Abstract
Tempered fractional diffusion equations (TFDEs) involving tempered fractional derivatives on the whole space were first introduced in Sabzikar et al. (J Comput Phys 293:14–28, 2015), but only the finite-difference approximation to a truncated problem on a finite interval was proposed therein. In this paper, we rigorously show the well-posedness of the models in Sabzikar et al. (2015), and tackle them directly in infinite domains by using generalized Laguerre functions (GLFs) as basis functions. We define a family of GLFs and derive some useful formulas of tempered fractional integrals/derivatives. Moreover, we establish the related GLF-approximation results. In addition, we provide ample numerical evidences to demonstrate the efficiency and “tempered” effect of the underlying solutions of TFDEs.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The normal diffusion equation \(\partial _t p(x,t)=\partial _x^2p(x,t)\) can be derived from the Brownian motion which describes the particle’s random walks. Over the last few decades, a large body of literature has demonstrated that anomalous diffusion, in which the mean square variance grows faster (super-diffusion) or slower (sub-diffusion) than in a Gaussian process, offers a superior fit to experimental data observed in many important practical applications, e.g., in physical science [14, 17,18,19], finance [10, 16, 24], biology [4, 12] and hydrology [3, 7, 8]. The anomalous diffusion equation takes the form
where \(0<\nu \le 1\) and \(0<\mu < 2\) (cf. [17] for a review on this subject), whose solution exhibits heavy tails, i.e., power law decays at infinity. In order to “temper” the power law decay, the authors of [22] incorporated an exponential factor \(e^{-\lambda |x|}\) into the particle jump density, and showed that the Fourier transform of the tempered probability density function p(x, t) takes the form
where \(0\le p\le 1,~q=1-p\), D is a constant and
Moreover, they defined tempered fractional derivative operators \(\partial _{\pm ,x}^{\mu ,\lambda }\) through Fourier transform: \(\mathscr {F}{[\partial _{\pm ,x}^{\mu ,\lambda } u]}(\omega )=A_{\pm }^{\mu ,\lambda }(\omega ){\mathscr {F}}[u](\omega )\), and derived the tempered fractional diffusion equation:
It is believed that tempered anomalous diffusion models have advantages over the normal diffusion models in some applications in geophysics [15, 30] and finance [5].
It is challenging to numerically solve the tempered fractional diffusion equation (1.3), partially due to (i) the non-local nature of tempered fractional derivatives; and (ii) the unboundedness of the domain. In [22], a finite-difference method was applied to (1.3) on a truncated (finite) interval. In [28], the authors considered tempered derivatives on a finite interval and derived an efficient Petrov–Galerkin method for solving tempered fractional ODEs by using the eigenfunctions of tempered fractional Sturm–Liouville problems. In [11], the authors used Laguerre functions to approximate the substantial fractional ODEs, which are similar to those we consider in Sect. 3, on the half line. In order to avoid the difficulty of assigning boundary conditions at the truncated boundary, we shall deal with the unbounded domain directly in this paper.
Since the tempered fractional diffusion equation is derived from the random walk on the whole line, one is tempted to use Hermite polynomials/functions which are suitable for many problems on the whole line [25]. Unfortunately, due to the exponential factor in the tempered fractional derivatives, Hermite polynomials/functions are not suitable basis functions. Instead, as we will show in Sect. 3, properly defined GLFs enjoy particularly simple form under the action of tempered fractional derivatives, just as the relations between generalized Jacobi functions and usual fractional derivatives [6]. Hence, the main goal of this paper is to design efficient spectral methods using GLFs as basis functions to solve the tempered fractional diffusion equation (1.3) in various situations. However, Laguerre polynomials/functions are mutually orthogonal on the half line, how do we use them to deal with (1.3) on the whole line? We shall first consider special cases of (1.3) with \(p=1, \, q=0\) or \(p=0, \, q=1\). In these cases, we can reduce (1.3) to the half line, and the GLFs can be naturally used. For the general case, we shall employ a two-domain spectral-element method, and use GLFs as basis functions on each subdomain.
The rest of the paper is organized as follows. In the next section, we present the definition of tempered fractional derivatives, and recall some useful properties of Laguerre polynomials. In Sect. 3, we define a class of generalized Laguerre functions, study its approximation properties, and apply it for solving simple one sided tempered fractional equations. In Sect. 4, we develop a spectral-Galerkin method for solving a tempered fractional diffusion equation on the half line. Finally, we present a spectral-Galerkin method for solving the tempered fractional diffusion equation on the whole line in Sect. 5. Some concluding remarks are given in the last section.
2 Preliminaries
Let \({\mathbb {N}}\) and \({\mathbb {R}}\) be respectively the sets of positive integers and real numbers. We further denote
2.1 Usual (Non-tempered) Fractional Integrals and Derivatives
Recall the definitions of the fractional integrals and fractional derivatives in the sense of Riemann–Liouville (see e.g., [20]).
Definition 2.1
(Riemann–Liouville fractional integrals and derivatives) For \(a,b\in {\mathbb {R}} \) or \(a=-\infty , b=\infty ,\) and \(\mu \in {\mathbb {R}}^+,\) the left and right fractional integrals are respectively defined as
For real \(s\in [k-1, k)\) with \(k\in {\mathbb {N}},\) the left-sided Riemann–Liouville fractional derivative (LRLFD) of order s is defined by
and the right-sided Riemann–Liouville fractional derivative (RRLFD) of order s is defined by
From the above definitions, it is clear that for any \(k\in {\mathbb {N}}_0,\)
Therefore, we can express the RLFD as
According to [9, Thm. 2.14], we have that for any finite a and any \(f\in L^1(\Lambda ),\) and real \(s\ge 0,\)
Note that by commuting the integral and derivative operators in (2.6), we define the Caputo fractional derivatives:
For an affine transform \(x=\lambda t,~\lambda >0\), on account of
and \(\frac{\hbox {d}}{{\hbox {d}}t}=\lambda \frac{\hbox {d}}{{\hbox {d}}x},\) we derive from Definition 2.1 that
Similarly, we have the following identities for the right fractional derivative:
2.2 Tempered Fractional Integrals and Derivatives on \({\mathbb {R}}\)
Recently, Sabzikar et al. [22, (19)–(23)] introduced the tempered fractional integrals and derivatives on the whole line.
Definition 2.2
(Tempered fractional integrals) For \(\lambda \in {\mathbb {R}}^+_0\), the left tempered fractional integral of a suitable function u(x) of order \(\mu \in {\mathbb {R}}^+\) is defined by
and the right tempered fractional integral of order \(\mu \in {\mathbb {R}}^+\) is defined by
It is evident that by (2.2) and (2.11)–(2.12), we have
and
As shown in [22], the tempered fractional derivative can be characterized by its Fourier transform. Recall that, for any \(u\in L^2({\mathbb {R}}),\) its Fourier transform and inverse Fourier transform are defined by
There holds the well-known Parseval’s identity:
where \({\bar{v}}\) is the complex conjugate of v. Let H(x) be the Heaviside function, i.e., \(H(x)=1\) for \(x\ge 0,\) and vanishing for all \(x< 0.\) Then we can reformulate the left tempered fractional integral as
Note that K(x) is related to the particle jump density (cf. [22, (8)]). Using the formula: \( {{\mathscr {F}}}[K](\omega )= {(\lambda +\mathrm{i}\omega )^{-\mu }},\) and the convolution property of Fourier transform (see, e.g., [23, 26]), we derive
Similarly, the Fourier transform of the right tempered fractional integral is
In view of (2.18)–(2.19), Sabzikar et al. [22] then introduced the left and right tempered fractional derivatives as follows.
Definition 2.3
(Tempered fractional derivatives) For \( \lambda \in {\mathbb {R}}^+_0\), the left and right tempered fractional derivatives of order \(\mu \in {\mathbb {R}}^+\) of a suitable function u(x), are defined by
that is, for any \(x\in {\mathbb {R}}\),
Introduce the space
Thanks to the Parseval’s identity (2.16), the above tempered fractional derivatives are well-defined for any \(u\in W^{\mu ,2}_{\lambda }({\mathbb {R}}).\) Moreover, one verifies from (2.18)–(2.21) that
Similar to (2.14), we have the following explicit representations (see [13, Lemma 1 and Remark 2]).
Proposition 2.1
For any \(u\in W^{\mu ,2}_{\lambda }({\mathbb {R}}),\) with \( \lambda \in {\mathbb {R}}^+_0\), the left and right tempered fractional derivatives of order \(\mu \in [k-1,k)\) with \(k\in {\mathbb {N}},\) have the explicit representations:
where \({{}_{-\infty }}\mathrm{D}_{x}^{\mu }\) and \({{}_{x}}\mathrm{D}_{\infty }^{\mu }\) are the Riemann–Liouville fractional derivative operators in Definition 2.1. Alternatively, we have
We collect below some useful properties (see [22]).
Lemma 2.1
Given \(\lambda >0\) and \(\mu \in [k-1,k),~k\in {\mathbb {N}}\), the tempered fractional derivative
In addition, we have
where \( \mu ,\nu \ge 0\).
Remark 2.1
For a suitable function \(f(x),~x\in {\mathbb {R}}^+\), its reflection \(g(y)=f(-y),~y\in {\mathbb {R}}^-\) satisfies
Hence, we can use (2.26) and derivative relation \(\dfrac{{\mathrm{d}}^k}{{\mathrm{d}y^k}}=(-1)^k\dfrac{{\mathrm{d}}^k}{{\mathrm{d}x}^k}\) to obtain the tempered derivative relation
\(\square \)
2.3 Laguerre Polynomials and Some Useful Formulas
For any \(a\in {\mathbb {R}}\) and \(j\in {\mathbb {N}}_0,\) we recall that the rising factorial in the Pochhammer symbol and the Gamma function have the relation:
Recall the hypergeometric function (cf. [1]):
If \(b-a>0,\) then \({}_1F_1(a;b; x)\) is absolutely convergent for all \(x\in {\mathbb {R}}.\) If a is a negative integer, then it reduces to a polynomial.
The Laguerre polynomial with parameter \(\alpha >-1\) is defined as in Szegö [27, (5.3.3)]:
and \(L_0^{(\alpha )}(x)\equiv 1.\) Note that
and the Laguerre polynomials (with \(\alpha >-1\)) are orthogonal with respect to the weight function \(x^\alpha e^{-x},\) namely,
They are eigenfunctions of the Sturm–Liouville problem:
We have the following relations:
In particular, for \(\alpha =-k,~k=1,2,\ldots \) (See Szegö [27, (5.2.1)]),
For notational convenience, we denote
We present below some formulas related to Laguerre polynomials and fractional integrals and derivatives, which play an important role in the algorithm development and analysis later. We provide their derivations in “Appendix A”.
Lemma 2.2
For \(\mu \in {\mathbb {R}}^+,\) we have
and
Moreover, we have that for \(k\in {\mathbb {N}}\) and \(\alpha >k-1\),
3 Generalized Laguerre Functions
In this section, we introduce the generalized Laguerre functions (GLFs), and study its approximation properties. In what follows, the operators \({}_{0}\mathrm{I}_{x}^{\mu ,\lambda }, {{}_{0}}\mathrm{D}_{x}^{\mu ,\lambda }\) on the half line should be understood as 0 in place of \(-\infty \) in (2.11) and (2.24)–(2.25).
3.1 Definition and Properties
We first introduce the GLFs and their associated properties related to tempered fractional integrals/derivatives.
Definition 3.1
For real \(\alpha \in {\mathbb {R}}\) and \(\lambda >0,\) we define the GLFs as
for all \(x\in {\mathbb {R}}^+\) and \(n\in {\mathbb {N}}_0.\)
Remark 3.1
It’s noteworthy that Zhang and Guo [29] introduced the GLFs
where the scaling factor \(\beta >0.\) It is seen that we modified the definition in the range of \(0<\alpha <1\) (with \(\beta =2\lambda \)). This turns out to be essential for the numerical solution of FDEs of order \(\mu \in (0,1),\) as we shall see in the subsequent sections. \(\square \)
We next present the basic properties of GLFs. Firstly, one verifies readily from the orthogonality (2.36) and Definition 3.1 that for \(\alpha \in {\mathbb {R}}\) and \(\lambda >0,\)
where \(\gamma _n^{|\alpha |}\) is defined in (2.36).
We have the following important (left) “tempered” fractional integral and derivative rules.
Lemma 3.1
For \(\mu ,\nu ,\lambda ,x \in {\mathbb {R}}^+_0,\) we have
and
where \(h^{a,b}_n\) is defined in (2.41).
Proof
We obtain from (2.11) and (2.24)–(2.25) (with replacing \(-\infty \) by 0) that
and
Thus, from (2.9) and Lemma 2.2, we obtain (3.4)–(3.5).
Using (3.5) and the derivative relation (2.40) (with \(\alpha =\mu \)), we obtain
This leads to (3.6). \(\square \)
Similarly, we have the following rules of the (right) “tempered” fractional integrals and derivatives.
Lemma 3.2
For \( \mu ,\nu ,\lambda , x \in {\mathbb {R}}^+_0\), we have
Proof
Identities (3.7) and (3.8) can be easily derived from (2.9), (2.10) and Lemma 2.2. \(\square \)
We highlight the fractional derivative formulas, which play an important role in the forthcoming algorithm and analysis.
Theorem 3.1
Let \(k\in {\mathbb {N}}\) and \(k-\nu \le 0\),
Proof
From Lemma 2.2 and relations
we obtain that for \(k-\nu \le 0\),
and
This ends the proof. \(\square \)
Another attractive property of GLFs is that they are eigenfunctions of Sturm–Liouville problem.
Theorem 3.2
Let \(s,\nu ,x\in {\mathbb {R}}^+_0\) and \(n\in {\mathbb {N}}_0\). Then,
and
where the corresponding eigenvalues \(\lambda _{n,-}^{s,\nu }=(2\lambda )^s h_n^{\nu ,s}\) and \(\lambda _{n,+}^{s,\nu }=(2\lambda )^s h_n^{\nu +s,s}\).
Proof
It’s straightforward to obtain that
Similarly, we have
This ends the derivation. \(\square \)
Remark 3.2
The above identities can be viewed as an extension of the standard Sturm–Liouville problem of generalized Laguerre functions (cf. (2.37)) to the tempered fractional derivative. We derive immediately from (3.13), (3.14) and the Stirling’s formula (see (3.23)) that for fixed s and \(\nu \),
When \(s\rightarrow 1\) and \(\lambda =1/2\), it recovers the O(n) growth of eigenvalues of the standard Sturm–Liouville problem. \(\square \)
3.2 Approximation by GLFs
3.2.1 Approximation by \(\big \{{\mathcal {L}}^{(-\nu ,\lambda )}_{n}{(x)}:~\nu >0\big \}_{n=0}^\infty \)
Denote by \({\mathcal {P}}_N\) the set of all polynomials of degree at most N, and define the finite dimensional space
Define the \(L_\omega ^2({\mathbb {R}}^+)\) with the inner product and norm:
where \(\omega (x)\) be a generic weight function and \({\bar{g}}\) is the conjugate of the function g. In particular, we omit \(\omega \) when \(\omega \equiv 1.\)
To characterize the approximation errors, we define the non-uniformly weighted Sobolev space
equipped with the norm and semi-norm
where the weight function \(\omega ^{a}(x)=x^a.\)
Consider the orthogonal projection \(\pi _N^{-\nu ,\lambda }:\,{L^2_{\omega ^{-\nu }}}({\mathbb {R}}^+)\rightarrow \mathcal {F}^{\nu ,\lambda }_N({\mathbb {R}}^+)\) defined by
Then, by the orthogonality (3.3), u and its \(L^2\)-orthogonal projection can be expanded as
where
Theorem 3.3
For \(\lambda , \nu >0\), we have that for any \(u\in A^m_{\nu ,\lambda }({\mathbb {R}}^+) \) with \(m\le N+1\),
and for any \(k\le m,\)
where \(c\approx 1\) for large N.
Proof
By (3.20), we have
By the orthogonality (3.3) and (3.6),
where we denoted \(d_{n,k}^{\nu ,\lambda }:=(2\lambda )^k\,h^{\nu ,\nu }_n\) and used the fact:
Thus we can obtain
Then one verifies readily that
Recall the property of the Gamma function (see [1, (6.1.38)]):
One can then obtain that for any constants a, b, and for \(n\ge 1, \) \(n+a>1\) and \(n+b>1,\)
where
Therefore,
where \(\nu _n^{2-m,2+\nu }\approx 1\) and \(\nu _n^{2-m,2-k}\approx 1\) for fixed m and \(n\ge N\gg 1 \). Then (3.21)–(3.22) follow. \(\square \)
3.2.2 Approximation by \(\big \{{\mathcal {L}}^{(\nu ,\lambda )}_{n}{(x)}: \nu \ge 0\big \}_{n=0}^\infty \)
Introduce the non-uniformly weighted Sobolev space:
endowed with the norm and semi-norm
Consider the orthogonal projection \(\Pi _N^{\nu ,\lambda }:~{L^2_{\omega ^{\nu }}}({\mathbb {R}}^+)\rightarrow {\mathcal {F}}^{0,\lambda }_N({\mathbb {R}}^+),\) defined by
Theorem 3.4
Let \(\lambda ,r ,\nu >0\). For any \(u\in B^r_{\nu ,\lambda }({\mathbb {R}}^+) \) with \(0\le s\le r\le N,\) we have
where \(c\approx 1\) for large N.
Proof
Note that by definition,
Then by (3.8), and the orthogonality,
we can derive
Then,
where by (3.24)–(3.25) and an argument similar to (3.26), we obtain
Consequently, we have
This ends the proof. \(\square \)
3.3 A Model Problem and Numerical Results
In what follows, we consider the GLF approximation to a model tempered fractional equation of order \(s\in [k-1,k)\) with \(k\in \mathbb {N}:\)
where \(f\in L^2({\mathbb {R}}^+)\) is a given function. Using the fractional derivative relation (2.7), one can find
where \(\{c_i\}\) can be determined by the conditions at \(x=0.\) In fact, we have all \(c_i=0,\) and
We see that if f(x) is smooth, then \(u(x)=x^s F(x),\) where F(x) is smoother than f(x). With this understanding, we construct the GLF Petrov–Galerkin approximation as: find \(u_N\in {\mathcal {F}}^{s,\lambda }_N({\mathbb {R}}^+)\) (defined in (3.15)) such that
We expand f and \(u_N\) as
Using the derivative relation (3.5), we find immediately that \(\hat{u}_n={\hat{f}_n}/{h_n^{s,s}}\) for \(n=0,1,\ldots ,N\), which also implies \({{}_{0}}\mathrm{D}_{x}^{s,\lambda } u_N=\pi ^{0,\lambda }_N f.\)
Moreover, we can show that the numerical solution \(u_N\) is precisely the orthogonal projection in the following sense:
To this end, we first show
Indeed, thanks to \(u^{(j)}(0)=0\) for \(j=0,\ldots ,k-1\), we have
Then,
so (3.36) is valid. In addition, thanks to Lemma 2.2 and (2.10), we have
Hence, (3.35) is valid.
Thanks to (3.35), we derive from Theorem 3.3 the following estimate where the convergence rate only depends on the regularity of the source term.
Theorem 3.5
Let u and \(u_N\) be respectively the solutions of (3.31) and (3.33). Then for \({{}_{0}}\mathrm{D}_{x}^{m,\lambda }f\in L^2_{\omega ^m}(I)\) with \(m\in {\mathbb {N}}_0,\) we have
where \(c\approx 1\) for large N.
We provide some numerical results to illustrate the convergence behaviour. We take \(f(x)= e^{-x}\sin x\) and then evaluate the exact solution by (3.32). Note that as \({{}_{0}}\mathrm{D}_{x}^{m,\lambda }f=e^{-\lambda x} {{}_{}}\mathrm{D}_{}^{m}\{e^{\lambda x}f\}\), a direct calculation leads to
We infer from (3.37) that the spectral accuracy can be achieved by the GLF approximation. Indeed, we observe from Fig. 1 such a convergence behaviour.
4 Application to Tempered Fractional Diffusion Equation on the Half Line
In this section, we apply the GLFs to solve a tempered fractional diffusion equation on the half line.
4.1 The Tempered Fractional Diffusion Equation on the Half Line
Consider the tempered fractional diffusion equation of order \(\mu \in (0,1)\) on the half line:
This equation models the particles jumping on the half line \({\mathbb {R}}^+\) with the probability density function (see [22, (8)]):
Remark 4.1
Note that (4.1) can be viewed as the TFDE (1.3) on the half line with
Indeed, we can show that for \(\mu \in (0,1)\) and real \(\lambda >0\),
where \({\tilde{u}}=u\) for \(x\in {\mathbb {R}}^+\) and \(\tilde{u}=0\) for \(x\in (-\infty ,0)\). Moreover, we have
This implies \({\tilde{u}}\in W^{\mu ,2}_\lambda ({\mathbb {R}})\) and the extended tempered fractional derivative \({{}_{0}}\mathrm{D}_{x}^{\mu ,\lambda }u\) can be understood in the sense of the original definition in [22].
4.2 Spectral-Galerkin Scheme
Observe from Remark 4.1 that the identities (2.29) are also valid on \({\mathbb {R}}^+\). A weak form of the problem (4.1) is to find \(u(\cdot ,t)\in W^{\mu /2,2}_{\lambda }({\mathbb {R}}^+)\) such that
where the space \(W^{\mu /2,2}_{\lambda }({\mathbb {R}}^+)\) consists of all functions whose zero extensions are in \( W_\lambda ^{\mu /2,2}({\mathbb {R}})\) (cf. (2.22)), and the bilinear form reads
The semi-discrete Galerkin approximation scheme is to find \(u_N(\cdot ,t)\in {\mathcal {F}}^{\nu ,\lambda }_N({\mathbb {R}}^+) \) such that
with
where the projection operator \(\pi _N^{-\nu ,\lambda }\) is defined in (3.19) with \(\max \big \{0,\mu -1/2\big \}<\nu \le 1.\) Note that the boundary condition \(u(0,t)=0\) is automatically met, and we choose the parameter \(\nu \) is to better fit the singularity behavior of the solution near \(x=0\).
Remark 4.2
We show in next section (see Theorem 5.1 and Remark 5.1) the positivity of the bilinear form, that is, for any \(0\not =v\in W^{\mu /2,2}_{\lambda }({\mathbb {R}}^+)\), \(a_{\mu ,\lambda }(v,v)>0\). Then we can show the stability of the solutions of (4.2) and (4.4) as in (5.9) (with \({\mathbb {R}}^+\) in place of \({\mathbb {R}}\)). Moreover, we can conduct the error analysis of the semi-discrete scheme by using the approximation results in Sect. 3, and following a standard argument for usual diffusion equations. \(\square \)
4.3 Numerical Algorithm
Now, set
We derive from the scheme (4.4) that
where for fixed \(t>0\), vectors
Note that for any \(u,v\in {\mathcal {F}}^{\nu ,\lambda }_N({\mathbb {R}}^+)\), there exists
The mass matrix \({\mathbf {M}}\) and the stiffness matrix \(\mathbf {S}\) can be computed by Laguerre-Gauss quadrature formula. In fact, by using the tempered fractional derivative relations (3.5), it’s straightforward to obtain that for \(m,n=0,1,2,\ldots ,N\),
where \(h^{\nu ,\mu }_n=\Gamma (n+1+\nu )/\Gamma (n+1+\nu -\mu )\) was defined in (2.41).
4.4 Numerical Results
Typically, we test three cases as follows.
-
(i)
Choose the exact solution to be \(u(x,t)= xe^{-\lambda x}\cos {(t)}\). By a direct calculation, the source term is given by
$$\begin{aligned} f(x,t)=-xe^{-\lambda x}\sin (t)+\Big (\dfrac{\Gamma (2)}{\Gamma (2-\mu )}x^{1-\mu }-\lambda ^\mu x\Big )e^{-\lambda x}\cos {(t)}. \end{aligned}$$Figure 2 (left) illustrates that the error decays to zero rapidly for the spectral method built upon the GLF basis with \(\nu =-1\) and \(N=50\), and the third-order explicit Runge-Kutta method in time with \(\lambda =\mu =2/3\) and time stepping size \(h\in (10^{-3},10^{-1})\).
-
(ii)
Set \(f(x,t)=\cos (x)e^{-x}\sin (t),\) and choose \(\lambda , \mu \) as above. Figure 2 (right) verifies that the solution is singular even though f(x, t) is a smooth function. Here, we compare the error with an reference “exact” solution computed with \(N=100.\)
-
(iii)
Consider \(f(x,t)\equiv 0\), and let \(\mu =2/3,~\lambda =2/3\) in (4.1). Figure 3 (left) exhibits the evolution of the tempered fractional diffusion model with the initial distribution \(u_0(x)=xe^{-x}\). Figure 3 (right) shows the convergence rate of the scheme, where the error is compared with the reference solution \(u_{100}(x,t)\) with different \(\nu \) at \(t=10\).
5 Tempered Fractional Diffusion Equation on the Whole Line
In this section, we present a multi-domain spectral-element method for the tempered fractional diffusion equation on the whole line originally proposed by [22].
5.1 Tempered Fractional Diffusion Equation
Consider the tempered fractional diffusion equation of order \(\mu \in (k-1,k),~k=1,2\) on the whole line:
where p, q are nonnegative constants such that \(p+q=1,\) and \(f,u_0\) are given functions. Here, we denote
where the involved fractional operators are
-
(i)
for \(0<\mu <1,\)
$$\begin{aligned} \begin{aligned} \partial _{+,x}^{\mu ,\lambda }u={{}_{-\infty }}\mathrm{D}_{x}^{\mu ,\lambda }u-\lambda ^\mu u,\quad \partial _{-,x}^{\mu ,\lambda }u={{}_{x}}\mathrm{D}_{\infty }^{\mu ,\lambda }u-\lambda ^\mu u; \end{aligned}\end{aligned}$$(5.3) -
(ii)
for \(1<\mu <2\),
$$\begin{aligned} \partial _{+,x}^{\mu ,\lambda }u={{}_{-\infty }}\mathrm{D}_{x}^{\mu ,\lambda }u-\mu \lambda ^{\mu -1}\partial _x u-\lambda ^\mu u,\quad \partial _{-,x}^{\mu ,\lambda }u={{}_{x}}\mathrm{D}_{\infty }^{\mu ,\lambda }u+\mu \lambda ^{\mu -1}\partial _x u-\lambda ^\mu u. \end{aligned}$$(5.4)
We refer to Definition 2.3 for the tempered derivative operators.
A weak form of (5.1) is to find \(u(\cdot , t)\in V_\lambda ^\mu ({\mathbb {R}})\) for \(0<t\le T,\) such that
where \((\cdot , \cdot )\) is the inner product of \(L^2({\mathbb {R}})\) as before. The space \(V_\lambda ^\mu ({\mathbb {R}})\) and the bilinear form \(a^{\mu ,\lambda }_{p,q}(\cdot ,\cdot )\) are defined as
-
(i)
for \(0<\mu <1,\) and \(u,v\in V_\lambda ^\mu ({\mathbb {R}})= W_\lambda ^{\mu /2,2}({\mathbb {R}}) \cap L^2({\mathbb {R}})\) (cf. (2.22)),
$$\begin{aligned} a^{\mu ,\lambda }_{p,q}(u,v)=p\big ({{}_{-\infty }}\mathrm{D}_{x}^{\mu /2,\lambda }u, {{}_{x}}\mathrm{D}_{\infty }^{\mu /2,\lambda } v\big )+q\big ({{}_{x}}\mathrm{D}_{\infty }^{\mu /2,\lambda }u,{{}_{-\infty }}\mathrm{D}_{x}^{\mu /2,\lambda }v\big ) -\lambda ^\mu \, (u,v),\ \end{aligned}$$(5.6) -
(ii)
for \(1<\mu <2,\) and \(u,v\in V_\lambda ^\mu ({\mathbb {R}})= W_\lambda ^{\mu -1,2}({\mathbb {R}})\cap H^1({\mathbb {R}})\),
$$\begin{aligned} \begin{aligned} a^{\mu ,\lambda }_{p,q}(u,v)=&-p({{}_{-\infty }}\mathrm{D}_{x}^{\mu -1,\lambda }u,{{}_{x}}\mathrm{D}_{\infty }^{1,\lambda }v) -q({{}_{x}}\mathrm{D}_{\infty }^{1,\lambda }u,{{}_{-\infty }}\mathrm{D}_{x}^{\mu -1,\lambda }v)\\&+\lambda ^\mu (u,v)+(p-q)\mu \lambda ^{\mu -1}(\partial _xu,v). \end{aligned} \end{aligned}$$(5.7)
Note that in (5.7), \({{}_{x}}\mathrm{D}_{\infty }^{1,\lambda }u=\lambda u-\partial _x u\) (cf. (2.25)).
Importantly, we can show that the involved bilinear form is strictly positive, so the well-posedness of (5.5) folows.
Theorem 5.1
For any \(0\not =v \in V_\lambda ^\mu ({\mathbb {R}})\) with \(\mu \in (0,1)\cup (1,2)\) and \(\lambda >0,\) we have
If \(u_0\in L^2({\mathbb {R}})\) and \(f\in L^2({\mathbb {R}}\times (0,T)),\) then the problem (5.5) has a unique solution \(u\in L^\infty (0,T; V_\lambda ^\mu ({\mathbb {R}}))\) such that
Proof
We first consider \(0<\mu <1.\) Using (2.20) and the Parseval’s identity (2.16), leads to
where \(\Theta (\omega )\) is the argument of \(\lambda +\mathrm{i}\omega ,\) i.e.,
It is evident that \(\Theta (\omega )\) is odd in \(\omega .\) In fact, for real function v, \(\big |{\mathscr {F}}[v](\omega ) \big |^2\) is an even function in \(\omega ,\) thanks to the property \({\mathscr {F}}[v](\omega ) =\overline{ {\mathscr {F}}[v](-\omega )}\), which can be derived from the definition (2.15) straightforwardly. Thus, we derive from (5.10) that
For notational convenience, we denote
Noting that
we find
As \(\Theta \in (0, \pi /2),\) \({\mathcal K}_\mu (\omega )\) is ascending with respect to \(\omega ,\) when \(\mu \in (0,1)\). Consequently, for \(\mu \in (0,1)\),
Combing (5.12) and (5.15) leads to
where in the last, we used the property \(|{\mathscr {F}}[v](\omega )|^2\) is even in \(\omega \). As \(p+q=1,\) we obtain from (5.6) and the above that
For \(1<\mu <2,\) we can follow the same derivation and show that
but by (5.14), we have \({\mathcal K}_\mu '(\omega )<0,\) so
Observe that
From (5.7) and (5.17)–(5.19), we obtain
This ends the proof of (5.8).
Next, taking \(v=u\) in (5.5), we obtain from the Cauchy–Schwarz inequality that
which, together with (5.8), implies
We immediately obtain (5.9), which implies the uniqueness of the solution. The existence follows form the equivalence of uniqueness and existence for linear problems. \(\square \)
Remark 5.1
We see from the proof that the same result is valid for \(\mu \in (0,1)\) and \(u(x,0)=0\) on the half line \({\mathbb {R}}^+\) through zero extension. Therefore, we can show the stability of the model in the previous section (see Remark 4.2). \(\square \)
5.2 A Two-Domain Spectral-Element Method
An interesting observation of the model in [22] (i.e., (5.1)) is its solution might have a limited regularity across \(x=0.\) This motivates us to use the Laguerre polynomial approximations on \((-\infty , 0)\) and \((0,\infty ),\) respectively. Thus, we decompose the whole line as
and denote \(u_{\Lambda _j}(x,t):=u(x,t)\big |_{\Lambda _j},~j=1,2\). Introduce the approximation space:
and define
where \({\mathcal {L}}^{(-1,\lambda )}_{n}{(}x)=e^{-\lambda x}xL_n^{(1)}(2\lambda x)\). One verifies readily that
Then, our semi-discrete spectral-Galerkin method is to find \(u_{\varvec{N}}(\cdot ,t)\in V^{\lambda }_{\varvec{N}}({\mathbb {R}}) \) such that
We provide below some details of the algorithm.
Let H(x) be the Heaviside function as before. Thanks to the tempered fractional derivative and integral relations with GLFs, and a reflected mapping from positive half line \({\mathbb {R}}^+\) to negative half line \({\mathbb {R}}^-\), we can derive the following identities (see “Appendix B”):
Then (5.26) leads to the linear system of ordinary differential equations:
where
and the matrices
with the entries
and \(\overrightarrow{{\mathbf {C}}}(0)\) is determined by the initial data.
The derivation of the tempered derivative relation (5.28), and of the entries of the matrix \({\mathbf {A}}\) can be found in “Appendix B”. Base on the semi-discrete scheme (5.29), we further use the third-order explicit Runge-Kutta method in time direction with step size \(h=10^{-3}\) to numerically solve the problem.
5.3 Numerical Results
We solve (5.1) with \(C_T=1\) and \(u_0=10e^{-5|x|}\) as the initial distribution by using the proposed method. We first test its accuracy. In Fig. 4, we plot the convergence rate of the spectral method at \(T=5\) with fixed time step \(h=10^{-3}\). We choose a slow decay \(f(x,t)= (1+x^2)^{-1},\) and an exponential decay \(f(x,t)= \cos {t} ~e^{-x^2}\), respectively. Observe from Fig. 4 an exponential convergence for the latter, but an algebraic convergence for the former. In fact, we expect the solution with \(f(x,t)= \cos {t} ~e^{-x^2}\) decays exponentially in space. Indeed, as the Fourier transform of the Gaussian \(e^{-x^2}\) is invariant, so we can use the model by using Fourier transform and then solve the resulted equation. However, the solution with \(f(x,t)= (1+x^2)^{-1}\) should decay very slowly. As a result, we observe a very different convergence behavior.
Next, we examine behaviors of the solution under various situations. In Fig. 5, we plot the snapshots at different times of the tempered fractional diffusion with \(p=1/3,~q=2/3\) and \(p=3/4,~q=1/4\), respectively. The case with \(p=q=1/2\) is plotted in Fig. 6.
-
The parameters p and q reflect the directional preference of the particle jumping. More precisely, if \(p>q\), the particles tend to jump to the right, and if \(p<q\), the particles tend to jump to the left, see Fig. 5. In particular, \(p=q\) produces a symmetric profile in the case of \(f(x,t)=0\), see the left in Fig. 6.
-
The parameter \(\lambda \) determines the probability of the jump distance of the particles. A larger \(\lambda \) indicates a shorter jump distance, see the right of Fig. 6.
-
To compare with the usual fractional diffusion equation, i.e., \(\lambda =0\), , we plot in Fig. 7 the particle distributions of the usual fractional diffusion and the tempered fractional diffusion with initial distribution \(u_0(x)=10e^{-4 x^2}\) at time \(t=10\). We observe that the tail of the tempered fractional diffusion behaves like \(|x|^{-\mu -1}e^{-\lambda |x|}\) for large |x| while that that of the usual fractional diffusion behaves like \(|x|^{-\mu -1}\).
6 Concluding Remarks
We presented in this paper efficient spectral methods using the generalized Laguerre functions for solving the tempered fractional differential equations on infinite intervals. Our numerical methods and analysis are based on an important observation that the tempered fractional derivative, when restricted to the half line, is intrinsically related to the generalized Laguerre functions that we defined in Sect. 3. By exploring the properties of generalized Laguerre functions, we derived optimal approximation results in properly weighted Sobolev spaces. In Sect. 4, we developed a spectral-Galerkin method for solving a tempered fractional diffusion equation on the half line. Finally, we presented a spectral-Galerkin method for solving the tempered fractional diffusion equation on the whole line in Sect. 5. More importantly, we rigorously showed the well-posedness of the tempered fractional model in [22]. Also to the best of our knowledge, this is perhaps the first attempt in solving this model directly on infinite intervals, and also show the well-posedness of the models in [22]. Indeed, a finite-difference approach on a truncated domain was employed in [22]. Moreover, our numerical results demonstrated some expected properties and behaviors of the underlying solution of such a tempered diffusion model.
References
Abramovitz, M., Stegun, I.A.: Handbook of Mathematical Functions. Dover, New York (1972)
Andrews, G.E., Askey, R., Roy, R.: Special Functions. Encyclopedia of Mathematics and its Applications, vol. 71. Cambridge University Press, Cambridge (1999)
Baeumer, B., Benson, D.A., Meerschaert, M.M., Wheatcraft, S.W.: Subordinated advection–dispersion equation for contaminant transport. Water Resour. Res. 37(6), 1543–1550 (2001)
Baeumer, B., Kovács, M., Meerschaert, M.M.: Fractional reproduction–dispersal equations and heavy tail dispersal kernels. Bull. Math. Biol. 69(7), 2281–2297 (2007)
Carr, P., Geman, H., Madan, D.B., Yor, M.: The fine structure of asset returns: an empirical investigation. J. Bus. 75(2), 305–333 (2002)
Chen, S., Shen, J., Wang, L.L.: Generalized Jacobi functions and their applications to fractional differential equations. Math. Comput. 85(300), 1603–1638 (2016)
Cushman, J.H., Ginn, T.R.: Fractional advection–dispersion equation: a classical mass balance with convolution-fickian flux. Water Resour. Res. 36(12), 3763–3766 (2000)
Deng, Z.Q., Bengtsson, L., Singh, V.P.: Parameter estimation for fractional dispersion model for rivers. Environ. Fluid Mech. 6(5), 451–475 (2006)
Diethelm, K.: The analysis of fractional differential equations. Lecture Notes in Math, vol. 2004. Springer, Berlin (2010)
Gorenflo, R., Mainardi, F., Scalas, E., Raberto, M.: Fractional calculus and continuous-time finance III: the diffusion limit. Math. Financ. 171–180 (2001)
Huang, C., Song, Q., Zhang, Z.: Spectral collocation method for substantial fractional differential equations. arXiv:1408.5997 [math.NA] (2014)
Jeon, J.H., Monne, H.M.S., Javanainen, M., Metzler, R.: Anomalous diffusion of phospholipids and cholesterols in a lipid bilayer and its origins. Phys. Rev. Lett. 109(18), 188103 (2012)
Li, C., Deng, W.H.: High order schemes for the tempered fractional diffusion equations. Adv. Comput. Math. 42(3), 543–572 (2016)
Magdziarz, M., Weron, A., Weron, K.: Fractional Fokker–Planck dynamics: stochastic representation and computer simulation. Phys. Rev. E 75(1), 016708 (2007)
Meerschaert, M.M., Zhang, Y., Baeumer, B.: Tempered anomalous diffusion in heterogeneous systems. Geophys. Res. Lett. 35, L17403 (2008). doi:10.1029/2008GL034899
Meerschaert, M.M., Scalas, E.: Coupled continuous time random walks in finance. J. Phys. A Stat. Mech. Appl. 370(1), 114–118 (2006)
Metzler, R., Klafter, J.: The random walk’s guide to anomalous diffusion: a fractional dynamics approach. Phys. Rep. 339(1), 1–77 (2000)
Metzler, R., Klafter, J.: The restaurant at the end of the random walk: recent developments in the description of anomalous transport by fractional dynamics. J. Phys. A Math. Gen. 37(31), R161 (2004)
Piryatinska, A., Saichev, A., Woyczynski, W.: Models of anomalous diffusion: the subdiffusive case. J. Phys. A Stat. Mech. A. 349(3), 375–420 (2005)
Podlubny, I.: Fractional Differential Equations: An Introduction to Fractional Derivatives, Fractional Differential Equations, to Methods of Their Solution and Some of Their Applications. Academic Press Inc., San Diego (1999)
Popov, G.Y.: Concentration of elastic stresses near punches, cuts, thin inclusions and supports. Nauka Mosc. 1, 982 (1982)
Sabzikar, F., Meerschaert, M.M., Chen, J.: Tempered fractional calculus. J. Comput. Phys. 293, 14–28 (2015)
Samko, S.G., Kilbas, A.A., Maričev, O.I.: Fractional Integrals and Derivatives. Gordon and Breach Science Publ., Philadelphia (1993)
Scalas, E.: Five years of continuous-time random walks in econophysics. In: The Complex Networks of Economic Interactions, pp. 3–16. Springer, Berlin (2006)
Shen, J., Tang, T., Wang, L.L.: Spectral Methods: Algorithms, Analysis and Applications. Series in Computational Mathematics, vol. 41. Springer, Berlin (2011)
Stein, E.M., Weiss, G.L.: Introduction to Fourier Analysis on Euclidean Spaces, vol. 1. Princeton University Press, Princeton (1971)
Szegö, G.: Orthogonal Polynomials, 4th edn. AMS Coll. Publ., (1975)
Zayernouri, M., Ainsworth, M., Karniadakis, G.E.: Tempered fractional Sturm–Liouville Eigen-problems. SIAM J. Sci. Comput. 37(4), A1777–A1800 (2015)
Zhang, C., Guo, B.Y.: Domain decomposition spectral method for mixed inhomogeneous boundary value problems of high order differential equations on unbounded domains. J. Sci. Comput. 53(2), 451–480 (2012)
Zhang, Y., Meerschaert, M.M.: Gaussian setting time for solute transport in fluvial systems. Water Resour. Res. 47, W08601 (2011). doi:10.1029/2010WR010102
Author information
Authors and Affiliations
Corresponding author
Additional information
This work is supported in part by NSFC grants 11371298, 11421110001, 91630204 and 51661135011. J.S. is partially supported by NSF grant DMS-1620262 and AFOSR grant FA9550-16-1-0102. L.W. is partially supported by Singapore MOE AcRF Tier 1 Grant (RG 15/12), and Singapore MOE AcRF Tier 2 Grant (MOE 2013-T2-1-095, ARC 44/13).
Appendices
Appendix A: Proof of Lemma 2.2
We first prove (2.42)–(2.43). Recall the fractional integral formula of hypergeometric functions see [2, P. 287]: for real \(b,\mu \ge 0,\)
Taking \(a=-n,~b=\alpha +1\) and using the hypergeometric representation (2.34) of the Laguerre polynomials, we obtain
which yields (2.42), i.e.,
Then, performing \({{}_{0}}\mathrm{D}_{x}^{\mu }\) on both sides and taking \(\alpha +\mu \rightarrow \alpha ,\) we derive from the relation (2.7) that for \(\alpha -\mu >-1,\)
This leads to (2.43).
We now turn to (2.44)–(2.45). According to [20, (6.146), P. 191 ] (or [21, (B-7.2), P. 307]), we have
Similarly, from the property: \({{}_{x}}\mathrm{D}_{\infty }^{\mu } {}_{x}\mathrm{I}_{\infty }^{\mu }\, u(x)=u(x),\) we derive
Finally, we prove (2.46). Noting that
(cf. [2, P. 191]), we derive from (2.34) and (A.2) that
Then acting the derivative \({{}_{}}\mathrm{D}_{}^{k}\) on (A.3) and using the identities (2.34), (A.2) again, we obtain
This ends the proof.
Appendix B: Derivation of (5.28) and the Entries of \({\mathbf {A}}\)
Derivation of (5.28)
-
for \(x\in {\mathbb {R}}^-\), \(0<s<1\),
$$\begin{aligned} \begin{aligned} {{}_{-\infty }}\mathrm{D}_{x}^{s,\lambda }\phi ^{*}(x)&={}_{-\infty }\mathrm{I}_{x}^{1-s,\lambda }{{}_{-\infty }}\mathrm{D}_{x}^{1,\lambda } \phi ^{*}(x)=\frac{e^{-\lambda x}}{\Gamma (1-s)}\int _{-\infty } ^x\frac{e^{\lambda \tau }(2\lambda )e^{\lambda \tau }}{(x-\tau )^s} \mathrm{d}\tau \\&\overset{t=x-\tau }{=}\frac{2\lambda e^{-\lambda x}}{\Gamma (1-s)}\int _0^{\infty }\frac{e^{2\lambda (x-t)}}{t^s} \mathrm{d}t =\frac{(2\lambda )^{s}e^{\lambda x}}{\Gamma (1-s)}\int _0^{\infty } e^{-2\lambda t}(2\lambda t)^{-s} \mathrm{d}(2\lambda t)\\&=(2\lambda )^{s}e^{\lambda x},\\ {{}_{-\infty }}\mathrm{D}_{x}^{s,\lambda }\phi ^{-}_{n_1}(x)&={}_{-\infty }\mathrm{I}_{x}^{1-s,\lambda } {{}_{-\infty }}\mathrm{D}_{x}^{1,\lambda }\phi ^-_{n_1}(x)\overset{(3.10)}{=}{}_{-\infty }\mathrm{I}_{x}^{1-s,\lambda }\{-({n_1}+1){\mathcal {L}}^{(0,\lambda )}_{{n_1}+1}{(-x)}\}\\&\overset{(2.44)}{=}-({n_1}+1)(2\lambda ) ^{s-1}L^{(s-1)}_{{n_1}+1}{(-2\lambda x)}e^{-\lambda x}\\ {{}_{-\infty }}\mathrm{D}_{x}^{s,\lambda }\phi ^{+}_{n_2}(x)&=0 \end{aligned} \end{aligned}$$(B.1) -
for \(x\in {\mathbb {R}}^+\), \(0<s<1\),
$$\begin{aligned} \begin{aligned} {{}_{-\infty }}\mathrm{D}_{x}^{s,\lambda }\phi ^{*}(x)&={}_{-\infty }\mathrm{I}_{x}^{1-s,\lambda } {{}_{-\infty }}\mathrm{D}_{x}^{1,\lambda }\phi ^{*}(x){=}\frac{e^{-\lambda x}}{\Gamma (1-s)}\int _{-\infty }^0\frac{2\lambda e^{2\lambda \tau }}{(x-\tau )^s} \mathrm{d}\tau \overset{\tau =x-t}{=}\frac{2\lambda e^{\lambda x}}{\Gamma (1-s)}\\&\quad \int _x^{\infty }\frac{ e^{-2\lambda t}}{t^s} \mathrm{d}t,\\ {{}_{-\infty }}\mathrm{D}_{x}^{s,\lambda }\phi ^{-}_{n_1}(x)&={}_{-\infty }\mathrm{I}_{x}^{1-s,\lambda } {{}_{-\infty }}\mathrm{D}_{x}^{1,\lambda }\phi ^{-}_{n_1}(x)\overset{(3.10)}{=} \frac{e^{-\lambda x}}{\Gamma (1-s)}\\&\int _{-\infty }^0\frac{-(n_1+1) L^{(0)}_{{n_1}+1}{(-2\lambda \tau )}e^{2\lambda \tau }}{(x-\tau )^s} \mathrm{d}\tau \\&\overset{\tau =x-t}{=}-e^{\lambda x}\frac{{n_1}+1}{\Gamma (1-s)} \int _x^{\infty }\frac{L^{(0)}_{{n_1}+1}{(2\lambda (t-x))}e^{-2\lambda t}}{t^s} \mathrm{d}t,\\ {{}_{-\infty }}\mathrm{D}_{x}^{s,\lambda }\phi ^{+}_{n_2}(x)&={}_{-\infty }\mathrm{I}_{x}^{1-s,\lambda }{{}_{-\infty }}\mathrm{D}_{x}^{1,\lambda }\phi ^{+}_{n_2}(x)\overset{(3.9)}{=}\frac{e^{-\lambda x}}{\Gamma (1-s)}\int _0^{x}\frac{({n_2}+1)L^{(0)}_{{n_2}}{(x)}}{(x-\tau )^s} \mathrm{d}\tau \\&=\frac{\Gamma (n_2+2)}{\Gamma (n_2+2-s)}x^{1-s}{\mathcal {L}}^{(1-s,\lambda )}_{{n_2}}{(x)}. \end{aligned} \end{aligned}$$(B.2)
The entries of matrix \({\mathbf {A}}\) with \(1<\mu =1+s<2\).
Since
then,
Similarly, we have
The entries of matrix \({\mathbf {A}}\) with \(0<\mu =s<1\).
Owing to
i.e.,
we obtain that
Similarly, we have
The above equations are enough to calculate out the matrix \({\mathbf {A}}\) due to some symmetric properties of the entries.
Rights and permissions
About this article
Cite this article
Chen, S., Shen, J. & Wang, LL. Laguerre Functions and Their Applications to Tempered Fractional Differential Equations on Infinite Intervals. J Sci Comput 74, 1286–1313 (2018). https://doi.org/10.1007/s10915-017-0495-7
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10915-017-0495-7
Keywords
- Tempered fractional differential equations
- Singularity
- Laguerre functions
- Generalized Laguerre functions
- Weighted Sobolev spaces
- Approximation results
- Spectral accuracy