Introduction
Scalar conservation laws augmented with nonlocal diffusion operators arise in a wide range of physical models, including overdriven detonations in gases [11 ] , anomalous diffusion in crystal growth and semiconductor processes [35 ] , nonlinear acoustics [34 ] , geophysical flows [12 , 31 ] , and plasma transport [14 ] . In many of these systems, the underlying transport processes deviate substantially from classical Fickian diffusion, exhibiting long-range interactions or heavy-tailed jump processes that are more accurately captured by fractional diffusion operators. As a consequence, the fractional Laplacian has become a central tool for modeling anomalous transport and nonlocal regularization.
We consider scalar conservation laws with a fractional diffusion term of order s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , written as
{ ∂ t u + ∇ ⋅ f ( u ) + ( − Δ ) s u = g , ( x , t ) ∈ D × R + , u ( x , t ) = 0 , ( x , t ) ∈ ∂ D × R + , u ( x , 0 ) = u 0 ( x ) , x ∈ D . (1) \begin{cases}
\partial_t u + \nabla\cdot\boldsymbol{f}(u) + (-\Delta)^s u = g, & (x,t) \in D \times\mathbb{R}^{+}, \\
u(x,t) = 0, & (x,t) \in\partial D \times\mathbb{R}^{+}, \\
u(x,0) = u_0(x), & x \in D. \tag*{(1)}
\end{cases} ⎩ ⎨ ⎧ ∂ t u + ∇ ⋅ f ( u ) + ( − Δ ) s u = g , u ( x , t ) = 0 , u ( x , 0 ) = u 0 ( x ) , ( x , t ) ∈ D × R + , ( x , t ) ∈ ∂ D × R + , x ∈ D . (1)
where D ⊂ R d D \subset\mathbb{R}^d D ⊂ R d , d = 1 , 2 , 3 d = 1,2,3 d = 1 , 2 , 3 , is an open bounded domain, u 0 u_0 u 0 is an initial condition, f : R → R d \boldsymbol{f} : \mathbb{R} \to\mathbb{R}^d f : R → R d is the flux function, and g ( x , t ) g(x,t) g ( x , t ) is a given scalar-valued source term. Here ( − Δ ) s (-\Delta)^s ( − Δ ) s is the spectral fractional Laplacian, see Section 2.
Despite its wide applicability, the numerical approximation of (1) presents several challenges. The fractional Laplacian is intrinsically nonlocal, leading to dense discretization matrices and a high computational cost. In addition, solutions of problems involving fractional diffusion often develop steep gradients or exhibit limited regularity, complicating the design of numerical schemes that are accurate and robust.
A number of numerical methods have been proposed to approximate fractional diffusion operators, including finite differences by Huang and Oberman [26 ] , spectral and Fourier-based approaches by Constantin et al. [13 , 12 ] , spectral vanishing viscosity techniques by Song and Karniadakis [33 ] , the Balakrishnan method by Bonito and Pasciak [3 , 4 ] , Bonito et al. [7 ] and the extension method by Nochetto et al. [30 ] ; we refer, for instance, to [6 , 15 ] for surveys. However, when coupled with nonlinear hyperbolic fluxes, additional difficulties arise: numerical oscillations may violate the discrete maximum principle, positivity, or the invariant domain structure of the PDE. Ensuring monotonicity becomes even more delicate when the diffusion operator is nonlocal.
The primary goal of this work is to construct and analyze a space–time discretization for scalar conservation laws with a fractional Laplacian that preserves the maximum principle at the discrete level. Our approach is based on continuous, piecewise-linear finite elements in space and strong-stability-preserving (SSP) explicit Runge–Kutta schemes in time. A key ingredient is a sinc-quadrature representation of the fractional Laplacian [29 ] applied to the Balakrishnan-type integral identity
( − Δ ) s u = c s ∫ − ∞ ∞ e s y ( e y − Δ ) − 1 ( − Δ u ) d y (2) (-\Delta)^s u = c_s \int_{-\infty}^{\infty} e^{sy}(e^y-\Delta)^{-1}(-\Delta u)\,\mathrm{d}y \tag*{(2)} ( − Δ ) s u = c s ∫ − ∞ ∞ e sy ( e y − Δ ) − 1 ( − Δ u ) d y (2)
which reduces the nonlocal operator to a weighted family of resolvent problems for the classical Laplacian. This representation enables a natural finite-element approximation that interacts cleanly with the underlying discrete structure of the method. The representation (2) was originally advocated by Bonito et al. [8 ] for the integral fractional Laplacian and subsequently by Bonito and Wei [5 ] for the spectral fractional Laplacian.
To enforce the maximum principle, we construct a first-order monotone scheme obtained by combining the resolvent-based discretization of the fractional Laplacian with a graph viscosity mechanism, see, e.g., [20 , 22 , 2 ] . When the mesh satisfies the acute-angle condition, we prove that this low-order scheme satisfies a discrete maximum principle under an explicit CFL constraint that accounts for both the hyperbolic and fractional diffusive contributions. This closes a gap left open in Bonito and Nazarov [2 ] , where a discrete maximum principle was established for the inviscid SQG model only, while in the presence of fractional diffusion only an invariant-domain property was obtained under a more restrictive time-step condition [2 Remark 3.3] . The key observation is that, under the acute-angle condition, the discrete resolvents appearing in the sinc quadrature are monotone, so that the fractional diffusion term can be written as a convex combination of nodal values; see Lemma 3.1 , Lemma 3.2, and Theorem 3.3 . We also treat Dirichlet boundary conditions, which were not considered in Bonito and Nazarov [2 ] and require a modification in the analysis.
Building upon this foundation, we design a high-order method obtained by adding entropy-consistent high-order corrections and applying the convex limiting framework of Guermond et al. [24 ] . The resulting algorithm achieves second-order accuracy in both space and time while preserving the maximum principle and, in the periodic setting, mass conservation. The limiter acts only where needed, allowing the high-order method to remain active in smooth regions without sacrificing non-oscillatory behavior near sharp layers.
The performance of the proposed method is illustrated through several benchmarks. We first verify the expected convergence rates for the fractional heat equation and for a manufactured Burgers’ problem with fractional diffusion. We then apply the algorithm to the Surface Quasi-Geostrophic (SQG) equation, a widely studied geophysical model whose solutions may develop strong fronts. The method robustly captures the formation of sharp gradients while preserving the maximum principle, even on unstructured meshes.
The paper is organized as follows. In §2 we review the sinc-quadrature approximation of the fractional Laplacian. The fully discrete scheme and the first-order invariant-domain preserving approximation are described in § 3 , together with the CFL analysis. High-order extensions based on SSP time stepping and convex limiting are presented in § 4 . Numerical experiments are provided in § 5 . Concluding remarks are given in § 6 .
The spectral fractional Laplacian and its approximations
We recall that ( − Δ ) : H 2 ( D ) ∩ H 0 1 ( D ) ⊂ L 2 ( D ) → L 2 ( D ) (-\Delta): H^{2}(D) \cap H_{0}^{1}(D) \subset L^{2}(D) \to L^{2}(D) ( − Δ ) : H 2 ( D ) ∩ H 0 1 ( D ) ⊂ L 2 ( D ) → L 2 ( D ) is an unbounded operator and has a compact inverse. Therefore, there is a countable collection of eigenpairs { λ i , e i } i ∈ N ⊂ R + × H 0 1 ( D ) \{\lambda_i,e_i\}_{i\in\mathbb{N}} \subset\mathbb{R}^{+} \times H_{0}^{1}(D) { λ i , e i } i ∈ N ⊂ R + × H 0 1 ( D ) such that { e i } i ∈ N \{e_i\}_{i\in\mathbb{N}} { e i } i ∈ N constitutes an orthonormal basis of L 2 ( D ) L^{2}(D) L 2 ( D ) as well as an orthogonal basis of H 0 1 ( D ) H_{0}^{1}(D) H 0 1 ( D ) . The spectral fractional Laplacian ( − Δ ) s (-\Delta)^s ( − Δ ) s , s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , is defined for smooth and compactly supported functions u ∈ C 0 ∞ ( D ) u\in C_{0}^{\infty}(D) u ∈ C 0 ∞ ( D ) by
( − Δ ) s u : = ∑ i ∈ N λ i s u i e i , with u i : = ∫ D u e i for i ∈ N . (3) (-\Delta)^s u := \sum_{i\in\mathbb{N}} \lambda_i^s u_i e_i,\qquad\text{with } u_i := \int_D u e_i \quad\text{for } i\in\mathbb{N}. \tag*{(3)} ( − Δ ) s u := i ∈ N ∑ λ i s u i e i , with u i := ∫ D u e i for i ∈ N . (3)
Furthermore, this definition extends to functions u u u in the interpolation space H 0 s ( D ) : = [ L 2 ( D ) , H 0 1 ( D ) ] s H_{0}^{s}(D):=[L^{2}(D),H_{0}^{1}(D)]_s H 0 s ( D ) := [ L 2 ( D ) , H 0 1 ( D ) ] s (complex interpolation method) by density.
In this section, we advocate for the equivalent representation (2) of the spectral fractional Laplacian. In § 2.1 we give a heuristic justification of (2) and refer to Lunardi [28 ] for a proof using functional calculus. Then, we approximate ( − Δ ) s (-\Delta)^s ( − Δ ) s by introducing a quadrature for (2) . The key here is that, once discretized, (2) only involves the Laplacian operator, which is then easy to handle using standard finite element techniques. The material in this section has been developed in [3 , 4 , 7 , 8 , 5 ] . We also mention that the case of periodic boundary conditions, which are used in the numerical experiments of § 5 , can be treated similarly. We do not pursue this option to avoid adding an additional layer of abstraction, but refer to Bonito and Nazarov [2 ] and Bonito and Lei [1 ] . The reader familiar with this material is invited to jump to § 3 .
Heuristics
One way to understand the origin of the expression (2) consists of realizing that the following identity holds true for any λ > 0 \lambda>0 λ > 0 and any s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) :
λ s = λ s c s ∫ 0 ∞ z s − 1 z + 1 d z = c s ∫ − ∞ ∞ e s y ( e y + λ ) − 1 λ d y , (4) \lambda^s = \lambda^s c_s \int_0^\infty\frac{z^{s-1}}{z+1}\,dz = c_s \int_{-\infty}^{\infty} e^{sy}(e^y+\lambda)^{-1}\lambda\,dy, \tag*{(4)} λ s = λ s c s ∫ 0 ∞ z + 1 z s − 1 d z = c s ∫ − ∞ ∞ e sy ( e y + λ ) − 1 λ d y , (4)
with c s − 1 = ∫ 0 ∞ z s − 1 z + 1 d z c_s^{-1}=\int_0^\infty\frac{z^{s-1}}{z+1}\,dz c s − 1 = ∫ 0 ∞ z + 1 z s − 1 d z (use the change of variables λ z = e y \lambda z=e^y λ z = e y ). Using the beta function B ( x , y ) = ∫ 0 ∞ t x − 1 ( 1 + t ) x + y d t = Γ ( x ) Γ ( y ) Γ ( x + y ) B(x,y)=\int_0^\infty\frac{t^{x-1}}{(1+t)^{x+y}}\,dt=\frac{\Gamma(x)\Gamma(y)}{\Gamma(x+y)} B ( x , y ) = ∫ 0 ∞ ( 1 + t ) x + y t x − 1 d t = Γ ( x + y ) Γ ( x ) Γ ( y ) , defined for ℜ ( x ) > 0 \Re(x)>0 ℜ ( x ) > 0 and ℜ ( y ) > 0 \Re(y)>0 ℜ ( y ) > 0 , we observe that c s − 1 = B ( s , 1 − s ) c_s^{-1}=B(s,1-s) c s − 1 = B ( s , 1 − s ) . Then, Euler’s reflection formula Γ ( 1 − z ) Γ ( z ) = π sin ( π z ) \Gamma(1-z)\Gamma(z)=\frac{\pi}{\sin(\pi z)} Γ ( 1 − z ) Γ ( z ) = s i n ( π z ) π , which holds for any z ∉ Z z\notin\mathbb{Z} z ∈ / Z , gives c s = sin ( π s ) π c_s=\frac{\sin(\pi s)}{\pi} c s = π s i n ( π s ) . Using the basis { e i } i ∈ N \{e_i\}_{i\in\mathbb{N}} { e i } i ∈ N made of eigenfunctions of ( − Δ ) (-\Delta) ( − Δ ) introduced above and in view of (3) , we have
( − Δ ) s u = ∑ i ∈ N u i ( − Δ ) s e i = ∑ i ∈ N u i λ i s e i (-\Delta)^s u=\sum_{i\in\mathbb{N}}u_i(-\Delta)^s e_i=\sum_{i\in\mathbb{N}}u_i\lambda_i^s e_i ( − Δ ) s u = i ∈ N ∑ u i ( − Δ ) s e i = i ∈ N ∑ u i λ i s e i
so that (4) , applied with λ = λ i \lambda=\lambda_i λ = λ i , gives (2) .
Generic quadrature
Let M , N M,N M , N be two positive integers. We set L : = { − M : N } \mathcal{L}:=\{-M:N\} L := { − M : N } , with the notation { a : b } : = { a , a + 1 , … , b − 1 , b } \{a:b\}:=\{a,a+1,\ldots,b-1,b\} { a : b } := { a , a + 1 , … , b − 1 , b } for any a , b ∈ Z a,b\in\mathbb{Z} a , b ∈ Z such that a ≤ b a\leq b a ≤ b . Let { ( ω l , y l ) } l ∈ L \{(\omega_l,y_l)\}_{l\in\mathcal{L}} {( ω l , y l ) } l ∈ L be the weights and nodes of a quadrature for ∫ − ∞ ∞ f ( t ) d t \int_{-\infty}^{\infty}f(t)\,dt ∫ − ∞ ∞ f ( t ) d t , then we approximate ( − Δ ) s u (-\Delta)^s u ( − Δ ) s u using (2) as follows:
( − Δ ) s u ≈ c s ∑ l ∈ L ω l e s y l ( e y l − Δ ) − 1 ( − Δ u ) . (5) (-\Delta)^s u \approx c_s\sum_{l\in\mathcal{L}}\omega_l e^{sy_l}(e^{y_l}-\Delta)^{-1}(-\Delta u). \tag*{(5)} ( − Δ ) s u ≈ c s l ∈ L ∑ ω l e s y l ( e y l − Δ ) − 1 ( − Δ u ) . (5)
In order to simplify this expression, we follow Bonito et al. [8 ] and define the mapping u ↦ w l ( u ) u \mapsto w_l(u) u ↦ w l ( u ) by setting ( e y l − Δ ) − 1 ( − Δ u ) = w l ( u ) + u (e^{y_l}-\Delta)^{-1}(-\Delta u)=w_l(u)+u ( e y l − Δ ) − 1 ( − Δ u ) = w l ( u ) + u , i.e.,
− Δ u = ( e y l − Δ ) ( w l ( u ) + u ) = ( e y l − Δ ) w l ( u ) + e y l u − Δ u . -\Delta u=(e^{y_l}-\Delta)(w_l(u)+u)=(e^{y_l}-\Delta)w_l(u)+e^{y_l}u-\Delta u. − Δ u = ( e y l − Δ ) ( w l ( u ) + u ) = ( e y l − Δ ) w l ( u ) + e y l u − Δ u .
In other words, the mapping w l : H − 1 ( D ) → H 0 1 ( D ) w_l:H^{-1}(D)\to H_0^1(D) w l : H − 1 ( D ) → H 0 1 ( D ) is defined by setting
( e y l − Δ ) w l ( u ) = − e y l u . (6) (e^{y_l}-\Delta)w_l(u)=-e^{y_l}u. \tag*{(6)} ( e y l − Δ ) w l ( u ) = − e y l u . (6)
for any u ∈ H − 1 ( D ) u\in H^{-1}(D) u ∈ H − 1 ( D ) . Hence,
( − Δ ) s u ≈ c s ∑ l ∈ L ω l e s y l ( u + w l ( u ) ) . (-\Delta)^s u\approx c_s\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}(u+w_l(u)). ( − Δ ) s u ≈ c s l ∈ L ∑ ω l e s y l ( u + w l ( u )) .
This leads to the following approximation of (1) :
∂ t u + ∇ ⋅ f ( u ) + ∑ l ∈ L c s ω l e s y l ( w l ( u ) + u ) = g , u ∣ ∂ D = 0 , u ( ⋅ , 0 ) = u 0 ( ⋅ ) . (7) \partial_tu+\nabla\cdot f(u)+\sum_{l\in\mathcal{L}}c_s\omega_l e^{s y_l}(w_l(u)+u)=g,\qquad u|_{\partial D}=0,\qquad u(\cdot,0)=u_0(\cdot). \tag*{(7)} ∂ t u + ∇ ⋅ f ( u ) + l ∈ L ∑ c s ω l e s y l ( w l ( u ) + u ) = g , u ∣ ∂ D = 0 , u ( ⋅ , 0 ) = u 0 ( ⋅ ) . (7)
More details on the way the sinc quadrature is constructed and its accuracy are given in the next section.
Accuracy of the sinc quadrature
We fix k > 0 k>0 k > 0 , M , N ∈ N M,N\in\mathbb{N} M , N ∈ N positive. The sinc quadrature corresponds to the choice ω l : = k \omega_l:=k ω l := k and y l : = l k y_l:=lk y l := l k for l ∈ L : = { − M : N } l\in\mathcal{L}:=\{-M:N\} l ∈ L := { − M : N } .
We now recall results about the sinc quadrature approximation
k ∑ l = − M N ( e s y l ( e y l − Δ ) − 1 ( − Δ w ) , v ) L 2 ( D ) k\sum_{l=-M}^{N}\left(e^{s y_l}(e^{y_l}-\Delta)^{-1}(-\Delta w),v\right)_{L^2(D)} k l = − M ∑ N ( e s y l ( e y l − Δ ) − 1 ( − Δ w ) , v ) L 2 ( D )
of
∫ − ∞ ∞ ( e s y ( e y − Δ ) − 1 ( − Δ w ) , v ) L 2 ( D ) d y \int_{-\infty}^{\infty}\left(e^{s y}(e^y-\Delta)^{-1}(-\Delta w),v\right)_{L^2(D)}\,dy ∫ − ∞ ∞ ( e sy ( e y − Δ ) − 1 ( − Δ w ) , v ) L 2 ( D ) d y
for w ∈ H δ ( D ) w\in H^\delta(D) w ∈ H δ ( D ) with δ > s \delta>s δ > s and v ∈ H s ( D ) v\in H^s(D) v ∈ H s ( D ) . Here ( ⋅ , ⋅ ) L 2 ( D ) (\cdot,\cdot)_{L^2(D)} ( ⋅ , ⋅ ) L 2 ( D ) denotes the L 2 L^2 L 2 -scalar product and we let
q ( y ; w , v ) : = ( e s y ( e y − Δ ) − 1 ( − Δ w ) , v ) L 2 ( D ) . q(y;w,v):=\left(e^{s y}(e^y-\Delta)^{-1}(-\Delta w),v\right)_{L^2(D)}. q ( y ; w , v ) := ( e sy ( e y − Δ ) − 1 ( − Δ w ) , v ) L 2 ( D ) .
The sinc quadrature formula takes advantage of (i) the analyticity of z ↦ q ( z ; w , v ) z\mapsto q(z;w,v) z ↦ q ( z ; w , v ) on the strip
D r : = { z ∈ C : ∣ ℑ ( z ) ∣ < r } , D_r:=\{z\in\mathbb{C}:|\Im(z)|<r\}, D r := { z ∈ C : ∣ℑ ( z ) ∣ < r } ,
where 0 < r < π 0<r<\pi 0 < r < π is fixed from now on, and (ii) the decay property
∣ q ( z ; w , v ) ∣ ≤ C ( r ) { e s ℜ ( z ) ∥ w ∥ L 2 ( D ) ∥ v ∥ L 2 ( D ) , when ℜ ( z ) < 0 , e ( s − δ ) ℜ ( z ) / 2 ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) , when ℜ ( z ) ≥ 0 , |q(z;w,v)|\leq C(r)\begin{cases}
e^{s\Re(z)}\|w\|_{L^2(D)}\|v\|_{L^2(D)}, & \text{when }\Re(z)<0,\\
e^{(s-\delta)\Re(z)/2}\|w\|_{H^\delta(D)}\|v\|_{H^s(D)}, & \text{when }\Re(z)\geq0,
\end{cases} ∣ q ( z ; w , v ) ∣ ≤ C ( r ) { e s ℜ ( z ) ∥ w ∥ L 2 ( D ) ∥ v ∥ L 2 ( D ) , e ( s − δ ) ℜ ( z ) /2 ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) , when ℜ ( z ) < 0 , when ℜ ( z ) ≥ 0 ,
where C ( r ) C(r) C ( r ) is a constant only depending on r r r which tends to infinity as r → π r\to\pi r → π . From now on C ( r ) C(r) C ( r ) denotes such a generic constant which may change from one relation to the next.
From (ii) we deduce that
∫ − r r ∣ q ( x + i y ; w , v ) ∣ d y ≤ C ( r ) ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) \int_{-r}^{r}|q(x+iy;w,v)|\,dy\leq C(r)\|w\|_{H^\delta(D)}\|v\|_{H^s(D)} ∫ − r r ∣ q ( x + i y ; w , v ) ∣ d y ≤ C ( r ) ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D )
and
∫ − ∞ ∞ ( ∣ q ( x + i r ; w , v ) ∣ + ∣ q ( x − i r ; w , v ) ∣ ) d x ≤ C ( r ) min ( δ − s , s ) ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) . \int_{-\infty}^{\infty}\left(|q(x+ir;w,v)|+|q(x-ir;w,v)|\right)\,dx\leq\frac{C(r)}{\min(\delta-s,s)}\|w\|_{H^\delta(D)}\|v\|_{H^s(D)}. ∫ − ∞ ∞ ( ∣ q ( x + i r ; w , v ) ∣ + ∣ q ( x − i r ; w , v ) ∣ ) d x ≤ min ( δ − s , s ) C ( r ) ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) .
Whence, Theorem 2.20 in Lund and Bowers [29 ] applies to guarantee
∣ ∫ − ∞ ∞ q ( y ; w , v ) d y − k ∑ j = − ∞ ∞ q ( k j ; w , v ) ∣ ≤ C ( r ) e − 2 π r / k ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) . \left|\int_{-\infty}^{\infty} q(y;w,v)\,\mathrm{d}y-k\sum_{j=-\infty}^{\infty}q(kj;w,v)\right|\le C(r)e^{-2\pi r/k}\lVert w\rVert_{H^\delta(D)}\lVert v\rVert_{H^s(D)}. ∫ − ∞ ∞ q ( y ; w , v ) d y − k j = − ∞ ∑ ∞ q ( k j ; w , v ) ≤ C ( r ) e − 2 π r / k ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) .
In addition, the decay property (ii) yields estimates for the tails
k ∑ j = − ∞ − M − 1 ∣ q ( k j ; w , v ) ∣ ≤ C ( r ) s e − s M k ∥ w ∥ L 2 ( D ) ∥ v ∥ L 2 ( D ) k\sum_{j=-\infty}^{-M-1}|q(kj;w,v)|\le\frac{C(r)}{s}e^{-sMk}\lVert w\rVert_{L^2(D)}\lVert v\rVert_{L^2(D)} k j = − ∞ ∑ − M − 1 ∣ q ( k j ; w , v ) ∣ ≤ s C ( r ) e − s M k ∥ w ∥ L 2 ( D ) ∥ v ∥ L 2 ( D )
and
k ∑ j = N + 1 ∞ ∣ q ( k j ; w , v ) ∣ ≤ C ( r ) δ − s e ( s − δ ) N k / 2 ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) . k\sum_{j=N+1}^{\infty}|q(kj;w,v)|\le\frac{C(r)}{\delta-s}e^{(s-\delta)Nk/2}\lVert w\rVert_{H^\delta(D)}\lVert v\rVert_{H^s(D)}. k j = N + 1 ∑ ∞ ∣ q ( k j ; w , v ) ∣ ≤ δ − s C ( r ) e ( s − δ ) N k /2 ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) .
Gathering all these estimates, we find that upon selecting
M = ⌈ 2 π r s k 2 ⌉ and N = ⌈ 4 π r ( δ − s ) k 2 ⌉ (8) M=\left\lceil\frac{2\pi r}{sk^2}\right\rceil
\qquad\text{and}\qquad
N=\left\lceil\frac{4\pi r}{(\delta-s)k^2}\right\rceil \tag*{(8)} M = ⌈ s k 2 2 π r ⌉ and N = ⌈ ( δ − s ) k 2 4 π r ⌉ (8)
we arrive at
∣ ∫ − ∞ ∞ q ( y ; w , v ) d y − k ∑ j = − M N q ( k j ; w , v ) ∣ ≤ C ( r , s , δ ) e − 2 π r / k ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) . (9) \left|\int_{-\infty}^{\infty}q(y;w,v)\,\mathrm{d}y-k\sum_{j=-M}^{N}q(kj;w,v)\right|\le C(r,s,\delta)e^{-2\pi r/k}\lVert w\rVert_{H^\delta(D)}\lVert v\rVert_{H^s(D)}. \tag*{(9)} ∫ − ∞ ∞ q ( y ; w , v ) d y − k j = − M ∑ N q ( k j ; w , v ) ≤ C ( r , s , δ ) e − 2 π r / k ∥ w ∥ H δ ( D ) ∥ v ∥ H s ( D ) . (9)
Full approximation
In this section we describe the time and space approximation of (7) .
Finite element setting
Let ( T h ) h > 0 (\mathcal{T}_h)_{h>0} ( T h ) h > 0 be a shape-regular sequence of subdivisions made of simplices (the case of quadrilaterals/hexahedra can be treated similarly). For simplicity, we assume that D D D is a polyhedron that is exactly matched by the subdivision T h \mathcal{T}_h T h , h > 0 h>0 h > 0 . We denote by K ^ \widehat{K} K the reference simplex and by T K : K ^ → K T_K:\widehat{K}\to K T K : K → K the generic affine map from the reference element to the physical element.
We approximate the solution of (7) with piecewise linear continuous Lagrange finite elements generated from ( T h ) h > 0 (\mathcal{T}_h)_{h>0} ( T h ) h > 0 . The approximation space is defined by
P ( T h ) = { v ∈ C 0 ( D ; R ) ∩ H 0 1 ( D ) ∣ v ∣ K ∘ T K ∈ P 1 , ∀ K ∈ T h } . (10) P(\mathcal{T}_h)=\{v\in C^0(D;\mathbb{R})\cap H_0^1(D)\mid v|_K\circ T_K\in P_1,\ \forall K\in\mathcal{T}_h\}. \tag*{(10)} P ( T h ) = { v ∈ C 0 ( D ; R ) ∩ H 0 1 ( D ) ∣ v ∣ K ∘ T K ∈ P 1 , ∀ K ∈ T h } . (10)
where P 1 P_1 P 1 is the set of polynomials of total degree no larger than 1. We denote by V + \mathcal{V}^+ V + the index set of all the vertices of T h \mathcal{T}_h T h , by V ∂ ⊂ V + \mathcal{V}^{\partial}\subset\mathcal{V}^+ V ∂ ⊂ V + the index set of the vertices located on ∂ D \partial D ∂ D , and by V : = V + ∖ V ∂ \mathcal{V}:=\mathcal{V}^+\setminus\mathcal{V}^{\partial} V := V + ∖ V ∂ the index set of the interior vertices. Let { φ i } i ∈ V + \{\varphi_i\}_{i\in\mathcal{V}^+} { φ i } i ∈ V + be the Lagrange basis functions associated with all the vertices. Then { φ i } i ∈ V \{\varphi_i\}_{i\in\mathcal{V}} { φ i } i ∈ V is a basis of P ( T h ) P(\mathcal{T}_h) P ( T h ) , so that card ( V ) = dim P ( T h ) \operatorname{card}(\mathcal{V})=\dim P(\mathcal{T}_h) card ( V ) = dim P ( T h ) , and ∑ i ∈ V + φ i = 1 \sum_{i\in\mathcal{V}^+}\varphi_i=1 ∑ i ∈ V + φ i = 1 (partition of unity). For any i ∈ V i\in\mathcal{V} i ∈ V , we let I ( i ) : = { j ∈ V + ∣ φ i φ j ≢ 0 } \mathcal{I}(i):=\{j\in\mathcal{V}^+\mid\varphi_i\varphi_j\not\equiv0\} I ( i ) := { j ∈ V + ∣ φ i φ j ≡ 0 } be the stencil of i i i ; note that I ( i ) \mathcal{I}(i) I ( i ) contains the boundary vertices adjacent to i i i . The coefficients of the functions in P ( T h ) P(\mathcal{T}_h) P ( T h ) are extended by zero on V ∂ \mathcal{V}^{\partial} V ∂ : for u h = ∑ j ∈ V U j φ j ∈ P ( T h ) u_h=\sum_{j\in\mathcal{V}}U_j\varphi_j\in P(\mathcal{T}_h) u h = ∑ j ∈ V U j φ j ∈ P ( T h ) , we set U j : = 0 U_j:=0 U j := 0 for j ∈ V ∂ j\in\mathcal{V}^{\partial} j ∈ V ∂ , so that u h = ∑ j ∈ V + U j φ j u_h=\sum_{j\in\mathcal{V}^+}U_j\varphi_j u h = ∑ j ∈ V + U j φ j . This is the homogeneous Dirichlet condition in (1.1). With this convention, the sums over j ∈ I ( i ) j\in\mathcal{I}(i) j ∈ I ( i ) below include the boundary vertices adjacent to i i i , where the coefficients vanish.
For any i ∈ V i\in\mathcal{V} i ∈ V and j ∈ V + j\in\mathcal{V}^+ j ∈ V + we set a i j : = ∫ D ∇ φ i ⋅ ∇ φ j d x a_{ij}:=\int_D\nabla\varphi_i\cdot\nabla\varphi_j\,\mathrm{d}x a ij := ∫ D ∇ φ i ⋅ ∇ φ j d x . The partition of unity property implies ∑ j ∈ I ( i ) a i j = 0 \sum_{j\in\mathcal{I}(i)}a_{ij}=0 ∑ j ∈ I ( i ) a ij = 0 for all i ∈ V i\in\mathcal{V} i ∈ V .
We assume that each subdivision T h \mathcal{T}_h T h , h > 0 h>0 h > 0 , is such that
a i j : = ∫ D ∇ φ i ⋅ ∇ φ j d x ≤ 0 , ∀ i ∈ V , ∀ j ∈ I ( i ) \ { i } . (11) a_{ij}:=\int_D \nabla\varphi_i\cdot\nabla\varphi_j\,\mathrm{d}x\leq0,\qquad\forall i\in\mathcal{V},\ \forall j\in\mathcal{I}(i)\backslash\{i\}. \tag*{(11)} a ij := ∫ D ∇ φ i ⋅ ∇ φ j d x ≤ 0 , ∀ i ∈ V , ∀ j ∈ I ( i ) \ { i } . (11)
This condition is known to be satisfied when d = 1 d=1 d = 1 and for simplices when d = 2 d=2 d = 2 or d = 3 d=3 d = 3 under the so-called acute angle condition cf. e.g., Brandts et al. [9 ] , Brandts et al. [10 §5.2] , Xu and Zikatanov [36 Eq. (2.5)] . The purpose of this paper is not to relax this condition. Note that, since I ( i ) \mathcal{I}(i) I ( i ) contains the boundary vertices adjacent to i i i , (11) also applies to the pairs formed by an interior and a boundary vertex.
We henceforth denote by M L \mathcal{M}^{L} M L the lumped mass matrix, i.e., the diagonal matrix with entries
m i : = ∫ D φ i ( x ) d x , i ∈ V . (12) m_i:=\int_D\varphi_i(\boldsymbol{x})\,\mathrm{d}x,\qquad i\in\mathcal{V}. \tag*{(12)} m i := ∫ D φ i ( x ) d x , i ∈ V . (12)
Notice that φ i ( x ) ≥ 0 \varphi_i(\boldsymbol{x})\geq0 φ i ( x ) ≥ 0 for all x ∈ D \boldsymbol{x}\in D x ∈ D and so m i > 0 m_i>0 m i > 0 for all i ∈ V i\in\mathcal{V} i ∈ V .
We also introduce the following two quantities which will play an important role in the rest of the paper
c i j : = ∫ D φ i ∇ φ j d x , n i j : = c i j ∥ c i j ∥ ℓ 2 , i ∈ V , j ∈ V + , (13) c_{ij}:=\int_D\varphi_i\nabla\varphi_j\,\mathrm{d}x,\qquad n_{ij}:=\frac{c_{ij}}{\lVert c_{ij}\rVert_{\ell^2}},\qquad i\in\mathcal{V},\ j\in\mathcal{V}^{+}, \tag*{(13)} c ij := ∫ D φ i ∇ φ j d x , n ij := ∥ c ij ∥ ℓ 2 c ij , i ∈ V , j ∈ V + , (13)
where ∥ ⋅ ∥ ℓ 2 \lVert\cdot\rVert_{\ell^2} ∥ ⋅ ∥ ℓ 2 stands for the Euclidean norm in R d \mathbb{R}^d R d . The partition of unity implies ∑ j ∈ V + c i j = 0 \sum_{j\in\mathcal{V}^{+}}c_{ij}=0 ∑ j ∈ V + c ij = 0 , i.e., ∑ j ∈ I ( i ) c i j = 0 \sum_{j\in\mathcal{I}(i)}c_{ij}=0 ∑ j ∈ I ( i ) c ij = 0 for i ∈ V i\in\mathcal{V} i ∈ V .
Approximation of w l w_l w l
We start by approximating the mapping w l w_l w l defined in (6) for any l ∈ L l\in\mathcal{L} l ∈ L . Let u h = ∑ i ∈ V U i φ i u_h=\sum_{i\in\mathcal{V}}U_i\varphi_i u h = ∑ i ∈ V U i φ i be any function in P ( T h ) P(\mathcal{T}_h) P ( T h ) . We define the discrete mapping w h l : P ( T h ) → P ( T h ) w_h^l:P(\mathcal{T}_h)\to P(\mathcal{T}_h) w h l : P ( T h ) → P ( T h ) by setting w h l ( u h ) : = ∑ i ∈ V W i l ( u h ) φ i w_h^l(u_h):=\sum_{i\in\mathcal{V}}W_i^l(u_h)\varphi_i w h l ( u h ) := ∑ i ∈ V W i l ( u h ) φ i , where the coefficients { W i l ( u h ) } i ∈ V \{W_i^l(u_h)\}_{i\in\mathcal{V}} { W i l ( u h ) } i ∈ V solve the following linear system:
m i e y l W i l ( u h ) + ∑ j ∈ I ( i ) a i j W j l ( u h ) = − m i e y l U i , i ∈ V , and W i l ( u h ) = 0 , i ∈ V ∂ . (14) m_i\mathrm{e}^{y_l}W_i^l(u_h)+\sum_{j\in\mathcal{I}(i)}a_{ij}W_j^l(u_h)=-m_i\mathrm{e}^{y_l}U_i,\qquad i\in\mathcal{V},\qquad\text{and}\qquad W_i^l(u_h)=0,\qquad i\in\mathcal{V}^{\partial}. \tag*{(14)} m i e y l W i l ( u h ) + j ∈ I ( i ) ∑ a ij W j l ( u h ) = − m i e y l U i , i ∈ V , and W i l ( u h ) = 0 , i ∈ V ∂ . (14)
Recall that a i j a_{ij} a ij are the coefficients of the stiffness matrix defined in (11) and m i m_i m i those of the lumped mass matrix. When there is no ambiguity, we write W j l W_j^l W j l instead of W j l ( u h ) W_j^l(u_h) W j l ( u h ) .
We have the following maximum principle.
Lemma 3.1 (Maximum principle). Let T h \mathcal{T}_h T h be such that (11) holds. Let u h = ∑ i ∈ V + U i φ i ∈ P ( T h ) u_h=\sum_{i\in\mathcal{V}^{+}}U_i\varphi_i\in P(\mathcal{T}_h) u h = ∑ i ∈ V + U i φ i ∈ P ( T h ) with U i = 0 U_i=0 U i = 0 when i ∈ V ∂ i\in\mathcal{V}^{\partial} i ∈ V ∂ and w h l = ∑ i ∈ V + W i l ( u h ) φ i ∈ P ( T h ) w_h^l=\sum_{i\in\mathcal{V}^{+}}W_i^l(u_h)\varphi_i\in P(\mathcal{T}_h) w h l = ∑ i ∈ V + W i l ( u h ) φ i ∈ P ( T h ) . Then we have for any l ∈ L l\in\mathcal{L} l ∈ L
− max j ∈ V + U j ≤ min j ∈ V + W j l ( u h ) ≤ max j ∈ V + W j l ( u h ) ≤ − min j ∈ V + U j . (15) -\max_{j\in\mathcal{V}^{+}}U_j\leq\min_{j\in\mathcal{V}^{+}}W_j^l(u_h)\leq\max_{j\in\mathcal{V}^{+}}W_j^l(u_h)\leq-\min_{j\in\mathcal{V}^{+}}U_j. \tag*{(15)} − j ∈ V + max U j ≤ j ∈ V + min W j l ( u h ) ≤ j ∈ V + max W j l ( u h ) ≤ − j ∈ V + min U j . (15)
Proof. Using that ∑ j ∈ I ( i ) a i j = 0 \sum_{j\in\mathcal{I}(i)}a_{ij}=0 ∑ j ∈ I ( i ) a ij = 0 , which follows from the partition of unity property satisfied by { φ i } i ∈ V + \{\varphi_i\}_{i\in\mathcal{V}^{+}} { φ i } i ∈ V + , we deduce
m i e y l W i l = − m i e y l U i + ∑ j ∈ I ( i ) \ { i } − a i j ( W j l − W i l ) , i ∈ V . m_i\mathrm{e}^{y_l}W_i^l=-m_i\mathrm{e}^{y_l}U_i+\sum_{j\in\mathcal{I}(i)\backslash\{i\}}-a_{ij}(W_j^l-W_i^l),\qquad i\in\mathcal{V}. m i e y l W i l = − m i e y l U i + j ∈ I ( i ) \ { i } ∑ − a ij ( W j l − W i l ) , i ∈ V .
Let i ∈ V + i\in\mathcal{V}^{+} i ∈ V + be the index where W j l W_j^l W j l is maximum over j ∈ V + j\in\mathcal{V}^{+} j ∈ V + , i.e., W i l = max j ∈ V + W j l W_i^l=\max_{j\in\mathcal{V}^{+}}W_j^l W i l = max j ∈ V + W j l . If i ∈ V ∂ i\in\mathcal{V}^{\partial} i ∈ V ∂ , then max j ∈ V + W j l = 0 ≤ − min j ∈ V + U j \max_{j\in\mathcal{V}^{+}}W_j^l=0\leq-\min_{j\in\mathcal{V}^{+}}U_j max j ∈ V + W j l = 0 ≤ − min j ∈ V + U j , since U j = 0 U_j=0 U j = 0 for j ∈ V ∂ j\in\mathcal{V}^{\partial} j ∈ V ∂ . Instead, for i ∈ V i\in\mathcal{V} i ∈ V , since − a i j ≥ 0 -a_{ij}\geq0 − a ij ≥ 0 for all j ∈ I ( i ) \ { i } j\in\mathcal{I}(i)\backslash\{i\} j ∈ I ( i ) \ { i } thanks to (11) , we have
max j ∈ V + m i W j l = m i W i l ≤ − m i U i . \max_{j\in\mathcal{V}^{+}}m_iW_j^l=m_iW_i^l\leq-m_iU_i. j ∈ V + max m i W j l = m i W i l ≤ − m i U i .
This means max j ∈ V + W j l ≤ − min j ∈ V + U j \max_{j\in\mathcal{V}^{+}} W_j^l \leq-\min_{j\in\mathcal{V}^{+}} U_j max j ∈ V + W j l ≤ − min j ∈ V + U j . Similarly, let i ∈ V + i\in\mathcal{V}^{+} i ∈ V + be the index where W j l W_j^l W j l is minimum over j ∈ V + j\in\mathcal{V}^{+} j ∈ V + , i.e., W i l = min j ∈ V + W j l W_i^l=\min_{j\in\mathcal{V}^{+}}W_j^l W i l = min j ∈ V + W j l . If i ∈ V ∂ i\in\mathcal{V}^{\partial} i ∈ V ∂ , then min j ∈ V + W j l = 0 ≥ − max j ∈ V + U j \min_{j\in\mathcal{V}^{+}}W_j^l=0\geq-\max_{j\in\mathcal{V}^{+}}U_j min j ∈ V + W j l = 0 ≥ − max j ∈ V + U j . Instead, when i ∈ V i\in\mathcal{V} i ∈ V we have
min j ∈ V + m i W j l = m i W i l ≥ − m i U i . \min_{j\in\mathcal{V}^{+}}m_iW_j^l=m_iW_i^l\geq-m_iU_i. j ∈ V + min m i W j l = m i W i l ≥ − m i U i .
Altogether this implies min j ∈ V + W j l ≥ − max j ∈ V + U j \min_{j\in\mathcal{V}^{+}}W_j^l\geq-\max_{j\in\mathcal{V}^{+}}U_j min j ∈ V + W j l ≥ − max j ∈ V + U j .
We will also use the following local representation of W i l + U i W_i^l+U_i W i l + U i , which involves the quantities Λ i : = a i i / m i > 0 \Lambda_i:=a_{ii}/m_i>0 Λ i := a ii / m i > 0 , i ∈ V i\in\mathcal{V} i ∈ V .
Lemma 3.2 (Local form of the fractional term). Let T h T_h T h be such that (11) holds. Let u h = ∑ i ∈ V U i φ i ∈ P ( T h ) u_h=\sum_{i\in\mathcal{V}}U_i\varphi_i\in P(T_h) u h = ∑ i ∈ V U i φ i ∈ P ( T h ) and W j l : = W j l ( u h ) W_j^l:=W_j^l(u_h) W j l := W j l ( u h ) . Then, for all i ∈ V i\in\mathcal{V} i ∈ V and l ∈ L l\in\mathcal{L} l ∈ L ,
− ( W i l + U i ) = ∑ j ∈ I ( i ) ∖ { i } θ i j l ( − W j l − U i ) , θ i j l : = − a i j / m i e y l + Λ i ≥ 0 , (16) -(W_i^l+U_i)=\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\theta_{ij}^l(-W_j^l-U_i),\qquad\theta_{ij}^l:=\frac{-a_{ij}/m_i}{e^{y_l}+\Lambda_i}\geq0, \tag*{(16)} − ( W i l + U i ) = j ∈ I ( i ) ∖ { i } ∑ θ ij l ( − W j l − U i ) , θ ij l := e y l + Λ i − a ij / m i ≥ 0 , (16)
and ∑ j ∈ I ( i ) ∖ { i } θ i j l = Λ i e y l + Λ i ∈ ( 0 , 1 ) \sum_{j\in\mathcal{I}(i)\setminus\{i\}}\theta_{ij}^l=\frac{\Lambda_i}{e^{y_l}+\Lambda_i}\in(0,1) ∑ j ∈ I ( i ) ∖ { i } θ ij l = e y l + Λ i Λ i ∈ ( 0 , 1 ) .
Proof. Dividing the i i i -th equation defining W l W^l W l (14) by m i e y l m_i e^{y_l} m i e y l and using ∑ j ∈ I ( i ) a i j = 0 \sum_{j\in\mathcal{I}(i)}a_{ij}=0 ∑ j ∈ I ( i ) a ij = 0 , we obtain
W i l + U i = − e − y l m i ∑ j ∈ I ( i ) a i j W j l = e − y l m i ∑ j ∈ I ( i ) ∖ { i } ( − a i j ) ( W j l − W i l ) . W_i^l+U_i=-\frac{e^{-y_l}}{m_i}\sum_{j\in\mathcal{I}(i)}a_{ij}W_j^l=\frac{e^{-y_l}}{m_i}\sum_{j\in\mathcal{I}(i)\setminus\{i\}}(-a_{ij})(W_j^l-W_i^l). W i l + U i = − m i e − y l j ∈ I ( i ) ∑ a ij W j l = m i e − y l j ∈ I ( i ) ∖ { i } ∑ ( − a ij ) ( W j l − W i l ) .
We write W j l − W i l = ( W j l + U i ) − ( W i l + U i ) W_j^l-W_i^l=(W_j^l+U_i)-(W_i^l+U_i) W j l − W i l = ( W j l + U i ) − ( W i l + U i ) and use ∑ j ∈ I ( i ) ∖ { i } ( − a i j ) = a i i \sum_{j\in\mathcal{I}(i)\setminus\{i\}}(-a_{ij})=a_{ii} ∑ j ∈ I ( i ) ∖ { i } ( − a ij ) = a ii to get
( 1 + e − y l Λ i ) ( W i l + U i ) = e − y l m i ∑ j ∈ I ( i ) ∖ { i } ( − a i j ) ( W j l + U i ) . (1+e^{-y_l}\Lambda_i)(W_i^l+U_i)=\frac{e^{-y_l}}{m_i}\sum_{j\in\mathcal{I}(i)\setminus\{i\}}(-a_{ij})(W_j^l+U_i). ( 1 + e − y l Λ i ) ( W i l + U i ) = m i e − y l j ∈ I ( i ) ∖ { i } ∑ ( − a ij ) ( W j l + U i ) .
which is (16) . The weights θ i j l \theta_{ij}^l θ ij l are nonnegative owing to (11) , and their sum is a i i / m i e y l + Λ i \frac{a_{ii}/m_i}{e^{y_l}+\Lambda_i} e y l + Λ i a ii / m i .
Observe that if i i i is an index corresponding to a vertex adjacent to one in ∂ D \partial D ∂ D , the sum in (16) contains the boundary vertices in j ∈ I ( i ) ∩ V ∂ j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial} j ∈ I ( i ) ∩ V ∂ , for which − W j l = 0 -W_j^l=0 − W j l = 0 . The homogeneous boundary value thus enters in (16) with the weights θ i j l \theta_{ij}^l θ ij l , j ∈ I ( i ) ∩ V ∂ j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial} j ∈ I ( i ) ∩ V ∂ .
Full approximation, low-order scheme
We now construct an approximation in time and space of (7) . The space approximation is performed using the finite element space P ( T h ) P(T_h) P ( T h ) while the forward Euler method is used for the time discretization.
Let τ > 0 \tau>0 τ > 0 and set t n : = n τ t_n:=n\tau t n := n τ , n ∈ N n\in\mathbb{N} n ∈ N . Let G n : = ∑ i ∈ V G i n φ i G^n:=\sum_{i\in\mathcal{V}}G_i^n\varphi_i G n := ∑ i ∈ V G i n φ i be an approximation of g g g at time t = t n t=t_n t = t n and u h n : = ∑ i ∈ V U i n φ i u_h^n:=\sum_{i\in\mathcal{V}}U_i^n\varphi_i u h n := ∑ i ∈ V U i n φ i be a (possibly high-order) approximation of u u u at time t n t_n t n . We define the low-order approximation of u u u at time t n + 1 t_{n+1} t n + 1 by u h L , n + 1 = ∑ i ∈ V U i L , n + 1 φ i u_h^{L,n+1}=\sum_{i\in\mathcal{V}}U_i^{L,n+1}\varphi_i u h L , n + 1 = ∑ i ∈ V U i L , n + 1 φ i , where the coefficient U i L , n + 1 U_i^{L,n+1} U i L , n + 1 , i ∈ V i\in\mathcal{V} i ∈ V , is given by the relation
m i U i L , n + 1 − U i n τ + ∑ j ∈ I ( i ) c i j ⋅ f ( U j n ) + c s m i ∑ l ∈ L ω l e s y l ( W i l , n ( u h n ) + U i n ) = m i G i n + ∑ j ∈ I ( i ) ∖ { i } d i j L , n ( U j n − U i n ) . (17) m_i\frac{U_i^{L,n+1}-U_i^n}{\tau}+\sum_{j\in\mathcal{I}(i)}\boldsymbol{c}_{ij}\cdot\boldsymbol{f}(U_j^n)+c_s m_i\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}(W_i^{l,n}(u_h^n)+U_i^n)
=m_iG_i^n+\sum_{j\in\mathcal{I}(i)\setminus\{i\}}d_{ij}^{L,n}(U_j^n-U_i^n). \tag*{(17)} m i τ U i L , n + 1 − U i n + j ∈ I ( i ) ∑ c ij ⋅ f ( U j n ) + c s m i l ∈ L ∑ ω l e s y l ( W i l , n ( u h n ) + U i n ) = m i G i n + j ∈ I ( i ) ∖ { i } ∑ d ij L , n ( U j n − U i n ) . (17)
The artificial viscosity coefficients d i j L , n d_{ij}^{L,n} d ij L , n originally proposed in [21 ] are added to deal with the instabilities inherent to the approximations of advection systems. They are given for i ∈ V i\in\mathcal{V} i ∈ V , j ∈ I ( i ) ∖ { i } j\in\mathcal{I}(i)\setminus\{i\} j ∈ I ( i ) ∖ { i } by
d i j L , n : = max ( λ max ( U i n , U j n , n i j ) ∥ c i j ∥ ℓ 2 , λ max ( U j n , U i n , n j i ) ∥ c j i ∥ ℓ 2 ) . (18) d_{ij}^{L,n}:=\max\left(\lambda_{\max}(U_i^n,U_j^n,\boldsymbol{n}_{ij})\lVert\boldsymbol{c}_{ij}\rVert_{\ell^2},\lambda_{\max}(U_j^n,U_i^n,\boldsymbol{n}_{ji})\lVert\boldsymbol{c}_{ji}\rVert_{\ell^2}\right). \tag*{(18)} d ij L , n := max ( λ m a x ( U i n , U j n , n ij ) ∥ c ij ∥ ℓ 2 , λ m a x ( U j n , U i n , n j i ) ∥ c j i ∥ ℓ 2 ) . (18)
where λ max ( U i n , U j n , n i j ) \lambda_{\max}(U_i^n,U_j^n,n_{ij}) λ m a x ( U i n , U j n , n ij ) is an upper bound of the maximum wave speed in the 1D Riemann problem ∂ t v + ∂ x ( f ( v ) ⋅ n i j ) = 0 \partial_t v+\partial_x(f(v)\cdot n_{ij})=0 ∂ t v + ∂ x ( f ( v ) ⋅ n ij ) = 0 , with left and right data U i n U_i^n U i n and U j n U_j^n U j n respectively. In view of the discussion in Section 79.2 in Ern and Guermond [16 ] , for instance for convex fluxes f f f , we have
λ max ( U i n , U j n , n i j ) = { ∣ f ( U i n ) ⋅ n i j − f ( U j n ) ⋅ n i j ∣ ∣ U i n − U j n ∣ if U i n > U j n , max ( ∣ f ′ ( U i n ) ⋅ n i j ∣ , ∣ f ′ ( U j n ) ⋅ n i j ∣ ) otherwise. \lambda_{\max}(U_i^n,U_j^n,n_{ij})=
\begin{cases}
\dfrac{\left|f(U_i^n)\cdot n_{ij}-f(U_j^n)\cdot n_{ij}\right|}{\left|U_i^n-U_j^n\right|} & \text{if } U_i^n>U_j^n,\\
\max\left(\left|f'(U_i^n)\cdot n_{ij}\right|,\left|f'(U_j^n)\cdot n_{ij}\right|\right) & \text{otherwise.}
\end{cases} λ m a x ( U i n , U j n , n ij ) = ⎩ ⎨ ⎧ U i n − U j n f ( U i n ) ⋅ n ij − f ( U j n ) ⋅ n ij max ( ∣ f ′ ( U i n ) ⋅ n ij ∣ , f ′ ( U j n ) ⋅ n ij ) if U i n > U j n , otherwise.
For any f f f (not necessarily convex), this choice for λ max \lambda_{\max} λ m a x guarantees that
U ‾ i j n : = 1 2 ( U j n + U i n ) − ( f ( U j n ) − f ( U i n ) ) ⋅ c i j 2 d i j L , n \overline{U}_{ij}^n:=\frac{1}{2}(U_j^n+U_i^n)-\left(f(U_j^n)-f(U_i^n)\right)\cdot\frac{c_{ij}}{2d_{ij}^{L,n}} U ij n := 2 1 ( U j n + U i n ) − ( f ( U j n ) − f ( U i n ) ) ⋅ 2 d ij L , n c ij
satisfies
U ‾ i j n ∈ [ min ( U i n , U j n ) , max ( U i n , U j n ) ] ; (19) \overline{U}_{ij}^n\in\left[\min(U_i^n,U_j^n),\max(U_i^n,U_j^n)\right]; \tag*{(19)} U ij n ∈ [ min ( U i n , U j n ) , max ( U i n , U j n ) ] ; (19)
see [16 Lemma 79.18, Remark 79.19] .
When i = j i=j i = j , we set d i i L , n : = − ∑ j ∈ I ( i ) ∖ { i } d i j L , n d_{ii}^{L,n}:=-\sum_{j\in\mathcal{I}(i)\setminus\{i\}}d_{ij}^{L,n} d ii L , n := − ∑ j ∈ I ( i ) ∖ { i } d ij L , n , which implies the following important properties of the artificial viscosity coefficients:
d i j L , n = d j i L , n , d i j L , n ≥ 0 ∀ i , j , i ≠ j , and ∑ j ∈ I ( i ) d i j L , n = 0 ∀ i . d_{ij}^{L,n}=d_{ji}^{L,n},\qquad d_{ij}^{L,n}\ge0\quad\forall i,j,\ i\ne j,\qquad\text{and}\qquad\sum_{j\in\mathcal{I}(i)}d_{ij}^{L,n}=0\quad\forall i. d ij L , n = d j i L , n , d ij L , n ≥ 0 ∀ i , j , i = j , and j ∈ I ( i ) ∑ d ij L , n = 0 ∀ i .
Recall that from (4) , Λ i s = c s ∫ − ∞ ∞ e s y ( e y + Λ i ) − 1 Λ i d y \Lambda_i^s=c_s\int_{-\infty}^{\infty}e^{sy}(e^y+\Lambda_i)^{-1}\Lambda_i\,\mathrm{d}y Λ i s = c s ∫ − ∞ ∞ e sy ( e y + Λ i ) − 1 Λ i d y , where Λ i = a i i / m i \Lambda_i=a_{ii}/m_i Λ i = a ii / m i is introduced before Lemma 3.2. This motivates the notation for i ∈ V i\in\mathcal{V} i ∈ V
Λ i , s i n c s : = c s ∑ l ∈ L ω l e s y l ( e y l + Λ i ) − 1 Λ i , (20) \Lambda_{i,\mathrm{sinc}}^s:=c_s\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}(e^{y_l}+\Lambda_i)^{-1}\Lambda_i, \tag*{(20)} Λ i , sinc s := c s l ∈ L ∑ ω l e s y l ( e y l + Λ i ) − 1 Λ i , (20)
to denote the sinc approximation of Λ i s \Lambda_i^s Λ i s .
The next result guarantees a local discrete maximum principle satisfied by the coefficients U i L , n + 1 U_i^{L,n+1} U i L , n + 1 under the time-step restriction
τ : = C F L max i ∈ V ( 1 m i ∑ j ∈ I ( i ) ∖ { i } d i j L , n + Λ i , s i n c s ) , (21) \tau:=\frac{\mathrm{CFL}}{\max_{i\in\mathcal{V}}\left(\frac{1}{m_i}\sum_{j\in\mathcal{I}(i)\setminus\{i\}}d_{ij}^{L,n}+\Lambda_{i,\mathrm{sinc}}^s\right)}, \tag*{(21)} τ := max i ∈ V ( m i 1 ∑ j ∈ I ( i ) ∖ { i } d ij L , n + Λ i , sinc s ) CFL , (21)
where C F L > 0 \mathrm{CFL}>0 CFL > 0 is a constant. We postpone to § 3.4 a discussion of the relation between (21) and the standard CFL condition for the transport equation.
We are now in a position to prove a discrete maximum property. Before proceeding, we recall the convention that U j n = W j L , n = 0 U_j^n=W_j^{L,n}=0 U j n = W j L , n = 0 whenever j ∈ V ∂ j\in\mathcal{V}^{\partial} j ∈ V ∂ indexes a boundary vertex. These boundary values enter the discrete maximum estimates stated in the following result.
Theorem 3.3 (Discrete Maximum Principle). Let u h n = ∑ i ∈ V U i n φ i u_h^n=\sum_{i\in\mathcal{V}}U_i^n\varphi_i u h n = ∑ i ∈ V U i n φ i be an approximation at time t n t_n t n and U i L , n + 1 U_i^{L,n+1} U i L , n + 1 be as in (17) . Assume that the time-step τ \tau τ satisfies (21) for some CFL number 0 < C F L ≤ 1 2 0<\mathrm{CFL}\le\frac{1}{2} 0 < CFL ≤ 2 1 . Then for any i ∈ V i\in\mathcal{V} i ∈ V we have
U i n , min + τ G i n ≤ U i L , n + 1 ≤ U i n , max + τ G i n , (22) U_i^{n,\min}+\tau G_i^n\le U_i^{L,n+1}\le U_i^{n,\max}+\tau G_i^n, \tag*{(22)} U i n , m i n + τ G i n ≤ U i L , n + 1 ≤ U i n , m a x + τ G i n , (22) where
U i n , min : = min ( min j ∈ I ( i ) U j n , min l ∈ L min j ∈ I ( i ) ∖ { i } ( − W j l , n ) ) , U_i^{n,\min}:=\min\left(\min_{j\in\mathcal{I}(i)}U_j^n,\min_{l\in\mathcal{L}}\min_{j\in\mathcal{I}(i)\setminus\{i\}}\left(-W_j^{l,n}\right)\right), U i n , m i n := min ( j ∈ I ( i ) min U j n , l ∈ L min j ∈ I ( i ) ∖ { i } min ( − W j l , n ) ) , and
U i n , max : = max ( max j ∈ I ( i ) U j n , max l ∈ L max j ∈ I ( i ) ∖ { i } ( − W j l , n ) ) . (23) U_i^{n,\max}:=\max\left(\max_{j\in\mathcal{I}(i)}U_j^n,\max_{l\in\mathcal{L}}\max_{j\in\mathcal{I}(i)\setminus\{i\}}\left(-W_j^{l,n}\right)\right). \tag*{(23)} U i n , m a x := max ( j ∈ I ( i ) max U j n , l ∈ L max j ∈ I ( i ) ∖ { i } max ( − W j l , n ) ) . (23) Here W j l , n : = W j l ( u h n ) W_j^{l,n}:=W_j^l(u_h^n) W j l , n := W j l ( u h n ) .
Proof. Let i ∈ V i \in\mathcal{V} i ∈ V . Using ∑ j ∈ I ( i ) c i j = 0 \sum_{j\in\mathcal{I}(i)}c_{ij}=0 ∑ j ∈ I ( i ) c ij = 0 , we have
U i L , n + 1 = U i n + τ m i ∑ j ∈ I ( i ) ∖ { i } ( − c i j ( f ( U j n ) − f ( U i n ) ) + d i j L , n ( U j n − U i n ) ) − τ c s ∑ l ∈ L ω l e s y l ( W i l , n + U i n ) + τ G i n . U_i^{L,n+1}=U_i^n+\frac{\tau}{m_i}\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\left(-c_{ij}\bigl(f(U_j^n)-f(U_i^n)\bigr)+d_{ij}^{L,n}(U_j^n-U_i^n)\right)-\tau c_s\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}\left(W_i^{l,n}+U_i^n\right)+\tau G_i^n. U i L , n + 1 = U i n + m i τ j ∈ I ( i ) ∖ { i } ∑ ( − c ij ( f ( U j n ) − f ( U i n ) ) + d ij L , n ( U j n − U i n ) ) − τ c s l ∈ L ∑ ω l e s y l ( W i l , n + U i n ) + τ G i n .
In view of (19) , we find that
U i L , n + 1 = U i n ( 1 − 2 τ m i ∑ j ∈ I ( i ) ∖ { i } d i j L , n ) + ∑ j ∈ I ( i ) ∖ { i } τ m i 2 d i j L , n U ‾ i j n − τ c s ∑ l ∈ L ω l e s y l ( W i l , n + U i n ) + τ G i n . (24) \begin{aligned}
U_i^{L,n+1}={}&U_i^n\left(1-2\frac{\tau}{m_i}\sum_{j\in\mathcal{I}(i)\setminus\{i\}}d_{ij}^{L,n}\right)+\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\frac{\tau}{m_i}2d_{ij}^{L,n}\overline{U}_{ij}^n\\
&-\tau c_s\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}\left(W_i^{l,n}+U_i^n\right)+\tau G_i^n. \tag*{(24)}
\end{aligned} U i L , n + 1 = U i n 1 − 2 m i τ j ∈ I ( i ) ∖ { i } ∑ d ij L , n + j ∈ I ( i ) ∖ { i } ∑ m i τ 2 d ij L , n U ij n − τ c s l ∈ L ∑ ω l e s y l ( W i l , n + U i n ) + τ G i n . (24)
Owing to Lemma 3.2 and the definition (20) of Λ i , s i n c s \Lambda_{i,\mathrm{sinc}}^s Λ i , sinc s , the contribution of the fractional Laplacian in (24) can be rewritten as
− τ c s ∑ l ∈ L ω l e s y l ( W i l , n + U i n ) = ∑ l ∈ L ∑ j ∈ I ( i ) ∖ { i } τ c s ω l e s y l θ i j l ( − W j l , n ) − τ Λ i , s i n c s U i n . -\tau c_s\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}\left(W_i^{l,n}+U_i^n\right)=\sum_{l\in\mathcal{L}}\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\tau c_s\omega_l e^{s y_l}\theta_{ij}^l\left(-W_j^{l,n}\right)-\tau\Lambda_{i,\mathrm{sinc}}^sU_i^n. − τ c s l ∈ L ∑ ω l e s y l ( W i l , n + U i n ) = l ∈ L ∑ j ∈ I ( i ) ∖ { i } ∑ τ c s ω l e s y l θ ij l ( − W j l , n ) − τ Λ i , sinc s U i n .
Inserting this identity into (24) , we obtain
U i L , n + 1 − τ G i n = ( 1 − 2 τ m i ∑ j d i j L , n − τ Λ i , s i n c s ) U i n + 2 τ m i ∑ j d i j L , n U ‾ i j n + ∑ l ∈ L ∑ j τ c s ω l e s y l θ i j l ( − W j l , n ) , U_i^{L,n+1}-\tau G_i^n=\left(1-\frac{2\tau}{m_i}\sum_jd_{ij}^{L,n}-\tau\Lambda_{i,\mathrm{sinc}}^s\right)U_i^n+\frac{2\tau}{m_i}\sum_jd_{ij}^{L,n}\overline{U}_{ij}^n+\sum_{l\in\mathcal{L}}\sum_j\tau c_s\omega_l e^{s y_l}\theta_{ij}^l\left(-W_j^{l,n}\right), U i L , n + 1 − τ G i n = ( 1 − m i 2 τ j ∑ d ij L , n − τ Λ i , sinc s ) U i n + m i 2 τ j ∑ d ij L , n U ij n + l ∈ L ∑ j ∑ τ c s ω l e s y l θ ij l ( − W j l , n ) ,
where all the sums over j j j run over I ( i ) ∖ { i } \mathcal{I}(i)\setminus\{i\} I ( i ) ∖ { i } . From Lemma 3.2, we find that
∑ j ∈ I ( i ) ∖ { i } θ i j l = Λ i e s y l + Λ i \sum_{j\in\mathcal{I}(i)\setminus\{i\}}\theta_{ij}^l=\frac{\Lambda_i}{e^{s y_l}+\Lambda_i} j ∈ I ( i ) ∖ { i } ∑ θ ij l = e s y l + Λ i Λ i
and thus the coefficients in front of U i n U_i^n U i n , U ‾ i j n \overline{U}_{ij}^n U ij n , and − W j l , n -W_j^{l,n} − W j l , n sum to one. They are nonnegative because (21) with C F L ≤ 1 2 \mathrm{CFL}\leq\frac{1}{2} CFL ≤ 2 1 implies 2 τ m i ∑ j ≠ i d i j L , n + τ Λ i , s i n c s ≤ 2 C F L ≤ 1 \frac{2\tau}{m_i}\sum_{j\neq i}d_{ij}^{L,n}+\tau\Lambda_{i,\mathrm{sinc}}^s\leq2\mathrm{CFL}\leq1 m i 2 τ ∑ j = i d ij L , n + τ Λ i , sinc s ≤ 2 CFL ≤ 1 . Hence U i L , n + 1 − τ G i n U_i^{L,n+1}-\tau G_i^n U i L , n + 1 − τ G i n is a convex combination of U i n U_i^n U i n , U ‾ i j n \overline{U}_{ij}^n U ij n , − W j l , n -W_j^{l,n} − W j l , n . Since U ‾ i j n ∈ [ min ( U i n , U j n ) , max ( U i n , U j n ) ] \overline{U}_{ij}^n\in[\min(U_i^n,U_j^n),\max(U_i^n,U_j^n)] U ij n ∈ [ min ( U i n , U j n ) , max ( U i n , U j n )] according to (19) , we obtain (22) for U i n , min U_i^{n,\min} U i n , m i n , U i n , max U_i^{n,\max} U i n , m a x given by (23) . □
Remark 3.4 (Local and global bounds). The limits (23) involve, in addition to the values of u h n u_h^n u h n on the stencil I ( i ) \mathcal{I}(i) I ( i ) , the neighboring nodal values − W j l , n -W_j^{l,n} − W j l , n , j ∈ I ( i ) ∖ { i } j\in\mathcal{I}(i)\setminus\{i\} j ∈ I ( i ) ∖ { i } , l ∈ L l\in\mathcal{L} l ∈ L , which are weighted averages of u h n u_h^n u h n over the domain. This is unavoidable since ( − Δ ) s (-\Delta)^s ( − Δ ) s is nonlocal and a maximum principle involving only the stencil I ( i ) \mathcal{I}(i) I ( i ) cannot hold in general. According to Lemma 3.1 , we have min j ∈ V + U j n ≤ U i n , min ≤ U i n , max ≤ max j ∈ V + U j n \min_{j\in\mathcal{V}^+}U_j^n\leq U_i^{n,\min}\leq U_i^{n,\max}\leq\max_{j\in\mathcal{V}^+}U_j^n min j ∈ V + U j n ≤ U i n , m i n ≤ U i n , m a x ≤ max j ∈ V + U j n , so that (22) implies the global maximum principle min j ∈ V + U j n + τ G i n ≤ U i L , n + 1 ≤ max j ∈ V + U j n + τ G i n \min_{j\in\mathcal{V}^+}U_j^n+\tau G_i^n\leq U_i^{L,n+1}\leq\max_{j\in\mathcal{V}^+}U_j^n+\tau G_i^n min j ∈ V + U j n + τ G i n ≤ U i L , n + 1 ≤ max j ∈ V + U j n + τ G i n ; recall that U j n = 0 U_j^n=0 U j n = 0 for j ∈ V ∂ j\in\mathcal{V}^{\partial} j ∈ V ∂ . The relations (23) are available at a negligible extra cost since W j l , n W_j^{l,n} W j l , n are needed anyway to evaluate (17) .
Estimation of the time-step restriction
For the local discrete maximum principle to hold (Theorem 3.3 ), the time-step τ \tau τ is required to satisfy (21) . Note that the first term in the denominator is the advection contribution, and the second term, Λ i , s i n c s \Lambda_{i,\mathrm{sinc}}^s Λ i , sinc s , corresponds to the fractional diffusion contribution. The following lemma shows that the sinc quadrature Λ i , s i n c s \Lambda_{i,\mathrm{sinc}}^s Λ i , sinc s of Λ i s \Lambda_i^s Λ i s , see (20) , behaves like Λ i s \Lambda_i^s Λ i s independently of the sinc quadrature parameters M M M and N N N .
Lemma 3.5 (Bound on Λ i , s i n c s \Lambda_{i,\mathrm{sinc}}^{s} Λ i , sinc s ). For all M , N ∈ N M,N\in\mathbb{N} M , N ∈ N and k > 0 k>0 k > 0 , if the sinc quadrature weight and points are given by ω l = k \omega_l=k ω l = k and y l = l k y_l=lk y l = l k , l = − M , … , N l=-M,\ldots,N l = − M , … , N , then
Λ i , s i n c s ≤ c s ( 1 s ( 1 − s ) + 2 k ) Λ i s ≤ 4 + 2 k π Λ i s , i ∈ V . \Lambda_{i,\mathrm{sinc}}^{s}\leq c_s\left(\frac{1}{s(1-s)}+2k\right)\Lambda_i^s\leq\frac{4+2k}{\pi}\Lambda_i^s,\qquad i\in\mathcal{V}. Λ i , sinc s ≤ c s ( s ( 1 − s ) 1 + 2 k ) Λ i s ≤ π 4 + 2 k Λ i s , i ∈ V . Proof. We use Λ e y + Λ ≤ min ( 1 , Λ e − y ) \frac{\Lambda}{e^y+\Lambda}\leq\min(1,\Lambda e^{-y}) e y + Λ Λ ≤ min ( 1 , Λ e − y ) . Let L ∈ Z L\in\mathbb{Z} L ∈ Z be such that e L k ≤ Λ i < e ( L + 1 ) k e^{Lk}\leq\Lambda_i<e^{(L+1)k} e L k ≤ Λ i < e ( L + 1 ) k . Then
Λ i , s i n c s c s ≤ ∑ l ≤ L k e s l k + Λ i + ∑ l ≥ L + 1 k e ( s − 1 ) l k = k e s L k 1 − e − s k + k Λ i e ( s − 1 ) ( L + 1 ) k 1 − e − ( 1 − s ) k ≤ Λ i s ( k 1 − e − s k + k 1 − e − ( 1 − s ) k ) . \frac{\Lambda_{i,\mathrm{sinc}}^s}{c_s}\leq\sum_{l\leq L}ke^{slk+\Lambda_i}+\sum_{l\geq L+1}ke^{(s-1)lk}=\frac{ke^{sLk}}{1-e^{-sk}}+\frac{k\Lambda_i e^{(s-1)(L+1)k}}{1-e^{-(1-s)k}}\leq\Lambda_i^s\left(\frac{k}{1-e^{-sk}}+\frac{k}{1-e^{-(1-s)k}}\right). c s Λ i , sinc s ≤ l ≤ L ∑ k e s l k + Λ i + l ≥ L + 1 ∑ k e ( s − 1 ) l k = 1 − e − s k k e s L k + 1 − e − ( 1 − s ) k k Λ i e ( s − 1 ) ( L + 1 ) k ≤ Λ i s ( 1 − e − s k k + 1 − e − ( 1 − s ) k k ) . where we used e s L k ≤ Λ i s e^{sLk}\leq\Lambda_i^s e s L k ≤ Λ i s and e ( s − 1 ) ( L + 1 ) k ≤ Λ i s − 1 e^{(s-1)(L+1)k}\leq\Lambda_i^{s-1} e ( s − 1 ) ( L + 1 ) k ≤ Λ i s − 1 . The first bound follows from x 1 − e − x ≤ 1 + x \frac{x}{1-e^{-x}}\leq1+x 1 − e − x x ≤ 1 + x for x > 0 x>0 x > 0 , applied with x = s k x=sk x = s k and x = ( 1 − s ) k x=(1-s)k x = ( 1 − s ) k . The second follows from c s = sin ( π s ) π ≤ 1 π c_s=\frac{\sin(\pi s)}{\pi}\leq\frac{1}{\pi} c s = π s i n ( π s ) ≤ π 1 and c s = sin ( π s ) π ≤ 4 s ( 1 − s ) π c_s=\frac{\sin(\pi s)}{\pi}\leq\frac{4s(1-s)}{\pi} c s = π s i n ( π s ) ≤ π 4 s ( 1 − s ) for s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . □
A few comments are in order. Since Λ i = a i i / m i ≤ C h − 2 \Lambda_i=a_{ii}/m_i\leq Ch^{-2} Λ i = a ii / m i ≤ C h − 2 , where h h h denotes the smallest element diameter (the mesh size for quasi-uniform subdivisions T h \mathcal{T}_h T h ) and C C C only depends on the subdivision shape-regularity constant, Lemma 3.5 implies that Λ i , s i n c s ≤ C h − 2 s \Lambda_{i,\mathrm{sinc}}^s\leq Ch^{-2s} Λ i , sinc s ≤ C h − 2 s for a constant C C C uniform in h h h , M M M and N N N . Hence the CFL condition (21) becomes
τ ≲ min ( h , h 2 s ) , (25) \tau\lesssim\min(h,h^{2s}), \tag*{(25)} τ ≲ min ( h , h 2 s ) , (25)
which is the expected stability condition. In particular, for s ≤ 1 2 s\leq\frac{1}{2} s ≤ 2 1 it is compatible with the usual CFL condition for transport equations.
We also observe that the condition τ Λ i , s i n c s ≲ 1 \tau\Lambda_{i,\mathrm{sinc}}^s\lesssim1 τ Λ i , sinc s ≲ 1 included in (21) with Λ i , s i n c s ≈ ( a i i / m i ) s \Lambda_{i,\mathrm{sinc}}^s\approx(a_{ii}/m_i)^s Λ i , sinc s ≈ ( a ii / m i ) s is the fractional analogue of the classical condition τ a i i / m i ≤ 1 \tau a_{ii}/m_i\leq1 τ a ii / m i ≤ 1 for the explicit approximation of the heat equation.
Finally, because the stability condition (25) is uniform in the sinc parameters M M M and N N N , their values only affect the accuracy. Therefore, they can be chosen as in (8) to balance the different quadrature errors without affecting the stability of the discrete scheme.
High-order accuracy
In this section, we show how to increase the accuracy in time and space of the method presented in § 3.3 .
Time approximation
Increasing the time accuracy of the method described in the above section while keeping the maximum principle can be done by using so-called strong stability preserving (SSP) time discretization methods. The key idea is to achieve higher-order accuracy in time by making convex combinations of forward Euler steps. More precisely, each time step of a SSP method is decomposed into substeps that are all forward Euler steps; the final update is constructed as a convex combination of the intermediate solutions. We refer the reader to [17 , 25 , 19 ] for detailed reviews on SSP methods.
Without claiming originality, let us illustrate the SSP concept. Let us consider a finite-dimensional vector space E E E , a subset A ⊂ E A\subset E A ⊂ E and a (possibly nonlinear) operator L : [ 0 , T ] × A → E L:[0,T]\times A\to E L : [ 0 , T ] × A → E . We are interested in approximating in time the following system of ordinary differential equations ∂ t u = L ( t , u ) \partial_tu=L(t,u) ∂ t u = L ( t , u ) with appropriate initial condition. We assume that this problem is well-posed (for instance L L L is continuous w.r.t. t t t and Lipschitz w.r.t. u u u ). We further assume that there exists a convex subset B ⊂ A B\subset A B ⊂ A and τ max > 0 \tau_{\max}>0 τ m a x > 0 such that
v + τ L ( t , v ) ∈ B , ∀ v ∈ B , ∀ t ∈ [ 0 , T ] , ∀ τ ∈ [ 0 , τ max ] . (26) v+\tau L(t,v)\in B,\qquad\forall v\in B,\ \forall t\in[0,T],\ \forall\tau\in[0,\tau_{\max}]. \tag*{(26)} v + τ L ( t , v ) ∈ B , ∀ v ∈ B , ∀ t ∈ [ 0 , T ] , ∀ τ ∈ [ 0 , τ m a x ] . (26)
Let us now consider an S S S -stage explicit Runge–Kutta method, and let us assume that this method has the following ( α − β ) (\alpha-\beta) ( α − β ) representation: There are real coefficients α i k \alpha_{ik} α ik , β i k \beta_{ik} β ik with k ∈ { 0 : i − 1 } k \in\{0:i-1\} k ∈ { 0 : i − 1 } and i ∈ { 1 : S } i \in\{1:S\} i ∈ { 1 : S } such that u n + 1 u^{n+1} u n + 1 is obtained by first setting w ( 0 ) = u n w^{(0)}=u^n w ( 0 ) = u n , then computing
w ( i ) = ∑ k = 0 i − 1 α i k w ( k ) + β i k τ L ( t n + γ k τ , w ( k ) ) , i ∈ { 1 : S } , (27) w^{(i)}=\sum_{k=0}^{i-1}\alpha_{ik}w^{(k)}+\beta_{ik}\tau L(t^n+\gamma_k\tau,w^{(k)}),\qquad i\in\{1:S\}, \tag*{(27)} w ( i ) = k = 0 ∑ i − 1 α ik w ( k ) + β ik τ L ( t n + γ k τ , w ( k ) ) , i ∈ { 1 : S } , (27)
and finally setting u n + 1 = w ( S ) u^{n+1}=w^{(S)} u n + 1 = w ( S ) ; here t n + γ k τ t^n+\gamma_k\tau t n + γ k τ is the time at which stage k k k is evaluated. We further assume that α i k ≥ 0 \alpha_{ik}\geq0 α ik ≥ 0 , β i k ≥ 0 \beta_{ik}\geq0 β ik ≥ 0 , ∑ k = 0 i − 1 α i k = 1 \sum_{k=0}^{i-1}\alpha_{ik}=1 ∑ k = 0 i − 1 α ik = 1 , and that β i k = 0 \beta_{ik}=0 β ik = 0 if α i k = 0 \alpha_{ik}=0 α ik = 0 , for all k ∈ { 0 : i − 1 } k\in\{0:i-1\} k ∈ { 0 : i − 1 } , i ∈ { 1 : S } i\in\{1:S\} i ∈ { 1 : S } . Not every S S S -stage explicit Runge–Kutta method admits an ( α − β ) (\alpha-\beta) ( α − β ) representation with nonnegative coefficients. Any Runge–Kutta method that admits an ( α − β ) (\alpha-\beta) ( α − β ) representation as defined above is said to be SSP for the reason stated in the following theorem.
Theorem 4.1 (Shu–Osher). Let c o s : = inf α i k ≠ 0 , 1 ≤ k + 1 ≤ i ≤ S α i k β i k − 1 c_{\mathrm{os}}:=\inf_{\alpha_{ik}\ne0,\ 1\leq k+1\leq i\leq S}\alpha_{ik}\beta_{ik}^{-1} c os := inf α ik = 0 , 1 ≤ k + 1 ≤ i ≤ S α ik β ik − 1 . Let B ⊂ A B\subset A B ⊂ A be convex. Let u n ∈ B u^n\in B u n ∈ B and let u n + 1 u^{n+1} u n + 1 be given by (27) . Assume that τ ≤ c o s τ max \tau\leq c_{\mathrm{os}}\tau_{\max} τ ≤ c os τ m a x , then u n + 1 ∈ B u^{n+1}\in B u n + 1 ∈ B .
For instance, the SSPRK(2, 2) method can be represented as follows:
w ( 1 ) = u n + τ L ( t n , u n ) , w ( 2 ) = w ( 1 ) + τ L ( t n + 1 , w ( 1 ) ) , u n + 1 = 1 2 u n + 1 2 w ( 2 ) . (28) \begin{aligned}
w^{(1)}&=u^n+\tau L(t^n,u^n),\\
w^{(2)}&=w^{(1)}+\tau L(t^{n+1},w^{(1)}),\\
u^{n+1}&=\frac{1}{2}u^n+\frac{1}{2}w^{(2)}. \tag*{(28)}
\end{aligned} w ( 1 ) w ( 2 ) u n + 1 = u n + τ L ( t n , u n ) , = w ( 1 ) + τ L ( t n + 1 , w ( 1 ) ) , = 2 1 u n + 2 1 w ( 2 ) . (28)
The SSPRK(3, 3) method can be represented as follows:
w ( 1 ) = u n + τ L ( t n , u n ) , z ( 1 ) = w ( 1 ) + τ L ( t n + τ , w ( 1 ) ) , w ( 2 ) = 3 4 u n + 1 4 z ( 1 ) , z ( 2 ) = w ( 2 ) + τ L ( t n + 1 2 τ , w ( 2 ) ) , u n + 1 = 1 3 u n + 2 3 z ( 2 ) . (29) \begin{aligned}
w^{(1)}&=u^n+\tau L(t^n,u^n),\qquad z^{(1)}=w^{(1)}+\tau L(t^n+\tau,w^{(1)}),\\
w^{(2)}&=\frac{3}{4}u^n+\frac{1}{4}z^{(1)},\qquad z^{(2)}=w^{(2)}+\tau L\left(t^n+\frac{1}{2}\tau,w^{(2)}\right),\\
u^{n+1}&=\frac{1}{3}u^n+\frac{2}{3}z^{(2)}. \tag*{(29)}
\end{aligned} w ( 1 ) w ( 2 ) u n + 1 = u n + τ L ( t n , u n ) , z ( 1 ) = w ( 1 ) + τ L ( t n + τ , w ( 1 ) ) , = 4 3 u n + 4 1 z ( 1 ) , z ( 2 ) = w ( 2 ) + τ L ( t n + 2 1 τ , w ( 2 ) ) , = 3 1 u n + 3 2 z ( 2 ) . (29)
Remark 4.2 (Literature). We remark that Theorem 4.1 has been established in a slightly different form in Shu and Osher [32 Prop. 2.1] without explicitly invoking convexity. The variant of this theorem presented here relies only on convexity.
Second-order approximation in space
In Section 3.3, we introduced a low-order method to approximate (1) while ensuring that the maximum principle is preserved. The low-order scheme (17) serves as a crucial component in the limiting technique discussed in the subsequent section. Our objective is to construct a method that is at least second-order accurate and that also preserves the maximum principle. To achieve this goal, we require two additional components: a high-order entropy consistent method (that may violate the maximum principle due to oscillations in its solutions), and a reliable limiting algorithm that uses both the low- and high-order solutions to construct a higher-order maximum principle preserving approximation.
We fix ( η , q ) (\eta,q) ( η , q ) to be an entropy pair for (1) , i.e., η ∈ C 1 ( R ) \eta\in C^1(\mathbb{R}) η ∈ C 1 ( R ) is convex and q ∈ C 1 ( R ; R d ) q\in C^1(\mathbb{R};\mathbb{R}^d) q ∈ C 1 ( R ; R d ) satisfies q ′ ( u ) = η ′ ( u ) f ′ ( u ) q'(u)=\eta'(u)f'(u) q ′ ( u ) = η ′ ( u ) f ′ ( u ) . The associated entropy residual is defined as
R ( u ) : = ( ∂ t u + ∇ ⋅ f ( u ) + ( − Δ ) s u − g ) η ′ ( u ) = ( ∂ t u + ( − Δ ) s u − g ) η ′ ( u ) + ∇ ⋅ q ( u ) R(u):=\left(\partial_tu+\nabla\cdot f(u)+(-\Delta)^s u-g\right)\eta'(u)=\left(\partial_tu+(-\Delta)^s u-g\right)\eta'(u)+\nabla\cdot q(u) R ( u ) := ( ∂ t u + ∇ ⋅ f ( u ) + ( − Δ ) s u − g ) η ′ ( u ) = ( ∂ t u + ( − Δ ) s u − g ) η ′ ( u ) + ∇ ⋅ q ( u )
and vanishes when u u u is a smooth solution to (1) .
In order to derive a discrete counterpart of the above entropy residual, we follow the strategy adopted in Guermond et al. [23 ] and Guermond et al. [24 ] . Given u h n = ∑ i ∈ V U i n φ i u_h^n=\sum_{i\in\mathcal{V}}U_i^n\varphi_i u h n = ∑ i ∈ V U i n φ i , an approximation of u u u at time t n t_n t n , we let U i G , n + 1 U_i^{G,n+1} U i G , n + 1 be defined as the low order approximation (17) but without stabilization terms, i.e., it satisfies
m i U i G , n + 1 − U i n τ = − ∑ j ∈ I ( i ) c i j ⋅ f ( U j n ) − c s m i ∑ l ∈ L ω l e s y l ( W i l , n ( u h n ) + U i n ) + m i G i n . (30) m_i\frac{U_i^{G,n+1}-U_i^n}{\tau}=-\sum_{j\in\mathcal{I}(i)}c_{ij}\cdot f(U_j^n)-c_sm_i\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}(W_i^{l,n}(u_h^n)+U_i^n)+m_iG_i^n. \tag*{(30)} m i τ U i G , n + 1 − U i n = − j ∈ I ( i ) ∑ c ij ⋅ f ( U j n ) − c s m i l ∈ L ∑ ω l e s y l ( W i l , n ( u h n ) + U i n ) + m i G i n . (30)
Then, the discrete entropy residual reads R h ( u ) : = ∑ i ∈ V R i n φ i R_h(u):=\sum_{i\in\mathcal{V}}R_i^n\varphi_i R h ( u ) := ∑ i ∈ V R i n φ i , where
R i n : = ( U i G , n + 1 − U i n τ ) η ′ ( U i n ) + 1 m i ∑ j ∈ I ( i ) c i j ⋅ q ( U j n ) + ( c s ∑ l ∈ L ω l e s y l ( W i l , n ( u h n ) + U i n ) − G i n ) η ′ ( U i n ) . (31) \begin{aligned}
R_i^n:=\left(\frac{U_i^{G,n+1}-U_i^n}{\tau}\right)\eta'(U_i^n)+\frac{1}{m_i}\sum_{j\in\mathcal{I}(i)}c_{ij}\cdot q(U_j^n) \\
&+\left(c_s\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}(W_i^{l,n}(u_h^n)+U_i^n)-G_i^n\right)\eta'(U_i^n). \tag*{(31)}
\end{aligned} R i n := ( τ U i G , n + 1 − U i n ) η ′ ( U i n ) + m i 1 j ∈ I ( i ) ∑ c ij ⋅ q ( U j n ) + ( c s l ∈ L ∑ ω l e s y l ( W i l , n ( u h n ) + U i n ) − G i n ) η ′ ( U i n ) . (31)
After multiplying (30) by η ′ ( U i n ) \eta'(U_i^n) η ′ ( U i n ) , we notice that the coefficient R i n R_i^n R i n in (31) reduces to
R i n = − 1 m i ∑ j ∈ I ( i ) c i j ⋅ f ( U j n ) η ′ ( U i n ) + 1 m i ∑ j ∈ I ( i ) c i j ⋅ q ( U j n ) = : R 1 , i n + R 2 , i n . (32) \begin{aligned}
R_i^n=-\frac{1}{m_i}\sum_{j\in\mathcal{I}(i)}c_{ij}\cdot f(U_j^n)\eta'(U_i^n)+\frac{1}{m_i}\sum_{j\in\mathcal{I}(i)}c_{ij}\cdot q(U_j^n) \\
&=:R_{1,i}^n+R_{2,i}^n. \tag*{(32)}
\end{aligned} R i n = − m i 1 j ∈ I ( i ) ∑ c ij ⋅ f ( U j n ) η ′ ( U i n ) + m i 1 j ∈ I ( i ) ∑ c ij ⋅ q ( U j n ) =: R 1 , i n + R 2 , i n . (32)
Next, we propose a nodal residual indicator for each node i ∈ V i\in\mathcal{V} i ∈ V :
α i n : = ∣ R 1 , i n + R 2 , i n ∣ ∣ R 1 , i n ∣ + ∣ R 2 , i n ∣ + ϵ V R . \alpha_i^n:=\frac{\left|R_{1,i}^n+R_{2,i}^n\right|}{\left|R_{1,i}^n\right|+\left|R_{2,i}^n\right|+\epsilon_{\mathcal{V}}^R}. α i n := R 1 , i n + R 2 , i n + ϵ V R R 1 , i n + R 2 , i n .
where ϵ V R : = 10 − 8 max j ∈ V ( ∣ R 1 , j n ∣ + ∣ R 2 , j n ∣ ) \epsilon_{\mathcal{V}}^R:=10^{-8}\max_{j\in\mathcal{V}}\left(\left|R_{1,j}^n\right|+\left|R_{2,j}^n\right|\right) ϵ V R := 1 0 − 8 max j ∈ V ( R 1 , j n + R 2 , j n ) is a small positive constant introduced to avoid division by zero. On the one hand, in regions where the solution is smooth, the entropy residual is small, i.e., R 1 , i n R_{1,i}^n R 1 , i n and R 2 , i n R_{2,i}^n R 2 , i n nearly cancel, and so the residual indicator is small. On the other hand, where the solution is discontinuous, the entropy residual is large and the residual indicator is close to one.
We are now in position to define the high-order graph viscosity coefficient and the high-order approximation. We set
d i j H , n : = d i j L , n max ( α i n , α j n ) when i ≠ j , d i i H , n : = − ∑ j ∈ I ( i ) ∖ { i } d i j H , n . (33) d_{ij}^{H,n}:=d_{ij}^{L,n}\max(\alpha_i^n,\alpha_j^n)\quad\text{when }i\ne j,\qquad d_{ii}^{H,n}:=-\sum_{j\in\mathcal{I}(i)\setminus\{i\}}d_{ij}^{H,n}. \tag*{(33)} d ij H , n := d ij L , n max ( α i n , α j n ) when i = j , d ii H , n := − j ∈ I ( i ) ∖ { i } ∑ d ij H , n . (33)
Since α j n \alpha_j^n α j n is only defined for j ∈ V j\in\mathcal{V} j ∈ V , we set α j n : = 0 \alpha_j^n:=0 α j n := 0 for j ∈ V ∂ j\in\mathcal{V}^{\partial} j ∈ V ∂ , so that only the indicator of the interior vertex i i i enters d i j H , n d_{ij}^{H,n} d ij H , n when j j j is a boundary vertex. Note that, similarly to the low-order graph viscosity, we have d i j H , n = d j i H , n d_{ij}^{H,n}=d_{ji}^{H,n} d ij H , n = d j i H , n , d i j H , n ≥ 0 d_{ij}^{H,n}\ge0 d ij H , n ≥ 0 for all i ≠ j i\ne j i = j , and ∑ j ∈ I ( i ) d i j H , n = 0 \sum_{j\in\mathcal{I}(i)}d_{ij}^{H,n}=0 ∑ j ∈ I ( i ) d ij H , n = 0 .
Given an approximation u h n = ∑ i ∈ V U i n φ i u_h^n=\sum_{i\in\mathcal{V}}U_i^n\varphi_i u h n = ∑ i ∈ V U i n φ i of the solution u u u to (7) at time t n t_n t n , the high-order approximation u h H , n + 1 = ∑ i ∈ V U i H , n + 1 φ i u_h^{H,n+1}=\sum_{i\in\mathcal{V}}U_i^{H,n+1}\varphi_i u h H , n + 1 = ∑ i ∈ V U i H , n + 1 φ i of u u u at time t n + 1 t_{n+1} t n + 1 is defined by the relations
∑ j ∈ I ( i ) m i j U j H , n + 1 − U j n τ + ∑ j ∈ I ( i ) c i j ⋅ f ( U j n ) + c s m i ∑ l ∈ L ω l e s y l ( W i l , n ( u h n ) + U i n ) = m i G i n + ∑ j ∈ I ( i ) ∖ { i } d i j H , n ( U j n − U i n ) . (34) \sum_{j\in\mathcal{I}(i)}m_{ij}\frac{U_j^{H,n+1}-U_j^n}{\tau}+\sum_{j\in\mathcal{I}(i)}c_{ij}\cdot f(U_j^n)+c_sm_i\sum_{l\in\mathcal{L}}\omega_l e^{s y_l}(W_i^{l,n}(u_h^n)+U_i^n)
=m_iG_i^n+\sum_{j\in\mathcal{I}(i)\setminus\{i\}}d_{ij}^{H,n}(U_j^n-U_i^n). \tag*{(34)} j ∈ I ( i ) ∑ m ij τ U j H , n + 1 − U j n + j ∈ I ( i ) ∑ c ij ⋅ f ( U j n ) + c s m i l ∈ L ∑ ω l e s y l ( W i l , n ( u h n ) + U i n ) = m i G i n + j ∈ I ( i ) ∖ { i } ∑ d ij H , n ( U j n − U i n ) . (34)
Compared to the low-order approximation (17) , where a lumped mass matrix is used, here m i j : = ∫ D φ j ( x ) φ i ( x ) d x m_{ij}:=\int_D\varphi_j(x)\varphi_i(x)\,\mathrm{d}x m ij := ∫ D φ j ( x ) φ i ( x ) d x are the coefficients of the consistent mass matrix.
Relation between high- and low-order solutions
To create a limiter matrix, it is necessary to establish the relationship between the high-order solution and the low-order solution while keeping in mind that the low-order solution maintains the maximum principle whereas the high-order solution may not.
Let us write the first-order and the high-order schemes as
m i U i L , n + 1 = m i U i n + τ ∑ j ∈ I ( i ) ∖ { i } ( − c i j ⋅ f ( U j n ) + d i j L , n ( U j n − U i n ) ) − τ m i c s ∑ l ∈ L ω l ε s y l ( W i L , n + U i n ) + τ m i G i n , m_i U_i^{L,n+1}=m_i U_i^n+\tau\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\left(-c_{ij}\cdot f(U_j^n)+d_{ij}^{L,n}(U_j^n-U_i^n)\right)-\tau m_i c_s\sum_{l\in\mathcal{L}}\omega_l\varepsilon^{s y_l}(W_i^{L,n}+U_i^n)+\tau m_iG_i^n, m i U i L , n + 1 = m i U i n + τ j ∈ I ( i ) ∖ { i } ∑ ( − c ij ⋅ f ( U j n ) + d ij L , n ( U j n − U i n ) ) − τ m i c s l ∈ L ∑ ω l ε s y l ( W i L , n + U i n ) + τ m i G i n ,
and
∑ j ∈ I ( i ) m i j U j H , n + 1 = ∑ j ∈ I ( i ) m i j U j n + τ ∑ j ∈ I ( i ) ∖ { i } ( − c i j ⋅ f ( U j n ) + d i j H , n ( U j n − U i n ) ) − τ m i c s ∑ l ∈ L ω l ε s y l ( W i L , n + U i n ) + τ m i G i n . \sum_{j\in\mathcal{I}(i)}m_{ij}U_j^{H,n+1}=\sum_{j\in\mathcal{I}(i)}m_{ij}U_j^n+\tau\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\left(-c_{ij}\cdot f(U_j^n)+d_{ij}^{H,n}(U_j^n-U_i^n)\right)-\tau m_i c_s\sum_{l\in\mathcal{L}}\omega_l\varepsilon^{s y_l}(W_i^{L,n}+U_i^n)+\tau m_iG_i^n. j ∈ I ( i ) ∑ m ij U j H , n + 1 = j ∈ I ( i ) ∑ m ij U j n + τ j ∈ I ( i ) ∖ { i } ∑ ( − c ij ⋅ f ( U j n ) + d ij H , n ( U j n − U i n ) ) − τ m i c s l ∈ L ∑ ω l ε s y l ( W i L , n + U i n ) + τ m i G i n .
By subtracting the low-order scheme from the high-order scheme, we obtain:
∑ j ∈ I ( i ) m i j U j H , n + 1 = m i U i L , n + 1 + ∑ j ∈ I ( i ) m i j U j n − m i U i n + τ ∑ j ∈ I ( i ) ∖ { i } ( d i j H , n − d i j L , n ) ( U j n − U i n ) . \sum_{j\in\mathcal{I}(i)}m_{ij}U_j^{H,n+1}=m_iU_i^{L,n+1}+\sum_{j\in\mathcal{I}(i)}m_{ij}U_j^n-m_iU_i^n+\tau\sum_{j\in\mathcal{I}(i)\setminus\{i\}}(d_{ij}^{H,n}-d_{ij}^{L,n})(U_j^n-U_i^n). j ∈ I ( i ) ∑ m ij U j H , n + 1 = m i U i L , n + 1 + j ∈ I ( i ) ∑ m ij U j n − m i U i n + τ j ∈ I ( i ) ∖ { i } ∑ ( d ij H , n − d ij L , n ) ( U j n − U i n ) .
By observing that m i = ∑ j ∈ I ( i ) m i j m_i=\sum_{j\in\mathcal{I}(i)}m_{ij} m i = ∑ j ∈ I ( i ) m ij we add m i U i H , n + 1 − ∑ j ∈ I ( i ) m i j U j H , n + 1 m_iU_i^{H,n+1}-\sum_{j\in\mathcal{I}(i)}m_{ij}U_j^{H,n+1} m i U i H , n + 1 − ∑ j ∈ I ( i ) m ij U j H , n + 1 to both sides of the last equality and obtain
m i U i H , n + 1 = m i U i L , n + 1 + ∑ j ∈ I ( i ) m i j ( U i H , n + 1 − U j H , n + 1 ) + ∑ j ∈ I ( i ) m i j ( U j n − U i n ) + τ ∑ j ∈ I ( i ) ∖ { i } ( d i j H , n − d i j L , n ) ( U j n − U i n ) , m_iU_i^{H,n+1}=m_iU_i^{L,n+1}+\sum_{j\in\mathcal{I}(i)}m_{ij}(U_i^{H,n+1}-U_j^{H,n+1})+\sum_{j\in\mathcal{I}(i)}m_{ij}(U_j^n-U_i^n)
+\tau\sum_{j\in\mathcal{I}(i)\setminus\{i\}}(d_{ij}^{H,n}-d_{ij}^{L,n})(U_j^n-U_i^n), m i U i H , n + 1 = m i U i L , n + 1 + j ∈ I ( i ) ∑ m ij ( U i H , n + 1 − U j H , n + 1 ) + j ∈ I ( i ) ∑ m ij ( U j n − U i n ) + τ j ∈ I ( i ) ∖ { i } ∑ ( d ij H , n − d ij L , n ) ( U j n − U i n ) ,
which we write in the form
m i U i H , n + 1 = m i U i L , n + 1 + ∑ j ∈ I ( i ) A i j n + 1 , (35) m_iU_i^{H,n+1}=m_iU_i^{L,n+1}+\sum_{j\in\mathcal{I}(i)}A_{ij}^{n+1}, \tag*{(35)} m i U i H , n + 1 = m i U i L , n + 1 + j ∈ I ( i ) ∑ A ij n + 1 , (35)
where for i ∈ V i\in\mathcal{V} i ∈ V and j ∈ I ( i ) j\in\mathcal{I}(i) j ∈ I ( i )
A i j n + 1 : = ( m i δ i j − m i j ) ( U j H , n + 1 − U i H , n + 1 − ( U j n − U i n ) ) + τ ( d i j H , n − d i j L , n ) ( U j n − U i n ) . A_{ij}^{n+1}:=(m_i\delta_{ij}-m_{ij})\left(U_j^{H,n+1}-U_i^{H,n+1}-(U_j^n-U_i^n)\right)+\tau(d_{ij}^{H,n}-d_{ij}^{L,n})(U_j^n-U_i^n). A ij n + 1 := ( m i δ ij − m ij ) ( U j H , n + 1 − U i H , n + 1 − ( U j n − U i n ) ) + τ ( d ij H , n − d ij L , n ) ( U j n − U i n ) .
The equation (35) expresses the relation between the high-order and low-order solutions. We observe that the skew-symmetry property A i j n + 1 = − A j i n + 1 A_{ij}^{n+1}=-A_{ji}^{n+1} A ij n + 1 = − A j i n + 1 for i , j ∈ V i,j\in\mathcal{V} i , j ∈ V implies
∑ i ∈ V m i U i H , n + 1 = ∑ i ∈ V m i U i L , n + 1 + ∑ i ∈ V ∑ j ∈ I ( i ) ∩ V ∂ A i j n + 1 , (36) \sum_{i\in\mathcal{V}}m_iU_i^{H,n+1}=\sum_{i\in\mathcal{V}}m_iU_i^{L,n+1}+\sum_{i\in\mathcal{V}}\sum_{j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial}}A_{ij}^{n+1}, \tag*{(36)} i ∈ V ∑ m i U i H , n + 1 = i ∈ V ∑ m i U i L , n + 1 + i ∈ V ∑ j ∈ I ( i ) ∩ V ∂ ∑ A ij n + 1 , (36)
since the contributions of the pairs of interior vertices cancel. In the periodic setting, V ∂ = ∅ \mathcal{V}^{\partial}=\varnothing V ∂ = ∅ and both schemes conserve the mass, i.e., ∑ i ∈ V m i U i H , n + 1 = ∑ i ∈ V m i U i L , n + 1 = ∑ i ∈ V m i U i n \sum_{i\in\mathcal{V}}m_iU_i^{H,n+1}=\sum_{i\in\mathcal{V}}m_iU_i^{L,n+1}=\sum_{i\in\mathcal{V}}m_iU_i^n ∑ i ∈ V m i U i H , n + 1 = ∑ i ∈ V m i U i L , n + 1 = ∑ i ∈ V m i U i n . With homogeneous Dirichlet conditions, the mass is not conserved at the continuous level, since ∫ D ( − Δ ) s u d x ≠ 0 \int_D(-\Delta)^s u\,dx\ne0 ∫ D ( − Δ ) s u d x = 0 in general, and neither scheme conserves it: both exchange mass with the boundary through the pairs ( i , j ) (i,j) ( i , j ) with j ∈ V ∂ j\in\mathcal{V}^{\partial} j ∈ V ∂ . Relation (36) shows that the high-order correction only redistributes mass among the interior vertices, except along these pairs, where it modifies the boundary flux of the low-order scheme by ∑ i ∈ V ∑ j ∈ I ( i ) ∩ V ∂ A i j n + 1 \sum_{i\in\mathcal{V}}\sum_{j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial}}A_{ij}^{n+1} ∑ i ∈ V ∑ j ∈ I ( i ) ∩ V ∂ A ij n + 1 .
Convex limiting
The objective of convex limiting, see e.g., [24 Sec. 4] , is to determine a limiter matrix ( ℓ i j ) i ∈ V , j ∈ I ( i ) (\ell_{ij})_{i\in\mathcal{V},\,j\in\mathcal{I}(i)} ( ℓ ij ) i ∈ V , j ∈ I ( i ) such that ℓ i j ∈ [ 0 , 1 ] \ell_{ij}\in[0,1] ℓ ij ∈ [ 0 , 1 ] for i ∈ V i\in\mathcal{V} i ∈ V , j ∈ I ( i ) j\in\mathcal{I}(i) j ∈ I ( i ) , and ℓ i j = ℓ j i \ell_{ij}=\ell_{ji} ℓ ij = ℓ j i for i ∈ V i\in\mathcal{V} i ∈ V , j ∈ I ( i ) ∩ V j\in\mathcal{I}(i)\cap\mathcal{V} j ∈ I ( i ) ∩ V (symmetry is only required between interior degrees of freedom), and
m i U i n + 1 : = m i U i L , n + 1 + ∑ j ∈ I ( i ) ℓ i j A i j n + 1 , (37) m_i U_i^{n+1}:=m_i U_i^{L,n+1}+\sum_{j\in\mathcal{I}(i)}\ell_{ij}A_{ij}^{n+1}, \tag*{(37)} m i U i n + 1 := m i U i L , n + 1 + j ∈ I ( i ) ∑ ℓ ij A ij n + 1 , (37)
satisfies Ψ i ± ( U i n + 1 ) ≥ 0 \Psi_i^{\pm}(U_i^{n+1})\geq0 Ψ i ± ( U i n + 1 ) ≥ 0 , where
Ψ i − ( w ) : = U i n , max + τ G i n − w and Ψ i + ( w ) : = w − U i n , min − τ G i n . \Psi_i^{-}(w):=U_i^{n,\max}+\tau G_i^n-w
\qquad\text{and}\qquad
\Psi_i^{+}(w):=w-U_i^{n,\min}-\tau G_i^n. Ψ i − ( w ) := U i n , m a x + τ G i n − w and Ψ i + ( w ) := w − U i n , m i n − τ G i n .
Recall that U i n , max U_i^{n,\max} U i n , m a x and U i n , min U_i^{n,\min} U i n , m i n are defined in (23) .
It should be noted that when ℓ i j ≡ 1 \ell_{ij}\equiv1 ℓ ij ≡ 1 in (37) , we obtain the high-order solution U i H , n + 1 U_i^{H,n+1} U i H , n + 1 , which may potentially violate the maximum principle. In contrast, setting ℓ i j ≡ 0 \ell_{ij}\equiv0 ℓ ij ≡ 0 corresponds to the low-order solution U i L , n + 1 U_i^{L,n+1} U i L , n + 1 , which is known to satisfy the maximum principle. Additionally, similarly to (36) but using in addition the symmetry assumption ℓ i j = ℓ j i \ell_{ij}=\ell_{ji} ℓ ij = ℓ j i for i ∈ V i\in\mathcal{V} i ∈ V and j ∈ I ( i ) ∩ V j\in\mathcal{I}(i)\cap\mathcal{V} j ∈ I ( i ) ∩ V , we have
∑ i ∈ V m i U i n + 1 = ∑ i ∈ V m i U i L , n + 1 + ∑ i ∈ V ∑ j ∈ I ( i ) ∩ V ∂ ℓ i j A i j n + 1 . \sum_{i\in\mathcal{V}}m_iU_i^{n+1}
=\sum_{i\in\mathcal{V}}m_iU_i^{L,n+1}
+\sum_{i\in\mathcal{V}}\sum_{j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial}}\ell_{ij}A_{ij}^{n+1}. i ∈ V ∑ m i U i n + 1 = i ∈ V ∑ m i U i L , n + 1 + i ∈ V ∑ j ∈ I ( i ) ∩ V ∂ ∑ ℓ ij A ij n + 1 .
We now propose a limiter matrix and prove in Theorem 4.3 that it leads to a discrete maximum principle. Given u h n = ∑ i ∈ V U i n φ i u_h^n=\sum_{i\in\mathcal{V}}U_i^n\varphi_i u h n = ∑ i ∈ V U i n φ i an approximation at time t n t_n t n and u h L , n + 1 = ∑ i ∈ V U i L , n + 1 φ i u_h^{L,n+1}=\sum_{i\in\mathcal{V}}U_i^{L,n+1}\varphi_i u h L , n + 1 = ∑ i ∈ V U i L , n + 1 φ i a low-order approximation at time t n + 1 t_{n+1} t n + 1 satisfying the local discrete maximum principle (22) , for i ∈ V i\in\mathcal{V} i ∈ V , we set ℓ i j = 0 \ell_{ij}=0 ℓ ij = 0 when j ∉ I ( i ) j\notin\mathcal{I}(i) j ∈ / I ( i ) and
ℓ i j : = min ( ℓ j i , ℓ i j ) if j ∈ I ( i ) ∩ V , ℓ i j : = ℓ j i if j ∈ I ( i ) ∩ V ∂ , (38) \ell_{ij}:=\min(\ell_j^i,\ell_i^j)\quad\text{if }j\in\mathcal{I}(i)\cap\mathcal{V},\qquad
\ell_{ij}:=\ell_j^i\quad\text{if }j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial}, \tag*{(38)} ℓ ij := min ( ℓ j i , ℓ i j ) if j ∈ I ( i ) ∩ V , ℓ ij := ℓ j i if j ∈ I ( i ) ∩ V ∂ , (38)
with
ℓ j i : = { ∣ U i n , min + τ G i n − U i L , n + 1 ∣ ∣ P i j n + 1 ∣ if U i L , n + 1 + P i j n + 1 < U i n , min + τ G i n , ∣ U i n , max + τ G i n − U i L , n + 1 ∣ ∣ P i j n + 1 ∣ if U i n , max + τ G i n < U i L , n + 1 + P i j n + 1 , 1 otherwise , \ell_j^i:=
\begin{cases}
\dfrac{\left|U_i^{n,\min}+\tau G_i^n-U_i^{L,n+1}\right|}{\left|P_{ij}^{n+1}\right|}
&\text{if }U_i^{L,n+1}+P_{ij}^{n+1}<U_i^{n,\min}+\tau G_i^n,\\[1.2ex]
\dfrac{\left|U_i^{n,\max}+\tau G_i^n-U_i^{L,n+1}\right|}{\left|P_{ij}^{n+1}\right|}
&\text{if }U_i^{n,\max}+\tau G_i^n<U_i^{L,n+1}+P_{ij}^{n+1},\\[1.2ex]
1&\text{otherwise},
\end{cases} ℓ j i := ⎩ ⎨ ⎧ P ij n + 1 U i n , m i n + τ G i n − U i L , n + 1 P ij n + 1 U i n , m a x + τ G i n − U i L , n + 1 1 if U i L , n + 1 + P ij n + 1 < U i n , m i n + τ G i n , if U i n , m a x + τ G i n < U i L , n + 1 + P ij n + 1 , otherwise ,
where P i j n + 1 : = 1 λ j ( i ) m i A i j n + 1 P_{ij}^{n+1}:=\dfrac{1}{\lambda_j(i)m_i}A_{ij}^{n+1} P ij n + 1 := λ j ( i ) m i 1 A ij n + 1 with λ j ( i ) : = 1 card ( I ( i ) ) − 1 \lambda_j(i):=\dfrac{1}{\operatorname{card}(\mathcal{I}(i))-1} λ j ( i ) := card ( I ( i )) − 1 1 for i ∈ V i\in\mathcal{V} i ∈ V and j ∈ I ( i ) ∖ { i } j\in\mathcal{I}(i)\setminus\{i\} j ∈ I ( i ) ∖ { i } . Note that for j ∈ I ( i ) ∩ V ∂ j\in\mathcal{I}(i)\cap\mathcal{V}^{\partial} j ∈ I ( i ) ∩ V ∂ , ℓ i j \ell_i^j ℓ i j is not defined and the flux between i i i and j j j is then limited using only ℓ j i \ell_j^i ℓ j i . Also, when U i L , n + 1 + P i j n + 1 < U i n , min + τ G i n U_i^{L,n+1}+P_{ij}^{n+1}<U_i^{n,\min}+\tau G_i^n U i L , n + 1 + P ij n + 1 < U i n , m i n + τ G i n , because the low-order approximation is assumed to satisfy the local discrete maximum principle (22) , we deduce that
P i j n + 1 < U i n , min + τ G i n − U i L , n + 1 ≤ 0. (39) P_{ij}^{n+1}<U_i^{n,\min}+\tau G_i^n-U_i^{L,n+1}\leq0. \tag*{(39)} P ij n + 1 < U i n , m i n + τ G i n − U i L , n + 1 ≤ 0. (39)
Therefore, dividing by ∣ P i j n + 1 ∣ \left|P_{ij}^{n+1}\right| P ij n + 1 in the definition of ℓ j i \ell_j^i ℓ j i is justified, and
0 ≤ ℓ j i = ∣ U i n , min + τ G i n − U i L , n + 1 ∣ ∣ P i j n + 1 ∣ ≤ 1. 0\leq\ell_j^i=\frac{\left|U_i^{n,\min}+\tau G_i^n-U_i^{L,n+1}\right|}{\left|P_{ij}^{n+1}\right|}\leq1. 0 ≤ ℓ j i = P ij n + 1 U i n , m i n + τ G i n − U i L , n + 1 ≤ 1.
A similar argument shows that for the case U i n , max + τ G i n < U i L , n + 1 + P i j n + 1 U_i^{n,\max}+\tau G_i^n<U_i^{L,n+1}+P_{ij}^{n+1} U i n , m a x + τ G i n < U i L , n + 1 + P ij n + 1 , we have P i j n + 1 > 0 P_{ij}^{n+1}>0 P ij n + 1 > 0 and
0 ≤ ℓ j i = ∣ U i n , max + τ G i n − U i L , n + 1 ∣ ∣ P i j n + 1 ∣ ≤ 1. 0\leq\ell_j^i=\frac{\left|U_i^{n,\max}+\tau G_i^n-U_i^{L,n+1}\right|}{\left|P_{ij}^{n+1}\right|}\leq1. 0 ≤ ℓ j i = P ij n + 1 U i n , m a x + τ G i n − U i L , n + 1 ≤ 1.
In practice, we avoid numerical instabilities by replacing ∣ P i j n + 1 ∣ \left|P_{ij}^{n+1}\right| P ij n + 1 by ∣ P i j n + 1 ∣ + 10 − 14 max i ∈ V ∣ U i n ∣ \left|P_{ij}^{n+1}\right|+10^{-14}\max_{i\in\mathcal{V}}|U_i^n| P ij n + 1 + 1 0 − 14 max i ∈ V ∣ U i n ∣ in the denominator.
Theorem 4.3. Let u h n = ∑ i ∈ V U i n ϕ i u_h^n = \sum_{i\in\mathcal{V}} U_i^n \phi_i u h n = ∑ i ∈ V U i n ϕ i be an approximation at time t n t_n t n , U i L , n + 1 U_i^{\mathrm{L},n+1} U i L , n + 1 the low-order approximation given by (17) and U i n + 1 U_i^{n+1} U i n + 1 the convex limiting solution defined by (37) with the limiting matrix given in (38) . Assume that the time-step τ \tau τ satisfies (21) for some 0 < C F L ≤ 1 2 0 < \mathrm{CFL} \le\frac{1}{2} 0 < CFL ≤ 2 1 . Then we have
U i n , min + τ G i n ≤ U i n + 1 ≤ U i n , max + τ G i n , U_i^{n,\min}+\tau G_i^n \le U_i^{n+1} \le U_i^{n,\max}+\tau G_i^n, U i n , m i n + τ G i n ≤ U i n + 1 ≤ U i n , m a x + τ G i n , for every i ∈ V i\in\mathcal{V} i ∈ V .
Proof. We fix i ∈ V i\in\mathcal{V} i ∈ V and remark that the stated bounds are equivalent to showing that Ψ i ± ( U i n + 1 ) ≥ 0 \Psi_i^\pm(U_i^{n+1})\ge0 Ψ i ± ( U i n + 1 ) ≥ 0 . We only prove the result for Ψ i + \Psi_i^+ Ψ i + since the proof for Ψ i − \Psi_i^- Ψ i − is analogous.
We start by noting that (37) can be rewritten as a convex combination
U i n + 1 = ∑ j ∈ I ( i ) ∖ { i } λ j ( i ) ( U i L , n + 1 + ℓ i j P i j n + 1 ) . (40) U_i^{n+1}=\sum_{j\in\mathcal{I}(i)\setminus\{i\}}\lambda_j^{(i)}\left(U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1}\right). \tag*{(40)} U i n + 1 = j ∈ I ( i ) ∖ { i } ∑ λ j ( i ) ( U i L , n + 1 + ℓ ij P ij n + 1 ) . (40)
Furthermore, since the upper level set L ( Ψ i + ) : = { r ∈ R ∣ Ψ i + ( r ) ≥ 0 } L(\Psi_i^+) := \{r\in\mathbb{R}\mid\Psi_i^+(r)\ge0\} L ( Ψ i + ) := { r ∈ R ∣ Ψ i + ( r ) ≥ 0 } is convex, the function Ψ i + : R ↦ R \Psi_i^+:\mathbb{R}\mapsto\mathbb{R} Ψ i + : R ↦ R is quasi-concave. Therefore, if Ψ i + ( U i L , n + 1 + ℓ i j P i j n + 1 ) ≥ 0 \Psi_i^+(U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1})\ge0 Ψ i + ( U i L , n + 1 + ℓ ij P ij n + 1 ) ≥ 0 for all j ∈ I ( i ) ∖ { i } j\in\mathcal{I}(i)\setminus\{i\} j ∈ I ( i ) ∖ { i } , then the convex-combination representation (40) of U i n + 1 U_i^{n+1} U i n + 1 implies U i n + 1 ∈ L ( Ψ i + ) U_i^{n+1}\in L(\Psi_i^+) U i n + 1 ∈ L ( Ψ i + ) , i.e., Ψ i + ( U i n + 1 ) ≥ 0 \Psi_i^+(U_i^{n+1})\ge0 Ψ i + ( U i n + 1 ) ≥ 0 as desired.
It remains to show that Ψ i + ( U i L , n + 1 + ℓ i j P i j n + 1 ) ≥ 0 \Psi_i^+(U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1})\ge0 Ψ i + ( U i L , n + 1 + ℓ ij P ij n + 1 ) ≥ 0 , which is performed by distinguishing three cases.
When U i n , min + τ G i n ≤ U i L , n + 1 + P i j n + 1 ≤ U i n , max + τ G i n U_i^{n,\min}+\tau G_i^n\le U_i^{\mathrm{L},n+1}+P_{ij}^{n+1}\le U_i^{n,\max}+\tau G_i^n U i n , m i n + τ G i n ≤ U i L , n + 1 + P ij n + 1 ≤ U i n , m a x + τ G i n , we have
Ψ i + ( U i L , n + 1 + P i j n + 1 ) ≥ 0. \Psi_i^+\left(U_i^{\mathrm{L},n+1}+P_{ij}^{n+1}\right)\ge0. Ψ i + ( U i L , n + 1 + P ij n + 1 ) ≥ 0.
In addition, the maximum principle property (22) satisfied by the low-order coefficient U i L , n + 1 U_i^{\mathrm{L},n+1} U i L , n + 1 reads
Ψ i + ( U i L , n + 1 ) ≥ 0. \Psi_i^+\left(U_i^{\mathrm{L},n+1}\right)\ge0. Ψ i + ( U i L , n + 1 ) ≥ 0.
Therefore, since 0 ≤ ℓ i j ≤ 1 0\le\ell_{ij}\le1 0 ≤ ℓ ij ≤ 1 ,
U i L , n + 1 + ℓ i j P i j n + 1 = ( 1 − ℓ i j ) U i L , n + 1 + ℓ i j ( U i L , n + 1 + P i j n + 1 ) , U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1}=(1-\ell_{ij})U_i^{\mathrm{L},n+1}+\ell_{ij}\left(U_i^{\mathrm{L},n+1}+P_{ij}^{n+1}\right), U i L , n + 1 + ℓ ij P ij n + 1 = ( 1 − ℓ ij ) U i L , n + 1 + ℓ ij ( U i L , n + 1 + P ij n + 1 ) ,
is a convex combination in L ( Ψ i + ) L(\Psi_i^+) L ( Ψ i + ) and thus
Ψ i + ( U i L , n + 1 + ℓ i j P i j n + 1 ) ≥ 0. \Psi_i^+\left(U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1}\right)\ge0. Ψ i + ( U i L , n + 1 + ℓ ij P ij n + 1 ) ≥ 0.
When U i L , n + 1 + P i j n + 1 < U i n , min + τ G i n U_i^{\mathrm{L},n+1}+P_{ij}^{n+1}<U_i^{n,\min}+\tau G_i^n U i L , n + 1 + P ij n + 1 < U i n , m i n + τ G i n , we recall that P i j n + 1 < 0 P_{ij}^{n+1}<0 P ij n + 1 < 0 according to (39) . Whence, from the definition of ℓ j i \ell_j^i ℓ j i , we observe that
Ψ i + ( U i L , n + 1 + ℓ j i P i j n + 1 ) = Ψ i + ( U i L , n + 1 + U i L , n + 1 − U i n , min − τ G i n − P i j n + 1 P i j n + 1 ) = Ψ i + ( U i n , min + τ G i n ) = 0. \begin{aligned}
\Psi_i^+\left(U_i^{\mathrm{L},n+1}+\ell_j^iP_{ij}^{n+1}\right)
&=\Psi_i^+\left(U_i^{\mathrm{L},n+1}+
\frac{U_i^{\mathrm{L},n+1}-U_i^{n,\min}-\tau G_i^n}{-P_{ij}^{n+1}}P_{ij}^{n+1}\right)\\
&=\Psi_i^+\left(U_i^{n,\min}+\tau G_i^n\right)=0.
\end{aligned} Ψ i + ( U i L , n + 1 + ℓ j i P ij n + 1 ) = Ψ i + ( U i L , n + 1 + − P ij n + 1 U i L , n + 1 − U i n , m i n − τ G i n P ij n + 1 ) = Ψ i + ( U i n , m i n + τ G i n ) = 0.
When U i n , max + τ G i n < U i L , n + 1 + P i j n + 1 U_i^{n,\max}+\tau G_i^n<U_i^{\mathrm{L},n+1}+P_{ij}^{n+1} U i n , m a x + τ G i n < U i L , n + 1 + P ij n + 1 , we have this time that P i j n + 1 > 0 P_{ij}^{n+1}>0 P ij n + 1 > 0 and from the definition of ℓ j i \ell_j^i ℓ j i we deduce that
Ψ i + ( U i L , n + 1 + ℓ j i P i j n + 1 ) = Ψ i + ( U i L , n + 1 + U i n , max + τ G i n − U i L , n + 1 P i j n + 1 P i j n + 1 ) = Ψ i + ( U i n , max + τ G i n ) ≥ 0. \begin{aligned}
\Psi_i^+\left(U_i^{\mathrm{L},n+1}+\ell_j^iP_{ij}^{n+1}\right)
&=\Psi_i^+\left(U_i^{\mathrm{L},n+1}+
\frac{U_i^{n,\max}+\tau G_i^n-U_i^{\mathrm{L},n+1}}{P_{ij}^{n+1}}P_{ij}^{n+1}\right)\\
&=\Psi_i^+\left(U_i^{n,\max}+\tau G_i^n\right)\ge0.
\end{aligned} Ψ i + ( U i L , n + 1 + ℓ j i P ij n + 1 ) = Ψ i + ( U i L , n + 1 + P ij n + 1 U i n , m a x + τ G i n − U i L , n + 1 P ij n + 1 ) = Ψ i + ( U i n , m a x + τ G i n ) ≥ 0.
In cases 2 and 3, thanks to 0 ≤ ℓ i j ≤ ℓ j i 0 \le\ell_{ij} \le\ell_j^i 0 ≤ ℓ ij ≤ ℓ j i and 0 ≤ ℓ j i < 1 0 \le\ell_j^i < 1 0 ≤ ℓ j i < 1 (if ℓ j i = 0 \ell_j^i = 0 ℓ j i = 0 , then ℓ i j = 0 \ell_{ij} = 0 ℓ ij = 0 and there is nothing to prove),
U i L , n + 1 + ℓ i j P i j n + 1 = ( 1 − ℓ i j ℓ j i ) U i L , n + 1 + ℓ i j ℓ j i ( U i L , n + 1 + ℓ j i P i j n + 1 ) , U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1}=\left(1-\frac{\ell_{ij}}{\ell_j^i}\right)U_i^{\mathrm{L},n+1}+\frac{\ell_{ij}}{\ell_j^i}\left(U_i^{\mathrm{L},n+1}+\ell_j^iP_{ij}^{n+1}\right), U i L , n + 1 + ℓ ij P ij n + 1 = ( 1 − ℓ j i ℓ ij ) U i L , n + 1 + ℓ j i ℓ ij ( U i L , n + 1 + ℓ j i P ij n + 1 ) ,
is a convex combination between U i L , n + 1 U_i^{\mathrm{L},n+1} U i L , n + 1 and U i L , n + 1 + ℓ j i P i j n + 1 U_i^{\mathrm{L},n+1}+\ell_j^iP_{ij}^{n+1} U i L , n + 1 + ℓ j i P ij n + 1 . It remains to invoke again that the low-order approximation satisfies the discrete maximum principle, i.e., Ψ i + ( U i L , n + 1 ) ≥ 0 \Psi_i^+(U_i^{\mathrm{L},n+1}) \ge0 Ψ i + ( U i L , n + 1 ) ≥ 0 to deduce that Ψ i + ( U i L , n + 1 + ℓ i j P i j n + 1 ) ≥ 0 \Psi_i^+(U_i^{\mathrm{L},n+1}+\ell_{ij}P_{ij}^{n+1}) \ge0 Ψ i + ( U i L , n + 1 + ℓ ij P ij n + 1 ) ≥ 0 . □
Numerical examples
In this section, we test the performance of the proposed algorithm on several benchmarks. The first benchmark is designed to test the fractional diffusion term while the second benchmark incorporates the Burgers-type hyperbolic term f ( u ) = ( 1 2 u 2 , 1 2 u 2 ) T f(u) = (\frac{1}{2}u^2,\frac{1}{2}u^2)^{\mathsf{T}} f ( u ) = ( 2 1 u 2 , 2 1 u 2 ) T . In both cases, exact solutions can be derived upon appropriately choosing g g g .
Finally, we apply the proposed method to the Surface Quasi-Geostrophic (SQG) equation. The SQG model is well known and widely used to study large-scale atmospheric and oceanic fluid dynamics. Common numerical approaches for solving SQG equations include spectral methods [12 , 13 ] . A finite element method for the (mostly) inviscid SQG equation was first proposed in [2 ] . By solving a standard benchmark problem, we validate that the proposed method is accurate and preserves the maximum principle for several values of the dissipation coefficient.
The implementation uses the C++ interface of the FEniCS project [27 ] . We use the third-order strong stability preserving explicit Runge–Kutta method SSPRK(3, 3) of § 4 .
Fractional heat equation
To validate our implementation of the fractional Laplacian, we consider the fractional heat equation
{ ∂ t u ( x , t ) + κ ( − Δ ) s u ( x , t ) = 0 , ( x , t ) ∈ D × ( 0 , T ] , u ( x , 0 ) = u 0 ( x ) : = sin x 2 cos x 1 , x ∈ D , (41) \begin{cases}
\partial_t u(x,t)+\kappa(-\Delta)^s u(x,t)=0, & (x,t)\in D\times(0,T],\\
u(x,0)=u_0(x):=\sin x_2\cos x_1, & x\in D, \tag*{(41)}
\end{cases} { ∂ t u ( x , t ) + κ ( − Δ ) s u ( x , t ) = 0 , u ( x , 0 ) = u 0 ( x ) := sin x 2 cos x 1 , ( x , t ) ∈ D × ( 0 , T ] , x ∈ D , (41)
with D = [ 0 , 2 π ] 2 D=[0,2\pi]^2 D = [ 0 , 2 π ] 2 and periodic boundary conditions in both the x 1 x_1 x 1 - and x 2 x_2 x 2 -directions. The time step is set according to (21) with CFL = 0.4 =0.4 = 0.4 .
In this setting, the exact solution is given by
u ( x , t ) : = e − 2 s κ t sin x 2 cos x 1 . u(x,t):=e^{-2^s\kappa t}\sin x_2\cos x_1. u ( x , t ) := e − 2 s κ t sin x 2 cos x 1 .
We consider two values for the fractional power: s = 1 4 s=\frac{1}{4} s = 4 1 and s = 1 2 s=\frac{1}{2} s = 2 1 . The results of the numerical computations are presented in Table 5.1 for both the limited and unlimited cases. The unlimited case corresponds to the scheme (34) , while the limited solution is obtained using the convex limiting (37) . The sequence of unstructured meshes is generated using the mesh generator Gmsh [18 ] and is constructed to satisfy the acute angle condition.
Table 5.1. Fractional heat equation. Convergence history in the L 1 L^1 L 1 - and L 2 L^2 L 2 -norms for simulations using limited and unlimited approximations at T = 0.5 T=0.5 T = 0.5 and κ = 0.001 \kappa=0.001 κ = 0.001 , with two different fractional diffusion powers s = 1 4 s=\frac{1}{4} s = 4 1 and s = 1 2 s=\frac{1}{2} s = 2 1 .
We observe second-order convergence in both the L 1 L^1 L 1 - and L 2 L^2 L 2 -norms for the limited and unlimited simulations (the rates are computed with respect to the mesh size h ∝ ( # d o f s ) − 1 / 2 h\propto(\#\mathrm{dofs})^{-1/2} h ∝ ( # dofs ) − 1/2 ). Furthermore, the errors are nearly identical for both approaches on fine meshes, suggesting that the nonlinear viscosity approximation (34) essentially satisfies the bounds (22) , so that the limiter is rarely active in this regime.
Although not reported here, we observe that the convergence rate is not optimal in the L ∞ L^\infty L ∞ -norm for this set of unstructured meshes. This is in contrast to the optimal second-order rate observed in [2 Sec. 4.2] when using uniform meshes.
Burgers' equation with fractional Laplacian
We consider the following two-dimensional fractional diffusion Burgers’ equation:
∂ t u + ∇ ⋅ ( u 2 2 , u 2 2 ) T + κ ( − Δ ) s u = g , ( x , t ) ∈ D × ( 0 , T ] , u ( x , 0 ) = u 0 ( x ) , x ∈ D , (42) \begin{aligned}
\partial_t u+\nabla\cdot\left(\frac{u^2}{2},\frac{u^2}{2}\right)^{\mathsf T}+\kappa(-\Delta)^s u&=g,\qquad(x,t)\in D\times(0,T],\\
u(x,0)&=u_0(x),\qquad x\in D, \tag*{(42)}
\end{aligned} ∂ t u + ∇ ⋅ ( 2 u 2 , 2 u 2 ) T + κ ( − Δ ) s u u ( x , 0 ) = g , ( x , t ) ∈ D × ( 0 , T ] , = u 0 ( x ) , x ∈ D , (42)
where the domain is D = [ 0 , 2 π ] 2 D=[0,2\pi]^2 D = [ 0 , 2 π ] 2 with periodic boundary conditions, s = 1 2 s=\frac{1}{2} s = 2 1 , and T = 0.5 T=0.5 T = 0.5 . The time step is set according to (21) with CFL = 0.4 =0.4 = 0.4 . In addition, we let
g ( x , t ) : = 1 2 e − 2 s + 1 κ t ( − sin ( 2 x 1 ) sin 2 x 2 + cos 2 x 1 sin ( 2 x 2 ) ) , g(x,t):=\frac{1}{2}e^{-2s+1\kappa t}\left(-\sin(2x_1)\sin^2x_2+\cos^2x_1\sin(2x_2)\right), g ( x , t ) := 2 1 e − 2 s + 1 κ t ( − sin ( 2 x 1 ) sin 2 x 2 + cos 2 x 1 sin ( 2 x 2 ) ) ,
corresponding to the manufactured solution u ( x , t ) : = e − 2 s κ t sin x 2 cos x 1 u(x,t):=e^{-2s\kappa t}\sin x_2\cos x_1 u ( x , t ) := e − 2 s κ t sin x 2 cos x 1 .
The convergence history in two different norms is reported in Table 5.2 for two diffusion coefficients: κ = 0.01 \kappa=0.01 κ = 0.01 and κ = 0.001 \kappa=0.001 κ = 0.001 . The same sequence of unstructured meshes as in § 5.1 is used. We observe second-order convergence rates in both norms, as reported in the table.
Table 5.2. Fractional diffusion Burgers’ equation. Convergence history in the L 1 L^1 L 1 - and L 2 L^2 L 2 -norms for simulations using limited and unlimited approximations at T = 0.5 T = 0.5 T = 0.5 and s = 1 2 s = \frac{1}{2} s = 2 1 , with two different fractional diffusion coefficients κ = 0.01 \kappa= 0.01 κ = 0.01 and κ = 0.001 \kappa= 0.001 κ = 0.001 . Unlimited solution; Limited solution.
Surface Quasi-Geostrophic model
Consider the following two-dimensional Surface Quasi-Geostrophic (SQG) equation:
∂ t u + ∇ ⋅ ( v u ) + κ ( − Δ ) s u = 0 , ( x , t ) ∈ D × ( 0 , T ] , u ( x , 0 ) = u 0 ( x ) , x ∈ D , (43) \begin{aligned}
\partial_tu+\nabla\cdot(vu)+\kappa(-\Delta)^su&=0,\qquad(x,t)\in D\times(0,T],\\
u(x,0)&=u_0(x),\qquad x\in D, \tag*{(43)}
\end{aligned} ∂ t u + ∇ ⋅ ( v u ) + κ ( − Δ ) s u u ( x , 0 ) = 0 , ( x , t ) ∈ D × ( 0 , T ] , = u 0 ( x ) , x ∈ D , (43)
where u 0 ( x ) u_0(x) u 0 ( x ) is a given initial condition and periodic boundary conditions are imposed in both directions. The SQG model is widely used in geophysical fluid dynamics, in particular for atmospheric and oceanographic simulations. In (43) u u u denotes the surface temperature and κ ≥ 0 \kappa\ge0 κ ≥ 0 models the so-called Ekman pumping effect. The velocity field is defined via a fractional derivative of the solution u u u :
v = ∇ ⊥ ( − Δ ) − 1 2 u , (44) v = \nabla^\perp(-\Delta)^{-\frac{1}{2}}u, \tag*{(44)} v = ∇ ⊥ ( − Δ ) − 2 1 u , (44)
where for any scalar function w : R 2 → R w : \mathbb{R}^2 \to\mathbb{R} w : R 2 → R the orthogonal gradient operator is defined as ∇ ⊥ w : = ( − ∂ x 2 w , ∂ x 1 w ) T \nabla^\perp w := (-\partial_{x_2}w,\partial_{x_1}w)^\mathsf{T} ∇ ⊥ w := ( − ∂ x 2 w , ∂ x 1 w ) T . The negative fractional power can be computed with an approach similar to that of Section 2. A detailed discussion can be found in [2 Sec. 3.2.1] .
In this final numerical demonstration, we set D = [ 0 , 2 π ] 2 D = [0,2\pi]^2 D = [ 0 , 2 π ] 2 and use the initial condition u 0 ( x ) = sin ( x 1 ) sin ( x 2 ) + cos ( x 2 ) u_0(x) = \sin(x_1)\sin(x_2) + \cos(x_2) u 0 ( x ) = sin ( x 1 ) sin ( x 2 ) + cos ( x 2 ) . This initial condition is smooth, as shown in Figure 5.1 (a). This benchmark was previously used in [12 , 31 ] to investigate finite-time singularity formation in the solution of the SQG equation.
Figure 5.1. Snapshots of the solution of the viscous SQG with κ = 0.001 \kappa= 0.001 κ = 0.001 . The mesh consists of 148318 P 1 \mathrm{P}_1 P 1 nodes. Sharp layers are developing around time t = 8 t = 8 t = 8 .
Temporal snapshots of the numerical solution obtained using convex limiting with κ = 0.001 \kappa= 0.001 κ = 0.001 on a mesh consisting of 148318 vertices are presented in Figure 5.1 . We observe the development of a sharp layer around time t ≈ 8 t \approx8 t ≈ 8 . The algorithm successfully captures the formation of sharp layers and small-scale instabilities with mesh resolution comparable to those in the cited references, while also preserving the maximum principle.
In Figure 5.2 , we present solutions for different Ekman pumping effects. The time evolution of ∥ ∇ u ( t ) ∥ L 2 ( D ) \lVert\nabla u(t)\rVert_{L^2(D)} ∥ ∇ u ( t ) ∥ L 2 ( D ) for various values of κ \kappa κ is shown in Figure 5.3 . As reported in the literature [12 , 31 , 2 ] , we also observe a sharp increase in the gradient norm around times t = 7 t = 7 t = 7 and t = 8 t = 8 t = 8 for κ = 0.001 \kappa= 0.001 κ = 0.001 and κ = 0.01 \kappa= 0.01 κ = 0.01 . For more diffusive cases, no sharp layer forms, and the gradient norm decays over time.
Figure 5.2. Snapshots of the solution to the viscous SQG equation for various values of κ \kappa κ . The mesh consists of 148318 P 1 \mathrm{P}_1 P 1 nodes. The top row shows the solution at time t = 8 t = 8 t = 8 , while the bottom row depicts the solution at the final time t = 20 t = 20 t = 20 .
Figure 5.3. Viscous SQG with sharp layer: Evolution of ∥ ∇ u ( t ) ∥ L 2 ( D ) \lVert\nabla u(t)\rVert_{L^2(D)} ∥ ∇ u ( t ) ∥ L 2 ( D ) for different viscosity coefficients on the mesh consisting of 148318 P 1 \mathrm{P}_1 P 1 nodes.
Concluding remarks
We have proposed a fully discrete, explicit, continuous finite element approximation of scalar conservation equations with a fractional Laplacian that preserves the maximum principle. The fractional Laplacian is approximated by a sinc quadrature of a Balakrishnan-type integral representation, which reduces its evaluation to a family of independent reaction–diffusion problems. On meshes satisfying the acute-angle condition, the discrete resolvents are monotone, and the low-order scheme with graph viscosity satisfies a discrete maximum principle under an explicit time-step restriction that accounts for both transport and fractional diffusion. A second-order approximation, which is mass conservative in the periodic setting, is then obtained by convex limiting between this low-order scheme and a high-order scheme with entropy-based viscosity. The numerical experiments confirm second-order convergence on unstructured meshes and show that the method captures the formation of sharp fronts in the SQG model while preserving the maximum principle.
The time-step restriction associated with the fractional term scales like h 2 s h^{2s} h 2 s , independently of the number of quadrature nodes; it is compatible with the hyperbolic CFL condition for s ≤ 1 2 s \le\frac{1}{2} s ≤ 2 1 but becomes restrictive for s > 1 2 s > \frac{1}{2} s > 2 1 . Treating the fractional term implicitly while retaining the maximum principle, and relaxing the acute-angle condition, are natural directions for future work.
Declarations
Funding. This work was supported in part by NSF grant DMS-2409807 (AB) and by the Swedish Research Council (VR) grant 2021-04620 (MN).
Competing interests. The authors have no relevant financial or non-financial interests to disclose.
Data availability. The data supporting the findings of this study (convergence histories and figures) can be reproduced and are available from the corresponding author upon reasonable request.
Code availability. The code used to produce the numerical results is available from the corresponding author upon reasonable request.