An hp-Analysis of the Loal Disontinuous Galerkin Method for Diusion Problems Ilaria Perugia Dominik Shotzauy Journal of Sienti Computing, Vol. 17, pp. 561{571, 2002 Abstrat We present an hp-error analysis of the loal disontinuous Galerkin method for diusion problems, onsidering unstrutured meshes with hanging nodes and two- and three-dimensional domains. Our estimates are optimal in the meshsize h and slightly suboptimal in the polynomial approximation order p. Optimality in p is ahieved for mathing grids and polynomial boundary onditions. Key words: hp-FEM, loal disontinuous Galerkin methods 1 Introdution The aim of this paper is to present an hp-error analysis of the loal disontinuous Galerkin (LDG) method, introdued by Cokburn and Shu (1998), for the diusion problem r ( ru) = f in u = gD on ; (1) where is a bounded polygonal or polyhedral domain in Rd , d = 2; 3, and 2 L1 ( )dd a symmetri, uniformly positive denite diusion tensor. Department of Mathematis, University of Pavia, Via Ferrata 1, 27100 Pavia, Italy. E-mail: perugiadimat.unipv.it. Supported in part by the NSF (DMS-9807491) and by the University of Minnesota Superomputing Institute. This work was arried out when the author was visiting the Shool of Mathematis, University of Minnesota. y Department of Mathematis, University of Basel, Rheinsprung 21, 4051 Basel, Switzerland. E-mail: shotzaumath.unibas.h. Supported in part by NSF grant DMS0107609 and by the University of Minnesota Superomputing Institute. This work was arried out when the author was aÆliated with the Shool of Mathematis, University of Minnesota. 1 An hp-Analysis of the LDG Method for Diusion Problems 2 The1 right-hand side f belongs to L2 ( ), and the Dirihlet datum gD to H 2 ( ). Deriving error bounds that take into aount both the elemental meshsize and approximation order ompletes previous work by Castillo, Cokburn, Perugia and Shotzau (2000) on the LDG method applied to purely ellipti problems. This is relevant sine the LDG method, being based on disontinuous spaes, is ideally suited for hp-adaptivity. The analysis in this paper follows the same lines as the one developed in Perugia and Shotzau (2001) in the more ompliated situation of the low-frequeny time-harmoni Maxwell equations, and uses the general framework introdued in Arnold, Brezzi, Cokburn and Marini (2001). On meshes without hanging nodes and for polynomial boundary onditions, our setting immediately gives optimal hp-onvergene rates. Furthermore, on unstrutured meshes with hanging nodes, we prove error estimates that are optimal in the mesh-size h and slightly suboptimal in the polynomial degree p (half a power of p is lost). We point out that, for multi-dimensional ellipti problems on general unstrutured grids, no better p-bounds an be found in the literature (see, e.g., Riviere, Wheeler and Girault (1999), Prudhomme, Pasal, Oden and Romkes (2000), Houston, Shwab and Suli (2000), where the same bounds have been obtained for dierent disontinuous Galerkin methods and with dierent analysis tehniques). Optimal hpbounds have also been proved in Castillo, Cokburn, Shotzau and Shwab (2000) for one-dimensional onvetion-diusion problems and, reently, in Georgoulis and Suli (2001) for two-dimensional reation-diusion problems on aÆne quadrilateral meshes ontaining hanging nodes. The outline of the paper is as follows. In setion 2, we dene the LDG method for the diusion problem (1). We arry out the error analysis in setion 3, obtaining hp-error estimates in a problem-related energy norm, as well as in the L2 -norm. We end our presentation in setion 4 with onluding remarks. 2 LDG Disretization 2.1 Meshes and Finite Element Spaes We onsider shape-regular meshes Th that partition the domain into triangles and/or parallelograms, if d = 2, and tetrahedra and/or parallelepipeds, if d = 3, with possible hanging nodes and aligned with the possible disontinuities of the diusion tensor , so that is smooth within eah element of Th . We denote by hK the diameter of the element K 2 Th . We dene the (d 1)-dimensional faes of Th as follows. An interior fae of Th is the (non-empty) interior of K + \ K , where K + and K are two adjaent elements of Th , not neessarily mathing. Similarly, a boundary An hp-Analysis of the LDG Method for Diusion Problems 3 fae of Th is the (non-empty) interior of K \ , where K is a boundary element of Th . We denote by EI the union of all interior faes of Th , by ED the union of all boundary faes, and set E = EI [ ED . Let p = fpK gK 2Th be a degree vetor that assigns to eah element K 2 Th a polynomial approximation order pK 1. The generi hp-nite element spae of pieewise polynomials is given by S p (Th ) := fu 2 L2 ( ) : ujK 2 S pK (K ); 8K 2 Th g, where S pK (K ) is the spae P pK (K ) of polynomials of degree at most pK in K , if K is a triangle or a tetrahedron, and the spae QpK (K ) of polynomials of degree at most pK in eah variable in K , if K is a quadrilateral or a parallelepiped. 2.2 The Flux Formulation of the LDG Method By introduing the auxiliary variables q = s and s = ru as in Cokburn and Dawson (2000), the diusion problem (1) an be rewritten as q = s in r q = f in s = ru in u = gD on : We approximate the variables (q,s,u) by disrete funtions (q h ; sh ; uh) in the hp-nite element spae Qh Qh Vh , where Qh = S p (Th )d and Vh = S p (Th ), for a given degree distribution p. The LDG method then onsists in nding (qh ; sh ; uh) 2 Qh Qh Vh suh that for any (r; t; v) 2 Qh Qh Vh and for any element K 2 Th Z ZK Z K K q h r dx = sh t d x + q h rv d x Z Z K K Z sh r d x uh r t dx K Z K qbh nK v ds = ubh t nK ds = 0 Z K (2) f v dx: Here, nK denotes the outward unit normal vetor to K . The quantities ubh and qbh are the so-alled numerial uxes, whih are approximations to the traes of u and q on K , and are dened as follows. Consider an interior fae e shared by two elements K + and K . Denoting by v and r the traes on K of funtions v and r that are smooth in K , we dene the averages and jumps of v and r aross e by ffvgg = (v+ + v )=2 ffrgg = (r+ + r )=2 [ v ℄ = v + n K + + v nK [ r ℄ = r + nK + + r nK : If v 2 H 1 ( ), then [ v℄ = 0 on EI . Similarly, if r 2 H (div; ), then [ r℄ = 0 on EI . In partiular, for the exat solution u, we have [ ru℄ = 0 on EI . 4 An hp-Analysis of the LDG Method for Diusion Problems The numerial uxes are dened by ( ( f f ugg + b [ u℄ if e EI ffqgg a[ u℄ b[ q℄ if e EI b ubje = q j e = gD if e ED q a(u gD )n if e ED ; with parameters a and b to be properly hosen. This ompletes the denition of the LDG method. Notie that the ux in u is independent of q. This allows for an element-by-element elimination of the auxiliary variables q and s, giving rise to the so-alled primal formulation of the method in the variable u only. This loal solvability gives the name to the LDG method. We refer to Castillo (2001) for a disussion of this elimination proess from a omputational point of view. Let us also point out that the LDG method dened above is onsistent and, if a is stritly positive, denes a unique disrete solution (qh ; sh ; v) 2 Qh Qh Vh ; see Castillo et al. (2000). 2.3 The Primal Formulation We develop the error analysis of the LDG method in the framework introdued in Arnold et al. (2001), that is, by onsidering its primal formulation. For v belonging to V (h) := Vh + H 1 ( ), we dene L(v) 2 Qh by Z L(v) r dx = Z EI (ffrgg b [ r℄ ) [ v℄ ds + Z ED v r n ds 8r 2 Qh : Similarly, we dene the lifting G D 2 Qh of the boundary datum gD by Z G D r dx = Z ED gD r n ds 8r 2 Qh : Notie that, for the exat solution u, we have L(u) = G D . Adding the rst and seond equations in (2) over all elements and simple alulations yield (3) qh = rhuh L(uh) + GD ; 2 with denoting the L -projetion onto Qh and rh the elementwise gradient. Inserting this expression in the third equation of (2) yields the primal form of the LDG method: nd uh 2 Vh suh that Ah (uh ; v) + Ih (uh ; v) = Fh (v) 8v 2 Vh ; (4) where Z Ah (u; v) = rh u L(u) rh v L(v) dx Ih (u; v) = Fh (v) = Z E Z I a [ u℄ [ v ℄ ds + f v dx + Z Z a u v ds ED Z a gD v ds GD rh v L(v) dx: An hp-Analysis of the LDG Method for Diusion Problems 5 For disrete trial and test funtions, the primal form (4), together with identity (3), is equivalent to the original ux form (2) of the LDG method. However, unlike (2), the formulation (4) is no longer onsistent. Nevertheless, the form Ah + Ih has the ontinuity and oerivity properties that allow us to arry out an error analysis in a straightforward way by using Strang's lemma. 3 Error Analysis In this setion, we develop the hp-error analysis of the LDG method. We start by speifying the parameters a and b in the denition of the method. Then we prove abstrat error estimates in a broken norm and derive our atual error bounds. Finally, we address the issue of the stability of the LDG formulation. 3.1 The Disontinuity Stabilization Parameter We introdue the funtions h and p in L1 (E ) related to the loal meshsize and approximation degree as ( fhK ; hK 0 g if x in the interior of K \ K 0 x) := hmin if x in the interior of K \ K ( maxfpK ; pK 0 g if x in the interior of K \ K 0 p = p(x) := pK if x in the interior of K \ : h = h( Regarding the diusivity, we assume to be Lipshitz ontinuous in K , for any K 2 Th . This implies that jK an be extended up to K , and we denote this extension by K . Therefore, for any K 2 Th , there are positive onstants nK and NK suh that nK i (K (x)) NK for all x 2 K , where i (K (x)), i = 1; 2; 3, are the eigenvalues of K (x). For any K 2 Th , the onstants nK and NK are assumed to satisfy NK n K ; 8K 2 Th ; with a onstant > 0. Whenever is a pieewise onstant salar funtion, this assumption holds true with = 1. We set ( fjK (x)j; jK 0 (x)jg if x is in the interior of K \ K 0 x) := jmax K (x)j if x is in the interior of K \ ; where j (x)j is the spetral norm of the tensor (x). n = n( 6 An hp-Analysis of the LDG Method for Diusion Problems We dene the disontinuity stabilization parameter a 2 L1 (E ) in terms of h, p and n by p2 n ; a= h with > 0 independent of meshsize, approximation order and diusion. Moreover, the parameter b is taken suh that kbk1;EI ; with 0 independent of meshsize and approximation order. In partiular, the hoie b = 0 is possible. 3.2 Continuity and Stability We introdue the energy norm jj u jj2h = k 12 rh uk20; + kh 12 p n 21 [ u℄ k20;EI + kh 12 p n 21 uk20;ED : We have the following ontinuity and oerivity properties. Proposition 3.1 Assume the above hypotheses on and on the oeÆients in the denition of the numerial uxes. Then jAh (w; v) + Ih (w; v)j Cont jj w jjh jj v jjh 8w; v 2 V (h) Ah (v; v) + Ih (v; v) Coerjj v jj2h 8v 2 Vh ; with Cont and Coer only depending on , , and the shape-regularity of the mesh. Proof. With arguments similar to the ones in Perugia and Shotzau (2001, Proposition 4.2), we have i h k 12 L(v)k0; Clift ( + 1) kh 21 p n 21 [ v℄ k0;EI + kh 12 p n 21 vk0;ED (5) for all v 2 V (h). The onstant Clift > 0 only depends on the shaperegularity of the mesh. The ontinuity and oerivity of Ah + Ih are now easy onsequenes of estimate (5). 2 From Proposition 3.1 and Strang's lemma, we immediately have the following abstrat error bound. Theorem 3.1 Assume the above hypotheses on and on the oeÆients in the denition of the numerial uxes. Then we have jj u v jjh + C 1 sup jRjhj (wu;jjw)j ; jj u uh jjh 1 + CCont vinf 2V oer h oer w2Vh with the residual Rh (u; w) = Ah (u; w) + Ih (u; w) Fh (w). h 7 An hp-Analysis of the LDG Method for Diusion Problems Remark 3.1 For regular meshes without hanging nodes, onstant approximation orders pK = p and boundary onditions that are restritions to of nite element funtions, we an hoose v in the estimate of Theorem 3.1 as an optimal H 1 -onforming approximant of u, see, e.g., Babuska and Suri (1987, Theorem 4.6), and obtain estimates for jj u uh jjh that are optimal in h and p (the residual is optimally onvergent, see Lemma 3.2 below). Next, we derive hp-error estimates on general unstrutured meshes with hanging nodes. 3.3 hp-Error Estimates We assume that there exists a onstant ` > 0 suh that ` 1 hK hK 0 `hK and ` 1 pK pK 0 `pK for all K and K 0 sharing a (d 1)-dimensional fae. For K 2 Th , we set NÆK := maxfNK 0 : K and K 0 share at least one faeg. We have the following error bound. Theorem 3.2 Assume the above hypotheses on , the oeÆients in the denition of the numerial uxes and the meshes and polynomial degree distributions. Let the exat solution u satisfy ujK 2 H sK +1 (K ) and rujK 2 H sK (K ), for all K 2 Th , with sK 1. Then jj u uh jj2h C 2 min(pK ;sK ) h 1 k ruk2 i; hK 2 N k u k + ÆK sK +1;K sK ;K nK p2KsK 1 K 2Th X with C independent of hK and pK . Moreover, 2 min(pK ;sK ) h i hK NÆK kuk2sK +1;K +(1+ nK1)k ruk2sK ;K : 2 sK 1 pK K 2Th kq qh k20; C X Remark 3.2 The bounds in Theorem 3.2 are slightly suboptimal in the orders pK . On aÆne quadrilateral meshes in two dimensions with hanging nodes, the p-bounds an be improved by using the projetor of Georgoulis and Suli (2001), provided that u belongs to an augmented Sobolev spae. In order to prove Theorem 3.2, we need the following hp-approximation result (see Babuska and Suri (1987, Lemma 4.5)). Lemma 3.1 Let K 2 T and v 2 H K (K ), t 1. Then there exists a h t K sequene of polynomials v in S (K ), pK = 1; 2; : : :, satisfying 1 2 kv vkq;K + hK12 pK hK pK q hK pK pK min(pK +1;tK ) ptK q kv phKK vk0;K C hK q K q kvktK ;K 8 An hp-Analysis of the LDG Method for Diusion Problems for all q with 0 q tK . The onstant C is independent of v, hK and pK , but depends on the shape-regularity of the mesh and on tK . Let now hp v be given by hp vjK = phKK (vjK ), for any K 2 Th , with phKK from Lemma 3.1. For r = (r1 ; : : : ; rd ), we set hp r = (hp r1 ; : : : ; h rd ). Lemma 3.2 Let u be the exat solution. Assume ( ru)j 2 H K (K ) , s K K 2 Th , with loal regularity exponents sK 1. Then, for w 2 V (h), jRh (u; w)j C Proof. d 2 min(pK +1;sK ) 1 i 21 hK k ruk2sK ;K jj w jjh : 2 sK nK pK K 2Th h X Simple alulations lead to Rh (u; w) = Z EIZ + ff ru ( ru)gg b [ ru ( ru)℄℄ [ w℄ ds ED w ru ( ru) n ds; with being the L2-projetion onto Qh . By writing ru ( ru) = [ ru hp ( ru)℄ [ ru hp ( ru)℄, the estimate follows as in Perugia and Shotzau (2001, Lemma 4.11) from the triangle inequality, the CauhyShwarz inequality, inverse estimates for [ ru hp ( ru)℄ from K to K , K 2 Th , the L2-stability of and the approximation properties in Lemma 3.1. 2 We are now ready to prove Theorem 3.2. Proof of Theorem 3.2. We start by estimating jj u hp u jjh . Our assumptions on , meshes and polynomial degree distributions and the approximation properties in Lemma 3.1 yield jj u hp u jj2h C 2 min(pK ;sK ) hK 2 2sK 1 NÆK kuksK +1;K : p K K 2Th X By inserting this and the result of Lemma 3.2 in the bound of Theorem 3.1, we obtain the estimate of jj u uh jjh . To estimate kq qh k0; , we use (3), the triangle inequality and L(u) = G D to obtain kq qh k0; k ru ( rh uh)k0; + k( L(u uh))k0; . From the L2 -stability of and (5), k( L(u uh))k0; C jj u uh jjh . By the triangle inequality, the identity hp ( ru) = hp ( ru) , and again the L2 -stability of , we get k ru ( rh uh)k0; 2k ru hp ( ru)k0; + jj u uh jjh : 9 An hp-Analysis of the LDG Method for Diusion Problems Therefore, the desired result follows from the bound for jj u uh jjh and Lemma 3.1. 2 2 An estimate for the L -error in u an be obtained by using a standard duality argument. We assume that and are suh that the following ellipti regularity result holds true: for any 2 L2 ( ), the solution z to the problem r ( rz ) = in z = 0 on ; (6) satises z 2 H 2 ( ), rz 2 H 1 ( )d and kz k2; C kk0; , k rz k1; C kk0; , with a onstant C > 0. Theorem 3.3 With the same assumptions as in Theorem 3.2 and the above hypothesis on and , we have min(p;s)+1 ps+ 12 ku uh k0; C h kuks+1; + k ruks; ; with h = maxK 2Th hK , p = minK 2Th p 1 and s = minK 2Th sK 1. Proof. Let z be the solution to problem (6) with = u uh . Simple alulations give ku uhk20; = Ah (z; u uh)+ Ih (z; u uh) Rh(z; u uh). Sine Ah (zh; u uh) + Ih (zh ; u uh ) = Rh (u; zh), for any zh 2 Vh and Rh (u; zh) = Rh (u; z zh ), we obtain ku uh k20; =Ah (z zh; u uh ) + Ih (z zh ; u uh) Rh (u; z zh ) Rh (z; u uh ): Therefore, from Proposition 3.1, Lemma 3.2 and the regularity of z , i h ku uhk20; Cont jj z zh jjh + C hp k rz k1; jj u uh jjh min(p+1;s) + C h ps k ruks; jj z zh jjh : By hoosing zh = hp z , from the estimates in Lemma 3.1 and the ellipti regularity assumption, jj z zh jjh C hp 1 kz k2; C hp 1 ku uh k0; , in addition to k rz k1; C ku uhk0; . The result then follows from the estimate of jj u uh jjh in Theorem 3.2. 2 The stability of the LDG formulation with respet to the right-hand side, under mild smoothness assumptions, is implied by the following result. Proposition 3.2 Assume that and are suh that the solution z of (6) d and k rz k with right-hand side 2 V (h) satises rz 2 H s ( ) s; C kk0; for s > 12 . Then, jFh (v)j C [kf k20; + kh 12 p n 21 gD k20;ED ℄ 12 jj v jjh , for all v 2 V (h). An hp-Analysis of the LDG Method for Diusion Problems 10 The assertion follows from the broken Poinare inequality kvk0; 2 V (h), that an be proved following Arnold (1982), and the estimate k 12 G D k0; Cliftkh 21 p n 21 gD k0;ED , obtained as in Perugia and Shotzau (2001, Proposition 4.2). 2 Proof. C jj v jjh , v 4 Conlusions In this paper, we presented the rst hp-error analysis of the LDG method for diusion problems in several spae dimensions and extended the previous h-analysis in Castillo et al. (2000). Although we used the setting of Arnold et al. (2001) to ast the method in its primal form, we proposed a new tehnique to atually derive error estimates based on Strang's lemma. Referenes Arnold, D.N.: 1982, An interior penalty nite element method with disontinuous elements, SIAM J. Numer. Anal. 19, 742{760. Arnold, D.N., Brezzi, F., Cokburn, B. and Marini, L.D.: 2001, Unied analysis of disontinuous Galerkin methods for ellipti problems, SIAM J. Numer. Anal. 39, 1749{1779. Babuska, I. and Suri, M.: 1987, The hp-version of the nite element method with quasiuniform meshes, Model. Math. Anal. Numer. 21, 199{238. Castillo, P.: 2001, Performane of disontinuous Galerkin methods for ellipti partial dierential equations, Tehnial Report 1764, IMA, University of Minnesota, submitted. Castillo, P., Cokburn, B., Perugia, I. and Shotzau, D.: 2000, An a priori error analysis of the loal disontinuous Galerkin method for ellipti problems, SIAM J. Numer. Anal. 38, 1676{1706. Castillo, P., Cokburn, B., Shotzau, D. and Shwab, C.: 2000, Optimal a priori error estimates for the hp-version of the loal disontinuous Galerkin method for onvetion{diusion problems, Tehnial Report 1689, IMA, University of Minnesota, in press in Math. Comp. Cokburn, B. and Shu, C.-W.: 1998, The loal disontinuous Galerkin method for time{dependent onvetion{diusion systems, SIAM J. Numer. Anal. 35, 2440{2463. Cokburn, B. and Dawson, C.: 2000, Some extensions of the loal disontinuous Galerkin method for onvetion-diusion equations in multidimensions, in J. Whitemann (ed.), The Proeedings of the 10th Conferene on the Mathematis of Finite Elements and Appliations, Elsevier, pp. 225{238. Georgoulis, E.H. and Suli, E.: 2001, hp-DGFEM on shape{irregular meshes: reation{diusion, Tehnial Report NA 01{09, Oxford University Computing Laboratory. An hp-Analysis of the LDG Method for Diusion Problems 11 Houston, P., Shwab, C. and Suli, E.: 2000, Disontinuous hp nite element methods for advetion{diusion problems, Tehnial Report NA 00{15, Oxford University Computing Laboratory, in press in SIAM J. Numer. Anal. Perugia, I. and Shotzau, D.: 2001, The hp-loal disontinuous Galerkin method for low{frequeny time{harmoni Maxwell's equations, Tehnial Report 1774, IMA, University of Minnesota, in press in Math. Comp. Prudhomme, S., Pasal, F., Oden, J. and Romkes, A.: 2000, Review of a priori error estimation for disontinuous Galerkin methods, Tehnial Report 200027, TICAM, University of Texas at Austin. Riviere, B., Wheeler, M. and Girault, V.: 1999, Improved energy estimates for interior penalty, onstrained and disontinuous Galerkin methods for ellipti problems, Part I, Computational Geosienes 3 (4), 337{360.