CHAPTER

TWENTY
UNIFORM APROXIMATION
BY POLYNOMIALS

Introduction
20.1 Let KO be the interior of a compact set K in the complex plane. (By defini-
tion, KO is the union of all open discs which are subsets of K; of course, KO may
be empty even if K is not.) Let P(K) denote the set of all functions on K which
are uniform limits of polynomials in z.
Which functions belong to P(K)?
Two necessary conditions come to mind immediately: If ! E P(K), then
! E C(K) and! E H(KO).
The question arises whether these necessary conditions are also sufficient.
The answer is negative whenever K separates the plane (i.e., when the com-
plement of K is not connected). We saw this in Sec. 13.8. On the other hand, if K
is an interval on the real axis (in which case KO = 0), the Weierstrass approx-
imation theorem asserts that
P(K) = C(K).

So the answer is positive if K is an interval. Runge's theorem also points in this

direction, since it states, for compact sets K which do not separate the plane, that
P(K) contains at least all those! E C(K) which have holomorphic extensions to
some open set n => K.
In this chapter we shall prove the theorem of Mergelyan which states,
without any superfluous hypotheses, that the above-mentioned necessary condi-
tions are also sufficient if K does not separate the plane.
The principal ingredients of the proof are: Tietze's extension theorem, a
smoothing process invoving convolutions, Runge's theorem, and Lemma 20.2,
whose proof depends on properties of the class g which was introduced in Chap.
14.

386
 UNIFORM APPROXIMATION BY POLYNOMIALS 387

Some Lemmas

20.2 Lemma Suppose D is an open disc of radius r > 0, E c D, E is compact

and connected, n = S2 - E is connected, and the diameter of E is at least r.
Then there is afunction g E H(n) and a constant b, with the following property:
If
Q((, z) = g(z) + (( - b)g2(Z), (1)
the inequalities
100
IQ((, z)1 < - (2)
r

IQ((, z) - z ~ ( I< t;~;~3 (3)

holdfor all ZEn andfor all ( E D.

We recall that S2 is the Riemann sphere and that the diameter of E is the
supremum of the numbers 1Zl - z21, where Zl E E and Z2 E E.

PROOF We assume, without loss of generality, that the center of D is at the

origin. So D = D(O; r).
The implication (d)--+ (b) of Theorem 13.11 shows that n is simply con-
nected. (Note that 00 En.) By the Riemann mapping theorem there is there-
fore a conformal mapping F of U onto n such that F(O) = 00. F has an
expansion of the form

(w E U). (4)

We define
1
g(z) = - F -l(Z) (z En), (5)
a

where F - 1 is the mapping of n onto U which inverts F, and we put

mJrrzg(z) dz,
b = 21 . (6)

where r is the positively oriented circle with center 0 and radius r.

By (4), Theorem 14.15 can be applied to Fla. It asserts that the diameter
of the complement of (Fla)(U) is at most 4. Therefore diam E ~ 41 a I. Since
diam E ~ r, it follows that

(7)
388 REAL AND COMPLEX ANALYSIS

Since g is a conformal mapping of Q onto D(O; II Ia I), (7) shows that

4
I g(z) I <- (z E Q) (8)
r

and since r is a path in Q, of length 2nr, (6) gives

Ibl < 4r. (9)

If ( E D, then I(I < r, so (I), (8), and (9) imply

I QI S; -;4 + 5r(16) 100

r2 < -r-· (10)

This proves (2).

Fix (E D.
If z = F(w), then zg(z) = wF(w)la; and since wF(w)--+ a as w--+ 0, we have
zg(z)--+ 1 as z--+ 00. Hence g has an expansion of the form

1 Ad() ..1. 3 (0
g(z) = z _ ( + (z _ ()2 + (z - 0 3 + ... ( Iz - (I > 2r). (11)

Let robe a large circle with center at 0; (11) gives (by Cauchy's theorem) that

A.2«() = 21 .
1tl
r (z -
Jro ()g(z) dz = b - (. (12)

Substitute this value of A.i() into (11). Then (1) shows that the function

q>(z) = [ Q«(, z) - z ~ (}z - ()3 (13)

is bounded as z--+ 00. Hence q> has a removable singularity at 00. If

z E Q (') D, then I z - (I < 2r, so (2) and (13) give
I q>(z) I < 8r 3 I Q«(, z)1 + 4r2 < I,OOOr2. (14)
By the maximum modulus theorem, (14) holds for all z E Q. This proves (3).
IIII
20.3 Lemma Suppose f E C~(R2), the space of all continuously differentiable
functions in the plane, with compact support. Put

a=!(~+
2 ax
i~).
ay
(1)

Then the following" Cauchy formula" holds:

f(z) = -!
n
fTJR2 (-(afX() de d"
Z
(2)
 UNIFORM APPROXIMATION BY POLYNOMIALS 389

PROOF This may be deduced from, Green's theorem. However, here is a

simple direct proof:
Put cp(r, 0) = I(z + rei~, r > 0, 0 real. If,
= z + rel9, the chain rule gives

-
(af)m = :21 ei9[aar + ;i ao a] cp(r, 0). (3)

The right side of (2) is therefore equal to the limit, as e -+ 0, of

- -1 1i2"
211:.
00

0
-
(acp
ar
+ -i -acp) dO dr.
r ao
(4)

For each r > 0, cp is periodic in 0, with period 211:. The integral of acplaO is
therefore 0, and (4) becomes

1
- 211:
(2"
Jo dO •
1 00
acp 1
ar dr = 211:
(2"
Jo cp(e, lJ) dO. (5)

As e-+ 0, cp(e, 0)-+ I(z) uniformly. This gives (2). IIII

We shall establish Tietze's extension theorem in the same setting in which we
proved Urysohn's lemma, since it is a fairly direct consequence of that lemma.

20.4 Tietze's Extension Theorem Suppose K is a compact subset 01 a locally

compact Hausdorffspace X, and/E C(K). Then there exists an FE Cc(X) such
that F(x) = I(x) lor all x E K.

(As in Lusin's theorem, we can also arrange it so that IIF I x = II !II K .)

PROOF Assume I is real, -1 '5,1'5, 1. Let W be an open set with compact

closure so that K c: W. Put
K+ = {x E K:/(x) ~ t}, K- = {x E K:/(x) '5, -t}. (1)
Then K + and K - are disjoint compact subsets of W. As a consequence of
Urysohn's lemma there is a function 11 E Cc(X) such that 11(X) = on K+, t
11(X) = -t on K-, -t '5,/1(X) '5, t for all x E X, and the support of/1 lies in
W. Thus

I1- Itl '5, ~ on K, I/tl '5, t on X. (2)

Repeat this construction with 1-11 in place of I: There exists an 12 E

Cc(X), with support in W, so that

I1-11 - 12 I '5, (~)2 on K, 1/21'5, t . ~ on X. (3)

In this way we obtain functions In E Cc(X), with supports in W, such that

I1-11 - ... - In I '5, (~)n on K, (4)

390 REAL AND COMPLEX ANALYSIS

Put F = f1 + f2 + f3 + .... By (4), the series converges to f on K, and it

converges uniformly on X. Hence F is continuous. Also, the support of Flies
~W W
Mergelyan's Theorem

20.5 Theorem If K is a compact set in the plane whose complement is con-

nected, iff is a continuous complex function on K which is holomorphic in the
interior of K, and if e > 0, then there exists a polynomial P such that
I f(z) - P(z) I < efor all z E K.

If the interior of K is empty, then part of the hypothesis is vacuously satis-

fied, and the conclusion holds for every f E C(K). Note that K need not be con-
nected.

PROOF By Tietze's theorem, f can be extended to a continuous function in

the plane, with compact support. We fix one such extension, and denote it
again byf.
For any ~ > 0, let ro(~) be the supremum of the numbers

where Z1 and Z2 are subject to the condition I Z2 - z11 S;~. Since f is uni-
formly continuous, we have

lim ro(~) = 0. (1)

6 .... 0

From now on, ~ will be fixed. We shall prove that there is a polynomial
P such that

I f(z) - P(z) I < 1O,OOOro(~) (z E K). (2)

By (1), this proves the theorem.

Our first objective is the construction of a function <I> E C~(R2), such that
for all z

I f(z) - <l>(z) I S; ro(~), (3)

I(13<1>Xz) I < 2i~) , (4)

and

(5)
 UNIFORM APPROXIMATION BY pOLYNOMIALS 391

where X is the set of all points in the support of cI> whose distance from the
complement of K does not exceed (). (Thus X contains no point which is "far
within" K.)
We construct cI> as the convolution ofJwith a smoothing function A. Put
a(r) = 0 if r > (), put

3 ( r2)2 (0 ~ r ~ (), (6)

a(r) = n{)2 1 - {)2

and define

A(z) = a( I z I) (7)

for all complex z.1t is clear that A E C~(R2). We claim that

(8)

(9)

(10)
R2

The constants are so adjusted in (6) that (8) holds. (Compute the integral
in polar coordinates.) (9) holds simply because A has compact support. To
compute (10), express aA in polar coordinates, as in the proof of Lemma 20.3,
and note that iJA/iJ() = 0, IiJA/iJr I = -a'(r).
Now define

cI>(z) = II
R2
J(z - OA«() d~ d'1 = II
R2
A(z - ()J<O d~ d'1. (11)

Since J and A have compact support, so does cI>. Since

cI>(z) - J(z) = II
R2
[f(z - () - J(z)]A«() d~ d'1 (12)

and A«() = 0 if I (I > (), (3) follows from (8). The difference quotients of A
converge boundedly to the corresponding partial derivatives of A, since
392 REAL AND COMPLEX ANALYSIS

A E C~(R2). Hence the last expression in (11) may be differentiated under the
integral sign, and we obtain

(a<l>)(z) = ffR2
(aA)(z - OJ«() de d'1

= ff J(z - ()(aA)«() de d'1

R2

= ffR2
[f(z - () - J(z)](aA)(() de d'1. (13)

The last equality depends on (9). Now (10) and (13) give (4). If we write
(13) with <1>" and <1>, in place of 13<1>, we see that <I> has continuous partial
derivatives. Hence Lemma 20.3 applies to <1>, and (5) will follow if we can
show that 13<1> = 0 in G, where G is the set of all z E K whose distance from
the complement of K exceeds b. We shall do this by showing that

<l>(z) = J(z) (z E G); (14)

note that aJ = 0 in G, since J is holomorphic there. (We r.ecall that a is the

Cauchy-Riemann operator defined in Sec. 11.1.) Now if z E G, then z - ( is in
the interior of K for all ( with I (I < b. The mean value property for harmo-
nic functions therefore gives, by the first equation in (11),

ft! f2"
<l>(z) = Jo a(r)r dr Jo J(z - rei') dO

= 2nJ(z) f a(r)r dr = J(z) ff

R2
A = J(z) (15)

for all z E G.
We have now proved (3), (4), and (5).
The definition of X shows that X is compact and that X can be covered
by finitely many open discs D 1, ..• , Dn , of radius 2b, whose centers are not in
K. Since S2 - K is connected, the center of each Dj can be joined to 00 by a
polygonal path in S2 - K. It follows that each Dj contains a compact con-
nected set Ej , of diameter at least 2b, so that S2 - EJ is connected and so
that K n E j = 0.
We now apply Lemma 20.2, with r = 2b. There exist functions
 UNIFORM APPROXIMATION BY POLYNOMIALS 393

OJ E H(S2 - E j ) and constants bJ so that the inequalities

50
IQj(, z) I < --g-' (16)

IQj(, z) - 1
z_ (
I< I4,0000
z _ (13
2
(17)

hold for z ¢ E j and ( E Dj , if

(18)

Let n be the complement of E1 u ... u En. Then n is an open set which

contains K.
Put Xl = XII D1 and Xj = (X 11 D j ) - (Xl U ..• u X j - 1), for
2 :s; j :s; n. Define

R((, z) = Qj(, z) (19)

and

F(z) = ; II
x
(a<l>X()R((, z) de drt (z En). (20)

Since

F(z) = jt ;II XJ
(a<l»(()Qj(, z) de drt, (21)

(18) shows that F is a finite linear combination of the functions OJ and OJ.
Hence F E H(n).
By (20), (4), and (5) we have

I F(z) - <l>(z) 1< 2~0) II I

x
R((, z) - z ~ ( Ide drt (z En). (22)

Observe that the inequalities (16) and (17) are valid with R in place of QJ if
( E X and ZEn. For if ( E X then ( E Xj for some j, and then R((, z) =
Q/J, z) for all ZEn.
Now fix ZEn, put ( = z + peifJ, and estimate the integrand in (22) by (16)
if p < 40, by (17) if 40 :s; p. The integral in (22) is then seen to be less than the
sum of

2n Jof4d (50 1)p dp =

--g- + p 808no (23)
394 REAL AND COMPLEX ANALYSIS

and

2n 1 00

4.1
4,00002
p
3 p dp = 2,OOOno. (24)

Hence (22) yields

I F(z) - CIl(Z) I < 6,OOOw(0) (z En). (25)
Since F E H(n), Ken, and S2 - K is connected, Runge's theorem
shows that F can be uniformly approximated on K by polynomials. Hence
(3) and (25) show that (2) can be satisfied.
This completes the proof. IIII
One unusual feature of this proof should be pointed out. We had to prove
that the given function f is in the closed subspace P(K) of C(K). (We use the
terminology of Sec. 20.1.) Our first step consisted in approximating f by <1>. But
this step took us outside P(K), since <I> was so constructed that in general <I> will
not be holomorphic in the whole interior of K. Hence <I> is at some positive
distance from P(K). However, (25) shows that this distance is less than a constant
multiple of w(o). [In fact, having proved the theorem, we know that this distance
is at most w(o), by (3), rather than 6,000 W(o).J The proof of (25) depends on the
inequality (4) and on the fact that a<l> = 0 in G. Since holomorphic functions cp
are characterized by acp = 0, (4) may be regarded as saying that <I> is not too far
from being holomorphic, and this interpretation is confirmed by (25).

Exercises
I Extend Mergelyan's theorem to the case in which 8 2 - K has finitely many components: Prove
that then every f e C(K) which is holomorphic in the interior of K can be uniformly approximated on
K by rational functions.
2 Show that the result of Exercise 1 does not extend to arbitrary compact sets K in the plane, by
verifying the details of the following example. For n = 1,2,3, ... , let Dn = D(lXn; rn) be disjoint open
discs in U whose union V is dense in U, such that :Ern < 00. Put K = -0 - V. Let rand Yn be the
paths
os; t s; 21l,

and define

L(f) = If(Z) dz - t 1.f(Z) dz (fe C(K».

Prove that L is a bounded linear functional on C(K), prove that L(R) = 0 for every rational function
R whose poles are outside K, and prove that there exists anf e C(K) for which L(f) -# O.
3 Show that the function g constructed in the proof of Lemma 20.2 has the smallest supremum norm
among a1lf e H(n) such that zf(z)--+ 1 as z--+ 00. (This motivates the proof of the lemma.)
Show also that b = Co in that proof and that the inequality I b I < 4r can therefore be replaced by
I b I < r. In fact, b lies in the convex hull of the set E.