Journal of Modern Optics

ISSN: 0950-0340 (Print) 1362-3044 (Online) Journal homepage: www.tandfonline.com/journals/tmop20

Refraction and geometry in Maxwell's equations

A. J. Ward & J. B. Pendry

To cite this article: A. J. Ward & J. B. Pendry (1996) Refraction and geometry in Maxwell's
equations, Journal of Modern Optics, 43:4, 773-793, DOI: 10.1080/09500349608232782
To link to this article: https://doi.org/10.1080/09500349608232782

Published online: 03 Jul 2009.

Submit your article to this journal

Article views: 2781

View related articles

Citing articles: 26 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tmop20
JOURNAL OF MODERN OPTICS, 1996, VOL. 43, NO. 4, 773-793

Refraction and geometry in Maxwell’s equations

A. J. WARD and J. B. PENDRY

Condensed Matter Theory Group, T h e Blackett Laboratory,
Imperial College, London, SW7 2B2, UK

(Received 3 October 1995)

Abstract. Computational studies of Maxwell’s equations in complex geo-

metries encountered in photonic band structure calculations run into difficulties
when several length scales occur, such as the wavelength of light in free space
and the skin depth in metal. These problems are remedied by using an adaptive
co-ordinate system which expands or contracts length scales as necessary. Here
we show that moving to a general co-ordinate transformation is equivalent to
renormalizing E and p. This is an huge simplification because now we need only
write one computer code in a Cartesian system, and we can use this same code
to handle any co-ordinate system by adjusting the E and p we feed into the
calculation.

1. Introduction
Light propagation in complex structures has aroused considerable interest in
the past few years. This interest was largely stimulated by Yablonovitch and
co-workers [l) 23 who suggested that diffraction by periodic structures might lead
to a ‘photonic insulator’ in which light is forbidden to propagate in a certain range
of frequencies. T h e challenge to the theorist is to solve Maxwell’s equations in
these structures:

v x E = -ppo aH/at,
v x H = +CEO a q a t ,
where both E and p may depend on position. Various methodologies have been
developed to compute the dispersion relationships for these structures: expansion
of the wavefield in plane waves [3-51 or representation of the fields on a discrete
mesh of points [6-81. T h e latter method has particular advantages as it leads to
sparse matrices and lends itself to calculations at constant frequency using transfer
matrix methods. All these methods have been kept busy in the search for photonic
insulators, and in other tasks, but they do encounter severe problems when the
refractive index varies very strongly. For example, in a system comprising both
free space and metallic components there is difficulty in representing both the
relatively long wavelength variations in the vacuum which might be on the scale
of a few thousand Angstroms, and the field inside the metal which decays on the
scale of the skin depth, perhaps only a few tens of Angstroms. A plane wave
expansion would require many Fourier components with consequently heavy
computation overheads. Alternatively a uniform sampling grid would need to be
at least as finely spaced as the most rapid spatial fluctuation and hence this is
wasteful in the regions where the wavefield varies more smoothly. I n other
instances the system may have some intrinsic symmetry which is not compatible

0950-0340/96 512.00 0 1996 Taylor & Francis Ltd.

774 A . J . Ward and J . B . Pendry

Figui-e 1. Simple cubic lattice of points in one co-ordinate system (left) maps into a
distorted mesh in the other co-ordinate system (right). Choosing the co-ordinate
transformation correctly we can control the distortions and produce a mesh tailored
to our requirements.

with a uniform mesh of points. For example, a fibre with cylindrical symmetry
requires a cylindrical co-ordinate system before it can be exploited in a computa-
tion. These problems have long been recognized and the solution is to use a
non-uniform mesh on which to represent the wavefields: densely sampled where
the variation is rapid, less so elsewhere. However, whereas this may seem a simple
solution to the problem, construction of such meshes can be horribly complex,
and erecting a system of difference equations upon them is a programmer’s
nightmare. Here we present a simple observation that factorizes the problem into
two parts: having solved the problem of computing on a uniform Cartesian mesh
we can use exactly the same set of programs to compute for the non-uniform mesh.
Details of the mesh structure can all be subsumed into an effective E and p. In
practice this is a huge simplification in that we can divide our programming effort
into two distinct parts: the co-ordinate system and Maxwell’s equations. How does
this come about? Let us transform to a general system defined by

Q l b ,Y , 4, Q Zb, Y , z)
9 Q3(X, Y ?z). (2)
Lines of constant q2, q3 define the generalized q1 axis, and so on. T h u s if we define a
set of points by equal increments along the q l , q2, q3 axes, the mesh will appear
distorted in the original x, y, z co-ordinate frame: see figure 1.
Maxwell’s equations as defined above are written in the original Cartesian
system. What form does the new set of equations take when expressed in terms of
q l , q2, q3? T h e answer is surprisingly simple. Maxwell’s equations in the new
system of co-ordinates become
v, x € = -po/?aH*at,
(3)
v, x H = -E0i &*at,

where i, C; are in general tensors, and E, H are renormalized electric and magnetic
fields. All four quantities are simply related to the originals. In other words the
form of Maxwell’s equations is preserved by a co-ordinate transformation and we
have attained our stated objective of dividing the computation into two parts. T h e
co-ordinate transformation does not change the fact that we are still solving
Maxwell’s equations, it simply changes the definition of E , p.
 Refraction and geometry in Maxwell’s equations 77s

Not only does this transformation solve the problem of choosing an appropriate
discrete mesh on which to quantize the wavefield, it also helps methods in making
a Fourier expansion of the wavefield. Such expansions need not be confined to the
original co-ordinate system but may also be performed with respect to the new
co-ordinates. If these are chosen so that they make a more uniform sampling of
the wavefield then we can expect the Fourier expansion to be more convergent in
the new co-ordinate frame and therefore less demanding of computer time. All
this is reminiscent of a concept familiar in the calculation of electron wavefunctions:
the pseudopotential. Atomic potentials which are strongly singular near the origin
can be replaced by equivalent, or pseudo, potentials which have the same scattering
cross section, but are much more easily represented in a Fourier expansion. In the
following sections we derive the key results needed to implement these ideas in a
computation. This we do in two ways. Firstly we assume that the new mesh is
defined by an analytic co-ordinate transformation. This is much the easier
implementation because the co-ordinate transformation is defined in a compact
manner. In some circumstances analytic transformations may not be feasible and
in this case we give a more complex derivation in which we work directly with the
co-ordinates of the non-uniform mesh. Finally we give some examples of the
method at work.

2. Transformation to a general co-ordinate system

We define three units vectors, ul, u2, u3, to point along the generalized ql, q2, q3
axes. T h e length of a line element is given by

where

In particular we shall need the length of a line element directed along one of the
three axes
ds; = Q i dqi, (6)
where
Q? = 9.. (7)
for shorthand.
T o calculate V x E consider a small element, small enough that it resembles a
parallelepiped (figure 2). In this we assume that the transformation has no

Figure 2. Small element resembling a parallelepiped.

776 A . J . Ward and J . B . Pendry

ElQl4
Figure 3. Integration path for finding V x E.

singularities such as points or lines where the co-ordinate system suddenly heads
off in a different direction.
First we calculate the projection of V x E onto the normal to the uI-u2 plane
by taking a line integral around the u1-u2 parallelogram and applying Stokes'
theorem (figure 3). We define
El = E - u , ,
E2 = E-IJ,,
E3 = E - U , ,
so that

or
a i 2- -
(V x E)*(ul x u2)Q1Q2= - a i l = (Vnx t)3,
841 842

Note that the right-hand side of equation (10) is simple 'component 3' of curl
evaluated in the new co-ordinate system. Now applying Maxwell,
aH
(V x E).(uI x u d Q i Q , = - P O P ~ ' ( Ux~ ~ 2 1 Q i Q 2 .
We write H in terms of the contravariant components,
H = H'u, + H2u2+ H3u3,
which in turn can be expressed in terms of the covariant components,

where the first part defines g, and

H' = H - u l , Hz = H * u 2 , H3 = H *u3.
Inverting g gives
 Refraction and geometry in Maxwell's equations 777

Substituting equation (13) into (12) gives

aH
(V x E).(ui x ~ 2 1 Q i Q 2 = -POPT.(UI x ~2)QiQ2

Define

and

so that

(V x E).(ui x ~2)QiQz

Hence on substituting from equation (lo),

and by symmetry between E and H fields,

..a,@.
(V, x H)i= -&o c
3

j=l
& # I d ,
at

Note that these expressions simplify considerably if the new co-ordinate system
is orthogonal, e.g. cylindrical or spherical. Therefore equations (21) and (22) show
that we can forget the co-ordinate transformation as far as computations are
concerned, provided that we employ 2,6 everywhere in place of E , p.

3. A discrete general co-ordinate system

We shall now treat a general network of discrete points which need not have
an analytic definition. However, as we shall see, we still require that the mesh is
smooth and contains no sharp corners. Also we shall not assume that the E- and
H-fields are represented on the same mesh of points. These points need not be
regularly spaced, nor the bonds between the points orthogonal to one another, but
there are some restrictions we must impose. Firstly, it must be possible to divide
up the network into a series of layers, each point having four bonds linking it to
nearest neighbours within the layer and single bonds linking it to the layers above
and below (see figure 4). Secondly, each set of four bonds which form a closed
loop must enclose a flat (or at least approximately flat) surface. I n practice this
should not constitute too serious a restriction as long as we make our lattice points
close enough. At each point we associate three contravariant vectors which point
along the three bonds leading away from that point. (Each point has in total six
bonds touching it, three leading into the point and three leading away.) For the
E-field let these vectors be Qf(r)el(r), Qf(r)e,(r)and Qf(r)e)3(r). T h e em'sare
the local unit vectors along each bond and the Q f s are the distances to the next
mesh point.
Next we choose a set of mesh points on which to represent the magnetic field
778 A . J . Ward and J . B . Pendry

Figure 4. Schematic of a generalized discrete mesh. Each vertex has six bonds, four of
which join it to other mesh points within the layer, two of which join it to the layers
above and below. Note only a few of the inter-layer bonds have been shown.

Figure 5. Relative positions of the E-field and H-field meshes, chosen such that they
interlock, with H-points lying at the centre of the E-field cells.

H(r). These points are taken to lie at the centre of each E-mesh cell (see figure
5). For the H-field we must also define three vectors: Qy(r)hl(r), Qy(r)h,(r) and
Qy(r)h3(r). Notice that in most general cases, t h e vectors which join E-field
points are different from those which join H-field points.
We shall also need to define vectors reciprocal to the e’s,
e2W x e3m
el(r)= (24)
e1W0e2(r) x e3w’
etc. T h u s e;e8 = ~ 5 We ~ . do the same for ha. W e shall express the fields in
~ also
terms of the covariant components:
E(r) = El(r)el(r) + E2(r)e2(r) + E3(r)e3(r),
(25)
H(r) = Hl(r)hl(r) + H2(r)h2(r) + H3(r)h3(r).
We are now in a position to express Maxwell’s equations in the new co-ordinate
system. We start from the curl equations

V x E = --,
as
at
aD
VxH=-,
at
 Refraction and geometry in Maxwell's equations 77 9

r+a+t

Figure 6. Line integral for E is taken around the surface penetrated by L--e ird
component of B.

or in the integral form.

Starting with equation (29) we take the E-field line integral around the loop in
figure 6 and note that this surface is penetrated by the H-bond along h3.

E(r)-Qyel(r) + E(r + a)-Qfe,(r + a) - E(r + b).Qfel(r + b) - E(r)-Qfe,(r)

a E E
= -- B(r).(ei x e,>Qi Q 2 9 (31)
at

Qf(r)E,(r) + QfP + a)E,(r + a) - Qf (r + b)El(r + b) - Q f ( r ) W )

a
= -- B(r).e3(r)QfQflel.e2 x 8 3 1 . (32)
at
Write ,@" = Qf(r)EJr) and f i a = QF(r)Hu(r):

&(r) + ,@,(r + a ) - i , ( r + b) -I?,(') a

= --p,,p(r)H(r)-e3(r)Q1
E E
Q,le,.e, x e31.
at
(33)
We define the quantities g""D = ea.hP so we can write

H(r)-e3(r) = g3"(r)H,(r)
780 A . J . Ward and J . B . Pendry

Now define

which gives us

El(r) + E2(r + a) - E,(r + b) - k2(r) = --ata ~ ~ @ ~ ~ ( r ) f i , ( r ) .(36)

Repeating this for the other two surfaces penetrated by components of the H-field
gives us a set of three equations:

E2(r) + E3(r + b) - E2(r + c) - E3(r) = --ata po$'"(r)fia(r),

A

E3(r)
a
+ E1(r + c) - E3(r + a) - E,(r) = --po@2a(r)fia(r), (37)
at

&(r) + E2(r + a) - E,(r + b) - E2(r) = --ata p0k3,(r)fia(r).

By considering the three surfaces penetrated by E-field bonds and using equation
(30) we generate three more equations:

fi2(r - c)
* a Eo.31a(r)i!?a(r),
+ fi3(r) - fi,(r) - H(r - b) = -
at

a
- b) +
A

- H(r - a) = -eOPa(r)Ea(r),
- I?,@) at

where we have defined

These six equations are, of course, nothing more than the discrete form of
Maxwell's equations in the new co-ordinate system (compare with equations (21)
and (22)). In the special case where the E-field and H-field lattices are the same
then Qf = QY = Q , , e, = ha = U, and pp = gap. Then the expressions for 2 and
p revert to those in the previous section.

4. The transfer matrix in a general co-ordinate system

Having found expressions for Maxwell's equations in a general discrete
co-ordinate system, we can go on and try to find the transfer matrix in this scheme.
Others have already made progress in extending the Transfer Matrix Method
beyond the simple cubic mesh to a non-orthogonal mesh [9] but this goes one step
further and uses the general mesh. First we write out our six equations again,
 Refraction and geometry in Maxwell's equations 781

assuming a n explicit t i m e dependence exp [ - iwt]:

+ + b) - E2(r + c) - E3(r)= iopofila(r)fiK(r),
i 2 ( r ) k3(r (40)
E3(r)+ El(r + c) - E3(r + a) - El(r) = iwpofi2a(r)fia(r), (41)
El(r) + k2(r + a) - &(r + b) - k2(r) = iop,fi3"(r)da(r), (42)
fi2(r - c) + fi3(r) - fi2(r) - fi3(r - b) = -iweoilK(r)l?a(r), (43)
~ , ( r- a) + Bl(r) - fi3(r) - fil(r - c) = -iweoi2a(r)Ea(r),
A

(44)
H,(r - b) + H2(r)- fi,(r) - H2(r - a) = -iwe0i3"(r)Ea(r).
A A A

(45)
Then we use equations (42) and (45)to eliminate the third components of the
fields and rearrange:

1
x [fil(r - b) - fi,(r) + fi2(r)- fi2(r - a)] - i~e,i?~~(r+ a)
x [g,(r + a - b) - G,(r + a) + fi2(r + a) - fi2(r)l

+
+ E2(r + b, [ -i3'(r b)
] P'(r) @l3(r)
$33(r + b) + E 1 ( r ) [ m - m]
+ b)
+ El(r +
-.t31(r
b)[ i3'(r + b)
+

-1
fi13(r)
$33(r)

+ i ~ &1, i ~ ~[fi,(r
(r)
- b) - g,(r) + fi2(r) - fi2(r - a)]
1
-
i w ~ , i ~ ~+(b)
r
[fil(r) -gl(r +b) +g2(r + b) -G2(r -a +b)]
782 A. J . Ward and J . B . Pendry

+ G2(r- a + c)
1
c -p32(r- a c)
+
fi33(r- a c)
+ - ZZ3(r+ c)
t3'(r + c) 1
-
- 1

H1(r)+i~pop33(r
+c)
[&(r +c) -El@ + b +c ) + E2(r+ a +c) -E20 +c)]

1
-
i~p~p-a~+c)
~(r
[&(r -a +c) -&(r -a +b +c) +E2(r+ c) -E2(r -a +c)]

- iueO E,(r + c) I P(r

+ c)t31(r + c)
+ c) - B3(rB3(r + c) 1
+
t23(r ~ ) t ~ ~ +
c) ( r
t33(r c) +

+ + t13(r + c) + G2(r- b + c) -p3'(r - b + c)

fi2(r +c)[l+ @32(rc)
/i33(r c)+ -233(r+ c) 1 @33(r -b +c) 1
-t13(r + c) +
p37r c) - tI3(r + c)
+fi,(r-a+c)[ +
t 3 3 ( r + c ) ] + ~ 1 ( r + C ) [ $33(r c) t33(r+ c) 1
-p31(r - b c)+ t13(r+ c)
+ Gl(r - b + c) $33(r- b c)+
+

t33(r+ c) 1
1
- *
- H2(r)+ iWp0fi33(r+c)
[,@,(r+c)-%l(r+b+c)+E2(r +a + c)-E2(r+c)]

- 1
[&(r -b +c) -&(r +c) +E2(r+a -b +c)-g2(r - b +c)]
i ~ p ~ f-
i b~+c)
~(r

+ c)B1(r + c)
t13(r
t33(r+ c) 1
t13(r + c)B2(r + c)
(49)
233(r+ c)
 Refraction and geometry in Maxwell's equations 783

Finally we introduce the reduced magnetic field H' = (i/cix,,)H and the speed of
light ci = l / ~ ~ p ~ :

-t32(r + a)
+ i 2 ( r + a)[ +
t?33(r a)
+

-1
$23(r)
$33(r)

1
-- [fi;(r - b) - @(r)
E*33(r)
+ &(r) - @(r - a)]
784 A . J . Ward and J . B . Pendry

+ fi;(r + c)[ -B3(r + c) ]+ + c)[ $32(r + c) - B3(r+ c)

-b A33

E (r + c )
@(r
$33(r+ c) p3(r+ c) 1
+ + tZ3(r+ c)
+ fi;(r -a + c) -ji3% - a c)
p33(r- a + c) t33(r c)+ 1
= A;(r)
4 [&(r +c) -&(r + b + c) +E2(r+a +c) -i 2 ( r+c)]
+ 02433(r
+c)

- 4 +
[kl(r - a c) -El(r -a + b +c) +&(r +c) -i 2 ( r-a +c)]
02ji33(r-a +c)

+ c)Z3'(r + c)
+ El(r + c) iZ2l(r + c) - P3(r233(r+ c) I
+ c ) E * ~ ~ +( ~ c)
+ E2(r+ c) i;22(r + c) - P3(rE*33(r+ c)

p3'(r + c) 213(r + c)] +

+ c)[l + p33(r + c) + E*33(r + c) + &(r - b + c)
-,d32(r - b c)
p33(r- b c)+ 1
+ fi;(r -a + c) [
-E*"(r
A33
E (r
+c)
+c)
] + s ; ( r + [p3'(r ++
c)
$33(r c) .t33(r c)
+
+ 1
c) - t13(r c)

-p31(r - b + c)
+ &(r -b + c) li33(r- b c ) + + E*33(r+ c)
-
- Hi(r)+02p33(r
- " +c) PI(+ r c) - i,( + + +
r b c) &(r +a +c ) -E2(r+c 11

- 4 +
[&(r - b c) -il(r +c) +i2(r +a - b +c) -&(r - b + c)]
- b +c)
co2ji33(r

+ c)
P3(r+ c)P'(r
+
,f33(r c) I
+
E*I3(r c)B2(r + c) (53)
t33(r c) +
 Refraction and geometry in Maxwell’s equations 785

These equations have a more complex form than the conventional transfer
equations due to the non-zero off-diagonal elements of 2 and p. I n the equations
(50) and (51) for the E-field these extra terms are on the right-hand side and cause
no problems as they only introduce extra terms within the layer rather than linking
to the next layer. In other words, equations (50) and (51) are still transfer equations
+ +
as they allow us to find E,(r C) and E,(r C) only in terms of the fields in the
previous layer.
For the H-field (equations (52) and (53)) the situation is more difficult as extra
+ +
terms are introduced on the left-hand side. H,(r C) and H,(r C) can no longer
be expressed simply in terms of the fields in the previous layer. However, we can
proceed if we recognize equations (52) and (53) as a series of 2n coupled equations,
two for each point within the layer. Then, by inverting the 2n x 2n matrix formed
from the coefficients of the terms on the left-hand side, we can find H,(r C) and+
H2(r+ C ) in terms of the fields in the previous layer.
As a useful check we can take the special case that the discretization lattice is
simple cubic. In this limit, the equations we have derived here can be seen to revert
to the original simple cubic transfer equations [7]. With the transfer equations
complete we can continue just as we would in the simple cubic case and use them to
formally define a transfer matrix which relates the fields at one layer in the photonic
system to the fields at another. From this matrix we can then find the band
structure, transmission and reflection coefficients, etc., in the usual way.

5. Applications and examples

We will now allow the techniques developed to a few simple systems to test
and illustrate our ideas.

5.1. Triangular lattice

For a first test we take the case of a 2D triangular lattice of air holes
in a dielectric background. We choose this system because it has already been
extensively studied and there are some experimental results we can use for
comparison [lo].
T o perform the calculation first we need to choose a mesh. We pick vectors to
match the symmetry of the triangular lattice.
el = hi = i,
8, = h2 = j ,

4 3 . +$k,
e3= h3 =-I
2
where i, j and k are the Cartesian unit vectors. Each unit cell contains one cylinder
and is divided up into evently spaced mesh points (see figure 7). Bloch’s theorem
is used to set the boundary conditions at the edge of the cell. At each mesh point
we set the value for E^ and p using equations (35) and (39) and substitute the relevant
lengths and the values of E and p averaged over each mesh point. T h e cylindrical
holes are air ( E = 1.0) and the host dielectric PbO glass ( E = 2.6). T h e lattice spacing
is 1.17 pm and the diameter 0.9 pm. Figure 8 shows the first Brillouin zone for a
triangular lattice and marks the symmetry points. We have calculated band
structure along the T-X direction and it is given in figures 9 and 10.
786 A . J. Ward and J . B . Pendv

Figure 7. Sketch of primitive unit cell for a triangular lattice of cylinders. The unit cell
is divided up into evenly spaced mesh points, the value of the dielectric constant
assigned to each point being the average of the mesh points.

0
r

Figure 8. First Brillouin zone for a triangular lattice. The high symmetry points are
labelled.

5.2. Cylindrical waveguide

Let us try another simple calculation and try to find the modes of a perfect
metal cylindrical waveguide. We can then check our results as the solutions are
already well known [ll]. We will use cylindrical polar co-ordinates as the basis
for our mesh.
 Refraction and geometry in Maxwell's equations 787

Triangular array of Cylindrical holes in PbO glass

S polarization.

1.2

1.o

-.-
v)
c
c
3
I0.8
-
al
c
.-0
v)
c
E?
5 0.6
Y

E
c
JU
$ 0.4

0.2

0.0
-3.0 -2.0 -1.0 0.0 1.0 2.0 3.0
Wave Vector

Figure 9. Band structure for a triangular lattice of cylindrical holes in PbO glass;
S-polarization.

el = h, = P,
A

e, = h, = 8,
e 3 =h 3 = f .

These unit vectors are orthogonal so 1°F is a 3 x 3 unit matrix. This simplifies our
transfer equations considerably:

+ E*33(r1+ a) [@(r + a - b) - fi;(r + a) + &(r + a) - @(r)]

788 A . J . Ward and J . B . Pendry

Triangular array of Cylindrical air holes in PbO glass

P polarization.

1.2

1.o

.-dC
h

a
3 0.8
-
C
.-0
u)
C

gE 0.6
iC?
al
3
E 0.4
IL

0.2

0.0
-3.0 -2.0 -1.0 0.0 1.0 2.0 3.0
Wave Vector
Figure 10. Band structure for a triangular lattice of cylindrical holes in PbO glass
P-polarization.

1
-
&(r + c) = &r) - -[&(r - b) - &(r) + &(r) - k 2 ( r - a)]
- w r ) - -

1
+ -,,, . . [fi;(r) - f i { ( r + b) + f i i ( r + b) - &(r -a + b)]

x [&(r + c) - &(r + b + c) + &(r + a + c) - &(r + c)]

- ct
~ ~ f i-~a ~ c)
( r +
x [&(r - a + c) - E,(r - a + b + c) + &(r + c) - &(r -a + c)]
+ &r + c)P2(r + c), (56)
 Refraction and geometry in Maxwell's equations 789

Figure 11. Part of the mesh for a cylindrical waveguide; a, is the distance from the centre
to the first point, a the radial spacing, b the angular spacing and c the axial spacing (not
shown).

x [E,(r + c) - &(r + b + c) + g2(r + a + c) - &r + c)]

- c;
02fi33(r- b + c)
x [&(r - b + c) - i l ( r + c) + &(r +a -b + c) + i 2 ( r - b + c)]
- El(r + c).?*(r + c). (57)
We choose uniform intervals for the radial co-ordinate and assume the system
is cylindrically symmetric. Let a be the radial spacing, b be the angular spacing
and c the spacing along the axial direction (see figure 11). Therefore the lengths
we need are
Q f ( i r , iB, i z ) = Q y ( i r , iB, i z ) = a ,
+ ao)b,
Q f ( i r , iB, i z ) = ( ( i r - 1 ) a

( i i - 1 ) a + - + a.
U

2
Q f ( i r , iB, i z ) = Q y ( i r , iB, i z ) = c ,
where ir, iB and i z are the integers which label the mesh points. Finally, we need
to set the boundary conditions at the edge and centre of the guide. T h e outer surface
is terminated by a perfect metal so here we simply set all the fields to zero. T h e
centre is more difficult because the boundary conditions change depending on the
angular momentum of the mode. In all cases references to non-existent lattice
points are set to zero (i.e. points with the radial co-ordinate Y < 0). Ar Y = 0 the
radial and angular components of the fields are zero by symmetry. If the angular
momentum is non-zero then the axial component of the fields is also zero. This is
not the case if the angular momentum is zero.
790 A. J . Ward and J . B . Pendry

Modes of a cylindrical wave guide

m=O
0.020

0.015

%
v

0
%
5 0.010
3
0-
e!
L

0.005
+ Non-orthogonal mesh
x Orthogonal mesh
0 Analytical Result

0.000
-0 I10 -0.0005 0.0000 0.0005 0.0010
Wave Vector

Figure 12. Modes of a cylindrical waveguide. Angular momentum m = 0. Waveguide

radius = lOW4 m. The first curve is the analytical result, the second is the numerical
calculation using an orthogonal mesh, and the third is the calculation using the
distorted, non-orthogonal mesh.

Putting all this together we can calculate the modes. T h e radius of the
waveguide we have considered is m. T h e first set of curves in figures 12-14
give the analytic solutions for this system for three different values of the angular
momentum, while the results from our calculations are shown in the second set of
curves.

5.3. Distorted mesh for cylindrical waveguide

As a final calculation we shall take our cylindrical waveguide and distort
the mesh to remove the orthogonality from the unit vectors. P u t simply, we will
take hold of the centre of the cylindrical mesh and pull it along the axial
direction while keeping the outer surface fixed. T h i s is shown pictorially in
figure 15.
 Refraction and geometry in Maxwell's equations 791

Modes of a cylindrical wave guide

m=l
0.020

0.015

s
v

0
%
5 0.010
a
0-
2
LL

0.005

x Orthogond mesh
0 Analytical Result

+, + + +
0.000
-0.0010 -0.0005 0.0000 0.0005 0.0010
Wave Vector

Figure 13. Modes of a cylindrical waveguide; angular moment m = 1, waveguide radius =

m. The first curve is the analytical result, the second is the numerical calculation
using an orthogonal mesh, and the third is the calculation using the distorted,
non-orthogonal mesh.

Our unit vectors now become

el = h, = cos (4)i sin (4);,
A

02 = h2 = 8,

8 , = h, = z^.

If 4 = 0 then we revert to our previous calculation. For 4 = 30" the modes we

calculate are given in the third set of curves in figures 12-14. I t is clear that our
method has no difficulty in handling the distortion.

6. Conclusions
We have presented a formulation for expressing Maxwell's equations in a
general co-ordinate system. T h e form of the equations is preserved while all the
details of the co-ordinate transformation are absorbed into an effective 2 and $.
This enables us to extend the conventional transfer matrix method t o an arbitrary
792 A. J . Ward and J . B . Pendry

Modes of a cylindrical wave guide

m=2
0.020

0.015

z
Y

5i; 0.010
a
F
LA

0.005
t Non-orthogonal mesh
x Orthogonal mesh
0 Analytical Result

0.000
Wave Vector

Figure 14. Modes of a cylindrical waveguide. Angular momentum m = 2. Waveguide

radius = m. This first curve is the analytical result, the second is the numerical
calculation using an orthogonal mesh, and the third is the calculation using the
distorted, non-orthogonal mesh.

meshing scheme where the vectors joining mesh points need no longer be regularly
spaced or orthogonal. While the resulting transfer equations are more complex
than the standard ones, they can still be solved in a straightforward way on a
computer to give all the usual quantities: band structures, transmission and
reflection coefficients, etc.
I n order to test our new techniques, we have presented three simple
calculations:
(1) a triangular lattice of cylinders to test a non-orthogonal, uniform mesh,
(2) a cylindrical waveguide which uses orthogonal cylindrical polar co-
ordinates,
(3) a distorted cylindrical mesh which combines both ideas.
T h e results agree well with expectations giving confidence in the method. While
we may suspect that non-cubic mesh schemes will be more prone to numerical
difficulties than cubic ones we have yet to encounter such problems.
 Refraction and geometry in Maxwell’s equations 7 93

Figure 15. Diagram showing distortion of cylindrical mesh. The angle 4 parameterizes
the degree of distortion. For 4 = 0 we revert to the undistorted mesh.

It is hoped that these techniques will widen the range of applicatons for the
Transfer Matrix Method to include non-orthogonal lattices, fibres and tapers and
maybe systems with no underlying lattice symmetry.

Acknowledgments
A.J.W. is supported by an EPSRC CASE award in collaboration with GEC-
Marconi Materials Technology Limited. Many thanks are given to John Roberts
for numerous useful discussions.

References
T. J., 1989,Phys. Rev. Lett., 63, 1950.
E., and GMITTER,
c11 YABLONOVITCH,
E., GMITTER,
c21 YABLONOVITCH, T. J., and LEUNG,K. M., 1991, Phys. Rev. Lett., 67,
2295.
c31 LEUNG, K. M., and LIU,Y. F., 1990,Phys. Rev. Lett., 65, 2646.
c41 ZHANG, Z.,and SATPATHY, S., 1990,Phys. Rev. Lett., 65, 2650.
c51 Ho, K. M., CHAN,C. T., and SOUKOULIS, C. M., 1990,Phys. Rev. Lett., 65, 3152.
C6l PENDRY, J. B., and MACKINNON, A., 1992,Phys. Rev. Lett., 69, 2772.
c71 PENDRY, J. B., 1994,J . mod. Optics, 41, 209.
181 BELL,P. M., PENDRY, J. B., MART~N-MORENO, L., and WARD,A. J., 1995, Computer
Phys. Commun.,85, 306.
c91 ROBERTS, P. J., BIRKS, T. A., RUSSELL, P. ST J . , SHEPHERD,T. J., and ATKIN,D. M.,
1995,Optics Lett. (accepted).
P O I INOUE, K. WADA,M., SAKODA, K., YAMANAKA, A., HAYASHI, M., and HAUS,J., 1994,
Japan J . appl. Phys., Pt 2, 33, L1463.
c111 KAO,C. K., Optical Fibre (London: Peter Peregrinus on behalf of IEE), p. 18.