Anda di halaman 1dari 62

Introduction to Hardy spaces

1
CHAPTER 1

Preliminaries

1. Sequences and families of holomorphic functions

Given any sequence (fn ) of holomorphic functions in the open set U ⊂


C which converges uniformly on every compact subset of U to a function
f : U 7→ C, it follows that f is also holomorphic on U . We shall
list some basic results about uniform convergence on compacts. Their
proofs and additional information is defered to Appendix A.
Theorem 1.1. Let (fn ) be a sequence of holomorphic functions on the
open connected set U ⊂ C that converges uniformly on compacts to the
(holomorphic) function f . Let V be an open subset of U such that there
exists n0 ≥ 1 with fn (z) 6= 0 for all z ∈ V and n ≥ n0 . Then either f
is the constant function 0, or f has no zeros in V .
Definition 1.1. A family F of holomorphic functions on the open
set U ⊂ C is called normal if every sequence (fn ) in F contains a
subsequence (fnk ) that converges uniformly on compact subsets to some
(holomorphic) function f (not necessarily in F!)
Theorem 1.2. (Montel). A family F of holomorphic functions on the
open set U ⊂ C is normal if it is uniformly bounded on compacts, that
is, for every compact set K ⊂ U there exists a positive constant CK
such that for all f ∈ F and all z ∈ K we have
|f (z)| ≤ CK .

For uniform convergence on compacts we have the following useful cri-


terion.
Theorem 1.3. (Vitali) Let (fn ) be a sequence of holomorphic functions
on the open connected set U ⊂ C. Suppose that {fn , n = 1, 2, . . .} is
uniformly bounded on compacts and also that there is a nondiscrete
subset A of U such that (fn ) converges pointwise on A. Then (fn )
converges uniformly on compacts.

We want to apply this theorem to infinite products of bounded analytic


functions.
3
4 1. PRELIMINARIES

Proposition 1.1. Let (fn ) be a sequence of holomorphic functions on


the open set U ⊂ C with the property that |fn (z)| ≤ 1 for all z ∈ U and
all n ∈ N. Let (Fn ) be defined by
n
Y
Fn (z) = fk (z), z∈U
k=1

and assume that there is z0 ∈ U such that (Fn (z0 )) has a nonzero limit.
Then (Fn ) converges uniformly on compacts to a nonzero holomorphic
function F on U and each zero of F is the zero of some fn , n ∈ N.

Proof. Note that the family {Fn , n ≥ 1} is uniformly bounded


in U by 1, hence, it is a normal family by Montel’s theorem. As in
the proof of Vitali’s theorem it suffices to show that if (Fnk ) and (Fn0k )
subsequences that converge uniformly on compacts to G and H respec-
tively, then G = H. Since |Fn | ≥ |Fn+1 |, we can choose a subsequence
(n00k ) of (nk ) such that n00k ≥ n0k which implies
|G(z)| = lim |Fn00k (z)| ≤ lim |Fn0k (z)| = |H(z)|.
k→∞ k→∞

Analogously, one shows that |H| ≤ |G| and we obtain that |G| = |H|.
This implies that G = αH for some α ∈ C with |α| = 1. But, by
hypothesis, we have also G(z0 ) = H(z0 ) 6= 0 which implies α = 1
and G = H. To see the assertion about the zeros, note first that if
K ⊂ U is compact then K ∩ (∪n fn−1 ({0})) is finite, otherwise F would
be identically zero. Then if z is a zero of F that is not a zero of any fn
there is an open neighborhood of z that contains no zeros of fn , n ≥ 1
and this leads to a contradiction by Theorem 1.1. 

Let us denote from now on by D the (open) unit disc (i.e. the disc
centered at the origin and of radius one). Our next result will be fre-
quently used in what follows and is a direct consequence of Proposition
1.4 combined with the following simple but important identity:
If z, w ∈ C with wz 6= 1 then
z − w 2 (1 − |z|2 )(1 − |w|2 )

(1.2) 1− = .
1 − wz |1 − wz|2
Indeed, this can be verified by a direct calculation (exercise!).
P
Corollary 1.1. Let (an ) be a sequence in D\{0}. If n (1−|an |) = ∞
then the infinite product

Y −an z − an
B(z) =
n=1
|an | 1 − an z
1. SEQUENCES AND FAMILIES OF HOLOMORPHIC FUNCTIONS 5
P
converges uniformly to zero on compact subsets of D. If n (1−|an |) <
∞ then the product converges uniformly on compacts to a nonzero holo-
morphic function in D. Moreover, in this case, the zeros of B are pre-
cisely the points an , n ≥ 1 and for fixed m, the multiplicity of the zero
am equals the number of occurences of am in the sequence.
P
Proof. Assume that n (1 − |an |)
= ∞ and let w ∈ D. Note that
N N
! N  !
Y |w − an |2 X |w − an |2 X |w − an |2
2
= exp log 2
≤ exp 2
−1 .
n=1
|1 − a n w| n=1
|1 − a n w| n=1
|1 − a n w|

By (1.2) we have
|w − an |2 (1 − |w|2 )(1 − |an |2 ) (1 − |w|2 )(1 − |an |2 )
1− = ≥
|1 − an w|2 |1 − an w|2 (1 + |w|)2
which implies that
N  N
|w − an |2 1 − |w|2 X
X 
2
−1 ≤− 2
(1 − |an |2 ) → −∞
n=1
|1 − an w| (1 + |w|) n=1
when N → ∞, and hence, the product converges to zero uniformly on
compacts.
P
Now assume that n (1 − |an |) < ∞. From (1.2) we deduce that the
factors fn (z) = −a n z−an
|an | 1−an z
satisfy

(1 − |z|2 )(1 − |an |2 )


1 − |fn (z)|2 = >0
|1 − an z|2
for all z ∈ D, that is |fn (z)| < 1, z ∈ D. If the above
Qn sum converges,
it is a standard matter to conclude that limn→∞ k=1 |ak | exists and
is nonzero. Thus, the two assumptions in Proposition 1.1 are satisfied
with z0 = 0 and the convergence of the product follows. The statement
about the zeros of B is also an immediate consequence of Proposition
1.1. 

The functions considered in Corollary 1.1 are called Blaschke products


and they play an important role in the theory of analytic functions in
the unit disc. If we want to allow Blaschke products to have zeros at
the origin as well (and we should) then the general definition reads as
follows: A Blaschke product is a holomorphic function in D of the form

Y
m−an z − an
(1.3) B(z) = z ,
n=1
|a n | 1 − a n z
6 1. PRELIMINARIES
P
where m is a nonnegative integer, (an ) is a sequence in D with n (1 −
|an |) < ∞.

Exercise 1. Show that the following families of holomorphic functions


are normal.
(i) F1 = {fn , fn (z) = z n , |z| < 1, n ∈ N}.
(ii) The family F2 of all power series f (z) = n≥0 an z n with |an | ≤
P
4, n ≥ 0.
(iii) The family F3 of all holomorphic functions f in the open set U ⊂ C
with the property that
Z
|f (x + iy)|dxdy < 5.
U

Exercise 2. Show that if F is a normal family of analytic functions in


the open set U then the families F (n) = {f (n) , f ∈ F} are also normal
for all n ∈ N.

Exercise 3. Prove the converse of Montel’s theorem.

Exercise 4. Let f be analytic in D and continuous in D such that


|f (z)| ≤ 147 for all z ∈ D. Show that if a1 , a2 , . . . , an are zeros of f
and B is the (finite) Blaschke product with zeros ak , 1 ≤ k ≤ n, then
|f (z)| ≤ 147|B(z)| for all z ∈ D.

2. Boundary behavior of power series

This section is a primary on the subject. It is intended to provide some


intuition and motivation for the results that follow. For this reason, it
will contain a number (four) unproved results that are important for
our discussion.
Given a power series with a positive radius of convergence (in other
words, an analytic function in a disc) we would like to investigate its
behavior near the boundary points of the disc of convergence. As it
is well known from calculus in 2 variables, our observations may de-
pend heavily on the way we approach such points and this does indeed
happen as the following example shows.

Example 2.1. Let f (z) = exp(− 1+z


1−z
), z ∈ D. Then:
2. BOUNDARY BEHAVIOR OF POWER SERIES 7

(i) f is bounded in D and


lim f (r) = 0,
r→1−

(ii) For every angle A with vertex at 1, symmetric w.r.t. the segment
[0, 1) and with opening less than π we have
lim f (z) = 0,
z→1
z∈A∩D

(iii) For every w ∈ D there exists a sequence (zn ) in D with zn → 1 and


lim f (zn ) = w.
n→∞

1+z
To prove these assertions, let g(z) = 1−z , z ∈ D. Then clearly, |g(z)| →
∞ if and only if z → 1. Let us show first that g(D) equals the right
half plane. Indeed, if g(z) = w and |z| < 1 then
1+z Re(1 + z)(1 − z) 1 − |z|2
Rew = Re = = > 0.
1−z |1 − z|2 |1 − z|2
Conversely, if Re w > 0, a similar computation shows that |z| < 1.
(i) Clearly, |f | = exp(−Re g) ≤ 1 if Re g > 0. The second assertion is
1+r
obvious since Re g(r) = 1−r → ∞ when r → 1.
(ii) Every z = reit ∈ A ∩ D can be written in the form z = 1 + ρeiθ ,
where ρ > 0 and |θ − π| ≤ α < π/2. Then
1 + ρeiθ
r = (1 + ρ2 + 2ρ cos θ)1/2 , eit =
(1 + ρ2 + 2ρ cos θ)1/2
and
eit − 1 1 + ρeiθ − (1 + ρ2 + 2ρ cos θ)1/2
= −
1 − r2 (1 + ρ2 + 2ρ cos θ)1/2 ρ(ρ + 2 cos θ)
Now the inequality satisfied by the angle θ implies that this quantity
stays bounded when ρ → 0, or equivalently, z → 1, z ∈ A ∩ D. This is
further equivalent to the fact that
|t|
(2.1) lim sup < ∞.
reit →1 1−r
reit ∈A∩D

With this fact at hand, the proof of (ii) is immediate. Indeed, from
above we have that
1 − r2 1 − r2 1 − r2
Reg(reit ) = = =
|1 − reit |2 1 + r2 − 2r cos t (1 − r)2 + 4r sin2 ( 2t )
8 1. PRELIMINARIES

and from (2.1) it follows that


limit inf Reg(reit ) = ∞.
re →1
reit ∈A∩D

(iii) Clearly, it suffices to prove the statement for w ∈ D. Write w = reiθ


and let wn = − log r + i(θ + 2nπ). Then Rewn > 0 and there exists
zn ∈ D with g(zn ) = wn and since wn → ∞ when n → ∞, it follows
that limn→∞ zn = 1. Obviously, f (zn ) = w for all n.

Motivated by the above example we introduce the following terminol-


ogy.
Definition 2.1. Let f be a complex-valued function defined on D and
let ζ ∈ ∂D.
(i) We say that f has the radial limit L at ζ if
lim f (rζ) = L.
r→1

(ii) We say that a sequence (zn ) in D converges nontangentially to


a point ζ ∈ ∂D if there is an angle with vertex at ζ, symmetric w.r.t.
the ray joining ζ to the origin and with opening less than π such that
zn belongs to this angle for all n sufficiently large.
(ii) We say that the function f has the nontangential limit L at the
point ζ ∈ ∂D if for every sequence (zn ) in D that converges nontangen-
tially to ζ we have
lim f (zn ) = L
n→∞

To avoid confusions about the sets involved in the definition of non-


tangential convergence one usually considers Stolz angles which are sets
obtained in the following way: Take a disc of radius 0 < σ < 1 centered
at the origin and denote by Γσ (ζ) the union of all line segments that
join a point in that disc and ζ. In other words, Γσ (ζ) is the convex hull
of the above disc and the point ζ. Then a sequence (zn ) in D converges
nontangentially to a point ζ ∈ ∂D if there exists 0 < σ < 1 such that
zn ∈ Γσ (ζ) for all n sufficiently large.
Remark 2.1. Recall that in general, 1 − |z| ≤ |z − ζ| if z ∈ D and ζ ∈
∂D and of course, if z → ζ it may happen that 1 − |z| goes much faster
to 0 than |z − ζ|. The point is that if z approaches ζ nontangentially
this cannot happen, i.e. the quotient (1 − |z|)/|z − ζ| stays bounded
below. This statement is equivalent to the one in the next exercise and
a proof was already pointed out in the previous example.
2. BOUNDARY BEHAVIOR OF POWER SERIES 9

Exercise 1. Prove in detail that a sequence (zn ) in D with zn = rn eiθn ,


converges nontangentially to a point ζ = eiθ ∈ ∂D if and only if |θ1−r
n −θ|
n
stays bounded when n → ∞.

Exercise 2. Show that the power series



X
f (z) = z n! , z∈D
n=0

has no radial limit at any point eit with t/π ∈ Q.

Throughout in what follows m will denote the normalized arclength-


measure on the unit circle T = ∂D, which can be identified with the
normalized Lebesgue measure on [0, 2π].
It turns out that the radial limits of a power series provide much less
information about the function inside the disc than the nontangential
limits. Thus, for example, it is known that there exist analytic func-
tions f in the unit disc that are not identically zero and yet satisfy
lim f (reit ) = 0 a.e. on [0, 2π]
r→1

(see [Col]). An even more dramatic example has been constructed by


Kahane and Katznelson [KK]. They show that there exist such ”bad”
functions f that satisfy in addition a growth restriction of the form
|f (z)| ≤ C(1 − |z|)−α
for α > 0 fixed but arbitrary!
On the other hand, this cannot happen with nontangential limits. This
is essentially the content of a famous theorem of Privalov that is stated
below without proof (for a proof see [Koo] ).
Theorem 2.1. (Privalov) If f is analytic in D and has the nontangen-
tial limit zero on a set of positive measure on the unit circle T, then f
vanishes identically.

Exercise 3. Given an analytic function f in D show that the set of


points t ∈ [0, 2π] with the property that f has a nontangential limit at
eit is measurable.
Example 2.2. If µ > 1 is fixed then the infinite product
Y∞
f (z) = (1 + µz n )
n=0
10 1. PRELIMINARIES

converges uniformly on compacts on D to a nonzero holomorphic func-


tion with the property that the set of points t ∈ [0, 2π] such that f has
a nontangential limit at eit has measure zero.
To prove the convergence note first that every compact subset of D is
contained in some disc centered at the origin with radius r < 1. Now
for such r and |z| < r we have |1 + µz n | < 1 + µrn ≤ exp(µrn ), which
implies that the functions
N
Y
fN (z) = (1 + µz n )
n=0

satisfy

!
X µ
|fN (z)| < exp µrn = exp( ).
n=0
1−r
Thus, the family {fN , N ≥ 1} is uniformly bounded on compacts
and hence normal, by Montel’s theorem. Moreover, exactly the same
argument as above shows that (fN (r)) converges for every 0 < r < 1
and by Vitali’s theorem the convergence assertion follows. Finally,
using the inequality 1 + x ≥ e−x we see that for 0 < r < 1

!
X µ
f (r) ≥ exp − µrn = exp(− ),
n=0
1 − r
i.e. f is not identically zero.
On the other hand, every point
(2k+1)πi
znk = µ−1/n e n , n, k ∈ N
n
is a zero of the function f because znk = −1/µ. We claim that for
every t ∈ [0, 2π] there exists a subsequence of (znk )n,k≥1 that converges
nontangentially to eit . Indeed, given n ≥ 1 we can find 0 ≤ kn < n
(2kn +1)πi
such that |t − 2knn+1 π| < πn . Then clearly, znkn = µ−1/n e n → eit
and since
|t − 2knn+1 π| π π
−1/n
< −1/n

1−µ n(1 − µ ) log µ
it follows by Exercise 1 that (znkn ) converges nontangentially to eit .
This shows that whenever f has a nontangential limit at some point eit
the limit must be zero. But by Privalov’s theorem the set of these points
must have measure zero (recall that the set in question is measurable!).
Example 2.3. There exist analytic functions f in D that have a radial
limit but no nontangential limit at almost every boundary point!
2. BOUNDARY BEHAVIOR OF POWER SERIES 11

Indeed, if we consider a nonconstant function f that has radial limits


zero almost everywhere on the boundary, the set of boundary points
where f has nontangential limits must have measure zero otherwise
this would contradict Privalov’s theorem.

It will turn out that such an erratic behavior near the boundary points
as described in the previous examples, becomes impossible if functions
in question satisfy an appropriate growth restriction. This is an impor-
tant point of view because, as we shall see in the sequel, these growth
restrictions are very easy to define and provide large classes (spaces) of
analytic functions that behave nicely near the boundary.
Theorem 2.2. Let f be analytic and bounded in D. If f has a radial
limit at a boundary point eit then f has a nontangential limit at that
point and these limits coincide.

Proof. By a translation and a rotation it will be sufficient to prove


the following statement. For a > 0, 0 < b < π/2 let Γa,b be the set of
points z = reis ∈ C with 0 < r < a and |s| < b. If f is analytic and
bounded in Γa,b and limr→0 f (r) = L then for 0 < c < b
lim f (z) = L.
z→0
z∈Γa,c

To this end, consider the sequence (fn ) with fn (z) = f ( nz ), z ∈ Γa,b and
apply Vitali’s theorem. The family {fn , n ≥ 1} is uniformly bounded
in Γa,b and if 0 < r < a then
r
lim fn (r) = lim f ( ) = L.
n→∞ n→∞ n
Since the set (0, a) is not discrete in Γa,b Vitali’s theorem implies that
(fn ) converges uniformly on compacts to the constant function L. Now
let (zk ) be a sequence in Γa,c , 0 < c < b that converges to 0. We
want to show that f (zk ) → L. Choose a sequence (nk ) of integers such
that nk |zk | ∈ (a/2, a) for sufficiently large k (for example, the choice
nk + 1 = integer part of |zak | will do). Then the points wk = nk zk satisfy
a/2 < |wk | < a, | arg wk | < c for sufficiently large k which shows that
these points wk lie in a compact subset of Γa,b . Then by the above
reasoning we have
wk
L = lim fnk (wk ) = lim f ( ) = lim f (zk )
k→∞ k→∞ nk k→∞

and the result follows. 


12 1. PRELIMINARIES

Exercise 3. Prove the following statement which provides an improve-


ment of Theorem 2.2: Let γ be an arc (continuous image of an interval)
with one endpoint eit and all other points contained in some Stolz angle
Γσ (eit ). If f is bounded and analytic in D and f (z) has the limit L as
z approaches eit along γ, then f has the nontangential limit L at eit .

The natural and basic question that arises is when do nontangential


limits of an analytic function exist? The following result is a classical
sufficient condition for the existence of such a limit due to Abel.
n
P
Theorem 2.3. (Abel) Let f (z) = n≥0 an z be a convergent
P powern
series in D. Assume that ζ ∈ ∂D is such that the series n≥0 an ζ
converges and let L be the value of the sum. Then f has the nontan-
gential limit L at ζ.

P∞ We may assume without loss of generality that ζ = 1. Let


Proof.
Sn = k=n an and note that by hypothesis S0 = L and Sn → 0, n →
∞. In particular, (Sn ) is bounded, say |Sn | ≤ M for all n. Now write

X ∞
X
f (z) = an z n = (Sn − Sn+1 )z n .
n=0 n=0
P∞
Since |Sn | ≤ M it follows that n=0 Sn z n converges for all z ∈ D and
thus,

X X∞ ∞
X X∞
n n n
f (z) = (Sn −Sn+1 )z = Sn z − Sn+1 z = S0 + Sn (z n −z n−1 ).
n=0 n=0 n=0 n=1

This trick is frequently called summation by parts. Let ε > 0 and n0


be such that |Sn | < ε for n ≥ n0 . Recall also that S0 = L. Then from
the last equality we get
X∞ X n0 X∞
n−1 n−1
|f (z)−L| ≤ |z−1| |Sn ||z | ≤ M |z−1| |z| +ε|z−1| |z|n−1
n=1 n=1 n=n0 +1

≤ M n0 |z − 1| + ε|z − 1|(1 − |z|)−1 .


Now let (zk ) be a sequence in D that converges nontangentially to 1.
Recall that in this case there is a constant K > 0 such that |zk − 1|(1 −
|zk |)−1 < K and use the last estimate to obtain
lim sup |f (zk ) − L| ≤ εK,
k→∞

for every ε > 0, which finishes the proof. 


2. BOUNDARY BEHAVIOR OF POWER SERIES 13

Of course, at the first sight the condition in Abel’s theorem seems diffi-
cult to test and has little to do with a growth restriction. Nevertheless
there is a deep result in analysis that yields a global existence the-
orem for nontangential limits of power series with square summable
coefficients. The following famous theorem was proved by L. Carleson.
n
P
Theorem 2.4. (Carleson) Let f (z) = n≥0 an z be a power series
with square summable coefficients, that is,
X
|an |2 < ∞.
n≥0

an ζ n converges for almost every ζ ∈ ∂D.


P
Then the series n≥0

It is a simple exercise to show that such power series converge in the


unit disc. Thus, by Abel’s theorem every power series with square sum-
mable coefficients has a nontangential limit almost everywhere on the
unit circle. Let us now show that the condition on the coefficients really
is a growth restriction. This is a consequence of Parseval’s formula.
Theorem 2.5. Let f (z) = n≥0 an z n be a convergent power series in
P
D. Then
Z 2π Z 2π
it 2 dt dt
X
2
|an | = sup |f (re )| = lim |f (reit )|2 .
n≥0
0<r<1 0 2π r→1 0 2π

Proof. Write

! ∞ ! ∞
X X X
it 2 n int n −int
|f (re )| = an r e an r e = an am rm+n ei(n−m)t
n=0 n=0 m,n=0

and note that the last sum converges uniformly on [0, 2π]. But then
we can interchange limit and integration to obtain
Z 2π ∞ Z 2π ∞
it 2 dt dt
X X
m+n
|f (re )| = an am r ei(n−m)t = |an |2 r2n ,
0 2π m,n=0 0 2π n=0
R 2π ikt dt
because 0 e 2π = δk0 , that is, it equals 0 if k 6= 0 and 1 if k = 0.
The rest of the proof is a routine exercise based on the fact that the
integrals involved here increase with r as the last identity shows. 

Thus, we can reformulate the result on nontangential limits in terms


of the growth restriction we just found.
Corollary 2.1. Let f be analytic in D and suppose that
Z 2π
dt
sup |f (reit )|2 < ∞.
0<r<1 0 2π
14 1. PRELIMINARIES

Then f has nontangential limits a.e. on the unit circle.

The above result is classical, and, as we shall see in the sequel, its proof
does not need the ”heavy artillery” provided by Carleson’s theorem.
We close this section with a result which shows that, in terms of Taylor
coefficients, the square summability condition in Corollary 2.6 cannot
be improved. The theorem below is a special case of a much stronger
result of Khinchin and Kolmogorov (see [Du],p.).
Theorem 2.6. Let f (z) = n≥0 an z n be a convergent power series in
P
D such that X
|an |2 = ∞.
n≥0
Then there exists a sequence (εn ) with εn = ±1, n = 0, 1, 2 . . . such
that the power series X
g(z) = ε n an z n
n≥0
has no radial limit almost everywhere on the unit circle.
CHAPTER 2

Poisson integrals

1. Harmonic functions

Recall that a twice differentiable complex-valued function u defined


on some open set G ⊂ C is called harmonic if it satisfies the Laplace
equation
∂ 2u ∂ 2u
(1.1) ∆u = + = 0 in G.
∂x2 ∂y 2
Of course, holomorphic functions are harmonic. Moreover, a very im-
portant class of examples of real-valued harmonic functions are the
real (and imaginary) parts of holomorphic functions in G. This follows
from a very simple computation with the Cauchy-Riemann equations.
Conversely, if G is simply connected (roughly speaking, this means it
has no holes), it turns out that every real-valued harmonic function in
G is the real part of a holomorphic function in G. This is a classical re-
sult that needs to be proved. In order to avoid technicalities regarding
the definition of a simply connected domain we will use an equivalent
property.

1.1. Theorem Suppose that G is an open connected set in C with


the property that every holomorphic function on G has a primitive in
G. Then every real-valued harmonic function in G is the real part of a
holomorphic function in this open set.

Proof. Observe first that if u is harmonic on G then the function


h = ∂u∂x
− i ∂u
∂y
satisfies the Cauchy-Riemann equations in G. Indeed,
since u is harmonic we have
∂ 2u ∂ 2u
 
∂ ∂u ∂ ∂u
= =− 2 = − ,
∂x ∂x ∂x2 ∂y ∂y ∂y
and
∂ 2u
 
∂ ∂u ∂ ∂u
= =− − .
∂y ∂x ∂xy ∂x ∂y
15
16 2. POISSON INTEGRALS

Let f be a primitive of the holomorphic function h, i.e. f 0 = h. Then

∂ ∂u
Ref = Ref 0 =
∂x ∂x
and
∂ ∂u
Ref = −Imf 0 = .
∂y ∂y
Thus, Ref − u is constant in G and the result follows.

The simplest example that comes to mind of a domain G as above, is


a disc.
The theorem has the following reformulation: If G is as above and u is
harmonic in G then there exists a harmonic function v in G such that
u + iv is analytic in G.

Exercise 1. Show that v is harmonic and it is uniquely determined up


to an additive constant (i.e. if v1 , v2 have this property then v1 − v2 is
constant in G).

Such a function v is called a harmonic conjugate of u, while the asso-


ciation u 7→ f = u + iv is sometimes called analytic completion.

Exercise 2 Given function of the form

N
X
u(z) = akn z k z n , z∈C
k,n=0

show that u is harmonic if and only if akn = 0 whenever k and n are


both nonzero.

2. Representation by Poisson integrals

Since harmonic functions are so close to the analytic ones it is natural


to try to find an analogue of Cauchy’s formula for these functions. Here
is a first attempt for harmonic functions in the unit disc.
If u is harmonic and real-valued in a disc containing D and f is an
analytic completion of u then by Cauchy’s formula we have for all
2. REPRESENTATION BY POISSON INTEGRALS 17

z∈D
 Z
1
f (ζ)dζ
(2.1) u(z) = Rf (z) = R
|ζ|=1 ζ − z
2πi
 Z 2π
f (eit )eit dt

1
=R .
2π 0 eit − z
This is an integral representation of u which, unfortunately, involves
one of its harmonic conjugates and therefore needs to be improved.
Lemma 2.1. If u is harmonic in a disc containing D then for all z ∈ D
we have Z 2π
1 1 − |z|2
u(z) = it 2
u(eit )dt.
2π 0 |e − z|
Proof. The trick is to insert in the last integral in (2.1) the harm-
less term zf (eit )/(e−it −z), where f is again an analytic function whose
real part equals u. Note that for every integer n ≥ 0 we have
Z 2π int Z 2π i(n+1)t
ζ n dζ
Z
e dt e dt
= = −i = 0,
0 e−it − z 0 1 − eit z |ζ|=1 1 − ζz

by Cauchy’s Theorem. Thus, if we write



X
f (eit ) = an eint
n=0
then from the fact that the power series converges uniformly on the
unit circle we obtain
Z 2π ∞ Z 2π int
f (eit )zdt X e zdt
−it
= an = 0,
0 e −z n=0 0 e−it − z
which shows that the term inserted in the last integral in (2.1) is indeed
harmless. Therefore, from (2.1) we obtain


eit
 Z   
1 it z
u(z) = R f (e ) it
+ −it dt .
2π 0 e −z e −z
Now note also that
eit z 1 − |z|2
(2.2) + = , z 6= eit ,
eit − z e−it − z |eit − z|2
and that the right hand side of this equality is always real (it becomes
also positive if |z| < 1). Then
Z 2π
1 1 − |z|2
u(z) = Rf (eit )dt
2π 0 |eit − z|2
18 2. POISSON INTEGRALS

and the result follows. 

Observe that this lemma gives a much better integral representation


of a harmonic function. The only problem that remains is to weaken
the very restrictive hypothesis that u is actually harmonic in a larger
disc. A first step in this direction is an immediate application of the
theorem. is given in the next theorem.
Corollary 2.1. Let u be harmonic in D and continuous in the closed
disc D. Then Z 2π
1 1 − |z|2
u(z) = u(eit )dt.
2π 0 |eit − z|2

Proof. For 0 < r < 1 fixed but arbitrary, consider the dilation ur
of u defined in D by ur (z) = u(rz). Clearly, ur is harmonic in the disc
of radius 1/r centered at the origin, so that Lemma 2.1 gives for every
z∈D Z 2π
1 1 − |z|2
ur (z) = ur (eit )dt.
2π 0 |eit − z|2
Now let r → 1 and note that ur (z) → u(z), z ∈ D and also, that the
functions t 7→ ur (eit ), t ∈ [0, 2π] converge uniformly to t 7→ u(eit ), t ∈
[0, 2π], by the continuity assumption. This implies that for fixed z ∈ D
Z 2π Z 2π
1 − |z|2 it 1 − |z|2
it 2
u r (e )dt → it 2
u(eit )dt, r → 1
0 |e − z| 0 |e − z|
and we are done. 

The continuity assumption on u can be relaxed using some results from


integration theory and functional analysis.
Theorem 2.1. Let u be harmonic in D and suppose that there exist
constants C > 0 and 1 ≤ p ≤ ∞ such that for all r ∈ (0, 1) we have
Z 2π
dt
|u(reit )|p ≤ C,
0 2π
if p < ∞, or
sup |u(reit )| ≤ C,
t∈[0,2π]

if p = ∞. Then:
(i) If p > 1, there exists a unique ũ ∈ Lp (m) such that
Z 2π
1 − |z|2 dt
u(z) = it 2
ũ(eit ) .
0 |e − z| 2π
2. REPRESENTATION BY POISSON INTEGRALS 19

(ii) If p = 1, there exists a unique finite Borel measure µ on T such


that
1 − |z|2
Z
u(z) = 2
dµ(ζ).
T |ζ − z|

Proof. We shall prove first the existence part. The argument


goes exactly as the one in Corollary 2.1. More precisely, we consider
the dilations ur , 0 < r < 1 of u, apply Lemma 2.1 and let r → 1.
Clearly, for fixed z ∈ D ur (z) → u(z), when r → 1.
(i) Assume first that p < ∞, and use the assumption in the statement
together with the fact that Lp (m) is reflexive, to find a sequence (rn )
tending to 1 such that (urn ) converges weakly in Lp (m) to ũ ∈ Lp (m).
Since the functions
1 − |z|2
Pz (eit ) = it
|e − z|2
are bounded in t, we obtain the representation in the statement. When
p = ∞ we only use weak-star sequential compactness instead of reflex-
ivity to arrive at the same result. (ii) Here we use the fact that for
fixed z ∈ D the function Pz (·) is continuous on T and since the mea-
sures ur dm have bounded total variation, we can find a sequence that
converges weak-star in the dual of C(T) to some finite Borel measure
µ on T.
It remains to show that ũ, µ are unique. A direct computation shows
that

ze−it X
Pz (eit ) = 1 + 2R = 1 + 2R ∞z n e−int ,
1 − ze−it n=0

where the series converges uniformly on [0, 2π]. Since the equality in
the statement holds for all z ∈ D we see that the Fourier coefficients of
ũ, µ are uniquely determined, i.e. ũ, µ are unique. 

The kernel
it 1 − |z|2
Pz (e ) = it
|e − z|2
is called the Poisson kernel, and a harmonic function of the form de-
scribed in the above theorem is called the Poisson integral of ũ, or µ.
The simplest example that (i) cannot hold for p = 1 are the Poisson
kernels themselves. For example, Pz (1) satisfies the theorem with p = 1
and is the Poisson integral of the Dirac measure at 1 which is uniquely
determined by this assertion.
20 2. POISSON INTEGRALS

Corollary 2.2. Let u be harmonic and nonnegative in D. Then there


exists a unique nonnegative Borel measure µ on T such that µ(T) =
u(0) and
1 − |z|2
Z
u(z) = 2
dµ(ζ).
T |ζ − z|

Proof. A nonnegative harmonic function u in D satisfies the as-


sumption in part (ii) of Theorem 2.1 since
Z 2π Z 2π
it dt dt
|u(re )| = u(reit ) = u(0).
0 2π 0 2π
Since the measure µ goiven by the theorem is the weak-star limit of
nonnegative measures, it will be nonnegative as well. 
Corollary 2.3. Let f be analytic with nonnegative real part in D.
Then there exists a unique nonnegative Borel measure µ on T such
that µ(T) = Rf (0) and
Z
ζ +z
f (z) = dµ(ζ).
T ζ −z

The above result is called the Herglotz representation of functions with


positive real part.

Exercise 1 Prove Corollary 2.3.

Exercise 2 Prove that a harmonic function in an open connected sub-


set G of C which has a local minimum, or a local maximum in G is
constant.

3. The nontangential maximal function

Given a function f : D → C and σ ∈ (0, 1), the nontangential maximal


function of f is defined on T by
(3.3) f ∗ (ζ) = sup |f (z)|.
z∈Γσ (ζ)

The definition depends on σ, but is is customary to drop this parameter


in the notation. In all assertions σ ∈ (0, 1) is fixed but arbitrary, and
the constants apearing in estimates will depend on the choice of σ. The
aim of this section is to compare the nontangential maximal function of
3. THE NONTANGENTIAL MAXIMAL FUNCTION 21

Poisson integrals of functions h ∈ L1 (m) with the well-known Hardy-


Littlewood maximal function of h, defined by
1 θ+t
Z
† iθ
(3.4) h (e ) = sup |h(eis )ds.
t∈[0,π] 2t θ−t

The Hardy-Littlewood maximal function plays a crucial role in Lp -


estimates as it is shown by the following classical theorem essentially
due to Hardy and Littlewood, which we state without proof.
Theorem 3.1. For p > 1, h† ∈ Lp (m), whenever h ∈ Lp (m), and there
exists Cp > 0 such that
kh† kp ≤ Cp khkp .
If h ∈ L1 (m) there exists C1 > 0 such that for all t > 0 we have
khk1
m({ζ ∈ T : h† (ζ) > t}) < C1 .
t
We begin with some simple properties of the Poisson kernel which are
listed below.
Proposition 3.1. (i) Pz (eit ) > 0, for all z ∈ D and t ∈ [0, 2π].
(ii) For all t, θ ∈ [0, 2π) with t 6= θ
lim Pz (eit ) = 0.
z→eiθ

In fact, if we regard Pz (·) as a family of functions on T then the above


convergence is uniform on any set of the form {ζ ∈ T : |ζ − eit | > δ},
where δ > 0 is fixed.
(iii) For all z ∈ D, s ∈ [0, 2π], , 0 < r < 1, we have
Z 2π Z 2π
it dt dθ
Pz (e ) = Preiθ (eis ) = 1.
0 2π 0 2π
(iv) If z = reiθ then Pz (ei(θ+t) ) = Pz (ei(θ−t) ).
(v) If ζ ∈ T and z = reit ∈ Γσ (ζ), there exists C > 0 depending only
on σ such that for all λ ∈ T we have
Pz (λ) ≤ CPrζ (λ).
(vi) If r ∈ [0, 1), t ∈ [−π, 2π] then

t Pr (eit ) ≤ 0,
∂t
and consequently, there exists C > 0 such that
Z π

t Pr (e ) dt ≤ C,
it

∂t 2π
−π
for all r ∈ [0, 1).
22 2. POISSON INTEGRALS

Proof. (i), (ii), and (iv) are obvious. The first integral in (iii)
equals 1 by an application of Lemma 2.1u ≡ 1, the second equality
follows from Preiθ (t) = Preit (θ). To see (v), write
1 − r2 1 − r2 |λ − rζ|2 |λ − rζ|2
Pz (λ) = = = P rζ (λ) .
|λ − z|2 |λ − rζ|2 |λ − z|2 |λ − z|2
Moreover,
|λ − rζ| |λ − z| + |z − rz| + r|z − ζ| |z − ζ|
≤ ≤1+r+r ,
|λ − z| λ − z| 1−r
and the right hand side is bounded within Γσ (ζ) by a constant that
depends only on σ. Finally, the first part of (vi) is immediate since
∂ 2r(1 − r2 )t sin t
t Pt (eit ) = − ,
∂t |eit − r|2
while the second part follows from this together with integration by
parts:
Z π Z π

t Pr (e ) dt = − ∂ dt
it

∂t 2π t Pr (eit )
−π −π ∂t 2π
Z π
dt
= −Pr (−1) + Pr (eit )
−π 2π
1−r
=− + 1.
1+r


We can now turn to the main result of the section.


Theorem 3.2. There exists C > 0 such that, given h ∈ L1 (T) with
Poisson integral u, we have
u∗ (ζ) ≤ Ch† (ζ), ζ ∈ T.
In particular, if h ∈ Lp (m) for some p ∈ (1, ∞) then u∗ ∈ Lp (m), and
there exists Cp > 0, depending only on p, such that
ku∗ kLp (m) ≤ Cp khkLp (m)

Proof. Let z ∈ Γσ (eiθ ), with |z| = r ∈ [0, 1), and use Proposition
3.1 (v) to obtain the estimate
Z Z
(3.5) |u(z)| ≤ Pz (λ)|h(λ)|dm(λ) ≤ C Preiθ (λ)|h(λ)|dm(λ).
T T
4. BOUNDARY BEHAVIOR OF POISSON INTEGRALS 23

Using again this result we can write


Z Z π
dt
Preiθ (λ)|h(λ)|dm(λ) = Pr (eit )|h((ei(θ+t) )|

T
Z−π π
dt
= Pr (eit )[|h((ei(θ+t) )| − |h((ei(θ−t) )|] .
0 2π
Finally, integration by parts gives
Z Z π
dt
Preiθ (λ)|h(λ)dm(λ) = Pr (−1) |h((ei(θ+t) )|
T −π 2π
Z π Z t
∂ ds
− Pr (eit ) |h((ei(θ+s) )| .
0 ∂t −t 2π
Clearly,
Z π Z t
dt ds t
i(θ+t)
|h((e )| ≤ h† (eiθ ), |h((ei(θ+s) )| ≤ h† (eiθ ).
−π 2π −t 2π π
By Proposition 3.1 (vi) we have that
Z π Z t
∂ ds
− it
Pr (e ) |h((ei(θ+s) )| ≤ C 0 h† (eiθ ),
0 ∂t −t 2π
hence, Z
Preiθ (λ)|h(λ)|dm(λ) ≤ C 00 h† (eiθ ),
T
and the result follows by the estimate (3.5). 

4. Boundary behavior of Poisson integrals

In this section we will use the maximal function estimate to derive a


series of results about Poisson integrals of integrable functions.
Theorem 4.1. If u is the Poisson integral of h ∈ L1 (m), then u has
the nontangential limit h(ζ) at almost every ζ ∈ T.

Proof. Let ε > 0, and let g ∈ C(T) with kh − gk1 < ε. By the
Hardy-Littlewood theorem it follows that
√ √
m({ζ ∈ T : (h − g)† (ζ) > ε}) < C1 ε.
Let v denote the Poisson integral of g. By Theorem 3.2 we have that
√ √
m({ζ ∈ T : (u − v)∗ (ζ) > C ε}) < C1 ε,
where C > 0 is the constant given in that theorem. For a fixed integer
n > 0 consider the set An ⊂ T consisting of those points ζ for which
the nontangential ”limsup” of |u − h(ζ)| at ζ exceeds n1 . Since v is
24 2. POISSON INTEGRALS

continuous in D and equals g on T, it follows that if ε is sufficiently


small, An is contained in the set
√ √
{ζ ∈ T : (u − v)∗ (ζ) > C ε} ∪ {ζ ∈ T : |h(ζ) − g(ζ)| > ε}.

This gives that m(An ) < C1 ε, and since ε was arbitrary we conclude
that m(An ) = 0, which completes the proof. 

The result can be extended to Poisson integrals of finite Borel measures,


but this requires an additional step.
Lemma 4.1. If u is the Poisson integral of a finite Borel measure µ
on T which is singular w.r.t. m, then u has nontangential limits zero
m-a.e.

Proof. We may assume without loss of generality that the mea-


sure µ is positive. The the existence of the nontangential limits folows
from Theorem 4.1. Indeed, if f is analytic in D with Rf = u, then e−f
is analytic in and bounded D, in particular it is a bounded harmonic
function. Then it is the Poisson integral of a function in L∞ (m), hence
by the above result, it has nontangential limits a.e. on T. This implies
that u = − log |f | has nontangential limits u(ζ) ≥ 0 for almost every
ζ ∈ T. Now recall that our measure µ is the weak-star limit of the
measures ur dm, where ur (z) = u(rz), i.e. for every continous function
h on T we have that
Z Z
lim− hur dm = hdµ.
r→1 T
On the other hand, by Fatou’s lemma we obtain for every nonnegative
continuous function h on T,
Z Z Z
u(ζ)h(ζ)dm(ζ) ≤ lim inf

hur dm = hdµ.
T r→1 T
Since the characteristic function of any compact subset of T can be
approximated pointwise by a sequence of nonnegative continuous func-
tions, another application of Fatou’s lemma leads to
Z
udm ≤ µ(A),
A
for all compact subsets of T. From the fact that µ is singular we deduce
that u = 0, m-a.e. 

If u is the Poisson integral of the finite Borel measure µ on T we can


write
dµ = hdm + dν,
5. hp -SPACES 25


where h = dm ∈ L1 (m) is the Radon-Nykodim derivative of µ w.r.t.
m, and ν is singular w.r.t. m. By Theorem 4.1 and the above lemma
we obtain:
Corollary 4.1. If u is the Poisson integral of the finite Borel measure

µ on T then u has the nontangential limit dm (ζ) at almost every ζ ∈ T.

5. hp -spaces

Given 0 < p < ∞ we denote by hp the vector space of all harmonic


functions u in D with
Z 2π
p dt
(5.6) kukp = sup |u(reit )|p < ∞.
0<r<1 0 2π
h∞ will denote the space of bounded harmonic functions u in D with
kuk∞ = sup |u(z)|.
z∈D

We will be interested in the case p ≥ 1, when (5.6) defines a norm on


hp .
If p > 1, we know from the previous section that every u ∈ hp is the
Poisson integral of its boundary function defined a.e. on T by u(ζ) =
nontangential limit of u at ζ. We shall keep this notation for the
boundary values throughout in what follows. If u ∈ h1 we only know
that u is a Poisson integral of a (unique) finite Borel measure on T,
which we denote by µu . Its total variation is kµu k = |µu |(T).
Theorem 5.1. (i) If 1 < p ≤ ∞ and u ∈ hp we have
Z
p
kukp = kukLp (m) = |u|p dm,
p
T

and if p = 1, then
kuk1 = kµu k.
For 1 ≤ p ≤ ∞, hp is a Banach space with the norm (5.6).
(iii) If 1 < p < ∞, and u ∈ hp , ur (z) = u(rz), 0 ≤ r < 1, then
lim kur − ukp = 0.
r→1−

Proof. (i) The case p = ∞ is immediate


Z Z
|u(z)| ≤ Pz (ζ)|u(ζ)dm(ζ) ≤ kuk∞ Pz (ζ)dm(ζ) = kuk∞ .
T T
26 2. POISSON INTEGRALS

If 1 < p < ∞, note that u(ζ) = limr→1− u(rζ), m−a.e. on T, hence


Fatou’s lemma gives Z
|u|p dm ≤ kukpp .
T
On the other hand, since u is the Poisson integral of its boundary
values, we can apply by Hölder’s inequality to obtain
Z p
p

|u(z)| = Pz (ζ)u(ζ)dm(ζ)

T
Z  Z p−1
p
≤ Pz (ζ)|u(ζ)| dm(ζ) Pz (ζ)dm(ζ)
T T
Z
= Pz (ζ)|u(ζ)|p dm(ζ),
T
so that
Z 2π Z 2π Z Z
it p dt p dt
|u(re )| ≤ Pz (ζ)|u(ζ)| dm(ζ) = |u|p dm.
0 2π 0 T 2π T
The case when p = 1 is similar. Since µu is the weak-star limit of a
sequence urn dm, it follows that kµu k ≤ kuk1 . On the other hand, by
Fubini’s theorem
Z 2π Z 2π Z
it dt dt
|u(re )| ≤ Pz (ζ)d|µ| (ζ)dm(ζ) = kµu k.
0 2π 0 T 2π
(ii) assume (un ) is a Cauchy sequence in hp , and use (i) to conclude
that the corresponding sequence of boundary functions, or measures is
Cauchy in Lp (m), respectively in the space of finite Borel measures on
T. Since these are complete spaces, we have that (un ) converges to u in
Lp (m), or (µun ) converges in norm to µ and again by(i), (un ) converges
in hp to the Poisson integral of u, or µ. (iii) Almost everywhere on T
we have
|ur (ζ) − u(ζ)| ≤ u∗ (ζ), lim− ur (ζ) = u(ζ),
r→1
and the result follows by Theorem 3.2 together with the dominated
convergence theorem. 

Exercise 1 Show that part (iii) of Theorem 5.1 fails for p = ∞ when-
ever u has a discontinuous boundary function, or when p = 1, and u is
the Poisson integral of a measure with a nonzero singular part.

In the above proof we have used the important estimate


Z
(5.7) |u(z)| ≤ Pz (ζ)|u(ζ)|p dm(ζ),
p
T
6. EMBEDDINGS OF hp INTO Lp (ν) 27

which has on obvious version in the case when p = 1


Z
(5.8) |u(z)| ≤ Pz (ζ)d|µu |(ζ).
T

I)n particular, these estimates show that evaluation at a fixed point


z ∈ D is a bounded linear functional on hp .
Corollary 5.1. If z ∈ D and u ∈ hp , 1 ≤ p < ∞, then
 1/p
1 + |z|
|u(z)| ≤ kukp .
1 − |z|
In particular, for 1 ≤ p ≤ ∞, convergence in hp implies uniform con-
vergence on compact subsets of D.

Proof. The result follows from the simple estimate


1 − |z|2 1 + |z|
Pz (ζ) ≤ 2
= .
(1 − |z|) 1 − |z|


6. Embeddings of hp into Lp (ν)

This is an important problem with far-reaching consequences. We want


to characterize the positive Borel measures ν on D with the property
that hp ⊂ Lp (ν). An application of the closed graph theorem shows
that if this inclusion holds, then the inclusion map must be bounded,
that is
hp ⊂ Lp (ν) ⇐⇒ kukp ≤ CkukLp (ν) ,
for some C > 0 and all u ∈ hp . Indeed, if (un ) converges to u in hp and
to v in Lp (ν), then by Corollary 5.1, for each z ∈ D, (un (z)) converges
to u(z). Also, there exists a subsequence (unk ) which converges ν-a.e.
to v, hence u = v ν−a.e.. For fixed λ ∈ D, let
(1 − |λ|2 )1/p
uλ (z) = , z ∈ D.
(1 − λz)2/p
uλ is analytic, hence harmonic in D and for ζ ∈ T we have |uλ (ζ)| =
Pλ (ζ). Then the above inequality yields the following necessary condi-
tion for the inclusion hp ⊂ Lp (ν)
Z
(6.9) sup(1 − |λ| ) |1 − λz|−2 dν(z) < ∞.
2
λ∈D
28 2. POISSON INTEGRALS

A similar but unfortunately stronger condition is sufficient and follows


immediately from the estimates (5.8) and (5.7), namely
1 − |z|2
Z
(6.10) sup dν(z) < ∞.
ζ∈T |1 − ζz|2
For example, (5.7) and Fubini’s theorem give
Z Z Z
p
|u| dν ≤ Pz (ζ)|u(ζ)|p dm(ζ)dν(z)
Z T
1 − |z|2
Z
p
= |u(ζ)| dν(z)
T |1 − ζz|2
1 − |z|2
Z
p
≤ kukp sup dν(z).
ζ∈T |1 − ζz|2

Exercise 1 Show that if ν is the one-dimensional Lebesgue measure on


the segment (−1, 1) then (6.9) holds, but (6.10) fails .

It turns out that the necessary condition above is also sufficient for our
embedding. Let us rewrite this condition in a more transparent way.
Proposition 6.1. Given 0 < h < 1, and θ ∈ [0, 2π], let
Sh (eiθ ) = {reit : 1 − h < r < 1, |t − θ| < h}.
A positive Borel measure ν on D satisfies (6.9) if and only if
ν(Sh (eiθ )) ν({|z − ζ| < r})
sup < ∞, or sup < ∞.
0<h<1 h ζ∈T r
θ∈[0,2π] r>0

Proof. We leave it as an exercise to show that if one supremum


above is finite, then so is the other. Fix λ ∈ D \ {0} and note that if
λ
|z − |λ| | < (1 − |λ|, then
λ
|1 − λz| ≤ |λ||z − | + 1 − |λ| < (|λ| + 1)(1 − |λ|),
|λ|
hence
λ
2
Z
−2
(1 − |λ|2 )ν({|z − |λ|
| < (1 − |λ|)})
(1 − |λ| ) |1 − λz| dν(z) >
4(1 − |λ|)2
λ
ν({|z − |λ| | < (1 − |λ|)})
≥ ,
4(1 − |λ|)
6. EMBEDDINGS OF hp INTO Lp (ν) 29

which shows that if (6.9) holds, then the second supremum above is
finite. Conversely, fix λ ∈ D with |λ| ≥ 1/2, and let
En = {z ∈ D : 2n (1 − |λ|2 ) ≤ |1 − λ| < 2n+1 (1 − |λ|2 )}.
Then Z X ν(En )
(1 − |λ| )2
|1 − λz|−2 dν(z) ≤ .
n
22n (1 − |λ|2 )
Moreover, if z ∈ En , then
1 λ
|z − | < |1 − λz| + 1 − |λ| < 2n+2 (1 − |λ|2 ),
2 |λ|
hence,
ν({|z − ζ| < r})
ν(En ) ≤ 2n+2 (1 − |λ|2 ) sup ,
ζ∈T r
r>0

which gives
ν({|z − ζ| < r})
Z
2
(1 − |λ| ) |1 − λz|−2 dν(z) ≤ C sup ,
ζ∈T r
r>0

when λ ∈ D with |λ| ≥ 1/2. Since the integrals on the left are obviously
bounded or |λ| < 1/2, the result follows. 

The sets Sh have been used by Carleson who was the first to study
embeddings of this type. They will be called Carleson boxes, and the
measures satisfying one of the conditions in Proposition 6.1 are called
Carleson measures. The next theorem is essentially due to Carleson.
Theorem 6.1. If ν is a finite positive Borel measure on D, and 1 <
p < ∞ then hp ⊂ Lp (ν) if and only if ν is a Carleson measure.

Proof. The necessity part has already been proved. Let u ∈ hp ,


kukp = 1. We claim that there exists k > 0 such that if t > 0 is
sufficiently large then
ν({z ∈ D : |u(z)| > t}) ≤ km({ζ ∈ T : u∗ (ζ) > t}).
If this is achieved, it follows that
Z Z ∞
p
|u| dν = p tp−1 ν({z ∈ D : |u(z)| > t})dt
0
Z ∞
≤ kp tp−1 m({ζ ∈ T : u∗ (ζ) > t})dt
Z 0
= k (u∗ )p dm,
T
30 2. POISSON INTEGRALS

hence, by Theorem 3.2, we have u ∈ Lp (ν).


To see the claim, let ε > 0, and let G be an open subset of T such that
{ζ ∈ T : u∗ (ζ) > t} ⊂ G, m(G) < m({ζ ∈ T : u∗ (ζ) > t}) + ε.
If z ∈ D \ {0} is such that |u(z)| > t then z lies outside any Stolz angle
Γσ (ζ) (here σ is fixed), with u∗ (ζ) ≤ t. In particular, |z|
z
∈ G. Recall
that G is a countable union of disjoint open arcs In , n ≥ 1 so that, if
z
At = {z ∈ D : |u(z)| > t}, At,n = {z ∈ D : |u(z)| > t, ∈ In },
|z|
we have [
At = At,n .
n≥1

Moreover, if z ∈ At,n , and ζn is the endpoint of In closest to z then


m(In )
|z − ζn | < + 1 − |z|.
2
Also, recall that z ∈ / Γσ (ζn ). By Corollary 5.1 it follows that if t is
sufficiently large, then |z| > σ, so there exists c > 1, depending only
on σ, with |z − ζ| > c(1 − |z|. By replacing this in the last estimate we
obtain
m(In )
1 − |z| < (c − 1), z ∈ At,n .
2
Then if m(G) is sufficiently small , which happens when t is large and
ε is small, we can conclude that
At,n ⊂ Shn (ζn0 ),
where ζn0 is the middlepoint of In , and hn = c0 m(In ), for some fixed
c0 > 0. Then by the assumption on ν we have
ν(At,n ≤ Cm(In ),
and
X
m(At ) = m(At,n ) ≤ C(m(G)) < m({ζ ∈ T : u∗ (ζ) > t}) + ε,
n≥1

which implies the claim and completes the proof of the theorem. 
Example 6.1. As pointed out in Exercise 1, the one-dimensional Lebesgue
measure on the segment (−1, 1) is a Carleson measure, that is, there
exists C > 0 such that
Z 1
|u(x)|p dx ≤ Ckukpp ,
−1
6. EMBEDDINGS OF hp INTO Lp (ν) 31

for all u ∈ hp , 1 < p ≤ ∞. If we consider only analytic functions


the best constant is C = 12 and the estimate is called the Fejér-Riesz
inequality .
Remark 6.1. It is important to note that for a finite positive Borel
meausre ν the quantities considered in Proposition 6.1 are essentially
related to the behavior of µ near the boundary. In fact, it is an im-
mediate consequence of Corollary 5.1 that ν is a Carleson measure if
and only if its restriction to any annulus {z : 0 < r < |z| < 1} is a
Carleson meausre.

An important class of examples of Carleson measures are those of the


form
1
dν = |∇u|2 (z) log dA,
|z|
where u ∈ h∞ , A is the normalized 2-dimensional Lebesgue measure
restricted to D, dA = dxdy
π
, and ∇u stands for the gradient of u, so that
2 2
2
∂u ∂u
|∇u| = + .
∂x ∂y
This is a consequence of a more general result based on the following
identity of Littlewood and Paley.
Proposition 6.2. Let u ∈ h2 alnd λ ∈ D. Then

1 − λz
Z Z
2 2
(6.11) |u(ζ) − u(λ)| Pλ (ζ)dm(ζ) = |∇u| (z) log
dA(z).
T D z − λ

Proof. The proof is essentially based on the following direct ap-


plication of Green’s formula: If ϕ : D → C is twice continuously differ-
entiable then
Z Z
1
(6.12) ϕdm = ϕ(0) + ∆ϕ(z) log dA(z).
T D |z|
Moreover, another direct computation shows that if v is harmonic in
D, and ϕ(z) = |v(z)|2 , then
∆ϕ(z) = |∇v|2 (z).
Assume first that u is harmonic in a larger disc, and set v = u −
u(0), ϕ = |v|2 . Then (6.12) immediately implies (6.11) when λ = 0.
Thus, for any u ∈ h2 and 0 < r < 1, (6.11) holds for the dilation
ur (z) = u(rz) and λ = 0. Also, if 0 < r, ρ < 1 we can apply (6.11) to
obtain
Z
2 1
kur − uρ k2 = |∇ur (z) − ∇uρ (z)|2 log dA(z),
D |z|
32 2. POISSON INTEGRALS

and by Theorem 5.1 (iii) it follows that ∇ur − ∇uρ converges to zero in
1
L2 (D, log |z| dA), when r, ρ → 1− . This easily shows that when λ = 0,
(6.11) holds for every u ∈ h2 .
z−λ
If λ ∈ D is arbitrary, let φλ (z) = 1−λz . Clearly, φλ is an analytic
bijection from D onto itself. By what we have proved above, we have
for every u ∈ h2 ,
Z Z
2 1
|u ◦ φλ − u(λ)| dm = |∇u ◦ φλ |2 (z) log dA(z).
T D |z|
A direct computation similar to the above yields
|∇u ◦ φλ |2 (z) = |∇u|2 (φλ (z))|φ0λ (z)|2 ,
i.e.
Z Z
1
2
|u ◦ φλ − u(λ)| dm = |∇u|2 (φλ (z))|φ0λ (z)|2 log dA(z).
T D |z|
Then (6.11) follows directly from the change of variables ξ = φλ (ζ), w =
φλ (z). 

Exercise 1 Prove (6.12).

1
Corollary 6.1. Let u ∈ h2 . Then the measure |∇u|2 (z) log |z| dA, is
a Carleson measure if and only if
Z
sup |u(ζ) − u(λ)|2 Pλ (ζ)dm(ζ) < ∞.
λ∈D T

Proof. If the condition in the statement holds, we apply Propo-


sition 6.2 together with the inequality log x1 ≥ 1 − x, 0 < x ≤ 1, to
obtain
(1 − |λ|2 )(1 − |z|2 )
Z
(6.13) sup |∇u|2 (z) dA(z) < ∞,
λ∈D D |1 − λz|2
that is, (6.9) holds for the measure |∇u|2 (z)(1 − |z|2 )dA. Obviously,
1
this measure is comparable to the measure |∇u|2 (z) log |z| dA outside
1
the disc centered at the origin and of radius 2 , hence by Remark 6.1,
1
|∇u|2 (z) log |z| dA, is a Carleson measure. To see the converse, use
again the inequality log x1 ≥ 1 − x, 0 < x ≤ 1, to conclude that
|∇u|2 (z)(1 − |z|2 )dA is a Carleson measure, that is, (6.13) holds. We
want to prove that

1 − λz
Z
2
sup |∇u| (z) log
dA(z) < ∞.
λ∈D D z − λ
6. EMBEDDINGS OF hp INTO Lp (ν) 33

After the change of variables w = φλ (z) in both integrals above, we see


that it suffices to verify the estimate
Z Z
1
2
|∇u| (z) log dA(z) ≤ C |∇u|2 (z)(1 − |z|2 )dA(z),
D |z| D
for some C > 0 and all u ∈ h2 . This is a simple exercise based on
Parseval’s formula and is left to the reader. 
Obviously, bounded harmonic functions have the property stated in
Corollary 6.1, but there are also unbounded functions with this prop-
erty. An easy example whose verification is left to the reader is the
analytic function
f (z) = log(1 − z), z ∈ D.
There is an important intrinsic characterization of these functions in
terms of the boundary values alone. The Poisson integral u of a function
h ∈ L1 (m) satisfies the conditions in Corollary ?? if and only if
Z Z
1
h − 1
sup hdm dm < ∞,
I m(I) m(I)

I I
where the supremum is taken over all arcs I ⊂ T. We say that such a
function has bounded mean oscillation and denote the space of all such
functions by BM O.
The proof of the above assertions relies on a real-variable argument
which will not be covered here.
CHAPTER 3

Hardy spaces

The Hardy space H p , 0 < p ≤ ∞, is the subspace of hp consisting of


anlytic functions in D. A direct application of Corollary 5.1 shows that
when 1 ≤ p ≤ ∞, H p is a closed subspace of hp , in particular, it is
a Banach space. These spaces are the object of our study, and some
of the tools are provided by the material presented in the previous
chapter.

1. Outer functions.

In this section we shall apply the results on Poisson integrals in order to


construct analytic functions with prescribed boundary values of their
modulus.
Proposition 1.1. Let h be a nonnegative function on the unit circle
such that Z
| log h|dm < ∞.
T
Then function F defined in D by
Z 
ζ +z
F (z) = exp log h(ζ)dm(ζ)
T ζ −z

has nontangential limits almost everywhere on the unit circle. More-


over, if we denote by F (ζ) the nontangential limit of F at ζ ∈ T (when-
ever it exists) then |F (ζ)| = h(ζ) m−a.e. on T.

Proof. We leave it as an exercise to prove that F is analytic. Also


note that this function satisfies
Z  Z 
ζ +z
|F (z)| = exp Re log h(ζ)dm(ζ) = exp Pz (ζ) log h(ζ)dm(ζ) .
T ζ −z T

Consequently, by the results about Poisson integrals, |F | has the non-


tangential limit exp log h(ζ) = h(ζ), m−a.e.. It remains to show that F
has nontangential limits almost everywhere as well. This follows from
35
36 3. HARDY SPACES

the simple observation that if h is bounded above then the function F


is bounded in D. Indeed, if h ≤ M then
Z  Z 
|F (z)| = exp Pz (ζ) log h(ζ)dm(ζ) ≤ exp Pz (ζ) log M dm(ζ) = M.
T T

This means that the functions F1 , F2 defined by


Z 
ζ +z
F1 (z) = exp log min{h(ζ), 1}dm(ζ) ,
T ζ −z

and Z 
ζ +z
F1 (z) = exp log max{h(ζ), 1}dm(ζ) ,
T ζ −z
are analytic and bounded in D, hence both functions have nontangential
limits almost everywhere on the unit circle. By the above computation
we also know that these limits are nonzero m− a.e.. Thus F = F1 /F2
F has nontangential limits almost everywhere on the unit circle as
well. 

Of course, the function F constructed above, is not the unique ana-


lytic functions such that the modulus of its nontangential limits take
the prescribed value h a.e. If B is any finite Blaschke product then
|B| extends continuously to the boundary and takes the value 1 there.
Thus, BF shares the same property as F . It is also possible to con-
struct another analytic function G in D such that |G| has the same
nontangential limits as |F | a.e. and in addition, G has no zeros in D.
1+z
Indeed, if g(z) = exp(− 1−z ) then G = gF has the required property
because g extends continuously to ∂D \ {1} with |g(ζ)| = 1 for all
ζ ∈ ∂D \ {1}. Nevertheless, the function F constructed in the proof of
the above theorem plays a central role in what follows.
Definition 1.1. An analytic function F in D is called an outer func-
tion if there exists a nonnegative function h on the unit circle such
that Z
| log h|dm < ∞
T
and Z 
ζ +z
F (z) = exp log h(ζ)dm(ζ) .
T ζ −z
In this case, F is the outer function whose modulus equals h a.e. on
the boundary.
1. OUTER FUNCTIONS. 37

Exercise 1. Show that the product and quotient of two outer functions
is outer. Moreover, prove that if F is outer with |F (eit )| ≥ a > 0
a.e. then |F (z)| ≥ a for all z ∈ D. Finally, show that any function
F that is analytic in a larger disc (than D) and has no zeros in D is
an outer function, but the statement is no longer true if we replace the
assumption ”F analytic in a larger disc” by ”F continuous in D.
Proposition 1.2. Let h be a nonnegative function on the unit circle
such that Z
| log h|dm < ∞
T
and let Fh be the outer function whose modulus equals h a.e. on the
boundary. Then
Z
|Fh (z)| ≤ Pz (ζ)h(ζ)dm(ζ).
T

The proof is essentially based on the following simple lemma.


Lemma 1.1. If u, v : [a, b] 7→ R are integrable functions, v is nonnega-
Rb
tive with a v(x)dx = 1 then
Rb
Z b
u(t)v(t)dt
e a ≤ eu(t) v(t)dt.
a

Proof. The inequality is equivalent to


Z b  Rb 
1≤ v(t) eu(t)− a u(x)v(x)dx dt.
a
Since for all T ∈ R we have eT ≥ T + 1, we deduce that
Z b
u(t)− ab u(x)v(x)dx
R
e ≥ u(t) − u(x)v(x)dx + 1 , t ∈ [a, b].
a
Integration on [a, b] (w.r.t t) now yields
Z b   Z b  Z b 
u(t)− ab u(x)v(x)dx
R
v(t) e dt ≥ v(t) u(t) − u(x)v(x)dx + 1 dt = 1
a a a
and the result follows. 

Exercise 2. Prove the following generalization of Lemma 1.1 which is


called Jensen’s inequality.
Let φ : R 7→ R be a differentiable convex function, that is, φ satisfies
φ(x) − φ(y) ≥ φ0 (y)(x − y), x, y ∈ R.
38 3. HARDY SPACES
Rb
If u, v : [a, b] 7→ R are integrable functions, with v ≥ 0 and a
v(x)dx =
1 then Z b  Z b
φ u(t)v(t)dt ≤ φ(u(t))v(t)dt.
a a
R
Proof of Proposition 1.2. Recall first that T Pz dm = 1 for all z ∈ D
1
and set u(t) = log h(eit ), v(t) = 2π Pz (eit ). Then by Lemma 1.1 we
obtain
Z  Z
|Fh (z)| = exp Pz (ζ) log h(ζ)dm(ζ) ≤ Pz (ζ)h(ζ)dm(ζ).
T T

Corollary 1.1. Let h be a nonnegative function on the unit circle


such that Z
| log h|dm < ∞
T
and let Fh be the outer function whose modulus equals h a.e. on the
boundary. Then for 0 < p ≤ ∞, Fh ∈ H p if and only if h ∈ Lp (m),
and
kFh kp = khkLp (m) .

Proof. When 1 < p ≤ ∞, the result follows from the properties of


Poisson integrals. However, the following argument works for all p > 0.
The inequality
kFh kp ≥ khkLp (m) ,
follows by an application of Fatou’s lemma and Proposition 1.1. The
reverse inequality follows by Proposition 1.2 applied to the outer func-
tion Fhp , which is obviously well defined. 

2. Zeros of H p -functions

It turns out that in terms of zeros, the condition defining membership


in H p is quite restrictive, and in fact, it is equivalent to the Blaschke
condition. This has a number of important consequences for the devel-
opment of the theory.
We begin with a lemma on the integrals involved in the definition of
H p norms.
Lemma 2.1. If f is analytic in D and 0 < p < ∞, then
Z 2π
dt
Mp (r, f ) = |f (reit |p ,
0 2π
is an increasing function of r ∈ [0, 1).
2. ZEROS OF H p -FUNCTIONS 39

Proof. If p > 1, the result can be obtained using properties of


Poisson integrals. The argument below is valid for all values p > 0. It
is easy to verify that for 0 < r < 1,
Z
| log |f (rζ)||dm(ζ) < ∞,
T

so that we can consider the outer function Fr with |Fr (ζ)| = |f (rζ)|,
m-a.e. on T . Moreover, if f has no zeros on rT, then |Fr | is bounded
below, hence fr /Fr is bounded in D, where, as usual fr (z) = f (rz), z ∈
D. Since the nontangential limits of fr /Fr have modulus one a.e. on
the circle, it follows by the Poisson representation that |fr /Fr | ≤ 1,
that is,
|Fr (z)| ≥ |f (rz)|, z ∈ D.
Then for 0 < ρ < 1 we have by Corollary 1.1
Mp (rρ, f ) ≤ kfr kpp ≤ kFr kpp = Mp (r, f ).
Thus we have shown the inequality Mp (rρ, f ) ≤ Mp (r, f ) for all ρ ∈
(0, 1) and all r in a dense subset of [0, 1]). The result follows. 
Theorem 2.1. Let 0 < p ≤ ∞, and assume that f ∈ H p is not
identically zero. Let a1 , a2 , . . . be the zeros of f in D, repeated according
to their multiplicity. Then
X
(1 − |an |) < ∞.
n≥1

Moreover, if B is the Blaschke product with zeros a1 , a2 , . . . then f /B ∈


H p , and
||f /B||p = ||f ||p .

Proof. Denote by BN the partial products


N
Y
m −an z − an
BN (z) = z .
n=m+1
|an | 1 − an z

Then f /BN is analytic in D. We claim that f /BN ∈ H p with


kf /BN kp = kf kp .
If p = ∞ this follows immediately from the maximum principle together
with the fact that BN is continuous in the closed unit disc and |BN | = 1
on T. Indeed, if we pick a a sequence (zn ), |zn | → 1 such that
|f (zn )|
||f /BN ||∞ = lim ,
n→∞ |BN (zn )|
40 3. HARDY SPACES

from |BN (zn )| → 1, n → ∞, we obtain


||f /BN ||∞ = lim |f (zn )| ≤ ||f ||∞ .
n→∞

The reverse inequality is abvious, since |BN | ≤ 1 in D. The case when


0 < p < ∞ is similar, but the maximum principle is replaced by Lemma
2.1. By this result we have
kf /BN kpp = lim− Mp (r, f /BN ) ≤ lim sup Mp (r, f ) max(|BN (z)|−p = kf kp ,
r→1 r→1− |z|=r

and, again the reverse inequality is immediate from |BN | ≤ 1 in D.


With the claim in hand, we see that (BN ) cannot converge to zero
uniformly on compact subsets of D, since f does not vanish identically,
hence this would contradict the estimates
Mp (r, f /BN ) ≤ kf kpp , max |f /BN (z)| ≤ kf k∞ ,
|z|=r

valid for all r ∈ (0, 1). Thus the Blaschke condition in the statement
holds true. Letting N → inf ty, we obtain from the last estimates
Mp (r, f /B) ≤ kf kpp , max |f /B(z)| ≤ kf k∞ ,
|z|=r

i.e. f /B ∈ H p and kf /Bkp ≤ kf kp . The reverse inequality follows


again from |b(z)| < 1, z ∈ D, and the proof is complete. 

Exercise 1. In the case when p = ∞ Theorem 2.1 belongs to a very


classical circle of ideas which includes the famous Schwarz Lemma.
Prove the following assertions:
(i) (Schwarz Lemma ) Let f be analytic in D and satisfy |f (z)| ≤ 1
there. If f (0) = 0 then
|f (z)| ≤ |z|, z ∈ D,
and
|f 0 (0)| ≤ 1.
Moreover, if equality occurs in one of the above inequalities then f is a
rotation of the unit disc, i.e. f (z) = αz for some α ∈ C with |α| = 1.
(ii) (Schwarz-Pick Lemma ) Let f be analytic in D and satisfy
|f (z)| ≤ 1 there. Then if z, w ∈ D, z 6= w we have
|f (z) − f (w)| |z − w|

|1 − f (w)f (z)| |1 − wz|
3. APPLICATIONS 41

z−a
with equality if and only if f has the form f (z) = α 1−az for some fixed
a ∈ D and α ∈ ∂D. In particular, for all z ∈ D we have also
1
|f 0 (z)| ≤ .
1 − |z|2

Corollary 2.1. Let (an ) be a sequence in D such that


X
(1 − |an |) < ∞
n≥1

and let B be the corresponding Blaschke product. Then B has nontan-


gential limits of modulus one almost everywhere on the unit circle.

Proof. Clearly B is in H ∞ with kBk∞ ≤ 1. Let f be any bounded


analytic function in D and apply Theorem 2.1 to the function Bf . Then
the conclusion reads
||f B/B||∞ = ||f ||∞ = ||Bf ||∞ .
Now suppose there exist ε > 0 and a measurable set E ⊂ T, m(E) > 0
such that the nontangential limits B(ζ) satisfy |B(ζ)| < 1−ε for ζ ∈ E.
Let f be the outer function whose modulus equals 1 a.e. on E and 21
a.e. on its complement. Then by Corollary 5.3, Chapter I we have that
||f ||∞ = 1, but
1
||Bf ||∞ = essup|Bf | < max{ , 1 − ε} < 1
2
which contradicts the previous equality. Thus, for all ε > 0 the set of
points ζ where |B(ζ)| < 1 − ε has measure zero and hence, the set

[
{ζ, |B(ζ)| < 1} = {ζ, |B(ζ)| < 1 − 1/n}
n=1

has measure zero as well. 

3. Applications

Theorem 2.1 yields a very useful factorization of H p -functions, which


has a number of important consequences. According to this result,
every nonzero f ∈ H p , 0 < p ≤ ∞ cna be written in the form
f = Bg,
where B is a Blachke product and g ∈ H p is zero-free in D with kgkp =
kf kp .
42 3. HARDY SPACES

3.1. Boundary behavior. With the above factorization in hand,


we can easily apply the results obtained for Poisson integrals in Section
5 of the previous chapter. Indeed, if f = Bg ∈ H p , the zero-free factor g
belongs to H p so that, g p/2 belongs to H 2 ⊂ h2 . This simple observation
leads to a quick proof of the following theorem.
Theorem 3.1. (i) If f ∈ H p , 0 < p ≤ ∞, then f has nontangential
limits f (ζ) for m-a.e. ζ ∈ T and
kf kp = kf kLp (m) .
p
(ii) If f ∈ H , 0 < p ≤ ∞, then its maximal nontangential function
f ∗ belongs to Lp (m), and there exists cp > 0 such that
kf ∗ kLp (m) ≤ Cp kf kp .

Proof. (i) As pointed out above, if f = Bg with B a Blachke


product and g ∈ H p zero-free in D, then g p/2 ∈ H 2 , which implies
that g p/2 , and hence g has nontangential limits m-a.e.. Since B ∈ H ∞ ,
f = BG has nontangential limits m-a.e.. By Corollary 2.1 we have
|f | = |g|, m−a.e. on T, and using also Theorem 5.1 we obtain
2/p 2/p
kf kp = kgkp = kg p/2 k2 = kg p/2 kL2 (m) = kf kLp (m) .
(ii) If f = Bg as above f ∗ ≤ g ∗ = (g p/2 )∗ , and the result follows again
from the fact that g p/2 ∈ H 2 . 

The norm convergence of dilations also extends for all values of 0 <
p < ∞ as the following result shows.
Theorem 3.2. If f ∈ H p , 0 < p < ∞, and for 0 ≤ r < 1, fr (z) =
f (rz), z ∈ D, then
lim− kf − fr kp = 0.
r→1

Proof. By Theorem 3.1 (i) we have


Z
kf − fr kp = |f − fr |p dm,
p
T
and since |f (ζ) − f (rζ)| ≤ 2f ∗ (ζ), ζ ∈ T, the result follows by an
application of the dominated convergence theorem 

As pointed out in the previous chapter, the result fails when p = ∞.


Corollary 3.1. (i) For 1 ≤ p < ∞, H p is a separable Banach space.
(ii) For 0 < p < 1, H p is a complete separable space w.r.t. the metric
dp (f, g) = kf − gkpp .
3. APPLICATIONS 43

Proof. If f ∈ H p , and 0 < r < 1, then fr can be approximated


uniformly by polynomials, so that polynomials with rational coefficients
form a countable dense subset of H p . The verification of the remaining
assertions in (ii) is straightforward. 

3.2. Poisson representation of H 1 and the F. & M. Riesz


theorem. Since H p ⊂ hp , it follows that for 1 < p ≤ ∞, every function
in H p is the Poisson integral of its boundary function. Interesting
enough, this result continues to hold for H 1 as well, even if it fails for
the larger space h1 .
Theorem 3.3. If f ∈ H 1 then
Z
f (z) = Pz (ζ)f (ζ)dm(ζ), ζ ∈ D.
T

Proof. This is a direct application of Theorem 3.2. Since for 0 <


r < 1 fr ∈ H ∞ , we have
Z
f (rz) = Pz (ζ)f (rζ)dm(ζ), ζ ∈ D.
T

Let r → 1 , and apply Theorem 3.2 together with the fact that for
fixed z ∈ D, the Poisson kernel Pz (·) belongs to L∞ (m). 

The next theorem is a famous result due to the brothers Riesz with wide
applications in function theory. The original proof is quite different
from the one below and is of interest in its own right.
Theorem 3.4. Let µ be a finite Borel measure on T with the property
that Z
ζ n dµ(ζ) = 0,
T
for all nonnegative integers n. Then µ is absolutely continuous w.r.t.

m and there exists f ∈ H 1 with f (0) = 0 such that dm = f.

Proof. Let u be the Poisson integral of µ, and recall that u de-


termines µ uniquely. Clearly, u ∈ h1 , and u(0) = 0. We show that u is
analytic in D, hence u ∈ H 1 . For z ∈ D,
Z Z
ζ̄ + z̄ ζ +z
2u(z) = 2 Pz (ζ)dµ(ζ) = dµ(ζ) + intT dµ(ζ).
T T ζ̄ − z̄ ζ −z
If ζ ∈ T

ζ̄ + z̄ 1 + z̄ζ X
= =1+2 z̄ n ζ n ,
ζ̄ − z̄ 1 − z̄ζ n=1
44 3. HARDY SPACES

and the series converges uniformly on T, for fixed z ∈ D. Thus by


assumption, Z
ζ̄ + z̄
dµ(ζ) = 0,
T ζ̄ − z̄
and
ζ +z
2u(z) = +intT dµ(ζ),
ζ −z
so that u ∈ H 1 . Then the result follows by Theorem 3.3. 

3.3. Inner-outer factorization. The factorization used in the


previous section can be further refined. The crucial step is to construct
for nonzero f ∈ H p the outer function F| f |. To this end we need to
show that log |f | ∈ L1 (m), a quite remarkable property of H p -functions.
The classical approach to this fact is via the so-caled Jensen’s formula,
but we shall follow a different path here and begin with the following
inequality, which is an analogue of (5.7), Chapter 2.
Proposition 3.1. If f ∈ H p , 0 < p < ∞, then for all z ∈ D,
Z
|f (z)| ≤ Pz (ζ)|f (ζ)|p dm(ζ).
p
T
In particular,
 1/p
1 + |z|
|f (z)| ≤ kf kp , z ∈ D.
1 − |z|
Proof. Write again f = Bg, as in the previous proofs, with B
a Blachke product and g ∈ H p zero-free in D, and recall that |g| =
|f |, m−a.e. on T. Since g p/2 ∈ H 2 ⊂ h2 we can apply the estimate
(5.7) to obtain
Z Z
|f (z)| ≤ |g (z)| ≤ Pz (ζ)|g(ζ)| dm(ζ) = Pz (ζ)|f (ζ)|p dm(ζ).
p p/2 2 p
T T
The second part follows by the standard estimate of the Poisson kernel.


This inequality exhibits a harmonic majorant for H p -functions, and


it turns out that such harmonic majorants can be used to give an
alternative definition of Hardy spaces which extends to more general
plane domains as well. We shall not pursue this matter here.

Exercise 1. Show that if f ∈ H p , 0 < p < ∞, the harmonic function


Z
uf (z) = Pz (ζ)|f (ζ)|p dm(ζ)
T
3. APPLICATIONS 45

is the least harmonic majorant of |f |p in D, that is, if v is harmonic


in D with v ≥ |f |p , then v ≥ uf .

With Proposition 3.1 in hand we can turn to our original goal.


Proposition 3.2. If f ∈ H p , 0 < p ≤ ∞ is not identically zero
then log |f | ∈ L1 (m). Moreover, if F is the outer function with F | =
log |f |, m−a.e. on T then
|f (z)| ≤ |F (z)|, z ∈ D.

Proof. We start with the case when p = ∞ and assume without


loss of generality that kf k∞ ≤ 1. Write f = Bg with B a Blachke
product and g ∈ H p zero-free in D, and note that
| log |f (ζ)|| = − log |f (ζ)| = − log |g(ζ)|,
m−a.e. on T. By Fatou’s lemma
Z Z
− log |g(ζ)|dm(ζ) ≤ lim inf

− log |g(rζ)|dm(ζ) = − log |g(0)|,
T r→1 T
because log |g| is harmonic in D. If p ∈ (0, ∞), let G be the outer
function with |G(ζ)| = exp(|f (ζ)|p ), m−a.e. on T and note that:
1) |G(z)| ≥ 1, z ∈ D,
p
2) |f /G| is essentially bounded on T (ex > cp x),
From the first inequality we see that f /G ∈ H p and if we apply Propo-
sition 3.1 to f /G it follows easily that f /G ∈ H ∞ . Then by the first
part of the proof we have that
log |f | − log |G| = log |f | − |f |p ∈ L1 (m),
i.e. log |f | ∈ L1 (m). To see the second part of the statement, we con-
sider the sequence of outer functions (Fn ) with |Fn (ζ)| = log(|f (ζ)| +
1
n
), m−a.e. on T. Exactly the same reasoning as before shows that
f /Fn ∈ H ∞ with kf /Fn k∞ ≤ 1. Using the first part of the proof
we can easily show that Fn (z) → F (z) for all z ∈ D, which gives the
desired result. 

The outer function F in Proposition 3.2 will be called the outer factor
of f .
Definition 3.1. A bounded analytic function I in D is called inner if
its nontangential limits satisfy |I(ζ)| = 1, m− a.e. on T.

Clearly, inner functions are bounded by 1 in D. Note that unimodular


constant are inner functions. Less trivial examples are provided by
Blaschke products. According to Proposition 3.2 I = f /F is inner,
46 3. HARDY SPACES

whenever f ∈ H p . This function will be called the inner factor of


f ∈ H p . and the factorization
(3.14) f = IF
will be retfered to as the inner-outer factorization of f .
Proposition 3.3. The inner outer factorizarion (3.14) of f ∈ H p , 0 <
p ≤ ∞ is unique.

Proof. If f = IF = JG, with I, J inner and F, G outer then


I G
= ,
J F
which implies that the nontangential limits of G/F are unimodular a.e.
on T. Since G/F is outer it follows by definition that G/F = 1. 

At its turn, by Theorem 2.1 the inner factor can be further decomposed
as I = BS, where B is a Blaschke product and a function S which is
zero-free in D. Here we consider B = 1 if I has no zeros to begin with.
Definition 3.2. An inner function without zeros in D is called sin-
gular inner.

Singular inner functions have a very special form. If S is such a function


then log |S| is harmonic and negative in D, in particular it belongs to
h1 . Thus, by Corollary 2.2 in Chapter 2, there is a nonnegative finite
Borel measure µ on T such that
Z
log |S(z)| = − Pz (ζ)dµ(ζ).
T

Since log |S| has nontangential limits zero a.e. on T, it follows by


Corollary 4.1 in Chapter 2, that the measure µ is singular w.r.t. m.
This argument together with analytic completion and exponentiation
yields the following representation formula for singular inner functions.
Corollary 3.2. If S is a singular inner function then there exists
α ∈ R and a nonnegative finite Borel measure µ on T which is singular
w.r.t. m, such that
 Z 
ζ +z
S(z) = exp iα − dµ(ζ) .
T ζ −z

The above argument refines the inner outer factorization f = IF , of


f ∈ H p . The final result is stated below, but already proved above.
3. APPLICATIONS 47

Theorem 3.5. Let f ∈ H p , 0 < p ≤ ∞. If f is not identically zero, it


can be written uniquely in the form
f = BSF,
where B is a Blaschke product (or B = 1), S is a singular inner func-
tion, and F is the outer factor of f .

Exercise 2. Show that if f, 1/f ∈ H p for some p > 0, then f is outer.

Exercise 3. Show that a singular inner function cannot extend contin-


uously to the closed unit disc unless it is constant. Construct a function
that extends continuosly to D and yet has a nontrivial singular inner
factor. Similarily one can construct such a function that has infinitely
many zeros. Hint: Take the Blaschke product with zeros 1 − n−2 , and
multiply by (1 − z). Then show that the product is continuous.

Exercise 4. What is the cannonical factorization given by Theorem


3.5 of the following functions?
a) f (z) = (2z − 1)ez , z ∈ D.
b) g(z) = exp( z21−1 ), z ∈ D.
c) h(z) = sin z, z ∈ D.

3.4. Littlewood subordination. Given f, g analytic in D, we say


that f is subordinate to g, and write f ≺ g if there exists ϕ : D → D
analytic with ϕ(0) = 0 such that f = g ◦ ϕ. Note that in this case
f (D) ⊆ g(D), and by the Schwarz lemma we have f (rD) ⊆ g(rD),
for all r ∈ [0, 1]. J. E. Littlewood proved long ago that subordination
decreases the H p -norm, and his theorem is part of the following result.
Theorem 3.6. Let ϕ : D → D be analytic. Then for every f ∈ H p , 0 <
p < ∞, we have f ◦ ϕ ∈ H p and
 1/p
1 + |ϕ(0)|
kf ◦ ϕkp ≤ kf kp .
1 − |ϕ(0)|
Equality holds for some 0 < p < ∞ and all f ∈ H p , if and only if ϕ is
inner with ϕ(0) = 0.

Proof. The norm inequality in the statement is a direct conse-


quence of Proposition 3.1, since by that inequality we have for every
f ∈ H p, Z
|f (ϕ(z))|p ≤ Pϕ(z) (ζ)|f (ζ)|p dm(ζ).
T
48 3. HARDY SPACES

Then
Z Z Z
p
|f (ϕ(rz))| dm(z) ≤ Pϕ(rz) (ζ)|f (ζ)|p dm(ζ)dm(z)
T
ZT ZT
= |f (ζ)|p Pϕ(rz) (ζ)dm(z)dm(ζ),
T T
and since z → Pϕ(rz) (ζ) is harmonic in r−1 D, the inner integral is
1 + |ϕ(0)|
Z
Pϕ(rz) (ζ)dm(z) = Pϕ(0) (ζ) ≤ ,
T 1 − |ϕ(0)|
which proves the first part of the theorem. To see the second, assume
first that equality holds for some 0 < p < ∞ and all f ∈ H p , and let
f (z) = z, to obtain
Z
1 = kf kp = kf ◦ ϕkp = |ϕ(ζ)|p dm(ζ),
p p
T
which immediately |ϕ(ζ)| = 1, m−a.e., because, by the assumption of
the theorem we have |ϕ(ζ)| ≤ 1, m−a.e.. To see that ϕ(0) = 0, let
ζ ∈ T and f (z) = (ζ + z)2/p . Then
Z
2 = kf kp = kf ◦ ϕkp = |ζ + ϕ(z)|2 dm(z) = 1 + 2Re ζϕ(0) + 1.
p p
T
Since ζ ∈ T is arbitrary, we see that ϕ(0) = 0.
Conversely, assume that ϕ is inner with ϕ(0) = 0, and let Q(z) =
n
P
n an z be any polynomial. Then
X Z X
2
kQ ◦ ϕk2 = an ak ϕn (ζ)ϕk (ζ)dm(ζ) = |an |2 = kQk22 ,
n,k T n

since for n > k


Z Z
n
ϕ (ζ)ϕk (ζ)dm(ζ) = ϕn−k (ζ)dm(ζ) = 0.
T T
If f ∈ H is arbitrary, we approximate f in the norm of H 2 by a
2

sequence of polynomials (Qn ). By the first part of the theorem Qn ◦ϕ →


f ◦ ϕ in H 2 , which gives that kf k2 = kf ◦ ϕk2 . In particular, for every
inner function I we have
1 = kIk2 = kI ◦ ϕk2 ,
which implies that I ◦ ϕ is inner as well. If 0 < p < ∞ is arbitrary, and
f ∈ H p , we write f = Bg with B a Blaschke product and g p/2 ∈ H 2 .
Then B ◦ ϕ is inner and
2/p 2/p
kf ◦ ϕkp = kg p/2 ◦ ϕk2 = kg p/2 k2 = kf kp ,
and the proof is complete. 
3. APPLICATIONS 49

This result, especially the first part of it has numerous applications.


We shall only mention one of these below.
Corollary 3.3. Let f be analytic in D with Re f (z) ≥ 0, z ∈ D.
Then f is outer and belongs to H p for all p < 1.

1+φ(0)
Proof. If φ is an automorphism of the disc with f (0) 1−φ(0) , then
1+φ
f ≺ g = 1−φ . Indeed, g maps the disc onto the right half-plane, hence
g −1 ◦f maps D into itself with g −1 ◦f (0) = 0. We claim that g ∈ H p , 0 <
p < 1. In view of Theorem 3.6 it suffices to show that z → 1+z 1−z
belongs
to H p , 0 < p < 1, which follows by a direct calculation. Thus, another
application of Theorem 3.6 gives that f ∈ H p , 0 < p < 1. Note that
1/f has nonnegative real part as well, so 1/f ∈ H p , 0 < p < 1, hence
by Exercise 2 in the previous subsection, f is outer. 

3.5. The Phragmen-Lindelöf Principle. The type of problem


we want to investigate in this section is illustrated by the following
example.
Let 0 < a < π, and suppose that f is analytic in the the angle Γ =
{z = reit , r > 0, |t| < a}, and continuous in its closure, and assume
there exists M > 0 such that

|f (re±ia )| ≤ M, r ≥ 0.

It is not difficult to show by means of examples that such a function


f does not need to be bounded in Γ, i.e. the ”naive” form of the
maximum principle fails in this domain. Of course one would like
to add conditions that are as weak as possible, under which one can
conclude that |f | ≤ M in Γ.
A more general situation of this type can be described in the unit disc.
Consider an analytic function f in D which has the nontangential limit
f (ζ), for almost every ζ ∈ T. It turns out that the properties of f are
difficult to deduce from the behavior of the boundary function. For
example, if S is a nonconstant singular inner function, then f = 1/S
has unimodular nontangential limits a.e. on T, yet f is unbounded, in
fact, it doesn’t belong to any H p , p > 0. This is easily seen from the
inner-outer factorization, since 1/S = JF implies F = 1, hence SJ = 1,
and by the maximum principle, both S and J would be constant.
It turns out that this type of situation can be addressed with help of
the Phragmen-Lindelöf principle which we present here in the most
general form-
50 3. HARDY SPACES

Theorem 3.7. Let F be an outer function and f be an analytic func-


tion in D such that
|f (z)| ≤ |F (z)|
for all z ∈ D. Then f has the nontangential limit f (ζ), for almost
every ζ ∈ T, and if f ∈ Lp (m), for some p ∈ (0, ∞], then f ∈ H p .

Proof. By assumption we have that f /F ∈ H ∞ , so that f /F , and


hence f has nontangential limits a.e. on T. By Proposition 1.1 F has
nonzero nontangential limits a.e. on T, hence, so has f . The inner
-outer factorization of f /F gives f /F = IG with I inner and G outer,
so that f = IGF , and
|f | = |GF |,
m−a.e. on T. Then by Corollary 1.1 GF ∈ H p , and since |f | ≤ |GF |
in D, it follows that f ∈ H p . 
Corollary 3.4. Let 1 ≤ p ≤ ∞. The map which associates to f ∈ H p
its boundary function is a linear isometry from H p onto the closed
subspace of Lp (m) consisting of functions g with
Z
ζ n g(ζ)dm(ζ) = 0, n ≥ 1.
T
Its inverse is given by the Cauchy formula
Z
ζf (ζ)
f (z) = dm(ζ).
T ζ −z

Proof. It is clear that the boundary functions of H p -functions


belong to the subspace described in the statement , and also that this
subspace is closed. If g ∈ Lp (m) with
Z
ζ n g(ζ)dm(ζ) = 0, n ≥ 1,
T
then it is the boundary function of an H 1 -function, and the result
follows by Theorem 3.7. The Cauchy formula obviously holds for the
dilated function fr , f ∈ H p , and the formula in the statement follows
by letting r → 1− . 

We close with an application of the theorem to the situation described


at the beginning of the paragraph.
Theorem 3.8. Suppose that f is analytic in the the angle Γ = {z =
reit , r > 0, |t| < a}, where 0 < a < π is fixed. Assume there exists
M > 0 such that for every boundary point ζ = re±ia we have
lim sup |f (z)| ≤ M.
z→ζ
z∈Γ
3. APPLICATIONS 51

If f satisfies in addition an inequality of the form


|f (z)| ≤ b exp(c|z|d )
π
for some constants b, c > 0, 0 < d < 2a and all z ∈ Γ then f is bounded
in the whole angle with |f (z)| ≤ M, z ∈ Γ
Proof. Under our assumption on the constant d, there is c1 > 0
such that |z|d ≤ Re z d , hence
|f (z)| ≤ b exp(c2 Re z d ), z ∈ Γ.
2a/π
Let ψ : D → Γ, ψ(z) = ( 1+z 1−z
. The ψ maps D conformally onto Γ,
so that it suffices to show that f ◦ ψ ∈ H ∞ with kf ◦ ψk∞ ≤ M . By
the first inequality in this proof we have that
d (z)
|f ◦ ψ(z)| ≤ |eψ | = |F (z)|.
We claim that F is outer. If the claim holds, the result follows by
Theorem 3.7, since the nontangential limits of ψ lie on ∂Γ a.e. and
by assumption it follows that the nontangential limits of |f | are ≤ M
a.e.. To see the claim, recall from the previous subsection that ψ π/2a ∈
H p , 0 < p < 1, which implies that Re ψ d ∈ hq , for some q > 1, hence
it is the Poisson integral of its boundary function, which is obviously
integrable. Thus F is outer and the proof is complete. 
CHAPTER 4

The dual of H p

1. Basic reduction to the study of Cauchy integrals

We shall be concerned with the continuous linear functionals on the


Banach spaces H p , 1 ≤ p < ∞, that is the linear maps l : H p → C
with
|l(f )| ≤ Ckf kp ,
for some C > 0 and all f ∈ H p .
The description of these functionals is particulary simple when p = 2,
since H 2 is a Hilbert space with respect to the scalr prduct in L2 (m)
Z
hf, gi = f gdm, f, g ∈ H 2 .
T

According to the Riesz representation theorem, every continuous linear


functional l on H 2 has the form
Z
l(f ) = hf, gi = f gdm, f ∈ H 2 ,
T
2
for some fixed g ∈ H .
When p 6= 2, general functional analysis provides less information.
However we can use Corollary 3.4 to embed H p ito Lp (m) and then use
the Hahn-Banach theorem to represent every continuous linear func-
tional l on H p , 1 ≤ p < ∞ by
Z
l(f ) = f gdm,
T

with g ∈ Lq (m), p1 + 1q = 1. The problem that arises here, is that this


representation is not unique, it leads to a representation of the dual of
H p as a quotient space. This is a well known fact in functional analysis
and follows by another application of the Hahn-Banach theorem. The
proof of the proposition below is left as an exercise. As usual we denote
the dual of the normed space X by X 0 and if S ⊂ X,
S ⊥ = {l ∈ X 0 : l(S) = {0}}.
53
54 4. THE DUAL OF H p

Proposition 1.1. Let X be a Banach space and let Y be a closed


subspace. Then the map J : X 0 /Y ⊥ → Y 0 ,
J[l] = l|Y,
is a linear isometry.

According to this result, (H p )0 is isometrically isomorphic to the quo-


tient space Lq (m)/(H p )⊥ . Moreover, (H p )⊥ consists of functions g ∈
Lq (m) with
Z
ζ n g(ζ)dm(ζ) = 0, n ≥ 0,
T

hence by the F.& M. Riesz theorem, it follows easily that (H p )⊥ = H0q ,


the subspace of H q consisting of functions that vanish at the origin.
Can we ”single out” elements in the cosets [f ] ∈ Lq (m)/H0q in a mean-
ingful way?
There is a simple but powerful idea that can be used here, namely to
consider Cauchy integrals. Note that if g ∈ (H p )⊥ then
Z
ζ̄g(ζ)
dm(ζ) = 0, z ∈ D
T ζ̄ − z̄

since

X z̄ n X ∞
ζ̄
= n
= z̄ n ζ n , ζ ∈ T, z ∈ D,
ζ̄ − z̄ n=0
ζ̄ n=0

and the series converges uniformly on T when z ∈ D is fixed. This


gives the following result which is the starting point for the material of
this chapter.
Proposition 1.2. Let l be a continuous linear functional on H p , 1 ≤
p < ∞. Then there exists a unique analytic function h in D of the form
Z
ζg(ζ)
h(z) = dm(ζ), z ∈ D,
T ζ −z

where g ∈ Lq (m), kgkLq (m) ≤ 2klk, such that


Z
l(f ) = lim− f (ζ)h(rz)dm(z) f ∈ H p.
r→1 T

Proof. We can write


Z
l(f ) = f (ζ)g(ζ)dm(ζ),
T
1. BASIC REDUCTION TO THE STUDY OF CAUCHY INTEGRALS 55

with [g] ∈ Lq (m)/(H p )⊥ and kgkLq (m) ≤ 2k[g]k = klk. Moreover, since
fr → f in H p , when r → 1− we obtain easily that l(fr ) → l(f ), i.e.
Z
l(f ) = lim− f (rζ)g(ζ)dm(ζ).
r→1 T
Using the Cauchy formula as in Corollary 3.4 we have for ζ ∈ T
Z
zf (z)
f (rζ) = dm(z), ζ ∈ T, 0 < r < 1,
T z − rζ
and by an application of Fubini’s theorem we obtain
Z Z Z
g(ζ)
f (rζ)g(ζ)dm(ζ) = zf (z) dm(ζ)dm(z)
T T T z − rζ
Z Z
ζ̄g(ζ)
= f (z) dm(ζ)dm(z).
T T ζ̄ − rz̄
Then the result follows with
Z
ζg(ζ)
h(z) = dm(ζ), z ∈ D,
T ζ −z
using the considerations preceeding the proposition. 

According to this result, one way to describe the continuous linear


functionals on H p is to understand Cauchy integrals of Lq (m)-functions
1
p
+ 1q = 1, and this is precisely what will be done in the sequel.
The Cauchy integral, or the Cauchy transform of h ∈ L1 (m) is denoted
by b
h and is defined by
Z
h(ζ)
(1.15) h(z) =
b dm(ζ), z ∈ C \ T.
T ζ −z

This formula gives 2 analytic functions, one in D, and the other in C\D,
but we will be mostly concerned with the first, since the properties of
the second are very similar. In fact their boundary values are related
by the so-called ”jump theorem”.
Theorem 1.1. If h ∈ L1 (m)
1
h(rz) − b
lim− b h( z) = z̄h(z)
r→1 r
m−a.e. on T.

Proof. It suffices to check the simple identity


1 1
− = ζ̄Pz (ζ)
ζ − rz ζ − r−1 z
and apply the results about boundary behavior of Poisson integrals. 
56 4. THE DUAL OF H p

2. The M. Riesz theorem

In view of the representation as Cauchy integrals of H p -functions, 1 ≤


p ≤ ∞, proved in Corollary 3.4, a natural question that comes to
mind is whether the Cauchy transforms of Lp (m)-functions belong to
H p , 1 ≤ p ≤ ∞?
For real-valued functions, the above question turns into a question
about harmonic conjugation. Indeed, if h ∈ L1 (m) is real-valued with
Poisson integral u, then the analytic function
Z
ζ +z
f (z) = h(ζ)dm(ζ), z ∈ D,
T ζ −z

satisfies
Ref (z) = u(z), z ∈ D, , f (0) ∈ R.
Thus for such functions, we can reformulate the above in terms of real
parts:
If f is analytic in D and Ref ∈ hp , 1 ≤ p ≤ ∞, does f ∈ H p (or
Imf ∈ hp )?
Obviously the original question reduces to real-valued functions, so that
the two are equivalent.
It turns out that the answer is negative for the ”endpoints” p = 1 and
p = ∞. Indeed, if
1+z
f1 (z) = f2 (z) = i log(1 − z), z ∈ D,
1−z
then Ref1 (z) = Pz (1), so that Ref1 ∈ h1 , and Ref2 ∈ h∞ . However,
/ h1 , because its boundary function is not integrable, and f2 is ob-
f1 ∈
viously unbounded.

Nevertheless, the questions discussed above have an affirmative answer


for all 1 < p < ∞. This is the content of a famous theorem proved by
M. Riesz.
Theorem 2.1. (M. Riesz) For 1 < p < ∞, we have that if h ∈ Lp (m)
h ∈ H p , and there exists Cp > 0 such that
then b

kb
hkp ≤ Cp khkLp (m) .
Equivalently, if f is analytic in D, f (0) ∈ R and Ref ∈ hp , then
f ∈ H p and there exists Cp0 > 0 such that
kf kp ≤ Cp0 kRef kp .
2. THE M. RIESZ THEOREM 57

Proof. When 1 < p ≤ 2 we shall prove the second statement. Let


u = Ref, v = Imf . We use the Green formula in the form (6.12), i.e.
Z Z
1
ϕdm = ϕ(0) + ∆ϕ(z) log dA(z).
T D |z|
to compute for 0 < r < 1, kfr kpp , kur kpp . A straightforward computation
based on the Cauchy-Riemann equations yields
∆|fr |p = p2 |fr |p−2 |fr0 |2 ,
" 2  2 #
∂ur ∂ur
∆|ur |p = p(p − 1)up−2
r +
∂x ∂y
0 2
= p(p − 1)up−2
r |f | .

Thus for 1 < p ≤ 2 we have


p
∆|fr |p ≤ ∆|ur |p ,
p−1
hence by (6.12) and the fact that v(0) = 0 we obtain
p
kfr kpp ≤ kur kpp ,
p−1
and the result follows letting r → 1− .
In the remaining case when 2 < p < ∞, we shall prove the first state-
ment by duality. Let h ∈ Lp (m), and let g ∈ Lq (m), where p1 + 1q = 1.
For 0 < r < 1 we have by Fubini’s theorem,
Z Z Z
g(ζ)
h(rζ)g(ζ)dm(ζ) = h(z)
b dm(ζ)dm(z).
T T T z − rζ

As in the proof of Proposition 1.2 we have for z ∈ T,


Z Z
g(ζ) ζ̄g(ζ)
dm(ζ) = z̄ dm(ζ) = z gb1 (rz),
T z − rζ T ζ̄ − rz̄

where g1 (ζ) = ζg(ζ). Using the first part of the proof, the fact that
1 < q < 2, and the equivalence of the two statements, it follows that
G(z) = z gb1 (z) belongs to H q with
q
kGkqq ≤ kgkqLq (m) .
q−1
Thus by Hölder’s inequality
Z  1/q
q
h(rζ)g(ζ)dm(ζ) ≤ khkLp (m) kGkq ≤ q − 1 khkLp (m) kgkLq (m) ,
b
T
58 4. THE DUAL OF H p

which gives
 1/q
q
k(b
h)r kp ≤ khkLp (m) ,
q−1
and the result follows letting r → 1− . 
Combining the M. Riesz theorem with Proposition 1.2 we obtain im-
mediately a nice description of the dual of H p for 1 < p < ∞.
Corollary 2.1. For every continuous linear functional l on H p , 1 <
p < ∞, there exists a unique g ∈ H q , p1 + 1q = 1, such that
Z
l(f ) = f (ζ)g(ζ)dm(z), f ∈ H p .
T
Moreover, there exists a positive constant ap such that
ap kgkq ≤ klk ≤ kgkq .

Exercise 2. Prove the M. Riesz theorem in the case p = 2 using the


Parseval formula.

3. Thedual of H 1
APPENDIX A

Sequences and families of holomorphic functions

One of the basic applications of the Cauchy formula is the fact that
the set of holomorphic functions in a given region in the complex plane
is closed with respect to uniform convergence on compacts. That is,
given any sequence (fn ) of holomorphic functions in the open set U ⊂ C
which converges uniformly on every compact subset of U to a function
f : U 7→ C, it follows that f is also holomorphic on U . This is an
important result which, in particular, enables us to construct highly
nontrivial examples of holomorphic functions. Some of these are listed
below as exercises.

Exercise 1. Show that the following expressions define holomorphic


functions in the specified domain:


X 1
(i) ζ(z) = z
, Re z > 1, (Zeta function),
n=1
n

Z ∞
dt
(ii) Γ(z) = e−t tz , Re z > 0, (Gamma function),
0 t


X 1
(iii) f (z) = , z ∈ C \ N, (no special name),
n=1
(z − n)2

∞  
1 X 1 1
(iv) ℘(z) = 2 + − ,
z m,n∈Z
(z − m − in)2 (m + in)2
m2 +n2 6=0

z ∈ C \ Z + iZ, (Weierstrass function).

Theorem 0.1. Let (fn ) be a sequence of holomorphic functions on the


open connected set U ⊂ C that converges uniformly on compacts to the
(holomorphic) function f . Let V be an open subset of U such that there
59
60 A. SEQUENCES AND FAMILIES OF HOLOMORPHIC FUNCTIONS

exists n0 ≥ 1 with fn (z) 6= 0 for all z ∈ V and n ≥ n0 . Then either f


is the constant function 0, or f has no zeros in V .

Proof. Suppose f is not identically zero, but has a zero z0 ∈ V .


Choose a small disc ∆ centered at z0 such that ∆ ⊂ V and that f has
no zero on ∂∆. Since for n ≥ n0 fn has no zero in ∆ we have
fn0 (z)dz
Z
= 0, n ≥ n0 .
∂∆ fn (z)
fn0 0
On the other hand, converges uniformly on ∂∆ to ff , so that
fn

fn0 (z)dz f 0 (z)dz


Z Z
1 1
lim = ≥1
n→∞ 2πi ∂∆ fn (z) 2πi ∂∆ f (z)
which gives a contradiction and the theorem is proved. 

Let us now turn our attention to an important property of families of


holomorphic functions that resembles to (relative) compactness of sets
in Rn . We begin with the following general definition.
Definition 0.1. A family F of holomorphic functions on the open
set U ⊂ C is called normal if every sequence (fn ) in F contains a
subsequence (fnk ) that converges uniformly on compact subsets to some
(holomorphic) function f (not necessarily in F!)

The following classical result due to Montel characterizes the normal


families of holomorphic functions.
Theorem 0.2. (Montel). A family F of holomorphic functions on the
open set U ⊂ C is normal if it is uniformly bounded on compacts, that
is, for every compact set K ⊂ U there exists a positive constant CK
such that for all f ∈ F and all z ∈ K we have
|f (z)| ≤ CK .

Proof. The result will follow by a standard diagonalization pro-


cess, once we show that the family F is equicontinuous at every point
z0 ∈ U , that is, for every ε > 0 there exists δ > 0 such that for every
z ∈ U with |z − z0 | < δ and all f ∈ F we have
|f (z) − f (z0 )| < ε.
Assume this claim for a moment and proceed as follows. Choose a
countable dense set {zn , n = 1, 2, . . .} in U (i.e. for every z ∈ U and
every neighborhood V of z there exists n such that zn ∈ V ). Since the
sequences (fn (zj ))n≥1 are bounded in C we can extract a subsequence
A. SEQUENCES AND FAMILIES OF HOLOMORPHIC FUNCTIONS 61

(fn1k ) such that (fn1k (z1 )) converges. This sequence has at its turn a
subsequence (fn2k ) such that (fn2k (z2 )) converges. If we continue this
process indefinitely we obtain subsequences (fnj ) such that (fnj (zl ))
k k
converges for 1 ≤ l ≤ j. If we set nk = nkk then clearly, the subsequence
(fnk ) has the property that (fnk (zj )) converges for all j = 1, 2, . . .. This
is called a diagonalization process.
Let us now show that (fnk ) converges uniformly on compacts. Note
first that if w ∈ U and ε > 0 then by the claim (equicontinuity) there
is δ > 0 such that
|fnk (z) − fnk (w)| < ε
for all k ≥ 1 and z ∈ U with |z − w| < δ. Then, if z, ζ ∈ U satisfy
|z − w| < δ and |ζ − w| < δ, we have
|fnk (z) − fnk (ζ)| ≤ |fnk (z) − fnk (w)| + |fnk (ζ) − fnk (w)| < 2ε.
Fix a point zj (j depends on w) in the above set, such that |zj −w| < δ.
Then there exists k0 depending only on w such that for k, p ≥ k0
|fnk (zj ) − fnp (zj )| < ε.
Putting this together we obtain for all z ∈ U with |z − w| < δ that
|fnk (z)−fnp (z)| ≤ |fnk (z)−fnk (zj )|+|fnk (zj )−fnp (zj )|+|fnp (zj )−fnp (z)| < 3ε
for all k, p ≥ k0 .
Finally, let K ⊂ U be compact ε, δ > 0 as above and cover K by finitely
many sets of the form {z, |z − w| < δ} ∩ U, w ∈ K. Then the above
reasoning shows that there exists k1 ≥ 1 such that for all z ∈ K and
k, p ≥ k1 we have
|fnk (z) − fnp (z)| < 3ε.
Since K was arbitrary we obtain that (fnk ) converges pointwise to some
function f and from the above inequality, letting p → ∞ we get that
for every compact set K ⊂ U there exists k1 ≥ 1 such that for all
z ∈ K and k ≥ k1 we have
|fnk (z) − f (z)| < 3ε,
i.e. (fnk ) converges uniformly on compacts to f .
It remains to prove the claim from the beginning of the proof, i.e. the
equicontinuity of the family F. For z0 ∈ U let ∆0 be a disc of radius
r0 > 0 with ∆0 ⊂ U . By hypothesis, we know that there exists a
constant C0 such that
sup |f (z)| ≤ C0
z∈∆0
62 A. SEQUENCES AND FAMILIES OF HOLOMORPHIC FUNCTIONS

for all f ∈ F. By Cauchy’s formula we have


 
z − z0
Z Z
1 1 1 dζ
f (z)−f (z0 ) = f (ζ) − dζ = f (ζ)
2πi ∂∆0 ζ − z ζ − z0 2πi ∂∆0 (ζ − z)(ζ − z0 )
for all f ∈ F and all z ∈ ∆0 . This leads to the estimate
1 1
|f (z) − f (z0 )| ≤ |z − z0 | sup |f (ζ)| ≤ C0 |z − z0 | .
|ζ−z0 |=r0 |ζ − z| r0 − |z − z0 |
Thus, if ε > 0 is given, choose 0 < δ < r0 such that C0 δ/(r0 − δ) < ε.
Then for |z − z0 | < δ and f ∈ F we have
1
|f (z) − f (z0 )| ≤ C0 δ < ε.
r0 − δ
The claim now follows and the proof is complete. 
This is a famous theorem with a large number of applications in com-
plex analysis. One of these applications is a very useful criterion for
uniform convergence on compacts.
Theorem 0.3. (Vitali) Let (fn ) be a sequence of holomorphic functions
on the open connected set U ⊂ C. Suppose that {fn , n = 1, 2, . . .} is
uniformly bounded on compacts and also that there is a nondiscrete
subset A of U such that (fn ) converges pointwise on A. Then (fn )
converges uniformly on compacts.
Proof. By Montel’s theorem there is at least a subsequence (fnk )
that converges uniformly on compacts to a holomorphic function f
on U . Suppose the sequence itself does not converge uniformly on
compacts to f . Then there exists a compact subset K of U , an ε > 0
and a subsequence (fn0p ) of (fn ) such that
(0.16) sup |fn0p (z) − f (z)| ≥ ε
z∈K

But, again by Montel’s theorem, (fn0p ) contains at its turn a subse-


quence (fn00p ) that converges uniformly on compacts to some holomor-
phic function g on U . Then letting p → ∞ in (1.1) we see that g 6= f .
On the other hand, if A is the set in the statement we have, by hy-
pothesis, for all z ∈ A
g(z) = lim fn00p (z) = lim fn (z) = lim fnk (z) = f (z).
p→∞ n→∞ k→∞

Since A is not discrete we obtain g = f by the identity theorem and


we arrive at a contradiction that proves our result. 

Anda mungkin juga menyukai