DOI:10.1051/cocv/2012036 www.esaim-cocv.org
UNIFORMLY EXPONENTIALLY OR POLYNOMIALLY STABLE APPROXIMATIONS FOR SECOND ORDER EVOLUTION EQUATIONS
AND SOME APPLICATIONS
Farah Abdallah
1, Serge Nicaise
2, Julie Valein
3and Ali Wehbe
4Abstract.In this paper, we consider the approximation of second order evolution equations. It is well known that the approximated system by finite element or finite difference is not uniformly exponentially or polynomially stable with respect to the discretization parameter, even if the continuous system has this property. Our goal is to damp the spurious high frequency modes by introducing numerical viscosity terms in the approximation scheme. With these viscosity terms, we show the exponential or polynomial decay of the discrete scheme when the continuous problem has such a decay and when the spectrum of the spatial operator associated with the undamped problem satisfies the generalized gap condition.
By using the Trotter–Kato Theorem, we further show the convergence of the discrete solution to the continuous one. Some illustrative examples are also presented.
Mathematics Subject Classification. 65M60, 35L05, 35L15.
Received December 21, 2011. Revised May 7, 2012.
Published online June 3, 2013.
1. Introduction and main results
Let H be a complex Hilbert space with norm and inner product denoted respectively by. and (., .). Let A : D(A) → H be a densely defined self-adjoint and positive operator with a compact inverse in H. Let V =D(A12) be the domain ofA12.Denote byD(A12) the dual space ofD(A12) obtained by means of the inner product inH.
Furthermore, let U be a complex Hilbert space (which will be identified to its dual space) with norm and inner product denoted respectively by.U and (., .)U and letB∈ L(U, H).We consider the closed loop system
¨
ω(t) +Aω(t) +BB∗ω(t) = 0,˙
ω(0) =ω0,ω(0) =˙ ω1, (1.1)
Keywords and phrases.Stability, wave equation, numerical approximations.
1 Universit´e de Valenciennes et du Hainaut Cambr´esis, LAMAV, FR CNRS 2956, Institut des Sciences et Techniques of Valenciennes, 59313 Valenciennes Cedex 9, France. The Ph.D. studies of F. Abdallah are financed by the association Azm and Saad´e.Farah.Abdallah@meletu.univ-valenciennes.fr
2 Universit´e de Valenciennes et du Hainaut Cambr´esis, LAMAV, FR CNRS 2956, Institut des Sciences et Techniques of Valenciennes, 59313 Valenciennes Cedex 9, France.Serge.Nicaise@univ-valenciennes.fr
3 Institut Elie Cartan Nancy (IECN), Nancy-Universit´e & INRIA (Project-Team CORIDA), 54506 Vandoeuvre-l`es-Nancy Cedex France.Julie.Valein@iecn.u-nancy.fr
4 Universit´e Libanaise, Ecole Doctorale des Sciences et de Technologie, Hadath, Beyrouth, Liban.ali.wehbe@ul.edu.lb
Article published by EDP Sciences c EDP Sciences, SMAI 2013
where t∈[0,∞) represents the time,ω : [0,∞)→H is the state of the system. Most of the linear equations modeling the vibrations of elastic structures with feedback control (corresponding to collocated actuators and sensors) can be written in the form (1.1), whereω represents the displacement field.
We define the energy of system (1.1) at timet by E(t) = 1
2
ω(t)˙ 2+A12ω(t)2
. Simple formal calculations give
E(0)−E(t) = t
0
(BB∗ω(s),˙ ω(s)) ds,˙ ∀t≥0.
This obviously means that the energy is non-increasing.
In many applications, the system (1.1) is approximated by finite dimensional systems but usually if the continuous system is exponentially or polynomially stable, the discrete ones do no more inherit of this property due to spurious high frequency modes. Several remedies have been proposed and analyzed to overcome this difficulties. Let us quote the Tychonoff regularization [18,19,31,34], a bi-grid algorithm [16,28], a mixed finite element method [6,10,11,17,27], or filtering the high frequencies [22,25,35] (both methods providing good numerical results).
As in [31,34] our goal is to damp the spurious high frequency modes by introducing a numerical viscosity in the approximation schemes. Though our paper is inspired from [31], it differs from that paper on the following points:
(i) Contrary to [31] where the standard gap condition is required, we only assume that the spectrum of the operatorA1/2 satisfies the generalized gap condition, allowing to treat more general concrete systems;
(ii) we analyze the polynomial decay of the discrete schemes when the continuous problem has such a decay;
(iii) we prove a result about uniform polynomial stability for a family of semigroups of operators;
(iv) by using a general version of the Trotter–Kato Theorem proved in [23], we show that the discrete solution tends to the solution of (1.1) as the discretization parameter goes to zero and if the discrete initial data are well chosen.
Before stating our main results, let us introduce some notations and assumptions.
We denote by.V the norm
ϕV =
(A12ϕ, A12ϕ),∀ϕ∈V.
Remark that
ϕV =
(Aϕ, ϕ),∀ϕ∈ D(A).
We now assume that (Vh)h>0 is a sequence of finite dimensional subspaces ofD(A12). The inner product inVh is the restriction of the inner product of H and it is still denoted by (., .) (sinceVh can be seen as a subspace ofH). We define the operatorAh : Vh→Vh by
(Ahϕh, ψh) = (A12ϕh, A12ψh),∀ϕh, ψh∈Vh. (1.2) Leta(., .) be the sesquilinear form on Vh×Vh defined by
a(ϕh, ψh) = (A12ϕh, A12ψh), ∀(ϕh, ψh)∈Vh×Vh. (1.3) We also define the operatorsBh: U →Vhby
Bhu=jhBu, ∀u∈U, (1.4)
wherejhis the orthogonal projection ofH into Vh with respect to the inner product inH.
The adjointBh∗ ofBh is then given by the relation
Bh∗ϕh=B∗ϕh, ∀ϕh∈Vh.
We also suppose that the family of spaces (Vh)h approximates the space V = D(A12). More precisely, if πh denotes the orthogonal projection ofV =D(A12) ontoVh, we suppose that there existθ >0, h∗>0 andC0>0 such that, for allh∈(0, h∗), we have:
πhϕ−ϕV ≤C0hθAϕ,∀ϕ∈ D(A), (1.5) πhϕ−ϕ ≤C0h2θAϕ,∀ϕ∈ D(A). (1.6) Assumptions (1.5) and (1.6) are, in particular, satisfied in the case of standard finite element approximations of Sobolev spaces.
Denote by{λk}k≥1the set of eigenvalues ofA12 counted with their multiplicities (i.e.we repeat the eigenvalues according to their multiplicities). We further rewrite the sequence of eigenvalues{λk}k≥1 as follows:
λk1 < λk2 < . . . < λki< . . .
wherek1= 1, k2is the lowest index of the second distinct eigenvalue,k3is the lowest index of the third distinct eigenvalue, etc. For alli∈N∗, letli be the multiplicity of the eigenvalueλki,i.e.
λki−1 < λki =λki+1=. . .=λki+li−1< λki+li=λki+1.
We havek1= 1, k2= 1+l1, k3= 1+l1+l2,etc. Let{ϕki+j}0≤j≤li−1be the orthonormal eigenvectors associated with the eigenvalueλki.
Now, we assume that the following generalized gap condition holds:
∃M ∈N∗,∃γ0>0,∀k≥1, λk+M −λk≥M γ0. (1.7) Fix a positive real numberγ0 ≤γ0 and denote byAk, k= 1, . . . , M the set of natural numbers kmsatisfying (see for instance [5]) ⎧
⎨
⎩
λkm−λkm−1 ≥γ0
λkn−λkn−1 < γ0 form+ 1≤n≤m+k−1, λkm+k−λkm+k−1 ≥γ0.
Then one easily checks that
{km+j+l|km∈Ak, k∈ {1, . . . , M}, j∈ {0, . . . , k−1}, l∈ {0, . . . , lm+j−1}}=N∗.
Notice that some setsAk may be empty because, for the generalized gap condition, the choice ofM takes into account multiple eigenvalues. Forkn∈Ak, we defineBkn= (Bkn, ij)1≤i, j≤k the matrix of sizek×kby
Bkn, ij =
⎧⎪
⎪⎪
⎨
⎪⎪
⎪⎩
n+j−1
q=n+i−1q=n
(λkn+i−1−λkq)−1if i≤j,(i, j)= (1,1),
1 if (i, j) = (1,1),
0 else.
More explicitly, we have
Bkn=
⎛
⎜⎜
⎜⎜
⎜⎜
⎜⎝
1 λ 1
kn−λkn+1 1
(λkn−λkn+1)(λkn−λkn+2) · · · (λkn−λ 1
kn+1)···(λkn−λkn+k−1)
0 λ 1
kn+1−λkn
1
(λkn+1−λkn)(λkn+1−λkn+2) · · · (λkn+1−λkn)···(1λkn+1−λkn+k−1)
0 0 (λ 1
kn+2−λkn)(λkn+2−λkn+1) · · · (λ 1
kn+2−λkn)···(λkn+2−λkn+k−1)
... ... . .. ...
0 0 0 · · · (λkn+k−1−λkn)···(1λkn+k−1−λkn+k−2)
⎞
⎟⎟
⎟⎟
⎟⎟
⎟⎠ .
Lemma 1.1. The inverse matrix of Bkn is given by
B−kn1, ij =
⎧⎪
⎪⎪
⎨
⎪⎪
⎪⎩
n+i−2 q=n
(λkn+j−1−λkq)if i≤j, i= 1,
1 if i= 1,
0 else,
that is to say
Bk−n1=
⎛
⎜⎜
⎜⎜
⎝
1 1 · · · 1
0 (λkn+1−λkn)· · · (λkn+k−1−λkn)
0 0 · · · (λkn+k−1−λkn)(λkn+k−1−λkn+1)
... ... . .. ...
0 0 · · · (λkn+k−1−λkn)· · ·(λkn+k−1−λkn+k−2)
⎞
⎟⎟
⎟⎟
⎠,
and therefore
Bk−n1→
⎛
⎜⎜
⎝
1 1· · · 1 0 0· · · 0 ... . .. ...
0 · · · 0
⎞
⎟⎟
⎠,when n→+∞.
Proof. The form of Bk−n1 is obtained by induction on the size k of Bkn. The generalized gap condition (1.7) implies thatλkn+j−λkn→0 asn→+∞,∀0≤j≤k−1.This leads to the convergence ofB−kn1.
Now, forkn∈Ak, we define the matrixΦkn with coefficients inU and sizek×Ln, whereLn= k i=1
ln+i−1,as follows: for alli= 1, . . . , k, we set
(Φkn)ij =
B∗ϕkn+i−1+j−Ln, i−1−1ifLn, i−1< j≤Ln, i,
0 else,
where
Ln,0= 0, Ln, i= i i=1
ln+i−1fori≥1. (1.8)
For a vectorc= (cl)ml=1 in Um,we setcU,2 its norm inUmdefined by c2U,2=
m l=1
cl2U.
In this paper, we prove two results. The first result gives a necessary and sufficient condition to have the exponential stability of the family of systems
¨
ωh(t) +Ahωh(t) +BhB∗hω˙h(t) +hθAhω˙h(t) = 0
ωh(0) =ω0h∈Vh,ω˙h(0) =ω1h∈Vh, (1.9) in the absence of the standard gap condition assumed in [31]. Here and belowω0h(resp.ω1h) is an approximation ofω0(resp.ω1) inVh. For that purpose, we need to make the following assumption
∃α0>0,∀k∈ {1, . . . , M},∀kn∈Ak,∀C∈RLn,Bk−1
nΦknC
U,2≥α0C2, (1.10) where.2is the euclidian norm. The first main result is the following
Theorem 1.2. Suppose that the generalized gap condition (1.7)and the assumption(1.10)are verified. Assume that the family of subspaces (Vh) satisfies(1.5)and (1.6). Then the family of systems (1.9)is uniformly expo- nentially stable, in the sense that there exist constantsM, α, h∗>0 (independent ofh, ω0h, ω1h) such that for allh∈(0, h∗):
ω˙h(t)2+a(ωh(t), ωh(t))≤Me−αt(ω1h2+a(ω0h, ω0h)),∀t≥0.
Remark 1.3. If the standard gap condition
∃γ0>0,∀n≥1, λkn+1−λkn ≥γ0 (1.11) holds, thenA1=N∗ andB1= 1.In this case, the assumption (1.10) becomes
∃α0>0,∀kn≥1,∀C∈RLn,ΦknCU ≥α0C2.
Moreover, if the standard gap condition (1.11) holds and if the eigenvalues are simple, the assumption (1.10) becomes
∃α0>0,∀k≥1,B∗ϕkU ≥α0. (1.12) These assumptions are assumed in [31].
Remark 1.4. Note that Theorem1.2 is the discrete counterpart of the exponential decay of the solution of the continuous problem (1.1) under the assumptions (1.7) and (1.10), which follows Theorem 2.2 of [3] (see also [29]). Note that the assumption (H) from [3] here holds since A is a positive selfadjoint operator with a compact resolvent andB is bounded.
Remark 1.5. The uniform exponential stability of the family of systems (1.9) has been already proved in Theorem 7.1 of [14] without any assumption on the spectrum ofAand the dimension of the space. The proof of this theorem is based on decoupling of low and high frequencies. More precisely, the author combines a uniform observability estimate for filtered initial data corresponding to low frequencies (see [14], Thm. 1.3) together with a result of [15]. Indeed, in [15], after adding the numerical viscosity term, another uniform observability estimate is obtained for the high frequency components. The two established observability inequalities yield the uniform exponential decay of (1.9).
If the condition (1.10) is not satisfied, we may look at a weaker version. Namely if we assume that
∃l∈N∗,∃α0>0,∀k∈ {1, . . . , M},∀kn ∈Ak, ∀C∈RLn,Bk−n1ΦknC
U,2≥ α0
λlk
n
C2, (1.13) then we will obtain a polynomial stability for the family of systems
¨
ωh(t) +
1 +hθ−2
I+hθAh2l 2
Ahωh(t) +
I+hθAhl2 BhBh∗+hθA1+h 2l I+hθAh2l −1
˙
ωh(t) = 0, ωh(0) =ω0h∈Vh,ω˙h(0) =
1 +hθ−1
I+hθAhl2
ω1h∈Vh.
(1.14)
The structure of the above discrete system has been inspired from the one introduced in [31] for the exponential stability case where the authors have used system (1.9) corresponding to l = 0. In both cases, this choice is motivated by the corresponding observability estimates. The numerical viscosity term (I+hθAh2l)(BhB∗h+ hθA1+h l2)(I+hθAhl2)−1ω˙h(t) is added to damp the high frequency modes and as the set of high frequency modes is larger in the polynomial case, the viscosity term is naturally stronger. In the casel≥2 the powers of (I+hθAhl2) have been added to guarantee the boundedness of the resolvent of ˜Al,h (defined below) near zero. The question of the optimality of these viscosity terms remains open.
The second main result of our paper is the following one.
Theorem 1.6. Suppose that the generalized gap condition (1.7) and the Assumption (1.13) are verified with l∈N∗even. Assume that the family of subspaces(Vh)satisfies(1.5)and(1.6). Then the family of systems(1.14) is uniformly polynomially stable, in the sense that there exist constantsC, h∗>0(independent of h, ω0h, ω1h) such that for all h∈(0, h∗):
I+hθAh2l −1
˙ ωh(t)
2+a(ωh(t), ωh(t))≤ C
t2(ω0h, ω1h)2D( ˜A2ll,h),
I+hθAh2l −1
˙ ωh(t)
2+a(ωh(t), ωh(t))≤ C
t1l (ω0h, ω1h)2D(A˜l,h),∀t >0, ∀(ω0h, ω1h)∈Vh×Vh, where forq∈N∗,.D( ˜Aql,h) is the graph norm of the matrix operatorA˜ql,h given in(4.1) of Section4 below.
For a technical reason, we assumel to be even (see Lem. 4.4). If (1.13) holds forl odd, then it is also true forl+ 1 and we can apply the previous result withl+ 1.
Remark 1.7. If the standard gap condition (1.11) holds, the Assumption (1.13) becomes
∃l∈N∗,∃α0>0,∀kn ≥1,∀C∈RLn,ΦknCU ≥ α0
λlk
n
C2.
Moreover, if the standard gap condition (1.11) holds and if the eigenvalues are simple, the Assumption (1.13) becomes
∃l∈N∗,∃α0>0,∀k≥1,B∗ϕkU ≥ α0
λlk· (1.15)
Remark 1.8. As before, Theorem1.6 is the discrete counterpart of the polynomial decay of the solution of the continuous problem (1.1) under the Assumptions (1.7) and (1.13), that follows from Theorem 2.4 of [3] (see also [29]).
The paper is organized as follows: In Section2, we show that the generalized gap condition and the observ- ability conditions (1.10) and (1.13) remain valid for filtered eigenvalues. Section3 first recalls a result about uniform exponential stability for a family of semigroup of operators, and then extends such a result to the case of uniform polynomial stability. Some technical lemmas are proved in Section4. Sections5and6are devoted to the proof of Theorem1.2and1.6respectively. In Section7, we show that the solutionωh (resp.
I+hθAh2l)−1ω˙h tends toω, the solution of (1.1), (resp. ˙ω) inV (resp. inH) ashgoes to zero and if the discrete initial data are well chosen. Finally, we illustrate our results by presenting different examples in Section8.
2. Spectral analysis of the discretized problem
The eigenvalue problem of the discretized problem is the following one: findλk, h∈]0,+∞[, ϕk, h ∈Vh,such that
a(ϕk, h, ψh) =λ2k, h(ϕk, h, ψh),∀ψh∈Vh. (2.1) Let N(h) be the dimension ofVh. We denote by {λ2k, h}1≤k≤N(h) the set of eigenvalues of (2.1) counted with their multiplicities. Let{ϕk, h}1≤k≤N(h) be the orthonormal eigenvectors associated with the eigenvalueλ2k, h.
In this Section, we show that the generalized gap condition (1.7) and the observability conditions (1.10) and (1.13) still hold for the approximate problem (uniformly inh), provided that we consider only “low fre- quencies”. More precisely, we first have the following result:
Proposition 2.1. Suppose that the generalized gap condition (1.7) and the Assumption (1.10) are verified.
Then, there exist two constants >0 andh∗ >0, such that, for all 0< h < h∗ and for all k∈ {1, . . . , N(h)}
satisfying
hθλ2k≤, (2.2)
we have
∃M ∈N∗,∃γ >0, λk+M, h−λk, h≥M γ (2.3) and
∃α >0,∀p∈ {1, . . . , M},∀kn∈Ap,h,∀C∈RLn,Bk−1
nΦkn, hC
U,2≥αC2, (2.4) where α is independent of h, and where the matrix Φkn, h ∈ Mp, Ln(U), with coefficients in U, is defined as follows: for all i= 1, . . . , p, we set
(Φkn, h)ij =
B∗hϕkn+i−1+j−Ln, i−1−1, hif Ln, i−1< j≤Ln, i,
0 else,
whereLn, i−1 is defined by (1.8) and
Ap,h={kn∈Ap satisfying(2.2)and s.t.kn+p−1+ln+p−1−1≤N(h)}.
For the proof of this proposition, we need a result proved by Babuska and Osborn in [4]. For that purpose, we introduceh(n, j) such that
h(n, j) = inf
ϕ∈Mj(λkn) inf
vh∈Vhϕ−vhV,
where Mj(λkn) = {ϕ ∈ M(λkn) : a(ϕ, ϕkn, h) = . . . = a(ϕ, ϕkn+j−2, h) = 0} and M(λkn) = {ϕ : ϕis an eigenvector ofA12 corresponding toλkn,ϕ = 1}. The restrictions a(ϕ, ϕkn, h) = . . . = a(ϕ, ϕkn+j−2, h) = 0 are not imposed if j = 1. Then, we have the following estimate about the eigenvalue and eigenvector errors for the Galerkin method in terms of the approximability quantitiesh(n, j).
Theorem 2.2. There are positive constantsC andh0 such that
λkn+j, h−λkn+j≤C2h(n, j), ∀0< h≤h0, j= 0, . . . , ln−1, kn+j ≤N(h), n∈N∗ (2.5) and such that the eigenvectors{ϕkn+j}0≤j≤ln−1 ofA12 can be chosen so that
ϕkn+j, h−ϕkn+jV ≤Ch(n, j), ∀0< h≤h0, j= 0, . . . , ln−1, kn+j≤N(h), n∈N∗. (2.6) This result is proved by Babuska and Osborn in [4], p. 702. because
λ2kn+j, h−λ2kn+j = (λkn+j, h−λkn+j)(λkn+j, h+λkn+j)≥2λ1(λkn+j, h−λkn+j).
Remark 2.3. Notice that for everyϕ∈Mj(λkn) we have h(n, j)≤ inf
vh∈Vhϕ−vhV
≤C0hθAϕ by (1.5)
≤C0hθλ2knϕ=C0hθλ2kn+j.
(2.7) Proof of Proposition2.1. We begin with the proof of the generalized gap condition for the approximate eigen- valuesλk, h.First, we use the Min-Max principle (see [32]) to obtain
λk ≤λk, h,∀k∈ {1, . . . , N(h)}. (2.8) Second, we use the estimates (2.5) and (2.7) and we have
λk, h≤λk+C(C0hθλ2k)2≤λk+C(C0)2≤λk+CC02, (2.9) for allk∈ {1, . . . , N(h)} verifying (2.2) and≤1. Therefore, we may write
λk+M, h−λk, h≥λk+M −λk−CC02≥M γ0−CC02≥Mγ0
2 =:M γ for allk∈ {1, . . . , N(h)} satisfying (2.2) and for≤ 2MγCC02
0·
Now, we prove the estimate (2.4) which is the approximated version of (1.10). Notice that Φkn, h−ΦknU ≤C max
i=0,...,p−1 ln+i−1
j=0
B∗ϕkn+i+j, h−B∗ϕkn+i+j
U
≤C max
i=0,...,p−1 ln+i−1
j=0
B∗L(H, U)ϕkn+i+j, h−ϕkn+i+j
≤C max
i=0,...,p−1 ln+i−1
j=0
B∗L(H, U)ϕkn+i+j, h−ϕkn+i+j
V
≤C max
i=0,...,p−1 ln+i−1
j=0
h(n+i, j) by (2.6)
≤C max
i=0,...,p−1 ln+i−1
j=0
hθλ2kn+i+jby (2.7).
Thus, by (2.2), we get
Φkn, h−ΦknU ≤C. (2.10)
Therefore the triangular inequality leads to Bk−n1Φkn, hC
U,2=B−kn1ΦknC+Bk−n1(Φkn, h−Φkn)C
U,2
≥B−kn1ΦknC
U,2−Bk−n1(Φkn, h−Φkn)C
U,2
≥α0C2−Bk−n1(Φkn, h−Φkn)C
U,2
by (1.10). But, asBk−n1=
⎛
⎜⎜
⎝
1 1· · · 1 0 0· · · 0 ... . .. ...
0 · · · 0
⎞
⎟⎟
⎠+Rkn,withRkn→0,whenkn→+∞(see Lem.1.1), we obtain
Bk−n1(Φkn, h−Φkn)C
U,2≤
⎛
⎜⎜
⎝
1 1· · · 1 0 0· · · 0 ... . .. ...
0 · · · 0
⎞
⎟⎟
⎠(Φkn, h−Φkn)C
U,2
+Rkn(Φkn, h−Φkn)CU,2
≤CΦkn, h−ΦknUC2+ηnΦkn, h−ΦknUC2
≤C(1 +ηn)C2,
(2.11)
whereηn=Rkn →0.Thus
Bk−n1Φkn, hC
U,2≥(α0−C(1 +ηn))C2≥ α0
2 C2
for≤ 2C(1+maxα0
n (1 +ηn))·
For the polynomial stability, we have the same kind of result, but more filtering is necessary in order to have the discrete counterpart of the observability condition (1.13) (uniformly inh).
Proposition 2.4. Suppose that the generalized gap condition (1.7) and the Assumption (1.13) are verified.
Then, there exist two constants >0 and h∗ >0, such that, for all0< h < h∗ and for all k∈ {1, . . . , N(h)}, satisfying
hθλ2k ≤
λlk, (2.12)
we have (2.3)and
∃α >0, ∀p∈ {1, . . . , M},∀kn ∈A(p,hl),∀C∈RLn,B−kn1Φkn, hC
U,2≥ α
λlknC2, (2.13) whereA(p,hl) ={kn∈Ap satisfying (2.12)and s.t. kn+p−1+ln+p−1−1≤N(h)}.
Proof. The generalized gap condition for the approximate eigenvaluesλk, h is a consequence of Proposition2.1, becauseλk ≥λ1>0.
To prove the estimate (2.13) we notice that Φkn, h−ΦknU ≤C max
i=0,...,p−1 ln+i−1
j=0
hθλ2kn+i+j ≤C hθλ2kn+p−1. Moreover, by the triangular inequality and (1.13), we have
Bk−1
nΦkn, hC
U,2≥ λαl0
kn C2−Bk−1
n(Φkn, h−Φkn)C
U,2. By (2.11) and (2.12), we obtain
Bk−n1Φkn, hC
U,2≥
α0
λlkn −Cλ(1+l ηn) kn+p−1
C2
≥
α0
λlkn −λlC
kn+ρn(1 +ηn)
C2,withρn=λlkn+p−1−λlkn→0
≥ λαl kn C2
for an appropriate choice of >0.
3. Uniform stability results
3.1. Exponential stability result
The Proof of Theorem1.2is based on the following result (see Thm. 7.1.3 in [26]):
Theorem 3.1. Let (Th)h>0 be a family of semigroups of contractions on the Hilbert spaces (Xh)h>0 and let ( ˜Ah)h>0 be the corresponding infinitesimal generators. The family (Th)h>0 is uniformly exponentially stable, that is to say there exist constantsM >0, α >0 (independent ofh∈(0, h∗)) such that
Th(t)L(Xh)≤Me−αt,∀t≥0, if and only if the two following conditions are satisfied:
(i) For allh∈(0, h∗), iR is contained in the resolvent setρ( ˜Ah)of A˜h, (ii) sup
h∈(0,h∗),ω∈R
(iω−A˜h)−1
L(Xh)<+∞.
3.2. Polynomial stability result
The proof of Theorem1.6 is based on the results presented in this section by adapting the results from [9]
and from [24] to obtain the (uniform) polynomial stability of the discretized problem (1.14). Throughout this section, let (Th(t)) t≥0
h∈(0,h∗)
be a family of uniformly boundedC0 semigroups on the associated Hilbert spaces (Xh)h∈(0,h∗) (i.e., ∃M > 0, ∀h ∈ (0, h∗), Th(t)L(Xh) ≤ M ) and let ( ˜Ah)h∈(0,h∗) be the corresponding infinitesimal generators.
In the following, for shortness, we denote byR(λ,A˜h) the resolvent (λ−A˜h)−1; moreover, for any operator mapping Xh intoXh, we skip the indexL(Xh) in its norm, since in the whole section we work inXh.
Definition 3.2. Assuming that
iR⊆ρ( ˜Ah), ∀h∈(0, h∗), (3.1)
and that for allm≥1,there existsc=c(m)>0 such that sup
h∈(0,h∗)
|s|≤m
R(is,A˜h)L(Xh)≤c, (3.2)
we define the fractional power ˜A−αh forα >0 andh∈(0, h∗), according to [2] and [13], as A˜−αh = 1
2πi
Γλ−α(λ−A˜h)−1dλ,
where λ−α = e−αlogλ and R+ is taken as the cut branch of the complex log function and where the curve Γ =Γ1∪Γ2 is given by
Γ ={−+teiθ, t∈[0,+∞)} ∪ {−−te−iθ, t∈(−∞,0]} (3.3) for some >0 small enough independent ofhandθis a fixed angle in
0,π
4
.
Remark 3.3. Throughout this section, whenever ˜A−αh is mentioned, the Assumptions (3.1) and (3.2) are directly taken into consideration since otherwise ˜A−αh is not well defined.
In fact, under the Assumptions (3.1) and (3.2), for allm >0 there exists=(m)>0 such that
−μ+iβ∈ρ( ˜Ah), ∀h∈(0, h∗), ∀0≤μ≤,∀|β| ≤m.
Indeed, for all m >0 such that|β| ≤m, we have
(−μ+iβ−A˜h)−1= (iβ−A˜h)−1[Ih−μ(iβ−A˜h)−1]−1 and
μ(iβ−A˜h)−1 ≤μc.
Hence, if|β| ≤mandμ≤≤ 21c, then (−μ+iβ−A˜h) is invertible and we have
(−μ+iβ−A˜h)−1 ≤2(iβ−A˜h)−1 ≤2c, ∀h∈(0, h∗). (3.4) We choose m = (−+teiθ) = tanθ when (−+teiθ) = 0, i.e. when t = cosθ· Therefore, by (3.4), Assumptions (3.1) and (3.2) imply that there exists >0 independent ofhsuch that the curveΓ is included in ρ( ˜Ah) for anyh∈(0, h∗), and hence ˜A−αh is well defined. In fact, ifξ∈Γ such thatξ >0, then, by the Hille Yosida theorem,ξ∈ρ( ˜Ah), while if−≤ξ≤0, then, by (3.4),ξ∈ρ( ˜Ah).
Proposition 3.4. If, in addition to Assumptions (3.1)and(3.2), we have
h∈sup(0,h∗)
R(is,A˜h)L(Xh)=O(|s|α), |s| → ∞, (3.5)
thenA˜−αh is uniformly bounded independently ofh∈(0, h∗).