Jackson Notes 3 2012

1 Method of images
The method of images is a method that allows us to solve certain potential problems as well as obtaining a Greens function for certain spaces. Recall that the Greens function satises the equation 2 G (x0 , x) = 4 (x x0 ) (1) subject to the boundary conditions GD = 0 on S or and thus the solution is of the form G = C on S n 1 + (x0 , x) |x x0 | (2) (3) (4)

G (x0 , x) = where

2 (x0 , x) = 0 in V (5) Thus the Greens function is 40 times the potential at x due to a unit point charge at x0 in the volume V plus an additional term, with no sources in V, that xes up the boundary conditions, that is, a term due to sources outside of V. Thus we can make progress in nding the Dirichlet Greens function by nding the potential due to a point charge in V with grounded boundaries. 1.1

Plane boundary

Suppose the volume of interest is the half-space z > 0. A point charge q is placed at a distance d from the x y plane, which is a conducting boundary. What is the potential for z > 0? First, the conducting plane must be at a constant potential, which we may take to be zero. (If the potential is V0 , we can just add V0 to our solution at the end.) Then we may place an image charge q at distance d from the boundary, but on the opposite side. Then the potential due to these two charges everywhere on the boundary z = 0 is zero. Since the image charge we added is outside our volume, it does not contribute to the value of 2 in V. Thus, putting the z axis through q, the solution is q 1 q (x) = 40 |x dz | |x + dz | Properties of this solution are explored in Jackson problem 2.1. Note that the image charge represents the charge density drawn onto the plane through the ground wire. To obtain the Greens function for the half-space, we simply set q = 1 and multiply by 40 . Then 1 1 G (x, x0 ) = (6) |x x0 | |x x00 |

where x00 = (x0 , y 0 , z 0 ) is the position vector of the image point. This expression shows clearly that the physial dimensions of G are [1/length]. This Greens function (6) is explicitly in the form (4), but it is not very convenient to use. For example, suppose the problem of interest has potential V0 within a circle of radius a on the plane z = 0, with the rest of the plane grounded. Then we would need rst to compute (notes 2.5 equation 6) G G 2 z = = 3 0 n0 0 z 0 0 |x x |
z =0 z =0 z 0 =0

where r, are polar coordinates in the x y plane. This is ugly. (We encountered a similar integral in the bar magnet example, but that was only a 1/2 power. Later on we will identify some methods for doing this integral, but we will also nd more convenient expressions for G.) 1.2

(note that the outward normal on the plane at z = 0 is z ) and then evaluate Z V0 2z (x ) = dx0 dy 0 4 circle |x x0 |3 0 z =0 Z V0 z 1 = r0 dr0 d0 2 circle r2 + r02 2rr0 cos 0 + z 2 3/2

Images in a sphere

Now let the conducting boundary be a sphere of radius a and suppose we have a point charge q at a distance d from the center of the sphere, where d > a. We want to nd the potential outside the sphere. Learning from our experience above, we conjecture that we can place an image charge inside the sphere (and thus outside our volume V ) and form the potential in V as the sum of the potential due to the two charges. Let the image charge have magnitude q 0 and be at r = d0 . Then we have 0 q 1 q + (x) = 40 x d x d0 The boundary condition is that (x) = 0 everywhere on the surface of the sphere (r = a). The system has azimuthal symmetry about the line from the center of the sphere to the charge q. Rotate the system about this line and nothing changes. Thus the image charge q 0 must lie on this line at a distance d0 from the center. Then we have two unknowns in our potential: q 0 and d0 , and we need only pick two points on the sphere to solve for the two unknowns. The most convenient two points lie on the ends of the diameter through the image charge, as shown (P and Q in the diagram).

P = and Q = Then from equation (7) and from (8)

1 40 1 40

q q0 + 0 d+a d +a q q0 + d a a d0

=0 =0

(7) (8)

q (d0 + a) + q 0 (d + a) = 0 q (a d0 ) + q 0 (d a) = 0 Adding these two relations eliminates d0 and gives a 2qa + 2q 0 d = 0 q 0 = q (9) d This result has the nice property that the image charge is negative if q is positive, as we found in the planar case. Once again the image charge represents the charge drawn onto the surface of the sphere through the ground wire. Now we subtract the two relations to obtain an expression for d0 : a a2 q0 = a = (10) q d d The image charge is inside the sphere if d > a, as we need. Conversely, if d < a, the image charge is outside the sphere. (You are asked to conrm this result in Problem 2.2.) We can perform two checks on this result. First lets nd the potential at an arbitrary d0 = a 2qd0 + 2q 0 a = 0

point on the sphere. 40 (a, ) = = q q0 + 0 R R q qa/d q 2 d2 + a2 2ad cos (d0 ) + a2 2ad0 cos 1 q a q =0 q d 1 + a2 2 a cos (a2 /d)2 + a2 2 (a3 /d) cos d2 d

for all , as expected. Second, lets check that we get back the result from 1.1 as a (plane boundary). We have to be a bit careful here, because if we immediately let a , the point from which we are measuring our distances moves off innitely far to the left, and we will learn nothing. So rst we write our results in terms of distance from the surface of the sphere. The charge q is a distance d a = h from the surface, and then a q q 0 = q = q as a a+h 1 + h/a as required. The distance of the image from the surface is a2 ah h = = h as a a+h a+h 1 + h/a and this is the second required result. Finally we can write the Dirichlet Greens function for the region outside a sphere of radius a by setting q = 1 and multiplying by 40 : 1 1 a GD (x, x0 ) = |x x0 | r0 |x x00 | where x00 is the position vector of the image point. Again this is pretty ugly. h0 = a d0 = a 1.3

Images in a cylinder

The basic ideas and methods are the same as we have used in the plane and sphere cases. See Jackson problem 2.11. 1.4

Use of images to solve problems

Jackson P 2.10 asks us to compute the potential inside a parallel-plate capacitor with a small hemispherical boss on one plate.

The clue is in part (c) of the problem: We model the system by putting a point charge q at a very large distance d from the plane. Then we can use the image system shown to model the capacitor, since the image system puts potential zero on the lower plate. (q and q, q 0 and q 0 form pairs that make the potential on the plane zero; q and q 0 , q and q 0 form pairs that make the potential on the sphere zero.) Then we can show that as d we obtain a uniform eld at large distance from the lower plate, and we set that uniform eld equal to the given value of E0 , thus determining the necessary charge q. a2 0 Setup: From 1.2, the image charge q 0 = a d q and its distance from the plane is d = d . The potential due to the four charges (one real charge and three images) is ( ) r2 + d2 2rd cos 1/ r2q + d2 + 2rd cos 1/ q a (r, ) = kq (11) a4 a2 a4 a2 2 2 a d / r + d2 2r d cos + d / r + d2 + 2r d cos To see why this works, look at the potential for d r a. We drop terms in (a/d)2 to get q q r2 r r2 r kq 1/ d2 + 1 2 d cos 1/ d2 + 1 + 2 d cos q q (r, ) ' d a / 1 2 a2 cos + a / 1 + 2 a2 cos
r dr r rd 2

Next expand the square roots, dropping terms in (r/d) and (a/r) (a/d). a ao kq n r r (r, ) ' 1 + cos 1 cos + d d d r r kq r 2kq = 2 cos = 2 z to rst order in small quantities d d d

This corresponds to the given uniform eld provided that E0 = 2kq/d2 , or, if we choose q = E0 d2 /2k. (12)

Solve: (a) To nd the surface charge densities, begin with the eld components. From (11), we have ( r + d cos r d cos + = kq 3/2 3/2 2 2 2 r (r + d 2rd cos ) (r + d2 + 2rd cos ) a r a2 cos /d a r + a2 cos /d + 3/2 3/2 4 4 a2 a2 d r2 + a d r2 + a d2 2r d cos d2 + 2r d cos and at r = a Er = kq ( (a2 + d2 a d cos 2ad cos ) a a a2 cos /d
a4 d2
2

3/2

a + d cos

(a2

= kq

d a2 + (a2 + d2

2a a d cos
3/2

a d cos

= kq

(a2 + d2 2ad cos )3/2 (a2 + d2 + 2ad cos )3/2 # " d2 1 1 = kqa 1 2 3/2 3/2 a (a2 + d2 2ad cos ) (a2 + d2 + 2ad cos )

"

a3 (d2 + a2 2ad cos )

d a

da2 (d a cos )

2ad cos )

3/2 (a2

+ d2 + 2ad cos )3/2 2 a a + a cos /d + 3/2 4 a2 d a2 + a d2 + 2a d cos a + d cos + d2 + 2ad cos )3/2 da2 (d + a cos ) a3 (d2 + a2 + 2ad cos )
d a 3/2

3/2

a d cos

(d a cos )

a + d cos

(d + a cos )

and thus the charge density is (notes 1 eqn 5 with E = 0 inside the boss):

# " 1 d2 1 () = 0 Er = 0 kqa 1 2 a (a2 + d2 2ad cos )3/2 (a2 + d2 + 2ad cos )3/2 Analysis: Notice that is zero in the corners at = /2, .as expected. Also since

1 cos 0 on the boss and d > a, the chrage density is negative on the boss if q is positive, as expected. Solve: The total charge on the boss is Z /2 Qboss = 2 () a2 sin d 0 ! Z 1 2 3 1 1 d2 = d qa 1 2 3/2 3/2 4 a 0 (a2 + d2 2ad) (a2 + d2 + 2ad) !1 1 qa3 2 d2 2 = 1 2 2 a 2ad (a2 + d2 2da)1/2 (a2 + d2 + 2da)1/2
0

! q 2 1 d2 a2 1 Qboss = + 1 2 1/2 2 a d (a2 + d2 2ad)1/2 a2 + d2 (a2 + d2 + 2ad) 1 q 1 1 2 = a2 d2 + 2 2 d da a+d a + d2 2 2 d a Qboss = q 1 d a2 + d2 which is Jacksons result. Analysis: As d for xed q, Qboss 0. . Is this what you would have expected? The induced charge also goes to zero as a 0, .as the boss disappears in this case. Solve: The charge density on the plane is

z z=0 Here it is more convenient to express the potential in terms of z : () = 0 Ez = 0 (, z ) = kq ( 1 1 p p 2 2 2 2 2 z + + d 2zd z + + d2 + 2zd q d z 2 + 2 + a
a4 d2

2z a d

+ q d z 2 + 2 +

a4 d2

+ 2z a d

(13)

Thus () = 0 kq

"

(z 2

(z d) d2

where the second term is negligible if d a. The total charge on the plane is: Z 3 qd 1 a d Qplane = 2 3/2 4 3/2 2 2 2 3 2 a ( + d ) d +a 2 d qd a3 (2) 2 = 1 / 2 1 / 2 4 2 (2 + d2 ) d3 2 + a 2 d a 1 a3 = qd 1/2 4 1/2 (a2 + d2 ) d3 a2 + a d2

+ 4 a2 3/2 d z 2 + 2 + a d2 2z d qd a3 1 = 4 3/2 2 (2 + d2 )3/2 d3 2 + a d2

+ 2zd) a z a2 /d

3/2

(z 2

+ 2 + d2 + 2zd) 2 a z + a /d 4 a2 3/2 d z 2 + 2 + a + 2 z d2 d z =0

(z + d)

3/2

as expected. Solve: Now lets put in the value for q that gets us to the capacitor-plus-boss system (eqn 12): q = E0 d2 /2k. Then, for d a, the charge on the boss is: ! a 4 E0 a2 1 a2 1 a2 /d2 ' d2 1 1 2 Qboss = q 1 p 1 + O 2k d 2 d2 d a2 /d2 + 1 2 E0 3a 3 E0 2 = d2 = a = 30 E0 a2 2 2k 2d 4 k Analysis: This is Jacksons answer in (b). Notice that d disappears in the limit d , as required.

q d2 a2 = d d2 + a2 Analysis: Qplane q as a 0 (at plate) or d . The total induced charge is: d2 a2 1 d2 a2 Qboss + Qplane = q 1 q d d2 + a2 d a2 + d2 = q

Solve: The charge densities are: ! d2 d2 d2 a () = 0 E0 1 2 2 a (a2 + d2 2ad cos )3/2 (a2 + d2 + 2ad cos )3/2 ! 1 1 d2 a 1 2 = 0 E0 3/2 3/2 2d a (a2 /d2 + 1 2a/d cos ) (a2 /d2 + 1 + 2a/d cos ) 2 a d a a ' 0 E0 1 + 3 cos 1 3 cos 2 2d a d d 2 a a d = 0 E0 6 cos = 30 E0 cos 2 2d a d on the boss, and on the plane 3 3 a E0 d 1 () = 3/2 4 3/2 2 2 4k 3 2 ( + d ) d +a 2 d = 3 1 + E0 3 a4 32 a3 = 1 2 3 1 4k 2d 2 d2 2 3 a 0 E0 1 3 as d Analysis: Note that the charge density is zero where the boss meets the plate ( = a, = /2), as expected near a sharp hole in the conductor. Also 0 E0 as , the expected result for a at plate. The plot shows /0 E0 versus /a for > a and versus 2/ for 0 < < /2 E0 32 1 2 4k 2d a3
a4 d2 2

3/2

rho/a

s/s0

-1

-2

-3

Solve: The potential (13) is 1 1 kq q q (, z ) = z 2 2 2 2 d 1 + d + d 2z 1+ z + + 2z d d d d a 2 a 2

q + q z 2 a4 z a2 z 2 d + + 2 d + d + 4 2 d d d dd d n h io E0 d z a2 z z E0 a p ' 2 2 1+ 1 2 d d 2 d z + 2 z 2 + 2 ! a3 as d ' E0 z 1 3/2 (z 2 + 2 )

a4 d4

a + 2z d d2

Analysis: This function is plotted below. All distances are scaled by a.

Values of /E0 a are: black 0.2, red 0.5, green 1, purple 2

2 Expansion in orthogonal functions

To obtain a more useful form of the Greens function, well want to expand in orthogonal functions that are (relatively) easy to integrate. We begin with some basics and well see how to obtain solutions for the potential in boundary value problems with no charges inside the volume. Then we can move on to compute the Greens function. 2.1

Sturm-Liouville theory

See Lea Chapter 8 sections 1 and 2. 2.2

General method for nding the potential

1. Choose coordinates so that the boundaries of the region correspond to constant-coordinate surfaces. Then write the dening equation 2 = 0 10

in terms of the chosen coordinates, called u, v, and w, separate variables, and solve to determine the eigenfunctions. Youll get equations something like: Du U Dv V = ( + ) ; = ; U V Du is a differential operator of the form f (u) and similarly for Dv and Dw . 2Note which boundary has a non-zero value of potential. This boundary should be dened by the condition u = constant for one of the coordinates, u. You will want to choose your separation constants so that the sets of eigenfunctions in the other two coordinates v and w are orthogonal functions. 3Then use the boundaries on which = 0 to determine your eigenfunctions V and W, vV wW and eigenvalues, (which Ill call = DV and = DW ). If the boundaries are at a nite distance the eigenvalues will be countable and can be labelled with an integer: m . Otherwise they will form a continuous set. 4The last eigenfunction is determined from the differential equation in the nal coordinate, u: Du U = = U At this point you should have a general solution of the form: X Amn Umn (u : m n ) Vm (v : m ) Wn (w : n ) (u, v, w ) =
m,n

Dw W = W

d2 d + g (u) + h (u) du2 du

where the coefcients Amn are still undetermined.

5Now use the nal, non-zero, boundary condition together with orthogonality of the eigenfunctions Vm and Wn to nd the coefcients Amn 2.3

Rectangular coordinates

See Lea 8.2 and Grifths Ch 3 3, Jackson 2.9, or check here: http://www.physics.sfsu.edu/~lea/courses/ugrad/360notes7.PDF 2.3.1 Rectangular 2-D problems with non-zero potential on two sides. Since the general method allows for non-zero potential on only one side, we have to solve these problems by superposition. Suppose we have a rectangular region measuring a b with potential V on the sides at x = 0 and y = b, the sides at x = a and y = 0 being grounded. We solve two problems, each with three sides grounded, one having potential V at x = 0 and one having potential V at y = b. Then we add the results. The solution is = 1 + 2

11

where (Lea Example 8.1) (2m+1)y 4V X (2m + 1) x sinh 1 a 2 = sin m=0 2m + 1 a sinh (2m+1)b
a

and

(2n+1) (ax) 4V X 1 (2n + 1) y sinh b 1 = sin a n=0 2n + 1 b sinh (2n+1) b With a = 2b the potential looks like this (the axes are x/a and y/a):

0.5 0.4 0.3 y 0.2 0.1 0

0.2

0.4

0.6

0.8

Blue 0.75, black, 0.5, red 0.25, dashed 0.9 2.3.2 Continuous set of eigenvalues If one dimension of our rectangle becomes innite, then the set of eigenvalues is no longer countable but becomes a continuous set. Suppose we let b in the example above. With (x, y ) = X (x) Y (y ) , Laplaces equation takes the form X 00 Y 00 = k2 = X Y . The solutions to the differential equation that satisfy the boundary conditions = 0 at y = 0 and at x = a but non-zero at x = 0 are of the form But we no longer have an upper boundary in y to determine the values of k. (Effectively, 1 in the previous problem has gone to zero because of the b in the denominator of the argument, leaving 2 with an undetermined k.). However, since the eigenfunction is an even function of k, we need only positive values of k. Thus the solution is Z A (k ) sin ky sinh [k (a x)] dk (x, y ) =
0

sin ky sinh k (a x)

To nd the function A (k) , we use the known potential V (y ) at x = 0 : Z (0, y ) = V (y ) = A (k ) sin ky sinh ka dk
0

12

We now make use of the orthogonality of the functions sin ky by multiplying both sides by sin k 0 y and integrating over y : Z Z Z V (y ) sin k0 y dy = A (k) sin ky sinh ka dk sin k 0 y dy
0 0 0

For example, if V (y ) = V0 ey/c , then Z 2 sin ky V0 ey/c A (k ) = dy 0 sinh ka (ik1/c)y e(ik1/c)y e 1 V0 = i sinh ka ik 1 ik 1 c c 0 1 V0 1 1 = i sinh ka ik 1 ik + 1 c c 1 V0 2ic2 k = i sinh ka k 2 c2 + 1 and thus Z 2c2 k2 sin ky sinh [k (a x)] V0 dk (x, y ) = 0 k2 c2 + 1 k sinh ka Analysis: We can see right away that our result is dimensionally correct. Also, since x a, k(a x) ka and thus sinh [k(a x)] 1 sinh ka for all k, so the integral converges. I couldnt do this integral analytically, so lets compute some values in the case c = a: 13

Z 1 [cos (k k 0 ) y cos (k + k 0 ) y ] dy 2 0 Z0 o 0 0 0 1 n i(kk0 )y + ei(kk )y ei(k+k )y ei(k+k )y dy e = 4 0 Z 1 {exp [i (k k0 ) y ] exp [i (k + k 0 ) y ]} dy = 4 = [ (k k 0 ) (k + k 0 )] 2 where we used Lea eqn 6.16. Thus Z Z V (y ) sin k0 y dy = A (k ) sinh ka [ (k k 0 ) (k + k 0 )] dk 2 0 0 Since k and k 0 are both positive, only the rst delta function contributes, and this integral reduces to Z V (y ) sin k 0 y dy = A (k0 ) sinh k 0 a 2 0 and thus, dropping the primes, Z sin ky 2 V (y ) A (k ) = dy 0 sinh ka =

On the RHS, we have Z sin ky sin k 0 y dy

 a 2 a 2 , 2 /V0 = 0.314 45 = 0.20019 a 2 2 , a /V0 = 0.281 18 = 0.179 a 2 0.119 04 = 7.578 3 102 2 , 2a /V0 = The potential decreases rapidly for y > a 2.3.3 A 3-D problem Find the potential inside a cubical of side i a with grounded walls, except for the side at h box 2 x z = a which has potential V0 1 2a 1 . This potential increases from zero at the walls to V0 at x = a/2,as shown in the diagram.
1 0.8 0.6 ential 0.4 0.2

0.2

0.4

x/a

0.6

0.8

There is no charge inside the box, so the potential satises Laplaces equation: 2 = 0 2 + + = 0 x2 y 2 z 2
2 2

Try separating variables: = XY Z Then X 00 Y Z + XY 00 Z + XY Z 00 = 0 Y 00 Z 00 X 00 + + =0 (14) X Y Z For equation 14 to hold for all values of x, y, z we must have: Y 00 Z 00 X 00 = k1 , = k2 , = k3 and k1 + k2 + k3 = 0 X Y Z The potential is zero at x = 0 and at x = a, so we need k1 to be negative, k1 = 2 , which gives the solutions X = sin x and X = cos x, which have multiple zeros. We choose the sine function to make X (0) = 0, and then choose = n/a to make X (a) = 0. A similar reasoning gets Y = sin (my/a) . Then 2 k3 = n2 + m2 2 a 14 or

and to make Z (0) = 0 the appropriate solution for Z is the sinh . X p x z y (x, y, z ) = Anm sin n sin m sinh (n2 + m2 ) a a a n,m=1

We make use of the orhogonality of the sines by multiplying both sides by sin n0 x/a sin m0 y/a and integrating over x and y : 2 # Z aZ a " q x y 2x a2 V0 1 sin n0 1 sin m0 dxdy = An0 m0 sinh (n0 )2 + (m0 )2 a a a 4 0 0

Finally we evaluate the potential at z = a : " 2 # X p x 2x y V0 1 Anm sin n sin m sinh (n2 + m2 ) = 1 a a a n,m=1

Drop the primes, and let u = 2x/a 1, to get a Z 1 2 p u+1 a cos my/a 2 = Anm a sinh (n2 + m2 ) V 1 u sin n du 2 2 m/a 4 1 0 Z 1 m p a nu n nu n 1 (1) a2 2 ) V0 1 u2 sin (15) cos + cos sin du = Anm sinh (n2 + m 2 1 2 2 2 2 m/a 4 The result is zero unless m is odd. When we integrate over u the sine term gives zero because we have an odd integrand over an even range. For the cosine term, we do integation by parts: +1 Z 1 Z 1 nu nu 2 nu 2 u2 cos 1 u cos du = sin du 2 2 n 2 2 1 0 1 " # +1 Z 1 2 nu n nu 2 4 2 u sin sin 2 u2 sin du = n 2 n 2 0 n 0 2 Z nu 1 u2 cos du = 2 1
1

4 n sin n 2 +1 2 Z 1 2 2 nu 2 nu n 2 2 cos cos sin + 2u du 2 n 2 n 2 n 2 0 0 " # 3 3 2 n nu 2 +1 = 2 cos sin 2 2 n 2 0 (n )2 = 16 (n)

2

cos

n n 32 sin + 3 2 2 (n )

15

and so the left hand side of (15) is ! a2 n 32 16 n sin2 n V0 2 sin 2 cos 2 m (n )3 2 (n) ! 8 16 a2 = V0 (1 cos n) sin n m (n )3 (n)2 = V0 = V0 Thus for m odd and n odd Anm = V0 So
X 128 (x, y, z ) = 4 V0 X

a2 16 n (1 (1) ) m (n )3 32a2 for m and n odd mn3 4 128 1 mn3 4 sinh n2 + m2 1 nx my sinh n2 + m2 z a sin sin mn3 a a sinh n2 + m2

m=1,odd n=1,odd

Analysis: The result is dimensionally correct, and converges nicely. The plot shows the series up to m = n = 5. The black line is the potential as a function of z for x = y = a/2, and the red line for x = y = a/4. The potential increases faster near the center of the box, as expected, since the potential on the top side is maximum at x = a/2. .

0.8 0.6 phi/V0 0.4 0.2

0.2

0.4

z/a 0.6

0.8

At z = a/2, y = a/2 the potential versus x/a looks like this:

16

0.14 0.12 0.1 0.08 hi/V0 0.06 0.04 0.02 0

0.2

0.4

x/a

0.6

0.8

3 Conformal mapping
Recall that both the real and imaginary parts of an analytic function satisfy Laplaces equation in two dimensions. We may use this fact to solve 2-d potential problems. 3.0.4 Potential in a wedge. Suppose the region of interest is dened by the angular wedge 0 . Further suppose the potential is zero on the conducting boundaries at = 0 and = . Then the analytic function f (z ) = z / = r/ ei/ = r/ cos + i sin has imaginary part v (r, ) = r/ sin that satises the boundary conditions. If = /m for some integer m, then f (z ) is analytic everywhere. If is an arbitrary real number, then f (z ) may have a branch point at the origin, but we may choose the branch cut so that f (z ) is still analytic in our region everywhere except AT the origin. Thus v (r, ) also satises Laplaces equation in the volume, and so it is a solution of the correct form. In fact the function f (z ) = z n/ for any n has the same nice properties. Thus the potential in a wedge-shaped region with opening angle and conducting boundaries at potential V0 is described by the complex potential (z ) = A + iV0 + n an z n/ The imaginary part is n n= which is the potential in the region. The coefcients an must be chosen to satisfy any remaining boundary conditions in r. Compare with Jackson 2.72 which he derives using separation of variables in plane polar coordinates. The sum is over positive n if the origin is included within our region. The solution near the origin is dominated by the rst (n = 1) term. Then the eld near the origin has V0 + an rn/ sin 17
+ X

h i E = a1 r/1 sin r + cos Thus E 0 as r 0 if / > 1 but E as r 0 if / < 1. The eld is small in a hole ( < ) but very large near a spike ( > ). The diagram shows the cases of = /2 and 3/2. For = /2, the equipotential surfaces are r2 sin 2 = constant = 2xy Thus the equipotentials are hyperbolae.

components

y5
4 3 2 1 0 0 1 2 3 4 5

y 12
10 8 6 4 2 -6 -4 -2 -2 -4 -6 2 4 6 8 10 12

= 3/2 equipotentials and eld lines for = /2 We may use conformal mapping to nd the potential in a 2-d boundary value problem with more complex boundaries. See Lea Ch 2 section 2.8. We choose the mapping to simplify the boundary shape, solve the simpler problem, and then map back.

4 Finite element analysis

The big idea: we want a procedure for numerically solving a potential problem for a source g (x) in a region R with specied boundary conditions. Lets look at a Dirichlet problem in 2 dimensions. We start with two general relations. First, if 2 = g in R, then Z 2 + g dxdy = 0
R

for any function . The next relation we need is Greens identity from Chapter 1 (Notes 2.5 eqn 1). Z Z 2 + dV = dA = 0 S n where we have chosen the function to be zero on the boundary. (Note: in a two dimensional problem the volume is an area and the bounding surface S is actually a curve.) Now inserting our differential equation for , we get: Z g dxdy = 0 (16)
R

18

The next step is to set up a grid over the region R. We expand the desired solution in a set of functions ij each of which is zero except on a small region around the grid point (i, j ) . For example, let: ( |y y | xi | 1 |x 1 h j for |x xi | < h, |y yj | < h h ij = 0 otherwise

and then let (x, y ) =

Effectively, the function kl smears the potential value kl over one grid square, converting a set of numerical values kl into a continuous function. (x, y ) . Now stuff this assumed form into the integral 16, and take the function ij . Z X Z kl kl ij dxdy = ij g (x, y ) dxdy (17)
R k,l R

X
k,l

kl kl (x, y )

The integral on the right is non zero only on a square region surrounding (xi , yj ) , and because the function ij is peaky we can approximate the integral as: Z Z ij g (x, y ) dxdy ' g (xi , yj ) ij dxdy
R square

which may be evaluated as follows. Let u = x xi and v = y yj . Then: Z ! Z ! Z 0 Z 0 Z h h u u v v ij dxdy = 1 1+ 1 1+ du + du dv + dv h h h h square 0 h 0 h h h h h = + + = h2 2 2 2 2 Thus (17) becomes: Z X kl ij dxdy = h2 g (xi , yj ) kl (18)
k,l R

19

where the sum on the left is over all the grid points. However, the integrand is non-zero only on a small rectangular region 2h by 2h surrounding the grid point (i, j ). First note that inside this square region, 1 v u 1 ij = 1 x + 1 y h h h h where in the rst factor of each term we take the upper sign on the right half of our square (0 < u < h) and the bottom sign on the left (h < u < 0); in the second factor we take the top sign on the top of the box (0 < v < h) and the bottom sign on the bottom (h < v < 0). Thus if k = i and l = j : Z Z h Z h 1 u 2 v 2 ij i,j dxdy = + 1 1 dudv h2 0 0 h h box Z h Z 0 1 u 2 v 2 + 2 + 1 1+ dudv h 0 h h h Z 0 Z h 1 u 2 v 2 + 2 + 1+ 1 dudv h h 0 h h Z 0 Z 0 1 u 2 v 2 + 2 + 1+ 1+ dudv h h h h h Z 1Z 1 2 = 4 + 2 dd 1 4 3 1 8 = + 3 0 = 3 3 0 v v (terms1 and 3), = 1 + h (terms 2 and 4), where in the four terms we took = 1 h u u = 1 h (terms1 and 2) or = 1 + h (terms 3 and 4). Also if k = i + 1 and l = j : |x xi h| |y yj | i+1,j = 1 1 h h |u h| |v | = 1 1 for |u h| < h and |v | < h h h
0 0

and zero otherwise. The overlap region is on the right side of our box (0 < u < h) where 1 v u 1 i+1,j = 1 x + y h h h h

20

giving Z

box

ij i+1,j dxdy

1 1 2 3 3 + = 2 3 2 3 0 0 1 2 1 = 2 = 2 3 3 We get the same result in all the overlap regions. Thus equation (18) may be written as a matrix equation: K= G where the matrix K is described as sparse it has only a few non-vanishing elements and they are all near the diagonal, like this: 8 1 1 0 0 1 8 1 1 0 1 1 1 8 1 1 K= 3 0 1 1 8 1 0 0 1 1 8

v 2 u u 1 + 1 dudv h h h 0 0 Z 0 Z h u u v 2 1 + 1 1+ + 2 dudv h h 0 h h h Z 1Z 1 2 + (1 ) dd = = 1 h2 Z
h

Matrices of this type are relatively easy to invert numerically. The column vectors and G contain the values of the potential and the source at the grid points. and we have reduced the potential problem to a matrix inversion. = K1 G

For regions with odd shapes, the square grid we used above does not t very well. Triangles of arbitrary size and shape can be t onto a region of almost any shape. So now well modify the method above to use triangles instead of squares. By varying the sizes and shapes of the triangles we can also get better resolution where things change more rapidly. The basic triangular element has vertices at points that we label 1,2 and 3 with coordinates (x1 , y1 ) , (x2 , y2 ) , and (x3 , y3 ) . We approximate the potential solution (x, y ) in this triangle by a Taylor-series-type expansion of the form: (x, y ) = A + Bx + Cy The 3 values of the potential at the 3 vertices give enough information to evaluate the 3 coefcients A, B and C. To make the numerical computation more efcient, it is convenient to dene three shape functions Ni (x, y ) = ai + bi x + ci y that have the properties: Ni (xi , yi ) = 1 and Ni (xj , yj ) = 0, i 6= j These functions are the analogue of the ij that we used with the square grid. They have the 21

effect of smearing the potential at node i over the whole triangle. Then, for example, for N1 we have: a1 + b1 x1 + c1 y1 a1 + b1 x2 + c1 y2 a1 + b1 x3 + c1 y3 = 1 = 0 = 0

The determinant is:

This set of equations has a nontrivial solution for the coefcients a, b and c if the determinant: 1 x1 y1 D= 1 x2 y2 6= 0 1 x3 y3 D = x2 y3 x3 y2 + x3 y1 x1 y3 + x1 y2 x2 y1 = (x2 x1 ) (y3 y1 ) (x3 x1 ) (y2 y1 )

This is related to the area of the triangle. To see how, remember that we can write the area using the cross product: 1 A = 1 2 2 where 1 and 2 are vectors along the sides of the triangle. In terms of the coordinates:
1

and
2

= (x2 x1 ) x + (y2 y1 ) y = (x3 x1 ) x + (y3 y1 ) y

Then and so

2A = (x2 x1 ) (y3 y1 ) (x3 x1 ) (y2 y1 )

D = 2A and so is never zero. Then the solution for the coefcients is: 1 a1 = (x2 y3 x3 y2 ) 2A 1 b1 = (y2 y3 ) 2A 1 c1 = (x2 x3 ) 2A and similarly for the others. Now the procedure runs pretty much as before. In equation (16) we take: X j Nj =
j

where the sum is over all the vertices of all the triangles. Next take = Ni for one vertex of one triangle. Then relation (16) becomes: Z Z Z 3 X j gNi dxdy ' g (x, y ) Ni dxdy Ni Nj dxdy =
j =1 triangle

= g (x, y ) A (ai + bi x + ci y )

(19)

22

where x, y are the coordinates of the center of the triangle. Now using our solutions for the coefcients, we have: 1 (x1 + x2 + x3 ) (y1 + y2 + y3 ) a1 + b1 x + c1 y = (x2 y3 x3 y2 ) + (y2 y3 ) (x2 x3 ) 2A 3 3 D 1 1 = [x2 y3 x3 y2 + x1 y2 x1 y3 x2 y1 + x3 y1 ] = = 6A 6A 3 On the left hand side, we use the result that and is constant over the area of the triangle, so (19) becomes
3 X j =1

+ci y Ni = bi x

j (bi bj + ci cj ) A = g (x, y )

A 3

To combine the triangles, we dene kij = (bi bj + ci cj ) A (20) If i = j , kii refers to a single node. For each internal node i we sum over all the triangles connected to that node. X Kii = kii
triangles

The elements kij , i 6= j, are associated with two nodes, i.e. with the side of the triangle connecting the nodes. So we sum over all the triangles with a side along ij (usually two). X kij i<jN Kij =
triangles

If a node is on the boundary, the value of the potential there will be known. These terms are moved to the right hand side and serve as source terms. So we have: Gi =
N0 X 1 X At gt Kij j 3 triangles j =N +1

where the numbers N + 1 to N0 label the nodes on the boundary. Then combining the results for all the triangles, we have the matrix equation: K= G where again K is a sparse matrix. Example of how to get the kij . Consider an isoceles right triangle of sides 1,1 and 2. Put the origin at the right angle, and label the vertices 1 (the origin), 2 (on the y -axis) and 3 (on the x-axis). Then the area of the triangle is 1/2, and 2A = 1

23

. a1 = x2 y3 x3 y2 = 0 1 = 1 b1 = y2 y3 = 1 c1 = (x2 x3 ) = (1) = 1 b2 = y3 y1 = 0 c2 = (x3 x1 ) = 1

1 1 2 k11 = b1 + c2 (1 + 1) = 1 1 = 2 2 1 1 1 k12 = (b1 b2 + c1 c2 ) = (0 + 1 (1)) = 2 2 2 and so on. (See Jackson Fig 2.16). These coefcients depend on the shape of the triangle but not on its orientation or size.

Then from (20) we get

24