Journal of Computational and Applied Mathematics 235 (2011) 30793085

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

Shannon wavelet regularization methods for a backward heat equation

Jin-Ru Wang
Department of Applied Mathematics, Beijing University of Technology, Pingle Yuan 100, Beijing 100124, PR China

article

abstract

info

Article history:
Received 8 June 2009
Received in revised form 26 November
2010
MSC:
42C40

We consider the backward heat equation uxx (x, t ) = ut (x, t ), < x < , 0 t < T .
The solution u(x, t ) on the final value t = T is an known function gT (x). This is a typical
ill-posed problem, since the solution if it exists does not depend continuously on the
final data. In this paper, we shall give a Shannon wavelet regularization method and obtain
some quite sharp error estimates between the exact solution and the approximate solution
defined in the scaling space Vj .
2011 Elsevier B.V. All rights reserved.

Keywords:
Backward heat equation
Ill-posed problem
Regularization error estimates

1. Introduction
The backward heat conduction problem (BHCP) arises in the modeling of heat propagation in thermophysics and the
mechanics of continuous media. The determination of the unknown initial temperature from observable scattered final
temperature data is required in many real applications. The BHCP is typically ill posed in the sense of Hadamard [1], which
means that any small error in the collected data may induce enormous error in the solution. In general, the solution of
the BHCP does not exist, and even if it exists it is not continuously dependent on the final data; see Payne [2]. Uniqueness
conditions for the BHCP have been investigated in [3].
In the context of approximation methods for this problem, many approaches have been investigated. [4,5] have
approximated the BHCP by Fourier truncated methods. Numerical methods for the ill-posed BHCP have been given in [68].
A mollification method has been studied in [9]. Recently, wavelet methods have been developed to solve inverse parabolic
partial differential equations (PDEs) [1012]. It was demonstrated that wavelets methods can solve ill-posed PDEs efficiently.
However, to our knowledge, the theoretical results concerning the error estimates were unsatisfactory, i.e., the results were
not optimal in the sense of [13]. In this paper, we shall define a wavelet solution based on the Shannon wavelet and give a rule
for choosing an appropriate multi-resolution space Vj to get the optimal error estimates [13]. In particular, when the time
variable t = 0, and the Sobolev space index s > 0 or s = 0, the error estimates are also sharp. Therefore, we cannot expect
to find a numerical method for the approximating solution that satisfies a better error estimate. Furthermore, the results
available in the literature are mainly devoted to the case of constant coefficients, but the method can carry over to more
interesting problems, for example with variable coefficients and in more dimensions. In fact, we only need to generalize
Theorem 3.1 and use tensor wavelets in more dimensions. In this sense, we think we get better results.
We consider the following one-dimensional backward heat equation in an unbounded region:

uxx (x, t ) = ut (x, t ),

u(x, T ) = gT (x),

< x < , 0 t < T ,

< x < ,

E-mail address: wangjinru@bjut.edu.cn.

0377-0427/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2011.01.001

(1.1)

3080

J.-R. Wang / Journal of Computational and Applied Mathematics 235 (2011) 30793085

where we want to determine the temperature distribution u(x, t ) and assume that, for the exact data gT (x), the solution u
exists and satisfies the following a priori smoothness condition:

u(, 0)s E ,

s 0,

(1.2)

where s denotes the norm in Sobolev space H s (R) defined by

(1 + 2 )s |
f ( )|2 d

f s :=

21

In practice, due to the measurement errors, the solution has to be reconstructed from perturbed data gT , assumed to
satisfy

gT gT L2 .

(1.3)

Using the Fourier transform with respect to the variable x, we can get the Fourier transform of the solution u(x, t ) of
problem (1.1):
2
u ( , t ) = gT ( )e (T t ) .

Since u ( , t ) = gT ( )e (T t ) tends to infinity as | | , so the existence of a solution in L2 (R) depends on a rapid

decay of gT ( ) at high frequencies. However, in practice, gT ( ) is the measured data; it need not decay rapidly at the high
frequencies, and therefore the solution may not exist in L2 (R). Furthermore, if it exists, it does not depend continuously on
the final condition, i.e., (1.1) is ill posed.
The structure of the paper is as follows: in Section 2, we introduce some useful properties; in Section 3, we define a stable
Shannon wavelet solution; in Section 4, we obtain our main results.
2

2. Shannon scaling function and wavelet function

The Shannon scaling function is (x) =

1
) = 2 ,
(
0,

sin x
,
x

and its Fourier transform is given by

| | ,
otherwise.

The corresponding wavelet function (x) is given by

1 i 2
) = 2 e ,
(
0,

| | 2 ,
otherwise.

J
2

Let J ,k (x) = 2 (2J x k), and let the orthogonal projection of L2 (R) onto VJ

= span{J ,k ()}kZ , and V J =

span{ J ,k ()}kZ be denoted by PJ and PJ respectively; i.e., f () L2 (R),

PJ f (x) =

f , Jk Jk (x),

P J f ( ) =

f , Jk Jk ( ).
kZ

kZ

It is easy to see that P

J f ( ) = PJ f ( ), and we have the following properties.
(1) For any k Z,
supp J ,k ( ) = { : | | 2J },

J ,k ( ) = { : 2J | | 2J +1 };
supp
i.e., PJ can be considered a low-pass filter.
(2) f (x) L2 (R),
f (x) =

f , J ,k J ,k (x) +
f , l,k l,k (x).
kZ

lJ kZ

(3) For any J Z,

P
J f ( ) = f ( ),

| | 2J .

3. Regularization method
Let the infinite-dimensional matrix DJ be given by

{(DJ )l,k }lZ,kZ = {J,l (x), J ,k (x)}lZ,kZ ,

where (x) is the Shannon scaling function, J ,k (x) = 2J /2 (2J x k); then we have the following.

J.-R. Wang / Journal of Computational and Applied Mathematics 235 (2011) 30793085

3081

Theorem 3.1. (a) (DJ )l,k = (DJ )k,l ;

(b) DJ 2J ;
(c) let f () be a real continuous function on [ 2J , 2J ], and let {E }R denote the eigenprojection family of the Hermitian

2J

operator DJ . Define r (DJ ) =: 2J r ()dE . Then we have

r (DJ )

max
2J 2J

|r ()|.

Proof. The proof is very similar to that in [10].

Now, we have defined uJ (x, t ) VJ to be the Shannon wavelet solution of problem (1.1) for gT (x); i.e.,
uJ (x, t ) =:

ak (t )J ,k (x),

(3.1)

kZ

2J

where the infinite-dimensional vector a(t ) =: T 2J e(T t ) dE , {E }R denotes the eigenprojection family of the Hermitian operator DJ , and T = {T (k)}kZ =: {gT (x), J ,k (x)}kZ . Then we get the following.
Theorem 3.2. The infinite-dimensional vector a(t ) is the solution of the following system:

a (t ) = DJ a,
a(T ) = T .

0 t < T,

(3.2)

Proof. (i) First, we prove that

2J

a (t ) = T

e(T t ) dE .

2J

In fact,

J [

]
2J

2
e(T (t +t )) e(T t )

(T t )
dE T
e
dE
T

t
2J
2J

J [
]

2
e(T (t +t )) e(T t )

e(T t ) dE
= T

t
2

J [
]

2 e(T (t +t )) e(T t )

e(T t ) dE
T

2J
t
(T (t +t ))

e(T t )
T max
e(T t ) 0 (t 0).
t
|| 2j
So
2J

a (t ) = T

e(T t ) dE .

2J

On the other hand, by using the spectral theorem [14], it is easy to see that
2J

DJ a =

2J

= T

dE

2J

2J

2J
2J

e(T t ) dE

e(T t ) dE .

Therefore a (t ) = DJ a.
(ii) By the definition of the spectral family, we have
a(T ) = T

2J

(T T )

2J

dE = T

2J
2J

dE = T . 

Let uJ (x, t ) be the Shannon wavelet solution for the measured data gT (x), i.e.,
uJ (x, t ) =:

kZ

ak (t )J ,k (x),

(3.3)

3082

J.-R. Wang / Journal of Computational and Applied Mathematics 235 (2011) 30793085

2J

where the infinite-dimensional vector a (x) =: T 2J e(T t ) dE , {E }R denotes the eigenprojection family of the Hermitian operator DJ , and T = { (k)}kZ =: {gT (x), J ,k (x)}kZ . The following theorem indicates that the Shannon wavelet
solution which we have defined has stability.
Theorem 3.3. If gT (x) gT (x)L2 and

J = log2

1
T
ln

ln

2Ts 21

+ 1,

where [c ] denotes the largest integer less than or equal to c R, then we have the following estimate:
t

uJ (x, t ) uJ (x, t )L2 E 1 T T

1
T

ln

s(T2Tt )

t [0, T ),

where s 0 is a constant.
Proof. Since uJ (x, t ) and uJ (x, t ) are given by Eqs. (3.1) and (3.3), then from Plancherel formula and the spectral theorem
[14], we have

uJ (x, t ) uJ (x, t )L2

2J

(T t )
= a (t ) a(t )l2 = (T T )
e
dE

2
2J
l
J

e(T t ) dE l2 T T

2J
2

max |e(T t ) |PJ gT (t ) PJ gT (t )l2

|| 2J

max |e(T t ) |
|| 2J

e 2 (T t ) .
J

Using

J = log2

1
T
ln

ln

2Ts 21

s(T2Tt )

+ 1,

we get

uJ (x, t ) uJ (x, t )L2 E

1 Tt

t
T

1
T

ln

t [0, T ).

4. Main results
In this section, the main results will be given. First, we give the following useful lemma.
Lemma 4.1. Let u(x, t ) be the exact solution of problem (1.1); then PJ u(x, t ) satisfies the following equation:

(PJ u)t = PJ (PJ u)xx ,

PJ u(x, T ) = PJ gT (x),

< x < , 0 t < T ,

< x < .

Proof. Since u(x, t ) satisfies the following equation:

ut (x, t ) = uxx (x, t ),

u(x, T ) = gT (x),

< x < , 0 t < T ,

< x < ,

it is easy to see that PJ u(x, t ) satisfies

(PJ u)t = PJ (PJ u)xx + PJ [(I PJ )u]xx ,

PJ u(x, T ) = PJ gT (x),

< x < , 0 t < T ,

< x < .

J.-R. Wang / Journal of Computational and Applied Mathematics 235 (2011) 30793085

3083

( , t ) for | | 2J , and supp J ,k ( ) = { : | | 2J }, we know that

From the properties P
J u( , t ) = u
2
PJ [(I PJ )u]xx 2L2 = PJ (i )2 (u P
J u)L2
2
= PJ (i )2 (u P
J u)

| | 2J

2
+ PJ (i )2 (u P
J u)| |> 2J

= 0.
So PJ u(x, t ) satisfies the following equation:

(PJ u)t = PJ (PJ u)xx ,

PJ u(x, T ) = PJ gT (x),

Let PJ u(x, t ) =:

kZ

b (t ) = DJ b,
b(T ) = T ,

< x < , 0 t < T ,

< x < . 

bk (t )J ,k (x); then we get the following discrete equation:

0 t < T,

(4.1)

where the infinity vector

b(t ) = {bk (t )}kZ =: {u(x, t ), J ,k (x)}kZ ,

T = {T (k)}kZ =: {gT (x), J ,k (x)}kZ .

In this paper, we are interested in the L2 -norm estimate of the distance between the Shannon wavelet solution uJ (x, t ) for

the measured data gT (x) and the unknown solution u(x, t ) for the exact data gT (x). Let uJ (x, t ) denote the Shannon wavelet
solution for the exact data gT (x); then we have

u(x, ) uJ (x, )L2 u(, t ) PJ u(, t )L2 + PJ u(, t ) uJ (, t )L2 + uJ (, t ) uJ (, t )L2 .

Therefore it remains to obtain the last two estimates on the right-hand side.
Theorem 4.2. Let u(x, t ) be the exact solution of problem (1.1); then, for any J Z, PJ u(x, t ) = uJ (x, t ).
Proof. Define J (x, t ) =: PJ u(x, t ) uJ (x, t ); then

J (x, t ) =

bk (t )Jk (x)

ak (t )Jk (x)

kZ

kZ

[bk (t ) ak (x)]Jk (x)

kZ

k (t )Jk (x),

kZ

where k (x) = bk (x) ak (x), k Z.

By (3.2) and (4.1), we obtain that the infinite-dimensional vector (x) = {k (x)}kZ satisfies

(t ) = DJ ,
(T ) = 0.

0 t < T,

2J
As in the proof of Theorem 3.2, we have (t ) = T 2J e(T t ) dE and T = 0; hence = 0; i.e., PJ u = uJ .

Theorem 4.3. Let u(x, t ) be the exact solution of problem (1.1), let u(x, 0)H s E hold, and let

J = log2

1
T
ln

ln

2Ts 12

+ 1,

where [c ] denotes the largest integer less than or equal to c R. Then, we have
t
T

u(x, t ) PJ u(x, t ) E
where s 0 is a constant.

1 Tt

1
T

ln

s(T2Tt )

T ln E 2

,
2 ln s ln T1 ln E

t [0, T ),

3084

J.-R. Wang / Journal of Computational and Applied Mathematics 235 (2011) 30793085

( , t ) for | | 2J , due to the Parseval formula, we have

Proof. Since PJ is an orthogonal projection and P
J u( , t ) = u
2
u(x, t ) PJ u(x, t )2L2 = u( , t ) P
J u( , t )L2

= (I PJ )J+ ( )u( , t )2L2

J+ ( )u( , t )2L2 ,
where j+ ( ) denotes the characteristic function of the interval [ 2J , +). Note that the Fourier transform of the exact
2
solution u(x, t ) has the form gT ( )e (T t ) . Then the above inequality reduces to

u(x, t ) PJ u(x, t )L2

| | 2J

|gT ( )e

2 (T t ) 2

| e

2 (T t )

| d

2T

| | 2J

u ( , 0)|2 d

|et u ( , 0)|2 d
2

=
| | 2J

e t

max

(1 + 2 )

| | 2J

max

| | 2J

12

12

|u( , 0)|2 (1 + 2 )s d

12

t 2

| |s

| | 2J

s
2

12

2Js et ( 2 ) E ,
J 2

and using

J = log2

1
T
ln

ln

2Ts 21

+ 1,

we get

Tt

t ( 2J )2

2J

ln

2Tst

E E

1 Tt

1
T

ln

2Tst

4 T
2 ln E s ln

t
T

1
T

ln E

Therefore
t

u(x, t ) PJ u(x, t ) T E 1 T

1
T

ln

s(T2Tt )

T ln E 2

E
,
2 ln s ln T1 ln E

t [0, T ). 

By Theorems 3.3, 4.2 and 4.3, the following main result is immediately obtained.
Theorem 4.4. If gT (x) gT (x)L2 , u(x, 0)H s E hold, and

J = log2

1
T
ln

ln

2Ts 21

+ 1,

where [c ] denotes the largest integer less than or equal to c R, then we have

u(x, t ) uJ (x, t )L2 E

where s 0 is a constant.

1 Tt

t
T

1
T

ln

s(T2Tt )

T ln E 2

+ 1,
2 ln s ln T1 ln E

Remark 4.5. The error estimate (4.2) has optimal order [9].

t [0, T ),

(4.2)

J.-R. Wang / Journal of Computational and Applied Mathematics 235 (2011) 30793085

3085

Remark 4.6. When t = 0, s > 0, (4.2) becomes

u(x, t ) uJ (x, t )L2 E

1
T

ln

2s

(1 + o(1)) ,

( 0).

It is a sharp convergent estimate.

t

Remark 4.7. When s = 0, (4.2) becomes u(x, t ) uJ (x, t )L2 2E 1 T T .

Acknowledgements
The project was supported by Natural Science Foundation of Beijing (No. 1092003), National Natural Science Foundation
of China (No. 10871012), and Beijing Educational Committee Foundation (No. 00600054R1002). The author is grateful to
the anonymous referees for their valuable comments, leading to the improvement of this paper.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]

J. Hadamard, Lectures on Cauchy Problems in Linear Partial Differential Equations, Yale University Press, New Haven, CT, 1923.
L.E. Payne, Improperly posed problems in partial differential equations, Philadelphia, Society for Industrial and Applied Mathematics (1975).
W.L. Miranker, A well posed problem for the backward heat equation, Proc. Amer. Math. Soc. 12 (1961) 243247.
P.T. Nama, D.D. Trong, N.H. Tuan, The truncation method for a two-dimensional nonhomogeneous backward heat problem, Appl. Math. Comput. 216
(2010) 34233432.
C.L. Fu, X.T. Xiong, Z. Qian, Fourier regularization for a backward heat equation, J. Math. Anal. Appl. 331 (2007) 472480.
M. Li, T.S. Jiang, Y.C. Hon, A meshless method based on RBFs method for nonhomogeneous backward heat conduction problem, Eng. Anal. Bound.
Elem. 34 (2010) 785792.
X.T. Xiong, C.L. Fu, Z. Qian, Two numerical methods for solving a backward heat conduction problem, Appl. Math. Comput. 179 (2006) 370377.
Y.C. Hon, M. Li, A discrepancy principle for the source points location in using the MFS for solving the BHCP, Int. J. Comput. Methods 6 (2008) 181197.
Z. Qian, C.L. Fu, R. Shi, A modified method for a backward heat conduction problem, Appl. Math. Comput. 185 (2007) 564573.
T. Regiska, L. Eldn, Stability and convergence of a wavelet-Galerkin method for the sideways heat equation, J. Inverse Ill-Posed Probl. 8 (2000) 3149.
J.R. Linhares de Mattos, E.P. Lopes, A wavelet Galerkin method applied to partial differential equation with variable coefficients, Electron. J. Differential
Equations 10 (2003) 211225.
L. Eldn, F. Berntsson, T. Regiska, Wavelet and fourier methods for solving the sideways heat equation, SIAM J. Sci. Comput. 21 (2000) 21782205.
U. Tautenhahn, Optimality for ill-posed problems under general source conditions, Numer. Funct. Anal. Optim. 19 (1998) 377398.
E. Kreyszig, Introductory functional analysis with applications, High Educational Press, 1986 (in Chinese).