Numerical Computation of Heteroclinic Orbits
Numerical Computation of Heteroclinic Orbits
Numerical Computation of Heteroclinic Orbits
North-Holland
Numerical computation
of heteroclinic orbits
Eusebius J. DOEDEL *
Applied Mathematics 217-50, California Institute of Technology, Pasadena, CA 91125, U.S.A.
Mark J. FRIEDMAN
Department of Mathematics, University of Alabama in Huntsville, Huntsville, AL 35899, U.S.A.
Abstract: We give a numerical method for the computation of heteroclinic orbits connecting two saddle points in Iw*.
These can be computed to very high period due to an integral phase condition and an adaptive discretization. We can
also compute entire branches (one-dimensional continua) of such orbits. The method can be extended to compute an
invariant manifold that connects two fixed points in Iw”. As an example we compute branches of traveling wave front
solutions to the Huxley equation. Using weighted Sobolev spaces and the general theory of approximation of
nonlinear problems we show that the errors in the approximate wave speed and in the approximate wave front decay
exponentially with the period.
Keywords: Heteroclinic orbits, traveling waves, numerical computation and continuation, weighted Sobolev spaces,
approximation of nonlinear problems.
1. Introduction
The problem of finding traveling wave front solutions of constant speed to nonlinear parabolic
partial differential equations is equivalent to the problem of finding trajectories that connect two
fixed points of an associated system of ordinary differential equations. Such a trajectory is an
example of a heteroclinic orbit, i.e. an orbit with several fixed points on it. The period of such
orbits is necessarily infinite.
In this paper we give an accurate, robust, and systematic method for computing entire families
of orbits connecting two saddle points. In a forthcoming paper [g] we shall consider, more
generally, the computation of manifolds connecting two fixed points in Iw”.
Calculations using the numerical methods described here are easily carried out with existing
continuation software. The method is essentially very straightforward, but its particular formula-
tion of the problem of computing the heteroclinic connection makes it very powerful. Orbits of
* On leave from Computer Science Department, Concordia University, Montreal, Quebec, Canada. Supported in part
by NSERC (Canada), A4274 and FCAC (Quebec) EQ1438.
high period can be computed effectively, and more importantly, entire branches of such orbits
can be computed very efficiently. This is due to the use of adaptive mesh selection [19] and the
use of a phase condition that keeps the wave front at the same location.
In our applications we use the software package AUTO. This package incorporates algorithms
for the numerical bifurcation analysis of differential equations. The first reference to the package
is in [6]. The most complete description of AUTO is given in [7], which contains an overview of
the algorithms, a large number of illustrative applications, and a user manual with detailed
examples of actual use of the software.
Traveling wave solutions to nonlinear parabolic equations arise in numerous problems of
physical interest, for example, in chemical-biochemical systems [1,3,9,16,17,18], flame propoga-
tion [4,21], etc.. We next review briefly some numerical results. In Miura [16], solitary waves for
the Fitz-Hugh-Nagumo (FHN) equations were calculated by a variant of the Crank-Nicolson
scheme, where the interval - cc < x < cc was replaced by a finite interval with an adaptive
outgoing wave boundary condition, and the “wave integrals” were used to determine the wave
speed and to measure the closeness of the computed solutions to the exact solitary wave solution.
Another method used in [16] was to solve a boundary value problem (in a moving coordinate
system) on a finite interval with boundary conditions chosen as in Lentini and Keller [15]. A
similar method was used in Hassard [lo] to calculate traveling wave solutions to the Hodgkin-
Huxley equations by using higher order approximations of the stable and unstable manifolds.
Recently Keller and his students developed efficient methods to approximate systems of ODE
and PDE on infinite domains by the problems with appropriate boundary conditions on finite
domains (see e.g. Lentini and Keller [15] and Hagstrom and Keller [12,13] and the references
there). In particular, in [13] appropriate boundary conditions were derived to calculate traveling
waves by solving the original parabolic PDE.
The numerical method proposed here is based on ideas similar to those in [15] and is a
generalization of the methods in [16] and [lo]. In the derivation of error estimates, we use
weighted Banach spaces, combining an approach in BabuSka [2] and the general theory of
approximation of nonlinear problems in Keller [14] and in Descloux and Rappaz [5]. The
numerical method is described in Section 2. Section 3 contains applications. An error analysis is
given in Section 4.
2. Numerical method
(2-5)
Fig. 1. Geometric interpretation of equations (2.1)-(2.5). Consider A, as fixed here. For A, = XT we assume the
existence of the heteroclinic connection in (b). Generically, perturbation of A, will produce either (a) or (c), depending
on the sign of the perturbation. If ~a and et are sufficiently small, then there exists a A, close to A; for which the
equations (2.1)-(2.4) (and (2.6)) can be satisfied (Here in case (a)). Furthermore, the radii cO and el can be chosen
such that the period of the orbit equals a given large value T, and such that the phase condition (2.5) is satisfied.
Above ( . , a) is the d2 inner product in IR*, the corresponding norm will be denoted by 1).)I.
Equation (2.1) is the differential equation, with u(o), f( -, -) E R! 2. Note that the time variable t
has been scaled so that it varies from 0 to 1. The actual period T therefore appears explicitly in
(2.1). There are two problem parameters, viz., X, and X2, and we want to compute entire
brunches (one-parameter continua) of approximate heteroclinic orbits. Equation (2.2) defines two
fixed points, w0 and wr, of the vector field. These are assumed to be saddle points. Thus the
Jacobian matrix f,( w0, A) has one positive and one negative eigenvalue and the same holds for
fu(w1, A). For definiteness let the eigenvalue p0 of f,( wl, A) as defined by (2.3a) be negative and
let the eigenvalue pI of f,( wo, A) defined in (2.3b) be positive. The corresponding eigenvectors
are called u0 and ur, respectively. We have uO, ur, wO, wr E R2. Equation (2.6a) then requires
that the starting point u(0) of the orbit u(t) lie on the line L(C) = w0 + EZ+,and at distance co
from the fixed point w0 (see Fig. 1). Note that e(e) passes through w. and that it has the
direction of the eigenvector uO_Equation (2.6b) imposes the corresponding requirement on the
158 E.J. Doedel, M.J. Friedman / Computation of orbits
endpoint u(l) at wi. Thus the equations (2.6) require the points u(l) and u(O) to lie on the linear
approximation of the stable and unstable manifold at wi and wO, respectively. Finally (2.5)
represents a phase condition. Its significance will be discussed below.
The unknowns w. and wi can be eliminated entirely from (2.1)-(2.5) by using (2.6). Then
(2.1)-(2.5) consist of two coupled differential equations with eleven side conditions, of which
(2.5) is an integral constraint. Since we are interested in an entire branch of orbits, a formal
count shows that we should have ten scalar variables in addition to the vector function variable
u(t) E R 2. These scalar variables are
A,, A,, 60, El, PO, l-5 E R? 00, u1 E R2. (2.7)
The period T is kept fixed in this continuation. For T large and e. and e1 small, each solution
on the branch represents an approximate heteroclinic connection. Example 2 illustrates such a
computation. If we want to increase the period T, then we can replace one of the problem
parameters, say X2, by T and thus use the scalar variables
A,> T, co, ~1, PO, ~1 E R, uo, q E R2. (2.8)
Such a calculation is illustrated in Example 1.
A heteroclinic connection u(t), - CC< t < 00, of two saddle points of f( u(t), h) is not
uniquely defined because for any real u, u( t + a) is also a connection. This is very similar to the
phase shift invariance of periodic solutions. The indeterminacy persists in the truncated problem:
if u(t) is a heteroclinic connection then both u(t) on [0, T] and u(t) on [u, T + a] are truncated
solutions of integration length T. We remove the indeterminacy by adding an appropriate
constraint which we shall call a phase condition in analogy with the periodic case.
One simple way to fix the phase is to set one of the components of u(t) at t = $ equal to some
appropriate value in the time-scaled equations (2.1). However, it is easy to give examples where
this phase condition fails to work. It also leads to multi-point boundary conditions although this
is a much less serious disadvantage.
A better phase condition is obtained by requiring that the current heteroclinic orbit “look
like” the previously computed orbit as much as possible. To be more precise, let C(t) denote the
previous orbit on a branch of heteroclinic orbits. Let ii( t + u) be the continuum from which the
current orbit is to be selected. Since 11u’(t) 11+ 0 exponentially as 1t I + co, a good measure of
how close ii and C are is the integral
D(u) = Jm II ii’@ + u) - i;‘(t) /I* dt.
-CC
The necessary condition for a minimum is dD( Q/da = 0. With u(t) = ii( t + a*) this necessary
condition can be written as
Computationally, (2.9) and (2.10) ( in a truncated form) lead to the same results. However, (2.10)
is more convenient for the error analysis.
As in [7], the effect of the integral phase condition (2.5) is that it minimizes translation of wave
fronts along a solution branch. This facilitates the adaptive mesh selection. In practice it allows
much greater steps to be taken along a branch of orbits.
3. Examples
so that u(O) = - : and u(l) = :. Continuation with AUTO, using the scalar variables (2.Q gives
a branch of orbits of increasing period T. Some computed orbits are shown in Fig. 2. Note that
the phase condition has the effect of keeping the increasingly sharper front in the same location.
This facilitates the automatic mesh adaption and allows bigger steps to be taken along the
solution branch. The mesh adaption also enables the computation of very large period. In Fig. 2
the orbit with label 7 has the largest period (T = 10000). Since the time variable t has been
scaled to the unit interval, this orbit looks like a step function.
Example 2. The Huxley equation. We apply the algorithm of Section 2 to the problem of finding
traveling wave fronts in the Huxley equation. This is the problem for which we give an error
analysis in the next section. The equation is given by
Wt=W,,+f(a, w), -co <z< co, t>o,
(3.4)
f(a, w) = w(1 - w)(w - a), 0 <a < 1. 1
We look for solutions to (3.4) of the form w( z, t) = u( z + bt), where b is the wave speed. This
gives the first order system
u;(X) = u&),
(3.5)
u;(x) = bu,(x) -_/-(a, u,(x)),
160 E.J. Doedel, M.J. Friedman / Computation of orbits
1.0
0.5
0.0
-0.5
-10
=I
0 00 0.10 0.m 0.30 0 40 0.50 0.60 0 70 0.80 0 90 1.00
TIME (SCALED)
Fig. 2. Some orbits along the computed branch of solutions to (2.1)-(2.6) for problem (3.1). The scalar variables are
given by (2.8). The periods of the orbits shown are (1) T = log(3) (the starting orbit), (2) T = 2, (3) T = 3, (4) T = 5, (5)
T = 10, (6) T = 100, (7) T = 10000. The calculation was done using AUTO [7] with 25 mesh intervals, 4 orthogonal
collocation points per mesh interval, and adaptive mesh selection.
u2
0.20
1
!!0
0 15 -
010
0°0
005
05
0 0
-0
-0 10
-0.15
-020 ’ I I I I I I 1 I I
0.00 0.10 0.20 0 30 0 40 0.50 0.60 0.70 0.80 0.90 1.00
Fig. 3. Phase plane representation of (3.5) when a = 0.5 and b = 0. For this choice of parameters the solutions form a
family of periodic orbits. In the limit, as the period T goes to co, the orbits approach a heteroclinic cycle with rest
points (0,O) and (1,O). Each of the two (approximate) heteroclinic orbits can be continued in the scalar variables (2.7)
with X, = a and X, = b. The result of this continuation is shown in Fig. 4.
E.J. Doedel, M.J. Friedman / Computation of orbits 161
A
Fig. 4. The two branches of traveling wave front solutions to (3.4). Along each branch the orbits actually remain
unchanged. The exact representation of the two branches is b = & fi( a - t). The calculation was done using AUTO
with 50 mesh intervals, 4 orthogonal collocation points per mesh interval, and adaptive mesh selection. The (fixed)
period used in this calculation was T = 1000.
where x = z + bt, and ’ = d/dx. If a = 0.5 and b = 0 then (3.5) has a family of periodic orbits of
increasing period. Its phase portrait is shown in Fig. 3. One of the solutions is a double
heteroclinic connection (in fact a heteroclinic cycle). The first of the two heteroclinic orbits has
the exact representation
exp(tax)
u*(x)= 1 + exp(+fix) ’
u*(x) = u;(x), -00 -=cx < co. (3.6)
The second heteroclinic connection is obtained by reflecting the phase plane representation of
the first with respect to the horizontal axis u2 = 0 (see Fig. 3). The two exact solutions can be
used as starting points in the continuation algorithm defined by (2.1)-(2.6). The scalars are now
given by (2.7). The resulting two branches are shown in Fig. 4. It happens that the orbits actually
remain unchanged along the two branches. Furthermore, the branches have the analytical
representation b = + fi( a - t).
[20] and the discussion there). Our approach will be shown on a model problem, approximation
of wave fronts for the Huxley equation (3.4).
In the preceding section we computed solutions to (3.4) of the form
w(z, t) = n(z + bt), (4.1)
where b is the waue speed. It is well known (see e.g. [20]) that for given a there exists ii(x) which
satisfies for some b > 0
u
-” - bii’ +f( a, ii) = 0, -cfooxx<,
ii(-cc) =o, n(cc) =
P-2)
1,
where x = z + bt and ’ = d/dx. Let x E CM(R) be such that for some T, > 0, T_ -c0
1 forx>T+,
x(x> =
10 forx<T_.
Then U(X) = ii(x) - x(x) satisfies
-z/‘+bu’-f(a, u+x)
U(-cc)=z4(cc)=o.
We want to solve equation (4.3) in weighted Sobolev spaces. To determine the right weight
function, we need to know the asymptotic behavior of the solutions. Let (a,, b,, uO) be a
solution of (4.3). We first rewrite (4.3) as a first order system, setting u1 = uO,
24;=ZA,,
ZA;= b()U* -f( a,, zq+ x) - x” + box’. i (4.4)
0 i
4 = -fu,(ao>%b)
For x + - cc the matrix in (4.5) becomes
+x(x)) b0
ii
“1
v2
I
.
(4.5)
Its eigenvalues are given by the roots of p2 - hop - a,. Since lim u(x) = 0, we choose the
positive eigenvalue pi = $( b, + {R). Corresponding to this;iotmwe have a solution of the
form
vi = ePIX, v2 = j+ eP+. (4.6)
Similarly, for x + cc the matrix in (4.5) becomes
i -(-lo+n,, lo 1 .
Then equation (4.5) has a solution of the form
vi = e’Qx, v2 =pLlePoX, where p. = 5 ( b o-@TGi+o. (4.7)
For some p. < 0, pi > 0 we next set p = (po, ~i)~,
,-2PP x -=z0,
wJx)= ’ (4.8)
i e-%x 3 x 2 0.
E.J. Doedel, M.J. Friedman / Computation of orbits 163
Analogously to BabuSka [2] we now let He,cl = HL+(lR), L’> 0, integer, be the Banach space of all
functions u such that
Remark 1. In the solution of equation (4.3) we employ a continuation procedure with a and b as
parameters. Suppose we know the range 1, of a and the range I, of b. Then we shall need
u E H1,’ for all a E I,, all b E Ib. To achieve this we can choose p = (p,,, ~i)~, pLo< 0, pi > 0
from
b+\lb2 b-m +E
pl = inf - c, PO = sup > (4.9)
LlEI” 2 LlEI”
2
bs Ih btl,
We next want to reformulate equation (4.3) as an operator equation in HI+. We first define
the linearization
A(x)u = -u”+ bou’-fU(ao, sign x)u, (4.10)
where
I, x > 0,
sign x =
( 0, x 6 0,
about a solution (a,, b,, uo) of (4.3). We can now rewrite (4.3) as
B(U, u) = (Au + uu, u) = lw [u’u’+ bou’u -fU(uo, sign x)uu + uuu] dx. (4.13)
-LX
Problem (4.3) (or (4.11)) admits the variational formulation: given a, 0 -C a -C 1, find (b, u) E R
x H’+ from
By a slight modification of the proof of Theorem 3.1 in [2], using the inequality
By Lemma 2 we can now define the continuous linear operator T: HP’,’ + H’+ by
B(Tf, u) = (f, u), VUE H1,-p, Vf E H-l+. (4.17)
Then an equivalent form of the problem (4.14) is: given a, 0 < a < 1, find (b, u) E R X H”” from
F(a, b, u) = u + TG(a, b, u) = 0. (4.18)
Let (a,, b,, uO) E R x II3 x H’,’ be a solution of (4.18) i.e.,
F(%Y b,, uo> =0 (4.19)
and define the linearized operator H’,” + H’2P
D,F(a,, b,, u,,) -I+ TG,(a,, b,, uO)
= I- T[o +fU(aO, u0 + x) -fU(ao, sign x)]
=I- T(a+a”(*)), (4.20)
where
t(x) =f,(aO, u. + x) -fu(aO, sign x)
DenoteRu={cu: CER}.
Theorem 1. (i) D,F( a,, b,, uo) = I - T( a + a”) : H’+’ + H1” is singular and
HI = Ker D,F(a,, b,, uo) = W60,
which, obviously, holds. Next, (4.25b) follows from the decomposition (4.23), and (4.26) follows
from (4.29). 0
e(x)
= SW [~(t)&,(t) + u’(t)&(t)] dt, x = (b, u). (4.31)
--oo
For any 5 E R x H’+ we have
DX@(% x& = (d(5), D,F(a,, x,)0. (4.30b)
From (4.25b) and (4.26) by the Banach Theorem we deduce the following corollary.
Remark. In the derivation of error estimates we use the phase condition in the form
which differs from the condition (2.9) (or (2.10)) used in actual computations. In the notation of
this section (2.10) takes the form
00
$-fi;,)~;+($-~;)~~] dt=O.
J --oo](
Computations indicate that this condition leads to more accurate results than (4.31’). One can
justify the use of (2.10) instead of (4.31’) by assuming additional regularity of the solution.
Our numerical method (2.1)-(2.6) essentially consists in replacing for large 1t 1 the
[-fU(a,l)-a+f,(a,,I)]u+(b-b,)u’, t’T+
GT(t, a, b, U) = G(t, a, b, u>, T_<t<T,,
[ +,(a, 0) -(J t-f&o, O)]u + (b-b&‘, t< T-3
(4.32)
G(x) = j--~[u(t)&(t)
+u’(t)&(t)] dt, x = (6 a>. (4.36)
E.J. Doedel, M.J. Friedman / Computation of orbits 167
To establish existence, uniqueness and error estimates for the approximate problem, we shall
use a simplified version of the implicit function theorem in Descloux and Rappaz [S] and in
Keller [14].
Theorem 2. Let E, and E, be two real Banach spaces and let I C Iw and D c E, be open subsets of
R! and E,, respectively. Let @, @,, : IX D + E, be a C’ mapping and a family of mappings,
respectively, that satisfy
(i) @(a,, x0) = 0, (a,, x0) E I x D;
(ii> Q@(a,, x0) is an isomorphism of E, onto E,;
(iii) For an open subset A C I
lim sup 11D,@( a, x0) - D,@,(a, x0) II = 0;
fl+m aEA
(iv) lim sup (1@(a, x) - @,(a, x) (1 =O, Vx E D.
n-m a=A
Then the problem
@(a, x) =0
has a unique C’ solution branch ((x(a), a) : a E A } that contains (a,, x0), moreover, there exists
n so that for all n >, N the problem
@,(a, xn) = 0
has a unique C’ solution branch {(x,,(a), a) : a E A } and
IIx(a) -x,(a) II G C II @k44 a>II,
where C is a positive constant independent of n.
Lemma 3. Let x = (b, u) E 1, X HITP, a E I, for some finite intervals I,, I, c R, see (4.9). Then
we have
)I @(a, x) - Qr(a, x) 1)n~.~< ce(u, T, T+)(eT-‘O + eT+P1), (4.37)
where p,, < 0 and p, > 0 are as in (4.9), c does not depend on a, b, T-, T,, and E( u, T, T+) + 0
as T,+oo, T-4 -00;
]]D,@(a, x0) - D,@,(a, x0) ]I L(RxH~+)< c(eTmPO+ eT+‘l) (4.38)
uniformly in a E I,.
Proof. From the definitions (4.18) and (4.34) of F and F,, respectively, we have
F(a, x) - F,(a, x) = T(G(a, x) - G,(a, x)).
Using that T: HoSP + H1+“, defined by (4.17), is bounded and the definition (4.32) of G,,
IV%, x> -6-b x> II H’.‘G IITII IIGh x> - %-(a, x> II H”+((-w~T_]u[T+,~))
(4.39)
168 E.J. Doedel, M.J. Friedman / Computation of orbits
4suPIfuuh 4 ll~211H~.‘[7+,m)
-,y; l4Gl Ila+‘,T+,m). 0 (4.40)
4 +
Remark. Here, with some abuse of notation, we have assumed that I f,, I is bounded uniformly
in a, u. What we actually mean by this is the following. We show below, without using
I fuuI Gc, that the solution of the exact problem is bounded. Hence, without changing the
solution we can modify f for large I u 1, I a I so that ( f,, ( 4 c holds.
Similarly,
By the definitions (4.31), (4.36) and (4.21) of L’, 8, and &, respectively, for 5 = (b, U) E Ib X H”‘,
we arrive at
Now (4.37) follows from (4.41), (4.42) and the definitions (4.30a) and (4.35) of @ and @r,
respectively.
E.J. Doedel, M.J. Friedman / Computation of orbits 169
Similarly, (4.38) follows from (4.42) and the estimates for any u E H’,P of the type:
Theorem 3. Let x = (b, u) E [0, W) X II’*‘” and x0 = (b,, uo), where (a,,, x0> E (0, 1) X [O, m) X
H’+ is a solution of (4.3) or (4.11), or ((4.14) or (4.18)).
( (i) Then there exist positive constants Tz, - TO, Z,, a, K and two unique maps x(a),
x=(u) E R x EP”, ) a - a, I < ii, satisfying, respectively, the conditions
@(u, x(u))=O, Ijx(u)-xO1(R,H1.P<a for Iu-uOI <a,, (4.43)
Remark. The results of Theorem 3 apply to a branch (calculated in Section 3) which has the
analytical representation
-JZ(u-l/2), o<u<+,
b=
Iqu-l/2), :<a<1.
References
VI D.G. Aronson, Density dependent interaction-diffusion systems, in: Dynamics and Modelling of Reactive Systems
(Academic Press, New York, 1980) 161-176.
PI I. BabuSka, The finite element method for infinite domains, Math. Comp. 26 (1972) l-11.
131 M. Bieterman and I. BabuSka, An adaptive method of lines with error control for parabolic equations of the
reaction-diffusion type, J. Comput. Physics 63 (1986) 33-66.
141 J. Buckmaster and G.S.S. Ludford, Theory of Laminar Flames (Cambridge University Press, Cambridge, 1982).
151 J. Descloux and J. Rappaz, Approximation of solution branches of nonlinear equations, R.A.Z.R.O. 16 (1982)
319-349.
170 E.J. Doedel, M.J. Friedman / Computation of orbits
[6] A program for the automatic bifurcation analysis of autonomous systems, Cong. Numer. 30
(1981) 265-284.
[7] E.J. Doedel and J.P. Kernevez, AUTO: Software for continuation and bifurcation problems in ordinary
differential equations, Applied Mathematics Report, California Institute of Technology, 1986, 226 pages.
[8] E.J. Doedel and M.J. Friedman, Numerical computation and continuation of invariant manifolds connecting
fixed points, in preparation.
[9] E.J. Doedel and J.P. Kernkez, A numerical analysis of wave phenomena in a reaction diffusion model, in: H.G.
Othmer, Ed., Nonlinear Oscillations in Biology and Chemistry, Lecture Notes Biomath. 66 (Springer, Berlin, 1986)
261-273.
[lo] B.D. Hassard, Computation of invariant manifolds, in: P.J. Holmes, Ed., New Approaches to Nonlinear Problems
in Dynamics (SIAM, Philadelphia, PA, 1980) 27-42.
[ll] D. Henry, Geometric Theory of Semilinear Parabolic Equations, Lecture Notes Math. 840 (Springer, Berlin, 1981).
[12] T.M. Hagstrom and H.B. Keller, Asymptotic boundary conditions and numerical methods for nonlinear elliptic
problems on unbounded domains, Math. Comput. 48 (1987) 449-470.
[13] T.M. Hagstrom and H.B. Keller, The numerical calculation of traveling wave solutions of nonlinear parabolic
equations, SIAM J. Sci. Stat. Comput. 7 (1986) 978-988.
[14] H.B. Keller, Approximation methods for nonlinear problems with application to two-point boundary value
problems, Math. Comput. 29 (1975) 464-474.
[15] M. Lentini and H.B. Keller, Boundary value problems over semi infinite intervals and their numerical solution,
SIAM J. Numer. Anal. 17 (1980) 557-604.
[16] R.M. Miura, Accurate computation of the stable solitary wave for the Fitz-Hugh-Nagumo equation, J. Math.
Biology 13 (1982) 247-269.
[17] H.G. Othmer, Nonlinear wave propagation in reacting systems, J. Math. Biology 2 (1975) 133-163.
[18] J. Rinzel and D. Terman, Propagation phenomena in a bistable reaction-diffusion system, SIAM J. Appl. Math.
42 (1982) 1111-1137.
[19] R.D. Russell and J. Christiansen, Adaptive mesh selection strategies for solving boundary value problems, SIAM
J. Numer. Anal. 15 (1978) 59-890.
[20] D.H. Sattinger, On the stability of waves of nonlinear parabolic systems, Aduances in Mathematics 22 (1976)
312-355.
[21] D. Terman, Traveling wave solutions arising from a combustion model, IMA Preprint Series 216, University of
Minnesota, 1986.