S{I{R Epidemic Models with Directed Di usion

D. B. Meadey and F. A. Milneryz

Abstract. A generalization of the Gurtin{MacCamy model for an S{I{R epidemic is described. The
new model, which includes di usion away from overcrowded regions, retains many of the interesting qual-
itative features of previous models. In addition, the global{in{time existence of solutions is proved for a
special case. Qualitative properties of the solution are discussed and illustrated with a numerical example.

1. Introduction. Mathematical models for epidemics have been evolving in complex-

ity and realism during the past 60 years. From the simple, unstructured model of Kermack
and McKendrick of 1927 [7], many models incorporating, for example, age structure, time
delays corresponding to incubation periods, spatial di usion, and variable infectivity have
been proposed (see, e.g., [2], [3], [4], [6]). The e ects of dispersion on the spatial distribu-
tion of the subpopulations are of primary interest in this work.
In addition to allowing for more realistic descriptions of the observed phenomena, more
complex models have solutions which can be qualitatively very di erent from the solutions
of simpler ones. In this fashion the solutions can vary from exponentially increasing or
decreasing populations in the simplest setting, to periodic or oscillatory solutions when
delays are introduced, or to populations that \blow up" in a nite time for some nonlinear
models. In the case of spatially distributed populations, there is have a fairly complete
analysis for a model with \random" di usion [12]. Global existence and uniqueness re-
sults have been obtained for the case of equal di usivities in all classes, and all solutions
originating from non{negative (positive) initial data remain non{negative (positive). Fur-
thermore, the infection always dies out and the spatial distribution of susceptibles tends
to a constant function. For the case of unequal di usion coecients the questions of global
existence and asymptotic behavior are open. Just as is observed with the heat equation,
the speed of transmission of the infection is in nite, making this model rather unrealistic
for many applications.
From an epidemiological perspective, a more reasonable assumption than random mo-
tion of susceptibles is that they move away from infection. This is the basis for the
Gurtin{MacCamy model [8]. Solutions to this model may tend asymptotically to spa-
tially inhomogeneous distributions. The only analytic results for this model are positivity,
uniqueness and local existence of solutions [9]. Numerical evidence indicates that the so-
lution may actually \blow up" in nite time due to the accumulation of susceptibles at
the boundary of the support of the infected individuals (when the initial distribution of
y Department of Mathematics, Purdue University, West Lafayette, IN 47907, USA
z Dipartimento di Matematica, IIa Universita di Roma, Rome, Italy
the latter is compactly supported in the total domain) [10]. Although the disease propa-
gates with a nite speed, the particular form of the di usion is too restrictive for real-life
applications: the infection cannot spread to regions in which the disease was not present
initially.
In order to overcome this unwanted feature we propose a new model which incorporates
a \directed" di usion away from overcrowding. The new model also has more physical
propagation characteristics. Whereas the Gurtin{MacCamy model has nite propagation
speed it does not allow for the spread of an infection beyond its initial support. In contrast,
the new model retains the nite speed of transmission and does permit the spread of the
infection beyond its initial support.
2. Models with Directed Di usion. Although several types of di usion terms have
been proposed for various demographic models, for human epidemics it is fairly reasonable
to assume that uninfected individuals will move away from the infected, as considered by
MacCamy [8]. Models based on just this kind of motion (\directed away from infection")
may result in an unwanted accumulation of susceptibles on the boundary of the infected
region [10]. To overcome this limitation we propose a model which contains additional
di usion away from large concentrations of individuals.
Let N = N (x; t) be the spatial distribution of a population at time t  0 in the
one{dimensional domain = [0; L]. The population is subdivided into classes of suscep-
tible, infected (and infective, as we shall assume no incubation period), and removed or
\recovered" (usually by immunity) individuals with spatial distributions S (x; t), I (x; t),
and R(x; t), respectively. In this model we shall not consider the demography (that is,
there are no births or deaths) but just the transition from one epidemic class to another.
Consequently, the dynamics are given by the following system of PDE's:
8 S = k (SN ) + k (SI ) ? SI;
>
< t 1 xx 2 xx
(2.1) It = k1 (INx )x + SI ? I;
>
: R = k (RN ) + I;
t 1 x x

where the x and t subscripts indicate partial di erentiation in those variables, > 0 is a
mean infectivity per individual so that I is the \force of infection", > 0 is the recovery
rate, k1  0 is the coecient of di usion away from overcrowding and k2  0 the one away
from infection. The case k1 = 0, k2 > 0 corresponds to the Gurtin{MacCamy model [8].
We shall use \no ux" (homogeneous Neumann) boundary conditions to maintain a closed
population (no immigration or emigration):
 (k1 Nx + k2 Ix )(0; t) = (k1 Nx + k2 Ix )(L; t) = 0;
(2.2)
k1 Nx (0; t) = k1 Nx (L; t) = 0:
Note that, for k1 ; k2 > 0, these yield Nx (0; t) = Nx (L; t) = Ix (0; t) = Ix (L; t) = 0. We
shall assume that the initial distributions for the three classes are non{negative (for obvious
2
reasons) and C 2 on [0; L],
(2.3) S (x; 0) = S0 (x); I (x; 0) = I0 (x); R(x; 0) = R0 (x)  0:

A solution of (2.1){(2.3) on [0; T ), 0 < T  1, is a triple (S; I; R) of functions

C 1 ([0; T ); C 2 ( )) which are non{negative on  [0; T ) and satisfy (2.1){(2.3). Note
that this imposes some compatibility conditions on the initial and boundary data, e.g.
I0;x (0) = I0;x (L) = 0. The following lemma shows that any solution of the system (2.1){
(2.3) with non{negative initial data remains non{negative.
Lemma 2.1. Let (S; I; R) be a solution of (2.1){(2.3) on  [0; T ). Then, S; I , and R
are non{negative and, if S0 > 0, I0 > 0 everywhere, then S and I are positive on  [0; T ).
Proof. Consider the characteristic curve X (s; t), with parameter s 2 , which is the
solution of the initial value problem
 X (s; t) = ?k N (X (s; t); t) ? k I (X (s; t); t);
t 1 x 2 x
(2.4)
X (s; 0) = s:
Existence and uniqueness are guaranteed, for instance, if Nx and Ix are continuous in
 [0; T ), and uniformly Lipschitz continuous in x, which we have assumed. It follows
that
Xst (x; t) = ? (k1 Nxx (X (s; t); t) ? k2 Ixx (X (s; t); t)) Xs (s; t):
As in [9], let S(s; t) = S (X (s; t); t). Then
 X (s; t) 
St (s; t) = ? + I (X (s; t); t) S(s; t);
st
Xs (s; t)
with S(s; 0) = S (X (s; 0); 0) = S (s; 0) = S0 (s)  0. Also, since Xs (s; 0)  1 > 0, it follows
that
Xs (s; t) > 0; s2 ; 0  t < T:
Thus,  Zt 
S0 (s)
S(s; t) =
X (s; t)
exp ? I (X (s;  );  )d
s 0
is non{negative, and positive whenever S0 is.
Similarly, let Y (s; t), s 2 , be the solution of the initial value problem
 Yt (s; t) = ?k1 Nx (Y (s; t); t)
(2.5)
Y (s; 0) = s:
Then
Yst (s; t) = ?k1 Nxx Ys (s; t);
3
and I(s; t) = I (Y (s; t); t) is seen to satisfy
 Y (s; t) 
It (s; t) = ? ? S (X (s; t); t) + I(s; t);
st
Ys (s; t)
with I(s; 0) = I (Y (s; 0); 0) = I (s; 0) = I0 (s)  0. As above, we have Ys (s; t) > 0, s 2 ,
t 2 [0; T ), and thus
 Z 
I0 (s)
I(s; t) =
Ys (s; t)
exp t ? S (X (s;  );  )d
0
is always non-negative, and positive whenever I0 is.
Finally, with R(s; t) = R(Y (s; t); t) we obtain, as in the previous two cases,

Rt (s; t) = ? YYst((ss;;tt)) R(s; t) + I(s; t);

s

which gives
Z t
R (s; t) = Y (s; t) Ys (s;  )I(s;  )d  0;
s 0
since R0 = 0, Ys (s;  ) > 0, s 2 [0; L], t  0, and I  0 (as just shown). This completes the
proof.
Remark 2.1. For any xed s 2 , the characteristics X (s;
) and Y (s;
) remain in so
long as they exist. More speci cally, the no- ux boundary conditions (2.2) imply that any
characteristic originating at an endpoint must be vertical (i.e. X (0; t) = Y (0; t) = 0 and
X (L; t) = Y (L; t) = L for all t  0). Then, using the continuity with respect to s, it is
seen that each family of characteristic curves must cover all of for each t  0 for which
they exist. Note also that this con rms the claim that transmission of the disease occurs
with nite speed.
Next we observe that the boundary conditions, (2.2), imply that the total population,
Z
N (t) = N (x; t)dx;

is, as expected due to the absence of births, deaths, immigration, and emigration, constant.
Consequently,
N (t)  N 0 = kS0 kL1 ( ) + kI0 kL1 ( ) :
We now let
Z Z Z
(2.6) S (t) = S (x; t)dx; I (t) = I (x; t)dx; R(t) = R(x; t)dx;

4
denote, respectively, the total number of susceptible, infected, and removed individuals at
time t. It follows from (2.1) and (2.2) that
Z
S (t) = ? I (x; t)S (x; t)dx  0
0
(2.7)

since, by Lemma 2.1, I and S are non{negative. Under the assumption that the solution
to (2.1){(2.3) exists for all time, we see that the total number of susceptibles has a limit,
S 1 . Since S (0) = kS0 kL1 ( ) and S is non{negative and non{increasing, this limit satis es
Z
(2.8) 0  S 1 = tlim
!1
S (t) = tlim
!1
S (x; t)dx  kS0 kL1 ( ) :

We also obtain, from (2.1) and (2.2),

8 0 Z
< I (t) = I (x; t)S (x; t)dx ? I (t);
(2.9) : R0 (t) = I  0:

In particular, R is non{negative, non{decreasing, and bounded by N 0 , therefore it has a

limit R1 ,
(2.10) 0  R1 = tlim
!1
R(t)  N 0 :

Finally, I (t) = N 0 ? S (t) ? R(t) must also have a limit I 1 ,

(2.11) 0  I 1 = tlim
!1
I (t)  N 0 :

We can show now that, just as in other S ? I ? R model without demography, the disease
dies out everywhere.
Lemma 2.2. Assume the solution of (2.1){(2.3) exists for all time. Let S , I , R be the
functions de ned by (2.6) and let I 1 be given by (2.11). Then I 1 = 0, and tlim
!1
I (x; t) = 0
for all x 2 .
Proof. Integration of (2.9)2 yields
Z t
(2.12) R(t) = I ( )d:
0
Thus, from the existence of R1 , it is seen that I (t) ! 0 as t ! 1. The non-negativity of
I implies I (x; t) ! 0 almost everywhere and hence, by continuity, everywhere.
5
 Remark 2.2. Note that (2.12) says that the cumulative number of \recovered" individ-
uals is equal to the recovery rate times the cumulative number of infected individuals, as
one would intuitively expect.
Remark 2.3. A generalization of the well-known threshold phenomenon for the Kermack{
McKendrick
Z model is obtained from (2.7) and (2.9)1 : if there exists t  0 for which
0
S (x; t )I (x; t )dx > I (t ), then I (t ) > 0 and the total number of infectives in-
creases. This gives a sucient condition for a \local{in{time" epidemic. Note that this
does not say that the infectives increase pointwise Z on all of . Moreover, it is not clear
that this condition is also necessary, since, for S (x; t)I (x; t)dx  I (t) there may still
be regions of in which the infective population increases.
Further analytical information about the asymptotic behavior of solutions is more
dicult to obtain. It is interesting to note that once it is known that I ! 0 all of the
time-independent equations are formally satis ed. As a result there is no reason to believe
that the steady-state distribution of S or R will be homogeneous. This is illustrated in the
numerical example in the next section.
Next we concentrate on the questions of existence and uniqueness of solutions to (2.1){
(2.3). Numerical evidence [10] suggests that one should not expect global{in{time existence
of solutions without imposing some additional constraints on the data. We shall show that,
when k2 = 0 (that is, there is di usion directed away from overcrowding but not away from
infection) we do have global existence and uniqueness.
Theorem 2.1. Let k1 > 0 and k2 = 0 in (2.1) and (2.2). Then (2.1){(2.3) admits a
unique solution on  [0; 1) for any initial data such that S0 + I0 > 0.
Proof. In this case the total population, N , is a solution of the following initial{
boundary value problem for the porous medium equation:
8 k1 2
>
< Nt = 2 (N )xx ;
> x 2 ; t > 0;
(2.13) > Nx (0; t) = Nx (L; t) = 0; t  0;
>
: N (x; 0) = N (x) = S (x) + I (x) > 0; x 2 :
0 0 0
From the positivity of N0 it follows that there is a unique classical solution of (2.13)
which is positive (see, e.g., [1], [11]). Furthermore, both Nx and Nxx are continuous and
bounded. Thus S and I are solutions of the following almost linear rst{order system
(linear with quadratic perturbations in the zero{order terms):
 St = (k1 Nx )Sx + (k1 Nxx ? I )S;
(2.14)
It = (k1 Nx )Ix + (k1 Nxx + S ? )I:
Note that, in general, not even the single linear equation St = a(t; x)Sx admits a global
solution (consider, for example, a(x; t) = x2 ; then, no point (x1 ; t1 ) with x1 t1 > 1 lies on
6
a characteristic originating at (x0 ; 0), x0 > 0). Consequently, (2.1) may not be globally
solvable in general.
It is possible, however, to derive an a priori L1 {bound for W = S + I . In fact, (2.14)
implies that
(2.15) Wt = (k1 Nx )Wx + (k1 Nxx )W ? I:

From S; I  0 it follows that so long as W exists, it satis es the di erential inequality

Wt ? (k1 Nx )Wx  KW;

with K = supfk1 jNxx j; jk1 Nxx ? jg < +1.

Let x0 2 be given and consider the characteristic curve x = Y (x0 ; t). The function
W(t) = W (Y (x0 ; t); t) is non-negative and satis es the di erential inequality

d
W  K W:
dt
Therefore,
0  W(t)  W(0)eKt = (S0 (x0 ) + I0 (x0 ))eKt
 (kS0 kL1 + kI0 kL1 ( )) expf(k1kNxx kL1 + )tg:
Hence W = S + I will exist for all t  0, x 2 which can be reached via a Y -characteristic.
The fact that Nx is uniformly Lipshitz continuous in x and uniformly bounded implies the
global solvability of (2.5). Combined with the observations in Remark 2.1, W is seen to
exist globally in space and time.
To complete the proof, note that explicit formulae for S and I can be obtained from
(2.15) and the de nition of W :
( 1
I = (k (N W ) ? W );
1 x x t

S = W ? I:

Remark 2.4. The situation in which the initial data is not strictly positive is also of
interest, e.g. a compactly supported population. In this case solutions to (2.13) exist only
in a \weak" sense; discontinuities in Nx can occur at points where N = 0. The analysis of
the resulting almost linear system for S and I must also be modi ed.
3. Numerical Example. The families of characteristic curves are also well-suited
to numerical implementation [5]. In this section we describe a numerical method for the
solution of (2.1){(2.3) with k1 > 0 and k2 = 0, i.e. di usion to avoid overcrowding only.
7
An example which illustrate some of the properties discussed in the previous section is also
provided.
Let Y satisfy (2.5) and rede ne the auxiliary functions
S(s; t) = S (Y (s; t); t)Ys (s; t); I(s; t) = I (Y (s; t); t)Ys (s; t); N(s; t) = N (Y (s; t); t)Ys (s; t):
It is a simple exercise to verify that these new functions satisfy the equations
8 I(s; t)
>
>
> St (s; t) = ? S(s; t);
>
< Ys (s; t)
 
(3.1) S(s; t)
>
>
> It (s; t) = ? I(s; t);
> Y s (s; t)
:
Nt (s; t) = 0;
with initial conditions
(3.2) S(s; 0) = S0 (s); I(s; 0) = I0 (s); N(s; 0) = S0 (s) + I0 (s):

Moreover, the governing equation for the characteristic curves takes the form
 N (s; t) N(s; t)Y (s; t) 
(3.3) Yt (s; t) = ?k s
2 ? ss
:
Ys (s; t) Ys (s; t)3
The algorithm for the numerical solution of (3.1){(3.3) is a standard (and simple)
nite di erence method. Discretize both the time, t, and spatial parameter, s. The
explicit Euler method is used to advance from one timestep to the next. This proceeds
in two steps. Second-order centered di erence approximations of the spatial derivatives in
(3.3) are used to compute the position of each characteristic at the next timestep. Then
equations (3.1) can be \integrated" to obtain approximate values of S, I, and N along each
characteristic.
An example on = [0; 1] is used to illustrate the properties discussed in the preceding
section. Suppose the initial distributions of susceptibles and recovereds are constant (S0 =
6, R0 = 0) while the infectives are given by
8  2 
< 3 exp x
x2 ? 0:82
; 0:0 <x < 0:8;
I0 (x) =
:
0; 0:8 x  1:0:
Note that I0 2 C ( ). The values of the di usion coecients and infectivity and recovery
1

rates are selected to be k1 = 0:1, k2 = 0, = 1:0, and = 5:0. Snapshots of the solution
at several representative times are given in Figure 1.
It is interesting to note that the total population approaches the anticipated uniform
distribution very rapidly (see Figure 1.b). Although the threshold condition is satis ed
only in a very small neighborhood of t = 0, each of the graphs indicates some local growth
in the infection. Of course, all of the infectives eventually disappear.
The steady-state solution is essentially that shown in Figure 1.f. This gure is consis-
tent with the previous observation that the steady-state susceptible and recovered popu-
lations are not required to be constants.
8
 Conclusion. The Gurtin{MacCamy model has been generalized with the inclusion of
di usion to avoid overcrowding. While the analysis of the full model is far from complete,
the results for the special case without di usion to avoid concentrations of the infection
are promising.
It remains to analyze both the case with compactly supported initial data and the full
model with both di usions. Preliminary results on these problems indicate that most of
the features of the simple model are present in the more complicated models. The results
of these investigations are forthcoming.
References
[1] D. G. ARONSON, The Porous Medium Equation, in Nonlinear Di usion Problems, A.
Dold and B. Eckmann, eds., Lect. Notes in Math., 1224, Springer-Verlag, New
York{Heidelberg{Berlin, 1986, pp. 1{46.
[2] S. BUSENBERG AND C. CASTILLO-CHAVEZ, Interaction, Pair Formation and Force of
Infection Terms in Sexually Transmitted Diseases, 1989, (preprint).
[3] C. CASTILLO-CHAVEZ AND H. R. THIEME, On the Role of Variable Infectivity in the
Dynamics of the Human Immunode ciency Virus Epidemic, (preprint).
[4] M. E. GURTIN AND R. C. MACCAMY, On the Di usion of Biological Populations,
Math. Biosci., 33 (1977), pp. 35{49.
[5] M. E. GURTIN, R. C. MACCAMY, and E. SOCOLOVSKY, A Coordinate Transforma-
tion for the Porous Media Equation that Renders the Free-Boundary Stationary,
Quart. Appl. Math., 42 (1984), pp. 345{357.
[6] M. IANNELLI, R. LORO, F. A. MILNER, A. PUGLIESE, and G. RABBIOLO, An AIDS
Model with Distributed Incubation and Variable Infectivity: Applications to IV-
Drug Users in Latium, (preprint).
[7] W. O. KERMACK AND A. G. MCKENDRICK, A Contribution to the Mathematical
Theory of Epidemics, Proc. Roy. Soc. London Ser. A, A 115 (1927), pp. 700{721.
[8] R. C. MACCAMY, Simple Population Models with Di usion, Comput. Math. Appl.,
9 (1983), pp. 341{344.
[9] R. C. MACCAMY AND D. B. MEADE, An Epidemic Model with Directed Di usion,
in Biomedical Modelling and Simulation, J. Eisenfeld and D. S. Levine, eds.,
IMACS, Paris, 1989, pp. 197{199.
[10] D. B. MEADE, Analysis for an Epidemic Model with Di usion to Avoid Infection,
(submitted to J. Math. Biol.).
[11] O. OLEINIK, On some Degenerate Quasilinear Equations, in Instituto Nazionale di
Alta Matematica Seminari, 1965, pp. 355{371.
[12] G. WEBB, A Reaction-Di usion Model for a Deterministic Di usive Epidemic, J.
Math. Anal. Appl., 84 (1981), pp. 150{161.

9
 a. (t = 0:0) b. (t = 0:25)

c. (t = 0:5) d. (t = 1:0)

e. (t = 2:0) f. (t = 4:0)
Population distributions at selected times ranging from the initial distribution
(a.) to near steady-state (f.). The solid line indicates the total population. The infectives
occupy the region below the dashed line, the susceptibles the region between the dashed
and dotted lines, and the recovered the region between the dotted and solid lines.

10