# Generalized solutions of systems of nonlinear partial differential equations

Generalized solutions of systems of nonlinear partial differential equations by Jan Harm van der Walt Submitted in partial fulfilment of the requirements of the degree Philosophiae Doctor in the Department of Mathematics and Applied Mathematics in the Faculty of Natural and Agricultural Sciences University of Pretoria Pretoria February 2009 © University of Pretoria i DECLARATION I, the undersigned, hereby declare that the thesis submitted herewith for the degree Philosophiae Doctor to the University of Pretoria contains my own, independent work and has not been submitted for any degree at any other university. Name: Jan Harm van der Walt Date: February 2009 ii Title Generalized solutions of systems of nonlinear partial differential equations Name Jan Harm van der Walt Supervisor Prof E E Rosinger Co-supervisor Prof R Anguelov Department Mathematics and Applied Mathematics Degree Philosophiae Doctor Summary In this thesis, we establish a general and type independent theory for the existence and regularity of generalized solutions of large classes of systems of nonlinear partial differential equations (PDEs). In this regard, our point of departure is the Order Completion Method. The spaces of generalized functions to which the solutions of such systems of PDEs belong are constructed as the completions of suitable uniform convergence spaces of normal lower semi-continuous functions. It is shown that large classes of systems of nonlinear PDEs admit generalized solutions in the mentioned spaces of generalized functions. Furthermore, the generalized solutions that we construct satisfy a blanket regularity property, in the sense that such solutions may be assimilated with usual normal lower semi-continuous functions. These fundamental existence and regularity results are obtain as applications of basic topological processes, namely, the completion of uniform convergence spaces, and elementary properties of real valued continuous functions. In particular, those techniques from functional analysis which are customary in the study of nonlinear PDEs are not used at all. The mentioned sophisticated methods of functional analysis are used only to obtain additional regularity properties of the generalized solutions of systems of nonlinear PDEs, and are thus relegated to a secondary role. Over and above the mentioned blanket regularity of the solutions, it is shown that for a large class of equations, the generalized solutions are in fact usual classical solutions of the respective system of equations everywhere except on a closed, nowhere dense subset of the domain of definition of the system of equations. This result is obtained under minimal assumptions on the smoothness of the equations, and is an application of convenient compactness theorems for sets of sufficiently smooth functions with respect to suitable topologies on spaces of such functions. As an application of the existence and regularity results presented here, we obtain for the first time in the literature an extension of the celebrated Cauchy-Kovalevskaia Theorem, on its own general and type independent grounds, to equations that are not analytic. Preface For nearly four centuries, ordinary and partial differential equations have been one of the main tools by which scientists sought to describe the laws of nature in exact mathematical terms. At first, most of these equations were of a particular form, namely, linear and of second order. However, with the emergence of increasingly sophisticated scientific theories and state of the art technologies, in particular during the second half of the twentieth century, the interest of mathematicians, and scientists in general, shifted towards nonlinear equations. It became clear rather early on that the methods developed to deal with linear equations, such as the linear theory of distributions, in particular in the case of partial differential equations, are inappropriate for nonlinear equations. In fact, it is typically believed that a convenient and general theory for the solutions of nonlinear partial differential equations is impossible, or at best highly unlikely. This perception has lead to the development of several ad hoc solution methods for nonlinear partial differential equations, each developed with but a small class of equations, if not one single equation, in mind. While such methods may prove to be highly effective in those cases to which they apply, there is no attempt at a deeper understanding of the underlying nonlinear phenomena involved. The alternative to the mentioned ad hoc solution methods is to establish a general theory for the existence and regularity of solutions of nonlinear partial differential equations. To date there are three such general theories, namely, the theory of algebras of generalized functions introduced independently by Colombeau and Rosinger, the so called Cental Theory of partial differential equations developed by Neuberger, and the Order Completion Method developed by Oberguggenberger and Rosinger. These three theories, each based on different techniques and perspectives on partial differential equations, apply to large classes of nonlinear partial differential equations, and are not restricted to any particular type of partial differential equation. In this work, we present a fourth such general and type independent theory for the existence and regularity of solutions of systems of nonlinear partial differential equations. Our point of departure is the mentioned Order Completion Method. As such the theory that we present here may, to a certain extent, be considered also as a regularity theory for the solutions of systems of nonlinear partial differential equations delivered through the Order Completion Method. However, we go far beyond that basic theory by introducing new spaces of generalized functions, the elements iii PREFACE iv of which act as solutions of systems of nonlinear partial differential equations in a suitable extended sense. The mentioned spaces of generalized functions are constructed as the completion of suitable spaces of usual real valued functions, equipped with appropriate uniform convergence structures. Generalizations of the usual systems of partial differential equations are obtained by extending suitable mappings associated with such a given system of equations to the mentioned spaces of generalized functions. This is done in a consistent and rigorous way by ensuring that these mappings are suitably compatible with the mentioned uniform convergence structures on the spaces of functions. The existence of generalized solutions follows as an application of certain basic approximation results. The thesis is divided into two parts. Part I contains the introductory chapters, which include a historical overview of the subjects of nonlinear partial differential equations and topology. We also include chapters on real and interval valued functions, and the role of ordered structures in analysis and topology. These chapters contain some results and definitions that are relevant to the work presented in subsequent chapters. Part II contains our original contributions, which we now mention briefly. • In Chapter 6 we investigate the structure of the completion of a uniform convergence space. In particular, we consider uniform convergence structures that arise as initial structures with respect to families of mappings. • Nearly finite normal lower semi-continuous functions are introduced in Chapter 7. A uniform convergence structure is defined on a suitable space of such functions, and its completion is characterized. These spaces of normal lower semi-continuous functions are the fundamental spaces upon which the spaces of generalized functions used in this work are constructed. • The spaces of generalized functions that we introduce here are constructed in Chapter 8. In particular, Section 8.1 concerns the so called pullback type spaces of generalized functions, while Section 8.2 introduces the new Sobolev type spaces of generalized functions. In Section 8.3 we discuss the nonlinear partial differential operators which act on these spaces, as well as the extent to which the different types of spaces are related to one another. • Chapter 9 deals with the issues of existence of solutions of large classes of systems of nonlinear partial differential equations in the mentioned spaces of generalized functions. In Section 9.1 we give a number of approximation results for the solutions of such systems of equations. These are used in Sections 9.2 through 9.4 to prove the existence of solutions in the various spaces of generalized functions that are constructed in Chapter 8. • We proceed in Chapter 10 to show that a large class of equations admit solutions in the mentioned Sobolev type spaces of generalized functions, which are PREFACE v in fact classical solutions everywhere except on a closed nowhere dense subset of the domain of definition of system of equations. This regularity result is obtained as an application of certain compactness results for sets of sufficiently smooth functions, with respect to an appropriate topology on suitable spaces of such smooth functions. • Chapter 11 deals with the issues of boundary and / or initial conditions that may be associated with a given system of nonlinear partial differential equations. In this regard, we consider a general, nonlinear Cauchy problem. It is shown that such an initial value problem admits a solution in a suitable generalized sense. We also show that, under minimal assumptions on the smoothness of the initial data and the nonlinear partial differential operator that defines the system of equations, such an initial value problem admits a solution which is a classical solution everywhere except on a closed and nowhere dense subset of the domain of definition of the equations. At this point, a remark on the numbering of results is appropriate. This work contains 98 definitions, propositions, lemmas, corollaries, theorems, examples and remarks. These are numbered 1 through 98. All definitions in Part I are taken from the literature, and the relevant citations are indicated. Results form the literature are always marked with an asterisk (*), followed immediately by the relevant citation. Acknowledgements The author would like to thank the following people, each of whom, in some way or another, had made an invaluable contribution towards the completion of this thesis. My supervisors, Prof. Roumen Anguelov and Prof. Elemer Rosinger, with whom I had many insightful discussions, and who, after reading preliminary drafts of the thesis, made many useful suggestions. I would like to thank my family, most notably may parents, Peet and Susan van der Walt, my brother Pieter van der Walt and my sister Lydia du Toit for their support and encouragement. A special thanks is due to Lizelda Botha, who’s love and support has been a pilar of strength to me during the course of the past two and a half years. vi Contents Preface iii Acknowledgements vi I 1 Introduction 1 A Brief History of PDEs 1.1 From Newton to Schwarz . . . . . . . . . . . . 1.2 Weak Solution Methods . . . . . . . . . . . . 1.3 Differential Algebras of Generalized Functions 1.4 The Order Completion Method . . . . . . . . 1.5 Beyond distributions . . . . . . . . . . . . . . 2 Topological Structures in Analysis 2.1 Point-Set Topology: From Hausdorff to 2.2 The Deficiencies of General Topology . 2.3 Convergence Spaces . . . . . . . . . . . 2.4 Uniform Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 15 21 29 43 Bourbaki . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 45 50 54 66 . . . . . 3 Real and Interval Functions 77 3.1 Semi-continuous Functions . . . . . . . . . . . . . . . . . . . . . . . . 77 3.2 Interval Valued Functions . . . . . . . . . . . . . . . . . . . . . . . . 84 4 Order and Topology 90 4.1 Order, Algebra and Topology . . . . . . . . . . . . . . . . . . . . . . 90 4.2 Convergence on Posets . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5 Organization of the Thesis 96 5.1 Objectives of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 96 5.2 Arrangement of the Material . . . . . . . . . . . . . . . . . . . . . . . 98 vii CONTENTS II viii Convergence Spaces and Generalized Functions 6 Initial Uniform Convergence Spaces 6.1 Initial Uniform Convergence Structures . . 6.2 Subspaces of Uniform Convergence Spaces 6.3 Products of Uniform Convergence Spaces . 6.4 Completion of Initial Uniform Convergence 101 . . . . 102 . 102 . 105 . 108 . 109 . . . . 114 . 114 . 115 . 124 . 127 8 Spaces of Generalized Functions 8.1 The Spaces MLm T (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Sobolev Type Spaces of Generalized Functions . . . . . . . . . . . . 8.3 Nonlinear Partial Differential Operators . . . . . . . . . . . . . . . . 131 . 131 . 136 . 139 9 Existence of Generalized Solutions 9.1 Approximation Results . . . . . . . . . . . . . . . . 9.2 Solutions in Pullback Uniform Convergence Spaces 9.3 How Far Can Pullback Go? . . . . . . . . . . . . . 9.4 Existence of Solutions in Sobolev Type Spaces . . . 147 . 147 . 157 . 161 . 169 . . . . . . . . . . . . . . . . . . . . . Structures . 7 Order Convergence on ML (X) 7.1 Order Convergence and the Order Completion Method 7.2 Spaces of Lower Semi-Continuous Functions . . . . . . 7.3 The Uniform Order Convergence Structure on ML (X) 7.4 The Completion of ML (X) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Regularity of Generalized Solutions 175 10.1 Compactness Theorems in Function Spaces . . . . . . . . . . . . . . . 175 10.2 Global Regularity of Solutions . . . . . . . . . . . . . . . . . . . . . . 182 11 A Cauchy-Kovalevskaia Type Theorem 187 11.1 Existence of Generalized Solutions . . . . . . . . . . . . . . . . . . . . 187 11.2 Regularity of Generalized Solutions . . . . . . . . . . . . . . . . . . . 202 12 Concluding Remarks 213 12.1 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 12.2 Topics for Further Research . . . . . . . . . . . . . . . . . . . . . . . 216 Part I Introduction 1 Chapter 1 A Brief History of PDEs 1.1 From Newton to Schwarz The advent of the Differential and Integral Calculus in the later half of the seventeenth century heralded the start of a new age in Mathematics. This is true both in regards to the applications of Mathematics to other sciences, notably to Physics, Economics, Chemistry and lately also Biology, and also in respect of the development of Mathematics as such. Indeed, in connection with the second aspect of these new developments, we may note that from the Differential and Integral Calculus, and the new point of view it introduced in so far as mathematical functions are concerned, the vast and powerful field of Mathematical Analysis developed. On the other hand, and parallel to the above mentioned development of abstract Mathematics, Newton’s Calculus provides powerful new tools with which to solve so called “real world problems”. Indeed, it was exactly the consideration of physical problems, namely, The Laws of Motion, which lead Newton to the conception of the Calculus. With respect to the above mentioned power of the Calculus when it comes to the mathematical solution of physical problems, we need only note the following. Prior to the invention of the Differential Calculus, the only type of motion that could be described in a mathematically precise way was that of a particle moving uniformly along a straight line, and that of a particle moving at constant angular momentum along a circular path. In contradistinction with this rudimentary earlier state of affairs, and as is well known, Newton’s Differential and Integral Calculus provides the appropriate mathematical machinery for the formulation, in precise mathematical terms, and solution of the basic laws of nature, thus going incomparably farther than the mentioned simple motions. The mentioned mathematical expressions of these laws, using the Calculus, typically take the form of systems of ordinary differential equations (ODEs), or systems of partial differential equations (PDEs). At first, such systems of equations were mostly of a particular form, namely, linear and at most of second order. However, with the emergence of increasingly sophisticated physical theories, and the devel2 CHAPTER 1. A BRIEF HISTORY OF PDES 3 opment of state of the art technologies, in particular during the second half of the twentieth century, the need arose for more complex mathematical models, which typically take the form of systems of nonlinear ODEs or PDEs. As such it is clear that the theoretical treatment of nonlinear PDEs, in particular in connection with the existence of the solutions to such equations, and the properties of these solutions, should such solutions exists, is of major interest. In this regard, and for over a century by now, there have been general and type independent results on the existence and regularity of the solutions to both systems of ODEs and systems of PDEs. In particular, in the case of systems of ODEs, one may recall Picard’s Theorem [125]. Theorem 1 *[125] Consider any system of K ODEs in K unknown functions u1 , ..., uK of the form dy = F (t, y1 (t) , ..., yK (t)) , dt (1.1) where F : R×RK → RK is continuous on a neighborhood V of the point (t0 , (y1,0 , ..., yK,0 )) ∈ R × RK . Then there is some δ > 0 which depends continuously on the initial data (t0 , (y1,0 , ..., yK,0 )), and a solution y ∈ C 1 (t0 − δ, t0 + δ)K to (1.1) that also satisfies the initial condition y (t0 ) = (yi,0 ) . Furthermore, if F is Lipschitz on V , then the solution is unique. In modern accounts of the theory of ODEs, Theorem 1 is typically presented as an application of Banach’s fixed point principle in Banach spaces. This might lead to the impression that this result is obtained as an application of functional analysis. However, and as mentioned, Picard’s proof [125] of Theorem 1 predates the formulation of linear functional analysis, which culminated in Banach’s similar work [17]. In fact, Picard’s proof is based on techniques from the classical theory of functions, notably integration of usual smooth functions. One may also note that, around the same time that Picard proved Theorem 1, Peano [123] gave a proof of the existence of a solution of (1.1), which is based on the Arzellà-Ascoli Theorem. In the case of systems of PDEs, the first comparable general and type independent existence and regularity result is due to Kovalevskaia [86]. It is interesting, in view of the common perception that ODEs are far simpler objects than PDEs, to note that the Cauchy-Kovalevskaia Theorem precedes Picard’s Theorem by about twenty years, and as such is not, and in fact could not at the time, be based on more advanced mathematics. Theorem 2 *[86] Consider the system of K nonlinear partial differential equations of the form Dtm u (t, y) = G t, y, ..., Dtp Dyq ui (t, y) , ... (1.2) CHAPTER 1. A BRIEF HISTORY OF PDES 4 with t ∈ R, y ∈ Rn−1 , m ≥ 1, 0 ≤ p < m, q ∈ Nn−1 , p + |q| ≤ m and with the analytic Cauchy data Dtp u (t0 , y) = gp (y) , 0 ≤ p < m, (t0 , y) ∈ S (1.3) on the noncharacteristic analytic hypersurface S = {(t0 , y) : y ∈ Rn−1 }. If the mapping G is analytic, then there exists a neighborhood V of t0 in R, and an analytic function u : V × Rn−1 → CK that satisfies (1.2) and (1.3). It should be noted that the two general and type independent existence and regularity results, namely, Theorem 1 and Theorem 2, for systems of ODEs, respectively PDEs, predates the invention of functional analysis by nearly fifty years. Moreover, and as all ready mentioned in connection with Theorem 1, these results are based on some comparatively elementary mathematics. In particular, the only so called “hard mathematics” involved in the proof of Theorem 2 is the power series expansion for an analytic function, and certain Abel-type estimates for these expansions. Both of the existence results, Theorem 1 and Theorem 2, are local in nature. That is, the solution cannot be guaranteed to exists on the whole domain of definition of the respective systems of equations. Furthermore, and as can be seen from rather simple examples, this is not due to the limitations of the particular techniques used to prove these results, but may instead be attributed to the very nature of nonlinear ODEs and PDEs in such a general setup. The mentioned local nature of Theorems 1 and 2 is unsatisfactory in at least two respects. In the first place, in many of the physical problems that are supposed to be modeled by the respective system of ODEs or PDEs, we may be interested in solutions which exists on domains that are larger than that delivered by the respective existence results in Theorem 1 and Theorem 2, respectively. Secondly, and as can be seen from rather elementary examples, classical solutions, such as those obtained in the mentioned existence results, will in general fail to exists on the entire domain of physical interest. In this regard, a particularly simple, yet relevant, example is the conservation law Ut + Ux U = 0, t > 0, x ∈ R (1.4) U (0, x) = u (x) , x ∈ R. (1.5) with the initial condition If we assume that the function u in (1.5) is smooth enough, then a classical solution U , in fact an analytic solution, of (1.4) to (1.5) is given by the implicit equation U (t, x) = u (x − tU (t, x)) , t ≥ 0, x ∈ R. (1.6) CHAPTER 1. A BRIEF HISTORY OF PDES 5 According to the implicit function theorem, we can obtain U (s, y) from (1.6) for s and y in suitable neighborhoods of t and s, respectively, whenever tu0 (x − tU (t, x)) + 1 6= 0. (1.7) For t = 0 the condition (1.7) is clearly satisfied. As such, there is a neighborhood Ω ⊆ [0, ∞) × R of the x-axis R so that U (t, x) exists for (t, x) ∈ Ω. However, if for some interval I ⊆ R u0 (x) < 0, x ∈ I (1.8) then for certain values t > 0, the condition (1.7) may fail, irrespective of the domain or degree of smoothness of u. It is well known that the violation of the condition (1.7) may imply that the classical solution U fails to exists for the respective values of t and x. That is, the domain of existence Ω of the solution U will be strictly contained in [0, ∞) × R. As such, for certain x ∈ R, the solution U (t, x) does not exist for sufficiently large t > 0 so that the equation (1.4) fails to have a classical solution on the whole of its domain of definition. From a physical point of view, however, it is exactly the points (t, x) ∈ ([0, ∞) × R)\ Ω where the solution fails to exist that are of interest, as these points may represent the appearance and propagation of what are called shock waves. Under rather general conditions, see for instance [96] and [143], it is possible to define certain generalized solutions U for all t ≥ 0 and x ∈ R, which turn out to be physically meaningful, and which are in fact classical solutions everywhere on [0, ∞)×R except a suitable set of points Γ ⊂ [0, ∞) × R, where Γ consists of certain families of curves called shock fronts. As a clarification of the above mentioned lack of global smoothness of the solutions to (1.4) and (1.5), let us consider the example if x ≤ 0 1 1 − x if 0 ≤ x ≤ 1 . u (x) = (1.9) 0 if x ≥ 1 In this case, the shock front Γ is given by 1) t ≥ 1 , Γ = (t, x) 2) x = t+1 2 while the solution U (t, x) is given by 1 if x ≤ 0 x−1 if 0 ≤ x ≤ 1 U (t, x) = t−1 0 if x ≥ 1 (1.10) (1.11) CHAPTER 1. A BRIEF HISTORY OF PDES 6 when 0 ≤ t ≤ 1 and U (t, x) = 1 if x < t+1 2 0 if x > t+1 2 (1.12) when t ≥ 1. For (t, x) ∈ Γ one may define U (t, x) at will. In this example, the failure of the initial value u to be sufficiently smooth at x = 0 and x = 1 does not in any way contribute to the nonexistence of a solution U (t, x) which is classical on the whole domain of definition of the equations. Rather, the lack of global smoothness of the solution U (t, x) is due to the fact that the initial condition u satisfies (1.8) on the interval (0, 1) ⊂ R. In view of the local nature of the existence results in Theorem 1 and Theorem 2, as well as the occurrence of singularities in the solutions of nonlinear PDEs as demonstrated in the above example concerning the nonlinear conservation law (1.4), it is clear that there is both a physical and theoretical interest in the existence of solutions to such systems of PDEs that may fail to be classical on the whole domain of definition of the respective system of PDEs. The interest in such generalized solutions to PDEs is the main motivation for the study of generalized functions, that is, objects which retain certain essential features of usual real valued functions. In this regard, there have so far been two main approaches to constructing suitable spaces of generalized functions which may contain suitable generalized solutions to systems of PDEs, namely, the sequential approach and the functional analytic approach. The sequential approach, introduced by S L Sobolev, see for instance [148] and [149], is based on the concept of a weak derivative, which had been applied by Hilbert, Courant, Riemann and several others in the study of various classes of ODEs and PDEs. In this regard, we recall that a Lebesgue measurable function u : Ω → R is square integrable on Ω whenever Z u (x)2 dx < ∞ (1.13) Ω where the integral is taken in the sense of Lebesgue. The set of all square integrable functions on Ω is denoted L2 (Ω), and carries the structure of a Hilbert space under the inner product Z hu, viL2 = v (x) u (x) dx. Ω For u ∈ L2 (Ω) and α ∈ Nn , a measurable function v : Ω → R is called a weak derivative Dα u of u whenever ∀ K ⊂ Ω compact : ∀ ϕ ∈ D (K) : R R u (x) Dα ϕ (x) dx = (−1)|α| K v (x) ϕ (x) dx K (1.14) CHAPTER 1. A BRIEF HISTORY OF PDES 7 where D (K) is the set of C ∞ -smooth functions ϕ on K such that the closure of the set {x ∈ K : ϕ (x) 6= 0} is compact and strictly contained in K, or some other suitable space of test functions. In this sequential approach, generalized solutions to a nonlinear, or linear, PDE T (x, D) u (x) = 0, x ∈ Ω ⊆ Rn (1.15) are obtained by constructing a sequence of approximating equations Ti (x, D) ui (x) = 0, x ∈ Ω, i ∈ N (1.16) where the operators Ti (x, D) are supposed to approximate T (x, D) in a prescribed way, so that each ui is a classical solution of (1.16), and the sequence (ui ) converges in a suitable weak sense to a function u, for instance Z (ui (x) − u (x)) ϕ (x) dx → 0 (1.17) Ω for suitable test functions ϕ. The weak limit u of the sequence (ui ) is interpreted as a generalized solution of (1.15). In this regard, Sobolev introduced the space H 2,m (Ω), with m ≥ 1, which is defined as ∀ |α| ≤ m : 2,m . (1.18) H (Ω) = u ∈ L2 (Ω) Dα f ∈ L2 (Ω) That is, the Sobolev space H 2,m (Ω) consists of all square integrable functions u on Ω with all weak partial derivatives Dα u up to order m in L2 (Ω). An inner product may be defined on H 2,m (Ω) through the formula X hu, vim = hDα u, Dα viL2 . (1.19) |α|≤m so that the Sobolev space H 2,m (Ω) is a Hilbert space. In particular, a sequence (ui ) in H 2,m (Ω) converges to u ∈ H 2,m (Ω) if and only if X Z (Dα u (x) − Dα ui (x))2 dx → 0 |α|≤m Ω which, in particular, implies that the sequence (ui ) converges weakly to u, in the sense that ∀ ϕ R : P∈ D (Ω) α |α|≤m Ω (u (x) − ui (x)) D ϕ (x) dx → 0 CHAPTER 1. A BRIEF HISTORY OF PDES 8 The definition of the Hilbert space structure on H 2,m (Ω) is the essential feature introduced by Sobolev. Indeed, the concept of a weak derivative, and that of weak solution, had been used by many authors prior to Sobolev. However, the main difficulty in applying the techniques of modern analysis, in particular those connected with function spaces and topological structures on such spaces, to ODEs and PDEs, respectively, is that the differential operators on such spaces are typically not continuous with respect to the mentioned topological structures. In the case of the Sobolev spaces H 2,m (Ω), and in view of (1.19), it is clear that ∀ |α| ≤ m : ∀ u ∈ H 2,m (Ω) : hDα u, Dα uiL2 ≤ hu, uim so that each such differential operator Dα : H 2,m (Ω) → L2 (Ω) (1.20) is continuous with respect to the inner products on H 2,m (Ω) and L2 (Ω), respectively. As such, the powerful tools of analysis, and in particular linear functional analysis, may be applied to the study of such PDEs which admit a suitable weak formulation in terms of H 2,m (Ω) or other associated spaces. At this point it is worth noting that this sequential approach has as of yet not received any suitable general theoretic treatment. Nevertheless, it has resulted in a wide range of effective, though somewhat ad hoc, solution methods for both linear and nonlinear PDEs, see for instance [100]. The second major approach to establishing generalized solutions to PDEs within a suitable framework of generalized functions, namely, the functional analytic approach introduced by L Schwartz [144], [145] in the late 1940s is based on the the idea of generalizing the concept of a weak derivative through the machinery of linear function analysis. In this regard, recall that any open subset Ω of Rn may be expressed as the union of a countable and increasing family of compact sets. That is, there is some family {Ki ⊂ Ω compact : i ∈ N} so that ∀ i∈N: Ki ⊂ Ki+1 and Ω= [ Ki i∈N For each i, j ∈ N so that i < j we may define the injective mapping fi,j : D (Ki ) → D (Kj ) CHAPTER 1. A BRIEF HISTORY OF PDES 9 through fi,j u : Kj 3 x 7→ u (x) if x ∈ Ki 0 . (1.21) if x ∈ / Ki Furthermore, and in the same way as in (1.21), for each i ∈ N we may define the mapping fi : D (Ki ) → D (Ω) where ∃ K ⊂ Ω compact : ∞ D (Ω) = ϕ ∈ C (Ω) ϕ ∈ D (K) so that the diagram fi,j - D (Kj ) D (Ki ) @ @ @ @ fj @ fj (1.22) @ @ @ @ R @ D (Ω) commutes whenever i < j. As such, and in view of the injectivity of the mappings fi,j and fi , we may express D (Ω) as the strict inductive limit of the inductive system (D (Ki ) , fi,j )i,j∈N . Each of the spaces D (Ki ) carries in a natural way the structure of a Frechét space. Indeed, the family of seminorms {ρα }α∈Nn defined through ρα : D (Ki ) 3 u 7→ sup{|Dα u (x) | : x ∈ Ki }, where k·kK denotes the uniform norm on C 0 (K), defines a metrizable locally convex topology τi on D (Ki ). As such, and in view of the construction of D (Ω) as the strict inductive limit of the inductive system (D (Ki ) , fij ), the space D (Ω) may be equipped with the locally convex strict inductive limit of the family of Fréchet spaces (D (Ki ) , τi )i∈N . An intuitive feeling for this topology on D (Ω) may be obtained by considering convergent sequences. A sequence (ϕn ) in D (Ω) converges to ϕ ∈ D (Ω) if and only if ∃ K ⊂⊂ Ω : ∀ n∈N: suppϕn ⊆ K CHAPTER 1. A BRIEF HISTORY OF PDES 10 and ∀ α∈N: kDα ϕ − Dα ϕn kK → 0 The concept of weak derivative, such as in the sense of Sobolev [148], [149], is incorporated in the above functional analytic setting by associating with each u ∈ L2 (Ω) a continuous linear functional Tu : D (Ω) → R through Z Tu : ϕ 7→ u (x) ϕ (x) dx. (1.23) Ω As such, one obtains the inclusion 1) T linear 0 L2 (Ω) ⊆ D (Ω) = T : D (Ω) → R 2) T continuous In particular, if u ∈ H 2,m (Ω), then for each |α| ≤ m, we have Z Z |α| α TDα u : ϕ 7→ D u (x) ϕ (x) dx = (−1) u (x) Dα ϕ (x) dx. Ω (1.24) Ω Identifying with each u ∈ H 2,m (Ω) and every |α| ≤ m the functional Dα Tu ∈ D0 (Ω) which is defined as Dα Tu : ϕ 7→ TDα u ϕ (1.25) it is clear that D0 (Ω) also contains each weak partial derivative, up to order m, of functions in H 2,m (Ω). Generalizing the formula (1.24) to arbitrary continuous linear functionals in D0 (Ω), one may define generalized partial derivatives in D0 (Ω) of all orders for each T ∈ D0 (Ω) through Dα T : D (Ω) 3 ϕ 7→ (−1)|α| T (Dα ϕ) ∈ R (1.26) Not each continuous linear functional on D (Ω) can be described through (1.23). Indeed, suppose that Ω = Rn , and consider the linear functional δ, called the Dirac distribution, defined as δ : D (Rn ) 3 ϕ 7→ ϕ (0) ∈ R. (1.27) Clearly (1.27) defines a continuous linear functional on D (Rn ). However, there is no locally integrable function u on Rn so that Z δ : ϕ 7→ u (x) ϕ (x) dx. Ω CHAPTER 1. A BRIEF HISTORY OF PDES 11 In this regard, for a > 0, consider the function ϕa ∈ D (Rn ) which is defined as − 1 e 1−|x|/a if |x| < a ϕa (x) = 0 if |x| ≥ a Suppose that there is an integrable function u on Ω that defines δ through (1.23). For each a > 0 we have Z −1 e = δ (ϕa ) = u (x) ϕa (x) dx → 0 Rn which is absurd. The space D0 (Ω) of continuous linear functionals on D (Ω) is called the space of distributions on Ω and is denoted D0 (Ω). In view of the above example involving the Dirac distribution, it is clear that D0 (Ω) contains not only each of the Sobolev spaces H 2,m (Ω), for m ≥ 1, but also more general generalized functions. Indeed, every locally integrable function u, that is, every Lebesgue measurable function that satisfies ∀ K R ⊂ Ω compact : |u (x) |dx < ∞ K may be associated with a suitable element Tu of D0 (Ω) in a canonical way. The Schwartz linear theory of distributions has a rather natural position within the context of spaces of generalized functions which contain C 0 (Ω). In this regard, we should note that the main objective of the linear functional analytic approach of Schwartz [144] is the infinite differentiability of generalized functions, something that the sequential approach of Sobolev [148], [149] fails to achieve. In this regard, and in view of (1.26), it is clear that this aim is achieved within the setting of D0 distributions. In fact, from the point of view of the existence of partial derivatives, the space D0 (Ω) posses a canonical structure. Indeed, in the chain of inclusions C ∞ (Ω) ⊂ ... ⊂ C l (Ω) ⊂ ... ⊂ C 0 (Ω) ⊂ D0 (Ω) only C ∞ (Ω) is closed under arbitrary partial derivatives in the classical sense. However, identifying u ∈ C 0 (Ω) with Tu ∈ D0 (Ω) through (1.23), we may again perform indefinite partial differentiation on u, with the partial derivatives defined as in (1.26), which, however, are no longer classical. Obviously, in view of (1.24) and (1.25), if u ∈ C l (Ω), then for |α| ≤ l, the partial derivative Dα u is the classical one, that is, the weak and classical derivatives coincide for sufficiently smooth functions. The mentioned canonical structure of D0 (Ω) is the following: ∀ T ∈ D0 (Rn ) , K ⊂ Ω compact : ∃ u ∈ C 0 (Ω) , α ∈ Nn : T|K = Dα u|K (1.28) CHAPTER 1. A BRIEF HISTORY OF PDES 12 Here Dα is the weak partial derivative (1.26). In other words, D0 is a minimal extension of C 0 in the sense that locally, every distribution is the weak partial derivative of a continuous function. The linear theory of distributions, as shortly described above, as well as certain generalizations of it, see for instance [75], has proved to be a powerful tool in the study of PDEs, in particular in the case of linear, constant coefficient equations. Indeed, in view of the fact that D0 (Ω), as the dual of the locally convex space D (Ω), is a vector space with the usual operations, and since D0 (Ω) contains C ∞ (Ω) as a dense subspace, each constant coefficient linear partial differential operator X P (D) : C ∞ (Ω) 3 u 7→ aα Dα u ∈ C ∞ (Ω) (1.29) |α|≤m may be extended to the larger space D0 (Ω) through X P (D) : D0 (Ω) 3 T 7→ aα Dα T ∈ D0 (Ω) |α|≤m In this regard, the first major result is due to Ehrenpreis [52] and Malgrange [103]. Namely, for each linear, constant coefficient partial differential operator (1.29), the generalized equation P (D) T = δ admits a solution T ∈ D0 (Ω). From this it follows that, for any ϕ ∈ D (Ω), the equation P (D) u (x) = ϕ (x) , x ∈ Ω with P (D) defined as in (1.29), has a solution in D0 (Ω). This result alone justifies the use of the D0 -distributions in the study of linear, constant coefficient PDEs, and it has a variety of useful consequences and applications, see for instance [62] and [71]. In spite of the above mentioned canonical structure of the D0 -distributions in terms of partial differentiability, as well as the power of the linear theory of distributions in the context of linear, constant coefficient PDEs, the Schwartz distributions suffer from two major weaknesses. In the first place, we note that, for each u ∈ C ∞ (Ω) and each T ∈ D0 (Ω), we can define the product of u and T in D0 (Ω) as u × T : ϕ 7→ T (u × ϕ) (1.30) That is, each distribution T ∈ D0 (Ω) can be multiplied with any C ∞ -smooth function u. As such, and in view of the extension of the differential operators (1.26), every linear partial differential operator, say of order m, of the form X P (D) u (x) = aα (x) Dα u (x) |α|≤m CHAPTER 1. A BRIEF HISTORY OF PDES 13 where each coefficient aα is C ∞ -smooth, may be extended to the space D0 (Ω) of distributions on Ω. Indeed, in view of the linearity of the operator P (D), we may, for any T ∈ D0 (Ω), define the distribution P (D) T as X P (D) T : ϕ 7→ aα Dα T (ϕ) |α|≤m As such, the partial differential equation P (D) u (x) = g (x) , with g ∈ C ∞ (Ω), defined by the operator P (D), may be extended to a generalized equation in terms of distributions P (D) T = Tg (1.31) where Tg is the distribution associated with the C ∞ -smooth function g. Whenever the coefficient functions aα are constant, that is, ∀ |α| ≤ m : ∃ aα ∈ R : , aα (x) = aα , x ∈ Ω and the righthand term g has compact support, the generalized equation (1.31) admits a solution in D0 (Ω). Moreover, the existence of a solution holds also if the right hand term in (1.31) is any distribution with compact support. However, this result cannot be generalized to equations with nonconstant coefficients. In this regard, we may recall Lewy’s impossibility result [97], see also [88]. Lewy showed that for a large class of functions f1 , f2 ∈ C ∞ (R3 ), the system of first order linear PDEs − ∂x∂ 1 U1 + ∂ U ∂x2 2 − ∂x∂ 1 U2 ∂ U ∂x2 1 − 2x1 ∂x∂ 3 U2 − 2x2 ∂x∂ 3 U1 = f1 , + + 2x1 ∂x∂ 3 U1 − 2x2 ∂x∂ 3 U2 (1.32) = f2 which may be written as a single equation with complex coefficients, has no distributional solutions in any neighborhood of any point of R3 . The most interesting aspect of the example (1.32) is that it is not the typical, esoteric, counterexample type of equation, but appears rather naturally in connection with the theory of functions of several complex variables, see for instance [88]. In view of the rather natural, not to mention simple, form of Lewy’s example, and as will be elaborated upon further in the sequel, the Schwartz distributions prove to be insufficient from this point of view for large classes of equations. Over and above the mere insufficiency of the Schwartz distributions from the point of view of the existence of generalized solutions to systems of PDEs, the space D0 (Ω) suffers from serious structural deficiencies. In particular, and since CHAPTER 1. A BRIEF HISTORY OF PDES 14 the early 1950s, it is known that D0 (Ω) does not admit any reasonable concept of multiplication that extends the usual pointwise multiplication of smooth functions. Indeed, Schwartz [145], see also [140], proved the following result. Let A be an associative algebra so that C 0 (R) ⊂ A, and uv is the usual product of functions for each u, v ∈ C 0 (R). If D : A → A is a differentiation operator, that is, D is linear and satisfies the Leibnitz rule for product derivatives, so that D restricted to C 1 (Ω) ⊂ A is the usual differentiation operation, then there is no δ ∈ A, δ 6= 0, so that xδ = 0. (1.33) This result is usually interpreted as follows. If δ ∈ D0 (R) is the Dirac delta distribution, then, in view of (1.30), for any u ∈ C ∞ (R) we have uδ : D (R) 3 ϕ 7→ δ (uϕ) = u (0) ϕ (x) (1.34) Therefore, if we let x ∈ C ∞ (R) denote the identity function on R, then from (1.34) it follows that xδ : D (R) 3 ϕ 7→ δ (xϕ) = 0. That is, xδ is the additive identity in D0 (R). Therefore, in view of (1.33), it follows that δ = 0. But δ 6= 0 ∈ D0 (R), and hence there cannot be a suitable multiplication on D0 (R). In particular, this has the effect that one cannot formulate the concept of solution to nonlinear PDEs, such as (1.4) for instance, in the framework of the D0 distributions. In view of the above impossibility, in order to multiply arbitrary distributions in a consistent and meaningful way, it is necessary to find an embedding D0 (R) ,→ A (1.35) where A is a suitable algebra wherein multiplication may be performed. However, in this regard, there typically occur misinterpretations of the mentioned Schwartz impossibility. Indeed, Schwartz’s result is usually interpreted as stating that there cannot exists convenient algebras A such as in (1.35), and that a convenient multiplication of arbitrary distributions is not possible [71]. Furthermore, in view of the perceived impossibility of multiplying distributions in a convenient and meaningful way, it is widely believed that there can not be a general and convenient nonlinear theory of generalized functions. In particular, one cannot hope to obtain any significantly general and type independent theory for generalized solutions of nonlinear PDEs. It will be shown over and again in the sequel that this is in fact a misunderstanding, and suitable nonlinear theories of generalized functions can easily be constructed, see for instance [135] through [142], theories which deliver general existence and regularity results for large classes of systems of nonlinear PDEs. CHAPTER 1. A BRIEF HISTORY OF PDES 15 Over and above the mentioned deficiencies of the Schwartz linear theory of distributions, both from the point of view of existence of generalized solutions to PDEs, as well as in terms of its capacity to handle also nonlinear problems, the space D0 (Ω) suffers from several other serious weaknesses. In this regard, we mention only that D0 (Ω) fails to be a flabby sheaf [140], a property that is fundamental in connection with the study of singularities, and that the use of the D0 distributions is not convenient, from the point of view of exactness, to deal with sequential solutions, even to linear PDEs [140]. And now, taking into account the various weaknesses and deficiencies of the D0 distributions, when it comes to the study of generalized solutions to both linear and nonlinear PDEs, there appears to be only two possible ways forward. In the first place, and as is commonly believed, it may seem that there cannot be a general and convenient framework for such generalized solutions, and instead various ad hoc methods may be applied to different equations. On the other hand, a suitable extension of the theory of distributions may be pursued, such as may be provided by suitable embeddings of D0 (Ω) into convenient algebras of generalized functions, such as in (1.35). These two alternatives have been pursued for the last five decades, as will be explained in the subsequent sections. However, there is a third, if largely overlooked, possibility in pursuing a systematic account of generalized solutions to linear and nonlinear PDEs. In view of the above mentioned difficulties presented by the Schwartz distributions, in particular in connection with nonlinear PDEs, why should any theory of generalized solutions to systems of PDEs be restricted by the requirement that it must contain the D0 distributions as a particular case? Indeed, one may start from the very beginning and ask what exactly are the requirements of such a theory? In this regard, there have lately been two independent attempts at such a completely new theory of generalized solutions to systems of PDEs, namely, the Order Completion Method [119], and the Central Theory for PDEs [115] through [118]. Both these theories, although approaching the subject of generalized solutions to PDEs from rather different points of view, have delivered general and type independent existence and regularity results for generalized solutions of large classes of systems of nonlinear PDEs. 1.2 Weak Solution Methods In view of the deficiencies of the linear theory of D0 -distributions, in particular the impossibility of defining nonlinear operations on D0 (Ω), and the insufficiency of the D0 framework of generalized functions in the context of the existence of generalized solutions of PDEs, it is widely held that there cannot be a convenient nonlinear theory of generalized functions and, moreover, a general and type independent theory for the existence and regularity of generalized solutions of systems of nonlinear PDEs is impossible [12]. CHAPTER 1. A BRIEF HISTORY OF PDES 16 In view of the above remarks concerning the failures of the linear theory of distributions, rather than pursuing the essential features that are at work in regards to the existence of generalized solutions to linear and nonlinear PDEs, perhaps within a context other than the usual functional analytic one, the typical approach to the problems of existence and regularity of generalized solutions of nonlinear PDEs consists of a collection of rather ad hoc methods, each developed with a particular equation, or at best a particular type of equation, in mind. At this point it is worth noting that, ever since Sobolev [148], [149], the main, and to a large extent even exclusive, approach to solving linear and nonlinear PDEs has been that of functional analysis. In this regard, and as mentioned above, this approach consists of a collection of ad hoc methods, each applying to but a rather small class of equations. Furthermore, in particular in the case of nonlinear equations, this often leads to ill founded concepts of a solution of such equations. In this section we will briefly discuss the general framework of the more popular such methods for solving linear and nonlinear systems of equations, namely, the mentioned weak solution methods, see for instance [54] or [99], [100]. Furthermore, those particular difficulties that arise when applying such methods to nonlinear equations will be indicated [140]. In this regard, let us, at first, consider a linear partial differential equation of order m L (D) u (x) = f (x) , x ∈ Ω (1.36) where f : Ω → R is, say, a continuous function, and the partial differential operator is of the form X L (D) u (x) = aα (x) Dα u (x) , (1.37) |α|≤m where aα : Ω → R are sufficiently smooth functions, for instance aα ∈ C 0 (Ω). With the partial differential operator L (D) one associates a mapping L:A→B (1.38) where A is a vector space of sufficiently smooth functions on Ω, such as A = C m (Ω), and B is an appropriate linear space of functions which contains the righthand term f. The essential idea behind the so called weak solution methods is to construct an infinite sequence of partial differential operators Li (D) which approximate (1.37) in a suitable sense, so that each of the infinite sequence of PDEs Li (D) ui (x) = f (x) , x ∈ Ω (1.39) admits a classical solution ui ∈ A, i ∈ N (1.40) CHAPTER 1. A BRIEF HISTORY OF PDES 17 With a suitable choice for the approximate partial differential operators {Li : i ∈ N} in (1.39), and solutions {ui : i ∈ N} to (1.39), as well as appropriate linear space topologies on A and B, one has (Lui ) converges to f in B. (1.41) Furthermore, using some compactness, monotonicity or fixed point argument, one may often extract a Cauchy sequence from the sequence (ui ) ⊂ A, which we denote by (ui ) as well. Thus, we have now obtained a sequence (ui ) in A so that (ui ) converges to u] ∈ A] (1.42) (Lui ) converges to f ∈ B, (1.43) and where A] denotes the completion of A in its given vector space topology. Now, in view of (1.42) and (1.43), u] ∈ A] is considered a generalized solution to (1.36). Note that, in view of the typical difficulties involved in the steps (1.39) to (1.43), in particular when initial or boundary value problems are associated with the PDE (1.36), one ends up with very few Cauchy sequences in (1.42), if in fact not with a single such sequence. Furthermore, based solely on the very few, if in fact not the single Cauchy sequence in (1.42) and (1.43), the partial differential operator (1.37) is extended to a mapping L (D) : {u] } ∪ A → B. (1.44) As such, this customary method for finding generalized solutions to PDEs amounts to nothing but an ad hoc, pointwise extension of the partial differential operator L (D). The deficiency of the above solution method is clear. Indeed, the extension (1.44) of the partial differential operator L (D) is based on only very few Cauchy sequence (1.42) in A. Moreover, the sequence in (1.42), and therefore also the generalized solution u] of (1.36), is often obtained by an arbitrary subsequence selection from (1.40). In the case of linear PDEs such as (1.36), the rather objectionable construction of a generalized solution to such an equation may, to some extend, be justified. Indeed, in this case, with the coefficients aα in (1.37) sufficiently smooth, it happens that due to the phenomenon of automatic continuity of certain classes of linear operators, the above construction of a generalized solution to (1.36) is valid. Indeed, suppose that we obtain a Cauchy sequence (ui ) in A so that (ui ) converges to u] ∈ A] (1.45) (L (D) ui ) converges to f ∈ B. (1.46) and CHAPTER 1. A BRIEF HISTORY OF PDES 18 Given now any sequence (vi ) in A so that (vi ) converges to 0 ∈ A we may define the Cauchy sequence in A (wi ) = (ui + vi ) . From the linearity of the operator L (D) we now obtain (L (D) wi ) = (L (D) ui ) + (L (D) vi ) (1.47) If, as most often happens to be the case, the mapping L (D) is continuous, we have (L (D) vi ) converges to 0 ∈ B. (1.48) Now, in view of (1.47) and (1.48), it follows that (L (D) wi ) converges to f ∈ B Therefore, based solely on the single Cauchy sequence in (1.45) and (1.46), we have ∀ (ui ) ⊂ A : . (ui ) converges to u] ∈ A] ⇒ (L (D) ui ) converges to f (1.49) The relationships (1.47) to (1.49) affirms the interpretation of u] ∈ A] as a generalized solution to (1.36). In contradistinction with the case of linear PDEs, when the procedure (1.39) through (1.43) for establishing the existence of weak solutions is applied to a nonlinear PDE T (D) u (x) = f (x) , x ∈ Ω (1.50) one critical point is often overlooked, namely, the nonlinear operator T (D) is typically not compatible with the vector space topologies on A and B. In this case the claim that u] ∈ A] is a generalized solution to (1.50) is rather objectionable. Indeed, in the case of a nonlinear PDE (1.50) there is in fact a double breakdown in (1.45) to (1.49). In this regard, note that in case the linear partial differential operator L (D) in (1.36) is replaced with a nonlinear operator, such as in (1.50), both the crucial steps in (1.47) and (1.48) will in general break down. In that case, then, we cannot in general deduce from very few, if not in fact one single Cauchy sequence (ui ) ⊂ A such that (ui ) converges to u] ∈ A] (1.51) (T (D) ui ) converges to f ∈ B, (1.52) and CHAPTER 1. A BRIEF HISTORY OF PDES 19 that ∀ (ui ) ⊂ A : . (ui ) converges to u] ∈ A] ⇒ (L (D) ui ) converges to f In view of this double breakdown in the customary weak solution methods, when applied to nonlinear problems, it is clear that such methods are typically ill founded when applied to such nonlinear problems. It is exactly this double breakdown which, for a long time, was usually overlooked when applying solution methods for linear PDEs to nonlinear ones. In part, this oversight is perhaps due to the fact that, in the particular case of linear PDEs, the method (1.39) to (1.43) happens to be correct. However, and in view of the above remarks, it is clear that the extension of linear methods to nonlinear problems often require essentially new ways of thinking. For an excellent survey of the difficulties of several such well known extensions can be found in [140] or [165]. The careless application of essentially linear methods such as in (1.39) to (1.43) to nonlinear problems can, and in fact often does, lead to absurd conclusions, see for instance [140]. In this regard, a most simple example is given by the zero order nonlinear system of equations u = 0 (1.53) u 2 = 1 which, using the customary weak solution method (1.39) to (1.43), admit both weak and strong solutions, so that, apparently, we have proven the blatant absurdity 0 = 1 in R. Indeed, if we set A = C ∞ (R) with the topology induced by D0 (R), and we take B = D0 (R), then the sequence (vi ) defined as √ vi (x) = 2 cos (ix) , i ∈ N is Cauchy in A and (vi ) converges to u = 0 both weakly and strongly in A] = D0 (R). Furthermore, we shall also have vi2 converges to v = 1 both weakly and strongly in B = D0 (R). That is, the sequences (vi ) and (vi2 ) converge to their respective limits with respect to the usual topology on D0 and ∀ ϕ ∈ D (R) : 1) Tvi (ϕ) → 0 in R . 2) Tvi2 (ϕ) → 1 in R (1.54) CHAPTER 1. A BRIEF HISTORY OF PDES 20 As such, according to the typical weak solution method (1.39) to (1.43), the sequence (vi ) defines both a weak and strong solution to (1.53). Let us now take a closer look at how the above nonlinear stability paradox, namely, the existence of both weak and strong solutions to (1.53), effects the customary sequential approach (1.39) to (1.43). In this regard, consider the nonlinear partial differential operator given by T (D) u = L (D) u + u2 where L (D) is a linear partial differential operator such as in (1.37). Suppose that the system (1.53) has a solution in A, that is, ∃ (vi ) ⊂ A a Cauchy sequence : 1) (vi ) converges to 0 in A . 2) (vi2 ) converges to 1 in B (1.55) Let (ui ) be the Cauchy sequence in (1.51) and (1.52). Define the Cauchy sequence (wi ) in A as wi = ui + λvi , i ∈ N, for an arbitrary but fixed λ ∈ R. In view of (1.51) and (1.55) we have (wi ) converges to u] ∈ A] . (1.56) Hence the sequence (wi ) defines the same generalized function as (ui ) in (1.51). Now in view of (1.52) it follows that T (D) wi = T (D) ui + λL (D) vi + 2λui vi + λ2 vi2 . Hence, assuming that (ui vi ) is a Cauchy sequence in A with limit v ] ∈ A] , (1.52) and (1.55) yield (T (D) wi ) converges to f + λ2 + 2λv ] in B (1.57) if, as it often happens, (L (D) vi ) converges to 0 in B, such as for instance in the topology of D0 (Ω) when L (D) has C ∞ -smooth coefficients. Since λ ∈ R is arbitrary, (1.56) and (1.57) yields T (D) u] 6= f (1.58) In this way, the same u] ∈ A] that is a generalized solution to the equation T (D) u = f according to the customary interpretation of (1.51) and (1.52), now, in view of (1.58), no longer happens to be a solution to this equation. In view of the above example, it is clear that, as far as nonlinear partial differential equations are concerned, the extension of the concept of classical solution CHAPTER 1. A BRIEF HISTORY OF PDES 21 to that of the concept of generalized solution along the lines (1.39) to (1.43), is an improper generalization of various classical extensions, such as for instance the extension of the rational numbers into the real numbers. Over the last thirty years or so, there has been a limited awareness among the functional analytic school concerning the difficulties involved in extending linear methods to nonlinear problems, see for instance [14], [44], [48], [127], [147] and [151]. However, the techniques developed to overcome these difficulties, such as the Tartar-Murat compensated compactness and the Young measure associated with weakly convergent sequences of functions subject to differential constraints on an algebraic manifold, can deal only with particular types of nonlinear PDEs and sequential solutions. The effect of this limited approach is an obfuscation of the basic underlying reasons, whether these are of an algebraic or topological nature, for such problems as the nonlinear stability paradox discussed in this section. Not to mention that there is no attempt to develop a systematic nonlinear theory of generalized functions that would be able to accommodate large classes of nonlinear PDEs. 1.3 Differential Algebras of Generalized Functions As we have mentioned at the end of Section 1.1, and as demonstrated in Section 1.2, in order to solve large classes of linear and nonlinear PDEs, it is necessary to go beyond the usual functional analytic methods for PDEs, including the Schwartz D0 distributions. In particular, the Schwartz impossibility result places serious restrictions on use of distributions in the study of nonlinear PDEs. As such, in order to obtain a theory of generalized functions that would be able to handle large classes of nonlinear PDEs, and at the same time contain the D0 distributions, the underlying reasons for the mentioned Schwartz impossibility result, reasons that turn out to be rather algebraic than topological [140], should be clearly understood. In this regard [140], the difficulties involved in establishing a suitable framework of generalized functions for the solutions of nonlinear partial differential equations may be viewed as a consequence of a basic algebraic conflict between the trio of discontinuity, multiplication and differentiation. In order to illustrate how such a conflict might arise, we consider the most simple discontinuous function, namely, the Heaviside function H : R → R given by 0 if x ≤ 0 H (x) = (1.59) 1 if x > 0 When a discontinuous function such as H in (1.59) appears as a solution to a nonlinear PDEs, this function will necessarily be subjected to the operations of differentiation and multiplication. As such, a natural and intuitive setting for a theory of generalized functions that would include, in particular, the discontinuous function in (1.59), would be a ring of functions A ⊆ {u : R → R} (1.60) CHAPTER 1. A BRIEF HISTORY OF PDES 22 so that H ∈ A. (1.61) Furthermore, there should be a differential operator D:A→A (1.62) defined on A. That is, D is a linear operator that satisfies the Leibniz rule for product derivatives D (uv) = uDv + vDu (1.63) Already within this basic setup (1.60) through (1.63) we encounter rather surprising and undesirable consequences. Indeed, in view of (1.59) and (1.60), it follows that ∀ m ∈ N, m ≥ 1 : Hm = H (1.64) Furthermore, it follows from (1.60) that A is both associative and commutative. Hence (1.63) and (1.64) implies the relation ∀ m ∈ N, m ≥ 2 : . mH · DH = DH (1.65) From (1.65) it now follows that ∀ p, q ∈ N, p, q ≥ 2: p 6= q ⇒ p1 − 1q DH = 0 ∈ A which implies DH = 0 ∈ A. (1.66) However, in view of the fact that our theory should contain the D0 distributions we have that D0 (R) ⊂ A (1.67) and the differential operator D on A extends the distributional derivative. In particular, it follows from (1.67) that δ∈A where δ is the Dirac distribution. Furthermore, since H ∈ D0 (R) and DH = δ ∈ D0 (R) , CHAPTER 1. A BRIEF HISTORY OF PDES 23 it follows by (1.66) that δ=0∈A (1.68) which is of course false. It is now obvious that if we wish to define nonlinear operations, in particular unrestricted multiplication, on generalized functions that contain the D0 distributions in a consistent and useful way, some of the assumptions (1.60) to (1.63) must be relaxed. This can be done in any of several different ways. Indeed, while the algebra A should contain functions such as the Heaviside function H in (1.59), it need not be an algebra of functions from R to R. That is, A may contain elements that are more general than such functions u : R → R. Furthermore, the multiplication in A need not be so closely related to multiplication of real valued functions. As such, (1.60) need not necessarily hold. Regarding the differentiation operator D in (1.63), it is important to note that the requirement (1.62) is of highly restrictive assumption. Indeed, (1.62) implies that each u ∈ A is indefinitely differentiable. That is, ∀ m ∈ N, m ≥ 1 : ∀ u∈A: . Dm u ∈ A Of course, this is the case if A ⊆ C ∞ (R), which, in view of (1.61), is not possible. As such, one may also want to keep in mind the possibility that the differential operator may rather be defined as D:A→A where A is another algebra of generalized functions. In this case, the Leibniz Rule (1.63) may be preserved in this more general situation by assuming the existence of an algebra homomorphism A 3 u 7→ u ∈ A and rewriting (1.63) as D (u · v) = (Du) · v + u · (Dv) (1.69) where the product on the left of (1.69) is taken in A, and the product on the right is taken in A. The above two relaxations, namely, on the algebra A and the derivative operator D are sufficient in order to obtain generalized functions which extend the D0 distributions and admit generalized solutions to large classes of linear and nonlinear PDEs. Furthermore, the arguments leading to (1.66) are purely of an algebraic nature, and do not involve calculus or topology. This is precisely the reason for CHAPTER 1. A BRIEF HISTORY OF PDES 24 the power and usefulness of the so called ‘algebra first’ approach, [140]. We will shortly describe such an approach to constructing generalized functions, initiated by Rosinger [135], [136] and developed further in [137], [138] and [140]. In this regard, it is helpful to recall that generalized solutions to linear PDEs are typically constructed as elements of the completion of a suitably chosen locally convex topological, in particular metrizable, vector space A of sufficiently smooth functions u : Ω → R, see Section 1.2. From a more abstract point of view, the completion A] of the metrizable topological vector space A may be constructed as A] = S/V (1.70) V ⊂ S ⊂ AN (1.71) where with AN the set of sequences in A and S is the set of all Cauchy sequences in A. With termwise operations on sequences, AN is in a natural way a vector space, while S and V are suitable vector subspaces of AN . Therefore, the quotient space S/V is again a vector space. In the case of nonlinear PDEs, we shall instead be interested in a suitable algebra of generalized functions. In this regard, the above construction in (1.70) to (1.71) may be adapted for that purpose. Indeed, we may choose a suitable subalgebra of smooth functions A ⊆ C m (Ω) (1.72) I ⊂ S ⊂ AN (1.73) and where S is a subalgebra in AN , while I is an ideal in S. Then the quotient algebra A] = S/I (1.74) can offer the representation for our algebra of generalized functions. Furthermore, when A ⊆ C ∞ (Ω) , then for suitable choices of the subalgebra S ⊂ AN and the ideal I ⊂ S in (1.72) to (1.74), there is a vector space embedding D0 (Ω) → A] (1.75) However, in general the embedding (1.75) cannot be achieved in a canonical way, as can be seen by straightforward ring theoretic arguments connected with the existence CHAPTER 1. A BRIEF HISTORY OF PDES 25 of maximal off diagonal ideals [140]. Colombeau [39], [40] constructed such an algebra of generalized functions that admits a canonical embedding (1.75). Now it should be noted that, in view of (1.72), the algebra A is automatically both associative and commutative. Therefore, with the termwise operations on sequences, AN is also associative and commutative so that the algebra A] of generalized functions in (1.74) will also be associative and commutative. At first glance it may appear that the construction in (1.72) to (1.74) is rather arbitrary. However, see for instance [140], such concerns may be addressed and clarified to a good extent. In particular, the following points may be noted. First of all, it is a natural condition to impose on the differential algebra A that A ⊂ A] . (1.76) In particular, this addresses the issue of consistency of usual classical solutions to a nonlinear PDEs with generalized solutions in A] . In this regard, the purely algebraic neutrix condition will particularize the above framework (1.72) to (1.74) so as to incorporate (1.76). This purely algebraic condition characterizes the requirement (1.76), yet it has proven to be surprisingly powerful. In this regard, recall that in the case of a general vector space A of usual functions u : Ω → R we have taken vector subspaces V ⊂ S ⊂ AN and have defined A] = S/V as a space whose elements u] ∈ A] generalize the classical functions u ∈ A. As such, one should have a vector space embedding A ⊂ A] = S/V (1.77) which is defined by a linear injection ιA : A 3 u 7→ ιA (u) = (u) + V ∈ A] (1.78) where (u) is the constant sequence with all terms equal to u ∈ A. We may reformulate (1.77) to (1.78) in the following more convenient way. Let O ⊂ AN denote the null vector subspace, that is, the subspace consisting only of the constant zero sequence, and let UA,N = {ιA (u) : u ∈ A} (1.79) be the vector subspace in AN consisting of all constant sequences in A. That is, UA,N is the diagonal in the cartesian product AN . Then (1.77) to (1.78) is equivalent CHAPTER 1. A BRIEF HISTORY OF PDES 26 to the commutative diagram ⊂ V ⊂ - S 6 - AN 6 ⊂ ⊂ O ⊂ - (1.80) UA,N together with the co called off diagonal condition V ∩ UA,N = O (1.81) which is called the neutrix condition [140], the name being suggested by similar ideas introduced in [153] within a so called ‘neutrix calculus’ developed in connection with asymptotic analysis. Following the terminology in [153], see also [140], a sequence (ui ) ∈ V is called V-negligible. In this sense, for two functions u, v ∈ A, their difference u − v ∈ A is V-negligible if and only if ιA (u) − ιA (v) = ιA (u − v) ∈ V, which in view of (1.81) is equivalent to u = v. As such, the neutrix condition (1.81) simply means that the quotient structure S/V distinguishes between classical functions in A. Now, as mentioned, in the case of nonlinear PDEs we may be interested in constructing algebras of generalized functions such as is done in (1.72) to (1.74). In particular, (1.79) to (1.81) may be reproduced in this setting, given an algebra A of sufficiently smooth functions u : Ω → R, a subalgebra S of AN , and an ideal I ⊂ S, such that the inclusion diagram ⊂ I ⊂ - S 6 - AN 6 ⊂ ⊂ O ⊂ - UA,N (1.82) CHAPTER 1. A BRIEF HISTORY OF PDES 27 satisfies the off diagonal or neutrix condition I ∩ UA,N = O. (1.83) It follows that, similar to (1.77) through (1.78), for every quotient algebra A] = S/I we have the algebra embedding ιA : A 3 u 7→ ιA (u) = (u) + I ∈ A] (1.84) Furthermore, the conditions (1.80) to (1.83) are again necessary and sufficient for (1.84). The neutrix condition, although of a simple and purely algebraic nature, turns out to be highly important and powerful in the study of such algebras of generalized functions. In particular, variants of this condition characterize the existence and structure of so called chains of differential algebras [138], [140]. Furthermore, the neutrix condition determines the structure of ideals I which play an important role in the stability, exactness and generality properties of the algebras of generalized functions. The algebras of generalized functions A] constructed in (1.72) to (1.74) may be further particularized by introducing the natural requirement that nonlinear partial differential operators should be extendable to such algebras. In this regard, let us consider a polynomial type nonlinear partial differential operator ∞ T (D) : C (Ω) 3 u 7→ h X i=1 ci ki Y Dpij u ∈ C ∞ (Ω) (1.85) j=1 where for 1 ≤ i ≤ h we have ci ∈ C ∞ (Ω). For convenience, we shall consider the problem of extending the partial differential operator (1.85) to a differential algebra P (D) : A] → A] (1.86) where the original algebra of classical functions satisfies A ⊆ C ∞ (Ω) . (1.87) In this regard, we note that in order to obtain an extension (1.86), it is sufficient to extend the usual partial differential operators to mappings D p : A ] → A] , p ∈ N n . (1.88) This can easily be done by making the assumption ∀ p ∈ Nn : 1) Dp I ⊂ I 2) Dp S ⊂ A (1.89) CHAPTER 1. A BRIEF HISTORY OF PDES 28 In this case (1.86) can be defined as Dp U = Dp s + I ∈ A] = A/I, p ∈ Nn (1.90) for each U = s + I ∈ A] = S/I where for every sequence s = (un ) ∈ AN we define Dp s = (Dp un ) , p ∈ Nn . (1.91) The extension (1.86) of the nonlinear partial differential operator (1.85) can now be obtained as follows. For a given U = s + I ∈ A] = S/I, s = (un ) ∈ S we define T (D) U = T (D) t + I ∈ S/I = A] where t = (vn ) ∈ S, t − s ∈ I. (1.92) The construction (1.88) to (1.92) of the extension (1.88) may be replicated within a far more general setting. Indeed, the nonlinear operations, such as multiplication, are performed only in the range of the partial differential operator (1.85). As such, only the range need be an algebra, while we are more free in choosing the domain. In particular, we can replace the extensions (1.88) with D p : E ] → A ] , p ∈ Nn where E ] is a suitable vector space of generalized functions constructed in (1.77) to (1.78). Furthermore, since the partial differential operator (1.85) is of finite order, say m, one may relax the condition (1.87) by only requiring A ⊆ C m (Ω) while the assumption (1.89) must now obviously be replaced with ∀ p ∈ Nn , |p| ≤ m : 1) Dp I ⊂ I . p 2) D S ⊂ A (1.93) The algebras of generalized functions can be further particularized in connection with the important concepts of generality, exactness and stability [140]. CHAPTER 1. A BRIEF HISTORY OF PDES 29 The method described in this section for constructing algebras of generalized functions turns out to be particularly efficient in the study of nonlinear PDEs. In this regard, the particular version of this theory, developed in [39], [40] has proved to be highly successful, both in connection with the exact and numerical solutions to nonlinear PDEs. However, the more general version of the theory developed by Rosinger [135], [136], [137], [138] and [140] provides a better insight into the structure of what may be conveniently called all possible algebras of generalized functions. Furthermore, the more general theory in [140] has delivered deep results such as a global version of the Cauchy-Kovalevskaia Theorem [141]. Such a result cannot be replicated within the more specific framework of the Colombeau Algebras. This is due to the polynomial type growth conditions imposed on the generalized functions in Colombeau’s algebras. In particular, in this case the subalgebra in (1.73) is not equal to AN . As such, one cannot define arbitrary smooth operations on generalized functions, since an arbitrary analytic function may grow faster than any polynomial. 1.4 The Order Completion Method We have already mentioned the well known but often overlooked, if in fact not ignored, fact that the Schwartz distributions, and other similar spaces of generalized functions, suffer from certain structural weaknesses, and that these spaces fail to contain generalized solutions to a significantly general class of systems of PDEs. In particular, one may recall the Schwartz Impossibility Result, the Lewy Impossibility Result, as well as the nonlinear stability paradox discussed in Section 1.2. In view of such weaknesses, and as mentioned earlier, in order to develop a general framework for generalized solutions to PDEs, it is crucial that one goes beyond the usual distributions and related linear spaces that are customary in the study of nonlinear PDEs. In this regard, one may then chose to construct spaces in such a way as to extend the Schwartz distributions. On the other hand, in view of the insufficiency and structural weaknesses of the distributions, one may start all over, and construct convenient spaces of generalized solutions which do not necessarily contain the distributions. The approach mentioned first is pursued in its full generality through Rosinger’s Algebra First approach, as developed in [135] through [142], and discussed in Section 1.3, as well as the particular, yet highly important, case of that theory developed by Colombeau [39], [40]. The second possibility, that is, to define spaces of generalized functions without reference to the Schwartz distributions and other customary linear spaces of generalized functions, was pursued in a systematic and general way for the first time in [119], where spaces of generalized solutions are constructed through the process of Dedekind Order Completion of spaces of usual smooth functions on Euclidean domains. The advantage of this approach, in comparison with the usual functional analytic methods, in particular as far as its generality and type independent power is CHAPTER 1. A BRIEF HISTORY OF PDES 30 concerned, comes from the fact that it is formulated within the context of the more basic concept of order. In this regard, we may note that present day mathematics is a multi layered science, with successive and more sophisticated layers constructed upon one another, such as, for instance, is illustrated in the diagram below. Set Theory ↓ Binary Relations ↓ Order ↓ Algebra ↓ Topology ↓ Functional Analysis etc Traditionally, see also Sections 1.1 and 1.2, the problem of solving linear and nonlinear PDEs, is formulated in the context of the most sophisticated two levels, namely, that of topology and functional analysis. As a consequence of the almost exclusive use of the highly specialized tools of functional analysis, the basic underlying concepts involved in solving PDEs are only perceived through some of their most complicated aspects. Furthermore, the Order Completion Method, in contradistinction with the usual methods discussed in Section 1.2, applies to situations which are far more general than PDEs alone. Indeed, the method is based on general results on the construction of Dedekind order complete partially ordered sets, and the extension of suitable mappings between such ordered sets. This is exactly the reason for its type independent power. Now, as mentioned, the Order Completion Method goes far beyond the usual methods of functional analysis when it comes to the existence of generalized solutions to both linear and nonlinear PDEs. What is more, this method also produces a blanket regularity result for the solutions constructed. Furthermore, in view of the intuitively clear nature of the concept of order, one also gains much insight into the mechanisms involved in the solution of PDEs, as well as the structure of such generalized solutions. In this regard, the deescalation from the level of topology and functional analysis to the level of order proves to be particularly useful and relevant. However, when it comes to further properties of the solutions, such as for instance regularity of the solutions, functional analysis, or for that matter any mathematics, may yet play an important, but secondary role. As mentioned, the theory of Order Completion is based on rather basic constructions in partially ordered sets. In this regard, let (X, ≤X ) and (Y, ≤Y ) be partially ordered sets. A mapping T : X 3 x 7→ T x ∈ Y (1.94) CHAPTER 1. A BRIEF HISTORY OF PDES 31 which is injective, but not necessarily surjective, is an order isomorphic embedding whenever ∀ x 0 , x1 ∈ X : T x0 ≤Y T x1 ⇔ x0 ≤X x1 (1.95) A partially ordered set is called Dedekind Order Complete [101] if every set A ⊂ X which is bounded from above has a least upper bound, and every set which is bounded from below has largest lower bound. That is, ∀ A⊆X : ∃ x0 ∈ X : ∃ u0 ∈ X : 1) ⇒ x ≤X x0 , x ∈ A (x ≤X x0 , x ∈ A) ⇒ u0 ≤X x0 2) ∃ x0 ∈ X : x0 ≤X x, x ∈ A ⇒ ∃ u0 ∈ X : (x0 ≤X x, x ∈ A) ⇒ x0 ≤X u0 (1.96) . With every partially ordered set (X, ≤X ) one may associate a Dedekind Order Complete set X ] , ≤X ] and an order isomorphic embedding ιX : X → X ] , see [101], [102] and [119, Appendix A], such that ∀ x] ∈ X ] : 1) L x] = {x ∈ X : ιX x ≤X ] x] } 6= ∅ . (1.97) 2) U x] = {x ∈ X : x] ≤X ] ιX x} 6= ∅ 3) sup L x] = inf U x] = x] Furthermore, if X0] , ≤X ] is another partially ordered set, and ιX,0 : X → X0] 0 an order isomorphic embedding that satisfies (1.97), then there is a bijective order isomorphic embedding T : X ] → X0] , or shortly an order isomorphism, so that the diagram T X] - @ I @ X0] @ @ ι[email protected] ιX,0 @ @ @ @ X (1.98) CHAPTER 1. A BRIEF HISTORY OF PDES 32 commutes. That is, the Dedekind Order Completion of a partially ordered set is unique up to order isomorphism. Also, see [119, Appendix A], if T : X → Y is an order isomorphic embedding, then T extends uniquely to an order isomorphic embedding T ] : X] → Y ] so that the diagram X T ιX -Y ιY ? X ] T] ? -Y] commutes. Moreover, for any y0 ∈ Y , we have ∃! x] ∈ X ] : ∃ A⊆X : ⇔ T ] x] = y0 y0 = sup{T x : x ∈ A} (1.99) It is within this general context of partially ordered sets and order isomorphic embeddings that the Order Completion Method [119] for nonlinear PDEs is formulated. These basic results on the completion of partially ordered sets and the extensions of order isomorphisms, may be applied to large classes of nonlinear PDEs in so far as the existence, uniqueness and basic regularity of generalized solutions are concerned. In this regard we consider a general, nonlinear PDE T (x, D) u (x) = f (x) , x ∈ Ω nonempty and open (1.100) of order m ≥ 1 arbitrary but fixed. The right hand term f : Ω → R is assumed to be continuous, and the partial differential operator T (x, D) is supposed to be defined through a jointly continuous mapping F : Ω × RM → R (1.101) T (x, D) u (x) = F (x, u (x) , ..., Dp u (x) , ..) , x ∈ Ω (1.102) by the expression CHAPTER 1. A BRIEF HISTORY OF PDES 33 where p ∈ Nn satisfies |p| ≤ m. As is well known, see for instance Sections 1.1 and 1.2, a general nonlinear PDE of the form (1.100) through (1.102) will in general fail to have classical solutions on the whole domain of definition Ω. However, a necessary condition for the existence of a solution to (1.100) on a neighborhood of any x0 ∈ Ω may be formulated in terms of the mapping (1.101) in a rather straight forward way. In this regard, suppose that for some x0 ∈ Ω there is a neighborhood V of x0 and a function function u ∈ C m (V ) that satisfies (1.100). That is, ∃ V ∈ Vx0 nonempty, open : ∃ u ∈ C m (Ω) : , T (x, D) u (x) = f (x) , x ∈ V (1.103) where Vx0 denotes set of open neighborhoods of x0 . Then, in view of (1.103), it is clear that ∀ x∈V : ∃ (ξp (x))|p|≤m ∈ RM : . F (x, ..., ξp (x) , ...) = f (x) (1.104) Now, in view of (1.103) and (1.104) it is clear that the condition ∀ x ∈ Ω :n o M ξ = (ξ ) ∈ R f (x) ∈ F (x, ξ) p |p|≤m (1.105) is nothing but a necessary condition for the existence of a classical solution u ∈ C m (Ω) to (1.100). Many PDEs of applicative interest satisfies (1.105) trivially. Indeed, in this regard we may note that, since the mapping F is continuous, the set o n F (x, ξ) ξ = (ξp )|p|≤m ∈ RM must be an interval in R which is either bounded, half bounded or equals all of R. In the particular case of a linear PDE X T (x, D) u (x) = aα (x) Dα u (x) , x ∈ Ω |α|≤m that satisfies ∀ x∈Ω: ∃ α ∈ Nn , |α| ≤ m : aα (x) 6= 0 we have n o F (x, ξ) ξ = (ξp )|p|≤m ∈ RM = R. (1.106) CHAPTER 1. A BRIEF HISTORY OF PDES 34 This is also the case for most nonlinear PDEs of applicative interest, including large classes of polynomial nonlinear PDEs. As a mere technical convenience, we shall assume the slightly stronger condition ∀ x∈Ω: o n f (x) ∈ int F (x, ξ) ξ = (ξp )|p|≤m ∈ RM (1.107) In view of (1.106) it follows that (1.107) is also satisfied by the respective classes of linear and nonlinear PDEs. Subject to the assumption (1.107) one obtains the local approximation result ∀ ∀ ∃ ∃ x0 ∈ Ω : >0: δ>0: u ∈ C ∞ (Ω) : kx − x0 k ≤ δ ⇒ f (x) − ≤ T (x, D) u (x) ≤ f (x) (1.108) Indeed, from (1.107) it follows that, for > 0 small enough, there is some ξ ∈ RM so that (1.109) F (x0 , ξ ) = f (x0 ) − . 2 Choosing u ∈ C ∞ (Ω) in such a way that ∀ |p| ≤ m : , Dp u (x0 ) = ξp the result follows. At this junction we should note that, in contradistinction with rather difficult techniques of the usual functional analytic methods, the local approximation condition (1.108) follows by from basic properties of continuous functions on Euclidean space. Furthermore, a global version of (1.108) is obtained as a straightforward application of (1.108) and the existence of a suitable tiling of Ω by compact sets, see for instance [58]. In this regard, we have ∀ >0: ∃ Γ ⊂ Ω closed nowhere dense : ∃ u ∈ C ∞ (Ω \ Γ ) : x ∈ Ω \ Γ ⇒ f (x) − ≤ T (x, D) u (x) ≤ f (x) (1.110) The singularity set Γ in (1.110) is typically generated as the union of countably many hyperplanes. As such, each of the singularity sets Γ has zero Lebesgue measure, that is, mes (Γ ) = 0. From a topological point of view, the approximation result (1.110) is extraordinarily versatile. Indeed, since the singularity set Γ , for > 0, may be constructed CHAPTER 1. A BRIEF HISTORY OF PDES 35 so that mes (Γ ) = 0, the sequence (un ) of approximating functions, corresponding 1 to the sequence n = n of real numbers, satisfies ∃ E⊆Ω: 1) Ω \ E is of First Baire Category 2) mes (Ω \ E) = 0 3) x ∈ E ⇒ T (x, D) un (x) → f (x) (1.111) Over and above the mere pointwise convergence almost everywhere, the sequence (un ) also satisfies the stronger condition ∃ E⊆Ω: 1) Ω \ E is of First Baire Category 2) mes (Ω \ E) = 0 3) (T (x, D) un ) converges to f uniformly on E (1.112) Furthermore, since the singularity set Γ 1 associated with each function un in the n sequence (un ) is of measure 0, each such function is measurable on Ω. Moreover, one may construct each of the functions un in such a way that ∀ K R ⊂ Ω compact : |un (x) |dx < ∞ K (1.113) Then, in view of (1.112) and (1.113) it follows by rather elementary arguments in measure theory that ∀ K R ⊂ Ω compact : |T (x, D) un (x) − f (x) |dx → 0 K (1.114) The Order Completion Method, as developed in [119] and presented here, operates on a far more basic level than the respective topological interpretations (1.111) to (1.114) of the Global Approximation Result (1.110). In this regard, and in view of the closed nowhere dense singularity sets Γ associated with the approximations (1.110), the family of spaces ∃ Γu ⊂ Ω closed nowhere dense : m (1.115) Cnd (Ω) = u : Ω → R u ∈ C m (Ω \ Γu ) prove to be particularly useful. Indeed, in view of the continuity of the mapping F in (1.102) one may associate with the nonlinear partial differential operator T (x, D) a mapping m 0 (Ω) → Cnd (Ω) T : Cnd with the property m ∀ u ∈ Cnd (Ω) : ∀ Γ ⊂ Ω closed nowhere dense : u ∈ C m (Ω \ Γ) ⇒ T u ∈ C 0 (Ω \ Γ) (1.116) CHAPTER 1. A BRIEF HISTORY OF PDES 36 0 The space Cnd (Ω) contains certain pathologies. In particular, consider the simple example when Ω = R. For different values of α ∈ R, the functions 0 if x 6= α uα : R 3 x → 1 if x = α 0 each corresponds to a different elements in Cnd (Ω), although any two such functions are the same on some open and dense set. In order to remedy this apparent flaw, 0 one may introduce an equivalence relation on the space Cnd (Ω) through ∃ Γ ⊂ Ω closed nowhere dense : 1) u, v ∈ C 0 (Ω \ Γ) u∼v⇔ (1.117) 2) x ∈ Γ ⇒ u (x) = v (x) 0 (Ω) / ∼ is denoted M0 (Ω), and a partial order is defined The quotient space Cnd on it through ∃ u ∈ U, v ∈ V : ∃ Γ ⊂ Ω closed nowhere dense : U ≤V ⇔ (1.118) 1) u, v ∈ C 0 (Ω \ Γ) 2) x ∈ Ω \ Γ ⇒ u (x) ≤ v (x) It should be noted that with the order (1.118) the space M0 (Ω) is a lattice. Since m 0 ∀ u, v ∈ Cnd (Ω) ⊂ Cnd (Ω) : u ∼ v ⇒ Tu ∼ Tv one may associate with the mapping (1.4) in a canonical way a mapping m T : Cnd (Ω) → M0 (Ω) (1.119) so that the diagram T m Cnd (Ω) - C 0 (Ω) nd @ @ @ @ T @ q∼ (1.120) @ @ @ R @ M0 (Ω) commutes, with q∼ the quotient map associated with the equivalence relation (1.117). CHAPTER 1. A BRIEF HISTORY OF PDES 37 The equation T u = f, (1.121) which is a generalization of the PDE (1.100), does not fit in the framework (1.100) through (1.94). In particular, the mapping T is, except in extremely particular cases, neither injective, nor is it monotone. One may overcome these difficulties by associating an equivalence relation with the mapping T through m (Ω) : ∀ u, v ∈ Cnd u ∼T v ⇔ T u = T v (1.122) m (Ω) / ∼T by MTm (Ω), and order it through We denote the quotient space Cnd ∀ U, V ∈ MTm (Ω) : ∀ u ∈ U, v ∈ V : U ≤T V ⇔ Tu ≤ Tv (1.123) Clearly, one may now obtain, in a canonical way, an injective mapping Tb : MTm (Ω) → M0 (Ω) (1.124) so that the diagram m Cnd (Ω) T - C 0 (Ω) nd qT q∼ ? MLm (Ω) T Tb (1.125) ? - M0 (Ω) commutes, with qT the quotient mapping associated with the equivalence relation (1.122). In particular, the mapping Tb may be defined as ∀ U ∈ MTm (Ω) : Tb : U 7→ T u, u ∈ U (1.126) The mapping Tb clearly also satisfies (1.95) so that it is an order isomorphic embedding, and the setup (1.95) through (1.99) applies to the generalized equation TbU = f , U ∈ Mm (Ω) T (1.127) CHAPTER 1. A BRIEF HISTORY OF PDES 38 which, in view of the commutative diagrams (1.120) and (1.125), is equivalent to (1.121). In this regard, applying (1.110) and (1.99) leads to the fundamental existence and uniqueness result ∀ f ∈ C 0 (Ω) that satisfies (1.107) : ∃! U ] ∈ MTm (Ω)] : Tb] U ] = f (1.128) where Mm (Ω)] and M0 (Ω)] are the Dedekind completions of Mm (Ω) and M0 (Ω), T T respectively, and Tb] : MTm (Ω)] → M0 (Ω)] is the unique extension of Tb. It should be noted that, in contradistinction with the usual functional analytic and topological methods, the arguments leading to the above existence and uniqueness result are particularly clear and transparent. Moreover, these arguments remain valid in a far more general setting. Furthermore, and as we have mentioned, in view of the transparent and intuitively clear way in which the Dedekind order completion of a partially ordered set is constructed, the structure of the generalized solution U ] ∈ MTm (Ω)] to (1.100) is particularly clear. In this regard, we recall the abstract construction of the completion of a partially ordered set [101], [102], [119]. Consider a nonvoid partially ordered set (X, ≤X ) which, for convenience, we assume is without top or bottom. Two dual operations on the powerset P (X) of X may be defined through \ P (X) 3 A 7→ Au = [ai ∈ P (X) a∈A and P (X) 3 A 7→ Al = \ ha] ∈ P (X) a∈A where [ai and ha] are the half bounded intervals [ai = {x ∈ X : a ≤X x}, ha] = {x ∈ X : x ≤X a}. That is, Au is the set of upper bounds of A, and Al is the set of lower bounds of A. It is obvious that A u = X ⇔ Al = X ⇔ A = ∅ (1.129) CHAPTER 1. A BRIEF HISTORY OF PDES 39 while Au = ∅ ⇔ A unbounded from above Al = ∅ ⇔ A unbounded from below. (1.130) A set A ∈ P (X) is called a cut of the poset X if A = Aul = (Au )l . The set of all cuts of X is denoted X̃, that is, X̃ = {A ∈ P (X) : Aul = A}. In view of (1.129) and (1.130) we have ∅, X ∈ X̃ so that X̃ 6= ∅. Furthermore, for each A ∈ P (X) we have A ⊂ Aul ∈ X̃ (1.131) and Aul ul = Aul (1.132) so that, in view of (1.131) and (1.132) we have X̃ = {Aul : A ∈ P (X)}. In particular, for each x ∈ X the set hx] belongs to X̃. Furthermore, it is clear that ∀ x0 , x1 ∈ X : hx0 ] ⊆ hx1 ] ⇔ x0 ≤ x1 so that the mapping ιX : X 3 x 7→ hx] ∈ X̃ (1.133) is an order isomorphic embedding when X̃ is ordered through inclusion A ≤ B ⇔ A ⊆ B. The main theorem in this regard, due to MacNeille, is the following. Theorem 3 *[102] Let X be a partially ordered set without top or bottom, and X̃ the set of cuts in X, ordered through inclusion. Then the following statements are true: CHAPTER 1. A BRIEF HISTORY OF PDES 40 1. The poset X̃, ≤ is order complete. 2. The order isomorphic embedding ιX in (1.133) preserves infima and suprema, that is, ∀ A⊆X : 1) x0 = sup A ⇒ ιX x0 = sup ιX (A) 2) x0 = inf A ⇒ ιX x0 = inf ιX (A) 3. For A ∈ X̃, we have the order density property X in X̃, namely A = supX̃ {ιX x : x ∈ X ,hx] ⊆ A} = inf X̃ {ιX x : x ∈ X ,hx] ⊇ A} An easy corollary to MacNeille’s Theorem 3 is the following Dedekind Completion Theorem. Corollary 4 Suppose that the partially ordered set X is a lattice. Then the partially ordered set X ] = X̃ \ {X, ∅} ordered through inclusion is a Dedekind complete lattice. Furthermore, the mapping ιX : X 3 x 7→ hx] ∈ X ] is an order isomorphic embedding which preserves infima and suprema. Furthermore, the order density property A = supX̃ {ιX x : x ∈ X ,hx] ⊆ A} = inf X̃ {ιX x : x ∈ X ,hx] ⊇ A} also holds. Now, in view of the above general construction, we may interpret the existence and uniqueness result (1.128) for the solutions to continuous nonlinear PDEs of the form (1.100) through (1.102), subject to the assumption (1.107) as follows. From the approximation result (1.110) and the definition (1.126) it follows that f = sup{TbU : U ∈ Mm (Ω) , TbU ≤ f } T Then, in view of the definition (1.123) of the partial order on Mm (Ω), as well as T ] Corollary 4, the generalized solution U to (1.100) may be expressed as U ] = {U ∈ MTm (Ω) : TbU ≤ f } CHAPTER 1. A BRIEF HISTORY OF PDES 41 Finally, recalling the structure of the quotient space, and in particular the equivalence relation (1.122) one has m (Ω) : T u ≤ f } U ] = {u ∈ Cnd The unique generalized solution to (1.100) may therefore be interpreted as the tom (Ω) to (1.100) which includes also all exact classical tality of all subsolutions u ∈ Cnd m solutions, whenever such solutions exist, and all generalized solutions in Cnd (Ω). In this regard, we may notice that the notion of generalized solution through Order Completion is consistent with the usual classical solutions as well as with generalized m solutions in Cnd (Ω). The Order Completion Method also provides a blanket regularity, see [8] and [9], for the solutions to nonlinear PDEs, in the following sense, which is a consequence of the fact that the Dedekind completion of the space M0 (Ω) may be represented as the set Hnf (Ω) of all nearly finite Hausdorff continuous interval valued functions. In this regard, there is then an order isomorphism F0 : M0 (Ω)] → Hnf (Ω) (1.134) Then, since the mapping Tb] : MTm (Ω) → M0 (Ω)] is an order isomorphic embedding, it follows that Tb] ◦ F0 : Mm (Ω) → Hnf (Ω) T is an order isomorphic embedding. In this way, the generalized solution to (1.100) may be seen as being assimilated with usual Hausdorff continuous functions. At this point we have only considered existence and uniqueness of generalized solutions to free problems. That is, we have only solved the equation (1.100) without imposing any addition boundary and / or initial conditions. It is well know that, in the traditional functional analytic approaches to PDEs, in particular those that involve weak solutions and distributions, the further problem of satisfying such additional conditions presents difficulties that most often require entirely new and rather difficult techniques. This is particularly true when distributions, their restrictions to lower dimensional manifolds and the associated trace operators are involved. In contrast to the well known difficulties caused by the presence of boundary and / or initial conditions in the customary methods for PDEs, the Order Completion Method incorporates such conditions in a rather straight forward and easy way, as demonstrated by several examples, see for instance [119, Part II]. This is achieved by first obtaining an appropriate version of the global approximation result (1.110) that incorporates the respective boundary and / or initial value problem. The key is what amounts to a separation of the problem of satisfying the PDE and that of satisfying the additional condition. In this way, boundary and / or initial value problems are solved, essentially, by the same techniques that apply to the free problem. CHAPTER 1. A BRIEF HISTORY OF PDES 42 As a final remark concerning the theory for the existence and regularity of the solutions to arbitrary continuous nonlinear PDEs, as we have sketched it in this section, we mention certain possibilities for further enrichment of the basic theory. In particular, the following may serve as guidelines for such an enrichment. (A) The space of generalized solutions to (1.100) may depend on the PDE operator T (x, D) (B) There is no differential structure on the space of generalized solutions In order to accommodate (A), one may do away with the equivalence relation (1.122) m (Ω), and instead consider on Cnd ∃ Γ ⊂ Ω closed nowhere dense : 1) u, v ∈ C m (Ω \ Γ) u∼v⇔ 2) x ∈ Ω \ Γ ⇒ u (x) = v (x) to obtain the quotient space Mm (Ω). Furthermore, one may consider a partial order other than (1.123), which does not depend on the partial differential operator T (x, D). Indeed, in the original spirit of Sobolev spaces, one may consider the partial order ∀ U, V ∈ Mm (Ω) : ∀ |α| ≤ m : U ≤D V ⇔ Dα U ≤ Dα V which could also solve (B). However, such an approach presents several difficulties. In particular, the existence of generalized solutions in the Dedekind completion of the partially ordered set (Mm (Ω) , ≤D ) is not clear. In fact, the possibly nonlinear mapping T associated with the PDE (1.100) cannot be extended to the Dedekind completion in a unique and meaningful way, unless T satisfies some additional and rather restrictive conditions. We mention that the use of partial orders other than (1.123) was investigated in [119, Section 13], but the partial orders that are considered are still in some relation to the PDE operator T (x, D). Regarding (B), we may recall that there is in general no connection between the usual order on Mm (Ω) ⊂ M0 (Ω) and the derivatives of the functions that are its elements. One possible way of going beyond the basic theory of Order Completion is motivated by the fact that the process of taking the supremum of a subset A of a partially ordered set X is essentially a process of approximation. Indeed, x0 = sup A (1.135) means that the set A approximates x0 arbitrarily close from below. Approximation, however, is essentially a topological process. In this regard, the various topological interpretations (1.111) through (1.114) of the global approximation result (1.110) present a myriad of new opportunities. Therefore, and in connection with (1.135), perhaps the most basic approach, and the one nearest the basic Theory of Order Completion would comprise a topological type model for the process of Dedekind completion of M0 (Ω). CHAPTER 1. A BRIEF HISTORY OF PDES 1.5 43 Beyond distributions Ever since Schwartz [145] proved the so called Schwartz Impossibility Result, and Lewy [97] gave an example of a linear, variable coefficient PDE with no distributional solution, and in particular over the past four decades, there has been an increasing awareness that the usual methods of linear functional analysis, which are quite effective in the case of linear PDEs, in particular those with constant coefficients, cannot be reproduced in any general and consistent way when dealing with nonlinear PDEs. As such, a number of other methods, mostly based on linear functional analysis, were introduced. Many of these theories, it should be mentioned, depend far less on the sophisticated tools of functional analysis than is the case, for instance, with the D’ distributions. In this regard, we mention here the Theory of Monotone Operators [31], and the theory of Viscosity Solutions, see for instance [42]. These methods, however, were developed with particular types of equations in mind, and their powers are therefore limited to those particular types of equations for which they were designed. It should be noted that, in those cases when these methods do apply, they have proven to be effective beyond the earlier functional analytic methods. Recently, a general and type independent theory for the existence and regularity of generalized solutions of systems of nonlinear PDEs, based on techniques from Hilbert space, was initiated by J Neuberger, see [115] through [118]. This theory is based on a generalized method Steepest Descent in suitably constructed Hilbert spaces. Since this theory is not restricted to any particular class of nonlinear PDEs, that is, it is general and type independent, it bears comparison with methods developed in Part II of this work. As such, we include below a short account of the underlying ideas involved. In this regard, let H be a suitable Hilbert space of generalized functions, for instance, H might be one of the Sobolev spaces H 2,m (Ω). For a given nonlinear PDE T (x, D) u (x) = f (x) , (1.136) the method is supposed to produce a generalized solution in H. In order to obtain such a generalized solution, a suitable real valued mapping φT : H → R (1.137) is associated with with the nonlinear partial differential operator T (x, D) such that the critical points u ∈ H of φT correspond to the solutions of (1.136) in H. In this regard, recall that the derivative of a mapping φ:H→R CHAPTER 1. A BRIEF HISTORY OF PDES 44 at u ∈ H is a continuous linear mapping φ0u : H → R that satisfies |φ (u + v) − φ (u) − φ0u (v) | = 0. v→0 kvk lim We may associated with the function φ a mapping Dφ : H 3 u 7→ φ0u ∈ H 0 (1.138) where H 0 is the dual of the Hilbert space H. The mapping φ is C 1 -smooth on H whenever the mapping (1.138) is continuous. For such a C 1 -smooth mapping φ, the gradient of φ is the mapping ∇φ : H → H such that ∀ u, v ∈ H : . φ0u (v) = hv, ∇φ (u)iH (1.139) The gradient mapping ∇φ exists since Dφ is continuous. A critical point of φ is any u ∈ H such that (∇φ) (u) = 0. Neuberger’s method involves techniques to show the existence of critical points to mappings (1.137) associated with a nonlinear PDE, as well as effective numerical computation of such critical points. This involves, inter alia, the adaptation of the gradient mapping (1.139), as well as modifications of Newton’s Method of Steepest Descent to the particular problem at hand. It should be noted that the underlying ideas upon which these methods are based do not depend on any particular form of the nonlinear partial differential operator T (x, D). As such, the theory is, to a great extent, general and type independent. However, the relevant techniques involve several highly technical aspects, which have, as of yet, not been resolved for a class of equations comparable to that to which the Order Completion Method applies. On the other hand, the numerical computation of solutions, based on this theory, has advances beyond the scope of analytic techniques. In this regard, remarkable results have been obtained, see for instance [118]. Chapter 2 Topological Structures in Analysis 2.1 Point-Set Topology: From Hausdorff to Bourbaki Topology, generally speaking, may be described as that part of mathematics that deals with shape and nearness without explicit reference to magnitudes. The first results of a topological nature date back to Euler, who solved the now well known ‘Bridges of Köningstad’ problem. Cantor, however, gave the first description of the topology on R in the modern spirit of the subject. Namely, Cantor introduced the concept of an open set in R. It was only in 1906 when a general framework was introduced in which to describe such concepts as distance, nearness, neighborhood and convergence in an abstract setting. In this regard Fréchet [59] introduced the concept of a metric space, which generalizes the Euclidean spaces ∀ i ∈ N, i ≤ n : n ,n≥1 (2.1) R = x = (xi )i≤n x i ∈ Rn with the usual Euclidean metric ! 12 d2 : Rn × Rn 3 (x, y) 7→ X (xi − yi )2 ∈ R+ ∪ {0} (2.2) i≤n which, in the case n = 3, coincides with our everyday experience of the distance between two points in space. The concept of a metric space is a generalization of (2.1) to (2.2) in two different, yet equally important ways. In the first place, the set Rn of n-tuples of real numbers is replaced by an arbitrary, nonempty set X. Furthermore, the mapping d2 : Rn × Rn → R+ ∪ {0} is replaced by a suitable real valued mapping dX : X × X → R+ ∪ {0} 45 (2.3) CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 46 where for each x, y ∈ X, the real number dX (x, y) is interpreted as the distance between x and y. The properties that the metric dX in (2.3) is supposed to satisfy are suggested by our everyday experience with the distance between two points in three dimensional space, which may be expressed in mathematical form through (2.2). In particular, any two points in space are at a fixed, nonnegative distance from each other, which is expressed as the relation ∀ x, y ∈ X : . dX (x, y) ≥ 0 (2.4) Furthermore, the distance between any two distinct points is positive, that is, ∀ x, y ∈ X : . d (x, y) = 0 ⇔ x = y (2.5) Moreover, distance, as we commonly experience it, is a symmetric relation. That is, moving from point A to point B in three dimensional space along a straight line, one traverses the same distance as if we were moving from B to A. In general, this may be expressed as ∀ x, y ∈ X : . dX (x, y) = dX (y, x) (2.6) The fourth and final condition has rather strong geometric antecedents, as well as consequences. In this regard, consider two distinct points A and B in three dimensional space. The shortest path that can be traced out by a particle moving from the point A to the point B is the straight line connecting A and B. This may be generalized by the condition ∀ x, y, z ∈ X : . dX (x, y) ≤ dX (x, z) + dX (z, y) (2.7) Within such a general framework it is possible to describe a variety of concepts that are of importance in analysis. In this regard, one may formulate the notion of convergence of a sequence without reference to the particular properties of the elements of the underlying set X. Furthermore, Cantor’s concept of an open set in R has a natural and straight forward generalization to metric spaces. Moreover, within the setting of metric spaces, one may define the fundamental concept of continuity of a function u:X→Y in far more general cases than had previously been considered. However, the way in which metric spaces are defined in (2.3) to (2.7) places rather serious limitations on the possible structures that may be conceived of in this way. A most simple example will serve to illustrate the limitations of the concept of a metric space. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 47 Example 5 Consider the set RR of all functions u : R → R. A sequence (un ) in RR converges pointwise to u ∈ RR whenever ∀ x∈R: ∀ >0: . ∃ N (x, ) ∈ N : n ≥ N (x, ) ⇒ |u (x) − un (x) | < (2.8) There is no metric d on RR for which (2.8) is equivalent to ∀ >0: ∃ N () ∈ N : . n ≥ N () ⇒ d (u, un ) < (2.9) In this regard consider the function u : R 3 x 7→ /Q 0 if x ∈ 1 if x ∈ Q Write the set Q in the form of a sequence Q = {q1 , q2 , ...}. For each n ∈ N let Qn = {q1 , ..., qn }. For each n ∈ N, define the function un as / Qn 0 if x ∈ un : R 3 x 7→ 1 if x ∈ Qn Clearly the sequence (un ) converges pointwise to u. Let C 0 (R, R) denote the subspace of RR consisting of all continuous functions u : R → R. Suppose that there exists a metric d on RR so that (2.9) is equivalent to (2.8). Let C 0 (R)c denote the closure of C 0 (R) in RR with respect to the metric d, that is, 0 c 0 R ∃ (un ) ⊂ C (R) : C (R) = u ∈ R (un ) converges to u w.r.t. d Each function un is the pointwise limit of a sequence of continuous functions. As such, un ∈ C 0 (R)c for each n ∈ N, and consequently u ∈ C 0 (R)c . Therefore there exists a sequence (vn ) of continuous functions that converges pointwise to u. Then u is continuous everywhere except on a set of First Baire Category [121], which is clearly not the case. In view of Example 5 it is clear that, even though Fréchet’s theory of metric spaces provides a useful framework in which to formulate some problems in analysis, there are certain rather basic and important situations that cannot be described in terms of metric spaces. Furthermore, the axioms (2.4) through (2.7) of a metric CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 48 are all based on our geometric intuition, an intuition which is most typically based solely on our experience of the usual three dimensional space we believe ourselves to reside in. As such, it turns out to fail in capturing other, more general notions of space. In order to obtain more general structures than metric spaces, one may replace the conditions (2.3) to (2.7) on the mapping d : X × X → R that defines the metric space structure with less stringent ones to obtain, for instance, a pseudo-metric space. This, however, would miss the essence of the matter. F Hausdorff was the first to free topology from our geometric intuition. In particular, the fundamental concept in Hausdorff’s view of topology is not that of ‘distance’, but rather a generalization of Cantor’s definition of open set. Moreover, this is done in purely set theoretic terms, and, in general, is does not involve any notion of magnitude. In this regard, we recall what it means for two metrics d0 and d1 on a set X to be equivalent. Namely, d0 and d1 are considered equivalent whenever ∀ x0 ∈ X : ∃ αx0 ,0 , αx0 ,1 > 0 : . αx0 ,0 d0 (x0 , x) ≤ d1 (x0 , x) ≤ αx0 ,1 d0 (x0 , x) , x ∈ X (2.10) That is, every open ball with respect to d0 contains an open ball with respect to d1 and, conversely, every open ball with respect to d1 contains an open ball with respect to d0 . Then it is clear that, for any two metrics d0 and d1 on X we have ∀ U ⊆X : ⇔ d0 equivalent to d1 (2.11) U open w.r.t. d0 ⇔ U open w.r.t. d1 In view of the equivalence of (2.10) through (2.11), we may describe a metric space (X, d) uniquely by either specifying the metric explicitly, or by specifying the collection τ of open sets. Therefore, in order to generalize the concept of metric space, one may, as mentioned, relax some of the conditions (2.4) to (2.7), or consider a suitable family of mappings. On the other hand, one may generalize the concept of ‘open set’. Not only is the structure of the metric space uniquely and equivalently specified in terms of either one of these two concepts, but, and equally importantly, the continuous functions u : X → Y, where Y is another metric space, are also uniquely and equivalently specified. As mentioned, Hausdorff’s concept of topology is based on a generalization of the concept of an open set. In this regard, one may deduce, from the definition of an open set in a metric space, and the axioms of a metric (2.3), certain properties of open sets which may be stated without reference to the metric. First of all, for CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 49 some metric space (X, dX ), we note that, both the entire space X and the empty set ∅ are open. That is, X, ∅ ∈ τXd , (2.12) where τXd denotes the collection of open subsets of X, with respect to the metric dX . Indeed, ∀ x0 ∈ X : ∀ δ>0: Bδ (x0 ) ⊆ X and ∀ δ>0: /∅: =∅ x ∈∅ ∃ x∈ 0 dX (x0 , x) < 0 Furthermore, the collection of open sets are closed under the formation of finite intersections. That is, ∀ U1 , ..., TUn n ∈ τdX : U = i=1 Ui ∈ τdX (2.13) Indeed, is U either empty, or nonempty. In case U is nonempty, we have ∀ x0 ∈ U : ∀ i = 1, ..., n : ∃ δx0 ,i > 0 : Bδx0 ,i (x0 ) ⊆ Un Hence, upon setting δx0 = inf{δx0 ,i : i = 1, ..., n}, it is clear that Bδ (x0 ) ⊆ U , so that U ∈ τdX . Moreover, the union of any collection of open sets is open, so that ∀ {Ui : [ i ∈ I} ⊆ τdX : U= Ui ∈ τdX (2.14) i∈I It is these three properties (2.12), (2.13) and (2.14), together with the so called Hausdorff property ∀ x 0 , x1 ∈ X : ∃ U0 , U1 open sets : 1) U0 ∩ U1 = ∅ 2) x0 ∈ U0 , x1 ∈ U1 (2.15) that Hausdorff [70] set down as the axioms of his topology. In particular, Hausdorff defined a topology on a set X to be any collection τ of subsets of X that satisfies CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 50 (2.12), (2.13), (2.14) and (2.15), and termed the pair (X, τ ) a topological space. With the minor revision of omitting the Hausdorff Axiom (2.15), this definition has remained unchanged for nearly a century. Hausdorff’s theory was subsequently developed by several authors, chief among these being Kuratowski [93], [94], and the Bourbaki group [30]. In particular, the Bourbaki group, and most notably A Weil [160], introduced the highly important concept of a uniform space as a generalization of that of a metric space, within the context of topological spaces. The concept of uniform space allows for the definition of Cauchy sequences, or more generally Cauchy filters, as well as the associated concepts of completeness and completion. Hausdorff’s concept of topology, although rather abstract, proved extraordinarily useful, in particular in analysis. By the middle of the 20th century, mathematicians realized that Hausdorff’s topology could provide the framework within which Banach’s powerful results on normed vector spaces [17] could be generalized. With the subsequent development of the theory of locally convex spaces, the much sought generalizations were, to a limited extent, fulfilled. However, even at this early stage of development of topology, and its applications to mathematics in general, in particular analysis, certain deficiencies in general topology became apparent. 2.2 The Deficiencies of General Topology As mentioned, Hausdorff’s concept of topology proved to be particularly useful in generalizing classical result in analysis, for instance the powerful tools of linear functional analysis developed by Banach [17] within the setting of metric linear spaces. In spite of the great utility of these techniques, several serious deficiencies of General Topology had emerged by the middle of the twentieth century. In particular, the most important failure of the category of topological spaces is that it is not Cartesian closed, which is as much as to say that there is no natural topological structure for function spaces. In this regard, recall that if X, Y and Z are sets, then one has the relation Y Z X×Y ' Z X . (2.16) That is, there is a canonical one-to-one correspondence between functions f :X ×Y →Z (2.17) g : Y → Z X = {h : X → Z}. (2.18) and functions Indeed, with any function (2.17) we may associate the function f˜ : Y 3 y → f (·, y) ∈ Z X . (2.19) CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 51 That is, f˜ (y) : X 3 x 7→ f (x, y) ∈ Z (2.20) Conversely, with a function (2.18) we can associate the mapping g : X × Y 3 (x, y) 7→ g (y) (x) ∈ Z (2.21) Within the context of topological spaces (2.16) is naturally formulated in terms of continuous functions. In this case, the exponential law (2.16) may be expressed as C (X × Y, Z) ' C (Y, C (X, Z)) , (2.22) In general, (2.22) is not satisfied. That is, there are plenty of topological spaces X, Y and Z, that are of significant interest, such that there is no topology τ on the spaces of continuous functions in (2.22) for which (2.22) holds. Indeed, let f : X × Y → Z. (2.23) be a continuous map. With the mapping (2.23) we may associate a mapping Ff : Y 3 y 7→ Ff (y) ∈ C (X, Z) (2.24) defined as Ff : X 3 x 7→ f (x, y) ∈ Z Conversely, with a mapping F : Y → C (X, Z) (2.25) fF : X × Y → Z (2.26) we may associate a mapping defined as fF : X × Y 3 (x, y) 7→ (F (y)) (x) ∈ Z Suppose that C (X, Y ) is equipped with the compact-open topology, which is specified by the subbasis 1) K ⊆ X compact S (K, U ) 2) U ⊆ Y open CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 52 where S (K, U ) = {f ∈ C (X, Y ) : f (K) ⊆ U }. In view of the continuity of the mapping (2.23), it follows that the associated mapping (2.24) must also be continuous. Conversely, if X is locally compact and Hausdorff, then the mapping (2.26) associated with the mapping (2.25) is continuous whenever the mapping (2.25) is continuous. Therefore, in case X is locally compact and Hausdorff, (2.23) through (2.26) specifies a bijective mapping χ : C (X × Z, Y ) 3 f 7→ Ff ∈ C (Z, C (X, Y )) (2.27) Moreover, if Y and Z are also locally compact, the mapping (2.27) is a homeomorphism, which is as much as to say that the exponential law (2.22) holds for locally compact spaces X, Y and Z, and the compact open topology on the relevant spaces of continuous functions. This is known as the universal property of the compact open topology, within the class of locally compact spaces. However, when the assumption of local compactness on any of the spaces X, Y or Z is relaxed, then, in general, either the mapping (2.26) associated with (2.25) fails to be continuous, or the mapping (2.27) is no longer a homeomorphism, see for instance [110]. In particular, unless all the spaces X, Y and Z are locally compact, there is no topology on C (X, Y ) so that the above construction holds. Other rather unsatisfactory consequence of the mentioned categorical failure of topology, namely, that the category of topological spaces is not Cartesian closed, appears in connection with quotient mappings. Recall that a topological quotient map is a surjective map q:X→Y between topological space X and Y that satisfies ∀ U ⊆Y : U open in Y ⇔ q −1 (U ) open in X (2.28) Such maps appear frequently in topology, as well as in its applications to analysis. However, as mentioned, several irregularities appear in connection with quotient mappings [11], of which we mention only the following. Quotient maps are not hereditary. That is, if q : X → Y is a quotient mapping, and A a subset of Y , then the surjective, continuous mapping qA : q −1 (A) → A obtained by restriction q to q −1 (A) in X, is, in general, not a quotient map with respect to the subspace topologies on A and q −1 (A). To see that this is so, consider the following example [126]. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 53 Example 6 Consider the sets X = {0, 1, 2} and Y = {0, 1, 2, 3}. On X consider the topology τX = {{0, 2}, {1, 3}, X, ∅}, and equip Y with the topology τY = {Y, ∅}. Then the mapping 0 if x = 0 1 if x = 1 q : X 3 x 7→ 2 if x ∈ {2, 3} is continuous and surjective. In particular, q is a quotient map. Now consider the subset A = {0, 1} of Y , so that B = q −1 (A) = {0, 1}. The subspace topology on A is τA = {A, ∅}, while the subspace topology on B is τB = {{0}, {1}, B, ∅}. Clearly the mapping qA , which is simply the mapping q restricted to B, is not a quotient map. Furthermore, quotient maps are not productive. That is, if {Xi : i ∈ I} and {Yi : i ∈ I} are families of topological spaces, and for each i ∈ I, the mapping qi : Xi → Yi is a quotient map, then the product of the family of mappings {qi : i ∈ I}, which is defined as Y Y q: Xi 3 x = (xi )i∈I 7→ q (x) = (qi (xi ))i∈I ∈ Yi , (2.29) i∈I i∈I is not a quotient map with respect to the product topologies on Indeed, consider the following example [126]. Q Xi and Q Yi . Example 7 On R consider the equivalence relation ∼ defined through ∀ x 0 , x1 ∈ R : x0 ∼ x1 ⇔ x0 = x1 or {x0 , x1 } ⊂ Z (2.30) Let q∼ : R → R/ ∼ denote the canonical map associated with the equivalence relation (2.30), and equip R/ ∼ with the quotient topology ∀ U ⊆ R/ ∼π : − U ∈ τ ω ⇔ q∼ (U ) open in R so that q∼ is a quotient map. Let idQ denote the identity mapping on the rational numbers Q. Then the mapping q∼ × idQ : R × Q 3 (x, r) 7→ (q∼ x, r) ∈ R × Q is continuous and surjective. If q∼ × idQ were a quotient map, then it would map saturated closed sets onto closed sets. Recall that a closed set is saturated if it is the inverse image of a subset of R/ ∼ ×Q under q∼ × idQ . Let (an ) be a sequence CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 54 of irrational numbers converging to 0. For each n ∈ N, let (rn,m ) be a sequence of rational numbers converging to an . Set 1 A= n + , rn,m : n, m ∈ N and m > 1 . m A is closed and saturated in R × Q, but (q∼ × idQ ) (A) is not closed in R/ ∼ ×Q. As we have mentioned, the lack of a ‘universal topological structure’ for function spaces, as well as the well known difficulties that appear in connection with quotient mappings, is in fact only a concrete manifestation of the fundamental categorical flaw in Hausdorff’s topology. Namely, that the category TOP of all topological spaces with continuous mappings is not Cartesian closed. This flaw of TOP manifests itself even in the relatively simple setting of locally convex linear topological spaces. In particular, if E is a locally convex space with topological dual E ∗ , then, unless E is a normable space, there is no locally convex topology on the dual E ∗ so that the simple evaluation mapping ev : E × E ∗ 3 (x, x∗ ) 7→ x∗ (x) ∈ K (2.31) is continuous, with K the scalar field R or C. To see that this is so, suppose that E is a locally convex space so that the evaluation mapping (2.31) is continuous with respect to some vector space topology τLE on LE. Since (2.31) is continuous at (0, 0), there is a zero neighborhood W in LE, and a convex zero neighborhood U in E such that W (U ) is contained in the unit disc in K. Since W is absorbing, U is bounded in the weak topology and therefore bounded in E. Since E contains a bounded zero neighborhood, it is normed. In view of the remarks above, it is clear that Hausdorff’s concept of ‘topological space’ is not a satisfactory one. In particular, due to the categorial failures of the category TOP, there is no natural structure on function spaces. One solution to this problem is provided by the theory of convergence spaces [26]. 2.3 Convergence Spaces As was observed in Section 2.2, the Hausdorff-Kuratowski-Bourbaki concept of topology suffers from serious deficiencies, which manifest themselves even in the relatively simple setting of locally convex topological vector spaces. Over and above these basic flaws, this notion of topology is rather restrictive, which may be seen from the fact that several useful modes of convergence cannot be adequately described in terms of the usual topology. In this regard, one may recall the following examples, see [120], [126] and [154]. Example 8 Consider on the real line R the usual Lebesgue measure mes, and denote by M (R) the space of all almost everywhere (a.e.) finite, measurable functions on CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 55 R, with the conventional identification of functions a.e. equal. A natural notion of convergence on M (R) is that of convergence a.e. That is, ∀ (un ) ⊂ M (R) : ∀ u ∈ M (R) : (un ) converges a.e. to u ⇔ ∃ E ⊂ R, mes (E) = 0 : x ∈ R \ E ⇒ un (x) → u (x) (2.32) There is no topology τ on M (R) so that a sequence converges with respect to τ if and only if it converges a.e. to the same function. To see this, suppose that such a topology, say τae , exists, and let (un ) be a sequence which converges in measure to 0, but fails to converge a.e.. Then there is a τae neighborhood V of the constant zero function, and a subsequence (unm ) of (un ) so that ∀ nm ∈ N : unm ∈ /V (2.33) Since (un ) converges to 0 in measure, so does the subsequence (unm ). A well know theorem, see for instance [85], states that the subsequence (unm ) contains a further subsequence unmk that converges a.e. to 0. Therefore the sequence unmk is eventually in V , which contradicts (2.33). Example 9 Consider on the space C 0 (R) of all continuous, real valued functions on R the pointwise order ∀ x∈R: u≤v⇔ (2.34) u (x) ≤ v (x) With respect to this order, and the usual vector space operations, the space C (R) is an Archimedean vector lattice [101]. A sequence (un ) in C 0 (R) order converges to u ∈ C 0 (R) whenever ∃ (λn ) , (µn ) ⊂ C 0 (R) : 1) λn ≤ λn+1 ≤ un ≤ µn+1 ≤ µn , n ∈ N 2) sup{λn : n ∈ N} = u = inf{µn : n ∈ N} (2.35) There is no topology τo on C 0 (R) so that a sequence converges to u ∈ C 0 (R) with respect to τo if and only if it order converges to u. To see this, consider the sequence (un ) defined as 1 − n|x − qn | if |x − qn | < n1 un (x) = 0 if |x − qn | ≥ n1 Here Q ∩ [0, 1] = {qn : n ∈ N} is the set of rational numbers in the interval [0, 1], ordered as usual, so that the complement of any finite subset of Q ∩ [0, 1] is dense in [0, 1]. For any N0 ∈ N, we have sup{un : n ≥ N0 } (x) = 1, x ∈ R CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 56 and inf{un : n ≥ N0 } (x) = 0, x ∈ R so that this sequence does not order converge to 0. Suppose that there exists a topology τo on C 0 (R) that induces order convergence. Then there is some τo -neighborhood V of 0, and a subsequence (unm ) of (un ) which is always outside of V . Let (qnm ) denote the sequence of rational numbers associated with the subsequence (unm ). Since this sequence is bounded, there is a subsequence qnmk of it, and some real number q ∈ [0, 1] so that qnmk converges to q. Let unmk be the subsequence of (unm ) corresponding to the sequence of rational numbers qnmk . Then it is clear that ∀ >0: ∃ N ∈ N : nmk > N ⇒ unmk (x) = 0, |x − q| > Set l = 1 l for each l ∈ N. Define the sequence µnmk in C 0 (R) as µnmk 0 1 (x) = |q−x| l if |x − q| ≥ 2l if |x − q| ≤ l + 2 if l < |x − q| < 2l whenever Nl < nmk < Nl+1 . The sequence µnmk decreases to 0, and unmk ≤ µnmk , nmk ∈ N so that unmk order converges to 0. Therefore, it must eventually be in V , a contradiction. Therefore the topology τo cannot exists. We have given two useful and well known examples of concepts of sequential convergence that cannot be described in terms of the usual Hausdorff-KuratowskiBourbaki formulation of topology. Note that the concept of order convergence as introduced in Example 9 may be formulated in terms of an arbitrary partially ordered set (X, ≤). In particular, if X is the set M (R) of usual measurable functions on R, modulo almost everywhere equal functions, with the pointwise a.e. order, then (2.32) is identical with the order convergence. The order convergence is widely used, particularly in the theory of vector lattices, where, for instance, it appears in connection with σ-continuous operators, and in particular integral operators [163]. In view of Examples 8 and 9, as well as the lack of a natural topological structure for function spaces discussed in Section 2.2, a more general notion of topology may be introduced. In this regard, we recall that a given topological space (X, τ ) may be completely described by specifying the convergence associated with the topology CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 57 τ . More precisely, for each x ∈ X, we may specify the set λτ (x) consisting of those filters on X which converge to x with respect to τ . That is, λτ (x) = {F a filter on X : Vτ (x) ⊆ F} (2.36) where Vτ (x) denotes the τ -neighborhood filter at x ∈ X. In particular, if (xn ) is a sequence in X, then we may associate with it its Frechét filter h(xn )i = [{{xn : n ≥ k} : k ∈ N}]. If (xn ) converges to x ∈ X with respect to the topology τ on X, that is, ∀ V ∈ Vτ (X) : ∃ NV ∈ N : , n ≥ NV ⇒ x n ∈ V (2.37) then we must have Vτ (x) ⊆ h(xn )i Conversely, if Vτ (x) ⊆ h(xn )i, then we must have (2.37). As such, the definition (2.36) of filter convergence in a topological space is nothing but a straight forward generalization of the corresponding notion for sequences.1 Remark 10 Recall that a filter F on X is a nonempty collection of nonempty subsets of X such that ∀ F ∈F : ∀ G⊆X : F ⊆G⇒G∈F and ∀ F, G ∈ F : . F ∩G∈F A filter base for a filter is any collection B ⊆ F so that ∃ B∈B : = F. [B] = F ⊆ X B⊆F An ultrafilter on X is a filter which is not properly contained in any other filter. In particular, for each x ∈ X, the filter [x] = {F ⊆ X : x ∈ F } is an ultrafilter on X. The intersection of two filters F and G on X is defined as ∃ F ∈ F, G ∈ G : F ∩G = H ⊆X F ∪G⊆H and it is the largest filter, with respect to inclusion, contained in both G and F. A filter F is finer than G, or alternatively G is coarser than F, whenever G ⊆ F. 1 In the sequel, we will make no distinction between a sequence and its associated Frechét filter. As such, we will denote both entities by (xn ). The meaning will be clear from the context. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 58 More generally, a convergence structure [26] on a set X is defined as follows. Definition 11 A convergence structure on a nonempty set X is a mapping λ from X into the powerset of the set of all filters on X that, for each x ∈ X, satisfies the following properties. (i) [x] ∈ λ (x) (ii) If F, G ∈ λ (x) then F ∩ G ∈ λ (x) (iii) If F ∈ λ (x) and F ⊆ G then G ∈ λ (x). The pair (X, λ) is called a convergence space. When F ∈ λ (x) we say that F converges to x and write “F → x”. Concepts of convergences and convergence spaces that are more general than topological spaces were introduced and developed by several authors, see for instance [33], [37], [38], [51], [57], [78], [79] and [89]. The above Definition 11 is widely used and has proved to be a rather convenient one. It is clear that the mapping λτ associated with a topology τ on a set X through (2.36) is a convergence structure. However, the concept of a convergence structure is far more general than that of a topology. Indeed, convergence almost everywhere of measurable functions, and the order convergence of sequences in C (R) discussed in Examples 8 and 9, respectively, are induced by suitable convergence structures, but cannot be induced by a topology. The most striking generalization inherent in Definition 11 may be formulated as follows. For every x ∈ X, the set of filters λτ (x) that converge to x ∈ X with respect to the topology τ on X has a least element with respect to inclusion. Namely, the neighborhood filter Vτ (x) at x ∈ X. In particular, for each x ∈ X we have \ ∀ F ∈ λτ (x) : . (2.38) Vτ (x) = F = V ⊆X V ∈F F ∈λτ (x) More generally, for every subset {Fi : i ∈ I} of λτ (x), the filter ∀ i ∈ I : \ Fi = F ⊆ X ∃ Fi ∈ Fi : i∈I ∪i∈I Fi ⊆ F (2.39) converges to x with respect to the topology τ . Clearly this need not be the case for a convergence space in general, see for instance [26]. However, topological concepts such as open set, closure of a set and continuity generalize to the more general context of convergence spaces in a natural way. In this regard, if X and Y are convergence spaces with convergence structures λX and λY , respectively, then a mapping f :X→Y CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 59 is continuous at x ∈ X whenever F ∈ λX (x) ⇒ f (F) = [{f (F ) : F ∈ F}] ∈ λY (f (x)) and f is continuous on X if it is continuous at every point x of X. Furthermore, such a continuous mapping f is an embedding if it is injective with a continuous inverse defined on its image, and it is an isomorphism if it is also surjective. The open subsets of a convergence space X are defined through the concept of neighborhood. Note that in the topological case, it follows from (2.38) that for any x ∈ X, a set V ⊆ X is a neighborhood of x if and only if F ∈ λτ (x) ⇒ V ∈ F. (2.40) The definition of a neighborhood of a point x in an arbitrary convergence space is a straightforward generalization of (2.40). Namely, ∀ F ∈ λX (x) : V ∈ VλX (x) ⇔ V ∈F where VλX (x) denotes the neighborhood filter at x ∈ X with respect to the convergence structure λX .2 A set V ⊆ X is open if and only if it is a neighborhood of each of its elements. The generalization of the closure of a subset A of a topological space X within the context of convergence spaces is the adherence. In the case of a topological space X, the closure of a subset A of X consists of all cluster points of A, that is, ∀ V ∈ Vx : . (2.41) clτ (A) = x ∈ X V ∩ A 6= ∅ Therefore, for each x ∈ cl (X), the filter F = [{A ∩ V : V ∈ Vx }] converges to x, and A ∈ F. Conversely, if there is a filter F ∈ λτ (x) so that A ∈ F, then in view of (2.36) it follows that A meets every neighborhood of x, so that x ∈ cl (A). That is, the closure of A consists of all points x ∈ X so that A belongs to some filter F ∈ λτ (x). The generalization of (2.41) to convergence spaces gives rise to the concept of adherence. The adherence of a subset A of a convergence space X is the set ∃ F ∈ λX (x) : . (2.42) aλX (A) = x ∈ X A∈F The set A is called closed if aλX (A) = A.3 2 If the convergence structure or topology is clear from the context, we will use the simplified notation Vx for the neighborhood filter at x ∈ X. 3 Whenever there is no confusion, the adherence of a set A will simply be denoted by a (A) CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 60 Here we should point out that, although the concepts of open set, adherence and closed set coincide with the usual topological notions whenever the convergence space is topological, there are in general some important differences [26]. In particular, the neighborhood filter Vx at x ∈ X need not converge to x, while the adherence operator will typically fail to be idempotent, that is, a (A) 6= a (a (A)) A convergence space X that satisfies ∀ x∈X : Vx ∈ λX (x) is called pretopological, and the convergence structure λX is called a pretopology [26]. The customary constructions for producing new topological spaces form given ones, namely, initial and final structures, are defined for convergence spaces in the obvious way. Given a set X and a family of convergence structures (Xi , λXi )i∈I together with mappings fi : X → Xi , i ∈ I (2.43) the initial convergence structure λX on X with respect to the family of mappings (2.43) is the coarsest convergence structure on X making each of the mappings fi : X → Xi continuous. That is, for any other convergence structure λ on X such that all the fi are continuous, we have λ (x) ⊆ λX (x) , x ∈ X. (2.44) The initial convergence structure is defined as ∀ i∈I : F ∈ λX (x) ⇔ fi (F) ∈ λXi (x) Typical examples of initial convergence structures include the subspace convergence structure and the product convergence structure. Example 12 Let X be a convergence space, and A a subset of X. The subspace convergence structure λA induced on A from X is the initial convergence structure with respect to the inclusion mapping iA : A 3 x 7→ x ∈ X. That is, F ∈ λA (x) ⇔ ∃ F ∈F : G⊆X F ⊆G ∈ λX (x) . (2.45) CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 61 Example 13 Consider a family (Xi )i∈I of convergence spaces, and let X be the Cartesian product of the family Y X= Xi . i∈I The product convergence structure on X is the initial convergence structure with respect to the projection mappings πi : X 3 (xi )i∈I 7→ xi ∈ Xi , i ∈ I. The convergent filters in the product convergence structure may be constructed as follows: A filter F on X converges to x = (xi )i∈I ∈ X if and only if ∀ i∈I : ∃ Q Fi ∈ λXi (xi ) : . i∈I Fi ⊆ F (2.46) Q Q Here i∈I Fi denotes the Tychonoff product of the filters Fi , that is, i∈I Fi is the filter generated by ( ) Y Fi ∈ Fi , i ∈ I Fi . (2.47) Fi 6= Xi for only finitely many i ∈ I i∈I Final structures are constructed in a similar way, and include quotient convergence structures and convergence inductive limits as particular cases. In this regard, given a set X, a family of convergence spaces (Xi )i∈I , and mappings fi : Xi → X, i ∈ I (2.48) the final convergence structure λX on X is the finest convergence structure making all the mappings (2.48) continuous. That is, for every convergence structure λ on X for which each mapping fi is continuous, one has λX (x) ⊆ λ (x) , x ∈ X. In particular, the final convergence structure on X is defined through ∃ i1 , ..., ik ∈ I : ∃ xn ∈ Xin , n = 1, ..., k : [ F ∃ Fn ∈ λXin (xn ) , n = 1, ..., k : . λX (x) = {[x]} 1) fi (xn ) = x, i = 1, ..., k 2) fi1 (F1 ) ∩ ... ∩ fik (Fk ) ⊆ F (2.49) (2.50) Example 14 Let X be a convergence space, Y a set and q : X → Y a surjective mapping. The quotient convergence structure on Y is the final convergence structure CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 62 with respect to the mapping q. In particular, a filter F on Y converges to y ∈ Y with respect to the quotient convergence structure λq on Y if and only if ∃ x1 , ..., xk ∈ q −1 (y) ⊆ X : ∃ F1 , ..., Fk filters on X : . 1) Fi ∈ λX (xi ) , i = 1, ..., k 2) q (F1 ) ∩ ... ∩ q (Fk ) ⊆ F (2.51) If X and Y are convergence spaces, and q : X → Y a surjection so that Y carries the quotient convergence structure with respect to q, then q is called a convergence quotient mapping. Remark 15 In general, it is not true that a topological quotient mapping is a convergence quotient mapping. Indeed, if X and Y are topological spaces, and q:X→Y a continuous mapping, then q is a convergence quotient mapping if and only q is almost open [84], that is, ∀ y∈Y : ∃ x ∈ q −1 (y) : . ∃ Bx a basis of open sets at x : B ∈ Bx ⇒ q (B) open in Y Within the category CONV of convergence spaces, the most striking deficiencies of topological spaces are resolved. In particular, in contradistinction with TOP, the category CONV is cartesian closed. As such, within this larger category, there is a natural convergence structure for function spaces, namely, the continuous convergence structure [26], [28]. If X and Y are convergence spaces, then the continuous convergence structure λc on the set C (X, Y ) of continuous mappings from X into Y is the coarsest convergence structure making the evaluation mapping ωX,Y : C (X, Y ) × X 3 (f, x) 7→ f (x) ∈ Y continuous. That is, for each f ∈ C (X, Y ) and every filter H on C (X, Y ) we have ∀ x∈X : . H ∈ λc (f ) ⇔ ∀ F ∈ λX (x) : (2.52) ωX,Y (H × F) ∈ λY (f (x)) For convergence spaces X, Y and Z, the mapping 4 P : Cc (X × Y, Z) → Cc (X, Cc (Y, Z)) 4 It is customary in the literature to denote by Cc (X, Y ) the set of continuous functions from X into Y equipped with the continuous convergence structure. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 63 which is defined as P (f ) (x) : Y 3 y 7→ f (x, y) ∈ Z is a homeomorphism, which shows that the category CONV is indeed cartesian closed. Other difficulties encountered when working exclusively with topological spaces, such as for instance some of those mentioned in connection with quotient mappings, may also be resolved by considering the more general setting of convergence spaces. In this regard, we mention only the following. Example 16 In the category of convergence spaces quotient mappings are hereditary. Indeed, let X and Y be convergence spaces, and q:X→Y a surjective mapping so that Y carries the quotient convergence structure with respect to q. Consider any subspace A of Y , and the surjective mapping qA : q −1 (A) 3 x 7→ q (x) ∈ A. (2.53) Clearly the subspace convergence structure on A is coarser than the quotient convergence structure induced by the mapping (2.53). Let the filter F on A converge to y ∈ A with respect to the subspace convergence structure. That is, ∃ F1 , ..., Fk filters on X : ∃ x1 , ..., xk ∈ q −1 (y) : q (F1 ) ∩ ... ∩ q (Fk ) ⊆ FY where FY denotes the filter generated by F in Y . Equivalently, we may say ∀ F1 ∈ F1 , ..., Fk ∈ Fk : ∃ F ∈F : . q (F1 ) ∪ ... ∪ q (F1 ) ⊇ F (2.54) We may assume that each filter Fi has a trace on q −1 (A). That is, ∀ Fi ∈ Fi : . Fi ∩ q −1 (A) 6= ∅ As such, for each i = 1, ..., k the filter Fi|A generated in q −1 (A) by the family {Fi ∩ q −1 (A) : Fi ∈ Fi } converges to xi in q −1 (A). From (2.54) and the fact that q −1 (A) is saturated with respect to q, it follows that qA F1|A ∩ ... ∩ qA Fi|A ⊆ F so that F converges to y with respect to the quotient convergence structure on A. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 64 Example 17 In the category of convergence spaces, quotient mappings are productive. That is, if (Xi )i∈I and (Yi )i∈I are families of convergence spaces, and for each i ∈ I the mapping qi : Xi → Yi is a convergence quotient mapping, then the surjective mapping Y Y q:X= Xi 3 (xi )i∈I 7→ (qi xi )i∈I ∈ Y = Yi i∈I i∈I is a convergence quotient mapping. In this regard, and in view of (2.46) and (2.51) a filter F on Y converges to y = (yi )i∈I ∈ Y if and only if ∀ i∈I : ∃ xi,1 , ..., xi,ki ∈ qi−1 (yi ) : ∃ Fi,1 ∈QλXi (xi,1 ) , ..., Fi,ki ∈ λXi (xi,ki ) : F ⊇ i∈I (qi (Fi,1 ) ∩ ... ∩ qi (Fi,ki )) (2.55) where the product of filters in (2.55) is the Tychonoff product (2.47). An elementary, yet somewhat lengthy, computation shows that (2.55) coincides with the quotient convergence structure with respect to the mapping q. The theory of convergence spaces has proven to be particularly powerful in so far as its applications to topology and analysis are concerned [26]. This effectiveness of convergence structures is due mainly to the fact that, as we have mentioned, the category of convergence spaces is cartesian closed, thus providing a suitable topological structure for function spaces. In this regard, we mention only the following. The continuous convergence structure (2.52) yields a function space representation of a large class of topological and convergence spaces. In this regard, recall [26] that for a convergence space X, the mapping iX : X → Cc (Cc (X)) (2.56) defined through iX (x) : Cc (X) 3 f 7→ f (x) ∈ R, where Cc (X) is the set of real valued continuous functions on X equipped with the continuous convergence structure, is continuous. The convergence space X is called c-embedded whenever the mapping (2.56) is an embedding. A characterization of c-embedded spaces may be obtained through the concepts of functionally regular and functionally Hausdorff convergence spaces. Recall [26] that a convergence space X is functionally regular if the initial topology τ on X with respect to C (X) is regular, and it is called functionally Hausdorff if τ is Hausdorff. The main result in this regard is the following. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 65 Theorem 18 * [26] A convergence space X is c-embedded if and only if X is functionally regular, functionally Hausdorff and Choquet [26]. That is, a filter F converges to x ∈ X whenever every finer ultrafilter converges to x. In particular, iX (X) ⊂ Cc (Cc (X)) is the set of all continuous algebra homomorphisms from Cc (X) into R. The class of c-embedded convergence spaces includes, amongst others, all Tychonoff spaces. However, not every c-embedded topological space is a Tychonoff space. Furthermore, products, subspaces and projective limits of c-embedded convergence spaces are again c-embedded. Within the setting of functional analysis, in particular the theory of locally convex spaces, the theory of convergence spaces proves to be highly effective. In particular, the continuous convergence structure provides a natural structure for the topological dual of a locally convex space. In this regard, we may recall from Section 2.2 that for a locally convex space X, there is no locally convex topology on its dual LX so that the simple evaluation mapping ωX : X × LX 3 (x, ϕ) 7→ ϕ (x) ∈ K (2.57) is continuous unless X is normable. However, within the more general framework of convergence spaces, there is a natural dual structure available for locally convex spaces, namely, the continuous convergence structure. In this regard, we recall [26] that a convergence structure λX on a set X is a vector space convergence structure, and the pair (X, λX ) a convergence space, if X is a vector space over some field of scalars K, and the vector space operations + : X × X 3 (x, y) 7→ x + y ∈ X and · : K × X 3 (α, x) 7→ αx ∈ X are continuous. For a convergence vector space X, we denote by Lc X the convergence vector space of all continuous linear functionals on X into K equipped with the continuous convergence structure. For any convergence vector space X, the space Lc X is a convergence vector space, and we call it the continuous dual of X. In this regard, the main result [26] is that the evaluation mapping ωX : X × Lc X → K is jointly continuous. Consequently, the natural mapping iX : X → Lc Lc X (2.58) from X into its second dual, which is defined as iX (x) : Lc X 3 ϕ 7→ ϕ (x) ∈ K, (2.59) CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 66 is also continuous. In particular, in case X is a locally convex topological space, the mapping in (2.58) to (2.59) is actually an embedding. Furthermore, if X is complete, then this mapping is an isomorphism. Thus the continuous convergence structure provides a natural structure for the dual of a locally convex space. Beyond the basic duality result for locally convex spaces, convergence vector spaces prove to be a far more natural setting for functional analysis, in comparison with locally convex spaces. This is so even if one’s primary interest lies in the topological, locally convex case. In this regard, we may mention that the Pták’s Closed Graph Theorem, a technical and notoriously difficult result in locally convex spaces, becomes a transparent and natural result when viewed in the setting of convergence vector spaces, see [21], [22], [25] and [26]. Furthermore, the scope of the Banach-Steinhauss Theorem is greatly expanded by formulating the problem in terms of convergence vector spaces [24]. Moreover, common and important objects such as the inductive limit of a family of locally convex spaces seem to be far removed from its component spaces, when viewed in the setting of locally convex spaces. Consequently, properties of the component spaces rarely translate to properties of the limit, while properties of the limit is not easily lifted to that of the component spaces, see for instance [27] for an indication of such difficulties. In contradistinction with the locally convex topological case, when such constructions are performed in the context of convergence vector spaces, there is a clear connection between components and limits. Lastly, we mention that, owing to the remarkable categorical properties of convergence structures, the theory of convergence spaces has been applied with a good deal of success to difficult problems in point set-topology. In this regard, we mention the recent application to product theorems for topological spaces [51], [111], [112], [113]. 2.4 Uniform Structures We may recall from Section 2.1 that Hausdorff’s topological spaces were introduced as a generalization of Frechét’s metric spaces. Indeed, the concepts of open set, closed set, convergence of sequences, or more generally filters and nets, are extended in a straightforward way to this significantly more general class of spaces. However, certain aspects of the structure of a metric space are not preserved in this generalization, namely, the uniform structure. In this regard, recall that for a metric space X with metric dX , a sequence (xn ) on X is a Cauchy sequence if and only if ∀ >0: ∃ N ∈ N : . n, m ≥ N ⇒ dX (xn , xm ) < (2.60) CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 67 Furthermore, if Y is a another metric space, then a f : X → Y is uniformly continuous whenever ∀ >0: ∃ δ > 0 : dX (x, y) < δ ⇒ dY (f (x) , f (y)) < The space X is called complete if every Cauchy sequence in X converges to some x ∈ X. Moreover, with every metric space X one may associate a complete metric space X ] , with metric dX ] , which is minimal in the following sense: There exists a uniformly continuous embedding ιX : X → X ] so that ιX (X) is dense in X ] . Furthermore, for any complete metric space Y , and any uniformly continuous mapping f : X → Y, there exists a uniformly continuous mapping f ] : X] → Y so that the diagram ιX X - X] @ @ @ @ f] f @@ (2.61) @ @ R @ Y commutes. The above construction was given for the first time by Hausdorff [70]. The interest in such constructions may be seen from the fact that the set of real numbers may be constructed as the completion of a metric space, namely, the metric space of all rational numbers p : p, q ∈ Z Q= q with the usual metric d (x, y) = |x − y|. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 68 When one tries to extend the uniform concepts of Cauchy sequence, completeness and uniform continuity of functions to the more general setting of arbitrary topological spaces, there are several difficulties. These difficulties are all, to some extent, related to the following basic fact. Namely, that the uniform structure of a metric space, and in fact the metric itself, is defined through binary relations, whereas a topology is defined in a unary way. These underlying issues are identified within the most general context so far considered in [132], [133] and [134]. The structures introduced in [132] to [134], however, are far more general than the Hausdorff-Kuratowski-Bourbaki concept of topology, and contain it as a particular case. As a further clarification of the issue raised in the previous paragraph, let us consider the more formal way to introduce the uniform structure of a metric space. In this regard, we may consider the family UX,dX of subsets of X × X specified through ∃ >0: U ∈ UX,dX ⇔ (2.62) U ⊆ U where, for > 0, we define the set U as U = {(x, y) ∈ X × X : dX (x, y) < }. (2.63) It is clear now that, for any sequence (xn ) on X, the sequence is a Cauchy sequence if and only if ∀ >0: ∃ N ∈ N : . n, m > N ⇒ {(xn , xm ) : n, m > N } ⊂ U (2.64) Furthermore, for any other metric space Y , and any mapping f : X → Y , the mapping f is uniformly continuous if and only if ∀ U ∈ UY,dY : . (f −1 × f −1 ) (U ) = {(x, y) : (f (x) , f (y)) ∈ U } ∈ UX,dX (2.65) In view of (2.64) and (2.65) it is clear that, just as the open neighborhoods in X characterize the topology on X, the family of sets UX,dX completely determines the uniform structure of the metric space. In this regard, within the context of topological spaces, given a set X, the uniform structure shall consist of a family UX of subsets of X×X that satisfy certain purely set theoretic properties, properties that are intended to capture suitable topological properties. The concept of a uniform space is then nothing but a distillation of these purely set-theoretic properties of UX . In particular, the Bourbaki group [30], and most notably A Weil [160], defined a uniformity on a set X as follows. Definition 19 A uniformity on a set X is a filter U on X × X that satisfies the following properties CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 69 (i) ∆ ⊆ U for each U ∈ U. (ii) If U ∈ U then U −1 ∈ U. (iii) For each U ∈ U there is some V ∈ U so that V ◦ V ⊆ U . Remark 20 If U and V are subsets of the cartesian product X × X of X, then the inverse of U is defined as U −1 = {(x, y) ∈ X × X : (y, x) ∈ U } while the composition of U and V is specified through ∃ z∈X : U ◦ V = (x, y) ∈ X × X , (x, z) ∈ V , (z, y) ∈ U Furthermore, for any set A ⊆ X, the set U [A] is defined as ∃ x∈A: U [A] = y ∈ X . (x, y) ∈ U If A is the singleton {x}, then we simply write U [x] for U [A]. It is clear that the family (2.62) and (2.63) of subsets of the cartesian product X × X of a metric space X constitutes a uniformity. However, not every uniformity can be induced by a metric through (2.62) and (2.63). Conversely, every uniformity UX on X induces a topology τUX on X. In particular, τUX may be defined as ∀ x∈W : W ∈ τUX ⇔ ∃ U ∈ UX : . (2.66) U [x] ⊆ W Even though the topology (2.66) induced by a uniformity UX on X can in general not be induced by a metric, the conditions (i) to (iii) in Definition 19 have strong metric antecedents [81]. Indeed, the first condition ∆ ⊆ U , U ∈ UX is derived from the property (2.5), while the second condition U ∈ UX ⇒ U −1 ∈ UX merely reflects the symmetry condition (2.6) of a metric. Lastly, the third condition ∀ U ∈ UX : ∃ V ∈ UX : V ◦V ⊆U CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 70 is an abstraction of the triangle inequality (2.7), and states, roughly speaking, that for -balls there are /2-balls. In view of (2.64) it is clear that a sequence (xn ) in a metric space X is a Cauchy sequence if and only if UX,d ⊆ (xn ) × (xn ) , where (xn ) denotes also the Fréchet filter associated with the sequence. As such, we may generalize the concept of a Cauchy sequence to any uniform space X as (xn ) a Cauchy sequence ⇔ UX ⊆ (xn ) × (xn ) . More generally, given any filter F on X, we say that F is a Cauchy filter if and only if UX ⊆ F × F Furthermore, a uniform space X is complete if and only if every Cauchy filter on X converges to some x ∈ X. Other uniform concepts generalize in the expected way to uniform spaces. In this regard, we may recall [81] that a mapping f : X → Y, with X and Y uniform spaces, is uniformly continuous if and only if ∀ U ∈ UY : . (f −1 × f − ) (U ) ∈ UX Moreover, a uniformly continuous mapping is a uniform embedding whenever it is injective, and its inverse f −1 is uniformly continuous on the subspace f (X) of Y . Furthermore, f is a uniform isomorphism is a uniform embedding which is also surjective. The main result in connection with completeness of uniform spaces generalizes the corresponding result for metric spaces mentioned above, and is due to Weil [160]. However, it applies only to Hausdorff uniform spaces, that is, uniform spaces for which the induced topology (2.66) is Hausdorff. In this regard, for any Hausdorff uniform space X there is a complete, Hausdorff uniform space X ] and a uniform embedding ιX : X → X ] so that ιX (X) is dense in X ] . Furthermore, for any complete, Hausdorff uniform space Y , and any uniformly continuous mapping f :X→Y CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 71 there is a uniformly continuous mapping f ] : X] → Y so that the diagram ιX X - X] @ @ @ @ f] f @@ (2.67) @ @ R @ Y commutes. It should be noted that not every topology τX on a set X can be induced by a uniformity through (2.66). Indeed, Weil [160] showed that for a given topology τX on X, there is a uniformity on X that induces τX through (2.66) if and only if τX is completely regular. As such, the class of uniform spaces is rather small in comparison with the class of all topological spaces. In this regard, several generalizations of a uniform space have been introduced in the literature, including that of a quasiuniform space, which is related to the concept of nonsymmetric distance, see for instance [90]. Within the more general context of convergence spaces, a number of different concepts of ‘uniform space’ have been studied. The most successful of these are the so called uniform convergence spaces, and Cauchy spaces. The motivation for introducing such concepts within the setting of convergence spaces is twofold. First of all, it allows for the definition of uniform concepts, in particular that of completeness and completion, in this context, concepts which are fundamental in analysis. The second reason, and related to the first, is the mentioned relatively narrow applicability of uniform spaces within the context of the usual topology. A uniform convergence space generalizes the concept of a uniform space in the following ways. Every uniformity on a set X gives rise to a uniform convergence structure. Furthermore, and as will be shown shortly, every uniform convergence structure induces a convergence structure. This induced convergence structure need not be, and in general is not, topological, and satisfies rather general separation properties. In particular, even in case the induced convergence structure is topological, it need not be completely regular. The definition of a uniform convergence space is now as follows [26]. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 72 Definition 21 A uniform convergence structure on a set X is a family JX on X×X that satisfies the following properties. (i) [x] × [x] ∈ JX for every x ∈ X 5 . (ii) If U ∈ JX , and U ⊆ V, then V ∈ JX . (iii) If U, V ∈ JX , then U ∩ V ∈ JX . (iv) U −1 ∈ JX whenever U ∈ JX . (v) If U, V ∈ JX , then U ◦ V ∈ JX whenever the composition exists. If JX is a uniform convergence structure on X, then we refer to the pair (X, JX ) as a uniform convergence space. Remark 22 Let U and V filters on X × X. The filter U −1 is defined as U −1 = [{U −1 : U ∈ U}]. If the filters U and V satisfies ∀ U ∈U : ∀ V ∈V : , U ◦ V 6= ∅ then the filter U ◦ V exists, and is defined as U ◦ V = [{U ◦ V : U ∈ U, V ∈ V}]. As mentioned, every uniformity JX on a set X induces a uniform convergence structure JUX on X through U ∈ JUX ⇔ UX ⊆ U. (2.68) However, not every uniform convergence structure is of the form (2.68). In this regard, we may recall [26] that a uniform convergence structure JX on X induces a convergence structure λJX on X, called the induced convergence structure, through ∀ x∈X : ∀ F a filter on X : . F ∈ λJX (x) ⇔ F × [x] ∈ JX (2.69) The induced convergence structure need not be a completely regular topology on X. In fact, every convergence structure λX which is reciprocal, that is, ∀ x, y ∈ X : , λX (x) = λX (y) or λX (x) ∩ λX (y) = ∅ 5 In the original definition, this condition was replaced with the stronger one ‘[∆] ∈ JX ’. This definition, however, results in a category which is not cartesian closed. CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 73 is induced by a uniform convergence structure through (2.69). Indeed, the family of filters JλX on X × X, called the associated uniform convergence structure, specified by ∃ x1 , ..., xk ∈ X : ∃ F1 , ..., Fk filters on X : U ∈ JλX ⇔ (2.70) 1) Fi ∈ λX (xi ) for i = 1, ..., k 2) (F1 × F1 ) ∩ ... ∩ (Fk × Fk ) ⊆ U is a complete uniform convergence structure that induces the convergence structure λX whenever it is reciprocal. In particular, every Hausdorff convergence structure is induced by the associated uniform convergence structure (2.70). This is clearly far more general than the completely regular topologies that are induced by uniformities. The usual uniform concepts, namely, that of Cauchy filter, uniformly continuous function, completeness and completion extend in the natural way to uniform convergence spaces. In particular, if X and Y are uniform convergence spaces, then a mapping f :X→Y is uniformly continuous whenever ∀ U ∈ JX : (f × f ) (U) ∈ JY Furthermore, a uniformly continuous mapping is a uniformly continuous embedding if it is injective, and has a uniformly continuous inverse f −1 on the subspace f (X) of Y , and a uniformly continuous embedding is a uniformly continuous isomorphism if it is also surjective. The Cauchy sequences on a uniform convergence space are defined in the obvious way. A filter F on X is a Cauchy filter if and only if F × F ∈ JX , (2.71) and the uniform convergence space X is complete if every Cauchy sequence converges to some x ∈ X. Weil’s result on the completion of uniform spaces may be reproduced within the more general setting of uniform convergence spaces. In particular, Wyler [161] showed that with every Hausdorff uniform convergence space X one may associate a complete, Hausdorff uniform convergence space X ] , and a uniformly continuous embedding ιX : X → X ] so that ιX (X) is dense in X ] . Furthermore, the completion X ] satisfies the universal property that, given any other complete, Hausdorff uniform convergence space Y , and a uniformly continuous mapping f : X → Y, CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 74 there is a uniformly continuous mapping f ] : X] → Y ] so that the diagram ιX X - X] @ @ @ @ f] f @@ (2.72) @ @ R @ Y commutes. This completion is unique up to uniformly continuous isomorphism. As we have mentioned, a uniformity on X is a particular case of a uniform convergence structure. In particular, with each such uniformity UX on X we may associate a uniform convergence structure JUX on X in a natural way through (2.68). As such, for each Hausdorff uniform space X, we may construct two completions, namely, the uniform space completion of Weil [160], and the uniform convergence space completion of Wyler [161]. These two completions, let us denote them by ] ] XW e and XW y , respectively, are not the same. In particular, the Wyler completion ] XW y will typically not be a uniform space [161]. This apparent irregularity simply ] means that the Weil completion XW e will not satisfy the universal property enjoyed by the Wyler completion. More precisely, if X is a uniform space, and given a complete, Hausdorff uniform convergence space Y which is not a uniform space, and a uniformly continuous mapping f :X→Y then we will in general not be able to find a uniformly continuous extension f ] : ] ] XW e → Y of f to XW e . Closely related to the concept of a uniform convergence space is that of a Cauchy space. Roughly speaking, a Cauchy structure on a set X is supposed to be the family of Cauchy filters associated with a given uniform convergence structure, and were introduced in an attempt to axiomatize the concept of Cauchy filter. These structures were axiomatized by Keller [80] as follows. Definition 23 Consider a set X, and a family CX of filters on X. Then CX is a Cauchy structure if it satisfies the following conditions: CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 75 (i) [x] ∈ CX for each x ∈ X. (ii) If F ∈ CX and F ⊆ G, then G ∈ CX . (iii) If F, G ∈ CX and F ∨ G exists, then F ∩ G ∈ CX . The pair (X, CX ) is called a Cauchy space. Every uniform convergence structure JX on a set X induced a unique Cauchy structure CJX on X through F ∈ CJX ⇔ F × F ∈ JX . (2.73) Conversely, every Cauchy structure CX on X is induced by a uniform convergence structure. In particular, the family of filters JCX on X × X defined as ∃ F1 , ..., Fk ∈ CX : U ∈ JCX ⇔ (2.74) (F1 × F1 ) ∩ ... ∩ (Fk × Fk ) constitutes a uniform convergence structure on X. Furthermore, the Cauchy structure induced by (2.74) is exactly CX . It should be noted that, as is the case for uniform spaces, two different uniform convergence structures may induce the same Cauchy structures. In particular, the uniform convergence structure (2.74) is the largest uniform convergence structure, with respect to inclusion, that induces a given Cauchy structure. That is, if JX is a uniform convergence structure that induces a given Cauchy structure through (2.69), then JX ⊆ JCX . (2.75) It should be noted that several different ‘completions’ may be associated with a given Cauchy space, each with different properties [128]. Which completion is used is rather a matter of convenience. The Wyler completion is the unique completion that satisfies the universal extension property (2.72). However, the Wyler completion does not preserve compatibility with algebraic structures. In this regard, if X is a convergence vector space, then it carries in a natural way a uniform convergence structure. The underlying set associated with the Wyler completion X ] of X is a vector space in a straight forward way. However, in contradistinction with the Weil completion of a uniform space, the uniform convergence structure on the Wyler completion is not the one induced through by the algebraic structure on X ] . This is also true for convergence groups [61]. Throughout the current work, we will always use the Wyler completion. The role of uniform spaces, and more generally uniform convergence spaces, in analysis is well known. In particular, and most relevant to the current investigation, is the role played by such structures in the study of linear and nonlinear PDEs, CHAPTER 2. TOPOLOGICAL STRUCTURES IN ANALYSIS 76 as explained in Chapter 1, see also Chapter 6. Furthermore, these structures also appear in connection with the construction of compact spaces that contain a given topological space. In this regard, we may recall that Brummer and Hager [32] showed that, essentially, the Stone-Čech compactification of a completely regular topological space is in fact the completion of X equipped with a suitable uniformity. Chapter 3 Real and Interval Functions 3.1 Semi-continuous Functions The classical analysis of the nineteenth century was concerned mainly with sufficiently smooth functions, and in particular analytic functions. However, it is well known that even relatively simply constructions involving only continuous functions give rise to functions that are no longer continuous, see for instance [106] for an excellent historical overview of these and related matters. In this regard, consider the following example. Example 24 The pointwise limit of a sequence of continuous functions need not be continuous. Indeed, consider the sequence (un ) of continuous functions from R to R given by 0 if x ≤ − n1 nx+1 if |x| < n1 . (3.1) un (x) = 2 1 if x ≥ n1 Clearly the sequence (un ) of continuous functions converges pointwise to the function 0 if x < 0 1 if x = 0 u (x) = 2 1 if x > 0 which has a discontinuity at x = 0. Remark 25 It should be noted that the pointwise limit of a sequence of continuous functions, such as that constructed in Example 24, may have discontinuities on a dense subset of the domain of convergence. In particular, in the case of real valued functions of a real variable, the limit will in general be continuous only on a residual set, which may have dense complement, see for instance [121] and [77]. 77 CHAPTER 3. REAL AND INTERVAL FUNCTIONS 78 Furthermore, such discontinuous functions appear also in the applications of mathematics. In this regard, we may recall the discontinuous solution (1.12) to the nonlinear conservation law (1.4) to (1.5) which turns out to model the highly relevant physical phenomenon of shock waves. An important class of such discontinuous functions arises in a natural way as a particular case of Example 24, namely, the semi-continuous real valued functions. The concept of a semi-continuous function generalizes that of a continuous function, and was first introduced by Baire [13] in the case of real valued functions of a real variable. Subsequently, the definition was extended to real valued functions on an arbitrary topological space, as well as functions with more general ranges, notably extended real valued functions, and set valued functions. In this case we will restrict our attention to the situation which is most relevant to the current investigation, namely, the case of extended real valued functions u:X→R where X is a topological space, and R = R ∪ {±∞} is the extended real line. Recall that a function u : X → R is continuous at x ∈ X if and only if ∀ >0: ∃ V ∈ Vx : y ∈ V ⇒ |u (x) − u (y) | < (3.2) ∀ >0: ∃ V ∈ Vx : . y ∈ V ⇒ u (x) − < u (y) < u (x) + (3.3) which is equivalent to The concept of semi-continuity of a function u : X → R at x ∈ X is obtained by considering each of the two inequalities in (3.3) separately. In this regard, the standard definitions of lower semi-continuous function, and an upper semi-continuous function, respectively, are as follows. Definition 26 A function u : X → R is lower semi-continuous at x ∈ X whenever ∀ M < u (x) : ∃ V ∈ Vx : y ∈ V ⇒ M < u (y) or u (x) = −∞. If u is lower semi-continuous at every point of X, then it is lower semi-continuous on X. Definition 27 A function u : X → R is upper semi-continuous at x ∈ X whenever ∀ M > u (x) : ∃ V ∈ Vx : y ∈ V ⇒ M > u (y) or u (x) = +∞. If u is upper semi-continuous at every point of X, then it is upper semi-continuous on X. CHAPTER 3. REAL AND INTERVAL FUNCTIONS 79 Clearly, each continuous function u : X → R is both lower semi-continuous and upper semi-continuous on X. Conversely, a function u : X → R is both lower semicontinuous and upper semi-continuous on X, then it is continuous on X. Important examples of discontinuous semi-continuous functions include the indicator function χA of a set A ⊆ X, that is, 1 if x ∈ A χA (x) = (3.4) 0 if x ∈ /A If A is open, then χA is lower semi-continuous, and if A is closed, then χA is upper semi-continuous. Semi-continuous functions appear as fundamental objects in analysis and its applications. In particular, such functions play a basic role in optimization theory, since semi-continuous functions have certain useful properties that fail in the case of continuous functions. In this regard, we may recall that the supremum of a set of continuous functions need not be continuous. In this regard, consider the following. Example 28 Consider the set {uα : α > 1} of continuous, real valued functions on R, where each function uα is defined by 0 if |x| ≥ α |x|−α uα (x) = if α < |x| < 1 α−1 −1 if |x| ≤ 1 The pointwise supremum of the set {uα : α > 1} is the function u : R 3 x 7→ sup{uα (x) : α > 1} ∈ R which, in this case, is a well defined, real valued function given by if |x| > 1 0 uα (x) = −1 if |x| ≤ 1 which is not continuous on R.1 Remark 29 It should be noted that the function u constructed in Example 28 is continuous everywhere except on the closed nowhere dense set {±1} ⊂ R. In general, the pointwise supremum of a set of continuous functions may be discontinuous on a dense set. In particular, the set of points at which such a functions is continuous is in general only a residual set. 1 Similar examples may be constructed to show that the pointwise infimum of a set of continuous functions need not be continuous. CHAPTER 3. REAL AND INTERVAL FUNCTIONS 80 In contradistinction with the continuous case, semi-continuity of real valued functions is, to a certain extent, preserved when forming pointwise suprema and infima. In particular, if A is a set of lower semi-continuous functions on X, then the function u : X → R defined through u : X 3 x 7→ sup{v (x) : v ∈ A} ∈ R (3.5) is lower semi-continuous. Similarly, if B is a set of upper semi-continuous functions on X, then the function l : X 3 x 7→ inf{v (x) : v ∈ B} ∈ R (3.6) is upper semi-continuous. It should be noted that the infimum of a set of lower semi-continuous functions need not be lower semi-continuous, while the supremum of a set of upper semi-continuous functions is not always upper semi-continuous. A particular case of (3.5) and (3.6) above occurs when the sets of functions A and B consist of continuous functions. Indeed, since a continuous function is both lower semi-continuous and upper semi-continuous, it follows immediately from (3.5) and (3.6) that ∀ A ⊂ C (X) : 1) u : X 3 x 7→ sup{v (x) : v ∈ A} ∈ R lower semi-continuous 2) l : X 3 x 7→ inf{v (x) : v ∈ A} ∈ R upper semi-continuous Conversely, if X is a metric space, then for each lower semi-continuous function u : X → R we have ∃ A ⊂ C (X) : , u (x) = sup{v (x) : v ∈ A}, x ∈ X while for every upper semi-continuous function l : X → R we have ∃ B ⊂ C (X) : l (x) = inf{v (x) : v ∈ B}, x ∈ X Remark 30 In general it is not true that the pointwise supremum, respectively infimum, of a set of continuous functions is the supremum, respectively infimum, of such a set with respect to the pointwise order on C (X). Indeed, if for each n ∈ N we define the function un ∈ C (R) through 1 − n|x| if |x| < n1 , un (x) = 0 if |x| ≥ n1 then the pointwise infimum of the set {un : n ∈ N} is the function 1 if x = 0 , u (x) = 0 if x 6= 0 while the infimum of the set {un : n ∈ N} in C (R) is the function which is identically 0 on R. CHAPTER 3. REAL AND INTERVAL FUNCTIONS 81 As we have shown, semi-continuity of extended real valued functions is a far more general concept than continuity of such functions. However, such functions preserve certain useful properties of continuous functions. In particular, under certain mild assumptions on the function u, as well as the domain of definition X, the extreme value theorem for continuous functions may be generalized to semi-continuous functions. Indeed, if X is compact, and u : X → [−∞, +∞)2 is upper semi-continuous, then u attains a maximum value on X. That is, ∃ x0 ∈ X : . u (x) ≤ u (x0 ) , x ∈ X Similarly, if u : X → (−∞, +∞] is lower semi-continuous, then u attains a minimum on X ∃ x0 ∈ X : . u (x) ≥ u (x0 ) , x ∈ X Another useful property of continuous functions that extends to semi-continuous functions is concerned with the insertion of a continuous function in between two given functions. More precisely, given two continuous, real valued functions u and v on X such that u ≤ v, then it is trivial observation that ∃ w ∈ C (X) : . u≤w≤v (3.7) Indeed, we may simply take w to be the function (u + v) /2. In the nontrivial case when the continuous functions u and v in (3.7) are replaced with suitable semicontinuous functions, in particular u is upper semi-continuous and v is lower semicontinuous, (3.7) fails. However, a deep result due to Katětov [76] and Tong [152] characterizes normality of X in terms of such an insertion property. Namely, the topological space X is normal if and only if for each lower semi-continuous function v, and each upper semi-continuous function u such that u ≤ v, we have ∃ w ∈ C (X) : . u≤w≤v (3.8) This is a generalization of Hahn’s Theorem [68], which states that (3.8) holds if X is a metric space. Two fundamental operations associated with semi-continuous functions, and extended real valued functions in general, are the Baire Operators introduced by Baire [13] for real valued functions of a real variable. These operators were generalized to the case of extended real valued functions of a real variable by Sendov [146], and to functions defined on an arbitrary topological space by Anguelov [3]. Since these operators will be used extensively throughout the text, we include a detailed discussion of some of their more important properties. 2 Note that, in case the function is allows to assume the value +∞ on X, the result is trivial. CHAPTER 3. REAL AND INTERVAL FUNCTIONS 82 In this regard, we denote by A (X) the set of extended real valued functions on a topological space X. That is, A (X) = {u : X → R}. The Lower Baire Operator I and Upper Baire Operator S are, respectively, mappings I : A (X) → A (X) and S : A (X) → A (X) which are defined as I (u) (x) = sup{inf{u (y) : y ∈ V } : V ∈ Vx }, x ∈ X (3.9) S (u) (x) = inf{sup{u (y) : y ∈ V } : V ∈ Vx }, x ∈ X, (3.10) and respectively, with Vx denoting the neighborhood filter at x ∈ X. The connection of the operators (3.9) and (3.10) with semi-continuity of functions in A (X) may be seen immediately. Indeed, the Baire Operators characterize semi-continuity through u ∈ A (X) is lower semi-continuous ⇔ I (u) = u (3.11) u ∈ A (X) is upper semi-continuous ⇔ S (u) = u, (3.12) and respectively. Furthermore, for each u ∈ A (X) the function I (u) is lower semicontinuous, while the function S (u) is upper semi-continuous. From (3.11) and (3.12) it therefore follows that the operators I and S are idempotent, that is, for each u ∈ A (X) I (I (u)) = I (u) (3.13) S (S (u)) = S (u) . (3.14) and Moreover, from the definitions (3.9) and (3.10) it is clear that the operators I and S are also monotone with respect to the pointwise order on A (X) ∀ u, v ∈ A (X) : 1) I (u) ≤ I (v) . u≤v⇒ 2) S (u) ≤ S (v) (3.15) CHAPTER 3. REAL AND INTERVAL FUNCTIONS 83 It follows from (3.15) that compositions of the operators I and S must also be monotone so that ∀ u, v ∈ A (X) : 1) (I ◦ S) (u) ≤ (I ◦ S) (v) . u≤v⇒ 2) (S ◦ I) (u) ≤ (S ◦ I) (v) (3.16) Furthermore, the composite operators in (3.16) are also idempotent so that (I ◦ S) ((I ◦ S) (u)) = (I ◦ S) (u) (3.17) (S ◦ I) ((S ◦ I) (u)) = (I ◦ S) (u) (3.18) and for each u ∈ A (X). Indeed, by the obvious inequality I (u) ≤ S (u) , (3.19) as well as (3.13) to (3.14), we have (I ◦ S) ◦ (I ◦ S) (u) ≤ (I ◦ S) ◦ (S ◦ S) (u) = (I ◦ S) (u) and (I ◦ S) ◦ (I ◦ S) (u) ≥ (I ◦ I) ◦ (I ◦ S) (u) = (I ◦ S) (u) which implies (3.17). The identity (3.18) is obtained by similar arguments. A particularly useful class of semi-continuous functions is that of the normal semi-continuous functions. These functions were introduced by Dilworth [47] in connection with his attempts at obtaining a representation of the Dedekind order completion of spaces of continuous functions. In particular, Dilworth showed that the Dedekind order completion of the space Cb (X) of all bounded, real valued continuous functions on a completely regular topological space X is the set of bounded normal upper semi-continuous functions on X. A definition of normal semi-continuity for arbitrary real valued functions was given by Anguelov [3], which coincides with Dilworth’s definition in the case of bounded functions. This is the definition that we will use, and it is most simply stated in terms of the Baire Operators I and S. Namely, a real valued function u ∈ A (X) is normal lower semi-continuous whenever (I ◦ S) (u) = u (3.20) and it is normal upper semi-continuous whenever (I ◦ S) (u) = u (3.21) CHAPTER 3. REAL AND INTERVAL FUNCTIONS 84 From the time Dilworth introduced normal semi-continuous functions in the 1950s, up to very recently, there was rather limited interest in these functions. There are two possible reasons for such a lack of enthusiasm concerning normal semicontinuous functions functions. Firstly, Dilworth’s results on the order completion of spaces of continuous functions applies only in rather particular cases, namely, when the underlying space X is compact, or to the set of bounded continuous functions on a completely regular space. Moreover, it appeared to be rather difficult to extend Dilworth’s results to the case of unbounded functions. Several attempts at more general results concerning the order completion of spaces of continuous functions have resulted only in partial success, see for instance [56]. Furthermore, these functions appeared at the time not to have other significant applications. Recently, as will be discussed in more detail in Section 3.2, there has been a renewed interest in such functions. The current interest in these functions is due to two highly nontrivial applications of normal semi-continuous functions. Indeed, Anguelov [3] significantly extended Dilworth’s results on the Dedekind order completion of spaces of continuous functions, using spaces of functions that are essentially equivalent to normal semi-continuous functions. This, in turn, has lead to a significant improvement of the regularity of generalized solutions to large classes of nonlinear PDEs obtained through the Order Completion Method [8], [9]. 3.2 Interval Valued Functions The field of interval analysis, and in particular interval valued functions, is a subject that is traditionally associated with validated computing [2], [87], [146]. The central issue is to design algorithms generating bounds for exact solutions of mathematical problems, and such bounds may be represented as intervals, and interval valued functions. In this context, such functions appear in a natural way as error bounds for numerical and theoretical computations. Such interval valued functions have also been applied to approximation theory [146]. In fact, Sendov [146] introduced the concept of a Hausdorff continuous interval valued function in connection with Hausdorff approximations of real functions of a real argument. However, recent applications of interval valued functions to diverse mathematical fields previously considered to be unrelated to interval analysis, have lead to a renewed interest in these functions, as well as to a new point of view regarding them. Namely, the possible structures, of whatever appropriate kind (topological, algebraic or order theoretic), with which spaces of interval valued functions may be equipped. Let us now briefly recall the basic notations and concepts involved. In this regard, we denote by 1) a, a ∈ R IR = a = [a, a] 2) a ≤ a the set of extended, closed real intervals. The subset of IR consisting of finite and closed real intervals is denoted by IR. By identifying a point a ∈ R with the CHAPTER 3. REAL AND INTERVAL FUNCTIONS 85 degenerate interval [a, a] ∈ IR, we may consider the extended real line as a subset of IR R ⊂ IR. (3.22) The usual total order on R can be extended to a partial order on IR in several different ways. A particularly useful order was defined by Markov [104] though 1) a ≤ b (3.23) a≤b⇔ 2) a ≤ b Our main interest in this section is in functions whose values are extended real intervals. For a given set X we denote by A (X) the set of such interval valued functions on X, that is, A (X) = {u : X → IR}. (3.24) A convenient representation of interval valued functions is as pairs of point valued functions. In particular, with every u ∈ A (X) we associate the pair of point valued functions u, u ∈ A (X) such that u (x) = [u (x) , u (x)], x ∈ X. (3.25) Through the identification of points in R with intervals in IR, we may consider the set A (X) of extended real valued functions on X as a subset of A (X). Since the partial order (3.23) on IR extends the usual total order on R, the pointwise order on A (X), specified as ∀ x∈X : u≤v⇔ , (3.26) u (x) ≤ v (x) extends the pointwise order on A (X). Several concepts of continuity of interval valued functions, defined on a topological space X, have been introduced in the literature [6], [146]. Here we may recall the concepts of Hausdorff continuity, Dilworth continuity and Sendov continuity, all of which are closely linked to the concepts of semi-continuity of extended real valued functions discussed in Section 3.1. These continuity concepts are conveniently formulated in terms of extensions of the Baire operators (3.9) and (3.10). In this regard, we note that these operators act in a natural way also on interval valued functions. Indeed, for u ∈ A (X) we may define the operators I and S as I (u) (x) = sup{inf{u (y) : y ∈ V } : V ∈ Vx }, x ∈ X (3.27) S (u) (x) = inf{sup{u (y) : y ∈ V } : V ∈ Vx }, x ∈ X (3.28) and CHAPTER 3. REAL AND INTERVAL FUNCTIONS 86 which clearly coincides with (3.9) and (3.10), respectively, if u is point valued. It should be noted that the extended Baire operators (3.27) and (3.28) produce point valued functions. That is, I : A (X) → A (X) and S : A (X) → A (X) As such, and in view of (3.19), the Graph Completion Operator F : A (X) 3 u 7→ [I (u) , S (u)] ∈ A (X) , (3.29) introduced by Sendov [146] for finite interval valued functions of a real argument, see also [3], is a well defined mapping. From (3.15), (3.13) and (3.14) it follows that the operator F is both monotone and idempotent, that is, for all u ∈ A (X) F (F (u)) = F (u) , (3.30) u ≤ v ⇒ F (u) ≤ F (v) . (3.31) and for all u, v ∈ A (X) An important class of interval valued functions, namely, the Sendov continuous (Scontinuous) functions, is defined as the fixed points of the operator F . That is, u ∈ A (X) is S-continuous if and only if F (u) = u. (3.32) These functions, introduced by Sendov, play an important role in the theory of Hausdorff approximations [146]. The class of functions of main interest in the current context is that of the Hausdorff continuous functions, which are defined as follows [146], see also [3]. Definition 31 A function u ∈ A (X) is Hausdorff continuous (H-continuous) if F (u) = u and for each v ∈ A (X) ∀ x∈X : v (x) ⊆ u (x) (3.33) ⇒ F (v) = u. The set of all H-continuous functions on X is denoted as H (X). (3.34) CHAPTER 3. REAL AND INTERVAL FUNCTIONS 87 Remark 32 A remark on the particular meanings of conditions (3.33) and (3.34) is appropriate. Essentially, the condition (3.33) may be interpreted as a continuity requirement. Indeed, in view of the definition (3.29) of the Graph Completion Operator and the characterization (3.12) of semi-continuity in terms of the Baire operators, the condition (3.33) simply states that the function u may be represented by a pair of semi-continuous functions, namely, u is lower semi-continuous, and u is upper semi-continuous. The second condition (3.34) is in fact a minimality condition with respect to inclusion. That is, u is smallest S-continuous function, with respect to inclusion, included in u. H-continuous functions were first introduced by Sendov [146] for functions of a real variable in connection with applications to the theory of Hausdorff approximations. Recently [3], the definition was extended to functions defined on arbitrary topological spaces. The set H (X) of H-continuous functions inherits the partial order (3.26). With this order, the set H (X) is a complete lattice, that is, ∀ A ⊆ H (X) : ∃ u0 , l0 ∈ H (X) : . 1) u0 = sup A 2) l0 = inf A (3.35) Furthermore, the supremum and infimum in (3.35) may be described in terms of the pointwise supremum and infimum ϕ : X 3 x 7→ sup{u (x) : u ∈ A} ∈ R and ψ : X 3 x 7→ inf{u (x) : u ∈ A} ∈ R, respectively. Indeed, see [3], we have the following characterization of u0 and l0 in (3.35): u0 = F (I (S (ϕ))) , l0 = F (S (I (ψ))) To date, three important classes of H-continuous functions have been identified. These are the finite H-continuous functions Hf t (X), the bounded H-continuous functions Hb (X) and the nearly finite H-continuous functions Hnf (X). These classes of functions are defined as ∀ x∈X : Hf t (X) = u ∈ H (X) , (3.36) u (x) ∈ IR ∃ [a, a] ∈ IR : Hb (X) = u ∈ Hf t (X) u (x) ⊆ [a, a], x ∈ X (3.37) CHAPTER 3. REAL AND INTERVAL FUNCTIONS 88 and ∃ Γ ⊂ X closed nowhere dense : , Hnf (X) = u ∈ H (X) x ∈ X \ Γ ⇒ u (x) ∈ IR (3.38) respectively. The relevance of these classes of functions is evident from the recent and highly nontrivial applications to diverse branches of mathematics [3], [8], [9], [10]. One of the applications, namely, the the Order Completion Method for nonlinear PDEs [8], [9] is recounted in Section 1.4, while the discussion of an application to topological completion of C (X) [10] is postponed to Chapter 4. Here we proceed with a short account of the main results of Anguelov [3] concerning the Dedekind order completion of C (X). In this regard, we note that, as a simple corollary to (3.35), each of the spaces (3.36) to (3.38) is a Dedekind order complete lattice (1.96) with respect to the order induced from H (X). Furthermore, each continuous, real valued function on X is also H-continuous. Indeed, since each continuous, real valued function u is both lower semi-continuous and upper semi-continuous, it follows by (3.11) and (3.12) as well as the definition (3.29) of the Graph Completion Operator that F (u) = u. Furthermore, since u is point valued, it follows that the second condition (3.34) of Definition 31 also holds. Then the set of inclusions Cb (X) ⊆ Hb (X) , C (X) ⊆ Hf t (X) (3.39) is obvious. Moreover, in view of the fact that the order (3.26) extends the usual pointwise order on A (X), it is clear that the inclusions in (3.39) are in fact also order isomorphic embeddings (51). What remains to be verified is the denseness properties ∀ u ∈ Hf t (X) : u = sup {v ∈ C (X) : v ≤ u} (3.40) ∀ u ∈ Hb (X) : . u = sup {v ∈ Cb (X) : v ≤ u} (3.41) and The order denseness properties (3.41) holds for every topological space X. The property (3.40) is valid whenever X is a metric space, or when X is completely regular and satisfies ∀ u ∈ Hnf (X) : ∃ v ∈ C (X) : . u≤v In particular, in each of these cases, we may construct for each u ∈ Hf t (X) sequences (λn ) and (µn ) of continuous functions on X so that ∀ n∈N: λn ≤ λn+1 ≤ u ≤ µn+1 ≤ µn (3.42) CHAPTER 3. REAL AND INTERVAL FUNCTIONS 89 and ∀ x∈X : 1) sup{λn (x) : n ∈ N} = u (x) 2) inf{µn (x) : n ∈ N} = u (x) (3.43) Now, as mentioned, the H-continuous functions are essentially equivalent to suitable normal semi-continuous functions. In this regard, consider a finite H-continuous function u on X. Then clearly, the function u = (I ◦ S) (u) (3.44) is real valued and normal lower semi-continuous. Conversely, for a given real valued normal lower semi-continuous function v on X, the function v = F (v) is a finite H-continuous function. Indeed, the mapping F : N Lf t (X) → Hf t (X) (3.45) is a bijection, with N Lf t (X) the space of all real valued normal lower semi-continuous functions on X. Moreover, given normal lower semi-continuous functions u and v on X, we have u ≤ v ⇔ F (u) ≤ F (v) (3.46) so that F is in fact an order isomorphism. Remark 33 The same construction may be reproduced for normal upper semicontinuous functions. In this case, one may consider the mapping F : N U f t (X) → Hf t (X) instead of (3.45), where N U f t (X) denotes the set of real valued normal upper semicontinuous functions. The concept of an interval valued function, and in particular that of an Hcontinuous function, proves to be a useful concept in so far as the representations of certain extensions of spaces of continuous functions are concerned. Here we discussed only the extension through order, but recently the rational completion of the ring C (X) and other related extensions of this space were also characterized as spaces of H-continuous functions [4]. Moreover, and as will be discussed in more detail in Chapter 4, the convergence vector space completion of C (X) with respect to the so called order convergence structure may be constructed as the set Hf t (X). However, in view of the fact that interval valued functions are relatively unknown in analysis, our preference in the current investigation is rather to use the equivalent representation via normal lower semi-continuous functions. Chapter 4 Order and Topology 4.1 Order, Algebra and Topology Order, together with algebra and topology, are the most fundamental concepts in modern mathematics, and the rich mathematical field that may be broadly called ‘analysis’ is an example of the great power and utility of mathematical concepts that arise as combinations of these basic notions. In this regard, we may recall the theory of Operator Algebras, initiated by von Neumann, where all three these basic concepts are involved. In this section, we will discuss two examples of such fruitful interactions between the basic trio of order, algebra and topology which is relevant to the current investigation. Namely, the theory of ordered algebraic structures, in particular Riesz Spaces, and that of ordered topological spaces. Many of the interesting spaces in analysis, and in particular linear functional analysis, are equipped with partial orders in a natural way. Some of the most prominent examples include the following. The space C (X) of continuous, real valued functions on a topological space X ordered in the usual pointwise way ∀ u, v ∈ C (X) : ∀ x∈X : , u≤u⇔ u (x) ≤ v (x) and the space M (Ω) of real valued measurable functions on a measure space (Ω, Λ, µ), modulo functions that are almost everywhere equal, equipped with the almost everywhere pointwise order ∀ u, v ∈ M (Ω) ∃ E ⊂ Ω, µ (E) = 0 : u≤v⇔ u (x) ≤ v (x) , x ∈ Ω \ E as well as its important subspaces of p-integrable functions. These are defined as Z p Lp (Ω) = u ∈ M (Ω) : |u (x) | dx < ∞ Ω 90 CHAPTER 4. ORDER AND TOPOLOGY 91 for p ≤ 1 < ∞, and ∃ C>0: L∞ (Ω) = u ∈ M (Ω) ∃ E ⊂ Ω, µ (E) = 0 : u (x) ≤ C, x ∈ Ω \ E The spaces described above are all examples of Riesz spaces, also called vector lattices. A Riesz space is a real vector space L equipped with a partial order in such a way that L is a lattice and ∀ u, v, w ∈ L : ∀ α ∈ R, α ≥ 0 : . 1) u ≤ v ⇒ u + w ≤ v + w 2) u ≥ 0 ⇒ αu ≥ 0 (4.1) Riesz spaces were introduced independently, and more or less simultaneously, by F. Riesz [130], [131] , L. V. Kantorovitch [72], [73], [74] and H. Freudenthal [60]. The simple requirements on the compatibility of the order on L and its algebraic structure in (4.1) has some immediate and rather unexpected consequences. In this regard, we may recall [101] that any Riesz space L is a fully distributive lattice. That is, ∀ A⊂L: ∀ u∈L: inf{sup{u, v} : v ∈ A} = inf{u, sup A} provided that sup A exists in L. The utility and power of methods from Riesz space theory, when applied to problems in analysis and other parts of mathematics, is well documented, although often not fully appreciated. In this regard, we may recall Freudenthal’s Spectral Theorem [60], see also [101]. Roughly speaking, Freudenthal’s Theorem states that each u ∈ L may be approximated, in a suitable sense, by ‘step functions’. The importance of this result is clear from its applications. These include the following three fundamental results, namely, the Radon-Nykodym Theorem in measure theory, the Spectral Theorem for Hermitian operators and normal operators in Hilbert space, and the Poisson formula for harmonic functions on an open circle, see [101]. The theory of Riesz spaces has been extensively developed, and most of the basic questions have by now been settled for nearly twenty years. The most recent theoretical program was initiated by A. C. Zaanen in the later part of the 1980s. The aim of this program was to reprove the major results of the theory in elementary terms, without reference to the often highly complicated representation theorems for Riesz spaces. This area of research has culminated in 1997 in the excellent introductory text [164]. Current interest in Riesz spaces stem from their many applications, in particular in connection with stochastic calculus and martingale theory, see for instance [45], [91], [92] and [95]. CHAPTER 4. ORDER AND TOPOLOGY 92 A useful generalization of the concept of a Riesz space is that of a lattice ordered group, called an l-group for short. In this regard, recall [29] that an l-group is a commutative group G = (G, +, ) equipped with a lattice order ≤ such that ∀ f, g, h ∈ G : f ≤g⇒ 1) f + h ≤ g + h 2) h + f ≤ h + g . (4.2) Such an l-group is called Archimedean whenever ∀ f, g ≥ 0 : , nf ≤ g, n ∈ N ⇒ f = 0 (4.3) with 0 the group identity in G. Loosely speaking, the condition (4.3) means that the group G does not contain any ‘infinitely small’ of ‘infinitely large’ elements. Lastly, an element e ∈ G is an order unit if e > 0 and ∀ g∈G,g≥0: ∃ n∈N: , g ≤ ne while e is a weak order unit whenever e > 0 and ∀ g ∈ G, g ≥ 0 : . inf{e, g} = 0 ⇒ g = 0 A useful aspect of the theory of Archimedean l-groups is that, each such group admits a representation as a set of continuous, real valued functions on a suitable topological space. In particular, if G admits a distinguished weak order unit, then we may associate with G a completely regular topological space, call it XG , and an l-group homomorphism TG : G → C (XG ) . That is, every commutative l-group with distinguished weak order unit is isomorphic to an l-subgroup of the space of continuous functions on certain completely regular topological space XG , see [162] as well as [16] and the references cited there. Conversely, we may associate with each completely regular topological space X a commutative l-group with weak order unit GX , namely, the l-group C (X). This amounts to a relationship between the categories W of Archimedean l-groups with weak order unit, and the category R of completely regular topological spaces. As such, one may study the latter category through the more simple one W . On a partially ordered set X there are several ways to define a topology in terms of the order on X, see for instance [29]. Among these, we may mention the order topology [82], the interval topology [83], the Scott topology and the Lawson topology [66]. Such topologies turn out to be interesting from the point of view of applications. In this regard, we may recall that the Scott topology and the CHAPTER 4. ORDER AND TOPOLOGY 93 Lawson topology play an important role in the theory of continuous lattices, and its applications to theoretical computer science [66]. More generally, given a topological space X equipped with a partial order, we may require the order to be compatible with the topology in a suitable sense. One particularly useful requirement is that the mappings ∨ : X × X 3 (x, y) 7→ sup{x, y} ∈ X (4.4) ∧ : X × X 3 (x, y) 7→ inf{x, y} ∈ X (4.5) and be continuous on their respective domains of definition, with respect to the product topology on X × X. Such a space is termed an ordered topological space. The order topology, and the interval topology on a lattice X, see for instance [53], are both examples of ordered topological spaces. It turns out that, in this case, many properties of the topology τ on X may be characterized in terms of the partial order. This may represent a dramatic simplification in so far as topological concepts may be described at the significantly more simple level of order. 4.2 Convergence on Posets As mentioned in Chapter 2, we may associate with every topology τ on a set X a notion of convergence with respect to τ . However, one may also define several useful and important concepts of convergence that cannot be associated with a topology. In particular, we may, for example, associate with each element x of a set X a collection σ (x) of sequences on X, which is interpreted to mean that a sequence (xn ) converges to x if and only if (xn ) belongs to σ (x). Such an association of sequences with points is in general not determined by a topology. In the previous section, we discussed the idea of defining a topology in terms of a partial order. In this case, and in view of the above remark, we may associate with the partial order the convergence induced by the topology. More generally, we may define a notion of convergence of sequences in terms of a given partial order on a set X. Indeed, several such useful notions of convergence on partially ordered sets have been introduced in the literature, see for instance [29], [50], [101] and [105]. It often happens that these notions of convergence cannot be associated with any topology, see for instance [154]. One of the most well known such examples of convergence defined through a partial order that is, in general, not associated with any topology, is order convergence of sequences, see for instance [101] or [105]. In this regard, see also Example 9, for a given partially ordered set X, a sequence (xn ) order converges to x ∈ X whenever ∃ (λn ) , (µn ) ⊂ X : 1) n ∈ N ⇒ λn ≤ λn+1 ≤ xn+1 ≤ µn+1 ≤ µn 2) sup {λn : n ∈ N} = x = inf {µn : n ∈ N} (4.6) CHAPTER 4. ORDER AND TOPOLOGY 94 In general, there is no topology τ on X such that, for each x ∈ X, and every sequence (xn ) on X, (xn ) order converges to x if and only if (xn ) converges to x with respect to τ , see for instance [154]. Indeed, the following two properties of convergent sequences in topological spaces [107] fail to hold for the order convergence of sequences. Namely, the Divergence Axiom If every subsequence of (xn ) contains a subsequence which converges to x ∈ X, then (xn ) converges to x. and the Axiom of Iterated Limits If, for every n ∈ N, the sequence (xn,m ) converges to xn ∈ X, and the sequence (xn ) converges to x in X, then there is a strictly increasing sequence of natural numbers (mn ) so that the sequence (xn,mn ) converges to x in X. (4.7) In this regard, we may recall Example 9. Note, however, that the following version of the Axiom of Iterated Limits (4.7) remains valid under rather general conditions on the partially ordered set X. Proposition 34 *[10] Let L be a lattice with respect to a given partial order ≤. 1. For every n ∈ N, let the sequence (um,n ) in L be bounded and increasing and let un = sup{um,n : m ∈ N}, n ∈ N u0n = sup{um,n : m = 1, ..., n} If the sequence (un ) is bounded from above and increasing, and has supremum in L, then the sequence (u0n ) is bounded and increasing and sup{un : n ∈ N} = sup{u0n : n ∈ N} 2. For every n ∈ N, let the sequence (vm,n ) in L be bounded and decreasing and let vn = inf{vm,n : m ∈ N}, n ∈ N vn0 = inf{vm,n : m = 1, ..., n} If the sequence (vn ) is bounded from below and decreasing, and has infimum in L, then the sequence (vn0 ) is bounded and decreasing and inf{vn : n ∈ N} = inf{vn0 : n ∈ N} Since the usual Hausdorff concept of topology is insufficient to describe the order convergence of sequences, the question arises whether or not convergence structures provide a sufficiently general context for the study of the order convergence of sequences. Several authors have addressed this issue, and similar problems arising in CHAPTER 4. ORDER AND TOPOLOGY 95 connections with other types of convergence on partially ordered sets, see for example [55] and [67]. In this regard, R Ball [15] showed that the order convergence of (generalized) sequences, on an l-group is induced by a group convergence structure. In particular, it is shown that the convergence group completion of such an l-group with respect to the mentioned group convergence structure is the Dedekind order completion of G. Papangelou [122] considered a similar problem in the setting of sequential convergence groups. Recently [10], [155], it was shown that for any σ-distributive lattice X, that is, a lattice X that satisfies ∀ (xn ) ⊆ X : ∀ x∈X : , sup{xn : n ∈ N} = x0 ⇒ sup{inf{x, xn } : n ∈ N} = inf{x, x0 } there is a convergence structure on X that induces the order convergence of sequences. In particular, the convergence structure λo , specified as ∀ x∈X : ∀ F a filter on X: ∃ (λn ) , (µn ) ⊂ X : (4.8) 1) n ∈ N ⇒ λn ≤ λn+1 ≤ x ≤ µn+1 ≤ µn F ∈ λo (x) ⇔ 2) sup {λn : n ∈ N} = x = inf {µn : n ∈ N} 3) [{[λn , µn ] : n ∈ N}] ⊆ F is first countable, Hausdorff, and induces the order convergence of sequences. In particular, if X is an Archimedean Riesz space, then the convergence structure (4.8) is a vector space convergence structure, and X a convergence vector space. In this case, we may construct the convergence vector space completion of X, which is the Dedekind σ-completion of X [155], equipped with the order convergence structure (4.8). In particular, in case X is the set C (Z) of continuous functions on a metric space Z, then the completion may be constructed as the set Hf t (Z) of finite Hcontinuous interval valued functions on Z. Chapter 5 Organization of the Thesis 5.1 Objectives of the Thesis The aim of this thesis is to develop a general and type independent theory concerning the existence and regularity of generalized solutions of systems of nonlinear PDEs. In this regard, our point of departure is the Order Completion Method [119], which is discussed in Section 1.4. In particular, this includes, as a first and basic step, a reformulation of the Order Completion Method in the context of uniform convergence spaces. That is, the construction of generalized solutions to a nonlinear PDE (1.100) as an element of the Dedekind completion of the space Mm T (Ω) is interpreted in terms of the Wyler completion of a suitable uniform convergence space. Such a recasting of the Order Completion Method in terms of uniform convergence spaces allows for the application of convergence theoretic techniques to problems relating to the structure and regularity of generalized solutions, techniques that may turn out to be more suitable to the mentioned problems that the order theoretic methods involved in the Order Completion Method. The recasting of the Order Completion Method in the setting of uniform convergence spaces, instead of that of ordered sets and their completions, is the first and basic aim of this work. In this regard, appropriate uniform convergence spaces are introduced, and the completions of these spaces are characterized. The existence and uniqueness of generalized solutions of arbitrary continuous systems of nonlinear PDEs is proved within the context of the mentioned uniform convergence spaces. Furthermore, it is shown that these solutions may be assimilated with usual normal lower semi-continuous functions. That is, there is a natural injective and uniformly continuous mapping from the space of generalized solutions into the space of nearly finite normal lower semi-continuous maps. This provides a blanket regularity for the solutions. The regularity of the generalized solutions delivered through the Order Completion Method is dramatically improved upon in two ways. In the first place, it is shown that such generalized solutions may in fact be assimilated with functions that are smooth, up to the order of differentiability of the nonlinear partial differ96 CHAPTER 5. ORGANIZATION OF THE THESIS 97 ential operator T (x, D), everywhere except on a closed nowhere dense subset of the domain of definition of the system of equations. This result is based on the fact that any Hausdorff convergence structure admits a complete uniform convergence structure. As such, it seems unlikely that such a result can be obtained in terms of the purely order theoretic methods upon which the Order Completion Method is based. As mentioned in Section 1.4, the spaces of generalized functions delivered through the Order Completion Method are, to some extent, dependent on the particular nonlinear partial differential operator that defines the equation. For the spaces of generalized functions mentioned above, which are obtained as the Wyler completions of suitable uniform convergence spaces, this is also the case. The second major development we present here, regarding to the regularity of generalized solutions of systems of nonlinear PDEs, addresses this issue. In this regard, and in the original spirit of Sobolev, we construct spaces of generalized functions that do not depend in any way on a particular nonlinear partial differential operator. These spaces are shown to contain generalized solutions, in a suitable sense, of a large class of systems of nonlinear PDEs. This result provides some additional insight into the structure of unique generalized solutions constructed in the Order Completion Method. In particular, such a solution may be interpreted as nothing but the set of solutions in the new Sobolev type spaces of generalized functions. The solutions constructed in the Sobolev type spaces of generalized functions may be represented through their generalized partial derivatives as usual nearly finite normal lower semi-continuous functions. As such, the singularity set of such a generalized function, that is, the set of points where at least one of the generalized partial derivatives is not continuous, is a set of first Baire category. However, it should be noted that the generalized derivatives cannot be interpreted classically, that is, as usual partial derivatives, at those points where the generalized function is regular. In this regard, we show that, for a large class of equations, there are generalized solutions which are in fact classical solutions everywhere except on some closed nowhere dense set. This result is based on a suitable approximation of functions u:Ω→R that are C m -smooth everywhere except on a closed nowhere dense subset of Ω, by functions in C m (Ω), and on a result giving sufficient conditions for the compactness of a set in C m (Ω) with respect to a suitable topology. The last topic to be treated concerns initial and / or boundary value problems. The results discussed so far apply to systems of nonlinear PDEs without any additional conditions. In this regard, we show that the methods that have been developed here may be applied to initial and / or boundary value problems with only minimal modifications. In particular, we show that a large class of Cauchy problems admit solutions in the Sobolev type spaces of generalized functions. Furthermore, CHAPTER 5. ORGANIZATION OF THE THESIS 98 under only very mild assumptions regarding the smoothness of the nonlinear partial differential operator and the initial data, we show that a solution can be constructed which is in fact a classical solution everywhere except on a closed nowhere dense subset of the domain of definition of the system of equations. This result is a first in the literature. In particular, it is the first extension of the Cauchy-Kovalevskaia Theorem 2 on its own general and type independent grounds to equations that are not analytic. 5.2 Arrangement of the Material The results presented in this work are organized as follows. In Chapter 6 we obtain some preliminary results on the Wyler completion of Hausdorff uniform convergence spaces. In particular, we show that if a uniform convergence space X is a subspace of Y , then the completion X ] of X need not be a subspace of Y . However, the inclusion mapping i:X→Y extends to an injective, uniformly continuous mapping i] : X ] → Y ] . More generally, if X and Y are Hausdorff uniform convergence spaces, and ϕ:X→Y is a uniformly continuous embedding, then the unique uniformly continuous extension ϕ] : X ] → Y ] of ϕ to X ] is injective, but not necessarily an embedding. Products of uniform convergence structure are shown to Q be compatible the Wyler completion. In particular, the completion of the product i∈I Xi of a family (Xi )i∈I of Hausdorff uniform Q convergence spaces is the product i∈I Xi] of the completions Xi] of the Xi . These results are used to obtain a description of the completion of a uniform convergence space that is equipped with the initial uniform convergence structure with respect to a family of mappings ϕi : X → Xi , where each Xi is a Hausdorff uniform convergence space. In particular, we show that there is an injective, uniformly continuous mapping Y ] Φ : X] → Xi i∈I CHAPTER 5. ORGANIZATION OF THE THESIS 99 so that πi ◦ Φ = ϕ]i for each i ∈ I, with πi the projection. Chapter 7 concerns certain spaces of normal lower semi-continuous functions. In particular, we introduce the space N L (X) of all nearly finite normal lower semicontinuous functions, and the space ML (X) of nearly finite normal lower semicontinuous functions that are continuous and real valued everywhere except on a closed nowhere dense subset of X. This space also appears in connection with rings of continuous functions and their completions [56]. Some properties of the mentioned classes of functions are investigated. We introduce a uniform convergence structure on ML (X) in such a way that the induced convergence structure is the order convergence structure (4.8). The Wyler completion of ML (X) with respect to this uniform convergence structure is obtained as the set N L (X), equipped with a suitable uniform convergence structure. The uniform convergence spaces introduced in Chapter 7 form the point of departure for the construction of spaces of generalized functions in Chapter 8. In this regard, we construct the so-called pullback space of generalized functions N LT (Ω)K associated with a given system nonlinear PDEs T (x, D) u (x) = f (x) . (5.1) in Section 8.1. In Section 8.2 we introduce the Sobolev type spaces of generalized functions N Lm (Ω). These spaces are obtained as the completion of the set ∃ Γ ⊂ Ω closed nowhere dense : m ML (Ω) = u ∈ ML (Ω) u ∈ C m (Ω) equipped with a suitable uniform convergence structure. The structure of the generalized functions that are the elements of N Lm (Ω) are discussed, as well as the connection with the spaces N LT (Ω). In Section 8.3 we discuss the issue of extending a nonlinear partial differential operator to the Sobolev type spaces of generalized functions N Lm (Ω). In this regard, we show how such an operator may be defined on MLm (Ω)K , for a suitable K ∈ N. It is shown that every such operator is uniformly continuous on MLm (Ω)K , and as such it may be uniquely extended to the space N Lm (Ω)K . We also discuss the correspondence between generalized solutions in the pullback type spaces of generalized functions, and the Sobolev type spaces of generalized functions. It is shown that generalized solution to (5.1) in N Lm (Ω)K corresponds to the unique generalized solution in the pullback type space N LT (Ω), should such solutions exist. Chapter 9 addresses the issue of existence of generalized solutions in the spaces of generalized functions constructed in Chapter 8. In Section 9.1 we introduce certain basic approximation results. These include a multidimensional version of (1.110), as well as a suitable refinement of that result. Section 9.2 contains the first and basic existence and uniqueness result for generalized solutions in the pullback spaces of generalized functions N LT (Ω). This section also includes a detailed investigation of the structure of such generalized solutions. In Section 9.3 we investigate the effects CHAPTER 5. ORGANIZATION OF THE THESIS 100 of additional assumptions on the smoothness of the nonlinear partial differential operator T and the righthand term f on the regularity of generalized solutions in N LT (Ω). Here we obtain what may be viewed as a maximal regularity result for the solution in pullback type spaces of generalized functions. In particular, it is shown that if the nonlinear partial differential operator T, and the righthand term f are C k -smooth, for some k ∈ N, then the generalized solution in N LT (Ω) may be assimilated with functions in MLk (Ω)K . Section 9.4 contains existence results for generalized solutions of a large class of systems of nonlinear PDEs in the Sobolev type spaces of generalized functions. It is also shown that under additional assumptions on the smoothness of the nonlinear partial differential operator T and righthand term f, namely, that both are C k -smooth, we may obtain solutions in N Lm+k (Ω). In Chapter 10 we discuss further regularity properties of generalized solutions in Sobolev type spaces of generalized functions. Section 10.1 introduces suitable topologies on the space C m (Ω) which admit convenient conditions for a set A ⊂ C m (Ω) to be precompact. In particular, we show that any set F ⊂ C m+1 (Ω) that satisfies ∀ A ⊂ Ω compact : ∃ MA > 0 : ∀ |α| ≤ m + 1 : u ∈ A ⇒ |Dα u (x) | ≤ MA , x ∈ A is precompact in C m (Ω). This generalizes a well known result for the one dimensional case Ω ⊆ R, see for instance [49]. Using this result, it is shown in Section 10.2 that a large class of systems of nonlinear PDEs admit a generalized solution in N Lm (Ω)K that is in fact a classical solution everywhere except on a closed nowhere dense subset of Ω. Chapter 11 is dedicated to the study of a large class of initial value problems. In particular, we extend the Cauchy-Kovalevskaia Theorem 2 to systems of equations, and initial data, which are not analytic. In this regard, it is shown that such an initial value problem admits a generalized solution in N Lm (Ω)K which satisfies the initial condition in a suitable generalized sense. This is achieved through a slight modification of the methods introduced in Chapters 7 through 9. Indeed, the initial value problem is solved by essentially the same techniques that apply to the free problem. It is also shown that such a generalized solution to the Cauchy problem may be constructed which is a classical solution everywhere except on a closed nowhere dense subset of the domain of definition. Furthermore, this solution satisfies the initial condition in the usual sense. Chapter 12 contains some concluding remarks. In particular, we discuss some of the implications of the results obtained here. Directions for future research are also indicated. Part II Convergence Spaces and Generalized Functions 101 Chapter 6 Initial Uniform Convergence Spaces 6.1 Initial Uniform Convergence Structures As mentioned in Section 2.4, uniform spaces, and more generally uniform convergence spaces, appear in many important applications of topology, and in particular analysis. In this regard, the concepts of completeness and completion of a uniform convergence space play a central role. Indeed, Baire’s celebrated Category Theorem asserts that a complete metric space cannot be expressed as the union of a countable family of closed nowhere dense sets. The importance of this result is demonstrated by the fact that the Banach-Steinhauss Theorem, as well as the Closed Graph Theorem in Banach spaces follow from it. However, in many situations one deals with a space X which is incomplete, and in these cases one may want to construct the completion of X. In this regard, the main result, see for instance [63], [64] and [161] and Section 2.4, is that every Hausdorff uniform convergence space X may be uniformly continuously embedded into a complete, Hausdorff uniform convergence space X ] in a unique way such that the image of X in X ] is dense. Moreover, the following universal property is satisfied. For every complete, Hausdorff uniform convergence space Y , and any uniformly continuous mapping ϕ:X→Y 102 CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 103 the diagram ϕ X - Y @ @ @ @ @ ∃!ϕ] ι[email protected] (6.1) @ @ @ R @ X] commutes, with ϕ] uniformly continuous, and ιX the canonical embedding of X into its completion X ] . It is often not only the completion X ] of a uniform convergence space X that is of interest, but also the extension ϕ] of uniformly continuous mappings from X to X ] . In this regard, we recall that one of the major applications of uniform spaces, and recently also uniform convergence spaces, is to the solutions of PDEs. Indeed, let us consider a PDE T u = f, (6.2) with T a possibly nonlinear partial differential operator which acts on some relatively small space X of classical functions, u the unknown function, while the right hand term f belongs to some space Y . One usually considers some uniformities, or more generally uniform convergence structures, on X and Y in such a way that the mapping T :X→Y (6.3) is uniformly continuous. It is well known that the equation (6.2), or typically some suitable extension of it, can have solutions of physical interest which, however, may fail to be classical, in the sense that they do not belong to X. From here, therefore, the particular interest in generalized solutions to (6.2). Such generalized solutions to (6.2) may be obtained by constructing the completions X ] and Y ] of X and Y , respectively. The mapping (6.3) extends uniquely to a mapping T ] : X] → Y ] (6.4) CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 104 so that the diagram X T - Y ιX ιY ? X] T] ? -Y] commutes, with ιX and ιY the uniformly continuous embeddings associated with the completions X ] and Y ] of X and Y , respectively. One may now consider the extended equation T ] u] = f (6.5) where the solutions of (6.5) are interpreted as generalized solutions of (6.2). Note that the existence and uniqueness of generalized solutions depend on the properties of the mapping T ] and the uniform convergence structure on X ] and Y ] , as opposed to the regularity of the generalized solutions, which may be interpreted as the extent to which a generalized solution exhibits characteristics of classical solutions, which depends on the properties of the elements of the space X ] . It is therefore clear that not only the completion X ] of a u.c.s. X, but also the the associated extensions of uniformly continuous mappings, defined on X, are of interest. The example given above indicates a particular point of interest. The uniform convergence structure JX on the domain X of the PDE operator T is usually defined as the initial uniform convergence structure [26] with respect to some uniform convergence structure JY on Y , and a family of mappings (ψi : X → Y )i∈I (6.6) In the case of PDEs, the mappings ψi are typically usual partial differential operators, up to a given order m. A natural question arises as to the connection between the completion of X, and the completion of Y . More generally, consider a set X, a family of mappings (ψi : X → Xi )i∈I where each Xi is a uniform convergence space. If the family (ψi )i∈I separates the points of X, then the initial uniform convergence structure on X with respect to the family of mappings (6.6) is also Hausdorff, and we may consider its completion CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 105 X ] . It appears that the issue of the possible connections between the completion of X and that of the spaces Xi , respectively, has not yet been fully explored. We aim to clarify the possible connection between the completion X ] of X, and the completions Xi] of the Xi . 6.2 Subspaces of Uniform Convergence Spaces It can easily be shown that the Bourbaki completion of a uniform space X preserves subspaces. In particular, the completion Y ] of any subspace Y of X is isomorphic to a subspace of the completion X ] of X. For uniform convergence spaces in general, and the associated Wyler completion, this is not the case. In this regard, consider the following1 . Example 35 Consider the real line R equipped with the uniform convergence structure associated with the usual uniformity on R. Also consider the set Q of rational numbers equiped with the subspace uniform convergence structure induced from R. The Wyler completion Q] of Q is the set R equipped with a suitable uniform convergence structure. As such, the inclusion mapping i : Q → R extends to a uniformly continuous bijection i] : Q] → R (6.7) Furthermore, a filter F on Q] converges to x] if and only if [V x] |Q ] ∩ [x] ] ⊆ F where V x] is the neighborhood filter in R at x] , and V x] |Q denotes its trace on Q. As such, it is clear that the neighborhood filter at x] does not converge in Q] . Therefore the mapping (6.7) does not have a continuous inverse, so that it is not an embedding. In view of Example 35, it is clear that Wyler completion does not preserve subspaces. The underlying reason for this phenomenon is is twofold. In the first place, and as mentioned in Section 2.3, the adherence operator on a convergence space is in general not idempotent. Furthermore, and perhaps more fundamentally, for a subset Y of a set X, and a filter F on X, we have the inclusion F ⊆ [F|Y ]X , with equality only holding in case Y ∈ F. In terms of the underlying set associated with the uniform convergence space completion Y ] of a subspace Y of a uniform convergence space X, we may still say something. In particular, we have the following. 1 This example was communicated to the author by Prof. H. P. Butzmann CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 106 Proposition 36 Let Y be a subspace of the uniform convergence space X. Then there is an injective, uniformly continuous mapping i] : Y ] → X ] which extends the inclusion mapping i : Y → X. In particular, i] Y ] = aX ] (ιX (Y )) . Proof. In view of the fact that the inclusion mapping i : Y → X is a uniformly continuous embedding, we obtain a uniformly continuous mapping i] : Y ] → X ] (6.8) so that the diagram Y i - ιY X ιX ? Y] i] (6.9) ? - X] commutes. To see that the mapping (6.8) is injective, consider any y0] , y1] ∈ Y ] and suppose that ] ] ] (6.10) i y0 = i y1] = x] for some x] ∈ X ] . Since ιY (Y ) is dense in Y ] there exists Cauchy filters F and G on Y such that ιY (F) converges to y0] and ιY (G) converges to y1] . From the diagram above it follows that ιX (i (F)) and ιX (i (G)) converges to x] . Therefore the filter H = ιX (i (F)) ∩ ιX (i (G)) converges to x] in X ] . Note that the filter i−1 ι−1 X (H) ] ] is a Cauchy filter onY so that ιY i−1 ι−1 must X (H) converge in Y to some y . −1 −1 −1 −1 But ιY i ιX (H) ⊆ ιY (F) and ιY i ιX (H) ⊆ ιY (G) so that ιY (F) and ] ] ιY (G) must converge to y as well. Since Y is Hausdorff it follows by (6.10) that CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 107 y0] = y1] = y ] . Therefore i] is injective. Clearly i] Y ] ⊆ aX ] (ιX (Y )). To verify the reverse inclusion, consider any x] ∈ aX ] (ιX (Y )). Then ∃ F a filter on ιX (Y ) : . [F]X ] converges to x] in X ] Then there is a Cauchy filter G on X so that ιX (G) ∩ [x] ] ⊆ [F]X ] This implies that the Cauchy filter G has a trace H = G|Y on Y , which is a Cauchy filter on Y . The result now follows by the commutative diagram (6.9). The following is an immediate consequence of Proposition 36. Corollary 37 Let X and Y be uniform convergence spaces, and ϕ : X → Y a uniformly continuous embedding. Then there exists an injective uniformly continuous mapping ϕ] : X ] → Y ] , where X ] and Y ] are the completions of X and Y respectively, which extends F . It should be noted that Wyler completion is minimal, with respect to inclusion, among complete, Hausdorff uniform convergence on the set X ] , as demonstrated in the following proposition. We may obtain this as an easy conseqeunce of the universal property (6.1) and Corollary 37. Proposition 38 Consider a Hausdorff uniform convergence space X. For any complete, Hausdorff uniform convergence space X0] that contains X as a dense subspace, there is a bijective and uniformly continuous mapping ι]X,0 : X ] → X0] . Proof. Let X0] be a complete, Hausdorff uniform convergence space that contains X as a dense subspace, so that the inclusion mapping i : X 3 x 7→ x ∈ X0] (6.11) is a uniformly continuous embedding. It follows from Corollary 37 that the mapping (6.11) extends to an injective uniformly continuous mapping i : X ] 3 x] 7→ i] x] ∈ X0] . (6.12) It remains to verify that the mapping (6.12) is surjective. In this regard, consider any x]0 ∈ X0] . Since X is dense in X0] , there is a Cauchy filter F on X so that [F]X ] 0 converges to x]0 in X0] . As such, there exists x] ∈ X ] so that [F]X ] converges to x] . Therefore i] ([F]X ] ) converges to x]0 in X0] so that i] x] = x]0 . This completes the proof. For a subspace Y of a Hausdorff uniform convergence space X, this leads to the following. CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 108 Corollary 39 Let Y be a subspace of the Hausdorff uniform convergence space X. The uniform convergence structure on the Wyler completion Y ] of Y is the finest complete, Hausdorff uniform convergence structure on aX ] (Y ) so that Y is contained in it as a dense subspace. Remark 40 It should be noted, and as mentioned in Section 2.4, that the completion of a convergence vector space [65], the completion of a convergence group [61], and the Wyler completion [161] of a uniform convergence space are in general all different. Indeed, the Wyler completion is typically not compatible with the algebraic structure of a convergence group or convergence vector space [26], [65], while the convergence group completion of a convergence vector space does in general not induce a vector space convergence structure [21]. 6.3 Products of Uniform Convergence Spaces In this section we consider the completion of the product of a family of uniform convergence spaces. In contradistinction with subspaces of a uniform convergence space, products of uniform convergence spaces are well behaved with respect to the Wyler completion. In particular, it is well known [161] that the product of complete, Hausdorff uniform convergence structures are complete and Hausdorff. Furthermore, we obtain the following result. Theorem 41 Let (Xi )i∈I be a family of Hausdorff uniform convergence spaces, and let X denote their Cartesian product equipped with the product uniform convergence structure. Then the completion X ] of X is the product of the completions Xi] of the Xi . Q Proof. First note that i∈I Xi] is complete. For every i, let ιXi : Xi → Xi] be the uniformly continuous embedding associated with the completion Xi] of Xi . Define Q ] the mapping ιX : X → Xi through ιX : x = (xi )i∈I 7→ (ιXi (xi ))i∈I For each i, let πi : X → Xi be the projection. Since each ιXi is injective, so is ιX . Moreover, we have U ∈ JX ⇒ (πi × πi ) (U) ∈ JXi ⇒ (ιXi × ιXi ) ((πi × πi ) (U)) ∈ JX] i Y ⇒ (ιXi × ιXi ) ((πi × πi ) (U)) ∈ JQ] i∈I ⇒ (ιX × ιX ) (U) ∈ JQ] Q where JQ] denotes the product uniform convergence structure on i∈I Xi] . Hence ιX is uniformly continuous. Similarly, if the filter V on ιX (X) × ιX (X) belongs to CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 109 the subspace uniform convergence structure, then −1 (πi × πi ) (V) ∈ JX] i ⇒ ι−1 Xi × ιXi ((πi × πi ) (V)) ∈ JXi Y −1 ⇒ ι−1 Xi × ιXi ((πi × πi ) (V)) ∈ JX ⇒ i∈I ι−1 X × ι−1 X (V) ∈ JX so that ι−1 Hence ιX is a uniformly continuous embedding. X is uniformly continuous. Q ] That ιX (X) is dense in i∈I Xi follows by the denseness of ιXi (Xi ) in Xi] , for each i ∈ I. The extension property of uniformly continuous mappings into a complete u.c.s. follows in the standard way. 6.4 Completion of Initial Uniform Convergence Structures In view of the fact that the Wyler completion of uniform convergence spaces do not, in general, preserve subspace, initial structures are not invariant under the formation of completions. That is, if X carries the initial uniform convergence structure with respect to a family of mappings (ψi : X → Xi )i∈I into u.c.s.s Xi , then the completion X ] of X does not necessarily carry the initial uniform convergence structure with respect to ψi] : X ] → Xi] i∈I where ψi] denotes the uniformly continuous extension of ψi] to X ] . In this regard, one can only obtain a generalization of Proposition 36. The first, and in fact straight forward, result in this regard is the following. Proposition 42 Suppose that X is equipped with the initial uniform convergence structure with respect to a family of mappings (ϕi : X → Xi )i∈I , (6.13) where each uniform convergence space Xi is Hausdorff, and the family of mappings (8.22) separates the points on X. Then each mapping ϕi extends uniquely to a uniformly continuous mapping ϕ]i : X ] → Xi] (6.14) and the uniform convergence structure on X ] is finer than the initial uniform convergence structure with respect to the mappings (6.14). CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 110 Proof. It follows by the universal property (6.1) that each of the mappings (8.22) extend to a uniformly continuous mapping (6.14). From the continuity of the mappings (6.14) it follows that the uniform convergence structure on X ] is finer than the initial uniform convergence structure with respect to the mapping of mappings (6.14). In connection with the actual uniform convergence structure on the set X ] , we cannot in general make a stronger claim. However, it is possible to describe the structure of the set X ] itself in terms of the completions of the Xi . In this regard, we first note that the uniform convergence structure Q on X may be described in terms of the product uniform convergence structure on i∈I Xi . Proposition 43 For each i ∈ I, let Xi be a Hausdorff uniform convergence space, with uniform convergence structure JXi . Let the uniform convergence space X carry the initial uniform convergence structure JX with respect to the family of mappings (ψi : X → Xi )i∈I Assume that (ψi )i∈I separates the points of X. Then there exists a unique uniformly continuous embedding Y Xi (6.15) Ψ:X→ i∈I such that, for each i ∈ I, the diagram ψi - Xi X @ @ @ @ @ Ψ @ πi @ @ @ R @ Q Xi commutes, with πi the projection. Proof. Define the mapping Ψ as Ψ : X 3 x 7→ (ψi (x))i∈I ∈ Y i∈I Xi (6.16) CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 111 Since the family (ϕi )i∈I seperates the points of X, the mapping (6.16) is injective. Furthermore, the diagram ψi X - Xi @ @ @ @ @ Ψ@ πi (6.17) @ @ @ R @ Q Xi commutes for every i ∈ I. Suppose that U ∈ JX . Then ∀ i∈I : (ψi × ψi ) (U) ∈ JXi : and hence ∀ i∈I : . (πi × πi ) (Ψ × Ψ) (U) ∈ JXi : Therefore (Ψ × Ψ) (U) ∈ JQ , which is the product uniform convergence structure, so that Ψ is uniformly continuous. Q Q Let V ∈ JQ be a filter on i∈I Xi × i∈I Xi with a trace on Ψ (X) × Ψ (X). Then ∀ i∈I : a) (πi × πi ) (V) ∈ JXi b) W ∈ (πi × πi ) (V) ⇒ W ∩ (ψi (X) × ψi (X)) 6= ∅ so that ∀ i∈I : (ψi × ψi ) ((Ψ−1 × Ψ−1 ) (V)) ⊇ (πi × πi ) (V) Form the definition of an initial uniform convergence structure, and in particular the product uniform convergence structure, it follows that (Ψ−1 × Ψ−1 ) (V) ∈ JX . Hence Ψ is a uniformly continuous embedding. The uniqueness of the mapping Ψ is obvious from the construction of Ψ. The following now follows as an immediate consequence of Proposition 43. Theorem 44 For each i ∈ I, let Xi be a Hausdorff uniform convergence space, with uniform convergence structure JXi . Let the uniform convergence space X carry the initial uniform convergence structure JX with respect to the family of mappings (ψi : X → Xi )i∈I CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 112 Assume that (ψi )i∈I separates the points of X. Then there exists a unique injective, uniformly continuous mapping Y ] Ψ] : X ] → Xi (6.18) i∈I such that, for each i ∈ I, the diagram ψi] X] - X] i @ @ @ @ @ Ψ] @ πi @ @ @ R @ Q Xi] commutes, with πi the projection, and ψi] the unique extension of ψi to X ] . Proof. The result follows by Proposition 36, Theorem 41 and Proposition 43. Within the context of nonlinear PDEs, as explained in Section 6.1, Theorem 44 may be interpreted as a regularity result. Indeed, consider some space X ⊆ C ∞ (Ω) of classical, smooth functions on an open, nonempty subset Ω of Rn . Equip X with the initial uniform convergence structure JX with respect to the family of mappings D α : X → Y , α ∈ Nn (6.19) where Y is some space of functions on Ω that contains Dα (X) for each α ∈ Nn . In view of Theorem 44, the mapping D : X 3 u → (Dα u) ∈ Y N (6.20) is a uniformly continuous embedding, and as such (6.20) extends to an injective uniformly continuous mapping D] : X ] 3 u → (Dα u) ∈ Y ]N (6.21) CHAPTER 6. INITIAL UNIFORM CONVERGENCE SPACES 113 so that the diagram Dα] X] - Y] @ @ @ @ @ ] [email protected] πi (6.22) @ @ @ R @ Y ]N commutes. Here Dα] : X ] → Y ] , α ∈ Nn are the uniformly continuous extension of the mappings (6.19). As such, each generalized function u] ∈ X ] may be identified with D] u] ∈ Y ]N . The above interpretation of the completion of a uniform convergence space which is equipped with an initial structure is central to the theory of the solutions of nonlinear PDEs presented in the chapters to follow. In particular, we employ exactly the construction (6.22) to obtain our first and basic regularity properties for the solutions of such systems of equations. Chapter 7 Order Convergence on ML (X) 7.1 Order Convergence and the Order Completion Method We may recall from Section 1.4 that our approach to the enrichment of the Order Completion Method [119] is motivated by the fact that the process of taking the supremum of a subset A of a partially ordered set X is essentially a process of approximation. Such approximation-type statements, and in particular the process of forming the Dedekind completion of a partially ordered set, may be reformulated in terms of topological type structures, which may turn out to be more general than the usual Hausdoff-Kuratowski-Bourkabi concept of topology. In this regard, and as mentioned in Chapter 4, there are several useful modes of convergence on a partially ordered set which are defined in terms of the partial order, see for instance [29], [101] and [124]. A particularly relevant concept is that of the order convergence of sequences defined on a partially ordered set through (4.6). In general, and as mentioned in Section 4.2, there is no topology on a partially ordered set X that induces the order convergence of sequences. That is, for a partially ordered set X there is in general no topology τ on X such that the τ convergent sequences are exactly the order convergent sequences. However, the more general context of convergence structures and convergence spaces provides an adequate setting within which to describe the order convergence of sequences. Namely, if X is a σ-distributive lattice, then the convergence structure (4.8) induces the order convergence of sequences. In particular, and as is discussed in Section 4.2, every Archimedean vector lattice is fully distributive, and hence σ-distributive. In this case the convergence structure (4.8) is a vector space convergence structure, and as such it is induced by a uniform convergence structure [26]. In this case, the Cauchy filters may be defined through F a Cauchy filter on X ⇔ F − F ∈ λo (0) . Furthermore, the convergence vector space completion of an Archimedean vector lattice X, equipped with the order convergence structure λo , may be constructed as 114 CHAPTER 7. ORDER CONVERGENCE ON ML (X) 115 the Dedekind σ-completion X ] of X, equipped with the order convergence structure. If X is order separable, the completion of X is in fact its Dedekind completion. In the particular case when X = C (Y ), with Y a metric space, the convergence vector space completion is the set Hf t (X) of finite Hausdorff continuous functions on Y , which is the Dedekind completion of C (Y ). Let us now consider the possibility of applying the above results to the problem of solving nonlinear PDEs through the Order Completion Method. In this regard, consider a nonlinear PDE of the form (1.100), and the associated mapping T : Mm (Ω) → M0 (Ω) The Order Completion Method is based on the abundance of approximate solutions to (1.100), which are elements of Mm (Ω), and in general one cannot expect these approximations to be continuous, let alone sufficiently smooth, on the whole of Ω. Moreover, the space Hf t (Ω) does not contain the space M0 (Ω). On the other hand, the space M0 (Ω) is an order separable Archimedean vector lattice [119], and therefore one may equip it with the order convergence structure. The completion of this space will be its Dedekind completion M0 (Ω)] , as desired. However, there are several obstacles to applying the theory of the order convergence structure to the Order Completion Method. If one equips Mm (Ω) with the subspace convergence structure, then the nonlinear mapping T is not necessarily continuous. Moreover, the quotient space Mm T (Ω) is not a linear space, so that the completion process for convergence vector spaces does not apply. It is therefore necessary to develop a nonlinear convergence theoretic model for the Dedekind completion of M (Ω). 7.2 Spaces of Lower Semi-Continuous Functions We may recall from Section 3.1 that the notion of a normal lower semi-continuous function, respectively normal upper semi-continuous function, was introduced by Dilworth [47] in connection with the Dedekind completion of spaces of continuous functions. Dilworth introduced the concept for bounded, real valued functions. Subsequently the definition was extended to locally bounded functions [6]. The definition extends in a straightforward way to extended real valued functions. In particular, a function u : X → R, with X a topological space, is normal lower semi-continuous at x ∈ X whenever (I ◦ S) (u) (x) = u (x) (7.1) It is called normal lower semi-continuous on X if it is normal lower semi-continuous at every x ∈ X. Here I and S are the Lower- and Upper Baire Operators defined through (3.9) and (3.10), respectively. Note that if a function u is real valued and continuous at a point x ∈ X, then it is also normal lower semi-continuous at x. CHAPTER 7. ORDER CONVERGENCE ON ML (X) 116 In analogy with H-continuous interval valued functions, we call a normal lower semi-continuous function u nearly finite whenever the set {x ∈ X : u (x) ∈ R} is open and dense in X. We denote the space of all nearly finite normal lower semicontinuous functions by N L (X). The space N L (X) is ordered in a pointwise way through ∀ u, v ∈ N L(X) : ∀ x∈X : u≤v⇔ u (x) ≤ v (x) (7.2) The space N L (X) is the fundamental space upon which a convergence theoretic approach to nonlinear PDEs will be constructed. In this regard, the following basic order theoric properties of this space are fundamental. Theorem 45 The space N L (X) is Dedekind complete. Moreover, if A ⊆ N L (X) is bounded from above, and B ⊆ N L (X) is bounded from below, then sup A = (I ◦ S) (φ) inf B = (I ◦ S ◦ I) (ϕ) where φ : X 3 x 7→ sup{u (x) : u ∈ A} and ϕ : X 3 x 7→ inf{u (x) : u ∈ B} Proof. Consider a set A ⊂ N L (X) which is bounded from above. Then it follows by (3.17) and (3.15) that the function u0 = (I ◦ S) (ϕ) is nearly finite and normal lower semi-continuous. Furthermore, u0 is an upper bound for A, that is, ∀ u∈A: . u ≤ u0 Now suppose that u0 is not the least upper bound of A. That is, we assume ∃ w ∈ N L (X) : ∀ u∈A: . u ≤ w < u0 (7.3) ϕ (x) ≤ w (x) , (7.4) Then it follows that CHAPTER 7. ORDER CONVERGENCE ON ML (X) 117 so that (3.15) and (3.17) imply u0 ≤ w which contradicts (7.3). The existence of a greatest lower bound follows in the same way. We now proceed to establish further properties of the space N L (X) concerning the pointwise order (7.2). In this regard, the following result generalizes the well known property of continuous functions. If D is a dense subset of X, then ∀ u, v ∈ C (X) : ∀ x∈D: . ⇒u≤v u (x) ≤ v (x) Proposition 46 Consider any u ∈ N L (X). Then there is a set R ⊆ X such that X \ R is of First Baire Category and u is continuous at every x ∈ R. If v ∈ N L (X) and D ⊆ X is dense in X, then ∀ x∈D : ⇒u≤v u (x) ≤ v (x) Proof. Consider any u ∈ N L (X). Then u is lower semi-continuous on X, and real valued on some open and dense subset D of X. Fix > 0. We claim ∃ Γ ⊂ D closed nowhere dense : . 0 < S (u) (x) − u (x) < , x ∈ D \ Γ (7.5) In this regard, suppose that there is a nonempty, open subset V of D such that S (u) (x) ≥ u (x) + , x ∈ V. Since u is lower semi-continuous, so is the function u + . As such, it follows by (3.11) and (3.20) that u (x) ≥ u (x) + , x ∈ V, which is a contradiction. As such, the set of points {x ∈ D : 0 < S (u) (x) − u (x) < } is dense in D. That it is open follows by the semi-continuity of the functions u and S (u). Then we have ! [ u (x) = S (u) (x) , x ∈ R = D \ Γ1 . n n∈N CHAPTER 7. ORDER CONVERGENCE ON ML (X) 118 As such, and in view of (3.12) and (3.14) it follows that u is upper semi-continuous at every point of R. Since u is both lower semi-continuous and upper semi-continuous on R, it is continuous on R. Consider now any dense subset D of X, and any u, v ∈ N L (X) so that u (x) ≤ v (x) , x ∈ D. Take any x ∈ X arbitrary but fixed, and neighborhoods V1 and V2 of x. Since D is dense in X there is some z0 ∈ V1 ∩ V2 ∩ D so that inf{u (y) : y ∈ V1 } ≤ u (z0 ) ≤ v (z0 ) ≤ sup{v (y) : y ∈ V2 }. Since V1 and V2 are chosen independent of each other, and that x is arbitrary, we have I (u) (x) ≤ S (v) (x) , x ∈ X. From (3.13), (3.15) and (3.17) it follows that u = I (I (u)) ≤ I (S (v)) = v which completes the proof. Recall from Section 4.2 that the order convergence structure may be defined on an arbitrary lattice. However, this convergence structure induces the order convergence of sequences only on σ-distributive lattices. As such, the following property is essential. Proposition 47 The space N L (X) is a fully distributive lattice. Proof. tion Consider any u, v ∈ N L (X), and the normal lower semi-continuous funcw = (I ◦ S) (ϕ) where ϕ : X → R is the pointwise supremum of u and v, namely, ϕ : X 3 x 7→ sup{u (x) , v (x)}. Since both u and v are nearly finite, there is some open and dense subset D of X such that ϕ is finite on D. Note that both u and v must be locally bounded on D. As such, it follows that ∀ x∈D: ∃ V ∈ Vx : . ∃ M >0: −M < ϕ (y) < M , y ∈ V CHAPTER 7. ORDER CONVERGENCE ON ML (X) 119 Therefore we must have −M ≤ w (x) ≤ M , x ∈ V so that w is nearly finite. It now follows by Theorem 45 that w = sup{u, v}. The existence of inf{u, v} follows in the same way. Now let us show that N L (X) is distributive. Consider a set A ⊂ N L (X) such that sup A = u0 For v ∈ N L (X) we must show u0 ∧ v = sup{u ∧ v : u ∈ A} (7.6) Suppose that (7.6) fails for some A ⊂ N L (X) and some v ∈ N L (X). That is, ∃ w ∈ N L (X) : u ∈ A ⇒ u ∧ v ≤ w < u0 ∧ v (7.7) Clearly, u0 , v > w so that there is some u ∈ A such that w is not larger than u. In view of Proposition 46 ∃ V ⊆ X nonempty, open : x ∈ V ⇒ w (x) < u (x) (7.8) From (3.13), (3.14), (3.15) and Proposition 45 it follows that (v ∧ u) (x) > w (x) , x ∈ V. Hence (7.7) cannot hold. This completes the proof. It is a well known fact that a pointwise bounded subset of C (X) may fail to be uniformly bounded, even when X is compact. Furthermore, such a pointwise bounded set may not even be bounded with respect to the pointwise order on C (X). In this regard, consider the following. Example 48 Consider the sequence (un ) of continuous, real valued functions on R, defined through n − n2 |x − n1 | if |x − n1 | < n1 un (x) = 0 if |x − n1 | ≥ n1 Clearly the sequence (un ) is pointwise bounded on R. Indeed, ∀ x∈R: ∃ Nx ∈ N : n ≥ Nx ⇒ un (x) = 0 CHAPTER 7. ORDER CONVERGENCE ON ML (X) 120 which validates our claim. However, in view of 1 un = n, n ∈ N n it follows that there cannot be a continuous, real valued function u on R so that un ≤ u for each n ∈ N. Within the more general setting of spaces of normal lower semi-continuous functions there is quite a strong relationship between pointwise bounded sets and order bounded sets. In particular, we have the following. Proposition 49 Consider a set A ⊂ N L (X) that satisfies ∃ R ⊆ X a residual set : ∀ x∈R: . sup{u (x) : u ∈ A} < +∞ (7.9) If X is a Baire space, then ∃ µ ∈ N L (Ω) : . u ∈ A ⇒ u (x) ≤ µ (x) , x ∈ X (7.10) If X is a metric space, then ∃ Γ ⊂ X closed nowhere dense : ∃ µ ∈ N L (X) : 1) µ ∈ C (X \ Γ) 2) u ∈ A ⇒ u ≤ µ (7.11) The corresponding result for sets bounded from below is also true. Proof. Consider the function ϕ : X → R defined through ϕ (x) = sup{u (x) : u ∈ A}, x ∈ X. Since each u ∈ A is lower semi-continuous, it follows that ϕ is lower semi-continuous on X. Moreover, ϕ is finite on the residual set R. Set µ (x) = (I ◦ S) (ϕ) (x) . In view of the fact that I ◦ S is idempotent it follows that µ is normal lower semicontinuous and u ≤ µ for every u ∈ A. We claim that µ is nearly finite. Suppose this were not the case, so that ∃ V ⊂ X nonempty, open : . x ∈ V ⇒ µ (x) = +∞ Then it follows by the inequality I (S (ϕ)) ≤ S (ϕ) (7.12) CHAPTER 7. ORDER CONVERGENCE ON ML (X) 121 that ∀ x∈V : S (ϕ) (x) = +∞ Then, in view of (3.10), we have ∀ ∀ ∀ ∃ M >0: x0 ∈ V : W ∈ V x0 : xM ∈ V ∩ W : ϕ (xM ) > M Since ϕ is lower semi-continuous, we must have ∃ DM ⊆ V open and dense in V : x ∈ DM ⇒ ϕ (x) > M Therefore \ ϕ (x) = +∞, x ∈ R0 = DM M ∈N Since ϕ is finite on R, it follows that R ∩ V ⊆ V \ R0 Since X is a Baire space, V is a Baire space in the subspace topology, and R ∩ V is residual in V . But R0 is clearly also residual in V so that R ∩ V is of first Baire category, which is a contradiction. Therefore (7.12) cannot hold. Therefore µ is nearly finite, and we have proven (7.10). The validity of (7.11) follows by (3.42). The following related result provides a useful connection between pointwise convergence and order convergence in N L (X). Proposition 50 Let X be a Baire space. Consider a decreasing sequence (un ) in N L (X) which is bounded from below. Let u = inf{un : n ∈ N} ∈ N L (Ω) . Then the following holds: ∀ >0: ∃ Γ ⊆ Ω closed nowhere dense : ∃ N ∈ N : x ∈ Ω \ Γ ⇒ un (x) − u (x) < , n ≥ N The corresponding statement for increasing sequences is also true. CHAPTER 7. ORDER CONVERGENCE ON ML (X) Proof. 122 Take > 0 arbitrary but fixed. We start with the set ∀ n∈N: C= x∈X , un , u continuous at x the complement of which is a set of first Baire category. Hence C is dense. In view of Proposition 46, the set of points ∃ N ∈ N : C = x ∈ C un (x) − u (x) < , n ≥ N must be dense in C. From the continuity of u and the un on C it follows that ∀ x0 ∈ C : ∃ δx0 > 0 : x ∈ C, kx − x0 k < δx0 ⇒ x ∈ C Since C is dense in X, the result follows. The set Cnd (X) of all functions u : X → R that are continuous everywhere except on some closed nowhere dense subset of X, that is, ∃ Γu ⊂ X closed nowhere dense : u ∈ Cnd (X) ⇔ (7.13) u ∈ C (X \ Γu ) plays a fundamental role in the theory of Order Completion [119], as discussed in Section 1.4. In particular, one considers the quotient space M (X) = Cnd (X) / ∼, where the equivalence relation ∼ on Cnd (X) is defined by ∃ Γ ⊂ X closed nowhere dense : 1) x ∈ X \ Γ ⇒ u (x) = v (x) u∼v⇔ (7.14) 2) u, v ∈ C (X \ Γ) The canonical partial order on M (X) is defined as ∀ u ∈ U, v ∈ V : ∃ Γ ⊂ X closed nowhere dense : U ≤V ⇔ 1) u, v ∈ C (X \ Γ) 2) u (x) ≤ v (x) , x ∈ X \ Γ . An order isomorphic representation of the space M (X), consisting of normal lower semi-continuous functions, is obtained by considering the set ∃ Γ ⊂ X closed nowhere dense : ML (X) = u ∈ N L (X) u ∈ C (X \ Γ) The advantage of considering the space ML (X) rather than M (X) is that the elements of ML (X) are actual point valued functions on X, in contradistinction with the elements of M (X) which are equivalence classes of functions. In particular, CHAPTER 7. ORDER CONVERGENCE ON ML (X) 123 the singularity set Γ associated with a function u ∈ ML (X), as well as the values of u on Γ are fully specified. Hence the value u (x) of u ∈ ML (X) are completely determined. That is, for each x ∈ X, and every u ∈ ML (X), the value u (x) of u at x is a well defined element of R, which is not the case for an equivalence class in M (X). Indeed, for every U ∈ M (X), and each x ∈ X one may find u1 , u2 ∈ U so that u1 (x) 6= u2 (x) . Proposition 51 The mapping IS : M (X) 3 U 7→ (I ◦ S) (u) ∈ ML (X) , u ∈ U (7.15) is a well defined order isomorphism. Proof. First note that, in view of (3.17) and (7.13), the mapping IS does indeed take values in ML (X). Now we show that the mapping IS is well defined. That is, we show that IS (U ) does not depend on the particular representation u ∈ U that is used in (7.15). In this regard, consider some U ∈ M (X) and any u, v ∈ U . Let Γ ⊂ X be the closed nowhere dense set associated with u and v through (7.14). Since Γ is closed, it follows by (3.9), (3.10) and (7.13) that (I ◦ S) (u) (x) = (I ◦ S) (v) (x) , x ∈ X \ Γ (7.16) Since X \ Γ is dense in X, it follows by Proposition 46 that equality holds on the whole of X. It is obvious that the mapping IS is surjective. Indeed, each element u ∈ ML (X) generates an equivalence class U in ML (X), so that (7.15) and (3.17) implies that IS (U ) = u. To see that it is injective, consider any U, V ∈ M (X). From (7.14) it follows that ∀ u ∈ U, v ∈ V : ∃ A ⊆ X nonempty, open : ∃ >0: 1) x ∈ A ⇒ u (x) < v (x) − 2) u, v ∈ C (A) so that IS (U ) (x) < IS (V ) (x) − , x ∈ A It remains to verify ∀ U, V ∈ M (X) : U ≤ V ⇔ IS (U ) ≤ IS (V ) The implication U ≤ V ⇒ IS (U ) ≤ IS (V ) CHAPTER 7. ORDER CONVERGENCE ON ML (X) 124 follows by (3.20), Proposition 46 and (7.13). Conversely, suppose that IS (U ) ≤ IS (V ) for some U, V ∈ M (X). The result now follows in the same way as the injectivity of IS . This completes the proof. The following is immediate. Corollary 52 The space ML (X) is a fully distributive lattice. 7.3 The Uniform Order Convergence Structure on ML (X) As a consequence of Corollary 52 one may define the order convergence structure λo on the space ML (X). The order convergence structure induces the order convergence of sequences on ML (X) and is Hausdorff, regular and first countable. In order to define a uniform convergence structure on ML (X) that induces the order convergence structure, we introduce the following notation. For any open subset V of X, and any subset F of ML (X), we denote by F|V the restriction of F to V . That is, ∃ w∈F : F|V = v ∈ ML (V ) x ∈ V ⇒ w (x) = v (x) Definition 53 Let Σ consist of all nonempty order intervals in ML (X). Let Jo denote the family of filters on ML (X) × ML (X) that satisfy the following: There exists k ∈ N such that ∀ i = 1, ..., k : ∃ Σi = (Ini ) ⊆ Σ : i 1) In+1 ⊆ Ini , n ∈ N 2) ([Σ1 ] × [Σ1 ]) ∩ ... ∩ ([Σk ] × [Σk ]) ⊆ U (7.17) where [Σi ] = [{I : I ∈ Σi }]. Moreover, for every i = 1, ..., k and V ∈ τX one has ∃ T ui ∈ ML (X) : i n∈N In|V = {ui }|V or T i n∈N In|V =∅ (7.18) Before we proceed to establish that the family Jo of filters on ML (X) × ML (X) does indeed constitute a uniform convergence structure, let us recall the following useful technical lemma. Lemma 54 *[26] Let X be a set. (i) Consider filters U1 , ...Un and V1 , ..., Vm on X × X. Then the filter (U1 ∩ ... ∩ Un ) ◦ (V1 ∩ ... ∩ Vm ) CHAPTER 7. ORDER CONVERGENCE ON ML (X) 125 exists if and only if Ui ◦ Vj exists for some i = 1, ..., n and j = 1, ..., m. In this case, we have \ (U1 ∩ ... ∩ Un ) ◦ (V1 ∩ ... ∩ Vm ) = {Ui ◦ Vj : Ui ◦ Vj exists}. (ii) Consider filters F1 , F2 , G1 and G2 on X. Then (F1 × F2 ) ◦ (FG × G2 ) exists if and only if F1 ∨ G2 exists. If this is true, then (F1 × F2 ) ◦ (FG × G2 ) = G1 × F2 . Theorem 55 The family Jo of filters on ML (X) × ML (X) constitutes a uniform convergence structure. Proof. The first four axioms of Definition 21 are trivially fulfilled, so it remains to verify ∀ U, V ∈ Jo : U ◦ V exists ⇒ U ◦ V ∈ Jo (7.19) In this regard, take any U, V ∈ Jo such that U ◦ V exists, and let Σ1 , ..., Σk and Σ01 , ..., Σ0l be the collections of order intervals associated with U and V, respectively, through Definition 53. Set Φ = {(i, j) : [Σi ] ◦ [Σ0j ] exists} Then, by Lemma 54 (i) it follows that \ U ◦ V ⊇ {([Σi ] × [Σi ]) ◦ ([Σj ] × [Σj ]) : (i, j) ∈ Φ}. (7.20) Now (i, j) ∈ Φ if and only if ∀ m, n ∈ N : i Im ∩ Inj 6= ∅ For any (i, j) ∈ Φ, set Σi,j = (Ini,j ) where, for each n ∈ N Ini,j = [inf Ini ∧ inf Inj , sup Ini ∨ sup Inj ] Now, using (7.20), we find \ \ U ◦ V ⊇ {[Σi ] × [Σj ] : (i, j) ∈ Φ} ⊇ {[Σi,j ] × [Σi,j ] : (i, j) ∈ Φ} Clearly each Σi,j satisfies 1) of (7.17). Since N L (X) is fully distributive, see Proposition 47, (7.18) also holds. This completes the proof. An important fact to note is that the uniform order convergence structure Jo is defined solely in terms of the order on ML (X) and the topology on X. This CHAPTER 7. ORDER CONVERGENCE ON ML (X) 126 is unusual for a uniform convergence structure on a function space. Indeed, for a space of functions F (X, Y ), defined on some set X, and taking values in Y , one defines the uniform convergence structure either in terms of the uniform convergence structure on Y , or in terms of a convergence structure on F (X, Y ) which is suitably compatible with the algebraic structure of the space. Indeed, a convergence vector space carries a natural uniform convergence structure, where the Cauchy filters are determined by the linear structure. That is, F a Cauchy filter ⇔ F − F → 0 (7.21) The motivation for introducing a uniform convergence structure that does not depend on the algebraic structure of the set ML (X) comes from nonlinear PDEs, and in particular the Order Completion Method [119]. As mentioned in Chapter 1, as well as in Section 7.1, such linear topological structures are inappropriate when it comes to the highly nonlinear phenomena inherent in the study of nonlinear PDEs. Recall from Section 2.4 that every uniform convergence structure induces a convergence structure through (2.69). In the case of the uniform order convergence structure, this induced convergence structure on ML (X) may be characterized as follows. Theorem 56 A filter F on ML (X) belongs to λJo (u), for some u ∈ ML (X), if and only if there exists a family ΣF = (In ) of nonempty order intervals on ML (X) such that 1) In+1 ⊆ In , n ∈ N ∀ V T∈τ : 2) n∈N In|V = {u}|V and [ΣF ] ⊆ F. Proof. Let the filter F converge to u ∈ ML (X). Then, by (2.71), [u] × F ∈ Jo . Hence by Definition 53 there exist k ∈ N and Σi ⊆ Σ for i = 1, ..., k such that (7.17) through (7.18) are satisfied. Set Ψ = {i : [Σi ] ⊂ [u]}. We claim \ F⊃ [Σi ] (7.22) i∈Ψ Take a set A ∈ ∩i∈Ψ [Σi ]. Then for each i ∈ Ψ there is a set Ai ∈ [Σi ] such that A ⊃ ∪i∈Ψ Ai . For each i ∈ {1, ..., k} \ Ψ choose a set Ai ∈ [Σi ] with u ∈ / ML (X) \ Ai . Then (A1 × A1 ) ∪ ... ∪ (Ak × Ak ) ∈ ([Σ1 ] × [Σ1 ]) ∩ ... ∩ ([Σk ] × [Σk ]) ⊂ F × [u] and so there is a set B ∈ F such that B × {u} ⊂ (A1 × A1 ) ∪ ... ∪ (Ak × Ak ) CHAPTER 7. ORDER CONVERGENCE ON ML (X) 127 If w ∈ B then (u, w) ∈ Ai × Ai for some i. Since u ∈ Ai , we get i ∈ Ψ and so w ∈ ∪i∈Ψ Ai . This gives B ⊆ ∪i∈Ψ Ai ⊆ A and so A ∈ F so that (7.22) holds. Clearly, for each i ∈ Ψ, we have ∀ V ∈τ : i ∩n∈N In|V = {u}|V (7.23) Writing each Ini ∈ Σi in the form Ini = [λin , µin ], we claim sup{λin : n ∈ N} = u = inf{µin : n ∈ N} Suppose this were not the case. Then there exists v, w ∈ ML (X) such that λ n ≤ v < w ≤ µn , n ∈ N Then, in view of Proposition 46, there is some nonempty V ∈ τ such that v (x) < w (x) , x ∈ V which contradicts (7.18). Since ML (X) is fully distributive, the result follows upon setting 1) λn = inf{λin : i ∈ Ψ} ΣF = [λn , µn ] 2) µn = sup{µin : i ∈ Ψ} The converse is trivial. The following is now immediate Corollary 57 Consider a filter F on ML (X). Then F ∈ λJo (u) if and only if F ∈ λo (u). Therefore ML (X) is a uniformly Hausdorff uniform convergence space. In particular, a sequence (un ) on ML (X) converges to u if and only if (un ) order converges to u. 7.4 The Completion of ML (X) This section is concerned with the construction of the completion of the uniform convergence space ML (X). In this regard, recall that the completion of the convergence vector space C (X), equipped with the order convergence structure, is the set of finite Hausdorff continuous functions on X, see Section 4.3 and [10]. This space is order isomorphic to the set of all finite normal lower semi-continuous functions. Note, however, that functions u ∈ ML (X) need not be finite everywhere, but may, in contradistinction with functions in C (X), assume the values ±∞ on any closed nowhere dense subset of X. Hence we consider the space N L (X) of nearly finite normal lower semi-continuous functions on X. Following the results in Section 7.3, we introduce the following uniform convergence structure on N L (X). CHAPTER 7. ORDER CONVERGENCE ON ML (X) 128 Definition 58 A filter U on N L (Ω) × N L (Ω) belongs to the family Jo] whenever, for some positive integer k, we have the following: ∀ i = 1, ..., k : ∃ (λin ) , (µin ) ⊂ ML0 (Ω) : ∃ ui ∈ N L (Ω) : 1) λin ≤ λin+1 ≤ µin+1 ≤ µin , n ∈ N 2) sup{λin : n ∈ N} = ui = inf{µin : n ∈ N} Tk i i i i 3) i=1 (([Σ ] × [Σ ]) ∩ ([u ] × [u ])) ⊆ U (7.24) Here Σi = {Ini : n ∈ N} with Ini = {u ∈ ML0 : λin ≤ u ≤ µin }. The following now results by the same arguments and techniques used in Section 7.3, notably those employed in the proof of Theorems 55 and 56. Theorem 59 The family Jo] of filters on N L (X)×N L (X) is a Hausdorff uniform convergence structure. Theorem 60 A filter F on N L (X) belongs to λJo] if and only if ∃ (λn ) , (µn ) ⊂ ML (X) : 1) λn ≤ λn+1 ≤ µn+1 ≤ µn , n ∈ N , 2) sup{λn : n ∈ N} = u = inf{µn : n ∈ N} 3) [{In : n ∈ N}] ⊆ F where In = {v ∈ ML (X) : λn ≤ v ≤ µn }. We now proceed to show that N L (X) is the completion of ML (X). That is, we show that the following three conditions are satisfied: • The uniform convergence space N L (X) is complete • ML (X) is uniformly isomorphic to a dense subspace of N L (X) • Any uniformly continuous mapping ϕ on ML (X) into a complete, Hausdorff uniform convergence space Y extends uniquely to a uniformly continuous mapping ϕ] from N L (X) into Y . Proposition 61 The uniform convergence space N L (X) is complete. Proof. Clearly Jo] is simply the uniform convergence structure associated with the convergence structure described in Theorem 60. Therefore it is complete. Theorem 62 Let X be a metric space. Then the space N L (X) is the uniform convergence space completion of ML (X). CHAPTER 7. ORDER CONVERGENCE ON ML (X) 129 Proof. First we show that ι (ML (X)) is dense in N L (X), where ι : ML (X) → N L (X) is the inclusion mapping. To see this, consider any u ∈ N L (X), and set Du = {x ∈ X : u (x) ∈ R} Since Du is open, it follows that u restricted to Du is normal lower semi-continuous. Since u is also finite on Du it follows by (3.42) that there exists a sequence (un ) of continuous functions on Du such that u (x) = sup{un (x) : n ∈ N}, x ∈ Du (7.25) Consider now the sequence (vn ) = ((I ◦ S) (u0n )) where un (x) if x ∈ Du u0n (x) = 0 if x ∈ / Du Clearly vn (x) = un (x) for every x ∈ Du . We claim u = sup{vn : n ∈ N} (7.26) If (7.26) does not hold, then ∃ v ∈ N L (X) : n ∈ N ⇒ vn ≤ v < u But then, in view of Proposition 46, and the fact that Du is open and dense, there exists an open and nonempty set W ⊆ Du such that ∀ x∈W : n ∈ N ⇒ un (x) ≤ v (x) < u (x) which contradicts (7.25). Therefore (7.26) must hold. The sequence (ι (vn )) is clearly a convergent sequence in N L (X) so that ι (ML (X)) is dense in N L (X). Now let us show that the inclusion mapping is a uniformly continuous embedding. In this regard, it is sufficient to consider a filter [ΣF ] where ΣF = {In = [λn , µn ] : n ∈ N} is a family of nonempty order intervals in ML (X) that satisfies 1) of (7.17) as well as (7.18). We claim ∃ u ∈ N L (X) : . sup{λn : n ∈ N} = u = inf{µn : n ∈ N} (7.27) Since the sequence (λn ) is bounded from above, and the sequence (µn ) is bounded from below, it follows from the Dedekind completeness of N L (X), Theorem 45, that ∃ u, v ∈ N L (X) : . sup{λn : n ∈ N} = u ≤ v = inf{µn : n ∈ N} (7.28) CHAPTER 7. ORDER CONVERGENCE ON ML (X) 130 To see that (7.27) holds, we proceed by contradiction. Suppose that u 6= v. Then, by Proposition 46, we have ∃ W ⊂ X nonempty and open : . x ∈ W ⇒ u (x) < v (x) (7.29) We may assume that both u and v are finite on W . Since v is lower semi-continuous, ∀ x∈W : 1) ϕ ∈ C (W ) v (x) = sup ϕ (x) 2) ϕ (x) ≤ v (x) , x ∈ W Clearly, there is a function ϕ ∈ C (W ), and a nonempty open set A ⊆ W such that u (x) < ϕ (x) < v (x) , x ∈ A Applying the Katětov-Tong Theorem to the continuous function ϕ and the lower semi-continuous function v, one finds a function ψ ∈ C (A) such that u (x) < ϕ (x) < ψ (x) < v (x) , x ∈ A which contradicts (7.17). Therefore ι is uniformly continuous. That ι−1 is uniformly continuous follows immediately from (7.24). The extension property for uniformly continuous mappings on ML (X) follows in the standard way. Note that in the above proof, we actually showed that N L (X) is the Dedekind completion of ML (X). Hence the uniform order convergence structure provides a nonlinear topological model for the process of taking the Dedekind completion of ML (X). In view of Proposition 51, this extends a previous result of Anguelov and Rosinger [9] on the Dedekind completion of M (X). However, it should be noted that Theorem 62 is in fact more general than the result in [9]. Indeed, along with the uniform convergence space completion of ML (X) we obtain a class of mappings, namely, uniformly continuous mappings into any Hausdorff uniform convergence space Y , that can be extended uniquely to the completion of ML (X). In contradistinction with the uniform convergence space completion constructed in Theorem 62, the Dedekind completion result in [9] allows only for the extension of order isomorphic embeddings into partially ordered sets, see Section 1.4 and [119, Appendix A]. Chapter 8 Spaces of Generalized Functions 8.1 The Spaces MLm T (Ω) The aim of the current investigation is to enrich the basic theory of Order Completion for systems of nonlinear PDEs. In this regard we have two objectives, namely, to obtain a better understanding of the possible structure of generalized solutions, and to determine to what extent we may obtain stronger regularity properties of such generalized solutions. A first step in this direction is to recast the basic existence, uniqueness and regularity results in the Order Completion Method within the context of uniform convergence spaces. Such a reformulation of the basic results of the Order Completion Method in terms of uniform convergence spaces allows for the application of tools from the theory of convergence spaces to questions related to the structure and regularity of generalized solutions. Such convergence theoretic techniques may turn out to be more suited to address these issues than the basic order theoretic techniques upon which the Order Completion Method is based. In particular, our first efforts go towards the construction of the spaces of generalized functions as the completion of suitable uniform convergence spaces, rather than the Dedekind order completion of appropriate partially ordered sets as discussed in Section 1.4. Such a reformulation of the theory of Order Completion in topological terms is motivated by the difficulties, such as those mentioned at the end of Section 1.4, involved in going beyond the basic results presented in [119] in purely order theoretic terms. A key feature of the Order Completion Method is that, with the particular nonlinear partial differential operator that defines the equation, one associates a space of generalized functions. In particular, the partial order (1.123) on the space (Ω) is defined in exactly such a way as to make the nonlinear partial differential Mm T operator compatible with the given order structures on its domain and range. It is exactly this idea of defining the structure on the domain of the operator, in this case a uniform convergence structure, in such a way as to ensure a certain compatibility with the particular nonlinear mapping involved which we exploit. 131 CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 132 Consider now a system of K possibly nonlinear PDEs, each of order at most m, of the form T (x, D) u (x) = f (x) , x ∈ Ω, (8.1) where Ω ⊆ Rn is nonempty and open. The righthand term f is assumed to be a continuous mapping f : Ω → RK , with components f1 , ..., fK . The partial differential operator T (x, D) is supposed to be defined by a jointly continuous mapping F : Ω × RM → RK (8.2) T (x, D) u (x) = F (x, ..., ui (x) , ..., Dα ui (x) , ...) , |α| ≤ m; i = 1, ..., K (8.3) through where each component u1 , ..., uK of the unknown u belongs to C m (Ω). In view of the continuity of the mapping (8.2), we may associate with the nonlinear operator T (x, D) the mapping T : C m (Ω)K → C 0 (Ω)K (8.4) defined through Tu : Ω 3 x 7→ T (x, D) u (x) ∈ RK for each u ∈ C m (Ω)K . The mapping (8.4) associated with the system of equations (8.1) extends in a canonical way to a mapping between suitable spaces of normal lower semi-continuous functions. In this regard, we introduce, for an integer l ≥ 0, the following space of nearly finite normal lower semi-continuous functions ∃ Γ ⊂ Ω closed nowhere dense : l . (8.5) ML (Ω) = u ∈ ML (Ω) u ∈ C l (Ω \ Γ) Clearly, in case l = 0, we have recovered simply the space ML (Ω). We may also note that, in contradistinction with the space C l (Ω), for l ≥ 1, of smooth functions, each of the spaces MLl (Ω) is a fully distributive lattice with respect to the pointwise order (7.2). Proposition 63 For each l ≥ 0, the space MLl (Ω) is a fully distributive lattice with respect to the pointwise order (7.2). Proof. Consider any u, v ∈ MLl (Ω). Then there is a closed and nowhere dense subset Γ of Ω such that u, v ∈ C l (Ω \ Γ). Define open subsets U , V and W of Ω \ Γ through U = {x ∈ Ω \ Γ : u (x) < v (x)}, CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 133 V = {x ∈ Ω \ Γ : v (x) < u (x)} and W = int{x ∈ Ω \ Γ : u (x) = v (x)}, respectively. It is clear that the function ϕ : Ω 3 x 7→ sup{u (x) , v (x)} ∈ R is C l -smooth on U ∪ V ∪ W . Clearly the set U ∪ V ∪ W is dense in Ω \ Γ. As such, it follows by Theorem 45 that sup{u, v} belongs to MLl (Ω). The existence of the infimum of u, v ∈ MLl (Ω) in MLl (Ω) follows in the same way. The distributivity of MLl (Ω) now follows by Proposition 47. The usual partial differential operators Dα : C l (Ω) → C 0 (Ω) , |α| ≤ l (8.6) may be extended in a straightforward way to the larger space MLl (Ω). Indeed, in view of (8.5), it is clear that, for each u ∈ MLl (Ω), we have ∃ Γ ⊂ Ω closed nowhere dense : ∀ |α| ≤ l : Dα u|Ω\Γ ∈ C 0 (Ω \ Γ) (8.7) which allows for an extension of the mapping (8.6) to a mapping Dα : MLl (Ω) 7→ ML0 (Ω) (8.8) Dα : u 7→ (I ◦ S) (Dα u) . (8.9) through Indeed, in view of (8.7) and (7.1), the function Dα u is nearly finite and normal lower semi-continuous for every |α| ≤ l. Furthermore, each partial derivatives Dα u belongs to MLl (Ω). In particular, Dα u (x) = Dα u (x) , x ∈ Ω \ Γ, where Γ is the closed nowhere dense subset of Ω associated with u through (8.5). In order to now extend the mapping (8.4) to a mapping T : MLm (Ω)K → ML0 (Ω)K , (8.10) we express (8.4) componentwise as Tj : C m (Ω)K 3 u 7→ Fj (·, ..., ui , ..., Dα ui , ...) ∈ C 0 (Ω) (8.11) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 134 where F1 , ..., FK : Ω × RM → R are the components of the mapping (8.2). The components (8.11) extend in a straight forward way to mappings Tj : MLm (Ω)K → ML0 (Ω) which are defined as Tj : MLm (Ω)K 3 u 7→ (I ◦ S) (Fj (·, ..., ui , ..., Dα ui , ...)) ∈ ML0 (Ω) . (8.12) In view of (8.7), it follows by (7.1) and the continuity of each of the components F1 , ..., FK of the mapping (8.2) that the mapping (8.12) is well defined for each j = 1, ..., K. As such, we may define the extension (8.10) of the mapping (8.4) componentwise, with components defined in (8.12). That is, T : MLm (Ω)K 3 u 7→ (Tj u)j≤K ∈ ML0 (Ω)K . The mapping (8.10) extends the mapping (8.4) associated with the nonlinear partial differential operator (8.3). Therefore we may formulate the system of nonlinear PDEs (8.1) in the significantly more general framework of the spaces of normal lower semi-continuous functions MLm (Ω)K and ML0 (Ω)K . In particular, we formulate the generalized equation Tu = f (8.13) where the unknown u ranges over MLm (Ω)K . It should be noted that this extended formulation of the problem allows for functions with singularities on arbitrary closed nowhere dense subsets of the domain of definition Ω to act as global solutions of the system of nonlinear PDEs (8.1). This should be compared with the global version of the Cauchy-Kovalevskaia Theorem [141] which is also mentioned in Section 1.3. Furthermore, such a solution will in general not belong to any of the customary spaces of generalized functions, such as the Sobolev spaces H 2,m (Ω), or the space D0 (Ω) of distributions on Ω. Indeed, a function u ∈ MLm (Ω) will in general fail to be locally integrable on Ω, since it does not satisfy any growth conditions near the closed nowhere dense singularity set Γ associated with u through (8.5). Throughout this section, the space ML0 (Ω) is equipped with the uniform order convergence structure Jo , while the product space ML0 (Ω)K will carry the product uniform convergence structure JoK with respect to Jo . That is, ∀ i = 1, ..., K : K U ∈ Jo ⇔ (8.14) (πi × πi ) (U) ∈ Jo where πi denotes the projection πi : ML0 (Ω)K 3 u = (ui )i≤K 7→ ui ∈ ML0 (Ω) . The basic properties of the space ML0 (Ω)K that are relevant to this investigation are summarized in the following proposition. CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 135 Proposition 64 The uniform convergence space ML0 (Ω)K is first countable and Hausdorff. Furthermore, its completion is the space N L (Ω)K equipped with the product uniform convergence structure with respect to the uniform convergence structure Jo] . Proof. The assertions of the proposition follow immediately from Proposition 41, Corollary 57 and Theorem 62, respectively. Within the context of the nonlinear mapping associated with a given system of nonlinear PDEs introduced in this section, and in particular the extended mapping (8.10), the most simple way in which to define a suitable uniform convergence structure on MLm (Ω)K is to introduce the initial uniform convergence structure on MLm (Ω)K with respect to the mapping (8.10). However, the completion results for uniform convergence spaces discussed in Sections 2.4 and 6.1 apply to Hausdorff uniform convergence spaces only, while the initial uniform convergence structure on MLm (Ω)K with respect to the mapping (8.10) is Hausdorff if and only if the mapping (8.10) is injective, which is typically not the case. This difficulty can be overcome if we associate with the mapping (8.10) an equivalence relation on MLm (Ω)K through u ∼T v ⇔ Tu = Tv. (8.15) The mapping (8.10) induces an injective mapping b : MLm (Ω) → ML0 (Ω)K , T T (8.16) K m where MLm T (Ω) denotes the quotient space ML (Ω) / ∼T , such that the diagram T MLm (Ω)K - ML0 (Ω)K @ @ @ @ @ qT @ b T (8.17) @ @ @ R @ MLm T (Ω) commutes, where qT is the canonical quotient mapping associated with the equivalence relation (8.15). The diagram (8.17) amounts simply to a representation of the mapping T. In particular, the equation (8.13) is, in a certain precise sense, equivalent to the equation b = f, TU (8.18) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 136 with the unknown U ranging over MLm T (Ω). Indeed, in view of the diagram (8.17) and the surjectivity of qT , it follows that ∀ u ∈ MLm (Ω)K : b q u =f Tu = f ⇔ T T and ∀ U ∈ MLm T (Ω) : b = f ⇔ Tu = f, u ∈ q −1 (U) TU T Since the mapping (8.16) is injective it follows that the initial uniform convergence structure JT on MLm T (Ω) with respect to (8.16) is Hausdorff. In particular, b b U ∈ JT ⇔ T × T (U) ∈ JoK (8.19) b is in fact a uniformly continuous embedding. As such, and in view of so that T Proposition 36, the uniform convergence space completion N LT (Ω) of MLm T (Ω) K K 0 0 may be identified with a subspace of the completion N L (Ω) of ML (Ω) . In particular, the mapping (8.16) extends to an injective uniformly continuous mapping b ] : N L (Ω) → N L0 (Ω)K . T T (8.20) Within the context of the construction (8.15) to (8.20), we may formulate the generalized equation b ] U] = f T (8.21) corresponding to the equation (8.18), where the unknown U] ranges over N LT (Ω)K . In view of the equivalence of the equations (8.13) and (8.18), we will interpret any solution to (8.21) as a generalized solution of the system of nonlinear PDEs (8.1). The question of existence of solutions to (8.21) will be addressed in Section 9.2. 8.2 Sobolev Type Spaces of Generalized Functions The Order Completion Method [119] involves a construction of spaces of generalized functions which are associated with the particular nonlinear partial differential operator which defines the equation. The spaces of generalized functions constructed in Section 8.1 employ essentially the same technique, with the key difference that the spaces of generalized functions are obtained not through the process of order completion, but rather through the more general topological process of completion of a uniform convergence space. As mentioned, the spaces of generalized functions constructed in Section 8.1 are constructed with a particular nonlinear partial differential operator in mind. As CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 137 such, they may depend to a large extent on this operator. Furthermore, there is no concept of derivative of generalized functions. In this section we construct, in the original spirit of Sobolev [148] and [149], spaces of generalized functions which are independent of any particular nonlinear partial differential operator. Moreover, these spaces are equipped in a natural and canonical way with partial differential operators that extend the classical operators on spaces of smooth functions. Furthermore, and as we will show in Section 8.3, these spaces are, in a certain precise sense, compatible with the spaces constructed in Section 8.2. Recall that the Sobolev space H 2,l (Ω) may be constructed as the completion of C l (Ω) equipped with the initial vector space topology induced by the family of mappings Dα : C l (Ω) → L2 (Ω) |α|≤l where L2 (Ω) is the Hilbert space of square integrable functions on Ω. We follow a similar approach in constructing spaces of generalized functions. In this regard, we equip the space MLl (Ω), where l ≥ 1, with the initial uniform convergence structure JD with respect to the family of mappings Dα : MLl (Ω) → ML0 (Ω) |α|≤l (8.22) That is, for any filter U on MLl (Ω) × MLl (Ω), we have ∀ |α| ≤ l U ∈ JD ⇔ (Dα × Dα ) (U) ∈ Jo (8.23) Since the family of mappings (8.22) separates the elements of MLl (Ω), that is, ∀ u, v ∈ MLl (Ω) , u 6= v : , ∃ |α| ≤ l : Dα u 6= Dα v it follows that JD is uniformly Hausdorff. A filter F on MLl (Ω) is a Cauchy filter if and only if ∀ |α| ≤ l : Dα (F) is a Cauchy filter in ML0 (Ω) (8.24) In particular, a filter F on MLl (Ω) converges to u ∈ MLl (Ω) if and only if ∀ |α| ≤ l : Dα (F) ∈ λo (Dα u) (8.25) In view of the results in Chapter 6 on the completion of uniform convergence spaces, the completion of MLl (Ω) is realized as a subspace of N L (Ω)M , for an appropriate M ∈ N. In this regard, we note, see Proposition 43, that the mapping D : MLl (Ω) 3 u 7→ (Dα u)|α|≤l ∈ ML0 (Ω)M (8.26) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 138 is a uniformly continuous embedding. In particular, for each |α| ≤ l, the diagram D MLl (Ω) - ML0 (Ω)M @ @ @ @ Dα@ (8.27) πα @ @ @ @ R @ 0 ML (Ω) commutes, with πα the projection. This diagram amounts to a decomposition of u ∈ MLl (Ω) into its partial derivatives. In view of the uniform continuity of the mapping D and its inverse, it follows by Theorem 44 that D extends to an injective uniformly continuous mapping D] : N Ll (Ω) → N L (Ω)M (8.28) where N Ll (Ω) denotes the uniform convergence space completion of MLl (Ω). Moreover, since each mapping Dα is uniformly continuous, one obtains the commutative diagram D] l N L (Ω) - N L0 (Ω)M @ @ @ @ Dα] @ πα] (8.29) @ @ @ @ R @ 0 N L (Ω) where Dα] : N Ll (Ω) → N L0 (Ω) (8.30) is the extension through uniform continuity of the partial differential operator Dα . Since the mapping D] is injective and uniformly continuous, and in view of the CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 139 commutative diagram (8.29) above, each generalized function u] ∈ N Ll (Ω) may be uniquely represented by its generalized partial derivatives (8.31) u] 7→ D] u] = Dα] u] |α|≤l Each generalized partial derivative Dα] u] of u] is a nearly finite normal lower semi-continuous function. We note, therefore, that the set of singular points of each u] ∈ N Ll (Ω), that is, the set ∃ |α| ≤ l : x∈Ω Dα] u] not continuous at x is at most a set of first Baire category, that is, it is a topologically small set. However, this set may be dense in Ω. Furthermore, such a set may have arbitrarily large positive Lebesgue measure [121]. Highly singular objects, such as the generalized functions that are the elements of MLl (Ω) may turn out to model highly relevant real world situations, like turbulence or other chaotic phenomena. 8.3 Nonlinear Partial Differential Operators This section deals with the general class of nonlinear partial differential operators associated with systems of nonlinear PDEs of the form (8.1) to (8.3). In this regard, we investigate the properties of such operators in the context of the Sobolev type spaces of generalized functions introduced in Section 8.2, and in particular the extent to which such operators are compatible with the topological structures of these spaces. Furthermore, the extent to which the Sobolev type spaces are compatible with the ‘pull back’ spaces of generalized functions introduced in Section 8.1 are demonstrated. The first part of this section concerns the general class of nonlinear partial differential operators introduced in Section 8.1. It is shown that the mapping (8.10) induced by such an operator is uniformly continuous with respect to the Sobolev type uniform convergence structure on MLm (Ω), and the uniform order convergence structure on ML0 (Ω). It is also shown that the Sobolev type spaces of generalized functions are compatible with the pull back spaces. In the second part of this section we introduce additional smoothness properties on the nonlinear partial differential operators, and some basic properties of these operators are discussed. The approach to generalized solutions of nonlinear PDEs pursued in this work is based on extending nonlinear partial differential operators to the completion of a suitable uniform convergence space. As is mentioned in Section 1.2, some care must be taken in constructing such extensions. In particular, it is essential that the mapping associated with such a nonlinear operator is compatible with the relevant uniform convergence structures, namely, it must be uniformly continuous. CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 140 In this regard, consider a system of nonlinear PDEs of the form (8.1) through (8.3), and the mapping (8.10) associated with the system of equations, that is, the mapping T : MLm (Ω)K → ML0 (Ω)K The Cartesian product MLm (Ω)K will throughout be equipped with the product uniform convergence structure JDK with respect to the uniform convergence structure JD on MLm (Ω), that is, ∀ i = 1, ..., K : K U ∈ JD ⇔ . (8.32) (πi × πi ) (U) ∈ Jo Since MLm (Ω) is Hausdorff, so is the product. Furthermore, in view of Theorem 41, the completion of MLm (Ω)K is N Lm (Ω)K . Within the context of the Sobolev type uniform convergence structure (8.23) on MLm (Ω), and the uniform order convergence structure on ML0 (Ω), the basic result concerning the mapping (8.10) is the following. Theorem 65 Consider a mapping T : MLm (Ω)K → ML0 (Ω)K defined through a jointly continuous mapping (8.2) as in (8.12). Then this mapping is uniformly continuous. Proof. The mapping T may be represented through the diagram T MLm (Ω)K - ML0 (Ω)K @ @ @ @ [email protected] F (8.33) @ @ @ @ R @ ML0 (Ω)M ×K where F = F i i≤K is defined componentwise through F i : ML0 (Ω)M ×K 3 u 7→ (I ◦ S) (Fi (·, u1 , ..., uM )) ∈ ML0 (Ω) and D is defined as |α|≤m D : MLm (Ω)K 3 u 7→ (Dα ui )i≤K ∈ ML0 (Ω)M ×K (8.34) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 141 Clearly D is uniformly continuous, so in view of the diagram (8.33) it suffices to show that F is uniformly continuous with respect to the product uniform convergence structure on ML0 (Ω)M ×K . In this regard, we consider sequences of order intervals (Ini ) in ML0 (Ω), which, for i = 1, ..., M × K, satisfies condition 1) of (7.17) and (7.18). We claim ∀ n∈N: 0 K 1 ∃ Order Q intervals Jn , ..., Jn ⊆ ML (Ω) : M ×K i Fj In ⊆ Jnj , j = 1, ..., K i=1 (8.35) To verify (8.35), observe that there is a closed nowhere dense set Γn ⊆ Ω so that ∀ x∈Ω\Γ : ∃ a (x) > 0 : ∀ i = 1, ..., M × K : u ∈ Ini ⇒ |u (x) | ≤ a (x) (8.36) Since Fj : Ω × RM → R is continuous, it follows from (8.36) that ∀ x∈Ω\Γ : ∃ b (x) > 0 : ∀ i = 1, ..., M × K : ⇒ |Fj (x, u1 (x) , ..., uM (x)) | ≤ b (x) ui ∈ Ini (8.37) Therefore, in view of Proposition 49, our claim (8.35) holds. In particular, since N L (Ω) is Dedekind complete by Theorem 45 , we may set Jnj = [λjn , µjn ] where, for each n ∈ N and each j = 1, ..., K λjn = inf{F j u : u ∈ M ×K Y Ini } i=1 and µjn = sup{F j u : u ∈ M ×K Y Ini } i=1 The sequence (λjn ) and (µjn ) are increasing and decreasing, respectively. For each j = 1, ..., K we may consider sup{λjn : n ∈ N} = uj ≤ v j = inf{µjn : n ∈ N} We claim that uj = v j . To see this, we note that for each i = 1, ..., M × K there is some wi ∈ N L (Ω) so that sup{lni : n ∈ N} = wi = inf{uin : n ∈ N} CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 142 where Ini = [lni , uin ]. Applying Proposition 50 and the continuity of Fj our claim is verified. Applying the same technique as in the proof of Theorem 62, as well j as Proposition 34 we obtain a sequence I n of order intervals in ML0 (Ω) that satisfies 1) of (7.17), (7.18) and ! M Y j Fj Ini ⊆ I n i=1 This completes the proof. Since the mapping (8.10) is uniformly continuous, it extends in a unique way to a uniformly continuous mapping T] : N Lm (Ω)K → N L (Ω)K . (8.38) Therefore, one may formulate a generalized equation corresponding to (8.13) as T ] u] = f (8.39) where the unknown u] ranges over N Lm (Ω). In view of the fact that the mapping (8.10) is the unique uniformly continuous extension of (8.4), we interpret any solution to (8.39) as a generalized solution to the system of nonlinear PDEs (8.1). Recall also that the mapping (8.16) is a uniformly continuous embedding. As such, and in view of the commutative diagram (8.17), the canonical quotient mapping qT : MLm (Ω)K → MLm T (Ω) associated with the equivalence relation (8.15) is uniformly continuous, and extends in a unique way to a uniformly continuous mapping q ] : N Lm (Ω)K → N LT (Ω) T (8.40) In particular, the mapping (8.20) may be interpreted as a representation for the mapping (8.38) through the commutative diagram T] K m N L (Ω) - N L (Ω)K @ @ @ @ b] T @ q] @ T @ @ @ R @ N LT (Ω) (8.41) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 143 which is nothing but an extension of the diagram (8.17). Indeed, since the mapping (8.20) is an injective uniformly continuous mapping, it follows that ∀ u] , v] ∈ N Lm (Ω) : . T] u] = T] v] ⇔ q ] u] = q ] v] T T (8.42) In particular, q ] is nothing but the canonical quotient map associated with the T equivalence relation ∀ u] , v] ∈ N Lm (Ω) : . u] ∼T] v] ⇔ T] u] = T] v] (8.43) The meaning of (8.41) to (8.43) is clear. Indeed, any solution to the generalized equation (8.39) corresponds to a solution to (8.21). In particular, any generalized function U] ∈ q ] N Lm (Ω)K ⊆ N LT (Ω) T may be interpreted a ∼T] -equivalence class of generalized functions in N Lm (Ω)K . This may be interpreted as a regularity result for the generalized functions in N LT (Ω). However, from the diagram (8.41) we only obtain the inclusion K ] m q N L (Ω) ⊆ N LT (Ω) , (8.44) T and equality in (8.44) may not hold for all nonlinear partial differential operators T. In this regard, we will present sufficient conditions for equality to hold in (8.44) in Section 9.3. We have so far considered nonlinear partial differential operators which satisfy minimal assumptions on smoothness of the mapping (8.2). In particular, it is only assumed that the mapping (8.2) is continuous. However, it most often happens in practice that (8.2) satisfies additional smoothness conditions, namely, that it is continuously differentiable up to a given order. Such additional smoothness conditions will be exploited in Chapter 10 to obtain dramatic regularity results for the solutions of a large class of systems of nonlinear PDEs. In this regard, we consider now the case of a system of nonlinear PDEs of the form (8.1) to (8.3) where the mapping F : Ω×RM → RK which defines the nonlinear operator through (8.3), is assumed to be not only continuous, but also C k -smooth, that is, F ∈ C k Ω × RM , RK for some k ∈ N ∪ {∞}. Since C m+k (Ω) ⊂ C m (Ω), we may compute Tu for each u ∈ C m+k (Ω)K . In this case, in view of the chain rule of differentiation, it is clear that Tu ∈ C k (Ω)K , that is, T : C m+k (Ω)K → C k (Ω)K . (8.45) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 144 More generally, given any u ∈ MLm+k (Ω)K , applying the mapping (8.10) we obtain Tu ∈ MLk (Ω)K . That is, restricting (8.10) to MLm+k (Ω)K yields a mapping T : MLm+k (Ω)K → u ∈ MLk (Ω)K . (8.46) Indeed, in view of (8.5) we have, for each u ∈ MLm+k (Ω)K , ∃ Γ ⊂ Ω closed nowhere dense : ∀ i = 1, ..., K : ∀ |α| ≤ m : Dα ui ∈ C k (Ω \ Γ) From the smoothness of the mapping (8.2) and the chain rule, it follows that Tu ∈ C k (Ω \ Γ)K (8.47) which verifies (8.46). In case the nonlinear partial differential operator satisfies sufficient smoothness conditions, such as those introduced in (8.45) to (8.47), we may introduce a suitable notion of derivative of the partial differential operator T. Indeed, for each u ∈ MLm+k (Ω)K , we may calculate the partial derivatives Dβ Tj u ∈ ML0 (Ω) , |β| ≤ k, j = 1, ..., K where the Tj , for j ≤ K, are the components (8.12) of the mapping (8.46). In this regard, we may define a mapping Tk : MLm+k (Ω)K → ML0 (Ω)N , (8.48) for a suitable choice of N ∈ N ∪ {∞}, so that the diagram m+k ML Tk K (Ω) - ML0 (Ω)N @ @ @ (8.49) @ @ T @ D @ @ @ R @ MLk (Ω)K CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 145 commutes, with the mapping D defined through D : MLk (Ω)K 3 v 7→ Dβ vi |β|≤k i≤K ∈ ML0 (Ω)N (8.50) Applying the chain rule, we can obtain an explicit expression for the mapping (8.48) in terms of the mapping (8.2), which defines the partial differential operator (8.46), and its derivatives. Such a formula, however, is typically rather involved. As such, we will rather express it in terms of a suitable jointly continuous mapping Fk : Ω × R L → R N , (8.51) k of the for a suitable integer L. In particular, we may define the components Tj,β mapping (8.48) through k k Tj,β u = (I ◦ S) Fj,β (·, ..., ui , ..., Dα ui , ...) , |α| ≤ m + k; i = 1, ..., K (8.52) k where the Fj,β are components of the mapping (8.51). The main result concerning the mapping (8.46) is the following. Theorem 66 Let k be finite. Then the mapping (8.46) is uniformly continuous with respect to the Sobolev uniform convergence structures on MLm+k (Ω)K and MLk (Ω)K . Proof. The uniform continuity of the mapping (8.48) defined through (8.52) follows by the same arguments used in the proof of Theorem 65. Furthermore, the mapping (8.50) is clearly a uniformly continuous embedding. The uniform continuity of (8.46) now follows from the commutative diagram (8.49). In view of Theorem 66, the mapping (8.46) extends uniquely to a uniformly continuous mapping T] : N Lm+k (Ω)K → N Lk (Ω)K (8.53) Furthermore, both the mappings (8.48) and (8.50) are uniformly continuous, so that these mappings may be uniquely extended to uniformly continuous mappings Tk] : N Lm+k (Ω)K → N L0 (Ω)N , (8.54) and D] : N Lk (Ω)K 3 v 7→ Dβ] vi |β|≤k i≤K ∈ N L0 (Ω)N . (8.55) CHAPTER 8. SPACES OF GENERALIZED FUNCTIONS 146 As such, the diagram (8.49) extends to the commutative diagram Tk] N Lm+k (Ω)K - N L0 (Ω)N @ @ @ (8.56) @ @ ] D] T @ @ @ @ R @ N Lk (Ω)K Note that, in case both the nonlinear partial differential operator and the righthand term in the system of nonlinear PDEs (8.1) are C k -smooth, the extended equation (8.13) is equivalent to Tk u = Df. (8.57) In view of the extensions (8.53) and (8.55) of the smooth nonlinear partial differential operator, and the uniformly continuous embedding (8.50), respectively, we may formulate the equation corresponding to (8.57) as Tk] u] = D] f. (8.58) It should be noted that the generalized equation (8.39) corresponding to (8.13) is no longer equivalent to the equation (8.58). Indeed, a solution u] ∈ N Lm (Ω)K may not have generalized derivatives up to order m + k, which is required of any solution to (8.58). Such additional, and in fact rather minimal, smoothness conditions on the nonlinear partial differential operator turn out to be sufficient for particularly strong regularity properties of generalized solutions to large classes of systems of nonlinear PDEs. As will be shown in Section 10.2, only very basic assumptions of a simple topological nature are involved in the relevant regularity properties of generalized solutions of (8.1). Chapter 9 Existence of Generalized Solutions 9.1 Approximation Results In this section we obtain the basic approximation results used to prove the existence of solutions to the generalized equations (8.20) and (8.39). We also show that functions in MLm (Ω) may be suitably approximated by sequences of smooth functions. In particular, we show that C m (Ω) is dense in MLm (Ω). The first and basic approximation results are essentially multi dimensional versions of the fundamental approximation results (1.108) and (1.110) underlying the Order Completion Method. These results allow for the existence of generalized solutions to (8.1) in the space MLm T (Ω), that is, a solution to (8.21). Further specializations of these basic results will also be presented. In particular, under certain mild assumptions on the nonlinear partial differential operator (8.3) we obtain bounds for such approximate solutions. These bounds will be used to obtain generalized solutions to (8.1) in N Lm (Ω)K , that is, solutions to (8.39). Similar approximation results are also proved for equations that satisfy additional smoothness assumptions, namely, assumptions such as those introduced in Section 8.3. These approximation results for such equations that satisfy additional smoothness conditions result in a strong regularity property for solutions in the Sobolev type spaces of generalized functions. Finally, we investigate the extent to which functions in MLm (Ω) may be approximated by C m -smooth functions. We now again consider a system of K nonlinear PDEs of the form (8.1) through (8.3). Recall that the Order Completion Method, as discussed in Section 1.4, for single nonlinear PDEs of the form (1.100) through (1.102) is based on the simple approximation result (1.110). In this section we extend this result to the general K-dimensional case, for K ≥ 1 arbitrary but given, see [119] for a particular case of such an extension. A natural assumption on the function F : Ω × RM → RK , and hence the PDEoperator T (x, D), and the righthand term f is that, for every x ∈ Ω f (x) ∈ int{F (x, ξ1α , ..., ξiα , ...) : ξiα ∈ R, i = 1, ..., K, |α| ≤ m}, 147 (9.1) CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 148 which is a multidimensional version of (1.107). The condition (9.1) is noting but a sufficient condition for the system of nonlinear PDEs (8.1) to have usual classical solution on Ω. Note that (9.1) is of a technical nature, and hardly a restriction on the class of PDEs considered. In fact, every linear PDE, and also most nonlinear PDEs of applicable interest satisfy (9.1). It is in fact, as discussed in Section 1.4, a necessary condition for the existence of a classical solution to (8.3) in a neighborhood of x. Assuming that the condition (9.1) holds, we obtain the following basic result. Theorem 67 Consider a system of PDEs of the form (8.1) through (8.3) that also satisfies (9.1). For every > 0 there exists a closed nowhere dense set Γ ⊂ Ω with zero Lebesgue measure, and a function U ∈ C m (Ω \ Γ )K with components U,1 , ..., U,K such that fi (x) − ≤ Ti (x, D) U (x) ≤ fi (x) , x ∈ Ω \ Γ Proof. (9.2) Let Ω= [ Cν (9.3) ν∈N where, for ν ∈ N, the compact sets Cν are n-dimensional intervals Cν = [aν , bν ] (9.4) with aν = (aν,1 , ..., aν,n ), bν = (bν,1 , ..., bν,n ) ∈ Rn and aν,i ≤ bν,i for every i = 1, ..., n. We also assume that Cν , with ν ∈ N are locally finite, that is, ∀ x∈Ω: ∃ Vx ⊆ Ω a neighborhood of x : {ν ∈ N : Cν ∩ Vx 6= ∅} is finite (9.5) We also assume that the interiors of Cν , with ν ∈ N, are pairwise disjoint. We note that such Cν exist, see [58]. Let us now take > 0 given arbitrary but fixed. Let us take ν ∈ N and apply Proposition 68 to each x0 ∈ Cν . Then we obtain δx0 > 0 and Px0 ,1 , ..., Px0 ,K polynomial in x ∈ Rn such that fi (x) − ≤ Ti (x, D) Px0 (x) ≤ f (x) , x ∈ Ω ∩ B (x0 , δx0 ) and i = 1, ..., K (9.6) where Px0 : Rn → RK is the K-dimensional vector valued function with components Px0 ,1 , ..., Px0 ,K . Since Cν is compact, it follows that ∃ δ>0: ∀ x0 ∈ Cν : (9.7) ∃ Px0 ,1 , ..., Px0 ,K polynomial in x ∈ Rn : kx − x0 k ≤ δ ⇒ fi (x) − ≤ Ti (x, D) Px0 (x) ≤ f (x) , x ∈ B (x0 , δ) ∩ Cν where i = 1, ..., K. Subdivide Cν into n-dimensional intervals Iν,1 , ..., Iν,µ with diameter not exceeding δ such that their interiors are pairwise disjoint. If aj with CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 149 j = 1, ..., µ is the center of the interval Iν,j then by (9.7) there exists Paj ,1 , ..., Paj ,K polynomial in x ∈ Rn such that fi (x) − ≤ Ti (x, D) Paj (x) ≤ fi (x) , x ∈ Iν,j (9.8) where i = 1, ..., K. Now set Γν, = Cν \ µ [ ! intIν,j ! ∪ intCν (9.9) j=1 that is, Γν, is a rectangular grid generated as a finite union of hyperplanes. Furthermore, using (9.8), we find Uν, ∈ C m (Cν \ Γν, ) (9.10) fi (x) − ≤ Ti (x, D) Uν, (x) ≤ fi (x) , x ∈ Cν \ Γν, (9.11) such that In view of (9.5) it follows that [ Γ = Γν, is closed nowhere dense and mes (Γ ) = 0 (9.12) ν∈N From (9.3), (9.10) and (9.11) we obtain (9.2). The above proof relies on the following proposition which is in fact the basic approximation result. Proposition 68 Consider a system of PDEs of the form (8.1) through (8.3) that also satisfies (9.1). Then ∀ x0 ∈ Ω : ∀ >0: ∃ δ > 0, P1 , ..., PK polynomial in x ∈ Rn : x ∈ B (x0 , δ) ∩ Ω, 1 ≤ i ≤ k ⇒ fi (x) − ≤ Ti (x, D) P (x) ≤ fi (x) (9.13) Here P is the K-dimensional vector valued function with components P1 , ..., PK . Proof. For any x0 ∈ Ω and > 0 small enough it follows by (9.1) that there exist ξiα ∈ R with i = 1, ..., K and |α| ≤ m such that Fi (x0 , ..., ξiα , ...) = fi (x0 ) − 2 (9.14) (9.15) Now take P1 , ..., PK polynomials in x ∈ Rn that satisfy Dα Pi (x0 ) = ξiα for i = 1, ..., K and |α| ≤ m (9.16) CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 150 Then it is clear that (9.17) 2 where P is the K-dimensional vector valued function on Rn with components P1 , ..., PK . Hence (9.13) follows by the continuity of the fi and the Fi . Ti (x, D) P (x0 ) − fi (x0 ) = − It should be observed that, in contradistinction with the usual functional analytic methods, the local lower solution in Proposition 68 is constructed in a particularly simple way. Indeed, it is obtained by nothing but a straightforward application of the continuity of the mapping F. Using exactly these same techniques, one may prove the existence of the corresponding approximate upper solutions. Proposition 69 Consider a system of PDEs of the form (8.1) through (8.3) that also satisfies (9.1). Then ∀ x0 ∈ Ω : ∀ >0: ∃ δ > 0, P1 , ..., PK polynomial in x ∈ Rn : x ∈ B (x0 , δ) ∩ Ω, 1 ≤ i ≤ k ⇒ fi (x) < Ti (x, D) P (x) < fi (x) + (9.18) Here P is the K-dimensional vector valued function with components P1 , ..., PK . In connection with the global approximation result presented in Theorem 67, and as was mentioned in connection with Proposition 68, the approximation result above is based solely on the existence of a compact tiling of open subsets of Rn , the properties of compact subsets of Rn and the continuity of usual real valued functions. Hence it makes no use of so called advanced mathematics. In particular, techniques from functional analysis are not used at all. Instead, the relevant techniques belong rather to the classical theory of real functions. Note that Theorem 67 makes no claim concerning the convergence, or otherwise, of the sequence (un ) in MLm (Ω)K . Indeed, assuming only that (9.1) is satisfied, it is typically possible to construct a sequence (Un ) that satisfies Theorem 67, and is unbounded on every neighborhood of every point of Ω. This follows easily from the fact that, in general, for a fixed x0 ∈ Ω, the sets {ξ ∈ RM : F (x0 , ξ) = f (x0 )} may be unbounded. In view of the above remarks, it appears that a stronger assumption than (9.1) may be required in order to construct generalized solutions to (8.1) in N Lm (Ω)K . When formulating such an appropriate condition on the system of PDEs (8.1), one should keep in mind that the Order Completion Method [119], and in particular the pseudo-topological version of the theory developed in this work, is based on some basic topological processes, namely, the completion of uniform convergence spaces, and the simple condition (9.1), which is formulated entirely in terms of the usual real mappings F and f. In particular, (9.1) does not involve any topological structures CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 151 on function spaces, or mappings on such spaces. Furthermore, other than the mere continuity of the mapping F, (9.1) places no restriction on the type of equation treated. As such, it is then clear that any further assumptions that we may wish to impose on the system of equations (8.1) in order to obtain generalized solutions in N Lm (Ω)K should involve only basic topological properties of the mapping F, and should not involve any restrictions on the type of equations. In formulating such a condition on the system of PDEs (8.1) that will ensure the existence of a generalized solution in N Lm (Ω)K , it is helpful to first understand more completely the role of the condition (9.1) in the proof of the local approximation result Proposition 68. In particular, and as is clear from the proof of Proposition 68, the condition (9.1) relates to the continuity of the mapping F. Furthermore, and as has already been mentioned, the approximations constructed in Theorem 67 and Proposition 68 concern only convergence in the range space of the operator T associated with (8.1). Our interest here lies in constructing suitable approximations in the domain of T, and as such, properties of the inverse of the mapping F may prove to be particularly useful. In view of these remarks, we introduce the following condition. ∀ x0 ∈ Ω : ∃ ξ (x0 ) ∈ RM , F (x0 , ξ (x0 )) = f (x0 ) : ∃ V ∈ Vx0 , W ∈ Vξ(x0 ) : F : V × W → RK open (9.19) Note that the condition (9.19) above, although more restrictive than (9.1), allows for the treatment of a large class of equations. In particular, each equation of the form Dt u (x, t) + G (x, t, u (x, t) , ..., Dxα u (x, t) , ...) = f (x, t) with the mapping G merely continuous, satisfies (9.19). Indeed, in this case the mapping F : Ω×RM → RK that defines the equation through (8.3) is both open and surjective. Indeed, each component Fj of F is linear in ξj , where ξ = (ξ1 , ..., ξj , ..., ξM ) belongs to RM , from which our assertion follows immediately. Other classes of equations that satisfy (9.19) can be easily exhibited by using, for instance, various open mapping theorems, see for instance [19, 41.7]. The following is a specialization of the global approximation result Theorem 67. Theorem 70 Consider a system of nonlinear PDEs of the form (8.1) through (8.3) that also satisfies (9.19). Then there is a sequence (Γn ) of closed nowhere dense set Γn ⊂ Ω, which is increasing with respect to inclusion, and a sequence of function (Vn ) such that Vn ∈ C m (Ω \ Γn ) and ∀ j = 1, ..., K : . fj (x) − n1 ≤ Tj (x, D) Vn (x) ≤ fj (x) , x ∈ Ω \ Γn CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 152 Furthermore, for each |α| ≤ m and every i = 1, ..., K there are sequences λαn,i and µαn,i such that λαn,i , µαn,i ∈ C 0 (Ω \ Γn ) which sequences satisfy ∀ n∈N: ∀ |α| ≤ m : ∀ i = 1, ..., K : 1) λαn,i (x) < Dα Vn,i (x) < µαn,i (x) , x ∈ Ω \ Γn 2) λαn,i (x) < λαn+1,i (x) < µαn+1,i (x) < µαn,i (x) , x ∈ Ω \ Γn+1 and S ∀ x∈Ω\ : n∈N Γn ∀ |α| ≤ m : ∀ i = 1, ..., K : sup{λαn,i (x) : n ∈ N} = inf{µαn,i (x) : n ∈ N} Proof. Set Ω= [ Cν (9.20) ν∈N where, for ν ∈ N, the compact set Cν is an n-dimensional intervals Cν = [aν , bν ] (9.21) with aν = (aν,1 , ..., aν,n ), bν = (bν,1 , ..., bν,n ) ∈ Rn and aν,j ≤ bν,j for every j = 1, ..., n. We also assume that the collection of sets {Cν : ν ∈ N} is locally finite, that is, ∀ x∈Ω: ∃ V ⊆ Ω a neighborhood of x : {ν ∈ N : Cν ∩ V 6= ∅} is finite (9.22) Furthermore, let the interiors of the Cν , with ν ∈ N, be pairwise disjoint. Let Cν be arbitrary but fixed. In view of (9.19) and the continuity of f, we have ∀ x0 ∈ Cν : ∃ ξ (x0 ) ∈ RM , F (x0 , ξ (x0 )) = f (x0 ) : ∃ δ, > 0 : (9.23) kx − x0 k < δ 1) {(x, f (x)) : kx − x0 k < δ} ⊂ int (x, F (x, ξ)) kξ − ξ (x0 ) k < K 2) F : Bδ (x0 ) × B2 (ξ (x0 )) → R open For each x0 ∈ Cν , fix ξ (x0 ) ∈ RM in (9.23). Since Cν is compact, it follows from (9.23) that ∃ δ>0: ∀ x0 ∈ Cν : ∃ x0 > 0 : (9.24) kx − x0 k < δ 1) {(x, f (x)) : kx − x0 k < δ} ⊂ int (x, F (x, ξ)) kξ − ξ (x0 ) k < x0 2) F : Bδ (x0 ) × B2x0 (ξ (x0 )) → RK open CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 153 Subdivide Cν into n-dimensional intervals Iν,1 , ..., Iν,µν with diameter not exceeding δ such that their interiors are pairwise disjoint. If aν,j with j = 1, ..., µν is the center of the interval Iν,j then by (9.24) we have ∀ j = 1, ..., µν : ∃ ν,j > 0 : x ∈ Iν,j (9.25) 1) {(x, f (x)) : x ∈ Iν,j } ⊂ int (x, F (x, ξ)) kξ − ξ (aν,j ) k < ν,j 2) F : Iν,j × B2ν,j (ξ (aν,j )) → RK open Take 0 < γ < 1 arbitrary but fixed. In view of Proposition 68 and (9.25), we have ∀ x0 ∈ Iν,j : ∃ Ux0 = U ∈ C m (Rn )K : ∃ δ = δx0 > 0 : |α|≤m 1) (Dα Ui (x))i≤K ∈ Bν,j (ξ (aν,j )) x ∈ Bδ (x0 ) ∩ Iν,j ⇒ 2) i ≤ K ⇒ fi (x) − γ < Ti (x, D) U (x) < fi (x) As above, we may subdivide Iν,j into pairwise disjoint, n-dimensional intervals Jν,j,1 , ..., Jν,j,µν,j so that for k = 1, ..., µν,j we have ∃ Uν,j,k = U ∈ C m (Rn )K : ∀ x ∈ Jν,j,k : |α|≤m ∈ Bν,j (ξ (aν,j )) , |α| ≤ m 1) Dα Ui (x)i≤K 2) i ≤ K ⇒ fi (x) − γ < Ti (x, D) U (x) < fi (x) (9.26) Set Γ1 = Ω \ µν [ µν,j [ ν∈N j=1 k=1 [ !!! intJν,j,k . and V1 = µν,j µν X X X ν∈N j=1 !! χJν,j,k Uν,j,k k=1 where χJν,j,k is the characteristic function of Jν,j,k . Then Γ1 is closed nowhere dense, and V1 ∈ C m (Ω \ Γ1 )K . In view of (9.26) we have, for each i = 1, ..., K fi (x) − γ < Ti (x, D) V1 (x) < fi (x) , x ∈ Ω \ Γ1 Furthermore, for each ν ∈ N, for each j = 1, ..., µν , each k = 1, ..., µν,j , each |α| ≤ m and every i = 1, ..., K we have x ∈ intJν,j,k ⇒ ξiα (aν,j ) − < Dα V1,i (x) < ξiα (aν,j ) + CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 154 Therefore the functions λα1,i , µα1,i ∈ C 0 (Ω \ Γ1 ) defined as λα1,i (x) = ξiα (aj ) − 2ν,j if x ∈ intIν,j and µα1,i (x) = ξiα (aj ) + 2ν,j if x ∈ intIν,j , respectively, satisfy λα1,i (x) < Dα V1,i (x) < µα1,i (x) , x ∈ Ω \ Γ1 and µα1,i (x) − λα1,i (x) < 4ν,j , x ∈ intIν,j Applying (9.25) restricted to Ω\Γ1 , and proceeding in a fashion similar as above, we may construct, for each n ∈ N such that n > 1, a closed nowhere dense set Γn ⊂ Ω, so that Γn ⊆ Γn+1 , a function Vn ∈ C m (Ω \ Γn )K and functions λαn,i , µαn,i ∈ C 0 (Ω \ Γn ) so that, for each i = 1, ..., K γ fi (x) − < Ti (x, D) Vn (x) < fi (x) , x ∈ Ω \ Γn . (9.27) n and for every |α| ≤ m λαn−1,i (x) < λαn,i (x) < Dα Vn,i (x) < µαn,i (x) < µαn−1,i (x) , x ∈ Ω \ Γn (9.28) and µαn,i (x) − λαn,i (x) < 4ν,j , x ∈ (intIν,j ) ∩ (Ω \ Γn ) . n (9.29) This completes the proof. At this point we proceed to establish an approximation result for equations that satisfy addition smoothness conditions such as those introduced in Section 8.3. In particular, we will establish a version of Theorem 70 that incorporates also the derivatives of Tu, for a sufficiently smooth function u. Owing to the representation (8.49) of the operator (8.46), this result follows by the same elementary arguments that lead to Theorem 70. In this regard, we consider a system of nonlinear PDEs of the form (8.1) to (8.3) so that the mapping (8.2) is C k -smooth, for some k ∈ N. In view of the representation (8.49), the condition (9.19) on the mapping (8.2) is replaced with a suitable assumption on the mapping (8.51), namely, we assume ∀ x0 ∈ Ω : |α|≤m ∃ ξ (x0 ) ∈ RL , Fk (x0 , ξ (x0 )) = (Dα fi (x0 ))i≤K : . ∃ V ∈ Vx0 , W ∈ Vξ(x0 ) : Fk : V × W → RN open (9.30) The following now follows by the representation (8.49) and the same arguments used in the proof of Theorem 70. CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 155 Theorem 71 Consider a system of nonlinear PDEs of the form (8.1) through (8.3) with both the righthand term and the mapping (8.2) C k -smooth. Also assume that (9.30) holds. Then there is an increasing sequence (Γn ) of closed nowhere dense sets Γn ⊂ Ω and a sequence of function (Vn ) such that Vn ∈ C m+k (Ω \ Γn ) and ∀ i = 1, ..., K : ∀ |β| ≤ k : . 1 β β β D fi (x) − n ≤ D Ti (x, D) Vn (x) ≤ D fi (x) , x ∈ Ω \ Γn Furthermore, for each |α| ≤ m + k and every i = 1, ..., K there are sequences λαn,i and µαn,i so that λαn,i , µαn,i ∈ C 0 (Ω \ Γn ) which satisfy ∀ n∈N: ∀ |α| ≤ m + k : ∀ i = 1, ..., K : 1) λαn,i (x) < Dα Vn,i (x) < µαn,i (x) , x ∈ Ω \ Γn 2) λαn,i (x) < λαn+1,i (x) < µαn+1,i (x) < µαn,i (x) , x ∈ Ω \ Γn+1 and and S ∀ x∈Ω\ : n∈N Γn ∀ |α| ≤ m + k : . ∀ i = 1, ..., K : sup{λαn,i (x) : n ∈ N} = inf{µαn,i (x) : n ∈ N} By employing the representation (8.49), we may verify Theorem 71 by using exactly the same techniques and arguments as in the proof of Theorem 70. As such we omit it. Remark 72 It should be noted that Theorem 67 may be reproduced for nonlinear partial differential operators that satisfy additional smoothness conditions. In particular, if we assume that the mapping (8.2) as well as the righthand term in (8.1) are both C k -smooth, for some k ∈ N, then we may obtain version of Theorem 67 that also incorporates the derivatives of Tu up to order k. This is not a significant improvement, as it does not lead to a more general or powerful existence or regularity result than are already possible using only Theorem 67. We now turn to the final result of the section, namely, we show that each function u ∈ MLm (Ω) may be suitably approximated by functions in C m (Ω). Together with certain basic compactness results to be presented in Section 10.1, this results in a significant improvements on the regularity of the generalized solutions to a large class of equations. The result we present now is based on the well known principle of Partition of Unity. In this regard, we may recall the following version of this principle. CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 156 Theorem 73 *[150] Let O be a locally finite open cover of a smooth manifold M . Then there is a collection {ϕU : M → [0, 1] : U ∈ O} of C ∞ -smooth mappings ϕU such that the following hold: i) For each U ∈ O, the support of ϕU is contained in U . P ii) For each x ∈ M , we have U ∈O ϕU (x) = 1. A useful consequence of Theorem 73 concerns the separation of disjoint, closed sets by C ∞ -smooth, real valued mappings. In this regard, consider a nonempty, open set Ω ⊆ Rn . Let S and T be disjoint, nonempty, closed subsets of Ω. Then it follows from Theorem 73 that ∃ ϕ ∈ C ∞ (Ω, [0, 1]) : 1) x ∈ A ⇒ ϕ (x) = 1 . 2) x ∈ B ⇒ ϕ (x) = 0 (9.31) This leads to the following simple approximation result. Theorem 74 For any u ∈ MLm (Ω), denote by Γu ⊂ Ω the smallest closed nowhere dense set such that u ∈ C m (Ω \ Γu ). Then there exists a sequence (un ) in C m (Ω) such that ∀ A ⊂ Ω \ Γu compact : ∀ |α| ≤ m : . α α (D un ) converges uniformly to D u on A Proof. For each n ∈ N, we consider the set B 1 (Γ), which is the closure of the set n ∃ x0 ∈ Γ : x∈Ω kx − x0 k ≤ 1 2n and the set C 1 (Γ) = n ∀ x0 ∈ Γ : x∈Ω kx − x0 k ≥ 1 n Clearly, each of the sets B 1 (Γ) and C 1 (Γ) is closed, and for each n ∈ N, B 1 (Γ) and n n n C 1 (Γ) are disjoint. As such, by (9.31), there exists a function ϕn ∈ C ∞ (Ω, [0, 1]) so n that 0 if x ∈ B n1 (Γ) ϕn (x) = 1 if x ∈ C 1 (Γ) n CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 157 Clearly, each of the functions un = ϕn u is C m -smooth and satisfies if x ∈ B 1 (Γ) 0 n un (x) = u (x) if x ∈ C 1 (Γ) n Furthermore, \ B 1 (Γ) = Γ n n∈N and [ C 1 (Γ) = Ω \ Γ n n∈N which completes the proof. Remark 75 It should be noted that the approximations constructed in Theorems 67, 70 and 71 are in fact C ∞ -smooth everywhere except on a closed nowhere dense set. Indeed, each approximating functions is obtained by arranging, in an appropriate way, suitable functions obtained through Proposition 68, which are polynomials in x ∈ Rn . The approximation results presented in this section are fundamental to our approach to constructing generalized solutions to large classes of nonlinear PDEs. In this regard, and as we have mentioned already, it should be noted that none of the results are based on so called ‘advanced mathematics’. Indeed, functional analysis and topology are not used at all. Rather, the techniques used belong to the classical theory of real functions. 9.2 Solutions in Pullback Uniform Convergence Spaces In this section we present the first and basic existence result within the context of the spaces of generalized functions introduced in Chapter 8. In particular, we prove that every system of nonlinear PDEs of the form (8.1) to (8.3) that also satisfies the natural and rather minimal condition (9.1) will have a solution in the pullback type space of generalized functions associated with the particular nonlinear operator (8.10). As a consequence of the way in which the space of generalized functions is constructed, one also obtains immediately the uniqueness of a generalized solution to (8.1). This result amounts to a reformulation of the main existence and uniqueness result obtained through the Order Completion Method [119] in terms of uniform convergence spaces and their completions. Furthermore, and as mentioned in Section CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 158 8.1, such a recasting allows for the application of convergence theoretic techniques to questions related to the structure and regularity of generalized solutions. Such methods may prove to be more suitable to these problems than the order theoretic tools involved in the Order Completion Method. Recall that the space MLm T (Ω) associated with the mapping (8.10) consists of equivalence classes of functions in MLm (Ω)K under the equivalence relation (8.15). With the mapping (8.10) we associate in a canonical way the injective mapping (8.16). In view of the commutative diagram (8.17), the equations (8.13) and (8.18) are equivalent. Since the mapping (8.16) is injective, the initial uniform convergence structure (8.19) on MLm T (Ω) with respect to (8.16) is Hausdorff. As such, we may construct its completion N LT (Ω). In particular, we obtain a commutative diagram m b T MLT (Ω) φ - ML0 (Ω)K ϕ ? N LT (Ω) b] T (9.32) ? - N L (Ω)K where φ and ϕ are the canonical uniformly continuous embeddings associated with K 0 b ] is the unique the completions of MLm (Ω) and ML (Ω) , respectively, and T T b through uniform continuity. Note that, in view of the extension of the mapping T injectivity of the mapping (8.16), it is in fact a uniformly continuous embedding. As such, and as an immediate consequence of Corollary 37, it follows that the mapping b ] is injective. The existence and uniqueness result we present now follows by the T basic approximation result Theorem 67, and the diagram (9.32). Theorem 76 For every f ∈ C 0 (Ω)K that satisfies (9.1), there exists a unique U] ∈ N LT (Ω) such that b ] U] = f T (9.33) Proof. First let us show existence. For every n ∈ N, Theorem 67 yields a closed nowhere dense set Γn ⊂ Ω and a function un ∈ C m (Ω \ Γn ) that satisfies 1 x ∈ Ω \ Γn ⇒ fi (x) − ≤ Ti (x, D) un (x) ≤ fi (x) , i = 1, ..., K (9.34) n Since Γn is closed nowhere dense we associate un with a function vn ∈ MLm (Ω) in a unique way. Indeed, consider for instance the function un (x) if x ∈ Ω \ Γ wn : x 7→ 0 if x ∈ Γ CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 159 Now let vn be the K-dimensional vector valued function with components vni = (I ◦ S) (wni ). Denote by Vn the equivalence class generated by vn under the equivalence relation (8.15). In view of the the fact that each term in the sequence (un ) satisfies un ∈ C m (Ω \ Γn )K , it follows by (3.17), (8.9), (8.11), Proposition 46 and the continuity of the mapping (8.2) that ∀ i = 1, ..., K : . 1 fi − n ≤ Ti vn ≤ fi As such, and in view of the diagram (8.17), it is clear that for each i = 1, ..., K, the 0 b b sequence TVn,i order converges to fi in ML (Ω). Hence the sequence T (Vn ) converges to f in ML0 (Ω)K . It now follows that (Vn ) is a Cauchy sequence in ] MLm f (Ω) so that there exists U ∈ N LT (Ω) that satisfies (9.33). T b : MLm (Ω) → ML0 (Ω)K is a uniformly continuous embedSince the mapping T T ding, the uniqueness of the solution U] found above now follows by Corollary 37. The relative simplicity, and lack of technical difficulty, of the proof of Theorem 76 should be compared to the highly involved techniques used to prove the existence of generalized solutions of a single equation in the context of the usual functional analytic approach, including those involving weak solutions or distributions. Indeed, the existence result presented in Theorem 76 applies to what may be considered as all nonlinear partial differential equations. Furthermore, in contradistinction with the customary functional analytic methods, the nonlinearity of the partial differential operator does not give rise to any additional difficulties. Indeed, the Order Completion Method [119], as well as the theory presented here, do not make any distinction between linear and nonlinear equations, this being one of the main strengths of this approach. Let us now consider the structure of the space N LT (Ω). In this regard, we recall the construction of the completion of a Hausdorff uniform convergence space X [161]. One considers the set XC of all Cauchy filters on X, and an equivalence relation ∼C on XC , defined as ∃ H ∈ XC : F ∼C G ⇔ (9.35) H⊆F ∩G The space X is embedded in X ] = XC / ∼C through X 3 x 7→ λ (x) ∈ X ] CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 160 where λ is the induced convergence structure on X. The uniform convergence structure J ] on X ] is defined as ∃ V ∈ JX : ] U ∈ JX ⇔ [V]X ] ⊆ U In view of the above construction, the space N LT (Ω) consists of all filters F K 0 b on MLm T (Ω) such that the filter T (F) is a Cauchy filter in ML (Ω) , under the equivalence relation (9.35). In particular, the unique generalized solution U] to (8.1) may be represented as the set K b U] {F a filter on MLm T (Ω) : T (F) converges to f in N L (Ω) } (9.36) Note that each classical solution u ∈ C m (Ω)K to (8.1), and also each nonclassical u ∈ MLm (Ω)K , generates the Cauchy filter [U] = {F ⊆ MLm T (Ω) : U = qT u ∈ F } on N L f (Ω), which belongs to the set (9.36). Hence our concept of generalized soT lution is consistent with the usual classical and nonclassical solutions in MLm (Ω)K . Moreover, the generalized solution to (8.3) may be assimilated with usual, nearly finite normal lower semi-continuous functions on Ω, in the sense that there is an injective uniformly continuous mapping b ] : N L (Ω) → N L (Ω)K T T In this regard, we have a blanket regularity for the solutions of a rather large class of systems of nonlinear PDEs. It should be noted that this does not mean that the solution obtained in Theorem 76 is in fact a normal lower semi-continuous function, but rather that it may be constructed using such functions. In particular, since the mapping (8.20) is injective, the space N LT (Ω) of generalized functions may be considered as the subset b ] N L (Ω) T T of the set N L (Ω)K of K-tuples of normal lower semi-continuous functions, equipped with a suitable uniform convergence structure. In view of the above remarks concerning the structure of the unique generalized solution of (8.1), the uniqueness of the solution may be interpreted as follows. As mentioned, each classical solution u ∈ C m (Ω)K of (8.1), as well as each generalized solution u ∈ MLm (Ω)K to the extended equation (8.13), generates a Cauchy filter in MLm T (Ω). Such a Cauchy filter would then belong to the equivalence class (9.36), which is the representation of the generalized solution to (8.1). This class of Cauchy CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 161 filters will also include other, more general filters. In particular, and in view of the commutative diagrams (8.17) and (9.49) we have ∀ F a Cauchy filter on MLm (Ω)K : T (F) → f in ML0 (Ω)K ⇒ qT (F) ∈ U] so that U] also contains every generalized solution of (8.1) in the Sobolev type space of generalized functions N Lm (Ω)K . Therefore, we may interpret the unique generalized solution U] ∈ N LT (Ω) of (8.1) as the set of all solutions of (8.1) in the context of the spaces of generalized solutions associated with the theory of PDEs presented here. 9.3 How Far Can Pullback Go? In Section 9.2 we presented the first and basic existence, uniqueness and regularity result for the solutions of a large class of systems of nonlinear PDEs within the setting of the so called pullback spaces of generalized functions. This result essentially amounts to a reformulation of the fundamental results in the Order Completion Method [119] within the context of uniform convergence spaces. However, the underlying approach to constructing generalized solutions to systems of nonlinear PDEs presented in Sections 8.1 and 9.2 can result in significant improvements in the regularity of generalized solutions of (8.1). In this section we address the issue of improving upon the regularity of the generalized solutions obtained in Section 9.2 within that general and type independent setting. This is done by imposing rather minimal conditions on the smoothness of the nonlinear operator (8.10). In this regard, we consider a system of nonlinear PDEs of the form (8.1) to (8.3), with both the right hand term f in (8.1) as well as the mapping (8.2) are C k -smooth, for some k ∈ N ∪ {∞}. Recall that, in this case, we obtain the mapping (8.46) with domain MLm+k (Ω)K and range contained in MLk (Ω)K , rather than the mapping (8.4) with domain MLm (Ω)K and range contained in ML0 (Ω)K . In this case, we may reproduce the construction (8.15) through (8.17) as follows. We introduce an equivalence relation on MLm+k (Ω)K through u ∼T,k v ⇔ Tu = Tv. (9.37) Exactly as in Section 8.1, we may associate with the mapping (8.46) an injective mapping b k : MLm+k (Ω) → MLk (Ω)K T T,k (9.38) CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 162 in a canonical way so as to produce the commutative diagram T MLm+k (Ω)K - MLk (Ω)K @ @ @ (9.39) @ @ qT,[email protected] bk T @ @ @ R @ MLm+k T,k (Ω) Here qT,k is the canonical quotient mapping associated with the equivalence relation m+k (9.37), and MLm+k (Ω)K / ∼T,k . T,k (Ω) is the quotient space ML In introducing a suitable uniform convergence structure on MLk (Ω), and by implication also on MLm+k T,k (Ω), it should be noted that the Cauchy sequence (Vn ) constructed in Theorem 76 actually satisfies b k Vn converges to f in ML0 (Ω)K T (9.40) As such, there is in fact no need to go beyond the space ML0 (Ω)K when constructing the generalized solution of (8.1). Furthermore, we may observe that, as shown in Proposition 63, the space MLk (Ω) equipped with the usual pointwise order (7.2) is a sublattice of ML0 (Ω). As such, the order convergence structure (4.8) is a well defined convergence structure which induces the order convergence of sequences (2.35). Moreover, recall from Section 2.4 that every reciprocal convergence structure, and in particular every Hausdorff convergence structure, is induced by the complete uniform convergence structure (2.70) through (2.69). In view of (9.40), we equip the space MLk (Ω)K with the uniform convergence structure (2.70) associated with product convergence structure with respect to the order convergence structure λo on each copy of MLk (Ω). That is, ∃ u1 , ..., uk ∈ MLk (Ω)K : ∃ F1 , ..., Fk filters on MLk (Ω)K : . (9.41) U ∈ JλKo ⇔ 1) Fi converges to ui , i = 1, ..., k 2) (F1 × F1 ) ∩ ... ∩ (Fk × Fk ) ⊆ U CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 163 The space MLm+k T,k (Ω) is equipped with the initial uniform convergence structure JT,k with respect to the mapping (9.38). That is, bk × T b k (U) ∈ J K . U ∈ JT,k ⇔ T λo (9.42) Since the mapping (9.38) is injective, it is a uniformly continuous embedding, and the uniform convergence structure (9.42) is Hausdorff. As such, we may construct the completion of MLm+k T,k (Ω), which we denote by N LT,k (Ω), and a unique uniformly continuous mapping K k b] : N L T k T,k (Ω) → ML (Ω) (9.43) so that the diagram bk T MLm+k T,k (Ω) - MLk (Ω)K @ @ @ @ @ b] T k φ @ (9.44) @ @ @ R @ N LT,k (Ω) commutes, with φ the canonical uniformly continuous embedding associated with the completion N LT,k (Ω) of MLm+k T,k (Ω). In particular, in view of Corollary 37, the mapping (9.43) is injective. As in Sections 8.2 and 9.1, and in view of the diagram (9.39), we consider any solution U] ∈ N LT,k (Ω) of the equation b ] U] = f T k (9.45) as a generalized solution of (8.1). The main result of this section is now the following. Theorem 77 Consider a system of nonlinear PDEs of the form (8.1) through (8.2) that also satisfies (9.1). If both the righthand term f in (8.1) and the mapping (8.2) are C k -smooth, then there is a unique U] ∈ N LT,k (Ω) so that b ] U] = f T k CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 164 Proof. Let us first show existence. By Theorem 67, see also Remark 75, there is exists a sequence (Γn ) of closed nowhere dense subsets of Ω, and functions Un ∈ C m+k (Ω \ Γn )K so that ∀ i = 1, ..., K : ∀ x ∈ Ω \ Γn : fi (x) − n1 ≤ Ti (x, D) Un (x) ≤ fi (x) In view of (3.20), Proposition 46 and (8.12) it follows that ∀ i = 1, ..., K : fi − n1 ≤ Ti vn ≤ fi where vn ∈ MLm+k (Ω)K is the function with components vn,i defined through vn,i = (I ◦ S) (Un,i ) . Clearly each sequence (Ti vn ) converges to fi in MLk (Ω) so that the sequence (Tvn ) converges to f in MLk (Ω)K . As such, the sequence (Vn ) associated with (vn ) through (9.44) is a Cauchy sequence in MLm+k T,k (Ω). The existence of a solution now follows by the uniform continuity of the mapping (9.38). Since the mapping (9.38) is a uniformly continuous embedding, the uniqueness of the solution follows by Corollary 37. The structure of the generalized solution obtained in Theorem 77 may be explained in terms of the structure of the completion of a uniform convergence space. In particular, each element U] of the completion N LT,k (Ω) of MLm+k T,k (Ω) may be interpreted as consisting of the equivalence class of Cauchy filters n o b (F) converges to f F a Cauchy filter on MLm+k (Ω) : T (9.46) T,k under the equivalence relation (9.35). In view of (9.46), the unique generalized solution U] ∈ N LT,k (Ω) contains all possible sufficiently smooth solutions of (8.1) within the context of the Order Completion Method. In particular, each solution u ∈ MLm+k (Ω)K of the equation (8.13) generates a Cauchy sequence in MLm+k T,k (Ω) which belongs to the equivalence class (9.46). As such, this notion of solution is consistent with solutions in u ∈ MLm+k (Ω)K , which includes also all classical solutions of (8.1). Furthermore, since the mapping (9.38) is a uniformly continuous embedding, it follows by Corollary 37 that the extended mapping (9.43) associated with (9.38) is an injection. This may be interpreted as a regularity result for the unique generalized solution obtained in Theorem 77, in the sense that each generalized function in N LT,k (Ω) may be assimilated with usual functions in MLk (Ω)K . It should be noted that the generalized solution of (8.1) constructed in Theorem 76 contains also the solution obtained in Theorem 77. Indeed, since the uniform convergence structure (9.41) is finer than the subspace uniform convergence structure CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 165 induced from ML0 (Ω)K , the inclusion mapping i : MLk (Ω)K 3 u 7→ u ∈ ML0 (Ω)K is uniformly continuous. Combining the diagrams (8.17) and (9.44), we obtain an injective uniformly continuous mapping bi : MLm+k (Ω) → MLm (Ω) T T,k (9.47) so that the diagram MLm+k T,k bk T (Ω) bi - MLk (Ω)K i ? m MLT (Ω) b T (9.48) ? - ML0 (Ω)K commutes. Upon extension of the uniformly continuous mappings (8.20), (9.43) and (9.47) to the completions of their respective domains, one obtains the commutative diagram b] T k N LT,k (Ω) bi] - MLk (Ω)K i ? N LT (Ω) b] T (9.49) ? - N L (Ω)K corresponding to (9.48). Since the mappings Tb] , Tbk] and i are all injective by Corollary 37, it follows by the diagram (9.49) that the mapping bi] : N L T,k (Ω) → N LT (Ω) (9.50) CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 166 must also be injective. In particular, if U] ∈ N LT,k (Ω) is a solution of (9.45), then bi] U] ∈ N L (Ω) is a solution of (8.21). T The results on existence, uniqueness and regularity of generalized solutions of (8.1) obtained in this section are, to a certain extent, maximal with respect to the regularity of solutions within the framework of the so called pullback spaces of generalized functions. In this regard, let us now present the construction of generalized solution in an abstract framework. Consider spaces X and Y of functions g : Ω → RK such that f ∈ Y , and the nonlinear partial differential operator T associated with (8.1) acts as T : X → Y. (9.51) Also suppose that Y is equipped with a complete and Hausdorff uniform convergence structure JY which is first countable. Proceeding in the same way as is done in this section, we introduce an equivalence relation on X through u ∼T v ⇔ Tu = Tv, and associate with the mapping (9.51) the injective mapping b : X → Y, T T (9.52) where XT is the quotient space X/ ∼T . In particular, the mapping (9.52) is supposed to satisfy ∀ U ∈ XT : ∀ u∈U: . b Tu = TU If we equip XT with the initial uniform convergence structure JT with respect to the mapping (9.52), then JT is Hausdorff and first countable. In particular, the mapping (9.52) is a uniformly continuous embedding, and extends uniquely to a injective uniformly continuous mapping b ] : X ] → Y, T T (9.53) where X ] is the completion of XT . A generalized solution of (8.1) in this context T is any solution U] ∈ X ] of the equation T b ] U] = f. T (9.54) Note that, in view of the fact that the mapping (9.52) is a uniformly continuous embedding, and (9.53) therefore an injection, the equation (9.54) can have at most one solution. CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 167 Now, in order to obtain the existence of a solution of (9.54), we must construct a sequence (un ) in X so that (Tun ) converges to f in Y . In this regard, the most general such result is given by Theorem 67. As such, within such a general context as considered here, it follows that, if the mapping (8.2) is C k -smooth, we should have X ⊇ MLm+k (Ω)K . (9.55) It now follows by (9.51) and (9.55) that Y ⊇ MLk (Ω)K . (9.56) This may be summarized in the following commutative diagram. T m+k ML - MLk (Ω)K K (Ω) ⊂ ⊂ ? ? - Y T X (9.57) Combining the diagram (9.57) with (9.39) and T X -Y @ @ @ (9.58) @ @ [email protected] b T @ @ @ R @ XT we obtain an injective mapping ιT : MLm+k T,k (Ω) → XT (9.59) CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 168 so that the diagram MLm+k T,k bk T (Ω) - MLk (Ω)K ⊂ ιT b T ? XT (9.60) ? -Y commutes. In particular, if the subspace convergence structure induced on MLk (Ω)K from Y is coarser than the order convergence structure, then the mapping (9.59) is uniformly continuous. Furthermore, in this case the mapping (9.59) extends to an injective uniformly continuous mapping ι] : N LT,k (Ω) → X ] T T (9.61) so that the extended diagram b] T k N LT,k (Ω) ι] T - MLk (Ω)K ⊂ ? ] X T b] T (9.62) ? -Y commutes. The existence of the injective mapping (9.61) may be interpreted as follows. Any pullback type space of generalized functions X ] which is constructed as T above, and subject to the condition of generality of the nonlinear partial differential operator T must contain the space N LT,k (Ω). As such, within the context of general, continuous systems of nonlinear PDEs, the generalized functions in N LT,k (Ω) may be considered to be ‘more regular’ than those in any other space of generalized functions constructed in this way. CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 9.4 169 Existence of Solutions in Sobolev Type Spaces In the previous two section we obtained existence, uniqueness and regularity results for the generalized solutions of large classes of systems of nonlinear PDEs in the context of the so called pullback spaces of generalized functions. However, and as explained at the end of Section 9.3, it is not possible, in the general case of arbitrary systems of continuous nonlinear PDEs, to go beyond the basic regularity properties of such generalized solutions within the framework of the mentioned pullback type spaces of generalized functions. In this regard, there are two obstacles. In particular, the spaces of generalized functions N LT (Ω) are constructed with a given nonlinear operator T in mind. As such, both the generalized functions U] ∈ N LT (Ω), as well as the uniform convergence structure on N LT (Ω), may depend on this nonlinear mapping. A second difficulty, and connected with the first, is that there is no concept of generalized derivative on N LT (Ω). In fact, it is not clear how one should define the derivatives of the generalized functions in N LT (Ω). Within the context of the Sobolev type spaces of generalized functions introduced in Section 8.2, the difficulties discussed above are resolved. In particular, these spaces are constructed independent of any given nonlinear partial differential operator T. Furthermore, the usual partial differential operators Dα : MLm (Ω) → ML0 (Ω) extend uniquely to uniformly continuous mappings Dα] : N Lm (Ω) → N L (Ω) so that we may associate with each generalized function u] ∈ N Lm (Ω) the vector of generalized derivatives D] u = Dα] u |α|≤m ∈ N L (Ω)M . Note also that, in view of the commutative diagram (8.41), the space N Lm (Ω) provides also an additional clarification of the structure of generalized functions in the pullback type spaces of generalized functions, in case the generalized equation (8.39) admits a solution. In this section we investigate the existence of solutions to the generalized equation (8.39). In this regard, the main result is that a large class of systems of nonlinear PDEs have generalized solutions in the Sobolev type spaces of generalized functions N Lm (Ω). We also consider systems of equations that satisfy additional smoothness conditions, such as those introduced in Section 8.3, over and above the mere continuity of the mapping (8.2). Such equations turn out to admit solutions in the Sobolev type spaces of generalized functions N Lm+k (Ω)K , the elements of which have generalized partial derivatives up to order m + k. Here m is the order of the CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 170 system of equations (8.1) and k is the degree of smoothness of the righthand term f and the mapping (8.2). As mentioned, the central result of this section concerns the existence of solutions to (8.39) for a large class of nonlinear partial differential operators. This result, and as is also the case for the existence results presented in Sections 9.2 and 9.3, uses only rather basic topological processes associated with the completion of uniform convergence spaces, and the approximation results presented in Section 9.1, most notably Theorem 70. Theorem 78 Consider a system of nonlinear PDEs of the form (8.1) through (8.3) that satisfies (9.19). Then there is some u] ∈ N Lm (Ω)K such that T] u] = f. Proof. We may apply Theorem 70 to obtain a sequence (Γn ) of closed nowhere dense sets such that ∀ n∈N: , Γn ⊆ Γn+1 and a sequence of functions (Vn ) such that ∀ n∈N: . Vn ∈ C m (Ω \ Γn )K The sequence (Vn ) satisfies ∀ j = 1, ..., K : . fj (x) − n1 ≤ Tj (x, D) Vn (x) ≤ fj (x) , x ∈ Ω \ Γn (9.63) Furthermore, for each |α| ≤ m and every i = 1, ..., K there are sequences λαn,i and µαn,i so that λαn,i , µαn,i ∈ C 0 (Ω \ Γn ), and both ∀ n∈N: ∀ |α| ≤ m : ∀ i = 1, ..., K : 1) λαn,i (x) < Dα Vn,i (x) < µαn,i (x) , x ∈ Ω \ Γn 2) λαn,i (x) < λαn+1,i (x) < µαn+1,i (x) < µαn,i (x) , x ∈ Ω \ Γn+1 (9.64) S ∀ x∈Ω\ n∈N Γn : ∀ |α| ≤ m : ∀ i = 1, ..., K : sup{λαn,i (x) : n ∈ N} = inf{µαn,i (x) : n ∈ N} (9.65) and CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 171 are satisfied. Consider the sequence of functions (un ) in MLm (Ω)K , the components of which are defined through un,i = (I ◦ S) (Vn,i ) , i = 1, ..., K. In view of (9.63) it is clear that the sequence (Tun ) converges tof ∈ ML0 (Ω)K . α Now define, for each i = 1, ..., K and every |α| ≤ m, the sequences λn,i and µαn,i in ML0 (Ω) as α λn,i = (I ◦ S) λαn,i µαn,i = (I ◦ S) µαn,i and Applying (3.17), (3.20) and Propositions 46 to (9.64) it follows that, for each n ∈ N, α α λn,i ≤ λn+1,i ≤ Dα un,i ≤ µαn,i ≤ µαn,i . Furthermore, from (3.20), Definition 53 and (9.65) it follows that each of the filters α [{[λn,i , µαn,i ] : n ∈ N}] is a Cauchy filter in ML0 (Ω). As such, each of the sequences (Dα un,i ) is a Cauchy sequence in ML0 (Ω) so that the sequence (un ) is a Cauchy sequence in MLm (Ω)K . The result now follows by Theorem 65. Theorem 78 states that the generalized equation (8.39) corresponding to the system of nonlinear PDEs (8.1) has a solution in N Lm (Ω)K . Since the mapping (8.38) which defines the left hand side of the equation (8.39) is the unique uniformly continuous extension of the mapping (8.10), the solution u] ∈ N Lm (Ω)K to (8.39) is interpreted as a generalized solution to the system of PDEs (8.1). Furthermore, each of the partial differential operators (8.8) extends uniquely to uniformly continuous mapping (8.30) which represent the generalized derivatives of the generalized functions u] ∈ N Lm (Ω). In particular, and in view of the definition of the uniform convergence structure JD on MLm (Ω) as the initial uniform convergence structure with respect to the family of mappings (8.22), the mapping (8.26) is a uniformly continuous embedding of MLm (Ω) into ML0 (Ω)M . As such, and in view of Corollary 37, the mapping (8.26) extends uniquely to the injective uniformly continuous mapping (8.28). Thus, the commutative diagram (8.29) amounts to a representation of the generalized functions that are the elements of N Lm (Ω) in terms of their generalized derivatives Dα] u] ∈ N L (Ω). The representation of a generalized function u] ∈ N Lm (Ω) in terms of its generalized derivatives may be interpreted as a regularity result for the generalized solutions to (8.1) obtained in Theorem 78. Indeed, each generalized derivative Dα] u]i CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 172 of a component u]i of the solution u] to (8.39) is a nearly finite normal lower semicontinuous functions. As such, we have ∃ ∀ ∀ ∀ B ⊂ Ω of first Baire category : i = 1, ..., K : |α| ≤ m : x∈Ω\B : Dα u]i continuous at x That is, each generalized solution u] ∈ N Lm (Ω)K of (8.1) may be represented as a K-tuple of usual nearly finite normal lower semi-continuous functions which, in view of Proposition 46, are continuous and real valued on a residual subset of Ω. The existence of generalized solutions of (8.1) in the Sobolev type space of generalized functions N Lm (Ω)K also provides some insight into the structure of the generalized solutions in the pullback type spaces of generalized functions. In this regard, consider now a system of nonlinear PDEs of the form (8.1) such that the mapping (8.2) is both open and surjective. In that case, it follows by Theorems 76 and 78 that ∀ f ∈ C 0 (Ω)K : ∃! U] ∈ N LT (Ω) : b ] U] = f T and ∀ f ∈ C 0 (Ω)K : ∃ u] ∈ N Lm (Ω) : T ] u] = f In view of the commutative diagram (8.41) it follows that the unique generalized solution to (8.1) in N LT (Ω) consists precisely of all generalized solutions to (8.1) in N Lm (Ω)K . That is, n o K ] ] ] m ] U = u ∈ N L (Ω) : T u = f . Moreover, and as is explained in Section 8.3, the mapping (8.40) is the canonical quotient mapping associated with the equivalence relation (8.43) on N Lm (Ω)K . The existence result presented in Theorem 78 applies to a general class of systems of nonlinear PDEs. In particular, it requires rather minimal assumptions on the smoothness of the both the nonlinear partial differential operator T, as well as the righthand term f. In this regard, it is only assumed that the righthand term f and the mapping (8.2) that defines the nonlinear operator T through (8.12) are continuous. As is shown in Section 9.3 in connection with generalized solutions in the pullback type spaces of generalized functions, additional regularity assumptions on the CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 173 operator T and the righthand term f, such as those introduced in Section 8.3, may lead to significant improvements in the regularity of generalized solutions. As we shall see shortly, this is also the case for solutions constructed in the Sobolev type spaces of generalized functions. In this regard, we now consider a system of nonlinear PDEs of the form (8.1), with the mapping (8.2) which defines the nonlinear operator through (8.12) a C k smooth function, for some k ∈ N. We may recall from Section 8.3 that, in this case, we obtain a uniformly continuous mapping T : MLm+k (Ω)K → MLk (Ω)K . In particular, this mapping may be represented by the uniformly continuous mappings (8.48) and (8.50) in the commutative diagram (8.49). This shows that the equation (8.57) is equivalent to (8.13). Furthermore, and in view of the uniform continuity of the mappings (8.46), (8.48) and (8.50), each of these mappings extend uniquely the uniformly continuous mappings (8.53), (8.54) and (8.55), respectively. Moreover, since the mapping (8.55) is injective, one obtains also the representation (8.56) for the extended nonlinear partial differential operator (8.53). In this regard, it follows that the generalized equation (8.58) is equivalent to T] u] = f, (9.66) where the unknown u] is supposed to belong to the space N Lm+k (Ω)K . Note, however, that, as is mentioned in Section 8.3, the equivalence with the generalized equation (8.39) breaks down, since in that case the solution is only assumed to have generalized derivatives up to order m. Under assumptions similar to those required for Theorem 78, we now obtain the existence of a solution to the generalized equation (9.66). In this regard, the approximation result Theorem 71 is the key. Theorem 79 Consider a system of nonlinear PDEs of the form (8.1) to (8.3) with the mapping (8.2) and the righthand term f both C k -smooth for some k ∈ N. If the system satisfies (9.30), then there is some u] ∈ N Lm+k (Ω)K such that T] u] = f. Proof. The proof of this result utilizes exactly the same techniques by which Theorem 78 is verified. Hence we do not include it here. The structure of the generalized solution to (8.1) obtained in Theorem 79 may be explained by the same arguments used to describe the generalized solution constructed in Theorem 78. In particular, each solution u] ∈ N Lm+k (Ω)K to (9.66) may be uniquely represented through its generalized derivatives Dα] u]i , |α| ≤ m + k and i = 1, ..., K with each such generalized derivative a nearly finite normal lower semi-continuous function on Ω. CHAPTER 9. EXISTENCE OF GENERALIZED SOLUTIONS 174 The existence results presented in Sections 9.2 and 9.3 for generalized solutions to (8.1) within the context of pullback type spaces of generalized functions apply to a large class of such systems of equations. In particular, every system of linear PDEs, and more generally every system of polynomial type nonlinear PDEs satisfy the condition (9.1), see [119], so that Theorem 76 applies to all such systems of equations. In connection with the existence results presented in this section, namely Theorems 78 and 79, a large class of equations to which these results apply will be discussed in Section 10.2. Chapter 10 Regularity of Generalized Solutions 10.1 Compactness Theorems in Function Spaces In chapter 9 we obtained several existence results for generalized solutions of systems of nonlinear PDEs of the form (8.1) to (8.3). In particular, solutions are constructed in the pullback type spaces of generalized functions, the elements of which may be assimilated with usual nearly finite normal lower semi-continuous functions. Under minimal assumptions on the smoothness of the nonlinear partial differential operator, it is shown that such solutions may in fact be assimilated even with piecewise smooth functions. As is mentioned also in Section 9.3, this is to some extent the maximal regularity for solutions in these pullback type spaces of generalized functions. In Section 9.4 solutions to (8.1) are constructed in the Sobolev type spaces N Lm (Ω)K of generalized functions. These solutions are represented as nearly finite normal lower semi-continuous functions through the injective, uniformly continuous mapping (8.28). These solutions provide additional insight into the structure of the generalized solutions in the pullback type space N LT (Ω) of generalized functions through the commutative diagram (8.41). In particular, the unique generalized solution U] ∈ N LT (Ω) may be represented as the equivalence class n o u] ∈ N Lm (Ω)K : T] u] = f under the equivalence relation (8.43). As discussed in Section 8.2, the generalized derivatives Dα] u] of a generalized function u] ∈ N Lm (Ω) are normal lower semi-continuous functions. In particular, each such generalized derivative is continuous on a residual set, that is, ∃ B ⊂ Ω of first Baire category : ∀ |α| ≤ m : Dα u] is continuous at every x ∈ Ω \ B 175 CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 176 In general these generalized derivatives cannot be interpreted as usual derivatives of real functions. However, as we shall show in Section 10.2, under rather mild assumptions on the nonlinear partial differential operator (8.4), we can construct generalized solutions to (8.1) in N Lm (Ω)K such that ∃ u ∈ MLm (Ω)K : ∀ |α| ≤ m : ∀ i = 1, ..., K : Dα] u]i = Dα ui (10.1) The regularity property (10.1) for generalized solutions u] ∈ N Lm (Ω)K is obtained as an application of suitable compactness theorems in C k (Ω), these being the subject of the present section. Some of the results presented in this section can be found in [1]. We include the proofs, as these are of independent interest. In this regard, the notion of equicontinuity of sets of continuous functions is a key concept. Recall that for a topological space X, a set A ⊂ C (X) is equi-continuous at x0 ∈ X whenever ∀ >0: ∃ V ∈ V x0 : . ∀ u∈A: x ∈ V ⇒ |u (x0 ) − u (x) | < The set A is called equicontinuous on X if it is equicontinuous at each x0 ∈ X, see for instance [81]. Equicontinuity is closely related to compactness in C (X). In particular, the well known theorem of Arzellà-Ascoli is the standard result. Theorem 80 *[110] Consider a subset A of C (X). Then A has compact closure in the topology of uniform convergence on compacta in X whenever A is equicontinuous, and A (x) = {u (x) : u ∈ A} has compact closure in R for each x ∈ X. The converse holds whenever X is locally compact. The special case of Theorem 80 which is relevant in the context of nonlinear PDEs is when X is a suitable subset of Rn and A ⊂ C m (X). In this regard, we at first consider a compact, convex subset X of Rn with nonempty interior. We equip the space C m (X) with the norm 1) |α| ≤ m α (10.2) kukm = sup |D u (x) | . 2) x ∈ X Theorem 81 *[1] With the norm (10.2), the space C m (X) is a Banach space. CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 177 Proof. Let (un ) be a Cauchy sequence in C m (X). Then, in view of the completeness of C 0 (X) with respect to the uniform norm, it follows that ∀ |α| ≤ m : ∃ uα ∈ C 0 (X) : . α α (D un ) converges uniformly to u Denote by u the function uα for |α| = 0. We claim ∀ |α| ≤ m : D α u = uα (10.3) which would complete the proof. In this regard, fix some i0 ∈ {1, ..., n} and consider any c = (ci )i≤n ∈ intX. Define the nontrivial line segment Ii0 (c) as ∀ i 6= i0 : . Ii0 (c) = x ∈ X xi = ci Fix x0 ∈ Ii0 (c). By virtue of the Mean Value Theorem we have ∀ x ∈ Ii0 (c) : ∀ m, n ∈ N : ∃ y ∈ Ii0 (c) : 0 0 (um (x) − un (x)) n (x )) − (um (x ) − u ∂un ∂um = xi0 − x0i0 ∂x (y) − ∂x (y) i i 0 0 From this it follows that, whenever x0 6= x, we have um (x) − um (x0 ) un (x) − un (x0 ) ∂um ∂u n ≤ . − − 0 0 xi0 − xi0 xi0 − xi0 ∂xi0 ∂xi0 As such, and in view of the uniform convergence of the sequence of derivatives, it follows that ∀ ∃ ∀ >0: M ∈ N : m, n ≥ M : um (x)−um (x0 ) − xi0 −x0i0 . < un (x)−un (x0 ) xi0 −x0i 0 Therefore we have u (x) − u (x0 ) un (x) − un (x0 ) < , n ≥ M . − (10.4) 0 xi − x0 x − x i 0 0 i0 i0 ∂un Since the sequence ∂xi converges uniformly to uα , with α = (0, ..., 0, 1, 0, ..., 0), 0 it follows that ∃ N ∈N: ∂un 0 α 0 (x ) − u (x ) ∂xi < , n ≥ N 0 (10.5) CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 178 Set K = sup{M , N }. Since uK ∈ C m (X) it follows that ∃ δ (K) > 0 : ∀ x ∈ Ii0 (c) uK (x)−uK (x0 ) − xi0 −x0i0 ∂uK ∂xi0 . 0 (x ) < , 0 < |xi0 − xi0 | < δ (K) (10.6) 0 From the inequalities (10.4), (10.5) and (10.6) it follows that u (x) − u (x0 ) 0 α − u x < 3 xi − x0 0 i0 ∂u whenever 0 < |xi0 − x0i0 | < δ (K). This proves that ∂x (x0 ) = uα (x0 ). i0 This argument can be replicated for all x ∈ X and all |α| ≤ m. As such, (10.3) must hold, and the proof is complete. The main result of this section, in regard to the space C m (X), is a useful sufficient condition for a set A ⊆ C m (X) to be precompact. As mentioned, equicontinuity is closely connected with compactness in spaces of continuous functions. Indeed, this concept characterizes the compact sets in C (X) through the Arzellà-Ascoli Theorem 80. In this regard, within the context of sets of smooth functions discussed here, a useful class of equicontinuous sets may be easily described. Proposition 82 *[1] A subset A of C 1 (X) is equicontinuous whenever ∃ C>0: ∀ |α| = 1 : ∀ u∈A: kDα uk ≤ C (10.7) Proof. For u ∈ A, and c ∈ X, denote by Du (c) the Frechét derivative of u at x. That is, the linear functional defined through n Du (c) : R 3 x 7→ n X i=1 xi ∂u (c) ∈ R. ∂xi By the Mean Value Theorem [19, 40.4], it follows that ∀ u∈A: ∀ x, y ∈ X : ∃ z on the line segment from x to y : Du (z) (x − y) = u (x) − u (y) This leads to |u (x) − u (y) | ≤ kDu (z) k · kx − yk CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 179 where we take the supremum norm of Du (z). It now follows from (10.7) that ∀ u∈A: ∀ x, y ∈ X : . |u (x) − u (y) | ≤ Ckx − yk For a fixed x ∈ X we now have ∀ u∈A: |u (x) − u (y) | < Cδ whenever kx − yk < δ. As such, for every > 0, and if we choose δ < that ∀ u∈A: ∀ x, y ∈ X : kx − yk < δ ⇒ |u (x) − u (y) | < , M it follows which completes the proof. As an easy application of Proposition 82 we now obtain the following result on the compactness of sets in C m (X). Theorem 83 *[1] Consider a set A ⊆ C m+1 (X). If ∃ C>0: ∀ |α| ≤ m + 1 : , ∀ u∈A: kDα uk ≤ C (10.8) then A is precompact in C m (X), with respect to the topology induced by the norm (10.2). Proof. It is sufficient to show that A is sequentially precompact. In this regard, consider any sequence (un ) in A. From Proposition 82 it follows that, for each |α| ≤ m, the set {Dα u : u ∈ A} is equicontinuous. As such, and in view of (10.8) and Theorem 80, it follows that there exists a subsequence (unk ) of (un ), and functions uα ∈ C 0 (X), for |α| ≤ m, so that each sequence (Dα unk ) converges to uα . The result now follows by Theorem 81. The results obtained so far apply only to functions defined on a compact, convex subset of Rn , the interior of which is nonempty. As such, and in particular in connection with nonlinear PDEs, the power of the respective results resides rather in the sphere of local properties of solutions of a systems of nonlinear PDEs, as apposed to the global properties of such a solution. More precisely, in general the domain of definition Ω of a system of nonlinear PDEs (8.1) is in general neither convex, nor compact. In particular, Ω is typically some open subset of Rn , which may fail to be convex or bounded. In this regard, we introduce the following topology on C m (Ω), with Ω any nonempty and open subset of Rn . CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 180 Definition 84 Denote by τm the topology on C m (Ω) which is generated by the collection of subsets 1) A ⊂ Ω compact S A, {Uα }|α|≤m 2) Uα ⊆ R open, |α| ≤ m of C m (Ω), where for A ⊂ Ω compact and Uα ⊆ R, |α| ≤ m, open ∀ |α| ≤ m : m S A, {Uα }|α|≤m = u ∈ C (Ω) . Dα u (A) ⊆ Uα It is clear that τm does indeed define a topology on C m (Ω). Furthermore, a sequence (un ) in C m (Ω) converges to u ∈ C m (Ω) if and only if ∀ |α| ≤ m : ∀ A ⊂ Ω compact : . (Dα un ) converges Dα u uniformly on A Theorem 85 The topology τ m metrizable and complete. Proof. Let {Ai : i ∈ N} be a collection of compact, convex perfect subsets of Ω such that the family {intAi : i ∈ N} covers Ω, see for instance [58]. Then each of the sets C m (Ai ) is a complete metric space with respect to the metric induced through the norm (10.2). As such, the space Y C m (Ai ) i∈N is complete and metrizable in the product topology. This follows by the Urysohn Metrization Theorem, see for instance [110]. Consider the mapping Y C m (Ai ) E : C m (Ω) → i∈N defined through E (u) = u|Ai i∈N , (10.9) where u|Ai denotes the restriction of u to Ai . Clearly the mapping (10.9) is injective and continuous with a continuous inverse. As such, C m (Ω) is homeomorphic to the Q subspace E (C m (Ω)) of i∈N C m (Ai ), and hence it is metrizable. Completeness now follows by Theorem 81. Now, as mentioned, in the context of nonlinear PDEs, Theorem 83 is inappropriate, since the domain of definition of a system of nonlinear PDEs will in general fail to be compact and convex. However, the metrizable topology τm on C m (Ω), with Ω a nonempty and open subset of Rn , provides a suitable framework for proving similar results in the noncompact case. CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 181 Theorem 86 Let Ω be a nonempty and open subset of Rn . Suppose that the set A ⊂ C m+1 (Ω) satisfies ∀ ∃ ∀ ∀ A ⊂ Ω compact : MA > 0 : |α| ≤ m + 1 : . x∈A: u ∈ A ⇒ |Dα u (x) | < MA Then A is precompact in C m (Ω) with respect to the topology τm . Proof. Note that, by Proposition 82, every set Dα (A), with |α| ≤ m is equicontinuous, and hence, by Theorem 80, precompact in C 0 (Ω) with respect to the compact open topology. As such, each sequence (un ) in A contains a subsequence (unk ) such that ∀ |α| ≤ m : ∃ uα ∈ C 0 (Ω) : . ∀ A ⊂ Ω compact : (Dα unk ) converges uniformly to uα on A The result now follows by the same techniques used in the proof of Theorem 81. Theorems 83 and 86 provide a sufficient condition for a subset A of C m+1 (X), respectively C m+1 (Ω), to be compact in C m (X), respectively C m (Ω). It should be noted that, due to reasons from elementary Banach space theory, such sets need not be compact in C m+1 (X), C m+1 (Ω) respectively. Indeed, suppose sets A ⊂ C m+1 (X) which satisfy (10.8) are compact in C m+1 (Ω). Then the closed unit ball is also compact, so that C m+1 (X) is finite dimensional, which is obviously not the case. In order to obtain compactness of a set A ⊂ C m+1 (X), one must impose additional assumptions on the set A. In particular, in the one dimensional case when X is a compact interval in R, the compact subsets of C m (X) are characterized by the conditions 1) A is bounded w.r.t. the norm (10.2) , 2) {Dm u : u ∈ A} is equicontinuous see for instance [49]. This characterization can be generalized to the arbitrary n dimensional case studied here. However, within the context of nonlinear PDEs, and in particular the construction of generalized solutions though approximation by smooth functions, the condition of equicontinuity of the set of highest order derivatives is rather difficult to satisfy. Theorems 83 and 86 illustrate the phenomenon of ‘loss of smoothness’, which is well known in the field of partial differential equations. In this context, Theorem 83 states that, if you obtain a solution u of a PDE as the limit of a sequence (un ) of functions that are C m -smooth, then u will be only C m−1 -smooth. In this regard, CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 182 consider some iterative method for constructing successive approximations to the solution of a given nonlinear PDE T (x, D) u (x) = f (x) , x ∈ Ω. (10.10) Such an algorithm produces a sequence (un ) of approximate solutions to (10.10). It often happens, see for instance [108] , that if un ∈ C m (Ω) for some n ∈ N, then the next approximation un+1 in the sequence will be less smooth than un . That is, we will typically have un+1 ∈ C m−1 (Ω) \ C m (Ω). This has lead to the consideration of so called smoothing operators, which are supposed to restore the desired regularity of the approximations, see for instance [108], [114] and [117]. In this way, we may come to appreciate another novelty of the method of obtaining generalized solutions of systems of nonlinear PDEs of the form (8.1) presented here. Namely, that no such loss of smoothness of the approximating solutions occur. There is therefore no need to introduce any kind of smoothing operators. However, the approximate solutions are not smooth on the whole domain of definition of the system of equations. Indeed, each such approximate solution un ∈ MLm (Ω)K may be nonsmooth on some closed nowhere dense set Γn ⊂ Ω, and the sets ∪n≥K Γn , for K ∈ N, are typically dense in Ω. As such, one cannot apply the results of this section to obtain even just local regularity of generalized solutions. In the next section we shall present a way, based on Theorem 74, of going beyond these difficulties. 10.2 Global Regularity of Solutions As is mentioned in Section 10.1, in the method for obtaining generalized solutions of systems of nonlinear PDEs presented in this work, and in particular, the construction of solutions in the Sobolev type spaces of generalized functions in Sections 8.2 and Section 9.4, there is no loss of smoothness of the approximating functions. Indeed, recall that the generalized solutions to (8.1) in N Lm (Ω)K are constructed as the limits of sequences in MLm (Ω)K . As such, there is no need to introduce any kind of smoothing operator in order to restore the regularity of successive approximations. However, the results developed in Section 10.1, in particular Theorems 81, 83, 85 and 86 do not apply in the setting of the Sobolev type spaces of generalized functions, since the approximate solutions to (8.1) in MLm (Ω) allow singularities across arbitrary closed nowhere dense subsets of Ω. Furthermore, no suitable generalization of these results to the larger space MLm (Ω) seems possible. Indeed, the compactness results presented in Section 10.1 is based on the Arzellà-Ascoli Theorem 80, which requires pointwise boundedness and equicontinuity of the set of functions. However, note that a function u ∈ ML (Ω), as well as its derivatives, will typically become unbounded in every neighborhood of the singularity set Γu associated with it through (8.5). Furthermore, if a set A of real valued functions on Ω is equicontinuous on Ω, then we must have A ⊆ C 0 (Ω) , CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 183 which is is in general not the case for subsets of MLm (Ω). The aim of this section is to show that there exist generalized solutions to (8.1) in N Lm (Ω)K which are in fact classical solutions everywhere except on a closed nowhere dense set. This will follow as an application of Theorems 74 and 86. Note, however, that Theorem 83, and therefore Theorem 86, involves a loss of smoothness. In particular, given a sequence (un ) of C m -smooth functions on a compact, convex subset X of Rn with nonempty interior, which is bounded with respect to the norm (10.2), we are in general only able to extract a subsequence of (un ) which converges in C m−1 (Ω). As such, and in view of the results presented in Section 9.4, it is clear that some additional smoothness conditions on the nonlinear partial differential operator (8.10), beyond the mere continuity of the mapping (8.2), must be imposed in order to apply Theorem 86. In this regard, we consider a system of nonlinear PDEs of the form (8.1) such that the mapping (8.3), as well as the righthand term f are C k -smooth, for some k ≥ 1. Theorem 79 states that such a system of equations admits a solution u] ∈ N Lm+k (Ω)K whenever the condition (9.30) is satisfied. The main result of this section is a significant strengthening of Theorem 79 in terms of the regularity of the solution constructed. Theorem 87 Suppose that a system of nonlinear PDEs of the form (8.1) satisfies (9.30). Then there exists some u ∈ MLm+k−1 (Ω)K so that Tu = f Proof. By Theorem 79 we have ∃ u] ∈ N Lm+k (Ω)K : . T ] u] = f In particular, there exists a Cauchy sequence (un ) ⊂ MLm+k (Ω)K so that (Tun ) converges to f in N Lk (Ω)K . Furthermore, for each j = 1, ..., K and each |β| ≤ k we have Dβ fj − 1 ≤ Dβ Tj un ≤ Dβ fj . n (10.11) For each n ∈ N there is a closed nowhere dense set Γn ⊂ Ω such that un ∈ C m+k (Ω \ Γn )K . Therefore, in view of Theorem 74, we have ∀ ∃ ∀ ∀ ∀ n∈N: (un,r ) ⊂ C m+k (Ω)K : |α| ≤ m + k : i = 1, ..., K : A ⊂ Ω \ Γn compact : kDα un,r,i − Dα un,i kA → 0 (10.12) CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 184 where k · kA denotes the uniform norm on C 0 (A). It follows by the construction of the sequence (un ) in Theorem 71 that S ∀ x∈Ω\ n∈N Γn : ∀ |α| ≤ m + k : . (10.13) ∀ i = 1, ..., K : |Dα un,i (x) − Dα] ui (x) | → 0 Therefore, and in view of (10.12), it follows that there is a strictly increasing sequence of integers (rn ) so that S ∀ x ∈ Ω \ ( Γn ) : ∀ |α| ≤ m + k : (10.14) ∀ i = 1, ..., K : |Dα un,rn ,i (x) − Dα] uj (x) | → 0 From (10.11), as well as follows that ∀ ∀ ∀ the continuity of the mapping (8.2) and its derivatives, it S x ∈ Ω \ ( Γn ) : |β| ≤ k : j = 1, ..., K : |Dβ Tj un,rn (x) − Dβ fj (x) | → 0 (10.15) In view of Proposition 49 there is a function µ ∈ ML (Ω) so that ∀ n∈N: ∀ |β| ≤ k : . ∀ j = 1, ..., K : |Dβ Tj un,rn | ≤ µ As such, there is a closed nowhere dense set Γ ⊂ Ω so that ∀ ∃ ∀ ∀ A ⊂ Ω \ Γ compact : MA > 0 : |β| ≤ k : . j = 1, ..., K : kDβ Tj un,rn kA ≤ MA , n ∈ N As an application of Theorem 86 it follows that there is βa subsequence of (un,rn ), β which we dote by (vn ), so that D Tvn converges to D f uniformly on compact subsets of Ω \ Γ for each |β| ≤ k − 1. Hence the sequence (Tvn ) converges to f in MLk−1 (Ω)K . By similar arguments as those used above, it may be shown that there is a closed nowhere dense set Γ0 ⊂ Ω so that ∀ ∃ ∀ ∀ A ⊂ Ω \ Γ0 compact : MA0 > 0 : |α| ≤ m + k : i = 1, ..., K : kDα vn,i kA ≤ MA0 , n ∈ N CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 185 Applying Theorem 86, we find that there is a subsequence of (vn ), which we again denote by (vn ), and some v ∈ C m+k−1 (Ω \ Γ0 )K so that ∀ A ⊂ Ω \ Γ0 compact : ∀ |α| ≤ m + k − 1 : ∀ i = 1, ..., K : kDα vn,i − Dα vi kA → 0 Clearly the sequence (un ), the components of which are defined as un,i = (I ◦ S) (vn,i ) , converges in MLm+k−1 (Ω)K to the function u ∈ MLm+k−1 (Ω)K , the components of which are defined as ui = (I ◦ S) (vi ). The result now follows by the uniform continuity of the mapping T : MLm+k−1 (Ω)K → MLk−1 (Ω)K . Theorem 87 states that every system of nonlinear PDEs of the form (8.1) such that the mapping (8.2) and the righthand term f are C k -smooth, has a generalized solution u] ∈ N Lm (Ω)K such that u] ∈ MLm+k−1 (Ω), provided that the condition (9.30) is satisfied. That is, ∃ Γ ⊂ Ω closed nowhere dense : ∃ u ∈ C m+k−1 (Ω \ Γ)K : T (x, D) u (x) = f (x) , x ∈ Ω \ Γ A highly important particular case of Theorem 87 occurs when the system of equations is C 1 -smooth, in the sense that the mapping (8.2) and the righthand term f in (8.1) are C 1 -smooth. In this case, Theorem 87 may be stated as ∃ Γ ⊂ Ω closed nowhere dense : ∃ u ∈ C m (Ω \ Γ)K : T (x, D) = f (x) , x ∈ Ω \ Γ (10.16) We may recall [129] that a property of a system is called a strongly generic property of this system if and only if it holds on an open and dense subset of the domain of definition of that system. Therefore, in view of (10.16) the existence of a classical solution to a system of nonlinear PDEs (8.1) that satisfies (9.30) is a strongly generic property of such a system. A question naturally arises as to the actual scope of the result. That is, can we describe a significantly large class of systems of nonlinear PDEs to which Theorem 87 applies? To this question, the answer is affirmative. In this regard, note that the condition (9.30) is sufficient for the existence of classical solutions to (8.1) on an open and dense subset of the domain of definition Ω of the system of equations. As such, we need only demonstrate that this condition is satisfied. Furthermore, and as we shall shortly see, the condition (9.30) is, in many cases, rather easily verified through some standard techniques in real analysis. In particular, certain CHAPTER 10. REGULARITY OF GENERALIZED SOLUTIONS 186 open mapping type theorems [19] are useful in this regard. We shall exhibit one considerably general class of equations to which Theorem 87 applies. In this regard, we consider a system of K nonlinear PDEs of the form Dt u (x, t) + G (x, t, ..., Dα ui (x, t) , ...) = f (x, t) , i = 1, ..., K (10.17) n where (x, t) ∈ Ω × [0, ∞), with Ω ⊆ R nonempty and open, and G : Ω × [0, ∞) × RM → RK (10.18) a C 1 -smooth mapping. With the system of equations (10.17) we may associate a mapping T : MLm (Ω × [0, ∞))K → ML0 (Ω × [0, ∞))K . (10.19) 1 In particular, and in view of the fact that the mapping (10.21) is C -smooth, the mapping (10.19) satisfies T : MLm+1 (Ω × [0, ∞))K → ML1 (Ω × [0, ∞))K . (10.20) Theorem 88 Consider a system of nonlinear PDEs of the form (10.17). If both the mapping (10.21) and the righthand term f are C 1 -smooth, then the system of equations satisfies (9.30). Proof. Note that, for each β ∈ {0, 1}n , and every j = 1, ..., K there is a jointly continuous mapping Gβj : Ω × [0, ∞) × RL → R (10.21) so that, for each u ∈ C m+1 (Ω)K , we have Dβ Tj u (x, t) = Dβ Dt uj (x, t) + Gβj (x, t, ..., Dα ui (x, t) , ...) , |α| ≤ m + 1. As such, the K × 2n components of the mapping (8.51) may be expressed as Fj,β : (x, t, ξ) 7→ ξj + Gβj (x, t, ..., ξi , ...) , K × 2n < i ≤ L. (10.22) From (10.22) it is clear that the mapping (8.51) is both open and surjective. As such, the condition (9.30) is satisfied. The following is now a straight forward consequence of Theorems 87 and 88. Corollary 89 Consider any system of nonlinear PDEs of the form (10.17). Then there is some u ∈ MLm (Ω × [0, ∞))K such that Tu = f The results on the existence of generalized solutions to (8.1) presented in Chapter 9, as well as the regularity properties of such solutions obtained in this chapter, do not take into account any possible initial and / or boundary conditions that may be associated with a particular system of nonlinear PDEs. In the next chapter, we shall adapt the general method developed over the course of the last three chapters so as to also incorporate such additional conditions. We shall see that, in contradistinction with with usual functional analytic methods, in particular those involving distributions, boundary and / or initial value problems are solved by, essentially, the same techniques that apply to the free problem. Chapter 11 A Cauchy-Kovalevskaia Type Theorem 11.1 Existence of Generalized Solutions The first general and type independent existence and regularity result for the solutions of systems of nonlinear PDEs, namely, Theorem 2, dates back to Cauchy. The first rigorous proof of this result was given by Kovalevskaia [86] more than a century ago. It should noted that, and as mentioned in Section 1.1, the original proof of Theorem 2 does not involve any so called ‘advanced mathematics’. In particular, functional analysis is not used at all. As is well known, ever since Sobolev [148], [149] introduced the sequential method for solving linear and nonlinear PDEs in the setting of Hilbert spaces over 70 years ago, the main, and to some extent nearly exclusive, approach to PDEs has been that of linear functional analysis. However, during the nearly eighty years of functional analysis, the mentioned Cauchy-Kovalevskaia Theorem has not been extended on its own general and type independent grounds. It was only in the 1987 monograph [139], see also [141], that, based on algebraic rather than functional analytic methods, a global version of the local existence and regularity result in Theorem 2 was obtained. The mentioned global version of Theorem 2 still requires both the equation (1.2) and the initial data (1.3) to be analytic. As such, this does not present a generalization of the type of equations that may be solved, but rather the domain of definition of the solution is enlarged. In fact, and in view of Lewy’s impossibility result [97], see also [88], it may appear that an extension of Theorem 2 to nonanalytic equations is highly unlikely. As we shall see in the sequel, this is in fact a misunderstanding. In this section, we present a first in the literature. Namely, we show that a system of K nonlinear PDEs of the form Dtm u (t, y) = G t, y, ..., Dyq Dtp ui (t, y) , ... (11.1) with t ∈ R, y ∈ Rn−1 , m ≥ 1, 0 ≤ p < m, q ∈ Nn−1 , |q| + p ≤ m and with the 187 CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 188 Cauchy data Dtp u (t0 , y) = gp (y) , 0 ≤ p < m, (t0 , y) ∈ S (11.2) on the noncharacteristic analytic hypersurface S = {(t0 , y) : y ∈ Rn−1 } admits a generalized solution u] ∈ N Lm (Rn−1 × R), provided that the mapping G : Rn−1 × R × RM → RK (11.3) is jointly continuous, and the initial data (11.2) satisfies ∀ 0≤p<m: K gp ∈ C m−p (Rn−1 ) (11.4) That is, we give the first extension of the Cauchy-Kovalevskaia Theorem, on its own general and type independent grounds, to equations which are not analytic. Furthermore, if the mapping (11.3) is C 1 -smooth, and the initial condition (11.2) satisfies ∀ 0≤p<m: K gp ∈ C m−p+2 (Rn−1 ) (11.5) then the generalized solution of (11.1) through (11.2) is in fact a classical solution in the sense that ∃ Γ ⊂ Rn−1 × R closed nowhere dense : 1) Γ ∩ S closed nowhere dense in S 2) u] ∈ C m ((Rn−1 × R) \ Γ) (11.6) It is clear that the existence of a solution of the system of nonlinear PDEs (11.1) is a straight forward consequence of the general existence results proved in Chapter 9. Furthermore, the regularity property (11.6) of the solution follows easily from the results in Chapter 10. In order to also incorporate the initial data (11.2), the methods presented in the mentioned chapters need only be adapted slightly. In this way, we come to appreciate yet another key feature of the solution method for systems of nonlinear PDEs presented in Chapters 6 through 10. Namely, that initial and / or boundary value problems may be solved by essentially the same techniques that apply to the free problem. This should be compared with the customary functional analytic methods, in particular those involving distributions, where such additional conditions often lead to significant complications which often require entirely new techniques. CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 189 In order to incorporate the initial condition (11.2) into our solution method, we introduce the following spaces of functions. Denote by MLm g (Ω) the set ∀ i = 1, ..., K : ∀ 0≤p<m: K m ∀ q ∈ Nn−1 , 0 ≤ |q| + p ≤ m : MLm (Ω) = u ∈ ML (Ω) g qp 1) Dyt ui (y, t0 ) = Dq gp,i (y) , y ∈ Rn−1 qp 2) Dyt ui continuous at (y, t0 ) where Ω = Rn−1 ×R. For each i = 1, ..., K, every 0 ≤ p < m and each q ∈ Nn−1 such that 0 ≤ |q| + p ≤ m, we consider the space ML0i,q,p (Ω), which is defined through ∀ y ∈ Rn−1 : 1) u (y, t0 ) = Dq gp,i (y) ML0i,q,p (Ω) = u ∈ ML0 (Ω) . 2) u continuous at (y, t0 ) Clearly, for every 0 ≤ p < m, and p ∈ Nn−1 such that 0 ≤ |q| + p ≤ m, and each i = 1, ..., K we may define the partial differential operators qp 0 Di,yt : MLm g (Ω) → MLi,q,p (Ω) , (11.7) as in Chapter 8, through qp qp Di,yt u = (I ◦ S) Dyt ui . m , is defined in a similar way, namely, as The partial differential operators Di,t 0 m m Di,t : MLm g (Ω) 3 u 7→ (I ◦ S) (Dt ui ) ∈ ML (Ω) . (11.8) The method for constructing generalized solutions to the initial value problem (11.1) to (11.2) presented here is essentially the same as that used in the case of arbitrary systems of nonlinear PDEs, which is developed in Chapters 8 and 9. In particular, generalized solutions are constructed as elements of the completion of the space MLm g (Ω), equipped with a suitable uniform convergence structure. In this regard, the space ML0 (Ω) carries the uniform order convergence structure introduced in Chapter 7. We introduce the following uniform convergence structure on ML0i,q,p (Ω). Definition 90 Let Σ consist of all nonempty order intervals in ML0i,q,p (Ω). Let Ji,q,p denote the family of filters on ML0i,q,p (Ω) × ML0i,q,p (Ω) that satisfy the following: There exists k ∈ N such that ∀ j = 1, ..., k : ∃ Σj = (Inj ) ⊆ Σ : j 1) In+1 ⊆ Inj , n ∈ N 2) ([Σ1 ] × [Σ1 ]) ∩ ... ∩ ([Σk ] × [Σk ]) ⊆ U (11.9) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 190 where [Σj ] = [{I : I ∈ Σj }]. Moreover, for each j = 1, ..., k and every open subset V of Ω one has ∃ uj ∈ ML0i,q,p (Ω) : j ∩n∈N In|V = {uj }|V or j ∩n∈N In|V =∅ (11.10) Proposition 91 The family of filters Ji,q,p on ML0i,q,p (Ω) × ML0i,q,p (Ω) is a Hausdorff uniform convergence structure. Furthermore, a filter F on ML0i,q,p (Ω) converges to u ∈ ML0i,q,p (Ω) if and only if there exists a family ΣF = (In ) of nonempty order intervals on ML0i,q,p (Ω) such that 1) In+1 ⊆ In , n ∈ N ∀ V ⊆ Ω nonempty and open : 2) ∩n∈N In|V = {u}|V and [ΣF ] ⊆ F. Proof. The first four axioms of Definition 21 are clearly fulfilled, so it remains to verify ∀ U, V ∈ Jo : U ◦ V exists ⇒ U ◦ V ∈ Jo (11.11) In this regard, take any U, V ∈ Jo such that U ◦ V exists, and let Σ1 , ..., Σk and Σ01 , ..., Σ0l be the collections of order intervals associated with U and V, respectively, through Definition 90. Set Φ = {(l, j) : [Σl ] ◦ [Σ0j ] exists} Then, by Lemma 54 U ◦V ⊇ \ {([Σl ] × [Σl ]) ◦ ([Σj ] × [Σj ]) : (l, j) ∈ Φ} (11.12) Now (l, j) ∈ Φ if and only if ∀ m, n ∈ N : l Im ∩ Inj 6= ∅ For any (l, j) ∈ Φ, set Σl,j = Inl,j where, for each n ∈ N Inl,j = [inf Inl ∧ inf Inj , sup Inl ∨ sup Inj ] Now, using (11.12), we find \ \ U ◦ V ⊇ {[Σl ] × [Σj ] : (l, j) ∈ Φ} ⊇ {[Σl,j ] × [Σl,j ] : (l, j) ∈ Φ} Clearly each Σl,j satisfies 1) of (11.9). Since ML0 (Ω) is fully distributive, see Corollary 52, (11.10) follows by Lemma 92. The second part of the proposition follows by the same arguments used in the proof of Theorem 56. The proof of Proposition 91 relies on the following. CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 191 Lemma 92 The set ML0i,q,p (Ω) is a lattice with respect to the pointwise order. Proof. Consider any functions u, v ∈ ML0i,q,p (Ω), and set w = sup{u, v} ∈ ML0 (Ω). In view of Theorem 45 it follows that w (x) = (I ◦ S) (ϕ) (x) , x ∈ Ω where ϕ (x) = sup{u (x) , v (x)}, x ∈ Ω. Assume that ∃ y 0 ∈ Rn : ∃ a∈R: . w (y0 , t0 ) > a > Dq gp,i (y0 ) (11.13) It then follows that S (ϕ) (y0 , t0 ) > a > Dq gp,i (y0 ). Therefore ∀ δ>0: ∃ (yδ , tδ ) ∈ Bδ (y0 , t0 ) : ϕ (yδ , tδ ) > a > Dq gp,i (y0 ) so that we obtain a sequence (yn , tn ) in Ω which converges to (y0 , t0 ) and satisfies ∀ n∈N: u (yn , tn ) > a > Dq gi,p (y0 ) = u (y0 , t0 ) (11.14) ∀ n∈N: . v (yn , tn ) > a > Dq gi,p (y0 ) = v (y0 , t0 ) (11.15) or But both u and v are continuous at (y, t0 ) for each y ∈ Rn , which contradicts (11.14) to (11.15). Hence (11.13) cannot hold, so that w ∈ ML0i,q,p (Ω). The existence of the infimum of u and v follows in the same way. The completion of ML0i,q,p (Ω) may be represented as a suitable space of nearly finite normal lower semi-continuous functions. In particular, we consider the space ∃ λ, µ ∈ ML0i,q,p (Ω) : N Li,q,p (Ω) = u ∈ N L (Ω) . λ≤u≤µ Note that ML0i,q,p (Ω) ⊂ N Li,q,p (Ω). As such, in order to show that N Li,q,p (Ω) is the completion of ML0i,q,p (Ω), we must introduce a Hausdorff uniform conver] gence structure Ji,q,p on N Li,q,p (Ω) in such a way that the following conditions are satisfied: ] 1. N Li,q,p (Ω) is complete with respect to Ji,q,p . CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 192 2. N Li,q,p (Ω) contains ML0i,q,p (Ω) as a dense subspace. 3. If Y is a complete, Hausdorff uniform convergence space, then any uniformly continuous mapping ϕ : ML0i,q,p (Ω) → Y extends in a unique way to a uniformly continuous mapping ϕ] : N Li,q,p (Ω) → Y . In this regard, the definition of the uniform convergence structure on N Li,q,p (Ω) is similar to Definition 58 ] Definition 93 Let Ji,q,p denote the family of filters on N Li,q,p (Ω)×N Li,q,p (Ω) that satisfy the following: There exists k ∈ N such that ∀ j = 1, ..., k : ∃ (λjn ) , (µjn ) ⊆ ML0i,p (Ω) : ∃ uj ∈ ML0i,p (Ω) : 1) λjn ≤ λjn+1 ≤ µjn+1 ≤ µin , n ∈ N Tk 2) j=1 (([Σj ] × [Σj ]) ∩ ([uj ] × [uj ])) ⊆ U (11.16) where each uj ∈ N Li,q,p (Ω) satisfies uj = sup{λjn : n ∈ N} = inf{µjn : n ∈ N}. Here Σj = {Inj : n ∈ N} with Inj = {u ∈ MLi,q,p (Ω) : λjn ≤ u ≤ µjn }. ] That the family of filters Ji,q,p does indeed constitute a Hausdorff uniform con] vergence structure on N Li,q,p (Ω) can easily be seen. Indeed, Ji,q,p is nothing but the uniform convergence structure associated with the following Hausdorff convergence structure through (2.70): A filter F on N Li,q,p (Ω) converges to u ∈ N Li,q,p (Ω) if and only if ∃ (λn ) , (µn ) ⊂ ML0i,q,p (Ω) : 1) λn ≤ λn+1 ≤ µn+1 ≤ µn , n ∈ N . 2) ∩n∈N [λn , µn ]|V = {u}|V , V ⊆ Ω open 3) [{[λn , µn ] :n ∈ N}] ⊆ F Theorem 94 The space N Li,q,p (Ω) equipped with the uniform convergence struc] ture Ji,q,p is the uniform convergence space completion of ML0i,q,p (Ω). Proof. That N Li,q,p (Ω) is complete follows immediate by our above remarks. Furthermore, it is clear that the subspace uniform convergence structure on ML0i,q,p (Ω) is equal to Ji,q,p . The extension property for uniformly continuous mappings follows by a straight forward argument. An important property of the uniform convergence space ML0i,q,p (Ω) and its completion N Li,q,p (Ω) relates to the inclusion mapping i : ML0i,q,p (Ω) → ML0 (Ω) (11.17) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 193 and its extension through uniform continuity i] : N Li,q,p (Ω) → N L (Ω) . (11.18) Indeed, it is clear form Definitions 53 and 90 that the mapping (11.17) is in fact uniformly continuous. Similarly, the inclusion mapping i0 : N Li,q,p (Ω) → N L (Ω) (11.19) is uniformly continuous. Since the mappings (11.18) and (11.19) coincide on a dense subset of N Li,q,p (Ω), it follows that (11.18) is simply the inclusion mapping (11.19). This is related to the issue of consistency of generalized solutions of (11.1) to (11.2) that we construct in the sequel with solutions in the space N Lm (Ω)K , that is, solutions of the generalized equation (8.39). We will discuss this in some detail in what follows, after the uniform convergence structure on MLm g (Ω) has been introduced. In this regard, the uniform convergence structure Jg on MLm g (Ω) is defined as the initial uniform convergence structure with respect to the mappings (11.7) to m (11.8). That is, a filter U on MLm g (Ω) × MLg (Ω) belongs to Jg if and only if ∀ i = 1, ..., K: , m m Di,t × Di,t (U) ∈ Jo and ∀ 0≤p<m: ∀ q ∈ Nn−1 , 0 < |q| + p ≤ m : , ∀ i = 1, ..., K : qp qp Di,yt × Di,yt (U × U) ∈ Ji,q,p Clearly the family consisting of the mappings (11.7) through (11.8) separates the points of MLm g (Ω). As such, the uniform convergence structure Jg is uniformly Hausdorff. In particular, and in view of Theorem 44, the mapping Y m 0 D : MLg (Ω) → MLi,q,p (Ω) × ML0 (Ω)K which is defined through qp m D : u 7→ ..., Di,yt u, ...Di,t u, ... (11.20) is a uniformly continuous embedding. As such, it follows from Theorem 37 that the mapping (11.20) extends to an injective, uniformly continuous mapping Y D] : N Lm (Ω) → N L (Ω) × N L (Ω)K i,p g CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 194 m where N Lm g (Ω) denotes the uniform convergence space completion of MLg (Ω). In particular, for each i = 1, ..., K 0 ≤ p < m, and each q ∈ Nn−1 such that 0 ≤ p + |q| ≤ m the diagrams D] m N Lg (Ω) Q -( N Li,q,p (Ω)) × N L (Ω)K @ @ @ (11.21) @ @ qp] Di,yt @ πi,p,q @ @ @ R @ N Li,q,p (Ω) and D] N Lm g (Ω) Q -( N Li,q,p (Ω)) × N L (Ω)K @ @ @ (11.22) @ @ m] Di,t @ πi @ @ @ R @ N L (Ω) qp] m] commute, with πi,q,p and πi the projections, and Di,yt and Di,t the extensions through uniform continuity of the mappings (11.7) and (11.8), respectively. The meaning of the diagrams (11.21) and (11.22) is twofold. In the first instance, it explains the regularity of generalized functions in N Lm g (Ω). In particular, ] each generalized partial derivative of a generalized function u ∈ N Lm g (Ω) is a nearly finite normal lower semi-continuous function. Therefore, funcQ each such generalized 0≤p<m tion may be represented as an element of the space N Li,p (Ω) ×N L (Ω)L . i≤K Secondly, these diagrams state that each generalized function u] ∈ N Lm g (Ω) satisfies CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 195 the initial condition (11.2) in the sense that ∀ i = 1, ..., K : ∀ 0≤p<m: . ∀ q ∈ Nn−1 , 0 ≤ p + |q| ≤ m : p] ] Di,t u (t0 , y) = gp,i (t0 , y) , y ∈ Rn−1 (11.23) With the system of nonlinear PDEs (11.1) we may associate a mapping K 0 T : MLm g (Ω) → ML (Ω) , (11.24) the components of which are defined as in (8.12). Generalized solutions to the initial value problem (11.1) and (11.2) are obtained by suitably extending the mapping (11.24) to a mapping K T] : N Lm g (Ω) → N L (Ω) . (11.25) Such an extension is obtained through the uniform continuity of the mapping (11.24). In this regard, we have the following. Theorem 95 The mapping (11.24) is uniformly continuous. Proof. It follows from (11.17) through (11.18) that the inclusion mapping K m i : MLm g (Ω) → ML (Ω) is uniformly continuous. The result now follows from the commutative diagram T MLm g (Ω) - ML0 (Ω)K @ @ @ (11.26) @ @ i @ T0 @ @ @ R @ MLm (Ω)K and Theorem 65, with T0 the mapping defined on MLm (Ω)K through the nonlinear partial differential operator. In view of Theorem 95 the mapping (11.24) extends in a unique way to a uniK formly continuous mapping with domain N Lm g (Ω) and range contained in N L (Ω) , CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 196 which is the generalized nonlinear partial differential operator (11.25). As such, the generalized initial value problem corresponding to (11.1) and (11.2) is the single equation T] u] = 0, (11.27) where 0 denotes the the element in N L (Ω)K with all components identically 0. A solution to (11.27) is interpreted as a generalized solution to (11.1) through (11.2) based on the facts that the mapping (11.25) is the unique and canonical extension of (11.24), and each solution of (11.27) satisfies the initial condition in an extended sense, as mentioned in (11.23). Furthermore, in view of (11.17) to (11.19) and the diagram (11.26) we obtain the commutative diagram T] m N Lg (Ω) - N L (Ω)K @ @ @ @ @ (11.28) T]0 ] i @ @ @ @ R @ N Lm (Ω)K with i] injective and T]0 the uniformly continuous extension of the mapping T0 : MLm (Ω)K → ML0 (Ω)K associated with the system of nonlinear PDEs (11.1). In particular, the mapping i] is the inclusion mapping. As such, each solution u] ∈ N Lm g (Ω) of (11.27) is a solution of the system of nonlinear PDEs (11.1) in the sense of the Sobolev type spaces of generalized functions introduced in Section 8.2. In this regard, the main result of this section is the following. K Theorem 96 For each 0 ≤ p < m, let gp ∈ C m−p (Rn−1 ) . Then there is some u] ∈ N Lm g (Ω) so that T] u] = 0. Proof. Let us express Ω = Rn−1 × R as [ Ω= Cν ν∈N CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 197 where, for ν ∈ N, the compact sets Cν are n-dimensional intervals Cν = [aν , bν ] (11.29) with aν = (aν,1 , ..., aν,n ), bν = (bν,1 , ..., bν,n ) ∈ Rn and aν,j ≤ bν,j for every j = 1, ..., n. We assume that the Cν , with ν ∈ N, are locally finite, that is, ∀ x∈Ω: ∃ V ⊆ Ω a neighborhood of x : {ν ∈ N : Cν ∩ V 6= ∅} is finite (11.30) Such a partition of Ω exists, see for instance [58]. We also assume that, for each ν ∈ N, S ∩ Cν = ∅ or S ∩ IntCν 6= ∅ (11.31) where S is the noncharacteristic hypersurface S = {(y, t0 ) : y ∈ Rn−1 } For the sake of convenience, let us write x = (y, t) for each (y, t) ∈ Rn−1 × R. Let F : Ω × RM → RK is the mapping that defines the nonlinear operator T through T (x, D) u (x) = F (x, ..., Dα ui (x) , ...) . Fix ν ∈ N such that (11.31) is satisfied. In view of the fact that the mapping F is both open and surjective, we have ∀ x1 = (y1 , t1 ) ∈ Cν : ∃ ξ (x1 ) ∈ RM , F (x1 , ξ (x1 )) = 0 : ∃ δ, > 0 : (11.32) kx − x1 k < δ 1) {(x, 0) : kx − x1 k < δ} ⊂ int (x, F (x, ξ)) kξ − ξ (x1 ) k < 2) F : Bδ (x1 ) × B2 (ξ (x1 )) → RK open In particular, if t1 = t0 , we may take ξ (x1 ) = (ξiq,p , ξim ) such that ∀ i = 1, ..., K : ∀ 0≤p<m: ∀ q ∈ Nn−1 , 0 < |q| + p ≤ m : ξiq,p = Dq gp,i (y1 ) (11.33) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 198 For each x1 ∈ Cν , fix ξ (x1 ) ∈ RM in (11.32) so that (11.33) is satisfied in case t1 = t0 . Since Cν is compact, it follows from (11.32) that ∃ δ>0: ∀ x1 ∈ Cν : ∃ x1 > 0 : kx − x1 k < δ 1) {(x, 0) : kx − x1 k < δ} ⊂ int (x, F (x, ξ)) kξ − ξ (x1 ) k < x1 K 2) F : Bδ (x1 ) × B2x0 (ξ (x1 )) → R open (11.34) Subdivide Cν into n-dimensional intervals Iν,1 , ..., Iν,µν with diameter not exceeding δ such that their interiors are pairwise disjoint and, for each j = 1, ..., µν , Iν,j ∩ S = ∅ (11.35) intIν,j ∩ S 6= ∅ (11.36) or If aν,j with j = 1, ..., µν is the center of the interval Iν,j that satisfies (11.35), then by (11.34) we have ∃ ν,j > 0 : x ∈ Iν,j 1) {(x, 0) : x ∈ Iν,j } ⊂ int (x, F (x, ξ)) kξ − ξ (aν,j ) k < ν,j 2) F : Iν,j × B2ν,j (ξ (aν,j )) → RK open (11.37) On the other hand, if Iν,j satisfies (11.36), set aν,j equal to the midpoint of S ∩ Iν,j . Then we obtain (11.37) by (11.34) such that (11.33) also holds. Take 0 < γ < 1 arbitrary but fixed. In view of Proposition 68 and (11.37), we have ∀ x1 ∈ Iν,j : ∃ Ux1 = U ∈ C m (Rn )K : ∃ δ = δx1 > 0 : , |α|≤m α 1) (D Ui (x))i≤K ∈ Bν,j (ξ (aν,j )) x ∈ Bδ (x1 ) ∩ Iν,j ⇒ 2) i ≤ K ⇒ γ < Ti (x, D) U (x) < 0 with α = (q, p). Furthermore, if Iν,j satisfies (11.36), then we also have ∀ ∀ ∀ ∀ i = 1, ..., K : 0≤p<m: q ∈ Nn−1 , 0 < |q| + p ≤ m : . y ∈ Rn−1 : pq Dyt Ui (y, t0 ) = Dq gp,i (y) As above, we may subdivide Iν,j into pairwise disjoint, n-dimensional intervals Jν,j,1 , ..., Jν,j,µν,j so that for k = 1, ..., µν,j we have ∃ Uν,j,k = U ∈ C m (Rn )K : ∀ x ∈ Jν,j,k : |α|≤m 1) Dα Ui (x)i≤K ∈ Bν,j (ξ (aν,j )) , |α| ≤ m 2) i ≤ K ⇒ fi (x) − γ < Ti (x, D) U (x) < fi (x) (11.38) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 199 and Jν,j,k ∩ S = ∅ (11.39) intIν,j,k ∩ S 6= ∅. (11.40) or Furthermore, whenever Jν,j,k satisfies (11.40), we have ∀ ∀ ∀ ∀ i = 1, ..., K : 0≤p<m: q ∈ Nn−1 , 0 < |q| + p ≤ m : . y ∈ Rn−1 : qp Dyt Ui (y, t0 ) = Dq gp,i (y) In particular, in this case we may simply set Ui (y, t) = m−1 X (t − t0 )p gp,i (y) + wi (t) p=0 for a suitable function wi ∈ C m (R) that satisfies ∀ 0≤p<m: . (p) wi (t0 ) = 0 Set Γ1 = Ω \ µν [ µν,j [ ν∈N j=1 k=1 [ !!! . intJν,j,k and V1 = µν,j µν X X X ν∈N j=1 !! χJν,j,k Uν,j,k k=1 where χJν,j,k is the characteristic function of Jν,j,k . Then Γ1 is closed nowhere dense, and V1 ∈ C m (Ω \ Γ1 )K . Furthermore, S ∩ Γ1 is closed nowhere dense in S and ∀ ∀ ∀ ∀ i = 1, ..., K : 0≤p<m: q ∈ Nn−1 , 0 < |q| + p ≤ m : . (y, t0 ) ∈ S \ (S ∩ Γ1 ) : qp Dyt V1,i (y, t0 ) = Dq gp,i (y) In view of (11.38) we have, for each i = 1, ..., K −γ < Ti (x, D) V1 (x) < 0, x ∈ Ω \ Γ1 CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 200 Furthermore, for each ν ∈ N, for each j = 1, ..., µν , each k = 1, ..., µν,j , each |α| ≤ m and every i = 1, ..., K we have x ∈ intJν,j,k ⇒ ξiα (aν,j ) − < Dα V1,i (x) < ξiα (aν,j ) + (11.41) For 0 ≤ p < m, define the functions λα1,i , µα1,i ∈ C 0 (Ω \ Γ1 ), where α = (p, q) with |q| = 0, as α if x ∈ intIν,j,k and Iν,j,k ∩ S = ∅ ξi (aν,j ) − 2ν,j α λ1,i (x) = p Dt V1,i (y, t) − vν,j (t) if x ∈ intIν,j,k and Iν,j,k ∩ S 6= ∅ and µα1,i (x) = α ξi (aν,j ) + 2ν,j if x ∈ intIν,j,k and Iν,j,k ∩ S = ∅ Dtp V1,i (y, t) + vν,j (t) if x ∈ intIν,j,k and Iν,j,k ∩ S 6= ∅ Here vν,j is a continuous, real valued function on R such that vν,j (t0 ) = 0 (11.42) 0 < vν,j (t) < 2ν,j , t ∈ R (11.43) and For all other α, consider the functions λα1,i (x) = ξiα (aν,j ) − 2ν,j if x ∈ intIν,j and µα1,i (x) = ξiα (aν,j ) + 2ν,j if x ∈ intIν,j . Then it follows by (11.41) that λα1,i (x) < Dα V1,i (x) < µα1,i (x) , x ∈ Ω \ Γ1 and µα1,i (x) − λα1,i (x) < 4ν,j , x ∈ intIν,j Applying (11.37) restricted to Ω \ Γ1 , and proceeding in a fashion similar as above, we may construct, for each n ∈ N such that n > 1, a closed nowhere dense set Γn ⊂ Ω such that Γn ∩ S closed nowhere dense in S, CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 201 a function Vn ∈ C m (Ω \ Γn )K and functions λαn,i , µαn,i ∈ C 0 (Ω \ Γn ) so that, for each i = 1, ..., K − γ < Ti (x, D) Vn (x) < 0, x ∈ Ω \ Γn . n (11.44) and for every |α| ≤ m λαn−1,i (x) < λαn,i (x) < Dα Vn,i (x) < µαn,i (x) < µαn−1,i (x) , x ∈ Ω \ Γn (11.45) and µαn,i (x) − λαn,i (x) < 4ν,j , x ∈ (intIν,j ) ∩ (Ω \ Γn ) . n (11.46) Furthermore, for each 0 ≤ p < m and q ∈ Nn−1 so that 0 ≤ |q| + p ≤ m we have qp / S ∩ Γn Dyt Vn,i (y, t0 ) = λαn,i (y, t0 ) = µαn,i (y, t0 ) = Dq gp,i (y) , (y, t0 ) ∈ where α = (p, q). Notice that the functions un , the components of which are defined through un,i = (I ◦ S) (Vn,i ) α α belongs to MLm g (Ω). In view of (11.45) it follows that the functions λn,i , µn,i ∈ ML0 (Ω), which are defined as α λn,i = (I ◦ S) λαn,i , µαn,i = (I ◦ S) µαn,i , satisfies α α λn−1,i ≤ λn,i ≤ Dα un,i ≤ µαn,i ≤ µαn−1,i Furthermore, in case α = (p, q) with q ∈ Nn−1 such that 0 ≤ p + |q| ≤ m, then α λn,i , µαn,i ∈ ML0i,q,p (Ω). It now follows by (11.46) that the sequence (un ) is a Cauchy sequence in MLm g (Ω). Moreover, (11.44) implies that the sequence (Tun ) converges 0 to 0 in ML (Ω)K . The result now follows from Theorem 95. We have shown that the initial value problem (11.1) through (11.2) admits a generalized solution in the space N Lm g (Ω). In particular, and in view of the commutative diagram (11.26), the generalized solution constructed in Theorem 96 is a generalized solution of the system of nonlinear PDEs (11.1) in the sense of the Sobolev type spaces of generalized functions introduced in Section 8.2. Furthermore, this solution satisfies the initial data (11.2) in the sense that ∀ 0≤p<m: ∀ q ∈ Nn−1 , 0 ≤ |q| + p ≤ m : . ∀ y ∈ Rn−1 : qp] ] Dyt,i u (y, t0 ) = Dq gp,i (y) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 202 As such, it follows from Proposition 46 the singularity set ∃ |α| ≤ m : ∃ i = 1, ..., K : (y, t) ∈ Ω Diα] u not continuous at (y, t) of the solution is of first Baire category. This result is a first in the literature. Indeed, during the seventy years since Sobolev introduced functional analysis in the study of PDEs, the Cauchy-Kovalevskaia Theorem 2 has not been extended, in the context of any of the usual spaces of generalized functions, on its own general and type independent grounds. The only improvement upon this result which has been obtained to date is related to the domain of definition of the solutions. In particular, it has been shown [139] that the Cauchy problem (1.2) and (1.3) admits a generalized solution in a suitable algebra of generalized functions, which is defined on the whole domain of definition of the respective system of equations (1.2). Furthermore, such a solution is analytic everywhere except possibly for a closed nowhere dense set. However, the class of equations to which the result applies is the same as in the original version of the theorem, which was obtained more than a hundred years ago [86]. Theorem 96 delivers the existence of global generalized solutions of the initial value problem (11.1) and (11.2), as described above, provided only that the mapping (11.3) is continuous, and that the initial data satisfies rather obviously necessary smoothness conditions. As such, it is an extension of both the original Cauchy Kovalevskaia Theorem 2, and the global version of that result obtained in [139] in the context of the Sobolev type spaces of generalized functions introduced in Chapter 8. 11.2 Regularity of Generalized Solutions The results presented in the previous section, in particular Theorem 96, concern only the first and basic properties regarding existence and regularity of solutions of the Cauchy problem (11.1) to (11.2). In contradistinction with Theorem 2, and the global version of that result [139], the solution cannot be interpreted as a classical solution on any part of the domain of definition of the equation. However, and as we shall see in the sequel, such additional regularity properties of the solution may be obtained with only minimal additional assumptions on the nonlinear partial differential operator (11.24). In particular, such conditions do not involve any restrictions on the type of equation, but instead involves only very mild assumptions on the smoothness of the mapping (11.3) and the initial data (11.2). In this regard, consider now a system of nonlinear PDEs of the form (11.1) such that the mapping (11.3) is C 1 -smooth. Furthermore, we shall assume that the initial data (11.2) satisfies (11.5). In this case, and in view of the results presented in CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 203 Chapter 10, it is clear that the system of equations (11.1) admits a generalized solution u] ∈ N Lm (Ω)K that satisfies ∃ ∃ ∀ ∀ Γ ⊂ Rn−1 × R closed nowhere dense : U ∈ C m (Ω \ Γ)K : i = 1, ..., K : . |α| ≤ m : Dα] u]i (y, t) = Dα Ui (y, t) , (y, t) ∈ Ω \ Γ (11.47) That is, the solution u] is in fact a classical solution everywhere except for a closed nowhere dense set. Indeed, in this regard it is sufficient to show that the mapping F : Ω × RM → RK which defines the system of equations through (8.3) satisfies (9.30). This follows easily from the fact that the equation is linear in the terms Dtm ui . We now show that such a solution, that is, one that satisfies (11.47) may be constructed so as to also satisfy the initial condition (11.2). The idea is to apply the techniques from Chapter 10. In particular, we will construct a suitable generalized solution of the system of nonlinear PDEs (11.1) in the space N Lm+1 (Ω)K . Smooth approximations are then constructed using Theorem 74. Note, however, that this approach can, in its present form, deliver only the existence of solutions in MLm (Ω)K of the system of PDEs (11.1), solutions which may not satisfy the initial condition (11.2). Indeed, suppose that a generalized solution u] ∈ N Lm+1 (Ω)K of the system of equations (11.1) is constructed so as to also satisfy the initial condition (11.2) in the sense that ∀ ∀ ∀ ∀ 0≤p<m: q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : i = 1, ..., K : . y ∈ Rn−1 : qp] ] 1) Dyt,i u (y, t0 ) = Dq gp,i (y) qp] ] 2) Dyt,i u (y, t0 ) continuous at (y, t0 ) Such a solutions is constructed as the limit of a sequence (un ) in MLm+1 (Ω)K that satisfies ∀ ∀ ∀ ∀ 0≤p<m: q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : i = 1, ..., K : . y ∈ Rn−1 : qp 1) Dyt,i un (y, t0 ) = Dq gp,i (y) qp 2) Dyt,i un (y, t0 ) continuous at (y, t0 ) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 204 The next step is to approximate each function un by a sequence (un,r ) ⊂ C m+1 (Ω)K , in the sense that ∀ i = 1, ..., K : ∀ |α| ≤ m + 1 : ∀ A ⊂ Ω \ Γn compact : kDα un,i − Dα un,r,i kA → 0 where Γn ⊂ Ω is closed nowhere dense such that un ∈ C m+1 (Ω \ Γn )K . Using Proposition 49 and Theorem 86, one may extract a sequence (un,rn ) which converges to some function u ∈ MLm (Ω)K such that (Tun,rn ) converges to 0 in ML0 (Ω)K . In particular, the sequence (un,rn ) may be chosen in such a way that, for some closed nowhere dense set Γ ⊂ S ∀ ∀ ∀ ∀ 0≤p<m: q ∈ Nn−1 , 0 ≤ p + |q| ≤ m : i = 1, ..., K : . A ⊂ S compact : kDyq Dtp un,rn ,i − Dq gp,i kA → 0 However, the above construction does not imply that the solution u ∈ MLm (Ω)K of the system of PDEs (11.1) satisfies the initial condition (11.2). Indeed, the sequence (un,rn ) may be unbounded on every neighborhood of every point of S. In this regard, consider the following. Example 97 For each n ∈ N, consider the function un ∈ C 1 (R) given by 4 2 2 e(n t −1) if |t| < n1 un (t) = 0 if |t| ≥ n1 Clearly, un (0) = e for every n ∈ N. However, this sequence, and the sequence (u0n ) converge to 0 uniformly on every compact subset of R \ {0}, and the sequence of derivatives (u0n ) is unbounded on every neighborhood of 0. The difficulties mentioned above may be overcome by carefully constructing the original approximating sequence in MLm+1 (Ω). As such, the method used to construct the approximations in the proof of following result is slightly different from those used in the proofs of Theorems 76 and 96. Theorem 98 The nonlinear Cauchy problem Dtm u = G y, t, ..., Dyq Dtp ui (y, t) , ... Dtp u (y, t0 ) = gp (y) , with 0 ≤ p < m and q ∈ Nn−1 such that 0 ≤ p + |q| < m, admits a generalized solution u] ∈ N Lm (Ω) that also satisfies (11.47), provided that the mapping (11.3) is C 2 -smooth, and the initial data satisfies (11.5). CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 205 Proof. Write Rn−1 = [ Jν ν∈N where, for each ν ∈ N, Jν is a compact n − 1-dimensional interval [aν , bν ], with aν,i < bν,i for each i = 1, ..., n − 1. We also assume that the Jν are locally finite, and have pairwise disjoint interiors. Fix ν ∈ N and y0 ∈ Jν . Then it follows by Picard’s Theorem 1 and the compactness of Jν that there is some δν > 0 such that the system of ODEs has a solution v = vy0 F1 (y0 , t, ..., Dp viq (t) , ...) = 0 (11.48) |q|≤m+1 = vyq0 ,i i≤K ∈ C m+1 (t0 − δy0 , t0 + δν )L such that ∀ i = 1, ..., K : ∀ 0≤p<m: . ∀ q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : Dtp vyq0 ,i (t0 ) = Dyq gp,i (y0 ) (11.49) Here F1 : Ω × RL → RP is the continuous mapping such that Dβ (Dtm u + G (y, t, ..., Dtp Dq ui , ...)) |β|≤1 = F1 y, t, ..., Dyq Dtp ui , ... . Also note that since the mapping F1 , as well as the functions Dyp gp,i defining the initial data are C 1 -smooth, the solutions vy0 of (11.48) may be chosen in such a way that they depend continuously on y0 ∈ Jν , see for instance [69]. That is, ∀ ∃ ∀ ∀ ∀ ∀ >0: θ > 0 : i = 1, ..., K : 0≤p≤m+1 : q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : |t − t0 | < δν : ky0 − y1 k < θ ⇒ |Dtp vyq0 ,i (t) − Dtp vyq1 ,i (t) | < . (11.50) 2 Now define the functions Uy0 ,i ∈ C m+1 (Rn−1 × [t0 − δν , t0 + δν ]) through ! n−1 X Y qj q Uy0 ,i (y, t) = (yj − y0,j ) vy0 ,i |q|≤m+1 j=1 where q = (q1 , ..., qn−1 ). Then we have ∀ ∀ ∀ ∀ i = 1, ..., K : 0≤p≤m+1 : q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : |t − t0 | < δν Dtp Dyq Uy0 ,i (y0 , t) = Dp vyq0 ,i (t) (11.51) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 206 and ∀ i = 1, ..., K : ∀ 0≤p<m: ∀ q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : Dtp Dyq Uy0 ,i (y0 , t0 ) = Dyq gp,i (y0 ) (11.52) satisfies the system of ODEs As such, and in view of the fact that v = (viq )|q|≤m+1 i≤K (11.48), it follows that ∀ |β| ≤ 1 : ∀ |t − t0 | < δν : ∀ j = 1, ..., K : Dβ Tj (y, t, D) Uy0 (y0 , t) = 0 where Uy0 = (Uy0 ,i )i≤K and the Tj are the components of the partial differential operator T (y, t, D). As such, and in view of the continuity of the mapping F1 and the function Uy0 and its derivatives, it now follows that ∀ ∃ ∀ ∀ ∀ >0: δy0 > 0 : ky − y0 k < δy0 : . |t − t0 | < δν : |β| ≤ 1 : − < Dβ Tj (y, t, D) Uy0 (y, t) < (11.53) Furthermore, from (11.50) and (11.51) it follows that ∀ ∃ ∀ ∀ ∀ ∀ >0: δy0 > 0 : ky − y0 k < δy0 : 0≤p≤m+1 : . (11.54) q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : |t − t0 | < δν : qp qp 1) |D U (y, t) − D U (y, t) | < y y 0 1 yt,i yt,i ky0 − y1 k < δy0 ⇒ qp 2) |Dyt,i Uy1 (y, t) − Dtp vyq0 ,i (t) | < CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 207 Fix > 0. Since Jν is compact, it follows by (11.53) and (11.54) that ∃ ∀ ∃ ∀ ∀ ∀ ∀ ∀ ∀ δν , δ > 0 : y 0 ∈ Jν : Uy0 ∈ C m+1 (Jν × [t0 − δν , t0 + δν ])K : |β| ≤ 1 : 0≤p≤m+1 : q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : . i, j = 1, ..., K : ky − y0 k < δ : |t − t0 | < δν : 1) − < Dβ Tj (y, t, D) Uy0 (y, t) < 2) Dβ Tj (y, t, D) Uy0 (y, t) = 0 qp qp 3.1) |Dyt,i Uy0 (y, t) − Dyt,i Uy1 (y, t) | < 3) ky0 − y1 k < δ ⇒ qp 3.2) |Dyt,i Uy1 (y, t) − Dtp vyq0 ,i (t) | < Furthermore, (11.52) implies that ∀ ∀ ∀ ∀ y0 ∈ Rn−1 : i = 1, ..., K : 0≤p<m: . n q ∈ N , 0 ≤ p + |q| ≤ m + 1 : qp Dyt,i Uy0 (y0 , t0 ) = Dq gp,i (y0 ) Subdivide Jν into n − 1-dimensional, compact intervals Iν,1 , ..., Iν,γν with nonempty interiors, and diagonal not exceeding δ . In particular, the Iν,k must be locally finite with pairwise disjoint interiors. Let yν,k denote the midpoint of Iν,k . Then ∀ ∃ ∀ ∀ ∀ ∀ ∀ k = 1, ..., γν : Uν,k ∈ C m+1 (Iν,k × [t0 − δν , t0 + δν ])K : (y, t) ∈ Bν,k : 0≤p≤m+1 : q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : |β| ≤ 1 : , (11.55) i, j = 1, ..., K : 1) − < Dβ (Tj (y, t, D) Uν,k (y, t)) < 2) Dβ (Tj (y, t, D) Uν,k (yν,k , t)) = 0 qp p q 3) Dyt,i Uν,k (yν,k i (t) , |t − t0 | < δν , t) = Dt vqp qp 4.1) |Dyt,i Uy0 (y, t) − Dyt,i Uν,k (y, t) | < 4) y0 ∈ Iν,k ⇒ qp 4.2) |Dyt,i Uy0 (y, t) − Dtp vyqν,k (y, t) | < where Bν,k is the set Bν,k = Iν,k × [t0 − δν , t0 + δν ], CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 208 and ∀ i = 1, ..., K : ∀ 0≤p<m: ∀ q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : qp Dyt,i Uν,k (yν,k , t0 ) = Dq gp,i (yν,k ) (11.56) Consider the set Ω1 = [ (Jν × [t0 − δν , t0 + δν ]) ν∈N and the function V1 = γν X X ν∈N ! χν,k Uν,k k=1 where each χν,k denotes the characteristic function of the interior of Bν,k . The function V1 is clearly C m+1 -smooth everywhere on Ω1 except for a closed nowhere dense subset Γ1 of Ω1 which satisfies Γ1 ∩ S closed nowhere dense in S. It follows from (11.55) that the function V1 satisfies ∀ ∀ ∀ ∀ ∀ ∀ ∀ ν∈N: k = 1, ..., γν : |β| ≤ 1 : (y, t) ∈ intBν,k : i, j = 1, ..., K : 0≤p≤m+1 : , (11.57) q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : 1) − < Dβ (Tj (y, t, D) V1 (y, t)) < 2) Dβ (Tj (y, t, D) V1 (yν,k , t)) = 0 qp V1 (yν,k , t) = Dtp vyqν,k ,i (t) 3) Dyt,i qp qp 4.1) |Dyt,i V1 (y, t) − Dyt,i Uy0 (y, t) | < 4) y0 ∈ IntIν,k ⇒ qp 4.2) |Dyt,i V1 (y, t) − Dtp vyqν,k (y, t) | < and from (11.56) we obtain ∀ ∀ ∀ ∀ ∀ ν∈N: k = 1, ..., γν : i = 1, ..., K : . 0≤p<m: q ∈ Nn , 0 ≤ p + |q| ≤ m + 1 : qp Dyt,i V1 (yν,k , t0 ) = Dq gp,i (yν,k ) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 209 From (11.57) it follows that the function V1 satisfies ∀ ∀ ∀ ∀ ∀ ν∈N: k = 1, ..., γν : i = 1, ..., K : 0≤p≤m+1 : q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : qp p,q (y, t) ∈ Ω1 \ Γ1 ⇒ λp,q 1,i (y, t) ≤ Dyt,i V1 (y, t) ≤ µ1,i (y, t) (11.58) p,q 0 where λp,q 1,i , µ1,i ∈ C (Ω1 \ Γ1 ) are the functions p q λp,q 1,i (y, t) = Dt vyν,k ,i − 2, (y, t) ∈ IntBν,k (11.59) p q µp,q 1,i (y, t) = Dt vyν,k ,i + 2, (y, t) ∈ IntBν,k . (11.60) and Continuing in this way, we may construct a countable and dense subset A = {yk : k ∈ N} of Rn−1 , a sequence (Γn ) of closed nowhere dense subsets of Ω1 that satisfies Γn ∩ S closed nowhere dense in S and (yk , t0 ) ∈ / Γn , and functions Vn ∈ C m+1 (Ω1 \ Γn )K so that ∀ |β| ≤ 1 : ∀ (y, t) ∈ Ω1 \ Γn : ∀ j = 1, ..., K : − n < Dβ (Tj (y, t, D) Vn (y, t)) < . (11.61) n Furthermore, the sequence (Vn ) also satisfies ∀ k∈N: ∃ Nk ∈ N : ∀ i = 1, ..., K : (11.62) qp q 1) n ≥ Nk ⇒ Dyt,i Vn (yk , t0 ) = Dy gp,i (yk ) , 0 ≤ p < m, 0 ≤ p + |q| ≤ m + 1 qp 2) n ≥ Nk ⇒ Dyt,i Vn (yk , t) = Dtp vyqk ,i , 0 ≤ p ≤ m + 1, 0 ≤ p + |q| ≤ m + 1 and ∀ ∀ ∀ ∀ ∀ ν∈N: k = 1, ..., γν : i = 1, ..., K : 0≤p≤m+1 : q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : qp p,q (y, t) ∈ Ω1 \ Γ1 ⇒ λp,q n,i (y, t) ≤ Dyt,i Vn (y, t) ≤ µn,i (y, t) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 210 p,q 0 where λp,q n,i , µn,i ∈ C (Ω1 \ Γn ) are functions that satisfy p,q 0 < λp,q n,i (y, t) − µn,i (y, t) < 4 n (11.63) and p,q p,q p,q λp,q n+1,i (y, t) < λn,i (y, t) < µn+1,i (y, t) < µn+1,i (y, t) (11.64) for each n ∈ N. Let (Un ) denote the sequence of approximating solutions to the system of PDEs (11.1) constructed Theorem 71. That is, for each n ∈ N, we have Un ∈ C m+1 (Ω \ Γn )K , for some closed nowhere dense set Γn ⊂ Ω. Consider the functions Wn = χ1 Un + Vn where χ1 is the characteristic function of Ω \ Omega1 . Clearly, for each n ∈ N, we have Wn ∈ C m+1 (Ω \ Γ0n )K for some closed nowhere dense set Γ0n ⊆ Ω. In particular, Γ0n ∩ S closed nowhere dense in S and yk ∈ / Γ0n . Furthermore, it follows from (11.61) and the corresponding property of the functions Un , that the sequence (Wn ) satisfies ∀ |β| ≤ 1 : ∀ j = 1, ..., K : ∀ (y, t) ∈ Ω \ Γ0n − n1 < Dβ Tj (y, t, D) Wn (y, t) < (11.65) 1 n and (11.62) implies ∀ ∃ ∀ ∀ ∀ k∈N: Nk ∈ N : i = 1, ..., K : . 0≤p<m: q ∈ Nn−1 , 0 ≤ p + |q| ≤ m + 1 : qp n ≥ Nk ⇒ Dyt,i Wn (yk , t0 ) = Dyq gp,i (yk ) (11.66) Moreover, for 0 ≤ p < m + 1 and q ∈ Nn−1 such that 0 ≤ p + |q| ≤ m + 1 in (11.66) we have qp n ≥ Nk ⇒ Dyt,i Wn (yk , t) = Dp vyqk ,i (t) , (yk , t) ∈ S1 \ Γ0n (11.67) As such, it follows from (11.65), (11.66) and (11.67) that the sequence (un ) in MLm+1 (Ω)K , the components of which are defined as un,i = (I ◦ S) (Wn,i ) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 211 satisfies ∀ |β| ≤ 1 : − n1 ≤ Dβ Tj u < ∀ ∃ ∀ ∀ 1 n , (11.68) k∈N: Nk ∈ N : n ≥ Nk : (11.69) (yk , t) ∈ S1 \ Γ0n : 1) Dtp Dyq un,i (yk , t0 ) = Dyq gp,i (yk ) , 0 ≤ p < m, 0 ≤ p + |q| ≤ m + 1 2) Dtp Dyq un,i (yk , t) = Dtp vyqk ,i , 0 ≤ p ≤ m + 1, 0 ≤ p + |q| ≤ m + 1 and ∀ ∀ ∀ ∀ n∈N: i = 1, ..., K : 0≤p≤m+1 : . n−1 q ∈ N , 0 ≤ p + |q| ≤ m + 1 : q,p qp λn,i ≤ Dyt,i un ≤ µq,p n,i (11.70) where q,p λn,i = (I ◦ S) λq,p n,i and q,p µq,p n,i = (I ◦ S) µn,i . In particular, the sequence (un ) in MLm+1 (Ω)K is a Cauchy sequence, while the sequence (Tun ) converges to 0 in ML1 (Ω)K . It now follows by exactly the same arguments used in the proof of Theorem 87 that there is a sequence (vn ) in C m+1 (Ω)K , and a function u ∈ MLm (Ω)K such that (Tvn ) converges to 0 in ML0 (Ω)K , and (vn ) converges to u in MLm (Ω)K . In particular, there is a closed nowhere dense set Γ ⊂ Ω such that u ∈ C m (Ω \ Γ)K and ∀ A ⊂ Ω \ Ω compact : ∀ |α| ≤ m : . ∀ i = 1, ..., K : kDα vn,i − Dui kA → 0 (11.71) It now follows by Theorem 65 that Tu = 0. We claim Γ ∩ S closed nowhere dense in S. (11.72) CHAPTER 11. A CAUCHY-KOVALEVSKAIA TYPE THEOREM 212 In this regard, fix ν ∈ N and consider, for each i =, ..., K, every 0 ≤ p ≤ m and q ∈ Nn−1 such that 0 ≤ p + |q| ≤ m the function pq q wi,ν : intJν × [t0 − δν , t0 + δν ] 3 (y, t) 7→ Dtp vy,i (t) pq It follows from (11.50) that the function wi,ν is continuous at every point (y, t) ∈ Jν × [t0 − δν , t0 + δν ]. Furthermore, in view of (11.69), it is clear that the sequence (vn ) may be constructed in such a way that ∀ ∀ ∃ ∃ ∀ ∀ ∀ ∀ ν∈N y ∈ A ∩ intJν : δν > 0 : Ny ∈ N : n ≥ Ny : . 0≤p≤m: q ∈ Nn−1 , 0 ≤ p + |q| ≤ m : i = 1, ..., K : qp q Dyt,i vn (y, t) = Dtp vy,i (t) , |t − t0 | < δy As such, it follows that the solution u satisfies ∀ ∃ ∀ ∀ ∀ ∀ ν∈N δν > 0 : 0≤p≤m: y ∈ A ∩ intJν : . i = 1, ..., K : q ∈ Nn−1 , 0 ≤ p + |q| ≤ m : qp q Dyt,i u (y, t) = Dtp vy,i (t) , |t − t0 | < δy (11.73) Since A∩intJν is dense in intJν , our claim (11.72) follows from Proposition 46. That u satisfies the initial condition on S \ Γ follows by (11.73) and (11.49) It should also be mentioned that many of the interesting systems of PDEs that arise in applications may be written in the form (11.1). In particular, the equations of fluid mechanics typically take the form Dt u (y, t) + G y, t, ..., Dyq ui (y, t) , ... = 0 where G : Ω × RM → RK is often a C ∞ -smooth mapping. These include, amongst others, the Navier-Stokes equations which have attracted a lot of attention in recent years, see for instance [36], [41] and [98]. Although such equations may not be expressed in exactly the form (11.1), and while the additional conditions, such as boundary and / or initial conditions, may in general not take the form (11.2), the techniques presented here may be applied in these cases as well. Chapter 12 Concluding Remarks 12.1 Main Results We have constructed a general and type independent theory for the solutions of a large class of systems of nonlinear PDEs. The spaces of generalized functions upon which the theory is based are constructed as the Wyler completions of suitable uniform convergence spaces. A significant advantage of this method, when compared with the typical spaces of generalized functions used in the customary functional analytic methods is that the generalized functions introduced here may be represented with usual nearly finite normal lower semi-continuous functions. This provides a first basic, and so far unprecedented, blanket regularity for the generalized solutions of systems of nonlinear PDEs that are constructed. It should be noted that, in the basic construction of spaces of generalized functions, and the fundamental existence results for the solutions of systems of nonlinear PDEs, those functional analytic techniques that are typical in the study of PDEs, do not appear. However, this is not to say that functional analysis, or, for that matter, any mathematics, may not, and should not, be used in the study of nonlinear PDEs. Rather, the meaning of this is that such sophisticated mathematical tools should perhaps not form the basis for the study of the existence of solutions of nonlinear PDEs, but rather of their additional regularity properties, beyond the mentioned blanket regularity which anyhow results from the theory presented here. Indeed, perhaps the most dramatic results presented in this work, namely, the regularity results obtained for the solutions of a large class of systems of nonlinear PDEs in Chapter 10, and the Cauchy-Kovalevskaia type Theorem proved in Chapter 11, certainly make use of advanced tools from functional analysis. Namely, it is based on sufficient conditions for precompactness of sets in suitable Frechét spaces. However, these results arise as an application of the general existence and regularity theory presented in Chapters 7, 8 and 9, which is based on far simpler techniques. Let us now summarize the main results of this work. In Chapter 6 we present some auxiliary results on the completion of uniform convergence spaces. These results are used extensively in the text, in particular in regard to the interpretation 213 CHAPTER 12. CONCLUDING REMARKS 214 of generalized functions as nearly finite normal lower semi-continuous functions. Chapter 7 introduces suitable spaces of nearly finite normal lower semi-continuous functions. These are the fundamental spaces upon which the spaces of generalized functions studied here are constructed. It should be noted that the results obtained in Chapter 7, and especially those connected with the construction of the uniform order convergence structure and its completion, are of interest in their own right. Indeed, the uniform convergence structure Jo on ML (X) does not depend on the uniform structure on R, or the algebraic structure of ML (X). This might suggest more general result on constructing the Dedekind order completion of a partially ordered set as the completion of a suitable uniform convergence structure. Chapter 8 concerns the construction of spaces of generalized functions, and the action of nonlinear partial differential operators on the mentioned spaces of generalized functions. As mentioned, the generalized functions, which are the elements of these spaces, may be represented as usual nearly finite normal lower semi-continuous functions. This may be interpreted as a blanket regularity for these generalized functions. The development of pullback type spaces of generalized functions introduced in Section 8.1 comes down to a reformulation, in terms of uniform convergence spaces, of the construction of spaces of generalized functions in the Order Completion Method [119]. Such a recasting of the Order Completion Method in terms of uniform convergence spaces allows for the application of convergence theoretic techniques to problems related to the structure and regularity of generalized solutions. As is shown in this work, such tools turn out to be highly effective in this regard. The mentioned spaces are associated with a given nonlinear partial differential operator. In particular, one cannot, in general, define generalized derivatives of the elements in these spaces. The Sobelev type spaces of generalized functions are introduced in Section 8.2 in order to address these issues. In particular, the spaces are defined without reference to any particular nonlinear partial differential operator, which, to a certain extent, makes them universal. Furthermore, the generalized functions in these spaces may be uniquely represented through their generalized partial derivatives as nearly finite normal lower semi-continuous functions. The issue of existence of generalized solutions of systems of nonlinear PDEs in the spaces constructed in Chapter 8 is addressed in Chapter 9. Section 9.1 contains the approximation results upon which the theory is based. These include a multidimensional version of (1.110), as well as relevant refinements of this result. In Section 9.2, the basic existence and regularity result obtained in the Order Completion Method [119] is recast in the setting of the so called pullback uniform convergence spaces of generalized functions, while Section 9.3 deals with additional regularity properties of these solutions. In particular, it is shown that such solutions in the pullback spaces of generalized solutions may be assimilated with functions that are C k -smooth, for k ∈ N ∪ {∞}, everywhere except on a closed nowhere dense set, provided that the nonlinear operator is C k -smooth. It should be noted that such regularity results have so far not been obtained within the setting of the partially ordered sets within which the Order Completion Method [119] is formulated. In- CHAPTER 12. CONCLUDING REMARKS 215 deed, our result relies on the existence of a compatible complete, Hausdorff uniform convergence structure on a given Hausdorff convergence space [26]. Section 9.4 deals with the question of existence of generalized solutions in the Sobolev type spaces of generalized functions. It is shown that a large class of systems of nonlinear PDEs admit generalized solutions in this sense. This also provides additional insight into the structure of generalized solutions in the pullback type spaces of generalized functions. Indeed, each unique generalized solution in the pullback type spaces may be identified with the set of all solutions in the Sobolev type spaces of generalized functions. We also consider the effect of additional smoothness conditions on the nonlinear partial differential operator and the righthand term f of the system of equations on the regularity of generalized solutions. In this regard, it is shown that, under suitable conditions on the operator T, an equation of the form (8.1) admits a generalized solution in the Sobolev type space of order m + k, provided that the nonlinear operator T, and the righthand term f are C k -smooth. As mentioned, the generalized solutions constructed in the Sobolev type spaces of generalized functions may be uniquely represented through their generalized partial derivatives, which are nearly finite normal lower semi-continuous functions. As such, there is a set R ⊆ Ω with complement a set of first Baire category such that each generalized partial derivative is continuous and real valued at each x ∈ R. However, even in case the set R has nonempty interior, the generalized derivatives cannot, in general, be interpreted as usual partial derivatives at any point of R. In Chapter 10 it is shown that a large class of systems of nonlinear PDEs admit generalized solutions, in suitable Sobolev type spaces of generalized functions, which are in fact classical solutions everywhere except possibly on a closed nowhere dense set. This result is based on a useful sufficient condition for the precompactness of subsets of a suitable Frechét space of sufficiently smooth functions. In view of the various nonexistence results for certain partial differential equations, see for instance [97], this result is counter intuitive. Indeed, these results show that, contrary to common belief, most systems of nonlinear PDEs admit generalized solutions which are in fact classical solutions everywhere except on a closed nowhere dense subset of the domain of definition of the system. That is, the existence of a classical solution to such a system of nonlinear PDEs is a strongly generic property of that system [129]. The solution methods for systems of nonlinear PDEs developed in Chapters 8 to 10 do not take into account any possible additional conditions, such as initial and / or boundary conditions. However, and as is shown in Chapter 11, the theory developed in Chapters 8 to 10 may be applied to problems including such additional conditions with only minimal modifications. This results in the first extension of the Cauchy-Kovalevskaia Theorem 2 to systems of equations that may not be analytic, on its own, general and type independent grounds. In particular, it is shown that any initial value problem of the form (11.1) to (11.2) admits a generalized solution in a suitably constructed Sobolev type space of generalized functions. Furthermore, if the system of equations, and the initial data satisfy suitable smoothness conditions, such a solution can be constructed so that it is a classical solution everywhere except on CHAPTER 12. CONCLUDING REMARKS 216 a closed nowhere dense set. Furthermore, this solution satisfies the initial condition in the classical sense. It should be noted that these methods may be applied to many of the equations that arise in applications. In particular, the equations of fluid mechanics, including the Navier-Stokes equations, may be treated by similar techniques. 12.2 Topics for Further Research In this work we have initiated a general and type independent theory for the existence and regularity of generalized solutions for a large class of systems of nonlinear PDEs. The results obtained in this regard apply also to many of those equations that that have been proven to be unsolvable in the usual linear topological spaces of generalized functions, and are therefor generally believed to be unsolvable, such as the Lewy equation (1.32), see for instance [88] and [97]. As such, the issues of solvability of such systems of linear and nonlinear PDEs must be carefully reconsidered. Systems of linear and nonlinear PDEs appear frequently in the applications of mathematics to physics, chemistry, engineering and, recently, even biology. For such applications, knowledge of the qualitative properties of the solutions of such systems, and effective numerical computation of the solutions are required. The development of analytic and numerical tools for this purpose is an important issue. The spaces of generalized functions that we have constructed are not contained in any of the standard linear functional analytic spaces of generalized functions that are typical in the literature. In fact, even if some generalized function may be represented in, say, one of the Sobolev type spaces of generalized functions, and in one of the standard spaces, such as the D0 distributions, they may exhibit rather different properties. Indeed, the Heaviside function 1 if x ≤ 0 u (x) = (12.1) 0 if x > 0 belongs to both N L1 (R), and to D0 (R). However, in N L1 (R) its derivative is u0 (x) = 0, while in D0 (R) its derivative is u0 = δ, the Dirac distribution, which is not the 0 function. The exact clarification of the interrelations between the new spaces of generalized functions introduced here, and those of the classical theory of PDEs, is another interesting and important open problem. Bibliography [1] R A Adams : Sobolev spaces, Pure and Applied Mathematics, 65, Academic Press, 1975. [2] Alefeld G and Herzberger J : Introduction to interval computations, Academic Press, 1983. [3] Anguelov R : Dedekind order completion of C(X) by Hausdorff continuous functions, Quaestiones Mathematicae 27 (2004) 153 – 170. [4] Anguelov R : The rational extension of C (X) via Hausdorff continuous functions, Thai J Math 5 no 2 (2007) 267 – 272. [5] Anguelov R and Markov S : Extended segment analysis, Freiburger Interval Berichte 10 (1981) 1 – 63. [6] Anguelov R, Markov S and Sendov B : The set of Hausdorff continuous functions — the largest linear space of interval functions, Reliable Computing 12 (2006) 337 – 363. [7] Anguelov R, Markov S and Sendov B : Algebraic operations on the space of Hausdorff continuous interval functions, Proceedings of the International Conference on Constructive Theory of Functios, 1-7 June 2005, Varna, Bulgaria, Marin Drinov Acad. Publ. House, Sofia, 2006, 35 – 44. [8] Anguelov R and Rosinger E E : Hausdorff continuous solutions of nonlinear PDEs through the order completion method, Quaestiones Mathematicae 28 no 3 (2005) 271 – 285. [9] Anguelov R and Rosinger E E : Solving large classes of nonlinear systems of PDE’s, Computers and Mathematics with Applications 53 (2007) 491 – 507. [10] Anguelov R and van der Walt J H : Order convergence structure on C (X), Quaestiones Mathematicae 28 no. 4 (2005) 425 – 457. [11] Arhangel’skii A V : Some types of factor mappings, and the relations between classes of topological spcase, Dokl. Akad. Nauk SSSR 153 (1963) 743 – 746. 217 BIBLIOGRAPHY 218 [12] Arnold V I : Lectures on PDEs, Springer Universitext, 2004. [13] Baire R : Lecons sur les fonctions discontinues, Collection Borel, Paris, 1905. [14] Ball J : Convexity conditions and existence theorems in nonlinear elasticity, Arch Rat Mech Anal 63 (1977) 337 – 403. [15] Ball R N : Convergence and Cauchy structures of lattice ordered groups, Trans AMS 259 (1980) 357 – 392. [16] Ball R N and Hager A W : Epi-topology and Epi-convergence for Archimedean lattice-ordered groups with unit, Appl Categor Struct 15 (2007) 81 – 107. [17] Banach S : Théorie des opérations linéaires, Warsaw, 1932. [18] Bardi M and Capuzzo-Dolcetta I : Optimal control and viscosity solutions of Hamilton-Jacobi-Bellman equations, Birkhäuser, 1997. [19] Bartle R G : The elements or real analysis, John Wiley & Sons, 1964. [20] Bastiani A : Applications différentiables et varietés différentiables de dimension infinite, J Anal Math 23 (1964) 1 – 114. [21] Beattie R : Convergence spaces with webs, Math Nachr 116 (1984) 159 – 164. [22] Beattie R : A convenient category for the closed graph theorem, Categorical topology (Toledo, Ohio, 1983), 29 – 45, Helderman-Verlag, Berlin, 1984. [23] Beattie R : Continuous convergence and functional analysis, Proceedings of the International Convergence on Convergence Theory (Dijon, 1994), Topology Appl 70 no 2-3 (1996) 101 – 111. [24] Beattie R and Butzmann H-P : On the Banach-Steinhaus theorem and the continuity of bilinear mappings, Math Nachr 153 (1991) 297 – 312. [25] Beattie R and Butzmann H-P : Ultracomplete convergence spaces and the closed graph theorem, Topology Appl 111 (2001) 59 – 69. [26] Beattie R and Butzmann H-P : Convergence structures and applications to functional analysis, Kluwer Academic Plublishers, Dordrecht, Boston, London, 2002. [27] Bierstedt K-D : An introduction to locally convex inductive limits, Functional analysis and its applications, Lect Int Sch, Nice, France, 1986 (1988), 35 – 133. [28] Binz E : Continuous convergence on C (X), Lecture Notes in Mathematics 469, Springer-Verlag, Berlin, Heidelberg, New York, 1975. BIBLIOGRAPHY 219 [29] Birkhoff G : Lattice theory, AMS, Providence, Rhode Island, 1973. [30] Bourbaki N : General, Chapters 1 – 4, Springer-Verlag, Berline, Heidelberg, New York, 1998. [31] Brezis H, Crandall M G and Pazy A : Perturbations of nonlinear maximal monotone sets in Banach space, Comm Pure Appl Math 23 (1970) 123 – 144. [32] Brümmer G C L and Hager A W : Functorial uniformization of topological spaces, Topology Appl 27 (1987) 113 – 127. [33] Bucher W and Frölicher A : Calculus in vector spaces without norm, Lecture Notes in Mathematics 30, Springer-Verlag, Berlin, Heidelberg, New York, 1966. [34] Butzmann H-P : Dualitäten in Cc (X), PhD Thesis, Mannheim, 1971. [35] Butzmann H-P : An incomplete function space, Applied Categorical Structures 9 no 4 (2000) 595 – 606. [36] Caffarelli L, Kohn R and Nirenberg L : Partial regularity of suitable weak solutions of the Navier-Stokes equations, Comm Pure & Appl Math 35 (1982) 771 – 831. [37] Cartan G : Théorie des filters, C R Acad Sci 205 (1973) 595 – 598. [38] Choquet G : Convergences, Ann Univ Greoble- Sect Sci Math Phys II Ser 23 (1948) 57 - 112. [39] Colombeau J F : New generalized functions and multiplication of distributions, Noth Holland Mathematics Studies 84, 1984. [40] Colombeau J F : Elementary introduction to new generalized functions, North Holland Mathematics Studies 113, 1985. [41] Constantin P : Some open problems and research directions in the mathematical study of fluid dynamics, in ‘Mathematics Unlimited- 2001 and Beyond’, Springer Verlag, Berlin, 2001, 353 – 360. [42] Crandal M G, Ishii H and Lions P L : User’s guide to viscosity solutions of second order partial differential equations, Bull AMS 27 no 1 (1992) 1 – 67. [43] Cullender S F and Labuschagne C C A : Unconditinal Schauder decompositions and stopping times in the Lebesgue-Bochner spaces, To Appear in J Math Anal Appl. [44] Dacorogna B : Continuity and weak lower semicontinuity of nonlinear functionals, Lecture Notes in Mathematics 922, Springer, New York, 1982. BIBLIOGRAPHY 220 [45] De Pagter B and Grobler J J : Operators representable as multiplicationconditional expectation operators, J Operatro Theory 48 (2002) 15 – 40. [46] Del Prete I and Lignola M B : Uniform convergence structures in function spaces and preservation of continuity, Ricerche di Mthematica 30 no 1 (1987) 45 – 58. [47] Dilworth R P : The normal completion of the lattice of continuous functions, Trans AMS 68 (1950) 427 – 438. [48] Di Perna R J : Compensated compactness and general systems of conservation laws, Trans AMS 29 no 2 (1985) 383 – 420. [49] Dunford N and Schwartz T S : Linear operators part I: General theory, Pure and Applied Mathematics VII, Interscience Publishers, New York, London, Sydney, 1957. [50] Dobertin H, Erné M and Kent D C : A note on order convergence in complete lattices, Rocky Mountain J Math 14 no 2 (1984) 647 – 654. [51] Dolecki S and Mynard F : Convergence-theoretic mechanisms behind product theorems. Proceedings of the French-Japanese Conference “Hyperspace Topologies and Applications” (La Bussire, 1997). Topology Appl. 104 (2000) 67 – 99. [52] Ehrenpreis L : Solutions of some problems of division I, Amer J Math 76 (1954) 883 – 903. [53] Eilenberg S : Ordered topological spaces, Amer J Math 63 (1941) 39 – 45. [54] Evans L C : Partial differential equations, AMS Graduate Studies in Mathematics 19, AMS, 1998. [55] Feldman W A and Porter J F : A Marinescu structure for vector lattices, Proc Ams 41 no. 2 (1973) 602 – 608. [56] Fine N J, Gillman L and Lambek J : Rings of quotients of rings of functions, McGill University Press, Montreal, 1965. [57] Fischer H R : Limesräume, Math Ann 137 (1959) 269 – 303. [58] Forster O : Analysis 3, Intergralrechnung im Rn mit Anwendungen, Friedr Vieweg, Braunschweig, Wiesbaden, 1981. [59] Frechét M : Sur quelques points du calcul fonctionnel, Rend Circ Mat Palermo 22 (1906) 1 – 74. [60] Freudenthal H : Teilweise geordnete moduln, Proc Acad of Sc Amsterdam 39 (1936) 641 – 651. BIBLIOGRAPHY 221 [61] Fric R and Kent D C : Completion of pseudo-topological groups, Math Nachr 99 (1980) 99 – 103. [62] Friedlander F G : Introduction to the theory of distributions, Cambridge University Press, Cambridge, 1982. [63] Gähler W : Grundstrukturen der analysis I, Birkhäuser Verlag, Basel, 1977. [64] Gähler W : Grundstrukturen der analysis II, Birkhäuser Verlag, Basel, 1978. [65] Gähler S, Gähler W and Kneis G : Completion of pseudo-topological vector spaces, Math Nachr bf 75 (1976) 185 – 206. [66] Gierz G, Hofmann K H, Keimel K, Lawson J D, Mislove M and Scott D S : Continuous lattices and domains, Encyclopedia of mathematics and its applications 23, Cambridge University Press, 2003. [67] Gingras A R : Order convergence and order ideals, Proceedings of the converence on convergence spaces (Univ. Nevada, Reno, Nev, 1976), Dept Math, Univ Nevada, Reno, Nev, (1976) 45 – 59. [68] Hahn H : Reelle funktionen I, Wien, 1932. [69] Hartman P : Ordinary differential equations, John Wiley & Sons, 1964. [70] Hausdorff, F : Grundüge der mengenlehre, Veit, Leibzig 1914. [71] Hörmander L : Linear partial differential operators, Springer, New York, 1976. [72] Kantorovitch L V : Sur un espace des fonctions à variation bornée et la différentiation d’une séries terme à terme, Comptes Rendus de l’Acad. Sc Paris 201 (1935) 1457 – 1460. [73] Kantorovitch L V : Sur les propriétés des espaces semi-ordonnés linéaires, Comptes Rendus de l’Acad. Sc Paris 202 (1936) 813 – 816. [74] Kantorovitch L V : Linear partially ordered spaces, Mat Sbornik (N. S.) 44 (1937) 121 – 168. [75] Kashiwara M, Kawai T and Sato M : Hyperfunctions and pseudodifferential equations, Springer Lecture Notes in Mathematics 287, 1973. [76] Kaětov M : On real-valued functions in topological spaces, Fund Math 38 (1951) 85 - 91. [77] Kechris A S : Classical descriptive set theory, Graduate Texts in Mathematics 156, Springer-Verlag, New York, Berlin, Heidelberg, London, Paris, Tokyo, Hong Kong, Barcelona, Budapest, 1995. BIBLIOGRAPHY 222 [78] Keller H H : Differensierbarkeit in topologischen vektoräumen, Com Math Helv 38 (1964) 308 – 320. [79] Keller H H : Räume stetiger multilinearer Abbildungen als Limesräumen, Math Ann 159 (1965) 259-270. [80] Keller H H : Die limes-uniformisierbarkeit der limesräume, Math Ann 176 (1968) 334 – 341. [81] Kelley J : General topology, Van Nostrand, Princeton, 1955. [82] Kent D C : On the order topology in a lattice, Illinois J Math 10 (1966) 90 – 96. [83] Kent D C : The interval topology and order convergence as dual convergence structures, A Math Mon 74 (1967) 426 – 426. [84] Kent D C : Convergence quotient maps, Fund Math LXV (1969) 197 – 205. [85] Kolmogorov A N and Fomin S V : Functional analysis Vol. 2, Graylock, Albany, 1961. [86] Kovalevskaia S : Zur Theorie der partiellen differentialgleichung, Journal für die reine und angewandte Mathematik 80 (1875) 1 – 32. [87] Kraemer W and Wolff von Gudenberg J (Eds): Scientific computing, validated numerics, interval methods, Kluwer Academic, New York, Boston, Dordrecht, London, Moskow, 2001. [88] Kranz S G : Function theory of several complex variables, J Wiley, New York, 1982. [89] Kriegl A and Michor P W : The convenient setting of global analysis, Mathematical Serveys Monographs, Vol. 53, AMS, Providence, 1997. [90] Kunzi H P A : Nonsymmetric distances and their associated topologies: About the origins of basic ideas in the area of asymmetric topology, Handbook of the history of general topology, Vol 3, 853–968, Kluwer Acad. Publ., Dordrecht, 2001 [91] Kuo W -C, Labuschagne C C A and Watson B A : Conditional expectations on Riesz spaces, J Math Anal Appl 303 (2005) 509 – 521. [92] Kuo W -C, Labuschagne C C A and Watson B A : An upcrossing theorem for martingales on Riesz spaces, To Appear. [93] Kuratowski K : Topologie I, Espaces Mtrisables, Espaces Complets Monografie Matematyczne series, vol. 20, Polish Mathematical Society, WarszawaLww, 1948. BIBLIOGRAPHY 223 [94] Kuratowski K : Topologie II, Espaces Connexes, Plan Euclidien Monografie Matematyczne series, vol. 21, Polish Mathematical Society, Warszawa-Lww, 1950. [95] Labuschange C C A and Watson B A : Discreet time stochastic processes on Riesz spaces, To Appear in Indag Math. [96] Lax P D : The formation and decay of shock waves, Am Math Month (1972) 227 – 241. [97] Lewy H : An example of a smooth partial differential equation without solution, Ann Math 66 no 1 (1957) 155 – 158. [98] Lin F -H : A new proof of the Caffarelli-Kohn-Nirenberg theorem, Comm Pure & Appl Math 51 (1998), 241 – 257. [99] Lions J L : Une remarque sur les problemes D’evolution nonlineaires dans les domaines noncylindrique, Rev Romaine Math Pure Appl 9 (1964) 11 – 18. [100] Lions J L : Quelques methods de resolution des problemes aux limites nonlineaires, Dunod, Paris, 1969. [101] Luxemburg W A J and Zaanen A C : Riesz spaces I, North Holland, Amsterdam, 1971. [102] MacNeille H M : Partially ordered sets, Trans AMS 42 (1937) 416 – 460. [103] Malgrange B : Existence et approximation des equations aux derivees partielles et des equations de convolutions, Ann Inst Fourier 6 (1955-56) 271 – 355. [104] Markov S : Extended interval arithmetic involving infinite intervals, Math Balkanica 6 (1992) 269 – 304. [105] May R and McArthur C W : Comparison of two types of order convergence with topological convergence in an ordered topological vector space, Proc AMS 63 no 1 (1977) 49 – 55. [106] Medvedev F A Scenes from the history of real functions, Birkäuser, 1991. [107] Moore E and Smith H : A general theory of limits, Am J Math 44 (1922) 102 – 121. [108] Moser J : A new technique for the construction of solutions of nonlinear differential equations, Proc N A S 47 (1961) 1824 – 1831. [109] Murat F : Compacite par compensation: Conditions necessaire et sufficante de continuite faible sons une hypotheses de rang constant, Ann Scuola Norm Sup 8 (1981) 69 – 102. BIBLIOGRAPHY 224 [110] Munkres J R : Topoloy, 2nd Edition, Prentice Hall, 2000. [111] Mynard F : Coreflectively modified continuous duality applied to classical product theorems, Appl Gen Topol 2 (2001) 119 – 154. [112] Mynard F : More on strongly sequential spaces, Comment Math Univ Carolin 43 (2002) 525 – 530. [113] Mynard F : Productively Fréchet spaces, Czechoslovak Math J 54 (2004) 981 – 990. [114] Nash J : The imbedding problem for Riemannian manifolds, Ann Math 63 (1956) 20 – 63. [115] Neuberger J W : Sobolev gradients and differential equations, Springer Lecture Notes in Mathematics 1670, 1997. [116] Neuberger J W : Continuous Newton’s method for polynomials, Math. Inteligencer 21 (1999) 18 – 23. [117] Neuberger J W : A near minimal hypothesis Nash-Moser theorem, Int J Pure Appl Math 4 (2003) 269 – 280. [118] Neuberger J W : Prospects of a central theory of partial differential equations, Math Inteligencer 27 no 3 (2005) 47 – 55. [119] Oberguggenberger M B and Rosinger E E : Solution of continuous nonlinear PDEs through order completion, North-Holland, Amsterdam, London, New York, Tokyo, 1994. [120] Ordman E T : Convergence almost everywhere is not topological, Am Math Mon 73 (1966) 182 – 183. [121] Oxtoby J C : Meaure and category 2nd Edition, Springer-Verlag, 1980. [122] Papangelou F : Order convergence and topological completion of commutative lattice-groups, Math Ann 155 (1964) 81 – 107. [123] Peano G : Sull’integrabilità delle equazioni differenziali del primo ordine, Atti Accad Sci Torino 21 (1886) 677 – 685. [124] Peressini A : Ordered topological vector spaces, Harper & Row, New York, Evanston, London, 1967. [125] Picard E : Sur la forme des intégrales des équations différentielles du premier ordre dans le voisinage de certains points critiques, Bull Soc Math France 12 (1884) 48 – 51. [126] Preus G : Foundations of topology: An approach to convenient topology, Kluwer Academic Publishers, 2002. BIBLIOGRAPHY 225 [127] Rauch J and Reed M : Nonlinear superposition and absorption of delta waves in one space dimension, J Funct Anal 73 (1987) 152 – 178. [128] Reed E E : Completions of uniform convergence spaces, Math Ann 194 (1971) 83 – 108. [129] Richtmyer R D : Principles of advanced mathematical physics, vol 2, Springer, New York, 1981. [130] Riesz F : Sur la décomposition des oérations fonctionelles linéaires, Atti del Congr Internaz dei Mat, Bologna 1928, 3 (1930). [131] Riesz F : Sur quelques notions fondamentales dans al théorie générale des opérations lnéaires, Annals of Math 41 (1940). [132] Rosinger E E : Pseudotopological structures, Acad R P Romı̂ne Stud Cerc Mat 14 (1963) 223 – 251. [133] Rosinger E E : Pseudotopological structures II, Stud Cerc Mat 16 (1964) 1085 – 1110. [134] Rosinger E E : Pseudotopological structures III, Stud Cerc Mat 17 (1965) 1133 – 1143. [135] Rosinger E E : Embedding of the D0 distributions into pseudotopological algebras, Stud Cerc Math 18 no 5 (1966). [136] Rosinger E E : Pseudotopological spaces: The embedding of the D0 distributions into algebras, Stud Cerc Math 20 no 4 (1968). [137] Rosinger E E : distributions and nonlinear partial differential equations, Springer Lecture Notes in Mathematics 684, 1978. [138] Rosinger E E : Nonlinear partial differential equations, sequential and weak solutions, North Holland Mathematics Studies 44, 1980. [139] Rosinger E E : Generalized solutions of nonlinear partial differential equations, North Holland Mathematics Studies, 146, 1987. [140] Rosinger E E : Nonlinear partial differential equations, an algebraic view of generalized solutions, North Holland Mathematics Studies, 164, 1990. [141] Rosinger E E : Global version of the Cauchy-Kovaleskaia theorem for nonlinear PDEs, Acta Appl Math 21 (1990) 331 – 343. [142] Rosinger E E : Characterization of the solvability of nonlinear PDEs, Trans AMS 330 (1992) 203 – 225. [143] Scheaffer D G : A regularity theorem for conservation laws, Adv Math 11 no. 3 (1973) 368 – 386. BIBLIOGRAPHY 226 [144] Schwarz L : Théorie des distributions I & II, Hermann, Paris, 1950. [145] Schwarz L : Sur l’impossibilite de la multiplications des distributions, C R Acad Sci Paris 239 (1954) 847 – 848. [146] Sendov B : Hausdorff approximations, Kluwer Academic, Boston, 1990. [147] Slemrod M : Interrelationships among mechanics, numerical analysis, compensated compactness and oscillation theory, Oscillation theory, computation and methods of compensated compactness, Springer, New York, 1986. [148] Sobolev S L : Le probleme de Cauchy dans l’espace des functionelles, Dokl Acad Sci URSS 7 no. 3 (1935) 291 – 294. [149] Sobolev S L : Methode nouvelle a resondre le probleme de Cauchy pour les equations lineaires hyperbokiques normales, Mat Sbor 1 no. 43 (1936) 39 – 72. [150] Spivak M : A comprehensive introduction to differential geometry Vol 1, 3rd Ed, Publsih or Perish Inc, Houston, 2005. [151] Tartar L : Compensated compactness and applications to PDEs, Nonlinear analysis and mechanics: Herriot Watt Symposium 4, Pitman, 1979. [152] Tong H : Some characterizations of normal and perfectly normal spaces, Duke Math Journal 19 (1952) 289 - 292. [153] van der Corput J G : Introduction to neutrix calculus, J. D’Analyse Math 7 (1959) 281 – 398. [154] van der Walt J H : Order convergence in sets of Hausdorff continuous functions, Honors Essay, University of Pretoria, 2004. [155] van der Walt J H : Order convergence on Archimedean vector lattices with applications, MSc Thesis, University of Pretoria, 2006. [156] van der Walt J H : The uniform order convergence structure on ML (X), Quaestiones Mathematicae 31 (2008) 55 – 77. [157] van der Walt J H : The order completion method for systems of nonlinear PDEs: Pseudo-topological perspectives, Acta Appl Math 103 (2008) 1 – 17. [158] van der Walt J H : The order completion method for systems of nonlinear PDEs revisited, To Appear in Acta Appl Math. [159] van der Walt J H : On the completion of uniform convergence spaces and an application to nonlinear PDEs, Technical Report UPWT 2007/14, University of Pretoria, 2007. [160] Weil A : Sur les espaces á structures uniformes et sur la topologie générale, Hermann, Paris, 1937. BIBLIOGRAPHY 227 [161] Wyler O : Ein komplettieringsfunktor für uniforme limesräume, Math Nachr 40 (1970) 1 – 12. [162] Yoshida K : On the representation of the vector lattice, Proc Imp Acad (Tokyo) 18 (1942) 339 – 342. [163] Zaanen A C : Riesz spaces II, North Holland Publishing Company, 1983. [164] Zaanen A C : Introduction to operator theory in Riesz spaces, Springer, Berlin, Heidelberg, New York, 1997. [165] Zabuski N J : Computational synergies and mathematical innovation, J Comp Physics 43 (1981) 195 – 249.

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

### Related manuals

Download PDF

advertisement