Arch. Rational Mech. Anal. 117 (1992) 97-117.

9 Springer-Verlag 1992

Stability of Travelling Fronts

in a Model for Flame Propagation
Part I: Linear Analysis
HENRI BERESTYCKI, BERNARD LARROUTUROU
JEAN-MICHEL ROQUEJOFFRE

C o m m u n i c a t e d by H. BREZIS

Abstract
We investigate the stability of travelling wave solutions of the multidimen-
sional thermodiffusive model for flame propagation with unit Lewis number.
This model consists in a system of two nonlinear parabolic equations posed
in an infinite cylinder, with Neumann boundary conditions. In this paper, we
prove that every travelling wave solution of this model is linearly stable. Our
tools are exponential decay estimates for solutions of elliptic equations in a
cylinder, and the Maximum Principle for parabolic equations.

1. Introduction
This is the first of a series of articles aimed at proving linear and nonlinear
stability results for travelling wave solutions of a multidimensional flame pro-
pagation model. We are interested in solutions of the system of semilinear
parabolic equations
ut - • + o~(y) Ux = v f ( u ) ,
(1.9
v~ - Av + ~(y) vx = - i f ( u )
in a cylinder X = {(x, y) ~ • x a)}, co being an open bounded regular subset of
[RN-i, with N >1. System (1.1) is associated with boundary conditions which
will be made precise later.
Such equations are related to the classical thermodiffusive model for flame
propagation, which we briefly describe now. System (1.1) is a particular case
(when L e = 1) of the system
ut - Au + c~(y) ux --- i f ( u ) ,
(1.2) &v
vt - + o~(y) v~ = - v f ( u )
Le
in X.
98 H. BERESTYCKI,B. LARROVTtrgOU& J.-M. RoQw~orr~

System (1.2) only takes into account, for simplification, a one-step

chemical reaction of the form A - ~ B . The function u(t, x, y) is a normalized
temperature, and the function v(t, x, y) is the mass fraction of the reactant
A. The given function o~(y) is the velocity of the gas flow, which is assumed
to be parallel to the x-direction, and the term v f ( u ) represents the rate at
which the chemical reaction proceeds. Finally, the real number Le > 0 is the
Lewis number, i.e., the ratio between the thermal and molecular diffusion. In
the sequel we assume that Le = 1.
The thermodiffusive model is based on the isobaric approximation and the
constant-density approximation. Since flame typically propagates at a rate well
below the sonic speed, the spatial variations of the pressure are typically small,
and are neglected in the isobaric approximation. In one spatial dimension, the
introduction of a Lagrangian coordinate decouples the reaction-diffusion equa-
tions for the combustion variables (temperature and mass fraction of the reac-
tant) from the equations involving the hydrodynamic variables (density,
pressure, and velocity) and consequently permits a detailed, rigorous mathe-
matical analysis of the governing equations; see [15]. In higher dimensions the
mathematical analysis of the complete fully coupled system is still beyond
reach, even with the isobaric approximation. This difficulty can be overcome
by the introduction of an additional approximation that decouples the
hydrodynamical equations from the reaction-diffusion equations. This purpose
is served by the constant-density approximation, in which density fluctuations
of the gas are neglected. Thus the resulting thermodiffusive model is essentially
valid for situations in which the hydrodynamical effects play a role subservient
to those of the reactive and diffusive effects. Nevertheless, the thermodiffusive
model retains many important features of flame propagation. We refer to [21]
for the derivation of the equations and for the discussion of the thermodiffu-
sive model; see [5] for a summary of a few classical flame propagation models
and related mathematical resuks.
Under suitable assumptions on function f to be specified in Section 2,
system (1.1) admits travelling wave solutions, i.e., solutions of the form
(1.3) u(t, x, y) = O(x + ct, y), v(t, x, y) = q/(x + ct, y ) .

This result is proved in [6]; our goal is to prove that these travelling wave solu-
tions are stable under "small" perturbations. For this purpose we first linear-
ize system (1.1) formally about a travelling wave solution (~, qz); this yields
a new set of equations which can be written in the compact form

dU
-- + LU=O,
dt

where L is a second-order linear elliptic operator. A crucial step is the study

of the eigenvalues of operator L; the main result of the present paper is that
every eigenvalue of L except 0 has a positive real part. We prove this by deter-
mining the asymptotic behaviour as [x[ --' + oo of possible solutions of the
eigenvalue problem, and by using various versions of the Maximum Principle
for parabolic equations.
 Travelling Fronts in Flame Propagation 99

Before we state precisely the assumptions and the result, let us recall that
there is a vast literature devoted to this type of problem for a single equation
in the one-dimensional setting

u~ - Ux~ = g ( u ) in [R

with the usual boundary conditions

u(t,-oo)=O, u(t, + o o ) = t.

We refer here to the celebrated paper of KoLMoco~.ov, PETROVSKn, & PIS-

KUNOV [14], and to [12] and [19] for more recent global stability results with
various assumptions on the function g. See also [3], [13], and [20] for other
aspects of the theory.
Travelling-wave solutions of multidimensional parabolic problems have been
studied by several authors (see, for instance, [4]). Nevertheless, the results
stated in the present paper and in forthcoming papers appear to be the first
to address the question of stability of travelling waves in higher dimensions,
when there is a genuine y-dependence for the function o~. Let us also point
out that while the rigorous results in one space dimension have long been fore-
seen by formal asymptotic analyses for high activation energies (see, for exam-
ple, BUCKMASTER ~r LUDFORD [10]), no such formal calculations seem to be
available for the multidimensional problem.
This paper is organized as follows. In Section 2 we state the assumptions
and our main result. Section 3 is devoted to the study of the one-dimensional
case, where a classical, very simple proof of the result is available. In Section 4
we recall fundamental results on the asymptotic behaviour of solutions of
elliptic equations in a cylinder, and we prove in Section 5 the linear stability
of the travelling-wave solutions of (1.1).

2. Assumptions, Linearisation and Main Result

We investigate the following problem:

(2.1.a) u t - Au + a ( Y ) U x = vf(u) ~ for (x, y) ~ X ,

vt 5 v + a ( y ) vx = - v f ( u ) )

Ou Ov
(2.1.b) - -- 0 for (x, y) ~ ~ x 0 o ) ,
Ov Ov
u(t,-co, y)=0, u(t, +oo, y ) = 1 )
(2.1.c) for y E c~.
v(t, - o % y) = 1, v(t, + 0% y) oJ
Here v is the outward unit normal on OX. The Neumann boundary conditions
(2.1.b) mean that the walls of the tube are adiabatic.
100 H. BERESTYCKI,B. LARROUTUROU& J.-M. ROQUEJOFFRE

We impose the following conditions on functions c~ and f :

(2.2) ~ f C~ for some d > 0;
(2.3.a) f f CI([R), and 3 e > 0 such that ffC1/~([1 - e , 1 +el);
(2.3.b) 3 0 f ] 0 , 1[ such that f = 0 in [0, 0] and f > 0 in ]0, +co[.
The real number 0 is an "ignition temperature", below which no chemical
reaction may occur.
Let us comment on the assumptions. The regularity assumptions (2.2) and
(2.3.a) for functions c~ and f seem to be minimal (in the following paper [18],
in which nonlinear stability results are proved, further regularity is required).
Assumption (2.3.b) is crucial for the study of existence and uniqueness of
travelling-wave solutions of (2.1).
Set g(u) = f ( u ) (1 - u) ; if a solution (u(t,.), v(t, .)) satisfies u(0, .) +
v(0, . ) = 1, then u + v = - 1 , and u is a solution of the single equation

(2.4) ut - Au + c~(y) ux = g(u) for (x, y) f X ,

with the same boundary conditions as (2.1.b, c). Besides this, travelling wave
solutions (qL q/) of (2.3) satisfy
--&~b + (c + c~(y)) ~bx = gtf(th) "}
for (x, Y) f i x
(2.5) -/x~u + (c + ~(y)) qJx = - gtf(4~))
/

and (2.1.b, c). It is straightforward to show that q5 + q / = 1; therefore q~ is a

solution of
(2.6.a) -&4~ + (c + a(y)) qSx = g ( 4 ) for (x, y) f X ,

(2.6.b) - 0 for (x, y) ~ ~ x a c o ,

Ov
(2.6.c) t h ( - o o , y) = 0 , ~b(+oo, y) = 1 for yEc~.
In [6] BERESTYCKI,LARROUTUROtJ,& LIONS proved the existence of a solution
(c, q~) for problem (2.6); in [8] B~a~ESTYC~:I& NIRENBER6 proved the unique-
ness of the solution in the following sense: c is unique and q~ is unique up
to a translation in the x-direction. In the whole sequel we shall, without loss
of generality, work with the unique solution of (2.6.a-c), which satisfies
(2.6.d) maxO(0, y) = 0.
yE6)
Setting
1 ~c~(y) dy
~> = I~1

one has the two additional estimates:

(2.7.a) c + (a) > 0,
(2.7.b) 4~x > 0 in ~,.
 Travelling Fronts in Flame Propagation 101

Estimate (2.7.a) can be found in [6]; (2.7.b) is proved in [8], and shows that
condition (2.6.d) uniquely determines ~b.
Let us consider equations (2.1). We first rewrite them in the reference frame
of a travelling wave (q~, ~), and we linearize them about the (now steady) solu-
tion (q~, ~u). Specifically, we set

(u(t,x,y))=(q~(x,y) +a(t,x,y)).
v(t, x, y) k,~(x, y) + g(t, x, y)
System (2.1) then becomes

at - Aa + (c + o~(y)) ax = f ' ( r g,a +f(r g ) for (x, Y) fizZ,

(2.8) gt - A g + (c+ot(y))gx=_f }
together with (2.1.b), (2.1.c) with u and v replaced by a and g. The functions
a(t,.) and ~7(t, .) are the "variations" of the solution (u(t,.), v(t,.)) about
the equilibrium solution (~b, q/).
We investigate the "eigenmodes" of system (2.8), i.e., we look for solutions
of (2.8) in the form
a(t,x,y)) {e-ltu(x,y))
g(t, x, y) = ~e-Xtv(x, y) '
with )t ~ C. (Notice that the new functions u and v do not depend on t any
more.) Hence we are looking for perturbations of steady solutions of (2.1)
which grow (or decay) in time at the rate e-Xt; the functions u and v satisfy
the system

- A u + (c+a(y))u~-f'(ch)q/u-f(ch)v=2~l for
(x, y) eZ,
(2.9.a) - A v + (e+a(y))vx+f'(c~)gtu+f(dp)v ),
with boundary conditions

Ou Ov
(2.9.b) - - 0 for (x, y) e [Rx Oco.
Ov Ov
Since the solutions of the initial nonlinear problem have finite limits as j x[ --, + 0%
it would be natural to look for u and v satisfying

u(-o%y) = u(+Qo, y ) = 0 ~ for y f i ~ .

v ( - oo, y) v ( + oo, y) O)
In fact, we shall see that these conditions are automatically satisfied provided
that the functions u and v are bounded. Therefore, in the sequel, we shall say
that (&, u, v) is a solution of (2.9) if ()~, u, v) satisfy equations (2.9.a,b), if
(u, v) # (0, 0), and if u and v also satisfy
(2.9.c) u, V E W2(~P(~ ') t~ W l'~176
(~v') Vp _>1.
It is clear that (~bx, ~vx) is a solution of (2.9) with 2 = O. Our main result
is the following:
102 H. BERESTYCKI,B. LARROUTUROU& J.-M. ROQUEJOI~I~RE

Theorem 1. L e t (2, u, v) be a solution o f (2.9). Then either Re ;~ > 0 or 2 = O.

This means that the travelling-wave solutions of (2.1) are linearly stable.
In the following paper ([18]), it is proved that a nonlinear stability result holds
for the evolution problem (2.4) for these travelling waves: Any solution of
(2.4), with initial data close enough to the travelling front ~b, gt actually con-
verges towards a translate o f q~.
In the sequel, for convenience, we set

fl(y) = c+ a(y).

3. The One-dimensional Case

Here we have a ( y ) = constant; therefore we may, without loss of generali-

ty, assume that oe -- 0. Problem (2.9) is equivalent to finding a bounded solu-
tion of

- u " + cu' - f ' ( t b ) ~u - f ( 4 ~ ) v = 2"~ for x E JR,

(3.1) - v " + cv' + f ' ( r q/u + f ( r v = 2 v)

Moreover, we have c > 0 (see [7]).

Theorem 1 is proved in three steps. Assume that there exists a solution
(2, u, v) of (3.1) such that Re 2 _< 0 and 2 ae 0.
Step 1: Reduction o f problem (3.1). We set W = u + v; the function W ( x ) is a
solution of

(3.2) -W" + c W ' - 2W = 0 in R.

For any complex number z such that z r we as usual, denote by x/zz the
unique complex number y such that y Z = z and Re y > 0. Setting

e -T- x/~c2- 42
/~• (2) =
2

we immediately obtain
(3.3) W ( x ) = Ae ~+(x)x + Be ~-(~)x

with A, B E C. A straightforward computation yields Re/~+ (2) < 0, Re ~_ (2) > 0.

The function W being bounded, (3.3) implies W ~ 0, and therefore (recall that
q~ + q / = 0) problem (3.1) reduces to finding a bounded solution of

(3.4) - u " + cu' - g'(~b) u = 2u in R.

Step 2" 2 ( R. Since ~ ( 0 ) = 0, and since the function q~ is increasing, the term
g'(q~) vanishes for x =< 0. Since Re 2 =< 0, we get
(3.5) u ( x ) = u(O) e u-(~)x for x ~ R_.
 Travelling Fronts in Flame Propagation 103

Setting
r

Ul(X ) = e-iXu(x),
we rewrite equation (3.4) as

(3.6) -u~'+ [c---;-g'((o)Jux=2Ul in R.

Moreover, ul decays exponentially as I x l - ~ + oo. It is easy to see from (3.5)

that u~ decays exponentially as x ~ - c o ; similarly (3.4) implies that u' decays
exponentially as x---, + oo. We therefore may multiply (3.6) by Ul and integrate
the result on R; taking the imaginary parts of b o t h sides of the new equation
then yields I m 2 = 0.
Step 3: Conclusion. From Step 2 we infer that 2 < 0; furthermore, we may now
assume that u is real-valued. From the explicit expression (3.5) we deduce that
u/cb' vanishes as x ~ - oo. Setting
--C -4- N/C 2 -- 4 ( 2 Jr" g'(1))
r/~ (;L) =
2

we get, by the standard theory of ordinary differential equations (see [11]),

(3.7) u(x) = O(e -n+(~)x) for xfi N+.

Recall that from assumptions (2.3.a) and (2.3.b) we have g'(1) < 0. Therefore
u/4Y vanishes as x - ~ + oo. This time we set ul = u/4Y. ul is a solution of

(3.8.a) -u~' + c - 2 ~,---] u~ - 2u 1 = 0 in R,

(3.8.b) Ul(-~) = u l ( + oo) = 0.

Set M = sup u 1(x) ; assume that M . 0. Condition (3.8.b) yields the exis-
xfirE
tence of Xo such that us (x0) = M; moreover, u] (x0) = 0 and u~' (x0) -< 0. There-
fore M = ul(xo) < 0 from (3.8.a). Similarly, if we set m = inf ul(x), we get
x~lE
m > 0 if m * 0. Since u . 0, we infer that m > M, which is of course absurd.
Therefore 2 >_ 0, and Theorem 1 is proved in the one-dimensional case. []

We treated in detail this simple particular case chiefly because the proof
contains two of the main ideas of the p r o o f for the general case, i.e., that
a solution (2, u, v) of (2.9) with R e 2 < = 0 and 2 ~ : 0 is such that: (i)
u + v = 0, (ii) u decays faster than q~x as [x] ~ + oo. In the next section we
accordingly recall some general results on the exponential decay of solutions
of elliptic equations in a cylinder. However, the rest of the above demonstra-
tion does not apply, because of the y-dependance of the function ~; additional
ingredients will therefore be needed.
104 H. BERESTYCK92B. LARROUTUROU8r J.-M. ROQUEJOFFRE

4. K n o w n Results on the Exponential Decay

of Solutions of Elliptic Equations in a Cylinder

A very powerful tool for deriving some a priori information on the asymp-
totic behaviour of possible solutions of (2.9) is the theory of Ao~aoN &
NIRENBERa [2], and the extensions given by P~zY [16]. In this paper, we are
mainly interested in the solutions of second-order elliptic equations; therefore
the results that we are going to recall are only given in this framework. Never-
theless, they hold for much more general situations. In particular they are true
for elliptic equations of order 2m, and even for weighted elliptic equations (see
[2]); the last category includes parabolic equations.
We refer to [2] for the general results and their proofs; we also refer to
the beginning of [8], where one may find a summary of these results.
We set 27+ = ~+ • We are interested in the solutions of equations of
the following type:

(4.1 .a) -Au+B(y).~?u+a(y)u=O in 27+,

Obl
(4. l.b) -- = 0 in ~ + • 0co,
Ov

where a(y) and B(y) are HOlder-continuous functions. The function a(y) is
complex-valued, and the function B(y) takes its values in C u. O f course, all
the results that we are recalling have their analogues in 2;_ = ~ _ • ~o.
For B E [R and p > 1, let u E W~OP(s be a solution of (4.1) such that

em[[u(x,.)llLp(co) = o ( 1 ) as x ~ +oo.

Following BERESTYCKI& NIRENBERG [8] we denote the set of such solutions by

2 ( p , p, 27+). Finally, if X is a Banach space contained in cl(c~), we set

X = EX : - = 0 on 3 .
3v

An exponential solution of (4.1) is a solution o f the form

(4.2) e(x, y) = eZXp(x, y),

where p (x, y) is a polynomial in x with y-dependent coefficients. It is proved
in [2] that the set of exponents 3~ of exponential solutions is discrete; moreover,
in every strip of C of the form {Re 2 __<K}, there is only a finite number of
exponents corresponding to a nonzero exponential solution. The space of ex-
ponential solutions with a given exponent 2 is finite-dimensional.
Let (eDk___l be the sequence of all nonzero exponential solutions of (4.1),
written in the form ek(x, y) = e~Xpk(x, y), such that Re 2k _-< -/2 and such
that the sequence (Re2~)k__>1 is nonincreasing. The following theorem is
proved in [21.
 Travelling Fronts in Flame Propagation 105

Theorem 4.1. Let u be a solution of (4.1) in 2 ( l t , p, ~S+). Then

(i) u, considered as a function of x, with values in W2'p (co 0 ) , , is analytic.
(ii) There exists a sequence of exponential solutions (ek)k__>l such that the "for-
mal Fourier series"
+r
E ek(x'" )
k=l
converges to u as x ~ + co in the following sense: For a > l~, let {2~h_<l_<Lbe the
exponents of the exponential solutions ek in the Strip
-a < Re 2.~ < -/~,
and let (el, m)l<=m<=Ml be the sequence of exponential solutions with the same expo-
nent ~1. Let e > 0 be such that e - a < Re 2l for 1 <_ k <_L. Then the following
inequality holds:
(4.3)
oj [- L Ml -] 2 oi-1 u
--
3x j [ u ( x , ' ) - E ~ el,re(x, ' ) ~ 2 N
CE Ox,_ 1 ( 0 , ") e (e-a)x
l=1 m=l w'P(~) i=1 w2-i,P((,o)

for all integer j >_ O. The constant C depends on l~, j, P, co, e, and a but not on u.

We also have to take into account inhomogeneous problems of the form

(4.4.a) - A u + B(y) . VUx + a(y) u = f ( x , y) in 27+,
Ou
(4.4.b) -- = 0 in N+x0co,
Ov
with
IIf(x, Ce -bx
for some b > 0. We use the following result, proved by P~zy [16].

Theorem 4.2. Let uEL2([P+, H2(co)) be a solution of (4.4). Then u may be

written in the form
U : U 0 -~ U , ,

with the function u ~ satisfying (4.1), and the function u, satisfying (4.4). Further-
more, for every e > 0 the following estimate holds:
(4.5) I1u, (x,)11 cl, (Ex,x+l <= Cce (E-b)x
for all x > 0 and for all e > O.

We now turn to solutions of elliptic equations of the form

(4.6.a) - & u + b(y) Ux + a(y) u = 0 in X+,
Ou
(4.6.b) -- = 0 in ~ + x a c o ,
av
where the functions a(y) and b(y) are H01der-continuous and real-valued.
106 H. BERESTYCKI, B. LARROUTUROU& J.-M. ROQUEJOFFRE

Let L be any elliptic operator in co with Neumann boundary conditions on

309; we denote by pl(L) its first eigenvalue. BERESTVCKI&NIRENBERG [8]
proved that if Pl ( - A y + a) is nonnegative, then the exponential solutions of
(4.6) must have the form

ek(X, 9) = eXk~q/k(y), 4k ~ R.

Following [8] we call an exponent 4k a principal generalized eigenvalue if

q/k(Y) > 0 for all y ~ ~. We have

Theorem 4.3. ([8]) (i) If a =-- 0 and (b) > 0, then there exists a unique nonzero
principal generalized eigenvalue A_ > O.
(ii) If Pl ( - @ + a) > O, then there exist exactly two principal generalized eigen-
values A_ > )~+.
(iii) In both cases (i) and (ii) the space of the exponential solutions associated with
each generalized principal eigenvalue is one-dimensional.

From now on we denote by 4_ and - 4 + the positive and negative prin-

cipal generalized eigenvalues of problem (4.6) with b = fl, a = 0 and with
b = fl, a = -g'(1), respectively. Let eZ-~g/_(y) and e-X+xg/+ (y) be exponen-
tial solutions associated with 4_ and - 4 + , with ~u_(y), ~/+ (y) > 0.
Let ~b(x, y) be the solution of (2.6); in the sequel we need the following
estimates, proved in [5]. There exists an e > 0 such that

II~ <x, ) - e~-XvL[] c~(~o) <= Ce (z_+~)~,

(4.7.a)
[lCx(X,. ) -4-e~-~-Ilc~ ~ Ce (~-§
for x = O , and

II 1 - q S ( x , . ) - e-~+X~+llcl,~(~) ~ Ce -(~++~>x,
(4.7.b)

for x ~ O .

5. Proof of Theorem 1

Theorem t is proved through a careful study of the asymptotic behaviour,

as [x] ~ + 0% of possible solutions (2, u, v) of (2.9) such that Re 4 __<0, and
through the application of a maximum principle argument. Before we concen-
trate on problem (2.9), we want to prove a few additional lemmas which will
be useful in the sequel. Let us therefore consider a H61der-continuous, com-
plex-valued function a(y), with Re a(y) >__0 in ~ , a H61der-continuous, real-
valued function y(y), and a continuous, complex-valued function f ( x , y) satis-
fying
I f ( x , Y)I < C e - ~ for all (x, y) e 27+
 Travelling Fronts in Flame Propagation 107

for some /t > 0. We are interested in the solutions of the problem

(5.1.a) - A u + ?(y) Ux + a(y) u = f ( x , y) in 27+,

Ou
(5.1.b) -- = 0 in R+x0co.
Ov
Let e ~x ~ ( x , y) be an exponential solution of the homogeneous problem cor-
responding to (5.1).

L e m m a 5.1. Under the above assumptions, )~ has nonzero real part unless a =-O.
If a = 0 and if Re 2 = O, then )~ = 0; if furthermore (y) t O, then ~ =-- constant.

Proof. Assume that 2 has the form 2 = iq, and write ~ in the form

~ ( x , y) = ~ q/k(y)x k.
k=l

The function ~un does not identically vanish in co; moreover, q/n and ~Un_l are
solutions of

(5.2.a) ( - A y + iqy(y) + a(y) + q2) q/~ = 0 in co,

(5.2.b) ( - A y + iqy(y) + a(y) + q2) q/n-1 + (?(Y) - 2iq) nq/n = 0 in co,

0q/, 0~u~_1
(5.2.c) - - 0 on 0o2.
Ov Ov
Multiplying (5.2.a) by ~ and integrating the result over co, we get

~lV~,l=dy + ~(Rea(y) +q=)l~unl=dy=O.

(.o (2)

Since Re a is nonnegative, it follows that ~u, ~ constant and q = 0. Moreover,

q/n is a solution o f (5.2.a) only if a vanishes in co. I f we carry out the same
operation for (5.2.b) (remembering that ~, :r 0), we get ( y ) n = 0 and thus
n = 0 provided that ( y ) . 0. []

We now assume that a(y) does not identically vanish in co, and investigate
the asymptotic behaviour of the solutions of (5.1).

2,pc (27+) n W 1, o0 (27+) be a solution

L e m m a 5.2. For some p > N, let u (x, y) ~ Wlo
of (5.1). Moreover, assume that u is bounded in 27+. Then there exists an tl > 0
such that
lu( x, Y)I ~ Ce-eX for all (x, y) ~X+.

Proof. By L e m m a 5.1, we know that there exists p > 0 such that, for any ex-
ponential solution of the homogeneous problem, the corresponding exponent
lies outside the interval [ - p , p]. This comes from the fact that there is only
a finite number of exponents corresponding to nonzero exponential solutions
in every strip {Re z < K}.
108 H. BERESTYCKI, B. LARROUTUROU& J.-M. ROQOEJOFFRE

Select a real number r such that 0 < r < p. Let F(x) b e a nondecreasing
C~-function such that

F ( x ) = I01 ifx__< 1,
if x__>2.
Set
v(x, y) = e-rXF(x) u(x, y).
We have
A ( r ( x ) u) = r " u + 2r'ux - r ( x ) ( y ( y ) Ux + a(y) u).
Choose p > 0 and set

h(x' Y) = e-rX [ A (l'u) - 2r O(I'u~) + r217u - pu]

Then h ~ L 2 (27+) ~ L = (2;+) ; moreover, v satisfies

-Av+luv=h in Z+,
Ov
-- = 0 in R+x0co,
Ov
v--0 in [0, 1]•

It is possible to apply Nirenberg's method of translations (see [9], Chap. IX)

to the function v; we deduce that v (H2(Z+). We apply Theorem 4.2 to the
following equation, which is satisfied by v on [2, + co[ • co:
- A v + ( y ( y ) + 2r) vx + (a(y) + ry(y) - r 2) v = e-rXf(x, y),

with Neumann boundary conditions. As a consequence, v can be written in

the form
V = V0 + V , ,
with
Iv.(x, y)l -< Ce -(a+r-~)x in Z+,
for all e > 0, where v ~ is a bounded solution of the homogeneous problem.
Since the function u is bounded as x ~ + co, the function u ~ given by

u~ y) = er%~ y)
belongs to 2 ( - r , p, 2;+ ) ; moreover, u ~ is a solution of the homogeneous prob-
lem corresponding to (5.1). By the choice of r, we infer from Theorem 4.1 and
from the fact that p > N that
liu~ .)[[cl.~(~) < Ce -*x for all (x, y) ( X + ,
the real number ( - e ) being the (negative) real part of the first exponent of
the asymptotic expansion of u ~ Therefore,
lu( x, Y)I < Ce-"X for all (x, y) EZ+
for any 1/< rain (p, cx). Lemma 5.2 is proved. []
 Travelling Fronts in Flame Propagation 109

Now let (c, 4)) be the solution of (2.6.a-d). Since 4)~ > 0 in 2 , we have
g(4)(x,-)) =0 for x = 0 .
We now concentrate on the solutions of
(5.3.a) -~u + fl(y) Ux - g'(4)) u = 2u for (x, y) ~27,
Ou
(5.3.b) -- = 0 for (x, y) ~ • x 0o9,
Ov
(5.3.c) u E W~6P(27) n W 1'= (27) for all p > 1.

Let (u,)~) be a solution of (5.3) such that Re )~ < - g ' ( t ) . Let (ek)k>__l be the
sequence of the exponential solutions of
(5.4.a) - A u + fl(y) Ux - g'(1) u = 2u in 27+,
Ou
(5.4.b) -- = 0 in [P+x0co,
Ov
written in the form ek(x, y) = eZ~*pk(x, y ) , such that Re 2k < 0 and such that
the sequence (Re ,~k) is nonincreasing. Let ( - a l ) = max~__>t Re )~k; then a~ > 0.
Let (ek)a<_k<_n be the exponential solutions of (5.4) such that Re ,~k = - a v
Let - 2 + be the negative generalized principal eigenvalue of (5.4) with ,~ = 0 ;
this exponent is involved in estimate (4.7.b) for the exponential decay of
( 1 - 4)) and 4)x. From Lemma 5.1 we deduce:

Lemma 5.3. (i) / f 0 1 > ~ +, there exists an e > 0 such that

Ilu(x, y)llc~,~([x,~+l]x~o) = O(e -(z++~)x) f o r all x > O.

(ii) I f (71 ~ ,~+, there exist an e > 0 and n complex numbers at . . . . , an such that

U(X, y) -- ~ akek(X , y) ---- O(e -(Crl+e)x) for all x > O.

k=l C~'~([x,x+l]•

Proof. In either case the function u satisfies

- A u + f l ( y ) ux - (g'(1) + 2) u = (g'(4)) - g'(1)) u in X+

and condition (5.2.b). Since the function g is a Ca'~-function in a neighbour-
hood of 1, estimate (4.7.b) implies that
[g'(4)(x, y)) - g'(l)l __<Ce-'~z+ x

for x large enough. Notice that without loss of generality we may assume that
5 has the form
1
5= --, PEN.
P
Since Re 2 < - g ' ( 1 ) , we are in the framework of Lemma 5.2. Therefore there
exists an ~/> 0 such that (5.3) holds for u; consequently,
I [g'(4)(x, y)) - g'(1)] u(x, Y)I --< Ce-(~z++n)x in 2 + .
110 H. BERESTYCKI,B. LARROUTUROU& J.-M. ROQUEJOFFRE

Case 1:o-1 > 2+. Notice that P can be chosen so large that
P+I
- - 2+<0-1.
P
As a consequence o f Theorem 4.2 the function u may be written in the form
u = u ~ u,. The function u, is a solution of (5.3.a), (5.3.b) in 2;+ and
satisfies
(5.5.a) II U, (X, " )[[Cl,~([x,x+llx~o) ~= Cee-(~i+ +e)x
for small enough positive e and large enough positive x. The function u ~ is
a solution of (5.4); moreover, from Lemma 5.2 and (5.5.a) we have u ~
2 ( p , p, 27+) for all p > 1 and for some p > 0. We may then apply Theo-
rem 4.1 to u~ There exist n complex numbers: al . . . . . an such that, for
small enough e > 0,
(5.5.b)

u~ y) - ~ akek(x, y) = O(e -(~x+e)x) for all x > 0.

k=l
cl,~([x,x+l]x~)

Since o-1 > 2+, U satisfies (5.5.a). Therefore,

[ [g'(~(x, y)) - g'(1)] u(x, y)[ < Ce-(aaz++"lx in Z+
for x large enough. Now, apply the above procedure P times, with a suitable
choice of e at each iteration. By the choice of P, and after an additional ap-
plication of Theorem 4.2, this yields
11u(x, Y)llc~.~(tx,x+llx~) = O(e -(x++elx) for all x > 0.
Case 2: o 1 N 2+. The proofs is quite analogous to the proof of the preceding
case; one simply has to exchange the roles of 2+ and al in order complete
the proof. The details are left to the reader. []

Here comes the crucial part of the proof of Theorem 1. We first recall
some properties of the function Pl(P) given by

P ~ Px (P) = ,111( - - A y -]- p3(y)) ;

they are proved in [8]. Here we follow the notation of [8]: For any elliptic
operator L in co, with Neumann boundary conditions, we denote by /z1(L)
the first eigenvalue of L. Recall that if L has the form
L = --Ay + a(y),
the function a(y) being real-valued, we have the following variational charac-
terisation for /~ (L) :
(] VyU[ 2 + a(y) u 2) dy
69
(5.6) Pl (L) = min
"~r/i(~~ ~ u2 dy
co
 Travelling Fronts in Flame Propagation 111

From (5.6) one can prove

L e m m a 5.4. ([8]). The function Pl (P ) is Lipschitz-continuous. Moreover, the func-

Pl (P ) 9
tion p ~ - - ts nonincreasing and satisfies
P

(5.7) l i m p~ (p) - (fl).

p-*0 p

This l e m m a is, by the way, the key to the p r o o f of Theorem 4.3.

Let (u, 2) be a solution of (5.3). The following l e m m a is very important.

L e m m a 5.5. Assume that Re 2 < 0 and )~ t- O. Then there exists an e > 0 such
that, for all (x, y) f X _ , the following inequality holds:

lu(x, Y)I ~ Ce<X-+~)x for all (x, y) fiX_.

Proof. Since 4~ < 0 on 2;_ by (estimate (2.7.b)), u is a solution of

- A u + fl(y) Ux = 2u in 2J_,

au
-- = 0 in [R_xato.
Ov

By (5.3.c) we have that u f 2 ( - e , p, X_) for all e > 0 and for all p > 1; we
may then apply Theorem 4.1: u admits an asymptotic expansion in terms of
exponential solutions in the sense of Theorem 4.1. Let e ~ ( x , y) be one of
the terms of the asymptotic expansion o f u. Setting

B=p+iq,

we have p > 0 by L e m m a 5.1. Let us write

~ ( x , y) = ~ ~k(y)xk;
k=0

the function qJ~ is a solution of

(5.8.a) --Ayl[/n + (ltfl(y) - 2) ~u~ = / t 2 ~ in o9,

(5.8.b) 0~,~ _ 0 on 0o9.

Ov

Moreover ~,, does not identically vanish in w.

112 H. BERESTYCKI, B. LARROUTUROU& J.-M. ROQUEJOFFRE

We multiply (5.8.a) by ~ and integrate the resulting expression over co;

identifying the real parts of each side of the new equation yields
[] Vy~] 2 + (pfl(y) _ Re 2)] ~/~t2] dy = (p2 _ q2) j I ~'nl 2dy 9
co (2)

From the variational characterisation of Pt ( - / ~ y + p/3(y) - Re 2) of (5.6) we

derive the inequality

I11 (--Ay "[- p f l ( y ) -- Re 2) __<p2 _ q2.

We have now to study two cases.

Case 1: Re 2 < 0. We get

(5.9) ~1 (--Ay + p f l ( y ) ) < p2.

Recall (see [8]) that 2_ is obtained as the unique positive real number p satis-
fying the equation Pl(P) = p a . This follows from two facts. First, according
to Lemma 5.4, the function

p~--~ Pl(P) /-'/I ( -- ~y + P f i ( Y ) )

P P
is nonincreasing, whereas p ~-,p is increasing; second, (5.7) implies that
Pa (P) > p2 in a neighbourhood of 0 (see Fig. 5.1). This is sufficient to ensure
the existence and uniqueness of 2_. Now, since 2_ is unique, we see that
(5.10) cannot hold for p _ 2_. Therefore p > 2_.

2_
Fig. 5.1

Case 2: Re 2 = 0. Notice that in this case we cannot have q = 0. Let us indeed

denote by Im 2 the imaginary part of 2 and rewrite the equation for gtn

--&q/n + (Pfl(Y) -- i I m 2) ~un =p2q/n in co

with Neumann boundary conditions. Multiplying this equation by ~n, in-
tegrating the new equation over co, and identifying the imaginary parts yields

I m 2 ]" I g/n[2dy = O;
fo
 Travelling Fronts in Flame Propagation 113

we then have 2 = 0, which is, by assumption, impossible. Therefore q ~: 0. We

get l~1 ( - @ + Pfl(Y)) < p2, and as before we conclude that p > 2_.
Recall now that there is only a finite number of exponential solutions with
exponents p in the strip {2_ < R e p < ~_ + 1]; an additional application of
Theorem 4.1 yields Lemma 5.5. []

The following lemma is symmetrical to Lemma 5.5.

l ~ m m a 5.6. Under the assumptions of Lemma 5.5 there exists an e > 0 such that

lu(x, y)] =< Ce -(z++e)x for all (x, y) EX+.

Proof. First study the exponential solutions in 2 ( 0 , p, 27+) of the

homogeneous equation (5.4). We denote one of them by

e ~ ~ d~k(Y)x k.
k=O
Setting p = p + iq we find, by an easy adaptation of the beginning of the
proof of Lemma 5.5, that

/tl(--Ay +pfl(y) - (g'(1) + Re 2)) < p 2 _ q2.

This time we look for p < 0. By a discussion analogous to the discussion

in the preceding lemma, we get p < ( - 2 + ) . Reverting to the notations of
Lemma 5.3, let (e~kXpk(x, Y))k>__~ be the sequence o f the exponential solutions
of (5.4) such that Re 2k < 0 and such that the sequence (Re 2k) is nonin-
creasing. Let (-o-1) = maxk___l Re 2k; we have as > 2+, and we are thus in
case (i) of Lemma 5.3. This implies Lemma 5.6. []

Lemmas 5.5 and 5.6 enable us to prove:

Proposition 5.7. Let (2, u) be a solution of (5.3). Then either Re 2 > 0 or 2 = O.

We could partially infer Proposition 5.7 from Lemmas 5.5 and 5.6 and
from the proof given in [17] (Chap. 2, Sect. 8 "eigenvalues"); we prefer,
however, to use a more straightforward method.

Proof. Assume that there exists a solution (2, u) o f (5.3) with Re 2 =< 0 and
2 ~: 0. Let us write u = Ul + iu2; we always may suppose that Ul * 0. Indeed,
if we had ul ~ 0, we would just have to work with the function iu instead.
From Lemmas 5.5 and 5.6 we infer the existence of two real numbers k_
and k+, with k_ < 0 _< k+, such that

(5.10.a) k-qSx - ul <k+q~x in z~.

114 H. BERESTYCKI, B. LARROUTUROU ~; J.-M. ROQUEJOFFRE

Since /~1 =I=0, the real numbers k_ and k+ cannot be both zero; assume
therefore that k+ . 0. There exists (x+, y+) ~ such that

(5.10.b) k+Ox(X+, y+) = ul(x+, y+).

Let v(t, x, y) = Re[e-Ztu(x, y)]; the function v is the unique solution of the
parabolic problem

(5.11.a) v t - &v + p ( y ) vx - g'(qS) v = 0 for (x, y) E2;,

~V
(5.11.b) -- = 0 for (x, y) E N x a c o ,
av
(5.11.c) v(t, -co, y) = v(t, + 0% y) = 0 for yfi c~

with initial data

(5.11.d) v(0, x, y) = Ul(X, y).

We set 2 = 2 1 "[-i~2; we have an explicit expression for v(t, .):

v(t, x, y) = e-~lt[ul(x, y) cosX2t + 89 y) sin22t].

Once again we have to study two cases.

Case l: 21 < 0. The function q~x is a steady solution of (5.11.a-c); from
(5.10.a) and from an easy adaptation of the proof of the Weak Maximum
Principle (see, for instance, [10]), we infer that

k-r <--v(t, x, y) <=k+~x in 2 .

This is in contradiction with the explicit expression for the function v, which
shows that v ( t , . ) is unbounded as t ~ + c~.
Case 2 : 2 1 = 0. We now have 22 * 0 and

v(t, x, y) = us(x, y) cos22t + uz(x, y) sin22t.

This time nothing a priori contradicts the Weak Maximum Principle, but for
2~
/'1 = - - w e have
)~2 k+4~x(X+, y+) = V(tl, x+, y+).
If (x+, y+) E 27, then, by the Strong Maximum Principle, v -k+4~x---= 0, a con-
tradiction. Therefore (x+, y + ) 6 027; the H o p f Boundary Lemma for parabolic
equations (see, for example, [17]) yields
O(v - k+~x)
(tl, x+, y+) > 0,
Ov
which is impossible; Proposition 5.7 is proved. []

Proof of Theorem 1. Assume that there exists a solution (2, u, v) of (2.9) with
Re 2 ~ 0 and 2 . 0 . Let us set W = u + v; the function W is a solution of
 Travelling Fronts in Flame Propagation 115

the following linear elliptic boundary value problem:

- A W + f l ( y ) Wx=AW in 27,
aW
-- = 0 in IRxOog.
Ov
Moreover, since W is bounded we have

w~ ~ ( e , p, 2+) c~ 2 ( - ~ , p, 27)
for every e > 0 and for all p > l . Here we denote by ~z~(e,p, 2;+) and
2 ( - e , p, 2;_) the spaces of continuous functions defined in Z and satisfying
the definitions given at the beginning of Section 4. From Lemma 5.1 and The-
orem 4.1 applied to W in Z+ as well as in Z_, we infer that W and V W decay
exponentially when Ix I --* + 0% in particular, we have
WE H l ( z ~ ) ~ cl'O(z~).

We may then multiply the equation for W by W and integrate it on Z;

by taking the real parts on both sides of the equations we get

~ l V Wl 2 • dy - Re ,~ ~ l Wl 2 dx dy = O.
Z 2;

Since Re 2 _< 0, by assumption, we obtain

I l V w l 2 ~ dy =O,
Z

which implies that W-- 0, since W vanishes as Ixl ~ + ~. Therefore (2, u) is

a solution of problem (5.3), which is, by Proposition 5.7, impossible. This ends
the proof of Theorem 1. []

As a conclusion to this paper, let us point out that the sign of the function
c + a ( y ) does not interfere with the linear stability of the travelling waves. It
is proved in [6] that if the functions ot and g satisfy the technical condition

(5.12) max a ( y ) - min c~(y o0 - min o~(y < 2 ~ g ( u ) du,

yEff~ yE6~ yEff~ 0

then the solution (c, ~b) of (2.6) satisfies

c+~(y)>0 in ~ .
On the other hand, if (5.12) is not satisfied, the function c + a ( y ) may change
sign. This phenomenon, called "inversion of the velocity field," is discussed
in [6] (see also [22] for a similar phenomenon in other geometrical configura-
tions, such as for a counterflow diffusion flame).
One could have expected a loss of stability for the travelling-wave solutions
of (2.1), due to a possible inversion of the velocity field. Our results show that
they are always linearly stable, whether this phenomenon occurs or not. In
fact, the positivity of c + (o$ is the only fact that really matters.
116 H. BERESTYCKI,B. LARROUTUROU& J.-M. ROQUE/OFFRE

References

1. S. AGMON, A. DOUGLIS,& L. NIRENBERG,Estimates near the boundary for solutions

of elliptic partial differential equations satisfying general boundary conditions L II,
Comm. Pure Appl. Math., 12 (1959), 623-727; 17 (1964), 35-92.
2. S. AGMON& L. NIREN~Eg~, Properties of solutions of ordinary differential equations
in Banach space, Comm. Pure Appl. Math., 16 (1963), 121-239.
3. D. G. ARONSON& H. F. WEINBERGER, Nonlinear diffusion in population genetics,
combustion and nerve propagation, Partial differential equations and related topics,
Lect. Notes in Math., 446, Springer-Verlag, New York (1975), 5-49.
4. D. G. ARONSON& H. V. WEINBERGER, Multidimensional diffusion arising in popula-
tion genetics, Adv. Math., 30 (1978), 33-58.
5. H. BERESTYCKI& n. LARROUTUROU, Quelques aspects mathdmatiques de la propaga-
tion des flammes prdmdlangdes, Nonlinear Partial Differential Equations and their
Applications, Coll~ge de France Seminar, 10, H. BREZlS& J. L. LIONS, eds., Pitman,
Longman, Harlow (U.K), 1990.
6. H. BERESTYCKI, B. LARROUTUROU,& P. L. LIONS, Multidimensional travelling wave
solutions of a flame propagation model, Arch. Rational Mech. Anal., 111 (1990),
33 -49.
7. H. BERESTu B. NICOLAENKO& B. SCI-~URER, Travelling wave solutions to com-
bustion models and their singular limits, SIAM J. Math. Anal., 16 (1985),
1207-1242.
8. H. BERESTYEKI& L. NIRENBERG, Some qualitative properties of solutions of
semilinear equations in cylindrical domains, Analysis et Cetera (dedicated to
J. Moser), P. H. RAmNOWITZ& E. Z~HNDn~, eds., Academic Press, New York, 1990,
115-164.
9. H. BREZIS, Analyse fonctionnelle: thdorie et applications, Masson, Paris, 1983.
10. J.D. BUCKMASTER& G. S. S. LUDFORD, Theory of laminar flames, Cambridge Univ.
Press, Cambridge, t982.
11. E.A. CODDmGaON& N. LEVINSON, Theory of ordinary differential equations,
McGraw-Hill, New York, 1955.
12. P.C. Fn~ & J. B. McLEoo, The approach of solutions of nonlinear diffusion equa-
tions by travelling front solutions, Arch. Rational Mech. Anal., 65 (1977) 335-361.
13. P.S. HAGAN, Travelling wave and multiple travelling wave solutions of parabolic
equations, SIAM J. Math. Anal., 13 (1982), 717-738.
14. A.N. KOLMOGOROV, I. G. PETROVSKn,& N. S. PISKUNOV,A study of the equation of
diffusion with increase in the quantity of matter and its application to a biological
problem, Bjul. Moskovskogo Gos. Univ., 17 (1937), 1-26.
15. B. LARROUTUROU, The equations of one-dimensional unsteady flame propagation:
existence and uniqueness, SIAM J. Math. Anal., 19 (1988), 32-59.
16. A. PAZ% Asymptotic expansions of solutions of ordinary differential equations in
Hilbert space, Arch. Rational Mech. Anal., 24 (1967), 193-218.
17. M. PROTTER& H. F. WEINBERGER, Maximum principles in differential equations,
Prentice Hall, Englewood Cliffs, 1967.
18. J.M. ROQUEIOFrRE, Stability of travelling fronts in a model for flame propagation,
PartlI." Nonlinear stability, Arch. Rational Mech. Anal., 117 (1992), 119-153.
19. J.M. RoQurJorrgE, L. SAINSAtrLmU,& D. "lEgMAN, On the stability of steady planar
premixed flames, in preparation.
20. D.H. SATTnqGER, Stability of waves of nonlinear parabolic equations, Adv. Math.,
22 (1976), 312-355.
 Travelling Fronts in Flame Propagation 117

21. G. I. SIVASI-IINSKY, Instabilities, pattern formation and turbulence in flames, in

"Mod61isation des ph6nombnes de combustion", Direction des Etudes et Recher-
ches d'Electricit6 de France, Eyrolles, Paris, 1985.
22. E A. WrLLL~MS, Combustion theory, second edition, Benjamin Cummings, Menlo
Park, 1985.

Laboratoire d'Analyse Numdrique

Universit6 de Paris VI
4 Place Jussieu
75232 Paris Cedex 05
CERMICS, INRIA
Sophia Antipolis, 06565 Valbonne
France
and
Ecole Polytechnique
Centre de Math6matiques Appliqu6es
91128 Palaiseau Cedex
France

(Received March 3, 1991)