Available online at www.sciencedirect.

com
ScienceDirect

Mathematics and Computers in Simulation 159 (2019) 161–182

www.elsevier.com/locate/matcom

Original articles

Numerical simulations of spread of rabies in a spatially distributed

fox population
Khalaf M. Alanazia,b ,∗, Zdzislaw Jackiewicza,c , Horst R. Thiemea
a School of Mathematical and Statistical Sciences, Arizona State University, Tempe, AZ 85287, USA
b Northern Border University, Saudi Arabia
c Department of Applied Mathematics, AGH University of Science and Technology, Kraków, Poland

Received 18 February 2018; received in revised form 11 September 2018; accepted 13 November 2018
Available online 7 December 2018

Abstract
We describe a numerical algorithm for the simulation of the spread of rabies in a spatially distributed fox population. The
model considers both territorial and wandering rabid foxes and includes a latent period for the infection. The resulting systems are
mixtures of partial differential and integral equations. They are discretized in the space variable by central differences of second
order and by the composite trapezoidal rule. In a second step, they are discretized in time by explicit continuous Runge–Kutta
methods of fourth order for ordinary and delay differential systems. The results of the numerical calculations are compared for
latent periods of fixed and exponentially distributed length and for various proportions of territorial and wandering rabid foxes.
The speeds of spread observed in the simulations are compared to spreading speeds obtained by analytic methods and to observed
speeds of epizootic frontlines in the European rabies outbreak 1940 to 1980.
⃝c 2018 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
Keywords: Method of lines; Continuous Runge–Kutta method; Diffusing versus territorial rabid foxes; Latent period; Spreading speed

1. Introduction
In the last century, a rabies epizootic expanded over parts of Europe. It was mainly carried by foxes, started in
Poland in 1939 or 1940 and moved westward reaching Denmark in 1964, Belgium, Luxembourg, and Austria in 1966,
Switzerland in 1967, France in 1968, and Holland in 1974 [35]. It also moved to the east, to Hungary, the former
Czechoslovakia and the former Soviet Union [35]. Its speed ranged from 30 to 60 [km/year] according to [7,35] and
from 20 to 60 [km/year] according to [20]. There were important variations where the epidemic front retreated at
times in certain areas and moved up to 100 [km] in a given direction in other areas in a single year [35]. A study
performed in three areas in the state of Baden-Württemberg (Germany) from January 1963 to March 31, 1971, found
that the center of the frontwave moved at about 27 [km/year] [6] while the mean distance of new cases ahead of the
frontline within a month was approximately 4.8 [km] [6,21].
Since the rabies virus appears in the saliva one or two days before symptoms are evident in the infected fox [20, p.
247], the latent period is slightly shorter than the incubation period (transmission by biting). Fox rabies has a relatively
∗ Corresponding author.
E-mail addresses: kmalanaz@asu.edu (K.M. Alanazi), jackiewicz@asu.edu (Z. Jackiewicz), hthieme@asu.edu (H.R. Thieme).

https://doi.org/10.1016/j.matcom.2018.11.010
0378-4754/⃝ c 2018 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
162 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

long latent period with highly variable duration (from 12 to 110 days [2]) and a relatively short infection period (from
3 to 10 days [2]). While the literature seems to agree on the mean length of the infectious period (which is ended by
the death of the fox), 5 days, there seems to be disagreement on the mean length of the latent period: 35 days ([7]
and the references therein) and 28 days [21] and 28 to 30 days [2], and 25 and 26.5 days and one month according to
various sources cited in [35].
The above-mentioned fox epizootic has motivated quite a few mathematical studies [7,19,22,25,32]. Like this one,
they all assume that foxes that are susceptible or are in the latent period have home-ranges (unless they are migrating
juveniles looking for new home-ranges [25]). Differently from previous studies, our model assumes that some of the
rabid foxes essentially behave like susceptible and exposed foxes and keep their home-ranges, while the other rabid
foxes loose the attachment to their home-range and disperse by diffusion. We call the first ones territorial rabid
foxes and the second ones diffusing (wandering [35]) rabid foxes.
Mathematical models have so far assumed that either all rabid foxes are territorial [7] or all rabid foxes
diffuse [18,19,22,25]. The radio-tracking in [3] supports the existence of territorial rabid foxes, while the distances of
new rabies cases in [6, Table 2] seem to support the existence of diffusing foxes.
Since our model assumes that some rabid foxes are territorial and that the other rabid foxes diffuse, it is a system
of both partial differential and integral equations (with integration over space and, occasionally, also over time). We
propose the following methods for the numerical approximation of solutions. The partial differential and integral
equations are discretized in the space variable by central differences of second order and by the composite trapezoidal
rule. Next, they are discretized in time by explicit continuous Runge–Kutta methods of fourth order for ordinary and
delay differential systems. We also numerically solve equations for the spreading speed of the rabies epizootic that we
find analytically and compare them with the speeds observable in the numerical simulations. Our particular interest is
in how the partition of rabid foxes into territorial and diffusing rabid foxes influences the spreading speed, a question
we cannot answer by purely analytic means. We will restrict the numerical analysis to latent periods of fixed length
and to exponentially distributed latent periods.

2. A rabies model with diffusing and territorial rabid foxes

To study the spread of rabies in a spatially distributed fox population, we consider an open subset Ω of Rn which
represents the habitat of the foxes.
We consider an epidemic outbreak and assume that it is short enough that the natural turnover of the fox population
can be ignored: No foxes are born, and the only deaths are those of rabid foxes dying from rabies. While this
assumption certainly is an oversimplification made for the analytic methods used in [1] to work, the dissertation
by Hao Liu [19] about a model, in which all rabid foxes diffuse and foxes reproduce and also die from natural causes,
has some encouraging results that the spreading speed is similar to the minimal speed associated with wave-form
solutions.
Let S(x, t) denote the density of susceptible foxes (which are all territorial) at time t whose home-ranges center at
location x ∈ Ω . Further R1 (x, t) are the diffusing rabid foxes at location x and time t and R2 (x, t) the territorial rabid
foxes at time t whose home-ranges center at location x. Finally, L(x, t) are the foxes with home-range center at x at
time t that are in the latent period (they are all territorial), and I (x, t) is the transition rate at which foxes in the latent
period become infectious.
Let κ1 (x, z) denote the rate at which a fox with home-range center x visits the location z ∈ Ω . The rate at which a
susceptible fox with home-range center x meets a territorial rabid fox with home-range center z is given by
∫
κ2 (x, z) = κ1 (x, y)κ1 (z, y)dy, (2.1)
Ω
which means that it is the rate at which they both visit some common point y ∈ Ω (compare with Eq. (5.1) in [7]).
The model takes the form,
⎧ ∫
∂ κ κ
[ ]
⎪
⎪
⎪ t S(x, t) = −β S(x, t) 1 (x, z)R 1 (z, t) + 2 (x, z)R 2 (z, t) dz
Ω
⎪
⎪
⎪
⎪
⎨ =: −B(x, t),
(2.2)
∂t R1 (x, t) = D∇x2 R1 (x, t) + p1 I (x, t) − ν1 R1 (x, t),
⎪
⎪
⎪
⎪
⎪
⎪
∂t R2 (x, t) = p2 I (x, t) − ν2 R2 (x, t),
⎪
⎩
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 163

with given initial conditions

S(x, 0) = S0 (x), R1 (x, 0) = R1◦ (x), R2 (x, 0) = R2◦ (x), x ∈ Ω . (2.3)
If Ω ̸= Rn , the partial differential equation for R1 is accompanied by boundary conditions.
The parameters ν1 > 0 and ν2 > 0 are the per capita rabies death rates of diffusing and territorial rabid foxes,
respectively. p1 is the chance of a rabid fox to diffuse, and p2 the chance to be territorial, p j ≥ 0 and p1 + p2 = 1.
The radio-tracking in [3] supports the existence of territorial rabid foxes, while the distances of new rabies cases in [6,
Table 2] seem to support the existence of diffusing foxes. As for many infectious diseases, different individuals react
to the same disease in diverse ways. The easiest way to model the two types of behavior appears to assign them at
random. β > 0 is the rate at which the meeting of a susceptible and rabid fox leads to the infection of the susceptible
fox. B(x, t) is the incidence of the disease, i.e., the number of new cases per unit of time. D > 0 is the diffusion
∂x2i the Laplace operator. The nonnegative continuous functions S0 , R1◦ and R2◦ are the initial
∑n
constant and ∇x2 = i=1
densities of the susceptible and diffusing and territorial rabid foxes. Later, we will assume that S0 is constant. The
description of the model continues in the next subsections where the equations for I will depend on how the length of
the latent period is distributed.

2.1. Latent period with exponentially distributed length

Let θ > 0 be the constant per capita rate at which infected foxes transit from the latent to the rabid (infectious)
stage. In this case, we formulate a submodel for the infected foxes in the latent stage the number of which, at location
x and time t, is denoted by L(x, t). Since foxes are infected at rate B(x, t) and do not die during the latent period, we
have
∂t L(x, t) = B(x, t) − θ L(x, t) = −∂t S(x, t) − θ L(x, t). (2.4)
As before, we assume that there are no foxes in the latent period at time 0, i.e.,
L(x, 0) = 0. (2.5)
Since the rate of change of L(x, t) in time is the difference of the entry and exit rates of the latent stage and B(x, t) is
the entry rate, we obtain
I (x, t) = θ L(x, t). (2.6)
By (2.4), I satisfies the differential equation,
∂t I (x, t) = θ B(x, t) − θ I (x, t) = −θ ∂t S(x, t) − θ I (x, t), (2.7)
with the initial condition I (x, 0) = 0. We combine (2.7) and I (x, 0) = 0 with (2.2), which lead to the model
⎧ ∫
⎪ ∂t S(x, t) = −β S(x, t) κ1 (x, z)R1 (z, t) + κ2 (x, z)R2 (z, t) dz
[ ]
⎪
⎪
⎪
⎪
⎪ Ω
=: −B(x, t),
⎪
⎪
⎪
⎪
⎨
⎪ ∂t R1 (x, t) = D∇x2 R1 (x, t) + p1 I (x, t) − ν1 R1 (x, t), (2.8)
⎪
⎪
∂t R2 (x, t) = p2 I (x, t) − ν2 R2 (x, t),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
∂t I (x, t) = θ B(x, t) − θ I (x, t),
⎩

with given initial conditions

S(x, 0) = S0 (x), R1 (x, 0) = R1◦ (x), R2 (x, 0) = R2◦ (x), I (x, 0) = I0 (x), x ∈ Ω .

2.2. Latent period of fixed length

If τ > 0 is the fixed length of the latent period, then the individuals transiting from the latent to the infectious state
at time t > τ are exactly those that were infected τ time units ago, at time t − τ ,
I (x, t) = B(x, t − τ ) = −∂t S(x, t − τ ), t > τ. (2.9)
164 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

For simplicity of exposition and numerical calculation, we assume that the epizootic is started at time 0 by rabid foxes
with no foxes being in the latent stage, i.e.,
I (x, t) = 0, 0 ≤ t < τ. (2.10)
By combining (2.9) and (2.10) with (2.2), we obtain the model
⎧ ∫
∂ κ κ
[ ]
⎪
⎪
⎪ t S(x, t) = −β S(x, t) 1 (x, z)R 1 (z, t) + 2 (x, z)R 2 (z, t) dz
Ω
⎪
⎪
⎪
⎪
⎨ =: −B(x, t),
(2.11)
∂t R1 (x, t) = D∇x2 R1 (x, t) + p1 B(x, t − τ ) − ν1 R1 (x, t), t >τ
⎪
⎪
⎪
⎪
⎪
⎪
∂t R2 (x, t) = p2 B(x, t − τ ) − ν2 R2 (x, t),
⎪
⎩

with given initial conditions

S(x, 0) = S0 (x), R1 (x, 0) = R1◦ (x), R2 (x, 0) = R2◦ (x),
x ∈ Ω , and B(x, t − τ ) = 0 when 0 ≤ t < τ.

2.3. Latent periods with general length distribution

Let L(x, t) be again the number of foxes in the latent stage. Since foxes are infected at rate B(x, t) and, by
assumption, do not die during the latent period,
∂t L(x, t) = B(x, t) − I (x, t), (2.12)
where I (x, t) is the transition rate of foxes from the latent period to the rabid period appearing in (2.2).
Let Υ (a) be the probability that an infected fox is still in the latent stage a time units after infection. Then Υ is
decreasing and Υ (0) = 1.
We again assume that there are no foxes in the latent stage at time 0; so all foxes in the latent stage at time t > 0
have been infected at some t − a with 0 ≤ a ≤ t, and
∫ t
L(x, t) = B(x, t − a)Υ (a)da. (2.13)
0
After a substitution,
∫ t
L(x, t) = B(x, s)Υ (t − s)ds. (2.14)
0
Assume for a moment that Υ is continuously differentiable. By Leibniz differentiation rule, L has partial derivatives
with respect to time and we have
∫ t ∫ t
∂t L(x, t) = B(x, t) + B(x, s)Υ ′ (t − s)ds = B(x, t) + B(x, t − a)Υ ′ (a)da.
0 0
If Υ is not continuously differentiable, one can show that L(x, t) is absolutely continuous with respect to t and
∫ t
∂t L(x, t) = B(x, t) + B(x, t − a)dΥ (a),
0
for almost all t > 0, where the integral is a Stieltjes integral [33, Lem.B.29]. By (2.12),
∫ t
I (x, t) = − B(x, t − a)dΥ (a). (2.15)
0
This formula needs to be added to (2.2), if a latent period with general length distribution is considered.
If the latent period has a fixed length τ ,
1, 0 ≤ a < τ,
{
Υ (a) = (2.16)
0, a > τ,
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 165

and Υ (τ ) can be any number in [0, 1]. By (2.14),

∫ t
L(t, x) = B(x, s)ds, 0 ≤ t ≤ τ,
0
∫ t
(2.17)
L(t, x) = B(x, s)ds, t > τ.
t−τ

L is differentiable for 0 ≤ t ̸= τ , and

∂t L(x, t) =B(x, t), 0 ≤ t < τ,
(2.18)
∂t L(x, t) =B(x, t) − B(x, t − τ ), t > τ.
By (2.12),
I (x, t) =0, 0 ≤ t < τ,
(2.19)
I (x, t) =B(x, t − τ ), t > τ,
the same equations as (2.9) and (2.10).
If the latent period has exponentially distributed length with constant exit rate θ , then
Υ (a) = e−θa , (2.20)
and (2.14) is the solution to (2.4).

3. Spreading speeds

Assume that Ω = Rn and that S0 is constant and κ1 (x, z) = κ̃1 (|x − z|) ≥ 0 where | · | is the Euclidean norm on
n
R and
∫ ∫
κ̃1 (|y|)dy = 1, e−λy1 κ̃1 (|y|)dy⟨∞, λ⟩0. (3.1)
Rn Rn
Also assume that the length of the latent stage is arbitrarily distributed and Υ (a) is the probability that an infected fox
is still in the latent stage a time units after infection.
Let
S0
u(x, t) = ln , x ∈ Rn , t ≥ 0. (3.2)
S(x, t)
It can be shown [1] that u is nonnegative function and satisfies a scalar Volterra–Hammerstein integral equation
∫ t∫
u(x, t) = ξ (s, |x − y|)F(u(y, t − s))dsdy + u 0 (x, t). (3.3)
0 Rn
Here,
∫ s( ∫
ξ (s, |y|) = − β S0 p1 Γn (D(s − r ), |y − z|)e−ν1 (s−r ) κ1 (|z|) dz
0 Rn (3.4)
)
+ p2 e κ2 (|y|) dΥ (r ),
−ν2 (s−r )

∫ t ∫ (∫
u 0 (x, t) =β Γn (Ds, |y − z|)κ1 (|z|)R1◦ (y)e−ν1 s dy
0 Rn Rn (3.5)
)
+ κ2 (|z|)R2◦ (z)e−ν2 s dzds,

and F(u) = 1 − e−u , see also [4,8,9,15–17,30]. Recall that we have assumed for simplicity that there are no foxes in
the latent period at time t = 0.
It can be proved by the contraction mapping theorem that the integral equation (3.3) has a unique nonnegative
solution [31, Theorem 2.2]. Since S(t, x) = S0 e−u(x,t) , we also have a unique solution to (2.2).
166 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

The space–time Laplace transform Ξ (c, λ) is given by

∫ ∞∫
Ξ (c, λ) = e−λ(cs+y1 ) ξ (s, |y|)dsdy. (3.6)
0 Rn
Like the standard Laplace transform, the space–time Laplace transform converts time, space and space–time
convolutions into products of Laplace transforms [34, Proposition 4.1]. For the special form of (2.2) with arbitrarily
distributed length of the latent stage, this yields
p1 κ̂1 (λ) p2 (κ̂1 (λ))2 )
( ∫ ∞
Ξ (c, λ) = − − β S0 e−λca dΥ (a),
ν1 + λc − λ2 D ν2 + λc 0
∫ (3.7)
κ̂1 (λ) = e−λx1 κ1 (|x|)d x.
Rn
If the latent period has the fixed length τ, see (2.16), then
∫ ∞
− e−λca dΥ (a) = e−λcτ . (3.8)
0
If the length of latent period is exponentially distributed and θ is the constant exit rate, see (2.20), then
θ
∫ ∞ ∫ ∞
− e−λca dΥ (a) = e−λca θ e−θa da = . (3.9)
0 0 θ + λc
The basic reproduction number of rabies is given by
∫ ∞∫ (p
1 p2 )
R0 = ξ (s, |y|)dsdy = + β S0 , (3.10)
0 Rn ν1 ν2
see [17,34]. Here, 1/ν j is the average time a rabid fox was available for infecting
( others before
) disease-inflicted death
if it is territorial or diffusing, respectively. The weighted average of these, p1 /ν1 + p2 /ν2 is the time available for
a typical rabid fox. S0 is the density of susceptible foxes available to be infected, and β is the transmission rate.

Theorem 3.1. Let u(x, t) = ln(S0 /S(x, t)). Then

(a) If R0 < 1, u(x, t) → 0 as |x| → ∞ uniformly for t → ∞.
(b) If R0 ≤ 1, then, for any c > 0,
sup u(x, t) → 0, t → ∞.
|x|≥ct

If R0 > 1, the spreading speed (aka asymptotic speed of spread) is defined by

c∗ := inf{c ≥ 0; ∃λ > 0 : Ξ (c, λ) < 1}, (3.11)
[9,10,31]. If R0 ≤ 1, we define c∗ := 0.

3.1. The properties of the spreading speed

Using the theory developed in [31,34], one can show the following results.

Theorem 3.2. Let R0 > 1 and so c∗ > 0 and let u(x, t) = ln(S0 /S(x, t)). Let R ◦j be continuous on Rn for j = 1,2.
(a) Assume that, for any λ > 0, there exists some γ > 0 such that R1◦ (x) + R2◦ (x) ≤ γ e−λ|x| for all x ∈ Rn . Then,
for any c > c∗ ,
lim u(x, t) → 0.
t→∞, |x|≥ ct

(b) Assume that R1◦ (x) + R2◦ (x) > 0 for some x ∈ Rn . Then, for any c < c∗ ,
lim inf u(x, t) ≥ u ∗ ,
t→∞ |x|≤ ct
∗
where u ∗ > 0 is the unique solution of R0 (1 − e−u ) = u ∗ .
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 167

If R0 > 1, then c∗ > 0 and the spreading speed satisfies the equation [1,7,34]
∂
Ξ (c∗ , λ) = 1, Ξ (c∗ , λ) = 0. (3.12)
∂λ
This system has a unique solution (c∗ , λ) which can be numerically computed by Newton’s method.

3.2. The importance of latent periods with fixed length

Among all arbitrary length distributions with mean length τ , the distribution with fixed length τ is associated with
the smallest spreading speed. This follows from (3.11) and the fact that Ξ (c, λ) in (3.6) is a decreasing function of c
and from Jensen’s inequality [33, Thm.B.35]. Since the exponential function is convex, by Jensen’s inequality,
∫ ∞ (∫ ∞ )
− e−λca dΥ (a) ≥ exp λca dΥ (a) = e−λcτ , (3.13)
0 0
where
∫ ∞ ∫ ∞
τ =− adΥ (a) = Υ (a)da (3.14)
0 0
is the mean length of the latent stage [33, Sec.12.2] (compare with Eq. (3.8)).

3.3. An estimate

Let Ξ1 be Ξ with p1 = 1 and p2 = 0 and Ξ2 be Ξ with p1 = 0 and p2 = 1 and c∗j be the spreading speed
associated with Ξ j by (3.11) with Ξ j replacing Ξ . Then Ξ = p1 Ξ1 + p2 Ξ2 . So, we have the following theorem.

Theorem 3.3. c∗ ≥ min{c1∗ , c2∗ }.

Proof. We can assume that c∗j > 0 for j = 1,2. Set c• = min{c1∗ , c2∗ }. Let c ∈ (0, c• ). By (3.11)Ξ j (c, λ) ≥ 1 for all
λ ≥ 0. So Ξ (c, λ) ≥ 1 for all λ ≥ 0. By (3.11), c∗ ≥ c. Since c ∈ (0, c• ) has been arbitrary, c∗ ≥ c• . □

Remark 3.4. It is an open problem whether c∗ ≤ max{c1∗ , c2∗ }.

3.4. A final specialization for the numerical calculation of the spreading speed

For the numerical computation of c∗ , we assume that the movements of territorial foxes about the center of their
home-range are normally distributed, i.e.,
2 /(4b)
κ1 (|x|) = Γn (x, b) = (4π b)−n/2 e−|x| , x ∈ Rn , (3.15)
where | · | is the Euclidean norm on Rn , b > 0, and Γn is the fundamental solutions associated with the differential
2
operator ∂t − ∆x for n space dimensions. Then κ̂1 (λ) = ebλ in (3.7), see [34, Proposition 4.2], and
2 2 ∫ ∞
( p1 ebλ p2 e2bλ )
Ξ (c, λ) = − + β S0 e−λca dΥ (a). (3.16)
ν1 + λc − λ2 D ν2 + λc 0
Notice that Ξ does not depend on the space dimension! This observation is important because our numerical
simulations will be in one space dimension while the foxes live in two dimensional space and the parameters b
and D need to be estimated from two-dimensional data.

4. Existence and uniqueness of solutions

If one wants to see how solutions of (2.2) look like, one cannot solve it on Rn but rather solves it on a bounded
domain Ω of Rn with sufficiently smooth boundary and Dirichlet boundary conditions. The auxiliary variable
u(x, t) = ln(S0 /S(x, t)) ≥ 0 still satisfies a Volterra–Hammerstein integral equation, but now of the form
∫ t∫
u(x, t) = ξ (x, y, s)F(u(y, t − s))dsdy + u 0 (x, t), (4.1)
0 Ω
168 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

where ξ and u 0 are given in (3.4) and (3.5), respectively, but now ξ and u 0 incorporate the Green’s function associated
with ∂t − D∆x and Dirichlet boundary conditions [11] rather than the fundamental solution. Again, this equation has
a unique nonnegative solution u by the contraction mapping theorem [31, Theorem 2.2] which implies that (2.2) with
Dirichlet boundary conditions has a unique solution.
To find the solution numerically, one does not use (4.1), but discretizes (2.2), first in space and then in time.
By the maximum principle, the Green’s function lies below the fundamental solution on its domain of definition.
This implies that the integral kernel ξ associated with the Green’s function lies below the integral kernel associated
with the fundamental solution. Since F(u) = 1 − e−u is an increasing function of u, monotone iteration shows that the
solution u of (4.1) lies below the solution u of (3.3). So the epidemic model on a bounded domain Ω with Dirichlet
boundary conditions shows a less severe epidemic outbreak than the epidemic model on Rn , and the spread of the
disease modeled on Ω is not as fast as the spread of the disease modeled on Rn .

5. Discretization in space

We consider Ω to be a bounded domain on R to represent the habitat of foxes. We approximate first problems (2.8)
and (2.11) by restricting the spatial variable x to a symmetric interval [−a, a], a > 0, and we define the spatial grid
xi = −a + i∆x, i = 0, 1, . . . , N + 1,
where ∆x = 2a/(N + 1), and N is a positive integer. We are dropping the tilde and the absolute value from κ1 and
κ2 . Similarly as in [14], we replace (2.8) by the discrete problem
∂ S(xi , t)
⎧
⎪ = Ba (xi , t)
∂t
⎪
⎪
⎪
⎪
⎪
⎪ ∫ a
⎪ =: −β S(xi , t) κ1 (xi − z)R1 (z, t) + κ2 (xi − z)R2 (z, t) dz,
⎪
⎪ ( )
⎪
⎪
⎪
⎪
⎪ −a
⎪
⎪
⎨ ∂ R1 (xi , t) ∂ 2 R1 (xi , t)
⎪
⎪
=D + p1 I (xi , t) − ν1 R1 (xi , t), (5.1)
⎪
⎪ ∂t ∂x2
⎪
⎪
∂ R2 (xi , t)
⎪
⎪
= p2 I (xi , t) − ν2 R2 (xi , t),
⎪
⎪
∂t
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩ ∂ I (xi , t) = −θ Ba (xi , t) − θ I (xi , t),
⎪
⎪
⎪
⎪
∂t
i = 1, 2, . . . , N , t ≥ 0, with initial conditions
S(xi , 0) = S0 (xi ), R1 (xi , 0) = R1◦ (xi ), R2 (xi , 0) = R2◦ (xi ), I (xi , 0) = I0 (xi ).
We replace (2.11) by the discrete problems
⎧ ∂ S(x , t)
i
⎪ = Ba (xi , t)
∂t
⎪
⎪
⎪
⎪
⎪
⎪ ∫ a
=: −β S(xi , t) κ1 (xi − z)R1 (z, t) + κ2 (xi − z)R2 (z, t) dz,
⎪
⎪ ( )
⎪
⎪
⎪
⎪
⎨ −a
(5.2)
∂ R1 (xi , t) ∂ 2 R1 (xi , t)
− ν1 R1 (xi , t),
⎪
= D
⎪
⎪
∂t ∂x2
⎪
⎪
⎪
⎪
⎪
⎪
⎩ ∂ R2 (xi , t) = −ν R (x , t),
⎪
⎪
⎪
⎪
2 2 i
∂t
i = 1, 2, . . . , N , 0 ≤ t ≤ τ , with initial conditions
S(xi , 0) = S0 (xi ), R1 (xi , 0) = R1◦ (xi ), R2 (xi , 0) = R2◦ (xi ),
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 169

and
⎧ ∂ S(x , t)
i
⎪ = Ba (xi , t)
∂t
⎪
⎪
⎪
⎪
⎪
⎪ ∫ a
κ1 (xi − z)R1 (z, t) + κ2 (xi − z)R2 (z, t) dz,
⎪ ( )
=: −β S(x, t)
⎪
⎪
⎪
⎪
⎪
⎨ −a
(5.3)
∂ R1 (xi , t) ∂ 2 R1 (xi , t)
− p1 B(xi , t − τ ) − ν1 R1 (xi , t),
⎪
=D
⎪
⎪
∂t ∂x2
⎪
⎪
⎪
⎪
⎪
⎪
⎩ ∂ R2 (xi , t) = − p B(x , t − τ ) − ν R (x , t),
⎪
⎪
⎪
⎪
2 i 2 2 i
∂t
i = 1, 2, . . . , N , t > τ , where the functions S(xi , t − τ ), R1 (xi , t − τ ), and R2 (xi , t − τ ), appearing in Ba (xi , t − τ ),
satisfy the problem (5.2) for τ ≤ t ≤ 2τ . We assume that (5.1)–(5.3) satisfy the following boundary conditions

R1 (−a, t) = w1 (t), R1 (a, t) = w2 (t), t ≥ 0,

where w1 (t) and w2 (t), are given functions. We discretize next the integrals in (5.1)–(5.3) by the composite trapezoidal
rule
∆x (
κ1 (x − x0 )w1 (t) + κ2 (x − x0 )R2 (x0 , t)
)
T (x, t, ∆x) =
2
N
∑
κ1 (x − xk )R1 (xk , t) + κ2 (x − xk )R2 (xk , t)
( )
+ ∆x
k=1

∆x (
κ1 (x − x N +1 )w2 (t) + κ2 (x − x N +1 )R2 (x N +1 , t) ,
)
+
2
and ∂ 2 R1 (xi , t)/∂ x 2 by the central finite differences of the second order

∂ 2 R1 (xi , t) R1 (xi−1 , t) − 2R1 (xi , t) + R1 (xi+1 , t)

≈ .
∂x2 ∆x 2
Introducing the notation

Si (t) = S(xi , t), R1,i (t) = R1 (xi , t), R2,i (t) = R2 (xi , t), Ii (t) = I (xi , t),

i = 0, 1, . . . , N + 1, (5.1) can be approximated by the system of ordinary differential equations of the form

= −β Si (t)T (xi , t, ∆x),

⎧ ′
S (t)
⎪ i
⎪
⎪
⎪
⎪
⎪
⎪
′ R1,i−1 (t) − 2R1,i (t) + R1,i+1 (t)
⎪ R1,i (t) = D
⎪
⎪
⎪
⎪
⎨ ∆x 2

⎪ + p1 Ii (t) − ν1 R1,i (t), (5.4)

⎪
⎪
⎪
⎪ R2,i (t) = p2 Ii (t) − ν2 R2,i (t),
′
⎪
⎪
⎪
⎪
⎪
⎪
⎪
= θβ Si (t)T (xi , t, ∆x) − θ Ii (t),
⎩ ′
Ii (t)

i = 1, 2, . . . , N , t ≥ 0, with initial conditions

Si (0) = S0 (xi ), R1,i (0) = R1◦ (xi ), R2,i (0) = R2◦ (xi ), Ii (0) = I0 (xi ),
170 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

i = 1, 2, . . . , N . Similarly, (5.2) can be approximated by the system of ordinary differential equations

= −β Si (t)T (xi , t, ∆x),
⎧ ′
⎪ Si (t)
⎪
⎪
⎪
⎪
⎨ R ′ (t) = D R1,i−1 (t) − 2R1,i (t) + R1,i+1 (t)
⎪
⎪
⎪
1,i
∆x 2 (5.5)
− ν1 R1,i (t),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩ ′
R2,i (t) = −ν2 R2,i (t),
i = 1, 2, . . . , N , 0 ≤ t ≤ τ , with initial conditions
Si (0) = S0 (xi ), R1,i (0) = R1◦ (xi ), R2,i (0) = R2◦ (xi ),
i = 1, 2, . . . , N , and (5.3) can be approximated by the system of delay differential equations of the form
= −β Si (t)T (xi , t, ∆x),
⎧ ′
⎪ Si (t)
⎪
⎪
⎪
⎪
⎨ R ′ (t) = D R1,i−1 (t) − 2R1,i (t) + R1,i+1 (t)
⎪
⎪
⎪
1,i
∆x 2 (5.6)
+ p1 β Si (t − τ )T (xi , t − τ, ∆x) − ν1 R1,i (t),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
R2,i (t) = p2 β Si (t − τ )T (xi , t − τ, ∆x) − ν2 R2,i (t),
⎩ ′

i = 1, 2, . . . , N , t ≥ τ , where the functions Si (t − τ ), R1,i (t − τ ), and R2,i (t − τ ), satisfy the problem (5.5) for
τ ≤ t ≤ 2τ . Observe that R1,1 ′ ′
(t) and R1,N (t) depend on R1,0 (t) and R1,N +1 (t), which are equal to
R1,0 (t) = w1 (t), R1,N +1 (t) = w2 (t), t ≥ 0.
These systems (5.4)–(5.6) will be solved by the embedded pair of continuous Runge–Kutta method of fourth order
and discrete Runge–Kutta method of the third order with s = 6 stages, which is used for local error estimation. The
discretization in time is described in the next section.

6. Discretization in time

Introducing the notation

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
S1 (t) R1,1 (t) R2,1 (t)
S(t) = ⎣ ... ⎦ , R1 (t) = ⎣ ..
. ⎦ , R2 (t) = ⎣
..
. ⎦,
⎢ ⎥ ⎢ ⎥ ⎢ ⎥

S N (t) R1,N (t) R2,N (t)

⎡ ⎤ ⎡ ⎤
⎡
I1 (t)
⎤ S(t) S(0)
⎢ .. ⎥ ⎢ R1 (t) ⎥ ⎢ R1 (0)
I (t) = ⎣ . ⎦ , y(t) = ⎢⎣ R2 (t) ⎦ , y0 = ⎣ R2 (0)
⎥,
⎥
⎥ ⎢
⎦
I N (t) I (t) I (0)
the system (5.4) can be written in a compact form
{
y ′ (t) = f y(t) ,
( )
t ≥ 0,
(6.1)
y(0) = y0 ∈ R4N .

Here, the function f : R4N → R4N is defined by the right hand side of the problem (5.4). Similarly, the systems (5.5)
and (5.6) can be written as the systems of ordinary and delay differential equations of the form

⎨ y (t) = f ( y(t) , 0 ≤ t ≤ τ,
⎧ ′ ( )
⎪
y (t) = g y(t), y(t − τ ) , t > τ,
′
)
(6.2)
y(0) = y0 ∈ R ,
⎪
⎩ 3N
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 171

where now
S(t − τ )
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
S(t) S(0)
y(t) = ⎣ R1 (t) ⎦ , y(t − τ ) = ⎣ R1 (t − τ ) ⎦ , y0 = ⎣ R1 (0) ⎦ .
R2 (t) R2 (t − τ ) R2 (0)
Here, the function f : R3N → R3N is defined by the right hand side of the problem (5.5) and the function
g : R3N × R3N → R3N by the right hand side of the problem (5.6).
The problems (6.1) and (6.2) will be solved by the explicit continuous Runge–Kutta method of fourth order with
s = 6 stages, and the embedded discrete Runge–Kutta method of the third order, which are used for the estimation of
local discretization errors of the method of order four. This embedded pair was proposed by Owren and Zennaro [26–
28], see also [5]. The coefficients of this embedded pair are given by the Butcher table
0
1 1
6 6
11 44 369
37 1369 1369
c A 11 3388
− 8349 8140
yh b(θ ) =
17 4913 4913 4913
,
13 36764 767 32708 210392
15
− 408375 1125
− 136125 408375
ŷn+1 b̂ 1697 50653 299693 3375
1 18876
0 116160 1626240 11648
yh (tn + θ h n ) b1 (θ ) b2 (θ) b3 (θ ) b4 (θ ) b5 (θ ) b6 (θ )
101 1369 11849
ŷn+1 363
0 − 14520 14520
0 0
with continuous weights bi (θ ) given by
⎧
866577 4 1806901 3 104217 2
b1 (θ ) = − θ + θ − θ + θ,
⎪
⎪
824252 618189 37466
⎪
⎪
⎪
b2 (θ ) = 0,
⎪
⎪
⎪
⎪
⎪
12308679 4 2178079 3 861101 2
⎪
⎪
b3 (θ ) = θ − θ + θ ,
⎪
⎪
⎪
⎪
⎪
⎨ 5072320 380424 230560
7816583 4 6244423 3 63869 2 (6.3)
b4 (θ ) = − θ + θ − θ ,
10144640 5325936 293440
⎪
⎪
⎪
⎪
⎪ 624375 4 982125 3 1522125 2
θ + θ − θ ,
⎪
b5 (θ ) = −
⎪
⎪
⎪
⎪
⎪
⎪ 217984 190736 762944
296 4 461 3 165 2
⎪
⎪
θ − θ + θ .
⎪
⎪ b6 (θ ) =
⎪
⎩ 131 131 131
To implement this pair for delay differential equations we have to compute the approximate solution yh (tn − τ ), where
tn − τ is not, in general, a grid point. To compute this approximate solution we search for the index q such that
tn − τ ∈ (tq , tq+1 ], and compute yh (tn − τ ) from the formula
yh (tn − τ ) = b1 (θ )F1,q + b2 (θ )F2,q + b3 (θ )F3,q + b4 (θ )F4,q + b5 (θ )F5,q + b6 (θ )F6,q
if 0 ≤ tn − τ ≤ τ , or
yh (tn − τ ) = b1 (θ )G 1,q + b2 (θ )G 2,q + b3 (θ)G 3,q + b4 (θ )G 4,q + b5 (θ )G 5,q + b6 (θ )G 6,q
if tn − τ > τ . Here, bi (θ ) are given by (6.3),
tn − τ − tq
θ= ∈ (0, 1],
tq+1 − tq
Fk,q = f tq + ck h q , yh (tq + ck h q ) ,
( )

G k,q = g tq + ck h q , yh (tq + ck h q ), yh (tq + ck h q − τ ) ,

( )

k = 1, 2, 3, 4, 5, 6, h q = tq+1 − tq , and c = [c1 , c2 , c3 , c4 , c5 , c6 ]T .

172 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

The embedded pair of Runge–Kutta methods was implemented in a variable stepsize environment with the
estimates of the local discretization errors computed according to the formula
EST(tn+1 ) =  ŷn+1 − yh (tn+1 ) .
 
2
Following [12,29] the initial stepsize h 0 was computed from the formula
TOL1/5
{ }
h 0 = min 0.01 τ, ,
∥ f (0, y0 )∥2
where TOL is the accuracy tolerance prescribed by the user of the code. Then for n = 0, 1, . . . , the stepsize h n from
tn to tn+1 = tn + h n is accepted if
EST(tn+1 ) ≤ TOL,
and a new stepsize h n+1 from tn+1 to tn+2 = tn+1 + h n+1 is computed from the formula
( )1/5
TOL
h n+1 = η h n ,
EST(tn+1 )
where η is a safety coefficient to avoid too many rejected steps. In our implementation of the code we have chosen
η = 0.8. If
EST(tn+1 ) > TOL,
the stepsize is rejected, and the computations are repeated with a halved stepsize h n /2.
The construction of embedded pairs of continuous and discrete Runge–Kutta methods employed in this section,
and their convergence and order properties, are discussed in [26–28] and in the monographs [5,13].

7. Numerical experiments
In this section, we present the results of numerical experiments for (5.4)–(5.6) with continuous Runge–Kutta
methods of order four implemented in a variable stepsize environment as described in Section 6. We choose all
parameter values, initial conditions, and boundary conditions to be the same for all models (5.4)–(5.6). We compute
approximations to these systems for x ∈ [−a, a], a = 50 [km], and t ∈ [t0 , tend ], t0 = 0 [day], tend = 180 [day], and
p1 = p2 = 0.5. We choose N = 79 unless otherwise specified.
We assume the Dirichlet boundary conditions at x = −a and x = a, of the form
R1 (−a, t) = w1 (t) = 0 [fox/km], R1 (a, t) = w2 (t) = 0 [fox/km], t ∈ [0, tend ],
and initial conditions
Si (0) = S0 (xi ), R1,i (0) = R1◦ (xi ), R2,i (0) = R2◦ (xi ), Ii (0) = I0 (xi ),
i = 1, 2, . . . , N , where the functions S0 (x), R1◦ (x), R2◦ (x), and I0 (x) are defined by
{
0.2 [fox/km], −5 ≤ x ≤ 5,
R1◦ (x) =
0 [fox/km], otherwise,
R2◦ (x) = 0 [fox/km], I0 (x) = 0 [fox/km], x ∈ [−a, a].
2
Murray et al. let S0 = 4.6 [fox/km ], as in Fig. 2 of [24]. So, since we are working on R, S0 = 4.6 [fox/km] will be
used. The diffusion coefficient D, which measures the distance that a rabid fox can cross after the onset of clinical
disease, is assumed to be 200 [km2 /year] since it was used frequently in [23,24]. The mean length of the infectious
periods of diffusing rabid foxes 1/ν1 and territorial rabid foxes 1/ν2 is 5 [day] [2,24]. If rabid foxes exit the latent
period with a fixed length τ , then we assume τ = 28 [day] [21]. If the latent period has exponentially distributed
length with constant exit rate θ , then we have θ = 1/τ [1/day]. We also assume
1 −z 2
κ1 (z) = Γ1 (z, b) = √ e 4b . (7.1)
4πb
After a substitution to (2.1), it follows that
∫ ∫
κ2 (z) = κ1 (y + z)κ1 (y)dy = Γ1 (y + z, b)Γ1 (y, b)dy. (7.2)
R R
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 173

By the Chapman–Kolmogorov equation,

κ2 (z) = Γ1 (z, 2b). (7.3)
We set R0 = 1 in the formula (3.10), so ST is defined by
1( ν1 ν2 ) ν
1
ST = = . (7.4)
β ν2 p 1 + ν1 p 2 β
Following Anderson et al. [2], we will use the formula in (7.4) and the parameters values above with ST =
1 [fox/km2 ] [2,24] to find an estimate of the disease transmission parameter β. That gives β = 73 [km2 /year]. Again,
since we are working on R, β = 73 [km/year] will be used. With this choice of parameters, R0 = 4.6 > 1, so rabies
is going to spread among the fox population.
Now, we are working to find an estimate to b. We assume that foxes are circling around the center of their home-
ranges. So, the areas of these circles are equal to the average territory size A. We let Γ2 (t, x) to be the fundamental
solution of ∂t − ∆x in two space dimensions. Then the mean maximum distance of a territorial fox from the center of
its home-range (the mean radius of its home-range) is given by
∫ ∫
|x|(4bπ )−1 e−|x| /(4b) d x,
2
r0 = |x|Γ2 (x, b)d x =
R2 R2

with | · | being the Euclidean norm in R2 . We use polar coordinates x = (r cos θ, r sin θ) with r ≥ 0 and θ ∈ [0, 2π ).
By the change of variables formula,
∫ ∞
r 2 e−r /(4b) dr.
2
r0 = (2b)−1
0

We substitute r = s2b1/2 ,
∫ ∞
2
r0 = 4b 1/2
s 2 e−s ds = b1/2 π 1/2 .
0
So the mean area of the home range is
A = πr02 = bπ 2 .
So b = Aπ −2 [km2 ]. A is between 2 and 8 [km2 ] according to [35], the average area is taken as 5 [km2 ] by Källén
et al. [18] and Murray et al. [24]. Then b is 5π −2 [km2 ] ≈ 0.506605918 [km2 ].
The results of numerical simulations on the model (5.4) are presented in Figs. 7.1–7.5. The graphs of approxi-
mations Ih (x, t) to I (x, t) are similar to that of R2,h (x, t) and are not presented here. When the latent period has
exponentially distributed length, the contour plot in Fig. 7.5 shows that rabies propagates with asymptotic spreading
speed
c♦ ≈ 81 [km/year].
The results of numerical simulations on the model (5.5) and (5.6) are presented in Figs. 7.7–7.11. In the case that
the latent period has fixed length, the contour plot in Fig. 7.11 depicts that rabies spreads with asymptotic speed
c♦ ≈ 43 [km/year].
We also present in Figs. 7.6 and 7.12 the stepsize pattern for the algorithm described in Section 6 for the accuracy
tolerances TOL = 10−3 , 10−6 , 10−9 , and 10−12 . On these figures the rejected steps are denoted by ‘×’.

8. Discussion
We have solved the system (3.12) numerically by using Newton’s method, where the numerical values of the
parameters are given in Section 7, see also Fig. 8.1. With this choice of parameters, R0 = 4.6 > 1, so rabies is going
to spread among the fox population.
When the length of the latent period is exponentially distributed, the unique solutions of the system (3.12) are
(c∗ , λ) ≈ (0.182245 [km/day], 0.773071) ≈ (66.5195 [km/year], 0.773071)
174 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

Fig. 7.1. Surface plot of approximation Sh (x, t) to S(x, t) of the model (5.4).

Fig. 7.2. Surface plot of approximation R1,h (x, t) to R1 (x, t) of the model (5.4).

when p1 = 0,
(c∗ , λ) ≈ (0.253677 [km/day], 0.522531) ≈ (92.592 [km/year], 0.522531)
when p1 = 0.5, and
(c∗ , λ) ≈ (0.288236 [km/day], 0.484747) ≈ (105.206 [km/year], 0.484747)
when p1 = 1. Also, for this case, the contour plots in Fig. 7.5 and in [1] depict that c♦ ≈ 61 [km/year] when p1 = 0,
c♦ ≈ 81 [km/year] when p1 = 0.5, and c♦ ≈ 97 [km/year] when p1 = 1.
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 175

Fig. 7.3. Surface plot of approximation R2,h (x, t) to R2 (x, t) of the model (5.4).

Fig. 7.4. Approximations Sh (x, t), R1,h (x, t), and R2,h (x, t) to S(x, t), R1 (x, t), and R2 (x, t) at specific times t0 = 0 (thick dashed line), t = 2,
28, 50, 70, 100, 120, 150 (thin lines), and tend = 180 (thick solid line) of the model (5.4). Here N = 199.

When the latent has a fixed length, the unique solutions of the system (3.12) are
(c∗ , λ) ≈ (0.0774794 [km/day], 1.20104) ≈ (28.28 [km/year], 1.20104)
when p1 = 0,
(c∗ , λ) ≈ (0.129782 [km/day], 0.548928) ≈ (47.3705 [km/year], 0.548928)
176 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

Fig. 7.5. Contour plots of approximations Sh (x, t) (top), R1,h (x, t) (middle), and R2,h (x, t) (bottom) to S(x, t), R1 (x, t), and R2 (x, t) of the model
(5.4).

Fig. 7.6. Variable stepsize pattern for the algorithm based on continuous Runge–Kutta method of fourth order applied to the system (5.4) with
N = 119 for TOL = 10−3 , 10−6 , 10−9 , and 10−12 . Rejected steps are denoted by ‘×’. Almost all rejected steps occur for TOL = 10−3 .

when p1 = 0.5, and

(c∗ , λ) ≈ (0.145169 [km/day], 0.529397) ≈ (52.9868 [km/year], 0.529397)
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 177

Fig. 7.7. Surface plot of approximation Sh (x, t) to S(x, t) of the models (5.5) and (5.6).

Fig. 7.8. Surface plot of approximation R1,h (x, t) to R1 (x, t) of the models (5.5) and (5.6).

when p1 = 1. The contour plots in Fig. 7.11 and in [1] demonstrate that c♦ ≈ 26 [km/year] when p1 = 0,
c♦ ≈ 43 [km/year] when p1 = 0.5, and c♦ ≈ 47 [km/year] when p1 = 1.
Therefore, for latent periods of exponentially distributed length and fixed length, the asymptotic speeds c∗ , which
we get by solving the system (3.12), are quite close to asymptotic speeds c♦ , which we get from the contour plots.
In addition, the asymptotic speeds c∗ and c♦ confirm what we have discussed in Section 4 that the epidemic model
on a bounded domain Ω with Dirichlet boundary conditions shows a less severe epidemic outbreak than the epidemic
model on Rn , and the spread of the disease modeled on Ω is not as fast as the spread of the disease modeled on Rn .
178 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

Fig. 7.9. Surface plot of approximation R2,h (x, t) to R2 (x, t) of the models (5.5) and (5.6).

Fig. 7.10. Approximations Sh (x, t), R1,h (x, t), and R2,h (x, t) to S(x, t), R1 (x, t), and R2 (x, t) at specific times t0 = 0 (thick dashed line), t = 2,
28, 50, 70, 100, 120, 150 (thin lines), and tend = 180 (thick solid line) of the models (5.5) and (5.6). Here N = 199.

Furthermore, the numerical results confirm for the spreading speeds c∗ and c♦ that the latent period with fixed length
gives the smallest spreading speeds (Section 3.2) and the estimates in Section 3.3 hold.
Our results show that the spreading speed is a decreasing function of the mean length of the latent period τ, and
the spreading speed is an increasing function of β, S0 , D and the durations 1/νi of the infectious periods for diffusing
and territorial rabid foxes and also of b if κ1 is given by (3.15) [1]. b is a measure of how far territorial foxes move
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 179

Fig. 7.11. Contour plots of approximations Sh (x, t) (top), R1,h (x, t) (middle), and R2,h (x, t) (bottom) to S(x, t), R1 (x, t), and R2 (x, t) of the
models (5.5) and (5.6).

Fig. 7.12. Variable stepsize pattern for the algorithm based on continuous Runge–Kutta method of fourth order applied to the systems (5.5) and
(5.6) with N = 119 for TOL = 10−3 , 10−6 , 10−9 , and 10−12 . Rejected steps are denoted by ‘×’.

away from the center of their home-ranges if this distance is normally distributed (2b is the variance of the normal
distribution in each direction), and S0 is the initial value of the density of susceptible foxes. A study in three areas in
180 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

Fig. 8.1. Spreading speed c∗ dependence on p1 . We use the parameters S0 = 4.6 [fox/km], D = 200 [km2 /year], 1/ν1 = 1/ν2 = 5 [day],
β = 73 [km/year], τ = 28 [day], θ = 1/28 [1/day], and b = 5/π 2 [km2 ] with the analytic formula Ξ (c, λ) in (3.7) and (3.9) for the latent period
with exponentially distributed length, and Ξ (c, λ) in (3.7) with (3.8) for the latent period with fixed length to solve the system (3.12) for different
values of p1 ∈ [0, 1].

the state of Baden-Württemberg of Germany finds that the mean distance of new rabies cases ahead of the monthly
determined rabies frontline very slightly decreases if the hunting indicator of the fox density (foxes shot per km2
per year) increases [6, Table 3]. The likely explanation is that the home-range size is not independent of fox density
but depends on it in a decreasing fashion as in [7, Sec.7.2] where b is assumed to be proportional to 1/S0 [7, (7.6)].
Differently from [6,7], we assume b and S0 to be independent.
Also, c∗ increases as we increase the proportion of wandering rabid foxes p1 for both cases, as demonstrated by
the numerical simulations shown in Fig. 8.1. The last happens for what we believe is a realistic choice of parameters
b and D. In general, the monotone behavior of c∗ as a function of p1 depends on the relation between b and D, as
depicted in Fig. 8.2.
Murray et al. [23] assume that all rabid foxes diffuse, foxes reproduce and die from natural causes, and rabid foxes
exit the latent period with exponentially distributed length. Hence, we compare the asymptotic speeds in [23] to the
ones we have by solving the system (3.12) when the latent period has exponentially distributed length and p1 = 1. The
unique solutions of the system (3.12) show that the asymptotic velocity of rabies spread c∗ ≈ 105.2 [km/year], while
the speeds of initial waves are about 103 [km/year] with carrying capacity of 4.6 [fox/km2 ] and β = 80 [km2 /year]
in [23, Table 2] when there are no immune rabid foxes. This suggests that ignoring the natural turnover of fox
population does not have a huge impact on the speed of rabies spread.
When p1 = 0, all rabid foxes are territorial, and the asymptotic speeds of spread c∗ that we obtain by solving
the system (3.12) for a latent period of fixed length can be compared with the asymptotic speeds in [7]. There are
differences in some assumptions and in the determination of parameters, though; for instance, it is assumed in [7]
that the sizes of the home-ranges decrease with fox density while we assume them to be independent. For a fox
population density S0 of 4.6 [fox/km], we obtain an asymptotic speed of rabies spread c∗ ≈ 28.3 [km/year], while
Fig. 7 in [7] shows an asymptotic speed of about 33 [km/year] when S0 = 4.6 [fox/km2 ]. Furthermore, for this case,
the asymptotic speed of rabies spread c∗ compares quite well with the observed speeds about 27 [km/year] in [6] and
from 20 to 60 [km/year] according to [20].
For both cases, the numerical simulations depict that the density of diffusing rabid foxes R1 decays when t ≤ τ ,
then it increases for a period of time when the infected foxes leave the latent period at t > τ, as shown in Figs. 7.2,
7.4, 7.8 and 7.10. The decrease in the density of diffusing rabid foxes is because of the death from rabies ν1 , which
gives the rabid foxes as few as five days on average to live [2,22–24]. The density of territorial rabid foxes R2 , on
the other hand, remains zero when t ≤ τ because we assume there are no territorial rabid foxes initially at time zero.
When t > τ , we see R2 grows again before it loses some of its members with rate ν2 , as demonstrated in Figs. 7.3,
7.4, 7.9 and 7.10. Since, by assumptions, no foxes are born and infected foxes cannot recover and become susceptible
again, the densities of susceptible foxes in both cases continue to decrease, as shown in Figs. 7.1, 7.4, 7.7 and 7.10.
 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182 181

Fig. 8.2. Different monotonicity in the dependence of the spreading speed c∗ on p1 . We assume ν1 = ν2 , then we use Ξ (c, λ) in (3.7) with (3.8)
to solve (3.12) for different values of p1 ∈ [0, 1]. Here, 1/ν1 = 5 [day], b = 5/π 2 [km2 ], then bν1 = 36.98 [km2 /year], (a) D = 30 [km2 /year]
and (b) D = 50 [km2 /year].

References
[1] K.M. Alanazi, A Rabies Model with Distributed Latent Period and Territorial and Diffusing Rabid Foxes (Dissertation), Arizona State
University, Tempe, 2018.
[2] R.M. Anderson, H.C. Jackson, R.M. May, A.M. Smith, Population dynamics of fox rabies in Europe, Nature 289 (1981) 765–771.
[3] L.M. Andral, M.F.A. Artois, A.J. Blancou, Radio-pistage de renard enrages, Comp Immunol Microbiol Infect Dis 5 (1982) 284–291.
[4] C. Beaumont, J.-B. Burie, A. Ducrot, P. Zongo, Propagation of Salmonella within an industrial hens house, SIAM J. Appl. Math. 72 (2012)
1113–1148.
[5] A. Bellen, M. Zennaro, Numerical Methods for Delay Differential Equations, Oxford Science Publications, Clarendon Press, Oxford, 2003.
[6] K. Bögel, H. Moegle, F. Knorpp, A. Arata, K. Dietz, P. Diethelm, Characteristics of the spread of a wildlife rabies epidemic in Europe, Bull
World Health Organ 54 (1976) 433–447.
[7] F. van den Bosch, J.A.J. Metz, O. Diekmann, The velocity of spatial population expansion, J. Math. Biol. 28 (1990) 529–565.
[8] O. Diekmann, Limiting behaviour in an epidemic model, Nonlinear Anal. TMA 1 (1977) 459–470.
[9] O. Diekmann, Thresholds and travelling waves for the geographical spread of infection, J. Math. Biol. 6 (1978) 109–130.
[10] O. Diekmann, Run for your life. A note on the asymptotic speed of propagation of an epidemic, J. Differential Equations 33 (1979) 5873.
[11] M.G. Garroni, J.L. Menaldi, Green Functions for Second Order Parabolic Integro-Differential Problems, Longman Scientific & Technical,
Essex, 1992.
[12] I. Gladwell, L.F. Shampine, R.W. Brankin, Automatic selection of the initial step size for an ODE solver, J. Comput. Appl. Math. 18 (1987)
175–192.
[13] E. Hairer, S.P. Nø rsett, G. Wanner, Solving Ordinary Differential Equations I: Nonstiff Problems, Springer-Verlag, New York, 1993.
[14] Z. Jackiewicz, H. Liu, B. Li, Y. Kuang, Numerical simulations of traveling wave solutions in a drift paradox inspired diffusive delay
population model, Math. Comput. Simulation 96 (2014) 95–103.
[15] D.A. Jones, G. Röst, H.L. Smith, H.R. Thieme, On spread of phage infection of bacteria in a petri dish, SIAM J. Appl. Math. 72 (2012)
670–688.
[16] D.A. Jones, H.L. Smith, H.R. Thieme, Spread of viral infection of immobilized bacteria, Netw. Heterog. Media 8 (2013) 327–342.
[17] D.A. Jones, H.L. Smith, H.R. Thieme, Spread of phage infection of bacteria in a petri dish, Discrete Contin. Dyn. Syst. Ser. B 21 (2016)
471–496.
[18] A. Källén, P. Arcuri, J.D. Murray, A simple model for the spatial spread and control of rabies, J. Theoret. Biol. 116 (1985) 377–393.
[19] H. Liu, Spatial Spread of Rabies in Wildlife (Dissertation), Arizona State University, 2013.
[20] H.G. Lloyd, The Red Fox, Batsford LTD, London, 1980.
[21] H. Moegle, F. Knorpp, K. Bögel, A. Arata, K. Dietz, P. Diethelm, Zur epidemiologie der wildtiertollwut, Zbl. Vet. Med. B 21 (1974) 647–659.
[22] J.D. Murray, Mathematical Biology, Springer, Berlin Heidelberg, 1989.
[23] J.D. Murray, W.L. Seward, On the spatial spread of rabies among foxes with immunity, J. Theoret. Biol. 156 (1992) 327–348.
[24] J.D. Murray, E.A. Stanley, D.L. Brown, On the spatial spread of rabies among foxes, Proc R Soc Lond [Biol] (1986) 111–150.
[25] C. Ou, J. Wu, Spatial spread of rabies revisited: Influence of age-dependent diffusion on nonlinear dynamics, SIAM J. Appl. Math. 67 (2006)
138–163.
[26] B. Owren, M. Zennaro, Order barriers for continuous explicit Runge–Kutta methods, Math. Comp. 56 (1991) 645–661.
[27] B. Owren, M. Zennaro, Continuous explicit Runge–Kutta methods, in: Computational Ordinary Differential Equations, (London 1989), in:
Inst. Math. Appl. Conf. Ser. New Ser, vol. 39, Oxford University Press, New York, 1992, pp. 97–105.
[28] B. Owren, M. Zennaro, Derivation of efficient continuous explicit Runge–Kutta methods, SIAM J. Sci. Stat. Comput. 13 (1992) 1488–1501.
182 K.M. Alanazi, Z. Jackiewicz and H.R. Thieme / Mathematics and Computers in Simulation 159 (2019) 161–182

[29] L.F. Shampine, M.K. Gordon, Computer Solution of Ordinary Differential Equations: The Initial Value Problem, W. H. Freeman, San
Francisco, 1975.
[30] H.R. Thieme, A model for the spatial spread of an epidemic, J. Math. Biology 4 (1977) 337–351.
[31] H.R. Thieme, Asymptotic estimates of the solutions of nonlinear integral equations and asymptotic speeds for the spread of populations, J.
Reine Angew. Math. 306 (1979) 94–121.
[32] H.R. Thieme, Some mathematical considerations of how to stop the spatial spread of a rabies epidemic, in: W. Jäger, H. Rost, P. Tautu (Eds.),
Biological Growth and Spread, in: Lecture Notes in Biomathematics, vol. 38, Springer, 1980, pp. 310–319.
[33] H.R. Thieme, Mathematics in Population Biology, Princeton University Press, Princeton, 2003.
[34] H.R. Thieme, X.-Q. Zhao, Asymptotic speeds of spread and traveling waves for integral equations and delayed reaction–diffusion models, J.
Differential Equations 195 (2003) 430–470.
[35] B. Toma, L. Andral, Epidemiology of fox rabies, Adv Virus Res 21 (1977) 1–36.