Introduction
Elliptic boundary value problems with heterogeneous diffusion and reaction coefficients arise throughout scientific computing and engineering. In many applications these coefficients are not known exactly and are more naturally modelled as random fields. The computational task then becomes the estimation of expectations and other statistics of quantities of interest depending on the solution of a parameter-dependent PDE. For problems of realistic size, the dominant cost lies in the repeated numerical solution of the forward model.
Standard Monte Carlo methods are robust but become prohibitively expensive once fine spatial resolution is required. Multilevel Monte Carlo (MLMC) methods address this by combining many inexpensive coarse solves with far fewer expensive fine solves, and can reduce total work to near that of a single deterministic solve at the target accuracy [Hei01; Gil08; Gil15; BSZ11]. For elliptic PDEs with random coefficients, MLMC theory has been developed for conforming finite element discretisations on nested mesh hierarchies, including pathwise discretisation bounds and complexity results for uniformly elliptic and lognormal models [13 , 15 Te+13] .
Two related issues arise when MLMC is combined with spatially represented random fields on general meshes. The first is geometric, the forward discretisations may be non-nested or may be constructed by agglomeration. White-noise sampling and MLMC coupling across nested and non-nested finite element hierarchies have been developed in [14 ] , while recent work has combined MLMC with Virtual Element discretisations on general polytopal meshes using agglomerated coarse spaces [4 ] . The second issue is the representation of the random input itself. When a coefficient is generated through a spatially discretised latent field, its approximation becomes part of the overall discretisation error and its inter-level coupling directly affects the variance of the multilevel corrections. SPDE-based sampling of random-field coefficients has been studied in the multilevel quasi-Monte Carlo setting [10 ] , including coupling on non-nested finite element meshes. The present work focuses on this latter coefficient-representation issue in the setting of polytopic DG discretisations, where strong pairwise coupling remains essential for MLMC efficiency [12 ] .
Discontinuous Galerkin methods on polytopic meshes provide an attractive discretisation route in such settings. They offer geometric flexibility while retaining robust stability and high-order approximation on general polytopic mesh families [11 , 9 ] . Furthermore, the cardinality of local element bases is independent of the element shape. More broadly, recent work on polytopic discretisations has shown that this geometric setting supports a wider family of numerical methods, including recovery-based formulations and their extension to polytopic meshes [22 , 16 ] . In a related direction, recent composite and nodally bound-preserving constructions underline that polytopic meshes can support stable and structure-aware discretisations beyond the standard interior penalty framework [1 ] . At the same time, agglomeration-based multilevel constructions for DG on polytopic meshes have been developed successfully for linear solver design, including multigrid and Schwarz methods [2 , 5 , 3 ] .
The present work is motivated by the need to develop an MLMC framework for random-coefficient elliptic problems on general polytopic DG meshes in which level coupling is induced by the underlying random input, rather than by remapping already discretised coefficient fields between meshes. We consider coercive elliptic diffusion–reaction problems with random diffusion and reaction coefficients, discretised by lifting-based symmetric interior penalty DG methods on polytopic meshes. The FFT construction can be used on unrelated forward meshes, while the SPDE construction is developed on agglomerated hierarchies with a natural coarse-to-fine inclusion.
Our first contribution is analytical. We derive a Strang-type decomposition in the DG norm that separates the forward discretisation error from the error introduced by approximating the random coefficients. The formulation is written using a discrete lifting, equivalently an L 2 L^2 L 2 -projected numerical flux, so that coefficient traces are avoided in the consistency analysis. This is the key structural point for the multilevel analysis. Once the coefficients are generated only approximately, whether by representation of a spectral sample or by the numerical solution of an auxiliary stochastic PDE, the total bias must distinguish approximation of the forward solution from approximation of the data driving the forward problem.
Our second contribution is to instantiate this coefficient-aware framework with two different coefficient-generation and coupling strategies on polytopic DG meshes. The first is a spectral method in which a stationary Gaussian latent field is generated on a fixed Cartesian reference grid using FFT and circulant-embedding ideas [31 , 17 , 23 ] and then evaluated at the quadrature points required for DG assembly. In this way, all levels are driven by the same represented latent-field realisation, without transferring a discretised coefficient field between the forward meshes. The second is an SPDE-based Matérn construction in which the latent field is defined through an auxiliary elliptic stochastic PDE driven by spatial white noise [Whi54; Mat86; LRL11]. On an agglomerated hierarchy, the levels are coupled through the driving noise itself using the exact discrete relation between fine and coarse white-noise loads. The SPDE route is also informed by DG analysis in mesh-dependent norms for rough data, which provides a useful analytical precedent for discretisations driven by irregular right-hand sides [21 ] . Modern developments in the SPDE viewpoint and in scalable Markov and graph representations provide further context for this route [29 , 25 , 6 ] .
Our third contribution is to combine these ingredients in a single coefficient-aware multilevel error analysis. For the forward problem we obtain DG-norm bounds in which the influence of coefficient approximation appears explicitly through L 1 L^1 L 1 -type norms of the coefficient error. For the spectral sampler these terms are related to the representation and interpolation error of the latent field. For the SPDE sampler they are related to approximation properties of the DG resolvent acting on spatial white noise. This yields a unified MLMC framework in which both coefficient-generation strategies can be analysed using the same forward DG technology.
The distinction from the closest existing multilevel constructions is therefore in the role played by the random coefficient. White-noise sampling and coupling for finite element SPDE discretisations, including non-nested hierarchies, were developed in [14 ] , and this approach was subsequently used for SPDE-generated random-field coefficients in a multilevel quasi–Monte Carlo framework [10 ] . More recently, [4 ] developed an MLMC Virtual Element method on general polytopal meshes with agglomerated coarse spaces. In the latter setting the random diffusion coefficient is treated as a given continuum realisation and the multilevel analysis concerns the approximation of the corresponding forward problem.
Here the spatial representation of the random coefficient is included explicitly in the multilevel discretisation. The lifting-based DG formulation gives a coefficient-aware Strang decomposition separating forward spatial approximation, projected-flux consistency and coefficient-representation error without requiring traces of the represented coefficient. This permits two different input-space couplings to be treated within the same analysis. In the FFT construction, all forward levels are driven by a common represented latent field and the forward meshes need not be nested. In the SPDE construction, the agglomerated DG spaces are naturally nested and the coarse white-noise load is obtained by exact restriction of the fine-level load. Thus the coefficient representation controls a separate contribution to the bias, while the coupling of the latent random input controls the strength of the multilevel corrections.
The numerical experiments complement the analysis on general Voronoi and agglomerated polytopic meshes. They verify the expected spatial rates of the P1 DG discretisation, isolate the coefficient-representation contribution predicted by the Strang decomposition, and measure the weak and variance decay of the multilevel corrections. The coupling diagnostics show strong correlation between neighbouring levels for both random-field constructions. The final work experiments compare complete MLMC and single-level Monte Carlo calculations and illustrate how the observed variance decay and sample costs determine the practical multilevel gain.
The remainder of the paper is organised as follows. Section 2 introduces the random elliptic diffusion–reaction problem, the admissible random-coefficient models and the quantities of interest. Section 3 formulates the lifting-based polytopic DG discretisation and derives the Strang-type decomposition separating spatial discretisation, projected-flux and coefficient-representation errors. Section 4 develops the two random-field constructions: the common-reference FFT sampler and the SPDE Matérn sampler with white-noise coupling on agglomerated hierarchies. Section 5 combines these approximation estimates with the MLMC bias–variance decomposition, derives the corresponding error and work estimates, and describes the optimal sample allocation. Finally, Section 6 presents numerical experiments on Voronoi and agglomerated polytopic meshes, verifying the forward and coefficient-representation errors, the strength of the multilevel couplings, and the resulting work–accuracy behaviour.
Problem setup
Throughout, L p ( Ω ) L^p(\Omega) L p ( Ω ) and H s ( Ω ) H^s(\Omega) H s ( Ω ) denote the usual Lebesgue and Sobolev spaces, and ( Ξ , F , P ) (\Xi,\mathcal{F},\mathbb{P}) ( Ξ , F , P ) is a complete probability space. Bochner spaces are denoted by L p ( Ξ ; X ) L^p(\Xi;X) L p ( Ξ ; X ) , with expectation E \mathbb{E} E and variance V \mathbb{V} V in the usual sense.
Elliptic model with random coefficients
Elliptic model with random coefficients. Let
V : = H 0 1 ( Ω ) . (1) V := H_0^1(\Omega). \tag*{(1)} V := H 0 1 ( Ω ) . (1)
We model the diffusion and reaction coefficients as random fields
A : Ω × Ξ → R s y m d × d , c : Ω × Ξ → R , (2) A:\Omega\times\Xi\to\mathbb{R}_{\mathrm{sym}}^{d\times d}, \qquad c:\Omega\times\Xi\to\mathbb{R}, \tag*{(2)} A : Ω × Ξ → R sym d × d , c : Ω × Ξ → R , (2)
measurable in ξ ∈ Ξ \xi\in\Xi ξ ∈ Ξ and essentially bounded in x ∈ Ω x\in\Omega x ∈ Ω . We assume that there exist deterministic constants
0 < a ≤ a ‾ < ∞ , 0 < c ‾ ≤ c ‾ < ∞ , (3) 0<a\le\underline{a}<\infty,\qquad0<\underline{c}\le\overline{c}<\infty, \tag*{(3)} 0 < a ≤ a < ∞ , 0 < c ≤ c < ∞ , (3)
such that
a ‾ ∣ η ∣ 2 ≤ η T A ( x , ξ ) η ≤ a ‾ ∣ η ∣ 2 , c ‾ ≤ c ( x , ξ ) ≤ c ‾ , (4) \underline{a}\lvert\eta\rvert^2\le\eta^\mathsf{T}A(x,\xi)\eta\le\overline{a}\lvert\eta\rvert^2,\qquad\underline{c}\le c(x,\xi)\le\overline{c}, \tag*{(4)} a ∣ η ∣ 2 ≤ η T A ( x , ξ ) η ≤ a ∣ η ∣ 2 , c ≤ c ( x , ξ ) ≤ c , (4)
for almost every x ∈ Ω x\in\Omega x ∈ Ω , every η ∈ R d \eta\in\mathbb{R}^d η ∈ R d , and almost every ξ ∈ Ξ \xi\in\Xi ξ ∈ Ξ .
Given F ∈ L 2 ( Ω ) F\in L^2(\Omega) F ∈ L 2 ( Ω ) , for almost every ξ ∈ Ξ \xi\in\Xi ξ ∈ Ξ we consider the problem: find u ( ξ ) ∈ V u(\xi)\in V u ( ξ ) ∈ V such that
a ( u ( ξ ) , v ; A ( ξ ) , c ( ξ ) ) = ⟨ F , v ⟩ ∀ v ∈ V , (5) a(u(\xi),v;A(\xi),c(\xi))=\langle F,v\rangle\qquad\forall v\in V, \tag*{(5)} a ( u ( ξ ) , v ; A ( ξ ) , c ( ξ )) = ⟨ F , v ⟩ ∀ v ∈ V , (5)
where
a ( w , v ; A , c ) : = ∫ Ω A ( x ) ∇ w ⋅ ∇ v d x + ∫ Ω c ( x ) w v d x . (6) a(w,v;A,c):=\int_{\Omega} A(x)\nabla w\cdot\nabla v\,\mathrm{d}x+\int_{\Omega} c(x)wv\,\mathrm{d}x. \tag*{(6)} a ( w , v ; A , c ) := ∫ Ω A ( x ) ∇ w ⋅ ∇ v d x + ∫ Ω c ( x ) w v d x . (6)
By (4) , a ( ⋅ , ⋅ ; A ( ξ ) , c ( ξ ) ) a(\cdot,\cdot;A(\xi),c(\xi)) a ( ⋅ , ⋅ ; A ( ξ ) , c ( ξ )) is uniformly coercive,
a ( v , v ; A ( ξ ) , c ( ξ ) ) ≥ min { a ‾ , c ‾ } ∥ v ∥ H 1 ( Ω ) 2 ∀ v ∈ V , (7) a(v,v;A(\xi),c(\xi))\ge\min\{\underline{a},\underline{c}\}\lVert v\rVert_{H^1(\Omega)}^2 \qquad\forall v\in V, \tag*{(7)} a ( v , v ; A ( ξ ) , c ( ξ )) ≥ min { a , c } ∥ v ∥ H 1 ( Ω ) 2 ∀ v ∈ V , (7)
and uniformly continuous,
∣ a ( w , v ; A ( ξ ) , c ( ξ ) ) ∣ ≤ max { a ‾ , c ‾ } ∥ w ∥ H 1 ( Ω ) ∥ v ∥ H 1 ( Ω ) ∀ w , v ∈ V . (8) \lvert a(w,v;A(\xi),c(\xi))\rvert\le\max\{\overline{a},\overline{c}\}\lVert w\rVert_{H^1(\Omega)}\lVert v\rVert_{H^1(\Omega)} \qquad\forall w,v\in V. \tag*{(8)} ∣ a ( w , v ; A ( ξ ) , c ( ξ ))∣ ≤ max { a , c } ∥ w ∥ H 1 ( Ω ) ∥ v ∥ H 1 ( Ω ) ∀ w , v ∈ V . (8)
Hence, by the Lax–Milgram theorem, (5) admits a unique solution for almost every ξ ∈ Ξ \xi\in\Xi ξ ∈ Ξ . Moreover, there exists a constant C s t a b > 0 C_{\mathrm{stab}}>0 C stab > 0 , depending only on a ‾ \underline{a} a , a ‾ \overline{a} a , c ‾ \underline{c} c , c ‾ \overline{c} c and Ω \Omega Ω , such that
∥ u ( ξ ) ∥ H 1 ( Ω ) ≤ C s t a b ∥ F ∥ L 2 ( Ω ) for a.e. ξ ∈ Ξ . (9) \lVert u(\xi)\rVert_{H^1(\Omega)}\le C_{\mathrm{stab}}\lVert F\rVert_{L^2(\Omega)} \qquad\text{for a.e. }\xi\in\Xi. \tag*{(9)} ∥ u ( ξ ) ∥ H 1 ( Ω ) ≤ C stab ∥ F ∥ L 2 ( Ω ) for a.e. ξ ∈ Ξ. (9)
2.2. Remark (Almost sure well-posedness and moment bounds). Under the stated measurability assumptions, the solution map u : Ξ → V u : \Xi\to V u : Ξ → V is strongly measurable. Since (9) is uniform in ξ \xi ξ ,
u ∈ L ∞ ( Ξ ; V ) , ∥ u ∥ L ∞ ( Ξ ; V ) ≤ C s t a b ∥ F ∥ L 2 ( Ω ) . (10) u \in L^\infty(\Xi;V), \qquad\lVert u\rVert_{L^\infty(\Xi;V)} \le C_{\mathrm{stab}}\lVert F\rVert_{L^2(\Omega)}. \tag*{(10)} u ∈ L ∞ ( Ξ ; V ) , ∥ u ∥ L ∞ ( Ξ ; V ) ≤ C stab ∥ F ∥ L 2 ( Ω ) . (10)
Consequently, u ∈ L p ( Ξ ; V ) u \in L^p(\Xi;V) u ∈ L p ( Ξ ; V ) for every p ∈ [ 1 , ∞ ) p \in[1,\infty) p ∈ [ 1 , ∞ ) and
E [ ∥ u ∥ H 1 ( Ω ) p ] < ∞ ∀ p ∈ [ 1 , ∞ ) . (11) \mathbb{E}\left[\lVert u\rVert_{H^1(\Omega)}^p\right] < \infty\qquad\forall p \in[1,\infty). \tag*{(11)} E [ ∥ u ∥ H 1 ( Ω ) p ] < ∞ ∀ p ∈ [ 1 , ∞ ) . (11)
Random input models
The analysis below uses only the pathwise bounds (4). We therefore allow different constructions of the random coefficients, provided these bounds hold uniformly, either directly or after a deterministic truncation. We record here the models used subsequently in the analysis and numerical experiments.
A convenient scalar model is obtained from truncated lognormal fields. Let Y a Y_a Y a and Y c Y_c Y c be mean-zero Gaussian fields, which may be correlated, and let m a , m c ∈ L ∞ ( Ω ) m_a,m_c \in L^\infty(\Omega) m a , m c ∈ L ∞ ( Ω ) be deterministic. Define
Π [ − β , β ] ( y ) : = min { β , max { − β , y } } . (12) \Pi_{[-\beta,\beta]}(y) := \min\{\beta,\max\{-\beta,y\}\}. \tag*{(12)} Π [ − β , β ] ( y ) := min { β , max { − β , y }} . (12)
For truncation parameters β a , β c > 0 \beta_a,\beta_c > 0 β a , β c > 0 , set
a ( x , ξ ) : = exp ( m a ( x ) + Π [ − β a , β a ] Y a ( x , ξ ) ) , (13) a(x,\xi) := \exp\left(m_a(x)+\Pi_{[-\beta_a,\beta_a]}Y_a(x,\xi)\right), \tag*{(13)} a ( x , ξ ) := exp ( m a ( x ) + Π [ − β a , β a ] Y a ( x , ξ ) ) , (13)
and
c ( x , ξ ) : = exp ( m c ( x ) + Π [ − β c , β c ] Y c ( x , ξ ) ) . (14) c(x,\xi) := \exp\left(m_c(x)+\Pi_{[-\beta_c,\beta_c]}Y_c(x,\xi)\right). \tag*{(14)} c ( x , ξ ) := exp ( m c ( x ) + Π [ − β c , β c ] Y c ( x , ξ ) ) . (14)
Then
a ‾ : = exp ( ess inf Ω m a − β a ) , a ‾ : = exp ( ess sup Ω m a + β a ) , (15) \underline{a} := \exp\left(\operatorname*{ess\,inf}_{\Omega}m_a-\beta_a\right), \qquad\overline{a} := \exp\left(\operatorname*{ess\,sup}_{\Omega}m_a+\beta_a\right), \tag*{(15)} a := exp ( Ω ess inf m a − β a ) , a := exp ( Ω ess sup m a + β a ) , (15)
and
c ‾ : = exp ( ess inf Ω m c − β c ) , c ‾ : = exp ( ess sup Ω m c + β c ) . (16) \underline{c} := \exp\left(\operatorname*{ess\,inf}_{\Omega}m_c-\beta_c\right), \qquad\overline{c} := \exp\left(\operatorname*{ess\,sup}_{\Omega}m_c+\beta_c\right). \tag*{(16)} c := exp ( Ω ess inf m c − β c ) , c := exp ( Ω ess sup m c + β c ) . (16)
Thus the resulting coefficients satisfy (4) almost surely.
2.4. Remark (Scalar diffusion). In the scalar case we take
A ( x , ξ ) = a ( x , ξ ) I . (17) A(x,\xi)=a(x,\xi)I. \tag*{(17)} A ( x , ξ ) = a ( x , ξ ) I . (17)
The diffusion condition in (4) then reduces to
a ‾ ≤ a ( x , ξ ) ≤ a ‾ . \underline{a}\le a(x,\xi)\le\overline{a}. a ≤ a ( x , ξ ) ≤ a .
This is the setting used in the numerical experiments below. A single latent field may be used for both coefficients by taking Y a = Y c Y_a=Y_c Y a = Y c , while distinct or correlated fields are equally admissible for the analysis.
Quantities of interest
We consider quantities of interest of the form
Q : V → R , Q ⋆ ( ξ ) : = Q ( u ( ξ ) ) . (18) Q:V\to\mathbb{R}, \qquad Q^\star(\xi):=Q(u(\xi)). \tag*{(18)} Q : V → R , Q ⋆ ( ξ ) := Q ( u ( ξ )) . (18)
For the multilevel analysis we focus on functionals which are Lipschitz continuous with respect to the H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) norm. The stability and forward approximation estimates can then be transferred directly to the bias and to the multilevel corrections.
A canonical example is a bounded linear functional
Q ( u ) : = ⟨ g , u ⟩ V ∗ , V , g ∈ V ∗ . (19) Q(u):=\langle g,u\rangle_{V^\ast,V}, \qquad g\in V^\ast. \tag*{(19)} Q ( u ) := ⟨ g , u ⟩ V ∗ , V , g ∈ V ∗ . (19)
Indeed,
∣ Q ( w ) − Q ( z ) ∣ ≤ ∥ g ∥ V ∗ ∥ w − z ∥ H 1 ( Ω ) ∀ w , z ∈ V . (20) \lvert Q(w)-Q(z)\rvert\le\lVert g\rVert_{V^\ast}\lVert w-z\rVert_{H^1(\Omega)} \qquad\forall w,z\in V. \tag*{(20)} ∣ Q ( w ) − Q ( z )∣ ≤ ∥ g ∥ V ∗ ∥ w − z ∥ H 1 ( Ω ) ∀ w , z ∈ V . (20)
In particular, if g ∈ L 2 ( Ω ) g\in L^2(\Omega) g ∈ L 2 ( Ω ) , then
Q ( u ) = ( g , u ) L 2 ( Ω ) . Q(u)=(g,u)_{L^2(\Omega)}. Q ( u ) = ( g , u ) L 2 ( Ω ) .
Another useful example is a regional average. If ω ⊂ Ω \omega\subset\Omega ω ⊂ Ω is measurable with ∣ ω ∣ > 0 \lvert\omega\rvert>0 ∣ ω ∣ > 0 , define
Q ( u ) : = 1 ∣ ω ∣ ∫ ω u ( x ) d x . (21) Q(u):=\frac{1}{\lvert\omega\rvert}\int_{\omega}u(x)\,\mathrm{d}x. \tag*{(21)} Q ( u ) := ∣ ω ∣ 1 ∫ ω u ( x ) d x . (21)
Then
∣ Q ( w ) − Q ( z ) ∣ ≤ ∣ ω ∣ − 1 / 2 ∥ w − z ∥ L 2 ( Ω ) ≤ ∣ ω ∣ − 1 / 2 ∥ w − z ∥ H 1 ( Ω ) . (22) \lvert Q(w)-Q(z)\rvert\leq\lvert\omega\rvert^{-1/2}\lVert w-z\rVert_{L^2(\Omega)} \leq\lvert\omega\rvert^{-1/2}\lVert w-z\rVert_{H^1(\Omega)} . \tag*{(22)} ∣ Q ( w ) − Q ( z )∣ ≤ ∣ ω ∣ − 1/2 ∥ w − z ∥ L 2 ( Ω ) ≤ ∣ ω ∣ − 1/2 ∥ w − z ∥ H 1 ( Ω ) . (22)
Since point evaluation is not a continuous functional on H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) when d ≥ 2 d \geq2 d ≥ 2 , local averages of the form (21) provide a convenient replacement whenever a spatially localised output is required.
Polytopic DG discretisation and Strang decomposition
We formulate the polytopic DG method independently of the particular sampling strategy used for the random coefficients. On each level ℓ \ell ℓ , the forward problem is assembled using the represented coefficient pair
A ~ ℓ ∈ L ∞ ( Ω ; R s y m d × d ) , c ~ ℓ ∈ L ∞ ( Ω ) . (23) \widetilde{A}_{\ell} \in L^\infty\left(\Omega;\mathbb{R}_{\mathrm{sym}}^{d \times d}\right), \qquad\widetilde{c}_{\ell} \in L^\infty(\Omega). \tag*{(23)} A ℓ ∈ L ∞ ( Ω ; R sym d × d ) , c ℓ ∈ L ∞ ( Ω ) . (23)
These are the coefficient fields actually seen by the DG discretisation. Their construction is discussed in § 4 .
Throughout this section we assume the uniform bounds
a ‾ ∣ ξ ∣ 2 ≤ ξ T A ~ ℓ ( x ) ξ ≤ a ‾ ∣ ξ ∣ 2 , 0 < c ‾ ≤ c ~ ℓ ( x ) ≤ c ‾ , (24) \underline{a}\lvert\xi\rvert^2 \leq\xi^{\mathsf{T}}\widetilde{A}_{\ell}(x)\xi\leq\overline{a}\lvert\xi\rvert^2, \qquad0 < \underline{c} \leq\widetilde{c}_{\ell}(x) \leq\overline{c}, \tag*{(24)} a ∣ ξ ∣ 2 ≤ ξ T A ℓ ( x ) ξ ≤ a ∣ ξ ∣ 2 , 0 < c ≤ c ℓ ( x ) ≤ c , (24)
for almost every x ∈ Ω x \in\Omega x ∈ Ω , every ξ ∈ R d \xi\in\mathbb{R}^d ξ ∈ R d and every ℓ = 0 , … , L \ell= 0,\ldots,L ℓ = 0 , … , L , with constants independent of ℓ \ell ℓ .
Polytopic meshes and broken spaces
Let { T ℓ } ℓ = 0 L \{\mathcal{T}_{\ell}\}_{\ell=0}^{L} { T ℓ } ℓ = 0 L be a family of partitions of Ω \Omega Ω into non-overlapping open polytopes K K K such that
Ω ‾ = ⋃ K ∈ T ℓ K ‾ , K ∩ K ′ = ∅ for K ≠ K ′ . (25) \overline{\Omega} = \bigcup_{K \in\mathcal{T}_{\ell}} \overline{K}, \qquad K \cap K' = \varnothing\quad\text{for } K \ne K'. \tag*{(25)} Ω = K ∈ T ℓ ⋃ K , K ∩ K ′ = ∅ for K = K ′ . (25)
For K ∈ T ℓ K \in\mathcal{T}_{\ell} K ∈ T ℓ , let
h K : = diam ( K ) , h_K := \operatorname{diam}(K), h K := diam ( K ) ,
and set
h ℓ : = max K ∈ T ℓ h K . (26) h_{\ell} := \max_{K \in\mathcal{T}_{\ell}} h_K. \tag*{(26)} h ℓ := K ∈ T ℓ max h K . (26)
Let E ℓ \mathcal{E}_{\ell} E ℓ denote the set of open ( d − 1 ) (d-1) ( d − 1 ) -dimensional faces of T ℓ \mathcal{T}_{\ell} T ℓ , decomposed into interior and boundary faces,
E ℓ i : = { e = ∂ K − ∩ ∂ K + : K ± ∈ T ℓ , K − ≠ K + } , E ℓ b : = { e ⊂ ∂ K ∩ ∂ Ω : K ∈ T ℓ } . (27) \mathcal{E}_{\ell}^{i} := \{e = \partial K^{-} \cap\partial K^{+} : K^{\pm} \in\mathcal{T}_{\ell},\ K^{-} \ne K^{+}\}, \qquad\mathcal{E}_{\ell}^{b} := \{e \subset\partial K \cap\partial\Omega: K \in\mathcal{T}_{\ell}\}. \tag*{(27)} E ℓ i := { e = ∂ K − ∩ ∂ K + : K ± ∈ T ℓ , K − = K + } , E ℓ b := { e ⊂ ∂ K ∩ ∂ Ω : K ∈ T ℓ } . (27)
For each e ∈ E ℓ i e \in\mathcal{E}_{\ell}^{i} e ∈ E ℓ i , fix a unit normal n e n_e n e pointing from K − K^{-} K − to K + K^{+} K + , while on e ∈ E ℓ b e \in\mathcal{E}_{\ell}^{b} e ∈ E ℓ b we take n e n_e n e to be the outward unit normal.
The face scale is defined in terms of the neighbouring element diameters rather than the face measure. We set
h e : = { 2 h K − h K + h K − + h K + , e = ∂ K − ∩ ∂ K + ∈ E ℓ i , h K , e ⊂ ∂ K ∈ E ℓ b . (28) h_e := \begin{cases}
\dfrac{2h_{K^{-}}h_{K^{+}}}{h_{K^{-}}+h_{K^{+}}}, & e = \partial K^{-} \cap\partial K^{+} \in\mathcal{E}_{\ell}^{i},\\
h_K, & e \subset\partial K \in\mathcal{E}_{\ell}^{b}.
\end{cases} \tag*{(28)} h e := ⎩ ⎨ ⎧ h K − + h K + 2 h K − h K + , h K , e = ∂ K − ∩ ∂ K + ∈ E ℓ i , e ⊂ ∂ K ∈ E ℓ b . (28)
In particular, the stabilisation does not degenerate when a polytope contains faces of small measure [7 ] .
We impose a uniform chunkiness condition on the elements. Namely, there exists ϑ ∗ ∈ ( 0 , 1 ] \vartheta_{*} \in(0,1] ϑ ∗ ∈ ( 0 , 1 ] , independent of ℓ \ell ℓ , such that for every K ∈ T ℓ K \in\mathcal{T}_{\ell} K ∈ T ℓ there exist x K ∈ K x_K \in K x K ∈ K and ρ K ≥ ϑ ∗ h K \rho_K \geq\vartheta_{*}h_K ρ K ≥ ϑ ∗ h K satisfying
B ( x K , ρ K ) ⊂ K , (29) B(x_K,\rho_K) \subset K, \tag*{(29)} B ( x K , ρ K ) ⊂ K , (29)
and K K K is star-shaped with respect to this ball, in the sense that
[ x , y ] ⊂ K ∀ x ∈ B ( x K , ρ K ) , ∀ y ∈ K . [x,y] \subset K \qquad\forall x \in B(x_K,\rho_K),\quad\forall y \in K. [ x , y ] ⊂ K ∀ x ∈ B ( x K , ρ K ) , ∀ y ∈ K .
No lower bound is imposed on the measure of an individual face, and the number of faces of an element is not assumed to be uniformly bounded.
For a scalar field v v v defined elementwise, let v ± v^{\pm} v ± denote its traces on e = ∂ K − ∩ ∂ K + ∈ E ℓ i e = \partial K^{-} \cap\partial K^{+} \in\mathcal{E}_{\ell}^{i} e = ∂ K − ∩ ∂ K + ∈ E ℓ i and define
[ [ v ] ] : = v − − v + , { { v } } : = 1 2 ( v − + v + ) . (30) [\![v]\!] := v^{-} - v^{+}, \qquad\{\!\{v\}\!\} := \frac{1}{2}(v^{-}+v^{+}). \tag*{(30)} [ [ v ] ] := v − − v + , { { v } } := 2 1 ( v − + v + ) . (30)
On boundary faces we set
⟦ v ⟧ : = v , { { v } } : = v . (31) \llbracket v \rrbracket:= v,\qquad\{\!\{v\}\!\} := v. \tag*{(31)} [ [ v ] ] := v , { { v } } := v . (31)
For vector fields, averages are taken componentwise.
Fix a polynomial degree p ≥ 1 p \ge1 p ≥ 1 . On each level define
V ℓ : = { v ∈ L 2 ( Ω ) : v ∣ K ∈ P p ( K ) for all K ∈ T ℓ } , (32) V_\ell:= \{v \in L^2(\Omega) : v|_K \in P_p(K)\text{ for all }K \in\mathcal{T}_\ell\}, \tag*{(32)} V ℓ := { v ∈ L 2 ( Ω ) : v ∣ K ∈ P p ( K ) for all K ∈ T ℓ } , (32)
with broken gradient
( ∇ ℓ v ) ∣ K : = ∇ ( v ∣ K ) . (33) (\nabla_\ell v)|_K := \nabla(v|_K). \tag*{(33)} ( ∇ ℓ v ) ∣ K := ∇ ( v ∣ K ) . (33)
We also use
H 1 ( T ℓ ) : = { v ∈ L 2 ( Ω ) : v ∣ K ∈ H 1 ( K ) for all K ∈ T ℓ } . (34) H^1(\mathcal{T}_\ell) := \{v \in L^2(\Omega) : v|_K \in H^1(K)\text{ for all }K \in\mathcal{T}_\ell\}. \tag*{(34)} H 1 ( T ℓ ) := { v ∈ L 2 ( Ω ) : v ∣ K ∈ H 1 ( K ) for all K ∈ T ℓ } . (34)
Lemma 3.2 (Trace and inverse inequalities). Assume (29). There exists C t r > 0 C_{\mathrm{tr}} > 0 C tr > 0 , depending only on ϑ ∗ \vartheta_* ϑ ∗ and the dimension, such that
∥ v ∥ L 2 ( ∂ K ) 2 ≤ C t r [ h K − 1 ∥ v ∥ L 2 ( K ) 2 + h K ∥ ∇ v ∥ L 2 ( K ) 2 ] ∀ v ∈ H 1 ( K ) . (35) \|v\|_{L^2(\partial K)}^2 \le C_{\mathrm{tr}}\left[h_K^{-1}\|v\|_{L^2(K)}^2+h_K\|\nabla v\|_{L^2(K)}^2\right]\qquad\forall v \in H^1(K). \tag*{(35)} ∥ v ∥ L 2 ( ∂ K ) 2 ≤ C tr [ h K − 1 ∥ v ∥ L 2 ( K ) 2 + h K ∥∇ v ∥ L 2 ( K ) 2 ] ∀ v ∈ H 1 ( K ) . (35) Moreover , for fixed p ≥ 1 p \ge1 p ≥ 1 there exists C i n v > 0 C_{\mathrm{inv}} > 0 C inv > 0 , independent of K K K and ℓ \ell ℓ , such that
∥ ∇ v ∥ L 2 ( K ) ≤ C i n v p 2 h K ∥ v ∥ L 2 ( K ) ∀ v ∈ P p ( K ) . (36) \|\nabla v\|_{L^2(K)} \le C_{\mathrm{inv}}\frac{p^2}{h_K}\|v\|_{L^2(K)}\qquad\forall v \in P_p(K). \tag*{(36)} ∥∇ v ∥ L 2 ( K ) ≤ C inv h K p 2 ∥ v ∥ L 2 ( K ) ∀ v ∈ P p ( K ) . (36) Consequently ,
∥ v ∥ L 2 ( ∂ K ) ≲ h K − 1 / 2 ∥ v ∥ L 2 ( K ) ∀ v ∈ P p ( K ) , (37) \|v\|_{L^2(\partial K)} \lesssim h_K^{-1/2}\|v\|_{L^2(K)}\qquad\forall v \in P_p(K), \tag*{(37)} ∥ v ∥ L 2 ( ∂ K ) ≲ h K − 1/2 ∥ v ∥ L 2 ( K ) ∀ v ∈ P p ( K ) , (37) where the hidden constant may depend on p p p but is independent of the number and relative sizes of the faces of K K K .
Lifting operator and the DG form
Introduce the broken vector-polynomial space
Σ ℓ : = { τ ∈ L 2 ( Ω ; R d ) : τ ∣ K ∈ [ P p ( K ) ] d for all K ∈ T ℓ } . (38) \Sigma_\ell:= \{\tau\in L^2(\Omega;\mathbb{R}^d) : \tau|_K \in[P_p(K)]^d\text{ for all }K \in\mathcal{T}_\ell\}. \tag*{(38)} Σ ℓ := { τ ∈ L 2 ( Ω ; R d ) : τ ∣ K ∈ [ P p ( K ) ] d for all K ∈ T ℓ } . (38)
The lifting
R ℓ : L 2 ( E ℓ ) ⟶ Σ ℓ \mathcal{R}_\ell: L^2(\mathcal{E}_\ell) \longrightarrow\Sigma_\ell R ℓ : L 2 ( E ℓ ) ⟶ Σ ℓ
is defined by
∫ Ω R ℓ ( ϕ ) ⋅ τ d x = − ∑ e ∈ E ℓ ∫ e ϕ { { τ ⋅ n e } } d s ∀ τ ∈ Σ ℓ . (39) \int_\Omega\mathcal{R}_\ell(\phi)\cdot\tau\,\mathrm{d}x
=
-\sum_{e\in\mathcal{E}_\ell}\int_e \phi\{\!\{\tau\cdot n_e\}\!\}\,\mathrm{d}s
\qquad\forall\tau\in\Sigma_\ell. \tag*{(39)} ∫ Ω R ℓ ( ϕ ) ⋅ τ d x = − e ∈ E ℓ ∑ ∫ e ϕ { { τ ⋅ n e } } d s ∀ τ ∈ Σ ℓ . (39)
Lemma 3.4 (Stability of the lifting). Under (29),
∥ R ℓ ( ϕ ) ∥ L 2 ( Ω ) ≲ ( ∑ e ∈ E ℓ h e − 1 ∥ ϕ ∥ L 2 ( e ) 2 ) 1 / 2 , (40) \|\mathcal{R}_\ell(\phi)\|_{L^2(\Omega)}
\lesssim
\left(\sum_{e\in\mathcal{E}_\ell}h_e^{-1}\|\phi\|_{L^2(e)}^2\right)^{1/2}, \tag*{(40)} ∥ R ℓ ( ϕ ) ∥ L 2 ( Ω ) ≲ ( e ∈ E ℓ ∑ h e − 1 ∥ ϕ ∥ L 2 ( e ) 2 ) 1/2 , (40) with a constant independent of the number and relative sizes of the faces.
Proof. By (39), weighted Cauchy–Schwarz gives
∣ ∫ Ω R ℓ ( ϕ ) ⋅ τ d x ∣ ≤ ( ∑ e ∈ E ℓ h e − 1 ∥ ϕ ∥ L 2 ( e ) 2 ) 1 / 2 ( ∑ e ∈ E ℓ h e ∥ { { τ ⋅ n e } } ∥ L 2 ( e ) 2 ) 1 / 2 . \left|\int_\Omega\mathcal{R}_\ell(\phi)\cdot\tau\,\mathrm{d}x\right|
\le
\left(\sum_{e\in\mathcal{E}_\ell}h_e^{-1}\|\phi\|_{L^2(e)}^2\right)^{1/2}
\left(\sum_{e\in\mathcal{E}_\ell}h_e\|\{\!\{\tau\cdot n_e\}\!\}\|_{L^2(e)}^2\right)^{1/2}. ∫ Ω R ℓ ( ϕ ) ⋅ τ d x ≤ ( e ∈ E ℓ ∑ h e − 1 ∥ ϕ ∥ L 2 ( e ) 2 ) 1/2 ( e ∈ E ℓ ∑ h e ∥ { { τ ⋅ n e } } ∥ L 2 ( e ) 2 ) 1/2 .
Since h e ≲ h K h_e\lesssim h_K h e ≲ h K for every face of K K K , summing over the complete element boundaries and using (37) yields
∑ e ∈ E ℓ h e ∥ { { τ ⋅ n e } } ∥ L 2 ( e ) 2 ≲ ∑ K ∈ T ℓ h K ∥ τ ∥ L 2 ( ∂ K ) 2 ≲ ∥ τ ∥ L 2 ( Ω ) 2 . \sum_{e\in\mathcal{E}_\ell}h_e\|\{\!\{\tau\cdot n_e\}\!\}\|_{L^2(e)}^2
\lesssim
\sum_{K\in\mathcal{T}_\ell}h_K\|\tau\|_{L^2(\partial K)}^2
\lesssim
\|\tau\|_{L^2(\Omega)}^2. e ∈ E ℓ ∑ h e ∥ { { τ ⋅ n e } } ∥ L 2 ( e ) 2 ≲ K ∈ T ℓ ∑ h K ∥ τ ∥ L 2 ( ∂ K ) 2 ≲ ∥ τ ∥ L 2 ( Ω ) 2 .
Taking the supremum over τ ∈ Σ ℓ ∖ { 0 } \tau\in\Sigma_\ell\setminus\{0\} τ ∈ Σ ℓ ∖ { 0 } proves the result. □
In particular,
∥ R ℓ ( ⟦ v ℓ ⟧ ) ∥ L 2 ( Ω ) ≲ ( ∑ e ∈ E ℓ h e − 1 ∥ ⟦ v ℓ ⟧ ∥ L 2 ( e ) 2 ) 1 / 2 . (41) \lVert R_\ell(\llbracket v_\ell\rrbracket)\rVert_{L^2(\Omega)} \lesssim\left(\sum_{e\in\mathcal{E}_\ell} h_e^{-1}\lVert\llbracket v_\ell\rrbracket\rVert_{L^2(e)}^2\right)^{1/2}. \tag*{(41)} ∥ R ℓ ( [ [ v ℓ ] ] ) ∥ L 2 ( Ω ) ≲ ( e ∈ E ℓ ∑ h e − 1 ∥ [ [ v ℓ ] ] ∥ L 2 ( e ) 2 ) 1/2 . (41)
For a symmetric positive-definite coefficient field B B B , define the elemental diffusion scale
α K ( B ) : = ess sup x ∈ K λ max ( B ( x ) ) (42) \alpha_K(B) := \operatorname*{ess\,sup}_{x\in K}\lambda_{\max}(B(x)) \tag*{(42)} α K ( B ) := x ∈ K ess sup λ m a x ( B ( x )) (42)
and the corresponding face scale
α e ( B ) : = { max { α K − ( B ) , α K + ( B ) } , e = ∂ K − ∩ ∂ K + ∈ E ℓ i , α K ( B ) , e ⊂ ∂ K ∈ E ℓ b . (43) \alpha_e(B) :=
\begin{cases}
\max\{\alpha_{K^-}(B),\alpha_{K^+}(B)\}, & e=\partial K^-\cap\partial K^+\in\mathcal{E}_\ell^i,\\
\alpha_K(B), & e\subset\partial K\in\mathcal{E}_\ell^b.
\end{cases} \tag*{(43)} α e ( B ) := { max { α K − ( B ) , α K + ( B )} , α K ( B ) , e = ∂ K − ∩ ∂ K + ∈ E ℓ i , e ⊂ ∂ K ∈ E ℓ b . (43)
For γ > 0 \gamma>0 γ > 0 , set
τ e ( B ) : = γ p 2 h e α e ( B ) . (44) \tau_e(B):=\gamma\frac{p^2}{h_e}\alpha_e(B). \tag*{(44)} τ e ( B ) := γ h e p 2 α e ( B ) . (44)
Let
V ℓ : = V + V ℓ . V_\ell:=V+V_\ell. V ℓ := V + V ℓ .
For w , v ∈ V ℓ w,v\in V_\ell w , v ∈ V ℓ , define the lifting-based symmetric interior-penalty form
a ℓ d i f f ( w , v ; B ) : = ∫ Ω B ∇ ℓ w ⋅ ∇ ℓ v d x + ∫ Ω B ∇ ℓ w ⋅ R ℓ ( ⟦ v ⟧ ) d x + ∫ Ω B R ℓ ( ⟦ w ⟧ ) ⋅ ∇ ℓ v d x + ∑ e ∈ E ℓ ∫ e τ e ( B ) ⟦ w ⟧ ⟦ v ⟧ d s . (45) a_\ell^{\mathrm{diff}}(w,v;B):=\int_\Omega B\nabla_\ell w\cdot\nabla_\ell v\,dx+\int_\Omega B\nabla_\ell w\cdot R_\ell(\llbracket v\rrbracket)\,dx
+\int_\Omega BR_\ell(\llbracket w\rrbracket)\cdot\nabla_\ell v\,dx+\sum_{e\in\mathcal{E}_\ell}\int_e\tau_e(B)\llbracket w\rrbracket\llbracket v\rrbracket\,ds. \tag*{(45)} a ℓ diff ( w , v ; B ) := ∫ Ω B ∇ ℓ w ⋅ ∇ ℓ v d x + ∫ Ω B ∇ ℓ w ⋅ R ℓ ( [ [ v ] ] ) d x + ∫ Ω B R ℓ ( [ [ w ] ] ) ⋅ ∇ ℓ v d x + e ∈ E ℓ ∑ ∫ e τ e ( B ) [ [ w ] ] [ [ v ] ] d s . (45)
The full diffusion–reaction form is
a ℓ D G ( w , v ; B , d ) : = a ℓ d i f f ( w , v ; B ) + ∫ Ω d w v d x . (46) a_\ell^{\mathrm{DG}}(w,v;B,d):=a_\ell^{\mathrm{diff}}(w,v;B)+\int_\Omega dwv\,dx. \tag*{(46)} a ℓ DG ( w , v ; B , d ) := a ℓ diff ( w , v ; B ) + ∫ Ω d w v d x . (46)
To relate (45) to the usual face-flux formulation, let
Π ℓ : L 2 ( Ω ; R d ) ⟶ Σ ℓ \Pi_\ell:L^2(\Omega;\mathbb{R}^d)\longrightarrow\Sigma_\ell Π ℓ : L 2 ( Ω ; R d ) ⟶ Σ ℓ
denote the elementwise L 2 L^2 L 2 -orthogonal projector. Since R ℓ ( ⟦ v ⟧ ) ∈ Σ ℓ R_\ell(\llbracket v\rrbracket)\in\Sigma_\ell R ℓ ( [ [ v ] ] ) ∈ Σ ℓ ,
∫ Ω B ∇ ℓ w ⋅ R ℓ ( ⟦ v ⟧ ) d x = ∫ Ω Π ℓ ( B ∇ ℓ w ) ⋅ R ℓ ( ⟦ v ⟧ ) d x = − ∑ e ∈ E ℓ ∫ e { { Π ℓ ( B ∇ ℓ w ) ⋅ n e } } ⟦ v ⟧ d s . (47) \int_\Omega B\nabla_\ell w\cdot R_\ell(\llbracket v\rrbracket)\,dx
=\int_\Omega\Pi_\ell(B\nabla_\ell w)\cdot R_\ell(\llbracket v\rrbracket)\,dx
=-\sum_{e\in\mathcal{E}_\ell}\int_e\{\!\{\Pi_\ell(B\nabla_\ell w)\cdot n_e\}\!\}\llbracket v\rrbracket\,ds. \tag*{(47)} ∫ Ω B ∇ ℓ w ⋅ R ℓ ( [ [ v ] ] ) d x = ∫ Ω Π ℓ ( B ∇ ℓ w ) ⋅ R ℓ ( [ [ v ] ] ) d x = − e ∈ E ℓ ∑ ∫ e { { Π ℓ ( B ∇ ℓ w ) ⋅ n e } } [ [ v ] ] d s . (47)
Thus (45) is precisely a symmetric interior-penalty form in which the numerical normal flux is the L 2 L^2 L 2 projection of the physical flux. In particular, no trace of the possibly rough coefficient B B B is required in the definition of the method.
The natural norm is
∥ v ∥ D G , ℓ ; B , d 2 : = ∥ B 1 / 2 ∇ ℓ v ∥ L 2 ( Ω ) 2 + ∥ d 1 / 2 v ∥ L 2 ( Ω ) 2 + ∑ e ∈ E ℓ ∥ τ e ( B ) 1 / 2 ⟦ v ⟧ ∥ L 2 ( e ) 2 . (48) \lVert v\rVert_{\mathrm{DG},\ell;B,d}^2:=\lVert B^{1/2}\nabla_\ell v\rVert_{L^2(\Omega)}^2+\lVert d^{1/2}v\rVert_{L^2(\Omega)}^2+\sum_{e\in\mathcal{E}_\ell}\lVert\tau_e(B)^{1/2}\llbracket v\rrbracket\rVert_{L^2(e)}^2. \tag*{(48)} ∥ v ∥ DG , ℓ ; B , d 2 := ∥ B 1/2 ∇ ℓ v ∥ L 2 ( Ω ) 2 + ∥ d 1/2 v ∥ L 2 ( Ω ) 2 + e ∈ E ℓ ∑ ∥ τ e ( B ) 1/2 [ [ v ] ] ∥ L 2 ( e ) 2 . (48)
Theorem (Coercivity and continuity)
Theorem (Coercivity and continuity). Assume (29) and (24) . There exists γ 0 > 0 \gamma_0>0 γ 0 > 0 such that, for every γ ≥ γ 0 \gamma\ge\gamma_0 γ ≥ γ 0 ,
a ℓ D G ( v ℓ , v ℓ ; A ~ ℓ , c ~ ℓ ) ≥ C c o e r ∥ v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ 2 ∀ v ℓ ∈ V ℓ , (49) a_\ell^{\mathrm{DG}}(v_\ell,v_\ell;\widetilde{A}_\ell,\widetilde{c}_\ell)\ge C_{\mathrm{coer}}\lVert v_\ell\rVert_{\mathrm{DG},\ell;\widetilde{A}_\ell,\widetilde{c}_\ell}^2\qquad\forall v_\ell\in V_\ell, \tag*{(49)} a ℓ DG ( v ℓ , v ℓ ; A ℓ , c ℓ ) ≥ C coer ∥ v ℓ ∥ DG , ℓ ; A ℓ , c ℓ 2 ∀ v ℓ ∈ V ℓ , (49)
and
∣ a ℓ D G ( w , v ; A ~ ℓ , c ~ ℓ ) ∣ ≤ C c o n t ∥ w ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ∥ v ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ (50) \left|a_\ell^{\mathrm{DG}}(w,v;\widetilde{A}_\ell,\widetilde{c}_\ell)\right|\le C_{\mathrm{cont}}\lVert w\rVert_{\mathrm{DG},\ell;\widetilde{A}_\ell,\widetilde{c}_\ell}\lVert v\rVert_{\mathrm{DG},\ell;\widetilde{A}_\ell,\widetilde{c}_\ell} \tag*{(50)} a ℓ DG ( w , v ; A ℓ , c ℓ ) ≤ C cont ∥ w ∥ DG , ℓ ; A ℓ , c ℓ ∥ v ∥ DG , ℓ ; A ℓ , c ℓ (50)
for all w , v ∈ V ℓ w,v\in V_\ell w , v ∈ V ℓ . The constants are independent of ℓ \ell ℓ and of the number and relative sizes of the faces.
Proof. The volume and reaction terms are controlled directly by (48) . For the lifting terms, (40) and the uniform coefficient bounds give
∣ ∫ Ω A ~ ℓ ∇ ℓ w ⋅ R ℓ ( [ v ] ) d x ∣ ≲ ∥ w ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ∥ v ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . \left|\int_{\Omega} \widetilde{A}_{\ell}\nabla_{\ell}w\cdot\mathcal{R}_{\ell}([v])\,\mathrm{d}x\right|\lesssim\|w\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\|v\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}. ∫ Ω A ℓ ∇ ℓ w ⋅ R ℓ ([ v ]) d x ≲ ∥ w ∥ DG , ℓ ; A ℓ , c ℓ ∥ v ∥ DG , ℓ ; A ℓ , c ℓ .
The same estimate holds with w w w and v v v interchanged, proving continuity. Taking w = v = v ℓ w=v=v_{\ell} w = v = v ℓ and applying Young’s inequality to the two identical lifting terms shows that they can be absorbed by the volume and penalty terms once γ \gamma γ is sufficiently large, which proves coercivity. □
Given L ℓ ∈ V ℓ ′ L_{\ell}\in V_{\ell}^{\prime} L ℓ ∈ V ℓ ′ , find U ℓ ∈ V ℓ U_{\ell}\in V_{\ell} U ℓ ∈ V ℓ such that
a ℓ D G ( U ℓ , v ℓ ; A ~ ℓ , c ~ ℓ ) = L ℓ ( v ℓ ) ∀ v ℓ ∈ V ℓ . (51) a_{\ell}^{\mathrm{DG}}(U_{\ell},v_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})=L_{\ell}(v_{\ell})\qquad\forall v_{\ell}\in V_{\ell}. \tag*{(51)} a ℓ DG ( U ℓ , v ℓ ; A ℓ , c ℓ ) = L ℓ ( v ℓ ) ∀ v ℓ ∈ V ℓ . (51)
Equip V ℓ ′ V_{\ell}^{\prime} V ℓ ′ with
∥ L ℓ ∥ V ℓ ′ : = sup v ℓ ∈ V ℓ ∖ { 0 } ∣ L ℓ ( v ℓ ) ∣ ∥ v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . (52) \|L_{\ell}\|_{V_{\ell}^{\prime}}:=\sup_{v_{\ell}\in V_{\ell}\setminus\{0\}}\frac{|L_{\ell}(v_{\ell})|}{\|v_{\ell}\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}. \tag*{(52)} ∥ L ℓ ∥ V ℓ ′ := v ℓ ∈ V ℓ ∖ { 0 } sup ∥ v ℓ ∥ DG , ℓ ; A ℓ , c ℓ ∣ L ℓ ( v ℓ ) ∣ . (52)
Corollary 3.6. (Discrete well-posedness and stability.) Under the hypotheses of Theorem 3.5, (51) has a unique solution and
∥ U ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ ∥ L ℓ ∥ V ℓ ′ . (53) \|U_{\ell}\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim\|L_{\ell}\|_{V_{\ell}^{\prime}}. \tag*{(53)} ∥ U ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ ∥ L ℓ ∥ V ℓ ′ . (53)
We therefore define
S ℓ A ~ ℓ , c ~ ℓ : V ℓ ′ ⟶ V ℓ , S ℓ A ~ ℓ , c ~ ℓ L ℓ : = U ℓ . (54) S_{\ell}^{\widetilde{A}_{\ell},\widetilde{c}_{\ell}}:V_{\ell}^{\prime}\longrightarrow V_{\ell},\qquad S_{\ell}^{\widetilde{A}_{\ell},\widetilde{c}_{\ell}}L_{\ell}:=U_{\ell}. \tag*{(54)} S ℓ A ℓ , c ℓ : V ℓ ′ ⟶ V ℓ , S ℓ A ℓ , c ℓ L ℓ := U ℓ . (54)
For the forward problem,
F ℓ ( v ℓ ) : = ⟨ F , v ℓ ⟩ = ∫ Ω F v ℓ d x . (55) F_{\ell}(v_{\ell}):=\langle F,v_{\ell}\rangle=\int_{\Omega}Fv_{\ell}\,\mathrm{d}x. \tag*{(55)} F ℓ ( v ℓ ) := ⟨ F , v ℓ ⟩ = ∫ Ω F v ℓ d x . (55)
The same discrete solution operator will later be used for the latent SPDE with a Gaussian load functional.
Strang decomposition
Let
A ∈ L ∞ ( Ω ; R s y m d × d ) , c ∈ L ∞ ( Ω ) , A\in L^{\infty}(\Omega;\mathbb{R}_{\mathrm{sym}}^{d\times d}),\qquad c\in L^{\infty}(\Omega), A ∈ L ∞ ( Ω ; R sym d × d ) , c ∈ L ∞ ( Ω ) ,
satisfy
a ‾ ∣ ξ ∣ 2 ≤ ξ T A ( x ) ξ ≤ a ‾ ∣ ξ ∣ 2 , 0 < c ‾ ≤ c ( x ) ≤ c ‾ . (56) \underline{a}|\xi|^{2}\leq\xi^{\mathsf{T}}A(x)\xi\leq\overline{a}|\xi|^{2},\qquad0<\underline{c}\leq c(x)\leq\overline{c}. \tag*{(56)} a ∣ ξ ∣ 2 ≤ ξ T A ( x ) ξ ≤ a ∣ ξ ∣ 2 , 0 < c ≤ c ( x ) ≤ c . (56)
Let u ∈ V u\in V u ∈ V solve
a ( u , v ; A , c ) = ⟨ F , v ⟩ ∀ v ∈ V , (57) a(u,v;A,c)=\langle F,v\rangle\qquad\forall v\in V, \tag*{(57)} a ( u , v ; A , c ) = ⟨ F , v ⟩ ∀ v ∈ V , (57)
or equivalently
u = S A , c F . (58) u=S^{A,c}F. \tag*{(58)} u = S A , c F . (58)
The discrete solution is
u ℓ : = S ℓ A ~ ℓ , c ~ ℓ F ℓ . (59) u_{\ell}:=S_{\ell}^{\widetilde{A}_{\ell},\widetilde{c}_{\ell}}F_{\ell}. \tag*{(59)} u ℓ := S ℓ A ℓ , c ℓ F ℓ . (59)
For z ∈ V ℓ z\in V_{\ell} z ∈ V ℓ and w ℓ ∈ V ℓ w_{\ell}\in V_{\ell} w ℓ ∈ V ℓ , define
C ℓ ( z ; w ℓ ) : = F ℓ ( w ℓ ) − a ℓ D G ( z , w ℓ ; A ~ ℓ , c ~ ℓ ) . (60) \mathcal{C}_{\ell}(z;w_{\ell}):=F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(z,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell}). \tag*{(60)} C ℓ ( z ; w ℓ ) := F ℓ ( w ℓ ) − a ℓ DG ( z , w ℓ ; A ℓ , c ℓ ) . (60)
Theorem 3.8. (Abstract Strang lemma.) Under the hypotheses of Theorem 3.5,
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ inf v ℓ ∈ V ℓ { ∥ u − v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ + sup w ℓ ∈ V ℓ ∖ { 0 } ∣ C ℓ ( v ℓ ; w ℓ ) ∣ ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ } . (61) \|u-u_{\ell}\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim\inf_{v_{\ell}\in V_{\ell}}\left\{\|u-v_{\ell}\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}+\sup_{w_{\ell}\in V_{\ell}\setminus\{0\}}\frac{|\mathcal{C}_{\ell}(v_{\ell};w_{\ell})|}{\|w_{\ell}\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}\right\}. \tag*{(61)} ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ v ℓ ∈ V ℓ inf { ∥ u − v ℓ ∥ DG , ℓ ; A ℓ , c ℓ + w ℓ ∈ V ℓ ∖ { 0 } sup ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ ∣ C ℓ ( v ℓ ; w ℓ ) ∣ } . (61)
There are now two distinct consistency effects. The first is the small inconsistency introduced by the lifted, equivalently projected, numerical flux. Define
D ℓ ( u ) : = sup w ℓ ∈ V ℓ ∖ { 0 } ∣ F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) ∣ ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . (62) \mathcal{D}_{\ell}(u) := \sup_{w_{\ell}\in V_{\ell}\setminus\{0\}} \frac{\left|F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)\right|}{\lVert w_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}. \tag*{(62)} D ℓ ( u ) := w ℓ ∈ V ℓ ∖ { 0 } sup ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) . (62)
The second is the error caused by replacing ( A , c ) (A,c) ( A , c ) by the represented pair ( A ~ ℓ , c ~ ℓ ) (\widetilde{A}_{\ell},\widetilde{c}_{\ell}) ( A ℓ , c ℓ ) :
E ℓ ( u ) : = sup w ℓ ∈ V ℓ ∖ { 0 } ∣ a ℓ D G ( u , w ℓ ; A , c ) − a ℓ D G ( u , w ℓ ; A ~ ℓ , c ~ ℓ ) ∣ ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . (63) \mathcal{E}_{\ell}(u) := \sup_{w_{\ell}\in V_{\ell}\setminus\{0\}} \frac{\left|a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})\right|}{\lVert w_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}. \tag*{(63)} E ℓ ( u ) := w ℓ ∈ V ℓ ∖ { 0 } sup ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ a ℓ DG ( u , w ℓ ; A , c ) − a ℓ DG ( u , w ℓ ; A ℓ , c ℓ ) . (63)
Theorem 3.9. (Strang decomposition with representation defect). Under the preceding assumptions,
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ inf v ℓ ∈ V ℓ ∥ u − v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ + D ℓ ( u ) + E ℓ ( u ) . (64) \lVert u-u_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}} \lesssim\inf_{v_{\ell}\in V_{\ell}}\lVert u-v_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}+\mathcal{D}_{\ell}(u)+\mathcal{E}_{\ell}(u). \tag*{(64)} ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ v ℓ ∈ V ℓ inf ∥ u − v ℓ ∥ DG , ℓ ; A ℓ , c ℓ + D ℓ ( u ) + E ℓ ( u ) . (64)
Proof. For arbitrary v ℓ , w ℓ ∈ V ℓ v_{\ell},w_{\ell}\in V_{\ell} v ℓ , w ℓ ∈ V ℓ ,
C ℓ ( v ℓ ; w ℓ ) = [ F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) ] + [ a ℓ D G ( u , w ℓ ; A , c ) − a ℓ D G ( u , w ℓ ; A ~ ℓ , c ~ ℓ ) ] + a ℓ D G ( u − v ℓ , w ℓ ; A ~ ℓ , c ~ ℓ ) . \mathcal{C}_{\ell}(v_{\ell};w_{\ell})=\left[F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)\right]+\left[a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})\right]+a_{\ell}^{\mathrm{DG}}(u-v_{\ell},w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell}). C ℓ ( v ℓ ; w ℓ ) = [ F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) ] + [ a ℓ DG ( u , w ℓ ; A , c ) − a ℓ DG ( u , w ℓ ; A ℓ , c ℓ ) ] + a ℓ DG ( u − v ℓ , w ℓ ; A ℓ , c ℓ ) .
Apply (50) , divide by the DG norm of w ℓ w_{\ell} w ℓ , take the supremum, and substitute the result into (61) . □
Approximation and lifting consistency
The regularity assumptions can be expressed directly in W 1 + s , r W^{1+s,r} W 1 + s , r rather than obtained indirectly from Sobolev embedding. For r ∈ [ 2 , ∞ ] r\in[2,\infty] r ∈ [ 2 , ∞ ] and s > 0 s>0 s > 0 , write
∥ z ∥ W s , r ( T ℓ ) : = ( ∑ K ∈ T ℓ ∥ z ∥ W s , r ( K ) r ) 1 / r , \lVert z\rVert_{W^{s,r}(\mathcal{T}_{\ell})}:=\left(\sum_{K\in\mathcal{T}_{\ell}}\lVert z\rVert_{W^{s,r}(K)}^{r}\right)^{1/r}, ∥ z ∥ W s , r ( T ℓ ) := ( K ∈ T ℓ ∑ ∥ z ∥ W s , r ( K ) r ) 1/ r ,
with the usual modification when r = ∞ r=\infty r = ∞ .
For 0 < s ≤ p 0<s\le p 0 < s ≤ p and u ∈ W 1 + s , r ( Ω ) u\in W^{1+s,r}(\Omega) u ∈ W 1 + s , r ( Ω ) , standard polynomial approximation on uniformly star-shaped polytopes gives
inf v ℓ ∈ V ℓ ∥ u − v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ h ℓ s ∥ u ∥ W 1 + s , r ( Ω ) . (65) \inf_{v_{\ell}\in V_{\ell}}\lVert u-v_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim h_{\ell}^{s}\lVert u\rVert_{W^{1+s,r}(\Omega)}. \tag*{(65)} v ℓ ∈ V ℓ inf ∥ u − v ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ h ℓ s ∥ u ∥ W 1 + s , r ( Ω ) . (65)
The lifting consistency error is controlled by best approximation of the physical flux q : = A ∇ u \boldsymbol{q}:=A\nabla u q := A ∇ u . Let Π ℓ \Pi_{\ell} Π ℓ be the elementwise L 2 L^{2} L 2 projector introduced above.
Lemma 3.11. (Consistency of the projected flux). Let r ∈ [ 2 , ∞ ] r\in[2,\infty] r ∈ [ 2 , ∞ ] and 1 r < s ≤ p \frac{1}{r}<s\le p r 1 < s ≤ p .
Assume
q = A ∇ u ∈ W s , r ( T ℓ ) . \boldsymbol{q}=A\nabla u\in W^{s,r}(\mathcal{T}_{\ell}). q = A ∇ u ∈ W s , r ( T ℓ ) . Then, for every w ℓ ∈ V ℓ w_{\ell}\in V_{\ell} w ℓ ∈ V ℓ ,
F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) = − ∑ e ∈ E ℓ ∫ e { { ( q − Π ℓ q ) ⋅ n e } } ⟦ w ℓ ⟧ d s . (66) F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)=-\sum_{e\in\mathcal{E}_{\ell}}\int_{e}\{\!\{\left(\boldsymbol{q}-\Pi_{\ell}\boldsymbol{q}\right)\cdot\boldsymbol{n}_{e}\}\!\}\llbracket w_{\ell}\rrbracket\,\mathrm{d}s. \tag*{(66)} F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) = − e ∈ E ℓ ∑ ∫ e { { ( q − Π ℓ q ) ⋅ n e } } [ [ w ℓ ] ] d s . (66) Consequently,
D ℓ ( u ) ≲ h ℓ s ∥ A ∇ u ∥ W s , r ( T ℓ ) . (67) \mathcal{D}_{\ell}(u)\lesssim h_{\ell}^{s}\lVert A\nabla u\rVert_{W^{s,r}(\mathcal{T}_{\ell})}. \tag*{(67)} D ℓ ( u ) ≲ h ℓ s ∥ A ∇ u ∥ W s , r ( T ℓ ) . (67) Proof. Since u ∈ H 0 1 ( Ω ) u\in H_{0}^{1}(\Omega) u ∈ H 0 1 ( Ω ) , one has ⟦ u ⟧ = 0 \llbracket u\rrbracket=0 [ [ u ] ] = 0 . Elementwise integration by parts in (57) gives
F ℓ ( w ℓ ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − ∑ e ∈ E ℓ ∫ e { { q ⋅ n e } } ⟦ w ℓ ⟧ d s + ∫ Ω c u w ℓ d x . F_{\ell}(w_{\ell})=\int_{\Omega}\boldsymbol{q}\cdot\nabla_{\ell}w_{\ell}\,\mathrm{d}x-\sum_{e\in\mathcal{E}_{\ell}}\int_{e}\{\!\{\boldsymbol{q}\cdot\boldsymbol{n}_{e}\}\!\}\llbracket w_{\ell}\rrbracket\,\mathrm{d}s+\int_{\Omega}cuw_{\ell}\,\mathrm{d}x. F ℓ ( w ℓ ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − e ∈ E ℓ ∑ ∫ e { { q ⋅ n e } } [ [ w ℓ ] ] d s + ∫ Ω c u w ℓ d x . On the other hand, by (47) ,
a ℓ D G ( u , w ℓ ; A , c ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − ∑ e ∈ E ℓ ∫ e { { Π ℓ q ⋅ n e } } ⟦ w ℓ ⟧ d s + ∫ Ω c u w ℓ d x , a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)=\int_{\Omega}\boldsymbol{q}\cdot\nabla_{\ell}w_{\ell}\,\mathrm{d}x-\sum_{e\in\mathcal{E}_{\ell}}\int_{e}\{\!\{\Pi_{\ell}\boldsymbol{q}\cdot\boldsymbol{n}_{e}\}\!\}\llbracket w_{\ell}\rrbracket\,\mathrm{d}s+\int_{\Omega}cuw_{\ell}\,\mathrm{d}x, a ℓ DG ( u , w ℓ ; A , c ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − e ∈ E ℓ ∑ ∫ e { { Π ℓ q ⋅ n e } } [ [ w ℓ ] ] d s + ∫ Ω c u w ℓ d x , @page 9/25 starts_mid=0 ends_mid=1 There are now two distinct consistency effects. The first is the small inconsistency introduced by the lifted, equivalently projected, numerical flux. Define
D ℓ ( u ) : = sup w ℓ ∈ V ℓ ∖ { 0 } ∣ F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) ∣ ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . (68) \mathcal{D}_{\ell}(u) := \sup_{w_{\ell}\in V_{\ell}\setminus\{0\}} \frac{\left|F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)\right|}{\lVert w_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}. \tag*{(68)} D ℓ ( u ) := w ℓ ∈ V ℓ ∖ { 0 } sup ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) . (68) The second is the error caused by replacing ( A , c ) (A,c) ( A , c ) by the represented pair ( A ~ ℓ , c ~ ℓ ) (\widetilde{A}_{\ell},\widetilde{c}_{\ell}) ( A ℓ , c ℓ ) :
E ℓ ( u ) : = sup w ℓ ∈ V ℓ ∖ { 0 } ∣ a ℓ D G ( u , w ℓ ; A , c ) − a ℓ D G ( u , w ℓ ; A ~ ℓ , c ~ ℓ ) ∣ ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . (69) \mathcal{E}_{\ell}(u) := \sup_{w_{\ell}\in V_{\ell}\setminus\{0\}} \frac{\left|a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})\right|}{\lVert w_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}. \tag*{(69)} E ℓ ( u ) := w ℓ ∈ V ℓ ∖ { 0 } sup ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ a ℓ DG ( u , w ℓ ; A , c ) − a ℓ DG ( u , w ℓ ; A ℓ , c ℓ ) . (69) Theorem 3.9. (Strang decomposition with representation defect). Under the preceding assumptions,
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ inf v ℓ ∈ V ℓ ∥ u − v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ + D ℓ ( u ) + E ℓ ( u ) . (70) \lVert u-u_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}} \lesssim\inf_{v_{\ell}\in V_{\ell}}\lVert u-v_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}+\mathcal{D}_{\ell}(u)+\mathcal{E}_{\ell}(u). \tag*{(70)} ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ v ℓ ∈ V ℓ inf ∥ u − v ℓ ∥ DG , ℓ ; A ℓ , c ℓ + D ℓ ( u ) + E ℓ ( u ) . (70) Proof. For arbitrary v ℓ , w ℓ ∈ V ℓ v_{\ell},w_{\ell}\in V_{\ell} v ℓ , w ℓ ∈ V ℓ ,
C ℓ ( v ℓ ; w ℓ ) = [ F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) ] + [ a ℓ D G ( u , w ℓ ; A , c ) − a ℓ D G ( u , w ℓ ; A ~ ℓ , c ~ ℓ ) ] + a ℓ D G ( u − v ℓ , w ℓ ; A ~ ℓ , c ~ ℓ ) . \mathcal{C}_{\ell}(v_{\ell};w_{\ell})=\left[F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)\right]+\left[a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})\right]+a_{\ell}^{\mathrm{DG}}(u-v_{\ell},w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell}). C ℓ ( v ℓ ; w ℓ ) = [ F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) ] + [ a ℓ DG ( u , w ℓ ; A , c ) − a ℓ DG ( u , w ℓ ; A ℓ , c ℓ ) ] + a ℓ DG ( u − v ℓ , w ℓ ; A ℓ , c ℓ ) . Apply (50) , divide by the DG norm of w ℓ w_{\ell} w ℓ , take the supremum, and substitute the result into (61) . □
Approximation and lifting consistency The regularity assumptions can be expressed directly in W 1 + s , r W^{1+s,r} W 1 + s , r rather than obtained indirectly from Sobolev embedding. For r ∈ [ 2 , ∞ ] r\in[2,\infty] r ∈ [ 2 , ∞ ] and s > 0 s>0 s > 0 , write
∥ z ∥ W s , r ( T ℓ ) : = ( ∑ K ∈ T ℓ ∥ z ∥ W s , r ( K ) r ) 1 / r , \lVert z\rVert_{W^{s,r}(\mathcal{T}_{\ell})}:=\left(\sum_{K\in\mathcal{T}_{\ell}}\lVert z\rVert_{W^{s,r}(K)}^{r}\right)^{1/r}, ∥ z ∥ W s , r ( T ℓ ) := ( K ∈ T ℓ ∑ ∥ z ∥ W s , r ( K ) r ) 1/ r , with the usual modification when r = ∞ r=\infty r = ∞ .
For 0 < s ≤ p 0<s\le p 0 < s ≤ p and u ∈ W 1 + s , r ( Ω ) u\in W^{1+s,r}(\Omega) u ∈ W 1 + s , r ( Ω ) , standard polynomial approximation on uniformly star-shaped polytopes gives
inf v ℓ ∈ V ℓ ∥ u − v ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ h ℓ s ∥ u ∥ W 1 + s , r ( Ω ) . (71) \inf_{v_{\ell}\in V_{\ell}}\lVert u-v_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim h_{\ell}^{s}\lVert u\rVert_{W^{1+s,r}(\Omega)}. \tag*{(71)} v ℓ ∈ V ℓ inf ∥ u − v ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ h ℓ s ∥ u ∥ W 1 + s , r ( Ω ) . (71) The lifting consistency error is controlled by best approximation of the physical flux q : = A ∇ u \boldsymbol{q}:=A\nabla u q := A ∇ u . Let Π ℓ \Pi_{\ell} Π ℓ be the elementwise L 2 L^{2} L 2 projector introduced above.
Lemma 3.11. (Consistency of the projected flux). Let r ∈ [ 2 , ∞ ] r\in[2,\infty] r ∈ [ 2 , ∞ ] and 1 r < s ≤ p \frac{1}{r}<s\le p r 1 < s ≤ p .
Assume
q = A ∇ u ∈ W s , r ( T ℓ ) . \boldsymbol{q}=A\nabla u\in W^{s,r}(\mathcal{T}_{\ell}). q = A ∇ u ∈ W s , r ( T ℓ ) . Then, for every w ℓ ∈ V ℓ w_{\ell}\in V_{\ell} w ℓ ∈ V ℓ ,
F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) = − ∑ e ∈ E ℓ ∫ e { { ( q − Π ℓ q ) ⋅ n e } } ⟦ w ℓ ⟧ d s . (72) F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)=-\sum_{e\in\mathcal{E}_{\ell}}\int_{e}\{\!\{(\boldsymbol{q}-\Pi_{\ell}\boldsymbol{q})\cdot\boldsymbol{n}_{e}\}\!\}\llbracket w_{\ell}\rrbracket\,\mathrm{d}s. \tag*{(72)} F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) = − e ∈ E ℓ ∑ ∫ e { {( q − Π ℓ q ) ⋅ n e } } [ [ w ℓ ] ] d s . (72) Consequently,
D ℓ ( u ) ≲ h ℓ s ∥ A ∇ u ∥ W s , r ( T ℓ ) . (73) \mathcal{D}_{\ell}(u)\lesssim h_{\ell}^{s}\lVert A\nabla u\rVert_{W^{s,r}(\mathcal{T}_{\ell})}. \tag*{(73)} D ℓ ( u ) ≲ h ℓ s ∥ A ∇ u ∥ W s , r ( T ℓ ) . (73) Proof. Since u ∈ H 0 1 ( Ω ) u\in H_{0}^{1}(\Omega) u ∈ H 0 1 ( Ω ) , one has ⟦ u ⟧ = 0 \llbracket u\rrbracket=0 [ [ u ] ] = 0 . Elementwise integration by parts in (57) gives
F ℓ ( w ℓ ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − ∑ e ∈ E ℓ ∫ e { { q ⋅ n e } } ⟦ w ℓ ⟧ d s + ∫ Ω c u w ℓ d x . F_{\ell}(w_{\ell})=\int_{\Omega}\boldsymbol{q}\cdot\nabla_{\ell}w_{\ell}\,\mathrm{d}x-\sum_{e\in\mathcal{E}_{\ell}}\int_{e}\{\!\{\boldsymbol{q}\cdot\boldsymbol{n}_{e}\}\!\}\llbracket w_{\ell}\rrbracket\,\mathrm{d}s+\int_{\Omega}cuw_{\ell}\,\mathrm{d}x. F ℓ ( w ℓ ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − e ∈ E ℓ ∑ ∫ e { { q ⋅ n e } } [ [ w ℓ ] ] d s + ∫ Ω c u w ℓ d x . On the other hand, by (47) ,
a ℓ D G ( u , w ℓ ; A , c ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − ∑ e ∈ E ℓ ∫ e { { Π ℓ q ⋅ n e } } ⟦ w ℓ ⟧ d s + ∫ Ω c u w ℓ d x , a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)=\int_{\Omega}\boldsymbol{q}\cdot\nabla_{\ell}w_{\ell}\,\mathrm{d}x-\sum_{e\in\mathcal{E}_{\ell}}\int_{e}\{\!\{\Pi_{\ell}\boldsymbol{q}\cdot\boldsymbol{n}_{e}\}\!\}\llbracket w_{\ell}\rrbracket\,\mathrm{d}s+\int_{\Omega}cuw_{\ell}\,\mathrm{d}x, a ℓ DG ( u , w ℓ ; A , c ) = ∫ Ω q ⋅ ∇ ℓ w ℓ d x − e ∈ E ℓ ∑ ∫ e { { Π ℓ q ⋅ n e } } [ [ w ℓ ] ] d s + ∫ Ω c u w ℓ d x ,
which proves (66).
Weighted Cauchy–Schwarz yields
∣ F ℓ ( w ℓ ) − a ℓ D G ( u , w ℓ ; A , c ) ∣ ≤ ( ∑ e ∈ E ℓ h e ∥ { { ( q − Π ℓ q ) ⋅ n e } } ∥ L 2 ( e ) 2 ) 1 / 2 ( ∑ e ∈ E ℓ h e − 1 ∥ ⟦ w ℓ ⟧ ∥ L 2 ( e ) 2 ) 1 / 2 . \left|F_{\ell}(w_{\ell})-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)\right|
\leq\left(\sum_{e\in\mathcal{E}_{\ell}}h_{e}\left\|\left\{\!\left\{(q-\Pi_{\ell}q)\cdot n_{e}\right\}\!\right\}\right\|_{L^{2}(e)}^{2}\right)^{1/2}\left(\sum_{e\in\mathcal{E}_{\ell}}h_{e}^{-1}\left\|\llbracket w_{\ell}\rrbracket\right\|_{L^{2}(e)}^{2}\right)^{1/2}. F ℓ ( w ℓ ) − a ℓ DG ( u , w ℓ ; A , c ) ≤ ( e ∈ E ℓ ∑ h e ∥ { { ( q − Π ℓ q ) ⋅ n e } } ∥ L 2 ( e ) 2 ) 1/2 ( e ∈ E ℓ ∑ h e − 1 ∥ [ [ w ℓ ] ] ∥ L 2 ( e ) 2 ) 1/2 .
The second factor is controlled by the DG norm. Polynomial best approximation together with the trace estimate gives
( ∑ e ∈ E ℓ h e ∥ { { ( q − Π ℓ q ) ⋅ n e } } ∥ L 2 ( e ) 2 ) 1 / 2 ≲ h ℓ s ∥ q ∥ W s , r ( T ℓ ) . \left(\sum_{e\in\mathcal{E}_{\ell}}h_{e}\left\|\left\{\!\left\{(q-\Pi_{\ell}q)\cdot n_{e}\right\}\!\right\}\right\|_{L^{2}(e)}^{2}\right)^{1/2}\lesssim h_{\ell}^{s}\left\|q\right\|_{W^{s,r}(\mathcal{T}_{\ell})}. ( e ∈ E ℓ ∑ h e ∥ { { ( q − Π ℓ q ) ⋅ n e } } ∥ L 2 ( e ) 2 ) 1/2 ≲ h ℓ s ∥ q ∥ W s , r ( T ℓ ) .
Dividing by the DG norm of w ℓ w_{\ell} w ℓ and taking the supremum proves the result. □
Combining (65) and (67) gives
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ h ℓ s [ ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) ] + E ℓ ( u ) . (74) \left\|u-u_{\ell}\right\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim h_{\ell}^{s}\left[\left\|u\right\|_{W^{1+s,r}(\Omega)}+\left\|A\nabla u\right\|_{W^{s,r}(\mathcal{T}_{\ell})}\right]+\mathcal{E}_{\ell}(u). \tag*{(74)} ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ h ℓ s [ ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) ] + E ℓ ( u ) . (74)
Coefficient representation defect
The representation defect has a particularly simple form in the lifting formulation. Since ⟦ u ⟧ = 0 \llbracket u\rrbracket=0 [ [ u ] ] = 0 ,
a ℓ D G ( u , w ℓ ; A , c ) − a ℓ D G ( u , w ℓ ; A ~ ℓ , c ~ ℓ ) = ∫ Ω ( A − A ~ ℓ ) ∇ u ⋅ [ ∇ ℓ w ℓ + R ℓ ( ⟦ w ℓ ⟧ ) ] d x + ∫ Ω ( c − c ~ ℓ ) u w ℓ d x . (75) a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})
=\int_{\Omega}(A-\widetilde{A}_{\ell})\nabla u\cdot[\nabla_{\ell}w_{\ell}+\mathcal{R}_{\ell}(\llbracket w_{\ell}\rrbracket)]\,\mathrm{d}x
+\int_{\Omega}(c-\widetilde{c}_{\ell})uw_{\ell}\,\mathrm{d}x. \tag*{(75)} a ℓ DG ( u , w ℓ ; A , c ) − a ℓ DG ( u , w ℓ ; A ℓ , c ℓ ) = ∫ Ω ( A − A ℓ ) ∇ u ⋅ [ ∇ ℓ w ℓ + R ℓ ( [ [ w ℓ ] ] )] d x + ∫ Ω ( c − c ℓ ) u w ℓ d x . (75)
Thus
E ℓ ( u ) = sup w ℓ ∈ V ℓ ∖ { 0 } 1 ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ∣ ∫ Ω ( A − A ~ ℓ ) ∇ u ⋅ [ ∇ ℓ w ℓ + R ℓ ( ⟦ w ℓ ⟧ ) ] d x + ∫ Ω ( c − c ~ ℓ ) u w ℓ d x ∣ . (76) \mathcal{E}_{\ell}(u)=\sup_{w_{\ell}\in V_{\ell}\setminus\{0\}}\frac{1}{\left\|w_{\ell}\right\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}}\left|\int_{\Omega}(A-\widetilde{A}_{\ell})\nabla u\cdot[\nabla_{\ell}w_{\ell}+\mathcal{R}_{\ell}(\llbracket w_{\ell}\rrbracket)]\,\mathrm{d}x+\int_{\Omega}(c-\widetilde{c}_{\ell})uw_{\ell}\,\mathrm{d}x\right|. \tag*{(76)} E ℓ ( u ) = w ℓ ∈ V ℓ ∖ { 0 } sup ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ 1 ∫ Ω ( A − A ℓ ) ∇ u ⋅ [ ∇ ℓ w ℓ + R ℓ ( [ [ w ℓ ] ] )] d x + ∫ Ω ( c − c ℓ ) u w ℓ d x . (76)
Lemma (Lq ^{q} q representation estimate). Let q ∈ [ 2 , ∞ ] q\in[2,\infty] q ∈ [ 2 , ∞ ] and define r ∈ [ 2 , ∞ ] r\in[2,\infty] r ∈ [ 2 , ∞ ] by
1 q + 1 r + 1 2 = 1. (77) \frac{1}{q}+\frac{1}{r}+\frac{1}{2}=1. \tag*{(77)} q 1 + r 1 + 2 1 = 1. (77) Assume
u ∈ W 1 , r ( Ω ) u\in W^{1,r}(\Omega) u ∈ W 1 , r ( Ω ) , A − A ~ ℓ ∈ L q ( Ω ; R d × d ) A-\widetilde{A}_{\ell}\in L^{q}(\Omega;\mathbb{R}^{d\times d}) A − A ℓ ∈ L q ( Ω ; R d × d ) , c − c ~ ℓ ∈ L q ( Ω ) c-\widetilde{c}_{\ell}\in L^{q}(\Omega) c − c ℓ ∈ L q ( Ω ) .
Then
E ℓ ( u ) ≲ ∥ A − A ~ ℓ ∥ L q ( Ω ) ∥ ∇ u ∥ L r ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ∥ u ∥ L r ( Ω ) . (78) \mathcal{E}_{\ell}(u)\lesssim\left\|A-\widetilde{A}_{\ell}\right\|_{L^{q}(\Omega)}\left\|\nabla u\right\|_{L^{r}(\Omega)}+\left\|c-\widetilde{c}_{\ell}\right\|_{L^{q}(\Omega)}\left\|u\right\|_{L^{r}(\Omega)}. \tag*{(78)} E ℓ ( u ) ≲ A − A ℓ L q ( Ω ) ∥ ∇ u ∥ L r ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ∥ u ∥ L r ( Ω ) . (78)
Proof. From (75) and Hölder’s inequality,
∣ a ℓ D G ( u , w ℓ ; A , c ) − a ℓ D G ( u , w ℓ ; A ~ ℓ , c ~ ℓ ) ∣ ≤ ∥ A − A ~ ℓ ∥ L q ( Ω ) ∥ ∇ u ∥ L r ( Ω ) ∥ ∇ ℓ w ℓ + R ℓ ( ⟦ w ℓ ⟧ ) ∥ L 2 ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ∥ u ∥ L r ( Ω ) ∥ w ℓ ∥ L 2 ( Ω ) . \left|a_{\ell}^{\mathrm{DG}}(u,w_{\ell};A,c)-a_{\ell}^{\mathrm{DG}}(u,w_{\ell};\widetilde{A}_{\ell},\widetilde{c}_{\ell})\right|
\leq\left\|A-\widetilde{A}_{\ell}\right\|_{L^{q}(\Omega)}\left\|\nabla u\right\|_{L^{r}(\Omega)}\left\|\nabla_{\ell}w_{\ell}+\mathcal{R}_{\ell}(\llbracket w_{\ell}\rrbracket)\right\|_{L^{2}(\Omega)}
+\left\|c-\widetilde{c}_{\ell}\right\|_{L^{q}(\Omega)}\left\|u\right\|_{L^{r}(\Omega)}\left\|w_{\ell}\right\|_{L^{2}(\Omega)}. a ℓ DG ( u , w ℓ ; A , c ) − a ℓ DG ( u , w ℓ ; A ℓ , c ℓ ) ≤ A − A ℓ L q ( Ω ) ∥ ∇ u ∥ L r ( Ω ) ∥ ∇ ℓ w ℓ + R ℓ ( [ [ w ℓ ] ] ) ∥ L 2 ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ∥ u ∥ L r ( Ω ) ∥ w ℓ ∥ L 2 ( Ω ) .
By (41) , (24) and (48) ,
∥ ∇ ℓ w ℓ + R ℓ ( ⟦ w ℓ ⟧ ) ∥ L 2 ( Ω ) + ∥ w ℓ ∥ L 2 ( Ω ) ≲ ∥ w ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ . (79) \left\|\nabla_{\ell}w_{\ell}+\mathcal{R}_{\ell}(\llbracket w_{\ell}\rrbracket)\right\|_{L^{2}(\Omega)}+\left\|w_{\ell}\right\|_{L^{2}(\Omega)}\lesssim\left\|w_{\ell}\right\|_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}. \tag*{(79)} ∥ ∇ ℓ w ℓ + R ℓ ( [ [ w ℓ ] ] ) ∥ L 2 ( Ω ) + ∥ w ℓ ∥ L 2 ( Ω ) ≲ ∥ w ℓ ∥ DG , ℓ ; A ℓ , c ℓ . (79)
Division by the DG norm and taking the supremum gives (78) . □
We may therefore impose the W 1 + s , r W^{1+s,r} W 1 + s , r regularity required by the spatial approximation directly. If
u ∈ W 1 + s , r ( Ω ) , u \in W^{1+s,r}(\Omega), u ∈ W 1 + s , r ( Ω ) ,
then
E ℓ ( u ) ≲ [ ∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ] ∥ u ∥ W 1 + s , r ( Ω ) . (80) \mathcal{E}_{\ell}(u) \lesssim\left[\lVert A-\widetilde{A}_{\ell}\rVert_{L^q(\Omega)}+\lVert c-\widetilde{c}_{\ell}\rVert_{L^q(\Omega)}\right]\lVert u\rVert_{W^{1+s,r}(\Omega)}. \tag*{(80)} E ℓ ( u ) ≲ [ ∥ A − A ℓ ∥ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ] ∥ u ∥ W 1 + s , r ( Ω ) . (80)
Combining (74) and (80) yields
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ h ℓ s [ ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) ] + [ ∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ] ∥ u ∥ W 1 + s , r ( Ω ) . (81) \lVert u-u_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim h_{\ell}^{s}\left[\lVert u\rVert_{W^{1+s,r}(\Omega)}+\lVert A\nabla u\rVert_{W^{s,r}(\mathcal{T}_{\ell})}\right]+\left[\lVert A-\widetilde{A}_{\ell}\rVert_{L^q(\Omega)}+\lVert c-\widetilde{c}_{\ell}\rVert_{L^q(\Omega)}\right]\lVert u\rVert_{W^{1+s,r}(\Omega)}. \tag*{(81)} ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ h ℓ s [ ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) ] + [ ∥ A − A ℓ ∥ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ] ∥ u ∥ W 1 + s , r ( Ω ) . (81)
It is convenient to introduce the pathwise regularity quantity
M s , r ( u ; A ) : = ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) . (82) \mathcal{M}_{s,r}(u;A):=\lVert u\rVert_{W^{1+s,r}(\Omega)}+\lVert A\nabla u\rVert_{W^{s,r}(\mathcal{T}_{\ell})}. \tag*{(82)} M s , r ( u ; A ) := ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) . (82)
Then (81) may be written compactly as
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ [ h ℓ s + ∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ] M s , r ( u ; A ) . (83) \lVert u-u_{\ell}\rVert_{\mathrm{DG},\ell;\widetilde{A}_{\ell},\widetilde{c}_{\ell}}\lesssim\left[h_{\ell}^{s}+\lVert A-\widetilde{A}_{\ell}\rVert_{L^q(\Omega)}+\lVert c-\widetilde{c}_{\ell}\rVert_{L^q(\Omega)}\right]\mathcal{M}_{s,r}(u;A). \tag*{(83)} ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ ≲ [ h ℓ s + ∥ A − A ℓ ∥ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ] M s , r ( u ; A ) . (83)
The first contribution in (83) contains both the best-approximation error of the solution and the controlled consistency error introduced by the projected numerical flux. The remaining terms isolate the error caused by representing the random coefficients on level ℓ \ell ℓ . This is the decomposition used in the multilevel analysis below.
Sampling strategies for the random coefficients
This section describes two constructions of the represented coefficient pair ( A ~ ℓ , c ~ ℓ ) (\widetilde{A}_{\ell},\widetilde{c}_{\ell}) ( A ℓ , c ℓ ) appearing in the forward DG problem. Both constructions begin with a Gaussian latent field and then apply deterministic coefficient maps. They differ in the object used to couple adjacent MLMC levels.
In the spectral construction, a single Gaussian field is represented on a Cartesian reference grid and the same realised field is evaluated on every polytopic forward mesh. In the SPDE construction, the latent field is approximated on an agglomerated polytopic hierarchy and adjacent levels are coupled through consistent restrictions of the same spatial white noise.
These are alternative multilevel couplings and are not coupled pathwise to one another. When they are compared numerically, their correlation length, bulk variance and coefficient maps are chosen consistently. The periodic FFT model and the bounded-domain SPDE with homogeneous Dirichlet conditions nevertheless define different Gaussian measures near the boundary.
Throughout this section let
Φ A : Ω × R ⟶ R s y m d × d , Φ c : Ω × R ⟶ R (84) \Phi_A:\Omega\times\mathbb{R}\longrightarrow\mathbb{R}_{\mathrm{sym}}^{d\times d},\qquad\Phi_c:\Omega\times\mathbb{R}\longrightarrow\mathbb{R} \tag*{(84)} Φ A : Ω × R ⟶ R sym d × d , Φ c : Ω × R ⟶ R (84)
be measurable coefficient maps such that
A ( x , ξ ) = Φ A ( x , Y ( x , ξ ) ) , c ( x , ξ ) = Φ c ( x , Y ( x , ξ ) ) . (85) A(x,\xi)=\Phi_A(x,Y(x,\xi)),\qquad c(x,\xi)=\Phi_c(x,Y(x,\xi)). \tag*{(85)} A ( x , ξ ) = Φ A ( x , Y ( x , ξ )) , c ( x , ξ ) = Φ c ( x , Y ( x , ξ )) . (85)
The truncated lognormal models of § 2.2 are particular examples. We assume that, uniformly in x ∈ Ω x\in\Omega x ∈ Ω , the maps satisfy
∥ Φ A ( x , y 1 ) − Φ A ( x , y 2 ) ∥ ≤ L A ∣ y 1 − y 2 ∣ , ∣ Φ c ( x , y 1 ) − Φ c ( x , y 2 ) ∣ ≤ L c ∣ y 1 − y 2 ∣ . (86) \lVert\Phi_A(x,y_1)-\Phi_A(x,y_2)\rVert\le L_A\lvert y_1-y_2\rvert,\qquad\lvert\Phi_c(x,y_1)-\Phi_c(x,y_2)\rvert\le L_c\lvert y_1-y_2\rvert. \tag*{(86)} ∥ Φ A ( x , y 1 ) − Φ A ( x , y 2 )∥ ≤ L A ∣ y 1 − y 2 ∣ , ∣ Φ c ( x , y 1 ) − Φ c ( x , y 2 )∣ ≤ L c ∣ y 1 − y 2 ∣ . (86)
on the range relevant to the truncated latent field. The truncations are chosen so that the uniform bounds (24) hold for every represented coefficient field considered below.
Spectral sampling on a Cartesian reference grid
Let B : = [ 0 , L B ] d \mathcal{B}:=[0,L_{\mathcal{B}}]^d B := [ 0 , L B ] d be a Cartesian box containing Ω ‾ \overline{\Omega} Ω , interpreted periodically. For n ∈ Z d n\in\mathbb{Z}^d n ∈ Z d , set
k n : = 2 π L B n , (87) k_n:=\frac{2\pi}{L_{\mathcal{B}}}n, \tag*{(87)} k n := L B 2 π n , (87)
and let ϕ n ( x ) : = L B − d / 2 exp ( i k n ⋅ x ) \phi_n(x):=L_{\mathcal{B}}^{-d/2}\exp(\mathrm{i}k_n\cdot x) ϕ n ( x ) := L B − d /2 exp ( i k n ⋅ x ) denote the corresponding orthonormal Fourier basis of L 2 ( B ) L^2(\mathcal{B}) L 2 ( B ) .
We consider the real-valued, mean-zero periodic Gaussian field
Y ( x , ξ ) : = ∑ n ∈ Z d S n ζ n ( ξ ) ϕ n ( x ) , (88) Y(x,\xi):=\sum_{n\in\mathbb{Z}^d}\sqrt{S_n}\,\zeta_n(\xi)\phi_n(x), \tag*{(88)} Y ( x , ξ ) := n ∈ Z d ∑ S n ζ n ( ξ ) ϕ n ( x ) , (88)
where
S n : = c ν , d , B τ 2 ( κ 2 + ∣ k n ∣ 2 ) − ( ν + d / 2 ) . (89) S_n:=c_{\nu,d,\mathcal{B}}\tau^2\left(\kappa^2+\lvert k_n\rvert^2\right)^{-(\nu+d/2)}. \tag*{(89)} S n := c ν , d , B τ 2 ( κ 2 + ∣ k n ∣ 2 ) − ( ν + d /2 ) . (89)
Here κ > 0 \kappa>0 κ > 0 determines the correlation length, ν > 0 \nu>0 ν > 0 controls the regularity, τ > 0 \tau>0 τ > 0 determines the variance and c ν , d , B > 0 c_{\nu,d,\mathcal{B}}>0 c ν , d , B > 0 is the normalisation constant for the periodic model. The Gaussian coefficients satisfy E [ ζ n ] = 0 \mathbb{E}[\zeta_n]=0 E [ ζ n ] = 0 , E [ ζ n ζ m ] = δ n m \mathbb{E}[\zeta_n\zeta_m]=\delta_{nm} E [ ζ n ζ m ] = δ nm and ζ − n = ζ n ‾ \zeta_{-n}=\overline{\zeta_n} ζ − n = ζ n , with the zero mode real. Since ν > 0 \nu>0 ν > 0 , the series (88) converges in L 2 ( Ξ ; L 2 ( B ) ) L^2(\Xi;L^2(\mathcal{B})) L 2 ( Ξ ; L 2 ( B )) . Its restriction to Ω \Omega Ω defines the target latent field for the FFT construction.
Fix a reference resolution N r e f ∈ N N_{\mathrm{ref}}\in\mathbb{N} N ref ∈ N and set
δ r e f : = L B N r e f . (90) \delta_{\mathrm{ref}}:=\frac{L_{\mathcal{B}}}{N_{\mathrm{ref}}}. \tag*{(90)} δ ref := N ref L B . (90)
The Cartesian grid points are
x i : = δ r e f i , i ∈ { 0 , … , N r e f − 1 } d . (91) x_i:=\delta_{\mathrm{ref}}i,\qquad i\in\{0,\ldots,N_{\mathrm{ref}}-1\}^d. \tag*{(91)} x i := δ ref i , i ∈ { 0 , … , N ref − 1 } d . (91)
Let Λ r e f ⊂ Z d \Lambda_{\mathrm{ref}}\subset\mathbb{Z}^d Λ ref ⊂ Z d denote the Fourier modes represented by the corresponding N r e f d N_{\mathrm{ref}}^d N ref d -point discrete Fourier transform, with the standard Nyquist convention, and define
P r e f Y : = ∑ n ∈ Λ r e f S n ζ n ϕ n . (92) P_{\mathrm{ref}}Y:=\sum_{n\in\Lambda_{\mathrm{ref}}}\sqrt{S_n}\,\zeta_n\phi_n. \tag*{(92)} P ref Y := n ∈ Λ ref ∑ S n ζ n ϕ n . (92)
The reference-grid sample is the restriction of this same truncated realisation to the Cartesian grid,
( Y r e f ) i : = P r e f Y ( x i ) . (93) (Y_{\mathrm{ref}})_i:=P_{\mathrm{ref}}Y(x_i). \tag*{(93)} ( Y ref ) i := P ref Y ( x i ) . (93)
Equivalently, under the chosen discrete Fourier normalisation,
Y r e f = F − 1 ( S n W ^ r e f , n ) , W ^ r e f = F W r e f , (94) Y_{\mathrm{ref}}=\mathcal{F}^{-1}\left(\sqrt{S_n}\,\widehat{W}_{\mathrm{ref},n}\right),\qquad\widehat{W}_{\mathrm{ref}}=\mathcal{F}W_{\mathrm{ref}}, \tag*{(94)} Y ref = F − 1 ( S n W ref , n ) , W ref = F W ref , (94)
where W r e f W_{\mathrm{ref}} W ref is a real centred Gaussian white-noise vector and its Fourier coefficients are identified with the resolved coefficients in (92) . Thus (94) is the computational realisation of the Fourier truncation of the same field Y Y Y . The zero Fourier mode is retained.
Let I r e f \mathcal{I}_{\mathrm{ref}} I ref denote the Cartesian interpolation operator applied to the grid values (93) ; in the numerical experiments this is bilinear interpolation in two dimensions. Define
Y ~ r e f : = I r e f ( P r e f Y ) on Ω . (95) \widetilde{Y}_{\mathrm{ref}}:=\mathcal{I}_{\mathrm{ref}}(P_{\mathrm{ref}}Y)\quad\text{on }\Omega. \tag*{(95)} Y ref := I ref ( P ref Y ) on Ω. (95)
For a fixed MLMC hierarchy, the same realised function Y ~ r e f ( ⋅ , ξ ) \widetilde{Y}_{\mathrm{ref}}(\cdot,\xi) Y ref ( ⋅ , ξ ) drives every forward level:
Y ~ ℓ : = Y ~ r e f , ℓ = 0 , … , L . (96) \widetilde{Y}_{\ell}:=\widetilde{Y}_{\mathrm{ref}},\qquad\ell=0,\ldots,L. \tag*{(96)} Y ℓ := Y ref , ℓ = 0 , … , L . (96)
The represented coefficients are
A ~ ℓ ( x , ξ ) : = Φ A ( x , Y ~ r e f ( x , ξ ) ) , c ~ ℓ ( x , ξ ) : = Φ c ( x , Y ~ r e f ( x , ξ ) ) . (97) \widetilde{A}_{\ell}(x,\xi):=\Phi_A\left(x,\widetilde{Y}_{\mathrm{ref}}(x,\xi)\right),\qquad\widetilde{c}_{\ell}(x,\xi):=\Phi_c\left(x,\widetilde{Y}_{\mathrm{ref}}(x,\xi)\right). \tag*{(97)} A ℓ ( x , ξ ) := Φ A ( x , Y ref ( x , ξ ) ) , c ℓ ( x , ξ ) := Φ c ( x , Y ref ( x , ξ ) ) . (97)
Hence both forward solves in the coupled correction δ Q ℓ : = Q ℓ − Q ℓ − 1 \delta Q_{\ell}:=Q_{\ell}-Q_{\ell-1} δ Q ℓ := Q ℓ − Q ℓ − 1 use the same represented coefficient realisation. No coefficient field is transferred between the forward meshes.
By construction,
Y − Y ~ r e f = ( Y − P r e f Y ) + ( P r e f Y − I r e f ( P r e f Y ) ) on Ω . (98) Y-\widetilde{Y}_{\mathrm{ref}}=(Y-P_{\mathrm{ref}}Y)+(P_{\mathrm{ref}}Y-\mathcal{I}_{\mathrm{ref}}(P_{\mathrm{ref}}Y))\quad\text{on }\Omega. \tag*{(98)} Y − Y ref = ( Y − P ref Y ) + ( P ref Y − I ref ( P ref Y )) on Ω. (98)
Lemma (Propagation of latent-field error)
Lemma (Propagation of latent-field error). Under (86) , for every q ∈ [ 1 , ∞ ] q \in[1,\infty] q ∈ [ 1 , ∞ ] ,
∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ≲ ∥ Y − Y ~ r e f ∥ L q ( Ω ) (99) \left\| A-\widetilde{A}_{\ell}\right\|_{L^{q}(\Omega)}+\left\|c-\widetilde{c}_{\ell}\right\|_{L^{q}(\Omega)}\lesssim\left\|Y-\widetilde{Y}_{\mathrm{ref}}\right\|_{L^{q}(\Omega)} \tag*{(99)} A − A ℓ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ≲ Y − Y ref L q ( Ω ) (99)
almost surely, with a constant independent of ℓ \ell ℓ and N r e f N_{\mathrm{ref}} N ref .
Proof. The pointwise estimate follows from (85) , (97) and (86) . Taking the L q ( Ω ) L^{q}(\Omega) L q ( Ω ) norm gives the result.
Lemma (Reference-grid approximation)
Lemma (Reference-grid approximation). Let q ∈ [ 1 , ∞ ] q \in[1,\infty] q ∈ [ 1 , ∞ ] and suppose that, for some μ ∈ ( 0 , 2 ] \mu\in(0,2] μ ∈ ( 0 , 2 ] appropriate to the interpolation operator, P r e f Y ∈ W μ , q ( B ) P_{\mathrm{ref}}Y \in W^{\mu,q}(\mathcal{B}) P ref Y ∈ W μ , q ( B ) almost surely. Then
∥ Y − Y ~ r e f ∥ L q ( Ω ) ≲ ∥ Y − P r e f Y ∥ L q ( B ) + δ r e f μ ∥ P r e f Y ∥ W μ , q ( B ) . (100) \left\|Y-\widetilde{Y}_{\mathrm{ref}}\right\|_{L^{q}(\Omega)}\lesssim\left\|Y-P_{\mathrm{ref}}Y\right\|_{L^{q}(\mathcal{B})}+\delta_{\mathrm{ref}}^{\mu}\left\|P_{\mathrm{ref}}Y\right\|_{W^{\mu,q}(\mathcal{B})}. \tag*{(100)} Y − Y ref L q ( Ω ) ≲ ∥ Y − P ref Y ∥ L q ( B ) + δ ref μ ∥ P ref Y ∥ W μ , q ( B ) . (100)
Proof. Apply the triangle inequality to (98) and the Cartesian interpolation estimate to the second term.
Combining Lemmas 4.2 and 4.3 gives
∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ≲ ∥ Y − P r e f Y ∥ L q ( B ) + δ r e f μ ∥ P r e f Y ∥ W μ , q ( B ) . (101) \left\|A-\widetilde{A}_{\ell}\right\|_{L^{q}(\Omega)}+\left\|c-\widetilde{c}_{\ell}\right\|_{L^{q}(\Omega)}\lesssim\left\|Y-P_{\mathrm{ref}}Y\right\|_{L^{q}(\mathcal{B})}+\delta_{\mathrm{ref}}^{\mu}\left\|P_{\mathrm{ref}}Y\right\|_{W^{\mu,q}(\mathcal{B})}. \tag*{(101)} A − A ℓ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) ≲ ∥ Y − P ref Y ∥ L q ( B ) + δ ref μ ∥ P ref Y ∥ W μ , q ( B ) . (101)
For finite ϱ ≥ 1 \varrho\ge1 ϱ ≥ 1 , define
ε r e f : = ∥ ∥ Y − P r e f Y ∥ L q ( B ) + δ r e f μ ∥ P r e f Y ∥ W μ , q ( B ) ∥ L 2 ϱ ( Ξ ) . (102) \varepsilon_{\mathrm{ref}}:=\left\|\left\|Y-P_{\mathrm{ref}}Y\right\|_{L^{q}(\mathcal{B})}+\delta_{\mathrm{ref}}^{\mu}\left\|P_{\mathrm{ref}}Y\right\|_{W^{\mu,q}(\mathcal{B})}\right\|_{L^{2\varrho}(\Xi)}. \tag*{(102)} ε ref := ∥ Y − P ref Y ∥ L q ( B ) + δ ref μ ∥ P ref Y ∥ W μ , q ( B ) L 2 ϱ ( Ξ ) . (102)
Then
∥ ∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) ∥ L 2 ϱ ( Ξ ) ≲ ε r e f , ℓ = 0 , … , L . (103) \left\|\left\|A-\widetilde{A}_{\ell}\right\|_{L^{q}(\Omega)}+\left\|c-\widetilde{c}_{\ell}\right\|_{L^{q}(\Omega)}\right\|_{L^{2\varrho}(\Xi)}\lesssim\varepsilon_{\mathrm{ref}},\qquad\ell=0,\ldots,L. \tag*{(103)} A − A ℓ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) L 2 ϱ ( Ξ ) ≲ ε ref , ℓ = 0 , … , L . (103)
For a target hierarchy with finest scale h L h_{L} h L , the resolution N r e f = N r e f ( L ) N_{\mathrm{ref}}=N_{\mathrm{ref}}(L) N ref = N ref ( L ) is selected at the accuracy required for that hierarchy. The resulting reference realisation is then shared by every forward level 0 , … , L 0,\ldots,L 0 , … , L .
SPDE sampling on an agglomerated hierarchy
We next construct the latent field through the elliptic SPDE
( κ 2 − Δ ) Y = τ W in Ω , Y = 0 on ∂ Ω . (104) \left(\kappa^{2}-\Delta\right)Y=\tau W\quad\text{in }\Omega,\qquad Y=0\quad\text{on }\partial\Omega. \tag*{(104)} ( κ 2 − Δ ) Y = τ W in Ω , Y = 0 on ∂ Ω. (104)
Let L D : = κ 2 − Δ D \mathcal{L}_{D}:=\kappa^{2}-\Delta_{D} L D := κ 2 − Δ D and G : = L D − 1 \mathcal{G}:=\mathcal{L}_{D}^{-1} G := L D − 1 . Then
Y = τ G W . (105) Y=\tau\mathcal{G}W. \tag*{(105)} Y = τ G W . (105)
Spatial white noise is understood as an isonormal Gaussian process over L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) :
E [ ⟨ W , v ⟩ ⟨ W , w ⟩ ] = ( v , w ) L 2 ( Ω ) ∀ v , w ∈ L 2 ( Ω ) . (106) \mathbb{E}\left[\langle W,v\rangle\langle W,w\rangle\right]=(v,w)_{L^{2}(\Omega)}\qquad\forall v,w\in L^{2}(\Omega). \tag*{(106)} E [ ⟨ W , v ⟩ ⟨ W , w ⟩ ] = ( v , w ) L 2 ( Ω ) ∀ v , w ∈ L 2 ( Ω ) . (106)
In the whole-space or periodic setting, (104) has spectral density proportional to ( κ 2 + ∣ k ∣ 2 ) − 2 \left(\kappa^{2}+|k|^{2}\right)^{-2} ( κ 2 + ∣ k ∣ 2 ) − 2 . It therefore agrees with the Matérn spectrum (89) when
ν + d 2 = 2. (107) \nu+\frac{d}{2}=2. \tag*{(107)} ν + 2 d = 2. (107)
For other Matérn smoothness parameters a fractional or higher-order SPDE representation may be used.
The homogeneous Dirichlet condition modifies the covariance near ∂ Ω \partial\Omega ∂ Ω and, in particular, pins the latent field to zero on the boundary. It is part of the bounded-domain SPDE model and is not intended to reproduce the periodic FFT covariance close to the boundary.
Let { T ℓ } ℓ = 0 L \{\mathcal{T}_{\ell}\}_{\ell=0}^{L} { T ℓ } ℓ = 0 L be an agglomerated hierarchy such that every coarse cell K ∈ T ℓ − 1 K\in\mathcal{T}_{\ell-1} K ∈ T ℓ − 1 is a union of cells of T ℓ \mathcal{T}_{\ell} T ℓ :
K = ⋃ k ∈ C ℓ ( K ) k . (108) K=\bigcup_{k\in\mathcal{C}_{\ell}(K)}k. \tag*{(108)} K = k ∈ C ℓ ( K ) ⋃ k . (108)
For fixed polynomial degree this gives the natural inclusion
V ℓ − 1 ⊂ V ℓ , (109) V_{\ell-1}\subset V_{\ell}, \tag*{(109)} V ℓ − 1 ⊂ V ℓ , (109)
since a polynomial on a coarse agglomerate restricts to a polynomial of the same degree on every fine child.
The latent SPDE is discretised by the same lifting-based symmetric interior-penalty construction as the forward problem, with diffusion coefficient I I I and reaction coefficient κ 2 \kappa^2 κ 2 . Thus
a ℓ , D l a t ( w , v ) : = a ℓ , D D G ( w , v ; I , κ 2 ) . (110) a^{\mathrm{lat}}_{\ell,D}(w,v):=a^{\mathrm{DG}}_{\ell,D}(w,v;I,\kappa^2). \tag*{(110)} a ℓ , D lat ( w , v ) := a ℓ , D DG ( w , v ; I , κ 2 ) . (110)
In particular, the lifting and penalty terms include the boundary faces and impose the homogeneous Dirichlet condition weakly.
Define the discrete latent resolvent G ℓ : V ℓ ′ ⟶ V ℓ \mathcal{G}_{\ell}:V'_{\ell}\longrightarrow V_{\ell} G ℓ : V ℓ ′ ⟶ V ℓ by
a ℓ , D l a t ( G ℓ L ℓ , v ℓ ) = L ℓ ( v ℓ ) ∀ v ℓ ∈ V ℓ . (111) a^{\mathrm{lat}}_{\ell,D}(\mathcal{G}_{\ell}L_{\ell},v_{\ell})=L_{\ell}(v_{\ell})\qquad\forall v_{\ell}\in V_{\ell}. \tag*{(111)} a ℓ , D lat ( G ℓ L ℓ , v ℓ ) = L ℓ ( v ℓ ) ∀ v ℓ ∈ V ℓ . (111)
Discrete white noise and inter-level coupling
The canonical level-ℓ \ell ℓ white-noise functional is the restriction of the same continuum white noise to V ℓ V_{\ell} V ℓ :
η ℓ ( v ℓ ) : = ⟨ W , v ℓ ⟩ , v ℓ ∈ V ℓ . (112) \eta_{\ell}(v_{\ell}):=\langle W,v_{\ell}\rangle,\qquad v_{\ell}\in V_{\ell}. \tag*{(112)} η ℓ ( v ℓ ) := ⟨ W , v ℓ ⟩ , v ℓ ∈ V ℓ . (112)
It satisfies
E [ η ℓ ( v ℓ ) η ℓ ( w ℓ ) ] = ( v ℓ , w ℓ ) L 2 ( Ω ) . (113) \mathbb{E}[\eta_{\ell}(v_{\ell})\eta_{\ell}(w_{\ell})]=(v_{\ell},w_{\ell})_{L^2(\Omega)}. \tag*{(113)} E [ η ℓ ( v ℓ ) η ℓ ( w ℓ )] = ( v ℓ , w ℓ ) L 2 ( Ω ) . (113)
Let { φ i ( ℓ ) } i = 1 n ℓ \{\varphi_i^{(\ell)}\}_{i=1}^{n_\ell} { φ i ( ℓ ) } i = 1 n ℓ be a basis of V ℓ V_{\ell} V ℓ and define the mass matrix
( M ℓ ) i j : = ∫ Ω φ i ( ℓ ) φ j ( ℓ ) d x . (114) (M_{\ell})_{ij}:=\int_{\Omega}\varphi_i^{(\ell)}\varphi_j^{(\ell)}\,\mathrm{d}x. \tag*{(114)} ( M ℓ ) ij := ∫ Ω φ i ( ℓ ) φ j ( ℓ ) d x . (114)
The white-noise load vector
( b ℓ ) i : = η ℓ ( φ i ( ℓ ) ) (115) (b_{\ell})_i:=\eta_{\ell}(\varphi_i^{(\ell)}) \tag*{(115)} ( b ℓ ) i := η ℓ ( φ i ( ℓ ) ) (115)
therefore satisfies
b ℓ ∼ N ( 0 , M ℓ ) . (116) b_{\ell}\sim\mathcal{N}(0,M_{\ell}). \tag*{(116)} b ℓ ∼ N ( 0 , M ℓ ) . (116)
If v ℓ = ∑ i v i φ i ( ℓ ) v_{\ell}=\sum_i v_i\varphi_i^{(\ell)} v ℓ = ∑ i v i φ i ( ℓ ) , then
η ℓ ( v ℓ ) = b ℓ T v . (117) \eta_{\ell}(v_{\ell})=b_{\ell}^{\mathsf{T}}v. \tag*{(117)} η ℓ ( v ℓ ) = b ℓ T v . (117)
Thus b ℓ b_{\ell} b ℓ is the Gaussian load vector representing the action of white noise on V ℓ V_{\ell} V ℓ ; it is not the coefficient vector of an L 2 L^2 L 2 function.
Since the DG mass matrix is block diagonal by elements, (116) may be sampled locally. If M K = L K L K T M_K=L_KL_K^{\mathsf{T}} M K = L K L K T , then
b K = L K z K , z K ∼ N ( 0 , I ) , (118) b_K=L_Kz_K,\qquad z_K\sim\mathcal{N}(0,I), \tag*{(118)} b K = L K z K , z K ∼ N ( 0 , I ) , (118)
and the elemental load vectors are mutually independent.
Let P ℓ − 1 ℓ : R n ℓ − 1 ⟶ R n ℓ P^\ell_{\ell-1}:\mathbb{R}^{n_{\ell-1}}\longrightarrow\mathbb{R}^{n_\ell} P ℓ − 1 ℓ : R n ℓ − 1 ⟶ R n ℓ denote the coefficient prolongation associated with (109) . Thus if v ℓ − 1 ∈ V ℓ − 1 v_{\ell-1}\in V_{\ell-1} v ℓ − 1 ∈ V ℓ − 1 has coefficient vector v \mathbf{v} v , then P ℓ − 1 ℓ v P^\ell_{\ell-1}\mathbf{v} P ℓ − 1 ℓ v represents the same function as an element of V ℓ V_{\ell} V ℓ . Consequently,
M ℓ − 1 = ( P ℓ − 1 ℓ ) T M ℓ P ℓ − 1 ℓ . (119) M_{\ell-1}=(P^\ell_{\ell-1})^{\mathsf{T}}M_{\ell}P^\ell_{\ell-1}. \tag*{(119)} M ℓ − 1 = ( P ℓ − 1 ℓ ) T M ℓ P ℓ − 1 ℓ . (119)
Given a fine-level white-noise load b ℓ b_{\ell} b ℓ , define
b ℓ − 1 : = ( P ℓ − 1 ℓ ) T b ℓ . (120) b_{\ell-1}:=(P^\ell_{\ell-1})^{\mathsf{T}}b_{\ell}. \tag*{(120)} b ℓ − 1 := ( P ℓ − 1 ℓ ) T b ℓ . (120)
Then
Cov ( b ℓ − 1 ) = ( P ℓ − 1 ℓ ) T M ℓ P ℓ − 1 ℓ = M ℓ − 1 , (121) \operatorname{Cov}(b_{\ell-1})=(P^\ell_{\ell-1})^{\mathsf{T}}M_{\ell}P^\ell_{\ell-1}=M_{\ell-1}, \tag*{(121)} Cov ( b ℓ − 1 ) = ( P ℓ − 1 ℓ ) T M ℓ P ℓ − 1 ℓ = M ℓ − 1 , (121)
so the coarse marginal has exactly the required white-noise law. Moreover,
η ℓ − 1 ( v ℓ − 1 ) = η ℓ ( v ℓ − 1 ) ∀ v ℓ − 1 ∈ V ℓ − 1 , (122) \eta_{\ell-1}(v_{\ell-1})=\eta_{\ell}(v_{\ell-1})\qquad\forall v_{\ell-1}\in V_{\ell-1}, \tag*{(122)} η ℓ − 1 ( v ℓ − 1 ) = η ℓ ( v ℓ − 1 ) ∀ v ℓ − 1 ∈ V ℓ − 1 , (122)
where on the right-hand side v ℓ − 1 v_{\ell-1} v ℓ − 1 is identified with its prolongation to V ℓ V_{\ell} V ℓ . Thus adjacent levels are restrictions of one and the same isonormal Gaussian process.
The coupled latent fields are
Y ℓ : = τ G ℓ η ℓ , Y ℓ − 1 : = τ G ℓ − 1 η ℓ − 1 , (123) Y_{\ell}:=\tau\mathcal{G}_{\ell}\eta_{\ell},\qquad Y_{\ell-1}:=\tau\mathcal{G}_{\ell-1}\eta_{\ell-1}, \tag*{(123)} Y ℓ := τ G ℓ η ℓ , Y ℓ − 1 := τ G ℓ − 1 η ℓ − 1 , (123)
and the represented forward coefficients are
A ~ ℓ ( x , ξ ) : = Φ A ( x , Y ℓ ( x , ξ ) ) , c ~ ℓ ( x , ξ ) : = Φ c ( x , Y ℓ ( x , ξ ) ) . (124) \widetilde{A}_{\ell}(x,\xi):=\Phi_A(x,Y_{\ell}(x,\xi)),\qquad\widetilde{c}_{\ell}(x,\xi):=\Phi_c(x,Y_{\ell}(x,\xi)). \tag*{(124)} A ℓ ( x , ξ ) := Φ A ( x , Y ℓ ( x , ξ )) , c ℓ ( x , ξ ) := Φ c ( x , Y ℓ ( x , ξ )) . (124)
Approximation of the SPDE latent field
The strong approximation of elliptic SPDEs driven by spatial white noise by polytopic symmetric interior-penalty methods is analysed in [28 ] . The analysis uses the canonical restriction of the same continuum white-noise functional to the DG space and separates stochastic projection of the forcing from deterministic SIP approximation of the elliptic resolvent.
We assume that the polytopic hierarchy is quasi-uniform, independently of the chunkiness assumption used in the DG analysis: there exists c q u > 0 c_{\mathrm{qu}}>0 c qu > 0 , independent of ℓ \ell ℓ , such that
c q u h ℓ ≤ h K ≤ h ℓ ∀ K ∈ T ℓ . (125) c_{\mathrm{qu}}h_{\ell}\le h_K\le h_{\ell}\qquad\forall K\in\mathcal{T}_{\ell}. \tag*{(125)} c qu h ℓ ≤ h K ≤ h ℓ ∀ K ∈ T ℓ . (125)
We also assume the standard H 2 H^2 H 2 regularity estimate for the homogeneous Dirichlet reaction–diffusion problem,
∥ L D − 1 f ∥ H 2 ( Ω ) ≲ ∥ f ∥ L 2 ( Ω ) ∀ f ∈ L 2 ( Ω ) . (126) \left\|\mathcal{L}_{D}^{-1}f\right\|_{H^2(\Omega)}\lesssim\|f\|_{L^2(\Omega)}\qquad\forall f\in L^2(\Omega). \tag*{(126)} L D − 1 f H 2 ( Ω ) ≲ ∥ f ∥ L 2 ( Ω ) ∀ f ∈ L 2 ( Ω ) . (126)
Under these assumptions, [28 Theorem 3.15] gives the sharp strong estimate
∥ Y − Y ℓ ∥ L 2 ( Ξ ; L 2 ( Ω ) ) ≲ h ℓ α Y , α Y : = 2 − d 2 , d ∈ { 2 , 3 } . (127) \|Y-Y_{\ell}\|_{L^2(\Xi;L^2(\Omega))}\lesssim h_{\ell}^{\alpha_Y},\qquad\alpha_Y:=2-\frac{d}{2},\qquad d\in\{2,3\}. \tag*{(127)} ∥ Y − Y ℓ ∥ L 2 ( Ξ ; L 2 ( Ω )) ≲ h ℓ α Y , α Y := 2 − 2 d , d ∈ { 2 , 3 } . (127)
Since Y − Y ℓ Y-Y_{\ell} Y − Y ℓ is a centred L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -valued Gaussian random variable, equivalence of finite Gaussian moments yields, for every finite ϱ ≥ 1 \varrho\ge1 ϱ ≥ 1 ,
∥ Y − Y ℓ ∥ L ϱ ( Ξ ; L 2 ( Ω ) ) ≲ h ℓ α Y . (128) \|Y-Y_{\ell}\|_{L^{\varrho}(\Xi;L^2(\Omega))}\lesssim h_{\ell}^{\alpha_Y}. \tag*{(128)} ∥ Y − Y ℓ ∥ L ϱ ( Ξ ; L 2 ( Ω )) ≲ h ℓ α Y . (128)
Thus α Y = 1 \alpha_Y=1 α Y = 1 in d = 2 d=2 d = 2 and α Y = 1 2 \alpha_Y=\frac{1}{2} α Y = 2 1 in d = 3 d=3 d = 3 .
Lemma 4.7 (Propagation to the SPDE coefficients). Assume (80) and the uniform deterministic bounds on the exact and represented coefficients. Then, for every finite ϱ ≥ 1 \varrho\ge1 ϱ ≥ 1 ,
∥ A − A ~ ℓ ∥ L ϱ ( Ξ ; L 2 ( Ω ) ) + ∥ c − c ~ ℓ ∥ L ϱ ( Ξ ; L 2 ( Ω ) ) ≲ h ℓ 2 − d / 2 . (129) \|A-\widetilde{A}_{\ell}\|_{L^{\varrho}(\Xi;L^2(\Omega))}+\|c-\widetilde{c}_{\ell}\|_{L^{\varrho}(\Xi;L^2(\Omega))}\lesssim h_{\ell}^{2-d/2}. \tag*{(129)} ∥ A − A ℓ ∥ L ϱ ( Ξ ; L 2 ( Ω )) + ∥ c − c ℓ ∥ L ϱ ( Ξ ; L 2 ( Ω )) ≲ h ℓ 2 − d /2 . (129) Moreover, for every q ∈ [ 2 , ∞ ) q\in[2,\infty) q ∈ [ 2 , ∞ ) ,
∥ A − A ~ ℓ ∥ L ϱ ( Ξ ; L q ( Ω ) ) + ∥ c − c ~ ℓ ∥ L ϱ ( Ξ ; L q ( Ω ) ) ≲ h ℓ α s p d e , α s p d e : = 4 − d q . (130) \|A-\widetilde{A}_{\ell}\|_{L^{\varrho}(\Xi;L^q(\Omega))}+\|c-\widetilde{c}_{\ell}\|_{L^{\varrho}(\Xi;L^q(\Omega))}\lesssim h_{\ell}^{\alpha_{\mathrm{spde}}},\qquad\alpha_{\mathrm{spde}}:=\frac{4-d}{q}. \tag*{(130)} ∥ A − A ℓ ∥ L ϱ ( Ξ ; L q ( Ω )) + ∥ c − c ℓ ∥ L ϱ ( Ξ ; L q ( Ω )) ≲ h ℓ α spde , α spde := q 4 − d . (130) Proof. The L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) estimate follows from (80) and (122). For q ∈ [ 2 , ∞ ) q\in[2,\infty) q ∈ [ 2 , ∞ ) , interpolation gives
∥ v ∥ L q ( Ω ) ≤ ∥ v ∥ L 2 ( Ω ) 2 / q ∥ v ∥ L ∞ ( Ω ) 1 − 2 / q . \|v\|_{L^q(\Omega)}\le\|v\|_{L^2(\Omega)}^{2/q}\|v\|_{L^\infty(\Omega)}^{1-2/q}. ∥ v ∥ L q ( Ω ) ≤ ∥ v ∥ L 2 ( Ω ) 2/ q ∥ v ∥ L ∞ ( Ω ) 1 − 2/ q . The coefficient errors are uniformly bounded in L ∞ ( Ω ) L^\infty(\Omega) L ∞ ( Ω ) by the deterministic coefficient bounds. Applying this inequality pathwise, taking moments and using (122) gives α s p d e = 2 q ( 2 − d 2 ) = 4 − d q \alpha_{\mathrm{spde}}=\frac{2}{q}\left(2-\frac{d}{2}\right)=\frac{4-d}{q} α spde = q 2 ( 2 − 2 d ) = q 4 − d .
Transfer to the forward problem
We now combine the sampling estimates with the lifting-based Strang bound of § 3.7 . Let q , r ∈ [ 2 , ∞ ] q,r\in[2,\infty] q , r ∈ [ 2 , ∞ ] satisfy
1 q + 1 r + 1 2 = 1 , (131) \frac{1}{q}+\frac{1}{r}+\frac{1}{2}=1, \tag*{(131)} q 1 + r 1 + 2 1 = 1 , (131)
and let
1 r < s ≤ p , (132) \frac{1}{r}<s\le p, \tag*{(132)} r 1 < s ≤ p , (132)
where p p p is the polynomial degree of the forward DG space.
To make the level dependence explicit, define
M s , r , ℓ ( u ; A ) : = ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) (133) \mathcal{M}_{s,r,\ell}(u;A):=\|u\|_{W^{1+s,r}(\Omega)}+\|A\nabla u\|_{W^{s,r}(\mathcal{T}_{\ell})} \tag*{(133)} M s , r , ℓ ( u ; A ) := ∥ u ∥ W 1 + s , r ( Ω ) + ∥ A ∇ u ∥ W s , r ( T ℓ ) (133)
and set
M s , r : = sup 0 ≤ ℓ ≤ L M s , r , ℓ ( u ; A ) . (134) \mathcal{M}_{s,r}:=\sup_{0\leq\ell\leq L}\mathcal{M}_{s,r,\ell}(u;A). \tag*{(134)} M s , r := 0 ≤ ℓ ≤ L sup M s , r , ℓ ( u ; A ) . (134)
We assume
M s , r ∈ L 2 ϱ ( Ξ ) (135) \mathcal{M}_{s,r}\in L^{2\varrho}(\Xi) \tag*{(135)} M s , r ∈ L 2 ϱ ( Ξ ) (135)
for the probability exponent ϱ ∈ [ 1 , ∞ ) \varrho\in[1,\infty) ϱ ∈ [ 1 , ∞ ) under consideration. This is a regularity assumption on the random forward problem and is independent of the latent-field approximation estimates above. Define the level-ℓ \ell ℓ coefficient error
D ℓ : = ∥ A − A ~ ℓ ∥ L q ( Ω ) + ∥ c − c ~ ℓ ∥ L q ( Ω ) . (136) D_\ell:=\left\|A-\widetilde{A}_\ell\right\|_{L^q(\Omega)}+\left\|c-\widetilde{c}_\ell\right\|_{L^q(\Omega)}. \tag*{(136)} D ℓ := A − A ℓ L q ( Ω ) + ∥ c − c ℓ ∥ L q ( Ω ) . (136)
Theorem 4.10 (Transfer from coefficient approximation to forward error). Suppose that
∥ D ℓ ∥ L 2 ϱ ( Ξ ) ≲ ε ℓ . (137) \left\|D_\ell\right\|_{L^{2\varrho}(\Xi)}\lesssim\varepsilon_\ell. \tag*{(137)} ∥ D ℓ ∥ L 2 ϱ ( Ξ ) ≲ ε ℓ . (137) Then
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ; L ϱ ( Ξ ) ≲ ( h ℓ s + ε ℓ ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . (138) \left\|u-u_\ell\right\|_{DG,\ell;\widetilde{A}_\ell,\widetilde{c}_\ell;L^\varrho(\Xi)}
\lesssim\left(h_\ell^s+\varepsilon_\ell\right)\left\|\mathcal{M}_{s,r}\right\|_{L^{2\varrho}(\Xi)}. \tag*{(138)} ∥ u − u ℓ ∥ D G , ℓ ; A ℓ , c ℓ ; L ϱ ( Ξ ) ≲ ( h ℓ s + ε ℓ ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . (138)
Proof. The pathwise estimate (83) gives
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ≲ ( h ℓ s + D ℓ ) M s , r . \left\|u-u_\ell\right\|_{DG,\ell;\widetilde{A}_\ell,\widetilde{c}_\ell}\lesssim\left(h_\ell^s+D_\ell\right)\mathcal{M}_{s,r}. ∥ u − u ℓ ∥ D G , ℓ ; A ℓ , c ℓ ≲ ( h ℓ s + D ℓ ) M s , r .
Taking the L ϱ ( Ξ ) L^\varrho(\Xi) L ϱ ( Ξ ) norm and applying Hölder’s inequality to D ℓ M s , r D_\ell\mathcal{M}_{s,r} D ℓ M s , r yields
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ; L ϱ ( Ξ ) ≲ h ℓ s ∥ M s , r ∥ L ϱ ( Ξ ) + ∥ D ℓ ∥ L 2 ϱ ( Ξ ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . \left\|u-u_\ell\right\|_{DG,\ell;\widetilde{A}_\ell,\widetilde{c}_\ell;L^\varrho(\Xi)}
\lesssim h_\ell^s\left\|\mathcal{M}_{s,r}\right\|_{L^\varrho(\Xi)}+\left\|D_\ell\right\|_{L^{2\varrho}(\Xi)}\left\|\mathcal{M}_{s,r}\right\|_{L^{2\varrho}(\Xi)}. ∥ u − u ℓ ∥ D G , ℓ ; A ℓ , c ℓ ; L ϱ ( Ξ ) ≲ h ℓ s ∥ M s , r ∥ L ϱ ( Ξ ) + ∥ D ℓ ∥ L 2 ϱ ( Ξ ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) .
The result follows from (137) . □
FFT coefficients. For the FFT construction define
ε f f t : = ∥ Y − Y r e f ∥ L q ( Ω ) L 2 ϱ ( Ξ ) . (139) \varepsilon_{\mathrm{fft}}:=\left\|Y-Y_{\mathrm{ref}}\right\|_{L^q(\Omega)\,L^{2\varrho}(\Xi)}. \tag*{(139)} ε fft := ∥ Y − Y ref ∥ L q ( Ω ) L 2 ϱ ( Ξ ) . (139)
Since the same reference representation is used on every level,
∥ D ℓ ∥ L 2 ϱ ( Ξ ) ≲ ε f f t ∀ ℓ = 0 , … , L . (140) \left\|D_\ell\right\|_{L^{2\varrho}(\Xi)}\lesssim\varepsilon_{\mathrm{fft}}\qquad\forall\ell=0,\ldots,L. \tag*{(140)} ∥ D ℓ ∥ L 2 ϱ ( Ξ ) ≲ ε fft ∀ ℓ = 0 , … , L . (140)
Corollary 4.11 (FFT-based coefficient sampling). Under the hypotheses of Theorem 4.10,
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ; L ϱ ( Ξ ) ≲ ( h ℓ s + ε f f t ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . (141) \left\|u-u_\ell\right\|_{DG,\ell;\widetilde{A}_\ell,\widetilde{c}_\ell;L^\varrho(\Xi)}
\lesssim\left(h_\ell^s+\varepsilon_{\mathrm{fft}}\right)\left\|\mathcal{M}_{s,r}\right\|_{L^{2\varrho}(\Xi)}. \tag*{(141)} ∥ u − u ℓ ∥ D G , ℓ ; A ℓ , c ℓ ; L ϱ ( Ξ ) ≲ ( h ℓ s + ε fft ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . (141)
For a sequence of increasingly fine target hierarchies, one may choose N r e f = N r e f ( L ) N_{\mathrm{ref}}=N_{\mathrm{ref}}(L) N ref = N ref ( L ) so that, for example,
ε f f t ≲ h L α f f t . (142) \varepsilon_{\mathrm{fft}}\lesssim h_L^{\alpha_{\mathrm{fft}}}. \tag*{(142)} ε fft ≲ h L α fft . (142)
The same represented reference field is then used for every level in that hierarchy. In particular, α f f t ≥ s \alpha_{\mathrm{fft}}\geq s α fft ≥ s prevents reference-field approximation from degrading the forward spatial rate on the finest level.
Let u r e f u^{\mathrm{ref}} u ref denote the continuum forward solution driven by A ~ r e f : = Φ A ( ⋅ , Y r e f ) \widetilde{A}_{\mathrm{ref}}:=\Phi_A(\cdot,Y_{\mathrm{ref}}) A ref := Φ A ( ⋅ , Y ref ) and c ~ r e f : = Φ c ( ⋅ , Y r e f ) \widetilde{c}_{\mathrm{ref}}:=\Phi_c(\cdot,Y_{\mathrm{ref}}) c ref := Φ c ( ⋅ , Y ref ) . Every u ℓ u_\ell u ℓ in a fixed FFT hierarchy approximates this same represented continuum solution. Thus ε f f t \varepsilon_{\mathrm{fft}} ε fft contributes to the bias relative to the target continuum model, while it does not form a level-dependent representation defect in the correction Q ℓ − Q ℓ − 1 Q_\ell-Q_{\ell-1} Q ℓ − Q ℓ − 1 .
SPDE coefficients.
Corollary 4.12 (SPDE-based coefficient sampling). Let q ∈ [ 2 , ∞ ) q\in[2,\infty) q ∈ [ 2 , ∞ ) and assume (125) , (126) , Lemma 4.7 and the hypotheses of Theorem 4.10. Set
α s p d e : = 4 − d q . (143) \alpha_{\mathrm{spde}}:=\frac{4-d}{q}. \tag*{(143)} α spde := q 4 − d . (143) Then
∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ; L ϱ ( Ξ ) ≲ ( h ℓ s + h ℓ α s p d e ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . (144) \left\|u-u_\ell\right\|_{DG,\ell;\widetilde{A}_\ell,\widetilde{c}_\ell;L^\varrho(\Xi)}
\lesssim\left(h_\ell^s+h_\ell^{\alpha_{\mathrm{spde}}}\right)\left\|\mathcal{M}_{s,r}\right\|_{L^{2\varrho}(\Xi)}. \tag*{(144)} ∥ u − u ℓ ∥ D G , ℓ ; A ℓ , c ℓ ; L ϱ ( Ξ ) ≲ ( h ℓ s + h ℓ α spde ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) . (144)
Quantities of interest. Since u ℓ u_\ell u ℓ is discontinuous, a quantity of interest originally defined only on V = H 0 1 ( Ω ) V=H^1_0(\Omega) V = H 0 1 ( Ω ) must admit an extension to the discrete spaces. We therefore assume that Q Q Q is defined on V + V ℓ V+V_\ell V + V ℓ and satisfies
∣ Q ( w ) − Q ( v ) ∣ ≤ L Q ∥ w − v ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ (145) |Q(w)-Q(v)|\le L_Q\lVert w-v\rVert_{\mathrm{DG},\ell;\widetilde{A}_\ell,\widetilde{c}_\ell} \tag*{(145)} ∣ Q ( w ) − Q ( v ) ∣ ≤ L Q ∥ w − v ∥ DG , ℓ ; A ℓ , c ℓ (145)
whenever w ∈ V w\in V w ∈ V and v ∈ V ℓ v\in V_\ell v ∈ V ℓ . Regional averages and bounded L 2 L^2 L 2 functionals satisfy such an estimate directly. It follows that
∥ Q ( u ) − Q ( u ℓ ) ∥ L ϱ ( Ξ ) ≲ ∥ ∥ u − u ℓ ∥ D G , ℓ ; A ~ ℓ , c ~ ℓ ∥ L ϱ ( Ξ ) . (146) \lVert Q(u)-Q(u_\ell)\rVert_{L^\varrho(\Xi)}
\lesssim
\left\lVert\lVert u-u_\ell\rVert_{\mathrm{DG},\ell;\widetilde{A}_\ell,\widetilde{c}_\ell}\right\rVert_{L^\varrho(\Xi)}. \tag*{(146)} ∥ Q ( u ) − Q ( u ℓ ) ∥ L ϱ ( Ξ ) ≲ ∥ u − u ℓ ∥ DG , ℓ ; A ℓ , c ℓ L ϱ ( Ξ ) . (146)
Hence
∥ Q ( u ) − Q ( u ℓ ) ∥ L ϱ ( Ξ ) ≲ ( h ℓ s + ε f f t ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) (147) \lVert Q(u)-Q(u_\ell)\rVert_{L^\varrho(\Xi)}
\lesssim
\left(h_\ell^s+\varepsilon_{\mathrm{fft}}\right)\lVert\mathcal{M}_{s,r}\rVert_{L^{2\varrho}(\Xi)} \tag*{(147)} ∥ Q ( u ) − Q ( u ℓ ) ∥ L ϱ ( Ξ ) ≲ ( h ℓ s + ε fft ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) (147)
for the FFT construction, while
∥ Q ( u ) − Q ( u ℓ ) ∥ L ϱ ( Ξ ) ≲ ( h ℓ s + h ℓ ( 4 − d ) / q ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) (148) \lVert Q(u)-Q(u_\ell)\rVert_{L^\varrho(\Xi)}
\lesssim
\left(h_\ell^s+h_\ell^{(4-d)/q}\right)\lVert\mathcal{M}_{s,r}\rVert_{L^{2\varrho}(\Xi)} \tag*{(148)} ∥ Q ( u ) − Q ( u ℓ ) ∥ L ϱ ( Ξ ) ≲ ( h ℓ s + h ℓ ( 4 − d ) / q ) ∥ M s , r ∥ L 2 ϱ ( Ξ ) (148)
for the SPDE construction. These strong single-level estimates provide the bias and correction bounds used in the multilevel analysis below.
Multilevel estimator and total error decomposition
We now combine the single-level approximation estimates of § 4.7 with the standard MLMC bias–variance decomposition [20 ] . The two sampling strategies induce different coefficient approximations: the FFT construction uses one common represented field throughout a target hierarchy, while the SPDE construction uses level-dependent latent fields coupled through consistent restrictions of the same white noise.
Multilevel estimator
Let { T ℓ } ℓ = 0 L \{\mathcal{T}_\ell\}_{\ell=0}^{L} { T ℓ } ℓ = 0 L be the forward mesh hierarchy, let u ( ξ ) ∈ V u(\xi)\in V u ( ξ ) ∈ V denote the continuum solution and let u ℓ ( ξ ) ∈ V ℓ u_\ell(\xi)\in V_\ell u ℓ ( ξ ) ∈ V ℓ denote the corresponding DG solution. Define
Q ∗ ( ξ ) : = Q ( u ( ξ ) ) , Q ℓ ( ξ ) : = Q ( u ℓ ( ξ ) ) , ℓ = 0 , … , L . (149) Q^*(\xi):=Q(u(\xi)),\qquad Q_\ell(\xi):=Q(u_\ell(\xi)),\qquad\ell=0,\ldots,L. \tag*{(149)} Q ∗ ( ξ ) := Q ( u ( ξ )) , Q ℓ ( ξ ) := Q ( u ℓ ( ξ )) , ℓ = 0 , … , L . (149)
For ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , the pair ( Q ℓ , Q ℓ − 1 ) (Q_\ell,Q_{\ell-1}) ( Q ℓ , Q ℓ − 1 ) is generated from one coupled random input: a common reference-field realisation for FFT and a fine white-noise load together with its restriction (120) for SPDE.
Set
δ Q 0 : = Q 0 , δ Q ℓ : = Q ℓ − Q ℓ − 1 , ℓ = 1 , … , L . (150) \delta Q_0:=Q_0,\qquad\delta Q_\ell:=Q_\ell-Q_{\ell-1},\qquad\ell=1,\ldots,L. \tag*{(150)} δ Q 0 := Q 0 , δ Q ℓ := Q ℓ − Q ℓ − 1 , ℓ = 1 , … , L . (150)
The prescribed level marginals give
E [ Q L ] = ∑ ℓ = 0 L E [ δ Q ℓ ] . (151) \mathbb{E}[Q_L]=\sum_{\ell=0}^{L}\mathbb{E}[\delta Q_\ell]. \tag*{(151)} E [ Q L ] = ℓ = 0 ∑ L E [ δ Q ℓ ] . (151)
The coupling determines the correction variance V [ δ Q ℓ ] \mathbb{V}[\delta Q_\ell] V [ δ Q ℓ ] and hence the efficiency of the multilevel estimator [12 ] .
For independent samples { δ Q ℓ ( n ) } n = 1 N ℓ \{\delta Q_\ell^{(n)}\}_{n=1}^{N_\ell} { δ Q ℓ ( n ) } n = 1 N ℓ on each level, define
Q ^ L M L : = ∑ ℓ = 0 L 1 N ℓ ∑ n = 1 N ℓ δ Q ℓ ( n ) . (152) \widehat{Q}^{\mathrm{ML}}_L:=\sum_{\ell=0}^{L}\frac{1}{N_\ell}\sum_{n=1}^{N_\ell}\delta Q_\ell^{(n)}. \tag*{(152)} Q L ML := ℓ = 0 ∑ L N ℓ 1 n = 1 ∑ N ℓ δ Q ℓ ( n ) . (152)
Then
E [ Q ^ L M L ] = E [ Q L ] , V [ Q ^ L M L ] = ∑ ℓ = 0 L V [ δ Q ℓ ] N ℓ . (153) \mathbb{E}\left[\widehat{Q}^{\mathrm{ML}}_L\right]=\mathbb{E}[Q_L],\qquad
\mathbb{V}\left[\widehat{Q}^{\mathrm{ML}}_L\right]=\sum_{\ell=0}^{L}\frac{\mathbb{V}[\delta Q_\ell]}{N_\ell}. \tag*{(153)} E [ Q L ML ] = E [ Q L ] , V [ Q L ML ] = ℓ = 0 ∑ L N ℓ V [ δ Q ℓ ] . (153)
Let C ℓ C_\ell C ℓ denote the expected work of one sample of δ Q ℓ \delta Q_\ell δ Q ℓ , including generation of the coupled random input and the associated forward solves. The expected total work is
E [ W o r k ] = ∑ ℓ = 0 L N ℓ C ℓ . (154) \mathbb{E}[\mathrm{Work}]=\sum_{\ell=0}^{L}N_\ell C_\ell. \tag*{(154)} E [ Work ] = ℓ = 0 ∑ L N ℓ C ℓ . (154)
Minimising (154) subject to sampling variance ε s a m p 2 \varepsilon_{\mathrm{samp}}^{2} ε samp 2 gives
N ℓ ∝ V [ δ Q ℓ ] C ℓ (155) N_{\ell} \propto\sqrt{\frac{\mathbb{V}[\delta Q_{\ell}]}{C_{\ell}}} \tag*{(155)} N ℓ ∝ C ℓ V [ δ Q ℓ ] (155)
with integer allocation
N ℓ = ⌈ ε s a m p − 2 ( ∑ j = 0 L V [ δ Q j ] C j ) V [ δ Q ℓ ] C ℓ ⌉ . (156) N_{\ell} = \left\lceil\varepsilon_{\mathrm{samp}}^{-2}\left(\sum_{j=0}^{L}\sqrt{\mathbb{V}[\delta Q_{j}]C_{j}}\right)\sqrt{\frac{\mathbb{V}[\delta Q_{\ell}]}{C_{\ell}}}\right\rceil. \tag*{(156)} N ℓ = ε samp − 2 ( j = 0 ∑ L V [ δ Q j ] C j ) C ℓ V [ δ Q ℓ ] . (156)
Bias and total error
The mean-square error with respect to the continuum quantity E [ Q ⋆ ] \mathbb{E}[Q^{\star}] E [ Q ⋆ ] is
E [ ( Q L M L − E [ Q ⋆ ] ) 2 ] = ∣ E [ Q ⋆ ] − E [ Q L ] ∣ 2 + ∑ ℓ = 0 L V [ δ Q ℓ ] N ℓ . (157) \mathbb{E}\left[\left(Q_{L}^{\mathrm{ML}}-\mathbb{E}[Q^{\star}]\right)^{2}\right]=\left|\mathbb{E}[Q^{\star}]-\mathbb{E}[Q_{L}]\right|^{2}+\sum_{\ell=0}^{L}\frac{\mathbb{V}[\delta Q_{\ell}]}{N_{\ell}}. \tag*{(157)} E [ ( Q L ML − E [ Q ⋆ ] ) 2 ] = ∣ E [ Q ⋆ ] − E [ Q L ] ∣ 2 + ℓ = 0 ∑ L N ℓ V [ δ Q ℓ ] . (157)
Let q , r ∈ [ 2 , ∞ ] q,r\in[2,\infty] q , r ∈ [ 2 , ∞ ] and s > 0 s>0 s > 0 satisfy
1 q + 1 r + 1 2 = 1 , 1 r < s ≤ p , (158) \frac{1}{q}+\frac{1}{r}+\frac{1}{2}=1,\qquad\frac{1}{r}<s\le p, \tag*{(158)} q 1 + r 1 + 2 1 = 1 , r 1 < s ≤ p , (158)
and assume
M s , r ∈ L 2 ( Ξ ) . (159) \mathcal{M}_{s,r}\in L^{2}(\Xi). \tag*{(159)} M s , r ∈ L 2 ( Ξ ) . (159)
The DG-continuity of Q Q Q and Corollaries 4.11 and 4.12 then give the following bias bounds.
FFT bias
Corollary (FFT bias). For the common-reference FFT construction,
∣ E [ Q ⋆ ] − E [ Q L ] ∣ ≲ ( h L s + ε f f t ) ∥ M s , r ∥ L 2 ( Ξ ) . (160) \left|\mathbb{E}[Q^{\star}]-\mathbb{E}[Q_{L}]\right|\lesssim\left(h_{L}^{s}+\varepsilon_{\mathrm{fft}}\right)\left\|\mathcal{M}_{s,r}\right\|_{L^{2}(\Xi)}. \tag*{(160)} ∣ E [ Q ⋆ ] − E [ Q L ] ∣ ≲ ( h L s + ε fft ) ∥ M s , r ∥ L 2 ( Ξ ) . (160) If the master reference resolution is selected so that
ε f f t ≲ h L α f f t , (161) \varepsilon_{\mathrm{fft}}\lesssim h_{L}^{\alpha_{\mathrm{fft}}}, \tag*{(161)} ε fft ≲ h L α fft , (161) then
∣ E [ Q ⋆ ] − E [ Q L ] ∣ ≲ ( h L s + h L α f f t ) ∥ M s , r ∥ L 2 ( Ξ ) . (162) \left|\mathbb{E}[Q^{\star}]-\mathbb{E}[Q_{L}]\right|\lesssim\left(h_{L}^{s}+h_{L}^{\alpha_{\mathrm{fft}}}\right)\left\|\mathcal{M}_{s,r}\right\|_{L^{2}(\Xi)}. \tag*{(162)} ∣ E [ Q ⋆ ] − E [ Q L ] ∣ ≲ ( h L s + h L α fft ) ∥ M s , r ∥ L 2 ( Ξ ) . (162) In particular, α f f t ≥ s \alpha_{\mathrm{fft}}\ge s α fft ≥ s gives the spatial bias rate h L s h_{L}^{s} h L s .
FFT representation error and level corrections
Remark (FFT representation error and level corrections). Let u r e f u^{\mathrm{ref}} u ref denote the continuum solution driven by A r e f = Φ A ( ⋅ , Y r e f ) A^{\mathrm{ref}}=\Phi_{A}(\cdot,Y^{\mathrm{ref}}) A ref = Φ A ( ⋅ , Y ref ) and E c r e f = Φ c ( ⋅ , Y r e f ) \mathcal{E}_{\mathrm{c}}^{\mathrm{ref}}=\Phi_{\mathrm{c}}(\cdot,Y^{\mathrm{ref}}) E c ref = Φ c ( ⋅ , Y ref ) , and set Q r e f : = Q ( u r e f ) Q^{\mathrm{ref}}:=Q(u^{\mathrm{ref}}) Q ref := Q ( u ref ) . Both Q ℓ Q_{\ell} Q ℓ and Q ℓ − 1 Q_{\ell-1} Q ℓ − 1 approximate this same represented continuum output. Thus ε f f t \varepsilon_{\mathrm{fft}} ε fft controls the bias between Q r e f Q^{\mathrm{ref}} Q ref and Q ⋆ Q^{\star} Q ⋆ , while the coupled correction Q ℓ − Q ℓ − 1 Q_{\ell}-Q_{\ell-1} Q ℓ − Q ℓ − 1 measures spatial approximation of Q r e f Q^{\mathrm{ref}} Q ref .
SPDE bias
Corollary (SPDE bias). Let q ∈ [ 2 , ∞ ) q\in[2,\infty) q ∈ [ 2 , ∞ ) and set α s p d e : = ( 4 − d ) / q \alpha_{\mathrm{spde}}:=(4-d)/q α spde := ( 4 − d ) / q . Under the hypotheses of Corollary 4.12,
∣ E [ Q ⋆ ] − E [ Q L ] ∣ ≲ ( h L s + h L α s p d e ) ∥ M s , r ∥ L 2 ( Ξ ) . (163) \left|\mathbb{E}[Q^{\star}]-\mathbb{E}[Q_{L}]\right|\lesssim\left(h_{L}^{s}+h_{L}^{\alpha_{\mathrm{spde}}}\right)\left\|\mathcal{M}_{s,r}\right\|_{L^{2}(\Xi)}. \tag*{(163)} ∣ E [ Q ⋆ ] − E [ Q L ] ∣ ≲ ( h L s + h L α spde ) ∥ M s , r ∥ L 2 ( Ξ ) . (163)
Correction variance and computational complexity
Assume that the hierarchy is quasi-geometric,
h ℓ ≃ h 0 ρ − ℓ , ρ > 1 , (164) h_{\ell}\simeq h_{0}\rho^{-\ell},\qquad\rho>1, \tag*{(164)} h ℓ ≃ h 0 ρ − ℓ , ρ > 1 , (164)
and that, for exponents α , β , γ > 0 \alpha,\beta,\gamma>0 α , β , γ > 0 ,
∣ E [ Q ⋆ − Q ℓ ] ∣ ≲ h ℓ α , V [ δ Q ℓ ] ≲ h ℓ β , C ℓ ≲ h ℓ − γ . (165) \left|\mathbb{E}[Q^{\star}-Q_{\ell}]\right|\lesssim h_{\ell}^{\alpha},\qquad\mathbb{V}[\delta Q_{\ell}]\lesssim h_{\ell}^{\beta},\qquad C_{\ell}\lesssim h_{\ell}^{-\gamma}. \tag*{(165)} ∣ E [ Q ⋆ − Q ℓ ] ∣ ≲ h ℓ α , V [ δ Q ℓ ] ≲ h ℓ β , C ℓ ≲ h ℓ − γ . (165)
The bias estimates above give
α = min { s , α c o e f f } , α c o e f f = { α f f t , FFT , α s p d e , SPDE . (166) \alpha=\min\{s,\alpha_{\mathrm{coeff}}\},\qquad
\alpha_{\mathrm{coeff}}=
\begin{cases}
\alpha_{\mathrm{fft}}, & \text{FFT},\\
\alpha_{\mathrm{spde}}, & \text{SPDE}.
\end{cases} \tag*{(166)} α = min { s , α coeff } , α coeff = { α fft , α spde , FFT , SPDE . (166)
The exponents β \beta β and γ \gamma γ quantify, respectively, the decay of the coupled correction variance and the growth of coupled-sample work.
The following standard MLMC complexity result is recalled in the form used here; see [20 ] .
Theorem 5.7 (MLMC complexity). Assume (158)–(159) and
α ≥ 1 2 min { β , γ } . (167) \alpha\ge\frac{1}{2}\min\{\beta,\gamma\}. \tag*{(167)} α ≥ 2 1 min { β , γ } . (167) Choose L L L so that
h L α ≲ ε (168) h_L^\alpha\lesssim\varepsilon \tag*{(168)} h L α ≲ ε (168) and allocate samples according to (150) with sampling variance of order ε 2 \varepsilon^2 ε 2 . Then the resulting estimator has root-mean-square error of order ε \varepsilon ε and expected work satisfying
E [ W o r k ] ≲ { ε − 2 , β > γ , ε − 2 ( log ε − 1 ) 2 , β = γ , ε − 2 − ( γ − β ) / α , β < γ . (169) \mathbb{E}[\mathrm{Work}] \lesssim
\begin{cases}
\varepsilon^{-2}, & \beta>\gamma,\\
\varepsilon^{-2}(\log\varepsilon^{-1})^2, & \beta=\gamma,\\
\varepsilon^{-2-(\gamma-\beta)/\alpha}, & \beta<\gamma.
\end{cases} \tag*{(169)} E [ Work ] ≲ ⎩ ⎨ ⎧ ε − 2 , ε − 2 ( log ε − 1 ) 2 , ε − 2 − ( γ − β ) / α , β > γ , β = γ , β < γ . (169)
Remark. (Common-reference FFT cost). For the common-reference FFT construction, N r e f = N r e f ( L ) N_{\mathrm{ref}}=N_{\mathrm{ref}}(L) N ref = N ref ( L ) makes reference-field generation a hierarchy-dependent component of the coupled-sample cost. The complete measured costs used in § 6 include reference-field generation together with the forward solves.
Numerical experiments
The experiments examine the spatial accuracy of the polytopic DG method, approximation of the random coefficients, decay of the multilevel corrections, and the resulting MLMC work. All forward problems are posed on Ω = ( 0 , 1 ) 2 \Omega=(0,1)^2 Ω = ( 0 , 1 ) 2 and use the lifting-based symmetric interior-penalty method of § 3 with discontinuous P1 _1 1 polynomials and penalty parameter γ = 12 \gamma=12 γ = 12 . They are implemented within the package [EP26].
The finest meshes are clipped Voronoi tessellations. Coarser levels are obtained by exact agglomeration. The hierarchy used for the field visualisations contains 4, 16, 64, 256 and 1024 cells. The finest Voronoi mesh has between four and eight faces per cell, while the agglomerates have increasingly complex polytopic boundaries. The star-ball diagnostic is positive on every element of the displayed hierarchy.
Table 1. Agglomerated Voronoi hierarchy. The final two columns report the minimum chunkiness parameter and the mean number of polytopic faces.
Except in the verification experiment, we solve
− ∇ ⋅ ( a ( x , ξ ) ∇ u ( x , ξ ) ) + c ( x , ξ ) u ( x , ξ ) = 4 π 2 sin ( π x 1 ) sin ( π x 2 ) , u ∣ ∂ Ω = 0 , (170) -\nabla\cdot(a(x,\xi)\nabla u(x,\xi))+c(x,\xi)u(x,\xi)=4\pi^2\sin(\pi x_1)\sin(\pi x_2),\qquad u|_{\partial\Omega}=0, \tag*{(170)} − ∇ ⋅ ( a ( x , ξ ) ∇ u ( x , ξ )) + c ( x , ξ ) u ( x , ξ ) = 4 π 2 sin ( π x 1 ) sin ( π x 2 ) , u ∣ ∂ Ω = 0 , (170)
with
a ( x , ξ ) = exp ( 0.25 Π [ − 2.5 , 2.5 ] Y ( x , ξ ) ) , c ( x , ξ ) = 2 π 2 exp ( 0.15 Π [ − 2.5 , 2.5 ] Y ( x , ξ ) ) . (171) a(x,\xi)=\exp\left(0.25\Pi_{[-2.5,2.5]}Y(x,\xi)\right),\qquad c(x,\xi)=2\pi^2\exp\left(0.15\Pi_{[-2.5,2.5]}Y(x,\xi)\right). \tag*{(171)} a ( x , ξ ) = exp ( 0.25 Π [ − 2.5 , 2.5 ] Y ( x , ξ ) ) , c ( x , ξ ) = 2 π 2 exp ( 0.15 Π [ − 2.5 , 2.5 ] Y ( x , ξ ) ) . (171)
The quantity of interest is the domain average
Q ( u ) : = 1 ∣ Ω ∣ ∫ Ω u ( x ) d x . (172) Q(u):=\frac{1}{|\Omega|}\int_\Omega u(x)\,\mathrm{d}x. \tag*{(172)} Q ( u ) := ∣Ω∣ 1 ∫ Ω u ( x ) d x . (172)
Both random-field constructions use κ = 8 \kappa=8 κ = 8 and are normalised to unit bulk latent variance. The FFT sampler uses
S ( k ) ∝ ( κ 2 + ∣ k ∣ 2 ) − 2 (173) S(k)\propto\left(\kappa^2+|k|^2\right)^{-2} \tag*{(173)} S ( k ) ∝ ( κ 2 + ∣ k ∣ 2 ) − 2 (173)
on a periodic box of side length 1.5, with the zero Fourier mode retained. The SPDE sampler uses
( κ 2 − Δ ) Y = 4 π κ W (174) (\kappa^2-\Delta)Y=\sqrt{4\pi}\,\kappa W \tag*{(174)} ( κ 2 − Δ ) Y = 4 π κW (174)
with homogeneous Dirichlet conditions. The factor 4 π κ \sqrt{4\pi\kappa} 4 π κ gives unit variance for the corresponding whole-space field in two dimensions. Thus the two samplers have matched Matérn-type bulk spectra and identical coefficient maps, although the boundary condition modifies the SPDE covariance near ∂ Ω \partial\Omega ∂ Ω and the two constructions are not coupled pathwise to one another.
Coupled coefficient fields and solutions
Figure 1 compares representative coefficients and forward solutions on the three finest displayed levels, containing 64, 256 and 1024 polytopes. Panels (a) and (b) show FFT and SPDE sampling, respectively.
Figure 1. Coupled diffusion coefficients and P 1 P_1 P 1 DG solutions on agglomerated Voronoi meshes. Rows contain 64, 256 and 1024 cells from top to bottom; columns within each panel show the coefficient and solution. FFT uses a common reference-field realisation, while SPDE uses consistently restricted white-noise loads.
For FFT, one reference-field realisation is evaluated on every polytopic mesh. The coefficient is therefore pathwise common across the hierarchy. The corresponding solutions converge visibly as the polytopic mesh is refined.
For SPDE sampling, one fine white-noise load is restricted according to b ℓ − 1 = ( P ℓ − 1 ℓ ) T b ℓ b_{\ell-1} = (\mathbf{P}^{\ell}_{\ell-1})^{\mathsf{T}} b_{\ell} b ℓ − 1 = ( P ℓ − 1 ℓ ) T b ℓ , and the latent SPDE is solved separately on each level. The coefficient fields are therefore distinct approximations driven by the same underlying discrete white noise.
Forward discretisation verification
For spatial verification we solve
− Δ u + 2 π 2 u = 4 π 2 u , u ( x 1 , x 2 ) = sin ( π x 1 ) sin ( π x 2 ) , (175) -\Delta u + 2\pi^{2}u = 4\pi^{2}u,\qquad u(x_{1},x_{2}) = \sin(\pi x_{1})\sin(\pi x_{2}), \tag*{(175)} − Δ u + 2 π 2 u = 4 π 2 u , u ( x 1 , x 2 ) = sin ( π x 1 ) sin ( π x 2 ) , (175)
on independently generated Voronoi meshes with 4 4 4 –1024 1024 1024 cells; Figure 2 reports the resulting spatial convergence.
Figure 2. Spatial convergence on independently generated Voronoi meshes.
Coefficient representation
For the FFT sampler, lower-resolution fields are obtained by spectral restriction of common 128 2 128^2 12 8 2 master realisations. The latent and diffusion errors exhibit approximately first-order decay with the reference spacing, with fitted exponents close to 1.06 1.06 1.06 ; see Figure 3 (a).
Figure 3. FFT coefficient representation: (a) latent-field, diffusion coefficient and QoI errors under refinement of a common spectral representation; (b) interaction between forward discretisation and coefficient-representation errors.
Refining the forward mesh with a fixed, under-resolved FFT field produces a representation-error floor. Refining the random field together with the forward discretisation removes this floor, directly illustrating the two contributions in (77); see Figure 3 (b).
The corresponding strong approximation of the SPDE latent field is studied separately in [28 ] , where the sharp mean-square rate for the polytopic SIP approximation driven by canonical discrete white noise is established and verified numerically. We therefore focus here on the inter-level coupling and its effect on the forward multilevel corrections.
Multilevel correction decay
The correction experiment uses the six-level agglomerated hierarchy reported in Table 2 and Table 3 . For FFT, the fitted weak, variance and cost exponents are
Table 2. FFT multilevel correction statistics.
Table 3. Fitted rates: α = 1.40 \alpha= 1.40 α = 1.40 , β = 2.95 \beta= 2.95 β = 2.95 , γ = 1.29 \gamma= 1.29 γ = 1.29 . Table 3. SPDE multilevel correction statistics.
α ^ f f t = 1.45 , β ^ f f t = 3.12 , γ ^ f f t = 1.27 , (176) \widehat{\alpha}_{\mathrm{fft}} = 1.45,\qquad\widehat{\beta}_{\mathrm{fft}} = 3.12,\qquad\widehat{\gamma}_{\mathrm{fft}} = 1.27, \tag*{(176)} α fft = 1.45 , β fft = 3.12 , γ fft = 1.27 , (176)
while the SPDE coupling gives
α ^ s p d e = 1.39 , β ^ s p d e = 2.93 , γ ^ s p d e = 1.28. (177) \widehat{\alpha}_{\mathrm{spde}} = 1.39,\qquad\widehat{\beta}_{\mathrm{spde}} = 2.93,\qquad\widehat{\gamma}_{\mathrm{spde}} = 1.28. \tag*{(177)} α spde = 1.39 , β spde = 2.93 , γ spde = 1.28. (177)
Thus β ^ > γ ^ \widehat{\beta} > \widehat{\gamma} β > γ for both coupling strategies. Figure 4 , panels (a) and (b), shows the mean corrections and correction variances, respectively.
Figure 4. (a) Mean corrections. (b) Correction variances. (c) Coupling diagnostics. Multilevel correction decay and coupling, with FFT in the top row and SPDE in the bottom row. The columns show mean corrections, correction variances, and the correction-to-fine variance ratio together with 1 − ρ 1-\rho 1 − ρ .
The coupling becomes extremely strong on the finest pair. At the precision reported in Table 2 and Table 3 , the neighbouring-level correlations approach one, while
Var ^ [ δ Q ℓ ] Var ^ [ Q ℓ ] = 1.30 × 10 − 5 for FFT , 1.42 × 10 − 5 for SPDE (178) \frac{\widehat{\operatorname{Var}}[\delta Q_\ell]}{\widehat{\operatorname{Var}}[Q_\ell]} = 1.30 \times10^{-5}\quad\text{for FFT},\qquad1.42 \times10^{-5}\quad\text{for SPDE} \tag*{(178)} Var [ Q ℓ ] Var [ δ Q ℓ ] = 1.30 × 1 0 − 5 for FFT , 1.42 × 1 0 − 5 for SPDE (178)
on the finest pair. The decay of the variance ratios and 1 − ρ 1-\rho 1 − ρ is shown in Figure 4 (c).
End-to-end MLMC work
Finally, we compare complete MLMC estimators with finest-level Monte Carlo using measured sample costs. For each tolerance, pilot statistics determine the multilevel allocation and the complete estimators are then repeated independently to measure their realised RMSE.
The work–accuracy curves in Figure 5 (a) show that MLMC reduces the measured work for both coefficient samplers over the full tolerance range considered, with the gain increasing as the tolerance is tightened. This behaviour is consistent with the rapid decay of the coupled correction variances observed above. Panels (b) and (c) give the FFT and SPDE sample allocations at the tightest tolerance.
Figure 5. Measured MLMC performance. In (a), solid curves denote MLMC and dotted curves finest-level Monte Carlo, with a common colour for each sampler. Panels (b) and (c) show sample allocations at the tightest displayed tolerance.
Table 4. Realised RMSE and measured work for MLMC and finest-level Monte Carlo.
The experiments confirm the expected P 1 P_1 P 1 polytopic DG rates and demonstrate the distinct roles of coefficient representation and inter-level coupling. The FFT experiments exhibit the representation-error contribution predicted by the Strang estimate, while both coupling strategies produce rapidly decaying multilevel corrections. The observed correction-variance exponents exceed the measured growth of coupled-sample cost, and the resulting MLMC estimators substantially reduce the measured work relative to finest-level Monte Carlo.
Conclusions
We have developed a multilevel Monte Carlo framework for coercive elliptic problems with random diffusion and reaction coefficients on general polytopic meshes. The forward problem is discretised by a lifting-based symmetric interior penalty DG method, and the analysis separates spatial discretisation error from error introduced by the representation of the random coefficients. Two coefficient-generation strategies were considered: a common-reference FFT construction and an SPDE-based Matérn construction coupled through projected white-noise loads on agglomerated hierarchies. The numerical experiments on Voronoi and agglomerated polytopic meshes confirm the expected forward rates, illustrate the coefficient-representation contribution to the error, and show strong decay of the coupled multilevel corrections.
References
¹ Institute for Mathematical Innovation, University of Bath, Bath, UK. ² Department of Mathematical Sciences, University of Bath, Bath, UK.
The Maxwell Institute for Mathematical Sciences & Department of Mathematics,, Heriot-Watt University, Edinburgh EH14 4AS, UK, and, Department of Mathematics, School of Applied Mathematical and Physical Sciences,, National Technical University of Athens, Zografou 15780, Greece, and, IACM-FORTH, Greece
Email address: e.georgoulis@hw.ac.uk