April 19, 2007 Master Review Vol.

9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

The Nonlinear Schrödinger Equation and Applications in

Bose-Einstein Condensation and Plasma Physics

Weizhu Bao
Department of Mathematics
Center for Computational Science and Engineering
National University of Singapore
117543 Singapore
E-mail: bao@[Link],
URL: [Link]

Contents
1 Introduction 143
2 Derivation of NLSE from wave propagation 144
3 Derivation of NLSE from BEC 146
3.1 Dimensionless GPE 148
3.2 Reduction to lower dimension 148
4 The NLSE and variational formulation 150
4.1 Conservation laws 150
4.2 Lagrangian structure 151
4.3 Hamiltonian structure 152
4.4 Variance identity 153
5 Plane and solitary wave solutions of NLSE 157
6 Existence/blowup results of NLSE 158
6.1 Integral form 159
6.2 Existence results 159
6.3 Finite time blowup results 160
7 WKB expansion and quantum hydrodynamics 161
8 Wigner transform and semiclassical limit 162
9 Ground, excited and central vortex states of GPE 164
9.1 Stationary states 165

141
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

142 W. Bao

9.2 Ground state 165

9.3 Central vortex states 167
9.4 Variation of stationary states over the unit sphere 168
9.5 Conservation of angular momentum expectation 169
10 Numerical methods for computing ground states of GPE 171
10.1 Gradient flow with discrete normalization (GFDN) 171
10.2 Energy diminishing of GFDN 171
10.3 Continuous normalized gradient flow (CNGF) 173
10.4 Semi-implicit time discretization 174
10.5 Discretized normalized gradient flow (DNGF) 177
10.6 Numerical methods 178
10.7 Energy diminishing of DNGF 181
10.8 Numerical results 183
11 Numerical methods for dynamics of NLSE 189
11.1 General high-order split-step method 189
11.2 Fourth-order TSSP for GPE without external driving field 190
11.3 Second-order TSSP for GPE with external driving field 192
11.4 Stability 192
11.5 Crank-Nicolson finite difference method (CNFD) 195
11.6 Numerical results 195
12 Derivation of the vector Zakharov system 201
13 Generalization and simplification of ZS 205
13.1 Reduction from VZSM to GVZS 206
13.2 Reduction from GVZS to GZS 208
13.3 Reduction from GVZS to VNLS 210
13.4 Reduction from GZS to NLSE 211
13.5 Add a linear damping term to arrest blowup 212
14 Well-posedness of ZS 212
15 Plane wave and soliton wave solutions of ZS 213
16 Time-splitting spectral method for GZS 214
16.1 Crank-Nicolson leap-frog time-splitting spectral discretizations
(CN-LF-TSSP) for GZS 216
16.2 Phase space analytical solver + time-splitting spectral discretiza-
tions (PSAS-TSSP) 218
16.3 Properties of the numerical methods 221
16.4 Extension TSSP to GVZS 224
17 Crank-Nicolson finite difference (CNFD) method for GZS 226
18 Numerical results of GZS 227
References 233
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 143

1. Introduction
The Schrödinger equation was proposed to model a system when the quan-
tum effect was considered. For a system with N particles, the Schrödinger
equation is defined in 3N + 1 dimensions. With such high dimensions, even
use today’s supercomputer, it is impossible to solve the Schrödinger equa-
tion for dynamics of N particles with N > 10. After assumed Hatree
or Hatree-Fork ansatz, the 3N + 1 dimensions linear Schrödinger equa-
tion was approximated by a 3 + 1 dimensions nonlinear Schrödinger equa-
tion (NLSE) or Schrödinger-Poisson (S-P) system. Although nonlinearity
in NLSE brought some new difficulties, but the dimensions were reduced
significantly compared with the original problem. This opened a light to
study dynamics of N particles when N is large. Later, it was found that
NLSE had applications in different subjects, e.g. quantum mechanics, solid
state physics, condensed matter physics, quantum chemistry, nonlinear op-
tics, wave propagation, optical communication, protein folding and bending,
semiconductor industry, laser propagation, nano technology and industry,
biology etc. Currently, the study of NLSE including analysis, numerics and
applications becomes a very important subject in applied and computa-
tional mathematics. This study has very important impact to the progress
of other science and technology subjects.
A typical application of NLSE is for wave motion and interaction in
plasma physics where the Zakharov system (ZS) was derived by Zakharov
[121] in 1972 for governing the coupled dynamics of the electric-field ampli-
tude and the low-frequency density fluctuations of ions. Then it has become
commonly accepted that ZS is a general model to govern interaction of
dispersive wave and nondispersive (acoustic) wave. It has important appli-
cations in plasma physics (interaction between Langmuir and ion acoustic
waves [121, 101]), in the theory of molecular chains (interaction of the in-
tramolecular vibrations forming Davydov solitons with the acoustic distur-
bances in the chain [39]), in hydrodynamics (interaction between short-wave
and long-wave gravitational disturbances in the atmosphere [110, 40]), and
so on. In three spatial dimensions, ZS was also derived to model the col-
lapse of caverns [121]. Later, the standard ZS was extended to generalized
Zakharov system (GZS) [72, 73], vector Zakharov system (VZS) [113] and
vector Zakharov system for multi-component plasma (VZSM) [72, 73].
In this chapter, we first review derivation of NLSE from wave propaga-
tion and Bose-Einstein condensation (BEC). Then we present variational
formulation of NLSE including conservation laws, Lagrangian structure,
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

144 W. Bao

Hamiltonian structure and variance identity. Plane and soliton wave solu-
tions, existence/blowup results of NLSE are then presented. Ground, ex-
cited and central vortex states of NLSE with an external potential are
studied. We also study formally semiclassical limits of NLSE by WKB ex-
pansion and Wigner transform when the (scaled) Planck constant ε → 0.
In addition, numerical methods for computing ground states and dynamics
of NLSE are presented and numerical results are also reported. Then we
review derivation of VZS from the two-fluid model [113] for ion-electron
dynamics in plasma physics and generalize VZS to VZSM, reduce VZSM
to generalized vector Zakharov system (GVZS), GVZS to GZS or vector
nonlinear Schrödinger (VNLS) equations, and GZS to NLSE, as well as
generalize GZS and GVZS with a linear damping term to arrest blowup.
Conservation laws of the systems and well-posedness of GZS are presented,
and plane wave, soliton wave and periodic wave solutions of GZS are re-
viewed. Furthermore, we present a time-splitting spectral (TSSP) method
to discretize GZS and compare it with the standard Crank-Nicolson finite
difference (CNFD) method.
Throughout this notes, we use f ∗ , Re(f) and Im(f) denote the conjugate,
real part and imaginary part of a complex function f respectively. We also
adopt the standard Sobolev norms.

2. Derivation of NLSE from wave propagation

In this section, we review briefly derivation of NLSE from wave propaga-
tion, i.e. parabolic or paraxial approximation for forward propagation time
harmonic waves, to analyze high frequency asymptotics.
The wave equation
1 ∂ 2 u(x, t)
− 4u(x, t) = 0, x ∈ R3 , (2.1)
c2 ∂t2
where x = (x, y, z) is the Cartesian coordinate, t is time and c = c(x, |u|)
is the propagation speed, has time harmonic solutions of the form eiωt u(x)
with the complex wave function u satisfying the Helmholtz or reduced wave
equation
ω2
4u(x) + u = 0, x ∈ R3 . (2.2)
c2
Let c0 be a uniform reference speed, k0 = ω/c0 be the wave number and
n(x, |u|) = c0 /c(x, |u|) be the index of refraction. The reduced wave equa-
tion has then the form
4u(x) + k02 n2 (x, |u|)u = 0. (2.3)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 145

When wave propagates in a uniform medium, n(x, |u|) = 1; in a lin-

ear
p medium, n(x, |u|) = n(x); and in a Kerr medium, n(x, |u|) =
1 + 4n2 |u|2 /n0 with n0 linear index of refraction and n2 Kerr coefficient.
When waves are approximately plane and move in one direction pri-
marily, say the z direction, e.g. propagation of laser beams, we look for
solutions of the form
u(x, y, z) = eik0 z ψ(x, z) (2.4)
where x = (x, y) denotes the transverse variables. We insert (2.4) into the
reduced wave equation (2.3) and get
2ik0 ψz + 4⊥ ψ + k02 µ(x, z, |ψ|)ψ + ψzz = 0, (2.5)
where 4⊥ is the Laplacian in the transverse variables and µ(x, z, |ψ|) =
n2 (x, z, |ψ|) − 1 is the fluctuation in the refractive index. Note that the
direction of propagation z ploys the role of time and −k02 µ(x, z, |ψ|) is the
(time dependent) potential.
Introduce nondimensional variables:
x y z ψ(x, y, z)
x̃ = , ỹ = , t̃ = , ψ̃(x̃, ỹ, t̃) = , (2.6)
r0 r0 k0 r02 ψs
where r0 is the dimensionless length unit, e.g. width of the input laser beam,
and ψs is dimensionless unit for ψ to be determined. Plugging (2.6) into
(2.5), multiplying by r02 /2, and then removing all ˜, we get the following
dimensionless equation:
1 δ
iψt = − 4⊥ ψ + f (x, t, |ψ|)ψ − ψtt , (2.7)
2 2
where δ = 1/r02 k02 and the real-valued function f depends on µ. Due to the
input beam width r0  λ = 2π/k0 , we get
δ/2 = λ2 /8π 2 r02  1.
Thus we drop the nonparaxial term ψtt in (2.7) and obtain the NLSE:
1
iψt = − 4⊥ ψ + f (x, t, |ψ|)ψ, (2.8)
2
Of course (2.7) is only an approximation to the full reduced wave equation
and it is valid when the variations in the index of refraction are smooth and
the bulk of the wave energy is away from boundaries. This important and
very useful approximation for wave propagation is well suited for numerical
approximation since we now have an initial value problem for ψ, assuming
that ψ(x, 0) is known, rather than a boundary value problem for u.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

146 W. Bao

When n(x, z, |u|) = 1 in (2.3), then µ(x, z, |ψ|) = 0 in (2.5) and

f (x, t, |ψ|) = 0 in (2.7), thus (2.7) collapses to the free Schrödinger equa-
tion:
1
iψt = − 4⊥ ψ. (2.9)
2
When n(x, z, |u|) = n(x, z) in (2.3), then µ(x, z, |ψ|) = µ(x, z) in (2.5) and
f (x, t, |ψ|) = V (x, t) in (2.7), thus (2.7) collapses to a linear Schrödinger
equation with potential V (x, t):
1
iψt = − 4⊥ ψ + V (x, t)ψ. (2.10)
2
p
When n(x, z, |u|) = 1 + 4n2 |u|2 /n0 , i.e. laser beam in Kerr medium, then
µ(x, z, |ψ|) = 2n2 r02 k02 |ψ|2 /n0 in (2.5) and f (x, t, |ψ|) = −|ψ|2 in (2.7) by
√ √
choosing ψs = n0 /r0 k0 2n2 , thus (2.7) collapses to NLSE with a cubic
focusing nonlinearity:
1
iψt = − 4⊥ ψ − |ψ|2 ψ. (2.11)
2
The wave energy or power of the beam is conserved:
Z Z
N (ψ) = |ψ(x, t)|2 dx ≡ |ψ(x, 0)|2 dx, t ≥ 0. (2.12)
R2 R2

Remark 2.1: When we consider high frequency asymptotics which con-

cerns approximate solutions of (2.10) that are good approximations to os-
cillatory solutions. For such solutions the propagation distance is long com-
pared to the wavelength, the propagation time is large compared to the pe-
riod and the potential V (x) varies slowly. To make this precise, we introduce
slow time and space variables t → t/ε, x → x/ε with 0 < ε  1 the (scaled)
Planck constant and the scaled wave function ψ ε (x, t) = ψ(x/ε, t/ε) which
satisfies the NLSE in the semiclassical regime
ε2
iεψtε = − 4⊥ ψ ε + V ε (x, t)ψ ε , x ∈ R2 , t > 0, (2.13)
2
where V ε (x, t) = V (x/ε, t/ε).

3. Derivation of NLSE from BEC

Since its realization in dilute bosonic atomic gases [3, 26], BEC of alkali
atoms and hydrogen has been produced and studied extensively in the
laboratory [71], and has spurred great excitement in the atomic physics
community and renewed the interest in studying the collective dynamics of
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 147

macroscopic ensembles of atoms occupying the same one-particle quantum

state [99, 34, 68]. The condensate typically consists of a few thousands to
millions of atoms which are confined by a trap potential. In fact, beside
the effects of the internal interactions between the atoms, the macroscopic
behavior of BEC matter is highly sensitive to the shape of this external
trapping potential. Theoretical predictions of the properties of a BEC like
the density profile [19], collective excitations [43] and the formation of vor-
tices [105] can now be compared with experimental data [3]. Needless to
say that this dramatic progress on the experimental front has stimulated a
wave of activity on both the theoretical and the numerical front.
At temperatures T much smaller than the critical temperature Tc [84], a
BEC is well described by the macroscopic wave function ψ = ψ(x, t) whose
evolution is governed by a self-consistent, mean field NLSE known as the
Gross-Pitaevskii equation (GPE) [69, 103]
∂ψ(x, t) δH(ψ)
i~ =
∂t δψ ∗
~2 2
=− ∇ ψ(x, t) + V (x)ψ(x, t) + N U0 |ψ(x, t)|2 ψ(x, t), (3.1)
2m
where x = (x, y, z), m is the atomic mass, ~ is the Planck constant,
N is the number of atoms in the condensate, V (x) is an external trap-
ping potential. When a
harmonic trap potential is considered, V (x) =
m 2 2 2 2 2 2
2 ωx x + ωy y + ωz z with ωx , ωy and ωz being the trap frequencies
in x, y and z-direction, respectively. The Hamiltonian (or energy) of the
system H(ψ) per particle is defined as
Z  
∗ ~2 2
H(ψ) = ψ (x, t) − ∇ + V (x) ψ(x, t)dx
R3 2m
Z
1
+ ψ ∗ (x, t) ψ ∗ (x0 , t)Φ(x − x0 )ψ(x0 , t)ψ(x, t)dxdx0 , (3.2)
2 R3 ×R3
where the interaction potential is taken as the Fermi form
Φ(x) = (N − 1)U0 δ(x). (3.3)
U0 = 4π~2 as /m describes the interaction between atoms in the condensate
with the s-wave scattering length as (positive for repulsive interaction and
negative for attractive interaction). It is convenient to normalize the wave
function by requiring
Z
|ψ(x, t)|2 dx = 1. (3.4)
R3
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

148 W. Bao

3.1. Dimensionless GPE

In order to scale the Eq. (3.1) under the normalization (3.4), we introduce
x 3/2
p
t̃ = ωm t, x̃ = , ψ̃(x̃, t̃) = a0 ψ(x, t), with a0 = ~/mωm , (3.5)
a0
where ωm = min{ωx , ωy , ωz }, a0 is the length of the harmonic oscillator
ground state. In fact, we choose 1/ωm and a0 as the dimensionless time
and length units, respectively. Plugging (3.5) into (3.1), multiplying by
2 1/2
1/mωm a0 , and then removing all ˜, we get the following dimensionless
GPE under the normalization (3.4) in three dimension
∂ψ(x, t) 1
i = − ∇2 ψ(x, t) + V (x)ψ(x, t) + β |ψ(x, t)|2 ψ(x, t), (3.6)
∂t 2
U0 N 4πas N
where β = a30 ~ωm
= a0 and
1 2 2
ωα
V (x) = γx x + γy2 y 2 + γz2 z 2 , with γα = (α = x, y, z).
2 ωm
There are two extreme regimes of the interaction parameter β: (1) β =
o(1), the Eq. (3.6) describes a weakly interacting condensation; (2) β  1,
it corresponds to a strongly interacting condensation or to the semiclassical
regime.
There are two typical extreme regimes between the trap frequencies: (1)
γx = 1, γy ≈ 1 and γz  1, it is a disk-shaped condensation; (2) γx = 1,
γy  1 and γz  1, it is a cigar-shaped condensation. In these two cases,
the three-dimensional (3D) GPE (3.6) can be approximately reduced to a
2D and 1D equation respectively [85, 8, 5] as explained below.

3.2. Reduction to lower dimension

Case I (disk-shaped condensation):
ωx ≈ ωy , ωz  ωx , ⇐⇒ γx = 1, γy ≈ 1, γz  1.
Here, the 3D GPE (3.6) can be reduced to a 2D GPE with x = (x, y)
by assuming that the time evolution does not cause excitations along the
z-axis, since the excitations along the z-axis have large energy (of order
~ωz ) compared to that along the x and y-axis with energies of order ~ωx .
Thus, we may assume that the condensation wave function along the z-
axis is always well described by the harmonic oscillator ground state wave
function, and set
2
ψ = ψ2 (x, y, t)φho (z) with φho (z) = (γz /π)1/4 e−γz z /2
. (3.7)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 149

Plugging (3.7) into (3.6), multiplying by φ∗ho (z), integrating with respect to
z over (−∞, ∞), we get
∂ψ2 (x, t) 1 1 2 2

i = − ∇2 ψ2 + γx x + γy2 y 2 + C ψ2 + β2 |ψ2 |2 ψ2 , (3.8)
∂t 2 2
where
Z ∞ r Z ∞   !
γz  dφho 2
β2 = β 4
φho (z) dz = β , C= γz z |φho (z)| + 
2 2 2  dz.
−∞ 2π −∞ dz 
Ct
Since this GPE is time-transverse invariant, we can replace ψ2 → ψ e−i 2
so that the constant C in the trap potential disappears, and we obtain the
2D effective GPE:
∂ψ(x, t) 1 1 2 2

i = − ∇2 ψ + γx x + γy2 y 2 ψ + β2 |ψ|2 ψ. (3.9)
∂t 2 2
Note that the observables, e.g. the position density |ψ|2 , are not affected
by dropping the constant C in (3.8).
Case II (cigar-shaped condensation):
ωy  ωx , ωz  ωx ⇐⇒ γx = 1, γy  1, γz  1.
Here, the 3D GPE (3.6) can be reduced to a 1D GPE with x = x. Similarly
as in the 2D case, we can derive the following 1D GPE [85, 8, 5]:
∂ψ(x, t) 1 γ 2 x2
i = − ψxx (x, t) + x ψ(x, t) + β1 |ψ(x, t)|2 ψ(x, t), (3.10)
∂t 2 2
√
where β1 = β γy γz /2π.
The 3D GPE (3.6), 2D GPE (3.9) and 1D GPE (3.10) can be written
in a unified form:
∂ψ(x, t) 1
i = − ∇2 ψ + Vd (x)ψ + βd |ψ|2 ψ, x ∈ Rd , (3.11)
∂t 2
ψ(x, 0) = ψ0 (x), x ∈ Rd , (3.12)
with
√  2 2
 pγy γz /2π,  γx x /2,
d = 1,
βd = β γz /2π, Vd (x) = γx2 x2 + γy2 y 2 /2,
d = 2,
  2 2
1, γx x + γy2 y 2 + γz2 z 2 /2, d = 3,
(3.13)
where γx > 0, γy > 0 and γz > 0 are constants. The normalization condition
for (3.11) is
Z Z
N (ψ) = kψ(·, t)k2 = |ψ(x, t)|2 dx ≡ |ψ0 (x)|2 dx = 1. (3.14)
Rd Rd
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

150 W. Bao

Remark 3.1: When βd  1, i.e. in a strongly repulsive interacting con-

densation or in semiclassical regime, another scaling of the GPE (3.11) is
also very useful. In fact, after a rescaling in (3.11) under the normalization
−2/(d+2)
(3.14): x → ε−1/2 x and ψ → εd/4 ψ with ε = βd , then the GPE
(3.11) can be rewritten as

∂ψ(x, t) ε2
iε = − ∇2 ψ + Vd (x)ψ + |ψ|2 ψ, x ∈ Rd . (3.15)
∂t 2

4. The NLSE and variational formulation

Consider the following NLSE:
1
iψt = − 4 ψ + V (x)ψ + β|ψ|2σ ψ, x ∈ Rd , t ≥ 0, (4.1)
2
ψ(x, 0) = ψ0 (x), x ∈ Rd , (4.2)

where σ > 0 is a positive constant (σ = 1 corresponds to a cubic nonlinear-

ity and σ = 2 corresponds to a quintic nonlinearity), V (x) is a real-valued
potential whose shape is determined by the type of system under investi-
gation, β positive/negative corresponds to defocusing/focusing NLSE.

4.1. Conservation laws

Two important invariants of (4.1) are the normalization of the wave function
Z
N (ψ(·, t)) = |ψ(x, t)|2 dx ≡ N = N (ψ0 ), t≥0 (4.3)
Rd

and the energy

Z  
1 β
E(ψ(·, t)) = |∇ψ(x, t)|2 + V (x)|ψ(x, t)|2 + |ψ(x, t)|2σ+2 dx
Rd 2 σ+1
= E(ψ0 ), t ≥ 0. (4.4)

When V (x) ≡ 0, another important invariant of (4.1) is the momentum

Z
i
P(ψ(·, t)) = (ψ∇ψ ∗ − ψ ∗ ∇ψ) dx ≡ P(ψ0 ), t ≥ 0. (4.5)
2 Rd
Define the mass center
Z
1
x̄(t) = x|ψ(x, t)|2 dx. (4.6)
N Rd
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 151

Note that the mass center obeys

Z Z
dx̄ i
N = x∂t |ψ|2 dx = − x∇ · [ψ∇ψ ∗ − ψ ∗ 4 ψ] dx
dt 2
Z
i
= [ψ∇ψ ∗ − ψ ∗ 4 ψ] dx = P(ψ0 ) (4.7)
2
and thus moves at a constant speed.
To get more conservation laws, one can use the Noether theorem [113].

4.2. Lagrangian structure

Define the Lagrangian density L associated to (4.1) in terms of the real and
imaginary parts u and v of ψ, or equivalently in terms of ψ and ψ ∗ viewed
as independent variables in the form
i ∗ 1 β
L= (ψ ψt − ψψt∗ ) − ∇ψ · ∇ψ ∗ − V (x)ψψ ∗ − ψ σ+1 (ψ ∗ )σ+1 . (4.8)
2 2 σ+1
Consider the action
Z t1 Z
∗
S{ψ, ψ } = L dxdt (4.9)
t0 Rd

as a functional on all admissible regular function satisfying the prescribed

conditions ψ(x, t0 ) = ψ0 (x) and ψ(x, t1 ) = ψ1 (x). Its variation

δS = S{ψ + δψ, ψ ∗ + δψ ∗ } − S{ψ, ψ ∗ } (4.10)

for infinitesimal δψ and δψ ∗ reads

Z t1 Z  
∂L ∂L ∂L
δS = δψ + · ∇δψ + δψt dxdt + c.c.
t0 Rd ∂ψ ∂∇ψ ∂ψt
Z t1 Z     
∂L ∂L ∂L
= −∇· − ∂t δψ dxdt
t0 Rd ∂ψ ∂∇ψ ∂ψt
 t1
∂L
+ δψ + c.c. (4.11)
∂ψt t0

A necessary and sufficient condition for a function ψ(x, t) to lead to

an extremum for the action S among the functions with prescribed values
ψ(·, t0 ) and ψ(·, t1 ), thus reduces to the Euler-Lagrange equations
   
∂L ∂L ∂L
=∇· + ∂t , (4.12)
∂ψ ∂∇ψ ∂ψt
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

152 W. Bao

which, when the Lagrangian (4.8) is used, reduces to the NLSE (4.1). This
system is easily rewritten in terms of the real fields u = (ψ + ψ ∗ )/2 and
v = (ψ − ψ ∗ )/2i as
   
∂L ∂L ∂L
=∇· + ∂t , (4.13)
∂u ∂∇u ∂ut
   
∂L ∂L ∂L
=∇· + ∂t . (4.14)
∂v ∂∇v ∂vt

4.3. Hamiltonian structure

As usual, a Hamiltonian structure is easily derived from the existence of
a Lagrangian. Writing ψ = u + iv in order to deal with real fields, the
Hamiltonian density H = 2i (ψ ∗ ∂t ψ − ψ∂t ψ ∗ ) − L becomes
H = v∂t u − u∂t v − L. (4.15)
Introducing the canonical variables
∂L
q1 ≡ u, p1 ≡ , (4.16)
∂(∂t q1 )
∂L
q2 ≡ v, p2 ≡ , (4.17)
∂(∂t q2 )
it takes the form
X
H= pj ∂t qj − L. (4.18)
j

Define
∂L
ρj ≡ , (4.19)
∂(∇qj )
and rewrite the Euler-Lagrange equations as
∂L
= ∇ · ρj + ∂t pj . (4.20)
∂q
Using that
X ∂L ∂L ∂L
∂t L = ∂t qj + ∂t ∇qj + ∂tt qj , (4.21)
j
∂qj ∂∇qj ∂(∂t qj )

and the Euler-Lagrange equations, we get

X
∂t H = −∇ · ρj ∂t qj , (4.22)
j
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 153

which
R ensures that the conservation fo the Hamiltonian or energy H =
R d H dx.
Similarly, from the variation of the Lagrangian density
X δL δL δL
δL = δqj + ∇δqj + ∂t δqj , (4.23)
j
δq j δ∇q j δ(∂t qj )

the Euler-Lagrange equations, and the definition of pj we obtain the vari-

ation of the Hamiltonian H, in the form
XZ
δH = (∂t qj δpj − ∂t pj δqj )dx, (4.24)
j

which leads to the Hamilton equaitons

∂qj δH ∂pj δH
= , =− , (4.25)
∂t δpj ∂t δqj

or in complex form,

δH
i∂t ψ = . (4.26)
δψ ∗

4.4. Variance identity

Define the ‘variance’ (or ‘momentum of inertia’ in a context where N is
referred to as the mass of the wave packet) as
Z d
X Z
2 2
δV = |x| |ψ| dx = δj , δj = x2j |ψ|2 dx, j = 1, · · · , d (4.27)
Rd j=1 Rd

and the square width of the wave packet

Z Z
1 2 2 1 δ
δx = |x−x̄| |ψ| dx = (|x|2 −|x̄|2 )|ψ|2 dx = V −|x̄|2 . (4.28)
N Rd N Rd N

Here we use x = (x1 , · · · , xd ) ∈ Rd .

When V (x) ≡ 0 in (4.1), due to the conservation of the wave energy N
and of the linear momentum P, we have

d2 δx 1 d2 δV |P|2
= − 2 . (4.29)
dt2 N dt2 N2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

154 W. Bao

Lemma 4.1: Suppose ψ(x, t) be the solution of the problem (4.1), (4.2),
then we have
Z  
d2 δj (t) 2 2σβ 2σ+2 2
= 2|∂xj ψ| + |ψ| − 2xj |ψ| ∂xj (V (x)) dx, (4.30)
dt2 Rd σ+1
Z
(0)
δj (0) = δj = x2j |ψ0 (x)|2 dx, j = 1, · · · , d, (4.31)
R d
Z 
(1)
δj0 (0) = δj = 2 Im xj ψ0∗ ∂xj ψ0 dx . (4.32)
Rd

Proof: Differentiate (4.27) with respect to t, notice (4.1), integrate by parts,

we have
Z Z
dδj (t) d
= x2j |ψ(x, t)|2 dx = x2j (ψ ∂t ψ ∗ + ψ ∗ ∂t ψ) dx
dt dt Rd Rd
Z
i
= x2 (ψ ∗ 4 ψ − ψ 4 ψ ∗ ) dx
2 Rd j
Z


=i xj ψ ∂xj ψ ∗ − ψ ∗ ∂xj ψ dx. (4.33)
Rd

Similarly, differentiate (4.33) with respect to t, notice (4.1), integrate by

parts, we have
Z
d2 δj (t)  
=i xj ∂t ψ ∂xj ψ ∗ + ψ ∂xj t ψ ∗ − ∂t ψ ∗ ∂xj ψ − ψ ∗ ∂xj t ψ dx
dt2 d
Z R


= 2ixj ∂t ψ ∂xj ψ ∗ − ∂t ψ ∗ ∂xj ψ + i (ψ ∗ ∂t ψ − ψ ∂t ψ ∗ ) dx
d
ZR h

1
= −xj ∂xj ψ ∗ 4 ψ + ∂xj ψ 4 ψ ∗ − (ψ ∗ 4 ψ + ψ 4 ψ ∗ )
Rd 2


+2xj V (x) ψ ∂xj ψ ∗ + ψ ∗ ∂xj ψ + 2V (x)|ψ|2 + 2β|ψ|2σ+2

i
+2βxj |ψ|2σ ψ ∂xj ψ ∗ + ψ ∗ ∂xj ψ dx
Z h
= 2|∂xj ψ|2 − |∇ψ|2 − |ψ|2 ∂xj (2xj V (x)) + |∇ψ|2
Rd
2β i
− |ψ|2σ+2 + 2V (x)|ψ|2 + 2β|ψ|2σ+2 dx
σ+1
Z  
2σβ
= 2|∂xj ψ|2 + |ψ|2σ+2 − 2xj |ψ|2 ∂xj (V (x)) dx. (4.34)
Rd σ+1

Thus we obtain the desired equality (4.30). Setting t = 0 in (4.27) and

(4.33), we get (4.31) and (4.32) respectively. 
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 155

Lemma 4.2: When V (x) ≡ 0 in the NLSE (4.1), we have

Z
d2 δV 2β(dσ − 2)
= 4E(ψ0 ) + |ψ 2σ+2 dx. (4.35)
dt2 σ+1 Rd

Proof: Sum (4.30) from j = 1 to d, we get

Xd d Z  
d2 δV (t) d2 δj (t) X 2 2σβ 2σ+2
= = 2|∂xj ψ| + |ψ| dx
dt2 j=1
dt2 j=1 R
d σ+1
Z  
2 2dσβ 2σ+2
= 2|∇ψ| + |ψ| dx
Rd σ+1
Z
2β(σd − 2)
= 4E + |ψ|2σ+2 dx. (4.36)
σ+1 R d

Here we use conservation of energy of the NLSE. 

From this lemma, when V (x) ≡ 0 and at critical dimension, i.e. dσ −2 =
0, (4.35) reduces to
d2 δV
= 4E, (4.37)
dt2
leading to
δV (t) = 2Et2 + δV0 (0)t + δV (0). (4.38)
When the external potential V (x) is chosen as harmonic oscillator (3.13)
and σ = 1 in (4.1), we have

Lemma 4.3: (i) In 1D without interaction, i.e. d = 1 and β = 0 in (4.1),

we have
  (1)
E(ψ0 ) (0) E(ψ0 ) δx
δx (t) = + δ x − cos(2γx t) + sin(2γx t), t ≥ 0.
γx2 γx2 2γx
(4.39)
(ii) In 2D with radial symmetry, i.e. d = 2 and γx = γy := γr in (4.1),
for any initial data ψ0 (x, y) in (4.2), we have
  (1)
E(ψ0 ) (0) E(ψ0 ) δr
δr (t) = + δ r − cos(2γ r t) + sin(2γr t), t ≥ 0,
γr2 γr2 2γr
(4.40)
where
δr (t) = δx (t) + δy (t),
δr(0) := δr (0) = δx (0) + δy (0),
δr(1) := δr0 (0) = δx0 (0) + δy0 (0).
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

156 W. Bao

Furthermore, when d = 2 and γx = γy in (4.1) and the initial data ψ0 (x)

in (4.2) satisfying

ψ0 (x, y) = f (r)eimθ with m∈Z and f (0) = 0 when m 6= 0, (4.41)

we have for t ≥ 0
1
δx (t) = δy (t) = δr (t)
2
  (1)
E(ψ0 ) (0) E(ψ0 ) δx
= + δ x − cos(2γx t) + sin(2γx t). (4.42)
2γx2 2γx2 2γx
(iiii) For all other cases, we have for t ≥ 0
! (1)
E(ψ0 ) (0) E(ψ0 ) δj
δj (t) = + δ j − cos(2γx j t) + sin(2γxj t) + gj (t),
γx2j γx2j 2γxj
(4.43)
where gj (t) is a solution of the following problem

d2 gj (t) dgj (0)

+ 4γx2j gj (t) = fj (t), gj (0) = = 0, (4.44)
dt2 dt
with
Z h i
fj (t) = 2|∂xj ψ|2 − 2|∇ψ|2 − β|ψ|4 + (2γx2j x2j − 4V (x))|ψ|2 dx
Rd

satisfying

|fα (t)| < 4Eβ (ψ0 ), t ≥ 0.

Proof: (i) From (4.30) with d = 1 and β1 = 0, we have

d2 δx (t)
= 4E(ψ0 ) − 4γx2 δx (t), t > 0, (4.45)
dt2
δx (0) = δx(0) , δx0 (0) = δx(1) . (4.46)

Thus (4.39) is the unique solution of this ordinary differential equation

(ODE).
(ii). From (4.30) with d = 2, we have
Z
d2 δx (t) 2


2
= −2γ δ
x x (t) + 2|∂x ψ|2 + β|ψ|4 dx, (4.47)
dt Rd
Z
d2 δy (t) 2


2
= −2γy yδ (t) + 2|∂y ψ|2 + β|ψ|4 dx. (4.48)
dt Rd
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 157

Sum (4.47) and (4.48), notice (4.4) and γx = γy , we have the ODE for δr (t):
d2 δr (t)
= 4E(ψ0 ) − 4γx2 δr (t), t > 0, (4.49)
dt2
δr (0) = 2δx(0) , δr0 (0) = 2δx(1) . (4.50)
Thus (4.40) is the unique solution of the second order ODE (4.49) with
the initial data (4.50). Furthermore, when the initial data ψ0 (x) in (4.2)
satisfies (4.41), due to the radial symmetry, the solution ψ(x, t) of (4.1)-
(4.2) satisfies
ψ(x, y, t) = g(r, t)eimθ with g(r, 0) = f (r). (4.51)
This implies
Z Z ∞ Z 2π
δx (t) = x2 |ψ(x, y, t)|2 dx = r2 cos2 θ |g(r, t)|2 r dθdr
R2 0 0
Z ∞ Z ∞ Z 2π
=π r2 |g(r, t)|2 r dr = r2 sin2 θ |g(r, t)|2 r dθdr
0 0 0
Z
2 2
= y |ψ(x, y, t)| dx = δy (t), t ≥ 0. (4.52)
R2

Thus the equality (4.42) is a combination of (4.52) and (4.40).

(iii). From (4.30), notice the energy conservation (4.4) of the GPE (4.1),
we have
d2 δj (t)
= 4E(ψ0 ) − 4γx2j δj (t) + fj (t), t ≥ 0. (4.53)
dt2
Thus (4.43) is the unique solution of this ODE (4.53). 

5. Plane and solitary wave solutions of NLSE

For simplicity, we assume V (x) ≡ 0, d = 1 and σ = 1 in this section. In
this case, the NLSE (4.1) collapses to
1
iψt = − ψxx + β|ψ|2 ψ. (5.1)
2
To find the plane wave solution of (5.1), we take the ansatz
ψ = Aei(kx−ωt) , (5.2)
where A, ω and k are amplitude, angular frequency and wavenumber re-
spectively. Plugging (5.2) into (5.1), we get the dispersive relation
1 2
ω= k + β|A|2 (5.3)
2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

158 W. Bao

This implies that the dispersive relation depends on wavenumber and am-
plitude. Define the group velocity
dω
cg ≡ = k. (5.4)
dk
So the NLSE has the plane wave solution (5.2) provided the dispersive rela-
tion is satisfied. In fact, (5.3) can be viewed as zeroth-order approximation
of the NLSE (5.1), and (5.2) can be viewed as zeroth-order solution of the
NLSE (5.1).
To find the solitary wave solution, we take the ansatz

ψ = φ(ξ)ei(kx−ωt) , ξ = x − cg t, (5.5)

where φ is a real-valued function. Plugging (5.5) into (5.1), we get

1 d2 φ dφ
+ (ω − k 2 /2)φ − βφ3 + i(k − cg ) = 0. (5.6)
2 dξ 2 dξ
This implies

d2 φ
− + γφ + 2βφ3 = 0, γ = k 2 − 2ω > 0; cg = k. (5.7)
dξ 2
When β < 0, we have a solution for (5.7)
r r 
γ γ
φ(ξ) = ± dn (ξ − ξ0 ), k , (5.8)
−β(2 − k 2 ) 2 − k2

where dn is the Jacobian elliptic function. Letting k → 1, we have

r
γ √
φ(ξ) = ± sech γ(ξ − ξ0 ). (5.9)
−β

Thus we get a solitary wave solution for the NLSE (5.1) with β < 0:
r
γ √
ψ(x, t) = sech γ(x − t − x0 )ei[x−(1−γ)t/2] , (5.10)
−β

where γ > 0 is a constant.

For β > 0, one can get a traveling wave in a similar manner.

6. Existence/blowup results of NLSE

For simplicity, in this section, we assume V (x) ≡ 0 in (4.1).
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 159

6.1. Integral form

When β = 0 in (4.1), the free Schrödinger equation is solved as
ψ(x, t) = U (t)ψ0 (x), x ∈ Rd , t ≥ 0, (6.1)
where free Schrödinger operator U (t) = eit4/2 given by
 d/2 Z
1 |x−x0 |2
U (t)ψ0 (x) = ei 4t ψ0 (x0 ) dx0 (6.2)
4πit Rd

defines a unitary transformation group in L2 .

Theorem 6.1: (Decay estimates) For conjugate p and p0 ( p1 + p10 = 1),

with 2 ≤ p ≤ ∞, and t 6= 0, the transformation U (t) maps continuously
0
Lp (Rd ) into Lp (Rd ) and
1 1
kU (t)ψ0 kLp ≤ (4π|t|)−d( 2 − p ) kψ0 kLp0 . (6.3)

Proof (scratch): Use the conservation of L2 -norm kψ(t)kL2 = kψ0 kL2 , the
estimate |ψ(x, t)| ≤ (4π|t|)−d/2 kψ0 kL1 and the Riesz-Thorin interpolation
theorem. 

When β 6= 0 in (4.1), the problem is conveniently rewritten in the

integral form
Z t
ψ(t) = U (t)ψ0 − iβ U (t − t0 )|ψ(t0 )|2σ ψ(t0 ) dt0 . (6.4)
0

6.2. Existence results

Based on a fixed point theorem to (6.4), the following existence results for
NLSE is proved [113]:
2
Theorem 6.2: (Solution in H 1 ) For 0 ≤ σ < d−2 (no condition on σ
when d = 1 or 2) and an initial condition ψ0 ∈ H 1 (Rd ), there exists, lo-
cally in time, a unique maximal solution ψ in C((−T ∗ , T ∗ ), H 1 (Rd )), where
maximal means that if T ∗ < ∞, then kψkH 1 → ∞ as t approaches T ∗ . In
addition, ψ satisfies the normalization and energy (or Hamiltonian) con-
servation laws
Z
N (ψ) ≡ |ψ|2 dx = N (ψ0 ), (6.5)
R d
Z  
1 β
E(ψ) ≡ |∇ψ|2 + |ψ|2σ+2 dx = E(ψ0 ), (6.6)
Rd 2 σ+1
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

160 W. Bao

and depends continuously on the initial condition ψ in H 1 .

P
If in addition, the initial condition ψ0 belongs to the space = {f, f ∈
1 d 2 d 1
H (R ), |xf (x)| ∈ L (R )} of the functions in H with finite variance, the
P
above
R maximal solution belongs to C((−T ∗ , T ∗ ), ). The variance δV (t) =
Rd
|x|2 |ψ|2 dx belongs to C 2 (−T ∗ , T ∗ ) and satisfies the identity
Z
d2 δV 2β(dσ − 2)
= 4E(ψ 0 ) + |ψ|2σ+2 dx. (6.7)
dt2 σ+1 R d

Theorem 6.3: (Solution in L2 ) For 0 ≤ σ < 2d and an initial condition

ψ0 ∈ L2 (Rd ), there exist a unique solution ψ in C((−T ∗ , T ∗ ), L2 (Rd )) ∩
Lq ((−T ∗ , T ∗ ), L2σ+2 (Rd )) with q = 4(σ+1) 2
dσ , satisfying the L -norm conser-
vation law (6.5).

Theorem 6.4: (Global existence in H 1 ) Assume 0 ≤ σ < 2/(d−2) if β < 0

(attractive nonlinearity), or 0 ≤ σ < 2/d if β > 0 (repulsive nonlinearity).
For any ψ ∈ H 1 (Rd ), there exists a unique solution ψ in C(R, H 1 (Rd )). It
satisfies the conservation laws (6.5) and (6.6) and depends continuously on
initial conditions in H 1 (Rd ).

Theorem 6.5: (Global existence in L2 ) For 0 ≤ σ < 2/d and ψ0 ∈ L2 (Rd ),

there exists a unique solution ψ in C(R, L2 (Rd )) ∩ Lqloc (R, L2σ+2 (Rd )) with
q = 4(σ + 1)/dσ that satisfies the L2 -norm conservation (6.5) and depends
continuously on initial conditions in L2 .

6.3. Finite time blowup results

Classical blowup results are based on the “variance identity”, also known
as the “viral
R theorem”, and “uncertainty principle”. Define the variance
δV (t) = Rd |x|2 |ψ|2 dx, we have the identity
Z
d2 2β(dσ − 2)
δ (t) = 4E + |ψ|2σ+2 dx. (6.8)
dt2 V σ+1 R d

Theorem 6.6: Suppose that β < 0 and dσ ≥ 2. Consider an initial condi-

tion ψ0 ∈ H 1 with δV (0) finite that satisfies one of the conditions below:
(i) E(ψ0 ) < 0, R
(ii) E(ψ0 ) = 0 and δV0 (0) = 2 Rep Rd ψ0∗ (x · ∇ψ0 )dxp
< 0,
 0 
(iii) E(ψ0 ) > 0 and δV (0) ≥ 2 2E(ψ0 )δV (0) = 2 2E(ψ0 )kxψ0 kL2 .
Then, there exists a time t∗ < ∞ such that
lim k∇ψkL2 = ∞ and lim kψkL∞ = ∞. (6.9)
t→t∗ t→t∗
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 161

Proof: If β < 0 and dσ ≥ 2,

d2
δ (t) ≤ 4E, (6.10)
dt2 V
and by time integration,

δV (t) ≤ 2Et2 + δV0 (0)t + δV (0). (6.11)

Under any of the hypotheses (i)-(iii) of the above theorem, there exists
a time t0 such that the right-hand side of (6.11) vanishes, and thus also
t1 ≤ t0 such that

lim δV (t) = 0. (6.12)

t→t1

Furthermore, from the equality

Z Z Z
1 1
|f |2 dx = (∇ · x)|f |2 dx = − x · ∇(|f |2 ) dx, (6.13)
Rd d Rd d Rd
one gets the “uncertainty principle”
2
kf k2L2 ≤ k∇f kL2 kxf kL2 . (6.14)
d
When this inequality is applied to a solution ψ, one gets from (6.14)
and from the conservation of kψk2L2 , that there exists a time t∗ ≤ t1
such that limt→t∗ k∇ψkL2 = ∞. The conservation of E then ensures that
limt→t∗ kψk2σ+2 2
L2σ+2 = ∞, and since kψkL2 is conserved, this implies that
limt→t∗ kψkL∞ = ∞. 

7. WKB expansion and quantum hydrodynamics

In this section, we consider the NLSE in semiclassical regime
ε2
iεψtε = − 4 ψ ε + V (x)ψ ε + f (|ψ ε |2 )ψ ε , x ∈ Rd , t ≥ 0, (7.1)
2
ψ ε (x, 0) = ψ0ε (x), x ∈ Rd , (7.2)

where 0 < ε  1 is the (scaled) Planck constant, f (ρ) is a given real-valued

function; and find its semiclassical limit by using WKB expansion.
Suppose that the initial datum ψ0ε in (7.2) is rapidly oscillating on the
scale ε, given in WKB form:
 
i
ψ0ε (x) = A0 (x) exp S0 (x) , x ∈ Rd , (7.3)
ε
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

162 W. Bao

where the amplitude A0 and the phase S0 are smooth real-valued functions.
Plugging the radial-representation of the wave-function
   
i ε p i ε
ψ ε (x, t) = Aε (x, t) exp S (x, t) = ρε (x, t) exp S (x, t) (7.4)
ε ε
into (7.1), one obtains the following quantum hydrodynamic (QHD) form
of the NLSE for ρε = |Aε |2 , Jε = ρε ∇S ε [58, 41, 79]
ρεt + div Jε = 0, (7.5)
 ε ε
 2
J ⊗J ε
Jεt + div + ∇P (ρε ) + ρε ∇V = div(ρε ∇2 log ρε ); (7.6)
ρε 4
with initial data
ρε (x, 0) = ρε0 (x) = |A0 (x)|2 , Jε (x, 0) = ρε0 (x) ∇S0 (x) = |A0 (x)|2 ∇S0 (x),
(7.7)
(see Grenier [66], Jüngel [79, 80], for mathematical analyses of this system).
Here the hydrodynamic pressure P (ρ) is related to the nonlinear potential
f (ρ) by
Z ρ
P (ρ) = ρf (ρ) − f (s) ds, (7.8)
0
0
i.e. f is the enthalpy. Letting ε → 0+, one obtains formally the following
Euler system
ρt + div J = 0, (7.9)
 
J⊗J
Jt + div + ∇P (ρ) + ρ∇V = 0. (7.10)
ρ
which can be viewed formally as the dispersive (semiclassical) limit of the
NLSE (7.1). In the case f 0 > 0 we expect (7.9), (7.10) to be the ‘rigorous’
semiclassical limit of (7.1) as long as caustics do not occur, i.e. in the pre-
breaking regime. After caustics the dispersive behavior of the NLSE takes
over and (7.9), (7.10) is not correct any more.

8. Wigner transform and semiclassical limit

In this section, we consider the linear Schrödinger equation in semiclassical
regime
ε2
iεψtε = − 4 ψ ε + V (x)ψ ε , x ∈ Rd , t ≥ 0, (8.1)
2
ψ ε (x, 0) = ψ0ε (x), x ∈ Rd , (8.2)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 163

and find its semiclassical limit by using Wigner transformation.

Let f, g ∈ L2 (Rd ). Then the Wigner-transform of (f, g) on the scale
ε > 0 is defined as the phase-space function:
Z 
1 ∗ ε   ε  iy·ξ
wε (f, g)(x, ξ) = f x + y g x − y e dy (8.3)
(2π)d Rd 2 2
(cf. [58], [91] for a detailed analysis of the Wigner-transform). It is well-
known that the estimate
kwε (f, g)kE ∗ ≤ kf kL2 (Rd ) kgkL2(Rd ) (8.4)
holds, where E is the Banach space
E := {φ ∈ C0 (Rdx × Rdξ ) : (Fξ→v φ)(x, v) ∈ L1 (Rdv ; C0 (Rdx ))},
Z
kφkE := sup |(Fξ→v φ)(x, v)| dv,
Rd d
v x∈Rx

(cf.
R [91]). E ∗ denotes the dual space of E and (Fξ→v σ)(v) :=
Rd
σ(ξ) e−iv·ξ dξ the Fourier transform.
ξ
Now let ψ ε (t) be the solution of the linear Schrödinger equation (8.1),
(8.2) and denote
wε (t) := wε (ψ ε (t), ψ ε (t)). (8.5)
Then wε satisfies the Wigner equation
wtε + ξ · ∇x wε + Θε [V ]wε = 0, (x, ξ) ∈ Rdx × Rdξ , t ∈ R, (8.6)
ε ε
w (t = 0) = w (ψ0ε , ψ0ε ), (8.7)
where Θε [V ] is the pseudo-differential operator:
Z
i V (x + 2ε α) − V (x − 2ε α) ε
Θε [V ]wε (x, ξ, t) := d
ŵ (x, α, t)eiα·ξ dα,
(2π) Rdα ε
(8.8)
ε
here ŵ stands for the Fourier-transform
Z
ε
Fξ→α w (x, ·, t) := wε (x, ξ, t)e−iα·ξ dξ.
Rd
ξ

The main advantage of the formulation (8.6), (8.7) is that the semiclassical
limit ε → 0 can easily be carried out. Taking ε to 0 gives the Vlasov-equation
( or Liouville equation):
wt0 + ξ · ∇x w0 − ∇x V (x) · ∇ξ w0 = 0, (x, ξ) ∈ Rdx × Rdξ , t ∈ R, (8.9)
0 ε
w (t = 0) = w0 := lim w (ψ0ε , ψ0ε ), (8.10)
ε→0
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

164 W. Bao

(cf. [58], [91]), where

w0 := lim wε .
ε→0

Here, the limits hold in an appropriate weak sense (i.e. in E ∗ − ω ∗ ) and have
to be understood for subsequences (εnk ) → 0 of sequence εn . We recall that
w0 , w0 (t) are positive bounded measures on the phase-space Rdx × Rdξ .
When the initial Wigner distribution has the high frequency form

w0 = |A0 (x)|2 δ(ξ − ∇S0 (x)), (8.11)

then it is easy to see that the solution of (8.9) is given that

w0 (x, ξ, t) = |A(x, t)|2 δ(ξ − ∇S(x, t)), (8.12)

where A(x, t) is the solution of the transport equation

(|A|2 )t + ∇ · (|A|2 ∇S) = 0, |A(x, 0)|2 = |A0 (x)|2 (8.13)

and S(x, t) is the solution of the Eiconal equation

1
St + |∇S|2 + V (x) = 0, S(x, 0) = S0 (x). (8.14)
2
Define the moments
Z
ρ(x, t) = w0 (x, ξ, t) dξ, (8.15)
Rd
ξ
Z
J(x, t) = ξw0 (x, ξ, t) dξ. (8.16)
Rd
ξ

Then ρ and J satisfy the pressureless Euler equation:

ρt + div J = 0, (8.17)
 
J⊗J
Jt + div + ρ∇V = 0; (8.18)
ρ
with initial data

ρ(x, 0) = ρ0 (x) = |A0 (x)|2 , J(x, 0) = ρ0 (x) ∇S0 (x) = |A0 (x)|2 ∇S0 (x).
(8.19)

9. Ground, excited and central vortex states of GPE

For simplicity, in this section, we take σ = 1 and the potential V (x) as a
harmonic oscillator (4.1), i.e. NLSE is considered in terms of BEC setup.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 165

9.1. Stationary states

To find a stationary solution of (4.1), we write
ψ(x, t) = e−iµt φ(x), (9.1)
where µ is the chemical potential and φ is a function independent of time.
Inserting (9.1) into (4.1) gives the following equation for φ(x)
1
µ φ(x) = − 4 φ(x) + V (x) φ(x) + β|φ(x)|2 φ(x), x ∈ Rd , (9.2)
2
under the normalization condition
Z
2
kφk = |φ(x)|2 dx = 1. (9.3)
Rd
This is a nonlinear eigenvalue problem under a constraint and any eigen-
value µ can be computed from its corresponding eigenfunction φ by
Z  
1 2 2 4
µ = µ(φ) = |∇φ(x)| + V (x) |φ(x)| + β |φ(x)| dx
Rd 2
Z
β 4
= E(φ) + |φ(x)| dx. (9.4)
R d 2
In fact, the eigenfunctions of (9.2) under the constraint (9.3) are equivalent
to the critical points of the energy functional over the unit sphere S =
{φ | kφk = 1, E(φ) < ∞}. Furthermore, as noted in [6], they are equivalent
to the steady state solutions of the following continuous normalized gradient
flow (CNGF):
1 µ(φ)
∂t φ = 4 φ − V (x)φ − β |φ|2 φ + φ, x ∈ Rd , t ≥ 0, (9.5)
2 kφ(·, t)k2
φ(x, 0) = φ0 (x), x ∈ Rd with kφ0 k = 1. (9.6)

9.2. Ground state

The BEC ground state wave function φg (x) is found by minimizing the
energy functional E(φ) over the unit sphere S = {φ | kφk = 1, E(φ) < ∞}:
(V) Find (µgβ , φgβ ∈ S) such that
Z
g g g g g β g 2
Eβ = E(φβ ) = min E(φ), µβ = µ(φβ ) = E(φβ ) + |φβ | dx.
φ∈S R 2
d

(9.7)
In the case of a defocusing condensate, i.e. β ≥ 0, the energy functional
E(φ) is positive, coercive and weakly lower semicontinuous on S, thus the
existence of a minimum follows from the standard theory. For understanding
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

166 W. Bao

the uniqueness question note that E(αφgβ ) = E(φgβ ) for all α ∈ C with |α| =
1. Thus an additional constraint has to be introduced to show uniqueness.
For non-rotating BECs, the minimization problem (9.7) has a unique real
valued nonnegative ground state solution φgβ (x) > 0 for x ∈ Rd [87].
When β = 0, the ground state solution is given explicitly [5]

1  γx , d = 1,
µg0 = γx + γy , d = 2, (9.8)
2
γx + γy + γz , d = 3,

 γx1/4 e−γx x /2 ,
2
d = 1,
g 1  2 2
φ0 (x) = d/4 (γx γy )1/4 e−(γx x +γy y )/2 , d = 2, (9.9)
π 
 2 2 2
(γx γy γz )1/4 e−(γx x +γy y +γz z )/2 , d = 3.
In fact, this solution can be viewed as an approximation of the ground
state for weakly interacting condensate, i.e. |βd |  1. For a condensate
with strong repulsive interaction, i.e. β  1 and γα = O(1) (α = x, y, z),
the ground state can be approximated by the Thomas-Fermi approximation
in this regime [5]:
(q
(µTF
β − V (x))/β, V (x) < µTF
β ,
φTF
β (x) = (9.10)
0, otherwise,

(3βγx /2)2/3 , d = 1,
TF 1 1/2
µβ = (4βγx γy /π) , d = 2, (9.11)
2 2/5
(15βγx γy γz /4π) , d = 3.
Due to φTFβ is not differentiable at V (x) = µTF β , as noticed in [5, 8],
TF
E(φβ ) = ∞ and µ(φTF β ) = ∞. This shows that we can’t use (4.4) to define
the energy of the Thomas-Fermi approximation (9.10). How to define the
energy of the Thomas-Fermi approximation is not clear in the literatures.
Using (9.4), (9.11) and (9.10), here we present a way to define the energy
of the Thomas-Fermi approximation (9.10):
Z Z  
TF TF β TF 4 TF 2 β TF 4
Eβ = µβ − |φβ (x)| dx = V (x)|φβ (x)| + |φβ (x)| dx
Rd 2 Rd 2
d + 2 TF
= µ , d = 1, 2, 3. (9.12)
d+4 β
From the numerical results in [6, 5], when γx = O(1), γy = O(1) and γz =
O(1), we can get
Eβg − EβTF = E(φgβ ) − EβTF → 0, as βd → ∞.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 167

Any eigenfunction φ(x) of (9.2) under constraint (9.3) whose energy

E(φ) > E(φgβ ) is usually called as excited states in physical literatures.

9.3. Central vortex states

To find central vortex states in 2D with radial symmetry, i.e. d = 2 and
γx = γy = 1 in (4.1), we write

ψ(x, t) = e−iµm t φm (x, y) = e−iµm t φm (r)eimθ , (9.13)

where (r, θ) is the polar coordinate, m 6= 0 is an integer and called as

index or winding number, µm is the chemical potential, and φm (r) is a real
function independent of time. Inserting (9.13) into (4.1) gives the following
equation for φm (r) with 0 < r < ∞
     
1 d d 1 2 m2
µm φm (r) = − r + r + 2 + β2 |φm |2 φm , (9.14)
2r dr dr 2 r
φm (0) = 0, lim φm (r) = 0. (9.15)
r→∞

under the normalization condition

Z ∞
2π |φm (r)|2 r dr = 1. (9.16)
0

In order to find the central vortex state φm m

β (x, y) = φβ (r)e
imθ
with index
m
m, we find a real nonnegative function φβ (r) which minimizes the energy
functional

E m (φ(r)) = E(φ(r)eimθ )
Z ∞   
m2
=π |φ0 (r)|2 + r2 + 2 |φ(r)|2 + β2 |φ(r)|4 rdr, (9.17)
0 r
R∞
over the set S0 = {φ | 2π 0 |φ(r)|2 r dr = 1, φ(0) = 0, E m (φ) < ∞}. The
existence and uniqueness of nonnegative minimizer for this minimization
problem can be obtained similarly as for the ground state [87]. Note that
the set Sm = {φ(r)eimθ | φ ∈ S0 } ⊂ S is a subset of the unit sphere, so
φm
β (r)e
imθ
is a minimizer of the energy functional Eβ over the set Sm ⊂ S.
2
When β2 = 0 in (4.1), φm 0 (r) =
√ 1 r|m| e−r /2 [6].
π|m|!
Similarly, in order to find central vortex line states in 3D with cylindrical
symmetry, i.e. d = 3 and γx = γy = 1 in (4.1), we write

ψ(x, t) = e−iµm t φm (x, y, z) = e−iµm t φm (r, z)eimθ , (9.18)

April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

168 W. Bao

where m 6= 0 is an integer and called as index, µm is the chemical potential,

and φm (r, z) is a real function independent of time. Inserting (9.18) into
(4.1) with d = 3 gives the following equation for φm (r, z)
» „ « „ « –
1 ∂ ∂ ∂2 1 2 m2 2 2 2
µm φm = − r − + r + + γ z z + β |φ
3 m | φm , (9.19)
2r ∂r ∂r 2 ∂z 2 2 r2
φm (0, z) = 0, lim φm (r, z) = 0, −∞ < z < ∞, (9.20)
r→∞
lim φm (r, z) = 0, 0 ≤ r < ∞, (9.21)
|z|→∞

under the normalization condition

Z ∞Z ∞
2π |φm (r, z)|2 r drdz = 1. (9.22)
0 −∞

In order to find the central vortex line state φm m

β (x, y, z) = φβ (r, z)e
imθ
with
m
index m, we find a real nonnegative function φβ (r, z) which minimizes the
energy functional

E m (φ(r, z)) = E(φ(r, z)eimθ ) (9.23)

Z ∞Z ∞    
m2
=π |∂r φ|2 + |∂z φ|2 + r2 + γz2 z 2 + 2 |φ|2 + β3 |φ|4 r drdz,
0 −∞ r
R∞R∞
over the set S0 = {φ | 2π 0 −∞ |φ(r, z)|2 r drdz = 1, φ(0, z) = 0, −∞ <
z < ∞, Eβm (φ) < ∞}. The existence and uniqueness of nonnegative min-
imizer for this minimization problem can be obtained similarly as for the
ground state [87]. Note that the set Sm = {φ(r, z)eimθ | φ ∈ S0 } ⊂ S
is a subset of the unit sphere, so φm β (r, z)e
imθ
is a minimizer of the en-
ergy functional Eβ over the set Sm . When β3 = 0 in (3.11), φm 0 (r, z) =
γz1/4 |m| −(r 2
+γ z 2
)/2
3/4
√ r e z
[6].
π |m|!

9.4. Variation of stationary states over the unit sphere

For the stationary states of (9.2), we have the following lemma:

Lemma 9.1: Suppose β = 0 and V (x) ≥ 0 for x ∈ Rd , we have

(i) The ground state φg is a global minimizer of E(φ) over S.
(ii) Any excited state φj is a saddle point of E(φ) over S.

Proof: Let φe be an eigenfunction of the eigenvalue problem (9.2) and (9.3).

The corresponding eigenvalue is µe . For any function φ such that E(φ) < ∞
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 169

and kφe + φk = 1, notice (9.3), we have that

Z
kφk2 = kφ + φe − φe k2 = kφ + φe k2 − kφe k2 − (φ∗ φe + φφ∗e ) dx
Rd
Z
=− (φ∗ φe + φφ∗e ) dx. (9.24)
Rd
From (4.4) with ψ = φe + φ and β = 0, notice (9.3) and (9.24), integration
by parts, we get
Z  
1
E(φe + φ) = |∇φe + ∇φ|2 + V (x)|φe + φ|2 dx
Rd 2
Z   Z  
1 1
= |∇φe |2 + V (x)|φe |2 + |∇φ|2 + V (x)|φ|2 dx
Rd 2 Rd 2
Z  ∗   
1 1
+ − 4 φe + V (x)φe φ + − 4 φe + V (x)φe φ∗ dx
Rd 2 2
Z
= E(φe ) + E(φ) + (µe φ∗e + µe φe φ∗ ) dx
Rd
= E(φe ) + E(φ) − µe kφk2
= E(φe ) + [E(φ/kφk) − µe ] kφk2 . (9.25)
(i) Taking φe = φg and µe = µg in (9.25) and noticing E(φ/kφk) ≥
E(φg ) = µg for any φ 6= 0, we get immediately that φg is a global minimizer
of E(φ) over S.
(ii). Taking φe = φj and µe = µj in (9.25), since E(φg ) < E(φj ) and it
is easy to find an eigenfunction φ of (9.2) such that E(φ) > E(φj ), we get
immediately that φj is a saddle point of the functional E(φ) over S. 

9.5. Conservation of angular momentum expectation

Another important quantity for studying dynamics of BEC in 2&3d, espe-
cially for measuring the appearance of vortex, is the angular momentum
expectation value defined as
Z
hLz i(t) := ψ ∗ (x, t)Lz ψ(x, t) dx, t ≥ 0, d = 2, 3, (9.26)
Rd
where Lz = i (y∂x − x∂y ) is the z-component angular momentum.
Lemma 9.2: Suppose ψ(x, t) is the solution of the problem (4.1), (4.2)
with d = 2 or 3, then we have
Z
dhLz i(t)

= γx2 − γy2 δxy (t), δxy (t) = xy|ψ(x, t)|2 dx, t ≥ 0.
dt Rd
(9.27)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

170 W. Bao

This implies that, at least in the following two cases, the angular momentum
expectation is conserved:
i) For any given initial data ψ0 (x) in (4.2), if the trap is radial sym-
metric in 2d, and resp., cylindrical symmetric in 3d, i.e. γx = γy ;
ii) For any given γx > 0 and γy > 0 in (3.13), if the initial data ψ0 (x)
in (4.2) is either odd or even in the first variable x or second variable y.

Proof: Differentiate (9.26) with respect to t, notice (4.1), integrate by parts,

we have
Z
dhLz i(t)
= [(iψt∗ ) (y∂x − x∂y )ψ + ψ ∗ (y∂x − x∂y )(iψt )] dx
dt Rd
Z  
1 2 ∗
= ∇ ψ − V (x)ψ ∗ − β|ψ|2 ψ ∗ (y∂x − x∂y )ψ
Rd 2
 
1
+ψ ∗ (y∂x − x∂y ) − ∇2 ψ + V (x)ψ + β|ψ|2 ψ dx
2
Z
1 2 ∗ 
= ∇ ψ (y∂x − x∂y )ψ − ψ ∗ (y∂x − x∂y )∇2 ψ dx
Rd 2
Z
 ∗

+ ψ (y∂x − x∂y ) V (x)ψ + β|ψ|2 ψ
Rd


− V (x)ψ ∗ + β|ψ|2 ψ ∗ (y∂x − x∂y )ψ dx
Z


= |ψ|2 (y∂x − x∂y ) V (x) + β|ψ|2 dx
d
ZR Z
= |ψ|2 (y∂x − x∂y )Vd (x) dx = |ψ|2 (γx2 − γy2 )xy dx
Rd Rd
Z
2 2
= (γx − γy ) xy|ψ|2 dx, t ≥ 0. (9.28)
Rd

For case i), since γx = γy , we get the conservation of hLz i immediately

from the first order ODE:

dhLz i(t)
= 0, t ≥ 0. (9.29)
dt

For case ii), we know the solution ψ(x, t) is either odd or even in the
first variable x or second variable y due to the assumption of the initial
data and symmetry of V (x). Thus |ψ(x, t)| is even in either x or y, which
immediately implies that hLz i satisfies the first order ODE (9.29). 
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 171

10. Numerical methods for computing ground states of

GPE
In this section, we present the continuous normalized gradient flow (CNGF),
prove its energy diminishing and propose its semi-discretization for com-
puting ground states in BEC. For simplicity, we take σ = 1 in (4.1).

10.1. Gradient flow with discrete normalization (GFDN)

Various algorithms for computing the minimizer of the energy functional
E(φ) under the constraint (9.3) have been studied in the literature. For
instance, second order in time discretization scheme that preserves the nor-
malization and energy diminishing properties were presented in [2, 6]. Per-
haps one of the more popular technique for dealing with the normalization
constraint (9.3) is through the following construction: choose a time se-
quence 0 = t0 < t1 < t2 < · · · < tn < · · · with 4tn = tn+1 − tn > 0
and k = maxn≥0 4tn . To adapt an algorithm for the solution of the usual
gradient flow to the minimization problem under a constraint, it is natural
to consider the following splitting (or projection) scheme which was widely
used in physical literatures [6] for computing the ground state solution of
BEC:
1 δE(φ) 1
φt = − = 4 φ − V (x)φ − β |φ|2 φ,
2 δφ 2
x ∈ Ω, tn < t < tn+1 , n ≥ 0, (10.1)
4 φ(x, t−
n+1 )
φ(x, tn+1 ) = φ(x, t+
n+1 ) = , x ∈ Ω, n ≥ 0, (10.2)
kφ(·, t−
n+1 )k
φ(x, t) = 0, x ∈ Γ = ∂Ω, φ(x, 0) = φ0 (x), x ∈ Ω; (10.3)
where φ(x, t±n ) = limt→t±
n
φ(x, t), kφ0 k = 1 and Ω ⊂ Rd . In fact, the gra-
dient flow (10.1) can be viewed as applying the steepest decent method to
the energy functional E(φ) without constraint and (10.2) then projects the
solution back to the unit sphere in order to satisfying the constraint (9.3).
From the numerical point of view, the gradient flow (10.1) can be solved via
traditional techniques and the normalization of the gradient flow is simply
achieved by a projection at the end of each time step.

10.2. Energy diminishing of GFDN

Let
φ(·, t)
φ̃(·, t) = , tn ≤ t ≤ tn+1 , n ≥ 0. (10.4)
kφ(·, t)k
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

172 W. Bao

For the gradient flow (10.1), it is easy to establish the following basic facts:

Lemma 10.1: Suppose V (x) ≥ 0 for all x ∈ Ω, β ≥ 0 and kφ0 k = 1, then

(i). kφ(·, t)k ≤ kφ(·, tn )k = 1 for tn ≤ t ≤ tn+1 , n ≥ 0.
(ii). For any β ≥ 0,

E(φ(·, t)) ≤ E(φ(·, t0 )), tn ≤ t0 < t ≤ tn+1 , n ≥ 0. (10.5)

(iii). For β = 0,

E(φ̃(·, t)) ≤ E(φ̃(·, tn )), tn ≤ t ≤ tn+1 , n ≥ 0. (10.6)

Proof: (i) and (ii) follows the standard techniques used for gradient flow.
As for (iii), from (4.4) with ψ = φ̃ and β = 0, (10.1), (10.3) and (10.4),
integration by parts and Schwartz inequality, we obtain
Z  
d d |∇φ|2 V (x)φ2
E(φ̃) = + dx
dt dt Ω 2kφk2 kφk2
Z     Z  
∇φ · ∇φt V (x)φ φt d 2 |∇φ|2 V (x)φ2
=2 + dx − kφk + dx
Ω 2kφk2 kφk2 dt Ω 2kφk
4 kφk4
Z  1   Z 1 2 2
− 2 4 φ + V (x)φ φt d 2 2 |∇φ| + V (x)φ
=2 dx − kφk dx
Ω kφk2 dt Ω kφk4
 2 " Z 2 #
kφt k2 1 d 2 2 2 2
= −2 + kφk = φ φt dx − kφk kφt k
kφk2 2kφk4 dt kφk4 Ω

≤0, tn ≤ t ≤ tn+1 . (10.7)

This implies (10.6). 

Remark 10.2: The property (10.5) is often referred as the energy dimin-
ishing property of the gradient flow. It is interesting to note that (10.6)
implies that the energy diminishing property is preserved even with the
normalization of the solution of the gradient flow for β = 0, that is, for
linear evolution equations.

Remark 10.3: When β > 0, the solution of (10.1)-(10.3) may not preserve
the normalized energy diminishing property

E(φ̃(·, t)) ≤ E(φ̃(·, t0 )), 0 ≤ t0 < t ≤ t1

for any t1 > 0 [6].

From Lemma 10.1, we get immediately

April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 173

Theorem 10.4: Suppose V (x) ≥ 0 for all x ∈ Ω and kφ0 k = 1. For β = 0,

GFDN (10.1)-(10.3) is energy diminishing for any time step k and initial
data φ0 , i.e.

E(φ(·, tn+1 )) ≤ E(φ(·, tn )) ≤ · · · ≤ E(φ(·, 0)) = E(φ0 ), n = 0, 1, 2, · · · .

(10.8)

10.3. Continuous normalized gradient flow (CNGF)

In fact, the normalized step (10.2) is equivalent to solve the following ODE
exactly

φt (x, t) = µφ (t, k)φ(x, t), x ∈ Ω, tn < t < tn+1 , n ≥ 0, (10.9)

φ(x, t+
n) = φ(x, t−
n+1 ), x ∈ Ω; (10.10)

where
1
µφ (t, k) ≡ µφ (tn+1 , 4tn ) = − ln kφ(·, t− 2
n+1 )k , tn ≤ t ≤ tn+1 .
2 4 tn
(10.11)
Thus the GFDN (10.1)-(10.3) can be viewed as a first-order splitting
method for the gradient flow with discontinuous coefficients:
1
φt = 4 φ − V (x)φ − β |φ|2 φ + µφ (t, k)φ, x ∈ Ω, t ≥ 0, (10.12)
2
φ(x, t) = 0, x ∈ Γ, φ(x, 0) = φ0 (x), x ∈ Ω. (10.13)

Let k → 0, we see that

µφ (t) := lim µφ (t, k)

k→0+
Z  
1 1
= |∇φ(x, t)|2 + V (x)φ2 (x, t) + βφ4 (x, t) dx. (10.14)
kφ(·, t)k2 Ω 2
This suggests us to consider the following continuous normalized gradient
flow:
1
φt = 4 φ − V (x)φ − β |φ|2 φ + µφ (t)φ, x ∈ Ω, t ≥ 0, (10.15)
2
φ(x, t) = 0, x ∈ Γ, φ(x, 0) = φ0 (x), x ∈ Ω. (10.16)

In fact, the right hand side of (10.15) is the same as (9.2) if we view µφ (t)
as a Lagrange multiplier for the constraint (9.3). Furthermore for the above
CNGF, as observed in [6], the solution of (10.15) also satisfies the following
theorem:
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

174 W. Bao

Theorem 10.5: Suppose V (x) ≥ 0 for all x ∈ Ω, β ≥ 0 and kφ0 k = 1.

Then the CNGF (10.15)-(10.16) is normalization conservation and energy
diminishing, i.e.
Z
2
kφ(·, t)k = φ2 (x, t) dx = kφ0 k2 = 1, t ≥ 0, (10.17)
Ω
d 2
E(φ) = −2 kφt (·, t)k ≤ 0 , t ≥ 0, (10.18)
dt
which in turn implies
E(φ(·, t1 )) ≥ E(φ(·, t2 )), 0 ≤ t1 ≤ t2 < ∞.

Remark 10.6: We see from the above theorem that the energy diminishing
property is preserved in the continuous dynamic system (10.15).

Using argument similar to that in [88, 106], we may also get as t → ∞, φ

approaches to a steady state solution which is a critical point of the energy.
In non-rotating BEC, it has a unique real valued nonnegative ground state
solution φg (x) ≥ 0 for all x ∈ Ω [87]. We choose the initial data φ0 (x) ≥ 0
for x ∈ Ω, e.g. the ground state solution of linear Schrödinger equation
with a harmonic oscillator potential [5, 8]. Under this kind of initial data,
the ground state solution φg and its corresponding chemical potential µg
can be obtained from the steady state solution of the CNGF (10.15)-(10.16),
i.e.
Z
β
φg (x) = lim φ(x, t), x ∈ Ω, µg = µβ (φg ) = E(φg ) + |φg (x)|4 dx.
t→∞ 2 Ω
(10.19)

10.4. Semi-implicit time discretization

To further discretize the equation (10.1), we here consider the following
semi-implicit time discretization scheme:
φ̃n+1 − φn 1
= 4 φ̃n+1 − V (x)φ̃n+1 − β |φn |2 φ̃n+1 , x ∈ Ω, (10.20)
k 2
φ̃n+1 (x) = 0, x ∈ Γ, φn+1 (x) = φ̃n+1 (x)/kφ̃n+1 k , x ∈ Ω (10.21)
.
Notice that since the equation (10.20) becomes linear, the solution at
the new time step becomes relatively simple. In other words, in each discrete
time interval, we may view (10.20) as a discretization of a linear gradient
flow with a modified potential Ṽn (x) = V (x) + β|φn (x)|2 .
We now first present the following lemma:
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 175

Lemma 10.7: Suppose β ≥ 0 and V (x) ≥ 0 for all x ∈ Ω and kφn k = 1.

Then,
Z Z Z Z
|φ̃n+1 |2 dx ≤ φn φ̃n+1 dx, |φ̃n+1 |4 dx ≤ |φn |2 |φ̃n+1 |2 dx.
Ω Ω Ω Ω
(10.22)
Proof: Multiplying both sides of (10.20) by φ̃n+1 , integrating over Ω, and
applying integration by parts, we obtain
Z   Z  
1
|φ̃n+1 |2 − φn φ̃n+1 dx = −k |∇φ̃n+1 |2 + Ṽn (x)|φ̃n+1 |2 dx ≤ 0 ,
Ω Ω 2
which leads to the first inequality in (10.22). Similarly,
Z Z  2
 k 
|φ̃n+1 |2 |φn |2 dx = |φ̃n+1 |2 φ̃n+1 − 4 φ̃n+1 + k Ṽn (x)φ̃n+1  dx
Ω Ω 2
Z  
n+1 2 n+1 2 k n+1 n+1 n+1 2
= |φ̃ | |φ̃ | − 2 φ̃ 4 φ̃ + 2k Ṽn (x)|φ̃ | dx
Ω 2
Z  2
k 
+ |φ̃n+1 |2  4 φ̃n+1 − k Ṽn (x)φ̃n+1  dx
Ω 2
Z h i
= |φ̃n+1 |2 |φ̃n+1 |2 + 3k|∇φ̃n+1 |2 + 2k Ṽn (x)|φ̃n+1 |2 dx
Ω
Z  2
k 
+ |φ̃n+1 |2  4 φ̃n+1 − k Ṽn (x)φ̃n+1  dx
Ω 2
Z
≥ |φ̃n+1 |4 dx . (10.23)
Ω
This implies the second inequality in (10.22). 
Given a linear self-adjoint operator A in a Hilbert space H with inner
product (·, ·), and assume that A is positive definite in the sense that for
some positive constant c, (u, Au) ≥ c(u, u) for any u ∈ H. We now present
a simple lemma:
Lemma 10.8: For any k > 0, and (I + kA)u = v, we have
(u, Au) (v, Av)
≤ . (10.24)
(u, u) (v, v)
Proof: Since A is self-adjoint and positive definite, by Hölder inequality,
we have for any p, q ≥ 1 with p + q = pq, that
1/p 1/q
(u, Au) ≤ (u, u) (u, Aq u) ,
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

176 W. Bao

which leads to
1/2
1/2


(u, Au) ≤ (u, u) u, A2 u , (u, Au) u, A2 u ≤ (u, u) u, A3 u .
Direct calculation then gives
(u, Au) ((I + kA)u, (I + kA)u)
2

= (u, Au) (u, u) + 2k (u, Au) + k 2 (u, Au) u, A2 u



≤ (u, Au) (u, u) + 2k (u, u) u, A2 u + k 2 (u, u) u, A3 u
= (u, u) ((I + kA)u, A(I + kA)u) . (10.25)

Let us define a modified energy Ẽφn as
Z  
1
Ẽφn (u) = |∇u|2 + Ṽn (x)|u|2 dx
Ω 2
Z  
1
= |∇u|2 + V (x)|u|2 + β|φn |2 |u|2 dx ,
Ω 2

we then get from the above lemma that

Lemma 10.9: Suppose V (x) ≥ 0 for all x ∈ Ω, β ≥ 0 and kφn k = 1.

Then,
!
n+1 Ẽφn (φ̃n+1 ) φ̃n+1
Ẽφn (φ̃ )≤ = Ẽφn
kφ̃n+1 k kφ̃n+1 k
= Ẽφn (φn+1 ) ≤ Ẽφn (φn ) . (10.26)

Using the inequality (10.22), we in turn get:

Lemma 10.10: Suppose V (x) ≥ 0 for all x ∈ Ω and β ≥ 0, then,

Ẽ(φ̃n+1 ) ≤ Ẽ(φn ),
where
Z  
1
Ẽ(u) = |∇u|2 + V (x)|u|2 + β|u|4 dx .
Ω 2
Remark 10.11: As we noted earlier, for β = 0, the energy diminishing
property is preserved in the GFDN (10.1)-(10.3) and semi-implicit time
discretization (10.20)-(10.21). For β > 0, the energy diminishing property
in general does not hold uniformly for all φ0 and all step size k > 0,
a justification on the energy diminishing is presently only possible for a
modified energy within two adjacent steps.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 177

10.5. Discretized normalized gradient flow (DNGF)

Consider a discretization for the GFDN (10.20)-(10.21) (or a fully dis-
cretization of (10.15)-(10.16))
Ũ n+1 − U n Ũ n+1
= −AŨ n+1 , U n+1 = , n = 0, 1, 2, · · · ;
k kŨ n+1 k
(10.27)
where U n = (un1 , un2 , · · · , unM−1 )T , k > 0 is time step and A is an
(M − 1) × (M − 1) symmetric positive definite matrix. We adopt the in-
ner product, norm and energy of vectors U = (u1 , u2 , · · · , uM−1 )T and
V = (v1 , v2 , · · · , vM−1 )T as
M−1
X
(U, V ) = U T V = uj vj , kU k2 = U T U = (U, U ), (10.28)
j=1

E(U ) = U T AU = (U, AU ), (10.29)

respectively. Using the finite dimensional version of the lemmas given in
the previous subsection, we have

Theorem 10.12: Suppose kU 0 k = 1 and A is symmetric positive definite.

Then the DNGF (10.27) is energy diminishing, i.e.



E U n+1 ≤ E (U n ) ≤ · · · ≤ E U 0 , n = 0, 1, 2, · · · . (10.30)
Furthermore if I + kA is an M -matrix, then (I + kA)−1 is a nonnegative
matrix (i.e. with nonnegative entries). Thus the flow is monotone, i.e. if
U 0 is a non-negative vector, then U n is also a non-negative vector for all
n ≥ 0.

Remark 10.13: If a discretization for the GFDN (10.20)-(10.21) reads

Ũ n+1 − U n Ũ n+1
= −BU n , U n+1 = , n = 0, 1, 2, · · · . (10.31)
k kŨ n+1 k
For symmetric, positive definite B with ρ(kB) < 1 (ρ(B) being the spectral
radius of B), (10.30) is satisfied by choosing
1 
A= (I − kB)−1 − I = (I − kB)−1 B.
k
Remark 10.14: If a discretization for the GFDN (10.20)-(10.21) reads
Ũ n+1
Ũ n+1 = BU n , U n+1 = , n = 0, 1, 2, · · · . (10.32)
kŨ n+1 k
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

178 W. Bao

For symmetric, positive definite B with ρ(B) < 1, (10.30) is satisfied by

choosing
1

A= B −1 − I .
k
Remark 10.15: If a discretization for the GFDN (10.20)-(10.21) reads
Ũ n+1 − U n Ũ n+1
= −B Ũ n+1 − CU n , U n+1 = , n = 0, 1, 2, · · · .
k kŨ n+1 k
(10.33)
Suppose B and C are symmetric, positive definite and ρ(kC) < 1. Then
(10.30) is satisfied by choosing
−1
A = (I − kC) (B + C).

10.6. Numerical methods

In this section, we will present two numerical methods to discretize the
GFDN (10.1)-(10.3) (or a full discretization of the CNGF (10.15)-(10.16)).
For simplicity of notation we introduce the methods for the case of one
spatial dimension (d = 1) with homogeneous periodic boundary conditions.
Generalizations to higher dimension are straightforward for tensor product
grids and the results remain valid without modifications. For d = 1, we
have
1
φt = φxx − V (x)φ − β |φ|2 φ,
2
x ∈ Ω = (a, b), tn < t < tn+1 , n ≥ 0, (10.34)
4 φ(x, t−
n+1 )
φ(x, tn+1 ) = φ(x, t+
n+1 ) = − , a ≤ x ≤ b, n ≥ 0, (10.35)
kφ(·, tn+1 )k
φ(x, 0) = φ0 (x), a ≤ x ≤ b, φ(a, t) = φ(b, t) = 0, t ≥ 0; (10.36)
with
Z b
2
kφ0 k = φ20 (x) dx = 1.
a

We choose the spatial mesh size h = 4x > 0 with h = (b − a)/M and

M an even positive integer, the time step is given by k = 4t > 0 and define
grid points and time steps by
xj := a + j h, tn := n k, j = 0, 1, · · · , M, n = 0, 1, 2, · · ·
Let be the numerical approximation of φ(xj , tn ) and φn the solution
φnj
vector at time t = tn = nk with components φnj .
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 179

Backward Euler finite difference (BEFD) We use backward Euler

for time discretization and second-order centered finite difference for spatial
derivatives. The detail scheme is:

φ∗j − φnj 1  
2
= 2 φ∗j+1 − 2φ∗j + φ∗j−1 − V (xj )φ∗j − β φnj φ∗j ,
k 2h
j = 1, · · · , M − 1,
φ∗0 = φ∗M = 0, φ0j = φ0 (xj ), j = 0, 1, · · · , M,
φ∗j
φn+1
j = , j = 0, · · · , M, n = 0, 1, · · · ; (10.37)
kφ∗ k

PM−1
2
where the norm is defined as kφ∗ k2 = h j=1 φ∗j .
Time-splitting sine-spectral method (TSSP) From time t = tn to
time t = tn+1 , the equation (10.34) is solved in two steps. First, one solves

1
φt = φxx , (10.38)
2

for one time step of length k, then followed by solving

φt (x, t) = −V (x)φ(x, t) − β|φ|2 φ(x, t), tn ≤ t ≤ tn+1 , (10.39)

again for the same time step. Equation (10.38) is discretized in space by
the sine-spectral method and integrated in time exactly. For t ∈ [tn , tn+1 ],
multiplying the ODE (10.39) by φ(x, t), one obtains with ρ(x, t) = φ2 (x, t)

ρt (x, t) = −2V (x)ρ(x, t) − 2βρ2 (x, t), tn ≤ t ≤ tn+1 . (10.40)

The solution of the ODE (10.40) can be expressed as



 V (x)ρ(x, tn )

 V (x) 6= 0,
 (V (x) + βρ(x, tn )) e2V (x)(t−tn ) − βρ(x, tn )
ρ(x, t) =



 ρ(x, tn )
 , V (x) = 0.
1 + 2βρ(x, tn )(t − tn )
(10.41)
Combining the splitting step via the standard second-order Strang splitting
for solving the GFDN (10.34)-(10.36), in detail, the steps for obtaining φn+1
j
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

180 W. Bao

from φnj are given by

s

 V (xj )e−kV (xj )

 φnj V (xj ) 6= 0,

 −kV (xj ) )|φn |2
 V (xj ) + β(1 − e j
φ∗j =

 1

 q φnj , V (xj ) = 0,


 1+ βk|φnj |2
M−1
X 2
φ∗∗
j = e−kµl /2 φb∗l sin(µl (xj − a)), j = 1, 2, · · · , M − 1,
l=1
s

 V (xj )e−kV (xj )

 φ∗∗ V (xj ) 6= 0,

 −kV (xj ) )|φ∗∗ |2 j
 V (xj ) + β(1 − e j
φ∗∗∗
j =

 1

 φ∗∗
q

 ∗∗ 2
j , V (xj ) = 0,
1 + βk|φj |
φ∗∗∗
j
φn+1
j = , j = 0, · · · , M, n = 0, 1, · · · ; (10.42)
kφ∗∗∗ k

where U bl are the sine-transform coefficients of a real vector U =

(u0 , u1 , · · · , uM )T with u0 = uM = 0 which are defined as

M−1
X
πl bl = 2
µl = , U uj sin(µl (xj − a)), l = 1, 2, · · · , M − 1
b−a M j=1
(10.43)
and

φ0j = φ(xj , 0) = φ0 (xj ), j = 0, 1, 2, · · · , M.

Note that the only time discretization error of TSSP is the splitting error,
which is second order in k.

For comparison purposes we review a few other numerical methods

which are currently used for solving the GFDN (10.34)-(10.36). One is the
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 181

Crank-Nicolson finite difference (CNFD) scheme:

φ∗j − φnj 1  
= 2 φ∗j+1 − 2φ∗j + φ∗j−1 + φnj+1 − 2φnj + φnj−1
k 4h
 2
V (xj )  ∗ n
 β φnj   ∗ 
− φj + φj − φj + φnj , j = 1, · · · , M − 1,
2 2
φ∗0 = φ∗M = 0, φ0j = φ0 (xj ), j = 0, 1, · · · , M,
∗
φj
φn+1
j = ∗ , j = 0, · · · , M, n = 0, 1, · · · . (10.44)
kφ k
Another one is the forward Euler finite difference (FEFD) method:
φ∗j − φnj 1    2
= 2 φnj+1 − 2φnj + φnj−1 − V (xj )φnj − β φnj  φnj ,
k 2h
j = 1, · · · , M − 1,
φ∗0 = φ∗M = 0, φ0j = φ0 (xj ), j = 0, 1, · · · , M,
φ∗j
φn+1
j = , j = 0, · · · , M, n = 0, 1, · · · ; (10.45)
kφ∗ k

10.7. Energy diminishing of DNGF

First we analyze the energy diminishing of the different numerical methods
for linear case, i.e. β = 0 in (10.34). Introducing

T
Φn = φn1 , φn2 , · · · , φnM−1 ,

2, j = l,
1 
D = (djl )(M−1)×(M−1) , with djl = 2 −1, |j − l| = −1,
2h 
0, otherwise,
E = diag (V (x1 ), V (x2 ), · · · , V (xM−1 )) ,

T
F (Φ) = diag φ21 , φ22 , · · · , φ2M−1 , with Φ = (φ1 , φ2 , · · · , φM−1 ) ,
M−1
2 X πmj πml −kµ2m /2
G = (gjl )(M−1)×(M−1) , with gjl = sin sin e ,
M m=1 M M
 
H = diag e−kV (x1 )/2 , e−kV (x2 )/2 , · · · , e−kV (xM −1 )/2 .

Then the BEFD discretization (10.37) (called as BEFD normalized flow)

with β = 0 can be expressed as
Φ∗ − Φn Φ∗
= −(D + E)Φ∗ , Φn+1 = , n = 0, 1, · · · . (10.46)
k kΦ∗ k
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

182 W. Bao

The TSSP discretization (10.42) (called as TSSP normalized flow) with

β = 0 can be expressed as
Φ∗
Φ∗∗∗ = HΦ∗∗ = HGΦ∗ = HGHΦn , Φn+1 = , n = 0, 1, · · · .
kΦ∗ k
(10.47)
The CNFD discretization (10.44) (called as CNFD normalized flow) with
β = 0 can be expressed as
Φ∗ − Φn 1 1 Φ∗
= − (D+E)Φ∗ − (D+E)Φn , Φn+1 = , n = 0, 1, · · · .
k 2 2 kΦ∗ k
(10.48)
The FEFD discretization (10.45) (called as FEFD normalized flow) with
β = 0 can be expressed as
Φ∗ − Φn Φ∗
= −(D + E)Φn , Φn+1 = , n = 0, 1, · · · . (10.49)
k kΦ∗ k
It is easy to see that D and G are symmetric positive definite

2matrices.2
π
Furthermore D is also an M -matrix and ρ(D) = 1 + cos M /h < 2/h
−kµ21 /2
and ρ(G) = e < 1. Applying the theorem 10.12 and remarks 10.13,
10.14 and 10.15, we have

Theorem 10.16: Suppose V ≥ 0 in Ω and β = 0. We have that

(i). The BEFD normalized flow (10.37) is energy diminishing and
monotone for any k > 0.
(ii). The TSSP normalized flow (10.42) is energy diminishing for any
k > 0.
(iii). The CNFD normalized flow (10.44) is energy diminishing and
monotone provided that
2 2h2
k≤ = . (10.50)
2/h2 + maxj V (xj ) 2 + h2 maxj V (xj )
(iv). The FEFD normalized flow (10.45) is energy diminishing and
monotone provided that
1 h2
k≤ = . (10.51)
2/h2 + maxj V (xj ) 2 + h2 maxj V (xj )
For nonlinear case, i.e. β > 0, we only analyze the energy between two
steps of BEFD flow (10.37). In this case, consider
Φ̃n+1 − Φn Φ̃n+1
= − (D + E + βF (Φn )) Φ̃n+1 , Φn+1 = . (10.52)
k kΦ̃n+1 k
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 183

Lemma 10.17: Suppose V ≥ 0, β > 0 and kΦn k = 1. Then for the flow
(10.52), we have
 

Ẽ Φ̃n+1 ≤ Ẽ (Φn ) , ẼΦn Φn+1 ≤ ẼΦn (Φn ) (10.53)

where
M−1
X
Ẽ (Φ) = (Φ, (D + E + βF (Φ))Φ) = ΦT (D + E)Φ + β φ4j , (10.54)
j=1
n
Ẽ
Φn (Φ) = (Φ, (D + E + βF (Φ ))Φ)
M−1
X
2
= ΦT (D + E)Φ + β φ2j φnj . (10.55)
j=1

Proof: Combining (10.52), (10.27) and Theorem 10.12, we have

 
  Φ̃n+1 , (D + E + βF (Φn ))Φ̃n+1
Φ̃n+1 , (D + E + βF (Φn ))Φ̃n+1 ≤  
Φ̃n+1 , Φ̃n+1
(Φn , (D + E + βF (Φn ))Φn )
≤ = Ẽ (Φn ) . (10.56)
(Φn , Φn )
Similar to the proof of (10.22), we have
M−1
X
2  2 X
M−1 4
φnj φ̃n+1
j ≥ φ̃n+1
j . (10.57)
j=1 j=1

The required result (10.53) is a combination of (10.57), and (10.56). 

10.8. Numerical results

Here we report the ground state solutions in BEC with different potentials
by the method BEFD. Due to the ground state solution φg (x) ≥ 0 for
x ∈ Ω in non-rotating BEC [87], in our computations, the initial condition
(10.3) is always chosen such that φ0 (x) ≥ 0 and decays to zero sufficiently
fast as |x| → ∞. We choose an appropriately large interval, rectangle and
box in 1d, 2d and 3d, respectively, to avoid that the homogeneous periodic
boundary condition (10.36) introduce a significant (aliasing) error relative
to the whole space problem. To quantify the ground state solution φg (x),
we define the radius mean square
sZ
αrms = kαφg kL2 (Ω) = α2 φ2g (x) dx, α = x, y, or z. (10.58)
Ω
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

184 W. Bao

Example 1 Ground state solution of 1d BEC with harmonic oscillator

potential
x2 1 2
V (x) = , φ0 (x) = e−x /2 , x ∈ R.
2 (π)1/4
The CNGF (10.15)-(10.16) with d = 1 is solved on Ω = [−16, 16] with
mesh size h = 1/8 and time step  k = 0.1 by using BEFD. The steady state
solution is reached when max Φ  n+1
− Φn  < ε = 10−6 . Fig. 1 shows the
ground state solution φg (x) and energy evolution for different β. Tab. 1
displays the values of φg (0), radius mean square xrms , energy E(φg ) and
chemical potential µg .

β φg (0) xrms E(φg ) µg = µβ (φg )

0 0.7511 0.7071 0.5000 0.5000
3.1371 0.6463 0.8949 1.0441 1.5272
12.5484 0.5301 1.2435 2.2330 3.5986
31.371 0.4562 1.6378 3.9810 6.5587
62.742 0.4067 2.0423 6.2570 10.384
156.855 0.3487 2.7630 11.464 19.083
313.71 0.3107 3.4764 18.171 30.279
627.42 0.2768 4.3757 28.825 48.063
1254.8 0.2467 5.5073 45.743 76.312
Tab. 1: Maximum value of the wave function φg (0), root mean square size
xrms , energy E(φg ) and ground state chemical potential µg verus the inter-
action coefficient β in 1d.

The results in Fig. 1. and Tab. 1. agree very well with the ground state
solutions of BEC obtained by directly minimizing the energy functional [5].

Example 2 Ground state solution of BEC in 2d. Two cases are consid-
ered:
I. With a harmonic oscillator potential [5, 8, 44], i.e.
1 2 2

V (x, y) = γx x + γy2 y 2 .
2
II. With a harmonic oscillator potential and a potential of a stirrer
corresponding a far-blue detuned Gaussian laser beam [27] which is used
to generate vortices in BEC [27], i.e.
1 2 2
2 2
V (x, y) = γx x + γy2 y 2 + w0 e−δ((x−r0 ) +y ) .
2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 185

0.8
3 β=12.5484, E(U)
0.7 β=156.855, E(U)/10
β=627.4, E(U)/50
0.6 β=1254.8, E(U)/100
2.5

0.5
2

Energy
Φg(x)

0.4

0.3 1.5

0.2
1
0.1

0 0.5
0 2 4 6 8 10 12 14 0 0.1 0.2 0.3 0.4 0.5 0.6
a) x b) t

Fig. 1: Ground state solution φg in Example 1. (a). For β =

0, 3.1371, 12.5484, 31.371, 62.742, 156.855, 313.71, 627.42, 1254.8 (with de-
creasing peak). (b). Energy evolution for different β.

The initial condition is chosen as

(γx γy )1/4 −(γx x2 +γy y2 )/2
φ0 (x, y) = e .
π 1/2
For the case I, we choose γx = 1, γy = 4, w0 = δ = r0 = 0, β = 200
and solve the problem by BEFD on Ω = [−8, 8] × [−4, 4] with mesh size
hx = 1/8, hy = 1/16 and time step k = 0.1. We get the following results
from the ground state solution φg :
xrms = 2.2734, yrms = 0.6074, φ2g (0) = 0.0808,
E(φg ) = 11.1563, µg = 16.3377.
For case II, we choose γx = 1, γy = 1, w0 = 4, δ = r0 = 1, β = 200 and
solve the problem by TSSP on Ω = [−8, 8]2 with mesh size h = 1/8 and
time step k = 0.001. We get the following results from the ground state
solution φg :
xrms = 1.6951, yrms = 1.7144, φ2g (0) = 0.034,
E(φg ) = 5.8507, µg = 8.3269.
In addition, Fig. 2 shows surface plots of the ground state solution φg .

Example 3 Ground state solution of BEC in 3d. Two cases are consid-
ered:
I. With a harmonic oscillator potential [5, 8, 44], i.e.
1 2 2

V (x, y, z) = γx x + γy2 y 2 + γz2 z 2 .
2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

186 W. Bao

2
φg
0.05 0.04
2
φg

6
0 0.03
4
2 0.02
2

1 0.01 0
5
0 −2
0
0 −6 −4 −4
−1 −2 y
0 2 4 −6
−5 6
y
(a). −2 x (b). x

2
Fig. 2: Ground state solutions φg in Example 2, case I (a), and case II (b).

II. With a harmonic oscillator potential and a potential of a stirrer

corresponding a far-blue detuned Gaussian laser beam [27] which is used
to generate vortex in BEC [27], i.e.
1 2 2
2 2
V (x, y, z) = γx x + γy2 y 2 + γz2 z 2 + w0 e−δ((x−r0 ) +y ) .
2
The initial condition is chosen as
(γx γy γz )1/4 −(γx x2 +γy y2 +γz z2 )/2
φ0 (x, y, z) = e .
π 3/4
For case I, we choose γx = 1, γy = 2, γz = 4, w0 = δ = r0 = 0, β = 200
and solve the problem by TSSP on Ω = [−8, 8] × [−6, 6] × [−4, 4] with mesh
3
size hx = 81 , hy = 32 1
, hz = 16 and time step k = 0.001. The ground state
solution φg gives:

xrms = 1.67, yrms = 0.87, zrms = 0.49,

φ2g (0) = 0.052, E(φg ) = 8.33, µg = 11.03.

For case II, we choose γx = 1, γy = 1, γz = 2, w0 = 4, δ = r0 = 1,

β = 200 and solve the problem by TSSP on Ω = [−8, 8]3 with mesh size
h = 81 and time step k = 0.001. The ground state solution φg gives:

xrms = 1.37, yrms = 1.43, zrms = 0.70,

φ2g (0) = 0.025, E(φg ) = 5.27, µg = 6.71.

Furthermore, Fig. 3 shows surface plots of the ground state solution

φ2g (x, 0, z). BEFD gives similar results with k = 0.1.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 187

2
φg
0.03
φ2
g

0.02

0.04 0.01

0.02 0
5 4
0
2 2 5
1 0
0
0 0
−2
−1 x
−5 z
z x
a). −2 b). −4 −5

Fig. 3: Ground state solutions φ2g (x, 0, z) in Example 3. (a). For case I. (b).
For case II.

Example 4 2d central vortex states in BEC, i.e.

 
1 m2 1 2
V (x, y) = V (r) = 2
+ r 2
, φ0 (x, y) = φ0 (r) = √ rm e−r /2 , 0 ≤ r.
2 r πm!
The CNGF (10.15)-(10.16) is solved in polar coordinate with Ω = [0, 8] with
1
mesh size h = 64 and time step k = 0.1 by using BEFD. Fig. 4a shows the
ground state solution φg (r) with β = 200 for different index of the central
vortex m. Tab. 2 displays the values of φg (0), radius mean square rrms ,
energy E(φg ) and chemical potential µg .

Index m φg (0) rrms E(φg ) µg = µβ (φg )

1 0.0000 2.4086 5.8014 8.2967
2 0.0000 2.5258 6.3797 8.7413
3 0.0000 2.6605 7.0782 9.3160
4 0.0000 2.8015 7.8485 9.9772
5 0.0000 2.9438 8.6660 10.6994
6 0.0000 3.0848 9.5164 11.4664
Tab. 2: Numerical results for 2d central vortex states in BEC.

Example 5. The first excited state solution of BEC in 1d with a har-

monic oscillator potential, i.e.
√
x2 2 2
V (x) = , φ0 (x) = 1/4
x e−x /2 , x ∈ R.
2 (π)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

188 W. Bao

0.2 0.7

0.6
0.15
0.5
φ (r)
g

0.4

φ (x)
0.1

1
0.3

0.2
0.05
0.1

0 0

0 2 4 6 0 5 10
(a). r (b). x

Fig. 4: (a). 2d central vortex states φg (r) in Example 4. β =

200 for m = 1 to 6 (with decreasing peak). (b). First excited
state solution φ1 (x) (an odd function) in Example 5. For β =
0, 3.1371, 12.5484, 31.371, 62.742, 156.855, 313.71, 627.42, 1254.8 (with
decreasing peak).

The CNGF (10.15)-(10.16) with d = 1 is solved on Ω = [−16, 16] with mesh

size h = 1/64 and time step k = 0.1 by using BEFD. Fig. 4b shows the first
excited state solution φ1 (x) for different β. Tab. 3 displays the radius mean
square xrms = kxφ1 kL2 (Ω) , ground state and first excited state energies
E(φg ) and E(φ1 ), ratio E(φ1 )/E(φg ), chemical potentials µg = µβ (φg ) and
µ1 = µβ (φ1 ), ratio µ1 /µg .

E(φ1 ) µ1
β xrms E(φg ) E(φ1 ) E(φg ) µg µ1 µg
0 1.2247 0.500 1.500 3.000 0.500 1.500 3.000
3.1371 1.3165 1.044 1.941 1.859 1.527 2.357 1.544
12.5484 1.5441 2.233 3.037 1.360 3.598 4.344 1.207
31.371 1.8642 3.981 4.743 1.192 6.558 7.279 1.110
62.742 2.2259 6.257 6.999 1.119 10.38 11.089 1.068
156.855 2.8973 11.46 12.191 1.063 19.08 19.784 1.037
313.71 3.5847 18.17 18.889 1.040 30.28 30.969 1.023
627.42 4.4657 28.82 29.539 1.025 48.06 48.733 1.014
1254.8 5.5870 45.74 46.453 1.016 76.31 76.933 1.008
Tab. 3: Numerical results for the first excited state solution in 1d in Example
5.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 189

From the results in Tab. 3 and Fig. 4b, we can see that the BEFD can be
applied directly to compute the first excited states in BEC. Furthermore,
we have
E(φ1 ) µ1
lim = 1, lim = 1.
β→+∞ E(φg ) β→+∞ µg

These results are confirmed with the results in [5] where the ground and first
excited states are computed by directly minimizing the energy functional
through the finite element discretization.

11. Numerical methods for dynamics of NLSE

In this section we present time-splitting sine pseudospectral (TSSP) meth-
ods for the problem (4.1), (4.2) with/without external driven field with
homogeneous Dirichlet boundary conditions. For the simplicity of notation
we shall introduce the method for the case of one space dimension (d = 1).
Generalizations to d > 1 are straightforward for tensor product grids and
the results remain valid without modifications. For d = 1, the problem with
an external driven field becomes
1
i∂t ψ = − ∂xx ψ + V (x)ψ + W (x, t)ψ + β|ψ|2 ψ, a < x < b, t > 0, (11.1)
2
ψ(x, t = 0) = ψ0 (x), a ≤ x ≤ b, ψ(a, t) = ψ(b, t) = 0, t ≥ 0; (11.2)
where W (x, t) is an external driven field. Typical external driven fields used
in physical literatures include a far-blued detuned Gaussian laser beam
stirrer [27]
  
|x − xs (t)|2
W (x, t) = Ws (t) exp − , (11.3)
ws /2
with Ws the height, ws the width, and xs (t) the position of the stirrer; or
a Delta-kicked potential [77]
∞
X
W (x, t) = K cos(kx) δ(t − nτ ), (11.4)
n=−∞

with K the kick strength, k the wavenumber, τ the time interval between
kicks, and δ(τ ) is the Dirac delta function.

11.1. General high-order split-step method

As preparatory steps, we begin by introducing the general high-order split-
step method [53] for a general evolution equation
i ∂t u = f (u) = A u + B u, (11.5)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

190 W. Bao

where f (u) is a nonlinear operator and the splitting f (u) = Au + Bu can

be quite arbitrary, in particular, A and B do not need to commute. For
a given time step k = ∆t > 0, let tn = n k, n = 0, 1, 2, . . . and un be
the approximation of u(tn ). A second-order symplectic time integrator (cf.
[111]) for (11.5) is as follows:
u(1) = e−ik A/2 un ;
u(2) = e−ik B u(1) ;
un+1 = e−ik A/2 u(2) . (11.6)
A fourth-order symplectic time integrator (cf. [120]) for (11.5) is as follows:
u(1) = e−i2w1 k A un ;
u(2) = e−i2w2 k B u(1) ;
u(3) = e−i2w3 k A u(2) ;
u(4) = e−i2w4 k B u(3) ;
u(5) = e−i2w3 k A u(4) ;
u(6) = e−i2w2 k B u(5) ;
un+1 = e−i2w1 k A u(6) ; (11.7)
where
w1 = 0.33780 17979 89914 40851, w2 = 0.67560 35959 79828 81702,
w3 = −0.08780 17979 89914 40851, w4 = −0.85120 71979 59657 63405.

11.2. Fourth-order TSSP for GPE without external driving

field
We choose the spatial mesh size h = 4x > 0 with h = (b − a)/M for M
an even positive integer, and let xj := a + j h, j = 0, 1, · · · , M . Let ψjn
be the approximation of ψ(xj , tn ) and ψ n be the solution vector at time
t = tn = nk with components ψjn .
We now rewrite the GPE (11.1) without external driven field, i.e.
W (x, t) ≡ 0, in the form of (11.5) with
1
Aψ = V (x)ψ(x, t) + β|ψ(x, t)|2 ψ(x, t), Bψ = − ∂xx ψ(x, t). (11.8)
2
Thus, the key for an efficient implementation of (11.7) is to solve efficiently
the following two subproblems:
1
i ∂t ψ(x, t) = Bψ = − ∂xx ψ, (11.9)
2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 191

and

i ∂t ψ(x, t) = V (x)ψ(x, t) + β|ψ(x, t)|2 ψ(x, t). (11.10)

Equation (11.9) will be discretized in space by the sine pseudospectral

method and integrated in time exactly. For t ∈ [tn , tn+1 ], the ODE (11.10)
leaves |ψ| invariant in t [11, 10] and therefore becomes

iψt (x, t) = V (x)ψ(x, t) + β|ψ(x, tn )|2 ψ(x, t) (11.11)

and thus can be integrated exactly.

From time t = tn to t = tn+1 , we combine the splitting steps via the
fourth-order split-step method and obtain a fourth-order time-splitting sine-
spectral (TSSP4) method for the GPE (10.34). The detailed method is
given by
(1) n 2
ψj = e−i2w1 k(V (xj )+β|ψj | )
ψjn ,
M−1
X
(2) 2 (1)
ψj = e−iw2 kµl ψbl sin(µl (xj − a)),
l=1
(2) 2
(3) (2)
ψj = e−i2w3 k(V (xj )+β|ψj | )
ψj ,
M−1
X
(4) 2 (3)
ψj = e−iw4 kµl ψbl sin(µl (xj − a)), j = 1, 2, · · · , M − 1,
l=1
(4) 2
(5) (4)
ψj = e−i2w3 k(V (xj )+β|ψj | )
ψj ,
M−1
X
(6) 2 (5)
ψj = e−iw2 kµl ψbl sin(µl (xj − a)),
l=1
(6) 2
(6)
ψjn+1 = e−i2w1 k(V (xj )+β|ψj | )
ψj , (11.12)

where U bl , the sine-transform coefficients of a complex vector U =

(U0 , U1 , · · · , UM ) with U0 = UM = 0, are defined as
M−1
X
πl bl = 2
µl = , U Uj sin(µl (xj −a)), l = 1, 2, · · · , M −1, (11.13)
b−a M j=1

with

ψj0 = ψ(xj , 0) = ψ0 (xj ), j = 0, 1, 2, · · · , M. (11.14)

Note that the only time discretization error of TSSP4 is the splitting error,
which is fourth order in k for any fixed mesh size h > 0.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

192 W. Bao

This scheme is explicit, time reversible, just as the IVP for the GPE.
Also, a main advantage of the time-splitting method is its time-transverse
invariance, just as it holds for the GPE itself. If a constant α is added to the
potential V1 , then the discrete wave functions ψjn+1 obtained from TSSP4
get multiplied by the phase factor e−iα(n+1)k , which leaves the discrete
quadratic observables unchanged. This property does not hold for finite
difference schemes.

11.3. Second-order TSSP for GPE with external driving

field
We now rewrite the GPE (11.1) with an external driven field

1
Aψ = − ∂xx ψ(x, t),
2
Bψ = V (x)ψ(x, t) + W (x, t)ψ(x, t) + β|ψ(x, t)|2 ψ(x, t). (11.15)

Due to the external driven field could be vary complicated, e.g. it may be
a Delta-function [77], here we only use a second-order split-step scheme
in time discretization. More precisely, from time t = tn to t = tn+1 , we
proceed as follows:
M−1
X
ψj∗ =
2
[
e−ikµl /4 (ψ n ) sin(µ (x − a)),
l j j = 1, 2, · · · , M − 1,
l
l=1
 Z tn+1 
ψj∗∗ = exp −ik(V (xj ) + β|ψjn |2 ) − i W (xj , t)dt ψj∗ ,
tn
M−1
X
ψjn+1 =
2
\
e−ikµl /4 (ψ ∗∗ ) sin(µ (x − a)).
l j (11.16)
l
l=1

Remark 11.1: If the integral in (11.16) could not be evaluated analyti-

cally, one can use numerical quadrature to evaluate, e.g.
Z tn+1
k
W (xj , t)dt ≈ [W (xj , tn ) + 4W (xj , tn + k/2) + W (xj , tn+1 )] .
tn 6

11.4. Stability
Let U = (U0 , U1 , · · · , UM )T with U0 = UM = 0, f (x) a homogeneous
periodic function on the interval [a, b], and let k · kl2 be the usual discrete
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 193

l2 -norm on the interval (a, b), i.e.,

v v
u u
u b − a M−1
X u b − a M−1
X
kU kl2 = t 2
|Uj | , kf kl2 = t |f (xj )|2 . (11.17)
M j=1 M j=1

For the stability of the second-order time-splitting spectral approxima-

tions TSSP2 (11.16) and fourth-order scheme (11.12), we have the following
lemma, which shows that the total charge is conserved.

Lemma 11.2: The second-order time-splitting sine pseudospectral scheme

(11.16) and fourth-order scheme (11.12) are unconditionally stable. In fact,
for every mesh size h > 0 and time step k > 0,
kψ n kl2 = kψ 0 kl2 = kψ0 kl2 , n = 1, 2, · · · (11.18)

Proof: For the scheme TSSP2 (11.16), noting (10.43) and (11.17), one has
M−1
1 1 X  n+1 2
kψ n+1 k2l2 = ψ
b−a M j=1 j
 2
1 X  X −ikµ2l /4 \ 
M−1 M−1
∗∗ 
=  e (ψ )l sin(µl (xj − a))
M j=1  
l=1
 
1 X  X −ikµ2l /4 \
M−1 M−1 2
∗∗ 
=  e (ψ )l sin(jlπ/M )
M  
j=1 l=1

1 X
M−1
 −ikµ2l /4 \
2
 1 X \
M−1 2

= e ∗∗
(ψ )l  = (ψ ∗∗ )l  . (11.19)
2 2
l=1 l=1

Plugging (10.43) into (11.19), we obtain

 2
1 1 X  2 M−1
M−1 X 

kψ n+1 k2l2 =  ψ ∗∗
sin(µ (x − a))
b−a 2  M j=1 j l j 
l=1  
 2
M−1  M−1  M−1
1 X  2 X ∗∗ 
 1 X  ∗∗ 2
= M ψj sin(ljπ/M ) = ψ
2  j=1  M j=1 j
l=1
M−1   Z tn+1  2
1 X  n 2 ∗

= exp −ik(V (xj ) + β|ψj | ) − i W (xj , t)dt ψj 
M j=1 tn

M−1 M−1
1 X  ∗ 2 1 X  n 2 1
= ψj = ψj = kψ n k2l2 . (11.20)
M j=1 M j=1 b−a
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

194 W. Bao

Here we used the identities

M−1
X 
0, k = l,
sin(ljπ/M ) sin(kjπ/M ) = (11.21)
M/2, k 6= l.
j=1

For the scheme TSSP4 (11.12), using (10.43), (11.17) and (11.21), we
get similarly
M−1
1 1 X  n+1 2
kψ n+1 k2l2 = ψ
b−a M j=0 j

1 X  −i2w1 k(V (xj )+β|ψj(6) |2 ) (6) 2

M−1
= e ψj 
M j=0

1 X  (6) 2
M−1
= ψ 
M j=0 j
 2
1 X  X −iw2 kµ2l b(5) 
M−1 M−1

=  e ψl sin(µl (xj − a))
M j=0  
l=1
 2
1 X  X −iw2 kµ2l b(5) 
M−1 M−1

=  e ψl sin(ljπ/M )
M  
j=0 l=1

1 X
M−1 
 −iw2 kµ2l b(5) 2 1 X  b(5) 2
M−1
= e ψl  = ψl  . (11.22)
2 2
l=1 l=1

Plugging (10.43)) into (11.22), we have

 2
1 1 X  2 M−1
M−1 X (5) 

kψ n+1 2
k l2 =  ψj sin(µl (xj − a))
b−a 2  M j=1
l=1  
 2
M−1  M−1  X  (5) 2
M−1
1 X  2 X (5) 
 = 1
=  ψj sin(jlπ/M ) ψ 
2  M j=1  M j=1 j
l=1

1 X  (4) 2 1 X  (3) 2 1 X  (2) 2

M−1 M−1 M−1
= ψj  = ψj  = ψ 
M j=1 M j=1 M j=1 j

1 X  (1) 2
M−1 M−1
1 X  n 2
= ψj  = ψ
M j=1 M j=1 j
1
= kψ n k2l2 . (11.23)
b−a
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 195

Thus the equality (11.18) can be obtained from (11.19) for the scheme
TSSP2 and (11.22) for the scheme TSSP4 by induction. 

11.5. Crank-Nicolson finite difference method (CNFD)

Another scheme used to disretize the NLSE (11.1) is the Crank-Nicolson fi-
nite difference method (CNFD). In this method one uses the Crank-Nicolson
scheme for time derivative and the second order central difference scheme
for spatial derivative. The detailed method is:
ψjn+1 − ψjn 1 n+1


i = − 2 ψj+1 − 2ψjn+1 + ψj−1
n+1 n
+ ψj+1 − 2ψjn + ψj−1
n
k 4h
V (xj ) n+1
β


+ ψj + ψjn + |ψjn+1 |2 + |ψjn |2 ψjn+1 + ψjn ,
2 2
j = 1, 2, · · · , M − 1, n = 0, 1, · · · , (11.24)
ψ0n+1 = ψM
n+1
= 0, n = 0, 1, · · · ,
ψj0 = ψ0 (xj ), j = 0, 1, 2, · · · , M.

11.6. Numerical results

In this subsection we present numerical results to confirm spectral accuracy
in space and fourth order accuracy in time of the numerical method (11.12),
and then apply it to study time-evolution of condensate width in 1D, 2D
and 3D.

Example 6 1d Gross-Pitaevskii equation, i.e. in (4.1) we choose d = 1

and γx = 1. The initial condition is taken as
1 2
ψ0 (x) = 1/4 e−x /2 , x ∈ R.
π
We solve on the interval [−32, 32], i.e. a = −32 and b = 32 with ho-
mogeneous Dirichlet boundary condition (11.2). We compute a numerical
1
solution by using TSSP4 with a very fine mesh, e.g. h = 128 , and a very
small time step, e.g. k = 0.0001, as the ‘exact’ solution ψ. Let ψ h,k denote
the numerical solution under mesh size h and time step k.
First we test the spectral accuracy of TSSP4 in space. In order to do
so, for each fixed β1 , we solve the problem with different mesh size h but a
very small time step, e.g. k = 0.0001, such that the truncation error from
time discretization is negligible comparing to that from space discretization.
Tab. 4 shows the errors kψ(t) − ψ h,k (t)kl2 at t = 2.0 with k = 0.0001 for
different β1 and h.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

196 W. Bao

mesh h=1 h = 12 h = 14 h = 18 h= 1
16
β1 = 10
√ 0.2745 1.081E-2 1.805E-6 3.461E-11
β1 = 20 2 1.495 0.1657 7.379E-4 7.588E-10
β1 = 80 1.603 1.637 6.836E-2 3.184E-5 3.47E-11

Tab. 4: Spatial error analysis: Error kψ(t) − ψ h,k (t)kl2 at t = 2.0 with
k = 0.0001 in Example 6.

Then we test the fourth-order accuracy of TSSP4 in time. In order to

do so, for each fixed β1 , we solve the problem with different time step k but
1
a very fine mesh, e.g. h = 64 , such that the truncation error from space
discretization is negligible comparing to that from time discretization. Tab.
1
5 shows the errors kψ(t) − ψ h,k (t)kl2 at t = 2.0 with h = 64 for different β1
and k.

1 1 1 1 1
time step k = 20 k = 40 k = 80 k = 160 k = 320
δ = 10.0
√ 1.261E-4 8.834E-6 5.712E-7 3.602E-8 2.254E-9
δ = 20 2 1.426E-3 9.715E-5 6.367E-6 4.034E-7 2.529E-8
δ = 80 4.375E-2 1.693E-3 8.982E-5 5.852E-6 3.706E-7

1
Tab. 5: Temporal error analysis: kψ(t) − ψ h,k (t)kl2 at t = 2.0 with h = 64
in Example 6.

As shown in Tabs. 4&5, spectral order accuracy for spatial derivatives

and fourth-order accuracy for time derivative of TSSP4 are demonstrated
numerically for 1d GPE, respectively. Another issue is how to choose mesh
size h and time step k in the strong repulsive interaction regime or semi-
classical regime, i.e. βd  1, in order to get “correct” physical observables.
In fact, after a rescaling in (4.1) under the normalization (4.3): x → ε−1/2 x
−2/(d+2)
and ψ → εd/4 ψ with ε = βd , then the GPE (4.1) can be rewritten
as
ε2 2
iε ∂t ψ(x, t) = −∇ ψ + Vd (x)ψ + |ψ|2 ψ, x ∈ Rd . (11.25)
2
As demonstrated in [10, 11], the meshing strategy to capture ‘correct’ phys-
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 197

1.6

1.4

1
1.2
0.5

|ψ|2
σx (or |ψ(0,t)| )
2

1 0
0
0.8
1

2
0.6
3 1
0.4 0.8
4
0.6
5 0.4
0.2
0 2 4 6 8 10 x 0.2
a) t b) 6 0 t

Fig. 5: Numerical results for Example 7: a) Condensate width σx (

‘—’ ) and central density |ψ(0, t)|2 ( ‘- - - ’ ). b) Evolution of the density
function |ψ|2 .

ical observables for this this problem is

h = O(ε), k = O(ε).
Thus the admissible meshing strategy for the GPE with strong repulsive
interaction is
   
2/(d+2) 2/(d+2)
h = O(ε) = O 1/βd , k = O(ε) = O 1/βd , d = 1, 2, 3.
(11.26)

Example 7 1d Gross-Pitaevskii equation, i.e. in (4.1) we choose d = 1.

The initial condition is taken as the ground-state solution of (4.1) under
d = 1 with γx = 1 and β1 = 20.0 [6, 5], i.e. initially the condensate is
assumed to be in its ground state. When t = 0, we double the trap frequency
by setting γx = 2.
3
We solve this problem on the interval [−12, 12] under mesh size h = 64
and time step k = 0.005 with homogeneous Dirichlet boundary condition.
Fig. 5 plots the condensate width and central density |ψ(0, t)|2 as functions
of time, as well as evolution of the density |ψ|2 in space-time. One can
see from this figure that the sudden change in the trap potential leads to
oscillations in the condensate width and the peak value of the wave function.
Note that the condensate width contracts in an oscillatory way (cf. Fig. 5a),
which agrees with the analytical results in (4.43).

Example 8 2d Gross-Pitaevskii equation, i.e. in (4.1) we choose d = 2.

The initial condition is taken as the ground-state solution of (4.1) under
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

198 W. Bao

t=5.4

1.3 σx

1.2

1.1

1 0.3
0.9 0.2

|ψ|2
0.8 0.1

0.7 0
2
0.6
4
0.5 1
2
0.4 0
σy 0
0.3
−1
−2
0 5 10 15 y x
a) t b) −2 −4

Fig. 6: Numerical results for Example 8: a) Condensate width. b) Sur-

face plot of the density |ψ|2 at t = 5.4.

d = 2 with γx = 1, γy = 2 and β2 = 20.0 [6, 5], i.e. initially the condensate

is assumed to be in its ground state. When at t = 0, we double the trap
frequency by setting γx = 2 and γy = 4.
1
We solve this problem on [−8, 8]2 under mesh size h = 32 and time step
k = 0.005 with homogeneous Dirichlet boundary condition. Fig. 6 shows
the condensate widths σx and σy as functions of time and the surface of
the density |ψ|2 at time t = 5.4. Fig. 7 the contour plots of the density |ψ|2
at different times. Again, the sudden change in the trap potential leads
to oscillations in the condensate width. Due to γy = 2γx , the oscillation
frequency of σy is roughly double that of σx and the amplitudes of σx are
larger than those of σy in general (cf. Fig. 6a). Again this agrees with the
analytical results in (4.43).

Example 9 3d Gross-Pitaevskii equation, i.e. in (4.1) we choose d = 3.

We present computations for two cases:

Case I. Intermediate ratio between trap frequencies along different axis

(data for 87 Rb used in JILA [3]). The initial condition is taken as the
ground-state solution of (4.1) under d = 3 with γx = γy = 1, γz = 4 and
β3 = 37.62 [6, 5]. When at t = 0, we four times the trap frequency by setting
γx = γy = 4 and γz = 16.
Case II. High ratio between trap frequencies along different axis (data for
23
Na used in MIT (group of Ketterle) [38]). The initial condition is taken
as the ground-state solution of (4.1) under d = 3 with γx = γy = 360 3.5 ,
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 199

t=0 t=0.9
4 2

3 1.5

2 1

1 0.5

0 0
y

y
−1 −0.5

−2 −1

−3 −1.5

−4 −2
−4 −2 0 2 4 −2 −1 0 1 2
a). x b) x

t=1.8 t=2.7
4 2

3 1.5

2 1

1 0.5

0 0
y

−1 −0.5

−2 −1

−3 −1.5

−4 −2
−4 −2 0 2 4 −2 −1 0 1 2
c). x d) x

t=3.6 t=4.5
4 1.5

3
1
2
0.5
1

0 0
y

−1
−0.5
−2
−1
−3

−4 −1.5
−4 −2 0 2 4 −1.5 −1 −0.5 0 0.5 1 1.5
e). x f) x

2
Fig. 7: Contour plots of the density |ψ| at different times in Example 8.
a). t = 0, b). t = 0.9, c). t = 1.8, d). t = 2.7, e). t = 3.6, f). t = 4.5.

γz = 1 and β3 = 3.083 [6, 5]. When at t = 0, we double the trap frequency

by setting γx = γy = 720
3.5 and γz = 2.

For case I, we solve the problem on [−6, 6] × [−6, 6] × [−3, 3] under mesh
3 3
size hx = hy = 32 and hz = 64 , and time step k = 0.0025 with homoge-
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

200 W. Bao

1.2

σx=σy
1

0.1
0.8
0.05

|ψ|2
0.6
0
1.5
0.4
σ 1 4
z
0.5
2
0.2
0
0
−0.5
0 −1 −2
0 0.5 1 1.5 2 z x
a) t b) −1.5 −4

1.8
σ =σ
x y
1.6

1.4
20
1.2
2

10
|ψ|

0.8 0
4
0.6
2 0.2
0.4
0.1
0
0.2 σ 0
z
−2 −0.1
0
0 1 2 3 4 z −0.2
x
c) t d) −4

Fig. 8: Numerical results for Example 9: Left column: Condensate width;

right column: Surface plot of the density in xz-plane, |ψ(x, 0, z, t)|2 . Case
I: a) and b) at t = 1.64. Case II: c) and d) at t = 4.5.

neous Dirichlet boundary condition. For case II, we solve the problem on
1
[−0.5, 0.5] × [−0.5, 0.5] × [−8, 8] under mesh size hx = hy = 128 and hz = 81 ,
and time step k = 0.0005 with homogeneous Dirichlet boundary condition
along their boundaries.
Fig. 8 shows the condensate widths σx = σy and σz as functions of
time, as well as the surface of the density in xz-plane |ψ(x, 0, z, t)|2 . Similar
phenomena in case I in 3d is observed as those in Example 8 which is in
2d (cf. Fig. 6a). The ratio between the condensate widths increases with
increasing the ratio between trap frequencies along different axis, i.e. it
becomes more difficult to excite oscillations for large trap frequencies. In
case II, the curves of the condensate widths are very well separated. This
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 201

behavior is one of the basic assumptions allowing the reduction of GPE to

2d and 1d in the cases one or two of the trap frequencies are much larger
than the others [85, 5, 8].

12. Derivation of the vector Zakharov system

In this section, we derive VZS from the two-fluid model [113] for ion-electron
dynamics in plasma physics. Here we will use a more formal approach based
on the multiple-scale modulation analysis. Following from [113], we will
consider a plasma as two interpenetrating fluids, an electron fluid and an
ion fluid, and denote the mass, number density (number of particles per
unit volume) and velocity of the electrons (respectively of the ions), by me ,
Ne (x, t) and ve (x, t) (respectively mi , Ni (x, t) and vi (x, t)). The continuity
equations for these fluids read
∂t Ne + ∇ · (Ne ve ) = 0, (12.1)
3
∂t Ni + ∇ · (Ni vi ) = 0, x∈R , t>0 (12.2)
and the momentum equations read
 
1
me Ne (∂t ve + ve · ∇ve ) = −∇pe − e Ne E + ve × B , (12.3)
c
 
1
mi Ni (∂t vi + vi · ∇vi ) = −∇pi + e Ni E + vi × B , (12.4)
c
where −e and e represent the charge of the electron and the ions assumed
to reduce to protons, respectively; pe and pi are the pressure. For small
fluctuations, we write ∇pe = γe Te ∇Ne and ∇pi = γi Ti ∇Ni , where γe and
γi denote the specific heat ratios of the electrons and the ions and Te and
Ti their respective temperatures in energy units. The electric field E and
magnetic field B are provided by the Maxwell equations
1 4π
− ∂t E + ∇ × B = j, ∇ · E = 4πρ, (12.5)
c c
1
∂t B + ∇ × E = 0, ∇ · B = 0, (12.6)
c
where ρ = −e(Ne − Ni ) and j = −e(Ne ve − Ni vi ) are the densities of total
charge and total current, respectively.

Equations (12.5) and (12.6) yield

1 4π
2
∂tt E + ∇ × (∇ × E) + 2 ∂t j = 0, (12.7)
c c
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

202 W. Bao

and using equations (12.1)-(12.4), we have

1 eNe 1
∂t j = e(∇ · (Ne ve )ve + Ne ve · ∇ve + ∇pe + (E + ve × B))
me me c
1 eNi 1
−e(∇ · (Ni vi )vi + Ni vi · ∇vi + ∇pi + (E + vi × B)). (12.8)
mi mi c

In order to get VZS from the two-fluid model just mentioned, as in [113],
we consider a long-wavelength small-amplitude Langmuir oscillation of the
form
ε
E = (E(X, T )e−iωe t + c.c.) + ε2 Ê(X, T ) + · · · , (12.9)
2
where the complex amplitude E depends on the slow variables X = εx and
T = ε2 t, the notation c.c. stands for the complex conjugate and Ê(X, T )
denotes the mean complex amplitude. It induces fluctuations for the density
and velocity of the electrons and of the ions whose dynamical time will be
seen to be τ = εt, thus shorter than T . We write
ε2
Ne = N0 + (Ñe (X, τ )e−iωe t + c.c.) + ε2 N̂e (X, τ ) + · · · , (12.10)
2
ε2
Ni = N0 + (Ñi (X, τ )e−iωe t + c.c.) + ε2 N̂i (X, τ ) + · · · , (12.11)
2
ε
ve = (ṽe (X, τ )e−iωe t + c.c.) + ε2 v̂e (X, τ ) + · · · , (12.12)
2
ε
vi = (ṽi (X, τ )e−iωe t + c.c.) + ε2 v̂i (X, τ ) + · · · , (12.13)
2
where N0 is the unperturbed plasma density.

From the momentum equation (12.3), considering the leading order and
noting that the magnetic field B is of order ε2 , we can easily get
 ε   ε 
me Ne iωe ṽe e−iωe t = e Ne E e−iωe t ,
2 2
thus the amplitude of the electron velocity oscillations is given by
ie
ṽe = − E. (12.14)
me ω e
Neglecting the velocity oscillations of the ions due to their large mass, we
take

ṽi = 0. (12.15)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 203

Applying (12.14) and (12.15) into the continuity equations (12.1) and
(12.2), at the order of ε2 , we have
ε2 −iωe t ε2
−iωe Ñee + N0 ∇ · ṽe e−iωe t = 0,
2 2
thus the density fluctuations are obtained as
N0 eN0
Ñe = −i ∇ · ṽe = − ∇ · E, (12.16)
ωe me ωe2
Ñi = 0. (12.17)

At leading order, the equation for the electric field (12.7) with j =
−e(Ne ve − Ni vi ) becomes
1 2 ε −iωe t 4π ε
ω E e
− + 2 iωe eN0 ṽe e−iωe t = 0,
c2 e 2 c 2
from which, with (12.14), we finally get the electron plasma frequency
s
4πe2 N0
ωe = . (12.18)
me
At the order of ε3 , if no large-scale magnetic field is generated, then the
equation (12.7) with (12.8) implies that
ωe ε3 −iωe t ε3 −iωe t
−2i ∂T E e + ∇ × (∇ × E) e
c2 2 2
4πe2 N0 γe Te ε3 −iωe t 4πe2 N̂e E ε3 −iωe t
− ∇(∇ · E) e + e = 0,
c2 m2e ωe2 2 c2 m e 2
and thus
ωe γe Te 4πe2
−2i ∂T E + ∇ × (∇ × E) − ∇(∇ · E) + N̂e E = 0, (12.19)
c2 m e c2 c2 m e
where, resulting from (12.15) and (12.17), the contribution of the ions is
negligible.

We rewrite the amplitude equation (12.19) as

c2 3v 2 ωe N̂e
i∂T E − ∇ × (∇ × E) + e ∇(∇ · E) = E, (12.20)
2ωe 2ωe 2 N0
where the electron thermal velocity ve is defined by
r
Te
ve = (12.21)
me
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

204 W. Bao

and γe is taken to be 3 [113].

It is seen from (12.3), (12.4) and (12.15) that the mean electron velocity
v̂e and the mean ion velocity v̂i satisfy
 
1 ∗ ∗ γe Te
me ∂τ v̂e + (ṽe · ∇ṽe + ṽe · ∇ṽe ) = − ∇N̂e − eÊ, (12.22)
4 N0
γi Ti
mi ∂τ v̂i = − ∇N̂i + eÊ, (12.23)
N0
where
1 e2
(ṽe · ∇ṽe∗ + ṽe∗ · ∇ṽe ) = ∇|E|2 , (12.24)
4 4m2e ωe2
and ṽe denotes the conjugate of ṽe and me ∂τ v̂e is negligible because of the
small mass of the electron. Furthermore, Ê denotes the leading contribution
(of order ε3 ) of the mean electron field. We thus replace (12.22) by
e2 γe Te
∇|E|2 = − ∇N̂e − eÊ. (12.25)
4me ωe2 N0

The system is closed by using the quasi-neutrality of the plasma in the

form
N̂e = N̂i , (12.26)
v̂e = v̂i , (12.27)
which we denote by N and v, respectively. Then from the continuity equa-
tions, one gets
∂τ N + N0 ∇ · v = 0. (12.28)
Adding (12.25) to (12.23) and noting (12.28), we have
c2s 1
∂τ v = − ∇N − ∇|E|2 , (12.29)
N0 16πmi N0
with the speed of sound cs ,
Te γe Te + γi Ti
c2s = η , η= . (12.30)
mi Te

Finally, we obtain the VZS [113] from equations (12.20), (12.28) and
(12.29) as
c2 3v 2 ωe N
i∂T E − ∇ × (∇ × E) + e ∇(∇ · E) = E, (12.31)
2ωe 2ωe 2 N0
1
ε2 ∂T T N − c2s 4 N = 4 |E|2 . (12.32)
16πmi
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 205

This VZS governs the coupled dynamics of the electric-field amplitude and
of the low-frequency density fluctuations of the ions and describes the dy-
namics of the complex envelope of the electric field oscillations near the
electron plasma frequency and the slow variations of the density perturba-
tions.

In order to obtain a dimensionless form of the system (12.31)-(12.32),

we define the normalized variables
2η 2 1/2 X
t0 = µm ωe T, x0 = (ηµm ) , (12.33)
3 3 ζd
 1/2
3 1 N 1 1 3
N0 = , E0 = E. (12.34)
4η µm N0 η µ1/2
m
64πN0 Te
with
r
Te me
ζd = , µm = , (12.35)
4πe2 N0 mi
where ζd is the Debye length and µm is the ratio of the electron to the ion
masses. Then defining
c2 c2
a= = (12.36)
3ve2 3ωe2 ζd2
and plugging (12.33)-(12.34) into (12.31)-(12.32), and then removing all
primes, we get the following dimensionless vector Zakharov system in three
dimension
i∂t E − a∇ × (∇ × E) + ∇(∇ · E) = N E, (12.37)
ε2 ∂tt N − 4N = 4|E|2 , x ∈ R3 , t > 0. (12.38)
In fact, the equation (12.37) is equalivent to
i∂t E + a 4 E + (1 − a)∇(∇ · E) = N E. (12.39)

13. Generalization and simplification of ZS

The VZS (12.37), (12.38) can be easily generalized to a physical situation
when the dispersive waves interact with M different acoustic modes, e.g. in
a multi-component plasma, which may be described by the following VZSM
[113, 72, 73]:
M
X
i ∂t E + a 4 E + (1 − a) ∇(∇ · E) − E NJ = 0, x ∈ Rd , t > 0, (13.1)
J=1
ε2J ∂tt NJ − 4NJ + νJ 4 |E| = 0,2
J = 1, · · · , M; (13.2)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

206 W. Bao

where the real unknown function NJ is the Jth-component deviation of

the ion density from its equilibrium value, εJ > 0 is a parameter inversely
proportional to the acoustic speed of the Jth-component, and νJ are real
constants.
The VZSM (13.1), (13.2) is time reversible and time transverse invariant,
and preserves the following three conserved quantities. They are the wave
energy
Z
V ZSM
D = |E(x, t)|2 dx, (13.3)
Rd
the momentum
 
Z Xd M
X 2


PV ZSM = i Ej ∇Ej∗ − Ej∗ ∇Ej −
εJ
NJ VJ  dx (13.4)
R d 2 j=1
νJ
J=1

and the Hamiltonian

Z h M
X
H V ZSM = a k∇Ek2l2 + (1 − a)|∇ · E|2 + NJ |E|2
Rd J=1
M  2
X i
1 εJ 1 2
− |VJ |2 + N dx; (13.5)
2 νJ νJ J
J=1

where the flux vector VJ = ((vJ )1 , · · · , (vJ )d )T for the Jth-component is

introduced through the equations
1
∂t NJ = −∇ · VJ , ∂t VJ = − ∇(NJ − νJ |E|2 ), J = 1, · · · , M.
ε2J
(13.6)

13.1. Reduction from VZSM to GVZS

In the VZSM (13.1)-(13.2), if we choose M = 2, and assume that 1/ε22 
1/ε21 , i.e. the acoustic speed of the second component is much faster than
the first component, then formally the fast nondispersive component N2
can be excluded by means of the relation
N2 = ν2 |E|2 + ε22 4−1 ∂tt N2 ≈ ν2 |E|2 + O(ε22 ), when ε2 → 0. (13.7)
Plugging (13.7) into (13.1), then the VZSM (13.1), (13.2) is reduced to
GVZS with N = N1 , ν = ν1 , ε = ε1 , λ = −ν2 and α = 1:
i ∂t E + a 4 E + (1 − a) ∇(∇ · E) − α N E + λ |E|2 E = 0, (13.8)
ε2 ∂tt N − 4N + ν 4 |E|2 = 0, x ∈ Rd , t > 0. (13.9)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 207

The GVZS (13.8), (13.9) is time reversible, time transverse invariant

and preserves the following three conserved quantities, i.e. the wave energy,
momentum and Hamiltonian:
Z
GV ZS
D = |E(x, t)|2 dx, (13.10)
Rd
 
Z Xd

αε2
PGV ZS = i Ej ∇Ej∗ − Ej∗ ∇Ej − N V dx, (13.11)
Rd 2 j=1 ν
Z 
GV ZS λ
H = a k∇Ek2l2 + (1 − a)|∇ · E|2 + αN |E|2 − |E|4
R d 2
2

α αε
− N2 − |V|2 dx; (13.12)
2ν 2ν
where the flux vector V = (v1 , · · · , vd )T is introduced through the equations
1
∂t N = −∇ · V, ∂t V = − ∇(N − ν|E|2 ). (13.13)
ε2
In the case of M = 2, ν = ν1 and ε = ε1 , N = N1 and V = V1 in
(13.4) and (13.5), and λ = −ν2 , α = 1 in (13.11), (13.12), letting ε2 → 0 and
noting (13.7), we get formally quadratic convergence rate of the momentum
and Hamiltonian from VZSM to GVZS in the ‘subsonic limit’ regime of the
second component, i.e., 0 < ε2  1:
 
Z Xd 2
i
ε
PV ZSM =  Ej ∇Ej∗ − Ej∗ ∇Ej − 1 N1 V dx
Rd 2 j=1 ν1
Z
ε2
− 2 N2 V2 dx
ν2 Rd
≈ PGV ZS + O(ε22 ), (13.14)
Z 
1
H V ZSM = a k∇Ek2l2 + (1 − a)|∇ · E|2 + N1 |E|2 − N2
Rd 2ν1 1
 Z  
ε2 1 2 ε2
− 1 |V1 |2 dx + N2 |E|2 − N2 − 2 |V2 |2 dx
2ν1 Rd 2ν2 2ν2
≈ H GV ZS + O(ε22 ). (13.15)
Choosing a = 1, α = 1, ν = −1 and λ = 0 in the GVZS (13.8)-(13.9), it
collapses to the standard VZS [113]
i ∂t E + 4E − N E = 0, x ∈ Rd , t > 0, (13.16)
2 2
ε ∂tt N − 4N − 4|E| = 0. (13.17)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

208 W. Bao

13.2. Reduction from GVZS to GZS

In the case when E2 = · · · = Ed = 0 and a = 1 in the GVZS (13.8), (13.9),
it reduces to the scalar GZS [113, 15],

i ∂t E + 4E − α N E + λ |E|2 E = 0, x ∈ Rd , t > 0, (13.18)

2 2
ε ∂tt N − 4N + ν 4 |E| = 0. (13.19)

The GZS (13.18), (13.19) is time reversible, time transverse invariant

and conserved the following wave energy, momentum and Hamiltonian:
Z
GZS
D = |E(x, t)|2 dx, (13.20)
Rd
Z  
i ε2 α
PGZS = (E∇E ∗ − E ∗ ∇E) − N V dx, (13.21)
Rd 2 ν
Z  
λ α 2 αε2
H GZS = |∇E|2 + αN |E|2 − |E|4 − N − |V|2 dx;(13.22)
Rd 2 2ν 2ν
where the flux vector V = (v1 , · · · , vd )T is introduced through the equations
1
Nt = −∇ · V, Vt = − ∇(N − ν|E|2 ). (13.23)
ε2
Choosing α = 1, ν = −1, ε = 1 and λ = 0 in the GZS (13.18)-(13.19), it
collapses to the standard ZS [113, 15, 121]. When λ 6= 0, a cubic nonlinear
term is added to the standard ZS.

Proof of the conservation laws in GZS: Multiplying (13.18) by E, the

conjugate of E, we get

iEt E ∗ + E ∗ 4 E − α N |E|2 + λ |E|4 = 0. (13.24)

Then calculating the conjugate of (13.24) and multiplying it by E, one finds

−iEt∗ E + E 4 E ∗ − α N |E|2 + λ |E|4 = 0. (13.25)

Subtracting (13.25) from (13.24) and then multiplying both sides by −i,
one gets

Et E ∗ + Et∗ E + i(E 4 E ∗ − E ∗ 4 E) = 0. (13.26)

Integrating over Rd , integration by parts, (13.26) leads to the conservation

of the wave energy
Z
d GZS d
D = |E(x, t)|2 dx = 0.
dt dt Rd
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 209

From (13.21), noting (13.23), (13.18), one has the conservation of the
momentum
Z
d GZS i
P = (Et ∇E ∗ + E∇Et∗ − E ∗ ∇Et − Et∗ ∇E) dx
dt 2 Rd
Z
ε2 α
− (Nt V + N Vt ) dx
ν Rd
Z Z
∗ ∗ ε2 α
=i (Et ∇E − Et ∇E) dx − (Nt V + N Vt ) dx
d ν Rd
ZR
=i ∇E ∗ (i 4 E − iαN E + iλ|E|2 E)dx
Rd
Z
ε2 α
− (Nt V + N Vt ) dx
ν d
Z R
−i ∇E(−i 4 E ∗ + iαN E ∗ − iλ|E|2 E ∗ ) dx
Rd
Z Z
2 ε2 α
=α N ∇|E| dx + V∇ · V dx
Rd ν Rd
Z
α
+ ∇(N − ν|E|2 )N dx = 0.
ν Rd

Noting (13.23), (13.19) and multiplying (13.18) by Et∗ , the conjugate of

Et , we write it
Z
 
T = i|Et |2 + Et∗ 4 E − αN EEt∗ + λ|E|2 EEt∗ dx = 0. (13.27)
Rd

Then the real part of T is

Z
 ∗ 
0 = Re(T ) = Re Et 4 E − αN EEt∗ + λ|E|2 EEt∗ dx
Rd
Z  
∗ α 2 α 2 λ 4
= Re −∇E∇Et − (N |E| )t + Nt |E| + (|E| )t dx
Rd 2 2 4
Z   
1 2 2 λ 4 α 2
=− |∇E| + αN |E| − |E| + Nt |E| dx
2 Rd 2 2
Z  t 
1 2 2 λ 4 α 2
=− |∇E| + αN |E| − |E| − |E| ∇ · V dx
2 Rd 2 t 2
Z   
1 2 2 λ 4 α 2
=− |∇E| + αN |E| − |E| + ∇|E| · V dx
2 Rd 2 t 2
Z   
1 λ α ε2 1
=− |∇E|2 + αN |E|2 − |E|4 + ( Vt + ∇N ) · V dx.
2 Rd 2 t 2 ν ν
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

210 W. Bao

Thus we have,
Z  
1 2 2 λ 4
0=− |∇E| + αN |E| − |E| dx
2 Rd 2
Z Z
αε2 1 α
+ (|V|2 )t dx − N ∇ · V dx
2ν Rd 2 2ν Rd
Z  
1 2 2 λ 4
=− ∂t |∇E| + αN |E| − |E| dx
2 Rd 2
Z Z
αε2 1 α
+ (|V|2 )t dx + N ∂t N dx
2ν Rd 2 2ν Rd
Z  
1 2 2 λ 4 α 2 αε2 2
=− ∂t |∇E| + αN |E| − |E| − N − |V| ) dx,
2 Rd 2 2ν 2ν
which implies the conservation of Hamiltonian
d GZS
H = 0.
dt

13.3. Reduction from GVZS to VNLS

In the “subsonic limit”, i.e. ε → 0 in GVZS (13.8), (13.9), which corresponds
to that the density fluctuations are assumed to follow adiabatically the
modulation of the Langmuir wave, it collapses to the VNLS equation. In
fact, letting ε → 0 in (13.9), we get formally
N = ν |E|2 + ε2 4−1 ∂tt N = ν |E|2 + O(ε2 ), when ε → 0. (13.28)
Plugging (13.28) into (13.8), we obtain formally the VNLS:
i ∂t E + a 4 E + (1 − a) ∇(∇ · E) + (λ − αν)|E|2 E = 0, x ∈ Rd , t > 0.
(13.29)
The VNLS (13.29) is time reversible, time transverse invariant and preserves
the following wave energy, momentum and Hamiltonian:
Z
DV N LS = |E(x, t)|2 dx, (13.30)
Rd
Z d
iX

PV N LS = Ej ∇Ej∗ − Ej∗ ∇Ej dx, (13.31)
Rd 2 j=1
Z  
αν − λ 4
H V N LS = a k∇Ek2l2 + (1 − a)|∇ · E|2 + |E| dx. (13.32)
Rd 2
Letting ε → 0 in (13.11), (13.12), noting (13.28), we get formally
quadratic convergence rate of the momentum and Hamiltonian from GVZS
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 211

to VNLS in the ‘subsonic limit’ regime, i.e., 0 < ε  1:

Z d Z
iX
αε2
PGV ZS = Ej ∇Ej∗ − Ej∗ ∇Ej dx − N V dx
Rd 2 j=1 ν Rd

≈ PV N LS + O(ε2 ),
Z  
αν − λ 4
H GV ZS = a k∇Ek2l2 + (1 − a)|∇ · E|2 + |E| dx
Rd 2
2 Z
αε 2
− |V| dx
2ν Rd
≈ H V N LS + O(ε2 ).

13.4. Reduction from GZS to NLSE

Similarly, in the “subsonic limit”, i.e. ε → 0 in GZS (13.18), (13.19), it
collpases to the well-known NLSE with a cubic nonlinearity. In fact, letting
ε → 0 in (13.19), we get formally
N = ν |E|2 + ε2 4−1 ∂tt N = ν |E|2 + O(ε2 ), when ε → 0. (13.33)
Plugging (13.33) into (13.18), we obtain formally the NLS equation:
i Et + 4 E + (λ − αν)|E|2 E = 0, x ∈ Rd , t > 0. (13.34)
The NLSE (13.34) is time reversible, time transverse invariant, and pre-
serves the following wave energy, momentum and Hamiltonian:
Z
DN LS = |E(x, t)|2 dx, (13.35)
Rd
Z  
i
PN LS = (E∇E ∗ − E ∗ ∇E) dx, (13.36)
Rd 2
Z  
αν − λ 4
H N LS = |∇E|2 + |E| dx. (13.37)
Rd 2
Similarly, letting ε → 0 in (13.21), (13.22), noting (13.33), we get formally
the quadratic convergence rate of the momentum and Hamiltonian from
GZS to NLSE in the ‘subsonic limit’ regime, i.e., 0 < ε  1:
Z Z
GZS i ∗ ∗ ε2 α
P = (E∇E − E ∇E) dx − N V dx
Rd 2 ν Rd
≈ PN LS + O(ε2 ), (13.38)
Z   2 Z
αν − λ 4 αε
H GZS = |∇E|2 + |E| dx − |V|2 dx
R d 2 2ν R d

≈ H N LS + O(ε2 ). (13.39)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

212 W. Bao

13.5. Add a linear damping term to arrest blowup

When d ≥ 2 and initial Hamiltonian H GZS < 0 in the GZS (13.18), (13.19),
mathematically, it will blowup in finite time [113]. However, the physical
quantities modeled by E and N do not become infinite which implies the
validity of (13.18), (13.19) breaks down near singularity. Additional physical
mechanisms, which were initially small, become important near the singular
point and prevent the formation of singularity. In order to arrest blowup, in
physical literatures, a small linear damping (absorption) term is introduced
into the GZS [64]:
i ∂t E + 4E − α N E + λ |E|2 E + i γ E = 0, (13.40)
2 2 d
ε ∂tt N − 4N + ν 4 |E| = 0, x∈R , t > 0; (13.41)
where γ > 0 is a damping parameter. The decay rate of the wave energy
DGZS of the damped GZS (13.40), (13.41) is
Z Z
DGZS (t) = |E(x, t)|2 dx = e−2γt |E(x, 0)|2 dx
Rd Rd
−2γt GZS
=e D (0), t ≥ 0. (13.42)
GV ZS
Similarly, when d ≥ 2 and initial Hamiltonian H < 0 in the GVZS
(13.8), (13.9) (or H V ZSM < 0 in the VZSM (13.1), (13.2)), mathematically,
it will blowup in finite time too. In order to arrest blowup, in physical
literatures, a small linear damping (absorption) term is introduced into the
GVZS (or VZSM):
i ∂t E + a 4 E + (1 − a) ∇(∇ · E) − α N E + λ |E|2 E + i γ E = 0, (13.43)
ε2 ∂tt N − 4N + ν 4 |E|2 = 0, x ∈ Rd , t > 0; (13.44)
where γ > 0 is a damping parameter. The decay rate of the wave energy
DGV ZS of the damped GVZS (13.43), (13.44) is
Z Z
2 2
DGV ZS (t) = |E(x, t)| dx = e−2γt |E(x, 0)| dx
Rd Rd
−2γt GV ZS
=e D (0), t ≥ 0. (13.45)

14. Well-posedness of ZS
Based on the conservation laws, the wellposedness for the standard ZS
(13.16)-(13.17) were proven [113, 112, 23, 24]
Theorem 14.1: In one dimension, for initial conditions, E 0 ∈ H p (R),
N 0 ∈ H p−1 (R), and N (1) ∈ H p−2 (R) with p ≤ 3, there exists a unique
solution E ∈ L∞ (R+ , H p (R)), N ∈ L∞ (R+ , H p−1 (R)) for (13.16)-(13.17).
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 213

Theorem 14.2: In dimensions 2 and 3, for initial conditions E 0 ∈

H p (Rd ), N 0 ∈ H p−1 (Rd ), and N (1) ∈ H p−2 (Rd ) with p ≤ 3, there exists a
unique solution E ∈ L∞ ([0, T ∗ ), H p (Rd )), N ∈ L∞ ([0, T ∗ ), H p−1 (Rd )) for
(13.16)-(13.17), where time T ∗ depends on the initial conditions.

15. Plane wave and soliton wave solutions of ZS

In one spatial dimension (1D), the GZS (13.40)- (13.41) collapses to
i Et + Exx − αN E + λ|E|2 E + iγ E = 0, a < x < b, t > 0, (15.1)
2 2
ε Ntt − Nxx + ν(|E| )xx = 0, a < x < b, t > 0, (15.2)
which admits plane wave and soliton wave solutions.

Firstly, it is instructive to examine some explicit solutions to (15.1)

and (15.2). The well-known plane wave solutions [97] can be given in the
following form:
N (x, t) = d, a < x < b, t ≥ 0, (15.3)
(
c ei( ),
2πrx
b−a −ω1 t γ = 0,
E(x, t) = “
2πrx λc2 −2γt
” (15.4)
−γt i b−a −ω2 t− 2γ (e −1)
ce e , γ 6= 0,
where r is an integer, c, d are constants and
4π 2 r2 4π 2 r2
ω1 = αd + − λc2 , ω2 = αd + .
(b − a)2 (b − a)2

Secondly, as is well known, the standard ZS is not exactly integrable.

Therefore the generalized ZS cannot be exactly integrable, either. However,
it has exact one-soliton solutions to (15.1) and (15.2) for γ = 0 [72] for
x ∈ R and t ≥ 0:
 −1/2
λ αν 2 2 −1
Es (x, t; η, V, ε, ν) = − 2 (1/ε − V ) Us , (15.5)
2 2ε
 
Us ≡ 2iη sech[2η(x − V t)] exp iV x/2 + i(4η 2 − V 2 /4)t + iΦ0 , (15.6)
ν
Ns (x, t; η, V, ε, ν) = 2 (1/ε2 − V 2 )−1 |Es |2 , (15.7)
ε
where η and V are the soliton’s amplitude and velocity, respectively, and
Φ0 is a trivial phase constant.

Finally, we will consider the periodic soliton solution with a period L

in 1d of the standard ZS, that is, d = 1, ε = 1, α = 1, λ = 0, γ = 0 and
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

214 W. Bao

ν = −1 in (13.40)-(13.41). The analytic solution of the ZS (15.1)-(15.2)

was derived [100] and used to test different numerical methods for the ZS
in [100, 28]. The solution can be written as

Es (x, t; v, Emax ) = F (x − vt) exp[iφ(x − ut)], (15.8)

Ns (x, t; v, Emax ) = G(x − vt), (15.9)

where
|F (x − vt)|2
F (x − vt) = Emax · dn(w, q), G(x − vt) = + N0 ,
v2 − 1
q
2
(Emax 2
− Emin )
Emax
w= p · (x − vt), q= ,
(2(1 − v 2 )) Emax
v v 2N0 E 2 + Emin 2
φ = v/2, L = 2πm, m = 1, 2, 3 · · · , u = + − max ,
2 2 v v(1 − v 2 )
p p  
2 2(1 − v 2 ) 2 2(1 − v 2 ) 0 Emin
L= K(q) = K ,
Emax Emax Emax

with dn(w, q) a Jacobian elliptic function, L the period of the Jaco-

bian elliptic functions, K and K 0 the complete
 elliptic integrals of the
p
0 2
first kind satisfying K(q) = K 1 − q , and N0 chosen such that
Z L
hNs i = L1 Ns (x, t) dx = 0.
0

16. Time-splitting spectral method for GZS

In this section we present new numerical methods for the GZS (13.40),
(13.41). For simplicity of notations, we shall introduce the method in one
space dimension (d = 1) of the GZS with periodic boundary conditions.
Generalizations to d > 1 are straightforward for tensor product grids and
the results remain valid without modifications. For d = 1, the problem
becomes

i ∂t E + ∂xx E − αN E + λ|E|2 E + iγ E = 0, a < x < b, t > 0, (16.1)

ε2 ∂tt N − ∂xx (N − ν |E|2 ) = 0, a < x < b, t > 0, (16.2)
(0) (0) (1)
E(x, 0) = E (x), N (x, 0) = N (x), ∂t N (x, 0) = N (x), (16.3)
E(a, t) = E(b, t), ∂x E(a, t) = ∂x E(b, t), t ≥ 0, (16.4)
N (a, t) = N (b, t), ∂x N (a, t) = ∂x N (b, t), t ≥ 0. (16.5)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 215

Moreover, we supplement (16.1)-(16.5) by imposing the compatibility con-

dition
E (0) (a) = E (0) (b), N (0) (a) = N (0) (b),
Z b
N (1) (a) = N (1) (b), N (1) (x) dx = 0. (16.6)
a
As is well known, the GZS has the following property
Z b Z b
DGZS (t) = |E(x, t)|2 dx = e−2γt |E (0) (x)|2 dx
a a
= e−2γt DGZS (0), t ≥ 0. (16.7)
When γ = 0, DGZS (t) ≡ DGZS (0), i.e., it is an invariant of the GZS [28].
When γ > 0, it decays to 0 exponentially. Furthermore, the GZS also has
the following properties for t ≥ 0
Z b Z b Z b
∂t N (x, t) dx = 0, N (x, t) dx = N (0) (x) dx = const. (16.8)
a a a
In some cases, the boundary conditions (16.4) and (16.5) may be re-
placed by
E(a, t) = E(b, t) = 0, N (a, t) = N (b, t) = 0, t ≥ 0. (16.9)
We choose the spatial mesh size h = 4x > 0 with h = (b − a)/M for M
being an even positive integer, the time step k = 4t > 0 and let the grid
points and the time step be
xj := a+j h, j = 0, 1, · · · , M ; tm := m k, m = 0, 1, 2, · · · .
Let Ejm and Njm
be the approximations of E(xj , tm ) and N (xj , tm ), re-
spectively. Furthermore, let E m and N m be the solution vector at time
t = tm = mk with components Ejm and Njm , respectively.
From time t = tm to t = tm+1 , the first NLS-type equation (16.1) is
solved in two splitting steps. One solves first
i ∂t E + ∂xx E = 0, (16.10)
for the time step of length k, followed by solving
i ∂t E = αN E − λ|E|2 E − iγ E, (16.11)
for the same time step. Equation (16.10) will be discretized in space by the
Fourier spectral method and integrated in time exactly. For t ∈ [tm , tm+1 ],
multiplying (16.11) by E, we get
i ∂t E E ∗ = αN |E|2 − λ|E|4 − iγ|E|2 . (16.12)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

216 W. Bao

Then calculating the conjugate of the ODE (16.11) and multiplying it by

E, one finds
−i ∂t E ∗ E = αN |E|2 − λ|E|4 + iγ|E|2 . (16.13)
Subtracting (16.13) from (16.12) and then multiplying both sides by −i,
one gets
∂t (|E(x, t)|2 ) = ∂t E(x, t)E(x, t)∗ + ∂t E(x, t)∗ E(x, t)
= −2γ|E(x, t)|2 (16.14)
and therefore
|E(x, t)|2 = e−2γ(t−tm ) |E(x, tm )|2 , tm ≤ t ≤ tm+1 . (16.15)
Substituting (16.15) into (16.11), we obtain
i∂t E(x, t) = αN (x, t)E(x, t) − λe−2γ(t−tm ) |E(x, tm )|2 E(x, t)
−iγE(x, t). (16.16)
Integrating (16.16) from tm to tm+1 , and then approximating the integral
of N on [tm , tm+1 ] via the trapezoidal rule, one obtains
R tm+1 −2γ(τ −tm ) 2
E(x, tm+1 ) = e−i tm [αN (x,τ )−λe |E(x,tm )| −iγ] dτ
E(x, tm )
( 2
e−ik[α(N (x,tm )+N (x,tm+1 ))/2−λ|E(x,tm )| ] E(x, tm ), γ = 0,
≈ 2 −2γk
e−γk−i[kα(N (x,tm )+N (x,tm+1 ))/2+λ|E(x,tm )| (e −1)/2γ]
E(x, tm ), γ =
6 0.

16.1. Crank-Nicolson leap-frog time-splitting spectral

discretizations (CN-LF-TSSP) for GZS
The second wave-type equation (16.2) in the GZS is discretized by pseudo-
spectral method for spatial derivatives, and then applying Crank-Nicolson
/leap-frog for linear/nonlinear terms for time derivatives:
Njm+1 − 2Njm + Njm−1 h

ε2 2
− D f
xx βN m+1 + (1 − 2β)N m + βN m−1
i k
−ν|E m |2 = 0, j = 0, · · · , M, m = 1, 2, · · · , (16.17)
x=xj
f
where 0 ≤ β ≤ 1 is a constant, Dxx , a spectral differential operator approx-
imation of ∂xx , is defined as
M/2−1
 X
f
Dxx U x=xj = − el eiµl (xj −a)
µ2l U (16.18)
l=−M/2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 217

el , the Fourier coefficients of a vector U = (U0 , U1 , U2 , · · · , UM )T with

and U
U0 = UM , are defined as
M−1
2πl e 1 X M M
µl = , Ul = Uj e−iµl (xj −a) , l = − , · · · , − 1.
b−a M j=0 2 2
(16.19)
When β = 0 in (16.17), the discretization (16.17) to the wave-type equa-
tion (16.2) is explicit and was used in [15, 114]. When 0 < β ≤ 1, the
discretization is implicit, but can be solved explicitly. In fact, suppose
M/2−1
X
Njm = g
(N m ) eiµl (xj −a) ,
l j = 0, · · · , M ; m = 0, 1, · · · ,
l=−M/2
(16.20)
Plugging (16.20) into (16.17), using the orthogonality of the Fourier basis,
we obtain for m ≥ 1
^
(N ^
m+1 ) − 2(N m ) + (N ^ m−1 ) h
ε2 l l l
+ µ2
β (N^ ^
m+1 ) + (1 − 2β)(N m)
l l l
k2
i M M
^
+β (N ^
m−1 ) − ν (|E m |2 )
l l = 0, l = − ,··· , − 1. (16.21)
2 2
Solving the above equation, we get
 
^ m+1
k 2 µ2l ^ ^
m ) − (N m−1 ) +
νk 2 µ2l ^ m |2 ) ,
(N )l = 2 − 2 (N l l (|E l
ε + βk 2 µ2l ε2 + βk 2 µ2l
l = −M/2, · · · , M/2 − 1; m = 1, 2, · · · . (16.22)

From time t = tm to t = tm+1 , we combine the splitting steps via the

standard Strang splitting for m ≥ 0:
M/2−1
X
Njm+1 = ^
(N m+1 ) eiµl (xj −a) , (16.23)
l
l=−M/2
M/2−1
X 2
Ej∗ = g
e−ikµl /2 (E m ) eiµl (xj −a) ,
l
l=−M/2
( m m+1 ∗ 2
e−ik[α(Nj +Nj )/2−λ|Ej | ] Ej∗ , γ = 0,
Ej∗∗ = m m+1 ∗ 2 −2γk
e−γk−i[kα(Nj +Nj )/2+λ|Ej | (e −1)/2γ]
Ej∗ , γ 6= 0,
M/2−1
X 2
Ejm+1 = g
e−ikµl /2 (E ∗∗ ) eiµl (xj −a) , 0 ≤ j ≤ M − 1;
l (16.24)
l=−M/2
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

218 W. Bao

where (N ^ m+1 ) is given in (16.22) for m > 0 and (16.27) for m = 0. The
l
initial conditions (16.3) are discretized as

Nj1 − Nj−1 (1)

Ej0 = E (0) (xj ), Nj0 = N (0) (xj ), = Nj , j = 0, 1, 2, · · · , M −1,
2k
(16.25)
where
 (1)
 N (xj ),
 0 ≤ j ≤ M − 2,
(1) M−2
X
Nj = (16.26)

 − N (1) (xl ), j = M − 1.
l=0

This implies that

 
] 1) =
k 2 µ2l ^ ^
(0) ) + k (N (1) )
(N l 1− (N l l
2(ε2 + βk 2 µ2l )
νk 2 µ2l ^ (0) |2 ) , l = −
M M
+ 2 (|E l ,··· , − 1. (16.27)
2(ε2 2
+ βk µl ) 2 2

This type of discretization for the initial condition (16.3) is equivalent to

use the trapezoidal rule for the periodic function N (1) and such that (16.8)
is satisfied in discretized level. The discretization error converges to 0 ex-
ponentially fast as the mesh size h goes to 0.
Note that the spatial discretization error of the method is of spectral-
order accuracy in h and time discretization error is demonstrated to be
second-order accuracy in k from our numerical results.

16.2. Phase space analytical solver + time-splitting spectral

discretizations (PSAS-TSSP)
Another way to discretize the second wave-type equation (16.2) in GZS
is by pseudo-spectral method for spatial derivatives, and then solving the
ODEs in phase space analytically under appropriate chosen transmission
conditions between different time intervals. From time t = tm to t = tm+1 ,
assume
M/2−1
X
N (x, t) = e m (t) eiµl (x−a) ,
N a ≤ x ≤ b, tm ≤ t ≤ tm+1 .
l
l=−M/2
(16.28)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 219

Plugging (16.28) into (16.2) and noticing the orthogonality of the Fourier
series, we get the following ODEs for m ≥ 0:
e m (t)
d2 N h i
ε2 l
+ µ2 em
N (t) − ν ^
(|E(t )|2)
m l = 0, tm ≤ t ≤ tm+1 , (16.29)
d t2 ( l l

^ (0)
elm (tm ) = (N )l ,
N
m = 0,
l = −M/2, · · · , M/2 − 1. (16.30)
e
N m−1
(tm ), m > 0,
l

For each fixed l (−M/2 ≤ l ≤ M/2 − 1), Eq. (16.29) is a second-order ODE.
It needs two initial conditions such that the solution is unique. When m = 0
in (16.29), (16.30), we have the initial condition (16.30) and we can pose
the other initial condition for (16.29) due to the initial condition (16.3) for
the GZS (16.1)-(16.5):
d e0 d e0 ^ (1) ) ,
N (t0 ) = N (0) = (N l = −M/2, · · · , M/2 − 1. (16.31)
dt l dt l l

Then the solution of (16.29), (16.30) with m = 0 and (16.31) is:


 ^ (0) ^ (1)
 (N )0 + t (N )0 ,


l = 0,

 
Ne 0 (t) = (16.32)
l
 ^
(N (0) ) − ν (|E^ ^
(0) |2 ) cos(µ t/ε) + ν (|E (0) |2 )

 l l l l


 ε ^
+ µl (N (1) )l sin(µl t/ε), l 6= 0,
0 ≤ t ≤ t1 , l = −M/2, · · · , M/2 − 1.

But when m > 0, we only have one initial condition (16.30). One can’t sim-
d em d e m−1
ply pose the continuity between dt Nl (t) and dt Nl (t) across the time
t = tm due to the last term in (16.29) is usually different in two adjacent
^
time intervals [tm−1 , tm ] and [tm , tm+1 ], i.e. (|E(t 2 ^2
m−1 )| ) 6= (|E(tm )| ) .
l l
Since our goal is to develop explicit scheme and we need linearize the non-
linear term in (16.2) in our discretization (16.29), in general,
d e m−1 − d e m−1 d em d em +
Nl (tm ) = lim− N (t) 6= lim+ N l (t) = N (t ), (16.33)
dt t→tm dt
l
t→tm dt dt l m
m = 1, · · · , l = −M/2, · · · , M/2 − 1.
d em + d e m−1 −
Unfortunely, we don’t know the jump dt Nl (tm ) − dt Nl (tm ) across the
time t = tm . In order to get a unique solution of (16.29), (16.30) for m > 0,
here we pose an additional condition:
elm (tm−1 ) = N
N e m−1 (tm−1 ), l = −M/2, · · · , M/2 − 1. (16.34)
l
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

220 W. Bao

The condition (16.34) is equivalent to pose the solution N e m (t) on the time
l
interval [tm , tm+1 ] of (16.29), (16.30) is also continuity at the time t = tm−1 .
After a simple computation, we get the solution of (16.29), (16.30) and
(16.34) for m > 0:
 h i

 Ne m−1 (tm ) + t−tm N e m−1 (tm ) − N e m−1 (tm−1 ) , l = 0,

 0 k 0 0




 h m−1 i
e (t) =
m e
N (t ) − ν ^
(|E m |2 ) cos(µ (t − t )/ε)
N l l m l
h
l m (16.35)



 +ν (|E^ m |2 ) + sin(µ l (t−tm )/ε) e
N m−1
(tm ) cos(kµl /ε)

 l sin(kµl /ε) l
i


 −N e m−1 (tm−1 ) + ν [1 − cos(kµl /ε)] (|E ^ m |2 ) , l 6= 0,
l l

tm ≤ t ≤ tm+1 , l = −M/2, · · · , M/2 − 1.

From time t = tm to t = tm+1 , we combine the splitting steps via the
standard Strang splitting for m ≥ 0:
M/2−1
X
Njm+1 = elm (tm+1 ) eiµl (xj −a) ,
N (16.36)
l=−M/2
M/2−1
X 2
Ej∗ = g
e−ikµl /2 (E m ) eiµl (xj −a) ,
l
l=−M/2
( m m+1 ∗ 2
e−ik[α(Nj +Nj )/2−λ|Ej | ] Ej∗ , γ = 0,
Ej∗∗ = m m+1 ∗ 2 −2γk
e−γk−i[kα(Nj +Nj )/2+λ|Ej | (e −1)/2γ]
Ej∗ , γ 6= 0,
M/2−1
X 2
Ejm+1 = g
e−ikµl /2 (E ∗∗ ) eiµl (xj −a) , 0 ≤ j ≤ M − 1;
l (16.37)
l=−M/2

where


 ^
(N ^
(0) ) + k (N (1) ) , l = 0, m = 0,

 0 0





 ^ ε ^
 (N
 (0) ) cos(kµ /ε) +
l l µl (N
(1) )
sin(kµl /ε) l 6= 0, m = 0,
l
e m (tm+1 )
N = ^ (16.38)
l
 +ν [1 − cos(kµl /ε)] (|E (0) |2 )l ,







 e m−1 (tm ) cos(kµl /ε) − N e m−1 (tm−1 ) m ≥ 1,
 2 Nl
 l
 ^ m |2 ) ,
+2ν [1 − cos(kµl /ε)] (|E l
The initial conditions (16.3) are discretized as
(1)
Ej0 = E (0) (xj ), Nj0 = N (0) (xj ), (∂t N )0j = Nj , j = 0, 1, 2, · · · , M − 1.
(16.39)
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 221

Note that the spatial discretization error of the above method is again of
spectral-order accuracy in h and time discretization error is demonstrated
to be second-order accuracy in k from our numerical results.

16.3. Properties of the numerical methods

(1). Plane wave solution: If the initial data in (16.3) is chosen as

E (0) (x) = c ei2πlx/(b−a) , N (0) (x) = d, N (1) (x) = 0,

a ≤ x ≤ b,
(16.40)
where l is an integer and c, d are constants, then the GZS (16.1)-(16.5)
admits the plane wave solution [97]

N (x, t) = d, a < x < b, t ≥ 0, (16.41)

(
ei( b−a “−ω1 t) ,
2πlx
c γ = 0,
E(x, t) = 2πlx λc2 −2γt
” (16.42)
−γt i b−a −ω2 t− 2γ (e −1)
ce e , γ 6= 0.
where
4π 2 l2 4π 2 l2
ω1 = αd + − λc2 , ω2 = αd + .
(b − a)2 (b − a)2
It is easy to see that in this case our numerical methods CN-LF-TSSP
(16.23), (16.24) and PAAS-TSSP (16.36), (16.37) give exact results pro-
vided that M ≥ 2(|l| + 1).
(2). Time transverse invariant: A main advantage of CN-LF-TSSP and
PAAS-TSSP is that if a constant r is added to the initial data N 0 (x) in
(16.3) when γ = 0 in (16.1), then the discrete functions Njm+1 obtained
from (16.23) or (16.36) get added by r and Ejm+1 obtained from (16.24)
or (16.37) get multiplied by the phase factor e−ir(m+1)k , which leaves the
discrete function |Ejm+1 |2 unchanged. This property also holds for the exact
solution of GZS, but does not hold for the finite difference schemes proposed
in [62, 28] and the spectral method proposed in [100].
(3). Conservation: Let U = (U0 , U1 , · · · , UM )T with U0 = UM , f (x) a
periodic function on the interval [a, b], and let k · kl2 be the usual discrete
l2 -norm on the interval (a, b), i.e.,
v v
u u
u b − a M−1
X u b − a M−1
X
kU kl2 = t 2
|Uj | , kf kl2 = t |f (xj )|2 . (16.43)
M j=0 M j=0

Then we have
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

222 W. Bao

Theorem 16.1: The CN-LF-TSSP (16.23), (16.24) and PSAS-TSSP

(16.36), (16.37) for GZS possesses the following properties (in fact, they
are the discretized version of (16.7) and (16.8)):

kE m k2l2 = e−2γtm kE 0 k2l2 = e−2γtm kE (0) k2l2 , m = 0, 1, 2, · · · , (16.44)

M−1
b−a X Njm+1 − Njm
= 0, m = 0, 1, 2, · · · (16.45)
M j=0
k

and
M−1 M−1 M−1
b−a X b−a X b−a X
Njm = Nj0 = N (0) (xj ), m ≥ 0.
M j=0 M j=0 M j=0
(16.46)

Proof: From (16.43), (16.37) and (16.19), using the orthogonality of the
discrete Fourier series and noticing the Pasavel equality, we have

M
M−1
X X 
M/2−1 2
∗∗ ) 
2
kE m+1 2
k l2 = |Ejm+1 |2 = M g
e−ikµl /2 (E l
b−a j=0 l=−M/2
M/2−1 M−1
X X
=M g
|(E ∗∗ ) |2 =
l |Ej∗∗ |2
l=−M/2 j=0
M−1
X M−1
X
−2γk
=e |Ej∗ |2 =e −2γk
|Ejm |2
j=0 j=0
M
= e−2γk kE m k2l2 , m ≥ 0. (16.47)
b−a
Thus (16.44) is obtained from (16.47) by induction. The equalities (16.45)
and (16.46) can be obtained in a similar way.
(4). Unconditional stability: By the standard Von Neumann analysis for
(16.23) and (16.36), noting (16.44), we get PSAS-TSSP and CN-LF-TSSP
with 1/4 ≤ β ≤ 1 are unconditionally stable, and CN-LF-TSSP with 0 ≤
β < 1/4 is conditionally stable with stability constraint k ≤ √ 2hε in
π d(1−4β)
d-dimensions ( d = 1, 2, 3). In fact, for PSAS-TSSP (16.36), (16.37), setting
^
(|E m |2 ) = 0 and plugging N e m (tm+1 ) = µN
e m−1 (tm ) = µ2 N
e m−1 (tm−1 )
l l l l
into (16.38) with |µ| the amplification factor, we obtain the characteristic
equation:

µ2 − 2 cos(kµl /ε)µ + 1 = 0. (16.48)

April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 223

This implies
µ = cos(kµl /ε) ± i sin(kµl /ε). (16.49)
Thus the amplification factor
q
Gl = |µ| = cos2 (kµl /ε) + sin2 (kµl /ε) = 1, l = −M/2, · · · M/2 − 1.
This together with (16.44) imply that PSAS-TSSP is unconditionally stable.
Similarly for CN-LF-TSSP (16.23), (16.24), noting (16.22), we have the
characteristic equation:
 
k 2 µ2l
µ2 − 2 − 2 µ + 1 = 0. (16.50)
ε + βk 2 µ2l
This implies
s 2
k 2 µ2l k 2 µ2l
µ=1− ± 1− − 1. (16.51)
2(ε + βk 2 µ2l )
2 2(ε + βk 2 µ2l )
2

When 1/4 ≤ β ≤ 1, we have

 
 k 2 µ2l 
1 −  ≤ 1, k > 0, l = −M/2, · · · M/2 − 1.
 2(ε + βk µl ) 
2 2 2

Thus
s  2
k 2 µ2l k 2 µ2l
µ=1− ±i 1− 1− . (16.52)
2(ε2 + βk 2 µ2l ) 2(ε2 + βk 2 µ2l )
This implies the amplification factor
s 2  2
k 2 µ2l k 2 µ2l
Gl = |µ| = 1− + 1 − 1 −
2(ε2 + βk 2 µ2l ) 2(ε2 + βk 2 µ2l )
= 1, l = −M/2, · · · M/2 − 1.
This together with (16.44) imply that CN-LF-TSSP with 1/4 ≤ β ≤ 1/2 is
unconditionally stable. On the other hand, when 0 ≤ β < 1/4, we need the
stability condition
 
 k 2 µ2l  2ε 2hε
1 −  ≤ 1 =⇒ k ≤ min p = √ .
 2
2(ε + βk µl ) 2 2  −M/2≤l≤M/2−1 2
(1 − 4β)µl π 1 − 4β
This together with (16.44) imply that CN-LF-TSSP with 0 ≤ β < 1/4 is
conditionally stable in one dimension with stability condition
2hε
k≤ √ . (16.53)
π 1 − 4β
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

224 W. Bao

All above stability results are confirmed by our numerical experiments.

(5). ε-resolution in the ‘subsonic limit’ regime (0 < ε  1): As our
numerical results suggest: The meshing strategy (or ε-resolution) which
guarantees good numerical approximations of our new numerical methods
PSAS-TSSP and CN-LF-TSSP with 1/4 ≤ β ≤ 1/2 in the ‘subsonic limit’
regime, i.e. 0 < ε  1, is: i). for initial data with O(ε)-wavelength: h = O(ε)
and k = O(ε); ii). for initial data with O(1)-wavelength: h = O(1) and
k = O(1). Where the meshing strategy for CN-LF-TSSP with 0 ≤ β < 1/4
is: h = O(ε) & k = O(hε) = O(ε2 ); h = O(1) & k = O(ε), respectively.

Remark 16.2: If the periodic boundary conditions (16.4) and (16.5) are
replaced by the homogeneous Dirichlet boundary condition (16.9), then
the Fourier basis used in the above algorithm can be replaced by the sine
basis [15] or the algorithm in section 4 for VZSM. Similarly, if homogeneous
Neumann conditions are used, then cosine series can be applied in designing
the algorithm.

16.4. Extension TSSP to GVZS

The idea to construct the numerical methods CN-LF-TSSP and PSAS-
TSSP for GZS (16.1)-(16.5) can be easily extended to the VZSM [114] in
three dimensions for M different acoustic modes in a box Ω = [a1 , b1 ] ×
[a2 , b2 ] × [a3 , b3 ] with homogeneous Dirichlet boundary conditions:
M
X
i∂t E + a 4 E + (1 − a) ∇(∇ · E) − αE NJ + λ|E|2 E + iγE = 0, (16.54)
J=1
ε2J ∂tt NJ − 4NJ + νJ 4 |E|2 = 0, J = 1, · · · , M, x ∈ Ω, t > 0, (16.55)
(0) (0) (1)
E(x, 0) = E (x), NJ (x, 0) = NJ (x), ∂t NJ (x, 0) = NJ (x), x ∈ Ω, (16.56)
E(x, t) = 0, NJ (x, t) = 0 (J = 1, · · · , M), x ∈ ∂Ω; (16.57)

where x = (x, y, z)T and E(x, t) = (E1 (x, t), E2 (x, t), E3 (x, t))T . Moreover,
we supplement (16.54)-(16.57) by imposing the compatibility condition
(0) (1)
E(0) (x) = 0, NJ (x) = NJ (x) = 0, x ∈ ∂Ω, J = 1, · · · , M. (16.58)
In some cases, the homogeneous Dirichlet boundary condition (16.57)
may be replaced by periodic boundary conditions:
With periodic boundary conditions for E, NJ (J = 1, · · · , M) on ∂Ω.
(16.59)
−a1 b2 −a2 b3 −a3
We choose the spatial mesh sizes h1 = b1M 1
, h 2 = M2 and h 3 = M3
in x-, y- and z-direction respectively, with M1 , M2 and M3 given positive
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 225

integers; the time step k = 4t > 0. Denote grid points and time steps as

xj := a1 + jh1 , j = 0, 1, · · · , M1 ; yp := a2 + ph2 , p = 0, 1, · · · , M2 ;
zs := a3 + sh3 , s = 0, 1, · · · , M3 ; tm := mk, m = 0, 1, 2, · · · .

Let Em m
j,p,s and (NJ )j,p,s be the approximations of E(xj , yp , zs , tm ) and
NJ (xj , yp , zs , tm ), respectively.
For simplicity, here we only extend PSAS-TSSP from GZS (16.1)-(16.5)
to VZSM (16.54)-(16.57) with homogeneous Dirichlet conditions. For peri-
odic boundary conditions (16.59) or extension of CN-LF-TSSP can be done
in a similar way. Following the idea of constructing PSAS-TSSP for GZS
and the TSSP for VZSM in [114] here we only present the numerical algo-
rithm. From time t = tm to t = tm+1 , the PSAS-TSSP method for VZSM
(16.54)-(16.57) reads:
X „ « „ « „ «
m ljπ pgπ srπ
(NJ )m+1 = ]
(NJ )l,g,r (tm+1 ) sin sin sin ,
j,p,s M1 M2 M3
(l,g,r)∈N
X „ « „ « „ «
ljπ pgπ srπ
E∗j,p,s = g
Bl,g,r (k/2) (Em)
l,g,r sin sin sin ,
M1 M2 M3
(l,g,r)∈N
8 m+1
−ik[α M m ∗ 2
P
> J=1 ((NJ )j,p,s +(NJ )j,p,s )/2−λ|Ej,p,s | ] E∗
<e j,p,s , γ = 0,
E∗∗
j,p,s =
>
: PM m m+1 ∗ 2 −2γk
e−γk−i[kα J=1 ((NJ )j,p,s +(NJ )j,p,s )/2+λ|Ej,p,s | (e −1)/2γ] ∗
Ej,p,s , γ 6= 0,
X „ « „ « „ «
ljπ pgπ srπ
Em+1
j,p,s =
g
Bl,g,r (k/2) (E∗∗ )
l,g,r sin sin sin ,
M1 M2 M3
(l,g,r)∈N

where

N = {(l, g, r) | 1 ≤ l ≤ M1 − 1, 1 ≤ g ≤ M2 − 1, 1 ≤ r ≤ M3 − 1}.
8
>
> ^ (0) ^ (1)
>
> (NJ )0,0,0 + k (NJ )0,0,0 , Rl,g,r = 0, m = 0,
>
>
>
>
>
>
>
> ^ (0)
>
> (NJ )l,g,r cos(kRl,g,r /εJ ) Rl,g,r 6= 0, m = 0,
>
>
>
>
>
> ^ (1)
m < + Rεl,g,r
J
(NJ )l,g,r sin(kRl,g,r /εJ )
]
(N ) (t
J l,g,r m+1 ) = ˆ ˜ ^
> +νJ 1 − cos(kRl,g,r /εJ ) (|E (0) |2 )
>
> l,g,r ,
>
>
>
>
>
> m−1
>
>
> ]
2(N J )l,g,r (tm ) cos(kRl,g,r /εJ ) m ≥ 1,
>
> ˆ ˜ ^
>
> m 2
>
> +2νJ 1 − cos(kRl,g,r /εJ ) (|E | )l,g,r
>
>
: ] m−1
−(NJ ) (tm−1 ),
l,g,r
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

226 W. Bao

and

 I3 ,


l = g = r = 0,

Bl,g,r (τ ) = " 2
#

 −iaτ R2l,g,r e−i(1−a)τ Rl,g,r − 1

e I3 + Al,g,r , otherwise,
R2 l,g,r

with
   
κ2l κl ζg κl ηr κl

2
Rl,g,r = κ2l + ζg2 + ηr2 , Al,g,r =  κl ζg ζg2 ζg ηr  =  ζg  κl ζg ηr ;
κl ηr ζg ηr ηr2 ηr

e l,g,r , the sine-transform coeffi-

where I3 is the 3 × 3 identity matrix, and U
cients, are defined as
MX1 −1 M
X2 −1 M
X3 −1      
e l,g,r = 8 ljπ pgπ srπ
U Uj,p,s sin sin sin ,
M1 M2 M3 j=1 p=1 s=1 M1 M2 M3
(16.60)
with
πl πg
κl = , l = 1, . . . , M1 − 1, ζg = , g = 1, . . . , M2 − 1,
b 1 − a1 b 2 − a2
πr
ηr = , r = 1, . . . , M3 − 1.
b 3 − a3
The initial conditions (16.56) are discretized as

E0j,p,s = E(0) (xj , yp , zs ),

(0)
(NJ )0j,p,s = NJ (xj , yp , zs ), j = 0, · · · , M1 , p = 0, · · · , M2 , s = 0, · · · , M3 ,
(1)
(∂t NJ )0j,p,s = NJ (xj , yp , zs ), J = 1, · · · , M.

The properties of the numerical method for GZS in section 3 are still valid
here.

17. Crank-Nicolson finite difference (CNFD) method for

GZS
Another method for the GZS (13.40)-(13.41) is to use centered finite dif-
ference for spatial derivatives and Crank-Nicolson for time derivative. For
simplicity of notations , here we only present the CNFD method for the
standard ZS [28] with homogeneous Dirichlet boundary condition (16.9),
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 227

i.e., in (16.1)-(16.2) with ε = 1, ν = −1, α = 1, λ = 0 and γ = 0:

!
Ejm+1 − Ejm 1
m+1
Ej+1 − 2Ejm+1 + Ej−1
m+1 m
Ej+1 − 2Ejm + Ej−1
m
i + +
k 2 h2 h2

1 m
= (N + Njm+1 )(Ejm+1 + Ejm ), j = 1, 2, · · · , M − 1,
4 j

Njm+1 − 2Njm + Njm−1 m+1

Nj+1 − 2Njm+1 + Nj−1
m+1
−θ
k2 h2
m−1
Nj+1 − 2Njm−1 + Nj−1
m−1 m
Nj+1 − 2Njm + Nj−1
m
−θ − (1 − 2θ)
h2 h2
m 2
|Ej+1 | − 2|Ejm |2 + |Ej−1
m 2
|
= 2
,
h
E0m+1 m+1
= EM = 0, N0m+1 = N0m+1 = 0, m = 0, 1 · · · .

where 0 ≤ θ ≤ 1 is a parameter. The initial conditions are discretized as:

Ej0 = E 0 (xj ), Nj0 = N 0 (xj ), j = 0, 1, · · · , M, (17.1)

h
k 2 Nj+1 − 2Nj0 + Nj−1
0 0
Nj1 = Nj0 + kN 1 (xj ) +
2 h2
0 2 0 2 0 2i
|Ej+1 | − 2|Ej | + |Ej−1 |
+ . (17.2)
h2
When θ = 0, the discretization (17.1) for wave-type equation is ex-
plicit; when θ > 0, it is implicit but can be solved explicitly when periodic
boundary conditions are applied. Generalization of the method to GZS are
straightforward. In our computations in next subsection, we choose θ = 0.5.

Remark 17.1: In [62, 63], convergence and error estimate of the CNFD
discretization (17.1), (17.1) are proved.

18. Numerical results of GZS

In this section, we present numerical results of GZS with a solitary wave so-
lution in one dimension to compare the accuracy, stability and ε-resolution
of different methods. We also present numerical examples solitary-wave col-
lisions in one dimension GZS.
In our computation, the initial conditions for (16.3) are always chosen
such that |E 0 |, N 0 and N (1) decay to zero sufficiently fast as |x| → ∞. We
always compute on a domain, which is large enough such that the periodic
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

228 W. Bao

boundary conditions do not introduce a significant aliasing error relative to

the problem in the whole space.

Example 10 The standard ZS with a solitary-wave solution, i.e., we

choose d = 1, α = 1, λ = 0, γ = 0 and ν = −1 in (13.40)-(13.41). The well-
known solitary-wave solution (15.5)-(15.7) of the ZS in this case is given in
[97, 73]
p 2 2
E(x, t) = 2B 2 (1 − ε2 C 2 ) sech(B(x − Ct)) ei[(C/2)x−((C/2) −B )t] , (18.1)
N (x, t) = −2B 2 sech2 (B(x − Ct)), −∞ < x < ∞, t ≥ 0, (18.2)
where B, C are constants. The initial condition is taken as
E (0) (x) = E(x, 0), N (0) (x) = N (x, 0), N (1) (x, 0) = ∂t N (x, 0), (18.3)
where E(x, 0), N (x, 0) and ∂t N (x, 0) are obtained from (18.1), (18.2) by
setting t = 0.
We present computations for two different regimes of the acoustic speed,
i.e. 1/ε:
Case I. O(1)-acoustic speed, i.e. we choose ε = 1, B = 1, C = 0.5 in
(18.1), (18.2). Here we test the spatial and temporal discretization errors,
conservation of the conserved quantities as well as the stability constraint
of different numerical methods. We solve the problem on the interval [-32,
32], i.e., a = −32 and b = 32 with periodic boundary conditions. Let Eh,k
and Nh,k be the numerical solution of (16.1), (16.5) in one dimension with
the initial condition (18.3) by using a numerical method with mesh size h
and time step k. To quantify the numerical methods, we define the error
functions as
e1 = kE(·, t) − Eh,k (t)kl2 , e2 = kN (·, t) − Nh,k (t)kl2 ,
kE(·, t) − Eh,k (t)kl2 kN (·, t) − Nh,k (t)kl2
e= +
kE(·, t)kl2 kN (·, t)kl2
e1 e2
= +
kE(·, t)kl2 kN (·, t)kl2
and evaluate the conserved quantities DGZS , P GZS and H GZS by using the
numerical solution, i.e. replacing E and N by their numerical counterparts
Eh,k and Nh,k respectively, in (13.20)-(13.22).
First, we test the discretization error in space. In order to do this, we
choose a very small time step, e.g., k = 0.0001 such that the error from time
discretization is negligible comparing to the spatial discretization error, and
solve the ZS with different methods under different mesh sizes h. Tab. 6
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 229

lists the numerical errors of e1 and e2 at t = 2.0 with different mesh sizes
h for different numerical methods.

Mesh h = 1.0 h = 21 h = 14
e1 9.810E-2 1.500E-4 8.958E-9
PSAS-TSSP
e2 0.143 1.168E-3 6.500E-8
CN-LF-TSSP e1 9.810E-2 1.500E-4 7.409E-9
(β = 0) e2 0.143 1.168E-3 3.904E-8
CN-LF-TSSP e1 9.810E-2 1.500E-4 8.628E-9
(β = 1/4) e2 0.143 1.168E-3 6.521E-8
CN-LF-TSSP e1 9.810E-2 1.500E-4 1.098E-8
(β = 1/2) e2 0.143 1.168E-3 6.326E-8
e1 0.491 0.120 2.818E-2
CNFD
e2 0.889 0.209 4.726E-2
Tab. 6: Spatial discretization error analysis: e1 , e2 at time t=2 under k =
0.0001.

Secondly, we test the discretization error in time. Tab. 7 shows the

numerical errors of e1 and e2 at t = 2.0 under different time steps k and
mesh sizes h for different numerical methods.
Thirdly, we test the conservation of conserved quantities. Tab. 8 presents
the quantities and numerical errors at different times with mesh size h = 81
and time step k = 0.0001 for different numerical methods.
Case II: ‘Subsonic limit’ regime, i.e. 0 < ε  1, we choose B = 1
and C = 1/2ε in (18.1), (18.2). Here we test the ε-resolution of different
numerical methods. We solve the problem on the interval [-8, 120], i.e.,
a = −8 and b = 120 with periodic boundary conditions. Fig. 9 shows the
numerical results of PSAS-TSSP at t = 1 when we choose the meshing
strategy h = O(ε) and k = O(ε): T0 = (ε0 , h0 , k0 ) = (0.125, 0.5, 0.04),
T0 /4, T0 /16; and h = O(ε) and k = 0.04-independent of ε: T0 = (ε0 , h0 ) =
(0.125, 0.5), T0 /4, T0 /16. CN-LF-TSSP with β = 1/4 or β = 1/2 gives
similar numerical results at the same meshing strategies, where CN-LF-
TSSP with β = 0 gives correct numerical results at meshing strategy h =
O(ε) and k = O(ε2 ) and incorrect results at h = O(ε) and k = O(ε) [15].
Furthermore, our additional numerical experiments confirm that PSAS-
TSSP and CN-LF-TSSP with 1/4 ≤ β ≤ 1/2 are unconditionally stable and
CN-LF-TSSP with β = 0 is stable under the stability constraint (16.53).
From Tabs. 6-8 and Fig. 9, we can draw the following observations:
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

230 W. Bao

1 1 1 1
h Error k = 100 k = 400 k = 1600 k = 6400
1
PSAS-TSSP 4 e1 4.968E-5 3.109E-6 1.944E-7 1.226E-8
e2 1.225E-4 7.664E-6 4.797E-7 3.871E-8

1
8 e1 4.968E-5 3.109E-6 1.944E-7 1.172E-8
e2 1.225E-4 7.664E-6 4.797E-7 3.157E-8
1
CN-LF-TSSP 4 e1 4.829E-5 3.022E-6 1.888E-7 1.156E-8
(β = 0) e2 1.032E-4 6.456E-6 4.041E-7 3.673E-8

1
8 e1 4.829E-5 3.022E-6 1.888E-7 1.100E-8
e2 1.032E-4 6.456E-6 4.043E-7 2.946E-8
1
CN-LF-TSSP 4 e1 5.679E-5 3.556E-6 2.224E-7 1.425E-8
(β = 1/4) e2 1.623E-4 1.015E-5 6.351E-7 4.970E-8

1
8 e1 5.679E-5 3.556E-6 2.224E-7 1.377E-8
e2 1.623E-4 1.015E-5 6.351E-7 4.356E-8
1
CN-LF-TSSP 4 e1 7.468E-5 4.678E-6 2.924E-7 1.868E-8
(β = 1/2) e2 2.232E-4 1.396E-5 8.732E-7 6.360E-8

1
8 e1 7.468E-5 4.678E-6 2.924E-7 1.841E-8
e2 2.232E-4 1.396E-5 8.732E-7 5.942E-8
1
CNFD 4 e1 0.802 3.480E-2 2.855E-2 2.820E-2
e2 0.674 9.012-2 5.005E-2 4.743E-2

1
8 e1 0.809 1.753E-2 7.363E-3 6.961E-3
e2 0.656 5.491E-2 1.427E-2 1.167E-2
Tab. 7: Time discretization error analysis: e1 , e2 at time t=2.

In O(1)-acoustic speed regime, our new methods PSAS-TSSP and CN-

LF-TSSP with β = 1/2 or 1/4 give similar results as the old method,
i.e. CN-LF-TSSP with β = 0, proposed in [15]: they are of spectral order
accuracy in space discretization and second-order accuracy in time, conserve
DGZS exactly and P GZS , H GZS very well (up to 8 digits). However, they
are improved in two aspects: (i) They are unconditionally stable where
the old method is conditionally stable under the stability condition k ≤
√ 2hε in d-dimensions (d = 1, 2 or 3); (ii) In the ‘subsonic limit’ regime
π d(1−4β)
for initial data with O(ε)-wavelength, i.e. 0 < ε  1, the ε-resolution of
our new methods is improved to h = O(ε) and k = O(ε), where the old
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 231

Time DGZS P GZS H GZS

PSAS-TSSP 1.0 3.0000000000 3.41181556 0.510202736
2.0 3.0000000000 3.41181562 0.510202765
β=0 1.0 3.0000000000 3.41181557 0.510202736
2.0 3.0000000000 3.41181562 0.510202766
β = 1/4 1.0 3.0000000000 3.41181556 0.510202740
2.0 3.0000000000 3.41181564 0.510202779
β = 1/2 1.0 3.0000000000 3.41181556 0.510202737
2.0 3.0000000000 3.41181562 0.510202768
CNFD 1.0 3.0000000000 3.394829741 0.510115589
2.0 3.0000000000 3.394791238 0.510076710

Tab. 8: Conserved quantities analysis: k = 0.0001 and h = 81 .

method required h = O(ε) and k = O(εh) = O(ε2 ). Thus in the following,

we only present numerical results by PSAS-TSSP. In fact, CN-LF-TSSP
with 1/4 ≤ β ≤ 1 gives similar numerical results at the same mesh size and
time step for all the following numerical examples.

Example 11 Soliton-soliton collisions in one dimension GZS, i.e., we

choose d = 1, ε = 1, α = −2 and γ = 0 in (13.40)-(13.41). We use the
family of one-soliton solutions (15.5)-(15.7) in [72] to test our new numerical
method PSAS-TSSP.
The initial data is chosen as
E(x, 0) = Es (x + p, 0, η1 , V1 , ε, ν) + Es (x − p, 0, η2 , V2 , ε, ν),
N (x, 0) = Ns (x + p, 0, η1 , V1 , ε, ν) + Ns (x − p, 0, η2 , V2 , ε, ν),
∂t N (x, 0) = ∂t Ns (x + p, 0, η1 , V1 , ε, ν) + ∂t Ns (x − p, 0, η2 , V2 , ε, ν),
where x = ∓p are initial locations of the two solitons.
In all the numerical simulations reported in this example, we set λ = 2,
and Φ0 = 0. We only simulated the symmetric collisions, i.e., the collisions
of solitons with equal amplitudes η1 = η2 = η and opposite velocities
V1 = −V2 ≡ V . Here, we present computations for two cases:
I. Collision between solitons moving with the subsonic velocities, V <
1/ε = 1, i.e. we take ν = 0.2, η = 0.3 and V = 0.5;
II. Collision between solitons in the transonic regime, V > 1/ε = 1, i.e.
we take ν = 2.0, η = 0.3 and V = 3.0.
We solve the problem on the interval [-128,128], i.e., a = -128 and b =
128 with mesh size h = 41 and time step k = 0.005. We take p = 10. Fig.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

232 W. Bao

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
a). 0 2 x 4 6 8 10
b). 0 2 x 4 6 8 10

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2
0.2

0
0
c). 10 15 x 20 25
d). 8 10 12 14
x
16 18 20 22 24

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2
0.2

0
0
e). 5860 62 x 64 66 68 70
f). x
58 60 62 64 66 68 70

Fig. 9: Numerical solutions of the electric field |E(x, t)|2 at t = 1 for Exam-
ple 10 in the ‘subsonic limit’ regime by PSAS-TSSP. ‘—’: exact solution,
‘+ + +’: numerical solution. Left column corresponds to h = O(ε) and
k = O(ε): a). T0 = (ε0 , h0 , k0 ) = (0.125, 0.5, 0.04); c). T0 /4; e). T0 /16.
Right column corresponds to h = O(ε) and k = 0.04-independent ε: b).
T0 = (ε0 , h0 ) = (0.125, 0.5); d). T0 /4; f). T0 /16.

10 shows the evolution of the dispersive wave field |E|2 and the acoustic
(nondispersive) field N .
Case I corresponds to a soliton-soliton collision when the ratio ν/λ is
small, i.e., the GZS (16.1), (16.2) is close to the NLSE. As is seen, the
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 233

1.5 0.4

1
2
|E|

0.2

N
0.5
0
0 30
30 25
−10 −10
25 20
20 0
0 15
x 15 t 10 10
10 t
a). 10 x

2
0
N

1.5
|E|2

1
−5
15

0.5 10 20

5t 15
40
0 t 20
10
0 0
−20 −10 0 10 20
b). x 5 −20 x

2
Fig. 10: Evolution of the wave field |E| (left column) and acoustic field N
(right column) in Example 11. a). For case I; b). For case II.

collision seems quite elastic (cf. Fig. 10a). This also validates the formal
reduction from GZS to NLSE in section 2.5. Case II corresponds to the
collision of two transonic solitons. Note that the emission of the sound
waves is inconspicuous at this value of V (cf. Fig. 10b).
From Figs. 9&10, we can see that the unconditionally stable numerical
method PSAS-TSSP can really be applied to solve solitary-wave collisions
of GZS.

References
1. S. Abenda, Solitary waves for Maxwell-Dirac and Coulomb-Dirac models,
Ann. Inst. H. Poincare Phys. Theor., 68 (1998), pp. 229–244.
2. A. Aftalion and Q. Du, Vortices in a rotating Bose-Einstein condensate:
Critical angular velocities and energy diagrams in the Thomas-Fermi regime,
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

234 W. Bao

Phys. Rev. A, 64 (2001), pp. 063603.

3. M. H. Anderson, J. R. Ensher, M. R. Matthews, C. E. Wieman and E A.
Cornell, Science, 269 (1995), 198.
4. W. Bao, Ground states and dynamics of multi-component Bose-Einstein
condensates, Multiscale Modeling and Simulation, 2 (2004), pp. 210-236.
5. W. Bao, Numerical methods for nonlinear Schrödinger equation under
nonzero far-field conditions, Methods Appl. Anal., 11 (2004), pp. 367-388.
6. W. Bao and Q. Du, Computing the ground state solution of Bose-Einstein
condensates by a normalized gradient flow, SIAM J. Sci. Comp., 25 (2004),
pp. 1674-1697.
7. W. Bao and D. Jaksch, An explicit unconditionally stable numerical meth-
ods for solving damped nonlinear Schrodinger equations with a focusing
nonlinearity, SIAM J. Numer. Anal., 41(2003), pp. 1406-1426.
8. W. Bao, D. Jaksch and P.A. Markowich, Numerical solution of the Gross-
Pitaevskii equation for Bose-Einstein condensation, J. Comput. Phys., 187
(2003), pp. 318-342.
9. W. Bao, D. Jaksch and P.A. Markowich, Three dimensional simulation of
jet formation in collapsing condensates, J. Phys. B: At. Mol. Opt. Phys., 37
(2004), pp. 329-343.
10. W. Bao, S. Jin and P.A. Markowich, On time-splitting spectral approxima-
tions for the Schrödinger equation in the semiclassical regime, J. Comput.
Phys., 175 (2002), pp. 487-524.
11. W. Bao, S. Jin and P.A. Markowich, Numerical study of time-splitting spec-
tral discretizations of nonlinear Schrödinger equations in the semi-clasical
regimes, SIAM J. Sci. Comp., 25 (2003), pp. 27-64.
12. W. Bao and X.-G. Li, An efficient and stable numerical method for the
Maxwell-Dirac system, J. Comput. Phys., 199 (2004), pp. 663-687.
13. W. Bao and J. Shen, A fourth-order time-splitting Laguerre-Hermite pseudo-
spectral method for Bose-Einstein condensates, SIAM J. Sci. Comput., 26
(2005), pp. 2010-2028.
14. W. Bao, F. Sun, Efficient and stable numerical methods for the generalized
and vector Zakharov system, SIAM J. Sci. Comput., 26 (2005), pp. 1057-
1088.
15. W. Bao, F.F. Sun and G.W. Wei, Numerical methods for the generalized
Zakharov system, J. Comput. Phys. 190 (2003), pp. 201 - 228.
16. W. Bao and W. Tang, Ground state solution of trapped interacting Bose-
Einstein condensate by directly minimizing the energy functional, J. Com-
put. Phys., 187(2003), pp. 230-254.
17. W. Bao and Y. Zhang, Dynamics of the ground state and central vor-
tex states in Bose-Einstein condensation, Math. Mod. Meth. Appl. Sci., 66
(2006), pp. 758-786.
18. W. Bao, H.Q. Wang and P.A. Markowich, Ground, symmetric and central
vortex states in rotating Bose-Einstein condensates, Commun. Math. Sci.,
3 (2005), pp. 57-88.
19. G. Baym and C. J. Pethick, Phys. Rev. Lett., 76 (1996), 6.
20. P. Bechouche, N. J. Mauser and F. Poupaud, (Semi)-nonrelativistic limits
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 235

of the Dirac equation with external time-dependent electromagnetic field,

Commun. Math. Phys., 197 (1998), pp. 405-425.
21. J. Bolte and S. Keppeler, A semiclassical approach to the Dirac equation,
Ann. Phys., 274 (1999), pp. 125-162.
22. H. S. Booth, G. Legg and P. D. Jarvis, Algebraic solution for the vector
potential in the Dirac equation, J. Phys. A: Math. Gen., 34 (2001), pp.
5667-5677.
23. J. Bourgain, On the Cauchy and invariant measure problem for the periodic
Zakharov system, Duke Math. J. 76 (1994), pp. 175-202.
24. J. Bourgain and J. Colliander, On wellposedness of the Zakharov system,
Internat. Math. Res. Notices, 11 (1996), pp. 515-546.
25. N. Bournaveas, Local existence for the Maxwell-Dirac equations in three
space dimensions, Comm. Partial Differential Equations, 21 (1996), pp. 693–
720.
26. C.C. Bradley, C.A. Sackett, R.G. Hulet, Bose-Einstein condensation of
lithium: Observation of limited condensate number, Phys. Rev. Lett. 78
(1997), pp. 985-989.
27. B. M. Caradoc-Davis, R. J. Ballagh, K. Burnett, Coherent dynamics of
vortex formation in trapped Bose-Einstein condensates, Phys. Rev. Lett.,
83 (1999), 895.
28. Q. Chang and H. Jiang, A conservative difference scheme for the Zakharov
equations, J. Comput. Phys., 113 (1994), pp. 309-319.
29. J. M. Chadam, Global solutions of the Cauchy problem for the (classical)
coupled Maxwell-Dirac equations in one space dimension, J. Funct. Anal.,
13 (1973), pp. 173-184.
30. T. F. Chan, D. Lee and L. Shen, Stable explicit schemes for equations of
the Schrödinger type, SIAM J. Numer. Anal., 23 (1986), pp. 274-281.
31. T. F. Chan and L. Shen, Stability analysis of difference scheme for variable
coefficient Schrödinger type equations, SIAM J. Numer. Anal., 24 (1987),
pp. 336-349.
32. Q. Chang, B. Guo and H. Jiang, Finite difference method for generalized
Zakharov equations, Math. Comp., 64 (1995), pp. 537-553.
33. J. Colliander, Wellposedness for Zakharov systems with generalized nonlin-
earity, J. Diff. Eqs., 148 (1998), pp. 351-363.
34. F. Dalfovo S. Giorgini, L.P. Pitaevskii, and S. Stringari, Rev. Mod. Phys.,
71 (1999), 463.
35. A. Das, General solutions of Maxwell-Dirac equations in 1+1-dimensional
space-time and a spatially confined solution, J. Math. Phys, 34 (1993), pp.
3986-3985.
36. A. Das, An ongoing big bang model in the special relativistic Maxwell-Dirac
equations, J. Math. Phys., 37 (1996), pp. 2253-2259.
37. A. Das and D. Kay, A class of exact plane wave solutions of the Maxwell-
Dirac equations, J. Math. Phys., 30 (1989), pp. 2280-2284.
38. K.B. Davis, M.O. Mewes, M.R. Andrews, N.J. van Druten, D.S. Durfee,
D.M. Kurn, and W. Ketterle, Phys. Rev. Lett, 75 (1995), 3969.
39. A.S. Davydov, Phys. Scr., 20 (1979), pp. 387.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

236 W. Bao

40. L.M. Degtyarev, V.G. Nakhan’kov and L.I. Rudakov, Dynamics of the for-
mation and interaction of Langmuir solitons and strong turbulence, Sov.
Phys. JETP, 40 (1974), pp. 264-268.
41. B. Desjardins, C. K. Lin and T. C. Tso, Semiclassical limit of the derivative
nonlinear Schrödinger equation, M3 AS, 10 (2000), pp. 261-285.
42. P. A. M. Dirac, Principles of Quantum Mechanics, Oxford University Press,
London (1958).
43. M. Edwards, P.A. Ruprecht, K. Burnett, R.J. Dodd, and C.W. Clark,
Phys. Rev. Lett., 77 (1996), 1671; D.A.W. Hutchinson, E. Zaremba, and
A. Griffin, Phys. Rev. Lett., 78 (1997), 1842.
44. M. Edwards and K. Burnett, Numerical solution of the nonlinear
Schrödinger equation for small samples of trapped neutral atoms, Phys.
Rev. A, 51 (1995), 1382.
45. M. Esteban, E. Sere, An overview on linear and nonlinear Dirac equations,
Discrete Contin. Dyn. Syst., 8 (2002), pp. 381-397.
46. C. Fermanian-Kammerer, Semi-classical analysis of a Dirac equaiton with-
out adiabatic decoupling, Monatsh. Math., to appear.
47. A. L. Fetter and A. A. Svidzinsky, Vortices in a trapped dilute Bose-Einstein
condensate, J. Phys.: Condens. Matter, 13 (2001), pp. 135-194.
48. G. Fibich, An adiabatic law for self-focusing of optical beams, Opt. Lett.,
21 (1996), pp. 1735-1737.
49. G. Fibich, Small beam nonparaxiality arrests self-focusing of optical beams,
Phys. Rev. Lett., 76 (1996), 4356-4359.
50. G. Fibich, V. Malkin and G. Papanicolaou, Beam self-focusing in the pres-
ence of small normal time dispersion, Phys. Rev. A, 52 (1995), 4218-4228.
51. M. Flato, J. C. H. Simon and E. Taflin, Asymptotic completeness, global
existence and the infrared problem for the Maxwell-Dirac equations, Mem.
Amer. Math. Soc., 127 (1997), pp. 311.
52. E. Forest and R.D. Ruth, Fourth order symplectic integration, Phys. D, 43
(1990), pp. 105.
53. B. Fornberg and T. A. Driscoll, A fast spectral algorithm for nonlinear wave
equations with linear dispersion, J. Comput. Phys., 155 (1999), pp. 456.
54. J. J. Garcia-Ripoll, V. M. Perez-Garcia and V. Vekslerchik, Construc-
tion of exact solutions by spatial translations in inhomogeneous nonlinear
Schrödinger equaitons, Phys. Rev. E, 64 (2001), 056602.
55. I. Gasser and P.A. Markowich, Quantum hydrodynamics, Wigner transforms
and the classical limit, Asymptotic Analysis, 14 (1997), pp. 97-116.
56. P. Gérard, Microlocal defect measures, Comm. PDE., 16 (1991), pp. 1761-
1794.
57. V. Georgiev, Small amplitude solutions of the Maxwell-Dirac equations, In-
diana Univ. Math. J., 40 (1991), pp. 845-883.
58. P. Gérard, P. A. Markowich, N. J. Mauser and F. Poupaud, Homogenization
limits and Wigner transforms, Comm. Pure Appl. Math., 50 (1997), pp. 321-
377.
59. J. Ginibre, Y. Tsutsumi and G. Velo, The Cauchy problem for the Zakharov
system, J. Funt. Anal., 151 (1997), pp. 384-436.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 237

60. R. Glassey, On the asymptotic behavior of nonlinear Schrödinger equations,

Trans. Amer. Math. Soc., 182 (1973), pp. 187-200.
61. R. Glassey, On the blowing-up of solutions to the Cauchy problem for non-
linear Schrödinger equations, J. Math. Phys., 18 (1977), pp. 1794-1797.
62. R. Glassey, Approximate solutions to the Zakharov equations via finite dif-
ferences, J. Comput. Phys., 100 (1992), pp. 377-383.
63. R. Glassey, Convergence of an energy-preserving scheme for the Zakharov
equations in one space dimension, Math. Comp., 58 (1992), pp. 83-102.
64. O. Goubet and I. Moise, Attractor for dissipative Zakharov system, Nonlin-
ear Anal., Theory, Methods & Applications, 31 (1998), pp. 823-847.
65. D. Gottlieb and S.A. Orszag, Numerical analysis of spectral methods, Soc.
for Industr. & Appl. Math., Philadelphia, 1977.
66. E. Grenier, Semiclassical limit of the nonlinear Schrödinger equation in small
time, Proc. Amer. Math. Soc., 126 (1998), pp. 523-530.
67. M. Greiner, O. Mandel, T. Esslinger, T.W. Hänsch, and I. Bloch, Quantum
phase transition from a superfluid to a mott insulator in a gas of ultracold
atoms, Nature, 415 (2002), pp. 39-45.
68. A. Griffin, D. Snoke, S. Stringaro (Eds.), Bose-Einstein condensation, Cam-
bridge University Press, New York, 1995.
69. E.P. Gross, Nuovo. Cimento., 20 (1961), pp. 454.
70. L. Gross, The Cauchy problem for coupled Maxwell and Dirac equaitons,
Comm. Pure Appl. Math., 19 (1966), pp. 1-15.
71. D.S. Hall, M.R. Mattthews, J.R. Ensher, C.E. Wieman and E.A. Cornell,
Dynamics of component separation in a binary mixture of Bose-Einstein
condensates, Phys. Rev. Lett., 81 (1998), pp. 1539-1542.
72. H. Hadouaj, B. A. Malomed and G.A. Maugin, Soliton-soliton collisions in
a generalized Zakharov system, Phys. Rev. A, 44 (1991), pp. 3932-3940.
73. H. Hadouaj, B. A. Malomed and G.A. Maugin, Dynamics of a soliton in a
generalized Zakharov system with dissipation, Phys. Rev. A, 44 (1991), pp.
3925-3931.
74. R.H. Hardin and F.D. Tappert, Applications of the split-step Fourier method
to the numerical solution of nonlinear and variable coefficient wave equa-
tions, SIAM Rev. Chronicle, 15 (1973), pp. 423.
75. Z.Y. Huang, S. Jin, P.A. Markowich, C. Sparber and C.X. Zheng, A time-
splitting spectral scheme for the Maxwell-Dirac system, preprint.
76. W. Hunziker, On the nonrelativistic limit of the Dirac theory, Commun.
Math. Phys., 40 (1975), pp. 215-222.
77. D. Jaksch, C. Bruder, J. I. Cirac, C. W. Gardiner, and P. Zoller, Cold bosonic
atoms in optical lattices, Phys. Rev. Lett., 81 (1998), pp. 3108-3111.
78. S. Jin, P.A. Markowich and C.X. Zheng, Numerical simulation of a general-
ized Zakharov system, J. Comput. Phys., to appear.
79. A. Jüngel, Nonlinear problems in quantum semiconductor modeling, Nonlin.
Anal., Proceedings of WCNA2000, to appear.
80. A. Jüngel, Quasi-hydrodynamic semiconductor equations, Progress in Non-
linear Differential Equations and Its Applications, Birkhäuser, Basel, 2001.
81. T. Kato, On nonlinear Schrödinger equations, Ann. Inst. H. Poincaré. Phus.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

238 W. Bao

Théor, 46 (1987), pp. 113-129.

82. M. Landman, G. Papanicolaou, C. Sulem and P. Sulem, Rate of blowup for
solutions of the nonlinear Schrödinger equation at critical dimension, Phys.
Rev. A, 38 (1988), pp. 3837-3843.
83. M. Landman, G. Papanicolaou, C. Sulem, P. Sulem and X. Wang, Stability
of isotropic singularities for the nonlinear Schrödinger equation, Physica D,
47 (1991), pp. 393-415.
84. L. Laudau and E. Lifschitz, Quantum Mechanics: non-relativistic theory,
Pergamon Press, New York, 1977.
85. P. Leboeuf and N. Pavloff, Phys. Rev. A 64, 033602 (2001); V. Dunjko,
V. Lorent, and M. Olshanii, Phys. Rev. Lett., 86 (2001), pp. 5413.
86. A. J. Leggett, Bose-Einstein condensation in the alkali gases: some funda-
mental concepts, Rev. Mod. Phys., 73 (2001), pp. 307-356.
87. E. H. Lieb, R. Seiringer, J. Yugvason, Bosons in a trap: a rigorous derivation
of the Gross-Pitaevskii energy functional, Phys. Rev. A, 61 (2000), pp. 3602.
88. F. Lin and T. C. Lin, Vortices in two-dimensional Bose-Einstein condensates,
Geometry and nonlinear partial differential equations (Hangzhou, 2001), pp.
87–114, AMS/IP Stud. Adv. Math., 29, Amer. Math. Soc., Providence, RI,
2002.
89. M.J. Landman, G.C. Papanicolaou, C. Sulem, P.L. Sulem, X.P. Wang, Sta-
bility of isotropic singularities for the nonlinear Schrödinger equation, Phys.
D, 47 (1991), pp. 393–415.
90. A. G. Lisi, A solitary wave solution of the Maxwell-Dirac equations, J. Phys.
A: Math. Gen., 28 (1995), pp. 5385-5392.
91. P.A. Markowich, N.J. Mauser and F. Poupaud, A Wigner function approach
to semiclassical limits: electrons in a periodic potential, J. Math. Phys., 35
(1994), pp. 1066-1094.
92. P.A. Markowich, P. Pietra and C. Pohl, Numerical approximation of
quadratic observables of Schrödinger-type equations in the semi-classical
limit, Numer. Math., 81 (1999), pp. 595-630.
93. N. Masmoudi and N. J. Mauser, The selfconsistent Pauli equaiton, Monatsh.
Math., 132 (2001), pp. 19-24.
94. R. McLachlan, Symplectic integration of Hamiltonian wave equations, Nu-
mer. Math., 66 (1994), pp. 465.
95. V. Masselin, A result of the blow-up rate for the Zakharov system in dimen-
sion 3, SIAM J. Math. Anal., 33 (2001), pp. 440-447.
96. B. Najman, The nonrelativstic limit of the nonlinear Dirac equaiton, Ann.
Inst. Henri Poincare Non Lineaire, 9 (1992), pp. 3-12.
97. P.K. Newton, Wave interactions in the singular Zakharov system, J. Math.
Phys., 32 (1991), pp. 431-440.
98. G.C. Papanicolaou, C. Sulem, P.L. Sulem, X.P. Wang, Singular solutions of
the Zakharov equations for Langmuir turbulence, Phys. Fluids B, 3 (1991),
pp. 969-980.
99. A.S. Parkins and D.F. Walls, Physics Reports, 303 (1998), pp. 1.
100. G.L. Payne, D.R. Nicholson, and R.M. Downie, Numerical solution of the
Zakharov equations, J. Comput. Phys., 50 (1983), pp. 482-498.
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

Nonlinear Schrödinger Equations and Applications 239

101. N.R. Pereira, Collisions between Langmuir solitons, The Physics of Fluids,
20 (1977), pp. 750-755.
102. C.J. Pethick and H. Smith, Bose-Einstein Condensation in Dilute Gases,
Cambridge University Press, 2002.
103. L.P. Pitaevskii, Zh. Eksp. Teor. Fiz., 40 (1961), pp. 646. (Sov. Phys. JETP,
13 (1961), pp. 451).
104. L. Pitaevskii and S. Stringari, Bose-Einstein condensation, Oxford Univer-
sity Press, Oxford, 2002.
105. D. S. Rokhsar, Phys. Rev. Lett., 79 (1997), pp. 2164; R. Dum, J.I. Cirac,
M. Lewenstein, and P. Zoller, Phys. Rev. Lett., 80 (1998), pp. 2972;
P. O. Fedichev, and G. V. Shlyapnikov, Phys. Rev. A, 60 (1999), pp. 1779.
106. L. Simon, Asymptotics for a class of nonlinear evolution equations, with
applications to geometric problems, Annals of Math., 118 (1983), pp. 525-
571.
107. E.I. Schulman, Dokl. Akad. Nauk. SSSR, 259 (1981), pp.579 [Sov. Phys.
Dokl., 26 (1981), pp. 691].
108. C. Sparber and P. Markowich, Semiclassical asymptotics for the Maxwell-
Dirac system, J. Math. Phys., 44 (2003), pp. 4555–4572.
109. H. Spohn, Semiclassical limit of the Dirac equaiton and spin precession,
Ann. Phys., 282 (2000), pp. 420-431.
110. L. Stenflo, Phys. Scr., 33 (1986), pp. 156.
111. G. Strang, On the construction and comparison of differential schemes,
SIAM J. Numer. Anal., 5 (1968), pp. 506.
112. C. Sulem and P.L. Sulem, Regularity properties for the equations of Lang-
muir turbulence, C. R. Acad. Sci. Paris Sér. A Math., 289 (1979), pp. 173-
176.
113. C. Sulem and P.L. Sulem, The nonlinear Schrödinger equation, Springer,
1999.
114. F.F. Sun, Numerical studies on the Zakharov system, Master thesis, Na-
tional University of Singapore, 2003.
115. T.R. Taha and M.J. Ablowitz, Analytical and numerical aspects of certain
nonlinear evolution equations, II. Numerical, nonlinear Schrödinger equa-
tion, J. Comput. Phys., 55 (1984), pp. 203.
116. B. Thaller, The Dirac equation, New York, Springer, 1992.
117. M. Weinstein, Nonlinear Schrödinger equations and sharp interpolation es-
timates, Comm. Math. Phys., 87 (1983), pp. 567-576.
118. M. Weinstein, Modulational stability of ground states of nonlinear
Schrödinger equations, SIAM J. Math. Anal., 16 (1985), pp. 472-490.
119. M. Weinstein, The nonlinear Schrödinger equations-singularity formation,
stability and dispersion, Contemporary mathematics, 99 (1989), pp. 213-
232.
120. H. Yoshida, Construction of higher order symplectic integrators, Phys. Lett.
A., 150 (1990), pp. 262-268.
121. V.E. Zakharov, Zh. Eksp. Teor. Fiz., 62 (1972), pp. 1745 [Sov. Phys. JETP,
35 (1972), pp. 908].
April 19, 2007 Master Review Vol. 9in x 6in – (for Lecture Note Series, IMS, NUS) mrv-main

240 W. Bao