A generalized finite element method for the strongly ...
Transcript of A generalized finite element method for the strongly ...
ESAIM: M2AN 55 (2021) 1375β1404 ESAIM: Mathematical Modelling and Numerical Analysishttps://doi.org/10.1051/m2an/2021023 www.esaim-m2an.org
A GENERALIZED FINITE ELEMENT METHOD FOR THE STRONGLYDAMPED WAVE EQUATION WITH RAPIDLY VARYING DATA
Per Ljung1,*, Axel Malqvist1 and Anna Persson2
Abstract. We propose a generalized finite element method for the strongly damped wave equation withhighly varying coefficients. The proposed method is based on the localized orthogonal decompositionintroduced in Malqvist and Peterseim [Math. Comp. 83 (2014) 2583β2603], and is designed to handleindependent variations in both the damping and the wave propagation speed respectively. The methoddoes so by automatically correcting for the damping in the transient phase and for the propagationspeed in the steady state phase. Convergence of optimal order is proven in πΏ2(π»
1)-norm, independent ofthe derivatives of the coefficients. We present numerical examples that confirm the theoretical findings.
Mathematics Subject Classification. 35K10, 65M60.
Received November 5, 2020. Accepted May 17, 2021.
1. Introduction
This paper is devoted to the study of numerical solutions to the strongly damped wave equation with highlyvarying coefficients. The equation takes the general form
ββ Β· (π΄β + π΅βπ’) = π, (1.1)
on a bounded domain Ξ©. Here, π΄ and π΅ represent the systemβs damping and wave propagation respectively, πdenotes the source term, and the solution π’ is a displacement function. This equation commonly appears in themodelling of viscoelastic materials, where the strong damping ββ Β· π΄β arises due to the stress being repre-sented as the sum of an elastic part and a viscous part [6, 13]. Viscoelastic materials have several applicationsin engineering, including noise dampening, vibration isolation, and shock absorption (see [20] for more applica-tions). In particular, in multiscale applications, such as modelling of porous media or composite materials, π΄and π΅ are both rapidly varying.
There has been much recent work regarding strongly damped wave equations. For instance, well-posednessof the problem is discussed in [7, 19, 21], asymptotic behavior in [3, 8, 30, 34] solution blowup in [2, 12], anddecay estimates in [18]. In particular, FEM for the strongly damped wave equation has been analyzed in [24]
Keywords and phrases. Strongly damped wave equation, multiscale, localized orthogonal decomposition, finite element method,reduced basis method.
1 Department of Mathematical Sciences, Chalmers University of Technology and University of Gothenburg,SE-412 96 Gothenburg, Sweden.2 Department of Mathematics, KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden.*Corresponding author: [email protected]
c The authors. Published by EDP Sciences, SMAI 2021
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0),which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
1376 P. LJUNG ET AL.
using the RitzβVolterra projection, and [23] uses the classical Ritz-projection in the homogeneous case withRayleigh damping. In these papers, convergence of optimal order is shown. However, in the case of piecewiselinear polynomials, the convergence relies on at least π»2-regularity in space. Consequently, since the π»2-normdepends on the derivatives of the coefficients, the error is bounded by βπ’βπ»2 βΌ max(πβ1
π΄ , πβ1π΅ ) where ππ΄ and ππ΅
denote the scales at which π΄ and π΅ vary respectively. The convergence order is thus only valid when the meshwidth β fulfills β < min(ππ΄, ππ΅). In other words, we require a mesh that is fine enough to resolve the variationsof π΄ and π΅, which becomes computationally challenging. This type of difficulty is common for equations withrapidly varying data, an issue for which several numerical methods have been developed (see e.g. [4, 5, 22, 28]).None of these methods are however applicable to the strongly damped wave equation, where two differentmultiscale coefficients have to be dealt with. In this paper, we propose a novel multiscale method based on thelocalized orthogonal decomposition (LOD) method.
The LOD method is based on the variational multiscale method presented in [17]. It was first introduced in[29], and has since then been further developed and analyzed for several types of problems (see e.g. [1, 15, 16,25,27]). In particular, Malqvist and Peterseim [26] studies the LOD method for quadratic eigenvalue problems,which correspond to time-periodic wave equations with weak damping. The main idea of the method is basedon a decomposition of the solution space into a coarse and a fine part. The decomposition is done by defining aninterpolant that maps functions from an infinite dimensional space into a finite dimensional FE-space. In thisway, the kernel of the interpolant captures the finescale features that the coarse FE-space misses, and hencedefines the finescale space. Subsequently, one may use the orthogonal complement to this finescale space withrespect to a problem-dependent Ritz-projection as a modified FE-space. In the case of time-dependent problems,the LOD method performs particularly well in the sense that the modified FE-space only needs to be computedonce, and can then be re-used in each time step.
Multiscale methods, as the localized orthogonal decomposition, are usually designed to handle problemswith a single multiscale coefficient. In this sense, the strongly damped wave equation is different, as an extracoefficient appears due to the strong damping. Hence, one of the main challenges for the novel method is howto incorporate the finescale behavior of both coefficients in the computation. Nevertheless, it should be notedthat existing multiscale methods are applicable for some special cases of this equation. An example is the caseof Rayleigh damping where the coefficients are proportional to each other. Other examples are the steady statecase, the transient phase in which the solution evolves rapidly in time, as well as the case of weak dampingwhere no spatial derivatives are present on the damping term.
In this paper we present a generalized finite element method (GFEM), with a backward Euler time steppingfor solving the strongly damped wave equation. The method uses both the damping and diffusion coefficientsto construct a generalized finite element space, similar to those in e.g. [25, 29]. The solution is then evaluatedin this space, but to account for the time dependence, an additional correction is added to it. However, thiscorrection is evaluated on the fine scale, and thus expensive to compute. To overcome this issue, we prove spatialexponential decay for the corrections so that we can restrict the problems to patches in a similar manner as forthe modified basis functions in [29]. The effect of the proposed method is that the multiscale basis compensatesfor the damping early on in the simulation when it is dominant and then gradually starts to compensate for thewave propagation which is dominant at steady state. This is done seamlessly and automatically by the method.Furthermore, we prove optimal order convergence in πΏ2(π»1)-norm for this method. Following this, we showthat it is sufficient to compute the finescale corrections for only a few time steps by applying reduced basis (RB)techniques. For related work on RB methods, see e.g. [9,10,14], and for an introduction to the topic we refer to[32].
The outline of the paper is as follows: In Section 2 we present the weak formulation and classical FEM forthe strongly damped wave equation, along with necessary assumptions. Section 3 is devoted to the generalizedfinite element method and its localization procedure. In Section 4 error estimates for the method are proven.Section 5 covers the details of the RB approach, and finally in Section 6 we illustrate numerical examples thatconfirm the theory derived in this paper.
GFEM FOR THE STRONGLY DAMPED EQUATION 1377
2. Weak formulation and classical FEM
We consider the wave equation with strong damping of the following form
ββ Β· (π΄β + π΅βπ’) = π, in Ωà (0, π ], (2.1)π’ = 0, on ΞΓ (0, π ], (2.2)
π’(0) = π’0, in Ξ©, (2.3)(0) = π£0 in Ξ©, (2.4)
where π > 0 and Ξ© is a polygonal (or polyhedral) domain in Rπ, π = 2, 3, and Ξ := πΞ©. The coefficients π΄ andπ΅ describe the damping and propagation speed respectively, and π denotes the source function of the system.We assume π΄ = π΄(π₯), π΅ = π΅(π₯) and π = π(π₯, π‘), i.e. the multiscale coefficients are independent of time.
Denote by π»10 (Ξ©) the classical Sobolev space with norm
βπ£β2π»1(Ξ©) = βπ£β2πΏ2(Ξ©) + ββπ£β2πΏ2(Ξ©)
whose functions vanish on Ξ. Moreover, let πΏπ(0, π ;β¬) be the Bochner space with norm
βπ£βπΏπ(0,π ;β¬) =(β« π
0
βπ£βπβ¬ dπ‘
)1/π
, π β [1,β),
βπ£βπΏβ(0,π ;β¬) = ess supπ‘β[0,π ]
βπ£ββ¬,
where β¬ is a Banach space with norm β Β· ββ¬. In this paper, the following assumptions are made on the data.
Assumption 2.1. The damping and propagation coefficients π΄, π΅ β πΏβ(Ξ©, RπΓπ) are symmetric and satisfy
0 < πΌβ := ess infπ₯βΞ©
infπ£βRπβ0
π΄(π₯)π£ Β· π£π£ Β· π£
< ess supπ₯βΞ©
supπ£βRπβ0
π΄(π₯)π£ Β· π£π£ Β· π£
=: πΌ+ < β,
0 < π½β := ess infπ₯βΞ©
infπ£βRπβ0
π΅(π₯)π£ Β· π£π£ Β· π£
< ess supπ₯βΞ©
supπ£βRπβ0
π΅(π₯)π£ Β· π£π£ Β· π£
=: π½+ < β.
In addition, we assume that π β πΏβ([0, π ]; πΏ2(Ξ©)) and π β πΏ2([0, π ]; πΏ2(Ξ©)).
For the spatial discretization, let π―ββ>0 denote a family of shape regular elements that form a partitionof the domain Ξ©. For an element πΎ β π―β, let the corresponding mesh size be defined as βπΎ := diam(πΎ), anddenote the largest diameter of the partition by β := maxπΎβπ―β
βπΎ . We now define the classical FE-space usingcontinuous piecewise linear polynomials as
πβ := π£ β π(Ξ©) : π£Ξ
= 0, π£πΎ
is a polynomial of partial degree β€ 1, βπΎ β π―β,
and let πβ = πβ β©π»10 (Ξ©). The semi-discrete FEM becomes: find π’β : [0, π ] β πβ such that
(β, π£) + π(β, π£) + π(π’β, π£) = (π, π£), βπ£ β πβ, π‘ β [0, π ], (2.5)
with initial values π’β(0) = π’0β and β(0) = π£0
β where π’0β, π£0
β β πβ are appropriate approximations of π’0 and π£0
respectively. Here (Β·, Β·) denotes the usual πΏ2-inner product, π(Β·, Β·) = (π΄βΒ·,βΒ·), and π(Β·, Β·) = (π΅βΒ·,βΒ·).For the temporal discretization, let 0 = π‘0 < π‘1 < . . . < π‘π = π be a uniform partition with time step
π‘π β π‘πβ1 = π . We apply a backward Euler scheme to get the fully discrete system: find π’πβ β πβ such that
(π2π‘ π’π
β, π£) + π(ππ‘π’πβ, π£) + π(π’π
β, π£) = (ππ, π£), βπ£ β πβ, (2.6)
1378 P. LJUNG ET AL.
for π β₯ 2. Here, the discrete derivative is defined as ππ‘π’πβ = (π’π
β β π’πβ1β )/π . The first initial value is given by
π’0β β πβ. The second initial value π’1
β should be an approximation of π’(π‘1) and could be chosen as π’1β = π’0
β +ππ£0β.
For results on regularity and error estimates for the FEM solution of the strongly damped wave equation, werefer to [23]. Moreover, existence and uniqueness of a solution to (2.6) is guaranteed by LaxβMilgram.
In the analysis, we use the notations βΒ·β2π := π(Β·, Β·), βΒ·β2π := π(Β·, Β·), as well as |||Β·|||2 = (Β·, Β·) := π(Β·, Β·)+ππ(Β·, Β·), andthe fact that these are equivalent with the π»1-norm. That is, there exist positive constants πΆπ, πΆπ, πΆ, ππ, ππ, π βR, such that
ππβπ£β2π»1 β€ βπ£β2π β€ πΆπβπ£β2π»1 , βπ£ β π»1(Ξ©),ππβπ£β2π»1 β€ βπ£β2π β€ πΆπβπ£β2π»1 , βπ£ β π»1(Ξ©), (2.7)
πβπ£β2π»1 β€ |||π£|||2 β€ πΆβπ£β2π»1 , βπ£ β π»1(Ξ©).
Theorem 2.2. The solution π’πβ to (2.6) satisfies the following bounds
βππ‘π’πββ2πΏ2
+πβ
π=2
πβππ‘π’πββ
2π»1 + βπ’π
ββ2π»1 β€ πΆ
πβπ=2
πβπ πβ2π»β1 + πΆ(βππ‘π’
1ββ2πΏ2
+ βπ’1ββ2π»1
), for π β₯ 2, (2.8)
πβπ=2
πβπ2π‘ π’π
ββ2πΏ2
+ βππ‘π’πββ2π»1 β€ πΆ
πβπ=2
πβπ πβ2πΏ2+ πΆ
(βππ‘π’
1ββ2π»1 + βπ’1
ββ2π»1
), for π β₯ 2. (2.9)
Proof. To prove (2.8), choose π£ = πππ‘π’πβ in (2.6) to get
π(π2
π‘ π’πβ, ππ‘π’
πβ
)+ πβππ‘π’
πββ2π + ππ
(π’π
β, ππ‘π’πβ
)= π
(ππ, ππ‘π’
πβ
). (2.10)
Due to CauchyβSchwarz and Youngβs inequality we have the following lower bound
π(π2π‘ π’π
β, ππ‘π’πβ) = βππ‘π’
πββ2πΏ2
β(ππ‘π’
πβ1β , ππ‘π’
πβ
)β₯ 1
2βππ‘π’
πββ2πΏ2
β 12βππ‘π’
πβ1β β2πΏ2
,
and similarly
ππ(π’πβ, ππ‘π’
πβ) β₯ 1
2βπ’π
ββ2π β12βπ’πβ1
β β2π .
Similar bounds will be used repeatedly throughout the paper. Summing (2.10) over π gives
12βππ‘π’
πββ2πΏ2
β 12βππ‘π’
1ββ2πΏ2
+πβ
π=2
πβππ‘π’πββ
2π +
12βπ’π
ββ2π β12βπ’1
ββ2π β€πβ
π=2
πβπ πβπ»β1βππ‘π’πββπ»1 .
Using the equivalence of the norms (2.7), CauchyβSchwarz and Youngβs (weighted) inequality to subtractβππ=2 πβππ‘π’
πββ2π»1 from both sides, we get exactly (2.8).
The proof of (2.9) is similar. We choose π£ = ππ2π‘ π’π
β in (2.6) and sum over π to get
πβπ=2
πβπ2π‘ π’π
ββ2πΏ2
+12βππ‘π’
πββ2π β
12βππ‘π’
1ββ2π +
πβπ=2
ππ(π’π
β, π2π‘ π’π
β
)β€
πβπ=2
πβπ πβπΏ2βπ2π‘ π’π
ββπΏ2 .
For the sum involving the bilinear form π(Β·, Β·) we use summation by parts to get
πβπ=2
ππ(π’π
β, π2π‘ π’π
β
)=
πβπ=3
βππ(ππ‘π’
πβ, ππ‘π’
πβ1β
)β π
(π’2
β, ππ‘π’1β
)+ π
(π’π
β, ππ‘π’πβ
).
GFEM FOR THE STRONGLY DAMPED EQUATION 1379
Using (2.8), the equivalence of the norms (2.7), and Youngβs weighted inequality we have πβπ=3
ππ(ππ‘π’
πβ, ππ‘π’
πβ1β
)+ π
(π’2
β, ππ‘π’1β
)β π
(π’π
β, ππ‘π’πβ
)β€ πΆ
πβπ=2
πβππ‘π’πββ
2π»1 + πΆ
(βπ’2
ββ2π»1 + βππ‘π’1ββ2π»1
)+ πΆβπ’π
ββ2π»1 + πΆπβππ‘π’πββ2π
β€ πΆ
πβπ=2
πβπ πβ2π»β1 + πΆ(βππ‘π’
1ββ2π»1 + βπ’1
ββ2π»1
)+ πΆπβππ‘π’
πββ2π.
Since πΆπ can be made arbitrarily small, it can be kicked to the left hand side. Using that βπ πβ2π»β1 β€ πΆβπ πβ2πΏ2
we deduce (2.9).
3. Generalized finite element method
This section is dedicated to the development of a multiscale method based on the framework of the standardLOD. First of all, we introduce some notation for the discretization. Let ππ» be a FE-space defined analogouslyto πβ in previous section, but with larger mesh size π» > β. Moreover, we assume that the corresponding familyof partitions π―π»π»>β is, in addition to shape-regular, also quasi-uniform. Denote by π© the set of interior nodesof ππ» and by ππ₯ the standard hat function for π₯ β π© , such that ππ» = span(ππ₯π₯βπ© ). Finally, we make theassumption that π―β is a refinement of π―π» , such that ππ» β πβ.
3.1. Ideal method
To define a generalized finite element method for our problem, we aim to construct a multiscale space πms ofthe same dimension as ππ» , but with better approximation properties. For the construction of such a multiscalespace, let πΌπ» : πβ β ππ» be an interpolation operator that has the projection property πΌπ» = πΌπ» βπΌπ» and satisfies
π»β1πΎ βπ£ β πΌπ»π£βπΏ2(πΎ) + ββπΌπ»π£βπΏ2(πΎ) β€ πΆπΌββπ£βπΏ2(π(πΎ)), βπΎ β π―π» , π£ β πβ, (3.1)
where π(πΎ) := πΎ β² β π―π» : πΎ β² β© πΎ = β . Furthermore, for a shape-regular and quasi-uniform partition, theestimate (3.1) can be summed into the global estimate
π»β1βπ£ β πΌπ»βπΏ2(Ξ©) + ββπΌπ»π£βπΏ2(Ξ©) β€ πΆπΎββπ£βπΏ2(Ξ©),
where πΆπΎ depends on the interpolation constant πΆπΌ and the shape regularity parameter defined as
πΎ := maxπΎβπ―π»
πΎπΎ , where πΎπΎ =diam(π΅πΎ)diam(πΎ)
Β·
Here π΅πΎ denotes the largest ball inside πΎ. A commonly used example of such an interpolant is πΌπ» = πΈπ» βΞ π» ,where Ξ π» is the piecewise πΏ2-projection onto π1(π―π»), the space of functions that are affine on each triangleπΎ β π―π» , and πΈπ» : π1(π―π») β ππ» is an averaging operator that, to each free node π₯ β π© , assigns the arithmeticmean of corresponding function values on intersecting elements, i.e.
(πΈπ»(π£))(π₯) =1
cardπΎ β π―π» : π₯ β πΎ
βπΎβπ―π» :π₯βπΎ
π£πΎ
(π₯).
For more discussion regarding possible choices of interpolants, see e.g. [11] or [31].Let the space πf be defined by the kernel of the interpolant, i.e.
πf = ker(πΌπ») = π£ β πβ : πΌπ»π£ = 0.
1380 P. LJUNG ET AL.
Figure 1. The modified basis function ππ₯ β ππ₯ and the Ritz-projected hat function ππ₯.(A) ππ₯ β ππ₯. (B) ππ₯.
That is, πf is a finescale space in the sense that it captures the features that are excluded from the coarseFE-space. This consequently leads to the decomposition
πβ = ππ» β πf ,
such that every function π£ β πβ has a unique decomposition π£ = π£π» + π£f , where π£π» β ππ» and π£f β πf .In the case of the LOD method for the standard wave equation (see [1]), one considers a Ritz-projection
based solely on the π΅-coefficient to construct a multiscale space. Instead, the goal is to define a multiscalespace based on the inner product π(Β·, Β·) + ππ(Β·, Β·) (for a fixed π) and add additional correction to account forthe time-dependency. This particular choice of scalar product comes from the backward Euler time-steppingformulation and both simplifies the analysis and is more natural in the implementation. Another possibility isto choose π(Β·, Β·) as scalar product. For π£ β ππ» , we consider the Ritz-projection π f : ππ» β πf defined by
π(π fπ£, π€) + ππ(π fπ£, π€) = π(π£, π€) + ππ(π£, π€), βπ€ β πf .
Using this projection, we may define the multiscale space πms := ππ» βπ fππ» such that
πβ = πms β πf , and π(π£ms, π£f) + ππ(π£ms, π£f) = 0. (3.2)
Note that dim(πms) = dim(ππ»), and hence we can view πms as a modified coarse space that contains finescaleinformation of π΄ and π΅. Next, we may use the Ritz-projection to define the basis functions for the space πms.For π₯ β π© , denote by ππ₯ := π fππ₯ β πf the solution to the (global) corrector problem
π(ππ₯, π€) + ππ(ππ₯, π€) = π(ππ₯, π€) + ππ(ππ₯, π€), βπ€ β πf . (3.3)
We can now construct our basis for πms as ππ₯ β ππ₯π₯βπ© which includes the behavior of the coefficients. Foran illustration of the Ritz-projected hat function, as well as the modified basis function for πms, see Figure 1.
We may now formulate our ideal (but impractical) method. Since the solution space can be decomposed asπβ = πms β πf , the idea is to solve a coarse scale problem in πms, and then add additional correction from aproblem on the fine scale. The method reads: find π’π
lod = π£π + π€π, where π£π β πms and π€π β πf such that
π(π2
π‘ π£π, π§)
+ π (π£π, π§) + ππ (π£π, π§) = π (ππ, π§) + π(π’πβ1
lod , π§), βπ§ β πms, (3.4)
π (π€π, π§) + ππ (π€π, π§) = π(π’πβ1
lod , π§), βπ§ β πf , (3.5)
for π β₯ 2 with initial data π’0lod = π’0
β β πms and π’1lod = π’1
β β πms. The initial data is chosen in πms to simplifythe implementation of the finescale correctors. We further discuss this choice in Section 4.4.
GFEM FOR THE STRONGLY DAMPED EQUATION 1381
Remark 3.1. Note that in (3.5), we do not take the source function nor the second derivative into account.This is because we can subtract an interpolant within the πΏ2-product, so that the corresponding error convergesat the same order as the method itself. Moreover, the π£π-part and π€π-part have been excluded from the bilinearform π(Β·, Β·) + ππ(Β·, Β·) in (3.4) and (3.5) respectively, due to the orthogonality between πms and πf .
Note that the multiscale space πms is created using (3.3) with small π . Thus, the π΄-coefficient dominates thesystem for short times. Moreover, we note from (3.5) that for π large enough, we reach a steady state so thatπ€π β π€πβ1 and π£π β π£πβ1. We get for π§ β πf
π(π€π , π§
)+ ππ
(π€π , π§
)β π
(π’π
lod, π§)
= π(π£π , π§
)+ π
(π€π , π§
)= βππ
(π£π , π§
)+ π
(π€π , π§
),
due to the orthogonality. Hence, by rearranging terms we have that
π(π£π , π§
)+ π
(π€π , π§
)= π
(π’π
lod, π§)β 0,
which shows that the solution converges to a state where it is orthogonal with respect to π΅.
3.2. Localized method
The method we have considered so far is based on the global projection (3.3) onto the finescale space πf ,which results in a large linear system that is expensive to solve. Moreover, the basis correctors yield a globalsupport that makes the linear system (3.4) not sparse, but dense. Hence, we wish to localize the computationsonto coarse grid patches in order to yield a sparse matrix system.
To localize the corrector problem, we first introduce the patches to which the support of each basis functionis to be restricted. For π β Ξ©, let π(π) := πΎ β π―π» : πΎ β© π = β , and define a patch ππ(π) of size π as
π1(π) := π(π),
ππ(π) := π(ππβ1(π)), for π β₯ 2.
Given these coarse grid patches, we may restrict the finescale space πf to them by defining
π πf,π := π£ β πf : supp(π£) β ππ(π),
for a subdomain π β Ξ©. In particular, we will commonly use π = π β π―π» and π = π₯ β π© .Next, define the element restricted Ritz-projection π π
f such that π πf π£ β πf is the solution to the system
π(π π
f π£, π§)
+ ππ(π π
f π£, π§)
=β«
π
(π΄ + ππ΅)βπ£ Β· βπ§ dπ₯, βπ§ β πf .
Note that we may construct the global Ritz-projection as the sum
π fπ£ =β
πβπ―π»
π πf π£.
For π β N, we may restrict the projection to a patch by letting π πf,π : ππ» β π π
f,π be such that π πf,ππ£ β π π
f,π solves
π(π π
f,ππ£, π§)
+ ππ(π π
f,ππ£, π§)
=β«
π
(π΄ + ππ΅)βπ£ Β· βπ§ dπ₯, βπ§ β π πf,π.
By summation we yield the corresponding global version as
π f,ππ£ =β
πβπ―π»
π πf,ππ£.
Finally, we may construct a localized multiscale space as πms,π := ππ» βπ f,πππ» , spanned by ππ₯βπ f,πππ₯π₯βπ© .In order to justify the act of localization, it is required that a corrector ππ₯ vanishes rapidly outside an area
of its corresponding node π₯. Indeed, the following theorem (see [27], Thm. 4.1) shows that the corrector ππ₯
satisfies an exponential decay away from its node, making the localization procedure viable.
1382 P. LJUNG ET AL.
Theorem 3.2. There exists a constant π β₯ (8πΆπΌπΎ(2 + πΆπΌ))β1, that only depends on the mesh constant πΎ, suchthat for any π β π―π» and any π£ β π»1
0 (Ξ©) the solution π β πf of the variational problem
(π, π€) =β«
π
(π΄βπ£
)Β· βπ€ dπ₯, βπ€ β πf
satisfies
βπ΄1/2βπβπΏ2(Ξ©βππ(π )) β€β
2 exp(βππΌβ+ππ½β
πΌ++ππ½+π)βπ΄1/2βπ£βπΏ2(π ), βπ β N,
where π΄ = π΄ + ππ΅.
With the space πms,π defined, we are able to localize the computations on the coarse scale system in (3.4)by replacing the multiscale space by its localized counterpart. It remains to localize the computations of thefinescale system in (3.5), which equivalently can be written as
π(ππ‘π€
π, π§)
+ π (π€π, π§) =1π
π(π£πβ1, π§
).
We replace the right hand side by its localized version π£πβ1π β πms,π and note that π£πβ1
π =βπ₯βπ© πΌπβ1
π₯ (ππ₯ βπ f,πππ₯). Thus, we seek our localized finescale solution as π€ππ =
βπ₯βπ© π€π
π,π₯, where π€ππ,π₯ β π π₯
f,π
solves
π(ππ‘π€
ππ,π₯, π§
)+ π
(π€π
π,π₯, π§)
=1π
π(πΌπβ1
π₯ (ππ₯ βπ f,πππ₯) , π§), βπ§ β π π₯
f,π, (3.6)
so that the computation of this equation is localized to a patch surrounding the node π₯ β π© . We introduce thefunctions ππ
π,π₯ β π π₯f,π as solution to the parabolic equation
π(ππ‘π
ππ,π₯, π§
)+ π
(πππ,π₯, π§
)= π
(1π
π1(π) (ππ₯ βπ f,πππ₯) , π§
), βπ§ β π π₯
f,π, (3.7)
for π = 1, 2, . . . , π with initial value π0π,π₯ = 0, and where π1(π) is an indicator function that equals 1 when π = 1
and 0 otherwise. We claim that π€ππ,π₯ =
βππ=1 πΌπβπ
π₯ πππ,π₯ is the solution to (3.6). This follows as for all π§ β π π₯
f,π
π(ππ‘π€
ππ,π₯, π§
)+ π
(π€π
π,π₯, π§)
= π
(ππ‘
πβπ=1
πΌπβππ₯ ππ
π,π₯, π§
)+ π
(πβ
π=1
πΌπβππ₯ ππ
π,π₯, π§
)
=πβ
π=2
πΌπβππ₯
(π(ππ‘π
ππ,π₯, π§
)+ π
(πππ,π₯, π§
))+ πΌπβ1
π₯
(π(ππ‘π
1π,π₯, π§
)+ π
(π1π,π₯, π§
))= 0 + π
(πΌπβ1
π₯ (ππ₯ βπ f,πππ₯) , π§).
With the localized computations established, the GFEM reads: find π’πlod,π = π£π
π + π€ππ , where π£π
π =βπ₯βπ© πΌπ
π₯ (ππ₯ βπ f,πππ₯) β πms,π solves
π(π2
π‘ π£ππ , π§)
+ π (π£ππ , π§) + ππ (π£π
π , π§) = π (ππ, π§) + π(π’πβ1
lod,π, π§)
, βπ§ β πms,π, (3.8)
and π€ππ =
βπ₯βπ©
βππ=1 πΌπβπ
π₯ πππ,π₯, where ππ
π,π₯ β π π₯f,π solves (3.7).
To justify the fact that we localize the finescale equation, we require a result similar to that of Theorem 3.2,but for the functions ππ
π₯ππ=1. We finish this section about localization by proving that these functions satisfy
the exponential decay required for the localization procedure to be viable.
GFEM FOR THE STRONGLY DAMPED EQUATION 1383
Theorem 3.3. For any node π₯ β π© , let πππ₯ β πf be the solution to
π(ππ‘π
ππ₯ , π§)
+ π (πππ₯ , π§) = π
(1π
π1(π) (ππ₯ βπ fππ₯) , π§
), βπ§ β πf ,
with initial value π0π₯ = 0. Then there exist constants π > 0 and πΆ > 0 such that for any π β₯ 1
βπππ₯βπ»1(Ξ©βππ(π₯)) β€ πΆπβππβππ₯βπ»1 ,
for sufficiently small time step π .
Proof. Assume π β₯ 5. First, we analyze the problem for the first time step, which when multiplied by π can bewritten as
π(π1π₯, π§)
+ ππ(π1π₯, π§)
= π (ππ₯ β ππ₯, π§) , βπ§ β πf , (3.9)
where ππ₯ = π fππ₯. We denote = π + ππ such that (ππ₯, π§) = (ππ₯, π§) for all π§ β πf . Furthermore we use theenergy norm |||Β·||| :=
β (Β·, Β·), and by |||Β·|||π· we denote the restriction of the norm onto a domain π·. As seen in
the proof of Theorem 4.1 in [27], the result in Theorem 3.2 can be written as
|||ππ₯|||Ξ©βππ(π₯) β€ πΆππβπ/4β|||ππ₯|||,
for some π < 1. Moreover we define the cut-off function ππ β ππ» by
ππ :=
1, in Ξ©βππ+1 (π₯),0, in ππ (π₯),
for π₯ β π© . Now let π = ππβ3. Then we have that
supp (π) = Ξ©βππβ3 (π₯) ,
supp (βπ) = ππβ2 (π₯) βππβ3 (π₯) .
With this setting, we note that π1π₯
2Ξ©βππ β€
β«Ξ©
ππ΄βπ1π₯ Β· βπ1
π₯ dπ₯ =β«
Ξ©
π΄βπ1π₯ Β· β
(ππ1
π₯
)dπ₯β
β«Ξ©
π΄βπ1π₯ Β· π1
π₯βπ dπ₯
β€β«
Ξ©
π΄βπ1π₯ Β· β (1β πΌπ»)
(ππ1
π₯
)dπ₯
β β
=:π1
+β«
Ξ©
π΄βπ1π₯ Β· βπΌπ»
(ππ1
π₯
)dπ₯
β β
=:π2
+β«
Ξ©
π΄βπ1π₯ Β· π1
π₯βπ dπ₯
β β
=:π3
,
where we have denoted π΄ = π΄ + ππ΅. We now proceed to estimate the terms π1, π2 and π3 separately. Forπ1, we use the problem (3.9) with π§ = (1β πΌπ»)
(ππ1
π₯
)β πf to get
π1 =β«
Ξ©
π΄β (ππ₯ β ππ₯) Β· β (1β πΌπ»)(ππ1
π₯
)dπ₯
=
πβ«
Ξ©βππβ3π΅β (ππ₯) Β· β (1β πΌπ»)
(ππ1
π₯
)dπ₯
,
where we have used the -orthogonality between πms and πf , that the integral is zero on supp (ππ₯), and thatthe support of the remaining integrand is Ξ©βππβ4. Thus, we get that
π1 β€ ππ½+
πΌβ|||ππ₯|||Ξ©βππβ4
π1π₯
Ξ©βππβ4 β€ π
π½+
πΌβπΆππβ
πβ44 β|||ππ₯|||
π1π₯
Ξ©βππβ4 .
1384 P. LJUNG ET AL.
Moreover, by similar calculations as in the proof of Theorem 4.1 in [27], from π2 and π3 we get
π2, π3 β€ πΆ
π1π₯
2ππβππβ4 ,
for a constant πΆ > 0. In total, for π β (0, 1), we find that π1π₯
2Ξ©βππ β€ π
π½+
πΌβπΆππβ
πβ44 β|||ππ₯|||
π1π₯
Ξ©βππβ4 + πΆ
π1π₯
2ππβππβ4
β€ (π½+πΆπ)2
πΌ2βπ
π2π2β πβ44 β|||ππ₯|||2 + π
π1π₯
2Ξ©βππβ4 + πΆ
( π1π₯
2Ξ©βππβ4 β
π1π₯
2Ξ©βππ
).
Let πΏ :=(π + πΆ
)(1 + πΆ
)β1
< 1, and set π = max (πΏ, π) < 1. Then, by rearranging the terms we get theinequality
π1π₯
2Ξ©βππ β€
(π½+πΆπ)2
πΌ2βπ(
1 + πΆ)π2π 2β πβ4
4 β|||ππ₯|||2 + π
π1π₯
2Ξ©βππβ4 .
Repeating the estimate, we end up with
π1π₯
2Ξ©βππ β€ π βπ/4β π1
π₯
2Ξ©
+(π½+πΆπ)2
πΌ2βπ(
1 + πΆ) |||ππ₯|||2
βπ/4ββ1βπ=0
π2π ππ 2β πβ4β4π4 β.
We proceed by estimating
π1π₯
Ξ©
. By choosing π§ = π1π₯ in (3.9) we get
π1π₯
2 β€ |||ππ₯ β ππ₯|||
π1π₯
β€ |||ππ₯|||
π1π₯
,
since
|||ππ₯ β ππ₯|||2 = (ππ₯ β ππ₯, ππ₯ β ππ₯) β€ |||ππ₯ β ππ₯||||||ππ₯|||.
Moreover, for π = 0, 1, 2, . . . , βπ/4β β 1, we note that
π π+2β πβ4β4π4 β β€ π βπ/4ββ1+2β πβ4β4(π/4β1)
4 β = π βπ/4ββ1 (3.10)
so in total we have the estimate π1π₯
Ξ©βππ β€
β1 + πΆ0π2π
12 βπ/4β|||ππ₯|||, with πΆ0 =
(π½+πΆπ)2 π β1
πΌ2βπ(
1 + πΆ) (βπ/4β β 1) .
Recall that this is for the first time step. In next time step, we consider the problem
π(π2π₯, π§)
+ ππ(π2π₯, π§)
= π(π1π₯, π§), βπ§ β πf .
As for the first time step, we split the estimate into the similar integrals π1, π2, and π3, and get
π1 β€ ππ½+
πΌβ
π1π₯
Ξ©βππβ4
π2π₯
Ξ©βππβ4 β€ π
π½+
πΌβ
β1 + πΆ0π2π
12β πβ4
4 β|||ππ₯|||
π2π₯
Ξ©βππβ4 ,
while π2 and π3 remain the same. In total, we get the estimate(1 + πΆ
) π2π₯
2Ξ©βππ β€
π½2+
πΌ2βπ
π2(1 + πΆ0π
2)π β
πβ44 β|||ππ₯|||2 +
(π + πΆ
) π2π₯
2Ξ©βππβ4 .
GFEM FOR THE STRONGLY DAMPED EQUATION 1385
Once again, by letting πΏ =(π + πΆ
)/(
1 + πΆ)
and since πΏ β€ π , we get
π2π₯
2Ξ©βππ β€
π½2+
πΌ2βπ(
1 + πΆ)π2
(1 + πΆ0π
2)π β
πβ44 β|||ππ₯|||2 + π
π2π₯
2Ξ©βππβ4
β€ π βπ/4β|||ππ₯|||2 +π½2
+
πΌ2βπ(
1 + πΆ) (1 + πΆ0π
2) βπ/4ββ1β
π=0
π2π ππ βπβ4β4π
4 β.
Once again we use (3.10) to conclude that π2π₯
Ξ©βππ β€
β1 + πΆ1π2 (1 + πΆ0π2)π
12 βπ/4β|||ππ₯||| =
β1 + πΆ1π2 + πΆ1π2πΆ0π2π
12 βπ/4β|||ππ₯|||,
where
πΆ1 =π½2
+π β1
πΌ2βπ(
1 + πΆ) (βπ/4β β 1) .
Inductively, we get for arbitrary time step π the estimate
|||πππ₯ |||Ξ©βππ β€ π
12 βπ/4β|||ππ₯|||
β―πβ1βπ=0
(πΆ1π2)π + (πΆ1π2)ππΆ0π2. (3.11)
Since π 12 βπ/4β β€ π
12 (π/4β1) = π β
12 πβ
18 log(1/π )π, and since the energy norm is equivalent to the π»1-norm, the
theorem holds for π β₯ 5. We show that the estimate (3.11) is still valid for π β€ 4. For the π:th time step, letπ§ = ππ
π₯ in (3.9). This yields the estimate
|||πππ₯ ||| β€
ππβ1π₯
β€ . . . β€
π1π₯
β€ |||ππ₯|||.
Furthermore if π < 4 we have that π 12 βπ/4β = 1, and under the assumption that πΆ1π
2 < 1, we have that
πβ1βπ=0
(πΆ1π
2)π
=
(πΆ1π
2)π β 1
πΆ1π2 β 1> 1,
which shows that the estimate holds for π < 4. If π = 4 we can bound π 12 βπ/4β β€ π
12 βπ/5β and repeat the same
argument.
Remark 3.4. Note that the sum that appears in (3.11) converges to
πβ1βπ=0
(πΆ1π
2)π βββββ
πββ
11β πΆ1π2
,
which means that the total constant in (3.11) behaves nicely for sufficiently small time steps. More specifically,for time steps
π β€
β―πΌ2βππ
(1 + πΆ
)π½2
+ (βπ/4β β 1)Β·
1386 P. LJUNG ET AL.
4. Error estimates
In this section we derive error estimates of the ideal method (3.4) and (3.5). The additional error due tolocalization can be controlled in terms of the localization parameter π. This is further discussed in Remark 4.9.We begin by considering an auxiliary problem.
4.1. Auxiliary problem
The auxiliary problem is defined as the standard variational formulation for the strongly damped waveequation, but we exclude the second order time derivative. Moreover, we let the starting time π‘ = π‘0 be generaland set the time discretization to π‘ = π‘0 < π‘1 < . . . < π‘π = π . Note that π here might be different from thediscretization of the fully discrete equation (2.6), but the time step π , and thus the multiscale space, are thesame. The auxiliary problem is to find ππ
β β πβ for π = 1, . . . , π , such that
π(ππ‘π
πβ , π£)
+ π (ππβ , π£) = (ππ, π£) , βπ£ β πβ, (4.1)
with initial value π0β β πms. Equivalently, multiply (4.1) by π and we may consider
π (ππβ , π£) + ππ (ππ
β , π£) = π (ππ, π£) + π(ππβ1
β , π£), βπ£ β πβ. (4.2)
Existence of a solution to this problem is guaranteed by LaxβMilgram. For simplicity, we make the assumptionthat the initial data for the damped wave equation (2.6) is already in the multiscale space πms, such that
π’0β = π’0
lod β πms, π’1β = π’1
lod β πms.
For general initial data we refer to Section 4.4 below. Furthermore, to limit the technical details in the proofwe have chosen to analyze the error in the πΏ2
(π»1)-norm instead of the pointwise (in time) π»1-norm.
The solution space can be decomposed as πβ = πmsβπf , such that the solution can be written as ππβ = π£π+π€π
where π£π β πms and π€π β πf . If we insert this into the system in (4.2) and consider test functions π§ β πms, theleft hand side becomes
π (ππβ , π§) + ππ (ππ
β , π§) = π (π£π, π§) + ππ (π£π, π§) ,
where we have used the orthogonality between πms and πf with respect to π (Β·, Β·) + ππ (Β·, Β·). Likewise, if testfunctions π§ β πf are considered, the left hand side becomes
π (ππβ , π§) + ππ (ππ
β , π§) = π (π€π, π§) + ππ (π€π, π§) .
With these findings, we define the approximation to the auxiliary problem as to find ππlod = π£π + π€π, where
π£π β πms and π€π β πf such that
π (π£π, π§) + ππ (π£π, π§) = π (ππ, π§) + π(ππβ1
lod , π§), βπ§ β πms, (4.3)
π (π€π, π§) + ππ (π€π, π§) = π(ππβ1
lod , π§), βπ§ β πf , (4.4)
with initial data π0lod β πms. Note that if π = 0, then ππ
β = ππlod for every π, meaning that the method
reproduces ππβ exactly. For the auxiliary problem, we prove the following error estimates.
Theorem 4.1. Let ππβ be the solution to (4.1) and ππ
lod the solution to (4.3) and (4.4). Assume that π0lodβπ0
β =0, and ππ β πΏ2 (Ξ©), for π β₯ 0, then the error is bounded by
βππβ β ππ
lodβπ»1 β€ πΆπ»
πβπ=1
πβπ πβπΏ2 . (4.5)
GFEM FOR THE STRONGLY DAMPED EQUATION 1387
In addition,
πβπ=1
πβππβ β ππ
lodβ2πΏ2β€ πΆπ»2
πβπ=1
πβπ πβ2πΏ2, (4.6)
and if ππ = ππ‘ππ, for some πππ
π=0 such that ππ β πβ, then
πβπ=1
πβππβ β ππ
lodβ2πΏ2β€ πΆπ»2
ββ πβπ=1
πβππβ2πΏ2+ βπ0β2πΏ2
ββ , (4.7)
where C does not depend on the variations in π΄ or π΅.
Proof. Since ππβ β πβ there are π£π β πms and π β πf such that ππ
β = π£π + π. Let ππ = ππβ β ππ
lod, andconsider
|||ππ|||2 := π (ππ, ππ) + ππ (ππ, ππ)= π (ππ, ππ) + π
(ππβ1
β , ππ)β π (π£π, ππ)β ππ (π£π, ππ)β π (π€π, ππ)β ππ (π€π, ππ) .
For π£π β πms we have due to the orthogonality and (4.3)
π (π£π, ππ) + ππ (π£π, ππ) = π (π£π, π£π β π£π) + ππ (π£π, π£π β π£π)= π (ππ, π£π β π£π) + π
(ππβ1
lod , π£π β π£π).
Similarly, for π€π β πf we use the orthogonality and (4.4) to get
π (π€π, ππ) + ππ (π€π, ππ) = π(ππβ1
lod , π β π€π).
Hence,
|||ππ|||2 = π (ππ, ππ) + π(ππβ1
β , ππ)β π (ππ, π£π β π£π)β π
(ππβ1
lod , π£π β π£π)β π
(ππβ1
lod , π β π€π)
= π (ππ, π β π€π) + π(ππβ1
β β ππβ1lod , ππ
).
The first term can be bounded by using the interpolation operator πΌπ»
π | (ππ, π β π€π) | β€ πβππβπΏ2βπ β π€π β πΌπ» (π β π€π) βπΏ2 β€ πΆπ»πβππβπΏ2βπ β π€πβπ»1
β€ πΆπ»πβππβπΏ2βππβπ»1 β€ πΆπ»πβππβπΏ2 |||ππ|||.
For the second term we note that ππβ1β β ππβ1
lod = ππβ1 so that
|||ππ||| β€ πΆπ»πβππβπΏ2 +
ππβ1
.
Using this bound repeatedly and π0 = 0 we get
|||ππ||| β€ πΆπ»
πβπ=1
πβπ πβπΏ2 .
This concludes the proof since βππβπ»1 β€ πΆ|||ππ|||.To prove the remaining bounds in πΏ2-norm, we define the forward difference operator ππ‘π₯
π =(π₯π+1 β π₯π
)/π
and consider the dual problem: find π₯πβ β πβ for π = πβ 1, . . . , 0, such that π₯π
β = 0 and
π(βππ‘π₯
πβ, π§)
+ π(π₯π
β, π§)
=(ππ+1, π§
), βπ§ β πβ. (4.8)
1388 P. LJUNG ET AL.
Note that this problem moves backwards in time. By choosing π§ = π₯πβ in (4.8) and performing a classical energy
argument, we deduce
βπ₯πββ
2π»1 +
πβπ=π
πβπ₯πββ2π»1 β€ πΆ
πβπ=π+1
πβππβ2πΏ2. (4.9)
Similarly, by choosing π§ = βππ‘π₯πβ, we achieve
βπ₯πββ
2π»1 +
πβ1βπ=π
πβππ‘π₯πββ2π»1 β€ πΆ
πβπ=π+1
πβππβ2πΏ2. (4.10)
Now, use (4.8) to getπβ
π=1
πβππβ2πΏ2=
πβπ=1
ππ(βππ‘π₯
πβ1β , ππ
)+ ππ
(π₯πβ1
β , ππ)
.
Summation by parts givesπβ
π=1
πβππβ2πΏ2=
πβπ=1
ππ(βππ‘π₯
πβ1β , ππ
)+ ππ
(π₯πβ1
β , ππ)
=πβ
π=1
ππ(π₯πβ1
β , ππ‘ππ)
+ ππ(π₯πβ1
β , ππ)
, (4.11)
where we have used π₯π = π0 = 0. Furthermore, we use the equations (4.1) and (4.3), and the orthogonality in(3.2), to show that the following Galerkin orthogonality holds for π§ms β πms
π(ππ‘π
π , π§ms
)+ π
(ππ , π§ms
)= π
(ππ‘π
πβ, π§ms
)+ π
(ππ
β, π§ms
)β 1
ππ(π£π , π§ms
)β π
(π£π , π§ms
)+
1π
π(ππβ1
lod , π§ms
)(4.12)
=(π π , π§ms
)β(π π , π§ms
)= 0.
Let π₯πβ = π₯π
ms + π₯πf , for some π₯π
ms β πms, π₯πf β πf . Using the orthogonality (4.12) and the equations (4.4) and
(4.1) we deduceπβ
π=1
ππ(π₯πβ1
β , ππ‘ππ)
+ ππ(π₯πβ1
β , ππ)
=πβ
π=1
ππ(π₯πβ1
f , ππ‘ππ)
+ ππ(π₯πβ1
f , ππ)
=πβ
π=1
ππ(π₯πβ1
f , ππ‘ππβ
)+ ππ
(π₯πβ1
f , ππβ
)=
πβπ=1
π(π₯πβ1
f , π π)
.
If π π β πΏ2 (Ξ©), then we may subtract πΌπ»π₯πβ1f = 0 and use (3.1) to achieve
πβπ=1
π(π₯πβ1
f , π π)β€ πΆπ»
πβπ=1
πβπ₯πβ1f βπ»1βπ πβπΏ2 β€ πΆπ»
ββ πβπ=1
πβπ₯πβ1f β2π»1
ββ 1/2ββ πβπ=1
πβπ πβ2πΏ2
ββ 1/2
.
Note that
π₯πβ1f
2+
π₯πβ1ms
2 β€ π₯πβ1β
2. Hence the energy estimate (4.9) can now be used to achieve (4.6).
If π π = ππ‘ππ one may use summation by parts to achieve
πβπ=1
πβππβ2πΏ2=
πβπ=1
π(π₯πβ1
f , ππ‘ππ)β€
πβπ=1
π(βππ‘π₯
πβ1f , ππ
)β(π₯0
f , π0)
β€ πΆπ»
πβπ=1
πβππ‘π₯πβ1f βπ»1βππβπΏ2 + πΆπ»βπ₯0
f βπ»1βπ0βπΏ2 ,
where we have used π₯πf = π₯π
β = 0. Using (4.10) we conclude (4.7).
GFEM FOR THE STRONGLY DAMPED EQUATION 1389
Remark 4.2. The bound in (4.6) is not of optimal order, but it is useful in the error analysis.
The next lemma gives error estimates for the discrete time derivative of the error. In the analysis of the (full)damped wave equation we use π = ππ‘π’β, see Lemma 4.5. If the initial data is nonzero we expect ππ‘π
1 below tobe of order π‘β1
1 in πΏ2-norm. A detailed explanation of this is given below. Hence, we have a blow up close tozero due to low regularity of the initial data. Therefore, we need to multiply the error by π‘π . This is similar tothe parabolic case for nonsmooth initial data see, e.g. [33].
Lemma 4.3. Let ππβ be the solution to (4.1) and ππ
lod the solution to (4.3) and (4.4). Assume π0lod β π0
β = 0.If ππ‘π
π β πΏ2 (Ξ©), for π β₯ 1, then
πβπ=2
πβππ‘
(ππ
β β ππlod
)β2πΏ2
β€ πΆπ»2
ββ πβπ=2
πβππ‘ππβ2πΏ2
+ βπ1β2πΏ2
ββ (4.13)
and if ππ = ππ‘ππ, for some πππ
π=0, such that ππ β πβ, then
πβπ=2
πβππ‘
(ππ
β β ππlod
)β2πΏ2
β€ πΆπ»2
ββ πβπ=2
πβππ‘ππβ2πΏ2
+ βππ‘π1β2πΏ2
ββ (4.14)
and, in addition, the following bound holds
πβπ=2
ππ‘2πβππ‘
(ππ
β β ππlod
)β2πΏ2
β€ πΆπ»2
ββ πβπ=2
πβππ‘ππβ2πΏ2
+ π‘21βππ‘π1β2πΏ2
ββ , (4.15)
where C does not depend on the variations in π΄ or π΅.
Proof. The proof of (4.13) is similar to (4.6). Let ππ = ππβ β ππ
lod and define the dual problem
π(βππ‘π₯
πβ, π§)
+ π(π₯π
β, π§)
=(ππ‘π
π+1, π§), βπ§ β πβ, π = πβ 1, . . . , 0, (4.16)
with π₯πβ = 0. Choosing π§ = ππ‘π
π+1 and performing summation by parts we deduce
πβπ=2
πβππ‘ππβ2πΏ2
=πβ
π=2
ππ(βππ‘π₯
πβ1β , ππ‘π
π)
+ ππ(π₯πβ1
β , ππ‘ππ)
(4.17)
=πβ
π=2
ππ(π₯πβ1
β , π2π‘ ππ)
+ ππ(π₯πβ1
β , ππ‘ππ)
+ π(π₯1
β, ππ‘π1),
where we used that π₯πβ = 0. Following the same argument as for (4.7), but with a difference quotient, we arrive
at
πβπ=2
βππ‘ππβ2πΏ2
=πβ
π=2
ππ(π₯πβ1
β , π2π‘ ππ)
+ ππ(π₯πβ1
β , ππ‘ππ)
+ π(π₯1
β, ππ‘π1)
=πβ
π=2
π(π₯πβ1
f , ππ‘ππ)
+ π(π₯1
β, ππ‘π1).
Using π0 = 0, we deduce
π(π₯1
β, ππ‘π1)
=1π
π(π₯1
β, π1)β€ πΆ
ππ»βπ₯1
ββπ»1πβπ1βπΏ2 β€ πΆπ»βπ₯1ββπ»1βπ1βπΏ2 ,
1390 P. LJUNG ET AL.
and with ππ‘ππ β πΏ2 (Ξ©) we get
πβπ=2
βππ‘ππβ2πΏ2
β€ πΆπ»
πβπ=2
πβπ₯πβ1f βπ»1βππ‘π
πβπΏ2 + πΆπ»βπ₯1ββπ»1βπ1βπΏ2 ,
and (4.13) follows by using an energy estimate of π₯πβ similar to (4.9), but with ππ‘π
π on the right hand side.If π π = ππ‘π
π we proceed as for (4.7) to achieveπβ
π=2
π(π₯πβ1
f , ππ‘ππ)β€ πΆπ»
πβπ=2
πβππ‘π₯πβ1f βπ»1βππ‘π
πβπΏ2 + πΆπ»βπ₯1f βπ»1βππ‘π
1βπΏ2
and (4.14) follows by using energy estimates similar to (4.10).For (4.15) we consider the dual problem
π(βππ‘π₯
πβ, π§)
+ π(π₯π
β, π§)
=(π‘π+1ππ‘π
π+1, π§), βπ§ β πβ, π = πβ 1, . . . , 0. (4.18)
A simple energy estimate shows
βπ₯πββ
2π»1 +
πβ1βπ=π
πβππ‘π₯πββ2π»1 β€ πΆ
πβ1βπ=π
ππ‘2π+1βππ‘ππ+1β2πΏ2
, π = 0, . . . , πβ 1. (4.19)
Now choosing π§ = π‘π+1ππ‘ππ+1 in (4.18) and performing summation by parts gives
πβπ=2
ππ‘2πβππ‘ππβ2πΏ2
=πβ
π=2
ππ(βππ‘π₯
πβ1β , π‘πππ‘π
π)
+ ππ(π₯πβ1
β , π‘πππ‘ππ)
(4.20)
=πβ
π=2
(ππ(π₯πβ1
β , π‘ππ2π‘ ππ)
+ ππ(π₯πβ1
β , π‘πππ‘ππ)
+ π(π₯πβ1
β , (π‘π β π‘πβ1) ππ‘ππβ1))
+ π(π₯1
β, π‘1ππ‘π1).
The first two terms of the sum can be handled similarly to (4.14),πβ
π=2
ππ(π₯πβ1
β , π‘ππ2π‘ ππ)
+ ππ(π₯πβ1
β , π‘πππ‘ππ)
=πβ
π=2
π(π₯πβ1
f , π‘πππ‘ππ)
.
Now, using summation by parts we achieveπβ
π=2
π(π₯πβ1
f , π‘πππ‘ππ)
=πβ
π=2
π((βππ‘π₯
πβ1f , π‘ππ
π)β(π₯π
f , ππ))
β(π₯1
f , π‘2π1)
β€ πΆπ»
ββ πβπ=2
π(π‘πβππ‘π₯
πβ1f βπ»1βπ πβπΏ2 + βπ₯π
f βπ»1βπ πβπΏ2
)+ π‘2βπ₯1
f βπ»1βπ1βπΏ2
ββ where we can use (4.19). Note that in the first term we can use the (crude) bound π‘2π β€ π‘2π and let the constantπΆ depend on π . Moreover, we use that π‘2 β€ 2π‘1. We get
πβπ=2
π(π₯πβ1
f , π‘πππ‘ππ)β€ πΆπ»
ββ πβπ=1
ππ‘2πβππ‘ππβ2πΏ2
ββ 1/2βββββ πβ
π=2
πβπ πβ2πΏ2
ββ 1/2
+ π‘1βπ1βπΏ2
βββ .
GFEM FOR THE STRONGLY DAMPED EQUATION 1391
For the third term in (4.20), we use π‘π β π‘πβ1 = π and once again perform summation by parts to get
πβπ=2
ππ(π₯πβ1
β , ππ‘ππβ1)
=πβ
π=2
ππ(βππ‘π₯
πβ1β , ππβ1
),
where we have used π₯πβ = π0 = 0. Combining (4.19) and (4.5) we get
πβπ=2
ππ(βππ‘π₯
πβ1β , ππβ1
)β€ πΆ max
π=1,...,πβππβπ»1
ββ πβπ=2
π
ββ 1/2ββ πβπ=2
πβππ‘π₯πβ1β β2π»1
ββ 1/2
β€ πΆπ»
πβπ=1
πβπ πβπΏ2
ββ πβπ=1
ππ‘2πβππ‘ππβ2πΏ2
ββ 1/2
.
For the last term in (4.20) we use (4.19) and (4.5) for π = 1 to achieve
π(π₯1
β, π‘1ππ‘π1)
= π(π₯1
β, π1)β€ πΆπ»
ββ πβπ=1
ππ‘2πβππ‘ππβ2πΏ2
ββ 1/2
π‘1βπ1βπΏ2 ,
and (4.15) follows by letting π π = ππ‘ππ .
4.2. The damped wave equation
For the error analysis of the full damped wave equation we shall make use of the projection correspondingto the auxiliary problem. For π’π
β β πβ, let ππ = πππ£ + ππ
π€ β πβ with πππ£ β πms and ππ
π€ β πf such that
π (πππ£ β π’π
β, π§) + ππ (πππ£ β π’π
β, π§) = π(ππβ1 β π’πβ1
β , π§), βπ§ β πms, (4.21)
π (πππ€, π§) + ππ (ππ
π€, π§) = π(ππβ1, π§
), βπ§ β πf . (4.22)
Note that since π’πβ solves (2.6), the system (4.21) and (4.22) is equivalent to
π (πππ£ , π§) + ππ (ππ
π£ , π§) = π(ππ β π2
π‘ π’πβ, π§)
+ π(ππβ1, π§
), βπ§ β πms, (4.23)
π (πππ€, π§) + ππ (ππ
π€, π§) = π(ππβ1, π§
), βπ§ β πf . (4.24)
That is, we may view π’πβ and ππ as the solution and approximation to the auxiliary problem with source data
ππ β π2π‘ π’π
β. We deduce following lemma.
Lemma 4.4. Let π’πβ be the solution to (2.6) and ππ the solution to (4.21) and (4.22). The error satisfies the
following bounds
βππ β π’πββπ»1 β€ πΆπ»
πβπ=2
πβπ π β π2π‘ π’π
ββπΏ2 , π β₯ 2, (4.25)
πβπ=2
πβππ β π’πββ
2πΏ2β€ πΆπ»2
ββ πβπ=2
π(βπ πβ2πΏ2
+ βππ‘π’πββ
2πΏ2
)+ βππ‘π’
1ββ2πΏ2
ββ , π β₯ 2, (4.26)
where C does not depend on the variations in π΄ or π΅.
Proof. We let the auxiliary problem (4.1) start at π‘1 with initial data π’1β, such that the error at the initial time
is zero, i.e. π0 = 0. The bound (4.25) now follows directly from (4.5) with ππ β π2π‘ π’π
β as right hand side. Thesecond bound (4.26) follows from (4.6) and (4.7) with ππ β πΏ2 (Ξ©) and ππ = ππ‘π’
π+1β .
1392 P. LJUNG ET AL.
In a similar way me may deduce bounds for the (discrete) time derivative of the error. As a direct consequenceof Lemma 4.3, we get the following result.
Lemma 4.5. Let π’πβ be the solution to (2.6) and ππ the solution to (4.21) and (4.22). The following bounds
hold
πβπ=3
πβππ‘
(ππ β π’π
β
)β2πΏ2
β€ πΆπ»2
ββ πβπ=3
π(βππ‘π
πβ2πΏ2+ βπ2
π‘ π’πββ
2πΏ2
)+ βπ2β2πΏ2
+ βπ2π‘ π’2
ββ2πΏ2
ββ , (4.27)
πβπ=3
ππ‘2πβππ‘
(ππ β π’π
β
)β2πΏ2
β€ πΆπ»2
ββ πβπ=3
π(βππ‘π
πβ2πΏ2+ βπ2
π‘ π’πββ
2πΏ2
)+ π‘22βπ2β2πΏ2
+ π‘22βπ2π‘ π’2
ββ2πΏ2
ββ , (4.28)
where C does not depend on the variations in π΄ or π΅.
Lemma 4.6. Let π’πβ and π’π
lod be the solutions to (2.6) and (3.4), (3.5), respectively. Assume that π’0 = π’1 = 0.The error is bounded by
πβπ=2
πβπ’πlod β π’π
ββ2π»1 β€ πΆπ»2
ββ πβπ=1
π(βπ πβ2πΏ2
+ βππ‘ππβ2πΏ2
)+ max
π=1,...,πβπ πβ2πΏ2
ββ ,
for π β₯ 2, where C does not depend on the variations in π΄ or π΅.
Proof. Begin by splitting the error into two contributions
π’πlod β π’π
β = π’πlod βππ + ππ β π’π
β =: ππ + ππ,
where ππ is the solution to the simplified problem in (4.21) and (4.22). By Lemma 4.4 ππ is bounded by
βππβπ»1 β€ πΆπ»
πβπ=2
π(βπ πβπΏ2 + βπ2
π‘ π’πββπΏ2
),
and we can now apply the energy bound (2.9). It remains to bound ππ. Recall that for any π§ β πβ we haveπ§ = π§ms + π§f for some π§ms β πms and π§f β πf . Using that π’π
lod = π£π + π€π satisfies (3.4) and the orthogonality(3.2) we get (
π2π‘ π’π
lod, π§ms
)+ π
(ππ‘π’
πlod, π§ms
)+ π (π’π
lod, π§ms) = (ππ, π§ms) +(π2
π‘ π€π, π§ms
).
Similarly, due to (3.5) and the orthogonality,(π2
π‘ π’πlod, π§f
)+ π
(ππ‘π’
πlod, π§f
)+ π (π’π
lod, π§f) =(π2
π‘ π’πlod, π§f
).
For ππ we use (4.21) and (4.22) and the orthogonality to deduce(π2
π‘ ππ, π§)
+ π(ππ‘π
π, π§)
+ π (ππ, π§) =(π2
π‘ ππ, π§)
+(ππ β π2
π‘ π’πβ, π§ms
), π§ β πβ,
Hence, ππ satisfies(π2
π‘ ππ, π§)
+ π(ππ‘π
π, π§)
+ π (ππ, π§) =(βπ2
π‘ ππ, π§)β(π2
π‘ π’πβ, π§f
)+(π2
π‘ π’πlod, π§f
)+(π2
π‘ π€π, π§ms
), π§ β πβ,
with π0 = π1 = 0, since π’0lod = π’0
β = π0 and π’1lod = π’1
β = π1. Let ππ =βπ
π=2 πππ . Multiplying by π andsumming over π gives(
ππ‘ππ, π§)
+ π (ππ, π§) + π(ππ, π§
)β€(βππ‘π
π, π§)β(ππ‘π’
πβ β ππ‘π’
1β, π§f
)+(ππ‘π’
πlod β ππ‘π’
1lod, π§f
)(4.29)
GFEM FOR THE STRONGLY DAMPED EQUATION 1393
+(ππ‘π€
π β ππ‘π€1, π§ms
),
where we have used that π1 = π0 = π1 = π0 = 0. Using the interpolant πΌπ» we deduce(ππ‘π’
πβ, π§f
)+(ππ‘π’
πlod, π§f
)+(ππ‘π€
π, π§ms
)β€ πΆπ»
(βππ‘π’
πββπΏ2 + βππ‘π’
πlodβπΏ2
)βπ§βπ»1 + πΆπ»βππ‘π’
πlodβπ»1βπ§msβπΏ2 ,
for 1 β€ π β€ π . Let πΌ (π) = βππ‘π’πββπΏ2 +βππ‘π’
πlodβπ»1 . Since βπ§msβπΏ2 β€ πΆβπ§βπ»1 and πΌ (1) = 0 due to the vanishing
initial data, we get (ππ‘π
π, π§)
+ π (ππ, π§) + π(ππ, π§
)β€(βππ‘π
π, π§)
+ πΆπ»πΌ (π) βπ§βπ»1 , π§ β πβ.
Now, choose π§ = ππ = ππ‘ππ in (4.29). We get
12βππβ2πΏ2
β 12βππβ1β2πΏ2
+ πβππβ2π +12βππβ2π β
12βππβ1β2π β€ πβππ‘π
πβπΏ2βππβπΏ2 + πΆπ»ππΌ (π) βππβπ»1 .
Summing over π gives
βππβ2πΏ2+
πβπ=2
πβππβ2π»1 + βππβ2π»1 β€πβ
π=2
πβππ‘ππβπΏ2βππβπΏ2 + πΆπ»
πβπ=2
ππΌ (π) βππβπ»1 .
Now using that βππβπΏ2 β€ βππβπ»1 and Youngβs weighted inequality, ππ can be kicked back to the left hand side.We deduce
πβπ=2
πβππβ2π»1 β€ πΆ
πβπ=2
πβππ‘ππβ2πΏ2
+ πΆπ»2πβ
π=2
ππΌ (π)2 .
Using Lemma 4.5 we have
πβπ=2
πβππβ2π»1 β€ πΆπ»2
ββ πβπ=2
π(βππ‘π
πβ2πΏ2+ βπ2
π‘ π’πββ
2πΏ2
)+ βπ2
π‘ π’2ββ2πΏ2
ββ + πΆπ»2πβ
π=2
ππΌ (π)2 .
To bound βπ2π‘ π’2
ββ2πΏ2, we consider (2.6) for π = 2 and choose π£ = π2
π‘ π’2β, which gives(
π2π‘ π’2
β, π2π‘ π’2
β
)+ π
(ππ‘π’
2β, π2
π‘ π’2β
)+ π
(π’2
β, π2π‘ π’2
β
)=(ππ‘π
2, π2π‘ π’2
β
).
Due to the vanishing initial data ππ‘π’2β = πβ1π’2
β and π2π‘ π’2
β = πβ2π’2β. We get
βπ2π‘ π’2
ββ2πΏ2+
1π3βπ’2
ββ2π +1π2βπ’2
ββ2π =(π2, π2
π‘ π’2β
), (4.30)
and we deduce
βπ2π‘ π’2
ββ2πΏ2β€ πΆβπ2β2πΏ2
.
All terms, exceptβπ
π=2 πβππ‘π’πlodβ2π»1 that appears in
βππ=2 πΌ2 (π), can now be bounded by using the regularity
in Theorem 2.2. To boundβπ
π=2 πβππ‘π’πlodβ2π»1 we choose π§ = ππ‘π£
π and π§ = ππ‘π€π in (3.4) and (3.5) respectively.
Adding the two equations and using the orthogonality between πms and πf we achieve(π2
π‘ π£π, ππ‘π£π)
+ π(ππ‘π’
πlod, ππ‘π’
πlod
)+ π
(π’π
lod, ππ‘π’πlod
)=(ππ, ππ‘π£
π)β€ πΆπβππβ2πΏ2
+ πβππ‘π£πβ2πΏ2
.
1394 P. LJUNG ET AL.
Note that βππ‘π£πβπΏ2 β€ πΆββππ‘π£
πβπΏ2 β€ πΆ
ππ‘π£π
= πΆ
ππ‘π’πlod
β€ πΆβππ‘π’
πlodβπ , so we may choose π small enough
such that
ππ‘π’πlod
can be kicked to the left hand side. As in the proof of Theorem 2.2 we may now deduce
βππ‘π£πβ2πΏ2
+πβ
π=2
πβππ‘π’πlodβ
2π»1 + βπ’π
lodβ2π»1 β€ πΆ
ββ πβπ=2
βπ πβ2πΏ2+ βππ‘π’
1ββ2πΏ2
+ βπ’1ββ2π»1
ββ , (4.31)
where we have used that π£1 = π’1lod = π’1
β β πms. However, we have assumed vanishing initial data so these termsdisappear. The lemma follows.
Lemma 4.7. Let π’πβ and π’π
lod be the solutions to (2.6) and (3.4), (3.5), respectively. Assume that π = 0. Theerror is bounded by
πβπ=2
ππ‘2πβπ’πlod β π’π
ββ2π»1 β€ πΆπ»2
(βππ‘π’
1ββ2π»1 + βπ’1
ββ2π»1 + βπ’0ββ2π»1
), π β₯ 2,
where C does not depend on the variations in π΄ or π΅.
Proof. We follow the steps in the proof of Lemma 4.6 to equation (4.29). Note that βππβπ»1 can be bounded byLemma 4.4 and the energy bound in (2.9) with π = 0.
Now, let ππ =βπ
π=2 πππ . Choose π§ = ππ = ππ‘ππ in (4.29) and multiply by ππ‘2π. We get
π‘2π2βππβ2πΏ2
βπ‘2πβ1
2βππβ1β2πΏ2
+ ππ‘2πβππβ2π +π‘2π2βππβ2π β
π‘2πβ1
2βππβ1β2π
β€ ππ‘2πβππ‘ππβπΏ2βππβπΏ2 + πΆπ»π‘2ππ (πΌ (π) + πΌ (1)) βππβπ»1 +
(π‘2π β π‘2πβ1
)2
βππβ1β2πΏ2+
(π‘2π β π‘2πβ1
)2
βππβ1β2π .
Summing over π and using π‘2π β π‘2πβ1 β€ 2ππ‘π gives
π‘2πβππβ2πΏ2+
πβπ=2
ππ‘2πβππβ2π»1 + π‘2πβππβ2π»1 β€ πΆ
πβπ=2
ππ‘2πβππ‘ππβπΏ2βππβπΏ2 + πΆπ»
πβπ=2
ππ‘2π (πΌ (π) + πΌ (1)) βππβπ»1 (4.32)
+ πΆ
πβπ=2
ππ‘πβππβ2πΏ2+ πΆ
πβπ=2
ππ‘πβππβ2π .
From the first two sums on the right hand side we can kick π‘πβππβπΏ2 β€ π‘πβππβπ»1 and π‘πβππβπ»1 to the left handside. The remaining two sums needs to be bounded by other energy estimates.
Multiply (4.29) by π and sum over π to get
(ππ, π§) + π(ππ, π§
)+ π
ββ πβπ=2
πππ , π§
ββ β€ (ππ, π§)β(π’π
β β π’1β, π§f
)+(π’π
lod β π’1lod, π§f
)+(π€π β π€1, π§ms
)(4.33)
+ π‘π((
ππ‘π’1β, π§f
)β(ππ‘π’
1lod, π§f
)β(ππ‘π€
1, π§ms
)).
where we have used π1 = π1 = 0. As in the proof of Lemma 4.6 we get
(π’πβ, π§f) + (π’π
lod, π§f) + (π€π, π§ms) β€ πΆπ» (βπ’πββπΏ2 + βπ’π
lodβπΏ2) βπ§βπ»1 + πΆπ»βπ’πlodβπ»1βπ§msβπΏ2 ,
for 1 β€ π β€ π . Let π½ (π) = βπ’πββπΏ2 +βπ’π
lodβπ»1 . Choose π§ = ππ = ππ‘
βππ=1 πππ . Similar to above energy estimates,
we get
βππβ2πΏ2+
πβπ=2
πβππβ2π + βπβ
π=2
πππβ2π β€ πΆ
πβπ=2
πβππβ2πΏ2+ πΆπ»2
πβπ=2
π (π½ (π) + π½ (1) + πΌ (1))2 . (4.34)
GFEM FOR THE STRONGLY DAMPED EQUATION 1395
Sinceβπ
π=2 ππ‘πβππβ2π β€ πΆ (π‘π)βπ
π=2 πβππβ2π we may use (4.34) in (4.32). This gives
π‘2πβππβ2πΏ2+
πβπ=2
ππ‘2πβππβ2π»1 + π‘2πβππβ2π»1 β€ πΆ
πβπ=2
π(π‘2πβππ‘π
πβ2πΏ2+ βππβ2πΏ2
)+ πΆ
πβπ=2
ππ‘πβππβ2πΏ2(4.35)
+ πΆπ»2πβ
π=2
π(π‘2π (πΌ (π) + πΌ (1))2 + (π½ (π) + π½ (1) + πΌ (1))2
).
It remains to bound πΆβπ
π=2 ππ‘πβππβ2πΏ2. For this purpose, choose π§ = ππ = ππ‘π
π in (4.33). Multiply by π‘ππ andsum over π to achieve
πβπ=2
ππ‘πβππβ2πΏ2+ π‘πβππβ2π +
πβπ=2
π‘πππ
(πβ
π=2
πππ, ππ‘ππ
)β€ πΆ
πβπ=1
ππ‘πβππβ2πΏ2+ πΆ
πβπ=2
πβππβ2π (4.36)
+ πΆπ»
πβπ=2
ππ‘π (π½ (π) + π½ (1) + πΌ (1)) βππβπ»1 .
Note that βππβπ»1 is the last sum in only present in the right hand side. The second term on the right hand sideis bounded by (4.34). For the term involving the bilinear form π (Β·, Β·) we use summation by parts to get
βπβ
π=2
ππ
(π‘π
πβπ=2
πππ, ππ‘
πβπ=2
πππ
)β€
πβπ=2
ππ
(π‘ππ
π +πβ
π=2
πππ, ππβ1
)β π
ββπ‘π
πβπ=2
πππ , ππ
ββ β€ πΆ
πβπ=2
ππ‘πβππβ2π + πΆβπβ
π=2
πππβ2π + πΆππ‘2πβππβ2π»1 .
Here the constant πΆπ can be made arbitrarily small due to Youngβs weighted inequality. The first two terms canbe bounded by (4.34). Thus, (4.36) becomes
πβπ=2
ππ‘πβππβ2πΏ2+ π‘πβππβ2π β€ πΆ
πβπ=1
π(π‘πβππβ2πΏ2
+ βππβ2πΏ2
)+ πΆπ»
πβπ=2
ππ‘π (π½ (π) + π½ (1) + πΌ (1)) βππβπ»1
+ πΆπ»2πβ
π=2
π (π½ (π) + π½ (1) + πΌ (1))2 + πΆππ‘2πβππβ2π»1 .
Using this in (4.35) we arrive at
π‘2πβππβ2πΏ2+
πβπ=2
ππ‘2πβππβ2π»1 + π‘2πβππβ2π»1 β€ πΆ
πβπ=2
π(π‘2πβππ‘π
πβ2πΏ2+ π‘πβππβ2πΏ2
)+ πΆπ»2
πβπ=2
π(π‘2π (πΌ (π) + πΌ (1))2 + (π½ (π) + π½ (1) + πΌ (1))2
)(4.37)
+ πΆπ»
πβπ=2
ππ‘π (π½ (π) + π½ (1) + πΌ (1)) βππβπ»1 + πΆππ‘2πβππβ2π»1 .
Using Lemmas 4.5 and 4.4 with π = 0 we deduce for the first two terms in (4.37)
πΆ
πβπ=2
π(π‘2πβππ‘π
πβ2πΏ2+ π‘πβππβ2πΏ2
)β€ πΆπ»2
ββ πβπ=2
πβπ2π‘ π’π
ββ2πΏ2
+ π‘22βπ2π‘ π’2
ββ2πΏ2+ βππ‘π’
1ββ2πΏ2
ββ ,
1396 P. LJUNG ET AL.
where we can use (2.9) for π = 2 and π = 0 to bound π2π‘ π’2
β. We get
π‘22βπ2π‘ π’2
ββ2πΏ2β€ πΆπ
(βππ‘π’
1ββ2π»1 + βπ’1
ββ2π»1
).
For the remaining terms in (4.37) we note that π‘πβππβπ»1 now may be kicked to left hand side using CauchyβSchwarz and Youngβs weighted inequality. The term involving πΆπ can also be moved to the left hand side. Allterms involving πΌ (π) and π½ (π) can be bounded by (2.8) and (4.31). This finishes the proof after using theregularity in Theorem 2.2 with π = 0.
4.3. Error bound for the ideal method
We get the final result by combining the two previous lemmas.
Corollary 4.8. Let π’πβ and π’π
lod be the solutions to (2.6) and (3.4), (3.5), respectively. The solutions can besplit into π’π
β = π’πβ,1 + π’π
β,2 and π’πlod = π’π
lod,1 + π’πlod,2, where the first part has vanishing initial data, and the
second part a vanishing right hand side. The error is bounded by
πβπ=2
πβπ’πβ,1 β π’π
lod,1β2π»1 β€ πΆπ»2
ββ πβπ=1
π(βπ πβ2πΏ2
+ βππ‘ππβ2πΏ2
)+ max
π=1,...,πβπ πβ2πΏ2
ββ , for π β₯ 2
andπβ
π=2
ππ‘2πβπ’πβ,2 β π’π
lod,2β2π»1 β€ πΆπ»2
(βππ‘π’
1ββ2π»1 + βπ’1
ββ2π»1 + βπ’0ββ2π»1
), for π β₯ 2.
Proof. This is a direct consequence of Lemmas 4.6 and 4.7 together with the fact that the problem is linear sothe error can be split into two contributions satisfying the conditions of each lemma.
Remark 4.9. The result from Corollary 4.8 is derived for the ideal method presented in (3.4) and (3.5). TheGFEM in (3.7) and (3.8) will yield yet another error from the localization procedure. However, due to theexponential decay in Theorems 3.2 and 3.3, it holds for the choice π β | log (π») | that the perturbation from theideal method is of higher order and the derived result in Corollary 4.8 is still valid. For the details regardingthe error from the localization procedure, we refer to [29].
4.4. Initial data
For general initial data π’0β, π’1
β β πβ we consider the projections π msπ’0β and π msπ’
1β, where π ms = πΌ β π f is
the Ritz-projection onto πms. Let π£ be the difference between two solutions to the damped wave equation withthe different initial data. From (2.8) it follows that
βπ£β2π»1 β€ πΆ(βππ‘
(π’1
β βπ msπ’1β
)β2πΏ2
+ βπ’1β βπ msπ’
1ββ2π),
where we have chosen to keep the π-norm. For the first term we may use the interpolant πΌπ» to achieve π». Forthe second term use
βπ’1β βπ msπ’
1ββ2π β€
π½+
πΌβ + ππ½β
(βπ’1
β βπ msπ’1ββ2π + πβπ’1
β βπ msπ’1ββ2π). (4.38)
If the initial data fulfills the following condition for some π β πΏ2 (Ξ©)
π(π’1
β, π£)
+ ππ(π’1
β, π£)
= (π, π£) , βπ£ β πβ, (4.39)
then we may deduce
βπ’1β βπ msπ’
1ββ2π + πβπ’1
β βπ msπ’1ββ2π =
(π, π’1
β βπ msπ’1β
)β€ πΆπ»βπβπΏ2βπ’1
β βπ msπ’1ββπ»1 .
GFEM FOR THE STRONGLY DAMPED EQUATION 1397
Figure 2. The behavior of the correction functions πππ,π₯ with increasing π. The time step is
π = 0.01 and π is here chosen so that the support covers the entire interval.
Hence, the error introduced by the projection of the initial data is of order π». The condition (4.39) appearswhen applying the LOD method to classical wave equations, see [1], where it is referred to as βwell prepareddataβ. We note in our case that if π΅ is small compared to π΄, that is if the damping is strong, then the constant in(4.38) is small. In some sense, this means that the condition in (4.39) is of βless importanceβ, which is consistentwith the fact that strong damping reduces the impact of the initial data over time.
5. Reduced basis method
The GFEM as it is currently stated requires us to solve the system in (3.7) for each coarse node in each timestep, i.e. π number of times. We will alter the method by applying a reduced basis method, such that it willsuffice to find the solutions for π < π time steps, and compute the remaining in a significantly cheaper andefficient way.
First of all, we note how the system (3.7) that πππ,π₯ solves resembles a parabolic type equation with no source
term. That is, the solution will decay exponentially until it is completely vanished. An example of how πππ,π₯
vanish with increasing π can be seen in Figure 2, where the coefficients are given as
π΄ (π₯) =(
2β sin(
2ππ₯ππ΄
))β1
and π΅ (π₯) =(
2β cos(
2ππ₯ππ΅
))β1
,
with ππ΄ = 2β4 and ππ΅ = 2β6.In Figure 2 it is also seen how the solutions decay with a similar shape through all time steps. This gives
the idea that it is possible to only evaluate the solutions for a few time steps, and utilize these solutions tofind the remaining ones. This idea can be further investigated by storing the solutions ππ
π,π₯ππ=1 and analyzing
the corresponding singular values. The singular values are plotted and seen in Figure 3. It is seen how thevalues decrease rapidly, and that most of the values lie on machine precision level. In practice, this means thatthe information in ππ
π,π₯ππ=1 can be extracted from only a few ππ
π,π₯βs. We use this property to decrease thecomputational complexity by means of a reduced basis method. We remark that singular value decompositionis not used for the method itself, but is merely used as a tool to analyze the possibility of applying reducedbasis methods.
The main idea behind reduced basis methods is to find an approximate solution in a low-dimensional spaceπ RB
π,π,π₯, which is created using a number of already computed solutions. More precisely, to construct a basisfor this space, one first computes π solutions ππ
π,π₯ππ=1, where π < π . By orthonormalizing these solutions
using e.g. GramβSchmidt orthonormalization, we yield a set of vectors πππ,π₯π
π=1, called the reduced basis.
1398 P. LJUNG ET AL.
Figure 3. The singular values obtained when performing a singular value decomposition ofthe matrix created by storing the finescale corrections ππ
π,π₯ππ=1 with π = 100.
Consequently, the reduced basis space becomes π RBπ,π,π₯ = span
(ππ
π,π₯ππ=1
)for each node π₯ β π© . With this
space created, the procedure of finding πππ,π₯π
π=1 is now reduced to finding πππ,π₯π
π=1, and then approximate theremaining solutions by ππ,rb
π,π₯ ππ=π+1 β π RB
π,π,π₯. More precisely, for π = π+1, π+2, . . . , π we seek ππ,rbπ,π₯ β π RB
π,π,π₯
such that
π(ππ‘π
π,rbπ,π₯ , π§
)+ π
(ππ,rbπ,π₯ , π§
)= π
(1π
π1 (π) (ππ₯ βπ f,πππ₯) , π§
), βπ§ β π RB
π,π,π₯. (5.1)
The matrix system to solve for a solution in π RBπ,π,π₯ is of dimension π Γπ , so when π is chosen small, the
last π βπ solutions are significantly cheaper to compute, which solves the issue of computing π problems onthe finescale space.
When constructing the reduced basis πππ,π₯π
π=1, it is important to be aware of the fact that the solutioncorrections ππ
π,π₯ππ=1 all show very similar behavior. In practice, this implies that many of the ππ
π,π₯βs are linearlydependent, hence causing floating point errors to become of significant size in the RB-space π RB
π,π,π₯. To workaround this issue, one may include a relative tolerance level that removes a vector from the basis if it is tooclose to being linearly dependent to one of the previously orthonormalized vectors. Note that this may implythat we get < π basis vectors in our RB-space instead of π . One may moreover use this tolerance level asa criterion for the amount of solutions, π , to pre-compute. That is, once the first vector is removed from theorthonormalization process, then the RB-space contains sufficient information and no more solutions need tobe added.
In total, the novel method first requires that we solve ππ» number of systems on the localized fine scale inorder to construct the multiscale space πms,π. Moreover, we require to solve a localized fine system ππ» timesfor π time steps to create the RB-space π RB
π,π,π₯ for each coarse node π₯ β π© . By utilizing the RB-space, theremaining π βπ finescale corrections are then solved for in an π Γπ matrix system, and we yield the soughtsolution π’π,rb
lod,π by computing a matrix system on the coarse grid with the multiscale space πms,π. We summarizethe reduced basis approach in Algorithm 1.
GFEM FOR THE STRONGLY DAMPED EQUATION 1399
Figure 4. The two different coefficients used for the numerical examples. The contrast isπΌ+/πΌβ = π½+/π½β = 104. (A) π΄(π₯, π¦). (B) π΅(π₯, π¦).
Algorithm 1: Summary how to compute the finescale correctors πππ,π₯π
π=1 with the RB-approach.
1 Pick π, relative tol;2 for all π₯ β π© do3 for π = 1, 2, . . . , π do4 Solve (3.7) for ππ
π,π₯ and store;
5 Compute reduced basis πππ,π₯
π=1 = gram schmidt(πππ,π₯π
π=1, relative tol);
6 Construct π RB,π,π₯
= span(ππ
π,π₯π=1
);
7 for π = π + 1, . . . , π do
8 Solve (5.1) for ππ,rbπ,π₯ and store;
6. Numerical examples
In this section we illustrate numerical examples for the novel method. At first, we present numerical examplesthat confirm the theoretical findings derived in this paper. In addition, we provide a practical example relatedto seismology where the Marmousi model is used together with the Ricker wavelet as source function.
6.1. Academic example
In this section we present numerical examples that illustrate the performance of the established theory. Forall examples, we consider the domain to be the unit square Ξ© = [0, 1] Γ [0, 1]. The coefficients π΄ (π₯, π¦) andπ΅ (π₯, π¦) used in these examples are generated randomly with values in the interval [10β1, 103], and examplesof such are seen in Figure 4. Moreover, as initial value for each example we set π’0 = π’1 = 0, and the sourcefunction is given by π = 1.
The first example is used to show how the performance is effected by the localization parameter π. Here, weevaluate the solution on the full grid, π’π
lod, and compare it with the localized solution, π’πlod,π, as π varies. For the
example the time step π = 0.02 was used and final time was set to π = 1. The fine and coarse meshes were setto β = 2β7 and π» = 2β4 respectively, and we let π = 2, 3, . . . , 7. The relative error between the functions can
1400 P. LJUNG ET AL.
Figure 5. Relative π»1-error βπ’πlodβπ’π
lod,πβπ»1/βπ’πlodβπ»1 between the non-localized and localized
method, plotted against the layer number π.
be seen in Figure 5. Here we can see how the error decays exponentially as π increases, verifying the theoreticalfindings regarding the localization procedure.
For the second example, the performance of the GFEM in (3.7) and (3.8) depending on the coarse meshwidth π» is shown. For this example, the fine mesh width is set to β = 2β8, and for each coarse mesh widththe localization parameter is set to π = log2 (1/π»). Moreover, the time step is set to π = 0.02 (for the GFEMas well as the reference solution) and the solution is evaluated at π = 1. To compute the error, we use a FEMsolution on the fine mesh as a reference solution. The error as a function of 1/π» can be seen in Figure 6. Here itis seen how the error for the novel method decays faster than linearly, confirming the error estimates derived inSection 4. For comparison, Figure 6 also shows the error of the standard FEM solution, as well as the solutionusing the standard LOD method with correction solely on π΄ and π΅ respectively, i.e. corrections based on thebilinear forms π (Β·, Β·) and π (Β·, Β·) respectively and without finescale correctors. As expected, the error of thesemethods stay at a constant level through all coarse grid sizes.
At last, we compute the solution where the system (3.7) is computed using the reduced basis approach. Forthis example, we let the number of pre-computed solutions π vary, and see how the error between the solutionsπ’π
lod,π and π’π,rblod,π behaves. In the example we have the fine mesh β = 2β8, the coarse mesh π» = 2β5, the time step
π = 0.02, and the final time π = 1. The result can be seen in Figure 7. Here it is seen how the error decreasesrapidly with the amount of pre-computed solutions. Note that it is sufficient to compute approximately 10solutions to yield an error smaller than the discretization error for the main method in Figure 6. This for thecase when the number of time steps are π = 50. We emphasize that a large increment in time steps does notimpact the number of pre-computed solutions π significantly, making the RB-approach relatively more efficientthe more time steps that are considered.
6.2. Marmousi model
We finish by demonstrating the novel method (with the reduced basis approach) on a more practical example.A commonly used model problem in seismology is the Marmousi model, which we use to construct our coefficientsπ΄ (π₯, π¦) and π΅ (π₯, π¦). This is done by applying midpoint quadrature on the image seen in Figure 8a so that thedimensions work with the fine mesh, which is set to β = 2β8. The scales on the coefficients are set to πΌβ = 2,πΌ+ = 5, π½β = 1 and π½+ = 10. As earlier, the spatial domain is set to the unit square and the temporal domain
GFEM FOR THE STRONGLY DAMPED EQUATION 1401
Figure 6. Relative π»1-error βπ’πref β π’π
lod,πβπ»1/βπ’πrefβπ»1 between the reference solution
and the approximate solution computed with the proposed method (without reduced basiscomputations).
Figure 7. Relative π»1-error βπ’πlod,π β π’π,rb
lod,πβπ»1/βπ’πlod,πβπ»1 between the solution with and
without the reduced basis approach, plotted against the number of pre-computed solutions.
to [0, 1], discretized with time step π = 0.02. As source function we use the Ricker wavelet defined as
π (π₯, π¦, π‘) = ππ (π₯, π¦)(
1β 2π2π2 (π‘β π‘β²)2)
πβπ2π2(π‘βπ‘β²)2
where π denotes the frequency, π‘β² is the center of the wavelet and ππ (π₯, π¦) is the indicator function equal to 1 onπ β Ξ© and 0 elsewhere. For our example we let π = 3, π‘β² = 0.5 and π = [0.5β2β, 0.5 + 2β]Γ [0.5β2β, 0.5 + 2β].The temporal behaviour of the wavelet can be seen in Figure 8b. The solution is computed for coarse mesh sizesπ» = 2β1, 2β2, . . . , 2β6, with localization parameter π = log2 (1/π»), and for the construction of each reducedbasis space π RB
π,π,π₯ we pre-compute 15 vectors, i.e. π = 15. The error is computed with respect to a reference
1402 P. LJUNG ET AL.
Figure 8. Left: Marmousi data used to create the coefficients π΄ (π₯, π¦) and π΅ (π₯, π¦). Right: howthe Ricker wavelet π (π‘) varies over time. (A) Marmousi data. (B) Ricker wavelet π(π‘).
Figure 9. Relative π»1-error βπ’πref βπ’π,rb
lod,πβπ»1/βπ’πrefβπ»1 between the full method with reduced
basis approach and the reference solution for the practical example using Marmousi data.
solution evaluated on the fine grid, and is displayed in Figure 9. It seen here how the full method convergesfaster than linearly.
References
[1] A. Abdulle and P. Henning, Localized orthogonal decomposition method for the wave equation with a continuum of scales.Math. Comput. 86 (2017) 549β587.
[2] G. Avalos and I. Lasiecka, Optimal blowup rates for the minimal energy null control of the strongly damped abstract waveequation. Ann. Sc. Norm. Super. Pisa Cl. Sci. 2 (2003) 601β616.
[3] J. Azevedo, C. Cuevas and H. Soto, Qualitative theory for strongly damped wave equations. Math. Methods Appl. Sci. 40(2017) 08.
[4] I. Babuska and R. Lipton, Optimal local approximation spaces for generalized finite element methods with application tomultiscale problems. SIAM J. Multiscale Model. Simul. 9 (2010) 373β406.
[5] I. Babuska and J.E. Osborn, Generalized finite element methods: their performance and their relation to mixed methods. SIAMJ. Numer. Anal. 20 (1983) 510β536.
GFEM FOR THE STRONGLY DAMPED EQUATION 1403
[6] E. Bonetti, E. Rocca, R. Scala and G. Schimperna, On the strongly damped wave equation with constraint. Commun. Part.Differ. Equ. 42 (2017) 1042β1064.
[7] A. Carvalho and J. Cholewa, Local well posedness for strongly damped wave equations with critical nonlinearities. Bull. Aust.Math. Soc. 66 (2002) 443β463.
[8] C. Cuevas, C. Lizama and H. Soto, Asymptotic periodicity for strongly damped wave equations. In: Vol. 2013 of Abstract andApplied Analysis. Hindawi (2013).
[9] N. Dal Santo, S. Deparis, A. Manzoni and A. Quarteroni, Multi space reduced basis preconditioners for large-scale parametrizedPDEs. SIAM J. Sci. Comput. 40 (2018) A954βA983.
[10] M. Drohmann, B. Haasdonk and M. Ohlberger, Adaptive reduced basis methods for nonlinear convection-diffusion equations.In: Vol. 4 of Finite Volumes for Complex Applications VI. Problems & Perspectives. Volume 1, 2, Springer Proc. Math.Springer, Heidelberg (2011) 369β377.
[11] C. Engwer, P. Henning, A. Malqvist and D. Peterseim, Efficient implementation of the localized orthogonal decompositionmethod. Comput. Methods Appl. Mech. Eng. 350 (2019) 123β153.
[12] F. Gazzola and M. Squassina, Global solutions and finite time blow up for damped semilinear wave equations. Ann. Inst. H.Poincare Anal. Non Lineaire 23 (2006) 185β207.
[13] P.J. Graber and J.L. Shomberg, Attractors for strongly damped wave equations with nonlinear hyperbolic dynamic boundaryconditions. Nonlinearity 29 (2016) 1171.
[14] B. Haasdonk, M. Ohlberger and G. Rozza, A reduced basis method for evolution schemes with parameter-dependent explicitoperators. Electron. Trans. Numer. Anal. 32 (2008) 145β161.
[15] P. Henning and A. Malqvist, Localized orthogonal decomposition techniques for boundary value problems. SIAM J. Sci.Comput. 36 (2014) A1609βA1634.
[16] P. Henning, A. Malqvist and D. Peterseim, A localized orthogonal decomposition method for semi-linear elliptic problems.ESAIM: M2AN 48 (2014) 1331β1349.
[17] T.J. Hughes, G.R. Feijoo, L. Mazzei and J.-B. Quincy, The variational multiscale method β a paradigm for computationalmechanics. Comput. Methods Appl. Mech. Eng. 166 (1998) 3β24.
[18] R. Ikehata, Decay estimates of solutions for the wave equations with strong damping terms in unbounded domains. Math.Methods Appl. Sci. 24 (2001) 659β670.
[19] V. Kalantarov and S. Zelik, A note on a strongly damped wave equation with fast growing nonlinearities. J. Math. Phys. 56(2015) 011501.
[20] P. Kelly, Solid Mechanics Part I: An Introduction to Solid Mechanics. University of Auckland (2019).
[21] A. Khanmamedov, Strongly damped wave equation with exponential nonlinearities. J. Math. Anal. Appl. 419 (2014) 663β687.
[22] M. Larson and A. Malqvist, Adaptive variational multiscale methods based on a posteriori error estimation: energy normestimates for elliptic problems. Comput. Methods Appl. Mech. Eng. 196 (2007) 2313β2324.
[23] S. Larsson, V. Thomee and L.B. Wahlbin, Finite-element methods for a strongly damped wave equation. IMA J. Numer. Anal.11 (1991) 115β142.
[24] Y. Lin, V. Thomee and L. Wahlbin, Ritz-Volterra projections to finite element spaces and applications to integro-differentialand related equations. Technical report (Cornell University. Mathematical Sciences Institute). Mathematical Sciences Institute,Cornell University (1989).
[25] A. Malqvist and A. Persson, Multiscale techniques for parabolic equations. Numer. Math. 138 (2018) 191β217.
[26] A. Malqvist and D. Peterseim, Generalized finite element methods for quadratic eigenvalue problems. ESAIM: M2AN 51(2017) 147β163.
[27] A. Malqvist and D. Peterseim, Numerical Homogenization beyond Periodicity and Scale Separation. To appear in SIAMSpotlight (2020).
[28] A. Malqvist and D. Peterseim, Computation of eigenvalues by numerical upscaling. Numer. Math. 130 (2012) 337β361.
[29] A. Malqvist and D. Peterseim, Localization of elliptic multiscale problems. Math. Comp. 83 (2014) 2583β2603.
[30] P. Massatt, Limiting behavior of strongly damped nonlinear wave equations. J. Differ. Equ. 48 (1983) 334β349.
[31] D. Peterseim, Variational multiscale stabilization and the exponential decay of fine-scale correctors. In: Vol. 114 of BuildingBridges: Connections and Challenges in Modern Approaches to Numerical Partial Differential Equations. Lect. Notes Comput.Sci. Eng. Springer, [Cham] (2016) 341β367.
[32] A. Quarteroni, A. Manzoni and F. Negri, Reduced Basis Methods for Partial Differential Equations: An Introduction. Springer(2016).
[33] V. Thomee, Galerkin Finite Element Methods for Parabolic Problems, 2nd edition. Springer Series in Computational Mathe-matics. Springer-Verlag, Berlin (2006).
[34] G.F. Webb, Existence and asymptotic behavior for a strongly damped nonlinear wave equation. Can. J. Math. 32 (1980)631β643.
1404 P. LJUNG ET AL.
This journal is currently published in open access with no charge for authors under a Subscribe-to-Open model (S2O).Open access is the free, immediate, online availability of research articles combined with the rights to use these articlesfully in the digital environment.
S2O is one of the transformative models that aim to move subscription journals to open access. Every year, as long asthe minimum amount of subscriptions necessary to sustain the publication of the journal is attained, the content forthe year is published in open access.
Ask your library to support open access by subscribing to this S2O journal.
Please help to maintain this journal in open access! Encourage your library to subscribe or verify its subscription bycontacting [email protected]
We are thankful to our subscribers and sponsors for making it possible to publish the journal in open access, free ofcharge for authors. More information and list of sponsors: https://www.edpsciences.org/en/maths-s2o-programme