Вы находитесь на странице: 1из 13

Mathematics and Computers in Simulation 54 (2000) 169181

Numerical and bifurcation analyses for a population


model of HIV chemotherapy
A.B. Gumel
a,
, E.H. Twizell
b
, P. Yu
c
a
Department of Mathematics, University of Manitoba, Winnipeg, Man., Canada R3T 2N2
b
Department of Mathematical Sciences, Brunel University, Uxbridge, Middlesex UB8 3PH, UK
c
Department of Applied Mathematics, University of Western Ontario, London, Ont., Canada N6A 5B7
Received 1 August 2000; accepted 11 August 2000
Abstract
A competitive implicit nite-difference method will be developed and used for the solution of a non-linear
mathematical model associatedwiththe administrationof highly-active chemotherapytoanHIV-infectedpopulation
aimed at delaying progression to disease. The model, which assumes a non-constant transmission probability,
exhibits two steady states; a trivial steady state (HIV-infection-free population) and a non-trivial steady state
(population with HIV infection). Detailed stability and bifurcation analyses will reveal that whilst the trivial steady
state only undergoes a static bifurcation (single zero singularity), the non-trivial steady state can not only exhibit
static and dynamic (Hopf) bifurcations, but also a combination of two types of bifurcation (a double zero singularity).
Althoughthe GaussSeidel-type method to be developedinthis paper is implicit byconstruction, it enables the var-
ious sub-populations of the model to be monitored explicitly as time t tends to innity. Furthermore, the method will
be seen to be more competitive (in terms of numerical stability) than some well-known methods in the literature. The
method is used to determine the impact of the chemotherapy treatment by comparing the population sizes at equilib-
rium of the treated and untreated infecteds. 2000 IMACS. Published by Elsevier Science B.V. All rights reserved.
Keywords: Finite-difference; Chemotherapy; Critical points; Bifurcation
1. Introduction
The Human Immunodeency Virus (HIV) is the causative agent of Acquired Immunodeency Syn-
drome in humans (AIDS). HIV is largely transmitted sexually, by IV drug use or through mother-to-child
transmission. Over the past decade, substantial progress has been made in terms of HIV drugs develop-
ment (see, for instance, [7]). Thanks to such advances, HIV replication and most opportunistic infections
can be controlled without eliminating the virus using highly-active anti-retroviral drugs (HAART) [5].

Corresponding author.
E-mail addresses: gumelab@cc.umanitoba.ca (A.B. Gumel), masteht@brunel.ac.uk (E.H. Twizell), pyu@pyu1.apmaths.uwo.ca
(P. Yu).
0378-4754/00/$20.00 2000 IMACS. Published by Elsevier Science B.V. All rights reserved.
PII: S0378- 4754( 00) 00222- 6
170 A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181
Because of the ever-increasing numbers of reported cases of HIV infection and AIDS, much collabo-
rative research is being conducted by mathematicians, biologists and physicians with the hope of gaining
better insight into the transmission dynamics of HIV and designing effective methods of control (see
[310,12,14]).
Anderson et al. [3], Kiessling et al. [8] and Knox [10] developed epidemiological models which break
down the total population into sub-groups according to their sexual preferences and rate of partner change.
Dietz et al. [4] generalized the classical Susceptible-Infective-Susceptible (SIS) models (that is, models
without immunity) by introducing heterogeneous contact and pairing rates. Ho et al. [7], Kirschner and
Webb [9] and Wei et al. [14] studied the dynamics of HIV infection in patients receiving drug treatment.
McLean et al. [12] provided a mathematical explanation of the co-existence between drug-resistant and
drug-sensitive virus in treated HIV patients.
Most of these studies, however, appear to give scant or no detail of the numerical methods used, together
with their associated stability analyses, to solve the resulting initial value problems (IVPs).
This paper focusses on the bifurcation and stability of the equilibrium states, and the development of a
GaussSeidel-like, implicit, nite-difference scheme (see [11,13]), for a modied homogeneous model
for the administration of anti-viral therapy developed by Gupta et al. [6] with a non-constant probability
of transmission. The model was obtained by sub-dividing the infected and uninfected populations into
treated and untreated categories. The type of treatment used was the administration of chemotherapy to
infectious persons (which acts to delay progression to disease). Chemotherapy here refers to any drug
treatment that is effective in the control of HIV replication, most importantly, HAART-combination
therapy. The model enables the populations of susceptibles, together with treated and untreated infecteds,
to be monitored separately as t .
It is shown that the model exhibits two steady states, one of which can only have a static bifurcation,
jumping to the other equilibrium solution and, consequently, the two states exchange their stability
properties on the same static stability boundary. The second equilibrium solution, on the other hand,
can have both a single zero singularity and a Hopf singularity, and may also have a co-dimension two
singularity (that is, a double zero bifurcation).
It is worth mentioning that discretizing the ordinary differential equations (ODEs) of the model by
explicit Euler nite-difference schemes, can result in contrived chaos whenever the discretization pa-
rameters exceed certain values [11,13]. The same applies to the use of explicit RungeKutta methods.
Although chaos can often be avoided, even for Euler methods, by using small time steps, the extra com-
puting costs incurred when examining the long-termbehaviour of a dynamical systemmay be substantial.
It is, therefore, essential to use a numerical method which allows the largest possible time steps that are
consistent with stability and accuracy.
Toavoidcontrivedchaos, whilst retainingaccuracyandnumerical stability, it maybe necessarytoforego
the ease-of-implementation of inexpensive, explicit, numerical methods in favour of implicit methods
(which are, at times, computationally-intensive). The novelty of the implicit nite-difference scheme to
be developed in this paper is that it is easy to implement, fast-convergent (expected of GaussSeidel-type
methods) and has very competitive stability properties.
2. Mathematical model
Since AIDS has no cure at the moment, anti-viral therapy is merely aimed at containing viral multipli-
cation, thereby delaying progression to disease. Following Anderson et al. [3] and Gupta et al. [6], the
A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181 171
infected and uninfected populations will be sub-divided into treated and untreated categories. The type
of treatment to be considered is effective chemotherapy (HAART) administered to infectious persons. It
is worth mentioning that in this study, the sole mode of HIV transmission is via sexual contact with an
infected individual.
2.1. S
u
(t): susceptibles
The sexually-active population, N
0
, where is the inverse of the average duration of sexual activity
and N
0
is the stable population size in the absence of infection, is recruited into this class. The susceptible
population is diminished both by natural rate of cessation of sexual activity (S
u
) and by infection which
may be acquired from each new sexual partner with probability .
Following Gupta et al. [6], the per partnership probability of transmission is given by
=

1
Y
u
+
2
Y
v
N
, (1)
where
1
is the probability of transmission from an untreated carrier of HIV,
2
is the probability of
transmission froman infected individual who has been subjected to therapy, and N = S
u
(t )+Y
u
(t )+Y
v
(t )
is the current size of the sexually active population. This suggests the non-linear IVP
dS
u
dt
f
1
(S
u
, Y
u
, Y
v
) = N
0
S
u
c
_

1
Y
u
N
+

2
Y
v
N
_
S
u
; t > t
0
, S
u
(t
0
) = S
0
u
, (2)
2.2. Y (t): infected individuals
These are categorized in two classes. Those who have received chemotherapy (drug treatment, for
example HAART), and those who are not receiving drug treatment.
2.2.1. Y
u
(t): untreated infected
This population increases through the infection of susceptibles (S
u
), and is diminished by the devel-
opment of AIDS at a per capita rate . It is further diminished by the administration of therapy at a per
capita rate .
2.2.2. Y
v
(t): infected receiving therapy
This increases by the application of therapy to the untreated population. It is assumed that chemotherapy
slows the rate of progression to AIDS. Thus, treated infecteds develop AIDS at a slower per capita rate
d (assumed to be less than the magnitude of ). Thus,
dY
u
dt
f
2
(S
u
, Y
u
, Y
v
) = c
_

1
Y
u
N
+

2
Y
v
N
_
S
u
( + +)Y
u
; t > t
0
, Y
u
(t
0
) = Y
0
u
, (3)
and
dY
v
dt
f
3
(S
u
, Y
u
, Y
v
) = Y
u
(d +)Y
v
; t > t
0
, Y
v
(t
0
) = Y
0
v
. (4)
In (2)(4), the initial conditions S
0
u
, Y
0
u
, Y
0
v
are needed for a numerical method to generate the solution
sequences {S
0
u
, S
1
u
, S
2
u
, . . . }, {Y
0
u
, Y
1
u
, Y
2
u
, . . . } and {Y
0
v
, Y
1
v
, Y
2
v
, . . . }.
172 A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181
It should be noted that the total population size is N = S
u
(t ) +Y
u
(t ) +Y
v
(t ) and that
dN(t )
dt
= N
0
S
u
(t ) ( +)Y
u
(t ) (d +)Y
v
(t ). (5)
It can be seen from (5) that in an infection-free population (Y
u
= Y
v
= 0), the steady-state value is
N
0
= S
u
(t ). The IVP (2)(4) is a modied version of the model presented by Gupta et al. [6] with their
S
v
, the population of vaccinated susceptibles, set to 0.
3. Stability analysis
3.1. Critical points
For convenience, let

1
= +d,
2
= + +, a
1
= c
1
, a
2
= c
2
, (6)
then Eqs. (2)(4) can be rewritten as
dS
u
dt
= N
0
S
u

S
u
(a
1
Y
u
+a
2
Y
v
)
S
u
+Y
u
+Y
v
,
dY
u
dt
=
S
u
(a
1
Y
u
+a
2
Y
v
)
S
u
+Y
u
+Y
v

2
Y
u
,
dY
v
dt
= Y
u

1
Y
v
. (7)
Setting (dS
u
/dt ) = (dY
u
/dt ) = (dY
v
/dt ) = 0 yields two critical points (equilibrium solutions) namely:
the trivial critical point (no infected population)
S
(1)
u
= N
0
, Y
(1)
u
= Y
(1)
v
= 0, (8)
and the non-trivial critical point
S
(2)
u
=
N
0
(
1
+)
W
, Y
(2)
u
=
N
0
(a
1

1
+a
2

1

2
)

2
W
,
Y
(2)
v
=
N
0
(a
1

1
+a
2

1

2
)

2
W
, (9)
where
W = ( +a
2
) +
1
(a
1
+
2
). (10)
By making the realistic assumption that all the parameters , , , d, c,
1
, and
2
, are positive, it then
follows that
1
,
2
, a
1
, and a
2
, are also positive.
3.2. Stability and bifurcation analyses of the critical points
To obtain the stability conditions for the two equilibrium solutions and possible bifurcations from the
equilibria, the Jacobian (J) given by
A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181 173
J =
_
_
_
_
_
_
_
_

(Y
u
+Y
v
)(a
1
Y
u
+a
2
Y
v
)
(S
u
+Y
u
+Y
v
)
2

S
u
[a
1
S
u
+(a
1
a
2
)Y
v
]
(S
u
+Y
u
+Y
v
)
2

S
u
[a
2
S
u
(a
1
a
2
)Y
u
]
(S
u
+Y
u
+Y
v
)
2
(Y
u
+Y
v
)(a
1
Y
u
+a
2
Y
v
)
(S
u
+Y
u
+Y
v
)
2
S
u
[a
1
S
u
+(a
1
a
2
)Y
v
]
(S
u
+Y
u
+Y
v
)
2

2
S
u
[a
2
S
u
(a
1
a
2
)Y
u
]
(S
u
+Y
u
+Y
v
)
2
0
1
_

_
,
(11)
must be evaluated at the critical points.
3.2.1. Stability of the trivial critical point
Evaluating J at the trivial critical point (8) gives
J
(1)
=
_
_
_
a
1
a
2
0 a
1

2
a
2
0
1
_

_
. (12)
The characteristic equation of the Jacobian J
(1)
, can be easily obtained as
( +)[
2
+(
1
+
2
a
1
) +
1
(
2
a
1
) a
2
] = 0. (13)
Clearly, one of the eigenvalues is = < 0 and thus the stability of this equilibrium depends upon
the other two zeros of the quadratic factor,
2
+ (
1
+
2
a
1
) +
1
(
2
a
1
) a
2
. Thus, when the
following conditions

1
+
2
a
1
> 0,
1
(
2
a
1
) a
2
> 0, (14)
are satised, the real parts of the two eigenvalues are negative, and, therefore, the trivial critical point (8)
is stable.
3.2.2. Stability of the non-trivial critical point
Similarly, evaluating the Jacobian (11) at the non-trivial critical point (9) gives
J
(2)
=
_
_
_
J
11
J
12
J
13
J
21
J
22
J
23
0
1
_

_
, (15)
where
J
11
=
(
1
a
1
+a
2
)
2
+
1
a
1
[(
1
+)(
1
+)
1

2
] +a
2
[(
1
+)(
1
+) +
2
]
W
1
,
J
12
=

1

2
[(a
1
a
2
) +
1

2
]
W
1
, J
13
=

2
1
(a
1
a
2

2
)
W
1
, J
21
=
[
1
(a
1

2
) +a
2
]
2
W
1
,
J
22
=

2
[a
2
( +
1
)
2

2
1
]
W
1
, J
23
= J
13
, W
1
= ( +
1
)(a
1

1
+a
2
), (16)
174 A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181
with the associated characteristic polynomial
P() =
3
+p
1

2
+p
2
+p
3
. (17)
In (17),
p
1
=
1
J
11
J
22
, p
2
= J
13

1
(J
11
+J
22
) +J
11
J
22
J
12
J
21
,
p
3
= (J
11
J
22
J
12
J
21
)
1
J
13
(J
11
+J
21
). (18)
Using RouthHurwitz stability criteria [11], the equilibrium solution (9) is stable provided
p
1
> 0, p
2
> 0, p
3
> 0 and p
1
p
2
p
3
> 0. (19)
3.3. Bifurcation analysis
3.3.1. Trival critical point
In general, violation of the stability conditions in (14) leads to the following possibilities:
1. Single zero singularity: when
1
(
2
a
1
) a
2
= 0, but
1
+
2
a
1
> 0. Here, a static bifurcation
occurs, i.e. the equilibrium solution (8) loses stability and jumps to the second equilibrium solution
(9).
2. Hopf singularity: when
1
+
2
a
1
= 0, but
1
(
2
a
1
) a
2
> 0. In this case, the system undergoes
a dynamics (Hopf) bifurcation. The equilibrium solution (8) loses stability and bifurcates to periodic
solutions.
3. Single zero plus Hopf singularity: When
1
+
2
a
1
= 0 and
1
(
2
a
1
) a
2
= 0. This is
a compound bifurcation point, and more complicated dynamical behaviour such as quasi-periodic
motions may occur.
It is worth investigating which of the aforementioned possibilities may occur from the rst equilibrium
(8). It can be shown that, because of the earlier assumption that all the model parameters are positive,
only the rst case is possible. In other words, the rst inequality in (14) cannot become 0 before the
second inequality does. This can be seen as follows: if the rst inequality becomes 0, which gives a
1
=

1
+
2
> 0, then
1
(
2
a
1
)a
2
=
2
1
a
2
< 0, showing that the second inequality has already been
violated.
Thus, the equilibrium (8) loses stability when
1
(
2
a
1
) a
2
= 0 and jumps to another equilibrium.
The question now is what the other equilibrium is. For the IVP system under consideration, there is only
one possibility, namely, the equilibrium solution (9). This can be veried as follows: Substituting the
condition
1
(
2
a
1
) a
2
= 0 into the stability conditions for the second equilibrium (9) yields
p
1
=
1
+
2
a
1
+, p
2
= (
1
+
2
a
1
), p
3
= 0, (20)
and thus, p
1
> 0, p
2
> 0 due to the assumptions
1
+
2
a
1
> 0 and > 0. This indicates that the two
equilibria (8) and (9) indeed share the same (static) stability boundary, dened by
1
(
2
a
1
) a
2
= 0,
and both lose stability on this boundary and jump from one to the other. Next, it is worth investigating
how the two equilibria change stability conditions. Suppose the model parameters are chosen such that
the expression
1
(
2
a
1
) a
2
becomes negative. The aim is to determine whether the equilibrium (9),
which bifurcates from the equilibrium (8), is stable. If it is stable, then the two equilibria indeed exchange
A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181 175
their stabilities on the (static) stability boundary. To illustrate this point, let
B =
1
+
2
a
1
, and b =
1
(
2
a
1
) a
2
+, (21)
then if
1
+
2
a
1
> 0 and
1
(
2
a
1
) a
2
= 0, it follows that B > 0 and > 0 (since b > 0).
Substituting (21) into (20) gives
W
2
p
1
= ( +
1
)[B
1

2
+(
1
+
2
) +(
1

2
+)] +(
1

2
+),
W
2
p
2
= ( +
1
){B
1

2
+[(
1
+
2
) +
1

2
]} +[(
1
+
2
) +
1

2
],
W
2
p
3
=
1

2
[( +
1
) +], W
2
= ( +
1
)(
1

2
+). (22)
Since all the parameters are positive, and, in addition, it is assumed that B > 0, > 0, so p
1
> 0,
p
2
> 0 and p
3
> 0. Furthermore, it is easy to prove that
p
1
p
2
p
3
=
( )( ) ( +
1
)(
1

2
+)
1

2
[( +
1
) +]
W
2
2
> 0, (23)
where ( )( ) represents the right-handsides of p
1
, andp
2
. It is worthmentioningthat the multiplication
of the terms in p
1
, and p
2
, marked by under-lines and over-lines (note that
1
+
2
= +
1
+ +
makes the second term in (23) vanish, which can be seen by splitting the second term of (23) into two
parts
( +
1
)(
1

2
+)
1

2
[( +
1
) +]
= ( +
1
)(
1

2
+)( +
1
)
1

2
+(
1

2
+)( +
1
). (24)
This proves that the rst equilibrium (8) loses stability on the static stability boundary

1
+
2
a
1
> 0,
1
(
2
a
1
) a
2
= 0, (25)
and bifurcates into a new stable equilibrium solution given by (9).
3.3.2. Non-trivial critical point
Returning, now, to the non-trivial solution (9), it should be noted rst of all that, owing to the use of
the notation B and , dened in (21), the stability conditions for the rst equilibrium (8) can be written as
B > 0 and < 0, (26)
where = 0, gives the static stability boundary. Then, as discussed earlier, it can be seen that the stable
equilibrium solution (9) loses stability when decreases to zero from a positive value, and bifurcates into
the rst equilibrium solution (8) which is stable for < 0. So, indeed, the two equilibrium solutions lose
stability on the static boundary (25) and exchange their stabilities.
It has been established that Hopf bifurcation fromthe rst equilibrium(8) is not possible, since Bcannot
cross zero before does. For the second equilibrium solution (9), however, Hopf bifurcation is possible
by keeping > 0 and changing, say, a
1
so that B < 0 (note that the rst equilibrium (8) is still unstable
under this condition). Additionally, it can be seen from (19) that, if any one of the rst three stability
conditions is allowed to be 0, the condition p
1
p
2
p
3
> 0 could be violated. Therefore, by choosing
the value of a
1
appropriately, it is possible to have Hopf bifurcation from the second equilibrium solution
(9).
176 A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181
Single zero and Hopf bifurcations are called co-dimension one singularities. An important question to
ask is whether or not this system has co-dimension two (or even three) singularities. For instance, is a
double zero or a zero plus Hopf singularities possible? These dynamical features are now investigated.
First, note that a zero singularity implies = 0, under which the stability conditions given in (22) reduce
to the conditions given by (20). To have a zero plus Hopf singularity, it is required that p
1
= 0 but p
2
> 0.
This is impossible since p
1
= 0 means a
1
=
1
+
2
+ and thus p
2
=
2
< 0. In other words,
p
2
becomes zero before p
1
does. However, this actually shows that a double zero singularity is possible
from the second equilibrium solution (9), because parameter values may be chosen such that p
2
= 0, but
p
1
= > 0. It is clear that the system cannot have p
1
= p
2
= p
3
= 0 simultaneously. Therefore, a
triple zero (co-dimension three) singularity is not possible.
In summary, the IVP system described by {(2),(3),(4)} or (7) has two equilibria given by (8) and
(9), respectively. Under the assumption that all the system parameters are positive, the rst equilibrium
solution (8) can only have static bifurcation, jumping to the second equilibrium solution (9) and they
exchange their stability properties on the static stability boundary. The second equilibrium solution (9),
on the other hand, can have single zero singularity (static bifurcation) and Hopf singularity (dynamic
bifurcation). Moreover, this equilibrium can have a double zero (co-dimension two) bifurcations. Further
bifurcations are possible, leading to quasi-periodic motions, which will be studied in a separate paper.
4. Numerical methods
4.1. Implicit method
4.1.1. Development
Starting with the initial value problem for S
u
given by (2), the development of numerical methods may
be based on approximating the time derivative by its rst-order, forward-difference approximation given
by
dS
u
(t )
dt
=
S
u
(t +) S
u
(t )

+O(
2
) as 0, (27)
where > 0 is an increment in t (the time step). Discretizing the interval t t
0
= 0 at the points t
n
= n
(n = 0, 1, 2,. . . ), the solutionof (2) at the gridpoint t
n
is S
u
(t
n
). The solutionof anapproximatingnumerical
method at the same grid point will be denoted by S
n
u
. A rst-order numerical method for solving S
u
in
(2) based on approximating the time derivative by (27) and making appropriate approximations for the
right-hand-side functions, is
S
n+1
u
S
n
u

= N
0
S
n+1
u

c
1
S
n+1
u
Y
n
u
S
n
u
+Y
n
u
+Y
n
v

c
2
S
n+1
u
Y
n
v
S
n
u
+Y
n
u
+Y
n
v
, (28)
Y
n+1
u
Y
n
u

=
c
1
S
n+1
u
Y
n+1
u
S
n+1
u
+Y
n
u
+Y
n
v
+
c
2
S
n+1
u
Y
n
v
S
n+1
u
+Y
n
u
+Y
n
v
( + +)Y
n+1
u
, (29)
Y
n+1
v
Y
n
v

= Y
n+1
u
(d +)Y
n+1
v
, (30)
A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181 177
which may be simplied to give, respectively,
S
n+1
u
=
S
n
u
+N
0
1 +[ +((c)/(S
n
u
+Y
n
u
+Y
n
v
))(
1
Y
n
u
+
2
Y
n
v
)]
, (31)
Y
n+1
u
=
Y
n
u
+[(c
2
S
n+1
u
Y
n
v
)/(S
n+1
u
+Y
n
u
+Y
n
v
)]
1 +( + +) [(c
1
S
n+1
u
)/(S
n+1
u
+Y
n
u
+Y
n
u
)]
, (32)
Y
n+1
v
=
Y
n
v
+Y
n+1
u
1 +(d +)
. (33)
This method, (31)(33), is denoted by GTY1.
4.1.2. Local truncation errors
The associated local truncation errors of the methods in GTY1 are, respectively,
L
S
u
= L
S
u
[S
u
, Y
u
, Y
v
, ] =
_
1 +
_
+
c
N
(
1
Y
u
(t ) +
2
Y
v
(t ))
__
S
u
(t +) S
u
(t ) N
0
,
L
Y
u
= L
Y
u
[S
u
, Y
u
, Y
v
, ] =
_
1 +( + +)
(c
1
S
u
(t +)
S
u
(t +) +Y
u
(t ) +Y
v
(t )
_
Y
u
(t +)
Y
u
(t )
_
c
2
S
u
(t +)Y
v
(t )
S
u
(t +) +Y
u
(t ) +Y
v
(t )
_
,
L
Y
v
= L
Y
v
[S
u
, Y
u
, Y
v
, ] = [1 +(d +)]Y
v
(t +) Y
v
(t ) Y
u
(t +), (34)
in which t = t
n
. It may be shown that the Taylor expansion of the functions in (34) about t lead to
L
S
u
=
_
1
2
S

u
+S

u
+
c
N
(
1
Y
u
+
2
Y
v
)S

u
_

2
+O(
3
) as 0,
L
Y
u
=
_
N
2
Y

u
+( + +)(NY

u
+S

u
Y
v
) +S

u
Y

u
c
1
(S
u
Y

u
+S

u
Y
u
) c
2
Y
u
S

u
_

2
+O(
3
) as 0,
L
Y
v
=
_
1
2
Y

v
+(d +)Y

v
Y

u
_

2
+O(
3
) as 0, (35)
indicating that the methods (31)(33) are rst-order accurate. It should be noted that, although the
rst-order method GTY1 (31)(33) is implicit by construction, it enables the three populations, S
u
,
Y
u
, Y
v
to be computed explicitly at every time step.
4.2. Euler method
The corresponding Euler method for solving the model IVP (see [11,13]), obtained by evaluating the
right-hand functions in (2)(4) at base time level t = t
n
, is given by, respectively,
S
n+1
u
= S
n
u
+
_
N
0
S
n
u
c
_

1
Y
n
u
S
n
u
+Y
n
u
+Y
n
v
+

2
Y
n
v
S
n
u
+Y
n
u
+Y
n
v
_
S
n
u
_
, (36)
Y
n+1
u
= Y
n
u
+
_
c
_

1
Y
n
u
S
n
u
+Y
n
u
+Y
n
v
+

2
Y
n
v
S
n
u
+Y
n
u
+Y
n
v
_
S
n
u
( + +)Y
n
u
_
, (37)
178 A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181
and
Y
n+1
v
= Y
n
v
+[Y
n
u
(d +)Y
n
v
]. (38)
This explicit method is rst-order accurate.
4.3. Second-order RungeKutta method
The second-order explicit RungeKutta method for solving (2)(4) denoted by RK2, is given by
S
n+1
u
= S
n
u
+
1
2
(k
1S
u
+k
2S
u
), Y
n+1
u
= Y
n
u
+
1
2
(k
1y
u
+k
2y
u
),
Y
n+1
v
= Y
n
v
+
1
2
(k
1y
v
+k
2y
v
), (39)
(see [11,13]), where
k
2s
u
= N
0
S
n
u
c
_

1
Y
n
u
S
n
u
+Y
n
u
+Y
n
v
+

2
Y
n
v
S
n
u
+Y
n
u
+Y
n
v
_
S
n
u
,
k
2y
u
= c
_

1
Y
n
u
S
n
u
+Y
n
u
+Y
n
v
_
S
n
u
( + +)Y
n
u
, k
2y
v
= Y
n
u
(d +)Y
n
v
, (40)
in which
S
n
u
= S
n
u
+k
1s
u
, Y
n
u
= Y
n
u
+k
1y
u
and Y
n
v
= Y
n
v
+k
1y
v
,
(41)
with
k
1s
u
= N
0
S
n
u
c
_

1
Y
n
u
S
n
u
+Y
n
u
+Y
n
v
+

2
Y
n
v
S
n
u
+Y
n
u
+Y
n
v
_
S
n
u
, (42)
k
1y
u
= c
_

1
Y
n
u
S
n
u
+Y
n
u
+Y
n
v
_
S
n
u
( + +)Y
n
u
, (43)
and
k
1y
v
= Y
n
u
(d +)Y
n
v
. (44)
5. Numerical experiments
To test the behaviour of the methods GTY1, Euler and RK2 for solving the non-linear IVP (2)(4) a
number of simulations were carried out with parameter values = 1/32,
1
= 0.6,
2
= 0.3, = 0.4,
d = 0.01, c = 4 with N
0
= 1000, S
0
u
= 500, Y
0
u
= 300 and Y
0
v
= 200 (see [13,6] and the references
therein).
The effect of the time-step on the three methods was monitored (by using the methods to solve the IVP
with various time steps) and the results are tabulated in Table 1. It is evident from Table 1 that method
GTY1 has a much better stability property than the Euler method (which failed when 1.4) and the
Runge-Kutta method (which failed when 1.6).
A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181 179
Table 1
Effect of time-step using c = 4
Euler RK2 GTY1
0.1 Convergence Convergence Convergence
1 Convergence Convergence Convergence
1.4 Divergence (method failed) Convergence Convergence
1.6 Divergence Divergence (method failed) Convergence
10 Divergence Divergence Convergence
100 Divergence Divergence Convergence
1000 Divergence Divergence Convergence
Table 2
Effect of number of sexual partners c
c Euler RK2 GTY1
4 Convergence Convergence Convergence
10 Oscillatory convergence Convergence Convergence
15 Divergence (method failed) Divergence (method failed) Convergence
30 Divergence Divergence Convergence
60 Divergence Divergence Convergence
To monitor the effect of the number of partners c, the three methods were used to solve the IVP using
the aforementioned parameter and initial values with various values of c. A time step = 0.5 was used.
The results, tabulated in Table 2, clearly conrm the superior stability property of method GTY1 in
comparison to the RK2 and Euler methods (both of which failed when c 15). Based on the numerous
simulations carried out with large parameter values, method GTY1 appears to be unconditionally stable.
The method GTY1 was used to study the effect of chemotherapy on the population by using various
values of . The steady-state values of the three populations (S
u
, Y
u
, Y
v
) in the cases when c = 5and c = 15
are tabulated in Tables 3 and 4, respectively. From these tables, it can be seen that as (the parameter
governing the administration of chemotherapy) increases, the populations of susceptible and treated
infecteds also increase, with a corresponding decrease in the untreated infected population. Furthermore, it
was clear that the ratio N

/N
A
of population sizes at equilibriumin the treated versus untreated populations
(obtained by setting = 0) was always greater than unity; thus, therapy is not detrimental to the society
[6]. If the ratio is less than unity, it could be assumed that therapy is detrimental to the society, despite
Table 3
Effect of chemotherapy: steady-state values for c = 5 using GTY1
S
u
Y
u
Y
v
N

N
A
N

/N
A
0.2 17 49 236 302 84 3.6
0.4 20 37 357 414 84 4.9
0.6 20 30 432 482 84 5.7
0.8 20 25 482 527 84 6.3
1.0 20 21 519 560 84 6.7
180 A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181
Table 4
Effect of chemotherapy: steady-state values c = 15 using GTY1
S
u
Y
u
Y
v
N

N
A
N

/N
A
0.2 6 49 239 294 76 3.9
0.4 6 37 362 405 76 5.3
0.6 7 30 438 475 76 6.3
0.8 7 25 489 521 76 6.9
1.0 7 22 526 555 76 7.3
being of benet to the individual. Other measures for determining the impact of treatment are death rate
from AIDS among the population, or the average life expectancy within the population [6].
Because of the superior stability and convergence properties of method GTY1, the steady-state values
given in Tables 3 and 4 can be relied upon when used to predict whether or not therapy was detrimental
to the society.
6. Conclusions
A competitive GaussSeidel-like, nite-difference method has been developed and used to solve
a non-linear mathematical model associated with the administration of anti-viral therapy to HIV in-
fected populations aimed at delaying progression to disease. This method proved to be much better,
in terms of numerical stability, than some well-known methods in the literature. The model, together
with its numerical solution, could be used to assess the impact of therapy to the HIV population in
particular, and to the society in general. Detailed stability and bifurcation analyses of the model were
reported.
Acknowledgements
Two of the authors (A.B.G. and P.Y.) acknowledge, with thanks, the support of the Natural Sciences
and Engineering Research Council of Canada (NSERC). The authors are grateful to M.L. Garba of the
Center for HIV/STDs and Infectious Diseases, University of North Carolina, for his helpful comments.
References
[1] R.M. Anderson, G.F. Medley, R.M. May, A.M. Johnson, A preliminary study of the transmission dynamics of the
Human Immunodeciency Virus (HIV) the causative agent of AIDS, IMA J. Math. Appl. Med. Biol. 3 (1986)
229263.
[2] R.M. Anderson, R.M. May, Understanding the AIDS pandemic, Sci. Am. 266 (1992) 5766.
[3] R.M. Anderson, S. Gupta, R.M. May, Potential of community-wide chemotherapy or immunotherapy to control the spread
of HIV-1, Nature 350 (1991) 356359.
[4] K. Dietz, K.P. Hadeler, Epidemiological models for sexually transmitted diseases, J. Math. Biol. 26 (1988) 125.
[5] D. Finzi, et al., Identication of reservoir for HIV-1 in patients on highly-active antiretroviral therapy. J. Am. Med. Assoc.
278 (5341) 12951300.
A.B. Gumel et al. / Mathematics and Computers in Simulation 54 (2000) 169181 181
[6] S. Gupta, R.M. Anderson, R.M. May, Mathematical models and the design of public health policy: HIV and anti-viral
therapy, SIAM Rev. 35 (1) (1993) 116.
[7] D.D. Ho, A.U. Neumann, A.S. Perelson, M. Markowitz, Rapid turnover of plasma virions and CD4 lymphocytes in HIV-1
infection, Nature 373 (1995) 123127.
[8] D. Kiessling, S. Stannat, I. Schedel, H. Deicher, Uberlegungen und Hochrechnungen zur Epidemiologie des Acquired
Immunodeciency Syndrome in der Bundesrepublik Deutschland, Infection 14 (1986) 217222.
[9] D.E. Kirschner, G.F. Webb, Understanding resistance for monotherapy treatment of HIV infection, Bull. Math. Biol. 59
(1997) 763785.
[10] E.G. Knox, A transmission model for AIDS, Eur. J. Epidemiol. 1 (1986) 165177.
[11] J.D. Lambert, Numerical Methods for Ordinary Differential Systems: The Initial Value Problem. Wiley, Chichester, UK,
1991.
[12] A.R. McLean, V.C. Emery, A. Webster, P.D. Grifths, Population dynamic of HIV within an individual after treatment with
zidovudine, AIDS 5 (1991) 485489.
[13] E.H. Twizell, Numerical Methods, With Applications in the Biomedical Sciences. Ellis Horwood, Chichester, UK, 1988.
[14] X. Wei, S.K. Ghosh, M.E. Taylor, et al., Viral dynamics in human immunodeciency virus type 1 infection, Nature 373
(1995) 117123.

Вам также может понравиться