1 Introduction
1.1 Mathematical statement
The mathematical model in this work is stated as follows:
(1) 
The domain has sufficiently smooth boundary, and with denotes the DjrbashyanCaputo fractional derivative, defined as
where is the gamma function. The lower bound is set to ensure the well definedness of the Ito integral and this can be seen in the next section. The operator is an elliptic operator defined as , and may be highly heterogeneous. The source term contains the targeted unknowns , while the spatial component is given.
is the white noise derived from the Brownian motion and then
constitutes an Ito process, see [41] for details.Our data is the moments of the realizations of on a single point with the restriction , which is different from [31]. This condition will make the inverse problem more challenging in mathematics, but is meaningful in practical application. For instance, regarding equation (1) as the contaminant diffusion system, solution will be the concentration of pollutant, and is the location of pollution source, in which it is severely polluted. If the pollutant is harmful for human body, it is not allowed to observe inside the support of considering staff’s health. Reflected on mathematics, we should set the restriction , even though it will increase the difficulty. Actually, given , this inverse problem will be reduced to Volterra integral equations, and the stability result and iterative algorithm follow straightforwardly. See Remark (1) for details.
The precise mathematical description of this inverse problem is given as follows: observe , then use the statistical moments of the measurements to reconstruct simultaneously.
1.2 Physical background and literature
In microscopic level, the random motion of a single particle can be viewed as a diffusion process. The classical diffusion equation can be deduced to describe the motion of particles, if we assume the key condition, the mean squared displacement of jumps after a long time is proportional to time, i.e. . However, recently, people found some anomalous diffusion phenomena [5, 8, 17, 26], in which the assumption is violated. Sometimes it may possess the asymptotic behavior of , i.e. The different rate will lead to a reformulation to the diffusion equation, introducing the time fractional derivative in it, and the corresponding equations are called fractional differential equations (FDEs). We list some applications of FDEs, to name a few, the thermal diffusion in media with fractal geometry [39], ion transport in column experiments [19], dispersion in a heterogeneous aquifer [1], nonFickian diffusion in geological formations [6], the analysis on viscoelasticity in material science [35, 49]. [37] provides an extensive list.
If uncertainty is added in the source term, the FDE system will become more complicated and meaningful. Since it is common to meet a diffusion source, which is defined as a stochastic process to describe the uncertain character imposed by nature. As a consequence, it is worth to investigate the diffusion system with a random source. In such situation the solution will be written as a stochastic process, which makes the analysis more challenging.
In addition, to deal with the case of highly heterogeneous medium , the Generalized Multiscale Finite Element Method (GMsFEM [13]) will be used to simulate the forward problem of equation (1). The introduction of GMsFEM will be given in section (4.3.1).
For a comprehensive understanding of fractional calculus and FDEs, see [25, 45, 7] and the references therein. For inverse problems in FDEs, [24] is an extensive review. See [14, 50, 40] for inverse source and coefficient problems; see [22] for unique continuation principle; see [21]
for Carleman estimate in FDEs; see
[18, 27, 42] for fractional Calderon problem. Furthermore, if we extend the assumption to a more general case , the multiterm fractional diffusion equations and even the distributedorder differential equations will be generated, [43, 47, 30, 29]. For numerical methods for inverse problems, see [4, 3] and the references therein. Literature about the GMsFEM and its applications can be found in [13, 12, 9, 11, 10].1.3 Main result and outline
Throughout this paper, the following restrictions on spatial component , observation point , and the unknowns are supposed to be valid.
Assumption 1.

, and set such that ;

changes its sign times on and ;

and for ;

, namely, .
Now we can state the main result, which says the unknowns can be limited by some statistical moments of observations .
Theorem 1.
In this theorem, means the fractional integral operator, and
are the notations for expectation and variance, respectively. These knowledge can be seen in section (
2). Noting that the stochastic process is independent of the sign of by the properties of in section (2), sequentially we consider instead of . That’s why the norm of is estimated.The remaining part of this manuscript is structured as follows. Section (2
) includes the preliminaries, such as the probability space
and the stochastic solution for equation (1). Also, some auxiliary results like the reverse convolution inequality and the maximum principles in fractional diffusion equations are collected. In section (3), we prove Theorem (1). After that the numerical reconstruction for the unknowns is investigated in section (4). We construct the regularized LevenbergMarquardt iteration (15), and prove its convergence–Proposition (2). Some numerical results generated by iteration (15) are also displayed. Furthermore, some brief knowledge of GMsFEM is provided in this section.2 Preliminaries
2.1 Brownian motion and Ito isometry formula
To state the Ito formula, firstly we need to give the setting of probability space.
Definition 1.
We call a probability space if denotes the nonempty sample space, is the algebra of and is the probability measure.
The Brownian motion , which is also called Wiener process in mathematics, has the following properties,


has continuous paths;
Now the essential tool, Ito isometry formula can be stated.
Lemma 1.
([41]). Let be a probability space and satisfy the following properties.

is measurable, where denotes the Borel algebra on

is adapted;

for some .
Then the Ito integral , where denotes the random measure derived from , is well defined, and it follows that
2.2 Stochastic weak solution
The randomness from means that we can not differentiate in for each . As a consequence, we will define the weak solution of equation (1) in the integral sense.
Firstly, the fractional integral operator and the corresponding Ito integral are given.
Definition 2.
The fractional integral operator is defined as
Then we define as
Now we explain the necessity of the restriction . For , from the conditions and , we have is squareintegrable on Then Lemma (1) yields that the Ito integral is well defined.
In addition, the direct calculation gives that
which implies the next definition of the weak solution for equation (1).
Definition 3 (Stochastic weak solution).
The stochastic process is called as a stochastic weak solution of equation (1) if for each and , it holds that
2.3 Auxiliary lemmas
Here we list some auxiliary lemmas which will be used later. First the reverse convolution inequality is given.
Lemma 2.
([33, Lemma 3.1]). Let and be arbitrarily given. Suppose that , and on .

If keeps its sign on , then
(2) 
If only keeps its sign on , then
(3)
The next lemmas are the maximum principles in FDEs.
Lemma 3.
(Maximum principle, [34, Theorem 2]). Fix let satisfy the following fractional diffusion equation
(4) 
and define . If , then
Lemma 4.
In addition, we gives a representation lemma for the weak solution .
3 Main result
3.1 Moments representation
With Lemmas (1) and (5), the moments we used can be represented in terms of the unknowns . See the lemma below.
Lemma 6.
(7)  
Proof.
Remark 1.
However, since in this work, we can only start the analysis from (7). Due to the convolution structure, the estimates of the unknowns on the partial interval are attained. See the next subsection for details.
3.2 Proof of Theorem (1)
Proof of Theorem (1) .
Proof of Theorem (1) .
Let be small and assume that changes sign on , for convenience, we set and . By (7), for , we can write
where
Using (3) to the above equality with , we can obtain that for ,
(8)  
For , from the condition that we have
(9) 
For , it holds that
Assumption (1) and Lemma (3) give that . Consequently,
(10) 
Inserting (9) and (10) into (8) yields that
(11) 
Fix , we have
(12) 
Comments
There are no comments yet.