Vol. XX (XXXX) No. X
THEORETICAL FOUNDATIONS OF INCORPORATING LOCAL BOUNDARY CONDITIONS INTO NONLOCAL PROBLEMS
Burak Aksoylu ∗ Department of Mathematics, TOBB University of Economics and Technology Sogutozu Caddesi No:43, Ankara, 06560, Turkey & Department of Mathematics, Wayne State University 656 W. Kirby, Detroit, MI 48202, USA. email: [email protected]
Horst Reinhard Beyer Department of Mathematics, TOBB University of Economics and Technology Sogutozu Caddesi No:43, Ankara, 06560, Turkey & Universidad Polit´ecnicade Uruapan Carretera Carapan, 60120 Uruapan Michoac´an,M´exico& Theoretical Astrophysics, IAAT, Eberhard Karls University of T¨ubingen T¨ubingen72076, Germany. email: [email protected]
Fatih Celiker † Department of Mathematics, Wayne State University 656 W. Kirby, Detroit, MI 48202, USA. email: [email protected] (Received 2016)
We study nonlocal equations from the area of peridynamics on bounded domains. We present four main results. In our recent paper, we discover that, on R, the governing operator in peridynamics, which involves a convolution, is a bounded function of the classical (local) governing operator. Building on this, as main result 1, we construct an abstract convolution operator on bounded domains which is a generalization of the standard convolution based on integrals. The abstract convolution operator is a function of the classical operator, defined by a Hilbert basis available due to the purely discrete spectrum of the latter. As governing operator of the nonlocal equation we use a function of the classical operator, this allows us to incorporate local boundary conditions into nonlocal theories. As main result 2, we prove that the solution operator can be uniquely decomposed into a Hilbert-Schmidt operator and a multiple of the identity operator. As main result 3, we prove that Hilbert-Schmidt operators provide a smoothing of the input data in the sense a square integrable function is mapped into a function that is smooth up to boundary of the domain. As main result 4, for the homogeneous nonlocal wave equation, we prove that continuity is preserved by time evolution. Namely, the solution is discontinuous if and only if the initial data is discontinuous. As a consequence, discontinuities remain stationary.
Keywords: Nonlocal wave equation, nonlocal operator, peridynamics, boundary condition, Hilbert-Schmidt operator, operator theory.
Mathematics Subject Classification (2000): 35L05, 74B99.
∗ Burak Aksoylu was supported in part by the National Science Foundation DMS 1016190 grant, European Commission Marie Curie Career Integration 293978 grant, and Scientific and Technological Research Council of Turkey (TUB¨ ITAK)˙ MFAG 115F473 grant. Research visit of Horst R. Beyer was supported in part by the TUB¨ ITAK˙ 2221 Fellowship for Visiting Scientist Program. Sabbatical visit of Fatih Celiker was supported in part by the TUB¨ ITAK˙ 2221 Fellowship for Scientist on Sabbatical Leave Program. † Fatih Celiker was supported in part by the National Science Foundation DMS 1115280 grant.
[1] B. Aksoylu, H.R. Beyer, and F. Celiker 2
1. Introduction There are indications that a development of nonlocal theories is necessary for description of certain natural phenomena. It is part of the folklore in physics that the point particle model, which is the root for locality in physics, is the cause of unphysical singular behavior in the description of phenomena. On the other hand, all fundamental theories of physics are local. There are many more alternatives for formulating nonlocal theories compared to local ones. Therefore, the task of formulating a viable nonlocal theory consistent with experiment seems much harder than that of a local one. In any case, for such formulation, an understanding of the capacities of nonlocal theories appears inevitable. Considering wave phenomena only partially successfully described by a classical wave equation, it seems reasonable to expect that a more successful model can be obtained by employing the functional calculus of self-adjoint operators, i.e., by replacing the classical governing operator A by a suitable function f(A). We call f the regulating function. By choosing different regulating functions, we can define governing operators tailored to the needs of the underlying application. Since classical boundary conditions (BC) is an integral part of the classical operator, these BC are automatically inherited by f(A). In this way, we vision to model wave phenomena by using appropriate f(A) and, as a consequence, need to study the effect of f(A) on the solutions. One advantage of our approach is that every symmetry that commutes with A also commutes with f(A). As a result, required invariance with respect to classical symmetries such as translation, rotation and so forth is preserved. The choice of regulating functions appropriate for the physical situation at hand is an object for future research. We are interested in studying instances of successful modeling by nonlocal theories of phenomena that cannot be captured by local theories. There are noteworthy developments in the area of nonlocal modeling. For instance, crack propagation [1] and viscoelastic damping [2] are modeled by peridynamics (PD) and fractional derivatives, respectively, both of which are nonlocal. PD is a nonlocal extension of continuum mechanics developed by Silling [1], is capable of quantitatively predicting the dynamics of propagating cracks, including bifurcation. Its effectiveness has been established in sophisticated applications such as Kalthoff-Winkler experiments of the fracture of a steel plate with notches [3,4], fracture and failure of composites, nanofiber networks, and polycrystal fracture [5–8]. Further applications are in the context of multiscale modeling, where PD has been shown to be an upscaling of molecular dynamics [9, 10] and has been demonstrated as a viable multiscale material model for length scales ranging from molecular dynamics to classical elasticity [11]. Also see the review and news articles [12–15] for a comprehensive discussion, and the recent book [16]. Disturbances in solids propagate in form of waves. The wave equation is the basic model for the descrip- tion of the evolution of deformations. The driving application is PD, whose equation of motion corresponds exactly to the nonlocal wave equation under consideration. The same operator is also employed in nonlo- cal diffusion [12, 17, 18]. Similar classes of operators are used in numerous applications such as population models, particle systems, phase transition, coagulation, etc. Nonlocal operators have also been used in image processing [19]. In addition, we witness a major effort to meet the need for mathematical theory for PD applications and related nonlocal problems addressing, for instance, conditioning analysis, domain decomposition and variational theory [20–22], discretization [22–25], nonlinear PD [26, 27], convergence of solutions [28–32]. Since PD is a nonlocal theory, one might expect only the appearance of nonlocal BC, and, indeed, so far the concept of local BC does not apply to PD. Instead, external forces must be supplied through the loading force density b [1]. On the other hand, we demonstrate that the anticipation that local BC are incompatible with nonlocal operators is not quite correct. Our approach presents an alternative to nonlocal BC and we hope that the ability to enforce local BC provides a remedy for surface effects seen in PD; see [16, Chap. 4, 5, 7, and 12] and [33, 34]. We also hope that the applicability of PD can be extended to problems that require local BC such as contact, shear, and traction. There are four main results in this paper. Throughout the paper, we label the related results with the following identifiers. • Main Result 1: On a bounded domain Ω, utilizing the eigenbasis of the classical operator which allows us to incorporate local BC, we construct a convolution operator C∗ that possesses the Hilbert- Schmidt property.
• Main Result 2: The solution operator of the nonlocal wave equation utt(x, t) + (c − C∗)u(x, t) = 0 B. Aksoylu, H.R. Beyer, and F. Celiker 3
can be uniquely decomposed into a Hilbert-Schmidt operator and a multiple of the identity operator.
• Main Result 3: Hilbert-Schmidt operators provide a smoothing of the input data in the sense an L2(Ω) function is mapped into a function that is smooth up to boundary of Ω.
• Main Result 4: The solution to the homogeneous nonlocal wave equation is discontinuous if and only if the initial data is discontinuous. Hence, discontinuities remain stationary. The rest of the article is structured as follows. Several aspects of constructing convolutions are studied in Sec. 2.. In Sec. 2.1., we introduce the proposed convolution definition on Hilbert spaces. The concept of regulating function is central to our study. In Sec. 2.2., we demonstrate that the regulating function for the bounded domain is similar to that for the unbounded domain. Hence, our construction of convolutions is an appropriate generalization. In Sec. 2.3., we study properties of operators which comprise Main Result 1, i.e., the convolution operator enjoys the Hilbert-Schmidt property. In Sec. 2.4., we provide diagonalizations of the classical governing operator A and functions, f(A), of A. The governing nonlocal operator is a function of A. On the other hand, the representation of the solution of the initial value problem corresponding to f(A) contains bounded functions of f(A). Hence, in Sec. 2.5., we prove that bounded functions of f(A) are bounded functions of A. When we refer to operators that are (bounded) functions of functions of A, we mean solution operators as in Sec. 2.5.. In Sec. 2.6., we choose an instance of f(A) that corresponds to the PD governing operator in the unbounded domain case. We present important properties such as boundedness and self-adjointness of the PD governing operator, and provide the conditions that make it a function of A. We also present the form the solution operators take with this concrete choice of f(A). The main purpose of this paper is to provide a rigorous construction which gives the ability to satisfy local BC for nonlocal operators. In Sec. 3., we find that Hilbert-Schmidt operators play a crucial role in satisfying BC. By reflecting on the representation of the solution, in Sec. 3.1., in order to satisfy BC, we outline a strategy built on obtaining the Hilbert-Schmidt property. In Sec. 3.2., we seek the tools to obtain the Hilbert-Schmidt property. We reveal that the aforementioned strategy is based on a unique decomposition of the solution operator into a Hilbert-Schmidt operator and a multiple of the identity operator, referred to as Main Result 2. The main tool to obtain Main Result 2 turns out to be holomorphic functional calculus. We prove that Hilbert-Schmidt operators remain to be Hilbert-Schmidt under the action of holomorphic functions that vanish at zero. In Sec. 3.3., we rigorously show that BC are satisfied. For satisfying BC, since the key ingredient turns out to be the Hilbert-Schmidt property, we prove how this property leads to a uniform convergence argument which allows us to interchange limits, thereby, automatically satisfying BC. In addition, the Hilbert-Schmidt property leads to smoothing of the input in the sense an L2 function is mapped into a function that is continuous up to the boundary, which is what we refer to as Main Result 3. With additional decay conditions of the eigenvalues, we reach a uniform convergence argument also for derivatives which allows us to interchange limits. As a consequence, BC that involve derivatives are also automatically satisfied (also, Main Result 3). We close with Main Result 4, namely with the proof that continuity is preserved by time evolution for the homogeneous nonlocal wave equation. In other words, the solution is discontinuous if and only if initial data is discontinuous. This important result indicates that discontinuities of the solution remain stationary. We conclude in Sec. 4. by depicting the solutions of the nonlocal wave equation with discontinuous and continuous initial data. We also numerically verify Main Result 4 by observing that, for the homogeneous equation, discontinuities enter the solution only when the initial data is discontinuous. As a consequence, discontinuities remain stationary. In addition, we also observe that away from discontinuities, the solution is smooth, which numerically verifies Main Result 3. A comprehensive study regarding the application and implementation of the proposed construction in this paper is reported in our paper [35]. To improve the accessibility of the paper, we collect the treatment of inhomogeneous nonlocal wave equation in a separate section, which comprises the Appendix. B. Aksoylu, H.R. Beyer, and F. Celiker 4
2. Construction for the Bounded Domain PD equation of motion falls into the following class of generalized wave equations for u, x ∈ R
utt(x, t) + f(A)u(x, t) = 0, (1) with a micromodulus function C ∈ L1(R), the governing operator is defined by Z Z f(A)u(x, t) := C(y)dy u(x, t) − C(x − y)u(y, t)dy. (2) R R All results in the paper are also valid for x ∈ Rn. Furthermore, our previous paper [36, Rem. 4] provides a basis for the treatment of u ∈ Rn. Carrying out the details is the subject of ongoing work. Our driving application is PD which requires the dimension of u to be equal to that of x; for the correspondence of (1)–(2) to the bond based linearized PD, see [37, Eq. 23] and [38, Eq. 3]. That is why, we restrict the presentation to u, x ∈ R throughout the paper. The case of u ∈ R and x ∈ Rn corresponds to nonlocal diffusion [12, 18] to which our results apply in a straightforward manner. In the unbounded domain case, in [36], we discovered that PD uses as governing operator a function f(A) of the classical (local) operator A. 1 Practical applications call for a bounded domain. The question is the generalization of convolutions to functions on bounded domains, while preserving the Banach algebra struc- ture of convolutions for functions on L1(R). Indeed, such a generalization is known for periodic functions. We aim to accommodate general BC, not just periodic, and we show this in Sec. 2.1..
2.1. Convolutions on Hilbert Spaces In the unbounded domain case, in [36], we discovered that the governing nonlocal operator is a function of a multiple of the classical governing operator. Therefore, for the bounded domain case, it is natural to define the governing operator as a function of the corresponding classical operator. This opens a gateway to incorporate local BC to nonlocal theories. For simplicity, we choose the classical operator to be a multiple of the Laplace operator with appropriate BC. In the bounded domain case, the spectrum of the Laplace operator with classical BC such as periodic, antiperiodic, Neumann, Dirichlet, and Robin is purely discrete. Furthermore, we can explicitly calculate the eigenfunctions ek corresponding to each BC and the subscript signifies the BC used; BC ∈ {p, a, N, D} where p, a, N, and D stand for periodic, antiperiodic, Neumann, Dirichlet, respectively. These eigenfunctions form a Hilbert basis (complete orthonormal basis) through which the abstract convolution can be defined as follows:
X 2 C ∗BC u := hek|Ci hek|ui ek, (3) k where Z 1 ∗ hek|ui := ek(y)u(y)dy. −1 Let Ω denote the interval (−1, 1) which we refer to as the bounded domain throughout the paper. Inspired by the governing equation (1) on the unbounded domain, we define the nonlocal wave equation
utt(x, t) + ϕ(ABC)u(x, t) = 0, x ∈ Ω, t > 0, (4) where ϕ : σ(ABC) → R is a bounded function and ABC is the classical operator with spectrum σ(ABC). For 3 instance, the convolution in (3) defines the governing operator c − C∗BC where c is a suitable constant regulating function ϕ : σ(ABC) → R is defined as
ϕ(ABC) := c − C ∗BC . (5)
1Throughout the paper, we refer to the Laplace operator as the classical operator. 2If there is a specified BC in the context, then the corresponding eigenfunctions are assumed to satisfy that BC. A pedantic P BC BC BC BC notation would be C ∗BC u := hek |Ci hek |ui ek . For brevity, we suppress the BC identifier of ek , use ek instead. 3 R In the PD context, the choice of c in practice is Ω C(y)dy with C ≥ 0. B. Aksoylu, H.R. Beyer, and F. Celiker 5
Representing u in this Hilbert basis, X u = hek|ui ek, we arrive at X X (c − C∗BC)u = [c − hek|Ci] hek|ui ek = ϕ(λk) hek|ui ek,
where (λk, ek) denotes an eigenpair of the classical operator. Since the last relation regulates which function of the classical operator is used to define the nonlocal operator, we call ϕ(λk) the regulating function:
ϕ(λk) = c − hek|Ci . (6)
A comparative study of nonlocal operators obtained by different regulating functions is reported in our paper [39].
2.2. Similarity Between Regulating Functions
We demonstrate the similarity between the regulating functions f(A) and ϕ(ABC) for the cases of periodic and Neumann BC, i.e., BC = p, N. By borrowing operator theory tools from our recent paper [36], we first 1 reveal the regulating function f(A) by diagonalizing it employing Fourier transforms F1 : L (R) → C∞(R, C) 2 2 and F2 : L (R) → L (R) where C∞(R, C) denotes the space of complex valued continuous functions on R vanishing at infinity. In addition, let Tλ2 denote the maximal multiplication operator by the spectral value of the classical operator λ2 where λ ∈ [0, ∞). First, we diagonalize A and f(A):
−1 A = F2 ◦ Tλ2 ◦ F2, −1 f(A) = F2 ◦ Tf(λ2) ◦ F2. (7) Then, we directly diagonalize f(A) by using the expression given in (2)
−1 f(A) = F2 ◦ TF1C(0)−F1C(λ) ◦ F2. (8) Therefore, a comparison of (7) and (8) yields
2 f(λ ) = F1C(0) − F1C(λ). (9)
From this point forward, we assume that C is an even function. Namely,
C(−y) = C(y). (10)
By explicitly rewriting the Fourier transform in (9) and using (10), we get Z Z f(λ2) = e−i0yC(y)dy − e−iλyC(y)dy ZR Z R = ei0yC(y)dy − eiλyC(y)dy. R R Hence, we conclude that 2 f(λ ) = he0 − eλ|Ci . (11) BC Note that for BC = p, N, we have e0 = 1. By recalling (6), we conclude
2 BC BC ϕ(λk) = he0 − ek |Ci . (12)
2 2 We see from (11) and (12) that ϕ(λk) and f(λ ) project the micromodulus function C on the difference between the lowest and k-th eigenmode, and λ-th spectral mode, respectively. In conclusion, we infer that the regulating function ϕ(ABC) obtained from the abstract convolution (3) has a similar structure to f(A). B. Aksoylu, H.R. Beyer, and F. Celiker 6
2.3. Properties of the Convolution Operator We aim to generalize convolutions on R to bounded domains by preserving the Banach algebra structure of convolutions for functions on L1(R). The following is a formal definition of convolutions on Hilbert spaces which leads to a Banach algebra structure. The conventional convolution in the governing equation is in the form of an integral as in (2). Typically, the kernel function C is taken to belong to L1(Ω). Here, we modify the function space to be L2(Ω) instead of L1(Ω), a departure from the conventional construction. The following abstract Hilbert space X is taken to be L2(Ω) when we deal with the classical operator. Hence, the concrete choice of X is L2(Ω).
Theorem 1 (Main Result 1: Convolutions on Hilbert Spaces) Let K ∈ {R, C}, (X, h | i) a non-trivial K-Hilbert space with induced norm k k and M ⊂ X a Hilbert basis. By X ξ ∗ η := he|ξi he|ηi e (13) e∈M(ξ)∩M(η)
for ξ, η ∈ X, where M(ξ) := {e ∈ M : he|ξi= 6 0},M(η) := {e ∈ M : he|ηi= 6 0},
and the sum over the empty set is defined as 0X , there exists a bilinear, commutative and associative map
∗ : X2 → X (ξ, η) 7→ ξ ∗ η
such that kξ ∗ ηk 6 kξkkηk for all ξ, η ∈ X and, as a consequence, (X, +,., ∗, k k) is a commutative Banach algebra.
Proof. We note for ξ, η ∈ X that, as intersection of at most countable subsets of X, M(ξ) ∩ M(η) is at most countable. Furthermore, for every finite subset S of M(ξ) ∩ M(η)
X 2 X 2 2 X 2 X 2 | he|ξi he|ηi | = | he|ξi | | he|ηi | 6 | he|ξi | | he|ηi | e∈S e∈S e∈S e∈S X 2 X 2 2 2 6 | he|ξi | | he|ηi | = kξk kηk . e∈M(ξ) e∈M(η)
2 Hence, the sequence (| he|ξi he|ηi | )e∈M(ξ)∩M(η) is (absolutely) summable. As a consequence, the sequence (he|ξi he|ηi e)e∈M(ξ)∩M(η) is summable, with a sum that is independent of the order of summation. Therefore, ξ ∗ η is well-defined and satisfies
2 X 2 2 2 kξ ∗ ηk = | he|ξi he|ηi | 6 kξk kηk . e∈M(ξ)∩M(η)
That ∗ is bilinear, commutative, and associative is obvious. Throughout the paper, we denote the classical operator by A and a function of A by f(A). We want to establish results for (4) which involve ABC and ϕ(ABC). When we emphasize that BC have been enforced on a bounded domain, we use ABC and ϕ(ABC) instead of A and f(A); see (1) and (4). The results we establish hold for general a self-adjoint operator A with purely discrete spectrum σ(A), a general f ∈ B(σ(A), R), and the abstract space X. Most of the time, we present the material using A and f(A) instead of ABC and ϕ(ABC). We present properties of operators induced by convolutions on Hilbert spaces. We prove that such operators are Hilbert-Schmidt and are bounded functions of A.
Corollary 1 (Main Result 1: Convolution operator is Hilbert-Schmidt) Let K, (X, h | i), k k, M and ∗ be as in Theorem1. Then for every ξ ∈ X, ξ ∗ · ∈ L(X,X) and, in addition, Hilbert-Schmidt and therefore also compact. B. Aksoylu, H.R. Beyer, and F. Celiker 7
Proof. Let ξ ∈ X. First, it follows from Theorem1 that ξ ∗· ∈ L(X,X). Furthermore, from the definition of ∗, it follows for every e ∈ M that ξ ∗ e = he|ξi e. In particular, this implies that the set
M(A) := {e ∈ M : ξ ∗ e 6= 0}