Professional Documents
Culture Documents
Discrete and Continuous Dynamical Systems Series S: Doi:10.3934/dcdss.2022029
Discrete and Continuous Dynamical Systems Series S: Doi:10.3934/dcdss.2022029
2022029
DYNAMICAL SYSTEMS SERIES S
Mourad Hrizi∗
Institut Supérieur des Mathématiques Appliquées et de l’Informatique de Kairouan
Avenue Assad Iben Fourat, 3100 Kairouan, Tunisie
1
2 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
we denote the integral fractional Laplacian of order s defined by (see, e.g., [20, 55])
f (x) − f (y)
Z
(−∆)s f (x) := cd,s P.V. d+2s
dy, (1)
Rd |x − y|
where the symbol P.V. denotes the principal value of the integral; i.e.
f (x) − f (y) f (x) − f (y)
Z Z
P.V. d+2s
dy = lim dy, (2)
Rd |x − y| →0 Rd \B (x) |x − y|d+2s
where T > 0 is a given fixed final time, and u0 is an initial condition belongs to the
finite energy space L2 (Ω). This model problem has been employed for describing
the anomalous diffusion in various fields of engineering applications such as porous
media, biology, quantum mechanics, · · · etc. From a probabilistic point of view the
previous forward problem can be interpreted as follows, for more information about
this model problem one can consult [21, Section 2] and references therein:
(i) The function u can be understood as the probability distribution of the posi-
tion of a particle moving randomly inside Ω according to a random walk with
arbitrarily long jumps.
(ii) The homogeneous Dirichlet condition u = 0 means that the particle is killed
when hitting the boundary ∂Ω.
(iii) The scalar function u0 represent the initial probability distribution of the
position of the particle.
Particularly, if we proceed with the above probabilistic interpretation of the con-
sidered model equation (4), the inverse problem we consider in this case consists
in recovering the initial probability distribution of the position of particles from a
given final time measurement. In a more general context, this inverse problem can
be formulated as follows:
Let U σ,T be a given noisy measurement of the solution u at the time t = T . We
aim to identify the initial data u0 in the nonlocal parabolic problem (4) such that
the final time solution u(., T ) approximate as most as possible (in a sense to be
precise later) the measured data U σ,T on the domain Ω.
To our knowledge, there are no works on the inverse problem of determining
an unknown initial data of space-fractional parabolic equation. While the classical
parabolic models, such as diffusion equations (with s = 1), this inverse problem
has been studied by many authors. In [13, 31, 37, 38, 50] the initial condition is
reconstructed from additional temperature data at a terminal time. In [43] the
authors recovered the initial condition from an interior observations in ω × (τ, T ),
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 3
this end, we begin by introducing the spectral decomposition of the solution of the
fractional parabolic problem (4).
Firstly, it is well-known that for a given initial condition u0 ∈ L2 (Ω), the direct
problem (4) admits a unique solution, see for example [42, Theorem 26]. In order
∞
to derive the spectral decomposition of the solution u, we denote by {(ϕk , λk )}k=1
s
the eigenpairs of the fractional Laplace operator (−∆) with Dirichlet condition.
That is for each k ∈ N∗ , the pair (ϕk , λk ) solves
(−∆)s ϕk = λk ϕk in Ω,
(
(5)
ϕk = 0 in Rd \Ω.
In addition, it is well-known that the fractional Laplacian operator has a strictly
increasing positive sequence of eigenvalues satisfying
0 < λ1 ≤ λ2 ≤ · · · ≤ λk ≤ · · · and lim λk = +∞.
k→+∞
∞
Besides, the set of eigenfunctions {ϕk }k=1 forms an orthonormal basis of L2 (Ω).
Remark 1. The eigenvalue problem for the fractional Laplace operator (−∆)s , s ∈
(0, 1) has been the subject of some research studies. For instance, Kwaśnicki [40]
derived some theoretical results describing the behavior of the eigenvalues as well
as the eigenfunctions with respect to k and s. Particularly, the author proved that
the eigenvalues of the one-dimensional fractional Laplace operator in the interval
(−1, 1) satisfies that there exist two constants c1 > 0 and c2 > 0 such that
2s
3
kπ (1 − s)π c1 (1 − s)
c 4s
λ k − − ≤ √ , for all k ≥ 2 .
2 4 k s s
The regularity of the solutions of the eigenvalues problem (5) was discussed in [12].
It is proved that the eigenfunctions ϕk , k ∈ N∗ belong to H s+1/2−ς (Rd ), with ς > 0
is an arbitrary small real number.
According to the previous notations, the solution u of the fractional parabolic
problem (4) admits the following decomposition
∞
X
u(x, t) = uk (t)ϕk (x), (6)
k=1
D E D E
where uk (t) = u(., t), ϕk , with ., . is the usual scalar product on L2 (Ω).
Since, at this formal stage, we have u(x, 0) = u0 (x), this representation yields the
following fractional initial value problem for uk
∂t uk (t) + λk uk (t) = 0,
(7)
uk (0) = u0k ,
D E
with u0k = u0 , ϕk . Then, we immediately deduce that the unique solution of (7)
is given by
uk (t) = u0k e−λk t , (8)
which leads to an explicit formula for the solution u as
X∞
u(x, t) = ϕk (x)u0k e−λk t . (9)
k=1
With the help of this decomposition, we establish the following uniqueness result:
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 5
Theorem 2.1. Let u1 and u2 be two solutions of the fractional parabolic problem
∂t u` + (−∆)s u` = 0 in Ω × (0, T ],
u` = 0 in (Rd \Ω) × (0, T ], (10)
u` (., 0) = u0` in Ω,
relative to the initial data u0` ∈ L2 (Ω), ` = 1, 2. Then, if the solutions u1 and u2
coincide at the final time, i.e.
u1 = u2 in Ω × {T }, (11)
we have
u01 = u02 in L2 (Ω).
Proof. Consider the difference U = u2 − u1 , which is the solution to
∂t U + (−∆)s U = 0 in Ω × (0, T ],
U = 0 in (Rd \Ω) × (0, T ], (12)
U (., 0) = U0 in Ω,
Therefore,
D E X∞ D E2
U (., t), U 0 = U 0 , ϕk e−λk t . (15)
k=1
Considering the above relation at the final time t = T and taking into account
U = 0 in Ω × {T }, one can get
X∞ D E2
U 0 , ϕk e−λk T = 0. (16)
k=1
−λk T
From the fact that e > 0, one can deduce that
D E
U 0 , ϕk = 0 for all k ≥ 1. (17)
Then, it follows
U 0 = 0 in L2 (Ω),
which implies the desired result.
under the constraint that w[v] solves the following fractional parabolic problem
∂t w[v] + (−∆)s w[v] = 0 in Ω × (0, T ],
w[v] = 0 in (Rd \Ω) × (0, T ], (20)
w[v](., 0) = v in Ω.
In the next section, we will analyze this optimization problem and we will present
some theoretical results.
4.1. Notation and preliminaries. The purpose of this section is to introduce the
notations and some preliminary results, which will be used in the sequel.
For 1 ≤ p < ∞, let Lp (Ω), H01 (Ω), and H 1 (Ω) be the usual classical Lebesgue
and Sobolev spaces (see, e.g., [14]). Consider an abstract Banach space Y with
norm k · kY . The space Lp (0, T ; Y) consists of measurable functions z : [0, T ] → Y
such that
!1/p
Z T
kzkLp (0,T ;Y) = kz(t)kpY dt < ∞. (23)
0
The space H 1 (0, T ; Y ) consists of functions z : [0, T ] → Y such that the weak
derivative ∂t z exists and
! 12
Z T
2
kzkH 1 (0,T ;Y) = kz(t)k2Y + k∂t z(t)kY dt < ∞. (25)
0
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 7
For any s ≥ 0, we define H s Rd , the Sobolev space of order s over Rd , by [57,
Definition 15.7]
n s/2 o
H s Rd := z ∈ L2 Rd : 1 + |ξ|2 F z ∈ L2 Rd ,
(26)
s
where F is the Fourier transform. The space H s Rd at hand, we define H (Ω) as
the closure of C0∞ (Ω) in H s Rd . This space can be equivalently characterized by
and to H s (Ω)∩H01 (Ω) for s ∈ (1, 3/2) [46, Theorem 3.33]. In addition, the fractional
Sobolev space H s (Ω) is defined by
( 12 )
|z(x) − z(y)|2
Z Z
s 2
H (Ω) = z ∈ L (Ω) : |z|H s (Ω) := d+2s
dx dy < ∞ . (28)
Ω×Ω |x − y|
s s
We denote by H −s (Ω) = (H (Ω))? the dual space of H (Ω). Besides, let h·, ·i−s,s
s
be the duality pairing between H −s (Ω) and H (Ω). Finally, we define the bilinear
form c ZZ
d,s (z(x) − z(y))(h(x) − h(y))
As z, h = dx dy, (29)
2 Rd ×Rd |x − y|d+2s
s
constitutes an inner product on H (Ω). The norm it induces, which is just a mul-
tiple of the H s (Rd )-seminorm, is equivalent to the full H s (Rd )-norm on this space,
because a Poincaré-type inequality holds in it. See, for instance, [3] for details.
To prove the well-posedness of the minimization problem (22), we recall an energy
estimate and review some regularity results for the solution of the nonlocal parabolic
problem (20). To this end, we introduce the notion of weak solutions to (20).
Definition 4.1. A weak formulation for problem (20) reads as follows: Find w[v] ∈
X such that w[v](., 0) = v and
Z TD E Z T
∂t w[v](., t), φ(., t) dt + As w[v](., t), φ(., t) dt = 0, (30)
0 −s,s 0
2 s
for all φ ∈ L (0, T ; H (Ω)). The functional space X is defined by
s
X := H 1 0, T ; H −s (Ω) ∩ L2 0, T ; H (Ω) ∩ C([0, T ]; L2 (Ω)).
(31)
On the basis of [25, Theorem 3.1], [1, Theorem 3.1.(a)] and [42, Theorem 26], we
summarize the well-posedness of the problem (20) as follows.
Lemma 4.2. For given v ∈ L2 (Ω), the problem (20) has a unique weak solution in
the sense of Definition 4.1 and is given by
∞
X D E
w[v](x, t) = ϕk (x) v, ϕk e−λk t . (32)
k=1
Moreover, there exists a constant c = c(Ω, d, s, T ) > 0 such that
+
w[v]
2 +
w[v]
≤ c
v
2 .
w[v]
1 s
H (0,T ;H −s (Ω)) L (0,T ;H (Ω)) C([0,T ];L2 (Ω)) L (Ω)
(33)
8 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
Consequently, we have
s
Kγ (v0 )
≤ for all n ∈ N.
vn
L2 (Ω) γ
Then, one can check that {vn }n is uniformly bounded in L2 (Ω). Therefore, there
exist v ∗ and a sub-sequence of {vn }n , still denoted by {vn }n , such that
vn * v ∗ in L2 (Ω) as n → ∞, (35)
where the symbol * denotes the weak convergence. Next, we prove that v ∗ is
indeed the unique minimizer to (22).
For each n ∈ N, let us consider w[vn ] the solution of the fractional parabolic
equation (20) with v = vn . Thanks to Lemma 4.2, we can deduce that the sequence
{w[vn ]}n is uniformly bounded in X . This indicates the existence of some w∗ ∈ X
and a sub-sequence of {w[vn ]}n , again still denoted by {w[vn ]}n such that
s
w[vn ] * w∗ in L2 (0, T ; H (Ω)) as n → ∞,
w[vn ] * w in H (0, T ; H −s (Ω)) as n → ∞,
∗ 1 (36)
w[vn ] * w∗ in C([0, T ]; L2 (Ω)) as n → ∞.
In the next step, we aim to show that w∗ = w[v ∗ ], with v ∗ is the initial condition
defined by (35). To this end, we multiply both sides of (20) ( w[v] is replaced by
s
w[vn ], v is replaced by vn ) by a test function ϕ ∈ L2 (0, T ; H (Ω)) and get
Z T D E Z T
∂t w[vn ](., t), ϕ(., t) dt + As w[vn ](., t), ϕ(., t) dt = 0. (37)
0 −s,s 0
We passe n to infinity in (37) and using the convergence result (36), one can obtain
Z T D E Z T
∂t w∗ (., t), ϕ(., t) dt + As w∗ (., t), ϕ(., t) dt = 0,
0 −s,s 0
s
∀ϕ ∈ L2 (0, T ; H (Ω)).
Further, we shall prove w∗ (., 0) = v ∗ , which together with the Definition 4.1 of weak
solution implies that
w∗ = w[v ∗ ]. (38)
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 9
(39)
Letting n → ∞ in the above equality and using the convergence results (35) and
(36), we have
Z TZ Z TZ Z
∂t w∗ η θ(t) dxdt = − w∗ η θ0 (t) dxdt − v ∗ (x)η(x)θ(0) dx. (40)
0 Ω 0 Ω Ω
On the other hand, by integration by parts with respect to t, we also have
Z TZ Z TZ Z
∗ ∗ 0
∂t w η θ(t) dxdt = − w η θ (t) dxdt − w∗ (x, 0)η(x)θ(0) dx. (41)
0 Ω 0 Ω Ω
∗ ∗
which together with (40) implies w (x, 0) = v (x).
By vn * v ∗ in L2 (Ω) and w[vn ] * w[v ∗ ] in C([0, T ]; L2 (Ω)) (from (36) and (38)),
we employ the lower semi-continuity of the L2 -norm to conclude
2
2
Kγ (v ∗ ) ≤ lim inf
w[vn ](., T ) − U σ,T (.)
2 + γ lim inf
vn
2 , (42)
n→∞ L (Ω) n→∞ L (Ω)
Then there exists a subsequence of {vn }n , still denoted by {vn }n , such that
vn → v ∗ strongly in L2 (Ω) as n → ∞, (46)
∗
where v is the minimizer of problem (22).
10 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
For any v ∈ L2 (Ω), again we take advantage of the lower semi-continuity of the
2
L -norm to deduce
2
2
Kγ (v ∗ ) =
w[v ∗ ](., T ) − U σ,T (.)
2 + γ
v ∗
2
L (Ω) L (Ω)
2
2
≤ lim inf
w[vn ](., T ) − Unσ,T (.)
2 + γ lim inf
vn
2
n→∞ L (Ω) n→∞ L (Ω)
2
2
σ,T
≤ lim inf
w[vn ](., T ) − Un (.)
2 + γ
vn
2
n→∞ L (Ω) L (Ω)
2
2
≤ lim
w[v](., T ) − Unσ,T (.)
+ γ
v
n→∞ L2 (Ω) L2 (Ω)
2
2
=
w[v](., T ) − U σ,T (.)
+ γ
v
= Kγ (v), ∀v ∈ L2 (Ω), (51)
L2 (Ω) L2 (Ω)
∗
which verifies that v is the minimizer of (22).
In the next step, we shall prove that {vn } converges to v ∗ strongly in L2 (Ω) by
contradiction. Assuming that it is not true; i.e.
6−→
v ∗
as n → ∞.
vn
2 2
L (Ω) L (Ω)
∗ 2
As vn * v in L (Ω), by the weak lower semi-continuity of the norm, we have
∗
≤ lim inf
vn
. (52)
v
2
L (Ω) n→∞ L2 (Ω)
By setting B := lim sup
vn
2 , we get
n→∞ L (Ω)
B = lim sup
vn
> lim inf
vn
≥
v ∗
(53)
n→∞ 2 L (Ω) 2n→∞ L (Ω)2 L (Ω)
Proof. From the fact that w[v] solves the space-fractional parabolic problem (20),
one can deduce that the map v 7−→ w[v] is linear with respect to v. Consequently
the Fréchet derivative of the solution operator v 7−→ L(v) in the direction δv, has
the form
∇L(v)δv := w[δv], ∀δv ∈ L2 (Ω).
On the other hand, the weak variational formulation of the sensitivity problem (55)
can be rewritten as follow
Z TZ Z T
s
∂t w[δv] φ dxdt + As (w[δv], φ)dt = 0, ∀φ ∈ L2 (0, T ; H (Ω)). (61)
0 Ω 0
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 13
By taking φ = z[v] in (61) as a test function and using (60), one can obtain
Z Z
σ,T
w[v](., T ) − U (.) w[δv](., T ) dx = z[v](., 0) δv dx. (62)
Ω Ω
Inserting (62) in (59), we get
Z
Kγ (v + δv) − Kγ (v) = 2 (z[v](., 0) + γv)δv dx + o
δv
L2 (Ω)
D Ω E
(63)
= 2(z[v](., 0) + γv), δv + o
δv
2 .
L (Ω)
Consequently, the functional Kγ is Fréchet differentiable and its gradient has the
form
∇Kγ (v) = 2 z[v](., 0) + γv , ∀v ∈ L2 (Ω).
The following corollary follows from the first order necessary optimality condition
for the minimization problem (22). It provides a simplified characterization of the
solution to the minimization problem (22).
Corollary 1. Let v ∗ ∈ L2 (Ω) be the solution to the regularized minimization prob-
lem (22). Then, the initial condition v ∗ satisfies the following Euler-type equation:
z[v ∗ ](., 0) + γ v ∗ = 0, (64)
∗ ∗
where z[v ] is the solution to the backward problem (56) which depends on v and
the measured data U σ,T .
Remark 2. The relation (64) gives the link between the unknown initial condition
and the measurement of the final time solution. From the fact that the quadratic
functional Kγ (.) is convex, one can easily show that this condition is sufficient [44].
5.2. Reconstruction algorithm. We develop in this paragraph a numerical re-
construction algorithm for identifying the unknown initial condition u0 from noisy
measurement of the final time state. The proposed algorithm is based on the con-
jugate gradient method and the Morozov’s discrepancy principle (see, e.g., [49]).
The employed technique involves a denoising procedure at each iteration step and
provides a sequence of initial condition approximations {vk }k≥0 converging in norm
to the actual solution u0 of the minimization problem (22). Let v0 is a given initial
guess. Then, the sequence of the reconstructed terms are defined by the following
relation
vk+1 = vk + ξk dk , k = 0, 1, 2, · · · , (65)
where ξk is the search step size and dk is the direction of descent defined by the
iterative relation
dk = −∇Kγ (vk ) + αk dk−1 , ∀k ≥ 1, (66)
starting from d0 = −∇Kγ (v0 ).
Concerning the conjugate coefficient αk , different expressions are available in the
literature. Here, we exploit the proposed Fletcher-Reeves approach [5, 19, 23]
0Z if k = 0,
2
[∇Kγ (vk )] dx
αk = Z Ω (67)
if k = 1, 2, · · · .
2
[∇Kγ (vk−1 )] dx
Ω
14 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
where U σ,T is the given noisy measurement of the final time solution.
For this iterative procedure, the key work is to find the best stopping criteria. In
this paper, we use the well-known Morozov’s discrepancy principle. More precisely,
the iterative procedure is stopped when the number of iterations k satisfying the
following inequality:
Rk 6 σ < Rk−1 , (71)
where > 1 is a constant and can be taken heuristically to be 1.01, as estimated by
Hanke et al. [28]. Whereas in the particular case when σ = 0 (i.e. without noise), we
stop the iterative procedure when the variable k reached a predetermined maximum
number. In the above stopping condition, the residual Rk is defined by
Rk =
w[vk ](., T ) − U σ,T (.)
2 . (72)
L (Ω)
Based on the above discussion, we summarize in the following algorithm the main
steps of our reconstruction approach.
Algorithm 1: Conjugate Gradient Method
1. Initialize v0 and set k = 0;
2. Solve the direct problem (20) with v = vk , and compute the residual
rk = w[vk ](x, T ) − U σ,T (x), x ∈ Ω.
3. Solve the adjoint problem (56) and evaluate the gradient ∇Kγ (vk );
4. Calculate the conjugate coefficient αk by (67) and the direction dk
by (66);
5. Compute wk via solving the sensitive problem (55) with δv = dk ;
6. Calculate the step size ξk by (70);
7. Renew the initial condition vk+1 by (65);
8. Increase k by one and go to Step (2), repeat the process until a stopping
condition is satisfied.
The combination of the conjugate gradient algorithm and the Morozov’s discrep-
ancy principle has been successfully employed for solving some inverse problems.
Indeed, it has been used for identifying an unknown zeroth-order term from bound-
ary measurement in [56], the detection of a time-dependent source term in [61] and
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 15
6.1.1. Time and space discretization. To introduce the employed discretization ap-
proach, we start by dividing the time interval [0, T ] into M equal sub-intervals
[tm , tm+1 ] such that
0 = t0 < t1 < · · · < tM = T,
and tm = m τ for each m ∈ {0, 1, · · · , M − 1} with τ = T /M is the time step.
In the following, we denote by wm the approximated solution at the time tm .
Based on the Euler explicit scheme, the time derivative is approximated as follow
wm+1 − wm
∂w
≈ , for 0 ≤ m ≤ M − 1.
∂t t=tm τ
For the space discretisation, we apply the P1 finite element method in the context
of fractional order derivatives, one can consult [1] and references therein for more
details. Let Xh be the standard nodal finite element space of piecewise linear
functions
Xh := {φ ∈ X ; φ|Th ∈ P1 } ,
where Th is a shape regular triangulation of Ω with a mesh size h.
Then, from the weak formulations of the direct and adjoint problems one can
obtain the following space-time discrete systems:
• Discrete version of the direct problem (20): given wh0 = v, find whm+1 ∈
Xh (m = 0, . . . , M − 1) such that
D E D E
τ As whm+1 , φh + whm+1 , φh = whm , φh , ∀φh ∈ Xh , (73)
where As (., .) is the bilinear form defined by (29) and whp is the approximated
direct problem solution at time t = tp .
• Discrete version of the adjoint problem (56): given zhM = 0, find zhm ∈ Xh (m =
0, . . . , M − 1) such that
D E D E D E
τ As zhm , φh + zhm , φh = τ fhm , φh + zhm+1 , φh , ∀φh ∈ Xh , (74)
16 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
where zhp is the approximated adjoint state at time t = tp and fhm is the
approximation of the source term in (56).
Note that in the adjoint problem (56) the source term contains the Dirac delta
function which is approximated by
1 2 2
δ(t − T ) ≈ √ e−(t−T ) /κ , (75)
κ π
where κ is a small positive constant taken as κ = 10−3 .
N
Now, let {ψi }i=1 be a basis of the discrete space Xh , where N denotes the number
of basis functions. Then, the approximated function whp at time t = tp , p = 0, ..., M
can be decomposed as
XN
whp [v](x) = wjp ψj (x),
j=1
where;
• W m+1 = (wjm+1 )1≤j≤N represents the unknown approximated solution at
time t = tm+1 ,
• Ash = (asi,j )1≤i,j≤N denotes the stiffness matrix, which depends on the lapla-
cian fractional order as follows
(ψi (x) − ψi (y)) (ψj (x) − ψj (y))
Z Z
c1,s
asi,j = dx dy,
2 R R |x − y|1+2s
• Mh = (msi,j )1≤i,j≤N is the mass matrix, defined by
Z
mi,j = ψi (x) ψj (x) dx,
Ω
m
• F = (Fim )1≤i≤N ∈R N
denotes the term source, which is defined by
F m = Mh W m , m = 0, . . . , M − 1.
Similarly, the discrete system (74) (which is associated with the adjoint problem)
can be gathered in the form of a linear matrix system.
where w is the solution to the nonlocal initial value problem (20), wh is the solution
to the associated discrete problem (73) and η(s) is the convergence rate with respect
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 17
to h given as
4s if s ∈ (0, 1/6) ,
η(s) = s + 1/2 − ε if s ∈ [1/6, 1/2) , (77)
1−ε if s ∈ [1/2, 1),
As one can check, if we choose the source term F and the initial condition w0 as
√ √
−e−t (1 − x2 )s 2−2s π −t 0 2−2s π
F (x, t) = + e and w (x) = (1 − x2 )s ,
Γ( 1+2s
2 )Γ(1 + s) Γ( 1+2s
2 )Γ(1 + s)
the solution of this problem admits the following analytic expression
√
e−t (1 − x2 )s 2−2s π
wex (x, t) = χ(−1, 1) (x).
Γ( 1+2s
2 )Γ(1 + s)
To examine the convergence of the approximate solution wh to the analytic one
wex , we take the time step τ = h and we introduce the following error function
Err(h) =
wex − wh
, (78)
2L (Ω×(0,T ))
which measures the discrepancy between the exact and approximate solutions for
each step size h. It is important to note here that in the case when τ = h, the
derived estimate (76) implies that the error function Err satisfies the following
behavior with respect to h
Err(h) = O hη(s) . (79)
Step size hj Err(hj ), s = 0.1 Err(hj ), s = 0.2 Err(hj ), s = 0.4 Err(hj ), s = 0.8
h1 = h0 3.61e − 1 1.56e − 1 1.24e − 1 8.69e − 2
h2 = h0 /2 2.75e − 1 1.13e − 1 6.15e − 2 4.41e − 2
h3 = h0 /4 1.70e − 1 6.76e − 2 3.11e − 2 2.14e − 2
h4 = h0 /8 1.33e − 1 4.28e − 2 1.85e − 2 1.13e − 2
h5 = h0 /16 9.59e − 2 2.44e − 2 9.51e − 3 5.02e − 3
h6 = h0 /32 6.74e − 2 1.36e − 2 5.01e − 3 2.49e − 3
Table 1. Values of the error function Err for each s in
{0.1, 0.2, 0.4, 0.8} relative to the mesh step size variation.
To show the quality of the obtained results, we plot in Figure 1, the exact wex as
well as the approximate solutions (wex and wh ) at time t = 0.5 for different values
of the fractional Laplacian order s ∈ {0.1, 0.2, 0.4, 0.8} at time t = 0.5.
As one can observe here, the numerical and theoretical convergence rates are
nearly identical for each value s ∈ {0.1, 0.2, 0.4, 0.8} of the fractional Laplacian
order, which implies the accuracy of the proposed approximation method.
6.3.3. Third example. In this test, we apply our numerical procedure to reconstruct
an initial condition defined by the following function
3
vtrue (x) = sin(4πx)ex + x3/2 (1 − x)5/2 + sin(x) cos(x)e−x + cos(x), ∀x ∈ Ω.
In Figure 5, we illustrate the comparison of recovered solutions with the true ones.
Figure 5. Exact (blue dashed line) and obtained (red line) results.
We summarize in Table 3 the values of the error function err after convergence
of the iterative process. It measures the relative error between the exact and recon-
structed initial conditions for the previous three test examples with respect to the
mesh size h = 2hj−1
0
, j = 1, . . . , 5, where h0 = 0.1.
As one can observe here, the proposed reconstruction algorithm provides a very
accurate results, with an error less that 1% compared with the exact ones when the
mesh step size h ≤ 0.05. Particularly, we have obtained a good approximation of
the sinusoidal-like curve.
In the next section, we show the influence of the regularization parameter γ.
Then, the effect of the fractional Laplacian order s is examined in Section 6.5,
whereas the influence of the noise level ρ on the accuracy of the reconstructed
results is discussed in Section 6.6. Finally, the performance of our algorithm for
reconstructing non-smooth initial conditions is shown in Section 6.7.
with vrec (γ) is the reconstructed initial condition related to a given γ. For this
purpose, we examine the variation of the error function Er with respect to γ. We
discuss the following three situations:
(i) When using an exact data: We aim to examine the effect of the regularization
parameter γ on the accuracy of the reconstruction procedure when the em-
ployed final measured data is exact (without noise; i.e. ρ = 0%). To this end,
we solve the regularized inverse problem for each regularization parameter
value γ belonging to the following set:
The obtained values of the relative error are illustrated in Table 4. The
behavior of the error function γ 7−→ Er(γ) is plotted in Figure 6.
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 23
6.5. Effect of the fractional-order parameter. This test is devoted to the in-
fluence of the fractional Laplacian order s on the reconstruction results. We apply
1
the optimization procedure to identify the initial condition vtrue (given in Example
1, Section 6.3) using different values of the derivative order s ∈ (0, 1). To avoid
the influence of the other parameters, we deal with an exact final time measure
(without noise ρ = 0%).
To evaluate such influence, we present in Figure 9 the variation of the L2 -error
function when the parameter s varies in the interval (0, 1). From Figure 9, we can
remark that the proposed initial condition reconstruction procedure works slightly
better when the fractional derive order is small. However, the values of the L2 -error
function remains acceptable (less than 1%) for all s ∈ (0, 1), which means that the
influence of the parameter s is not significant.
6.6. Influence of the noise level. To emphasize further the reliability of our
initial condition reconstruction procedure, we examine the influence of the noise
level on the accuracy of the reconstruction results. To this end, we apply our
2
numerical algorithm to reconstruct the initial condition defined by function vtrue
(see Example 2, Section 6.3) from a noisy measured data U σ,T , with different noise
levels. According to the numerical studies in Section 6.4, the parameter γ is fixed
as γ = 10−7 , which valid for low or high noise level.
The results of this numerical test are summarized in Figure 10. We show the
noise level influence for two values of the fractional laplacian order s = 0.3 (left)
and s = 0.7 (right). For both cases, we plot the exact solution (see black line) as
well as the reconstructed ones for ρ = 1% (blue curve), ρ = 5% (green curve) and
ρ = 10% (red curve).
From Figure 10, one can deduce that the proposed numerical procedure is robust
with respect to noise and able to provide a reasonably good reconstruction results.
Indeed, we have an accurate solution (compared to the exact one) when the noise
26 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
level is less than 5%. However, the difference between the exact and approximate
solutions becomes relatively large when the noise level exceeds 5%. Moreover, as
shown in the previous section, the optimization process is more accurate and efficient
when the fractional Laplacian order s is relatively small.
To evaluate the accuracy of our numerical procedure, we plot in Figure 11 the exact
initial condition vtrue (with dashed blue line) as well as the reconstructed ones (with
red line) for four values of the fractional Laplacian order; s = 0.1, 0.3, 0.6, 0.9.
Moreover, we illustrate in Table 7 the obtained values of the L2 −error function
measuring the difference between the exact and reconstructed solutions for the four
considered fractional Laplacian orders.
From Figure 11 and Table 7, one can observe that the presence of the singular
points affects the accuracy of the reconstructed initial condition vtrue . Indeed, the
efficiency of our iterative procedure decreases sharply near these points, mainly
for higher values of the fractional Laplacian order. In addition, as expected, the
accuracy of our algorithm decreases with the increase of the derivative order s.
This can be explained by the fact that when the order s increases, the solution of
the fractional parabolic equation will be more smooth. Particularly, the final time
solution u(., T ) (used as a measured data) will be smooth. Then, it is natural that
the efficiency of the optimization procedure decreases when the order s increases
because one cannot obtain an accurate approximation of a non-smooth function
from smooth data.
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 27
Acknowledgments. The authors thank the anonymous referees for valuable com-
ments and for numerous helpful remarks.
28 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
REFERENCES
[1] G. Acosta, F. M. Bersetche and J. P. Borthagaray, Finite element approximations for frac-
tional evolution problems, Fractional Calculus and Applied Analysis, 22 (2019), 767–794.
[2] G. Acosta, F. M. Bersetche and J. P. Borthagaray, A short FE implementation for a 2d
homogeneous Dirichlet problem of a fractional Laplacian, Computers & Mathematics with
Applications, 74 (2017), 784–816.
[3] G. Acosta and J. P. Borthagaray, A fractional Laplace equation: Regularity of solutions and
finite element approximations, SIAM Journal on Numerical Analysis, 55 (2017), 472–495.
[4] M. Ali, S. Aziz and S. A. Malik, Inverse source problem for a space-time fractional diffusion
equation, Fractional Calculus and Applied Analysis, 21 (2018), 844–863.
[5] O. M. Alifanov, Inverse Heat Transfer Problems, Springer Science & Business Media, 2012.
[6] O. M. Alifanov and E. A. Artioukhine, Extreme Methods for Solving Illposed Problems with
Applications to Inverse Heat Transfer Problems, Begell House, New York, 1995.
[7] L. Ambrosio, N. Fusco and D. Pallara, Functions of Bounded Variation and Free Discontinuity
Problems, The Clarendon Press, Oxford University Press, New York, 2000.
[8] H. Attouch, G. Buttazzo and G. Michaille, Variational Analysis in Sobolev and BV Spaces:
Applications to PDEs and Optimization, SIAM, Philadelphia, PA, 2014.
[9] G. Barles, E. Chasseigne and C. Imbert, On the Dirichlet problem for second-order elliptic
integro-differential equations, Indiana University Mathematics Journal, 57 (2018), 213–246.
[10] M. BenSalah and M. Hassine, Inverse source problem for a diffusion equation involving the
fractional spectral Laplacian, Mathematical Methods in the Applied Sciences, 44 (2021), 917–
936.
[11] A. Bonito, J. P. Borthagaray, R. H. Nochetto, E. Otàrola and A. J. Salgado, Numerical
methods for fractional diffusion, Computing and Visualization in Science, 19 (2018), 19–46.
[12] J. P. Borthagaray, L. M. Del Pezzo and S. Martı́nez, Finite element approximation for the
fractional eigenvalue problem, Journal of Scientific Computing, 77 (2018), 308–329.
[13] N. Boussetila and F. Rebbani, Optimal regularization method for ill-posed Cauchy problems,
Electronic Journal of Differential Equations, 147 (2006), 1–15.
[14] H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equations, Universi-
text. Springer, New York, 2011.
[15] L. Caffarelli and L. Silvestre, Regularity theory for fully nonlinear integro-differential equa-
tions, Communications on Pure and Applied Mathematics: A Journal Issued by the Courant
Institute of Mathematical Sciences, 62 (2009), 597–638.
[16] X. Cao and H. Liu, Determining a fractional Helmholtz system with unknown source and
medium parameter, preprint, arXiv:1803.09538, 2018.
[17] A. Carbotti, S. Dipierro and E. Valdinoci, Local density of solutions of time and space frac-
tional equations, preprint, arXiv:1810.08448, 2018.
[18] M. Cekić, Y.-H. Lin and A. Rüland, The calderón problem for the fractional Schrödinger
equation with drift, Calculus of Variations and Partial Differential Equations, 59 (2020), 46
pp.
[19] J. W. Daniel, The conjugate gradient method for linear and nonlinear operator equations,
SIAM Journal on Numerical Analysis, 4 (1967), 10–26.
[20] E. Di Nezza, G. Palatucci and E. Valdinoci, Hitchhiker’s guide to the fractional Sobolev
spaces, Bulletin des Sciences Mathèmatiques, 136 (2012), 521–573.
[21] S. Dipierro, X. Ros-Oton and E. Valdinoci, Nonlocal problems with Neumann boundary
conditions, Revista Mathematica Iberoamericana, 33 (2017), 377–416.
[22] S. Dipierro, O. Savin and E. Valdinoci, Local approximation of arbitrary functions by solutions
of nonlocal equations, The Journal of Geometric Analysis, 29 (2019), 1428–1455.
[23] R. Fletcher and C. M. Reeves, Function minimization by conjugate gradients, The Computer
Journal, 7 (1964), 149–154.
[24] J. C. Gilbert and J. Nocedal, Global convergence properties of conjugate gradient methods
for optimization, SIAM Journal on Optimization, 2 (1992), 21–42.
[25] C. Glusa and E. Otàrola, Error estimates for the optimal control of a parabolic fractional
pde, SIAM Journal on Numerical Analysis, 59 (2021), 1140–1165.
[26] C. W. Groetsch, The Theory of Tikhonov Regularization for Fredholm Equations of the First
Kind, Boston Pitman Publication, 1984
[27] G. Grubb, Fractional Laplacians on domains, a development of Hörmander’s theory of µ-
transmission pseudodifferential operators, Advances in Mathematics, 268 (2015), 478–528.
DETERMINATION OF THE INITIAL DENSITY IN NONLOCAL DIFFUSION 29
[28] M. Hanke, A. Neubauer and O. Scherzer, A convergence analysis of the Landweber iteration
for nonlinear ill-posed problems, Numerische Mathematik , 72 (1995), 21–37.
[29] D. N. Hào and N. T. N. Oanh, Determination of the initial condition in parabolic equations
from boundary observations, Journal of Inverse and Ill-Posed Problems, 24 (2016), 195–220.
[30] D. N. Hào and N. T. N. Oanh, Determination of the initial condition in parabolic equations
from integral observations, Inverse Problems in Science and Engineering, 25 (2017), 1138–
1167.
[31] D. N. Hào and N. Van Duc, Stability results for backward parabolic equations with time-
dependent coefficients, Inverse Problems, 27 (2011), 025003.
[32] D. N. Hào, N. Van Duc and D. Lesnic, Regularization of parabolic equations backward in
time by a non-local boundary value problem method, IMA Journal of Applied Mathematics,
75 (2010), 291–315.
[33] D. N. Hào, P. X. Thanh, B. Bin-Mohsin and N. H. Cong, Stable reconstruction of the initial
condition in parabolic equations from boundary observations, Computers & Mathematics with
Applications, 79 (2020), 3570–3587.
[34] B. Harrach and Y.-H. Lin, Monotonicity-based inversion of the fractional Schrödinger equation
I. Positive potentials, SIAM Journal on Mathematical Analysis, 51 (2019), 3092–3111.
[35] B. Harrach and Y.-H. Lin, Monotonicity-based inversion of the fractional Schödinger equation
II. General potentials and stability, SIAM Journal on Mathematical Analysis, 52 (2020),
402–436.
[36] M. V. Klibanov, Estimates of initial conditions of parabolic equations and inequalities via
lateral Cauchy data, Inverse Problems, 22 (2006), 495–514.
[37] M. V. Klibanov, Carleman estimates for the regularization of ill-posed Cauchy problems,
Applied Numerical Mathematics, 94 (2015), 46–74.
[38] M. V. Klibanov, A. V. Kuzhuget and K. V. Golubnichiy, An ill-posed problem for the Black-
Scholes equation for a profitable forecast of prices of stock options on real market data, Inverse
Problems, 32 (2015), 015010.
[39] M. V. Klibanov and A. V. Tikhonravov, Estimates of initial conditions of parabolic equations
and inequalities in infinite domains via lateral Cauchy data, Journal of Differential Equations,
237 (2007), 198–224.
[40] M. Kwasnicki, Eigenvalues of the fractional laplace operator in the interval, Journal of Func-
tional Analysis, 262 (2012), 2379–2402.
[41] M. M. Lavrentev, V. G. Romanov and S. P. Shishatskii, Ill-Posed Problems of Mathematical
Physics and Analysis, Translations of Mathematical Monographs, American Mathematical
Soc., 64 1986.
[42] T. Leonori, I. Peral, A. Primo and F. Soria, Basic estimates for solutions of a class of nonlocal
elliptic and parabolic equations, Discrete & Continuous Dynamical Systems, 35 (2015), 6031–
6068.
[43] J. Li, M. Yamamoto and J. Zou, Conditional stability and numerical reconstruction of initial
temperature, Communications on Pure & Applied Analysis, 8 (2009), 361–382.
[44] J. L. Lions, Optimal Control of Systems Governed by Partial Differential Equations,
Problèmes aux Limites, Springer, Berlin, 1971.
[45] A. Lischke, G. Pang, M. Gulian, F. Song, C. Glusa, X. Zheng, Z. Mao, W. Cai, M. M.
Meerschaert, M. Ainsworth, et al., What is the fractional laplacian? A comparative review
with new results, Journal of Computational Physics, 404 (2020), 109009.
[46] W. McLean, Strongly Elliptic Systems and Boundary Integral Equations, Cambridge univer-
sity press, 2000.
[47] R. Metzler, J. H. Jeon, A. G. Cherstvy and E. Barkai, Anomalous diffusion models and their
properties: Non-stationarity, non-ergodicity, and ageing at the centenary of single particle
tracking, Physical Chemistry Chemical Physics, 16 (2014), 24128–24164.
[48] D. Molina-Garcı́a, T. M. Pham, P. Paradisi, C. Manzo and G. Pagnini, Fractional kinetics
emerging from ergodicity breaking in random media, Physical Review E , 94 (2016), 052147.
[49] V. A. Morozov, Methods for Solving Incorrectly Posed Problems, Springer, New York, 1984.
[50] N. T. N. Oanh, A splitting method for a backward parabolic equation with time-dependent
coefficients, Computers & Mathematics with Applications, 65 (2013), 17–28.
[51] L. Plociniczak, Analytical studies of a time-fractional porous medium equation. derivation,
approximation and applications, Communications in Nonlinear Science and Numerical Sim-
ulation, 24 (2015), 169–183.
30 MOURAD HRIZI, MOHAMED BENSALAH AND MAATOUG HASSINE
[52] X. Ros-Oton and J. Serra, The Dirichlet problem for the fractional Laplacian: Regularity up
to the boundary, Journal de Mathèmatiques Pures et Appliquèes, 101 (2014), 275–302.
[53] X. Ros-Oton and J. Serra, The extremal solution for the fractional Laplacian, Calculus of
Variations and Partial Differential Equations, 50 (2014), 723–750.
[54] Y. A. Rossikhin and M. V. Shitikova, Application of fractional calculus for dynamic problems
of solid mechanics: Novel trends and recent results, Applied Mechanics Reviews, 63 (2010).
[55] L. Silvestre, Regularity of the obstacle problem for a fractional power of the Laplace opera-
tor, Communications on Pure and Applied Mathematics: A Journal Issued by the Courant
Institute of Mathematical Sciences, 60 (2007), 67–112.
[56] L. Sun and T. Wei, Identification of the zeroth-order coefficient in a time fractional diffusion
equation, Applied Numerical Mathematics, 111 (2017), 160–180.
[57] L. Tartar, An Introduction to Sobolev Spaces and Interpolation Spaces, Springer Science &
Business Media, 2007.
[58] S. Tatar and S. Ulusoy, A uniqueness result for an inverse problem in a space-time fractional
diffusion equation, Electronic Journal of Differential Equations, 258 (2013), 1–9.
[59] E. Valdinoci, From the long jump random walk to the fractional Laplacian, Boletin de la
Sociedad Espanola de Matematica Aplicada. SeMA, 49 (2009), 33–44.
[60] N. Van Thang, N. Van Duc, L. D. N. Minh and N. T. Thành, Identifying an unknown source
term in a time-space fractional parabolic equation, Applied Numerical Mathematics, 166
(2021), 313–332.
[61] T. Wei, X. L. Li and Y. S. Li, An inverse time-dependent source problem for a time-fractional
diffusion equation, Inverse Problems, 32 (2016), 085003.
[62] X. B. Yan and T. Wei, Inverse space-dependent source problem for a time-fractional diffusion
equation by an adjoint problem approach, Journal of Inverse and Ill-posed Problems, 27
(2019), 1–16.
[63] F. Yang, Q. Pu and X.-X. Li, The fractional Landweber method for identifying the space
source term problem for time-space fractional diffusion equation, Numerical Algorithms, 87
(2021), 1229–1255.
[64] Y. Zhang, X. Liu, M. R. Belic, W. Zhong, Y. Zhang and M. Xiao, Propagation dynamics of a
light beam in a fractional Schrödinger equation, Physical Review Letters, 115 (2015), 180403.
[65] Y. Zhang, M. M. Meerschaert and R. M. Neupauer, Backward fractional advection dispersion
model for contaminant source prediction, Water Resources Research, 52 (2016), 2462–2473.
[66] G.-H. Zheng and T. Wei, Recovering the source and initial value simultaneously in a parabolic
equation, Inverse Problems, 30 (2014), 065013.
Received August 2021; revised December 2021; early access February 2022.
E-mail address: mourad-hrizi@hotmail.fr
E-mail address: mohamed.bensalah@fsm.rnu.tn
E-mail address: maatoug.hassine@enit.rnu.tn