J.

Differential Equations 255 (2013) 1657–1683

Contents lists available at SciVerse ScienceDirect

Journal of Differential Equations

www.elsevier.com/locate/jde

Lyapunov function and spectrum comparison for

a reaction–diffusion system with mass conservation
Shuichi Jimbo a , Yoshihisa Morita b,∗
a
Department of Mathematics, Hokkaido University, Sapporo 060-0810, Japan
b
Department of Applied Mathematics and Informatics, Ryukoku University, Seta Otsu 520-2194, Japan

a r t i c l e i n f o a b s t r a c t

Article history: We are dealing with a two-component system of reaction–diffusion

Received 17 September 2012 equations with mass conservation in a bounded domain with the
Revised 24 February 2013 Neumann boundary conditions. We prove the global boundedness
Available online 12 June 2013
of the solution in L ∞ -norm for t  0 under a condition, and
MSC:
then the existence of a Lyapunov function. Moreover, by studying
35B32 the linearized eigenvalue problem of a nonconstant equilibrium
35B35 solution, we provide a comparison theorem for the spectrum
35B40 between the linearized operators of the system and an appropriate
35K57 nonlocal scalar equation. As an application of the comparison result
we obtain that any stable equilibrium solution must be monotone
Keywords: if the space dimension is one. It is also shown that a modified
Reaction–diffusion system system with a new parameter, which covers the present model,
Mass conservation possesses a Lyapunov function.
Equilibrium solution
© 2013 Elsevier Inc. All rights reserved.
Spectrum comparison
Linearized eigenvalue problem

1. Introduction

We consider the following system of reaction–diffusion equations:


ut = du − g (u + v ) + kv ,
(x ∈ Ω) (1.1)
v t =  v + g (u + v ) − kv

with the Neumann boundary conditions

* Corresponding author.
E-mail addresses: jimbo@math.sci.hokudai.ac.jp (S. Jimbo), morita@rins.ryukoku.ac.jp (Y. Morita).

0022-0396/$ – see front matter © 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.jde.2013.05.021
1658 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

∂u ∂u
= = 0 (x ∈ ∂Ω), (1.2)
∂ν ∂ν

where d, k are positive constants, g (u , v ) is sufficiently smooth function and Ω is a bounded domain
in Rn with the smooth boundary ∂Ω .
We may assume

g (s)  
g (s)/s  K (s > 0), lim = β > 0, g ∈ C 1 [0, ∞); R , (1.3)
s→+0 s

which ensure the nonnegativity of the solution of the initial value problem provided u (x, 0),
v (x, 0)  0 in Ω by the maximum principle applied to each equation of the system.
Since the total mass

 
u (x, t ) + v (x, t ) dx
Ω

is conserved in time evolution for the solution (u (x, t ), v (x, t )) of (1.1), we fix the condition as

 
s := u (·, 0) + v (·, 0) , (1.4)

where we use the notation


 1
f (·) := f (x) dx, |Ω|: the volume of Ω.
|Ω|
Ω

A specific example of g is given by

kw
g(w) =
(aw + 1)2

in [14] and [20] as a simple model exhibiting similar dynamical behaviors to those observed in a
more complicated model describing the cell polarity. In fact, for such a function g they show the
Turing type instability for appropriate parameter values d, k and s. They also exhibit by a numerical
simulation (under the periodic boundary conditions) that after an emergence of a wave pattern and
a long transient behavior, the solution converges to a stationary solution with the shape of a single
spike. By virtue of this concentration property their model equations could be helpful for the under-
standing of the phenomenon of cell polarity. However, they don’t show why such a spike pattern is
stable while the multi-spike pattern is unstable.
The purpose of this article is to reveal some dynamical property and the stability of equilibrium
solution of (1.1)–(1.2), which justifies the numerical observation about the asymptotic profile of the
solution. More precisely, we show that the system of (1.1) allows a Lyapunov function, by which we
can assert that the ω -limit set of any bounded orbit of the solution consists of equilibrium solutions.
In addition, by investigating the linearized eigenvalue problem for any equilibrium solution, we see
that the profile of a stable equilibrium solution in one-dimensional space must be monotone (or
unimodal, i.e., of a single spike if the periodic boundary conditions are assumed).
We remark that for f (u ) instead of g (u + v ) the similar dynamical property is also found in
the same literatures [14,20]. For some mathematical result justifying it is proposed in [19,18] by
converting the equations to the equations of the phase field type system as in [4] and [9]. The present
article is thereby a successive work for a class of conserved reaction–diffusion systems.
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1659

We also remark on some results for the general system

ut = du xx − f (u , v ), v t = v xx + f (u , v ),

which arises as a simple chemical conserved reaction model, called a closed reaction–diffusion system.
The existence of traveling waves for a specific nonlinear term is shown in [5] and [8]. Moreover, this
type of the system is used in modeling of a precipitation kinetics, and the formation of Liesegang
patterns is shown in [13] and [23]. As seen in [21], by putting w = u + v , v = v, the above system
can be converted to the one belonging to a class of the systems

w t = a (u , v ) w x + b (u , v ) v x x , v t = v xx + f̃ (u , w ).

Under some conditions on a(u , v ) and b(u , v ) the instability of a spike solution of the system in the
infinite interval or a sufficiently large interval with the Neumann boundary conditions is investigated
in [21] (see also [22] on the stability of layer solutions).
Coming back to the present system, we state our results precisely. We first provide the global
boundedness of the solution for t  0 under the condition (1.3).

Proposition 1.1. Assume (1.3) and that u 0 , v 0 ∈ C 0 (Ω, R), u 0 (x), v 0 (x)  0. Let (u (x, t ), v (x, t )) be a solu-
tion of (1.1) with (1.2) and (u (x, 0), v (x, 0)) = (u 0 (x), v 0 (x)). Then u (·, t ) L ∞ ,  v (·, t ) L ∞ ,
∂ u (·, t )/∂ x j  L ∞ , ∂ v (·, t )/∂ x j  L ∞ ( j = 1, 2, . . .) are uniformly bounded in t  0. Moreover, there is a posi-
tive M > 0, independent of the initial data, such that

n 
 
∂u ∂v
lim sup u (·, t ) L∞
+ v (·, t ) L∞
+ (·, t ) + (·, t )  Ms
t →∞ ∂xj L∞ ∂xj L∞
j =1

holds.

Next we show the system allows a Lyapunov function. Define the functional

  
d   
E (u , v ) := ∇(u + v )2 + 1 ∇(du + v )2 + Q (u + v ) dx, (1.5)
2 2
Ω
w
dk
Q ( w ) := (1 − d) g ( z) dz + w 2.
2

Then we have

Proposition 1.2. For a smooth solution (u (·, t ), v (·, t )) (0  t < ∞) to (1.1) with (1.2) (or the periodic bound-
ary conditions)

d    2
E u (·, t ), v (·, t ) = − (1 + d)(ut + v t )2 + k∇(du + v ) dx  0 (1.6)
dt
Ω

holds. Then dE (u (·, t ), v (·, t ))/dt = 0 (∀t ∈ R) implies that (u (x, t ), v (x, t )) is an equilibrium solution.

The proof of Propositions 1.1 and 1.2 will be given in Section 2.

We remark that if (1.3) is satisfied, then by the propositions, the orbit of the solution in an ap-
propriate phase space is precompact and the ω -limit set of the orbit consists of equilibrium solutions
1660 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

(see for instance [11] or [12]). In other words, the asymptotic state of any solution to (1.1) with (1.2)
and (1.3) is determined by the stationary problem.
Before going to the stationary problem, we covert the system by putting w = u + v as


w t = d w + (1 − d) v ,
(x ∈ Ω) (1.7)
v t =  v + g ( w ) − kv

with the Neumann boundary conditions and w (·, t ) = s for t  0.

We decompose the variables under the condition d = 1 as


w = w + W = s+ W,
(1.8)
v = v + V =ξ + V.

Then (1.7) is decomposed as

W t = d W + (1 − d) V , (1.9)

V t =  V + g (s + W ) − g (s + W ) − kV , (1.10)

ξt := g (s + W ) − kξ (1.11)

with

 
V (·, t ) = W (·, t ) = 0, (1.12)

namely the solutions ( w (x, t ), v (x, t )) to (1.7) and ( W (x, t ), V (x, t ), ξ(t )) to (1.10)–(1.11) have the 1–1
correspondence by

    
w (x, t ), v (x, t ) = s + W (x, t ), ξ(t ) + V (x, t ) , ξ(t ) = v (·, t ) ,

consequently, the solution (u (x, t ), v (x, t )) of (1.1) with (1.2) and (1.4) is given by

   
u (x, t ), v (x, t ) = s − ξ(t ) + W (x, t ) − V (x, t ), ξ(t ) + V (x, t ) .

We notice that the equations for W and V of (1.9) and (1.10) are independent of ξ , therefore it
suffices to consider (1.9) and (1.10) with the Neumann conditions and (1.12) unless d = 1. If d = 1,
then the equations turn to be quite simple, we thereby exclude this case in the present study.
Now consider the stationary problem, that is,


d W + (1 − d) V = 0,
 (x ∈ Ω), (1.13)
 V + g (s + W ) − g (s + W ) − kV = 0

with

∂W ∂V
= = 0 (x ∈ Ω), W = V = 0. (1.14)
∂ν ∂ν

By the first equation of (1.13)

dW + (1 − d) V = 0
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1661

follows. Hence, we see that (1.13) with (1.14) is equivalent to the next scalar equation

  
 W − α g (s + W ) − g (s + W ) − kW = 0 (x ∈ Ω), (1.15)
∂W
= 0 (x ∈ ∂Ω), W = 0, (1.16)
∂ν

where we put

1−d
α := (d = 1).
d

In fact if W ∗ is a solution of (1.15) with (1.16), then ( W ∗ , V ∗ ) = ( W ∗ , − W ∗ /α ) gives a solution to

(1.13) with the Neumann boundary conditions, thus an equilibrium solution of (1.1) with (1.2) and
(1.4) is obtained by

 
  1   1 1   d
u (x), v (x) = s − g s + W ∗ +
∗ ∗
W (x), g s + W ∗ −
∗
W ∗ (x) .
k 1−d k 1−d

The next result is on the spectrum comparison between the linearized operator at an equilibrium
solution in the system and that of the above scalar equation (1.15)–(1.16). The idea of the comparison
is due to [3] for the Cahn–Hilliard equation and the phase field system. Later, some improvement for
the phase field system is proposed in [18]. We will develop the study to our system. To state the
result, we introduce function spaces. Denote by L 2 (Ω) the space of square integrable functions with
the norm u  (u ∈ L 2 (Ω)) and

   1/2 
2
H m (Ω) = u ∈ L 2 (Ω): u  H m := u 2 + ∂ a u /∂ xa <∞ ,
|a|m


n
a = (a1 , . . . , an ), a j  0, a j ∈ Z, |a| = a j.
j =1

We also introduce a subspace of L 2 (Ω) and H m (Ω) with the average zero as

  
2 2
 
L (Ω) := u ∈ L (Ω): u dx = 0 , H m (Ω) := u ∈ H m (Ω): u ∈ L 2 (Ω) .
Ω

We let W ∗ be a solution of (1.15) with (1.16) and put w ∗ = s + W ∗ . We fix s and consider the
linearized operator A for (1.7)

   
φ dφ + (1 − d)ψ
A := − , (1.17)
ψ ψ + g ( w ∗ )φ − kψ

and the linearized operator L at W ∗ for (1.15)

      
L(φ) := −φ + α g w ∗ φ − g w ∗ φ + kφ, (1.18)

with the domain

1662 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

 
D(A) = (φ, ψ) ∈ H 2 (Ω) × H 2 (Ω): ∂φ/∂ ν = ∂ψ/∂ ν = 0 (x ∈ ∂Ω) ,

and

 
D(L) = φ ∈ H 2 (Ω): ∂φ/∂ ν = 0 (x ∈ ∂Ω) ,

respectively.
We compare the spectrum of A and L. It is easy to see that any eigenvalue is real for L. We let
{μ j } j =1,...,∞ be the set of all the eigenvalues of L arranged in increasing order with counting the
multiplicity as

μ1  μ2  · · ·  μm  · · · . (1.19)

On the other hand there is a possibility A allows a complex eigenvalue. Fortunately, any eigenvalue λ
of A is real if the real part of λ is less that k/2 (see Section 4), in the sequel it is possible to compare
all the real eigenvalues which are related to the stability of the solutions. We thereby let {λ j } j =1,..., N
be the set of all the eigenvalues of A less than k/2, arranged in increasing order with counting the
multiplicity as

λ1  λ2  · · ·  λ N < k/2. (1.20)

The next theorem gives the spectrum comparison between the linearized operators.

Theorem 1.3. The number of the negative eigenvalues of A and L coincides, and so the multiplicity of the zero
eigenvalue of both the operators. Moreover,

d
|λ j |  |μ j | ( j = 1 , 2 , . . . , N ) (1.21)
1+d

holds, where μ j and λ j are as in (1.19) and (1.20) respectively.

We give a remark that (1.15) with (1.16) is the Euler–Lagrange equation of

  
d
G ( w ) := |∇ w |2 + Q ( w ) dx, (1.22)
2
Ω

in the space H 1 (Ω), where Q is as in (1.5). By virtue of Theorem 1.3, noticing (L(φ), φ) L 2 is the
second variation of G ( w ), we can apply the stability result for (1.22) in [10] to the system (1.1) to
obtain the following result (see also [24], and as for the periodic boundary conditions see [18]):

Corollary 1.4. Consider the one-dimensional case, namely, the case that Ω is interval. Let (u (x), v (x)) be a
stable equilibrium solution to (1.1) with (1.2) and (1.4). Then u (x) and v (x) are monotone, namely, those are
constant or strictly monotone. If the boundary conditions are periodic, then u (x) and v (x) are constant or
unimodal.

We use the terminology “stable” of the above corollary in the sense of Lyapunov (see [12]). We
note that the linearized operator always possesses zero eigenvalue.
We state the idea of the proof of Theorem 1.3 in brief. Noticing that the equations of the eigen-
value problem of A can be written in a single equation for φ , we introduce a nonlocally weighted
eigenvalue problem for L with a free parameter β  0 and establish the comparison between the
eigenvalues of the nonlocal problem and the original L, which is the case β = 0 up to multiplication
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1663

of the constant d/(1 + d). By the continuity property of the spectra in β each eigenvalue of A less
than k/2 is given by choosing an appropriate β > 0. By this we obtain the desired assertion.
In the next section we prove Propositions 1.1 and 1.2. Then in Section 3 we discuss the nonlocally
weighted eigenvalue problem and prove Theorem 1.3 by the key lemmas, which will be proved in
Section 4.
In Section 5 we consider a modified system which is given by g (u + γ v ) with the parameter
γ ∈ [0, 1], instead of g (u + v ). This model system covers the present case. We show that the new
system also possesses a Lyapunov function for any γ ∈ [0, 1]. We conclude the paper by Appendices A
and B in which it is shown that the Turing type instability certainly takes place for a specific g =
kw /(aw + 1)2 , and the proof of Lemma 2.2 is given.

2. Proofs of Propositions 1.1 and 1.2

2.1. Proof of Proposition 1.1

We simply write the equations of (1.1) as

ut = du − f (u , v ), v t =  v + f (u , v ).

It is known that the solutions u (x, t ) and v (x, t ) are nonnegative and smooth in Ω × (0, ∞). We first
prove that there are c 0 , M 0 > 0 such that

0  u (x, t ), v (x, t )  c 0 e M 0 t . (2.1)

Indeed, from the assumption there are ai j > 0 (i , j = 1, 2) for which

− f (u , v )  a11 u + a12 v , f (u , v )  a21 u + a22 v

hold, thus, the solutions satisfy the differential inequalities

ut  du + a11 u + a12 v , v t   v + a21 u + a22 v .

With the aid of the solutions p (t ), q(t ) of the ODEs

pt = a11 p + a12 q, qt = a21 p + a22 q,

with the initial conditions

p (0) = sup u 0 (x) + 1, q(0) = sup v 0 (x) + 1,

x∈Ω x∈Ω

we can apply the comparison argument to show

0  u (x, t )  p (t ), 0  v (x, t )  q(t ).

Since the solutions p (t ) and q(t ) have at most the exponential growth, we obtain (2.1).
Before going to the proof of the global boundedness of the solutions, we prepare two lemmas.
1664 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

Lemma 2.1. There exists a constant K = K (Ω) > 0 such that

   
φL ∞ (Ω)  1/|Ω| φ L 1 (Ω) + K (Ω)∇φL ∞ (Ω) φ ∈ C 1 (Ω; R)

holds.

Proof. From the regularity of ∂Ω , there exist a constant c > 0 and a curve γ : [0, 1] → Ω with the
conditions

 
γ (0) = x, γ (1) = y , γ̇ (τ )  c (0  τ  1)

for arbitrarily given x, y ∈ Ω . Take any φ ∈ C 1 (Ω)

1 1
d    
φ(x) − φ( y ) = φ γ (τ ) dτ = ∇φ γ (τ ) · γ̇ (τ ) dτ .
dτ
0 0

Hence, we have

   
φ(x)  φ( y ) + c ∇φL ∞ (Ω) .

Integrate both sides of this inequality in y, we easily see the desired inequality follows. 2

Next, let G 1 (x, y , t ) and G 2 (x, y , t ) be the heat kernels for d and  with the Neumann boundary
condition respectively. We have the next estimate.

Lemma 2.2. For the heat kernels G j ( j = 1, 2)

  c1  
∇x G j (x, y , t )  exp −η1 t − |x − y |2 /η2 t (x, y ∈ Ω, t > 0, j = 1, 2)
t (n+1)/2

hold, where c 1 , η1 and η2 are positive constants.

We note η j ( j = 1, 2) depend on d. As for the proof of this lemma see Appendix B.

Using G 1 , G 2 , we have the integral expression for Eqs. (1.1)–(1.2), that is,


u (x, t ) = G 1 (x, y , t )u 0 ( y ) dy
Ω

t 
 
− G 1 (x, y , t − τ ) f u ( y , τ ), v ( y , τ ) dy dτ , (2.2)
0 Ω

v (x, t ) = G 2 (x, y , t ) v 0 ( y )
Ω

t 
 
+ G 2 (x, y , t − τ ) f u ( y , τ ), v ( y , τ ) dy dτ . (2.3)
0 Ω
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1665

Given a small constant δ ∈ (0, 1), we consider the solutions for t  δ and estimate u (t , x), v (t , x) in
[δ, ∞) × Ω . From (2.2) and (2.3) we see

 t −δ
∇ u (t , x) = ∇ G 1 (x, y , t )u 0 ( y ) dy − ∇ G 1 (x, y , t − τ ) f (u , v ) dy dτ
Ω 0 Ω

t 
− ∇ G 1 (x, y , t − τ ) f (u , v ) dy dτ ,
t −δ Ω

 t −δ
∇ v (t , x) = ∇ G 2 (x, y , t ) v 0 ( y ) dy + ∇ G 2 (x, y , t − τ ) f (u , v ) dy dτ
Ω 0 Ω

t 
+ ∇ G 2 (x, y , t − τ ) f (u , v ) dy dτ ,
t −δ Ω

where we simply write f (u , v ) = f (u ( y , τ ), v ( y , τ )).

Recall w (x, t ) = u (x, t ) + v (x, t ) and


w (x, t ) dx = |Ω|s (t  0).
Ω

Define a function

 
ρ (t ) := sup w (x, τ ): 0  τ  t , x ∈ Ω . (2.4)

Lemma 2.3. There is a small number δ > 0, which is independent of the initial data, such that the function ρ
defined by (2.4) satisfies

 
0  ρ (t )  max Ms, ρ (δ) (t  0),

where M depends on c 1 , η1 , η2 , δ, Ω and ai j (1  i , j  2) but independent of s.

Proof. We first estimate the solutions in the region t  δ ,

 t −δ
       
∇ u (x, t )  ∇ G 1 (x, y , t )u 0 ( y ) dy + ∇ G 1 (t − τ , x, y ) f (u , v ) dy dτ
Ω 0 Ω

t 
  
+ ∇ G 1 (t − τ , x, y ) f (u , v ) dy dτ
t −δ Ω

 t −δ
c1   c 1 e −η1 (t −τ )  
 u 0 ( y ) dy + a11 u ( y , τ ) + a12 v ( y , τ ) dy dτ
δ (n+1)/2 δ(n+1)/2
Ω 0 Ω
1666 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

t   
c1 |x − y |2  
+ exp − a11 u ( y , τ ) + a12 v ( y , τ ) dy dτ
(t − τ )(n+1)/2 η2 (t − τ )
t −δ Ω

t
c 1 u 0  L 1 c 1 s|Ω| c 1 dτ
 + max(a11 , a12 ) + (πη2 )n/2 √ · max(a11 , a12 )ρ (t )
δ (n+1)/2 δ (n+1)/2 η1 t−τ
t −δ
c1   √
 u 0 L 1 + s|Ω| max(a11 , a12 )/η1 + 2c 1 (πη2 )n/2 δ max(a11 , a12 )ρ (t ).
δ (n+1)/2
Similarly, we obtain

  c1  
∇ v (x, t )   v 0 L 1 + s|Ω| max(a21 , a22 )/η1
δ (n+1)/2
√
+ 2c 1 (πη2 )n/2 δ max(a21 , a22 )ρ (t ).

By Lemma 2.1, we have


1 c1  
0  u (x, t )  v (·, t ) + K (Ω) u 0 L 1 + s|Ω| max(a11 , a12 )/η1
|Ω| δ (n+1)/2
L1

√
+ 2c 1 (πη2 )n/2 δ max(a11 , a12 )ρ (t ) ,

1 c1  
0  v (x, t )  u (·, t ) + K (Ω)  v 0 L 1 + s|Ω| max(a21 , a22 )/η1
|Ω| L1 δ (n+1)/2

√
+ 2c 1 (πη2 )n/2 δ max(a21 , a22 )ρ (t )

for t  δ , x ∈ Ω . We choose δ > 0, which is independent of the initial data, so that

√
2K (Ω)c 1 (πη2 )n/2 δ max(ai j : 1  i , j  2)  1/4 (2.5)

is satisfied. Then

K (Ω)c 1 s   1
0  w (x, t )  s + η1 + 2|Ω| max(ai j : 1  i , j  2) + sup w (x, τ )
η1 δ (n+1)/2 2 x∈Ω,0τ t

for t  δ . By considering the non-blowup of the solutions in a finite time, a simple argument leads us
to
 
2K (Ω)c 1 s  
0  ρ (t )  max 2s + η1 + 2|Ω| max(ai j : 1  i , j  2) , ρ (δ)
η1 δ (n+1)/2
for t  0. It is clear by (2.5) that δ is chosen independently of the initial data. Consequently, we obtain
the desired assertion of the lemma. 2

By virtue of Lemma 2.3 we have the global boundedness of the orbit defined by the solutions
(u (t , x), v (t , x)) in L ∞ (Ω)2 . Applying this result to the estimating of |∇ u (x, t )| and |∇ v (x, t )| with the
aid of the integral equations, we easily get the global C 1 estimates for (u (t , x), v (t , x)), in consequence,
we conclude the proof of Proposition 1.1.
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1667

2.2. Proof of Proposition 1.2

We prove Proposition 1.2. Let (u (·, t ), v (·, t )) be a solution of (1.1) with (1.2). A straightforward
computation shows the following:

d  
E u (·, t ), v (·, t )
dt


= d∇(u + v ) · ∇(u + v )t + ∇(du + v ) · ∇(du + v )t
Ω

+ (1 − d) g (u + v ) + dk(u + v ) (u + v )t dx


= −d(u + v ) + (1 − d) g (u + v ) + dk(u + v ) (u + v )t
Ω

+ ∇(du + v ) · ∇(du + dv )t + (1 − d)∇(du + v ) · ∇ v t dx


=− d(u + v ) − (1 − d) g (u + v ) − dk(u + v ) (u + v )t
Ω

+ d(du + v )(u + v )t + (1 − d)(du + v ) v t dx.

Then we use

ut + dv t = d(u + v ) − (1 − d) g (u + v ) + k(1 − d) v ,

and

ut + v t = (du + v ),

to obtain

d   
E u (·, t ), v (·, t ) = − ut + dv t − k(1 − d) v − dk(u + v ) (u + v )t
dt
Ω

+ d(ut + v t )2 + (1 − d)(ut + v t ) v t dx

 
=− ut + dv t − k(du + v ) + (1 − d) v t (u + v )t + d(ut + v t )2 dx
Ω

 
=− (1 + d)(ut + v t )2 − k(du + v )(du + v ) dx
Ω

  2 
=− (1 + d)(ut + v t )2 + k∇(du + v ) dx  0.
Ω

Thus

d  
E u (·, t ), v (·, t ) = 0 (∀t ∈ R)
dt
1668 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

implies

(u + v )t = w t = 0 and du + v = dw + (1 − d) v = constant in x.

This yields W t = V t = 0 in (1.9). Then the solution ξ(t ) of (1.11) defined for all t must be a steady
state. This concludes the proof of Proposition 1.2. 2

3. Linearized eigenvalue problem

For the operator of (1.17) we consider the eigenvalue problem

   
φ φ
A =λ , (3.1)
ψ ψ

that is,

dφ + (1 − d)ψ = −λφ, (3.2)

 ∗

ψ + g w φ − kψ = −λψ. (3.3)

We need to prepare several lemmas to prove Theorem 1.3.

Lemma 3.1. The spectrum of A consists of eigenvalues.

Proof. Since the principal part of the operator A can be diagonalized as

  −1      
1 −1 1 1 d 0 −g (w∗) −g (w∗) + k
A =− − ,
0 1 0 1 0  g (w∗) g (w∗) − k

the assertion of the lemma immediately follows from [15] (Theorem 6.29, Chap. 3, §§6–8). 2

Lemma 3.2. Let λ be an eigenvalue of A. If Re λ < k/2, then λ is real.

Proof. Combining the two equations of (3.2)–(3.3), we have

 
dφ − (1 − d) g w ∗ φ + (1 − d)kψ = −λφ + (1 − d)λψ. (3.4)

By putting
 
1−d d
χ =φ+ ψ i.e. ψ = (χ − φ) ,
d 1−d

(3.2)–(3.3) are transformed to

dχ = −λφ, (3.5)

 ∗

dφ − (1 − d) g w φ + dk(χ − φ) = −λφ + dλ(χ − φ). (3.6)

We prove that if Im λ = 0, then Re λ  k/2. Set

λ = λ 1 + i λ2 , χ = χ1 + i χ2 , φ = φ1 + i φ 2 ,
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1669

and
 
J 1 := (χ1 φ1 + χ2 φ2 ) dx, J 2 := (χ1 φ2 − χ2 φ1 ) dx,
Ω Ω

where λ j ∈ R ( j = 1, 2) and χ j , φ j ( j = 1, 2) are real-valued. Then by simple computation we obtain

 
d ∇ χ1 2 + ∇ χ2 2 = (λ1 + i λ2 )( J 1 + i J 2 ),

    
d ∇φ1 2 + ∇φ2 2 + d(λ1 + i λ2 − k)( J 1 − i J 2 ) + (1 − d) g w ∗ φ12 + φ22 dx
Ω
  
= (1 + d)(λ1 + i λ2 ) − dk φ1 2 + φ2 2 .

Taking the real part and imaginary part of the equalities yields

 
d ∇ χ1 2 + ∇ χ2 2 = λ1 J 1 − λ2 J 2 ,

λ2 J 1 + λ1 J 2 = 0,
 
d λ2 J 1 + (k − λ1 ) J 2 = λ2 (1 + d) φ1 2 + φ2 2 .

From these equations

   
λ2 d ∇ χ1 2 + ∇ χ2 2 = − λ21 + λ22 J 2 ,
 
λ2 (1 + d) φ1 2 + φ2 2 = −d(2λ1 − k) J 2

follow. Hence by λ2 = 0 we get to

    
d2 (2λ1 − k) ∇ χ1 2 + ∇ χ2 2 = (1 + d) λ21 + λ22 φ1 2 + φ2 2  0,

which yields 2λ1 − k  0. Hence, we obtain the desired assertion. 2

We define the closed operator A 0 the extension of − with the Neumann boundary condition
in L 2 (Ω). Namely, A 0 satisfies
  
A 0 v = − v , v ∈ u ∈ H 2 (Ω; R): ∂ u /∂ ν = 0 (x ∈ ∂Ω), u dx = 0
Ω

and A 0s stands for the fractional power of A 0 with index s. We note that A 0 is self-adjoint, positive
and invertible in L 2 (Ω). Thus A 0−1 is defined as a bounded operator in L 2 (Ω).
By using A 0−1 , we obtain the equation

d λ
ψ − ψ =− φ+ A− 1
0 φ, (3.7)
1−d 1−d

from (3.2). Inserting this into (3.4) yields

  λ
φ − α g w ∗ φ − kφ + α (k − λ) ψ = − (1 + d) + (k − λ) A −
0
1
φ
d
1670 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

which is decomposed to

  
g w ∗ φ − (k − λ) ψ = 0, (3.8)

and

λ 
L(φ) = (1 + d) + (k − λ) A −
0
1
φ, (3.9)
d

where L is defined by (1.18). Given a solution (φ, λ) of (3.9), ψ is determined by (3.8). In the sequel
the eigenvalue problem of A is reduced to that of (3.9).

Lemma 3.3. The algebraic multiplicity and the geometric multiplicity of any eigenvalue λ of A coincide if
Re λ < k/2.

Proof. By Lemma 3.2 λ is real if Re λ < k/2. We let λ0 be any eigenvalue with λ0 < k/2 and (φ0 , ψ0 )
be a corresponding eigenfunction. It suffices to prove

Ker(A − λ0 I ) = Ker(A − λ0 I )m (m  2).

We first prove that

 
φ1
(A − λ0 I )2 =0
ψ1

implies (φ1 , ψ1 ) T is an eigenfunction corresponding to λ0 , where I is the identity. We consider the

equation
   
φ φ0
(A − λ0 I ) = ,
ψ ψ0

that is,

dφ + (1 − d)ψ + λ0 φ = −φ0 , (3.10)

 ∗

ψ + g w φ − kψ + λ0 ψ = −ψ0 , (3.11)

and prove (3.10)–(3.11) have no solution.

From (3.10) we see

1 d 1  −1 
ψ = − (dφ + λ0 φ + φ0 ), ψ − ψ =− φ+ λ0 A − 1
0 φ + A 0 φ0 .
1−d 1−d 1−d

Thus,

  λ0
−φ + α g w ∗ φ + kφ − α (k − λ0 ) ψ − (1 + d)φ + (k − λ0 ) A − 1
0 φ
d
1 k − λ0 − 1
= −α ψ0 + φ0 + A 0 φ0 .
d d

Decomposing the equation we obtain

 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1671

  
g w ∗ φ − (k − λ0 ) ψ = − ψ0 ,
λ0   1 k − λ0 −1
L(φ) − (1 + d) + (k − λ0 ) A −
0
1
φ = −α ψ0 − ψ0 + φ0 + A 0 φ0 . (3.12)
d d d

By

d λ0 − 1
ψ 0 − ψ0 = − φ0 + A φ0 ,
1−d 1−d 0

(3.12) turns to be

λ0 1+d k − 2 λ0 − 1
L(φ) − (1 + d) + (k − λ0 ) A −
0
1
φ= φ0 + A 0 φ0 . (3.13)
d d d

Invoking

λ0
L(φ0 ) − (1 + d)φ0 + (k − λ0 ) A − 1
0 φ 0 = 0,
d

the right hand side of (3.13) is orthogonal to φ0 in L 2 (Ω), namely,

 1 
(1 + d)(φ0 , φ0 ) L 2 + (k − 2λ0 ) A −
0 φ0 , φ0 L 2
−1/2 2
= (1 + d)φ0 2 + (k − 2λ0 ) A 0 φ0 = 0.

This is impossible since φ0 = 0 and λ0 < k/2.

For the general case (A − λ0 I )m φ = 0 (m  2), we can prove that it has only the solution φ0
up to multiplication of a constant by the induction argument. Indeed, if there is no solution of
(A − λ0 I )m−1 φ1 = φ0 , then the null space of (A − λ0 I )m consists of {c φ0 : c ∈ R} by the above ar-
gument. This completes the proof. 2

As for Eq. (3.3), decompose ψ = ψ + ψ̂ . Then the equation is also decomposed as

  
g w ∗ φ − k ψ = −λ ψ ,
    
ψ̂ + g w ∗ φ − g w ∗ φ − kψ̂ = −λψ̂, ψ̂ = 0.

Thus, it suffices to consider the equation for ψ̂ since ψ follows from the solution (ψ̂, λ).
Instead of A we may consider the eigenvalue problem of the operator

   
φ dφ + (1 − d)ψ
A1 := − (3.14)
ψ ψ + g ( w ∗ )φ − g ( w ∗ )φ − kψ

with the domain

 
D(A1 ) = (φ, ψ) ∈ H 2 (Ω) × H 2 (Ω): ∂φ/∂ ν = ∂ψ/∂ ν = 0 (x ∈ ∂Ω) .

Moreover, from the same computation found in the proof of the above lemma we see that the eigen-
value problem for A1 is reduced to the problem
1672 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

λ
L(φ) = (1 + d) + (k − λ) A −
0
1
φ (3.15)
d

in the domain D (L).

To study (3.15), we introduce a nonlocal eigenvalue problem

  ∂φ
L[φ] = Λ dc + β A −
0
1
φ (x ∈ Ω), = 0 (x ∈ ∂Ω), (3.16)
∂ν

where β is nonnegative parameter. This is an eigenvalue problem of L with the weight in L 2 (Ω).
Then we have the following comparison result for the eigenvalues of L and Λ’s of (3.16), which will
be proved in the next section.

Lemma 3.4. Let {μ j } j =1,2,... and {Λ j (β)} j =1,2,... be the sets of eigenvalues of L and (3.16) for β  0 arranged
in increasing order with counting the multiplicity respectively. If μ j > 0, then

μj μj
 Λ j (β)  , (3.17)
dc + β/σ2 dc

while if μ j < 0, then

μj μj
 Λ j (β)  (3.18)
dc dc + β/σ2

holds, where σ2 denotes the second eigenvalue of − with the Neumann boundary condition in the domain Ω .

Now turn to (3.14) or (3.15). We write the equation as

 
1+d k − λ −1
L[φ] = λ + A 0 φ.
d d

Compare this and (3.16) with dc = (1 + d)/d. If there is β ∗ solving the equation

k − Λ j (β)
= β, (3.19)
d

then Λ j (β ∗ ) and the corresponding eigenfunction φ j (·; β ∗ ) give solutions to (3.15). Hence, Λ j (β ∗ ) is
an eigenvalue of A1 , that is, A. The next lemma is crucial to realize this idea.

Lemma 3.5. Each Λ j (β) is continuous in β  0. Moreover, if μ j > 0, then Λ j (β) is strictly monotone de-
creasing while if μ j < 0, then Λ j (β) is strictly monotone increasing.

We will prove this key lemma in the next section. We apply Lemmas 3.4 and 3.5 to complete the
proof of Theorem 1.3.

Proof of Theorem 1.3. We first see that by Lemma 3.3 the multiplicity of zero eigenvalue for A and
L coincides. We put dc = (1 + d)/d and consider the eigenvalue problem (3.16). Assume that

μ1  μ2  · · ·  μ p < 0  μ p+1  · · ·  μN < k/2.

Since Λ j (0) = μ j /dc , Lemmas 3.4 and 3.5 tells that, for each j  p, Λ j (β) < 0 (β > 0) and k − Λ j (β)
is monotone decreasing so that a unique solution β = β ∗j of (3.19) exists. Moreover we see that
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1673

Fig. 1. The curves γ = γ j (β) := (k − Λ j (β))/d and γi (β) := (k − Λ j (β))/d for the case μ j < μi < 0 in (a) and 0 < μ j < μi
in (b). The horizontal axis is β -axis. The intersection points at the line γ = β give β ∗j and βi∗ respectively. Note Λn (0) = μn /dc .

0 < β1∗  β2∗  · · ·  β p∗

(see Fig. 1), consequently we obtain

     
Λ1 β1∗  Λ2 β2∗  · · ·  Λ p β p∗ < 0.

Conversely given an eigenvalue λm < 0 of A, we consider (3.15), i.e.

 
1+d k − λm −1
L[φm ] = λm + A 0 φm . (3.20)
d d

Put βm := (k − λm )/d. Then we easily see that λm ∈ {Λ j (βm )} j =1,2,... . The number of the negative
eigenvalues of L coincides with that of A.
If 0 < μ j < k/2, applying Lemma 3.5 similarly, we see that Λ j (β) > 0 (β > 0) and k − Λ j (β) is
monotone increasing so that there is a unique β ∗j > 0 solving (3.19). Since k − Λ j (β ∗j ) > 0, Λ j (β ∗j )
gives a positive eigenvalue of A. We also easily see that if A has q positive eigenvalues less than k/2,
then L has also so.
Finally, (1.21) immediately follows from Lemma 3.4. This completes the proof. 2

4. A nonlocal eigenvalue problem

Given a positive dc , we introduce the following inner product:

 −1/2 −1/2   
(u , v )β := dc (u , v ) L 2 + β A 0 u , A 0 v L 2 u , v ∈ L 2 (Ω; R) ,

and the norm

 −1/2 2 1/2
u β := dc u 2 + β A 0 u

for β  0. We note that

σ2 A − 1/2 2
0 u  u 2

holds, where σ2 be the second eigenvalue of − with the Neumann boundary condition. Thus, the
norm  · β gives an equivalent norm to the usual one of L 2 .
1674 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

In what follows we fix dc and we will observe the dependence of eigenvalues on β . Given a

real-valued continuous function a(x) of Ω , we consider the eigenvalue problem
⎧   
⎨ La [φ] := −φ + a(x)φ − a(·)φ = Λ dc + β A −
0
1
φ (x ∈ Ω),
∂φ (4.1)
⎩ =0 (x ∈ ∂Ω).
∂ν
Let {Λ j (β)} j =1,2,... be the set of eigenvalues of (4.1) arranged in increasing order with counting the
β
multiplicity and let {φ j (x)} j =1,2,... be the corresponding eigenfunctions. We formulate the variational
characterization of the eigenvalues.
Let Mn be the set of all the n-dimensional subspaces of X = H 1 (Ω; R), that is,

Mn = { X n ⊂ X: Xn is a linear subspace with dim X n = n}.

We introduce the min–max principle for the eigenvalue problem (see [7]). Define the Rayleigh quo-
tient as

K [φ]  
R [φ; β] := , K [φ] := ∇φ2 + a(·)φ, φ L2
.
φβ 2

Then we have

 
Λn (β) = inf sup R [φ; β]: φ ∈ X n , φ = 0 . (4.2)
Xn ∈Mn

Moreover, by the standard argument we have the alternative formulations as

 
Λn (β) = sup inf R [φ; β]: φ ∈ X n⊥−1 , φ = 0 (4.3)
X n−1 ∈Mn−1
   
= inf R [φ; β]: φ ∈ X , φ = 0, φ, φ j (·; β) β = 0 ( j = 1, 2, . . . , n − 1) , (4.4)

where

 
Y ⊥ := u: (u , v )β = 0 (∀ v ∈ Y ) .

Lemma 4.1. The zero eigenspace for (4.1) does not depend on β  0.

Proof. The zero eigenspace is the set of solutions of La [φ] = 0 which is independent of β  0. 2

Lemma 4.2. For 0 < β1  β2

φ2β2 φ2β1
φβ1  φβ2 ,  (φ ∈ X ),
β2 β1

and
 
β2
φ20  φ2β2  1 + φ20 (φ ∈ X )
σ2 d c
hold.
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1675

Proof. The first and the third inequalities immediately follows form the definitions of  · β and σ2 .
For the second inequality, given φ ∈ X , we obtain

φ2β2 dc −1/2 2 dc −1/2 2 φ2β1

= φ2 + A 0 φ  φ2 + A 0 φ = .
β2 β2 β1 β1

This completes the proof. 2

Lemma 4.3. If 0 < β1  β2 and Λn (β2 ) > 0, then Λn (β2 )  Λn (β1 ).

Proof. From the inf–sup type max–min principle (4.2) and the condition Λn (β2 ) > 0, there exists
δ > 0 such that

   
sup K [φ]/φ2β2 : φ ∈ E , φ = 0  δ E ∈ Mn .

By the assumption we can assert

   
sup K [φ]/φ2β2 : φ ∈ E , φ = 0 = sup K [φ]/φ2β2 : φ ∈ E , φ = 0, K [φ]  0 .

From the max–min principle (4.2) we see

 
K [φ]
Λn (β2 ) = inf n sup : φ ∈ E, φ = 0
E ∈M φ2β2
 
K [φ]
= inf n sup : φ ∈ E , φ = 0, K [φ]  0
E ∈M φ2β2
 
K [φ]
 inf n sup : φ ∈ E , φ = 0, K [φ]  0
E ∈M φ2β1
 
K [φ]
= inf n sup : φ ∈ E , φ = 0 = Λn (β1 ).
E ∈M φ2β1

This proves the desired inequality. 2

Lemma 4.4. If 0 < β1  β2 and Λn (β1 ) > 0, then 0 < Λn (β1 )  (β2 /β1 )Λn (β2 ).

Proof. By (4.2) and Λn (β1 ) > 0 there exists δ > 0 such that

   
sup K [φ]/φ2β1 : φ ∈ E , φ = 0  δ E ∈ Mn .

Then we see

   
sup K [φ]/φ2β1 : φ ∈ E , φ = 0 = sup K [φ]/φ2β1 : φ ∈ E , φ = 0, K [φ]  0 .

Using (4.2) and Lemma 4.2, we have

1676 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

 
K [φ]
Λn (β1 ) = inf n sup : φ ∈ E, φ = 0
E ∈M φ2β1
 
K [φ]
= inf n sup : φ ∈ E , φ = 0, K [φ]  0
E ∈M φ2β1
 
β2 K [φ]
 inf n sup : φ ∈ E , φ = 0, K [φ]  0
E ∈M β1 φ2β
2
 
β2 K [φ]
 inf sup : φ ∈ E , φ = 0 , K [φ]  0
β1 E ∈Mn φ2β2
 
β2 K [φ] β2
= inf sup : φ ∈ E , φ = 0 = Λn (β2 ).
β1 E ∈M n
φβ22 β1

This completes the proof. 2

Combining Lemmas 4.1, 4.3 and 4.4, we have the following result.

Lemma 4.5. If Λn (β0 ) > 0 for a β0 > 0, then Λn (β) > 0 for all β > 0. In addition, Λn (β0 ) < 0 for a β0 > 0
yields Λn (β) < 0 for all β > 0.

Proof. The first assertion follows directly from Lemmas 4.3 and 4.4. To see the second assertion, sup-
pose it is not true. Then there is β1 > 0 such that Λn (β1 )  0. By Lemma 4.1 Λn (β) = 0 if Λn (β1 ) = 0,
while Λ(β) > 0 if Λn (β1 ) > 0, so we get a contradiction. 2

We have the following inequalities for negative eigenvalues.

Lemma 4.6. Assume that there exists a number β0 > 0 for which Λn (β0 ) < 0 is met. Then

Λn (β1 )  Λn (β2 )  (β1 /β2 )Λn (β1 )

holds if 0 < β1  β2 .

Proof. We notice that the assumption implies Λn (β) < 0 for any β > 0 by Lemma 4.5. Since
Λn (β2 ) < 0, we see that arbitrarily given  ∈ (0, −Λn (β2 )), there exists E  ∈ Mn such that

 
sup K [φ]/φ2β2 : φ ∈ E  , φ = 0 < Λn (β2 ) +  .

Thus K [φ] < 0 for φ ∈ E  . With the aid of this and φ2β1  φ2β2 , we have

 
Λn (β1 ) = inf n sup K [φ]/φ2β1 : φ ∈ E , φ = 0
E ∈M
 
 sup K [φ]/φ2β1 : φ ∈ E  , φ = 0
 
 sup K [φ]/φ2β2 : φ ∈ E  , φ = 0 < Λn (β2 ) +  .

Taking  ↓ 0, we obtain Λn (β1 )  Λn (β2 ).

 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1677

We prove the latter inequality. By Λn (β1 ) < 0, for any  ∈ (0, −Λn (β1 )), there exists E  ∈ Mn such
that

 
sup K [φ]/φ2β1 : φ ∈ E  , φ = 0 < Λn (β1 ) +  .

Using K [φ] < 0 for any φ ∈ E  and 1/φ2β2  (β1 /β2 )(1/φ2β1 ), we have

 
Λn (β2 ) = inf n sup K [φ]/φ2β2 : φ ∈ E , φ = 0
E ∈M
 
 sup K [φ]/φ2β2 : φ ∈ E  , φ = 0
 
β1 K [φ] β1  
 sup : φ ∈ E  , φ = 0 < Λn (β1 ) +  .
β2 φβ2 β2
1

Taking  ↓ 0 yields Λn (β2 )  (β1 /β2 )Λn (β1 ). 2

Combining the above results leads to the following continuity property for the eigenvalues in β .

Proposition 4.7. Every eigenvalue Λn (β) of (4.1) is continuous in β > 0. More precisely, for each β0 > 0

(i) if Λn (β0 ) < 0, then Λn (β) < 0 for any β > 0 and

β0 β0  
Λn (β0 )  Λn (β0 + t )  Λn (β0 ) t ∈ (−β0 , β0 ) ;
β0 − |t | β0 + |t |

(ii) if Λn (β0 ) > 0, then Λn (β) > 0 for any β > 0 and

β0 β0  
Λn (β0 )  Λn (β0 + t )  Λn (β0 ) t ∈ (−β0 , β0 ) ;
β0 + |t | β0 − |t |

(iii) if Λn (β0 ) = 0, then Λn (β) = 0 for every β > 0.

Proof. The assertion of (iii) follows from Lemma 4.1. We prove (i) by Lemma 4.6. Since Λn (β) < 0,
we have

β0 β0
Λn (β0 )  Λn (β0 )  Λn (β0 + t )  Λn (β0 ),
β0 − t β0 + t

for t  0. If t < 0, then

β0 + t β0
Λn (β0 )  Λn (β0 + t ), Λn (β0 + t )  Λn (β0 )  Λn (β0 ).
β0 β0 − t

Combining those yields the inequalities of (i).

It is easy to verify (ii) by Lemmas 4.4 and 4.5 similarly. We leave the details to the readers. 2

The next proposition ensures the continuity at β = 0.

1678 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

Proposition 4.8. For the eigenvalue Λn (β)

(i) if Λn (0) < 0, then

Λn (0)
Λn (0)  Λn (β)  (for β > 0);
1 + (β/σ2 dc )

(ii) if Λn (0) > 0, then

Λn (0)
 Λn (β)  Λn (0) (for β > 0);
1 + (β/σ2 dc )

(iii) if Λn (0) = 0, then Λn (β) = 0 for β > 0.

The proof of this proposition is carried out similarly as done for Proposition 4.7. Indeed, using the
third inequality of Lemma 4.2, we see that the assertion of Lemma 4.3 is true for the case β1 = 0.
Similarly, with the aid of the same inequality, by replacing β2 /β1 by 1 + (β2 /σ2 dc ), the assertions
of Lemmas 4.4 and 4.6 are also true for β1 = 0. Hence it is not so difficult to complete the proof of
Proposition 4.8. We leave the details of the proof to the readers.

Proof of Lemmas 3.4 and 3.5. Put dc = (1 + d)/d. By Proposition 4.7 together with Lemmas 4.4, 4.5
and 4.6, we obtain Lemma 3.5. Invoking Λn (0) = μn /dc = dμn /(1 + d) and applying Proposition 4.8
yield Lemma 3.4. 2

Remark 4.1. We are able to prove that if Λn (β0 ) is simple for some β0 , then there exists τ > 0 such
that Λn (β) is differentiable in β in (β0 − τ , β0 + τ ) and

dΛn −Λn (β0 )( A − 1

0 φn (·, β), φn (·, β))
(β0 ) = .
dβ ((dc + β0 A − 1
0 )φn (·, β), φn (·, β))

We omit the proof, since it is not used in the proof of the main result of the present article.

5. Lyapunov function for a modified system

In the papers [19] and [18] the authors treat the case g (u ), instead of g (u + v ), namely

ut = du − g (u ) + kv ,
(x ∈ Ω, t > 0). (5.1)
v t =  v + g (u ) − kv

Then the system allows the Lyapunov function

  
d dk k
E0 (u , v ) := |∇ u |2 + G (u ) + u2 + (du + v )2 dx (5.2)
2 2 2(1 − d)
Ω

if 0 < d < 1, where

u
G (u ) := g (ξ ) dξ.

It is clear that the functional (5.2) differs from (1.5). However, motivated by the fact that both systems
possess the Lyapunov functions, we get an idea to consider the following modified system:
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1679


ut = du − g (u + γ v ) + kv ,
(x ∈ Ω, t > 0) (5.3)
v t =  v + g (u + γ v ) − kv

with the Neumann boundary conditions

∂u ∂v
= = 0 (x ∈ ∂Ω), (5.4)
∂ν ∂ν

where d, k are positive constants, γ is a constant in [0, 1]. This system coincides with (1.1) and (5.1)
when γ = 1 and γ = 0 respectively.
We show that the system possesses a Lyapunov function.

Proposition 5.1. Let 0 < d < 1 and 0  γ  1. Then the following functional gives a Lyapunov function
of (5.3) with (5.4):

 
d 
Eγ (u , v ) := ∇(u + γ v )2 + (1 − dγ )G (u + γ v ) + dk (u + γ v )2
2 2
Ω

k(1 − γ ) γ  2
+ (du + v )2 + ∇(du + v ) dx, (5.5)
2(1 − d) 2

namely, for a smooth solution (u (x, t ), v (x, t )),

 
  1 − d2 γ    
d
Eγ u (·, t ), v (·, t ) = − (u + γ v )t 2 dx − k(1 − dγ ) ∇(du + v )2 dx
dt 1 − dγ 1−d
Ω Ω

γ (1 − γ )  
− (du + v )t 2 dx  0
1 − dγ
Ω

holds.

We note (5.5) coincides with (5.2) and (1.5) for γ = 0 and γ = 1 respectively.

Proof. We may assume γ < 1 since the case γ = 1 is proved in Section 2. We introduce the new
variables

U = u + γ v, V = du + v .

By

1 1
u= (U − γ V ), v= (−dU + V ),
1 − dγ 1 − dγ

the equations are transformed into

1 − dγ γ (1 − d)
U t = dU +  V − (1 − dγ ) g (U ) − dkU + kV , (5.6)
1−γ 1−γ
1−d 1−γ
Ut + V t = V , (5.7)
1 − dγ 1 − dγ
1680 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

and (5.5) is written as

  
d 2 dk 2 k(1 − γ ) 2 γ 2
Eγ ( U , V ) = |∇ U | + (1 − dγ )G (U ) + U + V + |∇ V | dx.
2 2 2(1 − d) 2
Ω

For the solution (U (x, t ), V (x, t )) of (5.6)–(5.7) with the Neumann conditions

d  
Eγ U (·, t ), V (·, t )
dt
  
k(1 − γ )
= d∇ U · ∇ U t + (1 − dγ ) g (U )U t + kdU U t + V V t + γ ∇ V · ∇ V t dx
1−d
Ω
    
1 − dγ
= −dU + (1 − dγ ) g (U ) + kdU U t + kV  V − U t − γ V t  V dx
1−d
Ω
  
1 − dγ γ (1 − d)
= − Ut +  V + kV U t
1−γ 1−γ
Ω
 
k(1 − dγ ) 1−d 1−γ
− |∇ V |2 − kV U t − γ V t Ut + Vt dx
1−d 1 − dγ 1 − dγ
   
1 − dγ γ (1 − d) 1 − d 1−γ
= − U t2 + Ut + Vt Ut
1−γ 1−γ 1 − dγ 1 − dγ
Ω

γ (1 − d) γ (1 − γ ) 2 k(1 − dγ )
− Vt Ut − Vt − |∇ V |2 dx
1 − dγ 1 − dγ 1−d
    
1 − dγ γ (1 − d)2 2 γ (1 − γ ) 2 k(1 − dγ )
=− 1− Ut + Vt + |∇ V |2 dx.
1−γ (1 − dγ )2 1 − dγ 1−d
Ω

Since

(1 − d)2 (1 − γ )(1 − d2 γ )
1−γ = ,
(1 − dγ )2 (1 − dγ )2

we conclude the proof. 2

Acknowledgments

The first author was supported in part by the Grant-in-Aid for Scientific Research (C) No. 22540216,
Japan Society for the Promotion of Science. The second author was partially supported by the Grant-
in-Aid for Scientific Research (B) No. 22340022 and Challenging Exploratory Research No. 24654044,
Japan Society for the Promotion of Science. The authors express their thanks to the referee for valu-
able comments to the improvement of the manuscript.
 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1681

Appendix A. Turing-type instability

As for a specific example

w
g ( w ) = kg 0 ( w ), g 0 ( w ) := , (A.1)
(aw + 1)2

which is given in [14] and [20], we verify the condition for the instability of a constant steady state
of (1.1). Under the condition (1.4), we obtain a constant steady state

 
(u , v ) = s − g 0 (s), g 0 (s) . (A.2)

This solution is asymptotically stable for the diffusion-free equations

du dv
= −kg 0 (u + v ) + kv , = kg 0 (u + v ) − kv (A.3)
dt dt

under the constraint u + v = s because the equations are reduced to a scalar equation

du  
= k − g 0 (s) + s − u .
dt

The equilibrium u = s − g 0 (s) is asymptotically stable.

Let σ j be the j-th eigenvalue of − with the Neumann boundary conditions. The linearized eigen-
value problem

− dφ − kg 0 (s)(φ + ψ) + kψ = λφ,

− ψ + kg 0 (s)(φ + ψ) − kψ = λψ (x ∈ Ω),

∂φ/∂ ν = ∂ψ/∂ ν = 0 (x ∈ ∂Ω)

can be reduced to an infinite sequence of finite-dimensional eigenvalue problems

 
dσ j + kg 0 (s) kg 0 (s) − k
A j := ( j = 1, 2, . . .).
−kg 0 (s) σ j − kg0 (s) + k

Hence

 
λ2 − (d + 1)σ j + k λ + dkσ j σ j /k + 1 − (1 − 1/d) g 0 (s) = 0.

Considering

1 − aw
g0 ( w ) = ,
(aw + 1)3

we know that for s > 1/a there are unstable modes j’s if d is sufficiently small.
1682 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683

Appendix B. Proof of Lemma 2.2

Lemma 2.2 provides a point wise estimate of the gradient of the heat kernel (fundamental solution
of the Laplacian with the Neumann boundary condition) for all t > 0. It seems that this kind of the
estimate is known to the experts who are familiar to the Green function. Unfortunately, there is no
suitable literature for this result. We therefore show how this lemma can be proved in this section.
Since G 1 (x, y , t ) = G 2 (x, y , dt ), we deal with only G 2 , and put G (x, y , t ) = G 2 (t , x, y ) for simplicity
of notation.

Proof of Lemma 2.2. Short time estimate: We consider the interval 0 < t  1. From the process of the
construction of the fundamental solution for the initial boundary value problem of the heat equation
as in [16], we see that there exist constants c 1 > 0, c 2 > 0 such that

 2
 
∇x G (x, y , t )  c 1 exp − |x − y | (0 < t  1, x, y ∈ Ω).
n +1
t 1/2 c2t

Long time estimate: We consider the interval [1, ∞). The following expressions are well known:

∞

G (x, y , t ) = exp(−σk t )Φk (x)Φk ( y ),
k =1
∞
  
∇x G (x, y , t ) = exp(−σk t ) ∇x Φk (x) Φk ( y ),
k =1

where σk and Φk are the k-th eigenvalue and the corresponding eigenfunction of  with the Neu-
mann boundary condition. Here {Φk } is orthonormalized in L 2 (Ω). Note that σ1 = 0 and Φ1 (x) =
1/|Ω|1/2 (constant function), therefore the first term of ∇x G (x, y , t ) vanishes. The convergence of the
above infinite series is assured by the following estimates:

σk  c3k2/n − c4 , (B.1)
n/2 (n/2)+1
Φk L ∞ (Ω)  c 5 k ,
σ ∇Φk L ∞ (Ω)  c 6 σk (k  1), (B.2)

where c 3 , c 4 , c 5 and c 6 are positive constants depending only on Ω . Those inequalities in (B.1)–(B.2)
are verified as follows: The inequality (B.1) follows from the famous Weyl formula for the asymptotic
distribution of the eigenvalues (see [6])

|Ω|  
N (s) := max{k ∈ N | σk < s} = n/2 Γ ((n/2) + 1)
sn/2 + o sn/2 (s → ∞),
2n π

while the first inequality of (B.2) is due to the result in [17]. For the proof of the second inequality
in (B.2) we use the Sobolev inequality (see [1]) and the elliptic estimate (see [2]) which applies to
Φk = −σk Φk in Ω , ∂Φk /∂ ν = 0 on ∂Ω , in the manner

 
∇Φk L ∞ (Ω)  c 7 ∇Φk W 1, p (Ω)  c 8 Φk W 2, p (Ω)  c 9 σk Φk L p (Ω) + Φk L p (Ω)
(n/2)+1
 c 9 (σk + 1)|Ω|1/ p Φk L ∞ (Ω)  c 9 c 5 |Ω|1/ p σk ,

for p > n, where c 7 , c 8 and c 9 are appropriate positive numbers.

 S. Jimbo, Y. Morita / J. Differential Equations 255 (2013) 1657–1683 1683

Using (B.1)–(B.2), we carry out the following estimate:

∞
     
∇x G (x, y , t )  exp(−σk t )∇x Φk (x)Φk ( y )
k =2
∞
  
 exp(−σ2 t ) exp −(σk − σ2 )t ∇x Φk  L ∞ (Ω) Φk  L ∞ (Ω)
k =2
∞
  
 exp(−σ2 t ) exp −(σk − σ2 ) c 5 c 6 σkn+1 (t  1, x, y ∈ Ω).
k =2

n+1 ∞
Due to the estimates (B.1) the infinite series k=2 σk exp(−(σk − σ2 )) converges. Since Ω is
bounded, the above estimate for ∇x G (x, y , t ) is bounded by the right hand side of the inequality
of Lemma 2.2 in t  1 if c 1 , η1 and η2 in the lemma are appropriately chosen.
Combining the short time estimate and the long time estimate, we obtain the desired inequality
of Lemma 2.2. 2

References

[1] R. Adams, J. Fourier, Sobolev Spaces, Academic Press, 2003.

[2] S. Agmon, A. Douglis, L. Nirenberg, Estimates near the boundary of solutions of elliptic partial differential equations satis-
fying general boundary condition, Comm. Pure Appl. Math. 12 (1959) 623–727.
[3] P.W. Bates, P.C. Fife, Spectral comparison principles for the Cahn–Hilliard and phase-field equations, and time scales for
coarsening, Phys. D 43 (1990) 335–348.
[4] G. Caginalp, An analysis of a phase field model of a free boundary, Arch. Ration. Mech. Anal. 92 (1986) 205–245.
[5] X. Chen, Y. Qi, Travelling waves of auto-catalytic chemical reaction of general order—An elliptic approach, J. Differential
Equations 246 (2009) 3038–3057.
[6] R. Courant, D. Hilbert, Method of Mathematical Physics, vol. I, Wiley–Interscience, New York, 1953.
[7] E.B. Davies, Spectral Theory and Differential Operators, Cambridge University Press, 1995.
[8] Y. Hosono, Phase plane analysis of travelling waves for higher order autocatalytic reaction–diffusion systems, Discrete Con-
tin. Dyn. Syst. Ser. B 8 (2007) 115–125.
[9] G.J. Fix, Phase filed methods for free boundary problems, in: A. Fasano, M. Primiceri (Eds.), Free Boundary Problems: Theory
and Applications, Pitman, London, 1983, pp. 580–589.
[10] M.E. Gurtin, H. Matano, On the structure of equilibrium phase transitions within the gradient theory of fluids, Quart. Appl.
Math. 156 (1988) 301–317.
[11] J.K. Hale, Asymptotic Behavior of Dissipative Systems, Math. Surveys Monogr., vol. 25, AMS, 1988.
[12] D. Henry, Geometric Theory of Semilinear Parabolic Equations, Springer, Berlin, New York, 1981.
[13] D. Hilhorst, R. van der Hout, M. Mimura, I. Ohnishi, Fast reaction limits and Liesegang bands, in: Free Boundary Problems,
in: Internat. Ser. Numer. Math., vol. 154, Birkhäuser, Basel, 2007, pp. 241–250.
[14] S. Ishihara, M. Otsuji, A. Mochizuki, Transient and steady state of mass-conserved reaction–diffusion systems, Phys. Rev.
E 75 (2007) 015203(R).
[15] T. Kato, Perturbation Theory for Linear Operators, Springer, 1995.
[16] O. Ladyzenskaya, V. Solonnikov, N. Ural’tseva, Linear and Quasilinear Equation of Parabolic Type, Transl. Math. Monogr.,
vol. 23, AMS, 1968.
[17] P. Li, On the Sobolev constant and the p-spectrum of a compact Riemannian manifold, Ann. Sci. Ec. Norm. Super. 13 (1980)
451–469.
[18] Y. Morita, Spectrum comparison for a conserved reaction–diffusion system with a variational property, J. Appl. Anal. Com-
put. 2 (2012) 57–71.
[19] Y. Morita, T. Ogawa, Stability and bifurcation of nonconstant solutions to a reaction–diffusion system with conservation of
a mass, Nonlinearity 23 (2010) 1387–1411.
[20] M. Otsuji, S. Ishihara, C. Co, K. Kaibuchi, A. Mochizuki, S. Kuroda, A mass conserved reaction–diffusion system captures
properties of cell polarity, PLoS Comput. Biol. 3 (2007) 1040–1054.
[21] A. Pogan, A. Scheel, Instability of spikes in the presence of conservation laws, Z. Angew. Math. Phys. 61 (2010) 979–998.
[22] A. Pogan, A. Scheel, Layers in the presence of conservation laws, J. Dynam. Differential Equations 24 (2012) 249–287.
[23] A. Scheel, Robustness of Liesegang patterns, Nonlinearity 22 (2009) 457–483.
[24] T. Suzuki, S. Tasaki, Stationary Fix–Caginalp equation with non-local term, Nonlinear Anal. 71 (2009) 1329–1349.