The SLAB Geometry PN Equation

Department of Mathematics, ACE

June 18, 2014

Abstract
This paper explains the method to solve spherical harmonics mo-
ment systems, such as the time-dependent equations, of radiative
transfer. The method, which works for arbitrary moment order N,
makes use of the specific coupling between the moments in the PN
equations. This coupling naturally induces staggered grids in space
and time, which in turn give rise to a canonical, second order accurate
finite difference scheme. While the scheme does not possess TVD o
realizability limiters, its simplicity allows for a very efficient imple-
mentation in MATLAB. We present several test cases, some of which
demonstrate that the code solves poblems with ten million degees of
feedom in space angle, and time within a few seconds. The code
for the numerical scheme, called StaRMAP (Staggered grid Radiation
Moment Approximation), along with files for all presented test cases,
can be downloaded so that all results can be reproduced by the reader.

1 Introduction
The purpose of this paper is to present a simple and accurate solution method
for the PN equations of radiative transfer, and its efficient implementation in
MATLAB. The key idea is to make use of the specific coupling of unknowns
that is induce by the spherical harmonics being a family of orthogonal poly-
nomials. This leads to a natural staggered grid on which this equations are
discretized.

2 Derivation of Boltzmann Equations

A second order staggered grid method for spherical harmonies moment equa-
tions of radioactive transfer. The rate of change of the number of molecules
in the element is , where v̄ is the velocity vector with components (u,v,w) and
d¯(v) = dudvdw is the volume element in the velocity space; n is the number
of molecules and df (r) = dxdydz, r̄ is the location). The number density of
class v molecules within dr is nf (vdv).The number flux of molecules across
surface element dsr is equal to nf (vdv(v.er dsr )where dsr is the surface ele-
ment of dr and er is the unit vector normal to the surface and pointing out
of volume element dr. Total number flux through the surface dsr is equal to

Z
sr nf (v.er )dsr dv (1)
Z
=− dr 5 .(nf (v)d(drdv) = −.(v + nf 5 .v) (2)
= −.f (v) (3)

5.v = 0, (4)

is constant The net rate of change of number of molecules = −v. ∂r(nf∂)dvdr v

is constant since the spatial dependence of gas properties is sufficiently low
Now consider the two processes which are similar (i) Molecules moving with
velocity v leave physical space element dr. (ii)Molecules attain speed in the
range (v, v+dv) as a result of external force per unit mass. So we can write
he rate of change of molecules due to process (ii) as −F ∂ (nf )dvdr Process
(iii) Scattering of molecules in and out of dv due to collisions. As collisions
can be thought of as being instantaneous, collisions change only velocity, but
not the position of the molecules.
Hence, the change in the number of molecules dvdr space element happen
due to: Collisions evolving molecules with velocity in (v, v + dv) range.
Collisions that produce molecules with the velocity in (v, v + dv) range.
Consider the first type of collisions: v, v1 , v1
One molecule of class v moving with relative velocity vr w.r.t the molecules
of class v1 and scattering molecules into elementary solid angle d suffers
nf1 dvr dσ
d
r collisions per unit time. There are nf drdv such molecules in the
phase space element, so the total number of such collisions in the phase space
element per unit time is n2 f f i dσ
d
rddv1 dv.dr
Here we used the principle of molecular chaos to independently write
down the number of the molecules of classes vandv1 .
Similarly, formula can be written down for the collisions of the second
type: n2 f fi ∗ ( dσ
d
d)dv1 dvdr

2
 Because collisions of the first and the second kinds or symmetric |( dσd
d)dv1 dv| =
|( dσ
d
d)dv1 dv|The net rate of increase in the number of molecules of class v
in the phase space element dvdr due to the two types of collisions can be
written as: n2 (f f1 )vr dσ
d
ddv1 dvdr
To get the total rate of increase in the number of molecules of class v, we
need to integrate over all possibleRcollision partners of class v1 and over all
n (f f1 − f f1 )vr dσ
R 2
possible elementary solid angles d d
ddv1 dvdr
Combining
R R all2 three processes and cancel out dvdr we get ∂ (nf )+v. ∂ (nf )+
∂ dσ
F. (nf ) = n (f f1 − f f1 )vr d dd
The angular flux Ψ(x, t) describe the particle density at position x3 , di-
rection 2 at velocity v.
The time-dependent, isotrophically scattering linear Boltzmann Equa-
tions are

1∂ 
v4
(x, , t) + . 5 4(x, , t) + t (x, t)Ψ(x, , t) = 1
s(x,t)4π 2 Ψ(x,,t)d+ Ψπ Q(x,t)

Where

(1) temporal rate of change in Ψ

(2) leakage in drift temperature

(3) loss of particle due to scattering and absorption

(4) gain in particles

(5) additional particles

And s (x, t) → scattering cross section a (x, t) → absorption cross section

t (x, t) → s (x, t) + a (x, t) = total cross section Q(x, t) → internal source
and x3 , is the direction; 2 at velocity v
Equation (1) is first restricted to planar geometry
1∂ ∂ 
+ µ (x, µ, ϕ, t) + t (x, t)Ψ(x, µ, φ, t) = RR 1 (5)
vΨ Ψ s(x, t)Ψπ Ψ(x, µ, φ, t)dϕ + 4π
Q(x, t)
Integrating
R 1∂ w.r.t. φRfrom 0 to 2π;
t)dΨ+ µ ∂4 (x, µ, φ, t)dQ+ t (x, t)Ψ(x, µ, φ, t)dϕ, weget v1 ∂4 (x, µ, t)+
R
v4
(x, µ, φ,
µ ∂4 (x, µ, t) + t (x, t)Ψ(x, µ, t) = s(x,t)Qπ R 4(x,µ,t)dµ+

1
Q(x,t)
2
Multiplying by P n(µ) and integrating for −1 ≤ µ ≤ 1
to

3
 1 ∂
µ ∂ Ψ(x, µ, t)P ndµ +
R R R
vΨ
rt(x, µ, t)Pn dµ + t (x, t)Ψ(x, µ, t)P ndµ =

s(x,t)4n
RR 1
Ψ(x,µ,ϕ,t)P nd+ 4π Q(x,t)
(6)

R
Define the azimutually-integrated
R 1∂ angular flux Ψ(x, µ, t) := Ψ(x, µ, Q, t)dϕ

We obtain vΨ
(x, µ, t)P ndµ = s(x,t)2 R Ψ(x,µ,t)dµ+ 1
Q(x,t) 2
l l+1
Using the recursion relation for Pn (µ) = 2l+1 Pl−1 (µ)+ 2l+1 Pl+1 (µ) and us-
1∂ ∂ n+1
R
ing (x, t) = P n(µ)Ψ(x, µ, t)dµ,
n
R n ≥ 0, we have v (x, t)+ ( 2n+1 φn+1 (x, t)+
ϕ (x, t)) + t (x, t)ϕn = n,0 [s (x, t)ϕ0 (x, t) + Q(x, t)]
2n+1 n+1

3 The Slab geometry PN equations

Consider the radiative transfer equation given in the form ”Linear Transport
Theory”, K.M. caxetal

Z
∂t ψ(t, x, Ω)+Ω.∇x ψ(t, x, Ω)+Σt (t, x)ψ(t, x, Ω) = Σs (t, x, Ω, Ω0 )ψ(t, x, Ω0 )dΩ0 +q(t, x, Ω)
S2

where ψ = ψ((t > 0)x1 , x2 , x3 ) and Ω = (v1 , v2 ), ψ is the density of photons

that undergo scattering and absorption in a medium.
Let Σa be the absorption cross section and Sigmas be the scattering kernel.
Then Σt = Σso + Σa
Consider a slab between two infinite parallel plates. Let the Z-axis be per-
pendicular to the plates.If the setting is invariant under translations perpen-
dicular to and rotations around the Z-axis, then the unknown ψ depends
only on the Z-component of the spatial variable, and one angular variable
µ(= cos θ)
Expressing the angular dependence of the distribution function in terms of
a Fourier series,
2l + 1
ψ(t, z, µ) = Σ∞l=0 ψl (t, z) Pl (µ)
2

where
1 dl 2
Pl (µ) = l 2
(µ − 1)l ;
2 l! dµ
Legendre polynomial
These form an orthogonal basis of the space of polynomials w.r.t. the stan-
dard scalar product on [-1,1], when

4
where ψ is known, then ψl can be obtained as the Fourier Coefficient.

Z1
i.e.ψl = ψPl dµ
−1

Thus(1) =⇒

∞ 2l + 1 2l + 1 2l + 1
∂l [σl=0 ψl (t, z) Pl (µ)]+Ω.5x [Σ∞l=0 ψl (t, z) ψl (µ)]+Σt (t, x)[Sigma∞
l=0 ψl (t, z) Pl (µ)
2 2 2
Z
2l + 1
= Σs (t, x, Ω, Ω) Σ∞
l=0 ψl (t, z) Pl (µ) + q(t, x, Ω)
2
s2

Taking thelt h coefficient term on both sides, reducing to z,

Z1
∂t ψl + ∂z µPl ψdµ + σtl ψl = ql , l = 0, 1, ...
−1

Here
Z1 Z1
(2) → ∂t ψl + ∂z µPl ψdµ + [Σa + Σso − 2π Pl (µ)Σs (µ)dµ]ψl = ql
−1 −1

when l=2

Z1 Z1
∂t ψ2 + ψz µPl ψdµ + [Σa + Σso − 2π Pl Σs (µ)dµ]ψ2 = q2
−1 −1

Z1 Z1
1
= ∂l ψ2 + ∂z µ( (−1 + 3µ2 )dµ + [Σa + Σso − π (3µ2 − 1)Σs (µ)dµ]ψl = q2
2
−1 −1

using the reccurence relation,

i l+1
µPl (µ) = Pl−1 (µ) + Pl+1 (µ)
2l + 1 2l + 1

5
and truncating the expression at l =N,

Z1  
l l+1
(2) → ∂t ψl + ∂z Pl−1 (µ) + Pl+1 (µ) ψdµ + Σtl ψl = ql
2l + 1 2l + 1
−1

 1 
Z Z1
l l+1
i.e.∂t ψl + ∂t  Pl−1 (µ)ψdµ + Pl+1 (µ)ψdµ + Σtl ψl = ql
2l + 1 2l + 1
−1 −1

 
l l+1
i.e∂t ψl + ∂t ψl−1 + ψl+1 + Σtl ψl = ql → (3)
2l + 1 2l + 1

Now consider  
l+1 l
∂z ψl+1 + ψl−1
2l + 1 2l + 1

l = 0, the term is  
1
∂z ψ1 + 0 = ∂z (ψ1 )
1

l = 1, the term is  
2 1
∂z ψ2 + ψ0
3 3

l = 2, the term is  
3 2
∂z ψ3 + ψ1
5 5

l = 3, the term is  
4 3
∂z ψ4 + ψ2
7 7

consider the matrix

6
 
0 1 0 0 . . .
 1 0 2 0 . . . 
 3 2 3 3 
 0 0 5 0 . . 
 5 
 0 0 3 0 4
0 . 
 7 7 
 . . . . . . . 
 
 . . . . . . . 
. . . . . . .
Thus equation (3) can be represented as a matrix equation:

∂t ū + M.∂z ū + c.ū = q̄

whereū = ψl . 
0 1 0 . . .
 1 0 2 0 
M= 3 3 
 0 2 0 3 0 
5 5
0 0 73 0 47 0
 
Σt0 . . . . .
 . Σt1 . . . . 
 
 . . . Σt2 . . 
and C=  .

 . . . . . 

 . . . . . . 
. . . . . .
The PN approximations is based on the assumption is based on the as-
sumption that all ΨM l = 0, forl > N q
m m 2L+1 (l−m)! imφ m
The complex valued spherical harmonics are defined asYl = (−1) 4π (l+m)!
e Pl (µ), Ylm
form an orthogonal family on the unit sphere.
They satisfy
 recursion of the form 
m+1
−cm−1
l−1 Y m−1
l−1 + d m−1 m−1
l+1 Y l+1 + e m+1
l−1 Y l − 1 − f m+1 m+1
l+1 Y l
ΩYlm = 12  i(cm−1 m−1 m−1 m+1 m+1 m+1 m+1
 
l−1 − d l+1 Y l+1 + el−1 Y l − 1 − f l+1 Y l

m m m m
2(al−1 Yl−1 + bl+1 Yl=1 )
with coefficients
q q q
(l−m+1)(l+m+1) m (l−m)(l+m) l+m+1)(l+m+2)
aml = (2l+3)(2l+1)
, bl = ,
(2l+1)(2l−1) l
c m
=
q q q (2l+3)(2l+1)
(l−m)(l−m−1 m
dml = (2l+1)(2l−1)
, el = (l−m+1)(l−m+2)
(2l+3)(2l+1)
, flm = (l+m)(l+m−1)
(2l+1)(2l−1)
NowR consider the scattering term in 1
i.e. s2 Σs (t, x, Ω, Ω0 )Ψ(t, x, Ω0 )dΩ0
Multiplying this byYlm and integrating over Ω , we get
Y m (Ω) S 2 Σs (Ω.Ω0 )ψ(t, x, Ω0 )dΩ0 dΩ = Σsl Ψm
R R
S2 l R m l (t, x)
m
where Ψl (t, x) = s2 yl (Ω)ψ(t, x, Ω)dΩ

7
Hence the complex valued PN equations are
1
∂t ψlm + ∂x (−cm−1 m−1 m−1 m−1 m+1 m+1 m+1 m+1
l−1 ψl−1 + dl+1 ψl+1 + el−1 ψl−1 − fl+1 ψl+1 )
2

i
+ ψy (cm−1 m−1 m−1 m−1 m+1 m+1 m+1 m+1
l−1 ψl−1 + dl+1 ψl+1 + el−1 ψl−1 − fl+1 ψl+1 )
2

+∂z (am m m l m
l−1 ψl−1 + bl+1 ) + Σtl ψ1 = ql

for
0 ≤ l < ∞, −l ≤ m ≤ l

For a two dimensional domain D,consider the infinite cylinder D× ∈ R3 .

Let √ √
Ω = ( (1 − µ2 ) cos ϕ, (1 − µ2 ) sin ϕ, µ)T

We assume that all data are z-independent, then the solution ψ is z-independent
and additionally an even function is µ. If l+m is odd, the associated Legen-
dre polynomial Plm is an odd function in µ, and as a consequence positive
moments for which l+m is odd have to vanish. Thus the choice of l and m
are (0, 0), (1, −1), (1, 1), (2, −2), (2, 0), (2, 2) . . . thus we obtain the following
matrix formula
 
0
ψ 0
. d−1 −f11 .
 

ψ00

1 . .  ψ −1 
0 −2 0  1 
 ψ −1   e0 . . d2 −f2 0   ψ 1 

 1  1  . e−1
  1 
 1  0 . . .  −2 
∂t  ψ1  + ∂x  1   ψ2 
 ψ −2  2 .
 . . . . .   ψ0 

 2  −1 1 2 
..  . −c 1 e 1 . . .   ψ2 
. . 0 −c11 . .
 2 
..
.

8
  
ψ00
  −1 
. −d−1  ψ1 

1 −f11 . . .
0 2 0  ψ1 
 e0 . . −d2 −f2 0  1 
i −1
  ψ −2 
+∂y  . e1 0 . . .  2 

2  0 
 . e−1
1 e 1
1 . . .   ψ2 
 ψ2 
. 0 c11 . . .  2 
..
.
 
ψ00  
  −1
 ψ1  q00

Σt0

ψ11   −1
Σt1     q1
 
  −2 

 =  ψ2  =  q11

+ Σt1
 
  0   −2

  ψ2    q2

Σt2
 

Σt2
 ψ2  ..
 2  .
..
.
Hence the complex valued PN equations are
1 m−1 m−1 m−1 m−1 m+1 m+1 m+1 m+1
∂t Ψml + 2 ∂x (−cl−1 Ψl − 1m − 1+dl−1 +dl+1 Ψl+1 +el−1 Ψl − 1 −fl+1 Ψl+1 )
i m−1 m−1 m−1 m−1 m+1 m+1 m+1 m+1
+ 2 ∂y (cl−1 Ψl − 1m − 1 − dl−1 + dl+1 Ψl+1 + el−1 Ψl − 1 − fl+1 Ψl+1 )
+∂t (am m m l m
l−1 Ψl−1 + bl+1 ) + Σtl Ψl = ql , 0 ≤< ∞, −l ≤ m ≤ l

4 Two dimensional PN equations

For a two pdimensional domain
p D, consider the infinite cylinder DxR ⊂ R3 ,
Let Ω = ( 1 − µ2 cosφ, 1 − µ2 sinφ, µ)T
We assume that all data are z-independent , then the solution Ψ is Z-
independent and additionally an even function in µ.therefore if l + m is
odd, the associated Legendre polynomial is an odd function in µ, and as a
consequence the moments for which l+m is odd have to vanish. Thus the
choice of l and m are (0,0), (1,-1), (1,1), (2,-2) (2,0) (2,2).

Thus we obtain the following matrix formulations

 
ψ00
. d−1 −f11 .
 

ψ00

1 . .  ψ1−1 
e00
. . d−2 −f 0
0  
ψ1−1 ψ11
   2 2  
. e−1


ψ11
 1 1 0 . . . 
ψ2−2

∂t   + ∂x 
   
2 . . . . . .  
 ψ2−2   ψ20 
. −c−1
 1
e . . . 
  
..  1 1  ψ22 
. . 0 −c1 1
. .
 
..
.

9
  
ψ00
  −1 
. −d−1  ψ1 

1 −f11 . . .
0 2 0  ψ1 
 e0 . . −d2 −f2 0  1 
i −1
  ψ −2 
+∂y  . e1 0 . . .  2 

2  0 
 . e−1
1 e 1
1 . . .   ψ2 
 ψ2 
. 0 c11 . . .  2 
..
.
 
ψ00
  −1   0 
 ψ1  q0

Σt0
ψ 1   q −1
Σt1   

  1   1 
−2   1
ψ  =  q1

+ Σt1 = 
  2 
0   −2
 ψ2   q2

Σt2  
 

Σt2
 ψ2  ..
 2  .
..
.
Let
 
ψ00
d−1
−f11 .
 
. 1 . .  ψ1−1 
e00
. . d−2 −f 0
0  
ψ11
 2 2  
. e−1

1 0 . . . 
ψ2−2

Mxcomplex =  1  
2 . . . . . .  
 ψ20 
. −c−1

e11 . . .  
ψ22
 1  
1
. 0 −c1 . .
 
..
.

 
ψ00
. −d−1 ψ1−1

−f11

. . .

1  
 e00
. . −d2 −f20
2
0  ψ11 
i 
ψ2−2

MYcomplex =  . −1
e1 0 . . .  
2  
 . e−1
1 e 1
1 . . . 
 ψ20 

. 0 c11
. . .

 ψ22 

..
.
m −m
and the matrix containing Σtl by c . Note that Ψml = (−1) Ψl
m m
The real variables Rl (f or0 ≤ m ≤ l) and Il (f or0 < m ≤ l can be obtained
by setting Rlo = Ψol For m 6= 0,

(−1)m m
Rlm = √ (Ψl + (−1)m ψl−m )
2

10
 (−1)m i m
Ilm = √ (ψl − (−1)m Ψ−m
l
2

Z
ψlm = (Ω)ΨdΩ − − − −(A)
s2

s
m 2l + 1 (l − m) imφ m
Y l = (−1)m e Pl (µ) − − − − − − − − − − − (B)
4π (l + m)

(−1)m m
Rlm = √ [Ψl − Ψ−m
l ] − − − − − − − − − − − −(C)
2

Evaluation : −

T of indR00

l = 0, m = 0
r
2(0) + 1 i0φ 0
Y00 = (−1)0 e Po (µ)sincePo0 (µ) = 1
4π
0 1
Yo = √
Z 4π
1 Ω
Ψ00 = √ dΩ = √
4π 4π
0
√
(−1) 2 o
Therefore Roo = Ψo − Ψ−o
o ] = 0
[

T of indR11

l = 1, m = 1
q
1 1 3 0! iφ 1
Y1 = (−1) 4π 2!
e P1 (µ)
q p
3 iφ
= − 8π e (− −1 − µ2 )
q p
3 iφ
= 8π e ( −1 − µ2 )

11
 q
Y1−1 3 −iφ
p
=− 8π
e ( −1 − µ2 )
R q 3 iφ
p
Ψ11 = s2 8π
e ( −1 − µ2 ΨdΩ
q p
3 iφ
Ψ11 = 8π
e ( −1 − µ2 ΨΩ
q
−Ψ−1 3 −iφ
p
1 = 8π
e ( 1 − µ2 ΨΩ
1
Therefore R11 = (−1)
√ [ψ 1 − Ψ−1 ]
2 1 1
q q
−1
−1 3
p 3 −iφ
p
1 iφ 2
= √2 [Ψ1 = 8π e ( −1 − µ ΨΩ + Ψ1 = 8π e ( 1 − µ2 ΨΩ]
q p
−1 3
=√ 2 8π
( −1 − µ2 ΨΩ[2cosφ]
q p
3
R11 = 4π 1 − µ2 ΨΩcosφ

We always place the first component of the solution vector, i.e the scalar
flux R00 , on this cell-centered grid G11 .

E.g: Let nx = 5, ny = 3, Lx = 5, Ly = 3. Therefore ∆x = 1, ∆y = 2

Consider the following division of the rectangular domain into a 5 × 3 ar-
rangements of cells.
Now
     
1 1
G11 = i− ∆x, j − ∆y |i ∈ {1, 2, . . . , nx }, j ∈ {1, 2, . . . , ny }
2 2
    
1
G21 = i∆x, j − ∆y |i ∈ {px , . . . , nx }, j ∈ {1, 2, . . . , ny }
2
   
1
G12 = i− ∆x, j∆y |i ∈ {1, 2, . . . , nx }, j ∈ {py , . . . , ny }
2
G22 = {(i∆x, j∆y) |i ∈ {px , . . . , nx }, j ∈ {py , . . . , ny }}

0, extrapolation b.c. in x;
wherepx =
1, periodic b.c. in x.

0, extrapolation b.c. in y;
py =
1, periodic b.c. in y.

12
For our example
                 
1 1 1 3 3 3 9 9 9
G11 = ,1 , ,3 , ,5 , ,1 , ,3 , ,5 ,..., ,1 , ,3 , ,5
2 2 2 2 2 2 2 2 2
G21 = {(1, 1), (1, 3), (1, 5)(2, 1), (2, 3), (2, 5), . . . , (5, 1), (5, 3), (5, 5)}
                 
1 1 1 3 3 3 9 9 9
G12 = ,2 , ,4 , ,6 , ,2 , ,4 , ,6 ,..., ,2 , ,4 , ,6
2 2 2 2 2 2 2 2 2
G22 = {(1, 2), (1, 4), (1, 6)(2, 2), (2, 4), (2, 6), . . . , (5, 2), (5, 4), (5, 6)}

5 Numerical Method
We now develop a numerical scheme for linear system of hyperbolic balance
laws of the form
∂t ū + Mx · ∂x ū + My · ∂y ū + c · ū = q̄..........(8)
where the matrices Mx , My and C possess very specific patterns of their non
zero entries that admit the systematic placement of the components of the
solution vector ū(x, y, t) on staggered grids. Specifically let the solution (8)
have N components, i.e., ū(x, y, t) ∈ RN and the source q̄(x, y, t) ∈ RN and
the martrices Mx , My ∈ RN ×N and C(x, y, t) ∈ RN ×N are of appropriate
sizes. Moreover the matrix C(x, y, t) is diagonal and the matrices Mx , My
are constant-coefficient, and possesses patterns of their non-zero entries, as
described in lemma 2. Hence the real value PN equations are covered as are
the SPN equation.
Spatial Approximation on staggered grids. We consider the partial differ-
ential equation (8) to hold in the interior of a rectangular computational
domain Ω = (Lx , 0) × (0, Ly ) and on each of the two boundary directions, we
allow for one of two types of boundary conditions periodic or extrapolation.
The domain Ω is divided into nx ×ny rectangular cells of size ∆x×∆y, where
∆x = Lnxx and ∆y = Lnyy . The center points of these cells then lies on the grid
     
1 1 
G11 = i− ∆x, j − ∆y i ∈ {1, 2, . . . , nx }, j ∈ {1, 2, . . . , ny } .....(9)

2 2
To reiterate, the condition on the non zero entry patterns of Mx , My &C,
the components of S ū canSbe distributed
S into four disjoint sets, according to
{1, 2, . . . , N } = I11 I21 I12 I22 such that the following properties hold
[ [ [ [
(Mx )i,j =0∀(i, j) ∈ / (I11 × I21 ) (I21 × I11 ) (I12 × I22 ) (I22 × I12 )
[ [ [ [
(My )i,j =0∀(i, j) ∈ / (I11 × I12 ) (I12 × I11 ) (I21 × I22 ) (I22 × I21 )

13
Having defined these fully staggered grids, the solution components with
indices in Ikl are assigned to the corresponding grid Gkl , where k, l ∈ {1, 2} on
these staggered grids, spatial derivatives of a function ω can be approximated
by the half-grid central difference approximation
      
1 1 1 Z
∂x ω(i∆x, j∆y) ≈ ω i+ ∆x, j∆y −ω i− ∆x, j∆y ∀i, j ∈
∆x 2 2 2
        
1 1 1 Z
∂y ω(i∆x, j∆y) ≈ ω i∆x, j + ∆y −ω i∆x, j − ∆y ∀j, i ∈
∆y 2 2 2
We donote the resulting finite difference operators Dx &Dy , respectively with
these finite difference approxiamations, x-derivative of components on the
grid Gkl are associated with the grid G3−k,l and y derivatives of components
on the grid Gkl are placed on the grid Gk,3−l , where k, l ∈ {1, 2}
The non zero entry patterns (10) guarantee that the distribution of the indices
of Mx · Dx ū + My · Dy ū + C · ū. Moreover the components of the source vector
q̄ in (8) are placed on the same grids as the corresponding components of the
solution vector.
Boundary conditions: In each of the two coordiane direction(as in the
paper).

Periodic: The solution satisfies S ū(x + Lx , y, t) = ū(x, y, t).

Define for all components k ∈ G21S G22 set uk (0, j∆y) := uk (Lx , j∆y)∀j ∈
Iy and for all components k ∈ G11 G12 set uk (Lx + 12 ∆x, j∆y) := uk ( 21 ∆x, j∆y)∀j ∈
Iy where Iy = { 12 Py , 21 Py + 12 , . . . , ny − 12 , ny }
S
Extrapolation: For every solution component with k ∈ G11 G12 satisfy
∂x uk = 0 at the left and the right boundary. By copying grid data onto ghost
points, we define
1 1
uk (− ∆x, j∆y) := uk ( ∆x, j∆y)∀j ∈ Iy and
2 2
1 1
uk (L+ − ∆x, j∆y) := uk (Lx − ∆x, j∆y)∀j ∈ Iy
2 2
1 1 1 1
where Iy = { 2 Py , 2 Py + 2 , . . . , ny − 2 , ny }

The fig.1 shown an example with periodic boundary condition in the

x-direction(grid data that is wrapped around in depicted in blue) and ex-
trapolation boundary condition in the y-direction(ghost points are depicted
in red).

14
 Time Stepping:
BootshappingSis followed to resolve the time domain in equation(8) i.e data S on
the grids G11 G22 is updatedS alternatingly with data on the grids G21 SG12 .
The solution components I11 I22 depend on the components in I21 I12 .
The former is called as even components as i + j is even and the latter is
called as odd components accordingly. When using a time step ∆t, the even
components would be defined at times n∆t, while the odd components would
be defined at times (n + 21 )∆t. Now define the solution components on the
same time grid n∆t , and conduct a time-step fromt to t+∆ t as follows
ūe
Consider the solution vector in the block form ū =
ūo
where ūe is the vector of the even components and ūo is the vector of the odd
components. Similarly apply to the source vector q̄ and the matrices Mx , My
and C, the equation (8) becomes

0 Myeo
 e     e     e 
ū 0 Mxeo ū ū
∂t + · Dx + · Dx
ūo Mxeo 0 ūo Myeo 0 ūo
 e   e   e 
c 0 ū q̄
+ o · o = .........(14)
0 c ū q̄ o
Define two evolution operators. The first operator updates only the even
components by freezing the odd components and second updates only odd
components by freezing the even ones. In the first consider the odd compo-
nents to be constant on the interval [t, t + ∆t]. Similar procedure is applied
for the matrix C and the source vector q̄, both are evaluated at time t + 12 ∆t
with these assumption the bolck system (14) decouples into two ODE’s

∂t · ūe + C e · ūe = q̄ e − Mxeo · Dx ūo − −Myeo · Dy ūo

∂t · ūo + C 0 · ūo = q̄ o − Mxoe · Dx ūe − −Myoe · Dy ūe .........(15)

where C e , C o and the RHS are time independent.

Since C = diag(C1 , C2 , . . . , CN ) is diagonal, the equation in (15) decouple

into scalar equation to the form

∂r uk (x, y, τ ) + C¯k (x, y)uk (x, y, τ ) = r¯k (x, y).........(16)

which has to be solved from τ = t to τ = t + ∆t.

In (16) we have
1 1
C¯k (x, y) = Ck (x, y, t + ∆t); r¯u (x, y) = rk (x, y, t + ∆t)
2 2
15
The exact solution of (16) is
1
uk (x, y, t+∆t) = exp(−C¯k (x, y)∆tuk (x, y, t)− ¯ (1−exp(−C¯k (x, y)∆t))r¯k (x, y))..........(17)
Cu (x, y)
ec − 1
= uk (x, y, t) + ∆t(r¯k (x, y) − C¯k (x, y)uk (x, y, t)) ...........(18)
c
Using formula (18) we define the even evolution operation
 e   e   e 
e ū ū e r̄ − C̄ e · ūe
St+∆t,t = + ∆tE(−C̄ ∆t) ..............(19)
ūo ūo 0

and the odd evolution operation

 e   e   
o ū ū o 0
St+∆t,t = + ∆tE(−C̄ ∆t) ..............(20)
ūo ūo r̄o − C̄ o · ūo
exp(c)−1
where E(C) = C

A full solution step form t to t + ∆t is acheived via

o e e o
St+ 1 ∆t,t = St+ 1
∆t,t
· St+ 1
∆t,t
· St+ 1
∆t,t
· St+ 1
∆t,t
...........(21)
2 2 2 2 2

e e e
In general St+ 1
∆t,t
· St+ 1
∆t,t
6= St+ 1
∆t,t
2 2 2
Accuracy: The spatial truncation error is O(h2 ) , where h = max{∆x, ∆y}.
Due to the symmetry in the arragement of the sub-half steps in (21), the
splitting error in a single time step in O(∆t3 ) and the global temporal trun-
catin error due to the fractional steps is O(∆t2 ). This second order accuracy
in time is preserved because all temporal evaluated are done in a symmet-
ric fashion. The overall second order accuracy i.e the truncation error is
O(∆t2 ) + O(h2 ) is confirmed by the numerical result sec 6.1
Stability: Here we shall see the L2 stability namely the combination of
staggered spatial grids with the temporal splitting (21). Consider the P1
system in one space dimension with constant coefficients and without a source

T
q̄ = 0. Here the solution vector has two scalar fields ū = ue uo which
satisfy the equation
∂t ue + ∂x uo + ce ue = 0
∂t uo + ∂x ue + co uo = 0
Conducting Von Neumann stability analysis, let the numerical solution
at time t be represented interms of its Fourier coefficients.

16
 The even Fourier coefficient
X
ue (x, t) = aek (t)eikx
k

and the odd Fourier coefficient

X
uo (x, t) = aok (t)eikx
k

where the domain is of periodic length 2π. Due to the linearity of the updated
rule, the coefficients aek (t, ∆t) and aok (t, ∆t) depends only on the coefficients
aek (t) and aok (t) for a particular k. Hence Fourier modes of different frequences
do not mix. Consequently it is enough to analyse the growth factor for basic
wave solution
ue (x, t) = aek (t)eikx and uo (x, t) = aok (t)eikx ...........(22)
Let x = hj , j ∈ Z for the grid hj in ue and for uo , let x = h(j + 21 ), j ∈ Z.
Then the grid solution values are
1 1
ue (hj , t) = aek (t)eikhj and uo (h( j + ), t) = aek (t)eikh( j+ 2 )
2
The staggered grid derivative are
1 ae (t)eikh(j+1) − eikhj
(Dx ue )(h(j + ), t) = k
2 h
aek (t)eikhj ikh
= (e − 1)
h
h
aek (t)eikhj eik 2 ik h h
= (e 2 − e−ik 2 )
h
2i ikh(j+ 1 ) h e
= e 2 sin k a (t)
h 2 k
and
1 1
o aok (t)(eikh(j+ 2 ) − eikh(j− 2 ) )
(Dx u )(hj, t) =
h
i kh ikhj o
= 2 sin( )e ak (t)
h 2
Hence, for the basic waves (22) on the staggered grids, the even half-steps
solution operator (19) updates the Fourier coefficient by the linear operation
 e   e e  e 
ak 1 d f ak
o (t + ∆t) = (t)
ak 2 0 1 aok
= Ge1 ∆t
2

17
where de = exp(− 12 ce ∆t) and f e = −i ∆t
h
E(− 21 ce ∆t) sin( kh
2
).
Similarly the odd half-step solution operation (20) acts on the Fourier
coefficient as
 e    e 
ak 1 0 1 ak
(t + ∆t) = (t)
aok 2 d o
f o
aok
= Go1 ∆t
2

where do = exp(− 12 co ∆t) and f o = −i ∆t h

E(− 12 co ∆t) sin( kh
2
).
Consequently a full solution time step (21) acts on the Fourier coefficient
as
 e   e 
ak o e e o ak
o (t + ∆t) = G 1 ∆t · G 1 ∆t · G 1 ∆t · G 1 ∆t · (t)
ak 2 2 2 2 aok
   e 
(de )2 + (de + 1)f e f o do (de + 1)f o ak
= · (t)
(do + (de )2 )f o + (de + 1)f e (f o )2 (do )2 + do (de + 1)f e f o aok
p
The eigen value of the growth factor matrix G are λ1,2 = g± g 2 − (do de )2

18