See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/366722812

Reconstruction of spatial source term for space-fractional diffusion equation

from final observation data

Preprint · December 2022

DOI: 10.13140/RG.2.2.25341.44003

CITATIONS

4 authors:

Hrizi Mourad Mohamed Bensalah

Université de Kairouan University of Monastir
16 PUBLICATIONS 38 CITATIONS 8 PUBLICATIONS 5 CITATIONS

SEE PROFILE SEE PROFILE

Mohammed Sabah Hussein Hamed Ould Sidi

University of Baghdad University Hassan 1st Settat Morocco
30 PUBLICATIONS 230 CITATIONS 7 PUBLICATIONS 7 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Solving two dimensional inverse heat source problem with iterative procedure View project

inverse problem for partial differential equations involving fractional operators View project

All content following this page was uploaded by Hrizi Mourad on 31 December 2022.

The user has requested enhancement of the downloaded file.

 RECONSTRUCTION OF SPATIAL SOURCE TERM FOR
SPACE-FRACTIONAL DIFFUSION EQUATION FROM FINAL
OBSERVATION DATA

M. HRIZI, M. BENSALAH, M.S. HUSSEIN, AND H. OULD SIDI

Abstract. Reconstructing the unknown source function is to identify the nature of

diffusion process and that is one of the most important and well-studied problems in
many branches of engineering sciences. In the present work, we study the inverse source
problem of reconstructing of the spatial component in the source term of a diffusion
equation with space-fractional Laplacian (in a bounded domain) from a final observation
data. From this extra data we ensure the unique solvability for under investigation ill-
posed problem. The inverse problem viewed as nonlinear optimization problem. Since
the problem is ill-posed that mean it is sensitive for inclusion of noise in the input
data. In order to recover this instabilities a sort of regularization/stabilization should be
employed. We use Tikhonov’s regularization method which is based on adding Tikhonov
regularization term to naive cost functional.
In particular, the least-squares cost functional is defined to measure the misfit of the
numerically obtained solution from the model (space-fractional diffusion equation) and
the data taken from the final time measurements. It is proved that the Tikhonov’s
functional is Fréchet differentiable and a formula for the gradient is derived via an ad-
joint problem approach. Then the conjugate gradient method combined with Morozov’s
discrepancy principle as stopping criteria are exploited for building an iterative recon-
struction process. Several numerical examples for one- and two-dimensional case are
presented and discussed. The numerical finding showing the efficiency and accuracy of
the proposed method.

1. Introduction and problem setting

Fractional order operators have recently emerged as a modeling alternative in different

fields of sciences and engineering thanks to their ability to describe effects that classical
operators fail to capture. Particularly, nonlocal diffusion operators such as the fractional
Laplacian have emerged as an excellent alternative to model diffusion. Under a proba-
bilistic framework, this operator can be derived as the limit of the so-called long jump
random walk [55]. More applications of these models appear in geophysics [59], anomalous
transport and diffusion [10, 45], image denoising and phase-field modeling [2, 4], porous
media flow [56], data analysis [3], elasticity [19], and population dynamics [57].

The space-fractional diffusion equation (SFDE) arises by replacing the standard space
partial derivative in the diffusion equation with a space-fractional partial derivative (in
particular fractional Laplacian). This fractional model can be used to describe the anoma-
lous diffusion which corresponds to the fractional Brownian motion. In the past few years,
Date: December 31, 2022.
Key words and phrases. Inverse source problem, fractional diffusion equation, nonlocal operator,
Tikhonov regularization, finite element method, conjugate gradient method.
1
2

the direct problems, i.e. initial value problem and initial boundary value problem for
SFDE have been investigated thoroughly from both theoretical and numerical aspects.
On the other hand, inverse problems for SFDEs are a rather new research topic and there
are already a lot of studies [1, 5, 17, 52, 61].

In this paper, we discuss the inverse problem of reconstructing an unknown source in

a SFDE. Generally, reconstructing the unknown source is to obtain information about
the nature of a physical object or phenomena with observational data, and that is one of
the most important and well-studied problems in many branches of engineering sciences,
geophysical prospecting, electromagnetic theory, and pollutant detection. In order to
present the considered inverse source problem, we first introduce the governing equation
used. Let T > 0 and let Ω ⊂ Rd (d ≥ 1) be a bounded open set with boundary ∂Ω. For
0 < s < 1, by (−∆)s , we denote the integral fractional Laplacian of order s defined by
(see e.g. [9, 11, 13, 14])
h(x) − h(y)
Z
s
(−∆) h(x) := Cd,s P.V. d+2s
dy, (1.1)
Rd |x − y|
where the symbol P.V. denotes the principal value of the integral; i.e.
h(x) − h(y) h(x) − h(y)
Z Z
P.V. d+2s
dy = lim dy, (1.2)
Rd |x − y| →0 Rd \B (x) |x − y|d+2s


with B (x) is a ball of radius  centered at x. In addition, the normalized constant Cd,s is
given by
4s Γ π2 + s


Cd,s = d/2 , (1.3)
π |Γ(−s)|
where Γ denotes the Euler’s Gamma function. Note that when s approaches 1− the frac-
tional Laplacian (−∆)s converge to the classical Laplacian −∆ (see e.g. [47, Proposition
4.4]).

Consider the inverse source problem of finding a source term F ∗ for the SFDE

 ∂t u + (−∆)s u = F ∗ in Ω × (0, T ],
u = 0 in (Rd \Ω) × (0, T ], (1.4)
u(., 0) = 0 in Ω,


from the final observation data u(., T ) in the domain Ω.

The major difficulty of this inverse source problem concerns the unidentifiability of
general sources’ terms. In this paper, to overcome this difficulty, we propose to write the
source term by means of the separation of variables. More precisely, the source term F ∗
in the following variable separation form
F ∗ (x, t) = f ∗ (x) g(t), (x, t) ∈ Ω × (0, T ). (1.5)
where the component f ∗ models the unknown spatial counterpart whereas g is a given
temporal one that describes the time evolution pattern. Hence, the inverse source problem
we are concerned with is formulated as follows: given an observation data ϑ ∈ L2 (Ω),
recover the spatial variable f ∗ such that
ϑ = u(., T ) in Ω. (1.6)
For the sake of completeness, we briefly recall the physical meaning of the source term
F ∗ (x, t) = f ∗ (x) g(t). When a substance is a conductor, if we have an external electric
 3

field then the work of a current will increase the internal energy of the conductor and
heat it up. For this, let the conductor be modeled by the cylinder C defined by
n o
3 2 2 2
C = x = (x1 , x2 , x3 ) ∈ R : x2 + x3 ≤ r , (1.7)
where r > 0 and we denotes by (x1 , x2 , x3 ) the electric conductivity of the substance
at the position (x1 , x2 , x3 ). On the other hand, according to the Joule-Lenz law and the
Ohm law [50, Chapter 5, page 112], the source F ∗ can be the amount of heat liberated in
a unit volume of the substance per unit time which has the form
F ∗ (x1 , x2 , x3 , t) = (x1 , x2 , x3 )E 2 (x1 , x2 , x3 , t), (1.8)
where E is the external electric field. In addition, if we design the electric field E = σ0 ν(t)e,
where e = (1, 0, 0) and ν is known temporal function. From this, one can deduce that the
source F ∗ can be written as
F ∗ (x1 , x2 , x3 , t) = σ02 (x1 , x2 , x3 )g(t) with g(t) = ν 2 (t). (1.9)
In the particular case when we take a small r then we can approximate the electric
conductivity (x1 , x2 , x3 ) by (x1 , 0, 0) and we can write the source F ∗ as
F ∗ (x, t) ≈ f ∗ (x) g(t) with x = (x1 , x2 , x3 ) ∈ C, t ∈ (0, T ), and f ∗ (x) = σ02 (x1 , 0, 0),
which is our investigated model. For more information about the physical meaning of the
source (1.5), reader can refer to [54].

Recently, Trong et.al [54], discussed the same inverse problem of finding the space-
dependent source f ∗ for the one-dimensional SFDE (in an unbounded domain, i.e. Ω = R)
from the final data. They proved the uniqueness and conditional stability of the inverse
problem. In addition, they obtained an asymptotically optimal a priori error estimate
between the exact solution and its regularized approximation using Fourier truncation
method. On the other hand, the proposed mathematical analysis in that paper cannot
be applied when Ω is a bounded domain. In this paper, we focus on a multi-dimensional
problem in a general bounded domain by using the final data. To the best of our knowl-
edge, there exists no previous work investigating the reconstruction of an unknown spatial
component in the source term of SFDE in a bounded domain. While the classical para-
bolic models, such as the diffusion model (with s = 1), this inverse source problem has
been studied from different types of observational data such as:

• Reconstruction from final time observations: This case concerns the recovery of
the spatial component f ∗ of the source term F ∗ if u(x, T ) is given. The well-
posedness of this inverse source problem has been proved by Isakov [37, 38]. The
uniqueness question have been analyzed in [12]. Kamynin [41] proved the unique
solvability of the inverse problem of determining the space-dependent source f ∗
with the leading coefficient g depending on time and space variables. Johansson
and Lesnic in [40], to reconstruct the space-dependent source f ∗ , proposed an
iterative procedure based on a sequence of well-posed direct problems which are
solved at each iteration step using the boundary element method. In addition,
they proved that the instability is overcome by stopping the iterations at the first
iteration for which the discrepancy principle is satisfied. While in [39], the authors
for determining the spacewise dependent source proposed a variational conjugate
gradient-type iterative algorithm. Hasanov and Pektaş [33] designed an adjoint
problem approach with subsequent conjugate gradient algorithm to identify the
4

spatial component f ∗ . Related inverse source problems for nonlinear problems have
been studied in [23, 24].
• Reconstruction fromZ integral observations: It consists in identifying the spatial
component f ∗ , if µ(t)u(x, t)dx is given. Here, µ is in L∞ (0, T ) and non-negative.
Z TΩ
Furthermore, µ(t)dt > 0. Such an observation is called integral observation
0
and it is a generalization of the final observation (defined in the above paragraph),
when µ is an approximation to the delta function at t = T . The existence and the
uniqueness of this inverse problem can be found in [42, 43, 49] Erdem et. al in
[21], they reconstructed the space-dependent source f ∗ using a variational least-
squares minimization approach. In order to minimize this error function, they
applied the conjugate gradient method. In [31], the authors developed an identi-
fication method based on the finite element method combined with the conjugate
gradient method. Hào et. al in [30] identifying the spatial component from inte-
gral observations which can be regarded as generalizations of point-wise interior
observations.
• Reconstruction from boundary measurements: This case is devoted to the identifi-
cation of a space-dependent source term from an additional boundary observation.
A large number of works have been devoted to this kind of inverse source prob-
lem. We refer to Cannon [15] and Cannon with Ewing [15, 16] for the generic
well-posedness. We also refer to [18, 20, 60] for the uniqueness and stability, and
[27, 28, 29, 32, 34, 35, 53] etc. for the numerical reconstruction by various regu-
larization methods.

This paper is a generalisation of the works mentioned above where the same problem
was considered for the classical parabolic case, s = 1. We will solve the considered problem
by a variational method. More precisely, the inverse problem is then reformulated as a
least-squares problem in coupling with a Tikhonov regularization term which penalizes
the space-dependent source to be identified. Hence, the Fréchet gradient of the objective
functional is obtained. Then, we apply the conjugate gradient method to approximate
the minimizer of the corresponding least-squares problem and the discrepancy principle is
applied to find a suitable stopping step. The main difficulties of this paper (in comparison
to the same problem in the classical case) are as follows: (i) Nonlocal diffusion operator.
The fractional Laplacian (−∆)s is a nonlocal operator. This can be easily seen from its
definition. (ii) Exterior conditions in Rd \Ω and not boundary conditions on ∂Ω.

The remainder of this paper is organized as follows. We begin with Section 2 which
introduces the notations and some preliminary results. In Section 3, we prove an unique-
ness result for the considered inverse problem. While in Section 4, the inverse problem is
rewritten as an optimization one. The conjugate gradient algorithm based on the sensi-
tivity and adjoint problems is presented in Section 5. In Section 6, we approximate the
variational problem by the finite element approximation and prove the convergence of the
method. Finally, we will test our method for some problems are described in Section 7.
 5

2. Notation and preliminaries

In this section, to be convenient for the readers, we first go over some notation and
preliminaries that will be frequently used in the rest of the paper. To this end, for any
0 < s < 1 we define the fractional order Sobolev space
n o
W0s,2 (Ω) = v ∈ W s,2 Rd : v = 0 in Rd \Ω


(2.1)
and we endow it with the norm given by
Z Z |v(x) − v(y)|2  12
kvkW s,2 (Ω) := dxdy .
0
Rd Rd |x − y|d+2s
The Sobolev space W s,2 (Rd ) of order s over Rd is define by
n
s/2
o
W s,2 (Rd ) := v ∈ L2 Rd : 1 + |ξ|2 F v ∈ L2 Rd ,


(2.2)

where F is the Fourier transform. In addition, we denote by W −s,2 (Ω) the dual space of
W0s,2 (Ω). Besides, let h·, ·i−s,s be the duality pairing between W −s,2 (Ω) and W0s,2 (Ω). For
more details on fractional order Sobolev spaces one can consult for instance [25, 47, 58].

Next, we define the nonlocal operator (−∆)sD in L2 (Ω) as follows.

n o
s s,2 s 2
D ((−∆)D ) := u|Ω : u ∈ W0 (Ω) and (−∆) u ∈ L (Ω) , (2.3)
(−∆)sD u := ((−∆)s u)|Ω a.e. in Ω. (2.4)
Then (−∆)sD is the realization in L2 (Ω) of (−∆)s with the Dirichlet exterior condition
u = 0 in Rd \Ω. In addition, from [51], it is well-known that the operator (−∆)sD has a
compact resolvent. We have the following result taken from (see e.g. [7]).
Lemma 1. The
operator −(−∆)sD generates a strongly continuous sub-markovian semi-
s
group e−t(−∆)D t≥0 on L2 (Ω). Moreover, the operator (−∆)sD can be also viewed as a
bounded operator from W0s,2 (Ω) into W −s,2 this case −(−∆)sD also generates a

(Ω). In −s,2
−t(−∆)sD
strongly continuous semi-group e t≥0
on W (Ω).

In order to introduce the notion of weak solutions to the direct problem (1.4), we define
the bilinear form
  C Z Z (v(x) − v(y))(w(x) − w(y))
d,s
B v, w = dxdy. (2.5)
2 Rd Rd |x − y|d+2s
We are now in a position to give the definition of weak solutions to (1.4).
Definition 2. Let f ∗ ∈ L2 (Ω) and g ∈ C([0, T ]) be given. A function u ∈ X0 is said to
be a weak solution of the problem (1.4) if
D E   D E
∂t u(·, t), w + B u(·, t), w = f ∗ (·)g(t), w ,
−s,s
D E
for every w ∈ W0s,2 (Ω) and almost every t ∈ (0, T ), where ·, · is the usual scalar product
on L2 (Ω) and the space X0 is defined by
1
0, T ; L2 (Ω) ∩ C([0, T ]; W0s,2 (Ω)) ∩ C([0, T ]; L2 (Ω))


X0 := H0,0
6

with n o
1
H0,0 (0, T ; L2 (Ω)) 1 2
= v ∈ H (0, T ; L (Ω)) : v(·, 0) = 0 .

Next lemma gives the existence and uniqueness of a weak solution for problem (1.4).
Lemma 3. Let f ∗ ∈ L2 (Ω) and g ∈ C([0, T ]). Then, there exists a unique weak solution
u ∈ X0 to the problem (1.4) in the sense of Definition 2. Moreover, this solution is given
by
Z t
∗ s
u(x, t) = f (x) e−(t−τ )(−∆)D g(τ )dτ, (2.6)
0
s

where e−t(−∆)D t≥0 is the semi-group mentioned in the above lemma. In addition, there
is a constant c > 0 such that
u + u ≤ c f∗ . (2.7)
H 1 (0,T ;L2 (Ω)) L2 (0,T ;W0s,2 (Ω)) L2 (Ω)

Proof. The existence and uniqueness of a weak solution to (1.4) have been shown in
[44, Theorem 26]. Moreover, from Remark 9 in [44] this weak solution belongs to
C([0, T ]; L2 (Ω)). On the other hand, Biccari, Warma and Zuazua [7] they proved (by
using semi-group theory) that a weak solution to (1.4) enjoys the following regularity
u ∈ C([0, T ]; D ((−∆)sD )) ∩ H0,0
1
(0, T ; L2 (Ω)). (2.8)
Consequently, we deduce that the solution u ∈ X0 . 

3. Unique identifiability

The main objective of this section is to investigate the uniqueness of the solution to our
inverse problem. More precisely, we prove that the space-dependent source f ∗ is uniquely
determined from the final time measurement u(., T ).
Theorem 4. (Uniqueness). Let g ∈ C([0, T ]) be a nonzero nonnegative function. Let u`
with ` = 1, 2 be the solutions of the problems

 ∂t u` + (−∆)s u` = f`∗ g in Ω × (0, T ],
u` = 0 in (Rd \Ω) × (0, T ], (3.1)
u` (., 0) = 0 in Ω,


where f`∗ ∈ L2 (Ω) and such that

u1 (·, T ) = u2 (·, T ) in Ω. (3.2)
Then we have
f1∗ = f2∗ in L2 (Ω). (3.3)

Proof. Consider the difference u2,1 = u2 − u1 , which is the solution to

 ∗
 ∂t u2,1 + (−∆)s u2,1 = f2,1 g in Ω × (0, T ],
u2,1 = 0 in (Rd \Ω) × (0, T ], (3.4)
u2,1 (., 0) = 0 in Ω,

 7

∗
where f2,1 = (f2∗ − f1∗ ) ∈ L2 (Ω). Moreover, according to (3.2), we have
u2,1 = 0 in Ω × {T }. (3.5)
Thanks to Lemma 3, the problem (3.4) has a unique weak solution u2,1 ∈ X0 is given by
Z t
∗ s
u2,1 (x, t) = f2,1 (x) e−(t−τ )(−∆)D g(τ ) dτ. (3.6)
0

Using the relation (2.4), we are allowed to write

D E Z t D s ∗
E
∗
u2,1 (., t), f2,1 = g(τ ) e−(t−τ )(−∆) f2,1 ∗
, f2,1 dτ
0
Z t D E
[−(t−τ )(−∆)s ]/2 [−(t−τ )(−∆)s ]/2 ∗ ∗
= g(τ ) e e f2,1 , f2,1 dτ (3.7)
0
Z t D  ? E
s ∗ s ∗
= g(τ ) e[−(t−τ )(−∆) ]/2 f2,1 , e[−(t−τ )(−∆) ]/2 f2,1 dτ,
0
 ?  
[−(t−τ )(−∆)s ]/2 s
where e is the adjoint operator of e[−(t−τ )(−∆) ]/2 .

From [48, Chapter 1,? Corollary 10.6], in reflexive Banach spaces the generator of semi-
s s ?
group e[−(t−τ )(−∆) ]/2 is e[−(t−τ )((−∆) ) ]/2 . Moreover, the operator (−∆)s is self-adjoint,
we get   s ]/2 ? s )]/2
e[−(t−τ )(−∆) = e[−(t−τ )((−∆) . (3.8)
Inserting (3.8) into (3.7), we see that
D E Z t D E
∗ [−(t−τ )(−∆)s ]/2 ∗ [−(t−τ )(−∆)s ]/2 ∗
u2,1 (., t), f2,1 = g(τ ) e f2,1 , e f2,1 dτ
0
Z t 2
s ∗
= g(τ ) e[−(t−τ )(−∆) ]/2 f2,1 dτ.
0 L2 (Ω)

Considering the above relation at the final time t = T and taking into account u2,1 = 0
in Ω × {T }, we have
Z T 2
s ∗
g(τ ) e[−(t−τ )(−∆) ]/2 f2,1 dτ = 0. (3.9)
0 L2 (Ω)

Due to the fact that g ∈ C([0, T ]) and g > 0, we deduce that

s ]/2 2
∗
e[−(t−τ )(−∆) f2,1 = 0, (3.10)
L2 (Ω)
∗
which implies f2,1 = 0 in L2 (Ω) and consequently also f1∗ = f2∗ . The proof is finished. 

4. Minimization problem and its regularization

In this section, we focus on the second objective of this work: the development of effi-
cient methods capable of numerically reconstructing the spatial component in the source
term of the parabolic fractional heat equation (1.4) with additional information (1.6). We
achieve this by reformulating the inverse source problem into an equivalent minimization
problem where a least-squares functional is minimized.
8

Let f ∈ L2 (Ω) be an initial guess of f ∗ and we consider the potential u[f ] to be the
solution to the following parabolic fractional problem

 ∂t u[f ] + (−∆)s u[f ] = f g in Ω × (0, T ],
u[f ] = 0 in (Rd \Ω) × (0, T ], (4.1)
u[f ](., 0) = 0 in Ω.


Assuming that the final measured data ϑ = u|Ω×{T } is known, where u is the potential
related to the actual spatial component f ∗ (i.e. u solves system (1.4)). Hence, the inverse
problem to be solved consists in finding f ∗ ∈ L2 (Ω) such that the associated potential
u[f ]|Ω×{T } approximate as most as possible the final measured data ϑ in the domain Ω. To
this end, we introduce the observation operator in the form
Λ : f (x) ∈ L2 (Ω) 7−→ u[f ](x, T ). (4.2)
From Lemma 3, we know that u[f ] ∈ C([0, T ]; L2 (Ω)). Then, we deduce that u[f ](., T )
belongs to L2 (Ω). Moreover, the map f 7−→ u[f ] is linear with respect to f . Consequently,
the map f 7−→ Λ f is a bounded linear operator from L2 (Ω) into L2 (Ω).

In order to find an estimate of the actual space-dependent component f ∗ from the final
time information data ϑ, one can minimize the discrepancy
2
J(f ) = Λ f − ϑ (4.3)
L2 (Ω)

with respect to f ∈ L2 (Ω).

It is well known that this problem is unstable (under the perturbation of the final ob-
servation data) and there might be many minimizers to it. A possible way to recover
well-posedness for the minimization problem in L2 (Ω) is to introduce a Tikhonov regu-
larization term in the functional to minimize. The regularized problem reads:

Z 2
Z 2
Jα (f ) = Λ f (x) − ϑ(x) dx + α f (x) dx, (4.4)
Ω Ω

where α being positive Tikhonov regularization parameter.

According to this regularization process, the reconstruction problem can be formulated

as an optimization one. More precisely, the unknown space-dependent component f ∗ is
characterized as the solution to the following minimization problem:

Minimize
2
Jα (f ). (4.5)
f ∈L (Ω)

The regularized solution is defined as the minimizer of variational problem (4.4). Next,
we will prove the existence, stability and convergence of regularized solution. The proofs
of the results are similar to the ones in [36, Theorems 4.3 and 4.4].
Theorem 5. For every α > 0 there exists a unique regularized solution to the optimization
problem (4.5).
 9

Theorem 6. For a measurement data sequence {ϑn }n converging to ϑ, i.e., ϑn → ϑ in L2 (Ω)

as n → ∞, and {f n }n is a minimizer of Jα (f ) for (4.4) with ϑ replaced by ϑn . Then the
regularized solution is stable with respect to the data ϑ

5. Identification approach

In this section, we present a numerical reconstruction approach for solving the mini-
mization problem (4.5). We develop an iterative process based on the conjugate gradient
method for reconstructing the spatial component in the source term of the parabolic frac-
tional equation. To this end, we start our analysis by calculating the Fréchet derivative
of the functional to be minimized.

5.1. Fréchet derivative of Jα . In this subsection, we study the Fréchet differentiability

of the functional f 7−→ Jα (f ). It is very critical to construction the algorithm for finding
the minimizer of the optimization problem (4.5). To this end, we begin by introducing
the following lemma which describes the derivation of the solution operator f 7−→ u[f ],
where u[f ] is the solution to (4.1).
Lemma 7. The mapping f 7−→ u[f ] from L2 (Ω) to L2 (0, T ; W0s,2 (Ω)) is Fréchet differen-
tiable and we get
∇u[f ]δf = u[δf ], for all δf in L2 (Ω).

Proof. It is simply consequently of the linearity of the fractional parabolic problem (4.1).


In order to reduce the computational cost for computing the Fréchet derivative of Jα ,
let us introduce the following adjoint problem
 s
 −∂t p[f ] + (−∆) p[f ] = 0
 in Ω × [0, T ),
p[f ] = 
0  in (Rd \Ω) × [0, T ), (5.1)
p[f ](., T ) = u[f ](·, T ) − ϑ in Ω.


 
2 2
As u[f ](., T ) ∈ L (Ω), ϑ ∈ L (Ω) and then it follows that u[f ](., T ) − ϑ also belongs
to L2 (Ω). Hence, by reverting time direction, one can check that there exists a unique
solution in L2 (0, T ; W0s,2 (Ω)) to (5.1). See, for example [8, Theorem B.5].

In the following theorem, we prove that Jα is Fréchet differentiable and derive a formula
for its gradient
Theorem 8. The functional Jα is Fréchet differentiable and its gradient is
Z T 
∇Jα (f ) = 2 g(t)p[f ](x, t) dt + αf (x) , (5.2)
0

where p[f ] is the solution to the adjoint problem (5.1).

10

Proof. Taking a small variation δf ∈ L2 (Ω) of f , we have

2 2
Jα (f + δf ) − Jα (f ) = u[f + δf ](., T ) − ϑ − u[f ](., T ) − ϑ
L2 (Ω) L2 (Ω)
2 2  
+ α f + δf −α f + o kδf kL2 (Ω)
L2 (Ω) L2 (Ω)
Z   (5.3)
=2 u[f ](., T ) − ϑ u[δf ](., T )dx
ZΩ  
+ 2 f δf dx + o kδf kL2 (Ω) .
Ω

We have u[δf ] as the solution to the problem


 ∂t u[δf ] + (−∆)s u[δf ] = g δf in Ω × (0, T ],
u[δf ] = 0 in (Rd \Ω) × (0, T ], (5.4)
u[δf ](., 0) = 0 in Ω,


The variational formulation of (5.4) and (5.1) gives us

Z T  Z T   Z TZ
∂t u[δf ], ψ dx dt + B u[δf ], ψ dt = δf (x)g(t)ψ(x, t) dx dt, (5.5)
0 −s,s 0 0 Ω

for all ψ ∈ L2 (0, T ; W0s,2 (Ω)).

Z T  Z T
∂t p[f ], ψ dx dt + B(p[f ], ψ) dt = 0, for all ψ ∈ L2 (0, T ; W0s,2 (Ω)). (5.6)
0 −s,s 0

By taking ψ = p[f ] in (5.5) and by integration by parts with respect to t, we obtain

Z T  Z T   Z TZ
∂t p[f ], u[δf ] dxdt + B p[f ], u[δf ] dt = δf (x)g(t)p[f ](x, t)dxdt.
0 −s,s 0 0 Ω
(5.7)
By integrating by parts over t and replacing ψ = u[δf ] in (5.6) we get
Z T  Z T   Z  
∂t p[f ], u[δf ] dxdt + B p[f ], u[δf ] dt = u[f ](., T ) − ϑ u[δf ](., T )dx.
0 −s,s 0 Ω

Hence,
Z   Z T Z
u[f ](., T ) − ϑ u[δf ](., T ) = δf (x)g(t)p(x, t) dx dt. (5.8)
Ω 0 Ω
Therefore,
Z T Z Z  
Jα (f + δf ) − Jα (f ) = 2 δf (x)g(t)p[f ](x, t) dx dt + 2 f (x)δf (x) dx + o kδf kL2 (Ω)
0 Ω Ω
 Z T   
= 2 g p[f ]dt + αf, δf + o kδf kL2 (Ω) .
0
(5.9)
Consequently, the functional Jα is Fréchet differentiable and its gradient has the form
Z T 
∇Jα (f ) = 2 g(t) p[f ](x, t)dt + αf .
0

 11

5.2. Conjugate gradient method (CGM). To find the minimizer of Jα , we use the
conjugate gradient method (CGM). It proceeds as follows: Assume that at the kth itera-
tion we have f k . We use the following iterative scheme:
f k+1 = f k + ξk dk , (5.10)
with (
−∇Jα (f k ) if k = 0,
dk = (5.11)
−∇Jα (f k ) + βk dk−1 if k = 1, 2, · · · ,

2
∇Jα (f k )L2 (Ω)
βk = (5.12)
k∇Jα (f k−1 )k2L2 (Ω)
and
ξk = arg min Jα (f k + ξdk ),
ξ≥0

where,
2 2
Jα (f k + ξdk ) = Λf k + ξΛdk − ϑδ + α f k + ξdk .
L2 (Ω) L2 (Ω)

Differentiating Jα (f k + ξdk ) with respect to ξ, we get

∂Jα (f k + ξdk ) D E
= ξ kΛdk k2L2 (Ω) + Λdk , ϑ
∂ξ
D E
+ αξ kdk k2L2 (Ω) + α dk , f k .

∂Jα (f k + ξdk )
In order to determine the step size ξk , we put = 0, which implies
∂ξ
D E D E
Λdk , ϑδ + α dk , f k
L2 (Ω)
ξk = − 2 2 , k = 0, 1, 2, · · ·
Λdk + α dk
L(Ω) L2 (Ω)
D E
k
dk , ∇Jα (f )
L2 (Ω)
=− 2 2 .
Λdk + α dk
L2 (Ω) L2 (Ω)

On the other hand, we have

D E
dk = −∇Jα (f k ) + βk dk−1 , rk = −∇Jα (f k ), and rk , dk−1 = 0. (5.13)

Therefore;
2
rk
L2 (Ω)
ξk = 2 2 (5.14)
Λdk + α dk
L2 (Ω) L2 (Ω)

Therefore, we have the following CGM:

12

Algorithm 1: Conjugate Gradient Algorithm

• Set k = 0, initiate f 0 .
• Calculate r0 = ∇Jα (f 0 ) and set d0 = r0 .
• Evaluate
2
r0
L2 (Ω)
ξ0 = 2 2
Λd0 + α d0
L2 (Ω) L2 (Ω)
1 0
Set f = f + ξ0 d0
• For k = 1, 2, · · · . Calculate
rk = −∇Jα (f k ), dk = rk + βk dk−1
with
krk k2L2 (Ω)
βk = . (5.15)
krk−1 k2L2 (Ω)
• Calculate
2
rk
L2 (Ω)
ξk = 2 2 .
Λdk + α dk
L2 (Ω) L2 (Ω)
Update
f k+1 = f k + ξk dk .

6. Finite Element Method

In this section, we will discrete the nonlinear optimization problem (4.5) by the finite
element method and prove its convergence.

Thanks to Theorem 5, let f α ∈ L2 (Ω) be the unique solution of the optimization prob-
lem (4.5). Then, this solution f α is characterized by the first-order optimality condition
as follows  
Λ∗ Λ f α − ϑ + αf α = 0, (6.1)
where Λ∗ : L2 (Ω) 7−→ L2 (Ω) is the adjoint operator of the observation operator Λ defined
by Λ∗ q = ζ. Here ζ is the solution of the following adjoint problem

 −∂t ζ + (−∆)s ζ = 0 in Ω × (0, T ],
ζ = 0 in (Rd \Ω) × (0, T ], (6.2)
ζ(., T ) = q in Ω.


Next, we will approximate the optimality condition (6.1) by the finite element method.
More precisely, we will approximate the operators Λ and Λ∗ as follows

6.1. Finite element approximation of the operators Λ and Λ∗ . We present now a

fully discrete FE approximation for the above variational problem. Supposing that Ω is
a polyhedral domain, we triangulate Ω into a shape regular quasi-uniform mesh Th with
 13

mesh size hx . In addition, let Vh ⊂ W0s,2 (Ω) be the piecewise linear finite element space
associated with Th , namely
n o
Vh := vh ∈ C(Ω); vh |K ∈ P1 (K), ∀K ∈ Th . (6.3)

Here, P1 (K) is the space of linear polynomials on the element K. For fully discretization
we introduce a uniform partition of the integral [0, T ] : 0 = t0 < t1 < . · · · < tN , where
tn = n δt, n = 0, 1, . · · · , N with the time step size δt = T /N.

Then, the fully discrete approximation for the fractional parabolic problem (4.1) by
the backward Euler-Galerkin method with initial datum u(0) [f ] = 0 as follows: Find
unh [f ] ∈ Vh for n = 1, 2, · · · , N such that
D E   D E
dt unh [f ], v + B unh [f ], v = f g n , v ∀v ∈ Vh , (6.4)

unh [f ] − un−1
h [f ]
where dt unh [f ]
= , n = 1, 2, · · · , N, and g n = g(tn ). Moreover, we recall
δt
that the bounded bilinear form B(., .): W0s,2 (Ω) × W0s,2 (Ω) → R is W0s,2 (Ω)-elliptic, i.e.,
  C Z Z |v(x) − v(y)|2 2
d,s
B v, v = d+2s
dxdy ≥ C(d, s) v s,2 . (6.5)
2 Rd Rd |x − y| W0 (Ω)

The discrete variational problem (6.4) admits a unique solution unh [f ] ∈ Vh . Next, let
uh [f ](x, t) be the linear interpolation of unh [f ] with respect to t. We now define a dis-
cretization of the observation operator Λ, i.e. Λh f := uh [f ](., T ) in Ω.

Hence the discrete version of the optimization problem (4.5) reads

2 2
Jα,h (f ) = Λh f − ϑ +α f → min (6.6)
L2 (Ω) L2 (Ω)

which is characterized by the first-order optimality condition

 
Λ∗h Λh fhα − ϑ + αfhα = 0, (6.7)

where Λ∗h is the adjoint operator of the linear operator Λh .

Now for the finite element approximation of the adjoint problem (6.2) we define an
b ∗ of Λ defined by
approximation Λ h

b ∗ q = ζh for q ∈ L2 (Ω),
Λ (6.8)
h

where ζh ∈ Vh is the unique solution of the finite element approximation of (6.2). In

addition, if instead of the final time observation ϑ we get only the perturbation ϑδ satisfies
the following condition

ϑ − ϑδ ≤ δ with δ ≥ 0. (6.9)
L2 (Ω)

Hence, we arrive at the following variational problem

 
b ∗ Λh fbα − ϑδ + αfbα = 0.
Λ (6.10)
h h h
14

6.2. Convergence results. Recently, Glusa and Otárola [22] derived a theoretical L2 -
error estimate for the fully discrete variational problem (6.4). They analyzed the depen-
dence of the convergence rate with respect to the space step h on the fractional Laplacian
order s.
Lemma 9. (see [22, Theorem 5.8]). Let u[f ] and uh [f ] solve (4.1) and (6.4), respectively.
Then there exists a constant c > 0, independent on the time and space steps, such that
 
u[f ] − uh [f ] ≤ c δt + hγ(s) f , (6.11)
L2 (Ω×(0,T )) L2 (Ω)

where γ(s) is the convergence rate with respect to h given as


 4s if s ∈ (0, 1/6) ,
γ(s) = s + 1/2 − σ if s ∈ [1/6, 1/2) , (6.12)
 1−σ if s ∈ [1/2, 1),
with σ > 0 is an arbitrary small real number.

From this Lemma 9, we can conclude the following convergence results.

Theorem 10. We have
       
γ(s) ∗ ∗ γ(s)
Λh − Λ f = O δt + h and Λh − Λ q
b = O δt + h ,
L2 (Ω) L2 (Ω)

where γ(s) is defined in (6.12).

Proof. The proof of this theorem is a simple consequence of Lemma 9. 

Theorem 11. For α > 0, we have the following error estimate:
 
α α γ(s)
f − fhb ≤ c δt + h +δ , (6.13)
L2 (Ω)

where f α and fbhα solve respectively variational problems (6.1) and (6.10). In addition,
c > 0 is a constant depending on f α , ϑ and α.

Proof. In order to simplify the presentation, we denote all constants c1 , c2 , . · · · indepen-

dent of f α , α, and ϑ whose value changes according to the place it is used. From (6.7),
we note that
2 D E 2 D E
α α α α
α fh = ϑ, Λh fh − Λh fh ≤ ϑ, Λh fh ≤ c1 ϑ fhα
L2 (Ω) L2 (Ω) L2 (Ω) L2 (Ω)

Hence,
c1
fhα ≤ ϑ (6.14)
L2 (Ω) α L2 (Ω)

To derive the estimate (6.13), we first consider the error f α − fˆhα . From Equations
L2 (Ω)
(6.7) and (6.10), we have
     
α fhα − fˆhα = Λ̂∗h fhα Λh fˆhα − ϑ − Λ∗h fhα Λh fhα − ϑδ .
 15

Therefore,
2 D E
α fhα − fˆhα = Λ̂∗h Λh fˆhα − Λ∗h Λh fhα
+ − Λ̂∗h ϑ Λ∗h ϑ, fhα − fˆhα
L2 (Ω)
D E D E
= (Λ̂∗h Λ∗ )Λh fˆhα , fhα − fˆhα + Λ∗ Λ∗h fˆhα − Λ∗h Λh fhα , fhα − fˆhα
D E
+ Λ∗h ϑδ − Λ̂∗h ϑ, fhα − fˆhα .

Using Theorem 10 and inequality (6.14) we can estimate the first term on the right-hand
side of the last equality as follows:
D E D E
(Λ̂∗h ˆ
∗α α ˆα ∗ ∗ ˆ α α
− Λ )Λh fh , fh − fh = (Λ̂h − Λ )Λh (fh − fh ), fh − fh α ˆα

D E
+ (Λ̂∗h − Λ∗ )Λh fhα , fhα − fˆhα

≤ (Λ̂∗h − Λ∗ )Λh (fˆhα − fhα ) fhα − fˆhα

L2 (Ω) L2 (Ω)

+ (Λ̂∗h − Λ∗ )Λh fhα fhα − fˆhα

L2 (Ω) L2 (Ω)
2
≤ c2 (δt + hγ(s) ) fhα − fˆhα + c3 (δt + hγ(s) ) ϑδ fhα − fˆhα
L2 (Ω) L2 (Ω) L2 (Ω)

Similarly, we can estimate the second term, by using Theorem 10 and inequality (6.14),
D E D E D E
Λ∗ Λ∗h fˆhα − Λ∗h Λh fhα , fhα − fˆhα = Λh fˆhα , Λ(fhα − fˆhα ) − Λh fhα , Λh (fhα − fˆhα )
D E D E
= Λh (fˆhα − fhα ), Λ(fˆhα − fhα ) + Λh fhα , (Λ − Λh )(fhα − fˆhα )
D E 2
= (Λh − Λ)(fˆhα − fhα ), Λ(fhα − fˆhα ) − Λ(fhα − fˆhα )
L2 (Ω)
D E
+ Λh fhα , (Λ − Λh )(fhα − fˆhα )
2
≤ (c3 (δt + hγ(s) ) fhα − fˆhα
L2 (Ω)

+ c4 (δt + hγ(s) ) ϑδ L2 (Ω)

fhα − fˆhα
L2 (Ω)

Finally, using Theorem 10 and inequality (6.9), we obtain

D E D E D E
Λ∗h ϑδ − Λ̂∗h ϑ, fhα − fˆhα = (Λ − Λ∗h )ϑ, fhα − fˆhα + ϑδ − ϑ, Λ(fhα − fˆhα )
D E
δ α
+ ϑ , (Λh − Λ)(fh − fh ) ˆα

≤ c6 (δt + hγ(s) + δ) fhα − fˆhα

L2 (Ω)

Hence, we conclude that:

fhα − fˆhα ≤ c7 (δt + hγ(s) + ε) (6.15)

L2 (Ω)
16

Similarly, from (6.7), (6.10), (6.14) and Theorem 10, we obtain that:
2 D E
α f α − fˆhα = Λ̂∗h Λh fˆhα − Λ∗ Λh fˆhα + Λ∗ Λh fˆ − Λ∗ Λf α − Λ̂∗h ϑ + Λ∗ ϑδ , f α − fˆhα
L2 (Ω)
D E D E
= (Λ̂∗h − Λ∗ )Λh fˆhα , f α − fˆhα + (Λh fˆhα −α , Λ(f α − fˆhα )
D E
∗ δ ∗ α
+ Λ ϑ − Λ̂h ϑ, f − fh ˆα

D E D E
= (Λ̂∗h − Λ∗ )Λh fˆhα , f α − fˆhα + (Λh − Λ)fˆhα , Λ(f α − fˆhα )
D E D E
ˆα α α ˆα ∗
+ Λ(fh − f ), Λ(f − fh ) + (Λ − Λ̂h )ϑ, f − fh ∗ α ˆα

D E
+ ϑδ − ϑ, Λ(f α − fˆhα )

≤ c8 (δt + hγ(s) ) f α − fˆhα

L2 (Ω)

The assertion of the theorem thus follows. 

7. Numerical Experiments

In this section, we will apply the conjugate gradient algorithm established in the previ-
ous section to the numerical treatment of problem (4.5) in one and two spatial dimensions
cases, that is the reconstruction of the spatial component f of the source term for the
fractional heat equation (4.1).
In our numerical computation, the noisy data is generated by adding a random pertur-
bation, i.e.
ϑδ = ϑ + ε ϑ · (2 · rand(size(ϑ)) − 1), ∀x ∈ Ω,
where the magnitude ε represents the relative noise level. The corresponding noise level
is calculated by
δ = ϑδ − ϑ L2 (Ω) .
To show the accuracy of the numerical solution, we compute the approximate L2 error
denoted by
Ek = f k − fexact L2 (Ω) ,
where f k is the heat source reconstructed at the kth iteration, and fexact is the exact
solution. The residual Rk at the kth iteration is given by
Rk = u(f k ) − ϑδ L2 (Ω)
.
In an iteration algorithm, the most important work is to find a suitable stopping rule. In
this study we use the well-known Morozov’s discrepancy principle [46], i.e. we choose k
satisfying the following inequality:
Rk 6 τ δ < Rk−1 ,
where τ > 1 is a constant and can be taken heuristically to be τ = 1.01, as suggested by
Hanke and Hansen [26]. In the case of noise-free data i.e. ε = 0, we take k = 100.

Next, in the numerical implementation, the final time T and the time-dependent source
g(t) are always chosen as T = 1 and g(t) = e−t , respectively.
 17

7.1. One-dimensional case. Without loss of generality, the space domain Ω is taken as
1
Ω = [−1, 1]. In this case, the grid size for time and space variable are fixed as ∆t = 50
1
and ∆x = 100 , respectively.

7.1.1. Choice of the regularization parameter. In the literature, there already exists several
strategies for choosing an appropriate (not the optimal) regularization parameter α > 0.
Tikhonov regularization is a typical example of an a-priori parameter choice rule since
the choice of the parameter of regularization α > 0 is made a-priori (depends only on
the noise level δ > 0). However, in our numerical experiments, we take α = δ 2 for noisy
data, which is essentially an a-priori choice. In order to validate this choice, we apply
our proposed algorithm in reconstructing a source term defined on Ω by f 1 (x) = sin(πx)
for various noise levels in the cases of α = 0 (without regularization) and α = δ 2 (with
regularization) by taking k = 100.

In Figures 1a and 1b, we illustrate the comparisons of recovered solutions with the true
one in the case of α = 0 and α = δ 2 , respectively, with noise level greater than 5%.

(a) α = 0 (b) α = δ 2

Figure 1. Numerical approximations to f 1 with s = 0.5 and ε ∈

{5%, 10%, 15%, 20%}. Left: No regularization. Right: with regu-
larization.

The choices of ε and α in this test and the corresponding numerical performances are
listed in table 1.
ε 5% 10% 15% 20%
α=0 0.2071 0.4151 0.6232 0.8310
α = δ2 0.1131 0.1185 0.1855 0.2012
Table 1. Values of the error function with respect to ε and α.
18

Figures 2a and 2b compare the true source terms f 1 with the corresponding recon-
structions ones in the case of α = 0 and α = δ 2 , respectively, with noise level less than
1%.

(a) α = 0 (b) α = δ 2

Figure 2. Numerical approximations to f 1 with s = 0.5 and ε ∈

{0%, 0.1%, 0.5%, 1%}. Left: No regularization. Right: with regu-
larization.

The choices of ε and α in this case (ε ≤ 0.01) and the corresponding numerical perfor-
mances are listed in table 2.

ε 0% 0.1% 0.5% 1%
α=0 0.0001 0.0041 0.0205 0.0410
α = δ2 0.0001 0.0038 0.0202 0.0407
Table 2. Values of the error function with respect to ε and α.

From the obtained results, one can deduce the following observations:

• From Figure 1 and Table 1, one can deduce that when the relative noise level ε
is greater than 5% the regularization parameter is very necessary to find a stable
solution. More precisely, the numerical approximations to f 1 without using the
regularization (i.e. α = 0) have somewhat amplitude oscillations (see Figure 1a).
It means that the problem is unstable in this case. When using the proposed
regularization method, the numerical results are more accurate than the ones
without using regularization (see Figure 1b).
• When the considered inverse problem is slightly unstable (the parameter ε is less
than 1%) the regularization parameter α does not have a significant influence on
 19

the accuracy of our algorithm. More precisely, the numerical approximations to f 1

without using the regularization, i.e. α = 0, are almost similar to those obtained
with α = δ 2 (see Figure 2 and Table 2). It means that in this case, we can apply
our algorithm by taking α = 0.

Thus the regularization parameter α can be taken as δ 2 in this study. In fact, it is a

difficult problem to choose an appropriate regularization parameter.

7.1.2. Reconstruction results by using Morozov’s discrepancy principle. In this paragraph,

we test the performance of our iterative algorithm 1 in the reconstruction of four examples
of the source terms in the cases of s = 0.3 and s = 0.7 with various noise levels.

Example 1. Suppose f 1 (x) = sin(πx) and the final data u(x, T ) are obtained by solving
the direct problem (4.1). The numerical results for this example by using the discrepancy
principle for various noise levels in the cases of s = 0.3, s = 0.7 are shown in Figures 3a
and 3b, respectively, by taking α = δ 2 . We can see that the numerical results for this
example match the exact ones quite well even up to 10% noise added in the “exact” final
data ϑδ .

(a) s = 0.3 (b) s = 0.7

Figure 3. The numerical results for example 1 for various noise

levels.

Example 2. In this example, we test the performance of our iterative algorithm in the
2
reconstruction of the source term defined by: f 2 (x) = (x2 − 1) e−x , with different values
of noise levels. Figures 4a and 4b compares the true source terms with their corresponding
reconstructions in the case of s = 0.3 and s = 0.7, respectively.

Example 3. We test a nonsmooth example with a cusp. More precisely, we apply our
algorithm in reconstructing an unknown nonsmooth source term defined by f 3 (x) = 1−|x|.
In Figures 5a and 5b, we illustrate the comparisons of recovered solutions with the true
ones for various noise levels.
20

(a) s = 0.3 (b) s = 0.7

Figure 4. The numerical results for example 2 for various noise

levels.

(a) s = 0.3 (b) s = 0.7

Figure 5. The numerical results for example 3 for various noise

levels.

Example 4. In this paragraph, we test the performance of our iterative algorithm in the
reconstruction of a discontinuous source term. As a prototype example, we consider

4, x ∈ [−0.5, 0.5]

4
f (x) =

1, x ∈ [−1, −0.5) ∪ (0.5, 1].

In Figures 6a and 6b, we present the curves of the exact source term and their corre-
sponding reconstructions for different values of the parameter ε.
 21

(a) s = 0.3 (b) s = 0.7

Figure 6. The numerical results for example 4 for various noise

levels.

From Figures 3-6, we can see that the numerical results for examples 1-4 are quite
accurate up to 1% noise added in the exact final data u(x, T ). Moreover, one can deduce
the following remarks:

• Our algorithm works slightly better when the fractional-order s is small.

• The numerical results of nonsmooth source terms (examples 3 and 4) are less
accurate in comparison with the numerical results of smooth function (examples
1 and 2).
• The efficiency and the accuracy of our algorithm decrease with the increase of the
noise level.

(a) Ek (b) Rk

Figure 7. The errors Ek and the residuals Rk for example 1 with

different values of ε.
22

7.1.3. Convergence and stability results. This paragraph is concerned with the conver-
gence and stability of the proposed algorithm. Moreover, the effectiveness of the regular-
ization parameter in the Tikhonov regularization function is discussed. The approxima-
tion errors Ek and the residuals Rk for Example 1 with various noise levels are shown in
Figure 7 with s = 0.5.

From Figure 7a, it can be observed that the approximation errors Ek become smaller
as the noise levels decrease, and after a few iterations, the computed errors have slightly
increased, thus we have to stop at a suitable step. In Figure 7b, it can be seen that the
Morozov’s discrepancy principle provides the stopping steps where the residuals attain
nearly their minimums.

7.2. Two-Dimensional Case. Now we proceed to the more challenging two-dimensional

case, where we divide the space-time region Ω × [0, T ] := [0, 1]2 × [0, 1] into 50 × 50 × 50
equidistant meshes. Similarly to the one-dimensional examples, we will test the numerical
performance of algorithm 1 in the reconstruction of two examples of source terms in the
cases of s = 0.3, s = 0.7.

Example 5. Suppose that the source term to be reconstructed is defined by f 5 (x, y) =

sin(x) cos(y). The 3D surface of this function is plotted in Figure 8

Figure 8. 3D surface of the source term f 5

The numerical results of this example in the cases of s = 0.3 and s = 0.7 are illustrated
in Figures 9a and 9b, respectively by taking α = 0 and ε = 0. We can see that the
numerical results are quite accurate to the exact solutions, and the proposed algorithm is
effective.
 23

(a) s = 0.3 (b) s = 0.7

Figure 9. The numerical results for example 5 with α = 0.

(a) Absolute error for s = 0.3 (b) Absolute error for s = 0.7

Figure 10. Absolute error for s = 0.3, s = 0.3 for Example 5

Example 6. In this test, we take the exact spatial function

5
f 6 (x, y) = (x3 + y 3 ).
4

The variation of the source function f 6 is implemented in Figure 11.

24

Figure 11. 3D surface of the source term f 6

We use the finite difference method developed in [6, 36] to solve problem (4.1) to obtain
the exact measured u(x, y, T ). The regularized solutions of our minimization problem in
the cases of s = 0.3 and s = 0.7, by taking α = 0 and ε = 0, are shown in Figures 12a
and 12b, respectively.

(a) s = 0.3 (b) s = 0.7

Figure 12. The numerical results for example 6 with α = 0.

 25

(a) Absolute error for s = 0.3 (b) Absolute error for s = 0.7

Figure 13. Absolute error for s = 0.3 and s = 0.7 for Example 6

Figures 10 and 13 show the absolute error between the precise source function and the
numerical solutions for the values of s = 0.3 and s = 0.7, respectively. As can be seen,
the numerical results closely match the exact ones.

From Figures 9 and 12, It is readily seen that algorithm 1 also works efficiently and
accurately in the two-dimensional case.

8. Conclusion

This paper is mainly concerned with an inverse problem for a space-fractional diffusion
equation; that is, to determine the space-dependent source term from noisy final data. In
the theoretical aspect, the existence and uniqueness for the direct problem and the unique-
ness for the inverse problem are both proved. The considered ill-posed inverse problem is
reformulated as a minimization one. The existence, uniqueness, and stability of the so-
lution are discussed. We use the Tikhonov regularization method to solve the considered
inverse problem and use a conjugate gradient algorithm to find the approximation of the
regularized solution. In the numerical part, we consider the numerical reconstruction of
four examples in one-dimensional case and two examples in two-dimensional case to show
the accuracy and the efficiency of the proposed method.

References

[1] M. Abdelwahed, M. BenSaleh, N. Chorfi, and M. Hassine. An inverse problem study related to
a fractional diffusion equation. Journal of Mathematical Analysis and Applications, 512(2):126145,
2022.
[2] H. Antil and S. Bartels. Spectral approximation of fractional pdes in image processing and phase
field modeling. Computational Methods in Applied Mathematics, 17(4):661–678, 2017.
[3] H. Antil, T. Berry, and J. Harlim. Fractional diffusion maps. Applied and Computational Harmonic
Analysis, 54:145–175, 2021.
26

[4] H. Antil and C.N. Rautenberg. Sobolev spaces with non-muckenhoupt weights, fractional elliptic
operators, and applications. SIAM Journal on Mathematical Analysis, 51(3):2479–2503, 2019.
[5] M. BenSalah and M. Hassine. Inverse source problem for a diffusion equation involving the fractional
spectral laplacian. Mathematical Methods in the Applied Sciences, 44(1):917–936, 2021.
[6] U. Biccari and V. Hernández-Santamarı́a. Controllability of a one-dimensional fractional heat equa-
tion: theoretical and numerical aspects. IMA Journal of Mathematical Control and Information,
36(4):1199–1235, 2019.
[7] U. Biccari, M. Warma, and E. Zuazua. Local regularity for fractional heat equations. In Recent
advances in PDEs: analysis, numerics and control, pages 233–249. 2018.
[8] U. Biccari, M. Warma, and E. Zuazua. Control and numerical approximation of fractional diffusion
equations. Numerical Control: Part A, 23:1, 2022.
[9] C. Bjorland, L.A. Caffarelli, and A. Figalli. Nonlocal tug-of-war and the infinity fractional laplacian.
Communications on Pure and Applied Mathematics, 65(3):337–380, 2012.
[10] M. Bologna, C. Tsallis, and P. Grigolini. Anomalous diffusion associated with nonlinear fractional de-
rivative fokker-planck-like equation: Exact time-dependent solutions. Physical Review E, 62(2):2213,
2000.
[11] C. Bucur and E. Valdinoci. Nonlocal diffusion and applications, volume 20. Springer, 2016.
[12] I. Bushuyev. Global uniqueness for inverse parabolic problems with final observation. Inverse Prob-
lems, 11(4):L11, 1995.
[13] L.A. Caffarelli, J.-M. Roquejoffre, and Y. Sire. Variational problems with free boundaries for the
fractional laplacian. Journal of the European Mathematical Society, 12(5):1151–1179, 2010.
[14] L.A. Caffarelli and L. Silvestre. An extension problem related to the fractional laplacian. Commu-
nications in partial differential equations, 32(8):1245–1260, 2007.
[15] J.R. Cannon. Determination of an unknown heat source from overspecified boundary data. SIAM
Journal on Numerical Analysis, 5(2):275–286, 1968.
[16] J.R. Cannon and R.E. Ewing. Determination of a source term in a linear parabolic partial differential
equation. Zeitschrift für angewandte Mathematik und Physik ZAMP, 27(3):393–401, 1976.
[17] H. Cheng, C.-L. Fu, G.-H. Zheng, and J. Gao. A regularization for a riesz–feller space-fractional
backward diffusion problem. Inverse Problems in Science and Engineering, 22(6):860–872, 2014.
[18] M. Choulli and M. Yamamoto. Some stability estimates in determining sources and coefficients.
Journal of Inverse and Ill-posed Problems, 14:355–373, 2006.
[19] S. Dipierro, G. Palatucci, and E. Valdinoci. Dislocation dynamics in crystals: a macroscopic theory
in a fractional laplace setting. Communications in Mathematical Physics, 333(2):1061–1105, 2015.
[20] H.W. Engl, O. Scherzer, and M. Yamamoto. Uniqueness and stable determination of forcing terms in
linear partial differential equations with overspecified boundary data. Inverse Problems, 10(6):1253,
1994.
[21] A. Erdem, D. Lesnic, and A. Hasanov. Identification of a spacewise dependent heat source. Applied
Mathematical Modelling, 37(24):10231–10244, 2013.
[22] C. Glusa and E. Otárola. Error estimates for the optimal control of a parabolic fractional pde. SIAM
Journal on Numerical Analysis, 59(2):1140–1165, 2021.
[23] N.L. Gol’dman. Determination of the right-hand side in a quasilinear parabolic equation with a
terminal observation. Differential Equations, 41(3):384–392, 2005.
[24] N.L. Gol’dman. Finding the right-hand side in multidimensional parabolic equations with terminal
observation. Differential Equations, 43(8):1101–1110, 2007.
[25] P. Grisvard. Elliptic problems in nonsmooth domains. SIAM, 2011.
[26] M. Hanke and P.C. Hansen. Regularization methods for large-scale problems. Surv. Math. Ind,
3(4):253–315, 1993.
[27] D.N. Hào. A noncharacteristic cauchy problem for linear parabolic equations ii: A variational
method. Numerical functional analysis and optimization, 13(5-6):541–564, 1992.
[28] D.N. Hào. A noncharacteristic cauchy problem for linear parabolic equations iii: A variational
method and its approximation schemes. Numerical functional analysis and optimization, 13(5-6):565–
583, 1992.
[29] D.N. Hào. A non-characteristic cauchy problem for linear parabolic equations and related inverse
problems: I. solvability. Inverse problems, 10(2):295, 1994.
[30] D.N. Hào, B.V. Huong, N.T.N. Oanh, and P.X. Thanh. Determination of a term in the right-hand
side of parabolic equations. Journal of Computational and Applied Mathematics, 309:28–43, 2017.
 27

[31] D.N. Hào, P.X. Thanh, D. Lesnic, and M. Ivanchov. Determination of a source in the heat equation
from integral observations. Journal of Computational and Applied Mathematics, 264:82–98, 2014.
[32] H. Harbrecht and J. Tausch. An efficient numerical method for a shape-identification problem arising
from the heat equation. Inverse problems, 27(6):065013, 2011.
[33] A. Hasanov and B. Pektaş. A unified approach to identifying an unknown spacewise dependent
source in a variable coefficient parabolic equation from final and integral overdeterminations. Applied
Numerical Mathematics, 78:49–67, 2014.
[34] F. Hettlich and W. Rundell. Identification of a discontinuous source in the heat equation. Inverse
Problems, 17(5):1465, 2001.
[35] M. Hrizi, M. Hassine, and R. Malek. A new reconstruction method for a parabolic inverse source
problem. Applicable Analysis, 98(15):2723–2750, 2019.
[36] Mourad Hrizi, Mohamed BenSalah, and Maatoug Hassine. Determination of the initial density in
nonlocal diffusion from final time measurements. Discrete and Continuous Dynamical Systems-S,
15(6):1469, 2022.
[37] V. Isakov. Inverse source problems. Number 34. American Mathematical Soc., 1990.
[38] V. Isakov. Inverse parabolic problems with the final overdetermination. Communications on Pure
and Applied Mathematics, 44(2):185–209, 1991.
[39] B.T. Johansson and D. Lesnic. A variational method for identifying a spacewise-dependent heat
source. IMA Journal of Applied Mathematics, 72(6):748–760, 2007.
[40] T. Johansson and Daniel Lesnic. Determination of a spacewise dependent heat source. Journal of
computational and Applied Mathematics, 209(1):66–80, 2007.
[41] V.L. Kamynin. On the unique solvability of an inverse problem for parabolic equations under a final
overdetermination condition. Mathematical Notes, 73(1):202–211, 2003.
[42] V.L. Kamynin. On the inverse problem of determining the right-hand side of a parabolic equation
under an integral overdetermination condition. Mathematical notes, 77(3):482–493, 2005.
[43] V.L. Kamynin. On the solvability of the inverse problem for determining the right-hand side of a
degenerate parabolic equation with integral observation. Mathematical Notes, 98(5):765–777, 2015.
[44] T. Leonori, I. Peral, Ana Primo, and F. Soria. Basic estimates for solutions of a class of nonlocal
elliptic and parabolic equations. Discrete & Continuous Dynamical Systems, 35(12):6031, 2015.
[45] M.M. Meerschaert. Fractional calculus, anomalous diffusion, and probability. Fractional dynamics:
recent advances, pages 265–284, 2012.
[46] Vladimir Alekseevich Morozov. Methods for solving incorrectly posed problems. Springer Science &
Business Media, 2012.
[47] E. Di Nezza, G. Palatucci, and E. Valdinoci. Hitchhikers guide to the fractional sobolev spaces.
Bulletin des sciences mathématiques, 136(5):521–573, 2012.
[48] A. Pazy. Semigroups of linear operators and applications to partial differential equations, volume 44.
Springer Science & Business Media, 2012.
[49] A.I. Prilepko and D.S. Tkachenko. Properties of solutions to the parabolic equation and unique-
ness of the solution to the inverse source problem with integral overdetermination. Computational
mathematics and mathematical physics, 43(4):537–546, 2003.
[50] I.V. Saveliev. General physics course. Tutorial in, 3, 1989.
[51] R. Servadei and E. Valdinoci. On the spectrum of two different fractional operators. Proceedings of
the Royal Society of Edinburgh Section A: Mathematics, 144(4):831–855, 2014.
[52] C. Shi, C. Wang, G. Zheng, and T. Wei. A new a posteriori parameter choice strategy for the con-
volution regularization of the space-fractional backward diffusion problem. Journal of computational
and applied mathematics, 279:233–248, 2015.
[53] D.D. Trong, P.N. Dinh Alain, and P. Thanh Nam. Determine the special term of a two-dimensional
heat source. Applicable Analysis, 88(3):457–474, 2009.
[54] D.D. Trong, D.N.D. Hai, and N.D. Minh. Optimal regularization for an unknown source of space-
fractional diffusion equation. Applied Mathematics and Computation, 349:184–206, 2019.
[55] E. Valdinoci. From the long jump random walk to the fractional laplacian. Boletin de la Sociedad
Espanola de Matematica Aplicada. SeMA, 49.
[56] J.L. Vázquez. Nonlinear diffusion with fractional laplacian operators. In Nonlinear partial differential
equations, pages 271–298. 2012.
[57] G.M. Viswanathan, V. Afanasyev, S.V. Buldyrev, E.J. Murphy, P.A. Prince, and H.E. Stanley. Lévy
flight search patterns of wandering albatrosses. Nature, 381(6581):413–415, 1996.
 28

[58] M. Warma. The fractional relative capacity and the fractional laplacian with neumann and robin
boundary conditions on open sets. Potential Analysis, 42(2):499–547, 2015.
[59] C.J. Weiss, B.G. Waanders, and H. Antil. Fractional operators applied to geophysical electromag-
netics. Geophysical Journal International, 220(2):1242–1259, 2020.
[60] M. Yamamoto. Conditional stability in determination of force terms of heat equations in a rectangle.
Mathematical and Computer Modelling, 18(1):79–88, 1993.
[61] G.H. Zheng and T. Wei. Two regularization methods for solving a riesz–feller space-fractional back-
ward diffusion problem. Inverse Problems, 26(11):115017, 2010.

(M. Hrizi) Institut Supérieur des Mathématiques Appliquées et de l’Informatique de

Kairouan Avenue Assad Iben Fourat, 3100 Kairouan, Tunisie

Email address: mourad-hrizi@hotmail.fr

(M. BenSalah) Higher Institute of Applied Science and Technology of Sousse, Rue Tahar
Ben Achour, Sousse 4003, Tunisia

Email address: mohamed.bensalah@fsm.rnu.tn

(M.S. Hussein) Department of Mathematics, College of Science, University of Baghdad,

Baghdad, Iraq

Email address: mmmsh@sc.uobaghdad.edu.iq

(H. Ould Sidi) Department of Mathematics, Faculty of Sciences University of Nouak-

chott Al Aasriya, Nouakchott, Mauritania

Email address: hamedouldsidi@yahoo.fr

View publication stats