Introduction
Problem Setting. Let I = ( 0 , T ) I=(0,T) I = ( 0 , T ) be a finite time interval, let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a Lipschitz domain, and let n s y s n_{\mathrm{sys}} n sys denote the number of components. Set H = [ L 2 ( Ω ) ] n s y s \mathcal{H}=[L^2(\Omega)]^{n_{\mathrm{sys}}} H = [ L 2 ( Ω ) ] n sys , with inner product ( ⋅ , ⋅ ) Ω (\cdot,\cdot)_{\Omega} ( ⋅ , ⋅ ) Ω , and let V ⊂ [ H 1 ( Ω ) ] n s y s \mathcal{V}\subset[H^1(\Omega)]^{n_{\mathrm{sys}}} V ⊂ [ H 1 ( Ω ) ] n sys be a dense subspace that incorporates the boundary conditions. We assume that the time-independent first-order differential operator G : V → H G:\mathcal{V}\to\mathcal{H} G : V → H is skew-symmetric on V \mathcal{V} V , that is,
( G U , V ) Ω = − ( U , G V ) Ω ∀ U , V ∈ V (1) (GU,V)_{\Omega}=-(U,GV)_{\Omega}\quad\forall U,V\in\mathcal{V} \tag*{(1)} ( G U , V ) Ω = − ( U , G V ) Ω ∀ U , V ∈ V (1)
Given F : I → H F:I\to\mathcal{H} F : I → H and U 0 ∈ H U_0\in\mathcal{H} U 0 ∈ H , we seek U : [ 0 , T ] → H U:[0,T]\to\mathcal{H} U : [ 0 , T ] → H , with U ( t ) ∈ V U(t)\in\mathcal{V} U ( t ) ∈ V for t ∈ I t\in I t ∈ I , that satisfies the Friedrichs-type initial-value problem
∂ t U + G U = F in I (2) \partial_t U+GU=F\quad\text{in }I \tag*{(2)} ∂ t U + G U = F in I (2)
U ( 0 ) = U 0 (3) U(0)=U_0 \tag*{(3)} U ( 0 ) = U 0 (3)
For the stationary theory of Friedrichs systems, we refer to [18 , 20 ] ; time-dependent formulations are considered in [4 , 8 ] . In several space dimensions the operator acts on coupled fields, so the analysis must control the interaction between multistep time differences and the full system operator rather than a single scalar transport direction.
Skew-symmetry endows the continuous problem with a natural L 2 L^2 L 2 -energy and yields energy stability for standard Galerkin finite element semidiscretisations. For smooth solutions, however, the resulting error estimates are typically suboptimal by one order. Stabilised methods commonly recover half an order and, more importantly for wave propagation, better preserve the local propagation of perturbations. The resulting discretisation errors can therefore reflect more closely the dynamics of perturbations in the continuous system, whereas dispersive errors generated by an unstabilised Galerkin method may pollute spatially separated regions.
The Transient Stabilisation Problem. For a conforming finite element space V h ⊂ V V_h \subset V V h ⊂ V , the consistent semidiscrete Streamline Diffusion, or streamline upwind/Petrov–Galerkin (SUPG), formulation takes the following form: find U h ( t ) ∈ V h U_h(t) \in V_h U h ( t ) ∈ V h , with U h ( 0 ) = U h , 0 U_h(0) = U_{h,0} U h ( 0 ) = U h , 0 , where U h , 0 U_{h,0} U h , 0 is a suitable approximation of U 0 U_0 U 0 , such that
∫ Ω ( ∂ t U h + G U h ) ⋅ ( V h + δ G V h ) d x = ∫ Ω F ⋅ ( V h + δ G V h ) d x ∀ V h ∈ V h (4) \int_{\Omega} (\partial_t U_h + G U_h) \cdot(V_h + \delta G V_h)\,dx = \int_{\Omega} F \cdot(V_h + \delta G V_h)\,dx \quad\forall V_h \in V_h \tag*{(4)} ∫ Ω ( ∂ t U h + G U h ) ⋅ ( V h + δ G V h ) d x = ∫ Ω F ⋅ ( V h + δ G V h ) d x ∀ V h ∈ V h (4)
Here δ > 0 \delta> 0 δ > 0 is the stabilisation parameter, with the dimension of time. Consistency requires the full evolution residual to be paired with the perturbed test function. The resulting cross term ( ∂ t U h , δ G V h ) Ω (\partial_t U_h,\delta G V_h)_{\Omega} ( ∂ t U h , δ G V h ) Ω is the principal obstacle in the transient energy analysis.
This analytical difficulty is coupled to the algebraic character of the time step. For a general stabilisation parameter, an implicit Streamline Diffusion step is nonsymmetric. Explicit Petrov–Galerkin schemes instead inherit a generally nonsymmetric mass matrix and are subject to a hyperbolic Courant–Friedrichs–Lewy (CFL) condition. This distinction is especially relevant for nonlinear problems, where implicit methods require a nonlinear solve at every time step.
In the framework developed below, a method average with current coefficient b 0 b_0 b 0 admits the distinguished parameter choice δ = b 0 τ \delta= b_0\tau δ = b 0 τ . This choice arises from local-in-time residual minimisation [5 , 21 ] and makes the linear system at each time step a symmetric positive definite normal equation in the corresponding graph norm. Other choices of δ \delta δ remain useful for the stability analysis but do not in general produce a symmetric system. For an explicit Adams–Bashforth method, b 0 = 0 b_0 = 0 b 0 = 0 , so no positive stabilisation parameter yields this normal-equation structure.
Main Contributions. We address these issues through a common operator and energy framework for transient Streamline Diffusion methods combined with the θ \theta θ -method and high-order Adams schemes. Beyond order two, the classical A A A -stability route for linear multistep methods is unavailable [16 ] ; stability must instead be recovered from method-dependent discrete energies that remain compatible with the spatial residual. The principal contributions are as follows. Here and below, h h h denotes the mesh size, k k k the spatial polynomial degree, τ \tau τ the time step, and ν \nu ν the temporal order.
The fully discrete formulation separates the method average, the backward time difference, the material residual, and the perturbed test operator. This organisation exposes a common algebraic identity and distinguishes the residual-minimisation choice δ = b 0 τ \delta= b_0\tau δ = b 0 τ from a general SUPG parameter.
A direct dual-residual estimate leads to a fully discrete finite element error bound of order O ( h k + 1 / 2 + τ ν ) O(h^{k+1/2}+\tau^\nu) O ( h k + 1/2 + τ ν ) when δ ∼ h \delta\sim h δ ∼ h and τ ≲ h \tau\lesssim h τ ≲ h , under the stated regularity assumptions. The proof uses the ordinary L 2 L^2 L 2 -projection, skew-symmetry, and discrete summation by parts.
The abstract energy hypothesis is verified for the θ \theta θ -method and for the Adams–Moulton (AM) methods AM3–AM5. For general polynomial degree, the AM3 and AM4 results require strengthened CFL conditions. For continuous piecewise affine elements, an argument
based on elementwise constant gradients recovers stability and the error estimate under the standard hyperbolic CFL condition when τ ≲ δ ≲ h \tau\lesssim\delta\lesssim h τ ≲ δ ≲ h . AM5 requires only the standard hyperbolic CFL condition for every degree covered by the abstract framework. Independent energy identities establish stability and error estimates for the Adams–Bashforth (AB) methods AB3 and AB4 under a standard hyperbolic CFL condition.
The experiments separate temporal accuracy from spatial approximation. An order-matched one-dimensional acoustic test pairs a method of temporal order ν \nu ν with P ν − 1 \mathbb{P}_{\nu-1} P ν − 1 elements and recovers orders two through five in the fully discrete final and graph errors; the material-residual errors follow the corresponding half-order sequence. A discontinuous solution constructed by d’Alembert’s formula shows second-order local P 1 \mathbb{P}_1 P 1 convergence for normal-equation SUPG in smooth regions separated from the wave fronts, whereas the unstabilised local error converges with order close to 1 / 2 1/2 1/2 . A smooth two-dimensional compact-wave test provides complementary evidence of improved localisation.
Previous Work. Hughes and Brooks introduced the Streamline Upwind Petrov–Galerkin method for stationary hyperbolic problems in [7 ] . Johnson and co-workers subsequently gave a comprehensive analysis for stationary linear transport in [25 ] , which also discusses the application of Streamline Diffusion to Friedrichs systems.
The extension to transient problems is more delicate because consistency requires the full space–time residual to be tested against the perturbed test function. The resulting cross term couples the discrete time derivative to the spatial differential operator. Bochev and co-workers established invertibility of the linear system in [6 ] , but their result did not provide uniform control of error growth in time. For finite-difference time discretisations, tuning the stabilisation parameter was proposed as a means of controlling the cross term [26 ] . Space–time finite element methods with discontinuous Galerkin approximation in time offered another route [24 , 25 ] , at the cost of larger, fully coupled nonsymmetric systems. Higher-order space–time Streamline Diffusion analyses are available for specialised kinetic systems [1 ] , while an upwind discontinuous Galerkin time discretisation with a space–time SUPG term has recently been analysed for scalar advection–diffusion [3 ] . These methods discretise time variationally and do not produce the separated linear multistep formulation considered here.
Discontinuous Galerkin methods combined with Runge–Kutta time integration were developed in [14 ] and became widely used for hyperbolic problems. For continuous finite elements, symmetric alternatives include subgrid viscosity [23 ] , orthogonal subscales [15 ] , local projection [2 ] , and continuous interior penalty stabilisation [12 ] . When combined with explicit Runge–Kutta methods, these approaches reproduce important stability properties of discontinuous Galerkin schemes [10 ] , although the stabilisation generally enlarges the spatial matrix stencil. High-order continuous finite elements with SUPG and Runge–Kutta or deferred-correction time integration have also been studied by fully discrete Fourier analysis on two-dimensional triangular meshes [27 ] . That analysis characterises linear stability and admissible CFL parameters, rather than deriving the system-level a priori energy and error estimates pursued here.
Uniform stability of the transient SUPG method for linear transport was established by Burman [9 ] . Testing with a linear combination of the solution and its time derivative yielded a stability estimate for any A A A -stable time discretisation, including Crank–Nicolson and the second-order backward differentiation formula. The analysis was subsequently extended to convection–diffusion problems [13 ] and time-dependent coefficients [28 ] . More recently, stability has been proved for third- and fourth-order Runge–Kutta methods combined with Streamline Diffusion in space [19 ] . Thus the closest rigorous results either stop at second-order linear multistep schemes or use Runge–Kutta, discontinuous-in-time, or coupled space–time formulations. To the best of our knowledge, the present work provides the first stability and a priori error analysis of strongly consistent Streamline Diffusion finite elements combined with linear multistep schemes of order greater than two for multidimensional first-order hyperbolic systems. This is not merely a change of time integrator: the loss of A A A -stability beyond order two leads to the method-dependent discrete energy mechanisms developed here for AM3–AM5 and AB3–AB4.
Outline. Section 2 introduces the operator setting, the time average and its adjoint, and the operators T δ T_\delta T δ , S δ S_\delta S δ , A τ A_\tau A τ , and P δ P_\delta P δ . Section 3 develops the abstract stability and error framework. Section 4 proves the θ \theta θ -method energy identity; Section 5 analyses AM3–AM5; and Section 6 derives the AB3 and AB4 energy and stability estimates. Section 7 presents the one- and two-dimensional acoustic experiments, and Section 8 summarises the results and open problems. Appendix A contains the horizon-uniform proof of the dual-residual estimate.
We formulate the method for a real Hilbert space H \mathcal{H} H with inner product ( ⋅ , ⋅ ) Ω (\cdot,\cdot)_\Omega ( ⋅ , ⋅ ) Ω and a dense subspace V ⊂ H \mathcal{V} \subset\mathcal{H} V ⊂ H . The notation introduced here is used in all subsequent sections. The relations a ≲ b a \lesssim b a ≲ b and a ∼ b a \sim b a ∼ b mean, respectively, that a ≤ C b a \le Cb a ≤ C b and c b ≤ a ≤ C b cb \le a \le Cb c b ≤ a ≤ C b , with positive constants c c c and C C C independent of the mesh size, time step, and final computed index unless stated otherwise.
Spatial Operator and Skew-Symmetry. Let G G G be a linear operator satisfying
G : V ⟶ H (5) G : \mathcal{V} \longrightarrow\mathcal{H} \tag*{(5)} G : V ⟶ H (5)
and assume that the boundary conditions and the choice of V \mathcal{V} V give
( G U , V ) Ω = − ( U , G V ) Ω ∀ U , V ∈ V (6) (GU,V)_\Omega= -(U,GV)_\Omega\quad\forall U,V \in\mathcal{V} \tag*{(6)} ( G U , V ) Ω = − ( U , G V ) Ω ∀ U , V ∈ V (6)
Consequently,
( G U , U ) Ω = 0 ∀ U ∈ V (7) (GU,U)_\Omega= 0 \quad\forall U \in\mathcal{V} \tag*{(7)} ( G U , U ) Ω = 0 ∀ U ∈ V (7)
The analysis below is for a time-independent operator. Thus, every time difference and time average commutes with G G G . Space- or time-dependent coefficients require additional commutator bounds and are not covered by the present statements.
Example: Acoustic Waves. Let c > 0 c > 0 c > 0 be constant. Starting from the physical variables p p p and v v v , introduce the scaled velocity u = c v u = cv u = c v . The acoustic system becomes
∂ t p + c ∇ ⋅ u = f in Ω × ( 0 , T ) (8) \partial_t p + c\nabla\cdot u = f \qquad\text{in } \Omega\times(0,T) \tag*{(8)} ∂ t p + c ∇ ⋅ u = f in Ω × ( 0 , T ) (8)
∂ t u + c ∇ p = c g in Ω × ( 0 , T ) (9) \partial_t u + c\nabla p = cg \qquad\text{in } \Omega\times(0,T) \tag*{(9)} ∂ t u + c ∇ p = c g in Ω × ( 0 , T ) (9)
Set U = ( p , u ) U = (p,u) U = ( p , u ) , F = ( f , c g ) F = (f,cg) F = ( f , c g ) , and
H = L 2 ( Ω ) × L 2 ( Ω ) d (10) \mathcal{H} = L^2(\Omega) \times L^2(\Omega)^d \tag*{(10)} H = L 2 ( Ω ) × L 2 ( Ω ) d (10)
with the product inner product. The operator is
G ( p , u ) = ( c ∇ ⋅ u , c ∇ p ) (11) G(p,u) = (c\nabla\cdot u,c\nabla p) \tag*{(11)} G ( p , u ) = ( c ∇ ⋅ u , c ∇ p ) (11)
For periodic boundary conditions, or reflecting-wall boundary conditions for which the boundary pairing vanishes, integration by parts proves (2.2). The corresponding norm is
∥ U ∥ Ω 2 = ∥ p ∥ Ω 2 + ∥ u ∥ Ω 2 = ∥ p ∥ Ω 2 + c 2 ∥ v ∥ Ω 2 (12) \lVert U\rVert_\Omega^2 = \lVert p\rVert_\Omega^2 + \lVert u\rVert_\Omega^2 = \lVert p\rVert_\Omega^2 + c^2\lVert v\rVert_\Omega^2 \tag*{(12)} ∥ U ∥ Ω 2 = ∥ p ∥ Ω 2 + ∥ u ∥ Ω 2 = ∥ p ∥ Ω 2 + c 2 ∥ v ∥ Ω 2 (12)
Thus the powers of c c c in (2.4)–(2.5), the operator, and the energy are consistent.
Time Differences, Averages, and Their Adjoints
Let N N N denote the final computed index, let τ > 0 \tau> 0 τ > 0 , and set t n = n τ t^n = n\tau t n = n τ for 0 ≤ n ≤ N 0 \le n \le N 0 ≤ n ≤ N , with t N ≤ T t^N \le T t N ≤ T . Define the backward and forward time differences by
D τ − U n = τ − 1 ( U n − U n − 1 ) (13) D_\tau^- U^n = \tau^{-1}(U^n-U^{n-1}) \tag*{(13)} D τ − U n = τ − 1 ( U n − U n − 1 ) (13)
D τ + U n = τ − 1 ( U n + 1 − U n ) (14) D_\tau^+ U^n = \tau^{-1}(U^{n+1}-U^n) \tag*{(14)} D τ + U n = τ − 1 ( U n + 1 − U n ) (14)
For coefficients b 0 , … , b ℓ b_0,\ldots,b_\ell b 0 , … , b ℓ , define the method average
M τ U n = ∑ j = 0 ℓ b j U n − j (15) \mathcal{M}_\tau U^n = \sum_{j=0}^{\ell} b_j U^{n-j} \tag*{(15)} M τ U n = j = 0 ∑ ℓ b j U n − j (15)
Thus ℓ + 1 \ell+1 ℓ + 1 is the stencil length, and b 0 b_0 b 0 is the coefficient of the current value. The same notation is used for sampled exact functions and for the data. The first computed index is denoted by n 0 n_0 n 0 , with n 0 ≥ ℓ n_0 \ge\ell n 0 ≥ ℓ , and all earlier values required by the stencil are prescribed start-up data.
The average M τ \mathcal{M}_\tau M τ is not self-adjoint on a finite time interval. Extend a test sequence by zero outside n 0 ≤ n ≤ N n_0 \le n \le N n 0 ≤ n ≤ N , and define its discrete adjoint average by
M τ ∗ V m = ∑ j = 0 ℓ b j V m + j (16) \mathcal{M}_\tau^* V^m = \sum_{j=0}^{\ell} b_j V^{m+j} \tag*{(16)} M τ ∗ V m = j = 0 ∑ ℓ b j V m + j (16)
A change of summation index gives the exact identity
∑ n = n 0 N ( M τ U n , V n ) Ω = ∑ m = n 0 − ℓ N ( U m , M τ ∗ V m ) Ω (17) \sum_{n=n_0}^{N}(\mathcal{M}_\tau U^n,V^n)_\Omega
=
\sum_{m=n_0-\ell}^{N}(U^m,\mathcal{M}_\tau^*V^m)_\Omega \tag*{(17)} n = n 0 ∑ N ( M τ U n , V n ) Ω = m = n 0 − ℓ ∑ N ( U m , M τ ∗ V m ) Ω (17)
In particular, M τ ∗ V m \mathcal{M}_\tau^*V^m M τ ∗ V m cannot be replaced by M τ V m \mathcal{M}_\tau V^m M τ V m without additional symmetry and endpoint terms. The corresponding summation-by-parts identity is
τ ∑ n = n 0 N ( D τ − U n , V n ) Ω = ( U N , V N ) Ω − ( U n 0 − 1 , V n 0 ) Ω (18) \tau\sum_{n=n_0}^{N}(D_\tau^-U^n,V^n)_\Omega
=
(U^N,V^N)_\Omega-(U^{n_0-1},V^{n_0})_\Omega \tag*{(18)} τ n = n 0 ∑ N ( D τ − U n , V n ) Ω = ( U N , V N ) Ω − ( U n 0 − 1 , V n 0 ) Ω (18)
− τ ∑ n = n 0 N − 1 ( U n , D τ + V n ) Ω (19) -\tau\sum_{n=n_0}^{N-1}(U^n,D_\tau^+V^n)_\Omega \tag*{(19)} − τ n = n 0 ∑ N − 1 ( U n , D τ + V n ) Ω (19)
Both endpoint terms in (2.14) are kept explicit whenever this formula is used.
Approximation and Local Truncation Error. Let T h \mathcal{T}_h T h be a shape-regular mesh of Ω \Omega Ω , set h = max K ∈ T h diam ( K ) h=\max_{K\in\mathcal{T}_h}\operatorname{diam}(K) h = max K ∈ T h diam ( K ) , and let P j ( K ) \mathbb{P}_j(K) P j ( K ) denote the scalar polynomials of total degree at most j j j on K K K . Let V h ⊂ V V_h\subset V V h ⊂ V be a conforming finite element space of componentwise polynomial degree at most k k k , and let Π h : H → V h \Pi_h:\mathcal{H}\to V_h Π h : H → V h be the L 2 L^2 L 2 -orthogonal projection. If s ≥ 1 s\ge1 s ≥ 1 is the spatial Sobolev regularity index, set r = min ( s , k + 1 ) r=\min(s,k+1) r = min ( s , k + 1 ) . We assume that
∥ U − Π h U ∥ Ω + h ∥ G ( U − Π h U ) ∥ Ω ≲ h r ∥ U ∥ H r ( Ω ) (20) \lVert U-\Pi_hU\rVert_\Omega+h\lVert G(U-\Pi_hU)\rVert_\Omega\lesssim h^r\lVert U\rVert_{H^r(\Omega)} \tag*{(20)} ∥ U − Π h U ∥ Ω + h ∥ G ( U − Π h U ) ∥ Ω ≲ h r ∥ U ∥ H r ( Ω ) (20)
Because Π h \Pi_h Π h is time independent, it commutes with the operators in (2.9) and (2.11). Define the local truncation residual by
ρ n = D τ − U ( t n ) + G M τ U ( t n ) − M τ F ( t n ) (21) \rho^n=D_\tau^-U(t^n)+G\mathcal{M}_\tau U(t^n)-\mathcal{M}_\tau F(t^n) \tag*{(21)} ρ n = D τ − U ( t n ) + G M τ U ( t n ) − M τ F ( t n ) (21)
For a method of temporal order ν \nu ν , we assume the standard local estimate
∥ ρ n ∥ Ω ≤ C τ ν − 1 / 2 C n , ν + 1 ( U , F ) (22) \lVert\rho^n\rVert_\Omega\le C\tau^{\nu-1/2}C_{n,\nu+1}(U,F) \tag*{(22)} ∥ ρ n ∥ Ω ≤ C τ ν − 1/2 C n , ν + 1 ( U , F ) (22)
where C n , ν + 1 ( U , F ) C_{n,\nu+1}(U,F) C n , ν + 1 ( U , F ) is an L 2 L^2 L 2 -in-time regularity norm on the stencil interval. The Adams estimate is derived in Section 5.1. Define the corresponding global regularity quantity by
C T , ν + 1 ( U , F ) = ( ∑ n = n 0 N C n , ν + 1 ( U , F ) 2 ) 1 / 2 (23) C_{T,\nu+1}(U,F)=\left(\sum_{n=n_0}^{N}C_{n,\nu+1}(U,F)^2\right)^{1/2} \tag*{(23)} C T , ν + 1 ( U , F ) = ( n = n 0 ∑ N C n , ν + 1 ( U , F ) 2 ) 1/2 (23)
The stencil intervals have uniformly bounded overlap (with a bound depending only on the fixed method), so this quantity is controlled by the appropriate global L 2 ( 0 , T ) L^2(0,T) L 2 ( 0 , T ) norm of the time derivatives occurring in the local remainder. In particular, (2.18) gives
∥ ρ ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ τ ν C T , ν + 1 ( U , F ) (24) \|\rho\|_{\ell^2_\tau(n_0,N;H)}\lesssim\tau^\nu C_{T,\nu+1}(U,F) \tag*{(24)} ∥ ρ ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ τ ν C T , ν + 1 ( U , F ) (24)
For the dual-residual estimate we use the corresponding differentiated consistency bounds
max n 0 ≤ n ≤ N ∥ ρ n ∥ Ω ≲ τ ν C T , ν + 2 ♯ ( U , F ) (25) \max_{n_0\le n\le N}\|\rho^n\|_{\Omega}\lesssim\tau^\nu C^\sharp_{T,\nu+2}(U,F) \tag*{(25)} n 0 ≤ n ≤ N max ∥ ρ n ∥ Ω ≲ τ ν C T , ν + 2 ♯ ( U , F ) (25)
∥ D τ + ρ ∥ ℓ τ 2 ( n 0 , N − 1 ; H ) ≲ τ ν C T , ν + 2 ♯ ( U , F ) (26) \|D^+_\tau\rho\|_{\ell^2_\tau(n_0,N-1;H)}\lesssim\tau^\nu C^\sharp_{T,\nu+2}(U,F) \tag*{(26)} ∥ D τ + ρ ∥ ℓ τ 2 ( n 0 , N − 1 ; H ) ≲ τ ν C T , ν + 2 ♯ ( U , F ) (26)
Here C T , ν + 2 ♯ ( U , F ) C^\sharp_{T,\nu+2}(U,F) C T , ν + 2 ♯ ( U , F ) contains one additional time derivative compared with C T , ν + 1 ( U , F ) C_{T,\nu+1}(U,F) C T , ν + 1 ( U , F ) , together with the endpoint regularity required for the maximum in (2.21). These estimates follow from the same integral-remainder argument as (2.18), applied also to the first forward difference of the remainder. They are verified below for the θ \theta θ - and Adams methods.
Shift and Method Operators
For a stabilisation parameter δ > 0 \delta>0 δ > 0 , define the spatial test operator, time shift, method operator, and combined test operator by
T δ = Id + δ G (27) T_\delta=\operatorname{Id}+\delta G \tag*{(27)} T δ = Id + δ G (27)
S δ = M τ + δ D τ − (28) S_\delta=M_\tau+\delta D^-_\tau \tag*{(28)} S δ = M τ + δ D τ − (28)
A τ = D τ − + G M τ (29) A_\tau=D^-_\tau+GM_\tau \tag*{(29)} A τ = D τ − + G M τ (29)
P δ = T δ S δ (30) P_\delta=T_\delta S_\delta \tag*{(30)} P δ = T δ S δ (30)
Commutation with G G G gives
P δ = M τ + δ D τ − + δ G M τ + δ 2 G D τ − (31) P_\delta=M_\tau+\delta D^-_\tau+\delta GM_\tau+\delta^2GD^-_\tau \tag*{(31)} P δ = M τ + δ D τ − + δ G M τ + δ 2 G D τ − (31)
Skew-symmetry gives the graph-inner-product identity
( T δ U , T δ V ) Ω = ( U , V ) Ω + δ 2 ( G U , G V ) Ω (32) (T_\delta U,T_\delta V)_\Omega=(U,V)_\Omega+\delta^2(GU,GV)_\Omega \tag*{(32)} ( T δ U , T δ V ) Ω = ( U , V ) Ω + δ 2 ( G U , G V ) Ω (32)
and hence
∥ T δ U ∥ Ω 2 = ∥ U ∥ Ω 2 + δ 2 ∥ G U ∥ Ω 2 (33) \|T_\delta U\|_\Omega^2=\|U\|_\Omega^2+\delta^2\|GU\|_\Omega^2 \tag*{(33)} ∥ T δ U ∥ Ω 2 = ∥ U ∥ Ω 2 + δ 2 ∥ G U ∥ Ω 2 (33)
We use the graph inner product and norm
( U , V ) δ = ( U , V ) Ω + δ 2 ( G U , G V ) Ω (34) (U,V)_\delta=(U,V)_\Omega+\delta^2(GU,GV)_\Omega \tag*{(34)} ( U , V ) δ = ( U , V ) Ω + δ 2 ( G U , G V ) Ω (34)
∥ U ∥ δ 2 = ( U , U ) δ = ∥ T δ U ∥ Ω 2 (35) \|U\|_\delta^2=(U,U)_\delta=\|T_\delta U\|_\Omega^2 \tag*{(35)} ∥ U ∥ δ 2 = ( U , U ) δ = ∥ T δ U ∥ Ω 2 (35)
The time-discrete SUPG method is: given the start-up values, find U n ∈ V h U^n\in V_h U n ∈ V h such that
( A τ U n , T δ V ) Ω = ( M τ F n , T δ V ) Ω ∀ V ∈ V h (36) (A_\tau U^n,T_\delta V)_\Omega=(M_\tau F^n,T_\delta V)_\Omega\quad\forall V\in V_h \tag*{(36)} ( A τ U n , T δ V ) Ω = ( M τ F n , T δ V ) Ω ∀ V ∈ V h (36)
This definition allows δ \delta δ to be chosen independently of the method coefficient b 0 b_0 b 0 .
Residual-Minimisation Case. At a fixed time level, consider
U n = argmin V n ∈ V h ∥ D τ − V n + G M τ V n − M τ F n ∥ Ω 2 (37) U^n = \underset{V^n \in\mathcal{V}_h}{\operatorname{argmin}} \left\|D_\tau^- V^n + G\mathcal{M}_\tau V^n - \mathcal{M}_\tau F^n\right\|_{\Omega}^{2} \tag*{(37)} U n = V n ∈ V h argmin D τ − V n + G M τ V n − M τ F n Ω 2 (37)
The derivative of the residual with respect to its current value is τ − 1 T b 0 τ \tau^{-1}T_{b_0\tau} τ − 1 T b 0 τ . Therefore the Euler–Lagrange equation is
( A τ U n , T b 0 τ V ) Ω = ( M τ F n , T b 0 τ V ) Ω ∀ V ∈ V h (38) (A_\tau U^n,T_{b_0\tau}V)_\Omega= (\mathcal{M}_\tau F^n,T_{b_0\tau}V)_\Omega\quad\forall V \in\mathcal{V}_h \tag*{(38)} ( A τ U n , T b 0 τ V ) Ω = ( M τ F n , T b 0 τ V ) Ω ∀ V ∈ V h (38)
Thus (36) is a normal equation precisely when δ = b 0 τ \delta= b_0\tau δ = b 0 τ .
Define the history part of the average by
M τ U n = ∑ j = 1 ℓ b j U n − j (39) \mathcal{M}_\tau U^n = \sum_{j=1}^{\ell} b_j U^{n-j} \tag*{(39)} M τ U n = j = 1 ∑ ℓ b j U n − j (39)
Separating the current value in (38) gives the update equation with the correct signs,
( T b 0 τ U n , T b 0 τ V ) Ω = ( U n − 1 − τ G M τ U n + τ M τ F n , T b 0 τ V ) Ω (40) (T_{b_0\tau}U^n,T_{b_0\tau}V)_\Omega= (U^{n-1}-\tau G\mathcal{M}_\tau U^n+\tau\mathcal{M}_\tau F^n,T_{b_0\tau}V)_\Omega \tag*{(40)} ( T b 0 τ U n , T b 0 τ V ) Ω = ( U n − 1 − τ G M τ U n + τ M τ F n , T b 0 τ V ) Ω (40)
The left-hand side is symmetric positive definite in the graph norm (33) . For a general stabilisation parameter δ ≠ b 0 τ \delta\ne b_0\tau δ = b 0 τ , the SUPG system need not be symmetric.
Acoustic-Wave Theta Method
For 1 / 2 ≤ θ ≤ 1 1/2 \le\theta\le1 1/2 ≤ θ ≤ 1 , set
M θ U n = θ U n + ( 1 − θ ) U n − 1 (41) \mathcal{M}_\theta U^n = \theta U^n + (1-\theta)U^{n-1} \tag*{(41)} M θ U n = θ U n + ( 1 − θ ) U n − 1 (41)
This is the θ \theta θ -method specialisation of the general average M τ \mathcal{M}_\tau M τ . For U h n = ( p h n , u h n ) U_h^n=(p_h^n,u_h^n) U h n = ( p h n , u h n ) , define the two current residual components by
R p n = D τ − p h n + c ∇ ⋅ M θ u h n − M θ f n (42) R_p^n = D_\tau^-p_h^n + c\nabla\cdot\mathcal{M}_\theta u_h^n-\mathcal{M}_\theta f^n \tag*{(42)} R p n = D τ − p h n + c ∇ ⋅ M θ u h n − M θ f n (42)
R u n = D τ − u h n + c ∇ M θ p h n − c M θ g n (43) R_u^n = D_\tau^-u_h^n+c\nabla\mathcal{M}_\theta p_h^n-c\mathcal{M}_\theta g^n \tag*{(43)} R u n = D τ − u h n + c ∇ M θ p h n − c M θ g n (43)
The component form of (36) is
( R p n , q h + δ c ∇ ⋅ w h ) Ω + ( R u n , w h + δ c ∇ q h ) Ω = 0 ∀ ( q h , w h ) ∈ V h (44) (R_p^n,q_h+\delta c\nabla\cdot w_h)_\Omega+(R_u^n,w_h+\delta c\nabla q_h)_\Omega=0 \quad\forall(q_h,w_h)\in\mathcal{V}_h \tag*{(44)} ( R p n , q h + δ c ∇ ⋅ w h ) Ω + ( R u n , w h + δ c ∇ q h ) Ω = 0 ∀ ( q h , w h ) ∈ V h (44)
All terms in (44) are evaluated at the current residual level n n n ; the previous level occurs only through D τ − D_\tau^- D τ − and M θ \mathcal{M}_\theta M θ .
When δ = θ τ \delta=\theta\tau δ = θ τ , define
H n − 1 = U h n − 1 − τ ( 1 − θ ) G U h n − 1 + τ M θ F n (45) H^{n-1}=U_h^{n-1}-\tau(1-\theta)GU_h^{n-1}+\tau\mathcal{M}_\theta F^n \tag*{(45)} H n − 1 = U h n − 1 − τ ( 1 − θ ) G U h n − 1 + τ M θ F n (45)
Equation (40) becomes
( p h n , q h ) Ω + ( u h n , w h ) Ω + δ 2 c 2 ( ∇ p h n , ∇ q h ) Ω (46) (p_h^n,q_h)_\Omega+(u_h^n,w_h)_\Omega+\delta^2c^2(\nabla p_h^n,\nabla q_h)_\Omega \tag*{(46)} ( p h n , q h ) Ω + ( u h n , w h ) Ω + δ 2 c 2 ( ∇ p h n , ∇ q h ) Ω (46)
+ δ 2 c 2 ( ∇ ⋅ u h n , ∇ ⋅ w h ) Ω = ( H n − 1 , T δ ( q h , w h ) ) Ω (47) +\delta^2c^2(\nabla\cdot u_h^n,\nabla\cdot w_h)_\Omega=(H^{n-1},T_\delta(q_h,w_h))_\Omega \tag*{(47)} + δ 2 c 2 ( ∇ ⋅ u h n , ∇ ⋅ w h ) Ω = ( H n − 1 , T δ ( q h , w h ) ) Ω (47)
The current pressure and scaled velocity are uncoupled on the left-hand side, and the two diagonal blocks are symmetric positive definite. This conclusion uses both the rescaling u = c v u=cv u = c v and the normal-equation choice δ = θ τ \delta=\theta\tau δ = θ τ .
Abstract Stability and Error Analysis
This section turns the common algebraic identity into a stability and error framework. We first introduce sequence norms that accommodate the different method-dependent energies. We then formulate an abstract energy hypothesis, derive the error equation, and estimate its residual in both a strong norm and a lifted dual norm. The final parameter balance makes the resulting spatial order explicit. The later method-specific sections verify the energy hypothesis by exact one-step identities or by discrete summation by parts.
Fundamental Algebraic Identity. In (36) , choose W = S δ U n W=S_\delta U^n W = S δ U n , which is admissible because S δ U n ∈ V h S_\delta U^n \in V_h S δ U n ∈ V h . Since P δ = T δ S δ P_\delta=T_\delta S_\delta P δ = T δ S δ , multiplication by τ \tau τ gives
( τ A τ U n , P δ U n ) Ω = ( τ M τ F n , P δ U n ) Ω (48) (\tau A_\tau U^n,P_\delta U^n)_\Omega=(\tau M_\tau F^n,P_\delta U^n)_\Omega \tag*{(48)} ( τ A τ U n , P δ U n ) Ω = ( τ M τ F n , P δ U n ) Ω (48)
Using (31) , skew-symmetry, and commutation of G G G with the time operators gives
( τ A τ U n , P δ U n ) Ω = τ ( D τ − U n , M τ U n ) Ω + τ δ ∥ A τ U n ∥ Ω 2 (49) (\tau A_\tau U^n,P_\delta U^n)_\Omega=\tau(D_\tau^-U^n,M_\tau U^n)_\Omega+\tau\delta\lVert A_\tau U^n\rVert_\Omega^2 \tag*{(49)} ( τ A τ U n , P δ U n ) Ω = τ ( D τ − U n , M τ U n ) Ω + τ δ ∥ A τ U n ∥ Ω 2 (49)
+ τ δ 2 ( G M τ U n , G D τ − U n ) Ω (50) +\tau\delta^2(GM_\tau U^n,GD_\tau^-U^n)_\Omega \tag*{(50)} + τ δ 2 ( G M τ U n , G D τ − U n ) Ω (50)
In particular, the material-residual contribution is τ δ ∥ A τ U n ∥ Ω 2 \tau\delta\lVert A_\tau U^n\rVert_\Omega^2 τ δ ∥ A τ U n ∥ Ω 2 . This fixes the energy and dissipation scaling used below.
Sequence Norms and Abstract Stability
The energy estimate must control both the solution at every computed time and the accumulated dissipation. Their precise forms depend on the time-stepping method, so we encode them abstractly. Let e τ n ( V ) ≥ 0 e_\tau^n(V) \ge0 e τ n ( V ) ≥ 0 be a method-dependent energy and let d τ n ( V ) ≥ 0 d_\tau^n(V) \ge0 d τ n ( V ) ≥ 0 be its dissipation. For a horizon n 0 ≤ m ≤ N n_0 \le m \le N n 0 ≤ m ≤ N , define
∥ V ∥ E , m 2 = max n 0 ≤ j ≤ m e τ j ( V ) + τ ∑ n = n 0 m d τ n ( V ) (51) \lVert V\rVert_{\mathcal{E},m}^2=\max_{n_0\le j\le m}e_\tau^j(V)+\tau\sum_{n=n_0}^{m}d_\tau^n(V) \tag*{(51)} ∥ V ∥ E , m 2 = n 0 ≤ j ≤ m max e τ j ( V ) + τ n = n 0 ∑ m d τ n ( V ) (51)
Thus the sequence norm contains only computed levels. The prescribed levels are recorded separately in the start-up energy. Specifically, the start-up energy and the discrete L 2 L^2 L 2 -in-time norm are
I τ ( V ) = ∑ j = 0 n 0 − 1 ∥ T δ V j ∥ Ω 2 (52) \mathcal{I}_\tau(V)=\sum_{j=0}^{n_0-1}\lVert T_\delta V^j\rVert_\Omega^2 \tag*{(52)} I τ ( V ) = j = 0 ∑ n 0 − 1 ∥ T δ V j ∥ Ω 2 (52)
∥ R ∥ ℓ τ 2 ( n 0 , m ; H ) 2 = τ ∑ n = n 0 m ∥ R n ∥ Ω 2 (53) \lVert R\rVert_{\ell_\tau^2(n_0,m;H)}^2=\tau\sum_{n=n_0}^{m}\lVert R^n\rVert_\Omega^2 \tag*{(53)} ∥ R ∥ ℓ τ 2 ( n 0 , m ; H ) 2 = τ n = n 0 ∑ m ∥ R n ∥ Ω 2 (53)
The forcing term in the energy identity is naturally paired with P δ V n P_\delta V^n P δ V n . We therefore define the corresponding global lifted dual norm by
∥ R ∥ E ∗ , N = max n 0 ≤ m ≤ N sup V n = 0 for n ∉ { 0 , … , m } I τ ( V ) + ∥ V ∥ E , m > 0 τ ∑ n = n 0 m ( R n , P δ V n ) Ω ( I τ ( V ) + ∥ V ∥ E , m 2 ) 1 / 2 (54) \lVert R\rVert_{\mathcal{E}^*,N}=\max_{n_0\le m\le N}\sup_{\substack{V^n=0\ \text{for }n\notin\{0,\ldots,m\}\\ \mathcal{I}_\tau(V)+\lVert V\rVert_{\mathcal{E},m}>0}}\frac{\tau\sum_{n=n_0}^{m}(R^n,P_\delta V^n)_\Omega}{\left(\mathcal{I}_\tau(V)+\lVert V\rVert_{\mathcal{E},m}^2\right)^{1/2}} \tag*{(54)} ∥ R ∥ E ∗ , N = n 0 ≤ m ≤ N max V n = 0 for n ∈ / { 0 , … , m } I τ ( V ) + ∥ V ∥ E , m > 0 sup ( I τ ( V ) + ∥ V ∥ E , m 2 ) 1/2 τ ∑ n = n 0 m ( R n , P δ V n ) Ω (54)
The maximum over horizons is needed because the sequence norm contains a maximum in time. The finite start-up stencil is retained and controlled by I τ ( V ) \mathcal{I}_\tau(V) I τ ( V ) , and the test sequence is extended by zero outside 0 ≤ n ≤ m 0\le n\le m 0 ≤ n ≤ m before P δ P_\delta P δ is applied. The positive-denominator condition removes sequences invisible to the horizon norm. The factor τ \tau τ matches the tested equation (48) .
Energy Hypothesis. The hypothesis has two parts: coercivity of the method-dependent energy and dissipation, and a forced energy estimate uniform in the terminal horizon. We assume that there are constants c E , C E > 0 c_E,C_E>0 c E , C E > 0 , independent of h h h , τ \tau τ , and m m m , such that
c E ∥ T δ V m ∥ Ω 2 ≤ e τ m ( V ) (55) c_E\lVert T_\delta V^m\rVert_\Omega^2 \le e_\tau^m(V) \tag*{(55)} c E ∥ T δ V m ∥ Ω 2 ≤ e τ m ( V ) (55)
c E δ ∥ A τ V n ∥ Ω 2 ≤ d τ n ( V ) (56) c_E\delta\lVert A_\tau V^n\rVert_\Omega^2 \le d_\tau^n(V) \tag*{(56)} c E δ ∥ A τ V n ∥ Ω 2 ≤ d τ n ( V ) (56)
and every solution of
( A τ V n , T δ W ) Ω = ( R n , T δ W ) Ω ∀ W ∈ V h (57) (A_\tau V^n,T_\delta W)_\Omega=(R^n,T_\delta W)_\Omega\quad\forall W\in\mathcal{V}_h \tag*{(57)} ( A τ V n , T δ W ) Ω = ( R n , T δ W ) Ω ∀ W ∈ V h (57)
satisfies, for every m m m ,
e τ m ( V ) + τ ∑ n = n 0 m d τ n ( V ) ≤ C E I τ ( V ) + C E τ ∑ n = n 0 m ( R n , P δ V n ) Ω (58) e_\tau^m(V)+\tau\sum_{n=n_0}^{m}d_\tau^n(V)\le C_E I_\tau(V)+C_E\tau\sum_{n=n_0}^{m}(R^n,P_\delta V^n)_\Omega \tag*{(58)} e τ m ( V ) + τ n = n 0 ∑ m d τ n ( V ) ≤ C E I τ ( V ) + C E τ n = n 0 ∑ m ( R n , P δ V n ) Ω (58)
+ C E τ ∥ R ∥ ℓ τ 2 ( n 0 , m ; H ) 2 (59) {}+C_E\tau\lVert R\rVert_{\ell_\tau^2(n_0,m;\mathcal{H})}^2 \tag*{(59)} + C E τ ∥ R ∥ ℓ τ 2 ( n 0 , m ; H ) 2 (59)
The right-hand side separates start-up data, the forcing functional measured by the lifted dual norm above, and a strong forcing defect. The last term permits the endpoint defects produced when increments are estimated in a multistep method; it is absent from the θ \theta θ -method identity.
Theorem 3.1 (Abstract stability). Assume (55) –(59) . Then
∥ V ∥ E , N 2 ≲ I τ ( V ) + ∥ R ∥ E ∗ , N 2 + τ ∥ R ∥ ℓ τ 2 ( n 0 , N ; H ) 2 (60) \lVert V\rVert_{\mathcal{E},N}^2\lesssim I_\tau(V)+\lVert R\rVert_{\mathcal{E}^*,N}^2+\tau\lVert R\rVert_{\ell_\tau^2(n_0,N;\mathcal{H})}^2 \tag*{(60)} ∥ V ∥ E , N 2 ≲ I τ ( V ) + ∥ R ∥ E ∗ , N 2 + τ ∥ R ∥ ℓ τ 2 ( n 0 , N ; H ) 2 (60)
where the hidden constant is independent of h h h and τ \tau τ .
Proof. Choose m ∗ m_* m ∗ , n 0 ≤ m ∗ ≤ N n_0\le m_*\le N n 0 ≤ m ∗ ≤ N , at which the computed-level energy is maximal. The energy inequality with m = m ∗ m=m_* m = m ∗ controls the maximum-energy term in (51) , while the inequality with m = N m=N m = N controls the full dissipation sum. By (54) , the forcing contribution is bounded by
τ ∑ n = n 0 m ( R n , P δ V n ) Ω ≤ ∥ R ∥ E ∗ , N ( I τ ( V ) + ∥ V ∥ E , m 2 ) 1 / 2 (61) \tau\sum_{n=n_0}^{m}(R^n,P_\delta V^n)_\Omega\le\lVert R\rVert_{\mathcal{E}^*,N}\left(I_\tau(V)+\lVert V\rVert_{\mathcal{E},m}^2\right)^{1/2} \tag*{(61)} τ n = n 0 ∑ m ( R n , P δ V n ) Ω ≤ ∥ R ∥ E ∗ , N ( I τ ( V ) + ∥ V ∥ E , m 2 ) 1/2 (61)
Add the two horizon inequalities. Their left-hand sides control ∥ V ∥ E , N 2 \lVert V\rVert_{\mathcal{E},N}^2 ∥ V ∥ E , N 2 up to a fixed constant. After bounding each horizon norm by ∥ V ∥ E , N \lVert V\rVert_{\mathcal{E},N} ∥ V ∥ E , N , Young’s inequality with a sufficiently small parameter absorbs the resulting contribution to ∥ V ∥ E , N 2 \lVert V\rVert_{\mathcal{E},N}^2 ∥ V ∥ E , N 2 . The accompanying start-up factor is bounded by C I τ ( V ) CI_\tau(V) C I τ ( V ) . This proves (60) . ■
Error Equation and Residual Bounds
We split the total error into an approximation error η n \eta^n η n , controlled by the projection estimates, and a discrete error ϕ n \phi^n ϕ n , to which the abstract stability theorem will be applied. Set
W n = Π h U ( t n ) , η n = U ( t n ) − W n , ϕ n = U h n − W n (62) W^n=\Pi_hU(t^n),\qquad\eta^n=U(t^n)-W^n,\qquad\phi^n=U_h^n-W^n \tag*{(62)} W n = Π h U ( t n ) , η n = U ( t n ) − W n , ϕ n = U h n − W n (62)
By (21) , the sampled exact solution satisfies
A τ U ( t n ) = M τ F ( t n ) + ρ n (63) A_\tau U(t^n)=\mathcal{M}_\tau F(t^n)+\rho^n \tag*{(63)} A τ U ( t n ) = M τ F ( t n ) + ρ n (63)
Subtracting the projected exact equation from the numerical method gives
( A τ ϕ n , T δ V ) Ω = ( R n , T δ V ) Ω ∀ V ∈ V h (64) (A_\tau\phi^n,T_\delta V)_\Omega=(R^n,T_\delta V)_\Omega\quad\forall V\in\mathcal{V}_h \tag*{(64)} ( A τ ϕ n , T δ V ) Ω = ( R n , T δ V ) Ω ∀ V ∈ V h (64)
with the residual
R n = M τ F n − A τ W n (65) R^n=\mathcal{M}_\tau F^n-A_\tau W^n \tag*{(65)} R n = M τ F n − A τ W n (65)
= A τ η n − ρ n (66) =A_\tau\eta^n-\rho^n \tag*{(66)} = A τ η n − ρ n (66)
The sign in (66) follows from (63) ; in particular, the truncation residual enters with a minus sign.
Strong Residual Bound. The abstract stability estimate contains the strong residual norm multiplied by τ 1 / 2 \tau^{1/2} τ 1/2 . The following bound controls that term; the sharper spatial rate will come from the dual estimate below. For v ∈ H 1 ( 0 , T ; H ) v \in H^1(0,T;\mathcal{H}) v ∈ H 1 ( 0 , T ; H ) , the endpoint trace inequality on each interval ( t n − 1 , t n ) (t^{n-1},t^n) ( t n − 1 , t n ) , followed by summation, gives
τ ∑ n = n 0 N ∥ v ( t n ) ∥ H 2 ≲ ∥ v ∥ L 2 ( 0 , T ; H ) 2 + τ 2 ∥ ∂ t v ∥ L 2 ( 0 , T ; H ) 2 (67) \tau\sum_{n=n_0}^{N} \|v(t^n)\|_{\mathcal{H}}^2 \lesssim\|v\|_{L^2(0,T;\mathcal{H})}^2+\tau^2\|\partial_t v\|_{L^2(0,T;\mathcal{H})}^2 \tag*{(67)} τ n = n 0 ∑ N ∥ v ( t n ) ∥ H 2 ≲ ∥ v ∥ L 2 ( 0 , T ; H ) 2 + τ 2 ∥ ∂ t v ∥ L 2 ( 0 , T ; H ) 2 (67)
For each fixed shift j j j in the finite stencil, changing the summation index in (67) gives
τ ∑ n = n 0 N ∥ v ( t n − j ) ∥ H 2 ≲ ∥ v ∥ L 2 ( 0 , T ; H ) 2 + τ 2 ∥ ∂ t v ∥ L 2 ( 0 , T ; H ) 2 (68) \tau\sum_{n=n_0}^{N} \|v(t^{n-j})\|_{\mathcal{H}}^2 \lesssim\|v\|_{L^2(0,T;\mathcal{H})}^2+\tau^2\|\partial_t v\|_{L^2(0,T;\mathcal{H})}^2 \tag*{(68)} τ n = n 0 ∑ N ∥ v ( t n − j ) ∥ H 2 ≲ ∥ v ∥ L 2 ( 0 , T ; H ) 2 + τ 2 ∥ ∂ t v ∥ L 2 ( 0 , T ; H ) 2 (68)
Here n 0 ≥ ℓ n_0 \geq\ell n 0 ≥ ℓ , so every sampled time belongs to [ 0 , T ] [0,T] [ 0 , T ] . Since the number of shifts and the method coefficients are fixed, it follows that
∥ M τ v ∥ ℓ 2 ( n 0 , N ; H ) ≲ ∥ v ∥ L 2 ( 0 , T ; H ) + τ ∥ ∂ t v ∥ L 2 ( 0 , T ; H ) (69) \|M_\tau v\|_{\ell^2(n_0,N;\mathcal{H})} \lesssim\|v\|_{L^2(0,T;\mathcal{H})}+\tau\|\partial_t v\|_{L^2(0,T;\mathcal{H})} \tag*{(69)} ∥ M τ v ∥ ℓ 2 ( n 0 , N ; H ) ≲ ∥ v ∥ L 2 ( 0 , T ; H ) + τ ∥ ∂ t v ∥ L 2 ( 0 , T ; H ) (69)
Apply (69) to v = G η = G ( Id − Π h ) U v=G\eta=G(\operatorname{Id}-\Pi_h)U v = G η = G ( Id − Π h ) U , which belongs to H 1 ( 0 , T ; H ) H^1(0,T;\mathcal{H}) H 1 ( 0 , T ; H ) under the regularity used below. The time independence of G G G and Π h \Pi_h Π h , together with (20) applied to U U U and ∂ t U \partial_tU ∂ t U , gives
∥ G M τ η ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (70) \|GM_\tau\eta\|_{\ell^2_\tau(n_0,N;\mathcal{H})} \lesssim h^{r-1}\|U\|_{L^2(0,T;H^r(\Omega))} \tag*{(70)} ∥ G M τ η ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω )) (70)
+ τ h r − 1 ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (71) +\tau h^{r-1}\|\partial_tU\|_{L^2(0,T;H^r(\Omega))} \tag*{(71)} + τ h r − 1 ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) (71)
On the other hand, writing D τ − η n D^-_\tau\eta^n D τ − η n as the average of ( Id − Π h ) ∂ t U (\operatorname{Id}-\Pi_h)\partial_tU ( Id − Π h ) ∂ t U over ( t n − 1 , t n ) (t^{n-1},t^n) ( t n − 1 , t n ) , and then using the Cauchy–Schwarz inequality, gives
∥ D τ − η ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (72) \|D^-_\tau\eta\|_{\ell^2_\tau(n_0,N;\mathcal{H})}\lesssim h^r\|\partial_tU\|_{L^2(0,T;H^r(\Omega))} \tag*{(72)} ∥ D τ − η ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) (72)
Consequently,
∥ A τ η ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (73) \|A_\tau\eta\|_{\ell^2_\tau(n_0,N;\mathcal{H})}\lesssim h^{r-1}\|U\|_{L^2(0,T;H^r(\Omega))} \tag*{(73)} ∥ A τ η ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω )) (73)
+ ( h r + τ h r − 1 ) ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (74) +(h^r+\tau h^{r-1})\|\partial_tU\|_{L^2(0,T;H^r(\Omega))} \tag*{(74)} + ( h r + τ h r − 1 ) ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) (74)
Together with (24) , this proves
∥ R ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω ) ) + ( h r + τ h r − 1 ) ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (75) \|R\|_{\ell^2_\tau(n_0,N;\mathcal{H})}\lesssim h^{r-1}\|U\|_{L^2(0,T;H^r(\Omega))}+(h^r+\tau h^{r-1})\|\partial_tU\|_{L^2(0,T;H^r(\Omega))} \tag*{(75)} ∥ R ∥ ℓ τ 2 ( n 0 , N ; H ) ≲ h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω )) + ( h r + τ h r − 1 ) ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) (75)
+ τ ν C T , ν + 1 ( U , F ) (76) +\tau^\nu C_{T,\nu+1}(U,F) \tag*{(76)} + τ ν C T , ν + 1 ( U , F ) (76)
No factor τ − 1 h r \tau^{-1}h^r τ − 1 h r occurs because the time difference of the projection error is estimated as a difference quotient of a smooth function.
Dual Residual Bound. The gain of one half order is obtained by estimating the residual functional itself, rather than its strong norm. Orthogonality of the L 2 L^2 L 2 -projection allows the functional to be reorganised in terms of precisely the quantities controlled by the energy and dissipation. Define
N r , 2 ( U ) = ∑ j = 0 2 ∥ ∂ t j U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (77) N_{r,2}(U)=\sum_{j=0}^{2}\|\partial_t^jU\|_{L^2(0,T;H^r(\Omega))} \tag*{(77)} N r , 2 ( U ) = j = 0 ∑ 2 ∥ ∂ t j U ∥ L 2 ( 0 , T ; H r ( Ω )) (77)
+ ∑ j = 0 1 ∥ ∂ t j U ∥ L ∞ ( 0 , T ; H r ( Ω ) ) (78) +\sum_{j=0}^{1}\|\partial_t^jU\|_{L^\infty(0,T;H^r(\Omega))} \tag*{(78)} + j = 0 ∑ 1 ∥ ∂ t j U ∥ L ∞ ( 0 , T ; H r ( Ω )) (78)
Proposition 3.2 (Dual residual estimate). Assume (3.8)–(3.9), 0 < δ ≤ 1 0 < \delta\le1 0 < δ ≤ 1 , the approximation estimate (2.16) for U U U , ∂ t U \partial_t U ∂ t U , and ∂ t 2 U \partial_t^2 U ∂ t 2 U , and the differentiated consistency estimates (2.21)–(2.22). Then
∥ A τ η − ρ ∥ E ∗ , N ≲ ( ( 1 + T 1 / 2 ) δ 1 / 2 h r − 1 + δ − 1 / 2 h r ) N r , 2 ( U ) (79) \lVert A_\tau\eta-\rho\rVert_{\mathcal{E}^*,N} \lesssim\left((1+T^{1/2})\delta^{1/2}h^{r-1}+\delta^{-1/2}h^r\right)\mathcal{N}_{r,2}(U) \tag*{(79)} ∥ A τ η − ρ ∥ E ∗ , N ≲ ( ( 1 + T 1/2 ) δ 1/2 h r − 1 + δ − 1/2 h r ) N r , 2 ( U ) (79)
∥ A τ η − ρ ∥ E ∗ , N + ( 1 + T 1 / 2 ) τ ν C T , ν + 2 ♯ ( U , F ) (80) \phantom{\lVert A_\tau\eta-\rho\rVert_{\mathcal{E}^*,N}}\quad+(1+T^{1/2})\tau^\nu\mathcal{C}^{\sharp}_{T,\nu+2}(U,F) \tag*{(80)} ∥ A τ η − ρ ∥ E ∗ , N + ( 1 + T 1/2 ) τ ν C T , ν + 2 ♯ ( U , F ) (80)
The hidden constant depends on the fixed time stencil and on the coercivity constants, but not on h h h , τ \tau τ , or δ \delta δ .
Proof. Because the projection is time independent, the projection error and all its time differences and method averages are L 2 L^2 L 2 -orthogonal to V h V_h V h . Expanding P δ = M τ + δ A τ + δ 2 G D τ − P_\delta=M_\tau+\delta A_\tau+\delta^2GD_\tau^- P δ = M τ + δ A τ + δ 2 G D τ − therefore gives the exact identity
( A τ η n , P δ V n ) Ω = ( δ A τ η n − M τ η n , A τ V n ) Ω + δ 2 ( A τ η n , G D τ − V n ) Ω (81) (A_\tau\eta^n,P_\delta V^n)_\Omega=(\delta A_\tau\eta^n-M_\tau\eta^n,A_\tau V^n)_\Omega+\delta^2(A_\tau\eta^n,GD_\tau^-V^n)_\Omega \tag*{(81)} ( A τ η n , P δ V n ) Ω = ( δ A τ η n − M τ η n , A τ V n ) Ω + δ 2 ( A τ η n , G D τ − V n ) Ω (81)
The first term is controlled directly by the dissipation quantity δ 1 / 2 A τ V \delta^{1/2}A_\tau V δ 1/2 A τ V . In the last term, G D τ − V = D τ − G V GD_\tau^-V=D_\tau^-GV G D τ − V = D τ − G V , so discrete summation by parts moves the time difference onto A τ η A_\tau\eta A τ η , with both endpoint terms retained. The same decomposition applied to ( ρ , P δ V ) Ω (\rho,P_\delta V)_\Omega ( ρ , P δ V ) Ω moves the last time difference onto ρ \rho ρ . Coercivity controls δ 1 / 2 A τ V \delta^{1/2}A_\tau V δ 1/2 A τ V , while (2.29) controls V V V and δ G V \delta GV δ G V at every computed and start-up level. Approximation and differentiated consistency then give (3.32)–(3.33). The complete horizon-uniform calculation is given in Appendix A. ■
Theorem 3.3 (Abstract error estimate). Assume Theorem 3.1, the regularity hypotheses of Proposition 3.2, and projected start-up data U h j = Π h U ( t j ) U_h^j=\Pi_hU(t^j) U h j = Π h U ( t j ) for 0 ≤ j < n 0 0\le j<n_0 0 ≤ j < n 0 . Then
max n 0 ≤ n ≤ N ∥ T δ ( U h n − U ( t n ) ) ∥ Ω ≲ ( h r + δ h r − 1 ) ∥ U ∥ L ∞ ( 0 , T ; H r ( Ω ) ) (82) \max_{n_0\le n\le N}\lVert T_\delta(U_h^n-U(t^n))\rVert_\Omega\lesssim(h^r+\delta h^{r-1})\lVert U\rVert_{L^\infty(0,T;H^r(\Omega))} \tag*{(82)} n 0 ≤ n ≤ N max ∥ T δ ( U h n − U ( t n )) ∥ Ω ≲ ( h r + δ h r − 1 ) ∥ U ∥ L ∞ ( 0 , T ; H r ( Ω )) (82)
max n 0 ≤ n ≤ N ∥ T δ ( U h n − U ( t n ) ) ∥ Ω + ( ( 1 + T 1 / 2 ) δ 1 / 2 h r − 1 + δ − 1 / 2 h r ) N r , 2 ( U ) (83) \phantom{\max_{n_0\le n\le N}\lVert T_\delta(U_h^n-U(t^n))\rVert_\Omega}\quad+\left((1+T^{1/2})\delta^{1/2}h^{r-1}+\delta^{-1/2}h^r\right)\mathcal{N}_{r,2}(U) \tag*{(83)} n 0 ≤ n ≤ N m a x ∥ T δ ( U h n − U ( t n )) ∥ Ω + ( ( 1 + T 1/2 ) δ 1/2 h r − 1 + δ − 1/2 h r ) N r , 2 ( U ) (83)
max n 0 ≤ n ≤ N ∥ T δ ( U h n − U ( t n ) ) ∥ Ω + ( 1 + T 1 / 2 ) τ ν C T , ν + 2 ♯ ( U , F ) (84) \phantom{\max_{n_0\le n\le N}\lVert T_\delta(U_h^n-U(t^n))\rVert_\Omega}\quad+(1+T^{1/2})\tau^\nu\mathcal{C}^{\sharp}_{T,\nu+2}(U,F) \tag*{(84)} n 0 ≤ n ≤ N m a x ∥ T δ ( U h n − U ( t n )) ∥ Ω + ( 1 + T 1/2 ) τ ν C T , ν + 2 ♯ ( U , F ) (84)
max n 0 ≤ n ≤ N ∥ T δ ( U h n − U ( t n ) ) ∥ Ω + τ 1 / 2 ( h r + τ h r − 1 ) ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (85) \phantom{\max_{n_0\le n\le N}\lVert T_\delta(U_h^n-U(t^n))\rVert_\Omega}\quad+\tau^{1/2}(h^r+\tau h^{r-1})\lVert\partial_tU\rVert_{L^2(0,T;H^r(\Omega))} \tag*{(85)} n 0 ≤ n ≤ N m a x ∥ T δ ( U h n − U ( t n )) ∥ Ω + τ 1/2 ( h r + τ h r − 1 ) ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) (85)
max n 0 ≤ n ≤ N ∥ T δ ( U h n − U ( t n ) ) ∥ Ω + τ 1 / 2 h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω ) ) (86) \phantom{\max_{n_0\le n\le N}\lVert T_\delta(U_h^n-U(t^n))\rVert_\Omega}\quad+\tau^{1/2}h^{r-1}\lVert U\rVert_{L^2(0,T;H^r(\Omega))} \tag*{(86)} n 0 ≤ n ≤ N m a x ∥ T δ ( U h n − U ( t n )) ∥ Ω + τ 1/2 h r − 1 ∥ U ∥ L 2 ( 0 , T ; H r ( Ω )) (86)
max n 0 ≤ n ≤ N ∥ T δ ( U h n − U ( t n ) ) ∥ Ω + τ ν + 1 / 2 C T , ν + 2 ♯ ( U , F ) (87) \phantom{\max_{n_0\le n\le N}\lVert T_\delta(U_h^n-U(t^n))\rVert_\Omega}\quad+\tau^{\nu+1/2}\mathcal{C}^{\sharp}_{T,\nu+2}(U,F) \tag*{(87)} n 0 ≤ n ≤ N m a x ∥ T δ ( U h n − U ( t n )) ∥ Ω + τ ν + 1/2 C T , ν + 2 ♯ ( U , F ) (87)
The same right-hand side controls the accumulated material-residual dissipation in the form
( τ δ ∑ n ∥ A τ φ n ∥ Ω 2 ) 1 / 2 . \left(\tau\delta\sum_n\lVert A_\tau\varphi^n\rVert_\Omega^2\right)^{1/2}. ( τ δ n ∑ ∥ A τ φ n ∥ Ω 2 ) 1/2 .
Proof. Apply Theorem 3.1 to (3.17). The start-up term vanishes, and (3.19), (3.28)–(3.29), and Proposition 3.2 bound the strong and dual residual contributions. Finally,
U h n − U ( t n ) = φ n − η n (88) U_h^n-U(t^n)=\varphi^n-\eta^n \tag*{(88)} U h n − U ( t n ) = φ n − η n (88)
and (2.16) bounds the graph norm of η n \eta^n η n . More precisely, the time-uniform projection contribution is
max 0 ≤ n ≤ N ∥ T δ η n ∥ Ω ≲ ( h r + δ h r − 1 ) ∥ U ∥ L ∞ ( 0 , T ; H r ( Ω ) ) (89) \max_{0\le n\le N}\lVert T_\delta\eta^n\rVert_\Omega\lesssim(h^r+\delta h^{r-1})\lVert U\rVert_{L^\infty(0,T;H^r(\Omega))} \tag*{(89)} 0 ≤ n ≤ N max ∥ T δ η n ∥ Ω ≲ ( h r + δ h r − 1 ) ∥ U ∥ L ∞ ( 0 , T ; H r ( Ω )) (89)
The terms in (3.38)–(3.40) come from the explicit τ 1 / 2 ∥ R ∥ ℓ τ 2 \tau^{1/2}\lVert R\rVert_{\ell^2_\tau} τ 1/2 ∥ R ∥ ℓ τ 2 term in (3.13). In particular, (3.38) contains the additional term τ 3 / 2 h r − 1 ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) \tau^{3/2}h^{r-1}\lVert\partial_tU\rVert_{L^2(0,T;H^r(\Omega))} τ 3/2 h r − 1 ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) ; no parameter relation is used to discard it. ■
Parameter Balance
We now specialise the estimate to the parameter regimes used later. For the standard stabilised choice δ ∼ h \delta\sim h δ ∼ h , the spatial part of (82) –(87) is of order h r − 1 / 2 h^{r-1/2} h r − 1/2 when τ ≲ h \tau\lesssim h τ ≲ h . The projection term is then of the higher order h r h^r h r . Thus, with r = k + 1 r = k + 1 r = k + 1 , smooth degree-k k k approximation gives order h k + 1 / 2 h^{k+1/2} h k + 1/2 under the stated regularity assumptions.
The AM3–AM5 assumptions in Section 5 and the AB3–AB4 assumption in Section 6 already include τ ≲ h \tau\lesssim h τ ≲ h . For those methods, τ h r − 1 ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) \tau h^{r-1}\lVert\partial_t U\rVert_{L^2(0,T;H^r(\Omega))} τ h r − 1 ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) is therefore controlled by h r ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω ) ) h^r\lVert\partial_t U\rVert_{L^2(0,T;H^r(\Omega))} h r ∥ ∂ t U ∥ L 2 ( 0 , T ; H r ( Ω )) , so the previously stated asymptotic orders are unchanged. No condition τ ≲ h \tau\lesssim h τ ≲ h is added to the abstract stability theorem or to θ \theta θ -method stability. The θ \theta θ -method error estimate retains the explicit τ 3 / 2 h r − 1 \tau^{3/2}h^{r-1} τ 3/2 h r − 1 -term unless a separately stated parameter regime justifies its absorption. In particular, under δ ∼ h \delta\sim h δ ∼ h and τ ≲ h \tau\lesssim h τ ≲ h , the h r − 1 / 2 h^{r-1/2} h r − 1/2 -conclusion is unchanged.
For the normal-equation choice δ = b 0 τ \delta= b_0\tau δ = b 0 τ , the two dual contributions are τ 1 / 2 h r − 1 \tau^{1/2}h^{r-1} τ 1/2 h r − 1 and τ − 1 / 2 h r \tau^{-1/2}h^r τ − 1/2 h r . Formally balancing the latter with the time error gives
τ = h r / ( ν + 1 / 2 ) (90) \tau= h^{r/(\nu+1/2)} \tag*{(90)} τ = h r / ( ν + 1/2 ) (90)
This is an accuracy balance; stability must still be checked for the chosen time-stepping method. The Adams–Moulton theorem below includes the normal-equation choice under its method-dependent CFL conditions, while Proposition 3.2 applies in both parameter regimes.
An A-Stable Scheme: The Theta Method
The θ \theta θ -family gives the simplest realisation of the abstract framework: its energy balance follows from a one-step identity and requires no CFL condition. It includes Crank–Nicolson at θ = 1 / 2 \theta= 1/2 θ = 1/2 and backward Euler at θ = 1 \theta= 1 θ = 1 . Throughout this section, let 1 / 2 ≤ θ ≤ 1 1/2 \le\theta\le1 1/2 ≤ θ ≤ 1 and set
D τ − U n = τ − 1 ( U n − U n − 1 ) (91) D_\tau^- U^n = \tau^{-1}(U^n - U^{n-1}) \tag*{(91)} D τ − U n = τ − 1 ( U n − U n − 1 ) (91)
M θ U n = θ U n + ( 1 − θ ) U n − 1 (92) \mathcal{M}_\theta U^n = \theta U^n + (1-\theta)U^{n-1} \tag*{(92)} M θ U n = θ U n + ( 1 − θ ) U n − 1 (92)
δ = θ τ (93) \delta= \theta\tau \tag*{(93)} δ = θ τ (93)
Here the current method coefficient is b 0 = θ b_0 = \theta b 0 = θ , so the last choice is the normal-equation value b 0 τ b_0\tau b 0 τ . The corresponding acoustic-wave realisation is given in Section 2.4.
Energy and Dissipation. The energy is one half of the squared graph norm associated with the normal equation. Define
e τ n ( U ) = 1 2 ∥ T δ U n ∥ Ω 2 = 1 2 ∥ U n ∥ Ω 2 + 1 2 δ 2 ∥ G U n ∥ Ω 2 (94) e_\tau^n(U) = \frac{1}{2}\lVert T_\delta U^n\rVert_\Omega^2 = \frac{1}{2}\lVert U^n\rVert_\Omega^2 + \frac{1}{2}\delta^2\lVert G U^n\rVert_\Omega^2 \tag*{(94)} e τ n ( U ) = 2 1 ∥ T δ U n ∥ Ω 2 = 2 1 ∥ U n ∥ Ω 2 + 2 1 δ 2 ∥ G U n ∥ Ω 2 (94)
and
d τ n ( U ) = δ ∥ A τ U n ∥ Ω 2 + 2 θ − 1 2 τ ∥ D τ − U n ∥ Ω 2 (95) d_\tau^n(U) = \delta\lVert A_\tau U^n\rVert_\Omega^2 + \frac{2\theta-1}{2}\tau\lVert D_\tau^- U^n\rVert_\Omega^2 \tag*{(95)} d τ n ( U ) = δ ∥ A τ U n ∥ Ω 2 + 2 2 θ − 1 τ ∥ D τ − U n ∥ Ω 2 (95)
+ 2 θ − 1 2 τ δ 2 ∥ G D τ − U n ∥ Ω 2 (96) + \frac{2\theta-1}{2}\tau\delta^2\lVert G D_\tau^- U^n\rVert_\Omega^2 \tag*{(96)} + 2 2 θ − 1 τ δ 2 ∥ G D τ − U n ∥ Ω 2 (96)
Both quantities are nonnegative for 1 / 2 ≤ θ ≤ 1 1/2 \le\theta\le1 1/2 ≤ θ ≤ 1 . Consequently, (55) –(56) hold with the common constant c E = 1 / 2 c_E = 1/2 c E = 1/2 . At the Crank–Nicolson endpoint θ = 1 / 2 \theta= 1/2 θ = 1/2 , the two terms proportional to 2 θ − 1 2\theta- 1 2 θ − 1 vanish, while the material-residual term δ ∥ A τ U n ∥ Ω 2 \delta\lVert A_\tau U^n\rVert_\Omega^2 δ ∥ A τ U n ∥ Ω 2 remains.
Increment Identities. The elementary Hilbert-space identity for the θ \theta θ -average gives
τ ( D τ − U n , M θ U n ) Ω = 1 2 ( ∥ U n ∥ Ω 2 − ∥ U n − 1 ∥ Ω 2 ) (97) \tau(D_{\tau}^{-}U^{n},\mathcal{M}_{\theta}U^{n})_{\Omega}=\frac{1}{2}\left(\lVert U^{n}\rVert_{\Omega}^{2}-\lVert U^{n-1}\rVert_{\Omega}^{2}\right) \tag*{(97)} τ ( D τ − U n , M θ U n ) Ω = 2 1 ( ∥ U n ∥ Ω 2 − ∥ U n − 1 ∥ Ω 2 ) (97)
+ 2 θ − 1 2 τ 2 ∥ D τ − U n ∥ Ω 2 (98) +\frac{2\theta-1}{2}\tau^{2}\lVert D_{\tau}^{-}U^{n}\rVert_{\Omega}^{2} \tag*{(98)} + 2 2 θ − 1 τ 2 ∥ D τ − U n ∥ Ω 2 (98)
Applying the same identity to G U n GU^{n} G U n yields
τ δ 2 ( G M θ U n , G D τ − U n ) Ω = 1 2 δ 2 ( ∥ G U n ∥ Ω 2 − ∥ G U n − 1 ∥ Ω 2 ) (99) \tau\delta^{2}(G\mathcal{M}_{\theta}U^{n},GD_{\tau}^{-}U^{n})_{\Omega}=\frac{1}{2}\delta^{2}\left(\lVert GU^{n}\rVert_{\Omega}^{2}-\lVert GU^{n-1}\rVert_{\Omega}^{2}\right) \tag*{(99)} τ δ 2 ( G M θ U n , G D τ − U n ) Ω = 2 1 δ 2 ( ∥ G U n ∥ Ω 2 − ∥ G U n − 1 ∥ Ω 2 ) (99)
+ 2 θ − 1 2 τ 2 δ 2 ∥ G D τ − U n ∥ Ω 2 (100) +\frac{2\theta-1}{2}\tau^{2}\delta^{2}\lVert GD_{\tau}^{-}U^{n}\rVert_{\Omega}^{2} \tag*{(100)} + 2 2 θ − 1 τ 2 δ 2 ∥ G D τ − U n ∥ Ω 2 (100)
The two increment identities convert the first and third terms in the fundamental algebraic identity into telescoping energy differences and nonnegative dissipation. In particular, the factor τ δ 2 \tau\delta^{2} τ δ 2 in (99) matches (50) .
Proposition 4.1 (Energy–dissipation identity for the Theta method). The solution of (36) with (91) –(93) satisfies
e τ n ( U ) − e τ n − 1 ( U ) + τ d τ n ( U ) = ( τ M θ F n , P δ U n ) Ω (101) e_{\tau}^{n}(U)-e_{\tau}^{n-1}(U)+\tau d_{\tau}^{n}(U)=(\tau\mathcal{M}_{\theta}F^{n},P_{\delta}U^{n})_{\Omega} \tag*{(101)} e τ n ( U ) − e τ n − 1 ( U ) + τ d τ n ( U ) = ( τ M θ F n , P δ U n ) Ω (101)
Proof. Use S δ U n S_{\delta}U^{n} S δ U n as the test function in (36) . Equation (49) –(50) gives
( τ A τ U n , P δ U n ) Ω = τ ( D τ − U n , M θ U n ) Ω + τ δ ∥ A τ U n ∥ Ω 2 (102) (\tau A_{\tau}U^{n},P_{\delta}U^{n})_{\Omega}=\tau(D_{\tau}^{-}U^{n},\mathcal{M}_{\theta}U^{n})_{\Omega}+\tau\delta\lVert A_{\tau}U^{n}\rVert_{\Omega}^{2} \tag*{(102)} ( τ A τ U n , P δ U n ) Ω = τ ( D τ − U n , M θ U n ) Ω + τ δ ∥ A τ U n ∥ Ω 2 (102)
+ τ δ 2 ( G M θ U n , G D τ − U n ) Ω (103) +\tau\delta^{2}(G\mathcal{M}_{\theta}U^{n},GD_{\tau}^{-}U^{n})_{\Omega} \tag*{(103)} + τ δ 2 ( G M θ U n , G D τ − U n ) Ω (103)
Substitution of (97) –(100) and the definitions (94) –(96) proves (101) . ■
Corollary 4.2 (Theta-method stability). For 1 / 2 ≤ θ ≤ 1 1/2\leq\theta\leq1 1/2 ≤ θ ≤ 1 , the energy hypothesis (55) –(59) holds with the forcing-defect term omitted. Consequently,
∥ U h ∥ E , N 2 ≲ ∥ T δ U 0 ∥ Ω 2 + ∥ M θ F ∥ ∗ , N 2 (104) \lVert U_{h}\rVert_{\mathcal{E},N}^{2}\lesssim\lVert T_{\delta}U_{0}\rVert_{\Omega}^{2}+\lVert\mathcal{M}_{\theta}F\rVert_{\ast,N}^{2} \tag*{(104)} ∥ U h ∥ E , N 2 ≲ ∥ T δ U 0 ∥ Ω 2 + ∥ M θ F ∥ ∗ , N 2 (104)
Proof. Sum (101) from n = 1 n=1 n = 1 to an arbitrary horizon m m m . This is the abstract energy hypothesis with n 0 = 1 n_{0}=1 n 0 = 1 , R n = M θ F n R^{n}=\mathcal{M}_{\theta}F^{n} R n = M θ F n , and no strong forcing defect. Apply Theorem 3.1 with e τ n e_{\tau}^{n} e τ n and d τ n d_{\tau}^{n} d τ n defined in (94) –(96) . ■
Consistency and Error Estimate
To apply the abstract error theorem, it remains to determine the temporal order and verify the differentiated consistency bound. For the sampled exact solution, the differential equation gives
ρ n = D τ − U ( t n ) − M θ ∂ t U ( t n ) (105) \rho^{n}=D_{\tau}^{-}U(t^{n})-\mathcal{M}_{\theta}\partial_{t}U(t^{n}) \tag*{(105)} ρ n = D τ − U ( t n ) − M θ ∂ t U ( t n ) (105)
Taylor’s formula with an integral remainder yields
∥ ρ n ∥ Ω ≲ ∣ 2 θ − 1 ∣ τ 1 / 2 ∥ ∂ t 2 U ∥ L 2 ( t n − 1 , t n ; H ) (106) \lVert\rho^{n}\rVert_{\Omega}\lesssim\lvert2\theta-1\rvert\tau^{1/2}\lVert\partial_{t}^{2}U\rVert_{L^{2}(t^{n-1},t^{n};\mathcal{H})} \tag*{(106)} ∥ ρ n ∥ Ω ≲ ∣ 2 θ − 1 ∣ τ 1/2 ∥ ∂ t 2 U ∥ L 2 ( t n − 1 , t n ; H ) (106)
+ τ 3 / 2 ∥ ∂ t 3 U ∥ L 2 ( t n − 1 , t n ; H ) (107) +\tau^{3/2}\lVert\partial_{t}^{3}U\rVert_{L^{2}(t^{n-1},t^{n};\mathcal{H})} \tag*{(107)} + τ 3/2 ∥ ∂ t 3 U ∥ L 2 ( t n − 1 , t n ; H ) (107)
Thus the method has order ν = 1 \nu=1 ν = 1 for fixed θ > 1 / 2 \theta>1/2 θ > 1/2 , and order ν = 2 \nu=2 ν = 2 for θ = 1 / 2 \theta=1/2 θ = 1/2 or, more generally, θ = 1 / 2 + O ( τ ) \theta=1/2+O(\tau) θ = 1/2 + O ( τ ) . The dual-residual argument also requires a uniform bound for ρ \rho ρ and an ℓ τ 2 \ell_{\tau}^{2} ℓ τ 2 -bound for its forward difference. Applying the same integral-remainder formula to these quantities gives
max 1 ≤ n ≤ N ∥ ρ n ∥ Ω + ∥ D τ + ρ ∥ ℓ τ 2 ( 1 , N − 1 ; H ) ≲ ∣ 2 θ − 1 ∣ τ ( ∥ ∂ t 2 U ∥ L ∞ ( 0 , T ; H ) + ∥ ∂ t 3 U ∥ L 2 ( 0 , T ; H ) ) (108) \max_{1\leq n\leq N}\lVert\rho^{n}\rVert_{\Omega}+\lVert D_{\tau}^{+}\rho\rVert_{\ell_{\tau}^{2}(1,N-1;\mathcal{H})}\lesssim\lvert2\theta-1\rvert\tau\left(\lVert\partial_{t}^{2}U\rVert_{L^{\infty}(0,T;\mathcal{H})}+\lVert\partial_{t}^{3}U\rVert_{L^{2}(0,T;\mathcal{H})}\right) \tag*{(108)} 1 ≤ n ≤ N max ∥ ρ n ∥ Ω + ∥ D τ + ρ ∥ ℓ τ 2 ( 1 , N − 1 ; H ) ≲ ∣ 2 θ − 1 ∣ τ ( ∥ ∂ t 2 U ∥ L ∞ ( 0 , T ; H ) + ∥ ∂ t 3 U ∥ L 2 ( 0 , T ; H ) ) (108)
+ τ 2 ( ∥ ∂ t 3 U ∥ L ∞ ( 0 , T ; H ) + ∥ ∂ t 4 U ∥ L 2 ( 0 , T ; H ) ) (109) +\tau^{2}\left(\lVert\partial_{t}^{3}U\rVert_{L^{\infty}(0,T;\mathcal{H})}+\lVert\partial_{t}^{4}U\rVert_{L^{2}(0,T;\mathcal{H})}\right) \tag*{(109)} + τ 2 ( ∥ ∂ t 3 U ∥ L ∞ ( 0 , T ; H ) + ∥ ∂ t 4 U ∥ L 2 ( 0 , T ; H ) ) (109)
Hence (25) –(26) hold with ν = 1 \nu=1 ν = 1 for fixed θ > 1 / 2 \theta>1/2 θ > 1/2 , and with ν = 2 \nu=2 ν = 2 for θ = 1 / 2 \theta=1/2 θ = 1/2 or θ = 1 / 2 + O ( τ ) \theta=1/2+O(\tau) θ = 1/2 + O ( τ ) . The displayed norms are included in C T , ν + 2 ♯ ( U , F ) C_{T,\nu+2}^{\sharp}(U,F) C T , ν + 2 ♯ ( U , F ) .
Corollary 4.3 (Theta-method error estimate). *Assume the regularity hypotheses of Proposition 3.2 and projected initial data. Then (3.35)–(3.40) holds with ν = 1 \nu= 1 ν = 1 for fixed θ > 1 / 2 \theta> 1/2 θ > 1/2 , and with ν = 2 \nu= 2 ν = 2 for θ = 1 / 2 \theta= 1/2 θ = 1/2 or θ = 1 / 2 + O ( τ ) \theta= 1/2 + O(\tau) θ = 1/2 + O ( τ ) .
Proof. Combine Corollary 4.2, (4.16)–(4.17), (4.18)–(4.19), and Theorem 3.3. ■
Under the balanced scaling τ ∼ h \tau\sim h τ ∼ h , the normal-equation choice satisfies δ ∼ h \delta\sim h δ ∼ h . With r = k + 1 r = k+1 r = k + 1 , the parameter balance in Section 3.3 therefore yields the graph-norm order O ( h k + 1 / 2 + τ ν ) O(h^{k+1/2}+\tau^\nu) O ( h k + 1/2 + τ ν ) , with ν \nu ν as stated in Corollary 4.3.
Adams–Moulton Methods of Orders Three to Five
This section treats the implicit Adams–Moulton methods AM3–AM5. We first establish consistency and the increment bounds needed to control the multistep history. Exact energy decompositions then reveal the different stability mechanisms: AM3 and AM4 require strengthened sufficient CFL conditions for general polynomial degree, whereas AM5 is stable under the standard hyperbolic condition. For continuous piecewise affine finite elements, an elementwise cancellation recovers the standard hyperbolic regime also for AM3 and AM4. The Adams–Bashforth methods are defined and analysed independently in Section 6.
Method Coefficients. For all three methods, D τ U n = τ − 1 ( U n − U n − 1 ) D_\tau U^n = \tau^{-1}(U^n-U^{n-1}) D τ U n = τ − 1 ( U n − U n − 1 ) , while the coefficients in (2.11) are
AM3 : ( b 0 , b 1 , b 2 ) = ( 5 12 , 2 3 , − 1 12 ) (110) \operatorname{AM3}: \quad(b_0,b_1,b_2)=\left(\frac{5}{12},\frac{2}{3},-\frac{1}{12}\right) \tag*{(110)} AM3 : ( b 0 , b 1 , b 2 ) = ( 12 5 , 3 2 , − 12 1 ) (110)
AM4 : ( b 0 , b 1 , b 2 , b 3 ) = ( 3 8 , 19 24 , − 5 24 , 1 24 ) (111) \operatorname{AM4}: \quad(b_0,b_1,b_2,b_3)=\left(\frac{3}{8},\frac{19}{24},-\frac{5}{24},\frac{1}{24}\right) \tag*{(111)} AM4 : ( b 0 , b 1 , b 2 , b 3 ) = ( 8 3 , 24 19 , − 24 5 , 24 1 ) (111)
AM5 : ( b 0 , b 1 , b 2 , b 3 , b 4 ) = ( 251 720 , 646 720 , − 264 720 , 106 720 , − 19 720 ) (112) \operatorname{AM5}: \quad(b_0,b_1,b_2,b_3,b_4)=\left(\frac{251}{720},\frac{646}{720},-\frac{264}{720},\frac{106}{720},-\frac{19}{720}\right) \tag*{(112)} AM5 : ( b 0 , b 1 , b 2 , b 3 , b 4 ) = ( 720 251 , 720 646 , − 720 264 , 720 106 , − 720 19 ) (112)
Set n 0 ( 3 ) = 2 n_0(3)=2 n 0 ( 3 ) = 2 , n 0 ( 4 ) = 3 n_0(4)=3 n 0 ( 4 ) = 3 , and n 0 ( 5 ) = 4 n_0(5)=4 n 0 ( 5 ) = 4 . These are the first computed indices for AM3, AM4, and AM5, respectively. Define the backward increments recursively by
d 1 U n = U n − U n − 1 (113) d^1U^n=U^n-U^{n-1} \tag*{(113)} d 1 U n = U n − U n − 1 (113)
d j U n = d j − 1 U n − d j − 1 U n − 1 2 ≤ j ≤ 4 (114) d^jU^n=d^{j-1}U^n-d^{j-1}U^{n-1} \qquad2\leq j\leq4 \tag*{(114)} d j U n = d j − 1 U n − d j − 1 U n − 1 2 ≤ j ≤ 4 (114)
With U ~ n = ( U n + U n − 1 ) / 2 \widetilde{U}^n=(U^n+U^{n-1})/2 U n = ( U n + U n − 1 ) /2 , exact expansion of the coefficients in (5.1)–(5.3) gives
M AM3 U n = U ~ n − 1 12 d 2 U n (115) \mathcal{M}_{\operatorname{AM3}}U^n=\widetilde{U}^n-\frac{1}{12}d^2U^n \tag*{(115)} M AM3 U n = U n − 12 1 d 2 U n (115)
M AM4 U n = U ~ n − 1 12 d 2 U n − 1 24 d 3 U n (116) \mathcal{M}_{\operatorname{AM4}}U^n=\widetilde{U}^n-\frac{1}{12}d^2U^n-\frac{1}{24}d^3U^n \tag*{(116)} M AM4 U n = U n − 12 1 d 2 U n − 24 1 d 3 U n (116)
M AM5 U n = U ~ n − 1 12 d 2 U n − 1 24 d 3 U n − 19 720 d 4 U n (117) \mathcal{M}_{\operatorname{AM5}}U^n=\widetilde{U}^n-\frac{1}{12}d^2U^n-\frac{1}{24}d^3U^n-\frac{19}{720}d^4U^n \tag*{(117)} M AM5 U n = U n − 12 1 d 2 U n − 24 1 d 3 U n − 720 19 d 4 U n (117)
These centred forms isolate the higher backward increments that generate the bulk terms in the energy decompositions below.
Consistency of the Adams–Moulton Methods
Set σ ( t ) = F ( t ) − G U ( t ) = ∂ t U ( t ) \sigma(t)=F(t)-GU(t)=\partial_tU(t) σ ( t ) = F ( t ) − G U ( t ) = ∂ t U ( t ) . The Adams–Moulton formula is the interpolatory quadrature rule obtained by approximating this exact time derivative over one time step. Indeed, the exact equation implies
U ( t n ) − U ( t n − 1 ) = ∫ t n − 1 t n σ ( t ) d t (118) U(t^n)-U(t^{n-1})=\int_{t_{n-1}}^{t^n}\sigma(t)\,\mathrm{d}t \tag*{(118)} U ( t n ) − U ( t n − 1 ) = ∫ t n − 1 t n σ ( t ) d t (118)
For AMν \nu ν , let σ ν − 1 \sigma_{\nu-1} σ ν − 1 be the polynomial of degree at most ν − 1 \nu-1 ν − 1 interpolating σ \sigma σ at the nodes
t n − j 0 ≤ j ≤ ν − 1 (119) t^{n-j} \qquad0 \le j \le\nu-1 \tag*{(119)} t n − j 0 ≤ j ≤ ν − 1 (119)
Integrating the interpolant over the current step gives, by construction,
∫ t n − 1 t n σ ν − 1 ( t ) d t = τ M A M ν σ ( t n ) (120) \int_{t^{n-1}}^{t^n} \sigma_{\nu-1}(t)\,\mathrm{d}t = \tau\mathcal{M}_{\mathrm{AM}\nu}\sigma(t^n) \tag*{(120)} ∫ t n − 1 t n σ ν − 1 ( t ) d t = τ M AM ν σ ( t n ) (120)
Subtracting (5.11) from (5.9) gives
τ ρ n = ∫ t n − 1 t n ( σ ( t ) − σ ν − 1 ( t ) ) d t (121) \tau\rho^n = \int_{t^{n-1}}^{t^n} \left(\sigma(t)-\sigma_{\nu-1}(t)\right)\,\mathrm{d}t \tag*{(121)} τ ρ n = ∫ t n − 1 t n ( σ ( t ) − σ ν − 1 ( t ) ) d t (121)
ρ n = D τ − U ( t n ) + G M A M ν U ( t n ) − M A M ν F ( t n ) (122) \rho^n = D_\tau^- U(t^n) + G\mathcal{M}_{\mathrm{AM}\nu}U(t^n) - \mathcal{M}_{\mathrm{AM}\nu}F(t^n) \tag*{(122)} ρ n = D τ − U ( t n ) + G M AM ν U ( t n ) − M AM ν F ( t n ) (122)
The interpolation remainder and the Cauchy–Schwarz inequality yield
∥ ρ n ∥ Ω ≤ C ν τ ν − 1 / 2 ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν + 1 , t n ; H ) (123) \lVert\rho^n\rVert_{\Omega} \le C_\nu\tau^{\nu-1/2}\lVert\partial_t^\nu(F-GU)\rVert_{L^2(t^{n-\nu+1},t^n;H)} \tag*{(123)} ∥ ρ n ∥ Ω ≤ C ν τ ν − 1/2 ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν + 1 , t n ; H ) (123)
Thus (2.18) holds with the claimed order ν ∈ { 3 , 4 , 5 } \nu\in\{3,4,5\} ν ∈ { 3 , 4 , 5 } . More precisely, in (2.19) one may take
C n , ν + 1 ( U , F ) = C ν ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν + 1 , t n ; H ) (124) C_{n,\nu+1}(U,F)=C_\nu\lVert\partial_t^\nu(F-GU)\rVert_{L^2(t^{n-\nu+1},t^n;H)} \tag*{(124)} C n , ν + 1 ( U , F ) = C ν ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν + 1 , t n ; H ) (124)
These intervals have overlap at most ν \nu ν , so (2.20) follows with a constant depending only on the fixed order. The dual-residual estimate also requires uniform control of the residual and square-summable control of its first forward difference. Writing the interpolation remainder in integral form gives both bounds:
max n 0 ( ν ) ≤ n ≤ N ∥ ρ n ∥ Ω + ∥ D τ + ρ ∥ ℓ 2 ( n 0 ( ν ) , N − 1 ; H ) ≤ C ν τ ν ( ∥ ∂ t ν ( F − G U ) ∥ L ∞ ( 0 , T ; H ) + ∥ ∂ t ν + 1 ( F − G U ) ∥ L 2 ( 0 , T ; H ) ) (125) \max_{n_0(\nu)\le n\le N}\lVert\rho^n\rVert_{\Omega}+\lVert D_\tau^+\rho\rVert_{\ell^2(n_0(\nu),N-1;H)}
\le C_\nu\tau^\nu\left(\lVert\partial_t^\nu(F-GU)\rVert_{L^\infty(0,T;H)}+\lVert\partial_t^{\nu+1}(F-GU)\rVert_{L^2(0,T;H)}\right) \tag*{(125)} n 0 ( ν ) ≤ n ≤ N max ∥ ρ n ∥ Ω + ∥ D τ + ρ ∥ ℓ 2 ( n 0 ( ν ) , N − 1 ; H ) ≤ C ν τ ν ( ∥ ∂ t ν ( F − G U ) ∥ L ∞ ( 0 , T ; H ) + ∥ ∂ t ν + 1 ( F − G U ) ∥ L 2 ( 0 , T ; H ) ) (125)
Thus (2.21)–(2.22) hold for AM3–AM5 with these two norms included in C T , ν + 2 ( U , F ) C_{T,\nu+2}(U,F) C T , ν + 2 ( U , F ) .
Stability Assumptions and Increment Bounds
Assume the inverse inequality
∥ G V h ∥ Ω ≤ C G h − 1 ∥ V h ∥ Ω ∀ V h ∈ V h (126) \lVert GV_h\rVert_{\Omega}\le C_Gh^{-1}\lVert V_h\rVert_{\Omega}\qquad\forall V_h\in V_h \tag*{(126)} ∥ G V h ∥ Ω ≤ C G h − 1 ∥ V h ∥ Ω ∀ V h ∈ V h (126)
and the parameter relations
0 < δ ≤ C δ h , τ ≤ c 0 C G − 1 h (127) 0<\delta\le C_\delta h,\qquad\tau\le c_0C_G^{-1}h \tag*{(127)} 0 < δ ≤ C δ h , τ ≤ c 0 C G − 1 h (127)
where c 0 > 0 c_0>0 c 0 > 0 will be chosen sufficiently small. This is the standard hyperbolic CFL restriction used for all three methods. For the general-degree AM3 and AM4 estimates, the present proof also uses the respective additional conditions
τ ≤ C 4 / 3 h 4 / 3 for AM3 (128) \tau\le C_{4/3}h^{4/3}\qquad\text{for AM3} \tag*{(128)} τ ≤ C 4/3 h 4/3 for AM3 (128)
τ ≤ C 6 / 5 h 6 / 5 for AM4 (129) \tau\le C_{6/5}h^{6/5}\qquad\text{for AM4} \tag*{(129)} τ ≤ C 6/5 h 6/5 for AM4 (129)
No smallness is imposed on C 4 / 3 C_{4/3} C 4/3 or C 6 / 5 C_{6/5} C 6/5 ; these constants enter the final Gronwall constant. AM5 requires no strengthened condition.
Relation to the Normal Equation. Only the upper bound on δ \delta δ is used below. Together with the inverse inequality, it makes the graph norm uniformly equivalent to the L 2 L^2 L 2 -norm on V h V_h V h ; no lower bound proportional to h h h is needed. The stability proof therefore permits δ / h → 0 \delta/h \to0 δ / h → 0 . In particular, the normal-equation value δ = b 0 τ \delta=b_0\tau δ = b 0 τ satisfies (5.19) under the relevant time-step condition and is covered for AM3–AM5. Choices δ ≠ b 0 τ \delta\ne b_0\tau δ = b 0 τ remain generally nonsymmetric SUPG schemes.
The energy decompositions contain negative bulk terms involving the highest backward increments for AM3 and AM4, and a negative fourth-increment term for AM5. The following lemma relates these increments to earlier solution values and to the forcing.
Lemma 5.1 (Increment recurrence). Let ν ∈ { 3 , 4 , 5 } \nu\in\{3,4,5\} ν ∈ { 3 , 4 , 5 } , and let V h V_h V h solve the generally forced AMν \nu ν equation
( A τ V h n , T δ W ) Ω = ( R n , T δ W ) Ω ∀ W ∈ V h , n ≥ n 0 ( ν ) (130) (A_\tau V_h^n,T_\delta W)_\Omega=(R^n,T_\delta W)_\Omega\qquad\forall W \in V_h,\qquad n \ge n_0(\nu) \tag*{(130)} ( A τ V h n , T δ W ) Ω = ( R n , T δ W ) Ω ∀ W ∈ V h , n ≥ n 0 ( ν ) (130)
Define the remaining-stencil coefficients by
γ 0 = b 0 + b 1 (131) \gamma_0=b_0+b_1 \tag*{(131)} γ 0 = b 0 + b 1 (131)
γ q = b q + 1 1 ≤ q ≤ ν − 2 (132) \gamma_q=b_{q+1}\qquad1 \le q \le\nu-2 \tag*{(132)} γ q = b q + 1 1 ≤ q ≤ ν − 2 (132)
For every admissible increment order j ≥ 1 j \ge1 j ≥ 1 and every n ≥ n 0 ( ν ) + j − 1 n \ge n_0(\nu)+j-1 n ≥ n 0 ( ν ) + j − 1 ,
∥ d j V h n ∥ δ ≲ τ h ∑ q = 0 ν − 2 ∥ d j − 1 V h n − 1 − q ∥ δ (133) \lVert d^jV_h^n\rVert_\delta\lesssim\frac{\tau}{h}\sum_{q=0}^{\nu-2}\lVert d^{j-1}V_h^{n-1-q}\rVert_\delta \tag*{(133)} ∥ d j V h n ∥ δ ≲ h τ q = 0 ∑ ν − 2 ∥ d j − 1 V h n − 1 − q ∥ δ (133)
+ τ ∑ q = 0 j − 1 ∥ R n − q ∥ Ω (134) \qquad+\tau\sum_{q=0}^{j-1}\lVert R^{n-q}\rVert_\Omega \tag*{(134)} + τ q = 0 ∑ j − 1 ∥ R n − q ∥ Ω (134)
where d 0 V h n = V h n d^0V_h^n=V_h^n d 0 V h n = V h n . In particular,
∥ d 1 V h n ∥ δ ≲ τ h ∑ q = 0 ν − 2 ∥ V h n − 1 − q ∥ δ + τ ∥ R n ∥ Ω (135) \lVert d^1V_h^n\rVert_\delta\lesssim\frac{\tau}{h}\sum_{q=0}^{\nu-2}\lVert V_h^{n-1-q}\rVert_\delta+\tau\lVert R^n\rVert_\Omega \tag*{(135)} ∥ d 1 V h n ∥ δ ≲ h τ q = 0 ∑ ν − 2 ∥ V h n − 1 − q ∥ δ + τ ∥ R n ∥ Ω (135)
The local recurrence above is the basic estimate. Its iteration gives the bulk and endpoint bounds stated next. Early increments for which the formal difference would precede the first computed equation are estimated directly from the computed updates. Only their solution terms are absorbed into I τ ( V h ) I_\tau(V_h) I τ ( V h ) ; every forcing contribution is retained from level n 0 ( ν ) n_0(\nu) n 0 ( ν ) onward. Consequently, for ν = 3 , 4 \nu=3,4 ν = 3 , 4 ,
∑ n = ν − 1 m ∥ d ν − 1 V h n ∥ δ 2 ≲ τ C 2 2 ν − 3 2 ( ν − 1 ) / ( 2 ν − 3 ) ∑ n = 0 m − 1 ∥ V h n ∥ δ 2 (136) \sum_{n=\nu-1}^{m}\lVert d^{\nu-1}V_h^n\rVert_\delta^2\lesssim\tau C_2^{2\nu-3\over2(\nu-1)/(2\nu-3)}\sum_{n=0}^{m-1}\lVert V_h^n\rVert_\delta^2 \tag*{(136)} n = ν − 1 ∑ m ∥ d ν − 1 V h n ∥ δ 2 ≲ τ C 2 2 ( ν − 1 ) / ( 2 ν − 3 ) 2 ν − 3 n = 0 ∑ m − 1 ∥ V h n ∥ δ 2 (136)
+ τ 2 ∑ n = n 0 ( ν ) m ∥ R n ∥ Ω 2 + C I τ ( V h ) (137) \qquad+\tau^2\sum_{n=n_0(\nu)}^{m}\lVert R^n\rVert_\Omega^2+CI_\tau(V_h) \tag*{(137)} + τ 2 n = n 0 ( ν ) ∑ m ∥ R n ∥ Ω 2 + C I τ ( V h ) (137)
and, for AM5,
∑ n = 4 m ∥ d 4 V h n ∥ δ 2 ≲ c 0 2 ∑ n = 4 m − 1 ∥ d 3 V h n ∥ δ 2 (138) \sum_{n=4}^{m}\lVert d^4V_h^n\rVert_\delta^2\lesssim c_0^2\sum_{n=4}^{m-1}\lVert d^3V_h^n\rVert_\delta^2 \tag*{(138)} n = 4 ∑ m ∥ d 4 V h n ∥ δ 2 ≲ c 0 2 n = 4 ∑ m − 1 ∥ d 3 V h n ∥ δ 2 (138)
+ τ 2 ∑ n = n 0 ( 5 ) m ∥ R n ∥ Ω 2 + C I τ ( V h ) (139) \qquad+\tau^2\sum_{n=n_0(5)}^{m}\lVert R^n\rVert_\Omega^2+CI_\tau(V_h) \tag*{(139)} + τ 2 n = n 0 ( 5 ) ∑ m ∥ R n ∥ Ω 2 + C I τ ( V h ) (139)
Moreover, every horizon m ≥ n 0 ( ν ) m \ge n_0(\nu) m ≥ n 0 ( ν ) satisfies the endpoint bounds
∑ j = 1 ν − 2 ∥ d j V h m ∥ δ 2 ≲ c 0 2 max 0 ≤ n ≤ m ∥ V h n ∥ δ 2 + τ 2 ∑ n = n 0 ( ν ) m ∥ R n ∥ Ω 2 (140) \sum_{j=1}^{\nu-2} \lVert d^j V_h^m\rVert_\delta^2 \lesssim c_0^2 \max_{0\le n\le m} \lVert V_h^n\rVert_\delta^2 + \tau^2 \sum_{n=n_0(\nu)}^m \lVert R^n\rVert_\Omega^2 \tag*{(140)} j = 1 ∑ ν − 2 ∥ d j V h m ∥ δ 2 ≲ c 0 2 0 ≤ n ≤ m max ∥ V h n ∥ δ 2 + τ 2 n = n 0 ( ν ) ∑ m ∥ R n ∥ Ω 2 (140)
+ C L τ ( V h ) (141) + C L_\tau(V_h) \tag*{(141)} + C L τ ( V h ) (141)
and, for AM5,
∥ d 2 V h m − 1 ∥ δ 2 ≲ c 0 2 max 0 ≤ n ≤ m ∥ V h n ∥ δ 2 + τ 2 ∑ n = 4 m ∥ R n ∥ Ω 2 + C L τ ( V h ) (142) \lVert d^2 V_h^{m-1}\rVert_\delta^2 \lesssim c_0^2 \max_{0\le n\le m} \lVert V_h^n\rVert_\delta^2 + \tau^2 \sum_{n=4}^m \lVert R^n\rVert_\Omega^2 + C L_\tau(V_h) \tag*{(142)} ∥ d 2 V h m − 1 ∥ δ 2 ≲ c 0 2 0 ≤ n ≤ m max ∥ V h n ∥ δ 2 + τ 2 n = 4 ∑ m ∥ R n ∥ Ω 2 + C L τ ( V h ) (142)
Proof. We first derive the local recurrence, then sum its iterates to obtain the bulk estimates, and finally apply it at a fixed horizon to control the boundary increments. Apply d j − 1 d^{j-1} d j − 1 to (5.22) and multiply by τ \tau τ . Since d j − 1 V h n = d j − 1 V h n − 1 + d j V h n d^{j-1}V_h^n=d^{j-1}V_h^{n-1}+d^jV_h^n d j − 1 V h n = d j − 1 V h n − 1 + d j V h n , the current b 0 b_0 b 0 term can be collected exactly in the left-hand operator. The resulting equation is
( T b 0 τ d j V h n , T δ W ) Ω = − τ ( G ∑ q = 0 ν − 2 γ q d j − 1 V h n − 1 − q , T δ W ) Ω (143) (T_{b_0\tau}d^jV_h^n,T_\delta W)_\Omega=-\tau\left(G\sum_{q=0}^{\nu-2}\gamma_qd^{j-1}V_h^{n-1-q},T_\delta W\right)_\Omega \tag*{(143)} ( T b 0 τ d j V h n , T δ W ) Ω = − τ ( G q = 0 ∑ ν − 2 γ q d j − 1 V h n − 1 − q , T δ W ) Ω (143)
+ τ ( d j − 1 R n , T δ W ) Ω ∀ W ∈ V h (144) +\tau(d^{j-1}R^n,T_\delta W)_\Omega\qquad\forall W\in V_h \tag*{(144)} + τ ( d j − 1 R n , T δ W ) Ω ∀ W ∈ V h (144)
This displays the previous-value stencil explicitly. With W = d j V h n W=d^jV_h^n W = d j V h n , skew-symmetry makes the current increment coercive:
( T b 0 τ d j V h n , T δ d j V h n ) Ω = ∥ d j V h n ∥ Ω 2 + b 0 τ δ ∥ G d j V h n ∥ Ω 2 (145) (T_{b_0\tau}d^jV_h^n,T_\delta d^jV_h^n)_\Omega=\lVert d^jV_h^n\rVert_\Omega^2+b_0\tau\delta\lVert Gd^jV_h^n\rVert_\Omega^2 \tag*{(145)} ( T b 0 τ d j V h n , T δ d j V h n ) Ω = ∥ d j V h n ∥ Ω 2 + b 0 τ δ ∥ G d j V h n ∥ Ω 2 (145)
The inverse inequality and the upper bound δ ≤ C δ h \delta\le C_\delta h δ ≤ C δ h give the norm equivalence
∥ V h ∥ Ω ≤ ∥ V h ∥ δ ≲ ∥ V h ∥ Ω ∀ V h ∈ V h (146) \lVert V_h\rVert_\Omega\le\lVert V_h\rVert_\delta\lesssim\lVert V_h\rVert_\Omega\qquad\forall V_h\in V_h \tag*{(146)} ∥ V h ∥ Ω ≤ ∥ V h ∥ δ ≲ ∥ V h ∥ Ω ∀ V h ∈ V h (146)
Equations (5.35)–(5.38) therefore prove the local estimates (5.25)–(5.27).
Iterating the previous-value recurrence ν − 1 \nu-1 ν − 1 times gives the solution factor ( τ / h ) ν − 1 (\tau/h)^{\nu-1} ( τ / h ) ν − 1 . Every resulting solution value has a time index strictly smaller than that of the highest increment. After squaring and summing through m m m , this gives the sum only through m − 1 m-1 m − 1 in (5.28). Finite convolution of the forcing values retains precisely the computed levels n 0 ( ν ) , … , m n_0(\nu),\ldots,m n 0 ( ν ) , … , m . The exact power relations
( τ / h ) 4 ≤ C 4 / 3 3 τ under (5.20) (147) (\tau/h)^4\le C_{4/3}^3\tau\qquad\text{under (5.20)} \tag*{(147)} ( τ / h ) 4 ≤ C 4/3 3 τ under (5.20) (147)
( τ / h ) 6 ≤ C 6 / 5 5 τ under (5.21) (148) (\tau/h)^6\le C_{6/5}^5\tau\qquad\text{under (5.21)} \tag*{(148)} ( τ / h ) 6 ≤ C 6/5 5 τ under (5.21) (148)
prove (5.28)–(5.29). For AM5, applying the recurrence once with j = 4 j=4 j = 4 gives only the previous third differences d 3 V h n − 1 − q d^3V_h^{n-1-q} d 3 V h n − 1 − q . Squaring, summing, and using τ / h ≤ c 0 C G − 1 \tau/h\le c_0C_G^{-1} τ / h ≤ c 0 C G − 1 proves (5.30)–(5.31). Finally, a finite iteration of the same previous-value recurrence at a fixed horizon proves (5.32)–(5.33). For AM5, applying the order-two endpoint argument at horizon m − 1 m-1 m − 1 and using the update at level m m m proves (5.34). ■
For continuous piecewise affine functions and constant-coefficient G G G , G V h GV_h G V h is elementwise constant. The elementwise mean of d 1 V h n d^1V_h^n d 1 V h n therefore drops out after applying G G G , avoiding the recurrence that produced the strengthened CFL powers above.
Lemma 5.2 (Piecewise affine increment bound). Assume that V h V_h V h consists of continuous piecewise affine functions on a shape-regular quasi-uniform mesh and that G G G has constant coefficients, so that
G V h ∣ K ∈ [ P 0 ( K ) ] n s y s ∀ V h ∈ V h , K ∈ T h (149) GV_h|_K\in[\mathbb{P}_0(K)]^{n_{\mathrm{sys}}}\qquad\forall V_h\in V_h,\quad K\in\mathcal{T}_h \tag*{(149)} G V h ∣ K ∈ [ P 0 ( K ) ] n sys ∀ V h ∈ V h , K ∈ T h (149)
Let V h V_h V h solve (130) for ν ∈ { 3 , 4 } \nu\in\{3,4\} ν ∈ { 3 , 4 } , and assume
τ ≤ C τ δ δ , 0 < δ ≤ C δ h , τ ≤ c 0 C G − 1 h (150) \tau\le C_{\tau\delta}\delta,\qquad0 < \delta\le C_{\delta}h,\qquad\tau\le c_0 C_G^{-1}h \tag*{(150)} τ ≤ C τ δ δ , 0 < δ ≤ C δ h , τ ≤ c 0 C G − 1 h (150)
Then every admissible horizon satisfies
∑ n = ν − 1 m ∥ d ν − 1 V h n ∥ δ 2 ≲ C τ δ ( τ h ) 2 τ δ ∑ n = n 0 ( ν ) m ∥ A τ V h n ∥ Ω 2 (151) \sum_{n=\nu-1}^{m} \left\lVert d^{\nu-1}V_h^n \right\rVert_{\delta}^{2}
\lesssim C_{\tau\delta}\left(\frac{\tau}{h}\right)^2\tau\delta
\sum_{n=n_0(\nu)}^{m} \left\lVert A_\tau V_h^n \right\rVert_{\Omega}^{2} \tag*{(151)} n = ν − 1 ∑ m d ν − 1 V h n δ 2 ≲ C τ δ ( h τ ) 2 τ δ n = n 0 ( ν ) ∑ m ∥ A τ V h n ∥ Ω 2 (151)
+ τ 2 ∑ n = n 0 ( ν ) m ∥ R n ∥ Ω 2 + C I τ ( V h ) (152) \qquad+ \tau^2 \sum_{n=n_0(\nu)}^{m} \left\lVert R^n \right\rVert_{\Omega}^{2} + C I_\tau(V_h) \tag*{(152)} + τ 2 n = n 0 ( ν ) ∑ m ∥ R n ∥ Ω 2 + C I τ ( V h ) (152)
The hidden constant is independent of h h h , τ \tau τ , and the horizon.
Proof. Let Q 0 Q_0 Q 0 be the elementwise L 2 L^2 L 2 -projection onto discontinuous piecewise constants, let Π h \Pi_h Π h be the L 2 L^2 L 2 -projection onto V h V_h V h , and set
Z n = ( Id − Q 0 ) d 1 V h n (153) Z^n = (\operatorname{Id}-Q_0)d^1V_h^n \tag*{(153)} Z n = ( Id − Q 0 ) d 1 V h n (153)
Thus Z n Z^n Z n is precisely the elementwise fluctuation of the first increment. Multiplication of (130) by τ \tau τ gives
( d 1 V h n + τ G M A A M ν V h n , T δ W ) Ω = τ ( R n , T δ W ) Ω (154) (d^1V_h^n+\tau GM A_{\mathrm{AM}\nu}V_h^n,T_\delta W)_\Omega
=\tau(R^n,T_\delta W)_\Omega \tag*{(154)} ( d 1 V h n + τ GM A AM ν V h n , T δ W ) Ω = τ ( R n , T δ W ) Ω (154)
Choose W = Π h Z n W=\Pi_h Z^n W = Π h Z n . Since d 1 V h n ∈ V h d^1V_h^n\in V_h d 1 V h n ∈ V h , projection orthogonality gives ( d 1 V h n , Π h Z n ) Ω = ∥ Z n ∥ Ω 2 (d^1V_h^n,\Pi_h Z^n)_\Omega=\lVert Z^n\rVert_\Omega^2 ( d 1 V h n , Π h Z n ) Ω = ∥ Z n ∥ Ω 2 . Moreover, G M A A M ν V h n GM A_{\mathrm{AM}\nu}V_h^n GM A AM ν V h n is elementwise constant and is orthogonal to Z n Z^n Z n , while D − T h n D_-T_h^n D − T h n is orthogonal to Π h Z n − Z n \Pi_h Z^n-Z^n Π h Z n − Z n . Combining the two terms containing G Π h Z n G\Pi_h Z^n G Π h Z n into A τ V h n A_\tau V_h^n A τ V h n therefore gives the exact identity
∥ Z n ∥ Ω 2 = τ ( R n , T δ Π h Z n ) Ω − τ ( A τ V h n , Π h Z n − Z n ) Ω (155) \lVert Z^n\rVert_\Omega^2
=\tau(R^n,T_\delta\Pi_h Z^n)_\Omega
-\tau(A_\tau V_h^n,\Pi_h Z^n-Z^n)_\Omega \tag*{(155)} ∥ Z n ∥ Ω 2 = τ ( R n , T δ Π h Z n ) Ω − τ ( A τ V h n , Π h Z n − Z n ) Ω (155)
− τ δ ( A τ V h n , G Π h Z n ) Ω (156) -\tau\delta(A_\tau V_h^n,G\Pi_h Z^n)_\Omega \tag*{(156)} − τ δ ( A τ V h n , G Π h Z n ) Ω (156)
The L 2 L^2 L 2 -stability of Π h \Pi_h Π h , the inverse inequality, and δ ≲ h \delta\lesssim h δ ≲ h show that
∥ Z n ∥ Ω ≲ τ ( ∥ A τ V h n ∥ Ω + ∥ R n ∥ Ω ) (157) \lVert Z^n\rVert_\Omega
\lesssim\tau\left(\lVert A_\tau V_h^n\rVert_\Omega+\lVert R^n\rVert_\Omega\right) \tag*{(157)} ∥ Z n ∥ Ω ≲ τ ( ∥ A τ V h n ∥ Ω + ∥ R n ∥ Ω ) (157)
Take one backward difference of (154) . At every level for which both equations are computed,
( d 2 V h n + τ G M A A M ν d 1 V h n , T δ W ) Ω = τ ( d 1 R n , T δ W ) Ω (158) (d^2V_h^n+\tau GM A_{\mathrm{AM}\nu}d^1V_h^n,T_\delta W)_\Omega
=\tau(d^1R^n,T_\delta W)_\Omega \tag*{(158)} ( d 2 V h n + τ GM A AM ν d 1 V h n , T δ W ) Ω = τ ( d 1 R n , T δ W ) Ω (158)
With W = d 2 V h n W=d^2V_h^n W = d 2 V h n , skew-symmetry removes ( d 2 V h n , G d 2 V h n ) Ω (d^2V_h^n,Gd^2V_h^n)_\Omega ( d 2 V h n , G d 2 V h n ) Ω . On each element, G d 1 V h n = G Z n Gd^1V_h^n=GZ^n G d 1 V h n = G Z n , and the affine inverse estimate gives
∥ G M A A M ν d 1 V h n ∥ Ω ≲ h − 1 ∑ j = 0 ν − 1 ∥ Z n − j ∥ Ω (159) \left\lVert GM A_{\mathrm{AM}\nu}d^1V_h^n\right\rVert_\Omega
\lesssim h^{-1}\sum_{j=0}^{\nu-1}\left\lVert Z^{n-j}\right\rVert_\Omega \tag*{(159)} GM A AM ν d 1 V h n Ω ≲ h − 1 j = 0 ∑ ν − 1 Z n − j Ω (159)
Using (157) , squaring, and summing in time yields the following estimate. Here the finite difference d 1 R n d^1R^n d 1 R n is bounded by its two adjacent values, and the fixed Adams stencil has uniformly bounded overlap:
∑ n = n 0 ( ν ) + 1 m ∥ d 2 V h n ∥ δ 2 ≲ τ 4 h 2 ∑ n = n 0 ( ν ) m ∥ A τ V h n ∥ Ω 2 + τ 2 ∑ n = n 0 ( ν ) m ∥ R n ∥ Ω 2 + C I τ ( V h ) (160) \sum_{n=n_0(\nu)+1}^{m}\left\lVert d^2V_h^n\right\rVert_\delta^2
\lesssim\frac{\tau^4}{h^2}\sum_{n=n_0(\nu)}^{m}\left\lVert A_\tau V_h^n\right\rVert_\Omega^2
+\tau^2\sum_{n=n_0(\nu)}^{m}\left\lVert R^n\right\rVert_\Omega^2
+C I_\tau(V_h) \tag*{(160)} n = n 0 ( ν ) + 1 ∑ m d 2 V h n δ 2 ≲ h 2 τ 4 n = n 0 ( ν ) ∑ m ∥ A τ V h n ∥ Ω 2 + τ 2 n = n 0 ( ν ) ∑ m ∥ R n ∥ Ω 2 + C I τ ( V h ) (160)
The finitely many differences touching the start-up stencil are included in I τ ( V h ) I_\tau(V_h) I τ ( V h ) ; their forcing terms are retained in the displayed forcing sum. For AM3, this already controls the required second increment; for AM4, the third increment is a finite difference of d 2 V h d^2V_h d 2 V h . Finally, τ ≤ C τ δ δ \tau\le C_{\tau\delta}\delta τ ≤ C τ δ δ converts τ 4 h − 2 \tau^4h^{-2} τ 4 h − 2 into at most C τ δ ( τ / h ) 2 − δ τ δ C_{\tau\delta}(\tau/h)^{2-\delta}\tau\delta C τ δ ( τ / h ) 2 − δ τ δ . This proves (151) –(152) . ∎
Exact AM3–AM5 Energy Decompositions
The signs of the bulk terms are decisive. AM3 and AM4 produce negative terms in their highest relevant increment. AM5 instead produces a positive third-increment term that can absorb its negative fourth-increment term under the standard CFL restriction. Recall the first computed indices n 0 ( 3 ) = 2 n_0(3)=2 n 0 ( 3 ) = 2 , n 0 ( 4 ) = 3 n_0(4)=3 n 0 ( 4 ) = 3 , and n 0 ( 5 ) = 4 n_0(5)=4 n 0 ( 5 ) = 4 , and define
E N ( ν ) = ∑ n = n 0 ( ν ) N ( d 1 U n , M A M ν U n ) δ (161) \mathcal{E}^{(\nu)}_N = \sum_{n=n_0(\nu)}^N \left(d^1 U^n,\mathcal{M}_{\mathrm{AM}\nu}U^n\right)_\delta \tag*{(161)} E N ( ν ) = n = n 0 ( ν ) ∑ N ( d 1 U n , M AM ν U n ) δ (161)
The common identity (3.2)–(3.3) then gives
τ ∑ n = n 0 ( ν ) N ( A τ U n , P δ U n ) Ω = E N ( ν ) + τ δ ∑ n = n 0 ( ν ) N ∥ A τ U n ∥ Ω 2 (162) \tau\sum_{n=n_0(\nu)}^N \left(A_\tau U^n,P_\delta U^n\right)_\Omega= \mathcal{E}^{(\nu)}_N+\tau\delta\sum_{n=n_0(\nu)}^N\left\|A_\tau U^n\right\|_\Omega^2 \tag*{(162)} τ n = n 0 ( ν ) ∑ N ( A τ U n , P δ U n ) Ω = E N ( ν ) + τ δ n = n 0 ( ν ) ∑ N ∥ A τ U n ∥ Ω 2 (162)
Lemma 5.3 (Exact Adams–Moulton decompositions). For ν = 3 , 4 , 5 \nu=3,4,5 ν = 3 , 4 , 5 ,
E N ( ν ) = B N ( ν ) + B T N ( ν ) (163) \mathcal{E}^{(\nu)}_N=\mathcal{B}^{(\nu)}_N+\mathrm{BT}^{(\nu)}_N \tag*{(163)} E N ( ν ) = B N ( ν ) + BT N ( ν ) (163)
where the bulk and boundary terms are as follows.
AM3.
B N ( 3 ) = − 1 24 ∑ n = 2 N ∥ d 2 U n ∥ δ 2 (164) \mathcal{B}^{(3)}_N=-\frac{1}{24}\sum_{n=2}^N\left\|d^2U^n\right\|_\delta^2 \tag*{(164)} B N ( 3 ) = − 24 1 n = 2 ∑ N d 2 U n δ 2 (164)
B T N ( 3 ) = 1 2 ( ∥ U N ∥ δ 2 − ∥ U 1 ∥ δ 2 ) − 1 24 ( ∥ d 1 U N ∥ δ 2 − ∥ d 1 U 1 ∥ δ 2 ) (165) \mathrm{BT}^{(3)}_N=\frac{1}{2}\left(\left\|U^N\right\|_\delta^2-\left\|U^1\right\|_\delta^2\right)-\frac{1}{24}\left(\left\|d^1U^N\right\|_\delta^2-\left\|d^1U^1\right\|_\delta^2\right) \tag*{(165)} BT N ( 3 ) = 2 1 ( U N δ 2 − U 1 δ 2 ) − 24 1 ( d 1 U N δ 2 − d 1 U 1 δ 2 ) (165)
AM4.
B N ( 4 ) = − 1 48 ∑ n = 3 N ∥ d 3 U n ∥ δ 2 (166) \mathcal{B}^{(4)}_N=-\frac{1}{48}\sum_{n=3}^N\left\|d^3U^n\right\|_\delta^2 \tag*{(166)} B N ( 4 ) = − 48 1 n = 3 ∑ N d 3 U n δ 2 (166)
B T N ( 4 ) = 1 2 ( ∥ U N ∥ δ 2 − ∥ U 2 ∥ δ 2 ) − 1 24 ( ∥ d 1 U N ∥ δ 2 − ∥ d 1 U 2 ∥ δ 2 ) (167) \mathrm{BT}^{(4)}_N=\frac{1}{2}\left(\left\|U^N\right\|_\delta^2-\left\|U^2\right\|_\delta^2\right)-\frac{1}{24}\left(\left\|d^1U^N\right\|_\delta^2-\left\|d^1U^2\right\|_\delta^2\right) \tag*{(167)} BT N ( 4 ) = 2 1 ( U N δ 2 − U 2 δ 2 ) − 24 1 ( d 1 U N δ 2 − d 1 U 2 δ 2 ) (167)
− 1 24 ( ( d 1 U N , d 2 U N ) δ − ( d 1 U 2 , d 2 U 2 ) δ ) + 1 48 ( ∥ d 2 U 2 ∥ δ 2 − ∥ d 2 U N ∥ δ 2 ) (168) -\frac{1}{24}\left(\left(d^1U^N,d^2U^N\right)_\delta-\left(d^1U^2,d^2U^2\right)_\delta\right)+\frac{1}{48}\left(\left\|d^2U^2\right\|_\delta^2-\left\|d^2U^N\right\|_\delta^2\right) \tag*{(168)} − 24 1 ( ( d 1 U N , d 2 U N ) δ − ( d 1 U 2 , d 2 U 2 ) δ ) + 48 1 ( d 2 U 2 δ 2 − d 2 U N δ 2 ) (168)
AM5.
B N ( 5 ) = 3 160 ∑ n = 4 N − 1 ∥ d 3 U n ∥ δ 2 − 19 1440 ∑ n = 4 N ∥ d 4 U n ∥ δ 2 (169) \mathcal{B}^{(5)}_N=\frac{3}{160}\sum_{n=4}^{N-1}\left\|d^3U^n\right\|_\delta^2-\frac{19}{1440}\sum_{n=4}^N\left\|d^4U^n\right\|_\delta^2 \tag*{(169)} B N ( 5 ) = 160 3 n = 4 ∑ N − 1 d 3 U n δ 2 − 1440 19 n = 4 ∑ N d 4 U n δ 2 (169)
B T N ( 5 ) = 1 2 ( ∥ U N ∥ δ 2 − ∥ U 3 ∥ δ 2 ) − 1 24 ( ∥ d 1 U N ∥ δ 2 − ∥ d 1 U 3 ∥ δ 2 ) (170) \mathrm{BT}^{(5)}_N=\frac{1}{2}\left(\left\|U^N\right\|_\delta^2-\left\|U^3\right\|_\delta^2\right)-\frac{1}{24}\left(\left\|d^1U^N\right\|_\delta^2-\left\|d^1U^3\right\|_\delta^2\right) \tag*{(170)} BT N ( 5 ) = 2 1 ( U N δ 2 − U 3 δ 2 ) − 24 1 ( d 1 U N δ 2 − d 1 U 3 δ 2 ) (170)
− 1 24 ( d 1 U N , d 2 U N ) δ + 1 24 ( d 1 U 3 , d 2 U 3 ) δ − 1 48 ∥ d 2 U N ∥ δ 2 + 1 48 ∥ d 2 U 3 ∥ δ 2 + 19 1440 ∥ d 2 U N − 1 ∥ δ 2 − 19 1440 ∥ d 2 U 2 ∥ δ 2 -\frac{1}{24}\left(d^1U^N,d^2U^N\right)_\delta+\frac{1}{24}\left(d^1U^3,d^2U^3\right)_\delta
-\frac{1}{48}\left\|d^2U^N\right\|_\delta^2+\frac{1}{48}\left\|d^2U^3\right\|_\delta^2+\frac{19}{1440}\left\|d^2U^{N-1}\right\|_\delta^2-\frac{19}{1440}\left\|d^2U^2\right\|_\delta^2 − 24 1 ( d 1 U N , d 2 U N ) δ + 24 1 ( d 1 U 3 , d 2 U 3 ) δ − 48 1 d 2 U N δ 2 + 48 1 d 2 U 3 δ 2 + 1440 19 d 2 U N − 1 δ 2 − 1440 19 d 2 U 2 δ 2
− 19 720 ( d 1 U N , d 3 U N ) δ + 19 720 ( d 1 U 3 , d 3 U 3 ) δ − 11 1440 ∥ d 3 U N ∥ δ 2 + 19 720 ∥ d 3 U 3 ∥ δ 2 (171) -\frac{19}{720}\left(d^1U^N,d^3U^N\right)_\delta+\frac{19}{720}\left(d^1U^3,d^3U^3\right)_\delta
-\frac{11}{1440}\left\|d^3U^N\right\|_\delta^2+\frac{19}{720}\left\|d^3U^3\right\|_\delta^2 \tag*{(171)} − 720 19 ( d 1 U N , d 3 U N ) δ + 720 19 ( d 1 U 3 , d 3 U 3 ) δ − 1440 11 d 3 U N δ 2 + 720 19 d 3 U 3 δ 2 (171)
Proof. Substitute (5.6)–(5.8) into (161) . The required summation identities are
∑ n = 2 N ( d 1 U n , d 2 U n ) δ = 1 2 ( ∥ d 1 U N ∥ δ 2 − ∥ d 1 U 1 ∥ δ 2 ) + 1 2 ∑ n = 2 N ∥ d 2 U n ∥ δ 2 (172) \sum_{n=2}^{N} (d^{1}U^{n},d^{2}U^{n})_{\delta} = \frac{1}{2}\left(\lVert d^{1}U^{N}\rVert_{\delta}^{2}-\lVert d^{1}U^{1}\rVert_{\delta}^{2}\right)+\frac{1}{2}\sum_{n=2}^{N}\lVert d^{2}U^{n}\rVert_{\delta}^{2} \tag*{(172)} n = 2 ∑ N ( d 1 U n , d 2 U n ) δ = 2 1 ( ∥ d 1 U N ∥ δ 2 − ∥ d 1 U 1 ∥ δ 2 ) + 2 1 n = 2 ∑ N ∥ d 2 U n ∥ δ 2 (172)
∑ n = 3 N ( d 1 U n , d 3 U n ) δ = ( d 1 U N , d 2 U N ) δ − ( d 1 U 2 , d 2 U 2 ) δ − ∑ n = 3 N − 1 ∥ d 2 U n ∥ δ 2 − 1 2 ( ∥ d 2 U 2 ∥ δ 2 + ∥ d 2 U N ∥ δ 2 ) + 1 2 ∑ n = 3 N ∥ d 3 U n ∥ δ 2 (173) \begin{aligned}
\sum_{n=3}^{N}(d^{1}U^{n},d^{3}U^{n})_{\delta}
&=(d^{1}U^{N},d^{2}U^{N})_{\delta}-(d^{1}U^{2},d^{2}U^{2})_{\delta}\\
&\quad-\sum_{n=3}^{N-1}\lVert d^{2}U^{n}\rVert_{\delta}^{2}-\frac{1}{2}\left(\lVert d^{2}U^{2}\rVert_{\delta}^{2}+\lVert d^{2}U^{N}\rVert_{\delta}^{2}\right)+\frac{1}{2}\sum_{n=3}^{N}\lVert d^{3}U^{n}\rVert_{\delta}^{2} \tag*{(173)}
\end{aligned} n = 3 ∑ N ( d 1 U n , d 3 U n ) δ = ( d 1 U N , d 2 U N ) δ − ( d 1 U 2 , d 2 U 2 ) δ − n = 3 ∑ N − 1 ∥ d 2 U n ∥ δ 2 − 2 1 ( ∥ d 2 U 2 ∥ δ 2 + ∥ d 2 U N ∥ δ 2 ) + 2 1 n = 3 ∑ N ∥ d 3 U n ∥ δ 2 (173)
∑ n = 4 N ( d 1 U n , d 4 U n ) δ = ( d 1 U N , d 3 U N ) δ − ( d 1 U 3 , d 3 U 3 ) δ − 1 2 ( ∥ d 2 U N − 1 ∥ δ 2 − ∥ d 2 U 2 ∥ δ 2 ) − ∥ d 3 U 3 ∥ δ 2 − 1 2 ∥ d 3 U N ∥ δ 2 − 3 2 ∑ n = 4 N − 1 ∥ d 3 U n ∥ δ 2 + 1 2 ∑ n = 4 N ∥ d 4 U n ∥ δ 2 (174) \begin{aligned}
\sum_{n=4}^{N}(d^{1}U^{n},d^{4}U^{n})_{\delta}
&=(d^{1}U^{N},d^{3}U^{N})_{\delta}-(d^{1}U^{3},d^{3}U^{3})_{\delta}\\
&\quad-\frac{1}{2}\left(\lVert d^{2}U^{N-1}\rVert_{\delta}^{2}-\lVert d^{2}U^{2}\rVert_{\delta}^{2}\right)-\lVert d^{3}U^{3}\rVert_{\delta}^{2}-\frac{1}{2}\lVert d^{3}U^{N}\rVert_{\delta}^{2}\\
&\quad-\frac{3}{2}\sum_{n=4}^{N-1}\lVert d^{3}U^{n}\rVert_{\delta}^{2}+\frac{1}{2}\sum_{n=4}^{N}\lVert d^{4}U^{n}\rVert_{\delta}^{2} \tag*{(174)}
\end{aligned} n = 4 ∑ N ( d 1 U n , d 4 U n ) δ = ( d 1 U N , d 3 U N ) δ − ( d 1 U 3 , d 3 U 3 ) δ − 2 1 ( ∥ d 2 U N − 1 ∥ δ 2 − ∥ d 2 U 2 ∥ δ 2 ) − ∥ d 3 U 3 ∥ δ 2 − 2 1 ∥ d 3 U N ∥ δ 2 − 2 3 n = 4 ∑ N − 1 ∥ d 3 U n ∥ δ 2 + 2 1 n = 4 ∑ N ∥ d 4 U n ∥ δ 2 (174)
Collecting like terms gives (5.56)–(5.63). In particular, the AM4 bulk coefficient is − 1 / 48 -1/48 − 1/48 , and its initial energy is at level 2. These identities also display why AM5 has a different stability mechanism from AM3 and AM4. ■
Lower Bounds for the Energy Decompositions
The stability proof requires a lower bound for E N ( ν ) \mathcal{E}_{N}^{(\nu)} E N ( ν ) . We therefore estimate its bulk and boundary contributions separately and then recombine them in the stability theorem. The direction of the bulk estimate is essential.
Lemma 5.4 (Bulk lower bounds). Let V h V_h V h solve (5.22), and evaluate the bulk terms on this sequence. Under (5.19) and the relevant strengthened condition, the AM3 and AM4 bulk terms satisfy
B m ( ν ) ≥ − C τ C 2 ν − 3 2 / ( ν − 1 ) / ( 2 ν − 3 ) ∑ n = 0 m − 1 ∥ V h n ∥ δ 2 (175) \mathcal{B}_{m}^{(\nu)}\geq-C\tau C_{2\nu-3}^{2/(\nu-1)/(2\nu-3)}\sum_{n=0}^{m-1}\lVert V_h^n\rVert_{\delta}^{2} \tag*{(175)} B m ( ν ) ≥ − C τ C 2 ν − 3 2/ ( ν − 1 ) / ( 2 ν − 3 ) n = 0 ∑ m − 1 ∥ V h n ∥ δ 2 (175)
− C τ 2 ∑ n = n 0 ( ν ) m ∥ R n ∥ Ω 2 − C I τ ( V h ) (176) -C\tau^{2}\sum_{n=n_{0}(\nu)}^{m}\lVert R^{n}\rVert_{\Omega}^{2}-C\mathcal{I}_{\tau}(V_h) \tag*{(176)} − C τ 2 n = n 0 ( ν ) ∑ m ∥ R n ∥ Ω 2 − C I τ ( V h ) (176)
For AM5,
B m ( 5 ) ≥ ( 3 160 − C c 0 2 ) ∑ n = 4 m − 1 ∥ d 3 V h n ∥ δ 2 (177) \mathcal{B}_{m}^{(5)}\geq\left(\frac{3}{160}-Cc_{0}^{2}\right)\sum_{n=4}^{m-1}\lVert d^{3}V_h^n\rVert_{\delta}^{2} \tag*{(177)} B m ( 5 ) ≥ ( 160 3 − C c 0 2 ) n = 4 ∑ m − 1 ∥ d 3 V h n ∥ δ 2 (177)
− C τ 2 ∑ n = n 0 ( 5 ) m ∥ R n ∥ Ω 2 − C I τ ( V h ) (178) -C\tau^{2}\sum_{n=n_{0}(5)}^{m}\lVert R^{n}\rVert_{\Omega}^{2}-C\mathcal{I}_{\tau}(V_h) \tag*{(178)} − C τ 2 n = n 0 ( 5 ) ∑ m ∥ R n ∥ Ω 2 − C I τ ( V h ) (178)
Proof. For AM3 and AM4, combine the negative identities (5.56) and (5.58) with (5.28)–(5.29). This gives a lower bound, not an upper bound. For AM5, substitute (5.30)–(5.31) into (5.61). ■
Lemma 5.5 (Boundary lower bound). Let V h V_h V h solve (5.22). For every truncated horizon J J J , n 0 ( ν ) ≤ J ≤ N n_{0}(\nu)\leq J\leq N n 0 ( ν ) ≤ J ≤ N , choose m J m_J m J so that
∥ V h m J ∥ δ = max n 0 ( ν ) ≤ n ≤ J ∥ V h n ∥ δ (179) \lVert V_h^{m_J}\rVert_{\delta}=\max_{n_{0}(\nu)\leq n\leq J}\lVert V_h^n\rVert_{\delta} \tag*{(179)} ∥ V h m J ∥ δ = n 0 ( ν ) ≤ n ≤ J max ∥ V h n ∥ δ (179)
For c 0 c_0 c 0 sufficiently small,
B T m J ( ν ) ≥ ( 1 2 − C c 0 2 ) ∥ V h m J ∥ δ 2 − C I τ ( V h ) (180) \mathrm{BT}^{(\nu)}_{m_J} \ge\left(\frac{1}{2}-C c_0^2\right)\left\|V_h^{m_J}\right\|_{\delta}^{2}-C\mathcal{I}_{\tau}(V_h) \tag*{(180)} BT m J ( ν ) ≥ ( 2 1 − C c 0 2 ) ∥ V h m J ∥ δ 2 − C I τ ( V h ) (180)
− C τ 2 ∑ n = n 0 ( ν ) m J ∥ R n ∥ Ω 2 (181) -C\tau^2\sum_{n=n_0(\nu)}^{m_J}\left\|R^n\right\|_{\Omega}^{2} \tag*{(181)} − C τ 2 n = n 0 ( ν ) ∑ m J ∥ R n ∥ Ω 2 (181)
Proof. Apply Young’s inequality to the endpoint cross terms in (5.57), (5.59)–(5.60), and (5.62)–(5.63). The previous-value recurrence gives (5.32)–(5.34). It bounds every final increment occurring in the boundary formulas by the maximum graph energy, the full squared forcing contribution displayed in (5.75), and start-up data. The terms at levels 1, 2, or 3 have no fixed sign and are therefore bounded below by − C I τ ( V h ) -C\mathcal{I}_{\tau}(V_h) − C I τ ( V h ) . Young’s inequality then absorbs the remaining final-increment terms into the leading one-half of the final energy for c 0 c_0 c 0 sufficiently small. ■
Stability and Error Estimates
Adams Method-Adapted Norm. The norm must reflect the bulk term available for each method. For the AMν \nu ν operator A τ A_\tau A τ , use n 0 = n 0 ( ν ) n_0=n_0(\nu) n 0 = n 0 ( ν ) in I τ \mathcal{I}_\tau I τ , and define the common graph energy by
e τ , ν n ( V ) = ∥ T δ V n ∥ Ω 2 (182) e_{\tau,\nu}^{n}(V)=\left\|T_{\delta}V^{n}\right\|_{\Omega}^{2} \tag*{(182)} e τ , ν n ( V ) = ∥ T δ V n ∥ Ω 2 (182)
For AM3 and AM4, the dissipation consists only of the material residual:
d τ , ν n ( V ) = δ ∥ A τ V n ∥ Ω 2 ν ∈ { 3 , 4 } (183) d_{\tau,\nu}^{n}(V)=\delta\left\|A_{\tau}V^{n}\right\|_{\Omega}^{2}\qquad\nu\in\{3,4\} \tag*{(183)} d τ , ν n ( V ) = δ ∥ A τ V n ∥ Ω 2 ν ∈ { 3 , 4 } (183)
For AM5, the positive third-increment bulk term supplies an additional dissipation component, so we define
d τ , 5 4 ( V ) = δ ∥ A τ V 4 ∥ Ω 2 (184) d_{\tau,5}^{4}(V)=\delta\left\|A_{\tau}V^{4}\right\|_{\Omega}^{2} \tag*{(184)} d τ , 5 4 ( V ) = δ A τ V 4 Ω 2 (184)
d τ , 5 n ( V ) = δ ∥ A τ V n ∥ Ω 2 + τ − 1 ∥ d 3 V n − 1 ∥ δ 2 n ≥ 5 (185) d_{\tau,5}^{n}(V)=\delta\left\|A_{\tau}V^{n}\right\|_{\Omega}^{2}+\tau^{-1}\left\|d^{3}V^{n-1}\right\|_{\delta}^{2}\qquad n\ge5 \tag*{(185)} d τ , 5 n ( V ) = δ ∥ A τ V n ∥ Ω 2 + τ − 1 d 3 V n − 1 δ 2 n ≥ 5 (185)
Thus the corresponding sequence norms are exactly
∥ V ∥ E ν , m 2 = max n 0 ( ν ) ≤ n ≤ m ∥ T δ V n ∥ Ω 2 + τ δ ∑ n = n 0 ( ν ) m ∥ A τ V n ∥ Ω 2 ν ∈ { 3 , 4 } (186) \left\|V\right\|_{\mathcal{E}_{\nu,m}}^{2}
=
\max_{n_0(\nu)\le n\le m}\left\|T_{\delta}V^{n}\right\|_{\Omega}^{2}
+\tau\delta\sum_{n=n_0(\nu)}^{m}\left\|A_{\tau}V^{n}\right\|_{\Omega}^{2}
\qquad\nu\in\{3,4\} \tag*{(186)} ∥ V ∥ E ν , m 2 = n 0 ( ν ) ≤ n ≤ m max ∥ T δ V n ∥ Ω 2 + τ δ n = n 0 ( ν ) ∑ m ∥ A τ V n ∥ Ω 2 ν ∈ { 3 , 4 } (186)
∥ V ∥ E 5 , m 2 = max 4 ≤ n ≤ m ∥ T δ V n ∥ Ω 2 + τ δ ∑ n = 4 m ∥ A τ V n ∥ Ω 2 + ∑ n = 4 m − 1 ∥ d 3 V n ∥ δ 2 (187) \left\|V\right\|_{\mathcal{E}_{5,m}}^{2}
=
\max_{4\le n\le m}\left\|T_{\delta}V^{n}\right\|_{\Omega}^{2}
+\tau\delta\sum_{n=4}^{m}\left\|A_{\tau}V^{n}\right\|_{\Omega}^{2}
+\sum_{n=4}^{m-1}\left\|d^{3}V^{n}\right\|_{\delta}^{2} \tag*{(187)} ∥ V ∥ E 5 , m 2 = 4 ≤ n ≤ m max ∥ T δ V n ∥ Ω 2 + τ δ n = 4 ∑ m ∥ A τ V n ∥ Ω 2 + n = 4 ∑ m − 1 d 3 V n δ 2 (187)
With these method-dependent choices understood, define the corresponding lifted dual norm by
∥ R ∥ E ν , N ∗ = max n 0 ( ν ) ≤ m ≤ N sup V n = 0 for n ∉ { 0 , … , m } I τ ( V ) + ∥ V ∥ E ν , m 2 > 0 τ ∑ n = n 0 ( ν ) m ( R n , P δ V n ) Ω ( I τ ( V ) + ∥ V ∥ E ν , m 2 ) 1 / 2 (188) \left\|R\right\|_{\mathcal{E}^{*}_{\nu,N}}
=
\max_{n_0(\nu)\le m\le N}
\sup_{\substack{V^{n}=0\ \text{for }n\notin\{0,\ldots,m\}\\ \mathcal{I}_{\tau}(V)+\left\|V\right\|_{\mathcal{E}_{\nu,m}}^{2}>0}}
\frac{\tau\sum_{n=n_0(\nu)}^{m}(R^{n},P_{\delta}V^{n})_{\Omega}}
{\left(\mathcal{I}_{\tau}(V)+\left\|V\right\|_{\mathcal{E}_{\nu,m}}^{2}\right)^{1/2}} \tag*{(188)} ∥ R ∥ E ν , N ∗ = n 0 ( ν ) ≤ m ≤ N max V n = 0 for n ∈ / { 0 , … , m } I τ ( V ) + ∥ V ∥ E ν , m 2 > 0 sup ( I τ ( V ) + ∥ V ∥ E ν , m 2 ) 1/2 τ ∑ n = n 0 ( ν ) m ( R n , P δ V n ) Ω (188)
The finite start-up stencil is included in the denominator, and the test sequence is zero-extended after its horizon before P δ P_{\delta} P δ is applied. The shift in (5.79) makes the last term of (5.81) agree exactly with the positive AM5 bulk sum in (5.61).
Theorem 5.6 (Adams–Moulton stability). Assume the inverse inequality (5.18) and the CFL condition (5.19), with c 0 c_0 c 0 sufficiently small. For AM3 also assume (5.20), and for AM4 assume (5.21). If V h V_h V h solves (5.22), then
∥ V h ∥ E ν , N 2 ≲ I τ ( V h ) + ∥ R ∥ E ν , N ∗ 2 + τ ∥ R ∥ ℓ 2 ( n 0 ( ν ) , N ; H ) 2 (189) \left\|V_h\right\|_{\mathcal{E}_{\nu,N}}^{2}
\lesssim
\mathcal{I}_{\tau}(V_h)
+\left\|R\right\|_{\mathcal{E}^{*}_{\nu,N}}^{2}
+\tau\left\|R\right\|_{\ell^{2}(n_0(\nu),N;H)}^{2} \tag*{(189)} ∥ V h ∥ E ν , N 2 ≲ I τ ( V h ) + ∥ R ∥ E ν , N ∗ 2 + τ ∥ R ∥ ℓ 2 ( n 0 ( ν ) , N ; H ) 2 (189)
For AM3 and AM4, the hidden constant depends exponentially on T C 2 ( ν − 1 ) / ( 2 ν − 3 ) 2 ν − 3 T C_{2(\nu-1)/(2\nu-3)}^{2\nu-3} T C 2 ( ν − 1 ) / ( 2 ν − 3 ) 2 ν − 3 . For AM5 it is independent of a strengthened CFL constant. The theorem requires only 0 < δ ≤ C δ h 0 < \delta\le C_\delta h 0 < δ ≤ C δ h ; hence all three conclusions include the symmetric positive definite normal-equation choice δ = b 0 τ \delta= b_0\tau δ = b 0 τ under their stated CFL conditions.
Proof. The proof has two distinct branches. For AM3 and AM4, the negative bulk term is controlled by the increment recurrence and then by a discrete Gronwall argument. For AM5, the positive third-increment term absorbs the negative fourth-increment term, so no Gronwall step is needed. Use S δ V h n S_\delta V_h^n S δ V h n as the test function in (5.22), sum to an arbitrary horizon, and insert Lemma 5.3. Write
K ν = C 2 ( ν − 1 ) / ( 2 ν − 3 ) 2 ν − 3 , M J = max n 0 ( ν ) ≤ n ≤ J ∥ V h n ∥ δ 2 (190) K_\nu= C_{2(\nu-1)/(2\nu-3)}^{2\nu-3}, \qquad M_J = \max_{n_0(\nu)\le n\le J} \lVert V_h^n\rVert_\delta^2 \tag*{(190)} K ν = C 2 ( ν − 1 ) / ( 2 ν − 3 ) 2 ν − 3 , M J = n 0 ( ν ) ≤ n ≤ J max ∥ V h n ∥ δ 2 (190)
For each J J J , choose the running maximum index m J m_J m J from (5.73) and apply the energy identity on the truncated interval ending at m J m_J m J . For AM3 and AM4, Lemmas 5.4 and 5.5, together with the dual-norm definition and Young’s inequality, give
M J ≤ C A N + C τ K ν ∑ n = n 0 ( ν ) J − 1 M n (191) M_J \le C A_N + C\tau K_\nu\sum_{n=n_0(\nu)}^{J-1} M_n \tag*{(191)} M J ≤ C A N + C τ K ν n = n 0 ( ν ) ∑ J − 1 M n (191)
A N = I τ ( V h ) + ∥ R ∥ E ν ∗ , N 2 + τ ∥ R ∥ ℓ τ 2 ( n 0 ( ν ) , N ; H ) 2 (192) A_N = \mathcal{I}_\tau(V_h) + \lVert R\rVert_{\mathcal{E}_\nu^*,N}^2 + \tau\lVert R\rVert_{\ell_\tau^2(n_0(\nu),N;\mathcal{H})}^2 \tag*{(192)} A N = I τ ( V h ) + ∥ R ∥ E ν ∗ , N 2 + τ ∥ R ∥ ℓ τ 2 ( n 0 ( ν ) , N ; H ) 2 (192)
Crucially, the bulk estimate contains ∑ n = 0 m J − 1 ∥ V h n ∥ δ 2 \sum_{n=0}^{m_J-1}\lVert V_h^n\rVert_\delta^2 ∑ n = 0 m J − 1 ∥ V h n ∥ δ 2 , not the current value. Its start-up part is controlled by I τ ( V h ) \mathcal{I}_\tau(V_h) I τ ( V h ) , and its computed part is bounded by ∑ n = n 0 ( ν ) J − 1 M n \sum_{n=n_0(\nu)}^{J-1} M_n ∑ n = n 0 ( ν ) J − 1 M n . The dual forcing term is bounded by
∥ R ∥ E ν ∗ , N ( I τ ( V h ) + ∥ V h ∥ E ν , m J 2 ) 1 / 2 (193) \lVert R\rVert_{\mathcal{E}_\nu^*,N}\left(\mathcal{I}_\tau(V_h)+\lVert V_h\rVert_{\mathcal{E}_\nu,m_J}^2\right)^{1/2} \tag*{(193)} ∥ R ∥ E ν ∗ , N ( I τ ( V h ) + ∥ V h ∥ E ν , m J 2 ) 1/2 (193)
so Young’s inequality absorbs the running graph energy and the partial material-residual sum. The endpoint and bulk forcing defects are exactly C τ 2 ∑ n = n 0 ( ν ) m J ∥ R n ∥ Ω 2 C\tau^2\sum_{n=n_0(\nu)}^{m_J}\lVert R^n\rVert_{\Omega}^2 C τ 2 ∑ n = n 0 ( ν ) m J ∥ R n ∥ Ω 2 , which is bounded by the last term of A N A_N A N .
The standard discrete Gronwall lemma applied to (191) yields
M N ≤ C A N exp ( C T K ν ) (194) M_N \le C A_N \exp(C T K_\nu) \tag*{(194)} M N ≤ C A N exp ( C T K ν ) (194)
which displays the claimed dependence on the strengthened CFL constant. Applying the energy identity once more at the final horizon N N N , using (194) for all preceding solution terms, and again applying Young’s inequality controls the complete τ δ ∑ n = n 0 ( ν ) N ∥ A τ V h n ∥ Ω 2 \tau\delta\sum_{n=n_0(\nu)}^N\lVert A_\tau V_h^n\rVert_{\Omega}^2 τ δ ∑ n = n 0 ( ν ) N ∥ A τ V h n ∥ Ω 2 sum.
For AM5, choose c 0 c_0 c 0 so that the negative fourth-increment contribution in Lemma 5.4 is absorbed by the positive third-increment sum. The same running-maximum and dual-norm argument then controls the graph energy without Gronwall. Repeating it at N N N controls the full material-residual and third-increment sums. These arguments prove (5.83) for the general forcing R R R . ■
Corollary 5.7 (Piecewise affine stability under hyperbolic CFL). Assume the piecewise affine mapping property (5.41) and the parameter conditions (5.42), with c 0 c_0 c 0 sufficiently small. Then the AM3 and AM4 solutions satisfy (5.83) without the strengthened conditions (5.20)–(5.21). In particular, this includes the standard hyperbolic regime τ ≲ δ ≲ h \tau\lesssim\delta\lesssim h τ ≲ δ ≲ h , subject to the stated smallness of τ / h \tau/h τ / h .
Proof. At an arbitrary horizon, substitute Lemma 5.2 into the negative bulk terms (5.56) and (5.58). This gives
B m ( ν ) ≥ − C C τ δ 0 2 C δ τ δ ∑ n = n 0 ( ν ) m ∥ A τ V h n ∥ Ω 2 (195) \mathcal{B}_m^{(\nu)} \ge-C C_\tau\delta_0^2 C_\delta\tau\delta\sum_{n=n_0(\nu)}^m \lVert A_\tau V_h^n\rVert_{\Omega}^2 \tag*{(195)} B m ( ν ) ≥ − C C τ δ 0 2 C δ τ δ n = n 0 ( ν ) ∑ m ∥ A τ V h n ∥ Ω 2 (195)
− C τ 2 ∑ n = n 0 ( ν ) m ∥ R n ∥ Ω 2 − C I τ ( V h ) (196) {}- C\tau^2 \sum_{n=n_0(\nu)}^m \lVert R^n\rVert_{\Omega}^2 - C\mathcal{I}_\tau(V_h) \tag*{(196)} − C τ 2 n = n 0 ( ν ) ∑ m ∥ R n ∥ Ω 2 − C I τ ( V h ) (196)
Choose c 0 c_0 c 0 so that the first term is absorbed by the material-residual term in (5.54). Lemma 5.5 controls the endpoint terms under the same hyperbolic CFL condition. The maximum-horizon and final-horizon arguments in the proof of Theorem 5.6 now give (5.83); no Gronwall term and no strengthened CFL condition are needed. ■
For the original inhomogeneous A M ν \mathrm{AM}_{\nu} AM ν scheme, take R n = M A M ν F n R^n=\mathcal{M}_{\mathrm{AM}_{\nu}}F^n R n = M AM ν F n in (5.83). Thus its right-hand side is
I τ ( U h ) + ∥ M A M ν F ∥ ε , N 2 + τ ∥ M A M ν F ∥ η ( n 0 ( ν ) , N ; H ) 2 (197) \mathcal{I}_{\tau}(U_h)+\lVert\mathcal{M}_{\mathrm{AM}_{\nu}}F\rVert_{\varepsilon,N}^{2}+\tau\lVert\mathcal{M}_{\mathrm{AM}_{\nu}}F\rVert_{\eta(n_0(\nu),N;\mathcal{H})}^{2} \tag*{(197)} I τ ( U h ) + ∥ M AM ν F ∥ ε , N 2 + τ ∥ M AM ν F ∥ η ( n 0 ( ν ) , N ; H ) 2 (197)
Corollary 5.8 (Adams–Moulton error estimate). Assume either the stability hypotheses of Theorem 5.6 or, for A M 3 \mathrm{AM}3 AM 3 –A M 4 \mathrm{AM}4 AM 4 , the piecewise affine hypotheses of Corollary 5.7. Also assume the regularity hypotheses of Proposition 3.2 and projected start-up data. Then (3.35)–(3.40) holds with ν = 3 , 4 , 5 \nu=3,4,5 ν = 3 , 4 , 5 for A M 3 \mathrm{AM}3 AM 3 , A M 4 \mathrm{AM}4 AM 4 , and A M 5 \mathrm{AM}5 AM 5 , respectively.
Proof. Apply the stability estimate from Theorem 5.6 or Corollary 5.7 to the error equation (3.17) with general forcing R = A τ η − ρ R=A_{\tau}\eta-\rho R = A τ η − ρ . Use the residual identity (3.19) and (5.14)–(5.17). Proposition 3.2 applies to the lifted norm (5.82), since its denominator contains the graph maximum and the material-residual sum used in the proposition. Theorem 3.3 then gives (3.35)–(3.40) with the AM-specific norm (5.80) or (5.81). ■
Under the balanced choice δ ∼ h \delta\sim h δ ∼ h and τ ≲ h \tau\lesssim h τ ≲ h , this yields the concrete rate O ( h k + 1 / 2 + τ ν ) O(h^{k+1/2}+\tau^\nu) O ( h k + 1/2 + τ ν ) , under the relevant general-degree stability condition or the piecewise affine relaxation above.
Remark 5.9 (Scope of the piecewise affine relaxation). Corollary 5.7 depends essentially on affine elements and constant-coefficient first-order operators: it uses that the differentiated finite element function is elementwise constant. The abstract inverse inequality alone does not provide this cancellation at higher polynomial degree, for which the strengthened conditions therefore remain in the present proof. This distinction is consistent with the imaginary-axis stability picture for Adams methods [22 ] ; the strengthened exponents for skew-symmetric discretizations are discussed in [17 ] . Here the streamline-diffusion term supplies the additional control needed for P 1 \mathrm{P}_1 P 1 under the standard hyperbolic CFL condition. □
Explicit Schemes: Adams–Bashforth Methods
This section analyses A B 3 \mathrm{AB}3 AB 3 and A B 4 \mathrm{AB}4 AB 4 directly from their coefficient expansions. Because the current coefficient is b 0 = 0 b_0=0 b 0 = 0 , the methods are explicit and cannot be residual-minimisation normal equations. We first establish consistency, then derive exact energy decompositions and the increment bounds needed to control their negative bulk terms. This leads to stability and error estimates under the standard hyperbolic restriction τ ≲ h \tau\lesssim h τ ≲ h and the upper bound 0 < δ ≲ h 0<\delta\lesssim h 0 < δ ≲ h . The formulation remains nonsymmetric for every admissible δ \delta δ , since b 0 = 0 b_0=0 b 0 = 0 .
Coefficients and Increment Expansions. The first computed indices are n A B , 0 ( 3 ) = 3 n_{\mathrm{AB},0}(3)=3 n AB , 0 ( 3 ) = 3 and n A B , 0 ( 4 ) = 4 n_{\mathrm{AB},0}(4)=4 n AB , 0 ( 4 ) = 4 . The method averages in (2.11) have coefficients
A B 3 : ( b 0 , b 1 , b 2 , b 3 ) = ( 0 , 23 12 , − 16 12 , 5 12 ) (198) \mathrm{AB}3:\quad(b_0,b_1,b_2,b_3)=\left(0,\frac{23}{12},-\frac{16}{12},\frac{5}{12}\right) \tag*{(198)} AB 3 : ( b 0 , b 1 , b 2 , b 3 ) = ( 0 , 12 23 , − 12 16 , 12 5 ) (198)
A B 4 : ( b 0 , b 1 , b 2 , b 3 , b 4 ) = ( 0 , 55 24 , − 59 24 , 37 24 , − 9 24 ) (199) \mathrm{AB}4:\quad(b_0,b_1,b_2,b_3,b_4)=\left(0,\frac{55}{24},-\frac{59}{24},\frac{37}{24},-\frac{9}{24}\right) \tag*{(199)} AB 4 : ( b 0 , b 1 , b 2 , b 3 , b 4 ) = ( 0 , 24 55 , − 24 59 , 24 37 , − 24 9 ) (199)
Using the backward increments in (5.4)–(5.5), exact coefficient expansion gives
M A B 3 U n = U ~ n − 1 12 d 2 U n − 5 12 d 3 U n (200) \mathcal{M}_{\mathrm{AB}3}U^n=\widetilde{U}^n-\frac{1}{12}d^2U^n-\frac{5}{12}d^3U^n \tag*{(200)} M AB 3 U n = U n − 12 1 d 2 U n − 12 5 d 3 U n (200)
M A B 4 U n = U ~ n − 1 12 d 2 U n − 1 24 d 3 U n − 3 8 d 4 U n (201) \mathcal{M}_{\mathrm{AB}4}U^n=\widetilde{U}^n-\frac{1}{12}d^2U^n-\frac{1}{24}d^3U^n-\frac{3}{8}d^4U^n \tag*{(201)} M AB 4 U n = U n − 12 1 d 2 U n − 24 1 d 3 U n − 8 3 d 4 U n (201)
The relatively large coefficient − 3 / 8 -3/8 − 3/8 of the fourth increment in AB4 generates a negative fourth-increment bulk term, balanced by a positive third-increment term in the energy decomposition below.
Consistency of the Adams–Bashforth Methods
Set σ ( t ) = F ( t ) − G U ( t ) = ∂ t U ( t ) \sigma(t)=F(t)-GU(t)=\partial_tU(t) σ ( t ) = F ( t ) − G U ( t ) = ∂ t U ( t ) . For ABν \nu ν , let σ ν − 1 \sigma_{\nu-1} σ ν − 1 interpolate this exact time derivative at the past nodes
t n − j 1 ≤ j ≤ ν (202) t^{n-j}\qquad1\le j\le\nu \tag*{(202)} t n − j 1 ≤ j ≤ ν (202)
Unlike the Adams–Moulton rule, this interpolant uses no value at the current time. The Adams–Bashforth weights are the integrals of the corresponding extrapolating Lagrange basis over the current step. Hence,
∫ t n − 1 t n σ ν − 1 ( t ) d t = τ M A B ν σ ( t n ) (203) \int_{t^{n-1}}^{t^n}\sigma_{\nu-1}(t)\,\mathrm{d}t=\tau\mathcal{M}_{\mathrm{AB}\nu}\sigma(t^n) \tag*{(203)} ∫ t n − 1 t n σ ν − 1 ( t ) d t = τ M AB ν σ ( t n ) (203)
Subtracting this relation from (5.9) gives
ρ n = D τ − U ( t n ) + G M A B ν U ( t n ) − M A B ν F ( t n ) (204) \rho^n=D^-_\tau U(t^n)+G\mathcal{M}_{\mathrm{AB}\nu}U(t^n)-\mathcal{M}_{\mathrm{AB}\nu}F(t^n) \tag*{(204)} ρ n = D τ − U ( t n ) + G M AB ν U ( t n ) − M AB ν F ( t n ) (204)
∥ ρ n ∥ Ω ≤ C ν τ ν − 1 / 2 ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν , t n ; H ) , ν ∈ { 3 , 4 } (205) \lVert\rho^n\rVert_{\Omega}\le C_\nu\tau^{\nu-1/2}\left\lVert\partial_t^\nu(F-GU)\right\rVert_{L^2(t^{n-\nu},t^n;\mathcal{H})},\qquad\nu\in\{3,4\} \tag*{(205)} ∥ ρ n ∥ Ω ≤ C ν τ ν − 1/2 ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν , t n ; H ) , ν ∈ { 3 , 4 } (205)
Thus AB3 and AB4 have consistency orders three and four, respectively. In the notation of (2.19), take
C n , ν + 1 ( U , F ) = C ν ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν , t n ; H ) (206) C_{n,\nu+1}(U,F)=C_\nu\left\lVert\partial_t^\nu(F-GU)\right\rVert_{L^2(t^{n-\nu},t^n;\mathcal{H})} \tag*{(206)} C n , ν + 1 ( U , F ) = C ν ∥ ∂ t ν ( F − G U ) ∥ L 2 ( t n − ν , t n ; H ) (206)
The intervals overlap at most ν + 1 \nu+1 ν + 1 times, so this choice gives (2.20). As in the Adams–Moulton analysis, the dual-residual estimate also needs uniform control of the residual and square-summable control of its first forward difference. The integral remainder gives both bounds:
max ν ≤ n ≤ N ∥ ρ n ∥ Ω + ∥ D τ + ρ ∥ ℓ 2 ( ν , N − 1 ; H ) ≤ C ν τ ν ∥ ∂ t ν ( F − G U ) ∥ L ∞ ( 0 , T ; H ) (207) \max_{\nu\le n\le N}\lVert\rho^n\rVert_{\Omega}+\left\lVert D^+_\tau\rho\right\rVert_{\ell^2(\nu,N-1;\mathcal{H})}\le C_\nu\tau^\nu\left\lVert\partial_t^\nu(F-GU)\right\rVert_{L^\infty(0,T;\mathcal{H})} \tag*{(207)} ν ≤ n ≤ N max ∥ ρ n ∥ Ω + D τ + ρ ℓ 2 ( ν , N − 1 ; H ) ≤ C ν τ ν ∥ ∂ t ν ( F − G U ) ∥ L ∞ ( 0 , T ; H ) (207)
+ ∥ ∂ t ν + 1 ( F − G U ) ∥ L 2 ( 0 , T ; H ) (208) +\left\lVert\partial_t^{\nu+1}(F-GU)\right\rVert_{L^2(0,T;\mathcal{H})} \tag*{(208)} + ∂ t ν + 1 ( F − G U ) L 2 ( 0 , T ; H ) (208)
Consequently, (2.21)–(2.22) hold for AB3 and AB4 with the displayed norms included in C T , ν + 2 ♯ ( U , F ) C^\sharp_{T,\nu+2}(U,F) C T , ν + 2 ♯ ( U , F ) .
Exact AB3 and AB4 Energy Decompositions
Assume the inverse inequality (5.18) and
0 < δ ≤ C δ h , τ ≤ c A B C G − 1 h (209) 0<\delta\le C_\delta h,\qquad\tau\le c_{\mathrm{AB}}C_G^{-1}h \tag*{(209)} 0 < δ ≤ C δ h , τ ≤ c AB C G − 1 h (209)
where c A B > 0 c_{\mathrm{AB}}>0 c AB > 0 is sufficiently small. No lower bound proportional to h h h is imposed on δ \delta δ . For ν ∈ { 3 , 4 } \nu\in\{3,4\} ν ∈ { 3 , 4 } , define
E A B ν , m = ∑ n = ν m ( d 1 U n , M A B ν U n ) δ (210) \mathcal{E}_{\mathrm{AB}\nu,m}=\sum_{n=\nu}^{m}\left(d^1U^n,\mathcal{M}_{\mathrm{AB}\nu}U^n\right)_\delta \tag*{(210)} E AB ν , m = n = ν ∑ m ( d 1 U n , M AB ν U n ) δ (210)
The common tested identity gives
τ ∑ n = ν m ( A τ U n , P δ U n ) Ω = E A B ν , m + τ δ ∑ n = ν m ∥ A τ U n ∥ Ω 2 (211) \tau\sum_{n=\nu}^{m}\left(A_\tau U^n,P_\delta U^n\right)_\Omega=\mathcal{E}_{\mathrm{AB}\nu,m}+\tau\delta\sum_{n=\nu}^{m}\left\lVert A_\tau U^n\right\rVert_\Omega^2 \tag*{(211)} τ n = ν ∑ m ( A τ U n , P δ U n ) Ω = E AB ν , m + τ δ n = ν ∑ m ∥ A τ U n ∥ Ω 2 (211)
For both methods, the decomposition below pairs a positive bulk term in d ν − 1 U d^{\nu-1}U d ν − 1 U with a negative bulk term in d ν U d^\nu U d ν U . The increment estimate in Section 6.3 will show that the positive term absorbs the negative one under (6.12).
Lemma 6.1 (Exact Adams–Bashforth decompositions). For every admissible horizon m m m ,
E A B ν , m = B A B ν , m + B T A B ν , m ν ∈ { 3 , 4 } (212) \mathcal{E}_{\mathrm{AB}\nu,m}=\mathcal{B}_{\mathrm{AB}\nu,m}+\mathrm{BT}_{\mathrm{AB}\nu,m}\quad\nu\in\{3,4\} \tag*{(212)} E AB ν , m = B AB ν , m + BT AB ν , m ν ∈ { 3 , 4 } (212)
For AB3, the bulk term is
B A B 3 , m = 3 8 ∑ n = 3 m − 1 ∥ d 2 U n ∥ δ 2 − 5 24 ∑ n = 3 m ∥ d 3 U n ∥ δ 2 (213) \mathcal{B}_{\mathrm{AB}3,m}=\frac{3}{8}\sum_{n=3}^{m-1}\lVert d^{2}U^{n}\rVert_{\delta}^{2}-\frac{5}{24}\sum_{n=3}^{m}\lVert d^{3}U^{n}\rVert_{\delta}^{2} \tag*{(213)} B AB 3 , m = 8 3 n = 3 ∑ m − 1 ∥ d 2 U n ∥ δ 2 − 24 5 n = 3 ∑ m ∥ d 3 U n ∥ δ 2 (213)
and the boundary term is
B T A B 3 , m = 1 2 ( ∥ U m ∥ δ 2 − ∥ U 2 ∥ δ 2 ) − 1 24 ( ∥ d 1 U m ∥ δ 2 − ∥ d 1 U 2 ∥ δ 2 ) (214) \mathrm{BT}_{\mathrm{AB}3,m}=\frac{1}{2}\left(\lVert U^{m}\rVert_{\delta}^{2}-\lVert U^{2}\rVert_{\delta}^{2}\right)-\frac{1}{24}\left(\lVert d^{1}U^{m}\rVert_{\delta}^{2}-\lVert d^{1}U^{2}\rVert_{\delta}^{2}\right) \tag*{(214)} BT AB 3 , m = 2 1 ( ∥ U m ∥ δ 2 − ∥ U 2 ∥ δ 2 ) − 24 1 ( ∥ d 1 U m ∥ δ 2 − ∥ d 1 U 2 ∥ δ 2 ) (214)
− 5 12 ( ( d 1 U m , d 2 U m ) δ − ( d 1 U 2 , d 2 U 2 ) δ ) + 1 6 ∥ d 2 U m ∥ δ 2 + 5 24 ∥ d 2 U 2 ∥ δ 2 (215) -\frac{5}{12}\left(\left(d^{1}U^{m},d^{2}U^{m}\right)_{\delta}-\left(d^{1}U^{2},d^{2}U^{2}\right)_{\delta}\right)+\frac{1}{6}\lVert d^{2}U^{m}\rVert_{\delta}^{2}+\frac{5}{24}\lVert d^{2}U^{2}\rVert_{\delta}^{2} \tag*{(215)} − 12 5 ( ( d 1 U m , d 2 U m ) δ − ( d 1 U 2 , d 2 U 2 ) δ ) + 6 1 ∥ d 2 U m ∥ δ 2 + 24 5 ∥ d 2 U 2 ∥ δ 2 (215)
For AB4, the bulk term is
B A B 4 , m = 13 24 ∑ n = 4 m − 1 ∥ d 3 U n ∥ δ 2 − 3 16 ∑ n = 4 m ∥ d 4 U n ∥ δ 2 (216) \mathcal{B}_{\mathrm{AB}4,m}=\frac{13}{24}\sum_{n=4}^{m-1}\lVert d^{3}U^{n}\rVert_{\delta}^{2}-\frac{3}{16}\sum_{n=4}^{m}\lVert d^{4}U^{n}\rVert_{\delta}^{2} \tag*{(216)} B AB 4 , m = 24 13 n = 4 ∑ m − 1 ∥ d 3 U n ∥ δ 2 − 16 3 n = 4 ∑ m ∥ d 4 U n ∥ δ 2 (216)
and the boundary term is
B T A B 4 , m = 1 2 ( ∥ U m ∥ δ 2 − ∥ U 3 ∥ δ 2 ) − 1 24 ( ∥ d 1 U m ∥ δ 2 − ∥ d 1 U 3 ∥ δ 2 ) (217) \mathrm{BT}_{\mathrm{AB}4,m}=\frac{1}{2}\left(\lVert U^{m}\rVert_{\delta}^{2}-\lVert U^{3}\rVert_{\delta}^{2}\right)-\frac{1}{24}\left(\lVert d^{1}U^{m}\rVert_{\delta}^{2}-\lVert d^{1}U^{3}\rVert_{\delta}^{2}\right) \tag*{(217)} BT AB 4 , m = 2 1 ( ∥ U m ∥ δ 2 − ∥ U 3 ∥ δ 2 ) − 24 1 ( ∥ d 1 U m ∥ δ 2 − ∥ d 1 U 3 ∥ δ 2 ) (217)
− 1 24 ( ( d 1 U m , d 2 U m ) δ − ( d 1 U 3 , d 2 U 3 ) δ ) (218) -\frac{1}{24}\left(\left(d^{1}U^{m},d^{2}U^{m}\right)_{\delta}-\left(d^{1}U^{3},d^{2}U^{3}\right)_{\delta}\right) \tag*{(218)} − 24 1 ( ( d 1 U m , d 2 U m ) δ − ( d 1 U 3 , d 2 U 3 ) δ ) (218)
− 3 8 ( ( d 1 U m , d 3 U m ) δ − ( d 1 U 3 , d 3 U 3 ) δ ) − 1 48 ∥ d 2 U m ∥ δ 2 + 1 48 ∥ d 2 U 3 ∥ δ 2 (219) -\frac{3}{8}\left(\left(d^{1}U^{m},d^{3}U^{m}\right)_{\delta}-\left(d^{1}U^{3},d^{3}U^{3}\right)_{\delta}\right)-\frac{1}{48}\lVert d^{2}U^{m}\rVert_{\delta}^{2}+\frac{1}{48}\lVert d^{2}U^{3}\rVert_{\delta}^{2} \tag*{(219)} − 8 3 ( ( d 1 U m , d 3 U m ) δ − ( d 1 U 3 , d 3 U 3 ) δ ) − 48 1 ∥ d 2 U m ∥ δ 2 + 48 1 ∥ d 2 U 3 ∥ δ 2 (219)
+ 3 16 ∥ d 2 U m − 1 ∥ δ 2 − 3 16 ∥ d 2 U 2 ∥ δ 2 + 1 6 ∥ d 3 U m ∥ δ 2 + 3 8 ∥ d 3 U 3 ∥ δ 2 (220) +\frac{3}{16}\lVert d^{2}U^{m-1}\rVert_{\delta}^{2}-\frac{3}{16}\lVert d^{2}U^{2}\rVert_{\delta}^{2}+\frac{1}{6}\lVert d^{3}U^{m}\rVert_{\delta}^{2}+\frac{3}{8}\lVert d^{3}U^{3}\rVert_{\delta}^{2} \tag*{(220)} + 16 3 ∥ d 2 U m − 1 ∥ δ 2 − 16 3 ∥ d 2 U 2 ∥ δ 2 + 6 1 ∥ d 3 U m ∥ δ 2 + 8 3 ∥ d 3 U 3 ∥ δ 2 (220)
Proof. Insert (6.3) and (6.4) into (6.13). Apply the summation identities (5.64)–(5.68) with the AB3 lower index 3 and the AB4 lower index 4. For AB3, the interior coefficients are
5 12 − 1 24 = 3 8 (221) \frac{5}{12}-\frac{1}{24}=\frac{3}{8} \tag*{(221)} 12 5 − 24 1 = 8 3 (221)
− 1 2 + 5 12 = − 5 24 (222) -\frac{1}{2}+\frac{5}{12}=-\frac{5}{24} \tag*{(222)} − 2 1 + 12 5 = − 24 5 (222)
For AB4, the interior second-increment coefficient cancels, while
1 24 − 1 2 1 12 = 0 (223) \frac{1}{24}-\frac{1}{2}\frac{1}{12}=0 \tag*{(223)} 24 1 − 2 1 12 1 = 0 (223)
3 28 − 1 2 1 24 = 13 24 (224) \frac{3}{28}-\frac{1}{2}\frac{1}{24}=\frac{13}{24} \tag*{(224)} 28 3 − 2 1 24 1 = 24 13 (224)
− 1 2 3 8 = − 3 16 (225) -\frac{1}{2}\frac{3}{8}=-\frac{3}{16} \tag*{(225)} − 2 1 8 3 = − 16 3 (225)
Collecting the endpoints gives the boundary terms stated above. ■
Increment and Boundary Control
The explicit relation contains no current spatial term. After taking time differences, the highest increment is therefore controlled directly by the preceding increment, with the small factor τ / h \tau/h τ / h .
Lemma 6.2 (Explicit increment control). Let ν ∈ { 3 , 4 } \nu\in\{3,4\} ν ∈ { 3 , 4 } , and let V h V_h V h solve the generally forced ABν \nu ν equation
( A τ V h n , T δ W ) Ω = ( R n , T δ W ) Ω ∀ W ∈ V h , n ≥ ν (226) (A_\tau V_h^n,T_\delta W)_\Omega=(R^n,T_\delta W)_\Omega\quad\forall W\in V_h,\qquad n\ge\nu \tag*{(226)} ( A τ V h n , T δ W ) Ω = ( R n , T δ W ) Ω ∀ W ∈ V h , n ≥ ν (226)
Then
∑ n = ν m ∥ d ν V h n ∥ δ 2 ≲ c A B 2 ∑ n = ν m − 1 ∥ d ν − 1 V h n ∥ δ 2 + τ 2 ∑ n = ν m ∥ R n ∥ Ω 2 (227) \sum_{n=\nu}^{m}\lVert d^\nu V_h^n\rVert_\delta^2 \lesssim c_{\mathrm{AB}}^2\sum_{n=\nu}^{m-1}\lVert d^{\nu-1}V_h^n\rVert_\delta^2+\tau^2\sum_{n=\nu}^{m}\lVert R^n\rVert_\Omega^2 \tag*{(227)} n = ν ∑ m ∥ d ν V h n ∥ δ 2 ≲ c AB 2 n = ν ∑ m − 1 ∥ d ν − 1 V h n ∥ δ 2 + τ 2 n = ν ∑ m ∥ R n ∥ Ω 2 (227)
+ C I τ ( V h ) (228) {}+C I_\tau(V_h) \tag*{(228)} + C I τ ( V h ) (228)
At every horizon m ≥ ν m\ge\nu m ≥ ν ,
∑ j = 1 ν − 1 ∥ d j V h m ∥ δ 2 + ∥ d ν − 2 V h m − 1 ∥ δ 2 ≲ c A B 2 max 0 ≤ n ≤ m ∥ V h n ∥ δ 2 + τ 2 ∑ n = ν m ∥ R n ∥ Ω 2 (229) \sum_{j=1}^{\nu-1}\lVert d^jV_h^m\rVert_\delta^2+\lVert d^{\nu-2}V_h^{m-1}\rVert_\delta^2\lesssim c_{\mathrm{AB}}^2\max_{0\le n\le m}\lVert V_h^n\rVert_\delta^2+\tau^2\sum_{n=\nu}^{m}\lVert R^n\rVert_\Omega^2 \tag*{(229)} j = 1 ∑ ν − 1 ∥ d j V h m ∥ δ 2 + ∥ d ν − 2 V h m − 1 ∥ δ 2 ≲ c AB 2 0 ≤ n ≤ m max ∥ V h n ∥ δ 2 + τ 2 n = ν ∑ m ∥ R n ∥ Ω 2 (229)
+ C I τ ( V h ) (230) {}+C I_\tau(V_h) \tag*{(230)} + C I τ ( V h ) (230)
Proof. For n ≥ 2 ν − 1 n\ge2\nu-1 n ≥ 2 ν − 1 , apply d ν − 1 d^{\nu-1} d ν − 1 to (226) . Since the average has no current value, the exact differenced equation is
( d ν V h n , T δ W ) Ω = τ ( d ν − 1 R n , T δ W ) Ω (231) (d^\nu V_h^n,T_\delta W)_\Omega=\tau(d^{\nu-1}R^n,T_\delta W)_\Omega \tag*{(231)} ( d ν V h n , T δ W ) Ω = τ ( d ν − 1 R n , T δ W ) Ω (231)
− τ ( G M A B ν d ν − 1 V h n , T δ W ) Ω ∀ W ∈ V h (232) -\tau(GM_{\mathrm{AB}\nu}d^{\nu-1}V_h^n,T_\delta W)_\Omega\quad\forall W\in V_h \tag*{(232)} − τ ( G M AB ν d ν − 1 V h n , T δ W ) Ω ∀ W ∈ V h (232)
For AB4, the history term thus contains the previous third differences d 3 V h n − j d^3V_h^{n-j} d 3 V h n − j . Choose W = d ν V h n W=d^\nu V_h^n W = d ν V h n . Then skew-symmetry, (146) , and the fact that b 0 = 0 b_0=0 b 0 = 0 give
∥ d ν V h n ∥ δ ≲ τ h ∑ j = 1 ν ∥ d ν − 1 V h n − j ∥ δ (233) \lVert d^\nu V_h^n\rVert_\delta\lesssim\frac{\tau}{h}\sum_{j=1}^{\nu}\lVert d^{\nu-1}V_h^{n-j}\rVert_\delta \tag*{(233)} ∥ d ν V h n ∥ δ ≲ h τ j = 1 ∑ ν ∥ d ν − 1 V h n − j ∥ δ (233)
+ τ ∑ q = 0 ν − 1 ∥ R n − q ∥ Ω (234) {}+\tau\sum_{q=0}^{\nu-1}\lVert R^{n-q}\rVert_\Omega \tag*{(234)} + τ q = 0 ∑ ν − 1 ∥ R n − q ∥ Ω (234)
For n ≥ 2 ν − 1 n\ge2\nu-1 n ≥ 2 ν − 1 , every equation used in the difference is a computed AB equation. Squaring and summing proves the asserted bound over these indices. The finitely many indices ν ≤ n < 2 ν − 1 \nu\le n<2\nu-1 ν ≤ n < 2 ν − 1 follow by applying the explicit update successively; their solution terms are controlled by I τ ( V h ) I_\tau(V_h) I τ ( V h ) and their forcing terms by the displayed forcing sum. The endpoint estimate follows from the same previous-value recurrence applied successively for orders 1 1 1 through ν − 1 \nu-1 ν − 1 . ■
Lemma 6.3 (AB bulk and boundary lower bounds). Let V h V_h V h solve (226) , and evaluate the bulk and boundary terms on this sequence. Under (6.12),
B A B 3 , m ≥ ( 3 8 − C c A B 2 ) ∑ n = 3 m − 1 ∥ d 2 V h n ∥ δ 2 − C τ 2 ∑ n = 3 m ∥ R n ∥ Ω 2 − C I τ ( V h ) (235) \mathcal{B}_{\mathrm{AB3},m}\ge\left(\frac{3}{8}-C c_{\mathrm{AB}}^2\right)\sum_{n=3}^{m-1}\lVert d^2V_h^n\rVert_\delta^2-C\tau^2\sum_{n=3}^{m}\lVert R^n\rVert_\Omega^2-CI_\tau(V_h) \tag*{(235)} B AB3 , m ≥ ( 8 3 − C c AB 2 ) n = 3 ∑ m − 1 ∥ d 2 V h n ∥ δ 2 − C τ 2 n = 3 ∑ m ∥ R n ∥ Ω 2 − C I τ ( V h ) (235)
B A B 4 , m ≥ ( 13 24 − C c A B 2 ) ∑ n = 4 m − 1 ∥ d 3 V h n ∥ δ 2 − C τ 2 ∑ n = 4 m ∥ R n ∥ Ω 2 − C I τ ( V h ) (236) \mathcal{B}_{\mathrm{AB4},m}\ge\left(\frac{13}{24}-C c_{\mathrm{AB}}^2\right)\sum_{n=4}^{m-1}\lVert d^3V_h^n\rVert_\delta^2-C\tau^2\sum_{n=4}^{m}\lVert R^n\rVert_\Omega^2-CI_\tau(V_h) \tag*{(236)} B AB4 , m ≥ ( 24 13 − C c AB 2 ) n = 4 ∑ m − 1 ∥ d 3 V h n ∥ δ 2 − C τ 2 n = 4 ∑ m ∥ R n ∥ Ω 2 − C I τ ( V h ) (236)
For ν ∈ { 3 , 4 } \nu\in\{3,4\} ν ∈ { 3 , 4 } and every truncated horizon J J J , ν ≤ J ≤ N \nu\le J\le N ν ≤ J ≤ N , choose m ∗ ∈ { ν , … , J } m_*\in\{\nu,\ldots,J\} m ∗ ∈ { ν , … , J } so that ∥ V h m ∗ ∥ δ = max ν ≤ n ≤ J ∥ V h n ∥ δ \lVert V_h^{m_*}\rVert_\delta=\max_{\nu\le n\le J}\lVert V_h^n\rVert_\delta ∥ V h m ∗ ∥ δ = max ν ≤ n ≤ J ∥ V h n ∥ δ . Then
B T A B ν , m ∗ ≥ ( 1 2 − C c A B 2 ) ∥ V h m ∗ ∥ δ 2 − C I τ ( V h ) (237) \mathcal{B}_{\mathrm{TAB}\nu,m_*}\ge\left(\frac{1}{2}-C c_{\mathrm{AB}}^2\right)\lVert V_h^{m_*}\rVert_\delta^2-CI_\tau(V_h) \tag*{(237)} B TAB ν , m ∗ ≥ ( 2 1 − C c AB 2 ) ∥ V h m ∗ ∥ δ 2 − C I τ ( V h ) (237)
− C τ 2 ∑ n = ν m ∗ ∥ R n ∥ Ω 2 ν ∈ { 3 , 4 } (238) {}-C\tau^2\sum_{n=\nu}^{m_*}\lVert R^n\rVert_\Omega^2\qquad\nu\in\{3,4\} \tag*{(238)} − C τ 2 n = ν ∑ m ∗ ∥ R n ∥ Ω 2 ν ∈ { 3 , 4 } (238)
Proof. Substitute (227) –(228) into the negative highest-increment sums in (213) and (216) . This proves the bulk bounds.
For the boundary terms, apply Young’s inequality to every final cross term in (214) –(220) and use (229) –(230) . All start-up contributions in the exact formulas, regardless of their sign, are bounded below by − C I τ ( V h ) -C I_{\tau}(V_h) − C I τ ( V h ) . Forcing contributions from the endpoint increments enter the lower bound with the negative sign and the squared τ 2 \tau^2 τ 2 scaling shown in (238) . The final increment terms are absorbed into the leading final energy for c A B c_{\mathrm{AB}} c AB sufficiently small. ■
Stability and Error Estimates
AB Method-Adapted Norm. The positive bulk term controls d ν − 1 V d^{\nu-1}V d ν − 1 V , which must therefore be included in the method-adapted norm. For ν ∈ { 3 , 4 } \nu\in\{3,4\} ν ∈ { 3 , 4 } , use n 0 = ν n_0=\nu n 0 = ν in I τ I_{\tau} I τ , and define
e τ , A B ν n ( V ) = ∥ T δ V n ∥ Ω 2 (239) e_{\tau,\mathrm{AB}\nu}^{n}(V)=\lVert T_{\delta}V^{n}\rVert_{\Omega}^{2} \tag*{(239)} e τ , AB ν n ( V ) = ∥ T δ V n ∥ Ω 2 (239)
d τ , A B ν ν ( V ) = δ ∥ A τ V ν ∥ Ω 2 (240) d_{\tau,\mathrm{AB}\nu}^{\nu}(V)=\delta\lVert A_{\tau}V^{\nu}\rVert_{\Omega}^{2} \tag*{(240)} d τ , AB ν ν ( V ) = δ ∥ A τ V ν ∥ Ω 2 (240)
d τ , A B ν n ( V ) = δ ∥ A τ V n ∥ Ω 2 + τ − 1 ∥ d ν − 1 V n − 1 ∥ δ 2 n ≥ ν + 1 (241) d_{\tau,\mathrm{AB}\nu}^{n}(V)=\delta\lVert A_{\tau}V^{n}\rVert_{\Omega}^{2}+\tau^{-1}\lVert d^{\nu-1}V^{n-1}\rVert_{\delta}^{2}\qquad n\geq\nu+1 \tag*{(241)} d τ , AB ν n ( V ) = δ ∥ A τ V n ∥ Ω 2 + τ − 1 ∥ d ν − 1 V n − 1 ∥ δ 2 n ≥ ν + 1 (241)
The resulting sequence norm is
∥ V ∥ E A B ν , m 2 = max ν ≤ n ≤ m ∥ T δ V n ∥ Ω 2 + τ δ ∑ n = ν m ∥ A τ V n ∥ Ω 2 (242) \lVert V\rVert_{\mathcal{E}_{\mathrm{AB}\nu,m}}^{2}
=
\max_{\nu\leq n\leq m}\lVert T_{\delta}V^{n}\rVert_{\Omega}^{2}
+\tau\delta\sum_{n=\nu}^{m}\lVert A_{\tau}V^{n}\rVert_{\Omega}^{2} \tag*{(242)} ∥ V ∥ E AB ν , m 2 = ν ≤ n ≤ m max ∥ T δ V n ∥ Ω 2 + τ δ n = ν ∑ m ∥ A τ V n ∥ Ω 2 (242)
+ ∑ n = ν m − 1 ∥ d ν − 1 V n ∥ δ 2 (243) \qquad+\sum_{n=\nu}^{m-1}\lVert d^{\nu-1}V^{n}\rVert_{\delta}^{2} \tag*{(243)} + n = ν ∑ m − 1 ∥ d ν − 1 V n ∥ δ 2 (243)
Define its lifted dual norm by
∥ R ∥ E A B ν , N ∗ = max ν ≤ m ≤ N sup V n = 0 for n ∉ { 0 , … , m } I τ ( V ) + ∥ V ∥ E A B ν , m 2 > 0 τ ∑ n = ν m ( R n , P δ V n ) Ω ( I τ ( V ) + ∥ V ∥ E A B ν , m 2 ) 1 / 2 (244) \lVert R\rVert_{\mathcal{E}_{\mathrm{AB}\nu,N}^{*}}
=
\max_{\nu\leq m\leq N}
\sup_{\substack{V^{n}=0\ \text{for }n\notin\{0,\ldots,m\}\\ I_{\tau}(V)+\lVert V\rVert_{\mathcal{E}_{\mathrm{AB}\nu,m}}^{2}>0}}
\frac{\tau\sum_{n=\nu}^{m}(R^{n},P_{\delta}V^{n})_{\Omega}}
{\left(I_{\tau}(V)+\lVert V\rVert_{\mathcal{E}_{\mathrm{AB}\nu,m}}^{2}\right)^{1/2}} \tag*{(244)} ∥ R ∥ E AB ν , N ∗ = ν ≤ m ≤ N max V n = 0 for n ∈ / { 0 , … , m } I τ ( V ) + ∥ V ∥ E AB ν , m 2 > 0 sup ( I τ ( V ) + ∥ V ∥ E AB ν , m 2 ) 1/2 τ ∑ n = ν m ( R n , P δ V n ) Ω (244)
The denominator controls the finite start-up stencil, and the test sequence is zero-extended after its horizon before P δ P_{\delta} P δ is applied.
Theorem 6.4 (Adams–Bashforth stability). Assume (126) and (209) , with c A B c_{\mathrm{AB}} c AB sufficiently small. For A B ν \mathrm{AB}\nu AB ν , ν ∈ { 3 , 4 } \nu\in\{3,4\} ν ∈ { 3 , 4 } , let V h V_h V h solve (226) . Then
∥ V h ∥ E A B ν , N 2 ≲ I τ ( V h ) + ∥ R ∥ E A B ν , N ∗ 2 + τ ∥ R ∥ ℓ τ 2 ( ν , N ; H ) 2 (245) \lVert V_h\rVert_{\mathcal{E}_{\mathrm{AB}\nu,N}}^{2}
\lesssim I_{\tau}(V_h)+\lVert R\rVert_{\mathcal{E}_{\mathrm{AB}\nu,N}^{*}}^{2}
+\tau\lVert R\rVert_{\ell_{\tau}^{2}(\nu,N;\mathcal{H})}^{2} \tag*{(245)} ∥ V h ∥ E AB ν , N 2 ≲ I τ ( V h ) + ∥ R ∥ E AB ν , N ∗ 2 + τ ∥ R ∥ ℓ τ 2 ( ν , N ; H ) 2 (245)
Proof. The argument follows the same absorption mechanism as for AM5: the positive lower-increment bulk term controls the negative highest-increment term, so no discrete Gronwall step is required. Use (211) at an arbitrary horizon and insert Lemma 6.1. For each truncated horizon J J J , apply the identity at the running maximum index m ∗ m^{*} m ∗ from Lemma 6.3. The lemma then gives positive control of the running graph-energy maximum and of the d ν − 1 d^{\nu-1} d ν − 1 -increment sum when c A B c_{\mathrm{AB}} c AB is sufficiently small. The right-hand side is bounded by (244) and Young’s inequality; the start-up factor in the dual denominator is absorbed by I τ ( V h ) I_{\tau}(V_h) I τ ( V h ) . All increment and boundary forcing defects are bounded by
C τ 2 ∑ n = ν N ∥ R n ∥ Ω 2 = C τ ∥ R ∥ ℓ τ 2 ( ν , N ; H ) 2 (246) C\tau^{2}\sum_{n=\nu}^{N}\lVert R^{n}\rVert_{\Omega}^{2}
=C\tau\lVert R\rVert_{\ell_{\tau}^{2}(\nu,N;\mathcal{H})}^{2} \tag*{(246)} C τ 2 n = ν ∑ N ∥ R n ∥ Ω 2 = C τ ∥ R ∥ ℓ τ 2 ( ν , N ; H ) 2 (246)
Repeating the argument at the final horizon controls the complete material-residual sum and proves (245) . ■
For the original inhomogeneous A B ν AB\nu A B ν scheme, take R n = M A B ν F n R^n = \mathcal{M}_{AB\nu}F^n R n = M A B ν F n in (6.48).
Corollary 6.5 (Adams–Bashforth error estimate). Assume Theorem 6.4, the regularity hypotheses of Proposition 3.2, and projected start-up data. Then (3.35)–(3.40) holds with ν = 3 \nu= 3 ν = 3 for A B 3 AB3 A B 3 and ν = 4 \nu= 4 ν = 4 for A B 4 AB4 A B 4 .
Proof. Apply Theorem 6.4 to the error equation with general forcing R = A τ η − ρ R = A_{\tau}\eta- \rho R = A τ η − ρ . Use (6.8)–(6.11). Proposition 3.2 applies to (6.47), whose denominator contains the graph maximum and material-residual sum required by the proposition. Theorem 3.3 then gives the asserted estimate. ■
Under the balanced choice δ ∼ h \delta\sim h δ ∼ h and τ ≲ h \tau\lesssim h τ ≲ h , this yields the concrete rate O ( h k + 1 / 2 + τ ν ) O(h^{k+1/2} + \tau^\nu) O ( h k + 1/2 + τ ν ) , with ν = 3 \nu= 3 ν = 3 for A B 3 AB3 A B 3 and ν = 4 \nu= 4 ν = 4 for A B 4 AB4 A B 4 .
Remark 6.6 (The AB2 scheme). The second-order Adams–Bashforth scheme (AB2) can be analysed by arguments similar to those used above. For finite element spaces of arbitrary polynomial degree, they yield stability under the 4 / 3 4/3 4/3 -CFL condition τ ≲ h 4 / 3 \tau\lesssim h^{4/3} τ ≲ h 4/3 . For continuous piecewise affine approximation, the elementwise cancellation used in Corollary 5.7 instead yields stability under the standard hyperbolic CFL condition τ ≲ h \tau\lesssim h τ ≲ h . In the latter case, the second-order discretisation errors in space and time are balanced. For an analysis of AB2 with continuous interior penalty stabilisation, see [11 ] . □
Numerical Experiments
The experiments address three questions. First, do the time integrators attain their formal orders when spatial error is removed? Second, are the predicted fully discrete rates observed when temporal and spatial orders are matched? Third, does stabilisation confine errors generated by discontinuities and compactly supported waves? We use the method coefficients and parameter conventions of Sections 4, 5, and 6. Rates computed from successive refinements are empirical; they neither determine universal stability thresholds nor establish a general localisation theorem.
Implementation and Test Problems
We use the acoustic system (2.4)–(2.7) with unit wave speed. Thus the scaled and physical velocities agree, and U = ( p , u ) U = (p,u) U = ( p , u ) . The one-dimensional domain is ( 0 , 1 ) (0,1) ( 0 , 1 ) , and the two-dimensional domain is ( 0 , 1 ) 2 (0,1)^2 ( 0 , 1 ) 2 , in both cases with periodic boundary conditions. All components are approximated in the same continuous finite element space: periodic P 1 \mathrm{P}_1 P 1 through P 4 \mathrm{P}_4 P 4 spaces in one dimension and periodic triangular P 1 \mathrm{P}_1 P 1 spaces in two dimensions.
The assembly uses consistent mass matrices and Gaussian quadrature of sufficient order. Exact finite element histories are used in the purely temporal test, while the smooth PDE tests use L 2 L^2 L 2 -projected exact histories. The discontinuous one-dimensional data are also L 2 L^2 L 2 -projected, with the jumps aligned to element interfaces on every mesh. The compact initial pressure in the two-dimensional localisation test is instead interpolated at the nodes so that its nodal support is preserved. For each fixed run, one direct sparse factorisation is reused at every time step. There is no mass lumping, filtering, limiter, added viscosity, or retuning along a refinement path.
For every production run, a nominal step τ ∗ \tau_\ast τ ∗ is prescribed and the actual step is chosen as τ = T / N ≤ τ ∗ \tau= T/N \leq\tau_\ast τ = T / N ≤ τ ∗ , with integer N N N , so that the final time is an exact grid level.
Let e n = U h n − U ( t n ) e^n=U_h^n-U(t^n) e n = U h n − U ( t n ) . The three error quantities reported below are, respectively, the final L 2 L^2 L 2 error, the maximum graph error, and the accumulated material-residual error,
E L 2 = ∥ e N ∥ Ω (247) E_{L^2}=\lVert e^N\rVert_{\Omega} \tag*{(247)} E L 2 = ∥ e N ∥ Ω (247)
E G = max 0 ≤ n ≤ N ( ∥ e n ∥ Ω 2 + δ 2 ∥ G e n ∥ Ω 2 ) 1 / 2 (248) E_G=\max_{0\le n\le N}\left(\lVert e^n\rVert_{\Omega}^2+\delta^2\lVert G e^n\rVert_{\Omega}^2\right)^{1/2} \tag*{(248)} E G = 0 ≤ n ≤ N max ( ∥ e n ∥ Ω 2 + δ 2 ∥ G e n ∥ Ω 2 ) 1/2 (248)
E R = ( τ δ ∑ n = n 0 N ∥ D τ e n + G M τ e n ∥ Ω 2 ) 1 / 2 (249) E_R=\left(\tau\delta\sum_{n=n_0}^{N}\lVert D_\tau e^n+G\mathcal{M}_\tau e^n\rVert_{\Omega}^2\right)^{1/2} \tag*{(249)} E R = ( τ δ n = n 0 ∑ N ∥ D τ e n + G M τ e n ∥ Ω 2 ) 1/2 (249)
Here n 0 n_0 n 0 is the first computed time level and M τ \mathcal{M}_\tau M τ is the average associated with the active method. These definitions use the same graph and material-residual scaling as the energy analysis. Matrix assembly reproduced both the discrete skew-symmetric structure and, when δ = b 0 τ \delta=b_0\tau δ = b 0 τ , the normal-equation cancellation to floating-point precision. Pilot calculations were used only to select conservative production constants. Unstable pilot points were kept in the reproducibility record, and the scans are not interpreted as sharp or mesh-independent CFL bounds.
Temporal Accuracy
To remove spatial approximation error, we work on a periodic P 1 \mathbb{P}_1 P 1 mesh with three elements and prescribe the nontrivial finite element vector W h W_h W h by its nodal values
W h ( x i ) = ( sin ( 2 π x i ) + 0.3 cos ( 2 π x i ) 0.7 cos ( 2 π x i ) − 0.2 sin ( 2 π x i ) ) (250) W_h(x_i)=
\begin{pmatrix}
\sin(2\pi x_i)+0.3\cos(2\pi x_i)\\
0.7\cos(2\pi x_i)-0.2\sin(2\pi x_i)
\end{pmatrix} \tag*{(250)} W h ( x i ) = ( sin ( 2 π x i ) + 0.3 cos ( 2 π x i ) 0.7 cos ( 2 π x i ) − 0.2 sin ( 2 π x i ) ) (250)
The discrete manufactured data are
U h ( t ) = exp ( t ) W h (251) U_h(t)=\exp(t)W_h \tag*{(251)} U h ( t ) = exp ( t ) W h (251)
F h ( t ) = ∂ t U h ( t ) + G U h ( t ) (252) F_h(t)=\partial_t U_h(t)+G U_h(t) \tag*{(252)} F h ( t ) = ∂ t U h ( t ) + G U h ( t ) (252)
up to T = 1 T=1 T = 1 , with N = 10 , 20 , 40 , 80 , 160 N=10,20,40,80,160 N = 10 , 20 , 40 , 80 , 160 , and 320 320 320 time steps. For the θ \theta θ -methods, δ = θ τ \delta=\theta\tau δ = θ τ ; for all Adams methods, δ = h \delta=h δ = h . Table 1 reports the mean of the three finest pairwise rates unaffected by roundoff. Every method attains its formal order. For all methods except AM5, these are the final three rates. AM5 reaches a floating-point error floor on the last refinement, so that level remains in the reproducibility record while its rate is excluded from the mean.
Table 1. Formal temporal orders and mean observed rates over the three finest time-step refinements unaffected by roundoff.
Order-Matched One-Dimensional Convergence
The fully discrete convergence test is the homogeneous periodic right-moving wave
p ( x , t ) = u ( x , t ) = sin ( 2 π ( x − t ) ) , T = 1 (253) p(x,t)=u(x,t)=\sin\left(2\pi(x-t)\right),\qquad T=1 \tag*{(253)} p ( x , t ) = u ( x , t ) = sin ( 2 π ( x − t ) ) , T = 1 (253)
Under a hyperbolic scaling τ ∼ h \tau\sim h τ ∼ h , a method of temporal order ν \nu ν has temporal error O ( h ν ) O(h^\nu) O ( h ν ) . We therefore pair it with continuous P ν − 1 \mathbb{P}_{\nu-1} P ν − 1 elements, whose formal L 2 L^2 L 2 -approximation order is also ν \nu ν , so that spatial error does not mask the benefit of high-order time integration. The four method–space pairs are Crank–Nicolson/P 1 \mathbb{P}_1 P 1 , AM3/P 2 \mathbb{P}_2 P 2 , AM4/P 3 \mathbb{P}_3 P 3 , and AM5/P 4 \mathbb{P}_4 P 4 .
The meshes have 10, 20, 40, 80, and 160 elements. Every run uses the nominal parameters
τ ∗ = 0.1 h , δ = b 0 τ (254) \tau_* = 0.1h,\qquad\delta= b_0\tau \tag*{(254)} τ ∗ = 0.1 h , δ = b 0 τ (254)
Consequently, all four methods use the normal-equation formulation and a Cholesky factorisation. The start-up histories are L 2 L^2 L 2 -projections of the exact solution. For AM3/P 2 \mathbb{P}_2 P 2 and AM4/P 3 \mathbb{P}_3 P 3 , the standard hyperbolic scaling in (254) is less restrictive than the sufficient general-degree CFL conditions in Section 5. Their stable behaviour here is therefore numerical evidence, not an extension of the theorem.
Table 2 shows the rates on the finest mesh pair. The final L 2 L^2 L 2 and maximum graph errors recover orders 2, 3, 4, and 5, respectively. The accumulated material-residual rates recover the corresponding half-order sequence 3/2, 5/2, 7/2, and 9/2. This experiment therefore exhibits the fully discrete high-order behaviour in a PDE problem, rather than removing the spatial error as in the manufactured temporal test. Fixed-P 1 \mathbb{P}_1 P 1 comparisons were retained in the reproducibility record as spatial controls, but are not used to assess the advantage of the higher-order time methods.
Table 2. Observed finest-pair rates, from 80 to 160 elements, for the order-matched one-dimensional travelling-wave experiment.
Two-Dimensional Convergence
This test keeps the spatial degree fixed at P 1 \mathbb{P}_1 P 1 . It is therefore not designed to distinguish the formal orders of the time integrators; its purpose is to test the graph and material-residual rates in two dimensions and to compare implicit and explicit schemes on the same spatial family. The exact solution is the oblique periodic plane wave
κ = ( 1 , 1 ) (255) \boldsymbol{\kappa} = (1,1) \tag*{(255)} κ = ( 1 , 1 ) (255)
p ( x , y , t ) = sin ( 2 π ( x + y − 2 t ) ) (256) p(x,y,t) = \sin\left(2\pi(x+y-\sqrt{2}t)\right) \tag*{(256)} p ( x , y , t ) = sin ( 2 π ( x + y − 2 t ) ) (256)
u ( x , y , t ) = 2 − 1 / 2 ( 1 , 1 ) p ( x , y , t ) , T = 2 − 1 / 2 (257) u(x,y,t) = 2^{-1/2}(1,1)p(x,y,t),\qquad T = 2^{-1/2} \tag*{(257)} u ( x , y , t ) = 2 − 1/2 ( 1 , 1 ) p ( x , y , t ) , T = 2 − 1/2 (257)
which exercises both velocity components. Crank–Nicolson, AM3, AM5, and AB4 are run on slash-diagonal triangular meshes with 8, 16, 32, and 64 subdivisions in each coordinate direction. The nominal refinement paths are
The finest-pair rates in Table 3 and the curves in Figure 1 show graph rates near two and material-residual rates near 3 / 2 3/2 3/2 . In the figure, the AM3, AM5, and AB4 curves are summarised by their pointwise median and range because the individual curves are indistinguishable at the plotted scale. This reflects the dominance of the P 1 \mathbb{P}_1 P 1 spatial error once the temporal error is small. These are the principal comparisons with the error analysis. The multistep final L 2 L^2 L 2 rates near 2.87 are an observed structured-mesh superconvergence effect and are not claimed as a general P 1 \mathbb{P}_1 P 1 rate.
Table 3.
Table 3. Observed finest-pair rates for the two-dimensional oblique plane wave on the slash-diagonal production meshes.
Figure 1. Observed maximum graph errors and accumulated material-residual errors for the two-dimensional oblique plane wave. The orange curve is the pointwise median of the AM3, AM5, and AB4 errors, and the shaded band spans their full range; Table 3 gives the individual finest-pair rates. Reference slopes indicate h 2 h^2 h 2 and h 3 / 2 h^{3/2} h 3/2 , respectively.
The mesh-orientation check used slash, backslash, and alternating diagonals. Across these orientations, the CN graph rates range from 2.015 to 2.095, and its material-residual rates range from 1.502 to 1.506. The AM5 graph rates range from 2.056 to 2.168, and its material-residual rates range from 1.493 to 1.498. Error levels are orientation-sensitive: the largest ratios between orientations are approximately 2.44 for the graph error and 3.48 for the AM5 final L 2 L^2 L 2 error. The graph and material-residual rate trends nevertheless remain robust across the tested orientations.
One-Dimensional Localisation of Discontinuities
The preceding smooth tests do not address how discretisation error generated near a discontinuity propagates through the domain. We therefore ask whether optimal P 1 \mathbb{P}_1 P 1 convergence is recovered in regions that remain a fixed physical distance from the moving jumps. We prescribe zero initial velocity and the periodic pressure profile
f ( x ) = { 1 + 1 4 sin ( 2 π ( x − x L ) / ( x R − x L ) ) , x L ≤ x < x R , 0 , otherwise x L = 0.25 , x R = 0.45 (258) f(x)=
\begin{cases}
1+\frac{1}{4}\sin\left(2\pi(x-x_L)/(x_R-x_L)\right), & x_L\le x<x_R,\\
0, & \text{otherwise}
\end{cases}
\qquad x_L=0.25,\quad x_R=0.45 \tag*{(258)} f ( x ) = { 1 + 4 1 sin ( 2 π ( x − x L ) / ( x R − x L ) ) , 0 , x L ≤ x < x R , otherwise x L = 0.25 , x R = 0.45 (258)
The profile is smooth and nonconstant in the interior of its support but has a unit jump at each endpoint. For the one-dimensional acoustic system, d’Alembert’s formula gives the exact solution,
p ( x , t ) = 1 2 ( f ( x − t ) + f ( x + t ) ) (259) p(x,t)=\frac{1}{2}\left(f(x-t)+f(x+t)\right) \tag*{(259)} p ( x , t ) = 2 1 ( f ( x − t ) + f ( x + t ) ) (259)
u ( x , t ) = 1 2 ( f ( x − t ) − f ( x + t ) ) (260) u(x,t)=\frac{1}{2}\left(f(x-t)-f(x+t)\right) \tag*{(260)} u ( x , t ) = 2 1 ( f ( x − t ) − f ( x + t ) ) (260)
Thus the initial pulse splits into left- and right-moving waves with a known set of moving discontinuities,
J ( t ) = { x L + t , x R + t , x L − t , x R − t } m o d 1 (261) \mathcal{J}(t)=\{x_L+t,x_R+t,x_L-t,x_R-t\}\bmod1 \tag*{(261)} J ( t ) = { x L + t , x R + t , x L − t , x R − t } mod 1 (261)
We use periodic P 1 \mathbb{P}_1 P 1 elements on meshes with 40 through 2560 elements and Crank–Nicolson up to T = 0.15 T=0.15 T = 0.15 , with nominal step τ ∗ = 0.2 h \tau_\ast=0.2h τ ∗ = 0.2 h . The two calculations differ only in the stabilisation parameter: δ = 0 \delta=0 δ = 0 for the standard Galerkin method and δ = τ / 2 \delta=\tau/2 δ = τ /2 for the normal-equation SUPG method. The endpoints x L x_L x L and x R x_R x R are element interfaces on every mesh, and the initial pressure is the L 2 L^2 L 2 -projection of (7.12).
For the fixed physical buffer width ρ b u f = 0.04 \rho_{\mathrm{buf}}=0.04 ρ buf = 0.04 , define the smooth region by
Ω s m ( t ) = { x ∈ ( 0 , 1 ) : dist p e r ( x , J ( t ) ) > ρ b u f } (262) \Omega_{\mathrm{sm}}(t)=\left\{x\in(0,1):\operatorname{dist}_{\mathrm{per}}(x,\mathcal{J}(t))>\rho_{\mathrm{buf}}\right\} \tag*{(262)} Ω sm ( t ) = { x ∈ ( 0 , 1 ) : dist per ( x , J ( t )) > ρ buf } (262)
and measure the maximum local error,
E l o c = max 0 ≤ n ≤ N ∥ e n ∥ Ω s m ( t n ) (263) E_{\mathrm{loc}}=\max_{0\le n\le N}\left\|e^n\right\|_{\Omega_{\mathrm{sm}}(t^n)} \tag*{(263)} E loc = 0 ≤ n ≤ N max ∥ e n ∥ Ω sm ( t n ) (263)
The same moving region is used at every refinement level, rather than a mesh-dependent exclusion zone.
Table 4 shows the finest-pair results. The unstabilised local error converges at rate 0.4995, whereas normal-equation SUPG recovers the optimal P 1 \mathbb{P}_1 P 1 local rate 2.0006. On the finest mesh their maximum local errors are 5.6971 × 10 − 3 5.6971\times10^{-3} 5.6971 × 1 0 − 3 and 3.9288 × 10 − 7 3.9288\times10^{-7} 3.9288 × 1 0 − 7 , respectively, a factor of approximately 1.45 × 10 4 1.45\times10^4 1.45 × 1 0 4 . In the part of the smooth region where neither exact travelling pulse is present, the corresponding maximum-in-time L 2 L^2 L 2 norms of the numerical state are 2.9652 × 10 − 3 2.9652\times10^{-3} 2.9652 × 1 0 − 3 and 4.5629 × 10 − 15 4.5629\times10^{-15} 4.5629 × 1 0 − 15 . The full-domain errors converge much more slowly because they include the moving discontinuities. Their observed rates are reported without interpreting them as universal asymptotic exponents.
Table 4. Errors on the 2560-element mesh and rates from 1280 to 2560 elements for the discontinuous d’Alembert experiment. The local error is defined by (7.17).
The convergence curves in Figure 2 show that the stabilised calculation passes through a preasymptotic layer-decay regime before settling onto the h 2 h^2 h 2 reference line. The standard Galerkin local error instead follows the h 1 / 2 h^{1/2} h 1/2 line on the finest levels. The profiles in Figure 3 explain the contrast: the unstabilised dispersive error extends across the periodic interval, whereas the SUPG error remains concentrated around the exact wave fronts. This is numerical evidence for localisation in this linear one-dimensional configuration, not a general theorem for nonsmooth solutions.
Figure 2. Final full-domain errors and maximum smooth-region errors for the discontinuous d’Alembert experiment. The fixed smooth region is given by (262) .
Figure 3. Final pressure and pointwise state error on the 160-element mesh. In the right panel, the pointwise state errors are plotted against the logarithmic left-hand axis, while the exact state magnitude is shown in black against the right-hand axis. The unstabilised error is dispersive and global, while normal-equation SUPG confines the significant error to neighbourhoods of the exact jumps.
Two-Dimensional Localisation of a Compact Wave
Finally, we ask whether stabilisation also suppresses numerical energy outside the physical wave cone for smooth compact data in two dimensions. Let x 0 = ( 0.5 , 0.5 ) x_0 = (0.5, 0.5) x 0 = ( 0.5 , 0.5 ) , R 0 = 0.12 R_0 = 0.12 R 0 = 0.12 , and r p e r ( x ) = dist p e r ( x , x 0 ) r_{\mathrm{per}}(x) = \operatorname{dist}_{\mathrm{per}}(x,x_0) r per ( x ) = dist per ( x , x 0 ) . The initial data are
p ( x , 0 ) = { exp ( 1 − 1 1 − ( r p e r ( x ) / R 0 ) 2 ) , r p e r ( x ) < R 0 , 0 , r p e r ( x ) ≥ R 0 , u ( x , 0 ) = 0 (264) p(x,0) =
\begin{cases}
\exp\left(1-\dfrac{1}{1-(r_{\mathrm{per}}(x)/R_0)^2}\right), & r_{\mathrm{per}}(x) < R_0,\\
0, & r_{\mathrm{per}}(x) \ge R_0,
\end{cases}
\qquad u(x,0)=0 \tag*{(264)} p ( x , 0 ) = ⎩ ⎨ ⎧ exp ( 1 − 1 − ( r per ( x ) / R 0 ) 2 1 ) , 0 , r per ( x ) < R 0 , r per ( x ) ≥ R 0 , u ( x , 0 ) = 0 (264)
Thus the pressure is smooth, compactly supported, and equal to one at its centre. We use 64 subdivisions in each coordinate direction and Crank–Nicolson up to T = 0.25 T = 0.25 T = 0.25 , with nominal step τ ∗ = 0.1 h \tau_* = 0.1h τ ∗ = 0.1 h . The two runs use the same time grid and differ only in the stabilisation parameter: δ = 0 \delta= 0 δ = 0 and the normal-equation value δ = τ / 2 \delta= \tau/2 δ = τ /2 . With h = 2 / 64 h = \sqrt{2}/64 h = 2 /64 , define the exterior region by
Ω e x t ( t ) = { x ∈ Ω : dist p e r ( x , x 0 ) > R 0 + t + 2 h } (265) \Omega_{\mathrm{ext}}(t) = \left\{x \in\Omega: \operatorname{dist}_{\mathrm{per}}(x,x_0) > R_0 + t + 2h\right\} \tag*{(265)} Ω ext ( t ) = { x ∈ Ω : dist per ( x , x 0 ) > R 0 + t + 2 h } (265)
so the buffer outside the exact wave cone has width 2 h 2h 2 h . At final time the cone-plus-buffer radius is 0.414194 < 0.5 0.414194 < 0.5 0.414194 < 0.5 , and hence it has not reached a periodic image.
For this experiment, let
E ( t ) = 1 2 ∫ Ω ( p h 2 + ∣ u h ∣ 2 ) d x (266) E(t) = \frac{1}{2}\int_{\Omega}\left(p_h^2 + \lvert u_h\rvert^2\right)\,dx \tag*{(266)} E ( t ) = 2 1 ∫ Ω ( p h 2 + ∣ u h ∣ 2 ) d x (266)
E e x t ( t ) = 1 2 ∫ Ω e x t ( t ) ( p h 2 + ∣ u h ∣ 2 ) d x (267) E_{\mathrm{ext}}(t) = \frac{1}{2}\int_{\Omega_{\mathrm{ext}}(t)}\left(p_h^2 + \lvert u_h\rvert^2\right)\,dx \tag*{(267)} E ext ( t ) = 2 1 ∫ Ω ext ( t ) ( p h 2 + ∣ u h ∣ 2 ) d x (267)
Figure 4 and Figure 5 show the pressure fields on a common colour scale and the corresponding energy histories. The unstabilised final ratio E ( T ) / E ( 0 ) E(T)/E(0) E ( T ) / E ( 0 ) is 1.0000, whereas the normal-equation SUPG ratio is 0.993225. Their final exterior-energy fractions E e x t ( T ) / E ( 0 ) E_{\mathrm{ext}}(T)/E(0) E ext ( T ) / E ( 0 ) are, respectively, 4.7772 × 10 − 6 4.7772 \times10^{-6} 4.7772 × 1 0 − 6 and 1.2126 × 10 − 7 1.2126 \times10^{-7} 1.2126 × 1 0 − 7 , so the stabilised fraction is 39.4 times smaller. At T T T , the full-state L 2 L^2 L 2 difference between the two solutions is 3.1115 × 10 − 3 3.1115 \times10^{-3} 3.1115 × 1 0 − 3 . Thus normal-equation stabilisation improves localisation in this configuration, with modest dissipation; no general multidimensional result is claimed.
Figure 4. Crank–Nicolson pressure on the periodic 64 × 64 64 \times64 64 × 64 triangular grid at t = 0 t = 0 t = 0 , t = 0.125 t = 0.125 t = 0.125 , and t = 0.25 t = 0.25 t = 0.25 . The top row uses δ = 0 \delta= 0 δ = 0 , the bottom row uses the normal-equation value δ = τ / 2 \delta= \tau/2 δ = τ /2 , and all panels use the same colour scale.
Figure 5. Crank–Nicolson total-energy ratio and exterior-energy fraction for δ = 0 \delta= 0 δ = 0 and δ = τ / 2 \delta= \tau/2 δ = τ /2 on the periodic 64 × 64 64 \times64 64 × 64 triangular grid over 0 ≤ t ≤ 0.25 0 \leq t \leq0.25 0 ≤ t ≤ 0.25 . Both quantities are normalised by the initial total energy.
Conclusions
The common operator and energy framework developed here leads to five principal conclusions.
A common formulation clarifies the algebra. Separating the method average, the backward time difference, the material residual, and the perturbed test operator places all the time integrators considered here within the same structure. For implicit methods, the distinguished choice δ = b 0 τ \delta= b_0\tau δ = b 0 τ turns each step into a symmetric positive definite normal equation associated with residual minimisation in the graph norm. A general SUPG choice remains admissible, but the resulting system is not symmetric in general. For the explicit Adams–Bashforth methods, b 0 = 0 b_0 = 0 b 0 = 0 , so no positive δ \delta δ can yield this normal-equation structure.
Stability depends on the method and the discrete space. The framework establishes stability for the θ \theta θ -method and AM3–AM5; separate exact energy decompositions yield stability for AB3 and AB4. For general polynomial degree, AM3 and AM4 require the strengthened time-step restrictions in Section 5. For continuous piecewise affine elements, cancellation due to the elementwise constancy of the gradient restores the standard hyperbolic regime τ ≲ δ ≲ h \tau\lesssim\delta\lesssim h τ ≲ δ ≲ h . AM5, AB3, and AB4 are stable under standard hyperbolic CFL restrictions. Since the general-degree Adams–Moulton analysis imposes no lower bound on δ \delta δ , it includes the normal-equation choice whenever the relevant time-step restriction is satisfied.
The dual residual yields a sharp fully discrete estimate. A direct dual-residual argument gives the bound O ( h k + 1 / 2 + τ ν ) O(h^{k+1/2}+\tau^\nu) O ( h k + 1/2 + τ ν ) for δ ∼ h \delta\sim h δ ∼ h and τ ≲ h \tau\lesssim h τ ≲ h , subject to the stated stability and regularity assumptions. It uses the ordinary L 2 L^2 L 2 -projection, skew-symmetry, and discrete summation by parts.
Order matching reveals the high-order regime. The standalone temporal tests confirm the formal order of each scheme. In the principal one-dimensional acoustic experiment, pairing a method of temporal order ν \nu ν with P ν − 1 \mathbb{P}_{\nu-1} P ν − 1 elements, τ = 0.1 h \tau= 0.1h τ = 0.1 h , and δ = b 0 τ \delta= b_0\tau δ = b 0 τ produces convergence of orders two through five in the final-time and graph errors. The material-residual errors follow the half-order sequence 3 / 2 3/2 3/2 , 5 / 2 5/2 5/2 , 7 / 2 7/2 7/2 , and 9 / 2 9/2 9/2 . The two-dimensional P 1 \mathbb{P}_1 P 1 tests give graph and material-residual rates close to two and 3 / 2 3/2 3/2 , respectively.
Stabilisation substantially improves localisation. In the discontinuous d’Alembert experiment, the unstabilised P 1 \mathbb{P}_1 P 1 solution converges locally at rate 0.4995, whereas normal-equation SUPG recovers the optimal local rate 2.0006 away from the moving jumps. On the finest mesh, the maximum local error of the stabilised method is smaller by a factor of approximately 1.45 × 10 4 1.45 \times10^4 1.45 × 1 0 4 . In the smooth two-dimensional compact-wave test, the stabilised solution has a final exterior-energy fraction that is 39.4 times smaller, with only modest dissipation. Together, these experiments provide evidence for the proposed localisation mechanism, but not a general theorem for nonsmooth solutions.
Natural theoretical extensions include variable coefficients and the associated commutator terms, nonperiodic boundary conditions, and quantitative local error estimates for nonsmooth solutions in multidimensional and nonlinear systems. Computational priorities include robust multistep start-up procedures and efficient solvers and preconditioners for both the symmetric normal-equation systems and their generally nonsymmetric SUPG counterparts.
Proof of the Dual Residual Estimate
This appendix gives the complete proof of Proposition 3.2. The argument is based on the ordinary L 2 L^2 L 2 -projection, skew-symmetry, and discrete summation by parts.
Fix a terminal index m m m , with n 0 ≤ m ≤ N n_0 \le m \le N n 0 ≤ m ≤ N , and extend the test sequence by zero outside 0 ≤ n ≤ m 0 \le n \le m 0 ≤ n ≤ m . Set
D m ( V ) = ( I τ ( V ) + ∥ V ∥ E , m 2 ) 1 / 2 (268) \mathcal{D}_m(V) = \left(\mathcal{I}_\tau(V) + \lVert V \rVert_{\mathcal{E},m}^{2}\right)^{1/2} \tag*{(268)} D m ( V ) = ( I τ ( V ) + ∥ V ∥ E , m 2 ) 1/2 (268)
The coercivity estimates (3.8) and (3.9), together with (2.29), imply
max 0 ≤ j ≤ m ( ∥ V j ∥ Ω + δ ∥ G V j ∥ Ω ) ≲ D m ( V ) (269) \max_{0 \le j \le m} \left(\lVert V^j \rVert_{\Omega} + \delta\lVert G V^j \rVert_{\Omega}\right) \lesssim\mathcal{D}_m(V) \tag*{(269)} 0 ≤ j ≤ m max ( ∥ V j ∥ Ω + δ ∥ G V j ∥ Ω ) ≲ D m ( V ) (269)
∥ δ 1 / 2 A τ V ∥ ℓ 2 ( n 0 , m ; H ) ≲ D m ( V ) (270) \lVert\delta^{1/2} A_\tau V \rVert_{\ell^2(n_0,m;\mathcal{H})} \lesssim\mathcal{D}_m(V) \tag*{(270)} ∥ δ 1/2 A τ V ∥ ℓ 2 ( n 0 , m ; H ) ≲ D m ( V ) (270)
For 0 ≤ j < n 0 0 \le j < n_0 0 ≤ j < n 0 , the first bound follows from the start-up energy I τ ( V ) \mathcal{I}_\tau(V) I τ ( V ) . Since the time stencil is fixed, we also have
max n 0 ≤ n ≤ m ∥ M τ V n ∥ Ω ≲ D m ( V ) (271) \max_{n_0 \le n \le m} \lVert\mathcal{M}_\tau V^n \rVert_{\Omega} \lesssim\mathcal{D}_m(V) \tag*{(271)} n 0 ≤ n ≤ m max ∥ M τ V n ∥ Ω ≲ D m ( V ) (271)
The spatial approximation estimate, combined with the finite-stencil trace argument leading to (3.22), gives
∥ M τ η ∥ ℓ 2 ( n 0 , m ; H ) ≲ h r N r , 2 ( U ) (272) \lVert\mathcal{M}_\tau\eta\rVert_{\ell^2(n_0,m;\mathcal{H})} \lesssim h^r \mathcal{N}_{r,2}(U) \tag*{(272)} ∥ M τ η ∥ ℓ 2 ( n 0 , m ; H ) ≲ h r N r , 2 ( U ) (272)
∥ A τ η ∥ ℓ 2 ( n 0 , m ; H ) ≲ h r − 1 N r , 2 ( U ) (273) \lVert A_\tau\eta\rVert_{\ell^2(n_0,m;\mathcal{H})} \lesssim h^{r-1} \mathcal{N}_{r,2}(U) \tag*{(273)} ∥ A τ η ∥ ℓ 2 ( n 0 , m ; H ) ≲ h r − 1 N r , 2 ( U ) (273)
max n 0 ≤ n ≤ m ∥ A τ η n ∥ Ω + ∥ D τ + A τ η ∥ ℓ 2 ( n 0 , m − 1 ; H ) ≲ h r − 1 N r , 2 ( U ) (274) \max_{n_0 \le n \le m} \lVert A_\tau\eta^n \rVert_{\Omega} + \lVert D_\tau^+ A_\tau\eta\rVert_{\ell^2(n_0,m-1;\mathcal{H})} \lesssim h^{r-1} \mathcal{N}_{r,2}(U) \tag*{(274)} n 0 ≤ n ≤ m max ∥ A τ η n ∥ Ω + ∥ D τ + A τ η ∥ ℓ 2 ( n 0 , m − 1 ; H ) ≲ h r − 1 N r , 2 ( U ) (274)
Indeed, D τ − η n D_\tau^- \eta^n D τ − η n is an interval average of ( Id − Π h ) ∂ t U (\operatorname{Id}-\Pi_h)\partial_t U ( Id − Π h ) ∂ t U , whereas D τ + D τ − η D_\tau^+D_\tau^- \eta D τ + D τ − η is a difference of two such averages. The remaining contributions are fixed-stencil averages of G η G\eta G η and G D τ + η G D_\tau^+\eta G D τ + η . Hence (A.5)–(A.7) follow from (2.16), applied to U U U and its first two time derivatives.
Projection Contribution
Since Π h \Pi_h Π h is the spatial L 2 L^2 L 2 -projection, it commutes with the time operators and
( D τ − η n , M τ V n ) Ω = 0 (275) (D_\tau^- \eta^n,\mathcal{M}_\tau V^n)_\Omega= 0 \tag*{(275)} ( D τ − η n , M τ V n ) Ω = 0 (275)
( M τ η n , D τ − V n ) Ω = 0 (276) (\mathcal{M}_\tau\eta^n,D_\tau^- V^n)_\Omega= 0 \tag*{(276)} ( M τ η n , D τ − V n ) Ω = 0 (276)
Skew-symmetry of G G G therefore gives
( A τ η n , M τ V n ) Ω = − ( M τ η n , A τ V n ) Ω (277) (A_\tau\eta^n,\mathcal{M}_\tau V^n)_\Omega= -(\mathcal{M}_\tau\eta^n,A_\tau V^n)_\Omega \tag*{(277)} ( A τ η n , M τ V n ) Ω = − ( M τ η n , A τ V n ) Ω (277)
Combining (A.10) with P δ = M τ + δ A τ + δ 2 G D τ − P_\delta= \mathcal{M}_\tau+ \delta A_\tau+ \delta^2 G D_\tau^- P δ = M τ + δ A τ + δ 2 G D τ − yields (3.34). Notice that this step does not move M τ \mathcal{M}_\tau M τ between the two arguments.
For the projection contribution, define
F η , m ( V ) = τ ∑ n = n 0 m ( A τ η n , P δ V n ) Ω (278) \mathcal{F}_{\eta,m}(V) = \tau\sum_{n=n_0}^{m} (A_\tau\eta^n,P_\delta V^n)_\Omega \tag*{(278)} F η , m ( V ) = τ n = n 0 ∑ m ( A τ η n , P δ V n ) Ω (278)
Since G D τ − V = D τ − G V G D_\tau^- V = D_\tau^- G V G D τ − V = D τ − G V , the exact summation-by-parts formula yields
F η , m ( V ) = τ ∑ n = n 0 m ( δ A τ η n − M τ η n , A τ V n ) Ω + δ 2 ( A τ η m , G V m ) Ω − δ 2 ( A τ η n 0 , G V n 0 − 1 ) Ω − τ δ 2 ∑ n = n 0 m − 1 ( D τ + A τ η n , G V n ) Ω \begin{aligned}
\mathcal{F}_{\eta,m}(V) ={}& \tau\sum_{n=n_0}^{m} (\delta A_\tau\eta^n-\mathcal{M}_\tau\eta^n,A_\tau V^n)_\Omega\\
&+ \delta^2(A_\tau\eta^m,G V^m)_\Omega-\delta^2(A_\tau\eta^{n_0},G V^{n_0-1})_\Omega\\
&-\tau\delta^2\sum_{n=n_0}^{m-1}(D_\tau^+A_\tau\eta^n,G V^n)_\Omega
\end{aligned} F η , m ( V ) = τ n = n 0 ∑ m ( δ A τ η n − M τ η n , A τ V n ) Ω + δ 2 ( A τ η m , G V m ) Ω − δ 2 ( A τ η n 0 , G V n 0 − 1 ) Ω − τ δ 2 n = n 0 ∑ m − 1 ( D τ + A τ η n , G V n ) Ω
{#eq:A.12} {#eq:A.13}
Both endpoint terms are retained. The Cauchy–Schwarz inequality and (269) –(270) then yield, with T m = m τ ≤ T T_m=m\tau\le T T m = m τ ≤ T ,
∣ F η , m ( V ) ∣ ≲ ( δ 1 / 2 ∥ A τ η ∥ ℓ 2 ( n 0 , m ; H ) + δ − 1 / 2 ∥ M τ η ∥ ℓ 2 ( n 0 , m ; H ) ) D m ( V ) (279) |\mathcal{F}_{\eta,m}(V)| \lesssim\left(\delta^{1/2}\lVert A_\tau\eta\rVert_{\ell^2(n_0,m;\mathcal{H})}+\delta^{-1/2}\lVert M_\tau\eta\rVert_{\ell^2(n_0,m;\mathcal{H})}\right)\mathcal{D}_m(V) \tag*{(279)} ∣ F η , m ( V ) ∣ ≲ ( δ 1/2 ∥ A τ η ∥ ℓ 2 ( n 0 , m ; H ) + δ − 1/2 ∥ M τ η ∥ ℓ 2 ( n 0 , m ; H ) ) D m ( V ) (279)
+ δ ( 2 max n 0 ≤ n ≤ m ∥ A τ η n ∥ Ω + T m 1 / 2 ∥ D τ + A τ η ∥ ℓ 2 ( n 0 , m − 1 ; H ) ) D m ( V ) (280) +\delta\left(2\max_{n_0\le n\le m}\lVert A_\tau\eta^n\rVert_{\Omega}+T_m^{1/2}\lVert D_\tau^+A_\tau\eta\rVert_{\ell^2(n_0,m-1;\mathcal{H})}\right)\mathcal{D}_m(V) \tag*{(280)} + δ ( 2 n 0 ≤ n ≤ m max ∥ A τ η n ∥ Ω + T m 1/2 ∥ D τ + A τ η ∥ ℓ 2 ( n 0 , m − 1 ; H ) ) D m ( V ) (280)
Substituting (272) –(274) , together with 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , gives
∣ F η , m ( V ) ∣ ≲ ( ( 1 + T 1 / 2 ) δ 1 / 2 h r − 1 + δ − 1 / 2 h r ) N r , 2 ( U ) D m ( V ) (281) |\mathcal{F}_{\eta,m}(V)|\lesssim\left((1+T^{1/2})\delta^{1/2}h^{r-1}+\delta^{-1/2}h^r\right)\mathcal{N}_{r,2}(U)\mathcal{D}_m(V) \tag*{(281)} ∣ F η , m ( V ) ∣ ≲ ( ( 1 + T 1/2 ) δ 1/2 h r − 1 + δ − 1/2 h r ) N r , 2 ( U ) D m ( V ) (281)
Truncation Contribution
For the truncation contribution, set
F ρ , m ( V ) = τ ∑ n = n 0 m ( ρ n , P δ V n ) Ω (282) \mathcal{F}_{\rho,m}(V)=\tau\sum_{n=n_0}^{m}(\rho^n,P_\delta V^n)_\Omega \tag*{(282)} F ρ , m ( V ) = τ n = n 0 ∑ m ( ρ n , P δ V n ) Ω (282)
Expanding P δ P_\delta P δ and applying the same summation-by-parts formula give
F ρ , m ( V ) = τ ∑ n = n 0 m ( ρ n , M τ V n ) Ω (283) \mathcal{F}_{\rho,m}(V)=\tau\sum_{n=n_0}^{m}(\rho^n,M_\tau V^n)_\Omega \tag*{(283)} F ρ , m ( V ) = τ n = n 0 ∑ m ( ρ n , M τ V n ) Ω (283)
F ρ , m ( V ) = τ ∑ n = n 0 m ( ρ n , M τ V n ) Ω + τ δ ∑ n = n 0 m ( ρ n , A τ V n ) Ω + δ 2 ( ρ m , G V m ) Ω − δ 2 ( ρ n 0 , G V n 0 − 1 ) Ω − τ δ 2 ∑ n = n 0 m − 1 ( D τ + ρ n , G V n ) Ω (284) \begin{aligned}
\mathcal{F}_{\rho,m}(V)={}&\tau\sum_{n=n_0}^{m}(\rho^n,M_\tau V^n)_\Omega
+\tau\delta\sum_{n=n_0}^{m}(\rho^n,A_\tau V^n)_\Omega\\
&+\delta^2(\rho^m,GV^m)_\Omega-\delta^2(\rho^{n_0},GV^{n_0-1})_\Omega\\
&-\tau\delta^2\sum_{n=n_0}^{m-1}(D_\tau^+\rho^n,GV^n)_\Omega \tag*{(284)}
\end{aligned} F ρ , m ( V ) = τ n = n 0 ∑ m ( ρ n , M τ V n ) Ω + τ δ n = n 0 ∑ m ( ρ n , A τ V n ) Ω + δ 2 ( ρ m , G V m ) Ω − δ 2 ( ρ n 0 , G V n 0 − 1 ) Ω − τ δ 2 n = n 0 ∑ m − 1 ( D τ + ρ n , G V n ) Ω (284)
Consequently, the bounds (269) –(271) and the Cauchy–Schwarz inequality give
∣ F ρ , m ( V ) ∣ ≲ ( T m 1 / 2 ∥ ρ ∥ ℓ 2 ( n 0 , m ; H ) + δ 1 / 2 ∥ ρ ∥ ℓ 2 ( n 0 , m ; H ) ) D m ( V ) (285) |\mathcal{F}_{\rho,m}(V)|\lesssim\left(T_m^{1/2}\lVert\rho\rVert_{\ell^2(n_0,m;\mathcal{H})}+\delta^{1/2}\lVert\rho\rVert_{\ell^2(n_0,m;\mathcal{H})}\right)\mathcal{D}_m(V) \tag*{(285)} ∣ F ρ , m ( V ) ∣ ≲ ( T m 1/2 ∥ ρ ∥ ℓ 2 ( n 0 , m ; H ) + δ 1/2 ∥ ρ ∥ ℓ 2 ( n 0 , m ; H ) ) D m ( V ) (285)
+ δ ( 2 max n 0 ≤ n ≤ m ∥ ρ n ∥ Ω + T m 1 / 2 ∥ D τ + ρ ∥ ℓ 2 ( n 0 , m − 1 ; H ) ) D m ( V ) (286) +\delta\left(2\max_{n_0\le n\le m}\lVert\rho^n\rVert_{\Omega}+T_m^{1/2}\lVert D_\tau^+\rho\rVert_{\ell^2(n_0,m-1;\mathcal{H})}\right)\mathcal{D}_m(V) \tag*{(286)} + δ ( 2 n 0 ≤ n ≤ m max ∥ ρ n ∥ Ω + T m 1/2 ∥ D τ + ρ ∥ ℓ 2 ( n 0 , m − 1 ; H ) ) D m ( V ) (286)
The consistency estimates (24) –(26) , together with 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , therefore imply
∣ F ρ , m ( V ) ∣ ≲ ( 1 + T 1 / 2 ) τ ν C T , ν + 2 ( U , F ) D m ( V ) (287) |\mathcal{F}_{\rho,m}(V)|\lesssim(1+T^{1/2})\tau\nu\mathcal{C}_{T,\nu+2}(U,F)\mathcal{D}_m(V) \tag*{(287)} ∣ F ρ , m ( V ) ∣ ≲ ( 1 + T 1/2 ) τ ν C T , ν + 2 ( U , F ) D m ( V ) (287)
Finally, since R = A τ η − ρ R=A_\tau\eta-\rho R = A τ η − ρ , its residual functional is F η , m − F ρ , m \mathcal{F}_{\eta,m}-\mathcal{F}_{\rho,m} F η , m − F ρ , m . Combining (281) and (287) , dividing by D m ( V ) \mathcal{D}_m(V) D m ( V ) , taking the supremum over all admissible test sequences, and then maximising over n 0 ≤ m ≤ N n_0\le m\le N n 0 ≤ m ≤ N proves (79) –(80) .
Acknowledgements. The authors worked on this project in autumn 2025 while visiting Institut Mittag-Leffler for the semester programme “Interfaces and Unfitted Discretization Methods”. We gratefully acknowledge the support and hospitality of Institut Mittag-Leffler.
EB was supported in part by EPSRC grants EP/X042650/1 and EP/V050400/1. PH was supported by the Swedish Research Council under grant 2022-03908. MGL was supported in part by the Swedish Research Council under grants 2021-04925 and 2025-05562, the Knut and Alice Wallenberg Foundation under grant KAW 2025.0277, and the Swedish Research Programme Essence.
Use of Artificial Intelligence and Computational Tools. During the preparation and revision of this manuscript, the authors used OpenAI’s ChatGPT and Codex for language editing, organisation of the presentation, LaTeX drafting, and algebraic checks of the Adams stability identities. MATLAB and Python, using NumPy, SciPy, and Matplotlib, were used for numerical computations, independent verification, and figure preparation. All AI-assisted material was reviewed and verified by the authors, who take full responsibility for the content of the manuscript.
Code Availability. The numerical code and reference data used for the experiments in this paper are publicly available at github.com/mglarson1/symtransport-numerics. The repository contains MATLAB and Python implementations, fixed experiment configurations, scripts for reproducing the reported tables and figures, reference results, and automated verification tests. The software is distributed under the BSD 3-Clause License.
Authors’ addresses:
Erik Burman, Mathematics, University College London, UK e.burman@ucl.ac.uk
Peter Hansbo, Mechanical Engineering, Jönköping University, Sweden peter.hansbo@ju.se
Mats G. Larson, Mathematics and Mathematical Statistics, Umeå University, Sweden mats.larson@umu.se