Department of Mathematics, City University of Hong Kong, Kowloon Tong, Hong Kong SAR, China. Email: leiwei@cityu.edu.hk.
Corresponding author. Department of Mathematics, City University of Hong Kong, Kowloon Tong, Hong Kong SAR, China. Email: linazha@cityu.edu.hk. L. Zhao is partially supported by Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11302424, CityU 11303225) and Young Scientists fund, NSFC (Project No. 12401543).
Introduction
The quasi-static Biot model couples porous-solid deformation and pore-fluid flow, with applications in geomechanics and subsurface flow [10 , 14 ] . Robust discretizations and solvers must handle nearly incompressible solids, small or vanishing storage, and weak Darcy control caused by low permeability or small time steps. These challenges motivate parameter-robust formulations and block preconditioners based on parameter-weighted norms [29 , 27 ] .
For Biot’s equations, existing discretizations include classical displacement–pressure formulations [30 , 3 ] , locking-free total-pressure formulations [33 , 27 ] , and fully mixed formulations retaining stress and Darcy velocity [25 , 6 ] . The latter directly approximate both fields and, with suitable H ( div ) H(\operatorname{div}) H ( div ) -conforming spaces, preserve local momentum and mass balances. For plane elasticity, Arnold et al. [7 ] established mixed finite element error estimates uniform with respect to compressibility, including the incompressible limit. Weak stress symmetry is imposed through a rotation multiplier [8 ] . Before condensation, fully mixed formulations lead to large four- or five-field saddle-point systems. Local condensation and hybridization reduce the globally coupled unknowns while retaining conservation properties [32 , 24 ] .
Multipoint flux and stress methods solve local systems on interaction regions formed by cells or subcells sharing a mesh vertex [1 , 36 , 4 , 5 ] . In these multipoint mixed finite element methods, vertex quadrature produces block-diagonal flux or stress matrices, enabling local elimination around each vertex. The closely related coupled multipoint stress–multipoint flux mixed finite element method for Biot’s equations [6 ] uses lowest-order BDM stress and Darcy velocity, cellwise constant displacement and pressure, and continuous piecewise linear or bilinear rotations. Quadrature of the velocity, stress, and stress–rotation forms permits local elimination of stress and Darcy velocity followed by rotation, yielding a cell-centered displacement–pressure matrix that is block-skew-symmetric and positive definite. Fu and Zhao [17 ] instead developed a locking-free control-volume method for elasticity on quadrilateral grids, with local momentum conservation and piecewise constant stress, displacement, and rotation that simplify implementation. Their stress space localizes the compliance form to vertex interaction regions by construction, eliminating the need for vertex quadrature; in the macrocell-rotation variant, the globally coupled unknowns are the cell-centered displacement and rotation. We extend this variant to Biot’s equations using the same subcell-wise constant normal-flux space for stress rows and Darcy velocity, and macrocell-wise constant displacement, rotation, and pressure. The stress and scaled Darcy blocks are block diagonal over vertex interaction regions by construction and are eliminated independently, yielding a symmetric displacement–rotation–pressure system. A macrocell rotation-jump stabilization acts only in the weak-symmetry equation, preserving local momentum and mass balances.
Parameter-robust block preconditioners for Biot systems have been developed primarily within the operator-preconditioning framework and through closely related norm-based approaches. Operator preconditioning equips the underlying product space with parameter-weighted norms for which the variational operator satisfies parameter-independent continuity and inf–sup estimates. The inverse Riesz map associated with these norms, or a uniformly spectrally equivalent approximation, then defines a robust preconditioner [21 , 29 ] . This framework has been applied to total-pressure formulations [27 , 26 ] and to conservative four-field total-pressure formulations [11 ] . Causemann and Kuchta [12 ] developed a scalable, parameter-robust solver for cell-by-cell poroelasticity based on norm-equivalent preconditioning and algebraic multigrid approximations of the diagonal blocks. Related norm-based constructions have been proposed for classical two- and three-field formulations [13 ] , stabilized discretizations [2 ] , and hybridized systems [24 , 20 ] . Huang and Wang [22 ] introduced parameter-robust block-triangular preconditioners for two- and three-field poroelasticity formulations that avoid the explicit construction of a Schur complement. For statically condensed systems, robustness must be established for the reduced operator itself: one may reduce a full-system preconditioner under suitable uniform conditions [19 , 20 ] or approximate the condensed blocks directly [16 ] . We adopt the latter approach for the reduced displacement–rotation–pressure system. Exact Schur factorizations provide the algebraic benchmark for the mechanics-induced pressure correction, while weighted field-of-values (FOV) estimates are used to analyze the nonsymmetric triangularly preconditioned operators [31 , 23 , 28 ] .
Two issues must be resolved to construct robust preconditioners for the condensed system. First, separating displacement and rotation makes mechanics solves cheaper, but the compliance operator loses uniform coercivity on spherical stresses near incompressibility; uniform control of the displacement–rotation coupling is therefore needed. Second, when storage and Darcy control are weak, the natural pressure block may fail to control the mechanics–pressure coupling. The pressure-control criterion for an uncoupled strategy must therefore account jointly for storage, residual compliance, and Darcy contributions, with a bound independent of the mesh size and physical parameters. At zero storage, this strategy remains valid only if the natural pressure block is positive definite and the criterion holds uniformly. A Schur-consistent alternative approximates the mechanics-induced pressure correction explicitly. Our analysis distinguishes these strategies and their applicable parameter regimes. The main contributions are summarized as follows.
(i) We establish the well-posedness of the proposed fully discrete control-volume scheme for the Biot model using inf–sup estimates for stabilized linear elasticity and Darcy flow. Notably, the rotation-jump stabilization added to the weak-symmetry equation enables the elasticity inf–sup estimate and yields a condensed rotation block uniformly spectrally equivalent to a μ \mu μ -scaled mass matrix.
(ii) We prove spectral equivalence between the coupled displacement–rotation block and its block-diagonal part. For uncoupled preconditioners, we establish an exact discrete pressure-control criterion admitting c 0 ≍ K d r − 1 c_0 \asymp K_{\mathrm{dr}}^{-1} c 0 ≍ K dr − 1 , so storage may vanish as λ → ∞ \lambda\to\infty λ → ∞ .
(iii) We construct a computable Schur-consistent pressure block and prove its uniform spectral equivalence to the exact pressure Schur complement, yielding parameter-independent spectral bounds for block-diagonal preconditioning. We also establish uniform weighted FOV bounds for the exact and Schur-consistent lower-triangular preconditioners.
(iv) We develop tailored two-grid solvers for the displacement and pressure blocks and use an inexpensive mass matrix to approximate the rotation block. Experiments in 2D and 3D assess parameter robustness and computational cost.
The remainder of the paper is organized as follows. Section 2 presents the Biot model, introduces the control-volume discretization and proves well-posedness of the fully discrete scheme. Section 3 establishes the mechanics-block equivalence, analyzes the uncoupled block preconditioners, and develops the Schur-consistent block-diagonal and lower-triangular preconditioners together with their spectral and FOV estimates. Section 4 reports the numerical experiments, and Section 5 gives the conclusions.
The control-volume discretization for the Biot model.
The mathematical model.
Let d ∈ { 2 , 3 } d \in\{2,3\} d ∈ { 2 , 3 } , and let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a bounded polygonal domain for d = 2 d=2 d = 2 or a bounded polyhedral domain for d = 3 d=3 d = 3 . We denote by ( ⋅ , ⋅ ) D (\cdot,\cdot)_D ( ⋅ , ⋅ ) D the L 2 ( D ) L^2(D) L 2 ( D ) inner product and write ( ⋅ , ⋅ ) : = ( ⋅ , ⋅ ) Ω (\cdot,\cdot):=(\cdot,\cdot)_\Omega ( ⋅ , ⋅ ) := ( ⋅ , ⋅ ) Ω . Boldface denotes vectors, and underlined symbols denote matrix-valued quantities. Let σ ‾ e \underline{\sigma}_e σ e , u u u , z z z , and p p p denote the elastic stress, displacement, Darcy velocity, and pore pressure. The constitutive law is
A σ ‾ e = ε ( u ) , ε ( u ) : = 1 2 ( ∇ u + ∇ u T ) . (1) \mathcal{A}\underline{\sigma}_e=\varepsilon(u),\qquad\varepsilon(u):=\frac{1}{2}\left(\nabla u+\nabla u^{\mathsf T}\right). \tag*{(1)} A σ e = ε ( u ) , ε ( u ) := 2 1 ( ∇ u + ∇ u T ) . (1)
For homogeneous isotropic elasticity, the compliance operator is
A τ ‾ = 1 2 μ ( τ ‾ − λ 2 μ + d λ tr ( τ ‾ ) I ) , μ ≥ μ min > 0 , λ ≥ 0. (2) \mathcal{A}\underline{\tau}=\frac{1}{2\mu}\left(\underline{\tau}-\frac{\lambda}{2\mu+d\lambda}\operatorname{tr}(\underline{\tau})I\right),\qquad\mu\geq\mu_{\min}>0,\qquad\lambda\geq0. \tag*{(2)} A τ = 2 μ 1 ( τ − 2 μ + d λ λ tr ( τ ) I ) , μ ≥ μ m i n > 0 , λ ≥ 0. (2)
Here λ \lambda λ and μ \mu μ are the Lamé parameters describing volumetric and shear stiffness, respectively. The total poroelastic stress is
σ ‾ = σ ‾ e − α B p I , 0 < α B ≤ 1. (3) \underline{\sigma}=\underline{\sigma}_e-\alpha_B pI,\qquad0<\alpha_B\leq1. \tag*{(3)} σ = σ e − α B p I , 0 < α B ≤ 1. (3)
The Biot–Willis coefficient α B \alpha_B α B measures the coupling between pore pressure and stress. Given a body force f f f and a fluid source q q q , the quasi-static Biot system reads
− div σ ‾ = f in Ω × ( 0 , T ] , K − 1 z + ∇ p = 0 in Ω × ( 0 , T ] , ∂ t ( c 0 p + α B ∇ ⋅ u ) + ∇ ⋅ z = q in Ω × ( 0 , T ] . \begin{aligned}
-\operatorname{div}\underline{\sigma}&=f &&\text{in }\Omega\times(0,T],\\
K^{-1}z+\nabla p&=0 &&\text{in }\Omega\times(0,T],\\
\partial_t(c_0p+\alpha_B\nabla\cdot u)+\nabla\cdot z&=q &&\text{in }\Omega\times(0,T].
\end{aligned} − div σ K − 1 z + ∇ p ∂ t ( c 0 p + α B ∇ ⋅ u ) + ∇ ⋅ z = f = 0 = q in Ω × ( 0 , T ] , in Ω × ( 0 , T ] , in Ω × ( 0 , T ] .
Here c 0 ≥ 0 c_0 \ge0 c 0 ≥ 0 is the storage coefficient, and the hydraulic mobility tensor K K K is symmetric and uniformly positive definite, with K 1 ∣ ξ ∣ 2 ≤ K ( x ) ξ ⋅ ξ ≤ K 2 ∣ ξ ∣ 2 K_1|\boldsymbol{\xi}|^2 \le K(x)\boldsymbol{\xi}\cdot\boldsymbol{\xi} \le K_2|\boldsymbol{\xi}|^2 K 1 ∣ ξ ∣ 2 ≤ K ( x ) ξ ⋅ ξ ≤ K 2 ∣ ξ ∣ 2 for a.e. x ∈ Ω x \in\Omega x ∈ Ω and all ξ ∈ R d \boldsymbol{\xi} \in\mathbb{R}^d ξ ∈ R d , where 0 < K 1 ≤ K 2 0 < K_1 \le K_2 0 < K 1 ≤ K 2 . We impose homogeneous displacement and pressure boundary conditions and prescribe p ( 0 ) = p 0 p(0)=p^0 p ( 0 ) = p 0 .
To impose stress symmetry weakly, let m d = d ( d − 1 ) / 2 m_d=d(d-1)/2 m d = d ( d − 1 ) /2 and introduce the rotation variable γ ∈ R m d \gamma\in\mathbb{R}^{m_d} γ ∈ R m d . For d = 2 d=2 d = 2 , γ = 1 2 rot ( u ) \gamma=\frac{1}{2}\operatorname{rot}(\boldsymbol{u}) γ = 2 1 rot ( u ) and as ( τ ) = τ 21 − τ 12 \operatorname{as}(\boldsymbol{\tau})=\tau_{21}-\tau_{12} as ( τ ) = τ 21 − τ 12 , whereas for d = 3 d=3 d = 3 , γ = 1 2 curl ( u ) \gamma=\frac{1}{2}\operatorname{curl}(\boldsymbol{u}) γ = 2 1 curl ( u ) and as ( τ ) = ( τ 32 − τ 23 , τ 13 − τ 31 , τ 21 − τ 12 ) T \operatorname{as}(\boldsymbol{\tau})=(\tau_{32}-\tau_{23},\tau_{13}-\tau_{31},\tau_{21}-\tau_{12})^{\mathsf{T}} as ( τ ) = ( τ 32 − τ 23 , τ 13 − τ 31 , τ 21 − τ 12 ) T . Using {(2.1)–(2.3)} and ∇ ⋅ u = tr A ( σ + α B p I ) \nabla\cdot\boldsymbol{u}=\operatorname{tr}\mathcal{A}(\boldsymbol{\sigma}+\alpha_BpI) ∇ ⋅ u = tr A ( σ + α B p I ) , we obtain the following weakly symmetric five-field formulation: Find σ ∈ C ( [ 0 , T ] ; L 2 ( Ω ) d × d ) \boldsymbol{\sigma}\in C([0,T];L^2(\Omega)^{d\times d}) σ ∈ C ([ 0 , T ] ; L 2 ( Ω ) d × d ) , u ∈ C ( [ 0 , T ] ; H 0 1 ( Ω ) d ) \boldsymbol{u}\in C([0,T];H_0^1(\Omega)^d) u ∈ C ([ 0 , T ] ; H 0 1 ( Ω ) d ) , γ ∈ C ( [ 0 , T ] ; L 2 ( Ω ) m d ) \gamma\in C([0,T];L^2(\Omega)^{m_d}) γ ∈ C ([ 0 , T ] ; L 2 ( Ω ) m d ) , z ∈ C ( [ 0 , T ] ; L 2 ( Ω ) d ) \boldsymbol{z}\in C([0,T];L^2(\Omega)^d) z ∈ C ([ 0 , T ] ; L 2 ( Ω ) d ) , and p ∈ C ( [ 0 , T ] ; H 0 1 ( Ω ) ) p\in C([0,T];H_0^1(\Omega)) p ∈ C ([ 0 , T ] ; H 0 1 ( Ω )) such that, for all ( ψ , v , η , ξ , w ) ∈ L 2 ( Ω ) d × d × H 0 1 ( Ω ) d × L 2 ( Ω ) m d × L 2 ( Ω ) d × H 0 1 ( Ω ) (\boldsymbol{\psi},\boldsymbol{v},\boldsymbol{\eta},\boldsymbol{\xi},w)\in L^2(\Omega)^{d\times d}\times H_0^1(\Omega)^d\times L^2(\Omega)^{m_d}\times L^2(\Omega)^d\times H_0^1(\Omega) ( ψ , v , η , ξ , w ) ∈ L 2 ( Ω ) d × d × H 0 1 ( Ω ) d × L 2 ( Ω ) m d × L 2 ( Ω ) d × H 0 1 ( Ω ) ,
( A ( σ + α B p I ) , ψ ) − ( ∇ u , ψ ) + ( γ , as ( ψ ) ) = 0 , (\mathcal{A}(\boldsymbol{\sigma}+\alpha_BpI),\boldsymbol{\psi})-(\nabla\boldsymbol{u},\boldsymbol{\psi})+(\gamma,\operatorname{as}(\boldsymbol{\psi}))=0, ( A ( σ + α B p I ) , ψ ) − ( ∇ u , ψ ) + ( γ , as ( ψ )) = 0 ,
( σ , ∇ v ) = ( f , v ) , (\boldsymbol{\sigma},\nabla\boldsymbol{v})=(\boldsymbol{f},\boldsymbol{v}), ( σ , ∇ v ) = ( f , v ) ,
( as ( σ ) , η ) = 0 , (\operatorname{as}(\boldsymbol{\sigma}),\boldsymbol{\eta})=0, ( as ( σ ) , η ) = 0 ,
( K − 1 z , ξ ) + ( ∇ p , ξ ) = 0 , (K^{-1}\boldsymbol{z},\boldsymbol{\xi})+(\nabla p,\boldsymbol{\xi})=0, ( K − 1 z , ξ ) + ( ∇ p , ξ ) = 0 ,
c 0 ( ∂ t p , w ) + α B ( ∂ t A ( σ + α B p I ) , w I ) − ( z , ∇ w ) = ( q , w ) . c_0(\partial_t p,w)+\alpha_B(\partial_t\mathcal{A}(\boldsymbol{\sigma}+\alpha_BpI),wI)-(\boldsymbol{z},\nabla w)=(q,w). c 0 ( ∂ t p , w ) + α B ( ∂ t A ( σ + α B p I ) , w I ) − ( z , ∇ w ) = ( q , w ) .
Unless stated otherwise, the hidden constants in ≲ \lesssim ≲ , ≳ \gtrsim ≳ , and ≃ \simeq ≃ are independent of the mesh size; dependence on physical parameters and on Δ t \Delta t Δ t is indicated explicitly.
Discrete spaces and fully discrete formulation
Let T M \mathcal{T}_M T M be a shape-regular partition of Ω \Omega Ω into macro-cells. Each M ∈ T M M\in\mathcal{T}_M M ∈ T M is divided into 2 d 2^d 2 d subcells by connecting an interior point of M M M to the midpoints of its edges in 2D or to the corresponding lower-dimensional face centers in 3D. The resulting subcell mesh is denoted by T h \mathcal{T}_h T h . The split macro-facets form the set F p r 1 / 2 \mathcal{F}_{\mathrm{pr}}^{1/2} F pr 1/2 , the facets internal to the macro-cells form F d l \mathcal{F}_{\mathrm{dl}} F dl , and F h = F p r 1 / 2 ∪ F d l \mathcal{F}_h=\mathcal{F}_{\mathrm{pr}}^{1/2}\cup\mathcal{F}_{\mathrm{dl}} F h = F pr 1/2 ∪ F dl . For each interior facet F = ∂ E + ∩ ∂ E − F=\partial E^+\cap\partial E^- F = ∂ E + ∩ ∂ E − , let n \boldsymbol{n} n point from E + E^+ E + to E − E^- E − and write ⟦ ϕ ⟧ = ϕ + − ϕ − \llbracket\phi\rrbracket=\phi^+-\phi^- [ [ ϕ ] ] = ϕ + − ϕ − ; on boundary facets, n \boldsymbol{n} n is the outward unit normal and the jump denotes the trace. For M ∈ T M M\in\mathcal{T}_M M ∈ T M , let n M \boldsymbol{n}_M n M denote the outward unit normal on ∂ M \partial M ∂ M . The subcells sharing a primal vertex constitute an interaction region. For further details of the mesh construction, see [17 ] .
Define the normal-flux space
V h : = { ξ h ∈ L 2 ( Ω ) d : ξ h ∣ E ∈ P 0 ( E ) ∀ E ∈ T h , ⟦ ξ h ⋅ n ⟧ F = 0 ∀ F ∈ F p r 1 / 2 ∖ ∂ Ω } . V_h:=\left\{\boldsymbol{\xi}_h\in L^2(\Omega)^d:\boldsymbol{\xi}_h|_E\in P_0(E)\ \forall E\in\mathcal{T}_h,\ \llbracket\boldsymbol{\xi}_h\cdot\boldsymbol{n}\rrbracket_F=0\ \forall F\in\mathcal{F}_{\mathrm{pr}}^{1/2}\setminus\partial\Omega\right\}. V h := { ξ h ∈ L 2 ( Ω ) d : ξ h ∣ E ∈ P 0 ( E ) ∀ E ∈ T h , [ [ ξ h ⋅ n ] ] F = 0 ∀ F ∈ F pr 1/2 ∖ ∂ Ω } .
Its degrees of freedom are the oriented normal traces on the primal subfacets. Let P 0 ( T M ) : = { w h ∈ L 2 ( Ω ) : w h ∣ M ∈ P 0 ( M ) ∀ M ∈ T M } P_0(\mathcal{T}_M):=\{w_h\in L^2(\Omega):w_h|_M\in P_0(M)\ \forall M\in\mathcal{T}_M\} P 0 ( T M ) := { w h ∈ L 2 ( Ω ) : w h ∣ M ∈ P 0 ( M ) ∀ M ∈ T M } . The discrete spaces are Σ h = [ V h ] d \Sigma_h=[V_h]^d Σ h = [ V h ] d , Z h = V h Z_h=V_h Z h = V h , U h = [ P 0 ( T M ) ] d U_h=[P_0(\mathcal{T}_M)]^d U h = [ P 0 ( T M ) ] d , P h = P 0 ( T M ) P_h=P_0(\mathcal{T}_M) P h = P 0 ( T M ) , and Γ h = [ P 0 ( T M ) ] m d \Gamma_h=[P_0(\mathcal{T}_M)]^{m_d} Γ h = [ P 0 ( T M ) ] m d .
For τ h , ψ h ∈ Σ h \boldsymbol{\tau}_h,\boldsymbol{\psi}_h\in\Sigma_h τ h , ψ h ∈ Σ h , v h ∈ U h \boldsymbol{v}_h\in U_h v h ∈ U h , η h ∈ Γ h \boldsymbol{\eta}_h\in\Gamma_h η h ∈ Γ h , z h , ξ h ∈ Z h \boldsymbol{z}_h,\boldsymbol{\xi}_h\in Z_h z h , ξ h ∈ Z h , and r h , w h ∈ P h r_h,w_h\in P_h r h , w h ∈ P h , we define
a σ σ ( τ h , ψ h ) : = ( A τ h , ψ h ) , a σ p ( r h , ψ h ) : = ( α B A ( r h I ) , ψ h ) , a p σ ( τ h , w h ) : = a σ p ( w h , τ h ) , a_{\sigma\sigma}(\boldsymbol{\tau}_h,\boldsymbol{\psi}_h):=(\mathcal{A}\boldsymbol{\tau}_h,\boldsymbol{\psi}_h),\qquad a_{\sigma p}(r_h,\boldsymbol{\psi}_h):=(\alpha_B\mathcal{A}(r_hI),\boldsymbol{\psi}_h),\qquad a_{p\sigma}(\boldsymbol{\tau}_h,w_h):=a_{\sigma p}(w_h,\boldsymbol{\tau}_h), a σ σ ( τ h , ψ h ) := ( A τ h , ψ h ) , a σ p ( r h , ψ h ) := ( α B A ( r h I ) , ψ h ) , a p σ ( τ h , w h ) := a σ p ( w h , τ h ) ,
a p p ( r h , w h ) : = c 0 ( r h , w h ) + α B 2 ( A ( r h I ) , w h I ) , a z z ( z h , ξ h ) : = ( K − 1 z h , ξ h ) . a_{pp}(r_h,w_h):=c_0(r_h,w_h)+\alpha_B^2(\mathcal{A}(r_hI),w_hI),\qquad a_{zz}(\boldsymbol{z}_h,\boldsymbol{\xi}_h):=(K^{-1}\boldsymbol{z}_h,\boldsymbol{\xi}_h). a pp ( r h , w h ) := c 0 ( r h , w h ) + α B 2 ( A ( r h I ) , w h I ) , a z z ( z h , ξ h ) := ( K − 1 z h , ξ h ) .
The stress–displacement and Darcy–pressure couplings are
b σ u ( v h , ψ h ) : = ∑ F ∈ F d l ( v h , ⟦ ψ h n ⟧ ) F , b u σ ( τ h , v h ) : = ∑ F ∈ F p r 1 / 2 ( τ h n , ⟦ v h ⟧ ) F , b_{\sigma u}(\boldsymbol{v}_h,\boldsymbol{\psi}_h):=\sum_{F\in\mathcal{F}_{\mathrm{dl}}}(\boldsymbol{v}_h,\llbracket\boldsymbol{\psi}_h\boldsymbol{n}\rrbracket)_F,\qquad b_{u\sigma}(\boldsymbol{\tau}_h,\boldsymbol{v}_h):=\sum_{F\in\mathcal{F}_{\mathrm{pr}}^{1/2}}(\boldsymbol{\tau}_h\boldsymbol{n},\llbracket\boldsymbol{v}_h\rrbracket)_F, b σ u ( v h , ψ h ) := F ∈ F dl ∑ ( v h , [ [ ψ h n ] ] ) F , b u σ ( τ h , v h ) := F ∈ F pr 1/2 ∑ ( τ h n , [ [ v h ] ] ) F ,
b z p ( r h , ξ h ) : = ∑ F ∈ F d l ( r h , ⟦ ξ h ⋅ n ⟧ ) F , b p z ( z h , w h ) : = ∑ F ∈ F p r 1 / 2 ( z h ⋅ n , ⟦ w h ⟧ ) F . b_{zp}(r_h,\boldsymbol{\xi}_h):=\sum_{F\in\mathcal{F}_{\mathrm{dl}}}(r_h,\llbracket\boldsymbol{\xi}_h\cdot\boldsymbol{n}\rrbracket)_F,\qquad b_{pz}(\boldsymbol{z}_h,w_h):=\sum_{F\in\mathcal{F}_{\mathrm{pr}}^{1/2}}(\boldsymbol{z}_h\cdot\boldsymbol{n},\llbracket w_h\rrbracket)_F. b z p ( r h , ξ h ) := F ∈ F dl ∑ ( r h , [ [ ξ h ⋅ n ] ] ) F , b p z ( z h , w h ) := F ∈ F pr 1/2 ∑ ( z h ⋅ n , [ [ w h ] ] ) F .
The weak-symmetry forms are c σ γ ( η h , ψ h ) : = ( η h , as ( ψ h ) ) c_{\sigma\gamma}(\eta_h,\psi_h):=(\eta_h,\operatorname{as}(\psi_h)) c σ γ ( η h , ψ h ) := ( η h , as ( ψ h )) and c γ σ ( τ h , η h ) : = ( as ( τ h ) , η h ) c_{\gamma\sigma}(\tau_h,\eta_h):=(\operatorname{as}(\tau_h),\eta_h) c γ σ ( τ h , η h ) := ( as ( τ h ) , η h ) .
Let F M i n t \mathcal{F}^{\mathrm{int}}_M F M int denote the set of facets shared by two adjacent macrocells. For F = ∂ M + ∩ ∂ M − ∈ F M i n t F=\partial M^+\cap\partial M^-\in\mathcal{F}^{\mathrm{int}}_M F = ∂ M + ∩ ∂ M − ∈ F M int , fix the unit normal n F \mathbf{n}_F n F from M + M^+ M + to M − M^- M − and define ⟦ ϕ ⟧ = ϕ + − ϕ − \llbracket\phi\rrbracket=\phi^+-\phi^- [ [ ϕ ] ] = ϕ + − ϕ − using the traces from M ± M^\pm M ± . For any macro-facet or subfacet F F F , set h F : = diam F h_F:=\operatorname{diam}F h F := diam F . In both 2D and 3D, we use full-jump rotation stabilization, penalizing all rotation components on every interior macro-facet. Define
j γ ( η h , ξ h ) : = ∑ F ∈ F M i n t h F [ [ η h ] ] F ⋅ [ [ ξ h ] ] F , ∣ η h ∣ γ , h 2 : = j γ ( η h , η h ) . j_\gamma(\eta_h,\xi_h):=\sum_{F\in\mathcal{F}^{\mathrm{int}}_M}h_F\lbrack\![\eta_h]\!\rbrack_F\cdot\lbrack\![\xi_h]\!\rbrack_F,\qquad|\eta_h|_{\gamma,h}^2:=j_\gamma(\eta_h,\eta_h). j γ ( η h , ξ h ) := F ∈ F M int ∑ h F [ [ η h ] ] F ⋅ [ [ ξ h ] ] F , ∣ η h ∣ γ , h 2 := j γ ( η h , η h ) .
and, for a fixed parameter δ γ > 0 \delta_\gamma>0 δ γ > 0 , set s γ ( η h , ξ h ) : = δ γ μ j γ ( η h , ξ h ) s_\gamma(\eta_h,\xi_h):=\delta_\gamma\mu\,j_\gamma(\eta_h,\xi_h) s γ ( η h , ξ h ) := δ γ μ j γ ( η h , ξ h ) . The stabilization is symmetric and positive semidefinite. Integration by parts gives
b σ u ( v h , τ h ) = − b u σ ( τ h , v h ) , b z p ( w h , ξ h ) = − b p z ( ξ h , w h ) , c σ γ ( η h , τ h ) = c γ σ ( τ h , η h ) . (4) b_{\sigma u}(v_h,\tau_h)=-b_{u\sigma}(\tau_h,v_h),\qquad b_{zp}(w_h,\xi_h)=-b_{pz}(\xi_h,w_h),\qquad c_{\sigma\gamma}(\eta_h,\tau_h)=c_{\gamma\sigma}(\tau_h,\eta_h). \tag*{(4)} b σ u ( v h , τ h ) = − b u σ ( τ h , v h ) , b z p ( w h , ξ h ) = − b p z ( ξ h , w h ) , c σ γ ( η h , τ h ) = c γ σ ( τ h , η h ) . (4)
Let t n = n Δ t t^n=n\Delta t t n = n Δ t , and introduce the scaled Darcy variable z ^ n : = Δ t z n \widehat{z}^n:=\Delta t\,z^n z n := Δ t z n . Given ( σ h n − 1 , p h n − 1 ) (\sigma_h^{n-1},p_h^{n-1}) ( σ h n − 1 , p h n − 1 ) , one backward Euler step seeks ( σ h n , u h n , γ h n , z ^ h n , p h n ) ∈ Σ h × U h × Γ h × Z h × P h (\sigma_h^n,u_h^n,\gamma_h^n,\widehat{z}_h^n,p_h^n)\in\Sigma_h\times U_h\times\Gamma_h\times Z_h\times P_h ( σ h n , u h n , γ h n , z h n , p h n ) ∈ Σ h × U h × Γ h × Z h × P h such that, for all ( ψ h , v h , η h , ξ h , w h ) ∈ Σ h × U h × Γ h × Z h × P h (\psi_h,v_h,\eta_h,\xi_h,w_h)\in\Sigma_h\times U_h\times\Gamma_h\times Z_h\times P_h ( ψ h , v h , η h , ξ h , w h ) ∈ Σ h × U h × Γ h × Z h × P h ,
a σ σ ( σ h n , ψ h ) + a σ p ( p h n , ψ h ) − b σ u ( u h n , ψ h ) + c σ γ ( γ h n , ψ h ) = 0 , (5) a_{\sigma\sigma}(\sigma_h^n,\psi_h)+a_{\sigma p}(p_h^n,\psi_h)-b_{\sigma u}(u_h^n,\psi_h)+c_{\sigma\gamma}(\gamma_h^n,\psi_h)=0, \tag*{(5)} a σ σ ( σ h n , ψ h ) + a σ p ( p h n , ψ h ) − b σ u ( u h n , ψ h ) + c σ γ ( γ h n , ψ h ) = 0 , (5)
− b u σ ( σ h n , v h ) = ( f n , v h ) , (6) -b_{u\sigma}(\sigma_h^n,v_h)=(f^n,v_h), \tag*{(6)} − b u σ ( σ h n , v h ) = ( f n , v h ) , (6)
c γ σ ( σ h n , η h ) − s γ ( γ h n , η h ) = 0 , (7) c_{\gamma\sigma}(\sigma_h^n,\eta_h)-s_\gamma(\gamma_h^n,\eta_h)=0, \tag*{(7)} c γ σ ( σ h n , η h ) − s γ ( γ h n , η h ) = 0 , (7)
( ( Δ t K ) − 1 z ^ h n , ξ h ) + b z p ( p h n , ξ h ) = 0 , (8) ((\Delta t K)^{-1}\widehat{z}_h^n,\xi_h)+b_{zp}(p_h^n,\xi_h)=0, \tag*{(8)} (( Δ t K ) − 1 z h n , ξ h ) + b z p ( p h n , ξ h ) = 0 , (8)
a p σ ( σ h n − σ h n − 1 , w h ) + a p p ( p h n − p h n − 1 , w h ) + b p z ( z ^ h n , w h ) = Δ t ( q n , w h ) . (9) a_{p\sigma}(\sigma_h^n-\sigma_h^{n-1},w_h)+a_{pp}(p_h^n-p_h^{n-1},w_h)+b_{pz}(\widehat{z}_h^n,w_h)=\Delta t(q^n,w_h). \tag*{(9)} a p σ ( σ h n − σ h n − 1 , w h ) + a pp ( p h n − p h n − 1 , w h ) + b p z ( z h n , w h ) = Δ t ( q n , w h ) . (9)
Notably, the rotation-jump stabilization in the weak-symmetry equation (7) enables a mesh-uniform elasticity inf–sup estimate and yields a condensed rotation block uniformly spectrally equivalent to a μ \mu μ -scaled macrocell mass matrix. Taking v h = e i χ M v_h=e_i\chi_M v h = e i χ M , i = 1 , … , d i=1,\ldots,d i = 1 , … , d , and w h = χ M w_h=\chi_M w h = χ M in (6) and (9) , respectively, yields local momentum balance and fluid-mass conservation on each M ∈ T M M\in\mathcal{T}_M M ∈ T M . Here e i e_i e i is the i i i -th coordinate unit vector and χ M \chi_M χ M is the indicator function of M M M .
Five-field and condensed three-field matrices
Let { Ψ i } \{\Psi_i\} { Ψ i } , { V j } \{V_j\} { V j } , { χ k } \{\chi_k\} { χ k } , { ξ ℓ } \{\xi_\ell\} { ξ ℓ } , and { ϕ m } \{\phi_m\} { ϕ m } be bases of Σ h \Sigma_h Σ h , U h U_h U h , Γ h \Gamma_h Γ h , Z h Z_h Z h , and P h P_h P h , respectively. With the coefficient ordering X = ( σ , u , γ , z ^ , p ) ⊤ X=(\sigma,u,\gamma,\widehat{z},p)^\top X = ( σ , u , γ , z , p ) ⊤ , define
A i j : = a σ σ ( Ψ j , Ψ i ) , B j i : = b u σ ( Ψ i , V j ) , C k i : = c γ σ ( Ψ i , χ k ) , D m ℓ : = b p z ( ξ ℓ , ϕ m ) , E i m : = a σ p ( ϕ m , Ψ i ) , F ℓ r : = ( ( Δ t K ) − 1 ξ r , ξ ℓ ) , ( A p p ) m r : = α B 2 ( A ( ϕ r I ) , ϕ m I ) , ( M 0 ) m r : = ( ϕ r , ϕ m ) , M p : = c 0 M 0 , M : = M p + A p p . (10) \begin{aligned}
A_{ij}&:=a_{\sigma\sigma}(\Psi_j,\Psi_i),\quad B_{ji}:=b_{u\sigma}(\Psi_i,V_j),\quad C_{ki}:=c_{\gamma\sigma}(\Psi_i,\chi_k),\quad D_{m\ell}:=b_{pz}(\xi_\ell,\phi_m),\\
E_{im}&:=a_{\sigma p}(\phi_m,\Psi_i),\quad F_{\ell r}:=((\Delta t K)^{-1}\xi_r,\xi_\ell),\quad(A_{pp})_{mr}:=\alpha_B^2(\mathcal{A}(\phi_r I),\phi_m I),\\
(M_0)_{mr}&:=(\phi_r,\phi_m),\quad M_p:=c_0M_0,\quad M:=M_p+A_{pp}. \tag*{(10)}
\end{aligned} A ij E im ( M 0 ) m r := a σ σ ( Ψ j , Ψ i ) , B j i := b u σ ( Ψ i , V j ) , C k i := c γ σ ( Ψ i , χ k ) , D m ℓ := b p z ( ξ ℓ , ϕ m ) , := a σ p ( ϕ m , Ψ i ) , F ℓ r := (( Δ t K ) − 1 ξ r , ξ ℓ ) , ( A pp ) m r := α B 2 ( A ( ϕ r I ) , ϕ m I ) , := ( ϕ r , ϕ m ) , M p := c 0 M 0 , M := M p + A pp . (10)
Let J γ J_\gamma J γ and M γ M_\gamma M γ denote the rotation stabilization and mass matrices, respectively:
( J γ ) k ℓ : = s γ ( χ ℓ , χ k ) , ( M γ ) k ℓ : = ( χ ℓ , χ k ) . (11) (J_\gamma)_{k\ell}:=s_\gamma(\chi_\ell,\chi_k),\qquad(M_\gamma)_{k\ell}:=(\chi_\ell,\chi_k). \tag*{(11)} ( J γ ) k ℓ := s γ ( χ ℓ , χ k ) , ( M γ ) k ℓ := ( χ ℓ , χ k ) . (11)
For isotropic elasticity, A p p = c λ M 0 A_{pp}=c_\lambda M_0 A pp = c λ M 0 , where c λ : = α B 2 / K d r c_\lambda:=\alpha_B^2/K_{\mathrm{dr}} c λ := α B 2 / K dr and K d r : = λ + 2 μ / d K_{\mathrm{dr}}:=\lambda+2\mu/d K dr := λ + 2 μ / d . Here K d r K_{\mathrm{dr}} K dr is the drained bulk modulus. From (4) , the five-field system takes the form
\begin{bmatrix}
A & B^\top& C^\top& 0 & E\\
B & 0 & 0 & 0 & 0\\
C & 0 & -J_\gamma& 0 & 0\\
0 & 0 & 0 & -F & D^\top\\
E^\top& 0 & 0 & D & M
\end{bmatrix}
\begin{bmatrix}
\sigma\\
u\\
\gamma\\
\widehat{z}\\
p
\end{bmatrix}
=
\begin{bmatrix}
0\\
-F_u\\
0\\
0\\
r_p
\end{bmatrix}. \tag*{(12)}\end{bmatrix}
where ( F u ) j = ( f n , V j ) (F_u)_j=(f^{n},V_j) ( F u ) j = ( f n , V j ) and ( r p ) m = a σ p ( σ h n − 1 , ϕ m ) + a p p ( p h n − 1 , ϕ m ) + Δ t ( q n , ϕ m ) (r_p)_m=a_{\sigma p}(\sigma_h^{n-1},\phi_m)+a_{pp}(p_h^{n-1},\phi_m)+\Delta t(q^n,\phi_m) ( r p ) m = a σ p ( σ h n − 1 , ϕ m ) + a pp ( p h n − 1 , ϕ m ) + Δ t ( q n , ϕ m ) . The diagonal block − J γ -J_\gamma − J γ is the matrix representation of − s γ ( γ h , η h ) -s_\gamma(\gamma_h,\eta_h) − s γ ( γ h , η h ) .
Since A and F are symmetric positive definite and block diagonal over the interaction regions, their inverses are applied locally. Eliminating the stress and scaled Darcy flux gives
σ = − A − 1 ( B T u + C T γ + E p ) , z ^ = F − 1 D T p . (13) \sigma=-A^{-1}(B^{\mathsf T}u+C^{\mathsf T}\gamma+Ep),\qquad\widehat{z}=F^{-1}D^{\mathsf T}p. \tag*{(13)} σ = − A − 1 ( B T u + C T γ + E p ) , z = F − 1 D T p . (13)
Define the reduced blocks
K u u : = B A − 1 B T , K u γ : = B A − 1 C T , K u p : = B A − 1 E , K γ γ 0 : = C A − 1 C T , K γ γ : = K γ γ 0 + J γ , K γ p : = C A − 1 E , S p σ : = E T A − 1 E , S p z : = D F − 1 D T , T p : = M − S p σ + S p z . (14) \begin{aligned}
K_{uu}&:=BA^{-1}B^{\mathsf T},\qquad K_{u\gamma}:=BA^{-1}C^{\mathsf T},\qquad K_{up}:=BA^{-1}E,\\
K_{\gamma\gamma}^{0}&:=CA^{-1}C^{\mathsf T},\qquad K_{\gamma\gamma}:=K_{\gamma\gamma}^{0}+J_\gamma,\qquad K_{\gamma p}:=CA^{-1}E,\\
S_p^\sigma&:=E^{\mathsf T}A^{-1}E,\qquad S_p^z:=DF^{-1}D^{\mathsf T},\qquad T_p:=M-S_p^\sigma+S_p^z. \tag*{(14)}
\end{aligned} K uu K γ γ 0 S p σ := B A − 1 B T , K u γ := B A − 1 C T , K u p := B A − 1 E , := C A − 1 C T , K γ γ := K γ γ 0 + J γ , K γ p := C A − 1 E , := E T A − 1 E , S p z := D F − 1 D T , T p := M − S p σ + S p z . (14)
The pressure block T p T_p T p is positive semidefinite. Indeed, M p = c 0 M 0 ⪰ 0 M_p=c_0M_0\succeq0 M p = c 0 M 0 ⪰ 0 and S p z = D F − 1 D T ⪰ 0 S_p^z=DF^{-1}D^{\mathsf T}\succeq0 S p z = D F − 1 D T ⪰ 0 . Moreover, A p p − S p σ ⪰ 0 A_{pp}-S_p^\sigma\succeq0 A pp − S p σ ⪰ 0 , because it is the Schur complement with respect to A of the block matrix [ A E T E A p p ] \begin{bmatrix} A&E^{\mathsf T}\\ E&A_{pp}\end{bmatrix} [ A E E T A pp ] , whose quadratic form is ( A ( τ h + α B p h I ) , τ h + α B p h I ) (A(\tau_h+\alpha_Bp_hI),\tau_h+\alpha_Bp_hI) ( A ( τ h + α B p h I ) , τ h + α B p h I ) . Consequently, T p = M p + ( A p p − S p σ ) + S p z ⪰ 0 T_p=M_p+(A_{pp}-S_p^\sigma)+S_p^z\succeq0 T p = M p + ( A pp − S p σ ) + S p z ⪰ 0 . Substituting the local elimination formulas (13) for σ \sigma σ and z ^ \widehat{z} z into (12) gives the condensed displacement–rotation–pressure system
[ − K u u − K u γ − K u p − K u γ T − K γ γ − K γ p − K u p T − K γ p T T p ] [ u γ p ] = [ − F u 0 r p ] . \begin{bmatrix}
-K_{uu}&-K_{u\gamma}&-K_{up}\\
-K_{u\gamma}^{\mathsf T}&-K_{\gamma\gamma}&-K_{\gamma p}\\
-K_{up}^{\mathsf T}&-K_{\gamma p}^{\mathsf T}&T_p
\end{bmatrix}
\begin{bmatrix}u\\ \gamma\\ p\end{bmatrix}
=
\begin{bmatrix}-F_u\\ 0\\ r_p\end{bmatrix}. − K uu − K u γ T − K u p T − K u γ − K γ γ − K γ p T − K u p − K γ p T p u γ p = − F u 0 r p .
Collecting the two mechanics variables in m = ( u , γ ) T m=(u,\gamma)^{\mathsf T} m = ( u , γ ) T , set
K m 0 : = [ K u u K u γ K u γ T K γ γ 0 ] , K m : = K m 0 + diag ( 0 , J γ ) , D m : = diag ( K u u , K γ γ ) . (15) K_m^0:=\begin{bmatrix}K_{uu}&K_{u\gamma}\\ K_{u\gamma}^{\mathsf T}&K_{\gamma\gamma}^{0}\end{bmatrix},\qquad K_m:=K_m^0+\operatorname{diag}(0,J_\gamma),\qquad D_m:=\operatorname{diag}(K_{uu},K_{\gamma\gamma}). \tag*{(15)} K m 0 := [ K uu K u γ T K u γ K γ γ 0 ] , K m := K m 0 + diag ( 0 , J γ ) , D m := diag ( K uu , K γ γ ) . (15)
Thus the condensed matrix is
A r e d = [ − K m − L − L T T p ] with L = [ K u p K γ p ] . (16) A_{\mathrm{red}}=\begin{bmatrix}-K_m&-L\\ -L^{\mathsf T}&T_p\end{bmatrix}
\qquad\text{with}\qquad
L=\begin{bmatrix}K_{up}\\ K_{\gamma p}\end{bmatrix}. \tag*{(16)} A red = [ − K m − L T − L T p ] with L = [ K u p K γ p ] . (16)
Well-posedness of the fully discrete scheme
The following two mesh-uniform stability estimates are used to prove well-posedness of the fully discrete scheme. We use the following discrete norms:
∥ ϕ h ∥ 1 , h 2 : = ∑ F ∈ F p r 1 / 2 h F − 1 ∥ ⟦ ϕ h ⟧ ∥ L 2 ( F ) 2 , ∥ ( v h , η h ) ∥ h 2 : = ∥ v h ∥ 1 , h 2 + ∥ η h ∥ L 2 ( Ω ) 2 . (17) \lVert\phi_h\rVert_{1,h}^{2}:=\sum_{F\in\mathcal{F}_{pr}^{1/2}}h_F^{-1}\lVert\llbracket\phi_h\rrbracket\rVert_{L^2(F)}^{2},\qquad\lVert(v_h,\eta_h)\rVert_h^{2}:=\lVert v_h\rVert_{1,h}^{2}+\lVert\eta_h\rVert_{L^2(\Omega)}^{2}. \tag*{(17)} ∥ ϕ h ∥ 1 , h 2 := F ∈ F p r 1/2 ∑ h F − 1 ∥ [ [ ϕ h ] ] ∥ L 2 ( F ) 2 , ∥( v h , η h ) ∥ h 2 := ∥ v h ∥ 1 , h 2 + ∥ η h ∥ L 2 ( Ω ) 2 . (17)
Boundary jumps in (17) incorporate the homogeneous Dirichlet conditions.
For the stability analysis below, consider a uniformly shape-regular family of macrocell and subcell meshes. In 2D, the subcells in each quadrilateral macrocell M M M are O ( h M 2 ) O(h_M^2) O ( h M 2 ) perturbations of parallelograms [17 ] . In 3D, we consider Cartesian macrocells, each divided into eight corner cuboids by its coordinate midplanes. Here h M = diam M h_M=\operatorname{diam}M h M = diam M and h = max M h M h=\max_M h_M h = max M h M .
Lemma 2.1. (Darcy inf-sup condition). There exists c D > 0 c_D>0 c D > 0 , depending only on the shape regularity of the mesh family, such that
sup 0 ≠ ξ h ∈ Z h b z p ( w h , ξ h ) ∥ ξ h ∥ L 2 ( Ω ) ≥ c D ∥ w h ∥ 1 , h ∀ w h ∈ P h . (18) \sup_{0\ne\xi_h\in Z_h}\frac{b_{zp}(w_h,\xi_h)}{\lVert\xi_h\rVert_{L^2(\Omega)}}\ge c_D\lVert w_h\rVert_{1,h}\qquad\forall w_h\in P_h. \tag*{(18)} 0 = ξ h ∈ Z h sup ∥ ξ h ∥ L 2 ( Ω ) b z p ( w h , ξ h ) ≥ c D ∥ w h ∥ 1 , h ∀ w h ∈ P h . (18)
The Darcy inf–sup estimate follows directly from [17 eqs. (4.14)–(4.15)] , since Σ h = [ Z h ] d \Sigma_h = [Z_h]^d Σ h = [ Z h ] d and U h = [ P h ] d U_h = [P_h]^d U h = [ P h ] d , we therefore omit the detailed proof.
Lemma 2.2 (Elasticity inf-sup condition). There exist h 0 > 0 h_0 > 0 h 0 > 0 and c E > 0 c_E > 0 c E > 0 , independent of h h h , λ \lambda λ , and μ \mu μ , such that, for all 0 < h ≤ h 0 0 < h \le h_0 0 < h ≤ h 0 ,
( sup 0 ≠ τ h ∈ Σ h b u σ ( τ h , v h ) + c γ σ ( τ h , η h ) ∥ τ h ∥ L 2 ( Ω ) ) 2 + ∣ η h ∣ γ , h 2 ≥ c E 2 ∥ ( v h , η h ) ∥ h 2 (19) \left(\sup_{0 \ne\tau_h \in\Sigma_h} \frac{b_{u\sigma}(\tau_h,v_h)+c_{\gamma\sigma}(\tau_h,\eta_h)}{\|\tau_h\|_{L^2(\Omega)}}\right)^2 + |\eta_h|_{\gamma,h}^2 \ge c_E^2\|(v_h,\eta_h)\|_h^2 \tag*{(19)} ( 0 = τ h ∈ Σ h sup ∥ τ h ∥ L 2 ( Ω ) b u σ ( τ h , v h ) + c γ σ ( τ h , η h ) ) 2 + ∣ η h ∣ γ , h 2 ≥ c E 2 ∥ ( v h , η h ) ∥ h 2 (19) for all ( v h , η h ) ∈ U h × Γ h (v_h,\eta_h) \in U_h \times\Gamma_h ( v h , η h ) ∈ U h × Γ h .
Proof. Set K h : = { τ h ∈ Σ h : b u σ ( τ h , w h ) = 0 , ∀ w h ∈ U h } K_h := \{\tau_h \in\Sigma_h : b_{u\sigma}(\tau_h,w_h) = 0,\ \forall w_h \in U_h\} K h := { τ h ∈ Σ h : b u σ ( τ h , w h ) = 0 , ∀ w h ∈ U h } . In both dimensions we establish a stabilized divergence estimate and reconstruct a stress in K h K_h K h . Invoking [17 eqs. (4.14)–(4.15)] , we have
∥ v h ∥ 1 , h ≤ C sup 0 ≠ τ h ∈ Σ h ∣ b u σ ( τ h , v h ) ∣ ∥ τ h ∥ L 2 ( Ω ) for any v h ∈ U h . (20) \|v_h\|_{1,h} \le C \sup_{0 \ne\tau_h \in\Sigma_h} \frac{|b_{u\sigma}(\tau_h,v_h)|}{\|\tau_h\|_{L^2(\Omega)}} \quad\text{for any } v_h \in U_h. \tag*{(20)} ∥ v h ∥ 1 , h ≤ C 0 = τ h ∈ Σ h sup ∥ τ h ∥ L 2 ( Ω ) ∣ b u σ ( τ h , v h ) ∣ for any v h ∈ U h . (20)
We next estimate the rotation component. We begin with 2D. For every η h ∈ Γ h \eta_h \in\Gamma_h η h ∈ Γ h , we prove
∥ η h ∥ L 2 ( Ω ) ≤ C ( sup 0 ≠ τ h ∈ K h ∣ c γ σ ( τ h , η h ) ∣ ∥ τ h ∥ L 2 ( Ω ) + ∣ η h ∣ γ , h ) . (21) \|\eta_h\|_{L^2(\Omega)} \le C\left(\sup_{0 \ne\tau_h \in K_h} \frac{|c_{\gamma\sigma}(\tau_h,\eta_h)|}{\|\tau_h\|_{L^2(\Omega)}} + |\eta_h|_{\gamma,h}\right). \tag*{(21)} ∥ η h ∥ L 2 ( Ω ) ≤ C ( 0 = τ h ∈ K h sup ∥ τ h ∥ L 2 ( Ω ) ∣ c γ σ ( τ h , η h ) ∣ + ∣ η h ∣ γ , h ) . (21)
To prove this claim, let Q h : = { ρ h ∈ H 0 1 ( Ω ) 2 : ρ h ∣ M ∘ F M ∈ [ Q 1 ( M ^ ) ] 2 , ∀ M ∈ T M } Q_h := \{\rho_h \in H^1_0(\Omega)^2 : \rho_h|_M \circ F_M \in[Q_1(\widehat{M})]^2,\ \forall M \in\mathcal{T}_M\} Q h := { ρ h ∈ H 0 1 ( Ω ) 2 : ρ h ∣ M ∘ F M ∈ [ Q 1 ( M ) ] 2 , ∀ M ∈ T M } be the continuous bilinear quadrilateral finite element space on the macro-mesh, where F M : M ^ → M F_M : \widehat{M} \to M F M : M → M is the bilinear element map from the reference square M ^ \widehat{M} M . We first show that, for q h ∈ P 0 ( T M ) q_h \in P_0(\mathcal{T}_M) q h ∈ P 0 ( T M ) satisfying ( q h , 1 ) = 0 (q_h,1)=0 ( q h , 1 ) = 0 ,
∥ q h ∥ L 2 ( Ω ) ≤ C ( sup 0 ≠ ρ h ∈ Q h ∣ ( ∇ ⋅ ρ h , q h ) ∣ ∥ ρ h ∥ H 1 ( Ω ) + ∣ q h ∣ γ , h ) . (22) \|q_h\|_{L^2(\Omega)} \le C\left(\sup_{0 \ne\rho_h \in Q_h} \frac{|(\nabla\cdot\rho_h,q_h)|}{\|\rho_h\|_{H^1(\Omega)}} + |q_h|_{\gamma,h}\right). \tag*{(22)} ∥ q h ∥ L 2 ( Ω ) ≤ C ( 0 = ρ h ∈ Q h sup ∥ ρ h ∥ H 1 ( Ω ) ∣ ( ∇ ⋅ ρ h , q h ) ∣ + ∣ q h ∣ γ , h ) . (22)
Choose ρ ∈ H 0 1 ( Ω ) 2 \rho\in H^1_0(\Omega)^2 ρ ∈ H 0 1 ( Ω ) 2 with ∇ ⋅ ρ = q h \nabla\cdot\rho= q_h ∇ ⋅ ρ = q h and ∥ ρ ∥ H 1 ( Ω ) ≤ C ∥ q h ∥ L 2 ( Ω ) \|\rho\|_{H^1(\Omega)} \le C\|q_h\|_{L^2(\Omega)} ∥ ρ ∥ H 1 ( Ω ) ≤ C ∥ q h ∥ L 2 ( Ω ) , and let I h ρ ∈ Q h I_h\rho\in Q_h I h ρ ∈ Q h be a boundary-preserving Scott–Zhang-type quasi-interpolant. Such an operator is obtained from the local regularization construction for quadrilateral elements in [9 ] , using the boundary-preserving treatment of [35 ] . The local H 1 H^1 H 1 -stability and first-order L 2 L^2 L 2 -approximation properties of this operator, together with the scaled trace inequality and the uniformly bounded overlap of element patches implies ∑ F ∈ F M i n t h F − 1 ∥ ρ − I h ρ ∥ L 2 ( F ) 2 ≤ C ∥ ρ ∥ H 1 ( Ω ) 2 \sum_{F\in\mathcal{F}^{\mathrm{int}}_M} h_F^{-1}\|\rho-I_h\rho\|_{L^2(F)}^2 \le C\|\rho\|_{H^1(\Omega)}^2 ∑ F ∈ F M int h F − 1 ∥ ρ − I h ρ ∥ L 2 ( F ) 2 ≤ C ∥ ρ ∥ H 1 ( Ω ) 2 . Set e = ρ − I h ρ e=\rho-I_h\rho e = ρ − I h ρ . Since q h q_h q h is macrocellwise constant and e ∣ ∂ Ω = 0 e|_{\partial\Omega}=0 e ∣ ∂ Ω = 0 , integration by parts and the Cauchy–Schwarz inequality give
∥ q h ∥ L 2 ( Ω ) 2 ≤ ∣ ( q h , ∇ ⋅ I h ρ ) ∣ + ∑ F ∈ F M i n t ∣ ( ⟦ q h ⟧ , e ⋅ n F ) F ∣ ≤ ∣ ( q h , ∇ ⋅ I h ρ ) ∣ + ∣ q h ∣ γ , h ( ∑ F ∈ F M i n t h F − 1 ∥ e ∥ L 2 ( F ) 2 ) 1 / 2 ≤ ∣ ( q h , ∇ ⋅ I h ρ ) ∣ + C ∣ q h ∣ γ , h ∥ q h ∥ L 2 ( Ω ) , \begin{aligned}
\|q_h\|_{L^2(\Omega)}^2
&\le|(q_h,\nabla\cdot I_h\rho)|+\sum_{F\in\mathcal{F}^{\mathrm{int}}_M}|(\llbracket q_h\rrbracket,e\cdot n_F)_F|\\
&\le|(q_h,\nabla\cdot I_h\rho)|+|q_h|_{\gamma,h}\left(\sum_{F\in\mathcal{F}^{\mathrm{int}}_M}h_F^{-1}\|e\|_{L^2(F)}^2\right)^{1/2}\\
&\le|(q_h,\nabla\cdot I_h\rho)|+C|q_h|_{\gamma,h}\|q_h\|_{L^2(\Omega)},
\end{aligned} ∥ q h ∥ L 2 ( Ω ) 2 ≤ ∣ ( q h , ∇ ⋅ I h ρ ) ∣ + F ∈ F M int ∑ ∣ ( [ [ q h ] ] , e ⋅ n F ) F ∣ ≤ ∣ ( q h , ∇ ⋅ I h ρ ) ∣ + ∣ q h ∣ γ , h F ∈ F M int ∑ h F − 1 ∥ e ∥ L 2 ( F ) 2 1/2 ≤ ∣ ( q h , ∇ ⋅ I h ρ ) ∣ + C ∣ q h ∣ γ , h ∥ q h ∥ L 2 ( Ω ) ,
which proves (2.18).
We next prove (2.17) using (2.18). Write ∇ ⊥ r : = ( ∂ y r , − ∂ x r ) \nabla^\perp r := (\partial_y r,-\partial_x r) ∇ ⊥ r := ( ∂ y r , − ∂ x r ) for a scalar r r r . For ρ h ∈ Q h \rho_h \in Q_h ρ h ∈ Q h , define R h ρ h ∈ Σ h R_h\rho_h \in\Sigma_h R h ρ h ∈ Σ h through its row-wise normal traces on the primal subfacets:
∫ F ( R h ρ h ) i ⋅ n d s = ∫ F ∇ ⊥ ( ρ h ) i ⋅ n d s , i = 1 , 2 , F ∈ F p r 1 / 2 . (23) \int_F (R_h\rho_h)_i\cdot n\,\mathrm{d}s = \int_F \nabla^\perp(\rho_h)_i\cdot n\,\mathrm{d}s,\qquad i=1,2,\quad F\in\mathcal{F}_{\mathrm{pr}}^{1/2}. \tag*{(23)} ∫ F ( R h ρ h ) i ⋅ n d s = ∫ F ∇ ⊥ ( ρ h ) i ⋅ n d s , i = 1 , 2 , F ∈ F pr 1/2 . (23)
We use one common value of the flux degree of freedom on each interior primal half-edge, with opposite orientations on the two adjacent subcells. Thus (2.19) defines a member of Σ h \Sigma_h Σ h directly in physical coordinates. On the other hand, standard scaling gives ∥ R h ρ h ∥ L 2 ( Ω ) ≤ C ∥ ρ h ∥ H 1 ( Ω ) \|R_h\rho_h\|_{L^2(\Omega)} \le C\|\rho_h\|_{H^1(\Omega)} ∥ R h ρ h ∥ L 2 ( Ω ) ≤ C ∥ ρ h ∥ H 1 ( Ω ) . Moreover, for any M ∈ T M M \in\mathcal{T}_M M ∈ T M and i = 1 , 2 i=1,2 i = 1 , 2 , we have ∫ ∂ M ( R h ρ h ) i ⋅ n M d s = 0 \int_{\partial M}(R_h\rho_h)_i\cdot n_M\,ds=0 ∫ ∂ M ( R h ρ h ) i ⋅ n M d s = 0 . Since every w h ∈ U h w_h \in U_h w h ∈ U h is constant on M M M , summation over the macrocells yields b u σ ( R h ρ h , w h ) = 0 b_u\sigma(R_h\rho_h,w_h)=0 b u σ ( R h ρ h , w h ) = 0 , so R h ρ h ∈ K h R_h\rho_h \in K_h R h ρ h ∈ K h .
On a parallelogram M M M with the center-to-midpoint subdivision, let K v K_v K v be the corner subcell at vertex v v v . Since the edge traces of ( ρ h ) i (\rho_h)_i ( ρ h ) i are affine, (2.19) matches the normal components of ( R h ρ h ) i ∣ K v (R_h\rho_h)_i|_{K_v} ( R h ρ h ) i ∣ K v and ∇ ⊥ ( ρ h ) i ( v ) \nabla^\perp(\rho_h)_i(v) ∇ ⊥ ( ρ h ) i ( v ) on the two adjacent primal half-edges. Their normals are linearly independent, so ( R h ρ h ) i ∣ K v = ∇ ⊥ ( ρ h ) i ( v ) (R_h\rho_h)_i|_{K_v}=\nabla^\perp(\rho_h)_i(v) ( R h ρ h ) i ∣ K v = ∇ ⊥ ( ρ h ) i ( v ) , i = 1 , 2 i=1,2 i = 1 , 2 . Thus as ( ( R h ρ h ) ∣ K v ) = ∇ ⋅ ρ h ( v ) \operatorname{as}((R_h\rho_h)|_{K_v})=\nabla\cdot\rho_h(v) as (( R h ρ h ) ∣ K v ) = ∇ ⋅ ρ h ( v ) . The divergence of a bilinear Q 1 Q_1 Q 1 function, pulled back by the affine map from the reference square, belongs to Q 1 Q_1 Q 1 . Since the four subcells have equal area, the tensor-product trapezoidal rule is exact for this function, and hence
∫ M as ( R h ρ h ) d x = ∫ M ∇ ⋅ ρ h d x . (24) \int_M \operatorname{as}(R_h\rho_h)\,dx=\int_M \nabla\cdot\rho_h\,dx. \tag*{(24)} ∫ M as ( R h ρ h ) d x = ∫ M ∇ ⋅ ρ h d x . (24)
For a perturbed quadrilateral M M M , the preceding quadrature argument is not exact. Let M 0 M^0 M 0 be its parallelogram counterpart, with ∣ v − v 0 ∣ ≤ C h M 2 |v-v^0|\le Ch_M^2 ∣ v − v 0 ∣ ≤ C h M 2 at corresponding vertices, and let ρ h 0 \rho_h^0 ρ h 0 have the same nodal values as ρ h \rho_h ρ h . Subtracting the common value at one vertex v ∗ v_* v ∗ and applying finite-dimensional scaling gives
∣ ρ h 0 ∣ H 1 ( M 0 ) + max v ∣ ρ h ( v ) − ρ h ( v ∗ ) ∣ ≤ C ∣ ρ h ∣ H 1 ( M ) . (25) |\rho_h^0|_{H^1(M^0)}+\max_v|\rho_h(v)-\rho_h(v_*)|\le C|\rho_h|_{H^1(M)}. \tag*{(25)} ∣ ρ h 0 ∣ H 1 ( M 0 ) + v max ∣ ρ h ( v ) − ρ h ( v ∗ ) ∣ ≤ C ∣ ρ h ∣ H 1 ( M ) . (25)
On a corner subcell K v K_v K v , let N v N_v N v contain the integrated normals of its two primal half-edges and d v , i d_{v,i} d v , i their rotated-gradient fluxes. Use superscript 0 for M 0 M^0 M 0 , and write r v , i r_{v,i} r v , i and r v , i 0 r^0_{v,i} r v , i 0 for the constant reconstructed rows. The affine edge traces have matching endpoint values, so d v , i = d v , i 0 d_{v,i}=d^0_{v,i} d v , i = d v , i 0 and N v T r v , i = d v , i = ( N v 0 ) T r v , i 0 N_v^{\mathsf T}r_{v,i}=d_{v,i}=(N_v^0)^{\mathsf T}r^0_{v,i} N v T r v , i = d v , i = ( N v 0 ) T r v , i 0 . Uniform shape regularity, (2.21), and the vertex perturbation give
∣ d v , i ∣ ≤ C ∣ ρ h ∣ H 1 ( M ) , ∥ N v − N v 0 ∥ 2 ≤ C h M 2 , ∥ ( N v T ) − 1 ∥ 2 + ∥ ( ( N v 0 ) T ) − 1 ∥ 2 ≤ C h M − 1 , |d_{v,i}|\le C|\rho_h|_{H^1(M)},\qquad\|N_v-N_v^0\|_2\le Ch_M^2,\qquad\|(N_v^{\mathsf T})^{-1}\|_2+\|((N_v^0)^{\mathsf T})^{-1}\|_2\le Ch_M^{-1}, ∣ d v , i ∣ ≤ C ∣ ρ h ∣ H 1 ( M ) , ∥ N v − N v 0 ∥ 2 ≤ C h M 2 , ∥ ( N v T ) − 1 ∥ 2 + ∥ (( N v 0 ) T ) − 1 ∥ 2 ≤ C h M − 1 ,
where ∥ ⋅ ∥ 2 \|\cdot\|_2 ∥ ⋅ ∥ 2 is the induced Euclidean matrix norm. Thus ∣ r v , i − r v , i 0 ∣ ≤ C ∣ ρ h ∣ H 1 ( M ) |r_{v,i}-r^0_{v,i}|\le C|\rho_h|_{H^1(M)} ∣ r v , i − r v , i 0 ∣ ≤ C ∣ ρ h ∣ H 1 ( M ) and ∣ K v ∣ − ∣ K v 0 ∣ ≤ C h M 3 \left|K_v\right|-\left|K_v^0\right|\le Ch_M^3 ∣ K v ∣ − K v 0 ≤ C h M 3 . Because the reconstructed stresses are constant on the corner subcells, their integrals are sums of subcell areas times the corresponding asymmetries. Using ∣ r v , i 0 ∣ ≤ C h M − 1 ∣ ρ h ∣ H 1 ( M ) |r^0_{v,i}|\le Ch_M^{-1}|\rho_h|_{H^1(M)} ∣ r v , i 0 ∣ ≤ C h M − 1 ∣ ρ h ∣ H 1 ( M ) , the row difference and area bounds above give
∣ ∫ M as ( R h ρ h ) d x − ∫ M 0 as ( R h 0 ρ h 0 ) d x ∣ ≤ C h M 2 ∣ ρ h ∣ H 1 ( M ) . \left|\int_M\operatorname{as}(R_h\rho_h)\,dx-\int_{M^0}\operatorname{as}(R_h^0\rho_h^0)\,dx\right|\le Ch_M^2|\rho_h|_{H^1(M)}. ∫ M as ( R h ρ h ) d x − ∫ M 0 as ( R h 0 ρ h 0 ) d x ≤ C h M 2 ∣ ρ h ∣ H 1 ( M ) .
The divergence theorem gives the same bound for the difference of the divergence integrals: subtract the common nodal value, use (2.21) for the remaining affine edge traces, and note that corresponding integrated edge normals differ by O ( h M 2 ) O(h_M^2) O ( h M 2 ) . Combining the two bounds with (2.20) on M 0 M^0 M 0 yields
∣ ∫ M ( as ( R h ρ h ) − ∇ ⋅ ρ h ) d x ∣ ≤ C h M 2 ∣ ρ h ∣ H 1 ( M ) . (26) \left|\int_M\left(\operatorname{as}(R_h\rho_h)-\nabla\cdot\rho_h\right)\,dx\right|\le Ch_M^2|\rho_h|_{H^1(M)}. \tag*{(26)} ∫ M ( as ( R h ρ h ) − ∇ ⋅ ρ h ) d x ≤ C h M 2 ∣ ρ h ∣ H 1 ( M ) . (26)
Since q h q_h q h is constant on each macrocell, summation of (2.22) and the Cauchy–Schwarz inequality yield the global defect explicitly as follows. Writing q M = q h ∣ M q_M=q_h|_M q M = q h ∣ M , using ∥ q h ∥ L 2 ( M ) ≃ h M ∣ q M ∣ \|q_h\|_{L^2(M)}\simeq h_M|q_M| ∥ q h ∥ L 2 ( M ) ≃ h M ∣ q M ∣ , we obtain
∣ c γ σ ( R h ρ h , q h ) − ( ∇ ⋅ ρ h , q h ) ∣ ≤ C ∑ M ∈ T M ∣ q M ∣ h M 2 ∣ ρ h ∣ H 1 ( M ) ≤ C h ∑ M ∈ T M ∥ q h ∥ L 2 ( M ) ∣ ρ h ∣ H 1 ( M ) ≤ C h ∥ q h ∥ L 2 ( Ω ) ∣ ρ h ∣ H 1 ( Ω ) . (27) \begin{aligned}
|c_{\gamma\sigma}(R_h\rho_h,q_h)-(\nabla\cdot\rho_h,q_h)|
&\le C\sum_{M\in\mathcal{T}_M}|q_M|h_M^2|\rho_h|_{H^1(M)}\\
&\le Ch\sum_{M\in\mathcal{T}_M}\|q_h\|_{L^2(M)}|\rho_h|_{H^1(M)}
\le Ch\|q_h\|_{L^2(\Omega)}|\rho_h|_{H^1(\Omega)}. \tag*{(27)}
\end{aligned} ∣ c γ σ ( R h ρ h , q h ) − ( ∇ ⋅ ρ h , q h ) ∣ ≤ C M ∈ T M ∑ ∣ q M ∣ h M 2 ∣ ρ h ∣ H 1 ( M ) ≤ C h M ∈ T M ∑ ∥ q h ∥ L 2 ( M ) ∣ ρ h ∣ H 1 ( M ) ≤ C h ∥ q h ∥ L 2 ( Ω ) ∣ ρ h ∣ H 1 ( Ω ) . (27)
Write η h = η h 0 + η ˉ h \eta_h=\eta_h^0+\bar{\eta}_h η h = η h 0 + η ˉ h , where η ˉ h = ∣ Ω ∣ − 1 ∫ Ω η h \bar{\eta}_h=|\Omega|^{-1}\int_\Omega\eta_h η ˉ h = ∣Ω ∣ − 1 ∫ Ω η h . Set S ( a ) = a 2 ( 0 − 1 1 0 ) S(a)=\frac{a}{2}\begin{pmatrix}0&-1\\1&0\end{pmatrix} S ( a ) = 2 a ( 0 1 − 1 0 ) and m h = ∣ Ω ∣ − 1 ∫ Ω as ( R h ρ h ) m_h=|\Omega|^{-1}\int_\Omega\operatorname{as}(\mathcal{R}_h\rho_h) m h = ∣Ω ∣ − 1 ∫ Ω as ( R h ρ h ) . The constant stress S ( m h ) ∈ K h S(m_h)\in K_h S ( m h ) ∈ K h , since its normal flux sums to zero on every macrocell. Thus Z h ρ h : = R h ρ h − S ( m h ) ∈ K h Z_h\rho_h:=\mathcal{R}_h\rho_h-S(m_h)\in K_h Z h ρ h := R h ρ h − S ( m h ) ∈ K h , with ∫ Ω as ( Z h ρ h ) d x = 0 \int_\Omega\operatorname{as}(Z_h\rho_h)\,dx=0 ∫ Ω as ( Z h ρ h ) d x = 0 and ∥ Z h ρ h ∥ L 2 ( Ω ) ≤ C ∥ ρ h ∥ H 1 ( Ω ) \|Z_h\rho_h\|_{L^2(\Omega)}\le C\|\rho_h\|_{H^1(\Omega)} ∥ Z h ρ h ∥ L 2 ( Ω ) ≤ C ∥ ρ h ∥ H 1 ( Ω ) . Since as S ( m h ) = m h \operatorname{as}S(m_h)=m_h as S ( m h ) = m h and ∫ Ω η h 0 = 0 \int_\Omega\eta_h^0=0 ∫ Ω η h 0 = 0 , c γ σ ( Z h ρ h , η h ) = c γ σ ( R h ρ h , η h 0 ) c_{\gamma\sigma}(Z_h\rho_h,\eta_h)=c_{\gamma\sigma}(\mathcal{R}_h\rho_h,\eta_h^0) c γ σ ( Z h ρ h , η h ) = c γ σ ( R h ρ h , η h 0 ) . Define S h ( η h ) : = sup 0 ≠ τ h ∈ K h ∣ c γ σ ( τ h , η h ) ∣ ∥ τ h ∥ L 2 ( Ω ) \mathcal{S}_h(\eta_h):=\sup_{0\ne\tau_h\in K_h}\frac{|c_{\gamma\sigma}(\tau_h,\eta_h)|}{\|\tau_h\|_{L^2(\Omega)}} S h ( η h ) := sup 0 = τ h ∈ K h ∥ τ h ∥ L 2 ( Ω ) ∣ c γ σ ( τ h , η h ) ∣ . Taking q h = η h 0 q_h=\eta_h^0 q h = η h 0 in (27) and using ∥ Z h ρ h ∥ L 2 ( Ω ) ≤ C ∥ ρ h ∥ H 1 ( Ω ) \|Z_h\rho_h\|_{L^2(\Omega)}\le C\|\rho_h\|_{H^1(\Omega)} ∥ Z h ρ h ∥ L 2 ( Ω ) ≤ C ∥ ρ h ∥ H 1 ( Ω ) , we obtain, for every ρ h ∈ Q h \rho_h\in Q_h ρ h ∈ Q h ,
∣ ( ∇ ⋅ ρ h , η h 0 ) ∣ ≤ ∣ c γ σ ( Z h ρ h , η h ) ∣ + C h ∥ ρ h ∥ H 1 ( Ω ) ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + h ∥ η h 0 ∥ L 2 ( Ω ) ) ∥ ρ h ∥ H 1 ( Ω ) . \left|(\nabla\cdot\rho_h,\eta_h^0)\right|\le|c_{\gamma\sigma}(Z_h\rho_h,\eta_h)|+Ch\|\rho_h\|_{H^1(\Omega)}\|\eta_h^0\|_{L^2(\Omega)}
\le C\left(\mathcal{S}_h(\eta_h)+h\|\eta_h^0\|_{L^2(\Omega)}\right)\|\rho_h\|_{H^1(\Omega)}. ( ∇ ⋅ ρ h , η h 0 ) ≤ ∣ c γ σ ( Z h ρ h , η h ) ∣ + C h ∥ ρ h ∥ H 1 ( Ω ) ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + h ∥ η h 0 ∥ L 2 ( Ω ) ) ∥ ρ h ∥ H 1 ( Ω ) .
Moreover, ∣ η h 0 ∣ γ , h = ∣ η h ∣ γ , h |\eta_h^0|_{\gamma,h}=|\eta_h|_{\gamma,h} ∣ η h 0 ∣ γ , h = ∣ η h ∣ γ , h . Hence (22) implies ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + ∣ η h ∣ γ , h ) + C d e f h ∥ η h 0 ∥ L 2 ( Ω ) \|\eta_h^0\|_{L^2(\Omega)}\le C\left(\mathcal{S}_h(\eta_h)+|\eta_h|_{\gamma,h}\right)+C_{\mathrm{def}}h\|\eta_h^0\|_{L^2(\Omega)} ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + ∣ η h ∣ γ , h ) + C def h ∥ η h 0 ∥ L 2 ( Ω ) , where C d e f C_{\mathrm{def}} C def is independent of h h h . Choose h 0 > 0 h_0>0 h 0 > 0 so that C d e f h 0 ≤ 1 2 C_{\mathrm{def}}h_0\le\frac12 C def h 0 ≤ 2 1 . For h ≤ h 0 h\le h_0 h ≤ h 0 , moving the last term to the left gives ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + ∣ η h ∣ γ , h ) \|\eta_h^0\|_{L^2(\Omega)}\le C\left(\mathcal{S}_h(\eta_h)+|\eta_h|_{\gamma,h}\right) ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + ∣ η h ∣ γ , h ) .
We next bound η ˉ h \bar{\eta}_h η ˉ h . Since as S ( 1 ) = 1 \operatorname{as}S(1)=1 as S ( 1 ) = 1 , ∥ S ( 1 ) ∥ L 2 ( Ω ) = ∣ Ω ∣ / 2 \|S(1)\|_{L^2(\Omega)}=|\Omega|/2 ∥ S ( 1 ) ∥ L 2 ( Ω ) = ∣Ω∣/2 , and c γ σ ( S ( 1 ) , η h ) = ∫ Ω η h d x = ∣ Ω ∣ η ˉ h c_{\gamma\sigma}(S(1),\eta_h)=\int_\Omega\eta_h\,dx=|\Omega|\bar{\eta}_h c γ σ ( S ( 1 ) , η h ) = ∫ Ω η h d x = ∣Ω∣ η ˉ h , we have ∥ η ˉ h ∥ L 2 ( Ω ) = ∣ Ω ∣ 1 / 2 ∣ η ˉ h ∣ = 1 2 ∣ c γ σ ( S ( 1 ) , η h ) ∣ ∥ S ( 1 ) ∥ L 2 ( Ω ) ≤ 1 2 S h ( η h ) \|\bar{\eta}_h\|_{L^2(\Omega)}=|\Omega|^{1/2}|\bar{\eta}_h|=\frac{1}{\sqrt{2}}\frac{|c_{\gamma\sigma}(S(1),\eta_h)|}{\|S(1)\|_{L^2(\Omega)}}\le\frac{1}{\sqrt{2}}\mathcal{S}_h(\eta_h) ∥ η ˉ h ∥ L 2 ( Ω ) = ∣Ω ∣ 1/2 ∣ η ˉ h ∣ = 2 1 ∥ S ( 1 ) ∥ L 2 ( Ω ) ∣ c γ σ ( S ( 1 ) , η h ) ∣ ≤ 2 1 S h ( η h ) . Combining the bounds for η h 0 \eta_h^0 η h 0 and η ˉ h \bar{\eta}_h η ˉ h proves (21) .
By the definition of K h K_h K h , (21) gives
∥ η h ∥ L 2 ( Ω ) ≤ C ( sup 0 ≠ τ h ∈ Σ h ∣ b u σ ( τ h , v h ) + c γ σ ( τ h , η h ) ∣ ∥ τ h ∥ L 2 ( Ω ) + ∣ η h ∣ γ , h ) . (28) \|\eta_h\|_{L^2(\Omega)}\le C\left(\sup_{0\ne\tau_h\in\Sigma_h}\frac{|b_{u\sigma}(\tau_h,v_h)+c_{\gamma\sigma}(\tau_h,\eta_h)|}{\|\tau_h\|_{L^2(\Omega)}}+|\eta_h|_{\gamma,h}\right). \tag*{(28)} ∥ η h ∥ L 2 ( Ω ) ≤ C ( 0 = τ h ∈ Σ h sup ∥ τ h ∥ L 2 ( Ω ) ∣ b u σ ( τ h , v h ) + c γ σ ( τ h , η h ) ∣ + ∣ η h ∣ γ , h ) . (28)
By (20) and the Cauchy–Schwarz inequality, ∥ v h ∥ 1 , h \|v_h\|_{1,h} ∥ v h ∥ 1 , h is controlled by the coupling supremum in (28) and ∥ η h ∥ L 2 ( Ω ) \|\eta_h\|_{L^2(\Omega)} ∥ η h ∥ L 2 ( Ω ) . Combining this with (28) proves (19) in 2D.
For d = 3 d=3 d = 3 , let Q h ( 3 ) ⊂ H 0 1 ( Ω ) Q_h^{(3)}\subset H_0^1(\Omega) Q h ( 3 ) ⊂ H 0 1 ( Ω ) be the continuous trilinear Q 1 Q_1 Q 1 space on the Cartesian macro-mesh, and let Q h ( 3 ) , 3 × 3 Q_h^{(3),3\times3} Q h ( 3 ) , 3 × 3 be its matrix-valued counterpart. For matrix-valued fields, both div and curl act row by row. We first establish the vector-valued stabilized divergence estimate, for q h ∈ Γ h q_h\in\Gamma_h q h ∈ Γ h with zero mean componentwise,
∥ q h ∥ L 2 ( Ω ) ≤ C ( sup 0 ≠ W h ∈ Q h ( 3 ) , 3 × 3 ∣ ( q h , div W h ) ∣ ∥ W h ∥ H 1 ( Ω ) + ∣ q h ∣ γ , h ) . (29) \|q_h\|_{L^2(\Omega)}\le C\left(\sup_{0\ne W_h\in Q_h^{(3),3\times3}}\frac{|(q_h,\operatorname{div}W_h)|}{\|W_h\|_{H^1(\Omega)}}+|q_h|_{\gamma,h}\right). \tag*{(29)} ∥ q h ∥ L 2 ( Ω ) ≤ C 0 = W h ∈ Q h ( 3 ) , 3 × 3 sup ∥ W h ∥ H 1 ( Ω ) ∣ ( q h , div W h ) ∣ + ∣ q h ∣ γ , h . (29)
Choose W ∈ H 0 1 ( Ω ) 3 × 3 W\in H_0^1(\Omega)^{3\times3} W ∈ H 0 1 ( Ω ) 3 × 3 row by row so that div W = q h \operatorname{div}W=q_h div W = q h and ∥ W ∥ H 1 ( Ω ) ≤ C ∥ q h ∥ L 2 ( Ω ) \|W\|_{H^1(\Omega)}\le C\|q_h\|_{L^2(\Omega)} ∥ W ∥ H 1 ( Ω ) ≤ C ∥ q h ∥ L 2 ( Ω ) . The boundary-preserving quasi-interpolation construction of [15 Section 6] , applied componentwise and specialized to the conforming tensor-product Q 1 Q_1 Q 1 element on affine Cartesian cuboids, provides I h W ∈ Q h ( 3 ) , 3 × 3 I_hW\in Q_h^{(3),3\times3} I h W ∈ Q h ( 3 ) , 3 × 3 satisfying ∥ I h W ∥ H 1 ( Ω ) + ∑ F ∈ F M i n t h F − 1 ∥ W − I h W ∥ L 2 ( F ) ≤ C ∥ W ∥ H 1 ( Ω ) \|I_hW\|_{H^1(\Omega)}+\sum_{F\in\mathcal{F}^{\mathrm{int}}_M}h_F^{-1}\|W-I_hW\|_{L^2(F)}\le C\|W\|_{H^1(\Omega)} ∥ I h W ∥ H 1 ( Ω ) + ∑ F ∈ F M int h F − 1 ∥ W − I h W ∥ L 2 ( F ) ≤ C ∥ W ∥ H 1 ( Ω ) . Indeed, with e = W − I h W e=W-I_hW e = W − I h W and a fixed-size patch ω M \omega_M ω M , the local estimates ∥ e ∥ L 2 ( M ) ≤ C h M ∥ W ∥ H 1 ( ω M ) \|e\|_{L^2(M)}\le Ch_M\|W\|_{H^1(\omega_M)} ∥ e ∥ L 2 ( M ) ≤ C h M ∥ W ∥ H 1 ( ω M ) and ∥ ∇ e ∥ L 2 ( M ) ≤ C ∥ W ∥ H 1 ( ω M ) \|\nabla e\|_{L^2(M)}\le C\|W\|_{H^1(\omega_M)} ∥∇ e ∥ L 2 ( M ) ≤ C ∥ W ∥ H 1 ( ω M ) and the scaled trace inequality give, on every face F ⊂ ∂ M F\subset\partial M F ⊂ ∂ M , h F − 1 ∥ e ∥ L 2 ( F ) 2 ≤ C ( h M − 2 ∥ e ∥ L 2 ( M ) 2 + ∥ ∇ e ∥ L 2 ( M ) 2 ) ≤ C ∥ W ∥ H 1 ( ω M ) 2 h_F^{-1}\|e\|_{L^2(F)}^2\le C\left(h_M^{-2}\|e\|_{L^2(M)}^2+\|\nabla e\|_{L^2(M)}^2\right)\le C\|W\|_{H^1(\omega_M)}^2 h F − 1 ∥ e ∥ L 2 ( F ) 2 ≤ C ( h M − 2 ∥ e ∥ L 2 ( M ) 2 + ∥∇ e ∥ L 2 ( M ) 2 ) ≤ C ∥ W ∥ H 1 ( ω M ) 2 . Summation proves the stated face bound because the patches have uniformly bounded overlap. Since q h q_h q h is macrocellwise constant and W − I h W W-I_hW W − I h W has zero boundary trace, macrocell integration by parts gives ( q h , div e ) = ∑ F ∈ F M i n t ( [ q h ] , e n F ) F (q_h,\operatorname{div}e)=\sum_{F\in\mathcal{F}^{\mathrm{int}}_M}([q_h],e n_F)_F ( q h , div e ) = ∑ F ∈ F M int ([ q h ] , e n F ) F . The Cauchy–Schwarz inequality then yields
∥ q h ∥ L 2 ( Ω ) 2 ≤ ∣ ( q h , div I h W ) ∣ + ∣ q h ∣ γ , h ( ∑ F ∈ F M i n t h F − 1 ∥ W − I h W ∥ L 2 ( F ) 2 ) 1 / 2 . \|q_h\|_{L^2(\Omega)}^2\le|(q_h,\operatorname{div}I_hW)|+|q_h|_{\gamma,h}\left(\sum_{F\in\mathcal{F}^{\mathrm{int}}_M}h_F^{-1}\|W-I_hW\|_{L^2(F)}^2\right)^{1/2}. ∥ q h ∥ L 2 ( Ω ) 2 ≤ ∣ ( q h , div I h W ) ∣ + ∣ q h ∣ γ , h F ∈ F M int ∑ h F − 1 ∥ W − I h W ∥ L 2 ( F ) 2 1/2 .
This proves (29) . Define the linear map Ξ Φ : = Φ T − ( tr Φ ) I \Xi\Phi:= \Phi^{\mathsf{T}} - (\operatorname{tr}\Phi)I ΞΦ := Φ T − ( tr Φ ) I . For W h ∈ Q h ( 3 ) , 3 × 3 W_h \in Q_h^{(3),3\times3} W h ∈ Q h ( 3 ) , 3 × 3 , set Φ h : = Ξ − 1 W h = W h T − 1 2 ( tr W h ) I \Phi_h := \Xi^{-1}W_h = W_h^{\mathsf{T}} - \frac{1}{2}(\operatorname{tr}W_h)I Φ h := Ξ − 1 W h = W h T − 2 1 ( tr W h ) I . Using the definition of as \operatorname{as} as and the row-wise curl, we obtain
( as ( curl Φ h ) ) i = ∑ j = 1 3 ∂ j ( Φ h ) j i − ∂ i tr Φ h = ( div ( Ξ Φ h ) ) i , i = 1 , 2 , 3. (\operatorname{as}(\operatorname{curl}\Phi_h))_i = \sum_{j=1}^{3}\partial_j(\Phi_h)_{ji} - \partial_i\operatorname{tr}\Phi_h = (\operatorname{div}(\Xi\Phi_h))_i,\qquad i=1,2,3. ( as ( curl Φ h ) ) i = j = 1 ∑ 3 ∂ j ( Φ h ) j i − ∂ i tr Φ h = ( div ( Ξ Φ h ) ) i , i = 1 , 2 , 3.
Hence as ( curl Φ h ) = div W h \operatorname{as}(\operatorname{curl}\Phi_h)=\operatorname{div}W_h as ( curl Φ h ) = div W h .
For each corner cuboid E E E , let F j ( E ) F_j(E) F j ( E ) be its primal quarterface normal to direction j j j . We then define R h Φ h ∈ Σ h R_h\Phi_h\in\Sigma_h R h Φ h ∈ Σ h to be constant on each E E E , with
( R h Φ h ) i j ∣ E : = ∣ F j ( E ) ∣ − 1 ∫ F j ( E ) ( curl Φ h ) i j d s , i , j = 1 , 2 , 3. (R_h\Phi_h)_{ij}\big|_E := |F_j(E)|^{-1}\int_{F_j(E)}(\operatorname{curl}\Phi_h)_{ij}\,ds,\qquad i,j=1,2,3. ( R h Φ h ) ij E := ∣ F j ( E ) ∣ − 1 ∫ F j ( E ) ( curl Φ h ) ij d s , i , j = 1 , 2 , 3.
This uniquely determines the reconstructed stress and preserves the row-wise normal flux integrals on every primal quarterface. Across an interior primal quarterface, the normal curl flux is single-valued: it depends only on tangential derivatives of the common trace of Φ h \Phi_h Φ h . The reconstructed rows therefore have matching normal components. Let h E = diam E h_E=\operatorname{diam}E h E = diam E . Then
∥ R h Φ h ∥ L 2 ( E ) 2 ≤ C h E ∑ i , j = 1 3 ∥ ( curl Φ h ) i j ∥ L 2 ( F j ( E ) ) 2 ≤ C ∥ ∇ Φ h ∥ L 2 ( E ) 2 . \|R_h\Phi_h\|_{L^2(E)}^2\leq Ch_E\sum_{i,j=1}^{3}\|(\operatorname{curl}\Phi_h)_{ij}\|_{L^2(F_j(E))}^2\leq C\|\nabla\Phi_h\|_{L^2(E)}^2. ∥ R h Φ h ∥ L 2 ( E ) 2 ≤ C h E i , j = 1 ∑ 3 ∥ ( curl Φ h ) ij ∥ L 2 ( F j ( E )) 2 ≤ C ∥∇ Φ h ∥ L 2 ( E ) 2 .
Summing over E E E , and using that Ξ − 1 \Xi^{-1} Ξ − 1 is fixed and bounded, yields
∥ R h Φ h ∥ L 2 ( Ω ) ≤ C ∥ W h ∥ H 1 ( Ω ) . (30) \|R_h\Phi_h\|_{L^2(\Omega)}\leq C\|W_h\|_{H^1(\Omega)}. \tag*{(30)} ∥ R h Φ h ∥ L 2 ( Ω ) ≤ C ∥ W h ∥ H 1 ( Ω ) . (30)
By the definition of R h R_h R h , we have ∫ ∂ M ( R h Φ h ) i ⋅ n M d s = ∫ ∂ M ( curl Φ h ) i ⋅ n M d s = 0 \int_{\partial M}(R_h\Phi_h)_i\cdot n_M\,ds=\int_{\partial M}(\operatorname{curl}\Phi_h)_i\cdot n_M\,ds=0 ∫ ∂ M ( R h Φ h ) i ⋅ n M d s = ∫ ∂ M ( curl Φ h ) i ⋅ n M d s = 0 for every M ∈ T M M\in\mathcal{T}_M M ∈ T M and i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 . Therefore R h Φ h ∈ K h R_h\Phi_h\in K_h R h Φ h ∈ K h .
We next show that the reconstruction preserves the integral of the asymmetry on each Cartesian macrocell. Write g i j = ( curl Φ h ) i j g_{ij}=(\operatorname{curl}\Phi_h)_{ij} g ij = ( curl Φ h ) ij . Let F j − F_j^- F j − and F j + F_j^+ F j + be the opposite macrofaces normal to direction j j j , separated by the side length ℓ j \ell_j ℓ j . Each is partitioned into four primal quarterfaces. On a corner cuboid E E E adjoining a quarterface F F F normal to direction j j j , flux matching and ∣ E ∣ = ( ℓ j / 2 ) ∣ F ∣ |E|=(\ell_j/2)|F| ∣ E ∣ = ( ℓ j /2 ) ∣ F ∣ give ∫ E ( R h Φ h ) i j d x = ℓ j 2 ∫ F g i j d s \int_E(R_h\Phi_h)_{ij}\,dx=\frac{\ell_j}{2}\int_Fg_{ij}\,ds ∫ E ( R h Φ h ) ij d x = 2 ℓ j ∫ F g ij d s . Summing over the eight corner cuboids gives
∫ M ( R h Φ h ) i j d x = ℓ j 2 ( ∫ F j − g i j d s + ∫ F j + g i j d s ) = ∫ M g i j d x . (31) \int_M(R_h\Phi_h)_{ij}\,dx=\frac{\ell_j}{2}\left(\int_{F_j^-}g_{ij}\,ds+\int_{F_j^+}g_{ij}\,ds\right)=\int_Mg_{ij}\,dx. \tag*{(31)} ∫ M ( R h Φ h ) ij d x = 2 ℓ j ( ∫ F j − g ij d s + ∫ F j + g ij d s ) = ∫ M g ij d x . (31)
The final equality is the 1D trapezoidal identity in x j x_j x j : trilinearity of Φ h \Phi_h Φ h makes g i j g_{ij} g ij affine in that coordinate, although it may vary tangentially. Because rotations are constant on each macrocell, (31) and as ( curl Φ h ) = div W h \operatorname{as}(\operatorname{curl}\Phi_h)=\operatorname{div}W_h as ( curl Φ h ) = div W h imply
c γ σ ( R h Φ h , η h ) = ( div W h , η h ) , ∀ η h ∈ Γ h . (32) c_{\gamma\sigma}(R_h\Phi_h,\eta_h)=(\operatorname{div}W_h,\eta_h),\qquad\forall\eta_h\in\Gamma_h. \tag*{(32)} c γ σ ( R h Φ h , η h ) = ( div W h , η h ) , ∀ η h ∈ Γ h . (32)
Because W h W_h W h has zero boundary trace, ( div W h , η ˉ h ) = 0 (\operatorname{div}W_h,\bar{\eta}_h)=0 ( div W h , η ˉ h ) = 0 , where η ˉ h = ∣ Ω ∣ − 1 ∫ Ω η h \bar{\eta}_h=|\Omega|^{-1}\int_\Omega\eta_h η ˉ h = ∣Ω ∣ − 1 ∫ Ω η h . Applying (29) to η h 0 = η h − η ˉ h \eta_h^0=\eta_h-\bar{\eta}_h η h 0 = η h − η ˉ h , and then using (32) and (30) , gives ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + ∣ η h ∣ γ , h ) \|\eta_h^0\|_{L^2(\Omega)}\leq C(\mathcal{S}_h(\eta_h)+|\eta_h|_{\gamma,h}) ∥ η h 0 ∥ L 2 ( Ω ) ≤ C ( S h ( η h ) + ∣ η h ∣ γ , h ) . We next bound η ˉ h \bar{\eta}_h η ˉ h . For a ∈ R 3 a\in\mathbb{R}^3 a ∈ R 3 , the globally constant skew stress
S ( a ) = 1 2 ( 0 − a 3 a 2 a 3 0 − a 1 − a 2 a 1 0 ) S(a)=\frac{1}{2}\begin{pmatrix}
0 & -a_3 & a_2\\
a_3 & 0 & -a_1\\
-a_2 & a_1 & 0
\end{pmatrix} S ( a ) = 2 1 0 a 3 − a 2 − a 3 0 a 1 a 2 − a 1 0
belongs to K h K_h K h and satisfies as S ( a ) = a \operatorname{as}S(a)=a as S ( a ) = a . Choosing τ h = S ( η ˉ h ) \tau_h=S(\bar{\eta}_h) τ h = S ( η ˉ h ) in the definition of S h ( η h ) \mathcal{S}_h(\eta_h) S h ( η h ) gives ∥ η ˉ h ∥ L 2 ( Ω ) ≤ C S h ( η h ) \|\bar{\eta}_h\|_{L^2(\Omega)}\leq C\mathcal{S}_h(\eta_h) ∥ η ˉ h ∥ L 2 ( Ω ) ≤ C S h ( η h ) . Consequently (21) holds also in 3D.
Finally, for v h ∈ U h v_h\in U_h v h ∈ U h , prescribe τ v , h n = h F − 1 ⟦ v h ⟧ \tau_{v,h}n=h_F^{-1}\llbracket v_h\rrbracket τ v , h n = h F − 1 [ [ v h ] ] on each primal quarterface F F F , with boundary jumps interpreted as traces. The three linearly independent primal-quarterface normals uniquely determine a constant stress on each corner cuboid. The prescribed normal traces agree across interior primal quarterfaces, hence τ v , h ∈ Σ h \tau_{v,h} \in\Sigma_h τ v , h ∈ Σ h . Scaling gives b u σ ( τ v , h , v h ) = ∥ v h ∥ 1 , h 2 b_{u\sigma}(\tau_{v,h}, v_h) = \|v_h\|_{1,h}^2 b u σ ( τ v , h , v h ) = ∥ v h ∥ 1 , h 2 and ∥ τ v , h ∥ L 2 ( Ω ) ≤ C ∥ v h ∥ 1 , h \|\tau_{v,h}\|_{L^2(\Omega)} \le C\|v_h\|_{1,h} ∥ τ v , h ∥ L 2 ( Ω ) ≤ C ∥ v h ∥ 1 , h . This is the 3D displacement lift. Combining this displacement estimate with the preceding bounds for η h 0 \eta_h^0 η h 0 and η ˉ h \bar{\eta}_h η ˉ h proves (19) .
Theorem 2.3. (Well-posedness). The system (5) –(9) has a unique solution.
Proof. The system is square and finite dimensional, so it suffices to prove that its homogeneous problem has only the zero solution. Let ( σ h ∗ , u h ∗ , γ h ∗ , z ^ h ∗ , p h ∗ ) ∈ Σ h × U h × Γ h × Z h × P h (\sigma_h^*, u_h^*, \gamma_h^*, \widehat{z}_h^*, p_h^*) \in\Sigma_h \times U_h \times\Gamma_h \times Z_h \times P_h ( σ h ∗ , u h ∗ , γ h ∗ , z h ∗ , p h ∗ ) ∈ Σ h × U h × Γ h × Z h × P h satisfy the homogeneous system. Taking ψ h = σ h ∗ \psi_h = \sigma_h^* ψ h = σ h ∗ , v h = u h ∗ v_h = u_h^* v h = u h ∗ , and η h = γ h ∗ \eta_h = \gamma_h^* η h = γ h ∗ in (5) –(7) , respectively, and using (4) , we obtain
a σ σ ( σ h ∗ , σ h ∗ ) + a σ p ( p h ∗ , σ h ∗ ) + s γ ( γ h ∗ , γ h ∗ ) = 0. (33) a_{\sigma\sigma}(\sigma_h^*,\sigma_h^*) + a_{\sigma p}(p_h^*,\sigma_h^*) + s_\gamma(\gamma_h^*,\gamma_h^*) = 0. \tag*{(33)} a σ σ ( σ h ∗ , σ h ∗ ) + a σ p ( p h ∗ , σ h ∗ ) + s γ ( γ h ∗ , γ h ∗ ) = 0. (33)
Likewise, taking ξ h = z ^ h ∗ \xi_h = \widehat{z}_h^* ξ h = z h ∗ and w h = p h ∗ w_h = p_h^* w h = p h ∗ in (8) and (9) , and adding the two equalities, yields
( ( Δ t K ) − 1 z ^ h ∗ , z ^ h ∗ ) + a p σ ( σ h ∗ , p h ∗ ) + a p p ( p h ∗ , p h ∗ ) = 0. (34) ((\Delta t K)^{-1}\widehat{z}_h^*,\widehat{z}_h^*) + a_{p\sigma}(\sigma_h^*,p_h^*) + a_{pp}(p_h^*,p_h^*) = 0. \tag*{(34)} (( Δ t K ) − 1 z h ∗ , z h ∗ ) + a p σ ( σ h ∗ , p h ∗ ) + a pp ( p h ∗ , p h ∗ ) = 0. (34)
Adding (33) and (34) yields
0 = ∥ A 1 / 2 ( σ h ∗ + α B p h ∗ I ) ∥ L 2 ( Ω ) 2 + c 0 ∥ p h ∗ ∥ L 2 ( Ω ) 2 + ( ( Δ t K ) − 1 z ^ h ∗ , z ^ h ∗ ) + s γ ( γ h ∗ , γ h ∗ ) . 0 = \|A^{1/2}(\sigma_h^*+\alpha_B p_h^*I)\|_{L^2(\Omega)}^2 + c_0\|p_h^*\|_{L^2(\Omega)}^2 + ((\Delta t K)^{-1}\widehat{z}_h^*,\widehat{z}_h^*) + s_\gamma(\gamma_h^*,\gamma_h^*). 0 = ∥ A 1/2 ( σ h ∗ + α B p h ∗ I ) ∥ L 2 ( Ω ) 2 + c 0 ∥ p h ∗ ∥ L 2 ( Ω ) 2 + (( Δ t K ) − 1 z h ∗ , z h ∗ ) + s γ ( γ h ∗ , γ h ∗ ) .
By the positive definiteness of A A A and the nonnegativity of the stabilization, we obtain
σ h ∗ + α B p h ∗ I = 0 , z ^ h ∗ = 0 , ∣ γ h ∗ ∣ γ , h = 0. (35) \sigma_h^*+\alpha_B p_h^*I=0,\qquad\widehat{z}_h^*=0,\qquad|\gamma_h^*|_{\gamma,h}=0. \tag*{(35)} σ h ∗ + α B p h ∗ I = 0 , z h ∗ = 0 , ∣ γ h ∗ ∣ γ , h = 0. (35)
Together, (8) and (35) imply b z p ( p h ∗ , ξ h ) = 0 b_{zp}(p_h^*,\xi_h)=0 b z p ( p h ∗ , ξ h ) = 0 for all ξ h ∈ Z h \xi_h \in Z_h ξ h ∈ Z h . Hence, by (18) ,
c D ∥ p h ∗ ∥ 1 , h ≤ sup 0 ≠ ξ h ∈ Z h b z p ( p h ∗ , ξ h ) ∥ ξ h ∥ L 2 ( Ω ) = 0. c_D\|p_h^*\|_{1,h}\le\sup_{0\ne\xi_h\in Z_h}\frac{b_{zp}(p_h^*,\xi_h)}{\|\xi_h\|_{L^2(\Omega)}}=0. c D ∥ p h ∗ ∥ 1 , h ≤ 0 = ξ h ∈ Z h sup ∥ ξ h ∥ L 2 ( Ω ) b z p ( p h ∗ , ξ h ) = 0.
The discrete Poincaré inequality for P 0 ( T M ) P_0(\mathcal{T}_M) P 0 ( T M ) gives ∥ p h ∗ ∥ L 2 ( Ω ) ≲ ∥ p h ∗ ∥ 1 , h = 0 \|p_h^*\|_{L^2(\Omega)}\lesssim\|p_h^*\|_{1,h}=0 ∥ p h ∗ ∥ L 2 ( Ω ) ≲ ∥ p h ∗ ∥ 1 , h = 0 . Thus p h ∗ = 0 p_h^*=0 p h ∗ = 0 , and (35) yields σ h ∗ = 0 \sigma_h^*=0 σ h ∗ = 0 .
Finally, (5) reduces to b u σ ( ψ h , u h ∗ ) + c γ σ ( ψ h , γ h ∗ ) = 0 b_{u\sigma}(\psi_h,u_h^*)+c_{\gamma\sigma}(\psi_h,\gamma_h^*)=0 b u σ ( ψ h , u h ∗ ) + c γ σ ( ψ h , γ h ∗ ) = 0 for all ψ h ∈ Σ h \psi_h\in\Sigma_h ψ h ∈ Σ h . Together with ∣ γ h ∗ ∣ γ , h = 0 |\gamma_h^*|_{\gamma,h}=0 ∣ γ h ∗ ∣ γ , h = 0 , the stabilized condition (19) yields ∥ ( u h ∗ , γ h ∗ ) ∥ h = 0 \|(u_h^*,\gamma_h^*)\|_h=0 ∥ ( u h ∗ , γ h ∗ ) ∥ h = 0 . Hence ( σ h ∗ , u h ∗ , γ h ∗ , z ^ h ∗ , p h ∗ ) = 0 (\sigma_h^*,u_h^*,\gamma_h^*,\widehat{z}_h^*,p_h^*)=0 ( σ h ∗ , u h ∗ , γ h ∗ , z h ∗ , p h ∗ ) = 0 .
Block preconditioners and spectral analysis
This section develops block preconditioners for the reduced displacement–rotation–pressure matrix (16) . Starting from the parameter-robust equivalence K m ≃ D m K_m \simeq D_m K m ≃ D m , which justifies replacing the coupled mechanics block by separate displacement and rotation blocks, we study the inexpensive uncoupled block-diagonal preconditioners P b d T \mathcal{P}_{\mathrm{bd}}^{\mathrm{T}} P bd T and P b d P \mathcal{P}_{\mathrm{bd}}^{\mathrm{P}} P bd P . They retain the natural pressure blocks, and their stability is characterized by the exact discrete ratio Θ h \Theta_h Θ h , with the explicit sufficient estimate Θ h ≤ α B 2 / ( c 0 K d r ) \Theta_h \le\alpha_B^2/(c_0K_{\mathrm{dr}}) Θ h ≤ α B 2 / ( c 0 K dr ) when c 0 > 0 c_0>0 c 0 > 0 . The third subsection introduces the exact and Schur-consistent pressure blocks. We prove uniform spectral equivalence for the block-diagonal preconditioners P b d e x \mathcal{P}_{\mathrm{bd}}^{\mathrm{ex}} P bd ex and P b d s c \mathcal{P}_{\mathrm{bd}}^{\mathrm{sc}} P bd sc . Moreover, we derive weighted FOV estimates for the lower-triangular factors P l t e x \mathcal{P}_{\mathrm{lt}}^{\mathrm{ex}} P lt ex and P l t s c \mathcal{P}_{\mathrm{lt}}^{\mathrm{sc}} P lt sc . The distinction is cost versus robustness: the uncoupled preconditioners are cheaper but require pressure control through T p T_p T p , whereas the Schur-consistent preconditioners use a more expensive pressure block. Finally, we describe practical implementations of the displacement, rotation, and pressure inverses.
Mechanical block equivalence
The purpose of this subsection is to prove
K m ≃ D m , K_m \simeq D_m, K m ≃ D m ,
with constants independent of h h h , λ \lambda λ , and μ \mu μ . The upper bound is immediate once the condensed energies are written in terms of compliance-Riesz representatives. The essential point is the lower bound: the displacement and rotation representatives may in principle cancel in the coupled energy, while a direct uniform comparison of the compliance and L 2 L^2 L 2 stress norms is unavailable as λ → ∞ \lambda\to\infty λ → ∞ . The inf–sup estimate controls the rotation through the total stress representative and the rotation-jump stabilization term.
We use the compliance norm ∥ τ h ∥ A 2 : = a σ σ ( τ h , τ h ) \lVert\tau_h\rVert_A^2 := a_{\sigma\sigma}(\tau_h,\tau_h) ∥ τ h ∥ A 2 := a σ σ ( τ h , τ h ) . For τ h ∈ Σ h \tau_h \in\Sigma_h τ h ∈ Σ h , let τ h D : = τ h − 1 d tr ( τ h ) I \tau_h^D := \tau_h - \frac{1}{d}\operatorname{tr}(\tau_h)I τ h D := τ h − d 1 tr ( τ h ) I denote its deviatoric part. The L 2 L^2 L 2 -orthogonal decomposition into deviatoric and spherical parts gives ∥ τ h ∥ L 2 ( Ω ) 2 = ∥ τ h D ∥ L 2 ( Ω ) 2 + 1 d ∥ tr ( τ h ) ∥ L 2 ( Ω ) 2 \lVert\tau_h\rVert_{L^2(\Omega)}^2 = \lVert\tau_h^D\rVert_{L^2(\Omega)}^2 + \frac{1}{d}\lVert\operatorname{tr}(\tau_h)\rVert_{L^2(\Omega)}^2 ∥ τ h ∥ L 2 ( Ω ) 2 = ∥ τ h D ∥ L 2 ( Ω ) 2 + d 1 ∥ tr ( τ h ) ∥ L 2 ( Ω ) 2 . Combining this identity with the isotropic compliance tensor (2.2), we obtain
∥ τ h ∥ A 2 = 1 2 μ ∥ τ h D ∥ L 2 ( Ω ) 2 + 1 d ( 2 μ + d λ ) ∥ tr ( τ h ) ∥ L 2 ( Ω ) 2 . \lVert\tau_h\rVert_A^2 = \frac{1}{2\mu}\lVert\tau_h^D\rVert_{L^2(\Omega)}^2 + \frac{1}{d(2\mu+d\lambda)}\lVert\operatorname{tr}(\tau_h)\rVert_{L^2(\Omega)}^2. ∥ τ h ∥ A 2 = 2 μ 1 ∥ τ h D ∥ L 2 ( Ω ) 2 + d ( 2 μ + d λ ) 1 ∥ tr ( τ h ) ∥ L 2 ( Ω ) 2 .
Consequently, ∥ τ h D ∥ L 2 ( Ω ) ≤ 2 μ ∥ τ h ∥ A \lVert\tau_h^D\rVert_{L^2(\Omega)} \le\sqrt{2\mu}\,\lVert\tau_h\rVert_A ∥ τ h D ∥ L 2 ( Ω ) ≤ 2 μ ∥ τ h ∥ A , and ∥ τ h ∥ A ≤ 1 2 μ ∥ τ h ∥ L 2 ( Ω ) \lVert\tau_h\rVert_A \le\frac{1}{\sqrt{2\mu}}\lVert\tau_h\rVert_{L^2(\Omega)} ∥ τ h ∥ A ≤ 2 μ 1 ∥ τ h ∥ L 2 ( Ω ) .
For any coefficient vector m = ( u , γ ) \mathbf{m}=(\mathbf{u},\boldsymbol{\gamma}) m = ( u , γ ) , set u h : = ∑ j ( u ) j V j \mathbf{u}_h:=\sum_j(\mathbf{u})_jV_j u h := ∑ j ( u ) j V j and γ h : = ∑ k ( γ ) k χ k \gamma_h:=\sum_k(\boldsymbol{\gamma})_k\chi_k γ h := ∑ k ( γ ) k χ k using the bases { V j } \{V_j\} { V j } of U h U_h U h and { χ k } \{\chi_k\} { χ k } of Γ h \Gamma_h Γ h . Introduce the linear functionals
ℓ u ( τ h ) : = b u σ ( τ h , u h ) , ℓ γ ( τ h ) : = c γ σ ( τ h , γ h ) , ℓ m : = ℓ u + ℓ γ . \ell_u(\tau_h):=b_{u\sigma}(\tau_h,u_h),\qquad\ell_\gamma(\tau_h):=c_{\gamma\sigma}(\tau_h,\gamma_h),\qquad\ell_m:=\ell_u+\ell_\gamma. ℓ u ( τ h ) := b u σ ( τ h , u h ) , ℓ γ ( τ h ) := c γ σ ( τ h , γ h ) , ℓ m := ℓ u + ℓ γ .
With respect to the compliance bilinear form a σ σ ( ⋅ , ⋅ ) a_{\sigma\sigma}(\cdot,\cdot) a σ σ ( ⋅ , ⋅ ) , define the Riesz representatives ρ u , ρ γ , ρ m ∈ Σ h \rho_u,\rho_\gamma,\rho_m\in\Sigma_h ρ u , ρ γ , ρ m ∈ Σ h by
(3.1)
a σ σ ( ρ u , τ h ) = ℓ u ( τ h ) , a σ σ ( ρ γ , τ h ) = ℓ γ ( τ h ) , a σ σ ( ρ m , τ h ) = ℓ m ( τ h ) , ∀ τ h ∈ Σ h . a_{\sigma\sigma}(\rho_u,\tau_h)=\ell_u(\tau_h),\qquad a_{\sigma\sigma}(\rho_\gamma,\tau_h)=\ell_\gamma(\tau_h),\qquad a_{\sigma\sigma}(\rho_m,\tau_h)=\ell_m(\tau_h),\qquad\forall\tau_h\in\Sigma_h. a σ σ ( ρ u , τ h ) = ℓ u ( τ h ) , a σ σ ( ρ γ , τ h ) = ℓ γ ( τ h ) , a σ σ ( ρ m , τ h ) = ℓ m ( τ h ) , ∀ τ h ∈ Σ h .
By linearity and uniqueness of the Riesz representation, ρ m = ρ u + ρ γ \rho_m=\rho_u+\rho_\gamma ρ m = ρ u + ρ γ .
Lemma 3.1. For every m = ( u , γ ) \mathbf{m}=(\mathbf{u},\boldsymbol{\gamma}) m = ( u , γ ) ,
( K m m , m ) = ∥ ρ m ∥ A 2 + s γ ( γ h , γ h ) , ( D m m , m ) = ∥ ρ u ∥ A 2 + ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) . (K_m\mathbf{m},\mathbf{m})=\lVert\rho_m\rVert_A^2+s_\gamma(\gamma_h,\gamma_h),\qquad(D_m\mathbf{m},\mathbf{m})=\lVert\rho_u\rVert_A^2+\lVert\rho_\gamma\rVert_A^2+s_\gamma(\gamma_h,\gamma_h). ( K m m , m ) = ∥ ρ m ∥ A 2 + s γ ( γ h , γ h ) , ( D m m , m ) = ∥ ρ u ∥ A 2 + ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) .
Proof. Let y u \mathbf{y}_u y u , y γ \mathbf{y}_\gamma y γ , and y m \mathbf{y}_m y m be the coefficient vectors of ρ u \rho_u ρ u , ρ γ \rho_\gamma ρ γ , and ρ m \rho_m ρ m , respectively. By the block definitions (2.6), the coefficient vectors of ℓ u \ell_u ℓ u , ℓ γ \ell_\gamma ℓ γ , and ℓ m \ell_m ℓ m are b u = B T u \mathbf{b}_u=B^\mathsf{T}\mathbf{u} b u = B T u , b γ = C T γ \mathbf{b}_\gamma=C^\mathsf{T}\boldsymbol{\gamma} b γ = C T γ , and b m = B T u + C T γ \mathbf{b}_m=B^\mathsf{T}\mathbf{u}+C^\mathsf{T}\boldsymbol{\gamma} b m = B T u + C T γ . Thus the above Riesz relations are equivalent to A y u = b u A\mathbf{y}_u=\mathbf{b}_u A y u = b u , A y γ = b γ A\mathbf{y}_\gamma=\mathbf{b}_\gamma A y γ = b γ , and A y m = b m A\mathbf{y}_m=\mathbf{b}_m A y m = b m . It follows that
∥ ρ m ∥ A 2 = y m T A y m = b m T A − 1 b m = ( B T u + C T γ ) T A − 1 ( B T u + C T γ ) = ( K m 0 m , m ) . \lVert\rho_m\rVert_A^2=\mathbf{y}_m^\mathsf{T}A\mathbf{y}_m=\mathbf{b}_m^\mathsf{T}A^{-1}\mathbf{b}_m=(B^\mathsf{T}\mathbf{u}+C^\mathsf{T}\boldsymbol{\gamma})^\mathsf{T}A^{-1}(B^\mathsf{T}\mathbf{u}+C^\mathsf{T}\boldsymbol{\gamma})=(K_m^0\mathbf{m},\mathbf{m}). ∥ ρ m ∥ A 2 = y m T A y m = b m T A − 1 b m = ( B T u + C T γ ) T A − 1 ( B T u + C T γ ) = ( K m 0 m , m ) .
Since K m = K m 0 + diag ( 0 , J γ ) K_m=K_m^0+\operatorname{diag}(0,J_\gamma) K m = K m 0 + diag ( 0 , J γ ) , adding s γ ( γ h , γ h ) = ( J γ γ , γ ) s_\gamma(\gamma_h,\gamma_h)=(J_\gamma\boldsymbol{\gamma},\boldsymbol{\gamma}) s γ ( γ h , γ h ) = ( J γ γ , γ ) gives the first identity of the lemma. Applying the same argument to b u \mathbf{b}_u b u and b γ \mathbf{b}_\gamma b γ separately gives
∥ ρ u ∥ A 2 + ∥ ρ γ ∥ A 2 = ( K u u u , u ) + ( K γ γ 0 γ , γ ) . \lVert\rho_u\rVert_A^2+\lVert\rho_\gamma\rVert_A^2=(K_{uu}\mathbf{u},\mathbf{u})+(K_{\gamma\gamma}^0\boldsymbol{\gamma},\boldsymbol{\gamma}). ∥ ρ u ∥ A 2 + ∥ ρ γ ∥ A 2 = ( K uu u , u ) + ( K γ γ 0 γ , γ ) .
Adding the same stabilization energy gives the identity for D m D_m D m . ∎
The next estimate for the rotation representative is the key parameter-robust step. It uses only the deviatoric part of the rotation representative and therefore avoids the degenerating spherical weight in the compliance decomposition above.
Lemma 3.2. There exists a constant C a s > 0 C_{\mathrm{as}} > 0 C as > 0 , independent of h h h , λ \lambda λ , and μ \mu μ , such that
∥ ρ γ ∥ A ≤ C a s 2 μ ∥ γ h ∥ L 2 ( Ω ) . \lVert\rho_{\gamma} \rVert_{A} \le C_{\mathrm{as}}\sqrt{2\mu}\,\lVert\gamma_h \rVert_{L^2(\Omega)}. ∥ ρ γ ∥ A ≤ C as 2 μ ∥ γ h ∥ L 2 ( Ω ) .
Proof. Taking τ h = ρ γ \tau_h = \rho_{\gamma} τ h = ρ γ in (3.1) gives ∥ ρ γ ∥ A 2 = ( as ( ρ γ ) , γ h ) \lVert\rho_{\gamma} \rVert_A^2 = (\operatorname{as}(\rho_{\gamma}),\gamma_h) ∥ ρ γ ∥ A 2 = ( as ( ρ γ ) , γ h ) . Since as ( q I ) = 0 \operatorname{as}(qI) = 0 as ( q I ) = 0 for every scalar q q q , the rotation functional annihilates the spherical part, and hence ( as ( ρ γ ) , γ h ) = ( as ( ρ γ D ) , γ h ) (\operatorname{as}(\rho_{\gamma}),\gamma_h) = (\operatorname{as}(\rho_{\gamma}^D),\gamma_h) ( as ( ρ γ ) , γ h ) = ( as ( ρ γ D ) , γ h ) . Using the Cauchy–Schwarz inequality, we obtain ∥ ρ γ ∥ A 2 ≤ C a s ∥ ρ γ D ∥ L 2 ( Ω ) ∥ γ h ∥ L 2 ( Ω ) ≤ C a s 2 μ ∥ ρ γ ∥ A ∥ γ h ∥ L 2 ( Ω ) \lVert\rho_{\gamma} \rVert_A^2 \le C_{\mathrm{as}}\lVert\rho_{\gamma}^D \rVert_{L^2(\Omega)}\lVert\gamma_h \rVert_{L^2(\Omega)} \le C_{\mathrm{as}}\sqrt{2\mu}\,\lVert\rho_{\gamma} \rVert_A\lVert\gamma_h \rVert_{L^2(\Omega)} ∥ ρ γ ∥ A 2 ≤ C as ∥ ρ γ D ∥ L 2 ( Ω ) ∥ γ h ∥ L 2 ( Ω ) ≤ C as 2 μ ∥ ρ γ ∥ A ∥ γ h ∥ L 2 ( Ω ) . Dividing by ∥ ρ γ ∥ A \lVert\rho_{\gamma} \rVert_A ∥ ρ γ ∥ A when it is nonzero proves the estimate; the zero case is immediate.
Theorem 3.3. There exist constants c m , C m > 0 c_m,C_m > 0 c m , C m > 0 , independent of h h h , λ \lambda λ , and μ \mu μ , such that
c m ( D m m , m ) ≤ ( K m m , m ) ≤ C m ( D m m , m ) , ∀ m = ( u , γ ) . c_m(\mathsf{D}_m\mathsf{m},\mathsf{m}) \le(\mathsf{K}_m\mathsf{m},\mathsf{m}) \le C_m(\mathsf{D}_m\mathsf{m},\mathsf{m}), \qquad\forall\mathsf{m} = (\mathsf{u},\gamma). c m ( D m m , m ) ≤ ( K m m , m ) ≤ C m ( D m m , m ) , ∀ m = ( u , γ ) .
Proof. By Lemma 3.1,
( K m m , m ) = ∥ ρ u + ρ γ ∥ A 2 + s γ ( γ h , γ h ) ≤ 2 ∥ ρ u ∥ A 2 + 2 ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) ≤ 2 ( D m m , m ) . (\mathsf{K}_m\mathsf{m},\mathsf{m}) = \lVert\rho_u+\rho_{\gamma}\rVert_A^2+s_{\gamma}(\gamma_h,\gamma_h) \le2\lVert\rho_u\rVert_A^2+2\lVert\rho_{\gamma}\rVert_A^2+s_{\gamma}(\gamma_h,\gamma_h) \le2(\mathsf{D}_m\mathsf{m},\mathsf{m}). ( K m m , m ) = ∥ ρ u + ρ γ ∥ A 2 + s γ ( γ h , γ h ) ≤ 2 ∥ ρ u ∥ A 2 + 2 ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) ≤ 2 ( D m m , m ) .
For the lower bound, Lemma 2.2, the definition of ℓ m \ell_m ℓ m , and ∥ τ h ∥ A ≤ ( 2 μ ) − 1 / 2 ∥ τ h ∥ L 2 ( Ω ) \lVert\tau_h\rVert_A \le(2\mu)^{-1/2}\lVert\tau_h\rVert_{L^2(\Omega)} ∥ τ h ∥ A ≤ ( 2 μ ) − 1/2 ∥ τ h ∥ L 2 ( Ω ) give c E 2 ∥ ( u h , γ h ) ∥ h 2 ≤ 1 2 μ ∥ ρ m ∥ A 2 + ∣ γ h ∣ γ , h 2 c_E^2\lVert(u_h,\gamma_h)\rVert_h^2 \le\frac{1}{2\mu}\lVert\rho_m\rVert_A^2+\lvert\gamma_h\rvert_{\gamma,h}^2 c E 2 ∥( u h , γ h ) ∥ h 2 ≤ 2 μ 1 ∥ ρ m ∥ A 2 + ∣ γ h ∣ γ , h 2 . Since s γ ( γ h , γ h ) = δ γ μ ∣ γ h ∣ γ , h 2 s_{\gamma}(\gamma_h,\gamma_h)=\delta_{\gamma}\mu\lvert\gamma_h\rvert_{\gamma,h}^2 s γ ( γ h , γ h ) = δ γ μ ∣ γ h ∣ γ , h 2 , it follows, for fixed δ γ > 0 \delta_{\gamma}>0 δ γ > 0 , that 2 μ ∥ γ h ∥ L 2 ( Ω ) 2 ≲ ∥ ρ m ∥ A 2 + s γ ( γ h , γ h ) = ( K m m , m ) 2\mu\lVert\gamma_h\rVert_{L^2(\Omega)}^2\lesssim\lVert\rho_m\rVert_A^2+s_{\gamma}(\gamma_h,\gamma_h)=(\mathsf{K}_m\mathsf{m},\mathsf{m}) 2 μ ∥ γ h ∥ L 2 ( Ω ) 2 ≲ ∥ ρ m ∥ A 2 + s γ ( γ h , γ h ) = ( K m m , m ) . Lemma 3.2 then controls ∥ ρ γ ∥ A \lVert\rho_{\gamma}\rVert_A ∥ ρ γ ∥ A by ( K m m , m ) (\mathsf{K}_m\mathsf{m},\mathsf{m}) ( K m m , m ) ; the identity ρ u = ρ m − ρ γ \rho_u=\rho_m-\rho_{\gamma} ρ u = ρ m − ρ γ controls ∥ ρ u ∥ A \lVert\rho_u\rVert_A ∥ ρ u ∥ A , while s γ ( γ h , γ h ) ≤ ( K m m , m ) s_{\gamma}(\gamma_h,\gamma_h)\le(\mathsf{K}_m\mathsf{m},\mathsf{m}) s γ ( γ h , γ h ) ≤ ( K m m , m ) holds directly. Hence ( D m m , m ) = ∥ ρ u ∥ A 2 + ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) ≲ ( K m m , m ) (\mathsf{D}_m\mathsf{m},\mathsf{m})=\lVert\rho_u\rVert_A^2+\lVert\rho_{\gamma}\rVert_A^2+s_{\gamma}(\gamma_h,\gamma_h)\lesssim(\mathsf{K}_m\mathsf{m},\mathsf{m}) ( D m m , m ) = ∥ ρ u ∥ A 2 + ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) ≲ ( K m m , m ) . This gives the required equivalence.
Finally, we note that the stabilized rotation block is uniformly mass-equivalent.
Corollary 3.4. There exist c γ , C γ > 0 c_{\gamma},C_{\gamma}>0 c γ , C γ > 0 , independent of h h h , λ \lambda λ , μ \mu μ , such that c γ μ M γ ⪯ K γ γ ⪯ C γ μ M γ c_{\gamma}\mu\mathsf{M}_{\gamma}\preceq\mathsf{K}_{\gamma\gamma}\preceq C_{\gamma}\mu\mathsf{M}_{\gamma} c γ μ M γ ⪯ K γ γ ⪯ C γ μ M γ .
Proof. Let γ h ∈ Γ h \gamma_h\in\Gamma_h γ h ∈ Γ h be the macrocell rotation associated with γ \gamma γ , and let ρ γ \rho_{\gamma} ρ γ be its compliance-Riesz representative. Taking m = ( 0 , γ ) \mathsf{m}=(0,\gamma) m = ( 0 , γ ) in Lemma 3.1 gives E γ : = ( K γ γ γ , γ ) = ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) = ∥ ρ γ ∥ A 2 + δ γ μ ∣ γ h ∣ γ , h 2 E_{\gamma}:=(\mathsf{K}_{\gamma\gamma}\gamma,\gamma)=\lVert\rho_{\gamma}\rVert_A^2+s_{\gamma}(\gamma_h,\gamma_h)=\lVert\rho_{\gamma}\rVert_A^2+\delta_{\gamma}\mu\lvert\gamma_h\rvert_{\gamma,h}^2 E γ := ( K γ γ γ , γ ) = ∥ ρ γ ∥ A 2 + s γ ( γ h , γ h ) = ∥ ρ γ ∥ A 2 + δ γ μ ∣ γ h ∣ γ , h 2 . For the upper bound, Lemma 3.2 gives ∥ ρ γ ∥ A 2 ≤ 2 C a s 2 μ ∥ γ h ∥ L 2 ( Ω ) 2 \lVert\rho_{\gamma}\rVert_A^2\le2C_{\mathrm{as}}^2\mu\lVert\gamma_h\rVert_{L^2(\Omega)}^2 ∥ ρ γ ∥ A 2 ≤ 2 C as 2 μ ∥ γ h ∥ L 2 ( Ω ) 2 , while the face inverse estimate gives ∣ γ h ∣ γ , h 2 = ∑ F ∈ F M i n t h F ∥ J γ h K ∥ L 2 ( F ) 2 ≤ C J ∥ γ h ∥ L 2 ( Ω ) 2 \lvert\gamma_h\rvert_{\gamma,h}^2=\sum_{F\in\mathcal{F}_{M}^{\mathrm{int}}}h_F\lVert J\gamma_hK\rVert_{L^2(F)}^2\le C_J\lVert\gamma_h\rVert_{L^2(\Omega)}^2 ∣ γ h ∣ γ , h 2 = ∑ F ∈ F M int h F ∥ J γ h K ∥ L 2 ( F ) 2 ≤ C J ∥ γ h ∥ L 2 ( Ω ) 2 . Hence E γ ≤ ( 2 C a s 2 + δ γ C J ) μ ∥ γ h ∥ L 2 ( Ω ) 2 E_{\gamma}\le(2C_{\mathrm{as}}^2+\delta_{\gamma}C_J)\mu\lVert\gamma_h\rVert_{L^2(\Omega)}^2 E γ ≤ ( 2 C as 2 + δ γ C J ) μ ∥ γ h ∥ L 2 ( Ω ) 2 . For the lower bound, the rotation-energy estimate established in the proof of Theorem 3.3, applied with m = ( 0 , γ ) \mathsf{m}=(0,\gamma) m = ( 0 , γ ) , gives 2 μ ∥ γ h ∥ L 2 ( Ω ) 2 ≲ ( K m m , m ) = E γ 2\mu\lVert\gamma_h\rVert_{L^2(\Omega)}^2\lesssim(\mathsf{K}_m\mathsf{m},\mathsf{m})=E_{\gamma} 2 μ ∥ γ h ∥ L 2 ( Ω ) 2 ≲ ( K m m , m ) = E γ . Since ( M γ γ , γ ) = ∥ γ h ∥ L 2 ( Ω ) 2 (\mathsf{M}_{\gamma}\gamma,\gamma)=\lVert\gamma_h\rVert_{L^2(\Omega)}^2 ( M γ γ , γ ) = ∥ γ h ∥ L 2 ( Ω ) 2 , the two bounds prove the equivalence.
The uncoupled preconditioners
Condensation gives the natural pressure block T p = M p + ( A p p − S p ∘ ) + S p ∘ \mathsf{T}_p=\mathsf{M}_p+(\mathsf{A}_{pp}-\mathsf{S}_p^{\circ})+\mathsf{S}_p^{\circ} T p = M p + ( A pp − S p ∘ ) + S p ∘ , equivalently T p = c 0 M 0 + ( c λ M 0 − S p ∘ ) + S p ∘ \mathsf{T}_p=c_0\mathsf{M}_0+(c_{\lambda}\mathsf{M}_0-\mathsf{S}_p^{\circ})+\mathsf{S}_p^{\circ} T p = c 0 M 0 + ( c λ M 0 − S p ∘ ) + S p ∘ , whereas the simpler positive definite pressure block is P m z : = M p + A p p + S p z = T p + S p z \mathsf{P}_m^z:=\mathsf{M}_p+\mathsf{A}_{pp}+\mathsf{S}_p^z=\mathsf{T}_p+\mathsf{S}_p^z P m z := M p + A pp + S p z = T p + S p z . We consider the two block-diagonal preconditioners
P b d T p = diag ( D m , T p ) , P b d P m z = diag ( D m , P m z ) . (36) \mathcal{P}_{\mathrm{bd}}^{\mathrm{T}_p}=\operatorname{diag}(\mathsf{D}_m,\mathsf{T}_p), \qquad\mathcal{P}_{\mathrm{bd}}^{\mathsf{P}_m^z}=\operatorname{diag}(\mathsf{D}_m,\mathsf{P}_m^z). \tag*{(36)} P bd T p = diag ( D m , T p ) , P bd P m z = diag ( D m , P m z ) . (36)
We refer to these two preconditioners as uncoupled preconditioners, since their pressure blocks do not involve the mechanics-induced Schur correction. We first determine when the natural pressure block can be inverted. Define the normalized coercivity constant of the non-storage pressure part
δ h : = inf p ≠ 0 ( ( A p p − S p σ ) p , p ) + ( S p σ p , p ) ( M 0 p , p ) . \delta_h := \inf_{\mathbf{p}\ne0} \frac{\left(\left(A_{pp}-S_p^\sigma\right)\mathbf{p},\mathbf{p}\right)+\left(S_p^\sigma\mathbf{p},\mathbf{p}\right)}{(M_0\mathbf{p},\mathbf{p})}. δ h := p = 0 inf ( M 0 p , p ) ( ( A pp − S p σ ) p , p ) + ( S p σ p , p ) .
Hence, T p ≻ 0 T_p \succ0 T p ≻ 0 if and only if c 0 + δ h > 0 c_0+\delta_h>0 c 0 + δ h > 0 . For the remainder of this subsection, we assume T p ≻ 0 T_p\succ0 T p ≻ 0 .
To quantify separately the relative size of the stress-induced pressure block and the omitted mechanics–pressure coupling in the diag ( D m , T p ) \operatorname{diag}(D_m,T_p) diag ( D m , T p ) -scaled norm, introduce the following two dimensionless quantities:
Θ h : = ∥ T p − 1 / 2 S p σ T p − 1 / 2 ∥ 2 = sup p ≠ 0 ( S p σ p , p ) ( T p p , p ) , G h : = ∥ D m − 1 / 2 L T p − 1 / 2 ∥ 2 = sup ζ ≠ 0 p ≠ 0 ∣ ( L p , ζ ) ∣ ( D m ζ , ζ ) 1 / 2 ( T p p , p ) 1 / 2 . \Theta_h := \left\|T_p^{-1/2}S_p^\sigma T_p^{-1/2}\right\|_2 = \sup_{\mathbf{p}\ne0}\frac{(S_p^\sigma\mathbf{p},\mathbf{p})}{(T_p\mathbf{p},\mathbf{p})}, \qquad G_h := \left\|D_m^{-1/2}LT_p^{-1/2}\right\|_2 = \sup_{\substack{\boldsymbol{\zeta}\ne0\\\mathbf{p}\ne0}}\frac{\lvert(L\mathbf{p},\boldsymbol{\zeta})\rvert}{(D_m\boldsymbol{\zeta},\boldsymbol{\zeta})^{1/2}(T_p\mathbf{p},\mathbf{p})^{1/2}}. Θ h := T p − 1/2 S p σ T p − 1/2 2 = p = 0 sup ( T p p , p ) ( S p σ p , p ) , G h := D m − 1/2 L T p − 1/2 2 = ζ = 0 p = 0 sup ( D m ζ , ζ ) 1/2 ( T p p , p ) 1/2 ∣( L p , ζ )∣ .
Thus Θ h \Theta_h Θ h is the smallest constant for which
S p σ ⪯ Θ h T p . (37) S_p^\sigma\preceq\Theta_h T_p. \tag*{(37)} S p σ ⪯ Θ h T p . (37)
The next lemma connects these two quantities.
Lemma 3.5. For every pressure vector p \mathbf{p} p and mechanics vector ζ \boldsymbol{\zeta} ζ ,
∣ ( L p , ζ ) ∣ ≤ ( S p σ p , p ) 1 / 2 ( K m ζ , ζ ) 1 / 2 . (38) \lvert(L\mathbf{p},\boldsymbol{\zeta})\rvert\le(S_p^\sigma\mathbf{p},\mathbf{p})^{1/2}(K_m\boldsymbol{\zeta},\boldsymbol{\zeta})^{1/2}. \tag*{(38)} ∣( L p , ζ )∣ ≤ ( S p σ p , p ) 1/2 ( K m ζ , ζ ) 1/2 . (38) Consequently,
∣ ( L p , ζ ) ∣ ≤ C m Θ h ( T p p , p ) 1 / 2 ( D m ζ , ζ ) 1 / 2 , G h 2 ≤ C m Θ h . (39) \lvert(L\mathbf{p},\boldsymbol{\zeta})\rvert\le\sqrt{C_m\Theta_h}(T_p\mathbf{p},\mathbf{p})^{1/2}(D_m\boldsymbol{\zeta},\boldsymbol{\zeta})^{1/2}, \qquad G_h^2\le C_m\Theta_h. \tag*{(39)} ∣( L p , ζ )∣ ≤ C m Θ h ( T p p , p ) 1/2 ( D m ζ , ζ ) 1/2 , G h 2 ≤ C m Θ h . (39)
Proof. Writing ζ = ( v , η ) \boldsymbol{\zeta}=(\mathbf{v},\boldsymbol{\eta}) ζ = ( v , η ) and using the definition of L L L , we obtain ( L p , ζ ) = ( A − 1 / 2 E p , A − 1 / 2 ( B T v + C T η ) ) (L\mathbf{p},\boldsymbol{\zeta})=(A^{-1/2}E\mathbf{p},A^{-1/2}(B^{\mathsf{T}}\mathbf{v}+C^{\mathsf{T}}\boldsymbol{\eta})) ( L p , ζ ) = ( A − 1/2 E p , A − 1/2 ( B T v + C T η )) . Cauchy–Schwarz gives ∣ ( L p , ζ ) ∣ ≤ ( S p σ p , p ) 1 / 2 ( K m 0 ζ , ζ ) 1 / 2 ≤ ( S p σ p , p ) 1 / 2 ( K m ζ , ζ ) 1 / 2 \lvert(L\mathbf{p},\boldsymbol{\zeta})\rvert\le(S_p^\sigma\mathbf{p},\mathbf{p})^{1/2}(K_m^0\boldsymbol{\zeta},\boldsymbol{\zeta})^{1/2}\le(S_p^\sigma\mathbf{p},\mathbf{p})^{1/2}(K_m\boldsymbol{\zeta},\boldsymbol{\zeta})^{1/2} ∣( L p , ζ )∣ ≤ ( S p σ p , p ) 1/2 ( K m 0 ζ , ζ ) 1/2 ≤ ( S p σ p , p ) 1/2 ( K m ζ , ζ ) 1/2 , because K m = K m 0 + diag ( 0 , J γ ) ⪰ K m 0 K_m=K_m^0+\operatorname{diag}(0,J_\gamma)\succeq K_m^0 K m = K m 0 + diag ( 0 , J γ ) ⪰ K m 0 . This proves (3.4). Combining it with K m ⪯ C m D m K_m\preceq C_mD_m K m ⪯ C m D m and (3.3) gives (3.5). □ \square □
Since P m z = T p + S p σ P_m^z=T_p+S_p^\sigma P m z = T p + S p σ , relation (3.3) yields T p ⪯ P m z ⪯ ( 1 + Θ h ) T p T_p\preceq P_m^z\preceq(1+\Theta_h)T_p T p ⪯ P m z ⪯ ( 1 + Θ h ) T p . Introduce the energy norms induced by the two preconditioners,
∥ ( m , p ) ∥ D T p 2 : = ( D m m , m ) + ( T p p , p ) , ∥ ( m , p ) ∥ D P m z 2 : = ( D m m , m ) + ( P m z p , p ) . \|(\mathbf{m},\mathbf{p})\|_{D_{T_p}}^2 := (D_m\mathbf{m},\mathbf{m})+(T_p\mathbf{p},\mathbf{p}), \qquad\|(\mathbf{m},\mathbf{p})\|_{D_{P_m^z}}^2 := (D_m\mathbf{m},\mathbf{m})+(P_m^z\mathbf{p},\mathbf{p}). ∥ ( m , p ) ∥ D T p 2 := ( D m m , m ) + ( T p p , p ) , ∥ ( m , p ) ∥ D P m z 2 := ( D m m , m ) + ( P m z p , p ) .
Then
∥ ( m , p ) ∥ D T p 2 ≤ ∥ ( m , p ) ∥ D P m z 2 ≤ ( 1 + Θ h ) ∥ ( m , p ) ∥ D T p 2 . (40) \|(\mathbf{m},\mathbf{p})\|_{D_{T_p}}^2 \le\|(\mathbf{m},\mathbf{p})\|_{D_{P_m^z}}^2 \le(1+\Theta_h)\|(\mathbf{m},\mathbf{p})\|_{D_{T_p}}^2. \tag*{(40)} ∥ ( m , p ) ∥ D T p 2 ≤ ∥ ( m , p ) ∥ D P m z 2 ≤ ( 1 + Θ h ) ∥ ( m , p ) ∥ D T p 2 . (40)
The bilinear form associated with A r e d A_{\mathrm{red}} A red is
B r e d ( ( m , p ) , ( ζ , q ) ) : = − ( K m m , ζ ) − ( L p , ζ ) − ( L q , m ) + ( T p p , q ) . B_{\mathrm{red}}((\mathbf{m},\mathbf{p}),(\boldsymbol{\zeta},\mathbf{q})) := -(K_m\mathbf{m},\boldsymbol{\zeta})-(L\mathbf{p},\boldsymbol{\zeta})-(L\mathbf{q},\mathbf{m})+(T_p\mathbf{p},\mathbf{q}). B red (( m , p ) , ( ζ , q )) := − ( K m m , ζ ) − ( L p , ζ ) − ( L q , m ) + ( T p p , q ) .
Finally, set C T ( θ ) : = C m + 1 + ( C m − 1 ) 2 + 4 C m θ 2 C_T(\theta):=\frac{C_m+1+\sqrt{(C_m-1)^2+4C_m\theta}}{2} C T ( θ ) := 2 C m + 1 + ( C m − 1 ) 2 + 4 C m θ , and β T : = min { c m , 1 } \beta_T:=\min\{c_m,1\} β T := min { c m , 1 } .
Theorem 3.6. For every parameter choice satisfying T p ≻ 0 T_p\succ0 T p ≻ 0 , the preconditioner P b d T p \mathcal{P}_{\mathrm{bd}}^{T_p} P bd T p satisfies
∣ B r e d ( ( m , p ) , ( ζ , q ) ) ∣ ≤ C T ( Θ h ) ∥ ( m , p ) ∥ D T p ∥ ( ζ , q ) ∥ D T p , (41) \lvert B_{\mathrm{red}}((\mathbf{m},\mathbf{p}),(\boldsymbol{\zeta},\mathbf{q}))\rvert\le C_T(\Theta_h)\|(\mathbf{m},\mathbf{p})\|_{D_{T_p}}\|(\boldsymbol{\zeta},\mathbf{q})\|_{D_{T_p}}, \tag*{(41)} ∣ B red (( m , p ) , ( ζ , q ))∣ ≤ C T ( Θ h ) ∥ ( m , p ) ∥ D T p ∥ ( ζ , q ) ∥ D T p , (41) inf ( m , p ) ≠ 0 sup ( ζ , q ) ≠ 0 B r e d ( ( m , p ) , ( ζ , q ) ) ∥ ( m , p ) ∥ D T p ∥ ( ζ , q ) ∥ D T p ≥ β T . (42) \inf_{(\mathbf{m},\mathbf{p})\ne0}\sup_{(\boldsymbol{\zeta},\mathbf{q})\ne0}\frac{B_{\mathrm{red}}((\mathbf{m},\mathbf{p}),(\boldsymbol{\zeta},\mathbf{q}))}{\|(\mathbf{m},\mathbf{p})\|_{D_{T_p}}\|(\boldsymbol{\zeta},\mathbf{q})\|_{D_{T_p}}}\ge\beta_T. \tag*{(42)} ( m , p ) = 0 inf ( ζ , q ) = 0 sup ∥ ( m , p ) ∥ D T p ∥ ( ζ , q ) ∥ D T p B red (( m , p ) , ( ζ , q )) ≥ β T . (42)
The preconditioner P b d P m z P_{\mathrm{bd}}^{P_m^z} P bd P m z satisfies
∣ B r e d ( ( m , p ) , ( ζ , q ) ) ∣ ≤ C T ( Θ h ) ∥ ( m , p ) ∥ D p m z ∥ ( ζ , q ) ∥ D p m z , \left|B_{\mathrm{red}}((\mathbf{m},p),(\boldsymbol{\zeta},q))\right| \le C_T(\Theta_h)\left\|(\mathbf{m},p)\right\|_{D_{p_m^z}}\left\|(\boldsymbol{\zeta},q)\right\|_{D_{p_m^z}}, ∣ B red (( m , p ) , ( ζ , q )) ∣ ≤ C T ( Θ h ) ∥ ( m , p ) ∥ D p m z ∥ ( ζ , q ) ∥ D p m z ,
inf ( m , p ) ≠ 0 sup ( ζ , q ) ≠ 0 B r e d ( ( m , p ) , ( ζ , q ) ) ∥ ( m , p ) ∥ D p m z ∥ ( ζ , q ) ∥ D p m z ≥ β T 1 + Θ h . \inf_{(\mathbf{m},p)\ne0}\sup_{(\boldsymbol{\zeta},q)\ne0}\frac{B_{\mathrm{red}}((\mathbf{m},p),(\boldsymbol{\zeta},q))}{\left\|(\mathbf{m},p)\right\|_{D_{p_m^z}}\left\|(\boldsymbol{\zeta},q)\right\|_{D_{p_m^z}}}\ge\frac{\beta_T}{1+\Theta_h}. ( m , p ) = 0 inf ( ζ , q ) = 0 sup ∥ ( m , p ) ∥ D p m z ∥ ( ζ , q ) ∥ D p m z B red (( m , p ) , ( ζ , q )) ≥ 1 + Θ h β T .
Proof. Set a = ( D m m , m ) 1 / 2 a=(D_m\mathbf{m},\mathbf{m})^{1/2} a = ( D m m , m ) 1/2 , b = ( T p p , p ) 1 / 2 b=(T_p p,p)^{1/2} b = ( T p p , p ) 1/2 , c = ( D m ζ , ζ ) 1 / 2 c=(D_m\boldsymbol{\zeta},\boldsymbol{\zeta})^{1/2} c = ( D m ζ , ζ ) 1/2 , and d = ( T p q , q ) 1 / 2 d=(T_p q,q)^{1/2} d = ( T p q , q ) 1/2 . Theorem 3.3 and Lemma 3.5 yield
∣ B r e d ( ( m , p ) , ( ζ , q ) ) ∣ ≤ [ a b ] [ C m C m Θ h C m Θ h C m Θ h 1 ] [ c d ] . \left|B_{\mathrm{red}}((\mathbf{m},p),(\boldsymbol{\zeta},q))\right|\le[a\ \ b]
\begin{bmatrix}
\frac{C_m}{\sqrt{C_m\Theta_h}} & \sqrt{C_m\Theta_h}\\
\sqrt{C_m\Theta_h} & 1
\end{bmatrix}
\begin{bmatrix}
c\\
d
\end{bmatrix}. ∣ B red (( m , p ) , ( ζ , q )) ∣ ≤ [ a b ] [ C m Θ h C m C m Θ h C m Θ h 1 ] [ c d ] .
The spectral norm of the middle matrix is C T ( Θ h ) C_T(\Theta_h) C T ( Θ h ) , proving (41) . For the inf–sup bound choose ( ζ , q ) = ( − m , p ) (\boldsymbol{\zeta},q)=(-\mathbf{m},p) ( ζ , q ) = ( − m , p ) ; then B r e d ( ( m , p ) , ( − m , p ) ) = ( K m m , m ) + ( T p p , p ) ≥ β T ∥ ( m , p ) ∥ D T p 2 B_{\mathrm{red}}((\mathbf{m},p),(-\mathbf{m},p))=(K_m\mathbf{m},\mathbf{m})+(T_p p,p)\ge\beta_T\left\|(\mathbf{m},p)\right\|_{D_{T_p}}^2 B red (( m , p ) , ( − m , p )) = ( K m m , m ) + ( T p p , p ) ≥ β T ∥ ( m , p ) ∥ D T p 2 . This proves (42) . The two estimates for P b d P m z P_{\mathrm{bd}}^{P_m^z} P bd P m z follow from the first two estimates and (3.6).
The bounds in Theorem 3.6 are expressed in terms of the discrete ratio Θ h \Theta_h Θ h . We next derive a transparent sufficient bound in terms of the physical parameters. By the definition of δ h \delta_h δ h , T p ⪰ ( c 0 + δ h ) M 0 T_p\succeq(c_0+\delta_h)M_0 T p ⪰ ( c 0 + δ h ) M 0 , while S p z ⪯ A p p = c λ M 0 S_p^z\preceq A_{pp}=c_\lambda M_0 S p z ⪯ A pp = c λ M 0 . Therefore
Θ h ≤ c λ c 0 + δ h = α B 2 K d r ( c 0 + δ h ) . \Theta_h\le\frac{c_\lambda}{c_0+\delta_h}=\frac{\alpha_B^2}{K_{\mathrm{dr}}(c_0+\delta_h)}. Θ h ≤ c 0 + δ h c λ = K dr ( c 0 + δ h ) α B 2 .
For c 0 > 0 c_0>0 c 0 > 0 , define
χ λ , c 0 : = c λ c 0 = α B 2 c 0 K d r = d α B 2 c 0 ( 2 μ + d λ ) . \chi_{\lambda,c_0}:=\frac{c_\lambda}{c_0}=\frac{\alpha_B^2}{c_0K_{\mathrm{dr}}}=\frac{d\alpha_B^2}{c_0(2\mu+d\lambda)}. χ λ , c 0 := c 0 c λ = c 0 K dr α B 2 = c 0 ( 2 μ + d λ ) d α B 2 .
Then Θ h ≤ χ λ , c 0 \Theta_h\le\chi_{\lambda,c_0} Θ h ≤ χ λ , c 0 , and the following parameter-explicit consequence is immediate.
Corollary 3.7. Let P \mathcal{P} P be a family of parameter choices with c 0 > 0 c_0>0 c 0 > 0 such that
χ ∗ : = sup P α B 2 c 0 K d r < ∞ . (43) \chi_*:=\sup_{\mathcal{P}}\frac{\alpha_B^2}{c_0K_{\mathrm{dr}}}<\infty. \tag*{(43)} χ ∗ := P sup c 0 K dr α B 2 < ∞. (43) Then P b d T p P_{\mathrm{bd}}^{T_p} P bd T p and P b d P m z P_{\mathrm{bd}}^{P_m^z} P bd P m z are uniformly stable on P \mathcal{P} P . Their continuity and inf–sup constants depend only on c m c_m c m , C m C_m C m , and χ ∗ \chi_* χ ∗ , and are independent of h h h , Δ t \Delta t Δ t , K K K , λ \lambda λ , and c 0 c_0 c 0 .
Condition (43) is sufficient rather than necessary. In particular, it permits the critical scaling
c 0 = C α B 2 K d r , C > 0 , c_0=C\frac{\alpha_B^2}{K_{\mathrm{dr}}},\qquad C>0, c 0 = C K dr α B 2 , C > 0 ,
and hence, for fixed α B > 0 \alpha_B>0 α B > 0 and μ \mu μ , c 0 ≍ K d r − 1 ∼ λ − 1 c_0\asymp K_{\mathrm{dr}}^{-1}\sim\lambda^{-1} c 0 ≍ K dr − 1 ∼ λ − 1 in the nearly incompressible limit. More generally, Theorem 3.6 yields uniform stability bounds provided that T p ≻ 0 T_p\succ0 T p ≻ 0 and Θ h \Theta_h Θ h is uniformly bounded. Thus, when c 0 = 0 c_0=0 c 0 = 0 , the uncoupled analysis may still apply if δ h > 0 \delta_h>0 δ h > 0 and the residual compliance and Darcy terms provide a uniform bound for Θ h \Theta_h Θ h . If c 0 = δ h = 0 c_0=\delta_h=0 c 0 = δ h = 0 , then T p T_p T p is singular: P b d T p P_{\mathrm{bd}}^{T_p} P bd T p is not defined on the full pressure space, and the present Θ h \Theta_h Θ h -based argument does not yield a stability estimate for P b d P m z P_{\mathrm{bd}}^{P_m^z} P bd P m z .
It is useful to close this subsection by comparing the two uncoupled pressure blocks with the exact pressure Schur complement. Define the mechanics-induced pressure correction Q : = L T K m − 1 L Q:=L^{\mathsf T}K_m^{-1}L Q := L T K m − 1 L and the exact Schur complement S p : = T p + Q S_p:=T_p+Q S p := T p + Q . The following result compares T p T_p T p , S p S_p S p , and P b d P m z P_{\mathrm{bd}}^{P_m^z} P bd P m z .
Proposition 3.8. The mechanics-induced correction satisfies
0 ⪯ Q ⪯ S p σ . (44) 0 \preceq Q \preceq S_p^{\sigma}. \tag*{(44)} 0 ⪯ Q ⪯ S p σ . (44) Consequently,
T p ⪯ S p ⪯ P m z . (45) T_p \preceq S_p \preceq P_m^{z}. \tag*{(45)} T p ⪯ S p ⪯ P m z . (45)
Proof. Since K m ≻ 0 K_m \succ0 K m ≻ 0 , Q ⪰ 0 Q \succeq0 Q ⪰ 0 . For every p \mathbf{p} p , (38) gives ( Q p , p ) = sup ζ ≠ 0 ∣ ( L p , ζ ) ∣ 2 ( K m ζ , ζ ) ≤ ( S p σ p , p ) (Q\mathbf{p},\mathbf{p}) = \sup_{\boldsymbol{\zeta}\ne0} \frac{|(L\mathbf{p},\boldsymbol{\zeta})|^2}{(K_m\boldsymbol{\zeta},\boldsymbol{\zeta})} \le(S_p^{\sigma}\mathbf{p},\mathbf{p}) ( Q p , p ) = sup ζ = 0 ( K m ζ , ζ ) ∣ ( L p , ζ ) ∣ 2 ≤ ( S p σ p , p ) . Thus (3.10) holds; using S p = T p + Q S_p = T_p + Q S p = T p + Q and P m z = T p + S p σ P_m^{z} = T_p + S_p^{\sigma} P m z = T p + S p σ gives (3.11). □
The relative size of the exact mechanics-induced pressure correction is measured by Θ h s c h : = ∥ T p − 1 / 2 Q T p − 1 / 2 ∥ 2 \Theta_h^{\mathrm{sch}} := \left\|T_p^{-1/2}QT_p^{-1/2}\right\|_2 Θ h sch := T p − 1/2 Q T p − 1/2 2 . Combining Theorem 3.3 with Proposition 3.8 gives the hierarchy
c m Θ h s c h ≤ Θ h σ ≤ C m Θ h s c h ≤ C m Θ h ≤ C m α B 2 K d r ( c 0 + δ h ) . (46) c_m\Theta_h^{\mathrm{sch}} \le\Theta_h^{\sigma} \le C_m\Theta_h^{\mathrm{sch}} \le C_m\Theta_h \le\frac{C_m\alpha_B^2}{K_{\mathrm{dr}}(c_0+\delta_h)}. \tag*{(46)} c m Θ h sch ≤ Θ h σ ≤ C m Θ h sch ≤ C m Θ h ≤ K dr ( c 0 + δ h ) C m α B 2 . (46)
Indeed, c m D m ⪯ K m ⪯ C m D m c_mD_m \preceq K_m \preceq C_mD_m c m D m ⪯ K m ⪯ C m D m implies ( 1 / C m ) D m − 1 ⪯ K m − 1 ⪯ ( 1 / c m ) D m − 1 (1/C_m)D_m^{-1} \preceq K_m^{-1} \preceq(1/c_m)D_m^{-1} ( 1/ C m ) D m − 1 ⪯ K m − 1 ⪯ ( 1/ c m ) D m − 1 , while Q ⪯ S p σ Q \preceq S_p^{\sigma} Q ⪯ S p σ . Thus the three pressure blocks have a nested interpretation: the natural block T p T_p T p omits the mechanics correction, the exact Schur block S p S_p S p adds the true correction Q Q Q , and the simpler block P m z P_m^{z} P m z replaces Q Q Q by the larger stress-compliance envelope S p σ S_p^{\sigma} S p σ . Formula (3.12) also separates three levels of information: Θ h s c h \Theta_h^{\mathrm{sch}} Θ h sch is the exact Schur-correction ratio, Θ h \Theta_h Θ h is its stress-compliance envelope, and χ λ , c 0 \chi_{\lambda,c_0} χ λ , c 0 is the coarser storage-only envelope when c 0 > 0 c_0>0 c 0 > 0 . Hence growth of χ λ , c 0 \chi_{\lambda,c_0} χ λ , c 0 alone cannot prove deterioration of the actual coupling; it only shows that storage has ceased to provide a uniform proof. This motivates the Schur-consistent pressure blocks studied next, which approximate the correction Q Q Q instead of relying only on the envelope S p σ S_p^{\sigma} S p σ .
Spectral analysis of the Schur-consistent preconditioners
The exact pressure Schur complement is used only as an algebraic benchmark. We first analyze the exact block factors and then replace K m − 1 K_m^{-1} K m − 1 by the computable block-diagonal inverse D m − 1 D_m^{-1} D m − 1 . Recall the exact mechanics-induced pressure correction Q : = L T K m − 1 L Q := L^{\mathsf T}K_m^{-1}L Q := L T K m − 1 L and S p : = T p + Q S_p := T_p + Q S p := T p + Q , and define its Schur-consistent counterpart by
Q D : = L T D m − 1 L , P s c : = T p + Q D . (47) Q_D := L^{\mathsf T}D_m^{-1}L,\qquad P_{\mathrm{sc}} := T_p + Q_D. \tag*{(47)} Q D := L T D m − 1 L , P sc := T p + Q D . (47)
Since K m , D m ≻ 0 K_m,D_m \succ0 K m , D m ≻ 0 , T p ⪰ 0 T_p \succeq0 T p ⪰ 0 , and the condensed matrix A r e d A_{\mathrm{red}} A red is nonsingular by Theorem 2.3 , both S p S_p S p and P s c P_{\mathrm{sc}} P sc are positive definite. Define the exact and Schur-consistent block-diagonal preconditioners by
P b d e x : = [ K m 0 0 S p ] , P b d s c : = [ D m 0 0 P s c ] . (48) \mathcal{P}_{\mathrm{bd}}^{\mathrm{ex}} :=
\begin{bmatrix}
K_m & 0\\
0 & S_p
\end{bmatrix},
\qquad
\mathcal{P}_{\mathrm{bd}}^{\mathrm{sc}} :=
\begin{bmatrix}
D_m & 0\\
0 & P_{\mathrm{sc}}
\end{bmatrix}. \tag*{(48)} P bd ex := [ K m 0 0 S p ] , P bd sc := [ D m 0 0 P sc ] . (48)
and the lower-triangular factors by
P l t e x : = [ K m 0 L T − S p ] , P l t s c : = [ D m 0 L T − P s c ] . (49) \mathcal{P}_{\mathrm{lt}}^{\mathrm{ex}} :=
\begin{bmatrix}
K_m & 0\\
L^{\mathsf T} & -S_p
\end{bmatrix},
\qquad
\mathcal{P}_{\mathrm{lt}}^{\mathrm{sc}} :=
\begin{bmatrix}
D_m & 0\\
L^{\mathsf T} & -P_{\mathrm{sc}}
\end{bmatrix}. \tag*{(49)} P lt ex := [ K m L T 0 − S p ] , P lt sc := [ D m L T 0 − P sc ] . (49)
For the sign convention of A r e d A_{\mathrm{red}} A red , set P l t , r e d e x : = − P l t e x \mathcal{P}_{\mathrm{lt,red}}^{\mathrm{ex}} := -\mathcal{P}_{\mathrm{lt}}^{\mathrm{ex}} P lt , red ex := − P lt ex and P l t , r e d s c : = − P l t s c \mathcal{P}_{\mathrm{lt,red}}^{\mathrm{sc}} := -\mathcal{P}_{\mathrm{lt}}^{\mathrm{sc}} P lt , red sc := − P lt sc .
The following result is the bridge from the exact Schur complement to its computable counterpart. It shows that replacing K m − 1 K_m^{-1} K m − 1 by D m − 1 D_m^{-1} D m − 1 in the mechanics-induced pressure correction preserves the pressure energy uniformly.
Theorem 3.9. Let c m c_m c m and C m C_m C m be the constants in Theorem 3.3 , and set α s c : = min { 1 , C m − 1 } \alpha_{\mathrm{sc}} := \min\{1,C_m^{-1}\} α sc := min { 1 , C m − 1 } and β s c : = max { 1 , c m − 1 } \beta_{\mathrm{sc}} := \max\{1,c_m^{-1}\} β sc := max { 1 , c m − 1 } . Then
α s c P s c ⪯ S p ⪯ β s c P s c . (50) \alpha_{\mathrm{sc}}P_{\mathrm{sc}} \preceq S_p \preceq\beta_{\mathrm{sc}}P_{\mathrm{sc}}. \tag*{(50)} α sc P sc ⪯ S p ⪯ β sc P sc . (50) Proof. From Theorem 3.3 , we have C m − 1 D m − 1 ⪯ K m − 1 ⪯ c m − 1 D m − 1 C_m^{-1}D_m^{-1} \preceq K_m^{-1} \preceq c_m^{-1}D_m^{-1} C m − 1 D m − 1 ⪯ K m − 1 ⪯ c m − 1 D m − 1 . Addition of T p ⪰ 0 T_p \succeq0 T p ⪰ 0 yields T p + C m − 1 Q D ⪯ S p ⪯ T p + c m − 1 Q D T_p+C_m^{-1}Q_D \preceq S_p \preceq T_p+c_m^{-1}Q_D T p + C m − 1 Q D ⪯ S p ⪯ T p + c m − 1 Q D , which proves (50) since T p ⪰ 0 T_p \succeq0 T p ⪰ 0 and Q D ⪰ 0 Q_D \succeq0 Q D ⪰ 0 .
We next establish eigenvalue bounds for the ideal block-diagonal preconditioner of the reduced system, which contains the positive semidefinite pressure block T p T_p T p .
Theorem 3.10. Let φ = ( 1 + 5 ) / 2 \varphi=(1+\sqrt{5})/2 φ = ( 1 + 5 ) /2 . Then
spec ( ( P b d e x ) − 1 A r e d ) ⊂ [ − φ , − 1 ] ∪ [ φ − 1 , 1 ] . (51) \operatorname{spec}\left((P_{\mathrm{bd}}^{\mathrm{ex}})^{-1}A_{\mathrm{red}}\right) \subset[-\varphi,-1]\cup[\varphi^{-1},1]. \tag*{(51)} spec ( ( P bd ex ) − 1 A red ) ⊂ [ − φ , − 1 ] ∪ [ φ − 1 , 1 ] . (51) Consequently, for every vector x x x ,
φ − 1 ∥ x ∥ P e x ≤ ∥ A r e d x ∥ ( P b d e x ) − 1 ≤ φ ∥ x ∥ P e x . (52) \varphi^{-1}\lVert x\rVert_{P_{\mathrm{ex}}} \le\lVert A_{\mathrm{red}}x\rVert_{(P_{\mathrm{bd}}^{\mathrm{ex}})^{-1}} \le\varphi\lVert x\rVert_{P_{\mathrm{ex}}}. \tag*{(52)} φ − 1 ∥ x ∥ P ex ≤ ∥ A red x ∥ ( P bd ex ) − 1 ≤ φ ∥ x ∥ P ex . (52) Proof. Set H : = K m − 1 / 2 L S p − 1 / 2 \mathbf{H}:=K_m^{-1/2}L S_p^{-1/2} H := K m − 1/2 L S p − 1/2 , so that
0 ⪯ H T H = S p − 1 / 2 L T K m − 1 L S p − 1 / 2 ⪯ I . (53) 0\preceq\mathbf{H}^{\mathsf{T}}\mathbf{H}=S_p^{-1/2}L^{\mathsf{T}}K_m^{-1}L S_p^{-1/2}\preceq I. \tag*{(53)} 0 ⪯ H T H = S p − 1/2 L T K m − 1 L S p − 1/2 ⪯ I . (53) The preconditioned matrix is similar to the symmetric matrix
T b d e x : = ( P b d e x ) − 1 / 2 A r e d ( P b d e x ) − 1 / 2 = [ − I − H − H T I − H T H ] . T_{\mathrm{bd}}^{\mathrm{ex}}:=(P_{\mathrm{bd}}^{\mathrm{ex}})^{-1/2}A_{\mathrm{red}}(P_{\mathrm{bd}}^{\mathrm{ex}})^{-1/2}
=
\begin{bmatrix}
-I & -\mathbf{H}\\
-\mathbf{H}^{\mathsf{T}} & I-\mathbf{H}^{\mathsf{T}}\mathbf{H}
\end{bmatrix}. T bd ex := ( P bd ex ) − 1/2 A red ( P bd ex ) − 1/2 = [ − I − H T − H I − H T H ] . Let σ i ∈ ( 0 , 1 ] \sigma_i\in(0,1] σ i ∈ ( 0 , 1 ] be a positive singular value of H \mathbf{H} H , with corresponding left and right singular vectors q i q_i q i and p i p_i p i , and put ω i : = σ i 2 \omega_i:=\sigma_i^2 ω i := σ i 2 . The subspace spanned by ( q i , 0 ) T (q_i,0)^{\mathsf{T}} ( q i , 0 ) T and ( 0 , p i ) T (0,p_i)^{\mathsf{T}} ( 0 , p i ) T is invariant under T b d e x T_{\mathrm{bd}}^{\mathrm{ex}} T bd ex , and the restriction of T b d e x T_{\mathrm{bd}}^{\mathrm{ex}} T bd ex to this subspace is
[ − 1 − σ i − σ i 1 − σ i 2 ] . \begin{bmatrix}
-1 & -\sigma_i\\
-\sigma_i & 1-\sigma_i^2
\end{bmatrix}. [ − 1 − σ i − σ i 1 − σ i 2 ] . Its eigenvalues are ρ i ± = − ω i ± ω i 2 + 4 2 \rho_i^{\pm}=\frac{-\omega_i\pm\sqrt{\omega_i^2+4}}{2} ρ i ± = 2 − ω i ± ω i 2 + 4 . For 0 < ω i ≤ 1 0<\omega_i\le1 0 < ω i ≤ 1 , the negative branch lies in [ − φ , − 1 ) [-\varphi,-1) [ − φ , − 1 ) , and the positive branch lies in [ φ − 1 , 1 ) [\varphi^{-1},1) [ φ − 1 , 1 ) . It remains to treat the case σ i = 0 \sigma_i=0 σ i = 0 . For H T q = 0 \mathbf{H}^{\mathsf{T}}q=0 H T q = 0 , we have T b d e x ( q , 0 ) T = − ( q , 0 ) T T_{\mathrm{bd}}^{\mathrm{ex}}(q,0)^{\mathsf{T}}=-(q,0)^{\mathsf{T}} T bd ex ( q , 0 ) T = − ( q , 0 ) T , and if H p = 0 \mathbf{H}p=0 H p = 0 , then T b d e x ( 0 , p ) T = ( 0 , p ) T T_{\mathrm{bd}}^{\mathrm{ex}}(0,p)^{\mathsf{T}}=(0,p)^{\mathsf{T}} T bd ex ( 0 , p ) T = ( 0 , p ) T . Therefore, the remaining directions associated with zero singular values contribute only the eigenvalues 1 1 1 and − 1 -1 − 1 . This proves (51) .
Because T b d e x T_{\mathrm{bd}}^{\mathrm{ex}} T bd ex is symmetric, (51) implies φ − 1 ∥ y ∥ 2 ≤ ∥ T b d e x y ∥ 2 ≤ φ ∥ y ∥ 2 \varphi^{-1}\lVert y\rVert_2\le\lVert T_{\mathrm{bd}}^{\mathrm{ex}}y\rVert_2\le\varphi\lVert y\rVert_2 φ − 1 ∥ y ∥ 2 ≤ ∥ T bd ex y ∥ 2 ≤ φ ∥ y ∥ 2 . Taking y = ( P b d e x ) 1 / 2 x y=(P_{\mathrm{bd}}^{\mathrm{ex}})^{1/2}x y = ( P bd ex ) 1/2 x gives (52) .
Combining the exact benchmark with the uniform equivalences K m ≃ D m K_m\simeq D_m K m ≃ D m and S p ≃ P s c S_p\simeq P_{\mathrm{sc}} S p ≃ P sc transfers the same parameter-independent spectral separation to the Schur-consistent block-diagonal preconditioner.
Corollary 3.11. Set a b d : = min { c m , α s c } a_{\mathrm{bd}}:=\min\{c_m,\alpha_{\mathrm{sc}}\} a bd := min { c m , α sc } , b b d : = max { C m , β s c } b_{\mathrm{bd}}:=\max\{C_m,\beta_{\mathrm{sc}}\} b bd := max { C m , β sc } . Then
spec ( ( P b d s c ) − 1 A r e d ) ⊂ [ − φ b b d , − a b d / φ ] ∪ [ a b d / φ , φ b b d ] . (54) \operatorname{spec}\left((P_{\mathrm{bd}}^{\mathrm{sc}})^{-1}A_{\mathrm{red}}\right)\subset[-\varphi b_{\mathrm{bd}},-a_{\mathrm{bd}}/\varphi]\cup[a_{\mathrm{bd}}/\varphi,\varphi b_{\mathrm{bd}}]. \tag*{(54)} spec ( ( P bd sc ) − 1 A red ) ⊂ [ − φ b bd , − a bd / φ ] ∪ [ a bd / φ , φ b bd ] . (54) Proof. The block inequalities in Theorem 3.3 and Theorem 3.9 give a b d P b d e x ⪯ P b d s c ⪯ b b d P b d e x a_{\mathrm{bd}}P_{\mathrm{bd}}^{\mathrm{ex}}\preceq P_{\mathrm{bd}}^{\mathrm{sc}}\preceq b_{\mathrm{bd}}P_{\mathrm{bd}}^{\mathrm{ex}} a bd P bd ex ⪯ P bd sc ⪯ b bd P bd ex . Hence a b d ∥ x ∥ P s c ≤ ∥ x ∥ P e x ≤ b b d ∥ x ∥ P s c \sqrt{a_{\mathrm{bd}}}\lVert x\rVert_{P_{\mathrm{sc}}}\le\lVert x\rVert_{P_{\mathrm{ex}}}\le\sqrt{b_{\mathrm{bd}}}\lVert x\rVert_{P_{\mathrm{sc}}} a bd ∥ x ∥ P sc ≤ ∥ x ∥ P ex ≤ b bd ∥ x ∥ P sc , whereas inversion of the same block equivalence gives a b d ∥ v ∥ ( P e x b d ) − 1 ≤ ∥ v ∥ ( P s c b d ) − 1 ≤ b b d ∥ v ∥ ( P e x b d ) − 1 \sqrt{a_{\mathrm{bd}}}\lVert v\rVert_{(P_{\mathrm{ex}}^{\mathrm{bd}})^{-1}}\le\lVert v\rVert_{(P_{\mathrm{sc}}^{\mathrm{bd}})^{-1}}\le\sqrt{b_{\mathrm{bd}}}\lVert v\rVert_{(P_{\mathrm{ex}}^{\mathrm{bd}})^{-1}} a bd ∥ v ∥ ( P ex bd ) − 1 ≤ ∥ v ∥ ( P sc bd ) − 1 ≤ b bd ∥ v ∥ ( P ex bd ) − 1 . Combining these estimates with (52) yields a b d φ ∥ x ∥ P s c ≤ ∥ A r e d x ∥ ( P s c b d ) − 1 ≤ φ b b d ∥ x ∥ P s c \frac{a_{\mathrm{bd}}}{\varphi}\lVert x\rVert_{P_{\mathrm{sc}}}\le\lVert A_{\mathrm{red}}x\rVert_{(P_{\mathrm{sc}}^{\mathrm{bd}})^{-1}}\le\varphi b_{\mathrm{bd}}\lVert x\rVert_{P_{\mathrm{sc}}} φ a bd ∥ x ∥ P sc ≤ ∥ A red x ∥ ( P sc bd ) − 1 ≤ φ b bd ∥ x ∥ P sc . The matrix ( P b d s c ) − 1 / 2 A r e d ( P b d s c ) − 1 / 2 (P_{\mathrm{bd}}^{\mathrm{sc}})^{-1/2}A_{\mathrm{red}}(P_{\mathrm{bd}}^{\mathrm{sc}})^{-1/2} ( P bd sc ) − 1/2 A red ( P bd sc ) − 1/2 is symmetric, so the same norm bounds imply (54) .
A direct block multiplication gives the exact factorization identity
( P l t , r e d e x ) − 1 A r e d = [ I K m − 1 L 0 I ] . (55) (\mathcal{P}^{\mathrm{ex}}_{\mathrm{lt,red}})^{-1}A_{\mathrm{red}}=
\begin{bmatrix}
I & K_m^{-1}L\\
0 & I
\end{bmatrix}. \tag*{(55)} ( P lt , red ex ) − 1 A red = [ I 0 K m − 1 L I ] . (55)
In particular, every eigenvalue is equal to one and ( ( P l t , r e d e x ) − 1 A r e d − I ) 2 = 0 \left((\mathcal{P}^{\mathrm{ex}}_{\mathrm{lt,red}})^{-1}A_{\mathrm{red}}-I\right)^2=0 ( ( P lt , red ex ) − 1 A red − I ) 2 = 0 . With exact block solves, (3.21) implies that GMRES converges in at most two iterations in exact arithmetic. To analyze the exact and Schur-consistent triangular factors in a common framework, we derive weighted FOV and operator-norm bounds. For any SPD matrix N N N , write ( x , y ) N : = x ⊤ N y (x,y)_N:=x^{\top}Ny ( x , y ) N := x ⊤ N y , and ∥ x ∥ N 2 : = ( x , x ) N \|x\|_N^2:=(x,x)_N ∥ x ∥ N 2 := ( x , x ) N .
Theorem 3.12. Let N e x : = P b d e x N_{\mathrm{ex}}:=\mathcal{P}^{\mathrm{ex}}_{\mathrm{bd}} N ex := P bd ex and L e x : = ( P l t , r e d e x ) − 1 L_{\mathrm{ex}}:=(\mathcal{P}^{\mathrm{ex}}_{\mathrm{lt,red}})^{-1} L ex := ( P lt , red ex ) − 1 . Then, for every nonzero vector x x x ,
( L e x A r e d x , x ) N e x ∥ x ∥ N e x 2 ≥ 1 2 , ∥ L e x A r e d x ∥ N e x ∥ x ∥ N e x ≤ φ . \frac{(L_{\mathrm{ex}}A_{\mathrm{red}}x,x)_{N_{\mathrm{ex}}}}{\|x\|_{N_{\mathrm{ex}}}^{2}}\geq\frac{1}{2},\qquad
\frac{\|L_{\mathrm{ex}}A_{\mathrm{red}}x\|_{N_{\mathrm{ex}}}}{\|x\|_{N_{\mathrm{ex}}}}\leq\varphi. ∥ x ∥ N ex 2 ( L ex A red x , x ) N ex ≥ 2 1 , ∥ x ∥ N ex ∥ L ex A red x ∥ N ex ≤ φ .
Proof. With H from (3.19) and y = N e x 1 / 2 x = ( y m , y p ) ⊤ y=N_{\mathrm{ex}}^{1/2}x=(y_m,y_p)^{\top} y = N ex 1/2 x = ( y m , y p ) ⊤ , set T l t e x : = N e x 1 / 2 L e x A r e d N e x − 1 / 2 T_{\mathrm{lt}}^{\mathrm{ex}}:=N_{\mathrm{ex}}^{1/2}L_{\mathrm{ex}}A_{\mathrm{red}}N_{\mathrm{ex}}^{-1/2} T lt ex := N ex 1/2 L ex A red N ex − 1/2 . By (3.21),
T l t e x = [ I H 0 I ] T_{\mathrm{lt}}^{\mathrm{ex}}=
\begin{bmatrix}
I & H\\
0 & I
\end{bmatrix} T lt ex = [ I 0 H I ]
and ∥ H ∥ 2 ≤ 1 \|H\|_2\leq1 ∥ H ∥ 2 ≤ 1 . Set a = ∥ y m ∥ 2 a=\|y_m\|_2 a = ∥ y m ∥ 2 and b = ∥ y p ∥ 2 b=\|y_p\|_2 b = ∥ y p ∥ 2 . Then
( L e x A r e d x , x ) N e x = a 2 + ( H y p , y m ) + b 2 ≥ a 2 − a b + b 2 ≥ 1 2 ( a 2 + b 2 ) = 1 2 ∥ x ∥ N e x 2 , (L_{\mathrm{ex}}A_{\mathrm{red}}x,x)_{N_{\mathrm{ex}}}=a^2+(Hy_p,y_m)+b^2\geq a^2-ab+b^2\geq\frac{1}{2}(a^2+b^2)=\frac{1}{2}\|x\|_{N_{\mathrm{ex}}}^2, ( L ex A red x , x ) N ex = a 2 + ( H y p , y m ) + b 2 ≥ a 2 − ab + b 2 ≥ 2 1 ( a 2 + b 2 ) = 2 1 ∥ x ∥ N ex 2 ,
∥ L e x A r e d x ∥ N e x 2 = ∥ y m + H y p ∥ 2 2 + b 2 ≤ ( a + b ) 2 + b 2 ≤ φ 2 ( a 2 + b 2 ) = φ 2 ∥ x ∥ N e x 2 . \|L_{\mathrm{ex}}A_{\mathrm{red}}x\|_{N_{\mathrm{ex}}}^2=\|y_m+Hy_p\|_2^2+b^2\leq(a+b)^2+b^2\leq\varphi^2(a^2+b^2)=\varphi^2\|x\|_{N_{\mathrm{ex}}}^2. ∥ L ex A red x ∥ N ex 2 = ∥ y m + H y p ∥ 2 2 + b 2 ≤ ( a + b ) 2 + b 2 ≤ φ 2 ( a 2 + b 2 ) = φ 2 ∥ x ∥ N ex 2 .
The second estimate follows by taking square roots. □ \square □
Before estimating the Schur-consistent triangularly preconditioned operator itself, we record that the exact and Schur-consistent lower-triangular factors are uniformly comparable at the spectral level.
Corollary 3.13. The exact and Schur-consistent lower-triangular factors satisfy
spec ( ( P l t s c ) − 1 P l t e x ) ⊂ [ min { c m , α s c } , max { C m , β s c } ] . \operatorname{spec}\left((\mathcal{P}^{\mathrm{sc}}_{\mathrm{lt}})^{-1}\mathcal{P}^{\mathrm{ex}}_{\mathrm{lt}}\right)\subset\left[\min\{c_m,\alpha_{\mathrm{sc}}\},\max\{C_m,\beta_{\mathrm{sc}}\}\right]. spec ( ( P lt sc ) − 1 P lt ex ) ⊂ [ min { c m , α sc } , max { C m , β sc } ] .
Proof. The product ( P l t s c ) − 1 P l t e x (\mathcal{P}^{\mathrm{sc}}_{\mathrm{lt}})^{-1}\mathcal{P}^{\mathrm{ex}}_{\mathrm{lt}} ( P lt sc ) − 1 P lt ex is block lower triangular with diagonal blocks D m − 1 K m D_m^{-1}K_m D m − 1 K m and P s c − 1 S p P_{\mathrm{sc}}^{-1}S_p P sc − 1 S p , so its spectrum is the union of their spectra. Theorem 3.3 and Theorem 3.9 bound these spectra by [ c m , C m ] [c_m,C_m] [ c m , C m ] and [ α s c , β s c ] [\alpha_{\mathrm{sc}},\beta_{\mathrm{sc}}] [ α sc , β sc ] , respectively, proving the stated inclusion. □ \square □
The relative spectral comparison of the two triangular factors does not by itself yield a GMRES estimate; the following theorem therefore derives a direct FOV bound for the Schur-consistent factor. Let c m c_m c m be as in Theorem 3.3, and set N s c : = P b d s c N_{\mathrm{sc}}:=\mathcal{P}^{\mathrm{sc}}_{\mathrm{bd}} N sc := P bd sc and L s c : = ( P l t , r e d s c ) − 1 L_{\mathrm{sc}}:=(\mathcal{P}^{\mathrm{sc}}_{\mathrm{lt,red}})^{-1} L sc := ( P lt , red sc ) − 1 .
Theorem 3.14. For every nonzero vector x x x ,
( L s c A r e d x , x ) N s c ∥ x ∥ N s c 2 ≥ Σ s c , ∥ L s c A r e d x ∥ N s c ∥ x ∥ N s c ≤ 3 , (56) \frac{(L_{\mathrm{sc}}A_{\mathrm{red}}x,x)_{N_{\mathrm{sc}}}}{\|x\|_{N_{\mathrm{sc}}}^{2}}\geq\Sigma_{\mathrm{sc}},\qquad
\frac{\|L_{\mathrm{sc}}A_{\mathrm{red}}x\|_{N_{\mathrm{sc}}}}{\|x\|_{N_{\mathrm{sc}}}}\leq3, \tag*{(56)} ∥ x ∥ N sc 2 ( L sc A red x , x ) N sc ≥ Σ sc , ∥ x ∥ N sc ∥ L sc A red x ∥ N sc ≤ 3 , (56) where Σ s c : = ( 1 − 1 / 2 ) min { c m , 1 } > 0 \Sigma_{\mathrm{sc}}:=(1-1/\sqrt{2})\min\{c_m,1\}>0 Σ sc := ( 1 − 1/ 2 ) min { c m , 1 } > 0 .
Proof. Set R : = D m − 1 / 2 K m D m − 1 / 2 R:=D_m^{-1/2}K_mD_m^{-1/2} R := D m − 1/2 K m D m − 1/2 and G : = D m − 1 / 2 L P s c − 1 / 2 G:=D_m^{-1/2}LP_{\mathrm{sc}}^{-1/2} G := D m − 1/2 L P sc − 1/2 . The proof of Theorem 3.3 gives the sharper upper bound K m ⪯ 2 D m K_m\preceq2D_m K m ⪯ 2 D m , so c m I ⪯ R ⪯ 2 I c_mI\preceq R\preceq2I c m I ⪯ R ⪯ 2 I . Moreover, T p ⪰ 0 T_p\succeq0 T p ⪰ 0 and (3.13) imply G ⊤ G = P s c − 1 / 2 Q D P s c − 1 / 2 ⪯ I G^{\top}G=P_{\mathrm{sc}}^{-1/2}Q_DP_{\mathrm{sc}}^{-1/2}\preceq I G ⊤ G = P sc − 1/2 Q D P sc − 1/2 ⪯ I , and G ⊤ R G ⪯ 2 G ⊤ G ⪯ 2 I G^{\top}RG\preceq2G^{\top}G\preceq2I G ⊤ R G ⪯ 2 G ⊤ G ⪯ 2 I . A block calculation yields T l t s c : = N s c 1 / 2 L s c A r e d N s c − 1 / 2 = [ R G G ⊤ ( R − I ) I ] T_{\mathrm{lt}}^{\mathrm{sc}}:=N_{\mathrm{sc}}^{1/2}L_{\mathrm{sc}}A_{\mathrm{red}}N_{\mathrm{sc}}^{-1/2}=\begin{bmatrix} R & G\\ G^{\top}(R-I) & I \end{bmatrix} T lt sc := N sc 1/2 L sc A red N sc − 1/2 = [ R G ⊤ ( R − I ) G I ] . For y = Λ s c 1 / 2 x = ( y m , y p ) T y=\Lambda_{\mathrm{sc}}^{1/2}x=(y_m,y_p)^{\mathsf T} y = Λ sc 1/2 x = ( y m , y p ) T , let a 2 = ( R y m , y m ) a^2=(Ry_m,y_m) a 2 = ( R y m , y m ) and b = ∥ y p ∥ 2 b=\lVert y_p\rVert_2 b = ∥ y p ∥ 2 . Since R R R is symmetric, we have ( G y p , y m ) + ( G T ( R − I ) y m , y p ) = ( G y p , R y m ) (Gy_p,y_m)+(\mathrm{G}^{\mathsf T}(R-I)y_m,y_p)=(Gy_p,Ry_m) ( G y p , y m ) + ( G T ( R − I ) y m , y p ) = ( G y p , R y m ) . The Cauchy–Schwarz inequality in the R R R -inner product therefore gives
∣ ( G y p , R y m ) ∣ ≤ a ( G T R G y p , y p ) 1 / 2 ≤ 2 a b . \lvert(Gy_p,Ry_m)\rvert\leq a(\mathrm{G}^{\mathsf T}RGy_p,y_p)^{1/2}\leq\sqrt{2}ab. ∣( G y p , R y m )∣ ≤ a ( G T R G y p , y p ) 1/2 ≤ 2 ab .
Consequently, we obtain ( L s c A r e d x , x ) s c ≥ ( 1 − 1 2 ) ( a 2 + b 2 ) ≥ Σ s c ∥ x ∥ s c 2 (\mathcal{L}_{\mathrm{sc}}\mathcal{A}_{\mathrm{red}}x,x)_{\mathrm{sc}}\geq\left(1-\frac{1}{\sqrt{2}}\right)(a^2+b^2)\geq\Sigma_{\mathrm{sc}}\lVert x\rVert_{\mathrm{sc}}^2 ( L sc A red x , x ) sc ≥ ( 1 − 2 1 ) ( a 2 + b 2 ) ≥ Σ sc ∥ x ∥ sc 2 , where the last inequality follows from R ⪰ c m I R\succeq c_mI R ⪰ c m I and ∥ y ∥ 2 2 = ∥ x ∥ s c 2 \lVert y\rVert_2^2=\lVert x\rVert_{\mathrm{sc}}^2 ∥ y ∥ 2 2 = ∥ x ∥ sc 2 .
For the operator norm, 0 ≺ R ⪯ 2 I 0\prec R\preceq2I 0 ≺ R ⪯ 2 I gives ∥ R − I ∥ 2 ≤ 1 \lVert R-I\rVert_2\leq1 ∥ R − I ∥ 2 ≤ 1 , while G T G ⪯ I \mathrm{G}^{\mathsf T}\mathrm{G}\preceq I G T G ⪯ I gives ∥ G ∥ 2 ≤ 1 \lVert G\rVert_2\leq1 ∥ G ∥ 2 ≤ 1 . Therefore, we obtain
∥ T l t s c ∥ 2 ≤ ∥ [ R 0 0 I ] ∥ 2 + ∥ [ 0 G G T ( R − I ) 0 ] ∥ 2 = max { ∥ R ∥ 2 , 1 } + max { ∥ G ∥ 2 , ∥ G T ( R − I ) ∥ 2 } ≤ 3. \lVert T_{\mathrm{lt}}^{\mathrm{sc}}\rVert_2\leq\left\lVert\begin{bmatrix}R&0\\0&I\end{bmatrix}\right\rVert_2+\left\lVert\begin{bmatrix}0&G\\\mathrm{G}^{\mathsf T}(R-I)&0\end{bmatrix}\right\rVert_2
=\max\{\lVert R\rVert_2,1\}+\max\{\lVert G\rVert_2,\lVert\mathrm{G}^{\mathsf T}(R-I)\rVert_2\}\leq3. ∥ T lt sc ∥ 2 ≤ [ R 0 0 I ] 2 + [ 0 G T ( R − I ) G 0 ] 2 = max {∥ R ∥ 2 , 1 } + max {∥ G ∥ 2 , ∥ G T ( R − I ) ∥ 2 } ≤ 3.
∥ L s c A r e d x ∥ s c = ∥ T s c l t y ∥ 2 ≤ 3 ∥ y ∥ 2 = 3 ∥ x ∥ s c \lVert\mathcal{L}_{\mathrm{sc}}\mathcal{A}_{\mathrm{red}}x\rVert_{\mathrm{sc}}=\lVert T_{\mathrm{sc}}^{\mathrm{lt}}y\rVert_2\leq3\lVert y\rVert_2=3\lVert x\rVert_{\mathrm{sc}} ∥ L sc A red x ∥ sc = ∥ T sc lt y ∥ 2 ≤ 3 ∥ y ∥ 2 = 3 ∥ x ∥ sc , which completes the proof.
Practical implementation of the block preconditioners
The exact block inverses appearing in the preceding preconditioners are not formed explicitly. For the stabilized formulation, we use a vertex-patch two-grid approximation for the displacement block, a macrocell mass inverse for the rotation block, and a scalar two-grid approximation for the pressure block.
For the uncoupled preconditioners in (3.2), we use
B u − 1 ≈ K u u − 1 , B γ − 1 ≈ ( μ M γ ) − 1 ≈ K γ γ − 1 , B p − 1 ≈ T p − 1 or B p − 1 ≈ ( P m z ) − 1 . B_u^{-1}\approx K_{uu}^{-1},\qquad B_\gamma^{-1}\approx(\mu M_\gamma)^{-1}\approx K_{\gamma\gamma}^{-1},\qquad B_p^{-1}\approx T_p^{-1}\quad\text{or}\quad B_p^{-1}\approx(P_m^z)^{-1}. B u − 1 ≈ K uu − 1 , B γ − 1 ≈ ( μ M γ ) − 1 ≈ K γ γ − 1 , B p − 1 ≈ T p − 1 or B p − 1 ≈ ( P m z ) − 1 .
The vector P 1 \mathcal{P}^1 P 1 coarse basis comprises componentwise samples of the physical-coordinate modes 1 , x 1 , … , x d 1,x_1,\ldots,x_d 1 , x 1 , … , x d at macrocell centers within each nonoverlapping vertex-centered aggregate. Each vertex-patch Gauss–Seidel block contains displacement degrees of freedom of macrocells sharing a vertex. By Corollary 3.4, the stabilized rotation block is uniformly mass-equivalent. For the pressure blocks, we use two-grid solvers with forward/backward Gauss–Seidel pre-/post-smoothing.
The exact Schur complement S p S_p S p is used only in the analysis. In computation, the stabilized mechanics inverse is approximated by B m − 1 : = diag ( B u − 1 , B γ − 1 ) B_m^{-1}:=\operatorname{diag}(B_u^{-1},B_\gamma^{-1}) B m − 1 := diag ( B u − 1 , B γ − 1 ) , and the practical Schur-consistent pressure block is P ^ s c : = T p + L T B m − 1 L \widehat{P}_{\mathrm{sc}}:=T_p+L^{\mathsf T}B_m^{-1}L P sc := T p + L T B m − 1 L . After this pressure matrix is assembled, its inverse action is approximated by the scalar pressure two-grid solver. The lower-triangular preconditioner uses the same block solvers together with one lower-left coupling update in the usual block forward substitution.
Numerical experiments
This section presents 2D and 3D numerical tests demonstrating the robustness of the proposed algorithms with respect to physical and discretization parameters. Throughout, we use isotropic permeability K = κ I K=\kappa I K = κ I , backward Euler time stepping, and the rotation-jump stabilization parameter δ γ = 0.5 \delta_\gamma=0.5 δ γ = 0.5 . The condensed control-volume three-field systems are solved by preconditioned GMRES with relative residual tolerances 10 − 8 10^{-8} 1 0 − 8 in 2D and 10 − 6 10^{-6} 1 0 − 6 in 3D. In the iteration tables, an entry a ( b ) a\ (b) a ( b ) gives the GMRES iteration counts for P b d P m z \mathcal{P}_{\mathrm{bd}}^{P_m^z} P bd P m z and P b d s c \mathcal{P}_{\mathrm{bd}}^{\mathrm{sc}} P bd sc , respectively. We mainly present results for P b d P m z \mathcal{P}_{\mathrm{bd}}^{P_m^z} P bd P m z and P b d s c \mathcal{P}_{\mathrm{bd}}^{\mathrm{sc}} P bd sc . For the tested parameter ranges, P l t T p \mathcal{P}_{\mathrm{lt}}^{T_p} P lt T p and P l t s c \mathcal{P}_{\mathrm{lt}}^{\mathrm{sc}} P lt sc yield iteration counts very close to those of P b d P m z \mathcal{P}_{\mathrm{bd}}^{P_m^z} P bd P m z and P b d s c \mathcal{P}_{\mathrm{bd}}^{\mathrm{sc}} P bd sc , respectively; hence, the results for the former two are omitted for brevity.
Example 4.1 (2D manufactured solution on rectangular meshes). We consider a manufactured solution on Ω = [ 0 , 1 ] 2 \Omega=[0,1]^2 Ω = [ 0 , 1 ] 2 . The exact pressure is p ( x , y , t ) = p(x,y,t)= p ( x , y , t ) = e − t sin ( π x ) sin ( π y ) e^{-t}\sin(\pi x)\sin(\pi y) e − t sin ( π x ) sin ( π y ) . The displacement components are u 1 = e − t 2 [ sin 2 ( π x ) sin ( 2 π y ) + λ − 1 sin ( 2 π x ) sin 2 ( π y ) ] u_1 = \frac{e^{-t}}{2}[\sin^2(\pi x)\sin(2\pi y)+\lambda^{-1}\sin(2\pi x)\sin^2(\pi y)] u 1 = 2 e − t [ sin 2 ( π x ) sin ( 2 π y ) + λ − 1 sin ( 2 π x ) sin 2 ( π y )] and u 2 = e − t 2 [ − sin ( 2 π x ) sin 2 ( π y ) + sin 2 ( π x ) sin ( 2 π y ) ] u_2 = \frac{e^{-t}}{2}[-\sin(2\pi x)\sin^2(\pi y)+\sin^2(\pi x)\sin(2\pi y)] u 2 = 2 e − t [ − sin ( 2 π x ) sin 2 ( π y ) + sin 2 ( π x ) sin ( 2 π y )] . The remaining variables γ \gamma γ , z z z , σ \sigma σ , and the source terms are obtained from the Biot equations. The final time is T = 0.04 T=0.04 T = 0.04 . The reduced linear systems in the accuracy study are solved with the P b d z ~ P_{\mathrm{bd}}^{\tilde z} P bd z ~ preconditioner. Figure 1 checks the approximation accuracy and the locking-free behavior of the discretization. All plotted errors are absolute L 2 L^2 L 2 errors. The macroelement errors e M σ e_{M\sigma} e M σ and e M z e_{Mz} e M z are computed by replacing the four subcell values of σ h \sigma_h σ h and z h z_h z h on each macroelement by their macroelement averages and then taking the corresponding L 2 L^2 L 2 errors against the exact fields. The observed errors e u , e γ , e p e_u,e_\gamma,e_p e u , e γ , e p are O ( h 2 ) O(h^2) O ( h 2 ) , whereas the stress and Darcy-velocity errors e σ , e z e_\sigma,e_z e σ , e z are O ( h ) O(h) O ( h ) . The macroelement-averaged stress and Darcy velocity exhibit second-order superconvergence. Since the same rates are retained for λ = 10 6 \lambda=10^6 λ = 1 0 6 , including the low-permeability case κ = 10 − 6 \kappa=10^{-6} κ = 1 0 − 6 , the test indicates locking-free accuracy and no visible deterioration of the discretization error in these parameter regimes.
Figure 1. Accuracy results for Example 4.1. Left: λ = κ = 1 \lambda=\kappa=1 λ = κ = 1 ; center: λ = 10 6 \lambda=10^6 λ = 1 0 6 , κ = 1 \kappa=1 κ = 1 ; right: λ = 10 6 \lambda=10^6 λ = 1 0 6 , κ = 10 − 6 \kappa=10^{-6} κ = 1 0 − 6 .
We next assess the parameter robustness of P b d z ~ P_{\mathrm{bd}}^{\tilde z} P bd z ~ and P b d s c P_{\mathrm{bd}}^{\mathrm{sc}} P bd sc . Table 1 reports the GMRES iteration counts. Both preconditioners converge to the prescribed tolerance for all tested physical and discretization parameters. The counts remain between 25 and 35 across the tested meshes and parameter combinations. The Schur-consistent block P b d s c P_{\mathrm{bd}}^{\mathrm{sc}} P bd sc never requires more outer iterations than the uncoupled block P b d z ~ P_{\mathrm{bd}}^{\tilde z} P bd z ~ , although the difference is generally small. To compare this iteration reduction with the actual computational cost, Table 2 reports the GMRES CPU time (in seconds) for one backward Euler step. The ratio t s c / t z ~ t_{\mathrm{sc}}/t_{\tilde z} t sc / t z ~ is the Schur-consistent solve time divided by the uncoupled solve time. The Schur-consistent preconditioner is about 4.7–10.4 times more expensive in the reported tests because each outer application contains an inner pressure solve. Consequently, the less expensive pressure block of P b d z ~ P_{\mathrm{bd}}^{\tilde z} P bd z ~ gives the lower overall cost despite its slightly larger iteration counts.
Table 1. GMRES iterations for Example 4.1, with c 0 = 1 c_0=1 c 0 = 1 .
Table 2. GMRES CPU time (s) and time ratio for Example 4.1.
We further test the critically balanced scaling c 0 = K d r − 1 c_0=K_{\mathrm{dr}}^{-1} c 0 = K dr − 1 , with μ = α B = 1 \mu=\alpha_B=1 μ = α B = 1 , Parameter family 1: λ = κ = c 0 = 1 \lambda= \kappa= c_0 = 1 λ = κ = c 0 = 1 , Δ t = 10 − 2 \Delta t = 10^{-2} Δ t = 1 0 − 2 ; Parameter family 2: λ = 10 6 \lambda= 10^6 λ = 1 0 6 , κ = 10 − 6 \kappa= 10^{-6} κ = 1 0 − 6 , c 0 = 10 − 12 c_0 = 10^{-12} c 0 = 1 0 − 12 , Δ t = 10 − 4 \Delta t = 10^{-4} Δ t = 1 0 − 4 .
κ = 10 − 6 \kappa= 10^{-6} κ = 1 0 − 6 , and Δ t = 10 − 3 \Delta t = 10^{-3} Δ t = 1 0 − 3 . In 2D, K d r = λ + 1 K_{\mathrm{dr}} = \lambda+ 1 K dr = λ + 1 , so this choice gives χ λ , c 0 = 1 \chi_{\lambda,c_0} = 1 χ λ , c 0 = 1 while c 0 → 0 c_0 \to0 c 0 → 0 as λ → ∞ \lambda\to\infty λ → ∞ . Table 3 shows that all cases converge. The counts remain between 27 and 65, with no iteration blow-up as c 0 c_0 c 0 approaches zero. These results support the robustness predicted for the balanced storage scaling.
Table 3. GMRES iterations along the path c 0 = K d r − 1 c_0 = K_{\mathrm{dr}}^{-1} c 0 = K dr − 1 for Example 4.1.
Example 4.2
Example 4.2 (2D manufactured solution on perturbed quadrilateral meshes). We repeat the manufactured-solution test of Example 4.1, now on perturbed quadrilateral meshes. Let ( ξ i , η j ) = ( i h , j h ) (\xi_i, \eta_j) = (ih, jh) ( ξ i , η j ) = ( ih , j h ) , where h = 1 / N h = 1/N h = 1/ N , denote the vertices of the uniform logical grid. The physical vertices are defined by the smooth mapping
x i j = ξ i + h 2 sin ( 2 π ξ i ) sin ( 2 π η j ) , y i j = η j + h 2 sin ( 2 π ξ i ) sin ( 2 π η j ) . x_{ij} = \xi_i + h^2 \sin(2\pi\xi_i)\sin(2\pi\eta_j), \qquad y_{ij} = \eta_j + h^2 \sin(2\pi\xi_i)\sin(2\pi\eta_j). x ij = ξ i + h 2 sin ( 2 π ξ i ) sin ( 2 π η j ) , y ij = η j + h 2 sin ( 2 π ξ i ) sin ( 2 π η j ) .
Thus, the boundary of Ω = ( 0 , 1 ) 2 \Omega= (0,1)^2 Ω = ( 0 , 1 ) 2 is unchanged, whereas the interior vertices are displaced by O ( h 2 ) O(h^2) O ( h 2 ) , producing a perturbed family of general quadrilateral meshes. We use the same exact solution and fixed parameters μ = α B = 1 \mu= \alpha_B = 1 μ = α B = 1 as in Example 4.1. For both practical preconditioners, the displacement block retains the general-quadrilateral two-grid solver. This experiment therefore isolates the effect of the O ( h 2 ) O(h^2) O ( h 2 ) quadrilateral mesh perturbation on the performance of the stabilized formulation and its practical preconditioners. Table 4 shows that, across the tested meshes and parameter combinations, the uncoupled and Schur-consistent preconditioners require 26–104 and 25–53 iterations, respectively. No systematic iteration growth is observed as the mesh is refined and the parameters vary.
Table 4. GMRES iterations for Example 4.2: Δ t = 10 − 2 \Delta t = 10^{-2} Δ t = 1 0 − 2 .
Example 4.3 (Barry–Mercer localized-source problem). We then consider a 2D Barry–Mercer benchmark to illustrate the pressure behavior in a low-permeability, small-time-step regime. The domain is Ω = ( 0 , 1 ) 2 \Omega= (0,1)^2 Ω = ( 0 , 1 ) 2 , the initial data are zero, and the point source at ( x 0 , y 0 ) = ( 1 / 4 , 1 / 4 ) (x_0,y_0) = (1/4,1/4) ( x 0 , y 0 ) = ( 1/4 , 1/4 ) is represented by q h n = Q ( t n ) ∣ M 0 ∣ − 1 χ M 0 q_h^n = Q(t_n)|M_0|^{-1}\chi_{M_0} q h n = Q ( t n ) ∣ M 0 ∣ − 1 χ M 0 , where Q ( t ) = cos ( 2 π t ) Q(t) = \cos(2\pi t) Q ( t ) = cos ( 2 π t ) and M 0 M_0 M 0 is the source cell. Homogeneous boundary data are imposed for displacement and pressure. We let E E E and ν \nu ν denote Young’s modulus and Poisson’s ratio, respectively. Under the plane-strain assumption, the Lamé coefficients are λ = E ν / ( ( 1 + ν ) ( 1 − 2 ν ) ) \lambda= E\nu/((1+\nu)(1-2\nu)) λ = E ν / (( 1 + ν ) ( 1 − 2 ν )) and μ = E / ( 2 ( 1 + ν ) ) \mu= E/(2(1+\nu)) μ = E / ( 2 ( 1 + ν )) . Following the challenging parameter regime used in [34 ] , we take E = 10 5 E = 10^5 E = 1 0 5 , ν = 0.1 \nu= 0.1 ν = 0.1 , α B = 1 \alpha_B = 1 α B = 1 , c 0 = 10 − 8 c_0 = 10^{-8} c 0 = 1 0 − 8 , Δ t = 10 − 4 \Delta t = 10^{-4} Δ t = 1 0 − 4 , and κ = 10 − 6 \kappa= 10^{-6} κ = 1 0 − 6 . Figure 2 shows the computed pressure at T = 10 − 4 T = 10^{-4} T = 1 0 − 4 on the 64 × 64 64 \times64 64 × 64 mesh. The left panel gives the full pressure surface, whereas the right panel is the horizontal cut through the source row. Although the low permeability produces a sharply localized pressure peak, the pressure cut exhibits a clean single-peak profile followed by a smooth decay. The computed pressure remains positive, and no visible cell-to-cell saw-tooth oscillations occur near or away from the source. This indicates that the proposed control-volume formulation can effectively suppress nonphysical pressure oscillations without a separate pressure stabilization term in the flow equation.
Figure 2. Pressure behavior for Example 4.3 on the 64 × 64 64 \times64 64 × 64 mesh.
We also use this localized-source problem to probe the parameter behavior of the two block-diagonal preconditioners. We keep E = 10 5 E = 10^5 E = 1 0 5 , α B = 1 \alpha_B = 1 α B = 1 , and Δ t = T = 10 − 4 \Delta t = T = 10^{-4} Δ t = T = 1 0 − 4 , and choose ν ∈ { 0.1 , 0.4 , 0.49 } \nu\in\{0.1, 0.4, 0.49\} ν ∈ { 0.1 , 0.4 , 0.49 } . For the comparison, we fix κ = 10 − 6 \kappa= 10^{-6} κ = 1 0 − 6 and c 0 = 10 − 8 c_0 = 10^{-8} c 0 = 1 0 − 8 . Table 5 reports the reduced-system GMRES counts. Both preconditioners converge to the prescribed tolerance. As λ \lambda λ increases over the tested range, the uncoupled iteration counts decrease from 96 to 46 at N = 40 N = 40 N = 40 and from 84 to 40 at N = 160 N = 160 N = 160 . The Schur-consistent block-diagonal preconditioner remains particularly stable, requiring only 30–35 outer iterations throughout the reported parameter range.
Table 5. GMRES iterations for Example 4.3.
Example 4.4 (3D manufactured solution on structured cubical meshes). On
Ω = ( 0 , 1 ) 3 \Omega= (0,1)^3 Ω = ( 0 , 1 ) 3 , we take p = e − t sin ( π x ) sin ( π y ) sin ( π z ) p=e^{-t}\sin(\pi x)\sin(\pi y)\sin(\pi z) p = e − t sin ( π x ) sin ( π y ) sin ( π z ) , u 1 = p u_1=p u 1 = p , u 2 = 2 p cos ( π x ) u_2=2p\cos(\pi x) u 2 = 2 p cos ( π x ) , and u 3 = 2 p cos ( π y ) u_3=2p\cos(\pi y) u 3 = 2 p cos ( π y ) . The initial pressure and stress are obtained from these fields at t = 0 t=0 t = 0 . We set μ = α B = 1 \mu=\alpha_B=1 μ = α B = 1 and Δ t = 10 − 3 \Delta t=10^{-3} Δ t = 1 0 − 3 . Table 6 reports iteration counts for the block preconditioners applied to the three-field system as the physical parameters λ \lambda λ , κ \kappa κ , and c 0 c_0 c 0 vary. Both preconditioners reach the prescribed tolerance. Thus, neither large λ \lambda λ nor small κ \kappa κ causes an iteration blow-up. Even for c 0 = 0 c_0=0 c 0 = 0 , P b d z m \mathcal{P}_{\mathrm{bd}}^{zm} P bd z m still converges in 15–57 iterations.
Table 6. GMRES iterations for Example 4.4.
Table 7 reports the GMRES CPU times for two selected parameter families. The outer iteration counts of P b d s c \mathcal{P}_{\mathrm{bd}}^{sc} P bd sc do not exceed those of P b d z m \mathcal{P}_{\mathrm{bd}}^{zm} P bd z m in these tests. Nevertheless, P b d z m \mathcal{P}_{\mathrm{bd}}^{zm} P bd z m requires less CPU time on every tested mesh, with t s c / t z m t_{sc}/t_{zm} t sc / t z m ranging from 5.12 to 9.54.
Table 7. GMRES CPU time (s) and time ratio for Example 4.4.
Parameter family 1: λ = 10 2 \lambda=10^2 λ = 1 0 2 , κ = 10 − 2 \kappa=10^{-2} κ = 1 0 − 2 , c 0 = 1 c_0=1 c 0 = 1 , Δ t = 10 − 3 \Delta t=10^{-3} Δ t = 1 0 − 3 ; Parameter family 2: λ = 10 6 \lambda=10^6 λ = 1 0 6 , κ = 10 − 6 \kappa=10^{-6} κ = 1 0 − 6 , c 0 = 0 c_0=0 c 0 = 0 , Δ t = 10 − 3 \Delta t=10^{-3} Δ t = 1 0 − 3 .
Example 4.5 (2D and 3D footing problems). The footing problem is a classical Biot benchmark [18 ] . We follow the 2D initial–boundary value setting in [20 ] on Ω = ( − 50 , 50 ) × ( 0 , 75 ) \Omega=(-50,50)\times(0,75) Ω = ( − 50 , 50 ) × ( 0 , 75 ) , with downward traction σ n = ( 0 , − σ 0 ) ⊤ \sigma n=(0,-\sigma_0)^\top σ n = ( 0 , − σ 0 ) ⊤ , σ 0 = 10 4 \sigma_0=10^4 σ 0 = 1 0 4 , on Γ 1 = { ( x , 75 ) : ∣ x ∣ ≤ 50 / 3 } \Gamma_1=\{(x,75):|x|\le50/3\} Γ 1 = {( x , 75 ) : ∣ x ∣ ≤ 50/3 } . The remaining top boundary is traction-free, the sides and bottom are fixed, and p = 0 p=0 p = 0 on ∂ Ω \partial\Omega ∂ Ω . The body force, fluid source, and initial data are zero. For the representative solution in Figure 3 , we use the benchmark parameters E = 3 × 10 4 E=3\times10^4 E = 3 × 1 0 4 , ν = 0.4995 \nu=0.4995 ν = 0.4995 , κ = 10 − 4 \kappa=10^{-4} κ = 1 0 − 4 , c 0 = 10 − 3 c_0=10^{-3} c 0 = 1 0 − 3 , and α B = 0.1 \alpha_B=0.1 α B = 0.1 , together with N = 60 N=60 N = 60 , Δ t = 1 \Delta t=1 Δ t = 1 , and final time T = 50 T=50 T = 50 . The snapshots are shown at t = 1 , 10 , 25 , 50 t=1,10,25,50 t = 1 , 10 , 25 , 50 . The black segment on the top boundary marks the loaded region Γ 1 \Gamma_1 Γ 1 . The pressure remains smooth and symmetric, with its largest values beneath Γ 1 \Gamma_1 Γ 1 ; as time advances, its peak decreases and the low-pressure layers adjacent to the drained boundary extend inward. The logarithmic flux plots reveal drainage over several orders of magnitude, with the strongest early flow occurring near the boundary and the two endpoints of Γ 1 \Gamma_1 Γ 1 . These high-flux regions steadily weaken, with max ∣ q n ∣ \max|q_n| max ∣ q n ∣ decreasing from 1.34 × 10 − 5 1.34\times10^{-5} 1.34 × 1 0 − 5 at t = 1 t=1 t = 1 to 2.19 × 10 − 6 2.19\times10^{-6} 2.19 × 1 0 − 6 at t = 50 t=50 t = 50 , and no visible oscillatory pattern develops.
To test parameter robustness, we take Δ t = 10 − 2 \Delta t=10^{-2} Δ t = 1 0 − 2 and fix μ = 1 \mu=1 μ = 1 and α B = 0.1 \alpha_B=0.1 α B = 0.1 . Table 8 reports the GMRES iteration counts. The two preconditioners remain effective
Figure 3. Evolution of the pressure p h p_h p h (top row) and logarithmic Darcy-velocity magnitude log 10 ∣ z h ∣ \log_{10} |z_h| log 10 ∣ z h ∣ (bottom row) for the 2D setting of Example 4.5 at t = 1 , 10 , 25 , 50 t = 1, 10, 25, 50 t = 1 , 10 , 25 , 50 .
as λ \lambda λ , κ \kappa κ , and c 0 c_0 c 0 vary over the reported ranges, including the limiting case c 0 = 0 c_0 = 0 c 0 = 0 ; no parameter choice causes an iteration blow-up.
Table 8. GMRES iterations for the 2D setting of Example 4.5.
We then use the 3D footing geometry of [20 ] , with Ω = ( − 32 , 32 ) 2 × ( 0 , 64 ) \Omega= (-32, 32)^2 \times(0, 64) Ω = ( − 32 , 32 ) 2 × ( 0 , 64 ) and downward traction σ n = ( 0 , 0 , − σ 0 ) ⊤ \boldsymbol{\sigma}n = (0, 0, -\sigma_0)^\top σ n = ( 0 , 0 , − σ 0 ) ⊤ on Γ 1 = { ( x , y , 64 ) : ∣ x ∣ ≤ 16 , ∣ y ∣ ≤ 16 } \Gamma_1 = \{(x, y, 64) : |x| \le16, |y| \le16\} Γ 1 = {( x , y , 64 ) : ∣ x ∣ ≤ 16 , ∣ y ∣ ≤ 16 } . The rest of the top face is traction-free, and the sides and bottom are fixed. The drainage conditions, initial data, body force, and fluid source are the same as in 2D. We use uniform structured cubical meshes with N 3 N^3 N 3 cells and fix c 0 = 0.5 c_0 = 0.5 c 0 = 0.5 , μ = 1 \mu= 1 μ = 1 , α B = 0.5 \alpha_B = 0.5 α B = 0.5 , and σ 0 = 0.1 \sigma_0 = 0.1 σ 0 = 0.1 . Table 9 reports the GMRES iteration counts. The iteration counts range from 15 to 37 for P b d 2 m \mathcal{P}_{\mathrm{bd}}^{2m} P bd 2 m and from 15 to 32 for P b d s c \mathcal{P}_{\mathrm{bd}}^{\mathrm{sc}} P bd sc , with no deterioration as κ \kappa κ and Δ t \Delta t Δ t decrease over the tested ranges. Thus, both preconditioners remain effective under mesh refinement and the reported parameter variations.
Conclusions
We have developed and analyzed block preconditioners for a control-volume discretization of the quasi-static Biot equations. Starting from a five-field formulation, the stress and scaled Darcy velocity are eliminated by independent vertex-local solves, resulting in a symmetric condensed three-field system. We established the well-posedness of the fully discrete scheme and proved that the coupled mechanics block is uniformly spectrally equivalent to its displacement–rotation block-diagonal approximation.
Two classes of pressure preconditioners were then studied. For the inexpensive uncoupled strategy, we identified an exact discrete pressure-control condition and derived a sufficient condition. In particular, the analysis permits the critical scaling c 0 ≍ K d r − 1 c_0 \asymp K_{\mathrm{dr}}^{-1} c 0 ≍ K dr − 1 , so that the storage coefficient may tend to zero as the material approaches the nearly incompressible limit. For the Schur-consistent strategy, we introduced a computable pressure block and proved its uniform equivalence to the exact pressure Schur complement. This gives parameter-independent spectral bounds for Schur-consistent block-diagonal preconditioning and uniform weighted field-of-values estimates for Schur-consistent lower-triangular preconditioning. Practical implementations of the block preconditioners were constructed using a vertex-patch two-grid solver for displacement, a macrocell mass inverse for the stabilized rotation block, and scalar two-grid solvers for pressure. The numerical experiments in 2D and 3D verify the expected accuracy of the control-volume discretization and demonstrate stable performance under mesh refinement and broad variations of the physical parameters.
Table 9. GMRES iterations for the 3D setting of Example 4.5.
####### Conclusions
####### References