betaFeedback

WorksMathematics

A Geometry-Robust Nečas-Type Inequality for Stokes Preconditioning

Abstract

Pressure mass preconditioners for the incompressible Stokes equations can deteriorate on domains with narrow gaps. We develop a pressure preconditioner that uses the distance to the boundary to account for spatially varying geometric scales. Its application consists of a pressure mass solve and a weighted pressure Poisson solve, with the squared boundary distance as the diffusion coefficient. The theoretical basis is a distance-weighted pressure norm defined by splitting the pressure into an L2L^2 component and a component controlled by its weighted gradient. We prove that this norm is bounded by the negative norm of the pressure gradient on every bounded connected open set, with a constant depending only on the dimension. A reverse bound follows from the boundary Hardy inequality, giving norm equivalence uniformly over families with bounded Hardy constants. Two-dimensional finite element experiments on slender, perforated, and composite domains demonstrate stable performance across the tested geometric parameters and mesh refinements. The smallest eigenvalues of the preconditioned pressure Schur complement remain close to 0.540.54, and block preconditioning with algebraic multigrid auxiliary solves requires about 20–30 FGMRES iterations across the geometric sweeps.

Introduction

We study the efficient numerical solution of the steady Stokes equations on domains with challenging geometries. For unit viscosity and homogeneous Dirichlet velocity data, the equations take the form

−Δu+∇p=f,∇⋅u=0in Ω,u=0on ∂Ω,(1)-\Delta u + \nabla p = f,\qquad\nabla\cdot u = 0 \quad\text{in } \Omega,\qquad u = 0 \quad\text{on } \partial\Omega, \tag*{(1)}

where Ω⊂Rd\Omega\subset\mathbb{R}^d, d≥2d \ge2, is a bounded connected Lipschitz domain, uu is the velocity, and pp is the pressure, normalized to have zero mean. Mixed finite-element discretizations lead to saddle-point systems, for which effective pressure preconditioning is essential to efficient iterative solution [12, 5].

The viscous term provides direct control of ∥∇u∥L2(Ω)\|\nabla u\|_{L^2(\Omega)}, while the pressure acts as a Lagrange multiplier enforcing incompressibility and has no coercive term of its own. Pressure stability therefore depends on its coupling to the velocity through the divergence operator. The corresponding inf-sup condition requires that the L2L^2 norm of a mean-zero pressure be controlled by the negative norm of its gradient. Equivalently, one needs the classical Nečas inequality [24, 14, 6]:

∥q∥L2(Ω)≤CN(Ω)∥∇q∥−1,Ω,q∈L02(Ω),(2)\|q\|_{L^2(\Omega)} \le C_N(\Omega)\|\nabla q\|_{-1,\Omega},\qquad q\in L_0^2(\Omega), \tag*{(2)}

where L02(Ω)L_0^2(\Omega) is the space of square-integrable functions with zero mean, and the negative norm is dual to the gradient norm on H01(Ω)dH_0^1(\Omega)^d. This estimate also underpins the comparison of the pressure Schur complement with a pressure mass operator in block preconditioners [12, 19].

In the classical negative-norm theory, the Nečas inequality is also stated with a lower-order term,

∥q∥L2(Ω)≤C(Ω)(∥q∥H−1(Ω)+∥∇q∥−1,Ω),q∈L2(Ω).\lVert q\rVert_{L^2(\Omega)} \le C(\Omega)\left(\lVert q\rVert_{H^{-1}(\Omega)}+\lVert\nabla q\rVert_{-1,\Omega}\right), \qquad q\in L^2(\Omega).

On a bounded connected Lipschitz domain, a compactness argument removes the lower-order term after fixing the mean, giving (1.2); see [24, 14]. Bramble [6] gives a direct proof on Lipschitz domains and explains its equivalence to the Stokes inf-sup condition. Amrouche, Ciarlet and Mardare [3] place these estimates in a broader equivalence theorem linking the Lions lemma, the closed range of the gradient, solvability of the divergence equation, and a weak form of de Rham’s theorem.

The classical Nečas constant depends on the domain geometry and may become unbounded along degenerating families. This dependence has been studied extensively through the equivalent Ladyzhenskaya–Babuška–Brezzi (LBB) formulation. Chizhonkov and Olshanskii [8] obtained explicit, asymptotically accurate estimates on model domains such as strips and rings, and related the continuous constant to discrete stability. Dobrowolski [9] established aspect-ratio estimates on stretched domains; his subsequent work [10] used coarse pressure spaces to construct preconditioners with bounds independent of the aspect ratio. Our preprint [28] gives a sharp LBB estimate for tight domains with periodic pillars. These results quantify how geometric degeneration enters the classical pressure estimate and how its effect can be compensated.

To obtain uniform stability estimates, one can incorporate the geometric scale into the pressure norm itself. For slender domains of width ϵ\epsilon, Sande, Koch, Kuchta and Mardal [25] established a uniform equivalence between ∥∇q∥−1,Ω\lVert\nabla q\rVert_{-1,\Omega} and the KK-functional norm ∥q∥L2+ϵH1\lVert q\rVert_{L^2+\epsilon H^1}, using a coarse partition and quasi-interpolation. For periodic pillar arrays with hole size proportional to the period ϵ\epsilon, the sharp LBB estimate of [28] is complemented by the norm equivalence ∥∇q∥−1,Ω≍∥q∥L2+σϵH1\lVert\nabla q\rVert_{-1,\Omega}\asymp\lVert q\rVert_{L^2+\sigma_\epsilon H^1} in [29], where σϵ≍ϵ\sigma_\epsilon\asymp\epsilon in this regime. The latter proof combines a weighted Nečas estimate on a fixed filled domain with a dual pressure extension built from the restriction operators of [1, 2, 21]; its constants are uniform in the period for a fixed macroscopic domain and reference geometry. In Nečas language, these norm characterizations provide pressure estimates adapted to the scale of the specified geometric family.

Spatially varying geometric scales are common in practical flow domains, including branching vascular geometries [13], microfluidic devices employing deterministic lateral displacement (DLD) arrays [20, 26], and the pore spaces of natural rocks [23]. Sande et al. [25], Remark 3 and Example 5 also discuss a variable local width and the associated weighted pressure operator. The analytical question pursued here is to identify a canonical local scale and prove the corresponding Nečas-type inequality on arbitrary bounded connected open sets, with a constant depending only on the dimension. The boundary distance supplies such a scale intrinsically, without a prescribed channel width, period, or coarse partition.

Let d≥2d\ge2, let Ω⊂Rd\Omega\subset\mathbb{R}^d be a bounded connected open set, and set δ(x)=dist⁡(x,∂Ω)\delta(x)=\operatorname{dist}(x,\partial\Omega). Let Hδ(Ω)\mathcal{H}_\delta(\Omega) consist of the functions ϕ∈L2(Ω)∩Hloc1(Ω)\phi\in L^2(\Omega)\cap H^1_{\mathrm{loc}}(\Omega) with δ∇ϕ∈L2(Ω)d\delta\nabla\phi\in L^2(\Omega)^d. Define the distance-weighted KK-functional norm by

∥q∥Kδ:=inf⁡ϕ∈Hδ(Ω)(∥q−ϕ∥L2(Ω)2+∥δ∇ϕ∥L2(Ω)2)1/2.(3)\lVert q\rVert_{K_\delta}:=\inf_{\phi\in\mathcal{H}_\delta(\Omega)}\left(\lVert q-\phi\rVert_{L^2(\Omega)}^2+\lVert\delta\nabla\phi\rVert_{L^2(\Omega)}^2\right)^{1/2}. \tag*{(3)}

Our main result, Theorem 3.1, is the inequality

∥q∥Kδ≤Cd∥∇q∥−1,Ω,q∈L02(Ω),(4)\lVert q\rVert_{K_{\delta}} \le C_d \lVert\nabla q\rVert_{-1,\Omega}, \qquad q \in L^2_0(\Omega), \tag*{(4)}

with CdC_d depending only on the dimension. The geometric information is carried by δ\delta in the norm, while the constant is independent of aspect ratios, perforations, and contrasts between local length scales. Since ∥q∥Kδ≤∥q∥L2(Ω)\lVert q\rVert_{K_{\delta}} \le\lVert q\rVert_{L^2(\Omega)}, it controls pressure in a weaker, geometrically adapted norm.

The proof starts from the classical Nečas inequality on a cube, which controls a function up to its mean with a constant independent of the cube size. A Whitney decomposition allows us to apply this estimate on slightly enlarged cubes whose sizes are comparable to the local boundary distance. Each cube has its own mean, and these local constants need not agree. We blend them into a smooth function ϕ\phi using a partition of unity and use ϕ\phi as a competitor in (3). Both the error q−ϕq-\phi and the weighted gradient δ∇ϕ\delta\nabla\phi are controlled by the local oscillations: the factor δ\delta compensates for the derivatives of the partition functions. Bounded overlap and negative-norm localization then give (4).

If Ω\Omega also satisfies the boundary Hardy inequality ∥v/δ∥L2(Ω)≤CH∥∇v∥L2(Ω)\lVert v/\delta\rVert_{L^2(\Omega)} \le C_H\lVert\nabla v\rVert_{L^2(\Omega)} for v∈H01(Ω)v \in H^1_0(\Omega), then integration by parts gives a reverse estimate. Combining this estimate with (4) yields the norm equivalence in Corollary 3.7:

Cd−1∥q∥Kδ≤∥∇q∥−1,Ω≤1+CH2∥q∥Kδ,q∈L02(Ω).(5)C_d^{-1}\lVert q\rVert_{K_{\delta}} \le\lVert\nabla q\rVert_{-1,\Omega} \le\sqrt{1+C_H^2}\lVert q\rVert_{K_{\delta}}, \qquad q \in L^2_0(\Omega). \tag*{(5)}

The equivalence is uniform on any family with uniformly bounded Hardy constants. For example, Ancona’s bound CH≤4C_H \le4 for simply connected planar domains [4, 17] gives the reverse constant 17\sqrt{17}, independently of channel widths and their contrast; see Remark 3.8. Thus the additional geometric dependence in the two-sided norm characterization is isolated in the Hardy constant.

For the Stokes problem, this norm equivalence provides an analytical basis for pressure preconditioning through the operator-preconditioning framework of Mardal–Winther [22], building on Bramble–Pasciak [7] and Loghin–Wathen [19]. The squared negative norm of the pressure gradient is the quadratic form of the continuous pressure Schur complement, so an equivalent pressure norm supplies a corresponding Riesz map. The distance-weighted KK-functional motivates an inverse pressure block combining a mass inverse and a boundary-distance-weighted stiffness inverse, with no prescribed global length scale. Transferring uniform continuous bounds to a discretization requires suitable uniform discrete stability and comparison of the discrete weighted norm with KδK_{\delta}. Numerical results on slender, perforated, and composite domains show that the computed smallest eigenvalues of the preconditioned pressure Schur complement remain uniformly bounded away from zero across the tested geometric variations, consistent with the geometry-independent lower bound in the continuous estimate.

The remainder of the paper is organized as follows. Section 2 introduces the negative norm of the gradient and the distance-weighted KK-functional norm. Section 3 proves the geometry-robust weighted Nečas inequality through local estimates on Whitney cubes and a smooth partition of unity. Section 3.4 then uses the boundary Hardy inequality to derive the norm equivalence as a corollary. Section 4 develops the Stokes preconditioning application, and Section 5 reports numerical experiments. Section 6 concludes.

Mathematical Setting

Let Ω⊂Rd\Omega\subset\mathbb{R}^d, d≥2d \ge2, be a bounded connected open set. Write L02(Ω)={q∈L2(Ω):∫Ωq dx=0}L^2_0(\Omega)=\{q \in L^2(\Omega): \int_{\Omega} q\,\mathrm{d}x=0\}. For q∈L2(Ω)q \in L^2(\Omega), define the negative norm of its gradient by

∥∇q∥−1,Ω:=sup⁡v∈H01(Ω)d∖{0}∣⟨∇q,v⟩∣∥∇v∥L2(Ω),(6)\lVert\nabla q\rVert_{-1,\Omega} := \sup_{v\in H_0^1(\Omega)^d\setminus\{0\}} \frac{\lvert\langle\nabla q,v\rangle\rvert}{\lVert\nabla v\rVert_{L^2(\Omega)}}, \tag*{(6)}

where ⟨∇q,v⟩=−∫Ωq ∇⋅v dx\langle\nabla q,v\rangle=-\int_\Omega q\,\nabla\cdot v\,dx.

To define the geometrically adapted norm, let

δ(x):=dist⁡(x,∂Ω),x∈Ω.(7)\delta(x):=\operatorname{dist}(x,\partial\Omega),\qquad x\in\Omega. \tag*{(7)}

The natural weighted energy space is

Hδ(Ω):={ϕ∈L2(Ω): ϕ∈Hloc1(Ω), δ∇ϕ∈L2(Ω)d}.(8)\mathcal{H}_{\delta}(\Omega):=\left\{\phi\in L^2(\Omega):\ \phi\in H^1_{\mathrm{loc}}(\Omega),\ \delta\nabla\phi\in L^2(\Omega)^d\right\}. \tag*{(8)}

We define the distance-weighted KK-functional norm by

∥q∥Kδ:=inf⁡ϕ∈Hδ(Ω)(∥q−ϕ∥L2(Ω)2+∥δ∇ϕ∥L2(Ω)2)1/2,q∈L2(Ω).(9)\lVert q\rVert_{K_\delta}:=\inf_{\phi\in\mathcal{H}_{\delta}(\Omega)}\left(\lVert q-\phi\rVert_{L^2(\Omega)}^2+\lVert\delta\nabla\phi\rVert_{L^2(\Omega)}^2\right)^{1/2},\qquad q\in L^2(\Omega). \tag*{(9)}

The functional is unchanged when a constant is added to qq: the change of variables ϕ↦ϕ+c\phi\mapsto\phi+c shows ∥q+c∥Kδ=∥q∥Kδ\lVert q+c\rVert_{K_\delta}=\lVert q\rVert_{K_\delta}. Thus it is naturally a norm on L2(Ω)/RL^2(\Omega)/\mathbb{R}. For completeness, if ∥q∥Kδ=0\lVert q\rVert_{K_\delta}=0, there are ϕn∈Hδ(Ω)\phi_n\in\mathcal{H}_{\delta}(\Omega) such that ϕn→q\phi_n\to q in L2L^2 and δ∇ϕn→0\delta\nabla\phi_n\to0 in L2L^2. On every U⋐ΩU\Subset\Omega, δ\delta is bounded below, and hence ∇ϕn→0\nabla\phi_n\to0 in L2(U)L^2(U). It follows distributionally that ∇q=0\nabla q=0 in Ω\Omega; connectedness implies that qq is constant. Therefore (2.4) is a genuine norm on L02(Ω)L_0^2(\Omega).

A Geometry-Robust Weighted Nečas Inequality

We prove the following distance-weighted inequality.

Theorem 3.1 (Geometry-robust distance-weighted Nečas-type inequality). Let d≥2d\ge2, let Ω⊂Rd\Omega\subset\mathbb{R}^d be a bounded connected open set, and set δ(x)=dist⁡(x,∂Ω)\delta(x)=\operatorname{dist}(x,\partial\Omega). Then

∥q∥Kδ≤Cd∥∇q∥−1,Ω,q∈L02(Ω).(10)\lVert q\rVert_{K_\delta}\le C_d\lVert\nabla q\rVert_{-1,\Omega},\qquad q\in L_0^2(\Omega). \tag*{(10)}

The constant Cd>0C_d>0 depends only on dd, independently of the domain.

The proof combines local Nečas estimates on Whitney cubes with a smooth partition of unity. We first establish the covering and local estimates, then construct a competitor for the infimum in (2.4) to obtain (3.1). All constants in the proof depend only on dd.

Whitney covering and partition of unity

Lemma 3.2 (Whitney covering and partition of unity). There is a countable family W\mathcal{W} of closed cubes with disjoint interiors whose union is Ω\Omega. Write ℓQ\ell_Q for the side length of Q∈WQ\in\mathcal{W} and set UQ=int⁡(98Q)U_Q=\operatorname{int}(\frac{9}{8}Q), where the dilation is concentric. There are positive constants cdc_d, CdC_d, NdN_d, AdA_d, BdB_d, depending only on dd, such that the following properties hold. The sets UQ⋐ΩU_Q\Subset\Omega form a locally finite open cover, and

cdℓQ≤δ(x)≤CdℓQ(x∈UQ),∑Q∈WχUQ≤Nd.(11)c_d\ell_Q\le\delta(x)\le C_d\ell_Q\quad(x\in U_Q),\qquad\sum_{Q\in\mathcal{W}}\chi_{U_Q}\le N_d. \tag*{(11)}

There are functions ψQ∈Cc∞(UQ)\psi_Q\in C^\infty_c(U_Q) such that

0≤ψQ≤1,∑Q∈WψQ=1 in Ω,∣∇ψQ∣≤AdℓQ−1.(12)0\le\psi_Q\le1,\qquad\sum_{Q\in\mathcal{W}}\psi_Q=1\ \text{in }\Omega,\qquad\lvert\nabla\psi_Q\rvert\le A_d\ell_Q^{-1}. \tag*{(12)}

In particular, δ∣∇ψQ∣≤Bd\delta\lvert\nabla\psi_Q\rvert\le B_d in $\Omega.

Proof. The Whitney covering and the bounds in (11)–(12) are standard; see [27 Chapter VI, Section 1], in particular Section 1.3 for the partition of unity. Combining δ≤CdℓQ\delta\le C_d\ell_Q on UQU_Q with ∣∇ψQ∣≤AdℓQ−1|\nabla\psi_Q| \le A_d\ell_Q^{-1} gives δ∣∇ψQ∣≤Bd:=CdAd\delta|\nabla\psi_Q| \le B_d := C_dA_d there; outside UQU_Q, ∇ψQ=0\nabla\psi_Q=0.

Local Nečas estimates and negative-norm localization

For an open set U⊂ΩU\subset\Omega, let ∥∇q∥−1,U\|\nabla q\|_{-1,U} denote the negative norm in (2.1) with Ω\Omega replaced by UU.

Lemma 3.3 (Local Nečas estimate on a cube). For every open cube U⊂RdU\subset\mathbb{R}^d and every r∈L2(U)r\in L^2(U),

∥r−rU∥L2(U)≤Cd∥∇r∥−1,U,rU:=1∣U∣∫Ur dx.(13)\|r-r_U\|_{L^2(U)}\le C_d\|\nabla r\|_{-1,U},\qquad r_U:=\frac{1}{|U|}\int_U r\,\mathrm{d}x. \tag*{(13)}

The constant is independent of the position and side length of UU.

This is the classical Nečas inequality [24], scaled from a reference cube.

Lemma 3.4 (Negative-norm localization). Let U\mathcal{U} be a finite or countable family of open subsets of Ω\Omega with ∑U∈UχU≤M\sum_{U\in\mathcal{U}}\chi_U\le M almost everywhere. Then, for every q∈L2(Ω)q\in L^2(\Omega),

∑U∈U∥∇q∥−1,U2≤M∥∇q∥−1,Ω2.(14)\sum_{U\in\mathcal{U}}\|\nabla q\|_{-1,U}^2\le M\|\nabla q\|_{-1,\Omega}^2. \tag*{(14)}

Proof. Fix a finite subfamily and 0<η<10<\eta<1, and write aU=∥∇q∥−1,Ua_U=\|\nabla q\|_{-1,U}. Omitting terms with aU=0a_U=0, choose vU∈H01(U)dv_U\in H_0^1(U)^d with ∥∇vU∥L2(U)=1\|\nabla v_U\|_{L^2(U)}=1 and ∫Uq∇⋅vU dx≥(1−η)aU\int_U q\nabla\cdot v_U\,\mathrm{d}x\ge(1-\eta)a_U. Since U⊂ΩU\subset\Omega, the zero extension of each vUv_U lies in H01(Ω)dH_0^1(\Omega)^d, and its weak gradient is the zero extension of ∇vU\nabla v_U. Using the same notation for these extensions, set v=∑UaUvUv=\sum_U a_Uv_U. The overlap bound gives

∥∇v∥L2(Ω)2≤M∑UaU2,∫Ωq∇⋅v dx≥(1−η)∑UaU2.\|\nabla v\|_{L^2(\Omega)}^2\le M\sum_Ua_U^2,\qquad\int_\Omega q\nabla\cdot v\,\mathrm{d}x\ge(1-\eta)\sum_Ua_U^2.

The definition of the global negative norm implies

(1−η)∑UaU2≤∥∇q∥−1,Ω(M∑UaU2)1/2.(1-\eta)\sum_Ua_U^2\le\|\nabla q\|_{-1,\Omega}\left(M\sum_Ua_U^2\right)^{1/2}.

Letting η↓0\eta\downarrow0 proves the estimate for every finite subfamily. Increasing finite subfamilies exhaust the countable case.

Proof of the main theorem

Proof of Theorem 3.1. Take the covering and partition of unity from Lemma 3.2. For Q∈WQ\in\mathcal{W}, let mQ=∣UQ∣−1∫UQq dxm_Q=|U_Q|^{-1}\int_{U_Q}q\,\mathrm{d}x. Lemmas 3.3 and 3.4 give

∑Q∥q−mQ∥L2(UQ)2≤Cd∑Q∥∇q∥−1,UQ2≤Cd∥∇q∥−1,Ω2.(15)\sum_Q\|q-m_Q\|_{L^2(U_Q)}^2\le C_d\sum_Q\|\nabla q\|_{-1,U_Q}^2\le C_d\|\nabla q\|_{-1,\Omega}^2. \tag*{(15)}

Define

ϕ:=∑Q∈WmQψQ.(16)\phi:=\sum_{Q\in\mathcal{W}}m_Q\psi_Q. \tag*{(16)}

Local finiteness implies ϕ∈C∞(Ω)\phi\in C^\infty(\Omega). Since ∑QψQ=1\sum_Q \psi_Q = 1 and ∑Q∇ψQ=0\sum_Q \nabla\psi_Q = 0, almost everywhere in Ω\Omega we have

q−ϕ=∑QψQ(q−mQ),∇ϕ=∑Q∇ψQ(mQ−q).(17)q-\phi=\sum_Q\psi_Q(q-m_Q),\qquad\nabla\phi=\sum_Q\nabla\psi_Q(m_Q-q). \tag*{(17)}

The second identity uses only cancellation of the partition gradients and does not differentiate qq. Jensen’s inequality gives

∣q−ϕ∣2≤∑QψQ∣q−mQ∣2≤∑QχUQ∣q−mQ∣2.|q-\phi|^2\leq\sum_Q\psi_Q|q-m_Q|^2\leq\sum_Q\chi_{U_Q}|q-m_Q|^2.

At each point at most NdN_d summands are nonzero. The bound δ∣∇ψQ∣≤Bd\delta|\nabla\psi_Q|\leq B_d and the Cauchy–Schwarz inequality yield

∣δ∇ϕ∣2≤NdBd2∑QχUQ∣q−mQ∣2.|\delta\nabla\phi|^2\leq N_dB_d^2\sum_Q\chi_{U_Q}|q-m_Q|^2.

Integrating these two inequalities, using Tonelli’s theorem and (3.6), gives

∥q−ϕ∥L2(Ω)2+∥δ∇ϕ∥L2(Ω)2≤Cd∑Q∥q−mQ∥L2(UQ)2≤Cd∥∇q∥−1,Ω2.(18)\|q-\phi\|_{L^2(\Omega)}^2+\|\delta\nabla\phi\|_{L^2(\Omega)}^2\leq C_d\sum_Q\|q-m_Q\|_{L^2(U_Q)}^2\leq C_d\|\nabla q\|_{-1,\Omega}^2. \tag*{(18)}

In particular, ϕ∈L2(Ω)\phi\in L^2(\Omega) and δ∇ϕ∈L2(Ω)d\delta\nabla\phi\in L^2(\Omega)^d, so ϕ∈Hδ(Ω)\phi\in\mathcal H_\delta(\Omega) is an admissible competitor for the infimum in (2.4). Taking square roots in (3.9) proves (3.1).

The theorem requires neither boundary regularity nor a Hardy inequality. The norm equivalence follows by combining it with the Hardy reverse estimate below.

Norm equivalence under the Hardy inequality

The main theorem supplies one side of the equivalence between the distance-weighted norm and the negative norm of the gradient on every bounded connected open set. To obtain the reverse estimate, we now introduce the boundary Hardy inequality, which holds in particular on bounded Lipschitz domains.

Lemma 3.5 (Boundary Hardy inequality on Lipschitz domains). Let Ω⊂Rd\Omega\subset\mathbb R^d be a bounded Lipschitz domain. There exists a constant CH>0C_H>0, depending only on dd and on the Lipschitz character of Ω\Omega, such that

∥uδ∥L2(Ω)≤CH∥∇u∥L2(Ω),∀u∈H01(Ω).(19)\left\|\frac{u}{\delta}\right\|_{L^2(\Omega)}\leq C_H\|\nabla u\|_{L^2(\Omega)},\qquad\forall u\in H_0^1(\Omega). \tag*{(19)}

This is the case p=2p=2, β=0\beta=0 of the classical weighted Hardy inequality recalled in [18 Theorem 1.1], extended from Cc∞(Ω)C_c^\infty(\Omega) to H01(Ω)H_0^1(\Omega) by density; see also [15, 11].

Lemma 3.6 (Hardy reverse estimate). Let Ω⊂Rd\Omega\subset\mathbb R^d be a bounded connected open set satisfying (3.10) with constant CHC_H. Then

∥∇q∥−1,Ω≤1+CH2 ∥q∥Kδ,q∈L2(Ω).(20)\|\nabla q\|_{-1,\Omega}\leq\sqrt{1+C_H^2}\,\|q\|_{K_\delta},\qquad q\in L^2(\Omega). \tag*{(20)}

Proof. Let q∈L2(Ω)q\in L^2(\Omega), fix ϕ∈Hδ(Ω)\phi\in\mathcal H_\delta(\Omega), and first take v∈Cc∞(Ω)dv\in C_c^\infty(\Omega)^d. Since ϕ∈Hloc1(Ω)\phi\in H^1_{\mathrm{loc}}(\Omega), integration by parts gives

⟨∇q,v⟩=−∫Ω(q−ϕ)∇⋅v dx+∫Ω∇ϕ⋅v dx.\langle\nabla q,v\rangle=-\int_\Omega(q-\phi)\nabla\cdot v\,dx+\int_\Omega\nabla\phi\cdot v\,dx.

Hence

∣⟨∇q,v⟩∣≤∥q−ϕ∥L2(Ω)∥∇⋅v∥L2(Ω)+∥δ∇ϕ∥L2(Ω)∥vδ∥L2(Ω).|\langle\nabla q,v\rangle| \le\|q-\phi\|_{L^2(\Omega)}\|\nabla\cdot v\|_{L^2(\Omega)}+\|\delta\nabla\phi\|_{L^2(\Omega)}\left\|\frac{v}{\delta}\right\|_{L^2(\Omega)}.

For v∈H01(Ω)dv\in H^1_0(\Omega)^d, the div–curl identity reads

∥∇v∥L2(Ω)2=∥∇⋅v∥L2(Ω)2+12∑i,j=1d∥∂jvi−∂ivj∥L2(Ω)2.\|\nabla v\|_{L^2(\Omega)}^2=\|\nabla\cdot v\|_{L^2(\Omega)}^2+\frac{1}{2}\sum_{i,j=1}^{d}\|\partial_jv_i-\partial_iv_j\|_{L^2(\Omega)}^2.

In particular, ∥∇⋅v∥L2(Ω)≤∥∇v∥L2(Ω)\|\nabla\cdot v\|_{L^2(\Omega)}\le\|\nabla v\|_{L^2(\Omega)}. Using this estimate and the boundary Hardy inequality (19) componentwise,

∥vδ∥L2(Ω)≤CH∥∇v∥L2(Ω),\left\|\frac{v}{\delta}\right\|_{L^2(\Omega)}\le C_H\|\nabla v\|_{L^2(\Omega)},

we obtain

∣⟨∇q,v⟩∣≤(∥q−ϕ∥L2(Ω)+CH∥δ∇ϕ∥L2(Ω))∥∇v∥L2(Ω).|\langle\nabla q,v\rangle|\le\left(\|q-\phi\|_{L^2(\Omega)}+C_H\|\delta\nabla\phi\|_{L^2(\Omega)}\right)\|\nabla v\|_{L^2(\Omega)}.

By density of Cc∞(Ω)dC^\infty_c(\Omega)^d in H01(Ω)dH^1_0(\Omega)^d and continuity of the gradient pairing, the same bound holds for all v∈H01(Ω)dv\in H^1_0(\Omega)^d. Dividing by ∥∇v∥L2(Ω)\|\nabla v\|_{L^2(\Omega)} and taking the supremum over v≠0v\ne0 gives a bound for ∥∇q∥−1,Ω\|\nabla q\|_{-1,\Omega}. The Cauchy–Schwarz inequality in R2\mathbb{R}^2, followed by the infimum over ϕ∈Hδ(Ω)\phi\in\mathcal{H}_\delta(\Omega), yields

∥∇q∥−1,Ω≤1+CH2∥q∥Kδ.\|\nabla q\|_{-1,\Omega}\le\sqrt{1+C_H^2}\|q\|_{\mathcal{K}_\delta}.

This proves (20). □\square

Corollary 3.7 (Distance-weighted norm equivalence). Under the assumptions of Lemma 3.6,

cd∥q∥Kδ≤∥∇q∥−1,Ω≤1+CH2∥q∥Kδ,q∈L02(Ω),(21)c_d\|q\|_{\mathcal{K}_\delta}\le\|\nabla q\|_{-1,\Omega}\le\sqrt{1+C_H^2}\|q\|_{\mathcal{K}_\delta},\qquad q\in L^2_0(\Omega), \tag*{(21)}

where cd=Cd−1>0c_d=C_d^{-1}>0 depends only on dd. In particular, the equivalence holds on bounded connected Lipschitz domains, with the upper constant controlled by their Hardy constants.

Proof. Combine Theorem 3.1 with Lemma 3.6. □\square

Remark 3.8 (Uniform Hardy constants). The Hardy constant admits geometry-independent bounds in several important classes. For bounded convex domains in any dimension, one may take CH≤2C_H\le2 [17]. For bounded simply connected planar domains, Ancona’s result gives CH≤4C_H\le4 [4]; see also Laptev–Sobolev [17], who obtain improved bounds for geometrically restricted classes of simply connected planar domains. The bound CH≤4C_H\le4 is unaffected by variations in channel width, bends, or constrictions, provided the planar domain remains simply connected.

For sufficient conditions ensuring uniform Hardy constants in multiply connected domains and in higher dimensions, see Korte–Lehrbäck–Tuominen [16].

Application to Stokes Preconditioning

The Stokes problem and pressure Schur complement

We now apply the norm equivalence to preconditioning the steady incompressible Stokes problem on a bounded connected Lipschitz domain:

{−μΔu+∇p=fin Ω,∇⋅u=0in Ω,u=uDon ∂Ω.(22)\begin{cases} -\mu\Delta\mathbf{u}+\nabla p=\mathbf{f} & \text{in }\Omega,\\ \nabla\cdot\mathbf{u}=0 & \text{in }\Omega,\\ \mathbf{u}=\mathbf{u}_{D} & \text{on }\partial\Omega. \tag*{(22)} \end{cases}

Here μ>0\mu>0 is the viscosity, f\mathbf{f} is the body force, and uD\mathbf{u}_{D} is prescribed velocity data with zero net boundary flux. We set μ=1\mu=1 in all experiments. After lifting the boundary data, the variational formulation uses the velocity space V(Ω)=H01(Ω)dV(\Omega)=H_{0}^{1}(\Omega)^{d} with norm ∥v∥V=∥∇v∥L2(Ω)\lVert v\rVert_{V}=\lVert\nabla v\rVert_{L^{2}(\Omega)} and the pressure space Q(Ω)=L02(Ω)Q(\Omega)=L_{0}^{2}(\Omega). The classical Nečas inequality (1.2) is equivalent to the Stokes inf-sup condition, with CN(Ω)=β(Ω)−1C_{N}(\Omega)=\beta(\Omega)^{-1} for the optimal Nečas constant CN(Ω)C_{N}(\Omega) and the Stokes inf-sup constant β(Ω)\beta(\Omega); see [6].

Let Vh⊂V(Ω)V_{h}\subset V(\Omega) and Qh,0⊂Q(Ω)Q_{h,0}\subset Q(\Omega) be an inf-sup stable conforming finite-element pair. For unit viscosity, let AhA_{h} be the matrix of (∇uh,∇vh)(\nabla u_{h},\nabla v_{h}) and BhB_{h} the matrix of −(qh,∇⋅vh)-(q_{h},\nabla\cdot v_{h}). After lifting the boundary data, the discrete Stokes problem has the form

Ah[up]=[fhgh],Ah=[AhBhTBh0].(23)A_{h} \begin{bmatrix} \mathbf{u}\\ \mathbf{p} \end{bmatrix} = \begin{bmatrix} \mathbf{f}_{h}\\ \mathbf{g}_{h} \end{bmatrix}, \qquad A_{h}= \begin{bmatrix} A_{h} & B_{h}^{\mathsf{T}}\\ B_{h} & 0 \end{bmatrix}. \tag*{(23)}

Eliminating the velocity gives the pressure Schur complement Sh=BhAh−1BhTS_{h}=B_{h}A_{h}^{-1}B_{h}^{\mathsf{T}}. Its quadratic form is

qTShq=sup⁡vh∈Vh∖{0}(∫Ωqh∇⋅vh dx)2∥∇vh∥L2(Ω)2,qh∈Qh,0,(24)\mathbf{q}^{\mathsf{T}}S_{h}\mathbf{q} = \sup_{v_{h}\in V_{h}\setminus\{0\}} \frac{\left(\int_{\Omega}q_{h}\nabla\cdot v_{h}\,\mathrm{d}x\right)^{2}} {\lVert\nabla v_{h}\rVert_{L^{2}(\Omega)}^{2}}, \qquad q_{h}\in Q_{h,0}, \tag*{(24)}

where q\mathbf{q} is the coefficient vector of qhq_{h}. Thus a block solver needs an approximation to Sh−1S_{h}^{-1} as well as a velocity solver. Forming ShS_{h} is generally expensive because it contains Ah−1A_{h}^{-1}.

The corresponding continuous Schur complement SS has quadratic form ⟨Sq,q⟩=∥∇q∥−1,Ω2\langle Sq,q\rangle=\lVert\nabla q\rVert_{-1,\Omega}^{2}, by the definition in (2.1). A pressure mass inverse uses the L2L^{2} norm to approximate this form. The classical estimates give

β(Ω)2∥q∥L2(Ω)2≤⟨Sq,q⟩≤∥q∥L2(Ω)2,\beta(\Omega)^{2}\lVert q\rVert_{L^{2}(\Omega)}^{2} \leq \langle Sq,q\rangle \leq \lVert q\rVert_{L^{2}(\Omega)}^{2},

so the mass-based bound can deteriorate as β(Ω)\beta(\Omega) decreases. For the weighted pressure norm, Corollary 3.7 gives

cd2∥q∥Kδ2≤⟨Sq,q⟩≤(1+CH2)∥q∥Kδ2.(25)c_{d}^{2}\lVert q\rVert_{\mathcal{K}_{\delta}}^{2} \leq \langle Sq,q\rangle \leq (1+C_{H}^{2})\lVert q\rVert_{\mathcal{K}_{\delta}}^{2}. \tag*{(25)}

This identifies the distance-weighted norm as a pressure metric whose equivalence constants are controlled by the dimension and the Hardy constant. Approximating the Schur complement by the operator associated with this norm follows the operator-preconditioning framework of [22, 19].

From the weighted norm to a pressure preconditioner

Assume Qh,0=Qh∩L02(Ω)Q_{h,0}=Q_{h}\cap L_{0}^{2}(\Omega) for a pressure space Qh⊂H1(Ω)Q_{h}\subset H^{1}(\Omega) containing constants, and let δh\delta_{h} approximate the boundary distance. Since Ω\Omega is bounded, Qh⊂Hδ(Ω)Q_{h}\subset\mathcal{H}_{\delta}(\Omega). In a basis of the mean-zero pressure space, let MpM_{p} and Lp,δhL_{p,\delta_{h}} be the matrices of

m(qh,rh)=∫Ωqhrh dx,ℓδh(qh,rh)=∫Ωδh2∇qh⋅∇rh dx.m(q_{h},r_{h})=\int_{\Omega}q_{h}r_{h}\,\mathrm{d}x, \qquad \ell_{\delta_{h}}(q_{h},r_{h})=\int_{\Omega}\delta_{h}^{2}\nabla q_{h}\cdot\nabla r_{h}\,\mathrm{d}x.

All pressure matrices and inverses in this section are restricted to the mean-zero space; the weighted stiffness matrix is assumed positive definite there. The stiffness form imposes natural Neumann conditions on the auxiliary pressure problem.

Restricting the infimum in (9) to Qh,0Q_{h,0} and using the discrete distance δh\delta_h gives

∥qh∥Kδh,h2:=min⁡ϕh∈Qh,0(∥qh−ϕh∥L2(Ω)2+∥δh∇ϕh∥L2(Ω)2).(26)\|q_h\|^2_{K_{\delta_h,h}} := \min_{\phi_h\in Q_{h,0}} \left(\|q_h-\phi_h\|^2_{L^2(\Omega)}+\|\delta_h\nabla\phi_h\|^2_{L^2(\Omega)}\right). \tag*{(26)}

Restricting ϕh\phi_h to have zero mean does not change this minimum: removing its mean reduces the first term and leaves the second term unchanged. In coefficient vectors, (26) becomes

min⁡ϕ((q−ϕ)TMp(q−ϕ)+ϕTLp,δhϕ).\min_{\phi}\left((\mathbf{q}-\boldsymbol{\phi})^{\mathsf{T}}M_p(\mathbf{q}-\boldsymbol{\phi})+\boldsymbol{\phi}^{\mathsf{T}}L_{p,\delta_h}\boldsymbol{\phi}\right).

The minimizing vector satisfies Mp(q−ϕ)=Lp,δhϕ=:tM_p(\mathbf{q}-\boldsymbol{\phi})=L_{p,\delta_h}\boldsymbol{\phi}=:\mathbf{t}. Hence q=(Mp−1+Lp,δh−1)t\mathbf{q}=(M_p^{-1}+L_{p,\delta_h}^{-1})\mathbf{t}, and the minimum is

∥qh∥Kδh,h2=qTRδ,hq,Rδ,h=(Mp−1+Lp,δh−1)−1.(27)\|q_h\|^2_{K_{\delta_h,h}}=\mathbf{q}^{\mathsf{T}}R_{\delta,h}\mathbf{q},\qquad R_{\delta,h}=(M_p^{-1}+L_{p,\delta_h}^{-1})^{-1}. \tag*{(27)}

We use this norm matrix as the pressure preconditioner. Its inverse is the sum

Pδ:=Rδ,h,Pδ−1=Mp−1+Lp,δh−1.(28)P_\delta:=R_{\delta,h},\qquad P_\delta^{-1}=M_p^{-1}+L_{p,\delta_h}^{-1}. \tag*{(28)}

This additive inverse can be viewed in the space decomposition and subspace correction framework of Xu [30], with the two components qh−ϕhq_h-\phi_h and ϕh\phi_h measured by the mass and weighted stiffness energies, respectively. Applying Pδ−1P_\delta^{-1} amounts to adding the results of a pressure mass solve and a distance-weighted stiffness solve. The weight uses local geometric information directly, whereas a constant-scale inverse block Mp−1+(cLp)−1M_p^{-1}+(cL_p)^{-1} assigns one coefficient to the whole domain.

The matrix identity (27) is exact. To obtain uniform spectral bounds for Pδ−1ShP_\delta^{-1}S_h from (25), one additionally needs uniform comparison of the discrete supremum in (24) with the continuous negative norm, and of (26) with KδK_\delta. If these comparisons hold uniformly, then

cqTRδ,hq≤qTShq≤CqTRδ,hq,κ(Pδ−1Sh)≤C/c,(29)c\mathbf{q}^{\mathsf{T}}R_{\delta,h}\mathbf{q}\leq\mathbf{q}^{\mathsf{T}}S_h\mathbf{q}\leq C\mathbf{q}^{\mathsf{T}}R_{\delta,h}\mathbf{q},\qquad\kappa(P_\delta^{-1}S_h)\leq C/c, \tag*{(29)}

where κ\kappa denotes the ratio of the largest to smallest eigenvalue. The constants depend on the comparison bounds as well as the dimension and the Hardy constant. Thus the continuous result motivates the discrete block, while its spectral quality and the effect of approximate auxiliary solves are assessed in Section 5.

For the full Stokes system, we use an upper triangular block preconditioner [19], with PP approximating ShS_h:

PU=[A^BhT0−P],(30)P_U=\begin{bmatrix}\widehat{A}&B_h^{\mathsf{T}}\\0&-P\end{bmatrix}, \tag*{(30)}

where A^\widehat{A} approximates AhA_h.

Numerical Results

We begin with the unit square to verify the basic performance of the distance-weighted pressure block from Section 4 and examine its sensitivity to mesh refinement. Slender channels and periodic pillar arrays provide benchmarks with a natural global length scale, for which existing analyses motivate scale-dependent pressure preconditioners [25, 29]. Comparisons with the pressure mass matrix and a constant-scale sum-space block test whether the boundary-distance weight recovers the effectiveness of these scale-specific choices. To examine local adaptation, we extend these settings to tapered channels, arrays with nonuniform pillar counts, and composite porous–channel domains. These variants test whether the same weight remains effective when different parts of a domain have competing geometric scales, without choosing a new global scale for each configuration. Figure 1 and Figure 2 illustrate the test geometries.

Construction parameters for channels and pillar arrays

Figure 1. Construction parameters for channels and pillar arrays. Grey circles are solid pillars; schematics are not to scale.

Construction of the composite domain showing the porous region, chamber, throat, outlet, and $m \times m$ pillars

Figure 2. Construction of the composite domain. Grey regions are solid.

Common experimental setup

Test geometries

The square baseline (0,1)2(0,1)^2 provides a mesh-refinement control. Slender channels test aspect-ratio effects on pressure stability [9] and width-scaled preconditioning [25]. Tapered channels extend the test to variable local widths, as considered in [25 Remark 3 and Example 5]. Both channel families are described by

Ωε,ρ={(x,y):0<x<1, 0<y<ε[1−(1−ρ−1)x]}(31)\Omega_{\varepsilon,\rho} = \{(x,y) : 0 < x < 1,\ 0 < y < \varepsilon[1 - (1 - \rho^{-1})x]\} \tag*{(31)}

where ε\varepsilon is the maximum width and ρ≥1\rho\ge1 is the ratio of maximum to minimum width. The case ρ=1\rho= 1 gives the rectangular slender channel (0,1)×(0,ε)(0,1) \times(0,\varepsilon).

Periodic pillar arrays test the density-dependent stability and pore-scale pressure norms studied in [28, 29]. They are obtained by removing m2m^2 circular pillars from the unit square. Their centres are ((i+12)/m,(j+12)/m)((i+\frac{1}{2})/m,(j+\frac{1}{2})/m), 0≤i,j<m0 \le i,j < m, and their radius is r=κ/mr = \kappa/m, where κ\kappa is the radius relative to the period 1/m1/m. The graded-density array tests spatially varying pore scales, a feature of heterogeneous rock geometries [23]. It also lies in the unit square, with equal-radius pillars (r=0.005r = 0.005) in four columns at xi=(i+12)/4x_i = (i+\frac{1}{2})/4, i=0,1,2,3i = 0,1,2,3. Column ii contains ni=1+⌊i(nmax⁡−1)/3+1/2⌋n_i = 1 + \lfloor i(n_{\max}-1)/3 + 1/2 \rfloor pillars with vertical coordinates yij=(j+12)/niy_{ij} = (j+\frac{1}{2})/n_i, j=0,…,ni−1j = 0,\ldots,n_i-1. Thus the pillar count increases from one in the first column to nmax⁡n_{\max} in the last. Figure 1 illustrates the channel and pillar-array parameters.

The composite domain is motivated by DLD devices combining pillar arrays with open collection regions [20]. We use a simplified geometry with coexisting pore, chamber, and throat scales. It is formed by removing circular pillars and two rectangular wall blocks from (0,2.15)×(0,1)(0,2.15) \times(0,1). Its square porous region (0.10,1.10)×(0,1)(0.10,1.10) \times(0,1) contains m2m^2 pillars with centres and radius

xij=0.10+(i+12)1m,yij=r+w+j(2r+w),r=κm,x_{ij}=0.10+\left(i+\frac{1}{2}\right)\frac{1}{m}, \qquad y_{ij}=r+w+j(2r+w), \qquad r=\frac{\kappa}{m},

where i,j=0,…,m−1i,j=0,\ldots,m-1 and w=(1−2mr)/(m+1)w=(1-2mr)/(m+1), so the vertical clearances between rows and at the outer walls are equal. The wall blocks occupy

[1.50,1.70]×[0,(1−g)/2],[1.50,1.70]×[(1+g)/2,1],[1.50,1.70]\times[0,(1-g)/2], \qquad[1.50,1.70]\times[(1+g)/2,1],

leaving a throat of width gg centred at y=0.5y=0.5. An open chamber separates the porous region from the throat, followed by an open outlet; see Figure 2 for the construction. m2m^2 is the pillar count, κ\kappa is the radius relative to the horizontal pillar spacing, and gg is the throat width. Parameter ranges and fixed values for each experiment are given below.

Discretization and pressure blocks

All numerical experiments consider the two-dimensional Stokes equations (22). Velocity has prescribed parabolic profiles at the left and right boundaries, with homogeneous Dirichlet conditions on all remaining walls, including every obstacle. We use Taylor–Hood elements with continuous P2P_2 velocity and continuous P1P_1 pressure, restricted to

Qh,0={qh∈Qh:∫Ωqh dx=0}.Q_{h,0}=\left\{q_h\in Q_h:\int_{\Omega}q_h\,\mathrm{d}x=0\right\}.

The boundary distance is approximated by a continuous P2P_2 finite-element function δh\delta_h on triangular meshes generated with Gmsh. We use P2P_2 to avoid the identically zero P1P_1 distance interpolant on triangles whose three vertices lie on the boundary.

We use the pressure mass matrix MpM_p and distance-weighted stiffness matrix Lp,δhL_{p,\delta_h} from § 4.2, and let LpL_p denote the unweighted Neumann stiffness matrix. We compare the pressure preconditioners

PM=Mp,Pc=(Mp−1+(cLp)−1)−1,Pδ=(Mp−1+Lp,δh−1)−1.(32)P_M=M_p,\qquad P_c=\left(M_p^{-1}+(cL_p)^{-1}\right)^{-1},\qquad P_\delta=\left(M_p^{-1}+L_{p,\delta_h}^{-1}\right)^{-1}. \tag*{(32)}

Preconditioning applies P−1P^{-1} for each choice of PP. All inverses are interpreted on the mean-zero pressure space, with natural Neumann conditions for the auxiliary stiffness operators. We prescribe c=W02/12c = W_0^2/12, following the reduced channel model of Sande et al. [25 Section 5]. The factor 1/121/12 comes from averaging the parabolic velocity profile between two parallel no-slip walls. For channels, W0W_0 is the maximum width; for periodic arrays, it is the period; for the graded-density array, it is the horizontal distance between the centre lines of adjacent pillar columns; for the composite domain, it is the horizontal centre-to-centre pillar spacing 1/m1/m in its porous region; and for the unit-square baseline, it is one. Figure 3 illustrates these choices. The use of this channel coefficient on other geometries specifies a reproducible baseline, rather than an optimal coefficient.

Reference lengths for the constant-scale pressure block

Figure 3. Reference lengths W0W_0 (blue arrows) for the constant-scale pressure block. Grey regions are solid; schematics are not to scale.

For a more demanding constant-coefficient comparison, we also compute PαcP_{\alpha c} for α∈{0.1,0.3,1,3,10}\alpha\in\{0.1, 0.3, 1, 3, 10\} on every matrix. We denote by Pc∗P_c^* the best tested constant envelope. For each test case, we select the coefficient giving the smallest condition number among the five candidates. For the FGMRES comparisons with AMG auxiliary solves, we separately select the coefficient giving the fewest iterations among the same five candidates.

Spectral diagnostics

We examine the eigenproblem P−1Shq=λqP^{-1}S_h\mathbf{q}=\lambda\mathbf{q} on Qh,0Q_{h,0}, where Sh=BhAh−1BhTS_h=B_hA_h^{-1}B_h^{\mathsf{T}} is the pressure Schur complement. Its equivalent generalized form is

Shq=λPq.(33)S_h\mathbf{q}=\lambda P\mathbf{q}. \tag*{(33)}

For P=PMP=P_M, the right-hand matrix is MpM_p; for either sum-space block,

P=(Mp−1+L−1)−1,P−1=Mp−1+L−1,L=cLp or Lp,δh.P=\left(M_p^{-1}+L^{-1}\right)^{-1},\qquad P^{-1}=M_p^{-1}+L^{-1},\qquad L=cL_p\ \text{or}\ L_{p,\delta_h}.

The smallest and largest eigenvalues of P−1ShP^{-1}S_h on Qh,0Q_{h,0} are denoted by λmin⁡\lambda_{\min} and λmax⁡\lambda_{\max}, respectively. Their Rayleigh quotient characterizations are

λmin⁡=min⁡q≠0qTShqqTPq,λmax⁡=max⁡q≠0qTShqqTPq.\lambda_{\min}=\min_{\mathbf{q}\ne\mathbf{0}}\frac{\mathbf{q}^{\mathsf{T}}S_h\mathbf{q}}{\mathbf{q}^{\mathsf{T}}P\mathbf{q}},\qquad\lambda_{\max}=\max_{\mathbf{q}\ne\mathbf{0}}\frac{\mathbf{q}^{\mathsf{T}}S_h\mathbf{q}}{\mathbf{q}^{\mathsf{T}}P\mathbf{q}}.

Thus λmin⁡\lambda_{\min} and λmax⁡\lambda_{\max} are the optimal lower and upper constants comparing the Schur complement and preconditioner quadratic forms. We report their ratio C(P)=λmax⁡/λmin⁡C(P)=\lambda_{\max}/\lambda_{\min} as the spectral condition number, and list both endpoints for Pδ−1ShP_\delta^{-1}S_h. For P=PδP=P_\delta, the denominator is qTPδq=∥qh∥Kδh,h2\mathbf{q}^{\mathsf{T}}P_\delta\mathbf{q}=\lVert q_h\rVert_{K_{\delta_h,h}}^2 by (27). The corresponding continuous comparison in (25) gives

cd2≤∥∇q∥−1,Ω2∥q∥Kδ2≤1+CH(Ω)2,q∈L02(Ω)∖{0}.c_d^2\le\frac{\lVert\nabla q\rVert_{-1,\Omega}^2}{\lVert q\rVert_{K_\delta}^2}\le1+C_H(\Omega)^2,\qquad q\in L_0^2(\Omega)\setminus\{0\}.

Thus λmin⁡\lambda_{\min} and λmax⁡\lambda_{\max} correspond to the geometry-independent lower bound and the Hardy-dependent upper bound, respectively. We assume that the discrete negative and weighted norms are uniformly comparable to their continuous counterparts in both mesh size and geometry, so these bounds carry over up to the comparison constants.

Prescribed flow and Krylov solver

There is no body force. On a vertical inlet or outlet of height HH, starting at y=y0y=y_0, we prescribe

uD(y)=(6FHη(1−η),0)T,η=y−y0H.(34)\mathbf{u}_D(y)=\left(\frac{6F}{H}\eta(1-\eta),0\right)^{\mathsf{T}},\qquad\eta=\frac{y-y_0}{H}. \tag*{(34)}

The inlet mean speed is one, so F=HinF=H_{\mathrm{in}}. The same flux FF is used at the outlet to ensure compatibility with incompressibility. After lifting these boundary data, the velocity correction belongs to H01(Ω)dH_0^1(\Omega)^d. The lifting changes only the right-hand side, so the Schur complement uses the homogeneous velocity space introduced in § 4.1.

For each pressure block, the full system is solved with the upper triangular preconditioner (30). The exact stage uses sparse direct solves for the auxiliary blocks. The AMG stage uses one BoomerAMG V-cycle for each inverse application. Both stages use right-preconditioned FGMRES with a relative PETSc tolerance of 10−1010^{-10}; converged runs must also satisfy a measured true relative residual of 10−810^{-8}. All iteration counts refer to outer FGMRES iterations; the labels Exact and AMG distinguish the auxiliary solvers. All methods use the same convergence tolerances. NC denotes failure of FGMRES to converge within 500 iterations.

Refinement controls and square baseline

The square (0,1)2(0,1)^2 is tested on six mesh levels, s=1,2,4,8,16,32s=1,2,4,8,16,32; the slender, periodic, and composite controls are refined at fixed geometry and cc (Table 1). The refinement factor ss scales the target mesh size by 1/s1/s; hmax⁡h_{\max} is the maximum element diameter, and nq=dim⁡Qh,0n_q=\dim Q_{h,0} is the number of mean-zero pressure degrees of freedom.

Refinement at fixed geometry. The slender width is 1/321/32; the periodic parameters are m=4m=4, κ=0.25\kappa=0.25; the composite parameters are m=3m=3, κ=0.25\kappa=0.25, g=0.1g=0.1. The endpoints λmin⁡\lambda_{\min} and λmax⁡\lambda_{\max} are for Pδ−1ShP_{\delta}^{-1}S_h; C(P)C(P) is the spectral condition number. Exact and AMG give its FGMRES iteration counts for the corresponding auxiliary solves.

Geometrysshmax⁡h_{\max}nqn_qC(PM)C(P_M)C(Pc)C(P_c)C(Pc∗)C(P_c^*)C(Pδ)C(P_\delta)λmin⁡\lambda_{\min}λmax⁡\lambda_{\max}ExactAMG
Square10.2545434.464.434.353.360.5341.7981417
Square20.12141434.675.024.663.930.5392.1221718
Square40.06995094.825.404.844.520.5392.4371720
Square80.029419344.935.674.975.060.5392.7261821
Square160.016875465.015.875.075.530.5402.9831821
Square320.0081299835.086.015.155.980.5393.2271822
Slender10.00947991246.407.605.464.700.5322.4991622
Slender20.005127031246.398.755.975.210.5392.8061623
Periodic10.0216456189.7112.8811.345.030.5392.7071822
Periodic20.01051652990.3512.9012.215.480.5392.9521823
Composite10.034736871183.1729.4929.495.230.5432.8382023
Composite20.0175139521192.3329.4529.455.650.5433.0682024

Table 1.

The smallest eigenvalues of Pδ−1ShP_{\delta}^{-1}S_h in Table 1 remain near 0.539 across the different geometries and mesh levels, consistent with the geometry-independent lower bound in (25). The increase in λmax⁡\lambda_{\max} under refinement is consistent with richer finite-element spaces resolving additional pressure modes. On the convex domains, all computed upper endpoints remain below 5, the continuous upper bound obtained from CH≤2C_H \le2. Across all four domain types, the weighted condition numbers remain modest, ranging from 3.36 to 5.98, and the FGMRES iteration counts remain stable under refinement with both exact and AMG auxiliary solves.

Slenderness and spatially varying width

For slender channels (ρ=1)(\rho=1), we vary ε\varepsilon from $1/4toto1/128$, with s=2s=2 and c=ε2/12c=\varepsilon^2/12. In the spectral subset, narrowing the channel from 1/41/4 to $1/32 raises C(PM)C(P_M) from 20.96 to 1246.39, while C(Pc)C(P_c) remains between 8.13 and 8.75 and C(Pδ)C(P_\delta) between 5.01 and 5.21 (Table 2). Both sum-space blocks therefore compensate for uniform slenderness.

GeometryParameternqn_qC(PM)C(P_M)C(Pc)C(P_c)C(Pc∗)C(P_c^*)C(Pδ)C(P_\delta)λmin⁡\lambda_{\min}λmax⁡\lambda_{\max}
Slenderε=1/32\varepsilon= 1/3227031246.398.755.975.210.5392.806
Slenderε=1/16\varepsilon= 1/161417312.688.715.965.210.5392.806
Slenderε=1/8\varepsilon= 1/868279.258.595.885.150.5392.779
Slenderε=1/4\varepsilon= 1/436120.968.135.625.010.5392.697
Taperρ=1\rho= 1491620.969.336.205.970.5393.217
Taperρ=3\rho= 3825249.318.598.596.090.5393.278
Taperρ=5\rho= 51012565.3012.1612.166.100.5393.284
Taperρ=8\rho= 81190380.0119.9118.696.090.5403.290

Table 2. Channel spectral results. Slender channels use s=2s = 2; tapered channels use locally refined physical meshes at s=8s = 8 and maximum width 1/41/4. Taper contrasts are shown at ρ=1,3,5,8\rho= 1, 3, 5, 8. The condition numbers compare all four blocks; the last two columns give the endpoints of Pδ−1ShP_\delta^{-1} S_h.

We then fix ε=1/4\varepsilon=1/4 and introduce tapering, with width contrast ρ\rho up to 8 for the spectral tests (s=8)(s=8) and 48 for the FGMRES tests (s=12)(s=12). As ρ\rho increases from 1 to 8, C(Pc)C(P_c) rises from 9.33 to 19.91, and even the tuned C(Pc∗)C(P_c^*) increases from 6.20 to 18.69. In comparison, C(Pδ)C(P_\delta) stays between 5.97 and 6.10 (Table 2).

A single width provides an effective global scale for a uniform channel; the distance weight also follows the changing local width in a tapered channel. Across both families, λmin⁡(Pδ−1Sh)\lambda_{\min}(P_{\delta}^{-1}S_h) remains near 0.539, consistent with the geometry-independent lower bound. The FGMRES results in Table 3 show the same robustness to both uniform narrowing and spatial width variation. The mass block deteriorates under narrowing, while the constant-scale blocks lose robustness as the width contrast increases.

Slender ε\varepsilon1/41/41/61/61/81/81/121/121/161/161/241/241/321/321/481/481/641/641/961/961/1281/128
PMP_M293336455497155198262NCNC
PcP_c2424232626252526262626
Pc∗P_c^*2424232525252525252626
PδP_\delta2021212223232323232424
Taper ρ\rho1234681216243248
PMP_M2836384043454850535658
PcP_c2528303132343639414447
Pc∗P_c^*2528303131323536394244
PδP_\delta2324252525252625252625

Table 3. FGMRES iteration counts for channel problems with one-cycle AMG auxiliary solves. Slender channels use s=2s = 2; tapered channels use physical meshes at s=12s = 12. Each column is a parameter value; NC denotes failure to converge within 500 iterations.

Obstacle density and obstacle size

For the periodic arrays defined in Section 5.1.1, the period is ℓ=1/m\ell=1/m, the radius is r=κℓr=\kappa\ell, and we set c=ℓ2/12c=\ell^2/12. We separate changes in the overall pore scale from changes in the contrast between local scales. The density sweep fixes κ=0.25\kappa=0.25 and uses

m∈{2,3,4,6,8,10,12,16,24,32,48,64}.m \in\{2,3,4,6,8,10,12,16,24,32,48,64\}.

The obstacle count increases from 4 to 4096, with pillars and passages shrinking together. The size sweep fixes m=4m=4 and varies κ\kappa over twelve values between 0.1 and 0.495. As κ\kappa approaches 1/21/2, the passages between neighbouring pillars become arbitrarily narrow, while the wider pore regions retain a scale comparable to ℓ\ell. This increases the contrast between local boundary-distance scales within the same domain, which a single period-based coefficient cannot track.

Table 4 shows that the smallest eigenvalues of Pδ−1ShP_\delta^{-1}S_h remain between 0.538 and 0.539 in both sweeps. Both sum-space blocks remain stable as density increases. As the radius increases, however, the weighted condition number changes little while the tuned constant envelope deteriorates. The FGMRES results in Table 5 show the same distinction: at κ=0.495\kappa= 0.495, the weighted block retains stable convergence, while the mass block and all five constant candidates fail to converge within 500 iterations.

SweepParameternqn_qC(PM)C(P_M)C(Pc)C(P_c)C(Pc∗)C(P_c^*)C(Pδ)C(P_\delta)λmin⁡\lambda_{\min}λmax⁡\lambda_{\max}
Densitym=2m=2117340.6112.049.784.800.5392.589
Densitym=4m=4456189.7112.8811.345.030.5392.707
Densitym=8m=818112292.4513.1111.835.080.5392.737
Sizeκ=0.1\kappa=0.11539428.0610.268.815.510.5382.965
Sizeκ=0.25\kappa=0.25456189.7112.8811.345.030.5392.707
Sizeκ=0.4\kappa=0.416308606.0154.7521.745.260.5392.836
Sizeκ=0.45\kappa=0.45465612574.97202.1859.145.420.5392.921

Table 4. Obstacle spectral results. Density varies mm at κ=0.25\kappa=0.25; size varies κ\kappa at m=4m=4. The last two columns are the endpoints of Pδ−1ShP_\delta^{-1}S_h.

Density mm2346810121624324864
PMP_M536473121142153170197379NCNCNC
PcP_c353437353535353736363635
Pc∗P_c^*272931323231313231323131
PδP_\delta212222222323232324242424
Size κ\kappa0.10.20.250.30.350.40.4250.450.460.4750.490.495
PMP_M42617397140182268428NCNCNCNC
PcP_c283137435476111152220270NCNC
Pc∗P_c^*283131333745577689132283NC
PδP_\delta222222232323242526272525
Graded nmax⁡n_{\max}12481632648090969899
PMP_M30333540486483166234NCNCNC
PcP_c2828293035456375115202346NC
Pc∗P_c^*2526283035456375115202291NC
PδP_\delta222324232424272728282930

Table 5. FGMRES iteration counts for obstacle arrays with one-cycle AMG auxiliary solves. Each column is a parameter value; NC denotes failure to converge within 500 iterations. The density and size sweeps respectively fix κ=0.25\kappa=0.25 and m=4m=4; the graded-density sweep varies nmax⁡n_{\max} with four columns and equal radii.

Increasing pillar counts across columns

For the graded-density array defined in § 5.1.1, we increase nmax⁡n_{\max} from 1 to 99 while keeping the pillar radius and column positions fixed. This creates increasingly narrow gaps in the denser columns. The weighted block remains robust throughout the sweep, whereas the mass block and all five constant candidates fail to converge within 500 iterations at nmax⁡=99n_{\max}=99 Table 5. These results illustrate the benefit of tracking local scales when the pillar density varies within one domain.

Composite porous–channel geometry

Composite porous–channel geometry. Numerical simulations of microfluidic devices, such as deterministic lateral displacement (DLD) chips, often involve flow passages of different widths within the same geometry. We illustrate this feature with a simplified composite domain defined in § 5.1.1, combining a porous array, an open chamber, and a narrow passage. The prescribed constant uses the porous horizontal pitch, c=1/(12m2)c=1/(12m^2).

Three independent sweeps use twelve values each: (i) gg from 0.2 to 0.005 at m=3m=3, κ=0.25\kappa=0.25; (ii) m∈{2,3,4,5,6,8,10,12,16,24,32,64}m \in\{2,3,4,5,6,8,10,12,16,24,32,64\} at g=0.1g=0.1, κ=0.25\kappa=0.25; and (iii) κ\kappa from 0.1 to 0.495 at m=3m = 3, g=0.1g = 0.1. This design tests whether a local pressure weight can accommodate porous, open, and narrow-channel scales within the same domain. Figure 4 illustrates the computed flow and local boundary distance on a denser composite with m=8m = 8, κ=0.35\kappa= 0.35, and g=0.05g = 0.05.

Computed dense composite flow showing speed and streamlines, mean-zero pressure, and local boundary distance

Figure 4. Computed dense composite flow at g=0.05g = 0.05, m=8m = 8, κ=0.35\kappa= 0.35 (64 pillars): velocity magnitude and streamlines (left), mean-zero pressure (centre), and boundary distance δh\delta_h (right).

The composite gives the clearest separation between local and global scaling because pore gaps, wide chambers, and a narrow throat coexist. The weighted block remains robust under changes in throat width, obstacle density, and obstacle radius, while the mass and constant-scale blocks deteriorate markedly (Table 6). At κ=0.495\kappa= 0.495, the mass block and all five constant candidates fail to converge within 500 iterations. The spectral subset also shows a factor of 16.75 improvement over the five-candidate envelope at g=0.05g = 0.05.

Throat gg0.20.150.10.0750.050.040.030.0250.020.0150.010.005
PMP_M6868729081113166157207207439NC
PcP_c3334373944465255607094395
Pc∗P_c^*3334373944465255607094395
PδP_\delta222223232324252525252627
Density mm234568101216243264
PMP_M6072114168185235298347450NCNCNC
PcP_c363739414553606680117154294
Pc∗P_c^*333739414553606679109145269
PδP_\delta222323242424242425252729
Size κ\kappa0.10.150.20.250.30.350.40.450.460.4750.490.495
PMP_M56586472112112238NCNCNCNCNC
PcP_c34343537405066138140189463NC
Pc∗P_c^*3434353740445587101163376NC
PδP_\delta242423232223232424242626

Table 6. FGMRES iteration counts for composite domains with one-cycle AMG auxiliary solves. Each column is a parameter value; NC denotes failure to converge within 500 iterations. Fixed parameters are as in Table 7.

The refinement control in Table 1 supports the persistence of this robustness under mesh refinement. Table 7 records the corresponding spectral subset. The weighted lower endpoints again cluster near 0.54, ranging from 0.541 to 0.547, while the upper endpoints remain bounded over these sweeps.

SweepParameternqn_qC(PM)\mathcal{C}(P_M)C(Pc)\mathcal{C}(P_c)C(Pc∗)\mathcal{C}(P_c^*)C(Pδ)\mathcal{C}(P_\delta)λmin⁡\lambda_{\min}λmax⁡\lambda_{\max}
Throatg=0.05g=0.05138207767.48112.9595.245.690.5443.093
Throatg=0.1g=0.136871183.1729.4929.495.230.5432.838
Throatg=0.2g=0.22831283.9015.8115.815.060.5422.743
Densitym=2m=236541156.5639.7030.805.230.5432.838
Densitym=3m=336871183.1729.4929.495.230.5432.838
Densitym=4m=449491223.5725.6225.625.350.5412.895
Sizeκ=0.1\kappa=0.1173561185.7132.7932.795.690.5443.093
Sizeκ=0.25\kappa=0.2536871183.1729.4929.495.230.5432.838
Sizeκ=0.4\kappa=0.4118452105.6470.7162.535.630.5473.083

Table 7. Composite spectral results. The throat sweep fixes m=3m=3, κ=0.25\kappa=0.25; the density sweep fixes g=0.1g=0.1, κ=0.25\kappa=0.25; the size sweep fixes m=3m=3, g=0.1g=0.1. The last two columns are the endpoints of Pδ−1ShP_\delta^{-1}S_h. Larger iteration-only cases appear in Table 6.

Conclusion

Conclusion. The main result is a distance-weighted Nečas-type inequality on every bounded connected open set: ∥q∥Kδ≤Cd∥∇q∥−1,Ω\lVert q\rVert_{K_\delta} \le C_d\lVert\nabla q\rVert_{-1,\Omega}, with CdC_d depending only on the dimension. Local Nečas estimates on enlarged Whitney cubes and negative-norm localization control the local oscillations. A smooth partition of unity combines the cube averages into a competitor in the natural weighted energy space, with both terms in the KK-functional controlled by those oscillations.

Equivalence with the negative norm of the gradient follows as a corollary when the boundary Hardy inequality holds. The reverse estimate has constant 1+CH(Ω)2\sqrt{1+C_H(\Omega)^2}, so the Hardy constant is the only domain-dependent quantity in the resulting equivalence. The corollary is uniform on families with uniformly bounded Hardy constants; for all bounded simply connected planar domains, the reverse constant can be chosen as 17\sqrt{17}. It motivates a pressure preconditioner that adapts to the local boundary distance, as examined in § 5.

The central numerical finding is that λmin⁡(Pδ−1Sh)\lambda_{\min}(P_\delta^{-1}S_h) stays close to 0.54 across the reported geometries and mesh refinements. This common empirical lower spectral level is consistent with the geometry-independent lower bound in the continuous estimate. Across the geometric sweeps, the distance-weighted block consistently requires about 20–30 FGMRES iterations with AMG auxiliary solves. The mass and constant-scale blocks lose robustness as the geometry becomes more extreme; in the near-contact array and composite tests, the mass block and all five constant candidates fail to converge within 500 iterations. Extending these controls to graded DLD chips and geological porous media, with scalable weighted auxiliary solves, is a next step.

References

References

  1. [1]G. Allaire, Homogenization of the Navier–Stokes equations in open sets perforated with tiny holes I: Abstract framework, a volume distribution of holes, Archive for Rational Mechanics and Analysis, 113 (1991), pp. 209–259, https://doi.org/10.1007/BF00375065.
  2. [2]G. Allaire, Homogenization of the Navier–Stokes equations in open sets perforated with tiny holes II: Non-critical sizes of the holes for a volume distribution and a surface distribution of holes, Archive for Rational Mechanics and Analysis, 113 (1991), pp. 261–298, https://doi.org/10.1007/bf00375066.
  3. [3]C. Amirouche, P. G. Ciarlet, and C. Mardare, On a lemma of Jacques-Louis Lions and its relation to other fundamental results, Journal de Mathématiques Pures et Appliquées, 104 (2015), pp. 207–226, https://doi.org/10.1016/j.matpur.2014.11.007.
  4. [4]A. Ancona, On strong barriers and an inequality of Hardy for domains in ℝⁿ, Journal of the London Mathematical Society, 34 (1986), pp. 274–290, https://doi.org/10.1112/jlms/s2-34.2.274.
  5. [5]M. Benzi, G. H. Golub, and J. Liesen, Numerical solution of saddle point problems, Acta Numerica, 14 (2005), pp. 1–137, https://doi.org/10.1017/S0962492904000212.
  6. [6]J. H. Bramble, A proof of the inf-sup condition for the Stokes equations on Lipschitz domains, Mathematical Models and Methods in Applied Sciences, 13 (2003), pp. 361–371, https://doi.org/10.1142/S0218202503002544.
  7. [7]J. H. Bramble and J. E. Pasciak, A preconditioning technique for indefinite systems resulting from mixed approximations of elliptic problems, Mathematics of Computation, 50 (1988), pp. 1–17, https://doi.org/10.1090/S0025-5718-1988-0917816-8.
  8. [8]E. V. Chizhonkov and M. A. Olshanskii, On the domain geometry dependence of the LBB condition, ESAIM: Mathematical Modelling and Numerical Analysis, 34 (2000), pp. 935–951, https://doi.org/10.1051/m2an:2000110.
  9. [9]M. Dobrowolski, On the LBB constant on stretched domains, Mathematische Nachrichten, 254–255 (2003), pp. 64–67, https://doi.org/10.1002/mana.200310052.
  10. [10]M. Dobrowolski, On domain-robust preconditioners for the Stokes equations, Electronic Transactions on Numerical Analysis, 39 (2012), pp. 60–74, https://etna.ricam.oeaw.ac.at/vol.39.2012/pp60-74.dir/pp60-74.pdf.
  11. [11]M. Egert, R. Haller-Dintelmann, and J. Rehberg, Hardy’s inequality for functions vanishing on a part of the boundary, Potential Analysis, 43 (2015), pp. 49–78, https://doi.org/10.1007/s11118-015-9463-8.
  12. [12]H. C. Elman, D. J. Silvester, and A. J. Wathen, Finite Elements and Fast Iterative Solvers: With Applications in Incompressible Fluid Dynamics, Oxford University Press, 2014.
  13. [13]C. A. Figueroa, I. E. Vignon-Clementel, K. E. Jansen, T. J. R. Hughes, and C. A. Taylor, A coupled momentum method for modeling blood flow in three-dimensional deformable arteries, Computer Methods in Applied Mechanics and Engineering, 195 (2006), pp. 5685–5706, https://doi.org/10.1016/j.cma.2005.11.011.
  14. [14]V. Girault and P.-A. Raviart, Finite Element Methods for Navier–Stokes Equations: Theory and Algorithms, vol. 5 of Springer Series in Computational Mathematics, Springer-Verlag, 1986, https://doi.org/10.1007/978-3-642-61623-5.
  15. [15]P. Grisvard, Elliptic Problems in Nonsmooth Domains, vol. 69 of Classics in Applied Mathematics, SIAM, 2011. Originally published by Pitman, 1985.DOI
  16. [16]R. Korte, J. Lehrbäck, and H. Tuominen, The equivalence between pointwise Hardy inequalities and uniform flatness, Mathematische Annalen, 351 (2011), pp. 711–731, https://doi.org/10.1007/s00208-010-0617-0.
  17. [17]A. Laptev and A. V. Sobolev, Hardy inequalities for simply connected planar domains, American Mathematical Society Translations, Series 2, 225 (2008), pp. 133–140, https://doi.org/10.1090/trans2/225/09.
  18. [18]J. Lehrbäck, Weighted Hardy inequalities beyond Lipschitz domains, Proceedings of the American Mathematical Society, 142 (2014), pp. 1705–1715, https://doi.org/10.1090/S0002-9939-2014-11904-6.
  19. [19]D. Loghin and A. J. Wathen, Analysis of preconditioners for saddle-point problems, SIAM Journal on Scientific Computing, 25 (2004), pp. 2029–2049, https://doi.org/10.1137/S1064827502418203.
  20. [20]K. Loutherback, J. D’Silva, L. Liu, A. Wu, R. H. Austin, and J. C. Sturm, Deterministic separation of cancer cells from blood at 10 mL/min, AIP Advances, 2 (2012), 042107, https://doi.org/10.1063/1.4758131.
  21. [21]Y. Lu, Homogenization of Stokes equations in perforated domains: A unified approach, Journal of Mathematical Fluid Mechanics, 22 (2020), 44, https://doi.org/10.1007/s00021-020-00506-6.
  22. [22]K.-A. Mardal and R. Winther, Preconditioning discretizations of systems of partial dif- ferential equations, Numerical Linear Algebra with Applications, 18 (2011), pp. 1–40, https://doi.org/10.1002/nla.716.
  23. [23]P. Mostaghimi, M. J. Blunt, and B. Bijeljic, Computations of absolute permeability on micro-CT images, Mathematical Geosciences, 45 (2013), pp. 103–125, https://doi.org/10.1007/s11004-012-9431-4.
  24. [24]J. Nečas, Les méthodes directes en théorie des équations elliptiques, Academia, Prague, and Masson, Paris, 1967.
  25. [25]E. Sande, T. Koch, M. Kuchta, and K.-A. Mardal, On a robust inf-sup condition for the Stokes problem in slender domains – with application to preconditioning. arXiv preprint, 2025, https://arxiv.org/abs/2510.24590.
  26. [26]L. Shen, Y. Zhang, Y. Chen, X. Ding, P. Wen, C. Wang, P. Sun, S. Gong, J. Xu, J. Han, and Y. Chen, A high-fidelity 3D fluid-structure interaction framework for predictive microfluidic design, 2026, https://doi.org/10.64898/2026.04.28.721227. bioRxiv preprint.
  27. [27]E. M. Stein, Singular Integrals and Differentiability Properties of Functions, vol. 30 of Princeton Mathematical Series, Princeton University Press, Princeton, NJ, 1970.
  28. [28]Q. Xin, S. Gong, and J. Xu, Sharp inf-sup estimate for the Stokes equation in tight domains with periodic pillars and some numerical implications. arXiv preprint, 2026, https://arxiv.org/abs/2604.12643.
  29. [29]Q. Xin, Y. Xie, C.-S. Zhang, S. Gong, and J. Xu, Uniform inf-sup norm equivalence and robust operator preconditioning for Stokes flow in tight domains with periodic pillars. arXiv preprint, 2026, https://arxiv.org/abs/2609.06999.
  30. [30]J. Xu, Iterative methods by space decomposition and subspace correction, SIAM Review, 34 (1992), pp. 581–613, https://doi.org/10.1137/1034116.

Paper details

vlnc:W_qpc2qz4u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

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