Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Applicable Analysis

An International Journal

ISSN: 0003-6811 (Print) 1563-504X (Online) Journal homepage: https://www.tandfonline.com/loi/gapa20

On the backward problem for parabolic equations


with memory

Tran Nhat Luan & Tra Quoc Khanh

To cite this article: Tran Nhat Luan & Tra Quoc Khanh (2019): On the backward problem for
parabolic equations with memory, Applicable Analysis, DOI: 10.1080/00036811.2019.1643013

To link to this article: https://doi.org/10.1080/00036811.2019.1643013

Published online: 23 Jul 2019.

Submit your article to this journal

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=gapa20
APPLICABLE ANALYSIS
https://doi.org/10.1080/00036811.2019.1643013

On the backward problem for parabolic equations with memory


Tran Nhat Luana and Tra Quoc Khanh b,c

a Institute of Computational Science and Technology, Ho Chi Minh City, Vietnam; b Division of Computational
Mathematics and Engineering, Institute for Computational Science, Ton Duc Thang University, Ho Chi Minh City,
Vietnam; c Faculty of Mathematics and Statistics, Ton Duc Thang University, Ho Chi Minh City, Vietnam

ABSTRACT ARTICLE HISTORY


We study the backward problem for the parabolic equation with mem- Received 10 April 2019
ory. We prove that the nonlocal problem is well-posed for 0 < t < T and Accepted 8 July 2019
is ill-posed only at the initial time of t = 0. This result makes a big differ- COMMUNICATED BY
ence between the classical backward parabolic equation and the backward M. Klibanov
parabolic equation with memory. We further propose a spectral regulariza-
tion method to overcome the ill-posedness at the initial time. The Hölder KEYWORDS
convergence rate is obtained under both a priori and a posteriori parameter Backward parabolic
equations; memory effect;
choice rules.
regularization; stability
estimate
CLASSIFICATION AMS
2010
Primary 31B30; 47A52;
65F22; 65J20

1. Introduction
Let  ⊂ Rn , n ≥ 1 be an open bounded domain with a smooth boundary ∂ and T be a posi-
tive number. Put Q =  × (0, T) and  = ∂ × (0, T). Here,  is called the lateral boundary of
the cylinder Q. In the current paper, we consider the question of finding a function u(x, t) where
(x, t) ∈  × [0, T], satisfies the following problem
⎧ ∂u t

⎨ ∂t (x, t) − u (x, t) = β 0 u(x, s) ds, in Q,

(1)

⎪ u(x, t) = 0, in ,

u(x, T) = χ(x), in ,

where χ is given function, β is a positive real number. The first equation in problem (1) is called the
t
parabolic equation with memory where the nonlocal term 0 u(x, s) ds represents the memorial effect
from the past to the present. This type of equation and its variations appear in various fields of engi-
neering such as the modeling the heat conduction in materials with memory [1, 2], the compression
of poro-viscoelastic media [3], the analysis of space-time dependent nuclear reactor dynamics [4],
the compartment model of a double-porosity system, and epidemic phenomena in biology [5].
There are numerous investigations on the parabolic equation with memory terms from various
aspects. For instance, from the aspects of controllability [6–10], the numerical approximation [5, 11,
12], the Carleman estimates [13], etc. In [14, 15], the authors investigated the initial boundary value

CONTACT Tra Quoc Khanh traquockhanh@tdtu.edu.vn

© 2019 Informa UK Limited, trading as Taylor & Francis Group


2 T. NHAT LUAN AND T. QUOC KHANH

problem associated with problem (1), i.e. the following problem


⎧ ∂u t

⎨ ∂t (x, t) − u (x, t) = β 0 u(x, s) ds, in Q,

(2)

⎪ u(x, t) = 0, in ,

u(x, 0) = u0 (x), in ,

It was proved in [14, Lemma 1.4] that if u0 ∈ L2 (), then there is a unique solution u ∈
C([0, T]; L2 ()) ∩ L2 (0, T; H01 ()).
Although there are several results on the initial value problem (2), to the best of our knowledge,
there is not any result devoted to backward parabolic equations with the memory (1). In the case of
β = 0, problem (1) reduces to the classical backward problem for parabolic equations (i.e. the local
backward parabolic problem)
⎧ ∂u

⎨ ∂t (x, t) − u (x, t) = 0,
⎪ in Q,
(3)

⎪ u(x, t) = 0, in ,

u(x, T) = χ(x), in .

It is well-known in the theory of inverse problems that backward problem (3) is ill-posed in the
Hadamards sense for all 0 ≤ t < T. In fact, the literature on regularization methods for these kinds of
problem is a very rich. These methods include the quasi Tikhonov method [16], the quasi-reversibility
method [17–20], the quasi-boundary value method [21–26], the spectral method [27–29], central
difference method [30], the filter method [31, 32], etc. Thus, it is of interest to know whether the
ill-posedness still remains for backward parabolic equations with memory. And if it does, we wonder
whether this ill-posedness can be diminished with regularization techniques. Our major object in
this paper is to provide answers to these questions. The main difficulty in the mathematical analysis
of backward problem (1) is handling the nonlocal memory term. When the nonlocal term disappears
(i.e. β = 0), the first equation in (1) reduces to a linear partial differential equation. Thus, one may
easily employs the above mentioned methods to regularize the problem. However, with the presence
of the nonlocal term, problem (1) becomes an integro-differential problem meaning that it requires
not only the information at a point (or neighborhood of a point) but also at an appropriate time inter-
val, to formulate the problem. Hence, the method applies to the local problem may not directly apply
to the nonlocal problem. In fact, a new approach is required. Motivated by [14], in the present paper
we propose a simple approach to overcome such kinds of difficulties.
The paper is organized as follows. The preliminary section prepares some necessary materials
which will be technically used in the rest of the paper. In Section 3, we present some stability estimates
in regards to the main problem (1). We also compare these results with the corresponding results for
the local backward parabolic problem. Section 4 is devoted to the regularization method to knock
over the ill-posedness mentioned in Section 3. The Hölder convergence rate is obtained in this section
under both a priori and a posteriori regimes. Finally, we end up this paper with a conclusion section
to summarize the achievements as well as to discuss some possible extensions of the problem in the
future.

2. Preliminary
Since  is open and bounded, the eigenvalues of the linear unbounded operator (−), denoted by
{λn }∞
n=1 satisfy

0 < λ1 ≤ λ 2 ≤ λ 3 ≤ · · · , (4)
APPLICABLE ANALYSIS 3

and limn→∞ λn = +∞. In addition, there exists an orthonormal {φn }∞


n=1 of L () such that
2

− φn (x) = λn φn (x), in ,


φn (x) = 0, in ∂, (5)

for n ∈ N (see [33, Theorem 1, p.335]). Throughout this paper, we denote · the standard L2 norm
on  for simplicity. We also use the following usual Sobolev space
 
H p () = v ∈ L2 () : v H p () < ∞ ,

where p > 0 and



2p
v H p () := λn v, φn 2 .
n=1

For a Banach space X we denote by Lp (0, T; X) the Banach spaces of real functions u : [0, T] → X
measurable, such that
 T 1/p
p
u Lp (0,T;X) = u (·,) X dt < ∞, 1 ≤ p < ∞,
0
u L∞ (0,T;X) = ess sup0<t<T u (·,) X < ∞, p = ∞.

Motivated by [14], we propose the following transformation to deal with the nonlocal term
t
0 u(x, s) ds. Looking at the memory
 t
v(x, t) = u(x, s) ds. (6)
0

It is easy to see that v(·, 0) = 0 and for all t ∈ [0, T],


∂v
(x, t) = u(x, t). (7)
∂t
Thus the function v satisfies the following problem
⎧ 2

⎪ ∂ v
⎨ ∂t 2 (x, t) − vt (x, t) = βv(x, t), in Q,

(8)
⎪v(x, t) = 0,
⎪ in ,


v(x, 0) = 0, vt (x, T) = χ(x), in .

By using Fourier’s method, we are looking for a solution of (8) in the following form

v(x, t) = vn (t)φn (x), (9)
n=1

where vn := v(x, t), φn (x) satisfy the following system


⎧ 2
⎪ d vn dvn

⎨ dt 2 (t) + λn dt = βvn (t),

vn (0) = 0, (10)



⎩ dvn (T) = χ .
n
dt
4 T. NHAT LUAN AND T. QUOC KHANH

Denote
 
λ2n + 4β − λn + λ2n + 4β + λn
= μ−
n , μn = . (11)
2 2
By elementary calculations, one may obtain
 − +

eμn t e−μn t
vn (t) = − + χn . (12)
μ+ −μ+ T − −μ− T +
μn e−μn T + μ− −μ− T
n e n + μn e n ne n

Thus, it allows us to write



 − +

eμn t e−μn t
v(x, t) = + − − + − χn φn (x). (13)
n=1 μ+ −μ T − −μ T
n e n + μn e n μ+ −μ T − −μ T
n e n + μn e n

Finally by the relationship (7), we conclude that



 
μ+n +
μn (T−t) μ −
n −
μn (t−T)
u (x, t) =  − e +  + e χn φn (x). (14)
+ − μ+ n +μn T − + − μ− n +μn T
n=1 μ n + μ n e μn + μn e

After this step, we have obtained an analytical solution of backward problem (1). This formulation
is very important in the quantitative analysis of the paper such as the estimate of the degree of ill-
posedness, the convergence estimate. Before doing so, let us mention the following technical lemmas.

Lemma 2.1: Let a,x > 0, then


m!
e−ax ≤ for all m ∈ N.
am xm

Lemma 2.2: Let {λn } be the sequence of eigenvalues of (−) and μ+ −


n , μn be as in (11). The following
statements hold:

(a) For all 0 < t ≤ T, we have

μ−
n μ− (t−T)
supn∈N   e n ≤ 1,
+ − μ− +μ+
μ−
+ μn e
n
n n T

μ+n μ+ (T−t) 2
supn∈N   e n ≤ . (15)
+ − μ+ +μ− βt 2
μn + μn e n n T

(b) For all 0 < t ≤ T and n ∈ N, we have


 − 2
μn μ− (t−T) β
  e n ≤ ,
+ − μ− +μ+ λn
μ−
n + μn e n n T
 + 2
μn μ+ (T−t) 24
  e n ≤ . (16)
+ − μ+ +μ− βt 4 λn
μn + μn e n n T

(c) For all 0 ≤ t ≤ T and n ∈ N, there exists a positive constant C+ := C+ (λ1 , β) depends only on λ1
and β such that

μ+ + μ− μ−
n
 − eμn (T−t) + n
 + e
n (t−T) ≤ C+ λ2n . (17)
μ+ + − μ−
μ+
n + μ−
ne
n +μn T μ−
n + μn e n n+μ T
APPLICABLE ANALYSIS 5

(d) For all n ∈ N, there exists a positive constant C− := C− (λ1 , β, T) depends only on λ1 , β and T
such that

μ+ + μ− −μ− T
n
 − eμn T + 
n
 e n ≥ C− λ2n . (18)
μ+ + − μ− +μ+
μ+
n + μ−
ne
n +μn T μ−
n + μn e n n T

Proof: (a) We first easily see that for all n ∈ N,

μ− μ−
n (t−T)


n
+ e ≤ eμn (t−T) ≤ 1. (19)
+ − μ−
μ−
n + μn e n n+μ T

Next, by applying Lemma 2.1, we may write that

μ+ + μ+ μ+
n
 − eμn (T−t) ≤ 
n
− e
n (T−t)
μ+ − μ+
μ+
n + μ−
ne
n +μn T +μ
μn e n n T
 + 2
μ+
n −μ+ μ +
≤ −e
n t
= n e−μn t
μn β
2
≤ . (20)
βt 2

Since the right hand sides of (19) and (20) are independent of n, we get part (a) of the lemma.
(b) For n ∈ N, one has
 2
μ− − β
n
 + eμn (t−T) ≤ μ−
n ≤ .
μ− λn
μ− + −
n + μn e
n +μn T

Since μ+ −
n × μn = β, we have
 + 2 3

μn μ+ μ+ +
 n (T−t) ≤ n
e−μn t .
− e
μ+ β
μ+ −
n + μn e
n +μn T

In view of Lemma 2.1, one may write that


 2
μ+ + 4! 1 24 1
n
 − eμn (T−t) ≤ + ≤ 4 .
μ+ βt μn βt λn
μ+ − n +μn T
4
n + μn e

The proof of part (b) is complete.


(c) By Lemma 2.1 again, one has
 2
μ+ μ− λ2n + 4β + λn
n μ+ (T−t) n μ− (t−T)
  e n +   e n ≤1+
+ − μ+ +μ− − + − μ− +μ+ 4β
μn + μn e n n T
μn + μn e n n T

1 β 2 2
≤1+ 1+ λn
β λ1
≤ C+ λ2n ,
2
where C+ = 1
λ21
(1 + (β+λ
β
1)
). The proof of part (c) is complete.
6 T. NHAT LUAN AND T. QUOC KHANH

(d) We observe that

μ+ + μ− − μ+ μ+ T
n
 − eμn T + n
 + e−μn T ≥ n
  e n
μ+ μ− − μ+ +μ−
μ+
n + μ−
ne
n +μn T μ−
n + μ+
ne
− n +μn T μ+
n + μn e n n T

λn μ+
n T.
≥  + − e
β
μ+
n + μ+
e μn +μn T
n

+
2eμn T
By using Lemma 2.1 with m = 2, it is true that μ+
n ≤ T 2 μ+
and
n

2β 2β
μ−
n =  2 ≤ .
λn + 4β + λn λ21 + 4β + λ1

Thus, we arrive at the following estimate

μ+ + μ− − T2
n
 − eμn T + n
 + e−μn T ≥ λ2n
μ+ − μ+
n +μn T μ− + − μ−
n +μn T √ 22βT
n + μn e n + μn e
2 + T 2 βe λ1 +4β+λ1
= C− λ2n .

√ 2βT

where C− := T 2 (2 + T 2 βe λ21 +4β+λ1


)−1 . The lemma is completely proved. 

Lemma 2.3: Let {λn } be the sequence of eigenvalues of (−) and μ+ −


n , μn be as in (11). The following
statements hold:

(a) For all 0 ≤ t1 , t2 ≤ T and n ∈ N, there exists a positive constant C1 := C1 (λ1 , β, T) depends only
on λ1 , β and T such that
   
 μ+n (T−t1 ) μ+  −  μ− − 
μ+
n e − e n (T−t2 )  μ n e n (t1 −T) − eμn (t2 −T) 
 − +  + ≤ C1 λ3n |t1 − t2 | . (21)
− μ+ + − μ−
μ+ n + μn e n +μn T μ −
n + μ n e n +μn T

(b) For all 0 < t1 , t2 ≤ T and n ∈ N, there exists a positive constant C2 (t1 , t2 ) := C2 (t1 , t2 , β, T)
depend on t1 , t2 , λ1 , β and T such that
   
 μ+n (T−t1 ) μ+  −  μ− − 
μ+n e − e n (T−t2 )  μ n e n (t1 −T) − eμn (t2 −T) 
C2 (t1 , t2 )
 − +  − + ≤ |t1 − t2 | . (22)
+ − μ+ +μ − + − μ +μ λn
μn + μn e n n T
μn + μn e n n T

Proof: For all 0 ≤ t1 , t2 ≤ T, we first estimate that


   
 μ+n (T−t1 ) +   μ−n (t1 −T) − 
μ+
n e − eμn (T−t2 )  μ− n e − eμn (t2 −T) 
 − +  +
− μ+ + − μ−
μ+n + μn e
n +μn T μ− n + μn e
n +μn T

μ+ μ+nT   μ− −μ−nT  − 
ne  −μ+n t1 −μ+ 
n t2  + ne  μn t1 μ− 
n t2 
=  +  e − e   e − e
+ − μn +μ− − + − μ− +μ+
μn + μn e n T
μn + μn e n n T
 + 2    
μ  + +  −  − − 
≤ n e−μn t1 − e−μn t2  +e−μn T eμn t1 − eμn t2  .
β      
e+ (t1 ,t2 ) e− (t1 ,t2 )
APPLICABLE ANALYSIS 7

If t1 ≤ t2 , then we have
+ + +
 +

e+ (t1 , t2 ) = e−μn t1 − e−μn t2 = e−μn t1 1 − e−μn (t2 −t1 ) ≤ μ+ n |t2 − t1 | ,
− − −
 −

μ−
e− (t1 , t2 ) = eμn t2 − eμn t1 = eμn t2 1 − e−μn (t2 −t1 ) ≤ μ+ne
n T |t − t | .
2 1

In contrast for t1 > t2 , one has


+ + +
 +

e+ (t1 , t2 ) = e−μn t2 − e−μn t1 = e−μn t2 1 − e−μn (t1 −t2 ) ≤ μ+ n |t2 − t1 | ,
− − −
 −

μ−
e− (t1 , t2 ) = eμn t1 − eμn t2 = eμn t1 1 − e−μn (t1 −t2 ) ≤ μ+ne
n T |t − t | .
2 1

Thus we can write that

e+ (t1 , t2 ) ≤ μ+
n |t2 − t1 | ,

e− (t1 , t2 ) ≤ μ+
ne
μn T
|t2 − t1 | . (23)

By using (23), we obtain the following estimate


   
 μ+n (T−t1 ) +   μ−n (t1 −T) −    + 3 
μ+
n e − eμn (T−t2 )  μ−n e − eμn (t2 −T)  μn +
 − +  + ≤ +μn |t1 − t2 |
− μ+ + − μ− β
μ+ n + μn e
n +μn T μ−n + μn e
n +μn T

  2 3
1 1 λn + 4β+λn
≤ + |t1 − t2 |
β λ21 2
 √ 3
1 1 β
≤ + 2 1+ λ3n |t1 − t2 |
β λ1 λ1
= C1 λ3n |t1 − t2 | ,

where
 √ 3
1 1 β
C1 := + 2 1+ .
β λ1 λ1
The proof of part (a) is complete. For part (b) of the theorem, we first note that from Lemma 2.1, for
all positive t1 , t2 such that t1 ≤ t2 ≤ T,

+
 +
 + 24
e+ (t1 , t2 ) = e−μn t1 1 − e−μn (t2 −t1 ) ≤ e−μn t1 μ+
n |t1 − t2 | ≤  3 |t1 − t2 | ,
t14 μ+
n
  −
βeμn T
μ− −μ−
n (t2 −t1 ) μ−
e− (t1 , t2 ) = en t2 1−e ≤enT μ−
n |t1 − t2 | ≤ |t1 − t2 | .
λn

Similarly, for positive t1 , t2 such that t1 > t2 ,

+
 +
 + 24
e+ (t1 , t2 ) = e−μn t2 1 − e−μn (t1 −t2 ) ≤ e−μn t2 μ+
n |t1 − t2 | ≤  3 |t1 − t2 | ,
t24 μ+
n
  −
βeμn T
μ− −μ−
n (t1 −t2 ) μ−
e− (t1 , t2 ) = en t1 1−e ≤enT μ−
n |t1 − t2 | ≤ |t1 − t2 | .
λn
8 T. NHAT LUAN AND T. QUOC KHANH

Thus, one may write that

24
e+ (t1 , t2 ) ≤  + 3 |t1 − t2 | ,
t14 μn

βeμn T
e− (t1 , t2 ) ≤ |t1 − t2 | . (24)
λn

For this reason, the following estimate holds


   
 μ+n (T−t1 ) +   μ−n (t1 −T) −    
μ+
n e − eμn (T−t2 )  μ−
n e − eμn (t2 −T)  24 1 1 β
 − +  + ≤ max , + |t1 − t2 |
μ+ −
n + μn e
μ+
n +μn T μ− + −
n + μn e
μ−
n +μn T βμ+n t14 t24 λn
  
1 24 1 1
≤ max 4 , 4 + β |t1 − t2 |
λn β t1 t2
|t1 − t2 |
= C2 (t1 , t2 ) ,
λn

where the constant C2 is defined by


  
24 1 1
C2 (t1 , t2 ) = max 4 , 4 + β .
β t1 t2

The proof is fully completed. 

3. The stability of solution of problem (1)


Theorem 3.1: Let u be solution of the problem (1) and χ ∈ H q () for q ≥ 0, then the following
statements hold:

(a) For all t ∈ (0, T), we have



2
u (·, t) H q () ≤ 1+ χ H q () ,
βt
  
 du  24
 (·, t) ≤ β+ 4 χ H q () .
 dt  q+1 βt
H ()

In addition, u ∈ C((0, T]; H q ()) ∩ C1 ((0, T); H q ()).


(b) At t = 0, there exists a positive constant C+ := C+ (λ1 , β) depends only on λ1 and β such that

u (·, 0) H q () ≤ C+ χ H q+2 () .

Proof: (a) For all t ∈ (0, T). By (15) and (22), one has
   
 ∞
μ+
+
μn (T−t) μ−

μn (t−T) 
 q ne ne 
u (·, t) H q () = λn  − +  + χn φn (·)
 μ+ −

n=1 μ+ −
n + μn e μ− + − μn +μn T
n +μn T
n + μn e
∞ 
μ+  
n μ+ (T−t)  q 
≤ supn∈N  + −  e n  λn χn φn (·)
+ − μn +μn T
μn + μn e  
n=1
APPLICABLE ANALYSIS 9

 
μ−  ∞ 
n μ−
n (t−T)
 q 
+ supn∈N  − + e  λn χn φn (·)
μ− + − μn +μn
n + μn e
T  
n=1

2
≤ 1+ 2 χ H q () ,
βt
where the last inequality comes from Lemma 2.2. To prove the second estimate, we note that from (16)
and (22), one has
  ∞   2 μ+ (T−t)  − 2 μ− (t−T)  
 du   − μ+ μn e n 
 (·, t)  q+1 n e n 
 dt  q+1 =  λ n   +   χ φ
n n (·) 
 + − μ+ +μ− − + − μ− +μ+ 
H () n=1 μn + μn e n n T
μn + μn e n n T
 ∞ q+1  2 +  ∞ q+1  2 μ− (t−T) 
 λ n μ+ μn (T−t)   λ n μ− 
 n e   n e
n

≤  + −  χn φn (·) +   − +  χn φn (·)
 + − μn +μn T   − + − μn +μn T 
n=1 μn + μn e n=1 μn + μn e

 ∞ 

24  q 
≤ + β  λ n χn φ n (·) 
βt 4  
n=1

24
= + β χ H q () .
βt 4
Now, for a sequence {tn } ⊂ (0, T] such that tn → t as n → ∞. We have
 ⎛  + 
∞ μ+
+
eμn (T−tn ) − eμn (T−t)
 n
limn→∞ u (·, tn ) − u (·, t) H q () = limn→∞  λn ⎝
q
  −
− μ+
n=1 μ+n + μn e
n +μn T

 − ⎞ 
μ− μn (tn −T) − eμ− n (t−T) 
n e 
+  +
⎠ χn φn (·)
− + − μ− +μ
μn + μn e n n T 
   
∞ μ + eμ+ +
n (T−tn ) − eμn (T−t) 
 n 
≤ limn→∞   q
λn  χn φn (·)
+ − μ+ − 
n=1 +μ 
μn + μn e n n T

   
∞ μ− μ− −
n (tn −T) − eμn (t−T) 
 n e 
+ limn→∞   q
λn  χn φn (·)
− + − μ− + 
n=1 +μ 
μn + μn e n n T

∞ 
 
 q−1 
≤ 2 λn χn φn (·) limn→∞ C2 (tn , t) |tn − t| ,
 
n=1

where the last inequality comes from Lemma 2.3. This clearly forces u ∈ C((0, T]; H q ()). In the
same manner, we can see that
 
 du du 
limn→∞   dt (·, tn ) − (·, t) 
 q
dt H ()

∞  + 2  μ+ (T−t ) +
 

μn n − eμn (T−t)
 e n 
≤ limn→∞   q
λn  χn φn (·)
+ − μ+ − 
n=1 +μ 
μn + μn e n n T

 
 − 2 μ− (t −T)  
∞ μn

− eμn (t−T) 
 e n n 
+ limn→∞   λn
q
 χn φn (·)
− + − μ− + 
n=1 +μ 
μn + μn e n n T
10 T. NHAT LUAN AND T. QUOC KHANH

   
√   ∞ μ+ eμ+ + 
n (T−tn ) − eμn (T−t)
β  n 
limn→∞  
q+1
≤ 1+  λn  + −  χ φ
n n (·) 
λ1  μ+ − μn +μn T
n + μn e 
n=1
  −  
√  ∞ μ−

eμn (tn −T) − eμn (t−T) 
β  n 
+ 1+  q+1
λn  − χn φn (·)
limn→∞  + 
λ1 n=1 − +
μn + μn e− μn +μn T 

√  ∞ 
β  
q 
≤2 1+  λn χn φn (·) limn→∞ C (tn , t) |tn − t| .
λ1  
n=1

Observing that for all t > 0, limn→∞ C(tn , t)|tn − t| = t−4 × 0 = 0. This yields
 
 du du 

limn→∞  (·, tn ) − (·, t) = 0.
dt dt  q
H ()

In other words, u ∈ C1 ((0, T); H q ()).


(b) By using Lemma 2.2, it enables us to write
∞   
 μ+ μ+nT μ − e−μ−
nT 
 q−2 n e n 
u (·, 0) H q−2 () =  λn  + −  +  − +  χn φn (·)
 +
μn + μn e− μn +μn T − + − μn +μn T
μn + μn e 
n=1
∞ 
 
 
≤ C+ 
q
λn χn φn (·)
 
n=1
+
=C χ H q () ,

where C+ is the same constant as in (17). The proof of this theorem is complete. 

Theorem 3.2: For all 0 < t < T, the solution of problem (1) depends continuously on the data χ with
respect to the L2 -norm for both u and χ. In contrast, the initial distribution u(x, 0) satisfies (1) does not
depends continuously on the data χ with respect to the L2 -norm for both u(x, 0) and χ.

Proof: Let u1 , u2 are two solutions of the problem (1) corresponding to the data χ1 and χ2 ,
respectively. By Lemma 2.2, it follows that for all 0 < t < T,
∞   
 μ+ μ+n (T−t) μ− μ− n (t−T) 
 ne ne 
u1 (·, t) − u2 (·, t) =   + −  +  − +  (χ1n − χ2n ) φn (·)
 + − μn +μn T − + − μn +μn T 
n=1 μn + μn e μn + μn e
∞ 
 μ+ μ+n (T−t) 
 ne 
≤  + −  (χ1n − χ2n ) φn (·)
 + − μn +μn T 
n=1 μn + μn e
∞ 
 μ− μ−n (t−T) 
 n e 
+  − +  (χ1n − χ2n ) φn (·)
 − + − μn +μn T
μn + μn e 
n=1
 ∞
1
2
2 2
≤2 1+ 2 (χ1n − χ2n )
βt
n=1

2
=2 1+ χ1 − χ2 .
βt 2
APPLICABLE ANALYSIS 11

We get the first conclusion. In contrast, the stability of solution does not hold at t = 0. We will
construct an example to prove this claim. For an orthonormal basis {φn }∞ n=1 of L () satisfy (5),
2

denote

χ1 (x) = φ1 (x) ,
1
χm (x) = φ1 (x) + φm (x) , τ ∈ (0, 2) , m ≥ 2.
λτm

It allows us to write
1
limm→+∞ χm − χ1 = limm→+∞ = 0.
λτm

On the other hand, let um be the solution of problem (1) with respect to the data χm , it is true that
 + −

1 μ+
ne
μn T μ−
ne
−μn T
limm→+∞ um (·, 0) − u1 (·, 0) = limm→+∞ τ  − +  − 
λm μ+ μn +μ+
μ+ −
n + μn e
n +μn T μ− + −
n + μn e
n T

≥ C− limm→+∞ λ2−τ
m = +∞,

where the last inequality comes from part (d) of Lemma 2.2. The theorem is proved. 

Remark: It is noted that when β is positive and t ∈ (0, T), the backward problem is stable in the
sense that the solution depends continuously on the terminal data. This is quite different from the
classical backward parabolic problem (i.e. β = 0) where the instability in the solution happens, even
for the case t ∈ (0, T). This phenomenon can be explained as a result of the memory term in which
one could easily determine the previous status from its present available information. However, the
initial time t = 0 is an outlier. The backward problem at t = 0 is unstable. This is consistent with the
classical backward parabolic problem. It is noted that the situation here is very similar to backward
subdiffusion in the sense that the nonlocality in time stabilizes the originally exponentially ill-posed
problem (see [34, Section 3.1]).

4. The spectral regularization method and its convergence estimate


In physical terms, the final data χ is obtained by measurement which always contains an error. In
fact, this gives us an approximate version χ δ satisfying
 
χ − χ δ  2 ≤ δ, (25)
L ()

where the constant δ ∈ (0, 1) represents the error bound. By the result of Theorem 3.2, we know that
the problem of finding the initial distribution u(x, 0) from the final observation u(x, T) satisfies (1) is
ill-posed in the sense of Hadamard. Thus, there is a demand for regularization to jump over such kind
of ill-posedness. Otherwise, any computation on the solution may lead to an unreliable result. The
regularization method for the problem (1) at t = 0 is the main concern of this section. With this in
mind, we adopt in this section the spectral regularization method which is simple, but very effective
to deal with ill-posed problems. The corresponding convergences estimate are not only performed in
an a priori parameter choice rule, but also in an a posteriori one. For the definition of these parameter
choice rules, we refer the reader to [35, Def. 3.2, p.51]. The key result in this section is that Hölder
convergence rate can be obtained under both regimes.
12 T. NHAT LUAN AND T. QUOC KHANH

Denote PM the orthogonal projection onto the eigenspace span {φn |λn ≤ M}, i.e.
N
PN (w) = (w, φn ) φn for all w ∈ H, (26)
n=1

where
N = N (M) := max { n ∈ N| λn ≤ M} , (27)
will be chosen appropriately later. Let us consider the following approximate problem
⎧ δ
⎪ duP − uδ = β  t uδ (x, s) ds, in Q,


⎨ dt P 0 P
$ (28)
⎪uδP (x, t) = 0,
⎪ in

⎩ δ  
uP (x, T) = PN χ δ (x) , in .

Here and in what follows, we use the notation uδP to represent the solution of the regularized prob-
lem (28) with respect to the measured data χ δ . Applying the same technique as in Section 2, one may
easily obtain the analytical solution of problem (28)
N
 + −

μ+
ne
μn (T−t) μ−
ne
μn (t−T)
uδP (x, t) =  − +  −  χnδ φn (x) , (29)
μ+ μn +μ+
n=1 μ+ −
n + μn e
n +μn T μ− + −
n + μn e
n T

where μ+ − δ δ
n , μn as in (11) and χn is the nth Fourier coefficient of the measured data χ . Occasionally
in this section, we require the following a priori bound holds
 ∞
1
2
2p
u (·, 0) H p () := λn (un (0))2 ≤ E. (30)
n=1

Our first stability result in this section reads as follow.

Theorem 4.1: Let uP and uδP are two solutions of the problem (28) corresponding to data χ and χ δ
respectively. Then we have
 δ   
u (·, 0) − u (·, 0) ≤ C+ M2 χ δ − χ  .
P

Proof: By (17), one has


   
 
 δ  N +
μn e μ +T
n −
μn e −μ −T
n  δ  
u (·, 0) − uP (·, 0) =    +   χn − χn φn (·)
P  + − − + 
n=1 μ+ − μn +μn T
n + μn e μ− + − μn +μn T
n + μn e 
 
N 

+
 δ  
≤C  λn χn − χn φn (·)
2

n=1 
 
≤ C+ M2 χ δ − χ  .

The proof is completed. 

Now, we are ready to present the key results in this section. The first result concerning about the
convergence estimate under an a priori parameter choice of N .
APPLICABLE ANALYSIS 13

4.1. Convergence estimate under an a priori parameter choice rule


In this section, the error estimations will be given for uδP (·, 0) − u(·, 0) under a suitable choice of
regularization parameter.

Theorem 4.2: Let u given by (14) be exact solution of problem (1). Let χ δ satisfies the error bound
assumption (25) and the a priori condition (30) holds for p > 0. Let uδP be regularized solution as in (29).
If the regularization parameter is chosen by
 1
E 2+p
N := max { n ∈ N| λn ≤ M} where M = , (31)
δ
then the following error estimate holds
 δ    2 p
u (·, 0) − u (·, 0) ≤ 1 + C+ E 2+p δ 2+p . (32)
P

Proof: The proof starts with an observation that


 δ   
u (·, 0) − u (·, 0) ≤ uδ (·, 0) − uP (·, 0) + uP (·, 0) − u (·, 0) .
P P

It follows from Theorem 4.1 and (25) that


 δ   
u (·, 0) − uP (·, 0) ≤ C+ M2 χ δ − χ 
P
≤ C+ M2 δ.

On the other hand, the direct computation yields


 
 +∞ 
 
uP (·, 0) − u (·, 0) = 
 un (0) φn (·) 

n=N +1 
 
 +∞ 
1 


λn un (0) φn (·)
p
≤ p 
M  
n=N +1
1
≤ E.
Mp
From the above estimates, it follows that
 δ 
u (·, 0) − u (·, 0) ≤ C+ M2 δ + 1 E
P Mp
  2 p
= 1 + C+ E 2+p δ 2+p ,

where the last equality comes directly from the parameter choice (31). The proof of this theorem is
complete. 

4.2. Convergence estimate under an a posteriori parameter choice rule


This part is devoted to the derivation of the convergence rate of the regularized solution under the a
posteriori parameter choice rule. Denote

 
1 μ+ μ+
n (T−t) μ− μ−
n (t−T)
ne ne
(x, t) =  − +  + χn φn (x) , (33)
λ2 μ+ − μ+ + − μ−
n=1 n n + μn e
n +μn T μ−n + μn e
n +μn T
14 T. NHAT LUAN AND T. QUOC KHANH


 + −

δ 1 μ+
ne
μn (T−t) μ−
ne
μn (t−T)
(x, t) =  − +  + χnδ φn (x) . (34)
λ2n μ+ μ−
n=1 μ+ −
n + μn e
n +μn T μ− + −
n + μn e
n +μn T

Let n (t) and δn (t) be the nth Fourier’s coefficient of (x, t) and δ (x, t) respectively. Using these
notations, we rewrite the exact solution and regularized solution with respect to the measured
data as

u (x, t) = λ2n n (t) φn (x),
n=1
N
δ δ
u P (x, t) = λ2n n (t) φn (x). (35)
n=1

Again, the parameter N will be chosen later. It is noted that the way of chosen N in a posteriori
style is different from the above. Thus, to distinguish it with the previous regularized solution, we
denote uδP the regularized solution in this subsection. Motivated by [35, Section 4.3], we introduce
the discrepancy principle in the following form
    
(I − PN ) δ
(·, 0) ≤ ϑδ <  I − PN −1 δ
(·, 0) , (36)

where and ϑ > C+ and C+ is the positive constant specified in Lemma 2.2. Here, PN is defined as in
(26), i.e.
N
 δ
 δ
 δ

PN (x, 0) = n (t) φn (x), P0 (x, 0) = 0.
n=1

According to the following lemma, we know there exists a unique N for (36) if δ (·, 0) > ϑδ. In
fact, one has:

Lemma 4.1: Set SN = (I − PN −1 ) δ (·, 0) . If δ (·, 0) > ϑδ, then the following results hold:

(a) S1 = δ (·, 0) ;

(b) limN →+∞ SN = 0;


(c) {SN }+∞
N =1 is a decreasing sequence.

Proof: The proofs are straightforwardly obtained by the expression of

 ∞
1
    2 2
 I−P δ
(·, 0) = δ
N −1 n (0) .
n=N

Lemma 4.2: Let the data χ and χ δ satisfy (25). The following estimate holds
 
 δ
(·, 0) − (·, 0) ≤ C+ δ, (37)

where C+ is the positive constant as in Lemma 2.2.


APPLICABLE ANALYSIS 15

Proof: Due to (33), (34), and (17), one has


 δ 
 (·, 0) − (·, 0)
∞   
 μ+ μ+
nT μ− −μ−nT  δ  
 1 ne ne 
=  +  +   χ − χn φn (·) 
 λ2n μ+ − μn +μ− − + − μ− +μ+ n

n + μn e μn + μn e
n T n n T
n=1
∞ 
  δ  
 
≤ C+  χn − χn φn (·)
 
n=1
+
≤ C δ.
The proof is completed. 

Lemma 4.3: Assume that the a priori bound (30) holds and the regularization parameter N is chosen
by the discrepancy principle (36), then the upper bound of N is given by
  1
E 2+p
N ≤   . (38)
ϑ − C+ δ

Proof: We have
   
 ∞   ∞ 
   1 
 n (0) φn (·) =  un (0) φn (·)
   λn
2 
n=N n=N
 ∞ 
1  p


≤ 2+p  λn un (0) φn (·)
N  
n=N
∞ 
1  p


≤ 2+p  λn un (0) φn (·)
N  
n=1
1
= u (·, 0) H p ()
N 2+p
1
≤ 2+p E.
N
On the other hand,
 ∞     
   ∞   ∞   
   δ   δ 
 n (0) φn (·) ≥  (0) φn (·) − n (0) − n (0) φn (·)
   n
  
n=N n=N n=N
 ∞ 
      
 
=  I − PN −1 δ
(·, 0) −  δ
n (0) − n (0) φn (·)
 
n=N
 δ 
≥ ϑδ −  (·, 0) − (·, 0) 
 
≥ ϑ − C+ δ.
Therefore,
  1
E 2+p
N ≤   .
ϑ − C+ δ
The proof is completed. 
16 T. NHAT LUAN AND T. QUOC KHANH

Using the above supporting lemmas, we are now ready to present the convergence estimate under
the a posteriori parameter choice rule.

Theorem 4.3: Let C+ := C+ (λ1 , β) be the constant as in Lemma 2.2 and ϑ > C+ . Assume that
δ (·, 0) > ϑδ and the a priori bound (30) holds. Let the regularization parameter N is chosen by

the discrepancy principle (36). Then we obtain the following error estimates
⎛ ⎞
  +  
 δ  C p 2 p
u P (·, 0) − u (·, 0) ≤ ⎝   2 + C +ϑ
+ 2+p ⎠ E 2+p δ 2+p . (39)
ϑ − C+ 2+p

Proof: By the triangle inequality, we have


     
 δ     .
u P (·, 0) − u (·, 0) ≤ uδP (·, 0) − u P (·, 0) + u P (·, 0) − u (·, 0) (40)

We first give an estimate for the first term. By Theorem 4.1, and Lemma 4.3, we know
   
 δ 
u (·, 0) − u P (·, 0) ≤ C+ N 2 χ δ − χ 
P

C+ 2 p
≤  2 E 2+p δ 2+p . (41)
ϑ − C+ 2+p

Next, we estimate the second term. We observe that


 
 
   ∞ 
u P (·, 0) − u (·, 0) =  un (0) φn (·)
 
n=N +1 
 
 ∞  p   
 un (0) 2+p 2

= λ
p
u (0)
2+p
φ (·) 
λ2n
n n n 
n=N +1 
 
 ∞ 
 p   2 
= ( n (0)) 2+p p
λn un (0)
2+p
φn (·)

n=N +1 
  p   2
 ∞  2+p  ∞  2+p
   
≤   λn un (0) φn (·)
p
 n (0) φn (·)  
n=N +1  n=N +1 
⎛   ⎞ p
 ∞   ∞  2+p
     
≤ ⎝ δ   δ 
n (0) φn (·)⎠
 n (0) φn (·) +  n (0) −
n=N +1  n=N +1 
2
2+p
× u (·, 0) H p ()
⎛  ⎞ p
 ∞  2+p
     
= ⎝(I − PN ) δ
(·, 0) + 


δ
n (0) −

n (0) φn (·)⎠
n=N +1 
2
2+p
× u (·, 0) H p ()
  p 2 p
≤ C+ + ϑ 2+p E 2+p δ 2+p . (42)
APPLICABLE ANALYSIS 17

By combining (40), (41), and (42), we get the conclusion (39). We complete the proof of this
theorem. 

5. Conclusion
In the current paper, we study perhaps for the first time, the backward problem for the parabolic
equation in the presence of constant memory term. We observed that for 0 < t < T, the investigated
problem has some well-posedness property such that the classical backward parabolic problems do
not have. At the initial time t = 0, the investigated problem is ill-posed and hence, we propose a spec-
tral regularization method to overcome this issue. The Hölder convergence rate is obtained under
both a priori and a posteriori parameter choice rules. In the future work, we wish to extend the model
to nonlinear reaction case as well as to study the continuity of the solution with respect to the memory
parameter β.

Disclosure statement
No potential conflict of interest was reported by the authors.

ORCID
Tra Quoc Khanh http://orcid.org/0000-0001-8643-777X

References
[1] Nunziato JW. On heat conduction in materials with memory. Quart Appl Math. 1971;29:187–204.
[2] Yanik EG, Fairweather G. Finite element methods for parabolic and hyperbolic partial integro-differential
equations. Nonlinear Anal. 1988;12(8):785–809.
[3] Habetler GJ, Schiffman RL. A finite difference method for analyzing the compression of poro-viscoelastic media.
Computing (Arch Elektron Rechnen). 1970;6:342–348.
[4] Pachpatte BG. On a nonlinear diffusion system arising in reactor dynamics. J Math Anal Appl. 1983;94(2):501–508.
[5] Zhang NY. On fully discrete Galerkin approximations for partial integro-differential equations of parabolic type.
Math. Comp. 1993;60(201):133–166.
[6] Guerrero S, Imanuvilov OY. Remarks on non controllability of the heat equation with memory. ESAIM Control
Optim Calc Var. 2013;19(1):288–300.
[7] Tao Q, Gao H. On the null controllability of heat equation with memory. J Math Anal Appl. 2016;440(1):1–13.
[8] Tao Q, Gao H, Zhang B. Approximate controllability of a parabolic equation with memory. Nonlinear Anal Hybrid
Syst. 2012;6(2):839–845.
[9] Tao Q, Gao H, Zhang B, Yao Z. Approximate controllability of a parabolic integrodifferential equation. Math
Methods Appl Sci. 2014;37(15):2236–2244.
[10] Zhou X, Gao H. Interior approximate and null controllability of the heat equation with memory. Comput Math
Appl. 2014;67(3):602–613.
[11] Larsson S, Thomée V, Wahlbin LB. Numerical solution of parabolic integro-differential equations by the discon-
tinuous Galerkin method. Math Comp. 1998;67(221):45–71.
[12] Li X, Xu Q, Zhu A. Weak galerkin mixed finite element methods for parabolic equations with memory. Discrete
Cont Dyn Syst Ser S. 2019;12(3):513–531.
[13] Loreti P, Sforza D, Yamamoto M. Carleman estimates for integro-differential parabolic equations with singular
memory kernels. J Elliptic Parabol Equ. 2017;3(1–2):53–64.
[14] Li L, Zhou X, Gao H. The stability and exponential stabilization of the heat equation with memory. J Math Anal
Appl. 2018;466(1):199–214.
[15] Sakthivel K, Balachandran K, Nagaraj BR. On a class of non-linear parabolic control systems with memory effects.
Internat J Control. 2008;81(5):764–777.
[16] Cheng J, Liu JJ. A quasi Tikhonov regularization for a two-dimensional backward heat problem by a fundamental
solution. Inverse Probl. 2008;24(6):065012, 18.
[17] Jana A, Thamban Nair M. Quasi-reversibility method for an ill-posed nonhomogeneous parabolic problem.
Numer Funct Anal Optim. 2016;37(12):1529–1550.
[18] Latt R, Lions JL. Methode de quasi-reversibility et applications. Paris: Dunod; 1967.
[19] Miller K. Stabilized quasi-reversibility and other nearly-best-possible methods for non-well-posed problems.
Vol. 316, Lecture Notes in Math.; 1973. p. 161–176.
18 T. NHAT LUAN AND T. QUOC KHANH

[20] Showalter RE. Quasi-reversibility of first and second order parabolic evolution equations. Res. Notes in Math.,
No. 1; 1975. p. 76–84.
[21] Hào DN, Van Duc N. A non-local boundary value problem method for semi-linear parabolic equations backward
in time. Appl Anal. 2015;94(3):446–463.
[22] Hào DN, Van Duc N, Lesnic D. Regularization of parabolic equations backward in time by a non-local boundary
value problem method. IMA J Appl Math. 2010;75(2):291–315.
[23] Hào DN, Van Duc N, Sahli H. A non-local boundary value problem method for parabolic equations backward in
time. J Math Anal Appl. 2008;345(2):805–815.
[24] Khanh TQ, Van Hoa N. On the axisymmetric backward heat equation with non-zero right hand side: regulariza-
tion and error estimates. J Comput Appl Math. 2018;335:156–167.
[25] Showalter RE. The final value problem for evolution equations. J Math Anal Appl. 1974;47:563–572.
[26] Trong DD, Duy BT, Minh MN. Backward heat equations with locally Lipschitz source. Appl Anal.
2015;94(10):2023–2036.
[27] Jana A, Thamban Nair M. Truncated spectral regularization for an ill-posed nonhomogeneous parabolic problem.
J Math Anal Appl. 2016;438(1):351–372.
[28] Nam PT. An approximate solution for nonlinear backward parabolic equations. J Math Anal Appl.
2010;367(2):337–349.
[29] Zhang Y-X, Fu C-L, Ma Y-J. An a posteriori parameter choice rule for the truncation regularization method for
solving backward parabolic problems. J Comput Appl Math. 2014;255:150–160.
[30] Xiong X-T, Fu C-L, Qian Z. Two numerical methods for solving a backward heat conduction problem. Appl Math
Comput. 2006;179(1):370–377.
[31] Khanh TQ, Khieu TT, Van Hoa N. Stabilizing the nonlinear spherically symmetric backward heat equation via
two-parameter regularization method. J Fixed Point Theory Appl. 2017;19(4):2461–2481.
[32] Qin H-H, Wei T. Some filter regularization methods for a backward heat conduction problem. Appl Math Comput.
2011;217(24):10317–10327.
[33] Evans LC. Partial differential equations. Providence (RI): American Mathematical Society; 1998.
[34] Jin B, Rundell W. A tutorial on inverse problems for anomalous diffusion processes. Inverse Probl.
2015;31(3):035003, 40.
[35] Engl HW, Hanke M, Neubauer A. Regularization of inverse problems. Berlin: Springer Science & Business Media;
1996.

You might also like