Appl. Comput. Harmon. Anal.

27 (2009) 100–116

Contents lists available at ScienceDirect

Applied and Computational Harmonic Analysis

www.elsevier.com/locate/acha

Generalized Poincaré–Bertrand formula on a hypersurface

Fengbo Hang a,1 , Shidong Jiang b,∗,2
a
Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA
b
Department of Mathematical Sciences, New Jersey Institute of Technology, Newark, NJ 07102, USA

a r t i c l e i n f o a b s t r a c t

Article history: The Poincaré–Bertrand formula concerning two repeated Cauchy’s principal integrals on
Received 13 October 2008 a smooth curve in the plane is generalized to identities of singular integrals on smooth
Revised 1 December 2008 hypersurfaces in higher dimensions.
Accepted 7 December 2008
Published by Elsevier Inc.
Available online 24 December 2008
Communicated by Vladimir Rokhlin

Keywords:
Singular integrals
Riesz transforms
Poincaré–Bertrand formula

1. Introduction

The classical Poincaré–Bertrand formula states that (see [9])

       
1 φ(ξ, w ) φ(ξ, w )
p.v. p.v. dw dξ = −π 2 φ( z, z) + p.v. dξ dw , (1.1)
ξ −z w −ξ (ξ − z)( w − ξ )
γ γ γ γ

where γ is a smooth arc or closed curve in the plane, z, ξ, w are all viewed as complex numbers and φ is a Hölder
continuous function. The formula has been generalized to accommodate fairly general conditions on γ and φ and to domains
in Cn by various researchers (see [6,9] and more recent references [2,3,5,7,10]).
In this paper we extend the formula to hypersurfaces in higher dimensional Euclidean spaces. We first note that if γ is
a line segment, Cauchy’s principal integral is just the finite Hilbert transform [9] of the given function, and the Poincaré–
Bertrand formula simply computes the iterated finite Hilbert transform of the function φ . The Riesz transforms [8] are a
natural analogy of the Hilbert transform in higher dimensions. For a bounded smooth domain Ω ⊂ Rn , we may consider
the iterated principal integral


n    
zj − xj zj − yj
p.v. p.v. f ( y ) dy dz.
| z − x|n+1 | z − y |n+1
j =1 Ω Ω

A special case of our main result, Theorem 1.1 below, says for a sufficiently smooth function f ,

* Corresponding author.
E-mail addresses: [email protected] (F. Hang), [email protected] (S. Jiang).
1
Supported in part by NSF under Grant DMS-0647010 and by Sloan Research Fellowship.
2
Supported in part by NSF under Grant DMS-0715121.

1063-5203/$ – see front matter Published by Elsevier Inc.

doi:10.1016/j.acha.2008.12.001
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 101


n    
zj − xj zj − yj
p.v. p.v. f ( y ) dy dz
| z − x|n+1 | z − y |n+1
j =1 Ω Ω
   
z−x z− y π n+1
= p.v. · dz f ( y ) dy + f (x).
|z − x|n+1 | z − y |n+1 Γ ( n+2 1 )2
Ω Ω

Here Γ represents the Gamma function given by

∞
Γ (λ) = t λ−1 e −t dt for λ > 0.
0

We next note that for n  2, up to a constant the Green’s function of the Laplace operator on Rn+1 is given by
1
G (x, y ) = , (1.2)
|x − y |n−1
and if we identify Rn as the hyperplane Rn × {0} in Rn+1 , then for x, z ∈ Rn ,
z−x 1
∇ R G (x, z).
n
=
| z − x|n+1 1−n z

Here ∇zR denotes the surface gradient. Hence if M ⊂ Rn+1 is a smooth (curved) n-dimensional hypersurface, we may
n

consider the iterated principal integral

   
p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) f ( y ) dμ( y ) dμ(z), (1.3)
M M

where ∇ M denotes the surface gradient. Our main result is stated in the following theorem.

Theorem 1.1. Assume n  2, M n ⊂ Rn+1 is a smooth compact hypersurface with or without boundary. For x, y ∈ Rn+1 , let G (x, y ) =
1
|x− y |n−1
, then for 1 < p < ∞, f ∈ L p ( M ),
   
p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) f ( y ) dμ( y ) dμ(z)
M M
   
4π n+1
= p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z) f ( y ) dμ( y ) + f (x) (1.4)
Γ ( n−2 1 )2
M M

for a.e. x ∈ M. Here μ is the surface measure on M.

Remark 1.1. Our motivation to consider the extension of the formula in the above form (1.3) comes from the study of scat-
tering problems by open surfaces arising from various areas such as computational electromagnetics, fluid mechanics and
elasticity. The original Poincaré–Bertrand formula plays an important role in constructing a second kind integral equation
formulation for the Dirichlet problem of the Laplace equation when the boundary consists of a set of open curves in two
dimensions (see [1] for details). Here we are looking for stable, efficient, and accurate numerical algorithms for solving the
Dirichlet problem on open surfaces in higher dimensions. That is, if M has nonempty boundary, g is a smooth function on
M, we are looking for a continuous function u on Rn+1 such that u | M = g, u is harmonic in Rn+1 \ M and u (x) tends to 0 as
x tends to infinity. We may assume for x ∈ Rn+1 , u (x) is in the form of (1.3) for some unknown f , then Theorem 1.1 together
with the kernel estimate in Propositions 3.1 and 3.2 implies the Dirichlet problem is reduced to an integral equation of the
second kind.

A formulation which is more similar to the original Poincaré–Bertrand formula (1.1) is as follows.

Theorem 1.2. Assume n  2, M n ⊂ Rn+1 is a smooth compact hypersurface with or without boundary. For x, y ∈ Rn+1 , let G (x, y ) =
1
|x− y |n−1
. Assume 0 < α < 1, H ∈ C α ( M × M ), then for every x ∈ M \∂ M,
   
p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) H (z, y ) dμ( y ) dμ(z)
M M
   
4π n+1
= p.v. ∇zM G (x, z) · ∇zM G (z, y ) H (z, y ) dμ(z) dμ( y ) + H (x, x). (1.5)
Γ ( n−2 1 )2
M M
102 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

The paper is organized as follows. In Section 2, we first describe some basic notations and formulas to be used subse-
quently. We then derive some integral identities related to Riesz transforms and some basic facts about singular integrals on
hypersurfaces. In Section 3, we show that the kernel on the right-hand side of (1.2) is only weakly singular. This estimate
is not only needed later in proving (1.2), it also justifies the applicability of the formula in numerical computation. In Sec-
tion 4, we prove the validity of (1.2) for the case f = 1. This is an important step toward the derivation of the general case.
This special case is proved by an approximation procedure together with a somewhat lengthy calculation. After all these
preparations, in Section 5, we prove the above two main theorems.

2. Some preparations

As usual, if A and B are two quantities, we write A = O ( B ) to mean there exists a constant c > 0 such that | A |  c | B |,
where c depends only on some unimportant ingredients (e.g. the hypersurface M). We also write A = o ε (1) to mean
limε→0+ A = 0. Note here A can depend on some other variables.
For x0 ∈ Rm and r > 0, we denote the open ball B r (x0 ) = {x ∈ Rm : |x − x0 | < r }, sometime to emphasize the dimension
r (x0 ) instead. We also write B r = B r (0).
we use B m
Let M n ⊂ Rn+1 be a smooth surface, ν be the unit normal direction, the shape operator A is given by
A( X ) = D X ν for X ∈ T M ,
where D denotes the usual directional derivative. The eigenvalues of A are called principle curvatures of M and the trace
of A is the mean curvature, i.e. H = tr A. If f is a smooth function defined on an open neighborhood of M in Rn+1 , then
on M,
∂f ∂2 f
M f = f − H − 2.
∂ν ∂ν
Here M is the Laplace operator on M.
Let x0 ∈ M be an interior point, then near x0 , M is the graph of a smooth function defined on the tangent space T x0 M.
More precisely by translation and rotation, we can assume x0 = 0 and the tangent plane is given by Rn × {0}, there exists a
r0 > 0 and a smooth function ϕ : B n2r → R such that near 0, M is just the graph of ϕ , moreover
0

1
n

ϕ (u ) = κi u 2i + O |u |3 as u → 0.
2
i =1

Here κ1 , . . . , κn are the principle curvatures of M at 0.

Let μ be the surface measure on M. If f is a function on M, for every small ε > 0, f is integrable on M \ B nε+1 (x0 ), then
we denote
 
p.v. f dμ = lim f d μ.
ε→0+
M n +1
M\Bε (x0 )

2.1. Some identities related to Riesz transformation

The Riesz transforms R j ( j = 1, 2, . . . , n) are defined by the formula [8, Chapter III]

Γ ( n+2 1 ) uj − vj
( R j f )(u ) = p.v. f ( v ) dv ,
π
n +1
2 |u − v |n+1
Rn

for a function f defined on Rn . In terms of Fourier transforms, the Riesz transforms are given by
iξ j
R

j f (ξ ) = −

f (ξ ).
|ξ |
If h is a harmonic function on upper half space Rn++1 with suitable decay condition near infinity, then for 1  j  n, we
have

R j ∂n+1 h(u , 0) = ∂ j h(u , 0).
For ε > 0, if we choose the harmonic function
1
h(x) = ,
|x + (0, . . . , 0, ε )|n−1
then we get
 
ε uj
Rj n +1
= n +1
.
(|u |2 +ε 2) 2 (|u |2 + ε2 ) 2
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 103

n
Because j =1 RjRj f = −f ,


n  n +1
uj − vj vj π 2 ε
p.v. dv = − .
|u − v |n+1 (| v |2 + ε 2 ) n +1
2 Γ ( n+2 1 ) (|u |2 + ε2 )
n+1
2
j =1 Rn

Let ε → 0+ , then

n 
uj − vj vj
p.v. dv = 0 for u = 0. (2.1)
|u − v |n+1 | v |n+1
j =1 Rn
|u |
Indeed, fix a positive number δ < 2
, we have

uj − vj vj
p.v. dv
|u − v |n+1 (| v |2 + ε 2 ) n+2 1
Rn
     
uj − vj vj uj uj − vj uj vj
= + n +1
− n +1
dv + +
− + n+1
dv
|u − v | n 1
(| v |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2 |u − v |n 1 |u |n 1
(| v |2 + ε 2 ) 2
B δ (u ) Bδ

uj − vj vj
+ + n+1
dv
|u − v | n 1
(| v |2 + ε 2 ) 2
Rn \( B δ (u )∪ B δ )

uj − vj vj
→ p.v. dv
|u − v |n+1 | v |n+1
Rn

as ε → 0+ , by dominated convergence theorem.

As a consequence we get the following fact which will be useful in Section 3.

Lemma 2.1. Let e 1 = (1, 0, . . . , 0) ∈ Rn , then

 
1 1
dt = du .
(1 + |t |2 )n |u |n−1 |u − 2e 1 |n+1
Rn−1 u ∈R ,n
u 1 <1

Proof. It follows from (2.1) that


u u − 2e 1
p.v. · du = 0.
|u |n+1 |u − 2e 1 |n+1
Rn

By symmetry we see that

 
u u − 2e 1 u u − 2e 1
p.v. · du = p.v. · du = 0.
|u |n+1 |u − 2e 1 |n+1 |u |n+1 |u − 2e 1 |n+1
u ∈Rn u ∈Rn
u 1 <1 u 1 >1

Hence

u u − 2e 1
0 = lim · du
ε→0+ |u |n+1 |u − 2e 1 |n+1
u ∈Rn \ B ε
u 1 <1

1 1 u − 2e 1
= lim ∇ · du
ε→0+ 1 − n |u |n−1 |u − 2e 1 |n+1
u ∈Rn \ B ε
u 1 <1
   
1 1 1
= dt − du .
n−1 (1 + |t |2 )n |u |n−1 |u − 2e 1 |n+1
Rn−1 u ∈Rn
u 1 <1

Note in the last equality we have applied the integration by parts formulas. And the lemma follows. 2

Another fact which will be useful in Section 4 is the following identity.

104 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

Lemma 2.2. Assume h ∈ C ∞ (Rn ) has compact support, h(0) = 1. Let


h( v )
g (u ) = dv ,
|u − v |n−1
Rn

then
  
(n + 1)ε 2 1 4π n+1
lim − g (u ) du = .
(n − 1)Γ ( n−2 1 )2
n+3 n+1
ε→0+ (|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
Rn

Proof. Since

n
( R j R j h)(0) = −h(0) = −1,
j =1

we have
   
Γ ( n+2 1 )2 
n
uj uj − vj
p.v. p.v. h( v ) dv du = 1.
π n+1 j =1
|u |n+1 |u − v | n+1
Rn Rn

Hence
   
u u−v π n+1
· p.v. h( v ) dv du = + oε (1).
(|u |2 + ε 2 )
n +1
2 |u − v |n+1 Γ ( n+2 1 )2
Rn Rn

Using
  
u−v
∇ g (u ) = −(n − 1) p.v. n+1
h( v ) dv
|u − v |
Rn

and integration by parts we get

  
1 (n + 1)ε 2 1 π n+1
− g (u ) du = + oε (1).
Γ ( n+2 1 )2
n +3 n+1
n−1 (|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
Rn

And the lemma follows. 2

2.2. Singular integral on a hypersurface

Assume M n ⊂ Rn+1 is a smooth compact hypersurface without boundary. Denote


( T f )(x) = p.v. ∇ yM G (x, y ) f ( y ) dμ( y ). (2.2)
M

In order to find the most singular part of the kernel, we may assume x = 0 and near 0, M is the graph of a smooth
function ϕ (as described at the beginning of Section 2), then

M  
∇ G 0, u , ϕ (u ) − (1 − n)(u , 0)  c ( M ) .
y |u |n+1 |u |n−1

It follows from [4, Chapters IX, X] that for f ∈ C α ( M ), 0 < α < 1, T f ∈ C α ( M ) and

| T f |C α (M )  c ( M , α )| f |C α (M ) .
On the other hand, for f ∈ L p ( M ), 1 < p < ∞, T f ∈ L p ( M ) and

| T f | L p (M )  c ( M , p )| f | L p (M ) .
If M n ⊂ Rn+1 is a smooth compact hypersurface with nonempty boundary, we can fix a smooth compact hypersurface
without boundary, namely M n0 ⊂ Rn+1 , such that M is a domain in M 0 . Again we have the operator T defined by (2.2). For
a function f on M, we may set

f (x), x ∈ M,
f 0 (x) =
0, x ∈ M0 \M .
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 105

Then for 1 < p < ∞,



| T f | L p (M ) = p.v. ∇ y 0 G (x, y ) f 0 ( y ) dμ( y )
M
 c ( M , p )| f 0 | L p (M 0 ) = c ( M , p )| f | L p (M ) .
p
L x (M )
M0

The next two lemmas are about the validity of the Fubini theorem when singular integrals are involved.

Lemma 2.3. Assume M n ⊂ Rn+1 is a smooth compact hypersurface with or without boundary, 1 < p < ∞, Φ ∈ L p ( M × M ), then for
a.e . x ∈ M,
      
p.v. ∇ yM G (x, y ) Φ( y , z) dμ(z) dμ( y ) = p.v. ∇ yM G (x, y )Φ( y , z) dμ( y ) dμ( z).
M M M M

Proof. At first we verify the identity for Φ ∈ C ∞ ( M × M ). Indeed under this assumption, for x ∈ M \∂ M,
  
p.v. ∇ yM G (x, y ) Φ( y , z) dμ(z) dμ( y )
M M
     
= ∇ yM G (x, y) Φ( y , z) dμ(z) − Φ(x, z) dμ(z) dμ( y ) + Φ(x, z) dμ(z) · p.v. ∇ yM G (x, y ) dμ( y )
M M M M M
     

= d μ( z ) ∇ yM G (x, y ) Φ( y , z) − Φ(x, z) dμ( y ) + Φ(x, z) · p.v. ∇ yM G (x, y ) d μ( y ) d μ( z )
M M M M
   
= p.v. ∇ yM G (x, y )Φ( y , z) dμ( y ) dμ( z).
M M

To continue we observe that

   

p.v. ∇ M G (x, y ) Φ( y , z) dμ(z) dμ( y )  c ( M , p ) Φ( y , z) dμ(z)
 c ( M , p )|Φ| L p (M × M )
y p p
L x (M ) L y (M )
M M M

and
     

p .v. ∇ M
G ( x, y )Φ( y , z ) d μ ( y ) d μ ( z )  p.v. ∇ M G (x, y )Φ( y , z) dμ( y ) d μ( z )
y p y p
L x (M ) L x (M )
M M M M


 c(M , p) Φ( y , z) p dμ( z)  c ( M , p )|Φ| L p ( M × M ) .
L y (M )
M

Because C ∞ ( M × M ) is dense in L p ( M × M ), the lemma follows. 2

Lemma 2.4. Assume M n ⊂ Rn+1 is a smooth compact hypersurface with or without boundary, 1 < p < ∞, p
= p −1 , f ∈ L p ( M ),
p

g ∈ L p ( M ), then
       
g (x) p.v. ∇xM G (x, y ) f ( y ) dμ( y ) dμ(x) = f ( y ) p.v. ∇xM G (x, y ) g (x) dμ(x) dμ( y ).
M M M M

Proof. Denote

(
T g )( y ) = p.v. ∇xM G (x, y ) g (x) dμ(x),
M

then | T g | L p
( M )  c ( M , p )| g | L p
( M ) . Hence by approximation to prove the lemma we only need to consider the case when
f , g ∈ C ∞ ( M ). Under this assumption we have for ε > 0,
       
M M
g (x) ∇x G (x, y ) f ( y ) dμ( y ) dμ(x) = f ( y) ∇x G (x, y ) g (x) dμ(x) dμ( y ).
M y∈M M x∈ M
|x− y |>ε |x− y |>ε

Let ε → 0+ , we get the needed identity. 2

106 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

Finally we discuss an approximation of the singular integral by integral operators with smooth kernels. If M n ⊂ Rn+1 is
a smooth compact hypersurface without boundary, for x ∈ M, let ν (x) be the unit normal direction, then for ε > 0 small,
we can define


( T ε f )(x) = ∇ yM G x + εν (x), y f ( y ) dμ( y ).
M

Then for 1 < p < ∞,

| T ε f | L p (M )  c ( M , p )| f | L p (M )
and

Tε f → T f in L p ( M )

for every f ∈ L p ( M ).
y G (x + εν (x), y ) for y close to x. To this aim we can assume
Later on we will need to know the asymptotic formula of M
x = 0 and near 0, M is the graph of a smooth function ϕ (as described at the beginning of Section 2). Let y = (u , ϕ (u )),
then
  ∂G  ∂2G 
M
y G x + εν (x), y = y G x + εν (x), y − H ( y ) x + εν (x), y − 2
x + εν (x), y
∂ν ∂ν
n−1 (n − 1) H (u · ∇ ϕ − ϕ − ε )
= n+1
+ n+1
[|u |2 + (ϕ + ε )2 ] 2 1 + |∇ ϕ |2 [|u |2 + (ϕ + ε )2 ] 2
(n2 − 1)(u · ∇ ϕ − ϕ − ε )2
− n+3
,
(1 + |∇ ϕ |2 )[|u |2 + (ϕ + ε )2 ] 2

here ϕ = ϕ (u ), H = H (u , ϕ (u )). Denote r = |u |2 + ε 2 , then
  
2 2 2
ε ni=1 κi u 2i  2
|u | + (ϕ + ε ) = r 1 + 2
+ O r ,
r
 n 
1 1 n+1 ε i =1 κi u 2i 
= 1− + O r2 ,
[|u |2 + (ϕ + ε )2 ]
n+1
2 r n+1 2 r2

1 
n

u · ∇ ϕ − ϕ − ε = −ε + κi u 2i + O r 3 .
2
i =1

Based on these formulas, we get

 n
 n − 1 (n2 − 1)ε 2 (n − 1)ε ni=1 κi n2 − 1 ε i =1 κi u 2i
M G x + εν ( x), y = − − +
y
r n+1 r n+3 r n+1 2 r n+3
3
n 2  
(n − 1)(n + 3) ε
2
i =1 κi u i 1
+ n+5
+ O n−1 , (2.3)
2 r r

where r = | u |2 + ε 2 .

3. Estimate of the kernel

In this section, we consider the kernel K on the right-hand side of (1.4) defined by the formula

K (x, y ) = p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z).
M

We show that the kernel is only weakly singular along the diagonal x = y. The main results are summarized in Proposi-
tion 3.1 for the case of a closed surface and Proposition 3.2 for the case of a surface with nonempty boundary.

3.1. Compact hypersurface without boundary

Proposition 3.1. Assume that n  2, M n ⊂ Rn+1 is a smooth compact hypersurface without boundary. For x, y ∈ M, let

K (x, y ) = p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z).
M

Then
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 107


c ( M )(| log |x − y || + 1), if n = 2;
K (x, y ) 
c ( M )( |x− y1|n−2 + 1), if n  3.

Here c ( M ) is a positive constant depending only on M.

It is clear that we only need to verify the estimate when |x − y | is small. Using notations in Section 2, we may assume
x = 0, T 0 M = Rn × {0} and near 0, M is the graph of a smooth function ϕ defined on B n2r . Let e 1 = (1, 0, . . . , 0) ∈ Rn , we
0
may assume y = (2δ e 1 , ϕ (2δ e 1 )). Using the expansion formula of ϕ , we see when | y | is small enough,

2δ  | y |  2.1δ.

Note that in the integrand of the expression of K (x, y ), we have two singular points as z = x and z = y. The main idea
of proving the estimate of K (x, y ) comes from the proof of Lemma 2.1: we divide the integral into two pieces from the
“middle” of x and y, each of which contains a singular point and satisfies the needed estimate. More precisely, let
 
D= u , ϕ (u ) : u ∈ B nr0 (δ e 1 ) ,
 
Dl = u , ϕ (u ) : u ∈ B nr0 (δ e 1 ), u 1 < δ ,
 
Dr = u , ϕ (u ) : u ∈ B nr0 (δ e 1 ), u 1 > δ .

Since on M \ D, both ∇zM G (0, z) and ∇zM G ( z, y ) are uniformly bounded,

 
K (0, y ) = O (1) + p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ(z) + p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ(z).
Dl Dr

We claim
 
O (log 1δ ), if n = 2,
p.v. ∇ M G (0, z) · ∇ M G (z, y ) dμ(z) = (3.1)
z z O ( δn1−2 ), if n  3,
Dl
 
O (log 1δ ), if n = 2,
p.v. ∇ M G (0, z) · ∇ M G (z, y ) dμ(z) = (3.2)
z z O ( δn1−2 ), if n  3.
Dr

This clearly implies Proposition 3.1.

Below we only prove (3.1). The verification of (3.2) is almost identical. To derive (3.1), we scale the graph by 1δ . More
precisely, let
ϕ (δ u )
ψ(u ) = ,
δ
then
2
ψ(u )  c ( M )δ|u |2 , ∇ψ(u )  c ( M )δ|u |, ∇ ψ(u )  c ( M )δ.

Let
 
N= u , ψ(u ) : u ∈ B nr0 /δ (e 1 ), u 1 < 1 ,

y = 2e 1 , ψ(2e 1 ) .

N is simply the 1δ dilation of D l . Using integration by parts, we get


p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ(z)
Dl

1
= n
p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ N (z)
δ
N
 
1 ∂G 1
= G (0, z) (z, y ) dS (z) − n G (0, z) zN G ( z, y ) dμ N ( z).
δn ∂ ν∂ N δ
∂N N

Here μN is the surface measure on N and ν∂ N is the outer normal direction of N at ∂ N.

108 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

We will show
  
1 ∂G n−1 1 O (log 1δ ), if n = 2,
G (0, z) (z, y ) dS (z) = n dt + (3.3)
δn ∂ ν∂ N δ (1 + |t |2 )n O ( δn1−2 ), if n  3
∂N Rn−1

and
  
1 N n−1 1 O (log 1δ ), if n = 2,
G (0, z) G ( z, y ) dμ N ( z) = du + (3.4)
δn δn |u |n−1 |u − 2e 1 |n+1 1
O ( δn−2 ), if n  3.
N u ∈Rn ,
u 1 <1

Once we have (3.3) and (3.4), the inequality (3.1) follows from Lemma 2.1 (note that the terms of order δ1n cancels!).
To prove (3.3), we decompose ∂ N into two pieces, namely
   
Σ1 = u , ψ(u ) : u ∈ B nr0 /δ (e 1 ), u 1 = 1 = 1, t , ψ(1, t ) : t ∈ B nr0−/δ1 ,
 
Σ2 = u , ψ(u ) : u ∈ ∂ B nr0 /δ (e 1 ), u 1  1 .

Using estimates of |∇ψ| we see that


1 ∂G
G (0, z ) ( z , y ) dS ( z )  c ( M ).
δn ∂ν
∂N
Σ2

On the other hand, on Σ1 ,


z = 1, t , ψ(1, t ) , t ∈ B nr0−/δ1 ,

dS ( z) = 1 + |∇ψ|2 − (∂1 ψ)2 dt ,
 − n−2 1
G (0, z) = 1 + |t |2 + ψ 2 ,
(1 + |∇ψ| − (∂1 ψ) , −∂1 ψ∂2 ψ, . . . , −∂1 ψ∂n ψ, ∂1 ψ)
2 2
ν∂ N (z) =   .
1 + |∇ψ|2 1 + |∇ψ|2 − (∂1 ψ)2
Hence
∂G (n − 1)[1 + |∇ψ|2 − (∂1 ψ)2 + t 1 ∂1 ψ∂2 ψ + · · · + tn−1 ∂1 ψ∂n ψ − ∂1 ψ(ψ(1, t ) − ψ(2e 1 ))]
( z, y ) =   n+1
.
∂ ν∂ N 1 + |∇ψ|2 1 + |∇ψ|2 − (∂1 ψ)2 [1 + |t |2 + (ψ(1, t ) − ψ(2e 1 ))2 ] 2

Based on these formulas and estimates of ψ we see that when n = 2,


1 ∂G
G (0, z) (z, y ) dS (z)
δ2 ∂ ν∂ N
Σ1

r0 /δ  
1 1 1
=   dt + O log
δ2 1 + t 2 + ψ 2 1 + |∇ψ|2 [1 + t 2 + (ψ(1, t ) − ψ(2, 0))2 ] 2
3
δ
−r0 /δ

r0 /δ  
1 1 1
= dt + O log
δ2 (1 + t 2 )2 δ
−r0 /δ
∞  
1 1 1
= dt + O log ;
δ2 (1 + t 2 )2 δ
−∞

and when n  3,

1 ∂G
G (0, z) (z, y ) dS (z)
δ2 ∂ ν∂ N
Σ1
  
n−1 1 1
= n+1 
dt + O
δn (1 + |t |2 + ψ 2 )
n −1
2 [1 + |t |2 + (ψ(1, t ) − ψ(2e 1 ))2 ] 2 1 + |∇ψ|2 δn−2
B nr −/δ
1
0
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 109

  
n−1 1 1
= dt + O
δn (1 + |t |2 )n δn−2
B nr −/δ
1
0
  
n−1 1 1
= dt + O .
δn (1 + |t |2 )n δn−2
Rn−1

Hence we get (3.3).

To prove (3.4), we note that

z = u , ψ(u ) , u ∈ B nr0 /δ (e 1 ), u 1 < 1,

dμ N ( z) = 1 + |∇ψ|2 du ,
(∇ψ, −1)
ν N ( z) =  ,
1 + |∇ψ|2
 − n−2 1
G (0, z) = |u |2 + ψ 2 ,

here νN is the normal direction of N. To find the formula of N G (z, y ), we note that
∂G ∂2G
N G ( z, y ) = G ( z, y ) − H N ( z, y ) − 2 ( z, y )
∂ νN ∂ νN
n−1
= n+1
[|u − 2e 1 + (ψ(u ) − ψ(2e 1 ))2 ]
|2 2

(n − 1) H N [ψ(u ) − ψ(2e 1 ) − (u − 2e 1 ) · ∇ψ] (n2 − 1)[ψ(u ) − ψ(2e 1 ) − (u − 2e 1 ) · ∇ψ]2

− n+1
− n +3
.
1 + |∇ψ|2 [|u − 2e 1 |2 + (ψ(u ) − ψ(2e 1 ))2 ] 2 (1 + |∇ψ|2 )[|u − 2e 1 |2 + (ψ(u ) − ψ(2e 1 ))2 ] 2
Based on these formulas, the estimates for ψ and the basic fact | H N |  c ( M )δ we can derive (3.4) by calculations similar as
before.

3.2. Compact hypersurface with nonempty boundary

Proposition 3.2. Assume n  2, M n0 ⊂ Rn+1 is a smooth compact hypersurface without boundary, M ⊂ M n0 is a smooth domain in
M 0 . For x, y ∈ M, let

K (x, y ) = p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z).
M

Then
⎧ 
⎨ c ( M )(| log |x − y || + M 0 \ M |x−z|2 | y−z|2 dμ(z) + 1), if n = 2;
1 1

K (x, y )  
⎩ c ( M )( 1 n−2 + 1 1
dμ( z) + 1), if n  3.
|x− y | M 0 \ M |x− z|n | y − z|n

In particular,


K (x, y ) dμ( y )  c ( M ) 1 + log d(x, ∂ M ) 2
M

and
 

dμ(x) K (x, y ) dμ( y )  c ( M ) < ∞.
M M

The proposition easily follows from the estimate in the closed surface case. Indeed let

M M
K 0 (x, y ) = p.v. ∇z 0 G (x, z) · ∇z 0 G (z, y ) dμ(z),
M0

then
110 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116



K (x, y )  K 0 (x, y ) + M M
∇z 0 G (x, z) · ∇z 0 G (z, y ) dμ( z)

M0 \M

1
 K 0 (x, y ) + c ( M ) dμ( z).
|x − z|n | y − z|n
M0 \M

The first inequality in the proposition follows from this and Proposition 3.1. Next we have
  
1
K (x, y ) dμ( y )  c ( M ) + c ( M ) d μ( z ) d μ( y )
|x − z|n | y − z|n
M M0 \M M

1 + | log d( z, M )|
 c(M ) + c(M ) d μ( z )
|x − z|n
M0 \M
 2
 c ( M ) 1 + log d(x, ∂ M ) .
The last two inequalities in the Proposition 3.2 follows.

4. Poincaré–Bertrand formula with a constant density

4.1. Compact hypersurface without boundary

The main aim of this subsection is to prove the special case of (1.4) when the density function f = 1. It is an important
step in deriving the general Poincaré–Bertrand formula.

Proposition 4.1. Assume n  2, M n ⊂ Rn+1 is a smooth compact hypersurface without boundary. Then for x ∈ M,
   
p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) d μ( y ) d μ( z )
M M
   
4π n+1
= p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z) dμ( y ) + .
Γ ( n−2 1 )2
M M

Our main strategy to prove the identity is by an approximation procedure, more precisely we will replace ∇zM G (x, z)
by ∇zM G (x + εν (x), z) and check what happens when ε → 0+ . The advantage of this procedure is that we can reduce the
number of principal integrals in the expressions and apply the Fubini type results in Lemmas 2.3 and 2.4. After that by a
lengthy and careful local analysis we derive the needed identity through a limiting process.
Without losing of generality we may assume x = 0 and near 0, M is the graph of a smooth function ϕ : B n2r → R (as
0
described at the beginning of Section 2). Let en+1 = (0, . . . , 0, 1) ∈ Rn+1 and

K ε ( y ) = p.v. ∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z),
M

K ( y ) = p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ(z).
M

It follows from Proposition 3.1 that K ∈ L 1 ( M ). By Lemma 2.4,

    
K ε ( y ) d μ( y ) = ∇zM G (−εen+1 , z) · p.v. ∇zM G (z, y ) dμ( y ) dμ(z)
M M M
   
= p.v. ∇zM G (0, z) · p.v. ∇zM G (z, y ) dμ( y ) dμ(z) + oε (1).
M M

On the other hand, we will show

 
4π n+1
K ε ( y ) d μ( y ) = K ( y ) d μ( y ) + + oε (1), (4.1)
Γ ( n−2 1 )2
M M

and the proposition clearly follows from the above two equalities by letting ε → 0+ .
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 111

To prove (4.1), first we claim for every δ > 0,



K ε ( y ) − K ( y ) dμ( y ) → 0 as ε → 0+ . (4.2)
M \ B nδ +1

This claim means interesting things happen only near 0. To prove the claim we note that
 

∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z) → p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ(z) in L ∞ n
y M\Bδ ,

M ∩ B nδ/+21 M ∩ B nδ/+21

and for any 1 < p < ∞,

 
p
p.v. ∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z) → p.v. ∇zM G (0, z) · ∇zM G (z, y ) dμ(z) in L y M \ B nδ
M \ B nδ/+21 M \ B nδ/+21

as ε → 0+ . (4.2) follows.
Next we study K ε ( y ) for y near 0. For 0 < δ < 2r0 , we write
 
Dδ = u , ϕ (u ) : u ∈ B nδ .

Fix a cutoff function η ∈ C ∞ (Rn ) such that η| B 1/2 = 1 and η|Rn \ B 1 = 0. For δ > 0 small, define a function hδ on M by

η( uδ ), if y = (u , ϕ (u )), u ∈ B δ ;
hδ ( y ) =
0, otherwise.

For δ > 0 small and y ∈ D δ , we have

 
K ε ( y ) = p.v. ∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z) + ∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z).
D r0 M \ D r0

Note that
 
∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z) → ∇zM G (0, z) · ∇zM G (z, y ) dμ(z) (4.3)
M \ D r0 M \ D r0

uniformly for y ∈ D δ as ε → 0+ . Hence we only need to know what happens for


p.v. ∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z)
D r0

as ε → 0+ .
To continue we observe that
   
p.v. ∇zM G (−εen+1 , z) · ∇zM G (z, y ) dμ(z) hδ ( y ) dμ( y )
M D r0
   
= ∇zM G (−εen+1 , z) · p.v. ∇zM G (z, y )hδ ( y ) dμ( y ) dμ(z)
D r0 M

= ∇zM G (−εen+1 , z) · ∇zM f δ (z) dμ(z)
D r0
 
∂G 
=− f δ ( z) zM G (−ε en+1 , z) dμ( z) + f δ ( z) (−εen+1 , z) dS (z), (4.4)
∂ ν∂ D r0
D r0 ∂ D r0

where

f δ ( z) = G ( z, y )hδ ( y ) dμ( y )
M
112 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

is the single layer potential and ν∂ D r0 is the unit outer normal direction on ∂ D r0 . For z ∈ D 3 r , we can write z = (u , ϕ (u ))
2 0
and
  
η( vδ ) 1 + |∇ ϕ ( v )|2
f δ ( z) = G ( z, y )hδ ( y ) dμ( y ) = n −1
dv = F δ (u ). (4.5)
[|u − v |2 + (ϕ (u ) − ϕ ( v ))2 ] 2
Dδ Bδ

The last equality is the definition of F δ . In particular, for z ∈ D r0 , | f δ ( z)|  c ( M )δ , hence


∂G 
f δ ( z) (−εen+1 , z) dS (z) = O (δ). (4.6)
∂ ν∂ D r0
∂ D r0

On the other hand, by (2.3),


− f δ ( z) zM G (−ε en+1 , z) dμ( z)
D r0
   
(n + 1)ε 2 1
= (n − 1) n+3
− n +1
F δ (u ) 1 + |∇ ϕ |2 du
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
B r0
 n 
ε i =1 κi
+ (n − 1) n+1
F δ (u ) 1 + |∇ ϕ |2 du
(|u |2 + ε 2 ) 2
B r0
 n 
n2 − 1 ε i =1 κi u 2i
− n +3
F δ (u ) 1 + |∇ ϕ |2 du
2 (|u |2 + ε 2 ) 2
B r0
 n 
(n2 − 1)(n + 3) ε3 i =1 κi u 2i
− n +5
F δ (u ) 1 + |∇ ϕ |2 du + O (δ). (4.7)
2 (|u |2 + ε 2 ) 2
B r0

Calculation shows that


1 n Γ (λ − n2 ) n
du = π 2 for λ > .
(|u |2 + 1) λ Γ (λ) 2
Rn

Hence
   n+1
ε 1 π 2
F δ (u ) 1 + |∇ ϕ |2 du = F δ (0) du + oε (1) = F δ (0) + oε (1). (4.8)
Γ ( n+2 1 )
n +1 n+1
(|u |2 + ε 2 ) 2 (|u |2 + 1) 2
B r0 Rn

Similarly, for 1  i  n,
  n+1
ε u 2i 1 π 2
F δ (u ) 1 + |∇ ϕ |2 du = F δ (0) + oε (1), (4.9)
(|u |2 + ε 2 )
n +3
2 n + 1 Γ ( n+1 )
2
B r0

and
  n +1
ε3 u 2i 1 π 2
F δ (u ) 1 + |∇ ϕ |2 du = F δ (0) + oε (1). (4.10)
(|u |2 + ε 2 )
n +5
2 (n + 1)(n + 3) Γ ( n+2 1 )
B r0

Note that (4.8)–(4.10) implies the summation of those terms on the right-hand side of (4.7) containing principal curvature
κi ’s is equal to oε (1).
We still need to study the term
   
(n + 1)ε 2 1
n+3
− n +1
F δ (u ) 1 + |∇ ϕ |2 du
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
B r0

which is “hypersingular.” We will show

 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 113

   
(n + 1)ε 2 1
n+3
− n +1
F δ (u ) 1 + |∇ ϕ |2 du
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
B r0
 
4π n+1 δ2 
= + oε (1) + O (δ) + O + O ε1α . (4.11)
(n − 1)Γ ( n−2 1 )2 ε1
For this purpose we will compare F δ (u ) with the single layer potential g δ (u ) of the flat surface case given by

η( vδ )
g δ (u ) = du . (4.12)
|u − v |n−1
Bδ

First note that

   
(n + 1)ε 2 1
n+3
− n +1
F δ (u ) 1 + |∇ ϕ |2 du
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
B r0
  
(n + 1)ε 2 1
= n +3
− n +1
F δ (u ) du + O (δ). (4.13)
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
B r0

Next we observe that for a fixed 0 < α < 1,

c(M , α)
| f δ |C 1,α (M )  c ( M , α )|hδ |C α (M )  .
δα
Here C 1,α denotes the space of functions whose first order derivatives are Hölder continuous with order α . Hence
c(M , α)
| F δ | C 1, α ( B r )  .
0 δα
On the other hand, let g δ be the single layer potential in the flat case defined in (4.12), then
c (n)
| g δ | C 1, α ( B r )  .
0 δα
It follows that

n  
|u |1+α
( F δ − g δ )(u ) = ( F δ − g δ )(0) + ∂i ( F δ − g δ )(0)u i + O .
δα
i =1

On the other hand, using the Taylor expansion of ϕ , we have

  
v
η( δ ) 1 + |∇ ϕ ( v )|2 η( vδ ) 
( F δ − g δ )(u ) = dv − dv = O δ|u |2 + δ 3 .
[|u − v |2 + (ϕ (u ) − ϕ ( v ))2 ]
n−1
2 |u − v |n−1
Bδ Bδ

In particular,

( F δ − g δ )(0) = O δ 3 .
Hence for ε1 > 0 small,
    
(n + 1)ε 2 1  δ2
n+3
− n +1
F δ (u ) − g δ (u ) du = O (δ) + O , (4.14)
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2 ε1
B r 0 \ B ε1 δ

and
  
(n + 1)ε 2 1 
n +3
− n +1
F δ (u ) − g δ (u ) du
(|u |2 +ε 2) 2 (|u |2 +ε 2) 2
B ε1 δ
  
(n + 1)ε 2 1  α
= ( F δ − g δ )(0) n +3
− n +1
du + O ε 1
(|u |2 +ε 2) 2 (|u |2 +ε 2) 2
B ε1 δ
  
u  α
= ( F δ − g δ )(0) div n+1
du + O ε 1
(|u |2 + ε 2 ) 2
B ε1 δ
 
δ2 
=O + O ε1α . (4.15)
ε1
114 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

Finally by Lemma 2.2,

  
(n + 1)ε 2 1
n+3
− n +1
g δ (u ) du
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
B r0
  
(n + 1)ε 2 1
= n +3
− n +1
g δ (u ) du + O (δ)
(|u |2 + ε 2 ) 2 (|u |2 + ε 2 ) 2
Rn
4π n+1
= + oε (1) + O (δ). (4.16)
(n − 1)Γ ( n−2 1 )2
Combining (4.13)–(4.16), (4.11) follows.
In view of (4.2)–(4.11) and Proposition 3.1 we have
  

K ε ( y ) d μ( y ) = K ε ( y ) 1 − h δ ( y ) d μ( y ) + K ε ( y )hδ ( y ) dμ( y )
M M M
  
4π n+1 δ2 
= K ( y ) d μ( y ) + + oε (1) + O (δ) + O + O ε1α .
Γ ( n−2 1 )2 ε1
M

Hence
   
4π n+1 2
lim sup K ε ( y ) d μ( y ) − K ( y ) d μ( y ) −  c(M ) δ + δ + εα .
Γ ( n−1 )2
1
+ ε→0 2
ε 1
M M

Let δ → 0+ , then ε1 → 0+ , we get

 
4π n+1
K ε ( y ) d μ( y ) → K ( y ) d μ( y ) + as ε → 0+ ,
Γ ( n−2 1 )2
M M

which is exactly (4.1). And Proposition 4.1 follows.

4.2. Compact hypersurface with nonempty boundary

Here we derive the same identity for the case of a surface with nonempty boundary.

Proposition 4.2. Assume n  2, M n ⊂ Rn+1 is a smooth compact hypersurface with nonempty boundary. Then
   
p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) d μ( y ) d μ( z )
M M
   
4π n+1
= p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z) dμ( y ) +
Γ ( n−2 1 )2
M M

for x ∈ M \∂ M.

Basically Proposition 4.2 follows from Proposition 4.1. Indeed we may find a smooth compact hypersurface without
boundary, namely M n0+1 ⊂ Rn+1 such that M is a domain in M 0 . For x ∈ M \∂ M, the same argument as in the proof of
Lemma 2.3 gives us
   
M M
p.v. ∇z 0 G (x, z) · ∇ z 0 G ( z , y ) d μ( y ) d μ( z )
M M0 \M
   
M M
= p.v. ∇z 0 G (x, z) · ∇z 0 G (z, y ) dμ( z) dμ( y ). (4.17)
M0 \M M

By the Fubini theorem we have

  
M M
∇z 0 G (x, z) · ∇ z 0 G ( z , y ) d μ( y ) d μ( z )
M0 \M M
   
M M
= ∇z 0 G (x, z) · ∇z 0 G (z, y ) dμ(z) dμ( y ). (4.18)
M M0 \M
 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116 115

Using Lemma 2.4 we have

   
M M
∇z 0 G (x, z) · p.v. ∇ z 0 G ( z , y ) d μ( y ) d μ( z )
M0 \M M0 \M
   
M M
= p.v. ∇z 0 G (x, z) · ∇z 0 G (z, y ) dμ(z) dμ( y ). (4.19)
M0 \M M0 \M

Finally, it follows from Proposition 4.1 that

   
M M
p.v. ∇z 0 G (x, z) · p.v. ∇z 0 G (z, y ) dμ( y ) dμ(z)
M0 M0
   
M M 4π n+1
= p.v. ∇z 0 G (x, z) · ∇z 0 G (z, y ) dμ(z) dμ( y ) + . (4.20)
Γ ( n−2 1 )2
M0 M0

Subtracting (4.17)–(4.19) from (4.20), we obtain the identity in Proposition 4.2.

5. Proof of the Poincaré–Bertrand formula

After all the preparations in Sections 2–4, we can now easily verify Theorems 1.1 and 1.2.

Proof of Theorem 1.2. Since H is Hölder continuous, we can show that both sides of (1.5) are continuous functions in x
when x ∈ M \∂ M. Hence we only need to verify (1.5) for a.e. x ∈ M. We have for a.e. x ∈ M,
   
p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) H (z, y ) dμ( y ) dμ(z)
M M
  

= p.v. ∇zM G (x, z) · ∇zM G (z, y ) H (z, y ) − H (z, z) dμ( y ) dμ(z)
M M
   

+ ∇zM G (x, z) H (z, z) − H (x, x) p.v. ∇zM G (z, y ) dμ( y ) dμ(z)
M M
   
+ H (x, x)p.v. ∇zM G (x, z) · p.v. ∇zM G (z, y ) d μ( y ) d μ( z )
M M
   

= p.v. ∇zM G (x, z) · ∇zM G (z, y ) H (z, y ) − H (z, z) dμ(z) dμ( y )
M M
   

+ p.v. ∇zM G (x, z) · ∇zM G (z, y ) H (z, z) − H (x, x) dμ(z) dμ( y )
M M
   
4π n+1
+ H (x, x) p.v. ∇zM G (x, z) · ∇zM G (z, y ) dμ(z) dμ( y ) + H (x, x)
Γ ( n−2 1 )2
M M
   
4π n+1
= p.v. ∇zM G (x, z) · ∇zM G (z, y ) H (z, y ) dμ(z) dμ( y ) + H (x, x).
Γ ( n−2 1 )2
M M

Note that in the second step we have used Lemmas 2.3, 2.4, Propositions 4.1 and 4.2. 2

Theorem 1.1 is basically a corollary of Theorem 1.2.

Proof of Theorem 1.1. Since smooth functions are dense in L p ( M ), we only need to verify (1.4) for f ∈ C ∞ ( M ). Under this
assumption, (1.4) follows from (1.5) by choosing H ( z, y ) = f ( y ). 2
116 F. Hang, S. Jiang / Appl. Comput. Harmon. Anal. 27 (2009) 100–116

6. Conclusions

We have generalized the classical Poincaré–Bertrand formula to the case of singular integrals on smooth hypersurfaces
in higher dimensions. The classical Poincaré–Bertrand formula is fairly easy to prove since one may either apply techniques
from complex analysis or take advantage of the fact that the kernel is a simple rational function. Our proof of the generalized
formula is rather lengthy. The main difficulties lie on the fact that we are dealing with much more complicated kernels on
curved hypersurfaces instead of flat Euclidean spaces. The proof is based entirely on the local analysis and thus can be easily
extended to handle other kernels with similar singularities. Specifically, we may replace the Green’s function of the Laplace
equation by the Green’s function of other elliptic partial differential equations (for example, the Helmholtz equation). Finally,
the formula is expected to be used in the construction of second kind integral equation formulations for the open surface
problems in higher dimensions ( 3). This application is currently under investigation and will be reported in a later date.

References

[1] S. Jiang, V. Rokhlin, Second kind integral equations for the classical potential theory on open surfaces I: Analytical apparatus, J. Comput. Phys. 191
(2003) 40–74.
[2] A.M. Kytmanov, B.B. Prenov, N.N. Tarkhanov, The Poincaré–Bertrand formula for the Martinelli–Bochner integral, Izv. Vyssh. Uchebn. Zaved.
Mat. 1992 (11) (1993) 29–34 (in Russian); translation in: Russian Math. (Izv. VUZ) 36 (11) (1992) 27–32.
[3] L.Y. Lin, C.H. Qiu, The singular integral equation on a closed piecewise smooth manifold in Cn , Integral Equations Operator Theory 44 (3) (2002)
337–358.
[4] S.G. Mikhlin, S. Prossdorf, Singular Integral Operators, Springer, 1986.
[5] I.M. Mitelman, M.V. Shapiro, Formulae of changing of integration order and of inversion for some multidimensional singular integrals and hypercomplex
analysis, J. Nat. Geom. 5 (1) (1994) 11–27.
[6] N.I. Muskhelishvili, Singular Integral Equations, Dover, New York, 1991.
[7] B. Schneider, Singular integrals of the time harmonic Maxwell equations theory on a piecewise Liapunov surface, Appl. Math. E—Notes 7 (2007)
139–146 (electronic).
[8] E.M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton University Press, Princeton, NJ, 1970.
[9] F.G. Tricomi, Integral Equations, Dover, New York, 1985.
[10] T.D. Zhong, L.P. Chen, The Poincaré–Bertrand formula for the Bochner–Martinelli integral, Integral Equations Operator Theory 54 (4) (2006) 585–595.