Jump to ContentJump to Main Navigation
Show Summary Details
More options …

Computational Methods in Applied Mathematics

Editor-in-Chief: Carstensen, Carsten

Managing Editor: Matus, Piotr

IMPACT FACTOR 2018: 1.218
5-year IMPACT FACTOR: 1.411

CiteScore 2018: 1.42

SCImago Journal Rank (SJR) 2018: 0.947
Source Normalized Impact per Paper (SNIP) 2018: 0.939

Mathematical Citation Quotient (MCQ) 2018: 1.22

See all formats and pricing
More options …
Volume 16, Issue 3


Space-Time Discontinuous Galerkin Discretizations for Linear First-Order Hyperbolic Evolution Systems

Willy Dörfler / Stefan Findeisen / Christian Wieners
Published Online: 2016-04-07 | DOI: https://doi.org/10.1515/cmam-2016-0015


We introduce a space-time discretization for linear first-order hyperbolic evolution systems using a discontinuous Galerkin approximation in space and a Petrov–Galerkin scheme in time. We show well-posedness and convergence of the discrete system. Then we introduce an adaptive strategy based on goal-oriented dual-weighted error estimation. The full space-time linear system is solved with a parallel multilevel preconditioner. Numerical experiments for the linear transport equation and the Maxwell equation in 2D underline the efficiency of the overall adaptive solution process.

Keywords: Space-Time Methods; Discontinuous Galerkin Finite Elements; Linear Hyperbolic Systems; Transport Equation; Wave Equation; Maxwell’s Equations

MSC 2010: 65N30

1 Introduction

Space-time methods for time-dependent PDEs discretize the full problem in the space-time cylinder, and then the corresponding large algebraic system is also solved for the full problem. This is in contrast to the method of lines or Rothe’s method, which first use a discretization either in space or in time and then apply standard techniques for the other variable. Our methods are based on treating space and time simultaneously in a variational manner. Depending on the choice of the ansatz and the test spaces, the methods become either explicit or implicit. Explicit methods are computationally efficient but suffer from severe limitations for the time step size, where the length of the time edge of the space-time elements is restricted by the smallest local resolution scale in space. To circumvent these restrictions, we focus on implicit methods.

A fully implicit space-time approach allows for flexible adaptive discretizations which combine adaptivity in space with local time stepping. A further motivation for developing space-time methods is the design of modern computer facilities with an enormous number of processor cores, where the parallel realization of conventional methods becomes inefficient. Since these machines allow a fully implicit space-time approach, new parallel solution techniques are required to solve the huge linear systems, particularly for time-dependent applications in three spatial dimensions.

In recent years, discontinuous Galerkin (DG) methods in space have become very popular, see, e.g., [18] for time-dependent first-order systems, where this discretization is coupled with explicit time integration. An application of this method to acoustic and elastic waves is considered in [7] combined with an adaptive space-time hp-strategy. Here, we extend these spatial DG discretization by a Petrov–Galerkin method in time with continuous ansatz space and discontinuous test space (see, e.g., [3]). A space-time method for elastic waves with a second-order formulation in space and implicit discontinuous Galerkin time discretization is considered in [22].

An alternative discontinuous Petrov–Galerkin (DPG) approach is proposed by L. Demkowicz (see [6] for an overview and [9] for space-time applications) for general linear first-order systems, where weak approximations are constructed by introducing skeleton variables. The application of this technique to the time-harmonic case is analyzed in [32]. For acoustic and elastic waves, the hybridization in space (applied to the second-order formulation) is presented in [26], and a hybrid space-time discontinuous Galerkin method is proposed in [29]. Both methods are implicit in every time slab, and only Dirichlet traces are used for the hybrid coupling. Space-time (Trefftz) discontinuous Galerkin methods for wave problems are analyzed in [8, 23].

Error estimation for linear wave equations (considered as second-order equations) is studied in [3, 27, 17] and for more general hyperbolic systems in [21]. Simple residual error indications are not sufficient for wave problems since, in the hyperbolic case, the error is transported and thus not correlated to large local residuals. Reliable error control requires the adjoint problem, as it is introduced for goal-oriented techniques in [3], to be solved. This technique requires a variational approach, since this allows for an error representation with respect to a given linear error functional.

In principle, all parallel solution methods in space apply also to implicit time integration schemes. Parallel strategies in time are studied extensively on the basis of the ‘parareal’ idea [24, 2, 16]. A general overview over the most popular algorithms and software packages is given in [13]. Methods such as MGRID [12] and PFASST [10] were developed under the aspect that they can be easily incorporated into existing time sequential code. In addition, solution concepts specially adapted to the full space-time problem were proposed. E.g., the wavefront method extends a spatial domain decomposition into time slices, see [14] for an application to the one-dimensional wave equation. In [15] a space-time multigrid method for parabolic problems is analyzed. A multigrid method for higher order discontinuous Galerkin discretizations of advection problems is proposed in [28].

In this paper we present a fully implicit and parallel adaptive space-time discontinuous Galerkin discretization for linear first-order hyperbolic problems. The paper is structured as follows. In Section 2 we introduce a setting for linear hyperbolic operators by reference to applications in the field of linear transport and acoustic and electro-magnetic waves, and we establish the well-posedness of the space-time variational problem based on a technique developed in [31]. In Section 3, following the setting established in [20], we consider a semi-discrete discontinuous Galerkin discretization in spatial direction with upwind flux. On this basis we define an implicit Petrov–Galerkin space-time discretization in Section 4, and we prove well-posedness of the discrete method and convergence on tensor product space-time meshes. In Section 5 we propose a goal-oriented space-time error indicator based on the explicit computation of the dual solution. In Section 6 a multilevel preconditioner with semi-coarsening first in time and then in space is defined. Within the parallel finite element software system M++ [30] the adaptive method and the multilevel solution method are realized in Section 7. Moreover, the efficiency of the full scheme is demonstrated for two models, the linear transport equation and Maxwell’s equations in 2D.

2 A Space-Time Setting for Linear Hyperbolic Operators

Let ΩD be a bounded Lipschitz domain, and let HL2(Ω)J be a Hilbert space with weighted inner product (𝐯,𝐰)H=(M𝐯,𝐰)0,Ω, where ML(Ω)J×J is uniformly positive and symmetric. We consider a linear operator A:D(A)H with domain D(A)H. For given initial function 𝐮0D(A), final time T>0 and right-hand side 𝐟L2(0,T;H), we study the evolution equation


We specialize A to the case of linear balance laws determined by a flux function 𝐅(𝐯)=[B1𝐯,,BD𝐯] with symmetric matrices BjL(Ω)J×J such that


Since the matrices Bd are symmetric, any linear combination n1B1++nDBD for 𝐧=(n1,,nD)D is diagonalizable with real eigenvalues, so that (2.1) is a linear hyperbolic system [11, Section 7.3]. We discuss the following examples.

Linear Transport.

For a scalar model problem (J=1), we consider the transport equation to determine u:Ω×(0,T) such that

ρtu+div(u𝐪)=fon Ω×(0,T),u(0,)=u0,

for a given vector field 𝐪W1(Ω)D with div𝐪=0 and a density distribution ρL(Ω) satisfying ρρ0 a.e. for some ρ0>0. This defines the inflow and outflow boundary


(where 𝐧 is the outer unit normal), the flux function 𝐅(u)=u𝐪, hence Au=div(u𝐪) with domain

D(A)={uH1(Ω):u=0 on Γin},

H=L2(Ω), and Mu=ρu. For the adjoint operator A* the roles of the inflow and outflow boundary are interchanged and hence A*u=-div(u𝐪) with domain

D(A*)={u*H1(Ω):u*=0 on Γout}.

Acoustic Waves.

Acoustic waves in isotropic and homogeneous media (with density ρ1) are described by


for the pressure p:Ω×(0,T) and the velocity 𝐯:Ω×(0,T)D. We set 𝐮=(𝐯,p), H=L2(Ω)D+1 and M(𝐯,p)=(𝐯,p). The operator A is defined by A(𝐯,p)=(p,div𝐯) which corresponds to the flux function 𝐅 given by Bj=𝐞j𝐞D+1+𝐞D+1𝐞j. In the case of homogeneous Dirichlet boundary conditions, the domain is given by D(A)=D(A*)=H(div,Ω)×H01(Ω) with A*(𝐯,p)=(-p,-div𝐯).

Electro-Magnetic Waves.

For given permeability μ and permittivity ε, electro-magnetic waves are determined by the first-order system for the electric field 𝐄:Ω×(0,T)3 and magnetic field 𝐇:Ω×(0,T)3:


for the J=6 components (𝐄,𝐇). Here, we set H=L2(Ω)3×L2(Ω)3, M(𝐄,𝐇)=(ε𝐄,μ𝐇), and the operators A(𝐄,𝐇)=(-curl𝐇,curl𝐄)=-A*(𝐄,𝐇) in D(A)=D(A*)=H0(curl,Ω)×H(curl,Ω) for a perfect conducting boundary. Here, the matrices Bj are given by Bj(𝐄,𝐇)=(-𝐞j×𝐇,𝐞j×𝐄). The divergence constraints require the compatibility condition div𝐟=ρ for the right-hand side. Note that in case of polarized electro-magnetic waves this 3D setting can be reduced to a 2D setting.

The Variational Setting.

In the abstract setting, we consider the operator L=Mt+A on the space-time cylinder Q=Ω×(0,T) with the domain V=D(L), where V is the closure of {𝐯C1(0,T;D(A)):𝐯(0)=} with respect to the weighted graph norm 𝐯V2=(M𝐯,𝐯)0,Q+(M-1L𝐯,L𝐯)0,Q. The corresponding dual space V* is the closure of {𝐯*C1(0,T;D(A*)):𝐯*(T)=}. Then we define


with the weighted norm 𝐰W2=(M𝐰,𝐰)0,Q. Note that in terms of this definition, the norm in V also reads 𝐯V2=𝐯W2+M-1L𝐯W2.

In the subsequent analysis, we assume homogeneous initial and boundary conditions that are included in the domain D(L). Our considerations extend to initial values 𝐮0 by replacing 𝐟(t) with 𝐟(t)-A𝐮0 in (2.1). Also inhomogeneous boundary conditions can be analyzed by modifying the right-hand side when the existence of a sufficiently smooth extension of the boundary data can be assumed.

We define the bilinear form b:V×W with b(𝐯,𝐰)=(L𝐯,𝐰)0,Q, and we establish the standard Babuška setting (see, e.g., [5, Theorem III.3.6]).

Assume that (A𝐯,𝐯)0,Ω0 for 𝐯D(A). Then, the bilinear form b(,) is continuous and inf-sup stable in V×W with β=(4T2+1)-1/2, i.e.,



The continuity follows from the upper bound |b(𝐯,𝐰)|𝐯V𝐰W. To prove the inf-sup condition we first note that for all 𝐯C1(0,T;D(A)) with 𝐯(0)= we have


This yields 𝐯W2TM-1L𝐯W for 𝐯V. Let 𝐯V{} and take 𝐰=M-1L𝐯W{}, then


where the final inequality follows from


The inf-sup stability ensures that the operator L(V,W) is injective and that the range is closed. Thus, the operator is surjective by construction and the inverse L-1 is bounded in (W,V). This yields directly the following result [5, Theorem III.3.6].

For given 𝐟L2(Q)J there exists a unique solution 𝐮V of


satisfying the a priori bound 𝐮V4T2+1M-1/2𝐟0,Q.

3 A Semi-Discrete Discontinuous Galerkin Discretization in Space

In this section we consider the semi-discrete evolution equation


in a finite dimensional subspace HhH associated to the mesh size h of the underlying mesh defined below. The discrete operator Ah(Hh,Hh) will be constructed from a discontinuous Galerkin discretization. The discrete mass operator Mh(Hh,Hh) is the Galerkin approximation of M defined by


Note that the discrete mass operator Mh is represented by a block diagonal positive definite matrix.

We assume that Ω is a bounded polyhedral Lipschitz domain decomposed into a finite number of open elements KΩ such that Ω¯=K𝒦K¯, where 𝒦 is the set of elements in space. Let K be the set of faces of K, and for inner faces fK let Kf be the neighboring cell such that f=KKf, and let 𝐧K be the outer unit normal vector on K. The outer unit normal vector field on Ω is denoted by 𝐧.

Integration by parts of A𝐯=div𝐅(𝐯) gives for smooth ansatz functions 𝐯 and smooth test functions ϕK,


This formulation is now the basis for the discretization. We select polynomial degrees pK, and we define the local spaces Hh,K=pK(K)J and the global discontinuous Galerkin space

Hh={𝐯hL2(Ω)J:𝐯h|KHh,K for all K𝒦}.

For 𝐯hHh we define 𝐯h,K=𝐯h|KHh,K for the restriction to K.

We then define the discrete linear operator Ah(Hh,Hh) for 𝐯hHh and ϕh,KHh,K by


where 𝐧K𝐅Knum(𝐯h) is the upwind flux obtained from local solutions of Riemann problems, see [20, Section 2]. Again using integration by parts, we obtain


On inner faces f=KKf it is a consistency requirement that the difference 𝐧K(𝐅Knum(𝐯h)-𝐅(𝐯h,K)) only depends on [𝐯h]K,f=𝐯h,Kf-𝐯h,K, and that 𝐧K(𝐅Knum(𝐯)-𝐅(𝐯))=0 on all faces fK for 𝐯D(A). In particular, this yields




The upwind flux guarantees that the discrete operator is non-negative, i.e., (Ah𝐯h,𝐯h)0,Ω0 for 𝐯hHh. For the examples in Section 2, the numerical upwind flux in homogeneous media is given as follows (see [20] for the explicit solution of Riemann problems in heterogeneous media).

Linear Transport.

We have 𝐧𝐅(u)=u𝐧𝐪 and


with [uh]K,f=-2uh,K on fΓin, and [uh]K,f=0 on fΩΓin.

Acoustic Waves.

We have 𝐧𝐅(𝐯,p)=(p𝐧,𝐧𝐯) and


On Dirichlet boundary faces f=KΩ, we set [ph]K,f=2ph,K and [𝐯h]K,f𝐧K=0.

Electro-Magnetic Waves.

We have 𝐧𝐅(𝐄,𝐇)=(-𝐧×𝐇,𝐧×𝐄) and


The perfect conducting boundary conditions on f=KΩ are modeled by the (only virtual) definition of 𝐧K×𝐄h,Kf=-𝐧K×𝐄h,K and 𝐧K×𝐇h,Kf=𝐧K×𝐇h,K, i.e., 𝐧K×[𝐄]K,f=-2𝐧K×𝐄h,K and 𝐧K×[𝐇]K,f=.

4 A Petrov–Galerkin Space-Time Discretization

Let Q¯=RR¯ be a decomposition of the space-time cylinder into space-time cells R=K×I with KΩ and I=(t-,t+)(0,T); denotes the set of space-time cells. For every R we choose local ansatz and test spaces Vh,R,Wh,RL2(R)J with Wh,RtVh,R, and we define the global ansatz and test space

Vh={𝐯hH1(0,T;H):𝐯h(𝐱,0)= for a.a. 𝐱Ω and 𝐯h,R=𝐯h|RVh,R},Wh={𝐰hL2(0,T;H):𝐰h,R=𝐰h|RWh,R}.

By construction, functions in Wh are discontinuous in space and time, and functions in Vh are continuous in time, i.e., 𝐯h(𝐱,) is continuous on [0,T] for a.a. 𝐱Ω.

In addition we aim for dim(Vh)=dim(Wh), which restricts the choice of Vh,R. In the most simple case this can be achieved for a tensor product space-time discretization with a fixed mesh 𝒦 in space and a time series 0=t0<t1<<tN=T, i.e., =K𝒦n=1NK×(tn-1,tn). Then, we can select a discrete space Hh with Hh,K=p(K)J independently of t, and in every time slice we define Wh,R=Hh,K constant in time on R=K×(tn-1,tn). This yields in this case piecewise linear approximations in time,

Vh={𝐯hH1(0,T;H):𝐯h(𝐱,0)=,𝐯h(𝐱,tn)Hh for a.a. 𝐱Ω and n=1,,N, and 𝐯h(𝐱,t)=tn-ttn-tn-1𝐯h(𝐱,tn-1)+t-tn-1tn-tn-1𝐯h(𝐱,tn) for t(tn-1,tn)}.

In the general case, we select locally in space and time polynomial degrees pR and qR in R, and we set for the local test space Wh,R=(pR(K)qR-1)J. Then we define for R,


This yields 𝐯h,R(𝐱,)qRJ for 𝐯h,RVh,R and (𝐱,)R.

The discontinuous Galerkin operator in space is extended to the space-time setting defining Ah𝐯hWh by


for 𝐯hVh and 𝐰hWh. We define the discrete space-time operator Lh(Vh,Wh) and the corresponding discrete bilinear form bh(,)=(Lh,)0,Q by (Lh𝐯h,𝐰h)0,Q=(Mht𝐯h+Ah𝐯h,𝐰h)0,Q.

In order to show that a solution to our Petrov–Galerkin scheme exists, we check the inf-sup stability of the discrete bilinear form bh(,) with respect to the discrete norm


By construction, bh(,) is bounded in Vh×Wh, i.e.,


For the verification of the inf-sup stability, we introduce the L2-projection Πh:WWh which is defined by (Πh𝐯,𝐰h)0,Q=(𝐯,𝐰h)0,Q for 𝐰hWh. Then, by construction, ΠhAh=Ah and ΠhLh=Lh. Moreover, we define the non-negative weight function in time dT(t)=T-t, and we observe


Assume that


Then, the bilinear form bh(,) is inf-sup stable in Vh×Wh with β=1/1+4T2, i.e.,



Transferring the proof of Lemma 2.1 to the discrete setting yields


This yields 𝐯hW2TMh-1Lh𝐯hW and thus 𝐯hVh1+4T2Mh-1Lh𝐯hW, which implies the inf-sup stability using bh(𝐯h,𝐰h)=(Lh𝐯h,𝐰h)0,Q=(Mh-1Lh𝐯h,𝐰h)W and inserting 𝐰h=Mh-1Lh𝐯h:


Referring to Theorem 2.2, Lemma 4.1 shows the existence of a unique discrete Petrov–Galerkin solution (provided that the assumptions in Lemma 4.1 are satisfied).

For given 𝐟L2(Q)J there exists a unique solution 𝐮hVh of


satisfying the a priori bound 𝐮hVh4T2+1Mh-1Πh𝐟W.

The convergence will be analyzed with respect to the discrete norm Vh. For 𝐯V the consistency of the numerical flux in (4.1) yields (Ah𝐯,𝐰h)0,Q=(div𝐅(𝐯),𝐰h)0,Q so that Ah𝐯=Πhdiv𝐅(𝐯). This shows that Ah and thus also Vh can be evaluated in V+Vh and that bh(,) is continuous with respect to this extension.

Let 𝐮V be the solution of (2.2) and 𝐮hVh its approximation solving (4.3). Then, we have


If in addition the solution is sufficiently smooth, we obtain the a priori error estimate


for t, x and p,q1 with tt+-t-, xdiam(K), ppR and qqR for all R=K×(t-,t+).


The consistency (3.1) of the discontinuous Galerkin method yields


and thus also consistency of the Petrov–Galerkin setting, i.e.,


This gives for all 𝐯hVh and 𝐰hWh,


and thus


Now we assume that the solution is regular satisfying 𝐮Hq+1(0,T;L2(Ω)J)L2(0,T;Hp+1(Ω)J). We have by consistency Ah𝐯h=A𝐯h for all 𝐯hVhH1(Ω)J, so that the error estimate yields


where Ih:VVhH1(Ω)J is a suitable Clément-type interpolation operator. By standard assumptions on the right-hand side and the mesh regularity we obtain a bound depending on t in time and x in space. ∎

We check the assumptions of Lemma 4.1 for the special case of a tensor product discretization where the polynomial degrees in time are fixed on every time slice I=(tn-1,tn)(0,T) and the polynomial degrees in space are fixed on every KΩ. Then we have the local spaces Wh,R=pK(K)qI-1 and VR=pK(K)qI on a space-time cell R=K×I, i.e., pR=pK and qR=qI. Note that for qI1 the Petrov–Galerkin method in time is equivalent to the implicit midpoint rule, see also [3].

In the case of tensor product space-time discretizations, condition (4.2) is satisfied, i.e., for 𝐯hVh we have



Let Hh be the discontinuous Galerkin space in Ω with Hh,K=pK(K). In the tensor product case, for 𝐯hVh and 𝐰hWh representations exist in the form


with orthonormal Legendre polynomials λI,kk in L2(I) and ψI,k,ϕI,kHh. We observe


i.e., t𝐯h,RWh,R and thus Πht𝐯h=t𝐯h. Furthermore, we have


since (dTtλI,k,λI,qI)0,I=0 for k<qI and (dTtλI,qI,λI,qI)0,I=-(ttλI,qI,λI,qI)0,I=-qI (see Lemma A.1 in the appendix for a proof). From


we obtain in the tensor product case Ah=AhΠh and thus


since both matrices with entries (AhψI,k,ψI,j)Ω and (λI,k,dTλI,j)0,I, respectively, are positive semi-definite. ∎

5 Duality Based Goal-Oriented Error Estimation

In order to develop an adaptive strategy for the selection of the local polynomial degrees pR, qR we derive an error indicator with respect to a given linear goal functional EW. Following the framework in [4], we define the adjoint problem and solve the dual problem. Then, the error is estimated in terms of the local residual and the dual weight.

The adjoint operator L*=-t+A* in space and time is defined on the adjoint Hilbert space

V*={𝐰W:there exists 𝐠W such that (L𝐯,𝐰)0,Q=(𝐯,𝐠)0,Q for all 𝐯V}

and is characterized by


Note that we have 𝐰(T)= for 𝐰V*, so that the adjoint space-time problem can be solved backward in time. In case of the linear hyperbolic problems discussed in Section 2 it holds


so that we have A*=-A on VV*.

For the evaluation of the error functional E we introduce the dual solution 𝐮*V* with


Let 𝐮V be the solution of (2.2), and 𝐮hVh its approximation solving (4.3). Now we derive an exact error representation for the error functional in the case that the dual solution is sufficiently smooth such that 𝐮*(,t)|fL2(f)J for all faces fh and a.a. t(0,T). Inserting the consistency of the numerical flux (3.2) and using (5.1) yields for all 𝐰hWh,


From this error representation, inserting some projection 𝐰h=Πh𝐮*, we obtain the estimate


However, this bound cannot be used since it depends on the unknown function 𝐮*. In applications, the following heuristic error bound is used instead. Let 𝐮h*Wh be a numerical approximation of the dual solution given by


(using the transposed finite element matrix). Then we replace the projection error 𝐮*-Πh𝐮* by Ih𝐮h*-𝐮h*, where Ih is a higher-order recovery operator (or a lower order interpolation operator). Then, the right-hand side of the error bound (5.2) is replaced by RηR with


These terms contain the given data functions 𝐟 and M and are computed by a quadrature formula. Alternatively a term 𝐟-𝐟h-(M-Mh)t𝐮h0,R could be separated to control this data error. Usually, this error contribution is of minor importance. This is especially the case in our numerical examples.

The error indicator construction extends to nonlinear goal functionals EC2(W). Then, the dual solution 𝐮*V* depends on the primal solution, i.e.,


The estimate (5.2) applies also to |E(𝐮)-E(𝐮h)|, since we have [19]


and the second term is quadratic in 𝐮-𝐮h0,Q and will thus be neglected. In our numerical examples E′′ is constant.

6 Space-Time Multilevel Preconditioner

In this section we address the numerical aspects in particular solution methods for the discrete hyperbolic space-time problem. First we describe the realization of our discretization using nodal basis functions in space and time, and then a multilevel preconditioner is introduced, and it is tested for different settings to derive a suitable solution strategy.

Nodal Discretization.

Here we consider the case of a tensor product space-time mesh =n=1Nn with time slices n=K𝒦K×(tn-1,tn) and variable polynomial degrees pR,qR in every space-time cell R. Let {ψR,jn}j=1,,dimWh,R be a basis of Wh,R and define


Then, 𝐯hVh is represented by

𝐯h(𝐱,t)=tn-ttn-tn-1𝐰hn-1(𝐱,tn-1)+t-tn-1tn-tn-1𝐰hn(𝐱,t)for (𝐱,t)K×(tn-1,tn)

with 𝐰h0= and 𝐰hnWhn for n=1,,N. The corresponding coefficient vector is denoted by v¯=(v¯1,,v¯N), where v¯ndimWhn is the coefficient vector of


With respect to this basis, the discrete space-time system (2.2) has the matrix representation L¯u¯=f¯ with the block matrix


with matrix entries


and the right-hand side f¯=(f¯1,,u¯N) with f¯j,Rn=(𝐟,ψR,jn)0,R. Sequentially, this system can be solved by a block-Gauss–Seidel method (corresponding to implicit time integration),


provided that D¯n can be inverted efficiently. In parallel, this requires a distribution only in space (see Figure 1). Here, we discuss parallel multilevel preconditioners with a distribution of the full space-time mesh, cf. Figure 2.

Spatial distribution of mesh cells to 4 processes and required communication
Figure 1

Spatial distribution of mesh cells to 4 processes and required communication (arrows).

Space-time distribution of mesh cells to 16 processes and required
communication (arrows).
Figure 2

Space-time distribution of mesh cells to 16 processes and required communication (arrows).

Multilevel Methods.

For space-time multilevel preconditioners we consider hierarchies in space and time. Therefore, let 0,0 be the coarse space-time mesh, and let l,k be the discretization obtained by l=1,,lmax uniform refinements in space and k=1,,kmax refinements in time. Let Vl,k be the approximation spaces on l,k with fixed polynomial degrees pRp and qRq. Let L¯l,k be the corresponding matrix representations of the discrete operator Lh in Vl,k.

The multilevel preconditioner combines smoothing operations on different levels and requires transfer matrices between the levels. Since the spaces are nested, we can define prolongation matrices P¯l-1,kl,k and P¯l,k-1l,k representing the natural injections Vl-1,kVl,k in space and Vl,k-1Vl,k in time. Correspondingly, the restriction matrices R¯l-1,kl,k and R¯l,k-1l,k represent the L2-projections of the test spaces Wl,kWl-1,k and Wl,kWl,k-1.

For the smoothing operations on level (l,k) we consider the block-Jacobi preconditioner or the block-Gauss–Seidel preconditioner (where all components corresponding to a space-time cell R build blocks)


with damping parameter θl,k. The corresponding iteration matrices are given by


and the number of pre- and postsmoothing steps is denoted by νl,kpre and νl,kpost, respectively.

Now, the multilevel preconditioner B¯l,kML is defined recursively. On the coarse level, we use a parallel direct linear solver B¯0,0ML=(L¯0,0)-1. Then, we have two options: restricting in time defines B¯l,kML by


with Jacobi smoothing (cf. Figure 3), and restricting in space yields


with Gauss–Seidel smoothing, cf. Figure 4 for an illustration of the two options and Algorithm 1 for the recursive realization of the multilevel preconditioner.

Two level in time coarsening strategy.
Figure 3

Two level in time coarsening strategy.

Two level in space coarsening strategy.
Figure 4

Two level in space coarsening strategy.

(Multilevel preconditioner c¯l,k=B¯l,kMLr¯l,k with smoother B¯l,kSM=B¯l,kJ or B¯l,kGS)

The different multilevel strategies are tested for the linear transport equation with fixed polynomial degrees (p,q)=(2,2). We consider a divergence-free vector field 𝐪(𝐱)=2π(-𝐱2,𝐱1) on Ω=(-10,10)2 with homogeneous right-hand side f=0, constant density ρ1, final time T=1, and starting with a 2D Gaussian pulse u0(𝐱)=exp(-1.4((𝐱1-5)2+𝐱22)).

Several tests indicate that a block-Jacobi smoother with νl,k=2 smoothing steps and damping parameter θJ=0.5 in time, and a block-Gauss–Seidel smoother with νh=5 pre- and postsmoothing steps and no damping (θGS=1) in space is a suitable choice. The contraction number of the two-level method on different space-time meshes l,k is estimated by the averaged convergence rate of the preconditioned linear iteration


see Table 1.

Table 1

Degrees of freedom of the transport example in the space-time domain on different space-time levels (starting with 128=16×8 space-time cells in 0,0), and iteration steps and averaged rates for a residual reduction by the factor 10-8 ofthe linear iteration with two-level multilevel preconditioners in time or space.

One observes that coarsening in time leads to stable multilevel behavior (the number of iteration steps is bounded by a constant) as long as k-l2, i.e., the ratio between t and x is bounded. For coarsening in space, we observe that the iteration steps are independent of the time level k, but not bounded in l. At least the increase is small enough to achieve a benefit by using a multilevel method. For higher spatial dimensions D>1 coarsening in space is cheaper than coarsening in time, since refining in time doubles the effort whereas refining in space increases the effort by a factor 2D>2.

This and the previous observations motivate a strategy for the space-time multilevel solver, where we at first only coarse in space until the lowest spatial level is reached. Afterwards we coarse in time up to a lowest temporal level where t/x is still small enough. The full multilevel V-cycle is illustrated in Figure 5.

Full space-time coarsening strategy: coarsening in space up to level (0,k)${(0,k)}$,
then coarsening in time up to level (0,0)${(0,0)}$; finally, solve exact on the
coarsest level.
Figure 5

Full space-time coarsening strategy: coarsening in space up to level (0,k), then coarsening in time up to level (0,0); finally, solve exact on the coarsest level.

Table 2

Iteration steps and averaged rates for a full space-time multilevel method for the transport problem.Smoother: Jacobi (νk,l=2, θl,k=0.5) in time, Gauss–Seidel (νl,k=5) in space.

The results for this strategy applied to the test problem are given in Table 2. Due to the problems, observed for the two-level in space strategy, we achieve a moderate growth of iteration steps, when refining in space. We observe the same behavior for a 2D Maxwell test problem in Q=(0,1)2×(0,1), where the initial and boundary conditions are given by


see Table 3.

Table 3

Degrees of freedom for the Maxwell example in the space-time domain on different space-time meshes (starting with 64=8×8 space-time cells in 0,0), and iteration steps and averaged rates for a full space-time multilevel method for the Maxwell example with Jacobi smoothing in time (νl,k=2, θl,k=0.5) and Gauss–Seidel smoothing in space (νl,k=5).

In the adaptive case a coarse cell may correspond to a set of fine space-time cells of different polynomial degrees. To set up a polynomial distribution on the subspaces Vl-1,k or Vl,k-1 (and correspondingly on Wl-1,k or Wl,k-1) which does not impair the convergence rate, we apply the following strategy. For every coarse cell we use the highest polynomial degree in space and time on the subset of fine cells. Hence, we interpolate all solutions on the fine cells to this highest polynomial degree and use the restriction or prolongation matrices of the uniformly refined case. For the adaptive computations in the next section we observe the same (cf. Tables 4 and 5) or slightly better (cf. Tables 6 and 7) convergence behavior of the multilevel preconditioner.

Table 4

Results for the transport equation with uniform mesh with 524 288=4 096×128space-time cells and different polynomial degrees.

Table 5

Adaptive refinement on a mesh with 524 288=4 096×128 space-time cells (ϑ=1e-4).

Table 6

Uniform vs. adaptive refinement on 606 208=2 368×256 space-time cellsdistributed to 256 processes (ϑ=1e-3, Eex=4.0234e-1).

Table 7

Uniform vs. adaptive refinement on 4 849 664=9 472×512 space-time cells distributedto 1024 processes (ϑ=1e-3, Eex=4.0257e-1). On level 3 and 4 the effort is estimated, since wewere not able to compute the reference values with uniform refinement.

7 Numerical Tests for Space-Time Adaptivity

Finally, we present results for the full adaptive method. We test the convergence properties for two examples, the linear transport equation for a configuration with known solution, which serves as a test problem to verify our methods, and a more sophisticated configuration for electro-magnetic waves in two spatial dimensions which is closer to practical applications. Here we use a generalized minimal residual solver (GMRES) equipped with the multilevel preconditioner from Section 6 and a residual reduction of 10-8 as stopping criterion. The adaptive strategy is described in Algorithm 2 depending on a parameter ϑ<1 for the adaptive selection criterion.

(Adaptive algorithm)

Linear Transport.

In the following numerical example we investigate the performance and reliability of our p-adaptive algorithm in comparison with uniform refinement for the example on the previous section. Since the characteristics for the transport vector 𝐪(𝐱)=2π(-𝐱2,𝐱1) are circles, we find u(𝐱,1)=u0(𝐱). We start with an initial coarse mesh with 1024=64×16 space-time cells which is refined 3 times in space and time to 524 288 cells. The coarse problem is solved by using a parallel direct solver [25]. Furthermore we use low-order polynomial degrees (p,q)=(1,1) as initial distribution on Q. In this test we aim to minimize the error E=|E(u)-E(uh)| towards the quadratic energy functional


using the dual error indicator derived in Section 5. Hence the adaptive strategy minimizes the energy error in Q.

The exact solution of the dual problem

-ρtu*-div(u*𝐪)=ρuon Ω×(0,T),u*(T)=0,

with homogeneous Dirichlet boundary conditions is given by u*(𝐱,t)=(T-t)u(𝐱,t), since for all wV,


Thus, u* also corresponds to a Gaussian pulse (traveling backwards in time).

Figure 6 shows the adaptive solution in the space time domain. In comparison with Figure 7 we see that highest polynomial degrees are only used in areas where the pulse is actually located, whereas lowest polynomial degrees are used everywhere else. The adaptive results are given in Table 5 and Figure 8. First we observe that the estimation for the dual error eh*=uh*-Ihuh* approximates the exact dual error e*=u*-uh* well. Using the solutions u and uh, the exact errors E=|E(u)-E(uh)| and uh(T)-u(T)0,Ω can be computed. Furthermore E can be estimated using (5.2) with approximations eh* and uh. These results, denoted as Eh, almost coincide with E. Finally, the sum over all cell-wise estimated errors η=RηR computed by (5.2) shows the same asymptotic behavior, which is required for reliable error estimation.

Solution of the transport equation in the space-time domain Q, sliced at times t=0,0.3,0.6,1${t=0,0.3,0.6,1}$.
Figure 6

Solution of the transport equation in the space-time domain Q, sliced at times t=0,0.3,0.6,1.

Location of the highest polynomial degrees in the space-time domain Q.
Figure 7

Location of the highest polynomial degrees in the space-time domain Q.

The benefit of adaptive strategies becomes clear in Figure 9, where we compare the adaptive solution with a uniformly refined solution (see Table 4). On the last refinement level we achieve the same errors E and uh(T)-u(T)0,Ω by using only approximately 3.3 million degrees of freedom. This corresponds to a reduction of about 90% compared to the uniformly refined case. The benefit depends on the underlying problem. But if the solution is strongly located (e.g., a Gaussian pulse or a single wavefront) or one is only interested in small parts of the solution (as in our next example), it is possible to save a large amount of computational resources.

Error estimation.
Figure 8

Error estimation.

Adaptive performance.
Figure 9

Adaptive performance.

Electro-Magnetic Waves.

We consider a 2D transverse electric wave 𝐮=(𝐇1,𝐇2,𝐄3) with wavelength λ=1. It is scattered by a double slit with slit gap a=3 and slit width b=1. The scattered wave enters the computational domain Q=(0,6)×(-6,6)×(0,8) on the left (see Figure 10). Furthermore we apply constant material parameters μ=ε=1 and reflecting boundary conditions. Behind the double slit one observes a diffraction pattern with several local intensity extrema. In applications one is often only interested in certain small parts of the scattered wave. Therefore we choose the region of interest as S=(5.5,6)×(0,2)×(0,8) to resolve the first minimum at the right end of our computational domain as good as possible. Hence the energy error functional is given as


This corresponds to a screen or receiver somewhere in S to receive and measure the scattered wave as illustrated in Figure 10. Since in this setting the exact value of E(𝐮) is not known, we approximate E by extrapolation:


We perform two tests on two different levels with 256 and 1024 processes, respectively. In the first case the initial coarse mesh consists of 9 472=148×64 space-time cells and is refined 2 times in space and time up to 606 208 cells. We use 256 processes to compute a uniform and an adaptive refined solution. Similarly to the first example we observe from Table 6 that the estimated value of the error functional coincidences in both cases and we are able save about 73% of the degrees of freedom.

Schematic illustration of the double-slit experiment. Slit gap a=3${a=3}$ and
slit width b=1${b=1}$.
Figure 10

Schematic illustration of the double-slit experiment. Slit gap a=3 and slit width b=1.

Scattered wave solution (left) and used polynomial degrees (right) at
different times. Solved on 1024 processes.
Figure 11

Scattered wave solution (left) and used polynomial degrees (right) at different times. Solved on 1024 processes.

In the second case we use 1024 processes for a problem that is refined once more in space and time up to 4 849 664 cells. To be able do some reasonable load balancing according to the degrees of freedom on each cell, we have to refine the coarse mesh too (i.e., 592×128 cells). For all tests up 109 space-time degrees of freedom, we see from Table 7 that we save about 79% of the degrees of freedom and hence are still able to compute an accurate adaptive solution (with respect to E). In both cases we used a Gauss–Seidel preconditioned GMRES solver for the coarse problem. Figure 11 shows the time evolution of the scattered wave computed on 1024 processes. The diffraction pattern and the result of the adaptive error estimation are clearly visible. The adaptive solution uses highest polynomial degrees in areas where it is necessary to have a high resolution in S and lowest polynomial everywhere else.

All numerical results where computed with 256 or 1024 processes on the ForHLR cluster at KIT, where a node contains two Intel Xeon E5-2670 v2 (2.5 GHz, 10 cores) and 64 GB memory.

8 Conclusion

We have demonstrated for the linear transport equation and for polarized waves in 2D that discontinuous Galerkin methods in space combined with a Petrov–Galerkin discretization in time yield a stable scheme. The numerical results confirm that a dual weighted error estimator together with a space-time multigrid strategy is efficient. It remains an open question to provide convergence estimates for the adaptive scheme and to derive bounds for the condition number of the multigrid preconditioner. Moreover, the extension to 3D simulation will be a challenge for the next generation of massive parallel machines.

A Appendix

Let λI,kk be the orthonormal Legendre polynomials with respect to the inner product in L2(I) in the interval I=(tn-1,tn).

We have (ttλI,k,λI,k)0,I=k for k0.


We prove the result for the orthonormal Legendre polynomials λkk in L2(-1,1); then, the general case follows directly from the relation


Starting with λ-10 and λ01/2, we obtain recursively


see [1, Lemma 8.5.3]. We have tλ00. For k0 from (k+1)λk+1(t)=(2k+1)tλk(t)-kλk-1(t) we obtain


Subtracting ktλk+1(t)=(2k+1)ttλk(t)-(k+1)tλk-1(t) results in tλk+1(t)=(2k+1)λk(t)+tλk-1(t). This yields the assertion by



  • [1]

    Abramowitz M. and Stegun I. A., Handbook of Mathematical Functions: With Formulas, Graphs, and Mathematical Tables, Appl. Math. Ser., Dover Publications, New York, 1964.  Google Scholar

  • [2]

    Ascher U. M., Ruuth S. J. and Wetton B. T. R., Implicit-explicit methods for time-dependent partial differential equations, SIAM J. Numer. Anal. 32 (1995), no. 3, 797–823.  Google Scholar

  • [3]

    Bangerth W. and Rannacher R., Finite element approximation of the acoustic wave equation: Error control and mesh adaptation, East-West J. Numer. Math. 7 (1999), no. 4, 263–282.  Google Scholar

  • [4]

    Bangerth W. and Rannacher R., Adaptive Finite Element Methods for Differential Equations, Birkhäuser, Basel, 2003.  Google Scholar

  • [5]

    Braess D., Finite Elements. Theory, Fast Solvers, and Applications in Solid Mechanics, 3rd ed., Cambridge University Press, Cambridge, 2007.  Google Scholar

  • [6]

    Demkowicz L. F. and Gopalakrishnan J., An overview of the discontinuous Petrov–Galerkin method, Recent Developments in Discontinuous Galerkin Finite Element Methods for Partial Differential Equations, IMA Vol. Math. Appl., Springer, Cham (2014), 149–180.  Google Scholar

  • [7]

    Dumbser M., Käser M. and Toro E. F., An arbitrary high-order discontinuous Galerkin method for elastic waves on unstructured meshes – V. Local time stepping and p-adaptivity, Geophys. J. Int. 171 (2007), 695–717.  Web of ScienceGoogle Scholar

  • [8]

    Egger H., Kretzschmar F., Schnepp S. M. and Weiland T., A space-time discontinuous Galerkin–Trefftz method for time dependent Maxwell’s equations, SIAM J. Sci. Comput. 37 (2015), no. 5, 689–711.  Web of ScienceGoogle Scholar

  • [9]

    Ellis T. E., Demkowicz L. F., Chan J. L. and Moser R. D., Space-time DPG: Designing a method for massively parallel CFD, ICES report 14-32, Institute for Computational Engineering and Sciences, The University of Texas at Austin, Austin, 2014.  Google Scholar

  • [10]

    Emmett M. and Minion M. L., Toward an efficient parallel in time method for partial differential equations, Commun. Appl. Math. Comput. Sci. 7 (2012), 105–132.  Google Scholar

  • [11]

    Evans L. C., Partial Differential Equations, 2nd ed., American Mathematical Society, Providence, 2010.  Google Scholar

  • [12]

    Falgout R. D., Friedhoff S., Kolev T. V., MacLachlan S. P. and Schroder J. B., Parallel time integration with multigrid, SIAM J. Sci. Comput. 36 (2014), no. 6, 635–661.  Google Scholar

  • [13]

    Gander M. J., 50 years of time parallel time integration, Multiple Shooting and Time Domain Decomposition, Contrib. Math. Comput. Sci., Springer, Basel (2015), 69–113.  Google Scholar

  • [14]

    Gander M. J., Halpern L. and Nataf F., Optimal Schwarz waveform relaxation for the one dimensional wave equation, SIAM J. Numer. Anal. 41 (2003), no. 5, 1643–1681.  Google Scholar

  • [15]

    Gander M. J. and Neumüller M., Analysis of a new space-time parallel multigrid algorithm for parabolic problems, preprint 2014, http://arxiv.org/abs/1411.0519.  

  • [16]

    Gander M. J. and Vandewalle S., Analysis of the parareal time-parallel time-integration method, SIAM J. Sci. Comput. 29 (2007), no. 2, 556–578.  Web of ScienceGoogle Scholar

  • [17]

    Grote M. J. and Schötzau D., Optimal error estimates for the fully discrete interior penalty DG method for the wave equation, J. Sci. Comput. 40 (2009), no. 1–3, 257–272.  Web of ScienceGoogle Scholar

  • [18]

    Hesthaven J. S. and Warburton T., Nodal Discontinuous Galerkin Methods, Springer, New York, 2008.  Google Scholar

  • [19]

    Heuveline V. and Rannacher R., Duality-based adaptivity in the hp-finite element method, J. Numer. Math. 11 (2003), 95–113.  Google Scholar

  • [20]

    Hochbruck M., Pazur T., Schulz A., Thawinan E. and Wieners C., Efficient time integration for discontinuous Galerkin approximations of linear wave equations, ZAMM Z. Angew. Math. Mech. 95 (2015), no. 3, 237–259.  Google Scholar

  • [21]

    Houston P. and Süli E., hp-adaptive discontinuous Galerkin finite element methods for first-order hyperbolic problems, SIAM J. Sci. Comput. 23 (2006), no. 4, 1226–1252.  Google Scholar

  • [22]

    Köcher U. and Bause M., Variational space-time methods for the wave equation, J. Sci. Comput. 61 (2014), no. 2, 424–453.  Web of ScienceGoogle Scholar

  • [23]

    Kretzschmar F., Moiola A., Perugia I. and Schnepp S. M., A priori error analysis of space-time Trefftz discontinuous Galerkin methods for wave problems, IMA J. Numer. Anal. (2015), 10.1093/imanum/drv064.  Google Scholar

  • [24]

    Lions J.-L., Maday Y. and Turinici G., A parareal in time discretization of PDE’s, C. R. Acad. Sci. Paris Ser. I 332 (2001), no. 7, 661–668.  Google Scholar

  • [25]

    Maurer D. and Wieners C., A parallel block LU decomposition method for distributed finite element matrices, Parallel Comput. 37 (2011), no. 12, 742–758.  Google Scholar

  • [26]

    Nguyen N. C., Peraire J. and Cockburn B., High-order implicit hybridizable discontinuous Galerkin methods for acoustics and elastodynamics, J. Comput. Phys. 230 (2011), no. 10, 3695–3718.  Web of ScienceGoogle Scholar

  • [27]

    Oden J. T., Prudhomme S. and Demkowicz L., A posteriori error estimation for acoustic wave propagation problems, Arch. Comput. Methods Eng. 12 (2005), no. 4, 343–389.  Google Scholar

  • [28]

    van der Vegt J. J. W. and Rhebergen S., hp-multigrid as smoother algorithm for higher order discontinuous Galerkin discretizations of advection dominated flows. Part I: Multilevel analysis, J. Comput. Phys. 231 (2012), no. 22, 7537–7563.  Web of ScienceGoogle Scholar

  • [29]

    Wang D., Tezaur R. and Farhat C., A hybrid discontinuous in space and time Galerkin method for wave propagation problems, Internat. J. Numer. Methods Engrg. 99 (2014), no. 4, 263–289.  Web of ScienceGoogle Scholar

  • [30]

    Wieners C., A geometric data structure for parallel finite elements and the application to multigrid methods with block smoothing, Comput. Vis. Sci. 13 (2010), 161–175.  Google Scholar

  • [31]

    Wieners C. and Wohlmuth B., Robust operator estimates and the application to substructuring methods for first-order systems, ESAIM Math. Model. Numer. Anal. 48 (2014), 161–175.  Web of ScienceGoogle Scholar

  • [32]

    Zitelli J., Muga I., Demkowicz L., Gopalakrishnan J., Pardo D. and Calo V. M., A class of discontinuous Petrov–Galerkin methods. Part IV: The optimal test norm and time-harmonic wave propagation in 1D, J. Comput. Phys. 230 (2011), no. 7, 2406–2432.  Web of ScienceGoogle Scholar

About the article

Received: 2015-12-22

Revised: 2016-03-16

Accepted: 2016-03-18

Published Online: 2016-04-07

Published in Print: 2016-07-01

Funding Source: Deutsche Forschungsgemeinschaft

Award identifier / Grant number: RTG 1294

Award identifier / Grant number: CRC 1173

We gratefully acknowledge financial support by the Deutsche Forschungsgemeinschaft (DFG) through RTG 1294 and CRC 1173 and by the Sino-German Science Center (grant id 1228) on the occasion of the Chinese-German Workshop on Computational and Applied Mathematics in Augsburg 2015.

Citation Information: Computational Methods in Applied Mathematics, Volume 16, Issue 3, Pages 409–428, ISSN (Online) 1609-9389, ISSN (Print) 1609-4840, DOI: https://doi.org/10.1515/cmam-2016-0015.

Export Citation

© 2016 by De Gruyter.Get Permission

Citing Articles

Here you can find all Crossref-listed publications in which this article is cited. If you would like to receive automatic email messages as soon as this article is cited in other publications, simply activate the “Citation Alert” on the top of this page.

Ulrich Langer, Svetlana Matculevich, and Sergey Repin
Computers & Mathematics with Applications, 2019, Volume 78, Number 8, Page 2641
Sriram Nagaraj, Jacob Grosek, Socratis Petrides, Leszek F. Demkowicz, and Jaime Mora
Journal of Computational Physics: X, 2019, Volume 2, Page 100002
Julia Brunken, Kathrin Smetana, and Karsten Urban
SIAM Journal on Scientific Computing, 2019, Volume 41, Number 1, Page A592

Comments (0)

Please log in or register to comment.
Log in