This section is devoted to the asymptotic quasi-optimal convergence analysis of the adaptive eigenvalue computation based on exact solutions of the algebraic eigenvalue problems. At first the approximation class Asis defined and its properties are described.
Lemma 3.5.2 shows an estimator reduction which is used in the proof of the contraction property in Lemma 3.5.3. The contraction property and the bulk criterion are key arguments in the proof of the quasi-optimality in Theorem 3.5.4.
Definition 3.5.1 (Approximation class). For an initial triangulation T0 and for s > 0
let the approximation class be defined by
As:= v ∈ V : |v|As := sup ε>0 ε inf Tε:|||v−vε|||≤ε (|Tε| − |T0|) s < ∞ .
The infimum is taken over all refinements Tεof T0 computed by the refinement algorithm
of Section 3.2 with |||v − vε||| ≤ ε and vε ∈ Vε.
Notice that As contains all functions that can be approximated within pre-described
tolerance ε > 0 in a finite element space Vε, |||v − vε||| ≤ ε for some vε ∈ Vε, based
on the triangulation Tε with |Tε| − |T0| ≤ ε−1/s|v| 1/s
As. For uniform refinement classical a priori estimates show that for 0 < r ≤ 1, H1+r(Ω) ∩ V ⊂ A
r/n, but the class contains
much more functions which motivates the use of adaptivity. Due to [102] an equivalent formulation, similar to that of [41], reads
As:= v ∈ V : sup N ∈N Ns inf Tε:|Tε|−|T0|≤N |||v − vε||| < ∞ .
In the following the marking strategy of Section 3.2 is a key argument in the proofs. Lemma 3.5.2. Let (λℓ, uℓ) and (λℓ+1, uℓ+1) be discrete eigenpairs on the levels ℓ and
ℓ + 1 to the continuous eigenpair (λ, u), then there exists some Λ > 0, such that, for all levels ℓ ≥ 0 and 0 < θ ≤ 1, it holds that
ηℓ+1(λℓ+1, uℓ+1) ≤
(1 − θ(1 − 2−2/n))η
ℓ(λℓ, uℓ) + Λ|||uℓ+1− uℓ|||.
Proof. As in the proof of [33, Lemma 5.1], Young’s inequality [52], some discrete inverse
inequalities and the bulk criterion of Section 3.2 lead to
for any 0 < δ from Young’s inequality, 0 < θ ≤ 1 bulk parameter and 0 < Λ from application of various discrete inverse inequalities. Thereby, the factor 2−2/n results from at least one bisection of refined elements or edges. The choice
δ = Λ|||uℓ+1− uℓ|||
(1 − θ(1 − 2−2/n))η
ℓ(λℓ, uℓ)
proves the assertion.
Lemma 3.5.3 (Contraction property). Let (λℓ, uℓ) and (λℓ+1, uℓ+1) be discrete eigen-
pairs on the levels ℓ and ℓ+1 to the same continuous eigenpair (λ, u) and let the mesh-size Hℓ be sufficiently small, then there exist constants 0 < ϱ < 1 and γ > 0, such that, for
all ℓ = 0, 1, 2, . . ., it holds that
γηℓ+12 (λℓ+1, uℓ+1) + |||u − uℓ+1|||2 ≤ ϱ
γη2ℓ(λℓ, uℓ) + |||u − uℓ|||2
. (3.3)
Proof. Theorem 5.3 of [33] shows for 0 < ρ < 1 that
γη2ℓ+1(λℓ+1, uℓ+1) + |||eℓ+1|||2 ≤ ρ
γηℓ2(λℓ, uℓ) + |||eℓ|||2
+ 3λℓ+1∥eℓ+1∥2 + 3λℓ∥eℓ∥2.
Lemma 3.3.3 and 3.3.4 show
∥u − uℓ∥2 ≤ σ(Hℓ)2|||u − uℓ|||2, (3.4)
where σ(Hℓ) := 2(1 + M )CapxCregHℓr.
Hence, for sufficiently small mesh-size H0, (3.3) follows with the constant
0 < ϱ := ρ + 3λ0σ(Hℓ)
2
1 − 3λ0σ(Hℓ)2
< 1.
Theorem 3.5.4. Suppose that (λℓ, uℓ) is a discrete eigenpair to the continuous eigenpair
(λ, u) with u ∈ As and that the initial mesh-size H0 is sufficiently small. Then λℓ and
uℓ from the AFEM converge quasi-optimal in the sense that
|||eℓ|||2+ |λ − λℓ| . (|Tℓ| − |T0|)−2s . Nℓ−2s.
Proof. First it is shown that for a set Mℓof marked edges and elements from the marking
strategy of Section 3.2, based on the bulk criterion, ηℓ(λℓ, uℓ) and a bulk parameter θ > 0,
it holds that
|Mℓ| . |||eℓ|||−1/s|u| 1/s As.
Note that it is sufficient that Mℓ is a set with almost minimal cardinality, i.e. minimal
that fulfils the bulk criterion. Suppose Tℓ+ε is any refinement of Tℓ such that
|||eℓ+ε||| ≤ ρ|||eℓ|||
for some 0 < ρ < 1. Suppose that Hℓ and θ are sufficiently small, such that
0 < θ ≤ (1 − ρ
2)
C2 relCeff2
− λσ(Hℓ)2,
where σ(Hℓ) from Lemma 3.5.3 tends to zero as Hℓ → 0. Using the efficiency estimates
of Remark 3.4.3 together with the quasi-orthogonality of Lemma 3.3.1 yields (1 − ρ2)ηℓ2(λℓ, uℓ)/Ceff2 ≤ (1 − ρ
2)|||e
ℓ|||2 ≤ |||eℓ|||2− |||eℓ+ε|||2
= |||uℓ+ε− uℓ|||2+ λ∥eℓ∥2− λ∥eℓ+ε∥2− λℓ+ε∥uℓ+ε− uℓ∥2.
Let Mε:= (Tℓ\Tℓ+ε) ∪ (Eℓ\Eℓ+ε), then the reliability of Remark 3.4.3 and (3.4) yield
(1 − ρ2)η2ℓ(λℓ, uℓ)/Ceff2 ≤ C 2 relη 2 ℓ(λℓ, uℓ; Mε) + λ∥eℓ∥2 ≤ Crel2 η2ℓ(λℓ, uℓ; Mε) + λσ(Hℓ)2Crel2 η 2 ℓ(λℓ, uℓ).
Therefore Mε satisfies the bulk criterion. Since Mℓ is the set with almost minimal
cardinality that fulfils the bulk criterion, it holds that |Mℓ| . |Mε| . |Tℓ+ε| − |Tℓ|.
Let Tε be an optimal mesh with smallest cardinality such that
|||eε||| ≤ ρ|||eℓ|||.
The definition of the approximation space As shows that
|Tε| − |T0| ≤ ρ−1/s|||eℓ|||−1/s|u| 1/s As.
Let Tℓ+ε be the smallest common refinement of Tε and Tℓ. The overlay estimate of
Lemma 2.5.2 yields
|Mℓ| . |Tℓ+ε| − |Tℓ| = |Tε⊕ Tℓ| − |Tℓ| ≤ |Tε| − |T0| . |||eℓ|||−1/s|u| 1/s As. This and the boundedness of closure in Lemma 2.5.1 yield
|TL| − |T0| . L−1 ℓ=0 |Mℓ| . |u| 1/s As L−1 ℓ=0 |||eℓ|||−1/s.
The efficiency estimate of Remark 3.4.3 yields
γηℓ2(λℓ, uℓ) + |||u − uℓ|||2 ≤
Thus, |||u − uℓ|||−1/s ≤ 1 + γCeff2 1/(2s)γηℓ2(λℓ, uℓ) + |||u − uℓ|||2 −1/(2s) . Lemma 3.5.3 leads to γηℓ2(λℓ, uℓ) + |||u − uℓ|||2 −1/(2s) ≤ ϱ1/(2s) γηℓ+12 (λℓ+1, uℓ+1) + |||u − uℓ+1|||2 −1/(2s) .
Exploiting the reliability of the estimator and a geometric series argument yields that |TL| − |T0| is, up to a generic multiplicative constant, bounded by
|u|1/sAs 1 + γCeff2 1/(2s)γηL2(λL, uL) + |||u − uL|||2
−1/(2s)L ℓ=1 ϱℓ/(2s) . |u|1/sAs 1 + γC2 eff 1 + γ/C2 rel 1/(2s) (1 − ϱ1/(2s))−1|||u − uL|||−1/s.
Note that Euler’s formula [51] shows (|Tℓ| − |T0|) ≈ Nℓ. Finally equation (3.2) proves
|λ − λℓ| . (|Tℓ| − |T0|)−2s.