Inverse Problems 14 (1998) 113137.

Printed in the UK PII: S0266-5611(98)85902-9

Direct and inverse scattering for transient electromagnetic
waves in nonlinear media
Gerhard Kristensson and David J N Wall
Deptartment of Electromagnetic Theory, Lund Institute of Technology, PO Box 118,
S-221 00 Lund, Sweden
Department of Mathematics and Statistics, University of Canterbury, Christchurch,
New Zealand
Received 11 July 1997
Abstract. Nonlinear propagation of electromagnetic waves is an important problem in optics.
Often the properties of the nonlinear media are not fully understood. The solution of an inverse
problem can provide an aid to that understanding.
An inverse transmission problem is posed; it is one of reconstructing the medium parameters,
by measurement of a wave that has been propagated through the nonlinear medium. The
nonlinear medium is assumed to be homogeneous and isotropic. The methods have application
to nonlinear optics, and the numerical results for both the direct and inverse problems presented
are based on the nonlinear Kerr effect, which is observed in the optical wavelength band.
However, the mathematical techniques that are developed are applicable to any set of nonlinear
rst-order equations. The method is therefore model independent.
1. Introduction
Inverse problems for a single, scalar, one-way nonlinear wave equation have recently been
considered in [3], and we will henceforth refer to this paper as (I).
One of the major results, given here in this paper, is to show that a system of two
rst-order partial differential equations can be effectively analysed via a single one-way
wave equation. This is because the wave splitting used to do this only allows coupling of
the two one-way wave equations through the wave speed functional. This is a signicant
result as it enables exact solutions to be formulated for both the direct and the inverse
problems. The methods utilized in this paper are applicable to other sets of quasi-linear
rst-order equations, and so may be applied to problems other than those of the nonlinear
optical problems that are considered here.
In section 2, we state the basic equations governing one-dimensional electromagnetic
wave transmission through a typical nonlinear optical medium as used in bre optics. The
effects of dissipation, in a passive nonlinear optical media, impose constraints on the form
that the constitutive relations can take. This is utilized to examine the allowable form of the
constitutive equations in section 2.1. These constitutive equations form the basic nonlinear
models that are utilized later in the paper.
It is shown, in section 3, that by wave splitting the Maxwell equations describing one-
dimensional wave propagation through an isotropic, dissipative nonlinear optical medium
can be reduced to two one-way wave (transport) equations. It is further shown that, for
the problem of relevance to this paper, consideration of just one one-way wave equation
0266-5611/98/010113+25$19.50 c 1998 IOP Publishing Ltd 113
114 G Kristensson and D J N Wall
is necessary. This is due to the wave splitting used in this paper, and it means that only
one one-way wave equation is needed in the subsequent analysis. This is true provided
that the initial reected wave is identically zero in space, and the solution to the nonlinear
system is unique. Therefore, the details of existence and uniqueness of solution, for the
original system of rst-order equations and the wave split equations, is essential for our
development; these results are stated in section 3.1, and proven in appendices A and B.
The various nonlinear models for optical media that are used in the results shown
in section 5 are presented in section 4. Our techniques in section 5 build on those
discussed in (I). Further problems involving source reconstruction for a linear one-way wave
equation are considered in [10]. In section 5, we pose an inverse transmission problem of
reconstructing the medium parameters from measurement of a propagated wave through the
nonlinear medium. The method of characteristics is then employed to solve both the direct
transmission problem and the inverse transmission problem, i.e. the problem of nding the
functional form of the wave speed function after having measured the transmitted eld.
One of the novel features of our inverse problems is that the inverse solution can be
written down in explicit form. Another novel feature is that the solution also depends
continuously on the given transmission data, and this is proven in section 5.
Our methods have application to nonlinear optics, and the numerical results for both the
direct and inverse problems presented in section 5.1 are based on the nonlinear Kerr effect
which is observed in the optical wavelength band. The analysis and results presented in this
paper are valid up to, but not beyond, any wavefunction shock or discontinuity that may
occur because of utilization of the Kerr model. When a more accurate model, such as the
Raman model, is utilized then dispersion alters the shock behaviour; the various available
models are discussed further in section 2.
2. Basic equations
We rst review here some of the nonlinear effects as observed in bre optics. When
an electromagnetic wave propagates through a dielectric medium then polarization of
the medium molecules occurs. This polarization is generally represented through the
polarization eld, P, which is related to the electric ux density, D, and the electric
eld intensity, E, as
D =
0
E +P
where
0
is the electric permittivity of free space. In general, evaluation of P will require
a quantum-mechanical approach. Although such an approach is often necessary, when the
optical frequencies are near to the resonant frequency of the medium, a classical approach
can be used far from medium resonance. This will be the case for optical bres in the
wavelength range 0.52 m, and this is the range that is of interest in the study of nonlinear
effects in optical bres. The polarization vector can be split into a linear and a nonlinear
part as
P = P
l
+P
nl
with the linear part being represented by a susceptibility kernel,
(1)
, through
P
l
i
(x, t ) =
0
_
t

(1)
ij
(t t

)E
j
(x, t

) dt

and the nonlinear part as (see, e.g., [2, 8, 9])

P
nl
i
(x, t ) =
0
_
t

_
t

_
t

(3)
ijkl
(t t
1
, t t
2
, t t
3
)E
j
(x, t
1
)E
k
(x, t
2
)
E
l
(x, t
3
) dt
3
dt
2
dt
1
.
Transient electromagnetic waves in nolinear media 115
In these equations, the elds are assumed to be quiescent before a xed time, and are
dependent on the spatial coordinates, x = (x, y, z) R
3
, and the temporal variable,
t R, where Cartesian tensor notation has been used to represent the tensor elds, so
that i, j, k, l {1, 2, 3}. Note that causality causes the upper limit of the integrals, in
the previous equations, to be truncated. Also observe that, in general, polarization effects
produce memory effects. The lowest-order nonlinear response, which is shown here, allows
for the molecular vibrations, which is attributed to the Raman effect. In general, both the
electrons and the nuclei respond to the optical eld, but the nuclei respond slower than the
electrons. Often the nonlinear effect is modelled by assuming an instantaneous response
as
P
nl
i
(x, t ) =
0

(3)
ijkl
E
j
(x, t )E
k
(x, t )E
l
(x, t ) (2.1)
and this amounts to neglecting the Raman effect. For silica bres the vibrational
or Raman response occurs over a time scale of 0.6 to 0.7 ps; this will mean
that equation (2.1) will be approximately valid for pulse widths greater than 1 ps.
This is the model of the Kerr effect utilized in this paper in sections 4.1, 4.2 and
5.1.
Non-resonant intensity dependent effects (incoherent) can be included by use of a
nonlinear susceptibility kernel of the form

(3)
ijkl
(t t
1
, t t
2
, t t
3
) =
(3)
ikl
R
j
(t t
1
)(t t
2
)(t
1
t
3
)
with
_

0
R
j
(t ) dt = 1
and then
P
nl
i
(x, t ) =
0

(3)
ikl
E
k
(x, t )
_
t

R
j
(t t

)E
j
(x, t

)E
l
(x, t

) dt

.
The response function R
j
should include electronic and vibrational (Raman) contributions,
and if the electronic effect is assumed to be instantaneous then (for notational convenience
we omit the index j)
R(t ) = (1 f
r
)(t ) +f
R
h
R
(t )
with
h
R
=

2
1
+
2
2

2
2
exp(t /
2
) sin(t /
1
)
where the R function parameters, f
r
, f
R
,
1
and
2
, are chosen so that the main memory
effect is for t < 0.7 ps (for silica bres). The form of this model, h
R
, is the well known
Lorentz or resonance model for the Raman gain spectrum. Extensions of our methods in
section 5 for this model will be considered in a later paper.
The Maxwell equations, in a source-free region, linking the electric eld, E, magnetic
eld intensity, H, with the magnetic and electric ux densities, B, and D, respectively,
are
E(x, t ) =

t
B(x, t )
H(x, t ) =

t
D(x, t ).
116 G Kristensson and D J N Wall
Furthermore, assume that all elds depend only on the coordinate z. The Maxwell equations
then become
J

z
E(z, t ) =

t
B(z, t )
J

z
H(z, t ) =

t
D(z, t )
where J = z I (a rotation /2 around the axis z), and I is the identity operator. Assuming
only transverse components of the electric eld and non-magnetic materials, rewrite these
equations as

z
E(z, t ) =
1
c
0

t
J H(z, t )

z
J H(z, t ) =
1
c
0

t
D(z, t ) (2.2)
where the parameter
0
=

0
/
0
is the impedance of free space, and c
0
= 1/

0
is
the wave speed of electromagnetic radiation in free space.
The Maxwell equations imply the Poynting theorem
S +E
t
D+H
t
B = 0
where the Poynting vector S is dened as
S = E H.
Let the constitutive relations specifying the relationship between the eld intensities and
the eld ux densities be written as
D(x, t ) =
0
F(E(x, ))(t ) B(x, t ) =
0
H(x, t ) (2.3)
and let them be assumed to hold at each point in space. We assume for optical media that
the polarization effect is purely due to the electric eld and can be represented through an
operator F. Then, the rst of these equations is a generalization of those considered earlier.
Dissipation of the medium implies, through integration of the Poynting theorem over space
and time that the energy density, E, through this region has the form [4]
E(t )
_
t

{E(t

)
t
D(t

) +H(t

)
t
B(t

)} dt

0 for all elds. (2.4)

Notice that this dissipation constraint includes only the electromagnetic part of dissipation.
If conversion of electromagnetic energy into mechanical energy, or other energy forms,
takes place in the medium, the dissipation constraints will look different.
In the remainder of this paper we assume that this medium is isotropic. This implies
that D and E have the same eld orientation, therefore scalar quantities sufce. It follows
from (2.4) that the appropriate functional inequality to study is
_
t

E(t

)
d
dt
F(E)(t

) dt

0 for all E X (2.5)

where the operator F : X X, and where the afne function space X is dened by
X = {f C
1
(R) : f (t ) = 0, t < t
0
, t
0
xed}
Note that J J = I
2
, where I
2
is the identity operator in the xy-plane.
As then, D
i
=
0
F(E
i
).
Transient electromagnetic waves in nolinear media 117
for t
0
bounded. We denote by C
n
() the space of n times continuously differentiable
functions on . It is further assumed that a Banach algebra is dened on X, so then
integration by parts of (2.5) gives
E(t )F(E)(t )
_
t

(t

)F(E)(t

) dt

0
where the prime, on the dependent variable, has been used to denote differentiation with
respect to its independent variable, t

. We can convert this expression into a more convenient

operator inequality by dening an operator G : X X as
G(E)(t ) =
_
t

(t

)F(E)(t

) dt

. (2.6)
Then
d
dt
G(E)(t ) = E

(t )F(E)(t )
and the energy equality (2.5) can then be rewritten as
E
E

d
dt
G(E) G(E) 0 for all E X. (2.7)
We observe that if the electric eld is independent of t , then the inequality reduces to a
vacuous equality.
2.1. Constitutive relations of the function typethe nite-dimensional case
In general, the operator F acts on the function E. However, this case will not be considered
further in this paper, we only consider the nite-dimensional case here.
Let F be a real, continuous function dened on R, i.e. F C(R). In this case
F(E)(t ) = F(E(t )) is just a functional composition. Therefore, replacing the notation F
by F and G by G, the basic inequality, (2.7), becomes
E(t )G

(E(t )) G(E(t )) 0 for all E X

where the prime on G denotes differentiation with respect to its argument, and where
G C
1
(R) is a primitive function to F, i.e. G

(E) = F(E), such that G(0) = 0.

Therefore, in this nite-dimensional case, the relevant question to ask to ensure that the
media is dissipative is:
Question. For what functions G(x) is the following inequality satised:
x
2
d
dx
_
G(x)
x
_
= xG

(x) G(x) 0 x R
This inequality is equivalent to
d
dx
_
G(x)
x
_
0 x R. (2.8)
Proposition 2.1. If F is an increasing function, with F C(R), for all x R, then F
denes a constitutive relation that satises the energy inequality (2.8).
Note that F C(R) is sufcient for proposition 2.1, but stronger conditions will be required in the remainder
of the paper, cf denition 2.1.
118 G Kristensson and D J N Wall
Proof. Since F(x) is a increasing function, we have F(x) F(y), whenever x y.
From this we easily get
_
x
0
F(t ) dt
_
x
0
F(x) dt = xF(x) x 0

_
x
0
F(t ) dt
_
0
x
F(x) dt = xF(x) x 0
or
_
x
0
F(t ) dt xF(x) for all x R.
Moreover, this inequality is equivalent to (2.8); in fact, with G(x) =
_
x
0
F(t ) dt we have
xG

(x) G(x).
The converse to proposition 2.1, that every function that satises the energy
inequality (2.8) is an increasing function, is not true. A simple counterexample is
F(x) = 0.8x +sin x.
It is readily seen that if F C
1
(R), the conditions of proposition 2.1 are satised if
F

> 0, i.e. F is a strictly increasing function. In the following, we need a slightly stronger
condition which motivates the following denition.
Denition 2.1. A dissipative medium is said to be positively dissipative iff F C
1
(R) and
F

(x) a > 0, i.e. it is a strictly increasing function with a derivative bounded away from
zero.
Media that are positively dissipative have the important property that F

(x) a > 0, i.e.

F

is bounded below, this will be important in section 3. In particular, in this paper we will
only consider problems for media that are positively dissipative.
We now consider some models for nonlinear media which are dissipative.
Example 1. The Kerr polynomial model. Let G(x) =

N
n=1
a
n
x
n
, with N bounded,
then the inequality (2.8) reduces to
N

n=1
(n 1)a
n
x
n2
0 x R.
This will be satised if the coefcients a
n
of the polynomial satisfy (sufcient conditions)
a
1
arbitrary
a
n
0 n even
a
n
= 0 n odd.
It follows that F(x) = G

(x) =

N
n=1
na
n
x
n1
. This is a generalization of the polynomial
relation that is commonly used for the constitutive model of the Kerr effect, cf equation (2.1)
and [6], i.e.
D
i
=
0
(
1
E
i
+
1
3

3
E
3
i
) i {1, 2, 3}
implies
F(x) =
1
x +
1
3

3
x
3
.
The result proven here shows that this model is guaranteed positively dissipative if
1
and

3
are positive constants, cf section 4.1.
Transient electromagnetic waves in nolinear media 119
Example 2. The Kerr saturation model. This model generalizes the Kerr effect to
include the possibility that for high values of electric eld the nonlinearity saturates. A
rational polynomial model for this can be written as
F(x) =
1
x +
x
3
x
2
+
2
which implies
F

(x) =
1
+
x
4
+3
2
x
2
(x
2
+
2
)
2
.
It then follows that this model is guaranteed positively dissipative if
1
and are positive
constants, cf also section 4.2.
3. Wave splitting
Throughout the following, we assume that the medium is homogeneous and isotropic and
then the constitutive relations become (2.3)
D
i
(x, t ) =
0
F(E
i
(x, ))(t )
B
i
(x, t ) =
0
H
i
(x, t ) i {1, 2, 3}. (3.1)
The assumption concerning the electromagnetic eld components in section 2 implies that
the electric and magnetic vector components lie in a plane tranverse to the z-axis (or the x
3
-
axis). Let w
1
be the transverse component of the electric eld E, and w
2
the corresponding
component of the eld
0
J H, then we can utilize these equations, and (3.1), in (2.2), to
yield the partial differential equation system
1
c
0

t
_
F(w
1
)(z, t )
w
2
(z, t )
_
=
_
0 1
1 0
_

z
_
w
1
(z, t )
w
2
(z, t )
_
.
When the constitutive relations are of the function case of section 2.1, then this system of
equations can be expressed as
1
c
0
_
F

(w
1
)
t
w
1

t
w
2
_
=
_
0 1
1 0
_

z
_
w
1
w
2
_
. (3.2)
For positively dissipative media, denition 2.1 ensures that F C
1
(R), F

(w
1
) a > 0,
so we can rewrite this equation as
1
c
0
_
F

(w
1
)
t
w
1

t
w
2
_
=
_
0 1/

(w
1
)
1/

(w
1
) 0
__
F

(w
1
)
z
w
1

z
w
2
_
. (3.3)
This form suggests that a transformation of the dependent variable will enable this system
to be expressed in a more convenient form. To facilitate this, introduce the anti-derivative
of

F

(w
1
), through
g

(x) =
_
F

(x).
Let it also be assumed that g be subject to g(0) = 0, so that
g(x) =
_
x
0
_
F

(t ) dt.
The properties just assumed for g

(through denition 2.1), and the inverse function theorem,

ensures that the C
1
function g has an inverse g
1
.
120 G Kristensson and D J N Wall
Proposition 3.1. The function g is a homeomorphism if it is associated with an F function
from a constitutive relation, for a positively dissipative medium.
Proof. For F a positively dissipative function, F C
1
(R), F

(x) a > 0, then from

elementary analysis with F strictly increasing this will imply g is strictly increasing and
hence the existence and continuity of g
1
, which is also strictly increasing. Moreover, we
have that g is a diffeomorphism, but this is not required later.
Now, by dening new dependent variables u R
2
u =
_
u
1
u
2
_
=
_
g(w
1
)
w
2
_
the underlying partial differential equation can be transformed from (3.2) into
u
t
= A(u)u
z
. (3.4)
The symmetric matrix A(u) is
A(u) = c
0
_
0 1/g

(g
1
(u
1
))
1/g

(g
1
(u
1
)) 0
_
c(u
1
)
_
0 1
1 0
_
where the wave speed c(u
1
) is
c(u
1
) =
c
0
g

(g
1
(u
1
))
= c
0
d
du
1
g
1
(u
1
). (3.5)
The positively dissipative medium assumption ensures that the function g

(x) is a
positive function bounded away from zero, i.e. g

(x)

a > 0, for all x R. Furthermore,

this implies that c is a function globally bounded above. The matrix A is, therefore, bounded
in the standard operator norm, specically, sup A(u) c
0
/

a. We summarize this result

for later reference in the following proposition.
Proposition 3.2. With g associated with an F in the positively dissipative class, the matrix-
valued operator is bounded as sup A(u) < c
0
/

a.
Equation (3.4) can now, in a special case, be solved by the method of characteristics,
after it has been converted to a diagonal form. This diagonalization can be motivated
physically by the concept of wave splitting; this is because under certain simplifying
hypotheses the dependent variables can then be interpreted as waves propagating in two
opposite directions. To this end, introduce two new dependent variables u

(z, t ) dened by
_
u
+
u

_
=
1
2
_
1 1
1 1
__
u
1
u
2
_
= P
_
u
1
u
2
_
where the transformation matrix has an inverse such that
_
u
1
u
2
_
=
_
1 1
1 1
__
u
+
u

_
= P
1
_
u
+
u

_
.
Substitution of these new dependent variables into (3.4) yields dynamics for the new elds
u

(z, t ) as

t
_
u
+
u

_
= PA(u
+
, u

)P
1

z
_
u
+
u

_
= c(u
+
+u

)
_
1 0
0 1
_

z
_
u
+
u

_
. (3.6)
It should be observed that the two component waves still couple through the functional
wave speed c(u
+
+u

), despite the fact that the wave speed matrix is diagonal. The effect
Transient electromagnetic waves in nolinear media 121
Figure 1. The geometry for the initial-boundary value problem.
of this coupling is to change the slope of the characteristic traces for each of the two wave
equations. This has the useful interpretation that if at some time there is no reected wave,
there will continue to be no reected wave for all future time, while the forward wave is
modied by its own presence in the medium.
Of interest for the remainder of this paper is a wave propagation problem in the half-
space, z 0. Appropriate initial and boundary values for this problem are (see gure 1)
u
+
(z, 0) = f (z)
u

(z, 0) = 0
_
, z 0 u
+
(0, t ) = h(t ), t 0. (3.7)
The system (3.6) is symmetric hyperbolic, so when the function c and its rst derivative are
globably bounded, the Cauchy problem specied by equation (3.6) with the side conditions
of equation (3.7) has at most one classical short time solution. This is shown in [5,
lemma 5.1.1] for the case where the solution has periodic initial conditions, and we extend
their result to apply to our problem in appendix A. However, it should be noted that not
all media have rst derivatives of c that are globally bounded, although two of the models
we consider in sections 4.1 and 4.2 do possess this property. The unique solution of this
system implies that u

(z, t ) 0, for all z 0 and t 0. Then the forward propagating

wave eld u
+
(z, t ) satises the quasi-linear transport equation

t
u
+
+c(u
+
)
z
u
+
= 0 (3.8)
which can be solved by the methods of characteristics. This means that the techniques
utilized in (I) can be applied to the system of rst-order partial differential equations
considered in this paper.
In theorem 3.1 of the next subsection we prove the existence and uniqueness of the
one-way wave equation (3.8), which is shown to be central to our development later in this
section for a wider range of media.
3.1. Solution of the one-way wave equation
We rst examine when the one-way wave equation (3.8) has a classical solution. This
enables us to nd a solution to the direct propagation problem for the partial differential
equation (3.4) with side conditions (3.7). For this section consider the initial-boundary value
122 G Kristensson and D J N Wall
problem on the quarter plane {(z, t ) R
2
|z > 0, t > 0} for the one-way wave equation (3.8)
with side conditions
u
+
(0, t ) = h(t ) t > 0
u
+
(z, 0) = f (z) z > 0.
Sufcient conditions on the boundary value h(t ) and on the initial condition f (z) for the
classical solution of the direct propagation problem to exist can be found in (I); we weaken
those conditions here. Let be a curve in the t z-plane, parametrized by the parameter ,
and dened by t = ( ), z = and the solution of
d( )
d
=
1
c(u
+
(, ( )))
. (3.9)
Then, by (3.8), the variation in u
+
along is
d
d
u
+
(, ( )) = D
1
u
+
(, ( )) +
1
c(u
+
(, ( )))
D
2
u
+
(, ( )) = 0
where D
1
u
+
and D
2
u
+
denote the partial derivatives of u
+
with respect to its rst and
second arguments, respectively. This equation shows that u
+
is constant along , and that
the characteristics dened by (3.9) are therefore straight lines.
The characteristic trace (not necessarily unique) that goes through the point (z, t ), t > 0,
can be described by
( ; z, t ) = t +
z
c(h((0; z, t )))
. (3.10)
Observe, for future reference, that the characteristic trace has slope 1/c. Later, we also
require the inverse of t

= ( ; z, t ), namely =
1
(t

; z, t )the existence of which is

assured for c > 0(see [1] for further use of this notation).
To simplify the notation in the remainder of this section, denote t

= (0; z, t ) by
(z, t ), where this is the intercept of the curve , which passes through the point (z, t ) with
the line z = 0. To determine = (z, t ) for a given value of (z, t ), it is necessary to solve
the equation which is obtained by setting = 0 in equation (3.10):
(t )c(h()) = z. (3.11)
In the previous analysis, it was assumed that (z, t ) > 0. Similar arguments can be used
to nd the solution when (z, t ) < 0, i.e. when the characteristic trace is back propagated
from (z, t ) and cuts the z-axis prior to the t -axis. In this case, the pertinent transcendental
equation is
t c(f ( )) + = z
where the solution = (z, t ) > 0 is the z-value where the characteristic curve intercepts
the z-axis.
The solution of the direct propagation problem, u
+
, is then determined from the
boundary value or the initial condition as
u
+
(z, t ) =
_
f ((z, t )) if (z, t ) < 0
h((z, t )) if (z, t ) > 0.
(3.12)
We formulate as an extension of this analysis a uniqueness and existence theorem, which
is proven in appendix B.
Transient electromagnetic waves in nolinear media 123
Theorem 3.1. Let f, h C
1
[0, ), and c C
1
(R) be given functions such that
f (0) = h(0) and c c
1
> 0, and let T 0 and Z 0 be determined by
1
Z
=
_

_
0 if dc(h(t ))/dt 0 for all t 0
sup
t 0
_

d
dt
1
c(h(t ))
_
otherwise
1
T
=
_

_
0 if dc(f (z))/dz 0 for all z 0
sup
z0
_

d
dz
c(f (z))
_
otherwise.
Furthermore, let = (z, t ) and = (z, t ), for a given value of (z, t ), be determined by
(t )c(h()) = z
t c(f ( )) + = z (3.13)
and let

1
= {(z, t ) : (z, t ) > 0, 0 z Z}

2
= {(z, t ) : (z, t ) < 0, 0 t T }.
Then in the two regions
1
and
2
the continuously differentiable function, u
+
(z, t ), dened
by
u(z, t ) =
_
h((z, t )) (z, t )
1
f ((z, t )) (z, t )
2
(3.14)
is the unique classical solution to

t
u +c(u)
z
u = 0 subject to
_
u(z, 0) = f (z) z 0
u(0, t ) = h(t ) t 0
in the respective regions
1
and
2
.
This theorem gives a maximal extension of the (z, t )-quarter plane for which no shocks
in the solution u
+
can occur. Sufcient conditions for no shocks to occur in u
+
on the whole
of the (z, t )-quarter plane were shown in proposition 2.1 of (I). The maximal extension of
the two regions is given by the rst part of the case statements for 1/Z and 1/T in the
theorem; this coincides with the conditions given in (I).
In section 5 the initial condition function, f , is assumed to be the zero function, and
this means that the consistency condition may be violated as h(0) may not be zero. This
implies that a discontinuous jump in the solution from u = 0 to u = h(0) occurs on the
initial causal wavefront, at time z/c(h(0)), so that the results of theorem 3.1 hold for all
later times.
4. Transmission problem
The results of the previous section are applied to the pulse transmission problem, for specic
cases of positively dissipative nonlinear optical media.
In a physical transmission problem an electromagnetic eld is incident from the left
onto an interface (the line z = 0) of a half space, which is the Kerr medium (z > 0), see
gure 1. It is assumed that the initial conditions are w
1
(z, 0) = w
2
(z, 0) = 0, inside the
Kerr medium, z > 0, for this problem. To the left of z = 0 (z < 0) only the u
+
wave is
124 G Kristensson and D J N Wall
present, and the results in section 3 ensure that u

is never excited for all z. Note in the

physical variables
_
u
+
u

_
=
1
2
_
g(w
1
) w
2
g(w
1
) +w
2
_
so it is observed, that if u

0, then the two components w

1
and w
2
are related through
w
2
= g(w
1
). (4.1)
The u
+
component can be physically measured from the fact that equation (4.1) implies
u
+
= w
2
. Remember that the w
2
component is related to the magnetic eld intensity.
4.1. Kerr polynomial medium
As an explicit example, take the Kerr medium dened by the constitutive function
F(E) =
1
E +
1
3

3
E
3
F

(E) =
1
+
3
E
2
and therefore
g

(x) =
_

1
+
3
x
2
(4.2)
which, for positive constants
1
and
3
, is bounded away from zero, and also has all higher
derivatives bounded. Furthermore, integration gives
g(x) =

1
2

3
ln
x

3
+
_

1
+
3
x
2

1
+
x
_

1
+
3
x
2
2
which is also bounded away from zero. Typical values for the parameters
1
and
3
are

1
= n
2
0

3
= 3Sn
4
0
where the zero-eld refractive index is n
0
= 1.5, and the parameter S is in the range
10
14
10
23
m
2
V
2
(SiO
2
, S = 1.3 10
22
m
2
V
2
and C
6
H
5
NO
2
,
3
=

1
2.89
10
18
m
2
V
2
). The nonlinear term is therefore usually very small, and propagation must
occur over a considerable distance for its effects on the propagating pulse to become very
marked.
The Kerr model can have higher-order dependence, on the eld, as shown in the next
equation for g

. In (4.3) g

is shown with a term

5
, which allows for the possibility of
fth-order dependence of F on the electric eld. This model is utilized in the numerical
reconstruction results presented in section 5.1:
g

(x) =
_

1
+
3
x
2
+
5
x
4
. (4.3)
It should be noted that unlike the third-order model used earlier in this section this modied
model will not have globally bounded derivatives.
It is convenient for numerical simulation to scale the problem through the scaling
transformations:
3

3
/a
2
,
5

5
/a
4
, t at , z az and u
+
au
+
. This ensures
that for our simulations z, t and the eld amplitudes are very many orders of magnitude
smaller than in a physical problem.
Two different diagrammatic presentations of the propagating pulse, through this medium,
are shown in gures 2 and 3. A graphical representation of g, g

and the resultant wave

speed c for this model can be seen in gures 7, 9 and 5, respectively, later.
For the example of SiO
2
with
3
= 0.1, as used subsequently in this paper, then a 10
11
.
Transient electromagnetic waves in nolinear media 125
Figure 2. The propagation of an incident pulse u
+
(0, t ) = 3H(t ) exp{(t 3)
2
} for the Kerr
polynomial model with
1
= 1.5,
3
= 0.1 and
5
= 0. Here, H(t ) denotes the Heaviside
function and the initial condition is f (z) = 0. Four different propagation times are shown,
t = 2, 4, 6 and 8.
4.2. Kerr saturation medium
Often nonlinear effects saturate, and the model discussed in this section will allow this at
eld amplitudes set by the parameter . The constitutive function is
F(E) =
1
E +
E
3
E
2
+
2
F

(E) =
1
+
E
4
+3
2
E
2
(E
2
+
2
)
2
and therefore
g

(x) =
_

1
+
E
4
+3
2
E
2
(E
2
+
2
)
2
. (4.4)
Similar scaling of the parameters, as carried out in the Kerr polynomial model, is performed
on this model. It is seen for positive constants in the model that g

is bounded away from

zero and all higher-order derivatives are also globally bounded.
A diagrammatic presentation of the propagating pulse through this medium is shown in
gure 4. Graphical representation of g, g

and the resultant wave speed c, for this model

can be seen in gures 8, 10 and 6, respectively, later.
Observe in gures 34 that the propagated pulses for = 4.2 and 0.6, respectively,
correspond to nearly the maximal existence distance for the classical solution of the one-
126 G Kristensson and D J N Wall
Figure 3. The propagation of an incident pulse u
+
(0, t ) = 3H(t ) exp{(t 3)
2
} for the Kerr
polynomial model with
1
= 1.5,
3
= 0.1 and
5
= 0. The initial condition f (z) = 0, and
three different propagation distances z = are shown, = 0, 2 and 4.2.
Figure 4. The propagation of an incident pulse u
+
(0, t ) = 3H(t ) exp{(t 3)
2
}. This is for
the Kerr saturation model with
1
= 1, = 4 and = 1. Three different propagation distances
z = values are shown, with = 0, 0.3 and 0.6.
way wave equation. For values of z greater than this distance a shock forms, which can be
seen to be almost occurring in these gures (the shock occurs in the Kerr polynomial and
saturation model for z equal to 4.23 and 0.65, respectively). Also, observe that the near
occurrence of a shock causes a steepening of the trailing edge of the pulse. One of the uses
of Kerr media is to sharpen transmitted pulses. This steepening can be reasoned physically
Transient electromagnetic waves in nolinear media 127
from the property that the wave speed functional increases with decreasing eld amplitude,
see the problems chosen here and shown in gures 5 and 6. It follows that the slope of the
characteristic traces attens with decreasing eld amplitude and therefore, as the travelling
pulse decreases in amplitude, on the trailing edge of the pulse the lower amplitude values
catch up with the higher amplitude parts! This should be contrasted with phenomena in
which the wavespeed increases with increasing amplitude, so causing the shock to form on
the leading edge. Reference [7] may be consulted to see that such leading edge shocks can
occur in shallow-water waves (p 108), and nonlinear transmission lines (p 38).
Figure 5. The true wave speed functional c() and the reconstructed functional (circles) when
the boundary value is h(t ) = 3H(t ) exp{(t 3)
2
}. This is for the Kerr polynomial model with

1
= 1.5,
3
= 0.1 and
5
= 0.
5. Inverse problems
Throughout this section it is assumed that u
+
(z, 0) = f (z) 0, for z > 0, and then the
inverse transmission problem is posed as follows.
Given the initial boundary values as in equation (3.7) for a one-way wave equation (3.8),
and the measured propagated eld at z = , namely u(t ) = u
+
(, t ), for /c(h(0)) < t < T
1
,
reconstruct the medium parameters.
Specically, in the rst instance we reconstruct the wave speed functional c() for as
large an interval of as possible. Here T
1
> /c(h(0)) is a time such that the interval
[/c(h(0)), T
1
] is non-zero. The time that the measurement of u starts is after the initial
wavefront arrives at z = , namely t = /c(h(0)).
128 G Kristensson and D J N Wall
Figure 6. The true wave speed functional c() and the reconstructed functional (circles) when
the boundary value is h(t ) = 3H(t ) exp{(t 3)
2
}. This is for the Kerr saturated model with

1
= 1.5, = 4, and = 1.
It is required that the measured propagated solution is continuously differentiable for
times after the initial wavefront has arrived, so that this imples < Z for all results in this
section.
An implicit solution to equation (3.8) can then be written as
u
+
(z, t ) = h(t z/c(u
+
))H(t z/c(u
+
))
rather than in the form of equation (3.12); here H denotes the Heaviside distribution.
This equation is obtained by rearranging (3.11). Then an inverse reconstruction map is
determined by the inverse of the map
u(t ) = h(t /c( u)) t > /c(h(0)) (5.1)
which is an implicit map from the wave speed c to the measurement u. This equation can
easily be inverted to read
c( u) = /(t h
1
( u(t ))) t > /c(h(0)) (5.2)
and this solution may also be written as
c() = /( u
1
() h
1
()) t > /c(
0
) (5.3)
with t = u
1
() and
0
= h(0). To use this form of solution it is required that u is strictly
monotone; this is in order to use its inverse function and that this can be the case is proven
in (I). However, there, it is required that the function h be monotone, which is not likely
in an inverse experiment. The boundary function, h, is usually only locally monotone over
Transient electromagnetic waves in nolinear media 129
an interval of t R. It is shown here that this implies that u is correspondingly locally
monotone, and it follows that the existence of piecewise local inverses for u and h is
assured. Lemma 2.2, from (I), is now generalized to prove that the transmission function u
is piecewise locally monotone.
Lemma 5.1. The function u is piecewise differentiable and piecewise strictly monotonic,
with the assumptions that the functions h and c are continuously differentiable and piecewise
strictly monotone.
Proof. Upon differentiation of (5.1) with respect to t it follows that
u

(t ) =
h

(t /c( u(t )))

1 c

( u(t ))h

(t /c( u(t )))/c

2
( u(t ))
t > /c(
0
) (5.4)
and if h

> 0, while c

< 0 no shock can occur and equation (5.4) implies u(t ) is strictly
increasing. However, if h

> 0 while c

> 0 the function u(t ) is again strictly increasing but

with the possibility of a shock occurring; in particular, when h

(t /c( u)) = c
2
( u)/(c

( u)),
cf equation (B.1), but because of our assumptions, this does not occur in
1
. Now, if h

< 0
it follows that u(t ) is strictly decreasing, with similar arguments on the dependence of the
sign of c

.
The time values for which the measurement function u is strictly increasing, and
decreasing, will not be the same as for h; for example, see gure 3. It can now be observed
from lemma 5.1 that if h is strictly monotone on the interval [0, T
2
], then u will also be so
on the interval [/c(h(0)), T
2
+/c(h(T
2
))], so implying that the locally monotone interval
for the two functions will not be the same unless c(h(0)) = c(h(T
2
)).
Corollary 5.1. The measurement function u is locally homeomorphic, when both c and
h C
1
are locally homeomorphic.
Proof. The results follow from lemma 5.1 and elementary analysis.
It is seen that with the assumptions made the solution of the inverse problem is unique
and that the solution depends continuously upon the measured data, u, provided these data
are exact.
Theorem. The wave speed functional c(u) can be uniquely reconstructed from the
transmission data u, over intervals in which both h and u are strictly monotone. Furthermore,
the wave speed functional depends continuously on the measurement u with a bounded
Lipschitz continuity coefcient.
Proof. From lemma 5.1 and corollary 5.1, together with equation (5.3), it follows that, if
given an interval [0, T
2
] in which h is strictly monotone, c is uniquely determined.
All that remains to be shown is the value of the Lipschitz constant. Differentiation of
(5.2) with respect to u shows
d
d u
c( u) =

(t h
1
( u))
2
_
dt
d u

d
d u
h
1
( u)
_
so that the appropriate Lipschitz constant

C is

C = sup

(t h
1
())
2
_
1
u

(t )

1
h

(h
1
())
_

with = u(t )
130 G Kristensson and D J N Wall
and t min{[/c(h(0)), T
2
+/c(h(0))], [/c(h(0)), T
2
+/c(h(T
2
))]}. It follows that the
Lipschitz constant exists, and is bounded for (, t )
1
.
It is noted that this proof also shows that c is monotone until h

(t ) = u

(t ), at which
point c

( u) = 0.
It was shown in section 3 that the function g can be related to the wave speed though
the initial-value differential equation, see (3.5),
d
d
g
1
() = c()/c
0
g
1
(0) = 0.
It follows that, once the wave speed has been reconstructed, g
1
can be found uniquely,
so enabling g to be calculated over the region for which g
1
is injective, which by
proposition 3.1 is R, and for which c is known.
Corollary 5.2. The material function g() can be uniquely reconstructed from the
transmission data, and it depends continuously on the measurement data u, provided these
data are exact.
5.1. Numerical results
All the inverse reconstruction experiments were carried out with an incident pulse
h(t ) = 3H(t ) exp{(t 3)
2
}
and this function is piecewise monotone, in two pieces, and satises the conditions required
in the previous section. Note that h(0) = 0 for this incident wave and therefore a nite
jump discontinuity propagates along the characteristic determined by c(h(0)) = z/t .
The solution to the direct transmission problem, as given by the implicit solution (5.1),
was found, as displayed in gures 24, by rst solving the nonlinear equation (3.11) and
then substituting into (3.12). This yields a unique solution up to the point z = Z, as proven
in theorem 3.1. The transmitted elds u for the two models considered in this paper, namely
a Kerr polynomial medium and a Kerr saturated medium, are shown in gures 3 and 4 for
various propagation lengths.
We recap the assumptions made for the inverse reconstruction problem.
Measure the transmitted eld, where this eld is assumed to be continuous, i.e. no
shock has occurred.
The incident pulse, h(t ), is known piecewise strictly monotone and it is continuously
differentiable for t [0, T
1
].
The transmission length is known. If this is not the case it will only effect the
scaling of c and subsequent parameters. The functional form of c can still be found.
The free space wave speed c
0
is known.
For the inverse reconstruction problem h is known, so that the inverse of this function
h
1
can be found over an appropriate interval. Lemma 5.1 ensures that the inverse function
to the measured transmission function, namely u
1
, can be found over an appropriate
interval. Once these inverse functions have been calculated equation (5.3) provides the
solution. The given boundary function h has two monotone sections, 0 < t < 3 and
3 < t < 6, so that (5.3) can be used over two separate intervals. When these two intervals
are used for the non-noisy data, as used here, this yields identical results for c. However,
The inverse function h
1
is found numerically by solving the nonlinear equation h() y = 0, for given y,
with the inverse function dened through = h
1
(y). A similar method is used for u
1
.
Transient electromagnetic waves in nolinear media 131
Figure 7. The true g() function and the reconstructed function (circles) when the boundary
value is h(t ) = 3H(t ) exp{(t 3)
2
}. This is for the Kerr polynomial model with
1
= 1.5,

3
= 0.1 and
5
= 0.
for noisy data, by performing two reconstructions this would enable some smoothing to be
achieved. Figures 5 and 6 show the reconstructed wave speed when = 4.2 and 0.6, for
the Kerr polynomial and the Kerr saturated models, respectively. The results when using
transmission data at = 2 and 0.3, for the Kerr polynomial and the Kerr saturated model,
respectively, cannot be distinguished from these. Note in these gures that c() can only
be found for 0 < < 3 as the applied boundary function only has this range.
In gures 7 and 8, the reconstructed g function is shown for the Kerr polynomial
( = 4.2) model and the Kerr saturated model ( = 0.6), respectively. Identical results are
obtained when = 2 and 0.3 for these respective models. Note in these gures that g()
can only be found for 0 < <
a
, where
a
= g
1
(3), as the applied boundary function
only has this range.
To reconstruct the material parameters {
1
,
3
,
5
} and {
1
, , } for the respective
models, it is more convenient to work with g

() as given by equations (4.3) and (4.4),

respectively, to nd the coefcients. Figures 9 and 10 show the true and reconstructed Kerr
polynomial and saturated Kerr model g

functions, respectively.
Let denote the set of these coefcients and g

denote the reconstructed g

function.
As g

denotes the model representation given by equations (4.3) and (4.4), the appropriate
coefcients can be calculated by the maximum likelihood principle. If g

has been
reconstructed at
i
, i = 0, . . . , N, values of through the previously specied algorithm,
and if
2
denotes the standard discrete
2
-norm over the N + 1 values of
i
, then the
problem of nding the maximum likelihood probability for the coefcients can be written
132 G Kristensson and D J N Wall
Figure 8. The true g() function and the reconstructed function (circles) when the boundary
value is h(t ) = 3H(t ) exp{(t 3)
2
}. This is for the Kerr saturated model with
1
= 1.5, = 4
and = 1.
as
min

2
where in this equation g

is the analytical representation of the model, either (4.3) or (4.4).

For the model given by equation (4.3) the method of the least sum of the squares of the errors
problem is linear, and it is a standard regression problem. However, for equation (4.4) the
method of the least squares problem is a nonlinear regression problem. For the reconstructed
g

functions shown in gures 9 and 10 the respective parameters were found to be

Figure 9:
1
= 1.5,
3
= 1.001,
5
= 0 with an
2
-error of 3.1 10
4
.
Figure 10:
1
= 1.0004, = 4.0006, = 1.0015 with an
2
-error of 1.1 10
3
.
If the wrong choice of models was made, and an attempt made to t a Kerr polynomial
model to the saturated g

data, values of a polynomial were found that were not positively

dissipative, and a
2
-error of 1.2 was obtained, so illustrating a false premise. Similarly, if
an attempt was made to t the saturated model to the reconstructed g

data from the Kerr

polynomial problem, then a least squares error measure of 9.110
4
was found; this was not
a clear false premise. This can be attributed to a well known result in approximation theory
that a rational approximation can provide a wider class of function t than a polynomial.
If the t is made to g
2
.
Transient electromagnetic waves in nolinear media 133
Figure 9. The true g

() function and the reconstructed function (circles) when the boundary

value is h(t ) = 3H(t ) exp{(t 3)
2
}. This is for the Kerr polynomial model with
1
= 1.5,

3
= 0.1 and
5
= 0.
6. Conclusions
This paper has shown that it is possible to reconstruct the material parameters for a nonlinear
medium from measurements of a propagated electromagnetic pulse through a medium. The
inverse algorithm is not model dependent and therefore useful if the underlying nonlinear
model of the material is not known. Specically, we have illustrated the use of the algorithm
for the Kerr medium. Furthermore, the solution depends continuously on the measurement
data. This is a rare occurrence in an inverse problem. However, it should be noted that
in optics the eld quantities are not directly accessible, but only observable through a
functional.
Acknowledgments
One of the authors (GK) wishes to express his gratitude to the Department of Mathematics,
University of Canterbury, Christchurch, New Zealand for the Visiting Erskine Fellowship
that initiated this project and collaboration. The work reported in this paper is partially
supported by a grant from the Swedish Research Council for Engineering Sciences and
their support is gratefully acknowledged.
134 G Kristensson and D J N Wall
Figure 10. The true g

() function and the reconstructed function (circles) when the boundary

value is h(t ) = 3H(t ) exp{(t 3)
2
}. This is for the Kerr saturated model with
1
= 1.5, = 4
and = 1.
Appendix A. Uniqueness result
In this appendix we extend the result of [5, lemma 5.1.1], for the case where the solution
has periodic initial conditions, to apply to our problem. It is shown that the direct problem
posed in section 3 has at most one classical solution. The requirement of global boundedness
of c needed for this result is true for the Kerr and saturated models used in this paper.
Theorem A.1. Let the matrix A(u) R
2,2
be diagonal of the form
A(u) = c(u)
_

1
0
0
2
_
where
1
> 0, i = 1, 2, and let c be bounded and have bounded rst-order derivatives, i.e.
A

< , DA

< . Then the initial-boundary-value problem for u = {u

+
, u

}
satisfying

t
_
u
+
u

_
= A(u)
z
_
u
+
u

_
_
t > 0
z > 0
subject to
u

(z, 0) = f

(z) z 0
u
+
(0, t ) = h(t ) t 0
lim
z
u

(z, t ) = 0 for all t 0

has at most one classical solution.
Transient electromagnetic waves in nolinear media 135
Proof. From the assumptions we have that
A(u) A(v) C
1
u v.
Now suppose u and v are two solutions and denote w = {w
+
, w

} = uv. Furthermore,
w satises the side conditions
w

(z, 0) = 0 z 0
w
+
(0, t ) = 0 t 0
lim
z
w

(z, t ) = 0 for all t 0.

Then

t
w = A(u)
z
u A(v)
z
v = A(u)
z
w +(A(u) A(v))
z
v
and it follows that
1
2
d
dt
(w, w) = (w, w
t
) = (w, A(u)w
z
) +(w, (A(u) A(v))v
z
).
Integration by parts gives, due to the homogeneous boundary condition of w
+
,
(w, A(u)w
z
) = c(u(z = 0))
2
w

(z = 0)
2
(w
z
, A(u)w) (w, A(u)
z
w)
which implies
(w, A(u)w
z
) =
1
2
(w, A(u)
z
w)
1
2
c(u(z = 0))
2
w

(z = 0)
2

1
2
(w, A(u)
z
w) C
2
w
2
.
Therefore, we get
d
dt
w
2
C
3
w
2
so by Gr onwalls lemma, and the initial conditions, this inequality proves that w = 0,
for all t 0, and the theorem is proven.
Appendix B. Proof of theorem 3.1
Proof. First examine the restricted region
1
, and observe that the characteristic trace can
be written as in equation (3.10). If two such traces intersect within the (z, t ) quarter plane
and they originate from the z = 0, t > 0 line, with values t
1
and t
2
, it is easily shown that
they intersect at Z where
1
Z
=
1
t
2
t
1
_
1
c(h(t
2
))

1
c(h(t
1
))
_
=
d
dt
1
c(h(t ))

t = t
for some t [t
1
, t
2
]. (B.1)
In such a case the region
1
is less than the whole quarter plane, and the largest value of the
right-hand side determines the smallest value of z where a shock can occur, and the result
follows; this occurs, for example, when c

< 0 and h

< 0. Consideration of the traces

emanating from the t = 0 line and using a similar argument denes
2
and the time T .
To prove that (3.14) is a uniquely dened solution to our problem we need to use the
implicit function theorem. For (z, t )
1
, the rst equation in (3.13) has a unique solution,
The scalar product is in the spatial variable z, i.e.
(u, v) =
_

0
(u
1
(z)v
1
(z) +u
2
(z)v
2
(z)) dz u =
_
(u, u).
136 G Kristensson and D J N Wall
by the implicit function theorem. In fact if F
1
(z, t ; ) = (t )c(h()) z = 0, then for
(z, t )
1
, which implies z Z,
D
3
F
1
(z, t ; ) = c(h()) +(t )
d
d
c(h())
= c(h())
_
z
d
d
1
c(h())
+1
_
c(h())
_

z
Z
+1
_
< 0 (B.2)
where D
3
denotes differentiation with respect to the third variable. This proves that for
all (z, t )
1
there is a unique solution = (z, t ) to (3.13). Similarly, F
2
(z, t ; ) =
t c(f ( )) + z = 0, gives for (z, t )
2
D
3
F
2
(z, t ; ) = t
d
d
c(f ( )) +1
t
T
+1 > 0 (B.3)
and for all (z, t )
2
, there is a unique solution = (z, t ) to (3.13). Notice that under
the assumptions made on c, f and h, (3.13) have unique solutions and in
1

2
, and
that therefore = 0 = 0, since c(h(0)) = c(f (0)) = z/t .
Furthermore, it is clear that u(0, t ) = h((0, t )) = h(t ) for t > 0, and u(z, 0) =
f ((z, 0)) = f (z) for z > 0; so (3.14) satises the side conditions.
We are now ready to prove that the solution (3.14) also satises the partial differential
equation. Differentiating (3.13) with respect to z and then t gives, respectively,

z
{(t )c

(h())h

() c(h())} = 1

z
{t c

(f ( ))f

( ) +1} = 1
and

t
{(t )c

(h())h

() c(h())} = c(h())

t
{t c

(f ( ))f

( ) +1} = c(f ( )).

Since (t )c

(h())h

() c(h()) = 0 for (z, t )

1
and t c

(f ( ))f

( ) + 1 = 0 for
(z, t )
2
, see (B.2) and (B.3), we get

z
c(h()) +
t
= 0
z
c(f ( )) +
t
= 0.
Therefore, nally from (3.14) it follows that this solution satises

t
u +c(u)
z
u=
_
_
_

t
h((z, t )) +c(h((z, t )))

z
h((z, t ))

t
f ((z, t )) +c(f ((z, t )))

z
f ((z, t ))
_
_
_
=0
_
(z, t )
1
(z, t )
2

References
[1]

Aberg I, Kristensson G and Wall D J N 1996 Transient waves in non-stationary media J. Math. Phys. 37
222952
[2] Agrawal G P 1995 Nonlinear Fiber Optics (San Diego, CA: Academic)
[3] Connolly T J and Wall D J N 1997 On some inverse problems for a nonlinear transport equation Inverse
Problems 13 28395
[4] Karlsson A and Kristensson G 1989 Constitutive relations, dissipation and reciprocity for the Maxwell
equations in the time domain Technical Report LUTEDX/(TEAT-7005)/135 Lund Institute of Technology,
Department of Electromagnetic Theory, PO Box 118, S-211 00 Lund, Sweden (Short published version
1992 J. Electron. Waves Appl. 6 53751)
Transient electromagnetic waves in nolinear media 137
[5] Kreiss H-O and Lorenz J 1989 Initial-Boundary Value Problems and the Navier-Stokes Equations (San Diego,
CA: Academic)
[6] Landau L D, Lifshitz E M and Pitaevski L P 1984 Electrodynamics of Continuous Media (Oxford: Pergamon)
[7] Remoissenet M 1996 Waves called SolitonsConcepts and Experiments 2nd edn (Berlin: Springer)
[8] Shen Y-R 1984 Principles of Nonlinear Optics (New York: Wiley)
[9] Shubert M and Wilhelmi B 1986 Nonlinear Optics and Quantuum Electronics (New York: Wiley)
[10] Wall D J N and Lundstedt J 1998 Inverse source problems involving the one-way wave equation: source
function reconstruction Wave Motion 27 5577