Introduction
In many applications the quantity of interest (QoI) is a scalar functional of the solution of a partial differential equation, such as a drag coefficient, an average over a subdomain, or a flux through an interface. We want to know how large the error of the computed value is. In the standard Galerkin setting, with the Galerkin approximation u h ∈ X h u_h \in X_h u h ∈ X h , γ = m ( u ) \gamma= m(u) γ = m ( u ) and the adjoint solution z z z satisfying a ( s , z ) = m ( s ) a(s,z)=m(s) a ( s , z ) = m ( s ) for all admissible s s s ,
γ − m ( u h ) = a ( u − u h , z − ψ h ) ∀ ψ h ∈ X h . \gamma-m(u_h)=a(u-u_h,z-\psi_h)\qquad\forall\psi_h\in X_h. γ − m ( u h ) = a ( u − u h , z − ψ h ) ∀ ψ h ∈ X h .
The dual-weighted residual method [4 , 2 ] replaces z z z by a more accurate approximation of the adjoint solution, computed in an enriched space or reconstructed from the Galerkin approximation z h ∈ X h z_h\in X_h z h ∈ X h of z z z . The resulting signed estimate is accurate in practice but not guaranteed, and its elementwise terms serve as refinement indicators. The other common approach takes ψ h = z h \psi_h=z_h ψ h = z h and bounds the two factors by estimators of the primal and the adjoint error [33 , 29 ] , ∣ γ − m ( u h ) ∣ ≤ C η p η d |\gamma-m(u_h)|\leq C\eta_p\eta_d ∣ γ − m ( u h ) ∣ ≤ C η p η d . This bound is reliable and is the basis of the convergence theory of goal-oriented adaptive methods [28 , 3 , 18 , 20 ] , but it can be very pessimistic. In this paper we obtain both a reliable bound and a signed estimate from the same first-order system residuals.
The first paper of this series [35 ] considers the general second-order elliptic equation with diffusion coefficient A A A , convection coefficient b b b and reaction coefficient c c c , which is also our model problem. That paper writes the primal problem and a physically meaningful adjoint problem as first-order systems for a flux and a potential. The norms of their residuals, the least-squares estimators η p L S \eta_p^{\mathrm{LS}} η p LS and η d L S \eta_d^{\mathrm{LS}} η d LS , are equivalent to the primal and adjoint errors for arbitrary approximations [10 , 11 , 6 , 7 , 37 , 14 ] . The same paper uses two consistent evaluators of γ \gamma γ . The first, which we call the potential evaluator, is the corrected functional of Giles and Süli [19 ] , γ 0 ( w , v ) : = m ( w ) + ℓ ( v ) − a ( w , v ) \gamma_0(w,v) := m(w) + \ell(v) - a(w,v) γ 0 ( w , v ) := m ( w ) + ℓ ( v ) − a ( w , v ) , where ℓ \ell ℓ is the right-hand side of the primal problem. It satisfies γ − γ 0 ( w , v ) = a ( u − w , z − v ) \gamma- \gamma_0(w,v) = a(u-w,z-v) γ − γ 0 ( w , v ) = a ( u − w , z − v ) for all admissible w w w and v v v , so we do not need Galerkin orthogonality. The second, γ 1 \gamma_1 γ 1 , also uses the flux approximations that a least-squares method computes along with the potentials. It expresses the gradients in γ 0 \gamma_0 γ 0 through these fluxes and adds two residual corrections. Another way to combine least squares with a quantity of interest is to build the quantity into the least-squares functional itself [17 ] .
The product of the two estimators gives a reliable upper bound for the errors of both evaluators, and its elementwise contributions [35 ] are good refinement indicators. In our experiments, however, the product exceeds the larger of these errors by factors of up to several hundred. If we stop when the product falls below a tolerance, we refine far more than needed. This overestimate does not come from the estimators. For the Poisson model problem each of them is asymptotically exact at the least-squares solution [16 ] , so there the whole overestimate comes from bounding the error a ( u − w , z − v ) a(u-w,z-v) a ( u − w , z − v ) of γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) by the product of the two error norms. To decide when to stop, we need to know how large the error actually is.
For this we compare two approximations of the same quantity. We compute γ \gamma γ twice from the same primal and adjoint approximations ( τ , w ) (\tau,w) ( τ , w ) and ( ρ , v ) (\rho,v) ( ρ , v ) , once as γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) from the potentials and once as γ 1 ( w , τ , v , ρ ) \gamma_1(w,\tau,v,\rho) γ 1 ( w , τ , v , ρ ) with the gradients expressed through the fluxes. We do not compute a second, more accurate approximation. We show that the difference of the two values, which we call the signed QoI gap , is exactly the A − 1 A^{-1} A − 1 inner product of the primal and adjoint constitutive residuals,
γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) , \gamma_1(w,\tau,v,\rho)-\gamma_0(w,v) = \left(A^{-1}R_p^{\mathrm{cst}}(\tau,w),R_d^{\mathrm{cst}}(\rho,v)\right), γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v ) ) ,
with R p c s t ( τ , w ) = f 2 + τ + A ∇ w R_p^{\mathrm{cst}}(\tau,w) = f_2 + \tau+ A\nabla w R p cst ( τ , w ) = f 2 + τ + A ∇ w and R d c s t ( ρ , v ) = g 2 + ρ + A ∇ v + b v R_d^{\mathrm{cst}}(\rho,v) = g_2 + \rho+ A\nabla v + bv R d cst ( ρ , v ) = g 2 + ρ + A ∇ v + b v , where f 2 f_2 f 2 and g 2 g_2 g 2 are the vector-valued data of the primal problem and of the quantity of interest. The identity is algebraic and holds for arbitrary primal and adjoint approximations. When γ \gamma γ lies between the two values, the QoI gap is between one and two times the larger of the two errors. Otherwise the QoI gap is smaller than the larger error, and much smaller only when the two values have nearly the same error.
The dual-weighted residual estimate also computes γ \gamma γ twice, but it evaluates γ 0 \gamma_0 γ 0 the second time at a more accurate adjoint approximation. It is a good estimate when the second value is much more accurate than the first. The signed QoI gap is a generalized dual-weighted residual estimate with flux-based weights. We do not assume that either of its two values is the more accurate, and we need no more accurate approximation. The fluxes need not be more accurate than the gradients of the potentials either. The signed QoI gap and the product bound come from the same residuals. For the QoI gap we take the inner product of the primal and adjoint constitutive residuals, and for the bound we take the product of the norms of the primal and adjoint residuals.
Since − A − 1 R p c s t ( τ , w ) = ∇ ( u − w ) + A − 1 ( σ − τ ) -A^{-1}R_p^{\mathrm{cst}}(\tau,w) = \nabla(u-w) + A^{-1}(\sigma-\tau) − A − 1 R p cst ( τ , w ) = ∇ ( u − w ) + A − 1 ( σ − τ ) and − R d c s t ( ρ , v ) = A ∇ ( z − v ) + b ( z − v ) + ( r − ρ ) -R_d^{\mathrm{cst}}(\rho,v) = A\nabla(z-v) + b(z-v) + (r-\rho) − R d cst ( ρ , v ) = A ∇ ( z − v ) + b ( z − v ) + ( r − ρ ) , the QoI gap contains the error a ( u − w , z − v ) a(u-w,z-v) a ( u − w , z − v ) of γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) without its reaction term, plus terms in the flux errors. Where reaction dominates, the reaction term can dominate the errors of both evaluators, and the signed QoI gap does not contain it. For this case we introduce the reaction-restored evaluator γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) , built from the fluxes alone. The reaction-restored QoI gap γ 2 ( τ , ρ ) − γ 0 ( w , v ) \gamma_2(\tau,\rho)-\gamma_0(w,v) γ 2 ( τ , ρ ) − γ 0 ( w , v ) then estimates the error of γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) . We refine the mesh with the least-squares estimators and stop when the magnitude of the signed QoI gap, or of the reaction-restored QoI gap where reaction dominates, is below a tolerance. In our experiments the QoI gaps follow the errors closely, even where the product bound is very pessimistic.
The construction does not depend on the discretization. We can take the potentials from any method that produces conforming potentials, and the fluxes from least-squares flux recovery [14 , 24 ] or from a mixed method. An augmented mixed method [27 , 25 ] gives both, and so does a least-squares finite element method. Our experiments include potentials from finite difference and finite volume schemes, equilibrated mixed fluxes for Kellogg's problem, a comparison with a dual-weighted residual computation, and computations with reaction coefficients from c = 1 c=1 c = 1 to c = 10 8 c=10^8 c = 1 0 8 .
Section 2 recalls the setting and the evaluators of [35 ] . Section 3 derives the QoI gap identity, shows what the signed QoI gap estimates, and introduces γ 2 \gamma_2 γ 2 and the reaction-restored QoI gap. Section 4 relates the QoI gaps to the product bound, the dual-weighted residual method, flux recovery and, for symmetric problems, the primal–dual bounds of [36 ] . Section 5 describes how we compute the fluxes and how we refine the meshes, and Section 6 reports the numerical experiments.
Primal and adjoint first-order systems and two evaluators
The primal problem, the quantity of interest and the adjoint problem
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d , d = 2 , 3 d=2,3 d = 2 , 3 , be a bounded Lipschitz polyhedral domain, whose boundary is split into relatively open, disjoint parts Γ D \Gamma_{D} Γ D and Γ N \Gamma_{N} Γ N with ∂ Ω = Γ ‾ D ∪ Γ ‾ N \partial\Omega=\overline{\Gamma}_{D}\cup\overline{\Gamma}_{N} ∂ Ω = Γ D ∪ Γ N , and let n \mathbf{n} n be the outward unit normal. The coefficients are A ∈ L ∞ ( Ω ) d × d A\in L^{\infty}(\Omega)^{d\times d} A ∈ L ∞ ( Ω ) d × d , symmetric and uniformly positive definite, b ∈ L ∞ ( Ω ) d \mathbf{b}\in L^{\infty}(\Omega)^{d} b ∈ L ∞ ( Ω ) d and c ∈ L ∞ ( Ω ) c\in L^{\infty}(\Omega) c ∈ L ∞ ( Ω ) . The data of the primal problem are f 1 ∈ L 2 ( Ω ) f_{1}\in L^{2}(\Omega) f 1 ∈ L 2 ( Ω ) , f 2 ∈ L 2 ( Ω ) d \mathbf{f}_{2}\in L^{2}(\Omega)^{d} f 2 ∈ L 2 ( Ω ) d , ϕ p ∈ H 1 / 2 ( Γ D ) \phi_{p}\in H^{1/2}(\Gamma_{D}) ϕ p ∈ H 1/2 ( Γ D ) and ψ p ∈ L 2 ( Γ N ) \psi_{p}\in L^{2}(\Gamma_{N}) ψ p ∈ L 2 ( Γ N ) , and those of the quantity of interest are g 1 ∈ L 2 ( Ω ) g_{1}\in L^{2}(\Omega) g 1 ∈ L 2 ( Ω ) , g 2 ∈ L 2 ( Ω ) d \mathbf{g}_{2}\in L^{2}(\Omega)^{d} g 2 ∈ L 2 ( Ω ) d , ϕ d ∈ H 1 / 2 ( Γ D ) \phi_{d}\in H^{1/2}(\Gamma_{D}) ϕ d ∈ H 1/2 ( Γ D ) and ψ d ∈ L 2 ( Γ N ) \psi_{d}\in L^{2}(\Gamma_{N}) ψ d ∈ L 2 ( Γ N ) . We write ( ⋅ , ⋅ ) (\cdot,\cdot) ( ⋅ , ⋅ ) and ∥ ⋅ ∥ 0 \|\cdot\|_{0} ∥ ⋅ ∥ 0 for the inner product and norm of L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) , and ⟨ ⋅ , ⋅ ⟩ Γ \langle\cdot,\cdot\rangle_{\Gamma} ⟨ ⋅ , ⋅ ⟩ Γ for the L 2 ( Γ ) L^{2}(\Gamma) L 2 ( Γ ) inner product on Γ ⊂ ∂ Ω \Gamma\subset\partial\Omega Γ ⊂ ∂ Ω , or the duality pairing of H − 1 / 2 H^{-1/2} H − 1/2 and H 1 / 2 H^{1/2} H 1/2 there. The spaces are
H D 1 ( Ω ) : = { s ∈ H 1 ( Ω ) : s = 0 on Γ D } , H N ( div ; Ω ) : = { μ ∈ H ( div ; Ω ) : μ ⋅ n = 0 on Γ N } , X p : = { w ∈ H 1 ( Ω ) : w = ϕ p on Γ D } , Σ p : = { τ ∈ H ( div ; Ω ) : τ ⋅ n = ψ p on Γ N } , X d : = { v ∈ H 1 ( Ω ) : v = ϕ d on Γ D } , Σ d : = { ρ ∈ H ( div ; Ω ) : ρ ⋅ n = ψ d on Γ N } . \begin{aligned}
H^{1}_{D}(\Omega)&:=\{s\in H^{1}(\Omega):s=0\text{ on }\Gamma_{D}\},&
H_{N}(\operatorname{div};\Omega)&:=\{\boldsymbol{\mu}\in H(\operatorname{div};\Omega):\boldsymbol{\mu}\cdot\mathbf{n}=0\text{ on }\Gamma_{N}\},\\
X_{p}&:=\{w\in H^{1}(\Omega):w=\phi_{p}\text{ on }\Gamma_{D}\},&
\Sigma_{p}&:=\{\boldsymbol{\tau}\in H(\operatorname{div};\Omega):\boldsymbol{\tau}\cdot\mathbf{n}=\psi_{p}\text{ on }\Gamma_{N}\},\\
X_{d}&:=\{v\in H^{1}(\Omega):v=\phi_{d}\text{ on }\Gamma_{D}\},&
\Sigma_{d}&:=\{\boldsymbol{\rho}\in H(\operatorname{div};\Omega):\boldsymbol{\rho}\cdot\mathbf{n}=\psi_{d}\text{ on }\Gamma_{N}\}.
\end{aligned} H D 1 ( Ω ) X p X d := { s ∈ H 1 ( Ω ) : s = 0 on Γ D } , := { w ∈ H 1 ( Ω ) : w = ϕ p on Γ D } , := { v ∈ H 1 ( Ω ) : v = ϕ d on Γ D } , H N ( div ; Ω ) Σ p Σ d := { μ ∈ H ( div ; Ω ) : μ ⋅ n = 0 on Γ N } , := { τ ∈ H ( div ; Ω ) : τ ⋅ n = ψ p on Γ N } , := { ρ ∈ H ( div ; Ω ) : ρ ⋅ n = ψ d on Γ N } .
The primal problem is
− div ( A ∇ u ) + b ⋅ ∇ u + c u = f 1 + div f 2 in Ω , u = ϕ p on Γ D , − ( A ∇ u + f 2 ) ⋅ n = ψ p on Γ N . (1) -\operatorname{div}(A\nabla u)+\mathbf{b}\cdot\nabla u+cu=f_{1}+\operatorname{div}\mathbf{f}_{2}\quad\text{in }\Omega,\qquad u=\phi_{p}\text{ on }\Gamma_{D},\qquad-(A\nabla u+\mathbf{f}_{2})\cdot\mathbf{n}=\psi_{p}\text{ on }\Gamma_{N}. \tag*{(1)} − div ( A ∇ u ) + b ⋅ ∇ u + c u = f 1 + div f 2 in Ω , u = ϕ p on Γ D , − ( A ∇ u + f 2 ) ⋅ n = ψ p on Γ N . (1)
With
a ( w , v ) : = ( A ∇ w , ∇ v ) + ( b ⋅ ∇ w + c w , v ) , ℓ ( v ) : = ( f 1 , v ) − ( f 2 , ∇ v ) − ⟨ ψ p , v ⟩ Γ N , (2) a(w,v):=(A\nabla w,\nabla v)+(\mathbf{b}\cdot\nabla w+cw,v),\qquad\ell(v):=(f_{1},v)-(\mathbf{f}_{2},\nabla v)-\langle\psi_{p},v\rangle_{\Gamma_{N}}, \tag*{(2)} a ( w , v ) := ( A ∇ w , ∇ v ) + ( b ⋅ ∇ w + c w , v ) , ℓ ( v ) := ( f 1 , v ) − ( f 2 , ∇ v ) − ⟨ ψ p , v ⟩ Γ N , (2)
its weak form is: find u ∈ X p u\in X_{p} u ∈ X p with a ( u , s ) = ℓ ( s ) a(u,s)=\ell(s) a ( u , s ) = ℓ ( s ) for all s ∈ H D 1 ( Ω ) s\in H^{1}_{D}(\Omega) s ∈ H D 1 ( Ω ) . We assume that a a a satisfies the inf-sup condition on H D 1 ( Ω ) × H D 1 ( Ω ) H^{1}_{D}(\Omega)\times H^{1}_{D}(\Omega) H D 1 ( Ω ) × H D 1 ( Ω ) , so that this problem and its adjoint problem are uniquely solvable. The flux σ : = − ( A ∇ u + f 2 ) \boldsymbol{\sigma}:=-(A\nabla u+\mathbf{f}_{2}) σ := − ( A ∇ u + f 2 ) belongs to Σ p \Sigma_{p} Σ p , and ( σ , u ) (\boldsymbol{\sigma},u) ( σ , u ) solves the first-order system
div σ + b ⋅ ∇ u + c u = f 1 , σ + A ∇ u = − f 2 in Ω , (3) \operatorname{div}\boldsymbol{\sigma}+\mathbf{b}\cdot\nabla u+cu=f_{1},\qquad\boldsymbol{\sigma}+A\nabla u=-\mathbf{f}_{2}\quad\text{in }\Omega, \tag*{(3)} div σ + b ⋅ ∇ u + c u = f 1 , σ + A ∇ u = − f 2 in Ω , (3)
whose first equation we call the balance equation and the second the constitutive equation.
The quantity of interest is
γ : = m ( u ) − ⟨ ( A ∇ u + f 2 ) ⋅ n , ϕ d ⟩ Γ D = m ( u ) + ⟨ σ ⋅ n , ϕ d ⟩ Γ D , m ( w ) : = ( g 1 , w ) − ( g 2 , ∇ w ) − ⟨ ψ d , w ⟩ Γ N . (4) \gamma:=m(u)-\langle(A\nabla u+\mathbf{f}_{2})\cdot\mathbf{n},\phi_{d}\rangle_{\Gamma_{D}}=m(u)+\langle\boldsymbol{\sigma}\cdot\mathbf{n},\phi_{d}\rangle_{\Gamma_{D}},\qquad m(w):=(g_{1},w)-(\mathbf{g}_{2},\nabla w)-\langle\psi_{d},w\rangle_{\Gamma_{N}}. \tag*{(4)} γ := m ( u ) − ⟨( A ∇ u + f 2 ) ⋅ n , ϕ d ⟩ Γ D = m ( u ) + ⟨ σ ⋅ n , ϕ d ⟩ Γ D , m ( w ) := ( g 1 , w ) − ( g 2 , ∇ w ) − ⟨ ψ d , w ⟩ Γ N . (4)
The four data of the quantity of interest determine what we measure. The datum g 1 g_{1} g 1 gives a weighted integral of u u u over Ω \Omega Ω , g 2 \mathbf{g}_{2} g 2 a weighted integral of ∇ u \nabla u ∇ u , ψ d \psi_{d} ψ d a weighted integral of u u u on Γ N \Gamma_{N} Γ N , and ϕ d \phi_{d} ϕ d a weighted integral of the normal flux σ ⋅ n \boldsymbol{\sigma}\cdot\mathbf{n} σ ⋅ n on Γ D \Gamma_{D} Γ D . For instance, with the indicator function 1 ω \mathbf{1}_{\omega} 1 ω of a subdomain ω \omega ω , the choice g 1 = 1 ω g_{1}=\mathbf{1}_{\omega} g 1 = 1 ω gives the integral of u u u over ω \omega ω , and in two dimensions g 2 = ( 1 ω , 0 ) T \mathbf{g}_{2}=(\mathbf{1}_{\omega},0)^{T} g 2 = ( 1 ω , 0 ) T gives − ∫ ω ∂ x u d x -\int_{\omega}\partial_{x}u\,dx − ∫ ω ∂ x u d x . The choice ψ d = − 1 \psi_{d}=-1 ψ d = − 1 gives the integral of u u u over Γ N \Gamma_{N} Γ N , and ϕ d = 1 \phi_{d}=1 ϕ d = 1 gives the total outward flux ∫ Γ D σ ⋅ n d s \int_{\Gamma_{D}}\boldsymbol{\sigma}\cdot\mathbf{n}\,ds ∫ Γ D σ ⋅ n d s through Γ D \Gamma_{D} Γ D .
The adjoint problem is: find z ∈ X d z\in X_{d} z ∈ X d with
a ( s , z ) = m ( s ) ∀ s ∈ H D 1 ( Ω ) . (5) a(s,z)=m(s)\qquad\forall s\in H^{1}_{D}(\Omega). \tag*{(5)} a ( s , z ) = m ( s ) ∀ s ∈ H D 1 ( Ω ) . (5)
Its strong form is the physically meaningful adjoint problem of [35 ] ,
− div ( A ∇ z + b z ) + c z = g 1 + div g 2 in Ω , z = ϕ d on Γ D , − ( A ∇ z + b z + g 2 ) ⋅ n = ψ d on Γ N , (6) -\operatorname{div}(A\nabla z+\mathbf{b}z)+cz=g_{1}+\operatorname{div}\mathbf{g}_{2}\quad\text{in }\Omega,\qquad z=\phi_{d}\text{ on }\Gamma_{D},\qquad-(A\nabla z+\mathbf{b}z+\mathbf{g}_{2})\cdot\mathbf{n}=\psi_{d}\text{ on }\Gamma_{N}, \tag*{(6)} − div ( A ∇ z + b z ) + cz = g 1 + div g 2 in Ω , z = ϕ d on Γ D , − ( A ∇ z + b z + g 2 ) ⋅ n = ψ d on Γ N , (6)
and the adjoint flux r : = − ( A ∇ z + b z + g 2 ) ∈ Σ d \mathbf{r}:=-(A\nabla z+\mathbf{b}z+\mathbf{g}_{2})\in\Sigma_{d} r := − ( A ∇ z + b z + g 2 ) ∈ Σ d and z z z solve
div r + c z = g 1 , r + A ∇ z + b z = − g 2 in Ω . (7) \operatorname{div}\mathbf{r}+cz=g_{1},\qquad\mathbf{r}+A\nabla z+\mathbf{b}z=-\mathbf{g}_{2}\quad\text{in }\Omega. \tag*{(7)} div r + cz = g 1 , r + A ∇ z + b z = − g 2 in Ω. (7)
By Green’s formula, (2.3) and (2.7), for every s ∈ H 1 ( Ω ) s\in H^{1}(\Omega) s ∈ H 1 ( Ω ) ,
a ( u , s ) = ℓ ( s ) − ⟨ σ ⋅ n , s ⟩ Γ D , a ( s , z ) = m ( s ) − ⟨ r ⋅ n , s ⟩ Γ D . (8) a(u,s)=\ell(s)-\langle\boldsymbol{\sigma}\cdot\mathbf{n},s\rangle_{\Gamma_{D}},\qquad a(s,z)=m(s)-\langle\mathbf{r}\cdot\mathbf{n},s\rangle_{\Gamma_{D}}. \tag*{(8)} a ( u , s ) = ℓ ( s ) − ⟨ σ ⋅ n , s ⟩ Γ D , a ( s , z ) = m ( s ) − ⟨ r ⋅ n , s ⟩ Γ D . (8)
Residuals, errors and least-squares estimators
For ( τ , w ) (\boldsymbol{\tau},w) ( τ , w ) and ( ρ , v ) (\boldsymbol{\rho},v) ( ρ , v ) in H ( div ; Ω ) × H 1 ( Ω ) H(\operatorname{div};\Omega)\times H^{1}(\Omega) H ( div ; Ω ) × H 1 ( Ω ) , the residuals of the constitutive and balance equations are
R p c s t ( τ , w ) : = f 2 + τ + A ∇ w , R p b a l ( τ , w ) : = div τ + b ⋅ ∇ w + c w − f 1 , R d c s t ( ρ , v ) : = g 2 + ρ + A ∇ v + b v , R d b a l ( ρ , v ) : = div ρ + c v − g 1 . (9) \begin{aligned}
R_{p}^{\mathrm{cst}}(\boldsymbol{\tau},w)&:=\mathbf{f}_{2}+\boldsymbol{\tau}+A\nabla w,&
R_{p}^{\mathrm{bal}}(\boldsymbol{\tau},w)&:=\operatorname{div}\boldsymbol{\tau}+\mathbf{b}\cdot\nabla w+cw-f_{1},\\
R_{d}^{\mathrm{cst}}(\boldsymbol{\rho},v)&:=\mathbf{g}_{2}+\boldsymbol{\rho}+A\nabla v+\mathbf{b}v,&
R_{d}^{\mathrm{bal}}(\boldsymbol{\rho},v)&:=\operatorname{div}\boldsymbol{\rho}+cv-g_{1}. \tag*{(9)}
\end{aligned} R p cst ( τ , w ) R d cst ( ρ , v ) := f 2 + τ + A ∇ w , := g 2 + ρ + A ∇ v + b v , R p bal ( τ , w ) R d bal ( ρ , v ) := div τ + b ⋅ ∇ w + c w − f 1 , := div ρ + c v − g 1 . (9)
They vanish at the exact solutions and are computable. For ( τ , w ) ∈ H ( div ; Ω ) × X p (\boldsymbol{\tau},w)\in H(\operatorname{div};\Omega)\times X_{p} ( τ , w ) ∈ H ( div ; Ω ) × X p and ( ρ , v ) ∈ H ( div ; Ω ) × X d (\boldsymbol{\rho},v)\in H(\operatorname{div};\Omega)\times X_{d} ( ρ , v ) ∈ H ( div ; Ω ) × X d we write throughout
E p : = σ − τ , e p : = u − w , E d : = r − ρ , e d : = z − v . (10) E_{p}:=\boldsymbol{\sigma}-\boldsymbol{\tau},\qquad e_{p}:=u-w,\qquad E_{d}:=\mathbf{r}-\boldsymbol{\rho},\qquad e_{d}:=z-v. \tag*{(10)} E p := σ − τ , e p := u − w , E d := r − ρ , e d := z − v . (10)
so that e p , e d ∈ H D 1 ( Ω ) e_p,e_d \in H_D^1(\Omega) e p , e d ∈ H D 1 ( Ω ) , and E p , E d ∈ H N ( div ; Ω ) E_p,E_d \in H_N(\operatorname{div};\Omega) E p , E d ∈ H N ( div ; Ω ) when τ ∈ Σ p \tau\in\Sigma_p τ ∈ Σ p , ρ ∈ Σ d \rho\in\Sigma_d ρ ∈ Σ d . Then
R p c s t ( τ , w ) = − ( E p + A ∇ e p ) , R p b a l ( τ , w ) = − ( div E p + b ⋅ ∇ e p + c e p ) , R d c s t ( ρ , v ) = − ( E d + A ∇ e d + b e d ) , R d b a l ( ρ , v ) = − ( div E d + c e d ) . (11) \begin{aligned}
R_p^{\mathrm{cst}}(\tau,w)&=-(E_p+A\nabla e_p),&\qquad R_p^{\mathrm{bal}}(\tau,w)&=-(\operatorname{div}E_p+\mathbf{b}\cdot\nabla e_p+ce_p),\\
R_d^{\mathrm{cst}}(\rho,v)&=-(E_d+A\nabla e_d+be_d),& R_d^{\mathrm{bal}}(\rho,v)&=-(\operatorname{div}E_d+ce_d). \tag*{(11)}
\end{aligned} R p cst ( τ , w ) R d cst ( ρ , v ) = − ( E p + A ∇ e p ) , = − ( E d + A ∇ e d + b e d ) , R p bal ( τ , w ) R d bal ( ρ , v ) = − ( div E p + b ⋅ ∇ e p + c e p ) , = − ( div E d + c e d ) . (11)
The least-squares functionals of [35 ] and their estimators are
LS p ( τ , w ) : = ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 2 + ∥ R p b a l ( τ , w ) ∥ 0 2 = η p L S ( τ , w ) 2 , LS d ( ρ , v ) : = ∥ A − 1 / 2 R d c s t ( ρ , v ) ∥ 0 2 + ∥ R d b a l ( ρ , v ) ∥ 0 2 = η d L S ( ρ , v ) 2 . (12) \begin{aligned}
\operatorname{LS}_p(\tau,w)&:=\left\|A^{-1/2}R_p^{\mathrm{cst}}(\tau,w)\right\|_0^2+\left\|R_p^{\mathrm{bal}}(\tau,w)\right\|_0^2=\eta_p^{\mathrm{LS}}(\tau,w)^2,\\
\operatorname{LS}_d(\rho,v)&:=\left\|A^{-1/2}R_d^{\mathrm{cst}}(\rho,v)\right\|_0^2+\left\|R_d^{\mathrm{bal}}(\rho,v)\right\|_0^2=\eta_d^{\mathrm{LS}}(\rho,v)^2. \tag*{(12)}
\end{aligned} LS p ( τ , w ) LS d ( ρ , v ) := A − 1/2 R p cst ( τ , w ) 0 2 + R p bal ( τ , w ) 0 2 = η p LS ( τ , w ) 2 , := A − 1/2 R d cst ( ρ , v ) 0 2 + R d bal ( ρ , v ) 0 2 = η d LS ( ρ , v ) 2 . (12)
Every bounded linear functional on H D 1 ( Ω ) H_D^1(\Omega) H D 1 ( Ω ) can be written as ( f 1 , ⋅ ) − ( f 2 , ∇ ⋅ ) (f_1,\cdot)-(f_2,\nabla\cdot) ( f 1 , ⋅ ) − ( f 2 , ∇ ⋅ ) . Since f 2 f_2 f 2 and g 2 g_2 g 2 enter the constitutive equations, the residuals are in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) for such data [37 ] . We write LS p 0 \operatorname{LS}_p^0 LS p 0 and LS d 0 \operatorname{LS}_d^0 LS d 0 for the functionals with homogeneous data,
LS p 0 ( μ , s ) = ∥ A − 1 / 2 ( μ + A ∇ s ) ∥ 0 2 + ∥ div μ + b ⋅ ∇ s + c s ∥ 0 2 , LS d 0 ( μ , s ) = ∥ A − 1 / 2 ( μ + A ∇ s + b s ) ∥ 0 2 + ∥ div μ + c s ∥ 0 2 , \operatorname{LS}_p^0(\mu,s)=\left\|A^{-1/2}(\mu+A\nabla s)\right\|_0^2+\left\|\operatorname{div}\mu+\mathbf{b}\cdot\nabla s+cs\right\|_0^2,\qquad
\operatorname{LS}_d^0(\mu,s)=\left\|A^{-1/2}(\mu+A\nabla s+bs)\right\|_0^2+\left\|\operatorname{div}\mu+cs\right\|_0^2, LS p 0 ( μ , s ) = A − 1/2 ( μ + A ∇ s ) 0 2 + ∥ div μ + b ⋅ ∇ s + cs ∥ 0 2 , LS d 0 ( μ , s ) = A − 1/2 ( μ + A ∇ s + b s ) 0 2 + ∥ div μ + cs ∥ 0 2 ,
so that, by (11) , LS p ( τ , w ) = LS p 0 ( E p , e p ) \operatorname{LS}_p(\tau,w)=\operatorname{LS}_p^0(E_p,e_p) LS p ( τ , w ) = LS p 0 ( E p , e p ) and LS d ( ρ , v ) = LS d 0 ( E d , e d ) \operatorname{LS}_d(\rho,v)=\operatorname{LS}_d^0(E_d,e_d) LS d ( ρ , v ) = LS d 0 ( E d , e d ) . We measure errors in the norm of H ( div ; Ω ) × H D 1 ( Ω ) H(\operatorname{div};\Omega)\times H_D^1(\Omega) H ( div ; Ω ) × H D 1 ( Ω ) weighted by A A A ,
∣ ∣ ( μ , s ) ∣ ∣ 2 : = ∥ A − 1 / 2 μ ∥ 0 2 + ∥ div μ ∥ 0 2 + ∥ A 1 / 2 ∇ s ∥ 0 2 + ∥ s ∥ 0 2 . (13) \left|\left|(\mu,s)\right|\right|^2:=\left\|A^{-1/2}\mu\right\|_0^2+\left\|\operatorname{div}\mu\right\|_0^2+\left\|A^{1/2}\nabla s\right\|_0^2+\left\|s\right\|_0^2. \tag*{(13)} ∣ ∣ ( μ , s ) ∣ ∣ 2 := A − 1/2 μ 0 2 + ∥ div μ ∥ 0 2 + A 1/2 ∇ s 0 2 + ∥ s ∥ 0 2 . (13)
Proposition 2.1 (Norm equivalence). There are constants c l s , C l s > 0 c_{\mathrm{ls}},C_{\mathrm{ls}}>0 c ls , C ls > 0 such that
c l s ∣ ∣ ( μ , s ) ∣ ∣ ≤ LS p 0 ( μ , s ) 1 / 2 ≤ C l s ∣ ∣ ( μ , s ) ∣ ∣ ∀ ( μ , s ) ∈ H N ( div ; Ω ) × H D 1 ( Ω ) , c_{\mathrm{ls}}\left|\left|(\mu,s)\right|\right|\leq\operatorname{LS}_p^0(\mu,s)^{1/2}\leq C_{\mathrm{ls}}\left|\left|(\mu,s)\right|\right|\qquad\forall\,(\mu,s)\in H_N(\operatorname{div};\Omega)\times H_D^1(\Omega), c ls ∣ ∣ ( μ , s ) ∣ ∣ ≤ LS p 0 ( μ , s ) 1/2 ≤ C ls ∣ ∣ ( μ , s ) ∣ ∣ ∀ ( μ , s ) ∈ H N ( div ; Ω ) × H D 1 ( Ω ) , and likewise for LS d 0 \operatorname{LS}_d^0 LS d 0 . Therefore, for arbitrary ( τ , w ) ∈ Σ p × X p (\tau,w)\in\Sigma_p\times X_p ( τ , w ) ∈ Σ p × X p and ( ρ , v ) ∈ Σ d × X d (\rho,v)\in\Sigma_d\times X_d ( ρ , v ) ∈ Σ d × X d ,
c l s ∣ ∣ ( E p , e p ) ∣ ∣ ≤ η p L S ( τ , w ) ≤ C l s ∣ ∣ ( E p , e p ) ∣ ∣ , c l s ∣ ∣ ( E d , e d ) ∣ ∣ ≤ η d L S ( ρ , v ) ≤ C l s ∣ ∣ ( E d , e d ) ∣ ∣ . (14) c_{\mathrm{ls}}\left|\left|(E_p,e_p)\right|\right|\leq\eta_p^{\mathrm{LS}}(\tau,w)\leq C_{\mathrm{ls}}\left|\left|(E_p,e_p)\right|\right|,\qquad
c_{\mathrm{ls}}\left|\left|(E_d,e_d)\right|\right|\leq\eta_d^{\mathrm{LS}}(\rho,v)\leq C_{\mathrm{ls}}\left|\left|(E_d,e_d)\right|\right|. \tag*{(14)} c ls ∣ ∣ ( E p , e p ) ∣ ∣ ≤ η p LS ( τ , w ) ≤ C ls ∣ ∣ ( E p , e p ) ∣ ∣ , c ls ∣ ∣ ( E d , e d ) ∣ ∣ ≤ η d LS ( ρ , v ) ≤ C ls ∣ ∣ ( E d , e d ) ∣ ∣ . (14)
In Proposition 2.1, the upper bound follows from the triangle inequality and the boundedness of the coefficients, and the lower bound is the coercivity of the least-squares functional of the flux–potential system. This formulation appears already in [21 ] , and its coercivity in H ( div ) × H 1 H(\operatorname{div})\times H^1 H ( div ) × H 1 is classical [32 , 10 ] . See also the monographs [22 , 7 ] and the survey [6 ] . For equations with convection and reaction terms it is proved in [10 ] , in a weighted form for the adjoint system in [9 Thm. 2.2] , with the convection term partly replaced by the flux in [24 Thm. 4.1] , and directly from the stability of the second-order problem in [37 ] . The bounds (14) follow from LS p ( τ , w ) = LS p 0 ( E p , e p ) \operatorname{LS}_p(\tau,w)=\operatorname{LS}_p^0(E_p,e_p) LS p ( τ , w ) = LS p 0 ( E p , e p ) and LS d ( ρ , v ) = LS d 0 ( E d , e d ) \operatorname{LS}_d(\rho,v)=\operatorname{LS}_d^0(E_d,e_d) LS d ( ρ , v ) = LS d 0 ( E d , e d ) . The least-squares functional already serves as an a posteriori estimator in [5 ] .
Remark 2.2 (Weighting the balance term). For b = 0 \mathbf{b}=0 b = 0 and c = 1 c=1 c = 1 the equivalence holds with both constants equal to one, because ( μ , ∇ s ) = − ( div μ , s ) (\mu,\nabla s)=-(\operatorname{div}\mu,s) ( μ , ∇ s ) = − ( div μ , s ) , so that the cross terms in LS p 0 \operatorname{LS}_p^0 LS p 0 cancel and LS p 0 = LS d 0 = ∣ ∣ ⋅ ∣ ∣ 2 \operatorname{LS}_p^0=\operatorname{LS}_d^0=\left|\left|\cdot\right|\right|^2 LS p 0 = LS d 0 = ∣ ∣ ⋅ ∣ ∣ 2 . For b = 0 \mathbf{b}=0 b = 0 and c ≥ c 0 > 0 c\geq c_0>0 c ≥ c 0 > 0 , if we weight the balance terms of (12) by c − 1 / 2 c^{-1/2} c − 1/2 , the same computation gives
∥ A − 1 / 2 ( μ + A ∇ s ) ∥ 0 2 + ∥ c − 1 / 2 ( div μ + c s ) ∥ 0 2 = ∥ A − 1 / 2 μ ∥ 0 2 + ∥ c − 1 / 2 div μ ∥ 0 2 + ∥ A 1 / 2 ∇ s ∥ 0 2 + ∥ c 1 / 2 s ∥ 0 2 \left\|A^{-1/2}(\mu+A\nabla s)\right\|_0^2+\left\|c^{-1/2}(\operatorname{div}\mu+cs)\right\|_0^2=\left\|A^{-1/2}\mu\right\|_0^2+\left\|c^{-1/2}\operatorname{div}\mu\right\|_0^2+\left\|A^{1/2}\nabla s\right\|_0^2+\left\|c^{1/2}s\right\|_0^2 A − 1/2 ( μ + A ∇ s ) 0 2 + c − 1/2 ( div μ + cs ) 0 2 = A − 1/2 μ 0 2 + c − 1/2 div μ 0 2 + A 1/2 ∇ s 0 2 + c 1/2 s 0 2
for every such c c c , so the weighted estimators equal the errors in the weighted norm. This is the weighting of the robust estimator of [1 ] for reaction–diffusion problems. The same weighting arises when we write these problems as convex minimization problems [36 ] , since the weighted functional is twice the primal–dual gap.
The evaluators γ 0 \gamma_0 γ 0 and γ 1 \gamma_1 γ 1
2.3. The evaluators γ 0 \gamma_0 γ 0 and γ 1 \gamma_1 γ 1 . Following Giles and Sili [19 ] , let
γ 0 ( w , v ) : = m ( w ) + ℓ ( v ) − a ( w , v ) , w , v ∈ H 1 ( Ω ) . (15) \gamma_0(w,v):=m(w)+\ell(v)-a(w,v),\qquad w,v\in H^1(\Omega). \tag*{(15)} γ 0 ( w , v ) := m ( w ) + ℓ ( v ) − a ( w , v ) , w , v ∈ H 1 ( Ω ) . (15)
Then
γ − γ 0 ( w , v ) = a ( e p , e d ) (16) \gamma-\gamma_0(w,v)=a(e_p,e_d) \tag*{(16)} γ − γ 0 ( w , v ) = a ( e p , e d ) (16)
for all w ∈ X p w\in X_p w ∈ X p , v ∈ X d v\in X_d v ∈ X d , with no orthogonality assumption [35 ] . If fluxes are available, we can express the gradients in (15) through them, ∇ u = − A − 1 ( σ + f 2 ) \nabla u=-A^{-1}(\sigma+f_2) ∇ u = − A − 1 ( σ + f 2 ) and A ∇ z + b z = − ( r + g 2 ) A\nabla z+bz=-(r+g_2) A ∇ z + b z = − ( r + g 2 ) . Substituting these expressions in m m m , ℓ \ell ℓ and a a a , we obtain m ~ \widetilde m m , ℓ ~ \widetilde\ell ℓ and a ~ \widetilde a a below. The second evaluator of [35 ] adds to m ~ + ℓ ~ − a ~ \widetilde{m}+\widetilde{\ell}-\widetilde{a} m + ℓ − a two corrections, the constitutive residuals tested with the fluxes, which vanish at the exact solutions:
γ 1 ( w , τ , v , ρ ) : = m ~ ( τ , w ) + ℓ ~ ( ρ , v ) − a ~ ( ( τ , w ) , ( ρ , v ) ) + ( R p c s t ( τ , w ) , A − 1 ρ ) + ( A − 1 τ , R d c s t ( ρ , v ) ) , (17) \begin{aligned}
\gamma_1(w,\tau,v,\rho) &:= \widetilde{m}(\tau,w)+\widetilde{\ell}(\rho,v)-\widetilde{a}((\tau,w),(\rho,v)) \\
&\quad+(R_p^{\mathrm{cst}}(\tau,w),A^{-1}\rho)+(A^{-1}\tau,R_d^{\mathrm{cst}}(\rho,v)), \tag*{(17)}
\end{aligned} γ 1 ( w , τ , v , ρ ) := m ( τ , w ) + ℓ ( ρ , v ) − a (( τ , w ) , ( ρ , v )) + ( R p cst ( τ , w ) , A − 1 ρ ) + ( A − 1 τ , R d cst ( ρ , v )) , (17)
m ~ ( τ , w ) : = ( g 1 , w ) + ( A − 1 ( τ + f 2 ) , g 2 ) − ⟨ ψ d , w ⟩ Γ N , \widetilde{m}(\tau,w):=(g_1,w)+(A^{-1}(\tau+f_2),g_2)-\langle\psi_d,w\rangle_{\Gamma_N}, m ( τ , w ) := ( g 1 , w ) + ( A − 1 ( τ + f 2 ) , g 2 ) − ⟨ ψ d , w ⟩ Γ N ,
ℓ ~ ( ρ , v ) : = ( f 1 , v ) + ( A − 1 ( ρ + b v + g 2 ) , f 2 ) − ⟨ ψ p , v ⟩ Γ N , \widetilde{\ell}(\rho,v):=(f_1,v)+(A^{-1}(\rho+bv+g_2),f_2)-\langle\psi_p,v\rangle_{\Gamma_N}, ℓ ( ρ , v ) := ( f 1 , v ) + ( A − 1 ( ρ + b v + g 2 ) , f 2 ) − ⟨ ψ p , v ⟩ Γ N ,
a ~ ( ( τ , w ) , ( ρ , v ) ) : = ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( c w , v ) , \widetilde{a}((\tau,w),(\rho,v)) := (A^{-1}(\tau+f_2),\rho+g_2)+(cw,v), a (( τ , w ) , ( ρ , v )) := ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( c w , v ) ,
for ( τ , w ) , ( ρ , v ) ∈ H ( div ; Ω ) × H 1 ( Ω ) (\tau,w),(\rho,v)\in H(\operatorname{div};\Omega)\times H^1(\Omega) ( τ , w ) , ( ρ , v ) ∈ H ( div ; Ω ) × H 1 ( Ω ) . We need the boundary conditions w ∈ X p w\in X_p w ∈ X p and v ∈ X d v\in X_d v ∈ X d only for consistency, as in (16) . At the exact solutions R p c s t ( σ , u ) = R d c s t ( r , z ) = 0 R_p^{\mathrm{cst}}(\sigma,u)=R_d^{\mathrm{cst}}(r,z)=0 R p cst ( σ , u ) = R d cst ( r , z ) = 0 , and γ 1 ( u , σ , z , r ) = γ 0 ( u , z ) = γ \gamma_1(u,\sigma,z,r)=\gamma_0(u,z)=\gamma γ 1 ( u , σ , z , r ) = γ 0 ( u , z ) = γ .
The QoI gaps between the evaluators
The QoI gap identity
Theorem 3.1 (QoI gap identity). For all ( τ , w ) (\tau,w) ( τ , w ) and ( ρ , v ) (\rho,v) ( ρ , v ) in H ( div ; Ω ) × H 1 ( Ω ) H(\operatorname{div};\Omega)\times H^1(\Omega) H ( div ; Ω ) × H 1 ( Ω ) ,
γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) . (18) \gamma_1(w,\tau,v,\rho)-\gamma_0(w,v)=(A^{-1}R_p^{\mathrm{cst}}(\tau,w),R_d^{\mathrm{cst}}(\rho,v)). \tag*{(18)} γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v )) . (18) Proof. By the definitions of m ~ \widetilde{m} m and m m m , and of ℓ ~ \widetilde{\ell} ℓ and ℓ \ell ℓ ,
m ~ ( τ , w ) = m ( w ) + ( R p c s t ( τ , w ) , A − 1 g 2 ) , ℓ ~ ( ρ , v ) = ℓ ( v ) + ( R d c s t ( ρ , v ) , A − 1 f 2 ) . \widetilde{m}(\tau,w)=m(w)+(R_p^{\mathrm{cst}}(\tau,w),A^{-1}g_2),\qquad
\widetilde{\ell}(\rho,v)=\ell(v)+(R_d^{\mathrm{cst}}(\rho,v),A^{-1}f_2). m ( τ , w ) = m ( w ) + ( R p cst ( τ , w ) , A − 1 g 2 ) , ℓ ( ρ , v ) = ℓ ( v ) + ( R d cst ( ρ , v ) , A − 1 f 2 ) . Since A A A is symmetric, a ( w , v ) − ( c w , v ) = ( A ∇ w , A − 1 ( A ∇ v + b v ) ) a(w,v)-(cw,v)=(A\nabla w,A^{-1}(A\nabla v+bv)) a ( w , v ) − ( c w , v ) = ( A ∇ w , A − 1 ( A ∇ v + b v )) . We subtract (15) from (17) . The terms m ( w ) m(w) m ( w ) and ℓ ( v ) \ell(v) ℓ ( v ) cancel, and
γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( R p c s t ( τ , w ) , A − 1 g 2 ) + ( R p c s t ( τ , w ) , A − 1 ρ ) + ( R d c s t ( ρ , v ) , A − 1 f 2 ) + ( A − 1 τ , R d c s t ( ρ , v ) ) − ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( A ∇ w , A − 1 ( A ∇ v + b v ) ) . \begin{aligned}
\gamma_1(w,\tau,v,\rho)-\gamma_0(w,v)
&=(R_p^{\mathrm{cst}}(\tau,w),A^{-1}g_2)+(R_p^{\mathrm{cst}}(\tau,w),A^{-1}\rho)\\
&\quad+(R_d^{\mathrm{cst}}(\rho,v),A^{-1}f_2)+(A^{-1}\tau,R_d^{\mathrm{cst}}(\rho,v))\\
&\quad-(A^{-1}(\tau+f_2),\rho+g_2)+(A\nabla w,A^{-1}(A\nabla v+bv)).
\end{aligned} γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( R p cst ( τ , w ) , A − 1 g 2 ) + ( R p cst ( τ , w ) , A − 1 ρ ) + ( R d cst ( ρ , v ) , A − 1 f 2 ) + ( A − 1 τ , R d cst ( ρ , v )) − ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( A ∇ w , A − 1 ( A ∇ v + b v )) . We split R d c s t ( ρ , v ) = ( g 2 + ρ ) + ( A ∇ v + b v ) R_d^{\mathrm{cst}}(\rho,v)=(g_2+\rho)+(A\nabla v+bv) R d cst ( ρ , v ) = ( g 2 + ρ ) + ( A ∇ v + b v ) in the second bracket. Its first part cancels the third term, and we are left with
γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( R p c s t ( τ , w ) , A − 1 ( g 2 + ρ ) ) + ( A − 1 ( f 2 + τ + A ∇ w ) , A ∇ v + b v ) = ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) . \begin{aligned}
\gamma_1(w,\tau,v,\rho)-\gamma_0(w,v)
&=(R_p^{\mathrm{cst}}(\tau,w),A^{-1}(g_2+\rho))\\
&\quad+(A^{-1}(f_2+\tau+A\nabla w),A\nabla v+bv)\\
&=(A^{-1}R_p^{\mathrm{cst}}(\tau,w),R_d^{\mathrm{cst}}(\rho,v)).
\end{aligned} γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( R p cst ( τ , w ) , A − 1 ( g 2 + ρ )) + ( A − 1 ( f 2 + τ + A ∇ w ) , A ∇ v + b v ) = ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v )) . The proof uses only L 2 L^2 L 2 inner products, so the identity also holds for fluxes that are only in L 2 ( Ω ) d L^2(\Omega)^d L 2 ( Ω ) d , for instance − ( A ∇ w ~ + f 2 ) -(A\nabla\widetilde{w}+f_2) − ( A ∇ w + f 2 ) for another potential w ~ \widetilde{w} w , or the broken flux of a Crouzeix–Raviart approximation. We take all fluxes in H ( div ; Ω ) H(\operatorname{div};\Omega) H ( div ; Ω ) because the balance residuals need it, and so do the least-squares bounds and γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) , which use these residuals. The QoI gap itself uses only the L 2 L^2 L 2 errors of the fluxes.
Corollary 3.2. For ( τ , w ) ∈ Σ p × X p (\tau,w)\in\Sigma_p\times X_p ( τ , w ) ∈ Σ p × X p and ( ρ , v ) ∈ Σ d × X d (\rho,v)\in\Sigma_d\times X_d ( ρ , v ) ∈ Σ d × X d ,
∣ γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) ∣ ≤ ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 ∥ A − 1 / 2 R d c s t ( ρ , v ) ∥ 0 ≤ η p c s t ( τ , w ) η d c s t ( ρ , v ) . (19) \left|\gamma_1(w,\tau,v,\rho)-\gamma_0(w,v)\right|
\le\|A^{-1/2}R_p^{\mathrm{cst}}(\tau,w)\|_0\|A^{-1/2}R_d^{\mathrm{cst}}(\rho,v)\|_0
\le\eta_p^{\mathrm{cst}}(\tau,w)\eta_d^{\mathrm{cst}}(\rho,v). \tag*{(19)} ∣ γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) ∣ ≤ ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 ∥ A − 1/2 R d cst ( ρ , v ) ∥ 0 ≤ η p cst ( τ , w ) η d cst ( ρ , v ) . (19) Proof. By Theorem 3.1 , γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) \gamma_1(w,\tau,v,\rho)-\gamma_0(w,v) γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) is the inner product of A − 1 / 2 R p c s t ( τ , w ) A^{-1/2}R_p^{\mathrm{cst}}(\tau,w) A − 1/2 R p cst ( τ , w ) and A − 1 / 2 R d c s t ( ρ , v ) A^{-1/2}R_d^{\mathrm{cst}}(\rho,v) A − 1/2 R d cst ( ρ , v ) . The first inequality is the Cauchy–Schwarz inequality, and the second holds because each factor is a part of the corresponding estimator (12) . □
We call the computable quantity
δ ( w , τ , v , ρ ) : = γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( A − 1 ( f 2 + τ + A ∇ w ) , g 2 + ρ + A ∇ v + b v ) (20) \begin{aligned}
\delta(w,\tau,v,\rho):=\gamma_1(w,\tau,v,\rho)-\gamma_0(w,v) \\
&=(A^{-1}(f_2+\tau+A\nabla w),g_2+\rho+A\nabla v+bv) \tag*{(20)}
\end{aligned} δ ( w , τ , v , ρ ) := γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) = ( A − 1 ( f 2 + τ + A ∇ w ) , g 2 + ρ + A ∇ v + b v ) (20)
the signed QoI gap , or simply the QoI gap. By (11) ,
δ ( w , τ , v , ρ ) = ( A − 1 ( E p + A ∇ e p ) , E d + A ∇ e d + b e d ) (21) \delta(w,\tau,v,\rho)=(A^{-1}(E_p+A\nabla e_p),E_d+A\nabla e_d+b e_d) \tag*{(21)} δ ( w , τ , v , ρ ) = ( A − 1 ( E p + A ∇ e p ) , E d + A ∇ e d + b e d ) (21)
for ( τ , w ) ∈ H ( div ; Ω ) × X p (\boldsymbol{\tau},w) \in H(\operatorname{div};\Omega) \times X_p ( τ , w ) ∈ H ( div ; Ω ) × X p and ( ρ , v ) ∈ H ( div ; Ω ) × X d (\boldsymbol{\rho},v) \in H(\operatorname{div};\Omega) \times X_d ( ρ , v ) ∈ H ( div ; Ω ) × X d . The QoI gap is thus the A − 1 A^{-1} A − 1 inner product of the errors of the two constitutive equations. The error of the potential evaluator,
γ − γ 0 ( w , v ) = a ( e p , e d ) = ( ∇ e p , A ∇ e d + b e d ) + ( c e p , e d ) , \gamma-\gamma_0(w,v)=a(e_p,e_d)=(\nabla e_p,A\nabla e_d+b e_d)+(c e_p,e_d), γ − γ 0 ( w , v ) = a ( e p , e d ) = ( ∇ e p , A ∇ e d + b e d ) + ( c e p , e d ) ,
is the A − 1 A^{-1} A − 1 inner product of the potential parts of the two constitutive errors plus the reaction term. Since
γ − γ 0 ( w , v ) = δ ( w , τ , v , ρ ) + ( γ − γ 1 ( w , τ , v , ρ ) ) , \gamma-\gamma_0(w,v)=\delta(w,\boldsymbol{\tau},v,\boldsymbol{\rho})+(\gamma-\gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})), γ − γ 0 ( w , v ) = δ ( w , τ , v , ρ ) + ( γ − γ 1 ( w , τ , v , ρ )) ,
the computable QoI gap equals the error of γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) up to the error of γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho}) γ 1 ( w , τ , v , ρ ) .
Bounds and two-sided estimates
Corollary 3.3 (Bounds). For ( τ , w ) ∈ Σ p × X p (\boldsymbol{\tau},w) \in\Sigma_p \times X_p ( τ , w ) ∈ Σ p × X p and ( ρ , v ) ∈ Σ d × X d (\boldsymbol{\rho},v) \in\Sigma_d \times X_d ( ρ , v ) ∈ Σ d × X d ,
∣ γ − γ 0 ( w , v ) ∣ ≤ C η p L S ( τ , w ) η d L S ( ρ , v ) , ∣ γ − γ 1 ( w , τ , v , ρ ) ∣ ≤ C η p L S ( τ , w ) η d L S ( ρ , v ) . (22) \begin{aligned}
|\gamma-\gamma_0(w,v)| &\leq C\eta_p^{\mathrm{LS}}(\boldsymbol{\tau},w)\eta_d^{\mathrm{LS}}(\boldsymbol{\rho},v),\\
|\gamma-\gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})| &\leq C\eta_p^{\mathrm{LS}}(\boldsymbol{\tau},w)\eta_d^{\mathrm{LS}}(\boldsymbol{\rho},v). \tag*{(22)}
\end{aligned} ∣ γ − γ 0 ( w , v ) ∣ ∣ γ − γ 1 ( w , τ , v , ρ ) ∣ ≤ C η p LS ( τ , w ) η d LS ( ρ , v ) , ≤ C η p LS ( τ , w ) η d LS ( ρ , v ) . (22) Proof. We bound (16) by the boundedness of a a a and use (14) . For γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho}) γ 1 ( w , τ , v , ρ ) we use γ − γ 1 ( w , τ , v , ρ ) = γ − γ 0 ( w , v ) − δ ( w , τ , v , ρ ) \gamma-\gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})=\gamma-\gamma_0(w,v)-\delta(w,\boldsymbol{\tau},v,\boldsymbol{\rho}) γ − γ 1 ( w , τ , v , ρ ) = γ − γ 0 ( w , v ) − δ ( w , τ , v , ρ ) and (19) .
Whether a QoI gap measures the errors of the two evaluators depends on the signs of the errors. The argument uses only that a QoI gap is the difference of two approximations of γ \gamma γ , so we state it for two arbitrary approximations.
Proposition 3.4 (Two-sided estimate). Let γ A \gamma_A γ A and γ B \gamma_B γ B be two approximations of γ \gamma γ with errors e A : = γ − γ A e_A:=\gamma-\gamma_A e A := γ − γ A and e B : = γ − γ B e_B:=\gamma-\gamma_B e B := γ − γ B , labeled so that ∣ e A ∣ ≤ ∣ e B ∣ |e_A|\leq|e_B| ∣ e A ∣ ≤ ∣ e B ∣ . Then ∣ γ A − γ B ∣ ≤ ∣ e A ∣ + ∣ e B ∣ ≤ 2 ∣ e B ∣ |\gamma_A-\gamma_B|\leq|e_A|+|e_B|\leq2|e_B| ∣ γ A − γ B ∣ ≤ ∣ e A ∣ + ∣ e B ∣ ≤ 2∣ e B ∣ , and the following hold.
(a) If e A e B ≤ 0 e_Ae_B\leq0 e A e B ≤ 0 , that is, if γ \gamma γ lies between γ A \gamma_A γ A and γ B \gamma_B γ B , then ∣ γ A − γ B ∣ = ∣ e A ∣ + ∣ e B ∣ |\gamma_A-\gamma_B|=|e_A|+|e_B| ∣ γ A − γ B ∣ = ∣ e A ∣ + ∣ e B ∣ , and consequently ∣ e B ∣ ≤ ∣ γ A − γ B ∣ ≤ 2 ∣ e B ∣ |e_B|\leq|\gamma_A-\gamma_B|\leq2|e_B| ∣ e B ∣ ≤ ∣ γ A − γ B ∣ ≤ 2∣ e B ∣ .
(b) If e A e B > 0 e_Ae_B>0 e A e B > 0 , then ∣ γ A − γ B ∣ = ∣ e B ∣ − ∣ e A ∣ = ( 1 − ∣ e A ∣ / ∣ e B ∣ ) ∣ e B ∣ |\gamma_A-\gamma_B|=|e_B|-|e_A|=\left(1-|e_A|/|e_B|\right)|e_B| ∣ γ A − γ B ∣ = ∣ e B ∣ − ∣ e A ∣ = ( 1 − ∣ e A ∣/∣ e B ∣ ) ∣ e B ∣ .
In particular, γ \gamma γ lies between γ A \gamma_A γ A and γ B \gamma_B γ B if and only if ∣ γ A − γ B ∣ = ∣ e A ∣ + ∣ e B ∣ |\gamma_A-\gamma_B|=|e_A|+|e_B| ∣ γ A − γ B ∣ = ∣ e A ∣ + ∣ e B ∣ .
Proof. If e A e B ≤ 0 e_Ae_B\leq0 e A e B ≤ 0 then ∣ e B − e A ∣ = ∣ e A ∣ + ∣ e B ∣ |e_B-e_A|=|e_A|+|e_B| ∣ e B − e A ∣ = ∣ e A ∣ + ∣ e B ∣ , and if e A e B > 0 e_Ae_B>0 e A e B > 0 then ∣ e B − e A ∣ = ∣ e B ∣ − ∣ e A ∣ |e_B-e_A|=|e_B|-|e_A| ∣ e B − e A ∣ = ∣ e B ∣ − ∣ e A ∣ , which equals ∣ e A ∣ + ∣ e B ∣ |e_A|+|e_B| ∣ e A ∣ + ∣ e B ∣ only if e A = 0 e_A=0 e A = 0 , excluded in case (b). The remaining statements follow from ∣ e A ∣ ≤ ∣ e B ∣ |e_A|\leq|e_B| ∣ e A ∣ ≤ ∣ e B ∣ .
We apply the proposition to γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) and γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho}) γ 1 ( w , τ , v , ρ ) , with ∣ γ A − γ B ∣ = ∣ δ ( w , τ , v , ρ ) ∣ |\gamma_A-\gamma_B|=|\delta(w,\boldsymbol{\tau},v,\boldsymbol{\rho})| ∣ γ A − γ B ∣ = ∣ δ ( w , τ , v , ρ ) ∣ . The QoI gap then estimates the larger of the two errors. It does so within a factor two when γ \gamma γ lies between the two evaluators, and within the factor ( 1 − ∣ e A ∣ / ∣ e B ∣ ) − 1 \left(1-|e_A|/|e_B|\right)^{-1} ( 1 − ∣ e A ∣/∣ e B ∣ ) − 1 otherwise. This factor is large only when the two evaluators make nearly the same error, in sign and in size. In either case we do not need to know which evaluator is the more accurate. When one error is much smaller than the other, the QoI gap is a signed estimate of the error of the less accurate evaluator, since γ A − γ B = e B − e A ≈ e B \gamma_A-\gamma_B=e_B-e_A\approx e_B γ A − γ B = e B − e A ≈ e B .
The value that we report for the quantity of interest need not be one of the evaluators. In case (a), γ \gamma γ lies in the interval between them, so the length of the interval bounds the error of any value in it, and half of the length bounds the error of the midpoint. These errors can be much smaller than the QoI gap.
The reaction-restored evaluator γ 2 \gamma_2 γ 2
The signed QoI gap does not contain the reaction term ( c e p , e d ) (c e_p,e_d) ( c e p , e d ) of γ − γ 0 ( w , v ) \gamma-\gamma_0(w,v) γ − γ 0 ( w , v ) . The coefficient c c c occurs only in the balance equations of (3) and (7) , while the signed QoI gap pairs the constitutive residuals, so c c c enters it only through the computed fields. Since ∣ ( c e p , e d ) ∣ ≤ ∥ c ∥ 0 , ∞ ∥ e p ∥ 0 ∥ e d ∥ 0 \lvert(c e_p,e_d)\rvert\le\|c\|_{0,\infty}\|e_p\|_0\|e_d\|_0 ∣( c e p , e d )∣ ≤ ∥ c ∥ 0 , ∞ ∥ e p ∥ 0 ∥ e d ∥ 0 , this term is of higher order than the product of the H 1 H^1 H 1 errors when c c c is not large and the L 2 L^2 L 2 errors of the potentials are of higher order than their H 1 H^1 H 1 errors. For large c c c the factor ∥ c ∥ 0 , ∞ \|c\|_{0,\infty} ∥ c ∥ 0 , ∞ can make the term dominant, and the reaction term causes layers that the mesh may not resolve.
To restore the reaction term, we solve the balance equations for the potentials, as in the flux-only least-squares method for linear transport equations [26 ] . We express the convection term of the primal balance equation through the flux, b ⋅ ∇ u = − b ⋅ A − 1 ( σ + f 2 ) \mathbf{b}\cdot\nabla u=-\mathbf{b}\cdot A^{-1}(\boldsymbol{\sigma}+f_2) b ⋅ ∇ u = − b ⋅ A − 1 ( σ + f 2 ) , as in mixed methods, so that the balance equation becomes div σ − b ⋅ A − 1 ( σ + f 2 ) + c u = f 1 \operatorname{div}\boldsymbol{\sigma}-\mathbf{b}\cdot A^{-1}(\boldsymbol{\sigma}+f_2)+cu=f_1 div σ − b ⋅ A − 1 ( σ + f 2 ) + c u = f 1 , one of the first-order systems of [24 §4] . Its residual is R p b a l ( τ , w ) − b ⋅ A − 1 R p c s t ( τ , w ) R_p^{\mathrm{bal}}(\boldsymbol{\tau},w)-\mathbf{b}\cdot A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w) R p bal ( τ , w ) − b ⋅ A − 1 R p cst ( τ , w ) . The adjoint balance equation of (7) contains no convection term. For c ≥ c 0 > 0 c\ge c_0>0 c ≥ c 0 > 0 we solve these equations for the potentials: for τ ∈ Σ p \boldsymbol{\tau}\in\Sigma_p τ ∈ Σ p and ρ ∈ Σ d \boldsymbol{\rho}\in\Sigma_d ρ ∈ Σ d let
w ^ ( τ ) : = c − 1 ( f 1 − div τ + b ⋅ A − 1 ( τ + f 2 ) ) , v ^ ( ρ ) : = c − 1 ( g 1 − div ρ ) , (23) \widehat{w}(\boldsymbol{\tau}) := c^{-1}\left(f_1-\operatorname{div}\boldsymbol{\tau}+\mathbf{b}\cdot A^{-1}(\boldsymbol{\tau}+f_2)\right),\qquad\widehat{v}(\boldsymbol{\rho}) := c^{-1}\left(g_1-\operatorname{div}\boldsymbol{\rho}\right), \tag*{(23)} w ( τ ) := c − 1 ( f 1 − div τ + b ⋅ A − 1 ( τ + f 2 ) ) , v ( ρ ) := c − 1 ( g 1 − div ρ ) , (23)
so that u = w ^ ( σ ) u=\widehat{w}(\boldsymbol{\sigma}) u = w ( σ ) and z = v ^ ( r ) z=\widehat{v}(\mathbf{r}) z = v ( r ) .
By (8) and (4) , γ = a ( u , z ) + ⟨ σ ⋅ n , ϕ d ⟩ Γ D + ⟨ r ⋅ n , ϕ p ⟩ Γ D \gamma=a(u,z)+\langle\boldsymbol{\sigma}\cdot\mathbf{n},\phi_d\rangle_{\Gamma_D}+\langle\mathbf{r}\cdot\mathbf{n},\phi_p\rangle_{\Gamma_D} γ = a ( u , z ) + ⟨ σ ⋅ n , ϕ d ⟩ Γ D + ⟨ r ⋅ n , ϕ p ⟩ Γ D , with a ( u , z ) = ( ∇ u , A ∇ z + b z ) + ( c u , z ) a(u,z)=(\nabla u,A\nabla z+\mathbf{b}z)+(cu,z) a ( u , z ) = ( ∇ u , A ∇ z + b z ) + ( c u , z ) . The constitutive equations, ∇ u = − A − 1 ( σ + f 2 ) \nabla u=-A^{-1}(\boldsymbol{\sigma}+f_2) ∇ u = − A − 1 ( σ + f 2 ) and A ∇ z + b z = − ( r + g 2 ) A\nabla z+\mathbf{b}z=-(\mathbf{r}+g_2) A ∇ z + b z = − ( r + g 2 ) , give
γ = ( A − 1 ( σ + f 2 ) , r + g 2 ) + ( c u , z ) + ⟨ σ ⋅ n , ϕ d ⟩ Γ D + ⟨ r ⋅ n , ϕ p ⟩ Γ D . \gamma=\left(A^{-1}(\boldsymbol{\sigma}+f_2),\mathbf{r}+g_2\right)+(cu,z)+\langle\boldsymbol{\sigma}\cdot\mathbf{n},\phi_d\rangle_{\Gamma_D}+\langle\mathbf{r}\cdot\mathbf{n},\phi_p\rangle_{\Gamma_D}. γ = ( A − 1 ( σ + f 2 ) , r + g 2 ) + ( c u , z ) + ⟨ σ ⋅ n , ϕ d ⟩ Γ D + ⟨ r ⋅ n , ϕ p ⟩ Γ D .
For c = 0 c=0 c = 0 there is no reaction term, and this formula expresses γ \gamma γ through the fluxes alone. For c ≥ c 0 > 0 c\ge c_0>0 c ≥ c 0 > 0 , replacing u u u and z z z by w ^ ( σ ) \widehat{w}(\boldsymbol{\sigma}) w ( σ ) and v ^ ( r ) \widehat{v}(\mathbf{r}) v ( r ) gives γ = γ 2 ( σ , r ) \gamma=\gamma_2(\boldsymbol{\sigma},\mathbf{r}) γ = γ 2 ( σ , r ) , where the reaction-restored evaluator
γ 2 ( τ , ρ ) : = ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( c w ^ ( τ ) , v ^ ( ρ ) ) + ⟨ τ ⋅ n , ϕ d ⟩ Γ D + ⟨ ρ ⋅ n , ϕ p ⟩ Γ D (24) \gamma_2(\boldsymbol{\tau},\boldsymbol{\rho}) := \left(A^{-1}(\boldsymbol{\tau}+f_2),\boldsymbol{\rho}+g_2\right)+\left(c\widehat{w}(\boldsymbol{\tau}),\widehat{v}(\boldsymbol{\rho})\right)+\langle\boldsymbol{\tau}\cdot\mathbf{n},\phi_d\rangle_{\Gamma_D}+\langle\boldsymbol{\rho}\cdot\mathbf{n},\phi_p\rangle_{\Gamma_D} \tag*{(24)} γ 2 ( τ , ρ ) := ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( c w ( τ ) , v ( ρ ) ) + ⟨ τ ⋅ n , ϕ d ⟩ Γ D + ⟨ ρ ⋅ n , ϕ p ⟩ Γ D (24)
is a functional of the two fluxes alone, just as (15) uses only the potentials. The evaluator γ 0 \gamma_0 γ 0 pairs each Neumann datum with the trace of the other potential, and the potentials satisfy the Dirichlet conditions. The evaluator γ 2 \gamma_2 γ 2 pairs each Dirichlet datum with the normal trace of the other flux, and the fluxes satisfy the Neumann conditions. For potentials w ∈ X p w\in X_p w ∈ X p and v ∈ X d v\in X_d v ∈ X d , γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho}) γ 1 ( w , τ , v , ρ ) has the same form. It differs from γ 2 ( τ , ρ ) \gamma_2(\boldsymbol{\tau},\boldsymbol{\rho}) γ 2 ( τ , ρ ) only in the reaction term, which it evaluates with both the potentials w w w , v v v and the potentials w ^ ( τ ) \widehat{w}(\boldsymbol{\tau}) w ( τ ) , v ^ ( ρ ) \widehat{v}(\boldsymbol{\rho}) v ( ρ ) obtained from the fluxes:
γ 1 ( w , τ , v , ρ ) = ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( c w ^ ( τ ) , v ) + ( c w , v ^ ( ρ ) ) − ( c w , v ) + ⟨ τ ⋅ n , ϕ d ⟩ Γ D + ⟨ ρ ⋅ n , ϕ p ⟩ Γ D , (25) \begin{aligned}
\gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})={}&\left(A^{-1}(\boldsymbol{\tau}+f_2),\boldsymbol{\rho}+g_2\right)+\left(c\widehat{w}(\boldsymbol{\tau}),v\right)+\left(cw,\widehat{v}(\boldsymbol{\rho})\right)-\left(cw,v\right)\\
&+\langle\boldsymbol{\tau}\cdot\mathbf{n},\phi_d\rangle_{\Gamma_D}+\langle\boldsymbol{\rho}\cdot\mathbf{n},\phi_p\rangle_{\Gamma_D}, \tag*{(25)}
\end{aligned} γ 1 ( w , τ , v , ρ ) = ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( c w ( τ ) , v ) + ( c w , v ( ρ ) ) − ( c w , v ) + ⟨ τ ⋅ n , ϕ d ⟩ Γ D + ⟨ ρ ⋅ n , ϕ p ⟩ Γ D , (25)
γ 2 ( τ , ρ ) − γ 1 ( w , τ , v , ρ ) = ( c ( w ^ ( τ ) − w ) , v ^ ( ρ ) − v ) . \gamma_2(\boldsymbol{\tau},\boldsymbol{\rho})-\gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})=\left(c(\widehat{w}(\boldsymbol{\tau})-w),\widehat{v}(\boldsymbol{\rho})-v\right). γ 2 ( τ , ρ ) − γ 1 ( w , τ , v , ρ ) = ( c ( w ( τ ) − w ) , v ( ρ ) − v ) .
The second identity pairs the differences between the potentials obtained from the fluxes and the potentials w w w and v v v .
Proposition 3.6 (Reaction-restored evaluator). Let c ≥ c 0 > 0 c\ge c_0>0 c ≥ c 0 > 0 , τ ∈ Σ p \boldsymbol{\tau}\in\Sigma_p τ ∈ Σ p and ρ ∈ Σ d \boldsymbol{\rho}\in\Sigma_d ρ ∈ Σ d . Then
γ − γ 2 ( τ , ρ ) = − ( A − 1 E p , E d ) − ( c − 1 ( div E p − b ⋅ A − 1 E p ) , div E d ) . (26) \gamma-\gamma_2(\boldsymbol{\tau},\boldsymbol{\rho})=-\left(A^{-1}E_p,E_d\right)-\left(c^{-1}(\operatorname{div}E_p-\mathbf{b}\cdot A^{-1}E_p),\operatorname{div}E_d\right). \tag*{(26)} γ − γ 2 ( τ , ρ ) = − ( A − 1 E p , E d ) − ( c − 1 ( div E p − b ⋅ A − 1 E p ) , div E d ) . (26) and for all w ∈ X p w\in X_p w ∈ X p and v ∈ X d v\in X_d v ∈ X d , (25) holds,
γ 2 ( τ , ρ ) = γ 1 ( w , τ , v , ρ ) + ( c − 1 ( R p b a l ( τ , w ) − b ⋅ A − 1 R p c s t ( τ , w ) ) , R d b a l ( ρ , v ) ) = γ 0 ( w , v ) + ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) + ( c − 1 R p b a l ( τ , w ) , R d b a l ( ρ , v ) ) − ( c − 1 b ⋅ A − 1 R p c s t ( τ , w ) , R d b a l ( ρ , v ) ) , (27) \begin{aligned}
\gamma_2(\boldsymbol{\tau},\boldsymbol{\rho})={}&\gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})+\left(c^{-1}\left(R_p^{\mathrm{bal}}(\boldsymbol{\tau},w)-\mathbf{b}\cdot A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w)\right),R_d^{\mathrm{bal}}(\boldsymbol{\rho},v)\right)\\
={}&\gamma_0(w,v)+\left(A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{cst}}(\boldsymbol{\rho},v)\right)+\left(c^{-1}R_p^{\mathrm{bal}}(\boldsymbol{\tau},w),R_d^{\mathrm{bal}}(\boldsymbol{\rho},v)\right)\\
&-\left(c^{-1}\mathbf{b}\cdot A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{bal}}(\boldsymbol{\rho},v)\right), \tag*{(27)}
\end{aligned} γ 2 ( τ , ρ ) = = γ 1 ( w , τ , v , ρ ) + ( c − 1 ( R p bal ( τ , w ) − b ⋅ A − 1 R p cst ( τ , w ) ) , R d bal ( ρ , v ) ) γ 0 ( w , v ) + ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v ) ) + ( c − 1 R p bal ( τ , w ) , R d bal ( ρ , v ) ) − ( c − 1 b ⋅ A − 1 R p cst ( τ , w ) , R d bal ( ρ , v ) ) , (27) γ 2 ( τ , ρ ) − γ 0 ( w , v ) = a ( e p , e d ) + ( A − 1 E p , E d ) + ( c − 1 ( div E p − b ⋅ A − 1 E p ) , div E d ) . (28) \gamma_2(\boldsymbol{\tau},\boldsymbol{\rho})-\gamma_0(w,v)=a(e_p,e_d)+\left(A^{-1}E_p,E_d\right)+\left(c^{-1}(\operatorname{div}E_p-\mathbf{b}\cdot A^{-1}E_p),\operatorname{div}E_d\right). \tag*{(28)} γ 2 ( τ , ρ ) − γ 0 ( w , v ) = a ( e p , e d ) + ( A − 1 E p , E d ) + ( c − 1 ( div E p − b ⋅ A − 1 E p ) , div E d ) . (28)
Proof. We insert τ = σ − E p \boldsymbol{\tau}=\boldsymbol{\sigma}-E_p τ = σ − E p and ρ = r − E d \boldsymbol{\rho}=\mathbf{r}-E_d ρ = r − E d in (24) , so that c w ^ ( τ ) = c u + div E p − b ⋅ A − 1 E p c\widehat{w}(\boldsymbol{\tau})=cu+\operatorname{div}E_p-\mathbf{b}\cdot A^{-1}E_p c w ( τ ) = c u + div E p − b ⋅ A − 1 E p and c v ^ ( ρ ) = c z + div E d c\widehat{v}(\boldsymbol{\rho})=cz+\operatorname{div}E_d c v ( ρ ) = cz + div E d , and use γ = γ 2 ( σ , r ) \gamma=\gamma_2(\boldsymbol{\sigma},\mathbf{r}) γ = γ 2 ( σ , r ) and the constitutive equations. The linear terms in E p E_p E p are ( E p , ∇ z ) + ( div E p , z ) − ⟨ E p ⋅ n , ϕ d ⟩ Γ D (E_p,\nabla z)+(\operatorname{div}E_p,z)-\langle E_p\cdot\mathbf{n},\phi_d\rangle_{\Gamma_D} ( E p , ∇ z ) + ( div E p , z ) − ⟨ E p ⋅ n , ϕ d ⟩ Γ D , and those linear in E d E_d E d are ( ∇ u , E d ) + ( u , div E d ) − ⟨ E d ⋅ n , ϕ p ⟩ Γ D (\nabla u,E_d)+(u,\operatorname{div}E_d)-\langle E_d\cdot\mathbf{n},\phi_p\rangle_{\Gamma_D} ( ∇ u , E d ) + ( u , div E d ) − ⟨ E d ⋅ n , ϕ p ⟩ Γ D . Both vanish by Green’s formula, since E p ⋅ n = E d ⋅ n = 0 E_p\cdot\mathbf{n}=E_d\cdot\mathbf{n}=0 E p ⋅ n = E d ⋅ n = 0 on Γ N \Gamma_N Γ N , z = ϕ d z=\phi_d z = ϕ d and u = ϕ p u=\phi_p u = ϕ p on Γ D \Gamma_D Γ D . The remaining terms are quadratic and give (26) .
By (18) and (15) , γ 1 ( w , τ , v , ρ ) = m ( w ) + ℓ ( v ) − a ( w , v ) + ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) \gamma_1(w,\boldsymbol{\tau},v,\boldsymbol{\rho})=m(w)+\ell(v)-a(w,v)+(A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{cst}}(\boldsymbol{\rho},v)) γ 1 ( w , τ , v , ρ ) = m ( w ) + ℓ ( v ) − a ( w , v ) + ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v )) . We expand ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) (A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{cst}}(\boldsymbol{\rho},v)) ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v )) by (9) . The terms from ( g 2 , ∇ w ) (g_2,\nabla w) ( g 2 , ∇ w ) , ( f 2 , ∇ v ) (f_2,\nabla v) ( f 2 , ∇ v ) and ( A ∇ w , ∇ v ) (A\nabla w,\nabla v) ( A ∇ w , ∇ v ) cancel against those of m ( w ) m(w) m ( w ) , ℓ ( v ) \ell(v) ℓ ( v ) and a ( w , v ) a(w,v) a ( w , v ) , and Green’s formula applied to ( τ , ∇ v ) (\boldsymbol{\tau},\nabla v) ( τ , ∇ v ) and ( ∇ w , ρ ) (\nabla w,\boldsymbol{\rho}) ( ∇ w , ρ ) , with the boundary conditions of w w w , τ \boldsymbol{\tau} τ and ρ \rho ρ , gives (25) . Since c ( w ^ ( τ ) − w ) = − ( R p b a l ( τ , w ) − b ⋅ A − 1 R p c s t ( τ , w ) ) c(\widehat{w}(\tau)-w)=-(R_p^{\mathrm{bal}}(\tau,w)-\mathbf{b}\cdot A^{-1}R_p^{\mathrm{cst}}(\tau,w)) c ( w ( τ ) − w ) = − ( R p bal ( τ , w ) − b ⋅ A − 1 R p cst ( τ , w )) and v ^ ( ρ ) − v = − c − 1 R d b a l ( ρ , v ) \widehat{v}(\rho)-v=-c^{-1}R_d^{\mathrm{bal}}(\rho,v) v ( ρ ) − v = − c − 1 R d bal ( ρ , v ) , this gives the first equality in (27) , and (18) gives the second. Finally, we obtain (28) by subtracting (26) from (16) . □
Remark 3.7 (The case c = 0 c=0 c = 0 ). For c = 0 c=0 c = 0 the potentials (23) are not defined, but c w ^ ( τ ) = f 1 − div τ + b ⋅ A − 1 ( τ + f 2 ) c\widehat{w}(\tau)=f_1-\operatorname{div}\tau+\mathbf{b}\cdot A^{-1}(\tau+f_2) c w ( τ ) = f 1 − div τ + b ⋅ A − 1 ( τ + f 2 ) and c v ^ ( ρ ) = g 1 − div ρ c\widehat{v}(\rho)=g_1-\operatorname{div}\rho c v ( ρ ) = g 1 − div ρ are, and the proof of (25) applies unchanged:
γ 1 ( w , τ , v , ρ ) = ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( f 1 − div τ + b ⋅ A − 1 ( τ + f 2 ) , v ) + ( w , g 1 − div ρ ) + ⟨ τ ⋅ n , ϕ d ⟩ Γ D + ⟨ ρ ⋅ n , ϕ p ⟩ Γ D . \gamma_1(w,\boldsymbol{\tau},v,\rho)=(A^{-1}(\boldsymbol{\tau}+f_2),\rho+g_2)+(f_1-\operatorname{div}\boldsymbol{\tau}+\mathbf{b}\cdot A^{-1}(\boldsymbol{\tau}+f_2),v)+(w,g_1-\operatorname{div}\rho)
+\langle\boldsymbol{\tau}\cdot\mathbf{n},\phi_d\rangle_{\Gamma_D}+\langle\rho\cdot\mathbf{n},\phi_p\rangle_{\Gamma_D}. γ 1 ( w , τ , v , ρ ) = ( A − 1 ( τ + f 2 ) , ρ + g 2 ) + ( f 1 − div τ + b ⋅ A − 1 ( τ + f 2 ) , v ) + ( w , g 1 − div ρ ) + ⟨ τ ⋅ n , ϕ d ⟩ Γ D + ⟨ ρ ⋅ n , ϕ p ⟩ Γ D .
where the second and third terms test the balance residuals of the fluxes with the potentials. For fluxes that satisfy the balance equations, these two terms vanish. Then γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\rho) γ 1 ( w , τ , v , ρ ) does not depend on w w w and v v v , as γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) does not depend on the fluxes, and, as in the proof of Proposition 3.6, γ − γ 1 ( w , τ , v , ρ ) = − ( A − 1 E p , E d ) \gamma-\gamma_1(w,\boldsymbol{\tau},v,\rho)=-(A^{-1}E_p,E_d) γ − γ 1 ( w , τ , v , ρ ) = − ( A − 1 E p , E d ) . For b = 0 \mathbf{b}=0 b = 0 these are the equilibrated fluxes. The fluxes of a mixed method are such fluxes, up to the oscillation of f 1 f_1 f 1 and g 1 g_1 g 1 . Then γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\rho) γ 1 ( w , τ , v , ρ ) evaluates γ \gamma γ from the mixed fluxes alone, and the QoI gap compares it with γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) , which uses only the potentials. If the balance residuals of the fluxes are of order c c c , the reaction term ( c w ^ ( τ ) , v ^ ( ρ ) ) (c\widehat{w}(\tau),\widehat{v}(\rho)) ( c w ( τ ) , v ( ρ )) of (24) tends to zero as c → 0 c\to0 c → 0 , and γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) tends to the same expression.
The error (26) of γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) pairs the primal and adjoint flux errors, as the error a ( e p , e d ) a(e_p,e_d) a ( e p , e d ) of γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) pairs the potential errors. We call
δ 2 ( w , τ , v , ρ ) : = γ 2 ( τ , ρ ) − γ 0 ( w , v ) \delta_2(w,\boldsymbol{\tau},v,\rho):=\gamma_2(\tau,\rho)-\gamma_0(w,v) δ 2 ( w , τ , v , ρ ) := γ 2 ( τ , ρ ) − γ 0 ( w , v )
the reaction-restored QoI gap . By (28) it equals the whole error of the potential evaluator, reaction term included, up to the error of γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) . The three evaluators use more and more of the fluxes. By (18) and (27) , γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) \gamma_1(w,\boldsymbol{\tau},v,\rho)-\gamma_0(w,v) γ 1 ( w , τ , v , ρ ) − γ 0 ( w , v ) pairs the constitutive residuals, γ 2 ( τ , ρ ) − γ 1 ( w , τ , v , ρ ) \gamma_2(\tau,\rho)-\gamma_1(w,\boldsymbol{\tau},v,\rho) γ 2 ( τ , ρ ) − γ 1 ( w , τ , v , ρ ) pairs the balance residuals, with a convective coupling term, and δ 2 ( w , τ , v , ρ ) \delta_2(w,\boldsymbol{\tau},v,\rho) δ 2 ( w , τ , v , ρ ) pairs both. We measure both QoI gaps from the potential evaluator γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) . Proposition 3.4 applies to δ 2 ( w , τ , v , ρ ) \delta_2(w,\boldsymbol{\tau},v,\rho) δ 2 ( w , τ , v , ρ ) in the same way, with γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) and γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) as the two approximations, and also to γ 1 ( w , τ , v , ρ ) \gamma_1(w,\boldsymbol{\tau},v,\rho) γ 1 ( w , τ , v , ρ ) and γ 2 ( τ , ρ ) \gamma_2(\tau,\rho) γ 2 ( τ , ρ ) . We bound the reaction-restored QoI gap by the least-squares estimators with the balance terms weighted by c − 1 / 2 c^{-1/2} c − 1/2 , as in Remark 2.2,
η p , c L S ( τ , w ) : = ( ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 2 + ∥ c − 1 / 2 R p b a l ( τ , w ) ∥ 0 2 ) 1 / 2 , η d , c L S ( ρ , v ) : = ( ∥ A − 1 / 2 R d c s t ( ρ , v ) ∥ 0 2 + ∥ c − 1 / 2 R d b a l ( ρ , v ) ∥ 0 2 ) 1 / 2 . (29) \begin{aligned}
\eta_{p,c}^{\mathrm{LS}}(\boldsymbol{\tau},w)&:=\left(\|A^{-1/2}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w)\|_0^2+\|c^{-1/2}R_p^{\mathrm{bal}}(\boldsymbol{\tau},w)\|_0^2\right)^{1/2},\\
\eta_{d,c}^{\mathrm{LS}}(\rho,v)&:=\left(\|A^{-1/2}R_d^{\mathrm{cst}}(\rho,v)\|_0^2+\|c^{-1/2}R_d^{\mathrm{bal}}(\rho,v)\|_0^2\right)^{1/2}. \tag*{(29)}
\end{aligned} η p , c LS ( τ , w ) η d , c LS ( ρ , v ) := ( ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 2 + ∥ c − 1/2 R p bal ( τ , w ) ∥ 0 2 ) 1/2 , := ( ∥ A − 1/2 R d cst ( ρ , v ) ∥ 0 2 + ∥ c − 1/2 R d bal ( ρ , v ) ∥ 0 2 ) 1/2 . (29)
Proposition 3.8 (Reaction-restored QoI gap). Let c ≥ c 0 > 0 c\ge c_0>0 c ≥ c 0 > 0 , let ( τ , w ) ∈ Σ p × X p (\boldsymbol{\tau},w)\in\Sigma_p\times X_p ( τ , w ) ∈ Σ p × X p and ( ρ , v ) ∈ Σ d × X d (\rho,v)\in\Sigma_d\times X_d ( ρ , v ) ∈ Σ d × X d , and let ϑ : = ∥ c − 1 / 2 A − 1 / 2 b ∥ 0 , ∞ \vartheta:=\|c^{-1/2}A^{-1/2}\mathbf{b}\|_{0,\infty} ϑ := ∥ c − 1/2 A − 1/2 b ∥ 0 , ∞ . Then
δ 2 ( w , τ , v , ρ ) = ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) + ( c − 1 R p b a l ( τ , w ) , R d b a l ( ρ , v ) ) − ( c − 1 b ⋅ A − 1 R p c s t ( τ , w ) , R d b a l ( ρ , v ) ) (30) \delta_2(w,\boldsymbol{\tau},v,\rho)=(A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{cst}}(\rho,v))+(c^{-1}R_p^{\mathrm{bal}}(\boldsymbol{\tau},w),R_d^{\mathrm{bal}}(\rho,v))
-(c^{-1}\mathbf{b}\cdot A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{bal}}(\rho,v)) \tag*{(30)} δ 2 ( w , τ , v , ρ ) = ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v )) + ( c − 1 R p bal ( τ , w ) , R d bal ( ρ , v )) − ( c − 1 b ⋅ A − 1 R p cst ( τ , w ) , R d bal ( ρ , v )) (30)
and
∣ δ 2 ( w , τ , v , ρ ) ∣ ≤ ( 1 + ϑ ) η p , c L S ( τ , w ) η d , c L S ( ρ , v ) . (31) |\delta_2(w,\boldsymbol{\tau},v,\rho)|\le(1+\vartheta)\eta_{p,c}^{\mathrm{LS}}(\boldsymbol{\tau},w)\eta_{d,c}^{\mathrm{LS}}(\rho,v). \tag*{(31)} ∣ δ 2 ( w , τ , v , ρ ) ∣ ≤ ( 1 + ϑ ) η p , c LS ( τ , w ) η d , c LS ( ρ , v ) . (31)
For b = 0 \mathbf{b}=0 b = 0 , δ 2 ( w , τ , v , ρ ) \delta_2(w,\boldsymbol{\tau},v,\rho) δ 2 ( w , τ , v , ρ ) is the inner product in L 2 ( Ω ) d × L 2 ( Ω ) L^2(\Omega)^d\times L^2(\Omega) L 2 ( Ω ) d × L 2 ( Ω ) , with the weights A − 1 A^{-1} A − 1 and c − 1 c^{-1} c − 1 , of the residual pairs of the two problems. The norm induced by this inner product is the weighted estimator (29) , so ∣ δ 2 ( w , τ , v , ρ ) ∣ ≤ η p , c L S ( τ , w ) η d , c L S ( ρ , v ) |\delta_2(w,\boldsymbol{\tau},v,\rho)|\le\eta_{p,c}^{\mathrm{LS}}(\boldsymbol{\tau},w)\eta_{d,c}^{\mathrm{LS}}(\rho,v) ∣ δ 2 ( w , τ , v , ρ ) ∣ ≤ η p , c LS ( τ , w ) η d , c LS ( ρ , v ) .
Proof. The representation (30) is (27) , by the definition of δ 2 ( w , τ , v , ρ ) \delta_2(w,\boldsymbol{\tau},v,\rho) δ 2 ( w , τ , v , ρ ) . By the Cauchy–Schwarz inequality in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) and then in R 2 \mathbb{R}^2 R 2 ,
∣ ( A − 1 R p c s t ( τ , w ) , R d c s t ( ρ , v ) ) + ( c − 1 R p b a l ( τ , w ) , R d b a l ( ρ , v ) ) ∣ ≤ ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 ∥ A − 1 / 2 R d c s t ( ρ , v ) ∥ 0 + ∥ c − 1 / 2 R p b a l ( τ , w ) ∥ 0 ∥ c − 1 / 2 R d b a l ( ρ , v ) ∥ 0 ≤ η p , c L S ( τ , w ) η d , c L S ( ρ , v ) . \begin{aligned}
\left|(A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w),R_d^{\mathrm{cst}}(\rho,v))+(c^{-1}R_p^{\mathrm{bal}}(\boldsymbol{\tau},w),R_d^{\mathrm{bal}}(\rho,v))\right|
&\le\|A^{-1/2}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w)\|_0\|A^{-1/2}R_d^{\mathrm{cst}}(\rho,v)\|_0\\
&\quad+\|c^{-1/2}R_p^{\mathrm{bal}}(\boldsymbol{\tau},w)\|_0\|c^{-1/2}R_d^{\mathrm{bal}}(\rho,v)\|_0\\
&\le\eta_{p,c}^{\mathrm{LS}}(\boldsymbol{\tau},w)\eta_{d,c}^{\mathrm{LS}}(\rho,v).
\end{aligned} ( A − 1 R p cst ( τ , w ) , R d cst ( ρ , v )) + ( c − 1 R p bal ( τ , w ) , R d bal ( ρ , v )) ≤ ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 ∥ A − 1/2 R d cst ( ρ , v ) ∥ 0 + ∥ c − 1/2 R p bal ( τ , w ) ∥ 0 ∥ c − 1/2 R d bal ( ρ , v ) ∥ 0 ≤ η p , c LS ( τ , w ) η d , c LS ( ρ , v ) .
Pointwise,
∣ c − 1 / 2 b ⋅ A − 1 R p c s t ( τ , w ) ∣ = c − 1 / 2 ∣ ( A − 1 / 2 b ) ⋅ ( A − 1 / 2 R p c s t ( τ , w ) ) ∣ ≤ ϑ ∣ A − 1 / 2 R p c s t ( τ , w ) ∣ |c^{-1/2}\mathbf{b}\cdot A^{-1}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w)|=c^{-1/2}|(A^{-1/2}\mathbf{b})\cdot(A^{-1/2}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w))|\le\vartheta|A^{-1/2}R_p^{\mathrm{cst}}(\boldsymbol{\tau},w)| ∣ c − 1/2 b ⋅ A − 1 R p cst ( τ , w ) ∣ = c − 1/2 ∣ ( A − 1/2 b ) ⋅ ( A − 1/2 R p cst ( τ , w )) ∣ ≤ ϑ ∣ A − 1/2 R p cst ( τ , w ) ∣
so that
∣ ( c − 1 b ⋅ A − 1 R p c s t ( τ , w ) , R d b a l ( ρ , v ) ) ∣ ≤ ϑ ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 ∥ c − 1 / 2 R d b a l ( ρ , v ) ∥ 0 ≤ ϑ η p , c L S ( τ , w ) η d , c L S ( ρ , v ) . \left|(c^{-1}\mathbf{b}\cdot A^{-1}R^{\mathrm{cst}}_{p}(\tau,w),R^{\mathrm{bal}}_{d}(\rho,v))\right|\leq\vartheta\lVert A^{-1/2}R^{\mathrm{cst}}_{p}(\tau,w)\rVert_{0}\lVert c^{-1/2}R^{\mathrm{bal}}_{d}(\rho,v)\rVert_{0}
\leq\vartheta\eta^{\mathrm{LS}}_{p,c}(\tau,w)\eta^{\mathrm{LS}}_{d,c}(\rho,v). ( c − 1 b ⋅ A − 1 R p cst ( τ , w ) , R d bal ( ρ , v )) ≤ ϑ ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 ∥ c − 1/2 R d bal ( ρ , v ) ∥ 0 ≤ ϑ η p , c LS ( τ , w ) η d , c LS ( ρ , v ) .
Corollary 3.9. Let c ≥ c 0 > 0 c\geq c_{0}>0 c ≥ c 0 > 0 , ( τ , w ) ∈ Σ p × X p (\tau,w)\in\Sigma_{p}\times X_{p} ( τ , w ) ∈ Σ p × X p and ( ρ , v ) ∈ Σ d × X d (\rho,v)\in\Sigma_{d}\times X_{d} ( ρ , v ) ∈ Σ d × X d . Then
∣ γ − γ 2 ( τ , ρ ) ∣ ≤ C η p , c L S ( τ , w ) η d , c L S ( ρ , v ) . (32) |\gamma-\gamma_{2}(\tau,\rho)|\leq C\eta^{\mathrm{LS}}_{p,c}(\tau,w)\eta^{\mathrm{LS}}_{d,c}(\rho,v). \tag*{(32)} ∣ γ − γ 2 ( τ , ρ ) ∣ ≤ C η p , c LS ( τ , w ) η d , c LS ( ρ , v ) . (32) Proof. We bound each term of (3.9) by the Cauchy–Schwarz inequality, using ∣ c − 1 / 2 b ⋅ A − 1 E p ∣ ≤ ϑ ∣ A − 1 / 2 E p ∣ |c^{-1/2}\mathbf{b}\cdot A^{-1}E_{p}|\leq\vartheta|A^{-1/2}E_{p}| ∣ c − 1/2 b ⋅ A − 1 E p ∣ ≤ ϑ ∣ A − 1/2 E p ∣ , and use (2.14). Since c 0 ≤ c ≤ ∥ c ∥ 0 , ∞ c_{0}\leq c\leq\lVert c\rVert_{0,\infty} c 0 ≤ c ≤ ∥ c ∥ 0 , ∞ ,
min { 1 , ∥ c ∥ 0 , ∞ } η p L S ( τ , w ) 2 ≤ η p , c L S ( τ , w ) 2 ≤ max { 1 , c 0 − 1 } η p L S ( τ , w ) 2 , \min\{1,\lVert c\rVert_{0,\infty}\}\eta^{\mathrm{LS}}_{p}(\tau,w)^{2}\leq\eta^{\mathrm{LS}}_{p,c}(\tau,w)^{2}\leq\max\{1,c_{0}^{-1}\}\eta^{\mathrm{LS}}_{p}(\tau,w)^{2}, min { 1 , ∥ c ∥ 0 , ∞ } η p LS ( τ , w ) 2 ≤ η p , c LS ( τ , w ) 2 ≤ max { 1 , c 0 − 1 } η p LS ( τ , w ) 2 , and likewise for the adjoint estimators. □
Relation to other approaches
A posteriori error estimation rests on two principles. The first compares two approximations of the same quantity, as when we compare a quadrature rule on a mesh with the same rule on its refinement, and gives an estimate. The second evaluates the residual, as we do for a linear system A x = b Ax=b A x = b through A ( x − x h ) = b − A x h A(x-x_{h})=b-Ax_{h} A ( x − x h ) = b − A x h , and gives a reliable bound, possibly with an overestimating constant [24 ] . The dual-weighted residual estimate is an instance of the first principle and the product bound of the second. Recovery-based estimators belong to the first, since they compare the flux of the numerical potential with a recovered flux, and residual-based estimators to the second. The least-squares functional contains both. Its constitutive residual R p c s t ( τ , w ) R^{\mathrm{cst}}_{p}(\tau,w) R p cst ( τ , w ) is the difference of the computed flux τ \tau τ and the flux − ( A ∇ w + f 2 ) -(A\nabla w+\mathbf{f}_{2}) − ( A ∇ w + f 2 ) of the potential, as in the first principle. Its balance residual R p b a l ( τ , w ) R^{\mathrm{bal}}_{p}(\tau,w) R p bal ( τ , w ) is a residual in the sense of the second. Together they are reliable for arbitrary arguments (Proposition 2.1). In this paper we apply both principles to the same residuals. The QoI gap compares the two approximations γ 0 ( w , v ) \gamma_{0}(w,v) γ 0 ( w , v ) and γ 1 ( w , τ , v , ρ ) \gamma_{1}(w,\tau,v,\rho) γ 1 ( w , τ , v , ρ ) of γ \gamma γ without assuming that either of them is the more accurate, and it estimates the larger of their errors. The norms of the residuals, the least-squares estimators, give the reliable upper bound, and we use them to refine the mesh.
The dual-weighted residual estimate and the QoI gap
We build the dual-weighted residual estimate and the QoI gap in the same way: we compute γ \gamma γ twice and take the difference of the two values as the estimate. The dual-weighted residual method evaluates γ \gamma γ twice, with the same potential w w w and two adjoint approximations, v v v and a more accurate v ~ ∈ X d \widetilde{v}\in X_{d} v ∈ X d , computed in an enriched space or by patchwise higher-order interpolation of v v v [4 §5.1] :
η D W R : = γ 0 ( w , v ~ ) − γ 0 ( w , v ) = ℓ ( v ~ − v ) − a ( w , v ~ − v ) = a ( u − w , v ~ − v ) . (33) \eta_{\mathrm{DWR}}:=\gamma_{0}(w,\widetilde{v})-\gamma_{0}(w,v)=\ell(\widetilde{v}-v)-a(w,\widetilde{v}-v)=a(u-w,\widetilde{v}-v). \tag*{(33)} η DWR := γ 0 ( w , v ) − γ 0 ( w , v ) = ℓ ( v − v ) − a ( w , v − v ) = a ( u − w , v − v ) . (33)
that is, γ − γ 0 ( w , v ) = a ( u − w , z − v ) \gamma-\gamma_{0}(w,v)=a(u-w,z-v) γ − γ 0 ( w , v ) = a ( u − w , z − v ) with z − v z-v z − v replaced by v ~ − v \widetilde{v}-v v − v . If w w w is a Galerkin approximation and v v v lies in its test space, as for a Galerkin approximation v v v on the same mesh with ϕ d = 0 \phi_{d}=0 ϕ d = 0 , Galerkin orthogonality gives γ 0 ( w , v ) = m ( w ) \gamma_{0}(w,v)=m(w) γ 0 ( w , v ) = m ( w ) , so that (4.1) estimates the error of the direct output.
Since the estimate is the difference of two values of γ \gamma γ , other choices of the two values give variants of it. We do not need Galerkin orthogonality. The identity (2.16) holds for all w ∈ X p w\in X_{p} w ∈ X p and v ∈ X d v\in X_{d} v ∈ X d , so (4.1) holds for arbitrary approximations and then concerns the error of γ 0 ( w , v ) \gamma_{0}(w,v) γ 0 ( w , v ) . This value is the output m ( w ) m(w) m ( w ) corrected by the primal residual tested with v v v [19 ] , and the orthogonality only makes this correction vanish. We can also change the potentials. With a more accurate primal approximation w ~ ∈ X p \widetilde{w} \in X_p w ∈ X p , the difference γ 0 ( w ~ , v ) − γ 0 ( w , v ) = a ( w ~ − w , z − v ) \gamma_0(\widetilde{w},v)-\gamma_0(w,v)=a(\widetilde{w}-w,z-v) γ 0 ( w , v ) − γ 0 ( w , v ) = a ( w − w , z − v ) replaces u − w u-w u − w by w ~ − w \widetilde{w}-w w − w , and we can average the two one-sided differences. With both, γ 0 ( w ~ , v ~ ) − γ 0 ( w , v ) = γ − γ 0 ( w , v ) − a ( u − w ~ , z − v ~ ) \gamma_0(\widetilde{w},\widetilde{v})-\gamma_0(w,v)=\gamma-\gamma_0(w,v)-a(u-\widetilde{w},z-\widetilde{v}) γ 0 ( w , v ) − γ 0 ( w , v ) = γ − γ 0 ( w , v ) − a ( u − w , z − v ) . Finally, we can keep the potentials and use fluxes instead of more accurate potentials. If we replace the gradients ∇ w \nabla w ∇ w and A ∇ v + b v A\nabla v+bv A ∇ v + b v in γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) by − A − 1 ( τ + f 2 ) -A^{-1}(\tau+f_2) − A − 1 ( τ + f 2 ) and − ( ρ + g 2 ) -(\rho+g_2) − ( ρ + g 2 ) and add the two residual corrections of (17) , we obtain γ 1 ( w , τ , v , ρ ) \gamma_1(w,\tau,v,\rho) γ 1 ( w , τ , v , ρ ) , and the difference of the two values is the signed QoI gap δ ( w , τ , v , ρ ) \delta(w,\tau,v,\rho) δ ( w , τ , v , ρ ) .
The dual-weighted residual estimate and the QoI gap are both differences of two values of γ \gamma γ , so Proposition 3.4 applies to both. The two estimates differ in what they assume. The dual-weighted residual estimate η D W R = γ − γ 0 ( w , v ) − a ( u − w , z − v ~ ) \eta_{\mathrm{DWR}}=\gamma-\gamma_0(w,v)-a(u-w,z-\widetilde{v}) η DWR = γ − γ 0 ( w , v ) − a ( u − w , z − v ) estimates γ − γ 0 ( w , v ) \gamma-\gamma_0(w,v) γ − γ 0 ( w , v ) when γ 0 ( w , v ~ ) \gamma_0(w,\widetilde{v}) γ 0 ( w , v ) is much more accurate than γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) , while the QoI gap does not assume that either of its two values is the more accurate.
The QoI gap is a generalized dual-weighted residual estimate with flux-based weights. In γ − γ 0 ( w , v ) = ( ∇ e p , A ∇ e d + b e d ) + ( c e p , e d ) \gamma-\gamma_0(w,v)=(\nabla e_p,A\nabla e_d+be_d)+(ce_p,e_d) γ − γ 0 ( w , v ) = ( ∇ e p , A ∇ e d + b e d ) + ( c e p , e d ) , the dual-weighted residual estimate replaces z − v z-v z − v by v ~ − v \widetilde{v}-v v − v . The QoI gap replaces ∇ e p \nabla e_p ∇ e p by − A − 1 R p c s t ( τ , w ) = ∇ e p + A − 1 E p -A^{-1}R_p^{\mathrm{cst}}(\tau,w)=\nabla e_p+A^{-1}E_p − A − 1 R p cst ( τ , w ) = ∇ e p + A − 1 E p and A ∇ e d + b e d A\nabla e_d+be_d A ∇ e d + b e d by − R d c s t ( ρ , v ) = A ∇ e d + b e d + E d -R_d^{\mathrm{cst}}(\rho,v)=A\nabla e_d+be_d+E_d − R d cst ( ρ , v ) = A ∇ e d + b e d + E d , which are exact when the fluxes are exact. It also omits the reaction term, and the result is δ ( w , τ , v , ρ ) \delta(w,\tau,v,\rho) δ ( w , τ , v , ρ ) . The weights come from the fluxes, on both the primal and the adjoint side, and the fluxes need not be more accurate than the gradients of the potentials.
The QoI gap and flux recovery
For each problem separately, the constitutive residual is the correction that a flux recovery makes. The flux − ( A ∇ w + f 2 ) -(A\nabla w+f_2) − ( A ∇ w + f 2 ) of the numerical potential is in general discontinuous, τ \tau τ is a flux computed in H ( div ; Ω ) H(\operatorname{div};\Omega) H ( div ; Ω ) , and R p c s t ( τ , w ) = τ + A ∇ w + f 2 R_p^{\mathrm{cst}}(\tau,w)=\tau+A\nabla w+f_2 R p cst ( τ , w ) = τ + A ∇ w + f 2 is the correction that τ \tau τ makes to − ( A ∇ w + f 2 ) -(A\nabla w+f_2) − ( A ∇ w + f 2 ) , as a field, before we take any norm. Recovery-based estimators take a norm of it. They are the most widely used a posteriori estimators in practice. The best known is the Zienkiewicz–Zhu estimator [39 , 40 ] , which recovers a continuous gradient by local averaging or patch projection. For discontinuous coefficients the reconstruction should be H ( div ) H(\operatorname{div}) H ( div ) -conforming rather than continuous, since only the normal component of the flux is continuous across interfaces [12 , 13 ] . The least-squares recovery of [14 ] gives such a flux, and the non-intrusive estimator of [24 ] uses it. In this flux recovery the recovered flux solves a problem posed in another space, H ( div ) H(\operatorname{div}) H ( div ) . It need not approximate σ \sigma σ better than the flux − ( A ∇ w + f 2 ) -(A\nabla w+f_2) − ( A ∇ w + f 2 ) of the potential. The reason is that the least-squares functional, which adds the balance residual to ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 2 \lVert A^{-1/2}R_p^{\mathrm{cst}}(\tau,w)\rVert_0^2 ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 2 , is reliable and efficient by (14) for arbitrary arguments, so we need no superconvergence assumption.
The QoI gap uses the corrections themselves rather than their norms. By (18) , it pairs the primal and the adjoint flux corrections with their signs, so it is a recovery-based estimate for the quantity of interest. It has the same properties as this flux recovery: the fluxes come from another space and need not be more accurate than the gradients of the potentials, and (18) and (19) hold for arbitrary arguments. To combine two recovery estimators into a goal-oriented indicator, a common choice is the sum over the elements of the products of the local primal and adjoint indicators. This sum is never negative, and by the Cauchy–Schwarz inequality it does not exceed η p L S ( τ , w ) η d L S ( ρ , v ) \eta_p^{\mathrm{LS}}(\tau,w)\eta_d^{\mathrm{LS}}(\rho,v) η p LS ( τ , w ) η d LS ( ρ , v ) , which bounds the errors of the two evaluators up to a constant (Corollary 3.3). Unlike this sum, the QoI gap keeps the signs of the two recovery fields.
Symmetric problems and the primal–dual gap
For b = 0 \mathbf{b}=0 b = 0 the primal and adjoint problems are convex minimization problems, and the primal–dual theory of [36 ] gives estimators without any least-squares formulation. Let c ≥ c 0 > 0 c \geq c_0 > 0 c ≥ c 0 > 0 . The primal problem minimizes
J p ( w ) : = 1 2 ( A ∇ w , ∇ w ) + 1 2 ( c w , w ) − ( f 1 , w ) + ( f 2 , ∇ w ) + ⟨ ψ p , w ⟩ Γ N J_p(w):=\frac{1}{2}(A\nabla w,\nabla w)+\frac{1}{2}(cw,w)-(f_1,w)+(f_2,\nabla w)+\langle\psi_p,w\rangle_{\Gamma_N} J p ( w ) := 2 1 ( A ∇ w , ∇ w ) + 2 1 ( c w , w ) − ( f 1 , w ) + ( f 2 , ∇ w ) + ⟨ ψ p , w ⟩ Γ N
over X p X_p X p , and its dual problem in the sense of convex duality, posed over the fluxes, maximizes
J p ∗ ( τ ) : = − 1 2 ( A − 1 ( τ + f 2 ) , τ + f 2 ) − 1 2 ( c − 1 ( f 1 − div τ ) , f 1 − div τ ) − ⟨ τ ⋅ n , ϕ p ⟩ Γ D J_p^*(\tau):=-\frac{1}{2}(A^{-1}(\tau+f_2),\tau+f_2)-\frac{1}{2}(c^{-1}(f_1-\operatorname{div}\tau),f_1-\operatorname{div}\tau)-\langle\tau\cdot n,\phi_p\rangle_{\Gamma_D} J p ∗ ( τ ) := − 2 1 ( A − 1 ( τ + f 2 ) , τ + f 2 ) − 2 1 ( c − 1 ( f 1 − div τ ) , f 1 − div τ ) − ⟨ τ ⋅ n , ϕ p ⟩ Γ D
over all of Σ p \Sigma_p Σ p . We impose no equilibration constraint on τ \tau τ . Integration by parts gives, for all w ∈ X p w\in X_p w ∈ X p and τ ∈ Σ p \tau\in\Sigma_p τ ∈ Σ p ,
J p ( w ) − J p ∗ ( τ ) = 1 2 η p , c L S ( τ , w ) 2 = 1 2 ( ∥ A 1 / 2 ∇ e p ∥ 0 2 + ∥ c 1 / 2 e p ∥ 0 2 ) + 1 2 ( ∥ A − 1 / 2 E p ∥ 0 2 + ∥ c − 1 / 2 div E p ∥ 0 2 ) , J_p(w)-J_p^*(\tau)=\frac{1}{2}\eta_{p,c}^{\mathrm{LS}}(\tau,w)^2=\frac{1}{2}\left(\lVert A^{1/2}\nabla e_p\rVert_0^2+\lVert c^{1/2}e_p\rVert_0^2\right)+\frac{1}{2}\left(\lVert A^{-1/2}E_p\rVert_0^2+\lVert c^{-1/2}\operatorname{div}E_p\rVert_0^2\right), J p ( w ) − J p ∗ ( τ ) = 2 1 η p , c LS ( τ , w ) 2 = 2 1 ( ∥ A 1/2 ∇ e p ∥ 0 2 + ∥ c 1/2 e p ∥ 0 2 ) + 2 1 ( ∥ A − 1/2 E p ∥ 0 2 + ∥ c − 1/2 div E p ∥ 0 2 ) ,
with η p , c L S \eta_{p,c}^{\mathrm{LS}} η p , c LS of (29) . This is the generalized Prager–Synge identity of [36 ] , and the same identity holds for the adjoint problem. Since a a a is then the energy inner product, the Cauchy–Schwarz inequality, first in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) and then in R 2 \mathbb{R}^2 R 2 , applied to (16) and (26) , gives the bound
∣ γ − γ 0 ( w , v ) ∣ + ∣ γ − γ 2 ( τ , ρ ) ∣ ≤ η p , c L S ( τ , w ) η d , c L S ( ρ , v ) , |\gamma-\gamma_0(w,v)|+|\gamma-\gamma_2(\tau,\rho)|\leq\eta^{\mathrm{LS}}_{p,c}(\tau,w)\eta^{\mathrm{LS}}_{d,c}(\rho,v), ∣ γ − γ 0 ( w , v ) ∣ + ∣ γ − γ 2 ( τ , ρ ) ∣ ≤ η p , c LS ( τ , w ) η d , c LS ( ρ , v ) ,
and the bound of Proposition 3.8 for ∣ δ 2 ( w , τ , v , ρ ) ∣ |\delta_2(w,\tau,v,\rho)| ∣ δ 2 ( w , τ , v , ρ ) ∣ follows from it. For c = 0 c=0 c = 0 the dual variables must satisfy div τ = f 1 \operatorname{div}\tau=f_1 div τ = f 1 and div ρ = g 1 \operatorname{div}\rho=g_1 div ρ = g 1 , that is, they must be equilibrated, as in the bounds for outputs of [30 , 34 , 31 ] .
Then J p ( w ) − J p ∗ ( τ ) = 1 2 ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 2 J_p(w)-J_p^*(\tau)=\frac{1}{2}\lVert A^{-1/2}R_p^{\mathrm{cst}}(\tau,w)\rVert_0^2 J p ( w ) − J p ∗ ( τ ) = 2 1 ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 2 , and in the same way
∣ γ − γ 0 ( w , v ) ∣ + ∣ γ − γ 1 ( w , τ , v , ρ ) ∣ ≤ ∥ A − 1 / 2 R p c s t ( τ , w ) ∥ 0 ∥ A − 1 / 2 R d c s t ( ρ , v ) ∥ 0 . |\gamma-\gamma_0(w,v)|+|\gamma-\gamma_1(w,\tau,v,\rho)|\leq\lVert A^{-1/2}R_p^{\mathrm{cst}}(\tau,w)\rVert_0\lVert A^{-1/2}R_d^{\mathrm{cst}}(\rho,v)\rVert_0. ∣ γ − γ 0 ( w , v ) ∣ + ∣ γ − γ 1 ( w , τ , v , ρ ) ∣ ≤ ∥ A − 1/2 R p cst ( τ , w ) ∥ 0 ∥ A − 1/2 R d cst ( ρ , v ) ∥ 0 .
In these cases we do not need the least-squares functional. We can solve the primal problem and its dual problem separately, by any methods, while the least-squares functional couples the potential and the flux. So we can take the flux from the dual problem, for instance from a mixed method, which for c = 0 c=0 c = 0 gives the equilibrated flux required there. If we do not want to solve the mixed problem, we can use the least-squares recovery instead, with the least-squares functional in place of the primal–dual gap. For c > 0 c>0 c > 0 , by the identity above, minimizing η p , c L S ( τ , w ) 2 \eta^{\mathrm{LS}}_{p,c}(\tau,w)^2 η p , c LS ( τ , w ) 2 over τ ∈ Σ p \tau\in\Sigma_p τ ∈ Σ p for a fixed w ∈ X p w\in X_p w ∈ X p maximizes J p ∗ J_p^* J p ∗ , so the weighted least-squares recovery is the discrete dual problem and does not depend on w w w . We need the least-squares functional when the problem is not a convex minimization problem, as for b ≠ 0 b\neq0 b = 0 , where no primal–dual gap is available. For this reason we use the least-squares estimators in this paper. The QoI gap identity (18) and the bound (22) do not use the minimization structure.
Implementation
Theorem 3.1 makes no assumption on how we compute ( τ , w ) (\tau,w) ( τ , w ) and ( ρ , v ) (\rho,v) ( ρ , v ) or on how they are related. We can compute the QoI gap whenever the four fields exist. For the bound we only evaluate the least-squares functionals at these fields, so we do not need a least-squares discretization for either. For the upper bound to be good, however, all four fields must be good approximations. By (14) , η L S ( τ , w ) \eta^{\mathrm{LS}}(\tau,w) η LS ( τ , w ) is equivalent to the error of the flux and the potential together, so a poor flux makes it large even when the potential is accurate, and likewise for the adjoint pair. The fluxes and the potentials should therefore have the same approximation order. We obtain the fields from a least-squares finite element method, which computes them directly, or, for an arbitrary discretization, from the non-intrusive recovery of [14 , 24 ] or from another method that computes a flux, such as a mixed method. The four fields must also converge. The bounds hold for arbitrary arguments. They bound the errors of the fields, but they do not show that these errors tend to zero. For a Galerkin potential with the recovered flux, refined with its own estimator, [24 §7] proves plain convergence of the adaptive algorithm. The proof uses a priori quasi-optimality of the Galerkin method and of the recovery together with the reliability and efficiency of the estimator. The a posteriori bounds alone do not give convergence.
Let T h \mathcal{T}_h T h and T H \mathcal{T}_H T H be conforming, shape-regular triangulations of Ω \Omega Ω into simplices, with elements K K K , for the primal and the adjoint problem. The two triangulations need not be related. Throughout, X h ⊂ H 1 ( Ω ) X_h\subset H^1(\Omega) X h ⊂ H 1 ( Ω ) denotes a C 0 C^0 C 0 Lagrange space and Σ h ⊂ H ( div ; Ω ) \Sigma_h\subset H(\operatorname{div};\Omega) Σ h ⊂ H ( div ; Ω ) a Raviart–Thomas space on T h \mathcal{T}_h T h , and X H X_H X H and Σ H \Sigma_H Σ H the corresponding spaces on T H \mathcal{T}_H T H . We write S k S_k S k for the continuous piecewise polynomials of degree k k k , P 0 P_0 P 0 for the piecewise constants and R T 0 RT_0 R T 0 for the lowest-order Raviart–Thomas space. The affine subsets X h p X_h^p X h p and Σ h p \Sigma_h^p Σ h p carry the discrete primal boundary conditions, X H d X_H^d X H d and Σ H d \Sigma_H^d Σ H d the adjoint ones, and Σ h 0 : = Σ h ∩ H N ( div ; Ω ) \Sigma_h^0:=\Sigma_h\cap H_N(\operatorname{div};\Omega) Σ h 0 := Σ h ∩ H N ( div ; Ω ) and Σ H 0 : = Σ H ∩ H N ( div ; Ω ) \Sigma_H^0:=\Sigma_H\cap H_N(\operatorname{div};\Omega) Σ H 0 := Σ H ∩ H N ( div ; Ω ) are the linear subspaces.
Fluxes from a least-squares finite element method
If the underlying method is a first-order system least-squares method, the fluxes are part of the discrete solution. We compute
( σ h , u h ) = arg min ( τ , w ) ∈ Σ h p × X h p LS p ( τ , w ) , ( r H , z H ) = arg min ( ρ , v ) ∈ Σ H d × X H d LS d ( ρ , v ) , (34) (\sigma_h,u_h)=\underset{(\tau,w)\in\Sigma_h^p\times X_h^p}{\operatorname{arg\,min}}\ \operatorname{LS}_p(\tau,w),
\qquad
(r_H,z_H)=\underset{(\rho,v)\in\Sigma_H^d\times X_H^d}{\operatorname{arg\,min}}\ \operatorname{LS}_d(\rho,v), \tag*{(34)} ( σ h , u h ) = ( τ , w ) ∈ Σ h p × X h p arg min LS p ( τ , w ) , ( r H , z H ) = ( ρ , v ) ∈ Σ H d × X H d arg min LS d ( ρ , v ) , (34)
and insert them in (20) . This is the setting of [35 ] , which introduced γ 1 \gamma_1 γ 1 , used it together with γ 0 \gamma_0 γ 0 , and proved bounds of the form (22) . The estimators η p L S ( σ h , u h ) \eta_p^{\mathrm{LS}}(\sigma_h,u_h) η p LS ( σ h , u h ) and η d L S ( r H , z H ) \eta_d^{\mathrm{LS}}(r_H,z_H) η d LS ( r H , z H ) needed in Corollary 3.2 are the square roots of the minimum values. The same holds for any method whose unknowns are a conforming flux–potential pair, such as the augmented mixed methods of [27 , 25 ] , with the least-squares functionals evaluated at the computed pairs rather than minimized.
Non-intrusive least-squares flux recovery for an arbitrary potential
A conforming Galerkin method, or any other method that returns only a potential, gives no flux. We obtain one by minimizing the least-squares functional over the flux alone, with the computed potential fixed: given u h ∈ X h p u_h \in X_h^p u h ∈ X h p and z H ∈ X H d z_H \in X_H^d z H ∈ X H d , we set
σ h : = arg min τ ∈ Σ h p L S p ( τ , u h ) , r H : = arg min ρ ∈ Σ H d L S d ( ρ , z H ) . (35) \sigma_h := \underset{\tau\in\Sigma_h^p}{\operatorname{arg\,min}}\,\mathrm{LS}_p(\tau,u_h), \qquad r_H := \underset{\rho\in\Sigma_H^d}{\operatorname{arg\,min}}\,\mathrm{LS}_d(\rho,z_H). \tag*{(35)} σ h := τ ∈ Σ h p arg min LS p ( τ , u h ) , r H := ρ ∈ Σ H d arg min LS d ( ρ , z H ) . (35)
This is the least-squares flux recovery [14 , 24 ] discussed in § 4.2 . The first-order conditions are
( A − 1 R p c s t ( σ h , u h ) , χ ) + ( R p b a l ( σ h , u h ) , div χ ) = 0 for all χ ∈ Σ h 0 , (A^{-1}R_p^{\mathrm{cst}}(\sigma_h,u_h),\chi)+(R_p^{\mathrm{bal}}(\sigma_h,u_h),\operatorname{div}\chi)=0 \quad\text{for all }\chi\in\Sigma_h^0, ( A − 1 R p cst ( σ h , u h ) , χ ) + ( R p bal ( σ h , u h ) , div χ ) = 0 for all χ ∈ Σ h 0 ,
that is
b f l u x ( σ h , χ ) = − ( A − 1 ( f 2 + A ∇ u h ) , χ ) − ( b ⋅ ∇ u h + c u h − f 1 , div χ ) , (36) b_{\mathrm{flux}}(\sigma_h,\chi)=-(A^{-1}(f_2+A\nabla u_h),\chi)-(b\cdot\nabla u_h+cu_h-f_1,\operatorname{div}\chi), \tag*{(36)} b flux ( σ h , χ ) = − ( A − 1 ( f 2 + A ∇ u h ) , χ ) − ( b ⋅ ∇ u h + c u h − f 1 , div χ ) , (36)
and, for all χ ∈ Σ H 0 \chi\in\Sigma_H^0 χ ∈ Σ H 0 ,
b f l u x ( r H , χ ) = − ( A − 1 ( g 2 + A ∇ z H + b z H ) , χ ) − ( c z H − g 1 , div χ ) , (37) b_{\mathrm{flux}}(r_H,\chi)=-(A^{-1}(g_2+A\nabla z_H+bz_H),\chi)-(cz_H-g_1,\operatorname{div}\chi), \tag*{(37)} b flux ( r H , χ ) = − ( A − 1 ( g 2 + A ∇ z H + b z H ) , χ ) − ( c z H − g 1 , div χ ) , (37)
with, in both cases, the A A A -weighted H ( div ) H(\operatorname{div}) H ( div ) inner product
b f l u x ( μ , χ ) : = ( A − 1 μ , χ ) + ( div μ , div χ ) , b f l u x ( μ , μ ) = ∥ ∥ ( μ , 0 ) ∥ ∥ 2 . (38) b_{\mathrm{flux}}(\mu,\chi):=(A^{-1}\mu,\chi)+(\operatorname{div}\mu,\operatorname{div}\chi), \qquad b_{\mathrm{flux}}(\mu,\mu)=\left\Vert\left\Vert(\mu,0)\right\Vert\right\Vert^2. \tag*{(38)} b flux ( μ , χ ) := ( A − 1 μ , χ ) + ( div μ , div χ ) , b flux ( μ , μ ) = ∥ ∥ ( μ , 0 ) ∥ ∥ 2 . (38)
Both problems are therefore uniquely solvable. The potentials enter only the right-hand sides, so we can use potentials in H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) from any method and on any mesh in place of u h u_h u h and z H z_H z H . The error of the recovered flux in H ( div ; Ω ) H(\operatorname{div};\Omega) H ( div ; Ω ) is bounded by the best approximation error of σ \sigma σ in Σ h p \Sigma_h^p Σ h p plus a constant times the H 1 H^1 H 1 error of u h u_h u h . This is [24 Thm. 5.1] for Γ N = ∅ \Gamma_N=\varnothing Γ N = ∅ , and the following argument covers Γ N ≠ ∅ \Gamma_N\ne\varnothing Γ N = ∅ . Since R p c s t ( σ , u ) = R p b a l ( σ , u ) = 0 R_p^{\mathrm{cst}}(\sigma,u)=R_p^{\mathrm{bal}}(\sigma,u)=0 R p cst ( σ , u ) = R p bal ( σ , u ) = 0 , the exact flux satisfies (36) with u h u_h u h replaced by u u u , so that
b f l u x ( σ − σ h , χ ) = − ( ∇ ( u − u h ) , χ ) − ( b ⋅ ∇ ( u − u h ) + c ( u − u h ) , div χ ) ∀ χ ∈ Σ h 0 . b_{\mathrm{flux}}(\sigma-\sigma_h,\chi)=-(\nabla(u-u_h),\chi)-(b\cdot\nabla(u-u_h)+c(u-u_h),\operatorname{div}\chi) \qquad\forall\chi\in\Sigma_h^0. b flux ( σ − σ h , χ ) = − ( ∇ ( u − u h ) , χ ) − ( b ⋅ ∇ ( u − u h ) + c ( u − u h ) , div χ ) ∀ χ ∈ Σ h 0 .
Let P h σ ∈ Σ h p P_h\sigma\in\Sigma_h^p P h σ ∈ Σ h p be the b f l u x b_{\mathrm{flux}} b flux -orthogonal projection of σ \sigma σ , that is, b f l u x ( σ − P h σ , χ ) = 0 b_{\mathrm{flux}}(\sigma-P_h\sigma,\chi)=0 b flux ( σ − P h σ , χ ) = 0 for all χ ∈ Σ h 0 \chi\in\Sigma_h^0 χ ∈ Σ h 0 . We test with χ = P h σ − σ h \chi=P_h\sigma-\sigma_h χ = P h σ − σ h and obtain ∥ ∥ ( P h σ − σ h , 0 ) ∥ ∥ ≤ C ∥ ∥ ( 0 , u − u h ) ∥ ∥ \left\Vert\left\Vert(P_h\sigma-\sigma_h,0)\right\Vert\right\Vert\le C\left\Vert\left\Vert(0,u-u_h)\right\Vert\right\Vert ∥ ∥ ( P h σ − σ h , 0 ) ∥ ∥ ≤ C ∥ ∥ ( 0 , u − u h ) ∥ ∥ , with C C C depending on A A A , b b b and c c c . Since ∥ ∥ ( σ − P h σ , 0 ) ∥ ∥ \left\Vert\left\Vert(\sigma-P_h\sigma,0)\right\Vert\right\Vert ∥ ∥ ( σ − P h σ , 0 ) ∥ ∥ is the best approximation error, the triangle inequality gives the bound. This bound is of the lower of the two approximation orders, which is another reason to choose the flux space and the potential space of the same approximation order. We use R T 0 \mathrm{RT}_0 RT 0 with S 1 \mathcal{S}_1 S 1 throughout.
The flux need not come from the least-squares recovery. We can use any flux in Σ p \Sigma_p Σ p , for instance that of a mixed method for the primal problem, and likewise for the adjoint problem. For b = 0 b=0 b = 0 this is the setting of § 4.3 , in which the primal problem and its dual problem are decoupled and we compute the potential and the flux separately.
Remark 5.1 (Recovering the potential). For a method that delivers a flux but no conforming potential, such as a mixed method, we can obtain a potential in the same way, by minimizing L S p ( σ h , ⋅ ) \mathrm{LS}_p(\sigma_h,\cdot) LS p ( σ h , ⋅ ) over X h p X_h^p X h p . This is a second-order problem with the symmetric positive definite form ( A ∇ w , ∇ v ) + ( b ⋅ ∇ w + c w , b ⋅ ∇ v + c v ) (A\nabla w,\nabla v)+(b\cdot\nabla w+cw,b\cdot\nabla v+cv) ( A ∇ w , ∇ v ) + ( b ⋅ ∇ w + c w , b ⋅ ∇ v + c v ) . There are local alternatives. A direct interpolation of the discontinuous potential gives a conforming one. Where only the gradient is needed, as in the constitutive residual R p c s t ( τ , w ) R_p^{\mathrm{cst}}(\tau,w) R p cst ( τ , w ) , a recovery in H ( curl ) H(\operatorname{curl}) H ( curl ) as in [8 ] gives an approximation of ∇ u \nabla u ∇ u without a potential, but the evaluators also need w w w itself.
The adaptive procedure
On each mesh we compute the primal and adjoint approximations ( σ h , u h ) (\sigma_h,u_h) ( σ h , u h ) and ( r h , z h ) (r_h,z_h) ( r h , z h ) , by a least-squares finite element method or by another method followed by the flux recovery (35) , together with the least-squares estimators. When we solve both problems on the same mesh, we mark elements by Dörfler’s strategy applied to the balanced indicator of [35 ] ,
η K 2 : = 1 2 ( η d L S η p L S ( η p , K L S ) 2 + η p L S η d L S ( η d , K L S ) 2 ) , K ∈ T h , (39) \eta_K^2:=\frac{1}{2}\left(\frac{\eta_{d}^{\mathrm{LS}}}{\eta_{p}^{\mathrm{LS}}}(\eta_{p,K}^{\mathrm{LS}})^2+\frac{\eta_{p}^{\mathrm{LS}}}{\eta_{d}^{\mathrm{LS}}}(\eta_{d,K}^{\mathrm{LS}})^2\right), \qquad K\in\mathcal{T}_h, \tag*{(39)} η K 2 := 2 1 ( η p LS η d LS ( η p , K LS ) 2 + η d LS η p LS ( η d , K LS ) 2 ) , K ∈ T h , (39)
where η p L S = η p L S ( σ h , u h ) \eta_p^{\mathrm{LS}}=\eta_p^{\mathrm{LS}}(\sigma_h,u_h) η p LS = η p LS ( σ h , u h ) , η d L S = η d L S ( r h , z h ) \eta_d^{\mathrm{LS}}=\eta_d^{\mathrm{LS}}(r_h,z_h) η d LS = η d LS ( r h , z h ) , and η p , K L S \eta_{p,K}^{\mathrm{LS}} η p , K LS and η d , K L S \eta_{d,K}^{\mathrm{LS}} η d , K LS are the square roots of the least-squares functionals (12) restricted to K K K . For recovered pairs these are the local non-intrusive estimators of [24 ] . The weights in (39) are the optimal ones in Young’s inequality applied to the product η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS , so that ∑ K ∈ T h η K 2 = η p L S η d L S \sum_{K\in\mathcal{T}_h}\eta_K^2=\eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} ∑ K ∈ T h η K 2 = η p LS η d LS , similar to the weighted marking of [3 ] . On each element the balanced indicator is at least the product of the local indicators, and it is positive on every element where either of them is positive.
We stop when the difference of the two evaluators that we report is below a tolerance TOL chosen by the user. We report the interval between the two evaluators and its midpoint rather than either evaluator, because the computed quantities do not tell which of the two is the more accurate. For γ 0 \gamma_0 γ 0 and γ 1 \gamma_1 γ 1 the criterion is ∣ δ ( u h , σ h , z h , r h ) ∣ ≤ T O L |\delta(u_h,\sigma_h,z_h,r_h)| \leq\mathrm{TOL} ∣ δ ( u h , σ h , z h , r h ) ∣ ≤ TOL . Since ∣ δ ( u h , σ h , z h , r h ) ∣ ≤ η p L S ( σ h , u h ) η d L S ( r h , z h ) |\delta(u_h,\sigma_h,z_h,r_h)| \leq\eta_p^{\mathrm{LS}}(\sigma_h,u_h)\eta_d^{\mathrm{LS}}(r_h,z_h) ∣ δ ( u h , σ h , z h , r h ) ∣ ≤ η p LS ( σ h , u h ) η d LS ( r h , z h ) by Corollary 3.2, the criterion holds after finitely many steps of any refinement under which this product tends to zero. Where reaction dominates, the signed QoI gap does not contain the reaction term (Section 3.3) and can be small while the errors of γ 0 ( u h , z h ) \gamma_0(u_h,z_h) γ 0 ( u h , z h ) and γ 1 ( u h , σ h , z h , r h ) \gamma_1(u_h,\sigma_h,z_h,r_h) γ 1 ( u h , σ h , z h , r h ) are large and nearly equal. There we report γ 0 ( u h , z h ) \gamma_0(u_h,z_h) γ 0 ( u h , z h ) and γ 2 ( σ h , r h ) \gamma_2(\sigma_h,r_h) γ 2 ( σ h , r h ) and use the criterion ∣ δ 2 ( u h , σ h , z h , r h ) ∣ ≤ T O L |\delta_2(u_h,\sigma_h,z_h,r_h)| \leq\mathrm{TOL} ∣ δ 2 ( u h , σ h , z h , r h ) ∣ ≤ TOL , which ((31) ) makes reachable in the same way. We do not use γ 2 \gamma_2 γ 2 where ϑ \vartheta ϑ of Proposition 3.8 is large, since there the terms weighted by c − 1 c^{-1} c − 1 can make γ 2 ( σ h , r h ) \gamma_2(\sigma_h,r_h) γ 2 ( σ h , r h ) less accurate than γ 1 ( u h , σ h , z h , r h ) \gamma_1(u_h,\sigma_h,z_h,r_h) γ 1 ( u h , σ h , z h , r h ) (Remark 3.10). Proposition 3.4 relates the length of the reported interval to the errors of its endpoints.
Numerical experiments
Setup
We compute in two dimensions, with lowest-order elements on triangulations that we refine adaptively by newest-vertex bisection. Only the comparison with finite differences uses uniform grids. We use two kinds of pairs ( σ h , u h ) (\sigma_h,u_h) ( σ h , u h ) and ( r h , z h ) (r_h,z_h) ( r h , z h ) .
LS: the least-squares method (5.1) with R T 0 × S 1 RT_0 \times S_1 R T 0 × S 1 , for the primal and the adjoint problem (§ 5.1 );
G: conforming S 1 S_1 S 1 Galerkin approximations of u u u and z z z , with both fluxes in R T 0 RT_0 R T 0 from the least-squares recovery (5.2) (§ 5.2 ).
We also combine the G potentials with the fluxes of the mixed method R T 0 × P 0 RT_0 \times P_0 R T 0 × P 0 , and use pairs from finite difference and finite volume schemes and pairs from different methods or meshes. On each mesh we write
γ 0 , h : = γ 0 ( u h , z h ) , γ 1 , h : = γ 1 ( u h , σ h , z h , r h ) , δ h : = δ ( u h , σ h , z h , r h ) = γ 1 , h − γ 0 , h , \gamma_{0,h}:=\gamma_0(u_h,z_h), \qquad\gamma_{1,h}:=\gamma_1(u_h,\sigma_h,z_h,r_h), \qquad\delta_h:=\delta(u_h,\sigma_h,z_h,r_h)=\gamma_{1,h}-\gamma_{0,h}, γ 0 , h := γ 0 ( u h , z h ) , γ 1 , h := γ 1 ( u h , σ h , z h , r h ) , δ h := δ ( u h , σ h , z h , r h ) = γ 1 , h − γ 0 , h ,
and η p L S \eta_p^{\mathrm{LS}} η p LS , η d L S \eta_d^{\mathrm{LS}} η d LS , η p , c L S \eta_{p,c}^{\mathrm{LS}} η p , c LS , η d , c L S \eta_{d,c}^{\mathrm{LS}} η d , c LS for the estimators at ( σ h , u h ) (\sigma_h,u_h) ( σ h , u h ) and ( r h , z h ) (r_h,z_h) ( r h , z h ) . Unless stated otherwise, we solve the primal and adjoint problems on the same mesh sequence and refine by the procedure of § 5.3 , with Dörfler bulk parameter 0.5. We compute δ h \delta_h δ h by subtraction or by quadrature of the right-hand side of (20) . With the same quadrature rule for corresponding terms, the two agree up to rounding. Table 1 lists the reference values γ \gamma γ and how we compute them. On each edge of Γ N \Gamma_N Γ N we set the normal component of the fluxes in R T 0 RT_0 R T 0 to the value of ψ p \psi_p ψ p or ψ d \psi_d ψ d at the midpoint of the edge. We interpolate the Dirichlet data φ p \varphi_p φ p and φ d \varphi_d φ d at the boundary nodes, so in general u h ∉ X p u_h \notin X_p u h ∈ / X p and z h ∉ X d z_h \notin X_d z h ∈ / X d . For c > 0 c>0 c > 0 we let γ 2 , h \gamma_{2,h} γ 2 , h be the last expression in (3.10) at the computed fields and δ 2 , h : = γ 2 , h − γ 0 , h \delta_{2,h}:=\gamma_{2,h}-\gamma_{0,h} δ 2 , h := γ 2 , h − γ 0 , h . For u h ∈ X p u_h\in X_p u h ∈ X p and z h ∈ X d z_h\in X_d z h ∈ X d these are γ 2 ( σ h , r h ) \gamma_2(\sigma_h,r_h) γ 2 ( σ h , r h ) and δ 2 ( u h , σ h , z h , r h ) \delta_2(u_h,\sigma_h,z_h,r_h) δ 2 ( u h , σ h , z h , r h ) . The errors γ − γ i , h \gamma-\gamma_{i,h} γ − γ i , h , i = 0 , 1 , 2 i=0,1,2 i = 0 , 1 , 2 , then contain the same boundary term ⟨ r h ⋅ n , φ p − u h ⟩ Γ D + ⟨ σ h ⋅ n , φ d − z h ⟩ Γ D \langle r_h\cdot n,\varphi_p-u_h\rangle_{\Gamma_D}+\langle\sigma_h\cdot n,\varphi_d-z_h\rangle_{\Gamma_D} ⟨ r h ⋅ n , φ p − u h ⟩ Γ D + ⟨ σ h ⋅ n , φ d − z h ⟩ Γ D , which cancels in the QoI gaps δ h \delta_h δ h and δ 2 , h \delta_{2,h} δ 2 , h . We leave its treatment to the next paper of this series. We measure complexity by the number N N N of vertices of the mesh. The tolerance T O L \mathrm{TOL} TOL in the stopping criterion of § 5.3 is the user’s choice, and the criterion needs ∣ δ h ∣ |\delta_h| ∣ δ h ∣ to be of the size of the larger of the errors of γ 0 , h \gamma_{0,h} γ 0 , h and γ 1 , h \gamma_{1,h} γ 1 , h on whatever mesh the computation stops. For this reason we do not stop the computations at a particular tolerance. We run them to a prescribed size, given with each experiment, and compare ∣ δ h ∣ |\delta_h| ∣ δ h ∣ with the errors on every mesh.
Table 1. Quantities of interest and reference values. For P2–P7 we compute them by quadrature of the exact solutions on the subdomains where these are smooth. For P1 we use γ 0 ( w , v ) \gamma_0(w,v) γ 0 ( w , v ) with S 3 S_3 S 3 and with S 4 S_4 S 4 Galerkin approximations of the primal and adjoint problems, on meshes graded towards the four points where the hypotenuses of T p T_p T p and T d T_d T d meet ∂ Ω \partial\Omega ∂ Ω . The two values agree to 4 ⋅ 10 − 14 4\cdot10^{-14} 4 ⋅ 1 0 − 14 .
We report the signed errors γ − γ 0 , h \gamma-\gamma_{0,h} γ − γ 0 , h and γ − γ 1 , h \gamma-\gamma_{1,h} γ − γ 1 , h of the two evaluators, the QoI gap δ h \delta_h δ h , the midpoint γ ‾ h : = 1 2 ( γ 0 , h + γ 1 , h ) \overline{\gamma}_h:=\frac{1}{2}(\gamma_{0,h}+\gamma_{1,h}) γ h := 2 1 ( γ 0 , h + γ 1 , h ) , the product bound η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS and the effectivity index
eff : = ∣ δ h ∣ max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) (40) \operatorname{eff}:=\frac{|\delta_h|}{\max(|\gamma-\gamma_{0,h}|,|\gamma-\gamma_{1,h}|)} \tag*{(40)} eff := max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) ∣ δ h ∣ (40)
which is the quantity controlled by Proposition 3.4. To compare with it, we report the factor η p L S η d L S / ∣ γ − γ 0 , h ∣ \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}}/|\gamma-\gamma_{0,h}| η p LS η d LS /∣ γ − γ 0 , h ∣ by which the product bound exceeds the error of γ 0 , h \gamma_{0,h} γ 0 , h . We also record, a posteriori, which case of Proposition 3.4 occurs, and the ratio of the smaller to the larger error,
ratio : = min ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) , \operatorname{ratio}:=\frac{\min(|\gamma-\gamma_{0,h}|,|\gamma-\gamma_{1,h}|)}{\max(|\gamma-\gamma_{0,h}|,|\gamma-\gamma_{1,h}|)}, ratio := max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) min ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) ,
which is ∣ e A ∣ / ∣ e B ∣ |e_A|/|e_B| ∣ e A ∣/∣ e B ∣ in Proposition 3.4. Then eff = 1 + ratio \operatorname{eff}=1+\operatorname{ratio} eff = 1 + ratio in case (a) and eff = 1 − ratio \operatorname{eff}=1-\operatorname{ratio} eff = 1 − ratio in case (b). We take all entries of a table row from the same mesh, so that every row satisfies δ h = γ 1 , h − γ 0 , h \delta_h=\gamma_{1,h}-\gamma_{0,h} δ h = γ 1 , h − γ 0 , h .
Test problems P1–P7
We take problems P1–P6 from [35 ] , which shows the meshes and the convergence histories of the adaptive computations. We add P7 to compare equilibrated and recovered fluxes. We compute the data not stated from the exact solutions: f 1 f_1 f 1 , φ p \varphi_p φ p and ψ p \psi_p ψ p from u u u by (1) , and, in P2, g 1 g_1 g 1 , φ d \varphi_d φ d and ψ d \psi_d ψ d from z z z by (6) . All other data are zero.
P1. Mommer–Stevenson benchmark [28 ] . Ω = ( − 1 , 1 ) 2 \Omega= (-1,1)^2 Ω = ( − 1 , 1 ) 2 , Γ D = ∂ Ω \Gamma_D = \partial\Omega Γ D = ∂ Ω , A = I A = I A = I , b = 0 \mathbf b = \mathbf0 b = 0 , c = 0 c = 0 c = 0 , f 2 = ( 1 T p , 0 ) T \mathbf f_2 = (1_{T_p},0)^T f 2 = ( 1 T p , 0 ) T and g 2 = ( 1 T d , 0 ) T \mathbf g_2 = (1_{T_d},0)^T g 2 = ( 1 T d , 0 ) T , where T p T_p T p is the triangle with vertices ( 1 , 0 ) (1,0) ( 1 , 0 ) , ( 1 , 1 ) (1,1) ( 1 , 1 ) , ( 0 , 1 ) (0,1) ( 0 , 1 ) and T d T_d T d the triangle with vertices ( − 1 , − 1 ) (-1,-1) ( − 1 , − 1 ) , ( 0 , − 1 ) (0,-1) ( 0 , − 1 ) , ( − 1 , 0 ) (-1,0) ( − 1 , 0 ) . The primal and adjoint residuals concentrate near the two disjoint triangles, so a local product indicator is nearly zero everywhere. We do not know the exact solution in closed form, and Table 1 describes how we compute the reference value.
P2. Boundary layers. Ω = ( 0 , 1 ) 2 \Omega= (0,1)^2 Ω = ( 0 , 1 ) 2 , Γ D = ∂ Ω \Gamma_D = \partial\Omega Γ D = ∂ Ω , A = ϵ I A = \epsilon I A = ϵ I , b = ( b 1 , b 2 ) T = ( 2 , 3 ) T \mathbf b = (b_1,b_2)^T = (2,3)^T b = ( b 1 , b 2 ) T = ( 2 , 3 ) T , c = 1 c = 1 c = 1 , and u = ( x − e b 1 ( x − 1 ) / ϵ ) ( y 2 − e b 2 ( y − 1 ) / ϵ ) , z = ( 1 − x − e − b 1 x / ϵ ) ( ( 1 − y ) 2 − e − b 2 y / ϵ ) , u = \left(x-e^{b_1(x-1)/\epsilon}\right)\left(y^2-e^{b_2(y-1)/\epsilon}\right), \qquad z = \left(1-x-e^{-b_1x/\epsilon}\right)\left((1-y)^2-e^{-b_2y/\epsilon}\right), u = ( x − e b 1 ( x − 1 ) / ϵ ) ( y 2 − e b 2 ( y − 1 ) / ϵ ) , z = ( 1 − x − e − b 1 x / ϵ ) ( ( 1 − y ) 2 − e − b 2 y / ϵ ) , with layers at x = 1 x=1 x = 1 and y = 1 y=1 y = 1 in u u u and at x = 0 x=0 x = 0 and y = 0 y=0 y = 0 in z z z . Here ϵ = 0.1 \epsilon= 0.1 ϵ = 0.1 , except in the computations with varying c c c , where ϵ = 0.05 \epsilon= 0.05 ϵ = 0.05 . The adjoint Dirichlet datum ϕ d = z ∣ ∂ Ω \phi_d = z|_{\partial\Omega} ϕ d = z ∣ ∂ Ω is of size e − b 1 / ϵ e^{-b_1/\epsilon} e − b 1 / ϵ .
P3–P6. L-shaped domain. Ω = ( − 1 , 1 ) 2 ∖ ( [ 0 , 1 ] × [ − 1 , 0 ] ) \Omega= (-1,1)^2 \setminus([0,1] \times[-1,0]) Ω = ( − 1 , 1 ) 2 ∖ ([ 0 , 1 ] × [ − 1 , 0 ]) , Γ N = { ( x , 1 ) : − 1 < x < 1 } \Gamma_N = \{(x,1): -1 < x < 1\} Γ N = {( x , 1 ) : − 1 < x < 1 } , Γ D = ∂ Ω ∖ Γ N ‾ \Gamma_D = \partial\Omega\setminus\overline{\Gamma_N} Γ D = ∂ Ω ∖ Γ N , A = I A = I A = I , b = 0 \mathbf b = \mathbf0 b = 0 , c = 0 c = 0 c = 0 , and, in polar coordinates ( r , θ ) (r,\theta) ( r , θ ) about the origin, u = r 2 / 3 sin ( 2 θ / 3 ) u = r^{2/3}\sin(2\theta/3) u = r 2/3 sin ( 2 θ /3 ) , so that σ = − ∇ u \boldsymbol{\sigma} = -\nabla u σ = − ∇ u . The four problems share the primal problem and differ in the adjoint data: in P3, g 1 = 1 ω 0 g_1 = 1_{\omega_0} g 1 = 1 ω 0 with ω 0 = ( − 1 , − 1 2 ) 2 \omega_0 = (-1,-\frac{1}{2})^2 ω 0 = ( − 1 , − 2 1 ) 2 ; in P4, g 2 = ( 1 T 0 , 0 ) T \mathbf g_2 = (1_{T_0},0)^T g 2 = ( 1 T 0 , 0 ) T with T 0 T_0 T 0 the triangle with vertices ( − 1 , − 1 ) (-1,-1) ( − 1 , − 1 ) , ( − 1 , 0 ) (-1,0) ( − 1 , 0 ) , ( − 1 2 , − 1 2 ) (-\frac{1}{2},-\frac{1}{2}) ( − 2 1 , − 2 1 ) ; in P5, ψ d = − 1 \psi_d = -1 ψ d = − 1 ; in P6, ϕ d = y + 1 \phi_d = y+1 ϕ d = y + 1 on Γ L = { ( − 1 , y ) : − 1 < y < 1 } \Gamma_L = \{(-1,y): -1 < y < 1\} Γ L = {( − 1 , y ) : − 1 < y < 1 } and ϕ d = 0 \phi_d = 0 ϕ d = 0 on Γ D ∖ Γ L \Gamma_D \setminus\Gamma_L Γ D ∖ Γ L .
P7. Kellogg’s problem [23 ] , as in [12 ] . Ω = ( − 1 , 1 ) 2 \Omega= (-1,1)^2 Ω = ( − 1 , 1 ) 2 , Γ D = ∂ Ω \Gamma_D = \partial\Omega Γ D = ∂ Ω , b = 0 \mathbf b = \mathbf0 b = 0 , c = 0 c = 0 c = 0 , A = a I A = aI A = a I with a = R ≈ 161.4476387975881 a = R \approx161.4476387975881 a = R ≈ 161.4476387975881 in the first and third quadrants and a = 1 a = 1 a = 1 in the second and fourth, and u = r β μ ( θ ) u = r^{\beta}\mu(\theta) u = r β μ ( θ ) with β = 0.1 \beta= 0.1 β = 0.1 and μ \mu μ as in [12 ] . The solution satisfies u ∈ H 1 + t ( Ω ) u \in H^{1+t}(\Omega) u ∈ H 1 + t ( Ω ) only for t < β t < \beta t < β . The adjoint data are g 2 = ( 1 ω , 0 ) T \mathbf g_2 = (1_{\omega},0)^T g 2 = ( 1 ω , 0 ) T with ω = ( − 1 2 , 1 2 ) × ( 0 , 1 2 ) \omega= (-\frac{1}{2},\frac{1}{2}) \times(0,\frac{1}{2}) ω = ( − 2 1 , 2 1 ) × ( 0 , 2 1 ) , which straddles the interface x = 0 x=0 x = 0 , so that γ = − ( ∂ x u , 1 ω ) = ( A − 1 σ , g 2 ) \gamma= -(\partial_xu,1_{\omega}) = (A^{-1}\boldsymbol{\sigma},\mathbf g_2) γ = − ( ∂ x u , 1 ω ) = ( A − 1 σ , g 2 ) .
Least-squares pairs
Table 2 shows the values for LS pairs on the final mesh of each sequence, and Figure 1 the whole sequence of P2. We refine until the number of vertices first exceeds 2 ⋅ 10 5 2\cdot10^5 2 ⋅ 1 0 5 .
Table 2. LS pairs, final mesh of each sequence. N N N is the number of vertices, and we compute all entries of a row on the same mesh.
Figure 1. LS pairs for P2: the errors ∣ γ − γ 0 , h ∣ |\gamma- \gamma_{0,h}| ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ |\gamma- \gamma_{1,h}| ∣ γ − γ 1 , h ∣ of the two evaluators and ∣ γ − γ ˉ h ∣ |\gamma- \bar{\gamma}_h| ∣ γ − γ ˉ h ∣ of their midpoint, the QoI gap ∣ δ h ∣ |\delta_h| ∣ δ h ∣ and the product bound η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS against the number N N N of vertices.
On the final meshes case (a) occurs in every problem. Along the sequences, case (a) occurs on every mesh of P1 and P3–P6, with eff in [ 1.23 , 1.97 ] [1.23,1.97] [ 1.23 , 1.97 ] , [ 1.29 , 1.96 ] [1.29,1.96] [ 1.29 , 1.96 ] , [ 1.03 , 1.49 ] [1.03,1.49] [ 1.03 , 1.49 ] , [ 1.51 , 1.98 ] [1.51,1.98] [ 1.51 , 1.98 ] and [ 1.45 , 1.75 ] [1.45,1.75] [ 1.45 , 1.75 ] , respectively. P2, case (a) occurs on 31 of the 33 meshes, with eff ∈ [ 1.10 , 2.00 ] \in[1.10, 2.00] ∈ [ 1.10 , 2.00 ] , and case (b) on meshes 13 and 15, with eff = 0.46 and 0.44. From mesh 12 to mesh 15, γ − γ 1 , h \gamma- \gamma_{1,h} γ − γ 1 , h changes sign three times, and γ − γ 0 , h \gamma- \gamma_{0,h} γ − γ 0 , h then changes sign between meshes 15 and 16. The QoI gap changes sign once, between meshes 14 and 15, and in the other problems it does not change sign.
Over the last six meshes, the fitted rates against N N N in P1–P6 lie between −0.50 and −0.51 for η p L S \eta_p^{\mathrm{LS}} η p LS and η d L S \eta_d^{\mathrm{LS}} η d LS , between −1.00 and −1.01 for η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS , and between −0.99 and −1.11 for ∣ δ h ∣ |\delta_h| ∣ δ h ∣ . For ∣ γ − γ 0 , h ∣ |\gamma- \gamma_{0,h}| ∣ γ − γ 0 , h ∣ and ∣ γ − γ 1 , h ∣ |\gamma- \gamma_{1,h}| ∣ γ − γ 1 , h ∣ they lie between −0.85 and −1.18. On the final meshes eff lies between 1.08 and 1.72, while the product η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS exceeds ∣ γ − γ 0 , h ∣ |\gamma- \gamma_{0,h}| ∣ γ − γ 0 , h ∣ by factors between 3.2 and 295 (Table 2). In P2 this factor lies between 258 and 315 on the last ten meshes. The product exceeds ∣ δ h ∣ |\delta_h| ∣ δ h ∣ by the factors 13 and 123 in P1 and P2 and 17, 5.5, 4.4 and 2.1 in P3–P6. On the last ten meshes this factor stays within 10% of its final value in P1 and P3–P6 and lies between 88 and 157 in P2. The product and ∣ δ h ∣ |\delta_h| ∣ δ h ∣ both decrease like N − 1 N^{-1} N − 1 . If we stopped when the product, instead of ∣ δ h ∣ |\delta_h| ∣ δ h ∣ , fell below a tolerance, we would need about these factors more vertices, 123 times as many in P2. Of P1–P6, only P2 is convection-dominated, and it is the only problem in which η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS exceeds ∣ γ − γ 0 , h ∣ |\gamma- \gamma_{0,h}| ∣ γ − γ 0 , h ∣ by more than a factor 48.
P2, where c = 1 c = 1 c = 1 and ϑ = 11.4 \vartheta= 11.4 ϑ = 11.4 , shows why we do not use γ 2 \gamma_2 γ 2 when ϑ \vartheta ϑ is large. There γ 2 , h \gamma_{2,h} γ 2 , h is the least accurate evaluator on every mesh, by the factors 55 and 40 against γ 0 , h \gamma_{0,h} γ 0 , h and γ 1 , h \gamma_{1,h} γ 1 , h on the final mesh. The pair γ 0 , h , γ 2 , h \gamma_{0,h}, \gamma_{2,h} γ 0 , h , γ 2 , h is in case (b) on the first 15 meshes and in case (a) on the remaining 18. On these 18 meshes ratio ≤ 0.023 \le0.023 ≤ 0.023 and eff ∈ [ 1.003 , 1.023 ] \in[1.003, 1.023] ∈ [ 1.003 , 1.023 ] , so that ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ is the error of γ 2 , h \gamma_{2,h} γ 2 , h to within 2.3%. On the final mesh γ − γ 2 , h = 1.73 ⋅ 10 − 5 \gamma- \gamma_{2,h} = 1.73 \cdot10^{-5} γ − γ 2 , h = 1.73 ⋅ 1 0 − 5 and δ 2 , h = − 1.76 ⋅ 10 − 5 \delta_{2,h} = -1.76 \cdot10^{-5} δ 2 , h = − 1.76 ⋅ 1 0 − 5 , and the product η p , c L S η d , c L S \eta_{p,c}^{\mathrm{LS}}\eta_{d,c}^{\mathrm{LS}} η p , c LS η d , c LS , here equal to η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS , exceeds ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ by the factor 5.3, against 123 for ∣ δ h ∣ |\delta_h| ∣ δ h ∣ .
Other methods, other fluxes and unrelated pairs
In this subsection at least one of the two pairs is not a least-squares pair on the common mesh. Either the potentials come from a Galerkin, finite difference or finite volume method and we recover the fluxes or take them from a mixed method, or the two pairs come from different methods or meshes. Theorem 3.1 and Proposition 3.4 apply unchanged. We test whether the QoI gap stays accurate, and how this depends on the method that produced the potentials, on how we obtain the fluxes, and on the relation between the two pairs.
Conforming Galerkin potentials
For G pairs on a common mesh and ϕ d = 0 \phi_d = 0 ϕ d = 0 , γ 0 , h = m ( u h ) \gamma_{0,h} = m(u_h) γ 0 , h = m ( u h ) (§ 4.1 ) is the value of the quantity of interest computed directly from the Galerkin solution, and the recovered fluxes enter only through δ h \delta_h δ h , since γ 1 , h = m ( u h ) + δ h \gamma_{1,h} = m(u_h) + \delta_h γ 1 , h = m ( u h ) + δ h . In P6, where m = 0 m = 0 m = 0 and ϕ d ≠ 0 \phi_d \ne0 ϕ d = 0 , γ 0 , h = ℓ ( z h ) − a ( u h , z h ) \gamma_{0,h} = \ell(z_h) - a(u_h,z_h) γ 0 , h = ℓ ( z h ) − a ( u h , z h ) depends only on the boundary values of z h z_h z h and is the variationally consistent evaluation of the boundary flux. A dual-weighted residual estimate of γ − γ 0 , h \gamma- \gamma_{0,h} γ − γ 0 , h needs a second, enriched adjoint approximation, since by Galerkin orthogonality (33) vanishes for every v ~ \widetilde{v} v in the Galerkin space with the boundary values of z h z_h z h . The QoI gap needs the two recovered fluxes instead.
We compare the G pairs of Table 3 with the LS pairs of Table 2 . On the final meshes the case is the same in every problem, and the factor by which the product exceeds the QoI gap agrees to within 3%. The ratio of the two errors agrees to within 0.06 in all problems except P1, where it is not stable along either sequence. On the last 10 meshes of P1, ∣ γ − γ 0 , h ∣ / ∣ γ − γ 1 , h ∣ \lvert\gamma-\gamma_{0,h}\rvert/\lvert\gamma-\gamma_{1,h}\rvert ∣ γ − γ 0 , h ∣ / ∣ γ − γ 1 , h ∣ ranges over [ 0.41 , 2.4 ] [0.41, 2.4] [ 0.41 , 2.4 ] for G pairs and over [ 0.41 , 2.3 ] [0.41, 2.3] [ 0.41 , 2.3 ] for LS pairs. Along the sequences, case (a) occurs on every mesh of P1, P2, P3, P5 and P6, and on 30 of the 31 meshes of P4. The exception is mesh 12 of P4, at which the smaller error γ − γ 1 , h \gamma-\gamma_{1,h} γ − γ 1 , h changes sign, with ratio = 0.10 and eff = 0.90. On this mesh ∣ δ h ∣ \lvert\delta_h\rvert ∣ δ h ∣ underestimates max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) \max(\lvert\gamma-\gamma_{0,h}\rvert,\lvert\gamma-\gamma_{1,h}\rvert) max (∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣) by 10%. In P2 the QoI gap of the G pairs keeps its sign along the whole sequence, and no mesh of the G sequence is in case (b). As for LS pairs, γ 0 , h \gamma_{0,h} γ 0 , h is the more accurate on most meshes of P2, here on 27 of the 34. With ϕ d = 0 \phi_d = 0 ϕ d = 0 , in P1 and P3–P5, the computed γ 0 , h \gamma_{0,h} γ 0 , h agrees with m ( u h ) m(u_h) m ( u h ) on the final meshes to within 3 ⋅ 10 − 6 ∣ γ − γ ‾ h ∣ 3\cdot10^{-6}\lvert\gamma-\overline{\gamma}_h\rvert 3 ⋅ 1 0 − 6 ∣ γ − γ h ∣ .
Table 3. G pairs with least-squares flux recovery, final mesh of each sequence. We generate the meshes as for LS pairs, with the indicator (5.6) evaluated at the G pairs.
Equilibrated and recovered fluxes for Kellogg’s problem
For P7 we combine the G potentials u h u_h u h and z h z_h z h on each mesh with two kinds of fluxes in R T 0 RT_0 R T 0 :
the fluxes of the mixed method R T 0 × P 0 RT_0 \times P_0 R T 0 × P 0 for the primal and the adjoint problem, which are equilibrated, div σ h = f 1 \operatorname{div}\boldsymbol{\sigma}_h = f_1 div σ h = f 1 and div r h = g 1 \operatorname{div}\boldsymbol{r}_h = g_1 div r h = g 1 , both zero here;
the least-squares recovery (5.2) [14 ] .
We generate the meshes with the pairs (i) and the indicator (5.6). Both fluxes are H ( div ) H(\operatorname{div}) H ( div ) -conforming, as discontinuous coefficients require [12 , 15 ] , so Theorem 3.1 applies to each. On each mesh γ 0 , h \gamma_{0,h} γ 0 , h is the same for (i) and (ii), and the two QoI gaps differ only through the fluxes. We take the LS pair on the same meshes as a third variant. The balance residuals of the pairs (i) vanish here, and since b = 0 \boldsymbol{b} = \boldsymbol{0} b = 0 and c = 0 c = 0 c = 0 , P7 with the fluxes (i) is the equilibrated symmetric case of § 4.3 . We ask whether the accuracy of the QoI gap depends on how we obtain the flux.
We compute the sequence up to mesh 28, with N = 22856 N = 22856 N = 22856 , and Figure 2 shows mesh 19. Meshes refined for the energy error of u u u are radial about the origin [38 ] . Our meshes also resolve the adjoint solution, because the indicator (5.6) distributes η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS over the elements. The adjoint data g 2 = ( 1 ω , 0 ) T \boldsymbol{g}_2 = (1_\omega,0)^T g 2 = ( 1 ω , 0 ) T is discontinuous across the vertical edges of ω \omega ω , so z z z is singular at the corners of ω \omega ω . Near the corners in the second quadrant, where a = 1 a = 1 a = 1 , the meshes are much finer than near those in the first, where a = R a = R a = R . The three variants give nearly the same values on every mesh. Case (a) occurs on every mesh, with eff ∈ [ 1.08 , 1.84 ] \in[1.08,1.84] ∈ [ 1.08 , 1.84 ] . On mesh 28 the three rows of Table 4 agree in the digits shown, except γ − γ 1 , h \gamma-\gamma_{1,h} γ − γ 1 , h and ∣ γ − γ ‾ h ∣ \lvert\gamma-\overline{\gamma}_h\rvert ∣ γ − γ h ∣ in the third digit, and the exact errors of the primal fluxes agree as well, ( ∥ A − 1 / 2 E p ∥ 0 2 + ∥ div E p ∥ 1 / 2 2 ) 1 / 2 = 0.171 (\lVert A^{-1/2}E_p\rVert_0^2+\lVert\operatorname{div}E_p\rVert_{1/2}^2)^{1/2}=0.171 (∥ A − 1/2 E p ∥ 0 2 + ∥ div E p ∥ 1/2 2 ) 1/2 = 0.171 . On this problem the accuracy of the QoI gap does not depend on how we obtain the flux. The product exceeds the QoI gap only by factors between 1.7 and 2.7 on meshes 19–28, as in P6. The midpoint is more accurate than γ 0 , h \gamma_{0,h} γ 0 , h by the factor 12. The errors have not yet reached their asymptotic rates on this sequence. Over meshes 23–28 the fitted rates against N N N are −0.15 for the flux error, −0.58 for η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS , and −0.37 and −0.41 for ∣ δ h ∣ |\delta_h| ∣ δ h ∣ and max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) \max(|\gamma-\gamma_{0,h}|,|\gamma-\gamma_{1,h}|) max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) .
Table 4. P7 on the mesh sequence generated with the pairs (i), mesh 28, N = 22856 N=22856 N = 22856 . For G pairs, γ − γ 0 , h \gamma-\gamma_{0,h} γ − γ 0 , h is the same in both rows.
Figure 2. P7: adaptive mesh 19, with N = 3456 N=3456 N = 3456 . The refinement concentrates at the origin and near the corners of ω \omega ω .
Pairs from different methods and meshes, finite differences and finite volumes.
Theorem 3.1 requires no relation between the two pairs and nothing of the methods that produced them. We compute the QoI gap from
G–LS: a G primal pair and an LS adjoint pair on a common mesh, for P1–P6, with the marking of § 6.1 ;
LS–LS: LS primal and adjoint pairs on separate mesh sequences, each refined by Dörfler marking of its own least-squares functional, for P1 and P2. We pair the k k k -th meshes of the two sequences and evaluate γ 0 , h \gamma_{0,h} γ 0 , h , γ 1 , h \gamma_{1,h} γ 1 , h and δ h \delta_h δ h on their overlay, the coarsest common refinement, on which both pairs are piecewise polynomial;
FD: for P2, pairs from a finite difference scheme on uniform grids of mesh size h = 2 − 3 , … , 2 − 9 h=2^{-3},\ldots,2^{-9} h = 2 − 3 , … , 2 − 9 , without adaptivity. We apply the five-point scheme with upwinding of the convection term to the primal problem and to the adjoint problem (6) , in the latter with upwinding for − b -b − b . We take the grid values as a continuous piecewise linear function on the triangulation that cuts every grid square along the same diagonal, and obtain both fluxes in R T 0 RT_0 R T 0 by the recovery (35) , as for G pairs. For comparison we also compute G pairs on the same grids; FV: for P1–P7, potentials from the vertex-centered finite volume method (box method) on the barycentric dual mesh, with upwinding of the convection term, for the primal problem and for the adjoint problem (6) in conservative form. We take the nodal values as functions in S 1 S_1 S 1 , obtain both fluxes by the recovery (35) , and generate the meshes as for G pairs, with the indicator (39) evaluated at the finite volume pairs.
Table 5 shows the results on the final meshes.
Table 5. Pairs from different methods (G–LS), on separate meshes (LS–LS), from finite differences with G pairs on the same grid for comparison (FD), and from finite volumes (FV), final meshes, with h = 2 − 9 h=2^{-9} h = 2 − 9 for FD. N N N is the number of vertices of the common mesh, and for LS–LS the sum of the numbers of vertices of the primal and the adjoint mesh.
For the G–LS pairs the results are close to those of LS and of G pairs (Table 2 –Table 3 ). On the final meshes the case is the same, and the product exceeds the QoI gap by the factors 12.7, 142, 17.0, 5.5, 4.4 and 2.1 in P1–P6, as for LS pairs except in P2, where the factor is 123. Along the sequences, case (a) occurs on every mesh of P1, P3, P5 and P6, on 30 of the 31 meshes of P4, and in P2 from mesh 11 on. The first ten meshes of P2 are in case (b).
For the LS–LS pairs, the two sequences of P1 have the same numbers of vertices on every mesh, since the problem is symmetric under ( x , y ) ↦ ( − x , − y ) (x,y)\mapsto(-x,-y) ( x , y ) ↦ ( − x , − y ) , which exchanges T p T_p T p and T d T_d T d . Case (a) occurs on all 33 meshes, with e f f ∈ [ 1.23 , 1.97 ] \mathrm{eff}\in[1.23,1.97] eff ∈ [ 1.23 , 1.97 ] , and the product exceeds the QoI gap by factors in [ 11.5 , 15.2 ] [11.5,15.2] [ 11.5 , 15.2 ] , as for LS pairs on a common mesh, although we compute each pair on a mesh with 217872 vertices and the overlay has 390757. In P2 case (a) occurs on 30 of the 34 meshes, and case (b) on meshes 12–15, where δ h \delta_h δ h changes sign, as on a common mesh (Section 6.3). Against the total number of degrees of freedom of the two sequences, the fitted rates over the last six meshes are −0.98 and −1.07 for ∣ δ h ∣ |\delta_h| ∣ δ h ∣ and −1.00 for η p L S η d L S \eta_p^{\mathrm{LS}}\eta_d^{\mathrm{LS}} η p LS η d LS in P1 and P2.
The finite difference potentials satisfy no Galerkin relation. The scheme takes the right-hand sides at the grid points, lumps the reaction term and upwinds the convection term, and only its diffusion part is, up to the factor h 2 h^2 h 2 , the stiffness matrix of S 1 S_1 S 1 on this triangulation. Case (a) occurs on all seven grids for both kinds of pairs, with e f f ∈ [ 1.38 , 1.59 ] \mathrm{eff}\in[1.38,1.59] eff ∈ [ 1.38 , 1.59 ] for the finite difference pairs and e f f ∈ [ 1.24 , 1.36 ] \mathrm{eff}\in[1.24,1.36] eff ∈ [ 1.24 , 1.36 ] for the G pairs. The fitted rates of ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ |\gamma-\gamma_{1,h}| ∣ γ − γ 1 , h ∣ and ∣ δ h ∣ |\delta_h| ∣ δ h ∣ against N N N over the last four grids lie between −1.01 and −1.00 for both, although the nodal errors of the upwind scheme decrease only as h h h . On the final grid, ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ is 2.5 times and ∣ γ − γ 1 , h ∣ |\gamma-\gamma_{1,h}| ∣ γ − γ 1 , h ∣ 18 times that of the G pairs, and the product exceeds the QoI gap by the factor 70, against 233. For the finite difference pairs γ 0 , h \gamma_{0,h} γ 0 , h is the more accurate on every grid, with ∣ γ − γ 1 , h ∣ / ∣ γ − γ 0 , h ∣ = 2.66 |\gamma-\gamma_{1,h}|/|\gamma-\gamma_{0,h}|=2.66 ∣ γ − γ 1 , h ∣/∣ γ − γ 0 , h ∣ = 2.66 on the final grid, and for the G pairs γ 0 , h \gamma_{0,h} γ 0 , h , with ∣ γ − γ 0 , h ∣ / ∣ γ − γ 1 , h ∣ = 2.75 |\gamma-\gamma_{0,h}|/|\gamma-\gamma_{1,h}|=2.75 ∣ γ − γ 0 , h ∣/∣ γ − γ 1 , h ∣ = 2.75 . With central differences instead of upwinding, the errors, the QoI gap and the estimators agree with those of the G pairs to within 1% from h = 2 − 6 h=2^{-6} h = 2 − 6 on, so we do not list them.
For the FV pairs, the box method differs from the Galerkin method only where the problem has a convection or reaction term, or data that are not constant on the elements or on the boundary edges. In P1 and P7 its linear systems are those of the Galerkin method, since b = 0 \mathbf{b}=0 b = 0 , c = 0 c=0 c = 0 , f 1 = g 1 = 0 f_1=g_1=0 f 1 = g 1 = 0 , Γ N = ∅ \Gamma_N=\varnothing Γ N = ∅ and f 2 , g 2 f_2,g_2 f 2 , g 2 are constant on every element, and the results are those of the G pairs with least-squares recovery, in P1 those of Table 3 . In P3–P6, where it differs only in the integration of the data over the boxes and their boundary edges, the errors, the QoI gap and the product on the final meshes agree with those of Table 3 to within 2%. We list only P2. There case (a) occurs on 26 of the 28 meshes, all but the first two, with eff ∈ [ 1.02 , 1.67 ] \operatorname{eff} \in[1.02, 1.67] eff ∈ [ 1.02 , 1.67 ] . The fitted rates of ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ |\gamma-\gamma_{1,h}| ∣ γ − γ 1 , h ∣ and ∣ δ h ∣ |\delta_h| ∣ δ h ∣ over the last six meshes are − 1.05 -1.05 − 1.05 , and the product exceeds the QoI gap by factors between 11 and 14 on the last ten meshes, against 124 on the final mesh of the G pairs. As for the finite difference pairs, γ 0 , h \gamma_{0,h} γ 0 , h is the more accurate, on 25 of the 28 meshes and by the factor 6.6 on the final one.
The interval and its midpoint
Which of the two evaluators is the more accurate depends on the problem, the method and the mesh, and the computed quantities do not tell. In P2, γ 0 , h \gamma_{0,h} γ 0 , h is the more accurate on 20 of the 33 meshes of the LS sequence and γ 1 , h \gamma_{1,h} γ 1 , h on the other 13, and on the last ten meshes ∣ γ − γ 0 , h ∣ / ∣ γ − γ 1 , h ∣ |\gamma-\gamma_{0,h}|/|\gamma-\gamma_{1,h}| ∣ γ − γ 0 , h ∣/∣ γ − γ 1 , h ∣ lies between 0.49 and 1.14. On the uniform grids of Table 5 , γ 0 , h \gamma_{0,h} γ 0 , h is the more accurate on every grid for the finite difference potentials, and γ 1 , h \gamma_{1,h} γ 1 , h for the Galerkin potentials. For this reason we report the interval between γ 0 , h \gamma_{0,h} γ 0 , h and γ 1 , h \gamma_{1,h} γ 1 , h , not one of them.
In case (a), γ \gamma γ lies in the interval, and the midpoint γ ˉ h \bar{\gamma}_h γ ˉ h has the error
∣ γ − γ ˉ h ∣ = 1 2 ∣ ∣ γ − γ 0 , h ∣ − ∣ γ − γ 1 , h ∣ ∣ = 1 2 ( 1 − ratio ) max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) . |\gamma-\bar{\gamma}_h|=\frac{1}{2}\left||\gamma-\gamma_{0,h}|-|\gamma-\gamma_{1,h}|\right|=\frac{1}{2}(1-\operatorname{ratio})\max(|\gamma-\gamma_{0,h}|,|\gamma-\gamma_{1,h}|). ∣ γ − γ ˉ h ∣ = 2 1 ∣ ∣ γ − γ 0 , h ∣ − ∣ γ − γ 1 , h ∣ ∣ = 2 1 ( 1 − ratio ) max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 1 , h ∣ ) .
This error is at most 1 2 ∣ δ h ∣ \frac{1}{2}|\delta_h| 2 1 ∣ δ h ∣ , and it is smaller than both endpoint errors if and only if ratio > 1 3 \operatorname{ratio}>\frac{1}{3} ratio > 3 1 . On the final meshes of Table 2 , the midpoint is more accurate than both evaluators in P1, P2, P3, P5 and P6. In P4, where ratio = 0.08 \operatorname{ratio}=0.08 ratio = 0.08 , it is less accurate than γ 1 , h \gamma_{1,h} γ 1 , h . Along the sequences it is more accurate than both evaluators on 30 of the 31 meshes of P1, 27 of the 33 of P2, 30 of the 31 of P3, every mesh of P5 and P6, and 5 of the 31 meshes of P4. The bound 1 2 ∣ δ h ∣ \frac{1}{2}|\delta_h| 2 1 ∣ δ h ∣ exceeds the midpoint error by the factor ( 1 + ratio ) / ( 1 − ratio ) (1+\operatorname{ratio})/(1-\operatorname{ratio}) ( 1 + ratio ) / ( 1 − ratio ) , which is 1.2 in P4 and between 2.5 and 6.1 in the other problems of Table 2 . Since neither the case nor ratio is computable, we report the midpoint together with the interval.
Comparison with a dual-weighted residual computation
For P1, P2 and P7 we also compute the S 2 S_2 S 2 Galerkin approximation z ~ h \tilde{z}_h z ~ h of the adjoint problem, an approximation in an enriched space (§ 4.1 ), on the meshes of Table 3 and of § 6.6 , and the dual-weighted residual estimate η D W R = ℓ ( z ~ h − z h ) − a ( u h , z ~ h − z h ) = γ 0 ( u h , z ~ h ) − γ 0 , h \eta_{\mathrm{DWR}}=\ell(\tilde{z}_h-z_h)-a(u_h,\tilde{z}_h-z_h)=\gamma_0(u_h,\tilde{z}_h)-\gamma_{0,h} η DWR = ℓ ( z ~ h − z h ) − a ( u h , z ~ h − z h ) = γ 0 ( u h , z ~ h ) − γ 0 , h of (4.1). For γ 1 , h \gamma_{1,h} γ 1 , h we use the G pairs in P1 and P2 and the mixed fluxes of § 6.6 in P7. In P1 and P2 the Dirichlet data are zero or of size at most e − 20 \mathrm{e}^{-20} e − 20 , and in P7 the boundary term of § 6.1 is below 0.3% of ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ on the last ten meshes. Table 6 shows the results on the final meshes.
Table 6. The dual-weighted residual estimate with the S 2 S_2 S 2 adjoint approximation z ~ h \tilde{z}_h z ~ h and the QoI gap, for G potentials on the final meshes of Tables 3 and 4.
In P1 and P2, η D W R / ( γ − γ 0 , h ) = 1.00 \eta_{\mathrm{DWR}}/(\gamma-\gamma_{0,h})=1.00 η DWR / ( γ − γ 0 , h ) = 1.00 on the last ten meshes, because γ 0 ( u h , z ~ h ) \gamma_0(u_h,\tilde{z}_h) γ 0 ( u h , z ~ h ) is much more accurate than γ 0 , h \gamma_{0,h} γ 0 , h . In P2 it is more accurate by the factor 4 ⋅ 10 4 4\cdot10^4 4 ⋅ 1 0 4 , and its error decreases with the fitted rate − 1.95 -1.95 − 1.95 , against − 1.03 -1.03 − 1.03 for ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ . In P1 it reaches the accuracy of the reference value. The QoI gap is in case (a) in both problems and estimates the larger of the errors of γ 0 , h \gamma_{0,h} γ 0 , h and γ 1 , h \gamma_{1,h} γ 1 , h . In P7, η D W R / ( γ − γ 0 , h ) \eta_{\mathrm{DWR}}/(\gamma-\gamma_{0,h}) η DWR / ( γ − γ 0 , h ) lies in [ 0.22 , 0.23 ] [0.22,0.23] [ 0.22 , 0.23 ] on the last ten meshes, so that the error of γ 0 ( u h , z ~ h ) \gamma_0(u_h,\tilde{z}_h) γ 0 ( u h , z ~ h ) is 78% of that of γ 0 , h \gamma_{0,h} γ 0 , h . The adjoint solution has in general, like u u u , the singular behavior r 0.1 r^{0.1} r 0.1 at the origin, and on a given mesh a higher polynomial degree does little to improve the approximation of such a singularity. The QoI gap is in case (a) on every mesh of P7.
Dominating reaction and the reaction-restored evaluator γ 2 \gamma_2 γ 2
We test the reaction-restored evaluator γ 2 \gamma_2 γ 2 of (24) and the reaction-restored QoI gap on P2 with ϵ = 0.05 \epsilon= 0.05 ϵ = 0.05 . We vary the reaction coefficient from c = 1 c = 1 c = 1 to c = 10 8 c = 10^{8} c = 1 0 8 and keep the exact solutions u u u and z z z fixed, so that γ = γ ∣ c = 0 + c ( z , u ) \gamma= \gamma|_{c=0} + c\,(z,u) γ = γ ∣ c = 0 + c ( z , u ) . We compute the least-squares approximations with the balance terms weighted by c − 1 / 2 c^{-1/2} c − 1/2 , as in Remark 2.2, and refine the meshes with the balanced indicator of the weighted estimators. For each c c c we take all quantities from one mesh of its adaptive sequence, the one whose number of vertices is closest to 3140, the smallest final number of vertices among the nine sequences of 28 meshes. As c c c grows, the parameter ϑ = ∣ b ∣ / ϵ c \vartheta= |\mathbf{b}|/\sqrt{\epsilon c} ϑ = ∣ b ∣/ ϵc of Proposition 3.8 decreases from ϑ ≫ 1 \vartheta\gg1 ϑ ≫ 1 to ϑ ≪ 1 \vartheta\ll1 ϑ ≪ 1 , where reaction dominates.
The data f 1 f_1 f 1 and g 1 g_1 g 1 grow linearly in c c c , and so do the integrals of data in the evaluators, for instance ( c − 1 f 1 , g 1 ) (c^{-1}f_1,g_1) ( c − 1 f 1 , g 1 ) in γ 2 \gamma_2 γ 2 , which is of size c ( u , z ) c\,(u,z) c ( u , z ) . Their quadrature errors also grow linearly in c c c . Let Q ( u z ) Q(uz) Q ( u z ) be the integral of u z uz u z computed by the three-point rule of degree two. At c = 10 8 c = 10^{8} c = 1 0 8 , on the mesh of Table 7 , c ∣ ( u , z ) − Q ( u z ) ∣ c|(u,z)-Q(uz)| c ∣ ( u , z ) − Q ( u z ) ∣ is 0.51, more than three orders of magnitude larger than the error of γ 2 , h \gamma_{2,h} γ 2 , h . For this reason we compute all evaluators with the 19-point rule of degree nine. Replacing it by the rule of degree eight changes them by less than 10 − 3 ∣ γ − γ 2 , h ∣ 10^{-3}|\gamma-\gamma_{2,h}| 1 0 − 3 ∣ γ − γ 2 , h ∣ .
Table 7. P2 with ϵ = 0.05 \epsilon= 0.05 ϵ = 0.05 and c c c from 1 to 10 8 10^{8} 1 0 8 , least squares with the balance terms weighted by c − 1 / 2 c^{-1/2} c − 1/2 . For each c c c we use one mesh, with N N N vertices, and δ 2 , h = γ 2 , h − γ 0 , h \delta_{2,h} = \gamma_{2,h} - \gamma_{0,h} δ 2 , h = γ 2 , h − γ 0 , h .
Table 7 shows two ranges of c c c with different behavior. For c ≤ 10 2 c \le10^{2} c ≤ 1 0 2 , where ϑ ≥ 1.6 \vartheta\ge1.6 ϑ ≥ 1.6 , γ 2 , h \gamma_{2,h} γ 2 , h has the largest error of the three evaluators, by the factor 70 against γ 1 , h \gamma_{1,h} γ 1 , h at c = 1 c = 1 c = 1 . There ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ is close to the error of γ 2 , h \gamma_{2,h} γ 2 , h and exceeds ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ by factors between 14 and 255. For c ≥ 10 5 c \ge10^{5} c ≥ 1 0 5 the errors of γ 0 , h \gamma_{0,h} γ 0 , h and γ 1 , h \gamma_{1,h} γ 1 , h grow in proportion to c c c , from 9.5 ⋅ 10 − 4 9.5\cdot10^{-4} 9.5 ⋅ 1 0 − 4 at c = 10 5 c = 10^{5} c = 1 0 5 to 0.83 at c = 10 8 c = 10^{8} c = 1 0 8 , and for c ≥ 10 6 c \ge10^{6} c ≥ 1 0 6 they nearly coincide. The QoI gap ∣ δ h ∣ |\delta_h| ∣ δ h ∣ stays at 2.2 ⋅ 10 − 4 2.2\cdot10^{-4} 2.2 ⋅ 1 0 − 4 , because the signed QoI gap does not contain the reaction term ( c e p , e d ) (c e_p,e_d) ( c e p , e d ) (Section 3.3). With TOL = 10 − 3 = 10^{-3} = 1 0 − 3 , the criterion ∣ δ h ∣ ≤ T O L |\delta_h| \leq\mathrm{TOL} ∣ δ h ∣ ≤ TOL holds for every c ≥ 10 5 c \ge10^{5} c ≥ 1 0 5 , although the errors of γ 0 , h \gamma_{0,h} γ 0 , h and γ 1 , h \gamma_{1,h} γ 1 , h reach 0.83. For this reason we use ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ where reaction dominates (Section 5.3). The error of γ 2 , h \gamma_{2,h} γ 2 , h does not grow. It lies between 1.28 ⋅ 10 − 4 1.28\cdot10^{-4} 1.28 ⋅ 1 0 − 4 and 2.20 ⋅ 10 − 4 2.20\cdot10^{-4} 2.20 ⋅ 1 0 − 4 for all c ≥ 10 3 c \ge10^{3} c ≥ 1 0 3 , and at c = 10 8 c = 10^{8} c = 1 0 8 it is smaller than that of γ 0 , h \gamma_{0,h} γ 0 , h by the factor 3.8 ⋅ 10 3 3.8\cdot10^{3} 3.8 ⋅ 1 0 3 .
For large c c c , the reaction-restored QoI gap therefore estimates the error of γ 0 , h \gamma_{0,h} γ 0 , h . For c = 10 6 , 10 7 , 10 8 c = 10^{6},10^{7},10^{8} c = 1 0 6 , 1 0 7 , 1 0 8 the quotient ∣ δ 2 , h ∣ / ∣ γ − γ 0 , h ∣ |\delta_{2,h}|/|\gamma-\gamma_{0,h}| ∣ δ 2 , h ∣/∣ γ − γ 0 , h ∣ is 0.975, 0.997 and 0.9997, and ratio = ∣ γ − γ 2 , h ∣ / ∣ γ − γ 0 , h ∣ = |\gamma-\gamma_{2,h}|/|\gamma-\gamma_{0,h}| = ∣ γ − γ 2 , h ∣/∣ γ − γ 0 , h ∣ is 0.025, 0.0026 and 0.00026, so that ( 1 − ratio ) ∣ γ − γ 0 , h ∣ ≤ ∣ δ 2 , h ∣ ≤ ( 1 + ratio ) ∣ γ − γ 0 , h ∣ (1-\text{ratio})|\gamma-\gamma_{0,h}| \leq|\delta_{2,h}| \leq(1+\text{ratio})|\gamma-\gamma_{0,h}| ( 1 − ratio ) ∣ γ − γ 0 , h ∣ ≤ ∣ δ 2 , h ∣ ≤ ( 1 + ratio ) ∣ γ − γ 0 , h ∣ by Proposition 3.4. For c ≥ 10 6 c \geq10^{6} c ≥ 1 0 6 , where ϑ ≤ 0.016 \vartheta\leq0.016 ϑ ≤ 0.016 , ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ is 0.03 times η p , c L S η d , c L S \eta_{p,c}^{\mathrm{LS}}\eta_{d,c}^{\mathrm{LS}} η p , c LS η d , c LS , so the bound (31) is pessimistic by a factor of about 30. Between these two ranges, at c = 10 3 c = 10^{3} c = 1 0 3 and 10 4 10^{4} 1 0 4 , γ 0 , h \gamma_{0,h} γ 0 , h is the most accurate, by the factors 8.1 and 2.6 over γ 2 , h \gamma_{2,h} γ 2 , h , and ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ exceeds ∣ γ − γ 0 , h ∣ |\gamma-\gamma_{0,h}| ∣ γ − γ 0 , h ∣ by the factors 7.1 and 1.6.
Over all values of c c c , ∣ δ 2 , h ∣ |\delta_{2,h}| ∣ δ 2 , h ∣ lies between 0.62 and 1.08 times max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 2 , h ∣ ) \max(|\gamma-\gamma_{0,h}|,|\gamma-\gamma_{2,h}|) max ( ∣ γ − γ 0 , h ∣ , ∣ γ − γ 2 , h ∣ ) . The pair γ 0 , h , γ 2 , h \gamma_{0,h},\gamma_{2,h} γ 0 , h , γ 2 , h is in case (a) at c = 10 c = 10 c = 10 and 10 2 10^{2} 1 0 2 and in case (b) at the other values of c c c , with ratio ≤ 0.38 \leq0.38 ≤ 0.38 .
Concluding remarks
The product of the primal and adjoint least-squares estimators is a reliable bound for the error in the quantity of interest, and its local contributions are good refinement indicators. It is, however, too pessimistic for deciding when to stop. In our experiments it exceeds the larger of the errors of the two evaluators by factors of up to several hundred, and if we stopped with it we would refine far more than needed.
We compute the signed QoI gap from the same residuals. Like the dual-weighted residual estimate, it is the difference of two values of the quantity of interest, but the second value comes from the fluxes and not from a more accurate approximation. In our experiments the QoI gap is of the size of the larger of the errors of the two evaluators. We therefore refine with the least-squares estimators and stop with the QoI gap. Where reaction dominates, we stop with the reaction-restored QoI gap instead.
With the least-squares flux recovery, we can use the QoI gap with conforming potentials from any method. Our experiments include Galerkin, finite difference and finite volume potentials. We leave the boundary term that appears when we interpolate the Dirichlet data to the next paper of this series.