Introduction
The Landau–Lifshitz–Gilbert (LLG) equation is a fundamental model for the dynamics of the magnetisation in ferromagnetic materials; see, e.g., [5 , 17 , 19 , 20 ] and the references therein. At nonzero temperature, thermal fluctuations may be incorporated through a stochastic magnetic field, leading to the stochastic Landau–Lifshitz–Gilbert (sLLG) equation. This gives a strongly nonlinear SPDE with a non-convex pointwise constraint, in which the multiplicative noise interacts nontrivially with the geometric cross-product structure. The noise is naturally interpreted in the Stratonovich sense, and is compatible with the pointwise constraint on the magnetisation since it acts tangentially to the unit sphere.
In this work, we consider an exchange-only sLLG model on the bounded interval D = ( 0 , L ) D = (0,L) D = ( 0 , L ) . Let T > 0 T > 0 T > 0 and let ( Ω , F , { F t } t ∈ [ 0 , T ] , P ) (\Omega, \mathcal{F}, \{\mathcal{F}_t\}_{t\in[0,T]}, \mathbb{P}) ( Ω , F , { F t } t ∈ [ 0 , T ] , P ) be a filtered probability space satisfying the usual conditions. For α , κ > 0 \alpha,\kappa> 0 α , κ > 0 , the sLLG equation reads:
d m + α κ m × ( m × ∂ x x m ) d t − κ m × ∂ x x m d t = m × ∘ d W in ( 0 , T ) × D , (1) \mathrm{d}m + \alpha\kappa m \times(m \times\partial_{xx}m)\,\mathrm{d}t - \kappa m \times\partial_{xx}m\,\mathrm{d}t = m \times\circ\mathrm{d}W \quad\text{in } (0,T) \times D, \tag*{(1)} d m + α κm × ( m × ∂ xx m ) d t − κm × ∂ xx m d t = m × ∘ d W in ( 0 , T ) × D , (1)
∂ x m = 0 on ( 0 , T ) × ∂ D , (2) \partial_x m = 0 \quad\text{on } (0,T) \times\partial D, \tag*{(2)} ∂ x m = 0 on ( 0 , T ) × ∂ D , (2)
m ( 0 ) = m 0 , in D , (3) m(0) = m_0, \quad\text{in } D, \tag*{(3)} m ( 0 ) = m 0 , in D , (3)
where m : [ 0 , T ] × D × Ω → R 3 m : [0,T] \times D \times\Omega\to\mathbb{R}^3 m : [ 0 , T ] × D × Ω → R 3 is the magnetisation. Here, we assume m 0 ∈ H 3 ( D ; S 2 ) m_0 \in H^3(D;\mathbb{S}^2) m 0 ∈ H 3 ( D ; S 2 ) . In (1) , the first two terms in the drift describe damping and precession, respectively, while the Stratonovich term models thermal fluctuations. Lower-order anisotropy field can also be added to (1) without difficulties. We consider the simplified multiplicative noise m × ∘ d W m \times\circ\mathrm{d}W m × ∘ d W in the sense of
Acknowledgment. This research was initiated while the author was supported by the Commonwealth through an Australian Government Research Training Program (RTP) Scholarship (DOI: 10.82133/C42F-K220) at UNSW Sydney. The author has also been supported by an Australian Mathematical Society Lift-off Fellowship. The work was completed during the author’s postdoctoral appointment at the Institute of Analysis and Scientific Computing of TU Wien, supported by the Austrian Science Fund (FWF) through the international project I6802 “Functional error estimates for PDEs on unbounded domains” (DOI: 10.55776/I6802), led by Prof. Dirk Praetorius. Stratonovich, as in [7 , 9 , 18 ] , and assume sufficient spatial regularity of the coloured Wiener process W W W throughout the analysis. An important property of eq:1 is that the magnetisation satisfies a non-convex geometric constraint
∣ m ( t , x ) ∣ = 1 for a.e. ( t , x ) ∈ ( 0 , T ) × D , P -a.s. (4) \lvert m(t,x)\rvert= 1 \qquad\text{for a.e. }(t,x) \in(0,T) \times D,\qquad\mathbb{P}\text{-a.s.} \tag*{(4)} ∣ m ( t , x )∣ = 1 for a.e. ( t , x ) ∈ ( 0 , T ) × D , P -a.s. (4)
Our analysis also applies to the one-dimensional flat torus with periodic boundary condition. The restriction to one space dimension is motivated by both the analytical theory and applications. From the analytical viewpoint, the global pathwise well-posedness and maximal regularity required for a quantitative error analysis in the present paper are available for the sLLG equation on an interval; see, e.g., [14 , 18 ] . In higher dimensions, the analytical theory is substantially more delicate, owing to the possibility of singularity formation and the weaker regularity available for solutions. The one-dimensional setting is also a natural model for thin ferromagnetic wires, and has been considered in a number of mathematical studies of stochastic magnetisation dynamics [7 , 11 ] .
The deterministic LLG equation has been studied extensively from both analytical and numerical viewpoints; see, e.g., [1 , 4 , 12 , 15 , 17 , 20 ] and the references therein. In particular, the midpoint scheme for the deterministic problem was first analysed in [8 ] . The stochastic problem is substantially more delicate. In addition to the nonlinearity and non-convex sphere constraint, the analysis must accommodate multiplicative Stratonovich noise, limited solution regularity, and the interaction between the stochastic forcing and geometric structure of the equation [5 , 9 , 10 , 18 ] . Three types of convergent numerical methods for sLLG have been developed:
midpoint scheme introduced in [13 , 7 ] , which preserves the unit-length constraint at nodal points and converges to a weak martingale solution along a subsequence [6 ] ;
tangent plane scheme [3 ] on the Gilbert form, which extended the scheme in [2 ] to the stochastic case;
tangent plane scheme [16 ] on a random PDE which is obtained by means of the Doss–Sussmann transform.
For the one-dimensional problem, Dunst [14 ] subsequently proved convergence in probability with rate arbitrarily close to 1 / 2 1/2 1/2 for the time-semidiscrete midpoint scheme.
Extending such a rate estimate to a fully discrete method is not immediate: the finite element approximation introduces mass-lumping and interpolation defects, while the discrete Laplacian and the stochastic midpoint term require estimates which have no direct counterpart in the time-semidiscrete analysis. Indeed, the lack of suitable uniform control of the discrete Laplacian was already identified in [14 ] as a main obstruction to a fully discrete rate analysis. To the best of our knowledge, no convergence rate has previously been proved for a fully discrete structure-preserving approximation of the sLLG equation. Even for deterministic LLG, rigorous optimal convergence rates towards strong solutions for the corresponding midpoint scheme have only been obtained recently [21 ] . This highlights the difficulty of quantitative error analysis for the midpoint scheme, even before additional complications caused by multiplicative Stratonovich noise are introduced.
The aim of this work is to close this gap for the one-dimensional sLLG equation with coloured multiplicative noise. We analyse a mass-lumped P 1 P_1 P 1 finite element discretisation combined with the midpoint rule in time, which preserves the unit-length constraint exactly at every finite-element node. The main difficulty is to obtain quantitative error bounds without destroying this geometric structure. To this end, we first derive uniform higher-moment estimates for the discrete exchange energy and time increments. We then develop a localised error analysis in which the deterministic drift consistency error and the stochastic midpoint defect are treated separately, since the latter requires a delicate control of diagonal and off-diagonal quadratic Brownian increments. Under suitable initial data and noise regularity assumptions, this yields a fully discrete convergence rate in probability of order h + k γ h+k^\gamma h + k γ for every γ < 1 2 \gamma< \frac{1}{2} γ < 2 1 in the energy norm. In particular, this provides the first rigorous convergence rate result for a fully discrete structure-preserving scheme for sLLG.
Our main result shows that, under suitable regularity assumptions on the initial data and the noise, for every γ ∈ ( 0 , 1 2 ) \gamma\in(0,\frac{1}{2}) γ ∈ ( 0 , 2 1 ) and every fixed localisation level Λ > 0 \Lambda> 0 Λ > 0 ,
E [ 1 Ω Λ , J max 0 ≤ j ≤ J ∥ m ( t j ) − m h j ∥ L 2 2 ] + E [ 1 Ω Λ , J k ∑ j = 0 J − 1 ∥ ∂ x ( m ( t j + 1 ) − m h j + 1 ) ∥ L 2 2 ] ≤ C Λ , γ ( h 2 + k 2 γ ) . \mathbb{E}\left[1_{\Omega_{\Lambda,J}}\max_{0\leq j\leq J}\left\|\mathbf{m}(t_j)-\mathbf{m}_h^j\right\|_{L^2}^2\right]+\mathbb{E}\left[1_{\Omega_{\Lambda,J}}k\sum_{j=0}^{J-1}\left\|\partial_x\left(\mathbf{m}(t_{j+1})-\mathbf{m}_h^{j+1}\right)\right\|_{L^2}^2\right]\leq C_{\Lambda,\gamma}\left(h^2+k^{2\gamma}\right). E [ 1 Ω Λ , J 0 ≤ j ≤ J max m ( t j ) − m h j L 2 2 ] + E [ 1 Ω Λ , J k j = 0 ∑ J − 1 ∂ x ( m ( t j + 1 ) − m h j + 1 ) L 2 2 ] ≤ C Λ , γ ( h 2 + k 2 γ ) .
Together with the probability estimate for the localisation sets, this implies a convergence rate of order h + k γ h+k^\gamma h + k γ , for every γ < 1 2 \gamma< \frac{1}{2} γ < 2 1 , in probability in the discrete L ∞ ( 0 , T ; L 2 ) ∩ L 2 ( 0 , T ; H 1 ) L^\infty(0,T;L^2)\cap L^2(0,T;H^1) L ∞ ( 0 , T ; L 2 ) ∩ L 2 ( 0 , T ; H 1 ) energy norm. Thus, the spatial discretisation does not deteriorate the temporal rate obtained for the time-semidiscrete midpoint method in [14 ] , while yielding simultaneously an optimal spatial rate in the H 1 H^1 H 1 -norm. Since comprehensive numerical experiments for midpoint discretisation of sLLG are already available in [7 , 14 ] , we do not repeat such simulations here and focus instead on the rigorous fully discrete convergence-rate analysis.
The remainder of the paper is organised as follows. Section 2 introduces the notation, standing assumptions, and finite-element preliminaries. In Section 3, we define the fully discrete midpoint scheme, establish its nodal constraint preservation and moment stability, and derive the consistency and localised error estimates. Various technical discrete estimates used in the analysis are collected in the appendix.
Preliminaries
In this section, we collect the notation, assumption, and finite element tools used throughout the paper.
Notation
For p ∈ [ 1 , ∞ ] p \in[1,\infty] p ∈ [ 1 , ∞ ] and s ≥ 0 s \geq0 s ≥ 0 , set
L p : = L p ( D ; R 3 ) , W s , p : = W s , p ( D ; R 3 ) , H s : = W s , 2 . L^p := L^p(D;\mathbb{R}^3), \qquad W^{s,p} := W^{s,p}(D;\mathbb{R}^3), \qquad H^s := W^{s,2}. L p := L p ( D ; R 3 ) , W s , p := W s , p ( D ; R 3 ) , H s := W s , 2 .
For a Banach space X X X , the spaces L p ( 0 , T ; X ) L^p(0,T;X) L p ( 0 , T ; X ) , W s , p ( 0 , T ; X ) W^{s,p}(0,T;X) W s , p ( 0 , T ; X ) , and C γ ( [ 0 , T ] ; X ) C^\gamma([0,T];X) C γ ([ 0 , T ] ; X ) have their usual vector-valued meanings. The operators ∂ x \partial_x ∂ x and ∂ x x \partial_{xx} ∂ xx act componentwise.
The scalar product in a Hilbert space H H H is denoted by ⟨ ⋅ , ⋅ ⟩ H \langle\cdot,\cdot\rangle_H ⟨ ⋅ , ⋅ ⟩ H , with corresponding norm ∥ ⋅ ∥ H \|\cdot\|_H ∥ ⋅ ∥ H . When H = L 2 H=L^2 H = L 2 , we simply write
⟨ u , v ⟩ : = ∫ D u ⋅ v d x . \langle u,v\rangle:= \int_D u\cdot v\,\mathrm{d}x. ⟨ u , v ⟩ := ∫ D u ⋅ v d x .
We use the same notation for vector- and matrix-valued functions whenever the meaning is clear from the context.
Throughout, C > 0 C > 0 C > 0 denotes a generic constant that may depend on fixed problem data but is independent of the discretisation parameters h h h and k k k . Dependence on specific parameters is indicated when relevant.
Assumption
The following assumption on initial data and noise regularity is made throughout this paper.
Assumption 2.1. We assume that the initial data m 0 ∈ H 3 ( D ; S 2 ) \mathbf{m}_0 \in H^3(D;\mathbb{S}^2) m 0 ∈ H 3 ( D ; S 2 ) with ∂ x m 0 = 0 \partial_x\mathbf{m}_0 = 0 ∂ x m 0 = 0 . To describe the noise, let
W ( t ) = ∑ ℓ = 1 ∞ g ℓ β ℓ ( t ) . (5) W(t)=\sum_{\ell=1}^{\infty}\mathbf{g}_\ell\beta_\ell(t). \tag*{(5)} W ( t ) = ℓ = 1 ∑ ∞ g ℓ β ℓ ( t ) . (5)
where { β ℓ } ℓ = 1 ∞ \{\beta_\ell\}_{\ell=1}^{\infty} { β ℓ } ℓ = 1 ∞ are independent standard { F t } \{\mathcal{F}_t\} { F t } -Brownian motions and g ℓ ∈ H 3 \mathbf{g}_\ell\in H^3 g ℓ ∈ H 3 are deterministic, such that
Γ : = ∑ ℓ = 1 ∞ ∥ g ℓ ∥ H 3 < ∞ . (6) \Gamma:=\sum_{\ell=1}^{\infty}\|\mathbf{g}_\ell\|_{H^3}<\infty. \tag*{(6)} Γ := ℓ = 1 ∑ ∞ ∥ g ℓ ∥ H 3 < ∞. (6)
We assume the compatibility condition ∂ x g ℓ = 0 \partial_x g_{\ell}=0 ∂ x g ℓ = 0 on ∂ D \partial D ∂ D for all ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 . This boundary compatibility condition is only needed to ensure that the stochastic vector fields preserve the homogeneous Neumann condition of a sufficiently regular solution.
Assumption 2.1 guarantees the existence of a strong pathwise solution to the stochastic LLG equation eq:1.1 with the following regularity: for every q ≥ 1 q \ge1 q ≥ 1 and every γ ∈ ( 0 , 1 2 ) \gamma\in(0,\frac{1}{2}) γ ∈ ( 0 , 2 1 ) ,
E [ ∥ m ∥ L 2 q ( Ω ; L ∞ ( 0 , T ; H 3 ) ) 2 q + ∥ m ∥ C γ ( [ 0 , T ] ; H 2 ) 2 q ] ≤ C q , γ . (7) \mathbb{E}\left[\|m\|_{L^{2q}(\Omega;L^{\infty}(0,T;H^{3}))}^{2q}+\|m\|_{C^{\gamma}([0,T];H^{2})}^{2q}\right]\le C_{q,\gamma}. \tag*{(7)} E [ ∥ m ∥ L 2 q ( Ω ; L ∞ ( 0 , T ; H 3 )) 2 q + ∥ m ∥ C γ ([ 0 , T ] ; H 2 ) 2 q ] ≤ C q , γ . (7)
Such pathwise well-posedness and regularity are established in [14 ] for the periodic problem; also see [18 ] . The corresponding Neumann-boundary result follows by the same argument under the compatibility conditions imposed above.
Finite element approximation
Let { T h } h > 0 \{\mathcal{T}_{h}\}_{h>0} { T h } h > 0 be a family of shape-regular and quasi-uniform partitions of D D D into intervals, with maximal mesh-size h > 0 h>0 h > 0 . We denote by N h \mathcal{N}_{h} N h the set of nodes of T h \mathcal{T}_{h} T h . We use the lowest-order conforming Lagrange space
V h : = { ϕ h ∈ C 0 ( D ‾ ; R 3 ) : ϕ h ∣ K ∈ P 1 ( K ; R 3 ) , ∀ K ∈ T h } ⊂ H 1 . (8) V_{h}:=\left\{\phi_{h}\in C^{0}(\overline{D};\mathbb{R}^{3}):\left.\phi_{h}\right|_{K}\in\mathcal{P}_{1}(K;\mathbb{R}^{3}),\ \forall K\in\mathcal{T}_{h}\right\}\subset H^{1}. \tag*{(8)} V h := { ϕ h ∈ C 0 ( D ; R 3 ) : ϕ h ∣ K ∈ P 1 ( K ; R 3 ) , ∀ K ∈ T h } ⊂ H 1 . (8)
We equip V h V_{h} V h with the norm
∥ ϕ h ∥ 1 , h : = ∥ ϕ h ∥ h + ∥ ∂ x ϕ h ∥ L 2 . (9) \|\phi_{h}\|_{1,h}:=\|\phi_{h}\|_{h}+\|\partial_{x}\phi_{h}\|_{L^{2}}. \tag*{(9)} ∥ ϕ h ∥ 1 , h := ∥ ϕ h ∥ h + ∥ ∂ x ϕ h ∥ L 2 . (9)
For a functional R ∈ V h ′ \mathcal{R}\in V_{h}^{\prime} R ∈ V h ′ , we define the corresponding dual norm by
∥ R ∥ − 1 , h : = sup 0 ≠ ϕ h ∈ V h ∣ R ( ϕ h ) ∣ ∥ ϕ h ∥ 1 , h . (10) \|\mathcal{R}\|_{-1,h}:=\sup_{0\ne\phi_{h}\in V_{h}}\frac{|\mathcal{R}(\phi_{h})|}{\|\phi_{h}\|_{1,h}}. \tag*{(10)} ∥ R ∥ − 1 , h := 0 = ϕ h ∈ V h sup ∥ ϕ h ∥ 1 , h ∣ R ( ϕ h ) ∣ . (10)
Let { φ z } z ∈ N h \{\varphi_{z}\}_{z\in\mathcal{N}_{h}} { φ z } z ∈ N h be the scalar nodal basis and denote by I h I_{h} I h the nodal interpolation operator, for both scalar- and vector-valued functions. Since D ⊂ R D\subset\mathbb{R} D ⊂ R is a bounded interval, H 1 ( D ) ↪ C 0 ( D ‾ ) H^{1}(D)\hookrightarrow C^{0}(\overline{D}) H 1 ( D ) ↪ C 0 ( D ) . For continuous vector fields u , v ∈ C 0 ( D ‾ ; R 3 ) u,v\in C^{0}(\overline{D};\mathbb{R}^{3}) u , v ∈ C 0 ( D ; R 3 ) , define the mass-lumped inner product by
⟨ u , v ⟩ h : = ∫ D I h ( u ⋅ v ) d x = ∑ z ∈ N h β z u ( z ) ⋅ v ( z ) , β z : = ∫ D φ z d x . (11) \langle u,v\rangle_{h}:=\int_{D}I_{h}(u\cdot v)\,\mathrm{d}x=\sum_{z\in\mathcal{N}_{h}}\beta_{z}u(z)\cdot v(z),\qquad\beta_{z}:=\int_{D}\varphi_{z}\,\mathrm{d}x. \tag*{(11)} ⟨ u , v ⟩ h := ∫ D I h ( u ⋅ v ) d x = z ∈ N h ∑ β z u ( z ) ⋅ v ( z ) , β z := ∫ D φ z d x . (11)
We write
∥ v h ∥ h 2 : = ⟨ v h , v h ⟩ h . (12) \|v_{h}\|_{h}^{2}:=\langle v_{h},v_{h}\rangle_{h}. \tag*{(12)} ∥ v h ∥ h 2 := ⟨ v h , v h ⟩ h . (12)
The norm ∥ ⋅ ∥ h \|\cdot\|_{h} ∥ ⋅ ∥ h is uniformly equivalent to the standard L 2 L^{2} L 2 -norm on V h V_{h} V h , namely
c ∥ v h ∥ L 2 2 ≤ ∥ v h ∥ h 2 ≤ C ∥ v h ∥ L 2 2 , ∀ v h ∈ V h , (13) c\|v_{h}\|_{L^{2}}^{2}\le\|v_{h}\|_{h}^{2}\le C\|v_{h}\|_{L^{2}}^{2},\qquad\forall v_{h}\in V_{h}, \tag*{(13)} c ∥ v h ∥ L 2 2 ≤ ∥ v h ∥ h 2 ≤ C ∥ v h ∥ L 2 2 , ∀ v h ∈ V h , (13)
where c , C > 0 c,C>0 c , C > 0 are independent of h h h . The following discrete Agmon inequality is a consequence of (13) and the standard one-dimensional Agmon inequality:
∥ v h ∥ L ∞ 2 ≤ C ∥ v h ∥ h ∥ ∂ x v h ∥ L 2 + C ∥ v h ∥ h 2 , ∀ v h ∈ V h . (14) \|v_{h}\|_{L^{\infty}}^{2}\le C\|v_{h}\|_{h}\|\partial_{x}v_{h}\|_{L^{2}}+C\|v_{h}\|_{h}^{2},\qquad\forall v_{h}\in V_{h}. \tag*{(14)} ∥ v h ∥ L ∞ 2 ≤ C ∥ v h ∥ h ∥ ∂ x v h ∥ L 2 + C ∥ v h ∥ h 2 , ∀ v h ∈ V h . (14)
We further define the mass-lumped discrete Laplacian Δ h : V h → V h \Delta_{h}:V_{h}\to V_{h} Δ h : V h → V h by
⟨ Δ h v h , χ h ⟩ h = − ⟨ ∂ x v h , ∂ x χ h ⟩ , ∀ v h , χ h ∈ V h . (15) \langle\Delta_{h}v_{h},\chi_{h}\rangle_{h}=-\langle\partial_{x}v_{h},\partial_{x}\chi_{h}\rangle,\qquad\forall v_{h},\chi_{h}\in V_{h}. \tag*{(15)} ⟨ Δ h v h , χ h ⟩ h = − ⟨ ∂ x v h , ∂ x χ h ⟩ , ∀ v h , χ h ∈ V h . (15)
We shall use the following standard approximation and inverse estimates. For each s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , there exists C > 0 C>0 C > 0 such that, for all v ∈ H 1 + s v\in H^{1+s} v ∈ H 1 + s ,
∥ v − I h v ∥ L 2 + h ∥ ∂ x ( v − I h v ) ∥ L 2 ≤ C h 1 + s ∥ v ∥ H 1 + s . (16) \|v-I_{h}v\|_{L^{2}}+h\|\partial_{x}(v-I_{h}v)\|_{L^{2}}\le Ch^{1+s}\|v\|_{H^{1+s}}. \tag*{(16)} ∥ v − I h v ∥ L 2 + h ∥ ∂ x ( v − I h v ) ∥ L 2 ≤ C h 1 + s ∥ v ∥ H 1 + s . (16)
In one-dimensional domain, we have the H 1 H^{1} H 1 -stability estimate:
∥ I h v ∥ H 1 ≤ C ∥ v ∥ H 1 . (17) \|I_{h}v\|_{H^{1}}\le C\|v\|_{H^{1}}. \tag*{(17)} ∥ I h v ∥ H 1 ≤ C ∥ v ∥ H 1 . (17)
Moreover, for every v h ∈ V h v_{h}\in V_{h} v h ∈ V h , s ∈ { 0 , 1 } s\in\{0,1\} s ∈ { 0 , 1 } , and 1 ≤ q ≤ p ≤ ∞ 1\le q\le p\le\infty 1 ≤ q ≤ p ≤ ∞ , the inverse estimate
∥ v h ∥ W s , p ≤ C h − ( 1 q − 1 p ) ∥ v h ∥ W s , q . (18) \|v_{h}\|_{W^{s,p}}\le Ch^{-\left(\frac{1}{q}-\frac{1}{p}\right)}\|v_{h}\|_{W^{s,q}}. \tag*{(18)} ∥ v h ∥ W s , p ≤ C h − ( q 1 − p 1 ) ∥ v h ∥ W s , q . (18)
holds with C C C independent of h h h .
In V h V_h V h , we have the following quadrature estimates associated with the mass-lumped inner product [8 ] :
∣ ⟨ v h , χ h ⟩ h − ⟨ v h , χ h ⟩ ∣ ≤ C h ∥ v h ∥ L 2 ∥ χ h ∥ H 1 , ∀ v h , χ h ∈ V h , (19) \left|\langle v_h,\chi_h\rangle_h-\langle v_h,\chi_h\rangle\right|\le Ch\|v_h\|_{L^2}\|\chi_h\|_{H^1},\qquad\forall v_h,\chi_h\in V_h, \tag*{(19)} ∣ ⟨ v h , χ h ⟩ h − ⟨ v h , χ h ⟩ ∣ ≤ C h ∥ v h ∥ L 2 ∥ χ h ∥ H 1 , ∀ v h , χ h ∈ V h , (19)
∣ ⟨ I h f , χ h ⟩ h − ⟨ f , χ h ⟩ ∣ ≤ C h ∥ f ∥ H 1 ∥ χ h ∥ 1 , h , ∀ f ∈ H 1 , χ h ∈ V h . (20) \left|\langle I_hf,\chi_h\rangle_h-\langle f,\chi_h\rangle\right|\le Ch\|f\|_{H^1}\|\chi_h\|_{1,h},\qquad\forall f\in H^1,\ \chi_h\in V_h. \tag*{(20)} ∣ ⟨ I h f , χ h ⟩ h − ⟨ f , χ h ⟩ ∣ ≤ C h ∥ f ∥ H 1 ∥ χ h ∥ 1 , h , ∀ f ∈ H 1 , χ h ∈ V h . (20)
We shall also repeatedly use the following elementary product estimates [21 ] : for all v h , w h ∈ V h v_h,w_h\in V_h v h , w h ∈ V h ,
∥ I h ( v h × w h ) ∥ L 2 ≤ C ∥ v h ∥ L ∞ ∥ w h ∥ L 2 , (21) \|I_h(v_h\times w_h)\|_{L^2}\le C\|v_h\|_{L^\infty}\|w_h\|_{L^2}, \tag*{(21)} ∥ I h ( v h × w h ) ∥ L 2 ≤ C ∥ v h ∥ L ∞ ∥ w h ∥ L 2 , (21)
∥ ∂ x I h ( v h × w h ) ∥ L 2 ≤ C ( ∥ v h ∥ L ∞ ∥ ∂ x w h ∥ L 2 + ∥ w h ∥ L ∞ ∥ ∂ x v h ∥ L 2 ) , (22) \|\partial_xI_h(v_h\times w_h)\|_{L^2}\le C\left(\|v_h\|_{L^\infty}\|\partial_xw_h\|_{L^2}+\|w_h\|_{L^\infty}\|\partial_xv_h\|_{L^2}\right), \tag*{(22)} ∥ ∂ x I h ( v h × w h ) ∥ L 2 ≤ C ( ∥ v h ∥ L ∞ ∥ ∂ x w h ∥ L 2 + ∥ w h ∥ L ∞ ∥ ∂ x v h ∥ L 2 ) , (22)
∥ ∂ x I h ( v h × w h ) ∥ L 2 ≤ C ( ∥ w h ∥ L ∞ ∥ ∂ x v h ∥ L 2 + ∥ ∂ x w h ∥ L ∞ ∥ v h ∥ h ) . (23) \|\partial_xI_h(v_h\times w_h)\|_{L^2}\le C\left(\|w_h\|_{L^\infty}\|\partial_xv_h\|_{L^2}+\|\partial_xw_h\|_{L^\infty}\|v_h\|_h\right). \tag*{(23)} ∥ ∂ x I h ( v h × w h ) ∥ L 2 ≤ C ( ∥ w h ∥ L ∞ ∥ ∂ x v h ∥ L 2 + ∥ ∂ x w h ∥ L ∞ ∥ v h ∥ h ) . (23)
Furthermore, noting (15) , we have the following one-dimensional discrete cancellation estimate:
∣ ⟨ v h × η h , − Δ h v h ⟩ h ∣ = ∣ ⟨ ∂ x I h ( v h × η h ) , ∂ x v h ⟩ ∣ ≤ C ∥ v h ∥ L ∞ ∥ ∂ x η h ∥ L 2 ∥ ∂ x v h ∥ L 2 . (24) \left|\langle v_h\times\eta_h,-\Delta_hv_h\rangle_h\right|=\left|\langle\partial_xI_h(v_h\times\eta_h),\partial_xv_h\rangle\right|\le C\|v_h\|_{L^\infty}\|\partial_x\eta_h\|_{L^2}\|\partial_xv_h\|_{L^2}. \tag*{(24)} ∣ ⟨ v h × η h , − Δ h v h ⟩ h ∣ = ∣ ⟨ ∂ x I h ( v h × η h ) , ∂ x v h ⟩ ∣ ≤ C ∥ v h ∥ L ∞ ∥ ∂ x η h ∥ L 2 ∥ ∂ x v h ∥ L 2 . (24)
We shall also need the following product-interpolation estimates.
Lemma 2.2 (Product-interpolation estimates). Let a ∈ H 2 a\in H^2 a ∈ H 2 and ϕ h ∈ V h \phi_h\in V_h ϕ h ∈ V h . Then
∥ ∂ x [ I h ( ϕ h × I h a ) − ϕ h × a ] ∥ L 2 ≤ C h ∥ a ∥ H 2 ∥ ϕ h ∥ 1 , h , (25) \|\partial_x[I_h(\phi_h\times I_ha)-\phi_h\times a]\|_{L^2}\le Ch\|a\|_{H^2}\|\phi_h\|_{1,h}, \tag*{(25)} ∥ ∂ x [ I h ( ϕ h × I h a ) − ϕ h × a ] ∥ L 2 ≤ C h ∥ a ∥ H 2 ∥ ϕ h ∥ 1 , h , (25)
∥ ∂ x [ I h ( ( ϕ h × I h a ) × I h a ) − ( ϕ h × a ) × a ] ∥ L 2 ≤ C h ∥ a ∥ H 2 ∥ ϕ h ∥ 1 , h . (26) \|\partial_x[I_h((\phi_h\times I_ha)\times I_ha)-(\phi_h\times a)\times a]\|_{L^2}\le Ch\|a\|_{H^2}\|\phi_h\|_{1,h}. \tag*{(26)} ∥ ∂ x [ I h (( ϕ h × I h a ) × I h a ) − ( ϕ h × a ) × a ] ∥ L 2 ≤ C h ∥ a ∥ H 2 ∥ ϕ h ∥ 1 , h . (26)
Proof. For every z ∈ N h z\in\mathcal{N}_h z ∈ N h , we have I h a ( z ) = a ( z ) I_ha(z)=a(z) I h a ( z ) = a ( z ) . Hence, the two differences are the elementwise interpolation errors of ϕ h × a \phi_h\times a ϕ h × a and ( ϕ h × a ) × a (\phi_h\times a)\times a ( ϕ h × a ) × a , respectively. The elementwise H 1 H^1 H 1 -interpolation estimate, the fact that ϕ h \phi_h ϕ h is affine on each element, and the one-dimensional embeddings H 1 ↪ L ∞ H^1\hookrightarrow L^\infty H 1 ↪ L ∞ and H 2 ↪ W 1 , ∞ H^2\hookrightarrow W^{1,\infty} H 2 ↪ W 1 , ∞ yield the asserted bounds. □
Next, we record a moment estimate for the discrete Brownian increments. Set
g ℓ , h : = I h g ℓ , W h ( t ) : = ∑ ℓ = 1 ∞ g ℓ , h β ℓ ( t ) . (27) g_{\ell,h}:=I_hg_\ell,\qquad W_h(t):=\sum_{\ell=1}^{\infty}g_{\ell,h}\beta_\ell(t). \tag*{(27)} g ℓ , h := I h g ℓ , W h ( t ) := ℓ = 1 ∑ ∞ g ℓ , h β ℓ ( t ) . (27)
For j = 0 , … , J − 1 j=0,\ldots,J-1 j = 0 , … , J − 1 , define
Δ j β ℓ ^ : = β ℓ ( t j + 1 ) − β ℓ ( t j ) , Δ j W h ^ : = W h ( t j + 1 ) − W h ( t j ) = ∑ ℓ = 1 ∞ g ℓ , h Δ j β ℓ ^ . \widehat{\Delta_j\beta_\ell}:=\beta_\ell(t_{j+1})-\beta_\ell(t_j),\qquad\widehat{\Delta_jW_h}:=W_h(t_{j+1})-W_h(t_j)=\sum_{\ell=1}^{\infty}g_{\ell,h}\widehat{\Delta_j\beta_\ell}. Δ j β ℓ := β ℓ ( t j + 1 ) − β ℓ ( t j ) , Δ j W h := W h ( t j + 1 ) − W h ( t j ) = ℓ = 1 ∑ ∞ g ℓ , h Δ j β ℓ .
Under Assumption (5) , for every q ≥ 1 q\ge1 q ≥ 1 ,
E [ ∥ Δ j W h ^ ∥ W 1 , ∞ 2 q ] ≤ C q Γ 2 q k q . (28) \mathbb{E}\left[\left\|\widehat{\Delta_jW_h}\right\|_{W^{1,\infty}}^{2q}\right]\le C_q\Gamma^{2q}k^q. \tag*{(28)} E [ Δ j W h W 1 , ∞ 2 q ] ≤ C q Γ 2 q k q . (28)
Indeed, the W 1 , ∞ W^{1,\infty} W 1 , ∞ -stability of I h I_h I h and Minkowski’s inequality give
( E [ ∥ Δ j W h ^ ∥ W 1 , ∞ 2 q ] ) 1 2 q ≤ ∑ ℓ = 1 ∞ ∥ I h g ℓ ∥ W 1 , ∞ ( E ∣ Δ j β ℓ ^ ∣ 2 q ) 1 2 q ≤ C q Γ k 1 2 . \left(\mathbb{E}\left[\left\|\widehat{\Delta_jW_h}\right\|_{W^{1,\infty}}^{2q}\right]\right)^{\frac{1}{2q}}\le\sum_{\ell=1}^{\infty}\|I_hg_\ell\|_{W^{1,\infty}}\left(\mathbb{E}\left|\widehat{\Delta_j\beta_\ell}\right|^{2q}\right)^{\frac{1}{2q}}\le C_q\Gamma k^{\frac{1}{2}}. ( E [ Δ j W h W 1 , ∞ 2 q ] ) 2 q 1 ≤ ℓ = 1 ∑ ∞ ∥ I h g ℓ ∥ W 1 , ∞ ( E Δ j β ℓ 2 q ) 2 q 1 ≤ C q Γ k 2 1 .
Stability and error estimates of the fully discrete midpoint scheme
Let 0 = t 0 < t 1 < ⋯ < t J = T 0=t_0<t_1<\cdots<t_J=T 0 = t 0 < t 1 < ⋯ < t J = T be the uniform partition of [ 0 , T ] [0,T] [ 0 , T ] with time-step k = T / J k=T/J k = T / J . The fully discrete mass-lumped midpoint approximation is defined by the following algorithm.
Existence of an adapted solution { m h j } j = 0 J \{m_h^j\}_{j=0}^J { m h j } j = 0 J to Algorithm 3.1 was shown in [6 ] . In the subsequent analysis, we assume k ∈ ( 0 , 1 ] k\in(0,1] k ∈ ( 0 , 1 ] , which entails no restriction in the asymptotic error analysis.
Algorithm 3.1 Mass-lumped midpoint scheme
Initialization. Set m h 0 : = I h m 0 m_h^0 := I_hm_0 m h 0 := I h m 0 .
Time-stepping. For j = 0 , … , J − 1 j=0,\ldots,J-1 j = 0 , … , J − 1 , given m h j ∈ V h m_h^j \in V_h m h j ∈ V h , find m h j + 1 ∈ V h m_h^{j+1} \in V_h m h j + 1 ∈ V h such that, with
m h j + 1 2 : = 1 2 ( m h j + 1 + m h j ) (29) m_h^{j+\frac{1}{2}} := \frac{1}{2}\left(m_h^{j+1}+m_h^j\right) \tag*{(29)} m h j + 2 1 := 2 1 ( m h j + 1 + m h j ) (29)
there holds, for every ϕ h ∈ V h \phi_h \in V_h ϕ h ∈ V h ,
⟨ m h j + 1 − m h j , ϕ h ⟩ h + α κ k ⟨ m h j + 1 2 × [ m h j + 1 2 × Δ h m h j + 1 ] , ϕ h ⟩ h − κ k ⟨ m h j + 1 2 × Δ h m h j + 1 , ϕ h ⟩ h = ⟨ m h j + 1 2 × Δ ^ j W h , ϕ h ⟩ h . (30) \left\langle m_h^{j+1}-m_h^j,\phi_h\right\rangle_h+\alpha\kappa k\left\langle m_h^{j+\frac{1}{2}}\times\left[m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right],\phi_h\right\rangle_h
-\kappa k\left\langle m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1},\phi_h\right\rangle_h
=\left\langle m_h^{j+\frac{1}{2}}\times\widehat{\Delta}_jW_h,\phi_h\right\rangle_h. \tag*{(30)} ⟨ m h j + 1 − m h j , ϕ h ⟩ h + α κk ⟨ m h j + 2 1 × [ m h j + 2 1 × Δ h m h j + 1 ] , ϕ h ⟩ h − κk ⟨ m h j + 2 1 × Δ h m h j + 1 , ϕ h ⟩ h = ⟨ m h j + 2 1 × Δ j W h , ϕ h ⟩ h . (30)
Stability estimates
We first record the nodal constraint and the corresponding L ∞ \mathbb{L}^{\infty} L ∞ -bound.
Lemma 3.2 (Nodal constraint and L ∞ \mathbb{L}^{\infty} L ∞ -bound). Assume that ∣ m h 0 ( z ) ∣ = 1 \lvert m_h^0(z)\rvert=1 ∣ m h 0 ( z )∣ = 1 for all z ∈ N h z\in\mathcal{N}_h z ∈ N h . Then, for every j = 0 , 1 , … , J j=0,1,\ldots,J j = 0 , 1 , … , J , every solution of (30) satisfies
∣ m h j ( z ) ∣ = 1 , ∀ z ∈ N h , P -a.s. (31) \lvert m_h^j(z)\rvert=1,\qquad\forall z\in\mathcal{N}_h,\quad\mathbb{P}\text{-a.s.} \tag*{(31)} ∣ m h j ( z )∣ = 1 , ∀ z ∈ N h , P -a.s. (31)
Consequently, for every j = 0 , 1 , … , J j=0,1,\ldots,J j = 0 , 1 , … , J ,
∥ m h j ∥ L ∞ ≤ 1 , P -a.s. (32) \lVert m_h^j\rVert_{\mathbb{L}^{\infty}}\leq1,\qquad\mathbb{P}\text{-a.s.} \tag*{(32)} ∥ m h j ∥ L ∞ ≤ 1 , P -a.s. (32)
Proof. Fix a time level j j j and a node z ∈ N h z\in\mathcal{N}_h z ∈ N h . Choose ϕ h = φ z m h j + 1 2 ( z ) \phi_h=\varphi_zm_h^{j+\frac{1}{2}}(z) ϕ h = φ z m h j + 2 1 ( z ) in the discrete scheme. Since the mass-lumped inner product is diagonal and β z > 0 \beta_z>0 β z > 0 , all cross-product terms vanish and
0 = β z ( m h j + 1 ( z ) − m h j ( z ) ) ⋅ m h j + 1 2 ( z ) = β z 2 ( ∣ m h j + 1 ( z ) ∣ 2 − ∣ m h j ( z ) ∣ 2 ) . 0=\beta_z\left(m_h^{j+1}(z)-m_h^j(z)\right)\cdot m_h^{j+\frac{1}{2}}(z)=\frac{\beta_z}{2}\left(\lvert m_h^{j+1}(z)\rvert^2-\lvert m_h^j(z)\rvert^2\right). 0 = β z ( m h j + 1 ( z ) − m h j ( z ) ) ⋅ m h j + 2 1 ( z ) = 2 β z ( ∣ m h j + 1 ( z ) ∣ 2 − ∣ m h j ( z ) ∣ 2 ) .
This implies (31) . On every element, m h j m_h^j m h j is a convex combination of its two nodal values. Hence, ∣ m h j ( x ) ∣ ≤ 1 \lvert m_h^j(x)\rvert\leq1 ∣ m h j ( x )∣ ≤ 1 for all x ∈ D ‾ x\in\overline{D} x ∈ D , which proves (32) . □ \square □
We start with the following estimate for the stochastic term.
Lemma 3.3. Let p ≥ 2 p\geq2 p ≥ 2 and let { m h j } j = 0 J \{m_h^j\}_{j=0}^J { m h j } j = 0 J be an adapted solution of (30) . Then
E [ max 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 κ ⟨ ∂ x I h ( m h j × Δ ^ j W h ) , ∂ x m h j ⟩ ∣ p ] ≤ C p , T , κ , Γ ( 1 + ∑ j = 0 J − 1 k E [ ∥ ∂ x m h j ∥ L 2 2 p ] ) . (33) \mathbb{E}\left[\max_{0\leq r\leq J}\left\lvert\sum_{j=0}^{r-1}\kappa\left\langle\partial_xI_h\left(m_h^j\times\widehat{\Delta}_jW_h\right),\partial_xm_h^j\right\rangle\right\rvert^p\right]
\leq C_{p,T,\kappa,\Gamma}\left(1+\sum_{j=0}^{J-1}k\mathbb{E}\left[\left\lVert\partial_xm_h^j\right\rVert_{\mathbb{L}^2}^{2p}\right]\right). \tag*{(33)} E [ 0 ≤ r ≤ J max j = 0 ∑ r − 1 κ ⟨ ∂ x I h ( m h j × Δ j W h ) , ∂ x m h j ⟩ p ] ≤ C p , T , κ , Γ ( 1 + j = 0 ∑ J − 1 k E [ ∂ x m h j L 2 2 p ] ) . (33)
where C p , T , κ , Γ C_{p,T,\kappa,\Gamma} C p , T , κ , Γ is independent of h h h and k k k .
Proof. Set J j + 1 : = κ ⟨ ∂ x I h ( m h j × Δ ^ j W h ) , ∂ x m h j ⟩ \mathcal{J}_{j+1}:=\kappa\left\langle\partial_xI_h\left(m_h^j\times\widehat{\Delta}_jW_h\right),\partial_xm_h^j\right\rangle J j + 1 := κ ⟨ ∂ x I h ( m h j × Δ j W h ) , ∂ x m h j ⟩ . The sequence { J j + 1 } \{\mathcal{J}_{j+1}\} { J j + 1 } is a martingale difference sequence. By (22) , (28) , and independence of the Brownian increment from F t j \mathcal{F}_{t_j} F t j , we have
E [ ∣ J j + 1 ∣ p ] ≤ C p k p 2 E [ ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) p ] . \mathbb{E}\left[\lvert\mathcal{J}_{j+1}\rvert^p\right]\leq C_pk^{\frac{p}{2}}\mathbb{E}\left[\left(1+\left\lVert\partial_xm_h^j\right\rVert_{\mathbb{L}^2}^2\right)^p\right]. E [ ∣ J j + 1 ∣ p ] ≤ C p k 2 p E [ ( 1 + ∂ x m h j L 2 2 ) p ] .
The discrete BDG inequality followed by Lemma A.1 proves the result. □ \square □
We also have the following p p p -th moment bound.
Proposition 3.4 (p-th moment bound). Let p ≥ 2 p \ge2 p ≥ 2 , let α , κ > 0 \alpha, \kappa> 0 α , κ > 0 , and suppose Assumption 2.1 holds. Let { m h j } j = 0 J \{m_h^j\}_{j=0}^J { m h j } j = 0 J be an adapted solution of Algorithm 3.1. Then there exists a constant C p , T , α , κ , Γ > 0 C_{p,T,\alpha,\kappa,\Gamma} > 0 C p , T , α , κ , Γ > 0 , independent of h h h and k k k , such that
E [ max 0 ≤ j ≤ J ∥ ∂ x m h j ∥ L 2 2 p ] + E [ ( ∑ j = 0 J − 1 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 ) p ] + E [ ( ∑ j = 0 J − 1 k ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ) p ] ≤ C p , T , α , κ , Γ ( 1 + E [ ∥ ∂ x m h 0 ∥ L 2 2 p ] ) . (34) \mathbb{E}\left[\max_{0\le j\le J}\left\|\partial_xm_h^j\right\|_{L^2}^{2p}\right] + \mathbb{E}\left[\left(\sum_{j=0}^{J-1}\left\|\partial_x\left(m_h^{j+1}-m_h^j\right)\right\|_{L^2}^2\right)^p\right]
+ \mathbb{E}\left[\left(\sum_{j=0}^{J-1} k\left\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\right\|_h^2\right)^p\right]
\le C_{p,T,\alpha,\kappa,\Gamma}\left(1+\mathbb{E}\left[\left\|\partial_xm_h^0\right\|_{L^2}^{2p}\right]\right). \tag*{(34)} E [ 0 ≤ j ≤ J max ∂ x m h j L 2 2 p ] + E [ ( j = 0 ∑ J − 1 ∂ x ( m h j + 1 − m h j ) L 2 2 ) p ] + E [ ( j = 0 ∑ J − 1 k m h j + 2 1 × Δ h m h j + 1 h 2 ) p ] ≤ C p , T , α , κ , Γ ( 1 + E [ ∂ x m h 0 L 2 2 p ] ) . (34)
Proof. Since the mass-lumped scheme (30) holds nodewise and ∥ m h j + 1 2 ∥ L ∞ ≤ 1 \left\|m_h^{j+\frac12}\right\|_{L^\infty}\le1 m h j + 2 1 L ∞ ≤ 1 , we have
∥ m h j + 1 − m h j ∥ h ≤ C k ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h + C ∥ Δ j W h ^ ∥ L ∞ . (35) \left\|m_h^{j+1}-m_h^j\right\|_h\le Ck\left\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\right\|_h+C\left\|\widehat{\Delta_jW_h}\right\|_{L^\infty}. \tag*{(35)} m h j + 1 − m h j h ≤ C k m h j + 2 1 × Δ h m h j + 1 h + C Δ j W h L ∞ . (35)
Testing (30) with − κ Δ h m h j + 1 -\kappa\Delta_hm_h^{j+1} − κ Δ h m h j + 1 gives
κ 2 ( ∥ ∂ x m h j + 1 ∥ L 2 2 − ∥ ∂ x m h j ∥ L 2 2 + ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 ) + α κ 2 k ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 = κ ⟨ ∂ x I h ( m h j + 1 2 × Δ j W h ^ ) , ∂ x m h j + 1 ⟩ . (36) \frac{\kappa}{2}\left(\left\|\partial_xm_h^{j+1}\right\|_{L^2}^2-\left\|\partial_xm_h^j\right\|_{L^2}^2+\left\|\partial_x\left(m_h^{j+1}-m_h^j\right)\right\|_{L^2}^2\right)+\alpha\kappa^2k\left\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\right\|_h^2
=\kappa\left\langle\partial_xI_h\left(m_h^{j+\frac12}\times\widehat{\Delta_jW_h}\right),\partial_xm_h^{j+1}\right\rangle. \tag*{(36)} 2 κ ( ∂ x m h j + 1 L 2 2 − ∂ x m h j L 2 2 + ∂ x ( m h j + 1 − m h j ) L 2 2 ) + α κ 2 k m h j + 2 1 × Δ h m h j + 1 h 2 = κ ⟨ ∂ x I h ( m h j + 2 1 × Δ j W h ) , ∂ x m h j + 1 ⟩ . (36)
By using the identity
m h j + 1 2 = m h j + 1 2 ( m h j + 1 − m h j ) , m_h^{j+\frac12}=m_h^j+\frac12\left(m_h^{j+1}-m_h^j\right), m h j + 2 1 = m h j + 2 1 ( m h j + 1 − m h j ) ,
after adding and subtracting a common term, the right-hand side of (36) becomes
κ ⟨ ∂ x I h ( m h j × Δ j W h ^ ) , ∂ x m h j ⟩ + κ ⟨ ∂ x I h ( m h j × Δ j W h ^ ) , ∂ x ( m h j + 1 − m h j ) ⟩ + κ 2 ⟨ ∂ x I h ( ( m h j + 1 − m h j ) × Δ j W h ^ ) , ∂ x m h j ⟩ + κ 2 ⟨ ∂ x I h ( ( m h j + 1 − m h j ) × Δ j W h ^ ) , ∂ x ( m h j + 1 − m h j ) ⟩ = : I 1 + … + I 4 . (37) \begin{aligned}
&\kappa\left\langle\partial_xI_h\left(m_h^j\times\widehat{\Delta_jW_h}\right),\partial_xm_h^j\right\rangle
+\kappa\left\langle\partial_xI_h\left(m_h^j\times\widehat{\Delta_jW_h}\right),\partial_x\left(m_h^{j+1}-m_h^j\right)\right\rangle\\
&\quad+\frac{\kappa}{2}\left\langle\partial_xI_h\left(\left(m_h^{j+1}-m_h^j\right)\times\widehat{\Delta_jW_h}\right),\partial_xm_h^j\right\rangle\\
&\quad+\frac{\kappa}{2}\left\langle\partial_xI_h\left(\left(m_h^{j+1}-m_h^j\right)\times\widehat{\Delta_jW_h}\right),\partial_x\left(m_h^{j+1}-m_h^j\right)\right\rangle=: I_1+\ldots+I_4. \tag*{(37)}
\end{aligned} κ ⟨ ∂ x I h ( m h j × Δ j W h ) , ∂ x m h j ⟩ + κ ⟨ ∂ x I h ( m h j × Δ j W h ) , ∂ x ( m h j + 1 − m h j ) ⟩ + 2 κ ⟨ ∂ x I h ( ( m h j + 1 − m h j ) × Δ j W h ) , ∂ x m h j ⟩ + 2 κ ⟨ ∂ x I h ( ( m h j + 1 − m h j ) × Δ j W h ) , ∂ x ( m h j + 1 − m h j ) ⟩ =: I 1 + … + I 4 . (37)
We now estimate each term above, except for I 1 I_1 I 1 . The second term satisfies
∣ I 2 ∣ ≤ κ 12 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 + C ∥ Δ j W h ^ ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) . |I_2|\le\frac{\kappa}{12}\left\|\partial_x\left(m_h^{j+1}-m_h^j\right)\right\|_{L^2}^2+C\left\|\widehat{\Delta_jW_h}\right\|_{W^{1,\infty}}^2\left(1+\left\|\partial_xm_h^j\right\|_{L^2}^2\right). ∣ I 2 ∣ ≤ 12 κ ∂ x ( m h j + 1 − m h j ) L 2 2 + C Δ j W h W 1 , ∞ 2 ( 1 + ∂ x m h j L 2 2 ) .
For the term I 3 I_3 I 3 , (23) and (35) give
∣ I 3 ∣ ≤ κ 12 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 + α κ 2 4 k ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 + C ∥ Δ j W h ^ ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) . |I_3|\le\frac{\kappa}{12}\left\|\partial_x\left(m_h^{j+1}-m_h^j\right)\right\|_{L^2}^2+\frac{\alpha\kappa^2}{4}k\left\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\right\|_h^2+C\left\|\widehat{\Delta_jW_h}\right\|_{W^{1,\infty}}^2\left(1+\left\|\partial_xm_h^j\right\|_{L^2}^2\right). ∣ I 3 ∣ ≤ 12 κ ∂ x ( m h j + 1 − m h j ) L 2 2 + 4 α κ 2 k m h j + 2 1 × Δ h m h j + 1 h 2 + C Δ j W h W 1 , ∞ 2 ( 1 + ∂ x m h j L 2 2 ) .
Finally, using (24) and an elementary inequality ∥ m h j + 1 − m h j ∥ L ∞ ≤ 2 \left\|m_h^{j+1}-m_h^j\right\|_{L^\infty}\le2 m h j + 1 − m h j L ∞ ≤ 2 , we obtain
∣ I 4 ∣ ≤ κ 12 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 + C ∥ Δ j W h ^ ∥ W 1 , ∞ 2 . |I_4|\le\frac{\kappa}{12}\left\|\partial_x\left(m_h^{j+1}-m_h^j\right)\right\|_{L^2}^2+C\left\|\widehat{\Delta_jW_h}\right\|_{W^{1,\infty}}^2. ∣ I 4 ∣ ≤ 12 κ ∂ x ( m h j + 1 − m h j ) L 2 2 + C Δ j W h W 1 , ∞ 2 .
Combining these estimates and absorbing the corresponding terms into the right-hand side of (36) , noting (37) , yields
κ 2 ( ∥ ∂ x m h j + 1 ∥ L 2 2 − ∥ ∂ x m h j ∥ L 2 2 ) + κ 4 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 + α κ 2 2 k ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ≤ κ ⟨ ∂ x I h ( m h j × Δ j W h ^ ) , ∂ x m h j ⟩ + C ∥ Δ j W h ^ ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) . \frac{\kappa}{2}\left(\left\|\partial_xm_h^{j+1}\right\|_{L^2}^2-\left\|\partial_xm_h^j\right\|_{L^2}^2\right)+\frac{\kappa}{4}\left\|\partial_x\left(m_h^{j+1}-m_h^j\right)\right\|_{L^2}^2+\frac{\alpha\kappa^2}{2}k\left\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\right\|_h^2
\le\kappa\left\langle\partial_xI_h\left(m_h^j\times\widehat{\Delta_jW_h}\right),\partial_xm_h^j\right\rangle+C\left\|\widehat{\Delta_jW_h}\right\|_{W^{1,\infty}}^2\left(1+\left\|\partial_xm_h^j\right\|_{L^2}^2\right). 2 κ ( ∂ x m h j + 1 L 2 2 − ∂ x m h j L 2 2 ) + 4 κ ∂ x ( m h j + 1 − m h j ) L 2 2 + 2 α κ 2 k m h j + 2 1 × Δ h m h j + 1 h 2 ≤ κ ⟨ ∂ x I h ( m h j × Δ j W h ) , ∂ x m h j ⟩ + C Δ j W h W 1 , ∞ 2 ( 1 + ∂ x m h j L 2 2 ) .
Summing up to an arbitrary time level, taking the maximum over the partial sums, and then taking the p p p -th power gives
max 0 ≤ r ≤ J ∥ ∂ x m h r ∥ L 2 2 p + ( ∑ j = 0 J − 1 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 ) p + ( k ∑ j = 0 J − 1 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ) p ≤ C p ∥ ∂ x m h 0 ∥ L 2 2 p + C p max 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 k ⟨ ∂ x I h ( m h j × Δ j W h ^ ) , ∂ x m h j ⟩ ∣ p + C p ( ∑ j = 0 J − 1 ∥ Δ j W h ^ ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) ) p . \begin{aligned}
\max_{0\le r\le J}\|\partial_xm_h^r\|_{L^2}^{2p}+\left(\sum_{j=0}^{J-1}\|\partial_x(m_h^{j+1}-m_h^j)\|_{L^2}^2\right)^p+\left(k\sum_{j=0}^{J-1}\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\|_h^2\right)^p \\
&\le C_p\|\partial_xm_h^0\|_{L^2}^{2p}+C_p\max_{0\le r\le J}\left|\sum_{j=0}^{r-1}k\langle\partial_xI_h(m_h^j\times\widehat{\Delta_jW_h}),\partial_xm_h^j\rangle\right|^p \\
&+C_p\left(\sum_{j=0}^{J-1}\|\widehat{\Delta_jW_h}\|_{W^{1,\infty}}^2\left(1+\|\partial_xm_h^j\|_{L^2}^2\right)\right)^p.
\end{aligned} 0 ≤ r ≤ J max ∥ ∂ x m h r ∥ L 2 2 p + ( j = 0 ∑ J − 1 ∥ ∂ x ( m h j + 1 − m h j ) ∥ L 2 2 ) p + ( k j = 0 ∑ J − 1 ∥ m h j + 2 1 × Δ h m h j + 1 ∥ h 2 ) p ≤ C p ∥ ∂ x m h 0 ∥ L 2 2 p + C p 0 ≤ r ≤ J max j = 0 ∑ r − 1 k ⟨ ∂ x I h ( m h j × Δ j W h ) , ∂ x m h j ⟩ p + C p ( j = 0 ∑ J − 1 ∥ Δ j W h ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) ) p .
After taking expectations, Lemma [3 ] controls the martingale term. Moreover, Lemma A.1, the independence of Δ j W h ^ \widehat{\Delta_jW_h} Δ j W h from F t j \mathcal{F}_{t_j} F t j , and (28) imply
E [ ( ∑ j = 0 J − 1 ∥ Δ j W h ^ ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) ) p ] ≤ C p , T , Γ ( 1 + ∑ j = 0 J − 1 k E [ ∥ ∂ x m h j ∥ L 2 2 p ] ) . \mathbb{E}\left[\left(\sum_{j=0}^{J-1}\|\widehat{\Delta_jW_h}\|_{W^{1,\infty}}^2\left(1+\|\partial_xm_h^j\|_{L^2}^2\right)\right)^p\right]\le C_{p,T,\Gamma}\left(1+\sum_{j=0}^{J-1}k\mathbb{E}\left[\|\partial_xm_h^j\|_{L^2}^{2p}\right]\right). E [ ( j = 0 ∑ J − 1 ∥ Δ j W h ∥ W 1 , ∞ 2 ( 1 + ∥ ∂ x m h j ∥ L 2 2 ) ) p ] ≤ C p , T , Γ ( 1 + j = 0 ∑ J − 1 k E [ ∥ ∂ x m h j ∥ L 2 2 p ] ) .
An application of the discrete Gronwall lemma then proves (34) .
The following discrete increment estimates will be essential for our analysis.
Lemma 3.5. Assume (6) holds. Then there exists a constant C > 0 C>0 C > 0 , independent of h h h and k k k , such that
E [ ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ h 2 ] + 1 k E [ ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ h 4 ] ≤ C . (38) \mathbb{E}\left[\sum_{j=0}^{J-1}\|m_h^{j+1}-m_h^j\|_h^2\right]+\frac{1}{k}\mathbb{E}\left[\sum_{j=0}^{J-1}\|m_h^{j+1}-m_h^j\|_h^4\right]\le C. \tag*{(38)} E [ j = 0 ∑ J − 1 ∥ m h j + 1 − m h j ∥ h 2 ] + k 1 E [ j = 0 ∑ J − 1 ∥ m h j + 1 − m h j ∥ h 4 ] ≤ C . (38)
Consequently,
∑ j = 0 J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 1 2 ≤ C . (39) \sum_{j=0}^{J-1}\left(\mathbb{E}\left[\|m_h^{j+1}-m_h^j\|_h^4\right]\right)^{\frac12}\le C. \tag*{(39)} j = 0 ∑ J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 2 1 ≤ C . (39)
Proof. Our starting point is the estimate (35) . Taking expectations, using (34) with p = 2 p=2 p = 2 and Jensen’s inequality for the first term, and using (28) with q = 1 q=1 q = 1 for the second term, we obtain
E [ ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ h 2 ] ≤ C k + C ∑ j = 0 J − 1 k ≤ C , \mathbb{E}\left[\sum_{j=0}^{J-1}\|m_h^{j+1}-m_h^j\|_h^2\right]\le Ck+C\sum_{j=0}^{J-1}k\le C, E [ j = 0 ∑ J − 1 ∥ m h j + 1 − m h j ∥ h 2 ] ≤ C k + C j = 0 ∑ J − 1 k ≤ C ,
thus proving the estimate for the first term in (38) .
Next, raising (35) to the fourth power gives
∥ m h j + 1 − m h j ∥ h 4 ≤ C k 4 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 4 + C ∥ Δ j W h ^ ∥ L ∞ 4 . (40) \|m_h^{j+1}-m_h^j\|_h^4\le Ck^4\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\|_h^4+C\|\widehat{\Delta_jW_h}\|_{L^\infty}^4. \tag*{(40)} ∥ m h j + 1 − m h j ∥ h 4 ≤ C k 4 ∥ m h j + 2 1 × Δ h m h j + 1 ∥ h 4 + C ∥ Δ j W h ∥ L ∞ 4 . (40)
Summing from j = 0 j=0 j = 0 to J − 1 J-1 J − 1 , taking expectations, and using an elementary inequality
k 2 ∑ j = 0 J − 1 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 4 ≤ ( k ∑ j = 0 J − 1 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ) 2 , k^2\sum_{j=0}^{J-1}\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\|_h^4\le\left(k\sum_{j=0}^{J-1}\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\|_h^2\right)^2, k 2 j = 0 ∑ J − 1 ∥ m h j + 2 1 × Δ h m h j + 1 ∥ h 4 ≤ ( k j = 0 ∑ J − 1 ∥ m h j + 2 1 × Δ h m h j + 1 ∥ h 2 ) 2 ,
we obtain
E [ ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ h 4 ] ≤ C k 2 E [ ( k ∑ j = 0 J − 1 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ) 2 ] + C ∑ j = 0 J − 1 E [ ∥ Δ j W h ^ ∥ L ∞ 4 ] . ≤ C k 2 + C ∑ j = 0 J − 1 k 2 ≤ C k . (41) \begin{aligned}
\mathbb{E}\left[\sum_{j=0}^{J-1}\|m_h^{j+1}-m_h^j\|_h^4\right]\le Ck^2\mathbb{E}\left[\left(k\sum_{j=0}^{J-1}\|m_h^{j+\frac12}\times\Delta_hm_h^{j+1}\|_h^2\right)^2\right]+C\sum_{j=0}^{J-1}\mathbb{E}\left[\|\widehat{\Delta_jW_h}\|_{L^\infty}^4\right]. \\
&\leq C k^2 + C \sum_{j=0}^{J-1} k^2 \leq C k. \tag*{(41)}
\end{aligned} E [ j = 0 ∑ J − 1 ∥ m h j + 1 − m h j ∥ h 4 ] ≤ C k 2 E ( k j = 0 ∑ J − 1 ∥ m h j + 2 1 × Δ h m h j + 1 ∥ h 2 ) 2 + C j = 0 ∑ J − 1 E [ ∥ Δ j W h ∥ L ∞ 4 ] . ≤ C k 2 + C j = 0 ∑ J − 1 k 2 ≤ C k . (41)
where we also used the moment bound (34) with p = 2 p=2 p = 2 and (28) with q = 2 q=2 q = 2 . This proves (38) .
Finally, by Cauchy’s inequality in the time index,
∑ j = 0 J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 1 2 ≤ J 1 2 ( ∑ j = 0 J − 1 E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 1 2 ≤ ( T k ) 1 2 ( C k ) 1 2 ≤ C . \sum_{j=0}^{J-1} \left(\mathbb{E}\left[\left\|m_h^{j+1}-m_h^j\right\|_h^4\right]\right)^{\frac{1}{2}}
\leq J^{\frac{1}{2}}\left(\sum_{j=0}^{J-1}\mathbb{E}\left[\left\|m_h^{j+1}-m_h^j\right\|_h^4\right]\right)^{\frac{1}{2}}
\leq\left(\frac{T}{k}\right)^{\frac{1}{2}}(Ck)^{\frac{1}{2}}\leq C. j = 0 ∑ J − 1 ( E [ m h j + 1 − m h j h 4 ] ) 2 1 ≤ J 2 1 ( j = 0 ∑ J − 1 E [ m h j + 1 − m h j h 4 ] ) 2 1 ≤ ( k T ) 2 1 ( C k ) 2 1 ≤ C .
This proves (39) and completes the proof of the lemma. □
Localised error estimate
In this subsection, we prove a localised a priori error estimate in the natural energy norm. For brevity, we write
m j : = m ( t j ) , m ‾ j + 1 2 : = 1 2 ( m j + 1 + m j ) , j = 0 , … , J − 1. m^j:=m(t_j),\qquad\overline{m}^{j+\frac{1}{2}}:=\frac{1}{2}(m^{j+1}+m^j),\qquad j=0,\ldots,J-1. m j := m ( t j ) , m j + 2 1 := 2 1 ( m j + 1 + m j ) , j = 0 , … , J − 1.
For j = 0 , … , J j=0,\ldots,J j = 0 , … , J , define
ρ h j : = m j − I h m j , θ h j : = I h m j − m h j , (42) \rho_h^j:=m^j-I_hm^j,\qquad\theta_h^j:=I_hm^j-m_h^j, \tag*{(42)} ρ h j := m j − I h m j , θ h j := I h m j − m h j , (42)
so that m j − m h j = ρ h j + θ h j m^j-m_h^j=\rho_h^j+\theta_h^j m j − m h j = ρ h j + θ h j . By (16) , the interpolation error ρ h j \rho_h^j ρ h j satisfies
max 0 ≤ j ≤ J ∥ ρ h j ∥ L 2 2 + k ∑ j = 0 J − 1 ∥ ∂ x ρ h j + 1 ∥ L 2 2 ≤ C h 2 ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 2 . (43) \max_{0\leq j\leq J}\|\rho_h^j\|_{L^2}^2+k\sum_{j=0}^{J-1}\|\partial_x\rho_h^{j+1}\|_{L^2}^2\leq Ch^2\|m\|_{L^\infty(0,T;H^2)}^2. \tag*{(43)} 0 ≤ j ≤ J max ∥ ρ h j ∥ L 2 2 + k j = 0 ∑ J − 1 ∥ ∂ x ρ h j + 1 ∥ L 2 2 ≤ C h 2 ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 2 . (43)
To derive the error equation, we first compare the deterministic drift of the exact solution with the corresponding discrete drift evaluated at the projected exact solution. For j = 0 , … , J − 1 j=0,\ldots,J-1 j = 0 , … , J − 1 , define the drift residual R h j + 1 ∈ V h ′ \mathcal{R}_h^{j+1}\in V_h' R h j + 1 ∈ V h ′ by
R h j + 1 : = R h , d j + 1 + R h , p j + 1 , (44) \mathcal{R}_h^{j+1}:=\mathcal{R}_{h,d}^{j+1}+\mathcal{R}_{h,p}^{j+1}, \tag*{(44)} R h j + 1 := R h , d j + 1 + R h , p j + 1 , (44)
where the damping residual R h , d j + 1 \mathcal{R}_{h,d}^{j+1} R h , d j + 1 measures the difference between the discrete midpoint approximation of the damping drift and its exact time integral:
R h , d j + 1 ( ϕ h ) : = α κ k ⟨ I h m ‾ j + 1 2 × [ I h m ‾ j + 1 2 × Δ h I h m j + 1 ] , ϕ h ⟩ h − α κ ∫ t j t j + 1 ⟨ I h [ m ( s ) × ( m ( s ) × ∂ x x m ( s ) ) ] , ϕ h ⟩ h d s , \mathcal{R}_{h,d}^{j+1}(\phi_h):=\alpha\kappa k\left\langle I_h\overline{m}^{j+\frac{1}{2}}\times\left[I_h\overline{m}^{j+\frac{1}{2}}\times\Delta_hI_hm^{j+1}\right],\phi_h\right\rangle_h
-\alpha\kappa\int_{t_j}^{t_{j+1}}\left\langle I_h\left[m(s)\times\left(m(s)\times\partial_{xx}m(s)\right)\right],\phi_h\right\rangle_h\,\mathrm{d}s, R h , d j + 1 ( ϕ h ) := α κk ⟨ I h m j + 2 1 × [ I h m j + 2 1 × Δ h I h m j + 1 ] , ϕ h ⟩ h − α κ ∫ t j t j + 1 ⟨ I h [ m ( s ) × ( m ( s ) × ∂ xx m ( s ) ) ] , ϕ h ⟩ h d s ,
and the precession residual R h , p j + 1 \mathcal{R}_{h,p}^{j+1} R h , p j + 1 is defined analogously by
R h , p j + 1 ( ϕ h ) : = − κ k ⟨ I h m ‾ j + 1 2 × Δ h I h m j + 1 , ϕ h ⟩ h + κ ∫ t j t j + 1 ⟨ I h [ m ( s ) × ∂ x x m ( s ) ] , ϕ h ⟩ h d s . \mathcal{R}_{h,p}^{j+1}(\phi_h):=-\kappa k\left\langle I_h\overline{m}^{j+\frac{1}{2}}\times\Delta_hI_hm^{j+1},\phi_h\right\rangle_h
+\kappa\int_{t_j}^{t_{j+1}}\left\langle I_h\left[m(s)\times\partial_{xx}m(s)\right],\phi_h\right\rangle_h\,\mathrm{d}s. R h , p j + 1 ( ϕ h ) := − κk ⟨ I h m j + 2 1 × Δ h I h m j + 1 , ϕ h ⟩ h + κ ∫ t j t j + 1 ⟨ I h [ m ( s ) × ∂ xx m ( s ) ] , ϕ h ⟩ h d s .
Note that since I h m ( s ) I_hm(s) I h m ( s ) and m ( s ) m(s) m ( s ) , as well as g ℓ , h g_{\ell,h} g ℓ , h and g ℓ g_\ell g ℓ , agree at every finite element node, the mass-lumped inner product gives ⟨ I h m ( s ) × g ℓ , h , ϕ h ⟩ h = ⟨ m ( s ) × g ℓ , ϕ h ⟩ h \langle I_hm(s)\times g_{\ell,h},\phi_h\rangle_h=\langle m(s)\times g_\ell,\phi_h\rangle_h ⟨ I h m ( s ) × g ℓ , h , ϕ h ⟩ h = ⟨ m ( s ) × g ℓ , ϕ h ⟩ h . Hence, after nodal projection, the stochastic term may equivalently be written using W h W_h W h . Integrating the exact equation over [ t j , t j + 1 ] [t_j,t_{j+1}] [ t j , t j + 1 ] , applying the nodal interpolant I h I_h I h , then adding and subtracting the two discrete drift terms appearing above shows that the projected exact solution I h m j I_hm^j I h m j satisfies
⟨ I h m j + 1 − I h m j , ϕ h ⟩ h + α κ k ⟨ I h m ‾ j + 1 2 × [ I h m ‾ j + 1 2 × Δ h I h m j + 1 ] , ϕ h ⟩ h − κ k ⟨ I h m ‾ j + 1 2 × Δ h I h m j + 1 , ϕ h ⟩ h − ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) , ϕ h ⟩ h = R h j + 1 ( ϕ h ) . (45) \left\langle I_hm^{j+1}-I_hm^j,\phi_h\right\rangle_h
+\alpha\kappa k\left\langle I_h\overline{m}^{j+\frac{1}{2}}\times\left[I_h\overline{m}^{j+\frac{1}{2}}\times\Delta_hI_hm^{j+1}\right],\phi_h\right\rangle_h
-\kappa k\left\langle I_h\overline{m}^{j+\frac{1}{2}}\times\Delta_hI_hm^{j+1},\phi_h\right\rangle_h
-\left\langle\int_{t_j}^{t_{j+1}}I_hm(s)\times\circ\mathrm{d}W_h(s),\phi_h\right\rangle_h
=\mathcal{R}_h^{j+1}(\phi_h). \tag*{(45)} ⟨ I h m j + 1 − I h m j , ϕ h ⟩ h + α κk ⟨ I h m j + 2 1 × [ I h m j + 2 1 × Δ h I h m j + 1 ] , ϕ h ⟩ h − κk ⟨ I h m j + 2 1 × Δ h I h m j + 1 , ϕ h ⟩ h − ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) , ϕ h ⟩ h = R h j + 1 ( ϕ h ) . (45)
Therefore, (45) has the same deterministic structure as the scheme (29) , up to the drift residual R h j + 1 \mathcal{R}_h^{j+1} R h j + 1 . Subtracting the numerical scheme thus yields the discrete error equation directly. The stochastic term is kept in its exact integral form, since its midpoint consistency error requires a more delicate argument. Here, R h j + 1 \mathcal{R}_h^{j+1} R h j + 1 contains only the drift consistency error, while the stochastic part is treated separately in Lemma 3.7.
The following lemma estimates this drift residual in the dual norm (10) . This is the estimate needed when R h j + 1 \mathcal{R}_h^{j+1} R h j + 1 is paired with the error θ h j + 1 \theta_h^{j+1} θ h j + 1 in the localised energy argument later.
Lemma 3.6 (Drift consistency). Let γ ∈ ( 0 , 1 2 ) \gamma\in\left(0,\frac{1}{2}\right) γ ∈ ( 0 , 2 1 ) , and let R h j + 1 \mathcal{R}_h^{j+1} R h j + 1 be the drift residual defined by (44) . Then
E [ ∑ j = 0 J − 1 1 k ∥ R h j + 1 ∥ − 1 , h 2 ] ≤ C γ ( h 2 + k 2 γ ) . (46) \mathbb{E}\left[\sum_{j=0}^{J-1}\frac{1}{k}\left\|\mathcal{R}_h^{j+1}\right\|_{-1,h}^{2}\right] \le C_\gamma\left(h^2+k^{2\gamma}\right). \tag*{(46)} E [ j = 0 ∑ J − 1 k 1 R h j + 1 − 1 , h 2 ] ≤ C γ ( h 2 + k 2 γ ) . (46)
Proof. We split
R h , p j + 1 = R ~ h , p j + 1 + Q h , p j + 1 , R h , d j + 1 = R ~ h , d j + 1 + Q h , d j + 1 , \mathcal{R}_{h,p}^{j+1}=\widetilde{\mathcal{R}}_{h,p}^{j+1}+\mathcal{Q}_{h,p}^{j+1},\qquad
\mathcal{R}_{h,d}^{j+1}=\widetilde{\mathcal{R}}_{h,d}^{j+1}+\mathcal{Q}_{h,d}^{j+1}, R h , p j + 1 = R h , p j + 1 + Q h , p j + 1 , R h , d j + 1 = R h , d j + 1 + Q h , d j + 1 ,
where the quadrature residuals
Q h , p j + 1 ( ϕ h ) : = κ ∫ t j t j + 1 [ ⟨ I h ( m ( s ) × ∂ x x m ( s ) ) , ϕ h ⟩ h − ⟨ m ( s ) × ∂ x x m ( s ) , ϕ h ⟩ ] d s , \mathcal{Q}_{h,p}^{j+1}(\phi_h):=\kappa\int_{t_j}^{t_{j+1}}\left[\left\langle I_h\left(m(s)\times\partial_{xx}m(s)\right),\phi_h\right\rangle_h-\left\langle m(s)\times\partial_{xx}m(s),\phi_h\right\rangle\right]\mathrm{d}s, Q h , p j + 1 ( ϕ h ) := κ ∫ t j t j + 1 [ ⟨ I h ( m ( s ) × ∂ xx m ( s ) ) , ϕ h ⟩ h − ⟨ m ( s ) × ∂ xx m ( s ) , ϕ h ⟩ ] d s ,
Q h , d j + 1 ( ϕ h ) : = − α κ ∫ t j t j + 1 [ ⟨ I h [ m ( s ) × ( m ( s ) × ∂ x x m ( s ) ) ] , ϕ h ⟩ h − ⟨ m ( s ) × ( m ( s ) × ∂ x x m ( s ) ) , ϕ h ⟩ ] d s . \mathcal{Q}_{h,d}^{j+1}(\phi_h):=-\alpha\kappa\int_{t_j}^{t_{j+1}}\left[\left\langle I_h\left[m(s)\times\left(m(s)\times\partial_{xx}m(s)\right)\right],\phi_h\right\rangle_h-\left\langle m(s)\times\left(m(s)\times\partial_{xx}m(s)\right),\phi_h\right\rangle\right]\mathrm{d}s. Q h , d j + 1 ( ϕ h ) := − α κ ∫ t j t j + 1 [ ⟨ I h [ m ( s ) × ( m ( s ) × ∂ xx m ( s ) ) ] , ϕ h ⟩ h − ⟨ m ( s ) × ( m ( s ) × ∂ xx m ( s ) ) , ϕ h ⟩ ] d s .
Thus, the tilded residuals R ~ h , p j + 1 \widetilde{\mathcal{R}}_{h,p}^{j+1} R h , p j + 1 and R ~ h , d j + 1 \widetilde{\mathcal{R}}_{h,d}^{j+1} R h , d j + 1 are obtained by replacing the exact mass-lumped pairings by standard L 2 L^2 L 2 -pairings.
We first estimate the precession residual. By the definition of Δ h \Delta_h Δ h and applying integration by parts, we add and subtract a common term to obtain R ~ h , p j + 1 = R h , p , 1 j + 1 + R h , p , 2 j + 1 \widetilde{\mathcal{R}}_{h,p}^{j+1}=\mathcal{R}_{h,p,1}^{j+1}+\mathcal{R}_{h,p,2}^{j+1} R h , p j + 1 = R h , p , 1 j + 1 + R h , p , 2 j + 1 , where
R h , p , 1 j + 1 ( ϕ h ) : = κ k [ ⟨ ∂ x I h m j + 1 , ∂ x I h ( ϕ h × I h m ‾ j + 1 2 ) ⟩ − ⟨ ∂ x m j + 1 , ∂ x ( ϕ h × m ‾ j + 1 2 ) ⟩ ] , \mathcal{R}_{h,p,1}^{j+1}(\phi_h):=\kappa k\left[\left\langle\partial_xI_hm^{j+1},\partial_xI_h\left(\phi_h\times I_h\overline{m}^{j+\frac{1}{2}}\right)\right\rangle-\left\langle\partial_xm^{j+1},\partial_x\left(\phi_h\times\overline{m}^{j+\frac{1}{2}}\right)\right\rangle\right], R h , p , 1 j + 1 ( ϕ h ) := κk [ ⟨ ∂ x I h m j + 1 , ∂ x I h ( ϕ h × I h m j + 2 1 ) ⟩ − ⟨ ∂ x m j + 1 , ∂ x ( ϕ h × m j + 2 1 ) ⟩ ] ,
R h , p , 2 j + 1 ( ϕ h ) : = κ ∫ t j t j + 1 [ ⟨ ∂ x m j + 1 , ∂ x ( ϕ h × m ‾ j + 1 2 ) ⟩ − ⟨ ∂ x m ( s ) , ∂ x ( ϕ h × m ( s ) ) ⟩ ] d s . \mathcal{R}_{h,p,2}^{j+1}(\phi_h):=\kappa\int_{t_j}^{t_{j+1}}\left[\left\langle\partial_xm^{j+1},\partial_x\left(\phi_h\times\overline{m}^{j+\frac{1}{2}}\right)\right\rangle-\left\langle\partial_xm(s),\partial_x\left(\phi_h\times m(s)\right)\right\rangle\right]\mathrm{d}s. R h , p , 2 j + 1 ( ϕ h ) := κ ∫ t j t j + 1 [ ⟨ ∂ x m j + 1 , ∂ x ( ϕ h × m j + 2 1 ) ⟩ − ⟨ ∂ x m ( s ) , ∂ x ( ϕ h × m ( s ) ) ⟩ ] d s .
Using the interpolation estimate, (22) , and (25) , we obtain
∥ R h , p , 1 j + 1 ∥ − 1 , h ≤ C k h ∥ m j + 1 ∥ H 2 ∥ m ‾ j + 1 2 ∥ H 2 . (47) \left\|\mathcal{R}_{h,p,1}^{j+1}\right\|_{-1,h}\le Ckh\left\|m^{j+1}\right\|_{H^2}\left\|\overline{m}^{j+\frac{1}{2}}\right\|_{H^2}. \tag*{(47)} R h , p , 1 j + 1 − 1 , h ≤ C k h m j + 1 H 2 m j + 2 1 H 2 . (47)
Furthermore, the one-dimensional H 1 H^1 H 1 -product estimate gives, for s ∈ [ t j , t j + 1 ] s\in[t_j,t_{j+1}] s ∈ [ t j , t j + 1 ] ,
∣ ⟨ ∂ x m j + 1 , ∂ x ( ϕ h × m ‾ j + 1 2 ) ⟩ − ⟨ ∂ x m ( s ) , ∂ x ( ϕ h × m ( s ) ) ⟩ ∣ ≤ C ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 2 ) ( ∥ m ( s ) − m j + 1 ∥ H 1 + ∥ m ( s ) − m ‾ j + 1 2 ∥ H 1 ) ∥ ϕ h ∥ 1 , h . \left|\left\langle\partial_xm^{j+1},\partial_x\left(\phi_h\times\overline{m}^{j+\frac{1}{2}}\right)\right\rangle-\left\langle\partial_xm(s),\partial_x\left(\phi_h\times m(s)\right)\right\rangle\right|
\le C\left(1+\left\|m\right\|_{L^\infty(0,T;H^2)}^2\right)\left(\left\|m(s)-m^{j+1}\right\|_{H^1}+\left\|m(s)-\overline{m}^{j+\frac{1}{2}}\right\|_{H^1}\right)\left\|\phi_h\right\|_{1,h}. ⟨ ∂ x m j + 1 , ∂ x ( ϕ h × m j + 2 1 ) ⟩ − ⟨ ∂ x m ( s ) , ∂ x ( ϕ h × m ( s ) ) ⟩ ≤ C ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 2 ) ( m ( s ) − m j + 1 H 1 + m ( s ) − m j + 2 1 H 1 ) ∥ ϕ h ∥ 1 , h .
Consequently,
∥ R h , p , 2 j + 1 ∥ − 1 , h ≤ C k 1 + γ ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 2 ) ∥ m ∥ C γ ( [ 0 , T ] ; H 1 ) . (48) \left\|\mathcal{R}_{h,p,2}^{j+1}\right\|_{-1,h}\le Ck^{1+\gamma}\left(1+\left\|m\right\|_{L^\infty(0,T;H^2)}^2\right)\left\|m\right\|_{C^\gamma([0,T];H^1)}. \tag*{(48)} R h , p , 2 j + 1 − 1 , h ≤ C k 1 + γ ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 2 ) ∥ m ∥ C γ ([ 0 , T ] ; H 1 ) . (48)
Hence, from (47) and (48) , we get
E [ ∑ j = 0 J − 1 1 k ∥ R ~ h , p j + 1 ∥ − 1 , h 2 ] ≤ C γ ( h 2 + k 2 γ ) . (49) \mathbb{E}\left[\sum_{j=0}^{J-1}\frac{1}{k}\left\|\widetilde{\mathcal{R}}_{h,p}^{j+1}\right\|_{-1,h}^{2}\right]\le C_\gamma\left(h^2+k^{2\gamma}\right). \tag*{(49)} E [ j = 0 ∑ J − 1 k 1 R h , p j + 1 − 1 , h 2 ] ≤ C γ ( h 2 + k 2 γ ) . (49)
To estimate R ~ h , d j + 1 \widetilde{\mathcal{R}}_{h,d}^{j+1} R h , d j + 1 , we observe that by the scalar triple-product identity and the definition of Δ h \Delta_h Δ h ,
⟨ I h m ‾ j + 1 2 × [ I h m ‾ j + 1 2 × Δ h I h m ‾ j + 1 ] , ϕ h ⟩ h = − ⟨ ∂ x I h m j + 1 , ∂ x I h [ ( ϕ h × I h m ‾ j + 1 2 ) × I h m ‾ j + 1 2 ] ⟩ . \left\langle I_h\overline{m}^{j+\frac{1}{2}}\times\left[I_h\overline{m}^{j+\frac{1}{2}}\times\Delta_h I_h\overline{m}^{j+1}\right],\phi_h\right\rangle_h=-\left\langle\partial_xI_hm^{j+1},\partial_xI_h\left[\left(\phi_h\times I_h\overline{m}^{j+\frac{1}{2}}\right)\times I_h\overline{m}^{j+\frac{1}{2}}\right]\right\rangle. ⟨ I h m j + 2 1 × [ I h m j + 2 1 × Δ h I h m j + 1 ] , ϕ h ⟩ h = − ⟨ ∂ x I h m j + 1 , ∂ x I h [ ( ϕ h × I h m j + 2 1 ) × I h m j + 2 1 ] ⟩ .
Proceeding similarly as before, but using (26) , yields a decomposition R ~ h , d j + 1 = R h , d , 1 j + 1 + R h , d , 2 j + 1 \widetilde{\mathcal{R}}_{h,d}^{j+1}=\mathcal{R}_{h,d,1}^{j+1}+\mathcal{R}_{h,d,2}^{j+1} R h , d j + 1 = R h , d , 1 j + 1 + R h , d , 2 j + 1 with
∥ R h , d , 1 j + 1 ∥ − 1 , h ≤ C k h ∥ m j + 1 ∥ H 2 ∥ m ‾ j + 1 2 ∥ H 2 2 , \left\|\mathcal{R}_{h,d,1}^{j+1}\right\|_{-1,h}\leq Ckh\left\|m^{j+1}\right\|_{H^2}\left\|\overline{m}^{j+\frac{1}{2}}\right\|_{H^2}^{2}, R h , d , 1 j + 1 − 1 , h ≤ C k h m j + 1 H 2 m j + 2 1 H 2 2 ,
∥ R h , d , 2 j + 1 ∥ − 1 , h ≤ C k 1 + γ ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 3 ) ∥ m ∥ C γ ( [ 0 , T ] ; H 1 ) . \left\|\mathcal{R}_{h,d,2}^{j+1}\right\|_{-1,h}\leq Ck^{1+\gamma}\left(1+\|m\|_{L^\infty(0,T;H^2)}^3\right)\|m\|_{C^\gamma([0,T];H^1)}. R h , d , 2 j + 1 − 1 , h ≤ C k 1 + γ ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 2 ) 3 ) ∥ m ∥ C γ ([ 0 , T ] ; H 1 ) .
Therefore,
E [ ∑ j = 0 J − 1 1 k ∥ R ~ h , d j + 1 ∥ − 1 , h 2 ] ≤ C γ ( h 2 + k 2 γ ) . (50) \mathbb{E}\left[\sum_{j=0}^{J-1}\frac{1}{k}\left\|\widetilde{\mathcal{R}}_{h,d}^{j+1}\right\|_{-1,h}^{2}\right]\leq C_\gamma\left(h^2+k^{2\gamma}\right). \tag*{(50)} E [ j = 0 ∑ J − 1 k 1 R h , d j + 1 − 1 , h 2 ] ≤ C γ ( h 2 + k 2 γ ) . (50)
It remains to estimate the quadrature residuals Q h , p Q_{h,p} Q h , p and Q h , d Q_{h,d} Q h , d . By (20) , we have
∥ Q h , p j + 1 ∥ − 1 , h + ∥ Q h , d j + 1 ∥ − 1 , h ≤ C h ∫ t j t j + 1 ( ∥ m ( s ) × ∂ x x m ( s ) ∥ H 1 + ∥ m ( s ) × ( m ( s ) × ∂ x x m ( s ) ) ∥ H 1 ) d s ≤ C k h ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 3 ) 3 ) , \begin{aligned}
\left\|Q_{h,p}^{j+1}\right\|_{-1,h}+\left\|Q_{h,d}^{j+1}\right\|_{-1,h}\leq Ch\int_{t_j}^{t_{j+1}}\left(\|m(s)\times\partial_{xx}m(s)\|_{H^1}+\|m(s)\times(m(s)\times\partial_{xx}m(s))\|_{H^1}\right)\,\mathrm{d}s \\
&\leq Ckh\left(1+\|m\|_{L^\infty(0,T;H^3)}^3\right),
\end{aligned} Q h , p j + 1 − 1 , h + Q h , d j + 1 − 1 , h ≤ C h ∫ t j t j + 1 ( ∥ m ( s ) × ∂ xx m ( s ) ∥ H 1 + ∥ m ( s ) × ( m ( s ) × ∂ xx m ( s )) ∥ H 1 ) d s ≤ C k h ( 1 + ∥ m ∥ L ∞ ( 0 , T ; H 3 ) 3 ) ,
where in the last step we used the one-dimensional H 1 H^1 H 1 -product estimate and the regularity of m m m . This implies
E [ ∑ j = 0 J − 1 1 k ( ∥ Q h , p j + 1 ∥ − 1 , h 2 + ∥ Q h , d j + 1 ∥ − 1 , h 2 ) ] ≤ C h 2 . (51) \mathbb{E}\left[\sum_{j=0}^{J-1}\frac{1}{k}\left(\left\|Q_{h,p}^{j+1}\right\|_{-1,h}^{2}+\left\|Q_{h,d}^{j+1}\right\|_{-1,h}^{2}\right)\right]\leq Ch^2. \tag*{(51)} E [ j = 0 ∑ J − 1 k 1 ( Q h , p j + 1 − 1 , h 2 + Q h , d j + 1 − 1 , h 2 ) ] ≤ C h 2 . (51)
Combining (49) , (50) , and (51) proves (46) , as required. □
For each Λ > 0 \Lambda>0 Λ > 0 and j = 0 , … , J j=0,\ldots,J j = 0 , … , J , define a subset
Ω Λ , j : = { max 0 ≤ r ≤ j ∥ ∂ x m h r ∥ L 2 4 + sup 0 ≤ t ≤ t j ∥ m ( t ) ∥ H 2 4 ≤ Λ } . (52) \Omega_{\Lambda,j}:=\left\{\max_{0\leq r\leq j}\|\partial_xm_h^r\|_{L^2}^{4}+\sup_{0\leq t\leq t_j}\|m(t)\|_{H^2}^{4}\leq\Lambda\right\}. \tag*{(52)} Ω Λ , j := { 0 ≤ r ≤ j max ∥ ∂ x m h r ∥ L 2 4 + 0 ≤ t ≤ t j sup ∥ m ( t ) ∥ H 2 4 ≤ Λ } . (52)
Then, Ω Λ , j + 1 ⊂ Ω Λ , j \Omega_{\Lambda,j+1}\subset\Omega_{\Lambda,j} Ω Λ , j + 1 ⊂ Ω Λ , j and 1 Ω Λ , j \mathbf{1}_{\Omega_{\Lambda,j}} 1 Ω Λ , j is F t j \mathcal{F}_{t_j} F t j -measurable. Moreover, by (7) and (34) , there exists a constant C > 0 C>0 C > 0 independent of h h h , k k k , and Λ \Lambda Λ such that
P [ Ω Λ , J c ] ≤ C Λ . (53) \mathbb{P}\left[\Omega_{\Lambda,J}^{c}\right]\leq\frac{C}{\Lambda}. \tag*{(53)} P [ Ω Λ , J c ] ≤ Λ C . (53)
The consistency error in the stochastic part is handled by the following technical lemma.
Lemma 3.7 (Stochastic error term). Let ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) and Λ > 0 \Lambda>0 Λ > 0 be fixed. Let m m m be the strong solution of (1) and { m h j } j = 0 J \{m_h^j\}_{j=0}^{J} { m h j } j = 0 J be an adapted solution of (30) . Let W h W_h W h be given by (27) . For r = 0 , 1 , … , J r=0,1,\ldots,J r = 0 , 1 , … , J , define
S r : = ∑ j = 0 r − 1 1 Ω Λ , j [ ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) , θ h j + 1 ⟩ h − ⟨ m h j + 1 2 × Δ ^ j W h , θ h j + 1 ⟩ h ] . (54) S_r:=\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left[\left\langle\int_{t_j}^{t_{j+1}}I_hm(s)\times\circ\mathrm{d}W_h(s),\theta_h^{j+1}\right\rangle_h-\left\langle m_h^{j+\frac{1}{2}}\times\widehat{\Delta}_jW_h,\theta_h^{j+1}\right\rangle_h\right]. \tag*{(54)} S r := j = 0 ∑ r − 1 1 Ω Λ , j [ ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) , θ h j + 1 ⟩ h − ⟨ m h j + 2 1 × Δ j W h , θ h j + 1 ⟩ h ] . (54)
Then, for every δ > 0 \delta>0 δ > 0 and η > 0 \eta>0 η > 0 , there exists a constant C δ , η , ε , Λ , T > 0 C_{\delta,\eta,\varepsilon,\Lambda,T}>0 C δ , η , ε , Λ , T > 0 , independent of h h h and k k k , such that
E [ sup 0 ≤ r ≤ J ∣ S r ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + η E [ sup 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η , ε , Λ , T ( k 1 − ε + k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] ) . (55) \begin{aligned}
\mathbb{E}\left[\sup_{0\leq r\leq J}|S_r|\right]\leq\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+\eta\mathbb{E}\left[\sup_{0\leq j\leq J}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\right] \\
&+ C_{\delta,\eta,\varepsilon,\Lambda,T}\left(k^{1-\varepsilon}+k\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right]\right). \tag*{(55)}
\end{aligned} E [ 0 ≤ r ≤ J sup ∣ S r ∣ ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + η E [ 0 ≤ j ≤ J sup 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η , ε , Λ , T ( k 1 − ε + k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] ) . (55)
Proof. We begin the proof with several observations. First, note that applying (7) with the Hölder exponent 1 2 ( 1 − ε ) ∈ ( 0 , 1 2 ) \frac{1}{2}(1-\varepsilon)\in\left(0,\frac{1}{2}\right) 2 1 ( 1 − ε ) ∈ ( 0 , 2 1 ) gives
∑ j = 0 J − 1 ∫ t j t j + 1 E [ ∥ m ( s ) − m j ∥ H 1 2 + ∥ m ( s ) − m j + 1 ∥ H 1 2 ] d s ≤ C ε k 1 − ε . (56) \sum_{j=0}^{J-1}\int_{t_j}^{t_{j+1}}\mathbb{E}\left[\left\|m(s)-m^j\right\|_{H^1}^2+\left\|m(s)-m^{j+1}\right\|_{H^1}^2\right]\mathrm{d}s\leq C_\varepsilon k^{1-\varepsilon}. \tag*{(56)} j = 0 ∑ J − 1 ∫ t j t j + 1 E [ m ( s ) − m j H 1 2 + m ( s ) − m j + 1 H 1 2 ] d s ≤ C ε k 1 − ε . (56)
Next, the Stratonovich integral is rewritten as a vector-valued Itô integral to obtain
⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) , θ h j + 1 ⟩ h = ⟨ ∑ ℓ = 1 ∞ ∫ t j t j + 1 I h m ( s ) × g ℓ , h d β ℓ ( s ) , θ h j + 1 ⟩ h + 1 2 ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ ( I h m ( s ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s . (57) \left\langle\int_{t_j}^{t_{j+1}}I_hm(s)\times\circ\mathrm{d}W_h(s),\theta_h^{j+1}\right\rangle_h
=
\left\langle\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}I_hm(s)\times g_{\ell,h}\,\mathrm{d}\beta_\ell(s),\theta_h^{j+1}\right\rangle_h
+\frac{1}{2}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle\left(I_hm(s)\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\mathrm{d}s. \tag*{(57)} ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) , θ h j + 1 ⟩ h = ⟨ ℓ = 1 ∑ ∞ ∫ t j t j + 1 I h m ( s ) × g ℓ , h d β ℓ ( s ) , θ h j + 1 ⟩ h + 2 1 ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ ( I h m ( s ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s . (57)
On the other hand,
⟨ m h j + 1 2 × Δ ^ j W h , θ h j + 1 ⟩ h = ∑ ℓ = 1 ∞ Δ ^ j β ℓ ⟨ m h j × g ℓ , h , θ h j + 1 ⟩ h + 1 2 ∑ ℓ = 1 ∞ Δ ^ j β ℓ ⟨ ( m h j + 1 − m h j ) × g ℓ , h , θ h j + 1 ⟩ h . (58) \left\langle m_h^{j+\frac{1}{2}}\times\widehat{\Delta}_jW_h,\theta_h^{j+1}\right\rangle_h
=
\sum_{\ell=1}^{\infty}\widehat{\Delta}_j\beta_\ell\left\langle m_h^j\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h
+\frac{1}{2}\sum_{\ell=1}^{\infty}\widehat{\Delta}_j\beta_\ell\left\langle\left(m_h^{j+1}-m_h^j\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h. \tag*{(58)} ⟨ m h j + 2 1 × Δ j W h , θ h j + 1 ⟩ h = ℓ = 1 ∑ ∞ Δ j β ℓ ⟨ m h j × g ℓ , h , θ h j + 1 ⟩ h + 2 1 ℓ = 1 ∑ ∞ Δ j β ℓ ⟨ ( m h j + 1 − m h j ) × g ℓ , h , θ h j + 1 ⟩ h . (58)
We aim rewrite the last term in (58) . The fully discrete scheme (30) implies the following identity, which holds at every node:
m h j + 1 − m h j = − α κ k m h j + 1 2 × ( m h j + 1 2 × Δ h m h j + 1 ) + κ k m h j + 1 2 × Δ h m h j + 1 + m h j + 1 2 × Δ ^ j W h . m_h^{j+1}-m_h^j=-\alpha\kappa k\,m_h^{j+\frac{1}{2}}\times\left(m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right)+\kappa k\,m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}+m_h^{j+\frac{1}{2}}\times\widehat{\Delta}_jW_h. m h j + 1 − m h j = − α κk m h j + 2 1 × ( m h j + 2 1 × Δ h m h j + 1 ) + κk m h j + 2 1 × Δ h m h j + 1 + m h j + 2 1 × Δ j W h .
Therefore, using this in the last mass-lumped inner product term in (58) gives
1 2 ∑ ℓ = 1 ∞ Δ ^ j β ℓ ⟨ ( m h j + 1 − m h j ) × g ℓ , h , θ h j + 1 ⟩ h = k 2 ∑ ℓ = 1 ∞ ⟨ ( m h j + 1 2 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h + 1 2 ∑ ℓ = 1 ∞ ( ∣ Δ ^ j β ℓ ∣ 2 − k ) ⟨ ( m h j + 1 2 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h + κ k 2 ∑ ℓ = 1 ∞ Δ ^ j β ℓ ⟨ ( m h j + 1 2 × Δ h m h j + 1 − α m h j + 1 2 × ( m h j + 1 2 × Δ h m h j + 1 ) ) × g ℓ , h , θ h j + 1 ⟩ h + 1 2 ∑ ℓ = 1 ∞ ∑ q = 1 q ≠ ℓ ∞ Δ ^ j β ℓ Δ ^ j β q ⟨ ( m h j + 1 2 × g q , h ) × g ℓ , h , θ h j + 1 ⟩ h . (59) \begin{aligned}
\frac{1}{2}\sum_{\ell=1}^{\infty}\widehat{\Delta}_j\beta_\ell\left\langle\left(m_h^{j+1}-m_h^j\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h
={}&\frac{k}{2}\sum_{\ell=1}^{\infty}\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\\
&+\frac{1}{2}\sum_{\ell=1}^{\infty}\left(\left|\widehat{\Delta}_j\beta_\ell\right|^2-k\right)\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\\
&+\frac{\kappa k}{2}\sum_{\ell=1}^{\infty}\widehat{\Delta}_j\beta_\ell\left\langle\left(m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}-\alpha m_h^{j+\frac{1}{2}}\times\left(m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right)\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\\
&+\frac{1}{2}\sum_{\ell=1}^{\infty}\sum_{\substack{q=1\\q\ne\ell}}^{\infty}\widehat{\Delta}_j\beta_\ell\widehat{\Delta}_j\beta_q\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{q,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h. \tag*{(59)}
\end{aligned} 2 1 ℓ = 1 ∑ ∞ Δ j β ℓ ⟨ ( m h j + 1 − m h j ) × g ℓ , h , θ h j + 1 ⟩ h = 2 k ℓ = 1 ∑ ∞ ⟨ ( m h j + 2 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h + 2 1 ℓ = 1 ∑ ∞ ( Δ j β ℓ 2 − k ) ⟨ ( m h j + 2 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h + 2 κk ℓ = 1 ∑ ∞ Δ j β ℓ ⟨ ( m h j + 2 1 × Δ h m h j + 1 − α m h j + 2 1 × ( m h j + 2 1 × Δ h m h j + 1 ) ) × g ℓ , h , θ h j + 1 ⟩ h + 2 1 ℓ = 1 ∑ ∞ q = 1 q = ℓ ∑ ∞ Δ j β ℓ Δ j β q ⟨ ( m h j + 2 1 × g q , h ) × g ℓ , h , θ h j + 1 ⟩ h . (59)
Combining (57) –(59) , and using the fact that I h m j − m h j = θ h j I_hm^j-m_h^j=\theta_h^j I h m j − m h j = θ h j , we can decompose (54) as S r = ∑ ν = 1 6 S r ( ν ) S_r=\sum_{\nu=1}^{6}S_r^{(\nu)} S r = ∑ ν = 1 6 S r ( ν ) , where
S r ( 1 ) : = ∑ j = 0 r − 1 1 Ω Λ , j ⟨ ∑ ℓ = 1 ∞ ∫ t j t j + 1 I h ( m ( s ) − m j ) × g ℓ , h d β ℓ ( s ) , θ h j + 1 ⟩ h . (60) S_r^{(1)}:=\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\langle\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}I_h\left(m(s)-m^j\right)\times g_{\ell,h}\,\mathrm{d}\beta_\ell(s),\theta_h^{j+1}\right\rangle_h. \tag*{(60)} S r ( 1 ) := j = 0 ∑ r − 1 1 Ω Λ , j ⟨ ℓ = 1 ∑ ∞ ∫ t j t j + 1 I h ( m ( s ) − m j ) × g ℓ , h d β ℓ ( s ) , θ h j + 1 ⟩ h . (60)
S r ( 2 ) : = ∑ j = 0 r − 1 1 Ω Λ , j ⟨ θ h j × Δ ^ j W h , θ h j + 1 ⟩ h , (61) S_r^{(2)}:=\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\langle\theta_h^j\times\widehat{\Delta}_jW_h,\theta_h^{j+1}\right\rangle_h, \tag*{(61)} S r ( 2 ) := j = 0 ∑ r − 1 1 Ω Λ , j ⟨ θ h j × Δ j W h , θ h j + 1 ⟩ h , (61)
S r ( 3 ) : = 1 2 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ ( ( I h m ( s ) − m h j + 1 2 ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s , (62) S_r^{(3)} := \frac{1}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle\left(\left(I_hm(s)-m_h^{j+\frac{1}{2}}\right)\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\,\mathrm{d}s, \tag*{(62)} S r ( 3 ) := 2 1 j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ ( ( I h m ( s ) − m h j + 2 1 ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s , (62)
S r ( 4 ) : = − 1 2 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( m h j + 1 2 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h , (63) S_r^{(4)} := -\frac{1}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\left(\left|\widehat{\Delta_j\beta_\ell}\right|^2-k\right)\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h, \tag*{(63)} S r ( 4 ) := − 2 1 j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ( Δ j β ℓ 2 − k ) ⟨ ( m h j + 2 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h , (63)
S r ( 5 ) : = − k κ 2 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ Δ j β ℓ ^ ⟨ ( m h j + 1 2 × Δ h m h j + 1 − α m h j + 1 2 × ( m h j + 1 2 × Δ h m h j + 1 ) ) × g ℓ , h , θ h j + 1 ⟩ h , (64) S_r^{(5)} := -\frac{k\kappa}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\widehat{\Delta_j\beta_\ell}\left\langle\left(m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}-\alpha m_h^{j+\frac{1}{2}}\times\left(m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right)\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h, \tag*{(64)} S r ( 5 ) := − 2 k κ j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ Δ j β ℓ ⟨ ( m h j + 2 1 × Δ h m h j + 1 − α m h j + 2 1 × ( m h j + 2 1 × Δ h m h j + 1 ) ) × g ℓ , h , θ h j + 1 ⟩ h , (64)
S r ( 6 ) : = − 1 2 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∑ q = 1 q ≠ ℓ ∞ Δ j β ℓ ^ Δ j β q ^ ⟨ ( m h j + 1 2 × g q , h ) × g ℓ , h , θ h j + 1 ⟩ h . (65) S_r^{(6)} := -\frac{1}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\sum_{\substack{q=1\\q\ne\ell}}^{\infty}\widehat{\Delta_j\beta_\ell}\widehat{\Delta_j\beta_q}\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{q,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h. \tag*{(65)} S r ( 6 ) := − 2 1 j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ q = 1 q = ℓ ∑ ∞ Δ j β ℓ Δ j β q ⟨ ( m h j + 2 1 × g q , h ) × g ℓ , h , θ h j + 1 ⟩ h . (65)
We estimate these six terms separately. For S r ( 1 ) S_r^{(1)} S r ( 1 ) , write θ h j + 1 = θ h j + ( θ h j + 1 − θ h j ) \theta_h^{j+1}=\theta_h^j+\left(\theta_h^{j+1}-\theta_h^j\right) θ h j + 1 = θ h j + ( θ h j + 1 − θ h j ) . For the part paired with θ h j + 1 − θ h j \theta_h^{j+1}-\theta_h^j θ h j + 1 − θ h j , using Young’s inequality followed by the Itô isometry and (56) , we have
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ I h ( m ( s ) − m j ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h d β ℓ ( s ) ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ ∑ j = 0 J − 1 E [ ∥ ∑ ℓ = 1 ∞ ∫ t j t j + 1 I h ( m ( s ) − m j ) × g ℓ , h d β ℓ ( s ) ∥ h 2 ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ , ε k 1 − ε . \begin{aligned}
\mathbb{E}\left[\sup_{0\le r\le J}\left|\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle I_h(m(s)-m^j)\times g_{\ell,h},\theta_h^{j+1}-\theta_h^j\right\rangle_h\,\mathrm{d}\beta_\ell(s)\right|\right] \\
&\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+C_\delta\sum_{j=0}^{J-1}\mathbb{E}\left[\left\|\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}I_h(m(s)-m^j)\times g_{\ell,h}\,\mathrm{d}\beta_\ell(s)\right\|_h^2\right] \\
&\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+C_{\delta,\varepsilon}k^{1-\varepsilon}.
\end{aligned} E [ 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ I h ( m ( s ) − m j ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h d β ℓ ( s ) ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + C δ j = 0 ∑ J − 1 E ℓ = 1 ∑ ∞ ∫ t j t j + 1 I h ( m ( s ) − m j ) × g ℓ , h d β ℓ ( s ) h 2 ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + C δ , ε k 1 − ε .
The part paired with θ h j \theta_h^j θ h j is a martingale. By the discrete BDG inequality, (2.17), and Young’s inequality, we obtain for every η > 0 \eta>0 η > 0 ,
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ I h ( m ( s ) − m j ) × g ℓ , h , θ h j ⟩ h d β ℓ ( s ) ∣ ] ≤ η E [ sup 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C η , ε k 1 − ε . \begin{aligned}
\mathbb{E}\left[\sup_{0\le r\le J}\left|\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle I_h(m(s)-m^j)\times g_{\ell,h},\theta_h^j\right\rangle_h\,\mathrm{d}\beta_\ell(s)\right|\right] \\
&\le\eta\mathbb{E}\left[\sup_{0\le j\le J}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^j\right\|_h^2\right]+C_{\eta,\varepsilon}k^{1-\varepsilon}.
\end{aligned} E [ 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ I h ( m ( s ) − m j ) × g ℓ , h , θ h j ⟩ h d β ℓ ( s ) ] ≤ η E [ 0 ≤ j ≤ J sup 1 Ω Λ , j θ h j h 2 ] + C η , ε k 1 − ε .
Thus, we have for any δ , η > 0 \delta,\eta>0 δ , η > 0 ,
E [ sup 0 ≤ r ≤ J ∣ S r ( 1 ) ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + η E [ sup 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η , ε k 1 − ε . (66) \mathbb{E}\left[\sup_{0\le r\le J}\left|S_r^{(1)}\right|\right]\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+\eta\mathbb{E}\left[\sup_{0\le j\le J}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^j\right\|_h^2\right]+C_{\delta,\eta,\varepsilon}k^{1-\varepsilon}. \tag*{(66)} E [ 0 ≤ r ≤ J sup S r ( 1 ) ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + η E [ 0 ≤ j ≤ J sup 1 Ω Λ , j θ h j h 2 ] + C δ , η , ε k 1 − ε . (66)
For S r ( 2 ) S_r^{(2)} S r ( 2 ) , we note that ⟨ θ h j × Δ j W h ^ , θ h j + 1 ⟩ h = ⟨ θ h j × Δ j W h ^ , θ h j + 1 − θ h j ⟩ h \left\langle\theta_h^j\times\widehat{\Delta_jW_h},\theta_h^{j+1}\right\rangle_h=\left\langle\theta_h^j\times\widehat{\Delta_jW_h},\theta_h^{j+1}-\theta_h^j\right\rangle_h ⟨ θ h j × Δ j W h , θ h j + 1 ⟩ h = ⟨ θ h j × Δ j W h , θ h j + 1 − θ h j ⟩ h . Therefore, by Young’s inequality and the independence of Δ j W h ^ \widehat{\Delta_jW_h} Δ j W h from F t j \mathcal{F}_{t_j} F t j ,
E [ sup 0 ≤ r ≤ J ∣ S r ( 2 ) ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (67) \begin{aligned}
\mathbb{E}\left[\sup_{0\le r\le J}\left|S_r^{(2)}\right|\right]\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+C_\delta k\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^j\right\|_h^2\right] \\
&\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+C_\delta k\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right]. \tag*{(67)}
\end{aligned} E [ 0 ≤ r ≤ J sup S r ( 2 ) ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + C δ k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j h 2 ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + C δ k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] . (67)
where in the last step we used the fact that θ h 0 = 0 \theta_h^0 = 0 θ h 0 = 0 and Ω Λ , j ⊂ Ω Λ , j − 1 \Omega_{\Lambda,j} \subset\Omega_{\Lambda,j-1} Ω Λ , j ⊂ Ω Λ , j − 1 for j ≥ 1 j \ge1 j ≥ 1 .
Next, we estimate S r ( 3 ) S_r^{(3)} S r ( 3 ) . We first record the consequence of the product estimate (21) which will be used repeatedly below. For v h , w h ∈ V h v_h, w_h \in V_h v h , w h ∈ V h ,
∑ ℓ = 1 ∞ ∣ ⟨ ( v h × g ℓ , h ) × g ℓ , h , w h ⟩ h ∣ = ∑ ℓ = 1 ∞ ∣ ⟨ I h [ ( v h × g ℓ , h ) × g ℓ , h ] , w h ⟩ h ∣ ≤ C ∑ ℓ = 1 ∞ ∥ I h [ ( v h × g ℓ , h ) × g ℓ , h ] ∥ L 2 ∥ w h ∥ h ≤ C Γ ∥ v h ∥ h ∥ w h ∥ h . (68) \begin{aligned}
\sum_{\ell=1}^{\infty} \left|\left\langle(v_h \times g_{\ell,h}) \times g_{\ell,h}, w_h \right\rangle_h\right|
&= \sum_{\ell=1}^{\infty} \left|\left\langle I_h\left[(v_h \times g_{\ell,h}) \times g_{\ell,h}\right], w_h \right\rangle_h\right| \\
&\le C \sum_{\ell=1}^{\infty} \left\|I_h\left[(v_h \times g_{\ell,h}) \times g_{\ell,h}\right]\right\|_{L^2}\|w_h\|_h \\
&\le C_{\Gamma}\|v_h\|_h\|w_h\|_h . \tag*{(68)}
\end{aligned} ℓ = 1 ∑ ∞ ⟨ ( v h × g ℓ , h ) × g ℓ , h , w h ⟩ h = ℓ = 1 ∑ ∞ ⟨ I h [ ( v h × g ℓ , h ) × g ℓ , h ] , w h ⟩ h ≤ C ℓ = 1 ∑ ∞ ∥ I h [ ( v h × g ℓ , h ) × g ℓ , h ] ∥ L 2 ∥ w h ∥ h ≤ C Γ ∥ v h ∥ h ∥ w h ∥ h . (68)
where we also used the discrete norm equivalence (13) and the assumption (6) . Now, using
I h m ( s ) − m h j + 1 2 = I h ( m ( s ) − m j + 1 ) + θ h j + 1 + 1 2 ( m h j + 1 − m h j ) , I_hm(s)-m_h^{j+\frac{1}{2}}=I_h\left(m(s)-m^{j+1}\right)+\theta_h^{j+1}+\frac{1}{2}\left(m_h^{j+1}-m_h^j\right), I h m ( s ) − m h j + 2 1 = I h ( m ( s ) − m j + 1 ) + θ h j + 1 + 2 1 ( m h j + 1 − m h j ) ,
we estimate the three resulting contributions in S r ( 3 ) S_r^{(3)} S r ( 3 ) directly. For the first, estimate (68) , Young’s inequality, the L ∞ L^\infty L ∞ -stability of I h I_h I h , and (56) give
E [ sup 0 ≤ r ≤ J ∣ 1 2 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ ( I h ( m ( s ) − m j + 1 ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s ∣ ] ≤ C ε k 1 − ε + C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (69) \mathbb{E}\left[\sup_{0\le r\le J}\left|\frac{1}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle\left(I_h\left(m(s)-m^{j+1}\right)\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\,\mathrm{d}s\right|\right]
\le C_{\varepsilon}k^{1-\varepsilon}+Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right]. \tag*{(69)} E [ 0 ≤ r ≤ J sup 2 1 j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ ( I h ( m ( s ) − m j + 1 ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s ] ≤ C ε k 1 − ε + C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] . (69)
The contribution containing θ h j + 1 \theta_h^{j+1} θ h j + 1 satisfies
E [ sup 0 ≤ r ≤ J ∣ 1 2 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ ( θ h j + 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s ∣ ] ≤ C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . \mathbb{E}\left[\sup_{0\le r\le J}\left|\frac{1}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle\left(\theta_h^{j+1}\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\,\mathrm{d}s\right|\right]
\le Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right]. E [ 0 ≤ r ≤ J sup 2 1 j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ ( θ h j + 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s ] ≤ C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] .
Finally, (68) , Young’s inequality, and the increment moment bound (38) imply
E [ sup 0 ≤ r ≤ J ∣ 1 4 ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ∫ t j t j + 1 ⟨ ( ( m h j + 1 − m h j ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s ∣ ] ≤ C k ∑ j = 0 J − 1 E [ ∥ m h j + 1 − m h j ∥ h 2 ] + C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] ≤ C k + C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (70) \begin{aligned}
\mathbb{E}\left[\sup_{0\le r\le J}\left|\frac{1}{4}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\int_{t_j}^{t_{j+1}}\left\langle\left((m_h^{j+1}-m_h^j)\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}\right\rangle_h\,\mathrm{d}s\right|\right]
&\le Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\left\|m_h^{j+1}-m_h^j\right\|_h^2\right] \\
&\quad+Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right] \\
&\le Ck+Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right]. \tag*{(70)}
\end{aligned} E [ 0 ≤ r ≤ J sup 4 1 j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ∫ t j t j + 1 ⟨ ( ( m h j + 1 − m h j ) × g ℓ , h ) × g ℓ , h , θ h j + 1 ⟩ h d s ] ≤ C k j = 0 ∑ J − 1 E [ m h j + 1 − m h j h 2 ] + C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] ≤ C k + C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] . (70)
Combining (69) –(70) gives
E [ sup 0 ≤ r ≤ J ∣ S r ( 3 ) ∣ ] ≤ C ε k 1 − ε + C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (71) \mathbb{E}\left[\sup_{0\le r\le J}\left|S_r^{(3)}\right|\right]\le C_{\varepsilon}k^{1-\varepsilon}+Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right]. \tag*{(71)} E [ 0 ≤ r ≤ J sup S r ( 3 ) ] ≤ C ε k 1 − ε + C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] . (71)
For the term S r ( 4 ) S_r^{(4)} S r ( 4 ) , we decompose
θ h j + 1 = θ h j + ( θ h j + 1 − θ h j ) , m h j + 1 2 = m h j + 1 2 ( m h j + 1 − m h j ) . \theta_h^{j+1}=\theta_h^j+\left(\theta_h^{j+1}-\theta_h^j\right),\qquad m_h^{j+\frac{1}{2}}=m_h^j+\frac{1}{2}\left(m_h^{j+1}-m_h^j\right). θ h j + 1 = θ h j + ( θ h j + 1 − θ h j ) , m h j + 2 1 = m h j + 2 1 ( m h j + 1 − m h j ) .
We first estimate the part paired with θ h j + 1 − θ h j \theta_h^{j+1}-\theta_h^j θ h j + 1 − θ h j . By Young’s inequality, for any δ > 0 \delta>0 δ > 0 ,
1 Ω Λ , j E [ ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( m h j + 1 2 × g ℓ , h ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h ] ≤ δ 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 + C δ ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ∥ m h j + 1 2 ∥ h 2 . (72) \begin{aligned}
\mathbb{1}_{\Omega_{\Lambda,j}}\mathbb{E}\Bigg[\sum_{\ell=1}^{\infty}\left(\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k\right)\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}-\theta_h^j\right\rangle_h\Bigg]
&\leq\delta\mathbb{1}_{\Omega_{\Lambda,j}}\lVert\theta_h^{j+1}-\theta_h^j\rVert_h^2 \\
&\quad+C_\delta\left(\sum_{\ell=1}^{\infty}\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\lVert g_{\ell,h}\rVert_{L^\infty}^2\lVert m_h^{j+\frac{1}{2}}\rVert_h^2. \tag*{(72)}
\end{aligned} 1 Ω Λ , j E [ ℓ = 1 ∑ ∞ ( ∣ Δ j β ℓ ∣ 2 − k ) ⟨ ( m h j + 2 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h ] ≤ δ 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 + C δ ( ℓ = 1 ∑ ∞ Δ j β ℓ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ∥ m h j + 2 1 ∥ h 2 . (72)
We note that by the assumption (6) and the fact that E [ ∣ Δ j β ℓ ^ ∣ 2 − k ] 2 = 2 k 2 \mathbb{E}\left[\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right]^2=2k^2 E [ Δ j β ℓ 2 − k ] 2 = 2 k 2 , we have
E [ ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] 2 ≤ C k 2 . (73) \mathbb{E}\left[\left(\sum_{\ell=1}^{\infty}\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\lVert g_{\ell,h}\rVert_{L^\infty}^2\right]^2\leq Ck^2. \tag*{(73)} E [ ( ℓ = 1 ∑ ∞ Δ j β ℓ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] 2 ≤ C k 2 . (73)
Therefore, noting (32) , summing (72) over j j j and taking expectations, we obtain
E [ ∑ j = 0 J − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( m h j + 1 2 × g ℓ , h ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ k . (74) \begin{aligned}
\mathbb{E}\Bigg[\sum_{j=0}^{J-1}\mathbb{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\left(\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^{j+1}-\theta_h^j\right\rangle_h\Bigg]
\leq\delta\mathbb{E}\Bigg[\sum_{j=0}^{J-1}\mathbb{1}_{\Omega_{\Lambda,j}}\lVert\theta_h^{j+1}-\theta_h^j\rVert_h^2\Bigg]+C_\delta k. \tag*{(74)}
\end{aligned} E [ j = 0 ∑ J − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ( Δ j β ℓ 2 − k ) ⟨ ( m h j + 2 1 × g ℓ , h ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ k . (74)
Hence, thanks to the discrete BDG inequality (Lemma A.2 with H = R H=\mathbb{R} H = R and p = 2 p=2 p = 2 ), we have
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( m h j × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ∣ ] ≤ C ( ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( m h j × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ] 2 ) 1 2 ≤ C ( ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] 2 ) 1 2 = C ( ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] E [ ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] 2 ) 1 2 . (75) \begin{aligned}
\mathbb{E}\left[\sup_{0\leq r\leq J}\left|\sum_{j=0}^{r-1}\mathbb{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\left(\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\left\langle\left(m_h^j\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^j\right\rangle_h\right|\right]
&\leq C\left(\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\left(\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\left\langle\left(m_h^j\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^j\right\rangle_h\right]^2\right)^{\frac{1}{2}} \\
&\leq C\left(\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\lVert\theta_h^j\rVert_h^2\left(\sum_{\ell=1}^{\infty}\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\lVert g_{\ell,h}\rVert_{L^\infty}^2\right]^2\right)^{\frac{1}{2}} \\
&=C\left(\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\lVert\theta_h^j\rVert_h^2\right]\mathbb{E}\left[\left(\sum_{\ell=1}^{\infty}\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\lVert g_{\ell,h}\rVert_{L^\infty}^2\right]^2\right)^{\frac{1}{2}}. \tag*{(75)}
\end{aligned} E [ 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ( Δ j β ℓ 2 − k ) ⟨ ( m h j × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ] ≤ C j = 0 ∑ J − 1 E [ 1 Ω Λ , j ℓ = 1 ∑ ∞ ( Δ j β ℓ 2 − k ) ⟨ ( m h j × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ] 2 2 1 ≤ C j = 0 ∑ J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ( ℓ = 1 ∑ ∞ Δ j β ℓ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] 2 2 1 = C j = 0 ∑ J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] E [ ( ℓ = 1 ∑ ∞ Δ j β ℓ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] 2 2 1 . (75)
where in the penultimate step we used the Cauchy–Schwarz inequality, (32) , and (6) , while in the last step we used the fact that 1 Ω Λ , j ∥ θ h j ∥ h 2 \mathbb{1}_{\Omega_{\Lambda,j}}\lVert\theta_h^j\rVert_h^2 1 Ω Λ , j ∥ θ h j ∥ h 2 is F t j \mathcal{F}_{t_j} F t j -measurable and the Brownian increments over [ t j , t j + 1 ] [t_j,t_{j+1}] [ t j , t j + 1 ] are independent of F t j \mathcal{F}_{t_j} F t j . Therefore, continuing from (75) , we obtain by Young’s inequality and (73) that
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( m h j × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ∣ ] ≤ η E [ max 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C η k . (76) \begin{aligned}
\mathbb{E}\left[\sup_{0\leq r\leq J}\left|\sum_{j=0}^{r-1}\mathbb{1}_{\Omega_{\Lambda,j}}\sum_{\ell=1}^{\infty}\left(\left\lvert\widehat{\Delta_j\beta_\ell}\right\rvert^2-k\right)\left\langle\left(m_h^j\times g_{\ell,h}\right)\times g_{\ell,h},\theta_h^j\right\rangle_h\right|\right]
\leq\eta\mathbb{E}\left[\max_{0\leq j\leq J}\mathbb{1}_{\Omega_{\Lambda,j}}\lVert\theta_h^j\rVert_h^2\right]+C_\eta k. \tag*{(76)}
\end{aligned} E [ 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ = 1 ∑ ∞ ( Δ j β ℓ 2 − k ) ⟨ ( m h j × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ] ≤ η E [ 0 ≤ j ≤ J max 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C η k . (76)
It remains to estimate the part containing m h j + 1 − m h j \mathbf{m}_h^{j+1}-\mathbf{m}_h^j m h j + 1 − m h j . By (68) and Young’s inequality,
E [ ∑ j = 0 J − 1 1 Ω Λ , j [ ∑ ℓ = 1 ∞ ( ∣ Δ j β ℓ ^ ∣ 2 − k ) ⟨ ( ( m h j + 1 − m h j ) × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ] ] ≤ C ∑ j = 0 J − 1 E [ 1 Ω Λ , j ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ m h j + 1 − m h j ∥ h ∥ θ h j ∥ h ≤ C ∑ j = 0 J − 1 E [ ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ m h j + 1 − m h j ∥ h 2 + C ∑ j = 0 J − 1 E [ 1 Ω Λ , j ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ θ h j ∥ h 2 . (77) \begin{aligned}
\mathbb{E}\Biggl[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\Biggl[\sum_{\ell=1}^{\infty}(\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k)\Biggr.\Biggr.&\Biggl.\Biggl\langle((\mathbf{m}_h^{j+1}-\mathbf{m}_h^j)\times\mathbf{g}_{\ell,h})\times\mathbf{g}_{\ell,h},\theta_h^j\rangle_h\Biggr]\Biggr] \\
&\le C\sum_{j=0}^{J-1}\mathbb{E}[\mathbf{1}_{\Omega_{\Lambda,j}}(\sum_{\ell=1}^{\infty}\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k)\|\mathbf{g}_{\ell,h}\|_{L^\infty}^2]\|\mathbf{m}_h^{j+1}-\mathbf{m}_h^j\|_h\|\theta_h^j\|_h \\
&\le C\sum_{j=0}^{J-1}\mathbb{E}[(\sum_{\ell=1}^{\infty}\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k)\|\mathbf{g}_{\ell,h}\|_{L^\infty}^2]\|\mathbf{m}_h^{j+1}-\mathbf{m}_h^j\|_h^2 \\
&\quad+C\sum_{j=0}^{J-1}\mathbb{E}[\mathbf{1}_{\Omega_{\Lambda,j}}(\sum_{\ell=1}^{\infty}\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k)\|\mathbf{g}_{\ell,h}\|_{L^\infty}^2]\|\theta_h^j\|_h^2 . \tag*{(77)}
\end{aligned} E [ j = 0 ∑ J − 1 1 Ω Λ , j [ ℓ = 1 ∑ ∞ (∣ Δ j β ℓ ∣ 2 − k ) ⟨ (( m h j + 1 − m h j ) × g ℓ , h ) × g ℓ , h , θ h j ⟩ h ] ] ≤ C j = 0 ∑ J − 1 E [ 1 Ω Λ , j ( ℓ = 1 ∑ ∞ ∣ Δ j β ℓ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ m h j + 1 − m h j ∥ h ∥ θ h j ∥ h ≤ C j = 0 ∑ J − 1 E [( ℓ = 1 ∑ ∞ ∣ Δ j β ℓ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ m h j + 1 − m h j ∥ h 2 + C j = 0 ∑ J − 1 E [ 1 Ω Λ , j ( ℓ = 1 ∑ ∞ ∣ Δ j β ℓ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ θ h j ∥ h 2 . (77)
For the first term on the right-hand side, Hölder’s inequality and (73) yield
∑ j = 0 J − 1 E [ ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ m h j + 1 − m h j ∥ h 2 ≤ C k ∑ j = 0 J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 1 2 ≤ C k . (78) \begin{aligned}
\sum_{j=0}^{J-1}\mathbb{E}\left[\left(\sum_{\ell=1}^{\infty}\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k\right)\|\mathbf{g}_{\ell,h}\|_{L^\infty}^2\right]\|\mathbf{m}_h^{j+1}-\mathbf{m}_h^j\|_h^2
&\le Ck\sum_{j=0}^{J-1}\left(\mathbb{E}\left[\|\mathbf{m}_h^{j+1}-\mathbf{m}_h^j\|_h^4\right]\right)^{\frac{1}{2}} \\
&\le Ck . \tag*{(78)}
\end{aligned} j = 0 ∑ J − 1 E [ ( ℓ = 1 ∑ ∞ ∣ Δ j β ℓ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ m h j + 1 − m h j ∥ h 2 ≤ C k j = 0 ∑ J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 2 1 ≤ C k . (78)
where in the final step we also used the increment moment bound (39) . For the last term on the right-hand side of (77) , the factor involving the Brownian increment is independent of F t j \mathcal{F}_{t_j} F t j and has expectation bounded by C k Ck C k . Therefore,
∑ j = 0 J − 1 E [ 1 Ω Λ , j ( ∑ ℓ = 1 ∞ ∣ Δ j β ℓ ^ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ θ h j ∥ h 2 ≤ C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] . (79) \sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left(\sum_{\ell=1}^{\infty}\lvert\widehat{\Delta_j\beta_\ell}\rvert^2-k\right)\|\mathbf{g}_{\ell,h}\|_{L^\infty}^2\right]\|\theta_h^j\|_h^2
\le Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\right] . \tag*{(79)} j = 0 ∑ J − 1 E [ 1 Ω Λ , j ( ℓ = 1 ∑ ∞ ∣ Δ j β ℓ ∣ 2 − k ) ∥ g ℓ , h ∥ L ∞ 2 ] ∥ θ h j ∥ h 2 ≤ C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] . (79)
Combining (74) , (76) , (77) , (78) , and (79) , and absorbing harmless numerical constants into δ \delta δ and η \eta η , we obtain
E [ sup 0 ≤ r ≤ J ∣ S r ( 4 ) ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + η E [ sup 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η k + C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (80) \mathbb{E}\left[\sup_{0\le r\le J}\lvert S_r^{(4)}\rvert\right]\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}-\theta_h^j\|_h^2\right]+\eta\mathbb{E}\left[\sup_{0\le j\le J}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\right]+C_{\delta,\eta}k+Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}\|_h^2\right] . \tag*{(80)} E [ 0 ≤ r ≤ J sup ∣ S r ( 4 ) ∣ ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + η E [ 0 ≤ j ≤ J sup 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η k + C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (80)
We estimate S r ( 5 ) S_r^{(5)} S r ( 5 ) next. By the mass-lumped Cauchy–Schwarz inequality and the nodal bound (31) , we first have
∣ S r ( 5 ) ∣ ≤ C k ∑ j = 0 r − 1 1 Ω Λ , j ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h ∥ Δ j W ^ h ∥ L ∞ ∥ θ h j + 1 ∥ h . (81) \lvert S_r^{(5)}\rvert\le Ck\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\mathbf{m}_h^{j+\frac{1}{2}}\times\Delta_h\mathbf{m}_h^{j+1}\right\|_h\left\|\widehat{\Delta_jW}_h\right\|_{L^\infty}\left\|\theta_h^{j+1}\right\|_h . \tag*{(81)} ∣ S r ( 5 ) ∣ ≤ C k j = 0 ∑ r − 1 1 Ω Λ , j m h j + 2 1 × Δ h m h j + 1 h Δ j W h L ∞ θ h j + 1 h . (81)
Now, let ρ > 0 \rho>0 ρ > 0 . Young’s inequality with conjugate exponents ( 2 + ρ ) / ( 1 + ρ ) (2+\rho)/(1+\rho) ( 2 + ρ ) / ( 1 + ρ ) and 2 + ρ 2+\rho 2 + ρ , together with the uniform bound ∥ θ h j + 1 ∥ h ≤ C \|\theta_h^{j+1}\|_h\le C ∥ θ h j + 1 ∥ h ≤ C , gives
∣ S r ( 5 ) ∣ ≤ C ρ k ∑ j = 0 r − 1 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 + ρ 1 + ρ ∥ Δ j W ^ h ∥ L ∞ 2 + ρ 1 + ρ + C ρ k ∑ j = 0 r − 1 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 . (82) \lvert S_r^{(5)}\rvert\le C_\rho k\sum_{j=0}^{r-1}\left\|\mathbf{m}_h^{j+\frac{1}{2}}\times\Delta_h\mathbf{m}_h^{j+1}\right\|_h^{\frac{2+\rho}{1+\rho}}\left\|\widehat{\Delta_jW}_h\right\|_{L^\infty}^{\frac{2+\rho}{1+\rho}}+C_\rho k\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}\|_h^2 . \tag*{(82)} ∣ S r ( 5 ) ∣ ≤ C ρ k j = 0 ∑ r − 1 m h j + 2 1 × Δ h m h j + 1 h 1 + ρ 2 + ρ Δ j W h L ∞ 1 + ρ 2 + ρ + C ρ k j = 0 ∑ r − 1 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 . (82)
By Hölder’s inequality and (28) ,
k ∑ j = 0 J − 1 E [ ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 + ρ 1 + ρ ∥ Δ j W h ∥ L ∞ 2 + ρ 1 + ρ ] ≤ C ρ k 1 + 2 + ρ 2 + 2 ρ ∑ j = 0 J − 1 ( E [ ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ] ) 2 + ρ 2 + 2 ρ . k\sum_{j=0}^{J-1}\mathbb{E}\left[\left\|m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right\|_h^{\frac{2+\rho}{1+\rho}}\left\|\Delta_jW_h\right\|_{L^\infty}^{\frac{2+\rho}{1+\rho}}\right]\le C_\rho k^{1+\frac{2+\rho}{2+2\rho}}\sum_{j=0}^{J-1}\left(\mathbb{E}\left[\left\|m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right\|_h^2\right]\right)^{\frac{2+\rho}{2+2\rho}}. k j = 0 ∑ J − 1 E [ m h j + 2 1 × Δ h m h j + 1 h 1 + ρ 2 + ρ ∥ Δ j W h ∥ L ∞ 1 + ρ 2 + ρ ] ≤ C ρ k 1 + 2 + 2 ρ 2 + ρ j = 0 ∑ J − 1 ( E [ m h j + 2 1 × Δ h m h j + 1 h 2 ] ) 2 + 2 ρ 2 + ρ .
Indeed, the moment of the Wiener increment used here is of order 2 + 4 / ρ 2+4/\rho 2 + 4/ ρ . Since 2 + ρ 2 + 2 ρ ∈ ( 0 , 1 ) \frac{2+\rho}{2+2\rho}\in(0,1) 2 + 2 ρ 2 + ρ ∈ ( 0 , 1 ) , the discrete Jensen inequality (Lemma A.1) and Proposition 3.4 yield
∑ j = 0 J − 1 ( E [ ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ] ) 2 + ρ 2 + 2 ρ ≤ J ρ 2 + 2 ρ ( ∑ j = 0 J − 1 E [ ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ] ) 2 + ρ 2 + 2 ρ ≤ C ρ k − 1 . \sum_{j=0}^{J-1}\left(\mathbb{E}\left[\left\|m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right\|_h^2\right]\right)^{\frac{2+\rho}{2+2\rho}}\le J^{\frac{\rho}{2+2\rho}}\left(\sum_{j=0}^{J-1}\mathbb{E}\left[\left\|m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right\|_h^2\right]\right)^{\frac{2+\rho}{2+2\rho}}\le C_\rho k^{-1}. j = 0 ∑ J − 1 ( E [ m h j + 2 1 × Δ h m h j + 1 h 2 ] ) 2 + 2 ρ 2 + ρ ≤ J 2 + 2 ρ ρ ( j = 0 ∑ J − 1 E [ m h j + 2 1 × Δ h m h j + 1 h 2 ] ) 2 + 2 ρ 2 + ρ ≤ C ρ k − 1 .
This estimate, together with (82) , implies
E [ sup 0 ≤ r ≤ J ∣ S r ( 5 ) ∣ ] ≤ C ρ k 2 + ρ 2 + 2 ρ + C ρ k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] ≤ C ε k 1 − ε + C ε k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] , (83) \begin{aligned}
\mathbb{E}\left[\sup_{0\le r\le J}\left|S_r^{(5)}\right|\right]\le C_\rho k^{\frac{2+\rho}{2+2\rho}}+C_\rho k\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right] \\
&\le C_\varepsilon k^{1-\varepsilon}+C_\varepsilon k\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\right], \tag*{(83)}
\end{aligned} E [ 0 ≤ r ≤ J sup S r ( 5 ) ] ≤ C ρ k 2 + 2 ρ 2 + ρ + C ρ k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] ≤ C ε k 1 − ε + C ε k j = 0 ∑ J − 1 E [ 1 Ω Λ , j θ h j + 1 h 2 ] , (83)
where ρ > 0 \rho>0 ρ > 0 is chosen so that ρ / ( 2 + 2 ρ ) ≤ ε \rho/(2+2\rho)\le\varepsilon ρ / ( 2 + 2 ρ ) ≤ ε .
Finally, we estimate S r ( 6 ) S_r^{(6)} S r ( 6 ) . To avoid repeatedly writing the same off-diagonal coefficient, define
X j : = ∑ ℓ = 1 ∞ ∑ q = 1 q ≠ ℓ ∞ ∣ Δ j β ℓ ^ Δ j β q ^ ∣ ∥ g ℓ , h ∥ L ∞ ∥ g q , h ∥ L ∞ . \mathcal{X}_j:=\sum_{\ell=1}^{\infty}\sum_{\substack{q=1\\q\ne\ell}}^{\infty}\left|\widehat{\Delta_j\beta_\ell}\widehat{\Delta_j\beta_q}\right|\|g_{\ell,h}\|_{L^\infty}\|g_{q,h}\|_{L^\infty}. X j := ℓ = 1 ∑ ∞ q = 1 q = ℓ ∑ ∞ Δ j β ℓ Δ j β q ∥ g ℓ , h ∥ L ∞ ∥ g q , h ∥ L ∞ .
By Gaussian moments and independence of Brownian increments, Minkowski’s inequality in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , as well as assumption (6) , we have
E [ X j ] ≤ C k , E [ X j 2 ] ≤ C k 2 . (84) \mathbb{E}[\mathcal{X}_j]\le Ck,\qquad\mathbb{E}[\mathcal{X}_j^2]\le Ck^2. \tag*{(84)} E [ X j ] ≤ C k , E [ X j 2 ] ≤ C k 2 . (84)
By writing
m h j + 1 2 = m h j + 1 2 ( m h j + 1 − m h j ) and θ h j + 1 = θ h j + ( θ h j + 1 − θ h j ) , m_h^{j+\frac{1}{2}}=m_h^j+\frac{1}{2}(m_h^{j+1}-m_h^j)\quad\text{and}\quad\theta_h^{j+1}=\theta_h^j+(\theta_h^{j+1}-\theta_h^j), m h j + 2 1 = m h j + 2 1 ( m h j + 1 − m h j ) and θ h j + 1 = θ h j + ( θ h j + 1 − θ h j ) ,
we first estimate the part paired with the error increment θ h j + 1 − θ h j \theta_h^{j+1}-\theta_h^j θ h j + 1 − θ h j . The nodal bound (31) and Young’s inequality imply
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ , q ≥ 1 q ≠ ℓ Δ j β ℓ ^ Δ j β q ^ ⟨ ( m h j + 1 2 × g q , h ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + C δ k . \mathbb{E}\left[\sup_{0\le r\le J}\left|\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_{\Lambda,j}}\sum_{\substack{\ell,q\ge1\\q\ne\ell}}\widehat{\Delta_j\beta_\ell}\widehat{\Delta_j\beta_q}\left\langle\left(m_h^{j+\frac{1}{2}}\times g_{q,h}\right)\times g_{\ell,h},\theta_h^{j+1}-\theta_h^j\right\rangle_h\right|\right]
\le\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+C_\delta k. E 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ , q ≥ 1 q = ℓ ∑ Δ j β ℓ Δ j β q ⟨ ( m h j + 2 1 × g q , h ) × g ℓ , h , θ h j + 1 − θ h j ⟩ h ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j θ h j + 1 − θ h j h 2 ] + C δ k .
The leading part paired with θ h j \theta_h^j θ h j is a martingale because, for q ≠ ℓ q\ne\ell q = ℓ , we have E [ Δ j β ℓ ^ Δ j β q ^ ∣ F t j ] = 0 \mathbb{E}\left[\widehat{\Delta_j\beta_\ell}\widehat{\Delta_j\beta_q}\mid\mathcal{F}_{t_j}\right]=0 E [ Δ j β ℓ Δ j β q ∣ F t j ] = 0 . Moreover, for every scalar array { c ℓ q } ℓ ≠ q \{c_{\ell q}\}_{\ell\ne q} { c ℓ q } ℓ = q ,
E [ ∣ ∑ ℓ , q ≥ 1 q ≠ ℓ c ℓ q Δ j β ℓ ^ Δ j β q ^ ∣ 2 ] ≤ 2 k 2 ( ∑ ℓ , q ≥ 1 q ≠ ℓ ∣ c ℓ q ∣ ) 2 . \mathbb{E}\left[\left|\sum_{\substack{\ell,q\ge1\\q\ne\ell}}c_{\ell q}\widehat{\Delta_j\beta_\ell}\widehat{\Delta_j\beta_q}\right|^2\right]\le2k^2\left(\sum_{\substack{\ell,q\ge1\\q\ne\ell}}|c_{\ell q}|\right)^2. E ℓ , q ≥ 1 q = ℓ ∑ c ℓ q Δ j β ℓ Δ j β q 2 ≤ 2 k 2 ℓ , q ≥ 1 q = ℓ ∑ ∣ c ℓ q ∣ 2 .
Applying this conditionally on F t j \mathcal{F}_{t_j} F t j , followed by the discrete BDG and Young inequalities, yields
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ , q ≥ 1 q ≠ ℓ Δ ^ j β ℓ Δ ^ j β q ⟨ ( m h j × g q , h ) × g ℓ , h , θ h j ⟩ h ∣ ] ≤ η E [ sup 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C η k . \mathbb{E}\left[\sup_{0\leq r\leq J}\left|\sum_{j=0}^{r-1}\mathbb{1}_{\Omega_{\Lambda,j}}\sum_{\substack{\ell,q\geq1\\q\ne\ell}}\widehat{\Delta}_j\beta_\ell\widehat{\Delta}_j\beta_q\left\langle\left(m_h^j\times g_{q,h}\right)\times g_{\ell,h},\theta_h^j\right\rangle_h\right|\right]\leq\eta\mathbb{E}\left[\sup_{0\leq j\leq J}\mathbb{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\right]+C_{\eta}k. E 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ , q ≥ 1 q = ℓ ∑ Δ j β ℓ Δ j β q ⟨ ( m h j × g q , h ) × g ℓ , h , θ h j ⟩ h ≤ η E [ 0 ≤ j ≤ J sup 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C η k .
It remains to estimate the contribution arising from 1 2 ( m h j + 1 − m h j ) \frac{1}{2}(m_h^{j+1}-m_h^j) 2 1 ( m h j + 1 − m h j ) paired with θ h j \theta_h^j θ h j . Using the definition of X j \mathcal{X}_j X j and the mass-lumped Cauchy–Schwarz inequality, we obtain
E [ sup 0 ≤ r ≤ J ∣ ∑ j = 0 r − 1 1 Ω Λ , j ∑ ℓ , q ≥ 1 q ≠ ℓ Δ ^ j β ℓ Δ ^ j β q ⟨ ( ( m h j + 1 − m h j ) × g q , h ) × g ℓ , h , θ h j ⟩ h ∣ ] ≤ C ∑ j = 0 J − 1 E [ 1 Ω Λ , j X j ∥ m h j + 1 − m h j ∥ h ∥ θ h j ∥ h ] ≤ C ∑ j = 0 J − 1 E [ X j ∥ m h j + 1 − m h j ∥ h 2 ] + C ∑ j = 0 J − 1 E [ 1 Ω Λ , j X j ∥ θ h j ∥ h 2 ] . \begin{aligned}
\mathbb{E}\left[\sup_{0\leq r\leq J}\left|\sum_{j=0}^{r-1}\mathbb{1}_{\Omega_{\Lambda,j}}\sum_{\substack{\ell,q\geq1\\q\ne\ell}}\widehat{\Delta}_j\beta_\ell\widehat{\Delta}_j\beta_q\left\langle\left((m_h^{j+1}-m_h^j)\times g_{q,h}\right)\times g_{\ell,h},\theta_h^j\right\rangle_h\right|\right]
\leq C\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\mathcal{X}_j\|m_h^{j+1}-m_h^j\|_h\|\theta_h^j\|_h\right] \\
&\leq C\sum_{j=0}^{J-1}\mathbb{E}\left[\mathcal{X}_j\|m_h^{j+1}-m_h^j\|_h^2\right]+C\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\mathcal{X}_j\|\theta_h^j\|_h^2\right].
\end{aligned} E 0 ≤ r ≤ J sup j = 0 ∑ r − 1 1 Ω Λ , j ℓ , q ≥ 1 q = ℓ ∑ Δ j β ℓ Δ j β q ⟨ ( ( m h j + 1 − m h j ) × g q , h ) × g ℓ , h , θ h j ⟩ h ≤ C j = 0 ∑ J − 1 E [ 1 Ω Λ , j X j ∥ m h j + 1 − m h j ∥ h ∥ θ h j ∥ h ] ≤ C j = 0 ∑ J − 1 E [ X j ∥ m h j + 1 − m h j ∥ h 2 ] + C j = 0 ∑ J − 1 E [ 1 Ω Λ , j X j ∥ θ h j ∥ h 2 ] .
For the first term on the right-hand side, applying the Cauchy–Schwarz inequality, (84) , and (39) give
∑ j = 0 J − 1 E [ X j ∥ m h j + 1 − m h j ∥ h 2 ] ≤ ∑ j = 0 J − 1 ( E [ X j 2 ] ) 1 2 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 1 2 ≤ C k ∑ j = 0 J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 1 2 ≤ C k . \sum_{j=0}^{J-1}\mathbb{E}\left[\mathcal{X}_j\|m_h^{j+1}-m_h^j\|_h^2\right]\leq\sum_{j=0}^{J-1}\left(\mathbb{E}[\mathcal{X}_j^2]\right)^{\frac{1}{2}}\left(\mathbb{E}\left[\|m_h^{j+1}-m_h^j\|_h^4\right]\right)^{\frac{1}{2}}
\leq Ck\sum_{j=0}^{J-1}\left(\mathbb{E}\left[\|m_h^{j+1}-m_h^j\|_h^4\right]\right)^{\frac{1}{2}}\leq Ck. j = 0 ∑ J − 1 E [ X j ∥ m h j + 1 − m h j ∥ h 2 ] ≤ j = 0 ∑ J − 1 ( E [ X j 2 ] ) 2 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 2 1 ≤ C k j = 0 ∑ J − 1 ( E [ ∥ m h j + 1 − m h j ∥ h 4 ] ) 2 1 ≤ C k .
For the second term, both 1 Ω Λ , j \mathbb{1}_{\Omega_{\Lambda,j}} 1 Ω Λ , j and θ h j \theta_h^j θ h j are F t j \mathcal{F}_{t_j} F t j -measurable, whereas X j \mathcal{X}_j X j depends only on the Brownian increments over [ t j , t j + 1 ] [t_j,t_{j+1}] [ t j , t j + 1 ] and is therefore independent of F t j \mathcal{F}_{t_j} F t j . Hence, again by (84) ,
∑ j = 0 J − 1 E [ 1 Ω Λ , j X j ∥ θ h j ∥ h 2 ] = ∑ j = 0 J − 1 E [ X j ] E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] ≤ C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] , \sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\mathcal{X}_j\|\theta_h^j\|_h^2\right]=\sum_{j=0}^{J-1}\mathbb{E}[\mathcal{X}_j]\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\right]\leq Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}\|_h^2\right], j = 0 ∑ J − 1 E [ 1 Ω Λ , j X j ∥ θ h j ∥ h 2 ] = j = 0 ∑ J − 1 E [ X j ] E [ 1 Ω Λ , j ∥ θ h j ∥ h 2 ] ≤ C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] ,
where the last step follows from the fact that θ h 0 = 0 \theta_h^0=0 θ h 0 = 0 and Ω Λ , j ⊂ Ω Λ , j − 1 \Omega_{\Lambda,j}\subset\Omega_{\Lambda,j-1} Ω Λ , j ⊂ Ω Λ , j − 1 for j ≥ 1 j\geq1 j ≥ 1 . We now combine the above estimates to obtain
E [ sup 0 ≤ r ≤ J ∣ S r ( 6 ) ∣ ] ≤ δ E [ ∑ j = 0 J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + η E [ sup 0 ≤ j ≤ J 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η , ε k 1 − ε + C k ∑ j = 0 J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (85) \mathbb{E}\left[\sup_{0\leq r\leq J}|S_r^{(6)}|\right]\leq\delta\mathbb{E}\left[\sum_{j=0}^{J-1}\mathbb{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}-\theta_h^j\|_h^2\right]+\eta\mathbb{E}\left[\sup_{0\leq j\leq J}\mathbb{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\right]+C_{\delta,\eta,\varepsilon}k^{1-\varepsilon}+Ck\sum_{j=0}^{J-1}\mathbb{E}\left[\mathbb{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}\|_h^2\right]. \tag*{(85)} E [ 0 ≤ r ≤ J sup ∣ S r ( 6 ) ∣ ] ≤ δ E [ j = 0 ∑ J − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + η E [ 0 ≤ j ≤ J sup 1 Ω Λ , j ∥ θ h j ∥ h 2 ] + C δ , η , ε k 1 − ε + C k j = 0 ∑ J − 1 E [ 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 ] . (85)
Combining (66) , (67) , (71) , (80) , (83) , and (85) completes the proof of (55) . □
For a h , w h , ϕ h ∈ V h a_h,w_h,\phi_h\in V_h a h , w h , ϕ h ∈ V h , define the discrete Landau–Lifshitz drift form
A h ( a h ; w h , ϕ h ) : = α κ ⟨ a h × ( a h × Δ h w h ) , ϕ h ⟩ h − κ ⟨ a h × Δ h w h , ϕ h ⟩ h . (86) \mathcal{A}_h(a_h;w_h,\phi_h):=\alpha\kappa\langle a_h\times(a_h\times\Delta_hw_h),\phi_h\rangle_h-\kappa\langle a_h\times\Delta_hw_h,\phi_h\rangle_h. \tag*{(86)} A h ( a h ; w h , ϕ h ) := α κ ⟨ a h × ( a h × Δ h w h ) , ϕ h ⟩ h − κ ⟨ a h × Δ h w h , ϕ h ⟩ h . (86)
We are now in a position to prove our main theorem.
Theorem 3.8. Let α , κ > 0 \alpha,\kappa>0 α , κ > 0 and γ ∈ ( 0 , 1 2 ) \gamma\in(0,\frac{1}{2}) γ ∈ ( 0 , 2 1 ) . Let m m m be the pathwise strong solution of (1) , and let { m h j } j = 0 J \{m_h^j\}_{j=0}^{J} { m h j } j = 0 J be an adapted solution of (30) . For every fixed Λ > 0 \Lambda>0 Λ > 0 , there exist k Λ ∈ ( 0 , 1 ] k_\Lambda\in(0,1] k Λ ∈ ( 0 , 1 ] and C Λ , γ > 0 C_{\Lambda,\gamma}>0 C Λ , γ > 0 , independent of h h h and k k k , such that, whenever 0 < k ≤ k Λ 0<k\leq k_\Lambda 0 < k ≤ k Λ ,
E [ 1 Ω Λ , J max 0 ≤ j ≤ J ∥ θ h j ∥ h 2 ] + E [ 1 Ω Λ , J ∑ j = 0 J − 1 ∥ θ h j + 1 − θ h j ∥ h 2 ] + E [ 1 Ω Λ , J k ∑ j = 0 J − 1 ∥ ∂ x θ h j + 1 ∥ L 2 2 ] ≤ C Λ , γ ( h 2 + k 2 γ ) . (87) \mathbb{E}\left[1_{\Omega_{\Lambda,J}}\max_{0\leq j\leq J}\left\|\theta_h^j\right\|_h^2\right]+\mathbb{E}\left[1_{\Omega_{\Lambda,J}} \sum_{j=0}^{J-1}\left\|\theta_h^{j+1}-\theta_h^j\right\|_h^2\right]+\mathbb{E}\left[1_{\Omega_{\Lambda,J}}k\sum_{j=0}^{J-1}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2\right]\leq C_{\Lambda,\gamma}\left(h^2+k^{2\gamma}\right). \tag*{(87)} E [ 1 Ω Λ , J 0 ≤ j ≤ J max θ h j h 2 ] + E [ 1 Ω Λ , J j = 0 ∑ J − 1 θ h j + 1 − θ h j h 2 ] + E [ 1 Ω Λ , J k j = 0 ∑ J − 1 ∂ x θ h j + 1 L 2 2 ] ≤ C Λ , γ ( h 2 + k 2 γ ) . (87)
Consequently,
E [ 1 Ω Λ , J max 0 ≤ j ≤ J ∥ m j − m h j ∥ L 2 2 ] + E [ 1 Ω Λ , J k ∑ j = 0 J − 1 ∥ ∂ x ( m j + 1 − m h j + 1 ) ∥ L 2 2 ] ≤ C Λ , γ ( h 2 + k 2 γ ) . (88) \mathbb{E}\left[1_{\Omega_{\Lambda,J}}\max_{0\leq j\leq J}\left\|m^j-m_h^j\right\|_{L^2}^2\right]+\mathbb{E}\left[1_{\Omega_{\Lambda,J}}k\sum_{j=0}^{J-1}\left\|\partial_x(m^{j+1}-m_h^{j+1})\right\|_{L^2}^2\right]\leq C_{\Lambda,\gamma}\left(h^2+k^{2\gamma}\right). \tag*{(88)} E [ 1 Ω Λ , J 0 ≤ j ≤ J max m j − m h j L 2 2 ] + E [ 1 Ω Λ , J k j = 0 ∑ J − 1 ∂ x ( m j + 1 − m h j + 1 ) L 2 2 ] ≤ C Λ , γ ( h 2 + k 2 γ ) . (88)
Proof. Since m h 0 = I h m 0 m_h^0=I_hm_0 m h 0 = I h m 0 , we have θ h 0 = 0 \theta_h^0=\mathbf{0} θ h 0 = 0 . Subtracting (30) from (45) yields, for every ϕ h ∈ V h \phi_h\in V_h ϕ h ∈ V h ,
⟨ θ h j + 1 − θ h j , ϕ h ⟩ h + k [ A h ( I h m ‾ j + 1 2 ; I h m j + 1 , ϕ h ) − A h ( m h j + 1 2 ; m h j + 1 , ϕ h ) ] − ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) − m h j + 1 2 × Δ ^ j W h , ϕ h ⟩ h = R h j + 1 ( ϕ h ) , (89) \left\langle\theta_h^{j+1}-\theta_h^j,\phi_h\right\rangle_h+k\left[\mathcal{A}_h\left(I_h\overline{m}^{j+\frac{1}{2}};I_hm^{j+1},\phi_h\right)-\mathcal{A}_h\left(m_h^{j+\frac{1}{2}};m_h^{j+1},\phi_h\right)\right]
-\left\langle\int_{t_j}^{t_{j+1}}I_hm(s)\times\circ\mathrm{d}W_h(s)-m_h^{j+\frac{1}{2}}\times\widehat{\Delta}_jW_h,\phi_h\right\rangle_h=\mathcal{R}_h^{j+1}(\phi_h), \tag*{(89)} ⟨ θ h j + 1 − θ h j , ϕ h ⟩ h + k [ A h ( I h m j + 2 1 ; I h m j + 1 , ϕ h ) − A h ( m h j + 2 1 ; m h j + 1 , ϕ h ) ] − ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) − m h j + 2 1 × Δ j W h , ϕ h ⟩ h = R h j + 1 ( ϕ h ) , (89)
where A h \mathcal{A}_h A h is given by (86) . Now, choose ϕ h = 1 Ω Λ , j θ h j + 1 \phi_h=1_{\Omega_{\Lambda,j}}\theta_h^{j+1} ϕ h = 1 Ω Λ , j θ h j + 1 . By Lemma B.4, we have
1 Ω Λ , j ⟨ θ h j + 1 − θ h j , θ h j + 1 ⟩ h + c 0 k 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 ≤ C Λ k 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 + C Λ k ∑ ν = 1 3 ( ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν + ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ) + 1 Ω Λ , j R h j + 1 ( θ h j + 1 ) + 1 Ω Λ , j ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) − m h j + 1 2 × Δ ^ j W h , θ h j + 1 ⟩ h . (90) 1_{\Omega_{\Lambda,j}}\left\langle\theta_h^{j+1}-\theta_h^j,\theta_h^{j+1}\right\rangle_h+c_0k1_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2
\leq C_{\Lambda}k1_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2+C_{\Lambda}k\sum_{\nu=1}^{3}\left(\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}^{2\nu}+\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2\nu}\right)
+1_{\Omega_{\Lambda,j}}\mathcal{R}_h^{j+1}(\theta_h^{j+1})+1_{\Omega_{\Lambda,j}}\left\langle\int_{t_j}^{t_{j+1}}I_hm(s)\times\circ\mathrm{d}W_h(s)-m_h^{j+\frac{1}{2}}\times\widehat{\Delta}_jW_h,\theta_h^{j+1}\right\rangle_h. \tag*{(90)} 1 Ω Λ , j ⟨ θ h j + 1 − θ h j , θ h j + 1 ⟩ h + c 0 k 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 ≤ C Λ k 1 Ω Λ , j θ h j + 1 h 2 + C Λ k ν = 1 ∑ 3 ( I h ( m j + 1 − m j ) 1 , h 2 ν + m h j + 1 − m h j 1 , h 2 ν ) + 1 Ω Λ , j R h j + 1 ( θ h j + 1 ) + 1 Ω Λ , j ⟨ ∫ t j t j + 1 I h m ( s ) × ∘ d W h ( s ) − m h j + 2 1 × Δ j W h , θ h j + 1 ⟩ h . (90)
By the definition of the discrete dual norm and Young’s inequality,
1 Ω Λ , j ∣ R h j + 1 ( θ h j + 1 ) ∣ ≤ c 0 4 k 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 + C k 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 + C k ∥ R h j + 1 ∥ − 1 , h 2 . 1_{\Omega_{\Lambda,j}}\left|\mathcal{R}_h^{j+1}(\theta_h^{j+1})\right|\leq\frac{c_0}{4}k1_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2+Ck1_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2+\frac{C}{k}\left\|\mathcal{R}_h^{j+1}\right\|_{-1,h}^2. 1 Ω Λ , j R h j + 1 ( θ h j + 1 ) ≤ 4 c 0 k 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 + C k 1 Ω Λ , j θ h j + 1 h 2 + k C R h j + 1 − 1 , h 2 .
After absorbing the first term on the right and summing (90) from j = 0 j=0 j = 0 to r − 1 r-1 r − 1 , we obtain
∑ j = 0 r − 1 1 Ω Λ , j ⟨ θ h j + 1 − θ h j , θ h j + 1 ⟩ h + c 1 k ∑ j = 0 r − 1 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 ≤ C Λ k ∑ j = 0 r − 1 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 + C Λ k ∑ j = 0 r − 1 ∑ ν = 1 3 ( ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν + ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ) + C ∑ j = 0 r − 1 1 k ∥ R h j + 1 ∥ − 1 , h 2 + S r , (91) \sum_{j=0}^{r-1}1_{\Omega_{\Lambda,j}}\left\langle\theta_h^{j+1}-\theta_h^j,\theta_h^{j+1}\right\rangle_h+c_1k\sum_{j=0}^{r-1}1_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2
\leq C_{\Lambda}k\sum_{j=0}^{r-1}1_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2+C_{\Lambda}k\sum_{j=0}^{r-1}\sum_{\nu=1}^{3}\left(\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}^{2\nu}+\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2\nu}\right)
+C\sum_{j=0}^{r-1}\frac{1}{k}\left\|\mathcal{R}_h^{j+1}\right\|_{-1,h}^2+S_r, \tag*{(91)} j = 0 ∑ r − 1 1 Ω Λ , j ⟨ θ h j + 1 − θ h j , θ h j + 1 ⟩ h + c 1 k j = 0 ∑ r − 1 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 ≤ C Λ k j = 0 ∑ r − 1 1 Ω Λ , j θ h j + 1 h 2 + C Λ k j = 0 ∑ r − 1 ν = 1 ∑ 3 ( I h ( m j + 1 − m j ) 1 , h 2 ν + m h j + 1 − m h j 1 , h 2 ν ) + C j = 0 ∑ r − 1 k 1 R h j + 1 − 1 , h 2 + S r , (91)
where c 1 > 0 c_1>0 c 1 > 0 and S r S_r S r is defined by (54) .
By Lemma A.3, noting the monotonicity in r r r of the gradient sum, we obtain for every 1 ≤ n ≤ J 1\leq n\leq J 1 ≤ n ≤ J ,
max 1 ≤ r ≤ n [ ∑ j = 0 r − 1 1 Ω Λ , j ⟨ θ h j + 1 − θ h j , θ h j + 1 ⟩ h + c 1 k ∑ j = 0 r − 1 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 ] ≥ c max 1 ≤ r ≤ n 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 + c ∑ j = 0 n − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 + c k ∑ j = 0 n − 1 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 . (92) \begin{aligned}
\max_{1\leq r\leq n}\left[\sum_{j=0}^{r-1}1_{\Omega_{\Lambda,j}}\left\langle\theta_h^{j+1}-\theta_h^j,\theta_h^{j+1}\right\rangle_h+c_1k\sum_{j=0}^{r-1}1_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2\right] \\
&\ge c\max_{1\le r\le n}\mathbf{1}_{\Omega_{\Lambda,r-1}}\|\theta_h^r\|_h^2+c\sum_{j=0}^{n-1}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}-\theta_h^j\|_h^2+ck\sum_{j=0}^{n-1}\mathbf{1}_{\Omega_{\Lambda,j}}\|\partial_x\theta_h^{j+1}\|_{L^2}^2. \tag*{(92)}
\end{aligned} 1 ≤ r ≤ n max [ j = 0 ∑ r − 1 1 Ω Λ , j ⟨ θ h j + 1 − θ h j , θ h j + 1 ⟩ h + c 1 k j = 0 ∑ r − 1 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 ] ≥ c 1 ≤ r ≤ n max 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 + c j = 0 ∑ n − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 + c k j = 0 ∑ n − 1 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 . (92)
We apply (92) to (91) , take the maximum over r = 1 , … , n r=1,\ldots,n r = 1 , … , n , and then take expectations. We also aim to apply (55) with ε = 1 − 2 γ \varepsilon=1-2\gamma ε = 1 − 2 γ to estimate the last term in (91) . Since θ h 0 = 0 \theta_h^0=0 θ h 0 = 0 and Ω Λ , j ⊂ Ω Λ , j − 1 \Omega_{\Lambda,j}\subset\Omega_{\Lambda,j-1} Ω Λ , j ⊂ Ω Λ , j − 1 for j ≥ 1 j\ge1 j ≥ 1 , we have
sup 0 ≤ j ≤ n 1 Ω Λ , j ∥ θ h j ∥ h 2 ≤ max 1 ≤ r ≤ n 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 . \sup_{0\le j\le n}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^j\|_h^2\le\max_{1\le r\le n}\mathbf{1}_{\Omega_{\Lambda,r-1}}\|\theta_h^r\|_h^2. 0 ≤ j ≤ n sup 1 Ω Λ , j ∥ θ h j ∥ h 2 ≤ 1 ≤ r ≤ n max 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 .
Therefore, the terms multiplied by δ \delta δ and η \eta η in (55) can be absorbed into, respectively, the discrete time-increment term and the stopped maximum on the left-hand side by choosing δ , η > 0 \delta,\eta>0 δ , η > 0 sufficiently small. We then infer that
E [ max 1 ≤ r ≤ n 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 ] + E [ ∑ j = 0 n − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + E [ k ∑ j = 0 n − 1 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 ] ≤ C Λ k ∑ j = 0 n − 1 E [ max 1 ≤ r ≤ j + 1 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 ] + C E [ ∑ j = 0 n − 1 1 k ∥ R h j + 1 ∥ − 1 , h 2 ] + C Λ E [ k ∑ j = 0 n − 1 ∑ ν = 1 3 ( ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν + ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ) ] + C Λ , γ k 2 γ ≤ C Λ k ∑ j = 0 n − 1 E [ max 1 ≤ r ≤ j + 1 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 ] + C Λ , γ ( h 2 + k 2 γ ) . (93) \begin{aligned}
&\mathbb{E}\left[\max_{1\le r\le n}\mathbf{1}_{\Omega_{\Lambda,r-1}}\|\theta_h^r\|_h^2\right]
+\mathbb{E}\left[\sum_{j=0}^{n-1}\mathbf{1}_{\Omega_{\Lambda,j}}\|\theta_h^{j+1}-\theta_h^j\|_h^2\right]
+\mathbb{E}\left[k\sum_{j=0}^{n-1}\mathbf{1}_{\Omega_{\Lambda,j}}\|\partial_x\theta_h^{j+1}\|_{L^2}^2\right]\\
&\le C_\Lambda k\sum_{j=0}^{n-1}\mathbb{E}\left[\max_{1\le r\le j+1}\mathbf{1}_{\Omega_{\Lambda,r-1}}\|\theta_h^r\|_h^2\right]
+C\mathbb{E}\left[\sum_{j=0}^{n-1}\frac{1}{k}\|\mathcal{R}_h^{j+1}\|_{-1,h}^2\right]\\
&\quad+C_\Lambda\mathbb{E}\left[k\sum_{j=0}^{n-1}\sum_{\nu=1}^{3}\left(\|I_h(m^{j+1}-m^j)\|_{1,h}^{2\nu}+\|m_h^{j+1}-m_h^j\|_{1,h}^{2\nu}\right)\right]+C_{\Lambda,\gamma}k^{2\gamma}\\
&\le C_\Lambda k\sum_{j=0}^{n-1}\mathbb{E}\left[\max_{1\le r\le j+1}\mathbf{1}_{\Omega_{\Lambda,r-1}}\|\theta_h^r\|_h^2\right]+C_{\Lambda,\gamma}\left(h^2+k^{2\gamma}\right). \tag*{(93)}
\end{aligned} E [ 1 ≤ r ≤ n max 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 ] + E [ j = 0 ∑ n − 1 1 Ω Λ , j ∥ θ h j + 1 − θ h j ∥ h 2 ] + E [ k j = 0 ∑ n − 1 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 ] ≤ C Λ k j = 0 ∑ n − 1 E [ 1 ≤ r ≤ j + 1 max 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 ] + C E [ j = 0 ∑ n − 1 k 1 ∥ R h j + 1 ∥ − 1 , h 2 ] + C Λ E [ k j = 0 ∑ n − 1 ν = 1 ∑ 3 ( ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν + ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ) ] + C Λ , γ k 2 γ ≤ C Λ k j = 0 ∑ n − 1 E [ 1 ≤ r ≤ j + 1 max 1 Ω Λ , r − 1 ∥ θ h r ∥ h 2 ] + C Λ , γ ( h 2 + k 2 γ ) . (93)
where in the last step we used Lemma 3.6 and (107) . Now, choose k Λ > 0 k_\Lambda>0 k Λ > 0 so that C Λ k Λ ≤ 1 2 C_\Lambda k_\Lambda\le\frac{1}{2} C Λ k Λ ≤ 2 1 . The term with j = n − 1 j=n-1 j = n − 1 can then be absorbed into the left-hand side. The discrete Gronwall inequality then yields (87) . Finally, (88) follows as a consequence of (42) , (43) , and (87) . □
Corollary 3.9 (Convergence rate in probability). Under the assumptions of Theorem 3.8, let γ ∈ ( 0 , 1 2 ) \gamma\in(0,\frac{1}{2}) γ ∈ ( 0 , 2 1 ) and define
E h , k : = max 0 ≤ j ≤ J ∥ m j − m h j ∥ L 2 2 + k ∑ j = 0 J − 1 ∥ ∂ x ( m j + 1 − m h j + 1 ) ∥ L 2 2 . \mathcal{E}_{h,k}:=\max_{0\le j\le J}\|m^j-m_h^j\|_{L^2}^2+k\sum_{j=0}^{J-1}\|\partial_x(m^{j+1}-m_h^{j+1})\|_{L^2}^2. E h , k := 0 ≤ j ≤ J max ∥ m j − m h j ∥ L 2 2 + k j = 0 ∑ J − 1 ∥ ∂ x ( m j + 1 − m h j + 1 ) ∥ L 2 2 .
Then
lim M → ∞ lim sup h , k → 0 P [ E h , k > M ( h 2 + k 2 γ ) ] = 0. (94) \lim_{M\to\infty}\limsup_{h,k\to0}\mathbb{P}\left[\mathcal{E}_{h,k}>M\left(h^2+k^{2\gamma}\right)\right]=0. \tag*{(94)} M → ∞ lim h , k → 0 lim sup P [ E h , k > M ( h 2 + k 2 γ ) ] = 0. (94)
Proof. Fix Λ > 0 \Lambda>0 Λ > 0 . Since k Λ > 0 k_\Lambda>0 k Λ > 0 is fixed once Λ \Lambda Λ is fixed, one has k ≤ k Λ k\le k_\Lambda k ≤ k Λ for all sufficiently small k k k . Hence, Theorem 3.8, estimate (53) , and Markov’s inequality give, for every M > 0 M>0 M > 0 ,
P [ E h , k > M ( h 2 + k 2 γ ) ] ≤ P [ Ω Λ , J ∁ ] + P [ 1 Ω Λ , J E h , k > M ( h 2 + k 2 γ ) ] ≤ C Λ + C Λ , γ M . \mathbb{P}\left[\mathcal{E}_{h,k}>M\left(h^2+k^{2\gamma}\right)\right]\le\mathbb{P}\left[\Omega_{\Lambda,J}^{\complement}\right]+\mathbb{P}\left[\mathbf{1}_{\Omega_{\Lambda,J}}\mathcal{E}_{h,k}>M\left(h^2+k^{2\gamma}\right)\right]\le\frac{C}{\Lambda}+\frac{C_{\Lambda,\gamma}}{M}. P [ E h , k > M ( h 2 + k 2 γ ) ] ≤ P [ Ω Λ , J ∁ ] + P [ 1 Ω Λ , J E h , k > M ( h 2 + k 2 γ ) ] ≤ Λ C + M C Λ , γ .
Therefore,
lim sup h , k → 0 P [ E h , k > M ( h 2 + k 2 γ ) ] ≤ C Λ . \limsup_{h,k\to0}\mathbb{P}\left[\mathcal{E}_{h,k}>M\left(h^2+k^{2\gamma}\right)\right]\le\frac{C}{\Lambda}. h , k → 0 lim sup P [ E h , k > M ( h 2 + k 2 γ ) ] ≤ Λ C .
Since Λ > 0 \Lambda>0 Λ > 0 is arbitrary, letting Λ → ∞ \Lambda\to\infty Λ → ∞ proves (94) . □
Appendix A. Auxiliary inequalities
Lemma A.1 (Discrete Jensen inequality). Let q ≥ 1 q \ge1 q ≥ 1 and k > 0 k > 0 k > 0 . Let { a j } j = 0 J − 1 \{a_j\}_{j=0}^{J-1} { a j } j = 0 J − 1 be a non-negative sequence and set T = J k T = Jk T = J k . Then
( ∑ j = 0 J − 1 k a j ) q ≤ T q − 1 ∑ j = 0 J − 1 k a j q . (95) \left(\sum_{j=0}^{J-1} ka_j\right)^q \le T^{q-1}\sum_{j=0}^{J-1} ka_j^q. \tag*{(95)} ( j = 0 ∑ J − 1 k a j ) q ≤ T q − 1 j = 0 ∑ J − 1 k a j q . (95)
The sign in (95) is reversed if q ∈ ( 0 , 1 ) q \in(0,1) q ∈ ( 0 , 1 ) .
Lemma A.2 (Discrete BDG inequality). Let p ≥ 2 p \ge2 p ≥ 2 , let H H H be a Hilbert space, and let { Z j } j = 0 J − 1 \{Z_j\}_{j=0}^{J-1} { Z j } j = 0 J − 1 be an H H H -valued martingale difference sequence with respect to a filtration { F j } j = 0 J \{\mathcal{F}_j\}_{j=0}^{J} { F j } j = 0 J , that is, Z j Z_j Z j is F j + 1 \mathcal{F}_{j+1} F j + 1 -measurable and
E [ Z j ∣ F j ] = 0 , j = 0 , … , J − 1. \mathbb{E}[Z_j \mid\mathcal{F}_j] = 0,\qquad j=0,\ldots,J-1. E [ Z j ∣ F j ] = 0 , j = 0 , … , J − 1.
Assume that Z j ∈ L p ( Ω ; H ) Z_j \in L^p(\Omega;H) Z j ∈ L p ( Ω ; H ) for each j j j . Then there exists a constant C p > 0 C_p > 0 C p > 0 , depending only on p p p , such that
( E [ max 0 ≤ r ≤ J ∥ ∑ j = 0 r − 1 Z j ∥ H p ] ) 1 / p ≤ C p ( ∑ j = 0 J − 1 ( E [ ∥ Z j ∥ H p ] ) 2 / p ) 1 / 2 . (96) \left(\mathbb{E}\left[\max_{0\le r\le J}\left\|\sum_{j=0}^{r-1} Z_j\right\|_H^p\right]\right)^{1/p}
\le C_p\left(\sum_{j=0}^{J-1}\left(\mathbb{E}\left[\|Z_j\|_H^p\right]\right)^{2/p}\right)^{1/2}. \tag*{(96)} E 0 ≤ r ≤ J max j = 0 ∑ r − 1 Z j H p 1/ p ≤ C p ( j = 0 ∑ J − 1 ( E [ ∥ Z j ∥ H p ] ) 2/ p ) 1/2 . (96)
Lemma A.3. Let { Ω j } j = 0 J \{\Omega_j\}_{j=0}^{J} { Ω j } j = 0 J be a sequence of decreasing events. For any sequence { v j } j = 0 J ⊂ V h \{v^j\}_{j=0}^{J} \subset V_h { v j } j = 0 J ⊂ V h and any 1 ≤ n ≤ J 1 \le n \le J 1 ≤ n ≤ J ,
max 1 ≤ r ≤ n ∑ j = 0 r − 1 1 Ω j ⟨ v j + 1 − v j , v j + 1 ⟩ h ≥ 1 4 max 1 ≤ r ≤ n 1 Ω r − 1 ∥ v r ∥ h 2 + 1 4 ∑ j = 0 n − 1 1 Ω j ∥ v j + 1 − v j ∥ h 2 − 1 2 1 Ω 0 ∥ v 0 ∥ h 2 . \max_{1\le r\le n}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_j}\langle v^{j+1}-v^j,v^{j+1}\rangle_h
\ge\frac{1}{4}\max_{1\le r\le n}\mathbf{1}_{\Omega_{r-1}}\|v^r\|_h^2+\frac{1}{4}\sum_{j=0}^{n-1}\mathbf{1}_{\Omega_j}\|v^{j+1}-v^j\|_h^2-\frac{1}{2}\mathbf{1}_{\Omega_0}\|v^0\|_h^2. 1 ≤ r ≤ n max j = 0 ∑ r − 1 1 Ω j ⟨ v j + 1 − v j , v j + 1 ⟩ h ≥ 4 1 1 ≤ r ≤ n max 1 Ω r − 1 ∥ v r ∥ h 2 + 4 1 j = 0 ∑ n − 1 1 Ω j ∥ v j + 1 − v j ∥ h 2 − 2 1 1 Ω 0 ∥ v 0 ∥ h 2 .
Proof. For each 1 ≤ r ≤ n 1 \le r \le n 1 ≤ r ≤ n , the identity
2 ⟨ v j + 1 − v j , v j + 1 ⟩ h = ∥ v j + 1 ∥ h 2 − ∥ v j ∥ h 2 + ∥ v j + 1 − v j ∥ h 2 2\langle v^{j+1}-v^j,v^{j+1}\rangle_h=\|v^{j+1}\|_h^2-\|v^j\|_h^2+\|v^{j+1}-v^j\|_h^2 2 ⟨ v j + 1 − v j , v j + 1 ⟩ h = ∥ v j + 1 ∥ h 2 − ∥ v j ∥ h 2 + ∥ v j + 1 − v j ∥ h 2
and summation by parts give
2 ∑ j = 0 r − 1 1 Ω j ⟨ v j + 1 − v j , v j + 1 ⟩ h = 1 Ω r − 1 ∥ v r ∥ h 2 − 1 Ω 0 ∥ v 0 ∥ h 2 + ∑ j = 1 r − 1 ( 1 Ω j − 1 − 1 Ω j ) ∥ v j ∥ h 2 + ∑ j = 0 r − 1 1 Ω j ∥ v j + 1 − v j ∥ h 2 . 2\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_j}\langle v^{j+1}-v^j,v^{j+1}\rangle_h
=\mathbf{1}_{\Omega_{r-1}}\|v^r\|_h^2-\mathbf{1}_{\Omega_0}\|v^0\|_h^2
+\sum_{j=1}^{r-1}(\mathbf{1}_{\Omega_{j-1}}-\mathbf{1}_{\Omega_j})\|v^j\|_h^2
+\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_j}\|v^{j+1}-v^j\|_h^2. 2 j = 0 ∑ r − 1 1 Ω j ⟨ v j + 1 − v j , v j + 1 ⟩ h = 1 Ω r − 1 ∥ v r ∥ h 2 − 1 Ω 0 ∥ v 0 ∥ h 2 + j = 1 ∑ r − 1 ( 1 Ω j − 1 − 1 Ω j ) ∥ v j ∥ h 2 + j = 0 ∑ r − 1 1 Ω j ∥ v j + 1 − v j ∥ h 2 .
Since the events are decreasing, the penultimate term on the right is nonnegative. Hence,
∑ j = 0 r − 1 1 Ω j ⟨ v j + 1 − v j , v j + 1 ⟩ h + 1 2 1 Ω 0 ∥ v 0 ∥ h 2 ≥ 1 2 1 Ω r − 1 ∥ v r ∥ h 2 + 1 2 ∑ j = 0 r − 1 1 Ω j ∥ v j + 1 − v j ∥ h 2 . \sum_{j=0}^{r-1}\mathbf{1}_{\Omega_j}\langle v^{j+1}-v^j,v^{j+1}\rangle_h+\frac{1}{2}\mathbf{1}_{\Omega_0}\|v^0\|_h^2
\ge\frac{1}{2}\mathbf{1}_{\Omega_{r-1}}\|v^r\|_h^2+\frac{1}{2}\sum_{j=0}^{r-1}\mathbf{1}_{\Omega_j}\|v^{j+1}-v^j\|_h^2. j = 0 ∑ r − 1 1 Ω j ⟨ v j + 1 − v j , v j + 1 ⟩ h + 2 1 1 Ω 0 ∥ v 0 ∥ h 2 ≥ 2 1 1 Ω r − 1 ∥ v r ∥ h 2 + 2 1 j = 0 ∑ r − 1 1 Ω j ∥ v j + 1 − v j ∥ h 2 .
Taking the maximum over r = 1 , … , n r=1,\ldots,n r = 1 , … , n proves the claim. □
Appendix B. Technical consistency estimates
Lemma B.1. Assume the hypotheses of Proposition 3.4 hold. Then, for every integer p ≥ 1 p \ge1 p ≥ 1 ,
E [ ( ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ 1 , h 2 ) p ] ≤ C p . (97) \mathbb{E}\left[\left(\sum_{j=0}^{J-1}\|m_h^{j+1}-m_h^j\|_{1,h}^2\right)^p\right]\le C_p. \tag*{(97)} E [ ( j = 0 ∑ J − 1 ∥ m h j + 1 − m h j ∥ 1 , h 2 ) p ] ≤ C p . (97)
Proof. The nodewise estimate (35) gives
∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ h 2 ≤ C k ( k ∑ j = 0 J − 1 ∥ m h j + 1 2 × Δ h m h j + 1 ∥ h 2 ) + C ∑ j = 0 J − 1 ∥ Δ ^ j W h ∥ L ∞ 2 . \sum_{j=0}^{J-1}\left\|m_h^{j+1}-m_h^j\right\|_h^2 \le C k \left(k\sum_{j=0}^{J-1}\left\|m_h^{j+\frac{1}{2}}\times\Delta_hm_h^{j+1}\right\|_h^2\right)+C\sum_{j=0}^{J-1}\left\|\widehat{\Delta}_jW_h\right\|_{L^\infty}^2. j = 0 ∑ J − 1 m h j + 1 − m h j h 2 ≤ C k ( k j = 0 ∑ J − 1 m h j + 2 1 × Δ h m h j + 1 h 2 ) + C j = 0 ∑ J − 1 Δ j W h L ∞ 2 .
The first term on the right has moments of every order by Proposition 3.4. Moreover, the discrete Jensen inequality and (28) imply that
E [ ( ∑ j = 0 J − 1 ∥ Δ ^ j W h ∥ L ∞ 2 ) p ] ≤ J p − 1 ∑ j = 0 J − 1 E [ ∥ Δ ^ j W h ∥ L ∞ 2 p ] ≤ C p . \mathbb{E}\left[\left(\sum_{j=0}^{J-1}\left\|\widehat{\Delta}_jW_h\right\|_{L^\infty}^2\right)^p\right]\le J^{p-1}\sum_{j=0}^{J-1}\mathbb{E}\left[\left\|\widehat{\Delta}_jW_h\right\|_{L^\infty}^{2p}\right]\le C_p. E [ ( j = 0 ∑ J − 1 Δ j W h L ∞ 2 ) p ] ≤ J p − 1 j = 0 ∑ J − 1 E [ Δ j W h L ∞ 2 p ] ≤ C p .
All moments of the gradient-increment contribution is controlled by the second term in (34) . These imply (97) for p ≥ 2 p\ge2 p ≥ 2 . The case p = 1 p=1 p = 1 follows from the case p = 2 p=2 p = 2 and Jensen’s inequality. □
In the following, we will establish several technical lemmas controlling the midpoint defect in the drift term A h \mathcal{A}_h A h .
Lemma B.2. Let A h \mathcal{A}_h A h be the form given by (86) and Ω Λ , j \Omega_{\Lambda,j} Ω Λ , j be defined by (52) . For every Λ > 0 \Lambda>0 Λ > 0 and η > 0 \eta>0 η > 0 , there exists C Λ , η > 0 C_{\Lambda,\eta}>0 C Λ , η > 0 , independent of h h h , k k k , and j j j , such that
1 Ω Λ , j [ A h ( I h m h j + 1 ; I h m h j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) ] ≥ ( α κ − η ) 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 − C Λ , η 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 − C Λ , η ( ∥ I h ( m h j + 1 − m j ) ∥ 1 , h 4 + ∥ m h j + 1 − m h j ∥ 1 , h 4 ) . (98) \begin{aligned}
\mathbf{1}_{\Omega_{\Lambda,j}}\left[\mathcal{A}_h(I_hm_h^{j+1};I_hm_h^{j+1},\theta_h^{j+1})-\mathcal{A}_h(m_h^{j+1};m_h^{j+1},\theta_h^{j+1})\right]
&\ge(\alpha\kappa-\eta)\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2-C_{\Lambda,\eta}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\\
&\quad-C_{\Lambda,\eta}\left(\left\|I_h(m_h^{j+1}-m^j)\right\|_{1,h}^4+\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^4\right). \tag*{(98)}
\end{aligned} 1 Ω Λ , j [ A h ( I h m h j + 1 ; I h m h j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) ] ≥ ( α κ − η ) 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 − C Λ , η 1 Ω Λ , j θ h j + 1 h 2 − C Λ , η ( I h ( m h j + 1 − m j ) 1 , h 4 + m h j + 1 − m h j 1 , h 4 ) . (98)
Proof. Note that both I h m h j + 1 I_hm_h^{j+1} I h m h j + 1 and m h j + 1 m_h^{j+1} m h j + 1 have unit length at every node. Hence, we have the identities
θ h j + 1 ( z ) ⋅ I h m h j + 1 ( z ) = 1 2 ∣ θ h j + 1 ( z ) ∣ 2 , θ h j + 1 ( z ) ⋅ m h j + 1 ( z ) = − 1 2 ∣ θ h j + 1 ( z ) ∣ 2 , ∀ z ∈ N h . \theta_h^{j+1}(z)\cdot I_hm_h^{j+1}(z)=\frac{1}{2}\left|\theta_h^{j+1}(z)\right|^2,\qquad\theta_h^{j+1}(z)\cdot m_h^{j+1}(z)=-\frac{1}{2}\left|\theta_h^{j+1}(z)\right|^2,\qquad\forall z\in\mathcal{N}_h. θ h j + 1 ( z ) ⋅ I h m h j + 1 ( z ) = 2 1 θ h j + 1 ( z ) 2 , θ h j + 1 ( z ) ⋅ m h j + 1 ( z ) = − 2 1 θ h j + 1 ( z ) 2 , ∀ z ∈ N h .
Using the triple product identity a × ( a × z ) = ( a ⋅ z ) a − z \boldsymbol{a}\times(\boldsymbol{a}\times\boldsymbol{z})=(\boldsymbol{a}\cdot\boldsymbol{z})\boldsymbol{a}-\boldsymbol{z} a × ( a × z ) = ( a ⋅ z ) a − z for ∣ a ∣ = 1 |\boldsymbol{a}|=1 ∣ a ∣ = 1 , we obtain
⟨ I h m h j + 1 × ( I h m h j + 1 × Δ h I h m h j + 1 ) , θ h j + 1 ⟩ h − ⟨ m h j + 1 × ( m h j + 1 × Δ h m h j + 1 ) , θ h j + 1 ⟩ h = − ∥ ∂ x θ h j + 1 ∥ L 2 2 + 1 2 ⟨ Δ h I h m h j + 1 , I h ( ∣ θ h j + 1 ∣ 2 I h m h j + 1 ) ⟩ h + 1 2 ⟨ Δ h m h j + 1 , I h ( ∣ θ h j + 1 ∣ 2 m h j + 1 ) ⟩ h . (99) \begin{aligned}
\left\langle I_hm_h^{j+1}\times\left(I_hm_h^{j+1}\times\Delta_hI_hm_h^{j+1}\right),\theta_h^{j+1}\right\rangle_h
-\left\langle m_h^{j+1}\times\left(m_h^{j+1}\times\Delta_hm_h^{j+1}\right),\theta_h^{j+1}\right\rangle_h
&=-\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2\\
&\quad+\frac{1}{2}\left\langle\Delta_hI_hm_h^{j+1},I_h\left(\left|\theta_h^{j+1}\right|^2I_hm_h^{j+1}\right)\right\rangle_h\\
&\quad+\frac{1}{2}\left\langle\Delta_hm_h^{j+1},I_h\left(\left|\theta_h^{j+1}\right|^2m_h^{j+1}\right)\right\rangle_h. \tag*{(99)}
\end{aligned} ⟨ I h m h j + 1 × ( I h m h j + 1 × Δ h I h m h j + 1 ) , θ h j + 1 ⟩ h − ⟨ m h j + 1 × ( m h j + 1 × Δ h m h j + 1 ) , θ h j + 1 ⟩ h = − ∂ x θ h j + 1 L 2 2 + 2 1 ⟨ Δ h I h m h j + 1 , I h ( θ h j + 1 2 I h m h j + 1 ) ⟩ h + 2 1 ⟨ Δ h m h j + 1 , I h ( θ h j + 1 2 m h j + 1 ) ⟩ h . (99)
The elementwise product estimate (22) and the nodal unit-length bound (31) give
∥ ∂ x I h ( ∣ θ h j + 1 ∣ 2 I h m h j + 1 ) ∥ L 2 ≤ C ∥ θ h j + 1 ∥ L ∞ ∥ ∂ x θ h j + 1 ∥ L 2 + C ∥ θ h j + 1 ∥ L ∞ 2 ∥ ∂ x I h m h j + 1 ∥ L 2 , \left\|\partial_xI_h\left(\left|\theta_h^{j+1}\right|^2I_hm_h^{j+1}\right)\right\|_{L^2}\le C\left\|\theta_h^{j+1}\right\|_{L^\infty}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}+C\left\|\theta_h^{j+1}\right\|_{L^\infty}^2\left\|\partial_xI_hm_h^{j+1}\right\|_{L^2}, ∂ x I h ( θ h j + 1 2 I h m h j + 1 ) L 2 ≤ C θ h j + 1 L ∞ ∂ x θ h j + 1 L 2 + C θ h j + 1 L ∞ 2 ∂ x I h m h j + 1 L 2 ,
and analogously with I h m h j + 1 I_hm_h^{j+1} I h m h j + 1 replaced by m h j + 1 m_h^{j+1} m h j + 1 . Therefore, using the definition of Δ h \Delta_h Δ h in (15) to discretely integrate by parts,
∣ 1 2 ⟨ Δ h I h m h j + 1 , I h ( ∣ θ h j + 1 ∣ 2 I h m h j + 1 ) ⟩ h + 1 2 ⟨ Δ h m h j + 1 , I h ( ∣ θ h j + 1 ∣ 2 m h j + 1 ) ⟩ h ∣ ≤ C ( ∥ ∂ x I h m h j + 1 ∥ L 2 + ∥ ∂ x m h j + 1 ∥ L 2 ) ∥ θ h j + 1 ∥ L ∞ ∥ ∂ x θ h j + 1 ∥ L 2 + C ( ∥ ∂ x I h m h j + 1 ∥ L 2 2 + ∥ ∂ x m h j + 1 ∥ L 2 2 ) ∥ θ h j + 1 ∥ L ∞ 2 ≤ η ∥ ∂ x θ h j + 1 ∥ L 2 2 + C η ( 1 + ∥ ∂ x I h m h j + 1 ∥ L 2 4 + ∥ ∂ x m h j + 1 ∥ L 2 4 ) ∥ θ h j + 1 ∥ h 2 . (100) \begin{aligned}
\left|\frac{1}{2}\left\langle\Delta_hI_hm_h^{j+1},I_h\left(\left|\theta_h^{j+1}\right|^2I_hm_h^{j+1}\right)\right\rangle_h+\frac{1}{2}\left\langle\Delta_hm_h^{j+1},I_h\left(\left|\theta_h^{j+1}\right|^2m_h^{j+1}\right)\right\rangle_h\right|
&\le C\left(\left\|\partial_xI_hm_h^{j+1}\right\|_{L^2}+\left\|\partial_xm_h^{j+1}\right\|_{L^2}\right)\left\|\theta_h^{j+1}\right\|_{L^\infty}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}\\
&\quad+C\left(\left\|\partial_xI_hm_h^{j+1}\right\|_{L^2}^2+\left\|\partial_xm_h^{j+1}\right\|_{L^2}^2\right)\left\|\theta_h^{j+1}\right\|_{L^\infty}^2\\
&\le\eta\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2+C_\eta\left(1+\left\|\partial_xI_hm_h^{j+1}\right\|_{L^2}^4+\left\|\partial_xm_h^{j+1}\right\|_{L^2}^4\right)\left\|\theta_h^{j+1}\right\|_h^2. \tag*{(100)}
\end{aligned} 2 1 ⟨ Δ h I h m h j + 1 , I h ( θ h j + 1 2 I h m h j + 1 ) ⟩ h + 2 1 ⟨ Δ h m h j + 1 , I h ( θ h j + 1 2 m h j + 1 ) ⟩ h ≤ C ( ∂ x I h m h j + 1 L 2 + ∂ x m h j + 1 L 2 ) θ h j + 1 L ∞ ∂ x θ h j + 1 L 2 + C ( ∂ x I h m h j + 1 L 2 2 + ∂ x m h j + 1 L 2 2 ) θ h j + 1 L ∞ 2 ≤ η ∂ x θ h j + 1 L 2 2 + C η ( 1 + ∂ x I h m h j + 1 L 2 4 + ∂ x m h j + 1 L 2 4 ) θ h j + 1 h 2 . (100)
where in the last step we used the discrete Agmon inequality (14) and Young’s inequality.
For the precession part, we write
I h m j + 1 × Δ h I h m j + 1 − m h j + 1 × Δ h m h j + 1 = I h m j + 1 × Δ h θ h j + 1 + θ h j + 1 × Δ h m h j + 1 . I_h m^{j+1} \times\Delta_h I_h m^{j+1} - m_h^{j+1} \times\Delta_h m_h^{j+1} = I_h m^{j+1} \times\Delta_h \theta_h^{j+1} + \theta_h^{j+1} \times\Delta_h m_h^{j+1}. I h m j + 1 × Δ h I h m j + 1 − m h j + 1 × Δ h m h j + 1 = I h m j + 1 × Δ h θ h j + 1 + θ h j + 1 × Δ h m h j + 1 .
The second term is orthogonal to θ h j + 1 \theta_h^{j+1} θ h j + 1 at every node. Hence, we obtain by (22) ,
⟨ I h m j + 1 × Δ h I h m j + 1 − m h j + 1 × Δ h m h j + 1 , θ h j + 1 ⟩ h = − ⟨ ∂ x θ h j + 1 , ∂ x I h ( θ h j + 1 × I h m j + 1 ) ⟩ ≤ ∥ ∂ x I h m j + 1 ∥ L 2 ∥ θ h j + 1 ∥ L ∞ ∥ ∂ x θ h j + 1 ∥ L 2 ≤ η ∥ ∂ x θ h j + 1 ∥ L 2 2 + C η ( 1 + ∥ ∂ x I h m j + 1 ∥ L 2 2 ) ∥ θ h j + 1 ∥ h 2 , (101) \begin{aligned}
\left\langle I_hm^{j+1} \times\Delta_hI_hm^{j+1} - m_h^{j+1} \times\Delta_hm_h^{j+1},\theta_h^{j+1}\right\rangle_h
&= -\left\langle\partial_x\theta_h^{j+1},\partial_xI_h\left(\theta_h^{j+1}\times I_hm^{j+1}\right)\right\rangle\\
&\le\left\|\partial_xI_hm^{j+1}\right\|_{L^2}\left\|\theta_h^{j+1}\right\|_{L^\infty}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2} \\
&\le\eta\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2+C_\eta\left(1+\left\|\partial_xI_hm^{j+1}\right\|_{L^2}^2\right)\left\|\theta_h^{j+1}\right\|_h^2, \tag*{(101)}
\end{aligned} ⟨ I h m j + 1 × Δ h I h m j + 1 − m h j + 1 × Δ h m h j + 1 , θ h j + 1 ⟩ h = − ⟨ ∂ x θ h j + 1 , ∂ x I h ( θ h j + 1 × I h m j + 1 ) ⟩ ≤ ∂ x I h m j + 1 L 2 θ h j + 1 L ∞ ∂ x θ h j + 1 L 2 ≤ η ∂ x θ h j + 1 L 2 2 + C η ( 1 + ∂ x I h m j + 1 L 2 2 ) θ h j + 1 h 2 , (101)
where in the last step we again applied the discrete Agmon and Young inequalities. Now, on Ω Λ , j \Omega_{\Lambda,j} Ω Λ , j ,
∥ ∂ x I h m j + 1 ∥ L 2 4 ≤ C Λ + C ∥ I h ( m j + 1 − m j ) ∥ 1 , h 4 , \left\|\partial_xI_hm^{j+1}\right\|_{L^2}^4\le C_\Lambda+C\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}^4, ∂ x I h m j + 1 L 2 4 ≤ C Λ + C I h ( m j + 1 − m j ) 1 , h 4 ,
∥ ∂ x m h j + 1 ∥ L 2 4 ≤ C Λ + C ∥ m h j + 1 − m h j ∥ 1 , h 4 . \left\|\partial_xm_h^{j+1}\right\|_{L^2}^4\le C_\Lambda+C\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^4. ∂ x m h j + 1 L 2 4 ≤ C Λ + C m h j + 1 − m h j 1 , h 4 .
Moreover, the nodal unit-length property implies ∥ θ h j + 1 ∥ h 2 ≤ 4 ∣ D ∣ \left\|\theta_h^{j+1}\right\|_h^2\le4|D| θ h j + 1 h 2 ≤ 4∣ D ∣ . Combining these estimates with eq:3 –(101) , and relabelling η \eta η , proves (4) . □
Lemma B.3. Let A h \mathcal{A}_h A h be the form given by (86) and Ω Λ , j \Omega_{\Lambda,j} Ω Λ , j be defined by (52) . For every Λ > 0 \Lambda>0 Λ > 0 and η > 0 \eta>0 η > 0 , there exists C Λ , η > 0 C_{\Lambda,\eta}>0 C Λ , η > 0 , independent of h h h , k k k , and j j j , such that
1 Ω Λ , j ∣ A h ( I h m ‾ j + 1 2 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) ∣ ≤ η 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 + C Λ , η 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 + C Λ , η ∑ ν = 1 3 ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν , (102) \begin{aligned}
\mathbf{1}_{\Omega_{\Lambda,j}}\left|\mathcal{A}_h\left(I_h\overline{m}^{j+\frac12};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(I_hm^{j+1};I_hm^{j+1},\theta_h^{j+1}\right)\right|\\
\le\eta\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2+C_{\Lambda,\eta}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2+C_{\Lambda,\eta}\sum_{\nu=1}^{3}\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}^{2\nu}, \tag*{(102)}
\end{aligned} 1 Ω Λ , j A h ( I h m j + 2 1 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) ≤ η 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 + C Λ , η 1 Ω Λ , j θ h j + 1 h 2 + C Λ , η ν = 1 ∑ 3 I h ( m j + 1 − m j ) 1 , h 2 ν , (102)
and
1 Ω Λ , j ∣ A h ( m h j + 1 2 ; m h j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) ∣ ≤ η 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 + C Λ , η 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 + C Λ , η ∑ ν = 1 3 ∥ m h j + 1 − m h j ∥ 1 , h 2 ν . (103) \begin{aligned}
\mathbf{1}_{\Omega_{\Lambda,j}}\left|\mathcal{A}_h\left(m_h^{j+\frac12};m_h^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(m_h^{j+1};m_h^{j+1},\theta_h^{j+1}\right)\right|\\
\le\eta\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2+C_{\Lambda,\eta}\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2+C_{\Lambda,\eta}\sum_{\nu=1}^{3}\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2\nu}. \tag*{(103)}
\end{aligned} 1 Ω Λ , j A h ( m h j + 2 1 ; m h j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) ≤ η 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 + C Λ , η 1 Ω Λ , j θ h j + 1 h 2 + C Λ , η ν = 1 ∑ 3 m h j + 1 − m h j 1 , h 2 ν . (103)
Proof. We prove (102) , while noting that (103) follows by the same argument.
By noting that I h m ‾ j + 1 2 = I h m j + 1 − 1 2 I h ( m j + 1 − m j ) I_h\overline{m}^{j+\frac12}=I_hm^{j+1}-\frac12I_h(m^{j+1}-m^j) I h m j + 2 1 = I h m j + 1 − 2 1 I h ( m j + 1 − m j ) , we can then write
A h ( I h m ‾ j + 1 2 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) = κ 2 ⟨ I h ( m j + 1 − m j ) × Δ h I h m j + 1 , θ h j + 1 ⟩ h − α κ 2 ⟨ I h ( m j + 1 − m j ) × ( I h m j + 1 × Δ h I h m j + 1 ) , θ h j + 1 ⟩ h − α κ 2 ⟨ I h m j + 1 × ( I h ( m j + 1 − m j ) × Δ h I h m j + 1 ) , θ h j + 1 ⟩ h + α κ 4 ⟨ I h ( m j + 1 − m j ) × ( I h ( m j + 1 − m j ) × Δ h I h m j + 1 ) , θ h j + 1 ⟩ h . (104) \begin{aligned}
\mathcal{A}_h\left(I_h\overline{m}^{j+\frac12};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(I_hm^{j+1};I_hm^{j+1},\theta_h^{j+1}\right)
&=\frac{\kappa}{2}\left\langle I_h(m^{j+1}-m^j)\times\Delta_hI_hm^{j+1},\theta_h^{j+1}\right\rangle_h\\
&\quad-\frac{\alpha\kappa}{2}\left\langle I_h(m^{j+1}-m^j)\times\left(I_hm^{j+1}\times\Delta_hI_hm^{j+1}\right),\theta_h^{j+1}\right\rangle_h\\
&\quad-\frac{\alpha\kappa}{2}\left\langle I_hm^{j+1}\times\left(I_h(m^{j+1}-m^j)\times\Delta_hI_hm^{j+1}\right),\theta_h^{j+1}\right\rangle_h\\
&\quad+\frac{\alpha\kappa}{4}\left\langle I_h(m^{j+1}-m^j)\times\left(I_h(m^{j+1}-m^j)\times\Delta_hI_hm^{j+1}\right),\theta_h^{j+1}\right\rangle_h. \tag*{(104)}
\end{aligned} A h ( I h m j + 2 1 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) = 2 κ ⟨ I h ( m j + 1 − m j ) × Δ h I h m j + 1 , θ h j + 1 ⟩ h − 2 α κ ⟨ I h ( m j + 1 − m j ) × ( I h m j + 1 × Δ h I h m j + 1 ) , θ h j + 1 ⟩ h − 2 α κ ⟨ I h m j + 1 × ( I h ( m j + 1 − m j ) × Δ h I h m j + 1 ) , θ h j + 1 ⟩ h + 4 α κ ⟨ I h ( m j + 1 − m j ) × ( I h ( m j + 1 − m j ) × Δ h I h m j + 1 ) , θ h j + 1 ⟩ h . (104)
Using the definition of Δ h \Delta_h Δ h in (15) to integrate by parts discretely, together with the elementwise product estimate (22) and the nodal bound (31) , we obtain
∣ A h ( I h m ‾ j + 1 2 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) ∣ ≤ C ∥ ∂ x I h m j + 1 ∥ L 2 ∥ I h ( m j + 1 − m j ) ∥ L ∞ ∥ ∂ x θ h j + 1 ∥ L 2 + C ∥ ∂ x I h m j + 1 ∥ L 2 2 ∥ ∂ x θ h j + 1 ∥ L ∞ ∥ ∂ x I h ( m j + 1 − m j ) ∥ L 2 + C ∥ ∂ x I h m j + 1 ∥ L 2 2 ∥ θ h j + 1 ∥ L ∞ ∥ I h ( m j + 1 − m j ) ∥ L ∞ . (105) \begin{aligned}
\left|\mathcal{A}_h\left(I_h\overline{m}^{j+\frac{1}{2}};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(I_hm^{j+1};I_hm^{j+1},\theta_h^{j+1}\right)\right|
&\le C\left\|\partial_xI_hm^{j+1}\right\|_{L^2}\left\|I_h(m^{j+1}-m^j)\right\|_{L^\infty}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}\\
&\quad+C\left\|\partial_xI_hm^{j+1}\right\|_{L^2}^2\left\|\partial_x\theta_h^{j+1}\right\|_{L^\infty}\left\|\partial_xI_h(m^{j+1}-m^j)\right\|_{L^2}\\
&\quad+C\left\|\partial_xI_hm^{j+1}\right\|_{L^2}^2\left\|\theta_h^{j+1}\right\|_{L^\infty}\left\|I_h(m^{j+1}-m^j)\right\|_{L^\infty}. \tag*{(105)}
\end{aligned} A h ( I h m j + 2 1 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) ≤ C ∂ x I h m j + 1 L 2 I h ( m j + 1 − m j ) L ∞ ∂ x θ h j + 1 L 2 + C ∂ x I h m j + 1 L 2 2 ∂ x θ h j + 1 L ∞ ∂ x I h ( m j + 1 − m j ) L 2 + C ∂ x I h m j + 1 L 2 2 θ h j + 1 L ∞ I h ( m j + 1 − m j ) L ∞ . (105)
Now, on Ω Λ , j \Omega_{\Lambda,j} Ω Λ , j , we can estimate
∥ ∂ x I h m j + 1 ∥ L 2 ≤ C Λ + ∥ I h ( m j + 1 − m j ) ∥ 1 , h , \left\|\partial_xI_hm^{j+1}\right\|_{L^2}\le C_\Lambda+\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}, ∂ x I h m j + 1 L 2 ≤ C Λ + I h ( m j + 1 − m j ) 1 , h ,
while applying the discrete Agmon inequality to the terms with L ∞ L^\infty L ∞ norms. Young’s inequality applied term by term in (105) then yields eq:B.6 .
To show eq:B.7 , we use m h j + 1 2 = m h j + 1 − 1 2 ( m h j + 1 − m h j ) m_h^{j+\frac{1}{2}}=m_h^{j+1}-\frac{1}{2}(m_h^{j+1}-m_h^j) m h j + 2 1 = m h j + 1 − 2 1 ( m h j + 1 − m h j ) , noting that
∥ ∂ x m h j + 1 ∥ L 2 ≤ C Λ + ∥ m h j + 1 − m h j ∥ 1 , h \left\|\partial_xm_h^{j+1}\right\|_{L^2}\le C_\Lambda+\left\|m_h^{j+1}-m_h^j\right\|_{1,h} ∂ x m h j + 1 L 2 ≤ C Λ + m h j + 1 − m h j 1 , h
on Ω Λ , j \Omega_{\Lambda,j} Ω Λ , j , then repeat the same argument. This completes the proof. □
The above two lemmas are used to derive the following estimate.
Lemma B.4. Let A h \mathcal{A}_h A h be the form given by (86) and Ω Λ , j \Omega_{\Lambda,j} Ω Λ , j be defined by (52) . For every fixed Λ > 0 \Lambda>0 Λ > 0 , there exist c 0 > 0 c_0>0 c 0 > 0 and C Λ > 0 C_\Lambda>0 C Λ > 0 , independent of h h h , k k k , and j j j , such that
k 1 Ω Λ , j [ A h ( I h m ‾ j + 1 2 ; I h m j + 1 , θ h j + 1 ) − A h ( m h j + 1 2 ; m h j + 1 , θ h j + 1 ) ] ≥ c 0 k 1 Ω Λ , j ∥ ∂ x θ h j + 1 ∥ L 2 2 − C Λ k 1 Ω Λ , j ∥ θ h j + 1 ∥ h 2 − C Λ k ∑ ν = 1 3 ( ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν + ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ) . (106) \begin{aligned}
k\mathbf{1}_{\Omega_{\Lambda,j}}\left[\mathcal{A}_h\left(I_h\overline{m}^{j+\frac{1}{2}};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(m_h^{j+\frac{1}{2}};m_h^{j+1},\theta_h^{j+1}\right)\right]
&\ge c_0k\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\partial_x\theta_h^{j+1}\right\|_{L^2}^2-C_\Lambda k\mathbf{1}_{\Omega_{\Lambda,j}}\left\|\theta_h^{j+1}\right\|_h^2\\
&\quad-C_\Lambda k\sum_{\nu=1}^{3}\left(\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}^{2\nu}+\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2\nu}\right). \tag*{(106)}
\end{aligned} k 1 Ω Λ , j [ A h ( I h m j + 2 1 ; I h m j + 1 , θ h j + 1 ) − A h ( m h j + 2 1 ; m h j + 1 , θ h j + 1 ) ] ≥ c 0 k 1 Ω Λ , j ∂ x θ h j + 1 L 2 2 − C Λ k 1 Ω Λ , j θ h j + 1 h 2 − C Λ k ν = 1 ∑ 3 ( I h ( m j + 1 − m j ) 1 , h 2 ν + m h j + 1 − m h j 1 , h 2 ν ) . (106)
Moreover, for every γ ∈ ( 0 , 1 2 ) \gamma\in\left(0,\frac{1}{2}\right) γ ∈ ( 0 , 2 1 ) , there exists C γ > 0 C_\gamma>0 C γ > 0 such that
E [ k ∑ j = 0 J − 1 ∑ ν = 1 3 ( ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν + ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ) ] ≤ C γ k 2 γ . (107) \mathbb{E}\left[k\sum_{j=0}^{J-1}\sum_{\nu=1}^{3}\left(\left\|I_h(m^{j+1}-m^j)\right\|_{1,h}^{2\nu}+\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2\nu}\right)\right]\le C_\gamma k^{2\gamma}. \tag*{(107)} E [ k j = 0 ∑ J − 1 ν = 1 ∑ 3 ( I h ( m j + 1 − m j ) 1 , h 2 ν + m h j + 1 − m h j 1 , h 2 ν ) ] ≤ C γ k 2 γ . (107)
Proof. We add and subtract the two endpoint drift terms to write
A h ( I h m ‾ j + 1 2 ; I h m j + 1 , θ h j + 1 ) − A h ( m h j + 1 2 ; m h j + 1 , θ h j + 1 ) = A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) + [ A h ( I h m ‾ j + 1 2 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) ] − [ A h ( m h j + 1 2 ; m h j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) ] . \begin{aligned}
\mathcal{A}_h\left(I_h\overline{m}^{j+\frac{1}{2}};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(m_h^{j+\frac{1}{2}};m_h^{j+1},\theta_h^{j+1}\right)
&=\mathcal{A}_h\left(I_hm^{j+1};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(m_h^{j+1};m_h^{j+1},\theta_h^{j+1}\right)\\
&\quad+\left[\mathcal{A}_h\left(I_h\overline{m}^{j+\frac{1}{2}};I_hm^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(I_hm^{j+1};I_hm^{j+1},\theta_h^{j+1}\right)\right]\\
&\quad-\left[\mathcal{A}_h\left(m_h^{j+\frac{1}{2}};m_h^{j+1},\theta_h^{j+1}\right)-\mathcal{A}_h\left(m_h^{j+1};m_h^{j+1},\theta_h^{j+1}\right)\right].
\end{aligned} A h ( I h m j + 2 1 ; I h m j + 1 , θ h j + 1 ) − A h ( m h j + 2 1 ; m h j + 1 , θ h j + 1 ) = A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) + [ A h ( I h m j + 2 1 ; I h m j + 1 , θ h j + 1 ) − A h ( I h m j + 1 ; I h m j + 1 , θ h j + 1 ) ] − [ A h ( m h j + 2 1 ; m h j + 1 , θ h j + 1 ) − A h ( m h j + 1 ; m h j + 1 , θ h j + 1 ) ] .
We can now apply Lemma B.2 to the first term on the right-hand side and Lemma B.3 to the two remaining differences. Choosing η > 0 \eta>0 η > 0 sufficiently small, for instance 3 η ≤ 1 2 α κ 3\eta\le\frac{1}{2}\alpha\kappa 3 η ≤ 2 1 α κ , gives (106) with c 0 = 1 2 α κ c_0=\frac{1}{2}\alpha\kappa c 0 = 2 1 α κ after enlarging C Λ C_\Lambda C Λ . It remains to prove (107) . By the H 1 \mathrm{H}^{1} H 1 -stability of I h I_h I h and (7) , for ν = 1 , 2 , 3 \nu=1,2,3 ν = 1 , 2 , 3 , we have
E [ k ∑ j = 0 J − 1 ∥ I h ( m j + 1 − m j ) ∥ 1 , h 2 ν ] ≤ C E [ k ∑ j = 0 J − 1 ∥ m j + 1 − m j ∥ H 1 2 ν ] ≤ C γ k 2 ν γ ≤ C γ k 2 γ . (108) \mathbb{E}\left[k\sum_{j=0}^{J-1}\left\|I_h\left(m^{j+1}-m^j\right)\right\|_{1,h}^{2\nu}\right]\leq C\mathbb{E}\left[k\sum_{j=0}^{J-1}\left\|m^{j+1}-m^j\right\|_{\mathrm{H}^{1}}^{2\nu}\right]\leq C_{\gamma}k^{2\nu\gamma}\leq C_{\gamma}k^{2\gamma}. \tag*{(108)} E [ k j = 0 ∑ J − 1 I h ( m j + 1 − m j ) 1 , h 2 ν ] ≤ C E [ k j = 0 ∑ J − 1 m j + 1 − m j H 1 2 ν ] ≤ C γ k 2 ν γ ≤ C γ k 2 γ . (108)
Moreover, for the numerical increments, by Lemma B.1 we have
E [ k ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ 1 , h 2 ν ] ≤ k E [ ( k ∑ j = 0 J − 1 ∥ m h j + 1 − m h j ∥ 1 , h 2 ) ν ] ≤ C ν k . (109) \mathbb{E}\left[k\sum_{j=0}^{J-1}\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2\nu}\right]\leq k\mathbb{E}\left[\left(k\sum_{j=0}^{J-1}\left\|m_h^{j+1}-m_h^j\right\|_{1,h}^{2}\right)^{\nu}\right]\leq C_{\nu}k. \tag*{(109)} E [ k j = 0 ∑ J − 1 m h j + 1 − m h j 1 , h 2 ν ] ≤ k E [ ( k j = 0 ∑ J − 1 m h j + 1 − m h j 1 , h 2 ) ν ] ≤ C ν k . (109)
Since 0 < k ≤ 1 0<k\leq1 0 < k ≤ 1 and 2 γ < 1 2\gamma<1 2 γ < 1 , the required estimate (107) follows from (108) and (109) . □
References
Institute of Analysis and Scientific Computing, TU Wien, Wiedner Hauptstrasse 8–10, 1040 Vienna, Austria
Email address: agus.soenjaya@asc.tuwien.ac.at