Abstract
We focus on a timedependent onedimensional spacefractional diffusion equation with constant diffusion coefficients. An allatonce rephrasing of the discretized problem, obtained by considering the time as an additional dimension, yields a large block linear system and paves the way for parallelization. In particular, in case of uniform space–time meshes, the coefficient matrix shows a twolevel Toeplitz structure, and such structure can be leveraged to build adhoc iterative solvers that aim at ensuring an overall computational cost independent of time. In this direction, we study the behavior of certain multigrid strategies with both semi and fullcoarsening that properly take into account the sources of anisotropy of the problem caused by the grid choice and the diffusion coefficients. The performances of the aforementioned multigrid methods reveal sensitive to the choice of the time discretization scheme. Many tests show that Crank–Nicolson prevents the multigrid to yield good convergence results, while secondorder backwarddifference scheme is shown to be unconditionally stable and that it allows good convergence under certain conditions on the grid and the diffusion coefficients. The effectiveness of our proposal is numerically confirmed in the case of variable coefficients too and a twodimensional example is given.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Fractional diffusion equations (FDEs) generalize classical partial differential equations (PDEs), and their recent success is due to the nonlocal behavior of fractional operators that translates in an appropriate modeling of anomalous diffusion phenomena appearing in several applicative fields, like imaging or electrophysiology [2, 4]. The nonlocality of the fractional operators causes absence of sparsity in the discretization matrices. Fortunately, in presence of uniform grids, the discretization matrices show a Toeplitzlike structure and this paves the way for the design of iterative solvers specialized for structured linear systems.
In this regard, for onedimensional spaceFDE problems we mention the circulant preconditioning in [14], the multigrid method in [20], and the structure preserving tridiagonal preconditioners in [6]. The latter preconditioners were motivated by the spectral study of the coefficient matrices through the notion of symbol and by the socalled generalized locally Toeplitz theory [9].
In the twodimensional setting, classical preconditioners based on multilevel circulant matrices are not well suited, while multigrid methods, possibly used as preconditioners, can be effective and robust solvers. Some multigrid proposals for FDEs discretized with finite differences are given in [16,17,18]. When finite element or finite volume discretizations are adopted, multigrid methods are investigated in [13] and [7], respectively. In particular, in [18] the spectral approach presented in [6] has been extended to twodimensional FDEs and has been used to define a multigrid preconditioner which is particularly effective when the fractional orders are both close to 2. For fractional derivative orders that differ from each other, i.e., for FDEs that show an intrinsic anisotropy along the axes, we mention the “multigrid as smoother” (MGS) approach firstly proposed in [19] for integerorder problems and then adapted to fractional ones in [5]. Such method consists in a Vcycle with semicoarsening used as smoother inside an outer fullcoarsening, and in [5] it is shown to be a robust preconditioner in presence of strong anisotropies caused by the fractional orders.
Due to the sequentiality of time integration, with none of the aforementioned approaches we can aspire towards complete independence of time of the overall computational cost. By contrary, an allatonce rephrasing of the discretized problem over a uniform space–time grid, obtained by considering the time as an additional dimension, yields large (multilevel) Toeplitz linear systems and opens to parallelization.
In this regard, we mention the banded Toeplitz preconditioner proposed in [27] for solving nonlinear spaceFDEs, and the block structured preconditioner given in [3] for dealing with arbitrary dimensional space problems. In [10, 21] a Strangtype preconditioner for solving FDEs by boundary value methods has been proposed. In the case of equal left and right diffusion coefficients, we also mention the multigrid reduction in time (MGRIT) discussed in [26]. Therein, the authors consider finite elements in space and Crank–Nicolson in time, since the MGRIT is specifically tailored for one step methods.
The present paper fits within the latter framework. Precisely, our scope is to build a fast and efficient parallelintime structurebased multigrid solver. We fix our attention on a weighted and shifted Grünwald difference (WSGD) discretization of a onedimensional timedependent spaceFDE with constant diffusion coefficients. We stress that this onedimensional problem turns out to be already a tough one, due to the block structure of the coefficient matrix and to its possibly anisotropic character because of the grid choice and the diffusion coefficients.
As for the time discretization, we opt either for Crank–Nicolson (CN) or secondorder backwarddifference formula (BDF2) schemes. The unconditional stability of CNWSGD has already been proven in [23]. Concerning BDF2, in [15] it was combined with a central finite difference scheme for solving spaceFDEs with diffusion coefficients equal to 1. In that same paper, a proof of unconditional stability of the resulting method was given. We extend this result to the case where the space scheme is WSGD and the diffusion coefficients are not necessarily equal to each other.
Aiming at building a parallelintime multigrid, we consider block Jacobi as smoother, since it is parallelizable. Moreover, exploiting the Toeplitz structure of the coefficient matrices and the related symbols we define the projectors according to what has been done in the integerorder literature for both isotropic [1], and anisotropic Toeplitz linear systems [8].
The performances of the proposed multigrid strategies reveal sensitive to the choice of the time discretization scheme. Indeed, many tests, including a twodimensional example, show that Crank–Nicolson prevents the multigrid to yield good convergence results, while BDF2 scheme allows good convergence under certain conditions on the grid and the diffusion coefficients.
The paper is organized as follows. In Sect. 2, we introduce the problem setting and we describe both CNWGSD and BDF2WSGD approximation methods, by giving the formal expression and the structure of the resulting matrices. In Sect. 3, we prove that BDF2WSGD scheme is unconditionally stable. In Sect. 4, we perform an allatonce rephrasing of the original matrices and give some results on their spectra, which are leveraged in Sect. 5 for the design of proper multigrid strategies. Several numerical experiments, also in the case of variable diffusion coefficients, are reported in Sect. 6 for testing the performances of our proposals. Finally, in Sect. 7 we draw conclusions.
2 Preliminaries
In this section, we first introduce the FDE problem we are interested in (Sect. 2.1). Second, we recall the definition of multilevel Toeplitz matrix (Sect. 2.2). The latter will be needed in Sect. 4 where we perform on a allatonce rephrasing of the given problem and we study the related spectral properties. Finally, we briefly review the combination of the chosen finite difference spacediscretization with two different time discretization schemes (Sect. 2.3).
2.1 Problem setting: a onedimensional spaceFDE
In this work, we focus on the following onedimensional initialboundary value spaceFDE problem
where \(\varOmega =(a,b)\) is the space domain, \(d_\pm >0\) are the diffusion coefficients, v(x, t) is the forcing term, and \(\frac{\partial ^{\alpha }u(x,t)}{\partial _{\pm }x^\alpha }\) are the left \((+)\) and right \(()\) fractional derivative operators of order \(\alpha \in (1,2)\) with respect to variable x defined as
By considering the time like an additional dimension, the discretization of equation (1) yields a twolevel Toeplitz matrix. In the next subsection we clarify such structure.
2.2 Multilevel Toeplitz matrices and their symbol
Here we report the formal definition of a multilevel Toeplitz matrix.
Definition 1
Let \(f\in \mathrm {L}^1(\left[ \pi ,\pi \right] ^d)\) and let \({\lbrace f_{\varvec{k}}}\rbrace _{\varvec{k}\in \mathbb {Z}^d}\) be the sequence of its Fourier coefficients defined as
where \(\langle \varvec{k},\theta \rangle =\sum _{t=1}^d k_t\theta _t\) and \(\text {i}^2=1\). Then the dlevel Toeplitz matrix of partial orders \(\varvec{n}=\left( n_1,\ldots ,n_d\right)\) associated with f is
where \(N=\prod _{i=1}^dn_i\) is the order of the matrix. The function f is called the symbol of the matrixsequence \(\lbrace \mathrm {T}_N^{(d)}(f)\rbrace _N\).
To clarify the notation, a 2level Toeplitz matrix of order N generated by f is given by
or equivalently
where \(J_{n_i}^{[j_i]}\in \mathbb {R}^{n_i\times n_i}\) are matrices whose entry (s, t)th equals 1 if \(st=j_i\) and 0 elsewhere. In other words, a 2level Toeplitz matrix is a block Toeplitz whose blocks are Toeplitz. When \(d=1\), we simplify the notation using
2.3 Space–time discretizations: CNWSGD and BDF2WSGD
In the following, we briefly review the finite difference spacediscretization of problem (1) obtained using the second order accurate Weighted and Shifted Grünwald Difference (WSGD) scheme [23] combined with either Crank–Nicolson (CN) or secondorder backwarddifference formula (BDF2) schemes in time.
Let \(N,M\in \mathbb {N}\) and consider the following uniform space–time grid
According to [23], the discretization of \(\frac{\partial ^{\alpha }}{\partial _+ x^\alpha }\) yields a lower Hessenberg Toeplitz matrix \(A_N^\alpha = \mathrm {T}_N(f_\alpha )\), where
with
and \(g_k^{(\alpha )}=(1)^k{\alpha \atopwithdelims (){k}}\). A similar reasoning shows that the discretization of \(\frac{\partial ^{\alpha }}{\partial _ x^\alpha }\) yields an upper Hessenberg Toeplitz matrix, which coincides with \(A_N^T\).
The discretization of the forcing term returns vector \(v^m=[v(x_i,t^m)]_{i=1}^N\) and the application of CN and BDF2 schemes in time gives the following linear systems
respectively, where \(r=\frac{\varDelta t}{2\varDelta x^\alpha }\) and
Remark 1
In the case of BDF2, the solution \(u^1\) at time \(t^1\) is computed with CN and outside the allatonce linear system. Note that any other one step method could be used to compute \(u^1\). For example, although Implicit Euler is only first order accurate, if we only use it once, it will not compromise the global secondorder accuracy. Such a statement can be found in [24].
The following proposition, which plays an important role in the definition of the projectors for our multigrid strategy (see Sect. 5), defines the symbol of the spatial discretization.
Proposition 1
Let \(d_\pm =d\), then \(A_{x,N}=d\cdot \mathrm {T}_N(g_\alpha )\), where \(g_\alpha (x)=f_\alpha (x)+\overline{f}_\alpha (x)\) is nonpositive and has a zero of order \(\alpha\) at \(x=0\).
3 Stability of the BDF2WSGD scheme
In [23] the authors proved the unconditional stability of the CNWSGD scheme (2), in the constant diffusion coefficients case, as a consequence of the following theorem.
Theorem 1
[23] Let \(\lambda\) be an eigenvalue of \(A_{x,N}\), then \(\mathrm {Re}(\lambda )<0,\ \forall \alpha \in (1,2)\).
Indeed, in case of a one step scheme like CN, the stability relies on the spectral radius of the iterations matrix of the time stepping algorithm, which is required to be lower than 1. We now aim to prove the stability of the BDF2WSGD scheme. After the discretization in space we obtain an equation of the form
where \(\mathbf {u}(t),\mathbf {v}(t)\) are, respectively, the semidiscrete in space unknown and forcing term.
The region of absolute stability of a linear multistep method is defined as follows.
Definition 2
(Absolute stability region) The region of absolute stability for the linear multistep method
is the set of points \(z\in \mathbb {C}\) for which the roots \(\lbrace \zeta _j\rbrace _{j=1}^s\) of the polynomial
satisfies the following root conditions:

(a)
\(\left\zeta _j\right\le 1\), for \(j=1,\ldots ,s\),

(b)
if \(\zeta _j\) is a repeated root, then \(\left\zeta _j\right<1\).
Therefore, again as a consequence of Theorem 1, the following theorem holds.
Theorem 2
Let \(\alpha \in (1,2)\) and consider \(A_{x,N}\) to be diagonalizable, then BDF2WSGD scheme (3) is unconditionally stable.
Proof
Since \(A_{x,N}\) is diagonalizable there exists an invertible matrix V such that \(A_{x,N}=V^{1}\varLambda V\), where \(\varLambda\) is the diagonal matrix containing the eigenvalues of \(A_{x,N}\). Therefore, by introducing \(\tilde{u}=Vu\), Eq. (3) can be written as N uncoupled equations with respect to \(\tilde{u}^m\):
Let us fix a row index i, then, by definition of r,
which, in the polynomial form of Eq. (5), becomes
By defining \(\tilde{z}:=1\frac{2z}{3\varDelta x^\alpha }\), which is a complex number and can be written as \({\tilde{z}}=a+\text {i}b\), it follows that the roots are
From Theorem 1, we have that \(\text {Re}(\tilde{z})=a>1\) for \(\alpha \in (1,2)\), and the study of the maximum of function g(a, b) shows that \(\mathop {\sup }\nolimits _{a>1}\ g(a,b)<1\). \(\square\)
The following corollary exploits the density of diagonalizable matrices into the space of square matrices to remove the diagonalizability hypothesis in Theorem 2.
Corollary 1
Let \(\alpha \in (1,2)\), then BDF2WSGD scheme (3) is unconditionally stable.
Proof
Let us suppose that \(A_{x,N}\) is not diagonalizable, otherwise the thesis follows from Theorem 2. Let us consider the Schur decomposition \(QTQ^\mathrm {H}\) of \(A_{x,N}\), where Q and T are unitary and upper triangular matrices, respectively. Note that due to the structure of T, its diagonal elements are the eigenvalues of T and that, by similarity, they coincide with the eigenvalues of \(A_{x,N}\).
Since we are assuming that \(A_{x,N}\) is not diagonalizable, T has at least two diagonal elements that are equal. Let us then consider matrix \(B_N=Q\tilde{T}Q^\mathrm {H}\), where \(\tilde{T}\) is obtained from T by properly shifting its diagonal entries such that \(\tilde{T}\) becomes diagonalizable. More precisely, for \(\epsilon >0\) and for \(i=1,\ldots ,N\) the ith diagonal element of \(\tilde{T}\) is \(\tilde{t}_{ii}=t_{ii}\delta _i\) with \(0\le \delta _i<\epsilon\) such that \(\tilde{t}_{ii}\ne \tilde{t}_{jj},\ \forall i,j=1,\ldots ,N\) and \(i\ne j\).
Since \(B_N\) is diagonalizable and its eigenvalues have negative real part, Theorem 2 applies to \(B_N\), and since
by letting \(\epsilon \rightarrow 0\) the thesis is proven. \(\square\)
Remark 2
The unconditional stability of BDF2 combined with a central finite difference scheme for discretizing the fractional derivative operator, was given in [15]. Therein, the diffusion coefficients were both equal to 1. Under the diagonalizability hypothesis, Theorem 2 extends the unconditional stability of BDF2 to the case where the space scheme is WSGD and the diffusion coefficients are not necessarily equal to each other. Corollary 1 generalizes Theorem 2 to the case of a non diagonalizable matrix \(A_{x,N}\).
4 Allatonce rephrasing of our problem and related spectral study
Starting from Eqs. (2) and (3), and chaining the unknown \(u^{m}\) at each time step into a unique vector as
in the case of CN and BDF2, respectively, we can rephrase the original discretized problem as the following large block linear systems
where
and
Note that both coefficient matrices \(A_{\text {CN}}\), \(A_{\text {BDF2}}\) are twolevel Toeplitz according to Definition 1, and hence we can compute their symbol.
From Proposition 1 we recall that \(A_{x,N}=d\cdot \text{T}_N(g_\alpha )\), where \(g_\alpha (x)\) has a zero of order \(\alpha\) at \(x=0\). Then, by assuming \(d_\pm =d\) and \(N,M\rightarrow \infty\), we have
with \(h_\text {S}\) defined as follows:

(i)
if \(r\) is constant, then
$$\begin{aligned} \begin{aligned} \bullet&\quad h_{\text {CN}}(x,t)=1\text {e}^{\text {i}t}d\cdot rg_\alpha (x)\left( 1+\text {e}^{\text {i}t}\right) ,\\ \bullet&\quad h_{\text {BDF2}}(x,t)=1\frac{4}{3}\mathrm {e}^{\text {i}t}+\frac{1}{3}\mathrm {e}^{2\text {i}t} \frac{4}{3}d \cdot rg_\alpha (x), \end{aligned} \end{aligned}$$(7)and both functions have a unique zero at \((x,t)=(0,0)\) of order 1 and α in t and x, respectively;

(ii)
if \(r\rightarrow 0\), then from (7) we have
$$\begin{aligned} \bullet&\quad h_{\text {CN}}(x,t)=1\mathrm {e}^{\text {i}t},\\ \bullet&\quad h_{\text {BDF2}}(x,t)=1\frac{4}{3}\mathrm {e}^{\text {i}t}+\frac{1}{3}\mathrm {e}^{2\text {i}t}, \end{aligned}$$and both functions have a zero of order 1 at \(t=0,\ \forall x\);

(iii)
if \(r\rightarrow \infty\), by grouping up r in (7), we have
$$\begin{aligned} \bullet&\quad h_{\text {CN}}(x,t)=d(1+\mathrm {e}^{\text {i}t})g_\alpha (x),\\ \bullet&\quad h_{\text {BDF2}}(x,t)= \frac{4}{3}d\cdot g_\alpha (x), \end{aligned}$$where \(h_{\text {CN}}\) has a zero of order 1 at \(t=\pi ,\ \forall x\) and a zero of order \(\alpha\) at \(x=0,\ \forall t\), while \(h_{\text {BDF2}}\) vanishes only at \(x=0,\ \forall t\) with order \(\alpha\).
The presence of at least a line of zeros in the symbol is called anisotropy. The latter becomes stronger as the number of such lines increases. We expect then case (iii) for CN to be much harder to be numerically treated than all other cases.
Remark 3
If we suppose r to be constant and let \(d\rightarrow \infty\) or \(d\rightarrow 0\) then the same results as in case (ii) and (iii), with r in place of d, hold. In practice, since r, d are fixed coefficients, the anisotropy arises when \(d\cdot r\) is very large or very small.
Remark 4
The study of the symbol \(h_\text {S}\) can easily be extended to the case where \(d_+\ne d_\) using the results in [6].
5 Multigrid methods for allatonce systems
This section is devoted to the design of multigrid strategies based on the spectral study performed in Sect. 4 for the linear systems in (6). With this aim, we first recall the basics of the multigrid in Sect. 5.1, then we discuss our multigrid proposals in Sect. 5.2.
5.1 Multigrid idea and convergence results
Multigrid methods, introduced in [22], combine two iterative methods known as smoother and coarse grid correction (CGC). Given the linear system \(A_Nx=b\), \(A_N\in {\mathbb {R}}^{N\times N}\), the former is typically a stationary iterative method, which we denote with \(\mathcal {S}^{\nu }\left( x,A_N,b\right)\), where \(\nu\) is the number of iterations.
To give a precise account of what the CGC is, let us consider the most basic version of multigrid, often used for proving convergence results, i.e., the twogrid method (TGM). Given a fullrank matrix \(P_N\in {\mathbb {R}}^{N\times k}\), with \(k<N\), a step of TGM is defined by Algorithm 1.
Steps 1) to 5) define the CGC and step 6) is called postsmoother. To strengthen the algorithm, a second smoother, which is called presmoother, could be added before the CGC. VCycle is obtained from TGM by replacing Step 4) with a recursive call of TGM. The Garlerkin approach is useful for the convergence analysis, but in practice it could be too computationally expensive. Therefore, it is often replaced by a geometric approach, which consists in the rediscretization of the equation over a coarser grid. This leads to a less robust algorithm, but allows to maintain the same structure of the coefficient matrix at each level and is highly parallelizable.
Moreover, if the matrix \(A_N\) shows a block Toeplitzlike structure, by choosing the geometric approach the matrix–vector products can be performed by means of the fast Fourier transform algorithm at each coarser grid.
Let \(K_N\in \mathbb {R}^{N\times {\lfloor }{\frac{N}{2}}{\rfloor }}\) be the downsampling matrix, which keeps an entry every two, then the onedimensional grid transfer operator \(P_N\) is defined as
In more than one dimension, the projector is defined through the tensor product. For instance, in two dimensions the required projector \(P_{NM}\) is defined as
where
are the projectors in space and time generated by \(p_1(x)\) and \(p_2(t)\), respectively. Notice, that this is the case we are interested in. Indeed, by considering the time as an additional dimension, we have reinterpreted the original onedimensional problem as a twodimensional one.
The convergence of the Vcycle, proved in [1], requires the following condition on \(p(x,t)=p_1(x)p_2(t)\). Let \(f\ge 0\) be the symbol of \(A_N\), vanishing only at \((x_0,t_0)\), then p has to satisfy
where \(\mathscr {M}(x,t)=\{(x,\pi t),(\pi x,t),(\pi x,\pi t)\}\) is the set of the “mirror points" of (x, t).
In the case where the symbol f has a whole line of zeros along the axes, relation (8) does not hold anymore. In such a case, an efficient alternative to standard Vcycle is given by the socalled semicoarsening [8]. The projector in the semicoarsening approach is defined by considering the vanishing variable as a parameter, and then building the projector in the remaining variable according to the onedimensional version of relation (8). The projector in the other dimension is simply given by an identity matrix.
5.2 Multigrid methods for the allatonce systems
We now see how the results recalled in Sect. 5.1 apply to our case. In particular, we define the projector according to condition (8) and the properties of \(h_\text {S}\) defined in Sect. 4. Let us first introduce the polynomials
Note that:

\(p_1(x)\) has a zero of order 2 at \(x=\pi\);

\(p_{2}^+(t)\) has a zero of order 1 at \(t=\pi\);

\(p_{2}^(t)\) has a zero of order 1 at \(t=0\).
In the case of CN, according to the analysis in Sect. 4, we distinguish the following three cases:

(1)
If r is constant, then \(h_{\text {CN}}\) has a unique zero of minimum order 1 at \((x,t)=(0,0)\). The mirror points of \((x,t)=(0,0)\) are \((0,\pi ),(\pi ,0),(\pi ,\pi )\). Since \(p_1(x)\) and \(p_{2}^+(t)\) vanish at \(x=\pi\) and \(t=\pi\), respectively, \(p(x,t)=p_1(x)p_{2}^+(t)\) vanishes at the mirror points with a minimum order of 1 and hence satisfies relation (8).

(2)
In the anisotropic case where \(r\rightarrow 0\), \(h_{\text {CN}}\) is zero on the whole xaxis, then we opt for semicoarsening in time. Precisely, by considering variable x as a parameter, \(h_{\text {CN}}\) has a unique zero of order 1 at \(t=0\). Then, we generate \(P_{t,M}\) through \(p_{2}^+(t)\) that has a zero of order 1 at the mirror point \(t=\pi\). The projector \(P_{x,N}\) is given by the identity matrix \(I_N\).

(3)
In the anisotropic case, where \(r\rightarrow \infty\), \(h_{\text {CN}}\) is zero on both axes. The theory does not apply to this scenario. Nevertheless, as a first attempt to dominate (at least partially) this other kind of anisotropy, we use again standard semicoarsening in both time and space. Precisely,

When x is considered as a parameter, \(h_{\text {CN}}\) has a zero of order 1 at \(t=\pi\). Therefore, we perform semicoarsening in time replacing \(P_{x,N}\) with \(I_N\) and generating \(P_{t,M}\) through \(p_{2}^(t)\), whose first order zero at \(t=0\) satisfies relation (8) for x fixed, i.e., considering only the onedimensional problem in the variable t.

When t is considered as a parameter, \(h_{\text {CN}}\) has a zero of order \(\alpha\) at \(x=0\). Therefore, the semicoarsening in space is defined by replacing \(P_{t,M}\) with the identity matrix \(I_M\) and generating \(P_{x,N}\) through \(p_1(x)\), which has a zero of order 2 at the mirror point \(x=\pi\), again according to condition (8) for t fixed.

In the case of BDF2, items 1) and 2) are identical. Regarding item 3), i.e. when \(r\rightarrow \infty\), symbol \(h_{\text {BDF2}}\) vanishes only over the line \(x=0,\ \forall t\) with order \(\alpha\). Hence we have a standard anisotropy, like in item 2), but this time along the taxis. Therefore, we consider a semicoarsening in space by setting \(P_{t,M}=I_M\) and by generating \(P_{x,N}\) through \(p_1(x)\), whose position and order of the zero satisfies relation (8) for t fixed.
Concerning the smoother, we consider \(\omega\)weighted block Jacobi method (\(\omega\)BJ), where the diagonal blocks are of the form \(I\xi r A_{x,N}\) with \(\xi=1\) for CN and \(\xi=4/3\) for BDF2. The reason for such a choice is that it is parallelizable and it allows to exploit the structure of the coefficient matrix. Moreover \(\omega\)BJ converges for any \(\omega \in (0,1]\) whenever the blocks are of size \(N\times N\), hence the study of its relaxation parameter is only related to the smoothing property along the time axis (see Sect. 6.1 for a numerical discussion about this issue). In our tests, the inversion of the blocks in blockJacobi is performed through the Matlab function backslash.
We stress that, choosing standard Jacobi would decrease the computational cost and would also exploit the structure of the coefficient matrix, but its use would ask for a tougher study of the relaxation parameter in order to ensure the convergence.
Remark 5
Note that in case of BDF2, when \(d\rightarrow \infty \text { or }r\rightarrow \infty\), the coefficient matrix in Eq. (6) tends to a block diagonal Toeplitz matrix. This means that, when \(d\cdot r\) becomes large, using the multigrid is pointless since its smoother is already computing the solution accurately enough.
6 Numerical results
In this section we investigate the performances of \(\omega\)BJ, twogrid method (TGM) and VCycle (V), mainly used as standalone solvers for solving the linear system in (6). Few numerical results concerning the use of both TGM and V as preconditioners (only one iteration) for GMRES are also given.
Both TGM and VCycle will have one iteration of \(\omega\)BJ as postsmoother and no presmoothing iteration (the reason is given in Sect. 6.1). In VCycle we halt the coarsening at the 5th level, when the coefficient matrix has a minimum size of \(N_5\times M_5\) with \(N_5\ge \frac{N}{2^5}\) and \(M_5\ge \frac{N}{2^5}\), depending on the coarsening technique, and the solution on the coarsest level is performed through the backslash Matlab function, which is a direct solver.
As already clarified in Remark 1, the solution \(u^1\) at the first time step is computed outside the coefficient matrix. One could of course include the computation of \(u^1\) in the coefficient matrix as done in [11]. In our case, due to the computationally expensive smoother we use, the difference between the two approaches is negligible. A comparison between the two approaches in terms of iterations can be found in Sect. 6.6.
The section is organized as follows. In the first part we aim at explaining how we fix the fractional derivative order \(\alpha\), and the relaxation parameter \(\omega\) in our numerical examples. Precisely, in Sect. 6.1 we test the performances of ωBJ for two different values of \(\omega\) and we show that it generates jumps along the time axis, independently of \(\omega\). In Sect. 6.2, we check how much TGM is sensitive to \(\alpha\), and we numerically prove that its behavior is only slightly \(\alpha\)dependent.
Aside from \(\alpha\) and \(\omega\), we also need to clarify how we choose between the two projector generators \(p^+_2\) and \(p^_2\) discussed in Sect. 5.2 when performing semicoarsening in time for CN. This is the subject of Sect. 6.3. In Sects. 6.4–6.5 we perform few tests with large N, M to numerically check the robustness of TGM and V as N, M increase in both constant and variable diffusion coefficients cases. Finally, in Sect. 6.6 we provide a twodimensional example.
All our tests have been run on a server with Intel(R) Xeon(R) Silver 4114 at 2.20GHz with Matlab 2019b. For all methods we fix the tolerance to \(10^{7}\) and the initial guess as the null vector. We use the builtin gmres function, whose preconditioner is leftsided. For this reason we force GMRES to reach the required tolerance on the actual residual through a ‘by hand’ restart. A right preconditioned GMRES could be of course employed and it would basically give the same amount of iterations.
Notation. In the following, we denote with TGMp (resp. Vp), p\(\in \lbrace \text {x},\text {t},\text {xt}\rbrace\) the TGM (resp. V) that uses ωBJ as postsmoother and performs semicoarsening in space (\(\mathrm{p}=\mathrm{x}\)), time (\(\mathrm{p}=\mathrm{t}\)), or both space and time (\(\mathrm{p}=\mathrm{xt}\)). Precisely:

‘x’ denotes the space semicoarsening, whose projector is generated by \(p_1(x)\);

‘t\(_\pm\)’ denotes the time semicoarsening, whose projector is generated by \(p_2(t)^\pm\);

‘xt\(_\pm\)’ denotes the fullcoarsening, whose projector is generated by \(p_1(x)p_2(t)^\pm\).
The addition of ‘(G)’ after the solver name stands for ‘Galerkin approach’. In case nothing is specified, geometric approach is adopted. Finally, the presence of ‘(\(\mathcal {P}\))’ in the name of the solver means that the considered multigrid method is set as GMRES preconditioner.
We point out that due to space limitations, in the key of each figure we omit the name of the method and specify only the projector. For instance, we write simply xt\(_+\) in place of TGMxt\(_+\). The name of the method will be clear from the caption of the figure.
All the results contained in the Sects. 6.1–6.4 refer to the following example.
Example 1
In this example we assume the diffusion coefficients to be constant and equal, that is \(d_\pm =d\). The space and time domains in problem (1) are fixed as \(\varOmega =(0,2)\), and [0, 1] respectively, while the true solution and the solution at \(t=0\) are given by
The numerical approximation of v is computed starting from the discretized exact solution.
6.1 Behavior of ωBJ smoother
Here we test the “smoothing properties” of \(\omega\)BJ. Let us consider \(\alpha =1.5\), \(d=1\), \(N=63\) and, to better point out the behavior of ωBJ along the time axis, we fix \(M=7\ll N\).
Figure 1 shows the error, reshaped as a space–time surface, after one iteration of \(\omega\)BJ for both linear systems in Eq. (6).
We note that, in case of CN, one iteration of 1BJ generates a jump along the time axis from \(t^1\) to \(t^2\). In Fig. 1b, where the 2step method BDF2 is considered, such a jump involves also \(t^3\) due to the longer stencil of BDF2 with respect to CN. On the other hand, both surfaces in Fig. 1 do not show any jump along the xaxis.
We now analyze the jump in time varying the magnitude of \(\mu :=d\cdot r\), where r is the grid dependent scale parameter. We introduce the function
which measures the vertical displacement of the discrete error E(x, t) in the first three time steps at the midpoint \(x=1\). We note that \(\text {dist}(E)=0\) if and only if E is constant in the first three time steps. In other words, as far as \(\text {dist}(E)\approx 0\), E is smooth, and this indicates that \(\omega\)BJ is a good smoother.
Figures 2a and b show how \(\text {dist}(E)\) behaves for both CN and BDF2, fixed \(N=M=63\), \(\omega =1,0.5\), \(\alpha =1.1, 1.8\), and varying \(\mu \in [10^{5},10^5]\). As we can see, both discretizations are characterized by a region where the jump is negligible. In detail, the jump generated in CN is negligible only when \(\mu \approx 10\). In BDF2, instead, the jump becomes negligible as \(\mu\) increases. Moreover, for both CN and BDF2, the jump slightly moves while varying \(\alpha\), and it halves its magnitude when switching from \(\omega =1\) to \(\omega =0.5\).
In summary, in all the considered cases \(\omega\)BJ generates jumps along the time axis which means that the projection along such axis could be inaccurate. In order to face this drawback, in the following we only consider \(\omega\)BJ as postsmoother avoiding presmoothing iteration at the first iteration. This choice is supported by the idea that applying the CGC before the smoother could reduce the jump, preventing then the projection of a nonsmooth error.
6.2 Behavior of TGM varying \(\alpha\)
In Sect. 6.1, we observed that \(\text {dist}(E)\), in Eq. (9), slightly varies with \(\alpha\). This could lead to a difference in the behavior of the multigrid depending on \(\alpha\), when solving the two linear systems in (6). Here we perform few tests which show that the behavior of the proposed TGM is almost independent of \(\alpha\) and hence that justify the choice of a fixed value for \(\alpha\) in the reminder of the numerical tests.
In Figs. 3 and 4, we check the number of iterations of TGMx, TGMt, and TGMxt, with fixed \(N=M=2^61\), and varying \(\mu \in [10^{5},10^{5}]\), and \(\alpha \in (1,2]\). Concerning the choice of \(\omega\) in the \(\omega\)BJ smoother, several tests (not reported here because of space limitations) show that, in the case of CN, the choice of \(\omega =1\) causes bad convergence results for both TGMxt and TGMt. On the other hand, \(\omega =0.5\) provides a good convergence, according also to the analysis in Sect. 6.1, for both CN and BDF2. Therefore, in the rest of this section we fix \(\omega =0.5\). We stress that such discussion on the relaxation parameter is not intended as a substitute of a rigorous study, and that a theoretical approach to the subject will be investigated in a future work.
Figure 3, where we use CN scheme, shows that by increasing \(\alpha\) the optimal region of convergence (blue) shifts to the left for any of the considered algorithms. Regarding BDF2, instead, Fig. 4 shows that the blue region shifts to the left as \(\alpha\) increases only in the case of TGMx. In the other two cases, their number of iterations stays almost independent of \(\alpha\).
Summarizing, the width of the blue regions does not seem to significantly change while varying \(\alpha\). Therefore, in the following we restrict our analysis to the case where \(\alpha =1.5\).
6.3 Time projection performances for the CN scheme
In Sect. 4 we have shown that the symbol \(h_\text {CN}\) has a zero in t that moves from 0 to \(\pi\) depending on how r or d behave asymptotically, and then on the magnitude of \(\mu =d\cdot r\). As discussed in Sect. 5.2, this means that the projector in time must change as well from t\(_+\) to t\(_\) according to \(\mu\). Here, we show that the latter does not work satisfactorily in practical applications when \(\mu\) is large.
Let us fix \(\alpha =1.5\) and \(N=M=2^71\). Figure 5a shows the iterations to tolerance of TGMt\(_{\pm }\), TGMt\(_{\pm }\)(G), TGMt\(_{}\)(\(\mathcal {P}\)) while varying the magnitude of \(\mu \in [10^{5},10^5]\). We note, in line with the discussion in Sect. 5.2, that the Galerkin approach allows TGMt\(_+\)(G) to converge in a low amount of iterations when \(\mu \in [10^{5},1]\), that is for small values of \(\mu\). When considering the less robust geometric approach, TGMt\(_+\) still yields good convergence results in the same range of \(\mu\), even if the iteration number slightly increases.
In the case where \(\mu \gg 1\), the only working method is TMGt\(_\)(G). Unfortunately, the Galerkin approach is not of practical use since it is too computationally expensive. Regarding the geometric approach, TGMt\(_\) results unpractical also when used as GMRES preconditioner (refer to TGMt\(_\)(\(\mathcal {P}\)) in Fig. 5a).
The reason why geometric and Galerkin methods behave differently is due to the large difference between the matrices at the coarser level obtained with the two approaches. Indeed, the convergence condition given in (8) requires the Galerkin approach, which leads to a coarser matrix having a symbol that vanishes at the origin (see [1] for details). Differently, the geometric approach, which consists in discretizing the same problem over a coarser grid, builds a coarser matrix that vanishes again at \(t=\pi\) and that shows then opposite spectral behavior with respect to Galerkin.
In Fig. 5b, TGM with fullcoarsening is shown not to work in the anisotropic cases \(\mu <10^{1}\) and \(\mu >10\), independently of the time projectors and the approach for computing the matrix at the coarser level.
In conclusion, in the following we only consider the time projector given by t\(_+\) and we denote it simply with ‘t’, since it is the only projector that allows TGM with both semicoarsening in time and fullcoarsening to yield good convergence results for the geometric approach.
6.4 Comparison between CN and BDF2: TGM and Vcycle performances
Now we discuss how the performances of TGM and Vcycle with both semi and fullcoarsening vary depending on the adopted discretization scheme, i.e., CN or BDF2.
Let us discuss first the behavior of TGM. In Fig. 6, we compare the iterations to tolerance of TGMxt, TGMx, TGMt for both CN and BDF2 varying the magnitude of \(\mu \in [10^{5},10^5]\) and fixed \(N=M\in \{2^71,2^81\}\). In case of BDF2, the iteration number of 1BJ used as a standalone solver are displayed as well.
We note that, in the case of CN, the iterations to tolerance of TGMt look stable for \(\mu <1\) as N, M increase. The same holds for TGMxt and TGMx when \(\mu \approx 1\). Nothing seems to work when \(\mu > 10\) again independently of N, M, which is what we are expecting due to the strong anisotropy of this specific case discussed in Sect. 5.2.
In the case of BDF2, again according to our theoretical analysis, TGMx and TGMt yield good convergence results when \(\mu >1\) and \(\mu <1\), respectively, and both are stable as N, M increase. In line with what we observed in Sect. 6.1, the high number of iterations of TGMt, even if constant, could be due to the bad smoothing effects along the time axis of 0.5BJ. Concerning TGMxt, it yields the same iterations to tolerance as TGMt when \(\mu >10^{1}\). Note that for this example both TGMxt and TGMt work where they are not supposed to, i.e., in the anisotropic case \(\mu \gg 1\). This is due to the smoother that, according to Remark 5, is already a robust enough solver.
Due to the high computational cost of TGM, in Fig. 7 we switch from TGM to Vcycle and we check its behavior depending on the chosen time discretization scheme.
In the case of CN, conversely to the results obtained for TGM, the only projector which seems to allow Vcycle to converge in a reasonable amount of iterations, is the semicoarsening in space. However, it works in a really small region, i.e. when \(\mu \approx 1\), which is close to the region where one iteration of 0.5BJ yields a smooth solution (go back to Fig. 2).
Regarding BDF2, Vx converges in almost the same amount of iterations as TGMx when \(\mu > 1\). In particular, for \(\mu \in \left[ 1,10^3\right]\) the block diagonal part of \(A_\text {BDF2}\) is not dominant and hence 0.5BJ has a slow convergence, but when it is used as smoother in Vx we obtain a robust and fast convergent method. Moreover, we note that the region where 0.5BJ used as standalone solver is already enough robust becomes smaller as the meshsize increases. This is not the case for Vx, which is then faster than 0.5BJ in a wider range of \(\mu\) as N, M become large. Concerning Vt and Vxt, their plots are basically superposed independently of \(\mu\), and they perform well only for large values of \(\mu\) again because of the ωBJ smoother.
In the case of a semicoarsening in space only, since time interpolation is not involved, larger values of \(\omega\) could be used. Tests which are not reported here show a reduction in the iteration number of the multigrid with ωBJ, with \(\omega \approx 1\), when applied to both CN and BDF2 for almost the same values of \(\mu\) where it performs well with \(\omega =0.5\).
We note that, for both CN and BDF2, the iterations to tolerance of all the tested Vcycles stay almost stable as N, M increase. Moreover our results are in line with the results reported in Figure 2, Section 4.4 of [12], where multigrid with coloured pointwise Gauss–Seidel as smoother is used to solve a space–time linear system obtained from the discretization of a standard timedependent diffusion equation.
6.5 A variable diffusion coefficients example
We now consider the example taken from [14] in which the diffusion coefficients are not constant.
Example 2
We assume the space and time domains in problem (1) as \(\varOmega =(0,2)\), and [0, 1] respectively, and define the diffusion coefficients, the true solution and the solution at \(t=0\) as follows
where \(\varGamma\) is the gamma function and \(d>0\). When \(d=1\), the forcing term is given by
while in the remaining cases, the numerical approximation of v is computed starting from the discretized exact solution.
In Fig. 8, like in Sect. 6.4, we test the behavior of Vx, Vt, Vxt for two fine grids with \(N=M=2^71\) and \(N=M=2^81\).
We note that, as in the case of constant and equal diffusion coefficients, the results are not significantly sensitive to N, M. Moreover, like in Example 1, Vx is the only Vcycle, between the three tested, that yields good convergence results for both CN and BDF2. Finally, also in this variable coefficients example, the optimal convergence region, given by the magnitude of \(\mu =d\cdot r\), is much bigger in the case of BDF2 than of CN.
We note that, independently of the constant or variable diffusion coefficients character, none of the tested methods is robust enough to deal with the case where \(\mu <1\). Further tests, not reported here, show that this holds unchanged even when using the Vcycle as preconditioner for the GMRES. On the other hand, we stress that, since d is fixed, the choice of an opportune grid could lead to \(\mu >1\), choosing \(\varDelta t\) and \(\varDelta x\) such that \(d\varDelta t>2\varDelta x^\alpha\) and making Vcycle a suitable solver again.
6.6 Two dimensional case
We end the numerical section by providing numerical results in the twodimensional case. We consider the following extension of the onedimensional FDE in Eq. (1):
where \(\varOmega =(a_1,b_1)\times (a_2,b_2)\) is the space domain and \(d_\pm ,e_\pm >0\) are the diffusion coefficients.
The discretization follows from the onedimensional case and yields the same coefficient matrices as in Eq. (6), but with the extension through Kronecker product in two dimensions of each block. In the case where \(d_+=d_=e_+=e_=d\), \(\alpha =\beta\) and both spatial steps \(\varDelta x,\varDelta y\) are equal, it holds that the grid dependent scale factor which multiplies the matrix representing the discretization in space is the same as in the onedimensional case, i.e. \(\mu =d\frac{\varDelta t}{2\varDelta x^\alpha }\).
Example 3
For our test we extend Example 1 to the twodimensional case by assuming \(d_\pm =e_\pm =d\), \(\varOmega =(0,2)\times (0,2)\), and taking as final time step \(T=1\). The true solution and the solution at the initial time \(t=0\) are, respectively, given by
As done in Example 1, the numerical approximation of v is computed starting from the discretized exact solution.
Since the timecoarsening does not seems to be effective in the onedimensional case, here we only consider the coarsening in both spatial dimensions and we use 0.95BJ as postsmoother (higher weights seemed more suitable for this case).
Due to hardware limitations we cannot choose too dense grids, therefore we fix \(N_x=N_y=M=2^61\), where \(N_x\) and \(N_y\) are the amount of points over the grids in the first and second spatial dimensions and M are the points over the time grid. Moreover, when considering the BDF2 scheme, as in the onedimensional case we use CN to compute the solution at the first time step. As done in [11], we consider the case where the computation of the solution \(u^{1}\) at the first the step is included in the coefficient matrix (inner CN) and we compare it with the previously considered case (outer CN), where \(u^{1}\) is computed outside the coefficient matrix.
In Fig. 9 we show the iterations to tolerance of the multigrid used as standalone solver (denoted by ‘xy’) for solving Eq. (10) discretized with CN, BDF2 with inner CN and BDF2 with outer CN. In the case of BDF2 with outer CN, we compare the results with 1BJ.
We note that, even in the twodimensional case, multigrid applied to CN is efficient when \(\mu \approx 1\) and applied to BDF2 when \(\mu \ge 1\). Moreover, we observe that the plots of BDF2 with inner and outer CN overlap almost everywhere, therefore the addition of CN inside the coefficient matrix does not seem to compromise the convergence of multigrid.
When \(\mu \ge 10^2\), as in the onedimensional case, 1BJ is an efficient solver, since the coefficient matrix becomes block diagonally dominant.
7 Conclusions and future works
In this work we focused on an allatonce rephrasing of a timedependent onedimensional spaceFDE with constant diffusion coefficients discretized with WSGD in space and CN or BDF2 in time. The unconditional stability of the BDF2WSGD scheme has been proven, and the twolevel Toeplitz structure of the resulting linear systems has been leveraged to design multigrid strategies that use block Jacobi as smoother and whose projectors definition is driven by the symbol.
We have numerically shown that Vcycle with semicoarsening in x is the only multigrid, among all the tested ones, that yields good convergence results for both BDF2 and CN schemes. Moreover, it performs satisfactorily under less restrictive assumptions on the magnitude of \(\mu =d\cdot r\) in the case of BDF2 than in the case of CN, and this let us to conclude that BDF2 is a much better alternative to CN for parallelintime integration with multigrid, when \(\mu\) is large.
As future works, we plan to properly study the relaxation parameter of block Jacobi, as well as to consider alternative smoothers that could allow multigrid convergence even when \(\mu \ll 1\). Moreover, it would be interesting to investigate the stability of higher order BDF schemes for smooth solutions in time, which could potentially increase the effectiveness of a multigrid solver reducing, at the same time, the overall computational cost. Furthermore, we aim at providing a parallel implementation of such strategies and to extend our structurebased approach also to other stateoftheart solvers like parareal [25] or MGRIT [26]. Finally, we aim at extending our results also to the variable coefficients case and to higherdimensional problems. This opens to other kind of anisotropies for which the Vcycle with semicoarsening could not work anymore and that could, instead, be treated with the MGS already applied to twodimensional FDEs in [5].
References
Aricò, A., Donatelli, M., SerraCapizzano, S.: Vcycle optimal convergence for certain (multilevel) structured linear systems. SIAM J. Matrix Anal. Appl. 26, 186–214 (2004)
Bai, J., Feng, X.: Fractionalorder anisotropic diffusion for image denoising. IEEE Tran. Image Process. 16, 2492–2502 (2007)
Bertaccini, D., Durastante, F.: Block structured preconditioners in tensor form for the allatonce solution of a finite volume fractional diffusion equation. Appl. Math. Lett. 95, 92–97 (2019)
BuenoOrovio, A., Kay, D., Grau, V., Rodriguez, B., Burrage, K.: Fractional diffusion models of cardiac electrical propagation: role of structural heterogeneity in dispersion of repolarization. J. Royal Soc. Interface (2014)
Donatelli, M., Krause, R., Mazza, M., Trotti, K.: Multigrid preconditioners for anisotropic spacefractional diffusion equations. Adv. Comput. Math. 46, Article number 49 (2020)
Donatelli, M., Mazza, M., SerraCapizzano, S.: Spectral analysis and structure preserving preconditioners for fractional diffusion equations. J. Comput. Phys. 307, 262–279 (2016)
Donatelli, M., Mazza, M., SerraCapizzano, S.: Spectral analysis and multigrid methods for finite volume approximations of spacefractional diffusion equations. SIAM J. Sci. Comput. 40, A4007–A4039 (2018)
Fischer, R., Huckle, T.: Multigrid methods for anisotropic BTTB systems. Linear Algebra Appl. 417, 314–334 (2006)
Garoni, C., SerraCapizzano, S.: Generalized Locally Toeplitz Sequences: Theory and Applications, vol. I. Springer, Cham (2017)
Gu, X.M., Huang, T.Z., Zhao, X.L., Li, H.B., Li, L.: Strangtype preconditioners for solving fractional diffusion equations by boundary value methods. J. Comput. Appl. Math. 277, 73–86 (2015)
Gu, X.M., Zhao, Y.L., Zhao, X.L., Carpentieri, B., Huang, Y.Y.: A note on parallel preconditioning for the allatonce solution of Riesz fractional diffusion equations. Numer. Math. Theor. Methods Appl. 14, 893–919 (2021)
Horton, G., Vandewalle, S.: A space–time multigrid method for parabolic partial differential equations. SIAM J. Sci. Comput. 16, 848–864 (1995)
Jiang, Y., Xu, X.: Multigrid methods for space fractional partial differential equations. J. Comput. Phys. 302, 374–392 (2015)
Lei, S.L., Sun, H.W.: A circulant preconditioner for fractional diffusion equations. J. Comput. Phys. 242, 715–725 (2013)
Liao, H., Lyu, P., Vong, S.: Secondorder BDF time approximation for Riesz spacefractional diffusion equations. Int. J. Comput. Methods 95, 144–158 (2017)
Lin, X., Ng, M.K., Sun, H.: A multigrid method for linear systems arising from timedependent twodimensional spacefractional diffusion equations. J. Comput. Phys. 336, 69–86 (2017)
Lin, X., Ng, M.K., Sun, H.: A splitting preconditioner for Toeplitzlike linear systems arising from fractional diffusion equations. SIAM J. Matrix Anal. Appl. 38, 1580–1614 (2017)
Moghaderi, H., Dehghan, M., Donatelli, M., Mazza, M.: Spectral analysis and multigrid preconditioners for twodimensional spacefractional diffusion equations. J. Comput. Phys. 350, 992–1011 (2017)
Oosterlee, C.W.: The convergence of parallel multiblock multigrid methods. Appl. Numer. Math. 19, 115–128 (1995)
Pang, H., Sun, H.W.: Multigrid method for fractional diffusion equations. J. Comput. Phys. 231, 693–703 (2012)
Ran, M., Zhang, C.: A highorder accuracy method for solving the fractional diffusion equations. J. Comp. Math. 38, 239–253 (2020)
Ruge, J.W., Stüben, K.: Algebraic multigrid. In: McCormick, S.F. (ed.) Multigrid methods, Front. Math. Appl, vol. 3. SIAM, pp. 73–130 (1987)
Tian, W., Zhou, H., Deng, W.: A class of second order difference approximations for solving spacefractional diffusion equations. Math. Comput. 84, 1703–1727 (2015)
Thomee, V.: Galerkin Finite Element Methods for Parabolic Problems (vol. 25). Springer Ser. Comput. Math. (1997)
Wu, S.L., Zhou, T.: Fast parallel iterations for fractional diffusion equations. J. Comput. Phys. 329, 210–226 (2017)
Yue, X., Shu, S., Xu, X., Bu, W., Pan, K.: Parallelintime multigrid for spacetime finite element approximations of twodimensional spacefractional diffusion equations. Comput. Math. Appl. 78, 3471–3484 (2019)
Zhao, Y.L., Zhu, P.Y., Gu, X.M., Zhao, X.L., Jian, H.Y.: A preconditioning technique for allatonce system from the nonlinear tempered fractional diffusion equation. J. Sci. Comput. 83, Article number 10 (2020)
Acknowledgements
We gratefully acknowledge the support of GNCSINDAM (Italy) and Swiss National Science Foundation SNF via the projects StressBased Methods for Variational Inequalities in Solid Mechanics (Grant No. 186407) and ExaSolvers (Grant No. 162199).
Funding
Open access funding provided by Università degli Studi dell'Insubria within the CRUICARE Agreement.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Donatelli, M., Krause, R., Mazza, M. et al. Allatonce multigrid approaches for onedimensional spacefractional diffusion equations. Calcolo 58, 45 (2021). https://doi.org/10.1007/s10092021004363
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10092021004363
Keywords
 Fractional diffusion equations
 Toeplitz matrices
 Spectral distribution
 Allatonce systems
 Space–time multigrid