A Central Discontinuous Galerkin Method for Hamilton-Jacobi Equations

J Sci Comput (00) 45: 404 48 DOI 0.007/s095-009-9340-y A Central Discontinuous Galerkin Method for Hamilton-Jacobi Equations Fengyan Li Sergey Yakovlev Received: 6 August 009 / Revised: November 009 / Accepted: December 009 / Published online: December 009 Springer Science+Business Media, LLC 009 Abstract In this paper, a central discontinuous Galerkin method is proposed to solve for the viscosity solutions of Hamilton-Jacobi equations. Central discontinuous Galerkin methods were originally introduced for hyperbolic conservation laws. They combine the central scheme and the discontinuous Galerkin method and therefore carry many features of both methods. Since Hamilton-Jacobi equations in general are not in the divergence form, it is not straightforward to design a discontinuous Galerkin method to directly solve such equations. By recognizing and following a weighted-residual or stabilization-based formulation of central discontinuous Galerkin methods when applied to hyperbolic conservation laws, we design a high order numerical method for Hamilton-Jacobi equations. The L stability and the error estimate are established for the proposed method when the Hamiltonians are linear. The overall performance of the method in approximating the viscosity solutions of general Hamilton-Jacobi equations are demonstrated through extensive numerical experiments, which involve linear, nonlinear, smooth, nonsmooth, convex, or nonconvex Hamiltonians. Keywords Discontinuous Galerkin method Central scheme Hamilton-Jacobi equation Viscosity solution High order accuracy The research of F. Li was supported in part by the NSF under the grant DMS-06548, NSF CAREER award DMS-08474 and by an Alfred P. Sloan Research Fellowship. Additional support was provided by NSFC grant 06709 while Li was visiting Department of Mathematics at Nanjing University, China. The research of S. Yakovlev was supported in part by the NSF CAREER award DMS-08474. F. Li ( ) S. Yakovlev Department of Mathematical Sciences, Rensselaer Polytechnic Institute, Troy, NY 80, USA e-mail: lif@rpi.edu S. Yakovlev e-mail: yakovs@rpi.edu

J Sci Comput (00) 45: 404 48 405 Introduction Hamilton-Jacobi (H-J) equation t ϕ(x,t)+ H(x,ϕ(x,t), x ϕ(x,t))= 0, x R d (.) with suitable initial and boundary conditions arises in various applications, such as optimal control, differential games, seismic waves, crystal growth, robotic navigation, image processing and calculus of variations. The solution of such equation may develop discontinuous derivatives in finite time even when the initial data is smooth, indicating the classical solutions may not always exist. To ensure the existence of the solution as well as to single out the physically relevant solution, the concept of viscosity solutions was established for H-J equations. Such development is due to Crandall, Evans and Lions among many others, see e.g. [5, 7]. Throughout the paper, the Hamiltonian H(x,q,p) in (.) is called linear (resp. nonlinear, smooth, nonsmooth, convex, nonconvex) when it is linear (resp. nonlinear, smooth, nonsmooth, convex, nonconvex) with respect to p. Since the early work by Crandall and Lions [6] and Souganidis [38], there have been many numerical methods developed to solve for the viscosity solution of (.). Because of the focus of this paper, some of these methods with high order accuracy are listed here, such as essentially nonoscillatory (ENO) schemes [33, 34], weighted ENO (WENO) and Hermite WENO schemes [, 35, 40], central schemes [4, 3, 4], and discontinuous Galerkin methods [6, 0, 7]. The primary goal of this paper is to design a high order numerical method, namely, a central discontinuous Galerkin (DG) method, to solve H-J equations for their viscosity solutions. DG finite element methods were originally devised for conservation laws [7 9,,, 36] and later for many other applications, see, e.g. [, 0, 3]. Compared with other methods, DG methods have many attractive features such as being flexible with complicated geometries, different boundary conditions, and various local approximations [4, 7, 39], using compact stencils to achieve high order accuracy, and the easy parallel implementation. Till now there are mainly two DG methods available for solving H-J equations. Both were formulated when the Hamiltonian H in (.) is in the form of H(x, x ϕ(x,t)),thatis, t ϕ(x,t)+ H(x, x ϕ(x,t))= 0, x R d (.) and they can be extended to the general H-J equation (.). The first approach is based on the elegant connection between the H-J equation and the conservation law system. Notice by differentiating (.) with respect to x, one obtains a conservation system for the gradient field w = x ϕ, t w(x,t)+ x H(x, w(x,t))= 0, x R d. (.3) Different from the standard conservation system, the components of w are not independent. By following the derivation of DG methods for conservation laws, in addition with a least squares procedure in each time step (or each time stage depending on the particular time discretization used) if (.3) is multi-dimensional, a DG method was proposed in [0]. The method first solves for w, then the missing constant in ϕ is recovered based on the original equation (.). Later this method was reinterpreted and simplified by Li and Shu in [7], with the use of locally curl-free approximations for w in a standard DG framework. This is based on the intrinsic structure of w: curl x w = curl x x ϕ = 0. The new equivalent formulation in [7] avoids the use of the least squares procedure and involves fewer unknowns, and therefore reduces the overall computational complexity for multi-dimensional cases.

406 J Sci Comput (00) 45: 404 48 Moreover, it recasts the DG method for H-J equations in [0, 5] into the method of lines framework, which is more natural for analysis and implementation. The DG method reviewed above reliably captures the viscosity solutions of (.). In terms of accuracy, it is optimal for x ϕ and in general suboptimal for ϕ itself with respect to the approximation properties of discrete spaces. On the other hand, the method is based on differential equations for x ϕ, and in multi-dimensional cases, a single equation (.) is converted to a system of (.3). In addition, the solutions of (.) are often smoother than those of (.3). This motivates the development of the second DG method in [6] to directly solve (.). However it is not straightforward to design a direct method for H-J equations in the DG framework. With one dimension as an example, notice (.) contains H(x, x ϕ) and it is in general not in the divergence form. When (.) is multiplied by a test function and integrated, integration by parts can not further shift the differential operator in space from ϕ to the test function, a simple but commonly used step to define weak solutions and to design DG methods for many applications. On the other hand, the weighted-residual or stabilization-based formulations of DG methods for linear PDEs are discussed in [, 3], which reveal in many DG methods the balance of how the numerical solution satisfies the PDE locally, how the continuity of certain qualities is imposed across element interfaces and even how the boundary condition is satisfied. The method proposed in [6] can be regarded as a DG method with such structure yet for equations which might be nonlinear. For linear Hamiltonians with discontinuous coefficients or for nonlinear Hamiltonians, since the weighted-residual type method alone in [6] is of Roe type, the scheme may generate entropy violating solutions. To ensure the convergence to viscosity solutions, an additional entropy correction procedure was adopted. The L stability and the optimal error estimate are established when the Hamiltonian is linear with smooth coefficients. For nonlinear cases with convex Hamiltonians, numerical experiments demonstrate the stability and the optimal accuracy of the overall scheme. In this paper, we propose a high order numerical method, namely, a central DG method, to directly solve for the viscosity solutions of H-J equations. Central DG methods were introduced in [9 3] for hyperbolic conservation laws, and they combine the central scheme [, 8, 3] and the DG method. These methods evolve two copies of approximating solutions defined on overlapping meshes, and they avoid the use of Riemann solvers which can be complicated and costly for system of equations [6]. The co-existence of two numerical solutions also provides new opportunities, and one such example is the nonoscillatory hierarchical reconstruction based on both solutions discussed in [30]. Besides, the methods carry many features of standard DG methods. What motivates the scheme in this paper is the observation that the central DG method, when solving hyperbolic conservation laws [30], can be rewritten into a weighted-residual or stabilization-based formulation [, 3]. By following the structure of such formulation, and with the similar treatment as in [6] to impose the continuity of the solution across element interfaces, we introduce a central DG method for H-J equations (.). When H(x,ϕ, x ϕ) = c x ϕ with c being a constant field, the H-J equation is also a conservation equation, and the proposed method is the same as the one in [30]. When the method is applied to H-J equations with linear Hamiltonians, we establish the L stability and the error estimate. This estimate is suboptimal with respect to the approximation properties of discrete spaces, however, we always observe the optimal accuracy of the proposed method for smooth solutions. In addition, a nonlinear limiting strategy is introduced in order to capture the viscosity solutions for some examples with nonconvex Hamiltonians, see Sect.. and Sect. 4. The overall performance of the proposed method in terms of high order accuracy and reliability, when approximating the viscosity solutions of general H-J equations,

J Sci Comput (00) 45: 404 48 407 is demonstrated through a series of numerical experiments which involve linear, nonlinear, smooth, nonsmooth, convex, or nonconvex Hamiltonians. The rest of this paper is organized as follows. In Sect., a central DG method is motivated and formulated for H-J equations. Theoretical analysis in terms of the L stability and the error estimate for linear Hamiltonian cases is carried out in Sect. 3. In Sect. 4, a collection of one and two dimensional numerical examples are presented to demonstrate the performance of the proposed method, which are followed by concluding remarks in Sect. 5. Numerical Scheme To motivate the scheme, we start with a brief review of the central DG method when applied to a one dimensional scalar conservation equation t ϕ(x,t) + x f(ϕ(x,t))= 0, x (0,a). (.) For simplicity, the periodic boundary condition is assumed. The method can also be defined for multi-dimensional or system cases with general boundary conditions. Let x i } i be a partition of [0,a] with x i+ = (x i + x i+ ), I i = (x i,x i+ ) and I i+ = (x i,x i+ ). Two discrete spaces associated with overlapping meshes x i } i and x i+ } i are defined V h = V k h =v : v I i P k (I i ), i}, W h = W k h =v : v I i+ P k (I i+ ), i} where P k (I) is the set of polynomials with the degree at most k on I. We further denote w(x ± ) = lim ɛ 0± w(x + ɛ), and the jump of w at x i, x i+ as [w] i = w(x + i ) w(x i ), [w] i+ = w(x + ) w(x ) respectively. i+ i+ To solve (.), the central DG method uses both spaces V h and W h and its semi-discrete formulation is given as follows [30]: find ϕ h (,t) V h and ψ h (,t) W h, such that for any η V h, ξ W h and for all i, t ϕ h ηdx = (ψ h ϕ h )ηdx + f(ψ h )η x dx I i τ max I i I i f(ψ h (x i+, t))η(x ) + f(ψ i+ h (x i, t))η(x + ), (.) i I i+ t ψ h ξdx = τ max (ϕ h ψ h )ξdx + f(ϕ h )ξ x dx I i+ I i+ f(ϕ h (x i+, t))ξ(x i+ ) + f(ϕ h(x i, t))ξ(x + i ). (.3) Here τ max is the maximal time step allowed by the stability condition. It is more suitable to write τ max = τ max (t), as in implementation τ max is often dynamically chosen and it is closely related to the actual time step, see Sect. 4. One can refer to [8, 30] for the origin of τ max. The key observation that motivates the proposed scheme is the following reformulation of (.) (.3), after integration by parts with respect to x, ( t ϕ h + x f(ψ h ))ηdx = (ψ h ϕ h )ηdx [f(ψ h )] i η(x i ), (.4) I i τ max I i ( t ψ h + x f(ϕ h ))ξdx = (ϕ h ψ h )ξdx [f(ϕ h )] I τ i+ ξ(x i+ max I i+ ). (.5) i+

408 J Sci Comput (00) 45: 404 48 Notice that the first term in each of the above equations is a generalized residual in the sense that if ϕ h = ψ h, this term measures how ϕ h satisfies the PDE (.) locally. Equations (.4) and(.5) reflect a balance between such generalized residual, continuity of the flux function f( ) at the element interfaces, and the difference of two approximating solutions ϕ h and ψ h. DG methods with similar structures defined on one mesh, termed as DG methods in the weighted-residual or stabilization-based formulations, are discussed in [, 3] for linear PDEs. Based on the central DG method for hyperbolic conservation laws with the structure of (.4) (.5) as well as the direct DG method for solving H-J equations introduced by Cheng and Shu [6], with the same notations as introduced above for meshes and discrete spaces V h and W h, we propose the following central DG method for the one dimensional H-J equation (.): look for ϕ h (,t) V h and ψ h (,t) W h, such that for any η V h, ξ W h and for all i, ( t ϕ h + H(x,ψ h, x ψ h ))ηdx I i = τ max I i (ψ h ϕ h )ηdx H (x, ϕ h, x ϕ h ) x=xi [ψ h ] i η(x i ), (.6) I i+ ( t ψ h + H(x,ϕ h, x ϕ h ))ξdx = (ϕ h ψ h )ξdx H (x, ψ h, x ψ h ) x=xi+ [ϕ h ] τ i+ ξ(x max I i+ i+ ), (.7) where H (x,q,p)= H(x,q,p), the characteristic speed of (.) according to its characteristic equations [8]. Though the treatment for the jumps of ϕ h and ψ h at element interfaces p is similar to the one in [6], the terms involving ϕ h ψ h introduce additional stabilization mechanism and therefore the proposed scheme is reliable to capture the viscosity solutions without any additional entropy correction procedure, see Theorem 3. in Sect. 3 and numerical examples in Sect. 4. The method proposed above can also be defined for multi-dimensional H-J equations with structured or unstructured meshes. For simplicity, we here consider a two dimensional domain =[0,a] [0,b] with overlapping rectangular meshes. Let x i } i and y j } j be partitions of [0,a] and [0,b] respectively, with x i+ = (x i + x i+ ), I i = (x i,x i+ ), I i+ = (x i,x i+ ),andy j+ = (y j + y j+ ), J j = (y j,y j+ ), J j+ = (y j,y j+ ).Then D i,j } i,j and D i+,j+ } i,j define two overlapping meshes for, with D i,j = I i J j and D i+,j+ = I i+ J j+,seefig.. Associated with these meshes, one can define two discrete spaces: V h = V k h =v : v D i,j P k (D i,j ), i, j}, W h = W k h =v : v D i+,j+ P k (D i+,j+ ), i, j}. Here P k (D) is the set of polynomials with the total degree at most k on D. Similar to one dimensional cases, a central DG method for two dimensional H-J equations (.) is defined as follows: look for ϕ h (,t) V h and ψ h (,t) W h, such that for any

J Sci Comput (00) 45: 404 48 409 Fig. Two dimensional overlapping meshes η V h, ξ W h and for all i and j, ( t ϕ h + H(x,ψ h, x ψ h ))ηdx D i,j = (ψ h ϕ h )ηdx τ max D i,j H (x,ϕ h, x ϕ h ) x=xi [ψ h (,y)] i η(x i,y)dy J j H (x,ϕ h, x ϕ h ) y=yj [ψ h (x, )] j η(x,y j )dx, I i ( t ψ h + H(x,ϕ h, x ϕ h ))ξdx D i+,j+ = (ϕ h ψ h )ξdx τ max D i+,j+ H (x,ψ h, x ψ h ) x=xi+ [ϕ h (,y)] i+ ξ(x J j+ i+,y)dy H (x,ψ h, x ψ h ) y=yj+ [ϕ h (x, )] j+ ξ(x,y I i+ j+ )dx. Here (H (x, q, p), H (x, q, p)) = p H(x,q,p), which is related to the characteristic speed of H-J equations [8], and τ max is the maximal time step allowed by the stability condition [8, 30].

40 J Sci Comput (00) 45: 404 48 Remark. When H(x,ϕ, x ϕ) = c x ϕ with c being a constant field, the H-J equation is also a conservation equation, and the proposed method is the same as the one in [30]. When (.) is nonlinear, the proposed scheme is consistent only when k. Take one dimensional Burgers equation as an example, that is, H(x,ϕ, x ϕ) = ( xϕ) and H (x,ϕ, x ϕ) = x ϕ. When k = 0, x ϕ h Ii = x ψ h Ii+ = 0, i, all terms containing H(x,q,p) and H (x,q,p) in (.6) (.7) vanish, and this implies that the scheme is inconsistent to the Burgers equation. Such inconsistency can be avoided by reconstructing the numerical derivative for a piecewise constant function, and such remedy is not pursued in this paper.. Numerical Boundary Condition When the boundary condition is not periodic, the proposed scheme needs to be modified for the boundary elements. Without loss of generality, we consider the one dimensional computational domain (0,a)and assume x = 0andx N = a. To impose the boundary condition numerically, at x = 0, (.6) with i = will be replaced by x3/ x ( t ϕ h + H(x,ψ h, x ψ h ))ηdx x3/ = (ψ h ϕ h )ηdx τ max x H (x, ϕ h, x ϕ h ) x=x (ψ h (x,t) g L (t))η(x ) for the inflow boundary condition φ(0,t)= g L (t), and by x3/ x ( t ϕ h + H(x,ψ h, x ψ h ))ηdx = τ max x3/ x (ψ h ϕ h )ηdx for the outflow boundary condition. At x = a, (.6) with i = N will be replaced by xn x N ( t ϕ h + H(x,ψ h, x ψ h ))ηdx = xn (ψ h ϕ h )ηdx τ max x N H (x, ϕ h, x ϕ h ) x=xn (g R (t) ψ h (x N, t))η(x N ) for the inflow boundary condition φ(a,t)= g R (t), and by xn x N ( t ϕ h + H(x,ψ h, x ψ h ))ηdx = τ max xn x N (ψ h ϕ h )ηdx for the outflow boundary condition. The boundary condition being inflow or outflow is either given along with the equation, or can be numerically determined as follows: H (x, ϕ h, x ϕ h ) x=0 > 0 (resp. 0) implies an inflow (resp. outflow) boundary condition at x = 0, and H (x, ϕ h, x ϕ h ) x=a < 0 (resp. 0) implies an inflow (resp. outflow) boundary condition at x = a.

J Sci Comput (00) 45: 404 48 4. Nonlinear Limiter It is demonstrated that for some examples with nonconvex Hamiltonians, nonlinear limiters are needed in order for the proposed scheme to capture the viscosity solution. Based on numerical experiments, the following limiting strategy is introduced for P and P approximations: consider three neighboring cells C,C,C 3 }, which could be I i,i i+,i i+ } or I i,i i,i i+ } for some i. OnC i (i =,, 3), the numerical solution is represented as U i (x) = a 0,i + a,i ξ i for P 3ξi approximations, and U i (x) = a 0,i + a,i ξ j + a,i for P approximations. Here ξ i = x c i C i / with c i being the cell center, and a 0,i is the cell average of U i (x) on C i. Note the basis functions,x, 3x are the first three Legendre polynomials on [, ]. To apply the limiting procedure on C, we first compute two candidates for the slope of U (x) in C based on cell averages, s = a 0,3 a 0,, s = a 0, a 0, ; let ã, = minmod(a,,s,s ).Ifa, =ã,, no change is needed for U (x). Otherwise we replace U (x) by Ũ (x) = a 0, +ã, ξ, here mind,d,d 3 } if d,d,d 3 > 0, minmod(d,d,d 3 ) = maxd,d,d 3 } if d,d,d 3 < 0, 0 otherwise. In the actual implementation, a, =ã, can be replaced by a, ã, ɛ. For instance in Sect. 4, ɛ = 0 7 is used for Examples 4.7 and 4.8. This limiting strategy can be regarded as the minmod slope limiter in [9] yet based on two numerical solutions on overlapping meshes. For the P case, it is the same as the nonoscillatory hierarchical reconstruction procedure proposed in [30]..3 Time Discretization Up to now we only have the semi-discrete central DG method, resulting the method of lines ODE system t = L(, t), (.8) where the dependence of L(, ) on t comes from the numerical boundary condition. This ODE system can be further discretized by total variation diminishing (TVD) Runge-Kutta methods [9, 37]. In Sect. 4, the following second and third order Runge-Kutta methods are used when k =, inv k h (also in W k h ) respectively. n (resp. n+ ) is the numerical solution of (.8) att n (resp. t n+ ), and t n = t n+ t n. RK: () = n + t n L( n,t n ), n+ = n + () + t nl( (),t n+ ).

4 J Sci Comput (00) 45: 404 48 RK3: () = n + t n L( n,t n ), () = 3 4 n + 4 () + 4 t nl( (),t n+ ), n+ = 3 n + 3 () + ( 3 t nl (),t n + ) t n. 3 Theoretical Results In this section, the L stability and the error estimate will be established for the scheme proposed in Sect. when the Hamiltonian H in (.) is linear. The analysis is closely related to those in [6, 3]. Though only the results for one dimensional cases are presented, similar results can be obtained for multi-dimensional cases. 3. L Stability Theorem 3. Consider the one dimensional H-J equation (.) with H(x,ϕ, x ϕ) = c(x) x ϕ and with the periodic boundary condition on (0,a). Let ϕ h V h and ψ h W h be two numerical solutions of the semi-discrete central DG method (.6) (.7), and assume β = c x <. Then the following L stability holds T E(T ) + 0 τ max (t)e β(t T) dt E(0)e βt, (3.) where E(t) = (t) = a 0 a In particular, when c(x) is a constant, then 0 ϕh (x, t) + ψ h (x, t)dx, (ϕ h (x, t) ψ h (x, t)) dx. (3.) T E(T ) + 0 τ max (t)dt = E(0). (3.3) Proof By taking H(x,q,p)= c(x)p in the scheme of (.6) (.7), one gets t ϕ h ηdx = c(x) x ψ h ηdx I i I i + (ψ h ϕ h )ηdx c(x i )[ψ h ] i η(x i ), (3.4) τ max I i t ψ h ξdx = c(x) x ϕ h ξdx I i+ I i+ + (ϕ h ψ h )ξdx c(x τ i+ )[ϕ h ] max I i+ ξ(x i+ ). (3.5) i+

J Sci Comput (00) 45: 404 48 43 Let η = ϕ h in (3.4)andξ = ψ h in (3.5), and sum up (3.4) (3.5) over i, weget d dt a 0 ϕ h + ψ h dx a = (ψ h ϕ h ϕh τ + ϕ hψ h ψh )dx max 0 ( c(x) x ψ h ϕ h dx + c(x) x ϕ h ψ h dx i I i I i+ ) + c(x i )[ψ h ] i ϕ h (x i ) + c(x i+ )[ϕ h ] i+ ψ h (x i+ ) = a τ max 0 ( xi + i i (ϕ h ψ h ) dx x i c x (x)ψ h ϕ h dx + xi+ ( c(x)ϕh ψ h x i x i + c(x)ψ h ϕ h x i+ x i x i ) c x (x)ψ h ϕ h dx + c(x i )[ψ h ] i ϕ h (x i ) + c(x i+ )[ϕ h ] i+ ψ h (x i+ ) ) a a = (ϕ h ψ h ) dx + c x (x)ϕ h ψ h dx τ max 0 0 ( c(xi )ϕ h (x i )ψ h (x i ) c(x i )ψ h(x i )ϕ h (x + ) + c(x i i+ )ψ h (x i+ )ϕ h (x ) i+ i c(x i )ϕ h (x i )ψ h (x + i ) + c(x i )[ψ h ] i ϕ h (x i ) + c(x i+ )[ϕ h ] i+ ψ h (x i+ ) ) a a = (ϕ h ψ h ) dx + c x (x)ϕ h ψ h dx. τ max 0 0 Now with E(t) and (t) defined as (3.), one gets d dt E(t) + (t) = τ max If c(x) is a constant, (3.6) becomes a d dt E(t) + (t) = 0, τ max and integrating in time from 0 to T leads to (3.3). For more general c(x) with β = c x <, (3.6) becomes d dt E(t) + (t) βe(t), τ max 0 c x (x)ϕ h ψ h dx. (3.6) therefore d E(t) βe(t) (t). (3.7) dt τ max

44 J Sci Comput (00) 45: 404 48 Multiplying e βt to both sides of (3.7) and integrating in time from 0 to T, one gets (3.). As pointed out in [3], Theorem 3. indicates that the energy dissipation of the proposed scheme is closely related to a 0 (ϕ h(x, t) ψ h (x, t)) dx,thel difference of two numerical solutions ϕ h and ψ h. This is different from the standard DG method which evolves one numerical solution, and the energy dissipation is related to the jumps of the numerical solution at element interfaces. 3. Error Estimate Theorem 3. Consider the one dimensional H-J equation (.) with H(x,ϕ, x ϕ) = c x ϕ and with the periodic boundary condition on = (0,a). Let ϕ denote the exact solution with the initial condition ϕ(x,0) = ϕ 0 (x) H k+ ( ), and ϕ h V h = V k h, ψ h W h = W k h be two numerical solutions of the semi-discrete central DG method (.6) (.7). Then the following L error estimate holds ϕ ϕ h L ( ) + ϕ ψ h L ( ) Ch k. (3.8) Here the constant C depends on ϕ 0 H k+ ( ) and the final time T. Based on Remark., the proposed scheme is exactly the same as the one in [30] when applied to the linear H-J equation with constant coefficient, which is also the linear scalar conservation law equation. And the result (3.8) was established by Theorem. in [3]. Notice that with respect to the approximation properties of discrete spaces V k h and W k h, the error estimate (3.8) is suboptimal. In numerical examples however we always observe optimal convergence rates for smooth examples. 4 Numerical Examples In this section we report a sequence of one and two dimensional numerical examples to illustrate the high order accuracy and reliability of the proposed method when approximating the viscosity solutions of H-J equations (.). When discretizing (.6) (.7) (orthetwo dimensional scheme) from time t n to t n+ = t n + t n, τ max in the scheme and the time step t n are computed by τ max = dt n, t n = θdt n, with θ (0, ] as a parameter, and dt n is dynamically determined by dt n = C CFLh x λ x,n (4.) in one dimensional cases and by dt n = C CFL / ( λ x,n h x + λ ) y,n h y (4.) in two dimensional cases. Here λ x,n = H and λ y,n = H at t = t n,andtheyare computed numerically. h x = min i (min(i i,i i+ )) and h y = min j (min(j j,j j+ )). In the numerical experiments reported below, θ = is taken. Unless specified otherwise, the second order Runge-Kutta method is used for P approximations with C CFL = 0.45, and the third order Runge-Kutta method is used for P approximations with C CFL = 0.33 [3].

J Sci Comput (00) 45: 404 48 45 Table Errors and convergence orders for Example 4. on a uniform mesh of N cells, t = N L error Order L error Order P 0 0 3.57e 0 7.77e 0 0.7e 0.05 3.6e 0.0 40 8.35e 0.04.70e 0.09 80 4.3e 0.0 8.e 0.05 60.05e 0.0 4.04e 0.0 P 0 4.33e 0 8.7e 0 0.05e 0.05.e 0.04 40.55e 03.04 5.8e 03.04 80 6.09e 04.06.4e 03.06 60.63e 04.9 3.9e 04.9 P 0.68e 03 3.33e 03 0.7e 04.96 4.03e 04 3.05 40.58e 05 3.07 4.8e 05 3.07 80 3.40e 06.9 6.0e 06.98 60 4.7e 07 3.03 7.49e 07 3.0 4. One Dimensional Examples Example 4. We consider the linear advection equation ϕ t + ϕ x = 0, x (0, π) ϕ(x,0) = sin(x) with the smooth initial data and the periodic boundary condition. For this smooth example, numerical errors and convergence orders are reported in Table for P 0, P and P cases. Even though the error estimate in Sect. 3 shows that the accuracy is suboptimal, we observe optimal (k + )-st order of convergence rate when Vh k and Wh k are used. Recall that the proposed scheme is inconsistent when (.) is nonlinear (see Remark.), and this is the only example of which numerical results are reported in this paper when k = 0. Example 4. We consider the linear advection equation ϕ t + ϕ x = 0, x (, ) ϕ(x,0) = ϕ 0 (x)

46 J Sci Comput (00) 45: 404 48 Fig. Example 4.. t = (left), t = 8(right), N = 00, using P polynomials Fig. 3 Example 4.. t = (left), t = 8(right), N = 00, using P polynomials with the periodic boundary condition and the nonsmooth initial data ( 3 ϕ 0 (x) = + 9 + π 3 cos( 3πx ) 3, x< 3 ) 3 (x + ) + + 3cos(πx), 3 x<0 5 3cos(πx), 0 x< 3 8+4π+cos(3πx) + 6πx(x ), 3 3 x< This example is used to illustrate the dissipative properties of the proposed method. Figure (resp. Fig. 3) includes the exact solution as well as the P (resp. P ) numerical solutions after one time period (left) and after four time periods (right), and the features of the solution are well captured by the method. In addition, one can see that the scheme has smaller numerical dissipation when using higher degrees of polynomials as approximations. This is generally the case when lower order methods are compared with higher order methods. Example 4.3 We consider the following linear equation ϕ t + sign(cos(x))ϕ x = 0, x (0, π) ϕ(x,0) = sin(x) with the nonsmooth coefficient and the periodic boundary condition. The exact solution is piecewise smooth and its formula can be found in [6].

J Sci Comput (00) 45: 404 48 47 Table Errors and convergence orders for Example 4.3 on a uniform mesh of N cells, computed in the subregion (0, π)\[.5,.64], t = N L error Order L error Order P 0.0e 0.8e 0 0 3.80e 0.4 6.59e 0.47 40.8e 0.69.95e 0.76 80 3.4e 03.78 5.e 03.90 60 7.47e 04.9.5e 03.06 P 0 3.77e 0 5.8e 0 0 9.6e 03.97.35e 0.0 40.44e 03.98 3.4e 03.99 80 9.34e 04.38 9.5e 04.84 60.40e 04.74.0e 04.5 Fig. 4 Example 4.3. t =, N = 00, using P polynomials The numerical solution at t = is plotted in Fig. 4, and it approximates well the exact solution, which contains a shock in ϕ x at x = π and a rarefaction wave around x = 3π. Errors and convergence orders in Table are computed in (0, π)\[.5,.64] where the exact solution is H. The order of accuracy in this subregion is expected to be min(,k+ ). We want to mention that the direct DG method introduced in [6] needs an entropy correction procedure in order to capture the viscosity solution for this example. Example 4.4 We consider the Burgers equation ϕ t + (ϕ x) = 0, x (0, π) ϕ(x,0) = cos(x) with the periodic boundary condition and the smooth initial data. The Hamiltonian is nonlinear and convex.

48 J Sci Comput (00) 45: 404 48 Table 3 Errors and convergence orders for Example 4.4 on a uniform mesh of N cells, t = 0.5 N L error Order L error Order P 0.87e 0 4.97e 0 0 7.67e 03.9.35e 0.88 40.04e 03.9 3.6e 03.90 80 4.84e 04.08 8.56e 04.08 60.6e 04.06.05e 04.06 P 0.6e 03 3.09e 03 0 3.00e 04.9 4.07e 04.93 40 4.e 05.86 5.56e 05.87 80 5.04e 06 3.03 6.7e 06 3.05 60 6.53e 07.95 8.5e 07.98 Fig. 5 Example 4.4. t =, N = 00, using P polynomials At t = 0.5, the exact solution is smooth, and the optimal order of accuracy of the proposed scheme is observed for both P and P approximations in Table 3. Later a shock is formed in ϕ x, and it is approximated sharply by the scheme, see Fig. 5 for the solution at t =. In Table 4, errors and convergence orders at t = are reported, and they are computed in (0, π)\[3.0, 3.8]. The solution in this region is smooth and the numerical results again show optimal accuracy. Example 4.5 We consider the Eikonal equation ϕ t + ϕ x =0, x (0, π) ϕ(x,0) = sin(x) with the periodic boundary condition and the smooth initial data. The Hamiltonian is nonsmooth and convex. This example has the same exact solution as Example 4.3. Numerical errors and convergence orders are reported in Table 5, which are computed in

J Sci Comput (00) 45: 404 48 49 Table 4 Errors and convergence orders for Example 4.4 on a uniform mesh of N cells, computed in the subregion (0, π)\[3.0, 3.8], t = N L error Order L error Order P 0.88e 0 5.8e 0 0 8.67e 03.73.37e 0.9 40.7e 03.68 3.70e 03.89 80 4.4e 04.7 7.64e 04.8 60.03e 04.00.9e 04.00 P 0 3.57e 03 3.03e 03 0.06e 03.75 7.0e 04. 40 3.53e 04.59.84e 04.93 80.79e 05 4.3 8.57e 06 4.43 60 8.00e 07 4.48 6.08e 07 3.8 Table 5 Errors and convergence orders for Example 4.5 on a uniform mesh of N cells, computed in the subregion (0, π)\[.5,.64], t = N L error Order L error Order P 0.4e 0.3e 0 0.5e 0.8 4.4e 0.36 40 8.76e 03.5.34e 0.63 80.77e 03.66 3.8e 03.8 60 6.57e 04.07.03e 03.89 P 0 3.78e 0 6.44e 0 0.09e 0.79.65e 0.97 40 3.9e 03.77 4.70e 03.8 80.7e 03.44.46e 03.69 60.68e 04.3 3.87e 04.9 (0, π)\[.5,.64] where the exact solution is H. The order of accuracy in this subregion is expected to be min(,k+ ). Unlike in [6], the entropy correction procedure is not needed for the proposed method to capture the viscosity solution. Example 4.6 The following example ϕ t cos(ϕ x + ) = 0, x (, ) ϕ(x,0) = cos(πx) involves a nonconvex Hamiltonian with the smooth initial data. The boundary condition is periodic. At t = 0.5/π, the exact solution is smooth, and numerical results are presented in Table 6, demonstrating the optimal order of accuracy of the proposed method. By the time

40 J Sci Comput (00) 45: 404 48 Table 6 Errors and convergence orders for Example 4.6 on a uniform mesh of N cells, t = 0.5/π N L error Order L error Order P 0.0e 0.7e 0 0 4.63e 03.8 4.80e 03.7 40.4e 03.0.5e 03.06 80.76e 04.05.77e 04.05 60 6.9e 05.00 6.93e 05.00 P 0.38e 03.55e 03 0.40e 04.53.e 04.8 40.7e 05 3.4.35e 05 3.4 80 3.36e 06 3.0.86e 06 3.04 60 4.5e 07 3.0 3.45e 07 3.05 Fig. 6 Example 4.6. t =.5/π, N = 00, using P polynomials t =.5/π, nonsmooth features have developed in ϕ, and they are reliably captured, see Fig. 6. Example 4.7 We consider the Riemann problem with a nonconvex Hamiltonian ϕ t + 4 (ϕ x )(ϕ x 4) = 0, x (, ) ϕ(x,0) = x. Similarlyasobservedin[0], a nonlinear limiter is needed in order for the proposed scheme to capture the viscosity solution of this example. This can be seen from Fig. 7,which contains the exact solution as well as the numerical solutions with (left) and without (right) the limiting step. The limiter procedure used here is described in Sect... Due to the low regularity of the exact solution, both P and P approximations are of first order accuracy, see Table 7. ForP case, C CFL = 0. is taken to determine the time step by (4.). Note the P results have larger errors, and this can be explained by the nonlinear limiter adopted

J Sci Comput (00) 45: 404 48 4 Table 7 Errors and convergence orders for Example 4.7 on a uniform mesh of N cells, t =. The second order Runge-Kutta method is used for P approximations with C CFL = 0.45, and the third order Runge-Kutta method is used for P approximations with C CFL = 0.. The time step t n = θdt n is dynamically determined by (4.) with θ = N L error Order L error Order P 0.06e 0.6e 0 0.05e 0 0.97.4e 0.9 40 5.3e 0 0.98 5.67e 0.0 80.66e 0 0.99.78e 0.03 60.33e 0.00.38e 0.0 P 0 4.8e 0 5.33e-0 0.4e 0 0.90.57e-0.05 40.4e 0 0.98.4e-0.06 80 5.7e 0 0.99 6.06e-0.03 60.85e 0.00.98e-0.0 Table 8 Errors and convergence orders for Example 4.7 on a uniform mesh of N cells, t =. The third order Runge-Kutta method is used for both P and P approximations, with the fixed time step t = C CFL h x /λ x,and C CFL = 0.8, λ x = 3.5, h x = min i (min(i i,i i+ )) N L error Order L error Order P 0 3.74e 0 4.98e 0 0.84e 0.03.06e 0.7 40 9.3e 0 0.98.00e 0.04 80 4.67e 0.00 4.9e 0.03 60.33e 0.00.43e 0.0 P 0 3.77e 0 4.90e 0 0.87e 0.0.e 0. 40 9.48e 0 0.98.0e 0.05 80 4.75e 0.00 5.0e 0.03 60.37e 0.00.47e 0.0 in the computation. In fact, when this limiter is used in some element I i,thep numerical solution in I i will be replaced with a linear polynomial. On the other hand, it is observed that in the P simulation, the limiter was applied in the central region where the exact solution is not linear. This implies that P and P approximations will have comparable spatial errors for this example, and this is confirmed by the results in Table 8 when the same time discretization, the third order Runge-Kutta method with the fixed time-step t = C CFL h x /λ x as well as C CFL = 0.8 and λ x = 3.5 (an upper bound for H in the (x, t) domain), is applied in P and P simulations. The computation is carried out with the outflow numerical boundary condition (see Sect..). Other nonlinear limiters that were tested yet failed to capture the viscosity solution for both P and P computations include the minmod slope limiter in [9] and the nonoscillatory hierarchical reconstruction procedure in [30]. Further investigation is needed to identify other robust limiting strategies for the proposed scheme when Hamiltonians are nonconvex and when higher order polynomials are used.

4 J Sci Comput (00) 45: 404 48 Fig. 7 Example 4.7. t =, N = 00, using P polynomials with (left) and without limiting (right) Example 4.8 We consider another nonconvex example ϕ t + H(ϕ x ) = 0, x (0, ) ϕ(x,0) = ϕ 0 (x), with the Hamiltonian and the initial condition 4 u( u), u H(u)= 3 u(u ) + 6, u> ϕ 0 (x) = x 4, x 4 0, x > 4. The exact viscosity solution is composed (from left to right) of a linear part, a parabolic part and a constant state, and the nonlinear limiter turns out to be necessary for capturing this solution. This can be seen from Fig. 8, which contains the exact solution and the numerical solution with the limiting step (left), and the zoomed-in plots of the exact solution and numerical solutions with and without the limiting step (right). The limiter procedure used here is described in Sect... In the computation, the outflow numerical boundary condition is applied. Similar to Example 4.7, it is observed that the nonlinear limiter is used in the parabolic section of the solution in all simulations, and therefore we here only report the results (see Table 9) when both P and P computations are performed with the same time discretization, that is, the third order Runge-Kutta method with the fixed time-step t = C CFL h x /λ x as well as C CFL = 0.8 and λ x = 0.5 (an estimate for H in the (x, t) domain). Note both P and P approximations are comparable and they are of first order accuracy due to the low regularity of the exact solution. This example was also discussed in [5] to illustrate the role of nonlinear limiters in capturing viscosity solutions.

J Sci Comput (00) 45: 404 48 43 Table 9 Errors and convergence orders for Example 4.8 on a uniform mesh of N cells, t =. The third order Runge-Kutta method is used for both P and P approximations, with the fixed time step t = C CFL h x /λ x,and C CFL = 0.8, λ x = 0.5, h x = min i (min(i i,i i+ )) N L error Order L error Order P 0.5e 0.94e 0 0 5.73e 03.9.93e 03.7 40.84e 03.0.3e 03.5 80.40e 03.0 5.43e 04.8 60 6.87e 04.03.45e 04.5 P 0.57e 0.4e 0 0 5.7e 03.46 3.3e 03.09 40.8e 03.0.e 03.45 80.39e 03.0 5.35e 04.8 60 6.79e 04.03.4e 04.4 Fig. 8 Example 4.8. t =, N = 00, using P polynomials. The solution with limiting (left), zoomed in comparison of solutions with and without limiting (right) 4. Two Dimensional Examples Example 4.9 We consider the linear advection equation ϕ t + ϕ x + ϕ y = 0, (x,y) (, ) ϕ(x,y,0) = sin(π(x + y)) with the periodic boundary condition and the smooth initial data. This can be regarded as a rotated one dimensional advection equation in two dimensional space. Similarly as in one dimensional case, the numerical results reported in Table 0 show the optimal convergence property of the proposed method for this smooth example. Example 4.0 We consider the Burgers equation ϕ t + (ϕ x + ϕ y ) = 0, (x,y) (0, π) ϕ(x,y,0) = cos(x + y)

44 J Sci Comput (00) 45: 404 48 Table 0 Errors and convergence orders for Example 4.9 on a uniform mesh of N N cells, t = N L error Order L error Order P 0 5.3e 0 4.64e 0 0.55e 0.76.38e 0.75 40 4.05e 03.93 3.6e 03.93 80.0e 03.99 9.0e 04.99 60.56e 04.00.8e 04.00 P 0 3.4e 03.86e 03 0 4.8e 04.99 3.5e 04 3.0 40 5.4e 05 3.06 4.6e 05 3.05 80 6.53e 06.98 5.38e 06.98 60 8.5e 07.94 6.95e 07.95 Table Errors and convergence orders for Example 4.0 on a uniform mesh of N N cells, t = 0., periodic boundary condition N L error Order L error Order P 0.3e 0.3e 0 0 6.79e 03.7 4.9e 03.65 40.56e 03. 9.40e 04.6 80 3.67e 04.08.9e 04.0 60 9.09e 05.0 5.40e 05.0 P 0 3.45e 03.e 03 0 4.76e 04.86 3.0e 04.8 40 5.74e 05 3.05 3.44e 05 3.3 80 6.8e 06 3.07 3.9e 06 3.4 60 8.83e 07.95 5.08e 07.94 with the smooth initial data. We first assume the boundary condition is periodic. At t = 0., the exact solution is still smooth, and the proposed scheme demonstrates the optimal convergent behavior in Table. Later at t = 0.5, nonsmooth features form in the solution which are reliably approximated, see Fig. 9. For the same exact solution, the computation is also carried out with the non-periodic boundary condition based on Sect... The results at t = 0. are reported in Table, and they further confirm the optimal convergence property of the method. Example 4. We consider the following nonlinear example ϕ t + ϕ x ϕ y = 0, (x,y) ( π,π) ϕ(x,y,0) = sin(x) + cos(y)

J Sci Comput (00) 45: 404 48 45 Table Errors and convergence orders for Example 4.0 on a uniform mesh of N N cells, t = 0., non-periodic boundary condition N L error Order L error Order P 0.4e 0.34e 0 0 6.8e 03.7 4.e 03.67 40.57e 03. 9.46e 04.5 80 3.70e 04.08.0e 04. 60 9.6e 05.0 5.4e 05.0 P 0 3.46e 03.09e 03 0 4.79e 04.85 3.03e 04.79 40 5.80e 05 3.04 3.49e 05 3. 80 6.96e 06 3.06 4.03e 06 3. 60 9.e 07.9 5.3e 07.9 Fig. 9 Example 4.0. t = 0.5, N N = 40 40, using P polynomials with the periodic boundary condition. Different from Examples 4.9 and 4.0, this example is genuinely two dimensional. Figure 0 shows how the numerical solutions look like when the exact solution is smooth (left, at t = 0.8), and when singular features form in the solution (right, at t =.5). Errors and convergence orders are presented in Table 3 at t = 0.8, demonstrating the optimal convergence property of the proposed method. Example 4. This example is related to controlling optimal cost determination, see [34] ϕ t + sin(y)ϕ x + (sin(x) + sign(ϕ y ))ϕ y sin (y) + cos(x) = 0, (x,y) ( π,π) ϕ(x,y,0) = 0. The boundary condition is periodic and the Hamiltonian is nonsmooth. The numerical solution (left) and the optimal control ω = sign(ϕ y ) (right) at t = areshowninfig.. The results are comparable to those by the DG methods in [0] and the WENO methods in [40].

46 J Sci Comput (00) 45: 404 48 Table 3 Errors and convergence orders for Example 4. on a uniform mesh of N N cells, t = 0.8 N L error Order L error Order P 0.6e 0.87e 0 0 6.86e 03.93 4.74e 03.98 40.7e 03.00.8e 03.00 80 4.e 04.0.9e 04.0 60.05e 04.0 7.30e 05.00 P 0 3.9e 03.3e 03 0 4.46e 04 3.4.40e 04 3.5 40 6.30e 05.8 3.5e 05.93 80 7.99e 06.98 3.90e 06 3.0 60 9.96e 07 3.00 4.77e 07 3.03 Fig. 0 Example 4.. t = 0.8 (left), t =.5 (right), N N = 40 40, using P polynomials Fig. Example 4.. The numerical solution (left) and the optimal control sign(ϕ y ) (right). t =, N N = 40 40, using P polynomials 5 Concluding Remarks A central discontinuous Galerkin method is proposed in this paper to solve Hamilton-Jacobi equations. The method is motivated by the weighted-residual or stabilization-based formulation of central discontinuous Galerkin methods when applied to hyperbolic conservation laws [9 3]. The L stability and the error estimate are established for linear Hamiltonians. And the high order accuracy and reliability of the proposed method, when approxi-

J Sci Comput (00) 45: 404 48 47 mating the viscosity solutions of more general Hamilton-Jacobi equations, are demonstrated through a series of numerical experiments. More investigation is needed to identify other robust nonlinear limiting strategies particularly suitable for the proposed method to capture the viscosity solutions when Hamiltonians are nonconvex and when higher order polynomials are used. The suboptimal error estimate for linear cases can also be further improved. References. Arnold, D.N., Brezzi, F., Cockburn, B., Marini, L.D.: Unified analysis of discontinuous Galerkin methods for elliptic problems. SIAM J. Numer. Anal. 39, 749 779 (00/00). Ayuso, B., Marini, L.D.: Discontinuous Galerkin methods for advection-diffusion-reaction problems. SIAM J. Numer. Anal. 47, 39 40 (009) 3. Brezzi, F., Cockburn, B., Marini, L.D., Süli, E.: Stabilization mechanisms in discontinuous Galerkin finite element methods. Comput. Methods Appl. Mech. Eng. 95, 393 330 (006) 4. Bryson, S., Levy, D.: High-order semi-discrete central-upwind schemes for multidimensional Hamilton- Jacobi equations. J. Comput. Phys. 89, 63 87 (003) 5. Bryson, S., Levy, D.: Mapped WENO and weighted power ENO reconstruction in semi-discrete central schemes for Hamilton-Jacobi equations. Appl. Numer. Math. 56, 4 (006) 6. Cheng, Y., Shu, C.-W.: A discontinuous Galerkin finite element method for directly solving the Hamilton-Jacobi equations. J. Comput. Phys. 3, 398 45 (007) 7. Cockburn, B., Shu, C.-W.: TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws II: general framework. Math. Comput. 5, 4 435 (989) 8. Cockburn, B., Shu, C.-W.: The Runge-Kutta local projection P-discontinuous Galerkin finite element method for scalar conservation laws. Math. Model. Numer. Anal. 5, 337 36 (99) 9. Cockburn, B., Shu, C.-W.: The Runge-Kutta discontinuous Galerkin method for conservation laws V: multidimensional systems. J. Comput. Phys. 4, 99 4 (998) 0. Cockburn, B., Shu, C.-W.: Special Issue on Discontinuous Galerkin Methods. J. Sci. Comput 3 (005). Cockburn, B., Lin, S.-Y., Shu, C.-W.: TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws III: one-dimensional systems. J. Comput. Phys. 84, 90 3 (989). Cockburn, B., Hou, S., Shu, C.-W.: The Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws IV: the multidimensional case. Math. Comput. 54, 545 58 (990) 3. Cockburn, B., Karniadakis, G., Shu, C.-W.: The development of discontinuous Galerkin methods. In: Cockburn, B., Karniadakis, G., Shu, C.-W.: Discontinuous Galerkin Methods: Theory, Computation and Applications. Lecture Notes in Computational Science and Engineering, vol.. Springer, Berlin (000) 4. Cockburn, B., Li, F., Shu, C.-W.: Locally divergence-free discontinuous Galerkin methods for the Maxwell equations. J. Comput. Phys. 94, 588 60 (004) 5. Crandall, M., Lions, P.L.: Viscosity solutions of Hamilton-Jacobi equations. Trans. Am. Math. Soc. 77, 4 (983) 6. Crandall, M., Lions, P.L.: Two approximations of solutions of Hamilton-Jacobi equations. Math. Comput. 43, 9 (984) 7. Crandall, M.G., Evans, L.C., Lions, P.L.: Some properties of viscosity solutions of Hamilton-Jacobi equations. Trans. Am. Math. Soc. 8, 487 50 (984) 8. Evans, L.C.: Partial Differential Equations. American Mathematical Society, Providence (998) 9. Gottlieb, S., Shu, C.-W., Tadmor, E.: Strong stability-preserving high-order time discretization methods. SIAM Rev. 43, 89 (00) 0. Hu, C., Shu, C.-W.: A discontinuous Galerkin finite element method for Hamilton-Jacobi equations. SIAM J. Sci. Comput., 666 690 (999). Jiang, G.-S., Peng, D.: Weighted ENO schemes for Hamilton-Jacobi equations. SIAM J. Sci. Comput., 6 43 (000). Kurganov, A., Tadmor, E.: New high-resolution central schemes for nonlinear conservation laws and convection-diffusion equations. J. Comput. Phys. 60, 4 8 (000) 3. Kurganov, A., Tadmor, E.: New high-resolution semi-discrete central schemes for Hamilton-Jacobi equations. J. Comput. Phys. 60, 70 74 (000) 4. Kurganov, A., Noelle, S., Petrova, G.: Semidiscrete central-upwind schemes for hyperbolic conservation laws and Hamilton-Jacobi equations. SIAM J. Sci. Comput. 3, 707 740 (00) 5. Lepsky, O., Hu, C., Shu, C.-W.: Analysis of the discontinuous Galerkin method for Hamilton-Jacobi equations. Appl. Numer. Math. 33, 43 434 (000)

48 J Sci Comput (00) 45: 404 48 6. Leveque, R.J.: Finite Volume Methods for Hyperbolic Problems. Cambridge University Press, Cambridge (00) 7. Li, F., Shu, C.-W.: Reinterpretation and simplified implementation of a discontinuous Galerkin method for Hamilton-Jacobi equations. Appl. Math. Lett. 8, 04 09 (005) 8. Liu, Y.-J.: Central schemes on overlapping cells. J. Comput. Phys. 09, 8 04 (005) 9. Liu, Y.-J.: Central schemes and central discontinuous Galerkin methods on overlapping cells. In: Conference on Analysis, Modeling and Computation of PDE and Multiphase Flow, Stony Brook, NY, 004 30. Liu, Y.-J., Shu, C.-W., Tadmor, E., Zhang, M.: Central discontinuous Galerkin methods on overlapping cells with a nonoscillatory hierarchical reconstruction. SIAM J. Numer. Anal. 45, 44 467 (007) 3. Liu, Y.-J., Shu, C.-W., Tadmor, E., Zhang, M.: L stability analysis of the central discontinuous Galerkin method and a comparison between the central and regular discontinuous Galerkin methods. ESAIM, Math. Model. Numer. Anal. 4, 593 607 (008) 3. Nessyahu, H., Tadmor, E.: Non-oscillatory central differencing for hyperbolic conservation laws. J. Comput. Phys. 87, 408 463 (990) 33. Osher, S., Sethian, J.: Fronts propagating with curvature dependent speed: algorithms based on Hamilton-Jacobi formulations. J. Comput. Phys. 79, 49 (988) 34. Osher, S., Shu, C.-W.: High-order essentially nonoscillatory schemes for Hamilton-Jacobi equations. SIAM J. Numer. Anal. 8, 907 9 (99) 35. Qiu, J., Shu, C.-W.: Hermite WENO schemes for Hamilton-Jacobi equations. J. Comput. Phys. 04, 8 99 (005) 36. Reed, W.H., Hill, T.R.: Triangular mesh methods for the neutron transport equation. Technical Report LA-UR-73-479, Los Alamos Scientific Laboratory (973) 37. Shu, C.-W.: Total-Variation-Diminishing time discretizations. SIAM J. Sci. Stat. Comput. 9, 073 084 (988) 38. Souganidis, P.E.: Approximation schemes for viscosity solutions of Hamilton-Jacobi equations. J. Differ. Equ. 59, 43 (985) 39. Yuan, L., Shu, C.-W.: Discontinuous Galerkin method based on non-polynomial approximation spaces. J. Comput. Phys. 8, 95 33 (006) 40. Zhang, Y.-T., Shu, C.-W.: High-order WENO schemes for Hamilton-Jacobi equations on triangular meshes. SIAM J. Sci. Comput. 4, 005 030 (003)