An e ﬃ cient numerical method for a time-fractional telegraph equation

: In this paper a time-fractional telegraph equation is considered. First the time-fractional telegraph equation is transformed into an integral-di ﬀ erential equation with a weakly singular kernel. Then an integral-di ﬀ erence discretization scheme on a graded mesh is developed to approximate the integral-di ﬀ erential equation. The possible singularity of the exact solution is taken into account in the convergence analysis. It is proved that the scheme is second-order convergent for both the spatial discretization and the time discretization. Numerical experiments conﬁrm the validity of the theoretical results.

There are a few numerical methods to solve the TFTE. Wang [9] developed a method in the reproducing kernel space by piecewise technique to solve the TFTE. Hosseini et al. [10] applied the radial basis functions and finite difference method to solve the TFTE. Kumar et al. [11] described a finite difference scheme for the generalized TFTE. Hashemi and Baleanu [12] utilized a combination of method of line and group preserving scheme to solve the TFTE. Wei et al. [13] discussed a fully discrete local discontinuous Galerkin finite element method for the TFTE. Sweilam [14] introduced the Sinc-Legendre collocation method for solving the TFTE. Mollahasani et al. [15] presented an operational method based on hybrid functions of Legendre polynomials and Block-Pulse-Functions for solving the TFTE. Hafez and Youssri [16] used a shifted Jacobi collocation scheme for a multidimensional TFTE. Bhrawy et al. [17] proposed an accurate and efficient spectral algorithm for the numerical solution of the two-sided space-time TFTE with three types of non-homogeneous boundary conditions. Youssri et al. [18,19] also developed numerical spectral Legendre approaches to solve the TFTEs. Bhrawy and Zaky [20] used a method based on the Jacobi tau approximation for solving multi-term time-space TFTE. In [2,3] B-spline collocation methods were applied to solve the TFTEs. In [21,22] wavelet methods were used to solve the TFTEs. But these papers only study the case that the solutions of the TFTEs are sufficiently smooth. Special treatment techniques, such as graded meshes [23][24][25][26][27] and mapped basis functions [28] reflecting the characteristics of the singularity of the exact solution, need to be used to solve the fractional differential equations.
The aim of the present study is twofold. The first aim is to construct an integral-difference discretization scheme on a graded mesh to approximate the integral-differential equation with a weakly singular kernel transformed from the TFTE, which is a second-order convergence discretization scheme. The second aim is to take the possible singularity of the exact solution into account in the convergence analysis, where the singularity of the exact solution is reflected as

4)
∂ l+m u(x, t) ∂t l ∂x m ≤ C 1 + t γ−l , l = 0, 1, 2, 3, and m = 0, 1, 2, (1.5) which can be referred in the literature [25][26][27]29] for details. We show that our discretization scheme on a graded mesh is second-order convergent for both the spatial discretization and the time discretization, although the exact solution of the TFTE may have singularity. Numerical experiments confirm the effectiveness of the theoretical results, and also verify that this scheme is more accurate than the methods given in [2,3].
Remark 1.1 If b, c are variable coefficients and a is a variable coefficient with respect to x , the TFTE (1.1) can also be transformed into the same integral-differential equation as given in (1.2). If a is a variable coefficient with respect to t, we first approximate a with a piecewise linear interpolation function about t and then transform the approximate TFTE into an integral-differential equation as given in (1.2). If there is a nonlinear term in the TFTE, we first transform it into a linear equation by Newton iterative method and then transform it into an integral-differential equation as given in (1.2).
Notation. Throughout the paper, C is used to indicate the positive constant independent of the mesh, and C in different places can represent different values. In order to simplify the notation, the representation g j i = g(x i , t j ) is introduced for any function g at the mesh point (x i , t j ) ∈ (0, 1) × (0, T ].

Discretization scheme
We developed the discretization scheme on a graded meshΩ N, and time sizes t j = t j − t j−1 for 1 ≤ j ≤ K, where the discretization parameters N and K are positive integers. The following integral-difference discretization scheme where U j i denotes the numerical solution at the mesh point x i , t j and δ 2 . Thus, our discretization scheme for problem (1.2) is Then, from (1.2) and (2.2) we know that the error mesh function z satisfies Then, under the assumption (1.3) we can obtain the regularities (1.4) and (1.5), the assumption (1.3), the graded mesh (2.1), the inequality (see [30]) for α < 0, and the following estimates The following lemma gives a useful result for our convergence analysis, which is given in [31]. Lemma 2.1 Assume that {ω k } ∞ k=0 be a sequence of non-negative real numbers satisfying Then for any integer K > 0 and vector (V 1 , V 2 , . . . , V K ) T ∈ R K , the following inequality holds true. Next we show that the sequence β j,k j k=1 satisfies the conditions in Lemma 2.1 by using the technique in [32], where β j, j−k = t j ξ j,k+1 + η j,k with ξ j, j+1 = 0.
A modified Grönwall inequality given in [33,Lemma 3.3] also is needed in the convergence analysis of the scheme. Lemma 2.3 Suppose that α, C 0 , T > 0 and d k, j = C 0 t k+1 t j − t k α−1 (k = 0, 1, . . . , j − 1) for 0 = t 0 < t 1 < · · · < t K = T and j = 1, 2, . . . , K, where t k+1 = t k+1 − t k . Assume that g 0 is positive and the sequence ϕ j satisfies For analyzing the stability and estimating the error for the discrete scheme (2.2) we introduce the discrete inner product and discrete L 2 -norm · as follows where v and w are two mesh functions. We also introduce the notation δ By using the technique given in [34, Theorem 3.2] we will derive the following stability result of the discrete scheme (2.2).
Theorem 2.4 Let U j i |0 ≤ i ≤ N, 1 ≤ j ≤ K be the solution of the discrete scheme (2.2). Then the numerical solution U satisfies the following estimates where C is a positive constant independent of N and K.
Proof. First we rewrite (2.2) as the following equation for 1 ≤ i < N and 1 ≤ j ≤ K. Then, taking the inner produce of (2.12) with U j i , we can get By recursion, we also can get the following equations and the equality Adding up the above inequalities we can obtain where we have used Lemmas 2.1 and 2.2. From the above inequality we have where ξ j−1, j = 0. For each j, there exists j * (1 ≤ j * ≤ j) such that Since j in (2.13) is any integer from 0 to N, we have Combining the above inequality with Lemma 2.2 we have for 1 ≤ j * ≤ j ≤ K. Furthermore, we have where we have used the mean value theorem with θ k ∈ (t k−1 , t k ) and the inequality n ≤ 2 (n − 1) for n ≥ 2. Combining (2.14) and (2.15) we can get where Lemma 2.3 has been used. This inequality implies the theorem holds true.
Next we derive the error estimates for the integral-difference discretization scheme (2.2). Theorem 2.5 Under the regularity conditions (1.4) and (1.5) and the assumption (1.3), we have the following error estimates where C is a positive constant independent of N and K.
Proof. Applying Theorem 2.4 to the error equation (2.5) we can derive where we have used the estimates (2.6). From this we complete the proof.

Numerical experiments
In this section we present some numerical results to indicate experimentally the efficiency and accuracy of the integral-difference discretization scheme. Errors and convergence rates for the integraldifference discretization scheme are presented for two examples.
We measure the accuracy in the discrete L 2 -norm and L ∞ -norm and the convergence rate respectively. In order to further confirm that the convergence rate in the time direction is consistent with the theoretical convergence rate, we also measure the convergence rates by fixing a large N as followsr respectively. The numerical results for Example 4.1 are tabulated in Tables 1 and 2.
We also measure the accuracy in the discrete L 2 -norm e N,K L 2 , L ∞ -norm e N,K L ∞ and the convergence rates r N,K L 2 , r N,K L ∞ ,r N,K L 2 ,r N,K L ∞ as previously defined, respectively. The numerical results for Example 4.2 are tabulated in Tables 3 and 4.  4 show that the numerical solution of the integral-difference discretization scheme on a graded mesh converges to the exact solution with second-order accuracy for both the spatial discretization and the time discretization in the discrete L 2 -norm and in the discrete L ∞ -norm, respectively. Moreover, compared with the previous methods with only first-order convergence for the time discretization, our discretization scheme improves the previous results given in [2,3].

Conclusions and discussion
In this paper, the TFTE is transformed into an equivalent integral-differential equation with a weakly singular kernel by using the integral transformation. Then an integral-difference discretization scheme on a graded mesh is developed to approximate the integral-differential equation. The stability and convergence are proved by using the L 2 -norm. The possible singularity of the exact solution is taken into account in the convergence analysis. It is shown that the scheme is second-order convergent for both the spatial discretization and the time discretization. The numerical experiments demonstrate the validity of our theoretical results and also verify that this scheme is more accurate than the methods given in [2,3]. In future we will extend this method to variable-order fractional differential equations.