k1 : k N}, but supp(d) = { k1 : k N} {0} .

Whenever we say that d is compactly supported or that it has compact

support, we mean that supp(d) is a compact subset of Rn .
34

In case d = f dm is an absolutely continuous compex Borel measure with

density function f L1 (Rn ), then, clearly,

supp(f dm) = supp(f ) .

Let f, g L1 (Rn ). Then, the convolution

f g(x) = f (x y)g(y) dm(y)
Rn

is defined for almost every x and it is true that f g L1 (Rn ). Also,

kf gkL1 (Rn ) kf kL1 (Rn ) kgkL1 (Rn ) .

It is true that

f g = gf , f (g h) = (f g) h

and
supp(f g) supp(f ) + supp(g) ,
where A + B = {x + y : x A, y B} for A, B Rn .
The convolution is defined in other instances also.
Let f be measurable in Rn and integrable in all compact subsets of Rn .
We, then, say that f is locally integrable in Rn and the space of all such f
is denoted by L1loc (Rn ).
If f is locally integrable and g is integrable and compactly supported, then
f g(x) is, obviously, well-defined for almost every x.
The convolution is a useful device and one reason is that it preserves the
regularity properties of its component functions.
For instance,
Proposition 0.6 Suppose f is locally integrable and g is continuous and com-
pactly supported. Then f g is everywhere defined and continuous.
If, moreover, g is in C k (Rn ), then f g is, also, in C k (Rn ) and

D (f g) = f D g

for all multi-indices = (1 , . . . , n ) with order || k.

The same conclusions hold, if we assume that f is locally integrable and
compactly supported and g is continuous or in C k (Rn ).
Proof:
(i) Let supp(g) be contained in the compact K Rn . Then, for every x,

|f (x y)g(y)| dm(y) kgk |f (y)| dm(y) < + ,
Rn xK

since x K = {x y : y K} is compact.
(ii) For fixed x and for |x0 x| 1, the function g(x0 ) g(x ) has
its support contained in the compact subset L = {y : d(y, x K) 1} of Rn .
0.7. DISTRIBUTIONS 35

Therefore,

|f g(x0 ) f g(x)| |f (y)||g(x0 y) g(x y)| dm(y)
L

sup |g(a) g(b)| |f (y)| dm(y)
|ab| L

and, since g is uniformly continuous, the last quantity tends to 0 as 0.

Thus, f g is continuous at x.
(iii) Now, suppose that g C 1 (Rn ). Then, as before, for fixed x and for h R
with |h| 1,
 f g(x + he ) f g(x) 
 j g 
 f (y) (x y) dm(y)
h Rn xj
 g(x + he y) g(x y) 
 j g 
|f (y)| (x y) dm(y)
L h x j
 g 
 g 
sup  (a) (b) |f (y)| dm(y)
|ab| xj xj L

and, since g
xj is uniformly continuous, the last quantity tends to 0 as 0.
(f g)
xj (x) = f xj (x) for every x
and, from part (ii), f g C 1 (Rn ).
g
Thus,
Using induction, we extend to derivatives of higher order.
The proof of the second part is, after trivial modifications, similar to the
proof of the first part.

A useful technical tool is the function

{
exp( 1t ) , if t > 0
0 (t) =
0, if t 0 .

It is easy to prove that 0 is in C (R) and supp(0 ) = R+

0 . Therefore, if
C > 0, the function
{ 1
C exp( 1|x| 2) , if x B(0; 1)
(x) =
0, if x
/ B(0; 1)

has the properties:

1. is in C (Rn ) ,

2. is non-negative and supp() = B(0; 1) ,

3. is radial. I.e. (x) = (y) whenever |x| = |y| ,

4. Rn (x) dm(x) = 1, if we choose the constant C appropriately.
36

Now, the functions ( > 0)

1 x
(x) = ( ) , x Rn ,
n
have the same properties 1, 3 and 4 and property 2 replaced by

supp( ) = B(0; ) .

Definition 0.2 The family of functions { : > 0}, or any other similar
family coming from any with properties 1-4, is called an approximation to
the identity.
Starting with any radial function F , the following function is well-defined
for all r R+
0 for which there is at least one x in the domain of definition of F
with |x| = r :

F (r) = F (x) , for any x with r = |x| .

This is useful whenever we use polar coordinates to evaluate integrals.

Let f be measurable in , an open subset of Rn , and integrable in all
compact subsets of . We call f locally integrable in and the space of all
such f is denoted by L1loc ().

Proposition 0.7 If f is in L1loc (), g is in C k (Rn ) and supp(g) B(0; ),

then the convolution

f g(x) = f (x y)g(y) dm(y)
Rn

is defined for all x in the open set

= {x : d(x, ) > }

and it is in C k ( ). Moreover,

D (f g) = f D g

in , for all with || k.

Proof:
A minor modification of the proof of Proposition 0.6.

Definition 0.3 Let be an open subset of Rn . Then the space of test-

functions in , denoted by D(), is the space of all infinitely differentiable
complex-valued functions with compact support contained in .
D() is a linear space and f D() for every f C () and every
D(). In fact, since C () is an algebra, we see that D() is an ideal in
C ().
The next result is technically very helpful.
0.7. DISTRIBUTIONS 37

Lemma 0.2 Suppose that is an open subset of Rn and K is a compact subset

of . Then, there exists a D() so that

1. 0 1 everywhere in and

2. = 1 everywhere in K.

Proof:
1
Consider 0 < 4 d(K, ) and the compact set

K20 = {x : d(x, K) 20 } .

We also consider any approximation to the identity { : > 0} and the

convolution
= K20 0 ,
where K20 is the characteristic function of K20 .
By Proposition 0.6, C (Rn ) and supp() K20 + B(0; 0 ) which is a
compact subset of . Hence, D().
Also, for every x,

0 (x) = 0 (x y) dm(y) 0 (x y) dm(y) = 1 .
K20 Rn

Finally, if x K, then B(x; 0 ) K20 and, thus,

(x) = K20 (y)0 (xy) dm(y) = 0 (xy) dm(y) = 1 .
B(x;0 ) B(x;0 )

We define a notion of convergence for sequences in D() as follows.

Definition 0.4 Let {m } and be in D(). We write

m in D() ,

if

1. there exists some compact K so that supp(m ) K for all m and

2. D m D uniformly in for every multi-index .

Of course, 1 and 2 imply that supp() K .

Proposition 0.8 1. If m and m in D(), then m + m

+ and m m in D() for all , C .

2. If m in D() and f C (), then f m f in D() .

38

Proof:
If the supports of all m are contained in the compact K and the supports
of all m are contained in the compact L, then the supports of all m + m
are contained in the compact K L, the supports of all m m are contained in
the compact K L and the supports of all f m are contained in K.
The rest is an easy application of the rule of Leibniz for the derivatives of
products.

Proposition 0.9 For every in D() and any approximation to the identity
{ : > 0} we have that belongs to D() for small enough and

in D()

as 0+ .
Proof:
Observe that , defined to be equal to 0 outside , is in C (Rn ) . Therefore,
from Proposition 0.6,
D ( ) = D
in Rn .
If we consider K = supp() and 0 = 12 d(K, ) > 0, then, for all 0 ,
the support of is contained in the compact set L = K + B(0; 0 ) = {x :
d(x, K) 0 } .
Hence, belongs to D() for all 0 .
Now, for every x,
( )
|D (x) D (x)| = |D (x) D (x)|

|D (x y) D (x)| (y) dm(y)
Rn

= |D (x y) D (x)|1 (y) dm(y)

Rn
sup |D (a) D (b)|
|ab|

) 0 as 0 .
and the last term (tends to
Therefore, D D uniformly in Rn .

This notion of convergence defines a topology in the space D(), which then
becomes a locally convex topological vector space, but we shall not touch and
not need this more abstract matter.
We call distribution in any continuous linear functional T on D(). Thus,
Definition 0.5 A function

T : D() C

is called distribution in , if
0.7. DISTRIBUTIONS 39

1. T ( + ) = T () + T () for all , D() and all , C .

2. T (m ) T () whenever m in D() .

We denote by D () the set of all distributions in .
The space D () becomes a linear space, when we define

(T + S)() = T () + S()

for all D(). In fact, it is trivial to prove that T + S is a distribution in

whenever T and S are distributions in .
We can, also, define the notion of (w -)convergence in D () as follows.
Definition 0.6 We write

Tm T in D () ,

if
Tm () T ()
for all D() .
The two examples which will be introduced in Definitions 0.7 and 0.10 show
how certain concrete objects, like functions and measures, can be viewed as
distributions.
Definition 0.7 For every f L1loc (), we define

Tf () = (x)f (x) dm(x)

for all D().

It is trivial to prove that Tf is a distribution in .
The (obviously) linear map

L1loc () 3 f 7 Tf D ()

is a continuous injection.
This means that,
1. if Tf = Tg , then f = g almost everywhere in and

2. Tfm
Tf in D (), whenever fm f in Lloc () or, equivalently, when-
1

ever K |fm (x) f (x)| dm(x) 0 for every compact K .

The second statement is trivial to prove, and the first is equivalent to
Proposition 0.10 If f and g are locally integrable in and if

f (x)(x) dm(x) = g(x)(x) dm(x)

for all D(), then f = g almost everywhere in .

40

Proof:
If K is a compact subset of , consider Um = {x : d(x, K) < m
1
}.
Then, K = m=1 Um , implying dm(Um \ K) <  for large m.
+

By Lemma 0.2, there exists a D(Um ) so that

1. 0 1 everywhere and

2. = 1 in K.

Then,
 
 
 f (x)(x) dm(x) f (x) dm(x) |f (x)| dm(x) ,
K Um \K

with a similar inequality for g.

Letting  0 and using the hypothesis, we get

f (x) dm(x) = g(x) dm(x)
K K

for all K.
By the regularity of dm, the proof is complete.

We may, thus, identify locally integrable functions in with the corre-

sponding distributions in .

Definition 0.8 If for some distribution T there exists a (necessarily unique)

locally integrable f so that T = Tf , then we say that T is represented by f or
that T is identified with f .

Because of this identification, it is common practice to use the symbol f for

both the function and the distribution represented by it. We, thus, write

f () = (x)f (x) dm(x) .

Therefore, the symbol f has two meanings; the function meaning, when
f acts on points of , and the functional meaning, when f acts on functions
D().
Let d be a complex-valued function defined on the union of B(K) for all
compact K (where B(K) is the Borel -algebra of K), so that it is a
complex Borel measure in every such compact K.
To say the same thing in a different way,

Definition 0.9 Let d(A) be defined as a complex number for every Borel set
A contained in a compact subset of and let
(+
)
+
d Ak = d(Ak ) ,
k=1 k=1
0.7. DISTRIBUTIONS 41

whenever the Borel sets Ak are pairwise disjoint and are all contained in the
same compact subset of .
Every such d is called a locally finite complex Borel measure in and
the set of all such d is denoted Mloc ().

If d is a non-negative Borel measure in , then d(A) is defined for all

Borel subsets A of and, by definition, it is a (finite) non-negative number
whenever A is contained in a compact subset of . Thus, by just restricting its
domain of definition, d becomes a locally finite complex Borel measure in
with non-negative values.
The converse is the content of the next result.

Proposition 0.11 Suppose that d is a locally finite complex Borel measure in

with non-negative values.
Then, the domain of definition of d can be extended to B() so that d
becomes a non-negative Borel measure in .
The extension of d on B() is unique.

Proof:
If A is an arbitrary Borel subset of , define

d0 (A) = sup d(B)

over all Borel sets B A which are contained in compact subsets of .

(i) It is obvious that, if A itself is contained in a compact subset of , then
d0 (A) = d(A) and, thus, d0 is an extension of d.
(ii) Now, let the sets Ak , k N, be pairwise disjoint Borel subsets of .
If B +
k=1 Ak is contained in a compact subset of , then all B Ak are
contained in the same compact subset of and, thus,

+
+
d(B) = d(B Ak ) d0 (Ak ) .
k=1 k=1

This implies that

(+
)
+
d0 Ak d0 (Ak ) .
k=1 k=1

Now, for each k with 1 k m, consider arbitrary Borel sets Bk Ak con-

tained in compact subsets mof . Then m
k=1 Bk is, also, contained in a compact
k=1 Bk ) d0 (k=1
+
subset of and, hence, k=1 d(Bk ) = d(m mAk ).
Taking the supremum over each Bk independently, we find k=1 d0 (Ak )
k=1 Ak ) and, letting m +,
d0 (+

+ (+
)
d0 (Ak ) d0 Ak .
k=1 k=1
42

Therefore, d0 is a non-negative Borel measure in .

(iii) Now, let d1 be another extension of d and consider a compact exhaustion
{K(m) } of . Then, for every Borel subset A of , the sets Bm = K(m) A
increase towards A and each is contained in a compact subset of . Therefore,
d1 (A) = lim d1 (Bm ) = lim d(Bm ) = lim d0 (Bm ) = d0 (A).
m+ m+ m+

Because of the last proposition we shall never distinguish between non-

negative Borel measures in and locally finite Borel measures in with non-
negative values.
If d is a locally finite complex Borel measure in , then through the ex-
tension procedure of Proposition 0.11, we can define the non-negative Borel
measures |d|, d+ and d in satisfying |d|(A) = d+ (A) + d (A) for
all A B() and d(A) = d+ (A) d (A) for all A B() contained in a
compact subset of .
If f is a locally integrable function in , then a locally finite complex Borel
measure f dm in is defined, as usual, by

f dm(A) = f (x) dm(x)
A

for all Borel sets A which are contained in compact subsets of .

It is obvious that we may integrate any continuous function with compact
support contained in against any locally finite complex Borel measure in .
Definition 0.10 For every d Mloc () we define

Td () = (x) d(x)

for all D().

It is clear that Td is a distribution in and that the linear map
Mloc () 3 d 7 Td D ()
is a continuous injection.
This means that,
1. if Td = Td , then d = d and
2. Tdm Td in D (), whenever dm d in Mloc () or, equivalently,
whenever |dm d|(K) 0 for every compact K .
The first statement is the same as
Proposition 0.12 If d and d are two locally finite complex Borel measures
in and
(x) d(x) = (x) d(x)

for all D(), then d = d.
0.7. DISTRIBUTIONS 43

Proof:
The proof is identical to the proof of Proposition 0.10, replacing f (x)dm(x)
by d(x) and g(x)dm(x) by d(x) and using the regularity of Borel measures.

Therefore, we identify locally finite complex Borel measures in with the

corresponding distributions in and, as above for functions,

Definition 0.11 If for a distribution T in there exists some (necessarily

unique) locally finite complex Borel measure d in so that T = Td , then we
say that T is represented by d or that T is identified with d.

As in the case of functions and because of this identification between mea-

sures and distributions, it is common practice to use the same symbol d for
both the measure and the distribution represented by it. We, therefore, very
often write
d() = (x) d(x) .

Hence, the symbol d has two meanings; one is the meaning of a measure,
when d acts on certain Borel subsets of , and the other is the functional
meaning, when d acts on functions D().
A particular case is when the measure is the da , the Dirac-mass at a point
a . Then,

da () = Tda () = (x) da (x) = (a)

for all D().

Another particular case is when the measure is of the form f dm for some
locally integrable function f in . Then,

f dm() = Tf dm () = (x)f (x) dm(x) = Tf () = f ()

for all D(), implying Tf dm = Tf or, informally, f dm = f .

Definition 0.12 A distribution T in is called non-negative and we write

T 0,

if
T () 0
for all D() with 0 everywhere in .

Theorem 0.11 If T is a non-negative distribution in , then there exists a

non-negative Borel measure d in so that T = Td .

Proof:
Consider any open exhaustion {(m) } of and fix m.
44

By Lemma 0.2, there exists 0 D() such that 0 0 1 everywhere

and 0 = 1 in the compact (m) .
Then, for every D((m) ), we have || kk 0 everywhere in and,
since T 0,
|T ()| |T (<)| + |T (=)| 2T (0 )kk .
Therefore, T is a bounded linear functional on the linear subspace D((m) ) of
the normed space C((m) ). By the Hahn-Banach Theorem, T can be extended
as a bounded linear functional on C((m) ).
By the Representation Theorem of J. Radon and F. Riesz, there exists a
complex Borel measure dm supported in (m) so that

T () = (x) dm (x)
(m)

for all D((m) ).

If K is any compact subset of (m) , then, as in the proof of Proposition
0.10,( there exists ) D((m) ) so that 0 1 in (m) , = 1 in K and
dm supp() \ K < . This implies

|dm (K) T ()| |K (x) (x)| dm (x)  ,
(m)

and, since T () 0 and  is arbitrary,

dm (K) 0 .
By the regularity
of dm , we conclude thatit is a non-negative measure.
Now, since (m) (x) dm (x) = T () = (m+1) (x) dm+1 (x) for every
D((m) ), Proposition 0.12 implies that
dm = dm+1
in (m) . Therefore, the measures dm define a locally finite complex Borel
measure d in with non-negative values so that the restriction of d in each
(m) coincides with dm ; simply, define, for all Borel sets A which are contained
in compact subsets of ,
d(A) = dm (A)
for any m with A (m) .
Finally, if D(), then, since supp() is compact, D((m) ) for large
m and

Td () = (x) d(x) = (x) dm (x) = T () .
(m)

In the rest of this section we define two operations on distributions, the

differentiation and the convolution with infinitely differentiable functions with
compact support in Rn , and study their calculus.
0.7. DISTRIBUTIONS 45

Definition 0.13 If T is a distribution in , we define

T ( )
() = T
xj xj

for all D() and call T

xj the j-derivative of T .

T
Then, x j
is a new distribution in , as one can, easily, see.
Inductively, we see that a distribution in has derivatives of all orders and
these are, also, distributions in . In fact, for every multi-index ,

D T () = (1)|| T (D ) .

Definition 0.14 If Tf is the distribution identified with the locally integrable

T
function f , then the distribution xfj is called the distributional j-derivative
of f .

By Definitions 0.7 and 0.13, this means nothing more than that this distri-
butional j-derivative is a functional on D(), defined by
( )
Tf
() = Tf = f (x) (x) dm(x) .
xj xj xj

Therefore, a locally integrable function has distributional derivatives of all

orders which are, in general, distributions.
To avoid a conflict, we must find the relation between the distributional
derivative and the classical derivative, whenever this last one exists. Therefore,
f
assume that f is in C 1 () and let x j
be its usual j-derivative in . Then, for
every D(),

Tf f
() = f (x) (x) dm(x) = (x)(x) dm(x) = T f () .
xj xj xj xj

The second equality is the well-known Integration by Parts formula and

can be proved using Greens Formula in a bounded open set G with C 1 -boundary
and with supp() G G (so that and vanish in G).
Thus,

Proposition 0.13 The distributional j-derivative of any f in C 1 () is identi-

fied with its classical j-derivative.

Based on this last result and on the identification between f and Tf , we,
f T
often, use the customary symbol x j
to denote xfj , even though f may not be
differentiable in the classical sense. Thus, the informal notation is

f
() = f (x) (x) dm(x) .
xj x j
46

If T D (), then

2T T ( ) ( 2 )
() = = T .
x2j xj xj x2j

Therefore, the Laplacian of a distribution T is given by

T () = T ()

for all D().

Another case which we shall often consider is the distributional Laplacian
f of a locally integrable function. The correct notation is, of course, Tf and
it is the distribution which, for every D(), satisfies

Tf () = Tf () = f (x)(x) dm(x) .

The informal notation is, of course,

f () = f (x)(x) dm(x) ,

even if f is not differentiable in the classical sense.

In case f C k (), using the formula of Integration by Parts,

D Tf () = (1)|| Tf (D ) = (1)|| f (x)D (x) dm(x)

= D f (x)(x) dm(x) = TD f ()

for all D() and all multi-indices of order at most k. Thus,

Proposition 0.14 The distributional -derivative of an f in C k () is identi-

fied with its classical -derivative, for all with || k.

Again, the informal notation, even if f is not differentiable, is

D f () = (1)|| f (x)D (x) dm(x) .

Let > 0 and consider any f C (Rn ) with supp(f ) B(0; ). Denote

fe(x) = f (x) , x Rn .

Definition 0.15 For any T D () and any D( ), we define

T f () = T (fe ) .

T f is called the convolution of the distribution T and the function f .

0.7. DISTRIBUTIONS 47

This is well-defined. Indeed, fe C (Rn ) and supp(fe) B(0; )+

and, hence, fe D().
Proposition 0.15 If T D () and f C (Rn ) with supp(f ) B(0; ),
then T f is a distribution in .
Proof:
The linearity is trivial to prove and we consider m in D( ).
If all supp(m ) are contained in the compact K , then all supp(fe m )
are contained in the compact L = K + B(0; ) . Moreover,

|D (fe m )(x) D (fe )(x)|

|D m (y) D (y)||fe(x y)| dm(y)
Rn

sup |D m (y) D (y)|

|f (z)| dm(z) .
yRn Rn

and, hence, D (fe m ) D (fe ) uniformly in Rn .

This implies that fe m fe in D() and, since T D (), we, finally,
get
T f (m ) = T (fe m ) T (fe ) = T f () .

Guided by the principle that convolution preserves the regularity of its

components, we shall prove that T f can be identified with some function
g C ( ).
Remark
In the proof that follows there are four whys and hows. The interested
reader is advised to answer them, providing the easy but technical proofs.

Define f x (y) = f (x y) and observe that f x D(), if x . Therefore,

the function
g(x) = T (f x ) , x ,
is well-defined.
If xm x, then f xm f x in D() (why?) and, hence, g(xm ) g(x).
Therefore, g is continuous in .
Next, we see easily (how?) that, for all x , if h 0 in R, then

f x+ej h f x ( f )x

h xj

in D(), where ej is the j-coordinate unit vector.

From this,

g(x + ej h) g(x) ( f x+ej h f x ) (( f )x )

lim = lim T = T .
h0 h h0 h xj
48

Hence,
g (( f )x )
(x) = T , x ,
xj xj
implying that g has continuous partial derivatives of first order and, by induc-
tion, of any order:
g C ( ) .
Take, now, any D( ).
Since the function g is continuous with compact support contained in ,

(x)g(x) dm(x) = lim (xk )g(xk )m(xk ) ,

where the last sum is a Riemann-sum defined, as usual (how?), by a mesh of

small cubes and the limit is taken as the size of the cubes tends to 0.
Now, it is easy to prove (how?) that, also,

(x)f x () dm(x) = lim (xk )f xk ()m(xk ) in D() .

Finally,

(x)g(x) dm(x) = lim (xk )g(xk )m(xk )


= lim (xk )T (f xk )m(xk )
( )
= lim T (xk )f xk ()m(xk )
( )
= T (x)f x () dm(x)

= T (fe ) = T f () ,

where the third equality holds because the sum is finite.

We conclude that T f is identified with the function g C ( ). Thus,

Proposition 0.16 The convolution of a distribution in and an infinitely dif-

ferentiable function supported in B(0; ) is a distribution in identified with
a function in C ( ).

We can easily prove

Proposition 0.17
T (f g) = (T f ) g
for all T D () and f, g C (Rn ) with supp(f ) B(0; ) and supp(g)

B(0; ) .
In this equality, both sides are distributions in + .
0.7. DISTRIBUTIONS 49

Proof:
For any D(+ ), and since supp(f g) B(0; + ) ,
( ) ( ) ( )
T (f g) () = T (fg g) = T fe (eg )
( )
= (T f )(eg ) = (T f ) g () .

Proposition 0.18 For all T D () and f C (Rn ) with supp(f )

B(0; ),
D (T f ) = D T f = T D f ,
as distributions in .
Proof:
For every D( ),

D (T f )() = (1)|| (T f )(D ) = (1)|| T (fe D ) .

This is, on one hand, equal to

( )
(1)|| T D (fe ) = D T (fe ) = (D T f )()

and, on the other hand, equal to

g
T (D f ) = (T D f )() .

Proposition 0.19 If h L1loc () and f C (Rn ) with supp(f ) B(0; ),

then
Th f = Thf
as distributions in .
Proof:
By Proposition 0.7, the function h f is in C ( ) and, hence, both sides
are distributions in . If D( ) and supp() is contained in the compact
K , then supp(fe ) is contained in the compact L = K + B(0; )
and, using the Theorem of Fubini,

e
(Th f )() = Th (f ) = h(x) f (y x)(y) dm(y) dm(x)
K

= h(x) f (y x)(y) dm(y) dm(x)

L K

= (h f )(y)(y) dm(y) = Thf () .

K
50

Proposition 0.20 Consider any approximation to the identity { : > 0},

let T be a distribution in and fix an arbitrary  > 0.
Then, for all , T is a distribution in  and

T T in D ( )

as 0.
Proof:
Take any D( ) and write

f ) = T (()
(T )() = T ( e ) T () ,

e in D() as 0.
since, by Proposition 0.9, ()

0.8 Concavity
An extended-real-valued function f defined in an interval I [, +] is
called concave in I, if it takes only real values in the interior of I and

f (ta + (1 t)b) tf (a) + (1 t)f (b)

for all a, b I and all t (0, 1).

f is called convex in I, if f is concave in I.
If f is concave in the interval I, then f is continuous at every interior point
of I and, at each of the endpoints, the one-sided limit of f exists. In case the
endpoint is contained in I, this limit is no smaller than the value of f there.
If f is concave in the open interval I, then f is differentiable at all points
of a subset D of I, where I \ D is at most countable. At every a I \ D
0 0 0
the one-sided derivatives of f exist and f (a) f+ (a). The functions f and
0
f+ are decreasing in I and, hence, f 0 is decreasing in D. The corresponding
Lebesgue-Stieltjes measure df 0 is a non-positive Borel measure in I, which is the
zero measure in I if and only if f is a linear function in I.
Also, at every interior point a0 of I, f has a line supporting its graph from
above. This means that for some l, k R,

f (a) la + k

for every a I, and

f (a0 ) = la0 + k .
As l we may consider any number between the one-sided derivatives of f at a0 .
If f is twice continuously differentiable in the open interval I, then f is
concave in I if and only if f 00 0 everywhere in I.
Theorem 0.12 (Inequality of Jensen) Let d be a non-negative measure in the
measure space (X, A) with d(X) = 1 and be an extended-real-valued function
in X which is in L1 (d).
0.9. THE FOURIER TRANSFORM 51

Then, for every f which is concave in an interval containing the values (x)
for almost all (with respect to d) x X,
( )
f (x) d(x) f (x) d(x) .
X X

The inequality is reversed, if f is convex instead of concave.

Proof:
Let I be an interval such that (x) I for almost all (with respect to d)
x X.
If the real number a0 = X (x) d(x) is an endpoint of I, then is almost
everywhere (with respect to d) equal to a0 and Jensens inequality is trivial to
prove.
Now, suppose that a0 = X (x) d(x) is an interior point of I.
( la
Then, there exist l, k so that f (a0 ) = la0 + k and f (a) ) + k for all a I.
Hence, for almost all (with respect to d) x X, f (x) l(x) + k and,
integrating,
( )
f (x) d(x) l (x) d(x) + k = f (x) d(x) .
X X X

0.9 The Fourier Transform

In this section we shall state and prove only a few elementary properties of the
Fourier transform of functions and measures which we shall need later on in
these notes.
Definition 0.16 If f L1 (Rn ), we define

fb() = e2ix f (x) dm(x)
Rn

for all Rn .
It is clear that the Lebesgue integral of the definition is well-defined as a
complex number.
Definition 0.17 If f L1 (Rn ), then the function

fb : Rn C

is called the Fourier transform of f .

Example
If
f (x) = e|x| ,
2
x Rn ,
for any > 0, then
( ) n2 2
fb() = e ||2
, Rn .

52

For the proof we use the Theorem of Fubini to reduce the calculation to the
case of dimension n = 1 and, then,
+
fb() = e2ix ex dm(x) .
2

Taking derivative and using integration by parts we may, easily, prove that

fb 2 2 b
() = f () , R.

Therefore,

(b 2 2
) 2 2 b 2 2 2 2 2 2
f ()e = f ()e + fb() e = 0 , R.


Since fb(0) = , we find

2 2
fb() = e .

The proof of the next result is trivial.

Proposition 0.21 If f L1 (Rn ), then

sup |fb()| kf kL1 (Rn )

Rn

and, thus,
fb L (Rn ) .
Moreover, for every f1 and f2 in L1 (Rn ) and all 1 and 2 in C,

(1 f1 + 2 f2 )b = 1 fb1 + 2 fb2 .

Definition 0.18 The linear operator

F : L1 (Rn ) L (Rn )

defined by
F(f ) = fb
is called the Fourier transform on L1 (Rn ).

Proposition 0.22 The norm of the linear operator F : L1 (Rn ) L (Rn ) is

equal to 1.

Proof:
The norm is, by Proposition 0.21, at most 1.
0.9. THE FOURIER TRANSFORM 53

Taking any non-negative f L1 (Rn ), we have, for all Rn ,

b
|f ()| |f (x)| dm(x) = f (x) dm(x) = fb(0) .
Rn Rn

Therefore,
sup |fb()| = kf kL1 (Rn ) ,
Rn

implying that the norm of the Fourier transform on L1 (Rn ) is equal to 1.

Theorem 0.13 (The Lemma of Riemann and Lebesgue) If f L1 (Rn ), then

the function fb is continuous in Rn and

lim fb() = 0 .

Proof:
If k , then

( 2i x )
fb(k ) fb() = e k
e2ix f (x) dm(x) 0
Rn

by the Dominated Convergence Theorem and this proves the continuity of fb.
(i) If
f = I
is the characteristic function of an n-dimensional interval

I = [a1 , b1 ] [an , bn ] ,

then, for every = (1 , . . . , n ), we calculate

n
e2ij bj e2ij aj
cI () =
.
j=1
2ij

We, now, easily, show that

cI () = 0 .
lim

This property extends, by the linearity in Proposition 0.21, to all linear

combinations
m
f = k Ik .
k=1

(ii) If f L (R ) and  > 0 is arbitrary, we take

1 n
a linear combination of
m
characteristic functions of n-dimensional intervals g = k=1 k Ik so that

kf gkL1 (Rn )  .
54

Then, from Proposition 0.21 and the result of (i),

lim sup |fb()| lim sup |fb() gb()| + lim sup |b

g ()|

kf gkL1 (Rn )
.

Since  is arbitrary, the proof is complete.

Proposition 0.23 If f and g are in L1 (Rn ), then

fd
g = fbgb .

Proof:
In fact, by the Theorem of Fubini,

d
f g() = e 2ix
f (x y)g(y) dm(y) dm(x)
R
n Rn

= e2i(xy) f (x y) dm(x) e2iy g(y) dm(y)

R R
n n

= fb()e2iy g(y) dm(y)

Rn
= fb()b
g () .

Definition 0.19 If d is a complex Borel measure in Rn , then we define

c
d() = e2ix d(x)
Rn

for all Rn .
The function
c : Rn C
d
is called the Fourier transform of d M(Rn ).

Proposition 0.24 If d M(Rn ), then

c
sup |d()| kdk
Rn

and, thus,
c L (Rn ) .
d
Moreover, for all d1 and d2 in M(Rn ) and all 1 and 2 in C,

d1 + 2 d
(1 d1 + 2 d2 )b = 1 d d2 .

The proof is clear.

0.9. THE FOURIER TRANSFORM 55

Definition 0.20 The linear operator

F : M(Rn ) L (Rn ) ,

defined by
c,
F(d) = d
is called the Fourier transform on M(Rn ).

c is continuous in Rn .
Proposition 0.25 If d M(Rn ), then d

Proof:
The proof, based on the Dominated Convergence Theorem, is identical to
the proof of the first part of Theorem 0.13.

The second part of Theorem 0.13 is, in general, not true for measures.

Example
da () = e2ia
d
for all Rn and, hence,
da ()| = 1
|d
for all Rn .

Example
For every absolutely continuous measure f dm with density function f
L1 (Rn ), it is clear that
fd
dm = fb .

Proposition 0.26 The Fourier transform on M(Rn ) has norm equal to 1.

Proof:
The proof is identical to the proof of Proposition 0.22 and uses any non-
negative Borel measure in M(Rn ).

Definition 0.21 For every d and d in M(Rn ) and every Borel set A, we
define
d d(A) = A (x + y) d(x) d(y) ,
Rn Rn

where d d is the product measure and the double integral is well-defined as

a complex number by the Theorem of Fubini.
d d is called the convolution of the measures d and d.

The proof of the next result is easy and is based on the definition of total
variation.
56

Proposition 0.27 If d and d are in M(Rn ), then d d is -additive and,

thus,
d d M(Rn ) .
Moreover,
kd dk kdkkdk .

Proposition 0.28 If d, d and d are in M(Rn ), then

d d = d d

and
(d d) d = d (d d) .

Proof:
By the Theorem of Fubini, for every Borel set A,

d d(A) = A (x + y) d(x) d(y)
Rn Rn

= Ay (x) d(x) d(y)
R R
n n

= d(A y) d(y)
R
n

= d(A x) d(x)

Rn

= Ax (y) d(y) d(x)

R R
n n

= Ax (y) d(x) d(y)

R R
n n

= A (x + y) d(x) d(y)
Rn Rn
= d d(A) .

The proof of the second identity is similar and, even, easier.

Definition 0.22 If f L1 (Rn ) and d M(Rn ), we define

f d(x) = f (x y) d(y) , x Rn ,
Rn

and call the function f d the convolution of the function f and the measure
d.

It is trivial to prove that f d is in L1 (Rn ) and that

kf dkL1 (Rn ) kf kL1 (Rn ) kdk .

0.9. THE FOURIER TRANSFORM 57

Proposition 0.29 Suppose that d and d are in M(Rn ) and that d = f dm

is absolutely continuous with density function f . Then, d d is, also, abso-
lutely continuous and its density function is

f d() = f ( y) d(y) .
Rn

Proof:
For every Borel set A,

d d(A) = d(A y) d(y)
R
n

= f (x) dm(x) d(y)

Rn Ay

= f (x y) dm(x) d(y)
R A
n

= f (x y) d(y) dm(x) .
A Rn

Proposition 0.30 If d and d are in M(Rn ), d is supported in the Borel set

M and d is supported in the Borel set N , then d d is supported in M + N .
Hence,
supp(d d) supp(d) + supp(d) .
Proof:
If the Borel set A is disjoint from M + N and if y N , then
M (A y) =
and, hence,

( )
dd(A) = Ay (x) d(x) d(y) = d M (Ay) d(y) = 0 .
N M N

Lemma 0.3 If d M(Rn ), then

lim kd dB(0;R) k = 0 .
R+

Proof:
( )
kd dB(0;R) k = kdRn \B(0;R) k = |d| Rn \ B(0; R) 0 ,
since |d| is a finite non-negative Borel measure and B(0; R) Rn as R + .
Proposition 0.31 If d and d are in M(Rn ) and f is any bounded continuous
function in Rn , then

f (z) d d(z) = f (x + y) d(x) d(y) .
Rn Rn Rn
58

Proof:
By Lemma 0.3, we may take large enough R so that
kd dB(0;R) k  , kd dB(0;R) k  .
Because of the uniform continuity of f in B(0; 2R), we may, also, take a
linear combination
m
g = k Ak ,
k=1
where B(0; 2R) = m
k=1 Akis a partition of B(0; 2R) in pairwise disjoint Borel
subsets of small diameter, so that
sup |f (z) g(z)| 
zB(0;2R)

and, hence,
sup |g(z)| sup |f (z)| +  .
zRn zRn
The measure dB(0;R) dB(0;R) is supported in B(0; R)+B(0; R) = B(0; 2R).
Therefore,
 
 
 f (z) d B(0;R) d B(0;R) (z) g(z) dB(0;R) dB(0;R) (z)
Rn Rn
kdB(0;R) dB(0;R) k kdkkdk .
We, clearly, have that
 
 
 f (z) d d(z) f (z) dB(0;R) dB(0;R) (z)
Rn Rn
sup |f (z)|kd dB(0;R) kkdk
zRn
+ sup |f (z)|kdB(0;R) kkd dB(0;R) k
zRn
( )
 sup |f (z)| kdk + kdk .
zRn

In the same manner,

 
 
 g(z) d d(z) g(z) dB(0;R) dB(0;R) (z)
n n
R
( R
)
 sup |g(z)| kdk + kdk
zR n
( )( )
 sup |f (z)| +  kdk + kdk .
zRn

With similar calculations, we find that



 f (x + y) d(x) d(y)
B(0;R)B(0;R)


g(x + y) d(x) d(y)
B(0;R)B(0;R)
kdkkdk ,
0.9. THE FOURIER TRANSFORM 59

that


 f (x + y) d(x) d(y)
Rn Rn


f (x + y) d(x) d(y)
B(0;R)B(0;R)
( )
 sup |f (z)| kdk + kdk
zRn

and that


 g(x + y) d(x) d(y)
Rn Rn


g(x + y) d(x) d(y)
B(0;R)B(0;R)
( )( )
 sup |f (z)| +  kdk + kdk .
zRn

Finally, from the definition of d d and from linearity, we have

g(z) d d(z) = g(x + y) d(x) d(y) .
Rn Rn Rn

Combining all the above estimates with the last equality, we conclude that
 
 
 f (z) d d(z) f (x + y) d(x) d(y) C ,
Rn Rn Rn

where C is a constant independent of , and this finishes the proof.

Proposition 0.32 If d and d are in M(Rn ), then

dd c d
d = d c.

Also, if f L1 (Rn ) and d M(Rn ), then

fd
d = fbd
c.

Proof:
For the first part, we apply Proposition 0.31 to the function

f (z) = e2iz , z Rn .

For the last part, we just imitate the proof of Proposition 0.23.

Lemma 0.4 Suppose that f and g are in L1 (Rn ) and that Rn
g(x) dm(x) = 1.
Then,
f g f
in L1 (Rn ) as 0+ .
60

Proof:

|f g (x) f (x)| dm(x)

Rn

 ( ) 
=  f (x y) f (x) g(y) dm(y) dm(x)

Rn Rn
 
f (x y) f (x) dm(x) |g(y)| dm(y) .
Rn Rn

The function

 
(y) = f (x y) f (x) dm(x) , y Rn ,
Rn

has the properties:

1. supyRn | (y)| 2kf kL1 (Rn ) and
2. lim0 (y) = 0 for all y Rn .
By the Dominated Convergence Theorem,

lim |f g (x) f (x)| dm(x) = 0 .
0 Rn

Theorem 0.14 (The Inversion Formula) If f L1 (Rn ) and fb L1 (Rn ), then

f (x) = e2ix fb() dm()
Rn

for almost every x Rn .

Proof:
For every > 0 and every x Rn ,

e2ix e|| fb() dm()
2

Rn

e2i(yx) e|| dm() f (y) dm(y)

2
=
Rn Rn
( ) n2 2
e |xy| f (y) dm(y) .
2
=
Rn

Since fb L1 (Rn ), by the Dominated Convergence Theorem, the left side

converges to Rn e2ix fb() dm() as 0.
From Lemma 0.4, the right side converges to f in L1 (Rn ) as 0 and the
proof is complete.
0.9. THE FOURIER TRANSFORM 61

Theorem 0.15 (The Fourier transform is injective) If f and g are in L1 (Rn )

and fb = gb, then f = g.
c = d,
If d and d are in M(Rn ) and d c then d = d.

Proof:
It is, clearly, enough to consider the case when g = 0 and d is the zero
measure.
If fb = 0, then the Inversion Formula gives that f = 0.
Now, assume that d c = 0.
By Proposition 0.32, for all f L1 (Rn ),

fd
d = fbd
c = 0

and, thus, by the first part,

f d = 0 .
Considering any approximation to the identity { : > 0} and any
D(Rn ),

0 = d(x) (x) dm(x)
R
n

= (x)(y x) dm(x) d(y)

Rn Rn

(y) d(y) ,
Rn

as 0, by the Dominated Convergence Theorem and Proposition 0.9.

Therefore,
(y) d(y) = 0
Rn

for all D(Rn ) and, by Proposition 0.12, d is the zero measure.

Theorem 0.16 Suppose that (1 + | |)k f () L1 (Rn ). Then fb C k (Rn ) and

b
[ ]
D f () = e2ix (2ix) f (x) dm(x) = (2i) () f () b()
Rn

for all Rn and all with || k.

Proof:
By induction, it is enough to consider the case || = 1. We, thus, suppose
that (1 + | |)f () L1 (Rn ) and we get, for all j with 1 j n,

fb( + hej ) fb() e2ixj h 1
lim = lim e2ix f (x) dm(x)
h0 h h0 h
Rn
2ix
= e (2ixj )f (x) dm(x) .
Rn
62

The use of the Dominated Convergence Theorem in the above limit is per-
 2ixj h 
mitted by the inequality  e h 1  2|xj |.

Theorem 0.17 Suppose that f C k (Rn ) and D f L1 (Rn ) for all with
|| k. Then
d
D f () = (2i) fb()

for all Rn and all with || k.

Proof:
By induction, it is enough to consider the case || = 1.
We estimate
 f (x + hej ) f (x)

 f 
 e2ix dm(x) e2ix (x) dm(x)
Rn h Rn xj
 1( ) 
 f f 
=  e2ix (x + thej ) (x) dt dm(x)
R n 0 x j x j
1  
 f f 
 (x + thej ) (x) dm(x) dt
0 R n x j x j
f f

sup ( + y) () 1 n .
|y||h| xj xj L (R )

Therefore,

f (x + hej ) f (x) f
lim e2ix dm(x) = e2ix (x) dm(x) .
h0 Rn h Rn xj
On the other hand,

f (x + hej ) f (x)
lim e2ix dm(x)
h0 Rn h

e2ij h 1
= lim e2ix f (x) dm(x)
h0 Rn h
= 2ij fb()

and we conclude that

d
f
() = 2ij fb()
xj
for all Rn .

Corollary 0.1 1. If | |k f () L1 (Rn ) for all k 0, then fb C (Rn ) and

D fb L (Rn ) for all . In particular, if f is compactly supported in Rn , then
fb C (Rn ).
2. If f C (Rn ) and D f L1 (Rn ) for all , then | |k fb() L (Rn ) for
all k 0.
0.9. THE FOURIER TRANSFORM 63

The most symmetric result comes after the following definition.

Definition 0.23 f is called a Schwartz-function if it belongs to C (Rn )
and | |k |D f ()| L (Rn ) for all k 0 and all .
The space of all Schwartz-functions is denoted by S(Rn ) and it is called the
Schwartz-class.
For example, every D(Rn ) belongs to S(Rn ).

Theorem 0.18 The Fourier transform is a linear, one-to-one mapping from

S(Rn ) onto itself:
F : S(Rn ) S(Rn ) .
Proof:
If f S(Rn ), then | |k |f ()| L (Rn ) for all k 0 and, hence, | |k |f ()|
L (Rn ) for all k 0. By Corollary 0.1, fb belongs to C (Rn ).
1

A combination of Theorems 0.16 and 0.17 gives

[ { }]
(2i)|| D fb() = (2i)|| D () f () b()
{ }
for all and all and . From f S(Rn ) we get that D () f () L1 (Rn )
and, thus, | ||| D fb() L (Rn ) for all and .
Hence, fb S(Rn ) and F is a linear and, by Theorem 0.15, one-to-one
mapping of S(Rn ) into itself.
If g S(Rn ), we consider the function f = e gb. This belongs to S(Rn ) and,
by the Inversion Formula, g = f . b

There exists a natural metric on S(Rn ) and F is a homeomorphism of S(Rn ).

We shall not work in this direction.
64
 Part II

Main Theory

65
Chapter 1

Harmonic Functions

1.1 Definition
Suppose is an open subset of Rn .

Definition 1.1 A complex-valued function u defined in is called harmonic

in , if

1. u is continuous in and

2. u(x) = Mru (x) for all x and all r < d(x, )

or, if

1. u is continuous in and

2. u(x) = Aru (x) for all x and all r < d(x, ) .

Both conditions 2 are called the mean-value property of harmonic func-

tions. The first is the surface-mean-value property and the second the space-
mean-value property.
The two definitions are equivalent because, if u(x) = Mru (x) for all x
and all r < d(x, ), then

n r s n r
Aru (x) = n Mu (x)sn1 ds = n u(x)sn1 ds = u(x)
r 0 r 0

and, if u(x) = Aru (x) for all x and all r < d(x, ), then
r
rn u(x) = n Msu (x)sn1 ds
0

and, taking derivatives with respect to r, we get

nrn1 u(x) = nrn1 Mru (x)

67
68 CHAPTER 1. HARMONIC FUNCTIONS

and, hence,
u(x) = Mru (x) .
Properties of harmonic functions
(1) Linear combinations of harmonic functions are harmonic.
(2) Harmonic functions are preserved by rigid motions of the space.
Suppose u is harmonic in the open and x0 = O(x) + b is a rigid motion,
where O is an orthogonal linear transformation and b Rn .
Then 0 = O() + b = {O(x) + b : x } is open and u0 (x0 ) = u(x) is
defined and continuous on 0 . By the invariance of distances and of Lebesgue
measure under rigid motions,

0 0 1 1
u (x ) = u(x) = u(y) dm(y) = u0 (y 0 ) dm(y 0 )
Vn rn B(x;r) Vn rn B(x0 ;r)

for every x0 0 and every r < d(x0 , 0 ).

(3) Locally uniform limits of harmonic functions are harmonic.
Because, suppose uk u locally uniformly in , i.e. uniformly on every
compact subset of . Then, u is, obviously, continuous on . Also, S(x; r) is
compact in if x and r < d(x, ). Therefore,

u(x) uk (x) = Mruk (x) Mru (x) ,

because of uniform convergence on S(x; r).

(4) (Picard) If the real-valued u is harmonic in Rn and bounded from above or
from below, then it is constant.
It is enough to assume u 0 in Rn .
Take x 6= y and B(x; r1 ) B(y; r2 ). Let, in fact,

r2 = r1 + |x y| .

Then,

Vn r1n Aru1 (x) = u(z) dm(z) u(z) dm(z) = Vn r2n Aru2 (y)
B(x;r1 ) B(y;r2 )

and, hence,
r1n u(x) r2n u(y) .
Letting r1 , r2 +, we get, since r1
r2 1,

u(x) u(y)

and, symmetrically, u(y) u(x). Therefore,

u(x) = u(y)

and u is constant.
(5) If the real-valued harmonic u has a local extremum at x, then it is constant
in an open neighborhood of x.
1.2. MAXIMUM-MINIMUM PRINCIPLE 69

Suppose u(y) u(x) for all y B(x; r) and u is harmonic in an open set
containing B(x; r). Then,

1 1
u(x) = u(y) dm(y) u(x) dm(y) = u(x) .
Vn rn B(x;r) Vn rn B(x;r)

The inequality becomes equality only if u(y) = u(x) for almost every y in
B(x; r) and, since u is continuous,

u(y) = u(x)

for every y B(x; r).

1.2 Maximum-minimum principle

Theorem 1.1 (Maximum-Minimum Principle for Harmonic Functions) Let u
be real-valued and harmonic in the open Rn .
1. If u takes one of its extremum values at some x, then u is constant in the
connected component of which contains x.
2. If ( ) ( )
M = sup lim sup u(x) , m = inf lim inf u(x) ,
y 3xy y 3xy

then m u(x) M for all x in .

If u(x) = m or if u(x) = M at some x , then u is constant in the
connected component of which contains x.

Proof:
By the fifth property of harmonic functions and, since u is harmonic when
u is, the result is an immediate application of Proposition 0.5.
If O is any connected component of , then its closure O is compact and its
boundary O is a non-empty subset of . By Proposition 0.5,
( ) ( )
m inf lim inf u(x) u(x) sup lim sup u(x) M
yO O3xy yO O3xy

for all x O.

Corollary 1.1 Let u be harmonic in the open Rn . If lim3xy u(x) = 0

for all y , then u is identically 0 in .
Later on we shall study the

Problem of Dirichlet Given the open Rn and the complex-valued f

defined in , find u harmonic in so that

lim u(x) = f (y)

3xy
70 CHAPTER 1. HARMONIC FUNCTIONS

for all y . This is the Dirichlet Problem for with boundary func-
tion f .

The only remark we shall make at this point is that, if a solution exists,
then, by the Corollary 1.1, it is unique.

1.3 Differentiability of harmonic functions

If u is harmonic in the open subset of Rn and { : > 0} is any approxi-
mation to the identity, then, for every x ,

u(x) = u(x y) (y) dm(y)
B(0;)

= (r) u(x + rt) d(t) rn1 dr
0 S n1

= n1 Mru (x) (r) rn1 dr
0

= n1 u(x) (r) rn1 dr
0

= u(x) (y) dm(y)

B(0;)
= u(x) .

From this and from Proposition 0.7, we get that u is in C ( ) and, since
is arbitrary, that u is in C (). Therefore, we proved the

Theorem 1.2 If u is harmonic in the open set Rn , then u is infinitely

differentiable in .

Theorem 1.3 u is harmonic in the open set Rn if and only if u is in

C 2 () and satisfies the Laplace equation u = 0 everywhere in .

Proof:
Suppose u is harmonic in . By Theorem 1.2, u is in C 2 () and, taking any
B(x; R) and applying Greens Formula, we get

u
u(z) dm(z) = (y) dS(y)
B(x;R) S(x;R)

d( )
= Rn1 u(x + rt) r=R d(t)
n1 dr
(
S
)
n1 d
= R u(x + rt) d(t)
dr S n1 r=R
d ( )
= n1 Rn1 Mru (x) r=R
dr
1.4. HOLOMORPHY AND HARMONIC CONJUGATES 71

d( )
= n1 Rn1 u(x) r=R
dr
= 0.
Since u is continuous in and the ball is arbitrary, we conclude that

1
u(x) = lim u(z) dm(z) = 0
R0+ Vn Rn B(x;R)

for all x .
Now, suppose u = 0 everywhere in , fix an arbitrary x and take any
R < d(x, ).
By reversing the calculations above, we find
d( r )
Mu (x) r=R = 0
dr
and, therefore, MR
u (x) is constant in the interval 0 < R < d(x, ).
By the continuity of u at x,
u(x) = lim MR
u (x)
R0+

and we conclude that

u(x) = MR
u (x)
for all R < d(x, ).

Lemma 1.1 If f is twice continuously differentiable in some neighborhood of

the point x Rn , then
2n ( r ) 2(n + 2) ( r )
f (x) = lim 2
Mf (x) f (x) = lim 2
Af (x) f (x) .
r0+ r r0+ r
Proof:
We write the Taylor-expansion of order 2 of f at x and observe the resulting
cancellation in Mrf (x) and Arf (x).

In case f is not differentiable but satisfies the necessary integrability condi-

tions on spheres or balls, then either of the limits in Lemma 1.1 is used to define
the so-called generalized Laplacian of f .

1.4 Holomorphy and harmonic conjugates

The standard examples of harmonic functions in dimension 2 are the real and
imaginary parts of holomorphic functions in open sets.
Let f be holomorphic in the open set R2 . Then, if u = <f and
v = =f , the functions u and v are infinitely differentiable in and satisfy the
Cauchy-Riemann equations
u v u v
= , = .
x1 x2 x2 x1
72 CHAPTER 1. HARMONIC FUNCTIONS

Therefore,
u = v = 0
in .
In another way, using the Cauchys Formula,

1 f (y) 1
f (x) = dy = f (y) dS(y)
2i S(x;r) y x 1 r S(x;r)

and taking real and imaginary parts, we prove the mean-value property of u
and v.
At least locally, these are the only examples of harmonic functions in di-
mension 2. In fact, if u is real-valued and harmonic in an open rectangle
R = (a1 , b1 ) (a2 , b2 ), then there exists a holomorphic f in this rectangle
so that u = <f .
To see this, we fix a point (c1 , c2 ) R and define
X1 X2
u u
v(X1 , X2 ) = (x1 , c2 ) dx1 + (X1 , x2 ) dx2
c1 x2 c2 x1

for all X = (X1 , X2 ) R .

We, clearly, have

v u
(X1 , X2 ) = (X1 , X2 )
x2 x1
and
X2
v u 2u
(X1 , X2 ) = (X1 , c2 ) + (X1 , x2 ) dx2
x1 x2 c2 x21
X2 2
u u
= (X1 , c2 ) (X1 , x2 ) dx2
x2 c2 x22
u u u
= (X1 , c2 ) (X1 , X2 ) + (X1 , c2 )
x2 x2 x2
u
= (X1 , X2 ) .
x2
Therefore, the functions u and v satisfy the Cauchy-Riemann equations in
R and, thus, the function f = u + iv is holomorphic in R with u = <f there.
If g = u + iw is another function holomorphic in R with u = <g there, then,
by the Cauchy-Riemann equations, we get that v and w differ by some (real)
constant in R and, hence, f and g differ by an imaginary constant in R.
This result is, as we shall see in a moment, true, even when we replace the
rectangle by an arbitrary simply-connected open R2 . On the other hand,
the function u : R2 \ {0} R given by

u(x) = log |x| , x 6= 0

1.4. HOLOMORPHY AND HARMONIC CONJUGATES 73

is harmonic in R2 \ {0}, but there is no f holomorphic in R2 \ {0} so that

<f = u everywhere in R2 \ {0}.
In fact, if there was such an f , then we would have that
f (x) = log0 (x) + ia , x 0
for some constant a R, where 0 = R2 \ {x = (x1 , 0) : x1 0} and log0
is the principal branch of the logarithm in 0 . This, clearly, contradicts the
continuity of f at the points of the negative-x1 -axis.
Definition 1.2 Let u be real-valued and harmonic in the open R2 . If v is
real-valued and harmonic in and
u v u v
= , =
x1 x2 x2 x1
everywhere in , then v is called a harmonic conjugate of u in .
It is obvious that the existence of a harmonic conjugate of u in is equivalent
to the existence of a holomorphic function in whose real part is u. Further-
more, if is, also, connected, then any two harmonic conjugates of u differ by
a real constant in .
If v is a harmonic conjugate of u, then u is a harmonic conjugate of v.
The many-valued function called argument is defined in R2 \ {0} by
x
arg(x) = + k2, where is any number with ei = |x| and k takes all values
in Z.
If g : A R2 \ {0} is continuous in A, then any function v : A R which
is continuous in A and is such that
( ) g(x)
exp iv(x) =
|g(x)|
for all x A, is called a branch of the argument of g in A. It is easy to
see, by continuity, that if A is a connected set and a branch v of the argument
of g exists in A, then the totality of branches of the argument of g in A are the
functions v + k2, where k Z.
It is obvious that, through f = log |g| + iv every branch of the logarithm
of g defines a branch of the argument of g and vice versa.
Proposition 1.1 If g : R2 \ {0} is holomorphic in the connected open set
R2 , then every branch of the argument of g in is a harmonic conjugate of
log |g| in and, conversely, every harmonic conjugate of log |g| in is, except
for an additive constant, a branch of the argument of g in .
If v is any branch of the argument of g in , then, clearly, the function
f = log |g| + iv is a branch of the logarithm of g in and, hence, v is a
harmonic conjugate of log |g| in . Conversely, if v is a harmonic conjugate of
log |g| in , then the function f = log |g| + iv is holomorphic in and the
functions exp(f ) and g have the same absolute value everywhere in . It is, now,
easy to show, through the Cauchy-Riemann equations, that, in each component
of , exp(f ) = g, where = ei is a constant with || = 1. Hence, v is a
branch of the argument of g in .
74 CHAPTER 1. HARMONIC FUNCTIONS

Definition 1.3 An open set R2 is called simply-connected if is con-

nected and R2 \ is connected.

Lemma 1.2 Suppose that R1 and R2 are two open rectangles with sides parallel
to the x1 - and x2 -axes with one common side I and let R be the open rectangle
R1 R2 int(I). If u is harmonic in R and v1 is a harmonic conjugate of u in
R1 , then there exists a harmonic conjugate v of u in R so that v = v1 in R1 .

Proof:
Let v0 be any harmonic conjugate of u in R. This is, also, a harmonic
conjugate of u in R1 and, hence, v0 = v1 + a everywhere in R1 for some real
constant a. Then v = v0 a is the harmonic conjugate of u in R with v = v1
in R1 .

Theorem 1.4 Suppose that u is real-valued and harmonic in the simply-conne-

cted R2 . Then, there exists a harmonic conjugate of u in . Two such
harmonic conjugates of u differ by a real constant in .

Proof:
Fix a small > 0 and consider the collection of all closed squares

Qk = {x = (x1 , x2 ) : k1 x1 (k1 + 1), k2 x2 (k2 + 1)} ,

where k = (k1 , k2 ) Z Z.
Fix, also, a large K Z and consider the finite subcollection QK
of all Qk
with |k1 | K and |k2 | K which are contained in . Form the set

K
= int {Qk : Qk QK
} ,

which consists of the interiors of all Qk QK together with the interiors of all
common sides of all these squares.
We separate, in the obvious way, K into horizontal layers of height where
each layer consists of finitely many open rectangles arranged horizontally and
having no common sides. These layers are glued at the common parts of their
horizontal sides to form the set K .
Starting with the first (bottom) layer, we construct a harmonic conjugate of
u in each of its rectangles and we, thus, have a harmonic conjugate v1 of u in
this first layer.
Suppose that we have constructed a harmonic conjugate vk of u in the part
K,k
of K which lies below its k + 1 layer and consider the first (to the left)
rectangle R of the k + 1 layer of K . If R has no common side with any of the
rectangles of the k layer, we construct, arbitrarily, a harmonic conjugate of u in
R. If the lower side of R has a common part with the upper side of only one
rectangle R0 of the k layer, we use Lemma 1.2 (at most three times) to extend
v1 as a harmonic conjugate of u in the union K,k R.
1.5. FUNDAMENTAL SOLUTION 75

If the lower side of R has a common part with the upper side of at least
two rectangles R0 , R00 , . . . of the k layer, then we consider the sets Ok0 , Ok00 , . . .
defined as the components of K,k which contain R0 , R00 , . . ., respectivelly.
Observe, now, that these components are pairwise disjoint as a trivial ap-
plication of Theorem 0.8 shows!
We may now extend vk in the union of, say, Ok0 with R and, next, modify vk
in the other components Ok00 , . . . , adding an appropriate (for each component)
constant to it, so that the, already constructed extension of vk from Ok into R
coincides (in R) with its similarly constructed extensions from Ok00 , . . . into R.
After all this, we end up with a harmonic conjugate of u in the union K,k R.
We, now, continue with the second from the left rectangle in the k + 1 layer
and, with the same procedure, we modify and extend the already constructed
harmonic conjugate into this new rectangle. After finitely many steps, we shall
end up with a harmonic conjugate vk+1 of u in K,k+1 .
Finally, after finitely many steps we find a harmonic conjugate of u in K .
Now, we consider any open exhaustion {(m) } of so that every (m) is
connected. Each (m) is contained in some K , provided is small and K is
large. Therefore, there is some harmonic conjugate v(m) of u in (m) .
Since (m) is connected, modifying v(m+1) by some appropriate constant, we
may arrange so that v(m+1) = v(m) in (m) . Therefore, a function v is defined
in which, clearly, is a harmonic conjugate of u in .

Proposition 1.2 Consider the open subsets 1 and 2 of R2 , a function f

holomorphic in 1 and assume that f (1 ) 2 .
If u is harmonic in 2 , then u f is harmonic in 1 .
Proof:
Direct calculation, using the Cauchy-Riemann equations, shows that
(u f )(x) = |f 0 (x)|2 u(f (x)) = 0
for every x 1 .

Example
If the open set R2 does not contain 0 and u is harmonic in , we
consider the open set = {x : x1 } and define
(1)
u (x) = u
x
for all x . Then u is harmonic in .

1.5 Fundamental solution

Now, suppose that u is defined in a ring B(0; R1 , R2 ) = {x Rn : R1 < |x| <
R2 } and that it is a radial function:
u(x) = u(y)
76 CHAPTER 1. HARMONIC FUNCTIONS

for all x, y B(0; R1 , R2 ) with |x| = |y|.

Then the function

u (r) = u(x) , for any x B(0; R1 , R2 ) with |x| = r ,

is well-defined in the interval R1 < r < R2 .

By trivial calculations, we can prove that u is twice continuously differen-
tiable in B(0; R1 , R2 ) if and only if u is twice continuously differentiable in the
interval R1 < r < R2 and, if this is true,

d2 u n 1 du
u(x) = 2
(|x|) + (|x|)
dr r dr
identically in B(0; R1 , R2 ).

Proposition 1.3 Let u be a radial function defined in B(0; R1 , R2 ). Then u is

harmonic there if and only if
{ 1
A log |x| +B , n=2
u(x) = 1
A |x|n2 + B , n > 2

for some constants A and B.

If u is radial and harmonic in B(0; R), then u is constant there.

Proof:
We, easily, solve the second order ordinary differential equation

d2 u n 1 du
(r) + (r) = 0
dr2 r dr
in the interval R1 < r < R2 .
The last statement is proved from the first, using the continuity of u at 0.

Definition 1.4 We call the function

{ 1
log |x| , n=2
h(x) = 1
|x|n2 , n>2

the fundamental solution of the Laplace equation in Rn \ {0} .

We, also, define

hz (x) = h(x z) , x Rn ,

for every z Rn .

In any case the fundamental solution becomes + at 0, but observe a crucial

difference between n = 2 and n > 2. The fundamental solution is bounded from
below (by 0) when n > 2, while it tends to when x tends to in R2 .
1.5. FUNDAMENTAL SOLUTION 77

Lemma 1.3
{
1 h (R) , if |x| = r R
hx (y) dS(y) =
n1 Rn1 S(0;R) h(x) = h (r) , if |x| = r R .

Proof:
Consider the function

1
u(x) = hx (y) dS(y) , x Rn .
n1 Rn1 S(0;R)

It is clear that u is well-defined for all x

/ S(0; R), since hx is bounded in
S(0; R). For the same reason, using the ideas in the proof of Proposition 0.6,
we may interchange integration and differentiation and prove that u = 0 in
Rn \ S(0; R).
u is also well-defined for all x S(0; R), since, as it is easy to see, hx is
integrable in S(0; R).
If |x1 | = |x2 |, we consider any orthogonal transformation O so that x2 =
O(x1 ) and, since dS is rotation invariant and O preserves distances,

1
u(x2 ) = hO(x1 ) (y) dS(y)
n1 Rn1 S(0;R)

1 ( )
= n1
hO(x1 ) O(z) dS(z)
n1 R S(0;R)

1
= hx (z) dS(z)
n1 Rn1 S(0;R) 1
= u(x1 ) .

Hence, u is radial and this, by Proposition 1.4, implies that u is constant in

B(0; R). This constant is u(0) = h (R).
hx is harmonic in Rn \ {x} and, if |x| > R, we may apply the mean-value
property to get

1
hx (y) dS(y) = hx (0) = h(x) .
n1 Rn1 S(0;R)

Now, consider any x with |x| = R and the points r

R x for all r > R. It is
trivial to see that (r )
h x y h(x y)
R
for all y S(0; R) and all r > R.
Therefore, by the Dominated Convergence Theorem,
(r ) (r )
u(x) = lim u x = lim h x = h(x) = h (R)
rR+ R rR+ R
and this finishes the proof.
78 CHAPTER 1. HARMONIC FUNCTIONS

1.6 Potentials
Definition 1.5 Let d be a compactly supported complex Borel measure. The
function

Uhd (x) = h(x y) d(y) = hx (y) dm(y) , x Rn \ supp(d) ,
Rn Rn

is called the h-potential of d. More precisely, the logarithmic potential in

case n = 2 and the Newtonian potential in case n > 2.
In case the support of d is contained in a C 1 -hypersurface , then the h-
potential is, also, called single-layer potential of d.

Definition 1.6 Let d be a complex Borel measure with compact support con-
tained in an orientable C 1 -hypersurface . Then the function

hx
U (x) = (y) d(y) , x Rn \ supp(d) ,

where

is a continuous unit vector field normal to , is called double-layer
potential of d.

Comments A Newtonian potential can be defined for locally finite complex

Borel measures d which are not necessarily supported in compact subsets of
Rn , provided they are of a definite sign, non-negative or non-positive. This
is because the fundamental solution is positive and we end up integrating a
quantity of a definite sign. The integral is, then, well-defined for all x Rn ,
although it may take the value .
The situation is different for the logarithmic potential, because the funda-
mental solution is, on the one hand, bounded from below in every bounded
set but, on the other hand, is not bounded from below in any unbounded set.
Therefore, a Borel measure not supported in a compact subset of Rn , even if it
is of a definite sign, may not have a well-defined logarithmic potential.
Hence, we never consider logarithmic potentials of Borel measures which are
not compactly supported.

Example: The h-potential of the Dirac mass da is equal to the translate of

the fundamental solution at the point a :

Uhda = ha .

Proposition 1.4 Any of the above defined potentials of a compactly supported

complex Borel measure d is harmonic in the open set Rn \ supp(d).

Proof:
We, easily, prove that U (x) = 0 for every x / supp(d), by passing the
derivatives inside the integrals. To do this, we use a simple variant of the proof
of Proposition 0.6.
1.7. FLUX 79

1.7 Flux
The next three theorems in this and the next section are applications of the
Greens Formulas. The first is for a special kind of domains, but for general
harmonic functions. The second and third theorems together with their corollary
are quite general and fundamental.
Theorem 1.5 If u is harmonic in the ring B(x; r1 , r2 ), then the quantity

u
= (y) dS(y)
S(x;r)

is constant as a function of r in the interval r1 < r < r2 . Here,

is the con-
tinuous unit vector field normal to S(x; r) in the direction towards the exterior
of B(x; r).
Also, for some constant ,
{
1 log 1r + , if n = 2
Mu (x) =
r
(n2)n1 rn2 + , if n > 2
1

in the interval r1 < r < r2 .

If the ring becomes a ball B(x; r2 ), then = 0 in all the above.
Proof:
Take r1 < r < r0 < r2 and = B(x; r, r0 ). Then

u
(y) dS(y) = u(x) dm(x) = 0 ,

where
is the continuous unit vector field normal to in the direction towards
the exterior of . This implies

u u
(y) dS(y) + (y) dS(y) = 0 ,
S(x;r) S(x;r )
0

where, in each integral,

is the continuous unit vector field normal to the
sphere in the direction towards the exterior of the corresponding ball.
Therefore,
u
= (y) dS(y)
S(x;r)

is constant.
Now, the last equality, easily, becomes
d
Mru (x) =
dr n1 rn1
and this implies the formula for Mru (x).
If u is harmonic in B(x; r2 ), then Mru (x) = u(x) is constant as a function of
r and, hence, = 0.
80 CHAPTER 1. HARMONIC FUNCTIONS

Theorem 1.6 Suppose is an open subset of Rn , K is a compact subset of

and u is harmonic in the open set \ K. Consider a variable bounded open
with C 1 -boundary such that K . Then, the quantity

u
= (y) dS(y) ,

where is the continuous unit vector field normal to in the direction

towards the exterior of , does not depend on .
In case K is empty, which means that u is harmonic in , then = 0.
Proof:
We consider a third bounded open 1 with C 1 -boundary such that
K 1 1
and apply Greens theorem in \ 1 .
Then
u u
(y) dS(y) = (y) dS(y) .
1
Now, given two sets, and , as in the statement of the theorem, we can
use an 1 such that 1 and finish the proof, by comparing the s
for these two sets with the for the third set.
The last result is proved, by applying Greens Formula in .

Definition 1.7 Suppose is an orientable C 1 -hypersurface,

is a continuous
unit vector field normal to and u is harmonic in an open set containing .
Then the quantity
u
= (y) dS(y)

is called flux of u through in the direction determined by the vector
field
.

1.8 The representation formula

Notation {
1 , if n = 2
n =
(n 2)n1 , if n > 2 .
Theorem 1.7 If is a bounded open set with C 1 -boundary, u is harmonic in
and belongs to C 1 (), then
(
1 hx u )
u(x) = (y)u(y) hx (y) (y) dS(y)
n
for every x , where
is the continuous unit vector field normal to in
the direction towards the exterior of .
Therefore, u is represented as a difference between a double- and a single-
layer potential.
1.8. THE REPRESENTATION FORMULA 81

Proof:
Take x and B(x; r) and apply Greens Formula in the open set
= \ B(x; r) with f = u and g = hx :
(
hx u )
0 = u(y) (y) hx (y) (y) dS(y)

(
hx u )
+ u(y) (y) hx (y) (y) dS(y) ,
S(x;r)

where is the continuous unit vector field normal to in the direction towards
the exterior of and normal to S(x; r) in the direction towards the interior of
B(x; r).
In case n = 2, using Theorem 1.5, the second integral becomes

1 1 u
u(y) dS(y) log (y) dS(y) = 1 Mru (x) = 1 u(x) .
r S(x;r) r S(x;r)

Similarly, in case n > 2, the second integral becomes

1 1 u
(n 2) n1 u(y) dS(y) n2 dS(y)
r S(x;r) r S(x;r)

= (n 2)n1 Mru (x) = (n 2)n1 u(x) .

The result of the theorem is, now, obvious.

Corollary 1.2 Let d be a compactly supported complex Borel measure and

a bounded open set with C 1 -boundary such that supp(d) . Then

Uhd
= (y) dS(y) = n d(Rn ) ,

where Uhd is the h-potential of d.

Proof:

Uhd ( )
(y) dS(y) = hx () d(x) (y) dS(y)
supp(d)

hx
= (y) d(x) dS(y)
supp(d)

hx
= (y) dS(y) d(x) .
supp(d)

Observe that, if x supp(d), then x and, applying Theorem 1.7 with

u = 1,
hx
(y) dS(y) = n .

82 CHAPTER 1. HARMONIC FUNCTIONS

Therefore,

Uhd
(y) dS(y) = n d(x) = n d(Rn ) .
supp(d)

Combining Corrolary 1.2 with the last part of Theorem 1.6, we may state a
general

Principle: The flux of the h-potential of a compactly supported complex Borel

measure d through the boundary of a bounded, open set with C 1 -boundary
disjoint from supp(d) and in the direction towards the exterior of is equal to
a (negative) constant times the total mass of the part of the measure which lies
inside the open set.

1.9 Poisson integrals

Definition 1.8 If x 6= x0 , then the point

R2
x = x0 + (x x0 )
|x x0 |2

is called the symmetric of x with respect to the sphere S(x0 ; R).

We define as the symmetric of x0 and x0 as the symmetric of .

It is easy to see that x and x are on the same half-line having x0 as vertex
and the product of their distances from x0 is equal to R2 . This, in an extended
sense, happens, also, for the pair of x0 and .
Observe that x is the symmetric of x , x = (x ) , and that x = x if and
only if x S(x0 ; R).

Suppose u is harmonic in B(x0 ; R) and belongs to C 1 (B(x0 ; R)). Then,

applying Theorem 1.7 to = B(x0 ; R), we find, for every x B(x0 ; R),
( h
1 x u )
u(x) = (y)u(y) hx (y) (y) dS(y) ,
n S(x0 ;R)

where is the continuous unit vector field normal to S(x0 ; R) directed towards
the exterior of B(x0 ; R).
If x 6= x0 , consider, also, the symmetric x of x with respect to the sphere
S(x0 ; R). Then hx is harmonic in an open set containing B(x0 ; R) and Greens
Formula implies
( h
1 x u )
0 = (y)u(y) hx (y) (y) dS(y) .
n S(x0 ;R)

Let n = 2.
1.9. POISSON INTEGRALS 83

Then hx (y) hx (y) = log |x|xy|

y| = log
|xx0 |
R for all y S(x0 ; R) and,
subtracting the two equations above, we get

1 (hx hx )
u(x) = (y)u(y) dS(y)
2 S(x0 ;R)

1 |x x0 | u
+ log (y) dS(y) .
2 R S(x0 ;R)

The last integral is 0 and a trivial calculation of the directional derivative of

hx () hx () = log |x|x || in the direction of


gives

1 R2 |x x0 |2
u(x) = u(y) dS(y) .
1 R S(x0 ;R) |x y|2

Now, let n > 2.

Rn2
Then hx (y) |xx0|
n2 hx (y) = 0 for all y S(x0 ; R) and, if we subtract

Rn2
the two equations, after multiplying the second by the factor |xx0 |n2 , we find

1 ( Rn2 )
u(x) = hx hx (y)u(y) dS(y) .
n S(x0 ;R) |x x0 |n2

Again, a calculation gives

1 R2 |x x0 |2
u(x) = u(y) dS(y) .
n1 R S(x0 ;R) |x y|n

Of course, this formula covers the case n = 2 above and it is, trivially, true
for x = x0 .

Definition 1.9 If x B(x0 ; R), then the function

1 R2 |x x0 |2
P (y; x, x0 , R) = , y S(x0 ; R) ,
n1 R |x y|n

is called the Poisson kernel of the ball B(x0 ; R) with respect to x.

We have proved the

Theorem 1.8 If u is harmonic in B(x0 ; R) and it belongs to C 1 (B(x0 ; R)),

then

u(x) = P (y; x, x0 , R)u(y) dS(y) , x B(x0 ; R) .
S(x0 ;R)

This is called the Poisson integral formula.

84 CHAPTER 1. HARMONIC FUNCTIONS

Properties of the Poisson kernel

(1) P is positive.
(2) P is a harmonic function of x in B(x0 ; R).
This
is a matter of calculation of the Laplacian.
(3) S(x0 ;R) P (y; x, x0 , R) dS(y) = 1 .
We just apply the Poisson integral formula to u = 1.
(4) If V is an open neighborhood of y0 S(x0 ; R), then

lim P (y; x, x0 , R) = 0
B(x0 ;R)3xy0

uniformly in y S(x0 ; R) \ V .
If |y y0 | 0 for all y S(x0 ; R) \ V , then, when |x y0 | 21 0 , we have

1 R2 (R |x y0 |)2
0 P (y; x, x0 , R) 0
n1 R ( 12 0 )n

as |x y0 | 0.
Definition 1.10 Let f be integrable in S(x0 ; R) with respect to the surface mea-
sure dS and define

Pf (x) = Pf (x; x0 , R) = = Pf (x; B) = P (y; x, x0 , R)f (y) dS(y)
S(x0 ;R)

for all x B(x0 ; R).

This is called the Poisson integral of f in B = B(x0 ; R).

Theorem 1.9 Pf ( ; x0 , R) is harmonic in B(x0 ; R).

If f is continuous at some y0 S(x0 ; R), then

lim Pf (x; x0 , R) = f (y0 ) .

B(x0 ;R)3xy0

Therefore, if f is continuous in S(x0 ; R), then Pf ( ; x0 , R) is the solution

of the Dirichlet problem in B(x0 ; R) with boundary function f .

Proof:
The harmonicity of Pf ( ; x0 , R) results from property (2) of the Poisson
kernel.
For the limit, it is enough, by linearity, to assume that f is real-valued and,
then, to prove
lim sup Pf (x; x0 , R) f (y0 ) .
B(x0 ;R)3xy0

Indeed, we then apply this to f , find

f (y0 ) lim inf Pf (x; x0 , R)

B(x0 ;R)3xy0

and combine the two inequalities to get the equality we want to prove.
1.10. CONSEQUENCES OF THE POISSON FORMULA 85

Let M > f (y0 ) and fix an open neighborhood V of y0 so that f < M in

S(x0 ; R) V . Then,

Pf (x) = P (y; x, x0 , R)f (y) dS(y)
S(x0 ;R)V

+ P (y; x, x0 , R)f (y) dS(y) .
S(x0 ;R)\V

Now, by properties (1) and (3) of the Poisson kernel, the first integral is

M P (y; x, x0 , R) dS(y) = M M P (y; x, x0 , R) dS(y)
S(x0 ;R)V S(x0 ;R)\V

and, by property (4) of the Poisson kernel, the integrals over S(x0 ; R) \ V tend
to 0 as x y0 .
Hence,
lim sup Pf (x) M
B(x0 ;R)3xy0

and, since M can be taken arbitrarily close to f (y0 ), the proof is complete.

Remark There is a second way to prove the Poisson Integral Formula, without
going through Greens Formula (in case we wish to avoid its heavy technical
machinery).
Suppose u is harmonic in B(x0 ; R) and it belongs to C 1 (B(x0 ; R)) and con-
sider the function Pu ( ; x0 , R), the solution of the Dirichlet problem in B(x0 ; R)
with the restriction of u in S(x0 ; R) as boundary function. By the uniqueness
of the solution, we conclude that

u(x) = Pu (x; x0 , R) , x B(x0 ; R) .

1.10 Consequences of the Poisson formula

Proposition 1.5 If u is harmonic in the open Rn , then it is real-analytic
in .

Proof:
Let x0 and take B(x0 ; R) . Write u(x) = Pu (x; x0 , R) in B(x0 ; R)
and use the real-analyticity of P (y; x, x0 , R):

R2 |x x0 |2
1
P (y; x, x0 , R) =
n1 R |x y|n
1 R2 |x x0 |2
=
n1 R |(x x0 ) (y x0 )|n

= a (y)(x x0 ) ,

86 CHAPTER 1. HARMONIC FUNCTIONS

where the sum is over all multi-indices = (1 , . . . , n ), a is a polynomial

( ) in
y and the series converges uniformly in y S(x0 ; R) and in x B x0 ; 21 R .
Therefore,
( 1 )
u(x) = (x x0 ) , x B x0 ; R ,

2

where = S(x0 ;R) a (y)u(y) dS(y) and, hence, u is real-analytic at the arbi-
trary point x0 of .

Theorem 1.10 Suppose that u is harmonic in the connected open set Rn

and that u = 0 in an open subset O of . Then u = 0 identically in .
Proof:
It is obvious that, for all x O, D u(x) = 0 for every multi-index .
We consider the sets

A = {x : D u(x) = 0 for all } , B = \A .

If x0 B, then D0 u(x0 ) 6= 0 for some multi-index 0 and, by continuity,

this is true in some neighborhood of x0 . Hence, the set B is open.
If x0 A, then the Taylor-series expansion of u in a neighborhood of x0 has
all its coefficients equal to 0, and, thus, u = 0 in this neighborhood. Therefore,
this neighborhood is contained in A and A is an open set.
By the connectivity of , we get that = A and, thus, u = 0 in .

Theorem 1.11 (Weak mean-value property) If u is continuous in the open

Rn and if for every x there exists a sequence {rk (x)} such that rk (x) 0
r (x)
and u(x) = Muk (x) for all k, then u is harmonic in .
The same is true, if we replace the surface-means with the space-means in
the above statement.
Proof:
Take B(x0 ; R) and consider

(x) = u(x) Pu (x; x0 , R) , x B(x0 ; R) .

By the continuity of u and Theorem 1.9, we have that is continuous in

B(x0 ; R) and that limB(x0 ;R)3xy (x) = 0 for every y S(x0 ; R).
Since Pu ( ; x0 , R) is harmonic in B(x0 ; R), for every x B(x0 ; R) we have
that
r (x)
(x) = Mk (x)

for every k which is large enough so that B(x; rk (x)) B(x0 ; R).
From this, in the same way in which we proved the fifth property of harmonic
functions, we can prove that, if takes one of its extremal values at some point
in B(x0 ; R), then is constant in a neighborhood of this point.
1.10. CONSEQUENCES OF THE POISSON FORMULA 87

Now, applying Proposition 0.5 to and , we conclude that

(x) = 0

for every x B(x0 ; R).

Hence, u is equal to the harmonic Pu in B(x0 ; R) and, since the ball is
arbitrary in , we get that u is harmonic in .
Exactly the same argument applies when we replace Mru by Aru .

The following theorem gives a necessary and sufficient condition for a point
to be a removable isolated singularity of a harmonic function.

Theorem 1.12 (Riemann) Suppose u is harmonic in B(x0 ; R) \ {x0 } .

Then, u can be defined at x0 so that it becomes harmonic in B(x0 ; R) if and
only if
u(x)
lim = 0.
xx0 hx0 (x)

Proof:
The necessity of the condition is trivial.
Fix R1 < R and consider
( )
(x) = u(x) Pu (x; x0 , R1 )  hx0 (x) h (R1 ) , x B(x0 ; R1 ) \ {x0 } ,

for an arbitrary  > 0.

By Theorem 1.9, we have that is harmonic in = B(x0 ; R1 ) \ {x0 } and,
also, by the hypothesis, that lim sup3xy (x) 0 for every y .
Theorem 1.1 implies that (x) 0 for all x B(x0 ; R1 ) \ {x0 }. Hence,
( )
u(x) Pu (x; x0 , R1 ) +  hx0 (x) h (R1 )

for all x B(x0 ; R1 ) \ {x0 } and every  > 0 and, finally,

u(x) Pu (x; x0 , R1 ) , x B(x0 ; R1 ) \ {x0 } .

Working in the same manner with u, we find the opposite inequality.

Therefore,

u(x) = Pu (x; x0 , R1 ) , x B(x0 ; R1 ) \ {x0 } .

But, Pu ( ; x0 , R1 ) is harmonic in B(x0 ; R1 ) and, defining

u(x0 ) = Pu (x0 ; x0 , R1 ) = MR
u (x0 ) ,
1

u becomes harmonic in B(x0 ; R1 ) and, hence, in B(x0 ; R).

88 CHAPTER 1. HARMONIC FUNCTIONS

Theorem 1.13 (Reflection Principle of H. Schwartz) Let L be a hyperplane in

Rn and Rn be open and symmetric with respect to L. Let L+ and L be the
two open half-spaces on the two sides of L and let + = L+ , = L
and = L.
Suppose u is harmonic in + and
lim u(x) = 0
+ 3xy

for every y .
Then, u can be defined in so that it becomes harmonic in .
Proof:
Define {
0, if x
u(x) = ,
u(x ) , if x
where x is the symmetric of x with respect to L.
Since u is harmonic in + , it satisfies the mean-value property for all balls
which are contained in + .
Since symmetry with respect to a hyperplane is a rigid motion, we have, by
the second property of harmonic functions, that the extended u is harmonic in
and it satisfies the mean-value property for all balls contained in .
We also have that the extended u is continuous in and, if we take any
point x and any B(x; R) , then
u(x) = 0 = Aru (x)
for all r < R. This holds because the two integrals over the two half-balls,
B(x; r) L+ and B(x; r) L , cancel.
Therefore, by Theorem 1.11, u is harmonic in .

Theorem 1.14 (Harnacks Inequalities) Let u be positive and harmonic in the

open set Rn and let B(x0 ; R) . Then,
R |x x0 | u(x) R + |x x0 |
Rn2 Rn2
(R + |x x0 |) n1 u(x0 ) (R |x x0 |)n1
for all x B(x0 ; R).
Also,
u(x)
3n 3n
u(x0 )
for all x, x0 in B(x0 ; 21 R).
Moreover, if is connected and K is any compact subset of , then there is
a constant C = Cn,K, > 0 depending only on n, K and so that
1 u(x)
C
C u(x0 )
for all x, x0 in K.
1.10. CONSEQUENCES OF THE POISSON FORMULA 89

Proof:
The first set of inequalities is an immediate application of the Poisson integral
formula, when we make use of the

R |x x0 | |x y| R + |x x0 | , y S(x0 ; R) ,

of the positivity of u and of the mean-value property of u at x0 .

If |x x0 | 12 R, then from these inequalities we get

2n2 u(x)
3 2n2 .
3n1 u(x0 )

If, also, |x0 x0 | 12 R, then we immediately get

u(x)
3n 3n .
u(x0 )

For the last set of inequalities, we observe that it is enough to prove the
inequality in the right.
For each x K take B(x; R(x)) and find x1 , . . . , xN so that K
N 1
k=1 B(xk ; 2 R(xk )).
Now, set

M = max(u(x1 ), . . . , u(xN )) , m = min(u(x1 ), . . . , u(xN )) .

From the first part, we get that

u(x) M
32n
u(x0 ) m

for every x, x0 in K.
It is obvious that there is some p N, depending only on the points
x1 , . . . , xN and , so that for every two xk and xl there are at most p successive
points in , the first being xk and the last being xl and every two consecutive
ones of which are contained in the same closed ball whose double is contained
in .
Applying the first part, we see that, for all k and l, u(x k)
u(xl ) 3
pn
and, finally,

u(x)
3(p+2)n
u(x0 )

for all x, x0 in K.

Theorem 1.15 (Gradient Estimates) Let u be harmonic in the open set

Rn and B(x0 ; R) .

1. |grad u(x0 )| R
n
n1 Rn1 S(x0 ;R) |u(y)| dS(y) .
1
90 CHAPTER 1. HARMONIC FUNCTIONS

2. If, also, u is positive in S(x0 ; R), then |grad u(x0 )| n
R u(x0 ) .
n M m
3. If m u M in S(x0 ; R), then |grad u(x0 )| R 2 .
Proof:
If x0 = (x0,1 , . . . , x0,n ), then, by an easy calculation in the Poisson integral,

u n
(x0 ) = (x0,j yj )u(y) dS(y) .
xj n1 Rn+1 S(x0 ;R)
u
Multiplying both sides by x j
(x0 ), then summing over j = 1, . . . , n and
using Cauchys inequality inside the integral, we prove 1.
If u > 0 in S(x0 ; R), then

1
|u(y)| dS(y) = u(x0 )
n1 Rn1 S(x0 ;R)

and 2 is implied by 1.
Finally, if m u(y) M for all y S(x0 ; R), then we apply 1 to the
function u M +m
2 and prove 3.

1.11 Monotone sequences

Theorem 1.16 (Monotone sequences of harmonic functions) Suppose {um } is
a sequence of real-valued functions harmonic in the connected open set Rn
and u1 u2 . . . in .
Then, either um + uniformly on compact subsets of or there exists
some u harmonic in so that um u uniformly on compact subsets of .
There is a dual result for decreasing sequences of harmonic functions.

Proof:
Subtracting u1 from all um , we may assume that

0 u1 u2 u3 . . .

in . By the Maximum-Minimum Principle, we may even assume that all

inequalities are strict everywhere in .
(1) Let x0 with um (x0 ) + and consider any compact K . By
Theorem 1.14 applied to K {x0 }, we have that, for some C > 0 independent
of m,
1
um (x0 ) um (x)
C
for all x K. This implies that um + uniformly in K.
(2) Let x0 with um (x0 ) M for some real M and consider any compact
K . Again by Theorem 1.14, we have that

uk (x) um (x) C(uk (x0 ) um (x0 ))

1.12. NORMAL FAMILIES OF HARMONIC FUNCTIONS 91

for all x K and all m, k with m < k. This means that {um } is uniformly
Cauchy in K and, hence, it converges uniformly in K to some real-valued func-
tion.
If we, now, define u(x) = limm+ um (x) for all x , then um u
uniformly on compact subsets of . By the third property of harmonic functions,
u is harmonic in .

1.12 Normal families of harmonic functions

If we drop the assumption of monotonicity, the results are not that clear, but
we still get some normal families-type of results.
Theorem 1.17 If {um } is a sequence of harmonic functions in the open
Rn and um u uniformly on compact subsets of , then the derivatives of the
um converge to the corresponding derivatives of u uniformly on compact subsets
of .
Proof:
Take B(x0 ; R) and observe that, if x B(x0 ; 12 R), then B(x; 12 R)
B(x0 ; R). By Theorem 1.15(1),
2n
|grad(um u)(x)| sup |um u| .
R B(x0 ;R)

xj xj uniformly on B(x0 ; 2 R) and, since x0 is arbitrary,

Therefore, u m u 1

the convergence is uniform on all compact subsets of .

By Theorem 1.3, all derivatives of harmonic functions are harmonic and, by
induction, we can prove uniform convergence on compact sets for all derivatives.

Theorem 1.18 Let U be a family of harmonic functions in the open Rn

which are uniformly bounded on compact subsets of .
Then, their derivatives are also uniformly bounded on compact subsets of
and from every sequence in U we can extract a subsequence converging uniformly
on compact subsets of to some harmonic function.
Proof:
Take B(x0 ; R) . Then, there is some M = M (x0 , R) so that

|u(x)| M

for every x B(x0 ; R) and every u U.

By the same argument as in the proof of the previous theorem, we get that
2n 2n
|grad u(x)| sup |u| M
R B(x0 ;R) R

for every x B(x0 ; 12 R) and every u U.

92 CHAPTER 1. HARMONIC FUNCTIONS

Since x0 is arbitrary, we conclude that the derivatives of first order of the

functions in U are uniformly bounded on compact subsets of .
From the mean value theorem of the differential calculus,
2n
|u(x) u(x0 )| M |x x0 |
R
for every x B(x0 ; 12 R) and every u U. Therefore, U is equicontinuous (and
bounded) at every point in .
Now, take any compact exhaustion {K(m) } of .
Given {uk } in U, we use the Arzela-Ascoli theorem for each K(m) to extract
a subsequence converging uniformly on K(m) . Then, by the usual diagonal ar-
gument, we find a subsequence converging uniformly on every K(m) and, hence,
on every compact subset of . By the third property of harmonic functions, the
limit function is harmonic in .

Definition 1.11 Suppose F is a family of extended-real-valued functions de-

fined in the set E. Then, the function

F (x) = sup f (x) , xE ,

f F

is called the upper envelope of the family F.

The lower envelope is similarly defined.

Theorem 1.19 Suppose U is a family of positive harmonic functions in the

connected open Rn . Then, the upper envelope of the family is either iden-
tically + in or everywhere finite and continuous in .
In the first case, there exists a sequence in U diverging to + uniformly on
compact subsets of .
In the second case, from every sequence in U we can extract a subsequence
converging uniformly on compact subsets of to some harmonic function.
There is a dual statement for lower envelopes of families of negative har-
monic functions.

Proof:
Let U be the upper envelope of U.
(1) If there is some x0 with U (x0 ) = +, then there exists {um } in U so
that um (x0 ) +. By Theorem 1.14 and in the same manner as in the proof
of Theorem 1.16, we prove that um (x) + uniformly on compact subsets of
and, thus, U = + everywhere in .
(2) Now, suppose that U (x0 ) < + for some x0 .
Again by Theorem 1.14, for every compact K , there is a C = Cn,K, > 0
so that
u Cu(x0 ) CU (x0 )
everywhere in K for all u U and, thus,

U CU (x0 )
1.13. HARMONIC DISTRIBUTIONS 93

in K.
This implies that U is bounded on every compact subset of and, in par-
ticular, everywhere finite in . By Theorem 1.18, we immediately get that
from every sequence in U we can extract a subsequence converging uniformly
on compact subsets of to some harmonic function.
It, only, remains to prove the continuity of U .
Consider x0 and take {um } in U so that um (x0 ) U (x0 ). Then, for
every m, by the continuity of um ,
um (x0 ) = lim inf um (x) lim inf U (x)
xx0 xx0

and, letting m +,
U (x0 ) lim inf U (x) .
xx0

Consider M = lim supxx0 U (x).

Then, there exist xm x0 so that U (xm ) M and, by the definition of U ,
there exists {um } in U so that um (xm ) M .
Extracting, if necessary, a subsequence, we may assume that um converges
uniformly on compact subsets of to some u harmonic in . Then,
M = lim um (xm ) = u(x0 ) = lim um (x0 ) U (x0 )
m+ m+

and, thus,
lim sup U (x) U (x0 ) ,
xx0

implying the continuity of U at x0 .

The hypotheses of Theorem 1.19 can be slightly weakened. Instead of the

positivity of all the functions in U, it is enough to assume that on every compact
subset of the family U is uniformly bounded from below.
We then work with an open exhaustion {(m) } of . In each (m) the family
is uniformly bounded from below by some constant and we may apply Theorem
1.19 there. The passage from the (m) s to presents absolutely no difficulty,
except that we must apply a diagonal argument when we extract subsequences.
The interested reader may, easily, complete the details.

1.13 Harmonic distributions

And, now, we prove the famous
Theorem 1.20 (Lemma of H. Weyl and L. Schwartz) If T is a distribution in
the open Rn and T = 0, then T is identified with a harmonic function u
in . This means T = Tu or, more specifically,

T () = (x)u(x) dm(x)

for every D().

94 CHAPTER 1. HARMONIC FUNCTIONS

Proof:
Consider an approximation to the identity { : > 0} .
Fix > 0 and consider variable 1 , 2 < 12 .
Then T 1 and T 2 are both distributions in 21 and (T 1 ) 2 =
(T 2 ) 1 is a distribution in .
Observe that, by Proposition 0.16, all these distributions are identified with
infinitely differentiable functions.
By Proposition 0.18,
(T 1 ) = T 1 = 0
as a distribution in 21 .
Let v be the infinitely differentiable function which represents T 1 in 12 .
Then Tv = Tv = 0, implying v = 0 and, hence, v is harmonic in 21 .
From the proof of Theorem 1.2, we get that
v 2 = v in .
This implies, of course,
(T 1 ) 2 = Tv 2 = Tv2 = Tv = T 1 in
and, hence,
(T 2 ) 1 = T 1 in .
Now, let 1 0 and get
T 2 = T in .
The same argument applied to T 2 concludes that T 2 and, hence, T
is identified, as a distribution in , with some function u harmonic in .
If 0 < , then in , which is smaller than 0 , we have that u and u0
represent the same distribution. Therefore, u0 is an extension of u and, since
>0 = , we conclude that all the u s define a single u harmonic in which
T is identified to.

Suppose f is a locally integrable function in whose distributional Laplacian

vanishes in . I.e.
f (x)(x) dm(x) = 0

for all D().
The Lemma of Weyl and Schwartz implies that there exists some u harmonic
in so that Tf = Tu and, hence,
f (x) = u(x)
for almost every x .
In other words, we can change f in at most a set of measure zero and make
it harmonic in .
If f is continuous to begin with, then it is identical to u and, hence, it is
harmonic.
1.13. HARMONIC DISTRIBUTIONS 95

Proposition 1.6 If the distributional Laplacian of a locally integrable function

f in the open Rn is zero, then f is almost everywhere equal to a harmonic
function in .
If, in addition, f is continuous in , then it is harmonic in .
96 CHAPTER 1. HARMONIC FUNCTIONS
Chapter 2

Superharmonic Functions

2.1 Definition
Let be open in Rn .
Definition 2.1 A function u is called superharmonic in , if
1. u is lower-semicontinuous in ,
2. u is not identically + in any connected component of and
3. Mru (x) u(x) for all x and all r < d(x, ).
The function u is called subharmonic, if u is superharmonic.
Condition 3 is called the super-mean-value property while the corre-
sponding condition for subharmonic functions is called the sub-mean-value
property.
Comments
1. A superharmonic function is extended-real-valued and may take the value
+, but not the value .
2. If u is superharmonic in , then u(x) = lim inf yx u(y) for all x .
In fact, let lim inf yx u(y) > u(x) and consider a number between these
two quantities. Take so that u(y) > for all y B(x; ). Then Mru (x) >
u(x) for all r < , a contradiction to the definition.
3. From Proposition 0.2, we get that Mru (x) is well-defined, either as a real
number or as +, for all x and all r for which S(x; r) .
4. We, easily, see that limr0+ Mru (x) = u(x).
In fact, take < u(x) and, by the lower-semicontinuity, find > 0 so that
u(y) for all y B(x; ). Then u(x) Mru (x) for all r < .
5. If a function u satisfies conditions 1 and 3 of the definition, but it is identically
+ in some connected components of , then we may drop these components
and form the set as the union of the remaining connected components of .
Since all components are open sets, is open and u is superharmonic in .

97
98 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Properties of superharmonic functions

(1) Linear combinations with non-negative coefficients of superharmonic func-
tions are superharmonic and the same is true for subharmonic functions.
(2) The minimum of finitely many superharmonic functions is superharmonic.
There is a dual statement for subharmonic functions.
The lower-semicontinuity is taken care of by property (4) of lower-semiconti-
nuous functions. As for the super-mean-value property, if u1 , . . . , uk have it and
u = min(u1 , . . . , uk ), then for every j we have Mru (x) Mruj (x) uj (x) and,
hence, Mru (x) u(x).
(3) Increasing limits of superharmonic functions are superharmonic, dropping,
if necessary, the connected components where the limits are identically +.
There is a dual statement for subharmonic functions.
In fact, let {um } be an increasing sequence of superharmonic functions in
and let um (x) u(x) for all x . The lower-semicontinuity of u comes from
the third property of lower semi-continuous functions.
Also, for every m, Mrum (x) um (x) u(x) and we prove Mru (x) u(x),
using Proposition 0.2 and the Monotone Convergence Theorem.
(4) If u is superharmonic in the open Rn and if u has a local minimum at
x , then u is constant in some open neighborhood of x.
A dual result is true for subharmonic functions.
Suppose u(x) u(y) for all y B(x; R), where B(x; R) .
Then, for all r < R, u(x) Mru (x) u(x) and for equality to hold we must
have u(y) = u(x) for all y S(x; r), except for at most a set E S(x; r) of zero
surface measure.
Now, take y E. Then, there is some {ym } in S(x; r) \ E so that ym y.
Hence, u(x) u(y) lim inf zy u(z) limm+ u(ym ) = u(x) and we get
that u(y) = u(x) for all y S(x; r).
Therefore, since r is arbitrary with r < R, we conclude that u is constant in
B(x; R).
(5) Superharmonic and subharmonic functions are preserved by rigid motions
of Rn .
The proof is exactly the same as the proof of property (2) of harmonic
functions.

2.2 Minimum principle

Theorem 2.1 (The Minimum Principle for superharmonic functions) Suppose
u is superharmonic in the open Rn .

1. If u takes its minimum value at some x , then u is constant in the

connected component of which contains x.

2. If ( )
m = inf lim inf u(x) ,
y 3xy

then m u(x) for all x .

2.2. MINIMUM PRINCIPLE 99

If m = u(x) for some x, then u is constant, u = m, in the connected

component of which contains x.

Proof:
The proof is an application of the fourth property of superharmonic functions
and of Proposition 0.5 and is identical to the proof of Theorem 1.1.

The next result provides a characterization of superharmonicity and it is

fundamental.

Theorem 2.2 Suppose u is lower-semicontinuous in the open Rn and it

is not identically + in any connected component of . Then the following are
equivalent.

1. u is superharmonic in .

2. For every B(x0 ; R) and every v harmonic in B(x0 ; R), the validity
of lim inf B(x0 ;R)3xy (u(x) v(x)) 0 for all y S(x0 ; R) implies that
u v in B(x0 ; R).

Proof:
Since v being harmonic implies that v is superharmonic, one direction is a
trivial application of Theorem 2.1.
Now, take arbitrary B(x0 ; R) .
Proposition 0.3 implies that there exist fm continuous in S(x0 ; R) so that

fm (y) u(y)

for all y S(x0 ; R).

Consider Pfm ( ; x0 , R), the Poisson integral of fm in B(x0 ; R). Then, by
Theorem 1.9,

lim inf (u(x) Pfm (x; x0 , R)) u(y) fm (y) 0

B(x0 ;R)3xy

for all y S(x0 ; R).

By the hypothesis,

u(x0 ) Pfm (x0 ; x0 , R) = MR

fm (x0 ) .

Finally, by the Monotone Convergence Theorem and Proposition 0.2,

u(x0 ) MR
u (x0 )

and u is superharmonic in .

If u is superharmonic in an open set containing a ball B(x0 ; R), then it is

bounded from below on the ball and, hence, the Poisson integral of u at every
point of B(x0 ; R) is well-defined either as a real number or as +.
100 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Proposition 2.1 Let u be superharmonic in an open set containing B(x0 ; R).

Consider Pu ( ; x0 , R), the Poisson integral in B(x0 ; R) of the restriction of u
on S(x0 ; R). Then,
1. Pu (x; x0 , R) u(x) for every x B(x0 ; R).
2. Either Pu (x; x0 , R) = + for every x B(x0 ; R) or Pu ( ; x0 , R) is har-
monic in B(x0 ; R).
In particular, if MR
u (x0 ) = +, then u = + identically in B(x0 ; R).

Proof:
Consider fm continuous in S(x0 ; R) so that fm (y) u(y) for every y
S(x0 ; R) and their Poisson integrals Pfm ( ; x0 , R). Then,

lim inf (u(x) Pfm (x; x0 , R)) u(y) fm (y) 0

B(x0 ;R)3xy

for all y S(x0 ; R) and, from Theorem 2.1 or Theorem 2.2, we get

Pfm (x; x0 , R) u(x)

for every x B(x0 ; R). By the Monotone Convergence Theorem and the posi-
tivity of the Poisson kernel, we prove statement 1.
Let < m u(y) for every y S(x0 ; R) and write v = u m.
Then, either (i) MR u (x0 ) = + or (ii) m Mu (x0 ) < + and, hence,
R

either (i)Mv (x0 ) = + or (ii) 0 Mv (x0 ) < +.

R R

In case (i): let x B(x0 ; R) and consider k = minyS(x0 ;R) P (y; x, x0 , R) >
0. Then, since 0 v(y) for all y S(x0 ; R),

u(x) Pu (x; x0 , R) = Pv (x; x0 , R)+m k v(y) dS(y)+m = + .
S(x0 ;R)

In case (ii): Pu (x; x0 , R) = Pv (x; x0 , R) + m is harmonic in B(x0 ; R), since

v is integrable in S(x0 ; R).

2.3 Blaschke-Privaloff parameters

Definition 2.2 If f is extended-real-valued and f (x) is a real number, we de-
fine:
2n ( r ) 2(n + 2) ( r )
M f (x) = lim sup 2
Mf (x)f (x) , Af (x) = lim sup 2
Af (x)f (x)
r0+ r r0+ r
and
2n ( r ) 2(n + 2) ( r )
M f (x) = lim inf 2
Mf (x)f (x) , Af (x) = lim inf 2
Af (x)f (x) ,
r0+ r r0+ r
whenever the mean values that appear are defined for all small enough r.
These four numbers are called Blaschke-Privaloff parameters of f at x.
2.3. BLASCHKE-PRIVALOFF PARAMETERS 101

Lemma 1.1 says that, if f is twice continuously differentiable in some neigh-

borhood of x, then all four Blaschke-Privaloff parameters of f at x are equal to
f (x).
Lemma 2.1 If f is extended-real-valued and f (x) is real, the four Blaschke-
Privaloff parameters of f at x satisfy:

M f (x) Af (x) Af (x) M f (x) .

Proof:
The middle inequality is obvious and it is enough to prove the third one,
since the first is implied by this, using f in place of f .
In case M f (x) = +, the result is obvious. Therefore, assume M f (x) <
+.
Let M f (x) < , implying that, for some R,

2n ( r )
2
Mf (x) f (x)
r
for all r < R.
Then, for r < R,

2(n + 2) ( r ) 2(n + 2) ( n r s )
2
Af (x) f (x) = 2 n
Mf (x) sn1 ds f (x)
r r r 0

2(n + 2) n r s2 n1
s ds
r2 rn 0 2n
= .

Hence,
Af (x) .
Letting M f (x), we conclude that

Af (x) M f (x) .

The following is a general characterization of superharmonicity. Observe, to

begin with, that the Blaschke-Privaloff parameters are well-defined for a lower-
semicontinuous u, whenever u(x) < +.
Theorem 2.3 (Blaschke and Privaloff ) If u is lower-semicontinuous in the
open Rn and not identically + in any connected component of , then
u is superharmonic in if and only if M u (x) 0 for all x with u(x) < + .
Proof:
The necessity is trivial and, for the sufficiency, take arbitrary B(x0 ; R)
and consider fm continuous in S(x0 ; R), so that fm (y) u(y) for all y
S(x0 ; R).
102 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Consider, also, the auxiliary function

w(x) = |x x0 |2 R2 .

Using w, define, for every  > 0,

{
u(x) Pfm (x; x0 , R)  w(x) , x B(x0 ; R)
v(x) =
u(x) fm (x) , x S(x0 ; R) .

Then, v is lower-semicontinuous in B(x0 ; R) and, thus, takes a minimum

value in there. Its values on S(x0 ; R) are non-negative and, hence, if we assume
that its minimum value is < 0, then it is taken at some x B(x0 ; R).
But, then M v (x) 0, while M v (x) = M u (x) 0  2n  2n.
We, thus, get a contradiction and conclude that v(x) 0 for all x B(x0 ; R).
Now, letting first  0 and then m +, we find for x = x0 ,

u(x0 ) MR
u (x0 ) .

Corollary 2.1 Suppose u is lower-semicontinuous in the open Rn and not

identically + in any connected component of .
1. If, for every x , there is some sequence {rm (x)} so that rm (x) 0 and
r (x) r (x)
Mum (x) u(x) or Aum (x) u(x) for all m, u is superharmonic in .
2. If u is in C (), u is superharmonic in if and only if u(x) 0 for all
2

x .

In view of the extra regularity, Corollary 2.1(2) has an additional proof which
uses Greens Formula, in exactly the same way as in the proof of the similar
Theorem 1.3.
In fact, we use the formula

d
u(x) dm(x) = n1 rn1 Mru (x)
B(x;r) dr

which was derived in that proof.

If u 0 identically in , then, by the above formula, Mru (x) is decreasing
in the interval 0 < r < d(x, ) and, taking the limit as r 0+, we find
Mru (x) u(x) for every x and every r < d(x, ).
If, conversely, Mru( (x) u(x) for)every x and every r < d(x, ), then
u(x) = limr0+ 2n r 2 Mu (x) u(x) 0.
r

We, thus, get a weakened version of the original definition of superharmoni-

city, while the next result is the version of the original definition, having the
surface-mean-values replaced by the space-mean-values.
2.3. BLASCHKE-PRIVALOFF PARAMETERS 103

Theorem 2.4 Suppose u is lower-semicontinuous in the open Rn and not

identically + in any connected component of .
Then u is superharmonic in if and only if, for every x and all r <
d(x, ),
Aru (x) u(x) .

Proof: r
The necessity follows from Aru (x) = rnn 0 Msu (x) sn1 ds.
For the sufficiency, we may observe that the above assumption, together
with Lemma 2.1, implies M u (x) Au (x) 0 and the proof is concluded by
Theorem 2.3.
Or, else, we may use Theorem 2.2 as follows. Take B(x0 ; R) and
v harmonic in B(x0 ; R) so that lim inf B(x0 ;R)3xy (u(x) v(x)) 0 for all
y S(x0 ; R).
Now, Aru (x) u(x) implies

Ar(uv) (x) (u v)(x) ,

for all x B(x0 ; R) and all r < R |x x0 |.

But, from this, in the usual manner (we repeat the proof of the fourth
property of superharmonic functions; this is even simpler), we get that, if u v
takes a minimum value at some point in B(x0 ; R), then it is constant in a
neighborhood of this point.
Proposition 0.5 implies that u v everywhere in B(x0 ; R).
By Theorem 2.2, u is superharmonic in .

Example
If
u(x) = |x| , x Rn \ {0} ,
then, using Corollary 2.1(2) and the formula

d2 u n 1 du
u(x) = 2
(|x|) + (|x|)
dr r dr
which holds for all twice continuously differentiable radial functions, we find
that u is superharmonic if and only if 2 n 0 and subharmonic if and
only if 2 n or 0 .

Theorem 2.5 Suppose u is superharmonic in B(x0 ; R) \ {x0 } .

Then, u can be defined at x0 so that it becomes superharmonic in B(x0 ; R)
if and only if
u(x)
lim inf 0.
xx0 hx0 (x)

Proof:
The nessecity comes from the fact that if u is superharmonic in B(x0 ; R),
then it is bounded from below in some neighborhood of x0 .
104 CHAPTER 2. SUPERHARMONIC FUNCTIONS

For the sufficiency, fix r < R and consider functions fm continuous in S(x0 ; r)
so that fm (y) u(y) for every y S(x0 ; r).
We, also, consider the function
( )
(x) = u(x) Pfm (x; x0 , r) +  hx0 (x) h (r) , x B(x0 ; r) \ {x0 } ,

for an arbitrary  > 0.

By Theorem 1.9, we have that is superharmonic in = B(x0 ; r) \ {x0 }
and, also by the hypothesis, that lim inf 3xy (x) 0 for every y .
Theorem 2.1 implies that (x) 0 for all x B(x0 ; r) \ {x0 }. Hence,
( )
u(x) Pfm (x; x0 , r)  hx0 (x) h (r)

for all x B(x0 ; r) \ {x0 } and every  > 0 and, finally,

u(x) Pfm (x; x0 , r) , x B(x0 ; r) \ {x0 } .

By the Dominated Convergence Theorem,

u(x) Pu (x; x0 , r)

for all x B(x0 ; r) \ {x0 } and, thus,

lim inf u(x) Pu (x0 ; x0 , r) = Mru (x0 ) .

xx0

Now, since r is arbitrary, if we define u(x0 ) = lim inf xx0 u(x), we immedi-
ately conclude that u is lower-semicontinuous on B(x0 ; R) and an application
of Corollary 2.1(1) concludes the proof.
In fact, the extended u satisfies the super-mean-value property at x0 , and,
since it coincides with the original superharmonic u in B(x0 ; R)\{x0 }, it satisfies
the super-mean-value property at all other points of B(x0 ; R) with respect to
small enough balls centered at these points.

Theorem 2.6 If u is superharmonic in the open Rn , then,

1. u(x) < + for almost every x and
2. for every B(x; r) ,

Aru (x) < + , Mru (x) < + .

Therefore, u is locally integrable in .

Proof:
It is enough to work separately in the various connected components of
and, hence, suppose that is connected.
There exists some x so that u(x) < +. Therefore, Aru (x) < + for
r < d(x, ) and, hence, u is integrable in B(x; r) for all these r.
2.4. POISSON MODIFICATION 105

Define

A = {x : u is integrable in some neighborhood of x} ,

B = {x : u is not integrable in any neighborhood of x} .

A is open and non-empty and we shall prove that B is, also, open.
Take x0 B. Then, Aru (x0 ) = + for r < d(x0 , ).
Now, for all x B(x0 ; 2r ) it is true that x0 B(x; 2r ) and, hence,
r
u(x) Au2 (x) = + .

Therefore, B(x0 ; 2r ) B.
Since is connected, u is integrable in a neighborhood of any point in .
This implies that u(x) < + for almost every x . It, also, implies that
Aru (x) < + for every x and all r < d(x, ). In fact, as we proved above,
if Aru (x) = + for some x and some r, then u = + in B(x; 2r ).
If, for some B(x; r) , we have Mru (x) = +, then, from Proposition 2.1,
u = + identically in B(x; r) and this is false.

2.4 Poisson modification

Theorem 2.7 Suppose u is superharmonic in the open subset of Rn and
B(x0 ; R) . Define
{
Pu (x; x0 , R) , if x B(x0 ; R)
uB(x0 ;R) (x) =
u(x) , if x \ B(x0 ; R) .

Then

1. uB(x0 ;R) u in and

2. uB(x0 ;R) is superharmonic in and harmonic in B(x0 ; R).

Proof:
The first part and the harmonicity of uB(x0 ;R) in B(x0 ; R) are consequences
of Proposition 2.1 and Theorem 2.6.
That < uB(x0 ;R) (x) for all x , is obvious.
Now, take fm continuous in S(x0 ; R) so that fm (y) u(y) for all y
S(x0 ; R). Then, for any y S(x0 ; R),

fm (y) = lim inf Pfm (x; x0 , R) lim inf Pu (x; x0 , R)

B(x0 ;R)3xy B(x0 ;R)3xy

and, letting m +,

u(y) lim inf Pu (x; x0 , R) = lim inf uB(x0 ;R) (x) .

B(x0 ;R)3xy B(x0 ;R)3xy
106 CHAPTER 2. SUPERHARMONIC FUNCTIONS

We, also, have

u(y) lim inf u(x) lim inf uB(x0 ;R) (x) .

3xy \B(x0 ;R)3xy

Therefore,
u(y) lim inf uB(x0 ;R) (x)
3xy

for all y S(x0 ; R) and, thus, uB(x0 ;R) is lower-semicontinuous in .

As for the super-mean-value property, this is obvious, by the harmonicity of
the function in B(x0 ; R), at points x B(x0 ; R) with respect to small enough
balls centered at x.
If x \ B(x0 ; R), then

uB(x0 ;R) (x) = u(x) Mru (x) MruB(x (x) .

0 ;R)

Definition 2.3 Suppose that is an open subset of Rn and B(x0 ; R) . If

u is superharmonic or subharmonic in , then the function
{
Pu (x; x0 , R) , if x B(x0 ; R)
uB(x0 ;R) (x) =
u(x) , if x \ B(x0 ; R)

is called the Poisson modification of u with respect to B(x0 ; R).

Proposition 2.2 Superharmonicity is a local property: if u is superharmonic

in a neighborhood of every point of the open Rn , then it is superharmonic
in .

Proof:
It is immediate from Corrolary 2.1(1).

Example
Let f be holomorphic in the open R2 and not identically 0 in any
connected component of . Then the function log |f | is superharmonic in .
log |f | is defined as + at the points where f = 0. In fact, these points
are isolated and log |f | is not identically + in any connected component of
.
log |f | is, trivially, lower-semicontinuous in , it is, by Proposition 1.2,
harmonic in a neighborhood of every point at which f 6= 0 and it satisfies
the super-mean-value property at every x where f (x) = 0, simply because
log |f (x)| = +.

Example
Suppose u is real-valued and harmonic in the open Rn and is concave
in the real interval (m, M ), where m = inf u and M = sup u. Then u is
superharmonic in .
2.5. POTENTIALS 107

Without loss of generality, we suppose that is connected. If u = m or

u = M at some point, then u is constant and, no matter how is defined at
the endpoints m, M , u is constant and, hence, superharmonic. Therefore,
we may assume that m < u < M in .
is continuous in (m, M ), implying that the composition is also continuous.
As for the super-mean-value property, by the inequality of Jensen,
( )
u(x) = Mru (x) Mru (x)

for every x and all r < d(x, ).

Example
If > 0, p 1 and u is real-valued and harmonic in the open Rn , then
u , u , eu and |u|p are subharmonic in .
+

Example
Suppose u is superharmonic in the open Rn and is increasing and
concave in the real interal (m, M ], where m = inf u and M = sup u. Observe
that, necessarily, (M ) = limtM (t). Then u is superharmonic in .
Except for minor modifications, the proof is the same as the proof in the
second example.

Example
If u is subharmonic in the open Rn and > 0, then u+ , eu are
subharmonic in .
Also, if p > 0 and f is holomorphic in the open R2 and not identically
0 in any connected component of , then |f |p is subharmonic in . This is a
special case of the last example, when we use the increasing convex function
(t) = ept and the subharmonic log |f |.

Example
The function hx0 is superharmonic in Rn .
The function is continuous in Rn \ {x0 } and limxx0 hx0 (x) = +. Hence,
hx0 is lower-semicontinuous in Rn .
It is harmonic in Rn \{x0 } and, thus, satisfies the super-mean-value property
at every x 6= x0 with respect to all sufficiently small balls centered at x.
It satisfies the super-mean-value property, also, at x0 simply because its
value there is +.

2.5 Potentials
The next example is, in a sense, the most general and deserves to be stated as
a theorem.
108 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Theorem 2.8 Suppose d is any compactly supported non-negative Borel mea-

sure. Then, its h-potential

Uhd (x) = hx (y) d(y) , x Rn ,
Rn
n
is superharmonic in R .
In case n > 2, we get the same result for any non-negative Borel measure
d, assuming only that Uhd (x) < + for at least one point x.
In any case, Uhd is harmonic in Rn \ supp(d).
Proof:
Suppose that d has compact support in Rn .
The statement about harmonicity is the content of Proposition 1.4 and,
because of this harmonicity, Uhd is not identically + in Rn .
If d is such that supp(d) B(0; d), then, for any B(x0 ; r),

|h(x y)| d(y) dm(x)
B(x0 ;r) supp(d)

= |h(x y)| dm(x) d(y)
supp(d) B(x0 ;r)

= |h(x)| dm(x) d(y)
supp(d) B(x0 y;r)

|h(x)| dm(x) d(y)
supp(d) B(0;|x0 |+d+r)
= |x0 |+d+r,n d(Rn ) < + ,
where t,n is a finite number depending only on t and n.
Therefore, we may use Fubinis Theorem and, from the last example of
section 2.4,

ArU d (x0 ) = Arhy (x0 ) d(y)
h
K

hy (x0 ) d(y)
K
= Uhd (x0 ) .
By Lemma 1.3, there is another way to prove the super-mean-value property.

MrU d (x0 ) = Mrhy (x0 ) d(y)
h supp(d)

= h (r) d(y)
supp(d)B(x0 ;r)

+ hx0 (y) d(y)
supp(d)\B(x0 ;r)

Uhd (x0 ) ,
2.6. DIFFERENTIABILITY OF POTENTIALS 109

where the use of the Theorem of Fubini is justified by the proof of Lemma 1.3.
It remains to prove that Uhd is lower-semicontinuous and let xm x.
If |xm x| 1, then |xm y| |x| + d + 1 for all m and all y supp(d).
Therefore, the functions hxm are all bounded from below in supp(d) by the
same constant and an application of the Lemma of Fatou gives

lim inf hxm (y) d(y) hx (y) d(y) .
m+ supp(d) supp(d)

Thus, Uhd is lower-semicontinuous.

When n > 2, the function h is positive everywhere and, in this case, we
may freely interchange integrations and apply Fatous Lemma without having
to assume that d is supported in a compact set.
Or, in a different way, consider the restrictions dB(0;m) of d in the balls
B(0; m).
d
By the first part, Uh B(0;m) is superharmonic in Rn and harmonic in Rn \
( )
B(0; m) supp(d) and, hence, in Rn \ supp(d).
d
Since h is positive, Uh B(0;m) Uhd everywhere in Rn . By the assumption,
Uhd is not identically + and, finally, by the third property of superharmonic
functions and Theorem 1.16, Uhd is superharmonic in Rn and harmonic in
Rn \ supp(d).

Later on we shall prove the fundamental theorem of F. Riesz stating that

the most general superharmonic function is, more or less, the sum of a harmonic
function and the h-potential of a non-negative Borel measure.

2.6 Differentiability of potentials

Proposition 2.3 Under the hypotheses of Theorem 2.8, Uhd is absolutely con-
tinuous on almost every line parallel to the principal xj -axes, 1 j n, it has
partial derivatives at almost every point of Rn and these partial derivatives are
locally integrable.

Proof:
We fix an arbitrary m N and we shall work in the cube

Qm = {x = (x1 , . . . , xn ) : |xj | m for all j} .

In case n = 2, the measure is supported in a compact set. In case n 3 the

measure need not be supported in a compact set, but we may split d = dQ2m +
dRn \Q2m and observe that the h-potential of the second term is harmonic in
Qm and, hence, infinitely differentiable there.
We may, therefore, assume that d is supported in some compact set and,
in particular, that it is finite.
110 CHAPTER 2. SUPERHARMONIC FUNCTIONS

By a trivial calculation, there is a constant Cn , depending only on the di-

mension, so that
 h 
 y  Cn
 (x)
xj |x y|n1
for all x, y Rn .
Therefore,
 
 hy  1
 (x) d(y) dm(x) Cn dm(x) d(y)
Qm Rn xj Rn Qm |x y|n1
< Cn,m d(Rn ) < + .

Thus, the function

hy
uj (x) = (x) d(y) , x Qm ,
Rn xj
is integrable in Qm and, by Fubinis Theorem, it is integrable on almost every
line segment parallel to the xj -axis and extending between the two sides of Qm
which are perpendicular to this axis.
If [a, b] is any part of such a segment, then
b b
hy
uj (x) dxj = (x) dxj d(y)
n xj
a
R a
( )
= hy (b) hy (a) d(y)
Rn
= Uhd (b) Uhd (a) .

This says that Uhd is absolutely continuous on almost every line segment
parallel to the xj -axis and extending between the two sides of Qm and that
Uhd
xj = uj almost everywhere on such a line segment.
U d
By Fubinis Theorem, again, xhj is defined and equal to uj almost every-
where in Qm .
Since m is arbitrary, the proof is finished.

Definition 2.4 Suppose that the non-negative function f is locally integrable in

Rn . The h-potential of f is defined to be the h-potential of the non-negative
Borel measure f dm. We denote it by

Uhf = Uhf dm .

According to Theorem 2.8, if n = 2, we assume that f is compactly supported

and, if n 3, we assume only that Uhf (x) < + for at least one x.

Proposition 2.4 Under the hypotheses of the previous definition, if the non-
negative function f is in C k (Rn ), 0 k +, then the h-potential of f is in
C k+1 (Rn ).
2.6. DIFFERENTIABILITY OF POTENTIALS 111

Proof:
If n = 2, then f has compact support. If n 3, then f need not have
compact support, but, since we want to study the h-potential in a neighborhood
of an arbitrary point x, we may take a large R so that x B(0; R) and split
f = f + f (1 ), where is in D(Rn ) with = 1 identically in B(0; 2R) and
0 1 everywhere. The existence of is due to Lemma 0.2.
Then, the h-potential of f (1) is harmonic in B(0; R) and, hence, infinitely
differentiable in a neighborhood of x. Thus, we need to study the h-potential
of the function f which is in C k (Rn ) and has compact support.
We, therefore, assume that f has compact support.
Since h is locally integrable, by Proposition 0.6, the convolution Uhf = h f
is in C k (Rn ) and
( )
D Uhf = h D f
for all multiidices with || k.
Hence, the proof reduces to showing that, if f is in C(Rn ) with compact
support, then h f is in C 1 (Rn ).
We observe that
h xj
(x) = (n 2) n
xj |x|
for all x 6= 0 and, thus, h
xj is locally integrable.
We write, now,

h f (x + tej ) h f (x) h(x y + tej ) h(x y)
= f (y) dm(y) .
t Rn t
If z is not on the xj -axis, then

h(z + tej ) h(z) 1
h
= (z + stej ) ds
t 0 xj
and, hence,

h f (x + tej ) h f (x) 1
h
= (x y + stej ) ds f (y) dm(y)
t Rn 0 xj
1
h
= (x y + stej )f (y) dm(y) ds
0 Rn xj
1
h
= (y)f (x y + stej ) dm(y) ds .
0 Rn xj
Therefore, if |t| 1,
 h f (x + te ) h f (x) 
 j h 
 f (x)
t xj
1  
 h 
 (y) |f (x y + stej ) f (x y)| dm(y) ds
0 Rn xj
 h 
 
 (z) dm(z) sup |f (a) f (b)|
B(x;1)+supp(f ) xj |ab||t|
112 CHAPTER 2. SUPERHARMONIC FUNCTIONS

and, thus,
(h f ) h
= f .
xj xj
h
Since x j
is locally integrable and f is continuous with compact support, by
Proposition 0.6, the last convolution is continuous.

2.7 Approximation, properties of means

Theorem 2.9 (First property of the means) Let u be superharmonic in the open
Rn . Then, for every x , Mru (x) is a decreasing function of r in the
interval 0 < r < d(x, ). We, also, have that limr0+ Mru (x) = u(x).
Exactly the same results hold for the space-means Aru (x).

Proof:
Consider r1 < r2 < d(x, ) and the Poisson integral Pu ( ; x, r2 ) in the ball
B(x; r2 ). From Proposition 2.1, u() Pu ( ; x, r2 ) in this ball and, hence,

Mru1 (x) MrP1u ( ;x,r2 ) (x) = Pu (x; x, r2 ) = Mru2 (x) .

The equality limr0+ Mru (x) = u(x) is just Comment 4 after the definition
of superharmonic functions.
For the space-means,

n r1 r
Au (x) =
r1
Mu (x)rn1 dr
r1n 0

n r2 r1
r
= n Mur2 (x)rn1 dr
r2 0

n r2 r
Mu (x)rn1 dr
r2n 0
= Aru2 (x) ,

where in the last inequality we used the result about the surface-means.
The limit limr0+ Aru (x) = u(x) is trivial and can be proved in the same
way as the limit of the surface-means above.

Corollary 2.2 If u and v are superharmonic in the open Rn and u(x) =

v(x) for almost every x , then u and v are identically equal in .

Theorem 2.10 (Approximation) Suppose u is superharmonic in the open

Rn and { : > 0} is an approximation to the identity. Then, for every ,
the function u C ( ) is superharmonic in and,

u (x) u(x)

as 0, for every x .
2.7. APPROXIMATION, PROPERTIES OF MEANS 113

Proof:
From Proposition 0.7 and Theorem 2.6 we find that u is in C ( ).
For every x and B(x; R) ,

1
Mu (x) =
R
u(y z) (z) dm(z) dS(y)
n1 Rn1 S(x;R) B(0;)

1
= n1
u(y z) dS(y) (z) dm(z)
B(0;) n1 R
S(x;R)

u(x z) (z) dm(z)
B(0;)
= u (x) ,

where the last inequality is true because n11Rn1 S(x;R) u(y z) dS(y) is the
surface-mean of the function u( z) which is superharmonic in the open set
z containing B(x; R) whenever z B(0; ).
Therefore, u is superharmonic in .
Now, since is radial,

u (x) = n1 (r)Mru (x) rn1 dr .
0

By Theorem 2.9, for any < u(x) we have < Mru (x) u(x) for small
enough r. Thus, if is small,

n1 (r) rn1 dr u (x) u(x)n1 (r) rn1 dr ,
0 0

implying
u (x) u(x) .
We conclude that u (x) u(x) as 0.
Also, taking < 0 ,

u (x) = n1 (r)Mru (x) rn1 dr
0

1 r
= n1 n
1 ( )Mru (x) rn1 dr
0
0
1 r
r
= n1 0n
1 ( 0 )Mu0 (x) rn1 dr
0
0
1 r
n1 0n
1 ( 0 )Mru (x) rn1 dr
0
= u 0 (x) .

The last inequality is true by Theorem 2.9.

114 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Proposition 2.5 Consider the open subsets 1 and 2 of R2 , the holomorphic

function f in 1 and let f (1 ) 2 . If u is superharmonic in 2 , then u f
is superharmonic in 1 .
The only exception is when f is constant in some connected component of
1 and u takes the value + at this value of f .

Proof:
Assuming that u is twice continuously differentiable in 2 , we use the formula
(u f )(z) = |f 0 (z)|2 u(f (z)) and Corollary 2.1(2) to prove the result.
Otherwise,
( )consider an arbitrary closed disc B(z0 ; R) 1 and the compact
set f B(z0 ; R) 2 .
We consider a bounded open set V so that B(z0 ; R) V V and
using Theorem 2.10, we approximate u in V by an increasing sequence {um } of
functions which are twice continuously differentiable and superharmonic in V .
Then, the functions um f are, by the first part, superharmonic in B(z0 ; R)
and increase towards u f there.
Thus, by the third property of superharmonic functions, u f is superhar-
monic in B(z0 ; R) and, since the ball is arbitrary, u f is superharmonic in 1 .

Example
If u is superharmonic in the open R2 which does not contain 0 and if
in the set = {x : x1 } we define the function
(1)
u (x) = u ,
x
then u is superharmonic in .

Theorem 2.11 (Second property of the means) Let u be superharmonic in the

open Rn and B(x; R1 , R2 ) . Then,

Mru (x) < +

for all r with R1 < r < R2 and Mru (x) is a concave function of h(r) in the
interval R1 < r < R2 .
In particular, Mru (x) is a continuous function of r in the same interval.

Proof:
Assume, first, that u is in C 2 ().
Then, from Corrolary 2.1(2), u 0 everywhere in .
Consider R1 < r1 < r2 < R2 and apply Greens Formula in B(x; r1 , r2 ) with

being the continuous unit vector field normal to B(x; r1 , r2 ) in the direction
towards the exterior of this ring.

u u
u(z) dm(z) = (y) dS(y) + (y) dS(y)
B(x;r1 ,r2 ) S(x;r2 ) S(x;r1 )

d( )
= r2n1 u(x + ry) r=r2 d(y)
S n1 dr
2.7. APPROXIMATION, PROPERTIES OF MEANS 115

d( )
r1n1 u(x + ry) r=r1 d(y)
S n1 dr
n1 d
( )
= r2 n1 Mru (x) r=r2
dr
d( )
r1n1 n1 Mru (x) r=r1 .
dr

Now, u(z) dm(z) is a decreasing function of r2 and, hence,
B(x;r1 ,r2 )
d
rn1 dr Mru (x) is a decreasing function of r. Therefore,

d ( n1 d )
r Mru (x) 0
dr dr
in R1 < r < R2 .
If we write h = h(r), the last relation becomes

d2
Mh (x) 0 ,
dh2 u
implying
Mth
u
1 +(1t)h2
(x) tMhu1 (x) + (1 t)Mhu2 (x)
for all t (0, 1) and h1 = h(r1 ), h2 = h(r2 ) with R1 < r1 < r2 < R2 .
In the general case, we use the approximation Theorem 2.10 to get a se-
quence of functions um superharmonic and twice continuously differentiable in
B(x; r10 , r20 ), with R1 < r10 < r1 < r2 < r20 < R2 and such that um u in
B(x; r10 , r20 ).
We, then, apply the last inequality to each um and prove it for u by the
Monotone Convergence Theorem.
If we assume that, for some r0 (R1 , R2 ), Mru0 (x) = +, then, using the
above inequality, it is easy to show that Mru (x) = + for all r in (R1 , R2 ).
Taking R1 < r1 < r2 < R2 , we get

u(y) dm(y) = +
B(x;r1 ,r2 )

contradicting the local integrability of u which was proved in Theorem 2.6.

Thus,
Mru (x) < +
for all r in (R1 , R2 ), finishing the proof of the concavity.

Theorem 2.12 Suppose u is superharmonic in the open subset of Rn and

let B(x0 ; R) .
Then, the only function which is superharmonic in , harmonic in B(x0 ; R)
and coincides with u in \ B(x0 ; R) is the Poisson modification uB(x0 ;R) .
Also, uB(x0 ;R) is the upper envelope of the family of functions v which are
superharmonic in , harmonic in B(x0 ; R) and satisfy v u everywhere in .
116 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Proof:
Consider any v with all properties in the first part of the statement.
From the definition of the Poisson modification, we have

vB(x0 ;R) = uB(x0 ;R)

in and, hence, it is enough to prove that v = vB(x0 ;R) in B(x0 ; R).

From Theorem 2.7, we have that v vB(x0 ;R) in B(x0 ; R) and, if we prove
that v(x0 ) = vB(x0 ;R) (x0 ), then, since both functions are harmonic in B(x0 ; R),
the Maximum-Minimum Principle will finish the proof.
From the harmonicity of v and of vB(x0 ;R) in B(x0 ; R) and from Theorem
2.11,
v(x0 ) = lim Mrv (x0 ) = MR v (x0 ) = vB(x0 ;R) (x0 )
rR

and the proof of the first part is complete.

If v satisfies the assumptions of the second part, then, by the first part,

v = vB(x0 ;R) uB(x0 ;R)

in .

2.8 The Perron process

In the proof of the next result we introduce the important Perron process.

Theorem 2.13 Let V be a non-empty family of functions v subharmonic in the

open connected Rn with the following two properties

1. If v1 , v2 V, then max(v1 , v2 ) V.

2. If v V and B(x0 ; R) , then vB(x0 ;R) V.

Then, the upper envelope V of the family V is either identically + in or

it is harmonic in .
There is a dual result about lower envelopes of families of superharmonic
functions satisfying the duals of properties 1 and 2.

Proof:
Fix an arbitrary B(x0 ; R) and consider a countable set {xi : i N}
dense in B(x0 ; R).
(m) (m)
For each xi , take a sequence {vi } in the family V so that vi (xi ) V (xi )
as m + .
Modify, defining
(m) (1) (m)
ui = max(vi , . . . , vi ), mN.
(m) (m)
{ui } is an increasing sequence in V with ui (xi ) V (xi ) as m + .
2.9. THE LARGEST HARMONIC MINORANT 117

Modify the new sequence, defining

(1) (m) (m)
w(1) = u1 , w(m) = max(u1 , u2 , . . . , u(m)
m ) .

Now, {w(m) } is an increasing sequence in V with w(m) (xi ) V (xi ) for all xi .
Modify once more, taking the sequence

vm = (w(m) )B(x0 ;R) .

This is a new increasing sequence in V such that vm (xi ) V (xi ) for all xi ,
with the additional property that all vm are harmonic in B(x0 ; R).
Set
v(x) = lim vm (x)
m+

for all x in B(x0 ; R).

By Theorem 1.16, either v is harmonic in B(x0 ; R) or v = + identically in
B(x0 ; R). It is obvious that, if v = + in B(x0 ; R), then the same is true with
V.
Suppose, for the moment, that v is harmonic in B(x0 ; R) and, by its con-
struction, v(xi ) = V (xi ) for all xi .
Consider, now, a point x B(x0 ; R) different from all xi and repeat the
previous construction with the set {x} {xi : i N}. A new function v 0 will
be produced, harmonic in B(x0 ; R) with v 0 (xi ) = V (xi ) for all xi and, also,
v 0 (x) = V (x).
The functions v, v 0 which are continuous in B(x0 ; R) agree on the dense set
{xi : i N} of B(x0 ; R) and, hence, are identically equal on this ball. Therefore,
v(x) = v 0 (x) = V (x) at the additional point x. Since x is arbitrary, this proves
that V = v identically in B(x0 ; R) and, finally, that V is harmonic in B(x0 ; R).
Now, we define the sets

A = {x : V is harmonic in some neighborhood of x} ,

B = {x : V = + in some neighborhood of x} .

By what we proved before, = A B, A B = and both A and B are

open sets. Since is connected, either A = or B = .

2.9 The largest harmonic minorant

Definition 2.5 If f, g are extended-real-valued functions defined on the same
set E and f (x) g(x) for all x E, we say that f is a minorant of g and
that g is a majorant of f in E.
If the same inequality is true for all f in a family F, we say that g is a
majorant of F in E and, if the inequality is true for all g in a family G, we say
that f is a minorant of G in E.
118 CHAPTER 2. SUPERHARMONIC FUNCTIONS

Theorem 2.14 Let U be a non-empty family of functions superharmonic in the

open Rn . Suppose that there exists at least one subharmonic minorant of
U in .
Then the upper envelope of all subharmonic minorants of U is a function
harmonic in .

Proof:
Let V be the non-empty family of all subharmonic minorants of the family
U and let V be the upper envelope of V.
If u U, then V u and, hence, V (x) < + for almost every x .
Since it is very easy to see that the family V satisfies the assumptions of
Theorem 2.13 in all connected components of , we conclude that V is harmonic
in .

Definition 2.6 Let U be a non-empty family of superharmonic functions in the

open Rn and suppose that U has at least one subharmonic minorant in .
Then the upper envelope of all the subharmonic minorants of U is called the
largest harmonic minorant of U.
We, similarly, define the smallest harmonic majorant of a non-empty
family of subharmonic functions.
As an example, we prove the
Proposition 2.6 Let u be superharmonic in B(x0 ; R). Then u has at least one
subharmonic minorant if and only if limrR Mru (x0 ) > .
If this condition is satisfied, then the largest harmonic minorant of u is the
function limrR uB(x0 ;r) .
Proof:
If u has some subharmonic minorant v in B(x0 ; R), then, by Theorem 2.6,
for every r < R, Mru (x0 ) Mrv (x0 ) > . Since, by Theorem 2.9, the
left side decreases while the right side increases when r R, we see that
limrR Mru (x0 ) > .
Now, let limrR Mru (x0 ) > and fix some r0 < R.
It is clear from Theorem 2.7, that, when r0 < r < R, the functions uB(x0 ;r)
are superharmonic in B(x0 ; R), harmonic in B(x0 ; r0 ) and decrease as r in-
creases.
Since uB(x0 ;r) (x0 ) = Mru (x0 ), Theorem 1.16 implies that the function

V = lim uB(x0 ;r)

rR

is harmonic in B(x0 ; r0 ).
Since r0 is arbitrary, V is harmonic in B(x0 ; R).
Clearly, V uB(x0 ;r) u for all r < R and, hence, V is a harmonic minorant
of u in B(x0 ; R).
Suppose, now, that v 0 is any subharmonic minorant of u in B(x0 ; R). We,
easily, see that Theorem 2.7 implies v 0 vB(x 0
0 ;r)
uB(x0 ;r) for all r < R.
2.10. SUPERHARMONIC DISTRIBUTIONS 119

Therefore v 0 V in B(x0 ; R) and, finally, V is, exactly, the largest harmonic

minorant of u in B(x0 ; R).

2.10 Superharmonic distributions

The rest of this chapter is devoted to the study of the distributional Laplacian
of superharmonic functions and the relevant characterization of them and to the
proof of the related Decomposition Theorem of F. Riesz.
In the statement of the next theorem we denote the distributional Laplacian
of the function u by u, instead of the more correct Tu , but this, as we have
already noted, is allowed by the standard convention to identify a function with
the corresponding distribution. In the proof, though, we shall be more formal.

Proposition 2.7 If u is superharmonic in the open Rn , then its dis-

tributional Laplacian u is a non-positive distribution in and, hence, it is
identified with a non-positive Borel measure in .

Proof:
Let u be superharmonic in and consider any approximation to the identity
{ : > 0}. From Theorem 2.10, we know that u is in C ( ) and that
it is superharmonic in .
Corrolary 2.1(2) implies that (u )(x) 0 for all x .
Take, now, any D() with 0 everywhere in . Since supp()
if is small enough, using Proposition 0.9 and the calculus of distributions, we
find
( )
Tu () = lim Tu ( f ) = lim Tu ( f )
0+ 0+

= f ) =
lim Tu ( lim (Tu )()
0+ 0+
= lim Tu () = lim Tu ()
0+ 0+

= lim T(u ) () = lim (u )(x)(x) dm(x)
0+ 0+
0.

Therefore, Tu 0 and the last statement is a consequence of Theorem

0.11.

The next two results are the main examples.

Proposition 2.8 hx0 = n dx0 .

Proof:
Consider any D(Rn ). In the following calculations the third equality is
true because the integrand is in L1 (Rn ), the fourth equality is an application
of Greens Formula in the set B(x0 ; r, R) for some R which is large enough
120 CHAPTER 2. SUPERHARMONIC FUNCTIONS

so that and vanish on S(x0 ; R) and the sixth equality is true because
n1
limr0+ r h (r) = 0, grad is bounded in a neighborhood of x0 and is
continuous at x0 .

Thx0 () = Thx0 () = hx0 (x)(x) dm(x)
Rn

= lim hx0 (x)(x) dm(x)

r0+ Rn \B(x ;r)
(
0

= lim (y)hx0 (y) dS(y)
r0+ S(x0 ;r)
)
hx0
(y) (y) dS(y)
S(x0 ;r)
( )
d
= lim rn1 h (r) (x0 + rz) d(z) + n Mr (x0 )
r0+ S n1 dr
= n (x0 )
= n Tdx0 () .

Therefore, Thx0 = n Tdx0 or, less formally,

hx0 = n dx0 .

Theorem 2.15 Let d be a compactly supported non-negative Borel measure

and consider the superharmonic function Uhd , the h-potential of d. Then,

Uhd = n d .

In case n > 2, the non-negative Borel measure d need not be compactly

supported and we, only, assume that Uhd (x) < + for at least one x.
Proof:
Using the formal notation for distributions, for every D(Rn ),

TU d () = Uhd (x)(x) dm(x)
h
Rn

= h(x y)(x) dm(x) d(y)
supp(d) Rn

= Thy () d(y)
supp(d)

= Thy () d(y)
supp(d)

= n Tdy () d(y)
supp(d)
2.10. SUPERHARMONIC DISTRIBUTIONS 121

= n (y) d(y)
supp(d)
= n Td () .

The use of the Theorem of Fubini, justifying the second equality, is permitted
by the calculation at the beginning of the proof of Theorem 2.8 and the fifth
equality is just Proposition 2.8.
In case n > 2, even if supp(d) is not compact in Rn , the use of Fubinis The-
orem is still justified, since, by the positivity of h, the compactness of supp()
and the local integrability of the superharmonic Uhd ,

|h(x y)| d(y) |(x)| dm(x)
supp() Rn

M Uhd (x) dm(x) < + ,
supp()

where M is a bound of ||.

Observe that this result agrees with the fact that Uhd is harmonic in the set
Rn \ supp(d).

Theorem 2.16 Let T be a distribution in the open Rn . Then, T is iden-

tified with a superharmonic function in if and only if T 0 .

Proof:
One direction is just Proposition 2.7.
Hence, suppose T 0 and apply Theorem 0.11 to get a non-negative Borel
measure d in so that
1
Td = T .
n
Consider, now, an arbitrary open G so that G is a compact subset of , the
restriction dG of d in G and its h-potential UhdG .
If we define the distribution

SG = T TU dG ,
h

then, taking Laplacians, by Theorem 2.15, we have

SG = T n TdG = 0

as a distribution in G.
By Theorem 1.20, SG is identified with some harmonic function in G and,
hence, T = SG + TU dG is identified with a superharmonic function in G.
h
Now, consider an open exhaustion {(m) } of and apply the previous result
to G = (m) . For each m, there is a superharmonic um in (m) so that T = Tum
in (m) . By Proposition 0.10, um = um+1 in (m) and, hence, the um s define
122 CHAPTER 2. SUPERHARMONIC FUNCTIONS

a single function u superharmonic in which, for every m, coincides with um

on (m) .
Now, for every D() we have D((m) ) for large enough m and,
thus,

T () = Tum () = um (x)(x) dm(x) = u(x)(x) dm(x) = Tu () .
(m)

Hence, T is identified with u in .

2.11 The theorem of F. Riesz

Theorem 2.17 (Decomposition Theorem of F. Riesz) Suppose u is superhar-
monic in the open Rn . Then, there exists a unique non-negative Borel
measure d in so that, for every open G with G being a compact subset of ,

u(x) = UhdG (x) + vG (x)

for every x G, where dG is the restriction of d in G and vG is some

harmonic function in G.
In the case n > 2, we, also, have

u(x) = Uhd (x) + v(x)

for all x for some v harmonic in , provided Uhd (x) < + for at least
one x Rn .

Proof:
Most of the work was done in the proof of the previous theorem. In fact,

Tu = SG + TU dG ,
h

where SG is identified with a harmonic function, say vG , in G. I.e.

Tu = TvG + TU dG ,
h

as distributions in G.
Proposition 0.10 implies

u(x) = vG (x) + UhdG (x)

for almost every x G and, finally, Corollary 2.2 implies that the equality is
true everywhere in .
The uniqueness is proved by taking distributional Laplacians in G. By The-
orem 2.15,
Tu = n TdG
2.12. DERIVATIVES OF SUPERHARMONIC FUNCTIONS 123

in G. From this, the distribution TdG is uniquely determined in G, and, hence,

the restriction dG is, also, uniquely determined. Taking as Gs the terms of
any open exhaustion of , we prove the uniqueness of d in .
In case n > 2, we may adjust the proof of the previous Theorem.
Consider the non-negative Borel measure d in so that
1
Td = Tu .
n

Assuming Uhd (x) < + for at least one x, Theorem 2.15 implies that the
distribution S = Tu TU d satisfies S = 0 in and, by Theorem 1.20, S is
h
identified with a harmonic v in .
From this, as before,

u(x) = v(x) + Uhd (x)

for almost every x and, hence, for every x .

Or, in another way, we may start from the restricted result of the first part
applied to the terms of an open exhaustion {(m) } of ,
d(m)
u(x) = Uh (x) + v(m) (x) , x (m) .

For every x , because of the positivity of h,

d(m)
Uh (x) Uhd (x)

and the limit is finite for almost every x Rn . Therefore, {v(m) } is (eventually)
a decreasing sequence of harmonic functions in every fixed (k) with a limit v
which is finite almost everywhere in (k) . Therefore, v is harmonic in (k) and

u(x) = Uhd (x) + v(x)

everywhere in (k) and, since k is arbitrary, everywhere in .

2.12 Derivatives of superharmonic functions

Theorem 2.18 Let u be superharmonic in the open Rn . Then, u is
absolutely continuous on almost every line parallel to the principal xj -axes, 1
j n, it has partial derivatives at almost every point of and these partial
derivatives are locally integrable in .
If the distributional Laplacian of u is a (non-negative) function f C k (),
then u is in C k+1 ().

Proof:
This is a trivial application of the Representation Theorem of F. Riesz and
of Propositions 2.3 and 2.4.
124 CHAPTER 2. SUPERHARMONIC FUNCTIONS
Chapter 3

The Problem of Dirichlet

3.1 The generalized solution

Lets remember that the boundary and the closure of subsets of Rn is taken
relative to Rn .
Definition 3.1 Let be an open subset of Rn and f any extended-real-valued
function defined in .
We denote by f the family of all functions u defined in with the proper-
ties:
1. in each connected component of either u is superharmonic or u is iden-
tically +,
2. u is bounded from below in and
3. lim inf 3xy u(x) f (y) for all y .

The lower envelope of the family f is denoted by H f .
In a dual manner, f denotes the family of all v defined in such that:
1. in each conected component of either v is subharmonic or v is identically
,
2. v is bounded from above in and
3. lim sup3xy v(x) f (y) for all y .

The upper envelope of the family
f is denoted by H f .

Both families are non-empty, since + f and f .

Comment It is easy to see that if G is any connected component of the open set
G
Rn , then the restriction of H f in G coincides with H f and the restriction
of H G
f in G coincides with H f .
This allows us, in many problems, to reduce the study of these functions to
the case of the set being connected.

125
126 CHAPTER 3. THE PROBLEM OF DIRICHLET

Proposition 3.1 In each connected component of , H f is either harmonic or
identically + or identically .
The same is true for H f.

Proof:
Consider the family V of the subharmonic restrictions of the functions in
f in any particular connected component G of .
In case all functions in f are identically in G, then V is empty and,
obviously, H f = in G.

In case there is at least one function in f with subharmonic restriction to

G, then V is non-empty and it is almost obvious that V satisfies the assumptions
in Theorem 2.13 for the connected open G. Therefore, the upper envelope of V
in G, which coincides with the restriction of H f in G, is either harmonic in G
or identically + in G.

Proposition 3.2 H f H
f everywhere in .

Proof:
Fix an arbitrary connected component G of .

In case either H f = + everywhere in G or H f = everywhere in G,
then the result is obvious. Hence, assume that there is some u f which is
not + everywhere in G and some v f which is not everywhere in G.
Then, u v is superharmonic in G and
( )
lim inf u(x) v(x) 0
G3xy

for all y G.
This last inequality is obvious for every y for which f (y) is real. In case
f (y) = +, then the boundedness from above of v is used and, if f (y) = ,
then the boundedness from below of u has to be used.
From Theorem 2.1, we find that u v in G. Since u is arbitrary in f and

f , we conclude that H f H f in G.

v is arbitrary in

Definition 3.2 Let the extended-real-valued f be defined in , where Rn

is open.

1. If H f = H
f everywhere in , then this common function is denoted by

Hf .

2. f is called resolutive with respect to if

Hf = H
f
3.1. THE GENERALIZED SOLUTION 127

everywhere in and if this common function is not identically + or

in any connected component of .
This common function Hf is harmonic in and is called the generalized
solution to the Problem of Dirichlet in with boundary function
f.

To motivate this definition, suppose that f is real-valued in and that

the Problem of Dirichlet in with boundary function f is solvable. Therefore,
there exists some u harmonic in so that lim3xy u(x) = f (y) for all y .
The first thing to observe is that f is, then, continuous on . In fact, take
any sequence {ym } in with ym y. Then, there is another {xm } in
so that dS (xm , ym ) < m1
and |u(xm ) f (ym )| < m1
. The first inequality gives
dS (xm , y) 0 and, hence, u(xm ) f (y). Thus, the second inequality gives
f (ym ) f (y).
Since f is continuous in , it is bounded in and Theorem 1.1 implies
that u is bounded in .
Therefore, u f f and, hence,

u H
f Hf u .

This implies
u = Hf
in .
Summarizing: if the Problem of Dirichlet in with the real-valued boundary
function f is solvable, then f is continuous in and resolutive with respect
to and the solution to the Problem of Dirichlet coincides with the generalized
solution.
Therefore, to solve the Problem of Dirichlet, we must suppose that the given
boundary function f is continuous in and, then, prove

1. that f is resolutive and

2. that lim3xy Hf (x) = f (y) for all y .

In the case of any bounded open , item 1 will be answered completely

by a theorem of Wiener, which we shall prove in a while, and says that every
continuous f is resolutive.
On the other hand, even if f is continuous, the Problem of Dirichlet may
not be solvable. Here is an instructive example.

Example
Consider = {x Rn : 0 < |x| < 1} and
{
0 , if |y| = 1
f (y) =
1 , if y = 0 .
128 CHAPTER 3. THE PROBLEM OF DIRICHLET
( )
Take um (x) = 1
m h(x) h (1) for all x . Then um
f and, hence,

H f (x) lim um (x) = 0 .
m+

On the other hand, 0

f , implying

0 H
f (x) .

Thus, f is resolutive and continuous in , with Hf = 0 everywhere in ,

but
lim Hf (x) = 0 6= f (0) = 1 .
3x0

3.2 Properties of the generalized solution

In all statements that follow, the boundary functions are all extended-real-
valued.

f +c = H f + c for all c R.
1. H f +c = H f + c and H

2. H f = H f and H
f = H f for > 0 and H f = H f for < 0.
Under the convention 0() = 0, these formulas hold in case = 0 also.

3. If f1 f2 , then H f1 H f2 and H
f1 H f2 .

4. inf f H
f H f sup f .
All these properties are trivial to prove.

Convention When we add boundary functions f and g, we assign any

value, whatsoever, to the indeterminate forms () + ().

But, when we add H f and H g , we assign the value + to these forms
f and H g , we assign the value to them.
and, when we add H

In any case, we assign the value 0 to 0().

5. H f +g H f + H g and H
f + H g H f +g .

Working in each connected component of separately, it is enough to

assume that is connected. Observe that, by our convention, the first

inequality is trivial in case at least one of H f and H g is + identically
in .
Hence, take any u
f and any v g which are not identically +.


Then, u + v f +g and, thus, H f +g u + v. This is enough to conclude
the proof of the first inequality and the second is proved similarly.
3.2. PROPERTIES OF THE GENERALIZED SOLUTION 129

6. If both f and g are resolutive with respect to , then cf + dg is resolutive

+dg = cHf + dHg everywhere in , for all c, d R.

and Hcf
This is a simple combination of properties 2 and 5.

7. Let fm f and H f1 be not identically in a connected component G

of . Then H fm H f in G.
The dual result is, also, true.

It is enough to assume that = G is connected and that, for all m, H fm

is not identically + in . Therefore, we assume that H fm is harmonic
in , for all m.
Fix x0 and consider um
fm , superharmonic in with


um (x0 ) H fm (x0 ) + .
2m

Now, um H fm is superharmonic and non-negative in and, by the third
property of superharmonic functions, the series

+
( )
um H fm
m=1

either converges to a superharmonic function in or it diverges to +

everywhere in . But its value at x0 is  and, thus, it is superharmonic
in .

By Theorem 1.16, the limm+ H fm is either + identically in or it
is harmonic in . In the first case, what we want to prove is clear and we
assume that this limit is harmonic in .
Now, the function

+
( )
w = lim H fm + um H fm
m+
m=1

is superharmonic in . Furthermore, for every m,

( )
w H fm + um H fm = um .

Therefore, w is bounded from below in and, also,

lim inf w(x) lim inf um (x) fm (y)

3xy 3xy

for all y and, since m is arbitrary,

lim inf w(x) f (y)

3xy
130 CHAPTER 3. THE PROBLEM OF DIRICHLET

for all y .

Therefore, w
f , implying that w H f . Then

H f (x0 ) w(x0 ) lim H fm (x0 ) + 
m+

and, since  is arbitrary,

H f (x0 ) lim H fm (x0 ) .
m+

The inequality limm+ H fm H f is obvious and , by Theorem 1.1,

lim H fm = H f
m+

in .

8. Let G be an open subset of and f a boundary function in . If we

consider {
f, on G ,
F =
H f , on G ,
G
then H F = H f everywhere in G.
There is a dual result for H.
Since every connected component of G is contained in one of the connected
components of , it is enough to assume that both G and are connected.
If u
f , then, in case y G ,

lim inf u(x) lim inf u(x) f (y) = F (y) ,

G3xy 3xy

while, in case y G ,

lim inf u(x) u(y) H f (y) = F (y) .
G3xy

G
Hence, u G
F , implying u H F in G and, finally,

G
Hf HF

in G.

The opposite inequality is clear in case H f = identically in .

In case H f = + identically in , then F (y) = + for all y G .
Now, taking any u G
F , we have that lim inf G3xy u(x) = + for all
y G .
3.2. PROPERTIES OF THE GENERALIZED SOLUTION 131

Extending u as identically + on \ G, we either get a superharmonic

u in (if u is superharmonic in G) or the function + in (if u = +
identically in G). But, the first alternative is impossible, since this would
imply the existence of a superharmonic function in f.
Therefore, u = + identically in G and we get that
G
H F = + H f
in G.

Finally, suppose that H f is harmonic in .
Take any u G
F and define
{ ( )
min u, H f , in G,
V =
Hf , in \ G
which is superharmonic in .

Take arbitrary v
f and consider the function V + v H f which is
superharmonic and bounded from below in .
In case y \ G,
( )
lim inf V (x) + v(x) H f (x) = lim inf v(x) f (y) .
3xy 3xy

In case y G,
( )
lim inf V (x) + v(x) H f (x) = lim inf v(x) f (y)
\G3xy \G3xy

and ( )

lim inf V (x) + v(x) H f (x) f (y) .
G3xy

Hence, V + v H f H f in .

This implies that V H f in , which gives u H f in G and, finally,
G
HF Hf
in G.

9. Let G be an open subset of and f a boundary function on . If f is

resolutive with respect to , then the function
{
f, on G ,
F = H , on G ,
f

is resolutive with respect to G and HFG = Hf identically in G.

132 CHAPTER 3. THE PROBLEM OF DIRICHLET

10. Let each fm be resolutive and fm f uniformly in .

Then f is resolutive and Hfm Hf uniformly in .
Fix  > 0. For large m, fm  f fm +  everywhere in and, hence,

Hfm  H f H f Hfm +  in . Thus, 0 H f H f 2 and,

since  is arbitrary,

Hf = Hf

in and f is resolutive.
Also, f  fm f +  in implies Hf  Hfm Hf +  in ,
from which we get the uniform convergence Hfm Hf in .

3.3 Wieners Theorem

Lemma 3.1 Let be a bounded open subset of Rn . Then, every real-valued f
continuous in can be uniformly approximated in by the difference of the
restrictions in of two functions continuous in and superharmonic in .

Proof:
Since is a compact subset of Rn , by the Stone-Weierstrass Theorem, we
can approximate f uniformly in by a real-valued polynomial P (x1 , . . . , xn ).
Take a constant M > 0 so that P (x) M for all x .
Now, the difference of the two functions, P (x1 , . . . , xn ) M
2 x1 and 2 x1 ,
2 M 2

approximates f uniformly on , while both of them are superharmonic in .

Theorem 3.1 (N. Wiener) Let be a bounded open subset of Rn . Then, every
real-valued f continuous in is resolutive with respect to .
Proof:
Let the real-valued F be continuous in and superharmonic in .

Then, obviously, F F and, thus, F H F in . By property 4 in section

3.2, H F is bounded and harmonic in and, for every y ,

lim sup H F lim sup F (x) = F (y) ,
3xy 3xy

implying that H F
F . Hence,

HF = H
F

and the restriction of F in is resolutive with respect to .

By Lemma 3.1, there is a sequence {Fm Gm } so that all Fm and Gm are
continuous in , superharmonic in and the restrictions in of Fm Gm
converge to f uniformly in .
From the first part and property 6, the restrictions in of all Fm Gm are
resolutive with respect to . Therefore, from property 10, f is resolutive with
3.4. HARMONIC MEASURE 133

respect to .

Example
Let = R2 \ {0} and f (0) = 0, f () = 1.
If u
f , then, for any  > 0 we have that, for all large enough M and N ,

log |x| + log N

u(x) 
log M + log N

for every x with N1 |x| M . Let N + and then M + and get that
u(x) 1  for all x 6= 0, . Since  is arbitrary, we find u 1 and, thus,

H f = 1 in .
If v
f , then, for any  > 0 we have that, for all large enough M and N ,

log |x| + log N

v(x) +
log M + log N

for every x with N1 |x| M . Let M + and then N + and get that
v(x)  for all x 6= 0, . Since  is arbitrary, we find u 0 and, thus, H
f =0
in .
Therefore, f is not resolutive with respect to and Theorem 3.1 cannot be
extended to hold for unbounded open sets in R2 .

3.4 Harmonic measure

Let be a bounded open set and x0 . By Theorem 3.1,

C() 3 f 7 Hf (x0 ) C

defines a (complex-)linear functional on C().

This is, at first, defined for real-valued f , but it is, trivially, extended to
complex-valued f by Hf (x0 ) = H<f
(x0 ) + iH=f (x0 ).
It is easy to prove that
H1 (x0 ) = 1 ,
Hf (x0 ) 0
for all f C() with f 0 in and

|Hf (x0 )| kf k

for all f C(). In fact, the last inequality is straightforward for real-valued
f , while, for complex-valued f , we take so that Hf (x0 ) = ei |Hf (x0 )| and
write

|Hf (x0 )| = ei Hf (x0 ) = Hei f (x0 )

i
i f ) (x0 ) k<(e f )k kf k .

= H<(e
134 CHAPTER 3. THE PROBLEM OF DIRICHLET

Therefore, H() (x0 ) is a non-negative bounded linear functional on C()
with norm 1 and, from Theorem 0.10, there exists a unique non-negative Borel
measure d
x0 supported in so that

d
x0 () = 1

and
Hf (x0 ) = f (y) d
x0 (y)

for all f C().

The fact that dx0 is a non-negative Borel measure with total mass equal to
1 is described by calling it a Borel probability measure.
By the process of Caratheodory, the measure d x0 can be considered as
uniquely extended on the -algebra of its measurable sets. This -algebra is
larger than B() and a set A belongs to this -algebra if and only if
A = B N for some Borel set B and some N with d x0 (N ) = 0.
Also, dx0 is complete on the -algebra of its measurable sets.

Definition 3.3 Let Rn be a bounded open set and x0 .

The complete probability measure d x0 in , constructed above, whose -
algebra of measurable sets includes all Borel sets in and satisfies


Hf (x0 ) = f (y) dx0 (y)

for all f C() is called the harmonic measure in with respect to

and x0 .

Lemma 3.2 Suppose is a bounded open subset of Rn and consider f lower-

semicontinuous in . Then,

1. For every x0 ,

Hf (x0 ) = f (y) d
x0 (y) .

2. f is resolutive with respect to if and only if it is d

x0 -integrable for at
least one x0 in every connected component of .
The same is true, if f is upper-semicontinuous.

Proof:
Consider a sequence {fm } of functions continuous in with fm (y) f (y)
for every y .
By Theorem 3.1,

fm (x0 ) H f (x0 ) .
H fm (x0 ) = H
3.4. HARMONIC MEASURE 135

By property 7 of H, the continuity of fm and the Monotone Convergence

Theorem,


H f (x0 ) = lim H fm (x0 ) = lim fm (y) d
x0 (y) = f (y) d
x0 (y) .
m+ m+

From these two relations we get the first result.

Now, since
f is bounded from below in , we have that, for every x0 ,

f (y) d x0 (y) < + if and only if f is d
x0 -integrable.
If f is resolutive with respect to , then all integrals are finite and f is
dx0 -integrable for all x0 . On the other hand, if f is dx0 -integrable for at


least one x0 in some connected component, then Hf (x0 ) is finite and, hence,
Hf is harmonic in the same component.

Theorem 3.2 Suppose is a bounded open subset of Rn . For every extended-

real-valued function f defined in ,
1.


H f (x0 ) = f (y) d
x0 (y) , H
f (x0 ) = f (y) d
x0 (y) ,

for all x0 ,
2. f is resolutive with respect to if and only if it is d
x0 -integrable for at
least one x0 in every connected component of and
3. f is resolutive with respect to if and only if it is dx0 -integrable for every
x0 in and, in this case,

Hf (x0 ) = f (y) dx0 (y) ,

for all x0 .
In particular, an E is d x0 -measurable for every x0 in if and only
if E is resolutive with respect to and, in this case,

HE (x0 ) = d
x0 (E)

for every x0 in .
Proof:
1. It is enough to prove the first equality and we, first, see that, by Lemma 3.2,


f (y) d (y) dx0 (y) = inf H (x0 ) H f (x0 ) ,

x0 (y) = inf

where the infima are taken over all lower-semicontinuous in with f

everywhere in .
136 CHAPTER 3. THE PROBLEM OF DIRICHLET

The opposite inequality is obvious if H f (x0 ) = + and, thus, assume that

H f (x0 ) < +. Now, take arbitrary > H f (x0 ) and u
f so that

u(x0 ) .
The function defined by (y) = lim inf 3xy u(x) for all y is lower-
semicontinuous in and satisfies f there. Hence,

f (y) dx0 (y)

(y) dx0 (y) = H (x0 ) u(x0 )

and, since is arbitrary, we get

x0 (y) H f (x0 ) .
f (y) d

2 and 3. It is obvious, from 1, that if f is resolutive with respect to , then it

is dx0 -integrable for every x0 .
If f is d x0 -integrable for some x0 in , then, from 1, the two functions,

Hf and H f , are harmonic in the connected component of which contains
x0 and, by the Maximum-Minimum Principle, they are identically equal in the
same component.

Theorem 3.3 Suppose is a bounded open subset of Rn and x1 , x2 are in the

same connected component of . Then
1. d
x1 and dx2 have the same zero-sets.

2. d
x1 -measurable sets and functions are the same as the dx2 -measurable
sets and functions.
3. L1 (, d 1
x1 ) = L (, dx2 ) and the norms in these two spaces are equiv-
alent.
Proof:
1. Let d
x1 (N ) = 0. From Theorem 3.2,

0 = d x1 (N ) = N (y) d
x1 (y) = HN (x1 )

and, from the Maximum-Minimum Principle, HN = 0 identically in the con-

nected component of containing x1 . Now, the same set of equalities for x2
instead of x1 give that d x2 (N ) = 0.
2. If E is d x1 -measurable, then E = B N , for some Borel set B and some
N with d x1 (N ) = 0. From part 1, we get that d x2 (N ) = 0 and, hence, E is

dx2 -measurable.
3. If f is d
x1 -integrable, then


H|f (x
| 1 ) = |f (y)| d
x1 (y) < + .

3.4. HARMONIC MEASURE 137

Therefore, H|f | is harmonic in the connected component of containing x1
and, then
|f (y)| d
x2 (y) = H|f | (x2 ) < +

implies that f is d
x2 -integrable.
If f is dx -integrable for some x and, hence, for all x in the same
connected component G of , then the function

|f (y)| d
x (y) = H|f | (x) , xG,

is harmonic and non-negative in G. From Harnacks Inequalities we have that

for every x1 , x2 G there is a constant C, depending only on these two points
and on G, so that

1
|f (y)| d
x2 (y) |f (y)| d
x1 (y) C |f (y)| d
x2 (y)
C

for all f .

Proposition 3.3 Let be a bounded open subset of Rn .

If G is one of the connected components of and x0 G, then,

x0 ( \ G) = 0 .
d

Hence, the harmonic measure with respect to and x0 is supported in the

boundary of the connected component of which contains x0 .
Proof:
In fact, consider the function
{
0 , if x G
u(x) =
1 , if x \ G .

u is harmonic in and lim inf 3xy u(x) \G (y) for all y .

Therefore,

0 d
x0 ( \ G) H \G (x0 ) u(x0 ) = 0 .

Example
If f is defined and continuous on the sphere S(x0 ; R), then Pf ( , x0 ; R) is
the solution to the Problem of Dirichlet in B(x0 ; R) with boundary function f .
Hence,
B(x0 ;R)
Hf (x) = f (y)P (y; x, x0 , R) dS(y)
S(x0 ;R)

for all x B(x0 ; R).

138 CHAPTER 3. THE PROBLEM OF DIRICHLET

From the definition of harmonic measure,

dB(x
x
0 ;R)
= P ( ; x, x0 , R) dS .
B(x ;R)
This means that dx 0 and susface measure on S(x0 ; R) are mutually
absolutely continuous and the density-function (Radon-Nikodym derivative) of
B(x ;R)
dx 0 with respect to dS is exactly the Poisson kernel P ( , x, x0 ; R).
In particular,
1
dB(x
x0
0 ;R)
= dS .
n1 Rn1
The harmonic measure with respect to the ball and its center is the normalized
surface measure on its surface.

3.5 Sets of zero harmonic measure

Definition 3.4 Suppose that is a bounded open subset of Rn .
A set E is said to be of zero harmonic measure with respect to
x (E) = 0 for every x or, equivalently, for at least one x in every
, if d
connected component of .

Theorem 3.4 Suppose that is a bounded open subset of Rn and let E .

Then, the following are equivalent.
1. There exists a function u superharmonic in with u 0 in and
lim3xy u(x) = + for every y E.
2. E is of zero harmonic measure with respect to .

Proof:
1. Assume the existence of a u with the properties in the statement of the
1
theorem. Then m E for every m N and, hence, for every x ,
u

1
0 H
E (x) H E (x) u(x) .
m
Now, let m + and get

d
x (E) = HE (x) = 0

for every x .
2. Assume d xk (E) = 0 for at least one xk in each of the at most countably
many connected components k of . Therefore

HE (xk ) = 0

for the same points. This implies that, for every k and every m, there exists a
um,k E so that
1
um,k (xk ) m+1 .
2
3.5. SETS OF ZERO HARMONIC MEASURE 139

Now, modify these functions and define

vm = min(um,1 , . . . , um,m ) .

The functions vm are superharmonic and non-negative in with

lim inf vm (x) 1

3xy

for all y E and

1
vm (xk ) ,
2m+1
for k = 1, . . . , m.
Finally, define

+
v = vm .
m=1

Since v(xk ) < + for every k, by the third property of superharmonic

functions, v is superharmonic and non-negative in and

+
lim inf v(x) lim inf vm (x) = +
3xy 3xy
m=1

for every y E.

Theorem 3.5 Suppose that is a bounded open subset of Rn and let E be a

x -measurable for all x . In particular, E can be
subset of which is d
any Borel subset of . Then, the following are equivalent.
1. E is of zero harmonic measure with respect to .
2. For every u superharmonic and bounded from below in with

lim inf u(x) 0

3xy

for every y \ E, it is true that

u 0

everywhere in .
There is a dual statement for subharmonic functions.
Proof:
Let E be of zero harmonic measure with respect to .
Assume that u satisfies the hypotheses in 2 and take M > 0 so that u M
identically in . Then M
1
u E implying

1
u HE
M
140 CHAPTER 3. THE PROBLEM OF DIRICHLET

and, since the last function is identically 0 in , the proof of one direction is
complete.
Now, let dx0 (E) > 0 for at least one x0 .
Then

H E
(x0 ) = d x0 (E) < 0

and, hence, there exists u

E with u(x0 ) < 0.
This u is superharmonic and bounded from below in and
lim inf u(x) 0
3xy

for every y \ E.

Theorem 3.5 expresses an extension of the Minimum Principle for superhar-

monic functions. It appears that the sets of zero harmonic measure with respect
to are the negligible sets when testing the hypotheses of the Minimum Prin-
ciple. There is an extra mild hypothesis: the superharmonic function must
be bounded from below.

3.6 Barriers and regularity

Definition 3.5 Let be open in Rn and y0 .
We say that has a barrier at y0 , if there is an open neighborhood V of
y0 and a positive superharmonic function u in V so that
lim u(x) = 0 .
3xy0

This u is called barrier for at y0 .

y0 is called regular boundary point of , if there exists some barrier for
at y0 .
is called regular open set, if all its boundary points are regular boundary
points of .
Observe that the neighborhood of y0 in the definition may become as small
as we like and, hence, the part of outside an arbitrarily small neighborhood
of y0 does not play any role in whether y0 is regular or not. In other words the
definition of regularity of a boundary point is local in character.
The next result is a concrete characterization of regularity of boundary
points.
Lemma 3.3 Let be open in Rn , y0 and y0 6= .
Then the following are equivalent.
1. y0 is a regular boundary point of .
2. There is an R > 0 so that
B(y ;R)
lim H| y0 | 0 (x) = 0 .
3xy0
3.6. BARRIERS AND REGULARITY 141

3. The previous condition holds for all R > 0.

If n 3 and is not bounded, then is always a regular boundary point
of .
If n = 2 and is not bounded, then the following are equivalent
1. is a regular boundary point of .
2. There is an R > 0 so that
{x:|x|>R}
lim H1/|| (x) = 0 .
3x

3. The previous condition holds for all R > 0.

Proof:
(1) Let y0 6= and suppose that the condition in 2 is true for some R.
B(y ;R)
Since | y0 | is subharmonic, we have that H| y0 | 0 (x) |x y0 | > 0
B(y ;R)
for all x B(y0 ; R). Therefore, H| y0 | 0 is a barrier for at y0 .
Conversely, suppose that there is a neighborhood V of y0 and a positive
superharmonic u in V with lim3xy0 u(x) = 0.
We may assume that V = B(y0 ; r) with r < R and we consider an extra
< r.
Then S(y0 ; ) is an open subset of S(y0 ; ) and we may decompose it as
S(y0 ; ) = F A, where F is compact, A is open in S(y0 ; ), F A = and
dS(A) < n1 n1 r .
For this purpose, consider a compact exhaustion {K(m) } of and take F =
K(m) S(y0 ; ) for a large enough m.
Now, define the function
r
w(x) = u(x) + rPA (x; y0 , ) + , x B(y0 ; ) .
minF u
which is positive and superharmonic in B(y0 ; ) and consider an arbitrary
B(y ;r) ( )
v | y0 | 0 . Then, for every y B(y0 ; ) ,

lim sup v(x) lim inf w(x) .

B(y0 ;)3xy B(y0 ;)3xy

This is easy to prove, by considering the three cases: y B(y0 ; ),

y F and y A.
Therefore, v w everywhere in B(y0 ; ) and, thus,
B(y ;r)
H| y0 | 0 w

everywhere in B(y0 ; ).
This implies
B(y ;r) r
lim sup H| y0 | 0 (x) lim u(x) + rPA (y0 ; y0 , ) + 2
3xy0 minF u 3xy0
142 CHAPTER 3. THE PROBLEM OF DIRICHLET

and, since is arbitrarily small,

B(y ;r)
lim H| y0 | 0 (x) = 0 .
3xy0

We need to show the same thing, but with r replaced by R.

Since
B(y ;r)
|x y0 | H| y0 | 0 (x) r
for all x B(y0 ; r), we get
B(y ;r)
lim H| y0 | 0 (z) = r
B(y0 ;r)3zx

for all x S(y0 ; r). This implies that the function

{
B(y0 ;r)
R
r H| y0 | (x) , if x B(y0 ; r)
( )
R, if x B(y0 ; R) \ B(y0 ; r)

B(y ;R)
belongs to | y0 | 0 and, hence,

B(y ;R) R B(y0 ;r)

H| y0 | 0 H
r | y0 |
everywhere in B(y0 ; r). Therefore,
B(y ;R)
lim H| y0 | 0 (x) = 0 .
3xy0

(2) Now, if n 3 and is not bounded, the function

1
h(x) = , x,
|x|n2

is a barrier for at .
(3) In the case n = 2 we may either modify the proof of part (1) or, better,
consider the inversion x = x1 which transforms \ {0} with as boundary
point to the set = {x : x1 \ {0}} having 0 as boundary point.
This inversion, as we have already seen, preserves the properties of har-
monicity and superharmonicity. Hence, it is clear that is regular for if and
only if 0 is regular for and, also, that
{x:|x|>R}
lim H1 (x) = 0
3x ||

if and only if
B(0; R
1
)
lim H|| (x) = 0 .
3x0

Now, we may use the result of part (1) and complete the proof.
3.7. REGULARITY AND THE PROBLEM OF DIRICHLET 143

Theorem 3.6 (Bouligand) Let be an open subset of Rn and y0 . Then,

the following are equivalent.
1. y0 is a regular boundary point of .
2. For every open neighborhood V of y0 , there is a positive u superharmonic
in so that u = 1 identically in \ V and lim3xy0 u(x) = 0.
Proof:
One direction is trivial. Therefore, let y0 be regular and V be any open
neighborhood of y0 .
Assume, first, that y0 Rn and consider R > 0 small enough to have
B(y0 ; R) V . By Lemma 3.3,
B(y ;R)
lim H| y0 | 0 (x) = 0 .
3xy0

We, clearly, have

B(y ;R)
0 < |x y0 | H| y0 | 0 (x) R

for all x B(y0 ; R) and, thus,

B(y ;R)
lim H| y0 | 0 (z) = R
B(y0 ;R)3zx

for all x S(y0 ; R).

This implies that the function
{ 1 B(y0 ;R)
R H| y0 | (x) , if x B(y0 ; R)
u(x) =
1, if x \ B(y0 ; R)

has the desired properties.

If n 3 and y0 = , we consider R large enough so that {x : |x| > R} V
and, then, take
( Rn2 )
u(x) = min , 1 , x.
|x|n2
If, finally, n = 2 and y0 = , then, taking R large enough and
{
{x:|x|>R}
RH 1 (x) , if x and |x| > R
u(x) = ||
1, if x and |x| R ,

we conclude the proof.

3.7 Regularity and the problem of Dirichlet

Theorem 3.7 Suppose is a bounded open subset of Rn and y0 . Then
the following are equivalent.
144 CHAPTER 3. THE PROBLEM OF DIRICHLET

1. y0 is a regular boundary point of .

2. For every real-valued f defined and bounded in and continuous at y0

it is true that

lim H f (x) = lim H
f (x) = f (y0 ) .
3xy0 3xy0

That 1 implies 2 holds without the assumption of boundedness of .

Proof:
By the subharmonicity of the function f () = | y0 |, we find

0 < |x y0 | H
f (x) diam()

for all x . Therefore, if 2 holds, the function H

f is a barrier for at y0 and
y0 is a regular boundary point of .
Now, suppose that y0 is a regular boundary point of and consider any
real-valued f defined and bounded in and continuous at y0 .
Take any  > 0 and let V be a neighborhood of y0 so that |f (y) f (y0 )| < 
for all y V .
Theorem 3.6 implies that there is a positive superharmonic u in so that
lim3xy0 u(x) = 0 and u = 1 identically in \ V .
If |f (y)| M for all y , then the function
( )
w = M f (y0 ) u + f (y0 ) + 

f and, thus, lim sup3xy0 H f (x) f (y0 ) + . Since  is arbitrary,
belongs to

lim sup H f (x) f (y0 ) .
3xy0

Applying this to f we find

f (x) f (y0 )
lim inf H
3xy0

and, combining the two inequalities, we finish the proof.

Theorem 3.8 If is an open subset of Rn , then the regularity of implies

that the Problem of Dirichlet is solvable for every f C().
If is bounded, then the converse is, also, true.

Proof:
A direct consequence of Theorem 3.7.
3.8. CRITERIA FOR REGULARITY 145

3.8 Criteria for regularity

The next three results give three useful criteria for regularity of boundary points.
Much later we shall prove Wieners characterization of regularity of boundary
points.
Proposition 3.4 (The ball-criterion) Let Rn be open and y0 , y0
Rn . If there is a ball B Rn \ so that y0 B, then y0 is a regular boundary
point of .
If, in particular, is C 2 at y0 , then y0 is a regular boundary point of .
Therefore, if the bounded open has C 2 -boundary, then it is a regular set.
Proof:
If y1 is the center of B and r is its radius, then hy1 + h (r) is a barrier for
at y0 .

Proposition 3.5 (The continuum-criterion) Let R2 be open and y0 .

If there is a continuum containing y0 and contained in R2 \, then y0 is a regular
boundary point of .
In particular, if the complement of with respect to R2 has no component
reducing to only one point, then is a regular set.
Proof:
Assume, first, that y0 6= .
Let C be the continuum of the statement and consider y1 C R2 with y1 6=
y0 . If R = |y1 y0 |, then all connected components  open set B(y0 ; R) \ C
 of the
 xy1 
are simply-connected. The function u(x) = log xy0  , x B(y0 ; R) \ C , is
harmonic in B(y0 ; R) \ C and, by Theorem 1.4, there is a harmonic conjugate v
of it there.
Since B(y0 ; R) B(y0 ; R) \ C, it is clear that the function
( ) x y 
1  1
< , x and  >1,
u(x) + iv(x) x y0
is a barrier for at y0 .
If y0 = , we choose a y1 C with y1 6= and, then, all connected
components of R2 \ C are simply-connected. We define a harmonic conjugate
v of the harmonic function u(x) = log |x y1 | in R2 \ C.
Since R2 \ C, the function
( 1 )
< , x and |x y1 | > 1 ,
u(x) + iv(x)
is a barrier for at .

Proposition 3.6 (The cone-criterion) Let Rn be open and y0 ,

y0 Rn . If there is an open truncated cone F Rn \ with vertex y0 , then y0
is a regular boundary point of .
146 CHAPTER 3. THE PROBLEM OF DIRICHLET

If, in particular, is C 1 at y0 , then y0 is a regular boundary point of .

Hence, if the bounded open has C 1 -boundary, then it is a regular set.
Proof:
Let R be the height of the cone F . It is enough to find a barrier for the open
set B(y0 ; R) \ F at y0 and, by Lemma 3.3, it is enough to prove

B(y ;R)\F
lim H| y00 | (x) = 0 .
B(y0 ;R)\F 3xy0

Now, set
B(y ;R)\F
u(x) = H| y00 | (x)

for all x B(y0 ; R) \ F and, then, dilate F by a factor of two, producing the
cone
F 0 = y0 + 2(F y0 ) ,
and consider
( 1 )
v(x) = u y0 + (x y0 )
2
for all x B(y0 ; 2R) \ F 0 .
The function v is harmonic in B(y0 ; 2R) \ F 0 .
By the ball-criterion, every boundary point of B(y0 ; R) \ F is regular except,
perhaps, y0 and, therefore,

lim u(x) = |y y0 |
B(y0 ;R)\F 3xy

( )
for all y B(y0 ; R) \ F except, perhaps, y0 .
By the Maximum-Minimum Principle, we have that u < R in B(y0 ; R) \ F
and, thus,
sup v < R .
S(y0 ;R)\F

1
Hence, we can choose so that 2 < < 1 and
( )
lim u(x) v(x) 0
B(y0 ;R)\F 3xy

for all y S(y0 ; R) \ F . We, also, have

(1 )
lim u(x) v(x) = 0
B(y0 ;R)\F 3xy 2

for all y F \ {y0 }.

Therefore, ( )
lim u(x) v(x) 0
B(y0 ;R)\F 3xy
( )
for all y B(y0 ; R) \ F except, perhaps, y0 .
3.8. CRITERIA FOR REGULARITY 147

Since the function u v is bounded from below, we have that, for all  > 0,
( )
lim u(x) v(x) + (hy0 (x) h (R)) 0
B(y0 ;R)\F 3xy
( )
for all y B(y0 ; R) \ F .
By the Maximum-Minimum Principle

u(x) v(x) + (hy0 (x) h (R)) 0

for all x B(y0 ; R) \ F ,and, since  is arbitrary, we find

u v

in B(y0 ; R) \ F .
From this,
1 ( 1 )
lim sup u(x) lim sup u y0 + (x y0 )
B(y0 ;R)\F 3xy0
B(y0 ;R)\F 3xy0 2
1
= lim sup u(x) ,
B(y0 ;R)\F 3xy0

implying
lim u(x) = 0 .
B(y0 ;R)\F 3xy0

If a boundary point satisfies the ball-criterion, then it satisfies the cone-

criterion and, in case n = 2, if it satisfies the cone-criterion then it satisfies
the continuum-criterion. Therefore, the cone-criterion is the most useful in case
n 3 and the continuum-criterion is the most useful in case n = 2.
Proposition 3.7 If Rn is open and y0 is an isolated point of , then y0
is not a regular boundary point of . The only exception is when n 3 and
y0 = .
Proof:
Let y0 6= be a regular boundary point of and consider a small enough R
so that B(y0 ; 2R) \ {y0 } and a positive superharmonic u in B(y0 ; 2R) \ {y0 }
with limxy0 u(x) = 0.
Let m = minS(y0 ;R) u , being clear that m > 0.
By the Minimum Principle, for every < R,
h () h(x)
u(x) m
h () h (R)

for all x in B(y0 ; , R).

Now, let 0+ and find
u(x) m
148 CHAPTER 3. THE PROBLEM OF DIRICHLET

for all x B(y0 ; R) \ {y0 }, getting a contradiction.

In case n = 2 and y0 = we modify the previous proof, taking u positive
and superharmonic in {x : |x| > 12 R} with limx u(x) = 0, defining
m = min|x|=R u(x) > 0 and observing that

h(x) h (r)
u(x) m
h (R) h (r)

for all x with R < |x| < r.

We get a contradiction, letting r + and finding

u(x) m

for all x with R < |x|.

Chapter 4

The Kelvin Transform

4.1 Definition
Consider any ball B(x0 ; R) and the symmetric x of any x with respect to
S(x0 ; R),
R2
x = x0 + (x x0 ) .
|x x0 |2
As usual, we consider each of x0 and to be symmetric to the other.
Now, for every set A Rn , we define its symmetric with respect to S(x0 ; R)
by
A = {x : x A} .
The new set A contains x0 or if and only if A contains or x0 , respec-
tively.
A nice geometric property is that spheres are transformed, by symmetry,
onto spheres. In fact, doing some easy calculations, we can prove that if
the sphere S(x1 ; R1 ) does not contain x0 , then its symmetric, S(x1 ; R1 ) , is
2
the sphere having the point |x1 x2R 2 (x1 x0 ) as center and the number
0 | R
2
1
R 2 R1
as radius.
||x1 x0 |2 R12 |
If the interior B(x1 ; R1 ) contains x0 , then it is transformed onto the exterior
of the image sphere, while the exterior is transformed onto the interior.
If the interior does not contain x0 , then it is transformed onto the interior
of the image sphere, while the exterior is transformed onto the exterior.
If the sphere S(x1 ; R1 ) contains x0 , then it is transformed onto the hyper-
plane described by the equation (x x0 ) (x1 x0 ) = 12 R2 and the interior
and exterior of the sphere are transformed onto the two half-spaces determined
by this hyperplane.
Definition 4.1 Let be an open subset of Rn not containing x0 .
For every function f defined in , we call the function
|x x0 |n2 1
f (x ) = f (x) = f (x) ,
R2n4 |x x0 |n2

149
150 CHAPTER 4. THE KELVIN TRANSFORM

defined in , the Kelvin Transform of f with respect to S(x0 ; R).

Whenever we write about the Kelvin Transform without specifying the sphere,
we shall understand that the sphere is S(0; 1).

Proposition 4.1 Let be an open subset of Rn not containing x0 .

Then, u is harmonic or superharmonic or suharmonic in if and only if its
Kelvin Transform u with respect to S(x0 ; R) is harmonic or superharmonic or
subharmonic, respectively, in .

Proof:
If u is in C 2 (), then, by trivial calculations, we can prove

|x x0 |n+2
u (x ) = u(x)
R2n
for all x .
We conclude that, if u is harmonic or twice continuously differentiable and
superharmonic in , then u is harmonic or superharmonic, respectively, in .
For a general superharmonic u in and an arbitrary open ball B with B

, we consider the symmetric closed ball B and, through Theorem 2.10,
a sequence {um } of twice continuously differentiable superharmonic functions

in an open set 0 with B 0 which increase towards u in 0 .
Then, the functions um are superharmonic in 0 and, hence, in B and
increase towards u there.
By the third property of superharmonic functions, u is superharmonic in B
and, since the ball is arbitrary, superharmonic in .

4.2 Harmonic functions at

If V is an open neighborhood of x Rn , then the set V \ {x} is called a
punctured neighborhood of x.

Proposition 4.2 Let u be harmonic in a punctured neighborhood V \ {} of

and u its Kelvin Transform with respect to S(x0 ; R). We know that u is
harmonic in the punctured neighborhood V \ {x0 } of x0 .
Then, the following are equivalent.

1. u can be defined at x0 so that it becomes harmonic in V .

2.
lim u(x) = 0 in case n 3
x

and
u(x)
lim = 0 in case n = 2 .
x log |x|
4.2. HARMONIC FUNCTIONS AT 151

Proof:
In case n 3,
u (x )
lim = lim u(x) ,
x x 0 hx0 (x ) x

while, in case n = 2,
u (x ) u(x) u(x)
lim = lim = lim .
x x 0 hx0 (x ) x log |xx0 | x log |x|
R2

Theorem 1.12 concludes the proof.

Proposition 4.3 Suppose that n = 2 and f is holomorphic in B(x0 ; R) \ {x0 }.

Then, f can be extended as a holomorphic function in B(x0 ; R) if and only if
limxx0 (x x0 )f (x) = 0.
Proof:
The necessity of the condition is obvious.
Therefore, assume that limxx0 (x x0 )f (x) = 0, take r so that 0 < r < R
and consider the function

1 f (y)
g(x) = dy , x B(x0 ; r) ,
2i B(x0 ;r) y x

which is holomorphic in B(x0 ; r).

Now, fix an x B(x0 ; r) \ {x0 }, take  so that 0 <  < |x x0 | and apply
Cauchys Formula to f in B(x0 ; r) \ B(x0 ; ) to get

1 f (y) 1 f (y)
f (x) = dy dy
2i B(x0 ;r) y x 2i B(x0 ;) y x

1 f (y)
= g(x) dy .
2i B(x0 ;) y x

The last integral tends to 0 as  0+ and, thus, f (x) = g(x) for all
x B(x0 ; r) \ {x0 }.

Proposition 4.4 Suppose that n = 2, f is holomorphic in a punctured neigh-

borhood V \ {} of and f is its Kelvin Transform with respect to S(x0 ; R).
Then, f is holomorphic in the punctured neighborhood V \ {x0 } of x0 and it
can be extended as a holomorphic function in V if and only if limx f (x)
x = 0.

Proof:
The proof is a trivial application of Proposition 4.3.

Observe that condition 2 of Proposition 4.2 and the analogous condition of

Proposition 4.4 are independent of the sphere with respect to which we take
the Kelvin Transform. Therefore, in the following definition the use of S(0; 1)
is only for reasons of simplicity and the use of any other sphere is equivalent.
152 CHAPTER 4. THE KELVIN TRANSFORM

Definition 4.2 Suppose that the open set Rn contains and u is defined
in \ {}.
We say that u is harmonic in if it is harmonic in \ {} and there
is a punctured neighborhood V \ {} of so that the Kelvin Transform u ,
harmonic in V \ {0}, can be defined at 0 so that it is harmonic in V .
We call u harmonic at , if it is harmonic in some neighborhood of .
If n = 2, we, similarly, define holomorphic functions at .
Proposition 4.2 gives a necessary and sufficient condition on u, harmonic in
a punctured neighborhood of , so that it is harmonic at and Proposition
4.4 gives a necessary and sufficient condition on f , holomorphic in a punctured
neighborhood of , so that it is holomorphic at .
Theorem 4.1 Let u be harmonic in a punctured neighborhood V \ {} of .
If n 3, u can be extended as harmonic in V if and only if limx u(x) = 0.
u(x)
If n = 2, u can be extended as harmonic in V if and only if limx log |x| = 0
if and only if limx u(x) exists in C.
Suppose that n = 2 and f is holomorphic in a punctured neighborhood
V \ {} of . Then f can be extended as holomorphic in V if and only if
limx f (x)
x = 0 if and only if limx f (x) exists in C.

In case n 3, if u is harmonic at , then, by defining u() = 0, we

guarrantee that u is continuous at . Therefore, we may say that, if u is
defined and continuous in an open set containing , then it is harmonic in
if and only if it is harmonic in \ {} and u() = 0.
Writing u (x ) = u (0) + O(|x |) when x is near 0, we find that
u (0) ( 1 )
u(x) = + O
|x|n2 |x|n1
when x is near .
In case n = 2, we get in the same way that
( 1 )
u(x) = u (0) + O
|x|
when x is near , implying that, by defining u() = u (0), u becomes contin-
uous at . Therefore, we may say that, if u is defined and continuous in an
open set containing , then it is harmonic in if and only if it is harmonic
in \ {}.
In this case there is no universal value at for harmonic functions there,
as is the value 0 in case n 3.
We summarize.
Proposition 4.5 Suppose u is defined and continuous in an open set con-
taining .
If n 3, then u is harmonic in if and only if u is harmonic in \ {}
and u() = 0 if and only if u is harmonic
( in
) \{} and there is some complex
number a so that u(x) = |x|n2 + O |x|n1 when x is near .
a 1
4.2. HARMONIC FUNCTIONS AT 153

( )in \ {}
If n = 2, then u is harmonic in if and only if u is harmonic
if and only if u is harmonic in \ {} and u(x) = u() + O |x| 1
when x is
near .
If f is holomorphic in an open set containing and all x R2 with
|x| R, then a trivial use of the Kelvin Transform with respect to S(0; 1)
shows that f has a power series expansion

+
1
f (x) = an , |x| > R .
n=0
xn

Of course, a0 = f () and

f 0 () = a1 = lim x(f (x) f ())

x+

is called the (complex) derivative of f at . Observe that this coincides

with the derivative of f at 0.
Either using the Kelvin Transform and the analogous formulas for f or
integrating the power series of f , we may, easily, prove that

1
an = f (y)y n1 dy .
2i B(0;r)

The following is a direct application of the definitions and Theorem 1.5.

Theorem 4.2 Let u be harmonic in the neighborhood {x : R < |x|} {} of
.
In case n 3 we have

1. = S(0;r) u
(y) dS(y) is constant in the interval R < r < +.

2. Mru (0) = (n2)

n1 r
1
n2 in the same interval.

In case n = 2,

1. 0 = S(0;r) u
(y) dS(y) identically in the interval R < r < +.

2. Mru (0) = u() identically in the same interval.

Here,
is the continuous unit vector field which is normal to S(0; r) and
in the direction towards the exterior of B(0; r).
The next result regards the representation of a harmonic function as the
difference between a single- and a double-layer potential.
The function h is defined by
{
log |x| , if n = 2
h (x) = 1
, if n 3 ,
|x|n2

and it, also, is a fundamental solution of the Laplace equation in Rn \ {0} .

154 CHAPTER 4. THE KELVIN TRANSFORM

Theorem 4.3 Let be an open set containing and having C 1 -boundary and
let

be the continuous unit vector field normal to in the direction towards
the exterior of . Then,
1. For every x with x 6= ,
(
1 h (x ) u )
u(x) u() = u(y) (y) h (x y) (y) dS(y) .
n

2. If x1 is any point outside , then

(
1 h (x1 ) u )
u() = u(y) (y) h (x1 y) (y) dS(y) .
n

Proof:
The proof is a routine application of Greens formula in the open set
B(0; R) or B(x1 ; R), where R is large and eventually tends to +. It uses
the formulas in Theorem 4.1 and it is left to the interested reader.

Theorem 4.4 If u is harmonic in Rn , then, in case n 3, it is identically 0

and, in case n = 2, it is a constant function.
Similarly, if f is holomorphic in R2 , then f is a constant function.
Proof:
u is bounded from below in Rn and, from the Theorem of Picard, it is
constant.

4.3 Superharmonic functions at

The following is parallel to the definition of harmonicity at .
Definition 4.3 Suppose that the open set Rn contains and u is defined
in \ {}.
We say that u is superharmonic in if it is superharmonic in \ {} and
there is a punctured neighborhood V \ {} of so that the Kelvin Transform
u , superharmonic in V \ {0}, can be defined at 0 so that it is superharmonic
in V .
We say that u is superharmonic at , if it is superharmonic in some neigh-
borhood of .
The definition of subharmonicity is analogous.
If u is superharmonic at , then it is natural to admit

u() = lim inf u(x)

x

as its value at . This choise makes u lower-semicontinuous in the set where it

is superharmonic.
4.4. POISSON INTEGRALS AT 155

Theorem 4.5 Let u be superharmonic in a punctured neighborhood V \ {} of

. Then u can be extended at so that it becomes superharmonic in V if and
only if

1. lim inf x u(x) 0, in case n 3, and

u(x)
2. lim inf x log |x| 0, in case n = 2.

Proof:
The proof is a direct application of the definition and Theorem 2.5.

Theorem 4.6 u is superharmonic in R2 if and only if u is constant.

If n 3, then u is superharmonic in Rn if and only if its restriction to Rn
is a non-negative superharmonic function and u() = lim inf x u(x).

Proof:
The Kelvin Transform u (x ) = u(x) is superharmonic in R2 \{0} and it can
be extended at 0 so that it is superharmonic in R2 . Its value at 0 is, necessarily,
u (0) = lim inf x 0 u (x ) = lim inf x u(x).
If we define u() = u (0), then u is lower-semicontinuous in the compact
set R2 . Therefore, it takes a minimum value in R2 .
If this minimum value is taken at a point in R2 , then u is constant in R2 .
Otherwise, u , which is superharmonic in R2 , takes its minimum value at 0
and it is constant. Thus, u is also constant in R2 \ {0} and, hence, in R2 .
To deal with the case n 3, we just apply the Minimum Principle for u in
Rn and Theorem 4.4.

Corollary 4.1 A superharmonic function in R2 which is bounded from below

is constant.

4.4 Poisson integrals at

Definition 4.4 Let f be integrable on S(0; R) with respect to the surface mea-
sure. We define the Poisson integral of f in the exterior of B(0; R) by

1 |x|2 R2
Pf (x; , R) = f (y) dS(y) , |x| > R .
n1 R S(0;R) |x y|n

In order to have continuity at , the values that are assigned to the Pois-
son integral at the point x = are defined (with the help of the Dominated
Convergence Theorem) by
{
0, if n 3
Pf (; , R) = lim Pf (x; , R) = 1
f (y) dS(y) , if n=2.
x 1 R S(0;R)
156 CHAPTER 4. THE KELVIN TRANSFORM

It is obvious that, if f is integrable on S(0; R), then f is integrable on

S(0; R1 ) and trivial calculations result to the formula
( 1)
Pf (x; , R) = |x |n2 Pf x ; 0,
R
for all x with |x| > R.
This says that the Kelvin Transform of the Poisson integral of f in the
exterior of B(0; R) is equal to Poisson integral of the Kelvin Transform of f in
the symmetric ball B(0; R1 ).
Now, the following properties are straightforward, and can be proved either
directly, using the usual properties of the Poisson kernel, or using this last
formula.
1. Pf ( ; , R) is harmonic in Rn \ B(0; R).
This is obvious, since this function is the Kelvin Transform of a function
harmonic in B(0; R1 ).
2. If f is continuous at some y0 S(0; R), then

lim Pf (x; , R) = f (y0 ) .

xy0 ,|x|>R

Just observe that f is continuous at y0 and write

1
lim Pf (x; , R) = lim |x |n2 Pf (x ; 0, )
xy0 ,|x|>R 1
B(0; R )3x y0 R
= |y0 |n2 f (y0 ) = f (y0 ) .

3. Thus, if f is continuous on S(0; R), then Pf ( ; , R) is the unique solution

to the Problem of Dirichlet in Rn \ B(0; R) with f as boundary function.
The uniqueness is proved easily by taking Kelvin transforms and reducing
to the uniqueness of the Problem of Dirichlet in B(0; R1 ).

4.5 The effect of the dimension

Many of the properties of harmonic or superharmonic or subharmonic functions
that we have studied continue to hold when the domain of definition contains
as interior point. If this domain is the whole space Rn , then most of these
properties are trivial due to the Theorems 4.3, 4.5 and Corollary 4.1. If the
domain of definition, , misses some point x0 Rn , we, then, take the translates
u( + x0 ) which are defined in the set x0 not containing 0 and apply the
Kelvin Transform. This reduces our study to the case of a domain of definition,
( x0 ) , contained in Rn .
We shall describe, now, a difference between the cases n = 2 and n 3
which has already, to a certain degree, appeared in our results.
Take, for example, the Maximum-Minimum Principle for harmonic func-
tions.
4.6. DIMENSION 2, IN PARTICULAR 157

The function h is, in case n 3, harmonic in the open set Rn \ B(0; 1) with
boundary values 1 in S(0; 1). One would expect that the function is identically
1 in Rn \ B(0; 1), but it is not. In fact, its value at is 0, as is the value at
of every harmonic function there.
This is, best, explained using the Kelvin Transform h (x ) = |x|n2 h(x) = 1
which is harmonic in the symmetric set B(0; 1), has boundary values 1 on S(0; 1)
and is, indeed, identically 1 in B(0; 1).
Thus, in case n 3, the correct statement of the Maximum-Minimum
Principle is:
Let be open in Rn containing and let x0 / . If u is superharmonic in
and
lim inf |x x0 |n2 u(x) m
3xy

for every y , then

|x x0 |n2 u(x) m
for every x .
There are similar statements for subharmonic and harmonic functions.
The situation is simpler when n = 2. In this case, the formula of the Kelvin
Transform, u (x ) = u(x), does not contain the factor |x|n2 and all results
which hold for open subsets of R2 transfer, without any change, for open sets
in R2 containing .

4.6 Dimension 2, in particular

We state, below, the most important of the properties that hold in case n = 2
and remark that some of them hold in case n 3, also, while some others hold
after an appropriate modification, as explained a few lines above. It is left to
the interested reader to investigate the case n 3.

In all results below the open sets are subsets of R2 .

1. All versions of the Maximum/Minimum Principles are valid.

2. Locally uniform limits of harmonic functions are harmonic.
3. If 1 and 2 are open, f is meromorphic in 1 , f (1 ) 2 and u is
harmonic or superharmonic or subharmonic in 2 , then u f is harmonic or
superharmonic or subharmonic, respectively, in 1 (except if f is constant c in
some component of and u(c) = ).
4. If u is harmonic (superharmonic) in an open set containing all x with |x| R
together with , then the Poisson Formula

u(x) = () Pu (x; , R)

holds for all x with |x| > R.

5. If u is superharmonic in an open set , B denotes any open disc with B
(its center may well be ) and uB denotes the function which equals u in \ B
and equals the Poisson integral of u in B, then
158 CHAPTER 4. THE KELVIN TRANSFORM

1. u uB in ,
2. uB is superharmonic in and
3. uB is harmonic in B.
6. Harnacks inequalities hold in general: if u is positive and harmonic in
the open and K is a compact subset of , then
1 u(x)
C
C u(x0 )
for all x, x0 K, where C is a positive constant depending only on and K.
7. If {um } is an increasing sequence of harmonic functions in the connected
open set , then, either the um converge uniformly on compact subsets of
to some harmonic function in or they diverge to + uniformly on compact
subsets of .
8. The minimum of finitely many superharmonic functions is superharmonic.
9. Limits of increasing sequences of superharmonic functions in a connected
open set are either identically + or superharmonic.
10. The Perron Process: suppose that V is a non-empty family of subharmonic
functions in the connected open set so that V contains the maximum of every
two of its elements and that it contains vB (see 5 above) for all v V and all
closed discs B .
Then, the upper envelope of V is either identically + or harmonic in .
From this we get the corollary
11. Let U be a non-empty family of superharmonic functions in the open
having at least one subharmonic minorant. Then the upper envelope of all sub-
harmonic minorants of U is harmonic in and it is called the largest harmonic
minorant of U.

12. For every extended-real-valued f defined in the functions H f and H f
are defined in , each of them is, in every connected component of , either

identically + or identically or harmonic and they satisfy H f Hf .
If this inequality is equality in and the common function is harmonic in
, we, then, call f resolutive, denote this common function by Hf and call it
the generalized solution of the Problem of Dirichlet in with boundary function
f.
13. We have Wieners Theorem: if there is some disc disjoint from the open ,
then every function continuous in is resolutive.
We just translate so that the disc has center at 0 and, then, apply the Kelvin
Transform. Since the resulting open set is bounded, we may apply the original
version of Wieners Theorem.
The resulting functional

C() 3 f 7 Hf (x0 ) C

is linear, non-negative and bounded with norm 1.

14. For every open which is disjoint from some disc and every x0 (even
4.6. DIMENSION 2, IN PARTICULAR 159

) the harmonic measure d x0 is defined in . This is a complete probability

measure whose -algebra of measurable sets contains B().
Every extended-real-valued f in is resolutive if and only if it is d x-
integrable for all x and, in this case,


Hf (x) = f (y) d
x (y)

for all x .
If E , then E is of zero harmonic measure with respect to if and
only if there is a non-negative superharmonic function in having limit + at
every point of E.
Borel subsets of of zero harmonic measure with respect to are negligible
regarding the assumptions of all versions of the Maximum/Minimum Principle.
15. Regularity of boundary points is defined as originally and we have the
basic result that for any open , the Problem of Dirichlet is solvable for every
continuous boundary function if the set is regular.
The converse is, also, true, if is disjoint from some disc.
16. A useful sufficient condition for the regularity of a boundary point y0 of an
open set is that there is a continuum containing y0 and contained in R2 \ .
If, in particular, no component of R2 \ reduces to only one point, then
is a regular set.
160 CHAPTER 4. THE KELVIN TRANSFORM
Chapter 5

Greens Function

5.1 Definition
Definition 5.1 Suppose that is an open subset of Rn and let x0 . Con-
sider the family Ux0 of all functions u with the properties
1. u is superharmonic in ,
2. u is a majorant of hx0 in .
In case this family is non-empty we say that has a Greens function
with respect to the point x0 and, if Ux0 is its lower envelope, the function

G
x0 = hx0 + Ux0

is called the Greens function of with respect to the point x0 .

In case this family is empty, we say that has no Greens function with
respect to x0 .

Observe that hx0 is a harmonic majorant of itself in every connected com-

ponent of not containing the point x0 . Thus, the existence of G
x0 is guarran-
teed in all these components and it is identically 0 there.
For the same reason, if O is the connected component of which contains
the point x0 , then the existence of Gx0 in is equivalent to the existence of
GOx0 in O and, in this case,

G O
x0 (x) = Gx0 (x) , xO.

This remark helps us to reduce the study of the Greens function to the case
of connected open sets.
Proposition 5.1 If Rn has a Greens function, G
x0 , with respect to its
point x0 , then

x0 hx0 is harmonic in ,
1. G

161
162 CHAPTER 5. GREENS FUNCTION

x0 is superharmonic in and harmonic in \ {x0 },

2. G
3. G
x0 (x) > 0 for every x in the connected component of containing x0
and Gx0 (x) = 0 for all other x .

Proof:
1. By its definition, Gx0 hx0 is the least harmonic majorant of the subharmonic
function hx0 in .
2. This is obvious.
3. If O is any connected component of not containing x0 , then hx0 is, clearly,
the least harmonic majorant of itself in O. Therefore, G x0 = 0 identically in O.
If O is the connected component containing x0 , then G x0 0 everywhere in
O. In case G x0 (x) = 0 for at least one x O, then, by the Minimum Principle,
Gx0 = 0 identically in O, implying that hx0 is harmonic in O.

5.2 Greens function, the problem of Dirichlet

and harmonic measure
Proposition 5.2 If is any bounded open subset of Rn , then has a Greens
function with respect to every x0 and

Gx0 (x) = hx0 (x) + Hhx0 (x) = hx0 (x)

hx0 (y) d
x (y)

for all x .

Proof:
1. Since is bounded, there is a large enough constant playing the role of a
superharmonic majorant of hx0 in .
Therefore, by definition, has a Greens function with respect to x0 .
Since hx0 is continuous in , Wieners Theorem implies that this function
is resolutive with respect to . One can see this, directly, as follows.
hx0 is subharmonic in , bounded from above in and, hence, belongs to
hx . Therefore,
0

hx0 (x) H
hx0 (x)

for all x . This, easily, implies that H

hx0 is harmonic and, by the
Maximum-Minimum Principle, bounded from below in . Therefore H
hx0
belongs to
hx and, thus,
0

H hx0 H
hx0

everywhere in , implying that hx0 is resolutive with respect to .

By the continuity of hx0 in and the definition of harmonic measure,

Hh
x0
(x) = hx0 (y) d
x (y)

5.3. A FEW EXAMPLES 163

for all x .
2. Assume that u belongs to the family
hx , implying
0

( )
lim inf u(x) + hx0 (x) 0
3xy

for all y .
From the Minimum Principle, we have that u hx0 everywhere in and,
hence, u belongs to Ux0 .
If, conversely, u belongs to Ux0 , then it is automatically true that u is
bounded from below in and that

lim inf u(x) hx0 (y)

3xy

for all y . Therefore u hx0 .

Hence, the families Ux0 and

hx are identical and, thus,
0

Ux0 = Hh

x0

everywhere in .

5.3 A few examples

Proposition 5.3 1. R2 has no Greens function with respect to any point
of it.
n
2. If n 3, then for every x0 Rn , GR n
x0 = hx0 in R .

Proof:
1. Assume that there is a superharmonic majorant u of hx0 in R2 .
For the arbitrary ball B(x0 ; R), we, then, have u(x) log R for all x in
S(x0 ; R) and, by the Minimum Principle,

u(x) log R

in B(x0 ; R).
Since R is arbitrary, we get a contradiction.
2. Assume that n 3 and let u be any superharmonic majorant of hx0 in Rn .
For every ball B(x0 ; R), we, then, have u(x) Rn2
1
for all x S(x0 ; R)
and, by the Minimum Principle, u(x) Rn2 for all x B(x0 ; R). Therefore,
1

u(x) 0

for all x Rn , implying that 0 is the smallest superharmonic majorant of hx0

in Rn .
We conclude that n
GRx0 (x) = hx0 (x)
164 CHAPTER 5. GREENS FUNCTION

for all x Rn .

Example
Take = B(x1 ; R) and any x0 B(x1 ; R), x0 6= x1 and consider the
R2
symmetric x0 = x1 + |x0 x 1|
2 (x0 x1 ) of x0 with respect to S(x1 ; R).

|x1 x
0|
n2
If n 3, the function Rn2 hx0 is harmonic in Rn \ {x0 } and coincides
with hx0 in S(x1 ; R).
Therefore,
|x1 x0 |n2
GB(x
x0
1 ;R)
= hx0 hx0
Rn2
in B(x1 ; R).
If n = 2, then, similarly,
( |x x | )
1
GB(x
x0
1 ;R)
= hx0 hx0 log 0
R
in B(x1 ; R).
In case x0 = x1 , then

GB(x
x1
1 ;R)
= hx1 h (R)

in B(x1 ; R), which can be recognized as the limit of both previous cases as
x0 x1 .

5.4 Monotonicity
Theorem 5.1 If the open set Rn has a Greens function with respect to
some x0 and if 0 is another open set with

x 0 0 ,

then 0 has, also, a Greens function with respect to x0 and

0
x0 Gx0
G

everywhere in 0 .

Proof:
0
It is clear that every element of Ux0 belongs to Ux0 .

Corollary 5.1 If n 3, then every open subset of Rn has a Greens function

in all of its components.

Theorem 5.2 Let {m } be an increasing sequence of open subsets of Rn with

= +
m=1 m and x0 1 .
5.5. SYMMETRY 165

1. If all m have a Greens function G

x0 with respect to x0 , then either
m

G
x0
m
+
everywhere in the component of containing x0 and has no Greens
function with respect to x0 or, in the opposite case, has a Greens func-
tion with respect to x0 and
Gm
x0 G
x0

in .
2. If has a Greens function with respect to x0 , then all m have a Greens
function with respect to the same point and
Gm
x0 G
x0

in .
Proof:
It is obvious, from Proposition 5.1(3) and Theorem 5.1, that in every compo-
nent of which does not contain x0 the Greens functions of all sets considered
are identically 0. Therefore, the proof reduces to the case of a connected .
Assume that all m have a Greens function with respect to x0 .
Take an arbitrary B(x; r) . Then, for a large enough m0 , B(x; r) m0
and Proposition 5.1(1) and Theorem 5.1 imply that {G x0 hx0 }m=m0 is an
m +

increasing sequence of harmonic functions in B(x; r).

Therefore, by Theorem 1.16, every point of has some neighborhood where
the sequence {G x0 hx0 } , eventually, either converges to a harmonic function
m

or diverges to +. Since is connected, this sequence either converges to a

harmonic function everywhere in or diverges to + everywhere in .
In the first case the harmonic limit-function majorizes hx0 in and, hence,
belongs to Ux0 . Therefore has a Greens function with respect to x0 and the
above limit majorizes G x0 hx0 in .
From Theorem 5.1, the same limit-function is majorized by G x0 hx0 in
and we, finally, get
G
x0
m
Gx0
in .
Conversely, if has a Greens function with respect to x0 , then all m have
a Greens function with respect to x0 and Gx0 Gx0 in m for all m.
m
m
Therefore, the limit of Gx0 cannot be identically + in .

5.5 Symmetry
Theorem 5.3 Let x0 and x1 belong to the same component of the open Rn .
If has a Greens function with respect to x0 , then it has a Greens function
with respect to x1 and
G
x0 (x1 ) = Gx1 (x0 ) .
166 CHAPTER 5. GREENS FUNCTION

Proof:
1. Assume that is bounded. From Proposition 5.2, we have that

Gx0 (x) = hx0 (x)

hx0 (y) d
x (y)

for all x .
Now, observe, by interchanging differentiations and integration, that the
x0 (x1 ) hx0 (x1 ) is,
integral is harmonic as a function of x0 in and, hence, G
as a function of x0 , a harmonic majorant of hx1 in .
Therefore,
Gx1 (x0 ) Gx0 (x1 ) .

The reverse inequality is proved symmetrically.

2. If is not bounded, consider the sets m = B(0; m).
Assuming that has a Greens function with respect to x0 , we get that
Gx0 (x1 ) < +.
For large enough m, x0 and x1 are both included in the same component of
m and we apply part 1 for m :

x1 (x0 ) = Gx0 (x1 ) Gx0 (x1 ) < +

G m m

and, thus, the limit of G

x1 is not identically + in the component of con-
m

taining x1 . Theorem 5.2 implies that has a Greens function with respect to
x1 and

G
x1 (x0 ) = lim G
x1 (x0 ) =
m
lim G
x0 (x1 ) = Gx0 (x1 ) .
m
m+ m+

Observe that both sides of the equality of Theorem 5.3 are equal to 0, if x0
and x1 belong to different components of .

Definition 5.2 We say that the open set Rn has a Greens function in
any one of its connected components, if it has a Greens function with respect to
at least one of the points of that component.

5.6 Greens function and regularity

Theorem 5.4 Let Rn be an open set having a Greens function in every
one of its components and y0 . If

lim G
z (x) = 0
3xy0

for at least one z in every component of , then y0 is a regular boundary point.

If is bounded and y0 is a regular boundary point, then the above limit holds
for every z .
5.7. EXTENSIONS OF GREENS FUNCTION 167

Proof:
Assuming that y0 is a regular boundary point of the bounded , we get, by
Theorem 3.7, that

lim x0 (x) = hx0 (y0 )

G lim Hhx0 (x) = 0 .
3xy0 3xy0

Now, let lim3xy0 G z (x) = 0 for at least one z in every component of .

If has finitely many components Oj , 1 j M , and zj Oj is such that
lim3xy0 G zj (x) = 0, then we form the function u which coincides in each Oj
with G zj .
This u is, obviously, a barrier for at y0 .
If has infinitely many components Oj , j N, then we form the ( positive)
1
superharmonic function u in which coincides in each Oj with min G zj , j .
For arbitrary  > 0 we take j0 1
 and we have

j 1
lim u(x) = 0
j=1
0 Oj 3xy0

and
lim sup u(x)  .
+
j=j
Oj 3xy0
0

Hence,
lim sup u(x) 
3xy0

and, since  is arbitrary, u is a barrier for at y0 .

5.7 Extensions of Greens Function

In this section we shall describe two possible extensions of a Greens function
in the complement of its domain of definition . The second extension is for
general bounded open sets , while the first is for regular bounded open and
it is intuitively simpler.

Proposition 5.4 Suppose that is a bounded regular open set and x0 .

x0 , extended as identically 0 in R \ , has the following
Then, the function G n

properties.

1. It is positive in the component O of containing x0 and it is identically

0 in Rn \ O,

2. it is continuous and subharmonic in Rn \ {x0 } and

3. it is harmonic in O \ {x0 } and its difference with hx0 is harmonic in O.

168 CHAPTER 5. GREENS FUNCTION

Proof:
All statements are already known, except for the second. The continuity
is a corollary of Theorem 5.4 and the subharmonicity is a consequence of the
continuity and of the simple fact that, for every y O, the value of the function
is 0 while the area-means over every B(y; r) are, clearly, positive.

Assume, now, that is a bounded open set and x0 . By Proposition 5.2,

x0 (x) = hx0 (x)
G hx0 (y) d
x (y)

for all x .
Take x in the same connected component O of with x0 and, since G
x (x0 ) =

Gx0 (x), we have

x0 (x) = hx (x0 )
G hx (y) d
x0 (y)

for all x in the component O of containing x0 .

If x does not belong to O, the left side of the last formula is 0. Since,
by Proposition 3.3, dx0 is supported in O and the function hx is harmonic in
O and continuous in O, the right side of the last formula is, also, 0.
Observe that this right side is, for the same reason, 0 for every x
/ O.
Therefore,

G
x0 (x) = h x0 (x) hx (y) d
x0 (y)

for all x .
By Theorem 2.8, the above integral is, as a function of x, superharmonic in
Rn and harmonic in Rn \ O.
Hence, we have proved the

Proposition 5.5 Suppose that is a bounded open set and x0 . Then the
function

hx0 (x) hx (y) d
x0 (y) , x Rn ,

is an extension of G
x0 with the following properties.

1. It is positive in the component O of containing x0 and identically 0 in

Rn \ O,

2. it is subharmonic in Rn \ {x0 } and

3. it is harmonic in O \ {x0 } and its difference with hx0 is harmonic in O.

5.8. GREENS POTENTIALS 169

5.8 Greens potentials

Now, let be a bounded open set and d be a non-negative Borel measure with
compact support in .
Consider the function

Ud (x) = G
x (y) d(y) , x.

From Proposition 5.2,

Ud (x) = hx (y) d(y) hy (z) d
x (z) d(y)
supp(d) supp(d)

= Uhd (x) hz (y) d(y) dx (z) ,
supp(d)

where the interchange of integrations is trivial to justify, since supp(d) and

are a positive distance apart. For the same reason, the inner integral defines a
continuous function of z in and, thus, the last term is a harmonic function
of x in .
We conclude, by Theorem 2.8 and Theorem 2.15, that
1. Ud is superharmonic in and harmonic in \ supp(d),

2. Ud = n d as distributions in and

3. Ud 0 everywhere in .
Now, only assume that the open set has a Greens function in every one
of its components.
Consider any open exhaustion {(m) } of and the restrictions d(m) of
the non-negative Borel measure d in .
d
By Theorem 5.1 and by the previous discussion, the sequence {U(m+1) }m=k
(m) +

is an increasing sequence of superharmonic functions in (k+1) which are har-

monic in (k+1) \ supp(d). Therefore, it either diverges to + everywhere
in (k+1) or it converges to a superharmonic function in (k+1) which is har-
monic in (k+1) \supp(d). Since k is arbitrary, Theorem 5.2 and the Monotone
Convergence Theorem give the next result.
Theorem 5.5 Let Rn be an open set with a Greens function in every one
of its components and d be a non-negative Borel measure in .
Assume that G x (y) d(y) < + for at least one x in every connected
component of .
Then the function

d
U (x) = G
x (y) d(y) , x,

is a non-negative superharmonic function in , harmonic in \ supp(d).

170 CHAPTER 5. GREENS FUNCTION

Also,
Ud = n d
as distributions in

Proof:
Since only the distribution equality remains to be proved, we take any k and
d d
observe that, for all m k, the difference U(m+1)(m)
U(k+1)
(k)
is harmonic in
(k) . This is true because, by the discussion before the theorem,
( d d(k) )
(m)
U(m+1) U(k+1) = n (d(m) d(k) ) = 0

as distributions in (k) .
d
Therefore, by the monotonicity of the sequence, the function Ud U(k+1)
(k)

is, also, harmonic in (k) and, thus,

d
Ud = U(k+1)
(k)
= n d(k) = n d

as distributions in (k) . Since k is arbitrary, the proof is finished.

Definition 5.3 Let Rn be an open set with a Greens function in every

one of its components
and d be a non-negative Borel measure in .
Assume that G x (y) d(y) < + for at least one x in every connected
component of .
Then the superharmonic function

Ud (x) = G
x (y) d(y) , x,

is called the Greens potential of d with respect to .

If d = f dm for some non-negative f locally integrable in , then the Greens
potential is, also, denoted by Uf and it is called the Greens potential of f
with respect to .

Lemma 5.1 Let be a bounded regular open set, d be a non-negative Borel

measure in defining its Greens potential with respect to .
Then the largest harmonic minorant of Ud in is identically 0 in .

Proof:
Without loss of generality, we may assume that is connected.
Consider any x0 with Ud (x0 ) < + and an open exhaustion {(m) }
of . Then,

d
U (x0 ) = Gx0 (y) d(y) + Gx0 (y) d(y)
(m) \(m)
5.8. GREENS POTENTIALS 171

where, by the Monotone Convergence Theorem, the last term tends to 0 as

m +.
Choose m0 so that

Gx0 (y) d(y) <  .
\(m0 )

Let dm0 be the restriction of d on (m0 ) and dm0 be the restriction of

d on \ (m0 ) .
We, first of all, have
d
U m0 (x0 ) <  .
We, then, observe that, for each y (m0 ) , the function G
y is positive and
harmonic in \ (m0 ) , and, by Proposition 5.4, it can become continuous in
\(m0 ) with values 0 everywhere in . Therefore, there is some large k > m0
so that Gy (x) <  for some particular value of y (m0 ) and all x \ (k) .
From Harnacks inequalities we get that there is a constant C > 0 so that
G
y (x) = Gx (y) < C

for all y (m0 ) and all x \ (k) .

This implies
d
U m0 (x) < Cd((m0 ) ) 
for all x \ (k) .
Now, let u be the largest harmonic minorant of Ud in . Since the identi-
cally 0 function is a harmonic minorant of Ud in , we have
u 0
in .
d
By the continuity of U m0 in (k) ,

(k) d
H d m0
(x0 ) = U m0 (x) dx0(k) (x) .
U (k)

dm0
Also, by the superharmonicity of U in (k) ,
(k) dm0
H dm0 (x0 ) U (x0 ) .
U

Hence,
(k)
u(x0 ) = Hu (k) (x0 ) H d (x0 )
U

(k) (k)
= H d m
(x0 ) + H d m
(x0 )
U 0 U 0

d d
U m0 (x) dx0(k) (x) + U m0 (x0 )
(k)

Cd((m0 ) )  +  .
172 CHAPTER 5. GREENS FUNCTION

Since  is arbitrary, we get u(x0 ) = 0 and, from the Maximum-Minimum

Principle, u is identically 0 in .

5.9 The Decomposition Theorem of F. Riesz

Theorem 5.6 (F. Riesz Decomposition) Let Rn be any open set. Suppose
that there is a superharmonic function u in which is not harmonic in any
component of and let d = 1n u be the associated non-negative Borel measure
in . Suppose, also, that u has a subharmonic minorant in .
Then has a Greens function in each of its components, the Greens poten-
tial of d is defined in and
u = Ud + u
everywhere in , where u is the largest harmonic minorant of u in .
Proof:
Let {(m) } be an open exhaustion of all of whose terms are regular. In
fact, it is easy to see that the usual construction, given in section 0.1.1, produces
(m) which satisfy the ball-criterion at every one of their boundary points.
If dm is the restriction of d in (m) , then, by Theorem 2.17,

u = Uhdm + wm
in (m) , where wm is a harmonic function in (m) . Therefore,

u = Ud(m)
m
+ vm

in (m) , where vm is another harmonic function in (m) . To see this, we observe

that, by Theorems 2.15 and 5.5, the functions Uhdm and Ud(m)m
have the same
distributional Laplacian in (m) and, hence, by Theorem 1.20, they differ by a
function harmonic in (m) .
If u is the largest harmonic minorant of u in , we have
u vm Ud(m)
m

in (m) and, since the largest harmonic minorant of Ud(m)

m
in (m) is, by Lemma
5.1, identically 0,
vm u
in (m) , whence
Ud(m)
m
u u
in (m) .
Taking any x0 with u(x0 ) < +, by Theorem 5.2 and the Monotone Con-
vergence Theorem,


lim Gx0(m) (x) d(x) u(x0 ) u (x0 ) < + .
m+
5.10. GREENS FUNCTION AND HARMONIC MEASURE 173

This, by Theorem 5.2 again, implies that has a Greens function in its
component which contains x0 , and hence in every one of its components.
Also,
Ud u u
in and we get that Ud is well defined as a superharmonic function in .
The functions vm decrease towards some harmonic function v in with

u = Ud + v

in and, hence,
u v
in .
On the other hand, from u = Ud + v v, we get

u v

in and we conclude
v = u
in , finishing the proof.

Theorem 5.7 Let Rn be any open set with a Greens function in each of
its components. Then the following are equivalent.
1. u is superharmonic in with largest harmonic minorant identically 0 in
.
2. u is the Greens potential with respect to of some (unique) non-negative
Borel measure in .
Proof:
That 1 implies 2 is just a consequence of Theorem 5.6.
The only thing that we have to prove is that, if Ud is a Greens potential,
then its largest harmonic minorant in is the constant 0.
If u is the largest harmonic minorant of Ud in , then, by Theorem 5.6,

Ud = Ud + u

in , where d = 1n Ud = d.
Thus, u = 0 in .

5.10 Greens function and harmonic measure

Lemma 5.2 Suppose that 0 < d < 1 and is a function defined in S(0; 1),
integrable with respect to d with the properties
1. = 0 in S(0; 1) \ B(e1 ; d),
174 CHAPTER 5. GREENS FUNCTION

2. |(y)| |y e1 |2 for all y in S(0; 1) B(e1 ; d),

where e1 = (1, 0, . . . , 0).

Then,

|grad P ( ; 0, 1)(te1 )| C(n)d
for all t with 1 d < t < 1, where C(n) depends only on the dimension.

Proof:
We have that

1 1 |x|2
P (x; 0, 1) = (y) d(y) .
n1 S(0;1) |y x|n

An easy calculation gives

1 |x|2 2xj 1 |x|2

= + n(yj xj ) .
xj |y x|n |y x| n |y x|n+2

Therefore, if x = te1 and 2 j n,

1 | |2 1 t2
(te 1 ) = nyj
xj |y |n |y te1 |n+2

and
 P ( ; 0, 1) 
  n 1 t2
 (te1 ) |yj | |y e1 |2 d(y)
xj n1 S(0;1)B(e1 ;d) |y te1 |n+2

2nd(1 t) |y e1 |2
d(y)
n1 V1 |y te1 |n+2

2nd(1 t) |y e1 |2
+ d(y) ,
n1 V2 |y te1 |n+2

where
V1 = {y S(0; 1) : |y e1 | < 1 t}
and
V2 = {y S(0; 1) : 1 t |y e1 | < d} .
The first integral is

C C(n)
|y e1 |2 d(y) ,
(1 t)n+2 V1 1t

while, the second integral is

1 C(n)
C d(y) .
V2 |y e1 |n 1t
5.10. GREENS FUNCTION AND HARMONIC MEASURE 175

We, thus, get

 P ( ; 0, 1) 
 
 (te1 ) C(n)d
xj
for 2 j n.
If j = 1,

1 | |2 2t 1 t2
(te 1 ) = + n(y1 t) .
x1 |y |n |y te1 |n |y te1 |n+2
2
The absolute value of the second term is nd |yte1t
1|
n+2 for all y S(0; 1)

B(e1 ; d) and the previous argument applies, word for word, to show that its
integral is C(n)d.
The absolute value of the first term is |yte
2
1|
n and, hence,

 P ( ; 0, 1) 
  2n |y e1 |2
 (te1 ) d(y) + C(n)d
x1 n1 S(0;1)B(e1 ;d) |y te1 |n

2n |y e1 |2
= d(y)
V1 |y te1 |
n1 n

2n |y e1 | 2
+ d(y) + C(n)d
n1 V2 |y te1 |n

C(n)
|y e1 |2 d(y)
(1 t)n V1

1
+ C(n) d(y) + C(n)d
V2 |y e1 |n2
C(n)d .
 
 P ( ;0,1) 
The above estimates of  xj (te1 ) for 1 j n conclude the proof.

Lemma 5.3 Suppose that {m } is a sequence of functions integrable in S(0; 1)

with respect to d and that m in L1 (S(0; 1), d).
If 0 < d < 1 and m = 0 identically in S(0; 1) B(e1 ; d) for all m, then

grad Pm ( ; 0, 1) grad P ( ; 0, 1)

uniformly in B(e1 ; d2 ).
All these gradients at points of S(0; 1) B(e1 ; d) are normal to S(0; 1).
Proof:
Since all m are 0 in S(0; 1) B(e1 ; d), we( can easily show )that, firstly, all
Pm ( ; 0, 1) are harmonic in the open set Rn \ S(0; 1) \ B(e1 ; d) and, secondly,

Pm ( ; 0, 1) P ( ; 0, 1)
( )
uniformly on compact subsets of Rn \ S(0; 1) \ B(e1 ; d) .
176 CHAPTER 5. GREENS FUNCTION

This implies the first statement and the second is due to the fact that S(0; 1)
B(e1 ; d) is, by Theorem 1.9, a common level surface of all Pm ( ; 0, 1) and of
P ( ; 0, 1).

Theorem 5.8 Suppose that u is harmonic in the open set , is open

relative to and is C 2 at every point of .
Let
lim u(x) = 0
3xy

for all y .

Then, grad u can be continuously extended in and at each point of
it is normal to .

If, also, there is an open V so that u > 0 in V , then grad u is
non-zero at every point of and has the direction towards .

Proof:
Take an arbitrary x0 and let > 0 be small enough so that B(x0 ; )
and, also, so that there is some defining function C 2 (B(x0 ; )) for
. Because of continuity, we may assume, taking a smaller if necessary, that

for some constants M0 , m0 > 0, grad (y) m0 for all y B(x0 ; ) and
maxxB(x0 ;),||=2 |D (x)| M0 .
Regarding the last statement of the theorem, we may, also, assume that is
small enough so that u > 0 in B(x0 ; ) .
Considering only y in B(x0 ; 31 ) and looking at the discussion
( at the
) end of
m0 1
section 0.1.2, we see that there is a fixed radius r0 = min 2Cn M0 , 6 so that,
for every y B(x0 ; 13 ),
there are two open balls b+ and b with common radius
r0 and mutually tangent at the point y so that

b+ B(x0 ; ) , b B(x0 ; ) \ .

It is obvious that the open ball B , which has the same center as b and
radius R0 = 3r0 , contains the ball b+ and is contained in B(x0 ; ).
Now, consider the function F continuous in the closed ring B \b , harmonic
in its interior B \ b , identically 1 in B and identically 0 in b .
If M is an upper bound for u in B(x0 ; ) , applying the Maximum-
Minimum Principle to the functions M F u in B , we find

|u(z)| M F (z)

for all z B . By explicitly writing the formula of F , we see that

F (z) K0 |z y|2

for all z b+ , where K0 is a constant depending only on the fixed r0 .

Therefore,
|u(z)| M K0 |z y|2
5.10. GREENS FUNCTION AND HARMONIC MEASURE 177

for all z b+ .
Consider, now, an arbitrary sequence {xm } in B(x0 ; ) with xm y.
Let ym be a point in of minimum distance from xm .
Then,

|ym y| |ym xm | + |y xm | 2|y xm | 0 ,

implying that ym y. We may, thus, assume that all ym are contained in

B(x0 ; 13 ) and, hence, we may construct the balls b+,m , b,m and B,m corre-
sponding to ym , whose radii r0 and R0 do not depend on m.
It is easy to see that these balls converge towards the balls b+ , b and B ,
respectively.
Since |ym xm | |y xm | 0, it is, also, easy to see that, for large enough
m, xm belongs to the radius of b+,m which goes through ym .
Take 0 < d < 1 and apply Lemma 5.2 to the ball b+,m , after the appropriate
dilation and translation.
If d,m = b+,m B(ym ; r0 d) and vm , wm are the restrictions of u on d,m
and b+,m \ d,m , respectively, then

u(xm ) = Pu (xm ; b+,m ) = Pvm (xm ; b+,m ) + Pwm (xm ; b+,m )

and

grad u(xm ) = grad Pvm ( ; b+,m )(xm ) + grad Pwm ( ; b+,m )(xm ) .

From Lemma 5.2, we get that, if k, m are so large that |xk yk | < r0 d and
|xm ym | < r0 d, then
 
grad u(xk )
grad u(xm ) 2CM K0 r0 d
 
+grad Pwk ( ; b+,k )(xk ) grad Pwm ( ; b+,m )(xm ) .

By Lemma 5.3, the convergence of the balls and the uniform continuity of
u, we have that there is a vector

vd , normal to at y, so that

grad Pwm ( ; b+,m )(xm )

vd .

Therefore,
 
lim sup grad u(xk ) grad u(xm ) 2CM K0 r0 d
k,m+

and, since d is arbitrary, {grad u(xm )} converges to some vector

v.
Thus,
   
grad Pwm ( ; b1,m )(xm )

v  CM K0 r0 d + grad u(xm )

v.

This implies that

|

vd

v | CM K0 r0 d
178 CHAPTER 5. GREENS FUNCTION

and, finally, that

v is normal to at y.
It is easy to see (combining two sequences into a single sequence) that

v
does not depend upon the sequence {xm } and we conclude that, defining

grad u(y) =

v ,

grad u becomes continuous in {y}.

The continuous dependence of v upon y B(x0 ; 31 ) is clear and, thus,
( )
grad u is continuously extended in B(x0 ; 13 ) and, since x0 is

arbitrary, grad u is continuously extended in .
Now, suppose that u > 0 in B(x0 ; ) .
As before, consider y B(x0 ; 13 ) , the corresponding ball b+ and an-
other open ball b0+ with the same center as b+ and radius equal to 12 r0 . It is
obvious that the open ring b+ \ b0+ is contained in B(x0 ; ) and is tangent
to at y.
If
(y) is the unit vector which is normal to at y and in the direction
towards the exterior of , then the line l containing this vector contains, also,
the center of b+ .
Consider the function G which is continuous in b+ \b0+ , harmonic in b+ \b0+ ,
identically 1 in b0+ and identically 0 in b+ .
If m > 0 is a lower bound of u in b0+ , then, by the Maximum-Minimum
Principle,
u mG
everywhere in b+ \ b0+ and, since G(y) = u(y) = 0,

G(y) G(x) u(y) u(x)

0 > m lim lim
l3xy |x y| l3xy |x y|
0
= lim grad u(x ) (y) = grad u(y)

(y) ,
l3xy

where x0 = x0 (x) is a point of the segment [x, y].

Therefore,

grad u(y) 6= 0

and grad u(y) is in the direction opposite to
(y) .

We present two proofs of the next result. The first is more straightforward
and its main ingredient is Greens Formula. (There is only an unpleasant tech-
nical detail in this proof, which is left to the interested reader to deal with.)
The idea in the second proof is that a certain kernel associated to an open set
with C 2 -boundary behaves like the Poisson kernel associated to a ball. (All
details in this proof are, actually, presented.)
Theorem 5.9 Let be a connected bounded open set with C 2 -boundary. Sup-
pose, also, that x0 ,

is the continuous unit vector field which is normal to
and in the direction towards the exterior of and dS is the surface measure
in .
5.10. GREENS FUNCTION AND HARMONIC MEASURE 179

Then,
1 G
x0
d
x0 = dS .
n

Moreover, d
x0 and dS are mutually absolutely continuous.

First proof:
By Proposition 3.4, is regular and, by Proposition 5.4, G
x0 can be consid-

ered continuous in \ {x0 } and identically 0 in . By Theorem 5.8, grad Gx0
can be continuously extended in \ {x0 } .
Since G x0 > 0 in , by the same theorem,

G
x0

x0 (y) (y) < 0
(y) = grad G

for every y .
Consider the open set

 = {x : G
x0 (x) > } .

This has the following properties.

1.  = {x : G x0 (x) } .
In fact, if xm x and G x0 (xm ) >  for all m, then, by the continuity in
\ {x0 }, we get G x0 (x) .
If, conversely, G x0 (x) = , then in every neighborhood of x there are points
where G x0 becomes larger than  and points where it becomes smaller than 
and, hence, x is in the boundary of  . Otherwise, by the Maximum-Minimum
Principle, G x0 would be constant in a neighborhood of x, and, by Theorem 1.10,
it would be constant in \ {x0 }, something impossible.
2.  = {x : G x0 (x) = }.
This was proved in the previous paragraph.
3.  is connected and contains x0 .
That  contains x0 is clear. If O is a component of  not containing
x0 , then, by 1 and 2, G x0 =  identically in O. By the Maximum-Minimum
Principle, G x0 is identically  in O and, hence, by Theorem 1.10, in \ {x0 },
which is impossible.
4. If  is small enough,  has C -boundary.

By the continuity of grad G 
x0 and its non-vanishing in , there is some

> 0 so that grad Gx0 (x) 6= 0 for all x with d(x, ) < . Now, G

x0 has a
positive minimum value in the compact set {x : d(x, ) }. Hence, if 

x0 (x) =  implies d(x, ) < and, thus, grad Gx0 (x) 6= 0.

is small enough, G

Therefore, the function Gx0 is a C defining function for in a neighbor-
hood of every one of its boundary points.
Now, fix a small  so that property 4 holds.
180 CHAPTER 5. GREENS FUNCTION

x0 hx0 is harmonic in
Considering an arbitrary u harmonic in , since G
, we get, by Greens formula,

1 G x0 1 u
u(y) (y) dS(y) (y)Gx0 (y) dS(y)
n  n 

1 hx0 1 u
= u(y) (y) dS(y) (y)hx0 (y) dS(y) .
n  n 

Here,

is the continuous unit vector field normal to  and in the direction
towards the exterior of  .
Therefore, first by Theorem 1.7 and then by Theorem 1.6,

1 G
x0 1 u
u(x0 ) = u(y) (y) dS(y) (y)Gx0 (y) dS(y)
n  n 

1 G 1 u
= u(y) x0 (y) dS(y)  (y) dS(y)
n  n 

1 G
= u(y) x0 (y) dS(y) .
n 

Assume, now, that u is continuous in . By the continuity of u, of G

x0 and
G
of x0
in \ {x0 }, we find, when  0, that

1 G
x0
u(x0 ) = u(y) (y) dS(y) .
n

The proof of this is quite technical and the main idea is in the discussion in
paragraph 4 of section 0.1.2. There is no actual need to see the details.
If f is any function continuous in , then, by the regularity of , the
function u = Hf is harmonic in and, extended as f in , is continuous in
.
Hence,

1 G
Hf (x0 ) = f (y) x0 (y) dS(y) ,
n
implying that
1 G
x0
d
x0 = dS
n
in .
G
x0
Since is continuous and negative in , there are two constants C1 , C2
so that
1 G
x0
0 < C1 C2
n
everywhere in .
We conclude that d
x0 and dS are mutually absolutely continuous.
5.10. GREENS FUNCTION AND HARMONIC MEASURE 181

Second proof:
G
The idea of this second proof is to show that the kernel 1n x behaves like
the Poisson kernel. Namely, it has the following four properties.
G
1. 1n x > 0 everywhere in .
G
2. (y) is, for every y , a harmonic function of x in .
x

1
Gx
3. n (y) dS(y) = 1.
G
4. If V is any neighborhood of y0 , then lim3xy0 1n x (y) = 0,
uniformly for y \ V .
We have already proved the first property and the third is an immediate
application of Theorem 1.7 with u = 1.
The second property can be proved as follows. Fix y , an open exhaus-
tion {(k) } of , a compact subset K of (k) and a sequence {xm } in \ (k)
converging to y over the line containing

(y).
For a fixed x0 K,
0
Gxm (x ) Gx0 (xm ) G x0
= (y)
|xm y| |xm y|
{ G (x0 ) }
as m +. In particular, the sequence |xxm m
y| is bounded and, since
G
every |xmxy|
m
is a positive harmonic function in (k) , Harnacks Inequalities
imply that this sequence of harmonic functions is uniformly bounded in K.
Since K is arbitrary, from Theorem 1.18, we get that there is some subsequence
converging to some function harmonic in (k) . But, for every x (k) ,

G
xm (x) G
x
(y)
|xm y|
G
implying that () (y) is harmonic in (k) and, hence, in .
For the fourth property and for V = B(y0 ; R), consider the compact set
K = S(y0 ; 12 R) and the open set U = {x : d(x; K) < 14 R} . We, also, define
the set 0 = U , which is, also, a connected open set.
The parts of and 0 in B(y0 ; 14 R) coincide and y0 , being a regular boundary
point of , is, also, a regular boundary point of 0 . Therefore, for every fixed
z 0 K, 0 0
0
lim G x (z ) = lim G z 0 (x) = 0
3xy0 3xy0

and, applying Harnacks Inequalities to the compact subset K of 0 ,

0
lim G
x (z) = 0 ,
3xy0

uniformly for z K. Since 0 , Theorem 5.1 implies that

lim G
x (z) = 0 ,
3xy0
182 CHAPTER 5. GREENS FUNCTION

uniformly for z K. Finally, by the Maximum-Minimum Principle,

lim G
x (z) = 0 ,
3xy0

uniformly for z \ B(y0 ; 12 R) .

Now, for an arbitrary y \ V there is some open ball b contained in
Rn \ and having y in its boundary. It was proved in the first part of the
proof of Theorem 5.8 that, if y is contained in a small enough neighborhood
of any boundary point, then the radius of b can be considered bounded from
below by a positive constant. Covering the compact set \ V by finitely many
such neighborhoods, we conclude that there is some fixed r so that the ball b ,
corresponding to the arbitrary y \ V , has radius r. Since b can be taken
smaller, if necessary, we may assume that r < 16 R.
Together with b , we, also, consider the open ball B which has the same
center as b and radius 2r and it is easy to see that B \ B(y0 ; 21 R)
for all y \ V .
Let F be the function which is continuous in the closed ring B \ b , har-
monic in the open ring B \ b , identically 1 in B and identically 0 in b .
It is, then, easy to calculate the number

F
(y) ,

where
(y) is the unit vector normal to b at y and, at the same time, normal
to at y and directed towards the exterior of . This number is negative and
depends only on r and, hence, not on y \ V .
Take, now, an arbitrary  and let x be close enough to y0 so that

x (z) 
G

for every z \ B(y0 ; 12 R). From the Maximum-Minimum Principle, we find

that
Gx (z)  F (z)

for all z B . Since G

x (y) = F (y) = 0,

G F
0 x
(y)  (y)

for all y \ V . This finishes the proof of property 4.

Now, take an arbitrary f C() and consider the function

1 G
u(x) = x
(y)f (y) dS(y) , x.
n

Take an arbitrary  > 0 and a small neighborhood V of y0 so that

|f (y) f (y0 )| <  for all y V . From properties 1, 3 and 4 of the kernel
5.11. AS INTERIOR POINT. MAINLY, N = 2 183

1 Gx
n , we get

1 G
lim sup |u(x) f (y0 )| lim sup x
(y)|f (y) f (y0 )| dS(y)
3xy0 3xy0 n

1 G
lim sup x
(y)|f (y) f (y0 )| dS(y)
3xy0 n V

1 G
+ lim sup x
(y)|f (y) f (y0 )| dS(y)
3xy0 n \V

1 G
 x
(y) dS(y)
n V
.

This implies that

lim u(x) = f (y0 )
3xy0

for all y0 . Now, this, together with property 2 of the kernel, says that
u is the solution of the Problem of Dirichlet in with boundary function f .
Therefore, for every f C(),

1 Gx
Hf (x) = (y)f (y) dS(y)
n
for all x , completing the second proof.

5.11 as interior point. Mainly, n = 2

We shall, now, consider an open set which contains .
By Theorems 4.4 and 4.6, potential theory in Rn is a triviality. We, there-
fore, assume that is not identical to Rn and, hence,

X
/,

for some X Rn .
Now, if x0 Rn , we consider the function
{ |xX|
log |xx 0|
, if n = 2
hx0 ,X (x) = hx0 (x) hX (x) = 1
|xx0 |n2 1
|xX|n2 , if n 3 .

This is harmonic in Rn \ {x0 , X} and, hence, in \ {x0 } .

Similarly, if x0 = , we consider the function
{
log |x X| , if n = 2
h,X (x) = h (x X) = 1
if n 3 ,
|xX|n2 ,

which is harmonic in Rn \ {, X} and, hence, in \ {} .

184 CHAPTER 5. GREENS FUNCTION

In any case, hx0 ,X is superharmonic in .

We consider the family Ux0 ,X of all superharmonic majorants of the sub-
harmonic function hx0 ,X in and, if this family is non-empty, the smallest
harmonic majorant Ux0 ,X .
We, now, define the function

G
x0 = hx0 ,X + Ux0 ,X .

If we take another point X 0

/ , then the function hx0 ,X hx0 ,X 0 is harmonic
in . Therefore, the function u is a superharmonic majorant of hx0 ,X in
if and only if the function u + hx0 ,X hx0 ,X 0 is a superharmonic majorant of
hx0 ,X 0 in . This implies that

Ux0 ,X 0 = Ux0 ,X + hx0 ,X hx0 ,X 0

in and, hence, the function G x0 , defined above, does not depend upon the
choice of X / .
The function Gx0 , if it exists, is called the Greens function of with respect
to x0 .
Observe that, if n 3, then h,X is harmonic in and, hence, G = 0
identically in .
In the following we shall avoid certain complications arising in case n 3
(and described, to some extent, in the previous chapter) and we shall concen-
trate on the case n = 2. In this case all results in this chapter extend in a
straightforward manner. We, briefly, describe the situation.

In all that follows, is an open subset of R2 with and X

/ .

1. The function G x0 , if it exists, is superharmonic in and harmonic in \

{x0 }, it is positive in the component of containing x0 and it is identically 0
in every other component of .
Also, the function G x0 hx0 ,X is harmonic in .
2. If some disc in R2 is disjoint from , then has a Greens function with
respect to every x0 and

Gx0 (x) = hx0 ,X (x) + H
hx0 ,X (x) = hx0 ,X (x) hx0 ,X (y) d
x (y)

for all x .
3. The open set R2 \ {X} has no Greens function with respect to any one of
its points.
4. If = R2 \ B(X; R), then

x0 = hx0 hx
G 0
+ log R ,

if x0 \ {} and x0 is the symmetric of x0 with respect to S(X; R), and

= h,X log R .
G
5.11. AS INTERIOR POINT. MAINLY, N = 2 185

5. If x0 0 and has a Greens function with respect to x0 , then 0 has,

also, a Greens function with respect to x0 and
0
x0 Gx0
G

in 0 .
6. If m , x0 1 and all m have a Greens function with respect to x0 ,
then, either Gx0m + in the component of containing x0 and has no
Greens function with respect to x0 , or has a Greens function with respect to
x0 and
G m
x0 Gx0

in .
7. If x0 and x1 are in the same component of , then has a Greens function
with respect to x0 if and only it has a Greens function with respect to x1 and,
in this case,
G
x0 (x1 ) = Gx1 (x0 ) .

We say that has a Greens function in one of its components, if it has a

Greens function with respect to at least one x0 in this component.
8. Let y0 . If
lim G x (z) = 0
3zy0

for at least one x in every component of , then y0 is a regular boundary point

of .
If there is some disc disjoint from and y0 is a regular boundary point, then
lim G
x (z) = 0
3zy0

for all x .
9. If is regular and disjoint from some disc, x0 and O is the component
of containing x0 , then G x0 extended as identically 0 in R \ is subharmonic
2

in R \ {x0 } and identically 0 in R \ O.

2 2

10. If there is some disc disjoint from , x0 and O is the component of

containing x0 , then the function

hx0 ,X (x) hx,X (y) dx0 (y) , x R2 ,

is an extension of Gx0 in R which is subharmonic in R \ {x0 } and identically

2 2

0 in R2 \ O.
11. If the connected is disjoint from some disc and has C 2 -boundary, x0 ,

is the continuous unit vector field which is normal to directed towards the
exterior of and dS is the surface measure in , then,
1 G
x0
d
x0 = dS .
n
Moreover, d
x0 and dS are mutually absolutely continuous.
186 CHAPTER 5. GREENS FUNCTION
Chapter 6

Potentials

6.1 Definitions
We shall consider the following two types of kernels.
Definition 6.1 Kernels of first type.
Let {
log 1r , if n = 2
h (r) = 1
r n2 , if n 3 .
Take any non-constant increasing convex function H defined in (, +),
in case n = 2, or in (0, +), in case n 3, and define
( )
K (r) = H h (r) , 0 < r < + .

Hence, K is continuous and decreasing in (0, +) with limr0+ K (r) = +

and we, next, define the kernel
( ) ( )
K(x, y) = K (|x y|) = H h (|x y|) = H h(x y)

for all x, y Rn .
We postulate the following rules.
1
1. 0 K (r)rn1 dr < +.
2. limr+ K (r) 0, in case n = 2, or limr+ K (r) = 0, in case n 3.
K (r1)
3. If K > 0 in (0, +), then lim supr+ K (r) < +.

Kernels of second type.

If is an open subset of Rn with a Greens function in every one of its
connected components, we consider the kernel

G (x, y) = G
x (y) = Gy (x)

for all x, y .

187
188 CHAPTER 6. POTENTIALS

Comments 1. Observe that, in case n 3, all our kernels are non-negative.

Also, observe that all kernels are symmetric: K(x, y) = K(y, x).
2. The third rule is not needed if in the theory, which we shall develop, we
restrict to the consideration of measures with compact support.

Examples
1. The Riesz kernel of order is defined by
1
K (x, y) = K (|x y|) = ,
|x y|n
where 0 < < n.
If 0 < < 2, in case n = 2, or 0 < 2, in case n 3, then K is of first
type.
2. The classical kernels, i.e. the Newtonian, in case n 3, and the logarith-
mic, in case n = 2, are of first type.

Definition 6.2 K-potential for a kernel of second type.

Let Rn have a Greens function in all its components and d be a non-
negative Borel measure in . The corresponding kernel is K = G and we have
already defined the Greens potential

d d
UK (x) = U (x) = G (x, y) d(y) , x,

only in case this is finite for at least one x in each component of .

d
If d is a locally finite complex Borel measure in , we define UK by the
|d|
same formula (and linearity) only when UK (x) is finite for at least one x in
each component of .
Under these assumptions, we say that the K-potential is well-defined.
We know, from last chapter, that, if K is of second type, the K-potential of
a non-negative Borel measure in the associated is (if it is well-defined) super-
harmonic in and harmonic outside the support of the measure. Therefore, the
K-potential of a locally finite Borel measure in is (if it is well-defined) finite
almost everywhere in and is a linear combination of (four) superharmonic
functions.
Definition 6.3 K-potential for a kernel of first type.
If d is a non-negative Borel measure in Rn and K is a non-negative kernel of
first type, we define the K-potential of d by

d
UK (x) = K(x, y) d(y) , x Rn ,
Rn

only when this is finite for at least one x Rn .

If d is a locally finite complex Borel measure in Rn (and K is non-negative),
d |d|
we define UK by the same formula (and linearity), only when UK (x) is finite
for at least one x Rn .
6.2. POTENTIALS OF NON-NEGATIVE BOREL MEASURES 189

If K is of variable sign and d is a locally finite complex (non-negative, in

d
particular) Borel measure in Rn , we define UK as before, but only if d is
compactly supported.
Under these assumptions, we say that the K-potential is well-defined.
It is obvious that, if K is a non-negative kernel of first type and d is a non-
negative Borel measure, then the K-potential is defined everywhere as either
a non-negative number or as +. Proposition 6.1, which will be proved in a
moment, describes the situation more clearly. If d is a locally finite complex
Borel measure with real values (and K is non-negative), then its K-potential is
d+ d
defined at those points where not both UK and UK take the value +. A
similar comment can be made for a general locally finite complex Borel measure.

6.2 Potentials of non-negative Borel measures

Lemma 6.1 If K is a kernel of first type, then, for every y Rn , K( , y) is
continuous and subharmonic in Rn \ {y} .

Proof:
It is clear, since h is harmonic in Rn \{0} and H is convex in an open interval
containing the values of h.

Proposition 6.1 1. If d is a non-negative Borel measure in Rn , K is of first

d
type and UK is well-defined, then this K-potential is continuous and subhar-
monic in Rn \ supp(d).
If, in particular K = h, then the h-potential is superharmonic in Rn and
harmonic in Rn \ supp(d).
2. If d is a non-negative Borel measure in Rn , which has a Greens
function in all its components, and Ud is well-defined, then this K-potential is
superharmonic in and harmonic in \ supp(d).
Proof:
The case of second type is just Theorem 5.5 and the case K = h is only
Theorem 2.8. Hence, assume that K is of the first type and, to begin with, that
K is non-negative.
d
By definition, UK (x0 ) < + for some x0 Rn .
Therefore, for an arbitrary  > 0,

K(x0 , y) d(y)  ,
{y:|y|>R}

for all large R. Take any x Rn \ supp(d) and, by the third rule on K, find
a constant C > 0 and R so large that, besides the previous inequality,

K(x, y) CK(x0 , y)

for all y with |y| > R, is, also, true.

190 CHAPTER 6. POTENTIALS

This implies,
K(x, y) d(y) C .
{y:|y|>R}

Since K(x, ) is bounded on supp(d),

d
UK (x) K(x, y) d(y) + C < + .
{y:|y|R}

If, now, xm x, then K(xm , ) K(x, ) uniformly in B(0; R) supp(d)

and, thus,  d 
lim supUK d
(xm ) UK (x) 2C ,
m+
d
proving the continuity of UK at x.
From Lemma 6.1, for every B(x; r) Rn \ supp(d),

d
MrU d (x) = MrK( ,y) (x) d(y) K(x, y) d(y) = UK (x)
K supp(d) supp(d)

d
and UK is subharmonic in Rn \ supp(d).
If K is of variable sign, then, by definition, d is supported in a compact
set and we may choose the R above so that B(0; R) contains the support of d.
The proof of continuity of the K-potential in R2 \supp(d) is, now, easier, since
there is no tail-term in the integral.

Comment: Suppose that K is a non-negative kernel of first type and d is a

d
non-negative Borel measure in Rn . If UK (x0 ) < + for some x0 (the condition
for the K-potential to be well-defined), then the finiteness of the K-potential
at any other x depends only on the behaviour of d in a neighborhood of x.
In fact, for all large R, we have that {y:|y|>R} K(x0 , y) d(y) < 1. As in
the last proof, there is some C > 0 and some large R, so that, besides the last
inequality,
we also have K(x, y) CK(x0 , y) for all y with |y| > R. This implies
that {y:|y|>R} K(x, y) d(y) < C and, thus, the finiteness of the K-potential
depends on the restriction of d in B(0; R).
The same comment is valid for the Greens potentials.
In fact, suppose that Rn has a Greens function in all its components
and take an (m) from some open exhaustion of so that it contains x and x0
(assumed to be in the same component of ) and \(m) G (x0 , y) d(y) < 1.
exists a C > 0 so that G (x, y) CG (x0 , y)

By Harnacks Inequalities, there
for all y \ (m) , implying \(m) G (x, y) d(y) < C.

Proposition 6.2 (Lower-semicontinuity in the space-variable)

d
1. If d is a non-negative Borel measure in Rn , K is of first type and UK is
well-defined, then this K-potential is lower-semicontinuous in Rn .
2. If d is a non-negative Borel measure in Rn , which has a Greens
function in all its components, and Ud is well-defined, then this K-potential is
lower-semicontinuous in .
6.2. POTENTIALS OF NON-NEGATIVE BOREL MEASURES 191

Proof:
This is a simple application of Fatous Lemma.

Proposition 6.3 (Lower-semicontinuity in measure)

1. Let K be of first type, {dk } be a sequence of non-negative Borel measures
dk
with lim inf k+ UK (x) < + for at least one x. If this sequence converges
weakly on compact sets to some non-negative Borel measure d, then d has a
well-defined K-potential and
dk d
lim inf UK (x) UK (x)
k+

for every x Rn .
If the kernel is of variable sign, we, also, assume that all dk are supported
in a common compact subset of Rn .
2. Let Rn have a Greens function in all its components and {dk } be a
sequence of non-negative Borel measures in with lim inf k+ Udk (x) < +
for at least one x in each component of . If the sequence converges weakly on
compact subsets of to some non-negative Borel measure d in , then d has
a well-defined Greens potential and

lim inf Udk (x) Ud (x)

k+

for every x .
Proof:
1. Assume that K is of first type and non-negative and consider the truncated
kernel
KN (x, y) = min(K(x, y), N )
for all x, y Rn .
For every x, KN (x, ) is continuous in Rn and, taking an arbitrary R > 0,

dk
lim inf UK (x) lim inf KN (x, y) dk (y) = KN (x, y) d(y) .
k+ k+ B(0;R) B(0;R)

Now, letting R + and then N +, we conclude the proof in this case.

If K is of variable sign, then we repeat the same proof, replacing the arbitrary
ball with a single compact set F R2 so that all dk are supported in F .
2. We consider, again, the truncated kernels G N and any open exhaustion
{(m) } of .
Then, as before,

lim inf Udk (x) lim inf GN (x, y) d k (y) = GN (x, y) d(y)
k+ k+ (m) (m)

and the proof is finished, by letting m + and N +.

192 CHAPTER 6. POTENTIALS

6.3 The maximum principle for potentials

Lemma 6.2 Let be a regular bounded open set and d be a non-negative Borel
measure supported in a compact subset of . Then, lim3xy Ud (x) = 0, for
all y .

Proof:
For fixed z supp(d) and y , we have that lim3xy G z (x) = 0.
It is easy to see, by Harnacks Inequalities, that lim3xy G
z (x) = 0 uni-
formly for z supp(d) and the proof is, now, clear.

Theorem 6.1 (The Maximum Principle)

1. Let K be of first type and d be a non-negative Borel measure in Rn . If
d d
> 0 and UK (x) for all x supp(d), then, UK everywhere in Rn .
2. Let R have a Greens function in all its components and d be a non-
n

negative Borel measure in . If > 0 and Ud (x) for all x supp(d),

then, Ud everywhere in .

Proof:
1. Let K be of first type and suppose that d is compactly supported. We shall
consider, for any N > 0, the truncated kernel
{
K(x, y) , if K(x, y) N
K N (x, y) =
0, if K(x, y) > N

and the function

d
UK N (x) = K N (x, y) d(y) , x Rn .
Rn

Then, K N K and,
d d
UK N (x) UK (x)

for all x.
Applying Egoroffs Theorem, we find that, for every  > 0, there is a closed
set F supp(d) with d(F ) > d(Rn )  and
d
UK N U d

uniformly in F .
Consider, also, dF , the restriction of d in F .
Then, for every x,

dF
U dF (x) UK N (x) = K(x, y) dF (y)
{y:K(x,y)>N }

d d
K(x, y) d(y) = UK (x) UK N (x)
{y:K(x,y)>N }
6.3. THE MAXIMUM PRINCIPLE FOR POTENTIALS 193

and, thus,
dF dF
UK N UK
uniformly in F .
Therefore, for every > 0, if N is large enough, we have

K(x, y) dF (y)
{y:K(x,y)>N }

for all x F .
Fix, now, x F and take any {xm } converging to x. Then,

dF
lim sup UK (xm ) lim sup K(xm , y) dF (y)
m+ m+ {y:K(xm ,y)>N }

+ lim sup K N (xm , y) dF (y)
m+ Rn

= lim sup K(xm , y) dF (y)
m+ {y:K(xm ,y)>N }

+ K N (x, y) dF (y) .
Rn

It is, geometrically, clear that there is some number M , depending only on

the dimension n, with the property that, for every z, we can find closed convex
cones z1 , . . . , zM with the same vertex z and each having an opening of 6 from
its axis of symmetry so that

M z n
k=1 k = R .

Now, if y zk F and kz is a closest point of zk F from z, then

|z y| |kz y| .

Applying this to every z = xm ,

M

K(xm , y)dF (y) K(xm , y)dF (y)
{y:K(xm ,y)>N } k=1 {y:K(xm ,y)>N }x
k
m

M

K(kxm , y)dF (y)
k=1 {y:K(xm ,y)>N }x
k
m

M

K(kxm , y)dF (y)
xm
k=1 {y:K(k ,y)>N }

M .

Therefore, we have that

dF
lim sup UK (xm ) M + K(x, y) dF (y) .
m+ {y:K(x,y)N }
194 CHAPTER 6. POTENTIALS

Now, let N + and, then, 0 and get

dF dF d
lim sup UK (xm ) UK (x) UK (x) m m ,
m+
( )
where m = min 0, minz,ysupp(d) K(z, y) .
dF
Combining the last result with Proposition 6.2, we get that UK is contin-
uous at every x F and, by Proposition 6.1, it is continuous in Rn .
From the second rule on our kernels, we, also, have that
dF
lim sup UK (x) 0 .
x

dF
By the subharmonicity of UK in Rn \ F and the Maximum Principle, we
get
dF
UK m
everywhere in Rn .
Now, let x
/ supp(d) and let > 0 be the distance of x from supp(d).
Then,

d
UK (x) = K(x, y) d(y) + K(x, y) d(y)
F supp(d)\F
m + K ()

and, since  is arbitrary,

d
UK (x)
for all x.
If d is not compactly supported, in which case K is, necessarily, non-
dm
negative, we consider the restrictions dm in B(0; m) and we have UK (x)
d
UK (x) for all x supp(d) and, hence, for all x supp(dm ). From what
dm
we proved up to now, UK (x) everywhere in Rn and letting m +,
we finish the proof in this case.
2. Let have a Greens function in all its components and d be a non-
negative Borel measure with compact support contained in . From the first
part of the proof, we know that there is a compact subset F of supp(d) with
d(F ) d(Rn )  so that the h-potential UhdF is continuous in Rn . We
know, from Theorems 2.15 and 5.5, that UdF and UhdF have the same distri-
butional derivative in and, hence, they differ by a function harmonic in .
This implies that UdF is continuous in .
Assume, for the moment, that is a regular bounded set. From Lemma 6.2,

lim UdF (x) = 0

3xy

for all y . Since, also,

lim UdF (x) = UdF (y) Ud (y)

3xy
6.4. THE CONTINUITY PRINCIPLE FOR POTENTIALS 195

for all y F , we get, by the Maximum Principle, that

UdF (x)

for all x \ F and, hence, for all x .

For an arbitrary x \ supp(d), denote M = supysupp(d) G (x, y) <
+. Then,
Ud (x) UdF (x) + M  + M 
and, thus,
Ud (x)
for all x \ F and, thus, for all x .
To drop the assumption of regularity, take an open exhaustion {(m) } of
consisting of regular sets and large m so that supp(d) (m) . Since Ud(m)
Ud in supp(d) and (m) is regular, we get

Ud(m) (x)

for all x (m) . By Theorem 5.2,

Ud (x)

for all x .
Finally, if d does not have compact support in , we consider the restric-
tions dm in the terms (m) of some open exhaustion of and we conclude the
proof in the same manner as in the previous paragraph.

6.4 The continuity principle for potentials

Proposition 6.4 1. If K is a kernel of first type and d is a non-negative
Borel measure with compact support, there is some closed F supp(d) with
dF
arbitrarily small d(Rn \ F ) so that UK is continuous in Rn .
2. If R has a Greens function in all its components and d is a non-
n

negative Borel measure with compact support in , there is some closed F

supp(d) with arbitrarily small d( \ F ) so that UdF is continuous in .

Proof:
It is, actually, part of the proof of Theorem 6.1.

Theorem 6.2 (The Continuity Principle)

1. If K is a kernel of first type, d is a non-negative Borel measure with compact
d
support and UK , restricted to supp(d), is continuous in supp(d), then it is
continuous in Rn .
2. If Rn has a Greens function in all its components, d is a non-negative
Borel measure with compact support in and Ud , restricted to supp(d), is
continuous in supp(d), then it is continuous in .
196 CHAPTER 6. POTENTIALS

Proof:
1. Following the argument in the proof of the Theorem 6.1, we have that
d d d d
UK N UK , that each UK N is continuous everywhere and that UK , restricted

to supp(d), is continuous in supp(d).

By Dinis Theorem,
d d
UK N UK
uniformly in the compact supp(d).
Therefore, in the same proof, we do not need to reduce supp(d) to any
d
subset F . The final result there is that UK is continuous everywhere.
2. Since Ud and Uhd differ by a harmonic function in , the proof in this case
is straightforward, by applying the first part to Uhd .
Chapter 7

Energy

7.1 Definitions
Definition 7.1 K-energy for kernels of second type.
Let Rn be an open set with a Greens function in all its components and
d1 and d2 two non-negative Borel measures in . We define the mutual
K-energy or mutual -energy of d1 and d2 by

IK (d1 , d2 ) = I (d1 , d2 ) = G (x, y) d1 (x) d2 (y) ,

which is either a non-negative number or +.

If d1 and d2 are two locally finite complex Borel measures in , we de-
fine their mutual -energy by the same formula (and linearity), but only when
IK (|d1 |, |d2 |) < +.
In case d1 = d2 = d, we call K-energy or -energy of d the
I (d) = I (d, d) .
Definition 7.2 K-energy for kernels of first type.
Let K be a non-negative kernel of first type and d1 and d2 two non-negative
Borel measures in Rn . We define the mutual K-energy of d1 and d2 by

IK (d1 , d2 ) = K(x, y) d1 (x) d2 (y) .
Rn Rn

If d1 and d2 are two locally finite complex Borel measures in Rn , we

define their mutual K-energy by the same formula (and linearity), but only
when IK (|d1 |, |d2 |) < +.
In case d1 = d2 = d, we call K-energy of d the
IK (d) = IK (d, d) .
If K is of variable sign, we define the mutual K-energy and K-energy as
before, but only for compactly supported Borel measures.

197
198 CHAPTER 7. ENERGY

It is clear, by the Theorem of Fubini, that if IK (|d1 |, |d2 |) < +, then

d1
UK (x) is a finite number for almost every x with respect to d2 and vice-versa.

Example.
IK (da , db ) = K(a, b)
for all types of kernels and all a and b. In particular, the K-energy of a Dirac
mass is always +.

Proposition 7.1 (Lower-semicontinuity of energy in measure)

1. Let K be of first type and {d1m } and {d2m } be two sequences of non-negative
Borel measures in Rn converging to the non-negative Borel measures d1 and
d2 , respectively, weakly on compact sets. Then,

lim inf IK (d1m , d2m ) IK (d1 , d2 ) .

m+

In case the kernel K is of variable sign, we, also, assume that all d1m and
d2m are supported in a common compact subset of Rn .
2. Let Rn have a Greens function in all its components and {d1m } and
{d2m } be two sequences of non-negative Borel measures in converging to
the non-negative Borel measures d1 and d2 , respectively, weakly on compact
subsets of . Then

lim inf I (d1m , d2m ) I (d1 , d2 ) .

m+

Proof:
1. Assume that K is of first type and non-negative and consider the truncated
kernel
KN (x, y) = min(K(x, y), N )
for all x, y Rn .
Now, KN is continuous in Rn Rn and, taking an arbitrary R > 0,

lim inf IK (dm , dm ) lim inf
1 2
KN (x, y) d1m (x) d2m (y)
m+ m+ B(0;R) B(0;R)

= KN (x, y) d1 (x) d2 (y) ,
B(0;R) B(0;R)

since the product measures d1m d2m converge to d1 d2 weakly on compact

sets in Rn Rn .
Now, by the Monotone Convergence Theorem, letting N + and then
R +, we conclude the proof in this case.
If K is of variable sign, then we repeat the same proof, replacing the arbitrary
ball with a single compact set F so that all dm are supported in F .
2. Now, assuming that K = G , consider, besides KN , any open exhaustion
{(k) } of .
7.2. REPRESENTATION OF ENERGY: GREENS KERNEL 199

Then, as before,

lim inf I (d1m , d2m ) lim inf KN (x, y) d1m (x) d2m (y)
m+ m+ (k) (k)

= KN (x, y) d1 (x) d2 (y)
(k) (k)

and the proof is finished, by letting k + and N +.

7.2 Representation of energy: Greens kernel

Let be any open set in Rn with a Greens function in all its components. In
n
case n 3 and = Rn , the kernel GR x coincides with the Newtonian kernel
hx .
The basis of all results in this section is the following simple
Lemma 7.1 Suppose that is a bounded open set with C 2 -boundary. Let g1
and g2 be two non-negative functions in D().
Then

1
I (g1 dm, g2 dm) = grad Ug1 (x) grad Ug2 (x) dm(x) .
n
Proof:
By Theorems 2.18 and 5.5, the potential Ugi is in C () and
Ugi = n gi
as distributions in .
Since both sides in this equation are continuous functions, it holds in the
classical sense everywhere in .
We, also, get that Ugi is harmonic in the open set \ supp(gi ). Since
is a regular set, by Lemma 6.2, Ugi can be considered continuous in and
identically 0 in .

Therefore, by Theorem 5.8, grad Ugi extends continuously in .
Now, applying Greens Formula,

I (g1 dm, g2 dm) = G (x, y)g1 (x) dm(x) g2 (y) dm(y)

= Ug1 (y)g2 (y) dm(y)

1
= U g1 (y)Ug2 (y) dm(y)
n

1
= grad Ug1 (y) grad Ug2 (y) dm(y)
n

1 U g2
+ Ug1 (z) (z) dS(z)
n

1
= grad Ug1 (y) grad Ug2 (y) dm(y) ,
n
200 CHAPTER 7. ENERGY

being the continuous unit vector field normal to and directed towards the
exterior of .

Remark
Observe that, under the assumptions of Lemma 7.1, the same application of
Greens Theorem implies

G (x, y)g1 (x) dm(x) g2 (y) dm(y)


1
= grad Ug1 (y) grad Ug20 (y) dm(y) ,
n
whenever 0 is another bounded open set with C 2 -boundary.
Proposition 7.2 Suppose that the open set Rn has a Greens function
in all its components and that d is a non-negative Borel measure in with
I (d) < +.
Then Ud is a superharmonic function in and it is finite almost everywhere
with respect to d.
Proof:
If d is the zero measure in any of the components of , then the result is
obvious in those components. Otherwise, the result is only an application of
Fubinis Theorem.

Proposition 7.3 If the open set Rn has a Greens function in all its com-
ponents and d1 , d2 are two non-negative Borel measures in with I (d1 ) <
+ and I (d2 ) < +, then

1
I (d1 , d2 ) = grad Ud1 (x) grad Ud2 (x) dm(x) < + .
n
We, also, have that Ud1 is finite almost everywhere with respect to d2 and
vice-versa.
Proof:
1. Assume, first, that both di are supported in a compact subset A of and
that is bounded with C 2 -boundary.
Each Udi is superharmonic in , harmonic in \ A and, since is regular,
it can, by Lemma 6.2, be considered continuous in Rn \ A and identically 0 in
Rn \ .
Consider any approximation to the identity { : > 0} and the convolution

Udi (x) = Udi (x y) (y) dm(y) , x .
Rn

Then, the functions Udi are in C ( ) and, by Theorem 2.10, they

are superharmonic in and
Udi Udi
7.2. REPRESENTATION OF ENERGY: GREENS KERNEL 201

in as 0.
Taking < 13 d(A, ), we know, from the beginning of section 1.3, that
( )
U = Udi in \ A . Therefore, the function
di

{
Udi , if x
vi, =
Udi , if x \ ,

is in C (), is superharmonic in , harmonic in \ A and can be considered

continuous in Rn \ A and identically 0 in Rn \ . Also, vi, Udi in Rn as
0.
Consider, also, the convolution

di (x) = (x y) di (y) , x Rn .
Rn

The non-negative functions di are in C (Rn ) and they are supported

in A = A + B(0; ), a compact subset of .
Therefore, the functions Udi are superharmonic in , harmonic in \
A and, by Theorem 2.18, they are in C (). By Lemma 6.2, they can be
considered continuous in Rn and identically 0 in Rn \ .
Employing the informal notation for distributions,
( )
Udi = Udi = n di = Udi

in .
Therefore, the functions vi, and Udi differ by a function harmonic in
and, hence, in . Both functions are continuous in and identically 0 in
. By the Maximum-Minimum Principle,

vi, = Udi

identically in .
Apply, now, Lemma 7.1 to the functions gi = di i , for i < 13 d(A, ),
to get

d
U 1 1 (x)d2 2 (x) dm(x)

d
1 d1 1 2 2
= grad U (x) grad U (x) dm(x)
n

1
= grad v1,1 (x) grad v2,2 (x) dm(x) .
n
d
Since U 1 1 = v1,1 Ud1 in as 1 0, the left side of the last
equality tends to
Ud1 (x)d2 2 (x) dm(x)

as 1 0.
202 CHAPTER 7. ENERGY

Regarding the right side, it is true, by Theorem 2.18, that Udi has partial
derivatives at almost every point in which are locally integrable in . From
d
U i
Theorem 5.8, we have that xj can be continuously extended in \ A and,
hence, if we further extend it as identically 0 in Rn \ , it becomes a function
in L1 (Rn ). Applying Lemma 0.4, we find that there is a sequence {i } so that
d d
U i U i vi,i
i 0 and i almost everywhere in Rn . Since, =
( d ) xj xj xj
i d
U U i
i in i , we get
i
xj = xj

vi,i Udi

xj xj
almost everywhere in as i 0.
From Lemma 7.1,
 
1  2 di i
grad vi,i (x) dm(x) = U (x)di i (x) dm(x)
n

Udi (x)di i (x) dm(x)

di i
= U (x) di (x)

Udi (x) di (x)

= I (di ) < + .
From our last two results, we conclude that there is some sequence of i s so
that, for all j,
vi,i Udi

xj xj
weakly in L2 ().
Letting this sequence of 1 s tend to 0,

Ud1 (x)d2 2 (x) dm(x)


1
= grad Ud1 (x) grad v2,2 (x) dm(x) .
n
d2 2
The left side is equal to U (x) d1 (x) and, following the same pro-
cedure with 2 0, we find

1
I (d1 , d2 ) = grad Ud1 (x) grad Ud2 (x) dm(x) .
n
2. Modifying slightly the proof of part 1 and using the Remark after Lemma
7.1, we, easily, prove that, under the same assumptions,

1
G (x, y) d1 (x) d2 (y) = grad Ud1 (x) grad Ud0 2 (x) dm(x) ,
n
7.2. REPRESENTATION OF ENERGY: GREENS KERNEL 203

whenever 0 is another bounded open set with C 2 -boundary.

3. Now, assume that has a Greens function in all its connected components
and that d1 and d2 are two non-negative Borel measures supported in a
compact subset A of .
Consider any open exhaustion {(m) } of , each (m) having C 2 -boundary
and containing A.
From part 2, we have that
d

1
I(m) (d1 , d2 ) = 1
grad U(m) (x) grad Ud(m2 0 ) (x) dm(x) ,
n (m)

for every m, m0 with m m0 .

Since G(m) G in , the right side has I (d1 , d2 ) as its limit when
m +.
We, also, have that Ud(m)
i
= Uhdi + vm,i , where vm,i is harmonic in (m) .
Since Ud(m)
i
Udi as m +, we get that {vm,i } increases towards a har-
monic function vi in as m + and, hence, Udi = Uhdi + vi in . By

Theorems 1.16 and 1.17, grad vm,i grad vi everywhere in as m +,
implying
d

i
grad U(m) grad Udi
almost everywhere in as m +.
From part 1, we get

1 d 2
 i 
grad U(m) (x) dm(x) = G(m) (x, y) di (x) di (y) I (di ) .
n (m)

Hence, there is some sequence of ms so that

d

1
grad U(m) grad Ud1

weakly in L2 () and some sequence of m0 s so that

d

grad U(m2 0 ) grad Ud2

weakly in L2 ().
We, now, let, first, m0 + and, then, m + through these sequences
and get

1
I (d1 , d2 ) = grad Ud1 (x) grad Ud2 (x) dm(x) .
n
4. Consider, finally, the general case.
Take the measures dm,i , which are the restrictions of di to the terms of
an open exhaustion {(m) } of .
By the assumption I (di ) < + and Proposition 7.2, it is implied that
d
Udi and all U m,i are superharmonic in .
204 CHAPTER 7. ENERGY

dk,i dm,i
Therefore, for any k > m, U U is harmonic in (m) and

dk,i dm,i dm,i

U U Udi U

in (m) as k +. The last function is harmonic in (m) and, from Theorems

1.16 and 1.17,
(
d
)
d
( )
d
grad U k,i U m,i grad Udi U m,i

in (m) as k +.
dm,i
Since U has partial derivatives almost everywhere in , we get that

d
grad U k,i grad Udi

almost everywhere in (m) and, since m is arbitrary, the last limit holds almost
everywhere in .
From the equality of part 3,

1  
grad U dm,i (x)2 dm(x) = I (dm,i ) I (di ) ,

n

there is a sequence of ms so that

dm,i
U Udi

weakly in L2 ().
If in the equality of part 3,

G (x, y) dm1 ,1 (x) dm2 ,2 (y)
d
1 d
= grad U m1 ,1 (x) grad U m2 ,2 (x) dm(x) ,
n

we let, first, m1 and, then, m2 tend to +, we find

1
G (x, y) d1 (x) d2 (y) = grad Ud1 (x) grad Ud2 (x) dm(x) .
n

Regarding the last statement of the theorem, observe that, from the Cauchy-
Schwartz inequality,

I (d1 , d2 )2 I (d1 )I (d2 ) < +

and, hence, from Fubinis Theorem, Ud1 is finite almost everywhere with re-
spect to d2 and vice-versa.
7.3. MEASURES OF FINITE ENERGY: GREENS KERNEL 205

7.3 Measures of finite energy: Greens kernel

Definition 7.3 Let the open Rn have a Greens function in all its con-
nected components. Then,

W = {d : d is a locally finite measure in with I (|d|) < +}

is called the space of measures in of finite -energy.

Theorem 7.1 Let the open Rn have a Greens function in all its connected
components. Then, for all d1 and d2 in W the bilinear form I (d1 , d2 ) is
a complex number and

1
I (d1 , d2 ) = grad Ud1 (x) grad Ud2 (x) dm(x) .
n

Under this bilinear form, W becomes an inner product space.

Proof:
If d1 and d2 are two locally finite complex Borel measures in of finite
-energy, then, by Proposition 7.3 and the Cauchy-Schwartz inequality,

I (|d1 + d2 |) I (|d1 | + |d2 |)

= I (|d1 |) + 2I (|d1 |, |d2 |) + I (|d2 |)
 2
1  |d | 
= grad U 1 (x) dm(x)
n

2 |d | |d |
grad U 1 (x) grad U 2 (x) dm(x)
n
 2
1  |d | 
grad U 2 (x) dm(x)
n
1 1
I (|d1 |) + 2I (|d1 |) 2 I (|d2 |) 2 + I (|d2 |)
< + .

Thus, W is a linear space.

1 1
From Fubinis Theorem and from I (|d1 |, |d2 |) I (|d1 |) 2 I (|d2 |) 2 <
+, we get that I (d1 , d2 ) is a complex number for every d1 and d2 in
W .
If d is a locally finite Borel measure in with real values and I (|d|) <
+, then its non-negative and non-positive variations, d+ and d , satisfy
I (d+ ) < + and I (d ) < +. Proposition 7.3 implies that

I (d) = I (d+ ) 2I (d+ , d ) + I (d )

 2
1  d+ 
= grad U (x) dm(x)
n

2 +
+ grad Ud (x) grad Ud (x) dm(x)
n
206 CHAPTER 7. ENERGY
 2
1  d 
grad U (x) dm(x)
n
 2
1  d 
= grad U (x) dm(x) .
n

If d1 and d2 are two locally finite Borel measures in W with real values,
then, applying the last equality to d1 + d2 , we get

1
I (d1 , d2 ) = grad Ud1 (x) grad Ud2 (x) dm(x) .
n

By linearity, this extends to hold for all measures in W (d2 being replaced
by d2 .)
Thus, the bilinear form I (d1 , d2 ) is a complex number for all d1 and
d2 in W and, using the representation of the last formula, we get that

I (d) 0 for all non-negative d W .

Assume, now, that, for some d W ,

I (d) = 0 .

Then grad Ud (x) = 0 for almost every x and, since Ud is, by Theorem
2.18, absolutely continuous on almost every line parallel to any of the principal
axes, we easily get that Ud is equal to some constant c almost everywhere in
. Since this function is superharmonic, it is identically equal to c in .
Finally,
n d = Ud = 0
in in the sense of distributions and, hence, d is the zero measure.
Therefore, the bilinear form is an inner product.

7.4 Representation of energy: kernels of first

type
If K is any non-negative kernel of first type, then

lim K (r) = 0 .
r+

Lemma 7.2 Suppose that K is a kernel of first type with the property that, for
some R > 0,
K(x) = 0
for all x with |x| R.
Then K L1 (Rn ) and
b
K() > 0
7.4. REPRESENTATION OF ENERGY: KERNELS OF FIRST TYPE 207

for all Rn . There is, also, a constant C > 0 so that

b C
K()
1 + ||2

for all Rn .

Proof:
The kernel is non-negative and, by the first property of kernels of first type,
we have
R
K(x) dm(x) = n1 K (r)rn1 dr < + .
Rn 0

Now,

b
K() = e2ix K(x) dm(x)
Rn
+
0
= K (r) e2irx d(x0 ) rn1 dr
0 S n1
+
= K (r)J(r||)rn1 dr ,
0

where we use the notation x0 = xr = (x01 , . . . , x0n ) with r = |x| and

0
J(r) = e2irx1 d(x0 ) , r0.
S n1

0
Finally, writing x0 = (x01 , x00 ) with x00 = (1 x12 ) 2 y, y S n2 , and using
1

0
d(x0 ) = (1 x12 ) 2 d(y) dx01 , we get
n2

1
0 0
e2irx1 (1 x12 )
n2
J(r) = 2 d(y) dx01
1 S n2

2
= 2n2 cos(2r cos ) sinn2 d
0

and J has the following properties:

1. J(0) = n1 and J 0 (0) = 0.

2. J(r) < n1 for all r > 0.

( )
3. J(r) = 41 2 J 00 (r) + n1 0
r J (r) for all r 0.

4. limr+ J(r) = 0.

The first three properties are trivial to prove. The fourth is an application
1 0 0
of Theorem 0.13 to J(r) = n2 1 e2irx1 (1 x12 ) 2 dx01 .
n2
208 CHAPTER 7. ENERGY

Therefore,
( r )(
b 1 +
n 1 0 ) n1
K() = 2 n K J 00 (r) + J (r) r dr
4 || 0 || r
+ ( r )d(
1 )
= 2 n K J 0 (r)rn1 dr
4 || 0 || dr
+ ( )
1 0 r
= K J 0 (r)rn1 dr
4 2 ||n+1 0 ||
+
1 ( ) ( ( r )( r )n1 )
= n1 J(r) d K 0
4 || 0
2 2 || ||
where the integrations by parts are easy to justify.
But, if we set s = h (r) and K (r) = H(s), then, in case n 3, the last
integral becomes
+
( ) ( 0 n2 )
(n 2) n1 J(h1
(s)) d H (|| s)
0
1( ) ( 0 n2 )
(n 2) n1 J(h1
(s)) d H (|| s)
0
( ) 0 n2
(n 2) min n1 J(r) H (|| )
1r<+

and, in case n = 2, it becomes

+
( ) ( 0 )
1 J(h1
(s)) d H (s + log ||)

0 ( ) ( 0 )
1 J(h1
(s)) d H (s + log ||)

( )
min 1 J(r) H 0 (log ||) .
1r<+

Considering R large enough so that H 0 (Rn2 ) > 0 or H 0 (log R) > 0, respec-

tively, then, for all with || R, we have

b C
K() .
||2
+ ( ) ( 0 n2 )
In any case, both integrals (n 2) 0 n1 J(h1
(s)) d H (|| s)
+ ( 1
) ( 0 )
and 1 J(h (s)) d H (s + log ||) are positive for all 6= 0. Hence,

b
K() b
> 0 for all 6= 0. Since K is non-negative, K(0) = Rn K(x) dm(x) > 0.
The proof is concluded by the continuity of K b in B(0; R).

Lemma 7.3 Suppose that K is a non-negative kernel of first type with the prop-
erty that, for some R > 0,
K (r) = 0
7.4. REPRESENTATION OF ENERGY: KERNELS OF FIRST TYPE 209

for all r R.
If d1 and d2 are any two compactly supported non-negative Borel measures
in Rn with IK (d1 ) < + and IK (d2 ) < +, then

IK (d1 , d2 ) = b d
K() d2 () dm() .
d1 ()d
Rn

d1
Therefore, IK (d1 , d2 ) < + and UK is finite almost everywhere with
respect to d2 and vice-versa.

Proof:
Suppose that Aj = supp(dj ). By Proposition 6.4, we can find restrictions
d1,1 and d2,2 of d1 and d2 in compact sets A1,1 A1 and A2,2 A2 ,
respectively, so that, for both j = 1, 2,

kdj dj,j k 0
dj,
as j 0 and each UK j is a continuous potential.
We may, also, assume that Aj,j Aj as j 0, as we may easily see by
looking in the proof of Theorem 6.1. Hence,

IK (d1,1 , d2,2 ) IK (d1 , d2 )

as j 0.
Now introduce the functions

k (x) = k 2 ek|x|
n 2

for all k > 0 and all x Rn . Then,

k (x) dm(x) = 1
Rn

and
||2
ck () = e
k

dj,
for all Rn . Now, both functions k and UK j = K dj,j are in L1 (Rn )
and, hence,
dj,
UK j k L1 (Rn ) .
( dj, )
Furthermore, since UK j b L (Rn ) and ck L1 (Rn ), we get

( dj,j ) ( dj, )
UK k b = UK j b ck L1 (Rn ) .

Hence, by the Inversion Formula,

( dj, )
dj,
UK j k (x) = ck () dm()
e2ix UK j b()
Rn
210 CHAPTER 7. ENERGY

and

d
UK 1,1 k (x) d2,2 (x)

Rn
( d )
= e2ix UK 1,1 b() ck () dm() d2, (x)
2

Rn Rn
( d1,1 )
= UK ck ()d
b() d 2,2 () dm()
Rn
||2
= b d
e k K() d d
1,1 ()d2,2 () dm() .
Rn

In the same way, for each j = 1, 2,

||2  2
dj,
UK j k (x) dj,j (x) = e k b d
K() d
j,j ()
 dm() .
Rn Rn
dj,j
By the continuity of UK and the Maximum Principle for K-potentials
dj,
we have that UK j is, also, bounded and, hence,
dj,j dj,j
UK k UK

uniformly on compact sets. Therefore, the left side of the last equality tends to
dj,j
U
Rn K
(x) dj,j (x) as k +. By the Monotone Convergence Theorem,

 2
b  d 
K() dj,j () dm() =
dj,
UK j (x) dj,j (x) IK (dj ) .
Rn Rn

By the Cauchy-Schwartz inequality,

 
b d
K() d d 
1,1 ()d2,2 () dm() < +
Rn

and we conclude that

IK (d1,1 , d2,2 ) = b d
K() d d
1,1 ()d2,2 () dm() .
Rn

d
Since d d
j,j () dj () for all as j 0 we may, now, take a decreasing
sequence of j s so that
d
d d
j,j () dj ()

b
weakly in L2 (Kdm). Letting, first, 1 0 and, then, 2 0 through the
appropriate sequences, we find from the last equality,

IK (d1 , d2 ) = b d
K() d1 ()d
d2 () dm() .
Rn

The last statement of the theorem is, just, an application of Cauchy-Schwartz

inequality and the Theorem of Fubini.
7.4. REPRESENTATION OF ENERGY: KERNELS OF FIRST TYPE 211

Proposition 7.4 Suppose that K is a kernel of first type with the property that,
for some R > 0,
K (r) = 0
for all r R.
If d is a locally finite complex Borel measure, let dr be its restriction in
the ball B(0; r).
If IK (|d|) < +, then the definition

c =
d dr
lim d
r+

b
is justified as a limit in the space L2 (Kdm).
If IK (|d1 |) < + and IK (|d2 |) < +, then

IK (d1 , d2 ) = d1 ()d
d d2 ()K()
b dm() .
Rn

Proof:
1. Assume that d is a non-negative Borel measure. By Lemma 7.3, for every
r, r0 with r < r0 , we have

 
IK (dr dr ) =
0 d
d d 2 b
r 0 () dr () K() dm() .
Rn

On the other hand, we can easily prove that

IK (dr0 dr ) IK (dr0 ) IK (dr ) .

From the last two relations, since IK (dr ) IK (d), we see that, if IK (d) <
c = limr+ d
+, then d dr exists as a limit in L2 (Kdm).
b
Now, if d1 and d2 are non-negative Borel measures with IK (d1 ) < +
and IK (d2 ) < +, we apply Lemma 7.3 to their restrictions in B(0; r) and let
r + to get

IK (d1 , d2 ) = d1 ()d
d d2 ()K()
b dm() .
Rn

2. Now, let d be a locally finite complex Borel measure with IK (|d|) < +.
Then the non-negative and non-positive parts of the real and imaginary parts
of d all have finite K-energy. From part 1, we get that dc = limr+ d dr
2 b
exists as a limit in L (Kdm) and that

IK (d1 , d2 ) = d1 ()d
d d2 ()K()
b dm() ,
Rn

for all locally finite complex Borel measures d1 and d2 with IK (|d1 |) < +
and IK (|d2 |) < +.
212 CHAPTER 7. ENERGY

7.5 Measures of finite energy: kernels of first

type
Definition 7.4 Suppose that K is a non-negative kernel of first type and let

WK = {d : d is a locally finite Borel measure in Rn and IK (|d|) < +} .

WK is called the space of measures of finite K-energy.

Proposition 7.5 Suppose that K is a non-negative kernel of first type and let
d1 and d2 be any two non-negative Borel measures with IK (d1 ) < + and
IK (d2 ) < +. Then,
( ) 21 ( ) 12
IK (d1 , d2 ) IK (d1 ) IK (d2 )
d1
and UK is finite almost everywhere with respect to d2 and vice-versa.

Proof:
If the measures are supported in a compact set and K vanishes for all large
enough values of its argument, then the result is an application of the last
Lemma.
Assume, now, that d1 and d2 are supported in a compact set but that
K(x) > 0 for all x Rn . (Observe that, if K(x0 ) = 0 for some x0 , then
K(x) = 0 for all x with |x| |x0 |.)
Consider, for every > 0, the kernel
{
K(x, y) , if K(x, y)
K (x, y) =
0, if K(x, y) .

Since limr+ K (r) = 0, we see that K is a non-negative kernel of first

type and that K (x) = 0 for all large enough |x| and, hence, Lemma 7.3 applies
to K .
Therefore,
( ) 21 ( )1
IK (d1 , d2 ) IK (d1 ) IK (d2 ) 2 .

Since K K as 0, by the Monotone Convergence Theorem or by

Fatous Lemma, we get
( )1 ( )1
IK (d1 , d2 ) IK (d1 ) 2 IK (d2 ) 2 .

If the measures are not supported in a compact set, then apply the result to
the restrictions of the measures on the balls B(0; R) and let R +.

Theorem 7.2 Suppose that K is any non-negative kernel of first type.

Then, WK is a linear space. Also, for any d1 and d2 in WK , IK (d1 , d2 )
is well-defined as a complex number and, under this bilinear form, WK becomes
an inner product space
7.5. MEASURES OF FINITE ENERGY: KERNELS OF FIRST TYPE 213

d
Also, if d WK , then UK is finite almost everywhere with respect to all
d WK .

Proof:
If d1 WK and d2 WK , then, by the last Proposition,

IK (|d1 + d2 |) IK (|d1 | + |d2 |)

= IK (|d1 |) + 2IK (|d1 |, |d2 |) + IK (|d2 |)
( )1 ( )1
IK (|d1 |) + 2 IK (|d1 |) 2 IK (|d2 |) 2 + IK (|d2 |)
< + .

Therefore, WK is a linear space.

By Fubinis Theorem and from IK (|d1 |, |d2 |) < +, we, also, have that
d1
UK is finite almost everywhere with respect to d2 and vice-versa and that
IK (d1 , d2 ) is well-defined as a complex number.
It is obvious that IK ( , ) is a bilinear form and we, only, have to prove that
it is positive definite.
Now, take an arbitrary locally finite complex Borel measure d in Rn with
real values and with IK (|d|) < + and consider its non-negative and non-
positive variations d+ and d .
Then, we, also, have IK (d+ ) < + and IK (d ) < + and, hence,

IK (d) = IK (d+ ) 2IK (d+ , d ) + IK (d )

( )1 ( )1
IK (d+ ) 2 IK (d+ ) 2 IK (d ) 2 + IK (d )
0.

If d is a locally finite complex Borel measure in Rn with IK (|d|) < +

and d1 , d2 are its real and imaginary parts, then

IK (d) = IK (d1 ) + IK (d2 ) 0 .

By the standard argument, we can prove, now, that

( )2
IK (d1 , d2 ) IK (d1 )IK (d2 )

for all d1 and d2 in WK .

Now, let d be a locally finite complex Borel measure in Rn with IK (|d|) <
+ and assume that
IK (d) = 0 .
If K (r) > 0 for all r > 0 and K = H h , then there exists some t0 with
H 0 (t0 ) > 0. We define
{ 1( )
0
H1 (t) = 2 H(t) H (t0 )(t t0 ) H(t0 ) , if t > t0
0, if t t0
214 CHAPTER 7. ENERGY

and
H2 = H H1 .
We consider, next,
K1 = H1 h , K2 = H2 h .
The corresponding kernels K1 and K2 are non-negative and of first type and
K1 has the property that
K1 (r) = 0
for all r R = h1 (t0 ).
Now, for each i = 1, 2,
IKi (|d|) IK (|d|) < + .
Also,
IK1 (d) + IK2 (d) = IK (d)
and, since both terms in the left side are non-negative, we get
IK1 (d) = 0 .
Hence, we may assume that our kernel K has the property that, for some R,
K (r) = 0
for all r R.
Consider, now, an arbitrary D(Rn ) and the function

b
f = .
Kb
Lemma 7.2 together with Corollary 0.1 easily give that f belongs to S(Rn ).
Theorem 0.18 implies that there is a g S(Rn ) so that gb = f and, hence,
gK = .
Since K L1 (Rn ) and g L1 (Rn ) L (Rn ), it is immediate that
IK (|g|) < + .
Thus,
   
   
 (x) d(x) =  K g(x) d(x)
Rn Rn
( )1 ( )1
= |IK (d, g)| IK (d) 2 IK (g) 2 = 0 .
Therefore, d is the zero measure.

In the case of R2 and of kernels of first type and of variable sign the measures
considered are supported in compact sets.
7.5. MEASURES OF FINITE ENERGY: KERNELS OF FIRST TYPE 215

Definition 7.5 Assume that K is a kernel of first type and of variable sign.
We define
WK
0
= {d : d is a compactly supported complex Borel measure with
IK (|d|) < + and d(R2 ) = 0} .
Now, we may extend the results about non-negative kernels. For example,
here is the central result.
Theorem 7.3 Assume that K is a kernel of first type and of variable sign.
Then, WK 0
is a linear space. Also, for any d1 and d2 in WK0
, IK (d1 , d2 )
is well-defined as a complex number and, under this bilinear form, WK 0
becomes
an inner product space.
d
Moreover, if d WK 0
, then UK is finite almost everywhere with respect to
all d WK .
0

Proof:
If d1 and d2 are in WK 0
, we consider a large enough R so that the two
measures are supported in the disc B(0; R) and then consider the modified kernel
{
K(x, y) K (2R) , if |x y| 2R
KR (x, y) =
0, if |x y| > 2R .
The new kernel is non-negative and of first type with the property that it
vanishes for all large enough values of its argument.
We, also, have that, for all x B(0; R),

d1 d1
UK (x) = K(x, y) d1 (y) = KR (x, y) d1 (y) = UKR
(x)
B(0;R) B(0;R)

and, thus,

d1
IK (d1 , d2 ) = UK (x) d2 (x)
B(0;R)

d1
= UKR
(x) d2 (x)
B(0;R)
= IKR (d1 , d2 ) .
Applying, now, Theorem 7.2 to the kernel KR , we see that IK (d1 , d2 ) is
well-defined as a complex number, that
IK (d1 ) 0
and that IK (d1 ) = 0 if and only if d1 is the zero measure.
In fact, either by Lemma 7.3 or by Proposition 7.4, we have that

 2
IK (d) = c
d() KdR () dm(x) ,
R2

for all d WK
0
which are supported in B(0; R).

There is another special result for the logarithmic kernel in R2 .

216 CHAPTER 7. ENERGY

1
Theorem 7.4 If K(x, y) = log |xy| = h (|x y|), then the space of all com-
2
plex Borel measures d in R which are supported in the unit disc B(0; 1) and
satisfy Ih (|d|) < + becomes an inner product space under the bilinear form
Ih (d1 , d2 ).

Proof:
Consider the measure d on S 1 and its h-potential
{
1 1
log |x| , if |x| > 1
Uhd (x) = log d(y) =
S 1 |x y| 0 , if |x| 1 .

Now, for every d1 and d2 with Ih (|d1 |) < + and Ih (|d2 |) < +, we
define the measures
di (R2 )
di = di d .
2
Then, both di belong to Wh0 and, hence, Theorem 7.3 applies to them.
Now, observe that

d1 (R2 )
Ih (d1 , d2 ) = Ih (d1 , d2 ) + Ih (d, d2 )
2
d2 (R2 ) d1 (R2 ) d2 (R2 )
+ Ih (d1 , d) + Ih (d)
2 2 2
2
d1 (R )
= Ih (d1 , d2 ) + Uhd (x) d2 (x)
2 R2

d2 (R2 )
+ Uhd (x) d1 (x)
2 R 2

d1 (R2 ) d2 (R2 )
+ Uhd (x) d(x)
2 2 R2
= Ih (d1 , d2 ) .
Chapter 8

Capacity

8.1 Definitions
Let K be a kernel of either the first or the second type.
Remark:Whenever, in this chapter, K = G is a kernel of second type, we
shall understand that the set has a Greens function in all its components
and that all sets to be considered are subsets of . Also, all measures are Borel
measures in .
If the kernel is of first type, then all sets are subsets of Rn and all measures
are Borel measures in Rn .
Definition 8.1 Let E be a compact set. By E K we denote the family of all
non-negative Borel measures d which are supported in E and satisfy
d
UK 1
d
everywhere in E. Therefore, by the Maximum Principle of potentials, UK 1
n
everywhere in R , in case K is of first type, or in , in case K = G is of
second type.

Definition 8.2 We define the K-capacity of the compact E by

CK (E) = sup d(E) .

dE
K

We, also, define the inner K-capacity of the set A by

i
CK (A) = sup{CK (E) : E is a compact subset of A} .

Finally, the outer K-capacity of a set A is defined by

o
CK i
(A) = inf{CK (O) : O is open with A O} .
i
It is almost obvious that CK (A) CK
o
(A) for all A.

217
218 CHAPTER 8. CAPACITY

i o
Definition 8.3 A set A is called K-capacitable if CK (A) = CK (A).
The proof of the next result is trivial.
Proposition 8.1 1. If E1 and E2 are compact sets with E1 E2 , then
K K .
E 1 E2

2. If E1 and E2 are compact sets with E1 E2 , then CK (E1 ) CK (E2 ).

3. If A1 A2 , then CK
i
(A1 ) CK
i o
(A2 ) and CK (A1 ) CK
o
(A2 ).
The proof of the next result is based on Proposition 8.1(3) and is, also,
trivial.
Proposition 8.2 All open sets O are K-capacitable:
i o
CK (O) = CK (O) .

Proposition 8.3 Every compact set E is K-capacitable and

i o
CK (E) = CK (E) = CK (E) .

Proof:
i
The equality CK (E) = CK (E) is clear and comes from Proposition 8.1(2).
From the definitions, the inequality CK (E) CKo
(E) is, also, clear and for
the rest of the proof we may assume that CK (E) < +.
Consider the sets
E = {x : d(x, E) }
and
O = {x : d(x, E) < } ,
where is small enough so that E in case K = G .
The sets E are compact, the sets O are open and

E O E

and
E E
as 0.
Therefore, CK (E ) is decreasing as 0 and let

CK (E )

as 0, where is such that

CK (E) + .

Assume that CK (E) < and consider any with

CK (E) < < .

8.1. DEFINITIONS 219

Then, we can find non-negative Borel measures d supported in E so that

d (E )

and
d
UK 1
everywhere in R , in case K is of first type, or in , in case K = G .
n

Deviding d by an appropriate positive number we may assume, if neces-

sary, that
d (E ) + 1 < + .
Hence, there is some sequence of s tending to 0 so that

d d

weakly on compact sets, where d is a non-negative Borel measure.

It is clear that d is supported in E and that

d(E) = lim d (E ) > CK (E)

as 0 through this sequence.

Because of the lower-semicontinuity of potentials in measure,

d d
UK lim inf UK 1
0

everywhere.
This is a contradiction to the definition of CK (E) and, thus,

CK (E ) CK (E)

as 0.
Now,
i
CK (E) = CK (E) CK
i
(O ) CK
i
(E ) = CK (E ) ,

implying that
o
CK (E) = CK (E) .

We may, now, write,

Definition 8.4 If A is K-capacitable, we define its K-capacity by
i o
CK (A) = CK (A) = CK (A) .

It can be proved that all Borel sets are K-capacitable and, even more, that
all analytic sets are K-capacitable. We shall not work in this direction and
we shall have only in mind that all compact and all open sets are K-capacitable.
220 CHAPTER 8. CAPACITY

i
Lemma 8.1 Assume that CK (A) = 0 and d is any non-negative Borel mea-
d
sure supported in a compact set with UK being bounded from above in the set
A. Then d(E) = 0 for every compact E A.
Proof:
It is clear that we may suppose that A is bounded.
Assume that d(E) > 0 for some compact E A and consider the restriction
dE of d in E.
dE
Then UK is, also, bounded from above in the set A E and, hence,
dE
UK M
everywhere, for some M > 0. Now, define
1
d = dE .
M
Then, d E
K and
d(E) > 0 .
Therefore,
i
CK (A) CK (E) d(E) > 0 .
i
Lemma 8.2 If all Am are Borel sets and CK (Am ) = 0 for all m, then
i
CK (+
m=1 Am ) = 0 .

Proof:
Take an arbitrary non-negative Borel measure supported in a compact subset
of +
m=1 Am with
d
UK 1
everywhere.
By Lemma 8.1,
d(Am ) = 0
for every m and, thus,
d(+
m=1 Am ) = 0 .
By the definition,
i
CK (+
m=1 Am ) = 0 .

Proposition 8.4 Suppose that A = + k=1 Ak .

1. If K is a non-negative kernel, then

+
o
CK (A) o
CK (Ak ) .
k=1

2. If K is of variable sign and D = diam(A), then

CK o
(A)
+

o (A)K (D) o
CK (Ak ) ,
1 + CK
k=1

where K (D) = min(0, K (D)).

8.2. EQUILIBRIUM MEASURES 221

Proof:
o
1. It is enough to assume that CK (Ak ) < + for all k and, then, we take open
Uk R so that Ak Uk and CK (Uk ) CK
n o
(Ak ) + 2k .
For an arbitrary compact E k=1 Uk we take N so that E N
+
k=1 Uk and,
then, write E = N E
k=1 k , where each E k is a compact subset of Uk .
Now, we take d EK and observe that dEk K for every k. Therefore,
Ek

N
N
+
+
d(E) d(Ek ) CK (Ek ) CK (Uk ) o
CK (Ak ) +  .
k=1 k=1 k=1 k=1

This implies

+
o
CK (A) CK (+
k=1 Uk )
o
CK (Ak ) + 
k=1

and concludes the proof.

2. We repeat the same argument observing that, if d E
K , then

dEk
UK d
(x) UK (x) K (D)d(E) 1 + K (D) d(E)

for every x Ek .
The rest of the proof remains unchanged.

The proof of the next result is an immediate consequence of Proposition 8.4.

Theorem 8.1 Suppose that CK (Ak ) = 0 for all k.

1. If K is a non-negative kernel then, CK (+
k=1 Ak ) = 0.
2. If K is of variable sign and +
k=1 Ak is bounded, then CK (+
k=1 Ak ) = 0.

Definition 8.5 We say that some property holds quasi-almost everywhere

or q-a.e. in a set A, if it holds everywhere in A except in a subset of A of zero
K-capacity.

8.2 Equilibrium measures

Theorem 8.2 Suppose E is a compact set. Then the extremal problem

K (E) = inf IK (d)

over all probability Borel measures (i.e. non-negative Borel measures with total
mass equal to 1) supported in E has a solution d0 .
This satisfies
d0
1. UK K (E) everywhere,
d0
2. UK = K (E) quasi-almost everywhere in E.
222 CHAPTER 8. CAPACITY

Proof:
Assume, first, that
K (E) < + .
Consider a sequence of probability Borel measures dm supported in E so
that
IK (dm ) K (E) .
There is some subsequence converging weakly in E to some probability Borel
measure d0 supported in E.
By the lower-semicontinuity of energy in measure,
K (E) IK (d0 ) lim inf IK (dm ) = K (E) .
m+

Thus, d0 is a solution of the extremal problem.

d0
Claim: UK K (E) q-a.e. in E.
To prove this, consider  > 0 and assume that the compact set
d0
E  = {x E : UK (x) K (E) }
has positive K-capacity.
By definition, there is some non-negative Borel measure d supported in E 
with
d (E  ) > 0
and
d
UK 1
everywhere.
Define, now, the probability measure
1
d = d
d (E  )
supported in E  , satisfying
1
d
UK M =
d (E  )
everywhere.
We apply, now, a variational argument, considering 0 < < 1 and defining
d = (1 )d0 + d .
Then d is a probability measure supported in E and, hence,
K (E) IK (d )

d0
= (1 )2 IK (d0 ) + 2(1 ) UK (x) d(x) + 2 IK (d)
E
d0
= K (E) 2IK (d0 ) + 2 UK (x) d(x)
(
E
)
d0
+ 2 IK (d) + IK (d0 ) 2 UK (x) d(x) .
E
8.2. EQUILIBRIUM MEASURES 223

Thus,
( )
d0
0 2 + 2 IK (d) + IK (d0 ) 2 UK (x) d(x) .
E

This is, clearly, absurd for small > 0 and, thus,

CK (E  ) = 0 .

Since
d0 1
{x E : UK (x) < (K)} = +
m=1 E
m ,

Theorem 8.1 implies the claim.

d0
Claim: UK K (E) everywhere.
Consider an arbitrary x E and assume that
d0
UK (x) > K (E) .

Take an arbitrary 1 so that

d0
UK (x) > 1 > K (E)

and, by lower-semicontinuity,
d0
UK > 1

everywhere in some open neighborhood B(x; rx ) of x.

d0
The Borel set {x E : UK < K (E)} has zero K-capacity and, by Lemma
8.1, ( )
d0
d0 {x E : UK < K (E)} = 0 .
Then,

IK (d0 ) 1 d0 (E B(x; rx )) + K (E)d0 (E \ B(x; rx )) .

By IK (d0 ) = K (E), it is implied that

d0 (E B(x; rx )) = 0 .

Consider the open set

{ d0 }
O = B(x; rx ) : x E and UK (x) > K (E) .

If F is any compact subset of O, then we can cover F by finitely many balls

d0
of the form B(x; rx ) with x E and UK (x) > K (E) and, hence, d0 (F ) = 0.
Since F is arbitrary,
d0 (O) = 0 .
Therefore, d0 is supported in the compact set E \ O and
d0
UK K (E)
224 CHAPTER 8. CAPACITY

in this set. By the Maximum Principle of potentials, the claim is clear and,
together with the first claim, the proof of the theorem in case K (E) < + is
complete.
Assume, now, that K (E) = +.
It is, then, obvious that EK contains only the zero measure and, hence,
CK (E) = 0.
Therefore, every probability Borel measure supported in E (for example,
any dx with x E) is a solution of the extremal problem.

Theorem 8.3 If E is a compact set and K (E) < +, then the solution of
the extremal problem of Theorem 8.2 is unique.

Proof:
If d0 and d00 are two solutions of the extremal problem, then UK
d0
= K (E)
d00 d0
and UK = K (E) quasi-almost everywhere in E and UK K (E) and
d00
UK K (E) everywhere.
By Lemma 8.2,
( )
d00 {x E : UK
d0
< K (E)} = 0 ,

implying

IK (d0 d00 ) = IK (d0 ) 2 d0
UK (x) d00 (x) + IK (d00 )
E
= K (E) 2K (E) + K (E)
= 0.

By Theorems 7.1, 7.2 and 7.3, we get

d0 = d00 .

Definition 8.6 If E is any compact set, then the quantity

K (E) = inf IK (d)

over all probability Borel measures which are supported in E, is called the K-
energy of E.

Definition 8.7 If E is a compact set with K (E) < +, then the unique
probability Borel measure d0 with

IK (d0 ) = K (E)

is called the K-equilibrium measure of E.

Theorem 8.4 If E is compact, then K (E) = + if and only if CK (E) = 0.

8.2. EQUILIBRIUM MEASURES 225

Proof:
Assume that CK (E) > 0 and take a non-negative Borel measure d sup-
ported in E with d(E) > 0 and
d
UK 1

everywhere.
1
Then d = d(E) d is a probability Borel measure supported in E with

1
IK (d) = d
UK (x) d(x) < + .
E d(E)
Thus, K (E) < +.
Assume, conversely, that K (E) < + and d0 is the K-equilibrium mea-
sure of E.
If 0 < K (E) < +, then
1
d = d0
K (E)

is in E
K and
1
d(E) = > 0.
K (E)
Therefore, CK (E) > 0.
If K (E) 0, then d0 E
K and

CK (E) d0 (E) = 1 .

Proposition 8.5 If K is a non-negative kernel, then

K (E) > 0

for every compact set E.

This is, also, true if K is the logarithmic kernel in R2 and E B(0; 1).
Proof:
Let K (E) < + and consider the K-equilibrium measure d0 of E.
Then, by Theorems 7.1, 7.2 and 7.4,

K (E) = IK (d0 ) > 0 .

Consider, now, for any compact set E, the three extremal problems
(I) K (E) = inf IK (d) over all probability Borel measures d supported in E,
(II) A = inf d(Rn ) over all non-negative Borel measures d supported in a
d
compact set with UK 1 q-a.e. in E,
(III) CK (E) = sup d(E) over all non-negative Borel measures d supported in
d
E with UK 1 everywhere in E.
226 CHAPTER 8. CAPACITY

Theorem 8.5 Suppose that E is a compact set with K (E) > 0. Then, the
three extremal problems (I), (II) and (III) are equivalent in the sense

1
= A = CK (E) .
K (E)

In case 0 < K (E) < +, the problem (I) has a unique solution, the K-
equilibrium measure d0 of E, and (III) has, also, a unique solution, the measure
1
K (E) d0 . The same measure is, also, a solution of problem (II) whose solution,
though, may not be unique.
If K (E) 0, then the problem (I) has, again, the K-equilibrium measure
of E as its unique solution, but

A = CK (E) = +

and the problems (II) and (III) have no solution.

Proof:
1. Suppose that K (E) = +.
By Theorem 8.4, CK (E) = 0. Considering the zero measure, we find that
A = 0.
2. Now, let 0 < K (E) < +.
Consider the K-equilibrium measure d0 of E, which is the unique solution
of problem (I), and define

1
d1 = d0 .
K (E)

Then, d1 E
K and, thus,

1
CK (E) d1 (E) = .
K (E)

On the other hand, let d be any measure qualifying for problem (III).
( d1 )
By Lemma 8.1, d {x E : UK (x) < 1} = 0 and, hence,

1 d1
= d1 (E) d
UK (x) d1 (x) = UK (x) d(x) d(E) .
K (E) E E

Therefore,
1
CK (E)
K (E)
and, hence,
1
= CK (E) .
K (E)
Furthermore, d1 is a solution of problem (III).
8.2. EQUILIBRIUM MEASURES 227

Suppose, now, that d is another solution of problem (III). Then,

d1
IK (d1 d) = IK (d1 ) 2 UK (x) d(x) + IK (d)
E
CK (E) 2CK (E) + CK (E)
= 0.

By Theorems 7.1, 7.2 and 7.3, d = d1 .

The measure d1 qualifies for problem (II) and, thus,

A CK (E) .

Let d be any measure qualifying for problem (II). Then, by Lemma 8.1,

d1
CK (E) UKd
(x) d1 (x) = UK (x) d(x) d(Rn ) .
E Rn

Therefore,
CK (E) A ,
and the proof is complete in case K (E) > 0.
Now, assume that K (E) 0.
If we consider the K-equilibrium measure, d0 , of E, then, for every > 0,
the measure d0 belongs to E K and, thus,

CK (E) d0 (E) = .

Since is arbitrary,
CK (E) = +
and the problem (III) cannot have a solution.
On the other hand, suppose that there is some d qualifying for problem
(II).
Then, by Lemma 8.1,

d0
1 UKd
(x) d0 (x) = UK (x) d(x) K (E)d(Rn ) .
E Rn

Since 0 d(Rn ) < +, this is impossible and, hence, there is no measure

qualifying for problem (II). Thus,

A = + .

Example
Let E = S 1 in R2 and K (r) = log+ 3r .
By symmetry and uniqueness, the K-equilibrium measure of E must be
1
d0 = d
2
228 CHAPTER 8. CAPACITY

and
d0 1
UK (x) = K(x, y) d(y) .
2 S1
If |x| < 2, then

d0 1 1
UK (x) = log 3 + log d(y) = log 3 log+ |x| .
2 S1 |x y|
Therefore,
IK (d0 ) = log 3
and
1
CK (E) = .
log 3
Consider the measures
1 1
d1 = d0 = d
K (E) 2 log 3
and
1
d = d0 .
log 3
Then,
1 3
d
UK (x) = log+ = 1
log 3 |x|
when |x| = 1.
Hence, d1 and d are two solutions of problem (II).

Example
i
If A is a countable set, then CK (A) = 0 with respect to any kernel.
By Lemma 8.2, it is enough to consider the case A = {x0 } and, since all
non-negative measures supported in {x0 } are of the form
dx0 , 0,
we, then, have
dx0
UK = K(x, x0 ) .
This is bounded from above in {x0 } only if = 0 and, hence, the only
{x }
measure in K 0 is the zero measure.
This implies that
CK ({x0 }) = 0 .
Theorem 8.1 implies that CK (A) = 0 for every countable A and every non-
negative K and, also, for every bounded and countable A and every K of variable
sign.

Example
If the compact set E has positive Lebesgue measure, then CK (E) > 0 for all
kernels.
This is true because dm, restricted in E, defines a potential which is bounded
from above.
8.3. TRANSFINITE DIAMETER 229

8.3 Transfinite diameter

Theorem 8.6 If E is a compact set, define

1
m
Mm = sup inf K(x, xj )
x1 ,...,xm E xE m j=1

and
2
Dm = inf K(xi , xj ) .
x1 ,...,xm E m(m 1)
1i<jm

Then,
lim Mm = lim Dm = K (E) .
m+ m+

Proof:
We divide the proof into steps.
Step 1.
Since K is continuous in (Rn Rn ) \ {(x, x) : x Rn }, if K is of first type,
or in ( ) \ {(x, x) : x }, if K = G is of second type, it is easy to see
that there exist x1 , . . . , xm+1 E so that

2
Dm+1 = K(xi , xj ) .
(m + 1)m
1i<jm+1

Rewrite

2 (1 1 )
m+1 m m+1
Dm+1 = K(x1 , xj ) + K(xi , xj ) .
m + 1 m j=2 i=2
m j=i+1

Observe that the second

m+1sum in the last equality does not depend on x1 and,
1
hence, x1 minimizes m j=2 K(x, xj ).
Therefore,
1
m+1
K(x1 , xj ) Mm .
m j=2

Similarly,
1
K(xi , xj ) Mm
m
j6=i

for all i and, thus,

1
Dm+1 (Mm + + Mm ) = Mm
m+1
for all m.
Step 2.
230 CHAPTER 8. CAPACITY

If K (E) < +, consider the equilibrium measure d0 of E. Then, for all

x1 , . . . , xm E,
m
1 1
m
inf K(x, xj ) K(x, xj ) d0 (x) K (E)
xE m m j=1 E
j=1

and, hence,
Mm K (E)
for all m.
This is, also, true in case K (E) = +.
Step 3.
Consider x1 , . . . , xm E so that
1
Dm = K(xi , xj )
m(m 1)
i6=j

and define the probability Borel measure

m
1
dm = dxj .
j=1
m

Consider, also, the trunkated kernel

KN (x, y) = min(K(x, y), N ) .

Then,

1
m
KN (x, y) dm (x) dm (y) = KN (x, xj ) dm (x)
E E m j=1 E

1 ( 1 N)
m
K(xi , xj ) +
m j=1 m m
i6=j
1 N
K(xi , xj ) + .
m(m 1) m
i6=j

Choose any subsequence {mk } of ms so that

1. Dmk lim inf m+ Dm ,
2. dmk d weakly in E
as k +, where d is a Borel probability measure supported in E.
Then, by the the weak convergence dmk dmk d d in E E and
the continuity of KN in E E,

KN (x, y) d(x) d(y) = lim KN (x, y) dmk (x) dmk (y)
E E k+ E E
lim inf Dm .
m+
8.3. TRANSFINITE DIAMETER 231

Letting N +, we find

K (E) IK (d) lim inf Dm .

m+

After these three steps we have

K (E) lim inf Dm lim sup Dm lim sup Mm K (E)

m+ m+ m+

and

K (E) lim inf Dm lim inf Mm lim sup Mm K (E) .

m+ m+ m+

The proof is, clearly, complete.

Proposition 8.6 If E isa compact set with K (E) < +, x1 , . . . , xm E are

1 m 1
such that Dm = m(m1) i6=j K(xi , x j ) and dm = j=1 m dxj , then

dm d0

weakly in E, where d0 is the K-equilibrium measure of E.

Proof:
Since Dm K (E), in Step 3 of the proof of the Theorem 8.6 we proved
that, if d is any limit point of dm in the weak sense in E, then,

K (E) IK (d) lim inf Dm = K (E) .

m+

Thus,
K (E) = IK (d)
and, by the uniqueness of the K-equilibrium measure of E,

d = d0 .

Definition 8.8 If E is a compact set, then the number

dK (E) = eK (E)

is called the K-transfinite diameter of E.

Proposition 8.7 Let K be any kernel of first type, E and E 0 be two compact
sets and f : E E 0 be surjective and a contraction.
Then, dK (E 0 ) dK (E).

Proof:
We have to prove that K (E) K (E 0 ).
232 CHAPTER 8. CAPACITY

Consider points x01 , . . . , x0m E 0 so that the quantity Dm 0

, corresponding to
0
E , is
1
0
Dm = K(x0i , x0j ) .
m(m 1)
i6=j

Consider, also, points x1 , . . . , xm E so that

f (xj ) = x0j .

Since f is a contraction and K depends upon the distance of its two argu-
ments,
1 1
Dm K(xi , xj ) K(x0i , x0j ) = Dm
0
.
m(m 1) m(m 1)
i6=j i6=j

We finish the proof, by letting m +.

Example
If K is of first type, then rigid motions preserve K-transfinite diameter and
K-energy. It is, also, trivial to see through the definition that they preserve
K-capacity.

8.4 The Theorem of Evans

Theorem 8.7 (Evans) Suppose that E is a compact set with K (E) = +
or, equivalently, CK (E) = 0. Then, there exists a probability Borel measure d
supported in E so that
d
UK = +
identically in E.

Proof:
Consider points x1 , . . . , xm E so that

1
m
1
inf K(x, xj ) Mm
xE m j=1 2

and the probability Borel measure

1
m
dm = dxj .
m j=1

which is supported in E.
Since Mm +, there is a sequence {mk } so that

K(x, y) dmk (y) 2k
E
8.4. THE THEOREM OF EVANS 233

for all x E.
Now, construct the probability Borel measure

+
1
d = dmk
2k
k=1

which is, also, supported in E.

Then,
d

+
1 dmk
UK (x) = U (x) = +
2k K
k=1
for all x E.

Theorem 8.8 Suppose that d is a non-negative Borel measure supported in a

compact set and that
d
UK = +
identically in some set A.
Then, the bounded set A is K-capacitable and
CK (A) = 0 .
Proof:
Consider the open sets
d
Om = {x : UK (x) > m} ,
an arbitrary compact E Om and a non-negative Borel measure d E
K.
Then,

1 d 1 1
d(E) UK (x) d(x) = UKd
(x) d(x) d(Rn ) .
m E m Rn m
Hence, CK (E) 1
m d(Rn ) and, thus,
1
CK (Om ) d(Rn ) .
m
Therefore,
1
0 CK
i
(A) CK
o
(A) CK (Om ) d(Rn )
m
for all m, implying
i o
CK (A) = CK (A) = 0 .

Thus, Theorem 8.7 appears as a partial converse to Theorem 8.8: if the

compact set A satisfies CK (A) = 0, then it is the +-set of the K-potential of
a compactly supported non-negative Borel measure.
In section 9.7 this matter will be fully explored in the case K = h.
234 CHAPTER 8. CAPACITY

8.5 Kernels of variable sign

As Theorem 8.5 shows, if E is a compact set and K (E) > 0, then CK (E) =
K (E) . But, if K (E) 0, then CK (E) = +. Proposition 8.5 shows that
1

the situation is simple when the kernel is non-negative and, especially, when the
dimension is n 3.
To get around the complication arising in case of a kernel of variable sign
when the dimension is n = 2, we may use another definition of capacity.
Definition 8.9 If E is a compact set, the Robin-K-capacity of E is defined
by
CK,R (E) = eK (E) = dK (E) .
Observe that, if K (E) > 0, then
C 1
CK,R (E) = e K (E) ,

while, if K (E) 0, then

CK,R (E) 1 and CK (E) = + .

Therefore, the compact sets of zero K-capacity are the same as the compact
sets of zero Robin-K-capacity.
Also, if {Em } is a sequence of compact sets, then CK (Em ) 0 if and only
if CK,R (Em ) 0.
We may give the following definitions.
Definition 8.10 For every set A its inner Robin-K-capacity is defined by
i
CK,R (A) = sup{CK,R (E) : E is a compact subset of A}

and its outer Robin-K-capacity is defined by

o
CK,R i
(A) = inf{CK,R (O) : O is an open set with O A}

The inequality
i
CK,R (A) CK,R
o
(A)
is obvious, while the proof of the next result is trivial and is based on the
definitions.
Proposition 8.8 1. If E1 and E2 are compact sets with E1 E2 , then
K (E2 ) K (E1 ) and CK,R (E1 ) CK,R (E2 ).
2. If A1 A2 , then CK,R
i
(A1 ) CK,R
i o
(A2 ) and CK,R (A1 ) CK,R
o
(A2 ).
From Proposition 8.8(2) we get
Proposition 8.9 For every open set O,
i o
CK,R (O) = CK,R (O) .
8.5. KERNELS OF VARIABLE SIGN 235

Proposition 8.10 For every compact set E,

i o
CK,R (E) = CK,R (E) = CK,R (E) .

Proof:
From Proposition 8.8(1) we have
i
CK,R (E) = CK,R (E)

and from Proposition 8.8(2),

CK,R (E) CK,R

o
(E) .

Assume, now, that CK,R (E) < + and consider the compact sets

E = {x : d(x, E) }

and the open sets

O = {x : d(x, E) < } ,
where is small enough so that E in case K = G .
These satisfy
E O E .
As decreases K (E ) increases and let

K (E )

for some K (E).

Suppose that < K (E) and consider the K-equilibrium measures d0 of

E .
Then, there is some sequence k so that

d0k d

weakly in E, for some probability Borel measure d supported in E. From

Proposition 6.3,

K (E) IK (d) lim inf IK (d0k ) = .

k+

We, thus, get a contradiction and, hence,

K (E ) K (E) .

Therefore,
CK,R (E ) CK,R (E) ,
from which we, immediately, get that
i
CK,R (O ) CK,R (E)

and, finally,
o
CK,R (E) = CK,R (E) .
236 CHAPTER 8. CAPACITY

i
Definition 8.11 The set A is called Robin-K-capacitable, if CK,R (A) =
o
CK,R (A) and we, then, define its Robin-K-capacity by
i o
CK,R (A) = CK,R (A) = CK,R (A) .

The proof of the following is trivial.

i i
Proposition 8.11 If CK (A) < + or, equivalently, if CK,R (A) < 1, then the
set A is K-capacitable if and only if it is Robin-K-capacitable and, in this case,
C 1
CK,R (A) = e K (A) .

If CKi i
(A) = + or, equivalently, if CK,R (A) 1, then the set A is K-
capacitable and CK (A) = +.
Chapter 9

The Classical Kernels

In this chapter we study the particular case of the so-called classical kernels:
the Newtonian and the logarithmic kernels and, also, the Greens kernel related
to an open set with a Greens function in all its components.

9.1 Extension through sets of zero capacity

Theorem 9.1 Suppose that Rn is a bounded open set and E is a compact
subset of . Let K = h be the Newtonian or the logarithmic kernel or K = G .
Then K (E) = + or, equivalently CK (E) = 0, if and only if every function
which is harmonic and bounded in \ E can be extended in E so that it becomes
harmonic in .

Proof:
Suppose that K (E) = + and consider an open set 1 with C 1 -boundary
so that
E 1
and Hu1 , the solution of the problem of Dirichlet in 1 with the restriction of
u in 1 as boundary function.
The third example in section 8.2 implies that E = E. Therefore, by The-
orem 8.7 and Proposition 6.1, there is a superharmonic function v in 1 so
that
v 0
everywhere in 1 \ E and

lim v(x) = +
1 \E3xy

for all y E.
From Theorem 3.4, we have that E is of zero harmonic measure with respect
to 1 \ E as a subset of its boundary.

237
238 CHAPTER 9. THE CLASSICAL KERNELS

Now, Hu1 u is harmonic and bounded in 1 \ E. Also, by Proposition 3.6

and Theorem 3.7, ( 1 )
lim Hu u = 0
1 \E3xy

for all y 1 .
By Theorem 3.5,
u = Hu1
identically in 1 \ E.
Therefore, the extension of u which is harmonic in is
{
u, in \ E
Hu1 in E .

Now, assume that K (E) < + and let d0 be the K-equilibrium measure of
d0
E and UK the corresponding K-potential. By Theorem 8.2 and Proposition
6.1, this is a bounded harmonic function in \ E. If there is a function u
d0
harmonic in with u = UK everywhere in \ E, then the superharmonicity
d0
of UK in and the Minimum Principle for superharmonic functions imply
d0
that u = UK everywhere in .
d0
Thus, n d0 = UK = 0 in , contradicting that d0 is a probability
Borel measure.

9.2 Sets of zero harmonic measure

Definition 9.1 A compact set E Rn is said to be of zero harmonic mea-
sure, if, for every R with E B(0; R), E is of zero harmonic measure with
respect to the open set B(0; R) \ E.

Lemma 9.1 Suppose that Rn is a bounded open set and E is a compact

subset of . If E is of zero harmonic measure, then E = E is of zero
harmonic measure with respect to .

Proof:
Consider a large enough R so that B(0; R) and an arbitrary x0 .
Then
dB(0;R)\E
x0 (E) = 0 .
If
u B(0;R)\E
E
,
then, by the Minimum Principle for superharmonic functions, we get that u 0
in B(0; R) \ E and, hence,
u \E
E .

This implies that

dB(0;R)\E
x0 (E) d\E
x0 (E)
9.3. THE SET OF IRREGULAR BOUNDARY POINTS 239

and, thus,
d\E
x0 (E) = 0 .

Lemma 9.1 implies that a compact set E Rn is of zero harmonic measure

if and only if E is of zero harmonic measure with respect to every bounded
open set with E .

Theorem 9.2 Suppose that E is a compact set in Rn . Then E is of zero

harmonic measure if and only if Ch (E) = 0.

proof:
1. Let Ch (E) = 0 or, equivalently, h (E) = +.
We showed in the proof of Theorem 9.1 that, for every B(0; R) containing
E, E is of zero harmonic measure with respect to B(0; R) \ E and, hence, that
E is of zero harmonic measure.
2. Suppose, conversely, that E is of zero harmonic measure.
If h (E) < +, and d0 is the h-equilibrium measure of E, then, for all R
with E B(0; R), Uhd0 is a bounded harmonic function in B(0; R) \ E. This
is true due to Theorem 8.2 and the superharmonicity of Uhd0 .
By Theorem 3.5,

Uhd0 (x) M (R) = max Uhd0 (y)

yS(0;R)

for all x B(0; R) \ E.

If R + and n = 2, then M (R) and we get a contradiction, since
d0
Uh results to be identically in R2 \ E.
If R + and n 3, then M (R) 0 and we get that

Uhd0 = 0

in Rn \ E and, by the Minimum Principle for superharmonic functions, in Rn .

Therefore, d0 is the zero measure and we, again, arrive at a contradiction.

9.3 The set of irregular boundary points

Lemma 9.2 Suppose that is a bounded open set, x0 and > 0.
If = {x : G
x0 (x) > }, then is an open connected set containing
x0 and

x0 ( ) = 0 .
d

Moreover, the function

x0 (x) ,
G x ,

is the Greens function of with respect to x0 .

240 CHAPTER 9. THE CLASSICAL KERNELS

Proof:
1. That is open and contains x0 is clear.
Suppose that has a connected component O not containing x0 . By defi-
x0 hx0 is the least harmonic majorant of hx0 in .
nition, G
Consider the function
{
hx0 , if x O
u(x) =
Gx0 hx0 , if x \ O .

This is, obviously, a majorant of hx0 in and it is easy to see that it is

superharmonic in . Therefore, u G x0 hx0 everywhere in and this is false
in O.
2. Now, let

v

and consider the function

v1 = max(v, 0) .
Then,

1. v1 is subharmonic in ,

2. v1 is bounded from above in ,

3. lim 3xy v1 (x) = 0, if y \ ( ) = and

4. lim sup 3xy v1 (x) 1, if y .

Hence, the function

x0 (x) v1 (x) ,
V (x) = G x ,

is superharmonic in with

lim inf V (x) 0

3xy

for all y and, by the Minimum Principle for superharmonic functions,

V 0

in .
If, now, we define

V (x) = G
x0 (x) , x \ ,

then V is superharmonic and non-negative in and, hence, V hx0 is a super-

harmonic majorant of hx0 in .
Therefore,
V G x0
9.3. THE SET OF IRREGULAR BOUNDARY POINTS 241

in , implying that
v1 = 0

in and, hence,
v 0
in . Therefore,

x0 ( ) = H
d
(x0 ) = 0 .

3. Consider, now, a small > 0 so that B(x0 ; ) and the number M =

maxxS(x0 ;) G
x0 (x).
The set M is, as we showed in part 1, connected and, hence, it is contained
in B(x0 ; ). Therefore G x0 is bounded in \ B(x0 ; ). For the same reason,

Gx0 is bounded away from some neighborhood of x0 .
4. The function G x0 hx0 is a harmonic majorant of hx0 in and, hence,

x0 Gx0
G

in .
The function
x0 Gx0 +
G

x0 at all points of ,
is harmonic in and, by the continuity of G

( )
x0 (x) Gx0 (x) +
lim inf G 0

3xy

for all y .
From part 3, we have that this function is bounded in and, from part 2,
that is of zero harmonic measure as a subset of . We, thus, get
by Theorem 3.5 that,

x0 Gx0
G
in .

Theorem 9.3 If is a bounded open set, then the set of its irregular boundary
points is of zero h-capacity and, hence, of zero harmonic measure with respect
to .
Proof:
Let x0 and consider an arbitrary > 0 and the set

= {x : G
x0 (x) > } .

If A = , we shall prove that Ch (A ) = 0.

On the contrary, suppose h (A ) < + and let d0 be the h-equilibrium
measure of A .
Then,
242 CHAPTER 9. THE CLASSICAL KERNELS

1. Uhd0 h (A ) everywhere and

2. Uhd0 = h (A ) quasi-almost everywhere in A .

Now, consider some with 0 < < . By Lemma 9.2, Gx0 is the Greens
function of with respect to x0 .
By Proposition 5.5, the function


hx0 (x) hx (y) d
x0 (y) , x Rn \ {x0 } ,

is a subharmonic extension of Gx0 in R \{x0 } and, by upper-semicontinuity,

n

for all z A ,


hx0 (z) x0 (y) lim sup Gx0 (x) > 0 .
hz (y) d
xz

Therefore,
( )

hx0 (z) hz (y) d
x0 (y) d0 (z)
A


= Uhd0 (x0 ) Uhd0 (y) d
x0 (y) .

Every point of is a point of continuity of G x0 . Hence, by Theorem
5.4, every such point is a regular boundary point of . Therefore, by Theorem
3.7, the function


Uhd0 (x) Uhd0 (y) d
x (y) , x ,

which is harmonic and bounded in , has zero boundary limits at all points of
\ A . Since A is of zero harmonic measure with respect to the set ,

d0
Uh (x) Uhd0 (y) d
x (y) = 0

for all x .
Applying this to x = x0 , we get a contradiction.
Therefore,
Ch (A ) = 0 .

By Theorem 5.4, the set of irregular boundary points of is the union of

A m1 for all m N. Theorem 8.1 concludes the proof of the first statement.
The set of irregular boundary points of is a Borel set. By Theorem 9.2 and
Lemma 9.1, all its compact subsets are of zero harmonic measure with respect
to and, thus, the set itself is of zero harmonic measure with respect to .
9.4. THE SUPPORT OF THE EQUILIBRIUM MEASURE 243

9.4 The support of the equilibrium measure

Proposition 9.1 Suppose that E Rn is a compact set with h (E) < +.
Then, the h-equilibrium measure d0 of E is supported in the, so-called,
outer boundary of E; namely, the set O, where O is the unbounded connected
component of = Rn \ E.

Proof:
Since
Uhd0 = h (E)
quasi-almost everywhere in E, we get, by the last example in section 8.2, that

Uhd0 = h (E)

almost everywhere in E.
If O0 is a component of different from O, then Uhd0 is harmonic and
bounded in O0 and has boundary limits equal to h (E) in O0 except in a
subset of O0 of zero harmonic measure. Therefore,

Uhd0 = h (E)

everywhere in O0 .
We conclude that
Uhd0 = h (E)
almost everywhere, and, hence, everywhere in the interior of Rn \ O. This
implies that
n d0 = Uhd0 = 0
in the interior of Rn \ O.

Example.
Let E = B(x0 ; R).
Then, the h-equilibrium measure
( d0 )of B(x0 ; R) is supported in S(x0 ; R).
By the rotation invariance of h B(x0 ; R) = Ih (d0 ) and the uniqueness of the
h-equilibrium measure, we get that d0 must be rotation invariant and, hence,
1
d0 = dS
n1 Rn1

is the normalized surface area measure of S(x0 ; R).

Thus,

( ) d0
h B(x0 ; R) = Uh (x0 ) = h (R) d0 (y)
S(x0 ;R)
{
log R1 , if n=2
= h (R) = 1
Rn2 , if n3.
244 CHAPTER 9. THE CLASSICAL KERNELS

Example.
Let E = B(x0 ; r, R) with 0 < r < R.
Then, the h-equilibrium measure d0 of B(x0 ; R) is supported in S(x0 ; R)
and, exactly as before, we get that
( )
h B(x0 ; r, R) = h (R) .

Proposition 9.2 Suppose that E is a compact set with h (E) < + and let
the open Rn contain E.
Then, y E is a regular boundary point of \ E if and only if

lim Uhd0 (x) = h (E) ,

\E3xy

where d0 is the h-equilibrium measure of E.

Proof:
Since the notion of regularity has a local character, we may assume that
is equal to some large enough ball B(0; R).
Let V be the generalized solution of the Dirichlet problem in B(0; R) \ E
with boundary function
{
h (E) , if y E
Uhd0 (y) , if y S(0; R) .

If y E is a regular boundary point of B(0; R) \ E, then, by Theorem 3.7,

limB(0;R)\E3xy V (x) = h (E).
If, conversely, limB(0;R)\E3xy V (x) = h (E), then the function h (E) V
is a barrier at y with respect to B(0; R) \ E.
V and Uhd0 are bounded and harmonic in B(0; R) \ E and they have the
same boundary limits at all points of S(0; R).
V has boundary limit equal to h (E) at every y E, except at every irre-
gular y. But, by Theorem 9.3, the set of irregular boundary points of B(0; R)\E
is of zero harmonic measure with respect to B(0; R) \ E.
Uhd0 has, also, boundary limit h (E) at every y E, except at every
y E with Uhd0 < h (E). But, by Theorem 8.2, all these ys belong to a
Borel set of zero h-capacity and, hence, of zero harmonic measure with respect
to B(0; R) \ E.
By Theorem 3.5,
Uhd0 = V
in B(0; R) \ E and the proof is, now, clear.

9.5 Capacity and conformal mapping

We shall, now, study, in the framework of potential theory, a fundamental sub-
ject of complex analysis, namely the existence of conformal mapping between
simply connected open sets.
9.5. CAPACITY AND CONFORMAL MAPPING 245

Lemma 9.3 If I is any compact linear segment in R2 , then Ch (I) > 0.

Proof:
If d is the linear Lebesgue measure on I, then we easily estimate
Ih (d) < + .
Proposition 9.3 If E is any continuum in R2 , then Ch (E) > 0.
Proof:
Let a, b E with a 6= b and consider the segment I = [a, b].
If
P r : E E0
is the orthogonal projection of E on the line containing I, then this function is
a contraction and E 0 = P r(E) I. From Proposition 8.7,
dh (E) dh (E 0 ) dh (I) > 0 .

The following is just an extension of the corresponding definition for subsets

of R2 .
Definition 9.2 An open set R2 is called simply-connected if it is connected
and R2 \ is, also, connected.
It is clear that symmetry with respect to any circle preserves the property
of simple-connectedness.
Through the Kelvin Transform and in view of Proposition 9.3, the first part
of the next result is identical to the well known theorem of complex analysis.
The proof which is presented here is not the standard proof presented in the
elementary graduate courses of function theory. It is based on the existence of
the Greens function and, thus, the proof is reduced to a maximization problem,
exactly as the standard proof.
Theorem 9.4 (The Riemann Mapping Theorem) Suppose that is a simply-
connected open subset of R2 with and E is its (compact connected)
complement. Then, there is a conformal mapping of onto B(0; 1), i.e. a
function
: B(0; 1) ,
which is one-to-one in , onto B(0; 1) and holomorphic in , if and only if
Ch (E) > 0 .
In this case we can arrange it so that, also, () = 0 and, then necessarily,
|0 ()| = eh (E) = dh (E) = Ch,R (E)
and
1
G
(x) = log , x,
|(x)|
is the Greens function of with respect to .
246 CHAPTER 9. THE CLASSICAL KERNELS

Proof: 1. Suppose that

: B(0; 1)
is one-to-one, onto, conformal and that () = 0.
If E = {a} consists of only one point, then the Kelvin Transform of with
respect to any circle centered at a is bounded and holomorphic in R2 . Both its
real and its imaginary parts are harmonic functions bounded in R2 and, hence,
by the Theorem of Picard, and, therefore, is a constant function. Similarly,
if E is empty, then the restriction of in R2 is bounded and holomorphic in
R2 and, hence, is constant.
We conclude that E has more than one point and, since E is a continuum,
Proposition 9.3 implies that Ch (E) > 0.
2. Suppose, conversely, that Ch (E) > 0 and consider the h-equilibrium measure
d0 of E and the h-potential

1
Uhd0 (x) = log d0 (y) , x R2 .
E |x y|

After a translation, we may suppose that 0 E.

By Theorem 8.2, the function

|x|
U (x) = h (E) Uhd0 (x) = log |x| + h (E) log d0 (y)
E |x y|

is non-negative in \ {} .
Hence,

|x|
h (E) log d0 (y) log |x| = h (x)
E |x y|

for all x \{} and we observe that the left side of this inequality is harmonic
in \ {} and that its limit at is the finite number h (E). By Theorem 4.1,
it is a harmonic majorant of h in . Therefore, has a Greens function
with respect to and, moreover,

U
G

in .
We shall, in fact, prove that G = U in .
Since, by Proposition 3.5, all points of E are regular boundary points of ,
Proposition 9.2 implies that

lim Uhd0 (x) = h (E)

3xy

for every y in E. We may observe that in this particular situation, where all
points of E are regular boundary points of , the proof of Proposition 9.2
simplifies, as it does not need Theorem 9.3.
9.5. CAPACITY AND CONFORMAL MAPPING 247

Now, let u be any harmonic majorant of h in and consider the function

|x|
V (x) = u(x) + h (E) log d0 (y) , x \ {} .
E |x y|

V is harmonic in and

V (x) h (E) Uhd0 (x)

for all x \ {}. Hence,

lim sup V (x) 0

3xy

for all y E. We conclude that

V 0

in . Therefore,
|x|
h (E) log d0 (y)
E |x y|
is the smallest harmonic majorant of h in and, finally,

|x|
G (x) = h (E) log d0 (y)+h (x) = h (E)Uhd0 (x) , x.
E |x y|

3. We, now, define

|x|
h(x) = h (E) log d0 (y) , x.
E |x y|

We have, already, seen that h is the smallest harmonic majorant of h in

and that G (x) = log |x| + h(x) , x .
Since is simply connected, there is, by Theorem 1.4 through the Kelvin
Transform, a harmonic conjugate g of h in .
Consider, also, the many-valued function

F (x) = arg(x) + g(x) , x \ {} ,

and the single-valued

1 h(x)ig(x)
(x) = eG (x)iF (x) =

e , x \ {} ,
x
which, by Theorem 4.1, is analytic in , since the right side is analytic in
with value 0 at .
4. We consider, for every > 0, the open sets

= {x : G
(x) > } .
248 CHAPTER 9. THE CLASSICAL KERNELS

If x and G (x) = , then in any B(x; r) \ {} there are points of

. Otherwise, by the Maximum-Minimum Principle, G

would be constant
in B(x; r) and, by Theorem 1.10, in the connected \ {}. Hence, x belongs
to .
Conversely, if x belongs to , then x . Otherwise, x E and x
would be a regular boundary point of , implying that lim3zx G (z) = 0.
But lim 3zx G (z) , resulting to a contradiction. Thus,

= {x : G
(x) = } = {x : |(x)| = e }.

Since 0 is holomorphic in , its zeros (the critical points of ) are at most

countably many. Therefore, the set of critical values,

{w : w = (x) for some x with 0 (x) = 0} ,

is at most countable. Furthermore, the set

= { > 0 : e = |w| for some critical value w of }

is at most countable.
If > 0 and / , then the function e2 ||2 is a C -defining function

of at all its boundary points.
Now, is an open subset of with C 1 -boundary, is holomorphic in
and ( ) S(0; e ).
By Theorem 0.7 applied through the Kelvin Transform, for every value y
B(0; e ) the total multiplicity of all solutions of the equation (x) = y, x ,
is equal to the multiplicity of the only solution, , of (x) = 0. Since

|0 ()| = lim |x(x)| = eh() = eh (E) 6= 0 ,

x

this multiplicity is exactly 1. Hence, is a bijection of onto B(0; e ).

Considering a sequence of s in R+ \ converging to 0+, we conclude that
is a bijection of onto B(0; 1).
5. Suppose, now, that : B(0; 1) is another conformal mapping with
() = 0.
We can prove that
lim |(x)| = 1
3xy

for all y . In fact, let {xn } be some sequence in with xn y

and (xn ) w0 for some w0 B(0; 1). Take x0 so that (x0 ) (= w0 and)
consider a small disc B(x0 ; ) . Since is one-to-one, w0 / S(x0 ; )
and, by Theorem 0.7, every w which is close enough to w0 can be written as
w = (x) for some x B(x0 ; ). But, eventually, all wn = (xn ) are close to
w0 while xn / B(x0 ; ). This is impossible, since is one-to-one.
Now, since is holomorphic at and () = 0, we have that 0 () =
limx x(x) C and, therefore, the function
1
log log |x| , x,
|(x)|
9.6. CAPACITY AND GREENS FUNCTION IN R2 249

is harmonic in .
This implies that the function
1 1 1
log log (x) log
= G , x,
|(x)| |(x)| |(x)|

is harmonic in . Its boundary limits are all 0 and, hence, it is identically 0 in

.
Now, it is clear that
| 0 ()| = eh (E)
and
1
G
(x) = log , x.
|(x)|

Example
Consider the line segment I = [l, l].
If = R2 \ I, then the conformal mapping : B(0; 1) is the inverse
mapping of
l( 1)
x = y+ .
2 y
Therefore,
 l ( 1 ) |I|
  l
Ch,R (I) = dh (I) = |0 ()| = lim |x(x)| = lim y y+  = =
x y0 2 y 2 4

and, thus,
4
h (I) = log .
|I|

9.6 Capacity and Greens function in R2

We shall state and prove a characterization of all open subsets of R2 which have
a Greens function in each of their components.
We remember that the problem in Rn is completely solved, if n 3, since,
in this case, all open sets have a Greens function.
We, also, observe that, if the open set has more than one components,
then it has a Greens function in every one of its components. This is true
because, in this case, every component is disjoint from some ball.
Hence, it is no loss of generality to assume that the open set is connected.
It is equivalent, through an application of the Kelvin transform, to consider the
case of a connected open set containing .

Theorem 9.5 Suppose that the open connected set R2 contains and let
E be the (compact) complement of .
Then, has a Greens function if and only if Ch (E) > 0.
250 CHAPTER 9. THE CLASSICAL KERNELS

If Ch (E) > 0, then

d0
(x) = h (E) Uh (x) ,
G x,

where d0 is the h-eqilibrium measure of E.

Proof:
If = R2 , then we, already, know that has no Greens function and, also,
that E = has zero logarithmic capacity.
Hence, we assume that there is at least one point not in and, through a
translation, we may assume that 0 / .
1. Suppose, now, that Ch (E) > 0 or, equivalently, that

h (E) < + .

Consider the h-equilibrium measure d0 of E and the h-potential

d0 1
Uh (x) = log d0 (y) , x R2 .
E |x y|

By Theorem 8.2, the function

|x|
U (x) = h (E) Uhd0 (x) = log |x| + h (E) log d0 (y)
E |x y|

is non-negative in \ {} .
Hence,

|x|
h (E) log d0 (y) log |x| = h (x)
E |x y|

for all x \ {} and we observe that the right side of this inequality is
harmonic in \ {} and that its limit at is the finite number h (E). By
Proposition 4.7, it is a harmonic majorant of h in and, therefore, has a
Greens function with respect to and, moreover,

U
G

in .
2. If is a regular set, then we repeat part 2 of the proof of Theorem 9.4 and
conclude that G = U in .
In general, since E is capacitable, we may consider a sequence of open sets
Om so that E Om {x : d(x, E) < m 1
}, Ch (Om ) Ch (E) and Om+1 Om
for all m. We take, next, open sets Nm with C 1 -boundary so that Om+1
Nm Nm Om and let m = R2 \ Nm . Then, Ch (Nm ) Ch (E) and, by the
previous discussion,
d
G
m
= h (Nm ) Uh 0,m
everywhere in m , where d0,m is the h-equilibrium measure of Nm .
9.7. POLAR SETS AND THE THEOREM OF EVANS 251

Since m , Theorem 5.2 implies that G

G everywhere in .
m

Therefore,
d0,m
G (x) = h (E) lim Uh (x)
m+

for all x .
If we fix an x and, then, take an m0 so that m10 < 12 d(x, E), we have that
all measures d0,m , m m0 , are supported in the compact set {z : d(z, E)
1
2 d(x, E)}. Taking any subsequence d0,mk weakly converging in this compact
set to some probability Borel measure d, we easily get that d is supported in
E. By Proposition 7.1,
Ih (d) lim inf Ih (d0,mk ) = lim inf h (Nmk ) = h (E) .
k+ k+

By Theorem 8.2, Ih (d) = h (E) = Ih (d0 ) and, by Theorem 8.3, d = d0 .

Hence, every weakly convergent subsequence of {d0,m } has d0 as weak limit
and this implies that {d0,m } converges weakly in {z : d(z, E) 21 d(x, E)} to
d0 .
Since the function hx is continuous in {z : d(z, E) 12 d(x, E)},
d0,m
Uh (x) Uhd0 (x) .
Thus,
d0
(x) = h (E) Uh (x)
G
for every x .
3. Assume, conversely, that Ch (E) = 0.
If h has a harmonic majorant h in , then, for every R which is large
enough so that E B(0; R), we have that h()+log || is harmonic and bounded
from below in B(0; R) \ E and
h(y) + log |y| m(R) + log R
for all y S(0; R), where m(R) = minS(0;R) h.
Since, by Theorem 9.2, E = E is of zero harmonic measure with respect to
B(0; R) \ E, we get from Theorem 3.5 that, for every x B(0; R) \ E,
h(x) + log |x| m(R) + log R .
Now, since limR+ m(R) = |h()|, letting R +, we get
h(x) = + .
We, thus, arrive at a contradiction and, hence, has no Greens function.

9.7 Polar sets and the Theorem of Evans

Definition 9.3 A set A Rn is called locally polar if for every x A there
is some B(x; r) and a function u superharmonic in B(x; r) so that
u(x) = + , x A B(x; r) .
252 CHAPTER 9. THE CLASSICAL KERNELS

Lemma 9.4 Suppose that u is superharmonic in B(x0 ; R) and let 0 < r < R.
Then, there exists a function U superharmonic in Rn so that

U (x) = u(x) , x B(x0 ; r) .

If n 3, we may also have U be bounded from below in Rn .

Proof:
Take r1 , R1 so that r < r1 < R1 < R and cover the ring B(x0 ; r1 , R1 ) by
finitely many open balls B1 , . . . , BN all of which are contained in B(x0 ; r, R).
Then, the function v = min(uB1 , . . . , uBN ) is superharmonic in B(x0 ; R),
bounded in B(x0 ; r1 , R1 ) and satisfies

v(x) = u(x) , x B(x0 ; r) .

If M = supxB(x0 ;r1 ,R1 ) v(x) and m = inf xB(x0 ;r1 ,R1 ) v(x), find a R+ and
b R so that ah (r1 ) + b > M and ah (R1 ) + b < m.
Now, it is easy to see that the function

v(x) , if x B(x0 ; r1 )
U (x) = min(v(x), ah(x) + b) , if x B(x0 ; r1 , R1 )

ah(x) + b , if x
/ B(x0 ; R1 ) ,

concludes the proof.

Proposition 9.4 Let A Rn be locally polar and let x0 / A. Then, there

exists a function u superharmonic in Rn so that u(x0 ) < + and

u(x) = + , xA.

If n 3, we may also have u > 0 everywhere in Rn .

Proof:
We, first, observe that A 6= Rn . In fact, if we take an arbitrary x A, a
ball B(x; R) and a u superharmonic in B(x; R) which is = + identically in
A B(x; R), then u must be finite almost everywhere in B(x; R). Therefore
there is at least one point of this ball not belonging to A. (Continuing this
argument, we may, easily, show that A has zero Lebesgue measure.)
Fix x0
/ A and for each x A we consider a B(x; Rx ) not containing x0 and
a ux superharmonic in B(x; Rx ) so that u = + identically in A B(x; Rx ).
We, then, take rx < Rx and Lemma 9.4 provides us with a Ux superharmonic
in Rn which is = + identically in A B(x; rx ) and with Ux (x0 ) < +. In
case n 3, we may also suppose that Ux > 0 everywhere in Rn .
We may replace each B(x; rx ) by a smaller open ball containing x and having
rational radius and rational center. We enumerate these countably many balls
and we have, now, constructed: a sequence {Bk } of open balls covering A
and a sequence {Uk } of functions superharmonic in Rn with every Uk being
9.7. POLAR SETS AND THE THEOREM OF EVANS 253

identically + in A Bk and, in case n 3, positive everywhere in Rn . Also:

Uk (x0 ) < +.
Case 1: n 3. +
For each k we take k > 0 so that k=1 k Uk (x0 ) < +. Then, the function

+
u = k Uk
k=1

is superharmonic and positive in Rn and is identically + in A.

Case 2: n = 2.
For each k we set mk = minxB(x0 ;k) Uk (x) and we find k > 0 so that
+
k=1 k (Uk (x0 ) mk ) < +. Then, the function

+
u = k (Uk mk )
k=1

is superharmonic in Rn and is identically + in A. The reason for both is that

the terms of the series are, eventually, non-negative in every large ball B(0; R).

Definition 9.4 Suppose that A Rn . The set A is called polar if there is a

function u superharmonic in Rn so that

u(x) = + , xA.

Thus, Proposition 9.4 says that a set A is polar if and only if it is locally
polar.
Observe that the Theorem of Evans implies that every compact E Rn
with Ch (E) = 0 is polar.
Theorem 9.6 Suppose that A Rn .
1. If Ch (A) = 0, then A is polar.
2. If n 3, then the converse of 1 is, also, true.
If n = 2, then the converse of 1 is true, provided that A is bounded.
Proof:
1. Suppose that Ch (A) = 0 and let A be bounded. Consider a bounded open set
O Rn with A O and Ch (O) <  < 1. Now, take any compact exhaustion
{K(m) } of O. It is true that Ch (K(m) ) Ch (O). From Theorem 8.5, we, also,
have that h (K(m) ) = Ch (K1 (m) ) Ch1(O) .
For each m consider the h-equilibrium measure dm of K(m) which, after
Proposition 9.1, is supported in K(m) . Then, Uhdm = h (K(m) ) identically in
the interior of K(m) and Uhdm h (K(m) ) everywhere in Rn .
Replacing, if necessary, by a subsequence, assume that dm d weakly in
the compact set O, where d is a probability Borel measure in O.
For an arbitrary x O, take m0 so that x is in the interior of K(m0 ) .
Now, since hx is continuous in O \ int(K(m0 ) ) and since all dm , m m0 , are
254 CHAPTER 9. THE CLASSICAL KERNELS

supported in this compact set, we get Uhd (x) = limm+ Uhdm (x) = 1
Ch (O) .
Hence
1 1
Uhd (x) = , xO.
Ch (O) 
We, thus, find a decreasing sequence of bounded open sets {Ok } with A
Ok and a sequence of probability Borel measures {dk }, where each dk is
supported in Ok and

Uhdk (x) 2k , xA.

Now, consider the probability Borel measure

+
1
d = dk .
2k
k=1

It is obvious that
Uhd = +
everywhere in A and A is polar.
If A is unbounded, for each x A we consider a B(x; r) and, since AB(x; r)
is bounded with zero h-capacity, we have that it is a polar set. A is, thus, locally
polar and, hence, polar.
2. Let A be bounded and let u be superharmonic in Rn so that u = + iden-
tically in A. By Theorem 2.17, there exists a compactly supported probability
measure d so that
Uhd = +
everywhere in A.
For arbitrary k > 0, consider the bounded open set

O = {x : Uhd (x) > k}

and any compact E O with Ch (E) > 0.

If d0 is the h-equilibrium measure of E, then

1
= h (E) Uhd0 (x) d(x) = Uhd (x) d0 (x) k .
Ch (E) supp(d) E

Hence,
1
Cho (A) Ch (O)
k
and, since k is arbitrary, Ch (A) = 0.
If A is unbounded and n 3, we consider, for each k N, the polar sets
Ak = A B(0; k) which, by the previous argument, have Ch (Ak ) = 0 and, then,
use Theorem 8.1.
9.8. THE THEOREM OF WIENER 255

Theorem 9.7 Suppose that E Rn is compact with Ch (E) > 0 and x0 E

is a non-regular boundary point of Rn \ E.
1. There is a function u superharmonic in Rn so that u(x) = h (E) for every
x E and u h (E) everywhere in Rn .
2. There is a function v superharmonic in Rn so that v(x) = h (E) for every
x E \ {x0 }, v(x0 ) < h (E) and v h (E) everywhere in Rn .

Proof:
1. Let d0 be the h-equilibrium measure of E. Then, Uhd0 h (E) everywhere
in Rn and the set A = {x E : Uhd0 (x) < h (E)} is polar.
We take u0 superharmonic in Rn so that u0 > 0 in E and u0 = + in A.
Then, the function u = min(Uhd0 + u0 , h (E)) satisfies the properties in the
statement.
2. By Proposition 9.2, we have that

Uhd0 (x0 ) lim inf Uhd0 (x) lim inf Uhd0 (x) < h (E) .
xx0 E63xx0

Now, A \ {x0 } is, also, polar and, by Proposition 9.4, there is a u1 super-
harmonic in Rn with u1 > 0 in E, u1 = + in A \ {x0 } and u1 (x0 ) < +.
Replacing, if necessary, u1 by a small positive multiple of it, we may, also,
suppose that
Uhd0 (x0 ) + u1 (x0 ) < h (E) .

The function we want is v = min(Uhd0 + u1 , h (E)).

9.8 The theorem of Wiener

The next result is a characterization of the regular boundary points of an open
set.
We remark that, if n 3, then for every compact set E Rn , we have
h (E) > 0 and, hence, Ch (E) = h 1(E) . If n = 2 and 0 < < 1, then, for every
compact set E B(x0 ; ), we have h (E) h (B(x0 ; )) = h () > 0 and,
again, Ch (E) = h 1(E) .

Theorem 9.8 (N. Wiener) Let be an open set, x0 Rn and

Ek = {x
/ : k+1 |x x0 | k } ,

for all k N, where is any number with 0 < < 1.

Then, x0 is a regular boundary point of if and only if

+
+
h (k )
h (k )Ch (Ek ) = = + .
h (Ek )
k=1 k=1
256 CHAPTER 9. THE CLASSICAL KERNELS

Proof:
1. At first, we observe that, for every pair of and 0 in (0, 1), the two equalities
+ k + 0k 0
k=1 h ( )Ch (Ek ) = + and k=1 h ( )Ch (Ek ) = + are equivalent.
0
We suppose that 0 < 0 < < 1 and it is trivial to see that every Ek
is contained in a finite number m(, 0 ) of consecutive El s, where m(, 0 )
0
depends only upon and 0 . Proposition 8.4 implies that Ch (Ek ) is not more
than the sum of these consecutive Ch (El )s. (In case n = 2 it may be necessary
to drop the first few ks so that all ring domains are contained in the disc
B(0; 21 ).) It is, also, trivial to see that the quantity h (0k ) is comparable to the
correspoding h (l )s. This, of course, means that the quotients are bounded
both from above and from below by two constants depending only upon the
number m(, 0 ). From all this it is obvious that

+
0
+
h (0k )Ch (Ek ) C(, 0 ) h (l )Ch (El ) .
k=1 l=1
0
(Since consecutive Ek s
may have one common El used to cover them, each
term of the series in the right side of the above inequality is counted at most
twice.)
It is also obvious that every Ek is contained in the union of at most two
0
consecutive El s and, hence

+
+
0
0 0
h (k
)Ch (Ek ) C (, ) h (0l )Ch (El ) ,
k=1 l=1

where each term of the series in the right side of the above inequality is counted
at most m(, 0 ) times.
2. Suppose that

+
h (k )
= +
h (Ek )
k=1

and consider the h-equilibrium measure dk of Ek and its h-potential

Uhdk (x) = hx (y) dk (y) , x Rn .
Ek

We know that
Uhdk (x) h (Ek )
for every x Rn and we shall estimate Uhdk on every
Am = {x : m+1 |x x0 | m } .
If x Am , then

h (k+1 m ) , if k < m 1
dk
Uh (x) (Ek ) , if m 1 k m + 1
h m+1
h ( k ) , if m + 1 < k .
9.8. THE THEOREM OF WIENER 257

Moreover,
h (k ) Uhdk (x0 ) .
Given an arbitrary  > 0 and an arbitrary N N, there is a smallest M N
so that
1
M
h (k )
.
 h (Ek )
k=N

By the minimality of M ,

M
h (k ) 1 h (M ) 1
< + +1 .
h (Ek )  h (EM ) 
k=N

Therefore, we have arbitrarily large integers N and corresponding M N

so that
1 M
h (k ) 1
+1 .
 h (Ek ) 
k=N

Define, now,

M
1
U =  U dk .
h (Ek ) h
k=N

Then,

M
h (k )
U (x0 )  1.
h (Ek )
k=N

Besides the parameters , N, M we introduce, now, an integer l N having

in mind the following. If n = 2, then we fix = 21 and l = 1 and N will depend
upon  in a manner that will be made precise in a moment. If n 3, all these
parameters will depend upon  and we shall shortly see how.
If x Am , trivial estimates show that
h (k+1 m )
U (x)  + 1
h (Ek )
N kM,k<ml N kM,mlkm+l
h (m+1 k )
+
h (Ek )
N kM,m+l<k
{ ( )
3 + (1 + ) 1 + N2 , if n = 2
( )
1
(2l + 1) + (1 + ) n2 + l(n2) (11l )n2 , if n 3.

If N m + l + 1, then in the estimate of U (x) only the third sum exists and,
hence,

 M (m+2) log 2
(1 + ) m+2 N log
6 1
k=N h (Ek ) N 2 log |xx0 | , if n = 2,
U (x)
(1 + ) l n2 ,
l(n2)
if n 3.
(1 )
258 CHAPTER 9. THE CLASSICAL KERNELS

3. To prove that x0 is a regular boundary point of we shall examine the

function
B(x ;)
u = H1| x00 |

in B(x0 ; ).
Since the function 1| x0 | is superharmonic, we have that u 1| x0 |
in B(x0 ; ). Therefore, to prove that 1u is a barrier of at x0 , it is enough
to prove that
lim inf u(x) 1 .
3xx0

We shall compare u with the function

1
V = U .
U (x0 )

We consider the case n = 2 first.

Then V (x0 ) = 1 and we take N 6 .

1. V (x) 1 + 14
3 , if x B(x0 ; 12 ) \ {x0 } and

0|
2. V (x) 12
N log 1
|xx0 | 2 1|xx
|xx0 | , if x Am and m + 2 N .

1
Trivial estimates show that the function 1+10 V , which is harmonic in
1
B(x0 ; 2 ), satisfies
1
V 1 | x0 |
1 + 10
there. Therefore,
1
V u
1 + 10
in B(x0 ; 12 ), implying that

1
lim inf u(x) .
3xx0 1 + 10

Since  is arbitrary, we find

lim u(x) 1 .
3xx0

Now, let n 3.
We take l = [ 3 ] and = 1  3 .
2 1

We, again, have V (x0 ) = 1 and, for a constant C depending only upon n,
1
1. V (x) 1 + C 3 , if x B(x0 ; ) \ {x0 } and

2. V (x) C exp((n 2) 3 ), if x Am and m + l + 1 N .

1
9.8. THE THEOREM OF WIENER 259

1
Now, the function V 2C 3 is harmonic in B(x0 ; ) and satisfies
1
V (x) 2C 3 1 |x x0 |

for every x B(x0 ; ), provided that we take N C 0  3 for some constant

2

C 0 depending only upon n.

1
Indeed, if |x x0 | C 3 , then this inequality is immediate
( 1 ) from 1. above.
1
00 13
If |x x0 | C , then x Am for some m C  log
3
1 . Therefore, the
C 3
already stated choice of N together with 2. finish the proof of the inequality.
This, now, implies that
1
V (x) 2C 3 u(x)

for every x B(x0 ; ) and

1
lim inf u(x) 1 2C 3 .
3xx0

Since  is arbitrary, we find

lim u(x) 1 .
3xx0

This proves that x0 is a regular boundary point of . 4. Consider, now, for

1
all n the particular value = 21 and the sets Ek = Ek2 .
+ h ( 1 )
Suppose that k=1 h (E2kk ) < + and let K 2 be such that

+ 1
h ( 2k+1 )
< 1.
h (Ek )
k=K

We define the bounded open set

e = B(x0 ; 1 ) \ {x0 } \ + Ek .
k=K
2

Since e B(x0 ; 1K ) = B(x0 ; 1K ), to prove that x0 is not a regular

2 2
boundary point of it is enough to prove that it is not a regular boundary
e
point of .
For each k with h (Ek ) < +, let dk be the h-equilibrium measure of Ek
e so that
and let f (x) = 1 2|x x0 | for all x

e
1. 0 f 1 in ,

2. f (x0 ) = 1 and

e \ {x0 } \ +
3. f = 0 in k=0 Ek .
260 CHAPTER 9. THE CLASSICAL KERNELS

We define the function

+
1
V = max f Uhdk ,
k=K e
E k
h (E k )

where we simply omit all terms with h (Ek ) < +.

V is a superharmonic function which is harmonic in e and non-negative and
1
bounded from above in B(x0 ; 12 ), since, for all m 1 and all x Am = Am2
,
1
h ( 2k+1 21m )
V (x) + 1
h (Ek )
Kk,km2 Kk,m1km+1
1
h ( 2m+1 21k )
+
h (Ek )
Kk,m+2k
1
h ( 2k+2 ) h ( 21k )
+3+
h (Ek ) h (Ek )
Kk,km2 Kk,m+2k

+ 1
h ( 2k+2 )
3+ < + .
h (Ek )
k=K

Moreover,

+ 1
h ( 2k+1 )
V (x0 ) < 1.
h (Ek )
k=K

e with
Now, let v be subharmonic and bounded from above in

lim sup v(x) f (y)

e3xy

e
for all y .
Since all terms in the sum defining V are non-negative in B(x0 ; 12 ) and since
dk
Uh = h (Ek ) quasi-almost everywhere in Ek , it is clear that

lim sup v(x) V (y) lim inf V (x)

e3xy e3xy

e \ {x0 } except for a boundary subset of at most zero harmonic

for all y
measure with respect to .e Theorem 3.5 implies that

v V
e
in .
Therefore,
e V
Hf
e
in .
9.8. THE THEOREM OF WIENER 261

We assume that x0 is a regular boundary point of e and we shall arrive at

a contradiction.
e
e3xx0 Hf (x) = f (x0 ) = 1 we get 1 lim inf

From lim e3xx0 V (x). This,
immediately, implies

1
1 lim inf V (x) dm(x) .
k+ m(Ak \ Ek ) Ak \Ek

By the superharmonicity of V ,

1
lim V (x) dm(x) = V (x0 ) .
k+ m(Ak ) Ak

The last two relations together with the

1 1
V (x) dm(x) V (x) dm(x)
m(Ak ) Ak m(Ak ) Ak \Ek
(
m(Ek ) ) 1
= 1 V (x) dm(x)
m(Ak ) m(Ak \ Ek ) Ak \Ek

imply that
m(Ek )
lim inf 1 V (x0 ) > 0 .
k+ m(Ak )

By the definition of capacity,

1
h (Ek ) h (|x y|) dm(x)dm(y)
m(Ek )2 Ek Ek

1
h (|x y|) dm(x)dm(y)
m(Ek )2 Ak Ak

2k(n+2) m(Ek )2 A0 A0 h (|x y|) dm(x)dm(y) , if n 3
1

=
1
16k m(Ek )2 A0 A0
h (|x y|) dm(x)dm(y)

+ k log 2 m(A )2 ,
16k m(Ek )2 0 if n = 2 .
h ( 21k )
C ,
22kn m(Ek )2

for some constant C > 0 depending only upon n.

Therefore, for some other constant C 0 depending only upon n,

m(Ek )2 h ( 21k )
2
C0 .
m(Ak ) h (Ek )

+ h ( 1k )
This contradicts the convergence of the series 2
k=1 h (Ek ) .
262 CHAPTER 9. THE CLASSICAL KERNELS

9.9 Thin Sets

Definition 9.5 A set E Rn is called thin at x0 Rn , if either x0 is not an
accumulation point of E or x0 is an accumulation point of E and there exists a
superharmonic function u in Rn so that u(x0 ) < lim inf E\{x0 }3xx0 u(x).

Theorem 9.9 Suppose E Rn and x0 Rn . Let

1 1
Ek = {x E : |x x0 | },
2k+1 2k
for all k N. Then, the following are equivalent:
1. E
+is thin at x0 .
1 o
2. k=1 h ( 2k )Ch (Ek ) < +.

Proof:
If x0 is not an accumulation point of E, then E is, automatically, thin at
x0 and the series in 2 converges, since, then, Ek is empty for all large k. We
assume, therefore, that x0 is an accumulation point of E.
Suppose that the series in 2 converges. For every k N, we take open sets
Ok so that
( 1 1 )
Ek Ok B x0 ; k+2 , k1
2 2
and

+
1
h ( )Ch (Ok ) < + .
2k
k=1

We know, from the proof of Theorem 9.6, that there exist probability Borel
measures dk supported in Ok so that

1
Uhdk (x) =
Ch (Ok )

for every x Ok . It is easy to see that

1
Uhdk (x0 ) h ( ).
2k+2
We, now, take K 2 and define the function

+
u0 = Ch (Ok )Uhdk .
k=K

Since K 2, all terms in this series are non-negative in B(x0 ; 12 ) and we, also,
have that the series converges at x0 . Hence, u0 is superharmonic in B(x0 ; 12 )
and
u0 (x) 1
9.9. THIN SETS 263

for every x E B(x0 ; 0, 21K ). Taking K large enough we, also, have u0 (x0 ) < 1.
Thus,
u0 (x0 ) < lim inf u0 (x) .
E\{x0 }3xx0

Applying Lemma 9.4, we prove that E is thin at x0 .

Now, suppose that E is thin at x0 and take u superharmonic in Rn so that

u(x0 ) < lim inf u(x) .

E\{x0 }3xx0

Consider so that

u(x0 ) < < lim inf u(x)

E\{x0 }3xx0

and K large enough in order to have

u(x) >

for every x E B(x0 ; 0, 2K1

1
).
Define
O = {x Rn : u(x) > }
and
1 1
Ok = O B(x0 ; , ).
2k+2 2k1
Assume that

+
1
h ( )C o (Ek ) = +
2k h
k=1

and, hence,

+
1
h ( )Ch (Ok ) = + .
2k
k=1

Take, for each k K, compact sets Fk Ok so that

+
1
h ( )Ch (Fk ) = + .
2k
k=K

Consider the compact set

+
F = {x0 } Fk
k=K

and its complement = Rn \ F .

By Theorem 9.8, x0 is a regular boundary point of .
264 CHAPTER 9. THE CLASSICAL KERNELS

On the other hand, the superharmonic u is everywhere in F . If we

consider the function


, if x B(x0 ; 2K+1
1
),
f (x) = m, if x S(x0 ; 2K ) ,
1

2K+1 (m )|x x | + 2 m ,
0 if x B(x0 ; 2K+1
1
, 21K ) ,

where m = minyS(x0 ; 1
) u(y), then f is a continuous boundary function of the
2K
open set K = B(x0 ; 21K ) and

u f

everywhere in K except at the point x0 . Since {x0 } is of zero harmonic

measure with respect to K , this implies that

u HfK

everywhere in K . Therefore,

= f (x0 ) = lim HfK (x) lim inf u(x) = u(x0 )

K 3xx0 K 3xx0

and we get a contradiction.