MATHEMATICAL

COMPUTER
MODELLING
PERGAMON Mathematical and Computer Modelling 31 (2000) 75-81
www.elsevier.nl/locate/mcm

Reaction Transport Equations

in Biological Modeling
K. P. HADELER
Biomathematik, UniversitGt Tiibingen
Auf der Morgenstelle 10, D-72076 Tiibingen, Germany

Abstract-&action transport equations are models for interactions within or between species and
for spatial spread such that individuals assume positions in space and also their individual velocities.
These equations have much in common with reaction diffusion equations but, being hyperbolic rather
then parabolic, do not show the undesired feature of infinitely fast propagation. They are particularly
useful for biological situations where detailed description of moving individuals is important, e.g., in
the reaction of microorganisms to gradients of substances. Several modeling approaches by reaction
transport equations are outlined and insights into the qualitative behavior of such systems are given.
@ 2000 Elsevier Science Ltd. All rights reserved.

Keywords-Moving particles, Evolution of density, Reaction law, Reaction transport equation,

Boltzmann type equations.

We consider particles moving in n-dimensional space. Assume that particles are characterized
by their position x and their velocity vector s. Let ~(t,x,s) be the density of these particles
at time t. If the particles just move with their own velocities, i.e., on straight lines, then the
evolution of the density u is described by the “free stream” operator, i.e., the function u satisfies
the equation

$+&&0, (1)
z
i=l

which in the usual condensed notation assumes the form

ut + s. vu = 0. (2)
Now assume that particles can turn. One can imagine that a particle stops moving with a rate
p > 0 (thus, stopping is governed by a Poisson process) and then chooses a new velocity according
to a probability distribution which may or may not depend on the previous velocity. Such “run
and tumble” motion can be observed in many species of bacteria. The probability distribution
will be given by a kernel k(s, S) which satisfies Ic > 0 and s, Ic(s, S) ds = p where V is the set of
possible velocities, typically a spherical shell V = {s : y1 I IsI 5 yz} with 0 < y1 < ‘yz < 00 or a
sphere ,S”-l where 9-l is the unit sphere in Wn and y > 0. In the latter case, particles have
constant speed. Then the evolution equation for the density u assumes the form

ut+s~vu+pu-
s V
k(s, S)u(t, x, 5) dg = 0. (3)

0895-7177/00/L - see front matter @ 2000 Elsevier Science Ltd. All rights reserved. Typeset by dM-Tj$
PII: SO895-7177(00)00024-8
 76 K. P. HADELER

So far we have modeled spread of particles by a transport equation. Now introduce reactions.
Assume that, in the absence of motion, in a well-stirred situation, the reaction law is described
by an ordinary differential equation in = f(u). What is the appropriate system for these particles
if they move and react at the same time ? An immediate idea is the following: at any space
position 2, all particles react according to the reaction law independent of their velocities. This
idea leads to the equation

u~+s~vu+pL-
J k(s, a)u(t,x, 3;)dS = +p),
V

Ju(t,z,s)ds
where
ti(t,z) = (5)
V

is the total number of particles at position x and (VI = sv ds. We call an equation of this
form a reaction transport equation or an equation of Boltzmann type. The classical Boltzmann
equation [l] has a similar form, but instead of the terms modeling turning and reaction, there is a
nonlinearity describing collisions of particles. In connection with biological modeling, equations
of the form (4) are more appropriate. They have been called velocity jump processes [2,3].
In writing equation (4), we have tacitly made another assumption. We have assumed that
the change in the number of particles at location z depends on the total number zi of particles
present, but also that the produced particles are uniformly distributed to all velocities in V. This
choice is somewhat special. More generally, one can assume that the produced particles choose
their velocities according to some distribution IEi(s) where ki >_ 0 and Jv /cl(s) ds = 1. Then we

Jk(s,
+(t,
z,
a)
get an equation
zJt+s*vu++- dS = kl(s)f(ii). (6)
V
Of course, if ki is totally arbitrary, then there will be preferred directions, but a kernel of the
form ki(]s]) will lead to a perfectly symmetric model.
Still there is something “wrong” with equation (6). To make this clear, consider a reaction
equation in the form 6 = f(u) E mu - g(u)u, where m(u) > 0 is the density-dependent
fertility and g(zL) > 0 is the density-dependent mortality. The simplest example is the Verhulst
equation ti = a~(1 - U/K), w h ere we can split the nonlinearity, e.g., m(u) = a, g(u) = (a/j+
The point is that a newborn particle can choose its velocity but a dying particle cannot; at the
moment of extinction it is removed from its velocity class. Therefore, it is natural to replace the
nonlinearity in (6) by a law which distinguishes between production and extinction [4-71. Then,
instead of (6), the equation becomes

u~+s*vu+pu-
Jlc(s,
B)u(t,z,
V
S) di = k~(s)m(ti)ii - g(ii)u. (7)

In the simplest situation, the particles live on the whole space R”. In more realistic (and more
interesting) situations, they are confined to some bounded domain Sz in space. Then boundary
conditions have to be imposed. Since the partial differential equations are hyperbolic, boundary
conditions can only be prescribed at points of the boundary 80 where characteristic directions
are pointing inward. Let v be the outer normal vector on 80. A boundary condition of Dirichlet
type describes a situation where particles are absorbed at the boundary,

u(t,2, s) = 0, for vTs < 0. (3)

The Neumann boundary condition describes a situation where particles do not leave the domain.
They are “reflected” at the boundary. Of course, the law of reflection must be specified. Then
the boundary condition reads

u(t, z, s) = u(t, 2, s’), if VTs = --VTs’ , uTs < 0. (9)

 Reaction Transport Equations 77

Now consider the special case of space dimension n = 1 with V = {fy} consisting of two points
only. Then there are only two types of particles u+(t,~) = u(t,z, y) (moving to the right) and
u- (t, x) = uL(t,z, -7) (moving to the left), and the systems (6) and (7) read, with u = u+ + u-,

?A:+yu,+
= fP@- u+) + ff(47
- (10)
2% Q(uf - u-) + $(U)?
-y”,=2

uI’+yuf = $(u- - u+) + $r(u)u- g(u)u+,

- (11)
ut - yu, = &A+
2 - u-) + +z(u)u- g(u)u-.
In the case of pure motion (f = 0), the system (10) describes a correlated random walk with
speed y and turning rate /.L (In most papers on correlated random walks, the present p appears
as 2~. The present notation makes comparison to equation (3) easier. The notational difference
has a physical meaning: it makes a difference whether a particle at a stop must turn to the
opposite direction or if it can do so with probability l/2.) Correlated random walks have been
studied in detail in [8,9] and in many subsequent papers, see [6].
If fl = [0,1] is a bounded interval, then the zero Dirichlet boundary condition (8) reads

u+(t,O) = 0, u-(t, 1) = 0, (12)

and the homogeneous Neumann boundary condition (9) becomes

u+(t,x) = u-(t,x), x = 0,l. (13)

Using the flow v = y(u+ - u-), the system (10) can be written

Ut + 21, = f(u),
(14)
Avt + Du, + v = 0,
P

with D = 7’/~. The first equation is a balance law (a conservation law in the case f G 0), the
second equation is a feedback loop connecting the flux v to the gradient of u. A straightforward
generalization of this system to several space dimensions is the semilinear Cattaneo system

ut + div v = f(u),
(15)
Tut+Dgradu+v=O.

Here, u is a scalar density, and v is an n-dimensional vector field, and r > 0 is a time constant.
The linear Cattaneo system (f = 0) has been introduced in [lo] as a model for heat transfer; see
also [ll] for the general problem of appropriate modeling of heat transfer by hyperbolic equations.
For solutions (u, w) of the system (15), the function u is a solution to the semilinear telegraph
equation or damped wave equation

Tutt + (1 - Tf’(u))ut = DAu + f(u). (16)

Of course, this transition is valid also for the one-dimensional case (10). In particular, the
probability density for the position of a particle performing a correlated random walk satisfies
a linear one-dimensional telegraph equation [9]. However, for the more general equation (1 l), a
transition to a telegraph equation is not possible, except in the case where g is a constant.
 78 K. P. HADELER

If in (16) we take the formal limit for /.J --t co, D = con&, then we obtain a reaction diffusion
equation
ut = DAu + f(u). (17)

Thus, the reaction diffusion equation appears as the limiting case of the semilinear Cattaneo
system (15) or the random walk system (10) in the case n = 1 if particles move with high speed,
turn very frequently, and if the speed and the turning rate are appropriately scaled, or if the time
constant is small.
Very much is known about the qualitative behavior of biological models based on reaction
diffusion equations. We review some typical cases and compare them to the hyperbolic situation.
First look at the original interpretation of Fisher’s equation ut = u,, + f(u) with f being a
hump function like f(u) = u(l - u). A travelling front solution is a solution u(t,z) = 4(z - d)
where 4 is the shape of the front and c > 0 is its speed. Recall that in the parabolic case, particles
have no individual velocities. Here, 0 5 4(z) < 1 and 4(-co) = 1, 4(foo) = 0. It is well known
that there is minimal speed cp > 0 such that for every speed c 2 cp, there is a travelling front
solution, unique up to translation. The front with minimal speed cp is the most important since
it describes the advance of a front of particles into unoccupied territory. For concave functions f,
the minimal speed is given by cp = 2m.
By what has been said before, the hyperbolic analogue of this equation is the random walk
system (10). For this system and also for (ll), results on travelling fronts can be shown [4,12]
which parallel those for the parabolic case. Again, there is a minimal speed cH of travelling
fronts, and for every c E [cH, 7) there is a travelling front solution, unique up to translation.
No front travels faster than the individual particles, hence, c < y. The minimal speed depends
continuously on the parameters in the sense that cH -+ cp for p + co, y + oo, y2/p = D. The
results on travelling fronts in the hyperbolic case are not as complete as in the parabolic case
[12]. The stability problem for front solutions of a hyperbolic equation has been studied in [13].
It is easy to see that a travelling front of the one-dimensional problem (10) yields also a plane
wave front of the Cattaneo system (15), just by adding dummy variables perpendicular to the
direction of propagation. This behavior is in sharp contrast to the behavior of the more realistic
reaction transport equations or equations of Boltzmann type (4)-(7). In this case, the speed of
travelling fronts depends on the space dimension in an essential way. Consider again equations (4)
and (5) with constant speed of particles, i.e., V = ySnW1 is a sphere of radius y,

Ut + s * vu + pu - ha = &(a).

Let c, be the minimal speed of (any) fronts in dimension n, obtained from linearization at u = 0,
i.e., from f’(0). In [14] it has been shown that, for f’(O), y, and p fixed, c, is strictly decreasing
with dimension n. The numbers cl, c2 can be explicitly computed.
Another application of reaction diffusion equations is epidemic spread, the standard model
being a partial differential equation version of the Kermack-McKendrick SIR model

ut = -Puw + Duu,,,
(1%
wt=/?uw-aw+D,w,,.

Here, u, w are the susceptibles and infecteds, respectively, p is the transmission rate, cr the
recovery rate, and D,, D, are the diffusion rates. In order to understand what kind of travelling
fronts can be expected, consider for a moment the underlying ordinary differential equations
system (i.e., (19) without the terms containing spatial derivatives). There is a one-parameter
family of stationary points (u, 0), u > 0, and a family of trajectories connecting stationary points
(6, 0) to (2, 0) where a- (a//3) log% = a - (o/P) log a, 14 < (Y/P < ti. Thus, the epidemic starting
with ?i susceptibles leaves 21 susceptibles after it has passed.
 Reaction Transport Equations 79

For the parabolic system (19), only the “rabies case” D, = 0 is relatively easy [15,16]. Then
only infecteds migrate. For each choice of fi > o/P, there is a minimal speed cp, and for each
c 2 cp, there is a travelling front with limits (~1,w)(-co) = (g,O), (u,w)(+o;)) = (G,O). Thus,
the epidemic is moving into a territory with a level of susceptibles B, and it leaves susceptibles u
behind.
The natural hyperbolic analogue of (19) is the following system:

u; +ytLu; = fp&- -I&+) - pu+w,

Ut - -/l&u; = 1p (uf
-
2 u - u-) - pu-w,
(20)
wt’ + yww;= fpwcw-
- P(ouf + w+) + (1 - g)zL-)w - QW--,

Wt -Yww,=2 1, w @J+ - w-) + P((l - a)uf + g.u-)w - crw-,

where w = w+ + w-. Susceptibles and infecteds have speeds yU, yW and turning rates pL,, p,,,,
respectively. Susceptibles become infected by infecteds and choose a new direction with proba-
bility 1 - (T. Again, the case where only infecteds migrate is much easier than the general case.
In that case, there is only one type of susceptibles u. The proportion u in (20) has no meaning.
But now newly infecteds can, with a certain probability 7, choose the direction of the infector,

Ut = -puw,

(21)

One can show, that for each choice of ii > o//3, and p sufficiently large, there is an interval [CH, y)
of speeds of travelling fronts, whereby CH --+ cp for I_L+ co, p/y2 = D, (161. The limiting case,
where only susceptibles move, has been studied in [17].
Consider a system of the form (10) in the case of several species ui, i = 1, . . . , m. We introduce
vector notation. The densities for particles moving to the right or left are U+ = (ut, . . . , U&),
U- = (u;,.. . ,I&), respectively, and u = (~1,. . . , IL,) are the total particle densities. The
diagonal matrices r = (~i&k) and M = (pi&k) contain the particle speeds and turning rates.
The function f : R” + Rm describes the interactions between the m species in a “well-stirred”
situation, ti = f(u). Then we have the system of 2m hyperbolic equations

UI’ + ru,f= +qL- - u’) + +,.

(22)
u; - ru, = $qu+-u_) + ff(u),
which should be compared to a system of reaction diffusion equations of the form (17), with
D = (Di&,) being the diagonal matrix of diffusion coefficients. If one takes the diffusion limit
in (22), with p( + c=, “li + oo, $I/& + Di, then one arrives at (17). Also, a transition to a
system of semilinear telegraph equations of the form (16) is possible [5].
For the vector-valued reaction diffusion equation (17), the Turing phenomenon has been fre-
quently studied. Consider a system (17) where u is a vector of m species, on an interval (0,1] with
no-flux (zero Neumann) boundary conditions. Assume the underlying reaction equation it = f(u)
has an exponentially stable stationary point G, i.e., f(G) = 0, and the Jacobian matrix A = f’(G)
has only eigenvalues with negative real parts. Then u(t, z) = ii is a spatially constant stationary
solution of the partial differential system. Although this stationary solution is stable with respect
to the reaction system and diffusion is a dissipative process, the spatially constant solution of
80 K. P. HADELER

the reaction diffusion system may nevertheless become unstable. This is an entirely linear phe-
nomenon: if the elements of the matrix A satisfy certain inequalities (which are explicitly known
for m = 2,3) and if the diffusion coefficients are chosen appropriately, then for certain domain
lengths 1 there are unstable modes. For m = 2, the typical concept is that of an activator ~1 and
an inhibitor ~2 and D1 < 02 (“short range activator and long range inhibitor”). These concepts
can be carried over to hyperbolic systems of the form (22). In [18,19], the exact conditions for
Turing instability in (22), for m = 2, have been determined. It is somewhat astonishing that the
results can be formulated in terms of the quantities $/pi, i.e., those quantities which appear in
the diffusion limit.
An important problem is the investigation of the global behavior of systems of the form (22).
Consider the system (22) on a bounded interval with boundary conditions (8) or (9). For the
reaction diffusion equation (17) with the appropriate boundary conditions, it is known that in the
scalar case, m = 1, for space dimension n = 1, the limit sets have a rather simple structure, there
holds an analogue of the Poincar&Bendixson theorem (which describes limit sets of systems
of ordinary differential equations in the two-dimensional plane). In the case of several space
dimensions, the existence of a compact global attractor can be shown if the problem is interpreted
as a dynamical system in an appropriate Banach space of functions. In the case of the hyperbolic
system (lo), the existence of a compact global attractor can be proved via a Lyapunov function
argument [5], and these results can be extended to the Cattaneo system [20] and also to systems
of the form (22), provided the vector field f has gradient structure [5].
An important property of the linear diffusion equation is the maximum principle which, on the
one hand, guarantees the preservation of positivity, important from the modelling viewpoint, and
on the other hand, provides a strong tool for estimates and comparison of solutions. Hyperbolic
equations in general do not show such properties. This can be easily understood because second-
order hyperbolic equations are close to the wave equation or the telegraph equation (damped
wave equation) which, in the usual physics context, describe oscillation phenomena. On the
other hand, the systems (3) and (lo), with f E 0, do preserve positivity, and this fact leads to
comparison and invariance principles for systems of the form (lo), see [21].
In transport equations relating to applications in physics, the case has been studied where
the turning rate of particles depends on their individual velocity s. In a biological context
(modelling aggregation, chemotaxis, crowding and other density-dependent behavior), it makes
sense to assume that the model parameters depend on the particle density u. Then equation (10)
assumes the form
Ut’ + (y(u)U+)z = +)(1L- - u+) -I- if(U),
(23)
u; - (+f(U)u-)z = $&)(u+- u-) + g(u).
This problem has been studied in [22].
For a more detailed presentation of reaction random walk systems, see the forthcoming lec-
tures (231.
NOTE ADDED IN PROOF. The results in [12] have been completed in [24].

REFERENCES

1. C. Cercignani, Fl. Illner and M. Pulvirenti, The Mathematical Theory of Dilute Gases, Springer-Verlag,
(1994).
2. W. Alt, Biased random walk models for chemotaxis and related diffusion approximations, J. Math. Biol. 9,
147-177 (1980).
3. H.G. Othmer, S.R. Dunbar and W. Alt, Models of dispersal in biological systems, J. Math. Biol. 26, 263-298
(1988).
4. K.P. Hadebr, Travelling fronts for correlated random walks, Can.. Appl. Math. Quart. 2, 27-43 (1994).
5. T. Hillen, Qualitative analysis of hyperbolic random walk systems, Preprint, University of Tiibingen,
SFB 382, Report No. 43, (1995).
 Reaction Transport Equations 81

6. K.P. Hadeler, Reaction telegraph equations and random walk systems, In Stochastic and Spatial Structures
of Dynamical Systems, (Edited by S.J. van Strien and S.M. Verduyn Lunal), Kon. Nederlandse Akad.
Wetensch. Verhandl., Afd. Natuurkunde, 1 Serie, pp. 133-161, North-Holland, (1996).
7. H. Schwetlick, Reaction transport equations, Dissertation, Tiibingen, (1997).
8. S. Goldstein, On diffusion by discontinuous movements and the telegraph equation, Quart. J. Mech. Appl.
Math. 4, 129-156 (1951).
9. M. Kac, A stochastic model related to the telegrapher’s equation, (1956); reprinted in Rocky Mtn. Math. J.
4, 497-509 (1974).
10. C. Cattaneo, Sulla conduzione de1 calore, Atti del Semin. Mat. e Fis., Univ. Modena 3, 83-101 (1948).
11. D.D. Joseph and L. Preziosi, Heat waves, Reviews of Modern Physics 61, 41-73 (1988).
12. K.P. Hadeler, Travelling fronts in random walk systems, Forma (Tokyo) 10, 223-233 (1995).
13. Th. Gallay and G. Raugel, Stability of travelling waves for a damped hyperbolic equation, J. DZfl. Equ.
150, 42-97 (1998).
14. H. Schwetlick, On the minimal wave speed for reaction transport equations in dimensions n > 2. Preprint,
University of Tiibingen, SFB 382, Report No. 73, (1997).
15. A. KBllBn, Thresholds and travelling waves in an epidemic model for rabies, Nonlinear Analysis TMA 8,
851-856 (1984).
16. K.P. Hadeler, Travelling epidemic waves and correlated random walks, In Difleerential Equations and Ap-
plications to Biology and to Industry, (Edited by M. Martelli et al.), pp. 145-156, World Scientific, (1996).
17. K.P. Hadeler, Spatial epidemic spread, with slow infectives, In Ordinary and Partial Diflerential Equations,
hoc. Conf. Dundee, 1996, (Edited by R.J. Jarvis et al.), pp. 18-32, Pitman Research Notes, Math. Series
370, (1997).
18. T. Hillen, A Turing model with correlated random walk, J. Math. Biol. 35, 49-72 (1996).
19. T. Hillen, Nonlinear hyperbolic systems describing random motion and their application on the Turing
model, Dissertation Summarizes 1,121-128 (1996).
20. T. Hillen, Qualitative analysis of Cattaneo systems, Preprint, University of Tiibingen, SFB 382, Report
No. 59, (1997); Math. Models and Methods in Applied Sciences 3 (8) (to appear).
21. T. Hillen, Invariance principles for hyperbolic random walk systems, J. Math. Anal. Appl. 210, 360-374
(1997).
22. K.P. Hadeler, Reaction-telegraph equations with density-dependent coefficients, In Part. DZfl. Equ., Mod-
els in Physics and Biology, Mathematical Research, Vol. 82, (Edited by G. Lumer et al.), pp. 152-158,
Akademie-Verlag, Berlin, (1994).
23. K.P. Hadeler, Reaction transport systems in biological modelling, In CIME Lectures 1997, Mathematics
Inspired by Biology, (Edited by V. Capaaso and 0. Diekmann), Springer-Verlag, Florence, (1999).
24. K.P. Hadeler, Nonlinear propagation in reaction transport systems, Fields Institute Communications 21,
251-257 (1999).