Introduction
Poroelasticity models the coupled interaction of solid deformation and pore-fluid flow in a saturated porous medium. Its mathematical foundations trace back to Biot’s theory of consolidation [1 , 26 , 6 ] , which describes the time-dependent compaction of a porous elastic solid under fluid pressure. Systems of this kind arise in geomechanics, including subsurface fluid storage [27 ] , in bone mechanics [25 ] , and in consolidation caused by groundwater withdrawal [13 ] .
Analytical and numerical studies of the linear Biot system include [24 , 20 , 29 ] . A central difficulty in its numerical approximation is volumetric locking : when the solid is nearly incompressible (λ → ∞ \lambda\to\infty λ → ∞ ), standard low-order displacement approximations can lose accuracy. Numerical methods to circumvent this difficulty were proposed including stress-velocity mixed formulation [14 ] , mixed method with reduced integration [28 ] , nonconforming finite element method [15 ] , the total pressure formulation [21 , 17 ] , a stabilized method [22 ] , and different mixed/hybridized-mixed formulations [11 , 5 ] , to name a few. This list of locking-free methods is not exhaustive.
Optimal control of PDE systems is treated, for example, in [10 ] . For poroelasticity, well-posedness of optimal control problems was studied in [2 ] with the zero-storage assumption, while numerical methods for distributed-control problem was studied in [12 ] with general nonnegative storage coefficient. A recent work [9 ] discussed a three-field virtual element method for distributed and Neumann boundary controls for static poroelasticity problems. An extension to thermo-poroelasticity optimal control problems was also studied recently in [16 ] .
As observed in [24 ] , the quasi-static poroelasticity system can be reduced to a parabolic-type evolutionary equation of pressure via mechanical Schur complement. Thus, the analysis of parabolic optimal control problems provides a framework for the space-time discretization and error estimates considered here; see, e.g., [18 , 19 ] . For the quasi-static poroelasticity system, however, the regularity of the state and adjoint variables must be established through the saddle point problem coupling the displacement and total pressure. Moreover, eliminating the mechanical variables from the continuous and discrete formulations leads to different effective mass operators. The resulting consistency error cannot be estimated directly under the available time regularity and requires a duality argument combined with discrete summation by parts. The dependence of these estimates on the storage coefficient and the Lamé parameter must also be controlled.
In this work, we study the distributed optimal control of a linear poroelastic system in the total pressure formulation. The control is a distributed fluid-pressure source, and the objective functional combines running-cost tracking of the displacement and fluid pressure, together with Tikhonov regularization. The main contributions of this paper are as follows.
We prove well-posedness of the differentiated total pressure poroelasticity system for non-negative storage and prove existence and uniqueness of an optimal control.
By a stability analysis of the system via mechanical Schur complement, we characterize the parameter dependence of the effective mass and show that its uniform coercivity is lost in the simultaneous zero-storage (s 0 ↓ 0 s_0 \downarrow0 s 0 ↓ 0 ) and incompressible limit (λ → + ∞ \lambda\to+\infty λ → + ∞ ). Sufficient conditions for uniform coercivity of effective mass are also obtained.
We discretize the state and adjoint systems by stable mixed finite elements in space and the dG(0) method in time, and derive first-order a priori control and state error estimates. Adopting the duality argument of parabolic optimal control problems in [18 , 19 ] we obtain the error estimates with minimal regularity assumptions of solutions. These error estimates are also robust for parameters under the sufficient conditions for uniform coercivity of effective mass.
The remainder of the paper is organized as follows. Section 2 introduces the model, derives the total pressure weak formulation, and establishes well-posedness. Section 3 formulates the optimal control problem, proves existence of optimal controls, and derives first-order necessary optimality conditions. Section 4 introduces the fully discrete scheme and the discrete optimality system. Section 5 presents the a priori error analysis for the discrete controls and states. Section 6 reports convergence, parameter-robustness, and localized pressure-control experiments.
Notation and function spaces
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d , d ∈ { 2 , 3 } d \in\{2,3\} d ∈ { 2 , 3 } , be a bounded domain with Lipschitz boundary ∂ Ω \partial\Omega ∂ Ω . We decompose ∂ Ω = Γ ‾ D ∪ Γ ‾ N \partial\Omega= \overline{\Gamma}_{D} \cup\overline{\Gamma}_{N} ∂ Ω = Γ D ∪ Γ N with Γ D ∩ Γ N = ∅ \Gamma_{D} \cap\Gamma_{N} = \varnothing Γ D ∩ Γ N = ∅ , ∣ Γ D ∣ > 0 |\Gamma_{D}| > 0 ∣ Γ D ∣ > 0 , and ∣ Γ N ∣ > 0 |\Gamma_{N}| > 0 ∣ Γ N ∣ > 0 , where Γ D \Gamma_{D} Γ D is the Dirichlet part of the boundary for the displacement and Γ N \Gamma_{N} Γ N the Neumann (traction-free) part. The time interval is ( 0 , T ) (0,T) ( 0 , T ) with 0 < T < ∞ 0 < T < \infty 0 < T < ∞ .
For k ≥ 0 k \ge0 k ≥ 0 and 1 ≤ r ≤ ∞ 1 \le r \le\infty 1 ≤ r ≤ ∞ , W k , r ( Ω ) W^{k,r}(\Omega) W k , r ( Ω ) denotes the standard Sobolev space on Ω \Omega Ω with norm ∥ ⋅ ∥ W k , r \|\cdot\|_{W^{k,r}} ∥ ⋅ ∥ W k , r , and H k ( Ω ) : = W k , 2 ( Ω ) H^{k}(\Omega) := W^{k,2}(\Omega) H k ( Ω ) := W k , 2 ( Ω ) with norm ∥ ⋅ ∥ H k \|\cdot\|_{H^{k}} ∥ ⋅ ∥ H k and seminorm ∣ ⋅ ∣ H k |\cdot|_{H^{k}} ∣ ⋅ ∣ H k . We write L 2 ( Ω ) : = H 0 ( Ω ) L^{2}(\Omega) := H^{0}(\Omega) L 2 ( Ω ) := H 0 ( Ω ) and denote its inner product by ( ⋅ , ⋅ ) Ω (\cdot,\cdot)_{\Omega} ( ⋅ , ⋅ ) Ω and norm by ∥ ⋅ ∥ 0 : = ∥ ⋅ ∥ L 2 ( Ω ) \|\cdot\|_{0} := \|\cdot\|_{L^{2}(\Omega)} ∥ ⋅ ∥ 0 := ∥ ⋅ ∥ L 2 ( Ω ) . For k ≥ 1 k \ge1 k ≥ 1 we also abbreviate ∥ ⋅ ∥ k : = ∥ ⋅ ∥ H k \|\cdot\|_{k} := \|\cdot\|_{H^{k}} ∥ ⋅ ∥ k := ∥ ⋅ ∥ H k and ∣ ⋅ ∣ k : = ∣ ⋅ ∣ H k |\cdot|_{k} := |\cdot|_{H^{k}} ∣ ⋅ ∣ k := ∣ ⋅ ∣ H k . The space H 0 1 ( Ω ) H_{0}^{1}(\Omega) H 0 1 ( Ω ) is the closure of C c ∞ ( Ω ) C_{c}^{\infty}(\Omega) C c ∞ ( Ω ) in H 1 ( Ω ) H^{1}(\Omega) H 1 ( Ω ) . Its dual is denoted H − 1 ( Ω ) H^{-1}(\Omega) H − 1 ( Ω ) , with duality pairing ⟨ ⋅ , ⋅ ⟩ H − 1 , H 0 1 \langle\cdot,\cdot\rangle_{H^{-1},H_{0}^{1}} ⟨ ⋅ , ⋅ ⟩ H − 1 , H 0 1 . For vector-valued functions we write H k ( Ω ) : = ( H k ( Ω ) ) d \mathbf{H}^{k}(\Omega) := (H^{k}(\Omega))^{d} H k ( Ω ) := ( H k ( Ω ) ) d and use the same notation ( ⋅ , ⋅ ) Ω (\cdot,\cdot)_{\Omega} ( ⋅ , ⋅ ) Ω for the L 2 L^{2} L 2 -inner product. The symmetric gradient is ε ( v ) : = 1 2 ( ∇ v + ( ∇ v ) T ) \varepsilon(v) := \frac{1}{2}(\nabla v + (\nabla v)^{\mathsf{T}}) ε ( v ) := 2 1 ( ∇ v + ( ∇ v ) T ) for v ∈ H 1 ( Ω ) v \in\mathbf{H}^{1}(\Omega) v ∈ H 1 ( Ω ) .
For a Banach space X X X and 1 ≤ r ≤ ∞ 1 \le r \le\infty 1 ≤ r ≤ ∞ , L r ( 0 , T ; X ) L^{r}(0,T;X) L r ( 0 , T ; X ) denotes the Bochner space of strongly measurable functions v : ( 0 , T ) → X v : (0,T) \to X v : ( 0 , T ) → X with ∥ v ∥ L r ( 0 , T ; X ) < ∞ \|v\|_{L^{r}(0,T;X)} < \infty ∥ v ∥ L r ( 0 , T ; X ) < ∞ . We write L t 2 L x 2 : = L 2 ( 0 , T ; L 2 ( Ω ) ) L_{t}^{2}L_{x}^{2} := L^{2}(0,T;L^{2}(\Omega)) L t 2 L x 2 := L 2 ( 0 , T ; L 2 ( Ω )) for the space-time L 2 L^{2} L 2 space, with norm ∥ v ∥ L t 2 L x 2 2 = ∫ 0 T ∥ v ∥ 0 2 d t \|v\|_{L_{t}^{2}L_{x}^{2}}^{2} = \int_{0}^{T}\|v\|_{0}^{2}\,dt ∥ v ∥ L t 2 L x 2 2 = ∫ 0 T ∥ v ∥ 0 2 d t . More generally, L t r X : = L r ( 0 , T ; X ) L_{t}^{r}X := L^{r}(0,T;X) L t r X := L r ( 0 , T ; X ) . The Bochner–Sobolev space W k , r ( 0 , T ; X ) W^{k,r}(0,T;X) W k , r ( 0 , T ; X ) consists of functions whose distributional time derivatives up to order k k k belong to L r ( 0 , T ; X ) L^{r}(0,T;X) L r ( 0 , T ; X ) ; we write H k ( 0 , T ; X ) : = W k , 2 ( 0 , T ; X ) H^{k}(0,T;X) := W^{k,2}(0,T;X) H k ( 0 , T ; X ) := W k , 2 ( 0 , T ; X ) . The embedding H 1 ( 0 , T ; X ) ↪ C 0 ( [ 0 , T ] ; X ) H^{1}(0,T;X) \hookrightarrow C^{0}([0,T];X) H 1 ( 0 , T ; X ) ↪ C 0 ([ 0 , T ] ; X ) holds for any Hilbert space X X X [8 ] . We write v ˙ : = ∂ t v \dot{v} := \partial_{t}v v ˙ := ∂ t v for the time derivative.
In quasi-static poroelasticity equations with homogeneous boundary conditions we seek the solid displacement $u : (0,T) \times \Omega \to \mathbb{R}^{d} and fluid pressure p : ( 0 , T ) × Ω → R p : (0,T) \times\Omega\to\mathbb{R} p : ( 0 , T ) × Ω → R satisfying
− div ( 2 μ ε ( u ) + λ div u I − α p I ) = f ~ , (1) -\operatorname{div}(2\mu\boldsymbol{\varepsilon}(u) + \lambda\operatorname{div} u\,\mathbf{I} - \alpha p\mathbf{I}) = \tilde{f}, \tag*{(1)} − div ( 2 μ ε ( u ) + λ div u I − α p I ) = f ~ , (1)
s 0 p ˙ + α div u ˙ − div ( κ p ∇ p ) = g p , (2) s_0\dot{p} + \alpha\operatorname{div}\dot{u} - \operatorname{div}(\kappa_p\nabla p) = g_p, \tag*{(2)} s 0 p ˙ + α div u ˙ − div ( κ p ∇ p ) = g p , (2)
in ( 0 , T ) × Ω (0,T) \times\Omega ( 0 , T ) × Ω , with boundary conditions u = 0 u = 0 u = 0 on ( 0 , T ) × Γ D (0,T) \times\Gamma_D ( 0 , T ) × Γ D , zero traction on ( 0 , T ) × Γ N (0,T) \times\Gamma_N ( 0 , T ) × Γ N , and p = 0 p = 0 p = 0 on ( 0 , T ) × ∂ Ω (0,T) \times\partial\Omega ( 0 , T ) × ∂ Ω , together with initial conditions ( u , p ) ∣ t = 0 = ( u 0 , p 0 ) (u,p)|_{t=0} = (u_0,p_0) ( u , p ) ∣ t = 0 = ( u 0 , p 0 ) . Here μ > 0 \mu> 0 μ > 0 and λ > 0 \lambda> 0 λ > 0 are the Lamé parameters, α > 0 \alpha> 0 α > 0 is the Biot–Willis coupling coefficient, s 0 ≥ 0 s_0 \ge0 s 0 ≥ 0 is the constrained specific storage coefficient, and κ p > 0 \kappa_p > 0 κ p > 0 is the fluid conductivity. Here g p g_p g p is the fluid source and f ~ \tilde{f} f ~ is the body force. In the distributed optimal control problems that we consider later, g p g_p g p will be replaced by the control variable.
We define the function spaces
V : = { v ∈ H 1 ( Ω ) : v ∣ Γ D = 0 } , Q T : = L 2 ( Ω ) , Q : = H 0 1 ( Ω ) , (3) V := \{v \in H^1(\Omega) : v|_{\Gamma_D} = 0\}, \qquad Q_T := L^2(\Omega), \qquad Q := H^1_0(\Omega), \tag*{(3)} V := { v ∈ H 1 ( Ω ) : v ∣ Γ D = 0 } , Q T := L 2 ( Ω ) , Q := H 0 1 ( Ω ) , (3)
equipped with inner products ( u , v ) V : = ( 2 μ ε ( u ) , ε ( v ) ) Ω (u,v)_V := (2\mu\boldsymbol{\varepsilon}(u),\boldsymbol{\varepsilon}(v))_\Omega ( u , v ) V := ( 2 μ ε ( u ) , ε ( v ) ) Ω , ( p , q ) Q T : = ( p , q ) Ω (p,q)_{Q_T} := (p,q)_\Omega ( p , q ) Q T := ( p , q ) Ω , and ( p , q ) Q : = ( κ p ∇ p , ∇ q ) Ω (p,q)_Q := (\kappa_p\nabla p,\nabla q)_\Omega ( p , q ) Q := ( κ p ∇ p , ∇ q ) Ω , and induced norms
∥ v ∥ V : = ( v , v ) V , ∥ p ∥ Q : = ( p , p ) Q = ( κ p ∇ p , ∇ p ) Ω 1 / 2 . (4) \|v\|_V := \sqrt{(v,v)_V}, \qquad\|p\|_Q := \sqrt{(p,p)_Q} = (\kappa_p\nabla p,\nabla p)_\Omega^{1/2}. \tag*{(4)} ∥ v ∥ V := ( v , v ) V , ∥ p ∥ Q := ( p , p ) Q = ( κ p ∇ p , ∇ p ) Ω 1/2 . (4)
By Korn’s inequality, ∥ ⋅ ∥ V \|\cdot\|_V ∥ ⋅ ∥ V is equivalent to ∥ ⋅ ∥ H 1 \|\cdot\|_{H^1} ∥ ⋅ ∥ H 1 on V V V . By the Poincaré inequality, ∥ ⋅ ∥ Q \|\cdot\|_Q ∥ ⋅ ∥ Q is equivalent to ∥ ⋅ ∥ H 1 \|\cdot\|_{H^1} ∥ ⋅ ∥ H 1 on H 0 1 ( Ω ) H^1_0(\Omega) H 0 1 ( Ω ) , with constant depending on κ p \kappa_p κ p . We denote by V ∗ V^* V ∗ the dual of V V V and keep the notation ( g , v ) Ω (g,v)_\Omega ( g , v ) Ω for the V ∗ V^* V ∗ –V V V duality pairing when g ∈ V ∗ g \in V^* g ∈ V ∗ ; it coincides with the L 2 L^2 L 2 inner product when g ∈ L 2 ( Ω ) d g \in L^2(\Omega)^d g ∈ L 2 ( Ω ) d . The body force satisfies f ~ ∈ H 1 ( 0 , T ; V ∗ ) \tilde{f} \in H^1(0,T;V^*) f ~ ∈ H 1 ( 0 , T ; V ∗ ) .
When λ ≫ 1 \lambda\gg1 λ ≫ 1 (nearly incompressible solid), standard displacement-pressure finite elements applied directly to (1) suffer from volumetric locking : the discrete solution locks onto the constraint div u ≈ 0 \operatorname{div} u \approx0 div u ≈ 0 and the approximation deteriorates. To circumvent this, we introduce the total pressure
p T : = − λ div u + α p , (5) p_T := -\lambda\operatorname{div} u + \alpha p, \tag*{(5)} p T := − λ div u + α p , (5)
equivalently div u = λ − 1 ( − p T + α p ) \operatorname{div} u = \lambda^{-1}(-p_T + \alpha p) div u = λ − 1 ( − p T + α p ) , as an additional unknown, as in [21 , 17 ] .
We introduce the bilinear forms
a u ( u , v ) : = ( 2 μ ε ( u ) , ε ( v ) ) Ω , b ( q T , v ) : = − ( q T , div v ) Ω , (6) a_u(u,v) := (2\mu\boldsymbol{\varepsilon}(u),\boldsymbol{\varepsilon}(v))_\Omega, \qquad b(q_T,v) := -(q_T,\operatorname{div}v)_\Omega, \tag*{(6)} a u ( u , v ) := ( 2 μ ε ( u ) , ε ( v ) ) Ω , b ( q T , v ) := − ( q T , div v ) Ω , (6)
for ( u , v ) ∈ V × V (u,v) \in V \times V ( u , v ) ∈ V × V and ( q T , v ) ∈ Q T × V (q_T,v) \in Q_T \times V ( q T , v ) ∈ Q T × V . The form a u a_u a u is coercive on V V V by Korn’s inequality, and b b b is the divergence coupling form arising in mixed formulations. Here the coefficient s 1 s_1 s 1 is defined as
s 1 : = s 0 + λ − 1 α 2 . (7) s_1 := s_0 + \lambda^{-1}\alpha^2. \tag*{(7)} s 1 := s 0 + λ − 1 α 2 . (7)
The storage coupling of the total pressure and fluid pressure is represented by
S ( λ ) : = [ λ − 1 − λ − 1 α − λ − 1 α s 1 ] . (8) \mathbb{S}(\lambda) :=
\begin{bmatrix}
\lambda^{-1} & -\lambda^{-1}\alpha\\
-\lambda^{-1}\alpha& s_1
\end{bmatrix}. \tag*{(8)} S ( λ ) := [ λ − 1 − λ − 1 α − λ − 1 α s 1 ] . (8)
The associated quadratic form is precisely the storage contribution that appears when the total pressure constraint and fluid mass balance are combined. Its nonnegativity, established by the following lemma, is useful for identifying the parameter dependence of the storage block.
Lemma 2.1 (λ \lambda λ -uniform nonnegativity of S \mathbb{S} S ). For λ > 0 \lambda> 0 λ > 0 and s 0 ≥ 0 s_0 \ge0 s 0 ≥ 0 ,
S ( λ ) = λ − 1 w w T + diag ( 0 , s 0 ) , w = ( 1 , − α ) T , (9) \mathbb{S}(\lambda) = \lambda^{-1}ww^\mathsf{T} + \operatorname{diag}(0,s_0), \qquad w = (1,-\alpha)^\mathsf{T}, \tag*{(9)} S ( λ ) = λ − 1 w w T + diag ( 0 , s 0 ) , w = ( 1 , − α ) T , (9) and hence, for all ξ = ( ξ p T , ξ p ) T \xi= (\xi_{p_T},\xi_p)^\mathsf{T} ξ = ( ξ p T , ξ p ) T ,
ξ T S ( λ ) ξ = λ − 1 ( ξ p T − α ξ p ) 2 + s 0 ξ p 2 , (10) \xi^\mathsf{T}\mathbb{S}(\lambda)\xi= \lambda^{-1}(\xi_{p_T} - \alpha\xi_p)^2 + s_0\xi_p^2, \tag*{(10)} ξ T S ( λ ) ξ = λ − 1 ( ξ p T − α ξ p ) 2 + s 0 ξ p 2 , (10) ξ T S ( λ ) ξ ≥ s 0 ξ p 2 . (11) \xi^\mathsf{T}\mathbb{S}(\lambda)\xi\ge s_0\xi_p^2. \tag*{(11)} ξ T S ( λ ) ξ ≥ s 0 ξ p 2 . (11)
In particular the lower bound is independent of λ > 0 \lambda> 0 λ > 0 .
Differentiating the momentum balance (1a) in time (with f : = f ~ f := \tilde{f} f := f ~ ) and using p ˙ T = − λ div u ˙ + α p ˙ \dot{p}_T = -\lambda\operatorname{div}\dot{u} + \alpha\dot{p} p ˙ T = − λ div u ˙ + α p ˙ to eliminate div u ˙ \operatorname{div}\dot{u} div u ˙ from (1b) yields the differentiated total pressure system
− div ( 2 μ ε ( u ˙ ) − p ˙ T I ) = f , (12) -\operatorname{div}\left(2\mu\varepsilon(\dot{u})-\dot{p}_T\mathbb{I}\right)=f, \tag*{(12)} − div ( 2 μ ε ( u ˙ ) − p ˙ T I ) = f , (12)
− div u ˙ − λ − 1 p ˙ T + λ − 1 α p ˙ = 0 , (13) -\operatorname{div}\dot{u}-\lambda^{-1}\dot{p}_T+\lambda^{-1}\alpha\dot{p}=0, \tag*{(13)} − div u ˙ − λ − 1 p ˙ T + λ − 1 α p ˙ = 0 , (13)
− λ − 1 α p ˙ T + s 1 p ˙ − div ( κ p ∇ p ) = g p , (14) -\lambda^{-1}\alpha\dot{p}_T+s_1\dot{p}-\operatorname{div}(\kappa_p\nabla p)=g_p, \tag*{(14)} − λ − 1 α p ˙ T + s 1 p ˙ − div ( κ p ∇ p ) = g p , (14)
with s 1 s_1 s 1 defined in (6). All factors of λ \lambda λ now appear only as λ − 1 \lambda^{-1} λ − 1 , so (11) is a first-order evolution problem for ( u , p T , p ) (u,p_T,p) ( u , p T , p ) with coefficients bounded uniformly as λ → ∞ \lambda\to\infty λ → ∞ . The precise equivalence between (1) and (11) under compatible initial data is established in §2.3.
The continuous inf-sup condition on the full total pressure space is
inf 0 ≠ q T ∈ Q T sup 0 ≠ v ∈ V ( q T , div v ) Ω ∥ v ∥ ∥ q T ∥ ≥ β > 0. (15) \inf_{0\ne q_T\in Q_T}\sup_{0\ne v\in V}\frac{(q_T,\operatorname{div}v)_\Omega}{\lVert v\rVert\lVert q_T\rVert}\ge\beta>0. \tag*{(15)} 0 = q T ∈ Q T inf 0 = v ∈ V sup ∥ v ∥ ∥ q T ∥ ( q T , div v ) Ω ≥ β > 0. (15)
Since ∣ Γ N ∣ > 0 |\Gamma_N|>0 ∣ Γ N ∣ > 0 , the divergence operator div : V → L 2 ( Ω ) = Q T \operatorname{div}:V\to L^2(\Omega)=Q_T div : V → L 2 ( Ω ) = Q T is surjective.
Remark 2.2. The forms a u a_u a u and b b b , and the inf-sup constant in (12), are independent of λ \lambda λ and s 0 s_0 s 0 . The parameter dependence of the evolution estimates is determined separately by the effective mass introduced below.
The continuous weak formulation of (11) reads: find ( u , p T , p ) ∈ V × Q T × Q (u,p_T,p)\in V\times Q_T\times Q ( u , p T , p ) ∈ V × Q T × Q such that, for a.e. t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
a u ( u ˙ , v ) + b ( p ˙ T , v ) = ( f , v ) Ω , (16) a_u(\dot{u},v)+b(\dot{p}_T,v)=(f,v)_\Omega, \tag*{(16)} a u ( u ˙ , v ) + b ( p ˙ T , v ) = ( f , v ) Ω , (16)
b ( q T , u ˙ ) − λ − 1 ( p ˙ T , q T ) Ω + λ − 1 α ( p ˙ , q T ) Ω = 0 , (17) b(q_T,\dot{u})-\lambda^{-1}(\dot{p}_T,q_T)_\Omega+\lambda^{-1}\alpha(\dot{p},q_T)_\Omega=0, \tag*{(17)} b ( q T , u ˙ ) − λ − 1 ( p ˙ T , q T ) Ω + λ − 1 α ( p ˙ , q T ) Ω = 0 , (17)
− λ − 1 α ( p ˙ T , q ) Ω + s 1 ( p ˙ , q ) Ω + ( κ p ∇ p , ∇ q ) Ω = ( g p , q ) Ω (18) -\lambda^{-1}\alpha(\dot{p}_T,q)_\Omega+s_1(\dot{p},q)_\Omega+(\kappa_p\nabla p,\nabla q)_\Omega=(g_p,q)_\Omega \tag*{(18)} − λ − 1 α ( p ˙ T , q ) Ω + s 1 ( p ˙ , q ) Ω + ( κ p ∇ p , ∇ q ) Ω = ( g p , q ) Ω (18)
for all ( v , q T , q ) ∈ V × Q T × Q (v,q_T,q)\in V\times Q_T\times Q ( v , q T , q ) ∈ V × Q T × Q . The time regularity and well-posedness of this system are established in Sections 2.4–2.5.
The weak system (13) is obtained by differentiating momentum balance and the total pressure relation in time. We establish its equivalence with the original system in the continuous spaces V , Q T = L 2 ( Ω ) , Q V,Q_T=L^2(\Omega),Q V , Q T = L 2 ( Ω ) , Q .
Define, for ( u , p T , p ) ∈ C 1 ( 0 , T ; V × L 2 ( Ω ) × Q ) (u,p_T,p)\in C^1(0,T;V\times L^2(\Omega)\times Q) ( u , p T , p ) ∈ C 1 ( 0 , T ; V × L 2 ( Ω ) × Q ) , the total pressure residual
Z ( t ) : = p T ( t ) + λ div u ( t ) − α p ( t ) ∈ L 2 ( Ω ) , (19) Z(t):=p_T(t)+\lambda\operatorname{div}u(t)-\alpha p(t)\in L^2(\Omega), \tag*{(19)} Z ( t ) := p T ( t ) + λ div u ( t ) − α p ( t ) ∈ L 2 ( Ω ) , (19)
so that Z ( t ) = 0 Z(t)=0 Z ( t ) = 0 is precisely the algebraic total pressure relation (4). The differentiated formulation permits arbitrary initial data; compatibility is imposed when equivalence with the original Biot system is asserted.
The original system consists of the (non-differentiated) momentum and mass-balance equations: find ( u , p ) ∈ C 1 ( 0 , T ; V × Q ) (u,p)\in C^1(0,T;V\times Q) ( u , p ) ∈ C 1 ( 0 , T ; V × Q ) such that
a u ( u , v ) + λ ( div u , div v ) Ω − α ( p , div v ) Ω = ( f ~ , v ) Ω , (20) a_u(u,v)+\lambda(\operatorname{div}u,\operatorname{div}v)_\Omega-\alpha(p,\operatorname{div}v)_\Omega=(\tilde{f},v)_\Omega, \tag*{(20)} a u ( u , v ) + λ ( div u , div v ) Ω − α ( p , div v ) Ω = ( f ~ , v ) Ω , (20)
s 0 ( p ˙ , q ) Ω + α ( div u ˙ , q ) Ω + ( κ p ∇ p , ∇ q ) Ω = ( g p , q ) Ω , (21) s_0(\dot{p},q)_\Omega+\alpha(\operatorname{div}\dot{u},q)_\Omega+(\kappa_p\nabla p,\nabla q)_\Omega=(g_p,q)_\Omega, \tag*{(21)} s 0 ( p ˙ , q ) Ω + α ( div u ˙ , q ) Ω + ( κ p ∇ p , ∇ q ) Ω = ( g p , q ) Ω , (21)
for all ( v , q ) ∈ V × Q (v,q)\in V\times Q ( v , q ) ∈ V × Q and all 0 < t < T 0<t<T 0 < t < T , where f ~ ∈ H 1 ( 0 , T ; V ∗ ) \tilde{f}\in H^1(0,T;V^*) f ~ ∈ H 1 ( 0 , T ; V ∗ ) and f : = f ~ ˙ f:=\dot{\tilde{f}} f := f ~ ˙ is the forcing appearing in (13a). We say the initial data ( u ( 0 ) , p T ( 0 ) , p ( 0 ) ) (u(0),p_T(0),p(0)) ( u ( 0 ) , p T ( 0 ) , p ( 0 )) are compatible if both of the following hold:
Z ( 0 ) = p T ( 0 ) + λ div u ( 0 ) − α p ( 0 ) = 0 , (22) Z(0)=p_T(0)+\lambda\operatorname{div}u(0)-\alpha p(0)=0, \tag*{(22)} Z ( 0 ) = p T ( 0 ) + λ div u ( 0 ) − α p ( 0 ) = 0 , (22)
i.e., the algebraic total pressure relation (5) at t = 0 t=0 t = 0 , and
a u ( u ( 0 ) , v ) + b ( p T ( 0 ) , v ) = ( f ~ ( 0 ) , v ) Ω ∀ v ∈ V , (23) a_u(u(0),v)+b(p_T(0),v)=(\tilde{f}(0),v)_{\Omega}\qquad\forall v\in V, \tag*{(23)} a u ( u ( 0 ) , v ) + b ( p T ( 0 ) , v ) = ( f ~ ( 0 ) , v ) Ω ∀ v ∈ V , (23)
i.e., the momentum equation holds at t = 0 t=0 t = 0 with the initial total pressure p T ( 0 ) = − λ div u ( 0 ) + α p ( 0 ) p_T(0)=-\lambda\operatorname{div}u(0)+\alpha p(0) p T ( 0 ) = − λ div u ( 0 ) + α p ( 0 ) . Condition (23) is needed to recover the original (non-differentiated) momentum equation from its time-integrated form. It is automatically satisfied if ( u ( 0 ) , p ( 0 ) ) (u(0),p(0)) ( u ( 0 ) , p ( 0 )) solves the original static equilibrium at t = 0 t=0 t = 0 .
Proposition 2.3 (Equivalence of the original and differentiated formulations). Let V , Q T = L 2 ( Ω ) , Q = H 0 1 ( Ω ) V,Q_T=L^2(\Omega),Q=H^1_0(\Omega) V , Q T = L 2 ( Ω ) , Q = H 0 1 ( Ω ) be the continuous function spaces of (3) . Assume f ~ ∈ H 1 ( 0 , T ; V ∗ ) \tilde{f}\in H^1(0,T;V^*) f ~ ∈ H 1 ( 0 , T ; V ∗ ) and f = f ~ ˙ f=\dot{\tilde{f}} f = f ~ ˙ .
(i) (Differentiated ⇒ \Rightarrow ⇒ original.) Let ( u , p T , p ) ∈ C 0 ( [ 0 , T ] ; V × L 2 ( Ω ) × Q ) ∩ C 1 ( 0 , T ; V × L 2 ( Ω ) × Q ) (u,p_T,p)\in C^0([0,T];V\times L^2(\Omega)\times Q)\cap C^1(0,T;V\times L^2(\Omega)\times Q) ( u , p T , p ) ∈ C 0 ([ 0 , T ] ; V × L 2 ( Ω ) × Q ) ∩ C 1 ( 0 , T ; V × L 2 ( Ω ) × Q ) solve (13) with compatible initial data (16). Then Z ( t ) = 0 Z(t)=0 Z ( t ) = 0 for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , and ( u , p ) (u,p) ( u , p ) solves the original system (15) with
f ~ ( t ) : = f ~ ( 0 ) + ∫ 0 t f ( s ) d s . (24) \tilde{f}(t):=\tilde{f}(0)+\int_0^t f(s)\,\mathrm{d}s. \tag*{(24)} f ~ ( t ) := f ~ ( 0 ) + ∫ 0 t f ( s ) d s . (24) (ii) (Original ⇒ \Rightarrow ⇒ differentiated.) Conversely, let ( u , p ) ∈ C 2 ( 0 , T ; V × Q ) (u,p)\in C^2(0,T;V\times Q) ( u , p ) ∈ C 2 ( 0 , T ; V × Q ) solve (15), and define p T ( t ) : = − λ div u ( t ) + α p ( t ) p_T(t):=-\lambda\operatorname{div}u(t)+\alpha p(t) p T ( t ) := − λ div u ( t ) + α p ( t ) via (5) (so Z ( t ) ≡ 0 Z(t)\equiv0 Z ( t ) ≡ 0 ). Then ( u , p T , p ) ∈ C 1 ( 0 , T ; V × L 2 ( Ω ) × Q ) (u,p_T,p)\in C^1(0,T;V\times L^2(\Omega)\times Q) ( u , p T , p ) ∈ C 1 ( 0 , T ; V × L 2 ( Ω ) × Q ) solves (13) with f = f ~ ˙ f=\dot{\tilde{f}} f = f ~ ˙ .
Proof. (i) Test the p T p_T p T -equation of (13) with an arbitrary q T ∈ L 2 ( Ω ) q_T\in L^2(\Omega) q T ∈ L 2 ( Ω ) (this is valid at the continuous level where Q T = L 2 ( Ω ) Q_T=L^2(\Omega) Q T = L 2 ( Ω ) ):
− ( div u ˙ , q T ) Ω − λ − 1 ( p ˙ T , q T ) Ω + λ − 1 α ( p ˙ , q T ) Ω = 0 ∀ q T ∈ L 2 ( Ω ) . -(\operatorname{div}\dot{u},q_T)_{\Omega}-\lambda^{-1}(\dot{p}_T,q_T)_{\Omega}+\lambda^{-1}\alpha(\dot{p},q_T)_{\Omega}=0\qquad\forall q_T\in L^2(\Omega). − ( div u ˙ , q T ) Ω − λ − 1 ( p ˙ T , q T ) Ω + λ − 1 α ( p ˙ , q T ) Ω = 0 ∀ q T ∈ L 2 ( Ω ) .
Since this holds for all q T ∈ L 2 ( Ω ) q_T\in L^2(\Omega) q T ∈ L 2 ( Ω ) , it holds pointwise a.e. in Ω \Omega Ω :
p ˙ T = − λ div u ˙ + α p ˙ = d d t ( − λ div u + α p ) . \dot{p}_T=-\lambda\operatorname{div}\dot{u}+\alpha\dot{p}=\frac{\mathrm{d}}{\mathrm{d}t}(-\lambda\operatorname{div}u+\alpha p). p ˙ T = − λ div u ˙ + α p ˙ = d t d ( − λ div u + α p ) .
Hence Z ˙ ( t ) = 0 \dot{Z}(t)=0 Z ˙ ( t ) = 0 for all t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , so Z ( t ) = Z ( 0 ) = 0 Z(t)=Z(0)=0 Z ( t ) = Z ( 0 ) = 0 by (16), i.e., p T ( t ) = − λ div u ( t ) + α p ( t ) p_T(t)=-\lambda\operatorname{div}u(t)+\alpha p(t) p T ( t ) = − λ div u ( t ) + α p ( t ) for all t t t . Substituting p ˙ T = − λ div u ˙ + α p ˙ \dot{p}_T=-\lambda\operatorname{div}\dot{u}+\alpha\dot{p} p ˙ T = − λ div u ˙ + α p ˙ into the displacement equation (16) and integrating in time from 0 0 0 to t t t gives
a u ( u ( t ) , v ) + b ( p T ( t ) , v ) = a u ( u ( 0 ) , v ) + b ( p T ( 0 ) , v ) + ∫ 0 t ( f ( s ) , v ) Ω d s . a_u(u(t),v)+b(p_T(t),v)=a_u(u(0),v)+b(p_T(0),v)+\int_0^t(f(s),v)_{\Omega}\,\mathrm{d}s. a u ( u ( t ) , v ) + b ( p T ( t ) , v ) = a u ( u ( 0 ) , v ) + b ( p T ( 0 ) , v ) + ∫ 0 t ( f ( s ) , v ) Ω d s .
By the initial momentum condition (23) , a u ( u ( 0 ) , v ) + b ( p T ( 0 ) , v ) = ( f ~ ( 0 ) , v ) Ω a_u(u(0),v)+b(p_T(0),v)=(\tilde{f}(0),v)_{\Omega} a u ( u ( 0 ) , v ) + b ( p T ( 0 ) , v ) = ( f ~ ( 0 ) , v ) Ω , so
a u ( u ( t ) , v ) + b ( p T ( t ) , v ) = ( f ~ ( 0 ) + ∫ 0 t f ( s ) d s , v ) Ω = ( f ~ ( t ) , v ) Ω , a_u(u(t),v)+b(p_T(t),v)=\left(\tilde{f}(0)+\int_0^t f(s)\,\mathrm{d}s,v\right)_{\Omega}=(\tilde{f}(t),v)_{\Omega}, a u ( u ( t ) , v ) + b ( p T ( t ) , v ) = ( f ~ ( 0 ) + ∫ 0 t f ( s ) d s , v ) Ω = ( f ~ ( t ) , v ) Ω ,
which recovers (20) . Substituting into (18) and using s 1 = s 0 + λ − 1 α 2 s_1=s_0+\lambda^{-1}\alpha^2 s 1 = s 0 + λ − 1 α 2 (cf. (7) ) recovers (21) .
(ii) Conversely, with Z ≡ 0 Z\equiv0 Z ≡ 0 by definition of p T p_T p T , differentiating (20) in time gives (16) . Substituting div u ˙ = λ − 1 ( − p ˙ T + α p ˙ ) \operatorname{div}\dot{u}=\lambda^{-1}(-\dot{p}_T+\alpha\dot{p}) div u ˙ = λ − 1 ( − p ˙ T + α p ˙ ) into (21) and using the s s s -coefficient identity gives (18) ; and Z = 0 Z=0 Z = 0 implies (17) . Hence ( u , p T , p ) (u,p_T,p) ( u , p T , p ) solves (13). □
Remark 2.4 (Compatible initial data). The differentiated total pressure system is equivalent to the original poroelastic system under both parts of (16): the algebraic condition (22) ensures Z ( t ) ≡ 0 Z(t)\equiv0 Z ( t ) ≡ 0 , while the initial momentum condition (23) is needed to identify the integration constant when recovering the original momentum equation (20) . Under (22) , p T ( 0 ) p_T(0) p T ( 0 ) is determined by u ( 0 ) u(0) u ( 0 ) and p ( 0 ) p(0) p ( 0 ) through (5) , and then (23) is automatically satisfied if ( u ( 0 ) , p ( 0 ) ) (u(0),p(0)) ( u ( 0 ) , p ( 0 )) solves the original static momentum balance at t = 0 t=0 t = 0 .
A priori estimates and well-posedness via mechanical Schur complement
For a priori estimate we eliminate the differentiated mechanical variables ( u ˙ , P ˙ T ) (\dot{u},\dot{\mathcal{P}}_T) ( u ˙ , P ˙ T ) and use the elastic Schur complement as an additional contribution to the scalar mass form. This allows an estimate independent of s 0 − 1 s_0^{-1} s 0 − 1 ,
Since the divergence map div : V → L 2 ( Ω ) \operatorname{div}: V \to L^2(\Omega) div : V → L 2 ( Ω ) is surjective, there exists a bounded right inverse R : L 2 ( Ω ) → V R:L^2(\Omega)\to V R : L 2 ( Ω ) → V such that
div ( R q ) = q , ∥ R q ∥ V ≤ C D ∥ q ∥ 0 . (25) \operatorname{div}(Rq)=q,\qquad\lVert Rq\rVert_V\le C_D\lVert q\rVert_0. \tag*{(25)} div ( R q ) = q , ∥ R q ∥ V ≤ C D ∥ q ∥ 0 . (25)
Lemma 2.5 (auxiliary solution operator). For given p ∈ L 2 ( Ω ) p\in L^2(\Omega) p ∈ L 2 ( Ω ) , let ( U p , P T p ) ∈ V × Q T (U_p,\mathcal{P}_Tp)\in V\times Q_T ( U p , P T p ) ∈ V × Q T be the unique solution of
a u ( U p , v ) + b ( P T p , v ) = 0 , (26) a_u(U_p,v)+b(\mathcal{P}_Tp,v)=0, \tag*{(26)} a u ( U p , v ) + b ( P T p , v ) = 0 , (26) b ( q T , U p ) − λ − 1 ( P T p , q T ) + λ − 1 α ( p , q T ) = 0 (27) b(q_T,U_p)-\lambda^{-1}(\mathcal{P}_Tp,q_T)+\lambda^{-1}\alpha(p,q_T)=0 \tag*{(27)} b ( q T , U p ) − λ − 1 ( P T p , q T ) + λ − 1 α ( p , q T ) = 0 (27) for all ( v , q T ) ∈ V × Q T (v,q_T)\in V\times Q_T ( v , q T ) ∈ V × Q T . Then
∥ U p ∥ 1 + ∥ P T p ∥ 0 ≤ C ∥ p ∥ 0 , (28) \lVert U_p\rVert_1+\lVert\mathcal{P}_Tp\rVert_0\le C\lVert p\rVert_0, \tag*{(28)} ∥ U p ∥ 1 + ∥ P T p ∥ 0 ≤ C ∥ p ∥ 0 , (28) where C C C is independent of s 0 s_0 s 0 and can be chosen uniformly for λ \lambda λ in any interval [ λ 0 , ∞ ) [\lambda_0,\infty) [ λ 0 , ∞ ) , λ 0 > 0 \lambda_0>0 λ 0 > 0 . For p ∈ H 0 1 ( Ω ) p\in H^1_0(\Omega) p ∈ H 0 1 ( Ω ) , additionally ∥ P T p ∥ H 1 ≤ C ∥ p ∥ H 1 \lVert\mathcal{P}_Tp\rVert_{H^1}\le C\lVert p\rVert_{H^1} ∥ P T p ∥ H 1 ≤ C ∥ p ∥ H 1 , and for p ∈ H 2 ( Ω ) p\in H^2(\Omega) p ∈ H 2 ( Ω ) , ∥ P T p ∥ H 1 ≤ C ∥ p ∥ H 2 \lVert\mathcal{P}_Tp\rVert_{H^1}\le C\lVert p\rVert_{H^2} ∥ P T p ∥ H 1 ≤ C ∥ p ∥ H 2 , under the regularity assumption (2) of Appendix A, applied with F = 0 F=0 F = 0 and ζ = − λ − 1 α p \zeta=-\lambda^{-1}\alpha p ζ = − λ − 1 α p .
Proof. This is the saddle point problem of displacement and total pressure with right-hand side − λ − 1 α ( p , ⋅ ) -\lambda^{-1}\alpha(p,\cdot) − λ − 1 α ( p , ⋅ ) in the second equation. The Babuška–Brezzi theorem and the inf-sup condition give existence, uniqueness, and (20). □
Lemma 2.6. For U U U and P T \mathcal{P}_T P T in (19), define
m e f f ( p , q ) : = − λ − 1 α ( P T p , q ) Ω + s 1 ( p , q ) Ω . (29) m_{\mathrm{eff}}(p,q):=-\lambda^{-1}\alpha(\mathcal{P}_Tp,q)_\Omega+s_1(p,q)_\Omega. \tag*{(29)} m eff ( p , q ) := − λ − 1 α ( P T p , q ) Ω + s 1 ( p , q ) Ω . (29) Then m e f f ( ⋅ , ⋅ ) m_{\mathrm{eff}}(\cdot,\cdot) m eff ( ⋅ , ⋅ ) is symmetric and bounded. Moreover, for every s 0 ≥ 0 s_0\ge0 s 0 ≥ 0 ,
m e f f ( q , q ) = s 0 ∥ q ∥ 0 2 + a u ( U q , U q ) + λ − 1 ∥ P T q − α q ∥ 0 2 ≥ c e f f ∥ q ∥ 0 2 (30) m_{\mathrm{eff}}(q,q)=s_0\lVert q\rVert_0^2+a_u(U_q,U_q)+\lambda^{-1}\lVert\mathcal{P}_Tq-\alpha q\rVert_0^2\ge c_{\mathrm{eff}}\lVert q\rVert_0^2 \tag*{(30)} m eff ( q , q ) = s 0 ∥ q ∥ 0 2 + a u ( U q , U q ) + λ − 1 ∥ P T q − α q ∥ 0 2 ≥ c eff ∥ q ∥ 0 2 (30) where c e f f : = s 0 + α 2 λ + C D 2 c_{\mathrm{eff}}:=s_0+\frac{\alpha^2}{\lambda+C_D^2} c eff := s 0 + λ + C D 2 α 2 .
Proof. Symmetry of m e f f ( ⋅ , ⋅ ) m_{\mathrm{eff}}(\cdot,\cdot) m eff ( ⋅ , ⋅ ) follows by testing the two mixed problems (19) for ( U p , P T p ) (U_p,\mathcal{P}_Tp) ( U p , P T p ) and ( U q , P T q ) (U_q,\mathcal{P}_Tq) ( U q , P T q ) against one another. Testing (19a) with U p U_p U p and (19b) with P T p \mathcal{P}_Tp P T p gives
a u ( U p , U p ) = λ − 1 ( α ( p , P T p ) Ω − ∥ P T p ∥ 0 2 ) . a_u(U_p,U_p)=\lambda^{-1}\bigl(\alpha(p,\mathcal{P}_Tp)_\Omega-\lVert\mathcal{P}_Tp\rVert_0^2\bigr). a u ( U p , U p ) = λ − 1 ( α ( p , P T p ) Ω − ∥ P T p ∥ 0 2 ) . Expanding the square in (22) proves the identity in (22).
For the inequality in (22), note that (19b) also gives ( P T p − α p , q T ) Ω = − λ ( div ( U p ) , q T ) Ω (\mathcal{P}_Tp-\alpha p,q_T)_\Omega=-\lambda(\operatorname{div}(U_p),q_T)_\Omega ( P T p − α p , q T ) Ω = − λ ( div ( U p ) , q T ) Ω for every q T ∈ L 2 ( Ω ) q_T\in L^2(\Omega) q T ∈ L 2 ( Ω ) . Combining this with q T = div v q_T=\operatorname{div}v q T = div v and (19a) gives
a u ( U p , v ) + λ ( div U p , div v ) Ω = α ( p , div v ) Ω ∀ v ∈ V . a_u(U_p,v)+\lambda(\operatorname{div}U_p,\operatorname{div}v)_\Omega=\alpha(p,\operatorname{div}v)_\Omega\qquad\forall v\in V. a u ( U p , v ) + λ ( div U p , div v ) Ω = α ( p , div v ) Ω ∀ v ∈ V . Choose v = R p v=Rp v = R p . By Cauchy–Schwarz in the elastic energy and the boundedness of R R R ,
α ∥ p ∥ 0 2 ≤ ( a u ( U p , U p ) + λ ∥ div U p ∥ 0 2 ) 1 / 2 ( ∥ R p ∥ V 2 + λ ∥ p ∥ 0 2 ) 1 / 2 \alpha\lVert p\rVert_0^2\le\bigl(a_u(U_p,U_p)+\lambda\lVert\operatorname{div}U_p\rVert_0^2\bigr)^{1/2}\bigl(\lVert Rp\rVert_V^2+\lambda\lVert p\rVert_0^2\bigr)^{1/2} α ∥ p ∥ 0 2 ≤ ( a u ( U p , U p ) + λ ∥ div U p ∥ 0 2 ) 1/2 ( ∥ R p ∥ V 2 + λ ∥ p ∥ 0 2 ) 1/2 ≤ ( a u ( U p , U p ) + λ ∥ div U p ∥ 0 2 ) 1 / 2 ( C D 2 + λ ) 1 / 2 ∥ p ∥ 0 . \le\bigl(a_u(U_p,U_p)+\lambda\lVert\operatorname{div}U_p\rVert_0^2\bigr)^{1/2}(C_D^2+\lambda)^{1/2}\lVert p\rVert_0. ≤ ( a u ( U p , U p ) + λ ∥ div U p ∥ 0 2 ) 1/2 ( C D 2 + λ ) 1/2 ∥ p ∥ 0 . Hence a u ( U p , U p ) + λ ∥ div U p ∥ 0 2 ≥ α 2 λ + C D 2 ∥ p ∥ 0 2 a_u(U_p,U_p)+\lambda\lVert\operatorname{div}U_p\rVert_0^2\ge\frac{\alpha^2}{\lambda+C_D^2}\lVert p\rVert_0^2 a u ( U p , U p ) + λ ∥ div U p ∥ 0 2 ≥ λ + C D 2 α 2 ∥ p ∥ 0 2 , and this implies the last inequality in (22). □
For the differentiated body force f ( t ) f(t) f ( t ) , let ( u f ( t ) , p T , f ( t ) ) ∈ V × Q T (u_f(t),p_T,f(t)) \in V \times Q_T ( u f ( t ) , p T , f ( t )) ∈ V × Q T be the unique solution of
a u ( u f , v ) + b ( p T , f , v ) = ( f , v ) Ω , b ( q T , u f ) − λ − 1 ( p T , f , q T ) Ω = 0 (31) \begin{aligned}
a_u(u_f,v) + b(p_{T,f},v) &= (f,v)_\Omega,\\
b(q_T,u_f) - \lambda^{-1}(p_{T,f},q_T)_\Omega&= 0 \tag*{(31)}
\end{aligned} a u ( u f , v ) + b ( p T , f , v ) b ( q T , u f ) − λ − 1 ( p T , f , q T ) Ω = ( f , v ) Ω , = 0 (31)
for all ( v , q T ) ∈ V × Q T (v,q_T) \in V \times Q_T ( v , q T ) ∈ V × Q T . The same saddle-point argument gives
∥ u f ∥ V + ∥ p T , f ∥ 0 ≤ C ∥ f ∥ 0 , p T , f = − λ div u f . (32) \lVert u_f\rVert_V+\lVert p_{T,f}\rVert_0 \le C\lVert f\rVert_0,\qquad p_{T,f}=-\lambda\operatorname{div}u_f. \tag*{(32)} ∥ u f ∥ V + ∥ p T , f ∥ 0 ≤ C ∥ f ∥ 0 , p T , f = − λ div u f . (32)
Lemma 2.7. A sufficiently regular triple ( u , p T , p ) (u,p_T,p) ( u , p T , p ) solves (13a)–(13c) if and only if
u ˙ = u f + U p ˙ , p ˙ T = p T , f + P T p ˙ , (33) \dot{u}=u_f+U\dot{p},\qquad\dot{p}_T=p_{T,f}+P_T\dot{p}, \tag*{(33)} u ˙ = u f + U p ˙ , p ˙ T = p T , f + P T p ˙ , (33) and p p p solves the reduced scalar problem
m e f f ( p ˙ , q ) + ( κ p ∇ p , ∇ q ) Ω = ( g p , q ) Ω + λ − 1 α ( p T , f , q ) Ω ∀ q ∈ Q . (34) m_{\mathrm{eff}}(\dot{p},q)+(\kappa_p\nabla p,\nabla q)_\Omega=(g_p,q)_\Omega+\lambda^{-1}\alpha(p_{T,f},q)_\Omega\qquad\forall q\in Q. \tag*{(34)} m eff ( p ˙ , q ) + ( κ p ∇ p , ∇ q ) Ω = ( g p , q ) Ω + λ − 1 α ( p T , f , q ) Ω ∀ q ∈ Q . (34) Proof. Equations (13a)–(13b) and linearity of the mixed mechanical problem give (25). Substitution in (13c) gives (26). Conversely, reconstructing ( u ˙ , p ˙ T ) (\dot{u},\dot{p}_T) ( u ˙ , p ˙ T ) by (25) gives (13a)–(13b), and (26) is exactly (13c).
Lemma 2.8. Let s 0 ≥ 0 s_0\ge0 s 0 ≥ 0 , and let ( u , p T , p ) (u,p_T,p) ( u , p T , p ) solve the differentiated total pressure system with p ( 0 ) ∈ Q p(0)\in Q p ( 0 ) ∈ Q . For 0 < t ≤ T 0<t\le T 0 < t ≤ T ,
∥ p ˙ ∥ L 2 ( 0 , t ; L x 2 ) + ∥ p ∥ C 0 ( [ 0 , t ] ; Q ) + ∥ u ˙ ∥ L 2 ( 0 , t ; V ) + ∥ p ˙ T ∥ L 2 ( 0 , t ; L x 2 ) ≤ C e f f ( ∥ p ( 0 ) ∥ Q + ∥ f ∥ L 2 ( 0 , t ; L x 2 ) + ∥ g p ∥ L 2 ( 0 , t ; L x 2 ) ) , ∥ p ∥ C 0 ( [ 0 , t ] ; L x 2 ) + ∥ p ∥ L 2 ( 0 , t ; Q ) + ∥ u ∥ C 0 ( [ 0 , t ] ; V ) + ∥ p T ∥ C 0 ( [ 0 , t ] ; L x 2 ) ≤ ∥ u ( 0 ) ∥ V + ∥ p T ( 0 ) ∥ 0 + C ∥ p ( 0 ) ∥ 0 + C ∥ f ∥ L 1 ( 0 , t ; L x 2 ) + C e f f ( m e f f ( p ( 0 ) , p ( 0 ) ) ) 1 / 2 + ∥ f ∥ L 2 ( 0 , t ; L x 2 ) + ∥ g p ∥ L 2 ( 0 , t ; L x 2 ) . (35) \begin{aligned}
\lVert\dot{p}\rVert_{L^2(0,t;L^2_x)}+\lVert p\rVert_{C^0([0,t];Q)}+\lVert\dot{u}\rVert_{L^2(0,t;V)}+\lVert\dot{p}_T\rVert_{L^2(0,t;L^2_x)}
&\le C_{\mathrm{eff}}\bigl(\lVert p(0)\rVert_Q+\lVert f\rVert_{L^2(0,t;L^2_x)}+\lVert g_p\rVert_{L^2(0,t;L^2_x)}\bigr),\\
\lVert p\rVert_{C^0([0,t];L^2_x)}+\lVert p\rVert_{L^2(0,t;Q)}+\lVert u\rVert_{C^0([0,t];V)}+\lVert p_T\rVert_{C^0([0,t];L^2_x)}
&\le\lVert u(0)\rVert_V+\lVert p_T(0)\rVert_0+C\lVert p(0)\rVert_0+C\lVert f\rVert_{L^1(0,t;L^2_x)}\\
&\quad+C_{\mathrm{eff}}\bigl(m_{\mathrm{eff}}(p(0),p(0))\bigr)^{1/2}+\lVert f\rVert_{L^2(0,t;L^2_x)}+\lVert g_p\rVert_{L^2(0,t;L^2_x)}. \tag*{(35)}
\end{aligned} ∥ p ˙ ∥ L 2 ( 0 , t ; L x 2 ) + ∥ p ∥ C 0 ([ 0 , t ] ; Q ) + ∥ u ˙ ∥ L 2 ( 0 , t ; V ) + ∥ p ˙ T ∥ L 2 ( 0 , t ; L x 2 ) ∥ p ∥ C 0 ([ 0 , t ] ; L x 2 ) + ∥ p ∥ L 2 ( 0 , t ; Q ) + ∥ u ∥ C 0 ([ 0 , t ] ; V ) + ∥ p T ∥ C 0 ([ 0 , t ] ; L x 2 ) ≤ C eff ( ∥ p ( 0 ) ∥ Q + ∥ f ∥ L 2 ( 0 , t ; L x 2 ) + ∥ g p ∥ L 2 ( 0 , t ; L x 2 ) ) , ≤ ∥ u ( 0 ) ∥ V + ∥ p T ( 0 ) ∥ 0 + C ∥ p ( 0 ) ∥ 0 + C ∥ f ∥ L 1 ( 0 , t ; L x 2 ) + C eff ( m eff ( p ( 0 ) , p ( 0 )) ) 1/2 + ∥ f ∥ L 2 ( 0 , t ; L x 2 ) + ∥ g p ∥ L 2 ( 0 , t ; L x 2 ) . (35) Here C e f f C_{\mathrm{eff}} C eff depends on c e f f − 1 c_{\mathrm{eff}}^{-1} c eff − 1 in (22).
Proof. Test (26) first with q = p ˙ q=\dot{p} q = p ˙ . Since m e f f m_{\mathrm{eff}} m eff is coercive and ( κ p ∇ p , ∇ p ) Ω = 1 2 d d t ∥ p ∥ Q 2 (\kappa_p\nabla p,\nabla p)_\Omega=\frac{1}{2}\frac{d}{dt}\lVert p\rVert_Q^2 ( κ p ∇ p , ∇ p ) Ω = 2 1 d t d ∥ p ∥ Q 2 , the mechanical lift estimate and Young’s inequality yield (27), using (25) for u ˙ \dot{u} u ˙ and p ˙ T \dot{p}_T p ˙ T .
If we test (26) with q = p q=p q = p , then symmetry of m e f f m_{\mathrm{eff}} m eff gives
1 2 d d t m e f f ( p , p ) + ( κ p ∇ p , ∇ p ) Ω = ( g p , p ) Ω + λ − 1 α ( p T , f , p ) Ω . \frac{1}{2}\frac{d}{dt}m_{\mathrm{eff}}(p,p)+(\kappa_p\nabla p,\nabla p)_\Omega=(g_p,p)_\Omega+\lambda^{-1}\alpha(p_{T,f},p)_\Omega. 2 1 d t d m eff ( p , p ) + ( κ p ∇ p , ∇ p ) Ω = ( g p , p ) Ω + λ − 1 α ( p T , f , p ) Ω . Poincaré’s inequality and λ − 1 p T , f = − div u f \lambda^{-1}p_{T,f}=-\operatorname{div}u_f λ − 1 p T , f = − div u f give
d d t m e f f ( p , p ) + c ∥ p ∥ Q 2 ≤ C ( ∥ g p ∥ 0 2 + ∥ f ∥ 0 2 ) . \frac{d}{dt}m_{\mathrm{eff}}(p,p)+c\lVert p\rVert_Q^2\le C\bigl(\lVert g_p\rVert_0^2+\lVert f\rVert_0^2\bigr). d t d m eff ( p , p ) + c ∥ p ∥ Q 2 ≤ C ( ∥ g p ∥ 0 2 + ∥ f ∥ 0 2 ) . Integrating and using m e f f ( p , p ) ≥ c e f f ∥ p ∥ 0 2 m_{\mathrm{eff}}(p,p)\ge c_{\mathrm{eff}}\lVert p\rVert_0^2 m eff ( p , p ) ≥ c eff ∥ p ∥ 0 2 yields
∥ p ∥ C 0 ( [ 0 , t ] ; L x 2 ) + ∥ p ∥ L 2 ( 0 , t ; Q ) ≤ C e f f ( m e f f ( p ( 0 ) , p ( 0 ) ) ) 1 / 2 + ∥ f ∥ L 2 ( 0 , t ; L x 2 ) + ∥ g p ∥ L 2 ( 0 , t ; L x 2 ) . \lVert p\rVert_{C^0([0,t];L^2_x)}+\lVert p\rVert_{L^2(0,t;Q)}\le C_{\mathrm{eff}}\bigl(m_{\mathrm{eff}}(p(0),p(0))\bigr)^{1/2}+\lVert f\rVert_{L^2(0,t;L^2_x)}+\lVert g_p\rVert_{L^2(0,t;L^2_x)}. ∥ p ∥ C 0 ([ 0 , t ] ; L x 2 ) + ∥ p ∥ L 2 ( 0 , t ; Q ) ≤ C eff ( m eff ( p ( 0 ) , p ( 0 )) ) 1/2 + ∥ f ∥ L 2 ( 0 , t ; L x 2 ) + ∥ g p ∥ L 2 ( 0 , t ; L x 2 ) . Finally integrate the mechanical reconstruction in time:
u ( s ) = u ( 0 ) + ∫ 0 s u f ( τ ) d τ + U ( p ( s ) − p ( 0 ) ) , p T ( s ) = p T ( 0 ) + ∫ 0 s p T , f ( τ ) d τ + P T ( p ( s ) − p ( 0 ) ) . u(s)=u(0)+\int_0^s u_f(\tau)\,d\tau+U(p(s)-p(0)),\qquad p_T(s)=p_T(0)+\int_0^s p_{T,f}(\tau)\,d\tau+P_T(p(s)-p(0)). u ( s ) = u ( 0 ) + ∫ 0 s u f ( τ ) d τ + U ( p ( s ) − p ( 0 )) , p T ( s ) = p T ( 0 ) + ∫ 0 s p T , f ( τ ) d τ + P T ( p ( s ) − p ( 0 )) . The mechanical operator and load-lift bounds yield
∥ u ∥ C 0 ( [ 0 , t ] ; V ) + ∥ p T ∥ C 0 ( [ 0 , t ] ; L x 2 ) ≤ ∥ u ( 0 ) ∥ V + ∥ p T ( 0 ) ∥ 0 + C ∥ f ∥ L 1 ( 0 , t ; L x 2 ) + C ( ∥ p ∥ C 0 ( [ 0 , t ] ; L x 2 ) + ∥ p ( 0 ) ∥ 0 ) , \lVert u\rVert_{C^0([0,t];V)}+\lVert p_T\rVert_{C^0([0,t];L^2_x)}\le\lVert u(0)\rVert_V+\lVert p_T(0)\rVert_0+C\lVert f\rVert_{L^1(0,t;L^2_x)}+C\bigl(\lVert p\rVert_{C^0([0,t];L^2_x)}+\lVert p(0)\rVert_0\bigr), ∥ u ∥ C 0 ([ 0 , t ] ; V ) + ∥ p T ∥ C 0 ([ 0 , t ] ; L x 2 ) ≤ ∥ u ( 0 ) ∥ V + ∥ p T ( 0 ) ∥ 0 + C ∥ f ∥ L 1 ( 0 , t ; L x 2 ) + C ( ∥ p ∥ C 0 ([ 0 , t ] ; L x 2 ) + ∥ p ( 0 ) ∥ 0 ) , and thus (28).
Existence and uniqueness of weak solutions
Definition 2.9. For given ( u 0 , p T , 0 , p 0 ) ∈ V × L 2 ( Ω ) × Q (u_0,p_{T,0},p_0) \in V \times L^2(\Omega) \times Q ( u 0 , p T , 0 , p 0 ) ∈ V × L 2 ( Ω ) × Q , a triple ( u , p T , p ) ∈ H 1 ( 0 , T ; V ) × H 1 ( 0 , T ; L 2 ( Ω ) ) × ( H 1 ( 0 , T ; L 2 ( Ω ) ) ∩ L 2 ( 0 , T ; Q ) ) (u,p_T,p) \in H^1(0,T;V) \times H^1(0,T;L^2(\Omega)) \times(H^1(0,T;L^2(\Omega)) \cap L^2(0,T;Q)) ( u , p T , p ) ∈ H 1 ( 0 , T ; V ) × H 1 ( 0 , T ; L 2 ( Ω )) × ( H 1 ( 0 , T ; L 2 ( Ω )) ∩ L 2 ( 0 , T ; Q )) is a weak solution if
∫ 0 T a u ( u ˙ , v ) d t + ∫ 0 T b ( p ˙ T , v ) d t = ∫ 0 T ( f , v ) Ω d t , (36) \int_0^T a_u(\dot{u},v)\,\mathrm{d}t + \int_0^T b(\dot{p}_T,v)\,\mathrm{d}t = \int_0^T (f,v)_\Omega\,\mathrm{d}t, \tag*{(36)} ∫ 0 T a u ( u ˙ , v ) d t + ∫ 0 T b ( p ˙ T , v ) d t = ∫ 0 T ( f , v ) Ω d t , (36) ∫ 0 T [ b ( q T , u ˙ ) − λ − 1 ( p ˙ T , q T ) Ω + λ − 1 α ( p ˙ , q T ) Ω ] d t = 0 , (37) \int_0^T [b(q_T,\dot{u})-\lambda^{-1}(\dot{p}_T,q_T)_\Omega+\lambda^{-1}\alpha(\dot{p},q_T)_\Omega]\,\mathrm{d}t = 0, \tag*{(37)} ∫ 0 T [ b ( q T , u ˙ ) − λ − 1 ( p ˙ T , q T ) Ω + λ − 1 α ( p ˙ , q T ) Ω ] d t = 0 , (37) ∫ 0 T [ − λ − 1 α ( p ˙ T , q ) Ω + s 1 ( p ˙ , q ) Ω + ( κ p ∇ p , ∇ q ) Ω ] d t = ∫ 0 T ( g p , q ) Ω d t , (38) \int_0^T [-\lambda^{-1}\alpha(\dot{p}_T,q)_\Omega+s_1(\dot{p},q)_\Omega+(\kappa_p\nabla p,\nabla q)_\Omega]\,\mathrm{d}t = \int_0^T (g_p,q)_\Omega\,\mathrm{d}t, \tag*{(38)} ∫ 0 T [ − λ − 1 α ( p ˙ T , q ) Ω + s 1 ( p ˙ , q ) Ω + ( κ p ∇ p , ∇ q ) Ω ] d t = ∫ 0 T ( g p , q ) Ω d t , (38) for all temporally compactly supported test functions ( v , q T , q ) ∈ C 0 ∞ ( ( 0 , T ) ; V × Q T × Q ) (v,q_T,q) \in C_0^\infty((0,T);V \times Q_T \times Q) ( v , q T , q ) ∈ C 0 ∞ (( 0 , T ) ; V × Q T × Q ) .
Theorem 2.10. Let s 0 ≥ 0 s_0 \ge0 s 0 ≥ 0 , f , g p ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) f,g_p \in L^2(0,T;L^2(\Omega)) f , g p ∈ L 2 ( 0 , T ; L 2 ( Ω )) , and ( u 0 , p T , 0 , p 0 ) ∈ V × L 2 ( Ω ) × Q (u_0,p_{T,0},p_0) \in V \times L^2(\Omega) \times Q ( u 0 , p T , 0 , p 0 ) ∈ V × L 2 ( Ω ) × Q . Then there exists a unique weak solution in the sense of Definition 2.9, and it satisfies (27)–(28).
Proof. By Lemma 2.6, m e f f m_{\mathrm{eff}} m eff is a bounded symmetric coercive bilinear form on L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) for every fixed λ > 0 \lambda> 0 λ > 0 and every s 0 ≥ 0 s_0 \ge0 s 0 ≥ 0 . The form ( κ p ∇ , ∇ ) Ω (\kappa_p\nabla,\nabla)_\Omega ( κ p ∇ , ∇ ) Ω is coercive on Q = H 0 1 ( Ω ) Q = H^1_0(\Omega) Q = H 0 1 ( Ω ) . Therefore the standard theory for linear parabolic problems applied to (26) gives a unique
p ∈ H 1 ( 0 , T ; L 2 ( Ω ) ) ∩ L ∞ ( 0 , T ; Q ) p \in H^1(0,T;L^2(\Omega)) \cap L^\infty(0,T;Q) p ∈ H 1 ( 0 , T ; L 2 ( Ω )) ∩ L ∞ ( 0 , T ; Q )
with p ( 0 ) = p 0 p(0)=p_0 p ( 0 ) = p 0 . Define u ˙ \dot{u} u ˙ and p ˙ T \dot{p}_T p ˙ T by (25) and set
u ( t ) = u 0 + ∫ 0 t u ˙ ( s ) d s , p T ( t ) = p T , 0 + ∫ 0 t p ˙ T ( s ) d s . u(t) = u_0 + \int_0^t \dot{u}(s)\,\mathrm{d}s,\qquad p_T(t) = p_{T,0} + \int_0^t \dot{p}_T(s)\,\mathrm{d}s. u ( t ) = u 0 + ∫ 0 t u ˙ ( s ) d s , p T ( t ) = p T , 0 + ∫ 0 t p ˙ T ( s ) d s .
Lemma 2.5 and (24) imply u ∈ H 1 ( 0 , T ; V ) u \in H^1(0,T;V) u ∈ H 1 ( 0 , T ; V ) and p T ∈ H 1 ( 0 , T ; L 2 ( Ω ) ) p_T \in H^1(0,T;L^2(\Omega)) p T ∈ H 1 ( 0 , T ; L 2 ( Ω )) . Lemma 2.7 then shows that the reconstructed triple satisfies the three-field differentiated system and hence the weak formulation. Uniqueness follows first for p p p from the reduced parabolic problem and then for ( u ˙ , p ˙ T ) (\dot{u},\dot{p}_T) ( u ˙ , p ˙ T ) from the unique mechanical reconstruction; the prescribed initial data determine ( u , p T ) (u,p_T) ( u , p T ) uniquely. □ \square □
Optimal control problem
Existence of optimal control
Let l p < L p l_p < L_p l p < L p be prescribed lower and upper control bounds. The admissible control set is
M p , a d : = { m ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) : l p ≤ m ( t , x ) ≤ L p a.e. in ( 0 , T ) × Ω } . (39) \mathcal{M}_{p,ad} := \{m \in L^2(0,T;L^2(\Omega)) : l_p \le m(t,x) \le L_p\ \text{a.e. in }(0,T) \times\Omega\}. \tag*{(39)} M p , a d := { m ∈ L 2 ( 0 , T ; L 2 ( Ω )) : l p ≤ m ( t , x ) ≤ L p a.e. in ( 0 , T ) × Ω } . (39)
The set M p , a d \mathcal{M}_{p,ad} M p , a d is nonempty (the constant l p l_p l p belongs to it), closed and convex in L 2 ( 0 , T ; L 2 ( Ω ) ) L^2(0,T;L^2(\Omega)) L 2 ( 0 , T ; L 2 ( Ω )) , and bounded.
Our main goal is to choose a distributed control m p ∈ M p , a d m_p \in\mathcal{M}_{p,ad} m p ∈ M p , a d acting on the fluid pressure equation such that the corresponding state ( u , p T , p ) (u,p_T,p) ( u , p T , p ) best approximates prescribed target fields u C , p C u_C,p_C u C , p C . Let α u , C \alpha_{u,C} α u , C , α p , C ≥ 0 \alpha_{p,C} \ge0 α p , C ≥ 0 be tracking weights (with α u , C + α p , C > 0 \alpha_{u,C}+\alpha_{p,C} > 0 α u , C + α p , C > 0 ) and let γ p > 0 \gamma_p > 0 γ p > 0 be the control cost parameter. For given target fields u C , p C ∈ L t 2 L x 2 u_C,p_C \in L^2_tL^2_x u C , p C ∈ L t 2 L x 2 and an admissible tuple ( u , p T , p , m p ) (u,p_T,p,m_p) ( u , p T , p , m p ) , we define the objective functional J ( u , p , m p ) J(u,p,m_p) J ( u , p , m p ) as
J ( u , p , m p ) = α u , C 2 ∥ u − u C ∥ L t 2 L x 2 2 + α p , C 2 ∥ p − p C ∥ L t 2 L x 2 2 + γ p 2 ∥ m p ∥ L t 2 L x 2 2 . (40) J(u,p,m_p) = \frac{\alpha_{u,C}}{2}\lVert u-u_C\rVert_{L^2_tL^2_x}^2 + \frac{\alpha_{p,C}}{2}\lVert p-p_C\rVert_{L^2_tL^2_x}^2 + \frac{\gamma_p}{2}\lVert m_p\rVert_{L^2_tL^2_x}^2. \tag*{(40)} J ( u , p , m p ) = 2 α u , C ∥ u − u C ∥ L t 2 L x 2 2 + 2 α p , C ∥ p − p C ∥ L t 2 L x 2 2 + 2 γ p ∥ m p ∥ L t 2 L x 2 2 . (40)
The running tracking terms penalize the distance of the trajectory ( u , p ) (u,p) ( u , p ) to the prescribed target profiles ( u C , p C ) (u_C,p_C) ( u C , p C ) over the whole time interval ( 0 , T ) (0,T) ( 0 , T ) , and the last term is the cost of the control m p m_p m p .
Definition 3.1. Let m ^ p ∈ M p , a d \widehat{m}_p \in\mathcal{M}_{p,ad} m p ∈ M p , a d . The admissible tuple ( u ^ , p ^ T , p ^ , m ^ p ) (\widehat{u},\widehat{p}_T,\widehat{p},\widehat{m}_p) ( u , p T , p , m p ) is called an optimal control if
J ( u , p , m p ) ≥ J ( u ^ , p ^ , m ^ p ) (41) J(u,p,m_p) \ge J(\widehat{u},\widehat{p},\widehat{m}_p) \tag*{(41)} J ( u , p , m p ) ≥ J ( u , p , m p ) (41) for all admissible ( u , p T , p , m p ) (u,p_T,p,m_p) ( u , p T , p , m p ) with the same initial data ( u 0 , p T , 0 , p 0 ) (u_0,p_{T,0},p_0) ( u 0 , p T , 0 , p 0 ) .
Writing ( u ( m p ) , p ( m p ) ) (u(m_p),p(m_p)) ( u ( m p ) , p ( m p )) for the displacement and fluid pressure of the state associated with m p m_p m p (for fixed initial data and body force), we use the reduced functional
j ( m p ) : = J ( u ( m p ) , p ( m p ) , m p ) , m p ∈ M p , a d . (42) j(m_p) := J(u(m_p),p(m_p),m_p), \qquad m_p \in\mathcal{M}_{p,ad}. \tag*{(42)} j ( m p ) := J ( u ( m p ) , p ( m p ) , m p ) , m p ∈ M p , a d . (42)
Theorem 3.2. For any given initial data ( u 0 , p T , 0 , p 0 ) ∈ V × Q T × Q (u_0,p_{T,0},p_0) \in V \times Q_T \times Q ( u 0 , p T , 0 , p 0 ) ∈ V × Q T × Q there exists a unique optimal control m ^ p ∈ M p , a d \widehat{m}_p \in\mathcal{M}_{p,ad} m p ∈ M p , a d .
Proof. By Theorem 2.10 , the solution of the differentiated state system depends affinely and continuously on m p ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) m_p \in L^2(0,T;L^2(\Omega)) m p ∈ L 2 ( 0 , T ; L 2 ( Ω )) , for fixed initial data and body force. Denote its displacement and fluid-pressure components by ( u ( m p ) , p ( m p ) ) (u(m_p),p(m_p)) ( u ( m p ) , p ( m p )) and let j j j be the reduced functional (33). The admissible set M p , a d \mathcal{M}_{p,ad} M p , a d is nonempty, convex, closed and hence weakly closed in L 2 ( 0 , T ; L 2 ( Ω ) ) L^2(0,T;L^2(\Omega)) L 2 ( 0 , T ; L 2 ( Ω )) .
The control regularization with γ p > 0 \gamma_p>0 γ p > 0 makes j j j coercive; thus a minimizing sequence is bounded in L 2 ( 0 , T ; L 2 ( Ω ) ) L^2(0,T;L^2(\Omega)) L 2 ( 0 , T ; L 2 ( Ω )) and has a weakly convergent subsequence with admissible limit. The bounded affine state map preserves weak convergence. The tracking terms and the control cost are squared Hilbert-space norms of affine functions, and are weakly lower semicontinuous. Consequently the weak limit minimizes j j j . Finally, the control cost is strictly convex because γ p > 0 \gamma_p>0 γ p > 0 , while the tracking terms are convex. Hence j j j is strictly convex and its minimizer is unique. □
Necessary optimality conditions
We define the state residual as follows. For ( u , p T , p ) (u,p_T,p) ( u , p T , p ) , a control m p m_p m p , and a test triple ( v ∗ , q T ∗ , q ∗ ) ∈ V × Q T × Q (v^*,q_T^*,q^*) \in V \times Q_T \times Q ( v ∗ , q T ∗ , q ∗ ) ∈ V × Q T × Q , define
R ( ( u , p T , p ) , m p ; ( v ∗ , q T ∗ , q ∗ ) ) \mathcal{R}((u,p_T,p),m_p;(v^*,q_T^*,q^*)) R (( u , p T , p ) , m p ; ( v ∗ , q T ∗ , q ∗ ))
: = a u ( u ˙ , v ∗ ) + b ( p ˙ T , v ∗ ) + b ( q T ∗ , u ˙ ) − λ − 1 ( p ˙ T , q T ∗ ) Ω + λ − 1 α ( p , q T ∗ ) Ω := a_u(\dot{u},v^*) + b(\dot{p}_T,v^*) + b(q_T^*,\dot{u}) - \lambda^{-1}(\dot{p}_T,q_T^*)_\Omega+ \lambda^{-1}\alpha(p,q_T^*)_\Omega := a u ( u ˙ , v ∗ ) + b ( p ˙ T , v ∗ ) + b ( q T ∗ , u ˙ ) − λ − 1 ( p ˙ T , q T ∗ ) Ω + λ − 1 α ( p , q T ∗ ) Ω
− λ − 1 α ( p T , q ∗ ) Ω + s 1 ( p , q ∗ ) Ω + ( κ p ∇ p , ∇ q ∗ ) Ω − ( f , v ∗ ) Ω − ( m p , q ∗ ) Ω . (43) -\lambda^{-1}\alpha(p_T,q^*)_\Omega+ s_1(p,q^*)_\Omega+ (\kappa_p\nabla p,\nabla q^*)_\Omega- (f,v^*)_\Omega- (m_p,q^*)_\Omega. \tag*{(43)} − λ − 1 α ( p T , q ∗ ) Ω + s 1 ( p , q ∗ ) Ω + ( κ p ∇ p , ∇ q ∗ ) Ω − ( f , v ∗ ) Ω − ( m p , q ∗ ) Ω . (43)
Thus the weak state equations (13a)–(13c) are equivalent to R ( ( u , p T , p ) , m p ; ( v ∗ , q T ∗ , q ∗ ) ) = 0 \mathcal{R}((u,p_T,p),m_p;(v^*,q_T^*,q^*))=0 R (( u , p T , p ) , m p ; ( v ∗ , q T ∗ , q ∗ )) = 0 for every ( v ∗ , q T ∗ , q ∗ ) ∈ V × Q T × Q (v^*,q_T^*,q^*) \in V \times Q_T \times Q ( v ∗ , q T ∗ , q ∗ ) ∈ V × Q T × Q .
Using the residual R \mathcal{R} R of (34), we introduce the Lagrangian functional
L ( ( u , p T , p ) , ( v ∗ , q T ∗ , q ∗ ) , m p ) : = J ( u , p , m p ) − ∫ 0 T R ( ( u , p T , p ) , m p ; ( v ∗ , q T ∗ , q ∗ ) ) d t . \mathcal{L}((u,p_T,p),(v^*,q_T^*,q^*),m_p) := J(u,p,m_p) - \int_0^T \mathcal{R}((u,p_T,p),m_p;(v^*,q_T^*,q^*))\,\mathrm{d}t. L (( u , p T , p ) , ( v ∗ , q T ∗ , q ∗ ) , m p ) := J ( u , p , m p ) − ∫ 0 T R (( u , p T , p ) , m p ; ( v ∗ , q T ∗ , q ∗ )) d t .
where ( v ∗ , q T ∗ , q ∗ ) (v^*,q_T^*,q^*) ( v ∗ , q T ∗ , q ∗ ) are the adjoint (co-state) variables corresponding to the displacement, total pressure, and fluid pressure, respectively.
Setting the variation of L \mathcal{L} L with respect to the state ( u , p T , p ) (u,p_T,p) ( u , p T , p ) to zero and integrating the time-derivative terms by parts yields the adjoint system, whose time-derivative block is the algebraic transpose of that in the state equation: find ( v ∗ , q T ∗ , q ∗ ) ∈ V × Q T × Q (v^*,q_T^*,q^*) \in V \times Q_T \times Q ( v ∗ , q T ∗ , q ∗ ) ∈ V × Q T × Q such that
− a u ( v ˙ ∗ , v ) − b ( q ˙ T ∗ , v ) = α u , C ( u − u C , v ) Ω , (44) -a_u(\dot{v}^*,v) - b(\dot{q}_T^*,v) = \alpha_{u,C}(u-u_C,v)_\Omega, \tag*{(44)} − a u ( v ˙ ∗ , v ) − b ( q ˙ T ∗ , v ) = α u , C ( u − u C , v ) Ω , (44)
− b ( q T , v ˙ ∗ ) + λ − 1 ( q ˙ T ∗ , q T ) Ω + λ − 1 α ( q ˙ ∗ , q T ) Ω = 0 , (45) -b(q_T,\dot{v}^*) + \lambda^{-1}(\dot{q}_T^*,q_T)_\Omega+ \lambda^{-1}\alpha(\dot{q}^*,q_T)_\Omega= 0, \tag*{(45)} − b ( q T , v ˙ ∗ ) + λ − 1 ( q ˙ T ∗ , q T ) Ω + λ − 1 α ( q ˙ ∗ , q T ) Ω = 0 , (45)
− λ − 1 α ( q ˙ ∗ , q ) Ω − s 1 ( q ˙ ∗ , q ) Ω + ( κ p ∇ q ∗ , ∇ q ) Ω = α p , C ( p − p C , q ) Ω , (46) -\lambda^{-1}\alpha(\dot{q}^*,q)_\Omega- s_1(\dot{q}^*,q)_\Omega+ (\kappa_p\nabla q^*,\nabla q)_\Omega= \alpha_{p,C}(p-p_C,q)_\Omega, \tag*{(46)} − λ − 1 α ( q ˙ ∗ , q ) Ω − s 1 ( q ˙ ∗ , q ) Ω + ( κ p ∇ q ∗ , ∇ q ) Ω = α p , C ( p − p C , q ) Ω , (46)
for all ( v , q T , q ) ∈ V × Q T × Q (v,q_T,q) \in V \times Q_T \times Q ( v , q T , q ) ∈ V × Q T × Q and for all 0 < t < T 0<t<T 0 < t < T , with the homogeneous terminal condition
( v ∗ ( T ) , q T ∗ ( T ) , q ∗ ( T ) ) = ( 0 , 0 , 0 ) . (47) (v^*(T),q_T^*(T),q^*(T)) = (0,0,0). \tag*{(47)} ( v ∗ ( T ) , q T ∗ ( T ) , q ∗ ( T )) = ( 0 , 0 , 0 ) . (47)
The adjoint system (35)–(36) is a backward-in-time parabolic problem with homogeneous terminal condition (36).
Remark 3.3. The signs of the α \alpha α -coupling in (35b)–(35c) are those of the transpose of the storage block of the state system.
Well-posedness of (35) with (36) follows by time reversal and the same mechanical Schur-complement reduction as for the state system. After the change t ↦ T − t t \mapsto T-t t ↦ T − t , the sign of the scalar coupling in the mechanical block is reversed, so the mechanical lift of the scalar derivative changes sign; the coupling in the scalar adjoint equation changes sign at the same time. Consequently the two sign changes cancel after elimination and the reduced adjoint equation contains the same effective mass form m e f f m_{\mathrm{eff}} m eff as (26). Hence Lemma 2.6 gives backward well-posedness.
The following lemma characterizes the optimal control via a variational inequality and an equivalent pointwise projection formula.
Lemma 3.4. Let ( u ( m ^ p ) , p T ( m ^ p ) , p ( m ^ p ) ) (u(\widehat m_p),p_T(\widehat m_p),p(\widehat m_p)) ( u ( m p ) , p T ( m p ) , p ( m p )) be the state corresponding to m ^ p ∈ M p , a d \widehat m_p \in\mathcal{M}_{p,\mathrm{ad}} m p ∈ M p , ad , and let ( v ∗ ( m ^ p ) , q T ∗ ( m ^ p ) , q ∗ ( m ^ p ) ) (v^*(\widehat m_p),q_T^*(\widehat m_p),q^*(\widehat m_p)) ( v ∗ ( m p ) , q T ∗ ( m p ) , q ∗ ( m p )) be the corresponding adjoint solution of (35)–(36). Then m ^ p \widehat m_p m p is the optimal control if and only if the following variational inequality holds:
∫ 0 T ( γ p m ^ p + q ∗ ( m ^ p ) , m p − m ^ p ) Ω d t ≥ 0 ∀ m p ∈ M p , a d . (48) \int_0^T (\gamma_p\widehat m_p+q^*(\widehat m_p),m_p-\widehat m_p)_\Omega\,dt \ge0 \qquad\forall m_p\in\mathcal{M}_{p,\mathrm{ad}}. \tag*{(48)} ∫ 0 T ( γ p m p + q ∗ ( m p ) , m p − m p ) Ω d t ≥ 0 ∀ m p ∈ M p , ad . (48) Moreover, (37) is equivalent to the pointwise projection formula
m ^ p ( t , x ) = P M p , a d ( − 1 γ p q ∗ ( m ^ p ) ( t , x ) ) a.e. in ( 0 , T ) × Ω , (49) \widehat m_p(t,x)=\mathcal{P}_{\mathcal{M}_{p,\mathrm{ad}}}\left(-\frac{1}{\gamma_p}q^*(\widehat m_p)(t,x)\right)
\qquad\text{a.e. in }(0,T)\times\Omega, \tag*{(49)} m p ( t , x ) = P M p , ad ( − γ p 1 q ∗ ( m p ) ( t , x ) ) a.e. in ( 0 , T ) × Ω , (49) where the pointwise projection onto the admissible set is defined by
P M p , a d ( r ) ( t , x ) : = max { l p , min { r ( t , x ) , L p } } . (50) \mathcal{P}_{\mathcal{M}_{p,\mathrm{ad}}}(r)(t,x):=\max\{l_p,\min\{r(t,x),L_p\}\}. \tag*{(50)} P M p , ad ( r ) ( t , x ) := max { l p , min { r ( t , x ) , L p }} . (50)
Proof. With the sign convention L = J − ∫ 0 T R d t \mathcal{L}=J-\int_0^T R\,dt L = J − ∫ 0 T R d t , the variation with respect to the control gives gradient γ p m p + q ∗ \gamma_p m_p+q^* γ p m p + q ∗ . Let m ^ p \widehat m_p m p be the optimal control and let m p ∈ M p , a d m_p\in\mathcal{M}_{p,\mathrm{ad}} m p ∈ M p , ad be arbitrary. Since M p , a d \mathcal{M}_{p,\mathrm{ad}} M p , ad is convex, for any δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] the perturbed control m ^ p + δ ( m p − m ^ p ) ∈ M p , a d \widehat m_p+\delta(m_p-\widehat m_p)\in\mathcal{M}_{p,\mathrm{ad}} m p + δ ( m p − m p ) ∈ M p , ad . By optimality,
j ( m ^ p + δ ( m p − m ^ p ) ) ≥ j ( m ^ p ) , j(\widehat m_p+\delta(m_p-\widehat m_p))\ge j(\widehat m_p), j ( m p + δ ( m p − m p )) ≥ j ( m p ) ,
so dividing by δ > 0 \delta>0 δ > 0 and letting δ → 0 \delta\to0 δ → 0 yields j ′ ( m ^ p ) ( m p − m ^ p ) ≥ 0 j'(\widehat m_p)(m_p-\widehat m_p)\ge0 j ′ ( m p ) ( m p − m p ) ≥ 0 . Differentiating j j j in the direction m p − m ^ p m_p-\widehat m_p m p − m p and using the adjoint equations to express the state sensitivities, we obtain
j ′ ( m ^ p ) ( m p − m ^ p ) = ∫ 0 T ( γ p m ^ p + q ∗ ( m ^ p ) , m p − m ^ p ) Ω d t , j'(\widehat m_p)(m_p-\widehat m_p)=\int_0^T(\gamma_p\widehat m_p+q^*(\widehat m_p),m_p-\widehat m_p)_\Omega\,dt, j ′ ( m p ) ( m p − m p ) = ∫ 0 T ( γ p m p + q ∗ ( m p ) , m p − m p ) Ω d t ,
which gives (37). The converse direction (variational inequality implies optimality) follows because j j j is convex, since J J J is convex and the control-to-state map is affine; see [10 ] for general background.
The variational inequality (37) can be written pointwise a.e. in ( 0 , T ) × Ω (0,T)\times\Omega ( 0 , T ) × Ω as
( γ p m ^ p ( t , x ) + q ∗ ( t , x ) , η − m ^ p ( t , x ) ) ≥ 0 ∀ η ∈ [ l p , L p ] . (\gamma_p\widehat m_p(t,x)+q^*(t,x),\eta-\widehat m_p(t,x))\ge0
\qquad\forall\eta\in[l_p,L_p]. ( γ p m p ( t , x ) + q ∗ ( t , x ) , η − m p ( t , x )) ≥ 0 ∀ η ∈ [ l p , L p ] .
This is precisely the optimality condition for the scalar projection problem min η ∈ [ l p , L p ] 1 2 ∣ η + q ∗ / γ p ∣ 2 \min_{\eta\in[l_p,L_p]}\frac{1}{2}|\eta+q^*/\gamma_p|^2 min η ∈ [ l p , L p ] 2 1 ∣ η + q ∗ / γ p ∣ 2 , whose unique solution is η = P M p , a d ( − q ∗ / γ p ) \eta=\mathcal{P}_{\mathcal{M}_{p,\mathrm{ad}}}(-q^*/\gamma_p) η = P M p , ad ( − q ∗ / γ p ) as defined in (39). This gives (38). □ \square □
The optimality system is coupled through the adjoint right-hand side and the projection formula for the control.
Finite element discretization and discrete optimal control
Finite element spaces
Let { T h } h > 0 \{\mathcal{T}_h\}_{h>0} { T h } h > 0 be a family of shape-regular partitions of Ω ‾ \overline{\Omega} Ω into triangles (or tetrahedra for d = 3 d=3 d = 3 ) with mesh size h = max K ∈ T h h K h=\max_{K\in\mathcal{T}_h}h_K h = max K ∈ T h h K . We use the notation P k ( K ) \mathbb{P}_k(K) P k ( K ) for the space of polynomials of total degree at most k k k on K K K , and P k d c ( T h ) \mathbb{P}^{\mathrm{dc}}_k(\mathcal{T}_h) P k dc ( T h ) for the space of piecewise polynomials of degree k k k with no interelement continuity imposed.
For the displacement and total pressure pair ( u , p T ) (u,p_T) ( u , p T ) we use the lowest-order Taylor–Hood ( P 2 / P 1 ) (\mathbb{P}_2/\mathbb{P}_1) ( P 2 / P 1 ) pair:
V h : = { v h ∈ V : v h ∣ K ∈ P 2 ( K ) d ∀ K ∈ T h } , (51) V_h:=\{v_h\in V:v_h|_K\in\mathbb{P}_2(K)^d\ \forall K\in\mathcal{T}_h\}, \tag*{(51)} V h := { v h ∈ V : v h ∣ K ∈ P 2 ( K ) d ∀ K ∈ T h } , (51)
Q T , h : = { q h ∈ H 1 ( Ω ) : q h ∣ K ∈ P 1 ( K ) ∀ K ∈ T h } ⊂ H 1 ( Ω ) ⊂ L 2 ( Ω ) = Q T . (52) Q_{T,h}:=\{q_h\in H^1(\Omega):q_h|_K\in\mathbb{P}_1(K)\ \forall K\in\mathcal{T}_h\}\subset H^1(\Omega)\subset L^2(\Omega)=Q_T. \tag*{(52)} Q T , h := { q h ∈ H 1 ( Ω ) : q h ∣ K ∈ P 1 ( K ) ∀ K ∈ T h } ⊂ H 1 ( Ω ) ⊂ L 2 ( Ω ) = Q T . (52)
It is known that the pair ( V h , Q T , h ) (V_h,Q_{T,h}) ( V h , Q T , h ) satisfies the discrete inf-sup condition on the full pressure space Q T , h Q_{T,h} Q T , h :
inf 0 ≠ q T , h ∈ Q T , h sup v h ∈ V h ( q T , h , div v h ) Ω ∥ v h ∥ 1 ∥ q T , h ∥ 0 ≥ C > 0 , (53) \inf_{0\ne q_{T,h}\in Q_{T,h}}\sup_{v_h\in V_h}\frac{(q_{T,h},\operatorname{div}v_h)_\Omega}{\lVert v_h\rVert_1\lVert q_{T,h}\rVert_0}\ge C>0, \tag*{(53)} 0 = q T , h ∈ Q T , h inf v h ∈ V h sup ∥ v h ∥ 1 ∥ q T , h ∥ 0 ( q T , h , div v h ) Ω ≥ C > 0 , (53)
with C C C independent of h h h ; see, e.g., [3 ] .
The discrete inf-sup condition yields a right inverse R h : Q T , h → V h R_h:Q_{T,h}\to V_h R h : Q T , h → V h satisfying
b ( R h q T , h , q T , h ) = − ( r T , h , q T , h ) Ω ∀ r T , h , q T , h ∈ Q T , h , b(R_hq_{T,h},q_{T,h})=-(r_{T,h},q_{T,h})_\Omega\qquad\forall r_{T,h},q_{T,h}\in Q_{T,h}, b ( R h q T , h , q T , h ) = − ( r T , h , q T , h ) Ω ∀ r T , h , q T , h ∈ Q T , h ,
∥ R h q T , h ∥ V ≤ C R ∥ q T , h ∥ 0 , (54) \lVert R_hq_{T,h}\rVert_V\le C_R\lVert q_{T,h}\rVert_0, \tag*{(54)} ∥ R h q T , h ∥ V ≤ C R ∥ q T , h ∥ 0 , (54)
where C R C_R C R is independent of h h h , λ \lambda λ , and s 0 s_0 s 0 . It is the discrete analogue of (25) .
For p p p we use the standard linear Lagrange finite element:
Q h : = { q h ∈ H 1 ( Ω ) : q h ∣ K ∈ P 1 ( K ) ∀ K ∈ T h } ⊂ Q . (55) Q_h:=\{q_h\in H^1(\Omega):q_h|_K\in\mathbb{P}_1(K)\ \forall K\in\mathcal{T}_h\}\subset Q. \tag*{(55)} Q h := { q h ∈ H 1 ( Ω ) : q h ∣ K ∈ P 1 ( K ) ∀ K ∈ T h } ⊂ Q . (55)
Lemma 4.1. For given p ∈ L 2 ( Ω ) p\in L^2(\Omega) p ∈ L 2 ( Ω ) , let ( u h p , P T , h p ) ∈ V h × Q T , h (u_{h p},\mathcal{P}_{T,h}p)\in V_h\times Q_{T,h} ( u h p , P T , h p ) ∈ V h × Q T , h be the unique solution of
a u ( u h p , v h ) + b ( P T , h p , v h ) = 0 , (56) a_u(u_{h p},v_h)+b(\mathcal{P}_{T,h}p,v_h)=0, \tag*{(56)} a u ( u h p , v h ) + b ( P T , h p , v h ) = 0 , (56) b ( q T , h , u h p ) − λ − 1 ( P T , h p , q T , h ) Ω + λ − 1 α ( p , q T , h ) Ω = 0 (57) b(q_{T,h},u_{h p})-\lambda^{-1}(\mathcal{P}_{T,h}p,q_{T,h})_\Omega+\lambda^{-1}\alpha(p,q_{T,h})_\Omega=0 \tag*{(57)} b ( q T , h , u h p ) − λ − 1 ( P T , h p , q T , h ) Ω + λ − 1 α ( p , q T , h ) Ω = 0 (57) for all ( v h , q T , h ) ∈ V h × Q T , h (v_h,q_{T,h})\in V_h\times Q_{T,h} ( v h , q T , h ) ∈ V h × Q T , h . Then ∥ u h p ∥ V + ∥ P T , h p ∥ 0 ≤ C ∥ p ∥ 0 \lVert u_{h p}\rVert_V+\lVert\mathcal{P}_{T,h}p\rVert_0\le C\lVert p\rVert_0 ∥ u h p ∥ V + ∥ P T , h p ∥ 0 ≤ C ∥ p ∥ 0 , with C C C independent of h h h and s 0 s_0 s 0 , and uniform for λ ≥ λ 0 > 0 \lambda\ge\lambda_0>0 λ ≥ λ 0 > 0 . Define, for p h , q h ∈ Q h p_h,q_h\in Q_h p h , q h ∈ Q h ,
m e f f , h ( p h , q h ) : = − λ − 1 α ( P T , h p h , q h ) Ω + s 1 ( p h , q h ) Ω . (58) m_{\mathrm{eff},h}(p_h,q_h):=-\lambda^{-1}\alpha(\mathcal{P}_{T,h}p_h,q_h)_\Omega+s_1(p_h,q_h)_\Omega. \tag*{(58)} m eff , h ( p h , q h ) := − λ − 1 α ( P T , h p h , q h ) Ω + s 1 ( p h , q h ) Ω . (58) Then m e f f , h m_{\mathrm{eff},h} m eff , h is symmetric and bounded on Q h Q_h Q h , and for every q h ∈ Q h q_h\in Q_h q h ∈ Q h ,
m e f f , h ( q h , q h ) = s 0 ∥ q h ∥ 0 2 + a u ( U h q h , U h q h ) + λ − 1 ∥ P T , h q h − α q h ∥ 0 2 ≥ c e f f , h ∥ q h ∥ 0 2 (59) m_{\mathrm{eff},h}(q_h,q_h)=s_0\lVert q_h\rVert_0^2+a_u(\mathcal{U}_hq_h,\mathcal{U}_hq_h)+\lambda^{-1}\lVert\mathcal{P}_{T,h}q_h-\alpha q_h\rVert_0^2\ge c_{\mathrm{eff},h}\lVert q_h\rVert_0^2 \tag*{(59)} m eff , h ( q h , q h ) = s 0 ∥ q h ∥ 0 2 + a u ( U h q h , U h q h ) + λ − 1 ∥ P T , h q h − α q h ∥ 0 2 ≥ c eff , h ∥ q h ∥ 0 2 (59) where c e f f , h : = s 0 + α 2 λ + C R 2 c_{\mathrm{eff},h}:=s_0+\frac{\alpha^2}{\lambda+C_R^2} c eff , h := s 0 + λ + C R 2 α 2 with C R C_R C R from (54) .
Proof. Existence, uniqueness, and the bound for ( U h p , P T , h p ) (\mathcal{U}_hp,\mathcal{P}_{T,h}p) ( U h p , P T , h p ) follow from the discrete Babuška–Brezzi theory with (53) , as in Lemma 2.5. In the rest of proof, assume p = q h ∈ Q h p=q_h\in Q_h p = q h ∈ Q h . Then, symmetry follows by the argument of Lemma 2.6. Testing (45a) with U h q h \mathcal{U}_hq_h U h q h and (45b) with P T , h q h \mathcal{P}_{T,h}q_h P T , h q h gives
a u ( U h q h , U h q h ) = λ − 1 ( α ( q h , P T , h q h ) Ω − ∥ P T , h q h ∥ 0 2 ) . a_u(\mathcal{U}_hq_h,\mathcal{U}_hq_h)=\lambda^{-1}(\alpha(q_h,\mathcal{P}_{T,h}q_h)_\Omega-\lVert\mathcal{P}_{T,h}q_h\rVert_0^2). a u ( U h q h , U h q h ) = λ − 1 ( α ( q h , P T , h q h ) Ω − ∥ P T , h q h ∥ 0 2 ) . The first identity in (47) follows by the argument of Lemma 2.6. For the inequality in (47), since q h ∈ Q h ⊂ Q T , h q_h\in Q_h\subset Q_{T,h} q h ∈ Q h ⊂ Q T , h , R h q h R_hq_h R h q h is well-defined. Testing the first equation of (45) with v h = R h q h v_h=R_hq_h v h = R h q h and using
(43) with r T , h = P T , h q h r_{T,h}=\mathcal{P}_{T,h}q_h r T , h = P T , h q h gives a u ( U h q h , R h q h ) − ( P T , h q h , q h ) Ω = 0 a_u(\mathcal{U}_hq_h,R_hq_h)-(P_{T,h}q_h,q_h)_\Omega=0 a u ( U h q h , R h q h ) − ( P T , h q h , q h ) Ω = 0 . From this
α ∥ q h ∥ 0 2 = ( α q h − P T , h q h , q h ) Ω + a u ( U h , R h q h ) ≤ ( ∥ α q h − P T , h q h ∥ 0 + C R ∥ U h q h ∥ V ) ∥ q h ∥ 0 ≤ ( λ + C R 2 ) 1 2 ( λ − 1 ∥ α q h − P T , h q h ∥ 0 2 + ∥ U h q h ∥ V 2 ) 1 2 ∥ q h ∥ 0 . \begin{aligned}
\alpha\lVert q_h\rVert_0^2=(\alpha q_h-\mathcal{P}_{T,h}q_h,q_h)_\Omega+a_u(\mathcal{U}_h,R_hq_h) \\
&\leq\left(\lVert\alpha q_h-\mathcal{P}_{T,h}q_h\rVert_0+C_R\lVert\mathcal{U}_hq_h\rVert_V\right)\lVert q_h\rVert_0 \\
&\leq\left(\lambda+C_R^2\right)^{\frac{1}{2}}\left(\lambda^{-1}\lVert\alpha q_h-\mathcal{P}_{T,h}q_h\rVert_0^2+\lVert\mathcal{U}_hq_h\rVert_V^2\right)^{\frac{1}{2}}\lVert q_h\rVert_0.
\end{aligned} α ∥ q h ∥ 0 2 = ( α q h − P T , h q h , q h ) Ω + a u ( U h , R h q h ) ≤ ( ∥ α q h − P T , h q h ∥ 0 + C R ∥ U h q h ∥ V ) ∥ q h ∥ 0 ≤ ( λ + C R 2 ) 2 1 ( λ − 1 ∥ α q h − P T , h q h ∥ 0 2 + ∥ U h q h ∥ V 2 ) 2 1 ∥ q h ∥ 0 .
Combining it with the identity in (47), we get the inequality in (47).
Remark 4.2. The proof of Lemma 4.1 relies on Q h ⊂ Q T , h Q_h\subset Q_{T,h} Q h ⊂ Q T , h .
Time discretization and fully discrete state system
Let Δ t = T / N \Delta t=T/N Δ t = T / N be the time step size with positive integer N N N , and set t k = k Δ t t_k=k\Delta t t k = k Δ t for k = 0 , 1 , … , N k=0,1,\ldots,N k = 0 , 1 , … , N . We write I k : = ( t k − 1 , t k ] I_k:=(t_{k-1},t_k] I k := ( t k − 1 , t k ] , k = 1 , … , N k=1,\ldots,N k = 1 , … , N . For a time-dependent function z z z we denote its average over I k I_k I k by
z ˉ − k : = 1 Δ t ∫ I k z ( t ) d t . (60) \bar{z}^{-k}:=\frac{1}{\Delta t}\int_{I_k}z(t)\,\mathrm{d}t. \tag*{(60)} z ˉ − k := Δ t 1 ∫ I k z ( t ) d t . (60)
Superscripts without an overbar are reserved for discrete time levels (e.g. u h k u_h^k u h k ). For the control variable we use the space-time piecewise constant space
M p , h : = { m ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) : m ∣ I k × K ∈ P 0 ( I k ) ⊗ P 0 ( K ) ∀ k = 1 , … , N , K ∈ T h } , (61) M_{p,h}:=\left\{m\in L^2(0,T;L^2(\Omega)):m|_{I_k\times K}\in\mathbb{P}_0(I_k)\otimes\mathbb{P}_0(K)\ \forall\ k=1,\ldots,N,\ K\in\mathcal{T}_h\right\}, \tag*{(61)} M p , h := { m ∈ L 2 ( 0 , T ; L 2 ( Ω )) : m ∣ I k × K ∈ P 0 ( I k ) ⊗ P 0 ( K ) ∀ k = 1 , … , N , K ∈ T h } , (61)
i.e. functions that are piecewise constant on each space-time cell I k × K I_k\times K I k × K . For a sequence { g k } k ≥ 0 \{g^k\}_{k\geq0} { g k } k ≥ 0 we use the backward difference quotient
d t g k + 1 : = g k + 1 − g k Δ t . d_tg^{k+1}:=\frac{g^{k+1}-g^k}{\Delta t}. d t g k + 1 := Δ t g k + 1 − g k .
We treat { m p , h k } k = 1 N \{m_{p,h}^k\}_{k=1}^N { m p , h k } k = 1 N as space-time functions that are constant on each interval ( t k − 1 , t k ] (t_{k-1},t_k] ( t k − 1 , t k ] with values m p , h k m_{p,h}^k m p , h k , and similarly for the state and adjoint sequences. For a sequence { g k } k = 1 N \{g^k\}_{k=1}^N { g k } k = 1 N in a normed space X X X we write
∥ { g k } ∥ ℓ t 2 X 2 : = Δ t ∑ k = 1 N ∥ g k ∥ X 2 ; \left\|\{g^k\}\right\|_{\ell^2_tX}^2:=\Delta t\sum_{k=1}^N\lVert g^k\rVert_X^2; { g k } ℓ t 2 X 2 := Δ t k = 1 ∑ N ∥ g k ∥ X 2 ;
for piecewise constant-in-time functions this coincides with the L 2 ( 0 , T ; X ) L^2(0,T;X) L 2 ( 0 , T ; X ) norm. Suppose that the numerical initial data satisfy
∥ u ( 0 ) − u h 0 ∥ 1 + ∥ p T ( 0 ) − p T , h 0 ∥ 0 + ∥ p ( 0 ) − p h 0 ∥ 0 ≤ C i n i t h , \lVert u(0)-u_h^0\rVert_1+\lVert p_T(0)-p_{T,h}^0\rVert_0+\lVert p(0)-p_h^0\rVert_0\leq C_{\mathrm{init}}h, ∥ u ( 0 ) − u h 0 ∥ 1 + ∥ p T ( 0 ) − p T , h 0 ∥ 0 + ∥ p ( 0 ) − p h 0 ∥ 0 ≤ C init h ,
with C i n i t C_{\mathrm{init}} C init independent of h h h and Δ t \Delta t Δ t , and of λ ≥ λ 0 \lambda\geq\lambda_0 λ ≥ λ 0 for the λ \lambda λ -uniform estimates. Given m p ∈ M p , a d m_p\in\mathcal{M}_{p,\mathrm{ad}} m p ∈ M p , ad , the fully discrete state solution at the ( k + 1 ) (k+1) ( k + 1 ) -th time step is defined by backward Euler: find ( u h k + 1 , p T , h k + 1 , p h k + 1 ) ∈ V h × Q T , h × Q h (u_h^{k+1},p_{T,h}^{k+1},p_h^{k+1})\in V_h\times Q_{T,h}\times Q_h ( u h k + 1 , p T , h k + 1 , p h k + 1 ) ∈ V h × Q T , h × Q h such that
a u ( d t u h k + 1 , v h ) + b ( d t p T , h k + 1 , v h ) = 1 Δ t ∫ t k t k + 1 ( f , v h ) Ω d t , (62) a_u(d_tu_h^{k+1},v_h)+b(d_tp_{T,h}^{k+1},v_h)=\frac{1}{\Delta t}\int_{t_k}^{t_{k+1}}(f,v_h)_\Omega\,\mathrm{d}t, \tag*{(62)} a u ( d t u h k + 1 , v h ) + b ( d t p T , h k + 1 , v h ) = Δ t 1 ∫ t k t k + 1 ( f , v h ) Ω d t , (62)
b ( q T , h , d t u h k + 1 ) − λ − 1 ( d t p T , h k + 1 , q T , h ) Ω + λ − 1 α ( d t p h k + 1 , q T , h ) Ω = 0 , (63) b(q_{T,h},d_tu_h^{k+1})-\lambda^{-1}(d_tp_{T,h}^{k+1},q_{T,h})_\Omega+\lambda^{-1}\alpha(d_tp_h^{k+1},q_{T,h})_\Omega=0, \tag*{(63)} b ( q T , h , d t u h k + 1 ) − λ − 1 ( d t p T , h k + 1 , q T , h ) Ω + λ − 1 α ( d t p h k + 1 , q T , h ) Ω = 0 , (63)
− λ − 1 α ( d t p T , h k + 1 , q h ) Ω + s 1 ( d t p h k + 1 , q h ) Ω + ( κ p ∇ p h k + 1 , ∇ q h ) Ω = 1 Δ t ∫ t k t k + 1 ( m p , q h ) Ω d t , (64) -\lambda^{-1}\alpha(d_tp_{T,h}^{k+1},q_h)_\Omega+s_1(d_tp_h^{k+1},q_h)_\Omega+(\kappa_p\nabla p_h^{k+1},\nabla q_h)_\Omega=\frac{1}{\Delta t}\int_{t_k}^{t_{k+1}}(m_p,q_h)_\Omega\,\mathrm{d}t, \tag*{(64)} − λ − 1 α ( d t p T , h k + 1 , q h ) Ω + s 1 ( d t p h k + 1 , q h ) Ω + ( κ p ∇ p h k + 1 , ∇ q h ) Ω = Δ t 1 ∫ t k t k + 1 ( m p , q h ) Ω d t , (64)
for all ( v h , q T , h , q h ) ∈ V h × Q T , h × Q h (v_h,q_{T,h},q_h)\in V_h\times Q_{T,h}\times Q_h ( v h , q T , h , q h ) ∈ V h × Q T , h × Q h . For the discrete optimal control problem, m p = m p , h ∈ M p , a d h m_p=m_{p,h}\in\mathcal{M}_{p,\mathrm{ad}}^h m p = m p , h ∈ M p , ad h , and the right-hand side of (50c) is ( m p , h k + 1 , q h ) Ω (m_{p,h}^{k+1},q_h)_\Omega ( m p , h k + 1 , q h ) Ω .
Lemma 4.3 (discrete forward stability). Let ( w h l , ξ T , h l , r h l ) l = 0 N ⊂ V h × Q T , h × Q h (w_h^l,\xi_{T,h}^l,r_h^l)_{l=0}^N \subset V_h \times Q_{T,h} \times Q_h ( w h l , ξ T , h l , r h l ) l = 0 N ⊂ V h × Q T , h × Q h satisfy, for l = 1 , … , N l=1,\ldots,N l = 1 , … , N ,
a u ( d t w h l , v h ) + b ( d t ξ h l , v h ) = ( F w l , v h ) Ω , (65) a_u(d_t w_h^l,v_h)+b(d_t\xi_h^l,v_h)=(F_w^l,v_h)_\Omega, \tag*{(65)} a u ( d t w h l , v h ) + b ( d t ξ h l , v h ) = ( F w l , v h ) Ω , (65) b ( q T , h , d t w h l ) − λ − 1 ( d t ξ h l , q T , h ) Ω + λ − 1 α ( d t r h l , q T , h ) Ω = 0 , (66) b(q_{T,h},d_t w_h^l)-\lambda^{-1}(d_t\xi_h^l,q_{T,h})_\Omega+\lambda^{-1}\alpha(d_t r_h^l,q_{T,h})_\Omega=0, \tag*{(66)} b ( q T , h , d t w h l ) − λ − 1 ( d t ξ h l , q T , h ) Ω + λ − 1 α ( d t r h l , q T , h ) Ω = 0 , (66) − λ − 1 α ( d t ξ h l , q h ) Ω + s 1 ( d t r h l , q h ) Ω + ( κ p ∇ r h l , ∇ q h ) Ω = ( F r l , q h ) Ω . (67) -\lambda^{-1}\alpha(d_t\xi_h^l,q_h)_\Omega+s_1(d_t r_h^l,q_h)_\Omega+(\kappa_p\nabla r_h^l,\nabla q_h)_\Omega=(F_r^l,q_h)_\Omega. \tag*{(67)} − λ − 1 α ( d t ξ h l , q h ) Ω + s 1 ( d t r h l , q h ) Ω + ( κ p ∇ r h l , ∇ q h ) Ω = ( F r l , q h ) Ω . (67) Assume s 0 ≥ 0 s_0\ge0 s 0 ≥ 0 and zero initial data. Then
max 1 ≤ l ≤ N ( ∥ w h l ∥ V + ∥ ξ h l ∥ 0 + ∥ r h l ∥ 0 ) ≤ C e f f ( Δ t ∑ l = 1 N ( ∥ F w l ∥ V ∗ 2 + ∥ F r l ∥ 0 2 ) ) 1 / 2 , (68) \max_{1\le l\le N}\left(\lVert w_h^l\rVert_V+\lVert\xi_h^l\rVert_0+\lVert r_h^l\rVert_0\right)\le C_{\mathrm{eff}}\left(\Delta t\sum_{l=1}^{N}\left(\lVert F_w^l\rVert_{V^*}^2+\lVert F_r^l\rVert_0^2\right)\right)^{1/2}, \tag*{(68)} 1 ≤ l ≤ N max ( ∥ w h l ∥ V + ∥ ξ h l ∥ 0 + ∥ r h l ∥ 0 ) ≤ C eff ( Δ t l = 1 ∑ N ( ∥ F w l ∥ V ∗ 2 + ∥ F r l ∥ 0 2 ) ) 1/2 , (68) where C e f f C_{\mathrm{eff}} C eff is independent of h h h , Δ t \Delta t Δ t , but depends on c e f f , h − 1 c_{\mathrm{eff},h}^{-1} c eff , h − 1 in (59) .
Proof. At each time level, split the differentiated mechanical variables as
d t w h l = w F , h l + U h ( d t r h l ) , d t ξ h l = ξ F , h l + P T , h ( d t r h l ) , d_t w_h^l=w_{F,h}^l+U_h(d_t r_h^l),\qquad d_t\xi_h^l=\xi_{F,h}^l+P_{T,h}(d_t r_h^l), d t w h l = w F , h l + U h ( d t r h l ) , d t ξ h l = ξ F , h l + P T , h ( d t r h l ) , where ( w F , h l , ξ F , h l ) (w_{F,h}^l,\xi_{F,h}^l) ( w F , h l , ξ F , h l ) solves the discrete mechanical problem with right-hand side F w l F_w^l F w l and zero scalar datum. The discrete saddle-point estimate gives
∥ w F , h l ∥ V + ∥ ξ F , h l ∥ 0 ≤ C ∥ F w l ∥ V ∗ , \lVert w_{F,h}^l\rVert_V+\lVert\xi_{F,h}^l\rVert_0\le C\lVert F_w^l\rVert_{V^*}, ∥ w F , h l ∥ V + ∥ ξ F , h l ∥ 0 ≤ C ∥ F w l ∥ V ∗ , uniformly in h h h . Substituting into (67) gives
m e f f , h ( d t r h l , q h ) + ( κ p ∇ r h l , ∇ q h ) Ω = ( F r l , q h ) Ω + λ − 1 α ( ξ F , h l , q h ) Ω . m_{\mathrm{eff},h}(d_t r_h^l,q_h)+(\kappa_p\nabla r_h^l,\nabla q_h)_\Omega=(F_r^l,q_h)_\Omega+\lambda^{-1}\alpha(\xi_{F,h}^l,q_h)_\Omega. m eff , h ( d t r h l , q h ) + ( κ p ∇ r h l , ∇ q h ) Ω = ( F r l , q h ) Ω + λ − 1 α ( ξ F , h l , q h ) Ω . Taking q h = d t r h l q_h=d_t r_h^l q h = d t r h l , using Lemma 4.1, multiplying by Δ t \Delta t Δ t , summing, and applying Young’s inequality yield an ℓ t 2 L x 2 \ell_t^2L_x^2 ℓ t 2 L x 2 bound for d t r h d_t r_h d t r h . The mechanical decomposition and Lemma 4.1 give the corresponding bounds for d t w h d_t w_h d t w h and d t ξ h d_t\xi_h d t ξ h . Since the initial data vanish, discrete summation in time and Cauchy–Schwarz give (68) . □
Remark 4.4. The state system and the state-difference system have exactly the sign pattern of (51). The adjoint system, after time reversal, has the same stability structure as the forward state scheme.
Lemma 4.5 (Well-posedness and stability of the discrete state system). The fully discrete state system (50) is well-posed at each time step, and its solution with zero initial data satisfies
max 1 ≤ k ≤ N ( ∥ u h k ∥ V + ∥ p T , h k ∥ 0 + ∥ p h k ∥ 0 ) ≤ C ( Δ t ∑ k = 1 N ( ∥ f u k ∥ V ∗ 2 + ∥ f p k ∥ 0 2 ) ) 1 / 2 , (69) \max_{1\le k\le N}\left(\lVert u_h^k\rVert_V+\lVert p_{T,h}^k\rVert_0+\lVert p_h^k\rVert_0\right)\le C\left(\Delta t\sum_{k=1}^{N}\left(\lVert f_u^k\rVert_{V^*}^2+\lVert f_p^k\rVert_0^2\right)\right)^{1/2}, \tag*{(69)} 1 ≤ k ≤ N max ( ∥ u h k ∥ V + ∥ p T , h k ∥ 0 + ∥ p h k ∥ 0 ) ≤ C ( Δ t k = 1 ∑ N ( ∥ f u k ∥ V ∗ 2 + ∥ f p k ∥ 0 2 ) ) 1/2 , (69) where f u k : = 1 Δ t ∫ t k − 1 t k f d t f_u^k:=\frac{1}{\Delta t}\int_{t_{k-1}}^{t_k}f\,dt f u k := Δ t 1 ∫ t k − 1 t k f d t and f p k : = 1 Δ t ∫ t k − 1 t k m p d t f_p^k:=\frac{1}{\Delta t}\int_{t_{k-1}}^{t_k}m_p\,dt f p k := Δ t 1 ∫ t k − 1 t k m p d t are the right-hand sides of (62) and (64) at step k k k , with C C C independent of h h h , Δ t \Delta t Δ t .
Proof. It is straightforward from Lemma 4.3, so we omit details. □
Discrete optimal control problem
We define the discrete admissible set
M p , a d h : = M p , a d ∩ M p , h = { m h ∈ M p , h : l p ≤ m h ≤ L p a.e. } . \mathcal{M}_{p,ad}^h:=\mathcal{M}_{p,ad}\cap\mathcal{M}_{p,h}=\{m_h\in\mathcal{M}_{p,h}:l_p\le m_h\le L_p\ \text{a.e.}\}. M p , a d h := M p , a d ∩ M p , h = { m h ∈ M p , h : l p ≤ m h ≤ L p a.e. } .
Note that M p , a d h ⊂ M p , a d \mathcal{M}_{p,ad}^h\subset\mathcal{M}_{p,ad} M p , a d h ⊂ M p , a d since M p , h \mathcal{M}_{p,h} M p , h is a piecewise constant subspace of L t 2 L x 2 L_t^2L_x^2 L t 2 L x 2 .
We also introduce the L 2 ( ( 0 , T ) × Ω ) L^2((0,T)\times\Omega) L 2 (( 0 , T ) × Ω ) -orthogonal projection Π h M : L t 2 L x 2 → M p , h \Pi_h^M:L_t^2L_x^2\to\mathcal{M}_{p,h} Π h M : L t 2 L x 2 → M p , h , defined cellwise in space and time by
( Π h M m ) ∣ ( t k − 1 , t k ) × K : = 1 Δ t ∣ K ∣ ∫ t k − 1 t k ∫ K m d x d t , K ∈ T h , k = 1 , … , N . (70) \left.(\Pi_h^M m)\right|_{(t_{k-1},t_k)\times K}:=\frac{1}{\Delta t\lvert K\rvert}\int_{t_{k-1}}^{t_k}\int_K m\,dx\,dt,\qquad K\in\mathcal{T}_h,\ k=1,\ldots,N. \tag*{(70)} ( Π h M m ) ( t k − 1 , t k ) × K := Δ t ∣ K ∣ 1 ∫ t k − 1 t k ∫ K m d x d t , K ∈ T h , k = 1 , … , N . (70)
Since Π h M m \Pi_h^M m Π h M m is a cellwise space-time average of m m m , if l p ≤ m ≤ L p l_p \le m \le L_p l p ≤ m ≤ L p a.e. then l p ≤ Π h M m ≤ L p l_p \le\Pi_h^M m \le L_p l p ≤ Π h M m ≤ L p a.e.; hence Π h M \Pi_h^M Π h M maps M p , a d \mathcal{M}_{p,ad} M p , a d into M p , a d h \mathcal{M}_{p,ad}^h M p , a d h :
Π h M : M p , a d → M p , a d h . (71) \Pi_h^M:\mathcal{M}_{p,ad}\to\mathcal{M}_{p,ad}^h. \tag*{(71)} Π h M : M p , a d → M p , a d h . (71)
The discrete objective functional is
J h ( { u h k } k = 1 N , { p h k } k = 1 N , { m p , h k } k = 1 N ) J_h\left(\{u_h^k\}_{k=1}^{N},\{p_h^k\}_{k=1}^{N},\{m_{p,h}^k\}_{k=1}^{N}\right) J h ( { u h k } k = 1 N , { p h k } k = 1 N , { m p , h k } k = 1 N )
: = α u , C 2 Δ t ∑ k = 0 N − 1 ∥ u h k + 1 − u ˉ C k + 1 ∥ 0 2 + α p , C 2 Δ t ∑ k = 0 N − 1 ∥ p h k + 1 − p ˉ C k + 1 ∥ 0 2 + γ p 2 ∥ { m p , h k } k = 1 N ∥ ℓ t 2 L x 2 2 , (72) :=\frac{\alpha_{u,C}}{2}\Delta t\sum_{k=0}^{N-1}\left\|u_h^{k+1}-\bar{u}_C^{k+1}\right\|_0^2+\frac{\alpha_{p,C}}{2}\Delta t\sum_{k=0}^{N-1}\left\|p_h^{k+1}-\bar{p}_C^{k+1}\right\|_0^2+\frac{\gamma_p}{2}\left\|\{m_{p,h}^k\}_{k=1}^{N}\right\|_{\ell^2_tL_x^2}^2, \tag*{(72)} := 2 α u , C Δ t k = 0 ∑ N − 1 u h k + 1 − u ˉ C k + 1 0 2 + 2 α p , C Δ t k = 0 ∑ N − 1 p h k + 1 − p ˉ C k + 1 0 2 + 2 γ p { m p , h k } k = 1 N ℓ t 2 L x 2 2 , (72)
subject to the fully discrete state system (50), where ( u h N , p h N ) (u_h^N,p_h^N) ( u h N , p h N ) is the state at the final time step t N = T t_N=T t N = T . As in the continuous setting, we use the discrete reduced functional j h ( m p , h ) : = J h ( { u h k } , { p h k } , { m p , h k } ) j_h(m_{p,h}):=J_h(\{u_h^k\},\{p_h^k\},\{m_{p,h}^k\}) j h ( m p , h ) := J h ({ u h k } , { p h k } , { m p , h k }) on M p , a d h \mathcal{M}_{p,ad}^h M p , a d h , where ( u h k , p T , h k , p h k ) (u_h^k,p_{T,h}^k,p_h^k) ( u h k , p T , h k , p h k ) solves (50) with control m p , h m_{p,h} m p , h .
Theorem 4.6. There exists a unique solution m ^ p , h ∈ M p , a d h \widehat{m}_{p,h}\in\mathcal{M}_{p,ad}^h m p , h ∈ M p , a d h to the discrete optimal control problem (50)–(72) .
Proof. The proof follows the same lines as the continuous case in Theorem 3.2 , so we omit here. □ \square □
Discrete first-order optimality system
Theorem 4.7. A control m ^ p , h ∈ M p , a d h \widehat{m}_{p,h}\in\mathcal{M}_{p,ad}^h m p , h ∈ M p , a d h and the corresponding state { ( u h k + 1 , p T , h k + 1 , p h k + 1 ) } k = 0 N − 1 \{(u_h^{k+1},p_{T,h}^{k+1},p_h^{k+1})\}_{k=0}^{N-1} {( u h k + 1 , p T , h k + 1 , p h k + 1 ) } k = 0 N − 1 is the optimal solution to (50)–(72) if and only if there exists an adjoint state { ( v h ∗ , k , q T , h ∗ , k , q h ∗ , k ) } k = 0 N − 1 \{(v_h^{*,k},q_{T,h}^{*,k},q_h^{*,k})\}_{k=0}^{N-1} {( v h ∗ , k , q T , h ∗ , k , q h ∗ , k ) } k = 0 N − 1 satisfying the discrete adjoint equations
− a u ( d t v h ∗ , k + 1 , v h ) − b ( d t q T , h ∗ , k + 1 , v h ) = α u , C ( u h k + 1 − u ˉ C k + 1 , v h ) Ω , (73) -a_u(dt v_h^{*,k+1},v_h)-b(dt q_{T,h}^{*,k+1},v_h)=\alpha_{u,C}(u_h^{k+1}-\bar{u}_C^{k+1},v_h)_\Omega, \tag*{(73)} − a u ( d t v h ∗ , k + 1 , v h ) − b ( d t q T , h ∗ , k + 1 , v h ) = α u , C ( u h k + 1 − u ˉ C k + 1 , v h ) Ω , (73) − b ( q T , h , d t v h ∗ , k + 1 ) + λ − 1 ( d t q T , h ∗ , k + 1 , q T , h ) Ω + λ − 1 α ( d t q h ∗ , k + 1 , q T , h ) Ω = 0 , (74) -b(q_{T,h},dt v_h^{*,k+1})+\lambda^{-1}(dt q_{T,h}^{*,k+1},q_{T,h})_\Omega+\lambda^{-1}\alpha(dt q_h^{*,k+1},q_{T,h})_\Omega=0, \tag*{(74)} − b ( q T , h , d t v h ∗ , k + 1 ) + λ − 1 ( d t q T , h ∗ , k + 1 , q T , h ) Ω + λ − 1 α ( d t q h ∗ , k + 1 , q T , h ) Ω = 0 , (74) − λ − 1 α ( d t q T , h ∗ , k + 1 , q h ) Ω − s 1 ( d t q h ∗ , k + 1 , q h ) Ω + ( κ p ∇ q h ∗ , k , ∇ q h ) Ω = α p , C ( p h k + 1 − p ˉ C k + 1 , q h ) Ω , (75) -\lambda^{-1}\alpha(dt q_{T,h}^{*,k+1},q_h)_\Omega-s_1(dt q_h^{*,k+1},q_h)_\Omega+(\kappa_p\nabla q_h^{*,k},\nabla q_h)_\Omega=\alpha_{p,C}(p_h^{k+1}-\bar{p}_C^{k+1},q_h)_\Omega, \tag*{(75)} − λ − 1 α ( d t q T , h ∗ , k + 1 , q h ) Ω − s 1 ( d t q h ∗ , k + 1 , q h ) Ω + ( κ p ∇ q h ∗ , k , ∇ q h ) Ω = α p , C ( p h k + 1 − p ˉ C k + 1 , q h ) Ω , (75) for all ( v h , q T , h , q h ) ∈ V h × Q T , h × Q h (v_h,q_{T,h},q_h)\in V_h\times Q_{T,h}\times Q_h ( v h , q T , h , q h ) ∈ V h × Q T , h × Q h and k = 0 , 1 , … , N − 1 k=0,1,\ldots,N-1 k = 0 , 1 , … , N − 1 , where u ˉ C k + 1 \bar{u}_C^{k+1} u ˉ C k + 1 , p ˉ C k + 1 \bar{p}_C^{k+1} p ˉ C k + 1 denote the cell-average values of the target fields on the interval ( t k , t k + 1 ) (t_k,t_{k+1}) ( t k , t k + 1 ) .
v h ∗ , N = 0 , q T , h ∗ , N = 0 , q h ∗ , N = 0. (76) v_h^{*,N}=0,\qquad q_{T,h}^{*,N}=0,\qquad q_h^{*,N}=0. \tag*{(76)} v h ∗ , N = 0 , q T , h ∗ , N = 0 , q h ∗ , N = 0. (76) Together with the discrete variational inequality
Δ t ∑ k = 0 N − 1 ( γ p m ^ p , h k + 1 + q h ∗ , k ( m ^ p , h ) , z p k + 1 − m ^ p , h k + 1 ) Ω ≥ 0 ∀ z p ∈ M p , a d h . (77) \Delta t\sum_{k=0}^{N-1}\left(\gamma_p\widehat{m}_{p,h}^{k+1}+q_h^{*,k}(\widehat{m}_{p,h}),z_p^{k+1}-\widehat{m}_{p,h}^{k+1}\right)_\Omega\ge0\qquad\forall z_p\in\mathcal{M}_{p,ad}^h. \tag*{(77)} Δ t k = 0 ∑ N − 1 ( γ p m p , h k + 1 + q h ∗ , k ( m p , h ) , z p k + 1 − m p , h k + 1 ) Ω ≥ 0 ∀ z p ∈ M p , a d h . (77) Moreover, (77) is equivalent to the cellwise projection formula: for k = 1 , 2 , … , N k=1,2,\ldots,N k = 1 , 2 , … , N and K ∈ T h K\in\mathcal{T}_h K ∈ T h ,
m ^ p , h k ∣ K = max { l p , min { − 1 γ p ∣ K ∣ ∫ K q h ∗ , k − 1 ( m ^ p , h ) d x , L p } } . (78) \widehat{m}_{p,h}^k\big|_K=\max\left\{l_p,\min\left\{-\frac{1}{\gamma_p|K|}\int_Kq_h^{*,k-1}(\widehat{m}_{p,h})\,dx,L_p\right\}\right\}. \tag*{(78)} m p , h k K = max { l p , min { − γ p ∣ K ∣ 1 ∫ K q h ∗ , k − 1 ( m p , h ) d x , L p } } . (78) Proof. The proof follows the same lines as the continuous Lemma 3.4 . By the optimality of m ^ p , h \widehat{m}_{p,h} m p , h and the convexity of M p , a d h \mathcal{M}_{p,ad}^h M p , a d h , for any admissible perturbation direction and δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] the Gâteaux derivative of j h j_h j h satisfies j h ′ ( m ^ p , h ) ( z p − m ^ p , h ) ≥ 0 j_h'(\widehat{m}_{p,h})(z_p-\widehat{m}_{p,h})\ge0 j h ′ ( m p , h ) ( z p − m p , h ) ≥ 0 . Differentiating j h j_h j h using the discrete state equations (50) and the discrete adjoint equations (57)–(76) to represent state sensitivities gives (77) . Testing (77) with z p z_p z p equal to m ^ p , h \widehat{m}_{p,h} m p , h except on one space-time cell I k × K I_k\times K I k × K , where z p = η ∈ [ l p , L p ] z_p=\eta\in[l_p,L_p] z p = η ∈ [ l p , L p ] , and dividing by Δ t ∣ K ∣ > 0 \Delta t|K|>0 Δ t ∣ K ∣ > 0 gives the scalar variational inequality ( γ p m ^ p , h k ∣ K + ∣ K ∣ − 1 ∫ K q h ∗ , k − 1 d x ) ( η − m ^ p , h k ∣ K ) ≥ 0 (\gamma_p\widehat{m}_{p,h}^k|_K+|K|^{-1}\int_Kq_h^{*,k-1}\,dx)(\eta-\widehat{m}_{p,h}^k|_K)\ge0 ( γ p m p , h k ∣ K + ∣ K ∣ − 1 ∫ K q h ∗ , k − 1 d x ) ( η − m p , h k ∣ K ) ≥ 0 for all η ∈ [ l p , L p ] \eta\in[l_p,L_p] η ∈ [ l p , L p ] , whose unique solution is (78) . Conversely, (78) gives these cellwise inequalities, and summing them over all cells gives (77) . □ \square □
Remark 4.8. The control value on ( t k − 1 , t k ] (t_{k-1},t_k] ( t k − 1 , t k ] is paired with q h ∗ , k − 1 q_h^{*,k-1} q h ∗ , k − 1 . The terminal value q h ∗ , N = 0 q_h^{*,N}=0 q h ∗ , N = 0 serves only as starting datum for the backward adjoint recursion.
We record a stability estimate for the discrete adjoint system with general interior right-hand sides and homogeneous terminal data, which is the discrete analogue of the backward parabolic energy estimate. It is used in the sensitivity and adjoint error estimates of § 5 .
Lemma 4.9 (Discrete adjoint stability). Let ( v h ∗ , k , q T , h ∗ , k , q h ∗ , k ) k = 0 N − 1 (v_h^{*,k},q_{T,h}^{*,k},q_h^{*,k})_{k=0}^{N-1} ( v h ∗ , k , q T , h ∗ , k , q h ∗ , k ) k = 0 N − 1 solve the discrete adjoint recursion (57) with general interior data α u , C F u k + 1 \alpha_{u,C}F_u^{k+1} α u , C F u k + 1 and α p , C F p k + 1 \alpha_{p,C}F_p^{k+1} α p , C F p k + 1 and with homogeneous terminal condition. Then
max 0 ≤ k ≤ N − 1 ( ∥ v h ∗ , k ∥ V + ∥ q T , h ∗ , k ∥ 0 + ∥ q h ∗ , k ∥ 0 ) ≤ C e f f ( Δ t ∑ j = 1 N ( ∥ F u j ∥ 0 2 + ∥ F p j ∥ 0 2 ) ) 1 / 2 , (79) \max_{0\leq k\leq N-1}\left(\lVert v_h^{*,k}\rVert_V+\lVert q_{T,h}^{*,k}\rVert_0+\lVert q_h^{*,k}\rVert_0\right)\leq C_{\mathrm{eff}}\left(\Delta t\sum_{j=1}^{N}\left(\lVert F_u^j\rVert_0^2+\lVert F_p^j\rVert_0^2\right)\right)^{1/2}, \tag*{(79)} 0 ≤ k ≤ N − 1 max ( ∥ v h ∗ , k ∥ V + ∥ q T , h ∗ , k ∥ 0 + ∥ q h ∗ , k ∥ 0 ) ≤ C eff ( Δ t j = 1 ∑ N ( ∥ F u j ∥ 0 2 + ∥ F p j ∥ 0 2 ) ) 1/2 , (79) where C e f f C_{\mathrm{eff}} C eff is independent of h h h , Δ t \Delta t Δ t , but may depend on the lower bound of c e f f , h c_{\mathrm{eff},h} c eff , h .
Proof. Reverse time by setting ( v ~ h l , q ~ T , h l , q ~ h l ) = ( v h ∗ , N − l , q T , h ∗ , N − l , q h ∗ , N − l ) (\widetilde v_h^l,\widetilde q_{T,h}^l,\widetilde q_h^l)=(v_h^{*,N-l},q_{T,h}^{*,N-l},q_h^{*,N-l}) ( v h l , q T , h l , q h l ) = ( v h ∗ , N − l , q T , h ∗ , N − l , q h ∗ , N − l ) . The reversed equations are, for l = 1 , … , N l=1,\ldots,N l = 1 , … , N ,
a u ( d t v ~ h l , v h ) + b ( d t q ~ T , h l , v h ) = α u , C ( F u N − l + 1 , v h ) Ω , (80) a_u(\mathrm{d}t\widetilde v_h^l,v_h)+b(\mathrm{d}t\widetilde q_{T,h}^l,v_h)=\alpha_{u,C}(F_u^{N-l+1},v_h)_{\Omega}, \tag*{(80)} a u ( d t v h l , v h ) + b ( d t q T , h l , v h ) = α u , C ( F u N − l + 1 , v h ) Ω , (80)
b ( q T , h , d t v ~ h l ) − λ − 1 ( d t q ~ T , h l , q T , h ) Ω − λ − 1 α ( d t q ~ h l , q T , h ) Ω = 0 , (81) b(q_{T,h},\mathrm{d}t\widetilde v_h^l)-\lambda^{-1}(\mathrm{d}t\widetilde q_{T,h}^l,q_{T,h})_{\Omega}-\lambda^{-1}\alpha(\mathrm{d}t\widetilde q_h^l,q_{T,h})_{\Omega}=0, \tag*{(81)} b ( q T , h , d t v h l ) − λ − 1 ( d t q T , h l , q T , h ) Ω − λ − 1 α ( d t q h l , q T , h ) Ω = 0 , (81)
λ − 1 α ( d t q ~ T , h l , q h ) Ω + s 1 ( d t q ~ h l , q h ) Ω + ( κ p ∇ q ~ h l , ∇ q h ) Ω = α p , C ( F p N − l + 1 , q h ) Ω . (82) \lambda^{-1}\alpha(\mathrm{d}t\widetilde q_{T,h}^l,q_h)_{\Omega}+s_1(\mathrm{d}t\widetilde q_h^l,q_h)_{\Omega}+(\kappa_p\nabla\widetilde q_h^l,\nabla q_h)_{\Omega}=\alpha_{p,C}(F_p^{N-l+1},q_h)_{\Omega}. \tag*{(82)} λ − 1 α ( d t q T , h l , q h ) Ω + s 1 ( d t q h l , q h ) Ω + ( κ p ∇ q h l , ∇ q h ) Ω = α p , C ( F p N − l + 1 , q h ) Ω . (82)
Their mechanical block differs from the forward state only in the sign of the scalar coupling. Consequently,
d t v ~ h l = w F , h l − U h ( d t q ~ h l ) , d t q ~ T , h l = ξ F , h l − P T , h ( d t q ~ h l ) , \mathrm{d}t\widetilde v_h^l=w_{F,h}^l-U_h(\mathrm{d}t\widetilde q_h^l),\qquad\mathrm{d}t\widetilde q_{T,h}^l=\xi_{F,h}^l-P_{T,h}(\mathrm{d}t\widetilde q_h^l), d t v h l = w F , h l − U h ( d t q h l ) , d t q T , h l = ξ F , h l − P T , h ( d t q h l ) ,
where ( w F , h l , ξ F , h l ) (w_{F,h}^l,\xi_{F,h}^l) ( w F , h l , ξ F , h l ) is the mechanical lift of the reversed displacement datum. Substitution in the third reversed adjoint equation gives
m e f f , h ( d t q ~ h l , q h ) + ( κ p ∇ q ~ h l , ∇ q h ) Ω = α p , C ( F p N − l + 1 , q h ) Ω − λ − 1 α ( ξ F , h l , q h ) Ω . m_{\mathrm{eff},h}(\mathrm{d}t\widetilde q_h^l,q_h)+(\kappa_p\nabla\widetilde q_h^l,\nabla q_h)_{\Omega}=\alpha_{p,C}(F_p^{N-l+1},q_h)_{\Omega}-\lambda^{-1}\alpha(\xi_{F,h}^l,q_h)_{\Omega}. m eff , h ( d t q h l , q h ) + ( κ p ∇ q h l , ∇ q h ) Ω = α p , C ( F p N − l + 1 , q h ) Ω − λ − 1 α ( ξ F , h l , q h ) Ω .
Thus the same effective mass m e f f , h m_{\mathrm{eff},h} m eff , h occurs as in the forward reduction. Testing with q h = d t q ~ h l q_h=\mathrm{d}t\widetilde q_h^l q h = d t q h l , summing in time, and using Lemma 4.1 gives the scalar derivative bound without s 0 − 1 s_0^{-1} s 0 − 1 . Mechanical reconstruction gives the remaining derivative bounds. Since the reversed initial data are zero, discrete integration in time yields (61); reversing time completes the proof. □
A priori error analysis
The well-posedness and stability results above are valid for s 0 ≥ 0 s_0\geq0 s 0 ≥ 0 with constants independent of s 0 − 1 s_0^{-1} s 0 − 1 , at the price that the effective-mass coercivity constant may depend on λ \lambda λ . In this section, when a λ \lambda λ -uniform error constant is asserted, we use the alternative bound m e f f ( p , p ) ≥ s 0 ∥ p ∥ 0 2 m_{\mathrm{eff}}(p,p)\geq s_0\lVert p\rVert_0^2 m eff ( p , p ) ≥ s 0 ∥ p ∥ 0 2 and assume that s 0 s_0 s 0 is fixed positive. Thus storage-robustness and near-incompressibility robustness are distinguished explicitly.
In this section we derive a priori error estimates for the fully discrete scheme (50)–(60). Throughout this section, we assume that the family of triangulations is shape-regular and quasi-uniform.
Regularity assumption (Assumption R). Throughout § 5 we assume that the continuous optimal state and adjoint satisfy
u ^ , v ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; V ) , \widehat u,\widehat v^*\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;V), u , v ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; V ) , p ^ T , q ^ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) ∩ H 1 ( 0 , T ; Q T ) , \widehat p_T,\widehat q_T^*\in L^2(0,T;H^1(\Omega))\cap H^1(0,T;Q_T), p T , q T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω )) ∩ H 1 ( 0 , T ; Q T ) , p ^ , q ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) , (83) \widehat p,\widehat q^*\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;L^2(\Omega)), \tag*{(83)} p , q ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; L 2 ( Ω )) , (83)
with all constants uniform for λ ≥ λ 0 > 0 \lambda\ge\lambda_0 > 0 λ ≥ λ 0 > 0 . Sufficient conditions for (83) are established in Appendix A. The control regularity used below follows from the projection formula, as shown in Remark 5.13 .
For the continuous optimal solution we write ( u ^ , p ^ T , p ^ , m ^ p , v ^ ∗ , q ^ T ∗ , q ^ ∗ ) (\hat{u},\hat{p}_T,\hat{p},\hat{m}_p,\hat{v}^*,\hat{q}_T^*,\hat{q}^*) ( u ^ , p ^ T , p ^ , m ^ p , v ^ ∗ , q ^ T ∗ , q ^ ∗ ) , and for the discrete optimal solution { ( u h k + 1 , p T , h k + 1 , p h k + 1 , m p , h k + 1 , v h ∗ , k , q T , h ∗ , k , q h ∗ , k ) } \{(u_h^{k+1},p_{T,h}^{k+1},p_h^{k+1},m_{p,h}^{k+1},v_h^{*,k},q_{T,h}^{*,k},q_h^{*,k})\} {( u h k + 1 , p T , h k + 1 , p h k + 1 , m p , h k + 1 , v h ∗ , k , q T , h ∗ , k , q h ∗ , k )} .
Auxiliary discrete systems
For any m p ∈ M p , a d m_p \in\mathcal{M}_{p,ad} m p ∈ M p , a d we define the auxiliary fully discrete state solution ( u h k + 1 ( m p ) , p T , h k + 1 ( m p ) , p h k + 1 ( m p ) ) (u_h^{k+1}(m_p),p_{T,h}^{k+1}(m_p),p_h^{k+1}(m_p)) ( u h k + 1 ( m p ) , p T , h k + 1 ( m p ) , p h k + 1 ( m p )) by (50) with right-hand side 1 Δ t ∫ t k t k + 1 ζ t k t k + 1 ( m p , q h ) d t \frac{1}{\Delta t}\int_{t_k}^{t_{k+1}} \zeta_{t_k}^{t_{k+1}}(m_p,q_h)\,d t Δ t 1 ∫ t k t k + 1 ζ t k t k + 1 ( m p , q h ) d t . We also define the auxiliary discrete adjoint solution ( v h ∗ , k ( m p ) , q T , h ∗ , k ( m p ) , q h ∗ , k ( m p ) ) (v_h^{*,k}(m_p),q_{T,h}^{*,k}(m_p),q_h^{*,k}(m_p)) ( v h ∗ , k ( m p ) , q T , h ∗ , k ( m p ) , q h ∗ , k ( m p )) by (57)–(58) driven by the auxiliary state.
For m p , h ∈ M p , a d h m_{p,h} \in\mathcal{M}_{p,ad}^h m p , h ∈ M p , a d h , let q h ∗ ( m p , h ) q_h^*(m_{p,h}) q h ∗ ( m p , h ) denote the piecewise constant-in-time function equal to q h ∗ , k ( m p , h ) q_h^{*,k}(m_{p,h}) q h ∗ , k ( m p , h ) on I k + 1 I_{k+1} I k + 1 . The L t 2 L x 2 L_t^2L_x^2 L t 2 L x 2 -gradient of j h j_h j h in M p , h M_{p,h} M p , h is
j h ′ ( m p , h ) = γ p m p , h + Π h M q h ∗ ( m p , h ) , (84) j_h'(m_{p,h})=\gamma_p m_{p,h}+\Pi_h^M q_h^*(m_{p,h}), \tag*{(84)} j h ′ ( m p , h ) = γ p m p , h + Π h M q h ∗ ( m p , h ) , (84)
that is, j h ′ ( m p , h ) [ η h ] = ( γ p m p , h + q h ∗ ( m p , h ) , η h ) L t 2 L x 2 j_h'(m_{p,h})[\eta_h]=(\gamma_p m_{p,h}+q_h^*(m_{p,h}),\eta_h)_{L_t^2L_x^2} j h ′ ( m p , h ) [ η h ] = ( γ p m p , h + q h ∗ ( m p , h ) , η h ) L t 2 L x 2 for all η h ∈ M p , h \eta_h \in M_{p,h} η h ∈ M p , h .
Sensitivity of the discrete state and adjoint to the control
Let Φ f w d \Phi_{\mathrm{fwd}} Φ fwd denote the map m p ↦ ( u h , p T , h , p h ) m_p \mapsto(u_h,p_{T,h},p_h) m p ↦ ( u h , p T , h , p h ) defined by (50), and Φ a d j \Phi_{\mathrm{adj}} Φ adj the adjoint map (57). The following lemma quantifies the sensitivity of Φ f w d \Phi_{\mathrm{fwd}} Φ fwd and Φ a d j ∘ Φ f w d \Phi_{\mathrm{adj}}\circ\Phi_{\mathrm{fwd}} Φ adj ∘ Φ fwd to the control.
Lemma 5.1 (Control-to-state and control-to-adjoint sensitivity). Let m p , m ~ p ∈ M p , a d m_p,\widetilde{m}_p \in\mathcal{M}_{p,ad} m p , m p ∈ M p , a d be two admissible controls (continuous or discrete; M p , a d h ⊂ M p , a d \mathcal{M}_{p,ad}^h \subset\mathcal{M}_{p,ad} M p , a d h ⊂ M p , a d is the special case of piecewise-constant controls), and let the corresponding auxiliary fully discrete state and adjoint solutions be Φ f w d ( m p ) \Phi_{\mathrm{fwd}}(m_p) Φ fwd ( m p ) , Φ f w d ( m ~ p ) \Phi_{\mathrm{fwd}}(\widetilde{m}_p) Φ fwd ( m p ) and Φ a d j ∘ Φ f w d \Phi_{\mathrm{adj}}\circ\Phi_{\mathrm{fwd}} Φ adj ∘ Φ fwd applied to each. Define the control difference m p , d : = m p − m ~ p m_{p,d}:=m_p-\widetilde{m}_p m p , d := m p − m p and its cell average
m ‾ p , d k + 1 : = 1 Δ t ∫ t k t k + 1 m p , d ( t ) d t , \overline{m}_{p,d}^{k+1}:=\frac{1}{\Delta t}\int_{t_k}^{t_{k+1}}m_{p,d}(t)\,d t, m p , d k + 1 := Δ t 1 ∫ t k t k + 1 m p , d ( t ) d t , which is equal to the value of m p , d m_{p,d} m p , d on I k + 1 I_{k+1} I k + 1 when the controls are already piecewise constant in time. By Jensen’s inequality,
Δ t ∑ k = 0 N − 1 ∥ m ‾ p , d k + 1 ∥ 0 2 ≤ ∥ m p , d ∥ L t 2 L x 2 2 . (85) \Delta t\sum_{k=0}^{N-1}\|\overline{m}_{p,d}^{k+1}\|_0^2\leq\|m_{p,d}\|_{L_t^2L_x^2}^2. \tag*{(85)} Δ t k = 0 ∑ N − 1 ∥ m p , d k + 1 ∥ 0 2 ≤ ∥ m p , d ∥ L t 2 L x 2 2 . (85) Let ( u d , h k , p T , d , h k , p d , h k ) : = Φ f w d ( m p ) − Φ f w d ( m ~ p ) (u_{d,h}^k,p_{T,d,h}^k,p_{d,h}^k):=\Phi_{\mathrm{fwd}}(m_p)-\Phi_{\mathrm{fwd}}(\widetilde{m}_p) ( u d , h k , p T , d , h k , p d , h k ) := Φ fwd ( m p ) − Φ fwd ( m p ) and let ( v d , h ∗ , k , q T , d , h ∗ , k , q d , h ∗ , k ) (v_{d,h}^{*,k},q_{T,d,h}^{*,k},q_{d,h}^{*,k}) ( v d , h ∗ , k , q T , d , h ∗ , k , q d , h ∗ , k ) be the corresponding difference of the adjoint solutions Φ a d j ∘ Φ f w d \Phi_{\mathrm{adj}}\circ\Phi_{\mathrm{fwd}} Φ adj ∘ Φ fwd , with homogeneous terminal data. Then
max 1 ≤ k ≤ N ( ∥ u d , h k ∥ V + ∥ p T , d , h k ∥ 0 + ∥ p d , h k ∥ 0 ) ≲ ( Δ t ∑ k = 1 N ∥ m ‾ p , d k ∥ 0 2 ) 1 / 2 , (86) \max_{1\leq k\leq N}\left(\|u_{d,h}^k\|_V+\|p_{T,d,h}^k\|_0+\|p_{d,h}^k\|_0\right)\lesssim\left(\Delta t\sum_{k=1}^{N}\|\overline{m}_{p,d}^k\|_0^2\right)^{1/2}, \tag*{(86)} 1 ≤ k ≤ N max ( ∥ u d , h k ∥ V + ∥ p T , d , h k ∥ 0 + ∥ p d , h k ∥ 0 ) ≲ ( Δ t k = 1 ∑ N ∥ m p , d k ∥ 0 2 ) 1/2 , (86) Δ t ∑ k = 1 N [ a u ( u d , h k , u d , h k ) + ∥ p T , d , h k ∥ 0 2 + ∥ p d , h k ∥ 0 2 ] ≲ C T Δ t ∑ k = 1 N ∥ m ‾ p , d k ∥ 0 2 , (87) \Delta t\sum_{k=1}^{N}\left[a_{u}(u_{d,h}^k,u_{d,h}^k)+\|p_{T,d,h}^k\|_0^2+\|p_{d,h}^k\|_0^2\right]\lesssim C_T\Delta t\sum_{k=1}^{N}\|\overline{m}_{p,d}^k\|_0^2, \tag*{(87)} Δ t k = 1 ∑ N [ a u ( u d , h k , u d , h k ) + ∥ p T , d , h k ∥ 0 2 + ∥ p d , h k ∥ 0 2 ] ≲ C T Δ t k = 1 ∑ N ∥ m p , d k ∥ 0 2 , (87) Δ t ∑ k = 0 N − 1 ∥ q d , h ∗ , k ∥ 0 2 ≲ C T Δ t ∑ k = 1 N ∥ m ‾ p , d k ∥ 0 2 , (88) \Delta t\sum_{k=0}^{N-1}\|q_{d,h}^{*,k}\|_0^2\lesssim C_T\Delta t\sum_{k=1}^{N}\|\overline{m}_{p,d}^k\|_0^2, \tag*{(88)} Δ t k = 0 ∑ N − 1 ∥ q d , h ∗ , k ∥ 0 2 ≲ C T Δ t k = 1 ∑ N ∥ m p , d k ∥ 0 2 , (88) with constants independent of h h h , Δ t \Delta t Δ t , but may depend on c e f f , h − 1 c_{\mathrm{eff},h}^{-1} c eff , h − 1 .
Proof. Subtracting the auxiliary state systems for m p m_p m p and m ~ p \widetilde m_p m p , which have the same initial data and body force, the state difference satisfies
a u ( d t u d , h k + 1 , v h ) + b ( d t p T , d , h k + 1 , v h ) = 0 , (89) a_u(d_tu_{d,h}^{k+1},v_h)+b(d_tp_{T,d,h}^{k+1},v_h)=0, \tag*{(89)} a u ( d t u d , h k + 1 , v h ) + b ( d t p T , d , h k + 1 , v h ) = 0 , (89)
b ( q T , h , d t u d , h k + 1 ) − λ − 1 ( d t p T , d , h k + 1 , q T , h ) Ω + λ − 1 α ( d t p d , h k + 1 , q T , h ) Ω = 0 , (90) b(q_{T,h},d_tu_{d,h}^{k+1})-\lambda^{-1}(d_tp_{T,d,h}^{k+1},q_{T,h})_\Omega+\lambda^{-1}\alpha(d_tp_{d,h}^{k+1},q_{T,h})_\Omega=0, \tag*{(90)} b ( q T , h , d t u d , h k + 1 ) − λ − 1 ( d t p T , d , h k + 1 , q T , h ) Ω + λ − 1 α ( d t p d , h k + 1 , q T , h ) Ω = 0 , (90)
− λ − 1 α ( d t p T , d , h k + 1 , q h ) Ω + s 1 ( d t p d , h k + 1 , q h ) Ω + ( κ p ∇ p d , h k + 1 , ∇ q h ) Ω = ( m ‾ p , d k + 1 , q h ) Ω , (91) -\lambda^{-1}\alpha(d_tp_{T,d,h}^{k+1},q_h)_\Omega+s_1(d_tp_{d,h}^{k+1},q_h)_\Omega+(\kappa_p\nabla p_{d,h}^{k+1},\nabla q_h)_\Omega=(\overline m_{p,d}^{k+1},q_h)_\Omega, \tag*{(91)} − λ − 1 α ( d t p T , d , h k + 1 , q h ) Ω + s 1 ( d t p d , h k + 1 , q h ) Ω + ( κ p ∇ p d , h k + 1 , ∇ q h ) Ω = ( m p , d k + 1 , q h ) Ω , (91)
with zero initial data. This is (51) with F u k + 1 = 0 F_u^{k+1}=0 F u k + 1 = 0 and F r k + 1 = m ‾ p , d k + 1 F_r^{k+1}=\overline m_{p,d}^{k+1} F r k + 1 = m p , d k + 1 , so Lemma 4.3 gives (86) . Since Δ t ∑ k = 1 N a k ≤ T max 1 ≤ k ≤ N a k \Delta t\sum_{k=1}^{N}a_k\le T\max_{1\le k\le N}a_k Δ t ∑ k = 1 N a k ≤ T max 1 ≤ k ≤ N a k for nonnegative a k a_k a k , squaring (86) gives (87) . The adjoint difference solves (57) with interior data F u k + 1 = u d , h k + 1 F_u^{k+1}=u_{d,h}^{k+1} F u k + 1 = u d , h k + 1 , F p k + 1 = p d , h k + 1 F_p^{k+1}=p_{d,h}^{k+1} F p k + 1 = p d , h k + 1 and homogeneous terminal data. Lemma 4.9, together with ∥ u d , h k ∥ 0 ≤ C ∥ u d , h k ∥ V \|u_{d,h}^{k}\|_0\le C\|u_{d,h}^{k}\|_V ∥ u d , h k ∥ 0 ≤ C ∥ u d , h k ∥ V and (87) , bounds max 0 ≤ k ≤ N − 1 ∥ q d , h ∗ , k ∥ 0 \max_{0\le k\le N-1}\|q_{d,h}^{*,k}\|_0 max 0 ≤ k ≤ N − 1 ∥ q d , h ∗ , k ∥ 0 by C ( Δ t ∑ k ∥ m ‾ p , d k ∥ 0 2 ) 1 / 2 C(\Delta t\sum_k\|\overline m_{p,d}^{k}\|_0^2)^{1/2} C ( Δ t ∑ k ∥ m p , d k ∥ 0 2 ) 1/2 ; the same summation inequality yields (88) . □
Notation and projection operators
We use the cell averages (60) . The hat denotes optimal quantities: u ˉ ^ j \widehat{\bar u}^{j} u ˉ j is the continuous optimal state associated with the optimal control m ‾ p \overline m_p m p , and u ˉ ^ h j \widehat{\bar u}_h^{j} u ˉ h j is the discrete optimal state at time level j j j . Thus u ˉ ^ j \widehat{\bar u}^{j} u ˉ j , p ˉ ^ j \widehat{\bar p}^{j} p ˉ j , and q ˉ ^ ∗ , k + 1 \widehat{\bar q}^{*,k+1} q ˉ ∗ , k + 1 are cell averages of continuous optimal quantities, not discrete quantities. The error estimate in Theorem 5.14 is stated in terms of these cell averages.
We use the following standard interpolation and projection operators.
Definition 5.2. Define the following interpolation operators.
Π h u : H 2 ( Ω ) ∩ V → V h \Pi_h^u:H^2(\Omega)\cap V\to V_h Π h u : H 2 ( Ω ) ∩ V → V h is the standard Scott–Zhang (or Clément-type) quasi-interpolation operator onto the Taylor–Hood velocity space, satisfying ∥ v − Π h u v ∥ 1 ≤ C h ∣ v ∣ 2 \|v-\Pi_h^uv\|_1\le Ch|v|_2 ∥ v − Π h u v ∥ 1 ≤ C h ∣ v ∣ 2 for v ∈ H 2 ( Ω ) ∩ V v\in H^2(\Omega)\cap V v ∈ H 2 ( Ω ) ∩ V .
Π h p T : L 2 ( Ω ) → Q T , h \Pi_h^{p_T}:L^2(\Omega)\to Q_{T,h} Π h p T : L 2 ( Ω ) → Q T , h is the L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -orthogonal projection onto the pressure space Q T , h Q_{T,h} Q T , h , satisfying ∥ p T − Π h p T p T ∥ 0 ≤ C h ∣ p T ∣ 1 \|p_T-\Pi_h^{p_T}p_T\|_0\le Ch|p_T|_1 ∥ p T − Π h p T p T ∥ 0 ≤ C h ∣ p T ∣ 1 for p T ∈ H 1 ( Ω ) p_T\in H^1(\Omega) p T ∈ H 1 ( Ω ) .
P h : L 2 ( Ω ) → Q h P_h:L^2(\Omega)\to Q_h P h : L 2 ( Ω ) → Q h is the L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -orthogonal projection onto the diffusion space, defined by
( P h z , q h ) L 2 = ( z , q h ) L 2 ∀ q h ∈ Q h . (P_hz,q_h)_{L^2}=(z,q_h)_{L^2}\quad\forall q_h\in Q_h. ( P h z , q h ) L 2 = ( z , q h ) L 2 ∀ q h ∈ Q h .
Lemma 5.3. The following properties hold:
(a) Orthogonality: ( ( I − P h ) z , q h ) L 2 = 0 ((I-P_h)z,q_h)_{L^2}=0 (( I − P h ) z , q h ) L 2 = 0 for all q h ∈ Q h q_h\in Q_h q h ∈ Q h .
(b) L 2 L^2 L 2 -approximation: ∥ ( I − P h ) z ∥ 0 ≤ C h ∥ z ∥ H 1 \|(I-P_h)z\|_0\le Ch\|z\|_{H^1} ∥ ( I − P h ) z ∥ 0 ≤ C h ∥ z ∥ H 1 for z ∈ H 0 1 ( Ω ) z\in H_0^1(\Omega) z ∈ H 0 1 ( Ω ) .
(c) H 1 H^1 H 1 -approximation: ∥ ( I − P h ) z ∥ H 1 ≤ C h ∥ z ∥ H 2 \|(I-P_h)z\|_{H^1}\le Ch\|z\|_{H^2} ∥ ( I − P h ) z ∥ H 1 ≤ C h ∥ z ∥ H 2 for z ∈ H 2 ( Ω ) ∩ H 0 1 ( Ω ) z\in H^2(\Omega)\cap H_0^1(\Omega) z ∈ H 2 ( Ω ) ∩ H 0 1 ( Ω ) (see, e.g., Crouzeix–Thomée [7 ] ).
Lemma 5.4 (Fixed-control state discretization error). Let ( u , p T , p ) (u,p_T,p) ( u , p T , p ) be the continuous state associated with control m p ∈ M p , a d m_p\in M_{p,\mathrm{ad}} m p ∈ M p , ad , and let ( u h j , p T , h j , p h j ) (u_h^j,p_{T,h}^j,p_h^j) ( u h j , p T , h j , p h j ) be the fully discrete state (50) driven by the cell averages m ‾ p k + 1 \overline m_p^{k+1} m p k + 1 . Assume the corresponding state satisfies the regularity conditions in (83) . Then
( Δ t ∑ j = 1 N ∣ p ‾ j − p h j ∣ 0 2 ) 1 / 2 ≤ C ( h + Δ t ) , (92) \left(\Delta t\sum_{j=1}^{N}\left|\overline p^j-p_h^j\right|_0^2\right)^{1/2}\le C(h+\Delta t), \tag*{(92)} ( Δ t j = 1 ∑ N p j − p h j 0 2 ) 1/2 ≤ C ( h + Δ t ) , (92) ( Δ t ∑ j = 1 N ( ∥ u ‾ j − u h j ∥ 1 2 + ∥ p ‾ T j − p T , h j ∥ 0 2 ) ) 1 / 2 ≤ C ( h + Δ t ) , (93) \left(\Delta t\sum_{j=1}^{N}\left(\|\overline u^j-u_h^j\|_1^2+\|\overline p_T^j-p_{T,h}^j\|_0^2\right)\right)^{1/2}\le C(h+\Delta t), \tag*{(93)} ( Δ t j = 1 ∑ N ( ∥ u j − u h j ∥ 1 2 + ∥ p T j − p T , h j ∥ 0 2 ) ) 1/2 ≤ C ( h + Δ t ) , (93) where C > 0 C>0 C > 0 is independent of h h h and Δ t \Delta t Δ t but may depends on c e f f − 1 c_{\mathrm{eff}}^{-1} c eff − 1 , c e f f , h − 1 c_{\mathrm{eff},h}^{-1} c eff , h − 1 , and κ p \kappa_p κ p .
The proof is organized through a sequence of auxiliary lemmas (Lemmas 5.5–5.9). For the proof we use the continuous and discrete mechanical Schur complements introduced in Lemmas 2.5 and 4.1.
For each time slab I k + 1 = ( t k , t k + 1 ] I_{k+1}=(t_k,t_{k+1}] I k + 1 = ( t k , t k + 1 ] , set
f ˉ k + 1 : = 1 Δ t ∫ I k + 1 f ( t ) d t . \bar{f}^{k+1}:=\frac{1}{\Delta t}\int_{I_{k+1}}f(t)\,\mathrm{d}t. f ˉ k + 1 := Δ t 1 ∫ I k + 1 f ( t ) d t .
Let ( u f , h k + 1 , p T , f , h k + 1 ) ∈ V h × Q T , h (u_{f,h}^{k+1},p_{T,f,h}^{k+1})\in V_h\times Q_{T,h} ( u f , h k + 1 , p T , f , h k + 1 ) ∈ V h × Q T , h be the discrete mechanical load lift defined by
a u ( u f , h k + 1 , v h ) + b ( p T , f , h k + 1 , v h ) = ( f ˉ k + 1 , v h ) Ω , a_u(u_{f,h}^{k+1},v_h)+b(p_{T,f,h}^{k+1},v_h)=(\bar{f}^{k+1},v_h)_\Omega, a u ( u f , h k + 1 , v h ) + b ( p T , f , h k + 1 , v h ) = ( f ˉ k + 1 , v h ) Ω ,
b ( q T , h , u f , h k + 1 ) − λ − 1 ( p T , f , h k + 1 , q T , h ) Ω = 0 (94) b(q_{T,h},u_{f,h}^{k+1})-\lambda^{-1}(p_{T,f,h}^{k+1},q_{T,h})_\Omega=0 \tag*{(94)} b ( q T , h , u f , h k + 1 ) − λ − 1 ( p T , f , h k + 1 , q T , h ) Ω = 0 (94)
for all ( v h , q T , h ) ∈ V h × Q T , h (v_h,q_{T,h})\in V_h\times Q_{T,h} ( v h , q T , h ) ∈ V h × Q T , h . The continuous lift ( u ˉ f k + 1 , p ˉ T , f k + 1 ) ∈ V × Q T (\bar{u}_f^{k+1},\bar{p}_{T,f}^{k+1})\in V\times Q_T ( u ˉ f k + 1 , p ˉ T , f k + 1 ) ∈ V × Q T solves the same problem in the continuous spaces. By linearity of (31) , it is the time average of the load lift already defined in § 2.4 :
( u ˉ f k + 1 , p ˉ T , f k + 1 ) = 1 Δ t ∫ I k + 1 ( u f ( t ) , p T , f ( t ) ) d t . (\bar{u}_f^{k+1},\bar{p}_{T,f}^{k+1})=\frac{1}{\Delta t}\int_{I_{k+1}}(u_f(t),p_{T,f}(t))\,\mathrm{d}t. ( u ˉ f k + 1 , p ˉ T , f k + 1 ) = Δ t 1 ∫ I k + 1 ( u f ( t ) , p T , f ( t )) d t .
Likewise, let ( u f , h ( t ) , p T , f , h ( t ) ) (u_{f,h}(t),p_{T,f,h}(t)) ( u f , h ( t ) , p T , f , h ( t )) solve (72) with load f ( t ) f(t) f ( t ) in place of f ˉ k + 1 \bar{f}^{k+1} f ˉ k + 1 ; its time average is ( u f , h k + 1 , p T , f , h k + 1 ) (u_{f,h}^{k+1},p_{T,f,h}^{k+1}) ( u f , h k + 1 , p T , f , h k + 1 ) . Define the corresponding load functionals by
G f k + 1 ( q ) : = λ − 1 α ( p T , f k + 1 , q ) Ω , q ∈ Q , G_f^{k+1}(q):=\lambda^{-1}\alpha(p_{T,f}^{k+1},q)_\Omega,\qquad q\in Q, G f k + 1 ( q ) := λ − 1 α ( p T , f k + 1 , q ) Ω , q ∈ Q ,
G f , h k + 1 ( q h ) : = λ − 1 α ( p T , f , h k + 1 , q h ) Ω , q h ∈ Q h . (95) G_{f,h}^{k+1}(q_h):=\lambda^{-1}\alpha(p_{T,f,h}^{k+1},q_h)_\Omega,\qquad q_h\in Q_h. \tag*{(95)} G f , h k + 1 ( q h ) := λ − 1 α ( p T , f , h k + 1 , q h ) Ω , q h ∈ Q h . (95)
Lemma 5.5. The fully discrete scheme (50) is equivalent, after eliminating ( u h k + 1 , p T , h k + 1 ) (u_h^{k+1},p_{T,h}^{k+1}) ( u h k + 1 , p T , h k + 1 ) via Lemma 4.1, to the reduced scalar scheme: find p h k + 1 ∈ Q h p_h^{k+1}\in Q_h p h k + 1 ∈ Q h such that
m e f f , h ( d t p h k + 1 , q h ) + ( κ p ∇ p h k + 1 , ∇ q h ) Ω = ( m ˉ p k + 1 , q h ) Ω + G f , h k + 1 ( q h ) ∀ q h ∈ Q h , (96) m_{\mathrm{eff},h}(d_t p_h^{k+1},q_h)+(\kappa_p\nabla p_h^{k+1},\nabla q_h)_\Omega=(\bar{m}_p^{k+1},q_h)_\Omega+G_{f,h}^{k+1}(q_h)\qquad\forall q_h\in Q_h, \tag*{(96)} m eff , h ( d t p h k + 1 , q h ) + ( κ p ∇ p h k + 1 , ∇ q h ) Ω = ( m ˉ p k + 1 , q h ) Ω + G f , h k + 1 ( q h ) ∀ q h ∈ Q h , (96) where
m ˉ p k + 1 : = 1 Δ t ∫ I k + 1 m p ( t ) d t . \bar{m}_p^{k+1}:=\frac{1}{\Delta t}\int_{I_{k+1}}m_p(t)\,\mathrm{d}t. m ˉ p k + 1 := Δ t 1 ∫ I k + 1 m p ( t ) d t .
Proof. The first two equations of (50) are the mechanical saddle-point problem for ( d t u h k + 1 , d t p T , h k + 1 ) (d_tu_h^{k+1},d_tp_{T,h}^{k+1}) ( d t u h k + 1 , d t p T , h k + 1 ) , driven by d t p h k + 1 d_tp_h^{k+1} d t p h k + 1 and f ˉ k + 1 \bar{f}^{k+1} f ˉ k + 1 . By linearity and (94) ,
d t u h k + 1 = U h ( d t p h k + 1 ) + u f , h k + 1 , d t p T , h k + 1 = P T , h ( d t p h k + 1 ) + p T , f , h k + 1 . (97) d_tu_h^{k+1}=\mathcal{U}_h(d_tp_h^{k+1})+u_{f,h}^{k+1},\qquad d_tp_{T,h}^{k+1}=\mathcal{P}_{T,h}(d_tp_h^{k+1})+p_{T,f,h}^{k+1}. \tag*{(97)} d t u h k + 1 = U h ( d t p h k + 1 ) + u f , h k + 1 , d t p T , h k + 1 = P T , h ( d t p h k + 1 ) + p T , f , h k + 1 . (97)
Substitution into (50c) and the definitions of m e f f , h m_{\mathrm{eff},h} m eff , h and G f , h k + 1 G_{f,h}^{k+1} G f , h k + 1 give (96) . Conversely, given a solution of the reduced scheme and the initial mechanical variables, (75) reconstructs the mechanical increments and hence the full discrete state. The load term depends only on f ˉ k + 1 \bar{f}^{k+1} f ˉ k + 1 . Its continuous counterpart is G f k + 1 G_f^{k+1} G f k + 1 , and their difference gives the load residual R F k + 1 R_F^{k+1} R F k + 1 in the error equation below.
Lemma 5.6. Assume λ ≥ λ 0 > 0 \lambda\ge\lambda_0>0 λ ≥ λ 0 > 0 and the Stokes-type regularity assumption (2) of Appendix A. For z ∈ H 0 1 ( Ω ) z\in H_0^1(\Omega) z ∈ H 0 1 ( Ω ) and q h ∈ Q h q_h\in Q_h q h ∈ Q h ,
∣ m e f f ( z , q h ) − m e f f , h ( P h z , q h ) ∣ ≤ C h 2 ∥ z ∥ H 1 ∥ q h ∥ 0 , (98) \left|m_{\mathrm{eff}}(z,q_h)-m_{\mathrm{eff},h}(P_hz,q_h)\right|\le Ch^2\lVert z\rVert_{H^1}\lVert q_h\rVert_0, \tag*{(98)} ∣ m eff ( z , q h ) − m eff , h ( P h z , q h ) ∣ ≤ C h 2 ∥ z ∥ H 1 ∥ q h ∥ 0 , (98) with C C C independent of h h h , Δ t \Delta t Δ t , s 0 s_0 s 0 , and uniform for λ ≥ λ 0 \lambda\ge\lambda_0 λ ≥ λ 0 .
Proof. By (29) , (58) and property (a) of P h P_h P h , the storage part s 1 ( ( I − P h ) z , q h ) Ω s_1((I-P_h)z,q_h)_\Omega s 1 (( I − P h ) z , q h ) Ω vanishes, so
m e f f ( z , q h ) − m e f f , h ( P h z , q h ) = − λ − 1 α ( ( P T − P T , h ) z + P T , h ( I − P h ) z , q h ) Ω . m_{\mathrm{eff}}(z,q_h)-m_{\mathrm{eff},h}(P_hz,q_h)=-\lambda^{-1}\alpha((P_T-P_{T,h})z+P_{T,h}(I-P_h)z,q_h)_\Omega. m eff ( z , q h ) − m eff , h ( P h z , q h ) = − λ − 1 α (( P T − P T , h ) z + P T , h ( I − P h ) z , q h ) Ω .
By Lemma 4.1 and property (b) of P h P_h P h , ∥ P T , h ( I − P h ) z ∥ 0 ≤ C ∥ ( I − P h ) z ∥ 0 ≤ C h ∥ z ∥ H 1 \|P_{T,h}(I-P_h)z\|_0 \le C\|(I-P_h)z\|_0 \le Ch\|z\|_{H^1} ∥ P T , h ( I − P h ) z ∥ 0 ≤ C ∥ ( I − P h ) z ∥ 0 ≤ C h ∥ z ∥ H 1 . The pair ( U h z , P T , h z ) (U_hz,P_{T,h}z) ( U h z , P T , h z ) is the Taylor–Hood Galerkin approximation of ( U z , P T z ) (U_z,P_Tz) ( U z , P T z ) , so Babuška–Brezzi quasi-optimality, the Taylor–Hood approximation properties (Definition 5.2), and assumption (2) with F = 0 F=0 F = 0 , ζ = − λ − 1 α z ∈ H 0 1 ( Ω ) \zeta=-\lambda^{-1}\alpha z\in H_0^1(\Omega) ζ = − λ − 1 α z ∈ H 0 1 ( Ω ) give
∥ ( P T − P T , h ) z ∥ 0 ≤ C h ( ∥ U z ∥ H 2 + ∥ P T z ∥ H 1 ) ≤ C h ∥ z ∥ H 1 . \|(P_T-P_{T,h})z\|_0 \le Ch(\|U_z\|_{H^2}+\|P_Tz\|_{H^1}) \le Ch\|z\|_{H^1}. ∥ ( P T − P T , h ) z ∥ 0 ≤ C h ( ∥ U z ∥ H 2 + ∥ P T z ∥ H 1 ) ≤ C h ∥ z ∥ H 1 .
Using λ − 1 ≤ λ 0 − 1 \lambda^{-1}\le\lambda_0^{-1} λ − 1 ≤ λ 0 − 1 completes the proof. □
Lemma 5.7 (Stability of the reduced backward dual). Let φ n ∈ Q h \varphi^n\in Q_h φ n ∈ Q h , n = 1 , … , N n=1,\ldots,N n = 1 , … , N , with φ N + 1 : = 0 \varphi^{N+1}:=0 φ N + 1 := 0 , solve
m e f f , h ( φ n − φ n + 1 Δ t , q h ) + ( κ p ∇ φ n , ∇ q h ) Ω = ( ξ n , q h ) L 2 ∀ q h ∈ Q h , (99) m_{\mathrm{eff},h}\left(\frac{\varphi^n-\varphi^{n+1}}{\Delta t},q_h\right)+(\kappa_p\nabla\varphi^n,\nabla q_h)_\Omega=(\xi^n,q_h)_{L^2}\quad\forall q_h\in Q_h, \tag*{(99)} m eff , h ( Δ t φ n − φ n + 1 , q h ) + ( κ p ∇ φ n , ∇ q h ) Ω = ( ξ n , q h ) L 2 ∀ q h ∈ Q h , (99) for a given sequence ξ n ∈ Q h \xi^n\in Q_h ξ n ∈ Q h . Then
( Δ t ∑ n = 1 N ∥ φ n ∥ 0 2 ) 1 / 2 ≤ C ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 , (100) \left(\Delta t\sum_{n=1}^{N}\|\varphi^n\|_0^2\right)^{1/2}\le C\left(\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2\right)^{1/2}, \tag*{(100)} ( Δ t n = 1 ∑ N ∥ φ n ∥ 0 2 ) 1/2 ≤ C ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2 , (100) ( Δ t ∑ n = 1 N ∥ ∇ φ n ∥ 0 2 ) 1 / 2 ≤ C ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 , (101) \left(\Delta t\sum_{n=1}^{N}\|\nabla\varphi^n\|_0^2\right)^{1/2}\le C\left(\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2\right)^{1/2}, \tag*{(101)} ( Δ t n = 1 ∑ N ∥∇ φ n ∥ 0 2 ) 1/2 ≤ C ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2 , (101) ( Δ t ∑ n = 1 N ∥ φ n − φ n + 1 Δ t ∥ 0 2 ) 1 / 2 ≤ C ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 , (102) \left(\Delta t\sum_{n=1}^{N}\left\|\frac{\varphi^n-\varphi^{n+1}}{\Delta t}\right\|_0^2\right)^{1/2}\le C\left(\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2\right)^{1/2}, \tag*{(102)} ( Δ t n = 1 ∑ N Δ t φ n − φ n + 1 0 2 ) 1/2 ≤ C ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2 , (102) max 1 ≤ n ≤ N ∥ φ n ∥ 0 ≤ C ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 , (103) \max_{1\le n\le N}\|\varphi^n\|_0\le C\left(\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2\right)^{1/2}, \tag*{(103)} 1 ≤ n ≤ N max ∥ φ n ∥ 0 ≤ C ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2 , (103) with C C C independent of h h h and Δ t \Delta t Δ t . It depends on the coercivity of m e f f , h m_{\mathrm{eff},h} m eff , h and may therefore depend on λ \lambda λ when s 0 = 0 s_0=0 s 0 = 0 . For s 0 ≥ s ∗ > 0 s_0\ge s_\ast>0 s 0 ≥ s ∗ > 0 , the estimates are λ \lambda λ -uniform under the stated diffusion assumptions.
Proof. Set δ τ φ n : = ( φ n − φ n + 1 ) / Δ t \delta_\tau\varphi^n:=(\varphi^n-\varphi^{n+1})/\Delta t δ τ φ n := ( φ n − φ n + 1 ) /Δ t . For (78) and (79) test (77) with q h = φ n q_h=\varphi^n q h = φ n . Since m e f f , h m_{\mathrm{eff},h} m eff , h is symmetric and coercive on Q h Q_h Q h , it induces the norm ∥ φ ∥ m e f f , h : = m e f f , h ( φ , φ ) 1 / 2 \|\varphi\|_{m_{\mathrm{eff},h}}:=m_{\mathrm{eff},h}(\varphi,\varphi)^{1/2} ∥ φ ∥ m eff , h := m eff , h ( φ , φ ) 1/2 , and
m e f f , h ( δ τ φ n , φ n ) = 1 2 Δ t ( ∥ φ n ∥ m e f f , h 2 − ∥ φ n + 1 ∥ m e f f , h 2 + ∥ φ n − φ n + 1 ∥ m e f f , h 2 ) ≥ 1 2 Δ t ( ∥ φ n ∥ m e f f , h 2 − ∥ φ n + 1 ∥ m e f f , h 2 ) . m_{\mathrm{eff},h}(\delta_\tau\varphi^n,\varphi^n)=\frac{1}{2\Delta t}\left(\|\varphi^n\|_{m_{\mathrm{eff},h}}^2-\|\varphi^{n+1}\|_{m_{\mathrm{eff},h}}^2+\|\varphi^n-\varphi^{n+1}\|_{m_{\mathrm{eff},h}}^2\right)\ge\frac{1}{2\Delta t}\left(\|\varphi^n\|_{m_{\mathrm{eff},h}}^2-\|\varphi^{n+1}\|_{m_{\mathrm{eff},h}}^2\right). m eff , h ( δ τ φ n , φ n ) = 2Δ t 1 ( ∥ φ n ∥ m eff , h 2 − ∥ φ n + 1 ∥ m eff , h 2 + ∥ φ n − φ n + 1 ∥ m eff , h 2 ) ≥ 2Δ t 1 ( ∥ φ n ∥ m eff , h 2 − ∥ φ n + 1 ∥ m eff , h 2 ) .
Using ( κ p ∇ φ n , ∇ φ n ) Ω = ∥ φ n ∥ Q 2 (\kappa_p\nabla\varphi^n,\nabla\varphi^n)_\Omega=\|\varphi^n\|_{Q}^2 ( κ p ∇ φ n , ∇ φ n ) Ω = ∥ φ n ∥ Q 2 , Poincaré’s inequality on Q h Q_h Q h , and Young’s inequality, we have
( ξ n , φ n ) ≤ C ∥ ξ n ∥ 0 ∥ φ n ∥ H 1 ≤ ε ∥ φ n ∥ H 1 2 + C ε ∥ ξ n ∥ 0 2 . (\xi^n,\varphi^n)\le C\|\xi^n\|_0\|\varphi^n\|_{H^1}\le\varepsilon\|\varphi^n\|_{H^1}^2+C_\varepsilon\|\xi^n\|_0^2. ( ξ n , φ n ) ≤ C ∥ ξ n ∥ 0 ∥ φ n ∥ H 1 ≤ ε ∥ φ n ∥ H 1 2 + C ε ∥ ξ n ∥ 0 2 .
Multiplying by Δ t \Delta t Δ t , summing n = 1 , … , N n=1,\ldots,N n = 1 , … , N , using φ N + 1 = 0 \varphi^{N+1}=0 φ N + 1 = 0 , and choosing ε \varepsilon ε small:
1 2 ∥ φ 1 ∥ m e f f , h 2 + C Δ t ∑ n = 1 N ∥ φ n ∥ H 1 2 ≤ C Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 . \frac{1}{2}\|\varphi^1\|_{m_{\mathrm{eff},h}}^2+C\Delta t\sum_{n=1}^{N}\|\varphi^n\|_{H^1}^2\le C\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2. 2 1 ∥ φ 1 ∥ m eff , h 2 + C Δ t n = 1 ∑ N ∥ φ n ∥ H 1 2 ≤ C Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 .
The m e f f , h m_{\mathrm{eff},h} m eff , h -coercivity and Poincaré inequality give (78)–(79).
For (81), for arbitrary 1 ≤ m ≤ N 1\le m\le N 1 ≤ m ≤ N , sum the energy inequality from n = m n=m n = m to N N N . Using φ N + 1 = 0 \varphi^{N+1}=0 φ N + 1 = 0 :
1 2 ∥ φ m ∥ m e f f , h 2 + C Δ t ∑ n = m N ∥ φ n ∥ H 1 2 ≤ C Δ t ∑ n = m N ∥ ξ n ∥ 0 2 ≤ C Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 . \frac{1}{2}\|\varphi^m\|_{m_{\mathrm{eff},h}}^2+C\Delta t\sum_{n=m}^{N}\|\varphi^n\|_{H^1}^2\le C\Delta t\sum_{n=m}^{N}\|\xi^n\|_0^2\le C\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2. 2 1 ∥ φ m ∥ m eff , h 2 + C Δ t n = m ∑ N ∥ φ n ∥ H 1 2 ≤ C Δ t n = m ∑ N ∥ ξ n ∥ 0 2 ≤ C Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 .
Taking the maximum over m = 1 , … , N m=1,\ldots,N m = 1 , … , N and using m e f f , h m_{\mathrm{eff},h} m eff , h -coercivity:
max 1 ≤ n ≤ N ∥ φ n ∥ 0 2 ≤ C max 1 ≤ m ≤ N ∥ φ m ∥ m e f f , h 2 ≤ C Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 , \max_{1\le n\le N}\|\varphi^n\|_0^2\le C\max_{1\le m\le N}\|\varphi^m\|_{m_{\mathrm{eff},h}}^2\le C\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2, 1 ≤ n ≤ N max ∥ φ n ∥ 0 2 ≤ C 1 ≤ m ≤ N max ∥ φ m ∥ m eff , h 2 ≤ C Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ,
giving (81).
For (80), test (77) with q h = δ τ φ n q_h=\delta_\tau\varphi^n q h = δ τ φ n . By symmetry of ( κ p ∇ ⋅ , ∇ ⋅ ) Ω (\kappa_p\nabla\cdot,\nabla\cdot)_\Omega ( κ p ∇ ⋅ , ∇ ⋅ ) Ω :
( κ p ∇ φ n , ∇ δ τ φ n ) Ω = 1 2 Δ t ( ∥ φ n ∥ Q 2 − ∥ φ n + 1 ∥ Q 2 + ∥ φ n − φ n + 1 ∥ Q 2 ) ≥ 1 2 Δ t ( ∥ φ n ∥ Q 2 − ∥ φ n + 1 ∥ Q 2 ) . (\kappa_p\nabla\varphi^n,\nabla\delta_\tau\varphi^n)_\Omega=\frac{1}{2\Delta t}\left(\|\varphi^n\|_Q^2-\|\varphi^{n+1}\|_Q^2+\|\varphi^n-\varphi^{n+1}\|_Q^2\right)\ge\frac{1}{2\Delta t}\left(\|\varphi^n\|_Q^2-\|\varphi^{n+1}\|_Q^2\right). ( κ p ∇ φ n , ∇ δ τ φ n ) Ω = 2Δ t 1 ( ∥ φ n ∥ Q 2 − ∥ φ n + 1 ∥ Q 2 + ∥ φ n − φ n + 1 ∥ Q 2 ) ≥ 2Δ t 1 ( ∥ φ n ∥ Q 2 − ∥ φ n + 1 ∥ Q 2 ) .
Multiplying by Δ t \Delta t Δ t , summing n = 1 , … , N n=1,\ldots,N n = 1 , … , N , using φ N + 1 = 0 \varphi^{N+1}=0 φ N + 1 = 0 , m e f f , h m_{\mathrm{eff},h} m eff , h -coercivity, and Young’s inequality:
c Δ t ∑ n = 1 N ∥ δ τ φ n ∥ 0 2 ≤ Δ t ∑ n = 1 N ∥ ξ n ∥ 0 ∥ δ τ φ n ∥ 0 ≤ c 2 Δ t ∑ n = 1 N ∥ δ τ φ n ∥ 0 2 + C Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 , c\Delta t\sum_{n=1}^{N}\|\delta_\tau\varphi^n\|_0^2\le\Delta t\sum_{n=1}^{N}\|\xi^n\|_0\|\delta_\tau\varphi^n\|_0\le\frac{c}{2}\Delta t\sum_{n=1}^{N}\|\delta_\tau\varphi^n\|_0^2+C\Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2, c Δ t n = 1 ∑ N ∥ δ τ φ n ∥ 0 2 ≤ Δ t n = 1 ∑ N ∥ ξ n ∥ 0 ∥ δ τ φ n ∥ 0 ≤ 2 c Δ t n = 1 ∑ N ∥ δ τ φ n ∥ 0 2 + C Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ,
giving (80).
Lemma 5.8. Under the regularity (63) and the initial approximation assumption ∥ p h 0 − p ( 0 ) ∥ 0 = O ( h ) \|p_h^0-p(0)\|_0=O(h) ∥ p h 0 − p ( 0 ) ∥ 0 = O ( h ) , the discrete solution p h j p_h^j p h j from (74) satisfies
( ∑ j = 1 N ∫ I j ∥ p ( t ) − p h j ∥ 0 2 d t ) 1 / 2 ≤ C ( h + Δ t ) , (104) \left(\sum_{j=1}^{N}\int_{I_j}\|p(t)-p_h^j\|_0^2\,\mathrm{d}t\right)^{1/2}\le C(h+\Delta t), \tag*{(104)} ( j = 1 ∑ N ∫ I j ∥ p ( t ) − p h j ∥ 0 2 d t ) 1/2 ≤ C ( h + Δ t ) , (104) with C C C independent of h h h and Δ t \Delta t Δ t but depending on the lower bounds of c e f f c_{\mathrm{eff}} c eff and c e f f , h c_{\mathrm{eff},h} c eff , h .
Proof. Set p ˉ 0 : = p ( 0 ) \bar p^0:=p(0) p ˉ 0 := p ( 0 ) . For k = 1 , … , N k=1,\ldots,N k = 1 , … , N , let p ˉ k : = 1 Δ t ∫ I k p ( t ) d t \bar p^k:=\frac{1}{\Delta t}\int_{I_k}p(t)\,\mathrm{d}t p ˉ k := Δ t 1 ∫ I k p ( t ) d t and d k : = p ( t k ) − p ˉ k d^k:=p(t_k)-\bar p^k d k := p ( t k ) − p ˉ k , so d 0 = 0 d^0=0 d 0 = 0 .
For t ∈ I j t\in I_j t ∈ I j we decompose p ( t ) − p h j p(t)-p_h^j p ( t ) − p h j into
p ( t ) − p h j = ( I − P h ) p ( t ) + P h ( p ( t ) − p ˉ j ) + ( P h p ˉ j − p h j ) = : η ( t ) + μ j ( t ) + ξ j . p(t)-p_h^j=(I-P_h)p(t)+P_h(p(t)-\bar p^j)+(P_h\bar p^j-p_h^j)=:\eta(t)+\mu^j(t)+\xi^j. p ( t ) − p h j = ( I − P h ) p ( t ) + P h ( p ( t ) − p ˉ j ) + ( P h p ˉ j − p h j ) =: η ( t ) + μ j ( t ) + ξ j .
By property (b) in Lemma 5.3, ∫ I j ∥ η ( t ) ∥ 0 2 d t ≤ C h 2 ∥ p ∥ L t 2 H x 1 2 \int_{I_j}\|\eta(t)\|_0^2\,\mathrm{d}t\le Ch^2\|p\|_{L_t^2H_x^1}^2 ∫ I j ∥ η ( t ) ∥ 0 2 d t ≤ C h 2 ∥ p ∥ L t 2 H x 1 2 . We can also obtain ∑ j = 1 N ∫ I j ∥ μ j ∥ 0 2 d t ≤ C Δ t 2 ∥ p ˙ ∥ L t 2 L x 2 2 \sum_{j=1}^{N}\int_{I_j}\|\mu^j\|_0^2\,\mathrm{d}t\le C\Delta t^2\|\dot p\|_{L_t^2L_x^2}^2 ∑ j = 1 N ∫ I j ∥ μ j ∥ 0 2 d t ≤ C Δ t 2 ∥ p ˙ ∥ L t 2 L x 2 2 because ∥ p ( t ) − p ˉ j ∥ 0 ≤ ∫ I j ∥ p ˙ ∥ 0 d t \|p(t)-\bar p^j\|_0\le\int_{I_j}\|\dot p\|_0\,\mathrm{d}t ∥ p ( t ) − p ˉ j ∥ 0 ≤ ∫ I j ∥ p ˙ ∥ 0 d t . Thus, ∥ η ∥ L t 2 L x 2 2 + ∑ j = 1 N ∫ I j ∥ μ j ∥ 0 2 d t ≤ C ( h 2 + Δ t 2 ) \|\eta\|_{L_t^2L_x^2}^2+\sum_{j=1}^{N}\int_{I_j}\|\mu^j\|_0^2\,\mathrm{d}t\le C(h^2+\Delta t^2) ∥ η ∥ L t 2 L x 2 2 + ∑ j = 1 N ∫ I j ∥ μ j ∥ 0 2 d t ≤ C ( h 2 + Δ t 2 ) .
It remains to estimate Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 \Delta t\sum_{n=1}^{N}\|\xi^n\|_0^2 Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 . For this we derive an error equation and apply discrete duality.
Error equation of ξ j \xi^j ξ j : For the error equation for ξ k = P h p ˉ k − p h k ∈ Q h \xi^k=P_h\bar p^k-p_h^k\in Q_h ξ k = P h p ˉ k − p h k ∈ Q h , since d t ( P h p ˉ k + 1 ) = P h d t p ˉ k + 1 d_t(P_h\bar p^{k+1})=P_hd_t\bar p^{k+1} d t ( P h p ˉ k + 1 ) = P h d t p ˉ k + 1 and d t p ˉ k + 1 = p ˉ k + 1 − p ˉ k Δ t d_t\bar p^{k+1}=\frac{\bar p^{k+1}-\bar p^k}{\Delta t} d t p ˉ k + 1 = Δ t p ˉ k + 1 − p ˉ k , subtracting (74) from the time-averaged continuous equation gives, for all q h ∈ Q h q_h\in Q_h q h ∈ Q h :
m e f f , h ( d t ξ k + 1 , q h ) + ( κ p ∇ ξ k + 1 , ∇ q h ) Ω = − R M k + 1 ( q h ) − R S k + 1 ( q h ) − C τ k + 1 ( q h ) + R F k + 1 ( q h ) , (105) m_{\mathrm{eff},h}(d_t\xi^{k+1},q_h)+(\kappa_p\nabla\xi^{k+1},\nabla q_h)_\Omega=-R_M^{k+1}(q_h)-R_S^{k+1}(q_h)-C_\tau^{k+1}(q_h)+R_F^{k+1}(q_h), \tag*{(105)} m eff , h ( d t ξ k + 1 , q h ) + ( κ p ∇ ξ k + 1 , ∇ q h ) Ω = − R M k + 1 ( q h ) − R S k + 1 ( q h ) − C τ k + 1 ( q h ) + R F k + 1 ( q h ) , (105)
where
R M k + 1 ( q h ) : = m e f f ( p ˉ k + 1 , q h ) − m e f f , h ( P h p ˉ k + 1 , q h ) , R_M^{k+1}(q_h):=m_{\mathrm{eff}}(\bar p^{k+1},q_h)-m_{\mathrm{eff},h}(P_h\bar p^{k+1},q_h), R M k + 1 ( q h ) := m eff ( p ˉ k + 1 , q h ) − m eff , h ( P h p ˉ k + 1 , q h ) ,
R S k + 1 ( q h ) : = 1 Δ t ∫ I k + 1 ( κ p ∇ p ( t ) , ∇ q h ) Ω d t − ( κ p ∇ P h p ˉ k + 1 , ∇ q h ) Ω , R_S^{k+1}(q_h):=\frac{1}{\Delta t}\int_{I_{k+1}}(\kappa_p\nabla p(t),\nabla q_h)_\Omega\,\mathrm{d}t-(\kappa_p\nabla P_h\bar p^{k+1},\nabla q_h)_\Omega, R S k + 1 ( q h ) := Δ t 1 ∫ I k + 1 ( κ p ∇ p ( t ) , ∇ q h ) Ω d t − ( κ p ∇ P h p ˉ k + 1 , ∇ q h ) Ω ,
C τ k + 1 ( q h ) : = m e f f , h ( P h ( d k + 1 − d k Δ t ) , q h ) , C_\tau^{k+1}(q_h):=m_{\mathrm{eff},h}\left(P_h\left(\frac{d^{k+1}-d^k}{\Delta t}\right),q_h\right), C τ k + 1 ( q h ) := m eff , h ( P h ( Δ t d k + 1 − d k ) , q h ) ,
R F k + 1 ( q h ) : = λ − 1 α ( p ˉ T , F k + 1 − p T , F k + 1 , q h ) Ω , R_F^{k+1}(q_h):=\lambda^{-1}\alpha(\bar p_{T,F}^{k+1}-p_{T,F}^{k+1},q_h)_\Omega, R F k + 1 ( q h ) := λ − 1 α ( p ˉ T , F k + 1 − p T , F k + 1 , q h ) Ω ,
and p ˉ ˙ k + 1 : = 1 Δ t ∫ I k + 1 p ˙ ( t ) d t = p ( t k + 1 ) − p ( t k ) Δ t \dot{\bar p}^{k+1}:=\frac{1}{\Delta t}\int_{I_{k+1}}\dot p(t)\,\mathrm{d}t=\frac{p(t_{k+1})-p(t_k)}{\Delta t} p ˉ ˙ k + 1 := Δ t 1 ∫ I k + 1 p ˙ ( t ) d t = Δ t p ( t k + 1 ) − p ( t k ) . The load lifts in (72) and its continuous counterpart give R F k + 1 = G f k + 1 ∣ Q h − G f , h k + 1 R_F^{k+1}=G_f^{k+1}|_{Q_h}-G_{f,h}^{k+1} R F k + 1 = G f k + 1 ∣ Q h − G f , h k + 1 by (73). This residual is the mechanical Galerkin error in the total pressure load lift. Set ξ 0 : = P h p ( 0 ) − p h 0 \xi^{0} := P_{h}p(0)-p_{h}^{0} ξ 0 := P h p ( 0 ) − p h 0 . Since p h 0 ∈ Q h p_{h}^{0} \in Q_{h} p h 0 ∈ Q h we have P h p h 0 = p h 0 P_{h}p_{h}^{0}=p_{h}^{0} P h p h 0 = p h 0 , so ξ 0 = P h ( p ( 0 ) − p h 0 ) \xi^{0}=P_{h}(p(0)-p_{h}^{0}) ξ 0 = P h ( p ( 0 ) − p h 0 ) and hence ∥ ξ 0 ∥ 0 ≤ ∥ p ( 0 ) − p h 0 ∥ 0 = O ( h ) \|\xi^{0}\|_{0}\leq\|p(0)-p_{h}^{0}\|_{0}=O(h) ∥ ξ 0 ∥ 0 ≤ ∥ p ( 0 ) − p h 0 ∥ 0 = O ( h ) by the initial approximation assumption.
An equation of Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 \Delta t\sum_{n=1}^{N}\|\xi^{n}\|_{0}^{2} Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 via the dual equation: Consider a discrete dual problem: Let φ n ∈ Q h \varphi^{n}\in Q_{h} φ n ∈ Q h , n = 1 , … , N n=1,\ldots,N n = 1 , … , N , with φ N + 1 : = 0 \varphi^{N+1}:=0 φ N + 1 := 0 , solve (77) with right-hand side ξ n \xi^{n} ξ n . Then, Lemma 5.7 gives (78)–(81). Testing (83) at step n n n with q h = φ n q_{h}=\varphi^{n} q h = φ n , summing Δ t ∑ n = 1 N \Delta t\sum_{n=1}^{N} Δ t ∑ n = 1 N , and applying summation by parts using φ N + 1 = 0 \varphi^{N+1}=0 φ N + 1 = 0 ,
Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 = − Δ t ∑ n = 1 N [ R M n ( φ n ) + R S n ( φ n ) + C τ n ( φ n ) − R F n ( φ n ) ] + m e f f , h ( ξ 0 , φ 1 ) . (106) \Delta t\sum_{n=1}^{N}\|\xi^{n}\|_{0}^{2}=-\Delta t\sum_{n=1}^{N}\left[R_{M}^{n}(\varphi^{n})+R_{S}^{n}(\varphi^{n})+C_{\tau}^{n}(\varphi^{n})-R_{F}^{n}(\varphi^{n})\right]+m_{\mathrm{eff},h}(\xi^{0},\varphi^{1}). \tag*{(106)} Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 = − Δ t n = 1 ∑ N [ R M n ( φ n ) + R S n ( φ n ) + C τ n ( φ n ) − R F n ( φ n ) ] + m eff , h ( ξ 0 , φ 1 ) . (106)
The boundary term m e f f , h ( ξ 0 , φ 1 ) m_{\mathrm{eff},h}(\xi^{0},\varphi^{1}) m eff , h ( ξ 0 , φ 1 ) can be estimated by
∣ m e f f , h ( ξ 0 , φ 1 ) ∣ ≤ C ∥ ξ 0 ∥ 0 ∥ φ 1 ∥ 0 ≤ C ∥ ξ 0 ∥ 0 max 1 ≤ n ≤ N ∥ φ n ∥ 0 ≤ C h ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 \left|m_{\mathrm{eff},h}(\xi^{0},\varphi^{1})\right|\leq C\|\xi^{0}\|_{0}\|\varphi^{1}\|_{0}\leq C\|\xi^{0}\|_{0}\max_{1\leq n\leq N}\|\varphi^{n}\|_{0}\leq Ch\left(\Delta t\sum_{n=1}^{N}\|\xi^{n}\|_{0}^{2}\right)^{1/2} m eff , h ( ξ 0 , φ 1 ) ≤ C ∥ ξ 0 ∥ 0 ∥ φ 1 ∥ 0 ≤ C ∥ ξ 0 ∥ 0 1 ≤ n ≤ N max ∥ φ n ∥ 0 ≤ C h ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2
from the previous estimate of ∥ ξ 0 ∥ 0 \|\xi^{0}\|_{0} ∥ ξ 0 ∥ 0 and (81).
Estimates of residuals: We now prove error bounds of the residual terms R M R_{M} R M , R S R_{S} R S , C τ C_{\tau} C τ , R F R_{F} R F . We first estimate R S R_{S} R S . Since the diffusion form is time-independent and linear, and p ˉ k + 1 \bar{p}^{k+1} p ˉ k + 1 is the exact cell average of p p p over I k + 1 I_{k+1} I k + 1 :
1 Δ t ∫ I k + 1 ( κ p ∇ p ( t ) , ∇ φ k + 1 ) Ω d t = ( κ p ∇ p ˉ k + 1 , ∇ φ k + 1 ) Ω . \frac{1}{\Delta t}\int_{I_{k+1}}(\kappa_{p}\nabla p(t),\nabla\varphi^{k+1})_{\Omega}\,dt=(\kappa_{p}\nabla\bar{p}^{k+1},\nabla\varphi^{k+1})_{\Omega}. Δ t 1 ∫ I k + 1 ( κ p ∇ p ( t ) , ∇ φ k + 1 ) Ω d t = ( κ p ∇ p ˉ k + 1 , ∇ φ k + 1 ) Ω .
Hence R S k + 1 ( φ k + 1 ) = ( κ p ∇ ( I − P h ) p ˉ k + 1 , ∇ φ k + 1 ) Ω R_{S}^{k+1}(\varphi^{k+1})=(\kappa_{p}\nabla(I-P_{h})\bar{p}^{k+1},\nabla\varphi^{k+1})_{\Omega} R S k + 1 ( φ k + 1 ) = ( κ p ∇ ( I − P h ) p ˉ k + 1 , ∇ φ k + 1 ) Ω . Since p ˉ k + 1 ∈ H 2 ( Ω ) \bar{p}^{k+1}\in H^{2}(\Omega) p ˉ k + 1 ∈ H 2 ( Ω ) by Jensen’s inequality and p ∈ L t 2 H x 2 p\in L^{2}_{t}H^{2}_{x} p ∈ L t 2 H x 2 , property (c) in Lemma 5.3 gives ∥ ( I − P h ) p ˉ k + 1 ∥ H 1 ≤ C h ∣ p ˉ k + 1 ∣ H 2 \|(I-P_{h})\bar{p}^{k+1}\|_{H^{1}}\leq Ch|\bar{p}^{k+1}|_{H^{2}} ∥ ( I − P h ) p ˉ k + 1 ∥ H 1 ≤ C h ∣ p ˉ k + 1 ∣ H 2 . Multiplying by Δ t \Delta t Δ t , summing, Cauchy–Schwarz and Hölder inequalities with (79) give
∣ Δ t ∑ n = 1 N R S n ( φ n ) ∣ ≤ C h ∥ p ∥ L t 2 H x 2 ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 . (107) \left|\Delta t\sum_{n=1}^{N}R_{S}^{n}(\varphi^{n})\right|\leq Ch\|p\|_{L^{2}_{t}H^{2}_{x}}\left(\Delta t\sum_{n=1}^{N}\|\xi^{n}\|_{0}^{2}\right)^{1/2}. \tag*{(107)} Δ t n = 1 ∑ N R S n ( φ n ) ≤ C h ∥ p ∥ L t 2 H x 2 ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2 . (107)
We now estimate R F R_{F} R F . The continuous load-driven system for ( u ˉ f k + 1 , p ˉ T , f k + 1 ) (\bar{u}_{f}^{k+1},\bar{p}_{T,f}^{k+1}) ( u ˉ f k + 1 , p ˉ T , f k + 1 ) and its discrete counterpart ( u f , h k + 1 , p T , f , h k + 1 ) (u_{f,h}^{k+1},p_{T,f,h}^{k+1}) ( u f , h k + 1 , p T , f , h k + 1 ) both solve a perturbed Stokes problem with the same right-hand side ( f ˉ k + 1 , 0 ) (\bar{f}^{k+1},0) ( f ˉ k + 1 , 0 ) as in (72). By quasi-optimality for the Taylor–Hood pair, which bounds the pressure error by the best-approximation errors of both the displacement and the total pressure (in Definition 5.2), and the Stokes-type regularity assumption (2) of Appendix A (with F = f ˉ k + 1 F=\bar{f}^{k+1} F = f ˉ k + 1 , ζ = 0 \zeta=0 ζ = 0 ),
∥ p ˉ T , f k + 1 − p T , f , h k + 1 ∥ 0 ≤ C h ( ∥ u ˉ f k + 1 ∥ H 2 + ∥ p ˉ T , f k + 1 ∥ H 1 ) ≤ C h ∥ f ˉ k + 1 ∥ 0 . \|\bar{p}_{T,f}^{k+1}-p_{T,f,h}^{k+1}\|_{0}\leq Ch\left(\|\bar{u}_{f}^{k+1}\|_{H^{2}}+\|\bar{p}_{T,f}^{k+1}\|_{H^{1}}\right)\leq Ch\|\bar{f}^{k+1}\|_{0}. ∥ p ˉ T , f k + 1 − p T , f , h k + 1 ∥ 0 ≤ C h ( ∥ u ˉ f k + 1 ∥ H 2 + ∥ p ˉ T , f k + 1 ∥ H 1 ) ≤ C h ∥ f ˉ k + 1 ∥ 0 .
Hence, for the dual solution { φ n } ∈ Q h \{\varphi^{n}\}\in Q_{h} { φ n } ∈ Q h ,
∣ Δ t ∑ n = 1 N R F n ( φ n ) ∣ ≤ C λ − 1 α h ∥ f ∥ L t 2 L x 2 ( Δ t ∑ n = 1 N ∥ φ n ∥ 0 2 ) 1 / 2 ≤ C h ⋅ C d a t a ⋅ ( Δ t ∑ n = 1 N ∥ ξ n ∥ 0 2 ) 1 / 2 . (108) \left|\Delta t\sum_{n=1}^{N}R_{F}^{n}(\varphi^{n})\right|\leq C\lambda^{-1}\alpha h\|f\|_{L^{2}_{t}L^{2}_{x}}\left(\Delta t\sum_{n=1}^{N}\|\varphi^{n}\|_{0}^{2}\right)^{1/2}\leq Ch\cdot C_{\mathrm{data}}\cdot\left(\Delta t\sum_{n=1}^{N}\|\xi^{n}\|_{0}^{2}\right)^{1/2}. \tag*{(108)} Δ t n = 1 ∑ N R F n ( φ n ) ≤ C λ − 1 α h ∥ f ∥ L t 2 L x 2 ( Δ t n = 1 ∑ N ∥ φ n ∥ 0 2 ) 1/2 ≤ C h ⋅ C data ⋅ ( Δ t n = 1 ∑ N ∥ ξ n ∥ 0 2 ) 1/2 . (108)
by (78) where C d a t a : = λ − 1 α T 1 / 2 ∥ f ∥ L t 2 L x 2 C_{\mathrm{data}}:=\lambda^{-1}\alpha T^{1/2}\|f\|_{L^{2}_{t}L^{2}_{x}} C data := λ − 1 α T 1/2 ∥ f ∥ L t 2 L x 2 is λ \lambda λ -uniform for λ ≥ λ 0 \lambda\geq\lambda_{0} λ ≥ λ 0 .
We prove bounds for R M R_{M} R M and C τ C_{\tau} C τ . Let E h ( z , q h ) : = m e f f ( z , q h ) − m e f f , h ( P h z , q h ) E_{h}(z,q_{h}):=m_{\mathrm{eff}}(z,q_{h})-m_{\mathrm{eff},h}(P_{h}z,q_{h}) E h ( z , q h ) := m eff ( z , q h ) − m eff , h ( P h z , q h ) for z ∈ H 0 1 ( Ω ) z\in H_{0}^{1}(\Omega) z ∈ H 0 1 ( Ω ) . Since p ˉ n = d t p ˉ n + d t d n \bar{p}^{n}=d_{t}\bar{p}^{n}+d_{t}d^{n} p ˉ n = d t p ˉ n + d t d n with d t d n : = ( d n − d n − 1 ) / Δ t d_{t}d^{n}:=(d^{n}-d^{n-1})/\Delta t d t d n := ( d n − d n − 1 ) /Δ t , the definitions give
R M n ( q h ) + C τ n ( q h ) = E h ( d t p ˉ n , q h ) + m e f f ( d t d n , q h ) . R_{M}^{n}(q_{h})+C_{\tau}^{n}(q_{h})=E_{h}(d_{t}\bar{p}^{n},q_{h})+m_{\mathrm{eff}}(d_{t}d^{n},q_{h}). R M n ( q h ) + C τ n ( q h ) = E h ( d t p ˉ n , q h ) + m eff ( d t d n , q h ) .
Summation by parts with δ t φ n = ( φ n − φ n + 1 ) / Δ t \delta_t\varphi^n = (\varphi^n-\varphi^{n+1})/\Delta t δ t φ n = ( φ n − φ n + 1 ) /Δ t , φ N + 1 = 0 \varphi^{N+1}=0 φ N + 1 = 0 , p ‾ 0 = p ( 0 ) \overline{p}^0=p(0) p 0 = p ( 0 ) , and d 0 = 0 d^0=0 d 0 = 0 gives
Δ t ∑ n = 1 N ( R M n + C τ n ) ( φ n ) = Δ t ∑ n = 1 N [ E h ( p ‾ n , δ t φ n ) + m e f f ( d n , δ t φ n ) ] − E h ( p ( 0 ) , φ 1 ) . \Delta t\sum_{n=1}^{N}(R_M^n+C_\tau^n)(\varphi^n)=\Delta t\sum_{n=1}^{N}\left[E_h(\overline{p}^n,\delta_t\varphi^n)+m_{\mathrm{eff}}(d^n,\delta_t\varphi^n)\right]-E_h(p(0),\varphi^1). Δ t n = 1 ∑ N ( R M n + C τ n ) ( φ n ) = Δ t n = 1 ∑ N [ E h ( p n , δ t φ n ) + m eff ( d n , δ t φ n ) ] − E h ( p ( 0 ) , φ 1 ) .
Here p ‾ n ∈ H 0 1 ( Ω ) \overline{p}^n\in H_0^1(\Omega) p n ∈ H 0 1 ( Ω ) , and p ( 0 ) ∈ H 0 1 ( Ω ) p(0)\in H_0^1(\Omega) p ( 0 ) ∈ H 0 1 ( Ω ) by the embedding H 1 ( 0 , T ; L 2 ) ∩ L 2 ( 0 , T ; H 0 1 ) ⊂ C ( [ 0 , T ] ; H 0 1 ) H^1(0,T;L^2)\cap L^2(0,T;H_0^1)\subset C([0,T];H_0^1) H 1 ( 0 , T ; L 2 ) ∩ L 2 ( 0 , T ; H 0 1 ) ⊂ C ([ 0 , T ] ; H 0 1 ) . Writing d n = Δ t − 1 ∫ t n t n + 1 p ( s ) d s d^n=\Delta t^{-1}\int_{t_n}^{t_{n+1}}p(s)\,ds d n = Δ t − 1 ∫ t n t n + 1 p ( s ) d s , Cauchy–Schwarz yields Δ t ∑ n = 1 N ∥ d n ∥ 0 2 ≤ Δ t 2 ∥ p ∥ L t 2 L x 2 2 \Delta t\sum_{n=1}^{N}\lVert d^n\rVert_{0}^2\leq\Delta t^2\lVert p\rVert_{L_t^2L_x^2}^2 Δ t ∑ n = 1 N ∥ d n ∥ 0 2 ≤ Δ t 2 ∥ p ∥ L t 2 L x 2 2 . By Lemma 5.6, Δ t ∑ n ∥ p ‾ n ∥ H 1 2 ≤ ∥ p ∥ L t 2 H x 1 2 \Delta t\sum_n\lVert\overline{p}^n\rVert_{H^1}^2\leq\lVert p\rVert_{L_t^2H_x^1}^2 Δ t ∑ n ∥ p n ∥ H 1 2 ≤ ∥ p ∥ L t 2 H x 1 2 due to Jensen’s inequality, boundedness of m e f f m_{\mathrm{eff}} m eff , (80) and (81) yield
∣ Δ t ∑ n = 1 N ( R M n + C τ n ) ( φ n ) ∣ ≤ C ( h ∥ p ∥ L t 2 H x 1 + h ∥ p ( 0 ) ∥ H 1 + Δ t ∥ p ˙ ∥ L t 2 L x 2 ) ( Δ t ∑ n = 1 N ∥ ζ n ∥ 0 2 ) 1 / 2 . (109) \left|\Delta t\sum_{n=1}^{N}(R_M^n+C_\tau^n)(\varphi^n)\right|\leq C\left(h\lVert p\rVert_{L_t^2H_x^1}+h\lVert p(0)\rVert_{H^1}+\Delta t\lVert\dot{p}\rVert_{L_t^2L_x^2}\right)\left(\Delta t\sum_{n=1}^{N}\lVert\zeta^n\rVert_0^2\right)^{1/2}. \tag*{(109)} Δ t n = 1 ∑ N ( R M n + C τ n ) ( φ n ) ≤ C ( h ∥ p ∥ L t 2 H x 1 + h ∥ p ( 0 ) ∥ H 1 + Δ t ∥ p ˙ ∥ L t 2 L x 2 ) ( Δ t n = 1 ∑ N ∥ ζ n ∥ 0 2 ) 1/2 . (109)
The conclusion then follows by collecting (85), (86), (87), and the triangle inequality.
Lemma 5.9. Under the assumptions of Lemma 5.8, (71) holds.
Proof. We use the integrated mechanical reconstruction. For t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] and j = 1 , … , N j=1,\ldots,N j = 1 , … , N ,
u ( t ) = u ( 0 ) + ∫ 0 t u f ( s ) d s + U ( p ( t ) − p ( 0 ) ) , u(t)=u(0)+\int_0^t u_f(s)\,ds+\mathcal{U}(p(t)-p(0)), u ( t ) = u ( 0 ) + ∫ 0 t u f ( s ) d s + U ( p ( t ) − p ( 0 )) ,
p T ( t ) = p T ( 0 ) + ∫ 0 t p T , f ( s ) d s + P T ( p ( t ) − p ( 0 ) ) , p_T(t)=p_T(0)+\int_0^t p_{T,f}(s)\,ds+\mathcal{P}_T(p(t)-p(0)), p T ( t ) = p T ( 0 ) + ∫ 0 t p T , f ( s ) d s + P T ( p ( t ) − p ( 0 )) ,
u h j = u h 0 + Δ t ∑ ℓ = 1 j u f , h ℓ + U h ( p h j − p h 0 ) , u_h^j=u_h^0+\Delta t\sum_{\ell=1}^{j}u_{f,h}^{\ell}+\mathcal{U}_h(p_h^j-p_h^0), u h j = u h 0 + Δ t ℓ = 1 ∑ j u f , h ℓ + U h ( p h j − p h 0 ) ,
p T , h j = p T , h 0 + Δ t ∑ ℓ = 1 j p T , f , h ℓ + P T , h ( p h j − p h 0 ) . p_{T,h}^j=p_{T,h}^0+\Delta t\sum_{\ell=1}^{j}p_{T,f,h}^{\ell}+\mathcal{P}_{T,h}(p_h^j-p_h^0). p T , h j = p T , h 0 + Δ t ℓ = 1 ∑ j p T , f , h ℓ + P T , h ( p h j − p h 0 ) .
Here ( u f , h ℓ , p T , f , h ℓ ) (u_{f,h}^{\ell},p_{T,f,h}^{\ell}) ( u f , h ℓ , p T , f , h ℓ ) is the mechanical load lift in (72). Taking the cell average of the first two identities and comparing with the last two reduces the error to three contributions.
First, the initial-data errors are O ( h ) O(h) O ( h ) by the numerical initial approximation assumption. Second, the continuous and discrete load lifts satisfy, by mixed Galerkin quasi-optimality and the mechanical H 2 × H 1 H^2\times H^1 H 2 × H 1 regularity estimate,
∥ u f ( s ) − u f , h ( s ) ∥ V + ∥ p T , f ( s ) − p T , f , h ( s ) ∥ 0 ≤ C h ∥ f ( s ) ∥ 0 , \lVert u_f(s)-u_{f,h}(s)\rVert_V+\lVert p_{T,f}(s)-p_{T,f,h}(s)\rVert_0\leq Ch\lVert f(s)\rVert_0, ∥ u f ( s ) − u f , h ( s ) ∥ V + ∥ p T , f ( s ) − p T , f , h ( s ) ∥ 0 ≤ C h ∥ f ( s ) ∥ 0 ,
The time averages of these lifts are u f , h ℓ , p T , f , h ℓ u_{f,h}^{\ell},p_{T,f,h}^{\ell} u f , h ℓ , p T , f , h ℓ , as defined above. Replacing a cumulative integral at t j t_j t j by its average over I j I_j I j produces an additional O ( Δ t ) O(\Delta t) O ( Δ t ) term in ℓ t 2 \ell_t^2 ℓ t 2 , since
∥ 1 Δ t ∫ I j ∫ t t j u f ( s ) d s d t ∥ V ≤ Δ t 1 / 2 ∥ u f ∥ L 2 ( I j ; V ) , \left\lVert\frac{1}{\Delta t}\int_{I_j}\int_t^{t_j}u_f(s)\,ds\,dt\right\rVert_V\leq\Delta t^{1/2}\lVert u_f\rVert_{L^2(I_j;V)}, Δ t 1 ∫ I j ∫ t t j u f ( s ) d s d t V ≤ Δ t 1/2 ∥ u f ∥ L 2 ( I j ; V ) ,
and similarly for p T , f p_{T,f} p T , f . Consequently the total load-lift contribution is bounded in ℓ t 2 ( V × Q T ) \ell_t^2(V\times Q_T) ℓ t 2 ( V × Q T ) by C ( h + Δ t ) ∥ f ∥ L t 2 L x 2 C(h+\Delta t)\lVert f\rVert_{L_t^2L_x^2} C ( h + Δ t ) ∥ f ∥ L t 2 L x 2 .
Third, for y ∈ H 0 1 ( Ω ) y\in H_0^1(\Omega) y ∈ H 0 1 ( Ω ) , mixed Galerkin approximation, the L 2 L^2 L 2 -approximation property of P h P_h P h , and the regularity assumptions give
∥ y − y h P h y ∥ V + ∥ P T y − P T , h P h y ∥ 0 ≤ C h ∥ y ∥ H 1 . (110) \lVert y-y_hP_hy\rVert_V+\lVert\mathcal{P}_Ty-\mathcal{P}_{T,h}P_hy\rVert_0\leq Ch\lVert y\rVert_{H^1}. \tag*{(110)} ∥ y − y h P h y ∥ V + ∥ P T y − P T , h P h y ∥ 0 ≤ C h ∥ y ∥ H 1 . (110)
Apply this to v = p ˉ j − p ( 0 ) v=\bar{p}^{j}-p(0) v = p ˉ j − p ( 0 ) . By the L t 2 H x 2 ∩ H t 1 L x 2 L_{t}^{2}H_{x}^{2}\cap H_{t}^{1}L_{x}^{2} L t 2 H x 2 ∩ H t 1 L x 2 regularity in Assumption R, the embedding L t 2 H x 2 ∩ H t 1 L x 2 ↪ C 0 ( [ 0 , T ] ; H 1 ( Ω ) ) L_{t}^{2}H_{x}^{2}\cap H_{t}^{1}L_{x}^{2}\hookrightarrow C^{0}([0,T];H^{1}(\Omega)) L t 2 H x 2 ∩ H t 1 L x 2 ↪ C 0 ([ 0 , T ] ; H 1 ( Ω )) gives ∥ p ∥ C 0 ( [ 0 , T ] ; H 1 ) < ∞ \|p\|_{C^{0}([0,T];H^{1})}<\infty ∥ p ∥ C 0 ([ 0 , T ] ; H 1 ) < ∞ . Using the uniform stability of ( U h , P T , h ) (\mathcal{U}_{h},\mathcal{P}_{T,h}) ( U h , P T , h ) ,
∥ U ( p ˉ j − p ( 0 ) ) − U h ( p h j − p h 0 ) ∥ V + ∥ P T ( p ˉ j − p ( 0 ) ) − P T , h ( p h j − p h 0 ) ∥ 0 ≤ C h ∥ p ˉ j − p ( 0 ) ∥ H 1 + C ( ∥ P h p ˉ j − p h j ∥ 0 + ∥ P h p ( 0 ) − p h 0 ∥ 0 ) . \begin{aligned}
\|\mathcal{U}(\bar{p}^{j}-p(0))-\mathcal{U}_{h}(p_{h}^{j}-p_{h}^{0})\|_{V}+\|\mathcal{P}_{T}(\bar{p}^{j}-p(0))-\mathcal{P}_{T,h}(p_{h}^{j}-p_{h}^{0})\|_{0} \\
&\leq Ch\|\bar{p}^{j}-p(0)\|_{H^{1}}+C(\|P_{h}\bar{p}^{j}-p_{h}^{j}\|_{0}+\|P_{h}p(0)-p_{h}^{0}\|_{0}).
\end{aligned} ∥ U ( p ˉ j − p ( 0 )) − U h ( p h j − p h 0 ) ∥ V + ∥ P T ( p ˉ j − p ( 0 )) − P T , h ( p h j − p h 0 ) ∥ 0 ≤ C h ∥ p ˉ j − p ( 0 ) ∥ H 1 + C ( ∥ P h p ˉ j − p h j ∥ 0 + ∥ P h p ( 0 ) − p h 0 ∥ 0 ) .
The first and last terms are O ( h ) O(h) O ( h ) in ℓ t 2 \ell_{t}^{2} ℓ t 2 . For the middle term, the decomposition in the proof of Lemma 5.8 and Jensen’s inequality give
Δ t ∑ j = 1 N ∥ P h p ˉ j − p h j ∥ 0 2 ≤ ∑ j = 1 N ∫ I j ∥ p ( t ) − p h j ∥ 0 2 d t ≤ C ( h + Δ t ) 2 . \Delta t\sum_{j=1}^{N}\|P_{h}\bar{p}^{j}-p_{h}^{j}\|_{0}^{2}\leq\sum_{j=1}^{N}\int_{I_{j}}\|p(t)-p_{h}^{j}\|_{0}^{2}\,\mathrm{d}t\leq C(h+\Delta t)^{2}. Δ t j = 1 ∑ N ∥ P h p ˉ j − p h j ∥ 0 2 ≤ j = 1 ∑ N ∫ I j ∥ p ( t ) − p h j ∥ 0 2 d t ≤ C ( h + Δ t ) 2 .
Combining the three contributions proves (71).
Proof of Lemma 5.4. Estimate (70) follows directly from Lemma 5.8. The reduction to the scalar scheme (74) is established in Lemmas 2.5, 2.6, 4.1, and 5.5. Estimate (71) follows from Lemma 5.9.
Estimates for the control error
We first record a strong monotonicity property of the discrete reduced gradient j h ′ j_{h}^{\prime} j h ′ on M p , a d h \mathcal{M}_{p,ad}^{h} M p , a d h , and then a fixed-control estimate for the adjoint error, before combining them to bound the control error.
Lemma 5.10. For any m p , h , m ~ p , h ∈ M p , a d h m_{p,h},\widetilde{m}_{p,h}\in\mathcal{M}_{p,ad}^{h} m p , h , m p , h ∈ M p , a d h ,
( j h ′ ( m p , h ) − j h ′ ( m ~ p , h ) , m p , h − m ~ p , h ) L t 2 L x 2 ≥ γ p ∥ m p , h − m ~ p , h ∥ L t 2 L x 2 2 (111) \left(j_{h}^{\prime}(m_{p,h})-j_{h}^{\prime}(\widetilde{m}_{p,h}),m_{p,h}-\widetilde{m}_{p,h}\right)_{L_{t}^{2}L_{x}^{2}}\geq\gamma_{p}\|m_{p,h}-\widetilde{m}_{p,h}\|_{L_{t}^{2}L_{x}^{2}}^{2} \tag*{(111)} ( j h ′ ( m p , h ) − j h ′ ( m p , h ) , m p , h − m p , h ) L t 2 L x 2 ≥ γ p ∥ m p , h − m p , h ∥ L t 2 L x 2 2 (111) where j h ′ j_{h}^{\prime} j h ′ is the discrete gradient (64).
Proof. Let m p , d : = m p , h − m ~ p , h m_{p,d}:=m_{p,h}-\widetilde{m}_{p,h} m p , d := m p , h − m p , h and let ( u d , h k + 1 , p T , d , h k + 1 , p d , h k + 1 ) (u_{d,h}^{k+1},p_{T,d,h}^{k+1},p_{d,h}^{k+1}) ( u d , h k + 1 , p T , d , h k + 1 , p d , h k + 1 ) denote the corresponding difference of the discrete states u h k + 1 ( m p , h ) − u h k + 1 ( m ~ p , h ) u_{h}^{k+1}(m_{p,h})-u_{h}^{k+1}(\widetilde{m}_{p,h}) u h k + 1 ( m p , h ) − u h k + 1 ( m p , h ) , etc. The derivative of the affine control-to-state map in the direction m p , d m_{p,d} m p , d is ( u d , h , p T , d , h , p d , h ) (u_{d,h},p_{T,d,h},p_{d,h}) ( u d , h , p T , d , h , p d , h ) , independent of the base point. Since j h j_{h} j h is quadratic, we obtain
( j h ′ ( m p , h ) − j h ′ ( m ~ p , h ) , m p , d ) L t 2 L x 2 = α u , C Δ t ∑ k = 0 N − 1 ∥ u d , h k + 1 ∥ 0 2 + α p , C Δ t ∑ k = 0 N − 1 ∥ p d , h k + 1 ∥ 0 2 + γ p ∥ m p , d ∥ L t 2 L x 2 2 . \left(j_{h}^{\prime}(m_{p,h})-j_{h}^{\prime}(\widetilde{m}_{p,h}),m_{p,d}\right)_{L_{t}^{2}L_{x}^{2}}=\alpha_{u,C}\Delta t\sum_{k=0}^{N-1}\|u_{d,h}^{k+1}\|_{0}^{2}+\alpha_{p,C}\Delta t\sum_{k=0}^{N-1}\|p_{d,h}^{k+1}\|_{0}^{2}+\gamma_{p}\|m_{p,d}\|_{L_{t}^{2}L_{x}^{2}}^{2}. ( j h ′ ( m p , h ) − j h ′ ( m p , h ) , m p , d ) L t 2 L x 2 = α u , C Δ t k = 0 ∑ N − 1 ∥ u d , h k + 1 ∥ 0 2 + α p , C Δ t k = 0 ∑ N − 1 ∥ p d , h k + 1 ∥ 0 2 + γ p ∥ m p , d ∥ L t 2 L x 2 2 .
Since α u , C , α p , C ≥ 0 \alpha_{u,C},\alpha_{p,C}\geq0 α u , C , α p , C ≥ 0 and γ p > 0 \gamma_{p}>0 γ p > 0 , dropping the nonnegative tracking terms yields (89).
We next bound the adjoint error for a fixed control, which is the key ingredient for the control-error estimate.
Lemma 5.11. Let m ~ p ∈ M p , a d \widetilde{m}_{p}\in\mathcal{M}_{p,ad} m p ∈ M p , a d be the continuous optimal control, with continuous state ( u ^ , p ^ T , p ^ ) (\widehat{u},\widehat{p}_{T},\widehat{p}) ( u , p T , p ) and continuous adjoint ( v ^ ∗ , q ^ T ∗ , q ^ ∗ ) (\widehat{v}^{*},\widehat{q}_{T}^{*},\widehat{q}^{*}) ( v ∗ , q T ∗ , q ∗ ) solving (35)–(36). Let ( u h ( m ~ p ) , p T , h ( m ~ p ) , p h ( m ~ p ) ) (u_{h}(\widetilde{m}_{p}),p_{T,h}(\widetilde{m}_{p}),p_{h}(\widetilde{m}_{p})) ( u h ( m p ) , p T , h ( m p ) , p h ( m p )) be the auxiliary fully discrete state driven by m ~ p \widetilde{m}_{p} m p as in (50), and let ( v h ∗ ( m ~ p ) , q T , h ∗ ( m ~ p ) , q h ∗ ( m ~ p ) ) (v_{h}^{*}(\widetilde{m}_{p}),q_{T,h}^{*}(\widetilde{m}_{p}),q_{h}^{*}(\widetilde{m}_{p})) ( v h ∗ ( m p ) , q T , h ∗ ( m p ) , q h ∗ ( m p )) be the corresponding auxiliary fully discrete adjoint, defined by (57)–(58) with the right-hand sides of (57a) and (57c) and with homogeneous terminal condition (58). Under (63) in Assumption R, there holds
( Δ t ∑ k = 0 N − 1 ∥ q h ∗ , k ( m ~ p ) − q ‾ ∗ , k + 1 ∥ 0 2 ) 1 / 2 ≤ C ( h + Δ t ) , (112) \left(\Delta t\sum_{k=0}^{N-1}\|q_{h}^{*,k}(\widetilde{m}_{p})-\overline{q}^{*,k+1}\|_{0}^{2}\right)^{1/2}\leq C(h+\Delta t), \tag*{(112)} ( Δ t k = 0 ∑ N − 1 ∥ q h ∗ , k ( m p ) − q ∗ , k + 1 ∥ 0 2 ) 1/2 ≤ C ( h + Δ t ) , (112) where q ‾ ∗ , k + 1 \overline{q}^{*,k+1} q ∗ , k + 1 denotes the time-average of q ∗ q^{*} q ∗ over ( t k , t k + 1 ) (t_{k},t_{k+1}) ( t k , t k + 1 ) , and C > 0 C>0 C > 0 is independent of h h h and Δ t \Delta t Δ t , uniform for λ ≥ λ 0 > 0 \lambda\geq\lambda_{0}>0 λ ≥ λ 0 > 0 if s 0 ≥ s ∗ > 0 s_{0}\geq s_{*}>0 s 0 ≥ s ∗ > 0 and the assumed regularity bounds are uniform.
Proof. Introduce the auxiliary discrete adjoint ( v ~ h ∗ , k , q ~ T , h ∗ , k , q ~ h ∗ , k ) (\widetilde{v}_h^{*,k},\widetilde{q}_{T,h}^{*,k},\widetilde{q}_h^{*,k}) ( v h ∗ , k , q T , h ∗ , k , q h ∗ , k ) defined by (57)–(58) with the interior right-hand sides of (57a) and (57c) driven by the continuous optimal state ( u ^ , p ^ ) (\widehat{u},\widehat{p}) ( u , p ) (i.e. α u , C ( u ^ k + 1 − u ‾ k + 1 , v h ) Ω \alpha_{u,C}(\widehat{u}^{k+1}-\overline{u}^{k+1},v_h)_{\Omega} α u , C ( u k + 1 − u k + 1 , v h ) Ω in place of α u , C ( u h k + 1 ( m ^ p ) − u ^ k + 1 , v h ) Ω \alpha_{u,C}(u_h^{k+1}(\widehat{m}_p)-\widehat{u}^{k+1},v_h)_{\Omega} α u , C ( u h k + 1 ( m p ) − u k + 1 , v h ) Ω , and analogously for the p p p -equation). Both auxiliary adjoint systems use the same homogeneous terminal condition (58); hence their difference has homogeneous terminal data. We split
q h ∗ , k ( m ^ p ) − q ‾ ∗ , k + 1 = ( q h ∗ , k ( m ^ p ) − q ~ h ∗ , k ) + ( q ~ h ∗ , k − q ‾ ∗ , k + 1 ) = : E 1 q , k + E 2 q , k . q_h^{*,k}(\widehat{m}_p)-\overline{q}^{*,k+1}=(q_h^{*,k}(\widehat{m}_p)-\widetilde{q}_h^{*,k})+(\widetilde{q}_h^{*,k}-\overline{q}^{*,k+1})=:E_1^{q,k}+E_2^{q,k}. q h ∗ , k ( m p ) − q ∗ , k + 1 = ( q h ∗ , k ( m p ) − q h ∗ , k ) + ( q h ∗ , k − q ∗ , k + 1 ) =: E 1 q , k + E 2 q , k .
As the first step, we prove a bound on E 1 E_1 E 1 . Subtracting the two discrete adjoint systems (driven by u h k + 1 ( m ^ p ) u_h^{k+1}(\widehat{m}_p) u h k + 1 ( m p ) versus u ^ k + 1 \widehat{u}^{k+1} u k + 1 , etc. in the interior right-hand sides, with homogeneous terminal data) shows that ( E 1 v , E 1 q T , E 1 q ) : = ( v h ∗ ( m ^ p ) − v ~ h ∗ , q T , h ∗ ( m ^ p ) − q ~ T , h ∗ , q h ∗ ( m ^ p ) − q ~ h ∗ ) (E_1^v,E_1^{q_T},E_1^q):=(v_h^{*}(\widehat{m}_p)-\widetilde{v}_h^{*},q_{T,h}^{*}(\widehat{m}_p)-\widetilde{q}_{T,h}^{*},q_h^{*}(\widehat{m}_p)-\widetilde{q}_h^{*}) ( E 1 v , E 1 q T , E 1 q ) := ( v h ∗ ( m p ) − v h ∗ , q T , h ∗ ( m p ) − q T , h ∗ , q h ∗ ( m p ) − q h ∗ ) solves the discrete adjoint system (57) with interior right-hand sides
F u k + 1 = u h k + 1 ( m ^ p ) − u ~ k + 1 , F p k + 1 = p h k + 1 ( m ^ p ) − p ~ k + 1 , F_u^{k+1}=u_h^{k+1}(\widehat{m}_p)-\widetilde{u}^{k+1},\qquad F_p^{k+1}=p_h^{k+1}(\widehat{m}_p)-\widetilde{p}^{k+1}, F u k + 1 = u h k + 1 ( m p ) − u k + 1 , F p k + 1 = p h k + 1 ( m p ) − p k + 1 ,
and with homogeneous terminal condition (58). Lemma 4.9 directly gives
( Δ t ∑ k = 0 N − 1 ∥ E 1 q , k ∥ 0 2 ) 1 / 2 ≤ C ( Δ t ∑ j = 1 N ( ∥ u h j ( m ^ p ) − u ‾ j ∥ 0 2 + ∥ p h j ( m ^ p ) − p ‾ j ∥ 0 2 ) ) 1 / 2 . (113) \left(\Delta t\sum_{k=0}^{N-1}\left\|E_1^{q,k}\right\|_0^2\right)^{1/2}\leq C\left(\Delta t\sum_{j=1}^{N}\left(\left\|u_h^j(\widehat{m}_p)-\overline{u}^j\right\|_0^2+\left\|p_h^j(\widehat{m}_p)-\overline{p}^j\right\|_0^2\right)\right)^{1/2}. \tag*{(113)} ( Δ t k = 0 ∑ N − 1 E 1 q , k 0 2 ) 1/2 ≤ C ( Δ t j = 1 ∑ N ( u h j ( m p ) − u j 0 2 + p h j ( m p ) − p j 0 2 ) ) 1/2 . (113)
By Lemma 5.4 (applied to the auxiliary state driven by m ^ p \widehat{m}_p m p ), the right-hand side is bounded by C ( h + Δ t ) C(h+\Delta t) C ( h + Δ t ) . Therefore
( Δ t ∑ k = 0 N − 1 ∥ E 1 q , k ∥ 0 2 ) 1 / 2 ≤ C ( h + Δ t ) . (114) \left(\Delta t\sum_{k=0}^{N-1}\left\|E_1^{q,k}\right\|_0^2\right)^{1/2}\leq C(h+\Delta t). \tag*{(114)} ( Δ t k = 0 ∑ N − 1 E 1 q , k 0 2 ) 1/2 ≤ C ( h + Δ t ) . (114)
We now prove a bound on E 2 E_2 E 2 . Reverse time as in the proof of Lemma 4.9: q ~ h l : = q ~ h ∗ , N − l \widetilde{q}_h^l:=\widetilde{q}_h^{*,N-l} q h l := q h ∗ , N − l and q ~ ∗ ( τ ) : = q ^ ∗ ( T − τ ) \widetilde{q}^{*}(\tau):=\widehat{q}^{*}(T-\tau) q ∗ ( τ ) := q ∗ ( T − τ ) . The interval ( t k , t k + 1 ] (t_k,t_{k+1}] ( t k , t k + 1 ] corresponds to the reversed cell I l I_l I l with l = N − k l=N-k l = N − k (up to endpoints), so q ‾ ‾ k + 1 = q ‾ l \overline{\overline{q}}^{k+1}=\overline{q}^{\,l} q k + 1 = q l and E 2 q , k = q ~ h l − q ‾ l E_2^{q,k}=\widetilde{q}_h^l-\overline{q}^{\,l} E 2 q , k = q h l − q l . By (62), eliminating the mechanical variables yields a scheme of the form (74) for q ~ h l \widetilde{q}_h^l q h l , with the same effective mass m e f f , h m_{\mathrm{eff},h} m eff , h , zero initial value, and data given by the cell averages of the adjoint forcing evaluated at the continuous optimal state; the continuous reversed adjoint satisfies the corresponding reduced equation with m e f f m_{\mathrm{eff}} m eff and zero initial value. The load-lift term enters with the opposite sign, which does not affect the bound on R F R_F R F . Under Assumption R, Lemma 5.8 therefore applies to ( q ~ ∗ , q ~ h ) (\widetilde{q}^{*},\widetilde{q}_h) ( q ∗ , q h ) , and Jensen’s inequality, Δ t ∥ q ~ ∗ l − q ~ h l ∥ 0 2 ≤ ∫ I l ∥ q ~ ∗ ( τ ) − q ~ h l ∥ 0 2 d τ \Delta t\|\widetilde{q}^{*l}-\widetilde{q}_h^l\|_0^2\leq\int_{I_l}\|\widetilde{q}^{*}(\tau)-\widetilde{q}_h^l\|_0^2\,d\tau Δ t ∥ q ∗ l − q h l ∥ 0 2 ≤ ∫ I l ∥ q ∗ ( τ ) − q h l ∥ 0 2 d τ , gives
( Δ t ∑ k = 0 N − 1 ∥ E 2 q , k ∥ 0 2 ) 1 / 2 ≤ C ( h + Δ t ) . (115) \left(\Delta t\sum_{k=0}^{N-1}\left\|E_2^{q,k}\right\|_0^2\right)^{1/2}\leq C(h+\Delta t). \tag*{(115)} ( Δ t k = 0 ∑ N − 1 E 2 q , k 0 2 ) 1/2 ≤ C ( h + Δ t ) . (115)
Finally, by the triangle inequality, (90) follows from (92) and (93).
Main error estimate for the control
Theorem 5.12. Let m ^ p \widehat{m}_p m p and m ^ p , h \widehat{m}_{p,h} m p , h be the continuous and discrete optimal controls, respectively. Assume s 0 ≥ s ∗ > 0 s_0\geq s_* >0 s 0 ≥ s ∗ > 0 . Under Assumption R (63), with a regularity bound C r e g C_{\mathrm{reg}} C reg uniform for λ ≥ λ 0 > 0 \lambda\geq\lambda_0>0 λ ≥ λ 0 > 0 (depending on λ 0 \lambda_0 λ 0 ), and with independence of C r e g C_{\mathrm{reg}} C reg from κ p \kappa_p κ p requiring the elliptic regularity constants in (63) to be uniform in κ p \kappa_p κ p , there holds
∥ m ^ p − m ^ p , h ∥ L t 2 L x 2 ≤ C ( h + Δ t ) , (116) \|\widehat{m}_p-\widehat{m}_{p,h}\|_{L^2_tL_x^2}\leq C(h+\Delta t), \tag*{(116)} ∥ m p − m p , h ∥ L t 2 L x 2 ≤ C ( h + Δ t ) , (116) where C > 0 C>0 C > 0 depends on γ p , T \gamma_p,T γ p , T and C r e g C_{\mathrm{reg}} C reg , is independent of h h h and Δ t \Delta t Δ t , and is uniform for $\lambda\geq\lambda_0>0.
Proof. Write m = m ^ p m=\hat m_p m = m ^ p , m h = m ^ p , h m_h=\hat m_{p,h} m h = m ^ p , h . We use the L 2 ( ( 0 , T ) × Ω ) L^2((0,T)\times\Omega) L 2 (( 0 , T ) × Ω ) inner product; discrete state and adjoint quantities are identified with their piecewise-constant-in-time extensions. The discrete functional is evaluated at controls in M p , a d h \mathcal{M}_{p,ad}^h M p , a d h .
By the discrete strong monotonicity (Lemma 5.10 ) and the discrete variational inequality at m h m_h m h tested with Π h M m \Pi_h^M m Π h M m , setting e m h = m h − Π h M m e_m^h=m_h-\Pi_h^M m e m h = m h − Π h M m gives
γ p ∥ e m h ∥ L t 2 L x 2 2 ≤ ( j h ′ ( m h ) − j h ′ ( Π h M m ) , e m h ) L t 2 L x 2 ≤ − ( j h ′ ( Π h M m ) , e m h ) L t 2 L x 2 . (118) \gamma_p\|e_m^h\|_{L_t^2L_x^2}^2\leq(j_h'(m_h)-j_h'(\Pi_h^M m),e_m^h)_{L_t^2L_x^2}\leq-(j_h'(\Pi_h^M m),e_m^h)_{L_t^2L_x^2}. \tag*{(118)} γ p ∥ e m h ∥ L t 2 L x 2 2 ≤ ( j h ′ ( m h ) − j h ′ ( Π h M m ) , e m h ) L t 2 L x 2 ≤ − ( j h ′ ( Π h M m ) , e m h ) L t 2 L x 2 . (118)
Add and subtract the continuous reduced gradient j ′ ( m ) = γ p m + q ^ ∗ j'(m)=\gamma_p m+\hat q^* j ′ ( m ) = γ p m + q ^ ∗ (Lemma 3.4 ). Because m h m_h m h is admissible, the continuous variational inequality gives ⟨ j ′ ( m ) , m h − m ⟩ ≥ 0 \langle j'(m),m_h-m\rangle\geq0 ⟨ j ′ ( m ) , m h − m ⟩ ≥ 0 . Hence
γ p ∥ e m h ∥ L t 2 L x 2 2 ≤ − ( j h ′ ( Π h M m ) − j ′ ( m ) , e m h ) L t 2 L x 2 + ( j ′ ( m ) , Π h M m − m ) L t 2 L x 2 ≤ ∣ ( j h ′ ( Π h M m ) − j ′ ( m ) , e m h ) L t 2 L x 2 ∣ + ( j ′ ( m ) , Π h M m − m ) L t 2 L x 2 . (119) \begin{aligned}
\gamma_p\|e_m^h\|_{L_t^2L_x^2}^2
&\leq-(j_h'(\Pi_h^M m)-j'(m),e_m^h)_{L_t^2L_x^2}+(j'(m),\Pi_h^M m-m)_{L_t^2L_x^2}\\
&\leq|(j_h'(\Pi_h^M m)-j'(m),e_m^h)_{L_t^2L_x^2}|+(j'(m),\Pi_h^M m-m)_{L_t^2L_x^2}. \tag*{(119)}
\end{aligned} γ p ∥ e m h ∥ L t 2 L x 2 2 ≤ − ( j h ′ ( Π h M m ) − j ′ ( m ) , e m h ) L t 2 L x 2 + ( j ′ ( m ) , Π h M m − m ) L t 2 L x 2 ≤ ∣ ( j h ′ ( Π h M m ) − j ′ ( m ) , e m h ) L t 2 L x 2 ∣ + ( j ′ ( m ) , Π h M m − m ) L t 2 L x 2 . (119)
The L 2 L^2 L 2 orthogonality of the space-time projection and Π h M m = Π h M m \Pi_h^M m=\Pi_h^M m Π h M m = Π h M m give
( j ′ ( m ) , Π h M m − m ) L t 2 L x 2 = − ( ( I − Π h M ) j ′ ( m ) , ( I − Π h M ) m ) L t 2 L x 2 ≤ C ( h + Δ t ) 2 , (j'(m),\Pi_h^M m-m)_{L_t^2L_x^2}=-((I-\Pi_h^M)j'(m),(I-\Pi_h^M)m)_{L_t^2L_x^2}\leq C(h+\Delta t)^2, ( j ′ ( m ) , Π h M m − m ) L t 2 L x 2 = − (( I − Π h M ) j ′ ( m ) , ( I − Π h M ) m ) L t 2 L x 2 ≤ C ( h + Δ t ) 2 ,
by Assumption R and the derived regularity (117) .
Since e m h ∈ M p , h e_m^h\in M_{p,h} e m h ∈ M p , h , only the M p , h M_{p,h} M p , h -projection of the discrete–continuous gradient difference matters. Using (84) and the continuous gradient identity, and observing that Π h M ( Π h M m − m ) = 0 \Pi_h^M(\Pi_h^M m-m)=0 Π h M ( Π h M m − m ) = 0 , we obtain
Π h M ( j h ′ ( Π h M m ) − j ′ ( m ) ) = Π h M ( q h ∗ ( Π h M m ) − q h ∗ ( m ) ) + Π h M ( q h ∗ ( m ) − q ^ ∗ ) . \Pi_h^M(j_h'(\Pi_h^M m)-j'(m))=\Pi_h^M(q_h^*(\Pi_h^M m)-q_h^*(m))+\Pi_h^M(q_h^*(m)-\hat q^*). Π h M ( j h ′ ( Π h M m ) − j ′ ( m )) = Π h M ( q h ∗ ( Π h M m ) − q h ∗ ( m )) + Π h M ( q h ∗ ( m ) − q ^ ∗ ) .
The control-to-adjoint sensitivity estimate (88) , Jensen’s inequality, and the projection error of m m m imply
∥ q h ∗ ( Π h M m ) − q h ∗ ( m ) ∥ L t 2 L x 2 ≤ C ∥ Π h M m − m ∥ L t 2 L x 2 ≤ C ( h + Δ t ) . \|q_h^*(\Pi_h^M m)-q_h^*(m)\|_{L_t^2L_x^2}\leq C\|\Pi_h^M m-m\|_{L_t^2L_x^2}\leq C(h+\Delta t). ∥ q h ∗ ( Π h M m ) − q h ∗ ( m ) ∥ L t 2 L x 2 ≤ C ∥ Π h M m − m ∥ L t 2 L x 2 ≤ C ( h + Δ t ) .
The fixed-control adjoint error (112) , together with the time-averaging projection error of q ^ ∗ \hat q^* q ^ ∗ , gives
∥ Π h M ( q h ∗ ( m ) − q ^ ∗ ) ∥ L t 2 L x 2 ≤ C ( h + Δ t ) . \|\Pi_h^M(q_h^*(m)-\hat q^*)\|_{L_t^2L_x^2}\leq C(h+\Delta t). ∥ Π h M ( q h ∗ ( m ) − q ^ ∗ ) ∥ L t 2 L x 2 ≤ C ( h + Δ t ) .
Therefore
∣ ( j h ′ ( Π h M m ) − j ′ ( m ) , e m h ) L t 2 L x 2 ∣ ≤ C ( h + Δ t ) ∥ e m h ∥ L t 2 L x 2 . |(j_h'(\Pi_h^M m)-j'(m),e_m^h)_{L_t^2L_x^2}|\leq C(h+\Delta t)\|e_m^h\|_{L_t^2L_x^2}. ∣ ( j h ′ ( Π h M m ) − j ′ ( m ) , e m h ) L t 2 L x 2 ∣ ≤ C ( h + Δ t ) ∥ e m h ∥ L t 2 L x 2 .
Substitution into (119) gives
γ p ∥ e m h ∥ L t 2 L x 2 2 ≤ C ( h + Δ t ) ∥ e m h ∥ L t 2 L x 2 + C ( h + Δ t ) 2 , \gamma_p\|e_m^h\|_{L_t^2L_x^2}^2\leq C(h+\Delta t)\|e_m^h\|_{L_t^2L_x^2}+C(h+\Delta t)^2, γ p ∥ e m h ∥ L t 2 L x 2 2 ≤ C ( h + Δ t ) ∥ e m h ∥ L t 2 L x 2 + C ( h + Δ t ) 2 ,
and hence ∥ e m h ∥ ≤ C ( h + Δ t ) \|e_m^h\|\leq C(h+\Delta t) ∥ e m h ∥ ≤ C ( h + Δ t ) . Finally,
∥ m − m h ∥ L t 2 L x 2 ≤ ∥ m − Π h M m ∥ L t 2 L x 2 + ∥ e m h ∥ L t 2 L x 2 ≤ C ( h + Δ t ) , \|m-m_h\|_{L_t^2L_x^2}\leq\|m-\Pi_h^M m\|_{L_t^2L_x^2}+\|e_m^h\|_{L_t^2L_x^2}\leq C(h+\Delta t), ∥ m − m h ∥ L t 2 L x 2 ≤ ∥ m − Π h M m ∥ L t 2 L x 2 + ∥ e m h ∥ L t 2 L x 2 ≤ C ( h + Δ t ) ,
proving (116) .
Theorem 5.14. Under the assumptions of Theorem 5.12 and the regularity (83) , the state error satisfies
( Δ t ∑ j = 1 N ( ∥ u ˉ j − u ^ h j ∥ 1 2 + ∥ p ^ T j − p ^ T , h j ∥ 0 2 + ∥ p ˉ j − p ^ h j ∥ 0 2 ) ) 1 / 2 ≤ C ( h + Δ t ) , (120) \left(\Delta t \sum_{j=1}^{N} \left(\left\|\bar{u}^{j}-\hat{u}_{h}^{j}\right\|_{1}^{2}+\left\|\widehat{p}_{T}^{j}-\hat{p}_{T,h}^{j}\right\|_{0}^{2}+\left\|\bar{p}^{j}-\hat{p}_{h}^{j}\right\|_{0}^{2}\right)\right)^{1/2} \le C(h+\Delta t), \tag*{(120)} ( Δ t j = 1 ∑ N ( u ˉ j − u ^ h j 1 2 + p T j − p ^ T , h j 0 2 + p ˉ j − p ^ h j 0 2 ) ) 1/2 ≤ C ( h + Δ t ) , (120) where C > 0 C>0 C > 0 is independent of h h h and Δ t \Delta t Δ t , uniform for λ ≥ λ 0 > 0 \lambda\ge\lambda_{0}>0 λ ≥ λ 0 > 0 .
Proof. We split
u ˉ j − u ^ h j = ( u ˉ j − u h j ( m ^ p ) ) + ( u h j ( m ^ p ) − u ^ h j ) , \bar{u}^{j}-\hat{u}_{h}^{j}=\left(\bar{u}^{j}-u_{h}^{j}(\hat{m}_{p})\right)+\left(u_{h}^{j}(\hat{m}_{p})-\hat{u}_{h}^{j}\right), u ˉ j − u ^ h j = ( u ˉ j − u h j ( m ^ p ) ) + ( u h j ( m ^ p ) − u ^ h j ) ,
and analogously for p ^ T j \widehat{p}_{T}^{j} p T j and p ˉ j \bar{p}^{j} p ˉ j . The first term is the fixed-control discretization error, bounded by O ( h + Δ t ) O(h+\Delta t) O ( h + Δ t ) in the ℓ t 2 \ell_{t}^{2} ℓ t 2 sense by both components of Lemma 5.4 : (93) for the ( u , p T ) (u,p_{T}) ( u , p T ) -part and (92) for the p p p -part. Together with Lemma 4.3 , one obtains
( Δ t ∑ j = 1 N ( ∥ u ˉ j − u h j ( m ^ p ) ∥ 1 2 + ∥ p ˉ T j − p T , h j ( m ^ p ) ∥ 0 2 + ∥ p ˉ j − p h j ( m ^ p ) ∥ 0 2 ) ) 1 / 2 ≤ C ( h + Δ t ) . (121) \left(\Delta t \sum_{j=1}^{N} \left(\left\|\bar{u}^{j}-u_{h}^{j}(\hat{m}_{p})\right\|_{1}^{2}+\left\|\bar{p}_{T}^{j}-p_{T,h}^{j}(\hat{m}_{p})\right\|_{0}^{2}+\left\|\bar{p}^{j}-p_{h}^{j}(\hat{m}_{p})\right\|_{0}^{2}\right)\right)^{1/2} \le C(h+\Delta t). \tag*{(121)} ( Δ t j = 1 ∑ N ( u ˉ j − u h j ( m ^ p ) 1 2 + p ˉ T j − p T , h j ( m ^ p ) 0 2 + p ˉ j − p h j ( m ^ p ) 0 2 ) ) 1/2 ≤ C ( h + Δ t ) . (121)
The second term is the control-perturbation error; by the ℓ t 2 \ell_{t}^{2} ℓ t 2 estimates of Lemma 5.1 ,
( Δ t ∑ j = 1 N ( ∥ u h j ( m ^ p ) − u ^ h j ∥ V 2 + ∥ p T , h j ( m ^ p ) − p ^ T , h j ∥ 0 2 + ∥ p h j ( m ^ p ) − p ^ h j ∥ 0 2 ) ) 1 / 2 ≤ C ( h + Δ t ) , \left(\Delta t \sum_{j=1}^{N} \left(\left\|u_{h}^{j}(\hat{m}_{p})-\hat{u}_{h}^{j}\right\|_{V}^{2}+\left\|p_{T,h}^{j}(\hat{m}_{p})-\hat{p}_{T,h}^{j}\right\|_{0}^{2}+\left\|p_{h}^{j}(\hat{m}_{p})-\hat{p}_{h}^{j}\right\|_{0}^{2}\right)\right)^{1/2} \le C(h+\Delta t), ( Δ t j = 1 ∑ N ( u h j ( m ^ p ) − u ^ h j V 2 + p T , h j ( m ^ p ) − p ^ T , h j 0 2 + p h j ( m ^ p ) − p ^ h j 0 2 ) ) 1/2 ≤ C ( h + Δ t ) ,
where the last step uses Theorem 5.12. Combining with (121) gives (120) . □
Remark 5.15. The fixed-control state estimate of Lemma 5.4 achieves the rate O ( h + Δ t ) O(h+\Delta t) O ( h + Δ t ) . This follows because, with time-independent coefficients and the cell-average formulation, the stiffness residual R S k + 1 ( q h ) = ( κ p ∇ ( I − P h ) p ˉ k + 1 , ∇ q h ) Ω R_{S}^{k+1}(q_{h})=(\kappa_{p}\nabla(I-P_{h})\bar{p}^{k+1},\nabla q_{h})_{\Omega} R S k + 1 ( q h ) = ( κ p ∇ ( I − P h ) p ˉ k + 1 , ∇ q h ) Ω is purely spatial, giving O ( h ) O(h) O ( h ) . The mass residual and the commutator are estimated jointly: after summation by parts they contribute O ( h ) O(h) O ( h ) through Lemma 5.6 and O ( Δ t ) O(\Delta t) O ( Δ t ) through the cell-average defect d n d^{n} d n .
The final control and optimal-state estimates (Theorems 5.12–5.14 ) achieve the same rate O ( h + Δ t ) O(h+\Delta t) O ( h + Δ t ) . This follows because the adjoint fixed-control estimate of Lemma 5.11 is controlled entirely through the O ( h + Δ t ) O(h+\Delta t) O ( h + Δ t ) fixed-control state estimate, applied both to the forward-in-time system and, after time reversal, to the reduced scalar adjoint system, which has the identical symmetric structure.
Numerical experiments
We examine convergence of the fully discrete optimality system, accuracy as the material parameters vary, and a localized pressure-control problem. All results in this section were computed with NGSolve 6.2.2608 [23 ] . The error estimates are uniform in λ \lambda λ for storage bounded below by a positive constant, while the Schur-complement formulation remains well posed at s 0 = 0 s_{0}=0 s 0 = 0 for each fixed finite λ \lambda λ . The experiments examine both regimes, including parameter combinations for which the analysis does not establish simultaneous uniformity. The implementation, input configurations, and raw results are supplied in the accompanying supplementary code archive.
Throughout, Ω = ( 0 , 1 ) 2 \Omega=(0,1)^{2} Ω = ( 0 , 1 ) 2 , T = 1 T=1 T = 1 , μ = α = 1 \mu=\alpha=1 μ = α = 1 , and κ p = 0.1 \kappa_{p}=0.1 κ p = 0.1 . We impose u = 0 u=0 u = 0 on Γ D = { 0 } × ( 0 , 1 ) \Gamma_{D}=\{0\}\times(0,1) Γ D = { 0 } × ( 0 , 1 ) , zero traction on the remaining three sides, and p = 0 p=0 p = 0 on the whole boundary. Thus ∣ Γ N ∣ > 0 |\Gamma_{N}|>0 ∣ Γ N ∣ > 0 and the total-pressure space is the full continuous P 1 \mathbb{P}_{1} P 1 space, without a mean constraint or pressure pinning. The dG(0) time treatment of bounded controls is related to the fully discrete optimal-control analysis in [4 ] . Each square of a uniform n × n n\times n n × n grid is divided into two triangles, so that h = 2 / n h=\sqrt{2}/n h = 2 / n . The state and adjoint use the P 2 / P 1 / P 1 \mathbb{P}_{2}/\mathbb{P}_{1}/\mathbb{P}_{1} P 2 / P 1 / P 1 spaces of Section 4.1; controls are constant on each triangle and time slab. We vary n n n and N N N independently in the convergence tests; each table states both resolutions. The localized example uses n = N = 16 n=N=16 n = N = 16 and 32. Both tracking weights are one.
Convergence verification
We use a manufactured optimum that requires no additional fluid source. Define
B ( x , y ) = 4096 x 3 ( 1 − x ) 3 y 3 ( 1 − y ) 3 , w = 1 20 ( ∂ y B , − ∂ x B ) , B(x,y)=4096x^3(1-x)^3y^3(1-y)^3,\qquad w=\frac{1}{20}(\partial_yB,-\partial_xB), B ( x , y ) = 4096 x 3 ( 1 − x ) 3 y 3 ( 1 − y ) 3 , w = 20 1 ( ∂ y B , − ∂ x B ) ,
η ( t ) = t 2 ( 1 − t ) 3 , β ( t ) = ( 1 − t ) 2 . \eta(t)=t^2(1-t)^3,\qquad\beta(t)=(1-t)^2. η ( t ) = t 2 ( 1 − t ) 3 , β ( t ) = ( 1 − t ) 2 .
With γ p = 0.1 \gamma_p=0.1 γ p = 0.1 , the exact state, control, and adjoint are
u = η w , p = η B , p T = α p , m p = s 0 η ′ B − κ p η Δ B , q ∗ = − γ p m p , q T ∗ = − α q ∗ , v ∗ = β w . (122) \begin{aligned}
u&=\eta w, & p&=\eta B, & p_T&=\alpha p,\\
m_p&=s_0\eta'B-\kappa_p\eta\Delta B, & q^*&=-\gamma_pm_p, & q_T^*&=-\alpha q^*,\\
v^*&=\beta w. \tag*{(122)}
\end{aligned} u m p v ∗ = η w , = s 0 η ′ B − κ p η Δ B , = β w . p q ∗ = η B , = − γ p m p , p T q T ∗ = α p , = − α q ∗ , (122)
Since div w = 0 \operatorname{div}w=0 div w = 0 , the total-pressure relations hold for all λ \lambda λ . The bubble and its derivatives ensure the homogeneous boundary conditions, including zero traction on Γ N \Gamma_N Γ N . The initial state and terminal adjoint are zero. The mechanical forcing in the differentiated state equation and the targets are set to
f = − div ( 2 μ ε ( u ˙ ) ) + ∇ p ˙ T , R u = div ( 2 μ ε ( v ∗ ) ) − ∇ q ˙ T ∗ , f=-\operatorname{div}(2\mu\varepsilon(\dot{u}))+\nabla\dot{p}_T,\qquad R_u=\operatorname{div}(2\mu\varepsilon(v^*))-\nabla\dot{q}_T^*, f = − div ( 2 μ ε ( u ˙ )) + ∇ p ˙ T , R u = div ( 2 μ ε ( v ∗ )) − ∇ q ˙ T ∗ ,
u C = u − R u , R p = − s 0 q ˙ ∗ − κ p Δ q ∗ , p C = p − R p . u_C=u-R_u,\qquad R_p=-s_0\dot{q}^*-\kappa_p\Delta q^*,\qquad p_C=p-R_p. u C = u − R u , R p = − s 0 q ˙ ∗ − κ p Δ q ∗ , p C = p − R p .
The two α 2 / λ \alpha^2/\lambda α 2 / λ terms in the scalar adjoint equation cancel for this family. Thus the state equations, adjoint equations, and γ p m p + q ∗ = 0 \gamma_pm_p+q^*=0 γ p m p + q ∗ = 0 hold simultaneously. Strict convexity identifies (100) as the unique optimum. We solve the unconstrained problem here, equivalently an inactive sufficiently wide box problem. The exact fields and targets remain bounded independently of λ \lambda λ . This divergence-free family isolates parameter effects but does not sample general volumetric deformation.
For state variables, the reported errors are
E u 2 = Δ t ∑ k ∥ u ˉ k − u h k ∥ H 1 2 , E T 2 = Δ t ∑ k ∥ p ˉ T k − p T , h k ∥ 0 2 , E_u^2=\Delta t\sum_k\lVert\bar{u}^k-u_h^k\rVert_{H^1}^2,\qquad E_T^2=\Delta t\sum_k\lVert\bar{p}_T^k-p_{T,h}^k\rVert_{0}^2, E u 2 = Δ t k ∑ ∥ u ˉ k − u h k ∥ H 1 2 , E T 2 = Δ t k ∑ ∥ p ˉ T k − p T , h k ∥ 0 2 ,
E p 2 = Δ t ∑ k ∥ p ˉ k − p h k ∥ 0 2 , E m = ∥ m p − m p , h ∥ L t 2 L x 2 . E_p^2=\Delta t\sum_k\lVert\bar{p}^k-p_h^k\rVert_{0}^2,\qquad E_m=\lVert m_p-m_{p,h}\rVert_{L_t^2L_x^2}. E p 2 = Δ t k ∑ ∥ p ˉ k − p h k ∥ 0 2 , E m = ∥ m p − m p , h ∥ L t 2 L x 2 .
We set E s t a t e = ( E u 2 + E T 2 + E p 2 ) 1 / 2 E_{\mathrm{state}}=(E_u^2+E_T^2+E_p^2)^{1/2} E state = ( E u 2 + E T 2 + E p 2 ) 1/2 . The adjoint errors E v ∗ E_{v^*} E v ∗ , E q T ∗ E_{q_T^*} E q T ∗ , and E q ∗ E_{q^*} E q ∗ are true space–time norms in H 1 H^1 H 1 , L 2 L^2 L 2 , and L 2 L^2 L 2 , respectively, with z h k z_h^k z h k held constant on ( t k , t k + 1 ] (t_k,t_{k+1}] ( t k , t k + 1 ] . In particular, both the control and adjoint errors include the within-slab variation of the exact solution. The coupled refinements n = N = 8 , 16 , 32 , 64 n=N=8,16,32,64 n = N = 8 , 16 , 32 , 64 are retained in the supplementary data. Here we separate the spatial and temporal refinements to examine the contributions to the bounds in Theorems 5.12 and 5.14 .
Spatial convergence. To separate the spatial and temporal effects, we keep N = 2048 N=2048 N = 2048 fixed and use n = 8 , 16 , 32 n=8,16,32 n = 8 , 16 , 32 , with λ = s 0 = 1 \lambda=s_0=1 λ = s 0 = 1 . All errors in Table 1 are measured against the analytic optimum, with the same state cell-average and true space–time norms as above. Spatial rates are r x = log 2 ( E ( n , N ) / E ( 2 n , N ) ) r_x=\log_2(E(n,N)/E(2n,N)) r x = log 2 ( E ( n , N ) / E ( 2 n , N )) with N N N unchanged. The time-resolution check repeats the finest spatial level with N = 4096 N=4096 N = 4096 . The measured changes in all seven errors are reported below together with the parameter study; they check residual temporal contamination rather than assume it is negligible from the size of Δ t \Delta t Δ t alone.
Table 1.
Temporal convergence. At fixed n = 32 n=32 n = 32 , we use N = 8 , 16 , 32 , 64 , 128 N=8,16,32,64,128 N = 8 , 16 , 32 , 64 , 128 . Spatial discretization produces an error floor in the comparison with the analytic optimum. For example, a separate n = 64 n=64 n = 64 pilot gives analytic control errors 2.270 × 10 − 3 2.270\times10^{-3} 2.270 × 1 0 − 3 and 1.582 × 10 − 3 1.582\times10^{-3} 1.582 × 1 0 − 3 at N = 64 N=64 N = 64 and 128, respectively, an apparent temporal rate of only 0.52. To measure temporal error separately, we compare each
Table 1 . Spatial refinement at fixed N = 2048 N = 2048 N = 2048 , λ = s 0 = 1 \lambda= s_0 = 1 λ = s 0 = 1 : errors relative to the manufactured optimum. Rates in parentheses compare consecutive spatial levels.
solution with a reference optimum on the same spatial mesh, computed with N r e f = 4096 N_{\mathrm{ref}} = 4096 N ref = 4096 . These reference errors approximate errors relative to the continuous-in-time, spatially discrete optimum; they are distinct from the total errors against (100). For r = N r e f / N r = N_{\mathrm{ref}}/N r = N ref / N , the displacement reference error is
( E u t ) 2 = Δ t ∑ k = 1 N ∥ u h , N k − 1 r ∑ j = ( k − 1 ) r + 1 k r u h , N r e f j ∥ H 1 2 . (E_u^t)^2 = \Delta t \sum_{k=1}^{N} \left\lVert u_{h,N}^{k} - \frac{1}{r} \sum_{j=(k-1)r+1}^{kr} u_{h,N_{\mathrm{ref}}}^{j} \right\rVert_{H^1}^{2}. ( E u t ) 2 = Δ t k = 1 ∑ N u h , N k − r 1 j = ( k − 1 ) r + 1 ∑ k r u h , N ref j H 1 2 .
The other state reference errors use the corresponding L 2 L^2 L 2 norms. Control and adjoint reference errors integrate the difference of their piecewise-constant reconstructions on the common fine time grid, retaining variation within the coarse slabs. For example, E m t = ∥ m h , N − m h , N r e f ∥ L t 2 L x 2 E_m^t = \lVert m_{h,N} - m_{h,N_{\mathrm{ref}}} \rVert_{L_t^2L_x^2} E m t = ∥ m h , N − m h , N ref ∥ L t 2 L x 2 . Rates are r t = log 2 ( E t ( N ) / E t ( 2 N ) ) r_t = \log_2(E^t(N)/E^t(2N)) r t = log 2 ( E t ( N ) / E t ( 2 N )) at fixed n n n and N r e f N_{\mathrm{ref}} N ref .
Table 2 . Temporal refinement at fixed n = 32 n = 32 n = 32 , λ = s 0 = 1 \lambda= s_0 = 1 λ = s 0 = 1 : reference errors E t E^t E t using N r e f = 4096 N_{\mathrm{ref}} = 4096 N ref = 4096 on the same spatial mesh. Rates are in parentheses. These quantities approximate the temporal discretization errors relative to the continuous-in-time, spatially discrete optimum.
Table 2.
We repeat the reference comparison with N r e f = 2048 N_{\mathrm{ref}} = 2048 N ref = 2048 to quantify reference sensitivity. We also repeat the temporal refinement on the n = 16 n = 16 n = 16 mesh, with references computed on that same mesh. Figure 1 compares both temporal sequences, so that the observed temporal orders can be assessed on two fixed spatial discretizations. The analytic-error and reference-error data are both included in the supplementary results.
Figure 1. Independent refinements at λ = s 0 = 1 \lambda=s_0=1 λ = s 0 = 1 . Left: errors against the analytic optimum at fixed N = 2048 N=2048 N = 2048 . Right: temporal reference errors at fixed n = 16 n=16 n = 16 and 32 32 32 , each using its own same-mesh N r e f = 4096 N_{\mathrm{ref}}=4096 N ref = 4096 optimum.
Robustness with respect to λ \lambda λ and s 0 s_0 s 0
We next solve the manufactured optimal control problem for all pairs
λ ∈ { 1 , 10 2 , 10 4 , 10 6 } , s 0 ∈ { 1 , 10 − 2 , 10 − 4 , 0 } , \lambda\in\{1,10^2,10^4,10^6\}, \qquad s_0 \in\{1,10^{-2},10^{-4},0\}, λ ∈ { 1 , 1 0 2 , 1 0 4 , 1 0 6 } , s 0 ∈ { 1 , 1 0 − 2 , 1 0 − 4 , 0 } ,
using the same spatial sequence n = 8 , 16 , 32 n = 8,16,32 n = 8 , 16 , 32 at fixed N = 2048 N = 2048 N = 2048 and temporal sequence N = 8 , 16 , 32 , 64 , 128 N = 8,16,32,64,128 N = 8 , 16 , 32 , 64 , 128 at fixed n = 32 n = 32 n = 32 . Each parameter pair has its own same-mesh N r e f = 4096 N_{\mathrm{ref}} = 4096 N ref = 4096 reference optimum, with the N r e f = 2048 N_{\mathrm{ref}}=2048 N ref = 2048 comparison described above. The data are reconstructed from (100) for each s 0 s_0 s 0 ; for fixed s 0 s_0 s 0 , neither the exact solution nor the targets change with λ \lambda λ .
For each fixed positive storage coefficient, comparison across λ \lambda λ tests the nearly incompressible regime covered by the λ \lambda λ -uniform error analysis. Comparison at fixed λ \lambda λ tests the approach to zero storage. The runs with s 0 = 0 s_0=0 s 0 = 0 use the same matrices, boundary conditions, and optimization procedure: no positive storage floor or pressure penalty is inserted. Table 3 reports absolute error magnitudes and the final spatial and temporal rates separately. Here the superscript x x x denotes the analytic error at fixed fine Δ t \Delta t Δ t , while the superscript t t t denotes the temporal reference error at fixed h h h . The combined state error uses the three state norms defined above.
Table 3. Independent parameter study. Spatial errors use ( n , N ) = ( 32 , 2048 ) (n,N)=(32,2048) ( n , N ) = ( 32 , 2048 ) ; spatial rates compare n = 16 , 32 n=16,32 n = 16 , 32 at fixed N = 2048 N=2048 N = 2048 . Temporal reference errors use ( n , N ) = ( 32 , 128 ) (n,N)=(32,128) ( n , N ) = ( 32 , 128 ) and N r e f = 4096 N_{\mathrm{ref}}=4096 N ref = 4096 ; temporal rates compare N = 64 , 128 N=64,128 N = 64 , 128 at fixed n = 32 n=32 n = 32 .
The final spatial control rates across the parameter grid are 1.03–1.04, and the combined state rates are 1.94–1.95. The smooth state displays higher spatial orders than the first-order control, consistent with the different approximation spaces; this does not strengthen the theorem for minimally regular solutions. The final temporal reference rates are 1.00–1.02 for the control and 1.02 for the combined state.
For λ = s 0 = 1 \lambda=s_0=1 λ = s 0 = 1 , replacing N = 2048 N=2048 N = 2048 by 4096 at n = 32 n=32 n = 32 changes the seven analytic errors by at most 2.72%, while doubling the temporal reference changes the temporal errors by at most 3.26%. Across the parameter grid, at fixed n = 32 n=32 n = 32 , replacing N = 2048 N=2048 N = 2048 by 4096 changes the analytic errors by at most 10.17% across all seven variables and sixteen parameter pairs. Replacing the temporal reference N r e f = 2048 N_{\mathrm{ref}}=2048 N ref = 2048 by 4096 changes the reported temporal errors by at most 3.26% over all five temporal levels and all parameter pairs. These are measured resolution-sensitivity checks, not rigorous bounds on the remaining reference error.
For the control and combined state quantities displayed in Table 3 , the respective spatial and temporal sensitivity maxima are 1.20% and 3.23%. The largest individual-variable sensitivity concerns total pressure in the nearly incompressible cases. All component sensitivities are retained in the supplementary data; we do not infer equally resolved individual-variable rates for every parameter pair.
The nearly incompressible and zero-storage cases retain the observed control and combined-state orders in this manufactured family.
A separate algebraic check makes the limitation of this observation explicit. On the n = 8 n=8 n = 8 mesh, we assemble the discrete effective-mass matrix corresponding to m e f f , h m_{\mathrm{eff},h} m eff , h in (58) by eliminating the displacement and the total pressure variables, and compute its smallest generalized eigenvalue relative to the fluid-pressure L 2 L^2 L 2 mass matrix. At s 0 = 0 s_0=0 s 0 = 0 , its smallest eigenvalue is approximately 2.9595 × 10 − 1 2.9595 \times10^{-1} 2.9595 × 1 0 − 1 , 9.7625 × 10 − 3 9.7625 \times10^{-3} 9.7625 × 1 0 − 3 , 9.9976 × 10 − 5 9.9976 \times10^{-5} 9.9976 × 1 0 − 5 , and 1.0000 × 10 − 6 1.0000 \times10^{-6} 1.0000 × 1 0 − 6 for the four values of λ \lambda λ above. The effective mass is positive for every tested finite λ \lambda λ , while its coercivity decreases as λ \lambda λ grows. Consequently, the bounded errors in this smooth, divergence-free manufactured family do not imply uniform coercivity, or a uniform error theorem, in the simultaneous limit s 0 ↓ 0 s_0 \downarrow0 s 0 ↓ 0 , λ → ∞ \lambda\to\infty λ → ∞ .
A localized pressure-control example
Finally, we consider a problem without a prescribed analytic optimum. The body force and initial state vanish, u C = 0 u_C=0 u C = 0 , and
p C ( t , x , y ) = 0.15 t ( 2 − t ) sin ( π x ) sin ( π y ) . p_C(t,x,y)=0.15\,t(2-t)\sin(\pi x)\sin(\pi y). p C ( t , x , y ) = 0.15 t ( 2 − t ) sin ( π x ) sin ( π y ) .
The source is restricted to the mesh-aligned actuator ω = ( 1 / 4 , 3 / 4 ) 2 \omega=(1/4,3/4)^2 ω = ( 1/4 , 3/4 ) 2 : admissible controls satisfy 0 ≤ m p ≤ 0.4 0\le m_p\le0.4 0 ≤ m p ≤ 0.4 on ( 0 , 1 ) × ω (0,1)\times\omega ( 0 , 1 ) × ω and m p = 0 m_p=0 m p = 0 outside it. Equivalently, the distributed source is extended by zero from ω \omega ω ; the same discrete projection uses equal zero bounds on the other cells. This spatially restricted admissible set extends the constant-box setting of Section 3. The control may jump across ∂ ω \partial\omega ∂ ω , so this example illustrates the discrete optimality system without assuming global H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) control regularity or claiming the rate of Theorem 5.12. We choose λ = 10 4 \lambda=10^4 λ = 1 0 4 , s 0 = 0 s_0=0 s 0 = 0 , and γ p = 10 − 3 \gamma_p=10^{-3} γ p = 1 0 − 3 . All remaining coefficients and boundary conditions are unchanged. This finite-parameter example combines nearly incompressible mechanics, zero storage, and an active upper control bound.
Table 4 reports the objective, tracking norm, control norm, and upper-active fraction for two mesh/time resolutions. The reported objective uses the continuous target norm; it differs from the slab-average-target objective in (56) only by a known, control-independent constant. The two objectives therefore have the same minimizer and gradient. With zero control, the state is zero, which supplies the uncontrolled baseline J ( 0 ) = 1.5 × 10 − 3 J(0)=1.5\times10^{-3} J ( 0 ) = 1.5 × 1 0 − 3 . At n = N = 32 n=N=32 n = N = 32 , the objective is reduced by 98.72%. The optimized source reduces the tracking discrepancy, while its limited support, the upper bound, and regularization prevent exact matching. The active fraction is measured relative to the whole space–time domain; fixed-zero cells outside the actuator are excluded from the active-set statistics.
Table 4. Localized source control. The tracking norm includes displacement and pressure. Active fractions refer to the whole space–time domain and exclude fixed-zero cells.
Figure 2. Localized control at n = N = 32 n=N=32 n = N = 32 : last-slab source, terminal pressure, target at T = 1 T=1 T = 1 , and displacement magnitude. The source is shown with its exact constant value on each triangle, so individual elements remain visible near the center. The displacement is resolved with its own color scale because it is small in this nearly incompressible regime.
Appendix A. Regularity of the optimal state and adjoint
This appendix gives sufficient conditions for Assumption R.
The regularity package established below is
u ^ , v ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; V ) , (123) \hat{u},\hat{v}^{*}\in L^{2}(0,T;H^{2}(\Omega))\cap H^{1}(0,T;V), \tag*{(123)} u ^ , v ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; V ) , (123)
p ^ T , q ^ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) ∩ H 1 ( 0 , T ; Q T ) , (124) \hat{p}_{T},\hat{q}_{T}^{*}\in L^{2}(0,T;H^{1}(\Omega))\cap H^{1}(0,T;Q_{T}), \tag*{(124)} p ^ T , q ^ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω )) ∩ H 1 ( 0 , T ; Q T ) , (124)
p ^ , q ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) . (125) \hat{p},\hat{q}^{*}\in L^{2}(0,T;H^{2}(\Omega))\cap H^{1}(0,T;L^{2}(\Omega)). \tag*{(125)} p ^ , q ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; L 2 ( Ω )) . (125)
Here Q T = L 2 ( Ω ) Q_{T}=L^{2}(\Omega) Q T = L 2 ( Ω ) . In particular, (103) does not assert p ^ , q ^ ∗ ∈ H 1 ( 0 , T ; Q ) \hat{p},\hat{q}^{*}\in H^{1}(0,T;Q) p ^ , q ^ ∗ ∈ H 1 ( 0 , T ; Q ) ; such a statement would require additional compatibility assumptions on the time-differentiated system and is not needed for the dG(0) error estimate.
The basic state and adjoint time-regularity below follows from the Schur complement estimate for every s 0 ≥ 0 s_{0}\geq0 s 0 ≥ 0 . When these regularity bounds are combined with the λ \lambda λ -uniform error estimates of § 5 , we retain the positive-storage regime stated there.
We also assume the following scalar elliptic and Stokes-type regularity estimates.
Scalar H 2 H^{2} H 2 -regularity. For every F ∈ L 2 ( Ω ) F\in L^{2}(\Omega) F ∈ L 2 ( Ω ) , the solution w ∈ H 0 1 ( Ω ) w\in H^{1}_{0}(\Omega) w ∈ H 0 1 ( Ω ) of − div ( κ ∇ w ) = F -\operatorname{div}(\kappa\nabla w)=F − div ( κ ∇ w ) = F satisfies ∥ w ∥ H 2 ( Ω ) ≤ C e l l ∥ F ∥ L 2 ( Ω ) \lVert w\rVert_{H^{2}(\Omega)}\leq C_{\mathrm{ell}}\lVert F\rVert_{L^{2}(\Omega)} ∥ w ∥ H 2 ( Ω ) ≤ C ell ∥ F ∥ L 2 ( Ω ) . This holds for convex polygonal/polyhedral Ω \Omega Ω or for ∂ Ω ∈ C 1 , 1 \partial\Omega\in C^{1,1} ∂ Ω ∈ C 1 , 1 , for constant κ > 0 \kappa>0 κ > 0 or for sufficiently smooth, uniformly elliptic coefficients. It is applied with κ = κ p \kappa=\kappa_{p} κ = κ p .
Stokes-type H 2 × H 1 H^{2}\times H^{1} H 2 × H 1 -regularity. If ( w , r ) ∈ V × Q T (w,r)\in V\times Q_{T} ( w , r ) ∈ V × Q T solves a u ( w , v ) + b ( r , v ) = ( F , v ) Ω a_{u}(w,v)+b(r,v)=(F,v)_{\Omega} a u ( w , v ) + b ( r , v ) = ( F , v ) Ω and b ( q T , w ) − λ − 1 ( r , q T ) Ω = ( ζ , q T ) Ω b(q_{T},w)-\lambda^{-1}(r,q_{T})_{\Omega}=(\zeta,q_{T})_{\Omega} b ( q T , w ) − λ − 1 ( r , q T ) Ω = ( ζ , q T ) Ω for all ( v , q T ) ∈ V × Q T (v,q_{T})\in V\times Q_{T} ( v , q T ) ∈ V × Q T , then ∥ w ∥ H 2 ( Ω ) + ∥ r ∥ H 1 ( Ω ) ≤ C S t ( ∥ F ∥ L 2 ( Ω ) + ∥ ζ ∥ H 1 ( Ω ) ) . \lVert w\rVert_{H^{2}(\Omega)}+\lVert r\rVert_{H^{1}(\Omega)}\leq C_{\mathrm{St}}(\lVert F\rVert_{L^{2}(\Omega)}+\lVert\zeta\rVert_{H^{1}(\Omega)}). ∥ w ∥ H 2 ( Ω ) + ∥ r ∥ H 1 ( Ω ) ≤ C St (∥ F ∥ L 2 ( Ω ) + ∥ ζ ∥ H 1 ( Ω ) ) . We assume this estimate holds for the particular domain Ω \Omega Ω , boundary conditions, and Lamé parameters under consideration, with C S t C_{\mathrm{St}} C St independent of
λ \lambda λ for λ ≥ λ 0 > 0 \lambda\ge\lambda_0 > 0 λ ≥ λ 0 > 0 ; whether it holds on a given convex polyhedral domain with mixed boundary conditions depends on additional geometric and compatibility conditions. Note that ζ \zeta ζ enters through an H 1 H^1 H 1 -norm because it plays the role of a divergence datum; L 2 L^2 L 2 -control of ζ \zeta ζ alone is not sufficient for H 1 H^1 H 1 -regularity of r r r .
The argument proceeds in three steps: (i) basic energy regularity of the state, giving in particular p ˙ ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) \dot{p} \in L^2(0,T;L^2(\Omega)) p ˙ ∈ L 2 ( 0 , T ; L 2 ( Ω )) and p ∈ L ∞ ( 0 , T ; Q ) p \in L^\infty(0,T;Q) p ∈ L ∞ ( 0 , T ; Q ) (Proposition A.1); (ii) spatial H 2 × H 1 H^2 \times H^1 H 2 × H 1 -regularity of ( u , p T ) (u,p_T) ( u , p T ) via the Stokes-type estimate applied to the integrated mechanical equation (Proposition A.2); and (iii) spatial H 2 H^2 H 2 -regularity of p p p via the scalar elliptic estimate applied to the diffusion equation (Proposition A.4). The adjoint is handled in Proposition A.6 by time-reversal, applying steps (i)–(iii) to the reversed system and then reading off the spatial regularity.
Basic energy regularity of the state
Proposition A.1. Let f : = f ~ ˙ f := \dot{\tilde{f}} f := f ~ ˙ . Assume
f ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) , g p ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) , (126) f \in L^2(0,T;L^2(\Omega)^d), \qquad g_p \in L^2(0,T;L^2(\Omega)), \tag*{(126)} f ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) , g p ∈ L 2 ( 0 , T ; L 2 ( Ω )) , (126) and compatible initial data ( u 0 , p 0 ) ∈ V × Q (u_0,p_0) \in V \times Q ( u 0 , p 0 ) ∈ V × Q for the total pressure formulation. Let ( u , p T , p ) (u,p_T,p) ( u , p T , p ) be the weak solution of (11). Then
u ∈ H 1 ( 0 , T ; V ) , p T ∈ H 1 ( 0 , T ; Q T ) , u \in H^1(0,T;V), \qquad p_T \in H^1(0,T;Q_T), u ∈ H 1 ( 0 , T ; V ) , p T ∈ H 1 ( 0 , T ; Q T ) , p ∈ L ∞ ( 0 , T ; Q ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) , (127) p \in L^\infty(0,T;Q) \cap H^1(0,T;L^2(\Omega)), \tag*{(127)} p ∈ L ∞ ( 0 , T ; Q ) ∩ H 1 ( 0 , T ; L 2 ( Ω )) , (127) with
∥ u ˙ ∥ L t 2 V + ∥ p ˙ T ∥ L t 2 Q T + ∥ p ˙ ∥ L t 2 L 2 + ∥ p ∥ L t ∞ Q ≤ C e n , (128) \lVert\dot{u} \rVert_{L^2_t V}+\lVert\dot{p}_T \rVert_{L^2_t Q_T}+\lVert\dot{p} \rVert_{L^2_t L^2}+\lVert p \rVert_{L^\infty_t Q}\le C_{\mathrm{en}}, \tag*{(128)} ∥ u ˙ ∥ L t 2 V + ∥ p ˙ T ∥ L t 2 Q T + ∥ p ˙ ∥ L t 2 L 2 + ∥ p ∥ L t ∞ Q ≤ C en , (128) where C e n C_{\mathrm{en}} C en depends on the data and material parameters, is independent of h h h , Δ t \Delta t Δ t , and may depend on c e f f c_{\mathrm{eff}} c eff in (30) . Dependence on κ p \kappa_p κ p is through the weighted norms and elliptic regularity constants.
Proof. This is exactly the storage-robust estimate of Lemma 2.8. The mechanical Schur complement yields the coercive effective mass m e f f m_{\mathrm{eff}} m eff , which controls p ˙ \dot{p} p ˙ in L 2 ( 0 , T ; L 2 ( Ω ) ) L^2(0,T;L^2(\Omega)) L 2 ( 0 , T ; L 2 ( Ω )) even when s 0 = 0 s_0=0 s 0 = 0 . The mechanical reconstruction then gives the stated bounds for u ˙ \dot{u} u ˙ and p ˙ T \dot{p}_T p ˙ T . No estimate of p ˙ \dot{p} p ˙ in Q Q Q is obtained or needed. □ \square □
Spatial regularity of the subsystem of displacement and total pressure Proposition A.2. Assume the hypotheses of Proposition A.1. Assume additionally:
f ~ ∈ H 1 ( 0 , T ; L 2 ( Ω ) d ) \tilde{f} \in H^1(0,T;L^2(\Omega)^d) f ~ ∈ H 1 ( 0 , T ; L 2 ( Ω ) d ) . This assumption implies both f ~ ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) \tilde{f} \in L^2(0,T;L^2(\Omega)^d) f ~ ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) (used for the Stokes estimate) and that f ~ ( 0 ) ∈ L 2 ( Ω ) d \tilde{f}(0) \in L^2(\Omega)^d f ~ ( 0 ) ∈ L 2 ( Ω ) d is well-defined as a trace (used in the compatibility condition below).
The initial data satisfy the mechanical compatibility condition: with p T , 0 : = − λ div u 0 + α p 0 p_{T,0}:=-\lambda\operatorname{div}u_0+\alpha p_0 p T , 0 := − λ div u 0 + α p 0 ,
a u ( u 0 , v ) + b ( p T , 0 , v ) = ( f ~ ( 0 ) , v ) Ω ∀ v ∈ V . a_u(u_0,v)+b(p_{T,0},v)=(\tilde{f}(0),v)_\Omega\qquad\forall v\in V. a u ( u 0 , v ) + b ( p T , 0 , v ) = ( f ~ ( 0 ) , v ) Ω ∀ v ∈ V . Then
u ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) , p T ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) . (129) u \in L^2(0,T;H^2(\Omega)), \qquad p_T \in L^2(0,T;H^1(\Omega)). \tag*{(129)} u ∈ L 2 ( 0 , T ; H 2 ( Ω )) , p T ∈ L 2 ( 0 , T ; H 1 ( Ω )) . (129) Proof. Integrating the mechanical equation of (11) in time from 0 0 0 to t t t and using the mechanical compatibility condition (assumption 2 above) gives, for a.e. t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
a u ( u ( t ) , v ) + b ( p T ( t ) , v ) = ( f ~ ( t ) , v ) Ω ∀ v ∈ V . a_u(u(t),v)+b(p_T(t),v)=(\tilde{f}(t),v)_\Omega\qquad\forall v\in V. a u ( u ( t ) , v ) + b ( p T ( t ) , v ) = ( f ~ ( t ) , v ) Ω ∀ v ∈ V . From the total pressure definition (5) ,
b ( q T , u ( t ) ) − λ − 1 ( p T ( t ) , q T ) Ω = − λ − 1 α ( p ( t ) , q T ) Ω ∀ q T ∈ Q T . b(q_T,u(t))-\lambda^{-1}(p_T(t),q_T)_\Omega=-\lambda^{-1}\alpha(p(t),q_T)_\Omega\qquad\forall q_T\in Q_T. b ( q T , u ( t )) − λ − 1 ( p T ( t ) , q T ) Ω = − λ − 1 α ( p ( t ) , q T ) Ω ∀ q T ∈ Q T .
The Stokes-type regularity estimate stated at the beginning of this appendix (assumption (2) in the enumerated list) therefore applies with
F ( t ) : = f ~ ( t ) , ζ ( t ) : = − λ − 1 α p ( t ) . F(t) := \tilde{f}(t), \qquad\zeta(t) := -\lambda^{-1}\alpha p(t). F ( t ) := f ~ ( t ) , ζ ( t ) := − λ − 1 α p ( t ) .
Since Proposition A.1 gives p ∈ L ∞ ( 0 , T ; H 1 ( Ω ) ) p \in L^\infty(0,T;H^1(\Omega)) p ∈ L ∞ ( 0 , T ; H 1 ( Ω )) , we have ζ ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) \zeta\in L^2(0,T;H^1(\Omega)) ζ ∈ L 2 ( 0 , T ; H 1 ( Ω )) , and
∥ u ( t ) ∥ H 2 + ∥ p T ( t ) ∥ H 1 ≤ C S t ( ∥ f ~ ( t ) ∥ L 2 + ∥ p ( t ) ∥ H 1 ) . \lVert u(t)\rVert_{H^2}+\lVert p_T(t)\rVert_{H^1}\le C_{\mathrm{St}}\bigl(\lVert\tilde{f}(t)\rVert_{L^2}+\lVert p(t)\rVert_{H^1}\bigr). ∥ u ( t ) ∥ H 2 + ∥ p T ( t ) ∥ H 1 ≤ C St ( ∥ f ~ ( t ) ∥ L 2 + ∥ p ( t ) ∥ H 1 ) .
Squaring and integrating over ( 0 , T ) (0,T) ( 0 , T ) yields (129) . The H 1 H^1 H 1 -in-time parts follow from Proposition A.1.
Remark A.3. The H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) -norm of ζ \zeta ζ in assumption (2) above is essential here: ζ ( t ) = − λ − 1 α p ( t ) \zeta(t)=-\lambda^{-1}\alpha p(t) ζ ( t ) = − λ − 1 α p ( t ) lies in H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) because p ∈ L ∞ ( 0 , T ; Q ) = L ∞ ( 0 , T ; H 0 1 ( Ω ) ) p\in L^\infty(0,T;Q)=L^\infty(0,T;H^1_0(\Omega)) p ∈ L ∞ ( 0 , T ; Q ) = L ∞ ( 0 , T ; H 0 1 ( Ω )) . If only ζ ∈ L 2 ( Ω ) \zeta\in L^2(\Omega) ζ ∈ L 2 ( Ω ) were available, the Stokes estimate would not deliver r ∈ H 1 ( Ω ) r\in H^1(\Omega) r ∈ H 1 ( Ω ) .
Spatial regularity of the diffusion block
Proposition A.4 (Scalar elliptic H 2 H^2 H 2 regularity for p p p ). Assume the hypotheses of Proposition A.1. Then
p ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) . (130) p\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;L^2(\Omega)). \tag*{(130)} p ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; L 2 ( Ω )) . (130)
Proof. For a.e. t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , the state equation (2) (rewritten in total pressure form) gives
− div ( κ p ∇ p ( t ) ) = g p ( t ) + λ − 1 α p ˙ T ( t ) − s p ˙ ( t ) = : F p ( t ) . -\operatorname{div}(\kappa_p\nabla p(t))=g_p(t)+\lambda^{-1}\alpha\dot{p}_T(t)-s\dot{p}(t)=:F_p(t). − div ( κ p ∇ p ( t )) = g p ( t ) + λ − 1 α p ˙ T ( t ) − s p ˙ ( t ) =: F p ( t ) .
By Proposition A.1, p ˙ T ∈ L t 2 Q T \dot{p}_T\in L^2_tQ_T p ˙ T ∈ L t 2 Q T , p ˙ ∈ L t 2 L x 2 \dot{p}\in L^2_tL^2_x p ˙ ∈ L t 2 L x 2 , and g p ∈ L t 2 L x 2 g_p\in L^2_tL^2_x g p ∈ L t 2 L x 2 by assumption (126) . Hence F p ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) F_p\in L^2(0,T;L^2(\Omega)) F p ∈ L 2 ( 0 , T ; L 2 ( Ω )) . The scalar H 2 H^2 H 2 -regularity estimate (assumption (1) above) gives, for a.e. t t t , ∥ p ( t ) ∥ H 2 ≤ C e l l ∥ F p ( t ) ∥ L 2 \lVert p(t)\rVert_{H^2}\le C_{\mathrm{ell}}\lVert F_p(t)\rVert_{L^2} ∥ p ( t ) ∥ H 2 ≤ C ell ∥ F p ( t ) ∥ L 2 . Integrating over ( 0 , T ) (0,T) ( 0 , T ) gives p ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) p\in L^2(0,T;H^2(\Omega)) p ∈ L 2 ( 0 , T ; H 2 ( Ω )) . The H 1 ( 0 , T ; L 2 ( Ω ) ) H^1(0,T;L^2(\Omega)) H 1 ( 0 , T ; L 2 ( Ω )) part was already established in Proposition A.1.
Regularity of the optimal state
Corollary A.5. Let m ^ p ∈ M p , a d \hat{m}_p\in\mathcal{M}_{p,\mathrm{ad}} m ^ p ∈ M p , ad be the optimal control and ( u ^ , p ^ T , p ^ ) (\hat{u},\hat{p}_T,\hat{p}) ( u ^ , p ^ T , p ^ ) the corresponding state. Assume (126) with g p = m ^ p g_p=\hat{m}_p g p = m ^ p . Assume also the hypotheses of Proposition A.2, namely f ~ ∈ H 1 ( 0 , T ; L 2 ( Ω ) d ) \tilde{f}\in H^1(0,T;L^2(\Omega)^d) f ~ ∈ H 1 ( 0 , T ; L 2 ( Ω ) d ) and the mechanical compatibility condition a u ( u 0 , v ) + b ( p T , 0 , v ) = ( f ~ ( 0 ) , v ) Ω a_u(u_0,v)+b(p_{T,0},v)=(\tilde{f}(0),v)_\Omega a u ( u 0 , v ) + b ( p T , 0 , v ) = ( f ~ ( 0 ) , v ) Ω for all v ∈ V v\in V v ∈ V , with p T , 0 : = − λ div u 0 + α p 0 p_{T,0}:=-\lambda\operatorname{div}u_0+\alpha p_0 p T , 0 := − λ div u 0 + α p 0 . Then
u ^ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; V ) , \hat{u}\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;V), u ^ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; V ) ,
p ^ T ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) ∩ H 1 ( 0 , T ; Q T ) , \hat{p}_T\in L^2(0,T;H^1(\Omega))\cap H^1(0,T;Q_T), p ^ T ∈ L 2 ( 0 , T ; H 1 ( Ω )) ∩ H 1 ( 0 , T ; Q T ) ,
p ^ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) . (131) \hat{p}\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;L^2(\Omega)). \tag*{(131)} p ^ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; L 2 ( Ω )) . (131)
Proof. Combine Propositions A.1, A.2, and A.4.
Regularity of the optimal adjoint
Proposition A.6. Assume the state regularity of Corollary A.5 and additionally
u C ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) , p C ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) . (132) u_C\in L^2(0,T;L^2(\Omega)^d),\qquad p_C\in L^2(0,T;L^2(\Omega)). \tag*{(132)} u C ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) , p C ∈ L 2 ( 0 , T ; L 2 ( Ω )) . (132)
Then the optimal adjoint ( v ^ ∗ , q ^ T ∗ , q ^ ∗ ) (\hat{v}^*,\hat{q}_T^*,\hat{q}^*) ( v ^ ∗ , q ^ T ∗ , q ^ ∗ ) satisfies
v ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; V ) , \hat{v}^*\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;V), v ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; V ) ,
q ^ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) ∩ H 1 ( 0 , T ; Q T ) , \hat{q}_T^*\in L^2(0,T;H^1(\Omega))\cap H^1(0,T;Q_T), q ^ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω )) ∩ H 1 ( 0 , T ; Q T ) ,
q ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) . (133) \hat{q}^*\in L^2(0,T;H^2(\Omega))\cap H^1(0,T;L^2(\Omega)). \tag*{(133)} q ^ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; L 2 ( Ω )) . (133)
Proof. Set s : = T − t s := T-t s := T − t and define the reversed adjoint variables
v ˇ ∗ ( s ) : = v ∗ ( T − s ) , q ˇ T ∗ ( s ) : = q T ∗ ( T − s ) , q ˇ ∗ ( s ) : = q ∗ ( T − s ) . \check{v}^{*}(s) := v^{*}(T-s), \qquad\check{q}_{T}^{*}(s) := q_{T}^{*}(T-s), \qquad\check{q}^{*}(s) := q^{*}(T-s). v ˇ ∗ ( s ) := v ∗ ( T − s ) , q ˇ T ∗ ( s ) := q T ∗ ( T − s ) , q ˇ ∗ ( s ) := q ∗ ( T − s ) .
The terminal condition (47) becomes the homogeneous initial condition ( v ˇ ∗ ( 0 ) , q ˇ T ∗ ( 0 ) , q ˇ ∗ ( 0 ) ) = ( 0 , 0 , 0 ) (\check{v}^{*}(0),\check{q}_{T}^{*}(0),\check{q}^{*}(0)) = (0,0,0) ( v ˇ ∗ ( 0 ) , q ˇ T ∗ ( 0 ) , q ˇ ∗ ( 0 )) = ( 0 , 0 , 0 ) , and the reversed adjoint system has the same coercive saddle-point and storage-diffusion structure as the forward state system (cf. the time-reversal argument in Section 3.2). The right-hand sides in s s s are
F ~ v ( s ) : = α u , C ( u ^ ( T − s ) − u C ( T − s ) ) ∈ L s 2 L x 2 , \widetilde{F}_{v}(s) := \alpha_{u,C}(\widehat{u}(T-s)-u_{C}(T-s)) \in L_{s}^{2}L_{x}^{2}, F v ( s ) := α u , C ( u ( T − s ) − u C ( T − s )) ∈ L s 2 L x 2 ,
F ~ q ( s ) : = α p , C ( p ^ ( T − s ) − p C ( T − s ) ) ∈ L s 2 L x 2 , \widetilde{F}_{q}(s) := \alpha_{p,C}(\widehat{p}(T-s)-p_{C}(T-s)) \in L_{s}^{2}L_{x}^{2}, F q ( s ) := α p , C ( p ( T − s ) − p C ( T − s )) ∈ L s 2 L x 2 ,
by (132) and the state regularity. Applying Proposition A.1 to the reversed system gives v ˇ ∗ ∈ H s 1 V \check{v}^{*} \in H_{s}^{1}V v ˇ ∗ ∈ H s 1 V , q ˇ T ∗ ∈ H s 1 Q T \check{q}_{T}^{*} \in H_{s}^{1}Q_{T} q ˇ T ∗ ∈ H s 1 Q T , q ˇ ∗ ∈ L s ∞ Q ∩ H s 1 L x 2 \check{q}^{*} \in L_{s}^{\infty}Q \cap H_{s}^{1}L_{x}^{2} q ˇ ∗ ∈ L s ∞ Q ∩ H s 1 L x 2 .
For spatial regularity of ( v ˇ ∗ , q ˇ T ∗ ) (\check{v}^{*},\check{q}_{T}^{*}) ( v ˇ ∗ , q ˇ T ∗ ) , integrating (44) from t t t to T T T and using the homogeneous terminal condition gives, for a.e. t t t ,
a u ( v ˇ ∗ ( t ) , v ) + b ( q ˇ T ∗ ( t ) , v ) = ( ∫ t T α u , C ( u ^ ( s ) − u C ( s ) ) d s , v ) Ω . a_{u}(\check{v}^{*}(t),v)+b(\check{q}_{T}^{*}(t),v)=\left(\int_{t}^{T}\alpha_{u,C}(\widehat{u}(s)-u_{C}(s))\,ds,v\right)_{\Omega}. a u ( v ˇ ∗ ( t ) , v ) + b ( q ˇ T ∗ ( t ) , v ) = ( ∫ t T α u , C ( u ( s ) − u C ( s )) d s , v ) Ω .
Integrating (45) from t t t to T T T gives b ( q T , v ˇ ∗ ( t ) ) − λ − 1 ( q ˇ T ∗ ( t ) , q T ) Ω = λ − 1 α q ( q ˇ ∗ ( t ) , q T ) Ω b(q_{T},\check{v}^{*}(t))-\lambda^{-1}(\check{q}_{T}^{*}(t),q_{T})_{\Omega}=\lambda^{-1}\alpha_{q}(\check{q}^{*}(t),q_{T})_{\Omega} b ( q T , v ˇ ∗ ( t )) − λ − 1 ( q ˇ T ∗ ( t ) , q T ) Ω = λ − 1 α q ( q ˇ ∗ ( t ) , q T ) Ω . The Stokes-type estimate applies with
F ∗ ( t ) : = ∫ t T α u , C ( u ^ ( s ) − u C ( s ) ) d s ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) , F^{*}(t):=\int_{t}^{T}\alpha_{u,C}(\widehat{u}(s)-u_{C}(s))\,ds\in L^{2}(0,T;L^{2}(\Omega)^{d}), F ∗ ( t ) := ∫ t T α u , C ( u ( s ) − u C ( s )) d s ∈ L 2 ( 0 , T ; L 2 ( Ω ) d ) ,
ζ ∗ ( t ) : = λ − 1 α q q ˇ ∗ ( t ) ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) , \zeta^{*}(t):=\lambda^{-1}\alpha_{q}\check{q}^{*}(t)\in L^{2}(0,T;H^{1}(\Omega)), ζ ∗ ( t ) := λ − 1 α q q ˇ ∗ ( t ) ∈ L 2 ( 0 , T ; H 1 ( Ω )) ,
where ζ ∗ ∈ L t 2 H x 1 \zeta^{*}\in L_{t}^{2}H_{x}^{1} ζ ∗ ∈ L t 2 H x 1 because q ˇ ∗ ∈ L t ∞ Q = L t ∞ H 0 1 ( Ω ) \check{q}^{*}\in L_{t}^{\infty}Q=L_{t}^{\infty}H_{0}^{1}(\Omega) q ˇ ∗ ∈ L t ∞ Q = L t ∞ H 0 1 ( Ω ) . Hence v ˇ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) \check{v}^{*}\in L^{2}(0,T;H^{2}(\Omega)) v ˇ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) and q ˇ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω ) ) \check{q}_{T}^{*}\in L^{2}(0,T;H^{1}(\Omega)) q ˇ T ∗ ∈ L 2 ( 0 , T ; H 1 ( Ω )) .
For spatial regularity of q ˇ ∗ \check{q}^{*} q ˇ ∗ , the adjoint diffusion equation (46) can be written, for a.e. t t t , as a scalar elliptic equation:
( κ p ∇ q ˇ ∗ ( t ) , ∇ q ) Ω = α p , C ( p ^ ( t ) − p C ( t ) , q ) Ω + λ − 1 α q ( q ˇ ∗ ( t ) , q ) Ω + s 1 ( q ˇ ∗ ( t ) , q ) Ω . (\kappa_{p}\nabla\check{q}^{*}(t),\nabla q)_{\Omega}=\alpha_{p,C}(\widehat{p}(t)-p_{C}(t),q)_{\Omega}+\lambda^{-1}\alpha_{q}(\check{q}^{*}(t),q)_{\Omega}+s_{1}(\check{q}^{*}(t),q)_{\Omega}. ( κ p ∇ q ˇ ∗ ( t ) , ∇ q ) Ω = α p , C ( p ( t ) − p C ( t ) , q ) Ω + λ − 1 α q ( q ˇ ∗ ( t ) , q ) Ω + s 1 ( q ˇ ∗ ( t ) , q ) Ω .
The right-hand side contains p ^ − p C ∈ L 2 L x 2 \widehat{p}-p_{C}\in L^{2}L_{x}^{2} p − p C ∈ L 2 L x 2 (by Corollary A.5 and (132) ) and q ˇ ˙ ∗ , q ˇ ∗ ∈ L 2 L x 2 \dot{\check{q}}^{*},\check{q}^{*}\in L^{2}L_{x}^{2} q ˇ ˙ ∗ , q ˇ ∗ ∈ L 2 L x 2 (from the energy estimate on the reversed system). The scalar H 2 H^{2} H 2 -regularity estimate gives q ˇ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω ) ) \check{q}^{*}\in L^{2}(0,T;H^{2}(\Omega)) q ˇ ∗ ∈ L 2 ( 0 , T ; H 2 ( Ω )) .
Since time-reversal is an isometry on all relevant Bochner spaces, the established regularity for the reversed variables translates directly to ( v ˇ ∗ , q ˇ T ∗ , q ˇ ∗ ) (\check{v}^{*},\check{q}_{T}^{*},\check{q}^{*}) ( v ˇ ∗ , q ˇ T ∗ , q ˇ ∗ ) , yielding (133) . □
Declarations
OpenAI’s ChatGPT was used to explore preliminary duality-based error estimates for the analysis in § 5 . The resulting mathematical arguments were checked and revised by the author. ChatGPT was also used to assist in writing Python scripts for the numerical experiments implemented with NGSolve. The author takes full responsibility for the mathematical analysis, numerical results, and all content of this manuscript.
References