betaFeedback

WorksMathematics

Quantitative homogenization of Stokes flow in a perforated domain under low regularity of data and geometry

Abstract

We study the homogenization of the steady Stokes equations with no-slip conditions in a domain perforated by ε\varepsilon-periodic obstacles, whose limit is Darcy's law for the homogenized pressure p∗p^\ast. We prove the quantitative error estimate with rate ε\sqrt\varepsilon, natural in the presence of a boundary layer, only assuming that f−∇p∗∈H1f-\nabla p^\ast\in H^1. Such estimates were previously available only under strong smoothness assumptions, such as f∈C1,1/2f\in C^{1,1/2}. Our geometrical assumptions are also less stringent than in the previous analyses. The outer domain and the reference fluid cell need only be Lipschitz, the perforations may deviate from periodicity in a layer of width O(ε)O(\varepsilon) near the outer boundary, and the solid part may be connected (in three dimensions) and attached to the outer wall. In the absence of a boundary layer, the same argument yields the rate O(ε)O(\varepsilon) assuming only f∈H1f\in H^1.

Introduction

We consider the Stokes problem in the perforated domain Ωε:=Ω∖Bε‾\Omega^\varepsilon:= \Omega\setminus\overline{B^\varepsilon}, Ω⊂Rd\Omega\subset\mathbb{R}^d, d=2d=2 or 33: find uε:Ωε→Rdu^\varepsilon: \Omega^\varepsilon\to\mathbb{R}^d and pε:Ωε→Rp^\varepsilon: \Omega^\varepsilon\to\mathbb{R} solution of

−Δuε+∇pε=fin Ωε,div⁡uε=0in Ωε,uε=0on ∂Ωε.(1)\begin{aligned} -\Delta u^\varepsilon+ \nabla p^\varepsilon&= f &&\text{in } \Omega^\varepsilon,\\ \operatorname{div} u^\varepsilon&= 0 &&\text{in } \Omega^\varepsilon,\\ u^\varepsilon&= 0 &&\text{on } \partial\Omega^\varepsilon. \tag*{(1)} \end{aligned}

with a given f∈L2(Ω)df \in L^2(\Omega)^d. BεB^\varepsilon represents periodic perforations placed on a Cartesian grid of period ε\varepsilon. To define them, we start from the unit reference cell Y=(0,1)dY=(0,1)^d, introduce an open set B⊂YB \subset Y (the reference obstacle) and F=Y∖B‾\mathcal{F}=Y\setminus\overline{B} (the reference fluid domain); the precise geometric hypotheses are collected in Assumptions 1 and 2 below. For ε>0\varepsilon>0 and k∈Zdk\in\mathbb{Z}^d let Ykε=ε(Y+k)Y_k^\varepsilon=\varepsilon(Y+k), Bkε=ε(B+k)B_k^\varepsilon=\varepsilon(B+k), Fkε=ε(F+k)\mathcal{F}_k^\varepsilon=\varepsilon(\mathcal{F}+k). A typical choice is

I={k∈Zd:Ykε∩Ω≠∅},Bε=(⋃k∈IBkε‾∩Ω)∘,Ωε=Ω∖Bε‾.(2)\mathcal{I}=\{k\in\mathbb{Z}^d:Y_k^\varepsilon\cap\Omega\ne\varnothing\},\qquad B^\varepsilon=\left(\bigcup_{k\in\mathcal{I}}\overline{B_k^\varepsilon}\cap\Omega\right)^\circ,\qquad \Omega^\varepsilon=\Omega\setminus\overline{B^\varepsilon}. \tag*{(2)}

The actual definition of Ωε\Omega^\varepsilon may differ from (4) in a band of width ∼ε\sim\varepsilon near the outer boundary ∂Ω\partial\Omega.

To study the behavior of the solution for small ε\varepsilon, one introduces the cell Stokes problems: for i=1,…,di=1,\ldots,d, find wi:F→Rdw_i:\mathcal{F}\to\mathbb{R}^d and πi:F→R\pi_i:\mathcal{F}\to\mathbb{R}, Zd\mathbb{Z}^d-periodic, such that

−Δwi+∇πi=eiin F,div⁡wi=0in F,wi=0on ∂B,∫Fπi=0.(3)\begin{aligned} -\Delta w_i+\nabla\pi_i &= e_i &&\text{in } \mathcal{F},\\ \operatorname{div} w_i &= 0 &&\text{in } \mathcal{F},\\ w_i &= 0 &&\text{on } \partial B,\\ \int_{\mathcal{F}}\pi_i &= 0. \tag*{(3)} \end{aligned}

The formal two-scale asymptotic expansion, cf. e.g. [10, 17], suggests that (uε,pε)(u^\varepsilon,p^\varepsilon) is close to (u∗,p∗)(u^*,p^*) with

u∗=ε2wi(⋅ε)(fi−∂ip∗),(4)u^*=\varepsilon^2 w_i\left(\frac{\cdot}{\varepsilon}\right)(f_i-\partial_i p^*), \tag*{(4)}

where p∗p^* satisfies the Darcy problem on the whole (non-perforated) domain

div⁡A∗(f−∇p∗)=0in Ω,(5)\operatorname{div} A^*(f-\nabla p^*)=0 \quad\text{in }\Omega, \tag*{(5)}
n⋅A∗(f−∇p∗)=0on ∂Ω.(6)n\cdot A^*(f-\nabla p^*)=0 \quad\text{on }\partial\Omega. \tag*{(6)}

The constant matrix A∗A^* is defined by A∗ei=⟨wi⟩A^*e_i=\langle w_i\rangle, i=1,…,di=1,\ldots,d, where ⟨⋅⟩\langle\cdot\rangle stands for the average on F\mathcal{F}:

⟨v⟩=1∣F∣∫Fv(y) dy.(7)\langle v\rangle=\frac{1}{|\mathcal{F}|}\int_{\mathcal{F}}v(y)\,dy. \tag*{(7)}

The homogenization of the Stokes equations was first rigorously studied in [23], where the weak L2L^2 convergence of the velocity and the strong L2L^2 convergence of the pressure were established. The strong L2L^2 convergence of the velocity was later proven in [2], where the case of a connected BεB^\varepsilon was also considered (as opposed to the case of BεB^\varepsilon composed of isolated obstacles strictly inside each periodic cell YkεY_k^\varepsilon in the earlier literature). More recently, convergence rates of the type

1ε2∥uε−u∗∥L2(Ωε)+1ε∣uε−u∗∣H1(Ωε)+∥pε−p∗∥L2(Ωε)≤Cε(8)\frac{1}{\varepsilon^2}\lVert u^\varepsilon-u^*\rVert_{L^2(\Omega^\varepsilon)}+\frac{1}{\varepsilon}\lvert u^\varepsilon-u^*\rvert_{H^1(\Omega^\varepsilon)}+\lVert p^\varepsilon-p^*\rVert_{L^2(\Omega^\varepsilon)}\le C\sqrt{\varepsilon} \tag*{(8)}

have been obtained. The first such estimate is available in [15], with the sub-optimal rate ε6\sqrt[6]{\varepsilon} instead of ε\sqrt{\varepsilon}, for d=2d=2. The convergence with the correct order ε\sqrt{\varepsilon} has been proven in [12] for d=2d=2 and in [19, 4, 3] for any dd. Apart from their methods of proof (boundary layer correctors in [19] vs. cut-off techniques elsewhere), these articles differ in their assumptions on the perforations: in [12] and [19], BεB^\varepsilon is composed of isolated obstacles, whereas [4, 3] cover the technically more involved and physically more realistic case of a connected structure BεB^\varepsilon in dimension d=3d=3. All of these papers make rather strong smoothness assumptions on the data ff and on the boundary ∂Ω\partial\Omega. In [19], ff is assumed to be in C1,1/2(Ω‾)dC^{1,1/2}(\overline{\Omega})^d and the boundary ∂B\partial B (hence ∂Bε\partial B^\varepsilon) of class C1,αC^{1,\alpha}. In [4], f−∇p∗f-\nabla p^* is assumed to be in H2(Ω)d∩C1,α(Ω‾)dH^2(\Omega)^d\cap C^{1,\alpha}(\overline{\Omega})^d and ∂Bε\partial B^\varepsilon of class Cm+2C^{m+2} with m>d/2m>d/2. Obtaining an O(ε)O(\sqrt{\varepsilon}) estimate with less regularity on ff, than C1,1/2C^{1,1/2} is singled out as “an interesting and challenging problem” in [19]. Lowering the regularity assumptions on ∂Bε\partial B^\varepsilon is arguably even more important for applications in engineering and in numerical analysis, where BεB^\varepsilon typically has corners and edges.

The purpose of the present article is to establish an estimate of type (10) under weaker regularity assumptions on ff and ∂Bε\partial B^\varepsilon. We only assume that f−∇p∗∈H1(Ω)df-\nabla p^*\in H^1(\Omega)^d. We make no smoothness assumptions on Ω\Omega or on the periodic perforations beyond Lipschitz conditions plus some mild restrictions on the behavior of the perforations near ∂Ω\partial\Omega (where they are allowed to deviate from the periodic pattern). In particular, BεB^\varepsilon can be connected (in dimension d=3d=3) and touch the outer boundary ∂Ω\partial\Omega. We note in passing that the preceding works, even those allowing for a connected structure, assume that the structure is separated from the wall ∂Ω\partial\Omega by a free space (filled by fluid) of width ∼ε\sim\varepsilon, which is a non-physical situation of a structure floating freely in the fluid. The only exception is Appendix B of [3], where a connected BεB^\varepsilon touching ∂Ω\partial\Omega is allowed, but Ω\Omega is restricted to be a box ∏i=1d(0,niε)\prod_{i=1}^d(0,n_i\varepsilon) with integers nin_i.

Under the assumptions of the present article, the solutions (wi,πi)(w_i,\pi_i) of the cell problems have only the regularity H1(F)d×L2(F)H^1(\mathcal{F})^d\times L^2(\mathcal{F}), and one cannot guarantee that u∗u^*, as defined by (6), belongs to H1(Ωε)dH^1(\Omega^\varepsilon)^d. We circumvent this problem by replacing f−∇p∗f-\nabla p^* in the definition of u∗u^* with its interpolant by continuous piecewise polynomial functions on the mesh naturally associated with the periodic perforations, i.e. the uniform Cartesian mesh of size ε\varepsilon. We thus introduce the regularized homogenized velocity

u^∗=ε2wi(⋅ε)Iε0(fi−∂ip∗),(9)\hat{u}^*=\varepsilon^2 w_i\left(\frac{\cdot}{\varepsilon}\right)I_\varepsilon^0(f_i-\partial_i p^*), \tag*{(9)}

where Iε0I_{\varepsilon}^{0} is a Clément-type interpolation [7] onto the Q1Q_{1} finite element space on the ε\varepsilon-cells YkεY_{k}^{\varepsilon}. Iε0I_{\varepsilon}^{0} produces continuous functions vanishing on ∂Ω\partial\Omega and, if necessary, in a band of width ∼ε\sim\varepsilon near ∂Ω\partial\Omega. By construction, u^∗∈H01(Ωε)d\hat{u}^{*}\in H_{0}^{1}(\Omega^{\varepsilon})^{d}, so that the estimate

1ε2∥uε−u^∗∥L2(Ωε)+1ε∣uε−u^∗∣H1(Ωε)+∥pε−p∗∥L2(Ωε)≤Cε(10)\frac{1}{\varepsilon^{2}}\lVert u^{\varepsilon}-\hat{u}^{*}\rVert_{L^{2}(\Omega^{\varepsilon})}+\frac{1}{\varepsilon}\lvert u^{\varepsilon}-\hat{u}^{*}\rvert_{H^{1}(\Omega^{\varepsilon})}+\lVert p^{\varepsilon}-p^{*}\rVert_{L^{2}(\Omega^{\varepsilon})}\leq C\sqrt{\varepsilon} \tag*{(10)}

makes sense. This regularization by finite elements is similar to a device introduced in the theory of operator error estimates for elliptic homogenization [5, 24, 25]. There, the classical first-order approximation u0+εχ(⋅/ε)∇u0u_{0}+\varepsilon\chi(\cdot/\varepsilon)\nabla u_{0} is replaced by u0+εχ(⋅/ε)Sε∇u0u_{0}+\varepsilon\chi(\cdot/\varepsilon)S_{\varepsilon}\nabla u_{0}, where SεS_{\varepsilon} is the Steklov average over cubes of size ε\varepsilon. The purpose is the same as here: to make sense of the product χ(⋅/ε)Sε∇u0\chi(\cdot/\varepsilon)S_{\varepsilon}\nabla u_{0} under minimal regularity assumptions. The estimate O(ε)O(\sqrt{\varepsilon}) in the presence of boundary layers is also proved by these methods in [25, 16, 22], see also [18]. Using Q1Q_{1} interpolant in place of SεS_{\varepsilon} has certain advantages in the present setting: the cut-off near ∂Ω\partial\Omega is built into the interpolant, which vanishes outside the interior cells. With SεS_{\varepsilon}, on the contrary, one needs an extension of u0u_{0} beyond Ω\Omega and a separate cut-off function, as in [18].

Besides the differences in the regularity settings and geometrical assumptions, our contribution differs from the papers on periodic homogenization of Stokes equations on perforated domains mentioned above in its method of proof. It borrows ideas from the finite element analysis, introducing an inf-sup condition between velocity and pressure with respect to a mesh-dependent pressure norm, proved by a local argument in the spirit of the macroelement technique [21, 14]. This norm includes the jumps of the cell means of the pressure across the faces of the ε\varepsilon-mesh, which is particularly reminiscent of the stabilized Q1Q_{1}–P0P_{0} finite element methods [11, 20, 14], cf. also the textbook [8].

Assumptions and main result

Assumption 1 (periodic pattern). B⊂Y=(0,1)dB\subset Y=(0,1)^{d} is open and non-empty, F=Y∖B‾\mathcal{F}=Y\setminus\overline{B} is connected and Lipschitz. Moreover, for j=1,…,dj=1,\ldots,d, the fluid part of two cells adjacent in the direction eje_{j}, i.e. F(j)=(F∪(F+ej))\mathcal{F}^{(j)}=\left(\mathcal{F}\cup(\mathcal{F}+e_{j})\right), is connected and Lipschitz.

To formulate the second assumption, we recall the notations Ykε=ε(Y+k)Y_{k}^{\varepsilon}=\varepsilon(Y+k), Fkε=ε(F+k)\mathcal{F}_{k}^{\varepsilon}=\varepsilon(\mathcal{F}+k), for k∈Zdk\in\mathbb{Z}^{d}, and introduce the definition of a special class of domains, in which one can construct the inverse of operator div, cf. Lemma 5 below:

Definition 1 (John domains). A bounded domain ω⊂Rd\omega\subset\mathbb{R}^{d} is a John domain if there exist a point x0∈ωx_{0}\in\omega (the John center) and a constant J>0J>0 such that, for every x∈ωx\in\omega, there is a rectifiable curve γ:[0,L]⟶ω\gamma:[0,L]\longrightarrow\omega parametrized by the arc length and satisfying γ(0)=x\gamma(0)=x, γ(L)=x0\gamma(L)=x_{0}, and dist⁡(γ(t),∂ω)≥t/J\operatorname{dist}(\gamma(t),\partial\omega)\geq t/J for every t∈[0,L]t\in[0,L].

The class of John domains is bigger than that of Lipschitz, but it excludes certain geometrical features, like external cusps or narrow channels connecting two parts of the domain (assuming a moderate value for the John constant JJ).

Assumption 2 (perforated domain and mesh). Ω⊂Rd\Omega\subset\mathbb{R}^{d}, d=2d=2 or 33, is a bounded connected Lipschitz domain and Ωε⊂Ω\Omega^{\varepsilon}\subset\Omega is open. Moreover, Ω‾\overline{\Omega} is covered by a mesh Tε\mathcal{T}_{\varepsilon} of closed cells with pairwise disjoint interiors, Ω‾=⋃T∈TεT\overline{\Omega}=\bigcup_{T\in\mathcal{T}_{\varepsilon}}T, divided into two disjoint families Tε=Tεin∪Tεbl\mathcal{T}_{\varepsilon}=\mathcal{T}_{\varepsilon}^{\mathrm{in}}\cup\mathcal{T}_{\varepsilon}^{\mathrm{bl}} such that:

(a) (interior cells) Tεin={Ykε:k∈I0}\mathcal{T}_{\varepsilon}^{\mathrm{in}}=\{Y_{k}^{\varepsilon}:k\in\mathcal{I}_{0}\} for a set I0⊂Zd\mathcal{I}_{0}\subset\mathbb{Z}^{d}, and Ωε∩Ykε=Fkε\Omega^{\varepsilon}\cap Y_{k}^{\varepsilon}=\mathcal{F}_{k}^{\varepsilon} for all k∈I0k\in\mathcal{I}_{0};

(b) (boundary layer) every T∈TεblT\in\mathcal{T}_{\varepsilon}^{\mathrm{bl}} satisfies T⊂Ω‾T\subset\overline{\Omega} and dist⁡(x,∂Ω)≤c0ε\operatorname{dist}(x,\partial\Omega)\leq c_{0}\varepsilon for all x∈Tx\in T;

(c) (shape of the layer cells) every T∈TεblT\in\mathcal{T}_{\varepsilon}^{\mathrm{bl}} is of diameter ≤c2ε\leq c_{2}\varepsilon; ε−1T\varepsilon^{-1}T is Lipschitz with constants independent of ε\varepsilon; the fluid part of any such cell, i.e. T∩ΩεT\cap\Omega^{\varepsilon}, is a John domain of constant JJ

and of TT has measure ≥c1εd\ge c_{1}\varepsilon^{d}; moreover, for any other cell T′∈TεT' \in\mathcal{T}_{\varepsilon} adjacent to TT, i.e. such that the intersection T∩T′T \cap T' is of positive (d−1)(d-1)-measure, the fluid part (T∪T′)∘∩Ωε(T \cup T')^{\circ} \cap\Omega^{\varepsilon} is also a John domain of constant JJ. There are at most NiN_{\mathrm{i}} such cells T′T' for any T∈TεblT \in\mathcal{T}_{\varepsilon}^{\mathrm{bl}}.

The parameters c0,c1,c2,J,Nic_{0},c_{1},c_{2},J,N_{\mathrm{i}} are independent of ε\varepsilon.

Remark 1 (consequences of the assumptions)

(i) Assumption 1 implies that Fper=⋃k∈Zd(F+k)\mathcal{F}_{\mathrm{per}}=\bigcup_{k\in\mathbb{Z}^{d}}(\mathcal{F}+k) is connected, hence A∗A^{*} is symmetric positive definite. Indeed, ξ⋅A∗ξ=1∣F∣∫F∣∇wξ∣2\xi\cdot A^{*}\xi=\frac{1}{|\mathcal{F}|}\int_{\mathcal{F}}|\nabla w_{\xi}|^{2} with wξ=ξiwiw_{\xi}=\xi_{i}w_{i}, πξ=ξiπi\pi_{\xi}=\xi_{i}\pi_{i}. If this vanishes, then wξ=0w_{\xi}=0 and ∇πξ=ξ\nabla\pi_{\xi}=\xi on the connected set Fper\mathcal{F}_{\mathrm{per}}, so πξ(y)=ξ⋅y+c\pi_{\xi}(y)=\xi\cdot y+c, which is periodic only if ξ=0\xi=0. Assumption 1 covers isolated Lipschitz inclusions B‾⊂Y\overline{B}\subset Y in any dimension and, for d=3d=3, connected solid structure crossing the faces of YY, provided the fluid remains connected. It excludes tangential contacts between solid parts, which create cusps in the fluid.

(ii) Assumption 2(c) allows for a solid attached to the wall ∂Ω\partial\Omega but forbids tangential contacts or similar situations creating small features of length scale ≪ε\ll\varepsilon. The configuration where a strip of width ∼ε\sim\varepsilon without any solid in it, assumed in previous studies [12, 19, 4], is also allowed. Assumption 2(c) is essentially void in this case, except for the careful choice of the mesh cells TεblT_{\varepsilon}^{\mathrm{bl}} in the boundary layer, which is always possible.

Interpolation operator. Let Ωεin=(⋃k∈I0Ykε)∘\Omega_{\varepsilon}^{\mathrm{in}}=\left(\bigcup_{k\in I_{0}}Y_{k}^{\varepsilon}\right)^{\circ}. For a node a∈εZda\in\varepsilon\mathbb{Z}^{d}, let φa\varphi_{a} be the continuous Q1\mathbb{Q}_{1} hat function of aa, i.e. φa(x)=∏s=1dϕ((xs−as)/ε)\varphi_{a}(x)=\prod_{s=1}^{d}\phi((x_{s}-a_{s})/\varepsilon) with ϕ:R→[0,1]\phi:\mathbb{R}\to[0,1] the piecewise affine function on the grid Z\mathbb{Z} such that ϕ(t)=δt0\phi(t)=\delta_{t0} at all nodes t∈Zt\in\mathbb{Z}. Let ωa\omega_{a} be support of φa\varphi_{a}, i.e. the union of the 2d2^{d} cells of the grid {Ykε}k∈Zd\{Y_{k}^{\varepsilon}\}_{k\in\mathbb{Z}^{d}} having aa as a vertex. If a∈Ωεina\in\Omega_{\varepsilon}^{\mathrm{in}}, then ωa⊂Ωεin\omega_{a}\subset\Omega_{\varepsilon}^{\mathrm{in}}. For v∈L1(Ω)v\in L^{1}(\Omega) we define the Clément-type interpolant

Iε0v=∑a∈εZd∩Ωεin(1∣ωa∣∫ωav)φa.(11)I_{\varepsilon}^{0}v=\sum_{a\in\varepsilon\mathbb{Z}^{d}\cap\Omega_{\varepsilon}^{\mathrm{in}}}\left(\frac{1}{|\omega_{a}|}\int_{\omega_{a}}v\right)\varphi_{a}. \tag*{(11)}

Iε0vI_{\varepsilon}^{0}v is continuous on Ω‾\overline{\Omega}, a Q1\mathbb{Q}_{1} polynomial on every grid cell in Tεbl\mathcal{T}_{\varepsilon}^{\mathrm{bl}}, and vanishes outside Ωεin\Omega_{\varepsilon}^{\mathrm{in}}, so Iε0v∈W1,∞(Rd)I_{\varepsilon}^{0}v\in W^{1,\infty}(\mathbb{R}^{d}).

Theorem 1. Let Assumptions 1 and 2 hold and ε≤1\varepsilon\le1. Let (uε,pε)∈H01(Ωε)d×L2(Ωε)(u^{\varepsilon},p^{\varepsilon})\in H_{0}^{1}(\Omega^{\varepsilon})^{d}\times L^{2}(\Omega^{\varepsilon}) solve (1)–(3), and let p∗∈H1(Ω)p^{*}\in H^{1}(\Omega) be the weak solution of (5)–(6), normalized by ∫Ωεp∗=0\int_{\Omega^{\varepsilon}}p^{*}=0. Assume f−∇p∗∈H1(Ω)df-\nabla p^{*}\in H^{1}(\Omega)^{d}. Then, with u^∗\widehat{u}^{*} defined in (9),

1ε2∥uε−u^∗∥L2(Ωε)+1ε∣uε−u^∗∣H1(Ωε)+∥pε−p∗∥L2(Ωε)≤Cε∥f−∇p∗∥H1(Ω)(12)\frac{1}{\varepsilon^{2}}\lVert u^{\varepsilon}-\widehat{u}^{*}\rVert_{L^{2}(\Omega^{\varepsilon})}+\frac{1}{\varepsilon}|u^{\varepsilon}-\widehat{u}^{*}|_{H^{1}(\Omega^{\varepsilon})}+\lVert p^{\varepsilon}-p^{*}\rVert_{L^{2}(\Omega^{\varepsilon})}\le C\varepsilon\lVert f-\nabla p^{*}\rVert_{H^{1}(\Omega)} \tag*{(12)}

with a constant CC independent of ε\varepsilon. CC depends only on Ω\Omega, F\mathcal{F}, and the constants in Assumption 2.

Remark 2 (on the regularity hypothesis) If Ω\Omega is a domain of class C1,1C^{1,1}, and f∈H1(Ω)df\in H^{1}(\Omega)^{d}, then p∗∈H2(Ω)p^{*}\in H^{2}(\Omega) [9] and ∥f−∇p∗∥H1≤C∥f∥H1\lVert f-\nabla p^{*}\rVert_{H^{1}}\le C\lVert f\rVert_{H^{1}}. Hence, the error bound (12) holds assuming only f∈H1(Ω)df\in H^{1}(\Omega)^{d}. On a general Lipschitz domain, f−∇p∗∈H1(Ω)df-\nabla p^{*}\in H^{1}(\Omega)^{d} is an additional hypothesis on (f,p∗)(f,p^{*}): the Neumann problem (5)–(6) only gives p∗∈H3/2(Ω)p^{*}\in H^{3/2}(\Omega) in general.

Remark 3 (error without boundary layer) If the boundary layer is absent, i.e. if Ω\Omega is a box, consisting of whole ε\varepsilon-cells YkεY_{k}^{\varepsilon}, with periodic conditions on ∂Ω\partial\Omega, and assuming that u^∗\widehat{u}^{*} is based on the interpolant IεI_{\varepsilon} with the sum in (11) extended to all the nodes εZd\varepsilon\mathbb{Z}^{d}, the same proof as below gives

1ε2∥uε−u^∗∥L2(Ωε)+1ε∣uε−u^∗∣H1(Ωε)+∥pε−p∗∥L2(Ωε)≤Cε∥f−∇p∗∥H1(Ω)≤Cε∥f∥H1(Ω).\frac{1}{\varepsilon^{2}}\lVert u^{\varepsilon}-\widehat{u}^{*}\rVert_{L^{2}(\Omega^{\varepsilon})}+\frac{1}{\varepsilon}|u^{\varepsilon}-\widehat{u}^{*}|_{H^{1}(\Omega^{\varepsilon})}+\lVert p^{\varepsilon}-p^{*}\rVert_{L^{2}(\Omega^{\varepsilon})}\le C\varepsilon\lVert f-\nabla p^{*}\rVert_{H^{1}(\Omega)}\le C\varepsilon\lVert f\rVert_{H^{1}(\Omega)}.

It suffices to get rid of all the bounds related to the boundary layer, no longer needed, and to adapt the interpolation error estimates (21)–(24) gaining there an extra ε\sqrt{\varepsilon} (this is seen immediately in the proof of Lemma 9).

Auxiliary results

This section collects several Lemmas necessary for the proof of the main theorem in the forthcoming Section 4. We shall systematically use the letter CC to denote constants depending only on dd, F\mathcal{F}, Ω\Omega and the constants in Assumption 2; they never depend on ε\varepsilon or on the data. We assume ε≤1\varepsilon\le1.

Compact notation. For a Zd\mathbb{Z}^{d}-periodic function ϕ\phi, we systematically write (ϕ)ε(x)=ϕ(x/ε)(\phi)_{\varepsilon}(x)=\phi(x/\varepsilon). For a measurable set DD, ∥⋅∥D\|\cdot\|_{D} denotes the norm in L2(D)L^{2}(D). Summation over repeated indices ii is implied.

Notations related to the ε\varepsilon-mesh and fluid parts. We recall the notion of the mesh Tε\mathcal{T}_{\varepsilon} and its submeshes Tεin\mathcal{T}_{\varepsilon}^{\mathrm{in}}, Tεbl\mathcal{T}_{\varepsilon}^{\mathrm{bl}} from Assumption 2. A face is a set F=T∩T′F=T\cap T^{\prime} with T≠T′∈TεT\ne T^{\prime}\in\mathcal{T}_{\varepsilon} such that the (d−1)(d-1)-measure ∣F∣|F| is positive (∣F∣>0|F|>0). We denote by Eε\mathcal{E}^{\varepsilon} the set of faces and by Einε\mathcal{E}_{\mathrm{in}}^{\varepsilon} the set of faces between two interior cells T≠T′∈TεinT\ne T^{\prime}\in\mathcal{T}_{\varepsilon}^{\mathrm{in}}. For every face we fix once and for all an orientation, i.e. we call one of the two cells, sharing FF, TT and the other T′T^{\prime}, and denote by nFn_{F} the unit normal to FF pointing from TT to T′T^{\prime}.

For any cell T∈TεT\in\mathcal{T}_{\varepsilon} we shall write FTε\mathcal{F}_{T}^{\varepsilon} for its fluid part, i.e. FTε=T∩Ωε\mathcal{F}_{T}^{\varepsilon}=T\cap\Omega^{\varepsilon} (in particular, for the interior cells T=YkεT=Y_{k}^{\varepsilon}, k∈I0k\in I_{0}, the fluid part is FTε=Fkε\mathcal{F}_{T}^{\varepsilon}=F_{k}^{\varepsilon}). Similarly, for a face F=T∩T′F=T\cap T^{\prime} between two cells T,T′∈TεT,T^{\prime}\in\mathcal{T}_{\varepsilon}, we shall denote by FFε\mathcal{F}_{F}^{\varepsilon} the fluid part of the union of these cells, i.e. FFε=(T∪T′)∘∩Ωε\mathcal{F}_{F}^{\varepsilon}=(T\cup T^{\prime})^{\circ}\cap\Omega^{\varepsilon}.

Poincaré and trace inequalities

Lemma 2 (strip estimates on a Lipschitz domain) For t>0t>0, let Ωt={x∈Ω:dist⁡(x,∂Ω)<t}\Omega_{t}=\{x\in\Omega:\operatorname{dist}(x,\partial\Omega)<t\} be the boundary strip of width tt.

(a) if w∈H1(Ω)w\in H^{1}(\Omega), then ∥w∥Ωt≤Ct∥w∥H1(Ω)\|w\|_{\Omega_{t}}\le C\sqrt{t}\|w\|_{H^{1}(\Omega)};

(b) if w∈H01(Ω)w\in H_{0}^{1}(\Omega), then ∥w∥Ωt≤Ct∥∇w∥ΩCt\|w\|_{\Omega_{t}}\le Ct\|\nabla w\|_{\Omega_{Ct}}.

Proof. Part (a) is borrowed from [18 Proposition 3.1.7]. There, the statement is in fact stronger: ∥w∥Ωt≤Ct∥w∥W1,q(Ω)\|w\|_{\Omega_{t}}\le C\sqrt{t}\|w\|_{W^{1,q}(\Omega)} with q=2dd+1q=\frac{2d}{d+1}. Our statement follows from this since q<2q<2.

To prove part (b), we first consider a function v∈H1(Qr+)v\in H^{1}(Q_{r}^{+}), where Qr+={x=(x′,xd)∈Rd:∣x′∣<r,xd>0}Q_{r}^{+}=\{x=(x^{\prime},x_{d})\in\mathbb{R}^{d}:|x^{\prime}|<r,x_{d}>0\}, vanishing on {xd=0}\{x_{d}=0\}. For s>0s>0,

∫∣x′∣<r∣v(x′,s)∣2 dx′≤2∫0s∫∣x′∣<r∣v(x′,xd)∣∣∂dv∣ dx′ dxd.\int_{|x^{\prime}|<r}|v(x^{\prime},s)|^{2}\,\mathrm{d}x^{\prime}\le2\int_{0}^{s}\int_{|x^{\prime}|<r}|v(x^{\prime},x_{d})||\partial_{d}v|\,\mathrm{d}x^{\prime}\,\mathrm{d}x_{d}.

Integrating over s∈(0,t)s\in(0,t) gives

∫0t∫∣x′∣<r∣v∣2 dx′ dxd≤2t∫0t∫∣x′∣<r∣v∣∣∇v∣ dx′ dxd,\int_{0}^{t}\int_{|x^{\prime}|<r}|v|^{2}\,\mathrm{d}x^{\prime}\,\mathrm{d}x_{d}\le2t\int_{0}^{t}\int_{|x^{\prime}|<r}|v||\nabla v|\,\mathrm{d}x^{\prime}\,\mathrm{d}x_{d},

and, by the Cauchy–Schwarz inequality,

(∫0t∫∣x′∣<r∣v∣2 dx′ dxd)1/2≤2t(∫0t∫∣x′∣<r∣∇v∣2 dx′ dxd)1/2.\left(\int_{0}^{t}\int_{|x^{\prime}|<r}|v|^{2}\,\mathrm{d}x^{\prime}\,\mathrm{d}x_{d}\right)^{1/2}\le2t\left(\int_{0}^{t}\int_{|x^{\prime}|<r}|\nabla v|^{2}\,\mathrm{d}x^{\prime}\,\mathrm{d}x_{d}\right)^{1/2}.

Part (b) follows by covering ∂Ω\partial\Omega with coordinate patches in which Ω\Omega lies above the graph of a Lipschitz function γ\gamma, and applying the result above after the bi-Lipschitz change of variables (x′,xd)↦(x′,xd−γ(x′))(x^{\prime},x_{d})\mapsto(x^{\prime},x_{d}-\gamma(x^{\prime})), which maps Ωt\Omega_{t} into a strip of width CtCt.

Lemma 3 (Poincaré inequality) For ϕ∈H01(Ωε)\phi\in H_{0}^{1}(\Omega^{\varepsilon}), ∥ϕ∥Ωε≤Cε∣ϕ∣H1(Ωε)\|\phi\|_{\Omega^{\varepsilon}}\le C\varepsilon|\phi|_{H^{1}(\Omega^{\varepsilon})}.

Proof. Extend ϕ\phi by 00 to Ω\Omega. By Assumption 2, Ω⊂Ωc0ε‾∪⋃k∈I0Ykε‾\Omega\subset\overline{\Omega_{c_{0}\varepsilon}}\cup\bigcup_{k\in I_{0}}\overline{Y_{k}^{\varepsilon}} in the notations of Lemma 2. We can apply part (b) of this Lemma to bound ∥ϕ∥Ωc0ε≤Cε∣ϕ∣H1(Ωc0ε)\|\phi\|_{\Omega_{c_{0}\varepsilon}}\le C\varepsilon|\phi|_{H^{1}(\Omega_{c_{0}\varepsilon})}. On YkεY_{k}^{\varepsilon}, k∈I0k\in I_{0}, ϕ\phi vanishes on BkεB_{k}^{\varepsilon}, and the scaled Poincaré inequality on YY for functions vanishing on the non-empty open set BB gives ∥ϕ∥Ykε≤Cε∣ϕ∣H1(Ykε)\|\phi\|_{Y_{k}^{\varepsilon}}\le C\varepsilon|\phi|_{H^{1}(Y_{k}^{\varepsilon})}.

Lemma 4 (trace inequality on the cells). For every T∈TεT \in\mathcal{T}_{\varepsilon} and w∈H1(T∘)w \in H^{1}(T^{\circ}),

∥w∥∂T2≤C(ε−1∥w∥T2+ε∣w∣H1(T)2).\lVert w\rVert_{\partial T}^{2} \le C\left(\varepsilon^{-1}\lVert w\rVert_{T}^{2}+\varepsilon\lvert w\rvert_{H^{1}(T)}^{2}\right).

Consequently, for w∈H1(Ω)w \in H^{1}(\Omega),

∑F∈Eε∥w∥F2≤C(ε−1∥w∥Ω2+ε∣w∣H1(Ω)2).\sum_{F\in\mathcal{E}^{\varepsilon}}\lVert w\rVert_{F}^{2}\le C\left(\varepsilon^{-1}\lVert w\rVert_{\Omega}^{2}+\varepsilon\lvert w\rvert_{H^{1}(\Omega)}^{2}\right).

Proof. Lemma 4 follows by scaling from the trace inequality ∥w∥∂T^2≤C(∥w∥T^2+∥w∥T^∥∇w∥T^)\lVert w\rVert_{\partial\widehat{T}}^{2}\le C(\lVert w\rVert_{\widehat{T}}^{2}+\lVert w\rVert_{\widehat{T}}\lVert\nabla w\rVert_{\widehat{T}}) on the Lipschitz domain T^=ε−1T\widehat{T}=\varepsilon^{-1}T, whose constant depends only on the Lipschitz character of T^\widehat{T}: it is obtained by integrating div⁡(μw2)\operatorname{div}(\mu w^{2}) with a smooth vector field μ\mu transversal to ∂T^\partial\widehat{T}, constructed from the local graph representations of ∂T^\partial\widehat{T}, see [9 Lemma 1.5.1.9 and Theorem 1.5.1.10]. □\square

Pressure means and the inf-sup condition

Lemma 5 (Bogovskiĭ operator). Let ω⊂Rd\omega\subset\mathbb{R}^{d} be a John domain. For every q∈L02(ω)q \in L_{0}^{2}(\omega) with ∫ωq=0\int_{\omega}q=0 there is v∈H01(ω)dv \in H_{0}^{1}(\omega)^{d} with div⁡v=q\operatorname{div}v=q and ∣v∣H1(ω)≤CB∥q∥ω\lvert v\rvert_{H^{1}(\omega)}\le C_{B}\lVert q\rVert_{\omega}, where the constant CBC_{B} depends only on the John constant JJ of ω\omega as in Definition 1.

Proof. The operator defining vv is constructed in [6]. The fact that this construction is valid on John domains and that CBC_{B} depends only on JJ is proved in [1]. □\square

For q∈L2(Ωε)q \in L^{2}(\Omega^{\varepsilon}) and T∈TεT \in\mathcal{T}_{\varepsilon} let qT=∣FTε∣−1∫FTεqq_{T}=\lvert F_{T}^{\varepsilon}\rvert^{-1}\int_{F_{T}^{\varepsilon}}q and Πεq\Pi_{\varepsilon}q be the piecewise constant function on the mesh Tε\mathcal{T}_{\varepsilon} defined on every T∈TεT\in\mathcal{T}_{\varepsilon} by Πεq∣T=qT\Pi_{\varepsilon}q|_{T}=q_{T}. Πε\Pi_{\varepsilon} maps L2(Ωε)L^{2}(\Omega^{\varepsilon}) into itself (after restricting Πεq\Pi_{\varepsilon}q to the fluid part FTεF_{T}^{\varepsilon} on every cell TT). We introduce the jumps ⟦Πεq⟧F=qT−qT′\llbracket\Pi_{\varepsilon}q\rrbracket_{F}=q_{T}-q_{T'} across any face F∈EεF\in\mathcal{E}^{\varepsilon}, F=T∩T′F=T\cap T', using the conventions adopted in the beginning of this Section. Set

∥q∥ε2=ε∑F∈Eε∥⟦Πεq⟧F∥F2+∥(I−Πε)q∥Ωε2.(13)\lVert q\rVert_{\varepsilon}^{2}=\varepsilon\sum_{F\in\mathcal{E}^{\varepsilon}}\lVert\llbracket\Pi_{\varepsilon}q\rrbracket_{F}\rVert_{F}^{2}+\lVert(I-\Pi_{\varepsilon})q\rVert_{\Omega^{\varepsilon}}^{2}. \tag*{(13)}

Lemma 6 (discrete Poincaré inequality). For every q∈L02(Ωε)q\in L_{0}^{2}(\Omega^{\varepsilon}),

∥Πεq∥Ωε≤Cε−1∥q∥εand∥q∥Ωε≤Cε−1∥q∥ε.(14)\lVert\Pi_{\varepsilon}q\rVert_{\Omega^{\varepsilon}}\le C\varepsilon^{-1}\lVert q\rVert_{\varepsilon} \qquad\text{and}\qquad \lVert q\rVert_{\Omega^{\varepsilon}}\le C\varepsilon^{-1}\lVert q\rVert_{\varepsilon}. \tag*{(14)}

Proof. Let Q∈L2(Ω)Q\in L^{2}(\Omega) be equal to Πεq\Pi_{\varepsilon}q on Ωε\Omega^{\varepsilon} and to 00 in the solid. Then ∫ΩQ=∑T∈Tε∣FTε∣qT=∫Ωεq=0\int_{\Omega}Q=\sum_{T\in\mathcal{T}_{\varepsilon}}\lvert F_{T}^{\varepsilon}\rvert q_{T}=\int_{\Omega^{\varepsilon}}q=0 and ∥Q∥Ω=∥Πεq∥Ωε\lVert Q\rVert_{\Omega}=\lVert\Pi_{\varepsilon}q\rVert_{\Omega^{\varepsilon}}. Let V∈H01(Ω)dV\in H_{0}^{1}(\Omega)^{d} satisfy div⁡V=Q\operatorname{div}V=Q and ∥V∥Ω+∣V∣H1(Ω)≤C∥Q∥Ω\lVert V\rVert_{\Omega}+\lvert V\rvert_{H^{1}(\Omega)}\le C\lVert Q\rVert_{\Omega} (this VV is given by Lemma 5 since the Lipschitz domain Ω\Omega is also John, and the constant CC depends only on Ω\Omega through its John constant and its diameter, involved in the Poincaré–Friedrichs inequality). Observe

∥Πεq∥Ωε2=∑T∈TεqT∫FTεq=∑T∈TεqT∫Tdiv⁡V=∑F∈Eε⟦Πεq⟧FV⋅nF.\lVert\Pi_{\varepsilon}q\rVert_{\Omega^{\varepsilon}}^{2} =\sum_{T\in\mathcal{T}_{\varepsilon}}q_{T}\int_{F_{T}^{\varepsilon}}q =\sum_{T\in\mathcal{T}_{\varepsilon}}q_{T}\int_{T}\operatorname{div}V =\sum_{F\in\mathcal{E}^{\varepsilon}}\llbracket\Pi_{\varepsilon}q\rrbracket_{F}V\cdot n_{F}.

The last equality follows from the divergence theorem on every cell TT, rearranging the sums over the cell boundaries, and the fact that V⋅nV\cdot n vanishes on ∂Ω\partial\Omega, so that only the contributions on faces F∈EεF\in\mathcal{E}^{\varepsilon} survive. We now use the Cauchy–Schwarz inequality and Lemma 4:

∥Πεq∥Ωε2≤(ε∑F∈Eε∥⟦Πεq⟧F∥F2)1/2(ε−1∑F∈Eε∥V∥F2)1/2≤C∥q∥ε(ε−2∥V∥Ω2+∣V∣H1(Ω)2)1/2≤Cε−1∥q∥ε∥Πεq∥Ωε\begin{aligned} \lVert\Pi_{\varepsilon}q\rVert_{\Omega^{\varepsilon}}^{2} \le\left(\varepsilon\sum_{F\in\mathcal{E}^{\varepsilon}}\lVert\llbracket\Pi_{\varepsilon}q\rrbracket_{F}\rVert_{F}^{2}\right)^{1/2} \left(\varepsilon^{-1}\sum_{F\in\mathcal{E}^{\varepsilon}}\lVert V\rVert_{F}^{2}\right)^{1/2} \\ &\le C\lVert q\rVert_{\varepsilon}\left(\varepsilon^{-2}\lVert V\rVert_{\Omega}^{2}+\lvert V\rvert_{H^{1}(\Omega)}^{2}\right)^{1/2} \le C\varepsilon^{-1}\lVert q\rVert_{\varepsilon}\lVert\Pi_{\varepsilon}q\rVert_{\Omega^{\varepsilon}} \end{aligned}

since ε<1\varepsilon<1. This proves the first inequality in (16). The second inequality follows from ∥q∥Ωε2=∥Πεq∥Ωε2+∥(I−Πε)q∥Ωε2\lVert q\rVert_{\Omega^{\varepsilon}}^{2}=\lVert\Pi_{\varepsilon}q\rVert_{\Omega^{\varepsilon}}^{2}+\lVert(I-\Pi_{\varepsilon})q\rVert_{\Omega^{\varepsilon}}^{2} and ε<1\varepsilon<1. □\square

Lemma 7 (velocity-pressure inf-sup). For every p∈L02(Ωε)p \in L^2_0(\Omega^\varepsilon) there is w∈H01(Ωε)dw \in H^1_0(\Omega^\varepsilon)^d with

(p,div⁡w)Ωε=∥ ⁣∥ ⁣∥p∥ ⁣∥ ⁣∥ε2,β∣w∣H1(Ωε)≤∥ ⁣∥ ⁣∥p∥ ⁣∥ ⁣∥ε,(p,\operatorname{div} w)_{\Omega^\varepsilon}=\lVert\!\lVert\!\lVert p\rVert\!\rVert\!\rVert_\varepsilon^2,\qquad\beta|w|_{H^1(\Omega^\varepsilon)}\leq\lVert\!\lVert\!\lVert p\rVert\!\rVert\!\rVert_\varepsilon,

with β∈(0,1]\beta\in(0,1] independent of ε\varepsilon.

Proof. For F∈EεF \in\mathcal{E}^\varepsilon with cells T,T′T,T', Lemma 5 gives ΦF∈H01(Fε)d\Phi_F \in H^1_0(F^\varepsilon)^d with

div⁡ΦF=1TFε∣FTε∣−1T′Fε∣FT′ε∣,∣ΦF∣H1(Fε)≤CB(∣FTε∣−1+∣FT′ε∣−1)1/2≤Cε−d/2,(15)\operatorname{div}\Phi_F=\frac{\mathbf{1}_{T}^{F^\varepsilon}}{|F_T^\varepsilon|}-\frac{\mathbf{1}_{T'}^{F^\varepsilon}}{|F_{T'}^\varepsilon|},\qquad|\Phi_F|_{H^1(F^\varepsilon)}\leq C_B\left(|F_T^\varepsilon|^{-1}+|F_{T'}^\varepsilon|^{-1}\right)^{1/2}\leq C\varepsilon^{-d/2}, \tag*{(15)}

using ∣FTε∣≥cεd|F_T^\varepsilon|\geq c\varepsilon^d (Assumption 2(c)). Domain FεF^\varepsilon is indeed John with constant JJ by Assumption 2(c), if one of the cells T,T′T,T' is in Tb\mathcal{T}^{\mathrm{b}}. For patches FεF^\varepsilon composed of interior cells, the John constant is also independent of ε\varepsilon, cf. Assumption 1, since the John constant is scale invariant.

Set vp=ε∑F∈Eε∣F∣(pT−pT′)ΦFv_p=\varepsilon\sum_{F\in\mathcal{E}^\varepsilon}|F|(p_T-p_{T'})\Phi_F. Since (p,div⁡ΦF)Ωε=pT−pT′(p,\operatorname{div}\Phi_F)_{\Omega^\varepsilon}=p_T-p_{T'}, we get

(p,div⁡vp)Ωε=ε∑F∈Eε∣F∣(pT−pT′)2=ε∑F∈Eε∥ ⁣∥Πεp∥ ⁣∥F2.(p,\operatorname{div}v_p)_{\Omega^\varepsilon}=\varepsilon\sum_{F\in\mathcal{E}^\varepsilon}|F|(p_T-p_{T'})^2=\varepsilon\sum_{F\in\mathcal{E}^\varepsilon}\lVert\!\lVert\Pi_\varepsilon p\rVert\!\rVert_F^2.

The supports of the ΦF\Phi_F overlap boundedly and ∣F∣≤Cεd−1|F|\leq C\varepsilon^{d-1}, so

∣vp∣H1(Ωε)2≤Cε2∑F∈Eε∣F∣2(pT−pT′)2ε−d≤Cε∑F∈Eε∥ ⁣∥Πεp∥ ⁣∥F2.(16)|v_p|_{H^1(\Omega^\varepsilon)}^2\leq C\varepsilon^2\sum_{F\in\mathcal{E}^\varepsilon}|F|^2(p_T-p_{T'})^2\varepsilon^{-d}\leq C\varepsilon\sum_{F\in\mathcal{E}^\varepsilon}\lVert\!\lVert\Pi_\varepsilon p\rVert\!\rVert_F^2. \tag*{(16)}

For every T∈TεT\in\mathcal{T}_\varepsilon, (I−Πε)p(I-\Pi_\varepsilon)p is of zero mean on FTεF_T^\varepsilon. Thus, Lemma 5 gives vT∈H01(FTε)dv_T\in H^1_0(F_T^\varepsilon)^d with div⁡vT=(I−Πε)p\operatorname{div}v_T=(I-\Pi_\varepsilon)p on FTεF_T^\varepsilon and ∣vT∣H1(FTε)≤CB∥(I−Πε)p∥FTε|v_T|_{H^1(F_T^\varepsilon)}\leq C_B\lVert(I-\Pi_\varepsilon)p\rVert_{F_T^\varepsilon}. The constant CBC_B is ε\varepsilon-independent by the same considerations as above.

Let vp0=∑T∈TεvTv_{p_0}=\sum_{T\in\mathcal{T}_\varepsilon}v_T (each extended by zero) and w=vp+vp0w=v_p+v_{p_0}. Then

(p,div⁡w)Ωε=(p,div⁡vp)Ωε+(p−Πεp,div⁡vp0)Ωε=(p,div⁡vp)Ωε+∥(I−Πε)p∥Ωε2=∥ ⁣∥ ⁣∥p∥ ⁣∥ ⁣∥ε2,(p,\operatorname{div}w)_{\Omega^\varepsilon}=(p,\operatorname{div}v_p)_{\Omega^\varepsilon}+(p-\Pi_\varepsilon p,\operatorname{div}v_{p_0})_{\Omega^\varepsilon}=(p,\operatorname{div}v_p)_{\Omega^\varepsilon}+\lVert(I-\Pi_\varepsilon)p\rVert_{\Omega^\varepsilon}^2=\lVert\!\lVert\!\lVert p\rVert\!\rVert\!\rVert_\varepsilon^2,

since div⁡vp0\operatorname{div}v_{p_0} has zero mean on every FTεF_T^\varepsilon. Recalling ∣vp0∣H1(Ωε)≤CB∥(I−Πε)p∥Ωε|v_{p_0}|_{H^1(\Omega^\varepsilon)}\leq C_B\lVert(I-\Pi_\varepsilon)p\rVert_{\Omega^\varepsilon} and combining with (18) gives the announced bound on the norm of ww. □\square

Remark 4 An alternative definition of the mesh-dependent norm on the pressure space could be

∥ ⁣∥q∥ ⁣∥ε,alt2=ε2∑T∈Tε∥Πεq∥T2+∥(I−Πε)q∥Ωε2.(17)\lVert\!\lVert q\rVert\!\rVert_{\varepsilon,\mathrm{alt}}^2=\varepsilon^2\sum_{T\in\mathcal{T}_\varepsilon}\lVert\Pi_\varepsilon q\rVert_T^2+\lVert(I-\Pi_\varepsilon)q\rVert_{\Omega^\varepsilon}^2. \tag*{(17)}

With this definition, Lemma 6 would be a triviality, and Lemma 7 would be simpler to prove. One can work with such definition as well, and obtain almost the same results as with the present definition (15), but the regularity requirement would be stronger (typically f−∇p∗∈H3/2f-\nabla p^*\in H^{3/2}) and the treatment of the boundary layer contributions would be more involved.

We shall also use the combined velocity-pressure space X=H01(Ωε)d×L02(Ωε)\mathcal{X}=H^1_0(\Omega^\varepsilon)^d\times L^2_0(\Omega^\varepsilon), equipped with the norm ∥(v,q)∥ε\lVert(v,q)\rVert_\varepsilon defined for any (v,q)∈X(v,q)\in\mathcal{X} by

∥(v,q)∥ε2=∣v∣H1(Ωε)2+∥ ⁣∥q∥ ⁣∥ε2.(18)\lVert(v,q)\rVert_\varepsilon^2=|v|_{H^1(\Omega^\varepsilon)}^2+\lVert\!\lVert q\rVert\!\rVert_\varepsilon^2. \tag*{(18)}

Introduce the combined bilinear form c:X×X→Rc:\mathcal{X}\times\mathcal{X}\to\mathbb{R} associated to the Stokes equations by

c((u,p),(v,q))=(∇u,∇v)Ωε−(p,div⁡v)Ωε−(q,div⁡u)Ωε.c((u,p),(v,q))=(\nabla u,\nabla v)_{\Omega^\varepsilon}-(p,\operatorname{div}v)_{\Omega^\varepsilon}-(q,\operatorname{div}u)_{\Omega^\varepsilon}.

Lemma 8 (inf-sup for the combined bilinear form cc). For every (u,p)∈X(u,p)\in\mathcal{X}, there exists (v,q)∈X∖{0}(v,q)\in\mathcal{X}\setminus\{0\} such that

c((u,p),(v,q))∥(v,q)∥ε≥β∗∥(u,p)∥ε(19)\frac{c((u,p),(v,q))}{\lVert(v,q)\rVert_\varepsilon}\geq\beta_*\lVert(u,p)\rVert_\varepsilon \tag*{(19)}

with β∗=β2/(23)\beta_*=\beta^2/(2\sqrt{3}) and β\beta from Lemma 7.

Proof. Take any (u,p)∈X(u,p) \in\mathcal{X} and choose ww as in Lemma 7. Then

c((u,p),(−w,0))=−(∇u,∇w)Ωε+∥p∥ε2≥−1β∣u∣H1(Ωε)∥p∥ε+∥p∥ε2,c((u,p),(u,−p))=∣u∣H1(Ωε)2.c((u,p),(-w,0)) = -(\nabla u,\nabla w)_{\Omega_\varepsilon} + \|p\|_\varepsilon^2 \ge-\frac{1}{\beta}|u|_{H^1(\Omega_\varepsilon)}\|p\|_\varepsilon+ \|p\|_\varepsilon^2,\qquad c((u,p),(u,-p)) = |u|_{H^1(\Omega_\varepsilon)}^2.

With v=u−β2wv=u-\beta^2w and q=−pq=-p, we get

c((u,p),(v,q))≥∣u∣H1(Ωε)2−β∣u∣H1(Ωε)∥p∥ε+β2∥p∥ε2≥12∣u∣H1(Ωε)2+β22∥p∥ε2≥β22∥(u,p)∥ε2c((u,p),(v,q)) \ge|u|_{H^1(\Omega_\varepsilon)}^2-\beta|u|_{H^1(\Omega_\varepsilon)}\|p\|_\varepsilon+\beta^2\|p\|_\varepsilon^2 \ge\frac{1}{2}|u|_{H^1(\Omega_\varepsilon)}^2+\frac{\beta^2}{2}\|p\|_\varepsilon^2 \ge\frac{\beta^2}{2}\|(u,p)\|_\varepsilon^2

and

∥(v,q)∥ε2≤(∣u∣H1(Ωε)+β∥p∥ε)2+∥p∥ε2≤3∥(u,p)∥ε2.\|(v,q)\|_\varepsilon^2 \le\left(|u|_{H^1(\Omega_\varepsilon)}+\beta\|p\|_\varepsilon\right)^2+\|p\|_\varepsilon^2 \le3\|(u,p)\|_\varepsilon^2.

Interpolation and the flux corrector

Lemma 9 For v∈H1(Ω)v \in H^1(\Omega), with Iε0I_\varepsilon^0 defined in (13),

∥Iε0v∥Ω≤C∥v∥Ω,(20)\|I_\varepsilon^0v\|_\Omega\le C\|v\|_\Omega, \tag*{(20)}
∥v−Iε0v∥Ω≤Cε∥v∥H1(Ω),(21)\|v-I_\varepsilon^0v\|_\Omega\le C\sqrt{\varepsilon}\|v\|_{H^1(\Omega)}, \tag*{(21)}
∣Iε0v∣H1(Ω)≤Cε∥v∥H1(Ω),(22)|I_\varepsilon^0v|_{H^1(\Omega)} \le\frac{C}{\sqrt{\varepsilon}}\|v\|_{H^1(\Omega)}, \tag*{(22)}
(∑F∈Fεε∥∇Iε0v∥F2)1/2≤Cε∥v∥H1(Ω),(23)\left(\sum_{F\in\mathcal{F}_\varepsilon^{\varepsilon}}\|\nabla I_\varepsilon^0v\|_F^2\right)^{1/2}\le\frac{C}{\varepsilon}\|v\|_{H^1(\Omega)}, \tag*{(23)}
(∑F∈Fε∥v−Iε0v∥F2)1/2≤C∥v∥H1(Ω).(24)\left(\sum_{F\in\mathcal{F}^\varepsilon}\|v-I_\varepsilon^0v\|_F^2\right)^{1/2}\le C\|v\|_{H^1(\Omega)}. \tag*{(24)}

Proof. By scaling and (13), we have for any T∈TεinT\in\mathcal{T}_\varepsilon^{\mathrm{in}},

∥Iε0v∥T2≤Cεd∑a∣Iε0v(a)∣2≤C∑a∥v∥ωa2≤C∥v∥ωT,∣Iε0v∣H1(T)≤Cε∥Iε0v∥T≤Cε∥v∥ωT.(25)\|I_\varepsilon^0v\|_T^2\le C\varepsilon^d\sum_a|I_\varepsilon^0v(a)|^2\le C\sum_a\|v\|_{\omega_a}^2\le C\|v\|_{\omega_T},\qquad|I_\varepsilon^0v|_{H^1(T)}\le\frac{C}{\varepsilon}\|I_\varepsilon^0v\|_T\le\frac{C}{\varepsilon}\|v\|_{\omega_T}. \tag*{(25)}

where the sum is taken over all the vertices of TT inside Ωεin\Omega_\varepsilon^{\mathrm{in}}, and ωT\omega_T denotes the union of node-associated patches ωa\omega_a for all such vertices aa. Bound (22) follows by summing the first inequality in (27) over all T∈TεinT\in\mathcal{T}_\varepsilon^{\mathrm{in}}, since the patches ωT\omega_T overlap boundedly.

Let Tε,in,0\mathcal{T}_{\varepsilon,\mathrm{in},0} be the set of cells T∈TεinT\in\mathcal{T}_\varepsilon^{\mathrm{in}} whose 2d2^d vertices all lie inside Ωεin\Omega_\varepsilon^{\mathrm{in}}, and Tε,in,bl=Tεin∖Tε,in,0\mathcal{T}_{\varepsilon,\mathrm{in},\mathrm{bl}}=\mathcal{T}_\varepsilon^{\mathrm{in}}\setminus\mathcal{T}_{\varepsilon,\mathrm{in},0}. If T∈Tε,in,0T\in\mathcal{T}_{\varepsilon,\mathrm{in},0}, then ωT\omega_T is a cube of side 3ε3\varepsilon and Iε0c=cI_\varepsilon^0c=c on TT for any constant cc. Taking cc as the mean of vv on ωT\omega_T, (27) and the Poincaré inequality on ωT\omega_T give, on T∈Tε,in,0T\in\mathcal{T}_{\varepsilon,\mathrm{in},0},

∥v−Iε0v∥T=∥v−c−Iε0(v−c)∥T≤C∥v−c∥ωT≤Cε∥∇v∥ωT,\|v-I_\varepsilon^0v\|_T=\|v-c-I_\varepsilon^0(v-c)\|_T\le C\|v-c\|_{\omega_T}\le C\varepsilon\|\nabla v\|_{\omega_T},
∣Iε0v∣H1(T)=∣Iε0(v−c)∣H1(T)≤C∥∇v∥ωT.(26)|I_\varepsilon^0v|_{H^1(T)}=|I_\varepsilon^0(v-c)|_{H^1(T)}\le C\|\nabla v\|_{\omega_T}. \tag*{(26)}

If T∈Tε,in,blT\in\mathcal{T}_{\varepsilon,\mathrm{in},\mathrm{bl}}, then TT touches ∂Ωεin\partial\Omega_\varepsilon^{\mathrm{in}}, which lies at distance at most c0εc_0\varepsilon from ∂Ω\partial\Omega by Assumption 2(b). Hence TT and ωT\omega_T, as well as Ωεbl=Ω∖Ωεin\Omega_\varepsilon^{\mathrm{bl}}=\Omega\setminus\Omega_\varepsilon^{\mathrm{in}}, are contained in the strip Ωcε\Omega_{c\varepsilon}. Summing over all cells and recalling Iε0v=0I_\varepsilon^0v=0 on Ωεbl\Omega_\varepsilon^{\mathrm{bl}}, we obtain thanks to Lemma 2(a), and (27)–(28),

∥v−Iε0v∥Ω2=∑T∈Tε,in,0∥v−Iε0v∥T2+∑T∈Tε,in,bl∥v−Iε0v∥T2+∥v∥Ωεbl2≤C(ε2∣v∣H1(Ω)2+∥v∥Ωcε2)≤Cε∥v∥H1(Ω)2,\|v-I_\varepsilon^0v\|_\Omega^2=\sum_{T\in\mathcal{T}_{\varepsilon,\mathrm{in},0}}\|v-I_\varepsilon^0v\|_T^2+\sum_{T\in\mathcal{T}_{\varepsilon,\mathrm{in},\mathrm{bl}}}\|v-I_\varepsilon^0v\|_T^2+\|v\|_{\Omega_\varepsilon^{\mathrm{bl}}}^2\le C\left(\varepsilon^2|v|_{H^1(\Omega)}^2+\|v\|_{\Omega_{c\varepsilon}}^2\right)\le C\varepsilon\|v\|_{H^1(\Omega)}^2,

and similarly

∣Iε0v∣H1(Ω)2≤C(∣v∣H1(Ω)2+ε−2∥v∥Ωcε2)≤Cε∥v∥H1(Ω)2,|I_\varepsilon^0v|_{H^1(\Omega)}^2\le C\left(|v|_{H^1(\Omega)}^2+\varepsilon^{-2}\|v\|_{\Omega_{c\varepsilon}}^2\right)\le\frac{C}{\varepsilon}\|v\|_{H^1(\Omega)}^2,

which are (23) and (24).

It remains to prove the estimates on faces. For (25), a scaling argument on a cube YkεY_k^\varepsilon gives

∥∇Iε0v∥F2≤Cε−1∣Iε0v∣H1(Ykε)2\|\nabla I_\varepsilon^0v\|_F^2\le C\varepsilon^{-1}|I_\varepsilon^0v|_{H^1(Y_k^\varepsilon)}^2

for a face FF of YkεY_k^\varepsilon, k∈I0k\in I_0. Thus, the left-hand side of (25) is bounded by Cε−1/2∣Iε0v∣H1(Ωin)C\varepsilon^{-1/2}\lvert I_\varepsilon^0v\rvert_{H^1(\Omega_{\mathrm{in}})}, and (25) follows from (24). For (26), Lemma 4 gives ∑F∈Eε∥v−Iε0v∥F2≤C(ε−1∥v−Iε0v∥2+ε∣v−Iε0v∣H1(Ω))\sum_{F\in\mathcal{E}_\varepsilon}\lVert v-I_\varepsilon^0v\rVert_F^2 \le C(\varepsilon^{-1}\lVert v-I_\varepsilon^0v\rVert^2+\varepsilon\lvert v-I_\varepsilon^0v\rvert_{H^1(\Omega)} ), so that (26) follows from (23) and (24). □

We now recall the classical flux corrector of periodic homogenization [13, 18].

Lemma 10 (flux corrector). For i=1,…,di=1,\ldots,d there is a Zd\mathbb{Z}^d-periodic skew-symmetric matrix field ϕ(i)∈Hper2(Rd)d×d\phi^{(i)}\in H^2_{\mathrm{per}}(\mathbb{R}^d)^{d\times d} such that

wi−∣F∣A∗ei=div⁡ϕ(i),(div⁡ϕ(i))a=∑b=1d∂bϕab(i).w_i-\lvert\mathcal{F}\rvert A^*e_i=\operatorname{div}\phi^{(i)},\qquad\left(\operatorname{div}\phi^{(i)}\right)_a=\sum_{b=1}^{d}\partial_b\phi_{ab}^{(i)}.

We assume here that wiw_i is extended by 00 in B=Y∖FB=Y\setminus\mathcal{F}.

Proof. Since wi∈H1(F)w_i\in H^1(\mathcal{F}), div⁡wi=0\operatorname{div}w_i=0 in Fper\mathcal{F}_{\mathrm{per}} and wi=0w_i=0 on ∂Fper\partial\mathcal{F}_{\mathrm{per}}, extending wiw_i by 00 in BB gives a field in Hper1(Rd)dH^1_{\mathrm{per}}(\mathbb{R}^d)^d and its divergence vanishes everywhere. Its mean over YY is ∫Fwi=∣F∣A∗ei\int_{\mathcal{F}}w_i=\lvert\mathcal{F}\rvert A^*e_i.

Let ψ∈Hper3(Rd)d\psi\in H^3_{\mathrm{per}}(\mathbb{R}^d)^d solve −Δψ=wi−∣F∣A∗ei-\Delta\psi=w_i-\lvert\mathcal{F}\rvert A^*e_i and set ϕab(i)=∂aψb−∂bψa\phi_{ab}^{(i)}=\partial_a\psi_b-\partial_b\psi_a. Then (div⁡ϕ(i))a=∂adiv⁡ψ−Δψa\left(\operatorname{div}\phi^{(i)}\right)_a=\partial_a\operatorname{div}\psi-\Delta\psi_a. Observe −Δdiv⁡ψ=div⁡wi=0-\Delta\operatorname{div}\psi=\operatorname{div}w_i=0. Thus, div⁡ψ\operatorname{div}\psi is a periodic harmonic function, hence constant, and div⁡ϕ(i)=−Δψ\operatorname{div}\phi^{(i)}=-\Delta\psi. □

Corollary 11. Let hh be a function of class C2C^2 on a grid cell YkεY_k^\varepsilon. Then

ε2∫Ykε(wi)ε⋅∇h=ε2∣F∣∫YkεA∗ei⋅∇h+ε3∫∂Ykε(ϕ(i))ε∇τh⋅n.(27)\varepsilon^2\int_{Y_k^\varepsilon}(w_i)_\varepsilon\cdot\nabla h=\varepsilon^2\lvert\mathcal{F}\rvert\int_{Y_k^\varepsilon}A^*e_i\cdot\nabla h+\varepsilon^3\int_{\partial Y_k^\varepsilon}(\phi^{(i)})_\varepsilon\nabla_\tau h\cdot n. \tag*{(27)}

where ∇τ\nabla_\tau denote the tangential part of the gradient on ∂Ykε\partial Y_k^\varepsilon.

Proof. We write (wi)ε=∣F∣A∗ei+εdiv⁡(ϕ(i))ε(w_i)_\varepsilon=\lvert\mathcal{F}\rvert A^*e_i+\varepsilon\operatorname{div}(\phi^{(i)})_\varepsilon, and perform the integration by parts in the integral

ε3∫Ykεdiv⁡(ϕ(i))ε⋅∇h=ε3∫∂Ykε(ϕ(i))ε∇h⋅n−ε3∫Ykε∑a,b=1d(ϕab(i))ε∂a∂bh.\varepsilon^3\int_{Y_k^\varepsilon}\operatorname{div}(\phi^{(i)})_\varepsilon\cdot\nabla h=\varepsilon^3\int_{\partial Y_k^\varepsilon}(\phi^{(i)})_\varepsilon\nabla h\cdot n-\varepsilon^3\int_{Y_k^\varepsilon}\sum_{a,b=1}^{d}(\phi_{ab}^{(i)})_\varepsilon\partial_a\partial_bh.

The last integral vanishes since ϕ\phi is a skew-symmetric matrix and the matrix of second derivatives of hh is symmetric. Writing, on ∂Ykε\partial Y_k^\varepsilon, ∇h=(∇h⋅n)n+∇τh\nabla h=(\nabla h\cdot n)n+\nabla_\tau h, we conclude ∇h⋅(ϕ(i))εn=(ϕ(i))ε∇τh⋅n\nabla h\cdot(\phi^{(i)})_\varepsilon n=(\phi^{(i)})_\varepsilon\nabla_\tau h\cdot n since (ϕ(i))εn⋅n=0(\phi^{(i)})_\varepsilon n\cdot n=0 by skew-symmetry of ϕ(i)\phi^{(i)}. □

Remark 5 (products with oscillating functions). We shall frequently encounter the products of ε\varepsilon-periodic functions, eg. wi,∇wi,πiw_i,\nabla w_i,\pi_i rescaled by ε\varepsilon, with Q1Q_1 functions on mesh Tεh\mathcal{T}_\varepsilon^h. If one needs to calculate the norm of such a product, then multiplication by the periodic function will act as multiplication by a constant. Indeed, let ψ∈L2(F)\psi\in L^2(\mathcal{F}) be Zd\mathbb{Z}^d-periodic and hh be a polynomial of degree up to dd on YkεY_k^\varepsilon. Scaling estimates give then

∥(ψ)εh∥Fkε≤εd/2∥ψ∥F∥h∥L∞(Ykε)≤C∥ψ∥F∥h∥Ykε≤C∥h∥Ykε.\lVert(\psi)_\varepsilon h\rVert_{\mathcal{F}_k^\varepsilon}\le\varepsilon^{d/2}\lVert\psi\rVert_{\mathcal{F}}\lVert h\rVert_{L^\infty(Y_k^\varepsilon)}\le C\lVert\psi\rVert_{\mathcal{F}}\lVert h\rVert_{Y_k^\varepsilon}\le C\lVert h\rVert_{Y_k^\varepsilon}.

In the last inequality we made the generic constant CC absorb the norm ∥ψ∥F\lVert\psi\rVert_{\mathcal{F}} since, in all the cases encountered, it will depend only on F\mathcal{F}, and the constants CC are allowed to hide this dependence.

Proof of Theorem 1

Let us denote uˉ=f−∇p∗\bar{u}=f-\nabla p^* and recall Definition (6) of u^∗\hat{u}^*, which we rewrite as u^∗=ε2(wi)εIε0uˉ\hat{u}^*=\varepsilon^2(w_i)_\varepsilon I_\varepsilon^0\bar{u} and introduce the pressure reconstruction

pε,1=p∗+ε(πi)εIε0uˉi−mε,mε=1∣Ωε∣∫Ωεε(πi)εIε0uˉi.(28)p^{\varepsilon,1}=p^*+\varepsilon(\pi_i)_\varepsilon I_\varepsilon^0\bar{u}_i-m_\varepsilon,\qquad m_\varepsilon=\frac{1}{\lvert\Omega^\varepsilon\rvert}\int_{\Omega^\varepsilon}\varepsilon(\pi_i)_\varepsilon I_\varepsilon^0\bar{u}_i. \tag*{(28)}

Since Iε0uˉI^0_\varepsilon\bar{u} vanishes outside Ωεin\Omega^{\mathrm{in}}_\varepsilon, u^∗∈H01(Ωε)d\hat{u}^* \in H^1_0(\Omega^\varepsilon)^d. By construction, pε,1∈L02(Ωε)p^{\varepsilon,1} \in L^2_0(\Omega^\varepsilon). Take some z∈H01(Ωε)dz \in H^1_0(\Omega^\varepsilon)^d, to be chosen later in the proof. We have (uε−u^∗+z,pε−pε,1)∈X(u^\varepsilon-\hat{u}^*+z,p^\varepsilon-p^{\varepsilon,1})\in\mathcal{X}. By Lemma 8 there is (v,q)∈X(v,q)\in\mathcal{X} with ∥(v,q)∥ε=1\lVert(v,q)\rVert_\varepsilon=1 and

β∗∥(uε−u^∗+z,pε−pε,1)∥ε≤c((uε−u^∗+z,pε−pε,1),(v,q))=(∇(uε−u^∗+z),∇v)Ωε⏟I−(pε−pε,1,div⁡v)Ωε⏟II−(q,div⁡(uε−u^∗+z))Ωε⏟III,(29)\beta_* \lVert(u^\varepsilon-\hat{u}^*+z,p^\varepsilon-p^{\varepsilon,1})\rVert_\varepsilon\le c((u^\varepsilon-\hat{u}^*+z,p^\varepsilon-p^{\varepsilon,1}),(v,q)) = \underbrace{(\nabla(u^\varepsilon-\hat{u}^*+z),\nabla v)_{\Omega^\varepsilon}}_{I} - \underbrace{(p^\varepsilon-p^{\varepsilon,1},\operatorname{div}v)_{\Omega^\varepsilon}}_{II} - \underbrace{(q,\operatorname{div}(u^\varepsilon-\hat{u}^*+z))_{\Omega^\varepsilon}}_{III}, \tag*{(29)}

We shall now treat the terms on the right-hand side of (29) and choose zz, finally obtaining the announced error estimate.

The terms I−III-II. The weak form of (1) and f=uˉ+∇p∗f=\bar{u}+\nabla p^* give

(∇uε,∇v)Ωε=(uˉ,v)Ωε−(p∗,div⁡v)Ωε+(pε,div⁡v)Ωε.(\nabla u^\varepsilon,\nabla v)_{\Omega^\varepsilon}=(\bar{u},v)_{\Omega^\varepsilon}-(p^*,\operatorname{div}v)_{\Omega^\varepsilon}+(p^\varepsilon,\operatorname{div}v)_{\Omega^\varepsilon}.

We are now going to write out the analogue of this relation for u^∗\hat{u}^*.1

(∇u^∗,∇v)Ωε=ε((∇wi)ε,(Iε0uˉi)∇v)Ωε+ε2(∇Iε0uˉi,∇v(wi)ε)Ωε=ε((∇wi)ε,∇(vIε0uˉi))Ωε−ε(∇Iε0uˉi,(∇wi)εv)Ωε+ε2(∇Iε0uˉi,(∇v)(wi)ε)Ωε.\begin{aligned} (\nabla\hat{u}^*,\nabla v)_{\Omega^\varepsilon}=\varepsilon((\nabla w_i)_\varepsilon,(I^0_\varepsilon\bar{u}_i)\nabla v)_{\Omega^\varepsilon}+\varepsilon^2(\nabla I^0_\varepsilon\bar{u}_i,\nabla v(w_i)_\varepsilon)_{\Omega^\varepsilon} \\ &=\varepsilon((\nabla w_i)_\varepsilon,\nabla(vI^0_\varepsilon\bar{u}_i))_{\Omega^\varepsilon}-\varepsilon(\nabla I^0_\varepsilon\bar{u}_i,(\nabla w_i)_\varepsilon v)_{\Omega^\varepsilon}+\varepsilon^2(\nabla I^0_\varepsilon\bar{u}_i,(\nabla v)(w_i)_\varepsilon)_{\Omega^\varepsilon}. \end{aligned}

The cell problems (3), rescaled by ε\varepsilon, read in the weak form tested by vIε0uˉivI^0_\varepsilon\bar{u}_i,

ε((∇wi)ε,∇(vIε0uˉi))Ωε−ε((πi)ε,div⁡(vIε0uˉi))Ωε=(Iε0uˉ,v)Ωε.\varepsilon((\nabla w_i)_\varepsilon,\nabla(vI^0_\varepsilon\bar{u}_i))_{\Omega^\varepsilon}-\varepsilon((\pi_i)_\varepsilon,\operatorname{div}(vI^0_\varepsilon\bar{u}_i))_{\Omega^\varepsilon}=(I^0_\varepsilon\bar{u},v)_{\Omega^\varepsilon}.

The test function (Iε0uˉi)v(I^0_\varepsilon\bar{u}_i)v is admissible since Iε0uˉi∈W1,∞I^0_\varepsilon\bar{u}_i\in W^{1,\infty} and its support is in Ωεin‾\overline{\Omega^{\mathrm{in}}_\varepsilon}, where the perforations are periodic. Denoting σi=∇wi−πiI\sigma_i=\nabla w_i-\pi_iI, we obtain

(∇u^∗,∇v)Ωε=(Iε0uˉ,v)Ωε−ε(∇Iε0uˉi,(σi)εv)Ωε+ε2(∇Iε0uˉi,(∇v)(wi)ε)Ωε+ε((πi)εIε0uˉi,div⁡v)Ωε.(\nabla\hat{u}^*,\nabla v)_{\Omega^\varepsilon}=(I^0_\varepsilon\bar{u},v)_{\Omega^\varepsilon}-\varepsilon(\nabla I^0_\varepsilon\bar{u}_i,(\sigma_i)_\varepsilon v)_{\Omega^\varepsilon}+\varepsilon^2(\nabla I^0_\varepsilon\bar{u}_i,(\nabla v)(w_i)_\varepsilon)_{\Omega^\varepsilon}+\varepsilon((\pi_i)_\varepsilon I^0_\varepsilon\bar{u}_i,\operatorname{div}v)_{\Omega^\varepsilon}.

Thus,

I−II=(uˉ−Iε0uˉ,v)Ωε+ε(∇Iε0uˉi,(σi)εv)Ωε−ε2(∇Iε0uˉi,(∇v)(wi)ε)Ωε+(∇z,∇v)Ωε.I-II=(\bar{u}-I^0_\varepsilon\bar{u},v)_{\Omega^\varepsilon}+\varepsilon(\nabla I^0_\varepsilon\bar{u}_i,(\sigma_i)_\varepsilon v)_{\Omega^\varepsilon}-\varepsilon^2(\nabla I^0_\varepsilon\bar{u}_i,(\nabla v)(w_i)_\varepsilon)_{\Omega^\varepsilon}+(\nabla z,\nabla v)_{\Omega^\varepsilon}.

The constant mεm_\varepsilon in pε,1p^{\varepsilon,1} does not contribute above, since ∫Ωεdiv⁡v=0\int_{\Omega^\varepsilon}\operatorname{div}v=0. By Lemma 3, (21), (22), Remark 5, and ∣v∣H1(Ωε)≤∥(v,q)∥ε=1|v|_{H^1(\Omega^\varepsilon)}\le\lVert(v,q)\rVert_\varepsilon=1,

∣I−II∣≤C(ε∥uˉ−Iε0uˉ∥Ω+ε2∣Iε0uˉ∣H1(Ω)+∣z∣H1(Ωε))≤Cε3/2∥uˉ∥H1(Ω)+∣z∣H1(Ωε).(30)|I-II|\le C\left(\varepsilon\lVert\bar{u}-I^0_\varepsilon\bar{u}\rVert_\Omega+\varepsilon^2|I^0_\varepsilon\bar{u}|_{H^1(\Omega)}+|z|_{H^1(\Omega^\varepsilon)}\right)\le C\varepsilon^{3/2}\lVert\bar{u}\rVert_{H^1(\Omega)}+|z|_{H^1(\Omega^\varepsilon)}. \tag*{(30)}

The term IIIIII. Since div⁡uε=0\operatorname{div}u^\varepsilon=0,

−III=(Πεq,div⁡u^∗)Ωε+((I−Πε)q,div⁡u^∗)Ωε−(q,div⁡z)Ωε.(31)-III=(\Pi_\varepsilon q,\operatorname{div}\hat{u}^*)_{\Omega^\varepsilon}+((I-\Pi_\varepsilon)q,\operatorname{div}\hat{u}^*)_{\Omega^\varepsilon}-(q,\operatorname{div}z)_{\Omega^\varepsilon}. \tag*{(31)}

By Remark 5, (22), and ∥(I−Πε)q∥Ωε≤∥q∥ε=1\lVert(I-\Pi_\varepsilon)q\rVert_{\Omega^\varepsilon}\le\lVert q\rVert_\varepsilon=1,

∣III2∣≤ε2∥(I−Πε)q∥Ωε∥(wi)ε∇Iε0uˉi∥Ωε≤Cε2∣Iε0uˉ∣H1(Ω)≤Cε3/2∥uˉ∥H1(Ω).(32)|III_2|\le\varepsilon^2\lVert(I-\Pi_\varepsilon)q\rVert_{\Omega^\varepsilon}\lVert(w_i)_\varepsilon\nabla I^0_\varepsilon\bar{u}_i\rVert_{\Omega^\varepsilon}\le C\varepsilon^2|I^0_\varepsilon\bar{u}|_{H^1(\Omega)}\le C\varepsilon^{3/2}\lVert\bar{u}\rVert_{H^1(\Omega)}. \tag*{(32)}

Since div⁡u^∗=ε2(wi)ε⋅∇Iε0uˉi\operatorname{div}\hat{u}^*=\varepsilon^2(w_i)_\varepsilon\cdot\nabla I^0_\varepsilon\bar{u}_i vanishes in the solid and outside Ωεin\Omega^{\mathrm{in}}_\varepsilon, the identity (27) with h=Iε0uˉih=I^0_\varepsilon\bar{u}_i on every interior cell gives

III1=∑T∈TεinqTε2∫T(wi)ε⋅∇Iε0uˉi=ε2∣F∣∑T∈Tε∫Tdiv⁡(A∗Iε0uˉ)+∑T∈TεinqTε3∫∂T(ϕ(i))ε∇τIε0uˉi⋅n,(33)III_1=\sum_{T\in\mathcal{T}^{\mathrm{in}}_\varepsilon}q_T\varepsilon^2\int_T(w_i)_\varepsilon\cdot\nabla I^0_\varepsilon\bar{u}_i=\varepsilon^2|\mathcal{F}|\sum_{T\in\mathcal{T}_\varepsilon}\int_T\operatorname{div}(A^*I^0_\varepsilon\bar{u})+\sum_{T\in\mathcal{T}^{\mathrm{in}}_\varepsilon}q_T\varepsilon^3\int_{\partial T}(\phi^{(i)})_\varepsilon\nabla_\tau I^0_\varepsilon\bar{u}_i\cdot n, \tag*{(33)}

where the sum in III1′III_1' could be extended to all cells since Iε0uˉ=0I^0_\varepsilon\bar{u}=0 on the layer cells.

The term III1′III'_1. Since div⁡(A∗uˉ)=0\operatorname{div}(A^*\bar u)=0 in Ω\Omega, term III1′III'_1 can be rewritten as

III1′=ε2∣F∣∑T∈TεqT∫Tdiv⁡(A∗(Iε0uˉ−uˉ))=ε2∣F∣∑T∈TεqT∫∂T∩Ωn⋅A∗(Iε0uˉ−uˉ).III'_1=\varepsilon^2|F|\sum_{T\in\mathcal{T}_\varepsilon}q_T\int_T\operatorname{div}\left(A^*\left(I_\varepsilon^0\bar u-\bar u\right)\right)=\varepsilon^2|F|\sum_{T\in\mathcal{T}_\varepsilon}q_T\int_{\partial T\cap\Omega}n\cdot A^*\left(I_\varepsilon^0\bar u-\bar u\right).

The portions of the boundaries ∂T\partial T lying on ∂Ω\partial\Omega could be excluded from the last sum since Iε0uˉ=0I_\varepsilon^0\bar u=0 and n⋅A∗uˉ=0n\cdot A^*\bar u=0 on ∂Ω\partial\Omega. We now rewrite the last sum over the boundaries of cells T∈TεT\in\mathcal{T}_\varepsilon as the sum over the faces F∈EεF\in\mathcal{E}^\varepsilon, the Cauchy–Schwarz inequality and (26) give

∣III′∣=ε2∣F∣∣∑F∈Eε∫F⟦Πεq⟧FnF⋅A∗(Iε0uˉ−uˉ)∣≤Cε2(ε∑F∈Eε∥⟦Πεq⟧F∥F2)1/2(ε−1∑F∈Eε∥uˉ−Iε0uˉ∥F2)1/2≤Cε3/2∥q∥ε∥uˉ∥H1(Ω)≤Cε3/2∥uˉ∥H1(Ω).(34)\begin{aligned} |III'|=\varepsilon^2|F|\left|\sum_{F\in\mathcal{E}^\varepsilon}\int_F\llbracket\Pi_\varepsilon q\rrbracket_F n_F\cdot A^*\left(I_\varepsilon^0\bar u-\bar u\right)\right| \\ &\le C\varepsilon^2\left(\varepsilon\sum_{F\in\mathcal{E}^\varepsilon}\|\llbracket\Pi_\varepsilon q\rrbracket_F\|_F^2\right)^{1/2}\left(\varepsilon^{-1}\sum_{F\in\mathcal{E}^\varepsilon}\|\bar u-I_\varepsilon^0\bar u\|_F^2\right)^{1/2}\le C\varepsilon^{3/2}\|q\|_\varepsilon\|\bar u\|_{H^1(\Omega)}\le C\varepsilon^{3/2}\|\bar u\|_{H^1(\Omega)}. \tag*{(34)} \end{aligned}

since ∥q∥ε≤∥(v,q)∥ε=1\|q\|_\varepsilon\le\|(v,q)\|_\varepsilon=1.

Choice of zz. We now choose zz so that to cancel exactly in (33) the term III1′′III''_1 introduced in (35), i.e. so that

III1′′=(q,div⁡z)Ωε.(35)III''_1=(q,\operatorname{div}z)_{\Omega_\varepsilon}. \tag*{(35)}

We rewrite III1′′III''_1 as a sum over faces. For F∈EinεF\in\mathcal{E}_{\mathrm{in}}^\varepsilon, F=T∩T′F=T\cap T' with T=YkεT=Y_k^\varepsilon, T′=YlεT'=Y_l^\varepsilon, set

JF=ε3∫F(ϕ(i))ε∇τIε0uˉi⋅nF.J_F=\varepsilon^3\int_F(\phi^{(i)})_\varepsilon\nabla_\tau I_\varepsilon^0\bar u_i\cdot n_F.

Note that JFJ_F is unambiguously defined since the tangential derivatives of Iε0uˉiI_\varepsilon^0\bar u_i on FF are the same from both sides TT and T′T'. On a face of an interior cell which is not in Einε\mathcal{E}_{\mathrm{in}}^\varepsilon, Iε0uˉiI_\varepsilon^0\bar u_i and its tangential derivatives vanish. Passing from the sum over the boundaries of the interior cells to the sum over their faces, and recalling that nFn_F is the outer normal of TT and the inner normal of T′T', we get

III1′′=∑F∈Einε(qT−qT′)JF.III''_1=\sum_{F\in\mathcal{E}_{\mathrm{in}}^\varepsilon}(q_T-q_{T'})J_F.

On the other hand, the fields ΦF\Phi_F defined in the proof of Lemma 7 by (17) satisfy (q,div⁡ΦF)Ωε=qT−qT′(q,\operatorname{div}\Phi_F)_{\Omega_\varepsilon}=q_T-q_{T'}. Hence the choice

z=∑F∈EinεJFΦF∈H01(Ωε)d(36)z=\sum_{F\in\mathcal{E}_{\mathrm{in}}^\varepsilon}J_F\Phi_F\in H_0^1(\Omega_\varepsilon)^d \tag*{(36)}

gives (37). To bound zz, note that ∥(ϕ(i))ε∥F≤Cε(d−1)/2\|(\phi^{(i)})_\varepsilon\|_F\le C\varepsilon^{(d-1)/2}, so that ∣JF∣≤Cε3+(d−1)/2∥∇τIε0uˉ∥F|J_F|\le C\varepsilon^{3+(d-1)/2}\|\nabla_\tau I_\varepsilon^0\bar u\|_F by the Cauchy–Schwarz inequality. Since ∣ΦF∣H1(FFε)≤Cε−d/2|\Phi_F|_{H^1(\mathcal{F}_F^\varepsilon)}\le C\varepsilon^{-d/2}, we get with (25):

∣z∣H1(Ωε)≤C(ε−d∑F∈EinεJF2)1/2≤Cε5/2(∑F∈Einε∥∇τIε0uˉ∥F2)1/2≤Cε3/2∥uˉ∥H1(Ω).(37)|z|_{H^1(\Omega_\varepsilon)}\le C\left(\varepsilon^{-d}\sum_{F\in\mathcal{E}_{\mathrm{in}}^\varepsilon}J_F^2\right)^{1/2}\le C\varepsilon^{5/2}\left(\sum_{F\in\mathcal{E}_{\mathrm{in}}^\varepsilon}\|\nabla_\tau I_\varepsilon^0\bar u\|_F^2\right)^{1/2}\le C\varepsilon^{3/2}\|\bar u\|_{H^1(\Omega)}. \tag*{(37)}

Conclusion. With zz chosen above, (37) is satisfied, so that (33) and (35) give −III=III1′+III2-III=III'_1+III_2. Thus, ∣III∣|III| is bounded by the sum of (34) and (36). Combining this with (32), (39), and inserting into (31) gives

∥(uε−u^∗+z,pε−pε,1)∥ε≤Cε3/2∥uˉ∥H1(Ω).\left\|\left(u^\varepsilon-\hat u^*+z,p^\varepsilon-p^{\varepsilon,1}\right)\right\|_\varepsilon\le C\varepsilon^{3/2}\|\bar u\|_{H^1(\Omega)}.

This entails

∥uε−u^∗∥H1(Ωε)≤Cε3/2∥uˉ∥H1(Ω)+∥z∥H1(Ωε)≤Cε3/2∥uˉ∥H1(Ω),\lVert u^\varepsilon-\hat{u}^*\rVert_{H^1(\Omega^\varepsilon)} \le C\varepsilon^{3/2}\lVert\bar{u}\rVert_{H^1(\Omega)}+\lVert z\rVert_{H^1(\Omega^\varepsilon)} \le C\varepsilon^{3/2}\lVert\bar{u}\rVert_{H^1(\Omega)},

thanks again to (37). Lemma 3 gives

∥uε−u^∗∥Ωε≤Cε5/2∥uˉ∥H1(Ω).\lVert u^\varepsilon-\hat{u}^*\rVert_{\Omega^\varepsilon} \le C\varepsilon^{5/2}\lVert\bar{u}\rVert_{H^1(\Omega)}.

Since pε−pε,1∈L2(Ωε)p^\varepsilon-p^{\varepsilon,1}\in L^2(\Omega^\varepsilon), Lemma 6 gives

∥pε−pε,1∥Ωε≤Cε−1∥pε−pε,1∥ε≤Cε∥uˉ∥H1(Ω).\lVert p^\varepsilon-p^{\varepsilon,1}\rVert_{\Omega^\varepsilon}\le C\varepsilon^{-1}\lVert p^\varepsilon-p^{\varepsilon,1}\rVert_\varepsilon\le C\sqrt{\varepsilon}\lVert\bar{u}\rVert_{H^1(\Omega)}.

Finally, by (28) and (20),

∥pε,1−p∗∥Ωε≤2∥ε(πi)εIε0uˉi∥Ωε≤Cε∥Iε0uˉ∥Ω≤Cε∥uˉ∥Ω.\lVert p^{\varepsilon,1}-p^*\rVert_{\Omega^\varepsilon}\le2\lVert\varepsilon(\pi_i)_\varepsilon I_\varepsilon^0\bar{u}_i\rVert_{\Omega^\varepsilon}\le C\varepsilon\lVert I_\varepsilon^0\bar{u}\rVert_\Omega\le C\varepsilon\lVert\bar{u}\rVert_\Omega.

This proves (12). □

Use of large language models. Claude (model Opus 5.5) was used to develop the central arguments in the proof of Theorem 1. The article grew out of a first draft from July 2024, and has greatly benefited from interactions with Claude AI in many aspects. The most outstanding contributions were:

From human part: The idea of using the finite element framework on the ε\varepsilon-grid and the mesh-dependent norm (17) on the pressure. As mentioned in Remark 4, this choice lead to technical complications and necessitated stronger regularity assumptions.

From AI part: Introducing the additional velocity correction zz, as in (36).

From human part again: The idea of introducing gradient-like terms in the pressure norm (a priori inconsistent with the L2L^2 framework for the pressure). The norm (17) becomes (13).

Claude was also used for several other refinements, proof checks, bibliographical search, and language editing. The author has checked all mathematical arguments and takes full responsibility for the content of the paper.

Acknowledgements

The author warmly thanks Profs. Grégoire Allaire and François Murat for their interest in the first version of this article and enlightening discussions. This work is part of the Franco-German project ”Numerical homogenization techniques for multiscale Stokes problems” (PHC PROCOPE). The exchanges with German partners, in particular Moritz Hauck, were very helpful.

References

  1. [1]G. Acosta, R. G. Durán, and M. A. Muschietti, Solutions of the divergence operator on John domains, Advances in Mathematics, 206 (2006), pp. 373–401.DOI
  2. [2]G. Allaire, Homogenization of the Stokes flow in a connected porous medium, Asymptotic Anal., 2 (1989), pp. 203–222.DOI
  3. [3]L. Balazi, Multi-scale Finite Element Method for incompressible flows in heterogeneous media: Implementation and Convergence analysis, PhD thesis, 2024.DOI
  4. [4]L. Balazi, G. Allaire, and P. Omnes, Sharp convergence rates for the homogenization of the Stokes equations in a perforated domain. working paper or preprint, May 2024.DOI
  5. [5]M. Birman and T. Suslina, Second order periodic differential operators. threshold properties and homogenization, St. Petersburg Mathematical Journal, 15 (2004), pp. 639–714.DOI
  6. [6]M. E. Bogovskii, Solution of the first boundary value problem for the equation of continuity of an incompressible medium, in Dokl. Akad. Nauk SSSR, vol. 248, 1979, pp. 1037–1040.
  7. [7]P. Clément, Approximation by finite element functions using local regularization, Revue Française d’Automatique, Informatique, Recherche Opérationnelle. Analyse Numérique, 9 (1975), pp. 77–84.DOI
  8. [8]H. C. Elman, D. J. Silvester, and A. J. Wathen, Finite Elements and Fast Iterative Solvers: with Applications in Incompressible Fluid Dynamics, Numerical Mathematics and Scientific Computation, Oxford University Press, Oxford, 2nd ed., 2014.
  9. [9]P. Grisvard, Elliptic problems in nonsmooth domains, SIAM, 2011.DOI
  10. [10]U. Hornung, Homogenization and Porous Media, Interdisciplinary Applied Mathematics, vol. 6, Springer, 1997.DOI
  11. [11]T. J. R. Hughes and L. P. Franca, A new finite element formulation for computational fluid dynamics: VII. The Stokes problem with various well-posed boundary conditions: symmetric formulations that converge for all velocity/pressure spaces, Computer Methods in Applied Mechanics and Engineering, 65 (1987), pp. 85–96.DOI
  12. [12]G. Jankowiak and A. Lozinski, Non-Conforming Multiscale Finite Element Method for Stokes Flows in Heterogeneous Media. Part II: error estimates for periodic microstructure, Discrete and Continuous Dynamical Systems-B, 29 (2024), pp. 2298–2332; also available as preprint arXiv:1802.04389 (2018).
  13. [13]V. Jikov, S. Kozlov, and O. Oleinik, Homogenization of differential operators and integral functionals, Springer-Verlag, 1994.DOI
  14. [14]N. Kechkar and D. Silvester, Analysis of locally stabilized mixed finite element methods for the Stokes problem, Mathematics of Computation, 58 (1992), pp. 1–10.DOI
  15. [15]E. Marušić-Paloka and A. Mikelić, An error estimate for correctors in the homogenization of the stokes and navier-stokes equations in a porous medium, Boll. Unione Mat. Ital., 7 (1996), pp. 661–671.
  16. [16]M. A. Pakhinin and T. A. Suslina, Homogenization of the elliptic Dirichlet problem: error estimates in the (L₂ → H¹)-norm, Funktsional. Anal. i Prilozhen., 46 (2012), pp. 92–96.
  17. [17]E. Sánchez-Palencia, Non-homogeneous media and vibration theory, vol. 127 of Lecture notes in physics, 1980.
  18. [18]Z. Shen, Periodic Homogenization of Elliptic Systems, vol. 269 of Operator Theory: Advances and Applications, Birkhäuser, Cham, 2018.DOI
  19. [19]———, Sharp convergence rates for Darcy’s law, Communications in Partial Differential Equations, 47 (2022), pp. 1098–1123.arxiv.org/abs/2011.14169
  20. [20]D. J. Silvester and N. Kechkar, Stabilised bilinear-constant velocity-pressure finite elements for the conjugate gradient solution of the Stokes problem, Computer Methods in Applied Mechanics and Engineering, 79 (1990), pp. 71–86.
  21. [21]R. Stenberg, Analysis of mixed finite element methods for the Stokes problem: a unified approach, Mathematics of Computation, 42 (1984), pp. 9–23.
  22. [22]T. A. Suslina, Homogenization of the Dirichlet problem for elliptic systems: L₂-operator error estimates, Mathematika, 59 (2013), pp. 463–476.
  23. [23]L. Tartar, Incompressible fluid flow in a porous medium-convergence of the homogenization process, Appendix of [17], (1980).
  24. [24]V. Zhikov, On operator estimates in homogenization theory, in Doklady Mathematics, vol. 72, Pleiades Publishing, 2005, pp. 535–538.
  25. [25]V. Zhikov and S. Pastukhova, On operator estimates for some problems in homogenization theory, Russian Journal of Mathematical Physics, 12 (2005), pp. 515–524.

Paper details

vlnc:W_nggac2n5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents