Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
:
Application to Parabolic Equation
Calculations
Matthias Ehrhardt
a,∗,1
and Ronald E. Mickens
b,2
a
Institut f¨ ur Mathematik, Technische Universit¨ at Berlin, Straße des 17. Juni 136,
D–10623 Berlin, Germany
b
Department of Physics, Clark Atlanta University, Atlanta, GA 30314, USA
Abstract
In the case of the equidistant discretization of the Airy diﬀerential equation (“dis
crete Airy equation”) the exact solution can be found explicitly. This fact is used
to derive a discrete transparent boundary condition (TBC) for a Schr¨ odinger–type
equation with linear varying potential, which can be used in “parabolic equation”
simulations in (underwater) acoustics and for radar propagation in the troposphere.
We propose diﬀerent strategies for the discrete TBC and show an eﬃcient imple
mentation. Finally a stability proof for the resulting scheme is given. A numerical
example in the application to underwater acoustics shows the superiority of the new
discrete TBC.
Key words: discrete Airy equation, discrete transparent boundary condition,
diﬀerence equation, Schr¨ odinger–type equation, parabolic equation prediction
PACS: 02.70.Bf, 43.30.+m, 92.10.Vz
∗
Corresponding author.
Email addresses: ehrhardt@math.tuberlin.de (Matthias Ehrhardt),
rohrs@math.gatech.edu (Ronald E. Mickens).
URL: http://www.math.tuberlin.de/~ehrhardt/ (Matthias Ehrhardt).
1
Supported by the DFG Research Center “Mathematics for key technologies” (FZT
86) in Berlin.
2
Research supported by DOE and the MBRS–SCORE Program.
Preprint submitted to Journal of Computational and Applied Mathematics23 February 2004
1 Introduction
In this work we consider the Airy diﬀerential equation
d
2
y
dx
2
−xy = 0. (1)
The solutions of this secondorder diﬀerential equation (1) are called Airy
functions and can be expressed in terms of Bessel functions of imaginary
argument of order ν = ±
1
3
. They play an important role in the theory of
asymptotic expansions of various special functions and have a wide range of
applications in mathematical physics.
The two linearly independent solutions of (1), the Airy functions of the ﬁrst
and second kind Ai(x), Bi(x), respectively have the following asymptotic rep
resentation for large [x[ [11]:
Ai(x) =
1
2
√
π
x
−1/4
e
−2/3x
3/2
[1 + O([x[
−3/2
)], (2a)
Bi(x) =
1
√
π
x
−1/4
e
2/3x
3/2
[1 + O([x[
−3/2
)]. (2b)
To solve the Airy equation (1) numerically we introduce the uniform grid
points x
m
= m∆x, y
m
· y(x
m
), and consider the standard discretization :
y
m+1
−2 y
m
+ y
m−1
(∆x)
2
−x
m
y
m
= 0, (3)
which can be rewritten as the secondorder diﬀerence equation
y
m+1
−2 y
m
+ y
m−1
−c my
m
= 0, c = (∆x)
3
. (4)
In [17] Mickens derived the following asymptotic behaviour of the two linearly
independent discrete solutions to (4) with c = 1:
y
(1)
m
=
_
m
7/2
e
m
m
m
_
_
1 −
85
12m
+ O
_
1
m
2
__
, (5a)
y
(2)
m
=
_
m
m
e
−m
m
9/2
_
_
1 +
133
12m
+ O
_
1
m
2
__
. (5b)
It can easily be seen that the solutions to this discretization (4) do not have
the same asymptotic properties as the solutions of (1) which motivated the
construction of a nonstandard discretization scheme (cf. [17], [18]).
This paper is organized as follows: ﬁrst we discuss a generalization of the the
discrete Airy equation (4) and show how to ﬁnd exact and asymptotic solu
tions. Afterwards we present an application to a problem arising in “parabolic
2
equation” calculations in (underwater) acoustics and radar propagation in
the troposphere: we construct a socalled discrete transparent boundary condi
tion (DTBC) for a Schr¨ odinger equation with a linear potential term, discuss
diﬀerent approaches and present an eﬃcient implementation by the sumof
exponentials ansatz. Afterwards we analyze the stability of the resulting nu
merical scheme. In this case the Laplace transformed Schr¨ odinger equation
can be viewed as a general Airy equation. Discrete transparent boundary con
ditions for a Schr¨ odinger equation with constant potential were derived in [1]
and a generalization to a problem arising in (underwater) acoustics was pre
sented in [2]. The construction of DTBC to the nonstandard discretization
scheme and to wideangle parabolic equations will be a topic of future work.
Finally we illustrate the results with a numerical example from underwater
acoustics.
2 Exact and asymptotic solutions to the discrete Airy equation
In this section we consider the discrete Airy equation in the more general form
y
m+1
−2 y
m
+ y
m−1
−(d + c m) y
m
= 0, c, d ∈ C. (6)
and show how to determine asymptotic solutions. This is a nontrivial task since
equation (6) is not of Poincar´e type. For such an equation, the coeﬃcients
must approach constant values as m → ∞ and it is clearly seen that the
coeﬃcient of y
m
in equation (6) does not satisfy this condition. Afterwards we
will demonstrate how to obtain an explicit solution to (6).
The asymptotic solution. Since the classic theorems of Poincar´e and Perron
cannot be applied to (6) it is not straight forward to obtain information about
the asymptotic behaviour of the solutions to this equation. One possible ansatz
is the one of Batchelder [5] given in [17] (see (5) in the case c = 1, d = 0).
Here we want to apply the approach of Wong and Li [23] to obtain asymptotic
solutions to the second–order diﬀerence equation
y
m+2
+ m
p
a(m) y
m+1
+ m
q
b(m) y
m
= 0, (7)
where p, q are integers and a(m) and b(m) have power expansions of the form
a(m) =
∞
s=0
a
s
m
s
, b(m) =
∞
s=0
a
s
m
s
, (8)
with nonzero leading coeﬃcients: a
0
,= 0, b
0
,= 0. We increase the index of (6)
3
by one to put it in the form of (7) and make the following identiﬁcations:
a
0
= −c, a
1
= −(2 + c + d), a
s
= 0, s ≥ 2,
b
0
= 1, b
s
= 0, s ≥ 1, p = 1, q = 0.
Then the two formal series solutions (cf. [23]) are given by
y
(1)
m
=
c
−m
(m−2)!
m
−2−(2+d)/c
∞
s=0
c
(1)
s
m
s
, (9a)
y
(2)
m
= (m−2)! c
m
m
1+(2+d)/c
∞
s=0
c
(2)
s
m
s
. (9b)
To determine the values of the coeﬃcients c
(1)
1
, c
(1)
2
, c
(1)
3
, . . . we substitute the
decaying solution y
(1)
m
in (7):
1
cm
_
m + 1
m + 2
_
θ
∞
s=0
c
(1)
s
(m + 2)
s
+ c(m−1)
_
m + 1
m
_
θ
∞
s=0
c
(1)
s
m
s
=
_
c(θ −1) + cm
_
∞
s=0
c
(1)
s
(m + 1)
s
, (10)
with θ = 2 + (2 + d)/c. We now obtain after an Taylor expansion in 1/m
and setting all the linearly independent terms equal to zero, by a lengthy but
elementary calculation, the results:
c
(1)
1
= −c
(1)
0
_
c
−2
−θ +
θ
2
(θ −1)
_
, (11a)
c
(1)
2
=
c
(1)
0
2
_
θc
−2
+
θ
2
(θ −1) −
θ
6
(θ −1)(θ −2)
_
−
c
(1)
1
2
_
c
−2
−2 +
θ
2
(θ −1)
_
,
(11b)
c
(1)
3
= −
c
(1)
0
3
_
_
2θ +
θ
2
(θ −1)
_
c
−2
−
θ
6
(θ −1)(θ −2) −
θ
24
(θ −1)(θ −2)(θ −3)
_
+
c
(1)
1
3
_
(2 + θ)c
−2
−2 + θ +
θ
2
(θ −1) −
θ
6
(θ −1)(θ −2)
_
−
c
(1)
2
3
_
c
−2
−5 + θ +
θ
2
(θ −1)
_
, (11c)
etc..
4
Similarly we obtain for the increasing solution y
(2)
m
the ﬁrst three coeﬃcients
c
(2)
1
= c
(2)
0
_
c
−2
−η +
η
2
(η −1)
_
, (12a)
c
(2)
2
= −
c
(2)
0
2
_
(η −1)c
−2
−η + η(η −1) −
η
6
(η −1)(η −2)
_
(12b)
+
c
(2)
1
2
_
c
−2
+ 3 −2η −
η
2
(η −1)
_
,
c
(2)
3
=
c
(2)
0
3
_
_
1 +
η
2
(η −1)
_
c
−2
−η +
3
2
η(η −1) −
η
2
(η −1)(η −2)
+
η
24
(η −1)(η −2)(η −3)
_
−
c
(2)
1
3
_
(η −1)c
−2
−7 −6η + 2η(η −1) −
η
6
(η −1)(η −2)
_
(12c)
+
c
(2)
2
3
_
c
−2
+ 9 −3η +
η
2
(η −1)
_
,
with η = 1 + (2 + d)/c. Here c
(1)
0
, c
(2)
0
denote arbitrary constants.
Remark 1 We remark that in the special case c = 1, d = 0 we obtain
y
(1)
m
= c
(1)
0
m
−4
(m−2)!
_
1 −
3
m
+
21
2m
2
−
104
3m
3
+ O
_
m
−4
_
_
, (13a)
y
(2)
m
= c
(2)
0
(m−2)! m
3
_
1 +
1
m
−
3
2m
2
−
1
3m
3
+ O
_
m
−4
_
_
. (13b)
The explicit solution. In most cases, secondorder linear diﬀerence equations
with variable coeﬃcients cannot be solved in closed form. In this section we
show that in the special case of the discrete Airy equation (6), it is possible
to obtain an explicit solution. We present the derivation of this exact solution
and study its asymptotic behaviour.
If one wants to solve a diﬀerence equation with polynomial coeﬃcients (like
(6)), one approach is to ﬁnd the solution by the “method of generating func
tions” (cf. [10]); i.e., a generating function for a solution of (6) can be shown
to satisfy a diﬀerential equation, which may be solvable in terms of known
functions. To start with, deﬁne the generating function to be
g(ξ) =
∞
m=−∞
y
m
ξ
m
. (14)
We multiply (6) with ξ
m−1
and sum it up for m ∈ Z:
∞
m=−∞
y
m
ξ
m−2
−(2 + d)
∞
m=−∞
y
m
ξ
m−1
+
∞
m=−∞
y
m
ξ
m
−c
∞
m=−∞
my
m
ξ
m−1
= 0.
5
This results in the following ordinary diﬀerential equation for g:
g
(ξ) −
1 −(2 + d)ξ + ξ
2
cξ
2
g(ξ) = 0,
for which the solution is
g(ξ) = ξ
−
2+d
c
e
(ξ−
1
ξ
)/c
= ξ
−
2+d
c
∞
ν=−∞
J
ν
(
2
c
)ξ
ν
.
Hence, the exact decaying solution of (6) is the Bessel function J
ν
(
2
c
) (regarded
as function of its order ν), i.e. the discrete Airy equation is nothing else but
the recurrence relation for J
ν
(
2
c
). We note that it is wellknown ([11], [22])
that the recurrence equation for the Bessel functions
J
ν+1
(z) −2
ν
z
J
ν
(z) + J
ν−1
(z) = 0, (15)
still holds for complex orders ν and complex arguments z.
Thus the decaying solution to (6) can be represented as (cf. [22, Chapter 3.1]):
y
m
= J
m+
2+d
c
(
2
c
) =
1
c
m+
2+d
c
∞
n=0
(−1)
n
c
2n
n! Γ(m +
2+d
c
+ n + 1)
, c, d ∈ C. (16)
We also observe that (14) is not a generating function in the strict sense but
a Laurent series, which is uniformly convergent, i.e. diﬀerentiating each term
is permissible (cf. [22]). Note that this generating function approach is not
suitable for determining the growing solution of (6) for m →∞. This solution
is the socalled “NeumannFunction” (or Bessel function of the second kind)
which is known to also satisfy the recursion equation of the Bessel functions.
Remark 2 A diﬀerence equation more general than (6) was examined by
Barnes [4] in 1904. He also considered (6) and found (through a diﬀerent
construction) the solution (16).
The next step is to use (16) to examine the asymptotic behaviour of the dis
crete solutions. One can derive the following dominating series for the decaying
solution:
[y
m
[ ≤ Y
m
=
1
[c[
m+[
2+d
c
[
1
¸
¸
¸Γ(m +
2+d
c
+ 1)
¸
¸
¸
e
1
[
c
2
(m+
2+d
c
+1)
[
, (17)
which can be estimated using Stirling’s inequality
Y
m
<
1
√
2π
e
1
c
2
[c[
m+
2+d
c
_
e
m
_
m
m
−
1
2
Γ(m + 1)
Γ(m +
2+d
c
+ 1)
, (18)
6
if (2+d)/c is a positive real number. We want to compare this result concerning
the decay rate with the one derived by Mickens [17]. From setting c = 1, d = 0
in (18), it follows that
Y
m
<
e
√
2π
m
−5/2
e
m
m
m
, (19)
and we see that the decay rate of (18) and (5) diﬀers by a factor m
−6
, i.e. our
new bound Y
m
decays faster.
Another approach is to use an approximation for the exact solution of (6).
We use the following asymptotic representation of Bessel functions for large
values of the order ν (cf. [11]):
J
ν
(z) ≈
1
√
2π
e
ν+ν log(z/2)−(ν+1/2) log ν
, [ν[ →∞, [ arg ν[ ≤ π −δ. (20)
This enables us to give another asymptotic form of the recessive solution to
the discrete model equation (6) with c = 1, d = 0:
y
m
= J
m+2
(2) ≈
e
2
√
2π
e
m
(m + 2)
−(m+5/2)
, m →∞, (21)
and with Stirling’s formula this simpliﬁes to y
m
≈ 1/(m + 2)!, m →∞.
3 Application to Parabolic Equation Calculations
With the results from the previous Section we want to derive a discrete trans
parent boundary condition (DTBC) for the so–called standard “parabolic equa
tion” (SPE) [20], i.e. a one–way wave equation, arising for example in (under
water) acoustics and radiowave propagation problems. Here we concentrate
on the application to underwater acoustics.
The standard parabolic equation in underwater acoustics. A standard
task in oceanography is to calculate the acoustic pressure p(z, r) emerging
from a time–harmonic point source located in the water at (z
s
, 0). Here, r > 0
denotes the radial range variable and 0 < z < z
b
the depth variable (assuming
a cylindrical geometry). The water surface is at z = 0, and the (horizontal)
sea bottom at z = z
b
. We denote the local sound speed by c(z, r), the density
by ρ(z, r), and the attenuation by α(z, r) ≥ 0. The complex refractive index
is given by N(z, r) = c
0
/c(z, r) + iα(z, r)/k
0
with a reference sound speed c
0
and the reference wave number k
0
= 2πf/c
0
, where f denotes the frequency
of the emitted sound.
The SPE in cylindrical coordinates (z, r) reads:
2ik
0
ψ
r
(z, r) + ρ ∂
z
(ρ
−1
∂
z
)ψ(z, r) + k
2
0
(N
2
(z, r) −1) ψ(z, r) = 0, (22)
7
where ψ denotes the (complex valued) outgoing acoustic ﬁeld
ψ(z, r) =
_
k
0
r p(z, r) e
−ik
0
r
, (23)
in the far ﬁeld approximation (k
0
r ¸1). This Schr¨ odinger equation (22) is an
evolution equation in r and a reasonable description of waves with a propaga
tion direction within about 15
◦
of the horizontal.
Here, the physical problem is posed on the unbounded z–interval (0, ∞) and
one wishes to restrict the computational domain in the z–direction by intro
ducing an artiﬁcial boundary at the water–bottom interface (z = z
b
), where
the wave propagation in water has to be coupled to the wave propagation in
the the bottom. At the water surface one usually employs a Dirichlet (“pres
sure release”) BC: ψ(0, r) = 0.
Since the density is typically discontinuous at the water–bottom interface
(z = z
b
), one requires continuity of the pressure and the normal particle
velocity:
ψ(z
b
−, r) = ψ(z
b
+, r), (24a)
ψ
z
(z
b
−, r)
ρ
w
=
ψ
z
(z
b
+, r)
ρ
b
, (24b)
where ρ
w
= ρ(z
b
−, r) is the water density just above the bottom and ρ
b
denotes
the constant density of the bottom.
In this work we are especially interested in the case of a linear squared re
fractive index in the bottom region. For most underwater acoustics (and also
radiowave propagation) problems the squared refractive index in the exterior
domain increases with z. However, the usual transparent boundary condition
(see e.g. [2]) was derived for a homogeneous medium (i.e. all physical parame
ters are constant for z > z
b
). This TBC is not matched to the behaviour of the
refractive index and spurious reﬂections will occur. Instead we will derive a
TBC that matches the squared refractive index gradient at z = z
b
. We denote
the physical parameters in the bottom with the subscript b and assume that
the squared refractive index N
b
below z = z
b
can be written as
N
2
b
(z, r) = 1 + β + µ(z −z
b
), z > z
b
, (25)
with real parameters β and µ ,= 0, i.e. no attenuation in the bottom: α
b
= 0.
All other physical parameters are assumed to be constant in the bottom. Here,
the slope µ > 0 corresponds to a downwardrefracting bottom (energy loss)
and µ < 0 represents the upwardrefracting case, i.e. energy is returned from
the bottom.
One can easily derive an estimate for the L
2
–decay of solutions to the SPE,
posed on the halfspace z > 0. We assume ρ = ρ(z) and a simple calculation
8
(cf. [2]) gives
∂
r
ψ(., r)
2
= −2c
0
_
∞
0
α
c
[ψ[
2
ρ
−1
dz, (26)
for the weighted L
2
–norm (“acoustic energy”)
ψ(., r)
2
=
_
∞
0
[ψ(z, r)[
2
ρ
−1
(z) dz, (27)
i.e. in the dissipation–free case (α ≡ 0) ψ(., r) is conserved and for α > 0 it
decays.
4 Transparent Boundary Conditions
Transparent Boundary Conditions. In the following we will review from
[14] the derivation of the transparent boundary condition at z = z
b
for the
SPE with a linear squared refractive index. A transparent BC for the SPE (or
Schr¨ odinger equation) for a constant exterior medium was derived by several
authors from various application ﬁelds, e.g. in [1].
If we further assume that the initial data ψ
I
= ψ(z, 0), which models a point
source located at (z
s
, 0), is supported in the computational domain 0 < z < z
b
,
then a Laplace transformation in range of (22) for z > z
b
yields:
ˆ
ψ
zz
(z, s) + [µk
2
0
(z − ˜ z
b
) + 2ik
0
s]
ˆ
ψ(z, s) = 0, z > z
b
, (28)
with ˜ z
b
= z
b
−β/µ. Now the basic idea of the derivation is to explicitly solve the
equation in the exterior domain z > z
b
. Setting σ
3
= −µk
2
0
and τ = 2ik
0
/σ
2
(28) can be written as
ˆ
ψ
zz
(z, s) + σ
2
_
σ(z − ˜ z
b
) + τs
_
ˆ
ψ(z, s) = 0, z > z
b
. (29)
Introducing the change of variables ζ
s
(z) = σ(z −˜ z
b
) +τs, U(ζ
s
(z)) =
ˆ
ψ(z, s),
we can write (29) in the form of an Airy equation:
U
(ζ
s
(z)) + ζ
s
(z) U(ζ
s
(z)) = 0, z > z
b
. (30)
The solution of (30) which decays for z →∞, for ﬁxed s, Re s > 0 is
ˆ
ψ(z, s) = C
1
(s) Ai(ζ
s
(z)), z > z
b
, (31)
if we deﬁne the physically relevant branch of σ to be
σ =
_
_
_
(µk
2
0
)
1/3
e
−iπ/3
, µ > 0,
(−µk
2
0
)
1/3
, µ < 0.
(32)
9
Elimination of C
1
(s) gives
ˆ
ψ(z, s) =
ˆ
ψ(z
b
+, s)
Ai(ζ
s
(z))
Ai(ζ
s
(z
b
))
, z > z
b
. (33)
Finally, diﬀerentiation w.r.t. z yields with the matching conditions (24) the
transformed transparent BC at z = z
b
:
ˆ
ψ
z
(z
b
−, s) =
ρ
w
ρ
b
s
ˆ
ψ(z
b
−, s) W(s), W(s) = σ
Ai
(ζ
s
(z
b
))
s Ai(ζ
s
(z
b
))
, (34)
i.e. the transparent BC at z = z
b
reads:
ψ
z
(z
b
, r) =
ρ
w
ρ
b
_
r
0
ψ
r
(z
b
, r
) g
µ
(r −r
) dr
. (35)
The kernel g
µ
is obtained by an inverse Laplace transformation of W(s)
(cf. [7]):
g
µ
(r) = σ
_
_
_
Ai
(ζ
0
(z
b
))
Ai(ζ
0
(z
b
))
+
∞
j=1
e
−(a
j
−ζ
0
(z
b
))r/τ
a
j
−ζ
0
(z
b
)
_
_
_
, (36)
where ζ
0
(z
b
) = σβ/µ and the (a
j
) are the zeros of the Airy function Ai which
are all located on the negative real axis [19]. This BC is nonlocal in the range
variable r and can easily be discretized, e.g. in conjunction with a ﬁnite diﬀer
ence scheme for (22). The constant term in g
µ
acts like a Dirac function and
the inﬁnite series represents the continuous part. As Levy noted in [15] the
kernel g
µ
decays extremely fast for µ > 0 and for negative µ it decays slowly
at short ranges and then oscillates.
Numerical Implementation. Now we shall discuss how to solve (22) nu
merically with a CrankNicolson ﬁnite diﬀerence scheme which is of second
order (in ∆z and ∆r) and unconditionally stable. We will use the uniform
grid z
j
= jh, r
n
= nk with h = ∆z, k = ∆r and the approximations
ψ
(n)
j
≈ ψ(z
j
, r
n
), ρ
j
≈ ρ(z
j
). The discretized SPE (22) then reads:
−iR(ψ
(n+1)
j
−ψ
(n)
j
) = ρ
j
∆
0
z
(ρ
−1
j
∆
0
z
)(ψ
(n+1)
j
+ψ
(n)
j
)+w
_
(N
2
)
(n)
j
−1
_
(ψ
(n+1)
j
+ψ
(n)
j
),
(37)
with ∆
0
z
ψ
(n)
j
= ψ
(n)
j+1/2
−ψ
(n)
j−1/2
, the mesh ratio R = 4k
0
h
2
/k and w = k
2
0
h
2
.
Discretization of the continuous TBC. To incorporate the TBC (35)
in a ﬁnite diﬀerence scheme we make the approximation that ψ
r
(z
b
, r
) is
constant on each subinterval r
n
< r
< r
n+1
and integrate the kernel g
µ
exactly. In the following we review the discretization from [15] and start with
the discretization in range:
ψ
z
(z
b
, r
n
) =
ρ
w
ρ
b
n−1
m=0
ψ
(n−m)
b
−ψ
(n−m−1)
b
k
G
m
, (38)
10
where we have set ψ
(n)
b
= ψ(z
b
, r
n
) and G
m
is given by
G
m
=
_
r
m+1
rm
g
µ
(η) dη = kσ
Ai
(ζ
0
(z
b
))
Ai(ζ
0
(z
b
))
+
2ik
0
σ
∞
j=1
e
−(a
j
−ζ
0
(z
b
))r/τ
(a
j
−ζ
0
(z
b
))
2
¸
¸
¸
¸
r=r
m+1
r=rm
. (39)
This leads after rearranging to
k
ρ
b
ρ
w
ψ
z
(z
b
, r
n
) = −ψ
(0)
b
G
n
+ ψ
(n)
b
G
0
+
n−1
m=1
ψ
(n−m)
b
(G
m
−G
m−1
) (40)
In [15] Levy used an oﬀset grid in depth, i.e. ˜ z
j
= (j +
1
2
)h,
˜
ψ
(n)
j
≈ ψ(˜ z
j
, r
n
),
j = −1, . . . , J, where the water–bottom interface lies between the grid points
j = J −1 and J:
ψ
(m)
b
= ψ(z
b
, r
m
) ≈
˜
ψ
(m)
J
+
˜
ψ
(m)
J−1
2
, ψ
z
(z
b
, r
n
) ≈
˜
ψ
(n)
J
−
˜
ψ
(n)
J−1
h
. (41)
This yields ﬁnally (recall that
˜
ψ
(0)
J
=
˜
ψ
(0)
J−1
= 0) the following discretized TBC
for the SPE:
(1 −b
0
)
˜
ψ
(n)
J
−(1 + b
0
)
˜
ψ
(n)
J−1
=
n−1
m=1
b
m
(
˜
ψ
(n−m)
J
+
˜
ψ
(n−m)
J−1
), (42)
with
b
0
=
1
2
h
k
ρ
w
ρ
b
G
0
, b
m
=
1
2
h
k
ρ
w
ρ
b
(G
m
−G
m−1
). (43)
Note that the constant term in (39) enters only b
0
. Since a
j
∼ −
_
3π
8
(4j−1)
_
2/3
for j → ∞ the series (39) deﬁning G
m
has good convergence properties for
positive range r but for r = 0 the convergence is very slow. To overcome this
numerical problem we use the identity
∞
j=1
1
(a
j
−ζ
0
(z
b
))
2
=
_
Ai
(ζ
0
(z
b
))
Ai(ζ
0
(z
b
))
_
2
−ζ
0
(z
b
), (44)
which can be derived analogously to the one in [14].
In a numerical implementation one has to limit the summation in (36) and
therefore the TBC is not fully transparent any more. Moreover, the stability of
the resulting scheme is not clear since the discretized TBC (42) is not matched
to the ﬁnite diﬀerence scheme (37) in the interior domain. Instead of using
an ad–hoc discretization of the analytic transparent BC like in [15] we will
construct discrete TBCs of the fully discretized half–space problem with the
help of the results from Section 2.
Discrete transparent boundary conditions. To derive the discrete TBC
we will now mimic the derivation of the analytic TBC from Section 2 on a
11
discrete level. In analogy to the continuous problem we assume for the initial
data ψ
0
j
= 0, j ≥ J − 1 and use the linear potential term (N
2
)
(n)
j
− 1 =
β + µh(j −J), z
b
= Jh and solve the discrete exterior problem
−iR(ψ
(n+1)
j
−ψ
(n)
j
) = ∆
2
ψ
(n+1)
j
+∆
2
ψ
(n)
j
+w
_
β+µh(j−J)
_
(ψ
(n+1)
j
+ψ
(n)
j
), (45)
j ≥ J −1, ∆
2
ψ
(n)
j
= ψ
(n)
j+1
−2 ψ
(n)
j
+ ψ
(n)
j−1
, by using the Z–transformation:
Z¦ψ
(n)
j
¦ =
ˆ
ψ
j
(z) :=
∞
n=0
ψ
(n)
j
z
−n
, z ∈ C, [z[ > 1. (46)
Hence, the Z–transformed ﬁnite diﬀerence scheme (37), for j ≥ J, is a discrete
Airy equation
ˆ
ψ
j+1
(z) −2
_
1 −iζ(z) −µ
k
2
0
2
h
3
(j −J)
_
ˆ
ψ
j
(z) +
ˆ
ψ
j−1
(z) = 0, j ≥ J −1, (47)
with
ζ(z) =
R
2
z −1
z + 1
−i
β
2
k
2
0
h
2
. (48)
Comparing (47) with the recurrence relation of the Bessel function J
ν
(σ) yields
the condition
ν
σ
= 1 −iζ(z) −µ
k
2
0
2
h
3
(j −J)
!
=
j + oﬀset
σ
, (49)
and we conclude that the exact solution of (47) is
ˆ
ψ
j
(z) = J
ν
j
(z)
(σ), (50)
with
ν = ν
j
(z) = σ(1 −iζ(z)) + j −J, σ = −
_
µ
k
2
0
2
h
3
_
−1
∈ R. (51)
From (50) we obtain the transformed discrete TBC at z
b
= Jh:
ˆ
ψ
J−1
(z) = ˆ g
µ,J
(z)
ˆ
ψ
J
(z) with ˆ g
µ,J
(z) =
J
ν
J−1
(z)
(σ)
J
ν
J
(z)
(σ)
=
J
σ(1−iζ(z))−1
(σ)
J
σ(1−iζ(z))
(σ)
.
(52)
Finally, an inverse Z–transformation yields the discrete TBC
ψ
(n)
J−1
−
(0)
J
ψ
(n)
J
=
n−1
m=1
ψ
(n−m)
J
(m)
J
, (53)
with
(n)
J
= Z
−1
¦ˆ g
µ,J
(z)¦ =
τ
n
2π
2π
_
0
ˆ g
µ,J
(τe
iϕ
)e
inϕ
dϕ, n ∈ Z
0
, τ > 0. (54)
12
Since this inverse Ztransformation cannot be done explicitly, we use a nu
merical inversion technique based on FFT (cf. [9]); for details of this routine
we refer the reader to [8]. Note that the Bessel functions in (52) with complex
order and (possibly large) real argument can be evaluated by special software
packages (see e.g. [21]).
Since the magnitude of
(n)
J
does not decay as n → ∞ (Im
(n)
J
behaves like
const (−1)
n
for large n), it is more convenient to use a modiﬁed formulation
of the DTBCs (cf. [9]). We introduce the summed coeﬃcients
s
(n)
J
= Z
−1
¦ˆ s
J
(z)¦ , with ˆ s
J
(z) :=
z + 1
z
ˆ
J
(z), (55)
which satisfy
s
(0)
J
=
(0)
J
, s
(n)
J
=
(n)
J
+
(n−1)
J
, n ≥ 1. (56)
In physical space, the DTBC is:
ψ
(n)
J−1
−s
(0)
J
ψ
(n)
J
=
n−1
m=1
s
(n−m)
J
ψ
(m)
J
−ψ
(n−1)
J−1
, n ≥ 1. (57)
In the following, we will present two alternative approaches to obtain a DTBC:
asymptotic expansions and a continued fraction formula.
Asymptotic Expansions. For a second approach to obtain an approximated
DTBC, one can use asymptotic expansions. Using the asymptotic formula (20)
leads to the approximate DTBC
ˆ
ψ
J−1
(z) =
ˆ
h
µ,J
(z)
ˆ
ψ
J
(z) with
ˆ
h
µ,J
(z) =
2
e
σ
−1
_
ν
J
(ν
J
−1)
_
ν
J
ν
J
−1
_
ν
J
,
(58)
with ν
J
, σ given by (51).
Alternatively, one can use the asymptotic solution y
(1)
m
from (9a). The Z–
transformed scheme in the exterior j ≥ J −1, given by (47), is a discrete Airy
equation of the form (6) with
c = 2σ
−1
, d = −2iζ(z) −cJ, i.e. θ = 2 +
2 + d
c
= 2 + ν
0
,
and thus we obtain an approximation to the transformed DTBC of the form
ˆ
ψ
J−1
(z) =
ˆ
k
µ,J
(z)
ˆ
ψ
J
(z), (59a)
13
with
ˆ
k
µ,J
(z) =
y
(1)
J−1
y
(1)
J
= c(J −2)
_
J
J −1
_
θ
∞
s=1
c
(1)
s
(J−1)
s
∞
s=1
c
(1)
s
J
s
. (59b)
Continued fraction formula. Finally, for a third approach for an approxi
mation to the DTBC, we use a continued fraction formulation. Following [22,
Section 5.6] we can easily deduce an expression for the quotient of Bessel func
tions (like (52)) as a continued fraction from the recurrence formula (15). If
we rewrite (15) as
J
ν−1
(z)
J
ν
(z)
= 2νz
−1
−
1
Jν(z)
J
ν+1
(z)
,
it is obvious that
J
ν−1
(z)
J
ν
(z)
= 2νz
−1
−
1
2(ν + 1)z
−1
− ... −
1
2(ν + M)z
−1
−
J
ν+M−1
(z)
J
ν+M
(z)
.
This holds for general values of ν and it can be shown, with the help of the
theory of Lommel polynomials [22, Section 9.65], that when M →∞, the last
quotient may be neglected, so that
J
ν−1
(z)
J
ν
(z)
= 2νz
−1
−
1
2(ν + 1)z
−1
−
1
2(ν + 2)z
−1
− ...
. (60)
This continued fractions formula oﬀers another way to evaluate the quotient
of two Bessel functions needed in the transformed discrete TBC (52). For the
numerical implementation we use the modiﬁed Lentz’s method [13] which is
an eﬃcient general method for evaluating continued fractions.
Remark 3 Our practical calculations showed that the evaluation of the con
tinued fraction (60) is stable for all considered values of ν and z although we
cannot prove this yet.
Remark 4 For brevity of the presentation, we omit here the discussion of an
adequate discrete treatment of the typical density shock at z = z
b
and refer the
reader to [2] for a detailed discussion of various strategies.
5 Approximation by Sums of Exponentials
An adhoc implementation of the discrete convolution (57) with convolution
coeﬃcients s
(n)
J
from (55) (or obtained by any of the above approaches) has
still one disadvantage. The boundary condition is non–local and therefore
computationally expensive. In fact, the evaluation of (57) is as expensive as
14
for the discretized TBC (42). As a remedy, we proposed in [3] the sumof
exponentials ansatz (for a comparison of the computational eﬀorts see Fig. 6).
In the sequel we will brieﬂy review this approach.
In order to derive a fast numerical method to calculate the discrete convo
lutions in (57), we approximate the coeﬃcients s
(n)
J
by the following (sum of
exponentials):
s
(n)
J
≈ ˜ s
(n)
J
:=
_
¸
¸
_
¸
¸
_
s
(n)
J
, n = 0, 1
L
l=1
b
l
q
−n
l
, n = 2, 3, . . . ,
(61)
where L ∈ N is a ﬁxed number. Evidently, the approximation properties of
˜ s
(n)
J
depend on L, and the corresponding set ¦b
l
, q
l
¦. Below we propose a
deterministic method of ﬁnding ¦b
l
, q
l
¦ for ﬁxed L.
Let us ﬁx L and consider the formal power series:
g(x) := s
(2)
J
+ s
(3)
J
x + s
(4)
J
x
2
+ . . . , [x[ ≤ 1. (62)
If there exists the [L −1[L] Pad´e approximation
˜ g(x) :=
P
L−1
(x)
Q
L
(x)
(63)
of (62), then its Taylor series
˜ g(x) = ˜ s
(2)
J
+ ˜ s
(3)
J
x + ˜ s
(4)
J
x
2
+ . . . (64)
satisﬁes the conditions
˜ s
(n)
J
= s
(n)
J
, n = 2, 3, . . . , 2L + 1, (65)
due to the deﬁnition of the Pad´e approximation rule.
Theorem 5 ([3]) Let Q
L
(x) have L simple roots q
l
with [q
l
[ > 1, l =
1, . . . , L. Then
˜ s
(n)
J
=
L
l=1
b
l
q
−n
l
, n = 2, 3, . . . , (66)
where
b
l
:= −
P
L−1
(q
l
)
Q
L
(q
l
)
q
l
,= 0, l = 1, . . . , L. (67)
It follows from (65) and (66) that the set ¦b
l
, q
l
¦ deﬁned in Theorem 5 can be
used in (61) at least for n = 2, 3, .., 2L + 1. The main question now is: Is it
possible to use these ¦b
l
, q
l
¦ also for n > 2L+1? In other words, what quality
of approximation
˜ s
(n)
J
≈ s
(n)
J
, n > 2L + 1 (68)
15
can we expect?
The above analysis permits us to give the following description of the approx
imation to the convolution coeﬃcients s
(n)
J
by the representation (61) if we
use a [L − 1[L] Pad´e approximant to (62): the ﬁrst 2L coeﬃcients are repro
duced exactly, see (65); however, the asymptotic behaviour of s
(n)
J
and ˜ s
(n)
J
(as
n →∞) diﬀers strongly (algebraic versus exponential decay). A typical graph
of [s
(n)
J
− ˜ s
(n)
J
[ versus n for L = 27 is shown in Fig. 2 in Section 7.
Fast Evaluation of the Discrete Convolution. Let us consider the ap
proximation (61) of the discrete convolution kernel appearing in the DTBC
(57). With these “exponential” coeﬃcients the convolution
C
(n)
:=
n−1
m=1
˜ s
(n−m)
J
ψ
(m)
J
, ˜ s
(n)
J
=
L
l=1
b
l
q
−n
l
, (69)
[q
l
[ > 1, of a discrete function ψ
(m)
J
, m = 1, 2, . . . , with the kernel coeﬃ
cients ˜ s
(n)
J
, can be calculated by recurrence formulas, and this will reduce the
numerical eﬀort signiﬁcantly (cf. Fig. 6 in Section 7).
A straightforward calculation (cf. [3]) yields: The value C
(n)
from (69) for
n ≥ 2 is represented by
C
(n)
=
L
l=1
C
(n)
l
, (70)
where
C
(1)
l
≡ 0,
C
(n)
l
= q
−1
l
C
(n−1)
l
+ b
l
q
−1
l
ψ
(n−1)
J
, (71)
n = 2, 3, . . . l = 1, . . . , L.
Finally we summarize the approach by the following algorithm:
1. calculate s
(n)
J
, n = 0, . . . , N −1, via numerical inverse Ztransformation;
2. calculate ˜ s
(n)
J
via Pad´e–algorithm;
3. the corresponding coeﬃcients b
l
, q
l
are used for the eﬃcient calculation
of the discrete convolutions.
Remark 6 We note that the Pad´e approximation must be performed with high
precision (2L − 1 digits mantissa length) to avoid a ‘nearly breakdown’ by ill
conditioned steps in the Lanczos algorithm (cf. [6]). If such problems still occur
or if one root of the denominator is smaller than 1 in absolute value, the orders
of the numerator and denominator polynomials are successively reduced.
16
6 Stability Analysis of the numerical scheme
Here we analyze the stability of our numerical scheme for the SPE (37) along
with a surface condition and the the DTBC (53):
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
−iR(ψ
(n+1)
j
−ψ
(n)
j
) = ρ
j
∆
0
z
(ρ
−1
j
∆
0
z
)(ψ
(n+1)
j
+ψ
(n)
j
)+w
_
(N
2
)
(n)
j
−1
_
(ψ
(n+1)
j
+ψ
(n)
j
)
j = 1, . . . , J −1,
ψ
(0)
j
= ψ
I
(z
j
), j = 0, 1, 2, . . . , J −1, J;
with ψ
(0)
J−1
= ψ
(0)
J
= 0,
ψ
(n)
0
= 0,
ˆ
ψ
J−1
(z) = ˆ g
µ,J
(z)
ˆ
ψ
J
(z),
(72)
where ˆ g
µ,J
(z) is given by (52).
In the sequel we want to derive an apriori estimate of the discrete solution in
the discrete weighted
2
–norm:
ψ
(n)

2
2
:= h
J−1
j=1
[ψ
(n)
j
[
2
ρ
−1
j
, (73)
which is the discrete analogue to (27). The following theorem bounds the expo
nential growth of solutions to the numerical scheme for a ﬁxed discretization.
Theorem 7 (Growth condition) Let the boundary kernel ˆ g
µ,J
satisfy
Im ˆ g
µ,J
(γe
iϕ
) ≤ 0, ∀ 0 ≤ ϕ ≤ 2π, (74)
for some (suﬃciently large) γ ≥ 1 (i.e. the system is dissipative). Assume
also that ˆ g
µ,J
(z) is analytic for [z[ ≥ γ. Then, the solution of (72) satisﬁes
the apriori estimate
ψ
(n)

2
≤ ψ
0

2
γ
n
, n ∈ N. (75)
PROOF. The proof is based on a discrete energy estimate for the new vari
able
φ
(n)
j
:= ψ
(n)
j
γ
−n
,
17
which satisﬁes the equation
−iR(φ
(n+1)
j
−φ
(n)
j
) =
_
ρ
j
∆
0
z
(ρ
−1
j
∆
0
z
) + w
_
(N
2
)
(n)
j
−1
__
(φ
(n+1)
j
+ φ
(n)
j
) (76a)
+ (γ −1)
_
ρ
j
∆
0
z
(ρ
−1
j
∆
0
z
) + w
_
(N
2
)
(n)
j
−1
_
+ iR
_
φ
(n+1)
j
,
j = 1, . . . , J −1,
φ
(0)
j
= ψ
(0)
j
, j = 0, . . . , J, (76b)
φ
(n)
0
= 0, (76c)
∆
−
ˆ
φ
J
(z) = −(ˆ g
µ,J
(γz) −1)
ˆ
φ
J
(z). (76d)
In physical space, the bottom BC can be written as
∆
−
φ
(n)
J
= −φ
(n)
J
∗
˜
(n)
J
γ
n
= −
n
m=0
φ
(m)
J
_
˜
(n−m)
J
γ
m−n
_
. (77)
First we multiply (76a) by
¯
φ
(n)
j
ρ
−1
j
/γ and its complex conjugate by φ
(n+1)
j
ρ
−1
j
:
iR
_
[φ
(n)
j
[
2
−
¯
φ
(n)
j
φ
(n+1)
j
_
ρ
−1
j
=
¯
φ
(n)
j
_
∆
0
z
(ρ
−1
j
∆
0
z
) + w
_
(N
2
)
(n)
j
−1
_
ρ
−1
j
_ _
φ
(n+1)
j
+ φ
(n)
j
_
+(γ
−1
−1)
¯
φ
(n)
j
_
∆
0
z
(ρ
−1
j
∆
0
z
) +
_
w
_
(N
2
)
(n)
j
−1
_
−iR
_
ρ
−1
j
_
φ
(n)
j
,
(78a)
iR
_
[φ
(n+1)
j
[
2
−
¯
φ
(n)
j
φ
(n+1)
j
_
ρ
−1
j
=
φ
(n+1)
j
_
∆
0
z
(ρ
−1
j
∆
0
z
) + w
_
(N
2
)
(n)
j
−1
_
ρ
−1
j
_ _
¯
φ
(n+1)
j
+
¯
φ
(n)
j
_
+(γ −1)φ
(n+1)
j
_
∆
0
z
(ρ
−1
j
∆
0
z
) +
_
w
_
(N
2
)
(n)
j
−1
_
−iR
_
ρ
−1
j
_
¯
φ
(n+1)
j
.
(78b)
Next we subtract (78a) from (78b), sum from j = 1 to j = J −1, and apply
18
summation by parts:
iR
J−1
j=1
_
[φ
(n+1)
j
[
2
−[φ
(n)
j
[
2
_
ρ
−1
j
= −
J−1
j=1
_
[∆
−
φ
(n+1)
j
[
2
−[∆
−
φ
(n)
j
[
2
_
ρ
−1
j−1/2
+
_
φ
(n+1)
J−1
∆
−
_
¯
φ
(n+1)
J
+
¯
φ
(n)
J
_
−
¯
φ
(n)
J−1
∆
−
_
φ
(n+1)
J
+ φ
(n)
J
_
_
ρ
−1
J−1/2
+ w
J−1
j=1
__
(N
2
)
(n)
j
−1
_
([φ
(n+1)
j
[
2
+
¯
φ
(n)
j
φ
(n+1)
j
) −
_
(N
2
)
(n)
j
−1
_
([φ
(n)
j
[
2
+
¯
φ
(n)
j
φ
(n+1)
j
)
_
ρ
−1
j
+ w
J−1
j=1
_
(γ −1)
_
(N
2
)
(n)
j
−1
_
[φ
(n+1)
j
[
2
+ (1 −γ
−1
)
_
(N
2
)
(n)
j
−1
_
[φ
(n)
j
[
2
_
ρ
−1
j
−iR
J−1
j=1
_
(γ −1)[φ
(n+1)
j
[
2
+ (1 −γ
−1
)[φ
(n)
j
[
2
_
ρ
−1
j
−(γ −1)
J−1
j=1
[∆
−
φ
(n+1)
j
[
2
ρ
−1
j
−(1 −γ
−1
)
J−1
j=1
[∆
−
φ
(n)
j
[
2
ρ
−1
j
+
_
(γ −1)φ
(n+1)
J−1
∆
−
¯
φ
(n+1)
J
+ (1 −γ
−1
)
¯
φ
(n)
J−1
∆
−
φ
(n)
J
_
ρ
−1
J−1/2
(79)
Now, taking imaginary parts one obtains after a lengthy calculation:
J−1
j=1
_
[φ
(n+1)
j
[
2
−[φ
(n)
j
[
2
_
ρ
−1
j
= −(γ −1)
J−1
j=1
[φ
(n+1)
j
[
2
ρ
−1
j
−(1 −γ
−1
)
J−1
j=1
[φ
(n)
j
[
2
ρ
−1
j
−
w
γR
J−1
j=1
Im
_
(N
2
)
(n)
j
−1
_
[φ
(n)
j
+ γφ
(n+1)
j
[
2
ρ
−1
j
−
1
γRρ
J−1/2
Im
_
_
¯
φ
(n)
J
+ γ
¯
φ
(n+1)
J
_
∆
−
_
φ
(n)
J
+ γφ
(n+1)
J
_
_
.
(80)
Summing (80) from n = 0 to n = N yields (note that γ ≥ 1):
φ
(N+1)

2
2
≤ φ
(0)

2
2
−
wh
γ
2
R
N
n=0
J−1
j=1
Im
_
(N
2
)
(n)
j
−1
_
[φ
(n)
j
+ γφ
(n+1)
j
[
2
ρ
−1
j
−
h
γ
2
Rρ
J−1/2
Im
N
n=0
(
¯
φ
(n)
J
+ γ
¯
φ
(n+1)
J
)∆
−
(φ
(n)
J
+ γφ
(n+1)
J
)
= φ
(0)

2
2
−
kh
2γ
2
N
n=0
J−1
j=1
α
(n+1/2)
j
c
0
c
(n+1/2)
j
[φ
(n)
j
+ γφ
(n+1)
j
[
2
ρ
−1
j
+
k
4γ
2
k
0
hρ
J−1/2
Im
N
n=0
(
¯
φ
(n)
J
+ γ
¯
φ
(n+1)
J
) (φ
(n)
J
+ γφ
(n+1)
J
) ∗
˜
(n)
J
γ
n
.
(81)
19
For the last identity we used the bottom BC (77) and φ
(0)
0
= φ
(0)
J
= 0.
Since α
(n+1/2)
j
≥ 0, it remains to determine the sign of the last term in (81) to
ﬁnish the proof. To this end we deﬁne (for N ﬁxed) the two sequences,
u
(n)
:=
_
_
_
φ
(n)
J
+ γφ
(n+1)
J
, n = 0, . . . , N,
0, n > N,
v
(n)
:= u
(n)
∗
˜
(n)
J
γ
n
=
n
m=0
u
(m)
˜
(n−m)
J
γ
n−m
, n ∈ N
0
.
The Z–transform Z¦u
(n)
¦ = ˆ u(z) is analytic for [z[ > 0, since it is a ﬁnite
sum. The Z–transform Z¦v
(n)
¦ then satisﬁes ˆ v(z) = (ˆ g
µ,J
(γz) −1)ˆ u(z) and is
analytic for [z[ ≥ 1. Using Plancherel’s Theorem for Z–transforms we have
N
n=0
v
(n)
¯ u
(n)
=
∞
n=0
v
(n)
¯ u
(n)
=
1
2π
_
2π
0
ˆ v(e
iϕ
)ˆ u(e
iϕ
) dϕ
=
1
2π
_
2π
0
[ˆ u(e
iϕ
)[
2
_
ˆ g
µ,J
(γe
iϕ
) −1
_
dϕ.
(82)
Using (82) for the boundary term in (81) now gives:
φ
(N+1)

2
2
≤ φ
(0)

2
2
+
h
2πRγ
2
ρ
J−1/2
_
2π
0
[(1 + γe
iϕ
)
ˆ
φ
J
(e
iϕ
)[
2
Im
_
ˆ g
µ,J
(γe
iϕ
) −1
_
dϕ.
(83)
Our assumption on ˆ g
µ,J
therefore implies
φ
(N)

2
≤ φ
(0)

2
, ∀ N ≥ 0,
and the result of the theorem follows.
Remark 8 Above we have assumed that the Ztransformed boundary kernel
ˆ g
µ,J
(z) is analytic for [z[ ≥ β. Hence its imaginary parts is a harmonic
functions there. Since the average of ˆ g
µ,J
(z) on the circles z = βe
iϕ
equals
g
(0)
µ,J
= ˆ g
µ,J
(z = ∞), condition (74) implies Im ˆ g
µ,J
(z = ∞) ≤ 0. Then
we have the following simple consequence of the maximum principle for the
Laplace equation:
If condition (74) holds for some β
0
, it also holds for all β > β
0
.
20
7 Numerical Example
In the example of this Section we will consider the SPE for comparing the
numerical result from using our new (approximated) discrete TBC to the
solution using the discretized TBC of Levy [14]. We used the environmental
test data from [7] and the Gaussian beam from [12] as starting ﬁeld ψ
I
. Below
we present the socalled transmission loss −10 log
10
[p[
2
, where the acoustic
pressure p is calculated from (23). We computed a reference solution on a
three times larger computational domain conﬁned with the DTBC from [2].
Example. As an illustrating example we chose the typical downward refract
ing case (i.e. energy loss to the bottom): µ = 2 10
−4
m
−1
. The source at the
depth z
s
= 91.44 m is emitting sound with a frequency f = 300 Hz and the re
ceiver is located at the depth z
r
= 27.5 m. The TBC is applied at z
b
= 152.5 m
and the discretization parameters are given by ∆r = 10 m, ∆z = 0.5 m. It con
tains no attenuation: α = 0. We consider a rangeindependent situation for
0 < r < 50 km, i.e. 5000 range steps. The sound speed varies linearly from
c(0 m) = 1536.5 m s
−1
to c(152.5 m) = 1539.24 m s
−1
. The reference sound
speed c
0
is chosen to be equal to c(z
b
) such that β = 0 in (25).
For this choice of parameters the mesh ratio becomes R ≈ 0.12246 and the
parameter σ ≈ −53345.32; that is, the value of ν
J
deﬁned in (51) is much too
large for the routines like COULCC [21] for evaluating Bessel functions. On the
other hand, using the asymptotic formula (58) is not advisable since for large
ν
J
we have
ˆ
h
µ,J
(z) ∼ 2(1−iζ(z)) which is only the ﬁrst term in the continued
fraction expansion (60). Therefore, we decided to evaluate the ratio of the two
Bessel functions in (52) by the continued fraction formula (60) together with
the sumofexponentials ansatz (61). We note that all approaches fulﬁlled for
moderate choices of ν
J
the growth condition (74) needed for stability.
We computed the ﬁrst 1000 terms in the expansion (60) and used a radius
τ = 1.04 with 2
10
sampling points for the numerical inverse Z–transformation
(54). The choice of an appropriate radius τ is a delicate problem: it may not
be too close to the convergence radius of (60) due to the approximation error
and τ too large raises problems with rounding errors during the rescaling
process. For a discussion of that topic we refer the reader to [3, Section 2], [16]
and [25]. In order to calculate the convolution coeﬃcients b
n
(discretized TBC
of Levy) we used the MATLAB routine from [24] to compute the ﬁrst 100
zeros of the Airy function. Alternatively, using precomputed values from the
call evalf(AiryAiZeros(1..100)); in MAPLE with high precision yielded
indistinguishable results.
First we examine the convolution coeﬃcients of the two presented approaches.
Fig. 1 shows a comparison of the coeﬃcients b
n
from the discretized TBC (42)
21
with the coeﬃcients s
(n)
J
from the approximated discrete TBC. The coeﬃcients
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
10
−20
10
−15
10
−10
10
−5
10
0
n
convolution coefficients b
n
, s
J
(n) ~
approximated discrete TBC
discretized TBC of Levy
Fig. 1. Comparison of the convolution coeﬃcients b
n
of the discretized TBC (42)
and ˜ s
(n)
J
from the approximated discrete TBC (with L = 27).
b
n
decay even faster than the coeﬃcients s
(n)
J
. In Fig. 2 we plot both the exact
convolution coeﬃcients s
(n)
J
and the error [s
(n)
J
− ˜ s
(n)
J
[ versus n for L = 27
(observe the diﬀerent scales).
Now we investigate the stability of the approximated discrete TBC and check
the growth condition (74). For β = 1 we have max¦Im(ˆ g
µ,J
(β e
iϕ
)¦ = 0.153
and, with β = 1.01, we obtain max¦Im(ˆ g
µ,J
(β e
iϕ
)¦ = −0.002 (see Fig. 3).
This means that the Z–transformed kernel ˆ g
µ,J
(β e
iϕ
) of the approximated
discrete TBC satisﬁes the stability condition (74) for β ≥ 1.01 (for this dis
cretization).
In Fig. 4 and Fig. 5 we compare the transmission loss results for the discretized
TBC and the approximated discrete TBC in the range from 0 to 50 km. The
transmission loss curve of the solution using the approximated DTBC is indis
tinguishable from the one of the reference solution while the solution with the
discretized TBC still deviates signiﬁcantly from it (and is more oscillatory) for
the chosen discretization. The result in Fig. 5 does not change if we compute
more zeros of the Airy function.
Evaluating the convolution appearing in the discretized TBC (42) is quite ex
pensive for longrange calculations. Therefore we extended the range interval
up to 250 km and shall now illustrate the diﬀerence in the computational ef
22
0 20 40 60 80 100 120 140 160 180 200
0
0.03
0.06
0.09
n
s
J
(n)
 and error s
J
(n)
− s
J
(n) ~


s
J (
n
)

0 20 40 60 80 100 120 140 160 180 200
0
1
2
3
4
x 10
−5
e
r
r
o
r

s
J (
n
)
−
s
J (
n
)
~

Fig. 2. Convolution coeﬃcients s
(n)
J
(left axis, dashed line) and error s
(n)
J
− ˜ s
(n)
J
 of
the convolution coeﬃcients (right axis); (L = 27).
0 π/2 π 3/2π 2π
−1.5
−1
−0.5
0
0.5
ϕ
I
m
g ^
µ
,
J
(
z
)
L=27: Im g
^
µ,J
(z) on circle, β=1, β=1.01
β = 1
β = 1.01
Fig. 3. Growth condition ˆ g
µ,J
(z), z = β e
iϕ
(L = 27).
fort for both approaches in Fig. 6: The computational eﬀort for the discretized
TBC is quadratic in range, since the evaluation of the boundary convolutions
dominates for large ranges. On the other hand, the eﬀort for the approximated
discrete TBC only increases linearly. The line (  ) does not change consid
23
0 5 10 15 20 25 30 35 40 45 50
50
55
60
65
70
75
80
85
90
95
100
Range r [km]
T
r
a
n
s
m
i
s
s
i
o
n
L
o
s
s
[
d
B
]
Discretized TBC of Levy
Fig. 4. Transmission loss at z
r
= 27.5 m.
0 5 10 15 20 25 30 35 40 45 50
50
55
60
65
70
75
80
85
90
95
100
Range r [km]
T
r
a
n
s
m
i
s
s
i
o
n
L
o
s
s
[
d
B
]
Approximated Discrete TBC (L=27)
Fig. 5. Transmission loss at z
r
= 27.5 m.
erably for diﬀerent values of L since the evaluation of the sumofexponential
convolutions has a negligible eﬀort compared to solving the PDE in the interior
domain.
24
0 5000 10000 15000 20000 25000
0
50
100
150
200
250
300
350
e
l
a
p
s
e
d
c
p
u
t
i
m
e
time steps
discretized TBC vs. approximated discrete TBC (L=27), 25000 range steps
Fig. 6. Comparison of CPU times: the discretized TBC of Levy has quadratic eﬀort
(—), while the sumofexponential approximation to the discrete TBC has only
linear eﬀort (  ).
Conclusion
We have proposed a variety of strategies to derive an approximation to the
discrete TBC for the Schr¨ odinger equation with a linear potential term in the
exterior domain. The derivation was based on the knowledge of the exact solu
tion (including the asymptotics) to the discrete Airy equation. Our approach
has two advantages over the standard approach of discretizing the continu
ous TBC: higher accuracy and eﬃciency; while discretized TBCs have usually
quadratic eﬀort, the sumofexponential approximation to discrete TBCs has
only linear eﬀort. Moreover, we have provided a simple criteria to check the
stability of our method.
Acknowledgement
The ﬁrst author thanks Prof. A. Arnold for many helpful suggestions to this
article.
25
References
[1] A. Arnold, Numerically Absorbing Boundary Conditions for Quantum Evolution
Equations, VLSI Design 6 (1998), 313–319.
[2] A. Arnold and M. Ehrhardt, Discrete transparent boundary conditions for wide
angle parabolic equations in underwater acoustics, J. Comp. Phys. 145 (1998),
611–638.
[3] A. Arnold, M. Ehrhardt and I. Sofronov, Discrete transparent boundary
conditions for the Schr¨ odinger equation: Fast calculation, approximation, and
stability, Comm. Math. Sci. 1 (2003), 501–556.
[4] E.W. Barnes, On the homogeneous linear diﬀerence equation of the second order
with linear coeﬃcients, Messenger 34 (1904), 52–71.
[5] P.M. Batchelder, An Introduction to Linear Diﬀerence Equations, (Harvard
University Press, Cambridge, MA, 1927).
[6] A. Bultheel and M. Van Barel, Linear algebra, rational approximation and
orthogonal polynomials, (Studies in Computational Mathematics 6, North–
Holland, 1997).
[7] T.W. Dawson, D.J. Thomson and G.H. Brooke, Non–local boundary conditions
for acoustic PE predictions involving inhomogeneous layers, in 3rd European
Conference on Underwater Acoustics, FORTH–IACM, Heraklion, 1996, pp.
183–188.
[8] M. Ehrhardt, Discrete Artiﬁcial Boundary Conditions, Ph.D. Thesis,
Technische Universit¨ at Berlin, 2001.
[9] M. Ehrhardt and A. Arnold, Discrete Transparent Boundary Conditions for the
Schr¨ odinger Equation, Riv. Mat. Univ. Parma 6 (2001), 57–108.
[10] W.G. Kelley and A.C. Peterson, Diﬀerence equations: an introduction with
applications, (Academic Press, Boston, 1991).
[11] N.N. Lebedev, Special Functions and their Applications, Selected Russian
Publications in the Mathematical Sciences, Prentice–Hall, 1965.
[12] D. Lee and S.T. McDaniel, Ocean acoustic propagation by ﬁnite diﬀerence
methods, Comput. Math. Appl. 14 (1987), 305–423.
[13] W.J. Lentz, Generating Bessel Functions in Mie Scattering Calculations Using
Continued Fractions, Appl. Opt. 15 (1976), 668–671.
[14] M.F. Levy, Transparent Boundary Conditions for Parabolic Equation Solutions
of Radiowave Propagation Problems, IEEE Trans. Antennas Propag. 45 (1997),
66–72.
[15] M.F. Levy, Parabolic equation models for electromagnetic wave propagation,
(IEE Electromagnetic Waves Series 45, 2000).
26
[16] C. Lubich, Convolution Quadrature and Discretized Operational Calculus. II,
Numer. Math. 52 (1988), 413–425.
[17] R.E. Mickens, Asymptotic properties of solutions to two discrete Airy equations,
J. Diﬀerence Equ. Appl. 3 (1998), 231–239.
[18] R.E. Mickens, Asymptotic solutions to a discrete Airy equation, J. Diﬀerence
Equ. Appl. 7 (2001), 851–858.
[19] F.W.J. Olver, Asymptotics and Special Functions, (Academic Press, New York,
1974).
[20] F.D. Tappert, The parabolic approximation method, in: Wave Propagation
and Underwater Acoustics, Lecture Notes in Physics 70, eds. J.B. Keller and
J.S. Papadakis, (Springer, New York, 1977), 224–287.
[21] I.J. Thompson and A.R. Barnett, Coulomb and Bessel Functions of Complex
Arguments and Order, J. Comp. Phys. 46 (1986), 490–509.
[22] G.N. Watson, A treatise on the theory of Bessel functions, 2nd ed., (Cambridge
Univ. Press, Cambridge, 1966).
[23] R. Wong and H. Li, Asymptotic expansions for second–order linear diﬀerence
equations. II, Stud. Appl. Math. 87 (1992), 289–324.
[24] S. Zhang and J. Jin, Computation of special functions: with over 100 computer
programs in FORTRAN, John Wiley & Sons, New York, 1996). (MATLAB
routines: http://ceta.mit.edu/comp spec func/)
[25] A. Zisowsky, Discrete Transparent Boundary Conditions for Systems of
Evolution Equations, Ph.D. Thesis, Technische Universit¨ at Berlin, 2003.
27
1
Introduction
In this work we consider the Airy diﬀerential equation d2 y − xy = 0. dx2 (1)
The solutions of this secondorder diﬀerential equation (1) are called Airy functions and can be expressed in terms of Bessel functions of imaginary argument of order ν = ± 1 . They play an important role in the theory of 3 asymptotic expansions of various special functions and have a wide range of applications in mathematical physics. The two linearly independent solutions of (1), the Airy functions of the ﬁrst and second kind Ai(x), Bi(x), respectively have the following asymptotic representation for large x [11]: 1 3/2 −3/2 Ai(x) = √ x−1/4 e−2/3x [1 + O(x )], 2 π 1 3/2 Bi(x) = √ x−1/4 e2/3x [1 + O(x−3/2 )]. π (2a) (2b)
To solve the Airy equation (1) numerically we introduce the uniform grid points xm = m∆x, ym y(xm), and consider the standard discretization : ym+1 − 2 ym + ym−1 − xm ym = 0, (∆x)2 which can be rewritten as the secondorder diﬀerence equation ym+1 − 2 ym + ym−1 − c m ym = 0, c = (∆x)3. (4) (3)
In [17] Mickens derived the following asymptotic behaviour of the two linearly independent discrete solutions to (4) with c = 1:
(1) ym =
m7/2 em mm mm e−m (2) ym = m9/2
1−
1 85 +O 12m m2 133 1 1+ +O 12m m2
, .
(5a) (5b)
It can easily be seen that the solutions to this discretization (4) do not have the same asymptotic properties as the solutions of (1) which motivated the construction of a nonstandard discretization scheme (cf. [17], [18]). This paper is organized as follows: ﬁrst we discuss a generalization of the the discrete Airy equation (4) and show how to ﬁnd exact and asymptotic solutions. Afterwards we present an application to a problem arising in “parabolic 2
The asymptotic solution. d ∈ C. (7) where p. Here we want to apply the approach of Wong and Li [23] to obtain asymptotic solutions to the second–order diﬀerence equation ym+2 + mp a(m) ym+1 + mq b(m) ym = 0. (6) and show how to determine asymptotic solutions. 2 Exact and asymptotic solutions to the discrete Airy equation In this section we consider the discrete Airy equation in the more general form ym+1 − 2 ym + ym−1 − (d + c m) ym = 0. s s=0 m ∞ (8) with nonzero leading coeﬃcients: a0 = 0. d = 0). Afterwards we will demonstrate how to obtain an explicit solution to (6).equation” calculations in (underwater) acoustics and radar propagation in the troposphere: we construct a socalled discrete transparent boundary condition (DTBC) for a Schr¨dinger equation with a linear potential term. Discrete transparent boundary conditions for a Schr¨dinger equation with constant potential were derived in [1] o and a generalization to a problem arising in (underwater) acoustics was presented in [2]. Finally we illustrate the results with a numerical example from underwater acoustics. q are integers and a(m) and b(m) have power expansions of the form a(m) = as . s s=0 m ∞ b(m) = as . Since the classic theorems of Poincar´ and Perron e cannot be applied to (6) it is not straight forward to obtain information about the asymptotic behaviour of the solutions to this equation. Afterwards we analyze the stability of the resulting numerical scheme. In this case the Laplace transformed Schr¨dinger equation o can be viewed as a general Airy equation. We increase the index of (6) 3 . One possible ansatz is the one of Batchelder [5] given in [17] (see (5) in the case c = 1. b0 = 0. c. For such an equation. discuss o diﬀerent approaches and present an eﬃcient implementation by the sumofexponentials ansatz. This is a nontrivial task since equation (6) is not of Poincar´ type. the coeﬃcients e must approach constant values as m → ∞ and it is clearly seen that the coeﬃcient of ym in equation (6) does not satisfy this condition. The construction of DTBC to the nonstandard discretization scheme and to wideangle parabolic equations will be a topic of future work.
b0 = 1. a1 = −(2 + c + d). − 3 2 (1) etc. the results: θ (1) (1) c1 = −c0 c−2 − θ + (θ − 1) . s ≥ 1. . c2 . ms s=0 (1) To determine the values of the coeﬃcients c1 . s (m − 2)! s=0 m ∞ (1) ym = (9a) (9b) (2) ym = (m − 2)! cm m1+(2+d)/c c(2) s . (10) (m + 1)s s=0 ∞ (1) (1) 1 m+1 cm m + 2 θ ∞ θ ∞ = c(θ − 1) + cm with θ = 2 + (2 + d)/c. 2 2 6 2 2 (11b) (1) c3 c θ θ θ =− 0 2θ + (θ − 1) c−2 − (θ − 1)(θ − 2) − (θ − 1)(θ − 2)(θ − 3) 3 2 6 24 (1) c θ θ + 1 (2 + θ)c−2 − 2 + θ + (θ − 1) − (θ − 1)(θ − 2) 3 2 6 (1) θ c2 (11c) c−2 − 5 + θ + (θ − 1) . by a lengthy but elementary calculation. p = 1. Then the two formal series solutions (cf. as = 0. [23]) are given by ∞ (1) c−m cs m−2−(2+d)/c . q = 0. 4 . we substitute the (1) decaying solution ym in (7): c(1) m+1 s + c(m − 1) s m s=0 (m + 2) c(1) s ms s=0 c(1) s . We now obtain after an Taylor expansion in 1/m and setting all the linearly independent terms equal to zero. c3 . (11a) 2 (1) (1) θ c1 c0 θ θ (1) −2 c2 = θc + (θ − 1) − (θ − 1)(θ − 2) − c−2 − 2 + (θ − 1) . .by one to put it in the form of (7) and make the following identiﬁcations: a0 = −c.. s ≥ 2. bs = 0. .
To start with. d = 0 we obtain (1) ym = c 0 (2) ym (1) (2) (12a) (12b) (12c) m−4 3 21 104 1− + − + O m−4 . In this section we show that in the special case of the discrete Airy equation (6). 2 (2) c0 η (2) c2 = − (η − 1)c−2 − η + η(η − 1) − (η − 1)(η − 2) 2 6 (2) c η + 1 c−2 + 3 − 2η − (η − 1) . Here c0 . [10]). deﬁne the generating function to be ∞ g(ξ) = m=−∞ ym ξ m . c0 denote arbitrary constants. it is possible to obtain an explicit solution. one approach is to ﬁnd the solution by the “method of generating functions” (cf. a generating function for a solution of (6) can be shown to satisfy a diﬀerential equation. If one wants to solve a diﬀerence equation with polynomial coeﬃcients (like (6)). which may be solvable in terms of known functions. Remark 1 We remark that in the special case c = 1. 2 2 (2) c 3 η η (2) c3 = 0 1 + (η − 1) c−2 − η + η(η − 1) − (η − 1)(η − 2) 3 2 2 2 η + (η − 1)(η − 2)(η − 3) 24 (2) c η − 1 (η − 1)c−2 − 7 − 6η + 2η(η − 1) − (η − 1)(η − 2) 3 6 (2) c η + 2 c−2 + 9 − 3η + (η − 1) .. m=−∞ 5 . In most cases. 2 3 m 2m 3m (1) (13a) (13b) The explicit solution. i. 2 (m − 2)! m 2m 3m3 1 3 1 (2) = c0 (m − 2)! m3 1 + − − + O m−4 . secondorder linear diﬀerence equations with variable coeﬃcients cannot be solved in closed form. (14) We multiply (6) with ξ m−1 and sum it up for m ∈ Z: ∞ ∞ ∞ ∞ ym ξ m=−∞ m−2 − (2 + d) ym ξ m=−∞ m−1 + m=−∞ ym ξ − c m mym ξ m−1 = 0.e. We present the derivation of this exact solution and study its asymptotic behaviour.(2) Similarly we obtain for the increasing solution ym the ﬁrst three coeﬃcients η (2) (2) c1 = c0 c−2 − η + (η − 1) . 3 2 with η = 1 + (2 + d)/c.
cξ 2 e(ξ− ξ )/c = ξ − 1 2+d c ∞ ν=−∞ Jν ( 2 )ξ ν . One can derive the following dominating series for the decaying solution: 2+d 1 1 2 ym  ≤ Ym = e c (m+ c +1) . 2+d 2+d m+ c  c  Γ(m + c + 1) 1 (17) which can be estimated using Stirling’s inequality 1 e c2 Ym < √ 2+d 2π cm+ c 1 e m m m− 2 1 Γ(m + 1) . i. 2+d 2n n=0 c n! Γ(m + c + n + 1) ∞ c. [22]) c that the recurrence equation for the Bessel functions Jν+1 (z) − 2 ν Jν (z) + Jν−1 (z) = 0. z (15) still holds for complex orders ν and complex arguments z. The next step is to use (16) to examine the asymptotic behaviour of the discrete solutions. Remark 2 A diﬀerence equation more general than (6) was examined by Barnes [4] in 1904.e. the discrete Airy equation is nothing else but the recurrence relation for Jν ( 2 ). diﬀerentiating each term is permissible (cf.e.1]): ym = Jm+ 2+d ( 2 ) = c c 1 cm+ 2+d c (−1)n . [22]). (16) We also observe that (14) is not a generating function in the strict sense but a Laurent series.This results in the following ordinary diﬀerential equation for g: g (ξ) − for which the solution is g(ξ) = ξ − 2+d c 1 − (2 + d)ξ + ξ 2 g(ξ) = 0. This solution is the socalled “NeumannFunction” (or Bessel function of the second kind) which is known to also satisfy the recursion equation of the Bessel functions. We note that it is wellknown ([11]. Γ(m + 2+d + 1) c (18) 6 . Chapter 3. He also considered (6) and found (through a diﬀerent construction) the solution (16). [22. i. Note that this generating function approach is not suitable for determining the growing solution of (6) for m → ∞. c Hence. d ∈ C. which is uniformly convergent. the exact decaying solution of (6) is the Bessel function Jν ( 2 ) (regarded c as function of its order ν). Thus the decaying solution to (6) can be represented as (cf.
We use the following asymptotic representation of Bessel functions for large values of the order ν (cf.  arg ν ≤ π − δ. The water surface is at z = 0. where f denotes the frequency of the emitted sound. r) ≥ 0. From setting c = 1. (19) mm 2π and we see that the decay rate of (18) and (5) diﬀers by a factor m−6 . The standard parabolic equation in underwater acoustics. Another approach is to use an approximation for the exact solution of (6). r) = 0. 0). i. 2π m → ∞. the density by ρ(z. r) reads: 2 2ik0 ψr (z. We want to compare this result concerning the decay rate with the one derived by Mickens [17]. r) + iα(z. r) emerging from a time–harmonic point source located in the water at (zs . and the (horizontal) sea bottom at z = zb . a one–way wave equation. r > 0 denotes the radial range variable and 0 < z < zb the depth variable (assuming a cylindrical geometry). d = 0: e2 ym = Jm+2 (2) ≈ √ em (m + 2)−(m+5/2) . The complex refractive index is given by N (z. it follows that e m−5/2 em Ym < √ .if (2+d)/c is a positive real number. (20) This enables us to give another asymptotic form of the recessive solution to the discrete model equation (6) with c = 1. r). Here we concentrate on the application to underwater acoustics. d = 0 in (18). r) + k0 (N 2 (z. our new bound Ym decays faster. (21) and with Stirling’s formula this simpliﬁes to ym ≈ 1/(m + 2)!. 2π ν → ∞. The SPE in cylindrical coordinates (z. (22) 7 . r) = c0 /c(z. r) − 1) ψ(z. r). r)/k0 with a reference sound speed c0 and the reference wave number k0 = 2πf /c0 . A standard task in oceanography is to calculate the acoustic pressure p(z. i.e. arising for example in (underwater) acoustics and radiowave propagation problems.e. We denote the local sound speed by c(z. Here. [11]): 1 Jν (z) ≈ √ eν+ν log(z/2)−(ν+1/2) log ν . m → ∞. r) + ρ ∂z (ρ−1 ∂z )ψ(z. and the attenuation by α(z. 3 Application to Parabolic Equation Calculations With the results from the previous Section we want to derive a discrete transparent boundary condition (DTBC) for the so–called standard “parabolic equation” (SPE) [20].
This TBC is not matched to the behaviour of the refractive index and spurious reﬂections will occur. All other physical parameters are assumed to be constant in the bottom. the physical problem is posed on the unbounded z–interval (0. r) is the water density just above the bottom and ρb denotes the constant density of the bottom. energy is returned from the bottom. Here. [2]) was derived for a homogeneous medium (i. In this work we are especially interested in the case of a linear squared refractive index in the bottom region.where ψ denotes the (complex valued) outgoing acoustic ﬁeld ψ(z. r) = k0 r p(z. Instead we will derive a TBC that matches the squared refractive index gradient at z = zb . all physical parameters are constant for z > zb ). This Schr¨dinger equation (22) is an o evolution equation in r and a reasonable description of waves with a propagation direction within about 15◦ of the horizontal. Here. (23) in the far ﬁeld approximation (k0 r 1). We assume ρ = ρ(z) and a simple calculation 8 . ρw ρb (24a) (24b) where ρw = ρ(zb −. ∞) and one wishes to restrict the computational domain in the z–direction by introducing an artiﬁcial boundary at the water–bottom interface (z = zb ). However. For most underwater acoustics (and also radiowave propagation) problems the squared refractive index in the exterior domain increases with z. We denote the physical parameters in the bottom with the subscript b and assume that the squared refractive index Nb below z = zb can be written as 2 Nb (z. (25) with real parameters β and µ = 0. ψz (zb+. r) e−ik0 r . i. i. one requires continuity of the pressure and the normal particle velocity: ψ(zb −. z > zb .g. One can easily derive an estimate for the L2 –decay of solutions to the SPE. the usual transparent boundary condition (see e. At the water surface one usually employs a Dirichlet (“pressure release”) BC: ψ(0. r) = 0.e. posed on the halfspace z > 0. no attenuation in the bottom: αb = 0. r) = 1 + β + µ(z − zb). Since the density is typically discontinuous at the water–bottom interface (z = zb ). r).e. the slope µ > 0 corresponds to a downwardrefracting bottom (energy loss) and µ < 0 represents the upwardrefracting case.e. where the wave propagation in water has to be coupled to the wave propagation in the the bottom. r) = . r) ψz (zb −. r) = ψ(zb+.
0). s) = 0.g. in the dissipation–free case (α ≡ 0) ψ(. is supported in the computational domain 0 < z < zb . Setting σ 3 = −µk0 and τ = 2ik0 /σ 2 (28) can be written as ˆ ˆ ψzz (z. then a Laplace transformation in range of (22) for z > zb yields: 2 ˆ ˆ ψzz (z. (29) ˆ Introducing the change of variables ζs (z) = σ(z − zb ) + τ s.e. µ > 0. z > zb . In the following we will review from [14] the derivation of the transparent boundary condition at z = zb for the SPE with a linear squared refractive index.. z > zb . s). (27) i. for ﬁxed s. in [1].. Now the basic idea of the derivation is to explicitly solve the ˜ 2 equation in the exterior domain z > zb . r) 2 ∞ = −2c0 ∞ (26) ψ(. c 0 2 for the weighted L –norm (“acoustic energy”) ∂r ψ(. ˜ we can write (29) in the form of an Airy equation: U (ζs (z)) + ζs (z) U (ζs (z)) = 0. Re s > 0 is ˆ ψ(z. which models a point source located at (zs . s) = C1(s) Ai(ζs (z)). r) is conserved and for α > 0 it decays. ˜ z > zb . 0 (−µk 2 1/3 0 ) . 0). r)2 ρ−1 (z) dz.(cf. [2]) gives α ψ2 ρ−1 dz. s) + [µk0 (z − zb ) + 2ik0 s]ψ(z. U (ζs (z)) = ψ(z. (28) with zb = zb −β/µ. µ < 0. e. r) 2 = 0 ψ(z. (31) if we deﬁne the physically relevant branch of σ to be σ= (µk 2 )1/3e−iπ/3 .. A transparent BC for the SPE (or Schr¨dinger equation) for a constant exterior medium was derived by several o authors from various application ﬁelds. (32) 9 . ˜ z > zb . s) = 0. If we further assume that the initial data ψ I = ψ(z. 4 Transparent Boundary Conditions Transparent Boundary Conditions. (30) The solution of (30) which decays for z → ∞. s) + σ 2 σ(z − zb ) + τ s ψ(z.
(38) 10 . Numerical Implementation. s) = ρb i. rn = nk with h = ∆z.Elimination of C1(s) gives Ai(ζs (z)) ˆ ˆ ψ(z.e. rn ). The discretized SPE (22) then reads: −iR(ψj (n+1) The kernel gµ is obtained by an inverse Laplace transformation of W (s) (cf. ψz (zb −. In the following we review the discretization from [15] and start with the discretization in range: ρw ψz (zb . The constant term in gµ acts like a Dirac function and the inﬁnite series represents the continuous part. z yields with the matching conditions (24) the transformed transparent BC at z = zb : ρw ˆ ˆ s ψ(zb −. the transparent BC at z = zb reads: ψz (zb . As Levy noted in [15] the kernel gµ decays extremely fast for µ > 0 and for negative µ it decays slowly at short ranges and then oscillates. This BC is nonlocal in the range variable r and can easily be discretized. s Ai(ζs (zb )) (34) ψr (zb . (n) (n+1) +ψj )+w (N 2 )j −1 (ψj (n) (n) (n+1) (n) Discretization of the continuous TBC. r ) gµ (r − r ) dr . s) = ψ(zb +.g. (33) Finally. r ) is constant on each subinterval rn < r < rn+1 and integrate the kernel gµ exactly. (36) Ai(ζ0 (zb )) j=1 aj − ζ0 (zb ) −ψj ) = ρj ∆0 (ρ−1 ∆0 )(ψj z j z with ∆0 ψj z (n) +ψj ). r) = ρw ρb r 0 W (s) = σ Ai (ζs (zb )) . (37) (n) (n) 2 2 2 = ψj+1/2 − ψj−1/2. Ai(ζs (zb )) z > zb . ρj ≈ ρ(zj ).r. rn ) = ρb n−1 m=0 ψb (n−m) − ψb k (n−m−1) Gm . k = ∆r and the approximations (n) ψj ≈ ψ(zj . the mesh ratio R = 4k0 h /k and w = k0 h . (35) where ζ0 (zb ) = σβ/µ and the (aj ) are the zeros of the Airy function Ai which are all located on the negative real axis [19]. s) W (s). To incorporate the TBC (35) in a ﬁnite diﬀerence scheme we make the approximation that ψr (zb . e. s) . [7]): ∞ −(aj −ζ0 (zb ))r/τ Ai (ζ (z )) e 0 b gµ (r) = σ + .t. Now we shall discuss how to solve (22) numerically with a CrankNicolson ﬁnite diﬀerence scheme which is of second order (in ∆z and ∆r) and unconditionally stable. in conjunction with a ﬁnite diﬀerence scheme for (22). We will use the uniform grid zj = jh. diﬀerentiation w.
where the water–bottom interface lies between the grid points j = J − 1 and J : (m) ψb ˜(m) ˜(m) ψJ + ψJ−1 . rn ) = −ψb Gn + ψb G0 + ρw m=1 (40) ˜(n) In [15] Levy used an oﬀset grid in depth. zj = (j + 1 )h. . Moreover. ˜ z 2 j = −1. rn ) ≈ . . 2 k ρb (n) (n) n−1 m=1 ˜ b m (ψJ (n−m) ˜ + ψJ−1 ). In a numerical implementation one has to limit the summation in (36) and therefore the TBC is not fully transparent any more. the stability of the resulting scheme is not clear since the discretized TBC (42) is not matched to the ﬁnite diﬀerence scheme (37) in the interior domain. = ψ(zb. Since aj ∼ − 3π (4j −1) 8 for j → ∞ the series (39) deﬁning Gm has good convergence properties for positive range r but for r = 0 the convergence is very slow. (n−m) (42) bm = 1 h ρw (Gm − Gm−1 ). Instead of using an ad–hoc discretization of the analytic transparent BC like in [15] we will construct discrete TBCs of the fully discretized half–space problem with the help of the results from Section 2. rn ). ψj ≈ ψ(˜j . i. To derive the discrete TBC we will now mimic the derivation of the analytic TBC from Section 2 on a 11 . h (41) ˜(0) ˜(0) This yields ﬁnally (recall that ψJ = ψJ−1 = 0) the following discretized TBC for the SPE: ˜ ˜ (1 − b0)ψJ − (1 + b0)ψJ−1 = with b0 = 1 h ρw G0 . J . . To overcome this numerical problem we use the identity 1 = 2 j=1 (aj − ζ0 (zb )) ∞ Ai (ζ0 (zb )) Ai(ζ0 (zb )) 2 − ζ0 (zb ). (44) which can be derived analogously to the one in [14].e.where we have set ψb Gm = rm+1 rm (n) = ψ(zb . (39) r=rm This leads after rearranging to k n−1 ρb (n−m) (0) (n) ψb (Gm − Gm−1 ) ψz (zb. Discrete transparent boundary conditions. rn ) and Gm is given by Ai (ζ0 (zb)) 2ik0 + Ai(ζ0 (zb )) σ e−(aj −ζ0 (zb ))r/τ 2 j=1 (aj − ζ0 (zb )) ∞ r=rm+1 gµ (η) dη = kσ . 2 k ρb (43) 2/3 Note that the constant term in (39) enters only b0 . rm ) ≈ 2 ˜(n) ˜(n) ψJ − ψJ−1 ψz (zb . .
ˆ 0 n ∈ Z0 . (54) 12 . (48) 2 z+1 2 Comparing (47) with the recurrence relation of the Bessel function Jν (σ) yields the condition k2 ν ! j + oﬀset = 1 − iζ(z) − µ 0 h3 (j − J ) = . σ=− µ (51) Jν (z) (σ) Jσ(1−iζ(z))−1 (σ) ˆ ˆ ψJ−1 (z) = gµ. τ > 0. (45) (n) j ≥ J − 1.J (z) = J−1 ˆ ˆ = .discrete level. zb = J h and solve the discrete exterior problem −iR(ψj (n+1) −ψj ) = ∆2 ψj (n) (n) (n) (n+1) +∆2 ψj +w β+µh(j−J ) (ψj (n) (n) (n) (n+1) +ψj ). (47) Rz−1 β 2 − i k 0 h2 . an inverse Z–transformation yields the discrete TBC ψJ−1 − with (n) J (n) (0) (n) J ψJ n−1 = m=1 ψJ (n−m) (m) J . with 2 k0 3 −1 ∈ R. ∆2 ψj = ψj+1 − 2 ψj + ψj−1 . JνJ (z) (σ) Jσ(1−iζ(z)) (σ) (52) Finally. for j ≥ J . (46) Hence. z > 1. σ 2 σ and we conclude that the exact solution of (47) is ˆ ψj (z) = Jνj (z) (σ). In analogy to the continuous problem we assume for the initial (n) 0 data ψj = 0.J (z)ψJ (z) with gµ. by using the Z–transformation: ∞ (n) ˆ Z{ψj } = ψj (z) := ψj z −n . h 2 From (50) we obtain the transformed discrete TBC at zb = J h: (49) (50) ν = νj (z) = σ(1 − iζ(z)) + j − J. (53) =Z −1 τn {ˆµ. is a discrete Airy equation k ˆ ˆ ˆ ψj+1 (z) − 2 1 − iζ(z) − µ 0 h3 (j − J ) ψj (z) + ψj−1 (z) = 0. the Z–transformed ﬁnite diﬀerence scheme (37). 2 with ζ(z) = 2 j ≥ J − 1. n=0 (n) z ∈ C.J (τ eiϕ)einϕ dϕ.J (z)} = g 2π 2π gµ. j ≥ J − 1 and use the linear potential term (N 2 )j − 1 = β + µh(j − J ).
(56) In physical space. Note that the Bessel functions in (52) with complex order and (possibly large) real argument can be evaluated by special software packages (see e. (58) c = 2σ −1 .J (z) = σ −1 νJ (νJ − 1) h e νJ − 1 with νJ . For a second approach to obtain an approximated DTBC.Since this inverse Ztransformation cannot be done explicitly.e. c and thus we obtain an approximation to the transformed DTBC of the form ˆ ˆ ˆ ψJ−1 (z) = kµ. (1) Alternatively.J (z)ψJ (z). we use a numerical inversion technique based on FFT (cf. (57) In the following. 13 (59a) . one can use the asymptotic solution ym from (9a). J (n) (n) (n) z +1ˆ J (z). i. The Z– transformed scheme in the exterior j ≥ J − 1. We introduce the summed coeﬃcients sJ = Z −1 {ˆJ (z)} . is a discrete Airy equation of the form (6) with νJ . [21]). (n) (n) J (n−1) . σ given by (51). given by (47). Using the asymptotic formula (20) leads to the approximate DTBC 2 νJ ˆ ˆ ˆ ψJ−1 (z) = ˆ µ. we will present two alternative approaches to obtain a DTBC: asymptotic expansions and a continued fraction formula. with sJ (z) := s ˆ which satisfy sJ = (0) (0) J . for details of this routine we refer the reader to [8]. [9]).g. Since the magnitude of J does not decay as n → ∞ (Im J behaves like const ·(−1)n for large n). z (55) sJ = + n ≥ 1. one can use asymptotic expansions. (n−1) n ≥ 1. d = −2iζ(z) − cJ.J (z)ψJ (z) with hµ. [9]). θ = 2 + 2+d = 2 + ν0 . Asymptotic Expansions. it is more convenient to use a modiﬁed formulation of the DTBCs (cf. the DTBC is: ψJ−1 − sJ ψJ = (n) (0) (n) n−1 m=1 sJ (n−m) ψJ (m) − ψJ−1 .
with ˆ kµ. the last quotient may be neglected.6] we can easily deduce an expression for the quotient of Bessel functions (like (52)) as a continued fraction from the recurrence formula (15)... Following [22. with the help of the theory of Lommel polynomials [22. that when M → ∞. (59b) Continued fraction formula. For the numerical implementation we use the modiﬁed Lentz’s method [13] which is an eﬃcient general method for evaluating continued fractions. −1 − 2(ν + 2)z −1 − . 5 Approximation by Sums of Exponentials An adhoc implementation of the discrete convolution (57) with convolution (n) coeﬃcients sJ from (55) (or obtained by any of the above approaches) has still one disadvantage. Section 5. so that Jν−1 (z) 1 1 = 2νz −1 − .. Finally.. Section 9. Jν (z) 2(ν + 1)z (60) This continued fractions formula oﬀers another way to evaluate the quotient of two Bessel functions needed in the transformed discrete TBC (52). we use a continued fraction formulation. Remark 4 For brevity of the presentation.65]. Jν (z) 2(ν + 1)z −1 − . The boundary condition is non–local and therefore computationally expensive. Remark 3 Our practical calculations showed that the evaluation of the continued fraction (60) is stable for all considered values of ν and z although we cannot prove this yet. for a third approach for an approximation to the DTBC. − 2(ν + M )z −1 − Jν+M (z) This holds for general values of ν and it can be shown. If we rewrite (15) as 1 Jν−1 (z) = 2νz −1 − Jν (z) . Jν (z) Jν+1 (z) it is obvious that 1 Jν−1 (z) 1 Jν+M −1 (z) = 2νz −1 − . In fact.J (z) = (1) yJ−1 (1) yJ = c(J − 2) J J −1 cs θ s=1 (J−1)s ∞ (1) cs Js s=1 ∞ (1) . the evaluation of (57) is as expensive as 14 . we omit here the discussion of an adequate discrete treatment of the typical density shock at z = z b and refer the reader to [2] for a detailed discussion of various strategies.
ql } for ﬁxed L. ql} also for n > 2L + 1? In other words. 2L + 1. then its Taylor series ˜ ˜ g (x) = sJ + sJ x + sJ x2 + .. . Let us ﬁx L and consider the formal power series: g(x) := sJ + sJ x + sJ x2 + . . (67) It follows from (65) and (66) that the set {bl. l=1 where L ∈ N is a ﬁxed number. 2L + 1. . the approximation properties of (n) sJ depend on L. . (n) (n) (2) (3) (4) (2) (3) (4) x ≤ 1. . (62) PL−1 (x) QL(x) (63) (64) n = 2. what quality of approximation (n) (n) sJ ≈ sJ . we approximate the coeﬃcients sJ by the following (sum of exponentials): (n) s . 1 J (n) (n) L sJ ≈ sJ := ˜ (61) bl ql−n . 3. (65) due to the deﬁnition of the Pad´ approximation rule. . . . ql } deﬁned in Theorem 5 can be used in (61) at least for n = 2. n = 2. n = 2. 1. 3. . The main question now is: Is it possible to use these {bl . . As a remedy. e Theorem 5 ([3]) Let QL (x) have L simple roots ql with ql  > 1. Below we propose a ˜ deterministic method of ﬁnding {bl. . QL(ql ) l l = 1. 3. . . ˜ ˜ satisﬁes the conditions ˜ sJ = sJ . If there exists the [L − 1L] Pad´ approximation e g (x) := ˜ of (62). L. n = 0. . . Evidently. Then ˜ sJ = l=1 (n) L l = (66) bl ql−n . ˜ n > 2L + 1 (68) 15 . In order to derive a fast numerical method to calculate the discrete convo(n) lutions in (57). In the sequel we will brieﬂy review this approach. and the corresponding set {bl. ql }. . where bl := − PL−1 (ql ) q = 0. . . . we proposed in [3] the sumofexponentials ansatz (for a comparison of the computational eﬀorts see Fig. . .for the discretized TBC (42). 6). . L. .. . . 3. .
6 in Section 7). L. . Finally we summarize the approach by the following algorithm: 1. m = 1. + bl ql−1 ψJ (n−1) = ql−1 Cl (n−1) . via numerical inverse Ztransformation. n = 0. 2. [3]) yields: The value C (n) from (69) for n ≥ 2 is represented by C (n) = where Cl Cl (n) (1) L Cl . A typical graph (n) (n) of sJ − sJ  versus n for L = 27 is shown in Fig. . ql are used for the eﬃcient calculation of the discrete convolutions. . Let us consider the approximation (61) of the discrete convolution kernel appearing in the DTBC (57). [6]). If such problems still occur or if one root of the denominator is smaller than 1 in absolute value. 3. and this will reduce the ˜ numerical eﬀort signiﬁcantly (cf. . calculate sJ . . Fig. the orders of the numerator and denominator polynomials are successively reduced. Remark 6 We note that the Pad´ approximation must be performed with high e precision (2L − 1 digits mantissa length) to avoid a ‘nearly breakdown’ by ill conditioned steps in the Lanczos algorithm (cf.can we expect? The above analysis permits us to give the following description of the approx(n) imation to the convolution coeﬃcients sJ by the representation (61) if we use a [L − 1L] Pad´ approximant to (62): the ﬁrst 2L coeﬃcients are reproe (n) (n) ˜ duced exactly. (71) n = 2. the corresponding coeﬃcients bl . (n) 16 . . With these “exponential” coeﬃcients the convolution C (n) := n−1 m=1 (m) ˜ sJ (n−m) ψJ . ˜ Fast Evaluation of the Discrete Convolution. . calculate sJ via Pad´–algorithm. . l = 1. with the kernel coeﬃ(n) cients sJ . . (m) sJ = ˜ l=1 (n) L bl ql−n . l=1 (n) (70) ≡ 0. (n) 2. ˜ e 3. can be calculated by recurrence formulas. of a discrete function ψJ . . . the asymptotic behaviour of sJ and sJ (as n → ∞) diﬀers strongly (algebraic versus exponential decay). N − 1. (69) ql > 1. . A straightforward calculation (cf. . . see (65). however. . 2 in Section 7.
. ψ0 (n) (0) = ψ I (zj ). j (n) (73) which is the discrete analogue to (27). J − 1. The proof is based on a discrete energy estimate for the new variable φj (n) := ψj γ −n .J (γeiϕ ) ≤ 0.6 Stability Analysis of the numerical scheme Here we analyze the stability of our numerical scheme for the SPE (37) along with a surface condition and the the DTBC (53): −iR(ψ (n+1) −ψ (n)) j j (0) ψ j = ρj ∆0 (ρ−1 ∆0 )(ψj z j z (n+1) +ψj )+w (N 2 )j −1 (ψj j = 1. (72) ˆ ˆ ψJ−1 (z) = gµ. The following theorem bounds the exponential growth of solutions to the numerical scheme for a ﬁxed discretization. j = 0.e. n ∈ N. 1. 17 (n) . the system is dissipative). Then. (n) (n) (n+1) +ψj ) (n) with ψJ−1 = ψJ = 0. J − 1. ˆ ∀ 0 ≤ ϕ ≤ 2π. .J satisfy ˆ Im gµ. . the solution of (72) satisﬁes ˆ the apriori estimate ψ (n) 2 ≤ ψ0 2 γn. Theorem 7 (Growth condition) Let the boundary kernel gµ. (74) for some (suﬃciently large) γ ≥ 1 (i. (75) PROOF.J (z)ψJ (z). . . .J (z) is given by (52). J . . Assume also that gµ. (0) = 0. 2. ˆ In the sequel we want to derive an apriori estimate of the discrete solution in the discrete weighted 2 –norm: ψ (n) 2 2 J−1 := h j=1 ψj 2 ρ−1 . ˆ where gµ. .J (z) is analytic for z ≥ γ.
.which satisﬁes the equation −iR(φj (n+1) − φj ) = ρj ∆0 (ρ−1 ∆0 ) + w (N 2 )j −1 z j z j = 1. . . .J (γz) − 1) φJ (z). the bottom BC can be written as ∆− φJ = −φJ ∗ (n) (n) ˜(n) J γn n =− φJ m=0 (m) ˜(n−m) γ m−n . . . . (0) (n) (n) (φj (n+1) + φj ) (76a) (n+1) (n) + (γ − 1) ρj ∆0 (ρ−1 ∆0 ) + w (N 2 )j −1 + iR φj z j z j = 0. sum from j = 1 to j = J − 1. J − 1. J (77) (n+1) −1 ¯(n) First we multiply (76a) by φj ρ−1 /γ and its complex conjugate by φj ρj : j (n) ¯(n) (n+1) ρ−1 = iR φj 2 − φj φj j ¯(n) ∆0 (ρ−1 ∆0 ) + w (N 2 )(n) −1 ρ−1 φj j z j z j φj (n+1) + φj (n) (n) ¯ +(γ −1 − 1)φj (n) ∆0 (ρ−1 ∆0 ) + w (N 2 )j −1 − iR ρ−1 φj . (n) . ˆ = −(ˆµ. J. ∆0 (ρ−1 ∆0 ) + w (N 2 )j −1 − iR ρ−1 φj j z j z (78b) Next we subtract (78a) from (78b). j z z j (78a) (n) iR φj (n+1) 2 ¯(n) (n+1) ρ−1 =  − φj φj j φj (n+1) −1 ∆0 (ρj ∆0 ) + w (N 2 )j −1 ρ−1 z z j (n+1) (n) ¯(n+1) + φ(n) ¯ φj j +(γ − 1)φj (n) ¯(n+1) . (n) φ0 (0) (76b) (76c) (76d) = 0. and apply 18 . φj = ψ j . . g ˆ ∆− φJ (z) In physical space.
φJ + γ φJ J J (80) Summing (80) from n = 0 to n = N yields (note that γ ≥ 1): φ(N +1) 2 2 ≤ φ(0) − 2 2 − wh N J−1 (n) (n) (n+1) 2 −1 Im (N 2 )j −1 φj + γφj  ρj 2R γ n=0 j=1 N h Im 2 Rρ γ J−1/2 2 2 ¯(n) ¯(n+1) )∆−(φ(n) + γφ(n+1) ) ( φJ + γ φJ J J αj (n+1/2) = φ(0) − kh 2γ 2 n=0 N J−1 c0 (n+1/2) cj n=0 j=1 N n=0 φj + γφj (n) (n+1) 2 −1  ρj k Im + 2 4γ k0 hρJ−1/2 ¯(n) ¯(n+1) ) (φ(n) + γφ(n+1) ) ∗ ( φJ + γ φJ J J ˜(n) J γn .summation by parts: J−1 iR j=1 (n+1) 2 φj  − (n) φj 2 J−1 ρ−1 j + =− ¯(n) φJ j=1 ∆− φj (n+1) 2  − ∆− φj 2 ρ−1 j−1/2 (n) + (n+1) φJ−1 ∆− J−1 ¯(n+1) φJ (n) (n+1) (n) ¯(n) ρ−1 − φJ−1 ∆− φJ + φJ J−1/2 +w j=1 J−1 (N 2 )j −1 (φj (n) (n+1) 2 ¯(n) (n+1) ) − (N 2 )(n) −1 (φ(n) 2 + φ(n) φ(n+1) ) ρ−1 ¯  + φj φj j j j j j (n+1) 2 +w j=1 J−1 (γ − 1) (N 2 )j −1 φj (γ − 1)φj J−1 (n+1) 2  + (1 − γ −1 ) (N 2 )j −1 φj 2 ρ−1 j (n) (n) (n) − iR j=1  + (1 − γ −1 )φj 2 ρ−1 j J−1 − (γ − 1) j=1 ∆− φj (n+1) 2 −1  ρj (n+1) − (1 − γ −1 ) j=1 ∆−φj 2 ρ−1 j (n) ¯ + (γ − 1)φJ−1 ∆− φJ (n+1) (n) ¯(n) + (1 − γ −1 )φJ−1 ∆−φJ ρ−1 J−1/2 (79) Now. (81) 19 . taking imaginary parts one obtains after a lengthy calculation: J−1 j=1 φj (n+1) 2  − φj 2 ρ−1 = −(γ − 1) j − − w γR J−1 j=1 (n) (n) J−1 j=1 (n) φj (n+1) 2 −1  ρj J−1 − (1 − γ −1 ) j=1 φj 2ρ−1 j (n) Im (N 2 )j −1 φj + γφj (n+1) 2 −1  ρj 1 Im γRρJ−1/2 ¯(n) ¯(n+1) ∆− φ(n) + γφ(n+1) .
since it is a ﬁnite ˆ sum. 20 . Using Plancherel’s Theorem for Z–transforms we have N ∞ v n=0 (n) (n) u ¯ = n=0 v (n) u(n) = ¯ 1 2π 1 = 2π 2π 0 2π 0 u v(eiϕ )ˆ(eiϕ) dϕ ˆ (82) ˆ ˆ(e ) gµ. Since αj ≥ 0. . N. Since the average of gµ. condition (74) implies Im gµ. ∀ N ≥ 0. .J (z) on the circles z = βeiϕ equals ˆ (0) gµ.J (γeiϕ ) − 1 dϕ. u iϕ 2 iϕ Using (82) for the boundary term in (81) now gives: φ(N +1) 2 2 ≤ φ(0) + 2 2 2π 0 h 2ρ 2πRγ J−1/2 ˆ (1 + γeiϕ )φJ (eiϕ )2 Im gµ. n = 0. it also holds for all β > β0. n > N. The Z–transform Z{u(n) } = u(z) is analytic for z > 0.J (γz) − 1)ˆ(z) and is ˆ g u analytic for z ≥ 1. To this end we deﬁne (for N ﬁxed) the two sequences. Then ˆ ˆ we have the following simple consequence of the maximum principle for the Laplace equation: If condition (74) holds for some β0. u(n) := φ(n) + γφ(n+1) .J (z = ∞) ≤ 0. it remains to determine the sign of the last term in (81) to ﬁnish the proof. u γ n−m n ∈ N0 .J (γe ) − 1 dϕ. Hence its imaginary parts is a harmonic ˆ functions there.For the last identity we used the bottom BC (77) and φ0 = φJ = 0.J (z = ∞).J (z) is analytic for z ≥ β. Remark 8 Above we have assumed that the Ztransformed boundary kernel gµ. . ˆ (83) Our assumption on gµ. and the result of the theorem follows.J = gµ. v (n) := u (n) ∗ ˜(n) J γn n = m=0 ˜(n−m) (m) J . .J therefore implies ˆ φ(N ) 2 ≤ φ(0) 2 . The Z–transform Z{v (n)} then satisﬁes v(z) = (ˆµ. J J (n+1/2) (0) (0) 0.
We computed a reference solution on a three times larger computational domain conﬁned with the DTBC from [2]. we decided to evaluate the ratio of the two Bessel functions in (52) by the continued fraction formula (60) together with the sumofexponentials ansatz (61). [16] and [25].5 m. We consider a rangeindependent situation for 0 < r < 50 km. where the acoustic pressure p is calculated from (23).e.04 with 210 sampling points for the numerical inverse Z–transformation (54). i.5 m) = 1539. For this choice of parameters the mesh ratio becomes R ≈ 0. First we examine the convolution coeﬃcients of the two presented approaches. Section 2].24 m s−1 .32.12246 and the parameter σ ≈ −53345.44 m is emitting sound with a frequency f = 300 Hz and the receiver is located at the depth zr = 27. We note that all approaches fulﬁlled for moderate choices of νJ the growth condition (74) needed for stability.100))..e. that is. ∆z = 0. Alternatively.7 Numerical Example In the example of this Section we will consider the SPE for comparing the numerical result from using our new (approximated) discrete TBC to the solution using the discretized TBC of Levy [14]. We computed the ﬁrst 1000 terms in the expansion (60) and used a radius τ = 1.J (z) ∼ 2(1 − iζ(z)) which is only the ﬁrst term in the continued h fraction expansion (60).5 m s−1 to c(152. 1 shows a comparison of the coeﬃcients bn from the discretized TBC (42) 21 . The source at the depth zs = 91. The TBC is applied at zb = 152. Fig. Below we present the socalled transmission loss −10 log 10 p2. using precomputed values from the call evalf(AiryAiZeros(1. The sound speed varies linearly from c(0 m) = 1536. using the asymptotic formula (58) is not advisable since for large νJ we have ˆ µ. energy loss to the bottom): µ = 2 · 10−4 m−1 . As an illustrating example we chose the typical downward refracting case (i. In order to calculate the convolution coeﬃcients bn (discretized TBC of Levy) we used the MATLAB routine from [24] to compute the ﬁrst 100 zeros of the Airy function. It contains no attenuation: α = 0. the value of νJ deﬁned in (51) is much too large for the routines like COULCC [21] for evaluating Bessel functions. 5000 range steps. For a discussion of that topic we refer the reader to [3. On the other hand. Example. The choice of an appropriate radius τ is a delicate problem: it may not be too close to the convergence radius of (60) due to the approximation error and τ too large raises problems with rounding errors during the rescaling process. Therefore. in MAPLE with high precision yielded indistinguishable results.5 m. We used the environmental test data from [7] and the Gaussian beam from [12] as starting ﬁeld ψ I .5 m and the discretization parameters are given by ∆r = 10 m. The reference sound speed c0 is chosen to be equal to c(zb ) such that β = 0 in (25).
J (β e )} = −0. In Fig. 3). The coeﬃcients convolution coefficients b .01 (for this discretization). 5 does not change if we compute more zeros of the Airy function. Therefore we extended the range interval up to 250 km and shall now illustrate the diﬀerence in the computational ef22 (n) . we obtain max{Im (ˆµ. Now we investigate the stability of the approximated discrete TBC and check the growth condition (74). ~ (n) s n J (n) 10 0 approximated discrete TBC discretized TBC of Levy 10 −5 10 −10 10 −15 10 −20 0 500 1000 1500 2000 2500 n 3000 3500 4000 4500 5000 Fig. For β = 1 we have max{Im (ˆµ. 2 we plot both the exact (n) (n) (n) ˜ convolution coeﬃcients sJ and the error sJ − sJ  versus n for L = 27 (observe the diﬀerent scales). Comparison of the convolution coeﬃcients bn of the discretized TBC (42) (n) and sJ from the approximated discrete TBC (with L = 27). The result in Fig.002 (see Fig. ˜ bn decay even faster than the coeﬃcients sJ . g This means that the Z–transformed kernel gµ.J (β eiϕ ) of the approximated ˆ discrete TBC satisﬁes the stability condition (74) for β ≥ 1.with the coeﬃcients sJ from the approximated discrete TBC. Evaluating the convolution appearing in the discretized TBC (42) is quite expensive for longrange calculations.153 g iϕ and. In Fig. with β = 1.J (β eiϕ)} = 0. The transmission loss curve of the solution using the approximated DTBC is indistinguishable from the one of the reference solution while the solution with the discretized TBC still deviates signiﬁcantly from it (and is more oscillatory) for the chosen discretization. 4 and Fig. 5 we compare the transmission loss results for the discretized TBC and the approximated discrete TBC in the range from 0 to 50 km.01. 1.
^ L=27: Im gµ.J(z) on circle..03 1 0 0 20 40 60 80 100 n 120 140 160 180 0 200 Fig. β=1. 3. On the other hand. Convolution coeﬃcients sJ (left axis. The line (. (L = 27).5 0 π/2 π ϕ 3/2π 2π Fig.) does not change consid23 .5 β=1 β = 1.J (z). ˆ fort for both approaches in Fig. z = β eiϕ (L = 27). the eﬀort for the approximated discrete TBC only increases linearly. Growth condition gµ. dashed line) and error sJ − sJ  of ˜ the convolution coeﬃcients (right axis).5 −1 −1.J (z) −0. 2.01 (n) (n) error s(n) − ~ (n) s 0. since the evaluation of the boundary convolutions dominates for large ranges.01 0.09 3 2 0.~ s  and error s(n) − s(n) J J J (n) x 10 4 −5 0. β=1.06 s(n) J J J (n) 0 ^ Im g µ . 6: The computational eﬀort for the discretized TBC is quadratic in range.
Discretized TBC of Levy 50 55 60 65 Transmission Loss [dB] 70 75 80 85 90 95 100 0 5 10 15 20 25 30 Range r [km] 35 40 45 50 Fig.5 m. erably for diﬀerent values of L since the evaluation of the sumofexponential convolutions has a negligible eﬀort compared to solving the PDE in the interior domain. Approximated Discrete TBC (L=27) 50 55 60 65 Transmission Loss [dB] 70 75 80 85 90 95 100 0 5 10 15 20 25 30 Range r [km] 35 40 45 50 Fig. Transmission loss at zr = 27. Transmission loss at zr = 27.5 m. 5. 24 . 4.
while the sumofexponential approximation to the discrete TBC has only linear eﬀort (. Conclusion We have proposed a variety of strategies to derive an approximation to the discrete TBC for the Schr¨dinger equation with a linear potential term in the o exterior domain. we have provided a simple criteria to check the stability of our method.). Moreover. Comparison of CPU times: the discretized TBC of Levy has quadratic eﬀort (—).discretized TBC vs. Acknowledgement The ﬁrst author thanks Prof. Arnold for many helpful suggestions to this article. 6. while discretized TBCs have usually quadratic eﬀort.. Our approach has two advantages over the standard approach of discretizing the continuous TBC: higher accuracy and eﬃciency. approximated discrete TBC (L=27). A. the sumofexponential approximation to discrete TBCs has only linear eﬀort. 25000 range steps 350 300 250 elapsed cputime 200 150 100 50 0 0 5000 10000 time steps 15000 20000 25000 Fig. 25 . The derivation was based on the knowledge of the exact solution (including the asymptotics) to the discrete Airy equation.
M. 66–72. Numerically Absorbing Boundary Conditions for Quantum Evolution Equations. Technische Universit¨t Berlin. Messenger 34 (1904). Sci. Discrete Artiﬁcial Boundary Conditions. 52–71. 313–319. Levy. Discrete transparent boundary conditions for the Schr¨dinger equation: Fast calculation. Arnold. (Academic Press. Riv. [5] P. 1927). 45 (1997). Non–local boundary conditions for acoustic PE predictions involving inhomogeneous layers. [13] W. Boston. MA. IEEE Trans. [15] M. Appl. On the homogeneous linear diﬀerence equation of the second order with linear coeﬃcients. Linear algebra. 305–423.J. Ehrhardt and I. Discrete Transparent Boundary Conditions for the Schr¨dinger Equation. Comp. Prentice–Hall.J. Univ. Selected Russian Publications in the Mathematical Sciences.W. VLSI Design 6 (1998). [11] N. [7] T. Thomson and G. 1991).D. J. [8] M. Generating Bessel Functions in Mie Scattering Calculations Using Continued Fractions. a [9] M. Transparent Boundary Conditions for Parabolic Equation Solutions of Radiowave Propagation Problems. Dawson. Ocean acoustic propagation by ﬁnite diﬀerence methods. Parabolic equation models for electromagnetic wave propagation. Arnold. 14 (1987). (IEE Electromagnetic Waves Series 45.H. Ehrhardt and A.W. Math. and o stability. Ehrhardt. [14] M. Phys. Cambridge. Kelley and A.T. Parma 6 (2001). Arnold and M. Peterson. 611–638. rational approximation and orthogonal polynomials. in 3rd European Conference on Underwater Acoustics. Antennas Propag. 668–671. North– Holland. [3] A. pp.F.References [1] A. Ph. 501–556. D. Comm. [2] A.N. Bultheel and M. approximation. 1996. Batchelder. Van Barel. 57–108. Thesis. 1 (2003). 2000). Levy. [4] E. Barnes. 1965. [6] A. 2001. 183–188. Opt.C. Ehrhardt. Lebedev. Diﬀerence equations: an introduction with applications. FORTH–IACM. 15 (1976).G. 145 (1998). McDaniel. Lee and S. o [10] W. Special Functions and their Applications.M. 1997). Mat. Discrete transparent boundary conditions for wide angle parabolic equations in underwater acoustics. An Introduction to Linear Diﬀerence Equations. Math.F. 26 . [12] D. Arnold. Sofronov. Brooke. Heraklion. Lentz. Appl. (Studies in Computational Mathematics 6. (Harvard University Press. Comput.
W..E. (Springer. Asymptotic solutions to a discrete Airy equation. Asymptotic properties of solutions to two discrete Airy equations.R. The parabolic approximation method. Ph. Zisowsky. eds. Wong and H. Lubich. Li. 413–425. 289–324. (Academic Press. Lecture Notes in Physics 70. Diﬀerence Equ. [20] F. New York. Phys. Discrete Transparent Boundary Conditions for Systems of Evolution Equations. 3 (1998). Comp. Jin. 231–239.edu/comp spec func/) [25] A. Appl. 2nd ed.N. in: Wave Propagation and Underwater Acoustics. Appl.J. [21] I.[16] C. 851–858. Diﬀerence Equ. Cambridge. 1977). Thompson and A.mit. 490–509. 46 (1986). Coulomb and Bessel Functions of Complex Arguments and Order. 224–287. Watson. Computation of special functions: with over 100 computer programs in FORTRAN. 1996). Zhang and J. New York. Numer. 7 (2001). 87 (1992). Olver. Convolution Quadrature and Discretized Operational Calculus. 1974). (MATLAB routines: http://ceta. J. a 27 . Asymptotic expansions for second–order linear diﬀerence equations. Papadakis. A treatise on the theory of Bessel functions. 1966).S. II. Barnett. John Wiley & Sons. [17] R. [24] S. [19] F.B. 52 (1988). Appl. Technische Universit¨t Berlin. Math.D.D.E. Mickens. [22] G. J. New York. Mickens. II. Keller and J. Asymptotics and Special Functions. Press. Stud. (Cambridge Univ. Tappert. Thesis. [18] R. J. Math. 2003.J. [23] R. J.
This action might not be possible to undo. Are you sure you want to continue?