Acceleration of the Accelerating the Solution of SN Equations with Highly

Anisotropic Scattering Using the Fokker-Planck Equation Approximation

Japan K. Patel,1 James S. Warsa,2 and Anil K. Prinja3
1
Department of Mechanical and Aerospace Engineering, The Ohio State University, Columbus, OH
2
Transport Methods Group (CCS − 2), Los Alamos National Laboratory, Los Alalmos, NM
3
Department of Nuclear Engineering, The University of New Mexico, Albuquerque, NM

patel.3545@[Link], warsa@[Link], prinja@[Link]

Abstract: The discrete ordinates method can model forward-peaked transport problems accurately.
However, convergence of discrete ordinates solution can become arbitrarily slow upon use of standard it-
erative procedures like source iteration and GMRES. Standard zeroth and first moment-based acceleration
methods like nonlinear diffusion acceleration and diffusion synthetic acceleration are ineffective in accelerat-
ing such problems because these methods do not correct higher order Legendre-moments of angular flux. We
explore the idea of using Fokker-Planck as a preconditioner to accelerate forward-peaked transport problems
in this paper.

Keywords: radiation transport, highly anisotropic scattering, synthetic acceleration, Fokker-Planck

1 Introduction
Transport problems with forward-peaked scattering kernels are encountered in several applications related to
plasma physics, radiation sheilding, medical physics, and astrophysics. Such problems have extremely small
mean free paths and nearly singular differential scattering cross-sections in the forward direction. Use of
discrete ordinates method with standard methods like source iteration and GMRES can become extremely
inefficient due to these properties. Standard acceleration techniques like diffusion synthetic acceleration
(DSA) (Alcouffe, 1977) and nonlinear diffusion acceleration (NDA) (Smith et al., 2011) are ineffective in
accelerating such problems. These methods assume all moments higher than the zeroth moment are incon-
sequential to the convergence of the solution. We see, later in this paper, that such an assumption becomes
invalid for forward-peaked problems.

Several innovations have been made to accelerate the convergence of such problems. Valougeorgis,
Williams, and Larsen (Valougeorgis et al., 1988) presented their work on stability analysis of PL accel-
eration applied to anisotropic neutron transport problems. This paper presented an extremely valuable
framework for theoretical development and testing of future acceleration methods for transport problems
with anisotropic scattering. Khattab and Larsen presented their modified PL acceleration method that used
a modified form of PL equations with modified scattering cross section moments in (Khattab and Larsen,
1991). Morel and Manteuffel presented their angular multigrid method for solution of problems with high
anisotropy in (Morel and Manteuffel, 1991). The angular multigrid method proved to be effective in 1D with
a maximum spectral radius of 0.6. The method however had to be modified later to preserve stability for
problems in higher spatial dimensions in (Pautz et al., 1998). Turcksin and Morel integrated diffusion syn-
thetic acceleration and angular multigrid to develop their diffusion synthetic acceleration-angular multigrid
method in (Trucksin, 2012). Aristova and Gol’din developed a quasi-diffusion method for forward peaked
radiative transfer problems in (Aristova and Gol’din, 2000).

Several approximations to the transport equation have also been derived to tackle forward-peaked prob-
lems. Most prominent of these is the Fokker-Planck approximation which is an asymptotic limit of the
Boltzmann equation (Pomraning, 1992) as scattering angle and energy loss become small (Fokker-Planck
limit). Renormalization techniques can be applied for generating stable higher order approximations in this
limit to obtain generalized-Fokker-Planck equations (Pomraning, 1996), (Prinja and Pomraning, 2001), and
(Leakes and Larsen, 2001). The scattering kernel can be decomposed into smooth and singular parts (Caro
and Ligou, 1983), (Landesman and Morel, 1989), (Aristova and Gol’din, 1998), (Dixon, 2015) to derive the
Boltzmann-Fokker-Planck or Boltzmann-Fokker-Planck-like approximations.

In this paper primarily focuses on the acceleration side of the solving forward-peaked problems. Prob-
lems with forward-peaked scattering kernels require acceleration of all slowly-converging Legendre-moments
of angular flux with significant magnitudes. We develop and test a synthetic acceleration method - Fokker-
Planck synthetic acceleration (FPSA) - where the lower-order approximation for the error-correction stage
is obtained using asymptotic analysis (Bender and Orszag, 1978) in the Fokker-Planck limit. We call this
approach, to acceleration, asymptotics-based acceleration.

We organize the remainder of this paper as follows; in the next section, we introduce the FPSA and
describe how we discretize the angular Laplacian term of the Fokker-Planck equation - weighted finite differ-
ence (Morel, 1985) and moment preserving discretization (Warsa and Prinja, 2012). In section 3, we present
angularly-continuous and angularly-discrete Fourier analyses for FPSA and contrast them. Thenafter, in
section 4, we present an efficiency study for FPSA screened Rutherford kernel (SRK) (Pomraning, 1992), the
exponential kernel (EK) (Prinja et al., 1992) and Henyey-Greenstein kernel (HGK) (Henyey and Greenstein,
1941). We conclude this paper with a summary in section 5.

2 Fokker-Planck Synthetic Acceleration

The ideas presented here can be extended to problems with time and energy dependence, multiple dimen-
sions, and curvilinear coordinates. Our presentation, however, takes place in the context of steady-state,
monoenergetic, one-dimensional Cartesian coordinates. We use standard notation and assume cgs units
(Lewis and Miller, 1984), such that for a domain z ∈ [z0 , z1 ] the transport equation is
Z 1
∂
µ ψ(z, µ) + σt (z)ψ(z, µ) = dµ0 σs (z, µ0 )ψ(z, µ0 ) + q(z, µ), (1a)
∂z −1

with boundary conditions

ψ(z0 , µ) = ψL (µ) for µ > 0 and ψ(z1 , µ) = ψR (µ) for µ < 0. (1b)

The cross section σs (µ0 ) depends on the cosine of the laboratory frame scattering angle, µ0 = Ω̂0 · Ω̂, for a
particle traveling with incident direction Ω̂0 , and exiting after a scattering event in direction Ω̂. Typically,

2
this dependence is represented with an expansion in Legendre polynomials, whose expansion coefficients are
Z 1
σs,l (z) = dµ0 σs (z, µ0 )Pl (µ0 ). (2)
−1

Assuming the expansion is truncrated at order L, and using the addition theorem for the normalized spherical
harmonics, Eq. (1a) becomes

L
∂ X 2l + 1
µ ψ(z, µ) + σt (z)ψ(z, µ) = σs,l (z)Pl (µ)φl (z) + q(z, µ), (3)
∂z 2
l=0

where the scalar flux moments are Z 1

φl (z) = dµ Pl (µ)ψ(z, µ). (4)
−1

Under certain restrictions on the scattering cross section and its expansion, The Fokker-Planck equation is an
asymptotic limit of the Boltzmann transport equation when scattering is highly forward-peaked (Pomraning,
1992). The slab geometry Fokker-Planck equation is

∂ψ σtr (z)
µ + σa (z)ψ(z, µ) − LFP ψ(z, µ) = q(z, µ), (5a)
∂z 2

where
∂ ∂
LFP = (1 − µ2 ) (5b)
∂µ ∂µ
and the momentum transfer (or transport) cross section is σtr = σs,0 − σs,1 .

2.1 Source Iteration and Synthetic Acceleration

We demonstrate review the ideas behind source iteration and synthetic acceleration (Adams and Larsen,
2001) before describing FPSA. Source iteration is one of the simplest methods used to solve the Boltzmann
equations. We begin by rewriting (3):

Lψ(z, µ) = Sψ(z, µ) + q(z, µ), (6a)

where
L Z 1
∂ X 2l + 1
L=µ + σt (z) and S = σs,l (z)Pl (µ) dµ Pl (µ). (6b)
∂z 2 −1
l=0

Source iteration for the Boltzmann equation is then written as:

Lψ m+1 (z, µ) = Sψ m (z, µ) + q(z, µ), (7)

σs,l
where m is the iteration index. Fourier analysis of source iteration (Valougeorgis et al., 1988) returns σt as
the eigenvalues of the iteration matrix. Depending on the scattering kernel, the spectral radius can approach
unity. This could make solution extremely expensive. In order to optimize the performance of such schemes,
σs,l
the solution must be accelerated. The eigenvalues also suggest that for forward-peaked kernels, where σt
decay slowly, accelerating higher order moments with nonzero magnitudes becomes essential. In order to
understand synthetic acceleration (Kopp, 1963) consider (7). We break the solution procedure as follows:

3
 Predict
1
Lψ m+ 2 (z, µ) = Sψ m (z, µ) + q(z, µ), (8a)

Correct
1 1
ψ m+1 (z, µ) = ψ m+ 2 (z, µ) + F −1 S(ψ m+ 2 (z, µ) − ψ m (z, µ)) (8b)

Iterate if
||φm+1
l (z) − φm
l (z)||∞ > tolerence (8c)

Different choices of F operator return different synthetic acceleration schemes. For example, choosing F as
the diffusion operator the streaming-plus-collision operator of the diffusion equation returns DSA (Adams
and Larsen, 2001). For this paper FPSA, we choose F as the Fokker-Planck operator as follows:

∂ σtr (z)
F =µ + σa (z) − LFP (9)
∂z 2

2.2 Discretization
In order to discretize the Boltzmann and FP equations, we use linear discontinuous finite element discretiza-
tion (LD) in space (Warsa, 2014) and discrete ordinates (SN ) in angle (Lewis and Miller, 1984). Moreover
in order to discretize LFP , we use weighted finite difference (WFD) (Morel, 1985) and moment preserving
discretization (MPD) (Warsa and Prinja, 2012). For a more detailed presentation on spatial and angular
discretization, we refer the readers to (Patel, 2016). For convenience, we briefly review angular discretization
of FP equation based on (Morel, 1985) and (Warsa and Prinja, 2012). We use SN quadrature to discretize
the FP equation (5) in angle by collocating the angular flux at the directions µn

∂ψn
µ + σa (z)ψn (z) − ∇2n ψ(z) = qn (z), (10)
∂z

for n = 1, . . . , N . A way to discretize the term ∇2n ψ(z), which denotes the discrete form of the angular
Laplacian operator Eq. (5b) evaluated at angle n, has to be defined in terms of the SN quadrature points
and weights.

The three-point WFD scheme uses the same weights as the SN discretization of the one-dimensional
transport equation in spherical coordinates (Morel, 1985). The WFD scheme is given by the following
expressions.
∇2n ψ(z) = γn+1/2 ψ̇n+1/2 (z) − γn−1/2 ψ̇n−1/2 (z), (11)

where
ψn+1 (z) − ψn (z)
ψ̇n+1/2 (z) = , (12)
µn+1 − µn
γn+1/2 = γn−1/2 + νµn wn , with γn−1/2 = 0. (13)

and where the SN quadrature normalization is ν.

To develop the MPD method, we first recognize that in one dimension the Legendre polynomials are eigen-
functions of the Fokker-Planck operator

LFP Pl (µ) = −l(l + 1)Pl (µ). (14)

4
For a twice-differentiable function f (µ), integrating twice by parts shows that the angular Laplacian operator
is self-adjoint with respect to the Legendre polynomials:
Z 1 Z 1
[LFP Pl (µ)] f (µ) dµ = Pl (µ) [LFP f (µ)] dµ. (15)
−1 −1

Substituting (14) in (15), the following integral relationship is readily obtained

Z 1 Z 1
Pl (µ) [LFP f (µ)] dµ = −l(l + 1) Pl (µ)f (µ) dµ. (16)
−1 −1

We now evaluate this relationship with SN quadrature for the angular flux ψn (z) to get

N
X N
X
wn Pl (µn )∇2n ψ(z) = −l(l + 1) wn Pl (µn )ψn (z), (17)
n=1 n=1

for l = 0, . . . , N − 1. This defines an (N × N ) operator for the vector of N angular fluxes at the spatial
location z, Ψ(z), such that the result is the Fokker-Planck operator collocated at all N quadrature points
simultaneously. That is,
∇2 Ψ(z) = FΨ(z) (18a)

where ∇2 is the discrete approximation to LFP , and where

F = V−1 LV, (18b)

where the elements of V and L are

Vi,j = Pi−1 (µj )wj
(18c)
Li,i = −i(i − 1),
for i, j = 1, . . . , N . The method is a similarity transformation that by definition preserves the moments of
the flux up to order N −1. Using Eq. (18), the MPD method for the SN approximation to the Fokker-Planck
equation is, in operator notation,

∂ σtr (z)
H Ψ(z) + σa (z)Ψ(z) − FΨ(z) = Q(z), (19)
∂z 2

where the operator M is

H = diag (µn ),
n=1,N

and Q(z) is a vector of source terms qn (z) for n = 1, . . . , N .

Notice that we may also write the WFD operator in the manner of Eq. (18a), that is,

∇2 Ψ(z) = WΨ(z) (20a)

5
where W is the (N × N ) tridiagonal matrix whose elements are
  
1 γn−1/2
, j = i − 1, j > 1,


wn µn − µn−1





 1  γ 
n+1/2
Wi,j = , j = i + 1, j < N, (20b)

 w n µn+1 − µn
  

 1 γn−1/2 γn+1/2
− + , j = i,


wn µn − µn−1 µn+1 − µn

for i, j = 1, . . . , N . The WFD scheme for the SN approximation to the Fokker-Planck equation can then be
written in operator notation as

∂ σtr (z)
H Ψ(z) + σa (z)Ψ(z) − WΨ(z) = Q(z), (21)
∂z 2

As observed in (Morel, 1985), we see that the WFD scheme results in a diagonally-dominant M-matrix such
that the transport operator is inverse-positive (neglecting spatial discretization). In other words, the inverse
of the operator exists, and is zero or greater (Meyer and Stadelmaier, 1978). Even though the MPD operator
does not have a similar simple structure that allows us to show it so easily, we have observed numerically
that it is in fact inverse-positive.

3 Fourier Analysis for FPSA

Now, we analyze FPSA using angularly-continuous and angularly-discrete Fourier analysis in this section.
The goal of Fourier analysis is to theoretically determine how efficient our method is. We begin by deter-
mining the error equation for FPSA. We assume constant material properties throughout this exercise. We
also drop notation for z and µ dependence henceforth for convenience.
Consider Eq. 8b. Upon subtracting the exact solution ψ from both sides and adding and subtracting ψ
from the scattering term, we have:

1 1
ψ m+1 − ψ = ψ m+ 2 − ψ + F −1 S(ψ m+ 2 − ψ + ψ − ψ m ). (22)

Upon introduction of the following error definitions in Eq. (22):

1 1
ψ − ψ m+ 2 = m+ 2 , and ψ − ψ m = m , (23)

simplification, and rearrangement, we get the following error equation:

1 1
m+1 = m+ 2 − F −1 S(m − m+ 2 ). (24)

Based on whether we analyze the equation via Legendre-moments of angular error or by discretizing angular
error with SN quadrature, we get angularly-continuous or angularly-discrete analysis.

6
3.1 Angularly-Continuous Fourier Analysis
Angularly-continuous or PL -based Fourier analysis is inspired by (Valougeorgis et al., 1988). We begin by
defining error moments:
Z1 N
X
m
l = dµPl (µ)m = wn Pl (µn )m . (25)
−1 n=1

In terms of moments, the error equation is:

Z 1
m+ 1 1
m+1
l = l 2 − dµPl (µ)F −1 S(m − m+ 2 ). (26)
−1

We follow the following general steps:

m+ 1
1. Obtain an expression for l 2 by analyzing the predictor step.
R1 1
2. Obtain an expression for −1 dµPl (µ)F −1 S(m − m+ 2 ) by analyzing the corrector step.

3. Combine results from previous steps to obtain the iteration matrix IM such that the error is written
according to [m+1
l ] = IM [m
l ].

The spectral radius of IM determines the convergence rate of the iterative method (Hageman and Young,
1984). This is because of the following relation:

[m+1
l ] = IM [m m 0
l ] = IM [ ]. (27)

m+ 12
Step 1: In order to proceed, we note that l comes from the error moment equation of the predictor
step Eq. (8a). We obtain that equation by subtracting the exact transport equation from Eq. (8a), and
defining error according to Eq. (23):

1 L Z1
∂m+ 2 1
X 2l + 1
µ + σt m+ 2 = Pl (µ)σs,l dµ0 Pl (µ0 )m . (28)
∂z 2
l=0 −1

1
Now, we separate the error components into their angle and space dependent components by writing m+ 2
and m as Fourier integral (Adams and Larsen, 2001):
Z ∞
1
m+ 2 = m+1
dλˆλ (µ)eiλσt z , (29)
−∞

where, λ is the wave number. Substituting this form of error into the error equation Eq. (28) returns:
 
∞ L Z1
m+1 (µ)eiλσt z
Z
∂ˆ X 2l + 1
dλ µ λ
+ σt ˆm+1
λ (µ)eiλσt z = Pl (µ)σs,l dµ0 Pl (µ0 )ˆ
m
λ (µ)e
iλσt z 
. (30)
−∞ ∂z 2
l=0 −1

Simplifying the above equation and noting that Fourier modes, eiλσt z , are linearly independent for all λ, we

7
obtain (Adams and Larsen, 2001):

L Z1
X 2l + 1
(1 + iλµ)σt ˆm+1
λ (µ) = Pl (µ)σs,l dµ0 Pl (µ0 )ˆ
m
λ (µ) (31)
2
l=0 −1

Dropping λ and µ in notation of ˆ in Eq. (31) for convenience, and using definitions of error-moments, we
get:
L
1
X 2l + 1
(1 + iλµσt )σt ˆm+ 2 = Pl (µ)σs,l ˆm
l . (32)
2
l=0

Then rearranging the equation and taking nth Legendre moment of Eq. (32), we obtain the following:

Z 1 Z 1 L
m+ 1
X σs,l 2l + 1 Pl (µ) m
dµPn (µ)ˆ
λ 2 = dµPn (µ) ˆ . (33)
−1 −1 σt 2 1 + iλµσt l
l=0

Further rearrangement and use of definition of error moments returns:

L Z 1
m+ 12
X σs,l 2l + 1 Pn (µ)Pl (µ) m
ˆl = dµ ˆ . (34)
σt 2 −1 1 + iλµσt l
l=0

We note that Eq. (34) represents the following matrix equation:

m+ 21
[ˆ
l m
] = A[ˆl ], (35)

m
where, [ˆl ] is a vector of error-moments at iteration m, and

L Z 1
X σs,l 2l + 1 Pn (µ)Pl (µ)
A= dµ , (36)
σt 2 −1 1 + iλµσt
l=0

is an iteration matrix. We multiply Eq. (35) by eiλσt z and use Eq. (25) to get:

m+ 12
[l ] = A[m
l ]. (37)

m+ 12
Now that we have an equation for l , we move on to the next step.

1
Step 2: We begin from Eq. (24). We note that the correction υ m+1 = F −1 S(m − m+ 2 ) comes from
the solution of the following equation:

L
∂υ m+1 σtr ∂ ∂υ m+1 X 2l + 1 m+ 12
µ + σa υ m+1 − (1 − µ2 ) = Pl (µ)σs,l (m
l − l ) (38)
∂z 2 ∂µ ∂µ 2
l=0

Introducing the Fourier mode ansatz:

υ m+1 = υ̂λm+1 (µ)eiλzσt . (39)

8
Upon introduction of Eq. (25), and Eq. (39) in Eq. (38), we get:

∂ υ̂λm+1 (µ)eiλzσt σtr ∂ ∂ υ̂ m+1 (µ)eiλzσt

µ + σa υ̂λm+1 (µ)eiλzσt − (1 − µ2 ) λ
∂z 2 ∂µ ∂µ
L (40)
X 2l + 1 1
iλσt z m+ 2
= Pl (µ)σs,l (eiλσt z ˆm
l −e ˆl ).
2
l=0

Simplifying Eq. (40), taking its Legendre moment, and using the orthogonality property of Legendre poly-
nomials returns:

Z1 Z1 Z1
σtr ∂ ∂ υ̂ m+1 (µ)
iλσt dµPl (µ)µυ̂λm+1 (µ) + σa dµPl (µ)υ̂λm+1 (µ) − dµ (1 − µ2 ) λ
2 ∂µ ∂µ (41)
−1 −1 −1
m+ 12
m
= σs,l (ˆl −
ˆl ).

Now, using the recurrence relation for Legendre polynomials on the first term of Eq. (41), expanding υ̂λm+1 (µ)
in the third term of Eq. (41) using Legendre expansion, we get:

Z1 ∞
l m+1 l+1 m+1 σtr ∂ ∂ X 2l + 1
iλσt υ̂l−1 + iλσt υ̂l+1 + σa υ̂lm+1 − dµ (1 − µ2 ) Pn (µ)υ̂nm+1
2l + 1 2l + 1 2 ∂µ ∂µ n=0 2 (42)
−1
m+ 21
m
= σs,l (ˆl −
ˆl ).

Simple rearrangement of the third term in Eq. (42), followed by use of Legendre’s equation, and orthogonality
property of Legendre polynomials returns:

l m+1 l+1 m+1 σtr m+ 1

iλσt υ̂l−1 + iλσt υ̂l+1 + σa υ̂lm+1 + l(l + 1)υ̂nm+1 = σs,l (ˆ
m
l −
ˆl 2 ), (43)
2l + 1 2l + 1 2

where,
Z1
υ̂lm = dµPl (µ)υ̂λm (µ). (44)
−1

Eq. (43) can be written in matrix form as:

[υ̂lm+1 ] = B −1 XE[ˆ
m
l ] (45)

where,
X = diag(σs,l ), (46a)

E =I −A (46b)
σtr
Bl,l = σa + l(l + 1) (46c)
2
l+1
Bl,l+1 = iλσt (46d)
2l + 1
l
Bl−1,l = iλσt (46e)
2l + 1

9
Multiplying Eq. (45) with eiλzσt and using Eq. (39) returns:
Z 1
1
−1
[υlm+1 ] =B XE[m
l ] = dµPl (µ)F −1 S(m − m+ 2 ). (47)
−1

Step 3: Combining Eqs. (26), (37), and (47), returns:

[m+1
l ] = (A − B −1 XE)[m
l ]. (48)

Comparing Eq. (48) with Eq. (27) returns the iteration matrix IM = (A − B −1 XE). The spectral radius of
IM is the spectral radius of FPSA.

3.2 FPSA as a Special Case of PL Acceleration

Upon carrying out similar analysis for PL acceleration (Valougeorgis et al., 1988), we find that the iteration
matrix for PL acceleration has a similar form except, in this case, the definition of Bl,l is slightly different:

PL
Bl,l = σt − σs,l = σa + σs,0 − σs,l . (49)

That for FPSA is rewritten as:

F P SA σtr σs,0 − σs,1

Bl,l = σa + l(l + 1) = σa + l(l + 1). (50)
2 2

When we equate the two equations, we see that FPSA is a special case of PL acceleration when:

σs,0 − σs,1
σs,l = σs,0 − l(l + 1). (51)
2

Another way of obtaining this equivalence relation, is by noting that Legendre polynomials are eigenfunctions
of both Boltzmann scattering operator and the Fokker-Planck operator as done by Morel in (Morel, 1981):

ΓB Pl (µ) = (σs,l − σs,0 )Pl (µ), (52a)

(σs,0 − σs,1 )
ΓF P Pl (µ) = − l(l + 1)Pl (µ), (52b)
2
and equating the eigenvalues of Fokker-Planck and the Boltzmann scattering operators:

(σs,0 − σs,1 )
σs,l − σs,0 = − l(l + 1). (53)
2

Simple rearrangement of Eq. (53) returns Eq. (51). We will call these cross-section moments PL -equivalent
cross-section moments in this paper.

According to the SN −PL equivalence relation in slab geometry (Lewis and Miller, 1984), when N = L+1,
SN and PL equations are equivalent. Taking this and Eq. (51) into account, we note that FPSA will converge
in one iteration when it is analytically equivalent to PL acceleration. In other words, when the scattering
cross-section moments are according to Eq. (51), FPSA will converge in one iteration. Moreover, it would
be valid to think that the convergence will be rapid in case the cross-section moments are close to those
obtained from Eq. (51). However, in the case when we truncate scattering expansion arbitrarily and N

10
is no longer equal to L + 1, the FPSA-PL acceleration equivalence will no longer hold. This is due to the
inconsistent introduction of zero values for scattering cross section moments with N ≥ l > L (Patel, 2016).

3.3 Angularly-Discrete Fourier Analysis

Angularly-discrete analysis is carried out by using SN quadrature to approximate angular error. We need
angularly-discrete Fourier analysis to analyze FPSA because different discretizations of LFP preserve dif-
ferent number of moments. While WFD only preserves zeroth and first Legendre moments of the angular
flux (Morel, 1985), MPD preserves upto N Legendre moments (Warsa and Prinja, 2012). The angluarly-
continuous Fourier analysis (PL -based analysis) is moment-based; therefore, requires the numerical im-
plementation to preserve all relevant moments in order to get a consistent spectral radius measurement.
Moreover, in case of continuous transport, the transport equation only limits to the Fokker-Planck equa-
tion when a ”sufficient” number of Legendre moments are used to represent the angular flux (Patel, 2016).
This sufficient number of moments is scattering cross-section dependent. This, however, may not neces-
sarily be true in the discrete (SN ) case. This creates a discrepancy between the angularly-continuous and
angularly-discrete Fourier analyses when sufficient number of moments are not used to represent angular flux.
Therefore, in order to verify convergence rates of the numerical implementation, irrespective of how many
moments are used to represent angular flux, we introduce angularly-discrete analysis. First, we consider
Fourier analysis for FPSA with WFD for LFP .

3.3.1 Analysis with WFD

We follow the following analogous steps to do angularly discrete Fourier analysis:

1
1. Obtain an expression for m+ 2 .
1
2. Obtain an expression for F −1 S(m − m+ 2 ).

3. Obtain the overall matrix equation that is used to estimate the the spectral radius.

Step 1: Since we have already detailed angularly-continuous analysis, we skip furnishing the introduction
of Fourier mode assumption and simplification steps here. We also ignore notation of µ and z dependence of
relevant quantities for convenience. Taking the nth Legendre moment of Eq. (32), and using orthogonality
property of Legendre polynomial returns:

Z1 Z1
m+ 1
h i
dµPn (µ) (iλσt µ + σt )ˆ
l 2 = σs,l m
Pl (µ)ˆl . (54)
−1 −1

Now we write each integral as a discrete weighted-sum using SN quadrature:

N N
m+ 1
X h i X
Pl (µn )wn (iλσt µn + σt )ˆ
l 2 = σs,l Pl (µn )wn ˆm
l . (55)
n=1 n=1

Finally, we get the following matrix equation from Eq. (55):

1
[ˆ m ].
m+ 2 ] = Â[ˆ (56)

11
where,
 = Y −1 Z, (57)

and,
Yln = Pl (µn )wn (iλσt µn + σt ) and Zln = σs,l Pl (µn )wn . (58)
σ
Here, Â is the iteration matrix in angularly-discrete from. This returns σs,0
t
as the spectral radius, which
is consistent with the angularly-continuous analysis (Patel, 2016). We multiply Eq. (59) by the relevant
exponential from Fourier mode ansatz to get:

1
[m+ 2 ] = Â[m ]. (59)

Step 2: Upon introduction of Fourier mode assumption for υ in Eq. (38), taking Legendre moment of equa-
tion, using definition of error moments, carrying out the relevant spatial differentiation and simplifications,
we get:
Z1   Z1
σtr ∂ 2 ∂ m+1 1
dµPl (µ) iλσt µ + σa − (1 − µ ) υ̂ m − ˆm+ 2 ).
= σs,l dµPl (µ)(ˆ (60)
2 ∂µ ∂µ
−1 −1

Now we write each integral in the form of a weighted sum and the angular differential using the weighted
difference formulation (Morel, 1989) to obtain:

N
X N
X
iλσt Pl (µn )wn υ̂nm+1 + σa Pl (µn )wn υ̂nm+1
n=1 n=1
N N
σtr X
m+1 m+1
X 1
− bn υ̂nm+1 + cn υ̂n−1 m − ˆm+ 2 ), (61)


− Pl (µn )wn an υ̂n+1 = σs,l Pl (µn )wn (ˆ
2 n=1 n=1

where, an , bn , and cn are according to Eq. (20b). From Eq. (61), we get the following matrix equation:

[υ̂ m+1 ] = B̂ −1 Ĉ D̂[ˆ

m ], (62)

where,
D̂ = I − Â, Ĉl,n = σs,l Pl (µn )wn , ˆ + B2
and B̂ = B1 ˆ + B3,
ˆ (63)

with,
ˆ l,n = Pl (µn )wn (iλσt µn + σa + bn ) ,
B1 (64)
ˆ l,n+1 = Pl (µn )wn an ,
B2 (65)

and,
ˆ l,n−1 = Pl (µn )wn cn .
B2 (66)
1
We obtain the following expression for F −1 S(m − m+ 2 ):,

1
υ m+1 = υ̂ m+1 eiλσt z = F −1 S(m − m+ 2 ). (67)

Step 3: Combining Eqs. (66), (59), and (24) returns:

[m+1 ] = (Â − B̂ −1 Ĉ D̂)[m ], (68)

12
where, IM = Â − B̂ −1 Ĉ D̂ is the iteration matrix and its spectral radius determines the convergence rate of
FPSA with WFD. Now, we consider angularly-discrete analysis for FPSA with MPD.

3.3.2 Analysis with MPD

Angularly discrete Fourier analysis for FPSA with MPD is done in the same way as for FPSA with WFD.
The only difference will be how the Fokker-Planck operator is represented in step 2. Introducing angularly
discrete formulation for integrals and MPD formulation (Warsa and Prinja, 2012) for the angular Laplacian
in Eq. (60) returns:

N N N
X X σtr X
iλσt Pl (µn )wn υ̂nm+1 + σa Pl (µn )wn υ̂nm+1 + l(l + 1) Pl (µn )wn υ̂nm+1
n=1 n=1
2 n=1
N
1
X
= σs,l m − ˆm+ 2 ). (69)
Pl (µn )wn (ˆ
n=1

From Eq. (61), we get the following matrix equation:

[υ̂ m+1 ] = B̃ −1 C̃ D̃[ˆ

m ], (70)

where,
D̃ = I − Â, (71)

C̃l,n = σs,l Pl (µn )wn , (72)

and,
σtr
B̃l,n = iλσt wn Pl (µn )µn + σa wn Pl (µn ) + l(l + 1) wn Pl (µn ). (73)
2
1
We have the following expression for L−1 S(m − m+ 2 ):

1
υ m+1 = υ̂ m+1 eiλσt z = L−1 S(m − m+ 2 ). (74)

Step 3: Just like with previous analyses for FPSA, we get:

[m+1 ] = (Â − B̃ −1 C̃ D̃)[m ]. (75)

Thus the iteration matrix for FPSA with MPD is A − B −1 CD.

3.4 Comparison of Spectral Radii

In order to get a glimpse into how FPSA performs, we consider one problem with screened Rutherford kernel
(SRK) (Dixon, 2015), exponential kernel (EK) (Prinja et al., 1992) and Henyey-Greenstein kernel (HGK)
(Pomraning, 1992) each. We plot spectral radii of stand-alone SN and FPSA for each kernel in the figures
that follow. We choose L = 15, N = 16. We note that the spectral radius reduces significantly upon
introduction of FPSA. The spectral radius reduction is completely problem dependent. The spectral radius
can potentially change with N, L, and how close the scattering cross-section moments of the problem are to
the PL -equivalent moments.

13
Figure 1: Comparison of Source Iteration and FPSA - SRK - η = 2.836 × 10−5

Figure 2: Comparison of Source Iteration and FPSA - EK - ∆ = 10−5

14
 Figure 3: Comparison of Source Iteration and FPSA - HGK - g = 0.9

Next, we compare the numerically measured and theoretical (angularly-discrete Fourier analysis) spectral
radii. We analyze convergence rates for three scattering kernels - SRK, EK, and HGK. We choose L = 15,
N = 16. We use a slab of length, 100 cm, discretize it using 100 elements. We use vacuum boundaries for
numerical measurements of spectral radius. The theoretical and numerically measured spectral radii have
been presented in Table 1.

Kernel/Parameter ρMPD
FPSA -FA ρMPD
FPSA -Measured ρWFD
FPSA -FA ρWFD
FPSA -Measured
SRK/η = 2.83 × 10−5 0.4706 0.4706 0.2121 0.2120
EK/∆ = 10−5 0.1932 0.1954 0.6246 0.6327
HGK/g = 0.9 0.4304 0.4303 0.4177 0.4177

Table 1: Comparison of Numerical and Theoretical Spectral Radii

We obtain similar theoretical and measured spectral radii values for different scattering kernels with
varying parameters. This indicates a relatively accurate analysis of the method.

4 Efficiency Study
In this section we will assess how the reduction in spectral radius results in reduction in run-time of source
iteration (SI) and GMRES solves. We run all problems using MATLAB and track run-time using its tic-toc
functionality. We place tic and toc before and after the solver function calls respectively. In other words,
we do not include the stiffness matrix setup time in our calculation. We will only account for the solver
run-time. Specifically, choose problems with L = 15, and N = 16, 32. We use beam and vacuum boundaries.

15
We have a unit distributed source for problems with vacuum boundaries and a unit beam source with the
beam boundary. We do this for SRK with η = 2.83 × 10−5 , and for EK with ∆ = 10−5 . We solve the
Fokker-Planck error equation (invert the preconditioner) using LU factorization via factorize object (Davis,
2009) in MATLAB, and GMRES.

First, we compare unpreconditioned SI and GMRES solves. In order to compare these solves, we choose
η = 2.83 × 10−5 , L = 15, N = 16, H = 1cm, K = 100, tol = 10−10 . We do this to contrast source iteration
and GMRES solves. Table 2 and 3 present this data. It is clear that GMRES is more suitable than source
iteration for forward-peaked transport problems.

BC/Source Restart GMRES Iteration Count SI Iteration Count

Vacuum/Distributed > 150000
50 3305
100 2445
150 1875
200 1540
Beam/Zero > 150000
50 2602
100 2200
150 1895
200 1735

Table 2: SRK - Number of Iterations

BC/Source Restart GMRES Run-time SI Run-time

Vacuum/Distributed > 3000
50 64.97
100 50.41
150 37.76
200 32.71
Beam/Zero > 3000
50 50.68
100 43.99
150 41.29
200 36.11

Table 3: SRK - Solver Run-time [s]

Next, we compare solution rutimes and iteration counts. We will compare these for unpreconditioned
GMRES, FPSA-preconditioned SI, and FPSA-preconditioned GMRES solves. We do not include unpre-
conditioned source iteration in this study because its ineffectiveness for relevant problems has already been
demonstrated in Table 2 and 3. We will arbitrarily choose our restart parameter for this study to be 150.

4.1 Screened Rutherford Kernel

We compare efficiency of FPSA for problems involving SRK in this section. We choose a slab of unit length
discretized using hundred elements. We choose η = 2.83×10−5 and σa = 1. Scattering cross-section moments

16
come from SRK and their numerical values can be found in (Patel, 2016). Finally, L = 15, and N = 16 and
32. Number of iterations and overall run-time data has been presented in Table 4, 5, 6, and 7.

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 1487 9 6 14 10
15/32 1499 9 7 14 12
Factorize 15/16 9 7 14 10
15/32 9 8 14 12

Table 4: SRK - Vacuum Boundaries/Unit Distributed Source - Number of Iterations

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 28.75 8.79 6.15 12.01 5.98
15/32 28.12 36.76 18.36 54.13 19.77
Factorize 15/16 1.62 2.45 0.3373 0.2501
15/32 2.75 4.73 0.4244 0.3392

Table 5: SRK - Vacuum Boundaries/Unit Distributed Source - Run-time [s]

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 1357 12 8 21 13
15/32 1335 13 10 23 19
Factorize 15/16 12 9 21 13
15/32 13 11 23 19

Table 6: SRK - Beam Source - Number of Iterations

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 26.54 11.39 6.975 9.69 5.403
15/32 27 42.59 20.50 58.3 20.26
Factorize 15/16 1.687 2.547 0.4475 0.3074
15/32 2.927 5.061 0.611 0.4828

Table 7: SRK - Beam Source - Run-time [s]

We observe a significant decrease in the number of transport-sweeps required for convergence due to pre-
conditioning. Specifically, we observe upto three orders of magnitude decrease compared to unpreconditioned
GMRES and five orders of magnitude compared to SI. We also observe a decrease in overall solver run-times
due to preconditioning when FP-solve is done using LU factorization (by upto two orders of magnitude com-
pared to unpreconditioned GMRES). The FP-solve, however, can be extremely expensive and render this
preconditioner ineffective with respect to problem’s overall run-time if inefficient solvers are used. Here, the
number of iterations required for one FP-solve using GMRES was of the same order as an unpreconditioned
transport solve using GMRES. It is imperative that we find an effective preconditioner for FP-solves. We are
looking into this. The potential, however, of using FP as preconditioner for transport solves is amply evident
from the data presented in this section. Next, we look at efficiency data for problems with the exponential
kernel.

17
4.2 Exponential Kernel
We calculate scattering cross-section moments using EK for ∆ = 10−5 . The zeroth moment is calculated us-
ing SRK. The study is done using the same parameters as SRK except for scattering cross-section moments.
Number of iterations and overall run-time data has been presented in Table 8, 9, 10, and 11.

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 2217 7 8 9 15
15/32 2256 10 9 17 19
Factorize 15/16 7 9 9 15
15/32 10 10 17 19

Table 8: EK - Vacuum Boundaries/Unit Distributed Source - Number of Iterations

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 51.73 10.32 13.34 11.91 18.49
15/32 56.10 31.14 25.06 42.82 27.85
Factorize 15/16 2.3754 4.5586 0.3762 0.454
15/32 4.155 8.463 0.6098 0.6424

Table 9: EK - Vacuum Boundaries/Unit Distributed Source - Run-time [s]

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 2086 9 10 12 35
15/32 1932 14 12 24 28
Factorize 15/16 14 13 12 35
15/32 14 13 24 28

Table 10: EK - Beam Source - Number of Iterations

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 39.55 9.45 11.63 8.335 19.42
15/32 36.14 24.89 18.57 34.74 24.87
Factorize 15/16 2.842 6.657 0.2929 0.6853
15/32 2.898 6.799 0.6329 0.6585

Table 11: EK - Beam Source - Run-time [s]

We see similar behavior to what we saw in the case of SRK. The solver run-times differ due to difference
in rate at which FP-solve converges for this particular problem. Again, we note a significant decrease in
number of iterations but a decrease in solver run-time strongly depends on the efficiency of the FP-solve.

4.3 Henyey-Greenstein Kernel

In this section, we let the asymmetry parameter, g = 0.9999. The study is carried out in the same way
as the previously for SRK and EK. For this section, we will choose σa = 0.00001 cm−1 . The scattering
cross-section moments are calculated using HGK. We will choose slab length of 50 cm disretized using 200

18
elements. Number of iterations and overall run-time data has been presented in Table 12, 13, 14, and 15.

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 1150 10 7 28 18
15/32 1461 14 12 32 31
Factorize 15/16 10 8 28 18
15/32 14 13 32 31

Table 12: HGK - Vacuum Boundaries/Unit Distributed Source - Number of Iterations

Invert FP L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 83.66 288.1 160.9 931.9 373.4
15/32 102.5 1055 572.7 2618 1428
Factorize 15/16 6.390 12.27 2.193 1.385
15/32 11.75 28.820 2.651 2.5778

Table 13: HGK - Vacuum Boundaries/Unit Distributed Source - Run-time [s]

Invert FP L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 597 16 12 29 17
15/32 1634 21 19 32 31
Factorize 15/16 12 9 29 17
15/32 17 16 32 31

Table 14: HGK - Beam Source - Number of Iterations

FP-Solve L/N GMRES FPSAMPD

GMRES FPSAWFD
GMRES FPSAMPD
SI FPSAWFD
SI
GMRES 15/16 42.29 479.9 331.2 1040 285.7
15/32 115.1 1804 990.6 2408 1315
Factorize 15/16 6.795 12.41 2.131 1.316
15/32 12.59 29.69 2.945 2.452

Table 15: HGK - Beam Source - Run-time [s]

We note that, just like for SRK and EK, preconditioned schemes have significantly less iteration counts.
However, depending on how the Fokker-Planck error equation is solved, the preconditioning may or may
not be effective with respect to run-time reduction. Solving the FP equation with GMRES renders FPSA
scheme unviable, however use of factorization reduces to overall run-time significantly.

5 Summary and Future Work

We ran several numerical experiments and assessed the speed-ups in iteration count and solver run-time.
We saw that preconditioning transport solve using FP resulted in reduction in iteration count by upto three
orders (when compared to unpreconditioned GMRES solves). The overall run-time, however, depended
completely on how efficiently the FP preconditioner was solved. Direct factorization resulted in a run-time
reduction by upto two orders of magnitude. We observed that FP can be a very effective preconditioner

19
for transport solves with highly forward-peaked scattering. However, we must develop an effective solver
for FP-solve itself in order to make this an attractive preconditioning method. In future, we would like to
determine how do we optimize FP-solve. We would also like to test FPSA’s performance in energy dependent,
multi-D settings. Moreover, we would also like to develop a nonlinear version of this method which would
allow us not only to accelerate the convergence but also return a to obtain a Fokker-Planck equation that is
consistent with the relevant transport equation. For instance, approximations to the transport equation like
Fokker-Planck, Boltzmann-Fokker-Planck etc. do not return the same solution as the transport equation.
The novel nonlinear acceleration technique would return a modified Fokker-Planck equation that is discretely
consistent with the transport solve it is accelerating. Preliminary work on this has been presented in (Patel
et al., 2019). We plan to develop this method further.

Acknowledgments

This information has been co-authored by an employee or employees of the Los Alamos National Security,
LLC. (LANS), operator of the Los Alamos National Laboratory under Contract No. DE-AC52-06NA25396
with the U.S. Department of Energy.

References
1. C. Bogers, and E. W. Larsen “On Accuracy of Fokker-Planck and Fermi Pencil Beam for Charged
Particle Transport,” Medical Physics, (1996).
2. M. L. Adams and E. W. Larsen, “Fast Iterative Methods for Discrete Ordinates Particle Transport
Problems,” Progress in Nuclear Energy, (2002).
3. H. J. Kopp, “Synthetic Method Solution of the Transport Equation,” Nuclear Science and Engineering,
(1963).
4. M. Landesman and J. E. Morel, “Angular Fokker-Planck Decomposition and Representation Tech-
niques,” Nuclear Science and Engineering, (1989).
5. C. L. Leakeas and E. W. Larsen, “Generalized Fokker-Planck Approximations of Particle Transport
with Highly Forward-Peaked Scattering,” Nuclear Science and Engineering, (2001).

6. R. E. Alcouffe, “Diffusion Synthetic Acceleraton for Diamond-Differenced Discrete-Ordinates Equa-

tions,” Nuclear Science and Engineering, (1977).
7. K. Smith, D. A. Knoll, and H. Park, “Application of Jacobian Free Newton Krylov Method to Non-
linear Diffusion Acceleration of Transport Source Iteration in Slab Geometry,” Nuclear Science and
Engineering, (2011).
8. G. C. Pomraning, “The Fokker-Planck Operator as an Asymptotic Limit,” Mathematical Models and
Methods in Applied Sciences, (1992).
9. J. S. Warsa and A. K. Prinja, “A Moment Preserving SN Discretization for One-Dimensional Fokker-
Planck Equation,” Transactions of Americal Nuclear Society, (2012).

10. K. M. Khattab and E. W. Larsen, “Synthetic Acceleration Methods for Linear Transport Problems
with Highly Anisotropic Scattering,” Nuclear Science and Engineering, (1991).

20
11. D. Valougeorgis, M. Williams, and E. W. Larsen, “Stability Analysis of Synthetic Acceleration Methods
with Anisotropic Scattering,” Nuclear Science and Engineering, (1988).
12. C. M. Bender, and S. A. Orszag, Mathematical Methods for Scientists and Engineers - Asymptotic
Methods and Perturbation Theory, Mcgraw-Hill, New York, USA (1978).

13. E. E. Lewis, and W. F. Miller, Computational Methods of Neutron Transport, American Nuclear Society,
La Grange Park, USA (1993).
14. L. A. Hageman, and D. M. Young, Applied Iterative Methods, Academic Press, New York, USA (1981).
15. B. Trucksin, Acceleration Techniques for Discrete-Ordinates Transport Methods with Highly Forward-
Peaked Scattering, PhD Thesis, University of New Mexico, USA (2015).

16. D. A. Dixon, A Computationally Efficient Moment-Preserving Monte Carlo Transport Method with
Implementation in GEANT4, PhD Thesis, University of New Mexico, USA (2015).
17. J. K. Patel, Fokker-Planck-Based Acceleration for SN Equations with Highly Forward Peaked Scattering
in Slab Geometry, PhD Thesis, University of New Mexico, USA (2016).

18. J. E. Morel, “An Improved Fokker-Planck Angular Discretization Scheme,” Nuclear Science and Engi-
neering, (1989).
19. J. E. Morel, “Fokker-Planck Calculations Using Standard Discrete Ordinates Codes,” Nuclear Science
and Engineering, (1985).

20. J. E. Morel and T. A. Manteuffel, “An Angular Multigrid Acceleration Technique for SN Equations
with Highly Forward-Peaked Scattering,” Nuclear Science and Engineering, (1991).
21. E. N. Aristova and V. Ya. Gol’din, “Computation of Anisotropy Scattering of Solar Radiation in
Atmosphere (monoenergetic case),” Journal of Quantitative Spectropy and Radiative Transfer, (2000).
22. A. K. Prinja, G. C. Pomraning, and J. VanDenburg, “An Asymptotic Model for Spreading of a Colli-
mated Beam,” Nuclear Science and Engineering, (1992).
23. A. K. Prinja, and G. C. Pomraning “A Generalized Fokker-Planck Model for Transport of Collimated
Beams,” Nuclear Science and Engineering, (2001).
24. T. A. Davis, “Factorize: An Object Oriented Linear System Solver for MATLAB,” ACM Transcations
on Mathematical Software, (2009).
25. S. D. Pautz, Discrete Ordinates Transport Methods for Highly Forward Peaked Scattering, PhD Thesis,
Texas AM University, USA (2012).
26. J. K. Patel, J. J. Kuczek, and R. Vasques, “Nonlinear Fokker-Planck Acceleration for Forward-Peaked
Transport Problems in Slab Geometry”, Proceedings of The International Conference on Transport
Theory, September 22-27 2019, Paris, France.
27. C. D. Meyer and M. W. Stadelmaier, “Singular M-Matrices and Inverse Positivity”, Linear Algebra and
its Applications, (1978).

21