Next Article in Journal
Proposal of a General Identification Method for Fractional-Order Processes Based on the Process Reaction Curve
Next Article in Special Issue
Fractional Dual-Phase-Lag Model for Nonlinear Viscoelastic Soft Tissues
Previous Article in Journal
Fractional View Analysis of Swift–Hohenberg Equations by an Analytical Method and Some Physical Applications
Previous Article in Special Issue
A Mixed Finite Volume Element Method for Time-Fractional Damping Beam Vibration Problem
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Finite Element Approximations to Caputo–Hadamard Time-Fractional Diffusion Equation with Application in Parameter Identification

1
School of Mathematics, Shandong University, Jinan 250100, China
2
Department of Mathematics, University of South Carolina, Columbia, SC 29208, USA
3
School of Mathematical Sciences, Fudan University, Shanghai 200433, China
*
Author to whom correspondence should be addressed.
Fractal Fract. 2022, 6(9), 525; https://doi.org/10.3390/fractalfract6090525
Submission received: 8 August 2022 / Revised: 11 September 2022 / Accepted: 14 September 2022 / Published: 17 September 2022

Abstract

:
A finite element scheme for solving a two-timescale Hadamard time-fractional equation is discussed. We prove the error estimate without assuming the smoothness of the solution. In order to invert the fractional order, a finite-element Levenberg–Marquardt method is designed. Finally, we give corresponding numerical experiments to support the correctness of our analysis.

1. Introduction

Fractional differential equations have been shown to provide a competitive means in modeling long-time interactions, physics and engineering, see, for instance, [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17]. Compared to classical integer-order differential equations, the fractional equations provide a more desired descriptions of the diffusion due to the nature of the fractional operators. In particular, time-fractional partial differential equations are typically applied to model subdiffusion phenomena.
In this paper, we study a two-timescale Hadamard time-fractional diffusion equation which may describe the superslow diffusion as
t u + κ ( t ) a H D t α u Δ u = f ( x , t ) , ( x , t ) Ω × ( a , T ] ; u ( x , a + ) = u a ( x ) , x Ω ; u ( x , t ) = 0 , ( x , t ) Ω × [ a , T ] .
Here t u denotes the first derivative with respect to time, x : = ( x 1 , x 2 , , x d ) and the Hadamard fractional derivative is defined by [18,19,20]
a H D t α f ( t ) = 1 Γ ( 1 α ) a t log t s α f ( s ) d s
This model can simulate strongly anomalous diffusion [21,22,23], which means the mean square displacement of the model has the following form
< u ( t ) 2 > log μ t , μ > 0 .
An outline of our paper is as follows: We present some preliminaries and some notations in Section 2. In Section 3, we derive a finite element method for solving model problem (1) and present some numerical results to substantiate the mathematical and numerical analyses. In Section 4, we develop a finite-element Levenberg–Marquardt method to obtain the fractional order and give some numerical examples to show the utility of our method. We end this paper by giving a conclusion in the final section.

2. Preliminaries

We suppose that the domain Ω is a simply connected bounded domain and Ω R d ( d = 1 , 2 , 3 ) . We introduce the following notations and the corresponding norms:
w C ( Ω ) : = sup t | w ( t ) | , w C m ( I ) : = max 0 n m D n w C ( I ) .
The Sobolev space H r ( Ω ) with fractional index r is defined in [24,25]. Let H ˙ 2 ( Ω ) : = H 2 ( Ω ) H 0 1 ( Ω ) and H ˙ r ( Ω ) : = L 2 ( Ω ) , H ˙ 2 ( Ω ) r / 2 , 0 r 2 , be the complex interpolation space [25]. Then, H ˙ 0 ( Ω ) = L 2 ( Ω ) and H ˙ 1 ( Ω ) = H 0 1 ( Ω ) . The definition of the fractional Laplacian is given by [26,27]
( Δ ) s w ( x ) = C ( d , s ) 2 Ω 2 w ( x ) w ( x + y ) w ( x y ) | y | n + 2 s d y , s ( 0 , 1 ) ,
where C ( d , s ) is a dimensional constant. Let the operator L : = Δ , then the solutions { φ i } i = 1 of the following problem
L φ i ( x ) = λ i φ i ( x ) , x Ω ; φ i ( x ) = 0 , x Ω ;
constitute a basis [26]. We give the definition of the fractional Sobolev spaces [10,28]
H ˇ r ( Ω ) : = w L 2 ( Ω ) : | v | H ˇ r 2 : = ( Δ ) r w , w = i = 1 λ i r ( w , φ i ) 2 <
and the corresponding norm
w H ˇ r ( Ω ) : = w L 2 ( Ω ) 2 + | w | H ˇ γ ( Ω ) 2 1 / 2 .
The space H ˇ r ( Ω ) H r ( Ω ) [10,24,28] and
H ˇ r ( Ω ) = w H r ( Ω ) : ( Δ ) k w ( x ) = 0 , x Ω , k = 0 , 1 , , < r / 2
Hence | w | H ˇ r ( Ω ) and | w | H r ( Ω ) are equivalent in H ˇ r ( Ω ) .
We cite some theoretical results of problem (1) in [29].
Theorem 1.
Suppose κ ( t ) C [ a , T ] , | κ ( t ) | κ * for t [ 0 , T ] and α ( 0 , 1 ) . If u a H ˇ r + 2 ( Ω ) , f H ν ( H ˇ r ( Ω ) ) with r > d / 2 and ν > 1 / 2 , then problem (1) has a solution u satisfying u C 1 ( [ a , T ] ; H ˇ r ( Ω ) ) and the following holds for G = G ( α , κ C [ a , T ] , T )
u C 1 ( [ a , T ] ; H ˇ s ( Ω ) ) G u a H ˇ 2 + s ( Ω ) + | | f | | H ν ( H ˇ s ( Ω ) ) , 0 s r .
Theorem 2.
Suppose κ ( t ) C [ a , T ] , | κ ( t ) | κ * for t [ 0 , T ] and α ( 0 , 1 ) . If  u a H ˇ 4 + s ( Ω ) , f H ν ( H ˇ 2 + s ( Ω ) ) H 1 + ν ( H ˇ s ( Ω ) ) for some s 0 and ν > 1 / 2 , u C 2 ( ( a , T ] ; H ˇ s ( Ω ) ) and the following estimate holds
u C 2 ( ( a , T ] ; H ˇ s ( Ω ) ) G log t a α u a H ˇ 4 + s ( Ω ) + f H ν ( H ˇ 2 + s ( Ω ) ) + f H 1 + ν ( H ˇ s ( Ω ) ) ,
with G = G α , κ C 1 [ a , T ] , T .

3. Analysis of a Finite Element Scheme

In this section, we give a finite element method for solving (1) and analyze the convergence of the method.
Given an integer N > 0 . We discretize [ a , T ] by t n : = a + n Δ t for 0 n N with Δ t : = T a N . First, we discretize t u at t = t n by
t u ( x , t ) = u ( x , t n ) u ( x , t n 1 ) Δ t + 1 Δ t t n 1 t n t t u ( x , t ) ( t t n 1 ) d t = : δ t n u ( x , t n ) + E n ( x ) .
Next, we discretize a H D t α u ( x , t ) at time t = t n by
a H D t α u ( x , t n ) = 1 Γ ( 1 α ) k = 1 n t k 1 t k t k δ t k u ( x , t ) ( log t n log t ) α d t t + t k 1 t k ( t t u t k δ t k u ) ( x , t ) ( log t n log t ) α d t t = : δ t n α u ( x , t n ) + R n ( x ) .
where
δ t n α u ( x , t n ) : = 1 Γ ( 2 α ) k = 1 n t k ( log t n log t k 1 ) 1 α ( log t n log t k ) 1 α δ t k u ( x , t k ) = 1 Γ ( 2 α l ) k = 1 n b n , k u ( x , t k ) u ( x , t k 1 ) ,
R n : = k = 1 n R n , k : = 1 Γ ( 1 α ) k = 1 n t k 1 t k t t u t k δ t k u ( x , t ) ( log t n log t ) α d t t
The coefficients b n , k for 1 k n N are given by
b n , k : = t k ( log t n log t k 1 ) 1 α ( log t n log t k ) 1 α Δ t ,
which have the properties [30]
b n , n > b n , n 1 > > b n , k > b n , 1 > 0 .
In order to estimate the finite element error, we introduce the following Ritz projection P h : H 0 1 ( Ω ) S h ( Ω ) [31] and S h ( Ω ) H 0 1 ( Ω )
( ω P h ω ) , v = 0 , v S h ( Ω ) , for ω H 0 1 ( Ω )
and the corresponding error
ω P h ω L 2 ( Ω ) G h 2 ω H 2 ( Ω ) , ω H 2 ( Ω ) H 0 1 ( Ω ) .
Let u n : = u ( x , t n ) . We multiply Equation (1), incorporated with (5) and (6), by  w H 0 1 ( Ω ) . We give the weak form for (1) at t = t n for n = 1 , , N
δ t n u n , w + u n , w + κ ( t n ) δ t n α u n , w = ( f ( · , t n ) , w ) k ( t n ) R n , w ( E n , w ) .
We throw away the local truncation errors to arrive at a finite element scheme for (1):
δ t n u n , w + u n , w + κ ( t n ) δ t n α u n , w = ( f ( · , t n ) , w ) k ( t n ) R n , w ( E n , w ) , w S h , n = 1 , , N .

3.1. Analysis of Truncation Errors

We bound the errors E n and R n defined in (5) and (6), respectively, in this subsection.
Theorem 3.
Let the assumptions of Theorem 2 be satisfied. Then, we have the estimates for n = 1 , 2 , , N
E n L 2 ( Ω ) G G 0 n α N α 1 , R n L 2 ( Ω ) G G 0 n α N α 1 .
Here, G 0 = u a H ˇ 4 + f H γ ( H ˇ 2 ( Ω ) ) + f H 1 + γ ( L 2 ( Ω ) ) for some γ > 1 / 2 .
Proof of Theorem 3.
We use Theorem 2 to bound E n by
E n L 2 ( Ω ) G G 0 Δ t t n 1 t n log t a α ( t t n 1 ) d t G G 0 t n 1 t n log t a α d t G G 0 T t n 1 t n log t a α d t t = G G 0 T 1 α ( log t n a ) 1 α ( log t n 1 a ) 1 α = G G 0 T 1 α log ( 1 + n ( T a ) N a ) 1 α log ( 1 + ( n 1 ) ( T a ) N a ) 1 α ,
We use the Taylor expansion theorem
log ( 1 + θ ) = θ θ 2 2 ! + θ 3 3 ! + ,
to estimate (10) by
E n L 2 ( Ω ) G G 0 T 1 α log ( 1 + n ( T a ) N a ) 1 α log ( 1 + ( n 1 ) ( T a ) N a ) 1 α G G 0 T 1 α n N 1 α n 1 N 1 α G G 0 T 1 α n α N α 1 .
Next, we bound R n , 1 in (5)
R n , 1 L 2 ( Ω ) a t 1 ( log t n log t ) α t t u L 2 ( Ω ) + t 1 Δ t a t 1 t u ( · , s ) L 2 ( Ω ) d s d t t G G 0 T a t 1 ( log t n log t ) α d t t = G G 0 T 1 α log ( t n a ) 1 α log ( t n t 1 ) 1 α G G 0 T 1 α log ( t 1 a ) 1 α , n = 1 , G G 0 T 1 α log ( t n t 1 ) α log t 1 log a G G 0 ( n 1 ) α N 1 α , n > 1 .
We use the fact that for t [ t n 1 , t n ]
t t u t n δ t n u L 2 ( Ω ) t t u t n t u L 2 ( Ω ) + t n t u t n δ t n u L 2 ( Ω ) | t t n | t u L 2 ( Ω ) + t n t u δ t n u L 2 ( Ω ) Δ t G G 0 + G G 0 T Δ t u t t L 2 ( Ω ) G G 0 T Δ t log ( t n 1 a ) α
to bound R n , n for n 2 by
R n , n L 2 ( Ω ) G u C 2 ( [ t n 1 , t n ] ; L 2 ( Ω ) ) Δ t t n 1 t n ( log t n log t ) α d t t G G 0 log t n 1 a α Δ t log t n t n 1 1 α G G 0 ( n 1 ) α N α 1 N 2 α G G 0 n α N 2 α α G G 0 n α N 1 α .
We bound R n below (5) for n 3 as follows
k = n / 2 + 1 n 1 R n , k L 2 ( Ω ) G k = n / 2 + 1 n 1 u C 2 ( [ t k 1 , t k ] ; L 2 ) Δ t t k 1 t k ( log t n log t ) α d t t G G 0 log t n / 2 a α Δ t t n / 2 t n 1 ( log t n log t ) α d t t G G 0 log t n a α Δ t log t n a 1 α G G 0 n N α 1 N n N 1 α G G 0 n n N 2 α α ,
k = 2 n / 2 R n , k L 2 ( Ω ) G k = 2 n / 2 u C 2 ( [ t k 1 , t k ] ; L 2 ( Ω ) ) Δ t t k 1 t k ( log t n log t ) α d t t G G 0 k = 2 n / 2 log t k a α Δ t ( log t n t k 1 ) 1 α ( log t n t k ) 1 α ) G G 0 k = 2 n / 2 log t k a α Δ t ( log t n t k ) α ( log t k t k 1 ) G G 0 k = 2 n / 2 k α n α N 2 α α G G 0 n n N 2 α α .
Furthermore, we incorporate the estimates above to derive
G G 0 n n N 2 α α = G G 0 n α N 1 α n 1 α N 1 α G G 0 n α N 1 α
to obtain the estimate in (9).    □
Theorem 4.
Let η ( x , t ) : = ( I P h ) u ( x , t ) . Suppose that the assumptions of Theorem 2 are satisfied, then we have
δ t n η n L 2 ( Ω ) + δ t n α η n L 2 ( Ω ) G G 0 h 2 , 1 n N .
Proof of Theorem 4.
We apply (7) to obtain
δ t n η n L 2 ( Ω ) = 1 Δ t ( I P h ) t n 1 t n t u ( x , t ) d t L 2 ( Ω ) G h 2 u C 1 ( [ a , T ] ; H 2 ( Ω ) ) ,
δ t n α η n L 2 ( Ω ) = 1 Γ ( 2 α ) k = 1 n b n , k ( I P h ) t n 1 t n t u ( x , t ) d t L 2 ( Ω ) G Δ t h 2 u C 1 ( [ a , T ] ; H 2 ( Ω ) ) k = 1 n b n , k G h 2 u C 1 ( [ a , T ] ; H 2 ( Ω ) ) .
   □

3.2. Analysis of the Finite Element Method

We give the following error estimate of the finite element method (8), which could be proved by similar techniques as in Theorem 4.3 from [17], and interested readers can see the proof for more details.
Theorem 5.
Let the assumptions of Theorem 2 be satisfied, then we have the following results
u h u L ( a , T ; L 2 ( Ω ) ) : = max 1 n N u h , n u n L ( L 2 ( Ω ) ) G G 0 Δ t + h 2 .

3.3. Numerical Experiments of the Finite Element Method

We give some numerical examples to justify our numerical analysis above.
Example 1.
Let Ω = ( 0 , 1 ) , [ a , T ] = [ 1 , 1.2 ] , κ ( t ) = 1 , u a ( x ) = sin ( 2 π x ) and f ( x , t ) = 0 , and select the numerical solution u ^ with N = 2 10 and h = 1 / 32 to be a reference solution. We plot the first-order time difference quotient δ t n u h , n ( 1 / 4 , t n ) in Figure 1. It is clear that the temporal derivative of the solution exhibits a singular behavior near the initial time, which gets stronger as α increases. These findings numerically satisfy Theorem 2.
Example 2.
Let = ( 0 , 1 ) 2 , [ a , T ] = [ 1 , 2 ] , u a ( x ) = 0 and κ ( t ) = 1 . The exact solution to problem (1) is chosen to be u = ( log t ) 2 α sin ( 2 π x ) sin ( 2 π y ) and f is calculated accordingly. We measure the convergence rates κ (temporal rate) and γ (spatial rate) by
max 1 n N u n u h , n L 2 ( Ω ) Q ( N r t + M r s ) .
The numerical results listed in Table 1 and Table 2 show that the finite element method retains a first-order accuracy in time and a second-order accuracy in space as proved in Theorem 5.

4. An Inversion Algorithm to Evaluate the Fractional Order α

In many practical scenario problems, the fractional order α in model problem (1) is not clear. Therefore, we must use physical experiments to obtain the data.

4.1. L–M Regularization Method

We discuss an algorithm to simulate the fractional order α with the help of the finite element method (8) as follows: given the observation data { θ i } i = 1 N , the goal of the parameter identification of α is to find α i n v satisfying
α i n v = arg min α ( 0 , 1 ) F ( α ) : = 1 2 i = 1 N u ( x i , T ; α ) θ i 2 .
In order to solve minimization problem (11), one can use an iterative algorithm such as Newton’s method. We used the first and second derivatives of the function F ( α ) for minimizing (11)
α k + 1 = α k F ( α k ) F ( α k ) ,
Here, k represents the kth iteration. Actually, relation (12) is equivalent to solving
α k + 1 = α k ( J k J k ) 1 J k r k ,
where r k = ( r 1 , r 2 , , r N ) and r i = u ( x i , T ; α ) θ i
J k = u x 1 , T ; α α , , u x N , T ; α α R N .
We note that, in order to calculate the derivative u x i , T ; α α , we can use the finite difference approximation
u ( x i , T ; α + δ ) u ( x i , T ; α ) δ
with a small δ to approximate the derivatives. However, the Newton method may sometimes fail to work due to J k J k 0 . Usually, for dealing with this problem, we can apply the Levenberg–Marquardt algorithm to minimize (11) by
α k + 1 = α k J k J k + γ k 1 J k r k ,
where γ k is a positive penalty parameter. Thus, we give the program in Algorithm 1.
Algorithm 1:(A Levenberg–Marquardt Algorithm): Given the initial data, the boundary information and the observation data θ .
1:
For an initial guess α 0 and ρ ( 0 , 1 ) , σ ( 0 , 1 2 ) , γ 0 > 0 and a δ 1 .
2:
for   k = 1 , ,   Iterations do
3:
   Solve model problem (1) corresponding to α k and α k + δ , respectively, to get       u r ( · , T ; α k ) and u r ( · , T ; α k + δ ) .
4:
   Use formula (14) to calculate Jacobian J k and J k r k and update the search direction d k : = J k J k + γ k 1 J k r k .
5:
   Identify the search step ρ m by the following Armijo rule:
F α k + ρ m d k F α k + σ ρ m d k J k r k .
6:
   If ρ m d k tolerance , then stop and let α i n v : = α k . Otherwise update       α k + 1 : = α k + ρ m d k , γ k + 1 : = γ k / 2 , and go to Step 3.
7:
end for

4.2. Numerical Experiment of the Finite Element Levenberg–Marquardt Method

Next, we ran some numerical examples to test the utility of the inverse method in model problem (1). Let α * be the true order of model (1). Let α 0 be an initial guess for the optimization, α i n v be the approximation order and “Iter” be the number of iterations. We took the observation data θ to be the exact solution of the problem (1) at T with fractional order α * .
Test 1.
In this test, we considered the data for the uncontaminated case.
Let κ ( t ) = 1 , the exact solution can be chosen as
u ( x , t ) = ( log t ) 2 α sin ( 2 π x ) ,
and the right-hand side can be computed accordingly. In Algorithm 1, we set ρ = 0.75 , σ = 0.25 , δ = 10 4 and the tolerance as 10 10 . We chose N = M = 100 in the finite element method with different initial guesses α 0 ( 0 , 1 ) in the LM algorithm. We present the output α i n v and approximation error | α i n v α * | and we also give the total number of iterations inTable 3.
In Figure 2, Figure 3 and Figure 4, we plotted the variation of the errors and the cost functions with the number of iterations. From the figures, we can see that the proposed finite element LM method achieved a desired approximation α * for different initial guesses.
Test 2.
Because there may be noise in real problems, we tested cases in which the data were contaminated as follows:
θ ϵ ( x i ) = θ ( x i ) 1 + ϵ % · randn ( i ) , i = 1 , , M .
where ϵ represents the degree of noise, randn denotes the random noise generated by the Gaussian distribution. The exact solution, initial condition and source term in this Test were the same as those in Test 1. We listed the numerical results in Table 4.
In Figure 5, Figure 6 and Figure 7, we also present the variation of errors and values of the cost function with iterations under different degrees of the noise. From the figures, we can see that our algorithm can obtain satisfactory results with a smaller number of iterations.
Test 3.
For the next test, we considered the two-dimensional case. Let κ ( t ) = 1 , the exact solution was chosen to be
u ( x , y , t ) = ( log t ) 2 α sin ( 2 π x ) sin ( 2 π y ) ,
and the right-hand side can be computed accordingly. The other data were the same as in Test 1. We chose N = 100 and h x = h y = 1 / 100 in the finite element method with different initial guesses α 0 ( 0 , 1 ) in the LM algorithm. We present the numerical results in Table 5.
We also plotted the change of parameter errors and values of the cost function with iterations in Figure 8, Figure 9 and Figure 10 and observed that the optimization process took only a few iterations to reach the tolerance.
Test 4.
We also considered two-dimensional cases in which the data had a small random perturbation as follows:
θ ϵ ( x i , y j ) = θ ( x i , y j ) 1 + ϵ % · randn ( i , j ) , i , j = 1 , , M .
The exact solution, initial condition and source term in this Test were the same as those in Test 3. The numerical results are listed inTable 6.
From Figure 11, Figure 12 and Figure 13, we can see that when the observation data g are contaminated by random noise, the algorithm can still output satisfactory results in the two-dimensional case.

5. Conclusions

We developed a finite element scheme for the numerical solution of two-timescale fractional diffusion Equation (1) involving a so-called Hadamard time-fractional derivative. We also proved the optimal convergence order of the error estimates of the finite element scheme to (1) without smooth assumptions on the true solution. Then, we presented several numerical examples to substantiate the mathematical and numerical analyses. We accordingly developed a finite element Levenberg–Marquardt algorithm to simulate the fractional order. In the near future, we will consider the variable-order case of the model problem (1).

Author Contributions

S.C.: conceptualization, methodology, validation, writing—original draft preparation; N.D.: conceptualization, writing—review and editing, funding acquisition; H.W.: conceptualization, writing—review and editing, funding acquisition; Z.Y.: conceptualization, methodology, validation, writing—original draft preparation. All authors have read and agreed to the published version of the manuscript.

Funding

This research was partially funded by the ARO MURI grant W911NF-15-1-0562, by the National Science Foundation under grant DMS-2012291, by the National Natural Science Foundation of China under grants 11971272, 12071262, 12131014, 11831010, and the Natural Science Foundation of Shandong Province, China (ZR2020MA048).

Data Availability Statement

Data sharing not applicable to this article as no dataset was generated or analyzed during the current study.

Acknowledgments

The authors would like to express their most sincere thanks to the referees for their very helpful comments and suggestions, which greatly improved the quality of this paper.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Atangana, A. On the stability and convergence of the time-fractional variable order telegraph equation. J. Comput. Phys. 2015, 293, 104–114. [Google Scholar] [CrossRef]
  2. Bai, X.; Rui, H. A second-order space-time accurate scheme for Maxwell’s equations in a Cole–Cole dispersive medium. Eng. Comput. 2022, 1, 1–20. [Google Scholar] [CrossRef]
  3. Bai, X.; Huang, J.; Rui, H.; Wang, S. Numerical simulation for 2D/3D time fractional Maxwell’s system based on a fast second-order FDTD algorithm. J. Comput. Appl. Math. 2022, 416, 114590. [Google Scholar] [CrossRef]
  4. Ma, J.; Gao, F.; Du, N. Stabilizer-free weak Galerkin finite element method with second-order accuracy in time for the time fractional diffusion equation. J. Comput. Appl. Math. 2022, 414, 114407. [Google Scholar] [CrossRef]
  5. Gu, X.; Sun, H.; Zhao, Y.; Zheng, X. An implicit difference scheme for time-fractional diffusion equations with a time-invariant type variable order. Appl. Math. Lett. 2021, 120, 107270. [Google Scholar] [CrossRef]
  6. Gu, X.; Wu, S. A parallel-in-time iterative algorithm for Volterra partial integro-differential problems with weakly singular kernel. J. Comput. Phys. 2020, 417, 109576. [Google Scholar] [CrossRef]
  7. Jia, J.; Wang, H. A preconditioned fast finite volume scheme for a fractional differential equation discretized on a locally refined composite mesh. J. Comput. Phys. 2015, 299, 842–862. [Google Scholar] [CrossRef]
  8. Jia, J.; Wang, H. A fast finite volume method on locally refined meshes for fractional diffusion equations. East Asian J. Appl. Math. 2019, 9, 755–779. [Google Scholar]
  9. Liu, Z.; Li, X.; Zhang, X. A fast high-order compact difference method for the fractal mobile/immobile transport equation. Int. J. Comput. Math. 2020, 97, 1860–1883. [Google Scholar] [CrossRef]
  10. Luchko, Y. Initial-boundary-value problems for the one-dimensional time-fractional diffusion equation. Fract. Calc. Appl. Anal. 2012, 15, 141–160. [Google Scholar] [CrossRef]
  11. Metzler, R.; Klafter, J. The random walk’s guide to anomalous diffusion: A fractional dynamics approach. Phys. Rep. 2000, 339, 1–77. [Google Scholar] [CrossRef]
  12. Schumer, R.; Benson, D.A.; Meerschaert, M.M.; Baeumer, B. Fractal mobile/immobile solute transport. Water Resour. Res. 2003, 39, 1–12. [Google Scholar] [CrossRef]
  13. Zheng, X.; Wang, H. An error estimate of a numerical approximation to a hidden-Memory variable-order space-time fractional diffusion equation. SIAM J. Numer. Anal. 2020, 58, 2492–2514. [Google Scholar] [CrossRef]
  14. Zheng, X.; Wang, H. A hidden-memory variable-order fractional optimal control model: Analysis and approximation. SIAM J. Control Optim. 2021, 59, 1851–1880. [Google Scholar] [CrossRef]
  15. Zheng, X. Logarithmic transformation between (variable-order) Caputo and Caputo-Hadamard fractional problems and applications. Appl. Math. Lett. 2021, 121, 107366. [Google Scholar] [CrossRef]
  16. Zheng, X.; Wang, H. Optimal-order error estimates of finite element approximations to variable-order time-fractional diffusion equations without regularity assumptions of the true solutions. IMA J. Numer. Anal. 2021, 41, 1522–1545. [Google Scholar] [CrossRef]
  17. Yang, Z.; Zheng, X.; Wang, H. A variably distributed-order time-fractional diffusion equation: Analysis and approximation. Comput. Methods Appl. Mech. Eng. 2020, 367, 113118. [Google Scholar] [CrossRef]
  18. Kilbas, A.A.; Srivastava, H.M.; Trujillo, J.J. Theory and Applications of Fractional Differential Equations; Elsevier: Amsterdam, The Netherlands, 2006; Volume 204. [Google Scholar]
  19. Ma, L.; Li, C. On Hadamard fractional calculus. Fractals 2017, 25, 1750033. [Google Scholar] [CrossRef]
  20. Hadamard, J. Essai sur l’étude des fonctions, données par leure développement de Taylor. J. Pure Appl. Math. 1892, 4, 101–186. [Google Scholar]
  21. Denisov, S.I.; Kantz, H. Continuous-time random walk theory of superslow diffusion. Europhys. Lett. 2010, 92, 30001. [Google Scholar] [CrossRef]
  22. Dräger, J.; Klafter, J. Strong anomaly in diffusion generated by iterated maps. Phys. Rev. Lett. 2000, 84, 5998. [Google Scholar] [CrossRef] [PubMed]
  23. Yang, Z.; Zheng, X.; Wang, H. Well-posedness and regularity of Caputo—Hadamard fractional stochastic differential equations. Z. Angew. Math. Phys. 2021, 72, 141. [Google Scholar] [CrossRef]
  24. Adams, R.A.; Fournier, J.J.F. Sobolev Spaces; Elsevier: San Diego, CA, USA, 2003. [Google Scholar]
  25. Bennett, C.; Sharpley, R.C. Interpolation of Operators; Academic Press: Orlando, FL, USA, 1988. [Google Scholar]
  26. Evans, L.C. Partial Differential Equations; Graduate Studies in Mathematics, V 19; American Mathematical Society: Providence, RI, USA, 1998. [Google Scholar]
  27. Servadei, R.; Valdinoci, E. On the spectrum of two different fractional operators. Proc. R. Soc. Edinb. Sect. A. 2014, 144, 831–855. [Google Scholar] [CrossRef]
  28. Sakamoto, K.; Yamamoto, M. Initial value/boundary value problems for fractional diffusion-wave equations and applications to some inverse problems. J. Math. Anal. Appl. 2011, 382, 426–447. [Google Scholar] [CrossRef]
  29. Yang, Z.; Zheng, X.; Wang, H. Well-posedness and regularity of Caputo-Hadamard time-fractional diffusion equations. Fractals 2022, 30, 2250005. [Google Scholar] [CrossRef]
  30. Li, C.; Li, Z.; Wang, Z. Mathematical analysis and the local discontinuous Galerkin method for Caputo—Hadamard fractional partial differential equation. J. Sci. Comput. 2020, 85, 1–27. [Google Scholar] [CrossRef]
  31. Wheeler, M.F. A priori L2 error estimates for Galerkin approximations to parabolic partial differential equations. SIAM Numer Anal. 1973, 10, 723–759. [Google Scholar] [CrossRef]
Figure 1. Plots of δ t n u h , n ( 1 4 , t n ) : α = 0.1 , 0.5 , 0.9 .
Figure 1. Plots of δ t n u h , n ( 1 4 , t n ) : α = 0.1 , 0.5 , 0.9 .
Fractalfract 06 00525 g001
Figure 2. α * = 0.3 for uncontaminated observation data in Test 1.
Figure 2. α * = 0.3 for uncontaminated observation data in Test 1.
Fractalfract 06 00525 g002
Figure 3. α * = 0.6 for uncontaminated observation data in Test 1.
Figure 3. α * = 0.6 for uncontaminated observation data in Test 1.
Fractalfract 06 00525 g003
Figure 4. α * = 0.9 for uncontaminated observation data in Test 1.
Figure 4. α * = 0.9 for uncontaminated observation data in Test 1.
Fractalfract 06 00525 g004
Figure 5. α * = 0.75 for contaminated observation data with 0.01 % of noise in Test 2.
Figure 5. α * = 0.75 for contaminated observation data with 0.01 % of noise in Test 2.
Fractalfract 06 00525 g005
Figure 6. α * = 0.75 for contaminated observation data with 0.1 % of noise in Test 2.
Figure 6. α * = 0.75 for contaminated observation data with 0.1 % of noise in Test 2.
Fractalfract 06 00525 g006
Figure 7. α * = 0.75 for contaminated observation data with 1 % of noise in Test 2.
Figure 7. α * = 0.75 for contaminated observation data with 1 % of noise in Test 2.
Fractalfract 06 00525 g007
Figure 8. α * = 0.3 for uncontaminated observation data in Test 3.
Figure 8. α * = 0.3 for uncontaminated observation data in Test 3.
Fractalfract 06 00525 g008
Figure 9. α * = 0.6 for uncontaminated observation data in Test 3.
Figure 9. α * = 0.6 for uncontaminated observation data in Test 3.
Fractalfract 06 00525 g009
Figure 10. α * = 0.9 for uncontaminated observation data in Test 3.
Figure 10. α * = 0.9 for uncontaminated observation data in Test 3.
Fractalfract 06 00525 g010
Figure 11. α * = 0.75 for true data with 0.01 % of noise in Test 4.
Figure 11. α * = 0.75 for true data with 0.01 % of noise in Test 4.
Fractalfract 06 00525 g011
Figure 12. α * = 0.75 for true data with 0.1 % of noise in Test 4.
Figure 12. α * = 0.75 for true data with 0.1 % of noise in Test 4.
Fractalfract 06 00525 g012
Figure 13. α * = 0.75 for true data with 1 % of noise in Test 4.
Figure 13. α * = 0.75 for true data with 1 % of noise in Test 4.
Fractalfract 06 00525 g013
Table 1. Temporal convergence orders under different α .
Table 1. Temporal convergence orders under different α .
α = 0 . 1 α = 0 . 5 α = 0 . 9
N M = 2 10 r t M = 2 10 r t M = 2 10 r t
324.61  × 10 5 2.12  × 10 5 8.64  × 10 5
642.27  × 10 5 1.021.11  × 10 5 0.944.19  × 10 5 1.04
1281.13  × 10 5 1.015.79  × 10 6 0.942.07  × 10 5 1.02
2565.64  × 10 6 1.002.97  × 10 6 0.971.03  × 10 5 1.01
Table 2. Spatial convergence orders under different α .
Table 2. Spatial convergence orders under different α .
α = 0 . 1 α = 0 . 5 α = 0 . 9
M N = 2 12 r s N = 2 12 r s N = 2 12 r s
87.54  × 10 4 1.55  × 10 3 2.00  × 10 3
161.93  × 10 4 1.972.69  × 10 4 1.975.11  × 10 4 1.97
324.85  × 10 5 1.996.77  × 10 5 1.991.28  × 10 4 2.00
641.22  × 10 5 2.001.69  × 10 5 2.003.10  × 10 5 2.04
Table 3. Numerical observation of different α * = 0.3 , 0.6 , 0.9 with different initial guesses α 0 = 0.2 , 0.5 , 0.8 in Test 1.
Table 3. Numerical observation of different α * = 0.3 , 0.6 , 0.9 with different initial guesses α 0 = 0.2 , 0.5 , 0.8 in Test 1.
α * α 0 α i n v α * α i n v Iter.
0.2 2.9997  × 10 1 3.4048  × 10 5 11
0.3 0.5 2.9997  × 10 1 3.3971  × 10 5 11
0.8 2.9997  × 10 1 3.3984  × 10 5 12
0.2 5.9995  × 10 1 5.1359  × 10 5 12
0.6 0.5 5.9995  × 10 1 5.1364  × 10 5 11
0.8 5.9995  × 10 1 5.1287  × 10 5 11
0.9 0.2 8.9994  × 10 1 5.8136  × 10 5 12
0.5 8.9994  × 10 1 5.8134  × 10 5 12
0.8 8.9994  × 10 1 5.8135  × 10 5 11
Table 4. Numerical results of different α * = 0.75 with ϵ % -degree noise-contaminated data in Test 2.
Table 4. Numerical results of different α * = 0.75 with ϵ % -degree noise-contaminated data in Test 2.
ϵ % α 0 α i n v α * α i n v Iter.
0.2 7.4999  × 10 1 1.1803  × 10 5 12
0.01 0.5 7.4992  × 10 1 7.6152  × 10 5 12
0.8 7.4967  × 10 1 3.3354  × 10 4 12
0.2 7.5078  × 10 1 7.8329  × 10 4 12
0.1 0.5 7.5104  × 10 1 1.0446  × 10 3 12
0.8 7.5257  × 10 1 2.5709  × 10 3 12
1 0.2 7.3639  × 10 1 1.3607  × 10 2 12
0.5 7.4047  × 10 1 9.5338  × 10 3 12
0.8 7.4961  × 10 1 3.9271  × 10 4 12
Table 5. Numerical results of different α * = 0.3 , 0.6 , 0.9 with different initial guesses α 0 = 0.2 , 0.5 , 0.8 in Test 3.
Table 5. Numerical results of different α * = 0.3 , 0.6 , 0.9 with different initial guesses α 0 = 0.2 , 0.5 , 0.8 in Test 3.
α * α 0 α i n v α * α i n v Iter.
0.2 3.008  × 10 1 8.4911  × 10 4 11
0.3 0.5 3.008  × 10 1 8.4911  × 10 4 11
0.8 3.008  × 10 1 8.4911  × 10 4 12
0.2 6.008  × 10 1 8.3445  × 10 4 12
0.6 0.5 6.008  × 10 1 8.3445  × 10 4 11
0.8 6.008  × 10 1 8.3445  × 10 4 11
0.9 0.2 9.008  × 10 1 8.3297  × 10 4 12
0.5 9.008  × 10 1 8.3297  × 10 4 12
0.8 9.008  × 10 1 8.3297  × 10 4 11
Table 6. Numerical observation of different α * = 0.75 with data contaminated with ϵ % of noise in Test 4.
Table 6. Numerical observation of different α * = 0.75 with data contaminated with ϵ % of noise in Test 4.
ϵ % α 0 α i n v α * α i n v Iter.
0.2 7.5085  × 10 1 8.5595  × 10 4 12
0.01 0.5 7.5082  × 10 1 8.2265  × 10 4 12
0.8 7.5081  × 10 1 8.1000  × 10 4 12
0.2 7.5089  × 10 1 8.9797  × 10 4 12
0.1 0.5 7.5076  × 10 1 7.6190  × 10 4 12
0.8 7.5102  × 10 1 1.0166  × 10 3 12
1 0.2 7.5415  × 10 1 4.1538  × 10 3 12
0.5 7.4991  × 10 1 8.5762  × 10 5 12
0.8 7.5305  × 10 1 3.0522  × 10 3 12
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Cheng, S.; Du, N.; Wang, H.; Yang, Z. Finite Element Approximations to Caputo–Hadamard Time-Fractional Diffusion Equation with Application in Parameter Identification. Fractal Fract. 2022, 6, 525. https://doi.org/10.3390/fractalfract6090525

AMA Style

Cheng S, Du N, Wang H, Yang Z. Finite Element Approximations to Caputo–Hadamard Time-Fractional Diffusion Equation with Application in Parameter Identification. Fractal and Fractional. 2022; 6(9):525. https://doi.org/10.3390/fractalfract6090525

Chicago/Turabian Style

Cheng, Shijing, Ning Du, Hong Wang, and Zhiwei Yang. 2022. "Finite Element Approximations to Caputo–Hadamard Time-Fractional Diffusion Equation with Application in Parameter Identification" Fractal and Fractional 6, no. 9: 525. https://doi.org/10.3390/fractalfract6090525

APA Style

Cheng, S., Du, N., Wang, H., & Yang, Z. (2022). Finite Element Approximations to Caputo–Hadamard Time-Fractional Diffusion Equation with Application in Parameter Identification. Fractal and Fractional, 6(9), 525. https://doi.org/10.3390/fractalfract6090525

Article Metrics

Back to TopTop