You are on page 1of 16

4

Greens Functions

In this section, we are interested in solving the following problem. Let be an open, bounded subset of Rn . Consider u = f x Rn (4.1) u=g x .

4.1

Motivation for Greens Functions


y G(x, y ) = x G(x, y ) = 0 y y

Suppose we can solve the problem, (4.2)

for each x . Then, formally, we can say that for u a solution of (4.1), u(x) =

x u(y ) dy y G(x, y )u(y ) dy

= =

G (x, y )u(y ) dS (y ) u = G(x, y )y u(y ) dy + G(x, y ) (y ) dS (y ) G = G(x, y )f (y ) dy (y )g (y ) dS (y ). y G(x, y ) y u(y ) dy

G (x, y )u(y ) dS (y )

Now, we do know that the fundamental solution of Laplaces equation (y ) satises y (y ) = 0 and, moreover, y (x y ) = x . Of course, (x y ) does not satisfy our boundary conditions, but we will use that as a starting ground to try and construct a solution of (4.2), and, ultimately (4.1). Below, we will also make the formal argument given above more precise. Recalling the denition of distributional derivative, we will start by looking at (x y )y u(y ) dy.

We would like to integrate this term by parts. However, we know that (x y ) has a singularity at y = x. Therefore, in order to integrate by parts, we must proceed as follows. Fix x and > 0 such that dist(x, ) < and therefore, B (x, ) . Let V B (x, ).

x V

Let be the fundamental solution of Laplaces equation. That is, (x) = 21 ln |x|
1 1 n(n2)(n) |x|n2

n=2 n 3.

Suppose u C 2 (). By the Divergence Theorem, we have (y x)u(y ) dy =


V V

y (y x) y u(y ) dy +
V

(y x)

u dS (y )

=
V

y (y x)u(y ) dy
V

(y x)u(y ) dS (y )

+
V

u (y x) dS (y ).

where u denotes the derivative of u in the outer normal direction to V . Now on V , y (y x) = 0. Therefore, (y x)u(y ) dy =
V V

(y x)u(y ) dS (y ) +

(y x)
V

u dS (y ).

Now, we note that lim +


0

(y x)u(y ) dy =
V

(y x)u(y ) dy. 0+ .

We make the following claims about the limits of the other two terms as Claim 1. lim +
0

(y x)u(y ) dS (y ) = (y x)
V

(y x)u(y ) dS (y ) u(x). (y x) u dS (y ).

Claim 2. lim +
0

u (y ) dS (y ) =

Assuming these claims for a moment, we conclude that for any u C 2 (), u(x) =

(y x)

u (y ) (y x)u(y ) dS (y ) 2

(y x)u(y ) dy.

(4.3)

We would now like to use the representation formula (4.3) to solve (4.1). If we knew u on and u on and u on , then we could solve for u. But, we dont know all this information. We know u on and u on . We proceed as follows. For each x , we introduce a corrector function hx (y ) which satises the following boundary-value problem, y hx (y ) = 0 hx (y ) = (y x) y y . (4.4)

Now suppose we can nd such a (smooth) function hx which satises (4.4). Then using the same analysis as above, we have hx (y )u(y ) dy =

y hx (y ) y u(y ) dy +

hx (y )

u (y ) dS (y )

y hx (y )u(y ) dy

hx (y )u(y ) dS (y )

u + hx (y ) (y ) dS (y ). Now using the fact that hx is a solution of (4.4), we conclude that 0=

(y x)

hx u (y ) (y )u(y ) dS (y )

hx (y )u(y ) dy.

(4.5)

Now subtracting (4.5) from (4.3), we conclude that u(x) =

hx (y x) (y ) u(y ) dS (y )

[(y x) hx (y )]u(y ) dy.

Let G(x, y ) = (y x) hx (y ). Then, u can be written as u(x) =

G (x, y )u(y ) dS (y )

G(x, y )u(y ) dy.

(4.6)

We dene this function G as the Greens function for . That is, the Greens function for a domain Rn is the function dened as G(x, y ) = (y x) hx (y ) x, y , x = y,

where is the fundamental solution of Laplaces equation and for each x , hx is a solution of (4.5). We leave it as an exercise to verify that G(x, y ) satises (4.2) in the sense of distributions. Conclusion: If u is a (smooth) solution of (4.1) and G(x, y ) is the Greens function for , then G u(x) = (x, y )g (y ) dS (y ) + G(x, y )f (y ) dy. (4.7) 3

We will show below that conversely a function of the form (4.7) will give us a solution of (4.1). First, however, we prove the two claims given above. Proof of Claim 1.
V

(y x)u(y ) dS (y ) =

(y x)u(y ) dS (y ) + 1 y n(n) |y |n

B (x, )

(y x)u(y ) dS (y ).

Now y (y ) = and the outward normal on B (x, ) is = Therefore,

yx . |y x |

(y x) = y (y x) 1 yx yx = n(n) |y x|n |y x| 1 1 = . n(n) |y x|n1 Therefore, 1 (y x)u(y ) dS (y ) = n(n) 1 = n(n) =


B (x, )

B (x, )

B (x, ) n 1

1 u(y ) dS (y ) |y x|n1 u(y ) dS (y )

B (x, )

u(y ) dS (y ).

As

0+ ,
B (x, )

u(y ) dS (y ) u(x). (y x)u(y ) dS (y ) = u(x),

Therefore, we have lim +


0 B (x, )

and the claim follows. Proof of Claim 2. Now we know (y x)


V

u (y ) dS (y ) =

(y x)

u (y ) dS (y ) 4

(y x)
B (x, )

u (y ) dS (y ).

We just need to show that (y x)


B (x, )

u (y ) dS (y ) 0

as

0+ .

Substituting in the explicit formula for for n 3 (the case n = 2 can be handled similarly), we see that (y x)
B (x, )

u 1 1 u (y ) dS (y ) (y ) dS (y ) n(n)(n 2) B (x, ) |y x|n2 1 u dS (y ) L (B (x, )) n(n) n2 B (x, ) C


B (x, )

dS (y )

=C . Therefore, as 0+ , (y x)
B (x, )

u (y ) dS (y ) 0

as claimed. Therefore, the claim follows. Above we have proven the following theorem. Theorem 3. If u C 2 () is a solution of u = f u=g where f and g are continuous, then u(x) =

x Rn x ,

g (y )

G (x, y ) dS (y ) +

f (y )G(x, y ) dy

(4.8)

for x , where G(x, y ) is the Greens function for . Corollary 4. If u is harmonic in and u = g on , then u(x) =

g (y )

G (x, y ) dS (y ).

4.2

Finding Greens Functions

Finding a Greens function is dicult. However, for certain domains with special geometries, it is possible to nd Greens functions. We show some examples below.
2 Example 5. Let R2 + be the upper half-plane in R . That is, let 2 R2 + {(x1 , x2 ) R : x2 > 0}.

We will look for the Greens function for R2 + . In particular, we need to nd a corrector function hx for each x R2 , such that + y hx (y ) = 0 hx (y ) = (y x) y R2 + y R2 +.

Fix x R2 / , then + . We know y (y x) = 0 for all y = x. Therefore, if we choose z y (y z ) = 0 for all y . Now, if we choose z = z (x) appropriately, z / , such that x (y z ) = (y x) for y , then letting h (y ) = (y z (x)), we will have found a corrector function. Recall that for n = 2, (y z ) = 1 ln |y z |. 2

Therefore, (y z ) is a function of |y z |. Consequently, for x = (x1 , x2 ) R2 + , we see that for all y R2 , + |y x| = |(y1 , 0) (x1 , x2 )| = |(y1 , 0) (x1 , x2 )| = |y x| where x (x1 , x2 ) is the reection of x in the plane.

R+ x

y
~

Therefore, letting hx (y ) = (y x), we have found a corrector function for R2 + . Therefore, a Greens function for the upper half-plane is given by G(x, y ) = (y x) (y x) 1 = [ln |y x| ln |y x|] . 2 Example 6. More generally, for the upper half-space in Rn ,
n Rn + {(x1 , . . . , xn ) R : xn > 0},

the corrector function hx (y ) is given by hx (y ) = (y x) 6

where x (x1 , . . . , xn1 , xn ) is the reection of x in the plane. Therefore, a Greens function for the upper half-space Rn + is given by G(x, y ) = (y x) (y x).

Example 7. Let B2 (0, 1) be the unit ball in R2 . That is, let


2 B2 (0, 1) {(x1 , x2 ) R2 : x2 1 + x2 < 1}.

Fix x B2 (0, 1). We need to nd a corrector function hx for B2 (0, 1). Again, (y x) = 1 ln |y x|. 2

Therefore, (y x) is a function of |y x|. We need hx (y ) = (y x) for all y B2 (0, 1), that is, all y such that |y | = 1. Now for y B2 (0, 1), |y x|2 = (y x) (y x) = |y |2 2y x + |x|2 = |x|2 2x y + 1 = |x |2 |y |2 2 x y + 1 2x y 1 = |x |2 |y |2 + |x |2 |x |2 x |x |2 = |x |2 |y |2 2 y 2 + 4 |x | |x | 2 2 = |x | |y x | , where x = is called the point dual to x. x |x |2

B2(0,1) x

Notice that for x B2 (0, 1), x is not in B2 (0, 1). Consequently, we can conclude that (|x|(y x )) is harmonic for all y in . In addition, (|x|(y x )) = (y x) for all y B2 (0, 1). Therefore, letting hx (y ) = (|x|(y x )), we see that hx is a corrector function for the unit ball B2 (0, 1). Consequently, the Greens function for B2 (0, 1) is given by G(x, y ) = (y x) (|x|(y x )) 1 = [ln |y x| ln[|x||y x |]] . 2 Example 8. For the unit ball in Rn ,
2 Bn (0, 1) {(x1 , . . . , xn ) : x2 1 + . . . + xn = 1 } ,

a corrector function hx is given by hx (y ) = (|x|(y x )) where x = x |x |2

is the point dual to x. Therefore, a Greens function for Bn (0, 1) is given by G(x, y ) = (y x) (|x|(y x )).

4.3

Using Greens Functions to Solve Poissons Equation


u = f u=g x x ,

We have shown above that if u is a smooth solution of the Dirichlet problem

then u can be represented in terms of the Greens function for by (4.8). It remains to show the converse. That is, it remains to show that for continuous functions f , g and a given domain Rn , the representation formula (4.8) does give us a solution of the Dirichlet problem. First, however, we will use the representation formula (4.8) to write the proposed formula for the solution in the cases above, where we can explicitly calculate the Greens function for the domain .

n Example 9. Let Rn + be the upper half-space in R , 2 Rn + = {(x1 , . . . , xn ) R : xn > 0}.

From above, we calculated that G(x, y ) = (y x) (y x) is a Greens function for Rn + , where x = (x1 , . . . , xn1 , xn ) and is the fundamental solution of Laplaces equation in Rn . Now, from (4.8), our proposed solution has the form u( x ) =

Rn +

g (y )

G f (y )G(x, y ) dy. (x, y ) dS (y ) + Rn +

(Note: The analysis in Section 4.1 was carried out under the assumption that was a bounded domain. However, for now we will assume we can use the same representation formula to derive a solution for the unbounded half-space. Later, we will need to prove that this representation formula actually gives us a solution.) Now, we calculate G on {yn = 0} to nd an explicit formula for solutions to u = 0 u=g Now x x .

yn (y ) = . yn n(n)|y |n G (x, y ) = (y x) (y x ) yn yn yn x n y n xn = n n(n)|y x| n(n)|y x |n 2 xn = . n(n)|y x|n

Therefore, the normal derivative of G on {yn = 0} is given by

Therefore, if u is the solution of Laplaces equation on the upper half-space with Dirichlet boundary conditions, then we suspect that u will have the form u(x) = 2xn n(n) g (y ) dy. |y x |n (4.9)

n R+

This is called Poissons formula for the half-space Rn + . The function K (x, y ) = 2xn 1 n(n) |x y |n

is called Poissons kernel for the half-space Rn +. 9

Example 10. Let Bn (0, 1) be the unit ball in Rn . We look for a formula for the solution of Laplaces equation in Bn (0, 1) with Dirichlet boundary conditions, u = 0 u=g x Bn (0, 1) x Bn (0, 1). (4.10)

By (4.8), if u is a solution of (4.10), then u will have the form u(x) =


Bn (0,1)

g (y )

G (x, y ) dS (y ).

Now we just need to calculate G on Bn (0, 1) where G is a Greens function for Bn (0, 1). As shown above, G(x, y ) = (y x) (|x|(y x )) is a Greens function for the unit ball in Rn where x x = 2 |x | is the point dual to x. We consider the case when n 3. The case n = 2 can be handled similarly. For n 3, we have 1 1 (y ) = , n(n) |y |n2 which implies y (y ) = . n(n)|y |n Therefore, yx , y (y x) = n(n)|y x|n while (|x|(y x )) = 1 1 n(n) ||x|(y x )|n2 1 = n2 (y x ). |x |

Therefore, y x 1 y (|x|(y x )) = n2 |x | n(n)|y x |n y |x |2 x = n(n)||x|(y x )|n y |x |2 x . = n(n)|y x|n

Now, the unit normal to Bn (0, 1) is given by = y = y. |y | 10

Therefore, the normal derivative of G(x, ) on Bn (0, 1) is given by G (x, y ) = (y x) (|x|(y x )) yx y |x |2 x = y + y n(n)|y x|n n(n)|y x|n |y |2 + x y + |y |2 |x|2 x y = n(n)|y x|n 2 2 |y | (|x| 1) = n(n)|y x|n |x |2 1 . = n(n)|y x|n Therefore, the solution formula for (4.10) is given by u(x) = = 1 | x| n(n)
Bn (0,1) 2

g (y )

G (x, y ) dS (y ) g (y ) dS (y ). |y x |n

Bn (0,1)

We can use this formula to derive the solution formula for Laplaces equation on the ball of radius r with Dirichlet boundary conditions, u = 0 u=g x Bn (0, r) x Bn (0, r). (4.11)

Suppose u is a solution of (4.11), then u(x) = u(rx) is a solution of (4.10) with boundary data g (x) = g (rx). Therefore, by our work above, we see the formula for u is given by u(x) = g (y ) dS (y ) n Bn (0,1) |y x| g (ry ) = (1 |x|2 ) dS (y ) n Bn (0,1) |y x| g (y ) = (1 |x|2 ) dS (y ) n Bn (0,r) |y/r x| g (y ) dS (y ) = rn (1 |x|2 ) n Bn (0,r) |y rx| = Therefore, u(rx) = r2 |rx|2 n(n)r r2 |rx|2 n(n)r
Bn (0,r)

1 | x |2 n(n)

g (y ) dS (y ). |y rx|n g (y ) dS (y ), |y rx|n

Bn (0,r)

11

which implies the solution formula for (4.11) is given by u(x) = r 2 | x |2 n(n)r g (y ) dS (y ). |y x |n (4.12)

Bn (0,r)

This representation formula is called Poissons formula for the ball. The function K (x, y ) is called Poissons kernel for the ball. As mentioned above, formulas (4.9) and (4.12) will give solutions of u = 0 u=g x x r 2 | x |2 n(n)r|y x|n (4.13)

in the cases when is the half-space and the ball respectively, assuming a solution exists. We now will show that these formulas actually solve Laplaces equation with Dirichlet boundary conditions, u = g for x . In particular, we state the following theorems, (Ref: Evans, Sec. 2.2), Theorem 11. For g C (Rn1 ) L (Rn1 ) and u dened by (4.9), u satises the following,
n 1. u C (Rn + ) L (R+ )

2. u = 0 for x Rn + 3. xlim u(x) = g (x0 ) for all x0 Rn +. x


x
0 Rn +

Theorem 12. For g C (Bn (0, r)) and u dened by (4.12), u satises the following, 1. u C (Bn (0, r)) 2. u = 0 for x Bn (0, r) 3.
xx0 xBn (0,r)

lim

u(x) = g (x0 ) for all x0 Bn (0, r).

We will prove the rst of these theorems. The second one follows similarly. In order to prove this theorem, we make use of the following lemma. Lemma 13. Greens functions are symmetric. For all x, y , x = y , G(x, y ) = G(y, x).

12

Proof. (Ref: Evans, Sec. 2.2) Fix x, y , x = y . Let v (z ) G(x, z ) w(z ) G(y, z ). We will show that v (y ) = w(x), and, therefore, G(x, y ) = G(y, x). Recall that G(x, y ) = (y x) hx (y ) where hx (y ) satises y hx (y ) = 0 hx (y ) = (y x) Therefore, for z , v (z ) = G(x, z ) = (z x) hx (z ) = (z x) (z x) = 0 w(z ) = G(y, z ) = (z y ) hy (z ) = (z y ) (z y ) = 0. Further, z v = 0 for z = x and z w = 0 for z = y . Now v is smooth, except near z = x, while w is smooth, except near z = y . Dene the region V = [B (x, ) B (y, )] for > 0. On V , our functions are smooth. Therefore, we can use our integration by parts formula as follows, vw dz =
V V

y y .

v w dz +
V

v w dS (z )
V

=
V

v w dz
V

w dS (z ) + v

v w dS (z )

Using the fact that v = 0 = w on V , we conclude that v


V

w dS (z ) =

v w dS (z ). w v v w dS (z )

Using the fact that v = 0 = w on , we conclude that v w w v dS (z ) =

B (x, )

B (y, )

where denotes the inward pointing unit vector eld on B (x, ) B (y, ). Now we claim that as 0+ , the left-hand side converges to w(x), while the right-hand side converges to v (y ). For the terms on the left-hand side, we rst look at
B (x, )

w v dS (z ).

Now w is smooth near x. Therefore, w is bounded near B (x, ). Now v (z ) = (z x) hx (z ). Therefore, on B (x, ), v (z ) 1/ n2 . Therefore, w v dS (z ) C sup |v | B (x, ) =C
n1 B (x, )

dS (z )
B (x, )

B (x, )

sup |v | = O( ) 0 as 0.

13

Next, we look at
B (x, )

v w dS (z ).

Now

v hx w dS (z ) = (z x) (z ) w dS (z ). B (x, ) B (x, ) First, using the fact that hx is smooth and w is smooth near x, we see that
B (x, )

hx (z )w dS (z ) C C

dS (z )
B (x, ) n1

Therefore,
B (x, )

hx (z )w dS (z ) 0 as

0.

For the other term, we see that 1 (z x)w(z ) dS (z ) = n(n) B (x, ) 1 = n(n) =
B (x, )

B (x, ) n1

1 w(z ) dS (z ) |z x|n1 w(z ) dS (z )

B (x, )

w(z ) dS (z ) w(x) as

0.

Similarly, the right-hand side converges to v (y ). Therefore, the lemma is proven. We will now prove the rst of the theorems above. The second one follows similarly. Proof of Theorem 11. For u dened by (4.9), the Poisson kernel is given by K (x, y ) = Now, we know that y G(x, y ) = 0 for y = x. In addition, using the fact that G is symmetric, we see that x G(x, y ) = x G(y, x) = 0 for x = y. Therefore, x u(x) = x =

2xn 1 = G(x, y ). n(n) |x y |n yn

G(x, y )g (y ) dS (y ) yn Rn + x y G(x, y ) (y )g (y ) dS (y ) y x G(x, y ) (y )g (y ) dS (y ) = 0.

Rn + Rn +

14

Therefore, x u = 0. Using this same reasoning, and the fact that for each y = x, G(x, y ) C (Rn ) for x = y , we can show that u C . To show that lim u(x) = g (x0 ) for all x0 Rn +, xx
x
0 Rn +

we will use the fact that for each x Rn +,

Rn +

K (x, y ) dy = 1.

(4.14)

This can be seen by a direct calculation, and we omit this proof here. Using this fact, we proceed as follows. Fix x0 Rn + , > 0. We need to show that there exists a > 0 such that |u(x) g (x0 )| < for |x x0 | < . Now using (4.14), we see |u(x) g (x0 )| =

Rn + Rn +

K (x, y )g (y ) dy

Rn +

K (x, y )g (x0 ) dy

K (x, y )[g (y ) g (x0 )] dy .

Now, we know that K (x, y ) has a singularity at y = x0 , and we are considering K (x, y ) as x x0 . Therefore, we divide the integral into two pieces and handle them separately. K (x, y )[g (y ) g (x0 )] dy =
B (x0 , )

Rn +

K (x, y )[g (y ) g (x0 )] dy +

Rn + B (x0 , )

K (x, y )[g (y ) g (x0 )] dy

I + J. We rst look at term I . Using the assumption that g is continuous, we have |g (y ) g (x0 )|L (B (x0 , )) < and K (x, y ) dy
B (x0 , )

for suciently small,

Rn +

K (x, y ) dy = 1.

Therefore, |I | < 2 for suciently small. Now for this choice of , consider term J . J=

Rn + \B (x0 , )

K (x, y )[g (y ) g (x0 )] dy.

15

On J , |y x0 | > . Take |x x0 | < /2. Then |y x 0 | | y x | + |x x 0 | | y x | + 1 | y x | + |y x 0 |. 2 2

Therefore, |y x| 1 |y x0 | for |x x0 | < /2. This implies |y x|n 2n |y x0 |n . 2 Therefore, |J | 2|g |L

Rn + B (x0 , ) Rn + B (x0 , ) Rn + B (x0 , )

K (x, y ) dy xn dy n(n)|y x|n 2n xn dy n(n)|y x0 |n 1 dy 0 as x x0 . |y x0 |n

= 2 |g |L

2|g |L

2n+1 xn |g |L n(n)

Rn + B (x0 , )

Therefore, we can make J arbitrarily small by choosing x suciently close to x0 .

16