Revista Matematica Iberoamericana

Vol. 11, N.o 2, 1995

A remark on gradients
of harmonic functions
Wensheng Wang

Abstract. In any C 1 domain, there is nonzero harmonic function C 1

s

continuous up to the boundary such that the function and its gradient
on the boundary vanish on a set of positive measure.

1. Introduction.
In this note, we will extend Bourgain and Wol's result in 2] into
the general C 1 domain of R .
s d

Theorem. If D is a C 1 domain in R with s > 0 and d 3, then

s d

there is a harmonic function u : D ! R which is C 1 up to the boundary

and such that
jfQ 2 @D : u(Q) = 0 ru(Q) = 0gj > 0 :

The idea for the proof of this theorem follows from the argument
in 2]. We also need to use Aleksandrov-Kargaev function (see 1])
as our basic constructing factor. Since there is no reason to apply
directly Alexandrov's result to the arbitrary domain, we have to work
on the \almost at" domain rst and then get our nal result by Kelvin
transformation.

227
228 W. Wang

Notation. Let A and M be positive large numbers. Let  be a M

collection of C 1 domains in R which are of the form

s d

 = fX = (x x ) 2 R : x > '(x) x 2 R ;1 g

d
d
d
d

such that ' is some C 1 function on R ;1 satisfying '(0) = 0, r'(0) =

s d

0 and
kr'k1 + kr'k  e; Cs
M
and r2 '(x) = 0
when jX j > 1.
When X = (x x ) 2 @ , we denote by N the normal vector of @ 
at X , and n the normal vector of R ;1 at (x 0). By the assumption
d X
d

of , we know that jN ; n j  e; for any X = (x x ) 2 @ . We

x
M
X x d

use notation r u to denote the tangent gradient of u on @ . Finally,

T

we usual use B to denote the ball in R and Q to denote the cube on

d

R ;1 . If Q is the cube on R ;1 , then 'Q denotes its image on @  by

d d

'. C always denotes an absolute constant.

2. Several lemmas.
Lemma 1. Suppose ' is a C 1 function on R ;1 and  = fX =
s d

(x x ) 2 R : x > '(x) x 2 R ;1 g is a C 1 domain. Let G(X Y ) be

d
d
d
d s

the Green's function of . Then for any X , Y 2 @ , we have

(1) d d G(X Y )  C :
dN dNX Y jX ; Y j d

Proof. When  is a bounded domain, the result is known (see 4],

6]) but we do not nd a good reference for the proof. When  is
unbounded, it is not true in general. So we would like to give a proof
for such special case and one will see the proof still works for bounded
domains with a tiny correction.
Claim. Let X 2 @  and R > 0. If u is a harmonic function in
 \ B (X R), juj  1 on  \ @B (X R) and u = 0 on @  \ B (X R),
then jru(Z )j  C=R for all Z 2  \ B (X R=2) .
 A remark on gradients of harmonic functions 229
Proof of Claim. We may assume that X = 0. Let D =  \ B (0 R)
and D; =  \ B (0 R). Consider a map  : D; ! D dened by
c

(z z ) = (z 2 '(z) ; z ). Then the function u   solves

d d

div A0 r(u  ) = 0 in D; 
where
2 (r')?
 

A0 = ((0)?);1 (0 );1 = I

2 r' 1 + 4 jr'j2
is an elliptic matrix (see e.g. 3]). Let ! = u in D, ! = ;u   in D; .
Then with A = I in D and A0 in D; , ! solves div Ar! = 0 in B (0 R)
in the weak sense, because the function ! is an odd `reection' of u.
By the assumptions for the functions u and ', we know j!j  1 and
kAk  C e; 1 . If we dene functions v(Y ) = !(RY ) and B (Y ) =
Cs
M

minfR;  1g A(RY ) in B (1), then v solves div B rv = 0 in B (1) and B

s

has uniform C bound. So by 4, Lemma 3.1], we have jrv(Y )j  C .

s

Hence, jr!(Z )j  C=R. This proves the claim.

Now let us x X Y 2  and let R = jX ; Y j. If we apply the
claim to the function R ;2 G(Z1 Z2) for Z2 2 B (Y R=100) with Z1 2
d

B (X R=100) xed, then we have jr 2 G(Z1  Z2)j  C=R ;1 for all
Z
d

Z1 2 B (X R=100) and Z2 2 B (Y R=100). Similarly, if we apply the

claim to the function R ;1 r 2 G(Z1  Z2) for Z1 2 B (X R=100) with
d
Z

Z2 2 B (Y R=100) xed, then we get jr 1 r 2 G(Z1  Z2)j  C=R for

Z Z
d

all Z1 2 B (X R=100) and Z2 2 B (Y R=100). Finally let Z1 go to X

and let Z2 go to Y , to conclude.
Lemma 2. Let  be as in Lemma 1 with 0 2 @ . Suppose u is a
harmonic function in  which is C 1 up to the boundary. Assume the
s

restriction function of u onto @  is supported in @  \ B (0 1). Then

for all X 2 @  with jX j  2 ,

(2) d u(X )  C kuk :

dN X
C 1s @
( )

When X 2 @  with jX j 2, we have

d u(X )  C jX j; kuk d
:
dN X
C @ ( )
230 W. Wang

Proof. The rst result is a well known fact when  is either a bounded
domain or the upper half space. Here we would like to give a short proof
without using layer potential theory. Since u(Y ) has compact support
on @ , for X 2 ,

u(X ) =
Z
d G(X Y ) u(Y ) d (Y ) :
@  dN Y

If X 2 @  with jX j 2, then by (1) we have

d u(X ) = Z d2 G(X Y ) u(Y ) d (Y )

dN X  dN dZ @ X Y

 C kuk ( )
Z
1 d (Y )
jX ; Y j C @ d

Y 2 @  j j1
Y

 C jX j; kuk d
C @ ( ) :
This proves (3). Now let us x X 2 @  with jX j  2. Choose a
function
2 C01 (R ) with
(Z ) = 1 when jZ j  10 and
(Z ) = 0
d

when jZ j 20. Let us write

u (Y ) = u(Y ) ; u(X )
(Y ; X ) ; hr u(X ) Y ; X i
(Y ; X ) 
 T

which is supported in @  \ fjY j  15g and bounded by

kr uk T ( )
Cs @ jY ; X j1+  s
when jY ; X j  8 :

So we have

u(Z ) =
Z
d G(Z Y ) u (Y ) d (Y )
 dN

@ Y

+ u(X ) dNd G(Z Y )

(Y ; X ) d (Y )
Z

 @ Y

+
Z
d G(Z Y ) hr u(X ) Y ; X i
(Y ; X ) d (Y )
 dN
T
@ Y

= u1 (Z ) + u2 (Z ) + u3 (Z ) :
 A remark on gradients of harmonic functions 231
For u1,
d u (X ) = Z d2 G(X Y ) u (Y ) d (Y )
dN 1
X  dN dN @ X Y


C
Z
jX ; Y j1+ kr uk s

d (Y )
jX ; Y j d T Cs @( )
X 2 @  j ; j
X Y < 8
Z
1
+C jX ; Y j kuk d C1 @ ( )
X 2 @  j ; j8
X Y

 C kuk C 1s @ ( ) :
Notice that u2 is a bounded harmonic function whose boundary value
is 1 on @  \ B (0 4). So apply the claim in Lemma 1 to the function
u(X ) ; u2(), we have jru2(Z )j  C kuk1 for Z 2  \ B (0 2) so that
d u (X )  C kuk :
dN 2 X
C 1s @
( )

Finally consider the harmonic function u3 (Z ) ;hr u(X ) Z ; X i which T

is bounded in  \ B (0 4) and whose boundary value is zero on @  \

B (0 4). So again by using the claim in Lemma 1, we have
d u (X )  C kuk  C kuk 
dN 3 X
C1 @( ) C 1s @ ( )

since hr u(X ) Z ; X i is linear. This proves Lemma 2.

T

Now let a and " be two positive numbers. Let

a
" + x =a :
E (X ) = ;a jX=a d
"
+ "e j d
d

We denote by n the normal vector of R ;1 in R at x 2 R ;1 .

x
d d d

Lemma 3. We have the following properties for E (X ) : a

"

(4) rE
i a
(X )  C a; +1 min i
n
"; ; +1 
d i
X ; ; +1 o d i

"
a
232 W. Wang

for all X 2  and i = 0 1 2 .

Z 
d E (X ) ; 1 d (X )
1 + dN a
p 

"
'Q (0 b)

(5) ;
Z 
d 
1 + dn E (x) ; 1 dx
a
p

"
Q(0 b)

 C e; M ;1 a ;1 pM d d

if "  e; . M

(6)
Z 
d E (X ) ; 1 d (X )  ;  ; C a ;1 
1 + dN a
p   
d

'Q (0 b)
"
M
if p > 0 is small but independent of ", M , a . Here b = aM in (5) and
(6), and  is an absolute small positive number.
Proof. (4) follows directly from the calculation. After a change of
variable, the left hand side of (5) is the integral over a subset fjxj < bg
of R ;1 of the following integrand
d

(j1+ hrE (X ) N ij ; 1) (1+ jr'(x)j2)1 2 ; (j1+ hrE (x) n ij ; 1) :

a
" X
p = a
" x
p

If we introduce a term ;j1 + hrE (x) n ij (1 + jr'(x)j2)1 2 and sub-

a
" x
p =

tract it, then the integrand becomes

(j1 + hrE (X ) N ij ; j1 + hrE (x) n ij ) (1 + jr'(x)j2)1 2
a
" X
p a
" x
p =

+ j1 + hrE (x) n ij (1 + jr'(x)j2)1 2 ; 1)

a
" x
p =

; (1 + jr'(x)j2)1 2 ; 1) =

= I + II + III :
Up to multiplying a constant, I is bounded by
jhrE (X )N i ; hrE (x) n ij
a
" X
a
" x
p

 jhrE (X ) ; rE (x) n ij + jhrE (X ) N ; n ij

a
"
a
" x
p a
" X x
p

= I1 + I2 :
For I1, since X = (x '(x)), by (4),
I1 = jhrE (X ) ; rE (x) n ij a
"
a
" x
 A remark on gradients of harmonic functions 233

 a;2+1 xa
; ;2+1
jX ; xj
d

 a;1 xa
; ;1
j'(x)j
d

 e; xa :
M
; d

So Z

jI1 j dx  M (1; );1 e; a ;1 :

d p pM d

j jx b

Since jN ; n j  C kr'k1  C e; , again by (4),

X x
M

jI2j dx  C e; x ; dx = C M (1; );1 e; a ;1 :

pM
Z
dp
d p pM d

j jx b j j a x b

By a similar method, using (4) and the assumption of ',

Z

jIIj dx  C
Z 
x
1+ a
; ;
e dx
d p
M

j j x j j
b x b

 C a ;1 e; M (1; );1 : d M d p

And it is trivial to get

Z Z

jIIIj dx  C e; M
dx  C e; a ;1 M ;1 :
M d d

j j
x b j jx b

So combining those estimates of the integrations for I = I1 + I2 , II and

III, we have that the left hand side of (4) is bounded by
C a ;1 (M (1; );1 e;
d d p pM
+ M (1; );1 e;
d p pM
+ M ;1 e; )
d M

 C a ;1 M ;1 e; : d d pM

Let us now prove (6). From 1] or 2], we know that if p > 0 is
small enough then for all small ",
Z 
d E (x) ; 1dx  ; a ;1 
1 + dn a
p
d

R ;1
"
d

with a small absolute positive number . From this integral, we easily

get
Z 
d 
1 + dn E (x) ; 1 dxa
p

jj
"
x <b
234 W. Wang

 ; a d;1 +
Z
d E (x) ; 1 dx
1 + dn a
p

j Zj
"
x >b

(7)  ; a ;1 + C d
x ; ;1 dx d

jj x >b a
 ;( ; C M ) a ;1
; 2 d

by (4). So combine (7) and (5). We have that the left hand side of (6)
is less than or equal to

C  a ;1
;( ; C M ;2) a ;1 + C e; M ;1 a ;1 ;  ; M
d pM d d d

when M is large enough independently of a and " .

We state our main lemma.
Lemma 4. If p > 0 and 1=M > 0 are small enough, then for any
 2  and " > 0 and any cube Q = Q(0 l) on R ;1 with l < 1 there d

exists a harmonic function F which is C 1 up to  with supp F j 

M
Q Q
" @
'B (0 "l) such that
"

(8)
Z
d F (X ) d (X )  e;2 j'Qj 
1 + dN Q
p
p
"
'Q

Q
n
(9) rF (X )  C min ";  e;( ;1 2) d d = M

X ; + X ; +1 2 o 
d d =

"
l l
if X 2 @ , where  > 0 is such that
e;2 = 1 ; 2 e;( ;1) :
p d M

Proof. Let a = e; M
l and E (X ) be as above. Dene
a
"

I = fx 2 R ;1 : 2 ;1 l  jxj < 2 +1 lg
j
d j j

and  a cut-o function in R with supp  Q(0 2 +1l) n Q(0 2 ;1l)

j
d
j
j j

and jr  j  1=(2 l) , for i = 0 1 2 and for

i
j
j i
j = 0 1 : : : such that
P
 (X ) = 1 for jX j > 1. Let  = 1 ;  (X ). Denote by F
j
P
j
Q

the harmonic extension into  of  E (X )j  and Q the harmonic

j "
a a
" @ j "
 A remark on gradients of harmonic functions 235
extension into  of  E (X )j  . Let Q = Q . Then E = P
a a a a
j @

F +Q .
" " j j " "
a a
" "

An easy computation and (4) of Lemma 3 imply that

kQ a
j "
k1  C a (2 l); +1  d j d

(10) kr Q T
a
j "
k1  C a (2 l);  d j d

kr Q T
a
j "
k  C a (2 l); ; :
Cs
d j d s

If we let u(Y ) = Q (2 lY ), then the restriction function of u onto @ 

a j

is supported in @  \ B (0 1) and kuk 1  C a (2 l); +1 by the above

j "
d j d
C s

estimates. So apply Lemma 2 to u. When X 2 @  with jX j  2 +2 l , j

we get
d Q (X )  C a (2 l); :
a d j d

dN X
j "

When X 2 @  with jX j > 2 +2l , j

d Q (X )  C a jX j;  C a jX j; +1 2 (2 l);1 2 :
a d d d d = j =

dN X
j "

Hence
d Q (X )   X + X  d Q (X )
a a

dN "

j j 2 +2 j j 2 +2 dN
X < j l X > j l
j "

a (2 l); + C a (2 l);1 2jX j; +1 2

X X
C d j d d j = d =

j j
X < j +2 l
2 j j 2 +2 X > j l

(11) Ca d
jX j; + C a jX j; +1 2 l;1
d d d = = 2

C a X ; X ; +1 2 
d
+ a
d =


since a = e; l . We notice that if jX j  4 l , the process of estimates

M

above also give

(12) d Q (X )  C e; :
a M

dN "
236 W. Wang

Now we estimate (7). Let b = Ma .

Z 
d F (X )j ; 1 d (X )
1 + dN Q p


"
'Q

=
Z 
d (E (X ) ; Q (X )) ; 1 d (X )
1 + dN a a
p 

" "

'Q (0 b)

+
Z 
d F (X ) ; 1 d (X )
1 + dN Q
p 

"

'Q 'Qn (0 b)
(13) Z 
d E (X ) ; 1 d (X ) 

p
1 + dN a
"

'Q (0 b)

+
Z 
d 
1 + dN Q (X ) ; 1 d (X ) a
p

"

'Q (0 b)

+
Z 
d F (X ) ; 1 d (X )
1 + dN Q
p

"

'Q 'Qn (0 b)

by triangle inequality. When jxj > b, the integrand in the last integral
is bounded by C jx=aj; 2, by (4) and (11). So d=

(14)
Z 
d F (X ) ; 1 d (X )  C a ;1 M ;1 2
1 + dN Q
p 
d =
"

'Q 'Qn (0 b)

Since p < 1, by Holder inequality,

Z
d Q (X ) d a
p

) dN
"
'Q(0 b


X
Z
d Q (X ) d a
p

(0 ) dN
j "
'Q b
j

Cb (1;p=2)(d;1)
X
Z
d Q (X ) 2 d  2 :
a
p=

j
'Q (0 b) dN j "

Notice that ' is a Lipschitz graph with uniform bound. Theorem

2.2.6 of 5] shows that
Z
d Q a
2
d  C
Z

rQ a
2
d :
'  dN j "
'
T j "
 A remark on gradients of harmonic functions 237
So by (10),
Z
d Q (X ) d
a
p

'Q(0 b) dN "

Z
X 2 p=2
(1; 2)( ;1)
Cb p= d
rQ T
a
j " d
j
' 
Z
X p=2
(1;p=2)(d;1)
Cb ja (2 l); j2 d
d j d

(15) j
'Ij

 C a b(1; 2)(d;1)
(2 l); +p(d;1)=2
X
dp p= j dp

 C a ;1 (al) ( +1) 2 M (1;

d p d = p= 2)(d;1)

= C a ;1 e;
d pM d ( +1)=2
M (1; p= 2)(d;1)
:
Now apply (6) to the rst term in the left hand side of (13) and combine
(13), (14), and (15). The left hand side of (13) is less than or equal to

C  a ;1 + C a ;1 M ;1 2 + C a ;1 e; ( +1) 2 M (1; 2)( ;1)
; ; M d d = d pM d = p= d

 ; 2 a ;1 = ; 2 e; ( ;1) j'Qj : d M d

Hence
Z
d F (X ) d  1 ;  e;
1 + dN Q
p 
M d ( ;1)

j'Qj = e;2 j'Qj  p

'Q
"
2
with some  > 0.
Now we turn to prove (8). First we have
jrF (X )j  jrQ (X )j + jrE (X )j
a
"
a
"
a
"

X
C a + a
; X ; +1 2  d d =

 C e; ( ;1 2) Xl + Xl
 ; ; +1 2 d d =
M d
) =

by (4) and (11). In order to bound jrF (X )j by C "; , we notice thata d

when jX j < l " < l ,

"

d Q (X )  C a

dN "
238 W. Wang

is bounded by (12) and of course is less than C "; . The remaining d

part follows directly from calculation and (4).

Corollary 5. Let p, M and " be as in Lemma 4. Let  2 2 . For M

any X = (c  '(c )) 2 @  and s0 > 0, Q B (c  s0) is a cube in

R ;1 . For any function I : @  ! R with supp I ' such that
Q Q Q Q
d

jI (X );1 j kI () ; I (X )k1  81  e; ( ;1) 

Q Q
M d

there exists a harmonic function F in  which is C 1 up to  such

Q

that supp F 'B (c  " l(Q)) and

"
Q
" Q

I (X ) + I (X ) dF
Z Q
;2
dN (X ) d (X )  e jI (X )j j'Qj 
p
(16) Q
" p
Q
p

'Q

jrF (X )j
Q

(17)
"

 C min ";
n
d
 e; (
M d ;1 2)
=

X ; X ; + X ; X ; +1 2 o :
Q
d
Q
d =

l(Q) l(Q)
Proof. Under a new coordinate system such that X is the new origin
and the tangent space of @  at X is the new R ;1 ,  2  . Then
Q
d
Q M

this corollary follows directly from Lemma 4.

Let us rst prove a weaker version of our theorem.
Theorem 1. Let  be as in the Corollary 5. Then for any small
number s0 > 0, there is a harmonic function u :  ! R which is C 1 up
to  and such that
jfX 2 @  : jX j  s0 u(X ) = 0 ru(X ) = 0gj > 0 :

We will give the recursive construction. Let Q0 = Q0 (0 s0)

R ;1 . Let f g1
d
1 be a sequence of small numbers such that 
n
;1 2
Z which are chosen by induction later. Let fK g and f" g be two
n

n n

sequences of numbers which are decided later with K % +1 and n

" & 0. a will be a large universal number also decided later.

 A remark on gradients of harmonic functions 239
Let < be the collection of ;( ;1) cubes of side  in R ;1 whose
n n
d
n
d

union is Q0 . Let = be a subset of < such that

n n

du Z 1
;
dN d (X )  A e 
p =p
(18) n n

'Vn

where V = fQ : Q 2 = g and u is the C 1 function in  dened

S
s
n n n

by induction later. When n = 1, let 0 = s0 and =0 = fQ0g and u0

a harmonic function in  such that u0 j  2 C01 ('Q0=1000). Suppose @
s

we have = and u . Let  +1 with  +1 = 2 Z be such that  +1

n n n n n n

small enough and decided in the following lemmas. Let = +1 < +1 n n

be such that Q 2 = +1 satises n

(19) Q0 2 = for all Q0 2 < with Q Q0 

n n

(20)

1 Z
du d (X )1  K
n
p =p
e; : n

j'Qj 'Q dN n +1

Now let us dene

u (X ) = u ( X ) +
X du (X ) F n Q
(X ) 
2= +1 dN
+1
n n Q "n+1
Q n

where F Q
"n+1
is as in Corollary 5.
Lemma 6. For X 2 @  ,
(21)
X
jrF Q
(X )j  C e; M d ( ;1 = 2)

n +1
+

 +1 1 2 n
=

Q2= +1
"n+1
 
j; j
n
x cQ >

if  > C  +1 .n

(22) jru +1 (X ) ; ru (X )j  C K
n n n +1 e; ";+1 
n
n
d

if  +1 small enough.
n

Proof. By (17) of Corollary 5, the left hand side of (21) is bounded

by
X
e; (
M d ;1 = 2)

X ; c ; + X ; c ; +1 2 
Q
d
Q
d =

j;j
x c Q >

 +1n  +1 n

 C e; ( ;1 2)  +1 +  +1
  1 2 

=
M d = n n
240 W. Wang

if  > C  +1 . Let us prove (22) now. Since du =dN is continuous,

n n

after making  +1 small (20) will imply that

n

du (X 0 )  2 K e; 
n n

dN +1 Q n

for all Q0 2 = +1 . For X 2 @ , there is at most one Q 2 = +1 such

that for any other Q0 =
6 Q, Q0 2 = +1 , jx ; c 0 j >  . Then the left
n n

n Q

hand side of (15) is bounded by

X du (X 0 ) jrF 0
(X )j + du
dN (X ) jrF (X )j
n Q n Q

0
Q 2=n+1
dN Q "n+1 Q "n+1

0
Q 6=Q

 2 K +1 e; (C + C";+1 )
n
n
n
d

 C K +1 e; ";+1 
n
n
n
d

where we used (21) and (17) in the rst inequality.

The following lemma says our process in construction of u is pos- n

sible, i.e. (18) is true.

Lemma 7. There exists a large universal constant A such that
Z
du +1 d 1  A e;
n
p =p
n( +1)
:
'Vn+1 dN
Proof. As in 2], we rst state a claim as follows. One may nd the
proof in 2].
Claim. If  n +1 is small enough, then for all X 2 'Q, Q 2 = n +1 ,
X du (X 0 ) jrF
n Q
0
(X )j
0
Q 2=n+1
dN Q "n+1

0
Q 6=Q

 C e; M d ( ;1 2)
=
du (X ) + e;
n (
M d ;1 = 2)
e;4 n( +1)
:
dN

P
We would like to pointPout the idea. Divide the sum into two parts:
n+1 j ; j
< XQ +1
and j ; j +1 . For the rst term, use (21)
X <Ln XQ X Ln
 A remark on gradients of harmonic functions 241
and notice that if  +1 is small then the numbers du (X 0 )=dN are
n n Q

close to du (X )=dN up to an error term. For the second term, when

n xed, du (X 0 )=dN are bounded uniformly in Q0 6= Q. For the
n

n Q

remaining part, apply (21) again and let L be big.

Now let us take

I = Q 2 = +1 : du
n o
(X ) > e ;4 n ( +1)
n

dN n Q

and
! = = +1 n I : n

Let X 2 'Q for some Q 2 = +1 . Denote n

J (X ) = du du (X 0 ) d F 0
X

dN (X ) + (X ) :
n n Q

0 2= +1 dN dN Q "n+1

0 6=nQ
Q
Q

Then
du +1 (X ) = J (X ) + du (X ) d F (X )
n n Q

dN dN dN +1 Q "n

by the denition of u +1 . If Q 2 I , then after making  +1 small,

n n

J (X ) ; du
dN (X )n
 C e ; Q
du (X )
M d

dN
( ;1 = 2) n
Q

1 du
 8  e; ( ;1) dN (X )
M d n
Q

by claim and by letting M large. So apply Corollary 5,

Z
du +1 (X ) d (X )  e;2 j'Qj du (X )
n
p
p n
p

dN dN Q

(23) 'Q
Z
du d (X ) 
 e ;3 2
p
p= n

dN 'Q

if 
n +1 small again. When Q 2 !, write

J = du
dN (X ) + J1 
n
Q where jJ1j  C e; (
M d ;1 = 2)
e;4 ( +1)
n
242 W. Wang

by claim. So apply Corollary 5 again,

Z
du +1 (X ) d (X )
n
p

'Q dN
(24) 
 e;2 e;4 ( +1) + C e;4
p n p n ( +1)p

j'Qj
 C e;4 ( +1) j'Qj :
n p

Combine (23) and (24). By induction, we have

Z
du +1  e;3
n
p
p = 2
Z
du + C e;4
n
p
n ( +1)p
X
j'Qj
'Vn+1 dN dN Q 2Vn+1 \
2 \
'Q
Q Vn+1 I

 A (e;
p p = 2
+ C A; e;3 ( p p n +1)
) e; ( +1)
p n

A e (
p p n +1)

if A is large and independent of n .
Proof of Theorem 1. Choose K and " such that n n

"; K +1 e;  C0 < +1
X
d n
n +1 n

and
X
K ;+1 + " ;+11 
p d
s0;1 : d

n n
1000 C
Then the (21) of Lemma 6 implies that
jru +1 ; ru j  C K +1 ";+1 e;  C C0 < +1
X X
d n
n n n n

i.e. u = u is C 1 up to  . On the other hand, by the denition of

P
n

u and u +1 ,n

X
jfX 2 'Q0 : u(X ) 6= 0gj  jfX 2 'Q0 : u +1 (X ) 6= u (X )gj n n

n
X X
 j'(B (c ) " Q n +1 l(Q))j
n Q2= +1 n

;( ;1) ( ) ;1
X
 n+1
d
n +1 "n +1
d

" ;+11
X
=C d
n

s ;1 d

 1000 :
0
 A remark on gradients of harmonic functions 243
In order to estimate the gradient term, we notice that if a point X 2
'(Q0) is in some 'Q 's with innite many n, then we know by (20)
n

and the continuity of du=dN that du(X )=dN = 0 . So again by (20)

and (18),
n
du (X ) 6= 0o  X j'V n 'V
X 2 'Q0 : dN j
n n+1
n
X X
= j'Qj
n Q2 n
'Vn+1 'Vn


X
K; p
e; np
X
Z
un
p

n
n +1
Q2 n
'Vn+1 'Vn
'Q dN

X
K ;+1 e;p p n
Z
du n
p

n
n
'Vn dN
A K ;+1
X
 p
n
p

s ;1d

 1000 
0

by Lemma 7, this is because when Q 2 < +1 n= +1 , there exists Q0 2 =

such that Q0  Q but
n n n

Z
du > K
n
p
p
e; p n

'Q dN n +1

by denition. And so

j'Qj < K ; p
e; p n
Z
du : n
p

n+1
'Q dN
Finally we get
1 j'Q j > 0 :
jfX 2 'Q0 : u(X ) = 0 ru(X ) = 0gj 100 0

3. Proof of Theorem.
Case 1: D is a bounded domain. Then we can assume (0 0) 2 @D and
R ;1 is the tangent space of @D at (0 0) and D R + . So we may
d d
244 W. Wang

construct a domain  =  = f(x x ) : x > '(x)g where ' 2 2

' d d M

and D  and @D \ @   'Q0 for some Q0 = Q0 (0 s0) with s0 > 0.

Then apply Theorem 1 to , we get a harmonic function U in  which
is C 1 up to  such that
jfX 2 'Q0 : U (X ) = 0  rU (X ) = 0gj > 0 :
Now let u = U j , then this is the desired u .
D

Case 2: D = R n B for some bounded C 1 domain B . We assume

d s

that (0 0) 2 B . Let T : X ! X=jX j2 be the Kelvin transformation

and B~ = R j . Apply the result in Case 1 to B~ and get a harmonic
d

function U . Then u(X ) = U (TX )=jX j ;2 is the desired function for

TB
d

our domain D.
Case 3: D is a general C 1 domain. It is easy to nd a domain B D
s c

such that @B \ D contains some \ball" on @D. Then by the Case 2,

there is a harmonic function U in R n B which is C 1 up to @B and
d

jfX 2 \ball" : U = rU = 0gj > 0. Then this u(X ) = U j is needed.

D

Some Remarks. 1. The theorems are true for C 1 -Dini domains. The
proof follows our argumens with minor corrections.
2. It is not hard to see that for every " there exists a harmonic
function which is C 1 up to the boundary and such that
j@D n fX 2 @D : u(X ) 6= 0 or ru(X ) 6= 0gj  " :
3. We do not know if the theorem is true or not for Lipschitz
domains. In fact, our method does not work even for C 1 domains (and
even if we do not need the restriction u = 0).
4. We may also prove Lemma 4 by using the potential layer theory
as in 2]. But again this method does not work even for C 1 domains.

Acknowledgments. I would like to express my thanks to my teacher

and advisor, Proesor T. Wol, for bringing this problem to my atten-
tion as well as for explaining to me the idea in his paper 2].
 A remark on gradients of harmonic functions 245
References.
1] Aleksandrov, A. and Kargaev, P., private communication with T. Wol.
2] Bourgain, J. and Wol, T., A remark on gradients of harmonic functions
in dimension d 3. Colloq. Math. 40/41 (1990), 253-260.
3] Caarelli, L., Fabes, E. and Kenig, C. E., Completely singular elliptic
harmonic measures. Indiana Univ. Math. J. 30 (1981), 917-924.
4] Gruter, M. and Widman, K., The Green function for uniformly elliptic
equations. Manuscripta Math. 37 (1982), 303-342.
5] Kenig, C. E., Elliptic boundary value problems. Beijing Lectures in Har-
monic Analysis ed. E. M. Stein, Ann. of Math. Studies 112, Princeton
Univ. Press, 1986.
6] Krantz, S., Function theory of several complex variables. John Wiley
and sons, 1982.

Recibido: 30 de mayo de 1.993

Revisado: 18 de septiembre de 1.993

Wensheng Wang
Department of Mathematics
University of Southern California
Los Angeles CA 90089, U.S.A.
wang@math.usc.edu