The Pennsylvania State University The Graduate School
THE ANALYSIS OF THE PERIDYNAMIC THEORY OF SOLID
MECHANICS
A Dissertation in Mathematics by Kun Zhou
c 2012 Kun Zhou
Submitted in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy
May, 2012 The thesis of Kun Zhou was reviewed and approved∗ by the following:
Qiang Du Verne M. Willaman Professor of Mathematics Dissertation Advisor, Chair of Committee
Ludmil Zikatanov Professor of Mathematics
Wen Shen Professor of Mathematics
Xiantao Li Professor of Mathematics
Suzanne Shontz Professor of Computer Science and Engineering
Svetlana Katok Professor of Mathematics Graduate Program Chair
∗Signatures are on file in the Graduate School. Abstract
The peridynamic model proposed by Silling [1] is an integral-type nonlocal contin- uum theory. It depends crucially upon the non-locality of the force interactions and does not explicitly involve the notion of deformation gradients. It provides a more general framework than the classical theory for problems involving discontinuities or other singularities in the deformation. In this dissertation, we focus on the re- cent developed peridynamic models including the ordinary bond-based, state-based models. The linear ordinary bond-based peridynamic model is analyzed under a rigorous analytical framework. Meanwhile the relation between the peridynamic energy space and fractional sobolev spaces is established for various micromodulus functions. And for better assisting the nonlocal mechanical modeling and nonlocal mathematical analysis, a vector calculus for the nonlocal operators is developed. Nonlocal analogs of several theorem and identities of the vector calculus for differ- ential operators are also presented. Relationships between the nonlocal operators and their differential counterparts are established. We further apply the nonlocal vector calculus to express the constitutive relation for the ordinary state-based peridynamic elastic material. The linear peridynamic models and associated non- local volume-constraints problems are defined and analyzed within the nonlocal vector calculus framework. Especially, the well-posedness of the ordinary state- based peridynamic model for a linear homogeneous and anisotropic material is demonstrated. Moreover, we establish relation between the classical elasticity and nonlocal peridynamic theory as the nonlocal horizon converges to zero. And un- der the mathematical framework introduced, we conduct the numerical analysis of the finite-dimensional approximations to the bond-based peridynamic models. A posterior error estimator for the peridynamic model is also proposed and studied. To recover the full elasticity theory in its local limit, we also developed a peri- dynamic double-bonds model. Finally an open issue involved in the peridynamic crack nucleation theory is discussed.
iii Table of Contents
List of Figures viii
Acknowledgments ix
Chapter 1 Overview 1 1.1 Motivation to the peridynamic theory ...... 1 1.2 The current linear PD models ...... 2 1.2.1 The ordinary bond-based PD model ...... 3 1.2.2 The ordinary state-based PD model ...... 4 1.3 Content of the dissertation work ...... 6
Chapter 2 Mathematical Analysis of Bond-based Linear PD models 10 2.1 Introduction ...... 10 2.2 Bond-based PD model on Rd ...... 12 2.2.1 Mathematics analysis of the PD model ...... 12 2.2.2 Space equivalence for special influence functions ...... 19 2.2.3 Properties of stationary PD Model ...... 24 2.2.4 The time-dependent PD model ...... 26 2.3 The bond-based PD model on a finite bar ...... 29 2.3.1 The PD operators and related function spaces ...... 30 2.3.2 Space equivalence for special influence functions ...... 32 2.3.3 Properties of the stationary PD Model ...... 35 2.3.4 The time-dependent PD model on a finite bar ...... 37 2.4 Conclusion ...... 38
iv Chapter 3 A Nonlocal Vector Calculus, Nonlocal Balance Laws and Non- local Volume-Constrained Problems 39 3.1 Introduction ...... 39 3.1.1 Notation ...... 41 3.2 Nonlocal fluxes and nonlocal action-reaction principles ...... 43 3.2.1 Local fluxes ...... 43 3.2.2 Nonlocal fluxes ...... 44 3.3 Nonlocal operators ...... 45 3.3.1 Nonlocal point divergence, gradient, and curl operators . . . 45 3.3.2 Nonlocal adjoint operators ...... 46 3.3.3 Further observations and results about nonlocal operators . 48 3.3.3.1 A nonlocal divergence operator for tensor functions and a nonlocal gradient operator for vector functions 48 3.3.3.2 Nonlocal vector identities ...... 49 3.3.3.3 Nonlocal curl operators in two and higher dimensions 51 3.4 A nonlocal vector calculus ...... 52 3.4.1 Nonlocal interaction operators ...... 53 3.4.2 Nonlocal integral theorems ...... 54 3.4.3 Nonlocal Green’s identities ...... 57 3.4.4 Special cases of the vector calculus ...... 59 3.4.4.1 The free space vector calculus ...... 59 3.4.4.2 The vector calculus for interactions of infinite extent 60 3.4.4.3 The vector calculus for localized kernels ...... 60 3.5 Relations between nonlocal and differential operators ...... 61 3.5.1 Identification of nonlocal operators with differential opera- tors in a distributional sense ...... 62 3.5.2 Relations between weighted nonlocal operators and weak representations of differential operators ...... 65 3.5.2.1 Nonlocal weighted operators ...... 66 3.5.2.2 Relationships between weighted operators and dif- ferential operators ...... 68 3.5.3 A connection between the nonlocal and local Gauss theorems 74 3.6 The nonlocal volume-constrained problems ...... 75 3.7 Local and nonlocal balance laws ...... 78 3.7.1 Abstract balance laws ...... 79 3.7.1.1 Abstract local balance laws ...... 82 3.7.1.2 Abstract nonlocal balance laws ...... 83 3.7.2 The peridynamics nonlocal theory of continuum mechanics . 84
v Chapter 4 Application of the Nonlocal Vector Calculus to the Analysis of state-based Linear PD Materials 88 4.1 Introduction ...... 88 4.2 Useful identities in the nonlocal vector calculus ...... 89 4.3 Constitutive relations in peridynamic modeling ...... 90 4.4 Variational principles for linear peridynamic models ...... 92 4.4.1 Variation of the potential energy ...... 92 4.5 Well-posedness of the state-based PD solid models ...... 97 4.5.1 Decomposition of the solution space ...... 101 4.5.2 Nonlocal dual spaces and nonlocal trace spaces ...... 102 4.5.3 Well-posedness of variational problems ...... 103 4.5.4 An example of the peridynamic energy space ...... 104 4.6 Concluding remarks ...... 112
Chapter 5 Convergence of the PD models to the classical elasticity theory 114 5.1 Introduction ...... 114 5.2 Model reformulation ...... 115 5.3 Convergence of the bond-based PD model to the classical elasticity 116 5.4 The convergence of the ordinary state-based PD equation to the classical elasticity ...... 121 5.5 Conclusion ...... 126
Chapter 6 The Numerical Analysis of the bond-based PD model 127 6.1 Introduction ...... 127 6.2 Numerical Analysis of the Bond-based PD model on 1-D bar . . . . 128 6.2.1 Finite-dimensional approximations ...... 128 6.3 A posterior error estimate of bond-based PD equations ...... 136 6.3.1 Finite element discretization and a posterior error estimate . 138 6.3.2 Relation with local case ...... 141 6.4 Conclusion ...... 149
Chapter 7 The PD double-bonds model 151 7.1 Introduction ...... 151 7.2 The full nonlocal Peridynamic double-bonds model ...... 151 7.3 Relation between PD double-bonds and PD bond-based models . . 158 7.4 Conclusion ...... 159
vi Chapter 8 An open issue – The nonlinear simulation 160 8.1 Introduction ...... 160 8.2 A one dimensional experiment ...... 162 8.2.1 Simulation method ...... 163 8.2.2 Simulation results ...... 164 8.3 Open questions ...... 168
Bibliography 169
vii List of Figures
1.1 Point q interacts indirectly with x even though they are outside each other’s horizon, since they are both within the horizon of in- termediate points such as p ...... 4
3.1 Four of the possible configurations for Ωs and Ωc...... 53 3.2 For a localized kernel, the domain Ωs and the interaction regions Ωc and Ω whose thicknesses are given by the horizon ε...... 61 − 4.1 Four of the possible configurations for Ω = (Ω Ω ) (Ω Ω )0 . 93 s ∪ c ∪ s ∩ c
6.1 Ωs and Ωc...... 142 6.2 K = Kin Kout, an illustration in the two-dimensional space. . . . . 143 6.3 The δ-neighborhood∪ of an edge e...... 146 6.4 A projection of x onto e...... 146 6.5 An illustration of the integral domain SC(δ, δ + ξ) that denotes the spherical cap of a sphere with radius δ and height δ + ξ...... 147
8.1 Displacement field with δ = 0.1, h = 0.005, t = 0.00001 and t = 0.2 164 8.2 Compare the displacement under the same4 condition but different grid sizes ...... 165 8.3 Compare the displacement under the same condition but different time step ...... 166 8.4 P(x) for the particle in the middle of the bar, i.e. P (0), with mag- nitude of external force 85 ...... 166 8.5 Displacements at t = 0.15, t = 0.19303 and t = 0.24 ...... 167 8.6 Displacements at t = 0.15, t = 0.19303 and t = 0.24 ...... 167 8.7 ∂f/∂η at t = 0.15, t = 0.19303 and t = 0.24 ...... 168
viii Acknowledgments
It is a pleasure to thank those who made this thesis possible. First and foremost, I would like to show my deepest gratitude to my thesis advisor Professor Qiang Du. During my PhD studies, I received extraordinary support from him; his brilliant and patient guidance enabled me to develop a deep understanding of mathematics and a set of skills tackling quantitative challenges; his generous support exempted me from heavy teaching load and provided me plenty of time to think and con- duct my research project. Besides, his great personality and rigorous scholarship influence me a lot. I would also like to thank Professor Max Gunzburger and Dr. Richard B. Lehoucq for their invaluable educations on mathematics, writing, logicality and collaboration. Every communication with them through email, during the confer- ences is exciting and fruitful. They have made available their support in many ways. Moreover I would like to show my gratitude to Dr. Michael L. Parks for his insights and enthusiasm in the conferences in Philadelphia and Lincoln and during my visit to Sandia National Labs. Likewise, I am grateful to Professor Ludmil Zikatanov, Professor Xiantao Li and Professor Wen Shen from math department and Professor Suzanne Shontz from department of computer science and engineering for their time to serve on my committee and for their nice classes that I attended during my PhD studies. Further, I would like to give my appreciation to my friends in the group, Lei Zhang, Tianjiang Li, Manlin Li, Yanxiang Zhao, Yanping Ma and Li Tian as well as Professor Lili Ju, from University of South Carolina, for their help and support. We had a great time together. Last but not least, I am indebted to my parents, Yongqiang Zhou and Xin Huang, who have always gave me their trust and love, and to my wife, Jingyan Zhang, who has given me the strength and support needed during my academic life and being always there on my side.
ix Chapter 1
Overview
1.1 Motivation to the peridynamic theory
The peridynamic(PD) theory of mechanics attempts to unite the mathematical modeling of continuous media, fractures and particles within a single framework, see [1, 2, 3, 4]. It achieves this goal by replacing the partial differential equations of the classical elasticity theory with integral type equations. The model equations are based on the model of internal forces within a body in which material points interact with each other directly over finite distance. The classical elasticity theory is based on the assumption of a continuous distri- bution of mass within a body. And it assumes that the internal forces are contact forces that act across zero distance. The mathematical description of classical elas- ticity that follows from these assumptions relies on PDEs that additionally assume sufficient smoothness of the deformation for the well-posedness of the equation. The classical theory has been demonstrated to provide a good approximation to the response of real materials down to small length scales. Moreover, the PDEs of the classical elasticity do not apply directly on a crack or dislocation because of deformation is discontinuous on these features. Consequently, the techniques of fracture mechanics introduce relations that are extraneous to the basic field equations of the classical theory. For example, linear elastic fracture mechanics considers a crack to evolve according to a separate constitutive model that pre- dicts, on the basis of nearby conditions, how fast a crack grows, in what direction, whether it should arrest, branch, and so on. Although the methods of fracture 2 mechanics provide important and reliable tools in my applications, it is uncertain to what extent this approach can meet the future needs of fracture modeling in complex media under general conditions. Aside from requiring these supplemental constitutive equations for the growth of defects within linear elastic fracture mechanics, the classical theory predicts some well-known nonphysical features in the vicinity of these singularities. The unbounded stresses and energy densities predicted by the classical PDEs are con- ventionally treated in idealized cases by assuming that their effect is confined to a small process zone near the crack tip. However, the reasoning behind neglecting the singularities in this way becomes more troublesome as conditions and geometries become more complex. Molecular dynamics(MD) provides an approach to understanding the mechan- ics of materials at the smallest length scales that has met with important successes in recent years. However, even with the fastest computers, it is widely recognized that MD can not model systems of sufficient size to make it a viable replacement for continuum modeling. These considerations motivate the development of the PD theory which at- tempts to treat the evolution of discontinuities according to the same field equa- tions as for continuous deformation. The PD theory also has the goal of treating discrete particles according to the same field equations as for continuous media. The ability to treat both the nanoscale and the macroscale within the same mathe- matical system may make the method an attractive framework in which to develop multiscale and atomistic-to-continuum methods.
1.2 The current linear PD models
The peridynamic model [1, 2, 3] is a reformulation of solid mechanics in which a point in a continuum interacts directly with other points separated from it by a finite distance. The maximum interaction distance provides a length scale for a material model, called material horizon δ, although the model may additionally contain smaller length scales. The central assumption in the peridynamic model is that the strain energy density W (x) at a point x depends collectively on the deformation of all the points 3 in a neighborhood x with radius, δ > 0. Let x and y be the positions of two particles in the reference configuration, u(t, x) and u(t, y) denote the displacement of the two particles with respect to the reference configuration. And we denote f u(t, y) u(t, x), y x the interaction force density between x and y. Then the − − peridynamic equation of motion is
Z ρ(x)u¨(t, x) = f(u(t, y) u(t, x), y x) dy + b(t, x) (1.1) Ω − − where ρ is the mass density, b the external force density. The material horizon δ is built in the interaction force density f. In this dissertation, we mainly consider the linear peridynamic equation,
Z ρ(x)u¨(t, x) = C(x, y) u(t, y) u(t, x) dy + b(t, x) (1.2) Ω − where C(x, y) is a matrix-valued function obtained by taking the derivative of f u(t, y) u(t, x), y x with respect to the relative displacement u(t, y) u(t, x). − − − And the strain energy density is Z T W (t, u)(x) = u(t, y) u(t, x) C(x, y) u(t, y) u(t, x) dy. (1.3) Ω − −
The function C(x, y) is a 2nd order tensor, called micromodulus function which contains all the information including the material horizon δ that uniquely deter- mines the material properties. Depending on the form of C(x, y), the peridynamic model equations are divided into two groups, the ordinary bond-based PD model, the ordinary state-based PD model.
1.2.1 The ordinary bond-based PD model
The ordinary bond-based model interaction between two particles x and y only depends on the two particles themselves, independent of any others surrounding them. And the direction of the force is parallel to the bond x y. Then the − micromodulus can be expressed as
ω(x y) C(x, y) = − (y x) (y x) (1.4) y x 2 − ⊗ − | − | 4
where the function ω(x y) is called influence function defined in [2, Definition − 3.2], which measures the strength of the direct interaction between two particles, is normally defined as ( ω(x y) x y < δ ω(x y) = − | − | (1.5) − 0 otherwise. i.e. two particles lose interaction if their distance is larger than δ. Thus the bond-based PD strain energy density function is given by
1 Z ω(x y) 2 W (u) = − (u(t, y) u(t, x)) (y x) dy (1.6) b 2 y x 2 − · − Ω | − | 1.2.2 The ordinary state-based PD model
The state-based PD model assume that two particles can interact through an intermediate particle, see figure 1.2.2.
Figure 1.1. Point q interacts indirectly with x even though they are outside each other’s horizon, since they are both within the horizon of intermediate points such as p 5
In the figure, the point q interacts indirectly with x even though they are outside each other’s neighborhood with horizon δ since they are both within the horizon of the intermediate point such as p. It’s still a model for the ordinary material, so the direction of the force is also parallel to the bond x y. Based on − the above discussion the micromodulus function is given by
ω(x y) C(x, y) = − (y x) (y x) + C (x, y) (1.7) y x 2 − ⊗ − s | − | where Cs(x, y) describes the indirect interaction force between x and y whose ex- act expression will be derived in chapter 4.
To introduce the energy density function, we first review the peridynamic con- stitutive relation defined in [2, section 13]. Let y = u(t, y) + y u(t, x) x and | − − | x = y x for any given pair x and y. The fundamental deformation quantities | − | used in the peridynamic modeling can be expressed as follows:
extension scalar state e = y x, (1.8a) − weighted volume m = (ωx) x, (1.8b) • d dilatation θb = ωx e, (1.8c) m • θxb isotropic part of e ei = , (1.8d) d deviatoric part of e ed = e ei, (1.8e) − where d denotes the space dimension. The operation between different states • may be viewed as an inner product in a Hilbert space whose precise form is not required in the current dissertation; interested readers can refer to [2, Definition 2.6] for detailed explanations. Then the energy density function is defined as the following.
k(x)θb2 η(x) W (u) = + (ωed) ed. (1.9) s 2 2 •
The energy density is composed by the dilatation part and the shear deformation 6 part, i.e. deviatoric part in this language.
1.3 Content of the dissertation work
In this dissertation, we focus on various mathematical aspects of the PD models and nonlocal modeling.
In chapter 2, we study a linear PD model for a spring network systems in Rd, one dimensional PD bar with a periodic boundary condition, as some illustrative examples. The detailed analysis on the related PD operators and the associated functional spaces are conducted. And we prove the well-posedness of weak solu- tions to the PD equation, together with studies on the solution regularity. We point out, in particular, that for some special cases of singular micromodulus func- tions, the solution operators still share certain smoothing properties in fractional Sobolev spaces. These mathematical results can become useful in analyzing the output of numerical simulations based on the PD models and in assessing the qual- ity of the numerical solutions. Some of the materials can be found in [25] and [28], coauthored with my dissertation advisor Professor Qiang Du.
In chapter 3, we introduce a vector calculus for nonlocal operators that mimics the classical vector calculus for differential operators. We define nonlocal analogs of the divergence, gradient, and curl operators and deduce the corresponding nonlocal adjoint operators. Nonlocal analogs of the Gauss theorem and the Green’s identi- ties of the vector calculus for differential operators are also derived. We establish relationships between the nonlocal operators and their differential counterparts. The nonlocal vector calculus is also used to define nonlocal volume-constrained problems that are analogous to boundary-value problems for partial differential operators and to express the nonlocal balance laws1 in the sense that subregions not in direct contact may have a non-vanishing interaction. Our nonlocal vector calculus, then, provides an alternative to standard approaches for circumventing
1A balance law postulates that the rate of change of an extensive quantity over any subregion of a body is given by the rate at which that quantity is produced in the subregion minus the flux out of the subregion through its boundary. 7 the technicalities associated with the lack of sufficient regularity in local balance laws. Some of the materials can be found in [44], and coauthored with Professor Qiang Du from Penn State University, Professor Max Gunzburger from Florida State University and Dr. Rich. B. Lehoucq from Sandia National Laboratories.
In chapter 4, our major contribution here is the well-posedeness of the state- based linear peridynamic equilibrium equation, a first of a kind result. We rep- resent the peridynamic deformation fields such as the extension scalar, dilatation, etc., in terms of operators from the nonlocal vector calculus. Based upon the rep- resentation of the deformation state, we then formulate variational principles and rewrite both the linear bond-based and the linear state-based peridynamic elastic model via the nonlocal divergence operators and their corresponding adjoints. Our variational formulation coincides with the linearization of the original peridynamic equation derived in [1, 2]. Moreover, we prove the well-posedness of peridynamic “boundary”-value problem, or more appropriately, volume-constrained problem, that is, the linear peridynamic elastic model subject to suitable constraints that are either explicitly enforced on volumes of non-zero measure or naturally implied by the variational principle. Some of the materials can be found in [50], and coauthored with Professor Qiang Du from Penn State University, Professor Max Gunzburger from Florida State University and Dr. Rich. B. Lehoucq from Sandia National Laboratories.
In chapter 5, we demonstrate that the anisotropic PD operator including the ordinary bond-based and state-based operators, in the local limit, tends to the classical elasticity operator CE(u) = (C:( u + uT ). The clear relation − −∇ · ∇ ∇ between the nonlocal parameters, such as the influence function ω(x, y), and the 4th order stiffness tensor in the classical elasticity is established for the three kinds of models. We also showed that how the parameters relate to the coupling of elon- gations and shears in the limiting sense. As special cases, it is illustrated that the isotropic bond-based model converges the classical Navier equation with Poisson 1/4, and the isotropic ordinary state-based model, as a general model, converges to the Navier equation with arbitrary Possion ratio. Some of the materials can be found in [25, 28, 50], Professor Qiang Du from Penn State University, Professor 8
Max Gunzburger from Florida State University and Dr. Rich. B. Lehoucq from Sandia National Laboratories.
In chapter 6, based on the analytical framework established in chapter 2, we consider various finite dimensional approximations to the nonlocal BVPs. A num- ber of theorems are given on the convergence and error estimates of these approx- imations which appear to be the first of their kind in the literature. Estimates of the condition numbers of the resulting linear systems of equations are also provided here. Our findings are established in very general settings, and they are consistent with the results derived or observed for various specialized cases in earlier studies [5, 6, 7, 8]. In order to have a better understanding of the PD models, their nu- merical approximations, and their various applications, detailed analytical studies are becoming increasingly important. The results given in this chapter can be considered as another step forward in this direction. Some of the materials can be found in [28] and in [19]. The work is coauthored with Professor Qiang Du from Penn State University, Professor Lili Ju from University of South Carolina and Dr. Li Tian from Penn State University.
In chapter 7, we develop the PD Double-bonds model in order to recover the full elasticity in its local limit. In chapter 5, it is shown that the current PD models can not recover the 21 independent coefficients in the stiffness matrix in the classi- cal material. The PD double-bonds model effectively incorporate the interactions among elongations and shear deformations in different directions in a single-integral operator. We use Fourier analysis to show that the PD Double-bonds model con- verges to the classical elasticity model with 21 independent coefficients. And it can be reduced to the PD bond-based model by specifying the influence function ω.
In chapter 8, we list some open issues in the crack nucleation theory. We first review the analytical and numerical result in [63] for the crack nucleation condition. We conduct an similar one dimensional experiment in order to investigate more precise relation between the different physical quantities in the PD model. Some preliminary results are provided. It illustrates some different scenarios from the 9 discussion in [63] which is listed to be an open question at this moment. Chapter 2
Mathematical Analysis of Bond-based Linear PD models
2.1 Introduction
The effectiveness of PD model has already been demonstrated in several sophis- ticated applications, including the fracture and failure of composites, crack in- stability, fracture of polycrystals, and nanofiber networks. Yet, from a rigorous mathematical point of view, many important and fundamental issues remain to be studied. In this chapter, we intend to formulate a rigorous functional analyt- ical framework of the PD models so as to provide a better understanding of the PD model and to guide us in the development and analysis of the numerical al- gorithms. This in turn will help us utilize the PD theory for multiscale materials modeling. Indeed, PD can be effectively used in the multiscale modeling of mate- rials in different ways: it can serve as a bridge between molecular dynamics (MD) and continuum elasticity (CE) to help mitigate the difficulties encountered when one attempts to couple MD and CE directly [9, 10, 11, 12, 13] and, in some situa- tions, PD can be used as a stand-alone model to capture the behavior of materials over a wide range of spatial and temporal scales. For example, to study problems involving defects, one can use the same equations of motion over the entire body and no special treatment is needed near or at defects. Such properties make PD a powerful tool for modeling problems involving cracks, interfaces or defects, we 11 refer to [14] for a review of the recent applications of the PD framework. Parallel to the modeling and application to practical problems, there also have been efforts to establish a sound theoretical foundation for the PD model. For instance, an abstract variational formulation is presented in [15]. Some results on the existence and uniqueness of L2 solutions of the PD models associated with bounded integral PD operators have been given in [16]. Though much of the focus of [17] is on developing homogenization theory for the PD model, some existence and uniqueness results are also provided, again for bounded integral PD operators. In [18], a nonlocal vector calculus was developed which also provided a rigorous framework for studying the boundary value problems of the nonlocal peridynamic models. In this chapter, we study a linear peridynamic model for a spring network sys- tems in Rd and boundary value problem on a one-dimensional bar. The bond-based model in Rd is discussed in section 2.2 The detailed analysis on the related PD operators and the associated functional spaces are given in section 2.2.1 and 2.2.2. We consider the stationary and time-dependent PD models in section 2.2.3 and Section 2.2.4 respectively. Following the same line, we discuss the bond-based PD model in section 2.3. The properties of the models and their solutions depend crucially on the particular influence functions used to specify the spring network systems. Our results are valid for more general influence functions than consid- ered in the earlier literature. Indeed, the only essential assumption on the influence functions is that appropriate elastic modulus can be defined for the material model. For these more general cases, we prove the well-posedness of weak solutions to the peridynamic equation, together with studies on the solution regularity. We point out, in particular, that for some special cases of singular influence functions, the solution operators still share certain smoothing properties in fractional Sobolev spaces. These mathematical results can become useful in analyzing the output of numerical simulations based on the PD models and in assessing the quality of the numerical solutions. 12
2.2 Bond-based PD model on Rd
In this section, we analyze the bond-based PD model (1.6) on the whole space Rd. We set the micromodulus function
ω( x y ) C(x, y) = | − | (y x) (y x) y x 2 − ⊗ − | − | i.e. the influence function is set to be a radial, ω(x y) = ω( y x ), so the model − | − | under consideration is isotropic and homogenous.
Then, the bond-based PD equation of motion is give by d utt(t, x) = Lδu(t, x) + b(t, x) , (t, x) (0,T ) R ∀ ∈ × u(0, x) = g(x) , x Rd (2.1) ∀ ∈ d ut(0, x) = h(x) , x R ∀ ∈ where
Z ω( y x ) L u(x) = | − | (y x) (y x) u(y) u(x) dy. (2.2) δ y x 2 − ⊗ − − Bδ(x) | − | 2.2.1 Mathematics analysis of the PD model
To set up a suitable functional setting to discuss the well posedness and properties of peridynamic model equations, we first make some assumptions on the influence function ω: Z 2 ω( x ) > ρ( x ) 0, x Bδ(0) , and τδ := x ω( x ) dx < , (2.3) | | | | ≥ ∀ ∈ Bδ(0) | | | | ∞ where ρ( x ) is non-negative function belonging to L1 B (0) which is strictly pos- | | δ itive in at least a measure nonzero subset of B (0). The condition that ω( x ) > δ | | ρ( x ) is easily satisfied by all practical choices of the influence functions. | |
We note that, as pointed out in the literature (see for instance, equation (2.10) in [16], and equation (93) in [1]), the assumption on τδ being finite is needed in or- der to have a suitable definition of the elastic moduli for the corresponding material 13 under consideration. This assumption, in fact, allows us to study the bond-based PD models with much more general influence functions, and thus more general micromodulus functions, than those considered in the existing mathematical anal- ysis. As a notational convention, we use uˆ = uˆ(ξ) to denote the Fourier transform of u = u(x). Moreover, u¯ denotes the complex conjugate of u, and uT denotes the transpose of u. By performing the Fourier transform, we can introduce an equivalent definition of our multidimensional peridynamic operator,
Z 1 ix ξ ξ ξ · ξ Lδu(x) = d/2 Mδ( )uˆ( )e d (2.4) − (2π) Rd where M (ξ), a real-valued and symmetric positive semi-definite d d matrix, is δ × the Fourier symbol of the pseudo-differential operator, see [20], L : − δ Z ω( y x ) M (ξ) = | − | 1 cos(ξ y) y y dy (2.5) δ y x 2 − · ⊗ Bδ(0) | − | for any ξ Rd and for any PD horizon parameter δ > 0. Moreover referring to ∈ [16], the PD operator vanishes identically when applied to rigid body motions that read as u(x) = Rx + r for some skew-symmetric R and some vector r. Since here PD operator is defined on L2(Rd), its kernel contains only u(x) = 0.
By the equivalent form of peridynamic operator, we can define the following functional space, equipped with an associated norm:
d Definition 1. The space ω(R ), which depends on the influence function ω, M 2 d d consists of all the functions u L (R ) for which the ω(R ) norm ∈ M
1 Z 2
u ω = uˆ(ξ) (I + Mδ(ξ))uˆ(ξ)dξ , (2.6) M k k Rd · is finite. We also define the corresponding inner product associated with the Mω norm: Z
(u, v) ω = vˆ(ξ) (I + Mδ(ξ))uˆ(ξ)dξ , (2.7) M Rd · d 1 d for any u, v ω(R ). In addition, we use ω− (R ) to denote the dual space of ∈ M M d ω(R ). M 14
Remark: The norm is well-defined since I + Mδ(ξ) is real-valued symmetric pos- itive definite matrix and it is uniformly bounded below by I.
Meanwhile, we can have the following properties:
d Lemma 1. The space ω(R ) is a Hilbert space corresponding to the inner product M ( , ) ω . · · M d Proof: Let un be a Cauchy sequence in ω(R ). By definition, it is equivalent { } M to say n 1 o (I + Mδ) 2 uˆn is a Cauchy sequence in L2(Rd). So by the completeness of L2(Rd), there exists an element v L2(Rd), such that ∈
1 (I + M (ξ)) 2 uˆ (ξ) v(ξ) 2 0 k δ n − kL → as n . Then we set → ∞
1 1 u(x) = F − [(I + Mδ(ξ))− 2 v(ξ)],
1 where F − denotes the inverse Fourier transform. Then one can see that
1 2 2 un u ω = (I + Mδ(ξ)) (uˆn(ξ) uˆ(ξ)) L 0. k − kM k − k →
d So the space ω(R ) is complete, and it is thus a Hilbert space. M
d Lemma 2. The dual space of ω(R ) is the space of distributions: M Z 1 d 1 ω− (R ) = u : uˆ(ξ) (I + Mδ(ξ))− uˆ(ξ)dξ < , M Rd · ∞ equipped with the norm
1 Z 2 1 1 ξ ξ − ξ ξ u ω− = uˆ( ) (I + Mδ( )) uˆ( )d . k kM Rd · 15
d Proof: Let l = l(u) be a bounded linear functional on ω(R ), then by the M d Riesz Representation Theorem we know that there exists a unique w ω(R ) ∈ M d d such that l(u) = (u, w) ω(R ) for any u ω(R ). Using the inner product given M ∈ M in (2.7), we have Z l(u) = wˆ (ξ) (I + Mδ(ξ))uˆ(ξ)dξ . Rd · 1 2 d Let vˆ(ξ) = (I + Mδ(ξ))wˆ (ξ). We have (I + Mδ(ξ)) 2 wˆ (ξ) L (R ) since w ∈ ∈ d ω(R ). Thus, M
1 1 2 d (I + Mδ(ξ))− 2 vˆ(ξ) = (I + Mδ(ξ)) 2 wˆ (ξ) L (R ) . ∈
1 d So, v ω− (R ) and ∈ M Z l(u) = uˆ(ξ) vˆ(ξ)dξ . Rd · Again by the Riesz Representation Theorem, we have
2 2 l = w ω k k Zk kM = wˆ (ξ) (I + Mδ(ξ))wˆ (ξ)dξ d · ZR 1 = vˆ(ξ) (I + Mδ(ξ))− vˆ(ξ)dξ Rd · 1 = v ω− k kM
1 d d Meanwhile, if v ω− (R ), for any u ω(R ), ∈ M ∈ M Z (u, v)L2 = uˆ(ξ) vˆ(ξ) dξ | | | Rd { · } | Z n 1 1 o = uˆ(ξ) (I + Mδ(ξ)) 2 (I + Mδ(ξ))− 2 vˆ(ξ) dξ | Rd · d 1 6 u ω(R ) v −ω . k kM k kM
1 d d So, an element v ω− (R ) corresponds a bounded linear functional on ω(R ). ∈ M M
d Lemma 3. The peridynamic operator Lδ is self-adjoint on ω(R ). The oper- − M 16
d 1 d ator Lδ + I is also an isometry from ω(R ) to ω− (R ), and the norm and − M M d inner product in ω(R ) can also be formulated as M
1 1 2 2 u ω = (u, u) ω = [(u, u) + ( Lδu, u)] k kM M − 1 Z Z 2 2 ω( y x ) 2 = u L2 + | − 2| (u(y) u(x)) (y x) dydx k k d y x − · − R Bδ(x) | − | (2.8)
d for any u ω(R ). ∈ M Remark 1. The result of this Lemma is analogous to the classical result corre- sponding to the differential operator , that is, + I is self-adjoint and it is −4 −4 1 d 1 d an isometry from H (R ) to H− (R ).
Proof: By the equivalent definition of the PD operator in (2.4) as a pseudo- differential operator with a real, nonnegative symbol, we immediately see that d Lδ is self-adjoint in ω(R ). The fact that Lδ + I defines an isometry follows − M − d 1 d directly from the definitions of the norms in ω(R ) and ω− (R ). In addition, M M using the Parseval formula, we have Z ( Lδu, u) + (u, u) = ( Ldδu, uˆ) + (uˆ, uˆ) = uˆ(ξ) (I + Mδ(ξ)) uˆ(ξ)dξ , − − Rd · which implies the equation (2.8).
For the purpose of discussing the regularity of the weak solutions, we also need to define the following space Z 2 d 2 ω(R ) = u : uˆ(ξ) (I + Mδ(ξ)) uˆ(ξ)dξ < , M Rd · ∞ with the dual space Z 2 d 2 ω− (R ) = u : uˆ(ξ) (I + Mδ(ξ))− uˆ(ξ)dξ < M Rd · ∞ which share the similar properties as the ones in Lemma 1, Lemma 2 and Lemma 3. 17
Given any y B (0), let us define a difference operator D by ∈ δ y
D v(x) = v(x + y) v(x) y − for function v defined in a suitable function space. By the representation of Lδ given in (7.8), we have
Z Z ω( y ) 1 Dy + D y L = | | y yD dy = ω( y )(y y) − dy (2.9) δ y 2 ⊗ y 2 | | ⊗ y 2 Bδ(0) | | Bδ(0) | | which gives an interesting formulation of Lδ as a linear combination difference operators (first or second order). The matrix-valued weight ω( y )(y y) is in fact | | ⊗ in L1(Rd) under the assumption (2.3). Then,
Lemma 4. Let P be an scalar operator which commutes with difference operator D for all y B (0), then P also commutes with L . y ∈ δ δ
Proof: For u = u(x) with Lδu suitably defined, we have
Z ω( y ) Z ω( y ) | | y yP D u(x)dy = | | y yD P u(x)dy , y 2 ⊗ y y 2 ⊗ y Bδ(0) | | Bδ(0) | |
d That is, P (Lδu)(x) = Lδ(P u)(x), for any x R , so the lemma follows. ∈
We note that in particular, all scalar linear differential operators with constant coefficients, and their inverses (when they exist), commute with the difference operator Dy for any y, so we have,
Corollary 1. Let P be a scalar linear differential operator with constant coeffi- cients, then PLδ = LδP .
Let us remark that the results of Lemma 4 and Corollary 1 depend crucially on the facts that in the representation of the PD operator given by (7.8), the horizon parameter δ is a uniform constant in space and that the influence function ω = ω( y x ) is only a function of y x . | − | | − | The above corollary in particular allows us to have the well-posedness of the generalized (in distribution sense) solutions to the peridynamic equation with rough data. For instance, even when the external force b = b(t, x) is not in 18
L2, we may still get a unique solution to the equation (2.29) in the appropriate 1 1 weak sense by lifting b to P − b for some P , if P − b can be properly defined, to 1 1 get a generalized solution of (2.29) in the form P ( L + I)− P − b. − δ
d Now we can discuss the equivalence between the defined space ω(R ) and M standard Sobolev spaces. Similar results for the special case of scalar valued func- tions can be found in [21]. To treat the vector valued case, we adopt the convention that for two symmetric matrices A and B, A < B means B A is positive definite, − and A B means B A is positive semi-definite. Let , denote the continuous ≤ − → embedding of function spaces.
The following general embedding results can be shown:
Lemma 5. Let the influence function ω = ω( y ) satisfy the condition (2.3). Then | |
τδ 2 d 0 < Mδ(ξ) ξ I , ξ R (2.10) 6 2 | | ∀ ∈
Consequently, we have
1 d d 2 d 2 d H (R ) , ω(R ) ,H (R ) , ω(R ). → M → M
Proof: Using the inequality
(ξ y)2 0 1 cos(ξ y) · ξ 2 y 2/2, 6 − · 6 2 6 | | | | we have
Z 1 cos(ξ y) M (ξ) = ω( y ) − · y y dy δ | | y 2 ⊗ Bδ(0) | | Z 1 cos(ξ y) ω( y ) − · y 2dy I 6 | | y 2 | | Bδ(0) | | Z 1 2 2 6 ω( y ) ξ y dy I 2 Bδ(0) | | | | | | τ = δ ξ 2I 2 | | 19 which gives (2.10). The rest of the conclusions follow from the above inequality and the definitions of the relevant function spaces.
2.2.2 Space equivalence for special influence functions
We now focus on some influence functions with special properties to establish d relations between ω(R ) and the more conventional Sobolev spaces. To examine M more singular influence functions, we let
s d 2 d d s 2 d d H (R ) = u (L (R )) : ξ uˆ (L (R )) ∈ | | ∈ denote the fractional Sobolev space on Rd for s (0, 1). ∈
For convenience, we define the following functions for χ > 0 and α 0: ≥ Z 1 cos(z1) C (α, χ) = − min(z2)dz , (2.11) min z 2+d+2α Bχ(0) | | Z 1 cos(z1) C (α, χ) = − max(z2)dz . (2.12) max z 2+d+2α Bχ(0) | | Here, we have min(z2) = min z2 with z d being the components of z, and { i } { i}i=1 similarly, max(z2) = max z2 . { i } For influence functions ω such that ω( x ) L1(B (0)), there are already some | | ∈ δ 2 d results in [17, 16] showing that Lδ is a bounded linear operator from L (R ) to − L2(Rd). In fact, we have:
Lemma 6. Let ω = ω( y ) satisfy the additional condition that | | Z ω( y ) dy < , (2.13) Bδ(0) | | ∞ we then have 2 d d 2 d ω(R ) = ω(R ) = L (R ) , M M 20 and Z 1 2 d 2 d 2 2 u L 6 u ω(R ) 6 u L (1 + ω( y ) dy) , u L (R ) , (2.14) M k k k k k k Bδ(0) | | ∀ ∈ Z 2 d 2 2 d 2 u L 6 u ω(R ) 6 u L (1 + ω( y ) dy) , u L (R ) . (2.15) M k k k k k k Bδ(0) | | ∀ ∈ 1 Moreover, the operators L and ( L + I)− are bounded linear operators from − δ − δ L2(Rd) to L2(Rd).
Proof: Under the condition of ω, we can see
Z 0 < Mδ(ξ) 6 ω( y ) dy I . Bδ(0) | |
By Parseval identity, we have (2.14) and (2.15) which in turn implies that L − δ 1 2 d 2 2 1 d and ( Lδ + I)− are bounded operators from L (R ) to L (R ), and ω− (R ) = − M d 2 d ω(R ) = L (R ). M
For more general influence functions, i.e. ω satisfying (2.3), the PD operator 2 d Lδ may become unbounded in L (R ) when the function ω( x ) is no longer − | | 1 in L (Bδ(0)). Yet, as we demonstrate below, the basic existence and uniqueness results remain valid but with the discussion taking place in other function spaces d such as ω(R ), as defined earlier. This is due to the fact that Lδ becomes a M − d 1 d bounded operator from ω(R ) to ω− (R ). To see how such spaces are related M M to the conventional Sobolev spaces, we first consider the space equivalence for some special influence functions.
Lemma 7. Let the influence function ω = ω( y ) satisfy the assumption (2.3) and | | the condition
d 2β ω( y ) γ1 y − − , y δ (2.16) | | > | | ∀| | 6 for some exponent β (0, 1) and positive constant γ1, then we have ∈
d β d 2 d 2β d ω(R ) , H (R ) , ω(R ) , H (R ). M → M → 21
β d 1 d 2β d 2 d H− (R ) , ω− (R ) ,H− (R ) , ω− (R ). → M → M Moreover, we have
d β d d C1 u H (R ) 6 u ω(R ) , u ω(R ) , (2.17) k k k kM ∀ ∈ M
β d C u 1 d u β d , u H− ( ) , (2.18) 1 ω− (R ) 6 H− (R ) R k kM k k ∀ ∈ and 2 d ˜ 2β d 2 d C1 u H (R ) 6 u ω(R ) , u ω(R ) , (2.19) k k k kM ∀ ∈ M 2β d C˜ u 2 d u 2β d , u H− ( ) , (2.20) 1 ω− (R ) 6 H− (R ) R k kM k k ∀ ∈ 1 with constants C1 = min(1, (γ1Cmin(β, δ)) 2 ) and C˜1 = min(1, γ1Cmin(β, δ)).
Proof: Under the condition on σ, when ξ 1, we do a change of variable | | >
z = ξ Ry | | where R is an orthogonal matrix from Rd to Rd with the first row being ξ/ ξ . | | So we have
Z 1 cos(ξ y) M (ξ) = ω( y ) − · y y dy δ | | y 2 ⊗ Bδ(0) | | Z 1 cos(ξ y) γ1 − · y y dy > y 2+d+2β ⊗ Bδ(0) | | Z 2+2β 1 cos(z1) 1 T 1 T = γ1 ξ − R z R zdz 2+d+2β ξ ξ | | B ξ δ(0) z ⊗ | | | | | | | | Z 2β 1 cos(z1) T ξ 0 = γ1 −2+d+2β R z z Rdz | | B ξ δ(0) z ⊗ | | | | Z 1 cos(z ) ξ 2β 1 T 2 = γ1 −2+d+2β R diag(zi )Rdz | | B ξ δ(0) z | | | | Z 2β 1 cos(z1) 2 γ1 ξ − min(z )dz I > | | z 2+d+2β Bδ(0) | | 2β = γ1C (β, δ) ξ I . min | | 22 where the equality in the 5th line holds due to the symmetry of the integration domain. Thus the space embedding results follow. Moreover (2.17) and (2.19) are satisfied. The inequalities (2.18) and (2.20) follow by duality estimates. Meanwhile, we also have
Lemma 8. Let the influence function ω = ω( y ) satisfy the condition | |
d 2α ω( y ) γ2 y − − , y δ (2.21) | | 6 | | ∀| | 6 for some exponent α (0, 1) and positive constant γ2, then we have ∈
α d d 2α d 2 d H (R ) , ω(R ) ,H (R ) , ω(R ) , → M → M and 1 d α d 2 d 2α d ω− (R ) , H− (R ) , ω− (R ) , H− (R ). M → M → Moreover, we have
α d d α d u ω(R ) 6 C2 u H (R ), u H (R ) , (2.22) k kM k k ∀ ∈
1 d u α d C u 1 d , u − ( ) , (2.23) H− (R ) 6 2 −ω (R ) ω R k k k kM ∀ ∈ M and 2α d 2 d ˜ 2α d u ω(R ) 6 C2 u H (R ), u H (R ) , (2.24) k kM k k ∀ ∈ 2 d u 2α d C˜ u 2 d , u − ( ) , (2.25) H− (R ) 6 2 −ω (R ) ω R k k k kM ∀ ∈ M 1 with constants C2 = max(1, (γ2C (α, )) 2 ) and C˜2 = max(1, γ2C (α, )). max ∞ max ∞ Proof: Similar to the proof of the previous lemma, under the condition on σ and for ξ > 0, we make a change of variable z = ξ Ry where R is an orthogonal | | | | matrix from Rd to Rd with the first row ξ/ ξ . Then we have | | Z 1 cos(ξ y) M (ξ) = ω( y ) − · y y dy δ | | y 2 ⊗ Bδ(0) | | Z 1 cos(ξ y) γ2 − · y y dy 6 y 2+d+2α ⊗ Bδ(0) | | 23
Z 2+2α 1 cos(z1) 1 T 1 T = γ2 ξ − R z R zdz 2+d+2α ξ ξ | | B ξ δ(0) z ⊗ | | | | | | | | Z 2α 1 cos(z1) T ξ 0 = γ2 −2+d+2α R z z Rdz | | B ξ δ(0) z ⊗ | | | | Z 1 cos(z ) ξ 2α 1 T 2 = γ2 −2+d+2α R diag(zi )Rdz | | B ξ δ(0) z | | | | Z 1 cos(z ) ξ 2α 1 2 6 γ2 −2+d+2α max(z )dz I | | d z R | | 2α = γ2C (α, ) ξ I . max ∞ | | where the equality in the 5th line holds due to the symmetry of the integration domain. Thus the space embedding results follow from the respective definitions of the function spaces. Moreover, (2.22) and (2.24) are satisfied. The inequalities (2.23) and (2.25) then follow from duality estimates. Consequently, we see that under certain conditions on the influence function, d the space ω(R ) is equivalent to some standard fractional Sobolev spaces. M Theorem 1. Let the influence function ω = ω( y ) satisfy the condition | |
d 2α d 2α γ1 y − − ω( y ) γ2 y − − , y δ (2.26) | | 6 | | 6 | | ∀| | 6 for some exponent α (0, 1) and positive constants γ1 and γ2, then we have ∈
d α d 2 d 2α d ω(R ) = H (R ) , ω(R ) = H (R ). M M
d Moreover, for any u ω(R ), ∈ M
α d d α d C1 u H (R ) 6 u ω(R ) 6 C2 u H (R ) , (2.27) k k k kM k k
2 d and for any u ω(R ), ∈ M
˜ 2α d 2 d ˜ 2α d C1 u H (R ) 6 u ω(R ) 6 C2 u H (R ) , (2.28) k k k kM k k with the positive constants C1, C2, C˜1 and C˜2 defined in Lemmas 7 and 8. 24
We see from the above discussions that, under additional assumptions on the form of the influence function ω, we have the equivalence or continuous embedding d theories between ω(R ) and certain fractional Sobolev spaces. M
2.2.3 Properties of stationary PD Model
In this section, we give some results on the existence and uniqueness of weak so- lutions to the stationary (equilibrium) PD model with general influence functions, 2 d i.e. Lδ may be unbounded in L (R ):
L u + u = b (2.29) − δ and some convergence properties of the solution of the stationary PD model. The term u is added for two purposes, one is to eliminate the need to imposing far field conditions at infinity and the other is to eliminate the nonuniqueness of solution when no boundary condition is imposed.
First, we may also establish the corresponding variational theory and some regularity properties for the stationary PD model. Then, using the properties of the PD operator provided earlier, we have
1 d Lemma 9. Let ω = ω( y ) satisfy the condition (2.3), for any b ω− (R ), the | | ∈ M d problem (2.29) has a unique solution u ω(R ) which is the minimizer of the ∈ M functional:
1 2 E(u) = u d (u, b) 2 d ω(R ) L (R ) 2k kM − 1 Z = uˆ(ξ) (I + Mδ(ξ))uˆ(ξ)dξ (u, b)L2(Rd) (2.30) 2 Rd · −
d in ω(R ). M Proof: The conclusion follows directly from the fact that E = E(u) is a convex 1 d quadratical functional with Lδu + u b ω− (R ) being its variational deriva- − − ∈ M d tive at u ω(R ). ∈ M 25
We note that more general variational descriptions of the PD models can be found in [15, 18]. As for regularity, we have for some special influence functions that:
Lemma 10. Let ω = ω( y ) satisfy the condition (2.26), the problem (2.29) has a | | unique solution u Hm+2α(Rd), whenever b Hm(Rd) for any m 2α. ∈ ∈ > − Proof: Taking the Fourier transform of the equation (2.29), we get
(Mδ(ξ) + I)uˆ(ξ) = bˆ(ξ) . (2.31)
Then we have
2 m ˆ 2 2 m 1 d ((Mδ(ξ) + I)uˆ(ξ)) ((Mδ(ξ) + I)) uˆ(ξ)( ξ + 1) = b(ξ) ( ξ + 1) L (R ) . · | | | | | | ∈
By Theorem 1, we have
2 2 + 2 1 ( ξ + 1) α m uˆ(ξ) L (Rd) . | | | | ∈
So the result follows.
We now go back to the general influence functions to consider the regularity of solution of the equilibrium equations (2.29).
Lemma 11. Let the influence function ω = ω( y ) satisfy (2.3), the problem (2.29) | | d 1 d 2 d has a unique solution u ω(R ) for b ω− (R ). Moreover, if b L (R ), ∈ M ∈ M ∈ the solution of the equilibrium equation (2.29) satisfies
2 d u ω(R ). ∈ M
Proof: The first part follows from the isometry property given in the Lemma 3. The proof of regularity is similar to that of the Lemma 10.
By Lemma 4 and Corollary 1, we also have the following regularity
Lemma 12. Let P be a linear scalar operator with constant coefficients, if u is a solution of the equation (2.29) with a given function b, then P u is the solutions 26 of (2.29) with the right hand side being P b.
Proof: From Lemma 4 and Corollary 1, we know L P = PL . So − δ − δ
L (P u) = P ( L u) = P b. − δ − δ
This leads to the result of the lemma.
2.2.4 The time-dependent PD model
With the suitable function spaces for the PD operator and the stationary peridy- namic model given earlier, we now proceed to discuss the existence and uniqueness of the solutions of the time-dependent PD model (2.1) in these spaces, again for more general influence functions ω = ω( y ). | | Using the Fourier transform, we first rewrite the PD equation (2.1) as
ˆ uˆtt(t, ξ) + Mδ(ξ)uˆ(t, ξ) = b(t, ξ) , uˆ(0, ξ) = gˆ(ξ) , (2.32) uˆt(0, ξ) = hˆ(ξ) .
By Duhamel’s principle, we formally have p p sin ( Mδ(ξ)t) ˆ uˆ(t, ξ) = cos ( Mδ(ξ)t)gˆ(ξ) + p h(ξ) Mδ(ξ) Z t 1 p ˆ + p sin ( Mδ(ξ)s)b(t s, ξ)ds . (2.33) Mδ(ξ) 0 −
Then by taking the inverse Fourier transform, we can get
Z d Z u(t, x) = G(t, y)g(x y) dy + G(t, y)h(x y) dy Rd dt − Rd − Z t Z + G(t, y)b(t s, x y) dyds , (2.34) 0 Rd − − p 1 sin ( Mδ(ξ)t) where G(t, y) = F − ( p ), see also [22]. Mδ(ξ) 27
From the equation (2.33), we can see
Theorem 2. If the influence function ω = ω( y ) satisfies (2.3), and | |
d 2 d 2 2 d g ω(R ) , h L (R ) , b L (0,T ; L (R )), (2.35) ∈ M ∈ ∈ for some T > 0, then the PD equation (2.1) has a unique solution u = u(t, x) given by (2.34). Moreover,
d 2 2 d u C([0,T ]; ω(R )) , ut L (0,T ; L (R )). (2.36) ∈ M ∈
Proof: From (2.33) and (2.34), we can see the solution u can be expressed by the given quantities b, g and h, i.e. the solution of the PD equation (2.1) uniquely exists, so it is suffice to give the proper space that the solution belongs to.
First, we note that since Mδ(ξ) is real symmetric and positive definite, it can T be diagonalized by an orthogonal matrix Q = Q(ξ), i.e. Mδ(ξ) = Q diag(λδ,i)Q.
And we denote the three terms on the right hand side of (2.33) as uˆ1(t, ξ), uˆ2(t, ξ) and uˆ3(t, ξ) respectively so that
uˆ(t, ξ) = uˆ1(t, ξ) + uˆ2(t, ξ) + uˆ3(t, ξ) .
By the condition (2.35), we readily have Z 2 p p u1(t, x) ω = gˆ(ξ) cos( Mδ(ξ)t)(I + Mδ(ξ)) cos( Mδ(ξ)t)gˆ(ξ) dξ k kM Rd · Z T p T p = gˆ(ξ) Q diag(cos( λδ,it)Q(I + Mδ(ξ))Q diag(cos( λδ,it)Qgˆ(ξ) dξ Rd · Z T p p = gˆ(ξ) Q diag(cos( λδ,it)diag(1 + λδ,i)diag(cos( λδ,it)Qgˆ(ξ) dξ Rd · Z T 2 p = gˆ(ξ) Q diag((1 + λδ,i) cos ( λδ,it))Qgˆ(ξ) dξ Rd · Z T 6 gˆ(ξ) Q diag(1 + λδ,i)Qgˆ(ξ) dξ Rd · Z 2 = gˆ(ξ) (I + Mδ(ξ))gˆ(ξ) dξ = g ω . Rd · k kM 28
d So we can see that u1 = u1(t, x) is uniformly bounded in C([0,T ]; ω(R )). M Similarly, we can also deduce that
2 2 2 2 3 2 u2(t, x) ω (1 + T ) h L2( d) , and u3(t, x) ω (T + T ) b L2(0,T ;L2( d)) k kM ≤ k k R k kM ≤ k k R uniformly in [0,T ]. Therefore, it implies that u = u(t, x) is bounded uniformly in d d ω(R ) for any t [0,T ], i.e. u C([0,T ]; ω(R )). M ∈ ∈ M
Differentiating (2.33) with respect to t, we get
p p p uˆ (t, ξ) = M (ξ) sin ( M (ξ)t)gˆ(ξ) + cos ( M (ξ)t)hˆ(ξ) t − δ δ δ Z t p + cos ( Mδ(ξ)(t s))bˆ(s, ξ)ds . (2.37) 0 −
Then through a similar calculation as in the above, we can get that ut is uniformly bounded in L2(0,T ; L2(Rd)). These a priori estimates, together with standard PDE theory [23, 24], lead d to the existence and uniqueness of the solution u of (2.1) in C([0,T ]; ω(R )) M ∩ 1 2 H (0,T ; L (Rd)). Note that for the linear time-dependent equation, we can easily get the following lemma.
Lemma 13. Let P be a time-independent linear operator which commutes with
Lδ, then it commutes with the solution operator of the system (2.1).
Similar to the stationary case, it again allows us to establish the well-posedness of even more generalized solutions to (2.1).
Theorem 3. Let the influence function ω = ω( y ) satisfy (2.3), and P be an | | time-independent linear operator which commutes with Lδ. Then for the initial conditions and the forcing term satisfying
d 2 d 2 2 d P g ω(R ) ,P h L (R ) ,P b L (0,T ; L (R )), ∈ M ∈ ∈ the PD equation (2.1) has a unique solution u = u(t, x) with
d 1 2 d P u C([0,T ]; ω(R )) H (0,T ; L (R )). ∈ M ∩ 29
1/2 In particular, we can take P = ( L + I)− , then we get the existence and − δ uniqueness of weak solution u = u(t, x) to (2.1) with
2 d 1 1 d u C([0,T ]; L (R )) H (0,T ; ω− (R )) ∈ ∩ M
2 d 1 d 2 1 d for g L (R ), h ω− (R ) and b L (0,T ; ω− (R )). ∈ ∈ M ∈ M The proof is straightforward by verifying that P u is also the solution of the PD equation with the transformed data. Note that the theorem also implies the well-posedness of the Cauchy problem for the time-dependent PD equation even when the initial displacement is given as a distribution.
2.3 The bond-based PD model on a finite bar
In this section, we investigate the properties of the bond-based PD model defined on a finite bar, represented by the interval (0, π),
Z x+δ utt(t, x) = ω( y x ) u(t, y) u(t, x) dy, for x (0, π) (2.38) x δ | − | − ∈ − with the solution u = u(x) satisfying either
u is odd in ( δ, δ) and (π δ, π + δ) , (2.39) − − or u is even in ( δ, δ) and (π δ, π + δ) . (2.40) − − The condition (2.39) resembles (and thus is called) a Dirichlet-like condition, while (2.40) resembles a Neumann-like condition. Indeed, for smooth enough func- tions, (2.39) implies u(0) = u(π) = 0, while (2.40) leads to ux(0) = ux(π) = 0. In practice, one often also studies other displacement loading or force loading conditions. We note that the odd and even extensions given in (2.39) and (2.40) allow us to more easily formulate the spectrum of the corresponding PD operator than, for example, the case of force loading condition. We thus focus on the former cases in this work and leave discussions on other general BVPs to future works. 30
2.3.1 The PD operators and related function spaces
Similar to the section 2.2 and [25], we begin with a definition of the PD operators on (0, π).
Definition 2. For ω satisfying (2.3), the PD operator o is defined by −Lδ Z x+δ o δu(x) = ω( y x )(u(y) u(x)) dy x (0, π) (2.41) −L x δ | − | − ∀ ∈ − if u is an integrable function satisfying the Dirichlet-like displacement loading condition given in (2.39). Similarly, e is defined for functions satisfying the −Lδ Neumann-like displacement loading condition (2.40). More precisely, with the Fourier sine and cosine series expansions,
X o X e u(x) = uk sin(kx) and v(x) = vk cos(kx) k k with the coefficients uo and ve given by { k} { k} Z π Z π o 2 e 2 uk = u(x) sin(kx)dx , vk = v(x) cos(kx)dx , k 1 . π 0 π 0 ∀ ≥
We have the following representations of the PD operators o and e: −Lδ −Lδ X X ou(x) = η (k)uo sin(kx) , ev(x) = η (k)ve cos(kx) , (2.42) −Lδ δ k −Lδ δ k k k where ηδ is the Fourier symbol of the bond-based PD operator on the 1-D bar, defined as
Z δ ηδ(k) = ω( y )(1 cos(ky)) dy k 1 . (2.43) δ | | − ∀ ≥ − X Note that we have used to denote the sum over all positive integers. In k the case of the Neumann boundary condition with the Fourier cosine expansion, we have the constant term removed from the expansion so that the corresponding function always has zero mean. We now define the associated functional spaces and explicit spectra and eigenfunctions for the PD operators, similar to the study 31 for the PD equation in Rd in section 2.2 and in [25].
o Definition 3. The space Mω, which depends on the influence function ω, consists o of all functions u for which the Mω-norm
( )1/2 2 o 1/2 X o2 u o = [ ( u, u)] = η (k)u (2.44) k kMω −π Lδ δ k k
o is finite. The corresponding inner product in Mω is given by
X o o o (u, v) o := η (k)u v u, v M . (2.45) Mω δ k k ∀ ∈ ω k
so In addition, given an exponent s, we let Mω be generalized energy spaces with ( ) so 2 X s o2 M = u : u so = η (k)u < . ω k kMω δ k ∞ k
se Similarly, we have the definitions of Mω and their norms and inner products.
so We can see that Mω is a Hilbert space, with its dual space with respect to the 2 so o standard L duality pairing given by M − . Moreover, the PD operator is a ω −Lδ o o 1o self-adjoint operator, and it is also an isometry from Mω to Mω− = Mσ− . Let Hs denote the fractional Sobolev space on (0, π) for s [0, 1), which is o ∈ s the closure of Co∞ in H (0, π). Here Co∞ is the space of functions which are the s restrictions on (0, π) of functions in C∞(R) satisfying (2.39), and H (0, π) is the 0 2 0o standard fractional order Sobolev space [26]. Note that Ho = Lo = Mσ , and one s can characterize Ho and its norm in the Fourier space as ( ) X X Hs = v = v sin(kx) v 2 = v 2k2s < . o k | k ks | k| ∞ k k
Similarly, we have the same results for spaces M se related to e and Hs. The for- ω Lδ e mulations given in (2.42) are motivated by the explicit eigenfunctions and spectra of the PD-operators o and e. For o, it has a complete set of eigenfunc- −Lδ −Lδ −Lδ tions sin(kx) with the corresponding spectra η (k) . Analogously, e has a { } { δ } −Lδ complete set of eigenfunctions cos(kx) with spectra η (k) . { } { δ } 32
2.3.2 Space equivalence for special influence functions
o We now establish relations between Mω and the Sobolev spaces. Similar results e hold for the space Mω. Our discussion here is similar to that in section 2.1 for func- tions in Rd. We also refer to [21] for discussions of nonlocal spaces in more general non-Hilbert space settings and its application to nonlinear variational problems. Under the assumption (2.3) and with the symbol , being the conventional → notation for the continuous embedding between spaces, we first have what follows.
Lemma 14. Let ω = ω( y ) satisfy (2.3). We have for k 1, | | ≥ Z δ ηδ(k) inf (1 cos(ky))ρ( y )dy > 0 , (2.46) ≥ k 1 δ − | | ≥ − k4δ2τ 0 τ k2 2η (k) δ . (2.47) ≤ δ − δ ≤ 12
Moreover, H1 , M o , L2, H2 , M 2o , L2. Similar results hold for spaces o → ω → o o → ω → o associated with (2.40).
Proof. By the assumption on ω = ω( x ), we have for any k 1 that | | ≥ Z δ ηδ(k) (1 cos(ky))ρ( y )dy > 0 . ≥ δ − | | − Since ρ( x ) L1( δ, δ), by the Riemann lemma, we have | | ∈ − Z δ lim ρ( y ) cos(ky)dy = 0, k δ | | →∞ − which gives Z δ Z δ lim (1 cos(ky))ρ( y )dy = ρ( y )dy > 0 . k δ − | | δ | | →∞ − − Thus, we have the infimum attainable in (2.46), which remains strictly positive, thus implying that M o , L2. ω → o Next, using the inequality 1 cos(x) x 2/2, we have | − | ≤ | | Z δ 2 2 2 0 < 2ηδ(k) ω( y )k y dy τδk k 1. (2.48) ≤ δ | | | | ≤ ∀ ≥ − 33
We get the first inequality in (2.47). Using Taylor expansions of the cosine function, we can also get the second inequality in (2.47). The continuous space embedding results then follow from the definitions of the corresponding spaces. We omit the details here (for a proof of similar results in Rd, we refer to [25]).
The above lemma implies in particular that the solution spaces of the nonlocal BVPs of the PD model are subspaces of L2. In fact, for an influence function ω such that ω( x ) L1(0, δ), L2 is the natural solution space. As shown in [17, 25, 16] | | ∈ for problems defined in Rd, the linear PD operator is a bounded linear operator from L2 to itself. For the finite bar case, we also have what follows.
Lemma 15. Let ω = ω( y ) satisfy the additional condition that | |
ω( y ) L1(0, δ) ; (2.49) | | ∈ then M 2o = M o = L2 and o is a bounded linear operator from L2 to L2, and ω ω o −Lδ o o Z δ 0 < inf ηδ(k) ηδ(k) 4 ω( y )dy k 1. (2.50) k 1 ≥ ≤ ≤ 0 | | ∀ ≥
Similar results hold for spaces associated with (2.40).
Using the definition of ηδ(k) given in (2.43), it is obvious that the rightmost inequality of (2.50) holds, while the lower bound can be derived in the same way as (2.46) with ρ( y ) being replaced by ω( y ). | | | | We note that while the above lemma implies the equivalence of norms between M 2o, M o, and L2 and the boundedness of o from L2 to L2, such equivalence ω ω o −Lδ o o relations and the operator bound of o are not uniform with respect to the −Lδ horizon parameter δ. In fact, the upper bound in (2.50) tends to grow to infinity as δ 0. This is not surprising since, based on the analysis given later in the chapter → 5, the PD operator actually converges to a second order differential operator and 2 2 is thus unbounded from Lo to Lo in such a limit. For more general micromodulus functions, the PD operator may become −Lδ unbounded in L2, yet the basic existence and uniqueness results remain valid as we demonstrate below, with the discussion taking place in other function spaces 34 defined earlier. For convenience, we define the following function:
Z χ 1 cos(z) ωδ(α, χ) = − 1+2α dz . (2.51) χ z − | |
Lemma 16. Let ω = ω( y ) satisfy (2.3),and for some positive constant γ1, | |
1 2α ω( y ) γ1 y − − y δ (2.52) | | ≤ | | ∀| | ≤ for some exponent α (0, 1). Then we have ∈
0 η (k) Cδ(α)2k2α k 1 , (2.53) ≤ δ ≤ 1 ∀ ≥
δ 1 α o 2α 2o for C (α) = (γ1ω (α, )) 2 , and H , M and H , M . Moreover, 1 δ ∞ o → ω o → ω
δ α u o C (α) u u H , (2.54) k kMω ≤ 1 k kα ∀ ∈ o δ 2 2α u 2o C (α) u 2 u H . (2.55) k kMω ≤ 1 k k α ∀ ∈ o
Similar results hold for spaces associated with (2.40).
Proof. The lemma follows straightforwardly from the observation that for any k 1, the coefficient η (k) as defined in (2.43) satisfies ≥ δ Z δ 1 δ 2 2α 0 ηδ(k) γ1 (1 cos(ky)) 1+2α dy C1 (α) k , ≤ ≤ δ − y ≤ − | | which in turn leads to the continuous space embeddings using the definitions of the corresponding spaces.
Similarly, we also have the following lemma.
Lemma 17. Let ω = ω( y ) satisfy (2.3) and for some positive constant γ2, | |
1 2β ω( y ) γ2 y − − y δ (2.56) | | ≥ | | ∀| | ≤ for some exponent β < 1. Then we have
η (k) Cδ(β)2k2β k 1 , (2.57) δ ≥ 2 ∀ ≥ 35
δ 1 o β 2o 2β for C (β) = (γ2ω (β, δ)) 2 , and M , H and M , H . Moreover, 2 δ ω → o ω → o
δ o C (β) u u o u M , (2.58) 2 k kβ ≤ k kMω ∀ ∈ ω δ 2 2o C (β) u 2 u 2o u M . (2.59) 2 k k β ≤ k kMω ∀ ∈ ω
Similar results hold for spaces associated with (2.40).
The proof again follows from similar estimates on η (k) for any k 1: δ ≥ Z δ 1 δ 2 2β ηδ(k) γ2 (1 cos(ky)) 1+2β dy C2 (β) k . ≥ δ − y ≥ − | | We note that for β < 0, Lemma 15 can be applied, which gives stronger results. Based on the above discussion, we see that under suitable conditions on the influence function, for example, if (2.52) and (2.56) are satisfied with β = α ∈ o 2o (0, 1), then the spaces Mω and Mω are equivalent to standard fractional Sobolev α 2α spaces Ho and Ho , respectively. We caution again that the equivalence relations are not uniform as δ 0. → Let us also note that for many influence functions used in the existing studies of the linear PD models, conditions (2.49), (2.52), and (2.56) are often very relevant. 1 2β In fact, influence functions of the form ω( y ) = γ2 y − − with β < 1, especially | | | | 1 the case with β = 0 and ω = γ2 y − , have been frequently used in the literature | | [5, 7, 27].
2.3.3 Properties of the stationary PD Model
We now discuss the existence and uniqueness of weak solutions to the stationary o (equilibrium) PD model. We again focus on the solution in Mω of the equation
ou = f . (2.60) −Lδ
For completeness, parallel conclusions are also stated (without proof) on the solu- e tion in Mω of the equation eu = f . (2.61) −Lδ 36
Similar to elliptic equations, we may establish the variational theory and regularity properties for the stationary PD model. First, we obviously have what follows.
o Lemma 18. Let ω = ω( y ) satisfy (2.3); then for any f M − , the problem | | ∈ ω (2.60) has a unique solution u M o which is the minimizer of the functional ∈ ω
1 2 1X o2 u o (u, f)L2 = ηδ(k)u (u, f)L2 2k kMω − 2 k − k
o in Mω. A similar result holds for (2.61). As for the regularity, we have for some special micromodulus functions what follows.
Lemma 19. Let ω = ω( y ) satisfy (2.3) and (2.56). For β [0, 1), the problem | | ∈ (2.60) has a unique solution u = u(x) Hm+2β whenever f Hm for any m ∈ o ∈ o ≥ 2β. Moreover, − δ 2 u +2 C (β)− f . k km β ≤ 2 k km Similar results hold for (2.61).
Proof. Considering (2.60), with f o being the coefficients of the Fourier sine { k } expansion of f, obviously we have a unique solution u = u(x) with uo = f o/η (k) { k k δ } being the coefficients of its Fourier sine expansion. Moreover,
X X X Cδ(β)4k2m+4βuo2 η2(k)k2muo2 = k2mf o2 < . 2 k ≤ δ k k ∞ k k k
So, we have the solution u = u(x) Hm+2β with the desired bound. ∈ o The above lemma implies a lifting in regularity on the order of 2β (0, 2) for ∈ influence functions satisfying (2.52) and (2.56). We now go back to consider the regularity of solutions of (2.60) and (2.61) for more general cases.
Lemma 20. Let ω = ω( y ) satisfy (2.3); the problem (2.60) has a unique solution | | o o 2 u M for f M − , and u o = f o . Moreover, if f L , we have ∈ ω ∈ ω k kMω k kMω− ∈ o 2o u M and u 2o = f 0. Similar results hold for (2.61). ∈ ω k kMω k k Using the linearity of the model, we have the following corollary, similar as in [25]. We now let D be a difference operator given by D v(x) = v(x + y) v(x) for y y − 37 any function v defined in a suitable space. Then let P be a linear operator which 1 commutes with Dy for any y with P − its inverse (defined in a suitable space). Moreover, for any measurable function v that is odd in ( δ, δ) and (π δ, π + δ), − − we assume that P v, in the sense of distributions, remains odd in ( δ, δ) and − (π δ, π + δ). Examples of such operators include linear differential operators of − even orders with constant coefficients.
Corollary 2. Under the above assumptions on P , we have P o = oP and Lδ Lδ 1 o o 1 P − = P − . This in turn implies that if u is a solution of (2.60) with a Lδ Lδ 1 given function f, then P u and P − u are the solutions of (2.60) with the right 1 hand side being P f and P − f, respectively. A similar conclusion holds for (2.61).
Although the above corollary is trivial to derive, we note that problems with singular data are indeed of practical interests; see [6] for examples where Dirac- delta functions are used for f as concentrated forces. While o is a nonlocal integral operator, it shares some properties of an −Lδ elliptic differential operator such as those associated with the elliptic regularity, as shown earlier. As another illustration, we know that if u satisfies (2.60), with ω satisfying (2.49) and f being nonnegative, then
Z δ Z δ ω( y ) dy u(x) u(x + y)ω( y ) dy x (0, π) . δ | | ≥ δ | | ∀ ∈ − − Thus, we can readily get the following maximum principle.
Lemma 21. Let ω = ω( y ) satisfy (2.49), let f be nonnegative almost everywhere, | | and let u be a solution of (2.60). Then u is nonnegative in (0, π), and its maximum is attained in the interior or it is identically zero. A similar conclusion holds for (2.61).
2.3.4 The time-dependent PD model on a finite bar
Similar results can be established for the time-dependent PD models by adapting the approaches given in [25] to the setting of nonlocal BVPs as the case for the equilibrium models. We thus simply state a result without detailed derivation. 38
Theorem 4. Let ω = ω( y ) satisfy (2.3), and let P be a time-independent operator | | o that satisfies the same assumptions as those made in the Corollary 2. If P g0 M , ∈ ω 2 2 2 P g1 L , and P b L (0,T ; L ), then (2.38) with the nonlocal boundary condition ∈ o ∈ o (2.39), the initial conditions u(x, 0) = g0(x), and ut(x, 0) = g1(x) and the forcing term b = b(x, t) has a unique solution u = u(x, t) with P u C([0,T ],M o) ∈ ω ∩ H1(0,T ; L2). In particular, for P = I, we get u C([0,T ],M o) H1(0,T ; L2) o ∈ ω ∩ o o 2 2 2 2 with g0 M , g1 L , and b L (0,T ; L ). We also have u C([0,T ],L ) ∈ ω ∈ o ∈ o ∈ o ∩ 1 o 2 o 2 o H (0,T ; M − ) with g0 L , g1 M − , and b L (0,T ; M − ). Similar results ω ∈ o ∈ ω ∈ ω hold for problems associated with the nonlocal boundary condition (2.40).
2.4 Conclusion
In this chapter, a general functional analytic framework is provided for the mathe- matical and numerical analysis of the linear peridynamic models. For illustration, we focus on the case of the linear constitutive relations corresponding to the spring system in multi-dimensional space and one-dimensional boundary value problem. Various analytical issues are established here under the unified framework, extend- ing some of the results given in the literature. The techniques developed here can be extended to study more general nonlocal peridynamic state models [3]. We note that the analytic frameworks and the stud- ies of the solution regularity properties associated with the PD models can also be useful in establishing basic convergence and error estimates of their numerical ap- proximations such as the Galerkin finite element approximation [28] and chapter 6. While the Fourier based techniques similar to that developed here can still be used in the analysis of certain special nonlocal boundary value problems for the linear bond based PD models defined on box-like domains [28], other techniques need to be further developed in the future to treat more generic boundary conditions, arbitrary geometry and nonlinear models. Chapter 3
A Nonlocal Vector Calculus, Nonlocal Balance Laws and Nonlocal Volume-Constrained Problems
3.1 Introduction
In this chapter, we introduce a vector calculus for nonlocal operators that mimics the classical vector calculus for differential operators. We define nonlocal analogs of the divergence, gradient, and curl operators and deduce the corresponding nonlocal adjoint operators. Nonlocal analogs of the Gauss theorem and the Green’s identi- ties of the vector calculus for differential operators are also derived. We establish relationships between the nonlocal operators and their differential counterparts. The nonlocal vector calculus can be used to define nonlocal volume-constrained problems that are analogous to boundary-value problems for partial differential operators. In addition, The nonlocal vector calculus we develop has an impor- tant application to balance laws1 that are nonlocal in the sense that subregions not in direct contact may have a non-vanishing interaction. This is accomplished by defining a nonlocal flux in terms of interactions between disjoint regions of positive measure possibly sharing no common boundary. Our nonlocal vector cal-
1A balance law postulates that the rate of change of an extensive quantity over any subregion of a body is given by the rate at which that quantity is produced in the subregion minus the flux out of the subregion through its boundary. 40 culus, then, provides an alternative to standard approaches for circumventing the technicalities associated with the lack of sufficient regularity in local balance laws. Preliminary attempts at a nonlocal calculus are found in [29] and [18] which included applications to image processing and steady-state diffusion, respectively. In particular, the authors of [29] and [30] where a discrete nonlocal divergence and gradient are introduced within the context of machine learning; see also of of [31], [32], and [33] where a discrete calculus is also discussed. However, the discussion in these papers is limited to scalar problems. In contrast, this paper extends the ideas in [29] and [18] to vector and tensor fields and beyond the consideration of image processing and steady-state diffusion. For example, the ideas presented here enable an abstract formulation of the balance laws of momentum and energy and for the peridynamic theory for solid mechanics2 that parallels the classical vector calculus formulation of the balance laws of elasticity. The nonlocal vector calculus presented in this chapter, however, is sufficiently general that we envisage application to balance laws beyond those of elasticity, e.g., to the laws of fluid mechanics and electromagnetics. This chapter is organized as follows. The remainder of this section is devoted to establishing notation. In Section 3.2, the notions of local and nonlocal fluxes into or out off a region are briefly compared and contrasted. In Section 3.3, the nonlocal divergence, gradient, and curl operators are introduced as are the corre- sponding adjoint operators, several vector identities, and other results about the operators. The nonlocal vector calculus is developed in Section 3.4; in particular, nonlocal integral theorems and nonlocal Green’s identities are derived. In Sec- tion 3.5, connections between the nonlocal operators and distributional and weak representations of the associated classical differential operators are made. The connections made in Section 3.5 justify the use of the terminology “nonlocal di- vergence, gradient, and curl” to refer to the nonlocal operators defined in Section 3.3. Sections 3.6 and 3.7 deal with applications of the nonlocal vector calculus. In Section 3.6, examples are given of nonlocal volume-constrained problems formu- lated in terms of the nonlocal operators. Then, in Section 3.7, a brief review of the conventional notion of a balance law is provided after which abstract nonlocal
2Peridynamics was introduced in [1] and [2]; [4] reviews the peridynamic balance laws of mo- mentum and energy and provides many citations for the peridynamic theory and its applications. See Section 3.7.2 for a brief discussion. 41 balance laws are discussed. The notion of nonlocal fluxes discussed in Section 3.2 is used in developing nonlocal balance laws and the vector calculus developed in Section 3.4 plays a crucial role in transforming balance laws into field equations. Also, in Section 3.7, a brief discussion is given of the application of our nonlocal vector calculus to the peridynamic theory for continuum mechanics. Throughout, wherever it is illuminating, we associate definitions and results of the nonlocal vector calculus with the analogous definitions and results of the classical differential vector calculus.
3.1.1 Notation
We have need of two types of functions and two types of nonlocal operators. Point functions refer to functions defined at points whereas two-point functions refer to functions defined for pairs of points. Point operators map two-point functions to point functions whereas two-point operators map point functions to two-point func- tions so that the nomenclature for operators refer to their ranges. Point and two- point operators are both nonlocal. Point operators involve integrals of two-point functions whereas two-point operators explicitly involve point functions evaluated at two different points. We now make more precise the definitions given above. Points in Rn are denoted by the vectors x, y, or z and the natural Cartesian basis is denoted by e1,..., en. n m k Let m, k, and n denote positive integers. For Ω R , functions from Ω into R × ⊆ or Rm or R are referred to as point functions or point mappings and are denoted by Roman letters, upper-case bold for tensors, lower-case bold for vectors, and plain face for scalars, respectively, e.g., U(x), u(x), and u(x), respectively. Functions m k m from Ω Ω into R × or R or R are referred to as two-point functions or two- × point mappings and are denoted by Greek letters, upper-case bold for tensors, lower-case bold for vectors, and plain face for scalars, respectively, e.g., Ψ(x, y), ψ(x, y), and ψ(x, y), respectively. Symmetric and antisymmetric scalar two-point functions ψ(x, y) satisfy ψ(x, y) = ψ(y, x) and ψ(x, y) = ψ(y, x), respectively, − and similarly for vector and tensor two-point functions. A useful observation is 42 that Z Z if ψ(x, y) is antisymmetric, then ψ(x, y) dydx = 0 Ω Rn (3.1) Ω Ω ∀ ⊆ and similarly for antisymmetric vector and tensor two-point functions. For the sake of notational simplicity, in much of the rest of the chapter, we introduce the following notation:
α := α(x, y) α0 := α(y, x) ψ := ψ(x, y) ψ0 := ψ(y, x)
u := u(x) u0 := u(y) u := u(x) u0 := u(y) and similarly for other functions. The dot (or inner) product of two vectors u, v Rm is denoted by u v R; ∈ · ∈ m k k the dyad (or outer) product is denoted by u w R × whenever w R ; given ⊗ ∈ ∈ k m a second-order tensor (matrix) U R × , the tensor-vector (or matrix-vector) ∈ product is denoted by U v and is given by the vector whose components are the · dot products of the corresponding rows of U with v.3 For n = 3, the cross product of two vectors u and v is denoted by u v R3. The Frobenius product of two × ∈ m k m k second-order tensors A R × and B R × , denoted by A: B, is given by ∈ ∈ m m the sum of the element-wise product of the two tensors. The trace of B R × , ∈ denoted by tr B, is given by the sum of the diagonal elements of B. Inner products in L2(Ω) and L2(Ω Ω) are defined in the usual manner. For × example, for vector functions, we have Z (u, v)Ω = u v dx for u(x), v(x) Ω Ω · ∈ Z Z (µ, ν)Ω Ω = µ ν dydx for µ(x, y), ν(x, y) Ω × Ω Ω · ∈ with analogous expressions involving the Frobenius product and the ordinary prod- uct for tensor and scalar functions, respectively.
3In matrix notation, the inner, outer, matrix-vector products are given by x y = xT y, x y = xyT , and U v = Uv. · ⊗ · 43
3.2 Nonlocal fluxes and nonlocal action-reaction principles
A key concept in the development of a vector calculus is the notion of a flux which accounts for the interaction of points in a domain with points outside the domain. As a result, the notion of a flux is also fundamental to the understanding of balance laws in mechanics, heat transfer, and many other settings; see Section 3.7. In the classical setting of local interactions, that interaction occurs at the boundary of the domain, whereas in the nonlocal case, the interaction must occur over volumes external to the domain. In order to contrast the notion of a nonlocal flux with the classical local flux, we begin by briefly reviewing the latter notion.
3.2.1 Local fluxes
n n Let Ω1 R and Ω2 R denote two disjoint open regions. If Ω1 and Ω2 have a ⊂ ⊂ nonempty common boundary ∂Ω12 := Ω1 Ω2, then, for a vector-valued function ∩ q(x), the expression Z q ~n1 dA (3.2) ∂Ω12 · represents the classical local flux out of Ω1 into Ω2, where ~n1 denotes the unit normal on ∂Ω12 pointing outward from Ω1 and dA denotes a surface measure in n R ; q ~n1 is referred as the flux density along ∂Ω12. The vector q is often expressed · in terms of an intensive variable through a constitutive relation.4 The flux, then, conveys a notion of direction out of and into a region and is a proxy for the interaction between Ω1 and Ω2. It is important to note that the flux from Ω1 into
Ω2 occurs across their common boundary and that if the two disjoint regions have no common boundary, then the flux from one to the other is zero. The classical flux
(3.2) is then deemed to be local because there is no interaction between Ω1 and Ω2 when separated by a finite distance. The classical flux satisfies the action-reaction
4 For example, if q ~n1 denotes the heat flux density, then q is related to the temperature via Fourier’s heat law; see· Section 3.7.1.1. 44 principle5 Z Z q ~n1 dA + q ~n2 dA = 0, (3.3) ∂Ω12 · ∂Ω21 · where, of course, ∂Ω12 = ∂Ω21 and ~n2 = ~n1 denotes the unit normal on ∂Ω21 − R pointing outward from Ω2. In words, the flux q ~n1 dA from Ω1 into Ω2 across ∂Ω12 · R their common boundary ∂Ω12 is equal and opposite to the flux q ~n2 dA from ∂Ω21 · Ω2 into Ω1 across that same surface.
3.2.2 Nonlocal fluxes
We identify Z Z ψ(x, y) dydx (3.4) Ω1 Ω2 as a scalar interaction, or nonlocal flux, from Ω1 into Ω2, where ψ : (Ω1 Ω2) R ∪ × (Ω1 Ω2) denotes an antisymmetric function. We have that ψ(x, y) dy is R Ω2 ∪ → R the flux density into Ω2 from the point x Ω1 and likewise, ψ(x, y) dx is the ∈ Ω1 flux density into Ω1 from the point y Ω2. As is the case for the local flux density R ∈ q ~n1, the nonlocal flux density ψ(x, y) dy is related to an intensive variable · Ω2 through a constitutive relation; see Section 3.7.1.2. From (3.1), it is easily seen that the antisymmetry of ψ(x, y) is equivalent to the nonlocal action-reaction principle Z Z Z Z n ψ(x, y) dy dx + ψ(x, y) dy dx = 0 Ω1, Ω2 R ; (3.5) Ω1 Ω2 Ω2 Ω1 ∀ ⊂
(3.5) is the nonlocal analogue of (3.3). In words, (3.5) states that the flux (or interaction) from Ω1 into Ω2 is equal and opposite to the flux (or interaction) from
Ω2 into Ω1. The flux is nonlocal because, by (3.5), the interaction may be nonzero even when the closures of Ω1 and Ω2 have an empty intersection. This is in stark contrast to classical local interactions for which we have seen that the interaction between Ω1 and Ω2 vanishes if their closures have empty intersection, i.e., have no common boundary.
5An example is in mechanics for which Newton’s third law, i.e., the force exerted upon on object is equal and opposite to the force exerted by the object, is an action-reaction archtype. 45
3.3 Nonlocal operators
The nonlocal vector calculus developed in Section 3.4 involves nonlocal operators that mimic the classical local differential divergence, gradient, and curl operators. An important distinction between local and nonlocal operators is that the adjoint operators for the former involve the same operators, i.e., the adjoint of is , ∇· −∇ of is , and of is , whereas the adjoint of nonlocal operators involve ∇ −∇· ∇× ∇× differently defined nonlocal gradient, divergence, and curl operators. At this point, the association of the nonlocal operators given in Definition 4 and Theorem 5 with the operators of the classical differential vector calculus is purely conjectural. Justifications for making these associations are provided in Section 3.5.
3.3.1 Nonlocal point divergence, gradient, and curl opera- tors
The nonlocal point divergence, gradient, and curl operators map two-point func- tions to point functions and are defined in terms of their action on two-point functions as follows. These operators along with their adjoints are the building blocks of our nonlocal calculus.
Definition 4. [Nonlocal operators] Given the vector two-point function ν : Rn × Rn Rk and the antisymmetric vector two-point function α: Rn Rn Rk, the → × → action of the nonlocal point divergence operator on ν is defined as D Z n ν (x) := ν + ν0 α dy for x R , (3.6a) D Rn · ∈ where ν : Rn R. Given the scalar two-point function η : Rn Rn R and D → × → the antisymmetric vector two-point function β : Rn Rn Rk, the action of the × → nonlocal point gradient operator on η is defined as G Z n η (x) := η + η0 β dy for x R , (3.6b) G Rn ∈