School on Theory and Mathematics

of Ultra-Short Pulse Propagation

Arizona Center for Mathematical Sciences
March 20-22, 2011, Tucson
Computational Methods for Nonlinear PDEs
in Ultrashort Pulse Propagation
Lecture Notes
Miroslav Kolesik
College of Optical Sciences, University of Arizona
Abstract: The sole purpose of this text is to supplement the corresponding ACMS MURI
School lecture. It provides additional material for select parts of the course.
Contents
1 Two facets of simulation in extreme nonlinear optics. . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Propagation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Nonlinear Medium Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Brief overview: Standard medium model for femtosecond laments . . . . . . . . . . . . . 3
1.3.1 Linear material properties: chromatic dispersion and losses . . . . . . . . . . . . . 3
1.3.2 Nonlinear medium response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.3 Optical Kerr eect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.4 Third harmonic generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.5 Delayed response: stimulated Raman eect . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.6 Multiphoton and avalanche ionization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.7 De-focusing by free electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.8 Multiphoton ionization losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Maxwells Equations: The reduction problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1 Numerical solvers of Maxwells equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 The workhorse of nonlinear optics: Nonlinear Schrodinger Equation . . . . . . . . . . . . 11
2.3 Beyond the Nonlinear Schrodinger Equation: Correction terms . . . . . . . . . . . . . . . . 12
2.4 The reduction problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Unidirectional Pulse Propagation Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1 Time-propagated and space-propagated equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Derivation of T-propagated Unidirectional Propagation Equation . . . . . . . . . . . . . . 15
3.2.1 Projection operators for forward and backward propagating elds . . . . . . . 15
3.2.2 T-propagated UPPE as a projection of Maxwells equations . . . . . . . . . . . . . 17
3.3 Derivation of Z-propagated Unidirectional Propagation Equation . . . . . . . . . . . . . . 20
3.3.1 A warm-up exercise: One-dimensional Maxwells equations . . . . . . . . . . . . . . 20
3.3.2 Maxwell equations as a boundary value problem for pulsed beam
propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.3 Z-propagated UPPE for homogeneous media: General case. . . . . . . . . . . . . . 25
3.3.4 Z-propagated UPPE: Simplied, practical version. . . . . . . . . . . . . . . . . . . . . . 26
3.4 Other propagation models as approximations of UPPE . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.1 General method to obtain pulse propagation models from UPPE . . . . . . . . 28
3.4.2 Derivation of Non-Linear Schrodinger Equation from UPPE . . . . . . . . . . . . 29
3.4.3 Nonlinear Envelope Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4.4 Partially corrected NLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4.5 First-order propagation equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Contents VII
3.4.6 Forward Maxwell Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.7 Short-Pulse Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4.8 Maxwells equations versus Wave equation in pulse-propagation . . . . . . . . . 35
4 Numerical Methods for UPPE Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1 UPPE as a large system of ordinary dierential equations . . . . . . . . . . . . . . . . . . . . 37
4.2 Discretization and spectral transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Integration of evolution equations for spectral amplitudes . . . . . . . . . . . . . . . . . . . . . 40
4.4 Parallelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Contents 1
1
Two facets of simulation in extreme nonlinear optics
Simulations in nonlinear optics will in general reect two facets of a modeling problem. The rst
is related to the propagation part of Maxwells equations, while the second relates to constitutive
relations which describe response of a medium to electromagnetic elds. Thus, the rst challenge
is to achieve the capability to describe propagation-related evolution of optical pulse, while the
second is to come up with models of light-matter interactions.
In this course, we restrict our attention to the specic case of optical lamentation. This is
actually an example of a regime in which the two tasks appear appear in conict. Indeed, the
main challenge of computer modeling in this area is one of designing a compromise; Namely a
compromise between the level to which our model is based on rst principles, and the degree to
which such models can be eciently utilized in practical simulations.
A comprehensive approach must therefore address two computer modeling topics, that is
Propagation Models, and Medium Models. This short course concentrates on the former.
1.1 Propagation Models
Form the model point of view, Maxwells equations can be viewed as describing propagation
(as expressed in the curl equations), medium properties (in constitutive relation), and initial
constraints (expressed in the divergence equations). It is only natural to split an optical pulse
simulator framework accordingly. The propagation part of Maxwells equations, namely
H = +
t
D +J
E =
t
B (1.1)
has to be implemented into simulator core. In what follows we describe how this can be done
specically in the context of nonlinear optics and without sacricing any important physics.
Developments in nonlinear optics and particularly in optical lamentation over the last sesqui-
decade inspired a lot of eorts in computational methods aimed at ultrashort pulses. New regimes,
characterized by ultrafast dynamics, super-broad spectra, and approach to wave collapse, meant
that the traditional models based on envelope equations, and Nonlinear Schrodinger Equation in
particular, had to be replaced by more sophisticated approaches. While open problems remain in
emerging areas such as extreme nonlinear optics in structured media with complex geometries,
it is fair to say that at least for bulk media the central problem has been solved. State of the art
in simulation of femtosecond pulses reached a degree of maturity, and practitioners in the eld
have a good grasp on techniques which have been proven to work. This short course aims to be
an introduction into modern methods. However, while instructors describe in detail what works,
they will also illustrate how we got where we are today.
1.3 Brief overview: Standard medium model for femtosecond laments 3
1.2 Nonlinear Medium Models
The medium related part of Maxwell system describes magnetic and electric properties in terms of
so-called constitutive relations. Throughout this course we restrict our attention to non-magnetic
materials, so that we only need one relation that connects electric eld intensity with electric
induction, and a second relation between intensity and the current density induced in the mate-
rial:
D = E +P
NL
({E})
J = J({E}) (1.2)
Here, both nonlinear polarization P
NL
({E}) and current density J({E}) are functions of the
electric eld. It is the purpose of the light-matter interaction model to implement these relations.
Then, current and polarization terms become sources in the propagation equations which drive
creation of new frequency components and reshaping of an optical pulse waveform in general.
Current situation in this area is rather dierent than that in dealing with the pulse prop-
agation sub-problem. One can safely say that models to describe nonlinear interaction at time
scale of few femtoseconds are very far from satisfactory. There are several reasons this area
seems lagging behind the development of Propagation Models. The rst and most important is
that only recently experiments begin to show that what became a standard model of optical
lamentation needs improvement. The standard model, which is little more than a collection of
phenomenological ways to describe various aspects of interaction between light and matter in
laments, is still believed to capture the picture qualitatively correct. However, as experiments
become increasingly more quantitative, gaps become more evident. The second important issue is
that rst-principle models of light-matter interactions are too slow to be integrated into com-
prehensive ultrafast pulse simulations. It is therefore the grand challenge in this area to develop
self-consistent models which go back to rst principles, but are at the same time computation-
ally manageable to a degree that will allow their application as integral parts of light-and-matter
simulators.
1.3 Brief overview: Standard medium model for femtosecond laments
This short course concentrates on pulse propagation modeling, and our discussion is mostly inde-
pendent of what exactly are the nature and a concrete implementation of light-matter interaction
we want to include. Nevertheless, it will be useful to provide a brief overview of the standard
way the model is set up for numerical experiments with optical laments. This is the purpose of
this Section.
1.3.1 Linear material properties: chromatic dispersion and losses
In general, accurate capture of linear properties of the medium in which an ultra-short pulse
propagates is crucially important for modeling all nonlinear interactions. Because the UPPE
approach is spectral, the only potentially dicult issue is the availability of suitable experimental
data to model frequency dependent, complex-valued susceptibility. Within the bandwidth in
which such data can be obtained, the propagator will capture all linear propagation exactly.
We consider a nonmagnetic, dispersive medium with relative permittivity that is a function
of the transverse coordinates x, y and of the angular frequency
= (, x, y) , =
0
. (1.3)
4 1 Two facets of simulation in extreme nonlinear optics
This medium specication implies z-invariant geometry (where z is the propagation direction),
and includes any dispersive homogeneous medium, such as air or water, as well as structured
ber-like media such as photonic, microstructured and tapered optical bers.
In bulk media, an important notion, which appears in numerous places in derivation and
implementation of propagation equations, is the dispersion relation for a plane wave solution to
Maxwells equations, namely
K
z
(, k
x
, k
y
) =
_

2
()/c
2
k
2
x
k
2
y
(1.4)
where ks represent wave-numbers in corresponding transverse directions. We will often refer to
K
z
as propagation constant despite the obvious fact that it is a function of frequency and of
two transverse wave-vector components.
In some special cases, permittivity that appears in the above equation can be understood
as an eective quantity related to a given mode. For example, to describe propagation of light
conned to an eectively two-dimensional, nano-scale thickens glass membrane, becomes an
eective permittivity dependent on the mode conned in the membrane waveguide. Propagation
in the plane of the membrane, is then described by usual plane-wave expansion with an eective
dispersion relation
K
z
(, k
x
)
e
=
_

2
()
e
/c
2
k
2
x
(1.5)
and concrete ()
e
depends on the guided mode, e.g TE or TM.
The important aspect of the approach we present is that () (or ()) is fully incorporated
in the model as a complex-valued function of frequency for all modes of propagation represented
in numerics, and the implementation ensures that all these modes exhibit exact propagation
properties.
1.3.2 Nonlinear medium response
Nonlinear eects are usually described in terms of polarization P through the material constitu-
tive relation:
D =
0
E +P
NL
. (1.6)
The star in this formula represents a convolution integral with being the linear response func-
tion corresponding to the frequency dependent (, x, y) as described earlier. The non-linear
polarization is an arbitrary function of the electric eld P
NL
= P
NL
(E). We will also include
a current density that is driven by the optical eld
j = j(E) (1.7)
to describe interactions with plasma generated by the high-intensity optical pulse. It should
be noted that part of medium reaction to the presence of the light eld included in the current
density may be linear. For example, if free electrons are present, for example generated by another
pump pulse, the corresponding current density induced will be linear in E.
The main physical eects that inuence propagation of ultrashort, high-power light pulses
in nonlinear dispersive media include optical Kerr and stimulated Raman eects, free-electron
generation, defocusing by the generated free electrons and losses caused by avalanche and multi-
photon ionization (MPI). With minor modications, models including these eects can be used
for description of ultra-short optical pulses propagation in gases, condensed bulk media, and in
conventional, microstructured, and tapered bers as well as in ultra-thin silica wires and glass
membranes.
1.3 Brief overview: Standard medium model for femtosecond laments 5
Form the design point of view and for implementation of an UPPE simulator, the important
aspect of medium models is that they are prescriptions to evaluate material reactions formu-
lated in the real space (as opposed to spectral space). That is, given the eld history E(t) for a
given spatial point, these algorithms calculate either polarization or current density as function
of time, and return their results to the solver core.
1.3.3 Optical Kerr eect
The optical Kerr eect is due to electronic response and causes local modication of optical
susceptibility
P
NL
=
0
E (1.8)
which is proportional to light intensity I:
= 2n
b
n
2
I (1.9)
This is the most important eect responsible for the formation of femtosecond light laments.
It causes increase of refractive index where intensity of light is higher, which in turn results in
further (self-) focusing of light. In the absence of other eects this would lead to a runaway eect
called self-focusing collapse.
Note that the intensity I above is a time-averaged quantity, and the frequency content of
P
NL
is therefore essentially that of the driving pulse. However, because bound electrons react to
external elds on atomic time-scales, electronic Kerr eect can be considered as instantaneous.
To our knowledge, actual response times have not been detected, e.g. in two-beam coupling
experiments, and are believed to be well below femtosecond. That is why a model can implement
this eect as reacting to the instantaneous value of an electric eld. Importantly, this results in
a new source of third harmonic frequency.
1.3.4 Third harmonic generation
Consider an instantaneous (electronic) Kerr eect in an isotropic medium. The polarization
response that is third-order in E must be constructed solely from E taken at the given moment
in time, so there is a single vector to work with. Consequently, the only possible form of an
instantaneous third-order nonlinearity is
P
inst
= n
2
(E.E)E . (1.10)
Let us emphasize again, that this simple relation is the only possible for an instantaneous re-
sponse. As soon as there is memory, two frequency dependent components of third-order suscep-
tibility tensor are needed for full description. The frequency content of P
inst
consist of both the
fundamental frequency of E and its third harmonic. This is the main source of third harmonic
radiation observed in femtosecond laments.
It is worthwhile to remark that Readers will nd in the literature also a very dierent ap-
proach to modeling third-harmonic generation. In works and simulations based on application
of envelope pulse propagation equations, several authors used two envelope functions, one for
the fundamental frequency and one for the third harmonic generation. We emphasize that two-
envelope method is incorrect and should be avoided. This is because situations are frequently
encountered in which spectra become extremely broad. Then, distinction between fundamental
and third harmonic is impossible, and any two-envelope parametrization is therefore non-unique
and the model fundamentally inconsistent.
6 1 Two facets of simulation in extreme nonlinear optics
1.3.5 Delayed response: stimulated Raman eect
Air is by a large part made of two-atomic molecules with dierent polarizibilities parallel and
perpendicular to their symmetry axes. This leads to a nonlinear eect which is referred to
as stimulated Raman eect, although recently the nomenclature acknowledges the fact that
reorientation of molecules plays a central role in it.
The interaction energy of a molecule in an external eld is such that it prefers to align with
the elds direction. When a femtosecond pulse hits such a molecule, it excites rotational
motion; this is a stimulated Raman eect. Molecular rotation in turn changes the eective linear
polarizibility of the molecule as projected on the direction of the eld. The latter then experiences
a modied index of refraction.
Because the interaction Hamiltonian is quadratic in eld, the eect is of third-order. It is
therefore often considered a companion of the electronic Raman eect. Taken a very dierent
microscopic origins, this may seem arbitrary, but one has to keep in mind that manifestations of
the two eects (i.e. self-focusing) are very dicult to distinguish in relatively longer pulses.
A proper, rst-principles model would need to integrate quantum mechanical equations of
motion for a density matrix describing the rotational stale of an ensemble of molecules. Note
that such a system is to be solved at each spatial grid location, and at each propagation step!
Instead of this (relatively) dicult calculation, the Raman eect is approximately parametrized
with
(t) = 2n
b
n
R
2
_

0
R()I(t )d , (1.11)
where R() sin()e

is often sucient for ultrashort pulses. It is a memory function

which represents the response of the system to an excitation by a very short impulse. This
simple formula has the advantage of easy implementation that avoids explicit calculation of the
convolution integral 1.11. Sometimes an even simpler, exponential memory function R() e

has been used in simulations. If the actual memory function is suciently complex, as is the case
in silica for example, a numerical convolution approach may be better suited to calculate the
response (t).
1.3.6 Multiphoton and avalanche ionization
Because of high light intensities occurring in femtosecond pulses, free electrons are generated by
Multiphoton Ionization (MPI), tunneling, and avalanche mechanisms. Then it is necessary to
account for the response of the optical eld to the presence of a dilute plasma. Since the relevant
times scales are so short, electron diusion and ion motion are neglected, and the free-electron
density is usually obtained as a solution to an equation of the following form

t
= aI +b(I) c
2
. (1.12)
Here, I is the light intensity, a parametrizes the avalanche free-electron generation, and b(I)
represents the Multi Photon Ionization (MPI) rate that is a highly nonlinear function of the
intensity. The last term describes plasma recombination, and is often neglected in sub-hundred
femtosecond pulses.
Various representations exist for the ionization rate b(I). Frequently a power-law t to ex-
perimental data or a theoretic formula result is obtained in the form
b(i) =
_
I
I
0
_
K
(1.13)
1.3 Brief overview: Standard medium model for femtosecond laments 7
where and K (not necessarily an integer!) are ionization cross section and (eective) multi-
photon order. Because laments actually occur at intensities for which multi-photon ionization
crosses-over to tunneling regime, the eective order K is usually signicantly lower than the
number of photons needed to bridge the band-gap of the material.
In extreme intensity regimes, when ionization rate is mostly due to tunneling, b is given
as a function of the electric eld intensity. Then, because electrons are mostly generated at
moments when its amplitude reaches a maximum, a staircase-like temporal prole arises in the
free electron density and current. This gives rise to a secondary source of third (and higher)
harmonic generation which has recently been detected.
1.3.7 De-focusing by free electrons
In the spirit of a Drude model, it is assumed that a collective electron velocity v responds to the
optical eld, and the total current density is governed by the following simple equation
d
dt
j(t) =
e
2
m
e
(t)E(t) j(t)/
c
(1.14)
where
c
stands for the mean time between collisions experienced by electrons. Note that v
represents an average velocity induced in the ensemble of electrons by the driving eld E, it is
not an individual electron velocity. This equation is solved together with (1.12) to capture eects
of free electrons on the propagation of the optical eld, namely defocusing and losses.
Alternatively, one can treat free-electron induced eects as a susceptibility modication,
and lump them with the rest of P which in turn simplies numerical calculation. The price
for this is that one must neglect the chromatic dispersion induced by free electrons. Then,

t
P = j is interpreted as a time derivative of nonlinear polarization and its contribution to P
is approximated by
P =
0

pla
()E =
ie
2
m
e

R
(1/
c
i
R
)
E (1.15)
with
R
being a chosen reference angular frequency. It needs to be emphasized that this approxi-
mation completely neglects the plasma induced chromatic dispersion. This may be unacceptable
for some numerical experiments, for example pump-probe with fundamental and second har-
monic. In such a case, only the current density based description can properly reect the fact
that the second harmonics is experiencing four times weaker de-focusing caused by free elec-
trons. However, in many lamentation modeling scenarios, susceptibility based formulation gives
results in practical agreement with those obtained from the more accurate model based on J.
It is worthwhile to note that in the framework described so far, electrons interacting with a
light pulse are treated as either bound (in their respective atoms and molecules) or completely
free and oblivious to the presence of their parent ions. These free electrons are described in
terms of a Drude model which is more suitable for true plasma. Here we have to keep in mind
that during the rst few tens of femtoseconds there is not enough time for plasma to establish
itself as a collective ensemble of electrons and ions. Moreover, electrons liberated from atoms
and molecules are likely to interact with the parent ions - this eect, which among others leads
to High-Harmonic Generation, is completely neglected. These are obviously the issues the next-
generation modeling will have to account for.
1.3.8 Multiphoton ionization losses
Losses caused by multiphoton ionization are usually incorporated as either an equivalent current
or an imaginary susceptibility contribution that extracts from the eld the energy needed for the
8 1 Two facets of simulation in extreme nonlinear optics
free-electron generation.
J
loss
.E b(I)E
g
(1.16)
Here we equate the rate of energy loss suered by the eld with the rate of free electron generation
multiplied by an estimated energy cost E
g
of one free electron.
Note that although this is a universally utilized approximation in the femtosecond pulse
propagation area, it is a very crude model, very likely not adequate when the pulse spectrum
broadens in such a way that new frequencies carry a signicant portion of its energy. In such
a situation the absorption losses as well as MPI generation rates should be frequency selective.
This is just one aspect in which the currently standard medium model is less than satisfactory.
2
Maxwells Equations: The reduction problem
The Maxwells equations, while eciently solved by numerical solvers in many areas of com-
putational electromagnetics, are rather ill suited for straightforward computer simulations of
nonlinear phenomena in optical pulses of ultra-short duration which deliver TerraWatt-scale
powers. The following Sections explain why is it that direct Maxwell solvers are of limited util-
ity in extreme nonlinear optics. We review various ways this problem has been dealt with in
traditional nonlinear optics. We then formulate what we call the Maxwell reduction problem,
which is the need for an eective solver tailored specically to the pulse propagation context,
and which can model realistic situations without compromising underlying physics.
2.1 Numerical solvers of Maxwells equations
Throughout the several past decades, formulation of pulse propagation problems in nonlinear
optics has been most often based not directly on the physical elds that appear in Maxwells
equations, but on so called envelope equations. These are designed to eliminate the fast oscillatory
temporal and spatial variations related to the central frequency and wavenumber of a carrier light
wave. Eectively, physical elds are represented by slower evolving complex-valued envelopes.
The rationale behind such approximations is twofold. First, envelope equations are natural and
accurate approximations to Maxwells equation in many situations characteristic of nonlinear
optics, and provide a useful theoretic basis for building models for phenomena such as harmonic
generation. Second, such equations make it feasible to simulate many nonlinear phenomena
with sucient accuracy. While numerical methods for direct simulation of Maxwells equations
have been available, a closer look at the problem will quickly reveal that in a typical nonlinear
propagation context direct Maxwells solvers are of little practical use. One may question this
statement and argue that cheap, widely available computing power will surely make the direct
attack feasible in the nearest future. Actually, the answer is still negative, and it is the intent of
this section to demonstrate that the development of better theoretic models underlying numerical
simulators is a much more viable strategy than waiting for the computer power to catch up with
our requirements.
So, let us briey review, the general properties of direct Maxwell solvers in order to understand
the implications for description of nonlinear pulse propagation. We will restrict ourselves to the
so called Finite Dierence Time-Domain (FDTD) solvers because they are the most common
representatives of direct Maxwell solution methods. However, the discussed issues are relevant
to Maxwells equations simulators in general.
First, let us consider memory requirements. A direct solver works over a xed spatial domain,
and evolves the grid-based representation of the electric and magnetic elds in discrete time
10 2 Maxwells Equations: The reduction problem
steps. As a rule of thumb, for accurate simulations one typically needs about 30 grid points per
wavelength in space. About the same number of time-steps must be executed per a single wave
oscillation. Such a resolution may be practically achievable for radio- or micro-waves and small
simulated volumes (as measured in units of cubic wavelengths). But in the optics context, these
resolution requirements translate into sub-micrometer spatial, and sub-femtosecond temporal
resolution. If an optical pulse beam is only one centimeter wide, and propagates over a laboratory-
scale distance of only a few meters, a single snapshot of the eld will require of the order of 10
20
eld-variables to store in the memory. Moreover, for each meter of propagation, each of these
variables must be updated through tens of millions of integration steps! No matter how fast
typical computer memory and performance continue to increase, it becomes clear that a brute-
force approach in an optics context is not going to be feasible in any foreseeable future.
The second property which practically disqualies direct solvers for nonlinear propagation is
numerical dispersion. Similar to natural light waves, phase and group velocities of numerical
waves depend on their frequency - they exhibit so-called numerical dispersion. It is of course a
general property of each and every numerical method which simulates wave propagation, that
actual wave properties are warped, or modied because of the discrete nature of the simulation.
More precisely, the true relation between wave propagation velocity and its frequency, which is
usually called the dispersion relation, is replaced by an articial one which is signicantly dif-
ferent. Typically, a discrete numerical method can only mimic correct dispersion properties of
waves for small frequencies only. As a consequence, only a very small fraction of the bandwidth
which is available to the numerics can actually be utilized. This articial deformation of the nat-
ural relation between the wave velocity and its frequency is very restricting in all discretization
schemes for Maxwells equations, This is because numerical dispersion depends on the grid reso-
lution, and is in general extremely strong from the point of view of nonlinear optics. For example,
a numerical vacuum simulated by a direct solver will typically exhibit chromatic dispersion
orders of magnitude larger than gases and comparable to water or other transparent condensed
media. Surely, these unwanted eects decrease in magnitude as we increase the grid resolution.
But to get them fully under control, the required resolution would be enormous. Because in the
simulation of nonlinear optics it is absolutely crucial to capture the chromatic dispersion very
accurately, this problem becomes extremely serious.
Last but not least, it turns out that in direct solvers it is actually quite dicult to implement
models of nonlinear and dispersive media. The origin of the problem is that dispersion and often
also nonlinearity are connected to some kind of memory in the medium. This does not mesh well
with the fact that a direct solver scheme is naturally designed to store only a single temporal
snapshot of a eld conguration. If the reaction of the medium at any given point depends on the
history of the local eld, we must keep sucient information about this history available to the
numerical solver. This can easily multiply the memory needs. The problem gets even worse once
we consider that the frequency-dependent properties of the model media are also plagued by
numerical dispersion. The articial deformation of the actual numerical medium response can
be signicantly dierent from the targeted frequency-dependent properties for the same reason
we pointed out for the linear wave propagation.
Thus, it is clear that the direct numerical solution of Maxwells equations is not feasible
for many nonlinear optic phenomena. Next, we will have a look back into the recent history
and review how this problem has been by-passed through the use of envelope equations. As a
prototype of envelope equations utilized in optics, we rst turn our attention to the so called
Nonlinear Schrodinger Equation.
2.2 The workhorse of nonlinear optics: Nonlinear Schr odinger Equation 11
2.2 The workhorse of nonlinear optics: Nonlinear Schrodinger
Equation
As the name suggests, this equation is closely related to the Schrodinger equation of quantum
mechanics. Apart from the nonlinearity, the two are almost equivalent. Indeed, paraxial dirac-
tion of a continuous-wave optical beam is the same as the quantum-mechanical spreading of the
wave-packet of a free particle. Moreover, the nonlinear term can be often interpreted as a po-
tential that appears in the quantum mechanical version of the equation. However, nonlinearity
turns the optical version into a much richer equation. In fact, its applications in physics go way
beyond optics and apply to nonlinear dispersive waves in general. This is the reason behind the
enormous amount of related research in both mathematics and physics communities.
The mathematical form of the Nonlinear Schrodinger Equation (NLS) is rather simple:

z
A =
i
2k
0
A
k

2

tt
A+
i
0
n
2
c
|A|
2
A
Here, A represents the complex envelope of an optical pulse centered around the carrier angular
frequency
0
. The rst term on the right-hand-side, with k
0
standing for the propagation constant
corresponding to the carrier frequency, describes diraction. The next term captures chromatic
dispersion, with k being the group velocity dispersion. The nonlinear term represents the optical
Kerr eect in which the nonlinear modication of the refractive index is simply proportional to
the light intensity. The latter is here expressed by |A|
2
, and n
2
is the so-called nonlinear index
of the medium.
Equations of the same form, often with dierent dimensions of the Laplacian, arise in many
areas of physics. In optics, NLS is a prototype equation to describe pulse propagation in nonlin-
ear, dispersive transparent media. Taken its simplicity, it may come as a surprise that it captures
a wealth of nonlinear phenomena in bers and bulk media alike. Often, this equation works quali-
tatively very well even beyond the boundaries given by the paraxial and quasi-monochromatic
assumptions under which it is supposed to be valid. That is just one reason this and related
equations continue to receive signicant attention in the mathematical and physical literature.
We will discuss one of the many ways to derive this important equation in Section 4, where
we also specify precisely the physical assumptions that are necessary for its validity. However, at
this point it is instructive to be just a little more specic about the relation between the complex
amplitude A and its real physical eld counterpart E. For simplicity, we restrict ourselves to to
a one-dimensional, scalar case of linearly polarized light. Usually, the relation between the two
quantities is written as
E(x, t) = A(x, t)e
i0t+ik0x
+ c.c. .
Here,
0
and k
0
are the carrier frequency and wavenumber, and the purpose of the oscillatory
exponential is to account for the fast changing carrier-wave of the eld. The complex amplitude
A then can be viewed as a modulation of the carrier. This of course only makes sense if the
modulation speed is small in comparison to the angular frequency
0
. Said in an equivalent
way, the spectral content of A(x, t) must be narrow in comparison with
0
. And this is where
the need to improve this equation originates: In modern nonlinear optics, we often deal with
waveforms E(x, t) which change their amplitudes signicantly within a single oscillation of the
carrier. In such a situation even the carrier frequency becomes an ill dened notion. Roughly
speaking, this motivated researchers to propose a number of improved equations which were
intended specically for the modeling of ultra-fast optical pulses. And that is the topic of the
next section.
12 2 Maxwells Equations: The reduction problem
2.3 Beyond the Nonlinear Schrodinger Equation: Correction terms
With the advent of relatively available femtosecond lasers and subsequent accelerated expansion
of nonlinear optics into the ultra-fast domain, it became evident that NLS is a rather crude
model to describe phenomena such as supercontinuum generation. A single-color optical pulse
explosively broadening its spectrum such that it can cover the whole visible region and beyond
is arguably one of the most spectacular eects in nonlinear optics.Intuition alone tells us that de-
scribing such extreme events in terms of envelopes the way we sketched it in the previous Section
is not realistic. Moreover, it turns out that the white light generation, as the supercontinuum is
often called, is an extremely robust and universal phenomenon. The ubiquitous nature of such
a nonlinear interaction underlines the need to go beyond the NLS and introduce better pulse
propagation models.
Researchers therefore developed a whole series of equations which include the so-called cor-
rection terms. These are modications to the NLS equations intended to make it more robust,
and especially applicable to very short-duration, high-intensity optical pulses. Some of these
corrected propagation equation will be described in the next chapter in connection to our
nonlinear pulse propagation model. Here, we would like to restrict ourselves to a single example
to give the reader an idea about what types of corrections are possible to improve the NLS. The
following is the Nonlinear Envelope Equation (NEE), developed by Brabec and Krausz:

z
A+v
1
g

t
A = iD(i
t
)A+
i
2k
R
(1 +
i

t
)
1

A+
ik
R
2
0
n
2
b
(
R
)
(1 +
i

t
)P
Similarly to NLS, NEE is an envelope equation, but it is formally free of the quasi-monochromatic
approximation. Therefore it is much better suited for very short duration pulses. Without going
into details, let us note some of the corrections that take this model well beyond the NLS. The
rst is the occurrence of the dispersion operator D on the right hand side of the equation. This
is a formal innite series in temporal derivatives,
t
:
D(
R
) =

n=2
_

n
k

n
_
=R
(
R
)
n
n!
(2.1)
Setting aside the questions of how to properly dene and implement such an operator, let us just
say that this can in principle take into account chromatic properties of the medium in a wide range
of frequencies. Linear medium properties, specically its frequency-dependent index of refraction
and loss, are encoded in the dispersion operator through the innite series of coecients in (2.1).
The important point to note is that in comparison with NLS, NEE has a sucient number of
parameter slots (in contrast to only two in NLS) to incorporate a detailed medium description.
Another correction to note is the inverse operator in front of the diraction Laplacian. This
is the so-called spatial-temporal diraction correction term and it takes into account the depen-
dence of diraction on the color of light. This reects the well known fact that the longer the
wavelength, the more the light diracts and bends around obstacles.
Yet another important correction stands in the front of the nonlinear polarization P: It
is called the self-steepening term which is responsible for strong generation of new spectral
components in the trailing edge of a propagating pulse. Intuitively, this correction term can be
viewed as a modication of the propagation velocity which depends on the local light intensity.
The higher the intensity, the slower the propagation. This makes the intensity peaks to lag behind
their pedestals, and consequently steepens the trailing edge of the pulse. In turn, the steep edge
represents a wide spectral content. Because of this, this correction is to a large degree responsible
for capturing generation of the blue-shifted light as observed in many experiments.
2.4 The reduction problem 13
In comparison to NLS, the Nonlinear Envelope Equation is a very realistic model. Our super-
cial look at it already makes it evident that its robust numerical implementation requires close
attention to the correction terms. More importantly, it is fair to say that the original derivation
of the equation was based on intuitive, but, strictly speaking, unjustied arguments. This is
the case for a whole number of corrected equations presented in both optical and mathematical
literature. The discussion in later sections will identify cases in which weak arguments resulted
in rather poor equations: in one example, physicists were misled by intuition and possibly by
their desire for equations easily implemented in numerics which in turn lead to an outright in-
correct model. Mathematicians, on the other hand, may be willing to sacrice realistic models
to ones described by nice (for example integrable) equations which, however, may turn out to
be utterly unsuitable to the very application they were meant for. The celebrated Short Pulse
Equation serves as an example here. Because of built-in assumptions, which may seem benign
at a rst supercial look, this equation can not capture generic behaviour of short pulses in
supercontinuum generation in optical bers. Lessons like these lead us to ask an important ques-
tion; what kind of qualities do we need to require of our pulse propagation model if it has to
aspire to be a truly realistic model applicable in extreme nonlinear optics? We address this in
the following.
2.4 The reduction problem
Thus, we have identied the problem of Maxwell reduction, which means to replace the full
set of equations for electromagnetic eld with a reduced set capable of handling the extreme
conditions encountered in modern nonlinear optics. Specically, this requires:
Capability to capture, without any approximations, linear propagation regimes in transparent
media, i.e. in media with relatively small, but frequency-dependent absorption, but with
potentially complex chromatic dispersion landscapes.
Capability to handle spatio-temporal waveforms with extremely wide spectra extending over
several octaves of frequency. For this it is necessary to work on the level of physical elds,
and avoid introduction of envelopes and related articial parameters (such as the carrier
frequency that appears in envelope propagation equations).
It is important to implement such models of light-matter interactions that can be valid over
broad bandwidth and exhibit long response memory (i.e. dependence on the history of the
driving eld).
It is crucial that the pulse evolution equations can be implemented in a numerically ecient
way that allows simulations of nonlinear pulse propagation over laboratory-scale distances.
Unidirectional Pulse Propagation Equations represent the rst approach to put this problem
on rm mathematical ground. They are eective Maxwell solvers designed for the pulse propaga-
tion context, i.e. especially for situations in which the light propagates in a well dened direction.
This can be achieved without sacricing physics, while eliminating the practical limitations of
the direct Maxwell solvers described in the beginning of this Section. The next Section is devoted
to derivation of various forms of Unidirectional Pulse Propagation Equations.
3
Unidirectional Pulse Propagation Equations
In this chapter, we outline the key steps in deriving a physically self-consistent and robust ultra-
short pulse propagator that resolves the underlying optical carrier-wave while also enabling prop-
agation over many meter propagation lengths. Our goal is to retain the full rigor of Maxwells
equations while reducing the problem complexity by constraining the model to unidirectional
propagation. As our immediate interest is in very short-duration and high-intensity pulse propa-
gation in which the induced nonlinear polarization can have signicant impact on the dynamics,
we will need to accurately capture the very broad spectral landscape that the pulse experiences
during its interaction with a host dielectric material. In many cases, the spectral super-broadening
is such that the generated bandwidth exceeds the underlying carrier frequency, i.e /
0
>> 1.
In this regime, we expect the Nonlinear Schrodinger Equation (NLSE) to fail. Many attempts
have been made to derive nonlinear envelope models that go beyond NLSE. We discuss some of
these, and we show explicitly how each can be seamlessly derived from the Unidirectional Pulse
Propagation Equation (UPPE).
3.1 Time-propagated and space-propagated equations
Most of the pulse propagation problems in nonlinear optics are solved in one of two formulations:
Either the numerical evolution proceeds along the time coordinate, or it follows, or propagates
the wavepacket along one of the spatial coordinates, usually chosen as z, in the direction of the
laser beam. In the rst case, termed time-propagated evolution, one has an initial condition
(i.e. some description of the electric and magnetic elds) specied in all space for a given initial
time. The evolution is calculated along the time axis, and naturally reects the structure of the
Maxwells equations. In the second case, termed z-propagated evolution, the initial condition
is given as a function of the local pulse time and of the transverse coordinates (w.r.t. propagation
direction). The numerical evolution proceeds along the propagation axis. From the mathematical
point of view, this case is an initial value problem very much the same way the t-propagated case
is. However, from the physical point of view this is a rather subtle issue because the true initial
condition requires knowledge of the total eld in the past and in the future. This includes the
light which may be nonlinearly reected from the focal region of an experiment. Only if we
can assume that this is suciently weak, we can solve the corresponding initial value problem.
We will refer to the corresponding propagation models as time- and z-propagated equations.
The z-propagated approach is much more common in nonlinear optics simulations based on
envelope equations, and is often related to Nonlinear Schrodinger Equation. On the other hand,
the time-propagated approach is common for solvers based on direct integration of Maxwells
3.2 Derivation of T-propagated Unidirectional Propagation Equation 15
equations. The time-propagated versions of UPPE are more suitable for tight-focusing scenarios
when the non-paraxial eects start to play a role. On the other hand, the z-propagated equations
are easier to use in situations that allow us to neglect the longitudinal eld components as
contributing sources of nonlinear material responses. That is the main reason the z-propagated
approach is more common, specically in nonlinear optics.
First we focus our attention on the time-propagated UPPEs, because this approach is concep-
tually simpler. We start with derivation which is based on a formal projection technique designed
to separate what we call forward and backward propagating optical eld components. We discuss
its numerical implementation in Section 4, while its relation to previous envelope models will be
discussed together with the z-propagated methods later in the next section.
Most of the room in what follows is devoted to the z-propagated approach. Since the general
derivation method, while straightforward in principle, is rather involved, we rst show a simple
one-dimensional example of scalar propagation equation. This will reveal the mathematical na-
ture of certain assumptions one has to adopt; these are important in that they select from the set
of all Maxwells equations solutions those that fall into the category of pulsed beams. We then
follow with the general case, which is subsequently specialized to equations in the homogeneous
bulk media and to wave-guide, or ber-like structured media.
3.2 Derivation of T-propagated Unidirectional Propagation Equation
The version of UPPE that is numerically evolved along the time axis is termed a t-propagated
equation. It is, strictly speaking, a pair of equations, each describing a component of the optical
eld with a prevailing direction of propagation. The derivation is based on the notion of a
projection which selects only the forward (or backward) propagating part of the eld. It must
be emphasized that the forward-backward split is formally exact. However, approximations are
necessary for reducing the system into unidirectional equation, which we do by assuming that
the nature and strength of the nonlinear response is such that it can be accurately calculated
from the forward eld component only. This in turn implies that the gradients of the nonlinearly
modied medium properties are changing on length scales large in comparison to wavelength.
In the next Section, we describe the projector and its properties in detail. We then use it to
derive the time-propagated UPPE.
3.2.1 Projection operators for forward and backward propagating elds
Consider general electromagnetic eld E(r, t) H(r, t) and the corresponding induction, D(r, t),
for a homogeneous nonmagnetic medium characterized by its frequency dependent susceptibility
().
Let us denote E(k), H(k) and D(k) the spatial Fourier transforms of the corresponding
elds, e.g.
D(k) = F{D(r)}(k) D(r) = F
1
{D(k)}(r) (3.1)
In the Fourier, i.e. the spatial spectrum space, we can dene the following operators
P

_
D(k)
H(k)
_
=
1
2
_
D(k) sgn(k
z
)
1
(k)
k H(k)
H(k) sgn(k
z
)
(k)
k
2
k D(k)
_
. (3.2)
The reader will notice that this pair of operations is diagonal in the wave-vector k, and is closely
related to the properties of plane-wave solutions of linearized Maxwells equations. Next we show
that P

have the properties of projectors onto forward and backward propagating elds; they
must be idempotent in a certain sense, and have to constitute a decomposition of unity.
16 3 Unidirectional Pulse Propagation Equations
Obviously, the two operators P

provide a unity decomposition

P

+P
+
= 1 (3.3)
which follows directly from their denition.
Further, one can show that as long as they act over the divergence-free subspace, they behave
as projectors, namely P
2
= P. To see this, let U, V be two arbitrary divergence-free vector elds,
which satisfy .U = .V = 0. Then
P

_
U(k)
V (k)
_
=
1
2
P

_
U(k) sgn(k
z
)
1
(k)
k V (k)
V (k) sgn(k
z
)
(k)
k
2
k U(k)
_
=
1
4
_
U(k) sgn(k
z
)
1
(k)
k V (k)
V (k) sgn(k
z
)
(k)
k
2
k U(k)
_
+
1
4
_
sgn(k
z
)
1
(k)
k [V (k) sgn(k
z
)
(k)
k
2
k U(k)]
sgn(k
z
)
(k)
k
2
k [U(k) sgn(k
z
)
1
(k)
k V (k)]
_
=
1
4
_
U(k) sgn(k
z
)
1
(k)
k V (k)
V (k) sgn(k
z
)
(k)
k
2
k U(k)
_
+
1
4
_
sgn(k
z
)
1
(k)
k V (k)
sgn(k
z
)
(k)
k
2
k U(k)
_

1
4
_
1
k
2
k k U(k)
1
k
2
k k V (k)
_
=
1
2
_
U(k) sgn(k
z
)
1
(k)
k V (k)
V (k) sgn(k
z
)
(k)
k
2
k U(k)
_

1
4
_
1
k
2
kk.U(k)
1
k
2
kk.V (k)
_
=
1
2
_
U(k) sgn(k
z
)
1
(k)
k V (k)
V (k) sgn(k
z
)
(k)
k
2
k U(k)
_
=
P

_
U(k)
V (k)
_
(3.4)
Since the radiation elds D, H are divergence free, the above calculation shows that P

act as
projectors on these elds, with P

= P

.
Finally, it is easy to check that P
+
leaves invariant any plane wave solution to Maxwells
equations that propagates with k
z
wave-vector component larger than zero (in the positive
z-direction): At the same time, it annihilates all plane waves propagating in the negative z-
direction. To show this, we use the fact that in a plane wave there is a denite relation between
the electric and magnetic elds:
_
D(k)
H(k)
_
=
_

1
(k)
k H(k)
(k)
k
2
k D(k)
_
. (3.5)
Acting with P

on both sides of this equation one obtains

3.2 Derivation of T-propagated Unidirectional Propagation Equation 17
P

_
D(k)
H(k)
_
=
1
2
_
D(k) sgn(k
z
)
1
(k)
k H(k)
H(k) sgn(k
z
)
(k)
k
2
k D(k)
_
=
1
2
_
(1 sgn(k
z
))
1
(k)
k H(k)
(+1 sgn(k
z
))
(k)
k
2
k D(k)
_
=
+1 sgn(k
z
)
2
_
D(k)
H(k)
_
(3.6)
Here, the right-hand-side is either zero or equal to the original eld, depending on which direction
the plane wave propagates. Naturally, it follows the same way that P

leaves invariant the back-

ward propagating plane waves while cancelling all forward waves. Therefore, we can summarize
that P

each project out from an arbitrary radiation eld the component that propagates in the
positive (negative) z-direction.
3.2.2 T-propagated UPPE as a projection of Maxwells equations
The time-propagated UPPE describes the electric induction rather than the electric eld inten-
sity. The rationale behind this is that D is divergence free in the absence of free charges, but
this is not the case for E. From
0 = .D = .E +.P ,
we can see that the divergence of the electric vector eld is essentially the opposite of that
of the polarization. The latter in general possesses a non-zero divergence due to the nonlinear
interactions. For example, consider the optical Kerr eect, for which the modication of the index
of refraction is proportional to the light intensity. Because of the spatial prole of the latter, the
divergence of the polarization is
.P .(n
2
IE) = n
2
I.E +n
2
I.E
From this and the previous equations we can see that the gradient of the intensity gives rise to
the non-zero divergence of the E eld. That is why D is a more natural partner for H as both
vector elds are divergence free.
The use of the electric induction D as the primary representation of the electromagnetic eld
is similar in spirit to the approach utilized in many numerical solvers of Maxwell equations which
evolve D and calculate E from a given (nonlinear) constitutive relation at each step
D =
o
E +P
NL
(3.7)
Here, the D eld is supposed to be known in the whole space from previous calculations or from
an initial condition. P
NL
characterizes the medium nonlinearity and is usually specied in terms
of a function of the electric eld. Consequently, the above equation must be solved for the electric
eld intensity that becomes a function of the electric induction.
We will use the constitutive relation in the Fourier representation
D(k) =
o
((k)) E(k) +P
NL
(k) , (3.8)
where (k) is dened implicitly as an angular frequency of a plane wave with a wave-vector k,
(k)
2
((k)) = k
2
.
To isolate the linear part of the Maxwells equations from the nonlinear, let us introduce a
notation for the linear and nonlinear electric eld as a function of induction:
18 3 Unidirectional Pulse Propagation Equations
E
L
(k) =
1

0
((k))
D(k) (3.9)
E
NL
(k) = E(k) E
L
(k) =
1

0
((k))
P
NL
(k) . (3.10)
It must be emphasized that these equations are not expressions of any physics, but merely
denitions of the two auxiliary electric vector elds. The reader should note that because the
right hand side of the rst denition contains the full vector D, the quantity E
L
(k) can not
be understood as the electric eld in the linear regime. Rather, the subscript is only meant
to indicate that it is the linear algorithm which is used to calculate the electric eld from the
induction.
Next we use this formal splitting of the electric eld intensity in the Maxwells equations
written in the Fourier (plane-wave expansion) domain:

t
_
D(k)
H(k)
_
=
_
ik H(k)
i
0
k E(k)
_

_
ik H(k)
i
0
k E
L
(k)
_
+
_
0
i
0
k E
NL
(k)
_
. (3.11)
Now lets act on this equation with P
+
to project out what we will call the forward propagating
component of the pulse.

t
_
D
f
(k)
H
f
(k)
_

t
P
+
_
D(k)
H(k)
_
= P
+
_
ik H(k)
i
0
k E
(
k)
_
P
+
_
ik H(k)
i
0
k E
L
(k)
_
+P
+
_
0
i
0
k E
NL
(k)
_
(3.12)
The rst term on the right-hand side is simple to evaluate,
P
+
_
ik H(k)
i
0
k E
L
(k)
_
= i(k)
_
D
f
(k)
H
f
(k)
_
(3.13)
because our projector is diagonal in the plane-wave basis which in turn is the eigen-basis of
Maxwells equations in a homogeneous medium. If this equality seems counterintuitive at rst,
one has to keep in mind that E
L
(k) does contain contributions from nonlinear interactions. A
straightforward evaluation of both sides of this equation proves that it holds.
The second term on the right-hand-side of 3.12 is transformed by applying the denition of
the projector and performing a straightforward calculation:
P
+
_
0
i
0
k E
NL
(k)
_
=
_
i
20(k)
k k E
NL
(k)
i
20
k E
NL
(k)
_
=
_
i(k)
2k
2
k k P
NL
(k)
ic
2
2((k))
k P
NL
(k)
_
=
_
i
2
(k)P
NL
(k)
i
2
(k)
kk
k
2
.P
NL
(k)
ic
2
2((k))
k P
NL
(k)
_
(3.14)
After inserting expressions (3.13,3.14) into Eqn. (3.12), the rst (electric induction) equation
becomes the time-propagated UPPE:

t
D
f
(k) = i(k)D
f
(k) +
i
2
(k)
_
P
NL
(k)
1
k
2
k k.P
NL
(k)
_
. (3.15)
3.2 Derivation of T-propagated Unidirectional Propagation Equation 19
Note that the magnetic eld equation doesnt provide anything new after the projection. One
can check that it can be obtained from the induction equation the same way as the magnetic
eld is obtained from induction in the plane wave.
In the same way one can obtain the UPPE equation for the backward propagating eld
component

t
D
b
(k) = +i(k)D
b
(k)
i
2
(k)
_
P
NL
(k)
1
k
2
k k.P
NL
(k)
_
. (3.16)
The two UPPE equations are coupled through their polarization terms, because P
NL
(k) im-
plicitly depends on the total electric eld. We emphasize that at this stage this equation pair
is still exact; we have not made any approximations so far. However, neither of the equations
is unidirectional yet. To get there, we have to adopt an approximation which will de-couple the
forward and backward equations. Specically, we will restrict ourselves to those regimes in which
it is sucient to use only the forward propagating electric eld to evaluate an approximation for
the total nonlinear polarization. In other words, we require
P
NL
(k) P
NL
(D
f
+D
b
, k) P
NL
(D
f
, k) (3.17)
which means that the nonlinear polarization is calculated from the forward propagating compo-
nent of D only, and the second UPPE for D
b
can be dropped. This means that the contribution to
the nonlinear response from the backward propagating eld must be negligible. This is somewhat
implicit denition of the one-directional propagation regime, and it is dicult to say a-priori if
in a concrete situation this approximation holds. On the other hand, in practice this is actually
the assumption which underlines much of the work in the laboratory; it is known or rather
assumed based on experience that the nonlinear focal regions do not produce strong backward
radiation.
While form the practical point of view we know that the above uni-directional approximation
is safe to use in practical simulations, the precise characterization of admissible solutions is
a dicult, and open mathematical problem. Maybe even more importantly, it is also safe to
assume that sooner or later the extreme nonlinear optics research will create scenarios in which
the backward-directed radiation will not be negligible. On the theoretic level, this question was
already posed for the nonlinear Helmholtz equation exhibiting a self-focusing collapse, and it was
demonstrated that both directed eld components become signicant in the collapsing solution.
Thus, the detailed understanding of how two UPPE equations couple in the presence of nonlinear
interaction is of great interest.
Thus we have obtained a propagation equation for pulsed beam-like solutions of Maxwells
equations with minimal approximations. The adjective minimal means that any uni-directional
propagation equations will, in some form use the approximation expressed in Eqn. (3.17). It also
means that other optical pulse propagation models can be obtained as approximations to the
t-propagated UPPE. We will discuss several such examples in the following Chapter.
20 3 Unidirectional Pulse Propagation Equations
3.3 Derivation of Z-propagated Unidirectional Propagation Equation
The UPPE version which is numerically evolved along a spatial direction, say z, is quite similar
to its t-propagated counterpart discussed in the previous Sections. Not only the mathematical
form of the equation is formally the same, but, as we will see shortly, also the approximation
which makes it possible to separate two counter-propagation directions is equivalent. However,
derivations are rather dierent, and so far there is no proof that the two approaches are math-
ematically equivalent. In this case, the derivation is based on expansion into linear propagation
modes of the system, and on a coupled mode equation that describes the evolution of a set of
complex-valued spectral amplitudes. Similarly as in the t-propagated version, we end up with
a formally exact pair of equations, one for each direction of the laser beam. Reduction into
a uni-directional system requires that nonlinear response does not generate strong backward
component.
It can be explicitly showed that while the pair of forward and backward evolution equation is
formally exact as long as they are solved together, the set of all UPPE solutions is smaller than
the set of all solutions to Maxwells equations. It is interesting to note that similar reductions
is not encountered for the tpropagation. This is one of the reason we believe that the two
approaches are in fact not equivalent.
The most important ingredient in unidirectional evolution equations is the separation, or
elimination of the wave-form portion which propagates in the opposite direction. To emphasize
this, and also to make full derivation easier to digest, we start with a simplied case of one-
dimensional Maxwells equations with a xed, linear polarization of the electric eld. This is
of course equivalent to a scalar one-dimensional wave propagation. This simple case is free of
notational complications while it still contains all important steps of the fully vectorial treatment.
In the subsequent Section, we start by derivation of the coupled mode equations which consti-
tute a common starting point for all specialized z-propagated Unidirectional Pulse Propagation
Equations.
3.3.1 A warm-up exercise: One-dimensional Maxwells equations
The one dimensional Maxwells equations reduced to linearly polarized electric eld can be
written as

z
H =
t
E +
t
P

z
E =
t
H (3.18)
where z is the optical axis and E and H are implicitly understood to be orthogonal to each
other and to z. This system has harmonic waves as solutions in the linear regime when P = 0:
E

(, z, t) = E
0
exp[it +ik()z] H

(, z, t) = H
0
exp[it +ik()z] > 0 = 1
The direction indicator selects forward and backward (or left and right) propagating waves.
We can use these as a basis in which to express our full, nonlinear solution as
E =

=1
_
dA

(, z)E

(, z, t) H =

=1
_
dA

(, z)H

(, z, t)
Here, A

(, z) are spectral amplitudes for which we have to nd an evolution equation. Taking

(3.18) and multiplying with the above basal solutions we get
E

z
H = E

t
E E

t
P
3.3 Derivation of Z-propagated Unidirectional Propagation Equation 21
H

z
E = H

t
H (3.19)
In these equations and in following formulas, we assume that the arguments of E

and H

are
, z, t.
We now add these two equations and collect terms that constitute full derivatives:

z
[E

H +H

E] =
t
[E

H +H

E] E

t
P (3.20)
The next step is to integrate over the whole domain perpendicular to the direction of propagation.
In this simplied case it means the t domain alone. We can see that after t integration, we obtain
boundary terms at past and future temporal innity. To get rid of these, we will restrict our
solution space to those functions which satisfy
lim
t
[E

H(z, t) +H

E(z, t)] = 0 (3.21)

This condition eliminates the middle term in (3.20), and the rest can be transformed as follows.
First, in the left-hand side we use the fact that the basis solutions are orthogonal, and after time
integration they eliminate the sum over and the modal index :
_
dt
z
[E

(, z, t)H +H

(, z, t)E] =
=
_
dt
z
E

(, z, t)

=1
_
dA

(, z)H

(, z, t)+
+
_
dt
z
H

(, z, t)

=1
_
dA

(, z)E

(, z, t) =
= 2E
0
H
0

z
A

(, z) (3.22)
On the right-hand side of (3.20), the polarization term yields essentially a Fourier transform

_
dtE

(, z, t)
t
P = i exp[ik()z]

P(, z)
Collecting both sides, we arrive at an evolution equation for spectral amplitudes:

z
A

(, z) =
i
2E
0
H
0
exp[ik()z]

P(, z) (3.23)
To obtain a corresponding equation for the electric eld, we recall that

(, z) = A

(, z) exp[ik()z]
and expressing its z derivatives using the evolution equation for the spectral amplitudes (3.47)
we have:

z

E

(, z) = ik()

E

(, z) +i

2E
0
H
0

P(, z)
This is a pair of equations for forward and backward ( = 1) propagating elds. The two are
coupled through the polarization which depends on their sum. Explicitly,
P(z, t) = P[E
+
(z, t) +E

(z, t)] .
where the concrete functional form of this dependence is not important for the present purpose,
but as an example one can consider the instantaneous Kerr nonlinearity for which the polarization
22 3 Unidirectional Pulse Propagation Equations
is simply proportional to the cube of the electric eld (for more examples, see introductory
sections of these notes):
P(z, t) = P[E
+
(z, t) +E

(z, t)] [E
+
(z, t) +E

(z, t)]
3
We should note that as soon as nonlinear polarization exhibits memory eects the above expres-
sion does not capture the most general form of cubic nonlinearity in isotropic media.
The above derivation illustrates the scheme we will use in the next Section to derive the
general, fully vectorial Unidirectional Pulse Propagation Equation. The important point to note
here is that in the course of derivation we had to restrict the space of our solutions by the
condition (3.21). This means that the UPPE solution set is smaller than the set of all Maxwells
equations solutions. However, within the admissible subspace, unidirectional equations are exact,
as long as we can solve them simultaneously while always calculating the nonlinear polarization
term from the total eld.
So let us have a look at the assumption (3.21). What it says is that for every xed location z
along the laser beam axis, if we wait for a suciently long time, the eld at this location will tend
to zero. In other words, light energy will dissipate into positive and negative z-innities. This is
certainly very benign condition in the context of pulse propagation, because this is exactly what
happens to localized pulsed wavepackets - they eventually disappear from our sight. On the other
hand, one can easily think of an example which does not satisfy (3.21): a pulse trapped within a
perfect cavity lled with a loss-less medium. Obviously, its eld will continue to exhibit in-cavity
values which will not tend to zero at long times (they may not have such limits, of course). Of
course, a perfect cavity is an extreme idealization, as is the notion of a loss-less medium which
must be also dispersion-less and therefore in fact a vacuum. Thus we see that while there is a
condition on the UPPE solutions, it is not practically restricting. Situation will turn out to be
completely analogous in the fully three-dimensional case discussed in the next Section.
3.3.2 Maxwell equations as a boundary value problem for pulsed beam propagation
As a rst step in derivation of various versions of UPPE, we derive an exact coupled-modes
system of equations. Electromagnetic elds of a light pulse propagating along the z axis can be
expanded into modal contributions that reect the geometry of the waveguide (we can consider
a homogeneous medium as a special case of the latter).
E(x, y, z, t) =

m,
A
m
(, z)E
m
(, x, y)e
im()zit
H(x, y, z, t) =

m,
A
m
(, z)H
m
(, x, y)e
im()zit
(3.24)
Here, m labels all transverse modes, and an initial condition A
m
(, z = 0) is supposed to be given
or calculated from the known eld values at z = 0. Note that the above expansion is valid for the
transverse components only, and that the modal index m is a short hand for all quantities which
are required to specify a unique propagation mode. For example, in a homogeneous bulk medium,
the eigen modes are the well known plane waves, and the index m represents polarization, two
transverse wavenumbers, and a binary value selecting the forward or backward direction of
propagation.
To save space and reduce clutter, the following short-hand notation will be used below
E
m
E
m
(, x, y)e
im()zit
H
m
H
m
(, x, y)e
im()zit
. (3.25)
3.3 Derivation of Z-propagated Unidirectional Propagation Equation 23
We consider a non-magnetic medium ( =
0
) with a linear permittivity (, x, y) that
doesnt depend on the propagation coordinate z which coincides with what we consider forward
and backward propagation direction. Note that the permittivity or, equivalently, the index of
refraction may depend on the transverse coordinates x, y. That would be the case for example in
a micro-structured waveguide, or in a hollow-core ber or capillary; at this rst stage, we want to
treat bulk media and ber-like geometries together. Later we will branch and derive separate,
specialized equations for waveguides and for bulk media.
The starting point of our derivations is of course the Maxwells equations:
j +
t
P +
0

t
E = H

t
H = E (3.26)
where the star represents a convolution so that the term is a short hand for

t
E =
0

t
_

0
d()E(t )
Here () is the temporal representation of frequency-dependent permittivity (). The same
notation will be used for both quantities, and their arguments will serve to distinguish them
where needed.
As a rst step, we scalar-multiply Maxwells equations by complex conjugate modal elds
E

m
.(j +
t
P) +
0
E

m
.
t
E = E

m
.H

0
H

m
.
t
H = H

m
.E . (3.27)
Using the formula b.( a) = .(a b) + a.( b), we transform both right-hand sides to
obtain
E

m
.(j +
t
P) +
0
E

m
.
t
E = .[H E

m
] +H.[E

m
]

0
H

m
.
t
H = .[E H

m
] +E.[H

m
] . (3.28)
Now we can take advantage of the fact the modal elds themselves satisfy the Maxwells equations
E

m
=
0

t
H

m
H

m
=
0

t
E

m
, (3.29)
and therefore the previous equations can be written as
E

m
.(j +
t
P) +
0
E

m
.
t
E = .[H E

m
]
0
H.
t
H

0
H

m
.
t
H = .[E H

m
] +
0
E.
t
E

m
(3.30)
Next, we subtract the two equations and collect terms that constitute full time derivatives
E

m
.(j +
t
P) +
t
[
0
E

m
. E] = .[H E

m
]
t
[
0
H

m
.H] .[E H

m
] . (3.31)
Now we integrate over the whole xyt domain. Note that all terms except the rst and
z
, which
is implicit in the . operator, are derivatives that give rise to surface terms after integration
over x, y, t. These surface terms are supposed to vanish far from the axis of the laser beam, as
well as in past and future temporal innities. Intuitively, admissible solutions include spatially
and temporally localized pulse-like solutions, which we are interested in. As a consequence, the
only surviving derivatives will be
z
:
_
E

m
.(j +
t
P)dxdydt =
z
_
z.[H E

m
]dxdydt
z
_
z.[E H

m
]dxdydt (3.32)
24 3 Unidirectional Pulse Propagation Equations
Here and in what follows, t integrations are understood like this:
_
dt
1
T
_
+T/2
T/2
dt where
T is a large normalization volume, and integrals over x, y are understood in a similar way.
This will give us a convenient way to obtain the correct normalization and translate it into
implementation which will be in terms of numerical Fourier transforms.
Because only transverse eld components enter the above equation, we can use our modal ex-
pansion here (recall that those are only valid for transverse vector components):
_
E

m
.(j +
t
P)dxdydt =

z
_
z.[

n,
A
n
(, z)H
n
() E

m
()]e
in()zit
e
im()z+it
dxdydt

z
_
z.[

n,
A
n
(, z)E
n
() H

m
()]e
in()zit
e
im()z+it
dxdydt .
(3.33)
Integration over time gives a Kronecker delta between angular frequencies,

, which in turn
reduces the sum over :
_
E

m
.(j +
t
P)dxdydt =

z
_
z.[

n
A
n
(, z)H
n
(, x, y) E

m
(, x, y)]e
in()z
e
im()z
dxdy

z
_
z.[

n
A
n
(, z)E
n
(, x, y) H

m
(, x, y)]e
in()z
e
im()z
dxdy .
(3.34)
Collecting like terms results in an equation
_
E

m
.(j +
t
P)dxdydt =
z

n
A
n
(, z)e
in()z
e
im()z

_
z.[H
n
(, x, y) E

m
(, x, y) E
n
(, x, y) H

m
(, x, y)]dxdy . (3.35)
At this point we are going to use a general property of electromagnetic modal elds which
constitute an orthogonal basis: all radiative waveforms can be expressed as their linear combi-
nations. To calculate such expansions, one can utilize the following orthogonality relation
_
z.[E
m
H

n
H
m
E

n
] dxdy = 2
m,n
N
m
() (3.36)
Here N
m
() is a normalization constant, whose explicit functional form has to be derived for
each concrete set of modes.
Orthogonality of modes is used to reduce the sum over n in (3.35)
_
E

m
.(j +
t
P)dxdydt =
z

n
A
n
(, z)e
in()z
e
im()z
2
m,n
N
m
() , (3.37)
and we nally obtain an evolution equation for our expansion coecients:

z
A
m
(, z) =
1
2N
m
() XY T
_
+T/2
T/2
dt
_
+Y/2
Y/2
dy
_
+X/2
X/2
dx
e
im()z+it
E

m
(, x, y).[j(x, y, t) +
t
P(x, y, t)] (3.38)
This is the starting point for the z-propagated unidirectional equations. In the following Sections,
we will specialize this to the case of bulk media and then continue on with waveguides.
3.3 Derivation of Z-propagated Unidirectional Propagation Equation 25
3.3.3 Z-propagated UPPE for homogeneous media: General case
In this section, Eq. (3.38) is specialized for the case of a homogeneous medium. This is done by
inserting explicit expressions for a given family of modal elds.
In bulk media, eld modes are plane waves, They can be labeled by transverse wavenumbers
k
x
, k
y
, by a polarization index s = 1, 2, and by a sign signifying the direction of propagation
along the z direction. So the index m, which we used to label modes in the preceding Section, is
actually a list:
m k
x
, k
y
, s, . (3.39)
The following notation will be used for the frequency- and wavenumber-dependent propagation
constant of a plane wave characterized by its angular frequency :

kx,ky,s,
() k
z
(, k
x
, k
y
) =
_

2
()/c
2
k
2
x
k
2
y
, (3.40)
Electric and magnetic amplitudes in plane waves are mutually determined by the Maxwells
equations. We x them as
E
kx,ky,s,
= e
s
exp[ik
x
x +ik
y
y ik
z
(, k
x
, k
y
)] (3.41)
H
kx,ky,s,
=
1

k E
kx,ky,,s,
. (3.42)
We can choose one of the amplitudes (the electric one in this case) arbitrarily, and the choice will
be naturally reected in the modal normalization constants, but is otherwise inconsequential.
The polarization of modal elds is determined by polarization vectors e
s=1,2
which are of unit
length and are normal to the wave-vector
k = {k
x
, k
y
, k
z

_

2
()/c
2
k
2
x
k
2
y
} . (3.43)
Using the above formulas, it is straightforward to calculate the modal normalization constant
2N
kx,ky,s,
() =
_
z.[E
m
H

m
H
m
E

m
] dxdy =
2z.[e
s
(k e
s
)]
1

= 2k
z
(, k
x
, k
z
)
1

(3.44)
N
kx,ky,s,
() =
k
z
(, k
x
, k
z
)

. (3.45)
Now we can insert expressions for modal elds and the corresponding normalization constant
into coupled mode equation Eq. (3.38) to obtain

z
A
kx,ky,s,
(, z) =

0
2k
z
e
ikzz
_
dxdydt
L
x
L
y
T
e
i(tkxxkyy)

e
s
.[j(x, y, z, t) +
t
P(x, y, z, t)] (3.46)
The above integral is nothing but a spatial and temporal Fourier transform, so one can write it
down in the spectral domain as

z
A
kx,ky,s,+
(, z) =

2
0
c
2
k
z
e
ikzz
e
s
.[iP
kx,ky
(, z) j
kx,ky
(, z)] . (3.47)
26 3 Unidirectional Pulse Propagation Equations
This is the propagation equation that will actually be solved numerically because it is cast in
terms of the slowest variables our propagation problem has. We can see that the only source of
evolution in spectral amplitudes is nonlinearity.
For those who prefer to see evolution equations for electric elds proper, we express the above
in terms of the electric eld rather than in terms of modal expansion coecients. From a modal
expansion, the transverse part of the electric eld is
E

kx,ky,+
(, z) =

s=1,2
e

s
A
kx,ky,s,+
(, z)e
ikz(kx,ky,)z
, (3.48)
and therefore its z derivative reads

z
E

kx,ky,+
(, z) = ik
z
(k
x
, k
y
, )E

kx,ky,+
(, z)+

s=1,2
e

s

z
A
kx,ky,s,+
(, z)e
ikz(kx,ky,)z
(3.49)
Using Eq. (3.47), we obtain the full-vectorial UPPE for a homogeneous medium:

z
E

kx,ky,+
(, z) = +ik
z
E

kx,ky,+
(, z)+

s=1,2
e

s
e
s
.[
i
2
2
0
c
2
k
z
P
kx,ky
(, z)

2
0
c
2
k
z
j
kx,ky
(, z)] (3.50)
This is an exact system of equations that describes evolution of modal amplitudes along the z-
axis for the forward propagating eld. Of course, an analogous equation holds for the backward
propagating component:

z
E

kx,ky,
(, z) = ik
z
E

kx,ky,+
(, z)

s=1,2
e

s
e
s
.[
i
2
2
0
c
2
k
z
P
kx,ky
(, z)

2
0
c
2
k
z
j
kx,ky
(, z)] (3.51)
This pair of equation is exact and completely analogous to the pair of z-propagated equations
discussed in the previous Section. Because the nonlinear polarization in these equations results
as a response to the complete electric eld, they cant be used to calculate the forward eld in iso-
lation (i.e. without its backward propagating counterpart). The equation becomes unidirectional
only when the following approximation can be adopted:
P(E), j(E) P(E
f
), j(E
f
) (3.52)
In other words, to obtain a closed system which is restricted to a single direction and which can
be solved numerically, we must require that nonlinear polarization can be calculated accurately
from only the forward propagating eld. This means that UPPE is only applicable when the
back-reected portion of the eld is so small that its contribution to the nonlinearity can be
neglected.
3.3.4 Z-propagated UPPE: Simplied, practical version
Eq. 3.50, with nonlinear polarization approximated by Eq. 3.52 can easily become a rather large
system to solve numerically. This is especially true for experiments with wide-beam multi TW
lasers. Fortunately, in most cases transverse dimensions of resulting structures remain relatively
large in comparison to wavelength, and further approximations are possible. For example in
3.3 Derivation of Z-propagated Unidirectional Propagation Equation 27
femtosecond lamentation in gases, the typical diameter of the lament core is about hundred
micron which dimension is large in comparison with the laser wavelength. Consequently, the
longitudinal vector component of the electric eld is much smaller than the transverse (x, y)
components, and can be neglected in calculation of the nonlinear medium response. It thus
makes sense to take advantage of this fact to obtain a simpler equation.
Concretely, one can neglect the z components of the eld and polarization vectors. In such a
situation the sum over polarization vectors reduces approximately to unity

s=1,2
e

s
e
s
1 . (3.53)
To see this, it is enough to recall that the left-hand side constitutes a projector onto the wave-
vector (recall that these vectors are mutually orthogonal). As the wave-vector is pointing in the
direction almost parallel to the beam axis, it is also approximately a unity operator in the
vector subspace spanned by x, y.
Replacing the transverse projection by unity, the full UPPE simplies into an equation for
transverse component(s)

z
E
kx,ky
(, z) = ik
z
E
kx,ky
(, z) +
i
2
2
0
c
2
k
z
P
kx,ky
(, z)

2
0
c
2
k
z
j
kx,ky
(, z) ,
k
z
=
_

2
()/c
2
k
2
x
k
2
y
. (3.54)
This is the most useful form for practical calculation, and is therefore called simply UPPE in the
following. While we write it as a scalar equation, it should be understood that it is in general
coupled to its counterpart governing the other polarization. The two polarization components
of the electric eld both contribute to the nonlinear polarization and this is how they become
mutually coupled.
28 3 Unidirectional Pulse Propagation Equations
3.4 Other propagation models as approximations of UPPE
The previous Section showed the the Unidirectional Pulse Propagation Equations can be rigor-
ously derived under very general assumptions. They are based on a single approximation, an that
is that nonlinear interaction between light and matter occurs in such a regime that the medium
response to the eld can be calculated with sucient accuracy from only the forward-going eld
component. Because this is how all one-way pulse propagation equation include nonlinearity, one
can expect that other types of equations can be derived from UPPE. We want to demonstrate
this in what follows. We will in fact show a universal scheme to derive all other propagation
models.
The rst purpose of this Section is to illustrate that what we have identied as the Maxwell
reduction problem in Introduction, has been receiving a lot of attention in the literature for many
years. Researchers proposed a number of various solutions and new results continue to appear
in the literature.
The motivation for such research is obviously dierent in dierent research communities.
The optics community is mostly interested in practical wave propagation solvers as tools to
utilize in experimental and theoretic work. But there is also an applied mathematics community
which studies these equations as weakly nonlinear systems, not necessarily restricted to the
optical contexts. The goal is a rigorous analysis of resulting equations rather than their practical
application. These divergent interests are, in our opinion, one of the reasons why the Maxwell
reduction problem still enjoys attention of researchers.
The second purpose of the Section is to bring various derivations and dierent version of
propagation equations under one roof. We present a unifying view as a means to elucidate ex-
actly what approximations went into original derivations. Reader might expect that assumptions
behind an equation should be be clearly established, but it is actually not the case; derivations
often suer from lack of control over the neglected terms or eects. This is why we consider useful
to discuss some successful propagation models along with a couple of examples when the results
turn out to be less than solid. This should give our reader tools to evaluate the plethora of
propagation models scattered over the literature of last ten years.
3.4.1 General method to obtain pulse propagation models from UPPE
Several types of unidirectional propagation equation appear frequently in the literature on non-
linear optics. The most important examples are Non-Linear Schrodinger (NLS) equation [?],
Nonlinear Envelope Equation [?] (NEE), the First-Order Propagation equation [?] (FOP), For-
ward Maxwells equation [?] (FME), and several other equations that are closely related to these.
In this section, we explain a unied method which will be subsequently used to derive several
of the light-pulse propagation equations. The main benet of re-deriving known equations from
a common starting point while using the same method, is that it allows us to compare physical
assumptions and approximation underlying dierent equations. It also reveals relations between
equations which may not be obvious either because of their apparently dierent form, or because
of dierent methods used in the original derivations.
It is instructive to break the derivation procedure into several steps. As a rst step, we adopt a
scalar, one-component approximation and write the Unidirectional Pulse Propagation Equation
in the following form:

z
E
kx,ky
(, z) = iKE
kx,ky
(, z) +iQP
kx,ky
(, z) (3.55)
where
K(k
x
, k
y
, ) =
_

2
()/c
2
k
2
x
k
2
y
(3.56)
3.4 Other propagation models as approximations of UPPE 29
is the linear eld propagator in the spectral representation, and
Q(k
x
, k
y
, ) =

2
2
0
c
2
_

2
()/c
2
k
2
x
k
2
y
(3.57)
will be called nonlinear coupling term. In most cases, the concrete form of the nonlinear polar-
ization P is unimportant, and we will assume that it can be specied in terms of an algorithm
which accepts the electric eld (in general as a function of time), at a given spatial location as
its input.
Let us note that this one-component, or scalar representation can be still understood as a
description of a single polarization in a coupled system describing two transverse vector compo-
nents of an optical eld. While each equation appears scalar, the two become coupled through
the polarization term, for example due to the nonlinear birefringence. These coupling eects can
play a role even if a laser beam is much wider than the light wavelength, and can lead to a
rich polarization dynamics within femtosecond laments. What is neglected at this step is the
longitudinal part of the electric eld. That only becomes important when the beam focuses to
a size comparable with wavelength. However, in the naturally occurring laments such extreme
focusing is never achieved, because self-focusing collapse is always arrested either by chromatic
dispersion or by the free-electron induced de-focusing. Thus, for many practical purposes, the
above representation is suciently rich and accurate.
In the second derivation step, we replace couplings K and Q by suitable approximations.
In most cases, they are closely related to Taylor expansions in frequency and in transverse
wavenumbers. It is at this stage that articial parameters are introduced into a propagation
model (a typical example is the reference frequency). It is important to keep in mind that
knowledge extracted from simulations should not depend on such degrees of freedom. In this
respect, we aim to show that the improvements introduced into pulse evolution equations can
be often viewed as corrections which (partially) restore the invariance of a model with respect
to these free-will choices.
Having chosen our approximations for the linear and nonlinear coupling, we are still in the
real-eld representation. However, most of the published models are written using envelopes.
Thus, in the next step, we obtain envelope equations. To do this, one can expresses the eld in
terms of an envelope by factoring out the carrier wave at a chosen reference angular frequency

R
with the corresponding wave-vector k
R
= K(0, 0,
R
):
E(x, y, z, t) = A(x, y, z, t)e
i(kRzRt)
(3.58)
A similar factorization is of course introduced for the nonlinear polarization P(x, y, z, t) as well.
The nal step consist in transforming the equation from the spectral- to the real-space rep-
resentation. Mathematically, this is nothing but a Fourier transform, and the following standard
replacement rules for dierential operators provide quick and easy way to do this transformation:
(
R
) i
t
ik
x

x
ik
y

y

z
ik(
R
) +
z
(3.59)
Finally, in most cases we also transform to a frame moving with a suitable group velocity such
that the pulse remains close to the center of the computational domain.
3.4.2 Derivation of Non-Linear Schrodinger Equation from UPPE
The Nonlinear Schrodinger Equation (NLS) is a prototype propagation equation ubiquitous in
the nonlinear optics, and its importance can hardly be overstated. It is therefore only natural
30 3 Unidirectional Pulse Propagation Equations
to start our model discussions with this equation. It is also the simplest case to illustrate the
derivation method outlined above, and an opportunity to point out some important issues which
are also relevant in other pulse propagation models.
Following the general procedure, we replace the K and Q coecients with their appropri-
ate approximations. We denote by k
R
= k(
R
) the reference wavenumber corresponding to an
arbitrarily chosen reference frequency
R
, and take
K(k
x
, k
y
, ) =
_

2
()/c
2
k
2
x
k
2
y
k
R
+v
1
g
(
R
) +
k

2
(
R
)
2

1
2k
R
(k
2
x
+k
2
y
) (3.60)
This is a second-order Taylor expansion in
R
and in k
x
, k
y
. It should be noted that the
parameter
R
has nothing to do with the solution of the equation and in particular it does not
need to be equal to the central frequency of the simulated optical pulse. In fact, below we give
a very good reason why it should not, and why simulation runs obtained for dierent choices
should be compared to asses their robustness. The point is that this is an example when an
articial parameter, originally absent from Maxwells equations, makes its way into a theoretical
model. Such a parameter must be regarded as a gauge variable from whose concrete choice
all physical results must be independent. This is never the case, strictly speaking. In practice,
we choose a value and this choice has a measurable eect on the numerical solution: this is a
consequence of the equation not being strictly invariant. A blessing in disguise is that both the
accuracy of the numerical solution, as well as the robustness of the propagation equation itself,
can and should be measured in how big or small these unwanted parameter-dependencies are.
Clearly, the NLS equation will be manifestly dependent of our choice for the reference frequency
in the above approximation of K, and its solutions will inherit the same. In the next subsections,
it will become evident that the so called correction terms introduced into the NLS equation can
be actually viewed as corrections which eliminate at least partly the dependence of the equation
on
R
.
Now, let us return to our derivation. In the nonlinear coupling coecient, we neglect all
variable dependencies and replace it with its value at the reference frequency and zero transverse
wavenumbers:
Q(k
x
, k
y
, ) =

2
2
0
c
2
_

2
()/c
2
k
2
x
k
2
y

R
2
0
n(
R
)c
(3.61)
In the context of the Nonlinear Schrodinger Equation, the nonlinear polarization term only
accounts for the instantaneous optical Kerr eect. The polarization envelope is usually expressed
in terms of the light intensity I and of the eld-envelope A:
P = 2
0
n(
R
)n
2
IA (3.62)
Inserting this into (3.55,3.58) we obtain

z
A = +iv
1
g
(
R
)A +
ik

2
(
R
)
2
A
i
2k
R
(k
2
x
+k
2
y
)A+
i
R
c
n
2
IA (3.63)
The NLS is usually written in units suitable for the given geometry, be it a ber or a bulk
medium. In the latter, it is customary to normalize the envelope amplitude such that |A|
2
= I.
Using rules (3.59) we nally obtain the NLS equation:
(
z
+v
1
g

t
)A =
i
2k
R

A
ik

2

tt
A+
i
R
c
n
2
|A|
2
A (3.64)
3.4 Other propagation models as approximations of UPPE 31
The above derivation shows explicitly what approximations need to be adopted to obtain NLS:
Approximating K to second order in frequency and transverse wavenumber amounts to the
paraxial, and quasi-monochromatic approximations for the linear wave propagation. The ap-
proximation in the nonlinear coupling Q also requires a narrow spectrum in order to be able to
represent Q by a constant.
Despite the underlying assumption of a narrow spectrum, NLS is in practice often pushed
into a regime with signicant spectral broadening. This immediately invites a question: which
portion of the simulated spectrum can be correct?
One characteristic feature of NLS and of other envelope equations is the reference frequency.
It is usually chosen equal to the central frequency of the initial pulse, and beginners implementa-
tions often hard-code this particular choice in the program. This is not necessary, and not even a
good practice. It is useful to keep in mind that
R
is a free parameter, and as such it provides an
opportunity to test the robustness of obtained simulation results. Comparative runs with small
variations in reference frequency will readily reveal which parts of simulated spectra react to
these changes - these are artifacts and only the stable, robust portions of the solution can be
accepted. This is naturally applicable to any pulse propagation model that contains articial
parameters not present in the original Maxwells equations.
3.4.3 Nonlinear Envelope Equation
Nonlinear Envelope Equation [?] is a paraxial equation with some additional approximations
related to chromatic dispersion.
Once again, we follow the general procedure and approximate the linear propagator by its
paraxial version:
K(k
x
, k
y
, ) =
_

2
()/c
2
k
2
x
k
2
y
+k()
c
2n
b
(
R
)
(k
2
x
+k
2
y
) (3.65)
This is the second-order (paraxial) Taylor expansion in transverse wavenumbers with an addi-
tional approximation. Namely, we replaced n
b
() n
b
(
R
) in the denominator of the diraction
term, and thus partly neglected chromatic dispersion.
Further, the rst term in the above approximation, which is an exact propagation constant
for a plane wave propagating along the z axis, is re-expressed as a sum of its two lowest-order
Taylor expansion terms plus the rest:
k() = k(
R
) +v
1
g
(
R
) +D(
R
) (3.66)
where
D(
R
) =

n=2
_

n
k

n
_
=R
(
R
)
n
n!
(3.67)
This is formally exact and can be practically implemented in the spectral domain without further
approximations, but sometimes a nite number of series expansion terms is used to t the linear
chromatic dispersion of a medium or of a waveguide.
Next, we approximate the nonlinear coupling term. Unlike in NLS, we preserve the frequency
dependence exactly, but neglect the transverse wave-number dependence:
Q(k
x
, k
y
, )

2
2
0
c
2
_

2
()/c
2
k
2
x
k
2
y

(
R
) +
R
2
0
cn(
R
)
(3.68)
Here, as in the free propagation term, we neglect the chromatic dispersion of the background
index of refraction.
32 3 Unidirectional Pulse Propagation Equations
After inserting the above approximations for K and Q into the original UPPE, we obtain

z
A = iv
1
g
(
R
)A+iD(
R
)A

ic
2
R
n(
R
)
(1 +

R

R
)
1
(k
2
x
+k
2
y
)A
+
i
R
2
0
cn(
R
)
(1 +

R

R
)P (3.69)
Finally, transforming into the real-space representation, we arrive at NEE

z
A+v
1
g

t
A = iD(i
t
)A+
i
2k
R
(1 +
i

t
)
1

A+
ik
R
2
0
n
2
b
(
R
)
(1 +
i

t
)P (3.70)
Thus, approximations underlying the NEE are paraxiality both in the free propagator and in the
nonlinear coupling, and a small error in the chromatic dispersion introduced when the background
index of refraction is replaced by a constant, frequency independent value in both the spatio-
temporal correction term and in the nonlinear coupling term. Note that the latter approximations
are usually not serious at all.
An important point is that NEE should not be implemented in the form it was originally
introduced. Handling higher order time derivatives any other way than spectral would necessarily
cause additional numerical dispersion which may spoil linear propagation properties. However,
if implemented in a spectral representation, NEE transforms into another propagation model
discussed latter in these notes. Readers should be aware that authors in the literature sometimes
claim to use NEE even if, in the above sense, they are not.
As in all envelope equations a reference frequency and a reference wave-number appear in
the NEE. They are normally chosen equal to the central frequency and wave-number of the
input pulse. But as we pointed out in connection to NLS, one has to keep in mind that these
quantities are articial and arbitrary gauge parameters which do not appear in the Maxwells
equations. Consequently, numerical solutions should not depend on how the reference is chosen.
In other words a propagation equations should be reference-frequency-invariant. While NLS is
manifestly dependent on the reference choice, NEE is nearly invariant although
R
appears in it
several times. To appreciate this consider the spatio-temporal focusing correction term (operator)

1
R
(1 +
1
R

t
)
1
It seems to depend on the reference
R
, but it is in fact proportional to
1
as
long as it is implemented in the spectral domain which allows to include all orders of the series
expansion. Note that this (approximate) invariance is only achieved in the innite order, and can
be properly implemented only in the spectral representation. Truncating operators at a nite
order of series expansion breaks the invariance and causes undesirable artifacts. We illustrate
this point in the following example.
3.4.4 Partially corrected NLS
The Partially Corrected NLS (PC-NLS) equation can be viewed as a simplication of NEE. It
is derived from the UPPE in the same way, with one additional step. Namely, the following rst
order series expansion is applied in the correction term of the free propagator in Eqn.(3.69):
(1 +

R

R
)
1
(1

R

R
) (3.71)
This step is meant to make it easy to implement a numerical solver in the real space, as it results
in the equation that only contains simple dierential operators in the real-space representation:

z
A+v
1
g

t
A = iD(i
t
)A+
i
2k
R
(1
i

t
)

A+
ik
R
2
0
n
2
b
(
R
)
(1 +
i

t
)P (3.72)
3.4 Other propagation models as approximations of UPPE 33
While it may seem that the Partially Corrected NLS is essentially NEE with a little more
approximation, this equation is not to be recommended. Because of the arbitrary truncation of
an innite series, the dispersion properties of the linear part of this equation are unphysical.
While the PC-NLS provides better-than-NLS approximation around the reference frequency

R
, its dispersion properties become rather pathological around 2
R
where its diraction
term changes sign as a consequence of the truncated correction factor. Artifacts in the angular
distribution of the spectrum can be observed at high frequencies beyond 2
R
. Consequently,
this equation is only applicable in the same regime as the NLS, namely when the spectrum of
the pulse remains relatively narrow.
3.4.5 First-order propagation equation
The previous three examples represented eld envelope propagation equations. Next we discuss a
non-envelope, First-Order Propagation equation (FOP), introduced by Geissler et al. [?]. Though
it resolves the carrier wave, this equation is equivalent to NEE from the point of view of the
approximations required for its derivation as we shall see shortly. Following the original authors,
we neglect linear chromatic dispersion in order to obtain the same equation as Geissler et al.
In Eqn.(3.54) , we approximate
K(k
x
, k
y
, )
_

2
/c
2
k
2
x
k
2
y


c

c
2
(k
2
x
+k
2
y
) (3.73)
Note that the corresponding approximation made in NEE reduces to the present case if there is
no linear chromatic dispersion (vacuum).
Similarly,
Q(k
x
, k
y
, )

2
2
0
c
2
_

2
/c
2
k
2
x
k
2
y

2
0
c
(3.74)
is the same approximation as the one in NEE only with vacuum in the role of the linear medium.
Thus, the propagation equation obtained with these expressions is

z
E
kx,ky,
=
i
c
E
kx,ky,

ic
2
(k
2
x
+k
2
y
)E
kx,ky,
+
i
2
0
c
P
kx,ky,
(3.75)
which is equivalent to Eqn. (2) of Ref. [?].
When transforming into the real-space domain,
1
gives rise to an integral over time, and
we arrive at the FOP equation
(
z
+
1
c

t
)E(r

, t) =
c
2

_
t

dE(r

, )
1
2
0
c

t
P(r

, t) (3.76)
Comparing the derivation steps for NEE and FOP it becomes clear that despite of rather dierent
ways they were originally derived, these two equations become equivalent in a non-dispersive
medium.
3.4.6 Forward Maxwell Equation
Another non-envelope equation free of any reference frequency is the Forward Maxwell Equation,
introduced by Husakou and Herrmann [?]. Although it was written in a vector form, it was derived
from the wave equation with a neglected .E term. That is why vector nature of light is not
captured completely correctly. The result is a two-component equation rather than a true
34 3 Unidirectional Pulse Propagation Equations
vectorial one. The polarization scrambling terms only come from the nonlinear polarization term
in the equation. One can therefore derive the equation component-by-component in the same
way as previous propagation equations.
The linear propagator and the nonlinear coupling approximations are similar to those used
for NEE:
K(k
x
, k
y
, ) =
_

2
()/c
2
k
2
x
k
2
y
k()
c
2n
b
()
(k
2
x
+k
2
y
) (3.77)
and
Q(k
x
, k
y
, )

2
2
0
c
2
_

2
()/c
2
k
2
x
k
2
y

2
0
cn
b
()
(3.78)
The only dierence from NEE is that here chromatic dispersion of the index of refraction is
correctly preserved.
The resulting equation is then obtained by transforming to the real space in transverse
coordinates only, keeping the spectral frequency-time representation:

z
E(x, y, , z) = ik()E(x, y, , z) +
i
2k()

E(x, y, , z) +
i
0
c
2n
b
()
P(x, y, , z) (3.79)
This equation is equivalent to FME Eqn.(2) of Ref. [?]. A small dierence is that we have not
transformed the equation to the coordinate frame moving with the vacuum light velocity. The
reason is that in a strongly dispersive medium it is necessary to use the frame that moves with
a suitable group velocity.
Formulas (3.78) and (3.65,3.68) show that there is very little dierence between FME and
NEE. Because of the operator series expansion, the only acceptable way to solve NEE is in
the spectral domain. But there, the correct frequency dependence of the background index of
refraction can be easily taken into account - but that is when NEE becomes FME.
3.4.7 Short-Pulse Equation
The so called Short-Pulse Equation (SPE) was originally proposed by Alterman and Rausch in
2000, and re-derived by Schafer and Wayne four years later. Here we make a departure from the
scheme of the preceding sections. Instead we comment on the way the equation was obtained
originally.
One aim of this section is to illustrate how the applied mathematics community likes to treat
the Maxwell reduction problem. The Short Pulse Equation has beautiful mathematical properties
and is in fact an integrable equation. It is understandably very popular in mathematics literature.
From the physicist point of view, this equation has an extremely limited applicability, because
the Short Pulse Equation is not suitable to model short pulses!
The origin of this drawback is in how the equation was derived. An important ingredient is
selection of a medium that leads to a suitable wave equation. This is achieved by choosing a
special form of frequency dependent index of refraction, which in turn creates desired terms in
the wave equation and thus transforms it into a form that eventually leads to SPE. This results in
inherent restrictions, and the result may not be as generally applicable as one would wish, or as
the authors claimed. Encoding specic medium properties directly into SPE means that media
which do not satisfy this specic constraint can not be simulated at all. It can be shown that there
is no possibility for this equation to model a medium which exhibits a so-called zero dispersion
wavelength. That is a wavelength at which second-order chromatic dispersion vanishes, and which
is well-known to play a crucial role in dynamics of ultra-short pulses in bers. Roughly speaking,
much of the supercontinuum generation in bers depends on the existence of zero-dispersion
3.4 Other propagation models as approximations of UPPE 35
wavelength. So it occurs that despite the fact that SPE was designed specically for bers, it is
unable to describe supercontinuum generation in ultrashort pulses in bers. The lesson here is
that one of necessary attributes of a realistic model is an exact treatment of linear propagation.
3.4.8 Maxwells equations versus Wave equation in pulse-propagation
This subsection is devoted to the relation between the Unidirectional Pulse Propagation Equa-
tions and the wave equation. Naively, one may think that because the wave equation follows
directly from the Maxwells equations, we should be able to derive from it all pulse evolution.
Several previously published results discussed in previous subsections were obtained just like
that. One motivation for what follows is to demonstrate that such an approach has an inherent
problem, and that is the fact that a one-way equation must not be equivalent to the wave equa-
tion! To start, we recall that the total eld is a sum of left- and right-propagating components:
E(r, t, z) = E
+
(r, t, z) +E

(r, t, z) (3.80)
where each is expressed through Fourier amplitudes in transverse space (r) and time:
E
+
(r, t, z) =
_

E
+
(k, , z)e
it+ik.r
dk
2
d E

(r, t, z) =
_

E

(k, , z)e
it+ik.r
dk
2
d
(3.81)
The following is the total nonlinear polarization. Its functional dependence on the electric eld
is irrelevant for now; it may include the optical Kerr eect (i.e. self-focusing), the defocusing
induced by free electron, the nonlinear losses due to multiphoton and avalanche ionization, and
other eects...
P(r, t, z) =
_

P(k, , z)e
it+ik.r
dk
2
d (3.82)
(We assume that the linear portion of polarization has been included in the mediums frequency-
dependent permittivity (), and therefore it does not appear explicitly in our equations. ) The
nonlinear polarization serves as a source for one-way propagating elds:

z

E
+
(k, , z) = +ik
z
(, k)

E
+
(k, , z) +
i
2
2
0
c
2
k
z
(, k)

P(k, , z) (3.83)

z

E

(k, , z) = ik
z
(, k)

E

(k, , z)
i
2
2
0
c
2
k
z
(, k)

P(k, , z) (3.84)
where k
z
(, k) =
_

2
()
c
2
k
2
is a short-hand for the plan-wave propagation constant.
It is shown next that the unidirectional propagation equations give together a total eld which
satises the wave equation, while the single forward (or backward) propagating component does
not. First, let us evaluate the
zz
term in two
z
steps:

z
E
+
(r, t, z) =
_

z

E
+
(k, , z)e
it+ik.r
dk
2
d (3.85)
after using the propagation equation to express the z-derivative in the spectral representation
one gets

z
E
+
(r, t, z) =
_ _
+ik
z
(, k)

E
+
(k, , z) +
i
2
2
0
c
2
k
z
(, k)

P(k, , z)
_
e
it+ik.r
dk
2
d (3.86)
36 3 Unidirectional Pulse Propagation Equations
Add to this its backward-going counterpart to get the derivative of the total eld. Note that the
polarization terms cancel each other, and a forward-backward eld dierence appears:

z
E(r, t, z) =
z
(E
+
(r, t, z) +E

(r, t, z)) =
=
_
_
+ik
z
(, k)(

E
+
(k, , z)

E

(k, , z))
_
e
it+ik.r
dk
2
d (3.87)
From here, we take the second z-derivative to obtain:

zz
E(r, t, z) =
_
_
+ik
z
(, k)(
z

E
+
(k, , z)
z

E

(k, , z))
_
e
it+ik.r
dk
2
d (3.88)
which, after using the propagation equations once again, and grouping into terms of total eld
and polarization, yields:

zz
E(r, t, z) =
_ _
k
z
(, k)
2

E(k, , z)

2

0
c
2

P(k, , z)
_
e
it+ik.r
dk
2
d (3.89)
Having evaluated the
zz
term, we continue on with the remaining, transverse part of the
Laplacian:
(
xx
+
yy
)E(r, t, z) =
_
(k
2
x
+k
2
y
)

E(k, , z)e
it+ik.r
dk
2
d (3.90)
The remaining piece of the wave equation is the temporal derivative, which is also written in the
spectral domain:

tt
E(r, t, z) =
_

2
()

E(k, , z)e
it+ik.r
dk
2
d (3.91)
Now, put everything together to form left-hand side of a wave equation. In the process, terms
corresponding to the linear propagation cancel out, and only the nonlinear polarization survives:
(
zz
+
xx
+
yy
)E(r, t, z) 1/c
2

tt
E(r, t, z) =
_

2

0
c
2

P(k, , z)e
it+ik.r
dk
2
d =
1

0
c
2

tt
P(r, t, z) (3.92)
On the right-hand side of the above equation, we nd exactly what there should be, namely the
second temporal derivative of the polarization. We have shown that the sum of the forward and
backward solutions to the UPPE obeys the wave equation. Importantly, it is also evident that
each of the one-way elds alone does not. This observation explains the diculties encountered
in derivation of previous pulse propagation equations which started from the wave equation, and
worked with a single amplitude representing the one-directional eld. A prime example is the
Nonlinear Envelope Equation which was obtained for the price of more or less arbitrary neglecting
a number of undesirable terms. Now we can see that this was a consequence of an attempt to
satisfy, with a single envelope solutions, an equation which in fact must not be satised...
4
Numerical Methods for UPPE Solution
Having formulated our pulse propagation model in the previous Sections, we now address the
question of how to solve it numerically. To keep the notation simple, and equations readable, we
will restrict ourselves to the simplest version of UPPE, namely a one-component propagation
equation in a bulk medium. We also suppose that all nonlinear interactions are expressed in the
nonlinear polarization P and for simplicity omit from our equations the nonlinear current-density
term. Note that the numerical approach described in what follows translates directly to a general
case, and its practical implementation in software is essentially the same.
4.1 UPPE as a large system of ordinary dierential equations
We have written UPPE equation in a form which resembles the usual structure of pulse prop-
agation equations and to which simulation practitioners in the eld are most used to. Namely,
the UPPE expresses evolution of an electric eld, and its right-hand-side contains terms related
to the linear and nonlinear parts:

z
E
kx,ky
(, z) = ik
z
E
kx,ky
(, z) +
i
2
2
0
c
2
k
z
P
kx,ky
(, z) where k
z
=
_

2
()/c
2
k
2
x
k
2
y
.
(4.1)
Let us point out a few important points before going into details of a solver implementation.
First, unlike other propagation models, this is in a spectral representation. It describes evolution
of a Fourier spectrum, rather than that of a real physical eld. One important consequence is that
what he have is not a partial dierential equation anymore. Rather, it is a system of ordinary
dierential equations for spectral amplitudes, albeit a very large system. Thus, there are no
partial derivatives to approximate, which makes numerical solution conceptually very simple:
One can utilize any available library for ODE systems, and the only remaining thing to do
is to dene a right-hand-side calculation subroutine which will be fed to the chosen ODE solver.
Second, the above representation, which is in terms of spectral amplitudes for electric eld, is
not exactly one a numerical solver should work with because the slowest-evolving variables in this
problem only change in response to nonlinearity. And these are actually the native variables
A
kx,ky
which appeared in the course of UPPE derivation:
E
kx,ky
(, z) = A
kx,ky
(, z) exp[ik
z
z] A
kx,ky
(, z) exp[i
_

2
()/c
2
k
2
x
k
2
y
z] (4.2)
and their evolution equation (see previous Section) only contains nonlinear terms in its right-
hand-side:
38 4 Numerical Methods for UPPE Solution

z
A
kx,ky
(, z) = +
i
2
2
0
c
2
k
z
e
ikzz
P
kx,ky
(, z) with k
z
=
_

2
()/c
2
k
2
x
k
2
y
. (4.3)
It is obvious that spectral amplitudes A are the slowest variable in the pulse evolution problem,
because they do not change at all in a linear regime. This also means that UPPE equations
exactly solve the linear part of a problem, which is an extremely desirable property (reader is
encouraged to review various propagation equations specically to recall how much eort often
goes even into design of the linear part of all these equations ). The most important advantage
is the ability to model an arbitrary medium with frequency dependent index of refraction and
frequency dependent losses.
Alternatively, one can view (4.3) as Eqn. (4.1) to which integrating factor exp[ik
z
z] has been
applied. This cancels oscillations in the electric elds spectral amplitudes which are due to linear
propagation. Nonlinearity alone contributes to the evolution of the native UPPE variables A
kx,ky
,
and implementation based on them thus yields to faster numerical integration. This point of view
makes it evident that there is a degree of freedom in the relation between E and A amplitudes.
Namely, one could also use an integrating factor exp[ik
z
(z z
0
)], thus moving the point at which
A = E from z = 0 to z
0
. Lacking a better term, we say that E and A amplitudes are aligned
at z
0
.
It is imperative that a simulator utilizes this degree of freedom. At z = z
0
spectral amplitudes
of E and native variables A coincide, but as z increases, E and A diverge. Although always
connected by a simple complex phase change = k
z
(, k
x
, k
y
)(zz
0
), one must realize that the
latter can attain very large values. To avoid numerical diculties in handling the corresponding
exponentials, the relation between E and A has to be re-aligned after every integration step by
moving z
0
to the current propagation distance. After the new array of A is produced in the ODE
solver taking a step z, re-alignment with E is achieved by
A
new
kx,ky
(, z) = exp[ik
z
(, k
x
, k
y
)z]A
old
kx,ky
(, z) (4.4)
After this operation, E and A amplitudes coincide once again when the next integration step
is to be executed. Note that the above procedure corresponds to nothing but to the free, linear
propagation of the eld, and re-alignment therefore amounts to application of a linear propagator.
In what follows it is assumed that amplitudes are aligned before each step, which means that
z = 0 in (4.3) should be understood as a beginning of the current ODE solver step, and z values
are small, restricted to z < z.
Equation (4.3) is not completely explicit, because it hides the fact that the nonlinear polar-
ization P must be calculated from the current value of the electric eld. Indeed, polarization
is a functional of E(x, y, z, t) taken at a xed z value. As a rule, medium models are formulated
in real space. For example Raman-eect contribution to the change of refractive index can be
expressed as convolution in time direction calculated at a given spatial point (x, y, z). While con-
crete relation between electric and polarization elds is unimportant for how the UPPE solver is
designed, the fact that medium response is calculated in real space is crucial. We will therefore
assume that a function implementation P(x, y, z, t) = P
NL
{E(x, y, z, t)} is given which calculates
polarization as a function of time given a history of electric eld at a xed spatial point.
Let us incorporate this into our notation and write down an explicit denition of the UPPE
ODE system. The unknown functions are A
kx,ky
(, z) and they obey

z
A
kx,ky
(, z) = +
i
2
2
0
c
2
k
z
e
ikzz
P
kx,ky
(, z, {E(x, y, z, t)}) , k
z
=
_

2
()/c
2
k
2
x
k
2
y
(4.5)
where
4.2 Discretization and spectral transforms 39
P
kx,ky
(, z, {E(x, y, t)}) = (2)
3/2
_
e
+itikxxikyy
P
NL
({E(x, y, z, t)})dxdydt (4.6)
is a Fourier (or in general spectral) transform of P
NL
({E(x, y, z, t)}) which in turn encapsulates
nonlinear medium properties. (Note that a solver implementation does not need to, and in fact
should not know about its concrete functional form!) To evaluate the above, one rst needs to
calculate the real-space eld from the native computational variables through another (inverse)
Fourier transform
E(x, y, z, t) = (2)
3/2
_
e
it+ikxx+ikyy
A
kx,ky
(, z)e
ikzz
ddk
x
dk
y
(4.7)
Now it should be clear that when an ODE solver requests evaluation of its right-hand-side
function for a given value of z, and for a given grid representation of A
kx,ky
(, z), spectral
transforms will be invoked in both directions. One can view this as a price to pay for the
elimination of partial derivatives from propagation equations, and for the ability to solve the
linear problem exactly.
There is one more issue to clarify before transition to discretization, and that is that of a
moving frame. It is of course advantageous, and in fact necessary, to follow pulse evolution in a
frame of reference which moves with a velocity v
f
(w.r.t. the laboratory frame) chosen such that
at each location z in a lab, the pulse arrives at t 0, and thus stays located around the center of
the temporal computational domain. This is achieved by expressing time t through t = +z/v
f
where is our new temporal variable. Inspection of Eqns. (4.5-4.7) reveals that this amounts to
a simple modication of the linear propagator:
exp[ik
z
(, k
x
, k
y
)z] exp[ik
z
(, k
x
, k
y
)z z/v
f
] (4.8)
We emphasize that v
f
is an arbitrary parameter which does not reect any physics of the model.
It is merely an expression of what one deems to be the best reference frame. Quite often it is
reasonable to choose
1
v
f
=
1
v
group
=
k
z
(
pulse
, k
x
= 0, k
y
= 0)

(4.9)
which means that the computational frame of reference moves with the group velocity of the
pulse. Readers familiar with Nonlinear Schrodinger Equation should realize that it is exactly the
choice that makes a pulse described by NLS to stay localized in the vicinity of 0.
4.2 Discretization and spectral transforms
Because any UPPE solver is spectral in all dimensions, grid representation of both, real-space and
spectral space elds is determined by properties of discrete spectral transforms. Depending on
the symmetry of the problem, these are either variants of Fourier or discrete Hankel transforms.
Spatial (linear) axis
For a spatial dimension, say x, that spans one side of a computational domain box, the values
of coordinates (in real space) and transverse wavenumbers k
x
(in spectral space) are those of
ordinary Fourier transform sampling points. Both sets are equidistant and equal in size.
Temporal axis
A computational domain axis in time direction has its corresponding Fourier transform which
is slightly modied due to the fact that physical elds are real-valued. It is sucient to sample
40 4 Numerical Methods for UPPE Solution
spectral amplitudes A
kx,ky
() only for positive angular frequencies . Moreover, one can restrict
discrete sampling points to (
MIN
,
MAX
) if one only knows the medium susceptibility ()
in this interval. Only these discrete frequencies become active in the simulations in the sense
that they carry corresponding spectral amplitudes. When spectral-to-real transform is invoked,
the active set of frequency-samples is padded by zeros before a standard Fourier transform
is executed. This has the eect that the resulting real-space amplitude becomes the so-called
analytic signal. While its real part corresponds to physical electric eld, its modulus squared can
be interpreted (in suitable units) as the time-averaged light intensity. Both quantities are often
needed in the calculation of various nonlinear medium responses. Note that in this arrangement
the total number of discrete samples in the time dimension is more than twice the number of
active samples in the frequency dimensions.
Radial axis
For problems with axial symmetry, it is advantageous to utilize the radial discrete Hankel
transform instead of a two-dimensional Fourier transform. Because the discrete Hankel transform
is represented by a full matrix, it is not fast in the sense fast Fourier transforms are. Still,
the main computational savings are related to the reduced dimensionality of variable-arrays
representing physical elds. Sampling points in both real and spectral space are the same and
namely given by scaled zeros of Bessel function J
0
. For example N samples in real space are
r
i
= u
i
/u
N
R where J
0
(u
i
) = 0. Note that there is no radial sample located directly on the axis.
Readers concerned about the usage of slow spectral transform, should note that there are
fast discrete Hankel transforms. However, UPPE solver requires a truly orthogonal transform
implementation, because forward and inverse transformations are executed many times over the
same array. Only the proper Hankel transform is orthogonal (and in addition equal to its own
inverse) and should be preferred on grounds of numerical accuracy.
4.3 Integration of evolution equations for spectral amplitudes
The core of a UPPE solver can be based on essentially arbitrary ODE-solver library. A good
library should have the capability to choose between dierent algorithms, and it will also take care
of allocating auxiliary arrays based on the selected method. The advantage of using a canned
library over hard-coding a concrete algorithm into the solver implementation is the exibility in
the choice of method, and also the fact that everything concerning auxiliary variables involved in
ODE solution remains hidden. However, the UPPE system can contain several million variables,
and that an ODE solver will, depending on the requested method, allocate several auxiliary arrays
of the same size. As a consequence, the bulk of the memory allocated for the whole simulation
will actually be requested by the solver. Not only the memory needs will be several times larger
than those for one copy of all elds, but also that the method performance can be aected by
the size of the system solved. In practice simpler methods tend to perform better than the more
sophisticated ones. In particular, all methods that require calculation of a Jacobian are utterly
unsuitable for UPPE solution. Fortunately, the standard ODE solver work-horses such as various
Runge-Kutta methods work well.
Now, suppose we have calculated an array A
kx,ky
(, z) representing the solution at propa-
gation distance z. To keep the notation simple, discrete wavenumbers k
x
, k
y
and active angular
frequencies now represent array indices. To calculate A
kx,ky
(, z + z), an ODE solver is
invoked to produce it. Because we synchronize A and E representations after each step, we can
understand z = 0 as if the currently executed step was the very rst one. A pseudo-code for the
integration loop reads:
4.3 Integration of evolution equations for spectral amplitudes 41
Repeat for each step:
A) Invoke ODE Solver:
A
kx,ky
(, 0) A
kx,ky
(, z)
B) Re-align native and eld variables:
A
kx,ky
(, z) = exp[i(k
z
(, k
x
, k
y
) /v
f
)z]A
kx,ky
(, z)
Readers may note that the above integration may seem like an operator splitting method. It
does looks as if nonlinear and linear propagators were applied in turns the same way as in the
split-step approach. However, this is where the similarity ends. The second sub-step is nothing
but a shift of our reference frame. If our numerics did not suer from rounding, and could
evaluate exponentials with arbitrarily large arguments, this addition would not be necessary.
Behind the scenes, while executing A), ODE solver will invoke calculation of the right-hand-
side of our ODE system, i.e. it asks to evaluate
i
2
2
0
c
2
k
z
e
i(kz/v
f
)z
P
kx,ky
(, z, {E(x, y, z, t)})
for given spectral amplitudes A
kx,ky
(). Depending on the ODE algorithm, call of this function
occurs several times during a single integration step, each time with a dierent value of z. UPPE
solver implements the function call in the following steps:
1. Apply linear propagator to shift from z = 0 to z:
A
kx,ky
() A
kx,ky
()exp[+i(k
z
(, k
x
, k
y
) /v
f
)z]
2. Perform spectral transform from spectral to real space:
E(x, y, t) = FFT{A
kx,ky
()}
3. In real-space representation, calculate nonlinear medium response for a given eld E(x, y, t),
using a user-supplied medium-response implementing algorithm P
NL
:
P(x, y, t) = P
NL
{E(x, y, t)}
4. Perform spectral transform from the real to spectral space:
P
kx,ky
() = FFT
1
{P(x, y, t)}
5. Apply linear propagator to undo the previous shift in z:
P
kx,ky
() P
kx,ky
()exp[i(k
z
(, k
x
, k
y
) /v
f
)z]
6. Finally, multiply by the coupling factor and return result to the ODE solver:
P
kx,ky
()
i
2
2
0
c
2
k
z
P
kx,ky
()
The above is the most computation-intensive part of UPPE solution and thus merits attention
w.r.t. ecient parallel implementation. This and related issues are discussed in Sections devoted
to implementation in software.
42 4 Numerical Methods for UPPE Solution
4.4 Parallelization
Even if a problem has axial symmetry, a typical UPPE ODE system contains several million
variables. It is therefore more or less necessary that computations are parallelized. Let us briey
point out facts that may inuence our parallelization design decisions.
First, we have to take into account the fact that the UPPE framework is inherently spectral.
This means that each processor or a thread of execution will, at some point, require access to
distant locations in allocated arrays. The shared memory paradigm is therefore a natural way
to go, and the current UPPEcore implementation uses Pthreads. OpenMP (i.e. pragma based
loop parallelization) would probably work equally well for our purposes.
Second, one has to decide which parts of the code will actually execute in parallel. Of course,
ideally all of them, but the question is if it is worth of the trouble. It turns out that most of
the computational eort in UPPE is spent within the ODE system right-hand-side calculations
and in performing spectral transforms. This invites a parallel work crew strategy: A master
thread creates a family of workers or slaves and dispatches these to perform work as needed. The
master thread executes all work outside of the ODE solver loop which includes all initialization,
observation of results, input, and output. Master also runs the main ODE loop without help
from its slaves. (This is of course a compromise between between the achievable parallelization
eciency and complexity of the solution!) This means that almost any serial ODE solver library
can be used as a plug-in for the UPPE solver.
The parallel working crew enters a synchronization barrier immediately after their creation.
Here, they wait for commands from their master who species which is the next function to
execute in parallel. For example, this may be a spectral transform. Within the parallel section,
each of the workers takes over a proportional part of the load. When the parallel section is done,
workers meet again at the synchronization barrier, awaiting further commands from the master.
Obviously the biggest drawback of this design is that the ODE part of the code remains serial.
Although it is relatively small, it would limit possible parallel speed up with a large number of
threads. An alternative solution is akin to the domain decomposition strategy; the set of ODE
equations is evenly distributed between multiple instances of ODE solvers, each executed by an
independent thread. However, this solution requires a mild modication of the ODE routines
that control adaptive integration step. Some earlier versions of the UPPEcore were based on this
approach. Since the performance penalty with the working crew method is mostly negligible in
practice, it is preferred because it does not require open access to the code of the ODE solver.