A VERY SHORT SURVEY ON BOUNDARY BEHAVIOR OF HARMONIC

FUNCTIONS

BLACK

Abstract. This short expository paper aims to use Dirichlet boundary value problem to elab-
orate on some of the interactions between complex analysis, potential theory, and harmonic
analysis. In particular, I will outline Wiener’s solution to the Dirichlet problem for a general
planar domain using harmonic measure and prove some elementary results for Hardy spaces.

1. Introduction
Definition 1.1. Let Ω Ă R be an open set. A function u P C 2 pΩ; Rq is called harmonic if
2

B2 u B2 u
(1.1) ∆u “ ` 2 “ 0 on Ω.
Bx2 By
The notion of harmonic function can be generalized to any finite dimensional Euclidean space (or
on (pseudo)Riemannian manifold), but the theory enjoys a qualitative difference in the planar case
due to its relation to the magic properties of functions of one complex variable. Think of Ω Ă C
(in this paper I will use R2 and C interchangeably when there is no confusion). Then the Laplace
operator takes the form
ˆ ˙ ˆ ˙
B B B 1 B B B 1 B B
(1.2) ∆“4 , where “ ´i and “ `i .
Bz̄ Bz Bz 2 Bx By Bz̄ 2 Bx By
Let HolpΩq denote the space of holomorphic functions on Ω. It is easy to see that if f P HolpΩq,
then both <f and =f are harmonic functions. Conversely, if u is harmonic on Ω and Ω is simply
connected, we can construct a harmonic function ũ, called the harmonic conjugate of u, via
Hilbert transform (or more generally, Bäcklund transform), such that f “ u ` iũ P HolpΩq. If
f P HolpΩq, then f satisfies the maximal modulus principle and the Cauchy integral formula, and
f is complex analytic. In correspondence, it is well known from the theory of elliptic PDEs that if
u is harmonic, then u satisfies the maximal modulus principle and the mean value property (as a
counterpart of Cauchy integral formula), and u is real analytic. Another important result, which
mirrors Hurwitz’s theorem, states that local uniform convergence of a sequence of harmonic function
results in a harmonic function. This is called Harnack’s principle. We will use this important
property later in the paper.
One should also expect the qualitative difference between planar harmonic function theory and
higher-dimensional harmonic function theory by the fact that the symmetry group of Laplace equa-
tion in Rn is the conformal group of Rn , and the conformal groups of R2 and Rn for n ą 2 have
drastically different characterization. I will not dwell into this subject.

Key words and phrases. Dirichlet problem, harmonic measure, Hardy spaces.
The author uses the name Fushuai Jiang in his government documents. He is not supported by anybody and is
slowly running out of pennies.
1
2 BLACK

I am interested in the following

Definition 1.2. Let Ω Ă R2 be a domain (connected and open, here and below), and let Γ “ BΩ.
The Dirichlet problem consists of the following: given f P CpΓq, find u P C 2 pΩq X CpΩq such
that
(1.3) ∆u “ 0 in Ω
(1.4) u“f on Γ
A domain such that the Dirichlet problem can be solved for all f P CpΓq is called a Dirichlet
domain.
A simple example of a Dirichlet domain is in place.

2. The unit disc and the Riemann mapping theorem

2.1. Dirichlet problem on the unit disc. Let Ω “ D “ t|z| ă 1u be the unit disc. Then Γ “ T
is the unit circle. It is well known that the boundary value problem
(2.1) ∆u “ 0 in D
(2.2) u“f on T
has a unique solution for each f P CpTq. In particular, the solution has an elegant representation
formula. First consider the Poisson kernel given by
1 ´ r2
ˆ iθ ˙
e `z
(2.3) P pr, θq “ “ < iθ
1 ´ 2r cos θ ` r2 e ´z
for 0 ď r ă 1 and θ P T. Note that the Poisson kernel is a harmonic approximation to the identity.
The unique solution is given by convolving the boundary data with the Poisson kernel
ż ż 2
1 iθ 1 1 ´ |z|
(2.4) uf pr, φq “ f pe qP pr, φ ´ θqdθ “ f peiθ q 2 dθ.
2π T 2π T |eiθ ´ z|
Standard results about convolution tells us that uf converges uniformly to f as r Ñ 1.
Remark 2.1. The Poisson representation formula in essence owes its elegance to the L2 theory on
the torus and the nice form that Laplace operator takes on Euclidean domain. Also note that the
integral formula (2.4) makes sense even if f P L1loc pTq “ L1 pTq, so Equation 2.1 is satisfied. Since
L1 functions do not have pointwise value, Equation 2.2 of course does not make sense in general,
but it does at points where f is continuous. Moreover, in the disc model, since the convergence is
uniform, we can approach any boundary point along any curve. This turns out not to be true in
general. It seems like we need to synthesize a multifaceted problem:
‚ How does the regularity of Γ affect solvability?
‚ How does the regularity of f affect solvability?
In light of the Riemann mapping theorem (which I will present in the next section), I will add the
third
‚ How does the topology of Ω affect solvability?
 A VERY SHORT SURVEY ON BOUNDARY BEHAVIOR OF HARMONIC FUNCTIONS 3

2.2. Simply connected domain and boundary behavior of conformal maps. The cele-
brated Riemann mapping theorem states that every simply connected proper subdomain of C is
biholomorphic to the unit disc. This leads us to look at the boundary value problem (1.3) when Ω
is a simply connected subdomain. Obviously Equation 1.3 can be solved. However, we know that
conformal maps do not “see” boundaries when no conditions are put on the latter. It turns out
that as long as we put slightly more constraint on Γ, then the problem is resolved. First I start
with an important definition.
Definition 2.2. A curve γ Ă C is a Jordan curve if it is the image of an injective continuous
map T ÝÑ C. A Jordan arc is is the image of an injective continuous map of a closed interval
into the plane. Let Ω Ă C be a simply connected domain, then Ω is a Jordan domain if Γ “ BΩ
is a Jordan curve.
We have the celebrated
Theorem 2.3. (Carathéodory). Let φ be a conformal mapping from D onto a Jordan domain
Ω. Then φ has a continuous extension to D and the extension restricts to a homeomorphism from
T onto Γ.
Remark 2.4. The proof of the theorem is purely topological and can be found in [3].
In light of Theorem 2.3, any Jordan domain is a Dirichlet domain.
Let φ be as above. For further interest, we can investigate the relation between regularity of Γ
and that of φ on T. We can also study the relation between regularity of f P CpΓq and that of uf
at points of Γ. There are two results:
Theorem 2.5. (Kellogg.) Let φ be a conformal map from D onto a Jordan domain Ω. Let k ě 1
and α P p0, 1q. Then Γ P C k,α , if and only if arg φ1 P C k´1,α pTq, if and only if φ P C k,α pDq and
φ1 ‰ 0 on D.
Corollary 2.6. Suppose Γ P C k,α (k ě 1) and f P C m,β pΓq. Then uf P C n,σ pΩq where n “
min tk, mu and σ “ min tα, βu and uf is constructed in terms of conformal maps.
This finishes the discussion on Jordan domain.

3. A Potential Theory Approach

In this section, I would like to experiment with some weakened boundary regularity and topo-
logical constraints and sketch the proof for Wiener’s solution to the Dirichlet problem (for details,
see [3]) for an arbitrary domain. The machinery is heavy, so due to the length constraint, much
must be taken on faith.
Continuity does not see pathology, and the job is in turn done by measure theory.
3.1. Motivation. We can interpret the formula (2.4) as integrating f against a measure ωpz, Dq.
Let E be a finite union of open arcs on T. We define the harmonic measure of E at z P D to be
ż 2
1 1 ´ |z|
ωpz, E, Dq “ dθ.
2π E |eiθ ´ z|2
By property of Poisson kernel and Lebesgue integral, ωpz, ¨, Dq extends to a Borel probability
measure on T. Our goal is to construct harmonic measure for a large class of domains.
4 BLACK

3.2. Finitely connected domain. We have solved the Dirichlet problem for Jordan domain. One
should expect that if the domain contains slightly more holes, namely, if BΩ consists of finitely many
pairwise disjoint Jordan curves (in this case, we call Ω finitely connected Jordan domain), the
result is still true. Indeed, we have
Theorem 3.1. Finitely connected Jordan domains are Dirichlet.
Sketch of proof. We can find two points a, b P CzΩ and a Jordan curve γ connecting a and b that
cuts through all the “holes” of Ω such that Ωzγ consists of finitely many disjoint Jordan domains.
We can first solve the Dirichlet problem for one connected component by specifying a consistent
boundary data. By performing this procedure inductively while making sure that the boundary
data is consistent, we can solve the Dirichlet problem for Ω. Uniqueness is clear from maximum
principle. 
Fix a finitely connected domain Ω and z P Ω. Let uf be the unique solution to the Dirichlet
problem with boundary data f . Then the map CpBΩq ÝÑ R defined by f ÞÑ uf pzq is a bounded
linear functional. By Riesz-Markov-Kakutani theorem[4], there exists a unique Borel probability
measure such that
ż
(3.1) uf pzq “ f pζqdω.
BΩ

We call the quantity ωpEq “ ωpz, E, Ωq the harmonic measure of the set E Ă BΩ at z P Ω. By
uniqueness, we see that this definition agrees with the case when Ω “ D.

3.3. Logarithmic Capacity.

Definition 3.2. Let ν be a compactly supported signed measure. The energy integral of ν is
defined to be
ij
1
(3.2) Epνq “ log dνpzqdνpζq
|z ´ ζ|
whenever the integral is absolutely convergent. Let E Ă C be a compact set. Denote by PE the
space of Borel probability measure on E. The Robin’s constant of E is defined to be
(3.3) V pEq “ inf Epνq.
νPPE

The logarithmic capacity of E is defined to be CappEq “ e´V pEq .

Remark. In constrast with the Lebesgue measure, which is a notion of volume, the logarithmic
capacity is a notion of the ability of a set to hold a certain type of “potential energy” (the name
actually comes from the term “capacitance” in physics), depending on how we define the potential
energy. As an example, though the usual 1{3-Cantor set has Lebesgue measure zero, it has log-
arithmic capacity greater or equal to 1{9. It is not surprising that the notion of capacity should
relate to Hausdorff content. An introduction can be found in Carleson’s paper [5], and I will not
dwell too much into the subject. Note that logarithmic capacity also plays an important role of
complex dynamics, see[6].

3.4. Statement of Wiener’s result. Let Ĉ be the Riemann sphere constructed from one-point
compactification. The theorem states
Theorem 3.3. (Wiener, 1924). If Ω Ă Ĉ is such that CappCzΩq ą 0. Then Ω is Dirichlet.
 A VERY SHORT SURVEY ON BOUNDARY BEHAVIOR OF HARMONIC FUNCTIONS 5

Remark. Wiener’s idea is to exhaust the set Ω by finitely connected Jordan domains, and since
each solution on these nice domains is associated with a harmonic measure, one can expect a
harmonic measure in the limit. The hard part of the proof is the limiting argument, and I break it
down into two steps.
3.5. A technical lemma. One major difficulty in Wiener’s idea is to show that the limit does not
depend on our choice of exhaustion. This is taken care of by the following
Lemma 3.4. The infimum in (3.3) is uniquely attained on PE if CappEq ą 0.
Discussion of proof. Let Ω “ ĈzE be the component that includes 8. If Ω is finitely connected
and BΩ is smooth, then we can show that ωp8, ¨, Ωq is the unique minimizer. The idea is the
following: first construct a Green’s function1 gΩ associated with Ω with a pole at infinity (which
is possible by Theorem 3.1), and express Robin’s constant in terms of the Green’s function; the
regularity assumption on BΩ ensures that there is no boundary contribution to the area integral and
helps us conclude that ωp8, ¨, Ωq is the minimizer; the positive-definiteness of E helps us achieve
uniqueness, and during this step we have to interchange the order of integration, the validity of
which is justified by the finite capacity condition.
If E is any compact set, then we can consider a smooth 8-exhaustion by finitely connected
domain
ď8
(3.4) 8 P Ωn Ă Ωn Ă Ωn`1 , Ω “ Ωn , BΩn P C 8 .
n“1

Let En “ ĈzΩn . By the previous step, µn “ ωp8, ¨, Ωn q is the unique witness of the infimum of (3.3)
for En . By Banach-Alaoglu theorem (or Prokhorov’s theorem), we can pass onto a subsequence
and obtain a weak-star limit µ P PE (or in the language of probability, a vague limit). Since weak
star convergence is lower semi-continuous, µ is a minimizer. Uniqueness follows from a variation
argument. 
3.6. Sketch of Proof of Wiener’s theorem. Let Ω be smoothly exhausted as in (3.4). Con-
tinuously extend f to all of Ĉ in an arbitrary fashion. Let ωpz, ¨, Ωn q be the harmonic measure
associated with Ωn and set
ż
(3.5) un pzq “ f pζqdωpz, ζ, Ωn q.
BΩn
Let En “ CzΩn , E “ CzΩ, and let un “ µEn be defined as above, then µn á˚ µE . Therefore,
ż ż
(3.6) lim up8q “ lim f pζqdµn pζq “ f pζqdµE pζq.
nÑ8 nÑ8 BΩ BΩ
n

The last integral obviously does not depend on the choice of exhaustion of Ω or on the choice of
extension of f .
For 8 ‰ z P Ω, we can apply a Möbius transformation φ mapping z to 8. Möbius transforma-
tion preserves the positivity of capacity. Therefore, un pzq converges to a limit, which we call upzq
independently of the choice of exhaustion and boundary value extension. Note that the sequence un
is uniformly bounded and harmonic, so the convergence is locally uniform. By Harnack’s principle,
the limit function is also harmonic and has boundary value f . 

1The modern notion of a Green’s function is the fundamental solution of a differential operator, more precisely,
the function is mapped to the Dirac delta by the operator in the sense of distribution.
6 BLACK

4. Hardy and Dirichlet Spaces

In the previous section I mainly focus on recovering a harmonic function from its boundary data.
In this section, I will investigate the “dual” of the issue, namely, given a holomorphic or harmonic
function on a domain, how does it behave on the boundary? For an arbitrary domain the question
is very hard to answer. In this paper, I will again return to the toy model Ω “ D and Γ “ T. I
will mainly focus on the holomorphic case. The subject has been extensively studied in harmonic
analysis.

4.1. Hardy Space. We put functions in different classes in order to understand their behaviors.
For example, Lp -class measures the size of the function as well as their asymptotic decay behavior.
However, they do not give us any information about the oscillation of the function. To compensate
this, we can use the classical Sobolev space W k,p where k is the degree of weak differentiability
and it keeps track of the oscillation(frequency). The same principle carries through when we study
holomorphic functions. Given f P HolpDq, we have should have no worry of the regularity of f since
it is analytic. To capture the size, we consider the p-th moment radial maximal function
ˆ ˙1{p
Mp pf q “ sup 2π ˇf preiθ qˇp dθ
1
şˇ ˇ
(4.1) , 0ăpă8
0ďră1 T
M8 pf q “ max |f pzq| “ max ˇf peiθ qˇ .
ˇ ˇ
(4.2)
zPD θPT

and define the Hardy space on the unit disc as

(4.3) H p pDq “ tf P HolpDq : Mp pf q ă 8u , 0 ă p ď 8.
I will suppress the domain for the rest of the paper when the context is clear. If 1 ď p ď 8, H p
is a Banach space with norm }¨}H p “ Mp p¨q. When 0 ă p ă 1, H p is a complete metric space,
where the metric is again given by the maximal function raised to the p-th power. For expository
purpose, I will focus on the case when p “ 2 (many of the results carry through for p ‰ 2). H 2 is a
Hilbert space, with an inner product constructed in the usual way. To illustrate why Hardy space
is an appropriate choice for studying boundary behavior, I present the following
Theorem 4.1. If f P H 2 , then lim f preiθ q exists almost everywhere on T.
rÕ1

Proof. Expand f about the origin and we can write

8
ÿ
(4.4) f preiθ q “ an rn einθ .
n“0

Integrate over the circle and taking supremum, we have

8
ÿ 8
ÿ
2 2
(4.5) }f }H 2 “ sup |an | r2n “ |an | ă 8.
0ďră1 n“0 n“0
2
The sequence pan q then defines an L function on the circle, and this function is the almost every-
where limit. 
As a consequence, we can identify H 2 as a subspace of L2 pTq. We can denote this space by
2
H pTq. Another characterization by Fourier series is immediately in place.
Theorem 4.2. f P H 2 if and only if f is the Poisson integral of some f peiθ q P L2 pTq and fˆpnq “ 0
for all n ă 0.
 A VERY SHORT SURVEY ON BOUNDARY BEHAVIOR OF HARMONIC FUNCTIONS 7

Proof. One direction is clear. Suppose f P H 2 , then let f peiθ q be the radial limit as in Theo-
rem 4.1. Convolve it against the Poisson kernel as in (2.4) produces a holomorphic function, so its
power series expansion does not involve z̄, and hence f peiθ q cannot have negatively-indexed Fourier
coefficient. 

4.2. Dirichlet space. Hardy spaces, as we have seen above, only characterize the boundary be-
havior of holomorphic function up to size. To study smoothness, one candidate class is given by
the Dirichlet space. To define such class, first we consider the sesquilinear form
ż
(4.6) xf, gy1D “ f 1 pzqg 1 pzqdApzq for f, g P HolpDq
D
where dApzq “ π ´1 dxdy is the normalized area measure on the disc. The form gives rise to a
seminorm r¨sD in the natural way. Then the Dirichlet inner product can be defined to be
(4.7) xf, gyD “ xf, gy1D ` xf, gyH 2 pTq .
This inner product gives rise to a norm. The space D consists of all the holomorphic functions with
finite norm. Observe that if f P D, then f P H 2 pTq, so a Fourier characterization of the norm is
immediately in place, since
8 ˇ2 8 ˇ
ˇ ˆ ˇ2
ÿ ˇ ÿ ˇ
(4.8) n ˇfˆpnqˇ , }f }H 2 pTq “ ˇf pnqˇ
ˇ ˇ
rf sD “
n“1 n“0

and therefore, }¨}D “ H 1{2 pTq X HolpDq, where H 1{2 is the Sobolev space of square integrable
functions with 1{2-moment integrable “derivatives” (of course, characterized by Fourier series).
This new characterization exemplifies the phenomenon of loosing fractional derivatives when we
take the trace of Sobolev functions.
As an important remark, the Sobolev regularity exponent 1{2 is the borderline case for one
dimension, not enough for the boundary value to be continuous. However, the space H 1{2 pTq does
embed into BM OpTq, the space of functions with bounded mean oscillation. This class of functions
has been extensively studied in modern harmonic analysis. To further investigate along the line,
we need heavier machinery from some deep theories which ventures too far from either complex
analysis or the intended scope of this project2. As an example, we can exploit more of the nice
properties of Dirichlet spaces by considering how we can continuously embed D into L2 pD, µq, where
µ is a positive Borel measure. Of course, this is a criterion on the measure µ. As it turns out,
the class of measures such that this embedding is continuous (called Carleson measure) leads to
various important results. For instance, their weak-type estimate is comparable to the Lebesgue
weak-type estimate on the boundary. For more on BMO, Carleson measure, and other modern
aspects of harmonic analysis, see [2] and [7].

References
1. Q. Han, F. Lin, Elliptic Partial Differential Equations, American Mathematical Society, 2011
2. J. Garnett, Bounded Analytic Functions, 1981
3. J. Garnett, D. Marshall, Harmonic Measure, Cambridge University Press, 2005.
4. W. Rudin, Real And Complex Analysis
5. L. Carleson, On the connection between Hausdorff measure and Capacity, ARKIV FOR MATEMATIK, 1956
6. L. Carleson, T. Gamelin, Complex Dynamics, 1993
7. L. Grafakos, Modern Fourier Analysis, 2014

2In plain language, they have either nothing to do with the course, or I don’t understand anything about them.
8 BLACK

Department of Mathematics, University of California, Davis, CA 95616

E-mail address: fsjiang@math.ucdavis.edu