Topics in Spectral Theory Publicações Matemáticas Topics in Spectral Theory Carlos Tomei PUC-Rio o 30 Colóquio Brasileiro de Matemática Copyright 2015 by Carlos Tomei Impresso no Brasil / Printed in Brazil Capa: Noni Geiger / Sérgio R. Vaz 30o Colóquio Brasileiro de Matemática Aplicações Matemáticas em Engenharia de Produção - Leonardo J. Lustosa e Fernanda M. P. Raupp Boltzmann-type Equations and their Applications - Ricardo Alonso Dissipative Forces in Celestial Mechanics - Sylvio Ferraz-Mello, Clodoaldo Grotta-Ragazzo e Lucas Ruiz dos Santos Economic Models and Mean-Field Games Theory - Diogo A. Gomes, Levon Nurbekyan and Edgard A. Pimentel Generic Linear Recurrent Sequences and Related Topics - Letterio Gatto Geração de Malhas por Refinamento de Delaunay - Marcelo Siqueira, Afonso Paiva e Paulo Pagliosa Global and Local Aspects of Levi-flat Hypersurfaces - Arturo Fernández Pérez e Jiří Lebl Introdução às Curvas Elípticas e Aplicações - Parham Salehyan Métodos de Descida em Otimização Multiobjetivo - L. M. Graña Drummond e B. F. Svaiter Modern Theory of Nonlinear Elliptic PDE - Boyan Slavchev Sirakov Novel Regularization Methods for Ill-posed Problems in Hilbert and Banach Spaces - Ismael R. Bleyer e Antonio Leitão Probabilistic and Statistical Tools for Modeling Time Series - Paul Doukhan Tópicos da Teoria dos Jogos em Computação - O. Lee, F. K. Miyazawa, R. C. S. Schouery e E. C. Xavier Topics in Spectral Theory - Carlos Tomei Distribuição: IMPA Estrada Dona Castorina, 110 22460-320 Rio de Janeiro, RJ E-mail: [email protected] ISBN: 978-85-244-0413-9 http://www.impa.br i \Spectrum" | 2015/5/11 | 12:17 | page 3 | #3 i i i Contents 1 Introduction 5 1.1 Contents . .7 1.2 Texts . .8 1.3 Basic notation . .8 1.4 Why spectral theory? . .9 2 Some basic facts 11 2.1 Linear transformations, matrices . 11 2.2 Krylov spaces, companion matrices . 16 2.3 Lanczos's procedure, Jacobi matrices . 17 2.3.1 Jacobi inverse variables . 19 2.4 Genericity and density arguments . 20 2.4.1 The resultant . 20 2.4.2 Density arguments . 21 2.5 Tensors and spectrum . 23 2.6 Wedges . 26 2.7 Some applications . 27 2.7.1 Roots of polynomials are eigenvalues . 27 2.7.2 The resultant revisited . 28 2.7.3 Algebraic numbers form a field . 28 2.8 Some examples: adjacency matrices . 29 2.8.1 Polygons and second derivatives . 29 2.8.2 Regular polytopes . 32 2.8.3 Semi-regular polytopes . 36 3 i i i i i \Spectrum" | 2015/5/11 | 12:17 | page 4 | #4 i i i 4 CONTENTS 3 Some analysis 39 3.1 Algebras of matrices and operators . 39 3.2 Smoothness of eigenpairs . 42 3.2.1 Bi-orthogonality, derivatives of eigenpairs . 43 3.2.2 Continuity of eigenvalues . 45 3.3 Some variational properties . 45 3.4 Approximations of small rank . 48 3.5 Isospectral manifolds . 50 3.5.1 More isospectral manifolds . 52 3.5.2 Two functionals on SΛ .............. 53 4 Spectrum and convexity 57 4.1 The Schur-Horn theorem . 57 4.2 Mutations, the high and low roads . 65 4.3 Interlacing and more . 66 4.3.1 Rank one perturbations . 66 4.3.2 The sum of two Hermitian matrices . 71 4.3.3 Weinstein-Aronsjan, Sherman-Morrison . 71 5 The spectral theorem 74 5.1 The Dunford-Schwartz calculus . 74 5.2 Orthogonal polynomials . 80 5.3 A quadrature algorithm . 84 5.4 The spectral theorem | a sketch . 86 i i i i i \Spectrum" | 2015/5/11 | 12:17 | page 5 | #5 i i i Chapter 1 Introduction Linear relations are unavoidable | one doubles the cause and the effect doubles, at least on a first guess. A substantial amount of the mathematics used in modeling is linear, which does not mean that it is trivial. The calculus of many variables, like optimizing in hundreds, millions of unknowns, is frequently a linear algebra problem. Besides, nonlinear problems are hard and substantial information may be ob- tained by linearization at points of interest. It is not clear that this is how we approach the teaching of linear algebra however. Sometimes, students are introduced to the subject as a sophisticated version of analytic geometry, for essentially visual purposes. Few students have the opportunity of relating linear al- gebra to... all things linear. The task is considered in engineering courses, but rarely within mathematics departments. There are historical reasons for this attitude. Linear algebra at some moment must have looked like the golden opportunity to present students to the axiomatic approach. Possibly the very first conse- quence of this point of view was the utter separation between linear and nonlinear theory: Jacobians and Hessians from Calculus hardly relate to matrices in linear algebra courses. Then there were the computational difficulties | a 3 × 3 matrix should have a simple eigenvalue or somehow it is inappropriate to fit in an exercise. Galois theory provides one of the most interesting no-go theorems in mathematics: radicals and the usual arithmetic 5 i i i i i \Spectrum" | 2015/5/11 | 12:17 | page 6 | #6 i i i 6 [CHAP. 1: INTRODUCTION symbols are not sufficient to write down the solutions of a polyno- mial of degree 5 with integer coefficients. Still, this is not the end of the world | one might invent new symbols or simply live with arbitrarily good approximations. Few students (possibly few profes- sionals) are conscious of the fact that most real numbers cannot even be described, a simple cardinality argument. And worse, among the important concepts in linear algebra lie ... nonlinear objects, eigenvalues, functions and groups of matrices. Derek Hacon, a colleague from PUC-Rio, used to say that fiber bundle theory is linear algebra with parameters. Few math students know how to take a derivative of an eigenvalue λ(t) of a matrix M(t). The n standard analysis course in R interacts poorly with matrix theory. The highlights from last century | quantum mechanics 1, the role of spectral theory in pure and applied dynamical systems; the numerical analysis of differential equations; the spectral theory of graphs, or in a more general setting, numerical linear algebra as a whole, being confronted with larger symmetric and non-symmetric matrices | indicate a combination of theory, practice and technology which should be a source of enthusiasm to any mathematician. Just to stick to one inevitable example, any Google search is a very large (numerical) problem in spectral graph theory. There is so much to choose from, what should be said in five short lectures? The topics intend to stimulate the interaction among different disciplines within mathematics, having in mind a public of graduate students. There are arguments involving algebra, basic real analysis, geometry, some complex variable, a bit of measure theory, a couple of algorithms, differential equations... and extensive point- ers to more sophisticated material in algebraic topology, symplectic geometry, numerical and functional analysis. Alas, everything is deterministic: there is nothing about random matrices or eigenvalue distributions. Also, there is nothing about the integrable systems associated to spectral theory. Acknowledgements abound. The Departamento de Matem´aticaat PUC-Rio allowed me to teach a number of courses in these subjects, and students from different departments contributed with a large 1According to Reed and Simon ([45]), the fact that the point spectrum of the Schr¨odingeroperator of the hydrogen atom describes with spectacular precision the frequencies of its emission spectrum borders on the scientifically embarrassing. i i i i i \Spectrum" | 2015/5/11 | 12:17 | page 7 | #7 i i i [SEC. 1.1: CONTENTS 7 spectrum of opinions. Some colleagues are friends and mentors | Percy Deift, Charlie Epstein, Nicolau Saldanha. Peter Lax, Beresford Parlett, Barry Simon are sources of inspiration. Years of subsidies from CNPq, CAPES and FAPERJ are also gratefully acknowledged. 1.1 Contents There are threads across the text. Chapter 2 contains some basic algebraic constructions which are extensively used. The fact that matrices form an algebra lead to cyclic vectors, companion and Ja- cobi matrices, inverse variables for Jacobi matrices, an introduction to orthogonal polynomials and eventually an indication of how the spectral theorem for self-adjoint operators in infinite dimension fol- lows from its counterpart for tridiagonal matrices. Tensor products and their spectral properties give rise to the resultant, the SVD de- composition, small rank approximations of symmetric and nonsym- metric matrices. There is frequent interplay between the invariant formalism and the use of coordinates. From the very start, density arguments are used to simplify proofs: matrices with distinct eigen- values are usually easier to handle. The study of eigenvalues and eigenvectors as objects which depend smoothly on matrices extends to some basic geometry of isospectral manifolds, which in turn are presented as natural phase spaces of algorithms for the computation of spectrum. Other geometric meth- ods in the study of eigenvalues are exemplified by standard results | the Schur-Horn theorem, spectral interlacing | for which elementary proofs are given, as opposed to the current symplectic approach. The standard road to the spectral theorem is the construction of a powerful functional calculus. Since good presentations are available, we decided instead to stop along the road, in particular the Dunford- Schwartz calculus, for better appreciation of some details. The text contains a number of references for further study. And this is perhaps the real motivation for these notes: to convince the reader of the vitality of the subject. i i i i i \Spectrum" | 2015/5/11 | 12:17 | page 8 | #8 i i i 8 [CHAP. 1: INTRODUCTION 1.2 Texts There are periodicals, libraries, dedicated to the subject. Here I just quote a few classics, of great mathematical and pedagogical value.