Applied and Numerical Harmonic Analysis Series Editor John J. Benedetto University of Maryland Editorial Advisory Board Akram Aldroubi Vanderbilt University
Douglas Cochran Arizona State University
Ingrid Daubechies Princeton University
Hans G. Feichtinger University of Vienna
Christopher Heil Georgia Institute of Technology James McClellan Georgia Institute of Technology Michael Unser Swiss Federal Institute of Technology, Lausanne M. Victor Wickerhauser Washington University
Murat Kunt Swiss Federal Institute of Technology, Lausanne Wim Sweldens Lucent Technologies, Bell Laboratories Martin Vetterli Swiss Federal Institute of Technology, Lausanne
Ovidiu Calin Der-Chen Chang Kenro Furutani Chisato Iwasaki
Heat Kernels for Elliptic and Sub-elliptic Operators Methods and Techniques
Ovidiu Calin Department of Mathematics Eastern Michigan University Ypsilanti, MI 48197, USA
[email protected]
Kenro Furutani Department of Mathematics Science University of Tokyo 2641 Yamazaki, Noda Chiba 278-8510, Japan
[email protected]
Der-Chen Chang Department of Mathematics and Statistics Georgetown University Washington, DC 20057, USA
[email protected]
Chisato Iwasaki Department of Mathematical Science University of Hyogo 2167 Shosha Himeji 671-2201, Japan
[email protected]
ISBN 978-0-8176-4994-4 e-ISBN 978-0-8176-4995-1 DOI 10.1007/978-0-8176-4995-1 Springer New York Dordrecht Heidelberg London Library of Congress Control Number: 2010937587 Mathematics Subject Classication (2010): 22E25, 35B40, 35Q80, 35S05, 35R03, 35P10, 35K08, 35K30, 35K65, 35K05, 35F21, 35H10, 35J08, 49L99, 53B20, 53C15, 58J65, 53C17 c Springer Science+Business Media, LLC 2011 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper www.birkhauser-science.com
ANHA Series Preface
The Applied and Numerical Harmonic Analysis (ANHA) book series aims to provide the engineering, mathematical, and scientific communities with significant developments in harmonic analysis, ranging from abstract harmonic analysis to basic applications. The title of the series reflects the importance of applications and numerical implementation, but richness and relevance of applications and implementation depend fundamentally on the structure and depth of theoretical underpinnings. Thus, from our point of view, the interleaving of theory and applications and their creative symbiotic evolution is axiomatic. Harmonic analysis is a wellspring of ideas and applicability that has flourished, developed, and deepened over time within many disciplines and by means of creative cross-fertilization with diverse areas. The intricate and fundamental relationship between harmonic analysis and fields such as signal processing, partial differential equations (PDEs), and image processing is reflected in our state-of-theart ANHA series. Our vision of modern harmonic analysis includes mathematical areas such as wavelet theory, Banach algebras, classical Fourier analysis, time-frequency analysis, and fractal geometry, as well as the diverse topics that impinge on them. For example, wavelet theory can be considered an appropriate tool to deal with some basic problems in digital signal processing, speech and image processing, geophysics, pattern recognition, biomedical engineering, and turbulence. These areas implement the latest technology from sampling methods on surfaces to fast algorithms and computer vision methods. The underlying mathematics of wavelet theory depends not only on classical Fourier analysis, but also on ideas from abstract harmonic analysis, including von Neumann algebras and the affine group. This leads to a study of the Heisenberg group and its relationship to Gabor systems, and of the metaplectic group for a meaningful interaction of signal decomposition methods. The unifying influence of wavelet theory in the aforementioned topics illustrates the justification for providing a means for centralizing and disseminating information from the broader, but still focused, area of harmonic analysis. This will be a key role of ANHA. We intend to publish with the scope and interaction that such a host of issues demands. Along with our commitment to publish mathematically significant works at the frontiers of harmonic analysis, we have a comparably strong commitment to publish v
vi
ANHA Series Preface
major advances in the following applicable topics in which harmonic analysis plays a substantial role: Antenna theory Prediction theory Biomedical signal processing Radar applications Digital signal processing Sampling theory Fast algorithms Spectral estimation Gabor theory and applications Speech processing Image processing Time-frequency and Numerical partial differential equations time-scale analysis Wavelet theory The above point of view for the ANHA book series is inspired by the history of Fourier analysis itself, whose tentacles reach into so many fields. In the last two centuries Fourier analysis has had a major impact on the development of mathematics, on the understanding of many engineering and scientific phenomena, and on the solution of some of the most important problems in mathematics and the sciences. Historically, Fourier series were developed in the analysis of some of the classical PDEs of mathematical physics; these series were used to solve such equations. In order to understand Fourier series and the kinds of solutions they could represent, some of the most basic notions of analysis were defined, e.g., the concept of “function.” Since the coefficients of Fourier series are integrals, it is no surprise that Riemann integrals were conceived to deal with uniqueness properties of trigonometric series. Cantor’s set theory was also developed because of such uniqueness questions. A basic problem in Fourier analysis is to show how complicated phenomena, such as sound waves, can be described in terms of elementary harmonics. There are two aspects of this problem: first, to find, or even define properly, the harmonics or spectrum of a given phenomenon, e.g., the spectroscopy problem in optics; second, to determine which phenomena can be constructed from given classes of harmonics, as done, for example, by the mechanical synthesizers in tidal analysis. Fourier analysis is also the natural setting for many other problems in engineering, mathematics, and the sciences. For example, Wiener’s Tauberian theorem in Fourier analysis not only characterizes the behavior of the prime numbers, but also provides the proper notion of spectrum for phenomena such as white light; this latter process leads to the Fourier analysis associated with correlation functions in filtering and prediction problems, and these problems, in turn, deal naturally with Hardy spaces in the theory of complex variables. Nowadays, some of the theory of PDEs has given way to the study of Fourier integral operators. Problems in antenna theory are studied in terms of unimodular trigonometric polynomials. Applications of Fourier analysis abound in signal processing, whether with the fast Fourier transform (FFT), or filter design, or the adaptive modeling inherent in time-frequency-scale methods such as wavelet theory. The coherent states of mathematical physics are translated and modulated Fourier
ANHA Series Preface
vii
transforms, and these are used, in conjunction with the uncertainty principle, for dealing with signal reconstruction in communications theory. We are back to the raison d’ˆetre of the ANHA series! University of Maryland College Park
John J. Benedetto Series Editor
Preface
The Fourier transform is known as one of the most powerful and useful methods for finding fundamental solutions of operators with constant coefficients. However, in the general case this method has its own limitations. This book presents several other methods that can be used concurrently with the Fourier transform method to obtain heat kernels for elliptic and sub-elliptic operators. The text contains a large number of examples which facilitate understanding. An Overview for the Reader. The theory of parabolic operators describes the distribution of heat on a given manifold as well as evolution phenomena and diffusion processes. The solution of an initial value problem for a parabolic partial differential equation depends on its heat kernel, which is the fundamental solution of the associated parabolic operator. Hence the importance of finding explicit formulas for these kernels. This monograph presents several theories for finding explicit formulas for heat kernels for both elliptic and sub-elliptic operators. These methods are treated in distinct chapters. We shall find heat kernels for classical operators by several different methods. Some methods come from stochastic processes, others come from quantum physics, and others are purely mathematical. Depending on the symmetry, geometry and ellipticity, some methods are more suited for certain operators rather than others. This book is a perfect reference material for graduate students, researchers in pure and applied mathematics as well as theoretical physicists interested in understanding different ways of approaching evolution operators. Scientific Outline. Heat kernels arise naturally from probabilistic properties of stochastic processes. The transition density of a stochastic process provides the heat kernel for the associated generator operator, which is a second-order PDE operator. For instance, the generator of a Brownian motion in a plane is the operator 1 2 .@ C @2y /. On the other side, since the x- and y-components of the Brownian 2 x motion are independent, the joint transition density, given that the motion starts at .x0 ; y0 / at t D 0, is the product of two transition densities: .xx0 /2 .xx0 /2 1 1 1 1 Œ.xx0 /2 C.yy0 /2 e 2t e 2t p e 2t p D ; 2 t 2 t 2 t
which is the heat kernel for the aforementioned generator operator. ix
x
Preface
One of the large classes of operators studied in this book is the sum-of-squares operators. These operators might be either elliptic or sub-elliptic. The methods for finding the heat kernel depend on the commutativity condition of the operators. If the operators commute, then the heat kernel is the product of the heat kernels of the operators. If the operators do not commute, then the Trotter formula applies and the heat kernel is computed using the path integral method. This method was borrowed from quantum physics, and it was used to compute propagators for the Schr¨odinger equation. This technique of path integrals was initiated by Feynman in the early 1940s. Another class of operators investigated by this book is the sum between a second partial differential operator and a smooth potential. The case of linear and quadratic potentials can be solved explicitly by the path integral method, by van Vleck’s formula, or by geometric methods that encounter classical action and volume function. Finally, they can also be solved by means of pseudo-differential operators. For instance, the aforementioned operator 12 .@2x C @2y / has a heat kernel of the type K.x0 ; y0 ; x; yI t/ D V .t/e Scl .x0 ;y0 ;x;yIt / ; where V .t/ D
1 2 t
(0.0.1)
describes the density of geodesics emerging from .x0 ; y0 /, and
Scl .x0 ; y0 ; x; yI t/ D
1 1 Œ.x x0 /2 C .y y0 /2 D dist2 .x0 ; y0 /; .x; y/ 2t 2t
is the classical action on the Euclidean plane from .x0 ; y0 / to .x; y/ at time t. The key idea here is that the heat propagates mainly along the geodesics of the associated (sub-)Riemannian space. Then the heat density at a certain point in the space, which is the heat kernel of a certain operator, depends on the action along the geodesic and the density of geodesics at that point, given that all geodesics start at a given initial point. The density of geodesics is described by the volume function, which in certain special cases is given by a van Vleck determinant. This idea can be applied in general for elliptic operators to obtain locally closedform expressions for the heat kernels for certain operators. However, there are some limitations to the applicability of the method, and it depends on the nature of the potential function. The operators involving a power potential of degree greater than 2 do not in general have a closed-form solution. Geometrically speaking, this reduces to infinitely many geodesics between two points whose energies cannot be obtained in a closed-form solution in terms of the boundary points. None of the methods presented in this book can be applied to obtain exact solutions for these types of operators. For instance, the well-known problem of nonsolvability of the quartic oscillator problem is one of them. In the case of sub-elliptic operators, formula (0.0.1) no longer holds, due to the degeneracy of the operator, and another formula will take its place. This new formula will require a fiber integration along the characteristic variety (which for elliptic operators degenerate to a point). We note that pseudo-differential techniques and the path integration method do not, in general, provide easy ways of obtaining heat kernels for sub-elliptic operators.
Preface
xi
The novelty of this work is the diversity of methods aimed at computing heat kernels for elliptic and sub-elliptic operators. It is interesting that apparently distinct branches of mathematics, such as stochastic processes, differential geometry, special functions, quantum mechanics and PDEs, have all a common concept – the heat kernel. This concept unifies the aforementioned domains of mathematics, and hence deserves us dedicating our study to it. It is worth noting the relation of the material of this book with other previous books on the subject. One of the long-standing textbooks in the field is the well-known book of David Widder [111], which appeared in the mid-1970s. This treats the heat equation mainly in dimension 1, discussing boundary value problems, Green’s functions, integral transforms, theta-functions, the Huygens property, series expansions, heat polynomials, and other miscellaneous topics. Unlike the classical tract of the aforementioned reference, the present monograph covers the heat equation in several other contexts, such as geometric, stochastic, and quantic. However, the present work does not intend to replace Widder’s book, but to complement it with newer facts regarding the geometric and analytic contexts. This book contains most of the heat kernels computable by means of elementary functions. Future developments in this field can consider the possibility of closed-form expressions of heat kernels involving elliptic functions and hyperelliptic functions. These types of special functions have already appeared in the explicit computation of geodesics on certain sub-Riemannian manifolds, and we treated them in the monograph [27]. Similar types of functions also appeared in considering the heat kernel of operators with polynomial potential of degree greater than 2, one of the most famous being the “quartic oscillator” example. In general, the heat kernels of sub-elliptic operators associated with some sub-Riemannian manifolds of step larger than 2 may lead to the need for special functions. However, our present restrictive knowledge of the subject of hyperelliptic function theory indicates today’s limit of explicit computability of these types of heat kernels. The material was prepared as follows: Chaps. 1–8 were prepared by Ovidiu Calin; Chaps. 9–11 were written by Kenro Furutani, Chaps. 12–14 are attributed to Der-Chen Chang, while Chap. 15 was worked out by Chisato Iwasaki. Two of the authors (Calin and Chang) reside in the United States, while the others (Furutani and Iwasaki) live in Japan. Acknowledgments We wish to express our gratitude to Prof. Peter Greiner, who introduced us to the subject of this book. The monograph was written in 2009–2010 while the first author was on a sabbatical leave from Eastern Michigan University, and he was also partially supported by NSF Grant no. 0631541 and a fund from Tokyo University of Science during his stay there in the summer of 2009. The second author was partially supported by Hong Kong RGC Grant no. 600607, the Norwegian Research Council Research Grant no. 180275/D15 and a competitive research grant at Georgetown University. The third author was partially supported by the JSPS-grant and Grantsin-Aid for Scientific Research (C) no. 20540218, and the fourth author was partially supported by the Grants-in-Aid for Scientific Research (C) no. 21540194. Finally, we would like to express our thanks to the Birkh¨auser and ANHA editors, especially to J. J. Benedetto, in making this project possible.
Tokyo, 2009
xii
Preface The chapters’ diagram
Contents
Part I Traditional Methods for Computing Heat Kernels 1
Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3 1.1 Physical Significance of the Heat Equation.. . . . . . . .. . . . . . . . . . . . . . . . . 3 1.2 The Fundamental Solution . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 5 1.3 The Heat Equation in Its Setting . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 6 1.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 7 1.5 Theta-Functions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 7 1.6 Some Useful Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 10
2
A Brief Introduction to the Calculus of Variations . . . . . .. . . . . . . . . . . . . . . . . 2.1 Lagrangian Mechanics .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 2.1.1 The First Variation . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 2.1.2 The Second Variation . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 2.1.3 Geometrical Interpretation .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 2.1.4 The Case of Riemannian Geometry . . . . . . .. . . . . . . . . . . . . . . . . 2.1.5 Examples of Lagrangian Dynamics . . . . . . .. . . . . . . . . . . . . . . . . 2.2 Hamiltonian Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 2.3 The Hamilton–Jacobi Equation . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .
13 13 14 15 16 18 20 23 24
3
The Geometric Method . . . . . . P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.1 Heat Kernel for L D 12 ni;j D1 aij @xi @xj . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.2 Interpretation of the Volume Function . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . P 3.3 The Operator L D 12 niD1 bi @2xi . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.4 The Generalized Transport Equation.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.5 Solving the Generalized Transport Equation . . . . . . .. . . . . . . . . . . . . . . . . 3.6 The Operator L D 12 y m .@2x C @2y /, m 2 N . . . . . . . . .. . . . . . . . . . . . . . . . . 3.7 Heat Kernels on Curved Spaces .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.8 Heat Kernel at the Cut-Locus . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.9 Heat Kernel at the Conjugate Locus . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.10 Heat Kernel on the Half-Line . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.11 Heat Kernel on S 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3.12 Heat Kernel on the Segment Œ0; T . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .
27 27 30 31 33 38 40 45 46 47 47 48 49 xiii
xiv
Contents
3.13 3.14 3.15 3.16 3.17 3.18
Heat Kernel on the Cylinder .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . Operators with Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . The Linear Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . The Quadratic Potential.. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . P The Operator 12 @2xi ˙ 12 a2 jxj2 . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . The Operator L D 12 @2x C x2 . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .
50 52 54 58 60 64
4
Commuting Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 4.1 Commuting Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 4.1.1 The Operator L D 12 .@2x C @2y / . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 4.1.2 The Operator L D 12 .@2x C y@2y / . . . . . . . . . .. . . . . . . . . . . . . . . . . 4.1.3 The Operator L D 12 .x@2x C y@2y /. . . . . . . . .. . . . . . . . . . . . . . . . . 4.1.4 Sum of Squares of Linear Potentials . . . . . .. . . . . . . . . . . . . . . . . 4.1.5 The Operator L D 12 .x 2 C y 2 /.@2x C @2y / . . . . . . . . . . . . . . . . .
71 71 71 72 72 72 73
5
The Fourier Transform Method . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 5.1 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 5.2 Heat Kernel for the Grushin Operator . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 5.3 Heat Kernel for @2x C x@x @y . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 5.4 The Formula of Beals, Gaveau and Greiner . . . . . . . .. . . . . . . . . . . . . . . . . 5.5 Kolmogorov Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 5.6 The Operator 12 @2x C x@2y . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . P 5.7 The Generalized Grushin Operator 12 niD1 @2xi C 2jxj2 @2t . . . . . . . . . . 5.8 The Heisenberg Laplacian .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .
75 75 76 77 78 80 85 85 87
6
The Eigenfunction Expansion Method .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 89 6.1 General Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 89 6.2 Mehler’s Formula and Applications.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 91 6.3 Hille–Hardy’s Formula and Applications .. . . . . . . . . .. . . . . . . . . . . . . . . . . 94 6.4 Poisson’s Summation Formula and Applications . .. . . . . . . . . . . . . . . . . 98 6.4.1 Heat Kernel on S 1 . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 99 6.4.2 Heat Kernel on the Segment Œ0; T . . . . . . . .. . . . . . . . . . . . . . . . .100 6.5 Legendre’s Polynomials and Applications .. . . . . . . . .. . . . . . . . . . . . . . . . .102 6.6 Legendre Functions and Applications . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .104
7
The Path Integral Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .105 7.1 Introducing Path Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .105 7.2 Paths Integrals via Trotter’s Formula . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .107 7.3 Formal Algorithm for Obtaining Heat Kernels .. . . .. . . . . . . . . . . . . . . . .112 7.4 The Operator 12 @2x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .118 7.5 The Operator 12 .@2x C @x / . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .123 7.6 Heat Kernel for L D 12 .@2x @x / . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .125 7.7 Heat Kernel for L D 12 x 2 @2x C x@x . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .126 7.8 The Hermite Operator 12 .@2x a2 x 2 / . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .126
Contents
xv
7.9
Evaluating Path Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .132 7.9.1 Van Vleck’s Formula .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .132 7.9.2 Applications of van Vleck’s Formula . . . . .. . . . . . . . . . . . . . . . .133 7.9.3 The Feynman–Kac Formula . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .138 Non-Commutativity of Sums of Squares . . . . . . . . . . .. . . . . . . . . . . . . . . . .140 Path Integrals and Sub-Elliptic Operators . . . . . . . . . .. . . . . . . . . . . . . . . . .143
7.10 7.11 8
The Stochastic Analysis Method . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .145 8.1 Elements of Stochastic Processes . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .145 8.2 Ito Diffusion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .152 8.3 The Generator of an Ito Diffusion.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .154 8.4 Kolmogorov’s Backward Equation and Heat Kernel .. . . . . . . . . . . . . . .156 8.5 Algorithm for Finding the Heat Kernel . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .157 8.6 Finding the Transition Density . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .159 8.7 Kolmogorov’s Forward Equation . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .166 8.8 The Operator 12 @2x x@x . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .167 8.9 Generalized Brownian Motion . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .168 8.10 Linear Noise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .170 8.11 Geometric Brownian Motion .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .171 8.12 Mean Reverting Ornstein–Uhlenbeck Process . . . . .. . . . . . . . . . . . . . . . .174 8.13 Bessel Operator and Bessel Process . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .176 8.14 Brownian Motion on a Circle . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .181 8.15 An Example of a Heat Kernel . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .183 8.16 A Two-Dimensional Case . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .185 8.17 Kolmogorov’s Operator.. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .188 8.18 The Operator 12 x12 @2x2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .189 8.19 Grushin’s Operator .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .191 8.20 Squared Bessel Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .192 8.21 CIR Processes .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .193 8.22 Limitations of the Method .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .193 8.23 Operators with Potential and the Feynman–Kac Formula . . . . . . . . . .195
Part II Heat Kernel on Nilpotent Lie Groups and Nilmanifolds 9
Laplacians and Sub-Laplacians .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .201 9.1 Sub-Riemannian Structure and Heat Kernels .. . . . . .. . . . . . . . . . . . . . . . .201 9.1.1 Sub-Riemannian Structure .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .201 9.1.2 Heat Kernel of the Sub-Laplacian and Laplacian .. . . . . . . . .203 9.1.3 The Volume Form . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .206 9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups .. . . . . . . . . . . .207 9.2.1 Nilpotent Lie Groups.. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .207 9.2.2 The Heisenberg Group .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .209 9.2.3 Higher-Dimensional Heisenberg Groups .. . . . . . . . . . . . . . . . .211 9.2.4 Quaternionic Heisenberg Group .. . . . . . . . . .. . . . . . . . . . . . . . . . .213 9.2.5 Heisenberg-Type Lie Algebra . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .216
xvi
Contents
9.2.6 9.2.7
Free Two-Step Nilpotent Lie Algebra .. . . .. . . . . . . . . . . . . . . . .219 The Engel Group . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .220
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians . . . . . . . . . . . . . . . .225 10.1 Spectral Decomposition and Heat Kernel.. . . . . . . . . .. . . . . . . . . . . . . . . . .225 10.2 Complex Hamilton–Jacobi Theory.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .229 10.2.1 Path Integrals and Integral Expression of a Heat Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .230 10.2.2 A Solution of a Hamilton–Jacobi Equation . . . . . . . . . . . . . . . .232 10.2.3 The Generalized Transport Equation .. . . . .. . . . . . . . . . . . . . . . .237 10.2.4 Heat Kernel for the Sub-Laplacian . . . . . . . .. . . . . . . . . . . . . . . . .244 10.2.5 Heat Kernel for Laplacian I . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .251 10.2.6 Heat Kernel for Laplacian II . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .252 10.2.7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .255 10.3 Grushin-Type Operators and the Heat Kernel . . . . . .. . . . . . . . . . . . . . . . .263 10.3.1 Grushin-Type Operators . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .263 10.3.2 Heisenberg Group Case . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .265 10.3.3 Heat Kernel of the Grushin Operator .. . . . .. . . . . . . . . . . . . . . . .267 11 Heat Kernel for the Sub-Laplacian on the Sphere S 3 . . .. . . . . . . . . . . . . . . . .273 11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3 . .273 11.1.1 S 3 in the Quaternion Number Field . . . . . . .. . . . . . . . . . . . . . . . .273 11.1.2 Heat Kernel of the Sub-Laplacian on S3 . .. . . . . . . . . . . . . . . . .277 11.1.3 Spherical Grushin Operator and the Heat Kernel .. . . . . . . . .282 Part III Laguerre Calculus and the Fourier Method 12 Finding Heat Kernels Using the Laguerre Calculus . . . . .. . . . . . . . . . . . . . . . .289 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .289 12.2 Laguerre Calculus. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .291 12.2.1 Twisted Convolution . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .292 12.2.2 P.V. Convolution Operators . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .293 12.2.3 Laguerre Functions .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .294 12.2.4 Laguerre Calculus on H1 . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .295 12.2.5 Laguerre Calculus on Hn . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .296 12.2.6 Left Invariant Differential Operators . . . . . .. . . . . . . . . . . . . . . . .297 12.3 The Heisenberg Sub-Laplacian.. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .300 12.4 Powers of the Sub-Laplacian .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .303 12.5 Heat Kernel for the Operator Lm ˛ . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .306 12.6 Fundamental Solution of the Paneitz Operator . . . . .. . . . . . . . . . . . . . . . .309 12.6.1 Laguerre Tensor of the Paneitz Operator ... . . . . . . . . . . . . . . . .311 12.6.2 Fundamental Solution: The Case n 2 . . .. . . . . . . . . . . . . . . . .313 12.6.3 Fundamental Solution: The Case n D 1 . .. . . . . . . . . . . . . . . . .316 12.7 Heat Kernel of the Paneitz Operator . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .318 12.8 Projection and Relative Fundamental Solution .. . . .. . . . . . . . . . . . . . . . .321
Contents
xvii
12.9 The Kernel ‰m .z; t/ for m > n .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .326 12.10 The Isotropic Heisenberg Group . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .328 12.11 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .330 13 Constructing Heat Kernels for Degenerate Elliptic Operators .. . . . . . . . .333 13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .333 13.2 Finding Heat Kernels for Operators Lj , j D 1; : : : ; 5 .. . . . . . . . . . . . .335 13.3 Some Explicit Calculations .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .340 13.3.1 Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .341 13.3.2 Kolmogorov Operator .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .342 13.3.3 The Operator L3 . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .344 13.3.4 The Operator L4 . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .346 13.3.5 The Operator L5 . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .347 14 Heat Kernel for the Kohn Laplacian on the Heisenberg Group .. . . . . . . .349 14.1 The Kohn Laplacian on the Heisenberg Group .. . . .. . . . . . . . . . . . . . . . .349 14.2 Full Fourier Transform on the Group . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .351 14.3 Solving the Transformed Heat Equation Using Hermite Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .353 14.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .358 Part IV
Pseudo-Differential Operators
15 The Pseudo-Differential Operator Technique . . . . . . . . . . . .. . . . . . . . . . . . . . . . .361 15.1 Basic Results of Pseudo-Differential Operators .. . .. . . . . . . . . . . . . . . . .361 15.1.1 Definition of Pseudo-Differential Operators . . . . . . . . . . . . . .362 15.1.2 Calculus with Pseudo-Differential Operators .. . . . . . . . . . . . .363 15.1.3 Proof of Lemma 15.1.3 . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .366 15.2 Fundamental Solution by Symbolic Calculus: The Nondegenerate Case . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .372 15.3 Basic Results for Pseudo-Differential Operators of Weyl Symbols 376 15.3.1 Calculus with Pseudo-Differential Operators .. . . . . . . . . . . . .376 15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .382 15.4.1 Construction of the Symbol.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .382 P 15.4.2 The Symbol pm D 12 `j D1 qj2 . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .396 15.4.3 The Special Case of Quadratic Symbols . .. . . . . . . . . . . . . . . . .397 15.4.4 A Key Theorem for Eigenfunction Expansion .. . . . . . . . . . . .400 15.5 The Hermite Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .400 15.5.1 Exact Form of the Symbol of the Fundamental Solution .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .401 15.5.2 Eigenfunction Expansion . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .403 15.6 The Grushin Operator.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .407 15.7 Exact Form of the Symbol of the Fundamental Solution for the Sub-Laplacian .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .408
xviii
Contents
15.8 15.9
The Sub-Laplacian on Step-2 Nilpotent Lie Groups . . . . . . . . . . . . . . . .410 The Kolmogorov Operator . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .415
Appendix . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .417 Conclusions . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .419 List of Frequently Used Notations and Symbols . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .423 References .. . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .425 Index . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .431
Part I
Traditional Methods for Computing Heat Kernels
Chapter 1
Introduction
1.1 Physical Significance of the Heat Equation Consider the problem of finding the temperature of a gas contained in a volume V , which is supposed to be a bounded, open and connected set in R3 . Denote by u.x; t/ the gas temperature at the point x 2 V at time t. Let be a subdomain of V with smooth boundary @. Let be the unit normal to @; see Fig. 1.1a. In order to write the equation of temperature evolution, we shall consider the law of conservation of thermal energy on the domain , under the assumption that there is no external absorption or application of heat: The rate of change of the energy with respect to time in is equal to the net flow of energy across the boundary @. The total thermal energy (heat) in at time t is Z U.t/ D
u.x; t/ dx;
(1.1.1)
and the net flow of energy across the boundary is given by the Fourier law Z Q.t/ D @
@u.x; t/ d; @
where d is the area element on @. The normal derivative @u.x; t/ D hrx u.x; t/; .x/i; @
x 2 @;
is the flux density normal to the boundary @. If x 2 @, the only gas particles in a small neighborhood U of x that hit @ are those which have a normal direction to @. The other particles in U that do not have a perpendicular direction to @ do not have a contribution toward the term Q.t/, since they are colliding with other particles in U and are bouncing back: see Fig. 1.1b.
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 1, c Springer Science+Business Media, LLC 2011
3
4
1 Introduction
a
b
Fig. 1.1 .a/ The domain and its normal . .b/ The particles which do not have normal direction do not contribute to the heat exchanged through @
Applying Gauss’s formula, the heat changed through the boundary becomes Z
Z Q.t/ D
divrx u.x; t/ dx D
x u.x; t/ dx:
(1.1.2)
The aforementioned conservation of energy can be written as @ U.t/ D Q.t/I @t using (1.1.1) and (1.1.2), it becomes Z
@ u.x; t/ dx D @t
Z
x u.x; t/ dx:
Since is an arbitrary domain of V , it follows that @ u.x; t/ D x u.x; t/; @t
x 2 V; t 0;
which is called the heat equation for the Laplacian x on R3 . A similar approach can beapplied in the case when .V; gij / is a Riemannian manifold and ; .x1 ; : : :; xn / is a local chart on V . In this case the measure dx is p replaced by the volume element d vg D det.gij /dx1 ^ ^dxn and the divergence and the gradient are given by X @ 1 det.gij /X j ; div X D det.gij / @xi n
i D1
rx u D
X i;j
g ij
@u @ : @xj @xi
1.2 The Fundamental Solution
5
The Laplacian in this case becomes the elliptic operator LD
X
g ij
i;j
where ijk D
1 2
X `
g k`
@gi ` @xj
C
X @2 @ ijk ; @xi @xj @xk k
@gj ` @xi
@gij @x`
are the Christoffel symbols.
1.2 The Fundamental Solution Consider the following heat problem: Given two distinct points x0 ; x on a Riemannian manifold .V; g/, assume that a unit volume of heat is applied at the point x0 at time t D 0. What is the heat distribution at any time t > 0? Equivalently, what is volume of heat K.x0 ; xI t/ that transmits from x0 to x after time t? K.x0 ; xI t/ is defined on V V .0; 1/ and is called the fundamental solution of the heat operator, or the heat kernel, if it satisfies the equations @ K.x0 ; xI t/ D LK.x0 ; xI t/; @t lim K.x0 ; xI t/ D ıx0 :
x 2 V; t 0;
t &0
Since in the compact case (with no boundary) no heat is lost, the total amount of heat is preserved: Z Z K.x0 ; xI t/ dx D ıx0 .x/dx D 1: This is the reason why K.x0 ; xI t/ can be considered as a probability density function for any t > 0 for a certain random process discussed in Chap. 2. For instance, in the one-dimensional case, at t D 0 the density is the Dirac distribution ıx0 centered at x0 , while at t > 0, the distribution is Gaussian; see Fig. 1.2a, b.
a y
b y
± x0 t =0 x0
x
t>0 x0
Fig. 1.2 .a/ The Dirac distribution ıx0 . .b/ The one-dimensional Gaussian
p1 e 2 t
.xx0 /2 2t
x-
6
1 Introduction
The main use of the heat kernel is to find the temperature evolution in V , provided the initial distribution of temperature is given. For any continuous, bounded function , Cauchy’s problem @ u.x; t/ D Lu.x; t/; @t u.x; 0/ D .x/;
x 2 V; t 0;
Z
has the solution u.x; t/ D
V
K.x; yI t/.y/ d vg :
Since K.x; yI t/.y/ represents the volume of heat transmitted from y to x after time t, given the initial temperature .x/, the kernel K.x; yI t/ is sometimes also called a propagator.
1.3 The Heat Equation in Its Setting The heat equation is a particular case of a more general linear second-order partial differential equation of the type A@2x C B@x @t C C @2t C D@x C E@t C F D 0; where the coefficients A; B; C; D; E; F are functions on x and t only. The following classification is familiar from undergraduate courses on differential equations: Condition B 2 < 4AC B 2 > 4AC B 2 D 4AC
Type Elliptic Hyperbolic Parabolic
Example @2x u C @2t u D 0 @2x u @2t u D 0 @2x u @t u D 0
Name Laplace Wave Heat
The type is suggested by the corresponding conic curve defined by the principal symbol. The Laplace equation describes static equations such as minimal surfaces, gravitational potentials, electric potentials, etc., having the Riemannian geometry as its underlined geometry. The wave operators describe the propagation of mechanical, radio and electro-magnetic waves, and radiation. It is associated with a geometry of Lorentz type. The heat equation describes diffusion phenomena in classical physics, Schr¨odinger-type equations in quantum mechanics, and the Black–Scholes equation in finance. The reader can find an elementary treatment of these equations, for instance, in Folland [44].
1.5 Theta-Functions
7
1.4 Methodology The first part of this book provides several methods for computing heat kernels for both elliptic and sub-elliptic operators using the following methods:
Stochastic analysis of diffusion processes The Path integral method The Geometric method involving the action along the geodesics The Fourier transform method The eigenfunction expansion method
The stochastic analysis method associates with each Ito diffusion a generator operator. The heat kernel of this operator is given by a transition density for the associated Ito diffusion. Several methods for computing transition densities are provided. The path integral method stems from the computation of propagators in quantum mechanics and has the following approaches:
By direct computation By using Trotter’s formula By using van Vleck’s formula By using Feynman–Kac’s formula
The eigenfunction expansion is a method which can be applied as long as we are able to compute the eigenfunctions and eigenvalues of the operator. However, closed-form solutions can be provided only if one of the following bilinear generating formulas is used:
Mehler’s formula Hille–Hardy’s formula Poisson’s summation formula Any other bilinear generating function
The geometric method expresses the heat kernel in terms of the action and volume element. It is a method which can be applied successfully as long as we are able to find the explicit formula for the action and solve the transport equation. It can be applied for both elliptic and sub-elliptic operators.
1.5 Theta-Functions A theta-function is a special function that describes the evolution of temperature on a segment domain subject to certain boundary conditions. We shall use them in the sequel when several heat kernels will be expressed in terms of theta-functions.
8
1 Introduction
Consider the heat equation @ @2 2 D0 @t @x
(1.5.3)
on Œ0; , with constant diffusivity . There are two important boundary-type conditions which lead to several types of theta-functions. Dirichlet boundary condition. The heat is maintained at zero temperature at the endpoints at all times t: .0; t/ D .; t/ D 0: The separation-of-variables method provides the solution of (1.5.3) in the form 1 .x; t/ D 2
X
2 t
.1/n e .2nC1/
sin.2n C 1/x;
n0
also called the first theta-function. We note the temperature will decrease to zero in the long run, so limt !1 1 .x; t/ D 0. Neumann boundary condition. There is no heat exchange at the endpoints x D 0; . In this case we say the endpoints are insulated and the following condition holds at the boundary: @ @ .0; t/ D .; t/ D 0: @x @x In this case the solution of (1.5.3) is 4 .x; t/ D 1 C 2
X
.1/n e 4n
2 t
cos 2nx;
n1
called the fourth theta-function. Since the heat cannot escape the domain, in the long run the solution tends to a constant equilibrium solution. A translation of 1 and 4 with =2 in the direction of x yields two other thetafunctions: X 2 2 .x; t/ D 1 .x C =2; t/ D 2 e 4.nC1=2/ t cos.2n C 1/x; n0
3 .x; t/ D 4 .x C =2; t/ D 1 C 2
X
e 4n
2 t
cos 2nx:
n1
The theta-functions 1 and 2 are periodic with period 2, and the functions 3 and 4 are periodic with period .
1.5 Theta-Functions
9
A few variations of the definition of theta-functions can be found in the literature. If we let q D e 4t , then the aforementioned formulas become 1 .x; q/ D 2
X
2
.1/n q .nC1=2/ sin.2n C 1/x;
n0
2 .x; q/ D 2
X
2
q .nC1=2/ cos.2n C 1/x;
n0
3 .x; q/ D 1 C 2
X
2
q n cos 2nx;
n1
4 .x; q/ D 1 C 2
X
2
.1/n q n cos 2nx:
n1
The graphs of the theta-functions with q D 0:7 and x 2 Œ0; 12 can be seen in Fig. 1.3. In the aforementioned formulas the variable q can also be considered complex with jqj < 1. In this case, if we let q D e i , with 2 C, and I m. / > 0, we consider the notation X 2 3 .xj / D 1 C 2 e i n cos 2nx; n1
a
b
3
3
2
2
1
1 2
4
6
8
10
12
−1
−1
−2
−2
−3
−3
c
d
3.0
3.0
2.5
2.5
2.0
2.0
1.5
1.5
1.0
1.0
0.5
0.5 2
4
6
8
10
12
2
4
6
8
2
4
6
8
10
10
12
12
Fig. 1.3 The theta-functions plotted for q D 0:7: .a/ 1 .x; 0:7/, .b/ 2 .x; 0:7/, .c/ 3 .x; 0:7/ and .d/ 4 .x; 0:7/
10
1 Introduction
which is sometimes useful to be written as X 2 3 .xji / D 1 C 2 e n cos 2nx:
(1.5.4)
n1
Using Euler’s formula and the fact that sine is an odd function, we have the equivalent definition 1 X
3 .xj / D
2
q n e 2nxi D
nD1
1 X
2
e i n e 2nxi :
nD1
Similar considerations can be carried out for the other theta-functions. More properties of the theta-functions can be found in Chap. 1 of [86].
1.6 Some Useful Integrals When computing heat kernels, one often gets into integrals which can be reduced to one given by the following result. Proposition 1.6.1. We have Z
r
b2 e 4a ; a r Z 2 2 e 4a ; e ay Ci y dy D a e
ay 2 Cby
dy D
8a 2 R ; b 2 R; 8a 2 R ; 2 R:
Proof. Completing the square and using the Gaussian integral yields Z e
ay 2 Cby
dy D e
b2 4a
Z
p
e
p 2 ayb=.2 a/
(1.6.5)
dy D e
b2 4a
1 p a
Z e
R
v2
(1.6.6) 2
e v d v D r dv D
p
b2 e 4a : a
The second integral is obtained by formally replacing b by i . Next we shall present a complete proof of this result. Completing the square yields Z
e ay
2 Ci y
Z p p 2 p 2 2 ayi =.2 a/ ui =.2 a/ 4a 1 e e dy D e dyD e p du a Z 2 1 2 e z dz; (1.6.7) D p e 4a p a Im zD=.2 a/ 2
4a
Z
where I m z denotes the imaginary part of the complex number z. The integral can be brought down to an integral on the real line as follows. Since the function p 2 e z is holomorphic, its integral on the contour f.R; 0/; .R; 0/; .R; =.2 a//;
1.6 Some Useful Integrals
11
Fig. 1.4 The rectangular contour f.R; 0/;p a//; .R; 0/; .R; =.2 p .R; =.2 a//g
p 2 .R; =.2 a//g vanishes; see Fig. 1.4. Since the limit of the integral of e z on the vertical edges of the aforementioned rectangle vanishes as R ! 1, it follows that Z Z p 2 z2 e dz D e x dx D : p Im zD=.2 a/
R
Substituting in (1.6.7) yields the desired result.
Chapter 2
A Brief Introduction to the Calculus of Variations
The Lagrangian and Hamiltonian formalisms will be useful in the following chapters when the heat kernel will be computed using the path integral and geometric variational methods. In the following we shall present a brief overview of the variational theory needed in the sequel.
2.1 Lagrangian Mechanics In classical mechanics a moving particle is completely described at any instance of time t by its position x.t/ and its velocity x.t/. P The position x belongs to the coordinate space, which is, in general, a Riemannian manifold with the metric defined by the kinetic energy. The space of the positions and velocities .x; x/ P is called the phase space, and it is identified with the tangent bundle TM of the coordinate space M . The pair .x; x/ P is called the state of the particle. A real-valued function defined on the tangent bundle L W TM ! R is called Lagrangian. In classical mechanics the usual Lagrangian is given by the difference between the kinetic energy and the potential energy of the particle: L.x; x/ P D K.x/ P U.x/; P where K.x/ P D 12 niD1 xP i2 and U.x/ is usually a polynomial function of x. The trajectory of a particle x.t/ in the coordinate space is a curve parameterized by the time parameter t. The Lagrangian L.x; x; P t/ describes the dynamics of the particle, in the sense that the particle moves on a trajectory x.t/ such that the following action integral Z
S.x0 ; x; / D
L.x.t/; x.t/; P t/ dt
(2.1.1)
0
is locally minimized under small variations of the path x.t/. We shall investigate this problem in the next few sections.
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 2, c Springer Science+Business Media, LLC 2011
13
14
2 A Brief Introduction to the Calculus of Variations
2.1.1 The First Variation For our study it would be sufficient to investigate the case of Lagrangians that do not depend explicitly on time t. Let x.t/ be a smooth curve joining x0 D x.0/ and x D x./. The action along x.t/ is given by
S D S x.t/ D
Z
L x.s/; x.s/ P ds:
0
The main problem of Lagrangian formalism is formulated below: Given the fixed endpoints x0 and x, find the path x.t/ for which the functional x.t/ ! S x.t/ (a) Has a “critical point” (b) Has a minimum Customarily, this problem is approached by considering a fixed path x.t/ and a variation x .t/ D x.t/ C .t/ with fixed endpoints x .0/ D x.0/ D x0 ;
x ./ D x./ D x0 :
The vector field .t/ defined on Œ0; is an arbitrary variation vector field with homogeneous boundary conditions .0/ D ./ D 0:
(2.1.2)
Expanding S about the path x.t/ in powers of yields Z @L @L S x .t/ D S x.t/ C ; P C ; dt @xP @x 0 2 2 2 Z 2 @ L @ L @ L ; P C ; P P C 2 ; dt C O. 3 / C 2Š 0 @xP 2 @x@x P @x 2 2 2 ı S C O. 3 /; D S C ıS C 2Š The path x.t/ is a “critical point” for the action functional S x.t/ if and only if ıS D 0. Integrating by parts and using the boundary conditions (2.1.2) yields Z 0 D ıS D 0
Z @L @L d @L @L ; P C ; dt D C ; dt: @xP @x dt @xP @x 0
2.1 Lagrangian Mechanics
15
Since the vector field is arbitrary, we obtain the variational equations, which are the famous Euler–Lagrange equations @L d @L D ; dt @xP @x
(2.1.3)
where xP D .xP 1 ; : : : ; xP n / and x D .x1 ; : : : ; xn / denote the velocity and position of the particle, respectively. A solution of (2.1.3) is called a classical path and will play a central role in the rest of the book.
2.1.2 The Second Variation Assume x.t/ is a solution of the Euler–Lagrange equations; i.e., it is a classical path. If the second variation is positive definite along x.t/, i.e., P .t/i > 0; hı 2 S .t/;
8t 2 .0; /;
(2.1.4)
for any variation vector field .t/ with .0/ D ./ D 0, then the path x.t/ minimizes the action functional between x0 and x./. In order to have relation (2.1.4) satisfied, it suffices to prove that minfhı 2 S .t/; P .t/i; 8t 2 .0; /g > 0:
Let be a variation vector field for which the previous minimum is reached. Then will satisfy the Euler–Lagrange equations for the associated Lagrangian L.; / P D hı 2 S .t/; P .t/i; which are given by d dt
2
2 d @ L @2 L @2 L @2 L @ L P C 2 C P D 0; @xP 2 dt @x@xP @x @x@x P @x@xP
(2.1.5)
which is called the Jacobi equation. A solution .t/ of (2.1.5) is called a Jacobi vector field. If there is a value t1 2 .0; / such that a Jacobi vector field .t1 / D 0, then inequality (2.1.4) fails at t D t1 . A well-known result of variational calculus states that as long as the Jacobi vector field doesn’t vanish, the classical path is still minimizing the action. The classical path ceases to be a minimizer as soon as the Jacobi vector field vanishes the first time. The vanishing points of a Jacobi vector field are called conjugate points. This is equivalent to saying that the classical path is minimizing between two consecutive conjugate points and is no longer minimizing after that.
16
2 A Brief Introduction to the Calculus of Variations
2.1.3 Geometrical Interpretation Next we shall deal with the geometrical significance of the Jacobi vector fields. Consider the classical path starting at x0 and all the neighboring, classical paths x.p; t/ starting at the same point and parameterized by their initial momenta p; this is x.p; 0/ D x0 ; see Fig. 2.1. For the time being we neglect the fixed endpoint condition at t D . The separation between two classical paths is given by x.p C ; t/ D x.p; t/ C J.p; t/ C O. 2 /; with J.p; t/ D Jij .p; t/, where Jij .p; t/ D
@xi .p; t/ ; @pj
i; j D 1; : : : ; n:
(2.1.6)
Since x.p; t/ are classical paths, the Euler–Lagrange equations are satisfied: @L d @L
D ; dt @xP r @xr
r D 1; : : : ; n:
Using the chain rule @L @xi @L @xP j @L D C @pk @xi @pk @xP j @pk @L @L P D Ji k C Jjk @xi @xP j and differentiating in the Euler–Lagrange equations yields 2 2 @L @2 L @ L @ L P d @2 L d JPi k D 0: Ji k C Ji k C dt @xP r @xP i dt @xP r @xi @xr @xi @xP r @xi @xr @xP i (2.1.7)
Fig. 2.1 The Jacobi vector field along a classical path between two conjugate points
2.1 Lagrangian Mechanics
17
A standard result of ODEs shows that the second-order-system (2.1.7) together with the following 2n2 initial conditions @xi .p; 0/ D 0; @pk @xP i .p; 0/ D ıi k ; JPik .p; 0/ D @pk Jik .p; 0/ D
has a unique solution Jij .p; t/ 6D 0, for t 2 .0; /, with > 0 small enough. Let p D x.0/ P be fixed. For any vector v D .v1 ; : : : ; vn / 6D 0, we can define the vector field .t/ D J.t/v along the classical path x.t/, by i .t/ D Ji k .t/vk . If there is a T > 0 such that .T / D 0, then det Ji k .T / D 0. The point x.T / is conjugate with x0 D x.0/, in the sense of the definition given in the previous section. This fact will be shown next. Multiplying (2.1.7) by vk and summing over k yields the equation 2 2 @ L i d @2 L d @2 L @2 L @ L i C P D 0; P C dt @xP r @xP i dt @xP r @xi @xr @xi @xP r @xi @xr @xP i which is exactly the Jacobi equation (2.1.5). The vector field .t/ becomes a Jacobi vector field. Since Ji k is nondegenerate for t 2 .0; T /, the set of Jacobi vectors forms an n-dimensional space at every point along the path x.t/. The first point where one Jacobi vector (and hence all of them) vanishes is a conjugate point with x0 . The classical path x.t/ minimizes the action as long as it does not pass through a conjugate point. In the following we shall write the matrix J , which satisfies (2.1.7), in terms of the action S . Let x0 and x be two fixed endpoints and let x.t/ be a classical path such that x.0/ D x0 and x./ D x. The momentum along x.t/ at time t is denoted by p.x0 ; xI t/. Let e x .t/ be the reverted curve defined by e x .t/ D x. t/ D x.tQ/. The curves e x .t/ and x.t/ have the same endpoints, x0 D e x ./ D x.0/ and x D e x .0/ D x./. The following relation between the momenta along the curves x.t/ and e x .t/ holds: p.x0 ; xI t/ D p.x; x0 I t/ D p.x; x0 I tQ/: Using the well-known relation
@S @x
D p along x.t/ at instances t and yields
@S.x0 ; x.t/I t/ D p.x0 ; x.t/I t/; @x.t/
@S.x0 ; xI / D p.x0 ; xI /: @x
A similar argument for the curve e x .t/ yields @S.x; x0 I tQ/ D p.x; x0 I tQ/ D p.x0 ; xI t/: @x0
18
2 A Brief Introduction to the Calculus of Variations
Differentiating with respect to x yields @S @p.x0 ; xI t/ 1 D D : @x@x0 @x J.p; x/ This means that
@S @x i @x0k
(2.1.8)
and Ji k are inverse matrices. Since at conjugate points
det Ji k D 0, it follows that at conjugate points D D det
@S
D ˙1: @x@x0
One of the main properties of D is that it satisfies a continuity equation of the following type (see [54]): @D X @ C .xP k D/ D 0; @t @xk
(2.1.9)
k
which means that D can be interpreted as a density of paths. This fact will be useful in the future chapters that deal with path integration, van Vleck’s formula and the geometric method of computing heat kernels.
2.1.4 The Case of Riemannian Geometry P Consider the Lagrangian L.x; x/ P D 12 ni;j D1 gij .x/xP i xP j , with gij a positive definite and non-degenerate matrix at each point x 2 Rn . This Lagrangian can also be considered on the tangent bundle of an n-dimensional Riemannian manifold .M; g/. However, since we are studying local properties, we can make the simplifying assumption M D Rn . In this case the Euler–Lagrange equations become the familiar equations of geodesics xR k .t/ C
X
ijk .x.t//xP i .t/xP j .t/ D 0;
k D 1; : : : ; n:
i;j
The classical paths are called geodesics and satisfy the following local result. Theorem 2.1.1. Given a point x0 on a Riemannian manifold .M; g/, there is a neighborhood V of x0 such that for any x 2 V, there is a unique geodesic joining the points x0 and x. The aforementioned result does not necessarily hold globally for any Riemannian manifold. However, it holds on compact manifolds, and in general on metrically complete manifolds, as the Hopf–Rinov theorem states; see [79].
2.1 Lagrangian Mechanics
19
Moreover, any geodesic is locally minimizing the action functional. The distance on the Riemannian manifold .M; g/ is defined by d.A; B/ D inff`.x/ j x geodesic; x.0/ D A; x./ D Bg; where the length of x.t/ is Z `.x/ D 0
sX
gij .x.t//xP i .t/xP j .t/ dt:
i;j
The classical action in this case is given by the formula S.x0 ; xI t/ D
d 2 .x0 ; x/ : 2t
Along a geodesic we have rS D x.t/, P where r is the gradient in the metric g: g.U; rS/ D dS.U /; for any tangent vector field U ; see [24]. The Jacobi equation (2.1.5) on Riemannian manifolds takes the form .t/ R D R .t/x.t/ P x.t/; P
(2.1.10)
where R is the Riemannian curvature tensor induced by the Levi–Civita connection D: R.U; V /W D DŒU;V W ŒDU ; DV W: If the manifold has constant curvature K, then R.U; V /W D K g.W; U /V g.W; V /U : In this case, under the additional hypotheses that x.t/ is unit speed and the Jacobi vector field is normal to the geodesic x.t/, the Jacobi equation (2.1.10) can be written in the suggestive form .t/ R D K.t/;
(2.1.11)
with the initial condition .0/ D 0. Solving, we distinguish the following cases: (1) Euclidean case: K D 0, .t/ D ct, c constant. See Fig. 2.2a. (2) Elliptic case: K D k 2 > 0, .t/ D c sin.kt/. See Fig. 2.2b. (3) Hyperbolic case: K D k 2 < 0, .t/ D c sinh.kt/. See Fig. 2.2c. In cases (1) and (3) there are no conjugate points, while in case (2) there are infinitely many conjugate points to x.0/ that occur at tn D n=k, n D 1; 2; : : : :
20
2 A Brief Introduction to the Calculus of Variations
a
b
c
Fig. 2.2 .a/ Euclidean case: K D 0; .b/ elliptic case: K > 0; .c/ hyperbolic case: K < 0
This behavior, for instance, occurs on a sphere. In general, all manifolds in situation (2) are compact. Just for the record, we include here a generalization of this case. Recall the notation for the diameter of a manifold as diam.M / D supfd.p; q/I p; q 2 M g. Theorem 2.1.2 (Myers). Let .M; g/ be a complete, connected n-dimensional Riemannian manifold. If there is a positive number k > 0 such that Ric .n 1/k 2 g;
(2.1.12)
where Ric denotes the Ricci tensor of M , then the following relations hold: (i) diam.M / =k (ii) M is compact As a special case of the previous theorem, we have Corollary 2.1.3. If .M; g/ is a complete, connected n-dimensional Riemannian manifold with constant curvature K satisfying K k 2 > 0, then (i) diam.M / =k (ii) M is compact References for Riemannian geometry and its variational methods are the books [24, 79, 94].
2.1.5 Examples of Lagrangian Dynamics In the following examples we shall assume that the initial and final positions of the particle x0 D x.0/ and x D x./ are given and we shall determine its trajectory between these endpoints.
2.1 Lagrangian Mechanics
21
Example 2.1.4 (The natural Lagrangian). Let U.x/ be a smooth potential, and consider the Lagrangian L D 12 xP 2 U.x/. The Euler–Lagrange equation is @U : (2.1.13) x.t/ R D @x Given the boundary conditions x.0/ D x0 , x./ D x, (2.1.13) might not always have a unique solution, regardless of how close the boundary points x0 and x are: see [24]. If the potential is linear or quadratic, the aforementioned boundary value equation has a unique solution, and in this case the action is well defined. The Jacobi equation for the previous Lagrangian takes the form ˇ @2 U.x/ ˇˇ .t/ R D ˇ @x 2 ˇ
.t/:
(2.1.14)
x.t /
We note that if U.x/ is linear or quadratic in x, then the previous equation becomes similar to (2.1.11). Example 2.1.5 (The free particle). If the Lagrangian is L.x; P x/ D 12 jxj P 2 D P 2 1 P j , then the variational equation (2.1.13) is x.t/ R D 0, and the trajectories are j x 2 the straight lines t x.t/ D x0 C x x0 : (2.1.15) The Jacobi equation (2.1.14) is .t/ R D 0. Using the initial conditions .0/ D 0, .0/ P D 1 yields the Jacobi vector field .t/ D t. The classical action is obtained by integrating the Lagrangian along the classical path: Z
S.x0 ; xI / D 0
.x x0 /2 1 x x0 2 dt D : 2 2
We recuperate formula (2.1.8) by taking the mixed derivative of the action @2 S 1 1 D D 6D 0: @x@x0 ./ Hence there are no conjugate points to x0 in this case. Next we shall check the continuity equation (2.1.9). We have D D 1t , xP k .t/ D
x k x0k , t
X @D @ 1 X @ @xk .vk D/ D C C @t @t t @xk n
kD1
n
kD1
and then
x k x0k t2
!
1 1 D 2 C 2 D 0: t t Example 2.1.6 (Particle in constant gravitational field). The Lagrangian describing the dynamics of a free-falling particle under the action of a constant gravitational
22
2 A Brief Introduction to the Calculus of Variations
force is given by L.x; P x/ D 12 xP 2 kx, k > 0. The trajectory x.t/ satisfies Galileo’s equation xR D k with the solution k t k x.t/ D t 2 C .x x0 / C t C x0 : 2 2 Since the Jacobi equation is .t/ R D 0 with the Jacobi vector field .t/ D t 6D 0, for t > 0, there are no conjugate points along the trajectory. We leave the computation of the classical action as an instructive exercise to the reader. Example 2.1.7 (The linear oscillator). If L.x; P x/ D 12 xP 2 the Euler–Lagrange equation is x.t/ R D kx, with the solution
k 2 x , 2
k > 0, then
p p
sinh. kt/ x.t/ D x x0 cosh. k/ C x0 cosh. kt/: p sinh. k/
p
This represents the trajectory of a particle under an elastic force F .x/ D kx. The Jacobi equation (2.1.14) can be R D k.t/, and the Jacobi vector field pwritten as .t/ is given by .t/ D p1 sinh. kt/ 6D 0, for t > 0. Then there are no conjugate k points to x0 along the classical path x.t/. A tedious computation shows that the classical action in this case is p p
2 k p .x C x02 / cosh. k / 2xx0 : S.x0 ; xI / D 2 sinh. k / Differentiating yields p @2 S 1 k D ; p D @x@x0 .t/ sinh. k / which is (2.1.8). Example 2.1.8 (The simple pendulum). The dynamics of a unit mass pendulum bob with unit length pendulum string under the action of the gravitational force is modeled by the Lagrangian L.x; P x/ D 12 xP 2 k.1 cos x/, k > 0. Its variational equation is given by the second-order differential equation x.t/ R D k sin x that cannot be solved using elementary functions. A solution using elliptic functions can be found in [24], p. 38. In the case of small oscillations the dynamics is approximated by the linearized pendulum equation x.t/ R D kx, with the solution p p p
sin. k t/ x.t/ D x x0 cos. k / C x0 cos. k t/: p sin. k/ The Jacobi equation (2.1.14) becomes .t/ R D k.t/:
2.2 Hamiltonian Mechanics
23
p Under the standard initial conditions, the Jacobi vector field is .t/ D p1 sin. k t/. k p Hence the conjugate points to x0 will occur along x.t/ at instances tn D n= k, n D 1; 2; : : : . The classical action is p p p
2 k … fn= kg: p .x C x02 / cos. k / 2xx0 ; S.x0 ; xI / D 2 sin. k / Example 2.1.9 (Particle in a constant electric field). The Lagrangian of a particle in a one-dimensional electric potential U.x/ D k=x, x > 0, is L.x; P x/ D 12 xP 2 2 k=x, k > 0. The trajectories satisfy the equation xR D k=x , which cannot be integrated using elementary functions. The same occurs for the Jacobi equation R D
2k : x3
2.2 Hamiltonian Mechanics An alternate way of describing the dynamics of a particle is using the Hamiltonian function. A real-valued function defined on the cotangent bundle of the coordinate space T M is called a Hamiltonian function. There is an intimate relationship between the Lagrangian and the Hamiltonian associated with a moving particle. The Hamiltonian associated with a Lagrangian is obtained using Legendre’s transform H.x; p/ D p xP L.x; x; P t/; where xP is a function of the momentum p obtained by solving the equation pD
@L : @xP
(2.2.16)
For instance, if the Lagrangian L.x; x/ P D 12 xP 2 x m , then the Hamiltonian is H.x; p/ D
1 2 p C xm: 2
The procedure works vice versa, i.e., for a given Hamiltonian, the associated Lagrangian can be obtained by L.x; x; P t/ D p xP H.x; p/; with the momentum p given by (2.2.16). The dynamics of the particle is described in the cotangent space by the Hamiltonian system of equations xP D @p H; pP D @x H: A solution x.t/; p.t/ of the aforementioned system is called a bicharacteristic.
24
2 A Brief Introduction to the Calculus of Variations
If the Hamiltonian H does not depend explicitly on the time variable t, the system is said to be conservative, because in this case the Hamiltonian evaluated along the above solutions is constant. This can be verified by using the chain rule. If x.t/; p.t/ is a bicharacteristic curve, then d P @p H D 0: H x.t/; p.t/ D x.t/ P @x H C p.t/ dt The value of the Hamiltonian evaluated along the bicharacteristic is called the total energy of the particle. This is a first integral of motion. If .x.t/; p.t// is a bicharacteristic curve, then the component x.t/ is a solution of the Euler–Lagrange equation. Conversely, if x.t/ is a solution of the Euler–Lagrange equations, then x.t/; @L .t/ is a bicharacteristic curve, called the lift of x.t/. @xP
2.3 The Hamilton–Jacobi Equation The classical action associated with the Lagrangian L.x; x; P t/ is obtained by integrating the Lagrangian along the solution of the Euler–Lagrange equation Z L x.t/; x.t/; P t dt: (2.3.17) Scl .x0 ; xI / D 0
This assumes that there is only one solution x.t/ satisfying the boundary conditions x.0/ D x0 and x./ D x. The classical action is a solution of the following Hamilton–Jacobi equation: @ Scl C H.x; @x Scl / D 0: @
(2.3.18)
Since the above equation is nonlinear, there might be more than one solution. In the case of a conservative system, if E D E.x0 ; x; / D H denotes the energy along the solution, the equation becomes @ Scl D E; @ with the solution Scl .x0 ; x; / D E.x0 ; x; / C c.x0 ; x/;
(2.3.19)
where c.x0 ; x/ D Scl .x0 ; x; 0/ is the initial condition. The action plays a very important role in the geometric methods of finding the heat kernel. To conclude our brief discussion, there are three ways of computing the action. 1. Starting from the Lagrangian. After solving the Euler–Lagrange equations, we integrate the Lagrangian along solutions using formula (2.3.17). This is a robust method and works as long as we are able to solve the Euler–Lagrange system of equations explicitly.
2.3 The Hamilton–Jacobi Equation
25
2. Solving the Hamilton–Jacobi equation. Substitute p D @x S in the expression of the Hamiltonian H and solve (2.3.18). Because it is nonlinear, there are no standard methods to solve the aforementioned equation. Depending on the problem, one may try different strategies. For instance, we can try to look for particular solutions of the type S.x; t/ D a.t/Cb.x/. Substituted in (2.3.18), the functions a.t/ and b.x/ satisfy the separable equation a0 .t/ C H.x; @x b/ D 0; so there is a constant C such that a0 .t/ D C H) a.t/ D C t C a.0/; H.x; @x b/ D C:
(2.3.20)
If the Hamiltonian H does not depend explicitly on the variable x, (2.3.20) can sometimes be reduced to an eikonal equation. For instance, if H D 12 jpj2 , then the function b.x/ satisfies j@x bj2 D C =2;
C > 0:
This equation has infinitely many solutions of the type r b.x/ D
v u n u2 X 2 .xi xi0 /2 ; jx x0 j D t C C
x0 2 Rn :
i D1
3. The conservative Hamiltonian case. If the Hamiltonian does not depend explicitly on the time parameter t, the action function is given by (2.3.19), where the energy should be expressed in terms of the endpoints x.0/ and x./. Next we shall work out an example. The aforementioned methods will be used to find the action function for the free particle case; see Example 2.1.5. Differentiating in (2.1.15) yields 1 xP D x./ x.0/ : The Lagrangian along the solution is L.x; x/ P D
1 2 jx./ x.0/j2 ; xP D 2 2 2
and using (2.3.17) we obtain the action Z
S.x0 ; xI / D 0
jx x0 j2 jx./ x.0/j2 jx./ x.0/j2 D : dt D 2 2 2 2
26
2 A Brief Introduction to the Calculus of Variations
If we try to solve the Hamilton–Jacobi equation, using @x S D p D (2.3.18) becomes
@L @xP
D x, P
@ @ 1 1 jx x0 j2 2 D0” ” S C jx.t/j P S D @ 2 @ 2 2 @ @ jx x0 j2 : SD @ @ 2 Integrating yields the aforementioned relation for the action. The reader can find more advanced topics on the calculus of variations in [6, 24, 52].
Chapter 3
The Geometric Method
This chapter deals with a construction of heat kernels from the geometric point of view. Each operator will be associated with a geometry. Investigating the geodesic flow in this geometry, one can describe the heat kernels for a large family of operators. The idea behind this method is that the heat flow propagates along the geodesics of the associated geometry. The “density” of the heat flow is described by a volume function that satisfies a transport equation which is an analog of the continuity equation from fluid dynamics. This corresponds to the density of paths given by the van Vleck determinant in the path integral approach. This method works for elliptic operators with or without potentials or linear terms. The method can be modified to work even in the case of sub-elliptic operators, as the reader will become familiar with in Chaps. 9 and 10. This method was initially applied for the Heisenberg Laplacian; see, for instance [28].
3.1 Heat Kernel for L D
1 2
Pn
i;j D1
aij @xi @xj
P Consider the elliptic differential operator L D 12 ni;j D1 aij @xi @xj . Since the matrix aij is symmetric, nondegenerate and positive definite at each point, then .Rn ; aij / becomes a Riemannian space, where .aij /1 D .aij /. The study of the geometry of this space provides a method for finding the heat kernel of the operator L. The heat kernel models the physical phenomenon of heat propagation from a point x0 to another point x within time t. The main idea is that the heat flows along the geodesics of the space .Rn ; aij / from the initial point x0 to the final point x. For the sake of simplicity, we shall assume in Theorems 3.1.1 and 3.4.3 that the Riemannian space .Rn ; aij / has the property that any two points can be joined by a unique geodesic, i.e., the space is geodesically complete, and there are no conjugate or cut points along the geodesics. This condition is always satisfied locally on a Riemannian manifold, so the aforementioned theorems provide a local expression for the heat kernels. However, in some cases, for instance in the case of spaces with negative or zero curvature, this property is globally satisfied, and hence the expression for the heat kernels is globally defined.
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 3, c Springer Science+Business Media, LLC 2011
27
28
3 The Geometric Method
The case when the cut locus of x0 is not empty, i.e., there is more than one geodesic joining the points x0 and x, the analysis is more complicated and shall be treated separately. The geometric method described in this chapter is based on finding geodesics. Obtaining the geodesics. The geodesics of the Riemannian space .Rn ; aij / can be obtained in two ways. Consider the following Hamiltonian function obtained by taking the principal symbol of the operator L: H.x; p/ D
n 1 X aij pi pj : 2
(3.1.1)
i;j D1
One way to define the geodesics is by considering the x-component of the bicharacteristic curve that solves the Hamiltonian system xP D Hp ; pP D Hx with the boundary conditions x.0/ D x0 ;
x.t/ D x:
The solution will be a geodesic joining the points x0 and x. For a large class of ODE systems, the solution of the previous boundary value problem is unique. The alternate approach for finding the geodesics is by solving the Euler– Lagrange system for the Lagrangian associated with the Hamiltonian (3.1.1). Assuming that we were able to find the geodesic x.s/ in one way or another, we can measure its length by the formula v Z tuX u n t `.x/ D aij .x.t//x i .t/x j .t/ dtI 0
i;j D1
Denote by d.x0 ; x/ the Riemannian distance between the points x0 and x measured in the metric aij . This is the length of the shortest geodesic x.s/ which joins x0 D x.0/ with x D x.t/. This length is independent of the parameter t, in the sense that it is invariant when the geodesic changes its parameterization. One may show that the associated action is S.x0 ; xI t/ D
d 2 .x0 ; x/ I 2t
see, for instance, [24]. The aforementioned action satisfies the Hamilton–Jacobi equation @t S C H.rS / D 0I (3.1.2)
3.1 Heat Kernel for L D
1 2
Pn
i;j D1 aij @xi @xj
29
where H is the Hamiltonian (3.1.1). There is an alternate way of finding the action by integrating the Lagrangian along the geodesic Z
t
S.x0 ; xI t/ D 0
1 L x.s/; x.s/ P ds; 2
where x W Œ0; t ! Rn is the geodesic which satisfies x.0/ D x0 and x.t/ D x. There is an intimate relationship between the geometric method and the path integration technique. In Chap. 7 the previous action is called the classical action and is denoted by Scl . It will be shown that the heat kernel can be represented as a path integral by Z K.x0 ; xI t/ D
Px0 ;xIt
e S.;t / d m./ D e Scl .x0 ;xIt /
Z
e S. P0;0It
;t /
d m. /:
R We shall see that in several cases the path integral P0;0It e S. ;t / d m. / is a function of t only. In this case the heat kernel takes the friendly form K.x0 ; xI t/ D V .t/e Scl .x0 ;xIt / ;
t > 0;
(3.1.3)
so we may say that the geometric method is a particular case of the path integral method. More precisely, we have the following result. P Theorem 3.1.1. Consider the operator L D 12 ni;j D1 aij @xi @xj and assume that L.Scl / is a function of t only. If there is only one geodesic joining the points x0 and x within time t, then the heat kernel of L is given by (3.1.3), where V .t/ satisfies the following transport equation: V 0 .t/ C V .t/L.Scl / D 0;
(3.1.4)
with the initial condition Z
V .0/ D lim
t &0
1 e Scl .x;x0 I0/ dx
:
(3.1.5)
Proof. Let S D Scl .x0 ; xI t/ denote the classical action. Since @xi @xj e S D @xi e S @xj S D e S .@xi S @xj S @xi @xj S /; then L.e S / D
n 1 X aij @xi @xj e S D e S .H.rS / L.S // : 2 i;j D1
(3.1.6)
30
3 The Geometric Method
On the other hand, @t e S V .t/ D @t e S V .t/ C e S @t V .t/ D e S .@t V .t/ V .t/@t S / :
(3.1.7)
Subtracting (3.1.6) and (3.1.7) yields .@t L/ e S V .t/ D @t e S V .t/ L e S V .t/ D e S @t V C V .t/ L.S / V .t/ .@t S C H.rS // ƒ‚ … „ D0
S
De D 0;
.@t V C V .t/ L.S //
where we used the Hamilton–Jacobi equation (3.1.2) and the transport equation (3.1.4). The solution V .t/ of the transport equation is unique up to a multiplicative constant. Relation (3.1.5) is fixing this constant. The reason behind relation (3.1.5) is the normalization condition for the heat kernel Z K.x; x0 I t/ dx D 1: Remark 3.1.2. Since the differential operator L is quadratic, the term L.S / does not depend on x in the cases when the action S depends on x linearly or quadratically. Formula (3.1.3) resembles van Vleck’s formula with van Vleck’s determinant V .t/. The function V .t/ shows how the geodesics spread away from the point x0 as t increases. For this reason we shall refer to V as the volume function.
3.2 Interpretation of the Volume Function One of the well-known conservation equations in fluid dynamics is the continuity equation @ C div. vk / D 0; @t where v D .v1 ; : : : ; vn / denotes the velocity of the fluid and is the fluid density. If the fluid follows the direction of the geodesic flow of a Riemannian manifold, then the velocity of the flow is related to the action by v D rScl :
3.3 The Operator L D
1 2
Pn
2 iD1 bi @xi
31
Assuming the fluid is homogeneous, i.e., it has the same density .t/ at each point, the aforementioned equation becomes 0 .t/ C .t/divrScl D 0: Substituting V .t/ D
p
.t/, we get V 0 .t/ C V .t/LScl D 0;
where L D 12 . This is exactly the transport equation (3.1.4) in the case of the Laplace–Beltrami operator. Hence the square of the volume function describes the “density” of the geodesic flow, i.e., the way the geodesics spread away from a given geodesic. We also note the relation with the Jacobi vector field emphasized in Chap. 2. In the following we shall deal with explicit computations in a few particular cases.
3.3 The Operator L D Consider the operator L D Hamiltonian function
1 2
1 2
Pn
i D1
Pn
i D1
bi @x2 i
bi @2xi , with bi > 0 constants. Associate the 1X bi pi2 : 2 n
H.p/ D
i D1
Solving the associated Hamiltonian system xP i .s/ D bi pi ; pPi .s/ D 0; x.0/ D x0 ;
x.t/ D x;
yields xi .s/ D bi pi s C x0 D
xi xi0 s C xi0 ; t
i D 1; : : : ; n;
where x.t/ D x D .x1 ; : : : ; xn /, x.0/ D x 0 D .x10 ; : : : ; xn0 /, and the pi are constants. The associated Lagrangian is 1X 1 2 xP ; 2 bi i n
L.x; x/ P D
i D1
32
3 The Geometric Method
and the action becomes Z
t
S.x0 ; xI t/ D 0
Z t
n 1X 1 L x.s/; x.s/ P ds D 2 bi
0
i D1
xi xi0 t
2 ds
n X .xi xi0 /2 : D 2bi t i D1
Since n 1X bj @2xj LS D 2
n X .xi xi0 /2 2bi t
j D1
!
n ; 2t
D
i D1
the transport equation becomes V 0 .t/ C
n V .t/ D 0: 2t
Integrating yields
V .t/ D ct n=2 : P By Theorem 3.1.1, the heat kernel of 12 niD1 bi @2xi is K.x0 ; xI t/ D ct
n=2
Pn
e
2
i D1
.xi xi0 / 2bi t
:
The constant c is determined from the relation Z K.x0 ; xI t/ dx D 1;
(3.3.8)
Rn
by using the well-known formula Z
1
e 2
Pn
i;j D1
Mij ui uj
Rn
In our case ui D xi xi0 and Mij D grating in (3.3.8) yields Z 1D Rn
D ct
K.x0 ; xI t/ dx D ct
n=2
d u1 d un D
1 ı , bi t ij
n=2
.2/n=2 : .det M /1=2
with det M D t n .
Z e
1 2
Pn
i D1
n Y .2/n=2 n=2 D c.2/ bi 1=2 t
Qn1 n
i D1 bi
i D1
1 i D1 bi / .
2
.xi xi0 /
Rn
Qn
bi t
!1=2 :
dx1 dxn
Inte-
3.4 The Generalized Transport Equation
33
Hence n Y
c D .2/n=2
!1=2 bi
:
i D1
We have arrived at the following result. Theorem 3.3.1. Let bi > 0. Then the heat kernel of L D K.x; x0 I t/ D .2 t/
n=2
n Y
!1=2 bi
e
1 2
Pn
2 i D1 bi @xi
is given by
2
Pn
.xi xi0 /
i D1
2bi t
;
t > 0:
i D1
When bi D 1, we obtain the familiar formula for the heat kernel of the Laplacian. Corollary 3.3.2. The heat kernel for L D K.x; x0 I t/ D .2 t/
1 2
n=2
e
Pn
2 i D1 @xi
Pn
i D1
is given by 2
.xi xi0 / 2t
;
t > 0:
3.4 The Generalized Transport Equation In Theorem 3.1.1 we had assumed that L.Scl / depends on t only. In this case the transport equation (3.1.4) can easily be solved by integration. However, there are some cases when L.Scl / depends on both variables t and x. In these cases the transport equation is replaced by the generalized transport equation. This equation describes how the heat density evolves along the geodesics. We shall start with an example. One of the operators for which L.Scl / depends on both x and t is L D 12 x@2x . The following discussion applies to the domain fxI x > 0g, where the operator is elliptic. The associated Hamiltonian function is H.x; p/ D
1 2 xp : 2
The Hamiltonian system is xP D Hp D xp; 1 pP D Hx D p 2 ; 2 with boundary conditions x.0/ D x0 , x.t/ D x. First we solve for p. Integrating in the equation
pP 1 D 2 p 2
(3.4.9)
34
3 The Geometric Method
yields pD
1 ; s0 C 12 s
with s0 2 R. Substituting in the first Hamiltonian equation, we get 1 xP H) ln x.s/ D D x s0 C 12 s and hence
1 2 ds D ln s0 C s C C; 2 s0 C 12 s
Z
1 2 x.s/ D e C s0 C s : 2
It is worth noting that x.s/ > 0 for s > 0 and hence only points on the positive semiaxis can be joined by a geodesic, i.e., no heat propagates in the negative semi-axis, a fact which agrees with the domain fxI x > 0g considered before. Given the endpoints x0 ; x > 0, we shall determine the unique constants C and s0 such that x0 D x.0/ and x D x.t/. This will show that the geodesic between x0 and x within time t is unique. Setting s D 0 and s D t, we arrive at the system x0 D e C s02 ; 1 2 C s0 C t : xDe 2 The elimination method provides s0 D
t p ; 2. x=x0 1/
eC D
p 4 .x C x0 xx0 /: 2 t
(3.4.10)
The Lagrangian associated with the Hamiltonian (3.4.9) is obtained by applying the Legendre transform L.x; x/ P D p xP H D D
1 xP 2 xP xP x 2 x 2 x
1 xP 2 : 2 x
Integrating along the geodesic and using (3.4.10), we obtain the classical action Z Scl .x0 ; xI t/ D 0
D
t
1 xP 2 .s/ ds D 2 x.s/
Z
t 0
p 2 .x C x0 xx0 /: t
2 1 e C s0 C 12 s 1 ds D e C t 2 2 e C s0 C 1 s 2 2
3.4 The Generalized Transport Equation
35
A straightforward computation shows L.Scl / D
1 2 1 x0 1=2 : x@x Scl D 2 4t x
Since the function V also depends on x, the hypothesis of Theorem 3.1.1 is not satisfied, and therefore the transport equation (3.1.4) does not hold. In the following we shall find the equation satisfied by the volume function V .t; x/. This will be called the generalized transport equation. Assume the heat kernel has the form K D V .t; x/e Scl . Since @x K D e Scl .@x V V @x Scl /; @2x K D e Scl @2x V 2@x Scl @x V C V .@x Scl /2 V @2x Scl ; @t K D e Scl .@t V @t Scl V / ; we obtain 1 LK D @t K x@2x K (2 De
! !) 1 1 2 2 2 @t V V @t Scl C x.Sx / x @x V 2@x Scl @x V V @x Scl 2 2 „ ƒ‚ …
Scl
D e Scl
D0
1 2 2 @t V x @x V 2@x Scl @x V V @x Scl ; 2
where we used that Scl satisfies the Hamilton–Jacobi equation 1 @t Scl C x.Sx /2 D 0: 2 Since LK D 0 for any t > 0, it follows that V .t; x/ satisfies the following generalized transport equation: 1 @t V x @2x V 2@x Scl @x V V @2x Scl D 0: 2
(3.4.11)
An explicit computation of the coefficients of @x V and V in the above equation yields r 1 2 1 1 x0 p V D 0: (3.4.12) @t V x@x V C .2x xx0 /@x V C 2 t 4t x This equation does not have a straightforward solution. The function V should be looked for as a product between an exponential term and a modified Bessel function, as in (8.20.79), where the operator is associated with a squared Bessel process.
36
3 The Geometric Method
However, an easy explicit solution can be obtained if x0 D 0. In this particular case it is not hard to check that a solution is cx V .t; x/ D 2 ; t with c constant. In this case the classical action is Scl .0; xI t/ D 2t x. By Theorem 3.1.1, the heat kernel for 12 x@2x from the origin is K.0; xI t/ D V .t; x/e Scl D
cx 2 x e t : t2
The constant c is determined from the condition Z K.0; xI t/ dx D 1: x>0
Since integration by parts shows that Z 1 c cx 2 x e t dx D ; 2 t 4 0 it follows that c D 4. If x0 D 0, then s0 D 0, and e c D 4x=t 2 , and the geodesics joining x0 D 0 and x at time t is x.s/ D xs 2 =t 2 . Proposition 3.4.1. The heat kernel for 12 x@2x from the origin is given by K.0; xI t/ D
4x 2 x e t ; t2
t > 0; x > 0:
Proof. We have shown that 12 x@2x K D 0 for t > 0. We still need to check that lim K.0; xI t/ D ı0 in the distribution sense. For any test function , we have t &0
Z
1
lim K.0; xI t/./ D lim
t &0
K.0; xI t/.x/ dx
t &0 0
Z
1
D lim
t &0 0
Z
1
D lim
t &0 0
Z
D .0/ „0
4x 2 x 2x t e .x/ dx make u D t2 t Z 1 tu t tu 4 tu u u ue du e d u D lim 2 t 2 2 2 t &0 0 2
1
ue u d u D .0/ D ı0 ./: ƒ‚ … D1
If x0 > 0, x > 0, then the heat travels in infinitely many ways between x0 and x since it is reflected at the wall x D 0; the kernel will be given by a series in this case.
3.4 The Generalized Transport Equation
37
Remark 3.4.2. The change of variable x D r 2 transforms the operator 12 x@2x into an operator that looks just like a two-dimensional Bessel operator with a changed sign 1 2 1 2 1 @ @r : x@ D 2 x 8 r r The above calculations also hold in the general case. This is given by the following result. P Theorem 3.4.3. Consider the operator L D 12 ni;j D1 aij @xi @xj with .aij / a symmetric, non-degenerate matrix. Assume L.Scl / depends on both variables x and t. If there is a unique geodesic joining the points x0 and x within time t, then the heat kernel of L is given by K.x0 ; xI t/ D V .t; x/e Scl .x0 ;xIt / ; where V .t/ satisfies the following generalized transport equation: X aij @xi S @xj V LV C V LS D 0; @t V C
(3.4.13)
(3.4.14)
i;j
Z
with lim
t &0
V .t; x/e Scl .x;x0 It / dx D 1:
(3.4.15)
Proof. To simplify notations, we shall denote the classical action by S D Scl .x0 ; xI t/. Differentiating in (3.4.13) yields @xj @xi K D e S V @xi S @xj S C @xi @xj V V @xi @xj S .@xi S @xj V C @xi V @xj S / ; LK D e S fVH.rS / C LV V LS a.rS; rV /g ; @t K D e S f@t V V @t S g; where 1X aij @xi S @xj S; 2 X a.rS; rV / D aij @xi S @xj V: H.rS / D
Since S satisfies the Hamilton–Jacobi equation @t S C H.rS / D 0; and V satisfies the generalized transport equation (3.4.14), we get .@t L/K D e S f@t V C a.rS; rV / LV C V LS g D 0:
Remark 3.4.4. If V does not depend on x, the generalized transport equation (3.4.14) becomes the usual transport equation (3.1.4).
38
3 The Geometric Method
The first two terms of the generalized transport equation (3.4.11) can be written as a derivative along the classical path x.t/: X @V d @V @V xP k D V t; x.t/ D C C a.rV; x/ P dt @t @xk @t @V C a.rV; rS /: D @t Hence we obtain another form of the generalized transport equation, d V t; x.t/ C V LS D LV: dt
3.5 Solving the Generalized Transport Equation The generalized transport equation (3.4.14) is in general hard to solve explicitly. Its solution V .t; x/ might be represented as an integral formula rather than being an elementary function. The following computation holds in the case of an elliptic operator and uses a series expansion for t small. A parametrix expansion can also be found in [84]. We shall consider V .k/ .t; x0 ; x/ D .2 t/n=2
0 .x0 ; x/
C
1 .x0 ; x/t
CC
k .x0 ; x/t
k
;
which approximates the volume function locally for t small, as a correction to the Euclidean volume element. Then K .k/ .x0 ; xI t/ D Vk .t; x0 ; x/e Scl .x0 ;xIt / is an approximation of the heat kernel K.x0 ; xI t/ D K .k/ .x0 ; xI t/ C O.t k /; We need to solve for
m
t 0:
recursively for small t. A computation shows
n @t V .k/ D .2 t/n=2 2t
0C
n 1 2
1
n C t 2 2
2
n Ct k1 k k : 2 (3.5.16)
Since the classical action on a Riemannian manifold is related to the Riemannian distance by d 2 .x0 ; x/ ; Scl .x0 ; xI t/ D 2t then rScl D
1 rd 2 .x0 ; x/. 2t
Hence
a.rScl ; rV .k/ / D .2 t/n=2
k 1 X 2 a rd .x0 ; x/; r 2t j D0
j
t j:
(3.5.17)
3.5 Solving the Generalized Transport Equation
We also have LV
.k/
n=2
D .2 t/
39
k X
L
j .x0 ; x/ t
j
(3.5.18)
j D0
and V .k/ LScl D .2 t/n=2
k X 1 2 L d .x0 ; x/ 2t
j .x0 ; x/t
j
:
(3.5.19)
j D0
Using (3.5.16)–(3.5.19) and equating the coefficients of t j with 0 in the generalized transport equation for V .k/ , @t V .k/ C a rS; rV .k/ LV .k/ C V .k/ LS D O.t k /; yields the following system in the unknown functions 2 Ld n 2 Ld n C 2 2 Ld n C 4
Ld 2 n C 2k
0;
1; : : : :
0
C a.rd 2 ; r
0/
D 0;
1
C a.rd 2 ; r
1/
D 2L
0;
2
C a.rd 2 ; r
2/
D 2L
1;
:: :
k
D
C a.rd 2 ; r
k/
(3.5.20)
:: :
D 2L
k1 ;
together with the initial conditions j .x0 ; x0 / D 0, j D 1; : : : ; k. In general, solving the above system is as hard as solving the generalized transport equation. The system can be solved only in simple cases, as we shall see in the next example. Example 3.5.1. Let L D 12 @2x . Then d 2 .x0 ; x/ D .x x0 /2 , L d 2 .x0 ; x/ D 1, and rd 2 .x0 ; x/ D 2.x x0 /. The first equation of the above system becomes 2.x x0 / with the solution
0
1
C 2.x x0 /
with the solution 1 1 .x0 /
D 0;
(3.5.21)
D c0 , a constant. Then the second equation becomes 2
because
0 0
D
0 1
D 0;
c1 D 0; x x0
D 0. Inductively we obtain all
j
D 0, j 1.
In the following we shall discuss a few more examples.
40
3 The Geometric Method
3.6 The Operator L D 12 y m .@x2 C @y2 /, m 2 N We shall discuss the cases m D 0; 1 and 2. Case m D 0. In this case the operator is the usual two-dimensional Laplacian L D 1 2 .@ C @2y / with the heat kernel 2 x K.x0 ; xI t/ D
.xx0 /2 C.yy0 /2 1 2t : e .2 t/
Case m D 1. The operator becomes L D 12 y.@2x C @2y /. We shall consider it in the domain f.x; y/I y > 0g where the operator is elliptic. The Hamiltonian function in this case is 1 (3.6.22) H D y.p12 C p22 /; 2 and the associated Lagrangian is given by LD
1 2 .xP C yP 2 /: 2y
We are interested in the geodesics and the action from .x0 ; y0 / to .x; y/ within time t. From the Hamiltonian system we have xP D Hp1 D yp1 ; yP ; y pP1 D Hx D 0 H) p1 D k constant; 1 1 pP2 D Hy D p12 C p22 : 2 2 yP D Hp2 D yp2 H) p2 D
Since the Hamiltonian function H does not depend explicitly on the parameter s, it is preserved along the solutions; i.e., there is a constant C > 0, which depends on the boundary points .x0 ; y0 /; .x; y/ and t, such that H.x; y; p1 ; p2 / D
1 2 C : 2
Using the expression of the Hamiltonian (3.6.22) yields the following equation in y: yP 2 k2 C 2 D C 2: y Solving for yP yields yP 2 D C 2 y k 2 y 2 :
(3.6.23)
3.6 The Operator L D 12 y m .@2x C @2y /, m 2 N
41
Completing the above to a square, we obtain ( ) 4 2 2 1 C C y yP 2 D k 2 : 4k 4 2 k2 With the substitution u D y ( uP D k 2
2
1 C2 , 2 k2
the previous equation becomes
C2 u2 2k 2
2 )
r H) uP D ˙k
C2 u2 : 2k 2
Separating and integrating yields Z
u.s/
Z
du
s
k” 0 u2 2 2k 2k u.s/ D arcsin u.0/ ˙ ks ” arcsin C2 C2 C2 u.s/ D sin. ˙ ks/; 2k 2 u0
where
q
D˙
C2 2k 2 2
2k 2 D arcsin u.0/ : C2
Going back to the variable y, we obtain y.s/ D
1 C2 .1 C sin. ˙ ks// : 2 k2
Integrating in the first Hamiltonian equation xP D ky yields Z s x.s/ D x0 C k y 0 cos. ˙ ks/ k C2 : s D x0 C 2 k2 ˙k
(3.6.24)
(3.6.25)
Sign convention. If we choose the “plus” sign in front of k, it means that y.s/ is increasing, and hence y > y0 . If we choose the “negative” sign in front of k, it means that y.s/ is decreasing, and hence y < y0 . Relations (3.6.25) and (3.6.24) provide the equations of the geodesics. In the following we shall eliminate their trigonometric part. Rewrite the equations as k C2 1 C2 sD cos. ˙ ks/; 2 2 k ˙2 k 2 1 C2 1 C2 D sin. ˙ ks/: y.s/ 2 k2 2 k2
x.s/ x0
42
3 The Geometric Method
Summing the squares yields the following implicit equation: 2 2 k C2 1 C2 1 C4 x.s/ x0 s C y.s/ D : 2 k2 2 k2 4 k4
(3.6.26)
Making s D 0 yields 2 1 C4 1 C2 D ” y 2 k2 4 k4 C2 y0 y0 2 D 0: k Since y0 6D 0, it follows that
C2 D y0 : k2
(3.6.27)
Then (3.6.26) can be rewritten as 2 2 k 1 1 x.s/ x0 y0 s C y.s/ y0 D y02 ” 2 2 4 2 k x.s/ x0 y0 s C y.s/ y.s/ y0 D 0: 2
(3.6.28)
From (3.6.28) it follows that on the half-plane f.x; y/I y > 0g the solution y.s/ is a decreasing function of s. If by contradiction we assume that y.s/ is increasing, then 0 < y0 < y.s/ and hence y.y y0 / > 0 and (3.6.28) cannot hold. Hence the sign in front of k in formulas (3.6.24)–(3.6.25) must be negative. A particular family of geodesics is obtained when both terms of (3.6.28) vanish: k y0 s 2 x x0 s; D x0 C t y.s/ D y0 : x.s/ D x0 C
In this case the geodesics are lines and the action is S .x0 ; y0 /; .x; y0 /I t D
Z Z
t 0 t
D 0
1 P 2 / ds .x.s/ P 2 C y.s/ 2y.s/ 1 .x x0 /2 1 x x0 2 ds D : 2y0 t 2y0 t
3.6 The Operator L D 12 y m .@2x C @2y /, m 2 N
Since LS D
43
1 .x x0 /2 1 1 D ; y.@2x C @2y / 2 2y0 t 2t
the transport equation is 1 V .t/ D 0; 2t c with the solution V .t/ D t 1=2 , c a constant. Then the heat kernel between .x0 ; y0 / and .x; y0 / is given by Theorem 3.1.1: V 0 .t/ C
c 1 .xx0 /2 K .x0 ; y0 /; .x; y0 /I t D 1=2 e 2y0 t ; t Finding the constant c. Z 1D D
K dx D c t 1=2
c
Z
t 1=2
e
1 2y
0
.xx0 /2 t
t > 0:
dx
.2y0 t/1=2 D c.2y0 /1=2 H) c D
1 : .2y0 /1=2
Hence K .x0 ; y0 /; .x; y0 /I t D
2 1 .xx0 / 1 2y t 0 e ; .2y0 t/1=2
t > 0:
The general case. From the Hamiltonian equation xP D ky and y > 0, it follows that if k > 0, then x.s/ is increasing and hence x0 < x. And if k < 0, then x.s/ is decreasing and then x0 < x. Making s D t in (3.6.28), we obtain 2 k x x0 y0 t C y y y0 D 0; 2 and then solving for k, we get
kD
8 p 2 ˆ < y0 t x x0 C y.y0 y/ ;
if x > x0 I
p x x0 y.y0 y/ ;
if x < x0 :
ˆ :
2 y0 t
We note that y0 y always, since y.s/ is decreasing. Finding the action between .x0 ; y0 / and .x; y/ within time t. Since the Lagrangian along the geodesics is 1 2 1 2 2 .xP C yP 2 / D .y p1 C y 2 p22 / 2y 2y 1 1 D y.p12 C p22 / D H D C 2 ; 2 2
LD
44
3 The Geometric Method
integrating yields the action S .x0 ; y0 /; .x; y/I t D
Z
t
1 2 1 C2 2 1 k t D y0 tk 2 C tD 2 2 2k 2 0 8 p 2 < y2 t x x0 C y.y0 y/ ; if x > x0 I 0 D : 2 x x py.y y/2 ; if x < x : 0 0 0 y0 t LD
Since LS D 12 y.@2x C @2y /S depends on t, x and y, the heat kernel is given in this case by Theorem 3.4.3: 8 2 p ˆ 2 xx0 C y.y0 y/ ; if x > x0 I < V .t; x; y/e y0 t K .x0 ; y0 /; .x; y/I y D 2 p ˆ : 2 xx0 y.y0 y/ V .t; x; y/e y0 t ; if x < x0 ; where V .t; x; y/ satisfies the generalized transport equation (3.4.14). Case m D 2. The operator becomes L D 12 y 2 .@2x C @2y /, which is the Laplace– Beltrami operator on the upperhalf-plane U D f.x; y/I y > 0g, with the metric ds2 D y12 .dx2 C dy 2 /. The Hamiltonian function is H D
1 2 2 y .p1 C p22 /; 2
(3.6.29)
with the associated Lagrangian Any two points of U can be joined by a unique geodesic, which is either a vertical radius or a half-circle perpendicular on the line fy D 0g. Hence, by Theorem 3.4.3, the heat kernel can be represented as a product Scl .x0 ;y0 /;.x;y/It : (3.6.30) K .x0 ; y0 /; .x; y/I t D V .t; x; y/e Let dh denote the hyperbolic distance on the upperhalf-plane. If the points A.x0 ; y0 / and B.x; y/ belong to the same vertical line, then the distance between them is ˇ AM ˇ ˇ y ˇ ˇ ˇ ˇ ˇ ˇ 0ˇ ˇ ˇ dh .A; B/ D ˇ ln ˇ D ˇ ln ˇ D ˇ ln y ln y0 ˇI BM y see Fig. 3.1a. Otherwise, ˇ BN=BM ˇ ˇ A0 M NB 0 ˇ ˇ .K x C r/.x K C r/ ˇ 0 ˇ ˇ ˇ ˇ ˇ ˇ dh .A; B/ D ˇ ln ˇ D ˇ ln ˇ; ˇ D ˇ ln AN=AM AA0 BB 0 y0 y where r 2 D y02 C .K x0 /2 ; KD
1 .x x0 /2 C y 2 y02 I 2 x x0
3.7 Heat Kernels on Curved Spaces
a
45
b A(xo,yo ) B(x,y)
N(-r,0)
A(xo,0)
O(K,0)
B (x,0)
M(r,0)
Fig. 3.1 The hyperbolic distance dh .A; B/. (a) The case when A and B are on a vertical direction; (b) A and B are on a non-vertical direction
see Fig. 3.1b. Since dh is the associated Riemannian distance on the upper space U, we have dh2 .x0 ; y0 /; .x; y/ Scl .x0 ; y0 /; .x; y/I t D : 2t Since LScl does not depend on t only, the transport equation is messy and the heat kernel in this case is hard to find in the product form (3.6.30). However, McKean [89] found the integral representation p Z 1 s2 2 se 2t t =2 e p ds; (3.6.31) .2 t/3=2 cosh s cosh where D dh .x0 ; y0 /; .x; y/ . It is interesting for further investigations to clarify the relationship between the integral representation (3.6.31) and the product formula (3.6.30). As we shall see in the next section, this relation is obvious in the case of a three-dimensional hyperbolic space. K .x0 ; y0 /; .x; y/I t D
3.7 Heat Kernels on Curved Spaces Let g be the Laplace–Beltrami operator on a Riemannian space .M; gij / of dimension n. Let R D g ij Rij be the Ricci scalar curvature of the space, which will be assumed constant. Let d.x0 ; x/ denote the Riemannian distance between the points x0 and x. The heat kernel for the elliptic operator g is given by K.x0 ; x; t/ D where
d 2 .x0 ;x/ 1 g.x/1=4 D 1=2 g.x0 /1=4 e Rt e 4t ; n=2 .4 t/
2 @ Scl D D det @x0 @x
is the van Vleck determinant; see Schulman [102], Chap. 24.
(3.7.32)
46
3 The Geometric Method
Applying the aforementioned formula for the classical spaces with constant curvatures 0; 1; 1, we arrive at the following classical results. The three-dimensional hyperbolic space. The Laplace–Beltrami operator on the three-dimensional hyperbolic space, i.e., the domain f.x1 ; x2 ; x3 /jx3 > 0g endowed with the metric ds2 D .dx21 C dx22 C dx23 /=.x3 /2 , is given by D x32 .@2x1 C @2x2 C @2x3 / x3 @x3 : The heat kernel of is obtained from formula (3.7.32) by letting scalar curvature R D 1 and hyperbolic distance d.x0 ; x/ D , KD
2 1 e t e 4t : 3=2 .4 t/ sinh
This relation was obtained by a direct method in [39], p. 396. The three-dimensional unit sphere S 3 . The Laplace–Beltrami operator on S 3 in spherical coordinates is given by D
1 @2 @ C C cot @ 2 @ sin2
@2 @2 @2 C 2 cos 2 2 @ @ @@
:
In this case R D 1 and the heat kernel becomes K.x0 ; xI t/ D
t 2 1 e e 4t ; 3=2 .4 t/ sin
(3.7.33)
where D dS 3 .x0 ; x/. This was obtained for S U.2/ by Schulman [101], who also conjectured that this formula works in general for Lie groups. The three-dimensional Euclidean space. In the case of the Laplace operator D
@2 @2 @2 C C ; @x12 @x22 @x32
making the curvature R D 0 in (3.7.32) yields the familiar formula for the Euclidean heat kernel jxx0 j2 1 4t K.x0 ; xI t/ D e : .4 t/3=2
3.8 Heat Kernel at the Cut-Locus The point x belongs to the cut-locus of x0 if there is more than one geodesic between the points x0 and x in time t, and this number is finite. In this case the heat propagates from x0 to x in more than one way, each geodesic having its own contribution
3.10 Heat Kernel on the Half-Line
47
toward the heat kernel. If there are k geodesics between x0 and x, parameterized by Œ0; t, with the corresponding volume element Vk and classical action Sk , then the formula for the heat kernel is the sum of all contributions: K.x0 ; xI t/ D
k X
Vk .t; x0 ; x/e Sk :
(3.8.34)
j D1
The above sum has only one term in the case of elliptic operators. In the case of sub-elliptic operators the sum may become an infinite series, as in the case of the Grushin operator.
3.9 Heat Kernel at the Conjugate Locus If the points x0 and x are conjugate, there is a smooth variation with geodesics of the same length with fixed endpoints x0 and x. Consider the space x0 ;xIt D fc W Œ0; t ! M I c.0/ D x0 ; c.t/ D x; c geodesicg: Each element of the above space contributes to the propagation of heat between x0 and x. In this case the sum (3.8.34) becomes an integral over x0 ;xIt : Z
e S./ d ./;
K.x0 ; xI t/ D x0 ;xIt
where d ./ is the volume measure obtained by a limit process from Vk . This recovers a refined concept of path integral. In the particular case of homogeneous spaces, the volume element depends only on the Riemannian distance between x0 and x and does not depend in an explicit way on the end points. This can be seen in the cases of the Euclidean space and the three-dimensional hyperbolic space.
3.10 Heat Kernel on the Half-Line We interested in finding the heat kernel of the Laplacian
1 d2 2 dx2
on the domain Œ0; 1/,
D 0. This means that no heat can leak subject to the boundary condition across the point x D 0. Let x0 ; x > 0. The heat can travel from the point x0 to the point x in time t in two different ways: @ u.t; 0/ @x
On the shortest path between the points, of length d1 D jx x0 j, given by x.s/ D x0 C
.x x0 /s ; t
s 2 Œ0; t:
48
3 The Geometric Method
On a piecewise path that hits the origin at time and then is reflected toward x: .1 s /x0 ; 0 s ; x.s/ D x.s /=.t /; s t: of length d2 D x0 C x. The heat kernel is the amount of heat received by x at time t, which is the sum of the aforementioned amounts of heat: K.x0 ; xI t/ D p D p
d2 1
1
e 2t C p
2 t 1
e
2 t
.xx0 /2 2t
d2 2
1 2 t
e 2t
Cp
1 2 t
e
.xCx0 /2 2t
:
The following result will be needed in the next two applications. p Lemma 3.10.1. For any 2 R, T > 0, i D 1, we have X
1
2
e 2t .2nT C/ C
n1
X
1
2
2
e 2t .2nT / D 2e 2t
n1
X
e
2n2 T 2 t
cos
n1
2nT i : t
Proof. Expanding the binomial in the exponent and using Euler’s formula, we have X
1
2
e 2t .2nT C/ C
n1
D
X
X n1
e
2 2t
X
e
2 2 4n2tT
2 2t
X
e
2 2 2n t T
C
X
1
e
2nT t
cos
n1
Ce
2 T 2 C 2 4nT /
e 2t .4n
n1
n1
D 2e
2
1 2t .4n2 T 2 C 2 C4nT /
n1
De
1
e 2t .2nT /
2nT t
2
D 2e 2t
X
e
2n2 T 2 t
cosh
n1
2nT t
2nT i : t
3.11 Heat Kernel on S 1 Let X0 and X be two points on the unit circle S 1 with angular arguments s0 and s. The heat kernel K.X0 ; X I t/ depends on all geodesics that join the points X0 and X . The square of the lengths of both clockwise and counterclockwise geodesics joining X0 and X are `20 D .ss0 /2 ;
2 `2n D 2nC.ss0 / ;
2 `2n D 2n.ss0 / ;
n 2 N:
3.12 Heat Kernel on the Segment Œ0; T
49
Since the unitpcircle is locally Euclidean and one-dimensional, the volume element is V .t/ D 1= 2 t, and the heat kernel has the following expression: K.X0 ; X I t/ D 0 1
@e Dp 2 t
X
`2 k
V .t/e 2t D p
n2Z .ss /2 2t0
C
X
e
1 2t
1 X
1 2 t
1
e 2t
2
2nC.ss0 /
nD1
2nC.ss0 /
2
C
n1
X
e
1 2t
2n.ss0 /
2
1 A:
n1
Applying Lemma 3.10.1 with T D and D s s0 , we obtain 0 1 K.X0 ; X I t/ D p e 2 t
.ss /2 2t0
@1 C 2
X
e
2 2 2 t n
n1
1 2n .s s0 /i A : cos t
Choosing z D .s s0 /i=t and D 2=t in the definition of the theta-function 3 .zji / D 1 C 2
X
e n
2
cos.2nz/
n1
[see (1.5.4)], we arrive at the following formula for the heat kernel on S 1 : K.X0 ; X I t/ D p
1 2 t
e
.ss0 /2 2t
3
.s s0 /i ˇˇ 2 i : ˇ t t
(3.11.35)
Since there is no generating formula for 3 , this shows that even in the simplest compact case, the heat kernel does not have a neat formula. For an approach using eigenfunctions expansion and the Poisson summation formula, see Sect. 6.4.
3.12 Heat Kernel on the Segment Œ0; T The heat kernel of concern here is the one corresponding to the Neumann boundary @ @ conditions @x u.t; 0/ D @x u.t; T / D 0. This corresponds to the case when x D 0 and x D T are perfectly insulated reflecting walls. The distances traveled by heat from x0 to x in a direct way or using reflections in the walls are x x0 ;
x C x0 ;
2nT C x x0 ;
2nT x C x0 ;
2nT C x C x0 ;
2nT x x0 ; : : : ;
50
3 The Geometric Method
with n D 1; 2; : : : . These are the distances from x0 to the images of x seen in two parallel plane mirrors situated at coordinates 0 and T . Each of these distances has a contribution to the heat kernel as follows: X `2 n V .t/e 2t K.x0 ; xI t/ D n0
D p
2
1 2 t
4e
.xx0 /2 2t
C
X
C e
.xCx0 /2 2t
1
2
e 2t .2nT Cxx0 / C
n1
C
X
X
1
e 2t .2nT xCx0 /
2
n1
e
1 2t
.2nT CxCx0
n1
/2
C
X
3 e
1 2t
.2nT xx0
/2
5:
n1
Lemma 3.10.1 enables us to write the previous expression as 3 2 X 2n2 T 2 .xx0 /2 1 2nT K.x0 ; xI t/ D p e 2t 41 C 2 e t cos .x x0 /i 5 t 2 t n1 3 2 X 2n2 T 2 .xCx0 /2 2nT 1 e 2t 41 C 2 e t cos .x C x0 /i 5 ; Cp t 2 t n1 which, after using the definition of the 3 -function, can be expressed as ˇ 2T 2 i .xx0 /2 1 T ˇ e 2t 3 K.x0 ; xI t/ D p (3.12.36) .x x0 /i ˇ t t 2 t ˇ 2T 2 i .xCx0 /2 T 1 ˇ : (3.12.37) .x C x0 /i ˇ Cp e 2t 3 t t 2 t
3.13 Heat Kernel on the Cylinder The cylinder is two-dimensional and locally Euclidean, so the volume element is V .t/ D 21 t . Let .x0 ; y0 / and .x; y/ be two points on the cylinder. The geodesic joining the points is a spiral which makes a constant angle with the vertical. If the geodesic winds n times, the vertical distance is the sum of n equal increments plus a remainder amount (see Fig. 3.2): y y0 D 2n tan C .x x0 / tan : This shows that the angle D n must be quantized by the relation tan n D
y y0 ; 2n C .x x0 /
n 2 Z:
3.13 Heat Kernel on the Cylinder
51
Fig. 3.2 The geodesic between the points .x0 ; y0 / and .x; y/ on the cylinder R S1 tan
(x0 , y0)
x0
x0 mod
By convention, we consider the integer n positive if the winding is counterclockwise and negative otherwise. The length of the geodesic is the sum of n C 1 oblique segments: 2 n x x0 `n D C : cos n cos n Expressing cosine in terms of tangent function and using the aforementioned formula yields q p 2 `n D D 1 C tan2 n .2n C x x0 / D 2n C .x x0 / C .y y0 /2 : Then the heat kernel is given by the contribution of all geodesics from .x0 ; y0 / and .x; y/: K .x0 ; y0 /; .x; y/I t 1 X 2t1 .2nCxx0 /2 C.yy0 /2 1 X `2n e 2t D e D 2 t n2Z 2 t n2Z 1 .xx0 /2 C.yy0 /2 X 2t1 4n2 2 C4n.xx0 / 2t e D e 2 t n2Z 2n.x x0 /i 1 .xx0 /2 C.yy0 /2 X 2n2 2 2t t e cos e D 2 t t n2Z 2n.x x0 /i i .xx0 /2 C.yy0 /2 X 2n2 2 2t t e sin e C 2 t t n2Z 1 0 2 C.yy /2 X 2 2 .xx / 2n.x x0 /i A 1 2n 0 0 @1 C 2 2t e t cos e D 2 t t n1 1 .xx0 /2 C.yy0 /2 2t e 3 D 2 t
.x x0 /i ˇˇ 2 i : ˇ t t
52
3 The Geometric Method
It is worth noting that the previous heat kernel is the direct product of the heat kernels on R and S 1 . This formula could have been written directly since the cylinder is the product of these two manifolds.
3.14 Operators with Potentials This is an extension of the method described in the previous sections to operators that have a potential. We start the presentation with the heat kernel of the operator 1 d2 C U.x/, where U.x/ is a function of the variable x, called the potential. Since 2 dx the heat kernel will depend on the classical action between any two given points, in order to obtain closed-form formulas, we need to consider only those potentials U.x/ for which the classical action can be computed explicitly. This will occur in several cases, for instance, in the case of linear and quadratic potentials. In these cases there is a unique solution joining any two given points x0 and x within time t. The classical action cannot be obtained explicitly for potentials of the type U.x/ D a2 x m , with m 3. d 2 Consider the operator L D 12 dx C U.x/ with smooth potential U.x/. We associate the Hamiltonian function H.p; x/ D
1 2 p C U.x/: 2
(3.14.38)
Hamilton’s equations are xP D Hp D p; pP D Hx D U 0 .x/; and hence xR D pP D U 0 .x/: For any two given points x0 and x, the classical path joining them is obtained by solving the equation 8 0 ˆ ˆ <xR D U .x/; (3.14.39) x.0/ D x0 ; ˆ ˆ :x.t/ D x: We shall assume the potential U.x/ is such that the previous system has a unique solution. Since the Hamiltonian (3.14.38) does not depend explicitly on the variable t, it will be preserved along the solutions of (3.14.39): H D
xP 2 C U.x/ D E; 2
(3.14.40)
where E D E.x0 ; x; t/ is the constant of energy. Hence x.s/ verifies the integral equation Z x.s/ dw D ˙s; p 2E 2U.w/ x0
3.14 Operators with Potentials
53
where the positive (negative) sign is taken on the right-hand side if x > x0 .x < x0 /. The energy E D E.x0 ; x; t/ satisfies the equation Z x dw p D ˙t; 2E 2U.w/ x0 with the same sign convention. The action S is verifying the Hamilton–Jacobi equation @t S D E.x0 ; x; t/. Since along the solutions we have p D @x S , using xP D p, we get xP D @x S , and hence (3.14.40) becomes .@x S /2 D 2E 2U.x/:
(3.14.41)
We shall look for a fundamental solution of the type K D V .t; x/e S , where S D Scl is the classical action. A computation provides 0 V CE ; @t K D K V @x V C @x S ; @x K D K V 2 @ @x V xV 2 @x S C .@x S /2 @2x S : @2x K D K V V Let P D @t 12 @2x U.x/. Using the above formulas, we obtain PK D K DK
1 @t V 1 @2x V @x V @x S 1 CE C .@x S /2 C @2x S U.x/ V 2 V V 2 2 ! 1 2 @t V 2 @x V C @x V @x S 1 C @2x S ; V 2
where we used (3.14.41). Hence PK D 0 if and only if V satisfies the generalized transport equation 1 1 @t V @2x V C @x S @x V C @2x S V D 0: 2 2
(3.14.42)
Using that xP D @x S , we have d P x V D @t V C @x S @x V; V .t; x/ D @t V C x@ dt and hence (3.14.42) becomes d 1 1 V .t; x/ D @2x V .t; x/ @2x S V .t; x/: dt 2 2
(3.14.43)
54
3 The Geometric Method
If @2x S depends on t only, then it makes sense to look for a function V D V .t/ satisfying the simplified transport equation 1 @t V C @2x S V D 0: 2
(3.14.44)
The volume function V .t; x/ should satisfy the initial condition Z lim V .t; x/e Scl dx D 1: t &0
(3.14.45)
We conclude the above calculations with the following method for finding explicit formulas for heat kernels: 1. Assume the boundary value problem (3.14.39) satisfies the uniqueness condition; let x.s/ be its solution. 2. Let S D Scl .x0 ; xI t/ be the classical action along solution x.s/ between x0 and x in time t. 3. Consider the solution V .t; x/ of the transport equation (3.14.42) satisfying (3.14.45). 4. Then K.x0 ; xI t/ D V .t; x/e S is the heat kernel for the operator 12 @2x C U.x/. Unfortunately, there are potential functions U.x/ for which the boundary value problem (3.14.39) has a unique solution. This happens, for instance, in the case when U.x/ D x 4 ; see Calin and Chang [24]. Even if the solution is unique and the action function is found, the transport equation might be difficult to solve. It is worth noting that the potential function U.x/ does not appear explicitly in the expression of the generalized transport equation. It is even more interesting to observe that (3.14.42) is a particular form of the multidimensional generalized transport equation (3.4.11) with .aij / D 1 (one-dimensional matrix). In the following se shall consider the case of linear and quadratic potentials, for which all the computations can be done explicitly.
3.15 The Linear Potential Let U.x/ D ax, with a a real constant. The operator is L D associated Hamiltonian function is H.p; x/ D
1 2 p ax: 2
The Hamiltonian system becomes xP D Hp D p; pP D Hx D a:
1 d2 2 dx
ax and its
3.15 The Linear Potential
55
The associated Lagrangian function is L.x; x/ P D 12 xP 2 C ax. The Euler–Lagrange equation xR D a with the boundary conditions x.0/ D x0 , x.t/ D x has the unique solution 1 x.s/ D as 2 C bs C x0 ; (3.15.46) 2 with x x0 at : (3.15.47) bD t 2 The Lagrangian along the solution (7.9.43) is 1 P 2 C ax.s/ L x.s/; x.s/ P D x.s/ 2 1 1 2 D .as C b/2 C a as C bs C x0 2 2 1 D a2 s 2 C 2abs C ax0 C b 2 : 2 The classical action is obtained by integrating the Lagrangian along the solution: Z
t
S.x0 ; xI t/ D
L x.s/; x.s/ P ds
0
1 1 D a2 t 3 C abt 2 C ax0 t C b 2 t: 3 2 Substituting b from (7.10.67) yields 1 2 3 1 x x0 at 2 at 2 x x0 S.x0 ; xI t/ D a t C at C ax0 t C t 3 t 2 2 t 2 .x x0 /2 1 2 3 1 D (3.15.48) C a.x C x0 /t a t : 2t 2 24 There is an alternate method for finding the action using the Hamilton–Jacobi equation. For this we need first to find the energy along the solution as a function of the boundary points. In our case (3.14.40) becomes 1 x.s/ P 2 ax.s/ D E: 2 Using (7.9.43) yields ED
1 2 1 .as C b/2 a as C bs C x0 ; 2 2
8s 2 Œ0; t:
56
3 The Geometric Method
Taking the particular value s D 0 yields 1 2 b ax0 2 1 x x0 at 2 D ax0 2 t 2
ED
D
.x x0 /2 1 a2 t 2 a.x C x : / C 0 2t 2 2 8
It is not hard to check that the Hamilton–Jacobi equation @t S D E 1 a2 t 2 .x x0 /2 C / a.x C x ; D 0 2t 2 2 8 with the initial condition Sjt D0 D 0, has the solution S.x0 ; xI t/ D
1 .x x0 /2 1 C a.x C x0 /t a2 t 3 : 2t 2 24
(3.15.49)
In the following we shall find the volume function V . Since 12 @2x S D V D V .t/ and the transport equation is V 0 .t/ C
1 V .t/ D 0; 2t
with the solution V .t/ D
c t 1=2
:
The constant c can be found using (3.14.45): Z Z 1 D lim K.x0 ; xI t/ dx D lim V .t/ e S.x0 ;xIt / dx t &0
D lim
c
D lim
c
e
t &0 t 1=2
t &0
D lim
t
t &0
Z
.xx0 /2 1 2 3 1 2t 2 a.xCx0 /t C 24 a t
1
1
2t3
1
1
2t3
e 2 ax0 t C 24 a 1=2 c
t &0 t 1=2
e 2 ax0 t C 24 a
Z
e
.xx0 /2 2t
1
dx
1 2 a.xx0 /t
.2 t/1=2 e 8 a
dx
2t 3
D c.2/1=2 ; and hence c D .2/1=2 and the volume function is V .t/ D 1=.2 t/1=2 .
1 , 2t
then
3.15 The Linear Potential
57
We have arrived at the following result: Theorem 3.15.1. Let a 2 R. The heat kernel of the operator L D given by K.x; x0 ; t/ D p
1 2 t
e
.xx0 /2 2t
1 2 3 1 2 a.xCx0 /t C 24 a t
1 d2 2 dx
ax is
; t > 0:
Remark 3.15.2. In the case when a D 0 we recover the well-known heat kernel of 1 2 @ , 2 x
which is
p1 2 t
e
.xx0 /2 2t
.
Without muchPeffort, the previous result can be generalized to the n-dimensional P operator L D 12 niD1 @2xi niD1 ai xi . In this case the Hamiltonian function is H.p; x/ D
1X 2 X pi ai xi ; 2 n
n
i D1
i D1
and the Hamiltonian equations are xP j D pj ; pPj D aj ;
j D 1; : : : ; n:
The solution joining the points x0 and x is x j .s/ D
1 j aj s 2 C bj s C x0 ; 2
(3.15.50)
with bj D
x j x0j aj t ; t 2
j D 1; : : : ; n:
(3.15.51)
The Lagrangian L.x; x/ P D
1X 2 X 1 jxP i j2 C ha; xi D xP i C ai xi 2 2 n
n
i D1
i D1
evaluated along the previous solution is 1 L x.s/; x.s/ P D jaj2 s 2 C 2ha; bis C ha; x0 i C jbj2 : 2 Integrating the Lagrangian along the solution yields the following classical action: S.x0 ; xI t/ D
1 jx x0 j2 1 C ha; x C x0 it jaj2 t 3 : 2t 2 24
(3.15.52)
58
3 The Geometric Method
Similar calculations show that the generalized transport equation in this case is ! n n 1X 2 1 X 2 @t V @xi V C hrS; rV i C @xi S V D 0: 2 2 i D1 i D1 P n does not depend on x, it makes sense to look for a Since 12 niD1 @2xi S D 2t volume function of the form V D V .t/, in which case the previous equation takes the simplified form n V 0 .t/ C V .t/ D 0: 2t The solution is V .t/ D ct n=2 . Using a similar method as before, we find c D .2/n=2 . We conclude with the following result: P Theorem 3.15.3. Let a 2 Rn . The heat kernel of the operator L D 12 niD1 @2xi Pn i D1 ai xi is given by K.x; x0 ; t/ D
jxx0 j2 1 1 2 3 1 2t 2 ha; xCx0 it C 24 jaj t ; t > 0: e .2 t/n=2
3.16 The Quadratic Potential 2
d We consider the operator L D 12 . dx a2 x 2 /, with a constant, called the Hermite operator. The associated Hamiltonian is given by H.p; x/ D 12 p 2 12 a2 x 2 and the Hamiltonian system of equations is
xP D Hp D p; pP D Hx D a2 x: The geodesic between x0 and x within time t satisfies xR D a2 x; x.0/ D x0 ; x.t/ D x:
(3.16.53)
x x0 cosh.at/ sinh.as/ C x0 cosh.as/: sinh.at/
(3.16.54)
The solution is x.s/ D
The Lagrangian associated with the previous Hamiltonian is L.x; x/ P D p xP H D
1 2 1 2 2 xP C a x : 2 2
(3.16.55)
3.16 The Quadratic Potential
59
Integrating the solution (7.9.51) along the Lagrangian (7.9.52) and performing the same computations as in Sect. 7.8 yields S.x0 ; x; t/ D
2 a .x C x02 / cosh.at/ 2xx0 : 2 sinh.at/
(3.16.56)
There is an alternate way of finding the action using the Hamilton–Jacobi equation. The conservation of energy law for (7.9.50) is 1 2 1 xP .s/ a2 x 2 .s/ D E; 2 2 where E is the energy constant. This can also be written as p dx dx D 2E C a2 x 2 H) p D ds: ds 2E C a2 x 2 Integrating between s D 0 and s D t, with x.0/ D x0 and x.t/ D x, and solving for the energy yields a2 x 2 C x02 2xx0 cosh.at/ a2 .x x0 cosh.at//2 2 2 2E D a x0 D : sinh.at/2 sinh.at/2 The Hamilton–Jacobi equation becomes @t S D H.rS / D E a2 x 2 C x02 2xx0 cosh.at/ D 2 sinh.at/2
@ a 2 axx0 2 D : .x C x0 / coth.at/ @t 2 sinh.at/ Hence
a 2xx0 2 2 .x C x0 / coth.at/ S.x0 ; x; t/ D 2 sinh.at/ D
2
1 a x C x02 cosh.at/ 2xx0 : 2 sinh.at/
(3.16.57)
Since @2x S D a coth.at/ is a function of t only, the transport equation becomes 1 V 0 .t/ C a coth.at/V .t/ D 0: 2
60
3 The Geometric Method
Integrating, we obtain the solution c V .t/ D p ; sinh.at/ where c is a constant which will be determined later. Using K D V .t/e S , we get the fundamental solution K.x0 ; x; t/ D p
c sinh.at/
a
1
e 2 sinh.at / Œ.x
2 Cx 2 / cosh.at /2xx 0 0
:
In order to determine the constant c we write s 1 at 2 2 c at e 2t sinh.at / Œ.x Cx0 / cosh.at /2xx0 : K.x0 ; x; t/ D p at sinh.at/ Since at= sinh.at/ ! 1, for a ! 0, we can write 1 c 2 K.x0 ; x; t/ p e 2t .xx0 / : at
By comparison with the fundamental p solution for the usual heat operator p 1 2 .xx / 0 , we obtain c D a=2. To conclude, we have the following 1= 2 t e 2t result. 2
d Theorem 3.16.1. Let a be a constant. The heat kernel for the operator 12 . dx 2 2 a x / is s 1 at 2 2 1 at e 2t sinh.at / Œ.x Cx0 / cosh.at /2xx0 ; t > 0: K.x0 ; x; t/ D p 2 t sinh.at/
Using the substitutions a D i ˛ and cosh.i ˛t/ D cos.˛t/ and sinh.2i ˛t/ D i sin.2˛t/, we get 2
d Corollary 3.16.2. Let ˛ be a constant. The heat kernel for the operator 12 . dx C 2 2 ˛ x / is s 1 2˛t 2 2˛t 2 1 K.x0 ; x; t/ D p e 2t sin.2˛t / Œ.x Cx0 / cos.2˛t /2xx0 ; t > 0: 2 t sin.2˛t/
3.17 The Operator
1 2
P
@x2 i ˙ 12 a2 jxj2
Consider the operator 1 1 1 1 n a2 jxj2 D .@2x1 C C @2xn / a2 .x12 C C xn2 /; a 0: 2 2 2 2
3.17 The Operator
1 2
P
@2xi ˙ 12 a2 jxj2
61
The associated Hamiltonian is H D
1 2 1 .1 C C n2 / a2 .x12 C C xn2 /; 2 2
with the Hamiltonian system ( xP j D Hj D j ; Pj D Hxj D a2 xj ; j D 1; : : : ; n: The geodesic x.s/ starting at x0 D .x10 ; : : : ; xn0 / and having the final point x D .x1 ; : : : ; xn / satisfies the equations 8 2 ˆ ˆ <xR j D a xj ; xj .0/ D xj0 ; ˆ ˆ :x .t/ D x ; j D 1; : : : ; n: j
j
As in the one-dimensional case, we have the law of conservation of energy xP j2 .s/ a2 xj2 .s/ D 2Ej ; j D 1; : : : ; n; where Ej is the energy constant for the j th component. The total energy, which is the Hamiltonian, is given by H D
n X 1
1 xP j2 a2 xj2 2 2
j D1
D E1 C C En D E(constant):
Since the energy for the one-dimensional case is Ej D
a2 Œxj2 C .xj0 /2 2xj xj0 cosh.at/ 2 sinh2 .at/
;
we get the total energy H DED
n X j D1
where jxj2 D
Pn
j D1
Ej D
a2 Œjxj2 C jx0 j2 2hx; x0 i cosh.at/ ; 2 sinh2 .at/
xj2 and hx; x0 i D
Pn
0 j D1 xj xj .
Computing the classical action. The action between x0 and x within time t satisfies the equation @t@ S D E or @ a2 Œjxj2 C jx0 j2 2hx; x0 i cosh.at/ S D @t 2 sinh2 .at/
ahx; x0 i @ a 2 2 .jxj C jx0 j / coth.at/ D @t 2 sinh.at/
62
3 The Geometric Method
and hence Scl .x0 ; xI t/ D
a 1 .jxj2 C jx0 j2 / cosh.at/ 2hx; x0 i : 2 sinh.at/
(3.17.58)
We are looking for a kernel of the form K.x0 ; x; t/ D V .t/e kScl .x0 ;x;t / ;
k 2 R:
(3.17.59)
Since the Lagrangian is at most quadratically in x and x, P the function V .t/ is given by the van Vleck formula (see also Chap. 7) s 1 @2 Scl V .t/ D det : 2 @x @x0 Since in this case a @2 Scl D In ; @x @x0 sinh.at/ we obtain
V .t/ D
a 2 sinh.at/
n=2 :
Hence the heat kernel is K.x0 ; x; t/ D
a 2 sinh.at/
n=2
a
1
e 2 sinh.at /
.jxj2 Cjx0 j2 / cosh.at /2hx;x0 i
for t > 0. In a similar way, one can find the heat kernel for the operator 12 n C 12 a2 jxj2 . Formally, this reduces to changing sinh into sin and cosh into cos in the previous expression. An alternate method. Another method for finding the heat kernel of 1X 2 @xi 2ajxj2 2 n
LD
i D1
is described next by following reference [11]. Using the rotational symmetry of the operator L, it makes sense to state the following. Anzatz: The heat kernel of L starting at the origin is of the form 1
2
K.x; aI t/ D a .t/e 2 ˛a .t /jxj ;
t > 0:
3.17 The Operator
1 2
P
@2xi ˙ 12 a2 jxj2
63
Substituting in equation .@t L/K D 0 yields 0 2 1 2 .t/ n 0 2 ˛ .t/ C ˛ .t/ a jxj K D 0: ˛.t/ C .t/ 2 2 We shall look for functions and ˛ such that ˛ 0 .t/ D a2 ˛ 2 .t/; n 0 .t/ D ˛.t/: .t/ 2 Using the substitution ˛ D ˇ 0 =ˇ yields the equation ˇ 00 D a2 ˇ, with the solution ˇ.t/ D A cosh.at/ C B sinh.at/: Using limt &0C K.x; aI t/ D ıx yields 1 D lim ˛.t/ D t &0C
ˇ 0 .0/ 2aB D ; ˇ.0/ A
so A D 0. Hence ˛.t/ D a coth.at/. Integrating in
0 .t / .t /
D n2 a coth.at/ yields .t/ D
cn : sinh.at/n=2
The constant cn may be obtained from the condition Z
lim
t &0C
1 cn 2 e 2 ˛a .t /jxj dx D 1 ” sinh.at/n=2 cn 2 n=2 lim D1” ˛.t/ t &0C sinh.at/n=2 n=2 2 lim D cn1 ” t &0C a cosh.at/ a n=2 : cn D 2
Hence we obtain the heat kernel of L: 1
2
K.x; aI t/ D a .t/e 2 ˛a .t /jxj n=2 a 1 2 D e 2 a coth.at /jxj ; 2 sinh.at/
t > 0:
64
3 The Geometric Method
3.18 The Operator L D 12 @x2 C
x2
The operator is defined on the domain fx > 0g. This is an example where the transport equation reduces to a modified Bessel equation. This operator has been considered from the geometric point of view in reference [26]. First we find the classical action. The associated Hamiltonian is H D 12 p 2 C x 2 , a real constant. The geodesic joining the points x0 and x within time t satisfies the Euler–Lagrange equation associated with the Lagrangian L D 12 xP 2 x 2 : 2 ; x3 x.0/ D x0 ; x.t/ D x: xR D
Since the regions fx < 0g and fx > 0g are separated, in order to have connectivity, we have to assume that either x0 ; x > 0 or x0 ; x < 0. We can show that the energy is a first integral of motion, so that 1 2 xP .s/ C 2 D E; 2 x .s/
s 2 Œ0; t:
Under the assumption x0 ; x > 0, the previous relation becomes x.s/x.s/ P D
p
2Ex 2 .s/ 2 :
Let u.s/ D x 2 .s/, u0 D x02 , ut D x 2 . Then u.s/ verifies the ODE p uP D 2 2Eu 2 ; u.0/ D u0 ;
u.t/ D ut :
Integrating yields Z
ut
du p D 2t ” 2Eu 2 u0 p p 2Eut 2 2Eu0 2 D 2Et: Eliminating the square roots, we obtain 2 .u0 C ut / 2Et 2 D 4.u0 ut 2 t 2 /;
(3.18.60)
3.18 The Operator L D 12 @2x C
x2
65
assuming the following condition: <
x02 x 2 : 2t 2
Solving for E in (3.18.60) yields q x2 C x2 ED 0 2 2t
x02 x 2 2 t 2 t2
:
(3.18.61)
The classical action Scl satisfies the following Hamilton–Jacobi equation: @t Scl D E; with E given by (3.18.61). We can write Scl D S0 C S1 , where x02 C x 2 x2 C x2 H) S0 D 0 ; 2 2t 2t q x02 x 2 2 t 2 @t S1 D : t2
@t S0 D
We shall solve (3.18.63) as a homogeneous equation. Let D S1 . x0t x /. Then p 1 2 2 d S2 . / D : d
2 With the substitution D
p1 2
(3.18.62)
(3.18.63) t , x0 x
and S2 . / D
sin , integrating yields
Z p p Z 1 2 2 d D 2 cot2 d S2 . / D
2 p Z p D 2 1 cot0 d D 2 . C cot / ) ( p 2 p p 1 2
D 2 sin1 . 2 / C p 2
and hence S1 .x0 ; x; / D
q 2 x02 x 2 t 2 2t
p p t 2 sin1 2 : x0 x
(3.18.64)
66
3 The Geometric Method
From (3.18.64) and (3.18.62), we obtain the classical action
Scl .x0 ; x; t/ D
x02
Cx 2t 2
q 2 x02 x 2 t 2 2t
p p t 1 2 : (3.18.65) 2 sin x0 x
By the general theory, or by a direct computation, the classical action Scl satisfies the Hamilton–Jacobi equation 1 @t Scl C .@x Scl /2 C 2 D 0: 2 x
(3.18.66)
The transport equation. We shall assume the heat kernel for L of the type K.x0 ; x; t/ D V .x; t/e S.x0 ;x;t / : Then a computation shows @t K D e S .@t V V @t S / ; @2x K D e S @2x V 2@x V @x S C V .@x S /2 V .@2x S / ; and hence 1 .@t L/K D @t @2x 2 K 2 x ( " # 1 Œby .3:18:66/ D e S V @t S C .@x S /2 C 2 2 ƒ‚ x… „ D0
) 1 2 1 2 C@t V @x V C @x V @x S C V .@x S / : 2 2 We shall ask V to satisfy the following transport equation: 1 1 @t V @2x V C @x V @x S C V .@2x S / D 0: 2 2
(3.18.67)
Equation (3.18.67) might be hard to solve since the action S and its derivatives are complicated. In the following we shall consider a shortcut for these computations. We note that the action is the sum S D S0 C S1 , where the term S1 is a function of x0t x . Then e S D W
x x x02 Cx2 0 e 2t : t
3.18 The Operator L D 12 @2x C
x2
67
Then it makes sense now to look for a heat kernel of the type K.x0 ; x; t/ D V .x; t/e S0 D V .x; t/e where V .x; t/ D
1 p Z. x0t x / t
x 2 Cx 2 0 2t
;
satisfies the extended transport equation
1 1 1 @t V @2x V C @x V @x S0 C V .@2x S0 / V @t S0 C .@x S0 /2 C 2 D 0: 2 2 2 x (3.18.68) In the following we shall solve (3.18.68). Let D
x0 x . t
Then we have
1
V D t 2 Z. /; 1 3 0 2 Z. / C Z . / ; @t V D t 2 3
@x V D t 2 Z 0 . / x0 ; 3
@2x V D t 2 Z 00 . /
x02 : t
Since @t S D
x02 C x 2 ; 2t 2
@x S D
x ; t
@2x S D
1 ; t
(3.18.68) becomes, after cancelations,
x 2 1 1 x0 2 1 1 0 t 2 Z. / 2 D 0: t 2 Z 00 . / 2 t x 2 t 1
Multiplying by 2x 2 t 2 yields
2 Z 00 . / C Z. /Œ2 2 D 0: 1
Let U. / D 2 Z. /. A computation shows 1 3 3 1
2 Z. / 2 Z 0 . / C 2 Z 00 . /; 4 1 1 1
U 0 . / D 2 Z. / C 2 Z 0 . /; 2
2 U 00 . / D
(3.18.69)
68
3 The Geometric Method
and using (3.18.69), we have 1 2 00
U . / C U . / D
Z . / C Z. / 4 1 2 1 2 Z. / 2 C Z. / D
4 1 D U. / 2 2 C : 4 2
00
0
1 2
Hence U. / satisfies the modified Bessel equation
2 U 00 C U 0 C . 2 2 /U D 0; with D
1 2
p
1 8 . The general solution can be written as a linear combination U. / D ˛I . / C ˇJ . /;
˛; ˇ 2 R;
where I . / and J are the modified Bessel functions of the first and second types. Hence the general solution of (3.18.69) is given by Z. / D
p p p
U. / D ˛ I . / C ˇ J . /;
where r I . /
1 e ; 2
r J . /
1 as ! 1I 2 e
(3.18.70)
see, for instance, [41]. Consequently, the solution of the extended transport equation (3.18.68) will be given by V .x0 ; x; t/ D t
1=2
p xx xx xx0 0 0 Z. / D ˛I
C ˇJ
; t t t
(3.18.71)
with ˛; ˇ 2 R. Theorem 3.18.1. The heat kernel for the operator L D 12 @2x C x 2 , with 0 < < 1=8 and x > 0, is K.x0 ; xI t/ D
p x0 x x0 x x02 Cx2 2t e I
; t t
t > 0;
where I is the nonsingular modified Bessel function of order D
1 2
p
1 8 .
3.18 The Operator L D 12 @2x C
x2
69
Proof. We have shown already in the previous section that .@t L/K.x0 ; x; t/ D 0;
t > 0;
with K.x0 ; x; t/ D V .x0 ; x; t/e
x 2 Cx 2 0 2t
;
and V given by (3.18.71). We need to choose the constants ˛ and ˇ such that lim K.x0 ; x; t/ D ıx0
t &0
in the distributions sense. Let K D K1 C K2 with p
xx0 xx0 x2 Cx02 2t ; e I
t t p xx0 xx0 x2 Cx02 2t : e J
K2 .x0 ; x; t/ D ˇ t t K1 .x0 ; x; t/ D ˛
Then
p lim K1 .x0 ; x; t/ D lim ˛
t &0
t &0
D ˛ lim
!1
xx0 xx0 xx0 .xx0 /2 2t e t e I
t t
p
.xx0 /2 1
I . /e lim p e 2t t &0 t
.xx0 /2 1 1 lim p e 2t D ˛p 2 t &0 t D ˛ıx0 ;
where we have used the first relation of (3.18.70). Hence we shall choose ˛ D 1. A similar computation, using the second relation of (3.18.70) yields p xx0 xx0 xx0 .xx0 /2 2t e t e J
lim K2 .x0 ; x; t/ D lim ˇ t &0 t &0 t t D ˇ lim
!1
p
.xx0 /2 1
J . /e lim p e 2t t &0 t
.xx0 /2 p 1 D ˇ lim e 2 lim p e 2t !1 t &0 2 t D 0:
Hence lim K.x0 ; x; t/ D lim K1 .x0 ; x; t/ C lim K2 .x0 ; x; t/ D ˛ıx0 ;
t &0
t &0
t &0
70
3 The Geometric Method
so we need to choose ˛ D 1. In order to find ˇ, we shall consider the limit ! 0, in which case we recover the Gaussian kernel .xx0 /2 1 p D lim K.x0 ; x; t/ e 2t &0 2 t .xx0 /2 .xx0 /2 1 p 1 p D p I1=2 . /e e 2t C ˇ p J1=2 . /e e 2t t t .xx0 /2 .xx0 /2 p 1 1 D p e 2t C ˇ p J1=2 . /e e 2t ; t 2 t
since we take D 1=2 in I . / D p
1 2
e
1 . 2 1 / 2 4
C O.1= 2 /;
and hence we need to choose ˇ D 0 in relation (3.18.71).
Chapter 4
Commuting Operators
4.1 Commuting Operators If X and Y are two vector fields, or in general, two operators which commute, i.e., X Y D YX , then it is obvious that their squares also commute. If L D 12 .X 2 C Y 2 / with ŒX 2 ; Y 2 D 0, then the problem of finding the heat kernel for L is reduced to the same problem for the operators X 2 and Y 2 , with e t .X
2 CY 2 /
2
2
D e tX e t Y :
The aforementioned formula is reminiscent from stochastic calculus, where if the operators 12 X 2 and 12 Y 2 are generators for certain independent Ito diffusion processes, then their joint density function is a product of their individual density functions. A similar formula holds in the general case of n commuting vector fields. The nontrivial problem occurs when X 2 and Y 2 do not commute, in which case the heat kernel can be obtained either by using Trotter’s formula and path integrals (see Chap. 7), or by using the geometric method (see Chap. 3). We shall provide a few examples below that can be reduced to a commuting sum of squares of two vector fields.
4.1.1 The Operator L D 12 .@x2 C @y2 / This is an elementary example of reducing the two-dimensional Laplacian problem to a one-dimensional Laplacian problem. Since the operators @2x and @2y commute, we can write .xx0 /2 .yy0 /2 1 1 1 2 1 2 e t L D e 2 t @x e 2 t @y D p p e 2t e 2t 2 t 2 t
D
1 .xx0 /2 C.yy0 /2 2t e : 2 t
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 4, c Springer Science+Business Media, LLC 2011
71
72
4 Commuting Operators
4.1.2 The Operator L D 12 .@x2 C y@y2 / Let .x; y/ 2 R.0; 1/. Since the operators @2x and y@2y commute, using Proposition 3.4.1, the heat kernel for L between .x0 ; 0/ and .x; y/ within time t is t t 2 t 2 2 2 K .x0 ; 0/; .x; y/I t D e t L D e 2 .@x Cy@y / D e 2 @x e 2 y@y .xx0 /2 4y 1 2 2t e e t y 2 1=2 t .2 t/ r 2 y 2t1 4y.xx0 /2 D2 e ; t 3=2
D
t > 0:
4.1.3 The Operator L D 12 .x@x2 C y@y2 / The operator is defined on the domain .x; y/ 2 .0; 1/ .0; 1/. Since the operators x@2x and y@2y commute, using Proposition 3.4.1 we have t 2 t 2 K .0; 0/; .x; y/I t D e t L D e 2 x@x e 2 y@y D
4x 2 x 4y 2 y 4xy 2 e t e t D 4 e t .xCy/ ; t2 t2 t
t > 0:
The graph of this heat kernel for t D 0:25 and x > 0, y > 0 is given by Fig. 4.1a.
4.1.4 Sum of Squares of Linear Potentials The following example will be useful in the sequel. It is obtained by applying the Fourier transform to the Heisenberg operator. Consider the noncommutative
Fig. 4.1 Heat kernel profiles for different operators. (a) 12 .x@2x C y@2y /; (b) 12 2 C 12 ! C 12 jxj2 ; (c) 12 .x 2 C y 2 /.@2x C @2y /
4.1 Commuting Operators
73
operators L1 D @x1 C ax2 and L2 D @x2 ax1 , with a 2 R constant. Construct the operator 1 2 .L C L22 / 2 1 a a2 1 D .@2x1 C @2x2 / C .x2 @x1 x1 @x2 / C .x12 C x22 / 2 2 2 2 a 1 a D 2 C ! C jxj2 ; 2 2 2
LD
a
where ! stands for the angular momentum. We note that e t 2 ! ı.0;0/ D ı.0;0/ . Since the operators ! and 12 C a2 jxj2 commute, the heat kernel for L is 1 a a2 2 t 2 2 K.x; yI t/ D e 2 .L1 CL2 / ı.0;0/ D e t 2 2 C 2 !C 2 jxj ı.0;0/ 2 1 a2 2 t 1 2 C a2 jxj2 t a 2 De e 2 ! ı.0;0/ D e t . 2 2 C 2 jxj / ı.0;0/ t
D e 2 .2 Ca D
2 jxj2 /
ı.0;0/
1 2at 2 2 1 2at e 2t sin.2at / Œ.jxj Cjyj / cos.2at /2hx; yi ; 2 t sin.2at/
t > 0:
The shape of the heat kernel for the values a D 1 and t D 0:05 is given in Fig. 4.1b.
4.1.5 The Operator L D 12 .x 2 C y 2 /.@x2 C @y2 / Assume .x; y/; .x0 ; y0 / 6D .0; 0/. Using the representation in polar coordinates x D r cos , y D r sin , the operator can be written as a sum of two operators in variables r and : .x 2 C y 2 /.@2x C @2y / D .r 2 @2r C r@r / C @2 : Using formulas which will be proved in Sect. 6.3, we have t
K.x0 ; y0 ; x; yI t/ D e 2 .x
2 Cy 2 /.@2 C@2 / x y
t
2 2
t
2
D e 2 .r @r Cr@r / e 2 @ r r 1 r 2 Cr 2 . 0 /2 1 0 0 D e 2t I0 p e 2t 2t t 2 t r r 1 2 2 2 0 e 2t r0 Cr C. 0 / ; D 1=2 .2t/3=2 I0 t
t; r; r0 > 0;
74
4 Commuting Operators
where I0 is the Bessel function of the first kind of order zero. The profile of the heat kernel with r0 D 1, 0 D 0 and t D 0:2 is depicted in Fig. 4.1c.In the view of the geometric method (see Chap. 3), the term 1=2 .2t/3=2 I0 r0t r represents the volume element, while d 2 .r0 ; 0 /; .r; / D r02 C r 2 C . 0 /2 is the square of the associated Riemannian distance.
Chapter 5
The Fourier Transform Method
The Fourier transform has been known as one of the most powerful and useful methods of finding fundamental solutions for operators with constant coefficients. Sometimes the application of a partial Fourier transform might be more useful than the full Fourier transform. In this chapter, by the application of the partial Fourier transform, we shall reduce the problem of finding the heat kernel of a complicated operator to a simpler problem involving an operator with fewer variables. After solving the problem for this simple operator, the inverse Fourier transform provides the heat kernel for the initial operator represented under an integral form. In general, this integral cannot be computed explicitly, but in certain particular cases it actually can be worked out. We shall also apply this method to some degenerate operators.
5.1 The Algorithm If L is a partial differential operator in variables .x; y/ 2 Rn R, then the heat kernel K.x0 ; y0 ; x; yI t/ D v.x; yI t/ satisfies @t v D Lxy v; lim v.x; yI t/ D ıx0 .x/ ˝ ıy0 .y/;
t &0
where ıa is the Dirac distribution centered at a. Let u.x; I t/ D Fy .v/.x; I t/, where Fy denotes the partial Fourier transform with respect to y: Z .Fy v/.x; I t/ D
e iy v.x; yI t/ dy: R
Then u satisfies the equation @t u D Px u; u.x; I 0/ D ıx0 .x/ ˝ I ;
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 5, c Springer Science+Business Media, LLC 2011
75
76
5 The Fourier Transform Method
where Px is obtained from Lxy by replacing @2y by i . In order to recover v, we need to apply the inverse Fourier transform on u: v.x; yI t/ D .F1 u/.x; I t/: The following formulas will be useful in the sequel: 1 F .Lu/; where uL .x/ D u.x/; 2 1=2 2 2 e x =.4a/ ; a > 0; F .e a /.x/ D a F 1 .u/ D
F .ı/ D I;
F .I/ D 2ı:
5.2 Heat Kernel for the Grushin Operator The two-dimensional, two-step Grushin operator is defined as the sum of squares G D
1 2 1 .X C Y 2 / D .@2x C x 2 @2y /; 2 2
where X D @x and Y D x@y . The heat kernel satisfies @t v D G v;
t > 0;
lim v.x; yI t/ D ıx0 .x/ ˝ ıy0 .y/:
t &0
Let u D Fy .v/ be the partial Fourier transform of v with respect to y. Then u satisfies @t u D
1 2 .@ u x 2 2 u/; 2 x
t > 0;
u.x; I 0/ D ıx0 .x/ ˝ I : The solution is t
2
u.x; I t/ D e 2 .@x x
22 /
.ıx0 .x/ ˝ I /:
(5.2.1)
By Theorem 3.16.1, the kernel of the Hermite operator 12 @2x 12 2 x 2 is s K.x0 ; xI t/ D p
1 2 t
t 1 t e 2t sinh.t / sinh.t/
.x 2 Cx02 / cosh. t /2xx0
;
t > 0:
5.3 Heat Kernel for @2x C x@x @y
77
Applying an inverse Fourier transform with respect to in (5.2.1) yields 1 v.x; yI t/ D 2
Z
e iy u.x; I t/ d s ( Z ) t 1 2 Cx 2 / cosh. t /2xx 1 1 t 2t .x iy 0 0 sinh.t / e d e D p 2 2 t sinh.t/ Z r
.let D t /
=t 2 2 1 e iy=t 2 sinh .x Cx0 / cosh./2xx0 d .2 t/2 sinh Z r 1 1 2 2 p D e fiy 2 .x Cx0 / coth sinh xx0 g=t d: 3=2 sinh .2/ t
D
If we let 1 xx0 ; f .x0 ; x; yI / D iy C .x 2 C x02 / coth C 2 sinh r ; V ./ D sinh
(5.2.2) (5.2.3)
we obtain the following form: v.x; yI t/ D
Z
1
.2/
p 3=2
t
e f .x0 ;x;yI/=t V ./ d;
t > 0;
where f is called the modified complex action (see also [16]), and V ./ is the volume element. Formula (5.2.2) cannot be reduced to just an elementary function, since the Grushin operator does not have a heat kernel of the function type. This result is true for most sub-elliptic operators given as a sum of squares.
5.3 Heat Kernel for @x2 C x@x @y The heat kernel of the operator L D @2x C x@x @y satisfies @t v D @2x v C x@x @y v;
t > 0;
lim v.x; yI t/ D ıx0 .x/ ˝ ıy0 .y/:
t &0
Let u D Fy .v/ be the partial Fourier transform of v with respect to y. Then u satisfies @t u D @2x i x@x u; u.x; I 0/ D ıx0 .x/ ˝ I :
t > 0;
78
5 The Fourier Transform Method
The solution is 2
u.x; I t/ D e t .@x i x@x / .ıxx0 ˝ I /:
(5.3.4)
By Theorem 10.29 of [24], the fundamental solution of @t @2x C i x@x is s K.x0 ; xI t/ D p
1 4t
t 1 t e 4t sin.t / sin.t/
.x 2 Cx02 / cos. t /2xx0
;
t > 0:
Applying the inverse Fourier transform, we obtain 1 2
Z
e iy u.x; I t/ d s ( Z ) t 1 2 Cx 2 / cos. t /2xx 1 1 t 4t .x iy 0 0 sin.t / d p e e D 2 4 t sin.t/
v.x; yI t/ D
Z r
1 1 i y e t 4t sin 4.t/3=2 Z 1 D V ./e f =t d; 4.t/3=2
D
2xx .x 2 Cx02 / cot sin 0
.let D t /
d
(5.3.5)
with r V ./ D
; sin
f .x0 ; x; yI / D iy C
1 2 2xx0 .x C x02 / cot : 4 sin
It is remarkable that heat kernel formulas (5.2.2) and (5.3.5) are similar; just the modified complex actions and the volume elements are different. This leads to the question of whether this is a universal formula for sub-elliptic-type operators.
5.4 The Formula of Beals, Gaveau and Greiner In this section we shall consider a sub-elliptic operator with one degenerate direction of the type n 1 X LD aij @xi @xj C .x/@y : 2 i;j D1
In general, heat kernels in this case can be represented as an integral, which usually does not integrate to yield a function-type kernel. The degenerate direction can be
5.4 The Formula of Beals, Gaveau and Greiner
79
eliminated by applying a partial Fourier transform in the y-variable. Let v be the fundamental solution for the sub-elliptic operator L: Lv D 0;
t > 0;
lim v.x; yI t/ D ıx ˝ ıy :
t &0
Let u.x; I t/ D Fy v.x; yI t/ be the partial Fourier transform of v with respect to y. Then u satisfies Pu D 0; t > 0; lim u.x; I t/ D ıx ˝ I ;
t &0
where PD
n 1 X aij @xi @xj i .x/ 2 i;j D1
is an elliptic operator with potential U.x/ D i .x/. If we treat as a real parameter, then u becomes the heat kernel for the operator P P. Now we P shall assume that the potential is at most quadratic, i.e., .x/ D bij xi xj C cj xj C k, so the geodesic joining any two points is unique. The geometric method or the path integral method yields in this case a fundamental solution of the type u.x; ; t/ D V .t; /e Scl; where Scl is the classical action associated with the operator P. This is obtained by integrating the Lagrangian LD
n 1 X aij pi pj C i .x/ 2 i;j D1
along the geodesics. The parameter can be considered as a Lagrange multiplier that depends on the endpoints of the geodesic. Taking the inverse Fourier transform yields the heat kernel for the sub-elliptic operator L: v.x; y; t/ D F1 u.x; I t/ D
1 2
Z
e iy V .t; /e Scl .x0 ;x;t I/ d : (5.4.6)
We shall assume that in the formula of Scl the variables t and come as a product, so we can write D t. Let g.x0 ; x; / D tScl .x0 ; x; tI /:
80
5 The Fourier Transform Method
Under this hypothesis, the variables t and will also come as a product in the formula of V . This follows from the van Vleck formula s 1 @2 Scl V D det : 2 @x @x0 Consider the function W ./ D V .t; /. Then formula (5.4.6) becomes Z 1 W ./e iy.=t / e Scl .x0 ;x;/ d.=t/ 2 Z 1 W ./e Œiyg.x0 ;x;/=t d./ D 2 t Z 1 W ./e f .x0 ;x;/=t d./; D 2 t
v.x; yI t/ D
(5.4.7)
where f .x0 ; x; / D iy C g.x0 ; x; / is called the modified complex action function. The interested reader can find more details about finding the complex action function by methods of complex Hamiltonian mechanics in Chap. 5 of [28]. Formula (5.4.7) was first shown in the case of the Heisenberg operator in the work of Beals, Gaveau and Greiner [16, 48].
5.5 Kolmogorov Operators This section deals with the case of a degenerate operator for which the integral formula (5.4.7) actually does integrate exactly, providing a closed-form formula for the heat kernel. The operator LD
1 2 @ x@y 2 x
(5.5.8)
was first studied by Kolmogorov [80], who constructed its heat kernel as a transition probability of a diffusion process. We note that the operator is degenerate in the y-direction and is not a sum of squares of vector fields. This section is based on the results obtained in Calin and Chang [27]. Let K be the heat kernel for L, and denote by u D Fy K the partial Fourier transform with respect to the y-variable. Then 1 2 @ u i xu; 2 x u.x; yI t/ D ıx0 .x/ ˝ I : @t u D
5.5 Kolmogorov Operators
81
Denote a D i , and consider it a parameter. Then u is the heat kernel for the operator PD
1 2 @ ax: 2 x
Applying Theorem 3.15.1 yields u.x0 ; xI t/ D p D p
1 2 t 1 2 t
e e
.xx0 /2 2t
1 2 3 1 2 a.xCx0 /t C 24 a t
.xx0 /2 2t
1 2 3 1 2 i .xCx0 /t 24 t
:
The heat kernel of L can be obtained by taking the inverse Fourier transform: Z
K.x0 ; y0 ; x; yI t/ D
1 2
D
1 2
D
.xx0 /2 1 p e 2t 2 2 t
e i.yy0 / u.x; I t/ d Z p
1 2 t
e
.xx0 /2 1 2 3 24 t Ci Œ.yy0 / 1 2t 2 .xCx0 /t
Z
1
e 24
d
2 t 3 Ci Œ.yy / 1 .xCx /t 0 0 2
d :
(5.5.9) Let
1 A D A.x0 ; x; y0 ; y; t/ D .y y0 / .x C x0 /t: 2 If we complete the square, the exponent becomes
h i 1 1 t3 n 24 o 1 2 3 .i /2 Ci 3 A t Ci .y y0 / .xCx0 /t D 2 t 3 Ci A D 24 2 24 24 t 12 2 12A 2 o t 3 n i C 3 A D 24 t t3 12 2 6A2 t3 i C 3 A 3 : D 24 t t
The integral can be written as an integral along a line: Z e
1 2 t 3 Ci A 24
d D e
2
6A3 t
Z
3
e
t .i 12 A/2 24 3 t
d D e
2
6A3
Z
t3 2
e 24 z d z;
t
I m zDb
(5.5.10)
82
5 The Fourier Transform Method 3 2
where b D 12 A. Let R > 0. Using that f .z/ D e t z =24 is holomorphic, the t3 integral along the rectangular contour .R; 0/; .R; b/; .R; b/; .R; 0/ vanishes, so Z e I m zDb
t3 2 24 z
(Z
R
3
d z D lim
e
R!1
t 24 y2
Z
R
Z
t3
2
e 24 .RCi/ i d
0
)
0
C
b
dy C
3
e
t 24 .RCi/2
Z
i d D
3
e
t 24 y2
b
p 2 6 dy D 3=2 ; t (5.5.11)
since
ˇ t3 ˇ t3 2 2 ˇ 24 .RCi/2 ˇ ˇe ˇ D e 24 .R / ! 0;
for R ! 1:
We also note that the computation could be short-cut by the use of Proposition 1.6.1. Using (5.5.10) and (5.5.11), (5.5.9) yields p 2 .xx0 /2 1 6A3 2 6 2t t K.x0 ; y0 ; x; yI t/ D p e t 3=2 2 2 t p 2 2 0 / 6 yy 1 .xCx /t 3 .xx 0 2 0 2t 3 t D e : t2 Theorem 5.5.1. The heat kernel of the sub-elliptic operator 12 @2x x@y is given by p
K.x0 ; y0 ; x; yI t/ D
2 0/ 6 3 .xx 2t t3 e t2
yy0 1 2 .xCx0 /t
2 ;
t > 0:
One may consider Kolmogorov operators of the type 12 @2x x m @y . The previous method provides heat kernels only for the values m D 0; 1; 2. In the following we shall treat the case m D 0. We need the following result. Lemma 5.5.2. The heat kernel for the operator 12 @2x a, with a 2 C, is .xx0 /2 1 e 2t at : u.x0 ; x; y0 ; yI t/ D p 2 t
Proof. The associated Hamiltonian and Lagrangian are H D 12 p a and L D 1 2 xP C a. The Euler–Lagrange equation is xR D 0, with the boundary conditions 2 x.0/ D x0 , x.t/ D x. The classical action is Scl D
.x x0 /2 C at: 2t
5.5 Kolmogorov Operators
83
The fundamental solution has the form s u.x0 ; x; y0 ; yI t/ D
1 @2 Scl Scl det e 2 @x @x0
D p
1
e
2 t
.xx0 /2 2t
at
:
Let K be the heat kernel for the sub-elliptic operator 12 @2x @y . Then u D Fy K is the heat kernel for the operator 12 @2x i . Applying Lemma 5.5.2 with a D i yields Z
KD
1 2
D
1 2
D
.xx0 /2 1 1 p e 2t 3=2 .2/ t
e i.yy0 / u.x0 ; x; I t/ d Z e i.yy0 / p
1 2 t
e Z
.xx0 /2 i t 2t
d
e i .yy0 t / d
.xx0 /2 1 D p e 2t ˝ ıy0 Ct .y/; 2 t
where we have used the formula for the Dirac distribution Z
1.x/ D 2ı0 .x/: e i x d D b
To conclude, we have the following result: Proposition 5.5.3. The heat kernel for the operator 12 @2x @y is K.x0 ; x; y0 ; yI t/ D p
1 2 t
e
.xx0 /2 2t
˝ ıy0 Ct .y/:
We shall next find the heat kernel for the operator 12 @2x C ix 2 @y . After applying a partial Fourier transform with respect to y, we obtain the Hermite operator 1 2 @ x 2 with the heat kernel 2 x 1 u.x0 ; x; t/ D p 2
r
1 2 a 2 a e 2 sinh.at / Œ.x Cx0 / cosh.at /2xx0 ; sinh.at/
t > 0;
84
5 The Fourier Transform Method
with a2 D 2; see Theorem 3.16.1. The inverse Fourier transform yields the heat kernel for the operator 12 @2x C ix 2 @y : 1 KD 2
Z e i.yy0 / u.x0 ; x; I t/ d
r Z 2 a 2 2 a 1 1 i.yy0 / a2 1 2 sinh.at / Œ.x Cx0 / cosh.at /2xx0 D e a da 2 .2/3=2 sinh.at/ Z a2 1 a 2 2 1 a 3=2 ae i.yy0 / 2 2 sinh.at / Œ.x Cx0 / cosh.at /2xx0 da: D 5=2 2
We note that in this case the heat kernel has an integral representation that cannot be reduced to a function. In the rest of the section we shall study the sub-elliptic operator LD
1 2 .@ C @2x2 / C .x1 x2 /@y ; 2 x1
(5.5.12)
which can be regarded as a two-dimensional Kolmogorov operator. If K denotes the heat kernel for L, then u D Fy K is the heat kernel for the operator PD
1 2 .@ C @2x2 / C i .x1 x2 /: 2 x1
(5.5.13)
u can be obtained by letting a1 D a2 D i in Theorem 3.15.3: u.x0 ; x; I t/ D D
1 jxx0 j2 1 ha;xCx0 it C 1 jaj2 t 3 2t 2 24 e 2 t 1 jxx0 j2 i h;xCx0 it 2 t 3 2t 2 12 ; e 2 t
with D .1; 1/ 2 R2 . Let A D .y y0 / 12 h; x C x0 it. The heat kernel of (5.5.12) is obtained by taking the inverse Fourier transform: KD
Fy1 u
1 D 2 D
Z e i.yy0 / u.x0 ; x; I t/ d
jxx0 j2 1 2t e .2/2 t
Z
e
2t 3 12
The second formula of Proposition 1.6.1 yields Z
2 3
e
12t Ci A
De
2
3A3 t
p 2 3 ; t 3=2
Ci A
d :
(5.5.14)
5.7 The Generalized Grushin Operator
1 2
Pn
2 2 2 iD1 @xi C 2jxj @t
85
and substituting in (5.5.14), we arrive at the following result: Theorem 5.5.4. The heat kernel of the operator (5.5.12) is p
K.x0 ; y0 ; x; yI t/ D
jxx0 j2 2 3 33 Œ.yy0 / 1 2t 2 h;xCx0 it t e 3=2 5=2 2 t
with t > 0 and vector D .1; 1/.
5.6 The Operator 12 @x2 C x@y2 The operator L D 12 @2x C x@2y resembles the Kolmogorov operator, but it has a higher derivative in y. Its heat kernel can be obtained in a similar way, but it is not of a function type. Applying the Fourier transform with respect to y, we arrive at the operator P D
1 2 @ x 2 ; 2 x
which according to Theorem 3.15.1 has the heat kernel given by u.x0 ; x; tI / D p
1 2 t
e
.xx0 /2 2t
1 4 3 2 1 2 .xCx0 /t C 24 t
:
The heat kernel for the sub-elliptic operator L is obtained by applying the inverse Fourier transform: 1 2
Z
e i.yy0 / u.x0 ; x; tI / d Z .xx0 /2 1 1 1 3 4 2 2t e i.yy0 / 2 .xCx0 /t C 24 t d ; D e 3=2 1=2 .2/ t
K.x0 ; x; y0 ; y; t/ D
which is an integral that cannot be worked out by elementary functions since the exponent is of the fourth degree.
5.7 The Generalized Grushin Operator
1 2
Pn
i D1
@x2 i C 2jxj2 @t2
Consider the sub-elliptic operator 1 Ln D 2
n X kD1
! Xi2
CY
2
1X 2 @xi C 2jxj2 @2t ; 2 n
D
i D1
(5.7.15)
86
5 The Fourier Transform Method
where Xk D @xk , Y D 2jxj@t are vector fields on RnC1 . It is worth noting that L1 D 12 @2x C 2x 2 @2t is the two-dimensional Grushin operator treated in Sect. 5.2. In the following we shall find the heat kernel for Ln . Applying the Fourier transform in t, we obtain the operator 1X 2 @xi 2 2 jxj2 : 2 n
P D
i D1
Making a D 2 in Sect. 3.17, we obtain the following heat kernel for P : u.x0 ; x; tI / D
sinh.2t/
n=2
e sinh.2t / Œ.jxj
2 Cjx
2 0 j / cosh.2 t /2hx;x0 i
;
t > 0:
The inverse Fourier transform in t provides the heat kernel for the operator Ln : KD
1 2
1 D 2
Z e i.yy0 / u.x0 ; x; tI / d Z
sinh.2t/
n=2
e i.yy0 / sinh.2t / Œ.jxj
2 Cjx j2 / cosh.2 t /2hx;x i 0 0
d :
Changing the variable D t yields Z
n=2 2 2 2 2 1 e fi.yy0 /.jxj Cjx0 j / coth.2/C sinh.2 / hx;x0 ig=t d 2 t sinh.2/ t n=2 Z 2 2 2 2 1 D e fi.yy0 /.jxj Cjx0 j / coth.2/C sinh.2 / hx;x0 ig=t d : 1C n sinh.2/ 2 .2 t/
KD
1 2
We have arrived at the following result: Theorem 5.7.1. The heat kernel for the sub-elliptic operator 1X 2 @xi C 2jxj2 @2t 2 n
Ln D
i D1
is given by K.x0 ; x; t/ D
Z
1 1C n 2
.2 t/
V ./e f .x0 ;x;t;/=t d ;
(5.7.16)
with V ./ D
2 sinh.2/
n=2 ;
f .x0 ; x; t; / D i.y y0 / C .jxj2 C jx0 j2 / coth.2/
2 hx; x0 i: sinh.2/
5.8 The Heisenberg Laplacian
87
Formula (5.7.16) agrees with the formula obtained by Beals, Gaveau and Greiner for sub-elliptic operators [16], where f and V stand for the complex modified action and volume element, respectively.
5.8 The Heisenberg Laplacian Consider on R2nC1 the following 2n vector fields: Xj D @xj C 2xnCj @t ;
XnCj D @xnCj 2xj @t ;
j D 1; : : : ; n;
called the Heisenberg vector fields. The commutation relations are ŒXnCj ; Xj D 4@t ;
ŒXj ; Xk D 0 for jk j j 6D n:
The vector fields Xk together with @t are left invariant with respect to the following group law on R2nC1 : 0 .x; t/ ı .x 0 ; t 0 / D @x C x 0 ; t C t 0 C 2
n X
1 0 .xnCj xj0 xj xnCj /A :
j D1
This group is denoted by Hn and is called the Heisenberg group. The sub-elliptic operator H
2n 1X 2 D Xk 2 kD1
is called the Heisenberg Laplacian. In this section we shall determine the heat kernel for H . A computation yields H
2n 2n n X 1X 2 1X 2 2 2 D Xk D @xk C 2jxj @t C 2 xnCj @xj xj @xnCj @t 2 2 kD1
j D1
kD1
D Ln C 2 @t ; where Ln is the Grushin operator (5.7.15) and
D
n X j D1
xnCj @xj xj @xnCj
88
5 The Fourier Transform Method
is the angular momentum operator. A computation shows that ŒLn ; D 0. Using the property of the semigroup generated by the sum of two operators which commute, we have e t H D e t Ln C2t @t D e t Ln e 2t @t : Since e 2t @t D I at x D 0, it follows that e 2t @t ıx D Iıx D ıx . Then the heat kernel of H at x0 D 0 is K.0; x; t/ D e t H ı.x/ D e t Ln e 2t @t ı.x/ D e t Ln ı.x/ D u.0; x; t/; where u is the heat kernel of the operator Ln at x0 D 0, which is obtained from Theorem 5.7.1. Making x0 D y0 D 0 and replacing n by 2n in formula (5.7.16), we obtain Theorem 5.8.1. The heat kernel for the Heisenberg Laplacian H starting from the origin is given by K.0; x; t/ D
1 .2 t/nC1
Z
V ./e f .x;t;/=t d ;
with the volume element and complex modified action given by V ./ D
2 n ; sinh.2/
f .x; t; / D i.y/ C jxj2 coth.2/:
(5.8.17)
Chapter 6
The Eigenfunction Expansion Method
Finding the heat kernel of an elliptic operator on a compact manifold using the eigenvalues method is a well-known method in mathematical physics and quantum mechanics. Roughly speaking, the eigenvalues and eigenfunctions of an operator determine its heat kernel. The formula is an infinite series that involves products of eigenfunctions; see Theorem 6.1.1. It is interesting that in several cases this series can be written as an elementary function by using the associated bilinear generating function. We shall present applications of the bilinear generating formulas of Mehler, Hille–Hardy, and Poisson. The main disadvantage of this method is that in general the bilinear generating formula is missing, in which case the series that represents the heat kernel cannot be represented as a function. This fact is not surprising, since many heat kernels are not of the function type. Since this method involves a great deal of work with special functions, we shall refer the reader to the useful book of Erd´elyi [41].
6.1 General Results Consider the self-adjoint differential operator L defined on the interval I and let fi be its eigenfunctions Lfi D i fi ;
i D 0; 1; : : : ;
with i 2 R eigenvalues. We shall R assume that ffi gi 0 is a complete orthogonal system of L2 .I / D ff W I ! C I I jf j2 < 1g, i.e., Z I
fn .x/fm .x/ dx D ınm ;
and for any 2 L2 .I / with Z I
.x/fn .x/ dx D 0;
8n D 0; 1; : : :
H) D 0:
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 6, c Springer Science+Business Media, LLC 2011
89
90
6 The Eigenfunction Expansion Method
Then for any 2 L2 .I /, we can write .x/ D
X
hfi ; ifi .x/;
i 0
Z where hfi ; i D
I
fi .x/.x/ dx.
The next result provides a formal expression for the heat kernel. Proposition 6.1.1. Let ffi gi be an orthogonal complete system of L2 .I / of real eigenfunctions of operator L. Then the heat kernel of L is given by K.x0 ; x; t/ D
X
e i t fi .x0 /fi .x/:
(6.1.1)
i 0
Proof. We shall check the definition properties of the heat kernel .@t L/K.x0 ; x; t/ D
X i e i t fi .x0 /fi .x/ e i t fi .x0 /Lfi .x/ i 0
X i e i t fi .x0 /fi .x/ e i t fi .x0 /i fi .x/ D i 0
D 0: We also need to check the following limit in the distribution sense: lim K.x0 ; x; t/ D ıx0 :
t &0
For any smooth function with compact support, we have Z lim
t &0 I
K.x0 ; x; t/.x/ dx D lim
X
t &0
D
X Z X
Z
I
I
i 0
i 0
D
e
i t
fi .x/.x/ dx fi .x0 /
fi .x/.x/ dx fi .x0 /
hfi ; ifi .x0 / D .x0 / D ıx0 ./:
i 0
Hence the heat kernel is given by formula (6.1.1).
Remark 6.1.2. If the eigenfunctions fn are complex, the heat kernel of L is given by X e i t fi .x0 /fi .x/: K.x0 ; x; t/ D i 0
6.2 Mehler’s Formula and Applications
91
6.2 Mehler’s Formula and Applications In this section we show how a bilinear generating function involving Hermite polynomials, called Mehler’s formula, is reducing the series (6.1.1) to a function in the case of the Hermite operator. We shall start first with some prerequisites regarding Hermite polynomials. The Hermite polynomial Hn .x/ of order n is defined by 2 2 Hn .x/ D .1/n e x .d=dx/n e x D 2n x n C : : : : Using repetitive integration by parts, we obtain Z
8Z m m x 2 ˆ dx D 0; if n < m, ˆ < Hn .x/.1/ .d=dx/ e R
2
R
Hn .x/Hm .x/e x dx D
ˆ ˆ :
Z
R
p 2 .Hn /.n/ .x/e x D 2n nŠ ;
if n D m.
This can be written as Z
2
R
Hn .x/Hm .x/e x dx D 2n nŠ 1=2 ınm ;
where ınm denotes the Kronecker delta function. The graphs of a few Hermite polynomials are sketched in Fig. 6.1. Hence fn .x/ D
e x
2 =2
2n=2
p
Hn .x/
nŠ 1=4
n D 0; 1; : : : ;
;
(6.2.2)
forms an orthogonal system in L2 .R/. One may show that it is also a complete system.
a
b
c
60
100
40 20 -2
-1
-2 1
-20
500
50
2
-1
1 -50 -100
2
-2
-1
1 -500
Fig. 6.1 The graphs of a few Hermite polynomials Hn .x/: .a/ n D 4; .b/ n D 5; .c/ n D 6
2
92
6 The Eigenfunction Expansion Method
Lemma 6.2.1 (Generating function formula). We have X
Hn .x/
n0
yn 2 D e 2xyy : nŠ 2
Proof. Taylor’s formula for F .y/ D e .xy/ yields F .y/ D
X F .n/ .0/ nŠ
n0
.1/n y n D e x
2
X
Hn .x/
n0
yn ; nŠ
and hence X
Hn .x/
n0
yn 2 2 2 2 D e x F .y/ D e x .xy/ D e 2xyy : nŠ
The following result deals with the eigenfunctions and eigenvalues of the Hermite operator @2x x 2 . Proposition 6.2.2. The functions given by (6.2.2) satisfy .@2x x 2 /fn .x/ D .2n C 1/fn .x/; 2
Proof. Let F .y/ D e 2xyy D 2y
P n0
Hn .x/ n y nŠ
and
n D 0; 1; : : : :
(6.2.3)
D 12 x 2 C 2xy y 2 . Then
X 1 2 1 2 @ yn Hn .x/e 2 x 2n : e D 2ye 2 x F 0 .y/ D @y nŠ n0
(6.2.4)
And applying (6.2.4), we have .@2x x 2 C 1/e
De De
2 x
C
xx
x2 C 1
.x C 2y/2 x 2 D 2ye .2x 2y/
D 2y
X @ 1 2 yn e D Hn .x/e 2 x .2n/ : @y nŠ n0
(6.2.5)
On the other hand, we have .@2x x 2 C 1/e
i h 1 2 D @2x x 2 C 1 e 2 x F .y/ D
i y n X h 1 2 : @2x x 2 C 1 Hn .x/e 2 x nŠ n0
(6.2.6)
6.2 Mehler’s Formula and Applications
93
Comparing the coefficients of the power series (6.2.5) and (6.2.6) yields 2 1 2 1 2 @x x 2 Hn .x/e 2 x D .2n C 1/Hn .x/e 2 x ;
which is proportional to (6.2.3). Lemma 6.2.3 (Mehler’s formula). If Hn is the nth Hermite polynomial, then e .x
2 Cy 2 /
2 2 2 1 X zn e .x Cy 2xyz/=.1z / H .x/H .y/ D p : n n 2n nŠ 1 z2 nD0
For the proof, the interested reader can look in Erd´elyi [41], Vol. 2, p. 194. As an application of Mehler’s formula, we shall find a closed-form expression for the heat kernel K.x0 ; x; t/ of the operator @2x x 2 . The heat kernel as an eigenfunction expansion is given by Proposition 6.1.1. Using Proposition 6.2.2, the heat kernel is X K.x0 ; x; t/ D e n t fn .x0 /fn .x/ n0
D
X
1 1 1 2 2 e .2nC1/t e 2 .x0 Cx / p n Hn .x0 /Hn .x/ 2 nŠ n0
X .e 2t /n e t 1 2 2 D p e 2 .x0 Cx / Hn .x0 /Hn .x/: 2n nŠ n0
(6.2.7)
Making z D e 2t and applying Mehler’s formula, (6.2.7) becomes 2
2 2
.x Cx /z 2x0 xz 1 e t 1 2 2 0 1z2 e 2 .x0 Cx / e K.x0 ; x; t/ D p p 1 z2 2 1 e t Œ 1 .x 2 Cx 2 / 1Cz2 2x0 x z 2 : 1z 1z D p p e 2 0 1 z2
Since p
r
e t 1
z2
D
e 2t D 1 e 4t
r e 2t
1 1 D p ; 2t e 2 sinh.2t/
1 C e 4t e 2t C e 2t 1 C z2 D D D coth.2t/; 1 z2 1 e 4t e 2t e 2t z e 2t 1 ; D D 1 z2 1 e 4t 2 sinh.2t/
(6.2.8)
94
6 The Eigenfunction Expansion Method
formula (6.2.7) becomes the heat kernel of @2x x 2 : 1 1 1 2 2 e Œ 2 .x0 Cx / coth.2t /x0 x= sinh.2t / ; K.x0 ; x; t/ D p p 2 sinh.2t/
t > 0: (6.2.9)
6.3 Hille–Hardy’s Formula and Applications In this section we show how a bilinear generating function involving Laguerre polynomials, called Hille–Hardy’s formula, is reducing the series (6.1.1) to a function 2 in the case of the operator L D x@2x C @x 14 .b 2 x C ax /. Then we shall provide a couple of applications regarding the Bessel operator and an operator with gravitational potential in two dimensions. We shall first present some prerequisites regarding Laguerre polynomials. The Laguerre polynomial Lan of degree n and parameter a can be defined either by the Rodrigues formula Lan D
e x x a d n x nCa 1 x D .x/n C : : : ; n e nŠ dx nŠ
or equivalently, by the generating function formula X
Lan .x/y n D
n0
1 e xy=.y1/ : .1 y/aC1
(6.3.10)
It is worth noting that for the same degree n the graph does not depend essentially on the value of the parameter a; see Fig. 6.2. Applying integration by parts several times yields 8 Z 1 if n 6D m, < 0; Lan .x/Lam .x/e x x a dx D .a C n C 1/ : 0 ; if n D m. nŠ
a
b 40
600
100
20
400 200
50
10 2 -10
800
150
30
-2
c
4
6
8
10
12
5 -50
10
15
-200
5
10
15
20
25
-400 -600
Fig. 6.2 The graphs of the Laguerre polynomial La5 .x/ for a few distinct values of the parameter a: .a/ a D 0; .b/ a D 3; .c/ a D 9
6.3 Hille–Hardy’s Formula and Applications
95
Hence s fn .x/ D
nŠ e x=2 x a=2 Lan .x/; .a C n C 1/
n D 0; 1; : : : ;
(6.3.11)
forms an orthogonal system for L2 .0; 1/. One may show that the system is also complete. Proposition 6.3.1. The functions (6.3.11) are eigenfunctions for the operator L D x@2x C @x
1 a2 xC ; 4 x
with the corresponding eigenvalue n D n C
aC1 2
(6.3.12)
.
Proof. It comes from the generating formula (6.3.10) and the fact that y D Lan .x/ verifies the Laguerre equation xy 00 C .˛ C a x/y 0 C ny D 0: The following bilinear generating formula can be found in Erd´elyi [41], Vol. 2, p. 189. Proposition 6.3.2 (Hille–Hardy’s formula). If Lan denotes the Laguerre polynomial, and Ia is the modified Bessel function of first type of order a, then for jzj < 1 we have 1 X
.xCy/z nŠ 1 zn .xyz/a=2 Lan .x/Lan .y/ D e .1z/ Ia .n C a C 1/ .1 z/ nD0
p 2 xyz : 1z
In the following let z D e t and use that p
1 z D ; 1z 2 sinh.t=2/
1C
2z 1 C e t D coth.t=2/: D 1z 1 e t
By Proposition 6.1.1, the heat kernel for the operator (6.3.12) on .0; 1/ is given by K.x0 ; x; t/ D
X
e n t fn .x0 /fn .x/
n0
D
X
n0
a
t
e nt e 2 t e 2
p X 1 D e 2 .x0 Cx/ z
nŠ La .x0 /Lan .x/x0a=2 e x0 =2 x a=2 e x=2 .a C n C 1/ n
nŠ zn .x0 xz/a=2 Lan .x0 /Lan .x/ .a C n C 1/ n0
96
6 The Eigenfunction Expansion Method
p p z .x0 Cx/ z z 1 .x0 Cx/ 2 1z De Ia 2 x0 x e 1z 1z (by Hille–Hardy’s formula) 2z p 1 1 1 2 .x0 Cx/.1C 1z / Ia x0 x D e 2 sinh.t=2/ sinh.t=2/ p x0 x 1 1 .x0 Cx/ coth.t =2/ 2 Ia : e D 2 sinh.t=2/ sinh.t=2/ p
Hence the heat kernel of the operator LD
x@2x
1 a2 C @x xC 4 x
is given by the formula 1 1 K.x0 ; x; t/ D e 2 .x0 Cx/ coth.t =2/ Ia 2 sinh.t=2/
p x0 x ; sinh.t=2/
t > 0:
Next we shall present a slight generalization of the previous result. If we define L1 D L, then we shall find the kernel of the operator Lb D
x@2x
1 a2 2 b xC : C @x 4 x
Q 1 . Then If we take xQ D bx as a new variable, a computation shows that Lb D b L .e
t Lb
Z Q1 Q t bL f /.x0 / D e f .xQ 0 / D K.bx0 ; bx; bt/f .x0 /bdx0 ;
and hence the heat kernel of e t Lb is bK.bx0 ; bx; tb/; that is, b=2 b e 2 .x0 Cx/ coth.bt =2/ Ia Kb .x0 ; x; t/ D sinh.bt=2/
p b x0 x : sinh.bt=2/
(6.3.13)
Therefore, the heat kernel of the operator Lb is given by formula (6.3.13). We shall present next two applications. 2
a , then its heat kernel is Heat kernel for the Bessel operator. If L0 D x@2x C @x 4x obtained by making b ! 0 in the above relation: p 2 x0 x 1 1 K0 .x0 ; x; t/ D e t .x0 Cx/Ia ; (6.3.14) t t
where we used lim
b!0
bt=2 D 1: sinh.bt=2/
6.3 Hille–Hardy’s Formula and Applications
97
The change of variable x D r 2 in L0 yields a Bessel operator with potential a2 =r 2 : a2 a2 1 2 1 2 x@x C @x @ C @r 2 : D 4x 4 r r r Let D t=2. Since the following relation holds among the heat kernels e
a2 t x@2 x C@x 4x
De
t 4
2
1 2 a @2 r C r @r 2 r
De
2
1 2 a2 @2 r C r @r 2 r
;
2
then the heat kernel of 12 .@2r C 1r @2r ar 2 / is obtained from the heat kernel of x@2x C 2
a @x 4x by making the substitutions x0 D r02 , x D r 2 and D t=2 in formula (6.3.14). 2 Hence the heat kernel of 12 @2r C 1r @r ar 2 is
G.r0 ; r; / D
1 1 .r 2 Cr 2 / r0 r Ia : e 2 0 2
(6.3.15)
Operator with gravitational potential. In the following we shall present an application of formula (6.3.15) to a two-dimensional operator with gravitational potential; see [26]. We shall look for the heat kernel of the operator LD
2 1 2 ; @x1 C @2x2 2 2 x1 C x22
with 2 R:
(6.3.16)
In polar coordinates this becomes LD
2 1 2 1 1 @r C @r C 2 @2 2 : 2 r r r
If K is the heat kernel of the above operator, then applying a partial Fourier transform with respect to , we obtain that KO D F K is the heat kernel of the operator PD
1 1 2 1 a2 22 1 2 1 @r C @r 2 2 2 D @r C @r 2 ; 2 r r 2r 2 r r
with a2 D 22 C 2 . The heat kernel e t P can be obtained from the formula (6.3.14): O 0 ; r; I / D 1 e 21 .r02 Cr 2 / Ia r0 r : K.r 2
98
6 The Eigenfunction Expansion Method
Applying the inverse Fourier transform yields the heat kernel of the operator (6.3.16): Z r r 1 1 1 2 2 0 e i . 0 / e 2 .r0 Cr / I.22 C 2 /1=2 d K.r0 ; 0 ; r; ; / D 2 2 Z r r 1 1 .r 2 Cr 2 / 0 D e i. 0 / I.22 C 2 /1=2 d : e 2 0 4 Remark 6.3.3. In the case D 0 the above formula becomes Z r r 1 1 .r 2 Cr 2 / 0 2 0 d : e K.r0 ; 0 ; r; ; / D e i. 0 / I 4
(6.3.17)
On the other hand, the heat kernel for (6.3.16) with D 0 is given by 1 jxx0 j2 1 r02 Cr 2 2rr0 cos. 0 / 2 2 e e D : 2 t 2 t Comparing with (6.3.17) and making D 1 2
Z
r0 r
and u D 0 yields
e iu I . / d D e cos u :
6.4 Poisson’s Summation Formula and Applications Under some convergence and regularity conditions on the function f , we have the following summation formula, called Poisson’s summation formula (see [86]): 1 X nD1
f .x C 2n/ D
Z 1 1 X ikx 1 iky e e f .y/ dy: 2 1 kD1
As a consequence, we have the next result regarding the theta-function, which was introduced in Sect. 1.5. The result deals with the case of the 3 -function, but similar transformation formulas work for the other theta-functions. Proposition 6.4.1 (Jacobi’s transformation for 3 ). If t 0 D 1=t, we have 3 .zjt 0 / D .it/1=2 e itz
2 =
3 tzjt ;
(6.4.18)
where the third Jacobi theta-function is defined by 3 .zjt/ D
1 X nD1
e itn
2 C2nzi
:
(6.4.19)
6.4 Poisson’s Summation Formula and Applications
99
Proof. We note that we can write 3 .zjt/ D e
1 X
z2 =.i t /
it
2
e 4 .uC2n/ ;
nD1 2 =4
with u D 2z=t. Applying Poisson summation formula with f .u/ D e i t u
yields
Z 1 1 z2 = i t X iku 1 i ty 2 iky e e 4 dy e 3 .zjt/ D 2 1 D
1 z2 = i t e 2
D .i t/
kD1 1 X
e iku 2.i t/1=2 e i k
2 =t
kD1
1=2 z2 = i t
e
3 .z=t j 1=t/:
Replacing z by tz yields (6.4.18).
The interested reader can find details about the theta-functions in reference [86]. Next we shall present a few applications.
6.4.1 Heat Kernel on S 1 In this section we shall compute the heat kernel for the Laplacian on the unit circle S 1 using Jacobi’s transformation. If is the arc length on the unit circle, the d2 1 eigenfunctions of 12 d 2 on the circle S satisfy 1 d2 uk D k uk ; 2 d 2 where k D k 2 =2 and uk ./ D ck e ik . The constant ck can be obtained from the orthogonality condition Z 2 1 uk ./uk ./ d D 2ck2 H) ck D p : 1D 2 0 The functions uk ./ 2
D
p1 e ik 2
1
form a complete orthogonal system on
L .S ; d/. Therefore, the heat kernel for K.0 ; I / D D
X
e k
2 =2
1 d2 2 d 2
on the circle S 1 is given by 1 X k 2 =2 i k. 0 / e e 2 1 0 ˇˇ i ; 3 D ˇ 2 2 2
uk ./uk .0 / D
1 X k2 =2Ci k. 0 / e 2
100
6 The Eigenfunction Expansion Method
by the definition of the 3 given in (6.4.19). Denoting zD
0 ; 2
t0 D
i D 1=t; 2
t D 2 i=;
Jacobi’s transformation formula provides K.0 ; I / D
1 0 3 2 2
ˇ i 1 ˇ D 3 .zjt 0 / ˇ 2 2
1 2 .it/1=2 e itz = 3 .tzjt/ 2 ˇ 2 i i .0 /2 1 ˇ e 2 3 : . 0 / ˇ D p 2 D
(6.4.20)
This formula agrees with the geometric formula of the heat kernel for S 1 obtained in Sect. 3.11.
6.4.2 Heat Kernel on the Segment Œ0; T We shall determine the heat kernels on the segment Œ0; T with the Dirichlet and Neumann boundary conditions, respectively. Dirichlet boundary condition. In this case the temperature u.x; t/ is kept equal to zero at the endpoints x D 0; T all the time. A complete orthonormal system for the d2 Laplacian 12 dx 2 on Œ0; T is given by r uk .x/ D
kx 2 sin ; T T 2
2
with the corresponding eigenvalues k D k2T2 , k D 0; 1; : : : . For any two points x0 ; x 2 Œ0; T , the heat kernel can be expressed as K.x0 ; xI / D
X
e k uk .x0 /uk .x/
k0
D
2 X k2 22 kx0 kx e 2T sin sin T T T k1
D
1 X k2 22 h k.x C x0 / i k.x x0 / : (6.4.21) cos e 2T cos T T T k1
6.4 Poisson’s Summation Formula and Applications
101
Substituting z D .x ˙ x0 /=.2T / and t D =.2T 2 /, the previous series can be written in terms of theta-functions: " # X k2 2 k.x ˙ x0 / 1 2 e 2T cos 3 z j i t 1 D T 2 k1 " ! # .x ˙ x0 / ˇˇ i 1 3 1 ; D ˇ 2 2T 2T 2 and hence (6.4.21) becomes K.x0 ; xI / D
1 .x x0 / 3 2T 2T
ˇ i 1 .x C x0 / ˇ 3 ˇ 2 2T 2T 2T
ˇ i ˇ : (6.4.22) ˇ 2T 2
Even if (6.4.22) is a perfectly valid formula, we shall still transform it according to Poisson’s summation formula, obtaining the familiar term e i 0 / it. Taking z D .x˙x and t 0 D 2T 2 in formula (6.4.18) yields 2T 3
.x ˙ x0 / 2T
.x˙x0 /2 2
in front of
ˇ i 2T .x˙x0 /2 .x ˙ x0 /T i ˇˇ 2T 2 i ˇ 0 2 e 3 ; D 3 .zjt / D p ˇ ˇ 2T 2 2
and hence the heat kernel becomes " .xx0 /2 1 .x x0 /T i ˇˇ 2T 2 i 2 e 3 K.x0 ; xI / D p ˇ 2 # .xCx0 /2 .x C x0 /T i ˇˇ 2T 2 i 2 e 3 : ˇ Neumann boundary condition. In this case there is no leak of heat at the endpoints @ u.x; t/ D 0 for x D 0; T , and all of Œ0; T . This condition can be written as @x 2
d t 0. Then a complete orthonormal system for 12 dx 2 on Œ0; T with respect to the aforementioned boundary conditions is r kx 2 cos ; uk .x/ D T T 2
2
with the corresponding eigenvalues k D k2T2 , k D 0; 1; : : : . A computation similar to the one before provides the heat kernel: K.x0 ; xI / D
X
e k uk .x0 /uk .x/
k0
D
2 X k2 22 kx kx0 2 C cos e 2T cos T T T T k1
102
6 The Eigenfunction Expansion Method
20
1 X k2 2 / 1 4@ k.x C x 0 A 1C D e 2T 2 cos T T k1
0 C @1 C
X
e
2 2 k 2 2T
k1
D
1 3 T
.x C x0 / 2T
13 k.x x0 / A5 cos T
ˇ i 1 .x x0 / ˇ C ˇ 3 2T 2 T 2T
ˇ i ˇ : (6.4.23) ˇ 2T 2
Applying Poisson’s summation formula yields
K.x0 ; xI / D p
1 2
" e
.xCx0 /2 2
Ce
3
.xx0 /2 2
.x C x0 /T i
3
ˇ 2T 2 i ˇ ˇ
# .x x0 /T i ˇˇ 2T 2 i ; ˇ
which agrees with the geometric formula (3.12.37).
6.5 Legendre’s Polynomials and Applications Consider the operator @2 C cot @ :
(6.5.24)
If we make the change of variable cos D x, the operator becomes .1 x 2 /@2x 2x@x ;
x 2 .1; 1/:
(6.5.25)
The nth-degree Legendre polynomial Pn is defined by the Rodrigues formula Pn .x/ D
1 dn 2 .x 1/n : 2n nŠ dxn
Pn can also be defined by the following generating function: X n0
Pn .x/y n D
1 : .1 2xy C y 2 /1=2
6.5 Legendre’s Polynomials and Applications
a
b
1.0
103
c
1.0
0.6
0.8
0.5
0.4
0.6
0.2
0.4 -1.0
-0.5
0.5
0.2
1.0
-0.5
-1.0
-0.5
-0.2
-1.0 0.5
-0.5
1.0
0.5
1.0
-0.4 -0.6
-0.4
-1.0
-0.2
Fig. 6.3 The graphs of a few Legendre polynomials Pn .x/: .a/ n D 5; .b/ n D 10; .c/ n D 23
The graphs of three polynomials are contained in Fig. 6.3. One may show that Pn satisfies the equation .1 x 2 /@2x Pn .x/ 2x@x Pn .x/ D n.n C 1/Pn .x/; and hence Pn .x/ is an eigenfunction of the operator (6.5.25) with the corresponding eigenvalue n D n.n C 1/. Since Z
1 1
Pn .x/Pm .x/ dx D 0; Z
1 1
ŒPn .x/2 dx D
n 6D m;
1 ; n C 1=2
it follows that r fn .x/ D
1 n C Pn .x/; 2
n D 0; 1; 2 : : : ;
forms an orthogonal system on L2 .1; 1/. The system is also complete. By Proposition 6.1.1, the heat kernel for (6.5.25) on .1; 1/ is K.x; y; t/ D
X
.n C 1=2/e n.nC1/t Pn .x/Pn .y/;
n0
and hence the heat kernel for (6.5.24) is K.; ; t/ D
X
.n C 1=2/e n.nC1/t Pn .cos /Pn .cos /:
n0
In this case a bilinear generating function is missing, so the operator (6.5.24) does not have a function-type heat kernel.
104
6 The Eigenfunction Expansion Method
6.6 Legendre Functions and Applications In the following we shall deal with the heat kernel for the Laplacian on the sphere S 2 D fx 2 R3 I jxj D 1g. In spherical coordinates 2 .0; /, 2 Œ0; 2/, the Laplacian is given by S 2 D
1 @ @ 1 @2 : sin C sin @ @ .sin /2 @2 2
The eigenfunctions are given by the spherical harmonics Ymk .; / D cmk Pmjkj .cos /e i k ; where m D 0; 1; 2; : : : and k 2 f0; ˙ 1; : : : ; ˙ mg. Pmk denotes the Legendre function. Since S 2 Ymk D m.m C 1/Ymk ; the eigenvalue corresponding to Ymk is km D m2 m; see, for instance, [63], p. 74. The heat kernel is given by K.0 ; 0 I ; I t/ D D D
X X
k
e m t Ymk .; /Ymk .0 ; 0 / 2 e m.mC1/t Pmk .cos /e i k Pmk .cos 0 /e i k 0 cmk
X
e
m.mC1/t
m0
m X
2 cmk Pmk .cos /Pmk .cos 0 /e i k. 0 / :
kDm
A neater formula can be obtained if the kernel is represented in terms of the angle between x0 and x given by D cos1 .hx0 ; xi/. Even if the inverse cosine function is multi-valued, since the heat kernel is an even, 2-periodic function of , we may use any value of with cos D hx0 ; xi; see references [30, 106]. The heat kernel on S 2 can be written as Z 1 1 1 K.x0 ; xI t/ D p e t =4 p p 2 t 2 t cos cos
1 X 1 2 .1/n . 2 n/e 2t .2 n/ d : nD1
It is worth noting that in the case of the three-dimensional sphere there is a simple formula given by (3.7.33). To conclude this chapter, we note that the eigenfunction expansion technique is a robust method, which works as long as we are able to find an orthonormal set of functions and their eigenvalues. The heat kernel can be expressed in this case as an infinite series. There are only a few known cases when this series has a bilinear generating function, in which case the heat kernel has an elegant expression.
Chapter 7
The Path Integral Approach
7.1 Introducing Path Integrals In 1948 Feynman [42] provided an informal expression for the propagators of the famous Schr¨odinger equation, involving an integration over all the continuous paths with respect to a nonexistent infinite-dimensional Lebesgue measure. In fact, this is not really an integral, since there is no measure to give the integral. Since then, a large number of papers have tried to explain the precise mathematical meaning of the Feynman integral. In this chapter we do not attempt this direction. We are concerned only with the immediate applications of the concept of a Feynman integral to obtaining heat kernels for differential operators. In classical mechanics the particles follow trajectories provided by the Euler– Lagrange equations or Hamiltonian system. Given the initial conditions, standard theorems of ODE theory state the existence and uniqueness of the solution of the above systems of equations. This can be stated by saying that classical particles travel along deterministic trajectories. Unlike the aforementioned case, in quantum mechanics particles travel on nondeterministic trajectories. This was inferred from the double-slit experiment and was theoretically stated by the Heisenberg principle of uncertainty. Nondeterministic trajectories are better described by the notion of transition probability. Let P .c j a/ denote the probability that the particle is in state c given that it starts at state a. An application of the conditional probability rule yields that the previous probability depends on all the intermediate states b: P .c j a/ D
X
P .c j b/P .b j a/I
b
see Fig. 7.1b. In quantum mechanics the conditional probability P .c j a/ is replaced by the probability amplitude function 'ca , in which case the aforementioned rule becomes X 'cb 'ba : 'ca D b
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 7, c Springer Science+Business Media, LLC 2011
105
106
7 The Path Integral Approach
a
b
b
P(c lb)
P(b la) b`
c
c
b’’
Fig. 7.1 .a/ Quantum particles do not travel on deterministic paths. .b/ Nondeterministic trajectories
b b c b Fig. 7.2 Intermediate states and probability amplitude functions
Assume the particle starts at a at time zero and ends at b at time t. Consider the following n intermediate states between a and c: a D b0 ; b1 ; : : : ; bn ; bnC1 D c; such that the particle is in state bk at time tk , k D 0; : : : ; n; see Fig. 7.2. The probability amplitude function between a and c can be written as 'ca D
X
'cbn 'bn bn1 : : : 'b2 b1 'b1 a :
b1 ;:::;bn
When the number of the intermediate states n ! 1, the linear function ab1 b2 bn c tends to a continuous path x.s/ with x.0/ D a and x.t/ D c. The “transition probability” from a to c at time t is obtained by summing over all the continuous paths between a and c parameterized by Œ0; t: 'a;cIt D
X
'x.s/ :
x.s/
A breakthrough idea came from Dirac, who indicated that 'x.s/ D e iS=„ for t small, where S satisfies the Hamilton–Jacobi equation and „ denotes the Planck
7.2 Paths Integrals via Trotter’s Formula
107
constant. The idea was continued by Feynman, who provided a heuristical expression for the propagator of Schr¨odinger’s operator in the form X e iS=„ ; K.a; cI t/ D C x.s/
with C D .2 i „t/n=2 , for n large. This was expressed later as the path integral written in the form Z K.a; cI t/ D e iS.;t /=„ dm./; (7.1.1) P.a;cIt /
where P.a; cI t/ D f W Œ0; t ! R I .0/ D x0 ; .t/ D xg denotes the space of the continuous paths from x0 to x within time t. The “propagator” K.x0 ; xI t/ given by (7.1.1) satisfies Schr¨odinger’s equation n
i„
„2 @ K C x K D U.x/K; @t 2
(7.1.2)
where S.; t/ is the solution of the Hamilton–Jacobi equation @S C H.; @ S / D 0; @t with the Hamiltonian H.x; p/ D
1 2 jpj C U.x/: 2
Feynman’s formula for the propagator of the Schr¨odinger equation can be adapted for the heat kernels by replacing t by i t and U.x/ by V .x/ in (7.1.2). If for the sake of simplicity we choose m D „ D 1, then Schr¨odinger’s equation becomes the following heat equation: 1 @K D x K C V .x/K: @t 2 In this chapter we shall deal with finding heat kernels for second-order differential operators using path integrals. Since any heat kernel can be expressed formally as a path integral, the problem of finding the heat kernel is to compute the path integral explicitly. As we shall see throughout the chapter, even if the computation is sometimes tedious, this can be done for several examples of operators.
7.2 Paths Integrals via Trotter’s Formula This section deals with the application of Trotter’s formula to a formal construction of path integrals as an infinite limit of improper integrals. Consider the heat equation @ D L; @t
(7.2.3)
108
7 The Path Integral Approach
where L D C V .x/, with V a smooth function called the potential, and with the Laplacian D 12 @2x1 C C @2xn . If the initial condition is jt D0 D 0 , then the solution of (7.2.3) is given by .x; t/ D K.t/0 .x/;
t > 0;
where K.t/ D e t L D e t .CV .x//
(7.2.4)
is called the evolution operator or the propagator of the operator L. If 0 D ı0 is the Dirac distribution centered at x D 0, then the heat kernel is given by K.t/ı0 . We need to find an integral representation for the evolution operator K.t/ defined by (7.2.4). Before going on, we shall recall a few concepts regarding integral kernels. Let K be an integral operator on L2 .Rm /. Then its integral kernel, denoted by b K, is defined by Z b y/f .y/ dy; K.f /.x/ D K.x; Rm
b W Rm Rm ! R. with K m=2 jxyj2 ct .x; y/ D 1 e 2t and Example 7.2.1. The Gaussian kernel on Rm is K 2 t hence 1 m=2 Z jxyj2 .e t f /.x/ D e 2t f .y/ dy: (7.2.5) 2 t Rm The following result deals with the integral kernels of the product of two or more operators. Proposition 7.2.2. (1) Let K1 and K2 be two integral operators with the integral c1 and K c2 . Then the integral kernel of the product K D K1 K2 is kernels K b y/ D K.x;
Z Rm
c1 .x; u/K c2 .u; y/ du: K
(2) In the case of r operators K1 ; K2 ; : : : ; Kr , the integral kernel of the product operator K D K1 K2 Kr is Z Z c2 .u1 ; u2 / Kr1 .ur2 ; ur1 / c1 .x; u1 /K b y/ D K K.x; Rm
Rm
1
cr .ur1 ; y/ dur1 du1 : K
7.2 Paths Integrals via Trotter’s Formula
109
Let A and B be two self-adjoint operators on a Hilbert space H. There are two cases: 1. The operators commute. In this case ŒA; B D ABBA D 0 and e ACB D e A e B . Using Proposition 7.2.2, part (1), the integral kernel of e t .ACB/ is given by Z Rm
b
tA ec .x; u/e tB .u; y/ du:
Example 7.2.3. Let A D 12 @2x and B D 12 @2y . Let Z .yv/2 1 e 2t f .x; v/ dv: p .x; y/ D e tB f .x; y/ D 2 t Using that AB D BA, we have t 2 2 e 2 .@x C@y / f .x; y/ D e t .ACB/ f .x; y/ D e tA e tB f .x; y/ Z tA .xu/2 1 e 2t .u; y/ du D e .x; y/ D p 2 t Z Z .xu/2 .yv/2 1 1 D p e 2t f .u; v/ dv du p e 2t 2 t 2 t ZZ 2 2 .xu/ C.yv/ 1 2t D f .u; v/ dudv: e 2 t If we let f .u; v/ D ıx0 .u/ ˝ ıy0 .v/, we get the heat kernel t
2
2
e 2 .@x C@y / .ıx0 ˝ ıy0 / D
1 .xx0 /2 C.yy0 /2 2t : e 2
This is just the product of the heat kernels of A and B. 2. The operators do not commute. If ŒA; B 6D 0, then e ACB 6D e A e B . In this case the following Trotter product formula is needed. Theorem 7.2.4. Let A and B be self-adjoint operators on a separable Hilbert space H such that they are bounded from below and ACB is essentially self-adjoint. Then e ACB D lim
n!1
t n t e n Ae n B ;
with the convergence in the strong operator topology.1
1
limn!1 An D A in the strong operator topology if kAn f Af kH ! 0, for any f 2 H.
110
7 The Path Integral Approach
In our case the Hilbert space H is the space of test functions. We also denote A D t; B D tV .x/ Kn D e
t n
e
[we mean multiplication by tV .x/],
t n V .x/
:
Using the Trotter product formula, (7.2.4) becomes K.t/ D e ACB D lim Kn ı ı Kn D lim Knn : n!1
n!1
(7.2.6)
In order to compute the integral kernel of the product operator Knn , we first need to find the integral kernel for Kn . Substituting t D t=n in (7.2.5) yields t
.e n f /.x/ D
n m=2 Z njxyj2 e 2t f .y/ dy; 2 t
(7.2.7)
t
and substituting e n V .x/ f for f , we get the integral representation for Kn : t
t
.Kn f /.x/ D .e n e n V .x/ f /.x/ D or cn .x; y/ D K
n m=2 Z njxyj2 t e 2t e n V .y/ f .y/ dy; 2 t
n m=2 njxyj2 t e 2t e n V .y/ : 2 t
Using Proposition 7.2.2, part .2/, we obtain the integral kernel of the product operator Knn : cn .x; y/ D K n
Z
Z
b n .x; u1 /K b n .u1 ; u2 / : : : K b n .un1 ; y/ dun1 : : : du1 K
Z n nm=2 Z Pn1 n Pn1 2 t e 2t j D0 juj C1 uj j e n j D0 V .uj C1 / du D 2 t Z P n nm=2 Z n1 n t 2 e j D0 2t juj C1 uj j C n V .uj C1 / du D 2 t Z n nm=2 Z e Sn du; D 2 t where u0 D x, un D y, du D dun1 du1 and Sn D
n1 X j D0
n t juj C1 uj j2 V .uj C1 / : 2t n
(7.2.8)
7.2 Paths Integrals via Trotter’s Formula
111
Using the Trotter formula (7.2.6), the integral kernel of K.t/ is
bD K.t/
cn D lim lim K n
n!1
Z
n!1
Z
e Sn
n nm=2 du: 2 t
(7.2.9)
We shall make sense of the above integral as a Feynman path integral. The space of continuous, finite-energy paths from x to y in Rm within time t is defined by Z t P 2 < 1; .0/ D x; .t/ D y : jj Px;yIt D W Œ0; t ! Rn I 0
Let the points u0 .Dx/; u1 ; : : : ; un1 ; un .Dy/ 2 Rm be fixed. Consider the piecewise linear function n 2 Px;yIt such that n .0/ D x, n . kt / D uk , n .t/ D y, n for all k D 0; : : : ; n. Then (7.2.8) becomes Sn D
n1 X j D0
D
n1 X j D0
ˇ ! ˇ2 t .j C 1/t jt ˇˇ .j C 1/t n ˇˇ V n n n 2t ˇ n n ˇ n n ˇ ˇ2 ˇ ˇ .j C1/t ! n njt ˇ n t .j C 1/t 1 ˇn V n : 2 t=n n n
This is a Riemannian sum for the integral action Z t 1 P S.; t/ D j.s/j2 V .s/ ds; 2 0 which is associated with the Lagrangian 2 P D 1 j.s/j P L.; / V .s/ : 2 Since the associated Hamiltonian is H.p; q/ D
1 2 jpj C V .q/; 2
we note that S.; t/ satisfies the Hamilton–Jacobi equation @S 1 C jr S j2 C V .x/ D 0: @t 2 The measure element becomes d m.n / D
t n nm=2 n nm=2 .n 1/t dn : du D dn 2 t 2 t n n
112
7 The Path Integral Approach
When taking the limit, we assume n ! , Sn ! S , and d m.n / ! d m./, where 2 Px;yIt , S is the classical action, and d m is the Wiener measure defined on the space Px;yIt . A rigorous proof of the existence of the Wiener measure is beyond the goal of this exposition. With this preparation we can formally treat the limit of (7.2.9) as an integral: cn D lim lim K n
n!1
n!1
Z
Z
e
Sn
Z n nm=2 du D e S.;t / d m./: 2 t Px;yIt
This is called a Feynman path integral and is a formal expression that physicists use in their computations of propagators. We shall use it in this chapter to compute heat kernels. With this introduction, the integral kernel of K.t/ can be written as a Feynman path integral: Z
b
K.t/ D Px;yIt
e S.;t / d m./:
(7.2.10)
Anzatz: The above limit does not depend on the sequence of piecewise functions n chosen. Under this assumption, the path integral (7.2.10) makes sense. The concept of “Feynman integral” stems from Feynman’s 1948 paper [42], which contains a formula for the evolution of quantum systems. Making this idea rigorous has proven difficult. In this book we are using this concept just informally.
7.3 Formal Algorithm for Obtaining Heat Kernels The following result states that the heat kernel can be represented as a path integral. This means that the heat kernel from point x to point y at time t depends on all continuous curves between x and y parameterized by Œ0; t. Each curve is counted with a certain weight which is given as a solution of a transport equation. We shall just check the results informally, but the reader can elaborate a proof using the Trotter formula, in a similar way as we did in the case of the operator 12 C V .x/. Theorem 7.3.1. Let .aij / be a symmetric, nondegenerate, positive definite m m matrix. The heat kernel for the operator LD
m 1 X aij .x/@xi @xj C V .x/ 2 i;j D1
can be obtained by the following algorithm: 1. Associate the Hamiltonian as the principal symbol of the operator L: m 1 X H.p; x/ D aij .x/pi pj C V .x/: 2 i;j D1
7.3 Formal Algorithm for Obtaining Heat Kernels
113
2. Using the Legendre transform, associate the Lagrangian L.x; x/ P D
m 1 X ij a .x/xi xj V .x/; 2 i;j D1
where .aij / D .aij /1 . 3. Solve the Euler–Lagrange equation @L d @L D : ds @xP @x 4. Find the classical action S.x; t/ by integrating the Lagrangian along the solution of the Euler–Langrange equation: Z
t
S.x; t/ D
L x.s/; x.s/ P ds:
0
5. The action S.x; t/ also verifies the Hamilton–Jacobi equation @S C H.rx S; x/ D 0: @t 6. Then the integral kernel of the operator e t L is given by the path integral b yI t/ D K.x;
Z Px;yIt
e S.;t / d m./;
where the measure d m./ is determined as in step 8, and S.; t/ is the classical action S evaluated along the continuous curve 2 Px;yIt . 7. Consider the solution vn .t/ of the “transport equation” v0n .t/
m 1 X C aij .x/@xi @xj Sn vn .t/ D 0; 2n
n 1;
i;j D1
satisfying the boundary condition lim
t &0
2 t m=2 n
vn .t/ D 1;
where the Sn are the partial Riemannian sums of the integral action S.; t/. 8. The Wiener measure in the previous path integral is formally defined by dm./ D lim dm.n /; n!1
114
7 The Path Integral Approach
where d m.n / D vn .t/n dn
t .n 1/t dn : n n
Proof. We proceed by an informal computation, i.e., we shall assume, when necessary, that the conditions of the dominated convergence theorem always hold, so we can use the commutativity between the limit symbol and the integral. Using Sn D Sn .x; t/, we have L.e Sn / D
m 1 X aij @xi @xj e Sn C V .x/e Sn 2 i;j D1
8 9 m m = X 1 Sn < X D e aij .@xi Sn / .@xj Sn / aij @xi @xj Sn C V .x/e Sn : ; 2 i;j D1
i;j D1
8 9 m < = X 1 D e Sn H.rSn / aij @xi @xj Sn : : ; 2
(7.3.11)
i;j D1
Using (7.3.11) yields @t L e Sn vnn .t/ D @t e Sn vnn .t/ Le Sn vnn .t/ 8 9 m < = 0 X v 1 D e Sn vnn @t Sn n n e Sn vnn H.rSn / aij @xi @xj Sn : ; vn 2 i;j D1
( D
e Sn vnn
n .@t Sn C H.rSn // vn
m vn X C aij @xi @xj Sn 2n i;j D1 „ ƒ‚ …
!)
v0n
D0 by 7:
D e Sn vnn @t Sn C H.rSn / :
(7.3.12)
Assume there is a constant M > 0 such that je Sn vnn j < M for all n 1. Applying the operator @t L to the path integral b yI t/ D K.x;
Z e Px;yIt
S.;t /
Z d m./ D lim
n!1
Z
e Sn vnn .t/ dun1 du1
7.3 Formal Algorithm for Obtaining Heat Kernels
115
and using (7.3.12) yields b yI t/ D lim .@t L/K.x;
Z
Z
n!1
.@t L/ e Sn vnn .t/ dun1 du1
Z
Z
D lim
n!1
e Sn vnn @t Sn C H.rSn / dun1 du1
D 0; because lim
n!1
@t Sn C H.rSn / D @t S C H.rS / D 0;
by the Hamilton–Jacobi equation. The next theorem deals with second-order operators with a linear part.
Theorem 7.3.2. Let .aij / be a symmetric, nondegenerate, positive definite m m matrix. The heat kernel for the operator LD
m X 1 X aij .x/@xi @xj C Vj .x/@xj 2 i;j D1
j
can be obtained following this algorithm: 1. Associate the following Hamiltonian:2 H.p; x/ D
m X 1 X aij .x/pi pj Vj .x/pj : 2 i;j D1
j
2. Using the Legendre transform, associate the Lagrangian L.x; x/ P D
m 1 X ij a .x/ xi Vi .x/ xj Vj .x/ ; 2 i;j D1
where .aij / D .aij /1 . 3. Solve the Euler–Lagrange equation @L d @L D : ds @xP @x
2
Note the minus sign in front of the second term.
116
7 The Path Integral Approach
4. Find the classical action S.x; t/ by integrating the Lagrangian along the solution of the Euler–Lagrange equation: Z
t
S.x; t/ D
L x.s/; x.s/ P ds:
0
5. The action S.x; t/ also verifies the Hamilton–Jacobi equation @S C H.rx S; x/ D 0: @t 6. Then the integral kernel of the operator e t L is given by the path integral b yI t/ D K.x;
Z Px;yIt
e S.;t / d m./;
where the measure d m./ is determined as in step 8, and S.; t/ is the classical action S evaluated along the continuous curve 2 Px;yIt : 7. Consider the solution vn .t/ of the “transport equation” 1 0 m X 1 @ aij .x/@xi @xj Sn A vn .t/ D 0; v0n .t/ C 2n
n 1;
i;j D1
satisfying the boundary condition lim
t &0
2 t n
m=2 vn .t/ D 1;
where the Sn are the partial Riemannian sums of the integral action S.; t/. 8. The Wiener measure in the above path integral is formally defined by dm./ D lim dm.n /; n!1
where d m.n / D vn .t/ dn n
t n
dn
.n 1/t n
:
Proof. We shall proceed formally, providing a verification by a direct computation without worrying about the details. First we check the form of the Lagrangian given in step 2: Consider the vectors x D .x1 ; : : : ; xn / and V D .V1 ; : : : ; Vn /. The Hamiltonian can also be written as H D
1 hap; pi hV; pi: 2
7.3 Formal Algorithm for Obtaining Heat Kernels
117
One of the Hamiltonian equations yields xP D Hp D ap V H) p D a1 .xP C V /: The Legendre transform yields 1 L D p xP H D hp; ap V i hap; pi C hV; pi 2 1 1 D hap; pi D haa1 .xP C V /; a1 .xP C V /i 2 2 1 X ij 1 a .x/ xP i C Vi .x/ xP j C Vj .x/ : D h.xP C V /; a1 .xP C V /i D 2 2 i;j
By computation, we have m X 1 X L e Sn D aij .x/@xi @xj e Sn C Vj .x/@xj e Sn 2 i;j D1 j 8 9 m < = X X 1 D e Sn aij @xi Sn @xj Sn aij @xi @xj Sn : ; 2 i;j D1
e Sn
X
Vj .x/@xj Sn
j
D e Sn
8 <1 X :2
i;j
i;j
9 = X aij @xi Sn @xj Sn Vj .x/@xj Sn ; j
X 1 e Sn aij @xi @xj Sn 2 i;j 8 9 m < = X 1 D e Sn H.rSn / aij @xi @xj Sn : : ; 2
(7.3.13)
i;j
Using (7.3.13) yields @t L e Sn vnn .t/
8 9 < = 0 X v 1 aij @xi @xj Sn D e Sn vnn .t/ @t Sn n n e Sn vnn .t/ H.rSn / : ; vn 2 i;j
118
7 The Path Integral Approach
( D
e Sn vnn .t/
D
e Sn vnn .t/
n vn X v0n C aij @xi @xj Sn @t Sn C H.rSn / vn .t/ 2n i;j „ ƒ‚ …
!)
D0
@t Sn C H.rSn / :
The rest of the proof is the same as the last part of the proof of Theorem 7.3.1.
When det.aij .x// D 0, the matrix aij cannot be inverted, and we might have difficulty in assigning a Lagrangian. In this case we may skip from step 1 to step 5: Since the main point is finding the action, we may always do this skip as long as we are able to solve the Hamilton–Jacobi equation. We shall consider next several examples of operators for which we are able to compute the heat kernels using the previous theorems. We start the presentation with the easiest case. For more worked-out examples, the reader can consult, for instance [102]. We end this section by reminding the reader of two important integrals that will play a major role in computing path integrals in the sequel. Let M D .m/ij be a k k symmetric nonsingular matrix and .1 ; : : : ; k / 2 Rk be a given vector. Then Z Rk
Z
1
Rk
e 2
Pk
1
e 2
i;j D1
Pk
i;j D1
mij yi yj C
mij yi yj
Pk
i D1 i yi
D .2/k=2 .det M /1=2 ;
(7.3.14) 1
D .2/k=2 .det M /1=2 e 2
Pk
i;j D1 i .M
1 /
ij j
:
(7.3.15)
7.4 The Operator 12 @x2 We are interested in finding the heat kernel of the operator L D 12 @2x by using path integrals. The Hamiltonian is given by the principal symbol H.p; x/ D 12 p 2 and the associated Lagrangian is given by L.x; x/ P D 12 xP 2 . The Euler–Lagrange equation is xR D 0. The classical solution satisfying the boundary conditions x.0/ D x0 and x.t/ D x is xcl .s/ D
x x0 s C x0 : t
The classical action from x0 to x within time t is Z Scl .x; x0 ; t/ D 0
t
1 xP cl .s/2 ds D 2
Z
t 0
2 2 x x0 x.t/ x.0/ : ds D 2t 2 2t
7.4 The Operator 12 @2x
119
The heat kernel of the operator @t 12 @2x is given by the path integral Z K.x0 ; xI t / D
Px0 ;xIt
e S.;t / d m./:
In order to evaluate the above path integral, we shall change the integration variable 2 P0;0It as follows. For any 2 Px0 ;xIt , we consider the continuous curve defined by .s/ D xcl .s/ C
.s/:
(7.4.16)
Since .0/ D xcl .0/ D x0 and .t/ D xcl .t/ D x, it follows that .0/ D .t/ D 0. The reason for the change of variables (7.4.16) is to simplify a later computation from the more complicated integral (7.3.15) to the simpler one (7.3.14). This technique might not always work, in which case we need to use (7.3.15) directly. The value of the action integral along the curve is denoted by S.; t/ D S j .0/ D x0 ; .t/ D x : The action along can be written in terms of the action along
Z t 2 1P 2 1 xP cl .s/C P .s/ ds .s/ ds D 0 2 0 2 Z t Z t Z t 1 1P 2 xP cl .s/2 ds C .s/ ds D xP cl .s/ P .s/ ds C 0 2 0 0 2 Z x x0 t P .s/ ds C S j .0/ D 0; .t/ D 0 : D Scl .x; x0 ; t/ C t 0
S.; t/ D S j.0/Dx0 ; .t/ D x D
Z
. One obtains
t
The second term in the sum vanishes because Z t P .s/ ds D .t/
.0/ D 0:
0
Therefore, the action integral along is S.; t/ D
x x0 2t
2 CS
j .0/ D 0; .t/ D 0 ;
and substituting in our initial path integral yields Z e Px0 ;xIt
S.;t /
d m./ D e
.xx0 /2 2t
Z P0;0It
e S.
;t /
d m. /:
(7.4.17)
This way, the path integral over Px0 ;xIt was replaced by a path integral over the space of continuous loops P0;0It . We shall next compute the right-hand side of (7.4.17).
120
7 The Path Integral Approach
Let uj D .jt=n/, for j D 0; : : : ; n. Then the Riemannian sum Sn of the action integral along is Sn D
n1 X j D0
D
n1 X .uj C1 uj /2 1 uj C1 uj 2 t D 2 t=n n 2t=n j D0
1 ˚ .u1 u0 /2 C .u2 u1 /2 C C .un un1 /2 : (7.4.18) 2t=n
Since u0 D un D 0, we get Sn D D
1 f2u21 2u0 u1 C 2u22 2u1 u2 C C 2u2n1 g 2t=n n1 1 X bij ui uj ; 2t=n
(7.4.19)
i;j D1
where
0
2 1 B 1 2 B B B 0 1 .bij / D B :: B :: B : : B @ 0 0 0 0
0 1 2 :: :
:: :
0 0 0 :: :
0 0
2 1
0 0 0 :: :
1
C C C C C C C C 1 A 2
(7.4.20)
is an .n 1/ .n 1/ matrix. In order to construct the measure d m. /, we need to construct the volume elements vn given by part 8 of Theorem 7.3.1. First we need to construct the transport equation. Since x D un D .t/, differentiating in formula (7.4.18) by un yields @2x Sn D
2 n D : 2t=n t
Hence the transport equation given by part 8 of Theorem 7.3.1 is v0n .t/ C Integrating yields
1 vn .t/ D 0: 2t
cn vn .t/ D p : t
The constant cn is determined from the boundary condition 1 D lim
t &0
2 t 1=2 n
2 t 1=2 c n p : t &0 n t
vn .t/ D lim
7.4 The Operator 12 @2x
It follows that cn D
121
q
n . 2
The volume element becomes r vn .t/ D
n ; 2 t
and the measure element is r dm. / D lim vn .t/ du1 dun1 D lim n
n!1
n!1
n 2 t
n du1 dun1 :
Now we are ready to evaluate the path integral on the right side of (7.4.17). Using (7.4.19), we have Z e S. ;t / d m. / P0;0It
r
n n du1 dun1 n!1 2 t n Z r Z 1 Pn1 n e 2t =n i;j D0 bij ui uj du1 dun1 D lim n!1 2 t n Z r n 1 Pn1 D lim e 2 i;j D0 mij ui uj du1 dun1 n!1 2 t Rn1 n s r n .2/n1 D lim n!1 2 t det.mij / n=2 n 1 1 p : (7.4.21) D lim p n!1 t det.m 2 ij / Z
D lim
Z
e
Sn
where we let mij D bij =.t=n/. Using a property of determinants, det.mij / D
n n1 t
det.bij /:
(7.4.22)
We shall now compute det.bij /. Consider the following .n1/.n1/ determinant:
Bn1
ˇ ˇ 2 ˇ ˇ 1 ˇ ˇ ˇ 0 D ˇˇ : ˇ :: ˇ ˇ 0 ˇ ˇ 0
1 0 2 1 1 2 :: :: : : 0 0 0 0
:: :
0 0 0 :: :
2 1
ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ; ˇ ˇ ˇ 1 ˇˇ 2 ˇ 0 0 0 :: :
122
7 The Path Integral Approach
Expanding after the last row yields the following linear recursion equation: Bn1 D 2Bn2 Bn3 : The associated characteristic equation r 2 D 2r 1 D 0 has the double root r1 D r2 D 1. Then the general term of the sequence Bn can be written as Bn D r1n .C1 C nC2 / D C1 C nC2 ;
(7.4.23)
where C1 and C2 are two real constants which will be determined from the initial conditions. Since ˇ ˇ ˇ 2 1 ˇ ˇ D 3; B1 D j2j D 2; B2 D ˇˇ 1 2 ˇ the constants C1 , C2 satisfy the linear system C1 C C2 D 2; C1 C 2C2 D 3; with solution C1 D C2 D 1. Substituting in (7.4.23) yields Bn D n C 1. Hence det.bij / D Bn1 D n; and then (7.4.22) yields
n n1 det.mij / D n : t Substituting in (7.4.21), we obtain Z
e S.
P0;0It
;t /
1 n n=2 d m. / D lim p n!1 2 t D p
1 2 t
n1=2
1 n1 2 n t
:
(7.4.24)
Substituting in (7.4.17) yields the well-known heat kernel for 12 @2x : Z K.x0 ; xI t/ D
Px0 ;xIt
e S.;t / d m./ D p
1 2 t
e
.xx0 /2 2t
;
t > 0: (7.4.25)
7.5 The Operator 12 .@2x C @x /
123
7.5 The Operator 12 .@x2 C @x / Consider the operator L D 12 .@2x C@x /. Since the operator has a linear term, we shall use the method given by Theorem 7.3.2. The associated Hamiltonian is obtained by replacing @x by the momentum p and changing the sign3 in front of p: H.x; p/ D
1 2 .p p/: 2
The Hamiltonian system 1 1 H) p D C x; P 2 2 pP D Hx D 0 H) p constant; xP D Hp D p
with the boundary conditions x.0/ D x0 ; has the solution x.s/ D
x.t/ D x;
x x0 s C x0 : t
(7.5.26)
The Legendre transform provides the Lagrangian L.x; x/ P D p xP H D D
2 1 1 1 1 1 C xP xP C xP C C xP 2 2 2 2 2
xP 2 xP 1 C C : 2 2 8
The Euler–Lagrange equation is xR D 0 and its solution is given by (7.5.26). The classical action is Z t Z t 1 1 1 L xcl .s/; xP cl .s/ ds D Scl .x0 ; xI t/ D ds x.s/ P 2 C x.s/ P C 2 2 8 0 0 Z t x x0 1 .x x0 /2 C ds C D 2t 2 2t 8 0 D
3
1 .x x0 /2 t C .x x0 / C : 2t 2 8
This is the Hamiltonian of the conjugate operator.
(7.5.27)
124
7 The Path Integral Approach
Let be a curve satisfying .0/ D x0 and .t/ D x. We make a “change of variables,” introducing the curve , .s/ D xcl .s/ C
.s/:
We note that .0/ D .t/ D 0. Next we shall express the action along in terms of the action along : Z t P L .s/; .s/ ds S.; t/ D S. j .0/ D x0 ; .t/ D x/ D 0 Z t 1 1 P2 1P ds .s/ C .s/ C D 2 2 8 0 Z t Z t 1 1 1 2 D xP cl .s/ P .s/ ds ds C xP cl .s/ C xP cl .s/ C 2 2 8 0 0 Z t 1P 1 1 1 P2 .s/ C ds .s/ C C 2 2 8 8 0 Z t x x0 P .s/ ds C S j .0/ D 0; .t/ D 0 t : D Scl .x0 ; xI t/ C t 8 0 Rt Since 0 P .s/ ds D .t/ .0/ D 0, using relation (7.5.27), the above equation becomes S.; t/ D
.x x0 /2 1 C .x x0 / C S j .0/ D 0; .t/ D 0 : 2t 2
(7.5.28)
By Theorem 7.3.2, the heat kernel is Z K.x0 ; xI t / D
Px0 ;xIt
e S.;t / d m./ D e
.xx0 /2 2t
12 .xx0 /
Z P0;0It
e S.
;t /
d m. /: (7.5.29)
We still need to compute the path integral on the right side. We note that this is taken over the space of loops starting at the origin. Consider the intermediate points uj D .jt=n/, j D 0; : : : ; n, with u0 D .0/ D 0, un D .t/ D 0. The Riemannian sum of the action is given by Sn D
n1 X
j D0
D
1 uj C1 uj 1 1 .uj C1 uj /2 t C C 2 2 .t=n/ n 2 t=n 8
t n
n1 X
n1 1 .uj C1 uj /2 t 1X .uj C1 uj / C C 2 t=n 2 8 j D0 j D0 ƒ‚ … „ D0
D
1 2t=n
n1 X i;j D0
t bij ui uj C ; 8
(7.5.30)
7.6 Heat Kernel for L D 12 .@2x @x /
125
with .bij / given by (7.4.20). The same computation as before shows that the volume element is n 1=2 vn .t/ D 2 t and the measure is
n n=2 du1 dun1 : (7.5.31) 2 t Using (7.5.30) and (7.5.31), the path integral on the right side of (7.5.29) can be evaluated as in the following: Z e S. ;t / d m. / d m. / D
P0;0It
Z
Z
n n=2 du1 dun1 n!1 2 t Z n n=2 Z 1 Pn1 t e 2t =n i;j D0 bij ui uj du1 dun1 D e 8 lim n!1 2 t t 1 D e 8 ; .2 t/1=2 D lim
e Sn
by a previous calculation; see (7.4.24). Substituting in (7.5.29), we obtain the heat kernel for 12 .@2x C @x /: K.x0 ; xI t/ D p
1 2 t
e
.xx0 /2 2t
t 1 2 .xx0 / 8
;
t > 0:
(7.5.32)
7.6 Heat Kernel for L D 12 .@x2 @x / The Hamiltonian in this case is H.p; x/ D 12 .p 2 C p/. A similar computation as in the previous section shows that the associated Lagrangian is given by L.x; x/ P D
1 xP 2 xP C ; 2 2 8
with the classical action Scl .x0 ; xI t/ D
.x x0 /2 1 t .x x0 / C : 2t 2 8
The same computation as in the previous example yields the heat kernel for 1 2 .@ @x /: 2 x K.x0 ; xI t/ D p
1 2 t
e
.xx0 /2 2t
t C1 2 .xx0 / 8
;
t > 0;
(7.6.33)
126
7 The Path Integral Approach
7.7 Heat Kernel for L D 12 x 2 @x2 C x@x The operator is elliptic away from the origin. We shall find the heat kernel for x0 ; x > 0. We can obtain a similar formula for x0 ; x < 0. Note that there is no heat transfer between the regions x0 > 0 and x < 0. An efficient way to solve this problem is to reduce the operator to the previous operator. Changing the variable x D e u , and using x
@ @ D ; @x @u
x2
@2 @ @2 D 2 ; 2 @x @u @u
we obtain 1
et . 2 x
2 @2 Cx@ / x x
t
2
D e 2 .@u C@u / .uu0 /2 1 t 1 e 2t 2 .uu0 / 8 D p 2 t D p D p
1 2 t 1
e r
2 t
.ln xln x0 /2 2t
t 1 2 .ln xln x0 / 8
x0 1 .ln xln x0 /2 t ; 8 e 2t x
t >0
where we used formula (7.5.32).
7.8 The Hermite Operator 12 .@x2 a2 x 2 / Consider the operator L D 12 .@2x a2 x 2 /, with a constant. The associated Hamiltonian is H.p; x/ D 12 p 2 12 a2 x 2 and the Hamiltonian equations are xP D Hp D p; pP D Hx D a2 x: Then the classical path between x0 and x within time t satisfies xR D a2 x; x.0/ D x0 ; x.t/ D x: The solution is
x.s/ D A sinh.as/ C B cosh.as/:
From the boundary conditions, we get B D x0 ;
AD
x x0 cosh.at/ : sinh.at/
(7.8.34)
7.8 The Hermite Operator 12 .@2x a2 x 2 /
127
Therefore, the classical path is xcl .s/ D
x x0 cosh.at/ sinh.as/ C x0 cosh.as/: sinh.at/
The associated Lagrangian is
L.x; x/ P D p xP H D xP 2
1 2 1 2 2 xP a x 2 2
D
(7.8.35)
1 2 1 2 2 xP C a x ; 2 2
with the Euler–Lagrange equation (7.9.50). The classical action from x0 to x within time t is Z t 1 2 .s/ ds: xP cl .s/2 C a2 xcl Scl .x; x0 ; t/ D 0 2 We shall compute the above integrand first. In order to simplify the calculations, we let D sinh.at/; D cosh.at/: Using (7.8.35) yields xcl .s/ D
x x0 sinh.as/ C x0 cosh.as/;
!2 1 2 1 2 x x0 xP .s/ D a cosh.as/ C x0 sinh.as/ 2 cl 2 D
1 2 .x x0 /2 x x0 cosh2 .as/ C 2 a cosh.as/x0 sinh.as/ 2 2 ! C x02 sinh2 .as/ ;
1 2 2 .x x0 /2 1 x x0 a xcl .s/ D a2 cosh.as/x0 sinh.as/ sinh2 .as/ C 2 2 2 2 ! C x02 cosh2 .as/ : Using 1 C cosh.2as/ ; 2 adding the above relations yields cosh2 .as/ D
sinh2 .as/ D
1 C cosh.2as/ ; 2
.x x0 / 1 a2 .x x0 /2 1 2 2 .s/ D cosh.2as/ C 2x0 xP cl .s/ C a2 xcl sinh.2as/ 2 2 2 2 ! C x02 cosh.2as/ :
128
7 The Path Integral Approach
Integrating between 0 and t, we obtain the classical action Scl .x; x0 ; t/ D
a2 1 .x x0 /2 .x x0 / sinh.2at/ C 2x0 Œcosh.2at/ 1 2 2a 2 ! C x02 sinh.2at/
a2 1 .x x0 /2 .x x0 / 2 2 C 2x0 2 C x02 2 D 2 2a 2 ! a 2 2 2 2 D .x x0 / C 2x0 .x x0 / C x0 2 ! a 2 2 2 2 2 2 x 2xx0 C x0 C 2xx0 x0 D 2 ! a 2 2 2 2 2 2 x C x0 . / C 2xx0 . / D 2 ! a 2 2 x C x0 2xx0 ; D 2 where we used 2 2 D 1 Hence a Scl .x0 ; x; t/ D .x 2 C x02 / cosh.at/ 2xx0 : 2 sinh.at/
!
(7.8.36)
Let x W Œ0; t ! R be a path satisfying x.0/ D x0 and x.t/ D x. Consider the deviation v.s/ from the classical path x.s/ D xcl .s/ C v.s/; where v.0/ D v.0/ D 0. The action along x.s/ becomes Z t a2 1 2 S x j x.0/ D x0 ; x.t/ D x D xP .s/ C x.s/2 ds 2 0 2 Z t 2 2 a 1 2 D xcl .s/ C v.s/ xP cl .s/ C vP .s/ C ds 2 0 2 Z t Z t 1 1 2 a2 a2 2 2 2 D xP cl .s/ C xcl .s/ ds C vP .s/ C v.s/ ds 2 2 0 2 0 2 Z t xP cl .s/Pv.s/ C a2 xcl .s/v.s/ ds C 0 D Scl .x0 ; x; t/ C S v j v.0/ D 0; v.t/ D 0 Z t C xP cl .s/Pv.s/ C a2 xcl .s/v.s/ ds: 0
7.8 The Hermite Operator 12 .@2x a2 x 2 /
129
We shall show that the last integral vanishes: Z
t
xP cl .s/Pv.s/ C a2 xcl .s/v.s/ ds
0
Z
t
D 0
d .xP cl .s/v.s// ds ds
Z
t
xR cl .s/ a2 xcl .s/ v.s/ ds
0
D xP cl .t/v.t/ xP cl .0/v.0/ D 0; where we used v.0/ D v.t/ D 0 and the Euler–Lagrange equations (7.9.50). Hence S x.s/ D S x j x.0/ D x0 ; x.t/ D x D Scl .x0 ; x; t/CS v j v.0/ D 0; v.t/ D 0 D Scl .x0 ; x; t/ C S v.s/ : Using (7.8.36), Theorem 7.3.1 yields the heat kernel Z K.x0 ; xI t/ D
De
Px0 ;xIt
e S.x.s// d m.x/ D e Scl .x0 ;x;t /
a 2 2 2 sinh.at / .x Cx0 / cosh.at /2xx0
Z P0;0It
Z P0;0It
e S.v.s// d m.v/
e S.v.s// d m.v/: (7.8.37)
In order to compute the path integral on the right side of (7.8.37), we need to consider the intermediate points uj D v.jt=n/, j D 0; : : : ; n, with u0 D v.0/ D 0 and un D v.t/ D 0. The Riemannian sum associated with the action Z t SD 0
1 2 1 2 2 xP C a x 2 2
ds
is Sn D
n1 X j D0
D
X1 1 .uj C1 uj /2 t t C a2 u2j 2 .t=n/2 n 2 n n1
j D0
n1 n1 a2 t X 2 1 X bij ui uj C uj 2t=n 2 n i;j D0
D
n1 1 X aij .t/ ui uj ; 2t=n i;j D0
j D0
(7.8.38)
130
7 The Path Integral Approach
where aij .t/ D bij C 0
at 2 n
ıij
1 2 1 0 0 0 B 1 2 1 0 0 C B C B C 0 C at 2 B 0 1 2 0 B DB : :: :: C :: :: :: C C n ıij : B :: : : C : : B C @ 0 0 0 2 1 A 0 0 0 1 2 2 0 1 0 0 2 C at n B 2 B B 1 2 C at 1 0 n B 2 B B 0 0 1 2 C at B n DB :: :: :: :: :: B : B : : : : B 2 B 0 0 0 2 C at B n @ 0 0 0 1 is an .n 1/ .n 1/ matrix. Let fn .t/ D over the last row yields the recursion
t n
0
1
C C C C C C 0 C C :: C C : C C 1 C 2 A 2 C at n 0
det.aij .t//. Expanding the determinant
h at 2 i fn D 2 C fn1 fn2 : n This can be written as a difference equation: fn 2fn1 C fn2 D a2 fn1 : .t=n/2 Assume the pointwise limit f .t/ D limn!1 fn .t/ exists. Then the above difference equation becomes the ordinary differential equation f 00 .t/ D a2 f .t/; with solution f .t/ D A sinh.at/ C B cosh.at/;
A; B 2 R:
We shall determine the value of the constants A and B from the initial conditions.
7.8 The Hermite Operator 12 .@2x a2 x 2 /
131
Note that det.aij /.0/ D n, so fn .0/ D lim fn .t/ D lim t &0
t &0
t lim det.aij .t// D 0: n t &0
Hence, f .0/ D limn!1 fn .0/ D 0. It follows that B D 0 and f .t/ D A sinh.at/: Differentiating at t D 0, fn0 .0/ D lim
n t 0
t &0
n
nt d o o det aij .t/ det aij .t/ C lim t &0 n dt
D 1; since the second limit is zero. Therefore, f 0 .0/ D 1 and hence A D 1=a. It follows that 1 f .t/ D lim fn .t/ D sinh.at/: (7.8.39) n!1 a A similar computation as in a previous example shows that the volume elements are r n vn .t/ D ; 2 t and hence the Wiener measure becomes d m.v/ D
n n=2 du1 dun1 : 2 t
Next we evaluate the path integral on the right side of (7.8.37): Z P0;0It
e S.v.s// d m.v/ Z
Z
n n=2 du1 dun1 n!1 2 t Z n n=2 Z 1 Pn1 e 2t =n i;j D0 aij ui uj du1 dun1 D lim n!1 2 t n n=2 Z Pn1 1 D lim e 2 i;j D0 mij ui uj du1 dun1 ; n!1 2 t Rn1 D lim
e Sn
132
7 The Path Integral Approach
where mij D Z P0;0It
aij . t=n
s n n=2 .2/n1 e S.v.s// d m.v/ D lim n!1 2 t det.mij / n=2 t 1 1 r D lim p n!1 2 n det aij .t/ .t =n/n1
1 1 1 1 D lim p q D lim p p n!1 n!1 t 2 2 fn .t/ det aij .t/ n r a 1 D p ; sinh.at/ 2 by (7.8.39). Substituting in (7.8.37), we obtain the following heat kernel for the Hermite operator: K.x0 ; xI t/ D p
1 2 t
s
a at e 2 sinh.at / sinh.at/
.x 2 Cx02 / cosh.at /2xx0
; t > 0: (7.8.40)
7.9 Evaluating Path Integrals We have seen in the previous sections that the heat kernel can be expressed formally as a path integral. However, computing path integrals from the definition, even in the simplest cases, is a tedious job and reminds us of computing Riemann integrals starting from the Riemann sum. Since this is not always efficient, we shall next present a couple of methods that overpass the computation of the path integral.
7.9.1 Van Vleck’s Formula Given two points x; y in space and a time t > 0, the propagator from x to y within time t is given by the path integral Z K.x0 ; xI t/ D
Px;yIt
e S.;t / d m./:
(7.9.41)
This means that K depends on all continuous paths joining x and y parameterized by Œ0; t. Among all the possible paths between the aforementioned points, the classical path plays a distinguished role; see Fig. 7.3. This is the path on which a
7.9 Evaluating Path Integrals
133
Fig. 7.3 Classical and quantum paths between x and y
classical particle would travel on and is given by the solution of the Euler–Lagrange system of equations. It is a remarkable fact that for a classical Lagrangian, i.e., a Lagrangian which is at most quadratic in xP i and xi , the path integral (7.9.41) depends only on the classical action. This is the famous van Vleck formula which expresses the path integral in the following closed-form formula(see [109]): s
Z e Px;yIt
S.;t /
d m./ D
1 @2 Scl .x; y/ det e Scl .x;y;t / ; 2 @x @y
where
Z
t
Scl .x; y; t/ D
(7.9.42)
L x.s/; x.s/ P ds
0
is the classical action obtained by integrating the Lagrangian along the solution x.s/ of the Euler–Lagrange equation. The factor s V .t/ D
1 @2 Scl .x; y/ det 2 @x @y
is called the van Vleck determinant, and it plays the role of volume element in the geometric method described in Chap. 3. The aforementioned formula can be applied successfully in the cases of second-order elliptic operators with constant, linear and quadratic potential.
7.9.2 Applications of van Vleck’s Formula In the following we shall present some applications. In all of the next examples the heat kernel will be calculated using the formula K.x; yI t/ D V .t/e Scl .x;yIt / ; where V is the van Vleck determinant.
134
7 The Path Integral Approach
.a/ One-dimensional case. We have seen in Sect. 7.4 that the classical action associated with the operator L D 12 @2x is Scl .x0 ; x; t/ D
.x x0 /2 : 2t
2
@ Scl D 1t , formula (7.9.42) yields the following familiar formula of Since @x 0 @x the heat kernel:
K.x0 ; x; t/ D p
1
1
2 t
2
e 2t .xx0 / ;
t > 0:
.b/ One-dimensional case with linear part. The associated action with the operator L D 12 .@2x C @x / as provided by Sect. 7.5 is Scl .x0 ; x; t/ D Since
.x x0 /2 t 1 C .x x0 / C : 2t 2 8 1 @2 Scl D ; @x0 @x t
van Vleck’s formula (7.9.42) yields the following heat kernel: K.x0 ; x; t/ D p
1
1
2 t
2 1 .xx / t 0 2 8
e 2t .xx0 /
;
t > 0:
.c/ The case of linear potential. In the following we shall find the heat kernel of the operator L D 12 @2x ax, with a 2 R. The Hamiltonian function associated with the operator L is given by H.p; x/ D
1 2 p ax: 2
The Hamiltonian system becomes xP D Hp D p; pP D Hx D a: The associated Lagrangian function is L.x; x/ P D 12 xP 2 C ax. The Euler– Lagrange equation xR D a with the boundary conditions x.0/ D x0 , x.t/ D x has the unique solution x.s/ D
1 2 as C bs C x0 ; 2
(7.9.43)
7.9 Evaluating Path Integrals
135
with
x x0 at : t 2 The Lagrangian along the solution (7.9.43) is bD
(7.9.44)
1 L x.s/; x.s/ P D x.s/ P 2 C ax.s/ 2 1 2 1 2 as C bs C x0 D .as C b/ C a 2 2 1 D a2 s 2 C 2abs C ax0 C b 2 : 2 The classical action is obtained by integrating the Lagrangian along the solution: Z t S.x0 ; xI t/ D L x.s/; x.s/ P ds 0
1 1 D a2 t 3 C abt 2 C ax0 t C b 2 t: 3 2 Substituting b from (7.10.67) yields the classical action x x at 1 2 3 1 x x0 at 2 0 C ax0 t C t a t C at 2 3 t 2 2 t 2 2 .x x0 / 1 1 D (7.9.45) C a.x C x0 /t a2 t 3 : 2t 2 24
Scl .x0 ; xI t/ D
The heat kernel is now given by van Vleck’s formula: v u u K.x0 ; x; t/ D tdet D p
1 2 t
1 @2 Scl 2 @x @x0 e
.xx0 /2 2t
! e Scl .x0 ;x;t /
1 2 3 1 2 a.xCx0 /t C 24 a t
; t > 0:
.d / The n-dimensional linear potential. Consider the n-dimensional operator LD
X 1X 2 @xi ai xi ; 2 n
n
i D1
i D1
with ai 2 R. In this case the Hamiltonian function is H.p; x/ D
1X 2 X pi ai xi ; 2 n
n
i D1
i D1
(7.9.46)
136
7 The Path Integral Approach
and the Hamiltonian equations are xP j D pj ; pPj D aj ;
j D 1; : : : ; n:
The solution joining the points x0 and x is x j .s/ D
1 j aj s 2 C bj s C x0 ; 2
(7.9.47)
with j
bj D
x j x0 aj t ; t 2
j D 1; : : : ; n:
(7.9.48)
The Lagrangian L.x; x/ P D
1 1X 2 X xP i C ai xi jxP i j2 C ha; xi D 2 2 n
n
i D1
i D1
evaluated along the above solution is 1 L x.s/; x.s/ P D jaj2 s 2 C 2ha; bis C ha; x0 i C jbj2 : 2 Integrating the Lagrangian along the solution yields the following classical action: S.x0 ; xI t/ D
jx x0 j2 1 1 C ha; x C x0 it jaj2 t 3 : 2t 2 24
(7.9.49)
The van Vleck determinant is v ! u 2 u tdet 1 @ Scl .x0 ; x/ D .2 t/n=2 ; 2 @x0 @x so we conclude that the heat kernel of the operator (7.9.46) is K.x; x0 ; t/ D
jxx j2 1 1 2 3 2t0 1 2 ha; xCx0 it C 24 jaj t ; t > 0: e n=2 .2 t/
.e/ We shall next find the heat kernel for the Hermite operator 1 LD 2
! d 2 2 2 a x ; dx
7.9 Evaluating Path Integrals
137
with a 2 R. The associated Hamiltonian is given by H.p; x/ D
1 2 1 2 2 p a x 2 2
and the Hamiltonian system of equations is xP D Hp D p; pP D Hx D a2 x: The classical path between x0 and x at time t satisfies xR D a2 x;
(7.9.50)
x.0/ D x0 ; x.t/ D x; with the solution x.s/ D
x x0 cosh.at/ sinh.as/ C x0 cosh.as/: sinh.at/
(7.9.51)
The Lagrangian associated with the above Hamiltonian is given by the Legendre transform 1 1 L.x; x/ P D p xP H D xP 2 C a2 x 2 : (7.9.52) 2 2 Integrating the solution (7.9.51) along the Lagrangian (7.9.52) yields the classical action a .x 2 C x02 / cosh.at/ 2xx0 : (7.9.53) Scl .x0 ; x; t/ D 2 sinh.at/ The van Vleck determinant is v u u V .t/ D tdet
1 @2 Scl 2 @x @x0 s 1 at : D p sinh.at/ 2 t
!
Applying formula (7.9.42) yields the following heat kernel: K.x0 ; x; t/ D p where t > 0.
1 2 t
s
1 2 at 2 at e 2t sinh.at / Œ.x Cx0 / cosh.at /2xx0 ; sinh.at/
138
7 The Path Integral Approach
In spite of the fact that van Vleck’s formula is a fast way of obtaining functiontype heat kernels, it is not useful in the case when the potential is more than quadratic or when the operator is sub-elliptic, since in these cases the operators might not have function-type heat kernels.
7.9.3 The Feynman–Kac Formula This famous formula was first used by Kac in order to compute heat kernels for P 2 Laplace operators with potential; see [75]. Let m D 12 m @ i D1 xi be the Laplace operator on Rm and V be a smooth potential function. Let C0t denote the space of Rn -valued functions .s/ continuous on Œ0; t such that .0/ D 0. Then the following formula holds (see [73, 74]): .e
t .m V /
Z f /.x/ D
C0t
e
Rt 0
V .s/Cx ds
f .t/ C x d m.x/;
(7.9.54)
where m is the Wiener measure, which is the product of m one-dimensional Wiener measures. The right-hand side is a Wiener integral, which is computed as a path integral. Even if in the general case the integral cannot be computed explicitly, we shall work out just a few simple examples of heat kernels for which the Feynman– Kac formula can be successfully applied. 1. The operator 21 @x 2 . In this case V D 0 and m D 1 and hence formula (7.9.54) becomes Z t 2 .e 2 @x f /.x/ D f .t/ C x d m./ C0t
n n=2 Z Pn .uj 1 uj /2 e j D1 2.t =n/ f .un / dun du1 n!1 2 t Rn ( ) n n=2 Z Z P .uj 1 uj /2 n j D1 2.t =n/ D lim e dun1 du1 n!1 2 t R Rn1
D lim
f .un / dun n n=2 Z 2 t n1 .un u0 /2 2 D lim n1=2 e 2t f .un / dun n!1 2 t n R Z .vx/2 1 D e 2t f .v/ dv; (7.9.55) .2 t/1=2 R where we used that x D u0 . Making f .v/ D ıx0 .v/ yields the familiar formula of the heat kernel Z 2 .xx0 /2 1 1 t 2 .vx/ 2t 2t K.x0 ; xI t/ D .e 2 @x ıx0 /.x/ D e ı dv D e : x 0 .2 t/1=2 R .2 t/1=2
7.9 Evaluating Path Integrals
139
Equation (7.9.55) can be written in the equivalent form Z C0t
f .t/Cx d m./ D E x f .Wt Cx/ ;
where E x is the expectation operator and Wt is a Brownian motion starting at zero. This result holds in general for the right-hand side of (7.9.54): Z C0t
e
Rt
0
V .Xs / ds
h Rt i x 0 V .s/Cx ds f .t/ C x d m.x/ D E e f Xt ;
where E x is the expectation operator given that Xt is a Brownian motion starting at x. This shows that Feynman–Kac’s formula can be represented either as a path integral or as an expectation operator; see also formula (8.23.89). 2
2. The Hermite operator in m dimensions. In the case V .u/ D a2 juj2 , formula (7.9.54) becomes Z R a2 a2 t 2 2 e 2 0 j.s/Cxj ds f .t/ C x d m./ .e t .m 2 jxj / f /.x/ D C0t
D lim
n!1
2 t nm Z n
Rnm
e
Pn
j D1
juj 1 uj j2 2 C nt a2 2.t =n/
juj j2
f .un / du; 2 t nm Z e SnC1 f .un / du1 dun ; D lim n!1 n Rnm where SnC1 has the meaning given in formula (7.8.38). The rest of the path integral computation follows Sect. 7.8, with uj 2 Rm and x D u0 . The result will be the n-dimensional analog of formula (7.8.40): Z a2 2 e t .m 2 jxj / f .x/ D K.v; xI t/f .v/ dv; with K.x0 ; xI t/ D
1 .2 t/n=2 t > 0:
at n=2 a e 2 sinh.at / sinh.at/
.jxj2 Cjx0 j2 / cosh.at /2hx;x0 i
;
(7.9.56)
An application of the above formula for computing the heat kernel for the Heisenberg operator can be found in [112].
140
7 The Path Integral Approach
7.10 Non-Commutativity of Sums of Squares Let X and Y be two vector fields on Rm . We denote by X n D X X … the „ ƒ‚ n times
n-fold iterated composition of X . It is easy to note that if the vector fields X and Y commute, then the operators X n and Y n also commute; i.e., ŒX; Y D 0 H) ŒX n ; Y n D 0;
8n 1:
Since the case of concern is the sum of squares of vector fields, we shall limit ourselves to the case n D 2. Next we introduce a special class of noncommutative vector fields. Definition 7.10.1. Two vector fields X and Y satisfy condition N at the point p if ŒX 2 ; Y 2 p 6D 0. If the above condition holds at each point, the distribution D spanned by the vector fields X and Y is called an N-distribution, and the vector fields X and Y are called N-fields. Example 7.10.1. The vector fields X D @x and Y D @y C x@z satisfy condition N everywhere on R3.x;y;z/ . We have ŒX; Y D @z 6D 0; X 2 D @2z ; Y 2 D @2y C 2x@y @z C x 2 @2z ŒX 2 ; Y 2 D 4@x @y @z C 2@2z C 4x@x @2z 6D 0: The computation is left to the reader. In the following we shall provide a sufficient condition for a distribution to be of class N. First we need a few definitions. Definition 7.10.2. A distribution D D spanfX; Y g is called nilpotent if there is an integer k 1 such that all the Lie brackets of X and Y iterated k times vanish. The smallest integer k with this property is called the nilpotence class of D. For instance, the vector fields of Example 7.10.1 span a nilpotent distribution with nilpotence class k D 2: ŒX; ŒX; Y D 0;
ŒY; ŒX; Y D 0;
but ŒX; Y 6D 0:
The nilpotence class of a distribution is not the same thing as the step of the distribution. The nilpotence class describes the functional nature of the distribution (i.e., of polynomial, exponential type), while the step describes the non-holonomy of the distribution (i.e., the degree of non-integrability). More precisely, the step at the point p is the number of brackets plus 1 needed to span Rn at p. The nilpotence class
7.10 Non-Commutativity of Sums of Squares
141
and the step are the same thing only in the case when the distribution is generated by left invariant vector fields on nilpotent Lie groups (for instance, on the Heisenberg group). Example 7.10.2. For example, the distribution spanned by the vector fields X D @x C e y @z and Y D @y on R3 is bracket generating with step 2 everywhere, but it is not nilpotent. On the other side, the distribution spanned by the vector fields X D @x and Y D @y C zx@z is nilpotent with the nilpotence class 2. However, this distribution is not bracket generating on the plane fz D 0g; i.e., it does not have a finite step there. The following result can be found in [23]: Theorem 7.10.3. Any distribution D D spanfX; Y g with nilpotence class 2, is an N-distribution. Proof. Since the nilpotence class is 2, we have ŒX; ŒX; Y D 0;
ŒY; ŒY; X D 0;
ŒX; Y 6D 0:
(7.10.57)
The first two relations of (7.10.57) become X 2 Y C YX 2 D 2X YX; Y 2 X C X Y 2 D 2YX Y:
(7.10.58) (7.10.59)
Multiplying on the right of (7.10.58) by Y and of (7.10.59) by X yields X 2 Y 2 C YX 2 Y D 2.X Y /2 ; Y 2 X 2 C X Y 2 X D 2.YX /2 ; and subtracting, we have ŒX 2 ; Y 2 D X 2 Y 2 Y 2 X 2 D 2.X Y /2 YX 2 Y 2.YX /2 X Y 2 X D 2f.X Y /2 .YX /2 g C .X Y 2 X YX 2 Y /:
(7.10.60)
Multiplying on the left of (7.10.58) by Y and of (7.10.59) by X yields YX 2 Y C Y 2 X 2 D 2.YX /2; X Y 2 X C X 2 Y 2 D 2.X Y /2 ; and subtracting, we get ŒX 2 ; Y 2 D X 2 Y 2 Y 2 X 2 D 2.X Y /2 X Y 2 X 2.YX /2 YX 2 Y D 2f.X Y /2 .YX /2 g C .YX 2 Y X Y 2 X /:
(7.10.61)
142
7 The Path Integral Approach
Comparing (7.10.60) and (7.10.61) yields X Y 2 X YX 2 Y D YX 2 Y X Y 2 X ” X Y 2 X D YX 2 Y;
(7.10.62)
and hence relations (7.10.60) and (7.10.61) become ŒX 2 ; Y 2 D 2f.X Y /2 .YX /2 g:
(7.10.63)
If we let A D X Y and B D YX , using (7.10.62), the operators A and B commute, AB D .X Y /.YX / D X Y 2 X D YX 2 Y D .YX /.X Y / D BA; so A2 B 2 D .A B/.A C B/. Then (7.10.63) becomes ŒX 2 ; Y 2 D 2.X Y YX /.X Y C YX /:
(7.10.64)
We need to show that X and Y are N-fields; i.e., if ŒX; Y 6D 0, then ŒX 2 ; Y 2 6D 0. By contradiction, assume ŒX 2 ; Y 2 D 0. Then by (7.10.64), we have either: X Y YX D 0, i.e., ŒX; Y D 0, which is a contradiction. or: X Y C YX D 0, i.e., X Y D YX:
(7.10.65)
The remainder of the proof deals with showing that (7.10.65) cannot hold. By contradiction, we assume that (7.10.65) holds. Then ŒX; Y D X Y YX D 2X Y D 2YX: Therefore, we have ŒX; ŒX; Y D 0 H) ŒX; X Y D 0 H) X 2 Y D X YX; ŒY; ŒY; X D 0 H) ŒY; YX D 0 H) Y 2 X D YX Y:
(7.10.66) (7.10.67)
Using (7.10.66) and (7.10.67), we have X ŒX; Y D X.X Y YX / D X 2 Y X YX D 0; Y ŒX; Y D Y .X Y YX / D YX Y Y 2 X D 0: Combining the last two relations, we obtain ŒX; Y 2 D ŒX; Y ŒX; Y D .X Y YX /ŒX; Y D X.Y ŒX; Y / Y .X; ŒX; Y / D 0: Hence ŒX; Y D 0, which is a contradiction. It turns out that (7.10.65) cannot hold. It follows that X and Y are N-fields.
7.11 Path Integrals and Sub-Elliptic Operators
143
Next we present an example of a distribution with the nilpotence class 2. Example 7.10.3 (Heisenberg-type distributions). Consider the vector fields X; Y; T on R3 . If ŒX; Y D T; ŒX; T D 0; ŒY; T D 0; we say that D D spanfX; Y g is a Heisenberg-type distribution. Since ŒŒX; Y ; X D ŒT; X D 0;
ŒŒX; Y ; Y D ŒT; Y D 0;
it follows that D is a nilpotent distribution with the nilpotence class 2. One of the classical examples of vector fields with the above properties is X D @x C 2y@z ;
Y D @y 2x@z ;
T D 4@z :
It is left as an exercise for the reader to show that X 2 and Y 2 do not commute.
7.11 Path Integrals and Sub-Elliptic Operators In the following we shall assume that X1 and X2 are N-vector fields on R3 ; i.e., ŒX1 ; X2 6D 0 and ŒX12 ; X22 6D 0. We also assume that the distribution D D spanfX1 ; X2 g is step 2; i.e., fX1 ; X2 ; ŒX1 ; X2 g span R3 at each point. In this section we shall describe the heat kernel of the sub-elliptic operator X D 1 .X12 C X22 /. We are led to the study of the heat semigroup e t X , with t > 0. Since 2 the operators X12 and X22 do not commute, we need to invoke the Trotter formula. Hence we have t 2 t 2 n 1 2 1 2 K.t/ D e t X D e 2 tX1 C 2 tX2 D lim e 2n X1 e 2n X2 D lim Knn ; (7.11.68) n!1
n!1
where t
2
t
2
Kn D e 2n X1 e 2n X2 D Kn1 Kn2 ; t t 2 2 c1 , K c2 be the integral kernels of the cn , K with Kn1 D e 2n X1 , Kn2 D e 2n X2 . Let K n n 1 2 operators Kn , Kn , Kn , respectively. By Proposition 7.2.2, part .1/, we have
cn .x; y/ D K
Z R3
c2 .u; y/ du: c1 .x; u/K K n n
(7.11.69)
c1 and K c2 . Since the computacn , it suffices to know K Hence in order to compute K n n tions of the above integral kernels are virtually identical, it suffices to do it only for the first one.
144
7 The Path Integral Approach
c1 is the integral kernel for the operator @ K t n X12 D
3 X
ai aj @xi @xj C
i;j D1
t X 2. 2n 1
3 X
If X1 D
P3
i D1 ai @xi ,
ai .@xi aj / @xj
i;j D1
is a second-order operator with a linear term. Using Theorem 7.3.1 with L D yields a path integral form for the kernel c1 .x; yI t/ D K n
then
Z Px;yIt
e S.;t / d m./;
t X2 2n 1
(7.11.70)
where S satisfies the Hamilton–Jacobi equation (2.3.17). c2 yields K c1 and K cn by formula (7.11.69). Applying (7.11.68) yields Knowing K n n the integral kernel for @t X :
b
K.t/.x; y/
Z Z cn D lim cn .u1 ; u2 / K cn .un1 ; y/ dun1 du1 cn .x; u1 /K D lim K K n!1 n n!1 Z c2 .v ; u /K c1 .u ; v /K c2 .v ; u / K c1 .u ; v / c1 .x; v /K D lim K 1 n n 1 1 n 1 2 n 2 2 n n1 n n!1 R6n
c2 .v ; y/ dudv; K n n c with dudv D dun1 dvn1 du1 dv1 , and Knj given by (7.11.70). Even if the above computation is always theoretically possible, the computation complexity might make the integral almost impossible to evaluate, even in the case of relative simple vector fields such as X1 D @x1 and X2 D x1 @x2 . The conclusion is that the method of path integrals is not well suited to this type of operator.
Chapter 8
The Stochastic Analysis Method
This chapter deals with probabilistic methods of obtaining the heat kernel. The main idea of this subject is that the heat kernel can be represented as a transition density of an associated stochastic process, as pointed out by Kolmogorov [80] in the early 1930s. The probabilistic methods were also useful in obtaining the heat kernel of the Heisenberg Laplacian, as shown by Hulanicki [68] and Gaveau [49] in the late 1970s.
8.1 Elements of Stochastic Processes A stochastic process describes a random phenomenon that changes in time. For instance, the velocity of a small particle in a liquid, the price of a stock and the interest rates for bonds are random phenomena depending on time. A stochastic process is a family of random variables Xt depending smoothly on a parameter t 2 T . In our case, the parameter set T is usually the half-line Œ0; 1/. For each t, the random variable Xt is defined on a sample space , where we have a Borel field F of subsets of , called events, and a probability measure P on F . We shall assume the reader is already familiar with the basic notions of the probability space .; F ; P / the and random variables defined on it. For each state of the world ! 2 , the function t ! Xt .!/ is called a realization or sample path of the stochastic process. In Fig. 8.1 we have represented a few realizations of a process Xt . Stochastic processes are also called random processes or random functions. For convenience, a stochastic process will be denoted by Xt or X.t/. The process can be real-valued or vectorial. In the latter case Xt D .X1 .t/; : : : ; Xn .t// 2 Rn , for t 2 T . The joint distribution function of a real-valued stochastic process Xt is defined by F .x1 ; x2 ; : : : ; xn I t1 ; t2 ; : : : ; tn / D P Xt1 x1 ; Xt2 x2 ; : : : ; Xtn xn ;
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 8, c Springer Science+Business Media, LLC 2011
145
146
8 The Stochastic Analysis Method xt
Fig. 8.1 Three realizations of the process Xt
xt xt
x
xt
t
for t1 ; t2 ; : : : ; tn 2 T and n 1 integer. The previous function satisfies the following two conditions: (1) Symmetry: For any permutation .i1 ; i2 ; : : : ; in / of .1; 2; : : : ; n/ we have F .xi1 ; xi2 ; : : : ; xin I ti1 ; ti2 ; : : : ; tin / D F .x1 ; x2 ; : : : ; xn I t1 ; t2 ; : : : ; tn /: (2) Compatibility: For any k < n, we have F .x1 ; x2 ; : : : ; xk ; 1; : : : ; 1I t1 ; t2 ; : : : ; tk ; tkC1 ; : : : ; tn / D F .x1 ; x2 ; : : : ; xk I t1 ; t2 ; : : : ; tk /: In 1931 Kolmogorov proved that given any system of distribution functions F satisfying the previous conditions (1) and (2), there are a probability space .; F ; P / and a stochastic process Xt .!/ such that F .x1 ; x2 ; : : : ; xn I t1 ; t2 ; : : : ; tn / gives the joint distribution of the random variables Xt1 .!/; Xt2 .!/; : : : ; Xtn .!/: For a proof of Kolmogorov’s extension theorem, the reader can consult reference [85]. Using Kolmogorov’s result, one may define a stochastic process given the distribution function. We shall consider an important example below. The Gaussian process. This is a process Xt such that for any t1 ; : : : ; tn 2 T , the random variables Xt 1 ; X t 2 ; : : : ; X t n
8.1 Elements of Stochastic Processes
147
have an n-variate Gaussian distribution1 with E.Xtj / D m.tj /; Cov.Xtj ; Xtk / D E ŒXtj m.tj / ŒXtk m.tk / D .tj ; tk / D jk : Let D .jk / denote the covariance matrix, which is assumed nonsingular. The joint distribution function in this case is F .x1 ; x2 ; : : : ; xn I t1 ; t2 ; : : : ; tn / Z D
Z
x1
xn
1
1
8 n < 1 X jG.u/j1=2 exp Gij .u/ ui m.ti / : 2 .2/n=2 i;j D1 9 = du du1 ; uj m.tj / ; n
where .Gij / D G D 1 . Sometimes the following equivalent characterization of Gaussian processes is useful in practice. Proposition 8.1.1. The process Xt is Gaussian if and only if the linear combination Pn a X i D1 i ti is normally distributed for all n 1, all coefficients ai 2 R, and all times ti 2 T . The stationary process. A stochastic process Xt is called stationary in the strict sense if F .x1 ; x2 ; : : : ; xn I t1 C ; t2 C ; : : : ; tn C / D F .x1 ; x2 ; : : : ; xn I t1 ; t2 ; : : : ; tn /; for any t1 ; : : : ; tn ; 2 T . For these kinds of processes, P .X./ x/ D F .x; / D F .x; 0/ D F .x/; and the mean and the variance are constants: E.Xt / D m;
1
Var.Xt / D 2 :
This means that Xt1 ; Xt2 ; : : : ; Xtn have a joint normal distribution.
(8.1.1)
148
8 The Stochastic Analysis Method
Furthermore, the covariance function depends only on the difference jt sj: Z Z .s; t/ D xy dF.x; yI s; t/ D xy dF.x; yI 0; t s/ (8.1.2) R2
D .0; t s/:
R2
A process for which relations (8.1.1)–(8.1.2) hold is said to be stationary in the wide sense. Example. Let Xt D A1 cos t C A2 sin t, with A1 , A2 uncorrelated random variables, having mean 0 and variance 2 . Standard properties of random variables show that E.Xt / D E.A1 / cos t C E.A2 / sin t D 0; Var.Xt / D 2 cos2 t C 2 sin2 t D 2 ; .s; t/ D E.Xs Xt / D 2 .cos s cos t C sin s sin t/ D 2 cos.t s/: Hence Xt is a stationary process in the wide sense. The stationary Gaussian process. Another important process is the Gaussian process with constant mean m.t/ D m and covariance function .tj ; tk / D .tj tk /. This process is stationary in the wide sense. Since a Gaussian distribution is completely determined by its mean and variance, it follows that the process is also stationary in the strict sense. Independent increments process. This is a stochastic process Xt such that for any t1 < t2 < : : : < tn , the random variables Xt2 Xt1 ; Xt3 Xt2 ; : : : ; Xtn Xtn1 are mutually independent. The process is said to be stationary if the increment Xtj Xtj 1 has a distribution depending on the difference tj tj 1 . If the process Xt has independent increments and is stationary such that Xt Xs has a Gaussian distribution with mean E.Xt Xs / D 0 and variance
Var.Xt Xs / D 2 jt sj;
then Xt is called a Wiener–Einstein process. Brownian motion. A stochastic process Wt is called a Brownian motion if it satisfies the following properties: (1) The process starts at zero: W0 D 0 (2) For s < t, the random variable Wt Ws is normally distributed with mean 0 and variance t s
8.1 Elements of Stochastic Processes
149
(3) Wt has independent increments (4) The sample paths t ! Wt .!/ are continuous A Brownian motion is a stationary Gaussian process with independent increments. Its distribution function is given by Z
x
F .x; t/ D
p 1
and
P ˛ Wt Ws ˇ D p
1
2 =.2t /
2 t
e u
Z
1 2.t s/
ˇ
du;
e x
2 =2.t s/
dx;
˛
for 0 s < t and ˛ < ˇ. Since the process starts at zero, we note that Wt is normally distributed with mean 0 and variance t, see Fig. 8.2. An important fact is that the covariance Cov.Wt ; Ws / D minft; sg. This can be shown by using conditions (2) and (3). Assuming s < t, we have Cov.Wt ; Ws / D E.Ws Wt / D E .Ws W0 /.Wt Ws / C Ws2 D E.Ws W0 /E.Wt Ws / C E.Ws2 / D 0 C Var.Ws / D s D minft; sg:
The infinitesimal notation. If 0 D t0 < t1 < < tn , then the increment Wtj Wtj 1 is a random variable normally distributed with zero mean and variance tj tj 1 , and can be denoted by . W /tj . In the case when the difference dt D tj tj 1 is infinitesimal, the random variable . W /tj will be denoted by dW t . The random variable dW t has zero mean and variance p dt. It is a well-known result of stochastic analysis that dW t has the magnitude of dt; more precisely, the random variable dW 2t is completely predictable and we have dW 2t D dt in the mean sense. 5
200 -5
Fig. 8.2 A realization of a Brownian motion
-10
400
600
800
1000
150
8 The Stochastic Analysis Method
Markov processes. Brownian motion belongs to a larger class of “memoryless” processes, called Markov processes. A stochastic process Xt is said to be a Markov process if P .Xt j Xt1 ; Xt2 ; : : : ; Xtk / D P .Xt j Xtk /; for any ascending sequence t1 < t2 < < tk < t. This means that the probability that the particle has position Xt at time t, given that its position is Xti at time ti , for i D 1; : : : ; k, depends only on the last position Xtk of the particle. If Xt is a real-valued process, consider the function F .x0 ; t0 I x; t/ D P .Xt x j Xt0 D x0 /; called the transition distribution function of the process Xt . This function can be interpreted as the probability that the particle will be in a state included in .1; x, given that the particle starts at the point x0 at time t0 . Using the properties of the probability, we can check that F .x0 ; t0 I x; t/ 0;
F .x0 ; t0 I 1; t/ D 0;
F .x0 ; t0 I 1; t/ D 1:
Furthermore, the transition distribution function satisfies the Chapman–Kolmogorov equation Z F .x0 ; t0 I x; t/ D
1 1
F .y; sI x; t/ dy F .x0 ; t0 I y; s/;
t0 < s < t:
(8.1.3)
The derivative of the transition distribution function f .x0 ; t0 I x; t/ D
@ F .x0 ; t0 I x; t/ @x
(8.1.4)
is called the transition density function of Xt . It satisfies the Chapman–Kolmogorov equation Z f .x0 ; t0 I x; t/ D
1
1
f .y; sI x; t/ f .x0 ; t0 I y; s/ dy;
t0 < s < t:
(8.1.5)
The transition density function will play an important role in finding the heat kernel of certain second-order partial differential operators which generate stochastic processes. In the case t0 D 0, we shall denote pt .x0 ; x/ D f .x0 ; 0I x; t/. This makes sense in the case when the transition distribution function F .x0 ; t0 I x; t/ depends on the difference t t0 . These kinds of processes are called homogeneous. In this case F .x0 I x; t/ D P .Xt x j X0 D x0 /; and pt .x0 ; x/ D f .x0 I x; t/ D
@ @x F .x0 I x; t/.
8.1 Elements of Stochastic Processes
151
Stochastic differential equations. In the case of ordinary differential equations we deal with equations of the form dX t dY t D aX t C b ; dt dt
t 0;
which can also be written in the differential form dX t D aX t dt C b dY t ; where Yt is a given function. In the case of stochastic differential equations the continuous process t ! Xt is not differentiable, so one may not write dX t =dt. The stochastic processes of concern in this chapter have the form dX t D a.t; Xt /dt C b.t; Xt / dW t :
(8.1.6)
This can be interpreted as the infinitesimal notation for the finite difference equation Xt Ch Xt D a.t; Xt /h C b.t; Xt / .Wt Ch Wt /;
h & 0:
Another way of looking at (8.1.6) is by using integrals: Z
Z
t 0
dX s D
Z
t 0
a.s; Xs / ds C
0
t
b.s; Xs / dW s ;
where the last term on the right-hand side is an Ito integral. More details about stochastic differential equations and stochastic integrals can be found in references [95] and [83]. Ito’s formula. A process of the type (8.1.6) is called an Ito process. Consider the process Ft D F .t; Xt / derived2 from Xt , with F twice differentiable in both arguments. Ito’s formula provides the expression of dFt , which looks like a usual Taylor series up to the second derivative in Xt : dFt D
2 @F @F 1 @2 F dX t C dt C dX t : 2 @Xt @t 2 @Xt
Using the stochastic relations .dW t /2 D dt;
dt dW t D dW t dt D .dt/2 D 0;
If Xt is the stock price at time t , a financial instrument with the price Ft D F .t; Xt / derived from Xt is called a financial derivative.
2
152
8 The Stochastic Analysis Method
the aforementioned formula becomes the Ito formula
@F @F @F 1 @2 F 2 a.t; Xt / C b.t; Xt / dt C b.t; Xt / dW t : dF t D C 2 @Xt @t 2 @Xt @Xt (8.1.7) For instance, if Ft D F .t; Wt / D Wt2 is the square of the one-dimensional Brownian motion, then 1 dF t D 2Wt dW t C Œ2.dW t /2 D 2Wt dW t C dt: 2 There is also a multidimensional version of Ito’s formula. In this case W1 .t/; : : : ; Wm .t/ denote m independent Brownian motions and consider n Ito processes dX j .t/ D aj .t; Xt / dt C bj1 .t; Xt / dW 1 .t/ C C bj m .t; Xt / dW m .t/; j D 1; : : : ; n: (8.1.8) We have the following result. Lemma 8.1.2 (Ito). Let F .t; x/ D F1 .t; x/; : : : ; Fp.t; x/ be a C 2 map from Œ0; 1/ Rn into Rp , where Xt D X1 .t/; : : : ; Xn .t/ is given by (8.1.8). Then the process Yt D F .t; Xt / is also an Ito process satisfying d Yt D
X @F @F 1 X @2 F .t; X / dXk .t/ C .t; X / dXi .t/ dX j .t/; .t; X / dt C @xk @t 2 @xi @xj k
i;j
with the conventions dW i .t/ dW j .t/ D ıij dt, dW i .t/ dt D dt dW i .t/ D 0.
8.2 Ito Diffusion This section deals with a particular type of Ito process that is related to our study of heat kernels. Definition 8.2.1. A time-homogeneous Ito diffusion is an n-dimensional stochastic process t 0; X0 D x; (8.2.9) dX t D b.Xt /dt C .Xt / dW.t/; where W .t/ D W1 .t/; : : : Wm .t/ is an m-dimensional Brownian motion and b W Rn ! Rn and W Rn ! Rnm are measurable functions.
8.2 Ito Diffusion
153
In general, the stochastic differential equation (8.2.9) does not have a unique solution. For instance, the one-dimensional equation dX t D 3Xt1=3 dt C 3Xt2=3 dW t ;
X0 D 0;
(8.2.10)
has infinitely many solutions. If for any a > 0 we define 'a .x/ D
.x a/3 ; x a; 0; x < a; 2=3
then a straightforward computation provides the derivatives 'a0 .x/ D 3'a .x/ and 'a00 .x/ D 6'a1=3 .x/. Denoting Fa .t/ D 'a Wt , an application of Ito’s formula yields 1 dFt D 'a0 .Wt /dW t C 'a00 .Wt /dt; 2 D 3Fa .t/2=3 dW t C 3Fa .t/1=3 dt: Moreover, Fa .0/ D 'a .W0 / D 'a .0/ D 0. Hence the aforementioned equation has a solution Fa .t/ D 'a .Wt / for each a > 0. Sometimes the solution of (8.2.9) might explode in finite time. An application of Ito’s formula shows that Xt D 1=.1 Wt / is a solution for the following stochastic differential equation: dX t D Xt3 dt C Xt2 dW t ;
t 0; X0 D 1;
(8.2.11)
which blows up when the Brownian motion Wt reaches 1. As in the theory of ordinary differential equations, in order to ensure the existence and uniqueness of the solution of the stochastic differential equation (8.2.9), we need to require an additional Lipschitz condition on the functions b.x/ and .x/. Let jj2 D
X
ij2
denote the norm of the matrix function . The following result can be found, for instance, in reference [95], Theorem 5.2.1. Theorem 8.2.2. If there is a constant K > 0 such that jb.x/ b.y/j C j.x/ .y/j Kjx yj;
8x; y 2 Rn ;
(8.2.12)
then there is a unique stochastic process .Xt /t 0 satisfying the stochastic differential equation (8.2.9).
154
8 The Stochastic Analysis Method
Equation (8.2.10) does not satisfy condition (8.2.12). To show this, let’s assume there is a positive constant K such that jx 1=3 y 1=3 j C jx 2=3 y 2=3 j Kjx yj
(8.2.13)
for all x; y 2 R. Let u D x 1=3 and v D y 1=3 . Then (8.2.13) becomes ju vj C ju2 v2 j Kju3 v3 j;
8u; v 2 R:
Assuming u 6D v and dividing by ju vj, the preceding equation becomes 1 C ju C vj < Kju2 C uv C v2 j: This implies 1 < Kju2 C uv C v2 j or 1 < ju2 C uv C v2 j; K
u; v 2 R; u 6D v:
(8.2.14)
If > 0 and we let juj < , jvj < , then the right term is bounded by ju2 C uv C v2 j juj2 C juj jvj C jvj2 < 3 2 : p This contradicts the inequality of (8.2.14) if we choose < 1= 3K. Hence the inequality (8.2.14) does not hold for x and y small enough, with x 6D y. The Lipschitz condition (8.2.12) does not hold for the stochastic equation (8.2.10), and hence the uniqueness result given by Theorem 8.2.2 does not hold.
8.3 The Generator of an Ito Diffusion Let .Xt /t 0 be a stochastic process with X0 D x0 . In this section we shall deal with the operator associated with Xt . The generator of the stochastic process Xt is the second-order partial differential operator A defined by Af .x/ D lim
t &0
EŒf .Xt / f .x/ ; t
for any smooth function (at least of class C 2 ) with compact support f W Rn ! R. Here E stands for the expectation operator taken at t D 0; i.e., Z EŒf .Xt / D
Rn
f .y/pt .x; y/ dy;
where pt .x; y/ is the transition density of Xt .
8.3 The Generator of an Ito Diffusion
155
In the following we shall find the generator associated with the Ito diffusion (8.2.9). The main tool used in deriving the formula for A is Ito’s formula in several variables. If we let Ft D f .Xt /, then using Ito’s formula, we have dFt D
X @f 1 X @2 f j .Xt / dX it C .Xt / dX it dX t ; @xi 2 @xi @xj i
(8.3.15)
i;j
where Xt D .Xt1 ; : : : ; Xtn / satisfies the Ito diffusion (8.2.9) on components; i.e., dX it D bi .Xt /dt C Œ.Xt / dW.t/i X D bi .Xt /dt C i k dW k .t/:
(8.3.16)
k
Using the stochastic relations dt2 D dt dW k .t/ D 0 and dW k .t/ dW r .t/ D ıkr dt, a computation provides X X i k dW k .t/ bj dt C jk dW k .t/ dX it dX jt D bi dt C D
X
k
i k dW k .t/
X
k
jr dW r .t/
r
k
D
X
i k jr dW k .t/dW r .t/ D
X
k;r
i k jk dt
k
D . T /ij dt: Therefore, dX it dX jt D . T /ij dt: Substituting (8.3.16) and (8.3.17) into (8.3.15) yields " dFt D
# X @f 1 X @2 f T .Xt /. /ij C bi .Xt / .Xt / dt 2 @xi @xj @xi i;j
i
X @f .Xt /i k .Xt / dW k .t/: C @xi i;k
integrating, we obtain # Z t" X 2 X @f 1 @ f T . /ij C bi Ft D F0 C .Xs / ds 2 @xi @xj @xi 0 i;j i XZ t X @f C i k .Xs / dW k .s/: @x i 0 k
i
(8.3.17)
156
8 The Stochastic Analysis Method
Since F0 D f .X0 / D f .x/ and E.f .x// D f .x/, applying the ,expectation operator in the previous relation, we obtain 2 E.Ft / D f .x/ C E 4
Z
t 0
3 1 2 X @f X 1 @ f @ A .Xs / ds 5 : . T /ij C bi 2 @xi @xj @xi 0
i;j
i
Using the commutation between the operator E and the integral
Rt 0
(8.3.18) yields
1X E.Ft / f .x/ @2 f .x/ X @f .x/ D . T /ij C bk : t 2 @xi @xj @xk t &0 lim
i;j
k
We conclude the previous computations with the following result. Theorem 8.3.1. The generator of the Ito diffusion (8.2.9) is given by AD
X @2 @ 1X . T /ij C bk : 2 @xi @xj @xk i;j
(8.3.19)
k
Substituting (8.3.19) in (8.3.18), we obtain Dynkin’s formula: Z EŒf .Xt / D f .x/ C E
0
t
Af .Xs / ds ;
(8.3.20)
for any f 2 C02 .Rn /.
8.4 Kolmogorov’s Backward Equation and Heat Kernel For any function f 2 C02 .Rn /, let v.t; x/ D EŒf .Xt /, given that X0 D x. As usual, E denotes the expectation at time t D 0. Then v.0; x/ D f .x/, and differentiating in Dynkin’s formula (8.3.20), Z v.t; x/ D f .x/ C 0
t
EŒAf .Xs / ds;
yields @v D EŒAf .Xt / D AEŒf .Xt / D Av.t; x/: @t We have arrived at the following result.
8.5 Algorithm for Finding the Heat Kernel
157
Theorem 8.4.1 (Kolmogorov’s backward equation). For any f 2 C02 .Rn /, the function v.t; x/ D EŒf .Xt / satisfies the following Cauchy problem: @v D Av; @t v.0; x/ D f .x/;
t > 0;
where A denotes the generator of the Ito diffusion (8.2.9). Using the expression of the expectation operator, the previous solution can also be written in the familiar form Z f .y/ pt .x; y/ dy; v.t; x/ D Rn
which implies that the transition density pt .x; y/ of the process Xt is the heat kernel for the operator A. This observation is fundamental and will lead to a method of computing heat kernels for second-order partial differential operators. The formalization of this fact will be done as in the following. Let f .x/ D ıx be the Dirac distribution centered at x 2 Rn ; i.e., Z ıx .'/ D ıx .y/'.y/ dy D '.x/: Rn
Then for any x0 2 Rn , Z v.t; x0 / D
Rn
ıx .y/ pt .x0 ; y/ dy D pt .x0 ; x/
satisfies @v D Av; @t lim v.t; x/ D ıx I
t > 0;
t &0
i.e., pt .x0 ; x/ is the heat kernel of the operator A. Equivalently, pt .x0 ; x/ is the transition density function between the points x0 D X0 and x D Xt for the diffusion process (8.2.9).
8.5 Algorithm for Finding the Heat Kernel We shall conclude the last few sections with the following algorithm for finding heat kernels for second-order differential operators:
158
8 The Stochastic Analysis Method
1. Consider a second-order operator of the form AD
n n X @ 1 X ij @2 g .x/ C bk ; 2 @xi @xj @xk i;j D1
kD1
with det gij .x/ eventually zero at some points. 2. Decompose the matrix g.x/ 2 Rnn as g.x/ D .x/ T .x/, where .x/ 2 Rnm . 3. Construct the homogeneous Ito diffusion process dXt D b.Xt /dt C .Xt / dW.t/;
t 0; X0 D x0 ;
where W .t/ D W1 .t/; : : : ; Wm .t/ are m independent Brownian motions. 4. Solve the stochastic differential equation for Xt and find the transition density pt .x0 ; x/ with Xt D x, X0 D x0 . 5. The heat kernel (propagator) of the operator A is given by the transition density K.x0 ; xI t/ D pt .x0 ; x/. All the previous steps are almost obvious, with the exception of steps 2 and 4. When solving for Xt , we need also to check the uniqueness of solution by Theorem 8.2.2. The delicate part is to figure out the transition density associated with the process Xt , especially in cases when Xt has a complicated expression. Several methods of finding the transition density will be presented in the next section. We finalize this section with a simple application of the aforementioned algorithm for the operator 12 @2x . In this case n D m D 1 and b D 0, g D 1. Then D 1 and the associated Ito diffusion is dX t D dW t ;
X0 D x0 :
The solution is given by Xt D x0 C Wt , t 0. Since the Brownian motion Wt is normally distributed with mean 0 and variance t, it follows that Xt is normally distributed with mean x0 and variance t, so the transition density for Xt is pt .x0 ; x/ D p
1 2 t
e
.xx0 /2 2t
;
t > 0:
Hence the heat kernel for 12 @2x is given by K.x0 ; xI t/ D p
1 2 t
e
.xx0 /2 2t
;
t > 0:
8.6 Finding the Transition Density
159
8.6 Finding the Transition Density We concluded in the previous section that the problem of finding the heat kernel is now reduced to the problem of finding the transition density of an Ito diffusion process. Next we shall present a few techniques of obtaining the transition density that will be useful in the sequel when computing heat kernels. 1. The case of a well-known distribution. In this case Xt is one of the familiar stochastic processes for which the density is well known. We shall next consider a few examples. 2
1 If Xt D Wt is a Brownian motion, then pt .0; x/ D p2 e x =.2t / ; see t Fig. 8.3.a/. If Xt D x0 C Wt , i.e., a Brownian motion starting at x0 , then the transition 2 density is pt .x0 ; x/ D p 1 e .xx0 / =.2t / . 2 t
If Xt D e Yt , with Yt normally distributed with mean and variance 2 , then Xt is lognormally distributed, with .ln x/2 1 e 2 2 ; pD p x 2
x > 0:
For instance, in the particular case when Xt D e Wt Cc , with c 2 R given by c D ln x0 , the transition density is .ln xln x0 /2 1 2t ; e pt .x0 ; x/ D p x 2 t
x > 0I
(8.6.21)
see Fig. 8.3b. 2. Reduction to a well-known distribution. Here we shall distinguish between the cases of one and more variables.
b
a
0.4
1.2 1.0
0.3
0.8 0.2
0.6 0.4
0.1
0.2 −1.0
−0.5
0.5
1.0
2
4
6
8
Fig. 8.3 .a/ The standard normal distribution; .b/ the lognormal distribution with D 1, D 0:5
160
8 The Stochastic Analysis Method
The case of one variable. Let Xt D h.Wt C c/, with h an increasing, smooth function with the inverse h1 D g. The process starts at x0 D h.c/, so c D g.x0 /. If FXt and FWt denote the distribution functions of Xt and Wt , respectively, then we have FXt .x/ D P .Xt x/ D P .h.Wt C c/ x/ D P .Wt g.x/ c/ D FWt .g.x/ g.x0 //: Differentiating, we obtain the transition probability of Xt : d FX .x/ D FW0 t .g.x/ g.x0 // g 0 .x/ dx t .g.x/g.x0 //2 1 2t D p e g 0 .x/: 2 t
pt .x0 ; x/ D
We note that we recover formula (8.6.21) in the case h.x/ D ln x. The case of several variables. We shall next deal with the case of two variables. The general case of n variables can be treated in a similar way. Consider a continuous differentiable transformation of the region .x1 ; x2 / 2 R into the region .y1 ; y2 / 2 S given by y1 D g.x1 ; x2 /; y2 D h.x1 ; x2 /; with the nonzero Jacobian
@yi @.y1 ; y2 / D det 6D 0; @.x1 ; x2 / @xj i;j
8.x1 ; x2 / 2 R:
By the inverse function theorem, we can write x1 D G.y1 ; y2 /; x2 D H.y1 ; y2 /;
8.y1 ; y2 / 2 S:
Consider the two-dimensional stochastic process Xt D .Xt1 ; Xt2 / such that the vector .x1 ; x2 / is a possible value of Xt . Consider another stochastic process Yt D .Yt1 ; Yt2 / defined by Yt1 D g.Xt1 ; Xt2 /; Yt2 D h.Xt1 ; Xt2 /: Let pX .x/ and pY .y/ be the probability density functions of Xt and Yt , respectively. Since we have P .Xt 2 R/ D P .Yt 2 S/;
8.6 Finding the Transition Density
161
and the left side is given by ZZ P .Xt 2 R/ D
ZZ R
@xi pX .G.y/; H.y// det @y j S
pX .x/ dx D
while the right side is
dy;
ZZ P .Yt 2 S/ D
S
pY .y/ dy;
equating yields
ZZ S
pX .G.y/; H.y// det
@xi @yj
ZZ dy D S
pY .y/ dy:
Since this relation holds for any set S, it follows that pY .y/ D pX .G.y/; H.y// det
@xi @yj
:
Inverting, we get a formula that provides the density function of Xt in terms of the density function of Yt :
@yi : (8.6.22) pX .x/ D pY .g.x/; h.x// det @xj A similar relation holds for the n-dimensional case. If the process Xt can be transformed into the process Yt , and the joint density function for .Yt1 ; Yt2 / is known, then formula (8.6.22) provides a formula for the joint density function of .Xt1 ; Xt2 /. In the particular case when Yt1 and Yt2 are independent processes, then pY .y/ D pY 1 .y1 / pY 2 .y2 / and the aforementioned formula becomes
@yi : (8.6.23) pX .x/ D pY 1 .g.x//pY 2 .h.x// det @xj 3. Reduction to a Wiener integral. Sometimes we deal with stochastic differential equations of the form dX t D f .t/dW t ;
X0 D x0 ;
where Wt is a one-dimensional Brownian motion. The solution can be written in terms of a Wiener integral, Z t X t D x0 C f .s/ dW s : (8.6.24) 0
The following result can be found in [83], p. 11.
162
8 The Stochastic Analysis Method
Rb Theorem 8.6.1. Let f 2 L2 Œa; b. Then the Wiener integral a f .s/ dW s is a Rb Gaussian random variable with mean 0 and variance kf k2 D a f .s/2 ds. The idea of the proof is to show that the result holds for the step functions first. Then approximating any function f 2 L2 Œa; b by step functions, and using that the limit of a Gaussian random variable is Gaussian, one obtains the desired result. The proof details can be found in [83]. Using Theorem 8.6.1, R t the process Xt given by (8.6.24) is Gaussian, with mean x0 and variance 2 D 0 f .s/2 ds. Then the transition density of Xt is Z t
1=2 .xx0 /2 1 2 R t f 2 .s/ ds 2 0 pt .x0 ; x/ D 2 f .s/ ds e ;
t > 0:
(8.6.25)
0
4. The case of an integrated Brownian motion. The solution of the stochastic differential equation dXt D Wt dt;
X0 D x0
is given by
Z Xt D x0 C
t 0;
t 0
Ws ds
(8.6.26)
and is called an integrated Brownian motion. A realization of an integrated Brownian motion starting at x0 D 0 is given in Fig. 8.4. We shall find next the transition density of Xt .
y
0:8
0:5 Xt =
t 0
Ws ds
0:2 0:3
0:6 .
0:9 . t
Fig. 8.4 A realization of an integrated Brownian motion Xt D
Rt 0
Ws ds
8.6 Finding the Transition Density
163
For any smooth function f, the integration by parts for a Riemann–Stieltjes integral yields Z
Z
b a
f .s/ dW s D f .b/Wb f .a/Wa Z D f .b/Wb f .a/Wa
b
Ws df .s/
a b a
Ws f 0 .s/ ds:
Making f .s/ D s, a D 0, b D t, and using that W0 D 0, the previous relation implies Z t Z t Ws ds D tWt s dW s : 0
0
Then (8.6.26) becomes Z Xt D .x0 C tWt /
t
s dW s ;
0
t > 0:
(8.6.27)
Then Xt is a difference of two Gaussian random variables (see Theorem 8.6.1), and hence it will be Gaussian. Its mean is equal to Z E.Xt / D E.x0 C tWt / E
t 0
f .s/ dW s
D x0 :
A direct computation (see [99], p. 370) shows that the variance is Z Var.Xt / D Var
t 0
Ws ds D
t3 : 3
Hence the transition probability of Xt is given by r pt .x0 ; x/ D
3.xx0 /2 3 2t 3 e ; 2 t 3
t > 0:
(8.6.28)
5. Using the transition distribution function. If Xt is a one-dimensional continuous stochastic process with X0 D x0 , we recall that the transition distribution function is defined by F .x0 I x; t/ D P .Xt x j X0 D x0 /; where P denotes the probability. Then the transition density is obtained by differentiating the distribution function pt .x0 ; x/ D
d F .x0 I x; t/: dx
164
8 The Stochastic Analysis Method
A similar relation works in the case of several variables. For instance, consider a two-dimensional continuous process Xt D X1 .t/; X2 .t/ ; then define F .x10 ; x20 I x1 ; x2 I t/ D P .X1 .t/ x1 ; X2 .t/ x2 j X1 .0/ D x10 ; X2 .0/ D x20 /: Then the transition probability is obtained as the following mixed derivative: pt .x0 ; x/ D
@2 F .x10 ; x20 I x1 ; x2 I t/: @x1 @x2
6. Using the expectation operator. Assume again that Xt is a one-dimensional continuous stochastic process with X0 D x0 , and transition density pt .x0 ; x/. Then the expectation operator is given by Z (8.6.29) EŒf .Xt / D f .x/pt .x0 ; x/ dx; for any function f. This will lead to the following formula for computing the transition probability: pt .x0 ; x/ D EŒı.Xt x/; (8.6.30) where ı is the Dirac distribution and Xt is the associated stochastic process with X0 D x0 and Xt D x. The reason behind the aforementioned formula is the simple computation Z EŒı.Xt x/ D
ı.y x/pt .x0 ; y/ dy D pt .x0 ; x/I
see formula (8.6.29). We shall exemplify this in the case of the operator 12 @2x . The associated Ito diffusion satisfies dX t D dW t with the solution Xt D x0 C Wt . Consequently, Z pt .x0 ; x/ D EŒı.x0 C Wt x/ D
y2 1 e 2t dy ı.y .x x0 // p 2 t
.xx0 /2 1 e 2t : D p 2 t
However, for sub-elliptic operators, this easy computation does not work. Then we shall proceed as in the following. The trick is to use the expression of the Dirac distribution as the inverse Fourier transform of 1: Z 1 ı.y/ D e i y d : 2
8.6 Finding the Transition Density
165
Then we have Z
y2 1 e 2t dy ı.y x C x0 / p 2 t Ry Z Z y2 1 1 e i .yxCx0 / d e 2t dy D p 2 t Ry 2 R # Z "Z 2 1 1 i .yxCx0 / y2t e dy d D p 2 t 2 R R # "Z Z y2 1 1 D p e i .xx0 / e i y 2t dy d : 2 t 2 R R
pt .x0 ; x/ D
(8.6.31)
From relation (1.6.6), we have Z p y2 2 e i y 2t dy D 2 te t =2 : R
Substituting in (8.6.31) and using (1.6.6) yields pt .x0 ; x/ D
1 2
1 D 2
Z
t
2
e i .xx0 / e 2 d R
r
.xx0 /2 1 2 .xx0 /2 2t Dp e 2t : e t 2 t
(8.6.32)
The aforementioned computation can also be applied in the case of several variables. For instance, in the case of the operator 12 .@2x C @2y /, the associated stochastic process is .Xt ; Yt / D .x0 C Wt ; y0 C Wt / and the probability density can be computed as pt .x0 ; y0 I x; y/ D EŒı.Xt x/ı.Yt y/ D EŒı.x0 C Wt x/EŒı.y0 C Wt y/ Dp
1 2 t
e
.xx0 /2 2t
p
1 2 t
e
.yy0 /2 2t
D
1 1 Œ.xx0 /2 C.yy0 /2 : e 2t 2 t
The main ingredient of the computation was EŒX Y D EŒX EŒY for X and Y independent random variables. However, in general, Xt and Yt are not independent processes and the computation becomes more complicated. We shall deal with this case in Sect. 8.17. 7. Exponential martingale. Let Xt be a one-dimensional stochastic process satisRT fying EŒ 0 Xt2 dt < 1, with T 1. Define the process Zt D e
Rt
0
Xs dW s 1 2
Rt 0
Xs2 ds
:
166
8 The Stochastic Analysis Method
A sufficient condition that Zt be a martingale is Novikov condition (see [95], p. 55): 1 RT 2 E e 2 0 Xs ds < 1: This property has a useful consequence. Since Z0 D 1, from the martingale condition EŒZt D Z0 D 1, so i h Rt R 1 t 2 E e 0 Xs dW s =e 2 0 Xs ds D 1;
8t > 0:
If the process Xt is independent of Wt , then i h Rt R 1 t 2 E e 0 Xs dW s D e 2 0 Xs ds :
(8.6.33)
In particular, if Xt D ˛ is a constant process, the aforementioned relation becomes 1 2 EŒe ˛Wt D e 2 ˛ t .
8.7 Kolmogorov’s Forward Equation The adjoint operator of the operator Af .x/ D
X i;j
aij .x/
X @2 f @f C bk .x/ @xi @xj @xk k
is given by A .x/ D
X i;j
X @ @2 .aij / .bk /; @xi @xj @xk k
and satisfies the property hAf; iL2 D hf; A iL2 , for any f; 2 C02 . Let pt .x; y/ be a smooth transition density of the process Xt which satisfies the stochastic differential equation (8.2.9). Denote by A the generator of the Ito diffusion Xt . (x0 and x have been replaced in this section by x and y, respectively.) Theorem 8.7.1 (Kolmogorov’s forward equation.). The transition density pt .x; y/ satisfies the equation3 d pt .x; y/ D Ay pt .x; y/; dt
3
This equation is also called the Fokker–Plank equation.
8x; y 2 Rn :
8.8 The Operator 12 @2x x@x
167
Proof. Substituting the expectation operator EŒf .Xt / D Dynkin’s formula (8.3.20) yields Z
Z Rn
f .y/pt .x; y/ dy D f .x/ C
t
R
f .y/pt .x; y/ dy into !
Z
0
Rn
Rn
Ay f .y/ ps .x; y/ dy ds:
Differentiating with respect to t and using the definition of the adjoint, we obtain Z
d f .y/ pt .x; y/ dy D dt Rn D
and hence
d p .x; y/ dt t
Z ZR
n
Rn
Ay f .y/ pt .x; y/ dy f .y/ Ay pt .x; y/ dy;
8f 2 C02 .Rn /;
D Ay pt .x; y/.
In the following sections we shall consider several examples of second-order partial differential operators associated with an Ito diffusion for which we shall calculate the heat kernels following the algorithm presented in Sect. 8.5.
8.8 The Operator 12 @x2 x@x In this section we shall study an operator associated with Langevin’s equation in one dimension, dX t D Xt dt C dW t ;
t 0;
(8.8.34)
with X0 D x0 . The aforementioned equation is a homogeneous Ito diffusion with coefficients .x/ D 1 and b.x/ D x. Using formula (8.3.19), we obtain the generator for Langevin’s equation: AD
1 @2 @ x : 2 @x 2 @x
Next we shall solve (8.8.34) by multiplying by the integrating factor e t : e t dX t C Xt e t dt D e t dW t ” d.e t Xt / D e t dW t : Integrating yields e t Xt D x0 C
Z
t
e s dW s ” Z t Xt D e t x0 C e t e s dW s : 0
0
(8.8.35)
168
8 The Stochastic Analysis Method
Since Xt can be expressed in terms of Wiener integral, using Theorem 8.6.1 it follows that Xt is a Gaussian process, with mean E.Xt / D e t x0 and variance Var.Xt / D e 2t
Z
t
e 2s ds D
0
1 e 2t : 2
If we denote the Gaussian density with mean and variance t by Gt .; x/ D p
1 2 t
e
.x/2 2t
;
t > 0;
then Xt has density G.1e2t /=2 .e t x0 ; /. Following the algorithm of Sect. 8.5, the heat kernel of the operator (8.8.35) is given by K.x0 ; xI t/ D pt .x0 ; x/ D G.1e2t /=2 .e t x0 ; x/ .xx0 e t /2 1 e 1e2t D p .1 e 2t / 2
2
x Cx x x 1 0 Cx 2 C 0 D p e 1e2t 0 sinh t ; .1 e 2t /
t > 0:
8.9 Generalized Brownian Motion Consider the stochastic differential equation dX t D a dt C b dW t ;
X0 D x0 ;
(8.9.36)
with a; b 2 R constants, called the drift and diffusion coefficients. Sometimes this is also called Brownian motion with drift. Let X0 D x0 . Integrating in (8.9.36) yields Z Xt D x0 C at C b
t 0
dW s
D x0 C at C bWt ; which is Gaussian distributed with mean E.Xt / D x0 C at and variance Var.Xt / D b 2 t, so the transition density is Gb 2 t .x0 C at; x/ D
1 1 .xx0 at /2 e 2t b2 : p jbj 2 t
(8.9.37)
8.9 Generalized Brownian Motion
169
Using formula (8.3.19), the generator of the generalized Brownian motion (8.9.36) is given by AD
1 2 @2 @ Ca : b 2 2 @x @x
(8.9.38)
Using (8.9.37), it follows that the heat kernel of the operator (8.9.38) is given by 1 1 .xx0 at /2 e 2t b2 ; pt .x0 ; x/ D p b 2 t
t > 0;
(8.9.39)
where we assumed b > 0. We can easily extend this result to the case of n variables considering the operator LD
n n X 1 X 2 @2 @ bj 2 C aj : 2 @xj @xj j D1 j D1
(8.9.40)
The aforementioned operator is the generator of the n-dimensional generalized Brownian motion dX jt D aj dt C bj dW j .t/;
j D 1; : : : ; n:
(8.9.41)
Since W1 .t/; : : : ; Wn .t/ are independent Brownian motions, the processes Xt1 ; : : : ; Xtn are also independent, and hence the joint density function for .Xt1 ; : : : ; Xtn / is the product of the probability density functions associated with each Xtj : n Y ptj .x0j ; x j /; pt .x0 ; x/ D j D1
where ptj .x0j ; x j / D
1
j
2
2 .xj x0 aj t / 1 p e 2t bj ; jbj j 2 t
This is equivalent to the relation e Lt D e L1 t CCLn t D e L1 t e Ln t ; where Lj D
@ 1 2 @2 bj 2 C aj : 2 @xj @xj
t > 0:
170
8 The Stochastic Analysis Method
Proposition 8.9.1. The heat kernel of the operator (8.9.40) is 1 Pn 1 2t j D1 K.x0 ; xI t/ D e jb1 bn j.2 t/n=2
xj xj aj t 2 0 bj
;
t > 0: (8.9.42)
Taking b1 D D bn , we obtain the following result. P P 2 Corollary 8.9.2. The heat kernel of 12 njD1 @ 2 C njD1 aj @x@ is @xj
j
kxx0 k2 kak2 1 e 2t 2 t Cha; xx0 i ; n=2 .2 t/
K.x0 ; xI t/ D
t > 0;
where x0 D .x01 ; : : : ; x0n /, x D .x1 ; : : : ; xn /, a D .a1 ; an / and k k, h ; i denote the length and the inner product of vectors in Rn . Taking a D 12 , we obtain the next result. Corollary 8.9.3. The heat kernel of
1 @2 2 @x 2
1 @ 2 @x
is
.xx0 /2 1 1 1 K.x0 ; xI t/ D p e 2t 2 .xx0 / 8 t ; 2 t
t > 0:
8.10 Linear Noise The following Ito diffusion dX t D Xt dW t ;
X0 D x0 > 0;
(8.10.43)
is modeling linear noise. According to formula (8.3.19), the generator associated with the previous equation is AD
1 2 @2 ; x 2 @x 2
x > 0:
(8.10.44)
Following the algorithm of Sect. 8.5, we next need to solve (8.10.43). In order to do this, we shall consider the process Ft D ln Xt . Applying Ito’s lemma, we obtain 1 1 1 dX t dX 2 Xt 2 Xt2 t 1 1 1 2 D dX t X dt Xt 2 Xt2 t 1 D dW t dt; 2
dFt D
8.11 Geometric Brownian Motion
171
where we used (8.10.43). Integrating yields Z Ft D F0 C
t 0
1 dW s 2
ln Xt D ln x0 C Wt
Z
t
ds ” 0
t ” 2
t
t
Xt D e ln x0 CWt 2 D x0 e Wt 2 : Since Yt D ln x0 C Wt t=2 is normally distributed with mean D ln x0 t=2 and variance 2 D t, then following item 1 of Sect. 8.6, the process Xt has a lognormal distribution with the transition density pt .x0 ; x/ D
p
1
x 2 2
e
.ln x/ 2 2
2
.ln xln x0 C t /2 1 2 2t e D p x 2 t
D
1 t 1 2 1 p e 2t .ln xln x0 / 2 .ln xln x0 / 8 x 2 t
D
1 t x0 1 2 1 e 2t .ln xln x0 / 2 .ln xCln x0 / 8 ; p x 2 t
t > 0;
(8.10.45)
with x; x0 > 0. To conclude, the operator (8.10.44) has the heat kernel given by (8.10.45).
8.11 Geometric Brownian Motion The following stochastic differential equation dX t D ˛Xt dt C ˇXt dW t ;
X0 D x0 > 0;
(8.11.46)
is called geometric Brownian motion and is famous for its applications in the stock market.4 Both the drift and diffusion coefficients are linear, b.x/ D ˛x, .x/ D ˇx, and using (8.3.19) the generator associated with (8.11.46) becomes A D A˛;ˇ D
4
1 2 2 @2 @ ˇ x C ˛x : 2 @x 2 @x
In mathematical finance Xt represents the stock price with return ˛ and volatility ˇ.
(8.11.47)
172
8 The Stochastic Analysis Method
We shall first solve the stochastic equation (8.11.46). Let Ft D ln Xt , with F0 D ln x0 . An application of Ito’s lemma yields 1 1 1 dX t .dX t /2 Xt 2 Xt2 1 1 1 2 2 D .˛Xt dt C ˇXt dW t / ˇ Xt dt Xt 2 Xt2 1 D .˛ ˇ2 /dt C ˇ dW t : 2
dFt D
Integrate and get
Z t Z t 1 2 Ft D F0 C dW s ” ˛ ˇ ds C ˇ 2 0 0
1 ln Xt D ln x0 C ˛ ˇ 2 t C ˇWt ; 2 and hence the solution is 1
Xt D e ln x0 C.˛ 2 ˇ
2 /t CˇW
t
:
(8.11.48)
Let Ut D ln x0 C .˛ 12 ˇ2 /t C ˇWt . The process Ut is normally distributed with mean D ln x0 C .˛ 12 ˇ 2 /t and variance 2 D ˇt. Using the item 1 of Sect. 8.6, we obtain that Xt D e Ut is lognormally distributed with the transition density 2 1 1 .ln x/ 2 2 p e x 2 2 ˇ2 2 1 1 1 D p e 2ˇt Œln xln x0 .˛ 2 / ; x 2ˇt
pt .x0 ; x/ D
t > 0; (8.11.49)
with x; x0 > 0. To conclude, the heat kernel of the operator (8.11.47) is given by the relation (8.11.49). In the particular case when ˇ D 1 and ˛ D 12 , the Ito diffusion (8.11.48) becomes Xt D x0 e Wt ; see (Fig. 8.5). The generator operator A 1 ;1 D 2
1 2 @2 1 @ C x ; x 2 @x 2 2 @x
x > 0;
has the heat kernel K.x0 ; xI t/ D
1 1 1 2 p e 2t .ln xln x0 / ; x 2 t
t > 0:
8.11 Geometric Brownian Motion
173
y
0:8
X t = e Wt
0:5 0:2
0:3 t
Fig. 8.5 A realization of the geometric Brownian motion Xt D e Wt
This result can be generalized to n dimensions in a way similar to that presented at the end of Sect. 8.9. It follows that the operator n n 1 X 2 @2 1X @ xj 2 C xj ; 2 2 @xj @xj j D1 j D1
xj > 0;
has the heat kernel K.x0 ; xI t/ D
j 2 1 1 P 2t j .ln xj ln x0 / ; e x1 xn .2 t/n=2
t > 0:
Remark 8.11.1. The generator (8.11.47) is an Euler operator, which, after the change of variable y D ln x, becomes a second-order operator with constant coefficients AD D
@ 1 2 2 @2 C ˛x ˇ x 2 @x 2 @x @ 1 2 @2 C ˛ ˇ 2 =2 ˇ ; 2 2 @y @y
which is the generator of the Brownian motion with drift ˛ ˇ 2 =2 and diffusion coefficient ˇ given by d Yt D .˛ ˇ 2 =2/ dt C ˇ dW t :
174
8 The Stochastic Analysis Method
8.12 Mean Reverting Ornstein–Uhlenbeck Process In this section we shall study the heat kernel associated with the generator of the stochastic differential equation dXt D .m Xt /dt C dW t ;
X0 D x0 ;
m; 2 R:
(8.12.50)
Equation (8.12.50) has been used to model stochastic interest rates in finance that exhibit mean reverting properties. Since the drift is b.x/ D m x, the associated generator is given by AD
1 2 @2 @ C .m x/ : 2 2 @x @x
(8.12.51)
We may assume > 0. The next step in the algorithm of Sect. 8.5 is to solve the aforementioned stochastic equation. We proceed by writing the relation (8.12.50) in the form dX t C Xt dt D m dt C dW t : Multiplying by e t , we obtain the exact equation: d.e t Xt / D me t dt C e t dW t : Integrating yields e t Xt D x0 C m
Z
t
e s ds C
Z
t
e s dW s ” Z t e s dW s ; Xt D e t x0 C m.1 e t / C e t 0
0
(8.12.52)
0
which is the mean reverting Ornstein–Uhlenbeck process.5 We can see from formula (8.12.52) that Xt is the sum of a deterministic function of t and a Wiener integral. Using Theorem 8.6.1, we obtain that Xt is Gaussian with mean E.Xt / D e t x0 C m.1 e t / and variance 2 2t
Var.Xt / D e
5
Z 0
t
e 2s ds D 2
1 e 2t : 2
The name mean reverting comes from the fact that in the long run the mean E.Xt / tends to m.
8.12 Mean Reverting Ornstein–Uhlenbeck Process
175
Using the notation for the Gaussian distribution, the probability density of Xt is pt .x0 ; x/ D G 2 .1e2t /=2 e t x0 C m.1 e t /; x D D
p p
1 .1 e 2t / 1 .1e 2t /
e
e
1 Œxe t x0 m.1e t /2 2 .1e 2t /
1 Œ.xx0 /C.x0 m/.1e t /2 2 .1e 2t /
; t > 0: (8.12.53)
It follows by the last step of the algorithm in Sect. 8.5 that the operator (8.12.51) has the heat kernel given by (8.12.53). It is interesting to note that when t ! 1, the heat kernel tends to p1 .x0 ; x/ D
1 .xx20 /2 ; p e
which resembles a normal distribution. This shows that the distribution of heat does not go to zero in the long range, but it tends to a steady-state distribution. Particular cases: If we let D m D 1, we obtain that the heat kernel of the operator 1 @2 @ C .1 x/ 2 @x 2 @x is given by K.x0 ; xI t/ D p
1 .1
e 2t /
e
1 1e 2t
Œ.xx0 /C.x0 1/.1e t /2
;
If we take m D 0, we obtain that the operator AD
1 2 @2 @ x ; 2 @x 2 @x
has the heat kernel given by K.x0 ; xI t/ D
p
1 .1 e 2t /
e
.xe t x0 /2 2 .1e 2t /
;
t > 0:
t > 0:
176
8 The Stochastic Analysis Method
8.13 Bessel Operator and Bessel Process In this section we shall deal with the process satisfied by the Euclidean distance from the origin to a particle following a Brownian motion in Rn . More precisely, let W .t/ D .W1 .t/; : : : ; Wn .t// be a Brownian motion in Rn , n 2. Let R.t/ D dist.O; W .t// D
p W1 .t/2 C C Wn .t/2 :
The process R.t/ is not smooth at t D 0, but since W .t/ never hits the origin almost surely for n 2, we may still apply Ito’s formula and obtain n X Wj .t/ n1 dW j .t/ C dt; dR.t/ D R.t/ 2R.t/
(8.13.54)
j D1
called the n-dimensional Bessel process. In this case Wn W1 ; D ; ; R R
P
n1 ; b.R/ D 2R
T
D
Wi2 D 1: R2
Formula (8.3.19) provides the generator AD
n1 @ 1 @2 C ; 2 @ 2 2 @
> 0;
(8.13.55)
called the Bessel operator. The rest of this section deals with finding the transition density pt . 0 ; / of the Bessel process (8.13.54). The computation will be different for the cases 0 D 0 and 0 6D 0. In the case 0 D 0 we have the following result. The next computation can be found in reference [83], p. 133. Proposition 8.13.1. If 0 D 0, the transition density function of R.t/, t > 0, is given by 8 2 2 ˆ ˆ
n1 e 2t ; 0, < n=2 .2t/ .n=2/ pt .0; / D ˆ ˆ : 0;
< 0, with
n 2
D
8 n < . 2 1/Š :
. n2 1/. n2 2/ 32 12
for n even, p
; for n odd.
8.13 Bessel Operator and Bessel Process
177
Proof. Since the Brownian motions W1 .t/; : : : ; Wn .t/ are independent, their joint density function is fW1 Wn .t/ D fW1 .t/ fWn .t/ D
1 2 2 e .x1 CCxn /=.2t / ; n=2 .2 t/
t > 0:
In the next computation we shall use the following formula of integration which follows from the use of polar coordinates: Z Z f .x/ dx D .Sn1 / r n1 g.r/ dr; (8.13.56) 0
fjxjg n
where f .x/ D g.jxj/ is a function on R with spherical symmetry, and where .Sn1 / D
2 n=2 .n=2/
is the area of the .n 1/-dimensional sphere in Rn . Let 0. The distribution function of R.t/ is Z FR . / D P .R.t/ / D fW1 Wn .t/ dx1 dxn fR.t /g Z 1 2 2 D e .x1 CCxn /=.2t / dx1 dxn n=2 2 2 2 .2 t/ x1 CCxn ! Z Z 1 2 /=.2t / n1 .x12 CCxn D r e d dr n=2 0 S.0;1/ .2 t/ Z .Sn1 / n1 r 2 =.2t / D r e dr: .2 t/n=2 0 Differentiating yields d .Sn1 / n1 2 FR . / D
e 2t d .2 t/n=2 2 2
n1 e 2t ; D
> 0; t > 0: n=2 .2t/ .n=2/
pt .0; / D
In the two-dimensional case the aforementioned density becomes Wald’s distribution, (see Fig. 8.6a) ft .x/ D
1 x2 xe 2t ; t
x > 0; t > 0:
The case 0 > 0 is more complex and requires a few preliminary definitions.
178
8 The Stochastic Analysis Method
a
b 10
1.5
8 1.0
6 4
0.5
2 0.5
1.0
1.5
2.0
−4
−2
2
4
Fig. 8.6 .a/ Wald’s distribution; .b/ the Bessel function I0 .x/
The second-order differential equation z2
d2 dw .z2 C 2 /w D 0 Cz 2 dz dz
is called the modified Bessel equation. If D n is an integer, a fundamental system of solutions is given by the solutions In .z/ and Kn .z/. The modified Bessel function of the first kind has the expansion In .z/ D
X .z=2/2mCn I mŠ.m C n/Š m0
(8.13.57)
see reference [41], p. 9. The case D 0 serves our purposes, when the modified Bessel function of order zero of the first kind is given by I0 .z/ D
X .z=2/2m I .mŠ/2 m0
(8.13.58)
see Fig. 8.6b. Lemma 8.13.2. For any z 2 C we have Z
2 0
e z cos d D 2I0 .z/;
where I0 is the modified Bessel function of order zero of the first kind. Proof. Integrating in the Taylor expansion e z cos D
X zn cosn n0
nŠ
(8.13.59)
8.13 Bessel Operator and Bessel Process
179
yields Z
2
e z cos d D
0
X zn Z nŠ
n0
2
cosn d D
0
X z2m Z 2 .cos /2m d .2m/Š 0 m0
D4
X z2m Z =2 X z2m Z =2 .cos /2m dD4 .sin /2m d .2m/Š .2m/Š 0 0 m0 m0
D4
X z2m 1 3 5 .2m 1/ .2m/Š 2 4 .2m/ 2 m0
D 2
X z2m 1 X z2m .2m/Š D2 D2I0 .z/; .2m/Š Œ2 4 .2m/2 22m .mŠ/2
m0
m0
by virtue of (8.13.57). In the previous proof we also used the following relations, which are left as an exercise for the reader: Z
2
.cos /
2mC1
Z d D 0;
0
Z
.cos /
Z
=2
Z
=2
d D 4
.cos /2m d;
0
=2
f .sin / d;
0 =2
2m
0
f .cos / d D Z
2
0
.sin /2m d D
0
1 3 5 .2m 1/ : 2 4 .2m/ 2
Now we go back to the problem of finding the transition density pt . 0 ; /, with t; 0 ; > 0. The Brownian motion starts at the point x0 D .x10 ; : : : ; xn0 / at t D 0 and we let 0 D jx0 j. The conditional distribution function of R.t/ given R.0/ D 0 is given by FR.t /jR.0/ . j 0 / D P R.t/ jR.0/ D 0 Z D f 0 Z D D
fR.t /g
0 W1 Wn jW1 .0/Dx1 Wn .0/Dxn
2 2 g fx12 CCxn
1 .2 t/n=2
Z
.t/ dx1 dxn
Pn 1 1 e 2t j D1 n=2 .2 t/ 1
xj xj0
2
dx1 dxn
2
e 2t jxx0 j dx B.0;/
1 1 2 e 2t jx0 j D .2 t/n=2
Z
1
2
1
e 2t jxj e t hx;x0 i dx:
(8.13.60)
B.0;/
The integral over the ball B.0; / can be easily computed in the particular case n D 2. The general case is more complicated and we shall omit it. Using polar
180
8 The Stochastic Analysis Method
coordinates, the area element changes as dx1 dx2 D r d d, where r D jxj. Since the inner product is invariant under orthogonal transformations, if M 2 O2 , det M D 1, and then Z
1
1
2
e 2t jxj e t hx;x0 i dx D B.0;/
Z
1
2
1
e 2t jM xj e t hM x;M x0 i dx:
(8.13.61)
B.0;/
Choosing the orthogonal matrix M such that M x0 D jx0 je1 D 0 e1 , by using polar coordinates and Fubini’s theorem, the integral (8.13.61) becomes Z
Z
2
e 0
2
r2t
e
1 t r0
cos
Z
r dr d D
re
0
2
r2t
"Z
e
0
Z
1 t r0 cos
d dr
0
D
#
2
0
r2
re 2t I0
r 0
t
r;
by Lemma 8.13.2. Substituting back in (8.13.60) with n D 2 yields the following conditional distribution function: FR.t /jR.0/ . j 0 / D
1 1 2 e 2t 0 2 t
Z 0
r2
re 2t I0
r 0
t
dr:
Differentiating with respect to yields the transition density function pt . 0 ; / D
d 1 1 2 2
0 : FR.t /jR.0/ . j 0 / D e 2t 0 e 2t I0 d 2 t t
We have obtained the following result. Proposition 8.13.3. The heat kernel of the Bessel operator AD
1 @2 1 @ C ; 2 2 @ 2 @
> 0;
(8.13.62)
is given by pt . 0 ; / D
1 .2 C2 /
0 I0 ; e 2t 0 2 t t
t > 0; > 0:
(8.13.63)
Since the heat kernel of a sum of commuting operators is the product of their kernels, we obtain the following consequence.
8.14 Brownian Motion on a Circle
181
Corollary 8.13.4. The heat kernel of the operator n 1 X @2 1X 1 @ C 2 2 xi @xi @xi2 i D1n i D1
is K.x0 ; xI t/ D
n xi xi0 x1 xn 1 .jx0 j2 Cjxj2 / Y 2t ; e I 0 .2 t/n t
xi ; xi0 ; t > 0:
i D1
8.14 Brownian Motion on a Circle Let x0 D .cos 0 ; sin 0 / 2 S 1 be a fixed point on the unit circle. If Wt is the one-dimensional Brownian motion, consider the following system of stochastic differential equations: 1 dX 1 .t/ D X1 .t/ dt X2 .t/ dW t ; 2 1 dX 2 .t/ D X2 .t/ dt C X1 .t/ dW t ; 2
(8.14.64) (8.14.65)
to Theorem 8.2.2, the aforementioned with the initial condition X0 D x0 . According system has a unique solution X.t/ D X1 .t/; X2 .t/ . If we let Xt D X1 .t/; X2 .t/ D cos.0 C Wt /; sin.0 C Wt / ;
(8.14.66)
a direct application of Ito’s formula yields 1 dX 1 .t/ D cos.0 C Wt / dt sin.0 C Wt / dW t ; 2 1 dX2 .t/ D sin.0 C Wt / dt C cos.0 C Wt / dW t ; 2 which shows that (8.14.66) is the desired solution for the system (8.14.64)– (8.14.65). The process (8.14.66) is called the Brownian motion on the circle S 1 starting at x0 . The stochastic differential equation can also be written as dX t D b.Xt / dt C .Xt / dW t ; with the drift and diffusion given by
x1 =2 ; b.x/ D x2 =2
x2 .x1 ; x2 / D : x1
182
8 The Stochastic Analysis Method
Then T D
x1 x2 ; x12
x22 x1 x2
and hence the generator of the process Xt is AD
X @ 1X @2 . T /ij C bj .x/ 2 @xi @xj @xj i;j
j
D
i 1h 2 2 x2 @x1 C x12 @2x2 2x1 x2 @2x1 x2 x1 @x1 x2 @x2 2
D
1 .x1 @x2 x2 @x1 /2 : 2
In polar coordinates x1 D r cos , x2 D r sin , we have A D 12 @2 , which is the Laplace operator on S 1 . The heat kernel of A is given by the transition density of the process Xt D cos.0 C Wt /; sin.0 C Wt / : The distribution function given that the Brownian motion on the circle starts from 0 at t D 0 is F . j 0 / D P .0 C Wt C 2n < ; 8n 2 Z/ X X D P .Wt 2n C 0 / D FWt .2n C 0 /; n2Z
n2Z
and then the transition density is pt .0 ; / D
X 1 .2nC0 /2 d 2t e F .j0 / D p d 2 t n2Z
D p D p
1
X
2 t
n2Z
1 2 t
e
1
e 2t
.0 /2 2t
4n2 2 C.0 /2 C4n.0 /
X
e
2n2 2 t
D p
2 t
2n.0 / t
n2Z
0 1
e
e
.0 /2 2t
@1 C 2
X n1
e
2 2 2n t
1 2n. 0 / A cosh : t
This series can also be expressed as a theta-function, see (3.11.35).
8.15 An Example of a Heat Kernel
183
8.15 An Example of a Heat Kernel The following example follows [25]. Consider the stochastic differential equation dX t D
q
1C
Xt2
1 C Xt 2
dt C
q
1 C Xt2 dW t ;
X0 D x0 :
The drift and diffusion coefficients are given by b.x/ D p 1 C x 2 . The following estimations hold: p p j.x/ .y/j D j 1 C x 2 1 C y 2 j D p D p
p
(8.15.67)
1 C x 2 C 12 x, .x/ D
jx 2 y 2 j p 1 C x2 C 1 C y 2
jx C yj jx yj < 2jx yj; p 1 C x2 C 1 C y 2
since jx C yj jxj jyj p p p p Cp p 2 2 2 2 2 1Cx C 1Cy 1Cx C 1Cy 1 C x C 1 C y2 < p
jxj
Cp
jyj
< 1 C 1; 1 C x2 1 C x2 p p 1 jb.x/ b.y/j j 1 C x 2 1 C y 2 j C jx yj 2 5 1 < 2jx yj C jx yj D jx yj: 2 2 Hence jb.x/ b.y/j C j.x/ .y/j
9 jx yj; 2
8x; y 2 R;
and by Theorem 8.2.2 it follows that the stochastic equation (8.15.67) has a unique solution. Therefore, instead of solving the equation, it suffices to guess a solution. Let Ut D c Ct CWt , where c D sinh1 .x0 /. Then Xt D sinh.Ut /, with X0 D x0 , and by Ito’s formula we have 1 sinh Ut .d Ut /2 2 1 D cosh.Ut / .dt C dW t / C sinh.Ut / dt 2 1 D cosh.Ut / C sinh.Ut / dt C cosh.Ut / dW t 2
dX t D cosh.Ut / d Ut C
184
8 The Stochastic Analysis Method
q
q 1 1 C sinh2 .Ut / C sinh.Ut / dt C 1 C sinh2 .Ut / dW t 2
q q 1 D 1 C Xt2 C Xt dt C 1 C Xt2 dW t : 2 D
Hence Xt D sinh.c C t C Wt /;
c D sinh1 .x0 /
(8.15.68)
is the solution of (8.15.67). The generator of the aforementioned process Xt is the operator p 1 d2 x d A D .1 C x 2 / 2 C 1 C x2 C : (8.15.69) 2 2 dx dx The conditional distribution function of Xt is FXjX0 .xjx0 / D P .Xt x j X0 D x0 / D P .sinh.Ut / x j U0 D sinh1 x0 / D P .c C t C Wt sinh1 x j x D sinh1 x0 / D P .Wt sinh1 x c t j c D sinh1 x0 / D P .Wt sinh1 x sinh1 x0 t/ Z sinh1 xsinh1 x0 t 1 u2 D e 2t du; p 2 t 1 so the transition density of Xt is pt .x0 ; x/ D
.sinh1 xsinh1 x0 t /2 d 1 2t ; e FXjX0 .xjx0 / D p dx 2 t
t > 0: (8.15.70)
We may further transform this formula by using
q p sinh1 x sinh1 x0 D ln x C 1 C x 2 ln x0 C 1 C x02 p x C 1 C x2 ; q D ln x0 C 1 C x02 p p x C 1 C x2 1 C x2 1 1 2 2 xC .sinh x sinh x0 t/ D ln 2t ln C t 2: q q 2 2 x0 C 1 C x0 x0 C 1 C x0
8.16 A Two-Dimensional Case Fig. 8.7 The transition density as a function of x, with x0 D 0 in the cases: t D 1, t D 0:5 and t D 0:25. For t small, the graph tends to the Dirac distribution centered at x0 D 0
185
0.25
1
0.5
Then the transition density (8.15.70) becomes p p 1 t 2 xC 1Cx 2 1 C x 2 2t ln x0 Cp1Cx2 2 0 e ; pt .x0 ; x/ D p q 2 t x0 C 1 C x 2 0
1
xC
t > 0: (8.15.71)
To conclude, the heat kernel of the operator (8.15.69) is given by (8.15.71). Its graph is given by Fig. 8.7.
8.16 A Two-Dimensional Case Let W1 .t/ and W2 .t/ be two independent Brownian motions and consider the random process Xt D .X1 .t/; X2 .t// given by X1 .t/ D W1 .t/ C W2 .t/ C x10 ; X2 .t/ D sinh W2 .t/ C c ; c D sinh1 .x20 /: The process starts at X0 D .x10 ; x20 / D x0 . The associated stochastic differential equation is dX 1 .t/ D dW 1 .t/ C dW 2 .t/; 1 dX 2 .t/ D cosh W2 .t/ C c dW 2 .t/ C X2 .t/ dt 2 p 1 D 1 C X2 .t/2 dW 2 .t/ C X2 .t/ dt: 2
186
8 The Stochastic Analysis Method
In matrix form this becomes
1p 1 0 dX1 .t/ dW 1 .t/ D 1 dt C ; X .t/ dX2 .t/ dW 2 .t/ 0 1 C X2 .t/2 2 2 with the coefficients b.x/ D
0 ; 1 x 2 2
D
0
0
!
1q 1
1 C x22
;
T D @q
q 1
1 C x22
1 C x22 1 C x22
1 A:
Following relation (8.3.19), the generator of Xt is given by AD
X @ 1X @2 . T /ij C bj .x/ 2 @xi @xj @xj i;j
0
j
q
1
D
2 1 1 C x22 1@ A @ q C 2 @xi @xj 1 C x22 1 C x22
D
1 2
0 .@x1 ; @x2 / 1 x 2 2
q 2
@2 1 @2 2 @ 2 C 2 1 C x C .1 C x / C x2 @x2 : 2 2 2 2 @x1 @x2 2 @x1 @x2
In order to find the heat kernel of the operator A, we need to find the transition density of the process Xt D .X1 .t/;X2 .t//. This can be obtained from the transition density of the Brownian motion W1 .t/; W2 .t/ using formula (8.6.23) and the technique presented in Sect. 8.6, item 2. The transformation x1 D u1 C u2 C x10 ; x2 D sinh.u2 C c/;
c D sinh1 .x20 /;
has the inverse u1 D .x1 x10 / sinh1 .x2 / sinh1 .x20 / ;
(8.16.72)
u2 D sinh1 .x2 / sinh1 .x20 /;
(8.16.73)
with the Jacobian det
@.u1 ; u2 / 1 6D 0: D q @.x1 ; x2 / 1 C x22
8.16 A Two-Dimensional Case
187
The density function of Xt is @.u1 ; u2 / pt .x0 ; x/ D fX jX.0/Dx0 .xjx0 / D fW1 ;W2 u1 .x0 ; x/; u2 .x0 ; x/ det @.x1 ; x2 / @.u1 ; u2 / D fW1 u1 .x0 ; x/ fW2 u2 .x0 ; x/ det @.x1 ; x2 / 2 u2 u 1 1 1 1 2 e 2t p e 2t q D p 2 t 2 t 1 C x22 D
juj2 1 1 q e 2t ; 2 t 1 C x 2 2
with u given by (8.16.72)-(8.16.73). A computation provides juj2 D u21 C u22
D .x1 x10 /2 2.x1 x10 / sinh1 .x2 / sinh1 .x20 / 2 C2 sinh1 .x2 / sinh1 .x20 / q x2 C 1 C x22 2 D .x1 x10 /2 2.x1 x10 / ln C 2 sinh1 .x2 / sinh1 .x20 / q x20 C 1C.x20 /2
and hence
e
2 juj 2t
" D
0 q 1 # x1 x 2 t x2 C 1 C x22 1 0 2 1 1 1 0 q e 2t .x1 x1 / t sinh .x2 /sinh .x2 / : x20 C 1 C .x20 /2
We have arrived at the following result. Proposition 8.16.1. The heat kernel of the operator 1 AD 2
q 2
@2 1 @2 2 @ 2 C 2 1 C x2 C .1 C x2 / 2 C x2 @x2 2 @x1 @x2 2 @x1 @x2
is given by 0 q 1 # x1 x 2 t C 1 C x x 2 2 1 1 q pt .x0 ; x/ D q 2 t 1 C x 2 x 0 C 1 C .x 0 /2 2 2 2 2 1 0 2 1 1 1 0 e 2t .x1 x1 / t sinh .x2 /sinh .x2 / ; t > 0:
"
188
8 The Stochastic Analysis Method
8.17 Kolmogorov’s Operator One of the classical examples of heat kernels obtained by using the transition density of an associated stochastic process is the degenerate operator of Kolmogorov on R2 (see [80, 81]): 1 L D @2x1 x1 @x2 : (8.17.74) 2 Since
0 10 1 0 T ; bD D ; ; D x1 00 0 0 the operator L is the generator of the following two-dimensional Ito diffusion: dX t D b dt C dW
dW 1 .t/ 1 0 0 dt C ; D 0 0 X1 .t/ dW 2 .t/ where W1 .t/ and W2 .t/ are independent Brownian motions. If Xt D .X1 .t/; X2 .t//, then dX 1 .t/ D dW 1 .t/; dX 2 .t/ D X1 .t/ dt; with the solution X1 .t/ D x10 C W1 .t/; Z t Z t 0 X2 .t/ D x20 x1 C W1 .s/ ds D x20 x10 t W1 .s/ ds; 0
0
where X0 D .x10 ; x20 /. We shall use the expected value method presented in item 6 of Sect. 8.6. Denote W .t/ D W1 .t/: pt .x0 ; x/ D EŒı X1 .t/ x1 ı X2 .t/ x2 Z t D E ı.x1 x10 W .t//ı.x20 x2 x10 t W .s/ ds/ 0 Z Rt 1 0 i.x20 x2 x10 t 0 W .s/ ds/ d e D E ı.x1 x1 C W .t// 2 Z Rt 1 0 0 e i .x2 x2 x1 t / EŒı.x10 x1 C W .t//e i 0 W .s/ ds d: D 2 (8.17.75)
8.18 The Operator 12 x12 @2x2
189
Let v D v.x10 ; x1 ; tI / D EŒı.x10 x1 C W .t//e i
Rt 0
W .s/ ds
:
According to the Feynman–Kac formula (see Theorem 8.23.1) with the linear potential V .x/ D ax and a D i, the element v verifies the initial boundary problem @t v D
1 2
@2x1 ax1 v;
vjt D0 D ıx 0 .x/: 1
Theorem 3.15.1 yields vD p
1 2 t
e
.x1 x 0 /2 1 2t
1 2 3 0 1 2 a.x1 Cx1 /t C 24 a t
.x1 x 0 /2 i 1 t3 2 0 1 e 2t 2 .x1 Cx1 /t 24 : D p 2 t
Substituting back in (8.17.75) yields .x1 x 0 /2 1 1 1 pt .x0 ; x/ D e 2t p 2 2 t
Z
t3
e 24
2 CiŒx x 0 1 .x Cx 0 /t 2 1 2 2 1
d
r 0 /2 .x1 x1 1 24 243 Œx2 x20 12 .x1 Cx10 /t 2 1 2t e D e 4t p 2 2 t t3 p 0 2 1 / 6 Œx x 0 1 .x Cx 0 /t 2 3 .x1 x 2 1 2 2 1 2t t3 D e ; t > 0; t2
which is the heat kernel of the operator (8.17.74).
8.18 The Operator 12 x21 @x22 The operator A D 12 x12 @2x2 is the generator of the two-dimensional Ito diffusion dX 1 .t/ D 0; dX 2 .t/ D X1 .t/ dW t ;
t 0;
where Wt is a one-dimensional Brownian motion. Solving yields X1 .t/ D x10 ; X2 .t/ D x20 C x10 Wt :
190
8 The Stochastic Analysis Method
Assume x10 6D 0. Then the transition distribution function is FXjX0 .x1 ; x2 j x10 ; x20 / D P X1 .t/ x1 ; X2 .t/ x2 j X1 .0/ D x10 ; X2 .0/ D x20 D P .x10 x1 ; x20 C x10 Wt x2 /
x2 x20 0 D P .x1 x1 /P Wt x10 Z D Hx 0 .x1 / 1
x2 x 0 2 x0 1
1
where
Hx0 .x1 / D 1
p
1 2 t
u2
e 2t du;
1; x1 x10 , 0; x1 < x10 ,
is the Heaviside function. The transition density is obtained by differentiating the previous transition density: @2 Ft .x1 ; x2 j x10 ; x20 / @x1 @x2 Z x2 x20 0 @ @ 1 u2 x1 Hx 0 .x1 / p e 2t du D @x1 1 @x2 1 2 t 2 x2 x 0 1 2 1 2t x0 1 D ıx 0 .x1 / 0 p : e 1 x1 2 t
pt .x0 ; x/ D
Hence the heat kernel of A D 12 x12 @2x2 is given by 1 2t 1 K.x0 ; xI t/ D ıx 0 .x1 / ˝ 0 p e 1 x1 2 t
0 x2 x2 0 x 1
2 ;
for any initial point .x10 ; x20 / with x10 6D 0. If x10 D 0, the transition distribution is FXjX0 .x1 ; x2 j 0; x20 / D P X1 .t/ x1 ; X2 .t/ x2 j X1 .0/ D 0; X2 .0/ D x20 D P .0 x1 ; x20 x2 / D H0 .x1 / Hx 0 .x2 /; 2
so by taking the derivative, we get pt .0; x/ D ı0 .x1 /ıx 0 .x2 /: 2
8.19 Grushin’s Operator
191
8.19 Grushin’s Operator The Grushin operator G D
1 2 .@ C x12 @2x2 / 2 x1
is a sub-elliptic operator which is the generator for the Ito diffusion X1 .t/; X2 .t/ given by dX 1 .t/ D dW 1 .t/; dX 2 .t/ D X1 .t/ dW 2 .t/; with X1 .0/ D x10 , X2 .0/ D x20 and W1 .t/, W2 .t/ one-dimensional independent Brownian motions starting at zero. Solving yields X1 .t/ D x10 C W1 .t/; Z t X1 .s/ dW 2 .s/; X2 .t/ D x20 C 0
and the heat kernel of G is given by the following transition probability (see item 6 of Sect. 8.6): pt .x0 ; x/ D EŒı.x1 X1 .t//ı.x2 X2 .t// Z 1 D E ı.x1 X1 .t// e i.x2 X2 .t // d 2 Z i h Rt 1 0 D e i.x2 x2 / E ı.x1 X1 .t//e i 0 X1 .s/dW 2 .s/ d: (8.19.76) 2 The following idea of computing comes from [31]. The expectation operator E is the Wiener integral over W1 .t/ and W2 .t/. Since X1 .t/ is independent of W2 .t/, we may integrate first with respect to W2 .t/ and get h Rt i R 1 2 t 2 EW2 e i 0 X1 .s/ dW 2 .s/ D e 2 0 X1 .s/ ds ; see (8.6.33). Substituting back in (8.19.76) yields 1 pt .x0 ; x/ D 2 D
1 2
Z Z
0
1
e i.x2 x2 / EŒı.x1 X1 .t//e 2 0
e i.x2 x2 / v.x10 ; x1 ; tI / d;
2
Rt
0
X12 .s/I ds
d (8.19.77)
192
8 The Stochastic Analysis Method
where here E denotes the expectation with respect to W1 .t/ and h i R 1 2 t 2 v D v.x10 ; x1 ; tI / D E ı.x1 X1 .t//e 2 0 X1 .s/ ds i h R 1 2 t 0 2 D E ı.x1 .x10 C W1 .t///e 2 0 .x1 CW1 .s// ds : From the Feynman–Kac formula(see Theorem 8.23.1), with the quadratic potential V .x/ D 12 2 x 2 , the element v verifies the initial boundary problem @t v D
1 1 @ 2 2 x12 v; 2 x1 2
vjt D0 D ıx0 .x1 /: 1
From Theorem 3.16.1, we get vD p
s
1 2 t
t t 1 e 2t sinh.t / sinh.t/
h
x12 C.x10 /
2
cosh. t /2x1 x10
i
; t > 0:
Substituting back (8.19.77) yields 1 1 pt .x0 ; x/D p 2 2 t
Z s
1 t i.x2 x20 / 2t e sinh.t/
t sinh.t / Œ
2 x12 C.x10 / cosh. t /2x1 x10
d;
which, after the substitution D t, becomes 1 pt .x0 ; x/ D .2 t/3=2
Z r
1 e t f .x0 ;x;t I / ; sinh
with f .x0 ; x; tI / D i.x2 x20 /
i h 2 2 x1 C .x10 / coth 2x1 x10 sech : 2
8.20 Squared Bessel Process The operator 1 2 x > 0; x@ C @x ; 2 x is the generator of the following squared Bessel process: dX t D dt C
p Xt =2 dW t ;
t 0;
(8.20.78)
8.22 Limitations of the Method
193
with > 1=8. The heat kernel of (8.20.78) with zero boundaries at x D 0; 1 is given by the transition density of the aforementioned process in terms of the modified Bessel function(see [29]): K.x0 ; xI t/ D
p e .xCx0 /=.t =4/ x 2 1=2 I4 1 .8 xx0 =t/; t=4 x0
x; x0 > 0; t > 0: (8.20.79)
8.21 CIR Processes The operator 1 2 x@ C .0 1 x/@x ; 2 x
x > 0;
(8.21.80)
is the generator of the CIR process6 dX t D .0 1 x/dt C
p Xt =2 dW t ;
t > 0;
with 0 ; 1 > 0, constants. These types of processes have been used to model stochastic interest rates in finance: see [37]. The heat kernel of the operator (8.21.80), given as the transition density of the CIR process, is given for instance in [29]: !2 0 12 xe 1 t Cx0 41 e 1 t x 1 t 4
K.x0 ; xI t/ D t e e 1 e 1 t 1 e 1 1 x0 8 p 1 I4 0 1 t xx0 e 1 t ; e 1 1 x0 ; x > 0, t > 0.
8.22 Limitations of the Method The stochastic method presented in this chapter works for a large class of operators. However, it does not apply to all operators. There are several limitations of the method that will be discussed in the following. These limitations can be overcome by using some of the methods presented in the other chapters.
6
CIR stands for Cox, Ingersoll and Ross.
194
8 The Stochastic Analysis Method
1. Explicit solution but the transition density is hard to get. Consider the case of the one-dimensional linear noise with drift given by dX t D r dt C ˛Xt dW t ;
r; ˛ 2 R:
(8.22.81)
Given x0 2 R, Theorem 8.2.2 ensures the existence and uniqueness of a process Xt satisfying (8.22.81) with the initial condition X0 D x0 . According to Theorem 8.3.1, the generator associated with the process Xt is given by AD
@ 1 2 2 @2 Cr : ˛ x 2 2 @x @x
(8.22.82)
Without loss of generality, we may assume ˛ D 1. Equation (8.22.81) becomes dX t Xt dW t D r dt: 1
Multiplying by the integrating factor t D e Wt C 2 t yields t dX t t Xt dW t D rt dt:
(8.22.83)
We shall show that the left side is exact. Using the product rule and Ito’s formula yields dt D t dW t C t dt; dt dX t D t Xt dt; since dt2 D dt dW t D dW t dt D 0 and dW t dW t D dt. Then d.t Xt / D dt Xt C t dX t C dt dX t D t Xt dW t C t Xt dt C t dXt t Xt dt D t dX t t Xt dW t ; which is the left side of (8.22.83). Integrating in d.t Xt / D rt dt yields Z t Xt D 0 X0 C r
t
s ds ” Z t 1 Xt D 1 x C r s ds ” 0 t t 0 Z t 1 1 1 Xt D e Wt 2 t x0 C re Wt 2 t e Ws C 2 s ds; 0
0
(8.22.84)
8.23 Operators with Potential and the Feynman–Kac Formula
195
where 0 D 1 and X0 D x0 . If r D 0, then Xt has a lognormal distribution and this case was treated in Sect. 8.10. However, in the case r 6D 0, the transition density of the process (8.22.84) is not easy to obtain. 2. The stochastic differential equation has a unique solution, but it cannot be solved explicitly. Consider the following equation: dX t D dt C .cos Xt / dW t ; X0 D x0 : It is obvious that the existence and uniqueness condition of Theorem 8.2.2 applies and the equation has a unique solution Xt . However, a closed-form solution is hard to obtain. Hence the stochastic method is not suitable for finding the heat kernel of the operator AD
d 1 d2 .cos x/2 2 C : 2 dx dx
3. The stochastic differential equation might have multiple solutions. This is the case of equation 1=3
dX t D 3Xt
2=3
dt C 3Xt
dW t ;
(8.22.85)
X0 D 0; considered in Sect. 8.2. A direct application of the algorithm in Sect. 8.5 will not produce the heat kernel for AD
9 4=3 d 2 d x C 3x : 2 2 dx dx
The aforementioned operator is singular at x D 0 and each solution of (8.22.85) has a contribution to the kernel that starts at the origin.
8.23 Operators with Potential and the Feynman–Kac Formula In the previous sections we computed the heat kernel for a number of operators of the form X 1X @2 @ AD . T /ij C bk (8.23.86) 2 @xi @xj @xk i;j
k
196
8 The Stochastic Analysis Method
associated with the n-dimensional Ito diffusion process dX t D b.Xt / dt C .Xt / dW.t/;
(8.23.87)
which is supposed to satisfy the existence and uniqueness conditions of Theorem 8.2.2. In this section we deal with an extension of Kolmogorov’s backward equation (see Theorem 8.4.1), for operators of the type A V .x/, where V .x/ is a potential function. The first formula of this type deals with the n-dimensional Laplacian; it was found by Kac [75] in 1949. The result states that the equation @v 1 D v V .x/v; @t 2
t > 0;
v.0; x/ D f .x/; has the solution given by h i Rt v.t; x/ D E x f .Xt / e 0 V .Xs / ds ;
(8.23.88)
where E x is the conditional expectation given that the process Xt is a Brownian motion starting at x; i.e., Xt D Wt C x with W0 D 0. The extension of the aforementioned result to any operator of the type (8.23.86) is given below. For a proof the reader can consult the references [83, 95]. Theorem 8.23.1 (Feynman–Kac formula). Let Xt be the continuous solution of the stochastic differential equation (8.23.87) and let A be its generator. Let f 2 C02 .Rn / and V .x/ be lower-bounded continuous functions on Rn . Then the function v.t; x/ D E x Œf .Xt / e
Rt
0
V .Xs / ds
;
t 0; x 2 Rn ;
(8.23.89)
satisfies the Cauchy problem @v D Av V v; @t
t > 0;
v.0; x/ D f .x/: Here E x denotes the conditional expectation given that X0 D x. We note that if V D 0, then the Feynman–Kac formula becomes the Kolmogorov forward equation result; see Theorem 8.4.1. In spite of its simplicity, formula (8.23.89) cannot be worked out explicitly unless the potential V .x/ is of a very particular form. Its importance is more of a theoretical
8.23 Operators with Potential and the Feynman–Kac Formula
197
value, since it provides an integral representation of the heat equation even in cases that do not have closed-form solutions. For example, in the case of quartic potential, the solution of the equation 1 @v D @2x v x 4 v; @t 2 v.0; x/ D f .x/ is v.t; x/ D E x Œf .Wt C x/ e
Rt
4 0 .Ws Cx/
t > 0;
ds
;
t 0; x 2 R:
It is worth noting that the Feynman–Kac formula can also be expressed in terms of path integrals; see Chap. 7.
Part II
Heat Kernel on Nilpotent Lie Groups and Nilmanifolds
Chapter 9
Laplacians and Sub-Laplacians
9.1 Sub-Riemannian Structure and Heat Kernels We have seen in the previous chapters how an elliptic operator can be associated in a natural way with a geometric Riemannian structure. In a similar way sub-elliptic operators arise from similar structures, called sub-Riemannian structures, which will be discussed next. References for sub-Riemannian manifolds are [27] and [92].
9.1.1 Sub-Riemannian Structure Let M be a smooth manifold. We shall assume M without boundary throughout this chapter. A sub-bundle H of the tangent bundle T .M / is called nonholonomic if the vector fields of the sub-bundle H and a finite number of iterations of their Lie brackets span the whole tangent space at each point. The sub-bundle H is sometimes called the horizontal distribution. Definition 9.1.1. A connected manifold M is called a sub-Riemannian manifold if it has a nonholonomic sub-bundle H with a positive definite, nondegenerate metric on H. The aforementioned nonholonomic property of the sub-bundle H is also called the bracket generating property or H¨ormander’s condition [65]. Sometimes it is also called Chow’s condition [35]. Next we provide the physical interpretation for the previous definition. The coordinates of a physical system can be represented as a point on a manifold, called the configuration space. The states of the system, which consist of pairs of coordinates and momenta, form the cotangent bundle of a manifold, called the phase space. The physical system is Riemannian if the states “can move in any direction”; i.e., there are curves in the configuration space starting at a given point pointing in all directions. These systems have the freedom to move in all directions. However, there are examples of mechanical systems with nonholonomic constraints which are not of the type just described. A rolling ball on a surface or a O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 9, c Springer Science+Business Media, LLC 2011
201
202
9 Laplacians and Sub-Laplacians
penny rolling on a plane are two examples of these kind of systems. In this case the system is sub-Riemannian, which means that the structure moves only along horizontal directions, i.e., directions tangent to the sub-bundle H. Given an initial configuration and a final configuration, it is natural to ask whether it is always possible for a sub-Riemannian system to move from the initial position to the final position in the configuration space. Everyone has faced this problem while parking a car. Starting from a given position, the car has to move into a given final parking position. Since this motion cannot be done by only one smooth maneuver, the horizontal curve on the associated sub-Riemannian manifold will be just piecewise smooth. The following global connectedness result holds for subRiemannian manifolds: Theorem 9.1.2 (Chow’s theorem, [35]). Any two points of a sub-Riemannian manifold can be joined by a piecewise smooth horizontal curve, i.e., a curve tangent to the horizontal distribution H. Gromov showed that in some particular cases the “piecewise” attribute can be omitted, but in general it cannot. There are examples of physical systems which have states that cannot be connected by a horizontal curve. This is the case with thermodynamical systems where the horizontal curves correspond to adiabatic processes, i.e., processes along which there is no heat exchange. J. Charney was the first to note that there are always states which cannot be joined by an adiabatic curve, i.e., there are states which cannot be connected by a horizontal curve; see, for instance, [27], Chap. 3. We shall list below several basic definitions related to a sub-Riemannian structure. Definition 9.1.3. 1. The dimension dim T .M /=H is called the codimension of the sub-Riemannian structure H. 2. Let .M; T .M // be the space of all smooth vector fields on M; and set 0 .M; H/ D .M; H/, the space of smooth vector fields on M taking values in H. For any k 2 N, let k .M; H/ D 0 .M; H/ C Œ 0 .M; H/; k1 .M; H/: For any point x 2 M , we denote by x` .H/ the values of the tangent vectors in ` .M; H/ at the point x 2 M: If for some k > 0, xk .H/ = Tx .M / but xk1 .H/ 6D Tx .M /, then we say the sub-Riemannian structure is of step k C 1 at the point x. The sub-Riemannian structure is step k on the domain U if it is step k at each point x 2 U: 3. If there is an ascending sequence of sub-bundles fH` g of T .M / H D H0 H1 H1 Hk D T .M /
9.1 Sub-Riemannian Structure and Heat Kernels
203
such that ` .M; H/ D .M; H` /, then the sub-Riemannian structure is called regular. We have denoted by .M; H` / the space of smooth vector fields taking values in H` . Note that in this case all sub-bundles H` are nonholonomic. 4. A sub-Riemannian structure H is called minimal if there is no sub-bundle in H which satisfies the bracket generating property. Two additional conditions on the nonholonomic sub-bundle H shall be required throughout the book. (Sub-1) The horizontal sub-bundle H is trivial. This means that we can consider globally defined, nowhere-vanishing smooth vector fields fXi g, i D 1; : : : ; dim H, satisfying H¨ormander’s condition, which guarantees the hypoellipticity of the following second-order operator: dim XH
Xi Xi :
i D1
The second assumption is (Sub-2) There is a volume form dVM with respect to which the vector fields fXi g are all skew-symmetric. If the aforementioned conditions hold, we say that the sub-Riemannian structure is “trivializable,” or is of the strong sense. It is worth noting that even if the typical examples of sub-Riemannian manifolds are contact manifolds, they are not always trivializable structures. Furthermore, the associated sub-elliptic operator sub D
dim XH
Xi 2
i D1
is called the sub-Laplacian. Our next main purpose is to construct heat kernels for sub-Laplacians on sub-Riemannian manifolds. A special role will be played by the Grushin-type operators that will be treated in Sect. 10.3. Most of the operators in this chapter are sub-Laplacian operators on nilpotent Lie groups and their quotient spaces. It is a rather convenient assumption to consider the metric on the sub-bundle H in such a way that the vector fields fXi g are orthonormal.
9.1.2 Heat Kernel of the Sub-Laplacian and Laplacian Under the aforementioned conditions (Sub-1) and (Sub-2), the sub-Laplacian sub is symmetric and positive with respect to the volume form dM V , whose existence is assumed by the latter condition. If the metric is extended to the entire tangent
204
9 Laplacians and Sub-Laplacians
space, then the Laplacian is symmetric and positive with respect to the Riemannian volume form. The volume element that we assumed to exist for the sub-Riemannian manifold will coincide in this case with the Riemannian volume form, up to a multiplicative constant, as we shall see in the following examples. Moreover, we have Theorem 9.1.4 ([105]). If a sub-Riemannian metric on a noncompact manifold can be extended to a complete Riemannian metric, then the sub-Laplacian is essentially self-adjoint on the space of smooth functions with compact support. In all the cases treated in the sequel, we encounter left invariant operators on nilpotent Lie groups, for which the volume form dV M postulated by (Sub-2) is a Haar measure, and the conditions (Sub-1) and (Sub-2) are satisfied with respect to this measure. In this case, a left invariant sub-Riemannian metric defined on a left invariant nonholonomic sub-bundle can be extended to a left invariant complete metric on the entire space. This is the reason why in the following cases we shall not distinguish between the sub-Laplacian sub , respectively the Laplacian , both defined on C01 .M /, and their unique self-adjoint realizations on the space L2 .M; dM V /. Invoking the spectral decomposition theorem Z sub D
0
1
dE ;
Q g/ is the with fE g the spectral measure, we know that the heat kernel K.tI g; kernel distribution of the bounded operator e t sub D
Z 0
1
e t dE W L2 .M / ! L2 .M /;
t > 0:
This kernel is smooth because 1. sub is hypoelliptic 2. For any integer k, the operator subk ı e t sub D e t sub ı subk is defined 1 T on L2 .M / and e t sub maps continuously L2 .M / to domain of subk D kD1
C 1 .M / 3. e t sub can be extended to the entire space of distributions D0 .M / on M from L2 .M /. We have e t sub D e t =2sub ı e t =2sub as a map from D0 .M / to C 1 .M / Of course, the previous arguments are also valid for the case of Laplacians and the heat kernel for Laplacians on C 1 .RC M M /: We shall next review the purpose of our future exposition. (1) We are interested in the spectral decomposition of the sub-Laplacian and Laplacian in a multiplication operator form. This means finding a measure space .X; dm/, a (positive) function ' on X, and a unitary transformation
9.1 Sub-Riemannian Structure and Heat Kernels
205
U W L2 .M / ! L2 .X; dm/ such that sub D U1 ı M' ı U; where M' W L2 .X; dm/ 3 f 7! 'f 2 L2 .X; dm/ is the multiplication operator by the function '. (2) We shall provide an explicit construction for the heat kernel K.tI x; y/ of the sub-Laplacian sub (and also Laplacian). This means finding a solution K.tI x; y/ for @ sub C K.tI x; y/ D 0; @t
Z lim
t !0 M
K.tI x; y/f .y/dM V .y/ D f .x/:
In the case of nilpotent Lie groups, using the (left) invariance of the heat kernel K.tI g x; g y/ D K.tI x; y/;
8g; x; y 2 M;
it follows that the heat kernel K.tI x; y/ is of the form K.tI x; y/ D kt .y 1 x/; with kt .x/ a smooth function on RC M: (3) We shall find explicit expressions for the Green functions and fundamental solutions of the operators sub and . (4) We aim to obtain explicit expressions for the heat kernel of the Grushin-type operators defined in Sect. 10.3. There are many papers in the literature treating the hypoelliptic operators on nilpotent Lie groups (for instance, [15, 100] and papers cited therein). However, our concern here is to construct the heat kernel in terms of special functions. We shall use Hermite functions and hyperbolic functions for the two-step cases, while for the three-step cases the computation might be achieved by using elliptic functions and other special functions; see [21,22,58]. However, until now there are no explicit formulas for Laplacians and sub-Laplacians on nilpotent Lie groups with step greater than three (including the Engel group). Therefore, we are limited in our endeavor of finding- closed form formulas only for the two-step nilpotent Lie groups. Remark 9.1.5. If we have a spectral decomposition of (or sub ) in the multiplication operator form, then the heat kernel is expressed as e t D U ı e t M' ı U1 : This is similar to the case of Euclidean spaces where the operator U is the Fourier transform. So (1) leads to (2), but not the other way around.
206
9 Laplacians and Sub-Laplacians
Integrating the heat kernels, we obtain Green functions. If the heat kernel is K.tI x; y/ 2 C 1 .RC M M /; then the Green function G.x; y/ can be expressed as Z M
Z G.x; y/f .y/dM V .y/ D
0
1
Z G
K.tI x; y/f .y/dM V .y/ ^ dt:
Even if we have an explicit expression for the heat kernel, this does not provide the spectral decomposition of the (sub-)Laplacian in a multiplication operator form directly. For this purpose we are still required to find a measure space .X; dm/, a unitary transformation between L2 .G/ and L2 .X; dm/, and a positive function on X, with respect to which the Laplacian (sub-Laplacian) can be expressed as a multiplication operator.
9.1.3 The Volume Form In the definition of the sub-Riemannian manifold we have assumed that there exists a volume form dM V .x/ such that each vector field Xk is skew-symmetric with respect to the inner product induced by this volume form: Z
Z M
Xk .f /.x/g.x/dM V .x/ D
M
f .x/Xk .g/.x/dM V .x/;
f or g 2 C01 .M /:
For regular sub-Riemannian manifolds, we can construct a volume form in an intrinsic way (see [92] for Popp’s measure) based on the inner product defined on the nonholonomic sub-bundle, even if the vector fields taking values in the nonholonomic sub-bundle are not necessarily skew-symmetric. For the sake of simplicity, in this section we shall explain the construction of the volume form just for the case of two-step regular sub-Riemannian manifolds. Let X 2 Hx and Y 2 Hx be tangent vectors which are extended to the vector fields XQ and YQ on M taking values in H. Then the value of the bracket ŒXQ ; YQ 2 Tx .M / (mod Hx ) does not depend on the extensions XQ and YQ . Thus we have a well-defined bundle map W H ˝ H ! T .M /=H: The map is surjective by the bracket generating assumption .M; H/ C Œ.M; H/; .M; H/ D .T .M //: Now we can introduce a metric on T .M /=H such that it is isometric with the orthogonal complement of the kernel of the map . Note that H ˝ H has a tensorial
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
207
product metric from H. Then we have an exact sequence of vector bundles with metrics except total space T .M /: 0 ! H ! T .M / ! T .M /=H ! 0:
(9.1.1)
Since the isomorphism dim ^H
H ˝
dim M^ dim H
T .M /=H
dim ^M
Š
T .M /
does not depend on the choice of the splitting of the previous exact sequence, with dim VM the help of the metrics on H and T .M /=H, we can trivialize T .M / by choosing dual orthonormal bases of these two bundles at each point. This provides a volume form on M: Remark 9.1.6. In the following examples the vector fields trivializing a nonholonomic sub-bundle are skew-symmetric with respect to the volume form constructed by this way (up to a multiplicative constant).
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups In this section we shall explain the basic properties of nilpotent Lie groups and subRiemannian structures on such Lie groups. We shall also deal with several examples of nilpotent Lie groups and sub-Laplacians on them for which we shall construct heat kernels in later sections.
9.2.1 Nilpotent Lie Groups Let G be a nilpotent Lie group with the Lie algebra g. This means that if we put g1 D Œg; g, g2 D Œg; g1 , : : :, gk D Œg; gk1 , : : :, then there exists N 2 N such that gN D f0g. For an elementary treatment of nilpotent Lie groups, we refer the reader to [110]. We begin by reminding the reader of one of the most fundamental results. Theorem 9.2.1. Let G be a connected and simply connected nilpotent Lie group. Then the exponential map exp W g ! G is a diffeomorphism. Because of this fact, we can work on the linear space g using the Campbell– Hausdorff formula exp X exp Y D exp.X C Y C 1=2ŒX; Y C 1=12ŒŒX; Y ; Y X C : : :/I X; Y 2 g;
208
9 Laplacians and Sub-Laplacians
which says that we can introduce the group multiplication in the Lie algebra g through the identification exp W g ! G. The group law is defined on g as in the following: If exp X exp Y D exp Z; then we define X Y D Z; with Z given by Z D X C Y C 1=2ŒX; Y C 1=12ŒŒX; Y ; Y X C : : : (finite sum): Q : : : ; / 2 G Š g if we consider them We shall use the notations g; h; : : : (or g; Q h; as elements of the Lie group g, and X; Y; : : : 2 g if we consider them as elements of the Lie algebra g. This way we eliminate the confusion of whether we are dealing with g as a group or as a Lie algebra. e the left invariant vector field on the Lie group G Let X 2 g, and denote by X defined by e .f /.g/ D df .g exp tX / X ; dt jt D0 e , we have the following where f 2 C 1 .g/. By the correspondence g 3 X ! X trivialization of the tangent bundle T .G/ of the group G: G g ! T .G/; e g 2 Tg .G/: .g; X / 7! X We also have a trivialization of the cotangent bundle T .G/: G g ! T .G/; .g; / 7! e g 2 Tg .G/; where the left invariant one-form e for 2 g is defined by the formula e e g / D .X /; X 2 g: g .X g Let fXi gdim i D1 be a basis of the Lie algebra g. We consider the coordinates on the Lie group G D g introduced by the correspondence
.x1 ; : : : ; xdim g / $ g D
dim Xg
xi Xi :
i D1
Then we can take the differential form dg D dx1 ^ ^ dxdim g as a Haar measure.
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
209
e as a Proposition 9.2.2. Let X 2 g. If we consider the left invariant vector field X first-order differential operator on the group G, e W C 1 .G/ ! C 1 .G/; f 7! X e .f /; X then it is skew-symmetric with respect to the inner product defined by a Haar measure. Proof. Let '; Z
2 C01 .G/. We have
e .'/.g/ X G
d' g exp.tX / .g/dg .g/dg D lim t !0 G dt jt D0 Z d g exp.tX / D lim '.g/ dg t !0 G dt jt D0 Z e . /.g/dg; '.g/ X D Z
G
since the Haar measure is left and right invariant in the case of the nilpotent Lie groups. Let 1 W g ! g=Œg; g be the projection map and let fXi g`iD1 be linearly independent elements such that f1 .Xi /g`iD1 spans the quotient space g=Œg; g. Then the subspace fXQ i g`iD1 defines a regular sub-Riemannian structure on the group G. If the group is a k-step nilpotent Lie group, then this structure is minimal, nonholonomic, and of step k.
9.2.2 The Heisenberg Group Let H3 be the three-dimensional Heisenberg group identified by R3 . The product law is given by the formula R3 R3 3 .x; y; z/; .x; Q y; Q zQ/ 7! .x; y; z/ .x; Q y; Q zQ/ D x C x; Q y C y; Q z C zQ C 1=2.x yQ yx/ Q 2 R3 : Then its Lie algebra h3 is identified by itself; i.e., the exponential map is the identity. H3 can also be realized as a real 3 3 matrix space 80 9 1 < 1 x z = H3 Š @0 1 y A I x; y; z 2 R : : ; 0 0 1 The identification is given by the map 0
1 1 x z C xy 2 H3 3 .x; y; z/ 7! @0 1 y A: 0 0 1
210
9 Laplacians and Sub-Laplacians
Remark 9.2.3. There are also other ways to realize the group H3 as a matrix space. For example, if we let 0
0 B0 X DB @0 0
1 0 0 0
0 1 0 0 0 B C 0 0 0C ; Y DB @0 0 1A 0 0 0
0 0 0 0
0 1 0 0 0 B C 0 0 1C and Z D B @0 0 0A 0 0 0
1 0 0 0
0 0 0 0
1 2 0C C; 0A 0
we have ŒX; Y D Z, and the space of matrices 80 0 ˆ ˆ
x 0 0 0
9 1 y z > > = 0 y C C I x; y; z 2 R > 0 x A > ; 0 0
is isomorphic to h3 . So the three-dimensional Heisenberg group H3 is identified by a subgroup in 4 4 matrix space as 80 1 ˆ ˆ
x 1 0 0
9 1 y z > > = 0 y C C I x; y; z 2 R : > 1 x A > ; 0 1
The exponential map is given by 0
0 B0 exp W h3 3 B @0 0
x 0 0 0
0 1 0 y z B C 0 0 y C 7! Id C B @0 0 x A 0 0 0
x 0 0 0
1 0 1 y z B C 0 y C B0 D 0 x A @0 0 0 0
x 1 0 0
1 y z 0 y C C 2 H3 : 1 x A 0 1
In the following we shall consider the three-dimensional Heisenberg group as R3 . Let X D .1; 0; 0/, Y D .0; 1; 0/ and Z D .0; 0; 1/ be the basis of the Lie e, Y e and Z e are given by algebra h3 . The left invariant vector fields X eD @ y @; X @x 2 @z eD @ Cx @; Y @y 2 @z eD @: Z @z
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
211
A sub-Riemannian structure H on H3 in the strong sense is defined by a sube Y eg. We consider a left invariant metric on H such that the bundle spanned by fX; e and Y e are orthonormal at each point. In this case the sub-Laplacian vector fields X sub is e2 C Y e2 /: sub D .X (9.2.2) Furthermore, if we consider the left invariant Riemannian metric on H3 such that e, Y e and Z e are orthonormal at each point, then the metric tensor is given by X 0
1 y xy 4 2 C 2 1 C x4 x2 A ; x2 1
2
1C y B xy4 @ 4 y 2
with the inverse matrix
0
1 B @ 0 y2
0 1 x 2
y2 1C
1
C x A: 2 x 2 Cy 2 4
Hence the Laplacian H3 is
H3
@2 @2 @2 @2 x 2 C y 2 @2 D C 2 y Cx C 1C @x 2 @y @x@z @y@z 4 @z2 e2 Y e2 Z e 2 D sub Z e2 : D X
9.2.3 Higher-Dimensional Heisenberg Groups The higher-dimensional Heisenberg group is given by H2nC1 D R2nC1 D R2n R Š Rn Rn R together with the following group law:
< x; yQ > < y; xQ > ; .x; yI z/; .x; Q yI Q zQ/ 7 ! x C x; Q y C yI Q z C zQ C 2
for any .x; yI z/; .x; Q yI Q zQ/ 2 Rn Rn R. In the aforementioned formula, < ; > denotes the standard inner product on Rn . H2nC1 is realized in the matrix space
H2nC1
80 ˆ 1 x1 x2 ˆ ˆ ˆB0 1 0 ˆ ˆ B ˆ ˆB ˆ
9 1 > xn z > > C > > 0 y1 C > > C > > 0 y2 C = C n :: C I x; y 2 R ; z 2 R > 0 : C > C > > C > > > 1 yn A > > ; 0 1
212
9 Laplacians and Sub-Laplacians
by the map 0
H2nC1
1 x1 x2 B0 1 0 B B B0 0 1 B n n D R R R 3 .x; yI z/ 7! B B0 0 0 B B :: @: 0
1 xn z C <x;y> 2 C 0 y1 C C 0 y2 C C :: C: C 0 : C C A 1 yn 0 1
The Lie algebra of H2nC1 , which can also be regarded as R2nC1 , is denoted by h2nC1 . For 1 i n, the vectors Xi D .0; : : : ; 0 ; 1; 0; 0; : : : ; 0; I 0/; „ ƒ‚ … i 1 zero Yi D .0; : : : ; 0; 0; : : : ; 0; 1; 0; : : : I 0/; ƒ‚ … „ nCi 1 zero Z D .0; : : : ; 0; 0; : : : ; 0I 1/ ƒ‚ … „ 2n zero form a basis in the Heisenberg algebra h2nC1 . The corresponding left invariant vector fields are expressed as ei D @ X @xi ei D @ C Y @yi ei D @ : Z @z
yi 2 xi 2
@ ; @z @ ; @z
(9.2.3) (9.2.4) (9.2.5)
Now we have a left invariant sub-Riemannian structure H spanned by the left ei I i D 1; : : : ; ng. We shall consider the left invariei ; Y invariant vector fields fX ei ; Y ei I i D 1; : : : ; ng are orthonormal at each point ant metric on H such that fX of H2nC1 . The sub-Laplacian in this case is given by sub D
n X
ei 2 : ei 2 C Y X
(9.2.6)
i D1
We may also consider the left invariant Riemannian metric g on H2nC1 such that ei ; Y ei I i D 1; : : : ; ng and Z e are orthonormal. The the left invariant vectors fields fX metric tensor in this case is given by
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
213
1 0 @ @ g @x@ ; @y@ g @x@ ; @z g @x ; @x@ j i j i B i C C B @ @ @ @ @ @ C g gij D B ; ; ; g g B @yi @xj @yi @yj @yi @z C @ A @ @ @ @ ; @x@ ; @y@ ; @z g @z g @z g @z j
0 2 1 C y41 B :: B : B B : B : B : B B yn y1 B 4 B DB B B B B B B B B B @
j
:: :
:
::
1C
2 yn 4
x y i4 j
xi yj 4
yi 2
yi yj 4
yj yi 4
y1 yn 4
:: :
1C
yj xi 4
x12
xi yj 4
:: :
:: :
xj xi 4
::
x2i
4
1
xi xj 4
: x2
1 C 4n
C C C C C C C C C C x1 C : 2 C C C x2i C C C C C :: C : A 1 yi 2
Then the Laplacian H2nC1 is e2: H2nC1 D sub Z In terms of coordinates .x1 ; : : : ; xn ; y1 ; : : : ; yn ; z/, the Laplacian and sub-Laplacian can be expressed by X X @2 X @2 @2 @2 C C y x i i @xi 2 @yi 2 @xi @z @yi @z ! P 2 x i C yi 2 @2 ; (9.2.7) 4 @z2 X X @2 X @2 @2 @2 D C C y x i i @xi 2 @yi 2 @xi @z @yi @z ! P 2 xi C yi 2 @2 : (9.2.8) 1C 4 @z2
sub D
H2nC1
9.2.4 Quaternionic Heisenberg Group Let H be the field of quaternions over the real numbers with the basis f1; i; j; kg. Their products are given by the following usual relations:
214
9 Laplacians and Sub-Laplacians
1i D i1 D i; 1j D j1 D j; 1k D k1 D k; i2 D j2 D k2 D 1; ij D ji D k; jk D kj D i; ki D ik D j; which obviously satisfy the associativity but not the commutativity. The elements of H are denoted by h D x0 1 C x1 i C x2 j C x3 k, with real coefficients x0 ; x1 ; x2 ; x3 2 R. The “conjugate”of h is defined by h D x0 1 x1 i x2 j x3 k. Then we have hk D k h; 8h; k 2 H: The norm, or the absolute value, of the element h 2 H is denoted by jhj and is given by r p X xi 2 : jhj D hh D We shall now consider the antisymmetric bilinear form Im.h; k/ D
1 hk kh ; 2
which introduces a Lie bracket on the vector space H ˚ R3 by H H ! R3 Q 7! Œh; h Q D Im.h; h/: Q .h; h/ Then the space ˚ qh7 D H ˚ R3 D x0 1 C x1 i C x2 j C x3 kI z1 ; z2 ; z3 D .x0 ; x1 ; x2 ; x3 I z1 ; z2 ; z3 / j xi ; zi 2 R
becomes a two-step nilpotent Lie algebra with respect to the aforementioned bracket. We shall introduce a group multiplication on qh7 by H ˚ R3 H ˚ R3 3 .h; z/; .k; zQ/ 7! .h; z/ .k; zQ/; .h; z/ .k; zQ/ D h C k ˚ .z C zQ C Im.h; k/=2/ : We shall denote by qH7 the Lie group with this multiplication law. Let X0 D .1; 0; 0; 0I 0; 0; 0/ D 1 ˚ 0 2 H ˚ R3 ; X1 D .0; 1; 0; 0I 0; 0; 0/ D i ˚ 0 2 H ˚ R3 ; X2 D .0; 0; 1; 0I 0; 0; 0/ D j ˚ 0 2 H ˚ R3 ; X3 D .0; 0; 0; 1I 0; 0; 0/ D k ˚ 0 2 H ˚ R3
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
215
be elements in qh7 which, together with their bracket generate the entire space qh7 . The expressions of the corresponding left invariant vector fields are given by 3
X @ e0 D @ C 1 X xi ; @x0 2 @zi
(9.2.9)
i D1
e 1 D @ C x0 @ C x3 @ X @x1 2 @z1 2 @z2 @ @ x x 3 0 @ e2 D X C C @x2 2 @z1 2 @z2 e 3 D @ C x2 @ x1 @ C X @x3 2 @z1 2 @z2
x2 2 x1 2 x0 2
@ ; @z3 @ ; @z3 @ : @z3
(9.2.10) (9.2.11) (9.2.12)
We shall introduce on qH7 a sub-Riemannian structure H in the strong sense, i.e., a sub-bundle spanned by the left invariant vector fields fXQ i g3iD0 . The sub-Laplacian in this case is sub D
3 X
ei 2 X
i D0
3 2 X @2 @2 @2 @2 @ D C x x C x x i 0 3 2 @x0 @zi @x1 @z1 @x1 @z2 @x1 @z3 @xi2 i D0 @2 @2 @2 @2 @2 x0 x1 x2 C x1 @x2 @z1 @x2 @z2 @x2 @z3 @x3 @z1 @x3 @z2 ! 3 3 2 2 X X @ 1 @ x0 xi2 : (9.2.13) 2 @x3 @z3 4 @z i i D1 i D0 C x3
When considering a Riemannian metric on the entire space such that the vector ei g and fZ ej g are orthonormal at each point, this metric can be regarded as fields fX an extension of the metric defined on the sub-Riemannian structure. Consequently, the Laplacian in this case is given by D sub
@2 @2 @2 2 2: 2 @z1 @z2 @z3
Higher-dimensional quaternionic Heisenberg algebras qh4nC3 Š Hn ˚ R3 are defined by .h1 ; : : : ; hn I z/; .hQ 1 ; : : : ; hQ n I zQ/
! n X 1 Im.hi ; hQ i / ; 7 ! h1 C hQ 1 ; : : : ; hn C hQ n I z C zQ C 2 i D1
216
9 Laplacians and Sub-Laplacians
for all .h1 ; : : : ; hn I z/; .hQ 1 ; : : : ; hQ n I zQ/ 2 Hn ˚ R3 , and with notation Im hi ; hQ i D hei hN i /. The Lie group associated with the Lie algebra qh4nC3 will be denoted by qH4nC3 , and it will be called the .4n C 3/-dimensional quaternionic Heisenberg group. 1 .h e 2 i hi
9.2.5 Heisenberg-Type Lie Algebra Let n and z be two real vector spaces endowed with the inner products < ; >n and < ; >z , respectively. We shall assume that there is a bilinear form J W z n ! n such that (HT-1) W (HT-2) W
J.Z; J.Z; X // D < Z; Z >z X; < J.Z; X /; Y >n D < X; J.Z; Y / >n ;
(9.2.14) 8X 2 n; Z 2 z: (9.2.15)
The bilinear form J is sometimes called Clifford multiplication. Now we will define the skew-symmetric bilinear map B W n n ! z given by B.X; Y / W z ! R;
B.X; Y /.Z/ D< J.Z; X /; Y >n :
Proposition 9.2.4. If Z1 and Z2 2 z are orthogonal, then J.Z1 ; J.Z2 ; Y // C J.Z2 ; J.Z1 ; X // D 0: Proof. Using the bilinearity of the map J; we have J.Z1 C Z2 ; J.Z1 C Z2 ; X // D J.Z1 ; J.Z1 ; X //CJ.Z1 ; J.Z2 ; X //CJ.Z2 ; J.Z1 ; X //CJ.Z2 ; J.Z2 ; X // D < Z1 ; Z1 >z X CJ.Z1 ; J.Z2 ; X //CJ.Z2 ; J.Z1 ; X // < Z2 ; Z2 >z X D < Z1 C Z2 ; Z1 C Z2 ; >z X D < Z1 ; Z1 >z X < Z2 ; Z2 >z X; and hence J.Z1 ; J.Z2 ; X // C J.Z2 ; J.Z1 ; X // D 0: Consider the inner product < ; >z on z induced naturally by the inner product on z. We have the following result.
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
217
Proposition 9.2.5. Let ' 2 z and assume that for any fixed X 6D 0, < B.X; Y /; ' >z D 0;
8 Y 2 n:
Then ' = 0. Proof. Let fZk g be an orthonormal basis on z and denote by fZk g the dual basis on z . Then X < B.X; Y /; ' >z D < J.Zk ; X /; Y >n ak D 0; P where we expressed ' D ak Zk . If we take Y D J.Zj ; X /, then by property (9.2.15) and Proposition 9.2.4, we have < B.X; Y /; ' >z D D
X X
< J.Zk ; X /; Y >n ak < J.Zk ; X /; J.Zj ; X / >n < Zj ; Zj >z aj
k6Dj
D
X
< J.Zj ; X /; J.Zk ; X / >n < Zj ; Zj >z aj D 0:
k6Dj
This implies that aj D 0: Hence ' D 0.
Corollary 9.2.6. For any fixed nonzero X 2 n, the mapping n ! z ; Y 7! B.X; Y / is surjective. We now define a two-step nilpotent Lie algebra structure on V D n ˚ z (or on n ˚ z by identifying z Š z) in such a way that for X; Y 2 n; ŒX; Y D B.X; Y /; and all other brackets are zero. By Proposition 9.2.5, we have Proposition 9.2.7.
ŒV; V D z D center:
This induces a two-step nilpotent Lie group structure on V given by V V ! V; 1 .x; w/; .x; Q w/ Q 7! .x; w/ .x; Q w/ Q D x C x; Q w C wQ C B.X; Y / : 2
218
9 Laplacians and Sub-Laplacians
Remark 9.2.8. 1. Let C `Q .z/ denote the Clifford algebra with respect to the positive bilinear form Q.z/ D< z; z >z . This means that if T .z/ is the tensor algebra 3
T .z/ D R ˚ z ˚ z ˝ z˚ ˝ z ˚ ; and IQ is a two-sided ideal in T .z/ generated by the elements of the form z ˝ z C Q.z/; then C `Q .z/ is the quotient algebra T .z/=IQ : Conditions (9.2.14) and (9.2.15) required on the bilinear map J say that the vector space n has a Clifford module structure of the Clifford algebra C `Q .z/. 2. The Heisenberg-type Lie algebra was first defined in [77], where the Lie algebra g is two-step and is equipped with an inner product < ; > such that for any Z 2 Œg; g z D center, the map J.Z/ W z? ! z? defined by (Ka-1): < J.Z/.X /; Y >D< Z; ŒX; Y >D< ad.X / .Z/; Y >
(9.2.16)
satisfies (Ka-2): J.Z/2 D < Z; Z > Id:
(9.2.17)
We have constructed a Heisenberg-type Lie algebra from the two vector spaces n and z satisfying conditions (HT-1) and (HT-2). Conversely, let’s assume that g satisfies conditions (Ka-1) and (Ka-2). Set n D z? D Œg; g? , and define the bilinear map J W z z? ! z? by J.Z; X / D J.Z/.X /. Then it can easily be seen that z, z? and the bilinear map J satisfy conditions (HT-1) and (HT-2). 3. Heisenberg algebras are particular types of Heisenberg-type Lie algebras. Also, quaternionic Heisenberg algebras are Heisenberg-type Lie algebras. The latter case can be shown as in the following. Let H0 D fh 2 HI z D zg and consider the map H0 H ! H .z; h/ 7! z h: Properties (9.2.14) and (9.2.15) are satisfied. The Heisenberg-type Lie algebra defined by this bilinear map is the quaternionic Heisenberg Lie algebra. Higherdimensional quaternionic Heisenberg Lie algebras can be constructed in a similar way. 4. All the Heisenberg-type Lie algebras are realized in terms of the Clifford module following periodicity of Clifford algebras and modules [5]. We consider next a sub-Riemanian structure on the Heisenberg-type Lie group V D n ˚ z .
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
219
d For that purpose we choose arbitrary orthonormal bases fXi gN i D1 and fZk gkD1 of n and z , respectively, where N D dim n and d D dim z. These bases induce coordinates on the Lie group V D n ˚ z by
RN Cd 3 .x1 ; : : : ; xN I z1 ; : : : ; zd / $
X
xi Xi C
X
zj Zj 2 n ˚ z :
Then under the identification T .V / Š V V; we consider a sub-Riemannian struce i gN given by ture H spanned by the left invariant vector fields fX i D1 H D V n V V Š T .V /: Considering the constants of structure given by ŒXi ; Xj D B.Xi ; Xj / D the left invariant vector fields are
P
Cikj Zk ,
X @ k ei D @ 1 Ci;j xj ; X @xi 2 @zk jk
and hence the sub-Laplacian sub is given by sub D
X
ei 2: X
Remark 9.2.9. According to a theorem in [55], we can choose an orthonormal basis in the Heisenberg-type Lie algebra V D n ˚ z with respect to which the structure constants are one of f1; 0; 1g. In this case, the expressions for the Laplacian and sub-Laplacian become very simple.
9.2.6 Free Two-Step Nilpotent Lie Algebra 1
N 2 N.N 1/ , 1 i < Let fXi gN i D1 be a basis of R . Consider also a basis fZi j g of R j N . Define the Lie bracket relations by
ŒXi ; Xj D ŒXj ; Xi D
Zij ; for 1 i < j N; 0; otherwise.
Then we can introduce a Lie group structure on RN.N C1/=2 D RN ˚ RN C2 by N R ˚ RN C2 RN ˚ RN C2 3 x ˚ z; xQ ˚ zQ 0 1 X 1 7! x C xQ ˚ @z C zQ C xi xQ j xj xQ i Zij A : 2 1i <j N
This group is called a free two-step nilpotent Lie group.
220
9 Laplacians and Sub-Laplacians
Let g be a two-step nilpotent Lie algebra. Consider the bases fXi gN i D1 and d fZj gj D1 for the complement of the derived algebra Œg; g and the derived algebra Œg; g, respectively. Let the structure constants fCi kj g be given by ŒXi ; Xj D
X
Ci kj Zk :
k
Now let fgN CC 2 .CN2 D N.N 1/=2/ be the free two-step nilpotent Lie algebra N with the basis fXi ; Zi j g1i <j N such that ŒXi ; Xj D Zi j : We define a Lie algebra homomorphism by
X
W fgN CC 2 ! g; N X X X xi Xi C xi Xi C zi j Zi j D zi j Ci kj Zk :
Then we have the Lie group homomorphism between corresponding simply connected nilpotent Lie groups. In this sense, any two-step nilpotent Lie group (Lie algebra) is covered by a free two-step nilpotent Lie group (Lie algebra).
9.2.7 The Engel Group As we explained in Sect. 9.1.2, no one has been successful yet in expressing the heat kernel of a sub-Laplacian (or Laplacian) on nilpotent Lie groups of step greater than or equal to 3 in an explicit (integral) form. The explicit formula is missing even in the minimal dimensional case (dimension 4) nilpotent Lie group of step 3, called the Engel group. This group appears in various contexts and we shall also introduce it here, although we shall not give its heat kernel (for the Laplacian or sub-Laplacian) in any form. The crucial reason for which an explicit form of the heat kernel is missing is that we cannot solve the quartic oscillator in terms of any special functions, such as is done with the harmonic oscillator. Since the Engel group has codimension two and a three-step sub-Riemannian structure, we shall present here two sub-Laplacian operators and the Laplacian and note the difference from the two-step cases. Later, at the end of Sect. 10.3.3, we shall also express what the higher-step Grushin operators look like. Let e4 be a four-dimensional Lie algebra with the basis fX; Y; W; Zg with the following nonzero bracket relations: ŒX; Y D W;
ŒX; W D Z:
(9.2.18)
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
221
Let E4 (identified with R4 ) be the Lie group whose Lie algebra is e4 . Then the group multiplication is given by the formula 1 .x; y; w; z/ .x; Q y; Q w; Q zQ/ D x C x; Q y C y; Q wCw Q C .x yQ y x/ Q ; z C zQ 2 1 1 Q C .x x/ Q .x yQ y x/ Q : C .x wQ wx/ 2 12 This group is realized in 4 4 real matrix space as 80 ˆ 1 ˆ ˆ
x 1 0 0
x2 2
x 1 0
9 1 > z > > = C wC C I x; y; z; w 2 R : > yA > > ; 1
The identification is given by the map 0 1 B B0 .x; y; w; z/ 7! B @0 0
x 1 0 0
x2 2
x 1 0
2
z C xw C xy 2 3Š w C xy 2 y 1
1 C C C: A
If A 2 e4 is an element of Engel’s algebra, denote by AQ the left invariant vector field on the group E4 defined by Q /.g/ D d f .g exp tA/ A.f ; jt D0 dt
f 2 C 1 .E4 /:
Let H2 D spanfXQ ; YQ g. By the bracket relations (9.2.18), we obtain that H2 is a three-step, regular, minimal, codimension-2 sub-Riemannian structure on E4 in the Q YQ ; WQ g, we obtain a codimension-1, two-step strong sense. If we let H1 D spanfX; sub-Riemannian structure on E4 in the strong sense. The vector fields XQ , YQ , WQ , ZQ are expressed by @ @ @ y w ; XQ D @x @w @z @ @ YQ D Cx ; @y @w @ @ WQ D Cx ; @w @z @ ZQ D ; @z
222
9 Laplacians and Sub-Laplacians
while the sub-Laplacians associated with the sub-Riemannian structures H1 and H2 are given by .2/
sub D XQ 2 YQ 2 D
2 2 @2 @2 @2 2 2 @ 2 @ .x C y / w C 2y @x 2 @y 2 @w2 @z2 @x@w
C 2w .1/
@2 @2 @2 2yw 2x ; @x@z @w@z @y@w
.2/
sub D sub WQ 2 D
2 @2 @2 @2 @2 2 2 2 2 @ .x C y C 1/ .w C x / C 2y @x 2 @y 2 @w2 @z2 @x@w
C 2w
@2 @2 @2 2.yw C x/ 2x : @x@z @w@z @y@w
If we fix a left invariant Riemannian metric on the Engel group E4 in such a way e; Y e; W e ; Zg e are orthonormal at each point, then the Laplacian is given by that fX the formula 2 .1/ Q2 Q2 D .2/ sub W Z D sub Z
D
@2 @2 @2 @2 @2 2 .x 2 C y 2 C 1/ 2 .w2 C x 2 C 1/ 2 C 2y 2 @x @y @w @z @x@w
C 2w
@2 @2 @2 2.yw C x/ 2x : @x@z @w@z @y@w
When the aforementioned vector fields are considered on the group realized as a subgroup of the 4 4 matrix space, we have the following expressions: @ XQ D ; @x @ @ x2 @ YQ D Cx C ; @y @w 2 @z @ @ Cx ; WQ D @w @z @ ZQ D ; @z
9.2 Laplacian and Sub-Laplacian on Nilpotent Lie Groups
223
and .2/ sub D
@2 @2 x 4 @2 @2 2 x2 2 2 @x @y @w 4 @z2
@2 @2 @2 x2 x3 ; @y@w @y@z @w@z @2 @2 @2 x 4 @2 D 2 2 .x 2 C 1/ 2 x 2 C @x @y @w 4 @z2 2x
.1/
sub
@2 @2 @2 x2 .x 3 C 2x/ ; @y@w @y@z @w@z @2 @2 @2 x 4 @2 2 2 D 2 2 .x C 1/ 2 1 C x C @x @y @w 4 @z2 2x
2x
@2 @2 @2 x2 .x 3 C 2x/ : @y@w @y@z @w@z
More details about Engel’s group and its sub-Riemannian structure and a description of bicharacteristic curves of three-step Grushin operators can be found in Chap. 12 of the book [27] and in the paper [47].
Chapter 10
Heat Kernels for Laplacians and Step-2 Sub-Laplacians
10.1 Spectral Decomposition and Heat Kernel In this section we shall provide an explicit multiplication form of the Laplacian for a certain class of two-step nilpotent Lie groups including Heisenberg groups; see [46]. The result presented in this section is more precise than the explicit formula of the heat kernel given for the first time in the paper of Hulanicki [68], as explained in Remark 9.1.5. Let g be a two-step nilpotent Lie algebra such that g D gC ˚ g ˚ z; ŒgC ; g D z; Œg˙ ; g˙ D 0
(10.1.1)
z is the center and n D dim gC D dim g ; dim z D d: Let fXi gniD1 , fYi gniD1 and fZk gdkD1 be a basis of gC , g and z, respectively, with the k : structure constants Ci;j d X k ŒXi ; Yj D Ci;j Zk kD1
and all other brackets zero. For each 2 z , associate the matrix C./, C./i;j D
d X
k Ci;j .Zk /:
(10.1.2)
kD1
Assume the matrix C./ t C./ is diagonal: 0
c1 ./ B 0 B B C./ t C./ D B B @ 0
c2 ./
0
1 C C C C: C A
cn ./
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 10, c Springer Science+Business Media, LLC 2011
225
226
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
with all the diagonal elements ci ./ nondegenerate positive bilinear forms on z. We identify the Lie algebra g D gC ˚ g ˚ z and the Lie group exp W gC g z Š G through the exponential map. We denote an element g in G by gD
X
xi Xi C
X
yi Yi C
X
zk Zk D .x; y; z/:
The multiplication group law is defined as 1 g gQ D g C gQ C Œg; g: Q 2 Consider the left invariant vector fields fXQ i g, fYQj g and fZQ k g expressed by X @ ei D @ 1 Cikj yj X ; @xi 2 @zk X @ ej D @ C 1 ; Cikj xi Y @yj 2 @zk ek D @ : Z @zk The sub-Riemannian structure H on G is defined by the sub-bundle spanned by ei ; Y e j gn the vector fields fX i;j D1 . We choose the metric on H such that the aforementioned left invariant vector fields are orthonormal at each point. This left invariant metric on H can be extended to the entire tangent space as a left invariant ej ; Z e k g are orthonormal at each point. ei ; Y Riemannian metric by assuming that fX The sub-Laplacian sub and the Laplacian are given by the formulas sub D D
X
ei 2 C Y ei 2 ; X
X @2 X X @2 @2 @2 k k C Ci;j yj C Ci;j xi 2 2 @xi @zk @yj @zk @xi @yi 2 X X X 1 @ k1 k2 C Ci;j Ci;j yj1 yj2 1 2 2 @zk1 @zk2 i
C
k1
1X X X 2
j
k1
Cik11;j Cik22;j xi1 xi2
@2 @zk1 @zk2
91 8 0 12 !2 > ˆ n n n n = @2 < XB X X 1 X @X C k A k C yj Ci;j C xi Ci;j A 2: @1 C > 4ˆ @zk ; : i D1 j D1 j D1 i D1 k 0
10.1 Spectral Decomposition and Heat Kernel
227
We shall give two examples next of the Lie algebras satisfying the above assumptions. Example 10.1.1. The Heisenberg algebra of any dimension. Example 10.1.2. The Heisenberg-type algebra with the dimension of the center d D dim z D dimŒg; g D 0 .mod 4/: Let fZk gdkD1 be an orthonormal basis of the center z and consider T D J.Z1 / ı J.Z2 / ı ı J.Zd /: Using (9.2.4), we have T 2 D Id; T ı J.Z/ D J.Z/ ı T;
8Z 6D 0 2 z:
Now consider the vector spaces VC D fX 2 nI T .X / D X g;
V D fX 2 nI T .X / D X g:
Then for any nonzero Z 2 z, J.Z/ maps VC into V , and viceversa. We also have ŒVC ; V D 0. For the Laplacian associated with this group, we have an explicit spectral decomposition in a multiplication form; see [46]. Theorem 10.1.3. There exist a measure space .X; dm/, a positive function ' on X and a unitary operator U W L2 .G/ Š L2 .X; dm/ such that U1 ı M' ı U D : Using the explicit form of the function ' and of the measure dm given by (10.1.3) and (10.1.6), we obtain the following consequence. Corollary 10.1.4. The spectrum of the Laplacian is Œ0; 1/, and it consists of all continuous spectrum. Even if we do not intend to provide a proof for this theorem, we shall describe the measure space .X; dm/, the unitary transformation U and the positive function ' on X that appear in this case. Let k D .k1 ; : : : ; kn /; ki 2 N; ki 0, be a multi-index and define the measure space n p Y .Xk ; dmk / D gC z nf0g; ci ./dvd
i D1
! (10.1.3)
228
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
and consider the direct union .X; dm/ D
a
.Xk ; dmk /:
(10.1.4)
k2Nn
Let F be the partial Fourier transform F W C01 .gC g z/ ! C 1 .gC g z /; Z .nCd /=2 .F f /.x; ; / D .2/ e i.<;y>C<;z>/f .x; y; z/dyd z;
iD
p 1
and consider the restriction map R W L2 .gC g z / ! L2 .gC g .z nf0g//: We also let ˚ W gC gC .z nf0g/ ! gC g .z nf0g/ be a diffeomorphism defined by ˚.v; w; / D .x; ; /; x D v w; 1 < ; y > D < ; Œv C w; y > 2 D ;
ŒD T .v C w/.y/; y 2 g ;
where T W gC ! g is a isomorphism for any ¤ 0. Then the composition K = ˚ ı R ı F is a unitary transformation given by K D ˚ ı R ı F W L2 .gC g z/ ŒD L2 .G/ n Y p
! L2 gC gC .z nf0g/;
!
ci ./dv dw d :
i D1
Consider next the following first-order differential operators: Si D
p @ ci ./wi ; i D 1; : : : ; n; @wi
and let hi be the functions given by 1
hi .w; / D e 2
p
ci ./w2 i
:
For any fixed 6D 0 and each multi-index k D .k1 ; : : : ; kn /, let h.w; ; k/ denote the Hermite function of the variables w 2 gC : h.w; ; k/ D .Sk11 h1 /.w; / .Sknn hn /.w; /:
10.2 Complex Hamilton–Jacobi Theory
229
Now for a function f 2 C01 .gC gC .z nf0g//, we define 1 E.f /.v; ; k/ D Nk ./
Z gC
f .v; w; /h.w; ; k/d w 2 C 1 .gC .z nf0g//;
where Nk ./ is the L2 -norm of the function h.w; ; k/: Z 2
Nk ./ D
gC
jh.w; ; k/j2 d w1 d wn D 2jkj kŠ n=2
n Y
ci ./
ki 1 2
:
i D1
Then the operator E is extended to a unitary transformation from L2 .gC gC .z nf0g// to L2 .X; dm/, and the unitary operator U is defined by U D E ı ˚ ı R ı F:
(10.1.5)
Finally, let the function ' on X be '.v; ; k/ D jj2 C
n X
p .2ki C 1/ ci ./;
.v; / 2 gC .z nf0g/:
(10.1.6)
i D1
With this introduction, we get the explicit integral expression for the heat kernel by calculating the kernel distribution of the composition operator U1 ı e t M' ı U: The heat kernel is given by the following result. Theorem 10.1.5. K.t I g; g/ Q D K.t I x; y; z; x; Q y; Q zQ / Z p Q D .2/.nCd /=2 e 1<;QzzC1=2ŒxQ ;y1=2Œx;y>
(10.1.7)
z
e
t jj2
p p p c ./ cosh t c ./ ci ./ 4i sinh t pc i ./ f.xi xQi /2 C.yi yQi /2 g i d: p e 2 sinh t c ./ i i D1 n Y
10.2 Complex Hamilton–Jacobi Theory Complex Hamilton–Jacobi theory is different from the classical theory by the fact that the boundary values of the Hamiltonian system allow complex values, and hence the bicharacteristics and the action are complex. The fact that complex Hamilton–Jacobi theory is more suited to the study of sub-elliptic operators rather than classical theory was first noticed by Beals, Gaveau and Greiner in the 1990s; see [13, 14, 16]. The interested reader can also consult Chap. 5 of the book [28].
230
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
10.2.1 Path Integrals and Integral Expression of a Heat Kernel It is known from statistical mechanics that the heat kernel Kt .x; y/ must be expressed as a path integral Z e St . / d. /; K.t; x; y/ D Pt .x;y/
hopefully with a suitable infinite-dimensional measure d. /, where Pt .x; y/ denotes the path space connecting x to y at a time t. The function St . / is called the classical action and is given by Z 1 t 2 j .s/j P ds: St . / D 2 0 By normalizing the time parameter with t D 1, the aforementioned integral can also be written as Z S1 .t / 1 e 2t d. t /; with t . / D . t/: t N P1 .x;y/ In the case of the Laplacian the aforementioned integral has the following asymptotic expansion: K.t; x; y/
d.x;y/2 1 e 2t u0 .x; y/.1 C O.t//: n=2 .2 t/
Here d.x; y/ denotes the Riemannian distance between the points x and y. However, for sub-Laplacians the small time asymptotic expansion is more complicated: see [16]. In particular, when the Riemannian metric is Euclidean, there is only one geodesic (a line segment) that connects the points x and y, and hence the path integral reduces to the function 2 1 jxyj 2t e ; (10.2.8) .2 t/n=2 P @2 which is just the heat kernel of the Laplacian D 12 niD1 @x 2. i
There are also similar arguments in the Heisenberg group case that reduce the path integral formula to a certain formula which makes full mathematical sense, see [78]. The heat kernel of the sub-Laplacian on the three-dimensional Heisenberg group was first constructed by Hulanicki in the mid-1970s, who provided an explicit integral formula using a probabilistic argument. Several papers published later in the 1990s deal with similar problems involving the heat kernel for Laplacians and sub-Laplacians on nilpotent Lie groups; see [14–16, 48, 98]. There are also quite a few recent papers dealing with a similar subject and calculations; see [32, 33, 78].
10.2 Complex Hamilton–Jacobi Theory
231
Unlike in Riemannian geometry, in sub-Riemannian geometry there are many geodesics connecting two points even locally; see [27, 28]. Because of this specific behavior, the heat kernel for sub-Laplacians will necessarily have an integral expression. There are several ways to attack the problem of heat kernels for sub-Laplacians. Some methods involve Mehler’s formula (see [107]), which is an important tool in the construction of the heat kernel on Heisenberg group, as done by Hulanicki [68]. This formula was also used in the previous section of this chapter to calculate the integrals included in the formula U1 ı e t M' ı U. This formula provides the generating function of Hermite polynomials explicitly. Another method is the method of complex Hamilton–Jacobi theory originated by Beals, Gaveau and Greiner [14–16], where there is no need for the generating function formula of Hermite polynomials. The aforementioned authors obtained the same formula directly, using that the heat kernel itself can be seen as a generating function. Their method starts by assuming that the heat kernel kt .g/ has a certain integral expression [see (10.2.10)]. The physical significance of the formula is that the heat flows mostly along the geodesics starting from the identity element at time t D 0, where the density of heat is the Dirac’s ı-function. The total amount of heat at a point g at a time t, denoted by kt .g/, should be equal to the sum (integral) over a certain class of geodesics arriving at point g at a time t from somewhere. This class of geodesics is determined by solving the Hamiltonian system (bicharacteristics system) under an initial-boundary condition; i.e., we assume that the coordinates in g=Œg; g are zero at t D 0 and the endpoint g must be arbitrarily given in the space G. In the Euclidean case there is only one such geodesic arriving at the point g under this condition, so no integration is needed and we have the well-known formula (10.2.8). However, in the nilpotent (non-abelian) cases we must consider geodesics whose initial points are not the identity element. These will be parameterized by the dual space Œg; g for both the Laplacian and sub-Laplacian cases. The reason for which we need to consider such geodesics is that on our curved space (which is otherwise topologically Euclidean) the wavefront set of the ı-function produces influence into the direction Œg; g. However, it is not clear whether this interpretation suffices to study the construction of the heat kernel on nilpotent Lie groups of step 3 or higher step, under the assumption that it has a prescribed integral form. Inspired by the heat kernel formula provided in Theorem 10.1.5, we shall consider the following ansatz as one of the possibilities of constructing the heat kernel: Theorem 10.2.1 (Meta-theorem). We assume that the heat kernel of a general nilpotent Lie group has the following integral form, where f denotes the action function and W denotes the volume element: K.tI .x; z/; .x; Q zQ// D kt ..Zx; Q zQ/1 .x; z//; (10.2.9) f .x;z; / 1 e t W .x; z; /d; .x; z/ D g 2 g=Œg; g Œg; g ; kt .x; z/ D N t (10.2.10) with a specific order N . For two-step nilpotent groups, N D 12 dim g=Œg; g C dimŒg; g.
232
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
Consequently, the function f .g; / will include all the information of the real geodesics when t & 0, and the volume element W .g; / will reflect the weight of the energy of the geodesics arriving at the point g. We shall briefly present the content of the following few sections. In Sect. 10.2.2 we shall explain a method to construct a solution of a class of Hamilton–Jacobi equations that gives the action function f assumed to exist by the Meta-theorem. In Sect. 10.2.3 we shall assume that the action function f is a solution of the following generalized Hamilton–Jacobi equation: X
i
@f .g; / C H.gI rf / D f .g; /; @i
and we shall find an equation satisfied by the volume element W .g; /, called the generalized transport equation. In Sects. 10.2.4 and 10.2.5 we shall show that this general mechanism works well in the case of two-step nilpotent Lie groups. In this case we have heat kernels for both sub-Laplacians and Laplacians in the integral form stated by the Meta-theorem.
10.2.2 A Solution of a Hamilton–Jacobi Equation In this section we shall consider a class of Hamilton–Jacobi equations and construct solutions under a certain assumption. This assumption is satisfied in the case of the Hamiltonian being the principal symbol of (left or right invariant) sub-Laplacians or Laplacians on two-step nilpotent Lie groups. Let H.x; yI ; / be a polynomial of the variables .x; yI ; / 2 Rm Rd Rm d R and total degree 2 with respect to the variables 2 Rm and 2 Rd : H.x; yI ; / D
X
ai j .x; y/i j C
X
bi k .x; y/i k C
X
ck ` .x; y/k ` ; (10.2.11)
where ai j .x; y/, bi k .x; y/ and ck ` .x; y/ are polynomials of the variables x 2 Rm and y 2 Rd with real coefficients. We may also the functionsp ai j ; bi k and p allow 0 m ck ` to be entire functions of the variables x C 1x 2 C and y C 1y 0 2 Cd . This condition for the Hamiltonian is satisfied by the principal symbol of the (left)invariant sub-Laplacians and Laplacians on nilpotent Lie groups. Consider the Hamiltonian system xP D H D
@H.x; yI ; / ; @
@H.x; yI ; / ; P D Hx D @x
yP D H D
@H.x; yI ; / ; @
P D Hy D
@H.x; yI ; / ; @y
10.2 Complex Hamilton–Jacobi Theory
233
with the initial-boundary conditions 8 < x.0/ D 0; x.s/ D x; : .0/ D :
y.s/ D y;
The following assumption will play an important role in the sequel. Assumption 10.2.2 We assume that there exists an open conic domain D in Cd such that for any s 2 R, .x; y/ 2 Rm Rd and 2 D, there exists a unique global solution of the above system X.t/ D X.tI s; x; y; /; .t/ D .tI s; x; y; /;
Y .t/ D Y .tI s; x; y; /; .t/ D .tI s; x; y; /;
which is smooth with respect to the variables .s; x; y; / 2 R Rm Rd D. Under this assumption, let g be a function defined by the formula g.x; yI s; / D
d X
j Yj .0I s; x; y; /
j D1
C
Z sX
i .t/XPi .t/ C
X
j .t/YPj .t/ H.X.t/; Y.t/I .t/; .t//dt:
0
Then the function g D g.x; yI s; /, where the variable is considered a parameter, is smooth and satisfies the following Hamilton–Jacobi equation: Proposition 10.2.3. We have @g C H.x; yI rg/ D 0; @s where rg D
@g @g @x ; @y
(10.2.12)
.
Proof. This is proved by an explicit computation of the derivatives @g .x; yI s; /; @s
@g .x; yI s; /; @x
@g .x; yI s; /: @y
First we shall show that @g .x; yI s; / C H.x; yI .sI s; x; y; /; .sI s; x; y; // D 0: @s
(10.2.13)
234
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
We warn the reader that this will be a tedious computation: @g .x; yI s; / @s d X @Yj j .0I s; x; y; / D @s j D1 X X C i .s/XP i .s/ C .s/j YPj .s/ H.x; yI .sI s; x; y; /; .sI s; x; y; // Z s X @i .tI s; x; y; /XPi .tI s; x; y; / C @s 0 X @XPi .tI s; x; y; / C i .tI s; x; y; / @s X @j X @YPj C j .tI s; x; y; / .tI s; x; y; /YPj .tI s; x; y; / C .tI s; x; y; / @s @s X @H @Xi .tI s; x; y; / .X.t/; Y.t/I .t/; .t// @xi @s X @H @Yj .tI s; x; y; / .X.t/; Y.t/I .t/; .t// @yj @s X @H @i .X.t/; Y.t/I .t/; .t// .tI s; x; y; / @i @s X @H @j .tI s; x; y; / dt .X.t/; Y.t/I .t/; .t// @j @s D
D
d X
@Yj .0I s; x; y; / @s j D1 X X .s/j YPj .s/ H.x; yI .sI s; x; y; /; .sI s; x; y; // C i .s/XP i .s/ C ! ( Z s X Pi @i @ X .tI s; x; y; /XPi .tI s; x; y; / C i .tI s; x; y; / .tI s; x; y; / C @s @s 0 ! X @j Pj @ Y C .tI s; x; y; /YPj .tI s; x; y; / C j .tI s; x; y; / .tI s; x; y; / @s @s X X @Yj @Xi C Pi .tI s; x; y; / .tI s; x; y; / C P j .tI s; x; y; / .tI s; x; y; / @s @s X X @j @i YPj .tI s; x; y; / .tI s; x; y; / .tI s; x; y; / dt XP i .tI s; x; y; / @s @s d X
j
@Yj .0I s; x; y; / @s j D1 X X C i .s/XP i .s/ C .s/j YPj .s/ H.x; yI .sI s; x; y; /; .sI s; x; y; // ˇs X ˇs X @Yj @Xi ˇ ˇ .tI s; x; y; /ˇ C .tI s; x; y; /ˇ : j .t/ C i .t/ 0 0 @s @s j
10.2 Complex Hamilton–Jacobi Theory
235
Now from the initial-boundary conditions X.0I s; x; y; / D 0;
X.sI s; x; y; / D x;
Y .sI s; x; y; / D y;
we have @Xi XP i .sI s; x; y; / C .sI s; x; y; / D 0; @s @Yj .sI s; x; y; / D 0; YPj .sI s; x; y; / C @s @Xi .0I s; x; y; / D 0; @s which finally imply (10.2.13). Next we shall show that @g .x; yI s; / D i .sI s; x; y; /; @xk @g .x; yI s; / D l .sI s; x; y; /: @yl Another technical computation follows: @g .x; yI s; / @xk D
d X
j
j D1
Z
s
C 0
C
@Yj .0I s; x; y; / @xk X @i X @XPi .t I s; x; y; /XPi .t I s; x; y; / C i .t I s; x; y; / .t I s; x; y; / @xk @xk
X @j @xk X @H
.t I s; x; y; /YPj .t I s; x; y; / C
X
j .t I s; x; y; /
@Xi .X.t /; Y.t /I .t /; .t // .t I s; x; y; / @xi @xk X @H @Yj .X.t /; Y.t /I .t /; .t // .t I s; x; y; / @yj @xk X @H @i .X.t /; Y.t /I .t /; .t // .t I s; x; y; / @i @xk ! X @H @j .X.t /; Y.t /I .t /; .t // .t I s; x; y; / dt @j @xk
D
d X j D1
j
@Yj .0I s; x; y; / @xk
@YPj .t I s; x; y; / @xk
236
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
Z
s
C 0
X @i X @XPi .t I s; x; y; /XPi .t I s; x; y; / C i .t I s; x; y; / .t I s; x; y; / @xk @xk
X @j
X
@YPj .t I s; x; y; / @xk X X @Yj @Xi C Pi .t I s; x; y; / .t I s; x; y; / C Pj .t I s; x; y; / .t I s; x; y; / @xk @xk ! X X @ @ j i XPi .t I s; x; y; / .t I s; x; y; / YPj .t I s; x; y; / .t I s; x; y; / dt @xk @xk C
D
@xk
d X
j
j D1
.t I s; x; y; /YPj .t I s; x; y; / C
j .t I s; x; y; /
X @Yj @Yj ˇˇs @Xi ˇˇs X .0I s; x; y; / i .t I s; x; y; / j .t I s; x; y; / ˇ C ˇ @xk @xk 0 @xk 0
D k .sI s; x; y; /;
where we made use of the initial-boundary conditions @Xi .sI s; x; y; / D @xk @Yj .sI s; x; y; / D @xk
@xi D ıi;k ; @xk @yj D 0: @xk
@Xi .0I s; x; y; / D 0; @xk
Using the similar boundary conditions @Yj .sI s; x; y; / D @yl @Xi .sI s; x; y; / D @yl
@yj D ıj;l ; @yl @xi D 0; @yl
@Xi .0I s; x; y; / D 0; @yl
we have the following computation for the derivative of g: @g .x; yI s; / @yl D
d X
j
j D1
Z
C 0
C
s
@Yj .0I s; x; y; / @yl X X @i @XPi .tI s; x; y; /XPi .tI s; x; y; / C i .tI s; x; y; / .tI s; x; y; / @yl @yl
X @j @yl X @H
.tI s; x; y; /YPj .tI s; x; y; / C
X
j .tI s; x; y; /
@Xi .X.t/; Y .t/I .t/; .t// .tI s; x; y; / @xi @yl X @H @Yj .X.t/; Y .t/I .t/; .t// .tI s; x; y; / @yj @yl
@YPj .tI s; x; y; / @yl
10.2 Complex Hamilton–Jacobi Theory
X @H @i
.X.t/; Y .t/I .t/; .t//
237
@i .tI s; x; y; / @yl
! X @H @j .X.t/; Y .t/I .t/; .t// .tI s; x; y; / dt @j @yl D
d X
j
j D1
Z
@Yj .0I s; x; y; / @yl X @i X @XPi .tI s; x; y; /XPi .tI s; x; y; / C i .tI s; x; y; / .tI s; x; y; / @yl @yl
s
C 0
X @j
X
@YPj .tI s; x; y; / @yl @yl X X @Yj @Xi .tI s; x; y; / C P j .tI s; x; y; / .tI s; x; y; / C Pi .tI s; x; y; / @yl @yl X @H @i .X.t/; Y .t/I .t/; .t// .tI s; x; y; / @i @yl ! X @H @j .X.t/; Y .t/I .t/; .t// .tI s; x; y; / dt @j @yl C
D
d X j D1
C
j
X
.tI s; x; y; /YPj .tI s; x; y; / C
j .tI s; x; y; /
@Yj .0I s; x; y; / @yl
i .tI s; x; y; /
ˇs X ˇs @Yj @Xi ˇ ˇ .tI s; x; y; /ˇ C j .tI s; x; y; / .tI s; x; y; /ˇ 0 0 @yl @yl
D l .sI s; x; y; /:
Hence we conclude that g D g.x; yI s; / satisfies the equation @g C H.x; yI rg/ D 0: @s
10.2.3 The Generalized Transport Equation In this section we shall deduce an equation, called the generalized transport equation, which is satisfied by the volume element present in the integral expression of the heat kernel for the sub-Laplacian sub on a nilpotent Lie group given by the Meta-theorem 10.2.1. Let G be an n-dimensional, connected, simply connected nilpotent Lie group with the Lie algebra g, and let fXi gm i D1 be a basis of the complement of the first derived ideal Œg; g; with m D dim g=Œg; g. Let m X sub D XQ i2 (10.2.14) i D1
238
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
be the sum of the left invariant vector fields XQ i on the group G. Then sub is a sub-Laplacian satisfying the H¨ormander condition of hypoellipticity. To apply the result of Sect. 10.2.2, we fix a basis fZj gdjD1 of the derived ideal Œg; g, and we work with the coordinates G3 T .G/ 3
X
ei C xi X
X
X
ei C xi X
ej $ .x; y/ 2 Rm Rd ; yj Z
X
ej ; yj Z
X
i XQ i C
X
j ZQ j
$ .x; yI ; / 2 RmCd RmCd ei ; Z e j g the dual basis of fXi ; Zj g. The corresponding where we have denoted by fX left invariant 1-forms are given by fi ; j g. This splitting .x; y/ of the coordinates corresponds to the splitting of the variables in the preceding section. Let H be the Hamiltonian associated with the sub-Laplacian (10.2.14): H W T .G/ Š Rm Rd Rm Rd 3 .x; yI ; / 7! H.x; yI ; / 2 R; 1X i .XQ i /2 : 2 m
H.x; yI ; / D
i D1
Assume the function f D f .x; yI / is a solution of the generalized Hamilton– Jacobi equation H.x; yI rf / C
d X i D1
i
@f D f .x; yI 1 ; : : : ; d /; @i
(10.2.15)
obtained by setting s D 1 in the solution g of the Hamilton–Jacobi equation (10.2.3) given in Sect. 10.2.2 under assumption (10.2.2). To make our assumption clear, we state once again that the heat kernel K.tI x; y; x; Q y/ Q is supposed to be of the form Q y/ Q .x; y/ ; K.tI x; y; x; Q y/ Q D kt .x; Z f .x;yI / 1 kt .x; y/ D N e t W .x; y; /d; d t R
(10.2.16)
for some positive integer N > 0. In the case of a two-step nilpotent Lie group N is fixed and is given by N D 12 dim g=Œg; g C dimŒg; g D m=2 C d . However, the following calculations are valid for any d > 0 and N > 0. Let the characteristic variety of sub be Ch D f.x; y; ; /I H.x; y; ; / D 0g:
10.2 Complex Hamilton–Jacobi Theory
239
This is a sub-bundle in T G and is trivialized as G Œg; g ; which is embedded in T .G/ according to the splitting g D Œg; g ˚ ŒfXi g: Ch D G Œg; g T .G/I see also Remark 10.2.12. The dimension of the integration variable is d D dim Ch dim G = dim Œg; g. In the following we shall accept the reasonable assumption that the integrand / f .x;yW t e W .x; yI / decreases fast enough such that the partial integrations can be performed in a convenient way when jj ! 1; see also Theorem 10.2.7. Then we have sub .e
f .x;yI / t
W/
f X 1 1 ft H.x; yI rf / W e .f / W X .f /X .W / e t sub i i t2 t m
D
i D1
C sub .W /e
ft
;
and @ @t
1 tN D
!
Z Rd
1 tN
e
ft
Z Rd
W d1 dd
1 N f e t W d1 dd C N t t
Z Rd
f f e t W d1 dd : t2
Combining the last two relations, we obtain @ sub C kt .x; y/ @t Z f 1 1 .H.x; yI rf / f / e t W d D N 2 t t Rd ! Z X f 1 XQ i .f /XQ i .W / C .sub .f / N / W e t d t Rd i Z f sub .W /e t d Rd Z X f 1 1 @f e t W d D N 2 i t t Rd @i ! Z X f 1 XQ i .f /XQ i .W / .sub.f / C N / W e t d t Rd i Z f sub .W /e t d Rd 1 1 1 A2 A1 A0 : D N t t2 t
240
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
Now we assume W d D df ^ V
(10.2.17)
with the .d 1/-form1 V D
d X
b
.1/˛1 V˛ d1 ^ ^ d ˛ ^ ^ dd :
˛D1
In general, the coefficients Vi are functions of both space variables fxi ; yj g and the characteristic variety variables fj g. Therefore, the operation sub .V / is defined as sub .V / D
d X
b
.1/˛1 sub.V˛ /d1 ^ ^ d ˛ ^ ^ dd :
˛D1
The terms A1 and A2 will be computed under assumption (10.2.17) on the function W in the following way: 1 A2 t2
Z X @f f 1 e t df ^ V i D 2 t Rd @i Z X f @f 1 d e t ^ V i D t Rd @i X Z f 1 @f D i e t d V t Rd @i 0 0 1 1 ! Z 2 XX X @f f 1 @ f D dV A e t @df ^ V C @ i Vj A d C i t Rd @i @j @i j i i 0 1 Z Z 2 XX f =t 1 @ f ^V d e e f =t @ i Vj A d D t Rd @i @j Rd j i 0 1 X Z X @Vj @e f =t A d C i @ @i @j Rd j 0 1 Z 2 XX 1 @ f e f =t @ i Vj A d D t Rd @i @j j i 0 1 Z Z X X @2 Vj A d .d C 1/ e f =t @ i e f =t dV: d @ @ i j Rd R i j
1
P. Greiner worked out this form for the special case d D 1. See also Remark 10.2.4.
10.2 Complex Hamilton–Jacobi Theory
241
On the other hand, we have 1 1 A1 D t t
Z
X
Rd
Z
i
0
1 X @f XQ i .f /XQ i @ Vj A e f =t d @j j
1 sub .f / C N e f =t df ^ V t d Z RX 1 @ D H.x; yI rf / Vj e f =t d t Rd @j j Z Z XX @e f =t Q Q d C sub .f / CN d e f =t ^ V Xi .f /Xi .Vj / @j Rd j Rd i Z X @H.x; yI rf / 1 D e f =t Vj d t Rd @j j Z XX @ e f =t C XQi .f /XQ i .Vj / d @j Rd j i Z e f =t d sub .f / C N V :
Rd
Next we shall compute A0 C
1 A t2 2
1t A1 . Using that for any j
X @ @2 f H.x; yI rf / C i D 0; @j @j @i i
we have 1 A1 t Z X @f D sub Vj e f =t dt .d C 1/ e f =t dV @j X X Z @2 Vj f =t d i e @i @j Z XX @ e f =t XQi .f /XQi .Vj / d @j Z C e f =t d ..sub .f / C N / V / ! Z Z X @f f =t D e sub Vj d C .N d 1/ e f =t dV @j Z Z X X @2 Vj f =t i C e d.sub .f /V / e f =t d @i @j X X Z @ Q Xi .f /XQ i .Vj / d e f =t @j
A0 C
1 A2 t2 Z
242
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
@f X @f D e f =t sub .Vj / d Vj C sub @j @j Z Z C .N d 1/ e f =t dV C e f =t d.sub .f /V / X Z Z X X @2 Vj X @Vj d e f =t XQ i .f /XQ i i d e f =t @i @j @j Z Z D e f =t df ^ sub .V / e f =t .sub .f / N C d C 1/ dV X Z Z X X @2 Vj X @Vj e f =t i XQ i .f /XQ i d e f =t d @i @j @j Z Z Z D e f =t df ^ sub .V / C e f =t sub .f /dV C .N d 1/ e f =t dV Z Z X XQ i .f /XQ i .dV/: e f =t D.dV/ e f =t Z
Equating A0 C t12 A2 1t A1 D 0, it suffices for V to satisfy the following generalized transport equation: X XQ i .f / XQ i .dV/ df ^ sub .V / C C D.dV/ .sub .f / C N d 1/ dV D 0;
(10.2.18)
where D.V / is defined by D.V / D
X
i
b
XX @ @Vj .V / D .1/j 1 i d1 ^ ^ d j ^ ^ d @i @i
and we have D.dV/ D d D.V / dV: If we assume that all the coefficients Vi depend only on the variables j , j D 1; : : : ; d , then the equation reduces to the transport equation for the two-step case. Remark 10.2.4. Fundamental solutions have volume elements which do not depend on the number of missing directions (see [13]). The result here is similar for the heat kernel volume element W , under the assumption that the volume element is of the form W d D df ^ V ; see (10.2.17). When there is only one missing direction, which, in the case of a nilpotent Lie group, means dimŒg; g D 1, we recover an unpublished result of P. Greiner. Next we shall work out the first-order transport equation. We consider @ sub C kt .g/ @t ( Z f 1 1 .H.gI rf / f / e t W d D N 2 t t Rd
10.2 Complex Hamilton–Jacobi Theory
C
1 t Z
C D
1
Z
X
Rd
Rd
tN 1 C t Z C
i
Z
1 t2 Z
Rd
e t d
)
X Rd
X
Rd
f
XQ i .f /XQ i .W / .sub .f / C N / W
sub.W /e
(
243
!
ft
i
d
f @f e t W d @i
!
XQ i .f /XQ i .W / .sub .f / C N / W
i
f
e t d
)
sub.W /e
ft
d
( Z X @ f 1 1 e t W d D N i t t Rd @i ! Z X f 1 XQ i .f /XQ i .W / .sub .f / C N / W e t d C t Rd i ) Z ft C sub.W /e d Rd
D
1
(
tN 1 C t Z C
1 t Z
D
(
tN 1 C t Z C
f
Rd
e t X
Rd `
Rd
1
Z
i
Z
Rd
Rd
XQ i .f /XQ i .W / .sub .f / C N / W
f
Rd
ft
e t
X
! f
e t d
)
sub.W /e
1 t Z
X @ .i W /d @i
d
X
i
@W d @i
XQ i .f /XQ i .W / .sub .f / C N d / W
i
! f
e t d
)
sub.W /e
ft
d :
If the function W does not depend on the space variables, then W satisfies the firstorder transport equation X
i
@W .sub .f / C N d / W D 0: @i
(10.2.19)
244
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
10.2.4 Heat Kernel for the Sub-Laplacian This section deals with the complex Hamilton–Jacobi theory for two-step nilpotent Lie groups following [15]. In the sequel of this chapter we shall regard the heat equation as ! @ 1 C sub K D 0; @t 2 to avoid unnecessary constants in the description of the heat kernel. Let G be an .nCd /-dimensional connected, simply connected two-step nilpotent Lie group with center z D Œg; g of dimension dim z D d . We identify T G with g g . Let fXi gniD1 be a basis of the complement of the derived algebra Œg; g. Denote the coordinates on g g by .x; zI ; / by fixing a basis fZj gdjD1 in the center z D Œg; g. d P k Let ŒXi ; Xj D 2aij Zk , with aij D aj i , and let . / be a d d matrix kD1
with the entries
. /i j D
d X
aikj k :
(10.2.20)
kD1
For each Xj we denote by XQ j D
@ @xj
C
d n P P i D1 kD1
k aij xi @z@k the corresponding left
invariant vector field on the group G. Then the sum sub D
n X
XQ i2
i D1
is a sub-Laplacian which satisfies the “H¨ormander condition” of hypoellipticity as before. The Hamiltonian associated with the aforementioned sub-Laplacian is given by d n n X X 1X aijk xi k H.x; zI ; / D j C 2 j D1
i D1 kD1
!2
X 1X D
. /j i xi j 2 j
!2 :
i
We consider the Hamiltonian system 8 P ˆ xP j D Hj D j . /j;i xi ; ˆ ˆ ˆ i < zP k D Hk ; ˆ ˆ Pj D Hxj ; ˆ ˆ : P D H 0; zk k
(10.2.21)
10.2 Complex Hamilton–Jacobi Theory
245
with the following initial-boundary conditions: 8 ˆ x.0/ D 0; ˆ ˆ ˆ ˆ < x.s/ D x D .x1 ; : : : ; xn / 2 Rn ; z.s/ D p z D .z1 ; : : : ; zd / 2 Rd ; ˆ ˆ ˆ .0/ D 1; ˆ ˆ : D . ; : : : ; / 2 Rd ; 1 d
(10.2.22)
where s 2 R, and x and z are arbitrarily given. Since X d 2 xj .t/ dxi .t/ D 2
. /j i ; 2 dt dt i we have x.t/ P D e 2t . / .0/. Then by integrating the equation
. /x.t/ P D . /e 2t . / .0/;
. /x.t/ D 1=2 e 2t . / Id .0/: p Now bypthe condition that the value .0/ D 1 is pure imaginary, the matrix 1 ./ is self-adjoint. Hence the matrix p Z 1 p 1 1s ./ d p D p 1s ./ sinh 1s ./ 2 1 sinh we have
is well defined and invertible for any s 2 R and 2 Rd , so that we have a one-toone correspondence: between .0/ and x: p p 1 ./ s 1./ p x.s/; s 6D 0: (10.2.23) .0/ D e sinh s 1 ./ The p contour is taken to be suitably surrounding the spectrum of the matrix s 1 ./. Now we shall solve the following initial value problem: p P k p P 8 ˆ < xP j .t/ D Hj D j C 1 aij xi k D j 1 ./j i xi ; i i;k p P p P ˆ j 1 ./j ` x` ./ij ; : Pi .t/ D Hxi D 1 j
`
with the initial conditions ( x.0/ D 0; .0/ D e s
(10.2.24)
p 1./
p 1./ p x: sinh s 1./
(10.2.25)
246
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
The solutions are given by x.t/ D x.tI s; x; / D e
p .st / 1./ sinh t
p
1 ./ p x; sinh s 1 ./
.t/ D .t; s; x; / p p p p 1 ./ p e s 1./ Id e t 1./ sinh t 1 ./ x D sinh s 1 ./ ! ! p p p p 1 ./ t 1./ s 1./ D e x cosh t 1 ./ e p sinh s 1 ./ p p D e t 1./ cosh t 1 ./ .0/: The solution x.t/ satisfies the boundary condition x.s/ D x; and this leads to the solutions of the initial-boundary value problem (10.2.21) under the condition (10.2.22), together with the solutions ! Z tX X p 2u 1./ k .0/ ai j xi .u/ du; zk .t/ D zk C e s
.t/
j
j
p 1;
k D 1; : : : ; d;
i
D .1 ; : : : ; d / 2 Rd :
It can be inferred from the previous expression that the functions zk .t/, with k D 1; : : : ; d; are uniquely determined, but since we do not need their explicit form in the following calculations, their final form won’t be provided. Let g D g.sI x; z; / 2 C 1 .R Rn Rd Rd / be the complex action integral given by the formula d p X g.sI x; z; / D 1 zi .0I s; x; z; /
Z C
(10.2.26)
i D1 s
< .t/; x.t/ P > C < .t/; zP.t/ > H.x.t/; z.t/I .t/; .t//dt: 0
By Proposition 10.2.3, the action g satisfies the usual Hamilton–Jacobi equation @g C H.x; zI rg/ D 0: @s The function g also satisfies the relation g.sI x; z; ` / D
1 g.1I x; z; /: `
10.2 Complex Hamilton–Jacobi Theory
247
Next we shall provide an explicit determination of the function g.sI x; z; /. If we let p .t/ D x.t/ P D .t/ 1 ./x; then
p P D 2 1 ./.t/; .t/
and we have g.sI x; z; / D
d p X 1 i zi .0I s; x; z; / i D1
Z
s
C
< .t/; x.t/ P > C < .t/; zP.t/ > H.x.t/; z.t/I .t/; .t//dt 0
D
Z d p X 1 i zi C p
1
d X
< .t/; x.t/ P > 0
i D1
D
s
Z
s
i zi C
< .t/ C
p
1 < .t/; .t/ > dt 2
1 ./x;
0
i D1
1 < .t/; .t/ > dt 2 Z s d p p X 1 i zi C < .t/; .t/ > < x; 1 ./.t/ > dt D 1 0 2 .t/ >
i D1
D D
p
1
p 1
d X
Z
s
i zi C
1=2 < .t/; .t/ > C
i D1
0
d X
ˇs 1 ˇ i zi C < x; >ˇ 0 2
i D1
1 P > dt < x; .t/ 2
d E p X p 1 Dp D 1 i zi C 1 ./ coth 1s ./ x; x : 2 i D1
Let f D f .x; z; / be given by f .x; z; / D g.1I x; z; / D
p
1
d X
i zi C
i D1
E p 1 Dp 1 ./ coth. 1 .// x; x : (10.2.27) 2
Then f satisfies the identity f .x; z; s / D g.sI x; z; / s
248
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
and is a solution of the equation H.x; zI rf / C
d X
i
i D1
@f D f .x; z; 1 ; : : : ; d /; @i
called the generalized Hamilton–Jacobi equation. If denotes the group law on the group G, then the heat kernel K.tI .x; z/; .x; Q zQ// D kt ..x; Q zQ/1 .x; z// is given by a function kt .x; z/ of the form Z f .x;z; / 1 e t W .x; zI /d; d D dim Œg; g; kt .x; z/ D n=2Cd t Rd provided the function W .x; z; / satisfies the first-order transport equation (10.2.19): X @W X d XQ j .f /XQ j .W / sub .f / C i C W D 0: (10.2.28) @i 2 j
By noting that p 1 p tr 1 ./ coth 1 ./ 2 Z p 1 cosh 1 D tr 1 ./ d 2 sinh
sub .f / D
(10.2.29)
does not depend on the space variables .x; z/, we look for a solution of the transport equation (10.2.28) of the form W .x; z; / D W ./. In fact, the square root of the Jacobian of the correspondence (10.2.23) is a solution of the transport equation (10.2.28). Proposition 10.2.5. Let W ./ D det e
p 1./
p
1 ./ p sinh 1 ./
! 12
! 12 p 1 ./ p D det ; sinh 1 ./
where the branch is taken to be W .0/ D 1. Then the function W ./ is a solution of the transport equation (10.2.28). p 1t ./ p Proof. Let .t/ D W .t/2 D det . Then sinh
1t ./
d X @W d .t/
.t/ D 2W .t/ k dt @k kD1 0 ! p d 1t
./ D tr @ p dt sinh 1t ./
p sinh
1t ./ p 1t ./
!1 1 A .t/:
10.2 Complex Hamilton–Jacobi Theory
249
By making use of the resolvent equation, we have Z 1 p d sinh 1 d d
.t/ .t/ D tr 1t ./ dt d sinh Z 1 p sinh cosh sinh 1t ./ D tr d ; sinh2 and then d X
@W ./ @k kD1 Z 1 p sinh cosh sinh 1 d W ./: 1 ./ D tr 2 sinh2 k
Hence we have X @W ./ sub .f / W ./ k @k Z 1 p sinh cosh sinh 1 d W ./; D tr 1 ./ 2 sinh2 Z 1 p cosh 1 d W ./ C tr 1 ./ 2 sinh d D W ./; 2 which shows that W ./ is a solution of the transport equation (10.2.28).
Remark 10.2.6. The function W ./ is similar to the van Vleck determinant (see [109] and [38]). We arrive at the following result. Theorem 10.2.7. The function kt .x; z/ is given by the following integral: 1 kt .x; z/ D .2 t/n=2Cd
Z Rd
e
/ f .x;z; t
!1=2 p 1 ./ p det d; sinh 1 ./
where the action function f is given by (10.2.27): f .x; z; / D
d E p p X 1 Dp 1 i zi C 1 ./ coth. 1 .// x; x : 2 i D1
Proof. By the arguments of the last two sections, we know that ! @ 1 kt .x; z/ D 0; sub C 2 @t
250
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
since we have constructed the action function f and the volume form W in such a way that the function kt .x; z/ satisfies the heat equation. Using the following asymptotic behaviors, W ./ D O.jjj /; j > 0 is arbitrary; D p E p The bilinear form 1 ./ coth 1 .t/ x; x
(10.2.30)
is (strictly) positive definite and D p E p 1 ./ coth 1 .t/ x; x D O.jjjxj2 / (10.2.31) p (since nonzero eigenvalues of 1 ./ are proportional to jj ); we shall show that the Fourier inversion formula implies that 1 lim t !0 .2 t/n=2Cd
p
Z Z g
Rd
e
/ f .x;z; t
1 ./ p det sinh 1 ./
!1=2 d'.x; z/dxdz (10.2.32)
D '.0; 0/ for 8' 2 C01 .g/. The computation will be provided in the following. Let F.z/ .'/.x; / D .2/d=2
Z
p
Rd
e
1<;z>
'.x; z/d z
be the partial Fourier transformation of functions '.x; z/ with respect to the variable z. Then !1=2 p 1 ./ p e det d'.x; z/dxdz sinh 1 ./ g Rd Z Z 1 d=2 D F.z/ .'/.x; =t/ .2/ .2 t/n=2Cd !1=2 p p p 1 ./ ddx p e 1=.2t /< 1./ coth 1./x;x> det sinh 1 ./ Z Z p 1 d=2 F.z/ .'/. tx; / .2/ D n=2Cd .2 t/ !1=2 p p p 1 .t/ e 1=2< 1.t / coth 1.t /x;x> det t d d t n=2 dx p sinh 1 .t/ Z Z <x;x> 1 t !0 F.z/ .'/.0; /e 2 ddx D '.0; 0/: ! n=2Cd=2 .2/
1 .2 t/n=2Cd
Z Z
/ f .x;z; t
10.2 Complex Hamilton–Jacobi Theory
251
10.2.5 Heat Kernel for Laplacian I In this section we shall provide the heat kernel for the Laplacian on the general two-step nilpotent Lie groups, based on our arguments about the results of the last section. Let be a Laplacian on G D sub
d X
ZQ k2 ;
kD1
with ZQ k D @z@ . This means we are assuming that G is equipped with the left k invariant metric with respect to which fXi g and fZk g form an orthonormal basis at the identity element. The aforementioned operator becomes the Laplacian with respect to this Riemannian metric. Since we have Q D0 Œsub ; Z
.Z belongs to the center of g/;
the heat kernel K .tI .x; z/; .x; Q zQ// for the Laplacian is the kernel distribution [which belongs to C 1 .RC G G/] of the composed operator P 2 Q e t sub ı Id ˝e t Zk : Here Id denotes the identity operator on the space L2 .g=Œg; g/. We regard O 2 .Œg; g/ L2 .G/ Š L2 .g=Œg; g/˝L by the identification spanfXi g Š g=Œg; g:
Theorem 10.2.8. The heat kernel for the Laplacian on general two-step nilpotent Lie groups is given by Q zQ// K .tI .x; z/; .x; Z zj 2 1 jyQ 2t D K.tI .x; z/; .x; Q y// e dy .2 t/d=2 center Z Q z/1 .x;z/; / 1 f ..x;Q t e W ./d D .2 t/n=2Cd Rd Z f .xx;zQ Q z1=2Œx;x; Q / 1 t e W ./d; D n=2Cd .2 t/ Rd
252
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
where f .x; z; / D
p
1 < ; z > C1=2 <
p p 1 ./ coth 1 ./
x; x > C1=2jj2: Remark 10.2.9. We shall show in the next section that the function f is the complex action integral for the Laplacian [see also (10.2.35)]. Proof. Since K .tI .x; z/; .x; Q zQ// is of the form K .tI .x; z/; .x; Q zQ// D kt ..x; Q zQ/1 .x; z// with function kt .x; z/ 2 C 1 .RC g/, it will suffice to express this function as Z 2 1 jyj 2t dy kt .x; z/ D K.tI .x; z/; .0; y// e .2 t/d=2 center Z Z p p p / coth 1. /x;x> 1 1<;zy>C1=2< 1. t e W ./ D .2 t/n=2Cd Rd center jyj2 1 e 2t ddy d=2 .2 t/ Z Z p 2 1 1 1 jyj 2t e t <;y> dy e D n=2Cd d=2 d .2 t/ .2 t/ R center e
p p p 1<;z>C1=2< 1. / coth 1. /x;x> t
Z
2 j2tj
W ./d
p p p / coth 1. /x;x> 1<;z>C1=2< 1. t
1 e e W ./d .2 t/n=2Cd Rd Z / 1 f .x;z; t e W ./d: D .2 t/n=2Cd Rd The aforementioned formula coincides with the earlier particular case of Theorem 10.1.5, and is the heat kernel we want to construct. D
10.2.6 Heat Kernel for Laplacian II In this section we shall construct the heat kernel for the Laplacian using the complex Hamilton–Jacobi method as in Sect. 10.2.4. We shall describe the Hamiltonian system and the related quantities associated with the sub-Laplacian. 1. The Hamiltonian H of the Laplacian: H .x; zI ; / 1 0 !2 d n n X d X X 1 k 1 @X j C a xi k C k2 A D 2 2 ij j D1
i D1 kD1
kD1
10.2 Complex Hamilton–Jacobi Theory
253
0
X 1 X j D @
. /j i xi 2 j
i
D H.x; zI ; / C
1 2
d X
!2 C
d X
1 k2 A
kD1
k2 :
kD1
2. The Hamiltonian system: 8 xP i ˆ ˆ ˆ < zP k P ˆ ˆ j ˆ :P k
D H
D Hi D . /x; i
D H D Hk C k ; k D H
xj D Hxj ;
D H z 0:
(10.2.33)
3. The initial-boundary conditions: 8 ˆ x.0/ D 0; ˆ ˆ ˆ ˆ < x.s/ D x D .x1 ; : : : ; xn / 2 Rn ; z.s/ D zpD .z1 ; : : : ; zd / 2 Rd ; ˆ ˆ ˆ .0/ D 1; ˆ ˆ : D . ; : : : ; / 2 Rd ; 1 d
(10.2.34)
where s 2 R and x; z are arbitrarily given. In the Hamiltonian system (10.2.33), with the exception of the second equation, the other equations coincide with the corresponding equations in (10.2.21). Hence we have the same solutions x .t/ D x.t/ and .t/ D .t/ given by x .t/ D x .tI s; x; / D x.tI s; x; / p p sinh t 1 ./ D e .st / 1./ p x; sinh s 1 ./ .t/ D .t/ D .tI s; x; / p p p p 1 ./ D p e s 1./ Id e t 1./ sinh t 1 ./ x sinh s 1 ./ ! p p p p 1 ./ x p D e t 1./ cosh t 1 ./ e s 1./ sinh s 1 ./ p p D e t 1./ cosh t 1 ./ .0/;
254
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
and satisfy the system (10.2.24) under the initial conditions (10.2.25). The solution z .t/ is given by p z .t/ D z.t/ C 1.t s/k : Now the complex action integral g 2 C 1 .R Rn Rd Rd / is given by the formula g .sI x; z; / D
p
1
Z
d X
i z
i .0I s; x; z; /
i D1 s
< .t/; xP .t/ > C < .t/; zP .t/ >
C 0
H .x .t/; z .t/I .t/; .t//dt Z s d p X i zi C < .t/; x.t/ P > D 1
(10.2.35)
0
i D1
H.x.t/; z.t/I .t/; .t// C 1=2
d X
! k .t/
2
dt
kD1
D g.sI x; z; / C
d s X 2 k ; 2
(10.2.36)
kD1
where g D g.sI x; z/ is the complex action function constructed using (10.2.26). We conclude with the following result. Proposition 10.2.10. The following hold: 1. g satisfies the usual Hamilton–Jacobi equation @g
C H .x; zI rg / D 0: @s 2. The function g satisfies the relation g .sI x; z; ` / D
1
g .1I x; z; /: `
We can see that the function g .1I x; z; / coincides with the function which appears in the integrand of the heat kernel given in the last section. By the same reason as in the case of the sub-Laplacian, the function f .x; z; / D g .1I x; z; / satisfies the generalized Hamilton–Jacobi equation. This can also be proved directly as in the following: H .x; zI rf /C
d X i D1
i
d d 2 X X @f
@f X 2 D H.x; zI rf /1=2 k2 C i C k @i @i i D1 kD1 kD1 X D f .x; zI / C 1=2 k2 D f .x; zI /:
10.2 Complex Hamilton–Jacobi Theory
255
The Laplacian .f / is given by X k2 D sub .f /: .f / D f C 1=2 We also know that the volume form W ./ is the solution of the first-order transport equation (10.2.19): X X @W d
Q
Q
Q Q i Xj .f /Xj .W / C Zk .f /Zk .W / .f / C C W @i 2 j k X @W d W D 0: (10.2.37) D i sub .f / C @i 2
X
Theorem 10.2.11. The heat kernel K for the Laplacian is given by Z 1 Q zQz 1 Q 2 Œx;x/ Q zQ/; .x; z// D e f .xx; W ./d: (10.2.38) K .tI .x; n=2Cd .2 t/ Rd Remark 10.2.12. This integral form coincides with the one obtained in Theorem 10.2.8. In both of the expressions given by Theorem 10.2.7 (sub-Laplacian case) and Theorem 10.2.11 (Laplacian case) the integrands of kt .g/ and kt .g/ are defined on G Rd . We may identify this with the characteristic variety of the sub-Laplacian Ch D f.x; zI ; / j H.x; zI ; / D 0g through the following map: G Rd 3 .x; z; / 7! .x; zI ./x; / 2 T G: The characteristic variety is a sub-bundle in T G and the integral formula of the heat kernel can be seen as the fiber integration of the d -forms f .x;z; / 1 e t W ./d .2 t/n=2Cd
and
/ 1 f .x;z; t e W ./d .2 t/n=2Cd
on the characteristic variety Ch. On the other hand, for the special two-step case given by Theorem 10.1.3, the domain of integration in the expression of the heat kernel is the dual of the center Œg; g. It parameterizes the irreducible representation of G which appears in the description of the unitary transformation U as L2 .G/ ! L2 .X/.
10.2.7 Examples In this section we shall express the heat kernels for the sub-Laplacian and Laplacian on the Heisenberg group, quaternionic Heisenberg group, free two-step nilpotent
256
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
Lie group of dimension 6, and Heisenberg-type groups in general. We only need to determine the explicit formula for the matrices p
p p 1 ./ coth 1 ./
and
1 ./ p det sinh 1 ./
!1=2 :
1. The three-dimensional Heisenberg group. Since the matrix . / for the threedimensional Heisenberg group is
0
; 0
we have p p coth 1 0 : 1 ./ coth 1 ./ D sinh 0 1 p 1 ./ 1 0 p D sinh 0 1 sinh 1 ./ and then
!1=2 p 1 ./ D p det : sinh sinh 1 ./
The action function f = f .x; y; z; / is f .x; y; z; / D
p cosh 2 1z C .x C y 2 /: 2 sinh
Then the heat kernel K.tI x; y; z; x; Q y; Q zQ/ for the sub-Laplacian sub D XQ 2 YQ 2 on the three-dimensional Heisenberg group is K.tI x; y; z; x; Q y; Q zQ/ (10.2.39) Z C1 p1 zQzC yx Q xy Q cosh 2 2 C 2 sinh ..xx/ Q C.yy/ Q / 2 1 t e d: D 2 .2 t/ 1 sinh Q y; Q zQ/ for the Laplacian D XQ 2 YQ 2 ZQ 2 is The heat kernel K .tI x; y; z; x; given by K .tI x; y; z; x; Q y; Q zQ/ (10.2.40) 2 Z C1 p1 zQzC yx Q xy Q Q 2 C.yy/ Q 2 / C cosh ..xx/ 2 2 sinh 2 1 t D d: e .2 t/2 1 sinh
10.2 Complex Hamilton–Jacobi Theory
257
2. Higher-dimensional Heisenberg groups. Let H2nC1 be the .2n C 1/ -dimensional Heisenberg group described in Sect. 9.2.3 with the sub-Laplacian sub D
X
XQ i2
X
YQi2 :
In this case the matrix . /i j D ai j is given by
0 Idn
. / D ; Idn 0 where Idn is the n n identity matrix. We have the action function f expressed as f .x1 ; : : : ; xn ; y1 ; : : : ; yn ; z; / D
p cosh X 2 1z C xi C yi2 ; 2 sinh
and the volume element W ./ given by
n : sinh
Hence the heat kernel K.tI x; y; z; x; Q y; Q zQ/ is given by the formula K.tI x; y; z; x; Q y; Q zQ/ Z C1 p1 1 D e .2 t/nC1 1 n d: sinh
P P yQ i xi x Q i yi zQzC .xi x Q i /2 C.yi yQ i /2 C cosh 2 2 sinh t
. .
//
(10.2.41)
Q y; Q zQ/ for the Laplacian On the other side, the heat kernel K .tI x; y; z; x; D
X
XQ i2
X
YQi2 ZQ 2
is given by K .tI x; y; z; x; Q y; Q zQ/ Z
D
C1
1 e .2 t/nC1 1 n d: sinh
P p P P 2 yQ i xi x Q i yi C cosh .xi x Q i /2 C. yi yQ i /2 1 zQzC 2 2 sinh 2
.
/
t
(10.2.42)
258
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
3. The seven-dimensional quaternionic Heisenberg Lie group. Let qH7 (Š R7 ) be the seven-dimensional quaternionic Heisenberg group described in Sect. 9.2.4. e i , i D 0; 1; 2; 3, Recall the vector fields X X @ e0 D @ C 1 X xi ; @x0 2 @zi 3
i D1
e 1 D @ C x0 @ C x3 @ x2 X @x1 2 @z1 2 @z2 2 @ x x x @ @ 3 0 1 e2 D C C X @x2 2 @z1 2 @z2 2 e 3 D @ C x2 @ x1 @ C x0 X @x3 2 @z1 2 @z2 2
@ ; @z3 @ ; @z3 @ : @z3
The sub-Laplacian and Laplacian in this case are given by sub D
3 X
e i2 ; X
(10.2.43)
i D0
D
3 X
e i2 X
i D0
3 X @2 : @z2j j D1
(10.2.44)
Since the matrix . / has the expression 0
1 0 1 2 3 B 1 0 3 2 C C;
. / D B @ 2 3 0 1 A 3 2 1 0 we have
p
p 1 ./ coth . 1/ D jj coth jjId4 :
Hence the action function f is given by f .x; z; / D f .x0 ; x1 ; x2 ; x3 ; z1 ; z2 ; z3 ; 1 ; 2 ; 3 / D
p
1
3 X i D1
where jj D
qP
i2 and jxj D
qP
1 zi i C jj coth jj jxj2 ; 2
xi2 . The volume element in this case is
W ./ D
jj sinh jj
2 :
10.2 Complex Hamilton–Jacobi Theory
259
Hence the heat kernel K.t; x; z; x; Q zQ/ for the sub-Laplacian (10.2.43) is given by the formula 1 K.t; x; z; x; Q zQ/ D .2 t/5
Z R3
e
p
1 t f
.x;Q Q z/.x;z/;/
jj sinh jj
2 d1 d 2 d 3 ;
where f .x; Q Qz/ .x; z/; is given by p xQ 1 x0 xQ 0 x1 C xQ 3 x2 xQ 2 x3 f .x; Q Qz/ .x; z/; D 1 1 z1 zQ1 C 2 xQ 2 x0 xQ 0 x2 C xQ 1 x3 xQ 3 x1 C 2 z2 zQ2 C 2 xQ 3 x0 xQ 0 x3 C xQ 2 x1 xQ 1 x2 C 3 z3 zQ 3 C 2 1 Q 2: C jj coth jj jx xj 2 The heat kernel of the Laplacian D
3 X
e 2i X
i D0
3 X @2 @z2j j D1
is given by 1 K .t; x; z; x; Q zQ/D .2 t/5
Z e
p
1 t f
.x;Q Q z/.x;z/;
R3
jj sinh jj
2 d1 d 2 d 3 ;
Q zQ; /C1=2 12 C22 C32 . where the action function f is given by f D f .x; z; x; 4. Free two-step nilpotent Lie group of dimension 6. This type of structure was introduced in Sect. 9.2.6. Let fg .3/ be the free two-step nilpotent Lie algebra of dimension 3 C C32 D 6 with an orthonormal basis fZi j g1i <j 3 of the center and an orthonormal basis fXi g3iD1 of the orthogonal complement of the center. The Lie bracket relations are ŒXi ; Xj D ŒXj ; Xi D 2Zi j ; for 1 i < j N; all others being zero. We shall express the heat kernel of the sub-Laplacian X XQ i2 ; sub D and of the Laplacian D sub
X
ZQ i;j 2
260
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
where XQ i and ZQ i;j are the left invariant vector fields on the free two-step nilpotent Lie group fG .3/ Š fg .3/. It this case the matrix . / [see also (10.2.20)] is given as the most general 3 3 antisymmetric matrix 1 0 1 2
. / D @ 1 0 3 A ; 2 3 0 0
where D . 1 ; 2 ; 3 / 2 R3 is arbitrary. In the following we shall determine the volume element W ./ D det p 1=2 p P 1./ p and the action function f .x; z; / D k zk C 12 h 1 ./ sinh 1./ p p coth 1 ./ x; xi. Since q the eigenvalues of the self-adjoint matrix 1 ./ are f0; ˙jjg, with jj D 12 C 22 C 32 , we have !1=2 p 1 ./ jj det p D : sinh jj sinh 1 ./ p Let be suitably chosen such that it is enclosing the eigenvalues of the matrix 1 ./. In order to calculate the expression Z
1 p
2 1
1 p cosh d; 1 ./ sinh
it is enough to calculate the power series of the matrix p Id C
1 ./
2 C
2Š p
Id C
1 ./
2
3Š
C
p 4 1 ./ 4Š p 4 1 ./ 5Š
! C !1 C
:
For that purpose we consider the unitary matrix T : r
0 3 jj
B B B B B 2 T DB B jj B B B @ 1 jj
jj2 12 C22 32
12 C22 p 2.12 C22 /jj2 1jj1 C2 3
r
r
jj2 12 C22 32
p p 1jj2 1 3 2.12 C22 /jj2 1jj1 C2 3 jj2 12 C22 32
jj2 12 C22 32
2.12 C22 /jj2
r
1
2 C22 p 1 2.12 C22 /jj2 1jj1 C2 3 C
r
2.12 C22 /jj2
r
jj2 12 C22 32
p jj2 12 C22 32 1jj2 1 3 p 2.12 C22 /jj2 1jj1 C2 3
C C C C C; C C C C A
10.2 Complex Hamilton–Jacobi Theory
which satisfies T
p
261
0
1 0 0 0 1 ./T D @0 jj 0 A; 0 0 jj
for 1 6D 0. Then we have ! !1 1 p 1 p X X . 1 .//2k . 1 .//2k TT Id C TT Id C .2k/Š .2k C 1/Š kD1 kD1 0 1 1 0 0 AT D T @0 jj coth jj 0 0 0 jj coth jj 0 2 1 2 3 1 3 1 jj coth jj @ 3 D jj coth jjId 3 C 2 3 22 1 2 A : jj2 1 3 1 2 12
We can see that the resulting expression is also valid in the case 1 D 0. Now the action function is f .x1 ; x2 ; x3 ; z1 ; z2 ; z3 ; 1 ; 2 ; 3 / p X p p 1 D 1 k zk C < T T 1 ./ coth 1 ./ T T .x/; x > 2 p X 1h k zk C jj coth jj.x12 C x22 C x32 / D 1 2 2 i 1 jj coth jj C 3 x1 2 x2 C 1 x3 : 2 jj Hence the heat kernel K for the sub-Laplacian is given by the following expression: K.t; x; z; x; Q zQ/ 1 D .2 t/9=2 *0
Z R3
e
p 1
l1 .z1 Qz1 C 1=2.xQ2 x1 xQ1 x2 // C 2 .z2 Qz2 C 1=2.xQ3 x1 xQ1 x3 // C3 .z3 Qz3 C 1=2.xQ3 x2 xQ2 x3 // t
0
2
B 3 B B 1j j coth j j B B2 3 Bj j coth j jId3 C @ @ j j2
e
1 3
2t
2 3 22 1 2
11
+
1 3 CC CC C.xx/;x Q x Q 1 2 C AA 2 1
jj d1 d2 d3 : sinh jj
4. Heisenberg-type group. This structure was introduced in Sect. 9.2.5. Let V D n ˚ z be a Heisenberg-type algebra with the corresponding inner products < ; >n , < ; >z , and bilinear map J W z n ! n, with the notations dim n D n and dim z D d .
262
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
The Lie bracket Œ ; is defined by n n ! z; .X; Y / 7! B.X; Y / D ŒX; Y ;
B.X; Y /.Z/ D < J.Z; X /; Y >n :
We shall use the same notation for the Lie algebra V and for the Lie group V . Let fXi g and fZ k g be orthonormal bases on n and z , respectively. Denote by Q fXi g and fZQ k g the corresponding left invariant vector fields on the Lie group V , and consider the sub-Laplacian X
sub D and the Laplacian
D sub
XQ i2 ;
X
ZQ k2 : d P
If the structure constants are given by ŒXi ; Xj D 2
kD1
X
J.Z ; / D
Ci kj Z k , then
! D . /;
2Ci kj k
with
Z D
X
k Z k 2 z :
k
Hence we have . /2 D j j2 Id . Now we can determine the action function f and the volume element W as in the following: p X p 1 p f .x; z; / D 1 k zk C h 1 ./ coth 1 ./x; xi 2 p X 1 k zk C jj coth jj kxk2 ; D 1 2 s W ./ D
p
1 ./ D p sinh 1 ./
jj sinh jj
n=2 :
P P With x D xi Xi ; z D zi Z i , the heat kernel K.t; x; z; x; Q zQ/ for the subLaplacian sub is given by 1 .2 t/n=2Cd
Z Rd
e
p 1
P
Q i zi Qzi B.x;x/ t
e
j jcoth j j jxxj Q
2
2t
jj sinh jj
n=2 d z:
And the heat kernel for the Laplacian is expressed as 1 .2 t/n=2Cd
Z Rd
e
p 1
P
i zi Qzi B.x;x/ Q t
2
e
j2tj
j jcoth j j jxxj Q 2t
2
jj sinh jj
n=2 d z:
10.3 Grushin-Type Operators and the Heat Kernel
263
10.3 Grushin-Type Operators and the Heat Kernel In this section we shall introduce the Grushin-type operator associated with a sub-Riemannian structure and a submersion. This operator can be seen as a generalization of the Grushin operator 2 @2 2 @ C x : @x 2 @y 2
The heat kernel for such operators will be constructed in terms of the fiber integration of the heat kernel of a sub-Laplacian.
10.3.1 Grushin-Type Operators Let M be a manifold. We start by recalling the definition of a sub-Riemannian structure in a strong sense on M , that is, a sub-bundle H in the tangent bundle T .M / such that H is trivial as a vector bundle and all linear combinations of vector fields taking values in H and linear sums of finite numbers of their brackets span the entire tangent space T .M / at each point (such a sub-bundle is called nonholonomic). Denote the space of vector fields on M taking values in H by .M; H/. Usually this sub-bundle is equipped with a metric, and we call a manifold with such a sub-bundle a sub-Riemannian manifold in the strong sense. Let M be a sub-Riemannian manifold in the strong sense and let ' W M ! N be a (surjective) submersion. To define an operator on the manifold N in relation with H the sub-Laplacian on M , we assume that not only there are vector fields fXQ i gdim i D1 that trivialize the nonholonomic sub-bundle H, but also each Xi can be descended to the manifold N by the map '. This means: (Sub-1) The vector fields fXQ i g are linearly independent at each point on M and span the non-holonomic sub-bundle H. So they satisfy H¨ormander’s condition (bracket generating property; see [35] and [65]). Here we also consider a metric on H in such a way that the vector fields fXQ i g are orthonormal at each point on M . (Sub-2) We also assume that there is a volume form dM V on M with respect to which each vector field XQ i (i D 1; : : : ; dim H) is skew-symmetric. (Gru) If '.x/ D '.x 0 /, then d'x .XQ i / D d'x 0 .XQ i / for i D 1; : : : ; dim H. Let sub be the sub-Laplacian on M ; i.e., sub D
dim XH
XQ k2 :
kD1
By the assumptions (Sub-1) and (Sub-2), sub is formally symmetric and hypoelliptic.
264
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
Definition 10.3.1. An operator G on N of the type GD
dim XH
2 d'.XQ k /
kD1
is called a Grushin-type operator. Note that the commutators ŒXQ j ; XQ k can also be descended to N by the map ', so that we have the following results. Proposition 10.3.2. The operator G is hypoelliptic. Proposition 10.3.3. If each fiber of the map ' is compact, then G is formally symmetric with respect to the volume form ' .dM V / obtained by taking the push forward of the volume form dM V . Proof. Let f and g be in C01 .N /. We have Z
Z G .f /.y/ g.y/ .'/ .dM V /.y/ D ' .G.f //.x/ ' .g/.x/ dM V .x/ N M Z sub .' .f //.x/ ' .g/.x/ dM V .x/ D M Z D ' .f /.x/ sub .' .g//.x/ dM V .x/ M Z f .y/ G.g/.y/ .'/ .dM V /.y/: D N
e k g and Remark 10.3.4. The sub-Laplacian depends on the specific vector fields fX is not uniquely defined as it is in the case of the Laplacian on Riemannian manifolds. In concrete cases these vector fields will be defined in a geometric way and will have a certain meaning for each case. If the fibers of the map ' W M ! N are not compact, we will have a volume form dN V on N such that ' .dN V / ^ D dM V , with a differential form on M of degree dim M dim N . In our case the Grushin-type operators will be symmetric with respect to such a volume form. Similar considerations will apply to the Heisenberg group, the Engel group and the sphere. In the following we shall assume that the dimension dim H D dim N D n, Q are not linearly and let S be the submanifold on which T the vector fields fd'.Xk /g independent. This means that Hx Ker.d'x / 6D f0g for x 2 ' 1 .S/. So on the subset N nS the vector fields fd'.XQ i /g are linearly independent and span the tangent spaces there. We consider a Riemannian metric on the subset N nS in such a way that fd'.XQ k /g are orthonormal at each point of N nS: We call the manifold N with this Riemannian metric and the Grushin-type operator descended from a sub-Laplacian a singular Riemannian manifold (see [3]).
10.3 Grushin-Type Operators and the Heat Kernel
265
10.3.2 Heisenberg Group Case We shall show that the original Grushin operator is obtained in the way explained in the last section. Consider the three-dimensional Heisenberg group H3 , which is identified with R3 as a manifold, together with the following noncommutative group law: .x; y; z/ .x; Q y; Q zQ/ D x C x; Q y C y; Q z C zQ C 1=2.x yQ yx/ Q 2 R3 : The left invariant vector fields XQ and YQ are given by @ XQ D @x @ YQ D C @y
y @ ; 2 @z x @ : 2 @z
Q where ZQ D @ , the sub-bundle spanned by XQ Then from the relation ŒXQ ; YQ D Z, @z and YQ is nonholonomic and defines a sub-Riemannian structure on H3 together with the metric Q YQ > D 0; < X; < XQ ; XQ > D < YQ ; YQ > D 1: The sum .XQ 2 C YQ 2 / is the sub-Laplacian sub and is symmetric with respect to the volume form dx ^ dy ^ d z, which coincides with the volume form introduced in Sect. 9.1.3. This form is the Haar measure of the group. We take a subgroup NY D f.0; y; 0/jy 2 Rg and consider the left quotient space Y W H3 ! NY nH3 , which is realized as Y W H3 Š R3 ! NY nH3 Š R2 ; xy Y .x; y; z/ D .u; v/ D x; z C : 2 The left invariance of the vector fields XQ and YQ satisfies condition (Gru) and is @ descended by the projection map Y to the vector fields d Y .XQ / D @u and @ Q d Y .Y / D u @v , respectively. The resulting operator GD
2 @2 2 @ C u @u2 @v2
is the original Grushin operator. It is clear that this operator is symmetric with respect to the volume form du ^ d v and .du ^ d v/ ^ dy = dx ^ dy ^ d z, where dy is a left NY -invariant 1-form (see Remark 10.3.4). As is easy to see, the vector fields d Y .XQ / and d Y .YQ / are linearly dependent along the line u D 0. We can consider a singular metric on the plane R2 in such a way that the two tangent vectors d Y .XQ /.u;v/ and d Y .YQ /.u;v/ are orthonormal at each point except on the line u D 0. The plane endowed with this (singular) metric will be called the Grushin plane.
266
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
If we consider the subgroup NZ D f.0; 0; z/jz 2 Rg, which is the center of the Heisenberg group, then the vector fields XQ and YQ can be descended to the quotient space H3 =NZ and the resulting vector fields on Z W R3 Š H3 ! H3 =NZ Š R2 ;
.x; y; z/ 7! Z .x; y; z/ D .u; v/ D .x; y/;
are dZ .XQ / D
@ @u
and
d Z .YQ / D
@ : @v
So in this case the Grushin-type operator is just the Euclidean Laplacian and it does not degenerate at any point. More generally, we take a subgroup N.a;b;c/ of the form N.a;b;c/ D f.at; bt; ct/ j t 2 Rg; where a2 C b 2 6D 0. We shall investigate only the case a 6D 0. The other case can be treated in a similar way. Let 0 W H3 ! N.a;b;c/ nH3 Š R2 . Then the map 0 is expressed as x H3 Š R 3 .x; y; z/ 7! .u; v/; u D ay bx; v D z a 3
ay bx cC : 2
We have a trivialization of the principal bundle 0 W H3 ! N.a;b;c/ nH3 given by R R2 3 . I u; v/ ! .x; y; z/ 2 H3 ; where
u C b ; z D v C .c C u=2/: a We also have the coordinate expression of the cotangent bundle T .H3 / as x D a; y D
'0
R R3 R R3 Š T .H3 / Š R R3 R R3 ; ! .x; y; zI ; ; /; x ay bx x D ; u D ay bx; v D z cC ; a a 2 x ; ˇ D : D a C b C c C .ay bx/=2 ˛ D C a 2a . ; u; vI ; ˛; ˇ/
(10.3.45) (10.3.46) (10.3.47)
The vector fields XQ and YQ are descended to the quotient space N.a;b;c/ nH3 , and the resulting vector fields are given by d0 .XQ / D b
cCu @ @ ; @u a @v
d0 .YQ / D a
@ : @u
10.3 Grushin-Type Operators and the Heat Kernel
267
It is easy to see that these vector fields are antisymmetric with respect to the volume form du ^ d v. With a left N.a;b;c/ -invariant 1-form dx on H3 , we have (see also a Remark 10.3.4) dx D dx ^ dy ^ dz: 0 .du ^ d v/ ^ a The Grushin-type operator in this case is GH
@2 @ uCc @ 2 D b a2 2 ; C @u a @v @u
(10.3.48)
and it does degenerate along the line f.u; v/ju D cg. Next we shall make a remark regarding the classification of one-dimensional subgroups of H3 . The group Aut.H3 / of the automorphisms of the Heisenberg group H3 is given by 8 19 0 0 = g11 g12 < Aut.H3 / D S D @g21 g22 0 A ; ; : a b det g g11 g12 where g D 2 GL.2; R/ and .a; b/ 2 R2 are arbitrary vectors. For any g21 g22 subgroup N.a;b;c/ , with .a; b/ 6D 0, if we suitably choose an element S 2 Aut.H3 /, then S.N.a;b;c/ / is mapped to the subgroup N.0;1;0/ .
10.3.3 Heat Kernel of the Grushin Operator In this section we shall construct the heat kernel for Grushin operators coming from the sub-Laplacian sub on the three-dimensional Heisenberg group H3 by using the method of fiber integration. We shall use the same notations as in the last section. First, we shall construct the heat kernel of the original Grushin operator GD
@2 @2 u2 2 : 2 @u @v
Let K D K.t; x; y; z; x; Q y; Q zQ/ 2 C 1 .RC H3 H3 / be the heat kernel of the sub-Laplacian sub D XQ 2 YQ 2 on the three-dimensional Heisenberg group H3 : 1 KD .2 t/2
Z
C1
1
p Q xy Q .xx/ Q 2 C.yy/ Q 2 1 t .zQzC yx / coth 2t 2
e
e
d: sinh
268
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
Let KG du ^ d v be a 2-form on RC H3 .NY nH3 / given as the fiber integration of the 3-form Kdx ^ dy ^ d z on RC H3 H3 by the map Y W RC H3 H3 ! RC H3 .NY nH3 / Š RC R3 R2 ; xQ yQ I t; x; y; z; x; Q y; Q zQ 7! t; x; y; z; u; v ; u D x; Q v D zQ C 2 that is KG t; x; y; z; u; v du ^ d v D Y K t; x; y; z; : : : d xQ ^ d yQ ^ d zQ t; x; y; z; u; v ;
(10.3.49)
where the value at the point t; x; y; z; u; v 2 RC H3 .NY nH3 / is given by the integration along the fiber Y1 .u; v/ for each fixed .t; x; y; z/ 2 RC H3 . To do this integration, we shall describe the projection map Y as follows. Let a decomposition R R2 of the group H3 be D W R R2 Š R 3 Š H3 ;
xs .u; v; s/ 7! .x; y; z/ D u; s; v : 2
Then we have Y ı D.u; v; s/ D .u; v/; and dx ^ dy ^ d z D du ^ ds ^ d v: Hence Y K t; x; y; z; : : : d xQ ^ d yQ ^ d zQ t; x; y; z; u; v "Z # C1 D K t; x; y; z; u; s; v us=2 ds du ^ dv 1
D
1 .2 t/2
Z
C1
1
Z
C1
p
e
1 t .zvCus=2C sxuy / 2
1
dds du ^ dv sinh
Z C1 Z C1 p 1 2 1 2t s.xCu/ coth 2t s ds e e D .2 t/2 1 1 2 p 1 t zC xy v coth xu 2 2t e d du ^ d v e sinh
Z C1 p .uCx/2 1 zC xy v 1 1 8t t 2 coth e e p D .2 t/3=2 1 coth 2 coth 2t xu d du ^ d v e sinh e
coth .xu/2 C.ys/2 2t
10.3 Grushin-Type Operators and the Heat Kernel
D
1 .2 t/3=2
Z
C1
e
p .uCx/2 1 t zC xy 8t 2 v coth
e
1
269
e
coth xu 2t
2 r
d du ^ d v cosh sinh
D KG .t; x; y; z; u; v/du ^ d v: Since the heat kernel K of the sub-Laplacian sub on H3 is invariant under the left action of the group H3 in the sense that K.t; g .x; y; z/; g .x; Q y; Q zQ// D K.t; x; y; z; x; Q y; Q zQ/; 8g 2 H3 ; and because of the commutativity of the left actions with the projection map previously can be seen as a smooth function on Y , the function KG obtained RC NY nH3 NY nH3 . Also, the commutativity sub ı .Y / D .Y / ı G says that the function KG is the heat kernel of the Grushin operator. This can be written as Theorem 10.3.5. The function KG satisfies
Z lim
t !0
@ GC KG D 0; @t KG .t; uQ ; vQ ; u; v/f .u; v/dud v D f .Qu; vQ /; f 2 C01 .R2 /;
or, more generally, 8f 2 L2 .R2 ; dudv/: Next we shall consider the subgroup NZ D f.0; 0; z/ j z 2 Rg of H3 and the quotient space Z W H3 ! H3 =NZ Š R2 . In this case the resulting Grushin operator @2 @2 Z .XQ /2 Z .YQ /2 D 2 2 @u @v is just the Euclidean Laplacian and the heat kernel is well known in this case. This can also be deduced from the formula of the heat kernel on H3 by integrating along the fiber, as in the previous case. Then it suffices to calculate the integral Z
C1
K.t; x; y; z; x; Q y; Q zQ/d zQ 1
(10.3.50)
270
10 Heat Kernels for Laplacians and Step-2 Sub-Laplacians
by applying the inversion formula of the Fourier transform Z
C1
1
K t; x; y; z; x; Q y; Q zQ d zQ
Z C1 Z C1 p Q xy Q 1 .xx/ Q 2 C.yy/ Q 2 1 t .zQzC yx / coth 2t 2 dd zQ e e sinh .2 t/2 1 1 Z Z p p Q xy Q 1 .xx/ Q 2 C.yy/ Q 2 1 t Qz 1 t .zC yx / coth 2t 2 D e e e d d zQ sinh .2 t/2 R R Z p p Q xy Q zQ 1 .xx/ Q 2 C.yy/ Q 2 1 1 t .zC yx / coth 2t 2 e 2 F e d zQ D sinh t .2 t/2 R Q 2 C.yy/ Q 2 1 .xx/ 2t e D ; 2 t D
where F is the Fourier transform 1 F .f /. / D p 2
Z
p
e
1x
f .x/dx:
R
Remark 10.3.6. In the case of (10.3.49), the integrand of the heat kernel K p
e
sinh uy 2 p p coth coth 2 D e 1 2t s.xCu/ e 2t .ys/ e 1 t z 2 v e 2t xu
.xu/2 C.ys/2 1 t .zvCus=2C sxuy / coth 2t 2
e
sinh
can be integrated with respect to the variable s, keeping the other variables fixed. But in the case of (10.3.50), the order of integration of variables and s cannot be changed. We shall next consider the subgroup N.a;b;c/ . We shall treat only the case a 6D 0, since the case b 6D 0 can be treated in a similar way. The heat kernel K GH .t; u; v; uQ ; vQ / 2 C 1 RC .N.a;b;c/ nH3 / .N.a;b;c/ nH3 / of the Grushin-type operator GH provided by (10.3.48) is given by the integral 1 .2 t/2 e
Z Z e R
p 1 t zv.cCu=2/C .u=aCb2/xay
R
2 2 coth 2t ..xa / C.yu=ab / /
dd : sinh
Q and z D vQ C .c Q C uQ =2/, then this can be If we put x D auQ , y D .Qu=a C b / expressed in the form given in the following result.
10.3 Grushin-Type Operators and the Heat Kernel
Theorem 10.3.7. The heat kernel of the operator GH is K GH .t; uQ ; vQ ; u; v/ Z p 2 2 uu/ 1 1 t vQv b.Q .cC.Quu/ / .Quu/ 2 coth a.a 2 Cb 2 / e 2t .a 2 Cb 2 / coth 2t .a Cb 2 / e D 3=2 .2 t/ R 1=2 d: .a2 C b 2 / sinh cosh
271
Chapter 11
Heat Kernel for the Sub-Laplacian on the Sphere S 3
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3 This section deals with the study of the heat kernel of a sub-Laplacian on the three-dimensional sphere, and a Grushin-type operator on S 2 , called the spherical Grushin operator. This is a hypoelliptic operator on S 2 and is defined by the method explained in the previous chapter. The method of investigation is the explicit determination of eigenvalues and eigenfunctions of the Laplacian and sub-Laplacian in terms of harmonic polynomials (see [7–9, 97]).
11.1.1 S 3 in the Quaternion Number Field Let H be the quaternion number field over R, with the basis f1; i; j; kg specified in Sect. 9.2.4. We consider the three-dimensional sphere as the set of unit elements [Š Sp.1/] in H. Let h D x0 1 C x1 i C x2 j C x3 k 2 S 3 H and consider the curves fh exp tigt 2R , fh exp tjgt 2R and fh exp tkgt 2R in S 3 . Then these define left invariant vector fields Xi , Xj and Xk on S 3 (in fact, defined on the whole H and tangent to S 3 ) and are expressed as follows: @ @ @ @ C x0 C x3 x2 ; @x0 @x1 @x2 @x3 @ @ @ @ Xj D x2 x3 C x0 C x1 ; @x0 @x1 @x2 @x3 @ @ @ @ C x2 x1 C x0 : Xk D x3 @x0 @x1 @x2 @x3 Xi D x1
It is easy to check that the previous vector fields are orthonormal at each point x 2 S 3 with respect to the Euclidean metric. Considered first-order differential
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 11, c Springer Science+Business Media, LLC 2011
273
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
274
operators, the vector fields Xi , Xj and Xk are skew-symmetric with respect to the following Riemannian volume form dS.x/ on S 3 : dS.x/ D x0 dx1 ^ dx2 ^ x3 x1 dx0 ^ dx2 ^ dx3 C x2 dx0 ^ dx1 ^ dx3 x3 dx0 ^ dx1 ^ dx2 :
(11.1.1)
The Lie algebra of the three-dimensional sphere S 3 , considered a Lie group Sp.1/, is generated by the basis i; j; k. The Lie brackets are defined to be the commutators of the elements. In particular, we have Œk; i D ki ik D 2j and then ŒXi ; Xk D 2Xj : Therefore, the sub-bundle H in T .S 3 / generated by the two vector fields fXi ; Xk g defines a minimal sub-Riemannian structure in the strong sense. Consequently, the operator sub on S 3 defined by sub D Xi2 Xk2
(11.1.2)
is the sub-Laplacian. Also, the operator D Xi2 Xj2 Xk2 is the Laplacian on S 3 with respect to the standard Riemannian metric. Remark 11.1.1. The volume form constructed in Sect. 9.1.3 coincides with the Riemannian volume form (11.1.1) up to a multiplicative constant. Remark 11.1.2. The vector field N normal to S 3 is given by N WD x0
@ @ @ @ C x1 C x2 C x3 : @x0 @x1 @x2 @x3
We identify the set of quaternions H Š R4 with C2 by the following correspondence: (11.1.3) H 3 x D x0 1 C x1 i C x2 j C x3 k D .x0 1 C x2 j/ C i.x1 1 C x3 j/ p p $ .x0 C x2 1; x1 C x3 1/ DW w0 .x/; w1 .x/ D .w0 ; w1 / 2 C2 ; and we consider the space H as a (right-)complex vector space with the complex multiplication p H C 3 .x; / D .x0 1 C x1 i C x2 j C x3 k; a C 1b/ D .w0 ; w1 ; / 7! .w0 ; w1 / D x0 1 C x2 j C i.x1 1 C x3 j/ .a1 C bj/ 2 H:
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3
275
Under this identification, the one-parameter transformation group generated by the vector field Xj is the scalar multiplication by D a1 C bj 2 U.1/ from the right. The orbit space of this action on S 3 is the complex projective space P 1 C, and we have the following principal bundle, called the Hopf bundle: R W S 3 ! P 1 C: Let ` W U.1/ ! U.1/ be the character ` ./ WD ` . We denote the complex line bundle on P 1 C associated with the character ` by L` . Let PN Œ x0 ; x1 ; x2 ; x3 D PN Œ x be the space of homogeneous polynomials of degree N in the variables x0 ; x1 ; x2 and x3 . We denote by HN D HN Œ x the subspace of harmonic polynomials in PN Œ x : ( ) 3 ˇ X @2 p.x/ ˇ HN Œ x D p 2 PN Œ x ˇ D0 : (11.1.4) @xi2 i D0 Proposition 11.1.3. The following decomposition of PN Œ x holds: P N Œ x D HN Œ x C
3 X
! xi2
PN 2 Œ x :
i D0 3 The space HN Œ x restricted 2 to the2 sphere2 S , denoted by HN , is the eigenspace of the Laplacian D Xi C Xj C Xk with respect to the eigenvalue N D N.N C 2/. Moreover, its dimension is given by
dim HN D .N C 1/2 D dim HN Œ x : We shall express the elements of PN Œ x in terms of the complex and anticomplex variables w0 ; w1 ; w0 and w1 given by (11.1.3). We shall denote by P.n;m/ the subspace of PN Œ x = PN Œ w0 ; w1 ; w0 ; w1 (with n C m D N ) defined by P.n;m/ D
n
ˇ ˇ p. w0 ; w1 ; w0 ; w1 / 2 PN Œ w0 ; w1 ; w0 ; w1 ˇ p p p p p w0 e 1t ; w1 e 1t ; w0 e 1t ; w1 e 1t o p D e 1.nm/t p w0 ; w1 ; w0 ; w1 ; 8 t 2 R :
Let N D n C m and H.n;m/ WD P.n;m/
T
HN . Then we have
HN Œ x D HN Œ w0 ; w1 ; w0 ; w1 D
X nCmDN n0;m0
˚ H.n;m/ ;
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
276
and the following decomposition holds: P.n;m/ D H.n;m/ C w0 w0 C w1 w1 P.n1;m1/ : For any p 2 H.n;m/ , the following relation holds: p D Xi2 C Xk2 p C .n m/2 p D sub p C .n m/2 p:
(11.1.5)
The restriction of H.n;m/ to S 3 is denoted by H.n;m/ . Relation (11.1.5) implies the following result. Proposition 11.1.4 ([7, 9]). The space H.n;m/ is the eigenspace of the subLaplacian sub corresponding to the eigenvalue N.N C 2/ `2 D 4m2 C 4m.1 C j`j/ C 2j`j; with multiplicity dim H.n;m/ D j`j C 2m C 1; where ` D n m, N D n C m, n 0; m 0. Let F ` be the subspace of C 1 .S 3 / defined by n
ˇ f 2 C 1 .S 3 / ˇ f x .a C bj/ o p D .a C b 1/` f .x/; a2 C b 2 D 1; a; b 2 R :
F` D
Then F ` can be identified with the space of smooth sections .L` / of the complex line bundle L` . Due to this identification and since csub .F ` / F ` , the sub-Laplacian sub can be seen as a second-order differential operator acting on .L` /: D` W L` ! L` : The operator D` is elliptic and is called the horizontal Laplacian; see [7]. In fact, the principal symbol .D` /.x; / W L` ! L` , where .x; / 2 T .P 1 C/nf0g, is given by 2 2 .f //.x/ Q C Xk .R .f //.x/ Q Q D` .x; / D Xi .R .s/.x/; with s 2 .L` / and f 2 C 1 .P 1 C/, such that df .x/ D 2 Tx .P 1 C/; f .x/ D 0, and R .x/ Q D x. In other words, the principal symbol is given as the multiplication by the nonzero number
.f //.x/ Q Xi .R
2
2 C Xk .R .f //.x/ Q :
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3
277
Remark 11.1.5. For each ` 2 Z, the minimal eigenvalue of the horizontal Laplacian D` is 2j`j with multiplicity j`j C 1. Proposition 11.1.6. For fixed ` 2 Z, the sum X
˚ H.n;m/ D .L` / Š F `
mnD` n0;m0
is the eigenspace decomposition of .L` / with respect to the operator D` .
11.1.2 Heat Kernel of the Sub-Laplacian on S3 We start this section by recalling certain operators discussed in [97] (see also [8]) that are used to express the heat kernels of the horizontal Laplacian D` and the sub-Laplacian sub and also the Laplacian on S 3 . Define the map n o
S W T S 3 nf0g D .x; y/ 2 R4 R4 j jxj D 1; < x; y >D 0; y 6D 0 ! C4 nf0g by p T S 3 nf 0 g 3 .x; y/ 7! S .x; y/ D z D j y jx C 1y 2 C4 nf 0 g: The image XS D S T .S 3 /nf 0 g is a quadric hypersurface 4 ˇ n o X ˇ zi2 D 0 : XS D z D .z0 ; z1 ; z2 ; z3 / 2 C4 nf0g ˇ z2 D i D0
We identify the tangent bundle T .S 3 / and the cotangent bundle T .S 3 / through the standard Riemannian metric. Then under the isomorphism S , the symplectic form !S on T .S 3 /nf0g is expressed as a K¨ahler form v 1 u 3 uX p p jzi j2 A : !S D S @ 2 1 @ @t 0
i D0
Let PN Œ z D PN Œ z0 ; z1 ; z2 ; z3 denote the homogeneous polynomials of degree N in the Pcomplex variables .z0 ; z1 ; z2 ; z3 /. We fix an inner product < ; >XS on the space ˚PN Œ z of all polynomials restricted to the quadric hypersurface XS by N 0
Z < f; g >XS D
f .z/g.z/ exp „jzj jzjq S ; XS
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
278
where „ > 0; q > 3 are fixed constants and S D
1 !S ^ !S ^ !S 3Š
is the Liouville volume form on T .S 3 /nf0g [we identify T .S 3 /nf0g with XS via
S ]. With respect to this inner product, the spaces Pk Œz and Pl Œz (restricted to XS ) are orthogonal for k 6D l. On the space HN (the space of harmonic polynomials of degree N restricted to the sphere S 3 ), we consider the map AN W HN ! RPN Œ z ; ' 7! S 3 '.x/ < x; z >N dS.x/;
(11.1.6)
3 P where z 2 XS and < x; z > D xi zi is a bilinear form on R4 C4 . i D0 P 2 P @2 N N 2 zi for each fixed z 2 XS , the < x; z > D < x; z > Since 2 @x i
polynomial 'z .x/ D < x; z >N of the variable x is harmonic. Hence, if (k 6D N ), then Z S3
2 Hk
.x/ < x; z >N dS. x / D 0:
Next, we define a map BN by BN W PN Œ z ! RHN ; 7! XS .z/ < x; z >N exp „jzj jzjq S :
(11.1.7)
Since the integration in (11.1.7) is taken over XS , it can easily be checked that BN . / 2 HN . Proposition 11.1.7 ([8,97]). With the identity operator Id on HN , the composition BN ı AN has the form BN ı AN D aN Id; where the real constant aN is given by .2N C q C 3/ N C1 2N „ .N C 1/3 with 2 y 4Vol. S 3 / Vol. S 3 p / .1= 2/ VW D : „3Cq aN D V
(11.1.8)
Here we denote by Vol.S 3 / [respectively, Vol. S 3 p2/ /] the volume of the unit p.1= 3 p sphere S 3 [resp. the sphere S.1= of radius 1= 2]. 2/ From now on we shall work in the coordinates .w0 ; w1 / D .w0 .x/; w1 .x// described in (11.1.3) rather than using .x0 ; x1 ; x2 ; x3 / 2 R4 . We also define new
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3
279
coordinates .u0 ; u1 ; v0 ; v1 / in C4 [instead of .z0 ; z1 ; z2 ; z3 / by the coordinate transformation p p zi 1zi C2 zi C 1zi C2 ui D and vi D ; i D 0; 1: 2 2 P Then the bilinear map R4 C4 ! C .x; z/ 7! xi zi D< x; z > is expressed as < x; z > D w0 u0 C w0 v0 C w1 u1 C w1 v1 ; and
3 P i D0
zi 2 D 0 holds if and only if u0 v0 C u1 v1 D 0. The Laplacian on R4 in
coordinates .w0 ; w1 / is expressed in the form 4
@2 @2 C : @w0 @w0 @w1 @w1
Therefore, with the notations (11.1.4), we have Lemma 11.1.8. ( HN Œ x D
1 ˇ X @2 f ˇ f 2 PN Œ w0 ; w1 ; w0 ; w1 ˇ D0 @wi @wi
) :
i D0
With .r; s/ 2 N0 N0 , let Ar;s D Ar;s .w0 ; w1 ; w0 ; w1 I u0 ; u1 ; v0 ; v1 / be the function defined by r s Ar;s .w0 ; w1 ; w0 ; w1 I u0 ; u1 ; v0 ; v1 / D w0 u0 C w1 u1 w0 v0 C w1 v1 : Considered a function of the variables w0 ; w1 and w0 ; w1 , for any fixed .u0 ; u1 ; v0 ; v1 / with u0 v0 C u1 v1 D 0, it can be checked that Ar;s is a harmonic polynomial of degree r C s. Let Ar;s be the operator Ar;s W
X
˚HN Œ x ! f 7!
X Z
˚PN Œ u; v
r f .w0 ; w1 ; w0 ; w1 / w0 u0 C w1 u1 S3 s w0 v0 C w1 v1 dS.x/:
Proposition 11.1.9. Using the notations of Sect. 11.1.1, we have (i) Ar;s . HN / D 0 if N 6D r C s (ii) Ar;s . H.k;l/ / D 0 for r C s D k C l, but r s 6D k l
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
280
˛
˛
ˇ
ˇ
0 1 0 1 The image P Ar;s . H.r;s/ P / is spanned by the monomials of the form u0 u1 v0 v1 with ˛i D r and ˇi D s.
Put Pr;s Œ u; v WD Ar;s . H.r;s/ /. Then it follows that Pr;s Œ u; v and Pk;l Œ u; v are orthogonal when r C s 6D k C l or r s 6D k l in case r C s D k C l. We conclude these properties with the following result. Proposition 11.1.10. The map AN in (11.1.6) can be decomposed in the form N X
CNr AN r;r D AN ;
CNr D
where
rD0
NŠ : .N r/ŠrŠ
Moreover, we have Proposition 11.1.11. The inverse operator A1 r;s W Pr;s Œ u; v ! H.r;s/ is given by a constant Cr;s times the integral operator Br;s W Pr;s Œ u; v ! H.r;s/ : 1=2 , then Br;s is defined by If we put j.u; v/j WD ju0 j2 C ju1 j2 C jv0 j2 C jv1 j2 Br;s W
7! 2
qC3 2
Z .u; v/ XS
1 X
!r wi u i
i D0
Again, we have a decomposition
!s
p 2„ j.u;v/j
wi vi j.u; v/jq e
S .u; v/:
i D0 N P
rD0
BN ı AN D
1 X
N X
CNr BN r;r D BN and
CNr 2 BN r;r ı AN r;r D aN Id:
rD0
Moreover, on each subspace H.N r;r/ we have BN r;r ı AN r;r D
aN . r /2 .CN
We shall express the heat kernel k ` .tI x; y/ D k ` .tI w.x/; w.y// of the operator sub acting on the space F ` in terms of the variables p w.x/ D w0 .x/; w1 .x/ ; wi .x/ D xi C 1xi C2 ;
where i D 0; 1:
Since the vector fields Xi and Xk are left invariant, the kernel function k ` satisfies the invariance k ` tI g x; g y D k ` tI x; y
with
g; x; y 2 S 3 :
(11.1.9)
The function space .L` / is left invariant as well. In the following proposition we give an expression of the kernel k ` .tI x; y/ (with x; y 2 S 3 ) of the operator e tD` acting on the complex line bundle L` .
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3
281
Proposition 11.1.12. Let ` 0 and x; y 2 S 3 . Then the heat kernel k ` .tI x; y/ can be expressed as k ` tI x; y m 2 1 qC3 X C 2 2mC` 2 e t 4m C4m.1C`/C2` D2 a2mC` mD0 Z X mC` X m X mC` X m wi .x/ ui wi .y/ vi wi .x/ vi wi .y/ ui XS
p
j.u; v/jq e 2„j.u;v/j S .u; v/ m 2 1 qC3 X C 2 2mC` e t 4m C4m.1C`/C2` D2 2 a2mC` mD0 Z X mC` X m wi .y 1 x/ ui u0mC` vm wi .y 1 x/ vi 0 XS
p
j.u; v/jq e
2„j.u;v/j
S .u; v/:
Note that w0 .y 1 x/ D w0 .y/w0 .x/ C w1 .y/w1 .x/; w1 .y 1 x/ D w0 .y/w1 .x/ w1 .y/w0 .x/: The kernel function k ` satisfies the identities p p (i) k ` tI x e 1 ; y D e 1 ` k ` tI x; y ; p p (ii) k ` tI x; y e 1 D e 1 ` k ` tI x; y . In the case ` < 0 one has k ` t I x; y D2
qC3 2
m 2 1 X C2mCj`j
mD0
Z
X
XS
e t
4m2 C4m.1Cj`j/C2j`j
a2mCj`j m X mCj`j X m X mCj`j wi .x/ ui wi .x/ vi wi .y/ ui wi .y/ vi p
j.u; v/jq e 2„j.u;v/j S .u; v/ m 2 1 C2mCj`j t 4m2 C4m.1Cj`j/C2j`j qC3 X e D2 2 a2mCj`j mD0 Z X m X mCj`j mCj`j wi .y 1 x/ ui wi .y 1 x/ vi um 0 v0 XS
p
j.u; v/jq e
2„j.u;v/j
S .u; v/:
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
282
For ` < 0, the kernel k ` satisfies equivalence properties similar to (i) and (ii) above. The heat kernel Ksub . tI x; y / of the sub-Laplacian can be obtained as the sum over the kernels k ` . tI x; y /, ` 2 Z. Theorem 11.1.13. Ksub tI x; y D2
m 2 1 1 X X C2mC`
qC3 2
Z
e t
4m2 C4m.1C`/C2`
a2mC` mC` X m X mC` X m wi .x/ui wi .y/vi wi .x/vi wi .y/ui
`D0 mD0
X XS
p
j.u; v/jq e
2„j.u;v/j
S .u; v/ 2 1 1 m qC3 X X C 2 2mC` 2 C2 e t 4m C4m.1C`/C2` a2mC` `D1 mD0 Z X m X mC` X m X mC` wi .x/ui wi .y/vi wi .x/vi wi .y/ui XS
p
j.u; v/jq e
2„j.u;v/j
S .u; v/:
11.1.3 Spherical Grushin Operator and the Heat Kernel ˚ In this section we consider the left action of the group U.1/ Š e t j gt 2R on S 3 , and define a Grushin-type operator on P 1 .C/; i.e., we consider the Hopf bundle L W S 3 ! U.1/nS 3 Š P 1 C [= the orbit space by the left action of U.1/]. Then, by the associative law of the quaternion number field, the operators Xi and Xk commute with the left action and one obtains vector fields dL .Xi/ and dL .Xk / on P 1 C. Considered first-order operators, these vector fields are skew-symmetric with respect to the volume form .L / .dS.x// [the forward push of the volume form dS.x/ by the projection map L ; see Sect. 9.1.3]. This volume form also coincides with the standard volume form on S 2 except for a multiplicative constant. Since the vector fields Xi and Xk satisfy conditions (Sub-1), (Sub-2) and (Gru), we have Proposition 11.1.14. The operator GS WD dL .Xi /2 dL .Xk /2 is hypoelliptic on P 1 C. We shall call the operator GS the spherical Grushin operator, since it is con 2 @ @2 structed in a similar way as the Grushin operator GH WD @x 2 C x 2 @y 2 .
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3
283
Remark 11.1.15. All elements h 2 H with h2 D 1 are conjugate to each other. Therefore, taking another generator g0 (g0 2 D 1) of the group U.1/, the resulting operator on P 1 C is conjugate to GS as well. We determine the subset in P 1 C on which the operator GS degenerates, i.e., the subset on which the vector fields dL .Xi / and dL .Xk / are linearly dependent. This can be done by determining all points x D .x0 ; x1 ; x2 ; x3 / 2 S 3 for which a solution .a; b/ 2 R2 ; a2 C b 2 D 1, of j x0 C x1 i C x2 j C x3 k D x0 C x1 i C x2 j C x3 k ai C bk exists. A necessary condition for solving this equation turns out to be x02 C x22 D x12 C x32 D
1 2
and the solution .a; b/ is given by a D 2.x1 x2 C x0 x3 /; b D 2.x2 x3 x0 x1 /: p p p p If we put 2x0 D cos , 2x2 D sin , 2x1 D cos and 2x3 D sin , then the numbers a and b can be expressed in the following form: a D sin. C /; b D cos. C /: ˚ Let p.i; k/ D x 2 S 3 j x02 C x22 D 1=2 D x12 C x32 . It can be checked that p.i; k/ is invariant under the action of the group U.1/ to the left (and also to the right). Proposition 11.1.16. The operator GS degenerates on the set L .p.i; k// Š S 1 ; i.e., the vector fields dL .Xi / and dL .Xk / are linearly dependent (but do not vanish simultaneously). In order to make a comparison with the Heisenberg–Grushin operator GH , we shall present the spherical Grushin operator GS explicitly in two ways. First, we consider the map L given by L W H Š R4 ! R3 ; .x0 ; x1 ; x2 ; x3 / 7! .u1 ; u2 ; u3 /; u1 D x0 x1 x2 x3 ; u2 D x0 x3 C x1 x2 ; u3 D x02 C x22 1=2; 2 which does realize P 1 C as a sphere S.1=2/ of radius 1=2 in R3 . This map is characterized by the equation
u21 C u22 C u23 1=4 D x02 C x22 x02 C x12 C x22 C x32 1 :
(11.1.10)
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
284
Using (11.1.10) it can be easily checked that L .p.i; k// coincides with the equator. Moreover, S 3 is divided into two connected components having the twodimensional torus p.i; k/ as the common boundary. As stated in the beginning of this section, we can descend the right actions of fe t i g and fe t k g onto R3 through the map L . The resulting actions are the rotations with respect to the u2 -axis and u1 -axis, respectively. The vector fields dL .Xi / and dL .Xk / are of the form @ @ C 2u3 ; @u3 @u1 @ @ dL .Xk / D 2u2 C 2u3 : @u3 @u2
dL .Xi / D 2u1
(11.1.11) (11.1.12)
It is obvious that dL .Xi / and dL .Xk / are tangent to the sphere and are linearly dependent only along the equator f .u1 ; u2 ; 0/ j u21 C u22 D 1=4 g. The spherical Grushin operator GS is the restriction of 1˚ 1 dL .Xi/ 2 C dL .Xk / 2 GS D 4 4 2
@2 @2 @ @ @ C u1 2u3 D .u21 C u22 / 2 C u23 2 2 @u3 @u1 @u3 @u1 @u2 u2
@ @2 @2 2u1 u3 2u2 u3 @u2 @u1 @u3 @u2 @u3
to the sphere u21 C u22 C u23 D 1=4. Next, we shall provide an expression in local coordinates, with the exception of one point. Consider the following local trivialization: (11.1.13) ˆL W U.1/ C ! U0 ; p L W .; z/ D .; x C 1y/ 7! ! ! z Re p 1 C Re p i (11.1.14) 1 C jzj2 1 C jzj2 ! ! z j Im p k 2 S 3 H: C Im p 1 C jzj2 1 C jzj2 p Then the variable z D x C 1y is a local coordinate on the subset L .U0 /, and with this coordinate the vector fields dL .Xi / and dL .Xk / are expressed as @ @ dL .Xi / D 1 C x 2 y 2 C 2xy ; @x @y 2 @ @ C x y2 1 : dL .Xk / D 2xy @x @y
11.1 Sub-Riemannian Structure and Sub-Laplacian on the Sphere S 3
285
Only on the unit circle x 2 C y 2 D 1 are these vectors linearly dependent, and the spherical Grushin operator has the form @2 GS D 1 C 2.x 2 y 2 / C .x 2 C y 2 /2 @x 2 @2 @ @ @2 C 4x C 4y : C 1 2.x 2 y 2 / C .x 2 C y 2 /2 C 8xy @y 2 @x@y @x @y Remark 11.1.17. Recently, in [3] a structure on a manifold which is called almostRiemannian was defined. Roughly speaking, this structure is given through a set of Lie bracket generating vector fields without assuming the linear independence at every point (the number of the vector fields will be less than the dimension of the manifold). Moreover, a nondegeneracy condition for the sub-manifold on which 2 the vector ˚ fields are linearly dependent is assumed. The S -case carrying the vector fields dL .Xi /; dL .Xk / is mentioned as an example. Now using the heat kernel expression for the sub-Laplacian on S 3 given in the last section, we give an integral expression of the heat kernel for the spherical Grushin operator. Since Ksub . tI x; y / satisfies the following invariance: Ksub tI g x; g y ;
8g; x; y 2 S 3 ;
according to the construction of the spherical Grushin operator GS , the heat kernel KGS . t; L .x/; L .y/ / 2 C 1 RC P 1 C P 1 C of GS can be expressed in the following integral form. Theorem 11.1.18. 1 KGS t; L .x/; L .y/ D p 1
Z
d Ksub tI x; y : U.1/
We also have Theorem 11.1.19. The following trace formula holds: tr e t GS Z Z d 1 Ksub tI x 1 x; 1 dS.x/ D p 3 1 S U.1/ m 2 qC3 1 1 2 2 2 X X C2mC` D p e t 4m C4m.1C`/C2` 1 `D0 mD0 a2mC` Z Z h Z imC` jw0 .x/j2 C jw1 .x/j2 u0 C w0 .x/. /w1 .x/ u1 U.1/ S 3 XS
11 Heat Kernel for the Sub-Laplacian on the Sphere S 3
286
h
im mC` m u 0 v0 jw0 .x/j2 C jw1 .x/j2 v0 C w0 .x/. /w1 .x/ v1 p 2„j.u;v/j
j.u; v/jq e
d S .u; v/ dS.x/ 2
m qC3 1 1 2 2 2 X X C2mC` e t 4m C4m.1C`/C2` Cp 1 `D1 mD0 a2mC` Z Z Z h im jw0 .x/j2 C jw1 .x/j2 u0 C w0 .x/. /w1 .x/ u1 U.1/
h
S3
XS
imC` m mC` u 0 v0 jw0 .x/j2 C jw1 .x/j2 v0 C w0 .x/. /w1 .x/ v1 p 2„j.u;v/j
j.u; v/jq e
S .u; v/ dS.x/
d ;
where we have used the identities w0 . x 1 x / D jw0 .x/j2 C jw1 .x/j2 ; w1 . x 1 x / D w0 .x/ w1 .x/: Based on the previous information, the zeta-regularized determinant of the subLaplacian and other related operators is computed; see [10]. Similar computations can be found in [10] for the case of a two-step, codimension-1 sub-Riemannian structure on the seven-dimensional sphere S 7 . The problem of finding the heat kernel in the case of the two-step, codimension-3 sub-Riemannian structure on S 7 still remains an open problem. In the case of the sphere S 7 , expressing the zeta-regularized determinant is not easy. However, we know that the spectral zeta function is holomorphic at the origin.
Chapter 12
Finding Heat Kernels Using the Laguerre Calculus
12.1 Introduction In this chapter, we are going to use a harmonic analysis method to construct the heat kernels and fundamental solutions of the sub-Laplacian on the Heisenberg group. This method relies on Laguerre calculus. We shall start with a beautiful idea of Mikhlin from his 1936 study of convolution operators on R2 . Let K be a principal value (P.V.) convolution operator on R2 : Z K.f /.x/ D lim
"!0 jyj>"
K.y/f .x y/dy;
where f 2 C01 .R2 / and K 2 C 1 .R2 n f.0; 0/g/ is homogeneous of degree 2 with vanishing mean value; i.e., Z K.y/dy D 0: jyjD1
Thus we can write K.x/ D
f ./ ; r2
where f ./ D
x D x1 C ix2 D re i ; X
fm e i m :
m2Z;m¤0
Suppose that g is another smooth function on Œ0; 2 with g./ D
X
gm e i m :
m2Z;m¤0
Then g induces a principal value convolution operator G on R2 with kernel In [91], we found the following identity.
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 12, c Springer Science+Business Media, LLC 2011
g. / . r2
289
290
12 Finding Heat Kernels Using the Laguerre Calculus
Proposition 12.1.1. jmji jmj e i m jkji jkj e i k jm C ki jmCkj e i.mCk/ D : 2 2 r2 R 2 r2 2 r2 Here R2 stands for the standard convolution on R2 . X
Definition 12.1.2. Let f ./ D
fm e i m ;
m2Z;m¤0
inducing the principal value convolution operator K, with kernel Then the symbol .K/ of K is defined by .K/ D
X m2Z;m¤0
jmji jmj 2
f . / r2
on C01 .R2 /.
!1 fm e i m :
With this definition we may rewrite Proposition 12.1.1 as follows. Theorem 12.1.3. Let K and G be two principal value convolution operators on C01 .R2 /. Then K R2 G D .K .G : Now we shall generalize the result of Theorem 12.1.3 to the Heisenberg group. The nonisotropic Heisenberg group Hn is the Lie group with the underlying manifold Cn R D fŒz; t W z 2 Cn ; t 2 Rg and the multiplication law " Œz; t Œw; s D z C w; t C s C 2Im
n X
# aj zj w Nj ;
(12.1.1)
j D1
where a1 ; a2 ; : : : ; an are positive numbers. It is easy to check that the multiplication (12.1.1) does indeed make Cn R into a group whose identity is the origin e D Œ0; 0, and where the inverse is given by Œz; t1 D Œz; t. The Lie algebra of Hn is a vector space which, together with a Lie bracket operation defined on it, represents the infinitesimal action of Hn . Let hn denote the vector space of left invariant vector fields on Hn . Note that this linear space is closed with respect to the bracket operation ŒV1 ; V2 D V1 V2 V2 V1 : The space hn , equipped with this bracket, is referred to as the Lie algebra of Hn .
12.2 Laguerre Calculus
291
The Lie algebra structure of hn is most readily understood by describing it in terms of the following basis: Xj D
@ @ C 2aj yj ; @xj @t
Yj D
@ @ 2aj xj @yj @t
and T D
@ ; @t
(12.1.2)
where j D 1; 2; : : : ; n, z D .z1 ; z2 ; : : : ; zn / 2 Cn with zj D xj C iyj ; t 2 R. Note that we have the commutation relations ŒYj ; Xk D 4aj ıjk T
for j; k D 1; 2; : : : ; n:
(12.1.3)
Next, we define the complex vector fields @ 1 @ i aj zj ZN j D .Xj Ci Yj / D 2 @Nzj @t
and Zj D
@ 1 @ Ciaj zNj .Xj i Yj / D 2 @zj @t (12.1.4)
for j D 1; 2; : : : ; n. Here, as usual, @ 1 D @zj 2
@ @ i @xj @yj
and
@ 1 D @Nzj 2
@ @ Ci @xj @yj
:
The commutation relations (12.1.3) then become ŒZN j ; Zk D 2i aj ıjk T with all other commutators among the Zj , ZN k and T vanishing.
12.2 Laguerre Calculus Laguerre calculus is the symbolic tensor calculus on the Heisenberg group Hn . It was first introduced on H1 by Greiner [57] and extended to Hn and Hn Rd by Beals, Gaveau, Greiner and Vauthier [17]. The Laguerre functions have been used in the study of the twisted convolution, or equivalently, the Heisenberg convolution for several decades (see Geller [51] and Peetre [96]). The Laguerre functions also played an important role in the Fock–Bargmann and Schr¨odinger representations of the Heisenberg group (see Folland [43] for details). But it was in Greiner [57] that for the first time Laguerre functions were connected with left invariant convolution operators on H1 , and were used to invert some basic differential operators on H1 , namely, the Lewy operator and the Heisenberg sub-Laplacian. In this chapter we shall use the Laguerre calculus to find the heat kernel and fundamental solution of the Kohn Laplacian and Paneitz operator. Of course, this method has some limitations since it depends heavily on the group structure and
292
12 Finding Heat Kernels Using the Laguerre Calculus
orthogonality of the Laguerre functions. However, this method allows us to construct heat kernels for higher-order partial differential operators, like powers of the subLaplacian and Paneitz operator on the group. In order to introduce the Laguerre calculus, we first recall the definitions of the twisted and P.V. convolutions on the Heisenberg group. Consider the functions f; g 2 C01 .Hn /. The Heisenberg convolution is given by Z
f .y/g.y1 x/dV .y/I
f g.x/ D
(12.2.5)
Hn
here dV .y/ is the Haar measure on Hn and is exactly the Euclidean measure on R2nC1 .
12.2.1 Twisted Convolution We focus our attention on the phase space Rn Rn , which we identify with Cn via 2 Cn , D u C i v $ .u; v/ 2 Rn Rn . Let 2 3 a1 6 a2 7 6 7 AD6 7 :: 4 5 : an be a positive definite diagonal n n real matrix. Consider the symplectic form < ; > given by the Heisenberg group multiplication law (12.1.1) and defined by 0 N D 2Im @ < z; w > D 2Im.Az w/
n X
1 aj zj wN j A ;
j D1
where z, w 2 Cn . With a fixed real constant, we can define the twisted convolution of two functions F and G by Z .F G/.z/ D
Cn
F .z w/G.w/e i
d wI
(12.2.6)
here d w is the Euclidean measure on Cn . Notice that, in view of the antisymmetry of < ; >, we have < z w; w > D < w; z >; thus G F D F G; so the twisted product is not commutative.
12.2 Laguerre Calculus
293
The twisted convolution arises when we analyze the convolution of functions on the Heisenberg group in terms of the Fourier transform in the t-variable. To see this, let f .z; t/ be a test function on Hn . Define e .z/ D f
Z
f .z; t/e it dt:
(12.2.7)
R
Similarly define e g when g is another test function on Hn . Suppose f g is the convolution of f and g on Hn . Then
B
e e .f g/ D f g :
(12.2.8)
12.2.2 P.V. Convolution Operators In order to show the regularity of the solution operator S.f / D f ‰˛ , we need to introduce principal value convolution operators on Hn . These operators are the analog of Calder´on–Zygmund principal value convolution operators on Rn . As we know, the underlying manifold of Hn is R2nC1 ; but the role of the additive structure in R2nC1 is supplied by the Heisenberg group multiplication law (12.2.4). Moreover, the group law forces us to use nonisotropic dilations on Hn ; i.e., x 7! ı ı x D ı ı Œz; t D Œız; ı 2 t for all ı > 0. These dilations are automorphisms of the group Hn : ı ı .x y/ D .ı ı x/ .ı ı y/I but the standard isotropic dilations of R2nC1 are not automorphisms of Hn . A function f defined on Hn is said to be H-homogeneous of degree m on Hn if f .ı ı Œz; t/ D f .ız; ı 2 t/ D ı m f .z; t/ for all ı > 0. For example, the fundamental solution ‰m of Lm ˛ is H-homogeneous of degree 2n C 2m 2. Next we introduce the norm function given by 1=4 .x/ D kzk4 C t 2 ;
where kzk2 D
n X
aj jzj j2 :
(12.2.9)
j D1
Obviously, we have .x1 / D .x/ D .x/ and .ı ı x/ D ı.x/. In addition, the function satisfies the triangle inequality: .x y/ C1 f.x/ C .y/g
294
12 Finding Heat Kernels Using the Laguerre Calculus
for some universal constant C1 . The distance function, d.x; y/ of points x; y 2 Hn , is defined to be d.x; y/ D .y1 x/: It is clear that d.x; y/ satisfies the symmetric property d.x; y/ D d.y; x/. Suppose that K 2 C 1 .Hn n f0g/ is H-homogeneous of degree . Then K is locally integrable near the origin if > 2n 2. See Folland and Stein [45] for the proof of this statement. Definition 12.2.1. Let K 2 C 1 .Hn n f0g/ be H-homogeneous of degree 2n 2. Then K is said to have mean value zero property if Z K.x/d.x/ D 0;
(12.2.10)
.x/D1
where d.x/ is the induced measure on the Heisenberg unit sphere .x/ D 1. Using Theorem 3 and Corollary 5.24 of Chap. XII in Stein [105], we can get the basic estimate concerning P.V convolution operators on Hn : Theorem 12.2.2. Let K 2 C 1 .Hn n f0g/, H-homogeneous of degree 2n 2 with mean value zero property. Then K induces a principal value convolution operator, given by Z K.f /.x/ D .f K/.x/ D lim
"!0 d.x;y/>"
f .y/K.y1 x/dV .y/;
(12.2.11)
for f 2 C01 .Hn /. Moreover, the operator K given by (12.2.11) can be extended to a bounded operator from the Lp -Sobolev space Lpk .Hn / into itself, for 1 < p < 1 and k 2 ZC .
12.2.3 Laguerre Functions .˛/
Consider the generalized Laguerre polynomials Lk .x/ defined by their usual generating function formula 1 X kD1
.˛/
Lk .x/wk D
n xw o 1 exp ; for ˛ D 0; 1; 2; : : : ; x 0; .1 w/˛C1 1w and jwj < 1:
(12.2.12)
12.2 Laguerre Calculus
295
Definition 12.2.3. Let z D jzje i and k; p D 0; 1; 2; : : :. Then we define 1=2
.k C 1/ 2 .p/ e .p/ .z; / D 2jj W .2jjjzj2 /p=2 e ip e jjjzj Lk .2jjjzj2 /; k .k C p C 1/ (12.2.13) 1=2
.k C 1/ e .p/ .z; / D 2jj .1/p W k
.k C p C 1/ 2
.2jjjzj2 /p=2 e ip e jjjzj L.p/ .2jjjzj2 /: k
(12.2.14)
12.2.4 Laguerre Calculus on H1 e .p/ The most important property of W k .z; / is the following theorem of Greiner [57]: Theorem 12.2.4. Let p; k; q; m D 1; 2; : : :. Then e .pk/ jj W e .qm/ D ı .q/ W e .pm/ ; W .p^k/1 .q^m/1 k .p^m/1 .p/ e .qk/ e .qm/ D ım e .pk/ jj W W W .q^k/1 ; .p^k/1 .q^m/1
(12.2.15)
.q/
where a ^ b D min.a; b/ and ık denotes the Kronecker delta function; i.e., ık.q/ D 1 if k D q and vanishes otherwise. e .p/ is another function of the Thus the twisted convolution of two functions of W k same type. This surprising result justifies the use of Laguerre function expansion on the Heisenberg group in analogy with Mikhlin’s use of the spherical harmonics on Rn . .p/ Let Wk .z; t/; ˙ p; k D 0; 1; 2; : : :, be the inverse Fourier transform of .p/ e .z; / with respect to i.e., W k Wk.p/ .z; t/ D
1 2
Z
1 1
e .p/ .z; /d : ei t W k
These are the kernels of the generalized Cauchy–Szeg¨o operators on H1 . In particular, .0/
.0/
.0/
W0 .z; t/ D WC;0 .z; t/CW;0 .z; t/ D SC CS ; where S˙ D
1 1 2 .jzj2 i t/2 (12.2.16)
denotes the Cauchy–Szeg¨o kernels in H1 . The following result implies that the generalized Cauchy–Szeg¨o kernels indeed induce principal value convolution operators.
296
12 Finding Heat Kernels Using the Laguerre Calculus .p/
Theorem 12.2.5. The generalized Cauchy–Szeg¨o kernels Wk .z; t/ are in C 1 .H1 n f0g/ and have zero mean value property. Formula (12.2.15) is reminiscent of matrix multiplication. To show this similarity, Greiner [57] introduced the Laguerre matrix: Definition 12.2.6. We define the positive Laguerre matrix e .pk/ MC W .p^k/1 e of W to be the infinite matrix which has one at the intersection of the .p^k/1 pth row and kth column and zeros everywhere else. The negative Laguerre matrix .pk/ e M W .p^k/1 can be defined as the transpose of the positive Laguerre matrix. .pk/
Following this definition, Theorem 12.2.4 takes the following form: .pk/ .qm/ .qm/ e .q^m/1 e .pk/ jj W e .p^k/1 MC W e .q^m/1 : MC W D M W C .p^k/1 (12.2.17) The Laguerre matrix for any left invariant convolution operator can be defined as in [57] and [17]. In particular, we can define the Laguerre matrix for any left invariant differential operator on H1 , since we can write it in the form of convolution operators. We omit the details here, and the interested reader can consult [17] and Berenstein–Chang–Tie [19]. We only list the following results: Q Theorem 12.2.7. If F and G are two P.V. convolution operators on H1 , and M.F/ Q and M.G/ denote the Laguerre matrices of F and G, respectively, then e D M.e e D MC .e e ˚ M .e e (12.2.18) M.e F G/ F/ M.G/ F/ MC .G/ F/ M .G/: A simple consequence of this theorem is Corollary 12.2.8. Let I denote the identity operator on C01 .H1 /. Then I is induced .p/ by the identity Laguerre matrix W˙ .e I/ D .ık /.
12.2.5 Laguerre Calculus on Hn We define the n-dimensional version of the exponential Laguerre functions on Hn by the n-fold product: e .p/ .z; / D W k
n Y j D1
e .pj / .paj zj ; /; aj W kj
e .pj / .paj zj ; / is given by (12.2.13). where W kj
(12.2.19)
12.2 Laguerre Calculus
297
Composing two functions of the type (12.2.19) via twisted convolution yields 2 e .p/ W e .q/ 4 W m D k
n Y
3 .pj /
e aj W k
j
j D1
D
n Y j D1
aj2
Z
2
p . aj zj ; /5 4
3 j/ p e .q 5 aj W mj . aj zj ; /
j D1 .pj /
R2
n Y
e e 2i aj Im.zj wN j / W k
j
p j/ e .q . aj .zj wj /; /W mj
p . aj wj ; /d wj n Y j/ p e .pj / W e .q D aj W mj . aj zj ; /: k j D1
j
Consequently, Theorem 12.2.4 implies that Theorem 12.2.9. Let kj ; pj ; mj and qj D 1; 2; 3; : : : for j D 1; 2; : : : ; n. Then e .qm/ D ı .q/ W e .pm/ ; e .pk/ jj W W k .k^p/1 .m^q/1 .p^m/1 .p/ e .qk/ e .pk/ jj W e .qm/ D ım W .q^k/1 ; W .k^p/1 .m^q/1
(12.2.20)
where k D .k1 ; k2 ; : : : ; kn /; p D .p1 ; p2 ; : : : ; pn /; m D .m1 ; m2 ; : : : ; mn / and q D .q1 ; q2 ; : : : ; qn /. Here .k ^ p/ 1 D .min.p1 ; k1 / 1; : : : ; min.pn ; kn / 1/ .p/
and ık D
Qn
.pj / j D1 ıkj
is the n-fold Kronecker delta function.
Instead of the Laguerre matrix, Beals et al. [17] introduced the definition of a Laguerre tensor for the convolution operators on Hn . We omit the details here. Then we have the n-fold version of Theorem 12.2.7: Theorem 12.2.10. (The Laguerre calculus on Hn ) Let F and G induce the convolue denote the Laguerre tensors of F and tion operators on Hn . Let M.e F/ and M.G/ e e e e G, respectively. Then M.F G/ D M.F/ M.G/. Corollary 12.2.11. The identity operator I on C01 .Hn / is induced by the identity Laguerre tensor: I/ D .ık.p1 1 / ık.pnn / /: M˙ .e
12.2.6 Left Invariant Differential Operators A left invariant differential operator P on Hn is a polynomial P.X; Y; T/ with constant coefficients, or in complex coordinates, a polynomial in vector fields T, Zj ,
298
12 Finding Heat Kernels Using the Laguerre Calculus
and ZN j . We can have the following representation for P as a convolution operator on Hn : P D PI D
1 X jkjD0
PWk.0;:::;0/ ; ;:::;kn 1
where I D
1 X jkjD0
Wk.0;:::;0/ ;:::;kn 1
(12.2.21)
is the identity operator on C01 .Hn /. In particular, T, Zj , and ZN j , j D 1; 2; : : : ; n; can be represented as convolution operators and written in the Laguerre tensor forms. This is contained in the next proposition: Proposition 12.2.12. (1) M.e T/ is the i multiple of the identity Laguerre tensor: .p / .p / M˙ .e T/ D i.ık1 1 : : : ıknn /:
(2) Zj ; j D 1; 2; : : : ; n, has the following Laguerre tensor representation: M.e Zj / D MC .e Zj / ˚ M .e Zj /; where q .p / .p C1/ ;:::;pn / M .e Zj /k.p11;:::;k D 2aj pj jjık1 j : : : ık j : : : ık.pnn / n j
and MC .e Zj / D M .e Zj /t : N j , j D 1; 2; : : : ; n, has the following Laguerre tensor representation: (3) Z N j / D M.e Zj /t : M.e Z N T/ D P.Z1 ; : : : ; Zn ; ZN 1 ; : : : ; ZN n ; T/ denote a Theorem 12.2.13. Let P D P.Z; Z; left invariant differential operator on Hn i.e., P is a polynomial in the vector fields T, Zj , and ZN j , j D 1; 2; : : : ; n. Then e D P.M.e N i /; M.P/ Z/; M.e Z/;
(12.2.22)
N D .M.e Zn // and M.e Z/ ZN 1 /; : : : ; where we set M.e Z/ D .M.e Z1 /; : : : ; M.e e N n //. M.Z Example 12.2.14. Assume that n D 1 and a1 D 1. Then we have 2
0 6 p 60 MC .e Z1 / D 2jj 60 4 :: : and
p
1 p0 0 2 p0 0 0 0 3 :: :: :: : : :
t Z1 / D MC .e Z1 / : M .e
3 7 7 ; 7 5 :: :
12.2 Laguerre Calculus
299
Now we may set 2
0
6 p1 6 1 1 6 6 0 MC .K/ D p 6 2jj 6 6 0 4 :: : and
0 0 p1 2
0 :: :
3 0 0 7 7 7 0 7 7; 7 0 7 5 :: : : : :
0 0 0 p1 3
:: :
t M .K/ D MC .K/ :
Thus 1 X 1 e .1/ e ˙ .z; / D p 1 p K W ˙;k .z; /: 2jj kD0 k C 1 .1/
e ˙;k .z; /, we sum the series Using the definition of W 2
jjjzj e / D 2jjze K.z;
Z
1 1X 0
.1/
r k Lk .2jjjzj2 /dr:
kD0
Since 1 X
r k L.1/ .x/ D k
kD0
x ex e 1r ; .1 r/2
therefore, 2
jjjzj e / D 1 e K.z; ; zN
and 1 K.z; t/ D 2 2 zN
Z
C1
1
2
e i t jjjzj d D
z : 2 .jzj4 C t 2 /
Hence, we recover the Greiner, Kohn and Stein theorem [60]: .0/ D I S ; Z1 K D I W;0 .0/ KZ1 D I WC;0 D I SC :
Here S˙ are the Cauchy–Szeg¨o kernels which were defined by (12.2.16).
300
12 Finding Heat Kernels Using the Laguerre Calculus
12.3 The Heisenberg Sub-Laplacian The Heisenberg sub-Laplacian is the differential operator L˛ D
n 1X .Zj ZN j C ZN j Zj / C i ˛T 2
(12.3.23)
j D1
with Zj and ZN j given by (12.1.4). In the case of aj D 1 for all j , the operator L˛ was first introduced by Folland and Stein [45] in the study of @N b complex on a nondegenerate CR manifold. They found the fundamental solution of L˛ . Beals and Greiner [18] solved the case when the aj are different. Next we consider two operators related to L˛ . The first one is the powers of L˛ . In the special case of aj D 1=4 for all j D 1; 2; : : : ; n and ˛ D 0, Lm 0 , 1 m n, has been studied by Dolley, Benson and Ratcliff [40]. If all the aj are equal, the unitary group U.n/ acts on Hn via uŒz; t D Œu.z/; t
for u 2 U.n/;
Œz; t 2 Hn :
(12.3.24)
L˛ is invariant under the U.n/-action. The key idea in [40] is to exploit this invariance by expanding the fundamental solution in terms of U.n/-spherical functions ;m which are given by 2 =4
;m .z; t/ D e i t e jjjzj
.n1/
Lm1 .j jjzj2 =2/
.n1/
with 2 R n f0g, m 2 ZC . Here Lm denotes the generalized Laguerre polynomial. However, the invariance (12.3.24) does not hold for the nonisotropic case. Instead, we shall apply the Laguerre calculus to solve the general case. Since the Laguerre tensor of Lm ˛ is simply diagonal, its inverse is also simple and diagonal. As usual, taking the partial Fourier transform with respect to the t-variable, one has 3 2 1 n n X X Y 1 e e .0/ .paj zj ; / : 4 ZN j C e L˛ D e L˛e .e Zj e ZN j e aj W ID Zj / ˛ 5 kj 2 jkjD0
j D1
j D1
(12.3.25) A simple calculation yields 1 e .0/ .paj zj ; /: (12.3.26) e .0/ .paj zj ; / D .2k C 1/jjaj W ZN j C e ZN j e .e Zj e Zj /W k k 2 Thus (12.3.25) and (12.3.26) imply 1 0 1 n n X X Y e e .0/ .paj zj ; / : @ L˛ D .2kj C 1/jjaj ˛ A aj W kj jkjD0
j D1
j D1
(12.3.27)
12.3 The Heisenberg Sub-Laplacian
301
Consequently, the Laguerre tensor of the convolution operator induced by L˛ is 02
n X
M.e L˛ / D jj @4
3
1
.p / .p / .2kj C 1/aj ˛sgn./5 ık1 1 ıknn A ;
(12.3.28)
j D1
which is invertible as long as ˛ does not belong to the exceptional set ƒ˛ , where ƒ˛ D
n X
.2kj C 1/aj ;
k D .k1 ; k2 ; : : : ; kn / 2 .ZC /n :
j D1
According to Theorem 12.2.10, the inverse Laguerre tensor of (12.3.28) is 02 1 31 n X 1 B C .2kj C 1/aj ˛sgn./5 ık.p1 1 / ık.pnn / A : (12.3.29) M.e L˛ / D jj1 @4 j D1
If we write it in the Laguerre series expansion: 11 0 n n 1 X X Y 1 e e .0/ .paj zj ; /; A @ .z; / D jj .2k C 1/a ˛sgn./ aj W L1 j j ˛ k j D1
jkjD0
j D1
j
(12.3.30) then we can sum this series in the sense of Abel, take the inverse Fourier transform with respect to , and find the fundamental solution of L˛ . In fact, the above calculation makes sense for any polynomial of L˛ . We will carry out this extension, and N find the fundamental solution of Lm ˛ and the Paneitz operator L˛ L˛ , in the rest of this chapter. Remark 12.3.1. Being concerned about any possible confusions regarding the formulas (12.3.25) and (12.3.27), we shall clarify the notation one more time. The operators L˛ and P˛ are left invariant. So when they act on a function f , we may write L˛ .f / and P˛ .f /. Taking the partial Fourier transform with respect to f˛ fQ; and P f˛ fQ. Here the t-variable (whose dual variable is ), one has L means “twisted convolution,” concept defined in this chapter. So, removing the f˛ and P f˛ . We may decompose function variable fQ, the operators become L these operators into Laguerre tensor expansions, which are given by (12.3.25) and (12.3.27). Before going further, we shall say a few words about the exceptional set ƒ˛ of the operator L˛ . Basically, we look at the exceptional set from a different point of view. From the calculations of this section we know that ˛ ¤ ˙
n X
.2kj C 1/aj W
j D1
k D .k1 ; k2 ; : : : ; kn / 2 ZnC :
302
12 Finding Heat Kernels Using the Laguerre Calculus
Let us consider the “real part” of the operator L˛ : L D 4L0 D
2n X
Xj2
j D1
of the Kohn Laplacian and the operator T D @t@ . Since L and iT are essentially self-adjoint strongly commuting operators, there is a well-defined joint spectrum. We define the joint spectrum of the pair .L; iT / as the complement of the set of . ; / 2 C2 for which there exists Lp -bounded operators A and B with A. I L/ C B. I iT / D I: This implies that the spectrum should be the set of . ; / 2 C2 for which neither I L nor I iT is invertible. From the Laguerre calculus, we know that a convolution operator is invertible if and only if its Laguerre tensor is invertible. So we reduce the invertibility of an operator to that of its Laguerre tensor. In fact, we know that the Laguerre tensor of the operator I L is 0 1 n X
.p1 / M. e Ie L/ D @ jj aj .2k1 C 1/ ık1 ık.pnn / A j D1
and the Laguerre tensor of the operator I iT is
e/ D C ı .p1 / ı .pn / : M. e I iT k1 kn Therefore, . I L/
is invertible , jj
n X
aj .2k1 C 1/ ¤ 0
j D1
for all k 2 .ZC /n , 2 R and . I iT / is invertible , C ¤ 0: Hence the spectrum of L alone is the set of nonnegative numbers f 2 R W 0g, the spectrum of iT is the set of real numbers R, and the joint spectrum of .L; iT / is the union of 8 9 n < = X . ; / 2 C2 W D j j .2kj C 1/aj and 2 R : ; j D1
over the set k 2 .ZC /n . We can also write the joint spectrum in the form ( ) [ ˙ . ; / 2 C2 W 0; D Pn : .L; iT / D j D1 .2kj C 1/aj n k2.ZC /
(12.3.31) The set (12.3.31) (see Fig. 12.1) is called the Heisenberg brush (see [104]).
12.4 Powers of the Sub-Laplacian
303
Fig. 12.1 The Heisenberg brush in the isotropic case a1 D D a n D a
12.4 Powers of the Sub-Laplacian In this section we consider the powers of the Heisenberg sub-Laplacian Lm ˛ . For 1 m n, this work has been done by Dolley, Benson and Ratcliff [40] in the case when all the aj are equal. They used a different method. First we will find the Laguerre tensor of the operator Lm ˛ from the Laguerre tensor of L˛ . Similarly to (12.3.25), we can take the Fourier transform with respect to t, and write e Lm ˛ as a twisted convolution form: e eme Lm ˛ D L˛ I D
2
3m n n X Y 1 e .0/ .paj zj ; / : 4 Zj / ˛ 5 .e Zj e ZN j e aj W ZN j C e kj 2
1 X
j D1
jkjD0
j D1
(12.4.32) Then (12.3.26) yields e Lm ˛ D
1 X jkjD0
0 @
n X
1m .2kj C 1/jjaj ˛ A
j D1
n Y j D1
e .0/ .paj zj ; / : (12.4.33) aj W kj
Consequently, the Laguerre tensor of the convolution operator induced by Lm ˛ is 02 m M.e L˛ / D jjm @4
n X
3m
1
.p / .p / .2kj C 1/aj ˛sgn./5 ık1 1 ıknn A ;
(12.4.34)
j D1
which is invertible as long as ˛ does not belong to the singular set ƒ˛ , where ƒ˛ D
8 < :
˙
n X
.2kj C 1/aj W
j D1
k D .k1 ; k2 ; : : : ; kn / 2 ZnC
9 = ;
:
304
12 Finding Heat Kernels Using the Laguerre Calculus
According to Theorem 12.2.10, the inverse Laguerre tensor of (12.6.54) is 02 m M.e L˛ / D jjm @4
n X
1
3m
.p / .p / ık1 1 ıknn A ;
.2kj C 1/aj ˛sgn./5
j D1
(12.4.35) em .z; / in the following Laguerre series expansion: and we write its kernel ‰ 1m 0 n n 1 X X Y m e e .0/ .paj zj ; /: A @ .2kj C 1/aj ˛sgn./ aj W ‰m .z; / D jj k j D1
jkjD0
j D1
j
(12.4.36) To find the fundamental solution of Lm ˛ , we can sum this series and take the inverse Fourier transform with respect to . First we introduce the following integral representation of Am : 1 1 D m A
.m/
Z
1
s m1 e As ds
for Re.A/ > 0:
(12.4.37)
0
Then, we can write (12.6.56) in the following form: 1 Z 1 m X Pn em .z; / D jj ‰ s m1 e . j D1 .2kj C1/aj ˛sgn.//s ds
.m/ 0 jkjD0
n Y
e .0/ .paj zj ; /: aj W k
(12.4.38)
j
j D1
Next, we interchange the summation and the integration, and use the definition e .0/ : of W kj
em .z; / ‰ jjm D
.m/
Z
1 0
jjnm D n .m/
n Y j D1
s m1
1 X
Pn
e .
j D1 .2kj C1/aj ˛sgn./
/s ds
j D1
jkjD0
Z
1
s 0
m1
n Y
1 X
e .
Pn
jkjD0 2
2aj e aj jjjzj j L.0/ .2aj jjjzj j2 / k j
/s ds
j D1 .2kj C1/aj ˛sgn./
e .0/ .paj zj ; / aj W k j
12.4 Powers of the Sub-Laplacian
D
jjnm n .m/
1 X
Z
1
305 n Y
s m1 e ˛sgn./s
0
2
2aj e aj saj jjjzj j
j D1
.e 2aj s /kj L.0/ .2aj jjjzj j2 /ds; k j
kj D0
Applying the generating formula for the Laguerre polynomials 1 X
1 xz D exp .1 z/pC1 1z
L.p/ .x/zk k
kD0
em .z; /, we obtain to the last formula for ‰ Z
nm e m .z; / D jj ‰ n .m/
1
s
m1 ˛sgn./s
e
0
n Y 2aj e aj s 1 e 2aj s j D1
2e 2aj s exp aj jjjzj j 1 C ds 1 e 2aj s 3 2 n nm Z 1 Y jj aj 5 D n s m1 e ˛sgn./s 4 .m/ 0 sinh.aj s/ 2
j D1
8 <
exp jj :
n X j D1
9 =
aj jzj j2 coth.aj s/ ds: ;
P Letting .sI z/ D njD1 aj jzj j2 coth.aj s/ and taking the inverse Fourier transform with respect to , we obtain for any 0 m n
.n m C 1/ ‰m .z; t/ D 2 nC1 .m/ Z
1
s
m1 ˛s
e
s m1 e ˛s @
n Y
j D1
Z
1 1
n Y j D1
@
0
.n m/ D 2 nC1 .m/
0
0
0
1
C
"Z
0 @
1 aj ds A sinh.aj s/ Œ .sI z/ i tnmC1
1 ds aj A sinh.aj s/ Œ .sI z/ C i tnmC1
n Y
j D1
#
1 e ˛s s m1 ds aj A : sinh.aj s/ Œ .sI z/ i tnmC1
(12.4.39)
(12.4.40)
306
12 Finding Heat Kernels Using the Laguerre Calculus
In the above calculation, we applied the following integral formula: Z 1
.n m C 1/e ˛s
.n m C 1/e ˛s jjnm e ˛sgn./sCi t jj.sIz/ d D C nmC1 Œ .sI z/ i t Œ .sI z/ C i tnmC1 1 (12.4.41) to get (12.4.39), and then we substituted s with s in the second integral of (12.4.39) to obtain (12.4.40) using .sI z/ D .sI z/. Following [16] and [28], we introduce the complex distance and volume element on the Heisenberg group: g.sI z; t/ D .2sI z/ i t
and .s/ D
n Y j D1
2aj : sinh.2aj s/
(12.4.42)
We can write (12.4.40) in the closed form ‰m .z; t/ D
2m .n m/Š .2/nC1 .m/
Z
1
e 2˛s s m1
1
.s/ds : Œg.sI z; t/nmC1
(12.4.43)
When jzj D 0, t ¤ 0, then g.sI z; t/ D i t. The integrand of (12.4.43) is not integrable at s D 0. To regularize the integration, we must deform its path of integration from .1; 1/ to .1 C i "sgnt; 1 C i "sgnt/;
: 1j n 2aj
where 0 < " < min
We refer to [18] for the exact definition of this path. Finally, we have the formula ‰m .z; t/ D
2m .n m/Š .2/nC1 .m/
Z
1Ci "sgnt
e 2˛s s m1
1Ci "sgnt
.s/ds : Œg.sI z; t/nmC1
Remark 12.4.1. As we noted in the introduction, if all the aj are equal, the unitary group U.n/ acts on Hn via u.z; t/ D .uz; t/ for u 2 U.n/;
.z; t/ 2 Hn :
The operator Lm ˛ is invariant under the U.n/-action. The key idea in [40] is to exploit this invariance.
12.5 Heat Kernel for the Operator Lm ˛ In this section we shall compute the kernel of the fundamental solution for the powers of the Kohn Laplacian via the heat kernel hs .z; t/ D expfsL˛ gı0 . In the isotropic case, the heat kernel was independently studied by Gaveau [48] via
12.5 Heat Kernel for the Operator Lm ˛
307
the probability method and Hulanicki [68] using the Fourier transform on Hn and the basis of Laguerre functions. Later, Beals and Greiner [18] solved the general case by a different method. We will see that hs .z; t/ can be obtained easily using the Laguerre calculus. Our approach is more closely related to the method of Hulanicki. We first take the Fourier transform with respect to the t-variable and write the heat kernel e hs .z; t/ as a twisted convolution operator: 1 X
e L˛ ge ID hs .z; / D expfse
2 expfse L˛ g 4
j D1
jkjD0
D
1 X
e
s
Pn
j D1
n Y
aj jj.2kj C1/Cs˛
n Y j D1
jkjD0
3 p e .0/ 5 aj W kj . aj zj ; /
p e .0/ aj W kj . aj zj ; /:
Next, a similar computation as in the last section leads to 9 8 2 3 n n = < ˛s Y X e a jj j e 5 exp jj aj jzj j2 coth.aj jjs/ : hs .z; / D n 4 ; : sinh.aj jjs/ j D1
j D1
(12.5.44) Since n Y j D1
n Y aj jj aj D sinh.aj jjs/ sinh.aj s/
and jj coth.aj jjs/ D coth.aj s/;
j D1
we can simplify (12.5.44) by removing the absolute sign for and have 2 3 n ˛s Y a e j e 5 e .sIz/ ; hs .z; / D n 4 sinh.aj s/
(12.5.45)
j D1
P where, as before, .sI z/ D njD1 aj jzj j2 coth.aj s/. Now we take the inverse Fourier transform with respect to the -variable and obtain the heat kernel in the integral form 1 hs .z; t/ D 2 nC1
Z
C1
2 4
1
1 D 2.s/nC1
Z
n Y j D1
C1 1
2 4
3 aj 5 e ˛sCi t .sIz/ d sinh.aj s/
n Y j D1
3 aj 5 ˛Ci t .z;/ s s e d: sinh.aj /
308
12 Finding Heat Kernels Using the Laguerre Calculus
We substitute by 2 and rewrite the heat kernel in terms of the complex distance g and volume element of the Heisenberg group: hs .z; t/ D
1 .s/nC1
Z
C1
./ n e 2˛
2 s
g.Iz;t /
d :
(12.5.46)
1
Now we may combine the heat kernel hs .z; t/ and the formula m
1 D
.m/
Z
1
s m1 e s ds;
Re. / > 0
and Re.m/ > 0
0
to give another derivation of the fundamental solution of Lm ˛ , 0 m n. We can first write the inverse formally as Lm ˛ .z; t/ D
1
.m/
Z
1
s m1 e sL˛ ds
0
1 D nC1
.m/
Z
1
s
mn2
0
Z
C1
./ n e 2˛
2 s
g.Iz;t /
dds:
1
By Fubini’s theorem, if z ¤ 0, we may interchange the order of integration. Moreover, if Re.m/ > 0, one has Z
1
˛
s mn e s
0
ds D s2
Z
1
s nm e ˛s ds D
0
.n m C 1/ .n m/Š D nmC1 ; ˛ nmC1 ˛
so that we may rewrite Lm ˛ .z; t/ as 2m .n m/Š ‰m .z; t/ D .2/nC1 .m/
Z
1
1
./ m1 e 2˛ d : Œg.I z; t/nmC1
(12.5.47)
This is exactly (12.4.43) obtained in the last section. When m > n, we cannot apply Fubini’s theorem since the integral in the variable s diverges as s ! 1. Similarly, we may consider the kernel induced by the operator ‰ˇ .z; t/ D Liˇ ˛ ı0 . We may combine the heat kernel in (12.5.46) with the following formula: wiˇ D
1
.iˇ/
Z
1
s iˇ 1 e sw ds;
Re.w/ > 0
and Im.ˇ/ > 0:
0
This leads to Liˇ ˛ .z; t/
.n C iˇ C 1/ D nCiˇ C1 2
.iˇ/
Z
1 1
.s/s iˇ 1 e 2˛ ds: Œg.sI z; t/nCiˇ C1
(12.5.48)
12.6 Fundamental Solution of the Paneitz Operator
309
By analytic continuation, formula (12.5.48) remains true in the case of real ˇ. This operator was studied by M¨uller and Stein [93] in the case of aj D 1 for all j D 1; : : : ; n and ˛ D 0. They have shown that the weak-type (1,1)-norm of the operator .2nC1/=2 as jˇj ! 1. Liˇ 0 is bounded from below by C jˇj We shall mention another application of formulas (12.5.47) and (12.4.43). 1=2 Replacing m in (12.5.47) by 12 , we obtain the kernel for L0 . Then kernels for the 1
1
Riesz transforms Rj D Zj L0 2 and RnCj D ZN j L0 2 , j D 1; : : : ; n, which were first defined by Christ and Geller [36], can be obtained by composing Zj or ZN j to L12 0 . Here are the formulas:
1
Zj L0 2 D
.n C 12 /aj zNj 1
Z
3
2nC 2 nC 2
1
1
.s/s nC1 Œcoth.2aj s/ C 1 ds Œsg.sI z; t/nC3=2
(12.5.49)
.s/s nC1 Œcoth.2aj s/ 1 ds: Œsg.sI z; t/nC3=2
(12.5.50)
and 1
ZN j L0 2 D
.n C 12 /aj zj 2
nC 1 2
nC 3 2
Z
1 1
It is easy to see that these kernels are H-homogeneous of degree 2n 2 and satisfy the mean value zero property because the factor zN j or zj appears (see the next section for details).
12.6 Fundamental Solution of the Paneitz Operator In this section, we consider the Paneitz operator P˛ on the Heisenberg group, which is defined as 32 2 n X 1 P˛ D L˛ LN ˛ D 4 .Zj ZN j C ZN j Zj /5 C ˛ 2 T2 : 4
(12.6.51)
j D1
We shall about ˚ first recall ˚ n some background n the Paneitz operator. Consider the coframe ; !j ; !N j j D1 dual to T; Zj ; ZN j j D1 . Let ' 2 C01 .Hn / be a smooth function with compact support. It is easy to see that Ln ' D
n n X 1X N j Zj / C .i nT/.'/ D ZN j Zj .'/: .Zj ZN j C Z 2 j D1
j D1
310
12 Finding Heat Kernels Using the Laguerre Calculus
It follows that Pn ' D Ln LNn ' " n # i h 1 X D .Zj ZN j C ZN j Zj / C 2i nT .Zk ZN k C ZN k Zk / 2i nT ' 4 j;kD1
9 8 2 32 > ˆ n = <1 X N j Zj /5 C n2 T2 ' 4 .Zj ZN j C Z D > ˆ ; : 4 j D1 D4
n X
ZN k .Pk '/:
kD1
Here (see [87]) Pk ' D
n X
Zk .Zj ZN j '/;
k D 1; : : : ; n;
j D1
and P' D
n X
.Pk '/!k
kD1
is an operator that characterizes the CR-pluriharmonic functions. Note that a smooth real-valued function u on Hn is said to be a CR-pluriharmonic function if for any point p 2 Hn ; there are an open neighborhood U of p in Hn P and a smooth realvalued function v on U such that @N b .u C i v/ D 0. Here @N b u D njD1 .ZN j u/!N j . Moreover, one has Z Hn
hP ' C PN '; db 'i d D
Z Pn ' ' d ;
8' 2 C01 .Hn /:
Hn
Here db D ı d W C 1 .Hn / ! spanf!j ; !N j gnjD1 and is the orthogonal projection onto the subspace spanf!j ; !N j gnjD1 in the cotangent space. Moreover, a smooth real-valued function ' 2 L2 .Hn / satisfies Pn ' D 0 on Hn if and only if P ' D 0 on Hn . Note that via the CR-Paneitz operator Pn ; we are able to get a sub-gradient estimate and establish Liouville-type theorems of the CR-heat equation on Hn (see [32] and [34]). For complex geometric aspects about these operators, the reader is referred to Graham and Lee [56], Hirachi [64] and Lee [87].
12.6 Fundamental Solution of the Paneitz Operator
311
12.6.1 Laguerre Tensor of the Paneitz Operator First we will find the Laguerre tensor of the operator P˛ from the Laguerre tensor of L˛ . Similarly to (12.3.25), we can take the Fourier transform with respect to t, e˛ as a twisted convolution form: and write P e˛ D e P L˛ e LN ˛e I 32 3 2 1 n n X X X 1 1 4 .e Zj e ZN j e .e Zj e ZN j e Zj / ˛ 5 4 Zj / C ˛ 5 ZN j C e ZN j C e D 2 2 j D1
jkjD0
n Y
j D1
e .0/ .paj zj ; / : aj W k
j D1
(12.6.52)
j
Then (12.3.26) yields 82 9 32 ˆ > 1 n n < =Y X X 2 2 e .0/ .paj zj ; / : e 4 5 P˛ D .2kj C 1/jjaj ˛ aj W kj ˆ > ; j D1 jkjD0 : j D1 (12.6.53) Consequently, the Laguerre tensor of the convolution operator induced by P˛ is 020 3 1 12 n X 6 .p / C 2 7 .p / e˛ / D 2 B (12.6.54) M.P @4@ .2kj C 1/aj A ˛ 5 ık1 1 ıknn A ; j D1
which is invertible as long as ˛ does not belong to the exceptional set ƒ˛ , where 8 9 n < X = .2kj C 1/aj I k D .k1 ; k2 ; : : : ; kn / 2 ZnC : ƒ˛ D ˙ : ; j D1
According to Theorem 12.2.10, the inverse Laguerre tensor of (12.6.54) is 020 31 1 12 n X 7 .p / .p / C 2 B6@ e1 .2kj C 1/aj A ˛2 5 ık1 1 ıknn A ; (12.6.55) M.P @4 ˛ /D j D1
e˛ .z; / in the Laguerre series expansion: and we write its kernel ‰
e ˛ .z; / D 2 ‰
20 31 12 n 1 n X X Y 6@ 27 e .0/ .paj zj ; /: A .2k C 1/a ˛ aj W 4 5 j j kj
jkjD0
j D1
j D1
(12.6.56)
312
12 Finding Heat Kernels Using the Laguerre Calculus
In order to find the fundamental solution of P˛ , we may sum this series and take the inverse partial Fourier transform with respect to the -variable. Consider the integral representation of A1 given by (12.6.57): Z 1 1 1 e As ds for Re.A/ > 0: (12.6.57) D A
.m/ 0 Let A.k/ D
Pn
C 1/aj . Note that
j D1 .2kj
1 1 D A2 .k/ ˛ 2 2˛
1 1 : A.k/ ˛ A.k/ C ˛
Assume that j˛j < a1 C a2 C C an . Then we apply (12.6.57) and write (12.6.56) in the following form: e ˛ .z; / D ‰
1 Z i P 1 X 1 h .PnjD1 .2kj C1/aj ˛/s . n j D1 .2kj C1/aj C˛ /s ds e e 2˛ 2 0 jkjD0
n Y
e .0/ .paj zj ; /: aj W k
(12.6.58)
j
j D1
Next, we interchange the summation and integration, and use the definition e .0/ : of W k j
e˛ .z; / D ‰
1 2˛jj2
n Y j D1
D
n Y
1
0
Pn
e .
j D1 .2kj C1/aj ˛
Pn
/s e .
j D1 .2kj C1/aj C˛
/s
i
jkjD0
Z
1 1X 0
h
Pn
e .
j D1 .2kj C1/aj ˛
Pn
/s e .
j D1 .2kj C1/aj C˛
i /s ds;
jkjD0 2
2aj e aj jjjzj j L.0/ .2aj jjjzj j2 / k j
j D1
D
1 h X
e .0/ .paj zj ; / aj W kj
jjn2 2˛ n
Z
jjn2 2˛ n
Z
1 0
Œˆ˛ .z; I s/ ˆ˛ .z; I s/ ds;
where ˆ˛ .z; I s/ D e ˛s
n Y j D1
2
2aj e aj saj jjjzj j
1 X
.e 2aj s /kj L.0/ .2aj jjjzj j2 /: k
kj D0
j
12.6 Fundamental Solution of the Paneitz Operator
313
Applying the generating formula for the Laguerre polynomials 1 X
.p/ Lk .x/zk
kD0
1 xz D exp .1 z/pC1 1z
to ˆ.z; I s/, we obtain n Y 2e 2aj s 2aj e aj s 2 ˆ˛ .z; I s/ D e exp aj jjjzj j 1 C 1 e 2aj s 1 e 2aj s j D1 8 9 2 3 n n < = Y X aj 5 exp jj D e ˛s 4 aj jzj j2 coth.aj s/ : : ; sinh.aj s/ ˛s
j D1
This yields
j D1
3 2 n n2 Z 1 Y a jj j e˛ .z; / D 5 ‰ sinh.˛s/ 4 ˛ n 0 sinh.aj s/ j D1 8 9 n < = X exp jj aj jzj j2 coth.aj s/ ds: : ;
(12.6.59)
j D1
12.6.2 Fundamental Solution: The Case n 2 To simplify the notation, we introduce the function .z; s/ D
n X
aj jzj j2 coth.aj s/:
j D1
We next take the inverse Fourier transform with respect to and find 2 3 Z 1 Z 1 n Y a 1 j 5e i t jj.z;s/ ds d: ‰˛ .z; t/ D jjn2 sinh.˛s/4 2˛ nC1 1 sinh.aj s/ 0 j D1
Changing the order of integration, we obtain " n # Z Y
.n 1/ 1 aj sinh.˛s/ ‰˛ .z; t/ D 2˛ nC1 0 sinh.aj s/ j D1
1 1 ds: C Œ .z; s/ i tn1 Œ .z; s/ C i tn1
314
12 Finding Heat Kernels Using the Laguerre Calculus
For the second part of the integral, we substitute s with s, and note that sinh.˛s/ D sinh.˛s/ and .z; s/ D .z; s/ since sinh and coth are odd functions. Hence # Z 1"Y n aj sinh.˛s/ ds sinh.aj s/ Œ .z; t/ C i tn1 0 j D1 # Z 1 " Y n aj sinh.˛s/ .1/ds D sinh.aj s/ Œ .z; t/ C i tn1 0 j D1 # " n Z 0 Y .1/ aj sinh.˛s/ ds D nCn1 sinh.aj s/ Œ .z; t/ i tn1 1 .1/ j D1 # Z 0 "Y n aj sinh.˛s/ ds: D sinh.aj s/ Œ .z; t/ i tn1 1 j D1
We can write ‰˛ .z; t/ in the compact form ‰˛ .z; t/ D
.n 1/ 2˛ nC1
Z
1
2 4
1
3
n Y
j D1
aj sinh.˛s/ 5 ds: sinh.aj s/ Œ .z; s/ i tn1
(12.6.60)
As we mentioned before, we may rewrite the fundamental solution in terms of the modified complex action g.sI z; t/ and volume element .s/ which were defined in (12.4.42): n X
g.sI z; t/ D .2sI z/ i t D
aj jzj j2 coth.2aj s/ i t
and
j D1
.s/ D
n Y j D1
2aj : sinh.2aj s/
Substituting s with 2s, one obtains
.n 1/ ‰˛ .z; t/ D n nC1 2 ˛
.n 1/ D n nC1 2 ˛
Z
1
1
Z
1
1
"
n Y j D1
# sinh.2˛s/ 2aj ds sinh.2aj s/ Œg.sI z; t/n1
.s/ sinh.2˛s/ ds: Œg.sI z; t/n1
It seems impossible to find the exact formula for the above integral in general. We will consider the special case of aj D a for all j . In this case,
.n 1/ ‰˛ .z; t/ D 2˛ nC1
Z
1
1
h
in sinh.˛s/ a ds: 2 sinh.as/ Œajzj coth.as/ i tn1
12.6 Fundamental Solution of the Paneitz Operator
315
Differentiating with respect to t, we obtain @‰˛
.n/i an .z; t/ D @t 2˛ nC1
Z
1
1
sinh.˛s/ ds: Œajzj2 cosh.as/ i t sinh.as/n
Denote 1
D .a2 jzj4 C t 2 / 4 and e i D 2 .ajzj2 i t/ with 2 2 ; 2 . Using the identity cosh.s C i / D cosh.s/ cos C i sinh.s/ sin ; we can write @‰˛
.n/i an .z; t/ D @t 2˛ nC1 2n
Z
1 1
sinh.˛s/ ds: Œcosh.as i /n
Next, we apply the integral formula: Z
1 1
e ˛s 2 1 ˛ b 1 ˛ 1 ˛ a
: dx D e C cosh .as C b/ a ./ 2 2a 2 2a
Hence, Z
1
1
e ˛s e ˛s ds n Œcosh.as i/n 1 Œcosh.as i/ ˛ ˛
n2 C 2a 2n1 n2 2a ˛ ˛ e i a e i a D 2a .n/ ˛ ˛ h ˛ i
n2 C 2a 2n1 n2 2a i sin : D a .n/ a
sinh.˛s/ 1 ds D Œcosh.as i /n 2
Z
1
This yields ˛ ˛ ˛
n2 C 2a 2n1 n2 2a
.n/an @‰˛ .z; t/ D sin @t 2˛ nC1 2n a .n/ a ˛ ˛ ˛
n2 C 2a 2n2 an1 n2 2a sin : D ˛ nC1 2n a Recall that 4 D .a2 jzj4 C t 2 /1=4 and 2 e i D ajzj2 C it. We can write ‰˛ .z; t/ explicitly in terms of z and t by integrating with respect to t, requiring
316
12 Finding Heat Kernels Using the Laguerre Calculus
limt !1 ‰˛ .z; t/ D 0. In the case a D ˛, we can find a more explicit formula for the fundamental solution, noting that for a D ˛, ‰˛ .z; t/ D D D
D
D
Z in1 1
.n 1/ 1 h a ds 2 nC1 1 sinh.as/ Œajzj2 coth.as/ i tn1 Z 1
.n 1/an1 1 ds nC1 2 n1 2 1 Œajzj cosh.as/ i t sinh.as/ Z
.n 1/an1 1 1 ds nC1 2n2 n1 2 1 Œcosh.as i / i2 h n2 /
. n1 2
.n 1/an1 2 2 nC1 2n2 a .n 1/ i2 h / an2 2n3 . n1 2 : n1 2 nC1 a2 jzj4 C t 2
12.6.3 Fundamental Solution: The Case n D 1 In this case e˛ .z; / D ‰
1 ˛jj
Z
1
sinh.˛s/ 0
˚ a exp jjajzj2 coth.as/ ds: sinh.as/ (12.6.61)
We note that if we take the inverse Fourier transform directly, the integral will diverge, so the fundamental solution is not a regular function. We will provide a formula for the fundamental solution in the form of a generalized function. Recall the Fourier transform with respect to t and its inverse: e.z; / D f
Z
e i t f .z; /
and f .z; t/ D
R
1 2
Z
e.z; /d: ei t f R
We need to compute the inverse Fourier transform of e.z; / D 1 e .z;s/jj : f jj The following computation will yield the formula for the inverse Fourier transform. We apply the following formula for the inverse Fourier transform (see (33c), p. 153, of Kanwal [76]): F 1
1 jj
D
" ln jtj
and F 1 .e jj / D
; C t 2/
. 2
12.6 Fundamental Solution of the Paneitz Operator
317
where " is the Euler’s constant: Z
1
"D 0
1 cos y dy y
Z
1
1
cos y dy: y
e.z; / is the convolution Then, the inverse Fourier transform of f h
i ln jtj i h " D 2 . 2 C t 2 /
Z
1
1
" C ln juj du: 2 C .t u/2
We start by computing the integral: Z
1
1
We have Z 1 1
" du D 2 C .t u/2
ln juj du D 2 C .t u/2
Z 0
1
Z
1 1
2
" " du D : 2 Cu
ln u du C .u t/2 C 2
Z
1
ln u du: .u C t/2 C 2
0
Applying the residue theorem (see formula (7.2.11) of Antimirov, Kolyshkin and Vaillancourt [4]), yields Z
1 0
80 9 1 = < 2 1 .ln z/ ln u d u D Re @ „ƒ‚… Res A Res C „ƒ‚… .u t/2 C 2 2 : .z t/2 C 2 ; 1 D Re 2 D
(
zDt Ci
zDt i
ln2 .t i / ln2 .t C i / C 2i 2i
)
1 Refi Œln2 .t C i / ln2 .t i /g: 4
Since 0 arg.z/ < 2, then for t > 0, ln.t C i / D ln
p
t 2 C 2 C i
and
ln.t i / D ln
p
t 2 C 2 C i.2 /;
where D arctan. =t/ 2 Œ0; =2. Hence Z 1 ln u du D ln.t 2 C 2 /: 2 2 .u t/ C 2 0 Similarly, for t < 0, p ln.t Ci / D ln t 2 C 2 Ci. / and
ln.t i / D ln
p t 2 C 2 Ci. C/;
318
12 Finding Heat Kernels Using the Laguerre Calculus
where D arctan. =jtj/ 2 Œ0; =2. Then Z 1 ln u ln.t 2 C 2 /: du D 2 2 .u t/ C 2 0 Similarly, we have Z
1 0
Hence Z
1 0
8 < ln.t 2 C 2 /; t < 0; ln u 2 du D : ln.t 2 C 2 /; .u C t/2 C 2 t > 0: 4
ln u du C .u t/2 C 2
Z
1
0
ln u 2 du D ln.t 2 C 2 /; 2 2 .u C t/ C 4
where D arctan. =jtj/. Summarizing the calculations, we have " 1 ln jtj 2 2 C / : D ln.t " C . 2 C t 2 / 2 2 4 Hence the fundamental solution is Z 1 Z 1 sinh.˛s/ " C ln juj a duds ‰˛ .z; t/ D 3 .z; s/ 2 2 ˛ 0 sinh.as/ 1 .z; s/ C .t u/ Z Z a2 jzj2 1 1 sinh.˛s/ cosh.as/ " C ln juj duds D 2 3 2 4 ˛ sinh .as/ a jzj coth2 .as/ C .t u/2 0 1 Z Z a2 jzj2 1 1 sinh.˛s/ cosh.as/." C ln juj/ D duds 2 2 3 2 4 2 ˛ 0 1 a jzj cosh .as/ C .t u/ sinh .as/ Z 1 i sinh.˛s/ h a " C lnŒa2 coth2 .as/jzj4 C t 2 ds: D 3 ˛ 0 sinh.as/ 2 4
12.7 Heat Kernel of the Paneitz Operator In this section we will compute the heat kernel hs .z; t/ D expfsP˛ gı0 . We first take the Fourier transform with respect to the t-variable and write the heat kernel e hs .z; t/ as a twisted convolution operator: 2 3 1 n X Y p e e .0/ e˛ ge e˛ g 4 5 hs .z; / D expfs P ID expfs P aj W k . aj zj ; / jkjD0
D
1 X jkjD0
e
s 2 Œ
Pn
j D1
aj .2kj C1/ 2 Cs˛ 2 2
j D1 n Y j D1
j
p e .0/ aj W kj . aj zj ; /:
12.7 Heat Kernel of the Paneitz Operator
319
Then we apply the Fourier integral formula, e for D
Pn
j D1 .2kj
e s
2Œ
s 2
Dp
Z
1 4s
1
x2
e 4s ix dx;
1
C 1/aj and s 2 for s:
Pn
j D1 .2kj C1/aj
2
Dp
Z
1 4s 2
1
e
x2 4s 2
ix
Pn
j D1 .2kj C1/aj
dx:
1
We can write e hs .z; / D
1 X
e
s˛ 2 2
jkjD0
n Y j D1
p
"Z
1
1
e
4s 2
P x2 ix n j D1 .2kj C1/aj 4s 2
# dx
1
p e .0/ aj W kj . aj zj ; /
2 2
e s˛ D p 4s 2 2 2
e s˛ D p 4s 2
Z
1
e
x2 4s 2
1
Z
1
e
x2 4s 2
1
#
" "
1 Y n X
aj e
i.2kj C1/aj x
jkjD0 j D1 n Y j D1
aj
e .0/ .paj zj ; / W kj
# dx
1 jj aj jjjzj j2 i aj x X 2i kj aj x e e kj D0
.0/
Lkj .2aj jjjzj j2 / dx: We cannot sum up the series of Laguerre polynomials since je2i aj x j D 1. We can insert a convergence factor e kj and let " gs; .z; ; x/D
n Y j D1
# 1 jj aj jjjzj j2 i aj x X . C2i aj x/kj .0/ 2 aj e Lk .2aj jjjzj j / : e j kj D0
Then, applying the generating formula, we obtain gs; .z; ; x/ D
n Y
jj aj jjjzj j2 i aj x 1 e . C2i aj x/ 1 e j D1 ( ) e . C2i aj x/ 2 exp 2aj jjjzj j 1 e . C2i aj x/ aj
320
12 Finding Heat Kernels Using the Laguerre Calculus
( ) n jjn Y 1 C e . C2i aj x/ aj e i aj x 2 D n exp a jjjzj j . C2i aj x/ . C2i aj x/ j 1 e 1 e j D1 jjn e n =2 D 2n n
"
n Y j D1
( exp
jj
aj sinh. 2 C i aj x/
n X
2
aj jzj j coth
#
j D1
2
C i aj x
) :
Taking the limit ! 0C , we have jjn lim gs; .z; ; x/ D n n 2 !0C
"
n Y j D1
) # ( n X aj aj jzj j2 coth.i aj x/ : exp jj sinh.i aj x/ j D1
Since sinh.ix/ D i sin x and coth.ix/ D i cot x, we have 2 2
e s˛ jjn1 e hs .z; / D p 4s 2n n ( exp i jj
Z
1
e
x2 4s 2
1
n X
"
n Y j D1
aj i sin.aj x/
#
) aj jzj j2 cot.aj x/ dx:
j D1
Next, we take the inverse Fourier transform with respect to and obtain the heat kernel associated with the Paneitz operator: .i /n hs .z; t/ D p .2/nC1 s
Z
1
n1 e
0
˛ 2 s 2
Z
1
cos.t/ 1
2 4
n Y
j D1
3 aj 5 sin.aj x/
9 =
8 <
n X x2 C i aj jzj j2 cot.aj x/ dxd: exp ; : 4s 2 j D1
It seems impossible to find either of the integrals explicitly. In the special case of aj D a for all j D 1; 2; : : : ; n, we obtain Z 1 Z 1h a in .i /n n1 ˛2 s 2 hs .z; t/ D p e cos.t/ .2/nC1 s 0 1 sin.ax/
x2 exp C iajzj2 cot.ax/ dxd: 4s 2
12.8 Projection and Relative Fundamental Solution
321
The integration with respect to is the following: Z
1
n1 e
˛ 2 s 2
x2 4s 2
Ci ajzj2 cot.ax/
cos.t/d:
0
The above integral can eventually be written in terms of some special functions.
12.8 Projection and Relative Fundamental Solution In [61], Greiner and Stein have shown that
1 Nj Ce e .˙p/ .paj zj ; / D .2k C 1/ C p.1 ˙ sgn/ jjaj W e .˙p/ Z ZN j e .e Zj e Zj /W k k 2 p . aj zj ; / (12.8.62) for k; p D 0; 1; 2; : : :. This implies
n n Y 1Xe e e .˙pj / .paj zj ; / .Zj ZN j C e ZN j e Zj / W kj 2 j D1 j D1 ( n ) n X
Y .˙p / p j e D .2kj C 1/ C pj .1 ˙ sgn/ jjaj W . aj zj ; /; kj j D1
j D1
and the Paneitz operator P˛ D L˛ LN˛ satisfies 0 e˛ @ P
n Y j D1
1 e .˙pj / .paj zj ; /A W kj
80 9 12 ˆ > n n < X =Y
e .˙pj / .paj zj ; /: .2kj C1/Cpj .1 ˙ sgn/ aj A ˛2 D 2 @ W kj ˆ > : j D1 ; j D1 e˛ is the set Hence the spectrum of the operator P e˛ / D .P
8 ˆ < ˆ :
20
k;p
3 12 n X
6 7 D 2 4@ .2kj C 1/ C .jpj j C pj sgn/ aj A ˛ 2 5 ;
k 2 ZnC
j D1
9 > =
and p 2 Zn : > ;
322
12 Finding Heat Kernels Using the Laguerre Calculus
The operator P˛ is not invertible if ˛ 2 ƒ with ƒD
8 < :
˙
n X
.2kj C 1/ C .jpj j C pj sgn/ aj ; k 2 ZnC
j D1
This implies that if ˛k;p D ˙ some k 2 ZnC and p 2 Zn , then 0 e˛k;p @ P
n Y j D1
e˛ Let P
.C/
e˛ for > 0 and P e˛ DP n X
j D1
and p 2 Zn : ;
.2kj C 1/ C .jpj j C pj sgn/ aj for 1
e .pj / .paj zj ; /A W kj
./
˛k;p D ˙
Pn
9 =
D 0:
e˛ for < 0. Then for > 0, DP
.2kj C 1/ C .jpj j C pj / aj
j D1
8 P <˙ n 2.kj C pj / C 1aj ; j D1 D :˙ Pn .2k C 1/a ; j j j D1
pj 0; pj 0:
And for < 0, ˛k;p D ˙
n X
.2kj C 1/ C .jpj j pj / aj
j D1
8 P <˙ n .2kj C 1/aj ; pj 0; j D1 D :˙ Pn Œ2.k p / C 1a ; p 0: j j j j j D1 We summarize this result in the following proposition. e˛ u D 0 has the following linearly Proposition 12.8.1. (i) If > 0, then P 2 independent set of L -solutions: 8 n
e .pj / .paj zj ; /; k 2 ZnC and p 2 Zn satisfy W k
j D1
˙
j
n X
j D1
9 = .2kj C 1/ C .jpj j C pj / aj D ˛ : ;
12.8 Projection and Relative Fundamental Solution
323
e˛ u D 0 has the following linearly independent set of L2 (ii) If < 0, then P solutions: 8 n
˙
n X
j D1
9 =
.2kj C 1/ C .jpj j pj / aj D ˛ : ;
e ˛ u D 0 is quite complicated and Given ˛, the set that indexes L2 -solutions of P is given by 8 9 n < X =
†˛ D ˙ .2kj C 1/ C .jpj j C pj sgn.// aj D ˛; k 2 ZnC and p 2 Zn : : ; j D1
In the case when aj D 1 for all j , the set †˛ is empty if n is odd and ˛ is not an odd integer or n is even and ˛ is not an even integer. In general, this set cannot be classified by some simple rules. We now consider the case of n D 1 and a1 D 1. Then †˛ is empty if ˛ is not an odd integer. So we let ˛ D 2m C 1. Then for > 0, ( 2m C 1 D ˙Œ.2k C 1/ C .jpj C p/ D
˙Œ2.k C p/ C 1; p 0; ˙.2k C 1/;
p < 0:
Similarly, for < 0, ( 2m C 1 D ˙Œ.2k C 1/ C .jpj p/ D
˙Œ2.k p/ C 1; p < 0; ˙.2k C 1/;
p 0:
Let m 2 N. Then we have (i) If > 0 and ˛ D ˙.2m C 1/, then the set of linearly independent L2 -solutions e.C/ u D 0 is of P ˙.2mC1/ n
o n o e .p/ e .mk/ .z; /; k D 0; 1; 2; : : : ; m : W .z; /; p D 0; 1; 2; : : : [ W m k
(ii) If < 0 and ˛ D ˙.2m C 1/, then the set of linearly independent L2 -solutions e./ of P u D 0 is ˙.2mC1/ n
o n o e .km/ .z; /; k D 0; 1; 2; : : : ; m : e .p/ W m .z; /; p D 0; 1; 2; : : : [ W k
324
12 Finding Heat Kernels Using the Laguerre Calculus
e˛ / D 2 e ˛ given by M.P The Laguerre matrix of P diagonal. If ˛ D ˙.2m C 1/, we let
.2j 1/2 ˛ 2 ıjk is
1 1 X e .0/ .z; /: f W M 2mC1 D 2 .2j C 1/2 .2m C 1/2 k k6Dm
Then Laguerre calculus will yield e 2mC1 f M 2mC1 C e P S m D I; e m .z; /. We now sum up f where e Sm D W M 2mC1 : 1 1 X e .0/ .z; / f W M 2mC1 D 2 .2k C 1/2 .2m C 1/2 k k6Dm
D
L.0/ .2jjjzj2 / 1 2jj jjjzj2 X k e 2 4.k m/.k C m C 1/
D
X 1 1 1 e .0/ Lk .2jjjzj2 /: 2jj 2m C 1 km kCmC1
k6Dm
jjjzj2
k6Dm
We first consider the case of m D 0. We need to sum up the series F .!/ D
1 .0/ Lk .2jjjzj2 /: k kC1
1 X 1 kD1
Let ! D 2jjjzj2 . Then F .!/ D
kD1
D
1 .0/ .0/ X Lk .!/ Lk .!/ 1 .!/ D L.0/ k k kC1 k kC1
1 X 1
kD1
.0/ 1 X LkC1 .!/ kD0
kC1
D L.0/ 1 .!/ C
1 X kD1
.0/ 1 X Lk .!/ kD1
kC1
1 .!/: ŒL.0/ .!/ L.0/ kC1 kC1 k
We apply the formula .pC1/
!Lk
.p/
.p/
.!/ D .k C p C 1/Lk .!/ .k C 1/LkC1 .!/
12.8 Projection and Relative Fundamental Solution
325
with p D 0 to obtain F .!/ D
.0/ L1 .!/
1 X L.0/ .!/ k ! : .k C 1/2 kD1
Applying the formula 1 D .k C 1/2
Z
1
e .kC1/s sds
o
and the generating formula of the Laguerre polynomials yields F .!/ D
L.0/ 1 .!/
!
1 X
L.0/ k
Z
D
Z
"
1
!
se
e s.kC1/ sds
0
kD1 .0/ L1 .!/
1
1 X
s
0
# .0/ Lk .!/e sk
ds
kD1
e s 1 .1/ se s exp ! L .!/ ds 0 .1 e s /2 1 e s 0
Z 1 Z 1 e s se s .0/ .1/ se s ds! exp ! ds: D L1 .!/C!L0 .!/ .1e s /2 1e s 0 0
D L.0/ 1 .!/ !
Note that
Z
1
Z
1
se s ds D 1
0
.1/ and L.0/ 1 .!/ C !L0 .!/ D 1:
We obtain Z
1
F .!/ D 1 ! 0
se s e s exp ! ds: .1 e s /2 1 e s
This yields
Z 1 e !=2 1 C e s ! e !=2 ! se s f exp ds F .!/ D M 1 .z; / D 2jj 2jj 2jj 0 .1e s /2 1 e s 2 Z 2 s e jjjzj jzj2 1 2 D e jjjzj coth s ds: 2jj 0 .sinh s/2 Taking the inverse Fourier transform with respect to yields M1 .z; t/ D
Z 1 1 2 M 1 .z; /d D ei t f jj1 e i t jzj jj d 2 .2/ 1 1 Z Z 1 jzj2 1 s 2 2 e i t jjjzj coth s d ds: 2 0 .sinh s/2 1
1 2
Z
1
326
12 Finding Heat Kernels Using the Laguerre Calculus
The second part can be integrated with respect to : Z
1
2
e itjjjzj
coth s
d D
1
1 1 2jzj2 coth s : C D jzj2 coth s it jzj2 coth s C it jzj4 coth2 s C t 2
Hence, we have Z
1 M1 .z; t/ D .2/2
1
jj
1 i t jzj2 jj
e
1
jzj4 d 2
Z
1 0
jzj4
s coth s ds: cosh2 s C t 2 sinh2 s
We can compute the first integral using the convolution. It is similar to the integral we computed when we derived the fundamental solution. 1 .2/2
Z
1
1 i t jzj2 jj
jj 1
e
1 " ln jtj jzj2 d D 2 .jzj4 C t 2 / 1 2 4 D 3 " C ln.t C jzj / ; 2 2 4
where D arctan.jzj2 =jtj/. Summarizing the computation, we have the relative fundamental solution: 1 arctan.jzj2 =jtj/ 2 4 M1 .z; t/ D 3 " C ln.t C jzj / 2 2 4 Z s coth s jzj4 1 ds: 2 4 0 jzj cosh2 s C t 2 sinh2 s
(12.8.63) (12.8.64)
The integral converges for z 6D 0.
12.9 The Kernel ‰m .z; t/ for m > n In the case of m > n, we do not have a simple formula like (12.4.41). We shall calculate the following integral in the sense of distributions: Z
1
jjnm e itC˛sgn./sjj.sIz/ d
1
D e ˛s
Z
1 0
nm e it.sIz/ d C e ˛s
Z
1 0
nm e it.sIz/ d: (12.9.65)
12.9 The Kernel ‰m (z, t) for m > n
327
We apply the following formula (see p. 177, equation (14) of Gelfand and Shilov [50]): Z
1 0
.n/ a1
.n/ n1 n e i d D a0.n/ n1 a1 log. C i 0/;
i n1 D .n 1/Š
and
a0.n/
with
(12.9.66)
1 1 i n1 0 1C CC : C .1/ C i D .n 1/Š 2 n1 2 (12.9.67)
This yields Z
1
nm e i t .sIz/ d D .i .sI z/ C t/mn1
0 .mn/
Œa0
Z
1
.mn/
a1
log.i .sI z/ C t/; (12.9.68)
nm e i t .sIz/d D .i .sI z/ t/mn1
0 .n/
.mn/
Œa0 a1
log.i .sI z/ t/:
(12.9.69)
Taking the inverse Fourier transform of (12.4.43) and applying (12.9.65), (12.9.68) and (12.9.69), we obtain ‰m .z; t/ D
1 2 nC1 .m/
Z
1
s m1 e ˛s
0 mn1
.i .sI z/ C t/ C
1 2 nC1 .m/
Z
1 0
h
n Y j D1
.mn/ a0
s m1 e ˛s h
aj sinh.aj s/ .mn/
a1 n Y j D1
.mn/
.i .sI z/ t/mn1 a0
i log.i .sI z/ C t/ ds
aj sinh.aj s/
.mn/
a1
i log.i .sI z/ t/ ds;
.mn/ are given by (12.9.67). We can also rewrite the above where a0.mn/ and a1 equation in terms of the complex distance and volume element (12.4.42), but here we shall omit the details of algebraic manipulation:
‰m .z; t/ D
.mn/ 2m1 i mn a1 .2/n .m/
Z
0
s m1 e 2˛s .s/Œg.sI z; t/mn1 1 Z 1 2m i mn1 log.i g.sI z; t//ds C s m1 e 2˛s .s/ .2/nC1 .m/ 1 h i .mn/ .mn/ Œg.sI z; t/mn1 a0 a1 log.i g.sI z; t// ds: (12.9.70)
328
12 Finding Heat Kernels Using the Laguerre Calculus
12.10 The Isotropic Heisenberg Group The special case of ˛ D 0 and aj D 1=4 for all j D 1; 2; : : : ; n has been studied by Doley, Benson and Ratcliff [40], though they defined the Zj and ZN j differently. We shall summarize their result here and show that it coincides with our result in the special case. Let Z0j D 2
@ @ 1 i zNj @zj 2 @t
@ @ 1 C i zj ; and ZN 0j D 2 @Nzj 2 @t
j D 1; 2; : : : ; n:
Note that Z0j D 2Zj and ZN 0j D 2ZN j with aj D 1=4. They defined the Heisenberg sub-Laplacian to be Hn D
n 1 X N0 0 .Zj Zj C Z0j ZN 0j /: 2
(12.10.71)
j D1
It is easy to see that Hn D 4L0 . Let
jzj2 i t D re i ; 4
where r D
jzj4 C t2 16
1=2 and
< < : 2 2
Then they introduced the functions Gs ./ D e i.nmC1/
Z
s 0
1 sm
Z
s3 0
1 s2
Z
s2 0
.1
s1n1 2 m1 2 s1 / .s1 C
ds1 dsm
e 2i /nmC1
k
(12.10.72)
and m ./
D lim Re.Gs .//: s!1
Finally, they showed that ‰m .z; t/ D
.1/m .n m/Š 2n nC1 r nmC1
m ./
(12.10.73)
is a tempered fundamental solution for m Hn with singular support f.0; 0/g. Remark 12.10.1. In [40], the Haar measure of Hn is defined to be .2/n1 times the Euclidean measure on R2nC1 . Formula (12.10.73) is different from the one given in Theorem A of [40] since we normalize the Haar measure to be the Euclidean measure. The integration in the definition of Gs ./ should be ds1 dsm , not ds1 dsn . Note that our m is equal to their p.
12.10 The Isotropic Heisenberg Group
329
We will show that if we set aj D a D 14 and ˛ D 0 in (12.4.40), we can obtain the result of [40]. The relation m Hn D 4L0 implies that we must prove m Lm 0 .z; t/ D .4/ ‰m .z; t/:
(12.10.74)
We first rewrite formula (12.10.73). The crucial observation is that lim Gs ./ D
s!1
.1/m1 e i.nmC1/ .m 1/Š
Z
1
0
log s1 1 s12
m1
.s12
s1n1 ds1 : C e 2i /nmC1 (12.10.75)
Formula (12.10.75) can be proved via integration by parts. Equations (12.10.73) and (12.10.75) yield n .n m/Š Re e i.nmC1/ 2n nC1 .m/r nmC1
Z 1 .log u/m1 un1 du : 2 m1 .u2 C e 2i /nmC1 0 .1 u /
‰m .z; t/ D
(12.10.76)
In the case of aj D a for all j 1, then (12.4.40) becomes Lm 0 .z; t/
n a s m1 ds Œajzj2 coth.as/ i tnmC1 1 sinh.as/ m1 Z anmC1 .n m/Š 1 as D 2 nC1 .m/ 1 sinh.as/
.n m/Š D 2 nC1 .m/
Z
1
Œajzj2 cosh.as/ i t sinh.as/mn1 ds: To evaluate Lm 0 .z; t/ when jzj ¤ 0, we set re i D ajzj2 i t
1=2 with r D a2 jzj4 C t 2 and jj < : 2
Then the identity cosh.as i/ D cosh.as/ cos i sinh.as/ sin yields ajzj2 Œcosh.as/ i t sinh.as/mn1 D r mn1 Œcosh.as i/mn1 : This implies Lm 0 .z; t/
anmC1 .n m/Š D 2 nC1 .m/r nmC1
Z
1 1
as sinh.as/
m1
Œcosh.as i/mn1 ds: (12.10.77)
330
12 Finding Heat Kernels Using the Laguerre Calculus
If we let Z
1
F ./ D 0
as sinh.as/
m1
Œcosh.as C i/mn1 ds;
can be written as then Lm 0 Lm 0 .z; t/ D
anmC1 .n m/Š Re.F .//: nC1 .m/r nmC1
(12.10.78)
We now use the definitions of sinh and cosh to rewrite F ./ as F ./ D 2n e i.nmC1/
Z
1 0
m1 as e nas A ds: 1 e 2as Œe 2as C e 2i nmC1 (12.10.79)
Introducing the new variable u D e as , we reduce (12.10.79) to the form 2n .1/m1 e i.nmC1/ F ./ D a
Z
1 0
.log u/m1 un1 du: .1 u2 /m1 .u2 C e 2i /nmC1 (12.10.80)
Combining (12.10.78) and (12.10.80), we obtain Lm 0 .z; t/
2n .1/m1 anm .n m/Š D Re e i.nmC1/ nC1 .m/r nmC1
Z 1 .log u/m1 un1 du : (12.10.81) 2 m1 .u2 C e 2i /nmC1 0 .1 u /
By comparison with (12.10.76) and (12.10.81), the relation (12.10.74) holds by setting a D 1=4 in (12.10.81).
12.11 Conclusions Laguerre calculus is a powerful tool for harmonic analysis on the Heisenberg group. Many sub-elliptic partial differential operators can be inverted by Laguerre calculus. In this chapter we used the Laguerre calculus to find explicit kernels of the fundamental solution for the Paneitz operator and its heat equation. The Paneitz operator, which plays an important role in CR-geometry, can be written as " n #2 1 X N N N P˛ D L˛ L˛ D .Zj Zj C Zj Zj / C ˛ 2 T2 : 4 j D1
12.11 Conclusions
331
Here fZj gnjD1 is an orthonormal basis for the sub-bundle T .1;0/ of the complex tangent bundle TC .Hn / and T is the “missing direction.” The operator L˛ is the sub-Laplacian on the Heisenberg group, which is sub-elliptic if ˛ does not belong to an exceptional set ƒ˛ . We also constructed projection operators and a relative fundamental solution for the operator L˛ for ˛ 2 ƒ˛ .
Part III
Laguerre Calculus and the Fourier Method
Chapter 13
Constructing Heat Kernels for Degenerate Elliptic Operators
13.1 Introduction In this chapter we describe a method that was first studied by Beals [12] and Aar˜ao [1, 2] to construct heat kernels for a large class of operators that may or may not be group invariant, including certain degenerate elliptic and kinetic operators. Once again, we shall start with the Heisenberg group. Consider first the n-dimensional Hermite operator n n X X @2 2 xk2 : (13.1.1) 2 jxj2 D @xk2 kD1
kD1
Inspired by the Gaussian, we assume that the heat kernel for (13.1.1) has the form Pt .x; y/ D .t/e Qt .x;y/ ;
t > 0;
(13.1.2)
where Qt .x; y/ is a quadratic form in 2n variables. Taking symmetries into account, we expect Qt .x; y/ D
1 1 ˛.t/hx; xi C ˇ.t/hx; yi C ˛.t/hy; yi; 2 2
where h; i denotes the inner product. We apply the operator
@ C 2 jxj2 @t
to (13.1.2), multiply by e Qt .x;y/ , and set the coefficients of hx; xi, hx; yi, hy; yi and also the constant term equal to zero. The conditions on the coefficients are 1 ˛P D ˛2 2 ; 2 and
ˇP D 2˛ˇ;
1 ˛P D ˇ 2 2
1 P D n˛:
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 13, c Springer Science+Business Media, LLC 2011
(13.1.3)
(13.1.4)
333
334
13 Constructing Heat Kernels for Degenerate Elliptic Operators
Since P must act like an approximation to the identity as t ! 0, we need ˛ to blow up, and ˇ ˛, as t ! 0. The equation ˛P D 2˛ 2 22 is a Riccati equation which can be solved by means of the usual linearization ˛ D u1 C . Then the solution that blows up at D 0 is ˛.t/ D
cosh.2t/ : sinh.2t/
Then ˇP D 2˛ˇ, with boundary condition ˇ ˛ as t ! 0, and 1 P D n˛, normalized so that P is an approximation to the identity, has the solutions ˇ.t/ D
n
sinh.2t/
and
.t/ D
2 n
n
.2/ 2 sinh 2 .2t/
:
The remaining equation 12 ˛P D ˇ 2 is a consequence. The sub-Laplacian associated with the .2n C 1/-dimensional Heisenberg group has the form ! 2n n X X @ @ @2 @2 2 2 C a x C 2 Bx; ; (13.1.5) LD k @x @x0 @xk2 @x02 kD1
kD1
where B is a skew-symmetric 2n 2n matrix. We take the partial Fourier transform with respect to the x0 -variable to obtain in R2n the operator @ e ; L D a2 2 jxj2 C 2i Bx; @x P where x D .x1 ; : : : ; x2n / and, as usual, jxj2 D nkD1 xk2 . Because of the group structure, we may use the same method which has been discussed above to compute the kernel at the origin y D 0. The skew symmetry of B reduces the associated equation to the same form as (13.1.3) and (13.1.4). We take the inverse partial Fourier transform with respect to the -variable and rescale the variable of integration to obtain the heat kernel for the Heisenberg sub-Laplacian: P .x0 ; xI t/ D
1 .2 t/nC1
Z e R
.2ix0 a coth.2a /jxj2 / 2t
2a sinh.2a/
n d:
(13.1.6)
It can be shown that despite the use of the partial Fourier transform and its inverse, formula (13.1.6) is an integral that is absolutely convergent when jxj ¤ 0. We also know that the absolute convergence when jxj D 0, x1 ¤ 0 can be obtained by a change in the contour integration. This is a very powerful method which can be used to handle a class of operators that arise in fluid dynamics and statistical physics (see Aar˜ao [1]) either directly as in the case of (13.1.1) or via a partial Fourier transform as in the case of (13.1.5).
13.2 Finding Heat Kernels for Operators Lj , j D 1; : : : ; 5
335
This class of operators includes operators associated to general step-2 two nilpotent Lie groups and Grushin operators like @2 @2 C x12 2 ; 2 @x1 @x2
(13.1.7)
@ @2 @ bx1 c C ; @x2 @x1 @x12
(13.1.8)
L1 D and Kolmogorov-type operators like L2 D
@2 @ @ x1 C x2 C ; 2 @x2 @x1 @x1 2 @ @ @2 @ L4 D C 2 bx1 bx2 C ; @x1 @x2 @x12 @x2 2 @ @2 @2 @ C bx2 L5 D C C C : 2 2 2 @x3 @x1 @x2 @x3
L3 D
(13.1.9)
(13.1.10)
(13.1.11)
We will use this method to compute heat kernels for this class of operators in the next section.
13.2 Finding Heat Kernels for Operators Lj , j D 1; : : : ; 5 In general, we want to find the heat kernel for the following operator: @ @2 @2 @2 L D a1 2 C a2 2 C a3 2 C b1 x1 C b2 x2 C b3 x3 @x1 @x1 @x2 @x3 @ @ C b4 x1 C b5 x2 C b6 x3 C b7 x1 C b8 x2 C b9 x3 C @x2 @x3 @ @ @ C Bx; C ; (13.2.12) D A ; @x @x @x where
2 3 a1 0 0 A D 4 0 a2 0 5 ; 0 0 a3
2
3 b1 b2 b3 B D 4b4 b5 b6 5 ; b7 b8 b9
336
13 Constructing Heat Kernels for Degenerate Elliptic Operators
with a1 ; a2 ; a3 ; b1 ; : : : ; b9 and constants. Our goal is to obtain a formula for the heat kernel Pt .x; y/ which satisfies @ Pt .x; y/ D LPt .x; y/; @t lim Pt .x; y/ D ı.x y/;
t > 0; (13.2.13)
t !0C
where x D .x1 ; x2 ; x3 / and y D .y1 ; y2 ; y3 /. For convenience, we discuss this system in matrix notation. Since the situations for n D 2 and n D 3 are similar, we give a detailed discussion just for the case n D 2. We first apply conjugation by et to remove the constant term . As we mentioned in Sect. 13.1, we assume the kernel Pt .x; y/ has the following form: 1
1
1
Pt .x; y/ D .t/e Qt .x;y/ D .t/e 2 hMEv;Evi D .t/e 2 h˛.t /x;xihˇx;yi 2 h.t /y;yi ; (13.2.14) where
MD
˛ ˇ ˇT
and ˇ T is the transpose of the matrix ˇ. Here Ev D Œx; y T D Œx1 ; x2 ; y1 ; y2 T is a vector in R22 and ˛, ˇ, are 2 2 matrices, with ˛ 0, 0. We shall calculate the derivatives of Pt .x; y/ first: @Pt P v; Evie 12 hMEv;Evi P 12 hMEv;Evi hME D e @t 2! P P D hMEv; Evi Pt ; 2 and 1 @Pt D @x1 2
@ hMEv; Evi Pt : @x1
d . Since x1 is the first coordinate of Ev, then Here the dot denotes the differentiation dt @Ev E1 . Using the symmetry of M, we have @x1 D e
@ hMEv; Evi D hME e1 ; Evi C hMEv; eE1 i D 2hMEv; eE1 i: @x1 It follows that @Pt D hMEv; eE1 iPt : @x1
13.2 Finding Heat Kernels for Operators Lj , j D 1; : : : ; 5
337
Similarly, one has @Pt D hMEv; eE2 iPt ; @x2 and @2 Pt D hME e1 ; eE1 iPt C hMEv; eE1 i2 Pt @x12 D ˛11 C EvT ME e1 eE1T MEv; eE1 Pt D ˛11 C hME1 MEv; EviPt ; where ˛11 is the .1; 1/-component of the matrix M and 2
1 6 0 E1 D eE1 eE1T D 6 40 0
0 0 0 0
0 0 0 0
3 0 07 7: 05 0
Similarly, we have @2 Pt D ˛22 C hME2 MEv; EviPt ; @x22 where
2
0 6 0 E2 D eE2 eE2T D 6 40 0
0 1 0 0
0 0 0 0
3 0 07 7: 05 0
Applying @t@ C L to (13.2.14) gives P 1 P evi C a1 ˛11 C a2 ˛22 hMe evi D 0; (13.2.15) hMEv; Evi hMEv; BE AMEv; BE 2 e are given by where the matrices e A and B 2
a1 0 6 0 a2 e AD6 40 0 0 0
0 0 0 0
3 0 07 7D A 0 ; 05 0 0 0
2
b1 b 2 6b3 b4 eD6 B 40 0 0 0
0 0 0 0
Since e vi D h B eT MEv; Evi D hMBE ev; Evi; hMEv; BE we see that if we require that P C MB eCB eT M C 2Me AM D 0; M
3 0 07 7D B 0 : 05 0 0 0
338
13 Constructing Heat Kernels for Degenerate Elliptic Operators
and that P C a1 ˛11 C a2 ˛22 D 0; then (13.2.15) will hold for all Ev. More precisely, we have the following system: b b b b a1 0 ˛ ˛ 1 2 1 3 ˛ D 2˛A˛ ˛B C B T ˛ ; 0 a2 b3 b4 b2 b4 b b a 0 ˛ 1 3 ˇ D 2˛Aˇ B T ˇ; ˇP D 2ˇ 1 0 a2 b2 b4 0 a P D 2ˇ 1 ˇ T D 2ˇAˇT ; 0 a2 ˛P D 2˛
1 P D .a1 ˛11 C a2 ˛22 / D trace .A˛/:
(13.2.16)
We need to solve the first equation in (13.2.16). Once this is done, the solutions ˇ and are obtained in a simple manner. Indeed, let be a matrix solving P D b1 b2
D B satisfying the initial condition .0/ D 0 . Set ˇ D ˛ and b 3 b4 D T ˛ C M , with M a constant symmetric matrix. It is easy to see that as long as ˛ satisfies the first equation of (13.2.16), then ˇ and will satisfy the second and third equations of (13.2.16), respectively. So, the problem reduces to finding the matrices 0 and M . Since Pt .0; 0/ ! 1 as t ! 0C , we know that .t/ ! 1 as t ! 0C . Moreover, the second equation of (13.2.13) provides us with the boundary conditions ˛ ! C1 I;
as t ! 0C :
˛ ˇ ;
(13.2.17)
Assume that hMEv; Evi D h˛.x C y/; x C yi C hMy; yi: Again, the second equation of (13.2.13) tells us that Z
1
R2
Pt .x; y/dx D .t/e 2 hMy;yi
Z R2
1
e 2 h˛.xCy/;xCyi dx ! 1
as t ! 0C . Hence we conclude that M D 0. Finally, we want to conclude that
0 D .0/ D I . If not, then for some x we have x C 0 x ¤ 0, and hence h˛.x C x/; x C xi ! 1;
as t ! 0C :
13.2 Finding Heat Kernels for Operators Lj , j D 1; : : : ; 5
339
This implies that Pt .x; x/ converges to zero, when it should be going to infinity. This contradiction leads to the conclusion that 0 D I . As mentioned in [1], this argument can only be fully justified after we find ˛. Furthermore, 0 is nonsingular. If it were, then for some w ¤ 0, one has 0 w D 0, and then b1 b 2
.t/w D exp t
0 w D 0; b3 b4
8t > 0;
and this would imply that .t/ is singular for all t > 0. Now we shall solve the Riccati matrix equation in the system (13.2.16) for ˛: ˛P D 2˛
a1 0 b b b b ˛ ˛ 1 2 1 3 ˛ D 2˛A˛ ˛B C B T ˛ : b3 b4 b2 b4 0 a2 (13.2.18)
Conjugating the (13.2.18) by the matrix , we have
T ˛
P C T B T ˛ C T ˛B D 2 T ˛A˛ : Because P D B , we can rewrite the above equation as d T
˛ D 2 T ˛A˛ : dt Since D T ˛ , the above equation becomes P D 2A0 ; P , and so where A0 D 1 A. T /1 . Denote ! D 1 . Then one has P D ! !P D 2A0 : If we can solve the above equation with initial condition !.0/ D 0 (since ! D 1 ), we find that 1 : ˛ D . T /1 ! 1 1 D ! T In this case, the general self-adjoint solution of (13.2.18) is (see Beals [12]) 1 Z t T T ˛.t/ D e 2tB W C 2 e 2B Ae 2B d e 2tB : 0
(13.2.19)
340
13 Constructing Heat Kernels for Degenerate Elliptic Operators
Since ˇ D ˛ , then one obtains 1 Z t T ˇ.t/ D ˛.t/e 2tB D W C 2 e 2B Ae 2B d e 2tB ;
(13.2.20)
0
where W will be determined later. Moreover, it is immediate from (13.2.16) and (13.2.20) that the solution to the third equation P D 2ˇAˇT in (13.2.16) that satisfies the boundary condition is .t/ D e
2tB
˛.t/e
2tB T
D W C2
Z
t
e
2B
Ae
2B T
1 d
:
(13.2.21)
0
Combining (13.2.19)–(13.2.21), we may write the quadratic form as Qt .x; y/ D
! h i E 1 D .t/ e 2tB x y ; e 2tB x y : 2
Any solution of (13.2.16) fulfills the first requirement in (13.2.13). The delicate part is to choose the constants of integration: the matrix W in (13.2.19) to (13.2.21) so that it satisfies the boundary conditions in (13.2.13). In particular, W must be singular. We expect the heat kernel to provide a two-sided inverse for the heat equation for L. Therefore, P should satisfy analogous conditions in the y-variables for the adjoint operator L of L. These conditions provide another set of equations dual to (13.2.16). The latter equations can be shown to be compatible with (13.2.13). We illustrate these points with a few comments on examples.
13.3 Some Explicit Calculations We are very interested in finding the heat kernel for the operator L defined in (13.2.12): @ @2 @2 @2 L D a1 2 C a2 2 C a3 2 C b1 x1 C b2 x2 C b3 x3 @x1 @x1 @x2 @x3 @ @ C b4 x1 C b5 x2 C b6 x3 C b7 x1 C b8 x2 C b9 x3 C : @x2 @x3 Let us point out that for each of the operators (13.1.8)–(13.1.11), the major difficulty is choosing the correct constants of integration in order to achieve the initial conditions (13.2.17). Let us demonstrate this point by a detailed discussion of operators Lj , j D 1; : : : ; 5.
13.3 Some Explicit Calculations
341
13.3.1 Laplace Operator We start with the Laplace operator L D on R3 . In this case, A D I3 is the 3 3 identity matrix, and B D 0 is the zero matrix. Likewise, D 0. Hence, from (13.2.16), one has ˝ ˛ ˛P D 2 ˛; ˛ : Let D ˛1 . As we have seen before, the initial condition .0/ of ƒ must be zero. Moreover, ˛ D I implies that P D 2I , which provides .t/ D 2tI . Therefore, ˛.t/ D
1 I: 2t
Since P D B D 0, we have .t/ D .0/ D 0 . Since hMEv; Evi D h˛.x C y/; x C yi, we obtain 1
2
Pt .x; y/ D .t/e 4t jxCyj : However, we know that 0 D I from the discussion in Sect. 13.2. This implies that 1 2 Pt .x; y/ D .t/e 4t jxyj . Finally, let us calculate . From (13.2.16), we need to solve 3 P D trace .A˛/ D : 2t 3
Hence .t/ D ct 2 for some constant c. Since Z R3
1
3
2
e 4t jxyj dx D .4 t/ 2
independently of y, one concludes that c D
Pt .x; y/ D
1 . .4/3=2
Therefore,
1 1 2 e 4t jxyj : 3=2 .4 t/
In fact, a similar formula works for any n. In this case, ˛, ˇ and are all n n matrices. The system (13.2.16) would be the same and hence the above analysis n P will be the same. The only difference is that .t/ D 2t . Moreover, the operator is invariant under translations. Thus it suffices to consider y D 0, and the heat 1 2 kernel for the n-dimensional Laplace operator becomes Pt .x/ D .4 t1/n=2 e 4t jxj , which is the well-known Gaussian distribution.
342
13 Constructing Heat Kernels for Degenerate Elliptic Operators
Similarly, we may deal with hypoelliptic second-order operators associated with two-step nilpotent groups which lead to (13.1.5), whose solutions are similar to cosh.2 t/ ; sinh.2 t/ ; ˇ.t/ D sinh.2 t/ ˛.t/ D
n
.t/ D
2
n
Œ2 sinh.2 t/ 2
:
These kernels are similar to the formula n Z .2ix0 a coth.2a /jxj2 / 1 2a 2t e d; Pt .x0 ; xI 0; 0/ D .2 t/nC1 R sinh.2a/ which was obtained by several different methods.
13.3.2 Kolmogorov Operator Now let us turn to the operator L2 D 2
@2 @x12
bx1 @x@2 c @x@1 given in (13.1.8). We
t @ first conjugate it by expf 2cxc g to eliminate the term c @x . We may assume b ¤ 0 4 and rescale in x1 so that b D 1. This equation was first studied by Kolmogorov in [81]. Starting at a point x in the plane, a particle describes a motion .X1 .t/; X2 .t// in the following way. In the first coordinate, X1 .t/, it is just a Brownian motion starting at x1 . In the second coordinate, X2 .t/ obeys the equation dX2 D X1 .t/dt starting at x2 . In particular, if X1 < 0, then X2 decreases. In this situation, after a time t > 0 elapses, the probability that the particle is inside the open set R R2 is given by pt .x; y/dx. In [81], Kolmogorov used probability techniques to compute the fundamental solution for the operator L2 . Here we use the method which was discussed in the previous section to find its heal kernel (and hence the fundamental solution). Now, the operator L2 reduces to (13.2.16) with 0 0 1 0 : ; BD AD 1 0 0 0
We set D ˛ 1 and find P D 2A C B C B T : More precisely, this gives us 2 11 P 11 P 12 D : P 21 P 22 11 212
13.3 Some Explicit Calculations
343
It follows that 11 D 2t C 1 ;
12 D t 2 C 1 t C 2 ;
22 D
2 3 t C 1 t 2 C 22 t C 3 : 3
Here 1 , 2 and 3 are three constants. As before, we set .0/ D 0. This implies that 1 D 2 D 3 D 0. Hence " # " # 2 1 2 3t t32 1 t ˛.t/ D D D : (13.3.22) t32 t63 t 3t t62 Since B 2 D 0, the exponential 2
3 0 5
1
e 2B D I C 2B D 4 1 and thus
2 T 6 .t/ D e 2B ˛.t/e 2B D 4
and
" ˇ.t/ D .t/e
2B
D
1 t 3 t2
2 t
3 t2
3 t2
6 t3
t32 t63
3 7 5;
# :
It follows that 1 1 Qt .x; y/ D h˛.t/x; xi hˇ.t/x; yi h.t/y; yi 2 (*" 2 # + # + *" 2 1 t32 x1 1 t32 x1 x1 y t t D C2 ; ; 1 3 6 3 6 x2 x2 t 3 x2 y2 2 t2 t3 t2 # +) *" 2 3 y1 y t t2 ; 1 C 3 6 y y2 2 t2 t3 (* * + + 1 2t 2 x1 3tx2 x1 t 2 x1 3tx2 y C2 ; ; 1 D 3 3tx C 6x x 3tx 6x y 2t 1 2 2 1 2 2 * +) 2t 2 y1 C 3ty2 y C ; 1 3ty1 C 6y2 y2 1 n D 3 2t 2 x12 5tx1 x2 C 6x22 C 2 t 2 x1 y1 3tx2 y1 C 3tx1 y2 2t o 6x2 y2 C 2t 2 y12 C 6ty1 y2 C 6y2
344
13 Constructing Heat Kernels for Degenerate Elliptic Operators
2 o 1 n 2 2 2t x1 Cx1 y1 Cy12 C6t x1 Cy1 y2 x2 C6 x2 y2 3 2t
2 o 1 1 nt2 .x1 y1 /2 C 6 .x2 y2 / .x1 C y1 /t : D 3 2t 2 2
D
We also know that " 2 1 0 t A˛.t/ D 0 0 t32
t32 6 t3
# D
0 : 0 0
2 t
Hence
2 trace A˛.t/ D : t It follows that 1 P D trace A˛.t/ D 2t , which implies that .t/ D ct 2 . Now we need to calculate the constant c. However, a straightforward computation shows that Z t2 e Qt .x;y/ dx D p : 3 R2 Hence we have proved the following theorem. 2
@ @ Theorem 13.3.1. The heat kernel of the Kolmogorov operator @x 2 x1 @x2 is 1
p 2 1 / 3 .x y / 1 .x Cy /t 2 3 .x1 y Œ 2 2 2 1 1 : 4t t3 Pt .x1 ; x2 ; y1 ; y2 / D 2 e t
13.3.3 The Operator L3 Now we are going to study the operator L3 D Here we have
@ @ @2 C x1 x2 C : 2 @x2 @x1 @x1
1 0 ; AD 0 0
1 BD 0
0 : 1
Once again, set D ˛1 and find 2 212 11 22 P 11 P 12 D : P 21 P 22 11 22 212
13.3 Some Explicit Calculations
345
It follows that R 12 C 412 D 2; and we have 12 D
1 2
C 1 cos.2t/ C 2 sin.2t/. Hence
11 D 3 C t 1 cos.2t/ C 2 sin.2t/; 22 D 4 C t C 1 sin.2t/ 2 cos.2t/: Here j , j D 1; : : : ; 4, are constants. Since P 12 D 11 22 , we conclude that 3 D 4 . The initial condition .0/ D 0 tells us that 11 .0/ D 3 1 D 0 and 22 .0/ D 4 2 D 0. Hence, 1 D 2 D 3 D 4 D c, which provides " # 2t sin.2t/ 1 C cos.2t/ 1 ˛.t/ D : 2.t 2 sin2 t/ 1 C cos.2t/ 2t C sin.2t/ Next, we need to find . Since the matrix B is diagonal, we know that is also diagonal and can be calculated as # " e t 0 :
.t/ D 0 et Finally, the equation P C ˛11 D 0 can be written as 1 2t sin.2t/ P C D 0; 2 t 2 sin2 t which gives .t/ D p
c t2
sin2 t
:
A straightforward computation shows that Z R2
e Qt .x;y/ dx D 2
and so we conclude that c D
1 . 2
p t 2 sin2 t
We have proved the following theorem.
Theorem 13.3.2. The heat kernel of the operator L3 D
@ @ @2 C x1 x2 2 @x2 @x1 @x1
346
13 Constructing Heat Kernels for Degenerate Elliptic Operators
is Pt .x; y/ D
p
exp
1
2.t 2
2 1 .2t sin.2t//2 x1 C e t y1 2 sin t/
2 t 2 sin2 t 2.1 cos.2t//.x1 C e t y1 /.x2 C e t y2 / C .2t C sin.2t//2
.x2 C e t y2 /2 :
13.3.4 The Operator L4 The operator L4 is defined by
@2 @2 C L4 D @x12 @x22 Here we have AD
C bx1
1 0 ; 0 1
@ @ bx2 C : @x1 @x1
BD
b 0 : 0 b
Obviously, both matrices A and B are diagonal. Set D ˛ 1 and get 2 C 2b11 P 11 P 12 D 0 P 21 P 22
0 : 2 2b22
It follows that 11 D
1 2bt e 1 ; b
12 D 0;
Hence
" ˛.t/ D b
1 e2bt 1
0
22 D
0 1 1e2bt
1 1 e 2bt : b
# :
Using the same method as before, we know that .t/ D
b : 4 sinh.bt/
The matrix is diagonal because B is diagonal. In fact, one finds that bt 0 e :
D 0 e 2bt As a consequence, we have the following theorem.
13.3 Some Explicit Calculations
347
Theorem 13.3.3. The heat kernel of the operator
@2 @2 L4 D C @x12 @x22
C bx1
@ @ bx2 C @x1 @x1
is ( " #) b .x1 e bt y1 /2 .x2 e bt y2 /2 b exp C : Pt .x; y/ D 2 2.e bt e bt / e 2bt 1 1 e 2bt This recovers a result of Lingevitch and Bernoff in [88].
13.3.5 The Operator L5 The operator L5 is defined as follows:
@2 @2 @2 L5 D C C @x12 @x22 @x32 Here we have
2
3 1 0 0 A D 40 1 05 ; 0 0 1
2bx2
@ : @x3
2
3 0 0 0 B D 40 0 05 : 0 2b 0
Again set D ˛ 1 . With the initial condition .0/ D 0, we find that 2 3 2 3 P 11 P 12 P 13 2 0 2b12 4P 21 P 22 P 23 5 D 4 0 2 2b22 5 : P 31 P 32 P 33 2b12 2b22 2 4b23 It follows that 11 D 22 D 2t; b 2t 3 33 D 2t C 8 : 3 Thus
12 D 21 D 13 D 32 D 0;
23 D 2bt 2 ;
3 2 2 2 0 0 1 C b 3t 1 7 6 2 2 4 ˛.t/ D 0 1 C 4 b 3t bt 5 : b2 t 2 2t 1 C 3 0 bt 1
348
13 Constructing Heat Kernels for Degenerate Elliptic Operators
Furthermore,
2 3 1 0 0
D 4 0 1 0 5 : 0 2bt 1
One can also show that 1 q .t/ D p 8 3t 3 1 C
b2 t 2 3
:
Hence, we proved the following theorem. Theorem 13.3.4. The heat kernel of the operator L5 D
@2 @2 @2 C 2 C 2 2 @x1 @x2 @x3
2bx2
@ @x3
is 1 Pt .x; y/ D p q 8 3t 3 1 C
( b2 t 2 3
exp
"
b2t 2 1 C .x1 y1 /2 b2 t 2 3 2t 1 C 3 b2t 2 .x2 y2 /2 C .x3 y3 /2 C 1C4 3 C 2bt.x2 y2 /.x3 y3 / #)
1
C 2b 2 t 2 y2 .x2 y2 / C 2bty2 .x3 y3 /
:
Chapter 14
Heat Kernel for the Kohn Laplacian on the Heisenberg Group
14.1 The Kohn Laplacian on the Heisenberg Group We shall deal next with the nonsymmetric form of the Heisenberg group. The Heisenberg group considered in this section will be the set Hn D Rn Rn R with the following group law: .x; y; t/ .x 0 ; y 0 ; t 0 / D .x C x 0 ; y C y 0 ; t C t 0 C x y 0 /; where .x; y; t/; .x 0 ; y 0 ; t 0 / 2 Rn Rn R and x y0 D
n X
xk yk0 :
kD1
The leftinvariant vector fields for this group structure are Xj D
@ @ C yj ; @xj @t
Yj D
and T D
@ ; @yj
1 j n;
@ : @t
(14.1.1)
The group Hn can also be identified with the following hypersurface in CnC1 : Hn D
8 < :
.z1 ; : : : ; zn ; znC1 / 2 CnC1
9 n = 1X W Im.znC1 / D ŒIm.zj /2 : ; 2 j D1
Here we identify z1 ; : : : ; zn ; t C
i 2
Pn
2 j D1 ŒIm.zj /
2 Hn with
.z1 ; : : : ; zn ; t/ D .x1 ; : : : ; xn ; y1 ; : : : ; yn ; t/;
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 14, c Springer Science+Business Media, LLC 2011
349
350
14 Heat Kernel for the Kohn Laplacian on the Heisenberg Group
where zj D xj C iyj 2 C. With this identification, the leftinvariant vector fields of type .1; 0/ and .0; 1/ are, respectively, @ 1 zj zNj @ i Xj iYj D ; 2 @zj 4 @t 1 zj zNj @ @ i ZN j D Xj C iYj D 2 @Nzj 4 @t Zj D
(14.1.2)
for 1 j n. Hence, one has ŒZN j ; Zk D
i ıjk T; 2
j; k D 1; : : : ; n:
We make a further assumption that fZj ; : : : ; Zn ; ZN 1 ; : : : ; ZN n g form an orthonormal basis for the T .1;0/ ˚ T .0;1/ on Hn . Let #q be the set of all increasing q-tuples J D .j1 ; : : : ; jq /, 1 jk n. The Kohn Laplacian b D @N b @N b C @N b @N b acting on .0; q/-forms on Hn Ð Hn can be calculated in terms of the vector fields (14.1.1) and (14.1.2). Suppose gD
X
gJ !N J D
J 2#q
X
gj1 ;:::;jq !N j1 ^ ^ !N jq
J 2#q
is a .0; q/-form. Then it can be shown that (see Chap. 12 in Stein’s book [103]) b g D
X
Ln2q gJ !N J ;
(14.1.3)
J 2#q
where
1 0 n 1 @X 2 .Xj C Yj2 / C i ˛T A : L˛ D 4
(14.1.4)
j D1
The heat operator is defined on .0; q/-forms g on Hn with coefficient functions that depend on the time variable u 2 .0; 1/ and .x; y; t/ 2 Hn . This is given by @g C b g D 0: @u
14.2 Full Fourier Transform on the Group
351
As we can see from (14.1.3), the Kohn Laplacian b acts diagonally; we can look at the heat operator acting on each component of g; i.e., we are looking for the heat kernel Ku that satisfies (
@Ku @u
D L˛ Ku
.x; y; t/ 2 Hn ;
u > 0;
(14.1.5)
limu!0C Ku .x; y; t/ D ı0 .x; y; t/: Here ı0 .x; y; t/ is the Dirac delta function at the origin.
14.2 Full Fourier Transform on the Group We will use the full Fourier transform in the variables .x; y; t/. Let .; ; / be the dual variables of .x; y; t/ and define b.; ; / D f
Z
f .x; y; t/e i.xCyCt / dx dy dt: Hn
Then one may prove the following theorem. Theorem 14.2.1. For any ˛ 2 C, the Fourier transform of the heat kernel satisfying (14.1.5) is given by u
b u .; ; / D K where AD
sinh
e ˛ 4 n
cosh 2 . u / 2
A
u
2 ; cosh u 2
2 Cjj2 /CiB
e 2 .jj
BD
2 sinh2 cosh
;
(14.2.6)
u
u4 :
(14.2.7)
2
Proof. Recall that
1f .; ; / D X 2 j
b
@2 @ 2 2i j C 2 2 @j @j
! b; f
b; Yj2 f .; ; / D j2 f c .; ; / D i f b: Tf
(14.2.8)
Since the heat kernel on the n-dimensional space is the product of n copies of the one-dimensional heat kernel, we may first reduce the problem to the
352
14 Heat Kernel for the Kohn Laplacian on the Heisenberg Group
b
one-dimensional case. Denote by Ku˛;1 the one-dimensional version of (14.2.6); i.e., for 1 j n, e Kb . ; ; / D q ˛;j u
j
j
˛ u 4
A
2
2
e 2 .j Cj /CiBj j ;
cosh. u / 2
where A and B are given by (14.2.7). Then from (14.2.6), one has b u .; ; / D K
n Y
1 ˛ ;j
Kun .j ; j ; /;
(14.2.9)
j D1
where D .1 ; : : : ; n /;
D .1 ; : : : ; n / 2 Rn :
b
Then the problem reduces to showing that Ku˛;j satisfies the one-dimensional transformed heat equation; i.e., for 1 j n,
@ 1 c2 c2 b X C Yj C i ˛ T Ku˛;j .j ; j ; / D 0; @u 4 j
b
(14.2.10)
b
with the initial condition K0˛;j D 1. Then by using formula (14.2.9), we can conclude the result of this theorem. From now on, we are working on the one-dimensional Heisenberg group H1 . We may drop the index j from the kernel. Therefore, x, y, t and , , are all real variables. Define c˛ .; ; /e i : K˛ .; ; I u/ D K (14.2.11) u
Then, it is easy to verify the equations K˛ .; ; I 0/ D e i
(14.2.12)
and 1 2 @2 @K˛ 2 D ˛ K˛ : @u 4 @ 2
(14.2.13)
Equation (14.2.12) follows from the fact that the Fourier transform of the Dirac delta function is the constant 1. Equation (14.2.13) follows from (14.2.10) and (14.2.8). As in the paper of Boggess and Raich [20], we call (14.2.13) the transformed heat equation. It is easy to see that the differential operator (14.2.13) is a Hermite operator. Therefore, we may use the Hermite function to solve the problem.
14.3 Solving the Transformed Heat Equation Using Hermite Functions
353
14.3 Solving the Transformed Heat Equation Using Hermite Functions For k D 0; 1; 2; : : : and x 2 R, denote k .1/k x2 d 2 e x : hk .x/ D p e 2 p k k dx 2 kŠ
The functions hk .x/ satisfy the following relations: d C x hk .x/ L.hk /.x/ D dx # " k x2 d d .1/k x 2 2 e p D Cx p e dx dx k 2k kŠ " kC1 k .1/k x2 d x2 d 2 2 D p xe 2 e x e 2 e x p kC1 k dx dx 2k kŠ # k x2 d x 2 2 Cxe e dx k D D
p p
2.k C 1/ p
.1/kC1
p e 2kC1 .k C 1/Š
x2 2
d kC1 x 2 e dx kC1
2.k C 1/hkC1 .x/
and L .hk /.x/ D
p d C x hk .x/ D 2khk1 .x/: dx
(14.3.14)
(14.3.15)
d d C x and L D dx C x are called the creation and annihiThe operators L D dx lation operators in quantum mechanics. An easy calculation shows that
HD
d2 1 LL C L L D 2 C x 2 ; 2 dx
the Hermite operator. Hence, by (14.3.14) and (14.3.15), one has (see (1.1.28) in Thangavelu [107]) H.hk /.x/ D h00k .x/ C x 2 hk .x/ D .2k C 1/hk .x/:
(14.3.16)
354
14 Heat Kernel for the Kohn Laplacian on the Heisenberg Group
For ˛ 2 R, let Hk .x/
D
1 1
jj 4
hk
! x p : jj
It is known that fhk .x/g and hence fHk .x/g form an orthonormal system in L2 .R/ (see Chap. 1 in Thangavelu [107]). We first assume that > 0. Replacing x by p in (14.3.16) yields 2 2 @ 2 ˛ Hk D .2k C 1 C ˛/Hk ./: @ 2
(14.3.17)
In other words, Hk is an eigenfunction of the differential operator on the right-hand side of (14.2.13) with eigenvalue .2k C 1 C ˛/. Since fHk .x/g is an orthonormal basis for L2 .R/, K˛ .; ; I u/ can be expressed as K˛ .; ; I u/ D
1 X
1
ak .; /e 4 .2kC1C˛/u Hk ./;
(14.3.18)
kD0
where ak .; / are the Fourier–Hermite coefficients which will be determined later. Differentiating (14.3.18) with respect to the time variable u and using (14.3.17), one has 1 X 1 1 @K˛ .; ; I u/ D ak .; /e 4 .2kC1C˛/u .2k C 1 C ˛/ Hk ./ @u 4 kD0 1 2 1X 1 .2kC1C˛/u 2 @ 2 4 ak .; /e ˛ Hk ./ D 4 @ 2 kD0 # "X 1 1 2 @2 1 D 2 ˛ ak .; /e 4 .2kC1C˛/u Hk ./ 4 @ 2 kD0 1 2 @2 2 ˛ K˛ .; ; I u/: D 4 @ 2 Hence, K˛ satisfies the transformed heat equation (14.2.13). If the kernel K˛ satisfies the initial condition (14.2.12), one must have e i
D K˛ .; ; I 0/ D
1 X
ak .; /Hk ./:
kD0
Using the fact that the fHk .x/g is an orthonormal system, we have Z ak .; / D
R
e i
1
Hk ./d D 4
Z e R
i p
hk ./d :
14.3 Solving the Transformed Heat Equation Using Hermite Functions
355
The integral on the right-hand side is just the Fourier transform of the Hermite function hk ./ at the point p . At this point, we need to prove the following auxiliary lemma, which can be found in Thangavelu [107]. For self-containing reasons, we shall repeat the proof here. Lemma 14.3.1. The Hermite functions are eigenfunctions of the Fourier transform: p b hk ./ D .i /k 2hk ./: Proof. From (14.3.14), we know that p d C x hk .x/ D 2k C 2hk .x/: L.hk /.x/ D dx It is known that
b
p
d b .xf /./ D i 2 f ./; d
3 d
dx
p b ./: ./ D i 2 f
Now taking the Fourier transform on both sides of (14.3.14), we have p d i hk ./: C b hk ./ D 4.k C 1/b d If we assume that the lemma is true for hk , then it follows that p p d kC1 b 4.k C 1/hkC1 ./ D .i / 2 C hk ./ d p D .i /kC1 4.k C 1/hkC1 ./: p 1 2 p Therefore, it is enough to show that b h0 ./ D 2h0 ./. But h0 .x/ D e 2 x = 4 , p and hence by a straightforward calculations, we conclude that b h0 D 2h0 . Therefore, by Lemma 14.3.1, we know that p 1 ak .; / D .i /k 2 4 hk
p
:
Substituting this value of ak into the expression for K˛ , we have 1 X p 1 1 .1C˛/u k 4 .i / hk p hk p e 2 ku : K˛ .; ; I u/ D 2e kD0
356
14 Heat Kernel for the Kohn Laplacian on the Heisenberg Group
Now by (14.2.11), we have c˛ .; ; / D K˛ .; ; I u/e i K u D
1 X p 1 2e 4 .1C˛/u .i /k hk
kD0
p
hk
p
1
e 2 ku e i
:
1
Now let ‚ D e 2 u . Since ji ‚j < 1, we obtain "1 # X p 1 i p p c˛ .; ; / D 2‚ 2 .1C˛/ K e .i ‚/k hk p hk p u kD0
p 1 .1C˛/ 2 2 2 2‚ 2 1 1‚ . C / i . 2‚ C1/ D p e 2 1C‚2 e 1C‚2 : 1 C ‚2 Now r q u p p u p u 4u 2u 2 u 4 2 1C‚ D 1Ce De e Ce D 2e cosh : 2 Then we have
Next,
p 1 .1C˛/ u e ˛ 4 2‚ 2 p Dq : 1 C ‚2 cosh u 2 sinh u 1 e u 1 ‚2 2 D D 1 C ‚2 1 C e u cosh u 2
and u u u u 2 sinh2 u 2‚ .1 e 2 /2 e 2 .e 4 e 4 /2 4 : C1D D u u D u 2u 1 C ‚2 1 C e u 2 2 cosh e .e C e / 2 It follows that ˛ u 4
2 2 A c˛ .; ; / D q e 2 .jj Cjj /CiB ; K u u e cosh 2
where AD
sinh
u
cosh
2u ; 2
BD
2 sinh2 cosh
u
u4 : 2
c0 .; ; / D 1. The proof of the theorem for > 0 is therefore complete. Note that K
14.3 Solving the Transformed Heat Equation Using Hermite Functions
357
If D 0, the kernel in (14.2.6) becomes c˛ .; ; 0/ D e 4u . 2 C 2 / ; K u which is easily shown to satisfy
@ 1 c2 c2 c˛ .; ; 0/ D 0: b K C Y C i ˛T X u @u 4
If < 0, then is replaced by jj on the right side of (14.3.17), which will slightly change the previous calculation. However, the formula for the solution given in Theorem 14.2.1 also remains true in this case. If ˛ D 0, then the operator L˛ becomes the sum of the square of vector fields: LD
n 1X 2 .Xj C Yj2 /: 4
(14.3.19)
j D1
In this case, by Theorem 14.2.1, we know that the Fourier transform of the heat kernel of the operator L is 1
b u .; ; / D K cosh
1 2
A
. u / 2
2 Cjj2 /CiB
e 2 .jj
;
(14.3.20)
where A and B are given by (14.2.7): AD
sinh
u
cosh
2u ; 2
BD
2 sinh2 cosh
u
u4 : 2
It follows that the heat kernel for the sub-Laplacian is Z 1 b u .; ; /e i.xCyCt / d d d K .2/3 R3 Z A B 1 1 .x 2 Cy 2 /i xy 2.A2 CB 2 / e 2.A2 CB 2 / d : D p 1 2 R cosh 2 . u / A2 C B 2 2
Ku .; ; / D
By (14.2.7), it is an easy exercise to show that the coefficients of the real and imaginary parts of the exponent of the above integrand are B D 2 2 2.A C B / 2
and
u cosh u A 4 : D D coth B sinh. u / 4 4
358
14 Heat Kernel for the Kohn Laplacian on the Heisenberg Group
Consequently, the heat kernel should be Z
u 2 2 1 4 coth. 4 /.x Cy /i 2 xyCi t d u e sinh. / R 4 Z ˚ 4 2 2u coth.2/.x 2 Cy 2 /C2i xy4i t D e d : .2u/2 R sinh.2/
1 Ku .; ; / D 2 8
14.4 Conclusions Even if the method described in this chapter is not universally applicable, however, it is widely applicable. Not every operator of the form (13.2.12) has a well-behaved heat kernel. So far, we have not imposed any positivity conditions. For operators of the form (13.2.13) one has an additional problem of tracking behavior with respect to the Fourier transform variable and translating it into the behavior of the heat kernel itself. It may or may not be possible to express the integration with respect to the time variable t for obtaining the fundamental solution in closed form, and reading off the properties of the fundamental solution may pose new difficulties.
Part IV
Pseudo-Differential Operators
Chapter 15
The Pseudo-Differential Operator Technique
The pseudo-differential operator theory emerged from the theory of singular integrals and Fourier analysis, having Kohn and Nirenberg as initiators. The theory was later extended and developed by H¨ormander and has become an important tool in the theory of modern PDEs. In this chapter we shall study the construction of the fundamental solution for heat operators using the symbolic calculus of pseudo-differential operators. After we provide the definition of pseudo-differential operators, we shall deal with the symbol of the product of pseudo-differential operators and provide an estimate for their multi-product. In this chapter we use pseudo-differential operators with both the usual symbols and Weyl symbols. The main part is dedicated to the construction of the fundamental solution as a pseudo-differential operator with parameter t, for both nondegenerate and degenerate parabolic operators. In the case of the quadratic polynomial symbol of .x; /, the exact symbol of the fundamental solution is obtained. These results are then applied to Grushin, sub-Laplacian, and Kolmogorov operators. It is worth noting that the fundamental R c solution E.t/ obtained in this chapter is a smooth operator for any positive t and 0 E.t/ dt is a parametrix for any positive c. This method has proved useful in proving the index theorem in its local version; see Gilkey [53] and Iwasaki [70, 71].
15.1 Basic Results of Pseudo-Differential Operators We shall start with the definition of pseudo-differential operators following H¨ormander [66] and Kumano-go [82], and then we shall provide the estimation of the symbol of multi-product. These computations will play an important role in the construction of the fundamental solution.
O. Calin et al., Heat Kernels for Elliptic and Sub-elliptic Operators, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4995-1 15, c Springer Science+Business Media, LLC 2011
361
362
15 The Pseudo-Differential Operator Technique
15.1.1 Definition of Pseudo-Differential Operators m In the following we shall define the set S;ı of symbols of order m and type ., ı), and organize it as a Fr´echet space with respect to some semi-norms; see H¨ormander [66] and Kumano-go [82]. We recall the well-known notations
j˛j D ˛1 C C ˛n ; @˛1 @˛n @˛x f .x/ D f .x1 ; : : : ; xn /; ˛1 @x1˛n @x1 for any multi-index ˛ D .˛1 ; : : : ; ˛n / 2 Rn and x D .x1 ; : : : ; xn / 2 Rn . Definition 15.1.1. Let m, , ı be real numbers such that 0 ı 1 and ı < 1. m Then we denote by S;ı the set of all C 1 -functions p.x; / defined on Rn Rn satisfying the following estimates for any multi-indices ˛ D .˛1 ; ˛2 ; : : : ; ˛n / and ˇ D .ˇ1 ; ˇ2 ; : : : ; ˇn /; j@˛ @ˇx p.x; /j C˛;ˇ < >mj˛jCıjˇ j W with a positive constant C˛;ˇ , where < >D .1 C jj2 /1=2 . For simplicity we use the notation .˛/ .x; / D @˛ @ˇx p.x; /: p.ˇ / .m/
m For any symbol p.x; / 2 S;ı , we define the semi-norms jpj` , ` D 0; 1; 2; : : : , by .m/
jpj`
D
n max
sup
j˛jCjˇ j` .x;/2Rn Rn
o .˛/ jp.ˇ / .x; /j < >mCj˛jıjˇ j :
(15.1.1)
m becomes a Fr´echet space with respect to the semi-norms (15.1.1). The Then S;ı following notation will also be used: \ m S 1 D S1;0 : 1<m<1
In the following definition the prefix Os- will denote an oscillatory integral. More precisely, Z Os
Rn Rn
e
iy
Z f .y; / dy d D lim
!0 Rn Rn
e iy f .y; /" .y; / dyd;
where " .y; / D ."y; "/; with .y; / a rapidly decreasing function such that .0; 0/ D 1. It is left as an exercise for the reader to show that the previous definition is independent on the cut function .
15.1 Basic Results of Pseudo-Differential Operators
363
m A pseudo-differential operator P of symbol .P / D p.x; / 2 S;ı is an oscil-
latory integral defined for all u 2 S.Rn / D ff 2 C 1 .Rn / I supx2Rn jx ˛ @ˇx f .x/j < 1g, the set of rapidly decreasing functions, as Pu.x/ D p.x; D/u.x/ D Os .2/n n
Z Z
D Os .2/
Rn Rn Rn Rn
e iy p.x; /u.x C y/dy d e i.xy/ p.x; /u.y/dy d:
15.1.2 Calculus with Pseudo-Differential Operators This section will deal with some elements of pseudo-differential operator calculus, such as the product formula, the multi-product formula, and the estimates for symbols of the product of pseudo-differential operators. The product of two pseudo-differential operators P D p.x; D/ and Q D q.x; D/ is also a pseudo-differential operator, whose symbol is denoted by p ı q and is given by .p ı q/.x; D/ D p.x; D/q.x; D/: In fact, p ı q is given as an oscillatory integral of the form .p ı q/.x; / D Os .2/n
Z Rn Rn
e iy p.x; C /q.x C y; /dy d:
The following result can be found in Kumano-go [82]. m1 m2 and q 2 S;ı be two symbols. Then for any Theorem 15.1.2. Let p 2 S;ı integer N , we have the expansion
pıq D
N 1 X
sj .p; q/ C rN .p; q/;
j D0
where sj .p; q/ D
X .i /j˛j m.ı/j ; p .˛/ .x; /q.˛/ .x; / 2 S;ı ˛Š
j˛jDj m.ı/N S;ı ,
rN .p; q/ 2 holds for any `:
and there exist `0 and C such that the following estimate
C jrN jm.ı/N `
X j˛jDN
.m1 j˛j/ .m2 Cıj˛j/ jp .˛/ j`C` jq.˛/ j`C` : 0 0
364
15 The Pseudo-Differential Operator Technique
The following lemma is the key result for proving both the previous theorem and Theorem 15.1.4. We shall provide its proof at the end of this section by following Iwasaki [69]. Lemma 15.1.3. Let f .x 1 ; 1 ; : : : ; x ; / be a C 1 -function on R2n such that ˇ ˇ ˇ ˇY ˇ ˇ ˛j Y ˇ j 1 1 ˇ ˇ @ @xi f .x ; ; : : : ; x ; /ˇ ˇ i ˇ ˇj D1 j D1 M.˛ 1 ; ˛2 ; : : : ; ˛ ; ˇ 1 ; ˇ 2 ; : : : ; ˇ /
Y
< j >mj j˛
j jCıjˇ j j
j D1
for any sequence of multi-indices ˛1 ; ˛ 2 ; : : : ; ˛ ; ˇ 1 ; ˇ 2 ; : : : ; ˇ and constants M.˛1 ; ˛2 ; : : : ; ˛ ; ˇ 1 ; ˇ 2 ; : : : ; ˇ /: Set for 0 j 1, with j D 1; : : : ; 1, Z I D Os
0
Z Rn Rn
Rn Rn
exp @i
1 X
0
1
y j j A f @x; C 1 1 ; x C y 1 ;
j D1
C 2 2 ; x C y 1 C y 2 ; : : : ; C 1 1 ; x C
1 X
1 y j ; A dV;
j D1
where dV D dy1 d1 dy2 d2 dy1 d1 : P Then we can find `0 and a constant C0 that depend on j D1 jmj j but independent of and , such that the following estimate holds: jI j .C0 / M.`0 / < >m0 ; where m0 D
X
mj ;
M.`0 / D
j D1
max
j˛ j jCjˇ j j`0
M.˛ 1 ; ˛2 ; : : : ; ˛ ; ˇ 1 ; ˇ 2 ; : : : ; ˇ /:
Proof of Theorem 15.1.2: By the Taylor expansion we can write .p ı q/.x; / D Os .2/n D Os .2/
n
Z Z
Rn Rn Rn Rn
e iy p.x; C /q.x C y; /dy d e iy
X 1 p .˛/ .x; /˛ ˛Š
j˛j
15.1 Basic Results of Pseudo-Differential Operators
365
q.x C y; /dy d C Os .2/n Z
1
Z Rn Rn
X 1
Š
e iy
j jDN
N.1 /N 1 p . / .x; C /dq.x C y; / dy d
0
D
X .i /j˛j p .˛/ .x; /q.˛/ .x; / C rN .x; /; ˛Š
j˛j
where n
rN .x; / D .2/
.i / N
j jDN
Z Os
X Z
N
Rn Rn
1 0
.1 /N 1
Š
e iy p . / .x; C /q. / .x C y; / dy d d:
Applying Lemma 15.1.3 for rN .x; / with D 2, 1 D ; 2 D 0, and f .x 1 ; 1 ; x 2 ; 2 / D p . / .x 1 ; 1 /q. / .x 2 ; 2 /, we get the assertion. Next, we shall deal with a basic theorem of the symbol of the multi-product of pseudo-differential operators see [69]. m.j / Theorem 15.1.4. If pj 2 S;ı for j D 1; : : : ; , then the product of pseudodifferential operators p1 .x:D/ p .x; D/ is also a pseudo-differential operator whose symbol m p.x; / D .p1 .x:D/ p .x; D// 2 S;ı ;
with m D ˙jD1 m.j /, satisfies the following estimate for any `: .m/
jpj`
C
Y j D1
m.j /
jpj j`C`0 ;
where the constants C and `0 are independent of . Proof. We can write p1 .x; D/ p .x; D/u.x/ Z Z n D Os .2/ Rn Rn
Rn Rn
0 exp @i 0
p1 .x; 1 /p2 .x C y 1 ; 2 / p @x C 0 u @x C
X j D1
1
X
1 yj j A
j D1 1 X
1
yj ; A
j D1
y j A dy 1 d 1 dy 2 d 2 dy d :
366
15 The Pseudo-Differential Operator Technique
So the symbol p.x; / is given by p.x; / D Os .2/n.1/
Y
pj
xC
j D1
j 1 X
0
Z
Z Rn Rn
Rn Rn
exp @i
! y k ; C j
1 X
1 y j j A
j D1
dV;
kD1
where y 0 D 0;
D 0;
dV D dy1 d1 dy2 d2 dy1 d1 :
Now we can apply Lemma 15.1.3 to this formula, setting f .x 1 ; 1 ; : : : ; x ; / D Q j j j D1 pj .x ; / and j D 1. Then we get C jpj.m/ 0
Y j D1
/ jpj jm.j : `0
The proof for any ` can be done by using a similar argument. Remark 15.1.5. Even if it might look tempting to prove the previous estimation using the induction over , this method was proven not feasible for this case.
15.1.3 Proof of Lemma 15.1.3 We shall prepare the ground with a few propositions. In order to avoid repeating the same argument for each variable, for simplicity we may assume j D and ıj D ı for any j . For the proof of Lemma 15.1.3, we use integration by parts with respect to the variables j , j D 1; : : : ; 1. In this case we need the following propositions regarding F .I y/. Set for y; 2 Rn N
F .I y/ D .1C < >2ın0 jyj2n0 /1 ; where ıN D max.ı; 0/ and n0 D Œn=2 C 1. Then using the fact ˇ @ ˛1 @ ˛2 @ ˛n ˇ ˇ ˇ < > ˇ c˛ < >1j˛j ; ˇ @1 @2 @n
(15.1.2)
15.1 Basic Results of Pseudo-Differential Operators
367
we obtain the following result: Proposition 15.1.6. F .I y/ satisfies the inequality with constants C˛;ˇ ˇ ˇ @ ˛ @ ˇ N ˇ ˇ F .I y/ˇ C˛;ˇ F .I y/ < >j˛jCıjˇ j ˇ @ @y for all ˛ and ˇ. Proposition 15.1.7. The following inequality holds for some positive constant C : Z
F C 1 I z1 z0 F C 2 I z2 z1 dz1 F . C 1 I z2 z0 / 2 2 0 1 nıN 2 nıN CF C I z z < C > C : < C > F . C 2 I z2 z0 /
Rn
Proof. For each fixed pair .z0 ; z2 / 2 Rn Rn we divide Rn into two parts: 1 D fz1 2 Rn I jz1 z2 j jz0 z2 j=2g and 2 D Rn n 1 : For z1 2 1 , we have F . C 2 I z2 z1 / 22n0 F . C 2 I z2 z0 /
in 1 :
For z1 2 2 , we get F . C 1 I z1 z0 / 22n0 F . C 1 I z2 z0 /
in 2 :
Then we have Z F . C 1 I z1 z0 /F . C 2 I z2 z1 /d z1 Z 22n0 F . C 2 I z2 z0 / F . C 1 I z1 z0 /d z1 1 Z 2n0 1 2 0 C 2 F . C I z z / F . C 2 I z2 z0 /d z2 :
Rn
(15.1.3)
2
Since 2n0 > n, it is clear that Z Rn
N
F .I y/dy D c1 < >nı :
(15.1.4)
Then by (15.1.3) and (15.1.4), we get Z F . C 1 I z1 z0 /F . C 2 I z2 z1 /d z1 C fF . C 2 I z2 z0 / Rn
N
N
< C 1 >nı CF . C 1 I z2 z0 / < C 2 >nı g with C D 2n0 c1 :
368
15 The Pseudo-Differential Operator Technique
By (15.1.2), if jj c0 < >, there exists a constant c0 > 0 such that j< C>< >j
1 < >: 2
(15.1.5)
For the proof of Lemma 15.1.3, we also use integration by parts with respect to the variables y j , j D 1; : : : ; 1. The following two results deal with the integrand. Proposition 15.1.8. Let 2 Rn be fixed and consider K 0. Set N
I.K/./ D jj2K < C >m f< C > C < >g2K ı F . C I y/ N N < C >nı C < >nı F .I y/ and let N
I1 D f 2 Rn I jj c0 < >ı g; N
I2 D f 2 Rn I c0 < >ı jj c0 < >g; I3 D f 2 Rn I jj c0 < >g:
(15.1.6)
Then we have the constants b.Kk /, which depend on Kk , such that Z I.Kk /./d b.Kk / < >m ;
k D 1; 2; 3;
(15.1.7)
Ik
with K1 D 0;
K2 n=2;
N 0 C n/=2.1 ı/: N K3 .jmj C 1 C 2ın
(15.1.8)
Q Proof. If belongs to I1 or I2 , by (15.1.5), we have for some constant b.K/ N 2K Q I.K/./ b.K/jj < >.2Kn/ıCm ;
K 0:
Hence (15.1.7) is proved for k D 1 and 2 because we have Z Z
N Q 1 / < >nıCm I.K1 /./d b.K I1
Z I1
d; Z
N Q 2 / < >.2K2 n/ıCm I.K2 /./d b.K I2
jj2K2 d:
I2
If belongs to I3 , there is a constant b .K/ such that N
N
N ıKC2ın0 / I.K/ b .K/jj2KC.mC2 ;
m N D max.m; 0/;
(15.1.9)
15.1 Basic Results of Pseudo-Differential Operators
369
since the following system of inequalities holds: 8 << C > 1 C 1 jj; <> c0 N : F .CIy/ . jj /2ın0 C1 : F .Iy/
1 jj; c0
c0
Using (15.1.9), we get Z Z I.K3 /./d b .K3 / I3
N
N ın0 jj2.1ı/K3 CmC1C2 d;
I3
and hence inequality (15.1.7) holds for k D 3, with K3 chosen as above.
Proposition 15.1.9. For k D 1; 2; 3, let N
Jk D jj2Kk f< C > C < >g2ıKk < C >m F . C I z1 z0 /F .I z2 z1 /: Then we have Z Ik
Z Rn
Jk d z1 d B.Kk / < >m F .I z2 z0 /
with B.K/ D C b.K/ if Ik and Kk are defined as in (15.1.6) and (15.1.8). Proof. By means of Proposition 15.1.7, for 1 D ; 2 D 0, we get Z
N
Jk d z1 C jj2Kk f< C > C < >g2ıKk Rn F . C I z2 z0 / N N ın ın <> < C> C F .I z2 z0 / < C >m F .I z2 z0 / D CI.Kk /./F .I z2 z0 /;
k D 1; 2; 3:
Now, by Proposition 15.1.8, we get the assertion.
Now are able to provide the proof for Lemma 15.1.3.P Proof of Lemma 15.1.3. Set n0 D Œn=2 C 1, M D j D1 jmj j, K D Œ.M C N N 2ın0 C n C 1/=2.1 ı/ C 1, and consider the functions Kj D Kj .j I j C1 / with j D 1; : : : ; 1, given by Kj D 0 on Ij;1 ; Kj D n0 on Ij;2 and Kj D K on Ij;3 , where N
Ij;1 D fj 2 Rn I jj j C1 j c0 < C j C1 >ı g; N
Ij;2 D fj 2 Rn I c0 < C j C1 >ı jj j C1 j c0 < C j C1 >g; Ij;3 D fj 2 Rn I jj j C1 j c0 < C j C1 >g;
D 0:
370
15 The Pseudo-Differential Operator Technique
Set B D maxfB.0/; B.n0 /; B.K/g. Integrating by parts, we obtain 8 9 1 < X = D Os exp i y j j : ; Rn Rn Rn Rn Z
I
Z
j D1
1 Y
N
f1 C . j /n0 < C j >2ın0 g
j D1
0
N f1C < C j >2ın0 jy j j2n0 g1 f @x; C 1 1 ; x C y 1 ;
C 2 2 ; x C y 1 C y 2 ; : : : ; C 1 1 ; x C
1 X
1 y j ; A dV;
j D1
where y 0 D 0. Then by the change of variables xC
j X
y k D zj ;
j D 1; : : : ; 1;
kD1
we get 8 9 1 < X = D exp i zj .j j C1 / : ; Rn Rn Rn Rn Z
I
Z
j D1
1 Y
jj j C1 j2Kj . zj /Kj rdV;
j D1
where rD
1 Y
N
f1 C . j /n0 < C j >2ın0 g
j D1
1 Y
F . C j I zj zj 1 /f .x; C 1 1 ;
j D1
z1 ; C 2 2 ; z2 ; : : : : : : ; C 1 1 ; z1 ; /;
15.1 Basic Results of Pseudo-Differential Operators
371
z0 D x and D 0. Then from Proposition 15.1.6 and (15.1.2), we have with a constant C1 that ˇ ˇ 1 ˇ ˇY . zj /Kj rˇ .C1 / M.2.K C n0 // ˇ
(15.1.10)
j D1
1 Y
N
f< C j > C < C j C1 >g2ıKj
j D1
F . C j I zj zj 1 /
Y
< C j >mj ;
j D1
(15.1.11) where z0 D z D x; D 0. From (15.1.10) and Proposition 15.1.9, we get for k D 1, 2, 3 ˇ ˇ ˇ ˇ 1 Z Z Y ˇ ˇ j1 2 j2K1 ˇˇ . zj /Kj rˇˇ dz1 d1 I1;k Rn ˇ ˇj D1 .C1 / BM.2.K C n0 //
1 Y
N
f< C j > C < C j C1 >g2ıKj
j D2 1 Y
F . C j I zj zj 1 / < C j >mj
j D3
F . C 2 I z2 z0 / < C 2 >mQ 2 < >m ; where mQ2 D m1 C m2 . If we repeat the same argument, then we obtain Z
Z Rn Rn
1 Y
Rn Rnj D1
jj j C1 j2Kjj. zj /Kj rjdz1 dz2 dz 2 d1 d2 d 2
.C1 / B 1 M.2.K C n0 //j C1 j2K 1 N
f< C 1 > C < C >g2ıK 1 F . C 1 I z 1 z0 / < C 1 >mQ 1 < >m .z0 D z D x; D 0/;
where m Q 1 D
P1
j D1
Z Rn Rn
mj . Using Proposition 15.1.8, we get Z 1 Y K jj j C1 j2Kj j zjj rjdV Rn Rn j D1
.C1 / B M.2.K C n0 //< >m0 : Taking l0 D 2.K C n0 / and C0 D C1 B leads to the desired result.
372
15 The Pseudo-Differential Operator Technique
15.2 Fundamental Solution by Symbolic Calculus: The Nondegenerate Case This section deals with the construction of the fundamental solution for heat equations by the method of symbolic calculus in the nondegenerate case. We start reviewing the construction of the fundamental solution to the following Cauchy problem on .0; T / Rn following Tsutsumi [108] and Iwasaki [69]: d C P U.t/ D 0 on .0; T / Rn ; dt U.0/ D I on Rn ;
LU D
Here T is a finite positive number and P is a strongly elliptic P differential operator of order m defined on Rn with symbol p.x; /. Let p.x; / D m j D1 pj .x; /, where pj .x; / are homogeneous functions of order j with respect to . The fundamental solution U.t/ is constructed as a pseudo-differential operator with parameter t. For the construction of the fundamental solution, we need the estimate of the multi-product of pseudo-differential operators. Theorem 15.2.1. If there exist positive constants C and R such that pm .x; / C < >m for jj > R; the fundamental solution U.t/ is constructed as a pseudo-differential operator of 0 with parameter t and u.t/ 2 S 1 for t > 0. a symbol u.t/ belonging to S1;0 Moreover, u.t/ has the following expansion for any N : u.t/
N 1 X
N uj .t/ 2 S1;0 ;
j D0
u0 .t/ D e pt ;
j uj .t/ D fj .t/u0 .t/ 2 S1;0 ;
where fj .t/ are polynomials with respect to and t. Proof. Let fj .tI x; / for any j 1 be the solutions of the following ordinary differential equations with parameters .x; /: (
d C p.x; //.fj u0 / D qj .tI x; /; . dt fj jt D0 D 0;
where qj .tI x; / D
X kC` D j 0k j 1
s` .p; fk u0 /:
15.2 Fundamental Solution by Symbolic Calculus: The Nondegenerate Case
373
For example, we have 8 2 f1 D tp1 C t2 s1 .p2 ; p2 /; ˆ ˆ
˚ ˆ ˆ < f2 D tp0 C t 2 .p1 /2 C s1 .p1 ; p2 / C s1 .p2 ; p1 / C s2 .p2 ; p2 / 2 n 3 Pn 2 @ @ C t6 p2 @j@@k p2 s1 .p2 ; s1 .p2 ; p2 // ˆ j;kD1 @xj p2 @x ˆ k o ˆ ˆ 4 : 3p1 s1 .p2 ; p2 / C t8 fs1 .p2 ; p2 /g2 : We note that Z
t
uj .t/ D fj .tI x; /u0 .tI x; / D
u0 .t s/qj .s/ds;
j 1:
0
Set
(
.˛/
aj;˛;ˇ .tI x; / D uj.ˇ / .tI x; /e tp ;
j 0;
.˛/ qj.ˇ / .tI x; /e tp ;
j 1:
bj;˛;ˇ .tI x; / D
Then we have the following estimation for aj;˛;ˇ .tI x; / and bj;˛;ˇ .tI x; / by induction with respect to j and making use of the previous representation of qj .tI x; / and uj .tI x; /. Estimation I: (
jaj;˛;ˇ j Cj;˛;ˇ < >j j˛j !j;˛;ˇ ;
j 0;
0 jbj;˛;ˇ j Cj;˛;ˇ < >mj j˛j !j;˛;ˇ
j 1;
0 are defined by where !j;˛;ˇ and !j;˛;ˇ
!0;0;0 D 1;
! D < >m t;
!0;˛;ˇ D maxf!; ! j˛jCjˇ j g; 2
!j;˛;ˇ D maxf! ; !
2Cj˛jCjˇ j
j˛j C jˇj ¤ 0; g;
j 1;
0 !j;˛;ˇ D maxf! 2 ; ! 2j 1Cj˛jCjˇ j g
j 1:
j , since we have By Estimation I, we have uj 2 S1;0 mt
je tp j C0 e C <> with a positive constant C0 . For any N , we can write p ı uj .t/ D puj .t/ C
NX j 1 `D1
s` p; uj .t/ C rN;j .t/;
374
15 The Pseudo-Differential Operator Technique
mN with rN;j .t/ 2 S1;0 . So we have
0 1 N 1 X d Cp ı@ uj .t/A dt j D0 1 0 NX j 1 1 N 1 N X 1 X NX d A @ uj .t/ C s` p; uj .t/ C rN;j .t/ Cp D dt j D0
D
N 1 X j D0
D
N 1 X
j D0
d C p uj .t/ C dt
N 1 X
qj .t/ C
j D1
j D0
`D1
N 1 X
rN;j .t/
j D0
rN;j .t/:
j D0
Set gN .tI x; / D
N 1 X
uj .tI x; /; rN .tI x; / D
j D0
N 1 X
rN;j .tI x; /
j D0
for any positive integer N . Then gN .tI x; D/ D GN .t/ satisfies following equation with rN .tI x; D/ D RN .t/: (h
d dt
i C p.x; D/ GN .t/ D RN .t/; D I:
GN .0/ Now we will construct u.tI x; / of the form Z
t
u.tI x; D/ D
GN .t s/'.sI x; D/ds: 0
Then '.tI x; D/ D ˆ.t/ must satisfy Volterra’s integral equation Z
t
RN .t/ C ˆ.t/ C
RN .t s/ˆ.s/ds D 0: 0
Set ˆ1 .t/ D RN .t/ and for j 2; Z
t
ˆj .t/ D
ˆ1 .t s/ˆj 1 .s/ds 0 Z t Z s1 Z sj 2 D ˆ1 .t s1 /ˆ1 .s1 s2 / ˆ1 .sj 1 /dsj 1 ds2 ds1 : 0
0
0
15.2 Fundamental Solution by Symbolic Calculus: The Nondegenerate Case
375
Then we have l X
ˆj .t/ D ˆ1 .t/ C
j D1
l X
ˆj .t/
j D2
Z D RN .t/
t
RN .t s/ 0
l1 X
ˆj .s/ ds:
j D1
For .ˆj .t// D 'j .tI x; /; we have the following estimate that helps in solving the previous integral equation. Estimation II: If m N 0, then we have some constants B˛;ˇ independent of j such that ˇ ˇ j 1 t j 1 ˇ .˛/ ˇ .mN / < >mN j˛j : ˇ'j.ˇ / .tI x; /ˇ j'1 jl0 Cj˛jCjˇ j B˛;ˇ .j 1/Š Proof of Estimation II: Note that '1 .tI x; / D rN .tI x; /, We can apply Theorem 15.1.4 for the multi-product ˆ1 .sj 1 sj /: For any l, there exists l0 such that j 1 Z t Z s1 Z sj 2 / .mN / .0/ j C j' j j dsj 1 ds2 ds1 j' j'j .t/j.mN 1 l 1 l l 0
0
j 1 / C j'1 j.0/ C j'1 j.mN l0 l0
0
0
0
t j 1 : .j 1/Š
mN by Proof of Theorem 15.2.1: By Estimation II, we can define '.t/ 2 S1;0 1 X
'j .t/ D '.t/:
j D0
Then '.tI x; D/ D ˆ.t/ satisfies
Z
t
RN .t/ C ˆ.t/ C
RN .t s/ˆ.s/ds D 0; 0
and
ˇ ˇ ˇ .˛/ ˇ .mN / < >mN j˛j expŒB˛;ˇ t: ˇ'.ˇ / .t/ˇ C j'1 jl0
mNQ N For any N , choose NQ D N C m. Then for rNQ .t/ 2 S1;0 D S1;0 ; we can write
u.t/ D
Q 1 N X
uj .t/ C rNQ .t/
j D0
D
N 1 X j D0
uj .t/ C
Q 1 N X j DN
D gN .t/ C rN .t/;
uj .t/ C rNQ .t/
376
15 The Pseudo-Differential Operator Technique
P Q 1 N where rN .t/ D jNDN uj .t/ C rNQ .t/. It is clear that rN .t/ belongs to S1;0 , and hence we have arrived at the desired result. Remark 15.2.2. The kernel of the operator U.t/ D u.tI x; D/ is given by the integral Z K.t; x; y/ D .2/n e i.xy/ u.tI x; /d: Rn
For instance, in the case of the Laplacian P D 2
and u.t/ D e jj t .
Pn
@2 j D1 @xj 2 ,
the symbol is p D jj2 ,
15.3 Basic Results for Pseudo-Differential Operators of Weyl Symbols We shall start with the definition of pseudo-differential operators of Weyl symbols, which will play an important role in the sequel. Definition 15.3.1. Let 0 ı 1; ı < 1. A pseudo-differential operator P on m Rn of Weyl symbol p.x; / 2 S;ı is defined by the formula P u.x/ D p w .x; D/u.x/ Z D Os .2/n
y e iy p x C ; u.x C y/dy d: 2 Rn Rn Z xCy e i.xy/ p ; u.y/dy d: D Os .2/n 2 Rn Rn
In the rest of this chapter we shall use pseudo-differential operators of Weyl symbols. The reader can find details in the references H¨ormander [67]. and Iwasaki and Iwasaki [72].
15.3.1 Calculus with Pseudo-Differential Operators In this section we shall deal with the product formula, multi-product formula and estimates for symbols. The product of two pseudo-differential operators of Weyl
15.3 Basic Results for Pseudo-Differential Operators of Weyl Symbols
377
symbols P D p w .x; D/ and Q D q w .x; D/ is also a pseudo-differential operator of Weyl symbols given by w .p w .x; D/q w .x; D// D p ıw q; that is, p w .x; D/q w .x; D/ D .p ıw q/w .x; D/: Expressed as an oscillatory integral, p ıw q is given by .p ıw q/.x; / D Os .2/2n
Z
Z Rn Rn
Rn Rn
e i.y
1 1 Cy 2 2 /
y2 y1 p x ; C 1 q x C ; C 2 dy 1 d1 dy 2 d2 : 2 2 In fact, p ıw q can be given in the form provided by the following theorem. m1 m2 Theorem 15.3.2. Let p 2 S;ı and q 2 S;ı . Then for any integer N , we have the expansion N 1 X 1 j w j .p; q/ C rN .p; q/; p ıw q D 2i j D0
where j .p; q/ D
X j˛jCjˇ jDj
.1/jˇ j .˛/ .ˇ / m.ı/j .x; / 2 S;ı ; p .x; /q.˛/ ˛ŠˇŠ .ˇ /
m.ı/N w rN .p; q/ 2 S;ı and there exist `0 and C such that the following estimate holds for any `: .m.ı/N /
w j` jrN
C
X
.˛/ .m j˛jCıjˇ j/
j˛jCjˇ jDN
jp.ˇ / j`C`1 0
.ˇ / .m Cıj˛jjˇ j/
jq.˛/ j`C`2 0
:
Remark 15.3.3. Pseudo-differential operators of Weyl symbols are pseudodifferential operators. In fact, we have p w .x; D/ D q.x; D/ if q.x; / D Os .2/n
Z
and p.x; / D Os .2/
n
y e iy p x C ; C dy d 2 Rn Rn
Z
y e iy q x C ; C dy d: 2 Rn Rn
m m is equivalent to q.x; / 2 S;ı . This shows that the condition p.x; / 2 S;ı
For the terms j .p; q/ given in Theorem 15.3.2, we have the following remark.
378
15 The Pseudo-Differential Operator Technique
Remark 15.3.4. It is clear that j .p; q/ D .1/j j .q; p/ for any integer j : So we have j .p; p/ D 0 if j is an odd integer: Remark 15.3.5. In particular, j .p; q/, j D 1; 2, have the following nice representation that is used in the construction of the fundamental solution: 1 .p; q/ D < J rp; rq >; 1 2 .p; q/ D tr.JH p JH q /; 2 where rp Dt
@ @ @ @ p; : : : ; p; p; : : : ; p ; @x1 @xn @1 @n
J is the 2n 2n matrix defined by
0 I I 0
J D
;
and Hp denotes the Hessian matrix. That is, Hp D
@x @x p @ @x p
@x @ p @ @ p
:
Proof of Theorem 15.3.2: In order to show the formula for p ıw q, we shall start by writing p w .x; D/q w .x; D/u.x/ Z 2n D Os .2/ p
x C x1 2
Z
Rn Rn
Rn Rn
e i.xx
1 / 1 Ci.x 1 x 2 / 2
x 1 C x2 ; 1 q ; 2 u.x 2 / dx1 d 1 dx2 d 2 : 2
The integrand can be written as an oscillatory integral: p
x C x1 2
Z ; 1 D Os .2/n
Rn Rn
e i.x
1 w/ 2
p
By the change of variables w D x2 y 2;
x1 D x C y 1;
x C w 2
; 1 dw d2 :
15.3 Basic Results for Pseudo-Differential Operators of Weyl Symbols
379
we have .x x 1 / 1 C .x 1 x 2 / 2 .x 1 w/ 2 D .x x 2 /. 2 2 / y 1 .2 C 1 2 / y 2 2 D .x x 2 / y 1 1 y 2 2 if we set D 2 2 ;
1 D 2 C 1 2 :
Then we obtain the following expression: p w .x; D/q w .x; D/u.x/ Z D Os .2/3n
Z
Rn Rn
p
x C x2 y 2 2
Z Rn Rn
Rn Rn
e i.xx
2 / i.y 1 1 Cy 2 2 /
x C x2 C y 1 ; C 1 q ; C 2 u.x 2 / dy1 d1 dy2 d2 dx2 d: 2
So we have 2n
Z
Z
1
1
2
2
e i.y Cy / .p ıw q/.x; / D Os .2/ Rn Rn Rn Rn y2 y1 1 2 p x ; C q x C ; C dy1 d1 dy2 d2 : 2 2 Using the Taylor expansion, we can write .p ıw q/.x; / D Os .2/2n
Z Rn Rn
Z Rn Rn
e i.y
1 1 Cy 2 2 /
! ! y1 y2 1 2 ; C q x C ; C dy1 d1 dy2 d2 p x 2 2 Z Z 1 1 2 2 D Os .2/2n e i.y Cy / n n n n R R R R ! ! X 1 y1 y2 .˛/ p ; .1 /˛ q x C ; C 2 dy1 d1 dy2 d2 x ˛Š 2 2 j˛j
Š Rn Rn Rn Rn j jDN ! Z 1 2 y ; C 1 d N.1 /N 1 p . / x 2 0 ! y1 2 q x C ; C dy1 d1 dy2 d2 : 2
380
15 The Pseudo-Differential Operator Technique
The first term of the above equation can be written as Os .2/2n
Z
Z
1
1
2
2
e i.y Cy / X 1 y1 y2 .˛/ 1 ˛ 2 x p ; . / q x C ; C dy1 d1 dy2 d2 ˛Š 2 2 j˛j
Rn Rn
Rn Rn
X 1 j˛j X y2 1 .˛/ .ˇ / x p ; q.˛/ .x; /.2 /ˇ dy2 d2 2i ˛ŠˇŠ 2 j˛j
X
D
j˛jCjˇ j
1 2i
j˛jCjˇ j .1/ˇ
1 .˛/ .ˇ / w .1/.x; /; p .x; /q.˛/ .x; / C rN ˛ŠˇŠ
where w rN .1/.x; /
Z
D Os.2/ Z 0
1
n
Rn Rn
e
iy 2 2
X 1 j˛j 2i
j˛j
X j jDN j˛j
y2 1 .˛/ x p ; ˛Š Š 2
.N j˛j/.1 /N j˛j1 q.˛/ .x; C 2 /d.2 / dy2 d2 : . /
So we have .p ıw q/.x; / D
N 1 X j D1
1 j w w j .p; q/ C rN .1/.x; / C rN .2/.x; /; 2i
where w .2/.x; / rN
D Os .2/
2n
Z
Z Rn Rn
Rn Rn
e i.y
1 1 Cy 2 2 /
X 1 .1 /
Š
j jDN
15.3 Basic Results for Pseudo-Differential Operators of Weyl Symbols
381
1 y2 N 1 . / 1 N.1 / p x ; C d 2 0 y1 q x C ; C 2 dy 1 d1 dy 2 d2 : 2 Z
w w w .x; / D rN .1/.x; / C rN .2/.x; /, we get the Applying Lemma 15.1.3 for rN desired assertion.
We have the following theorem for the multi-product of pseudo-differential operators of Weyl symbols. m.j / Theorem 15.3.6. Let pj 2 S;ı for j D 1; : : : ; , and consider pseudodifferential operators of Weyl symbols p1w .x; D/; : : : ; pw .x; D/. Then the product p1w .x:D/ pw .x; D/ is also a pseudo-differential operator of Weyl symbols m p.x; / D w .p1w .x; D/ pw .x; D// 2 S;ı , with m D ˙jD1 m.j /, such that the following estimate is satisfied:
.m/ jpj`
C
Y j D1
m.j /
jpj j`C`0 ;
8`;
with the constants C and `0 independent of . Proof. Since the pseudo-differential operators of Weyl symbols are represented as pseudo-differential operators of usual symbols, applying Theorem 15.1.4, we obtain the desired assertion by transforming it into a pseudo-differential operator of Weyl symbols. Another variant of the proof is based on direct computation. The reader can find the details in the appendix of reference [72]. In this case p.x; / is given by p.x; / D Os .2/n
0
Y j D1
Z
Z Rn
0
Z Rn
j 1
Rn
1X k 1 pj @x C y 2 2 kD1
exp @i
X j D1
X
1 y j j A 1
y k ; C j A dV;
kDj C1
where dV D dy1 d1 dy2 d2 dy d :
382
15 The Pseudo-Differential Operator Technique
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case This section deals with the construction of the fundamental solution for degenerate heat equations via symbolic calculus of Weyl symbols. We shall provide the construction of the following fundamental solution E.t/ as a pseudo-differential operator of Weyl symbol e.tI x; /: d w e .tI x; D/ C p w .x; D/e w .tI x; D/ D 0 on .0; T / Rn ; dt e w .0I x; D/ D I; where T is a finite positive number, and p w .x; P D/ is a pseudo-differential opm erator of Weyl symbol p.x; /. Let p.x; / D j D1 pj .x; /, where pj .x; / are homogeneous functions of order j with respect to . In this section we consider a degenerate operator whose sub-elliptic estimation was provided in Melin [90].
15.4.1 Construction of the Symbol We shall provide in the following proof sketch for the construction of the fundamental solution of the heat equation according to [72]. m Theorem 15.4.1. Suppose that p.x; / 2 S1;0 satisfies the following conditions:
1. pm .x; / 0; 2. pm1 C 12 trC .A/ cjjm1
for jj large and some positive constant c on ˙,
where ˙ is the characteristic set of pm .x; /, A D iJH pm , trC .A/ is the sum of positive eigenvalues of A, the Hessian matrix is given by Hpm D
@x @x pm @ @x pm
and
J D
0 I
@x @ pm @ @ pm
I 0
;
:
Then we can construct the symbol ( e.tI x; / 2
0 S1=2;1=2 ; for t 0; for t > 0; S 1 ;
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
383
which satisfies the following expansion for any integer N: e.tI x; /
N 1 X
N=2
ej .tI x; / 2 S1=2;1=2 ;
j D0
n o e0 .tI x; / D exp '.tI x; / ; n o j=2 ej .tI x; / D fj .tI x; / exp '.tI x; / 2 S1=2;1=2 ; with At i it2 1 h < G.At=2/J rpm ; rpm >; '.tI x; / D pm t Cpm1 t C tr log cosh 2 2 4 where
G.x/ D .1 x 1 tanh x/=x:
Remark 15.4.2. We note that 2
e0 .tI x; / D
e pm t Ci t =4 ; p detfcosh.At=2/g
rpm D 0 on the characteristic set ˙ and
ˇ ˇ ˇp
1 detfcosh.At=2/g
ˇ C ˇ ˇ C e tr At =2 on ˙:
Proof. The proof is divided into four steps as in the following: In step 1 the construction of the main part of the fundamental solution is discussed. In step 2 we provide the estimation of the main part ofP all .x; / 2 Rn Rn . In step 3 we show that the symbol has the expansion j D0 ej .tI x; / with the terms of the form ej .tI x; / D fj .tI x; /e ' , where the fj are approximate solutions of (15.4.22). In the final step we discuss the construction of the fundamental solution by solving an integral equation. Step 1: Construction of the main part Assume the fundamental solution is a pseudo-differential operator of Weyl symbols, with e.tI x; / D e '.t Ix;/ . Applying Theorem 15.3.2, we have X 1 j @ e.t/ C j .p; e.t// D 0; @t 2i 1
j D0
e.0/ D 1:
384
15 The Pseudo-Differential Operator Technique
The following equations can be easily shown 1 .p; e.t// D < J rp; re.t/ > D < J rp; r' > e.t/; 1 2 .p; e.t// D 2 .p; '/e.t/ C < J r'; Hp J r' > e.t/ 2 1 D ftr.JHp JH' /C < J r'; Hp J r' >ge.t/: 2 So we have 2 X 1 j j .p; e.t// 2i j D0 1 1 1 D p < J rp; r' > tr.JH p JH ' / < J r'; Hp J r' > e.t/: 2i 8 8
Neglecting the terms j .p; e.t// for j 3, we get the following equation for ' with '.0/ D 0: @ 1 1 1 'pC < J rp; r' > C tr.JH p JH ' / C < J r'; Hp J r' >D 0: @t 2i 8 8 It might not be easy to find a solution of the above equation. But neglecting the derivatives of p and ' of order of greater than 3, we can find a suitable solution as in the following. Using r.< J rp; r' >/ D Hp J r' C H' J rp; r.< J r'; Hp J r' >/ D r.< r'; JH p J r' >/ D 2H' JH p J r'; we have 1 1 @ r' rp C fHp J r' C H' J rpg H' JH p J r' D 0 @t 2i 4 and
@ 1 1 H' Hp C fHp JH ' C H' JH p g H' JH p JH ' D 0: @t 2i 4 The above equation means that X D iJH ' satisfies @ 1 1 X A C .AX XA/ C XAX D 0; @t 2 4
X jt D0 D 0;
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
385
with A D iJH p . The unique solution of the equation is X D 2 tanh.At=2/: So we can write @ 1 1 J r' J rp C fAJ r' XJ rpg C XAJ r' D 0: @t 2 4 Put y D J r';
b D J rp:
Then we have @ 1 1 y b C fAy Xbg C XAy D 0; @t 2 4
yjt D0 D 0:
The unique solution of the above equation is y D A1 Xb: Now the equation for ' is written as @ i 1 1 ' p C < rp; y > tr.AX / C < y; Hp y >D 0; @t 2 8 8
'jt D0 D 0:
So we have 1 @ i i ' p < Jb; A1 Xb > tr.AX / C < A1 Xb; JXb > D 0: @t 2 8 8 Using the antisymmetry < JAu; v > D < J u; Av >; we have < Jh.A/u; v > D < Ju; h.A/v > if h.x/ is an odd function; < Jh.A/u; v > D < Ju; h.A/v > if h.x/ is an even function: If h.x/ is an even function, we have < Jh.A/u; u > D < Ju; h.A/u > D < J 2 u; J h.A/u > D < J h.A/u; u>:
386
15 The Pseudo-Differential Operator Technique
So we have < Jb; A1 Xb > D 0; < A1 Xb; JXb > D < A1 X 2 b; Jb > D < JA1 X 2 b; b>: ' is obtained by the following formula: 1 ' D pt C 8
Z 0
t
i tr.AX /ds 8
Z
t
< JA1 X 2 b; b > ds:
0
The integrals can be computed as 1 8
Z
Z 1 t tanh2 .As=2/ds 2 0 Z 1 t 1 D ds 1 2 0 cosh2 .As=2/ t 1 .At=2/1 tanh.At=2/ D 2
t
X 2 ds D 0
and 1 8
Z 0
t
Z 1 t AX ds D A=2 tanh.As=2/ds 2 0 it 1h logfcosh.As=2/g D 0 2 1 D logfcosh.At=2/g: 2
Then we have 1 ' D pt C trŒlogfcosh.At=2/g 2 1 D pt C trŒlogfcosh.At=2/g 2
it < JA1 1 .At=2/1 tanh.At=2/ b; b > 2 it2 < G.At=2/J rp; rp > : 4
The main part of ' is tr it2 At pm t C pm1 t C log cosh < G.At=2/J rpm ; rpm > 2 2 4 with A D iJHpm . Step 2: The estimation of the main part We note first that all the eigenvalues of A are real. If A has a nonzero eigenvalue , then is also an eigenvalue of A. These facts are shown in the following. Set
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
387
..u; v// D < iJ u; vN >: Then we have
N D ..u; Av//; ..Au; v// D ..u; Av//
and ..Au; u// D ..Hp u; u// 0: Assume Au D u, with ¤ 0. Then we have 0 < ..Hp u; u// D ..Au; u// D ..u; u//: This means that ..u; u// ¤ 0. On the other hands the following holds: N ..Au; u// D ..u; Au// D ..u; u// D ..u; u//; N We also have and hence we have D . ANu D Nu; and then is an eigenvalue of A. By Remark 15.4.2, we have the estimation ˇ ˇ ˇ ˇ 1 C ˇ ˇ ˇp ˇ C e tr At =2 ; ˇ detfcosh.At=2/g ˇ and by the assumptions on the characteristic set, we have '.x; / C < >m t '.x; / c < >m1 t
if pm .x; / cjjm ; in ˙:
We need precise arguments to obtain an estimate for ' near the set ˙. According to [72], we have 0 e ' 2 S1=2;1=2 : Step 3: The construction of the asymptotic term We shall give an idea of the method of construction of ej .tI x; /. The main difficulty of the construction appears on the characteristic set †. I this step we study how to construct ej .tI x; / near †. For the precise proof, see Iwasaki [72]. Let it2 At tr log cosh < G.At=2/J rp; rp >; '.tI x; / D pt C 2 2 4
388
15 The Pseudo-Differential Operator Technique
where
G.x/ D .1 x 1 tanh x/=x:
Proposition 15.4.3. We have @ ' X 1 j j .p; e ' / e C @t 2i 2
j D0
i D < .1 tanh.At=2//J rp; 2
i i > C tr.AY / < ; AJ 8 8
>;
where it2 At 1 ; (15.4.12) D r tr log cosh 2 2 4 Y D 2r.A1 tanh.At=2//J rp C J r :
(15.4.13)
Proof. By the definition of ', (15.4.12) and (15.4.13), we have At J rp C J ; J r' D 2A1 tanh 2 At JH ' D 2i tanh C Y: 2
(15.4.14) (15.4.15)
Equations (15.4.14) and (15.4.15) imply the following identities:
i 1 < J rp; r' > D i < A1 tanh.At=2/J rp; rp > < rp; J > 2i 2 i (15.4.16) D < J rp; >; 2
where we used that A1 tanh.At=2/ is an even function of A. Then 1 1 i tr.JH p JH ' / D tr.A tanh.At=2/=2/ C tr.AY / 8 2 8
(15.4.17)
and
1 < J r'; Hp J r' > 8 1 D < A1 tanh.At=2/J rp; Hp A1 tanh.At=2/J rp > 2 1 1 < A1 tanh.At=2/J rp; Hp J > < J ; Hp J 2 8
>
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
i < A1 tanh2 .At=2/J rp; rp > 2 i i < tanh.At=2/J rp; > < ; AJ >: 2 8
389
D
(15.4.18)
On the other hand, we have 1 tr.A tanh.At=2/=2/ 2 1 0 < G.At=2/ C .At=2/G .At=2/ J rp; rp > 2 1 D p C tr.A tanh.At=2/=2/ 2 i (15.4.19) < A1 tanh2 .At=2/J rp; rp >; 2
@ ' DpC @t it 2
where we used
x 0 x tanh2 x G .x/ D : 2 2 x By (15.4.16)–(15.4.19), we obtain the assertion. G.x/ C
Now fix an integer N 3. Since m ; p 2 S1;0
we have
0 e ' 2 S1=2;1=2 ;
1 j m j2 j .p; e ' / 2 S1=2;1=2 2i
and N X 1 j m 3 2 j .p; e ' / 2 S1=2;1=2 : 2i
j D3
Then e0 .tI x; / D e
'
satisfies the equation
@ ' X 1 j j .p; e ' / D g0 e ' ; e C @t 2i N
j D0
with
m 3
2 g0 e ' 2 S1=2;1=2 ;
390
15 The Pseudo-Differential Operator Technique
because g0 D
i < .1 tanh.At=2//J rp; 2 N X 1 j C j .p; e ' /e ' : 2i
i i > C tr.AY / < ; AJ 8 8
> (15.4.20)
j D3
Note that g0 is a polynomial of J rp whose coefficients are represented by functions of A and its derivatives are given by (15.4.12) and (15.4.13). For any smooth function h.x; /, we have X 1 j @ j .p; he ' / .he ' / C @t 2i N
j D0
N X @ 1 j .j .p; he ' / hj .p; e ' // h e ' C D @t 2i j D1 9 8 N = <@ X 1 j j .p; e ' / h C e ' C ; : @t 2i
j D0
D
X 1 j @ .j .p; he ' / hj .p; e ' //: h C hg0 e ' C @t 2i N
j D1
The main part of
PN
j D1
j 1 2i
.j .p; he ' / hj .p; e ' // is also given by
2 X 1 j .j .p; he ' / hj .p; e ' //: 2i
j D1
The following equations can easily be shown: 1 .p; he ' / h1 .p; e ' / D 1 .p; h/e ' D < J rp; rh > e ' ; 2 .p; he ' / h2 .p; e ' / D 2 .p; h/e ' < J rh; Hp J r' > e ' 1 D tr.JHp JHh / < J rh; Hp J r' > e ' : 2 Then we have 2 X 1 j .j .p; he ' / hj .p; e ' // 2i
j D1
D
1 < J rp; rh > e ' 2i
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
391
1 2 1 C tr.JHp JHh /e ' < J rh; Hp J r' > e ' 2i 2 D .Qh/e ' ; where Q is a second-order partial differential operator defined by 1 2 1 1 Qh D tr.JHp JHh / < J rh; Hp J r' > : < J rp; rh > C 2i 2i 2 So we have N X @ 1 j j .p; he ' / .he ' / C @t 2i j D0 X 1 j @ ' h C hg0 C Qh e C D .j .p; he ' / hj .p; e ' //: @t 2i 3j N
Next we shall approximate Q. Using J r' D 2A1 tanh.At=2/J rp C J ; we have 1 1 < iJ rp; rh > C < iJHp A1 tanh.At=2/iJ rp; rh > 2 2 i 1 tr.AJHh / C < AJ ; i rh > 8 4 1 1 i D < ftanh.At=2/ 1giJ rp; rh > C tr.AJHh / C < AJ ; i rh > 2 8 4 1 D Q0 h C < AJ ; i rh >; 4
Qh D
where
i 1 < ftanh.At=2/ 1giJ rp; rh > tr.AJHh /: 2 8 If f0 D 1, we formally have Q0 h D
1 X
@ @t
! fk e
kD0
D
'
N X 1 j j C 2i j D0
p;
1 X
! fk e
'
kD0
1 1 1 1 X X X X @ 1 fk C < AJ ; i rfk > e ' fk g0 C Q0 fk e ' ; C @t 4 kD1 1 X
C
kD0
kD1
kD1
X 1 j .j .p; fk e ' / fk j .p; e ' //: 2i
kD1 3j N
392
15 The Pseudo-Differential Operator Technique
Put n` .f1 ; f2 ; : : : ; f` / D
1 < AJ ; i rf` > C 4
X j Ck D`C3 1 k `, 3 j N
1 j j .p; fk e ' / fk j .p; e ' / e ' : 2i Then we have X 1 j X @ X fk e ' C j p; fk e ' @t 2i 1
D
kD0 1 X
kD1
1
N
j D0
@ fk C @t
1 X
kD0
fk g0 C
kD0
1 X
1 X ' Q0 fk e C n` .f1 ; f2 ; : : : ; f` /e ' :
kD1
`D1
It is natural to try to find solutions fk for the following equations: f1 is the solution of @ f1 C Q0 f1 C g0 D 0; @t
f1 jt D0 D 0:
(15.4.21)
fk , with k 2, is the solution of @ fk C fk1 g0 C Q0 fk C nk1 .f1 ; : : : ; fk1 / D 0; @t
fk jt D0 D 0: (15.4.22)
However, we cannot solve the aforementioned equations exactly. We can only find the approximate solution of the aforementioned equations. We prepare the ground for solving (15.4.21) by the following proposition. Let K.t/ be a smooth function valued in an ` ` matrix. Set .t; s) be the unique solution of d dt
C t K.t/
t
.t; s/ D 0;
(15.4.23)
.s; s/ D 0:
Proposition 15.4.4. Let g.t; / be a homogeneous polynomial of degree k in the variable D t .1 ; 2 ; : : : ; ` /, with coefficients smooth functions of t. Then the following polynomial defined by Z
t
h.t; / D
g.s; .t; s//ds 0
(15.4.24)
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
393
is also a homogeneous polynomial of degree k with respect to , and satisfies @ @ h.t; /C < K.t/; h.t; / > D g.t; /; @t @ h.0; / D 0; @ where @
h.t; / D t @ @1 h.t; /; @ @2 h.t; /; : : : ; @ @ h.t; / : `
Proof. Set G.t; / D
@ g.t; /. @
Then by (15.4.23), we have
@ h.t; / D g.t; / C @t
Z
t
d .t; s/; G.s; .t; s// > ds dt
< Z
0 t
D g.t; /
< .t; s/K.t/; G.s; .t; s// > ds Z t t D g.t; / < K.t/; .t; s/G.s; .t; s//ds > 0
0
@ D g.t; / < K.t/; h.t; / >; @ because
@ h.t; / D @
Z
t t
.t; s/G.s; .t; s//ds
0
holds by (15.4.24). Corollary 15.4.5. For an ` ` constant matrix B, set .t; s/ D .1 C e Bs /.1 C e Bt /1 and
Z
t
h.t; / D
g.s; .t; s//ds: 0
Then we have @ 1 @ h.t; / C < B.tanh.Bt=2/ 1/; h.t; / > D g.t; /; @t 2 @ h.0; / D 0:
Proof. Since in our case we have K.t/ D it suffices to show
1 B.tanh.Bt=2/ 1/; 2
Z
t
.t; s/ D exp s
K.r/dr :
394
15 The Pseudo-Differential Operator Technique
Since it is easy to see that Z
t 1 K.r/dr D B.t s/ logŒcosh.Bt=2/ C logŒcosh.Bs=2/; 2 s Z t t s exp K.r/dr D exp B cosh.Bs=2/.cosh.Bt=2//1 2 s
D .1 C e Bs /.1 C e Bt /1 ;
we get the desired assertion. Next we shall go back to solve (15.4.21).
Lemma 15.4.6. We can obtain an approximate solution of (15.4.21), in the sense @ f1 C Q0 f1 C g0 D G1 ; @t
f1 jt D0 D 0;
where G1 is a polynomial of b.D iJ rp/ with order less than g0 , in the sense that m2 : G1 e ' 2 S1=2;1=2 1=2
Moreover, e1 .tI x; / D f1 .t; x; ; b/e ' 2 S1=2;1=2 . Proof. Set b D iJ rp. We will prove the assertion by induction over the degree of g0 with respect to b. Assume that g0 .t; x; ; b/ is a homogenous polynomial of degree k with respect to b. Then Z f1 .t; x; ; b/ D
t
g0 .s; x; ; .t; s/b/ ds 0
is a homogeneous polynomial of degree k and satisfies @ r f1 .t; x; ; b/ D .rf1 /.t; x; ; b/ Ct A f1 .t; x; ; b/; @b where
.t; s/ D .1 C e As /.1 C e At /1 :
Using Corollary 15.4.5 yields 1 @ f1 .t; x; ; b/ C < .tanh.At=2/ 1/b; r f1 .t; x; ; b/ > @t 2 1 D g0 C < .tanh.At=2/ 1/b; .rf1 /.t; x; ; b/ >; 2 f1 .0; x; ; b/ D 0;
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
395
1=2
and we can show that f1 .t; x; ; b/e ' 2 S1=2:1=2 . We have the following fact: @2 t tr.AJHf1 / D tr AJ A 2 f1 .t; x; ; b/A C GQ 1 @b m2 with GQ 1 e ' 2 S1=2:1=2 . Then we have
i Q i @2 @ t f1 C Q0 f1 C g0 D G1 tr AJ A 2 f1 .t; x; ; b/A @t 8 8 @b 1 C < .tanh.At=2/ 1/b; .rf1 /.t; x; ; b/>: 2 m2 We can prove that h.tanh.At=2/ 1/b; .rf1 /.t; x; ; b/ie ' belongs to S1=2;1=2 , 2
@ and that tr.AJ t A @b 2 f1 .t; x; ; b/A/ is a homogenous polynomial of degree k 2 with respect to b. This leads to the desired conclusion.
Instead of (15.4.22) for fk .k 2/, we need to solve the following equations step by step: @ fk C Q0 fk C fk1 g0 C nk1 .f1 ; : : : ; fk1 / C Gk1 D Gk ; @t
fk jt D0 D 0; (15.4.25)
under the condition m3=2k=2 Gk e ' 2 S1=2;1=2 :
We can obtain the solution fk of (15.4.25) and also show that ej .tI x; / D j=2 by a similar method to the construction of f1 . fj .t; x; ; b/e ' 2 S1=2;1=2 Step 4: The construction of the fundamental solution We have constructed ej .tI x; /, for j D 1; 2; : : : ; N 1, such that 0 1 NX 1 @ w ejw .tI x; D/A D rN .tI x; D/; C p w .x; D/ @ @t j D0
mN=2 . So we can construct the symbol of the fundamental with rN .tI x; / 2 S1=2;1=2 solution of the form
e.t/ D
N 1 X
Z ej .t/ C
1 t N X
ej .t s/ ıw
.s/ds;
0 j D0
j D0
mN=2 if m N=2 0. In fact, we can construct .t/ as a unique with .t/ 2 S1=2;1=2 solution of the following equation by the similar method as in Theorem 15.2.1:
Z rN .t/ C
.t/ C
t
rN .t s/ ıw 0
.s/ds D 0:
396
15 The Pseudo-Differential Operator Technique
In this case we apply the estimate of symbols of the multi-product of the pseudomN=2 differential operators of Weyl symbols. Then we have the solution .t/ 2 S1=2;1=2 P 1 of the previous integral equation. So e.t/ has the expansion e.t/ jND0 ej .t/ 2
mN=2 j=2 S1=2;1=2 for any N . Since each ej .t/ belongs to S1=2;1=2 , we can show that e.t/ PN 1 N=2 j D0 ej .t/ 2 S1=2;1=2 . This completes the proof.
15.4.2 The Symbol pm D
1 2
P`
j D1
qj2
If the principal symbol pm has the exactly the double characteristic †, that is, P pm .x; / D 12 `j D1 qj2 .x; /, then the following theorem provides a closed-form expression for '. P Theorem 15.4.7. If pm .x; / D 12 `j D1 qj2 .x; /, then we can choose i t 1 h ' D pm1 t C tr logfcosh.M t=2/g C < F .M t=2/q; q >; 2 2 where
tanh x ; q D t .q1 ; : : : ; q` /; x and M is an ` ` Hermitian matrix defined by F .x/ D
M D .Mjk /;
Mjk D i < rqj ; J rqk >:
We note that if the qj are symbols of vector fields, then the Mjk are symbols for the corresponding commutators. Proof. The previous formula can be obtained in two ways. One is to repeat the proof under the assumption rp D
` X
qj rqj ;
Hp D
j D1
` X
rqjt .rqj /:
j D1
The other way is to use the result obtained in Theorem 15.4.1: ` it2 t X < G.At=2/.iJ rqj t=2/qj ; qk rqk > < G.At=2/J rp; rp > D 4 2 j;kD1
D
` t X qj qk .FQ .Mt=2//jk ; 2 j;kD1
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
397
where FQ .x/ D xG.x/ D 1 F .x/. Then the following identity holds: pt
` it2 t 1X 2 qj t < FQ .M t=2/q; q > < G.At=2/J rp; rp > D 4 2 2 j D1
t D < F .M t=2/q; q>: 2 On the other hand, we have AS D SM; where S is a .2n `/ matrix defined by S D .iJ rq1 ; iJ rq2 ; : : : ; iJ rq` /: So we get detfcosh.At=2/g D detfcosh.M t=2/g: This completes the proof.
15.4.3 The Special Case of Quadratic Symbols Let p.x; / be a polynomial in .x; / of degree at most 2. In this case the fundamental solution E.t/ is obtained as a pseudo-differential operator of Weyl symbol e.tI x; / D exp '.t; x; / , where '.tI x; / D pt C
it2 At tr log cosh < G.At=2/J rp; rp >; 2 2 4
because '.tI x; / is the exact solution of the equation
1 d' 1 1 C p 1 .p; '/ C 2 .p; '/ < J r'; Hp J r' >D 0: dt 2i 4 8
More precisely, we have the following result. This is the key theorem for the construction of the fundamental solution for the heat equation with polynomial coefficients. It is also useful for the construction of the fundamental solution of degenerate parabolic equations. Theorem 15.4.8. If p.x; / is a quadratic polynomial with respect to X D t.x; / 2 Rd Rd , 1 p D hX; HX i C i hX; p0 i C b; 2
398
15 The Pseudo-Differential Operator Technique
then E.t/ D e w .tI x; D/ is given by n e bt e.tI x; / D p exp i hJ tanh.At=2/X; X i det cosh.At=2/ C thJ tanh.At=2/.At=2/
1
o t2 X; Jp0 i C hJ G.At=2/Jp0 ; Jp0 i ; 4
where A D iJH is a 2d 2d matrix, J D and
0 I I 0
;
G.x/ D .1 x 1 tanh x/=x:
Proof. Since rp D HX C ip0 ;
Hp D H;
we have hG.At=2/Jp; pi D hG.At=2/JHX; HX i C hG.At=2/JHX; p0 i C i hG.At=2/Jp0 ; HX i hG.At=2/Jp0 ; p0 i D hJ G.At=2/JHX; JHX i C i hJ G.At=2/JHX; Jp0 i C i hJ G.At=2/Jp0 ; JHX i hJ G.At=2/Jp0 ; Jp0 i; and hence t2 hG.At=2/Jp; pi D hJ G.At=2/.At=2/X; .At=2/X i 4 t C hJ G.At=2/.At=2/X; Jp0i 2 t t2 C hJ G.At=2/Jp0 ; .At=2/X i hJ G.At=2/Jp0 ; Jp0 i: 2 4 Using that t
AJ D JA
and t
.J G.At=2// Dt G.At=2/t J D t G.At=2/J D J G.At=2/;
for the odd function G.x/, we have t2 hG.At=2/Jp; pi D hJ G.At=2/.At=2/2 X; X i C thJ G.At=2/.At=2/X; Jp0i 4 t2 hJ G.At=2/Jp0 ; Jp0 i: 4
15.4 Fundamental Solution by Symbolic Calculus: The Degenerate Case
Since x 2 G.x/ D x tanh x;
399
xG.x/ D 1 x 1 tanh x;
it holds that hJ G.At=2/.At=2/2 X; X i D hJ.At=2 tanh.At=2//X; Xi it D hHX; X i hJ tanh.At=2/X; X i 2 and hJ G.At=2/.At=2/X; Jp0i D hJX; Jp0 i hJ.At=2/1 tanh.At=2/X; Jp0 i D hX; p0 i hJ.At=2/1 tanh.At=2/X; Jp0i: Therefore, we have i
t t2 hG.At=2/Jp; pi D hHX; X i i hJ tanh.At=2/X; X i 4 2 C ithX; p0 i ithJ.At=2/1 tanh.At=2/X; Jp0 i t2 i hJ G.At=2/Jp0 ; Jp0 i 4 D pt bt i hJ tanh.At=2/X; X i ithJ.At=2/1 tanh.At=2/X; Jp0 i t2 i hJ G.At=2/Jp0 ; Jp0 i: 4
Finally, we have tr At Q D bt C log cosh. / C i hJ tanh.At=2/X; X i 2 2 C ithJ.At=2/1 tanh.At=2/X; Jp0 i t2 C i hJ G.At=2/Jp0 ; Jp0 i: 4 Noting that tr At 1 ; exp log cosh. / Dp 2 2 det cosh.At=2/ we get the assertion.
400
15 The Pseudo-Differential Operator Technique
15.4.4 A Key Theorem for Eigenfunction Expansion In a forthcoming section we shall show a method of obtaining eigenfunctions using the fundamental solution represented by pseudo-differential operators of Weyl symbols. The following result is the key of a future proof of the expansion for the kernel of the fundamental solution obtained as a pseudo-differential operator. Proposition 15.4.9. If the symbol of a pseudo-differential operator is of the form p.x; / D h.x; /g.x; /, then the kernel of operator p w .x; D/ in Rd is given by .2/d
Z
0
Rd
e i.xx / p
where gQ
x C x0 2
r 2
r @ r ˇˇ ; ; d D h ; i ;q ˇ gQ rDxCx 0 ;qDxx 0 2 @q 2
Z ; q D .2/d
Rd
e i q g
r
Proof. If the symbol has the expansion h.x; / D x C x 0 and q D x x 0 , we formally have Z .2/d
0
e i.xx / p
x C x0
Z ; d D .2/d
; d:
2 P
X
˛
a˛
2 Rd ˛ Z X r @ ˛ a˛ .2/d e i q g D i ; d 2 @q 2 Rd ˛ r @ r ˇˇ : ; i ;q ˇ Dh gQ rDxCx 0 ;qDxx 0 2 @q 2 Rd
a˛ .x/ ˛ , then with r D r r ˛ e i q g ; d 2 2
r
15.5 The Hermite Operator In the rest of this section we will provide symbols of the fundamental solution as pseudo-differential operators according to Theorem 15.4.8 for several examples that have been studied in previous chapters by different methods. First we will study the heat equation corresponding to the Hermite operator. We will also provide its eigenfunction expansion according to Proposition 15.4.9. The one-dimensional Hermite operator considered in this section is given by P D where c is a positive constant.
1 @2 2 C c2 x2 ; 2 @x
(15.5.26)
15.5 The Hermite Operator
401
15.5.1 Exact Form of the Symbol of the Fundamental Solution The next result deals with the fundamental solution E.t/ of the operator (15.5.26) as a pseudo-differential operator. We note that its kernel K.tI x; x 0 / is given by 1 K.tI x; x / D 2 0
Z
1
e
i.xx 0 /
1
x C x0 e tI ; d : 2
Theorem 15.5.1. The fundamental solution is E.t/ D e w .tI x; D/, with the Weyl symbol 1 tanh.ct=2/ 2 2 e.tI x; / D exp .c x C 2 / : cosh.ct=2/ c The kernel K.tI x; x 0 / is given by 0
r
K.tI x; x / D
n o c c 2 02 0 .x C x / cosh.ct/ 2xx : exp 2 sinh.ct/ 2 sinh.ct/
Proof. The symbol of the operator (15.5.26) is given by 1 2 . C c 2 x 2 /: 2
.P / D Then choosing in Theorem 15.4.8
H D
c2 0
0 1
and X Dt .x; /; we have 1 e.tI x; / D p expfi hJ tanh.At=2/X; X ig: det cosh.At=2/ If c D 0, then e.tI x; / D e jj
2 t =2
:
Assume that c ¤ 0. The 2 2 matrix A D iJH has eigenvalues c; c and eigenvectors U
C
D
1 ic
;
U
D
1 ic
:
402
15 The Pseudo-Differential Operator Technique
It is easy to see that
X D c1 U C C c2 U ;
with c1 D
x Ci ; 2 2c
c2 D
x i : 2 2c
Then we have iJ tanh.At=2/X D i tanh.ct=2/.c1 J U C c2 J U / D tanh.ct=2/W; where
W Dt cx; : c
We have 1 2 . C c 2 x 2 /; c p det cosh.At=2/ D cosh.ct=2/; tanh.ct=2/ 2 . C c 2 x 2 /: i hJ tanh.At=2/X; X i D c hW; X i D
Thus the kernel of the fundamental solution is given by Z 1 x C x0 i.xx 0 / e e tI K.tI z; s; z ; s / D ; d 2 R 2 Z 1 1 0 e i.xx / D 2 R cosh.ct=2/ tanh.ct=2/ 1 2 c .x C x 0 /2 C 2 d exp c 4 r c 1 D p 2 cosh.ct=2/ tanh.ct=2/ c .x x 0 /2 0 2 tanh.ct=2/.x C x / C : exp 4 tanh.ct=2/ 0
0
Using the equalities 1 2 D ; tanh.ct=2/ tanh.ct/ 1 2 tanh.ct=2/ D ; tanh.ct=2/ sinh.ct/ 1 sinh.ct=2/ cosh.ct=2/ D sinh.ct/ 2
tanh.ct=2/ C
15.5 The Hermite Operator
403
yields the following expression for the heat kernel: 0
r
K.tI x; x / D
n o c c 2 02 0 .x C x / cosh.ct/ 2xx : exp 2 sinh.ct/ 2 sinh.ct/
15.5.2 Eigenfunction Expansion We shall apply Proposition 15.4.9 to e.tI x; / given by the previous theorem. The Weyl symbol in the case of the Hermite operator can be written as e.tI x; / D
1 tanh.ct=2/ 2 2 exp .c x C 2 / cosh.ct=2/ c
D h.t; x; /e1 .tI x; /; where e1 .tI x; / D 2e and h.tI x; / D
ct =2
1 2 2 2 exp .c x C / ; c
n o 1 1 2 2 2 exp x C / : 1 tanh.ct=2/ .c 1 C e ct c
With the substitution
w D e ct ;
we have 1 2w @ @2 2 2 D c r =4 2 exp h tI r=2; i @q 1Cw c.1 C w/ @q 1 @ @ 2w D cr=2 C cr=2 : exp 1Cw c.1 C w/ @q @q If we let r D x C x 0 and q D x x 0 , we have cr=2 C and
1 @ D @q 2
@ 1 cr=2 D @q 2
@ @ C cx C 0 C cx 0 @x @x
@ @ 0 C cx C C cx : @x @x 0
404
15 The Pseudo-Differential Operator Technique
Then the following equation holds: @ 1 w h tI r=2; i D Bx C .Bx 0 / exp .Bx / C Bx 0 ; @q 1Cw .1 C w/ where
Note that
1 @ C cx ; Bx D p 2c @x ŒBx ; .Bx / D 1;
Set .x/ D
1 @ 0 C cx Bx 0 D p : 2c @x 0
ŒBx C .Bx 0 / ; .Bx / C Bx 0 D 0: c 14
We conclude with the following result.
cx 2 exp : 2
(15.5.27)
w .tI x; D/. Then we have Proposition 15.5.2. Let eQ1 .tI x; x 0 / be the kernel of e1
eQt;1 .tI x; x 0 / D
p w .x/ .x 0 /:
Proof. We have the following computation: Z 1 x C x0 i.xx 0 / e e1 tI ; d eQ1 .tI x; x / D 2 R 2 p p 2 w c c.x x 0 /2 c.x C x 0 /2 D exp 2 4 4 p D w .x/ .x 0 /: 0
next with the following expansion regarding the operator We shall @ deal h tI r=2; i @q . Proposition 15.5.3. (1) Let Cj .S / be the coefficients defined by the following expansion: 1 h w i X 1 Cj .S / j exp S D w : 1Cw .1 C w/ jŠ j D0
Then the following recursive formula holds: C`C1 .S / D S C` .S / .2` C 1/C` .S / `2 C`1 .S /; C0 .S / D 1: (2) If S D Bx C .Bx 0 / .Bx / C Bx0 , then Cj .S / .x/ .x 0 / D
j .x/
j .x
0
/;
` 0;
15.5 The Hermite Operator
405
where j .x/
D .B /j .x/:
Proof. (1) Consider the analytic function f .z/ defined by z 1 f .z/ D exp S : .1 C z/ 1Cz .1 C z/f .z/ D exp
Since we have
z S ; 1Cz
it follows that f 0 .z/.1 C z/2 C .1 C z/f .z/ D f .z/S;
f .0/ D 1:
Differentiating ` times, we obtain .1 C z/2 f .`C1/ .z/ C 2`.1 C z/f .`/ .z/ C `.` 1/f .`1/ .z/ C .1 C z/f .`/ .z/ C `f .`1/ .z/ D f .`/ .z/S: Making z D 0, we obtain the desired equation for C` .S / D f .`/ .0/. (2) We note that the following equations hold: (i) Bx .x/ D 0; (ii)
Bx
` .x/
D`
`1 .x/;
.Bx /k .Bx /k .x/ D kŠ .x/;
(iii) S
j .x/
j .x
0
/ D
j C1 .x/
Cj
2
j C1 .x
j 1 .x/
0
/ C .2j C 1/
j 1 .x
0
j .x/
j .x
0
/
/:
We shall proceed by induction with respect to j . The assertion obviously holds for j D 0, because C0 .S / D 1 for all S . Note that for any operator S; the operators Cj .S / satisfy part (1). Assume that formula (2) holds for all ` j . Then we have Cj C1 .S / .x/ .x 0 / D S Cj .S / .x/ .x 0 / .2j C 1/Cj .S / .x/ .x 0 / j 2 Cj 1 .S / .x/ .x 0 / D S j .x/ j .x 0 / .2j C 1/ j .x/ j .x 0 / j2 D
j 1 .x/ j 1 .x 0 j C1 .x/ j C1 .x /;
0
/
where the last equality is obtained by (iii). Thus the induction method provides the desired conclusion.
406
15 The Pseudo-Differential Operator Technique
We still have to prove formulas (i), (ii), and (iii). We shall next give the sketch of the proof for (i). Bx
D ŒBx ; .Bx /` .x/ C .Bx /` Bx .x/
` .x/
D `.Bx /`1 .x/ D ` `1 .x/: Formula (ii) is proved by using the following equation inductively: Bxk .Bx /k .x/ D Bxk D D We have .Bx / C Bx0 .
j .x/ j .x
0
// D .Bx / D
noting
k .x/ k1 Bx k k1 .x/ kBxk1 .Bx /k1
.Bx /
j .x/
j C1 .x/
j .x/
.x/:
D
j .x
j .x 0
0
/Cj
/C
j .x/
j .x/
Bx 0
j 1 .x
0
j .x
0
/
/;
j C1 .x/:
Then we have S
j .x/ j .x
0
/
D Bx C .Bx 0 / .
j C1 .x/
j .x
0
/Cj
j .x/
0
j 1 .x
0
D .j C 1/ j .x/ j .x / C j C j C1 .x/ j C1 .x 0 / C j j C1 .x/
//
0
D .Bx j C1 .x// j .x / C j.Bx j .x// j 1 .x / C j C1 .x/..Bx 0 / j .x 0 // C j j .x/..Bx 0 /
D
0
2
j 1 .x/ j .x/
0 j C1 .x / C .2j C 1/
j 1 .x 0
j .x j .x/
0
j 1 .x
0
//
/
/
j .x
0
/ C j2
j 1 .x/
j 1 .x
0
/:
The proof of (iii) is complete. Theorem 15.5.4. The kernel K.tI x; x 0 / of e w .tI x; D/ has the following expansion: 1 X 1 ct .j C1=2/ e K.tI x; x / D jŠ 0
j .x/
j .x
j D0
with
oj n 1 @ C cx D p .x/; @x 2c .x/ defined by (15.5.27). j .x/
with
0
/;
15.6 The Grushin Operator
407
Proof. By Propositions 15.4.9, 15.5.2, and 15.5.3, we have 1 X 1 j @ eQ1 .tI x; x 0 / D w Cj .S /eQ1 .tI x; x 0 / K.tI x; x 0 / D h tI r=2; i @q jŠ j D0
D
1 X j D0
1 X 1 j C1=2 1 j C1=2 Cj .S / .x/ .x 0 / D w w jŠ jŠ
j .x/
j .x
j D0
0
/:
The above theorem tells us that the Hermite operator has p the1eigenvalues fc.j C and the corresponding eigenfunctions f .x/= j Šgj D0. 1=2/g1 j j D0 These results are useful in the study of the following Grushin operator. Taking the Fourier transform, we obtain symbol similar to that of the Hermite operator.
15.6 The Grushin Operator In this section we shall provide the exact form for the symbol of the fundamental solution of the Grushin operator. This is a degenerate operator on R2 defined as 1 P D 2
2 @2 2 @ C x @x 2 @y 2
:
Theorem 15.6.1. We have E.t/ D e w .tI x; y; Dx ; Dy / with e.tI x; y; ; / D
1 tanh.t=2/ 2 exp . C x 2 2 / ; cosh.t=2/
and with its kernel given by K.tI x; y; x 0 ; y 0 / D
Z
1 3 2
0 v
e i.yy / t
r
v 2 sinh.v/
2 t R n o v 2 .x 2 C x 0 / cosh.v/ 2xx 0 dv: exp 2t sinh.v/ Proof. Since we have 1 2 . C x 2 2 /; 2 using Theorem 15.5.1 we easily get the symbol of the fundamental solution. The kernel of the fundamental solution is obtained as in the following: w .P / D
K.tI x; y; x 0 ; y 0 / D
1 2 Z x C x0 y C y 0 0 0 ; ; ; dd e i.xx /Ci.yy / e tI 2 2 2 R2
408
15 The Pseudo-Differential Operator Technique
Z
r
2 sinh.t/ R n o 2 .x 2 C x 0 / cosh.t/ 2xx 0 d exp 2 sinh.t/ Z r 1 v i.yy 0 / tv D e 3 2 sinh.v/ 2 t 2 R n o v 2 .x 2 C x 0 / cosh.v/ 2xx 0 dv: exp 2t sinh.v/
D
1 2
0
e i.yy /
15.7 Exact Form of the Symbol of the Fundamental Solution for the Sub-Laplacian Let x; y; s 2 R, z D x C iy 2 C and consider the following vector fields: ZD
@ @ C i aNz ; @z @s
@ @ ZN D i az ; @Nz @s
T D
@ : @s
This section is concerned with the sub-Laplacian on R3 defined by P D
1 N N Z Z C ZZ C i˛T: 2
Theorem 15.7.1. Let a 2 R and ˛ 2 C be constants such that j˛j < a and a > 0. Then E.t/ D e w .tI x; y; s; Dx ; Dy ; Ds / is obtained with e.tI x; y; s; ; ; / D
tanh.a t/ e ˛ t exp f. C 2ay/2 C . 2ax/2 g cosh.a t/ 4a
and its kernel is given by K.tI z; s; z0 ; s 0 / D
Z
1 1
a exp i fs s 0 C 2a.=.zzN0 //g 2 2 sinh.a t/
a jz z0 j2 C ˛ t d: tanh.a t/
Proof. We have w .L/ D
1 f. C 2ay/2 C . 2ax/2 g ˛; 4
15.7 Exact Form of the Symbol of the Fundamental Solution for the Sub-Laplacian
409
by Theorem 15.4.8, with 0
2a2 2 B 0 H DB @ 0 a
0 2a2 2 a 0
1 a 0 C C 0 A
0 a 1 2
1 2
0
and X D t .x; y; ; /I this yields the symbol e ˛ t expfi hJ tanh.At=2/X; X ig: e.tI x; / D p det cosh.At=2/ Assume that ¤ 0. The 4 4 matrix A D iJH has the eigenvalues 2a;
2a;
0;
0
and the corresponding eigenvectors 1 1 B i C C DB @ 2i a A; 2a 0
UC
1 1 B i C C DB @ 2i a A; 2a 0
U
1 1 B 0 C C U0 D B @ 0 A; 2a 0
1 0 B 1 C C UQ 0 D B @ 2a A: 0 0
It is easy to see that we have the decomposition X D c0 U0 C cQ0 UQ0 C c1 U C C c2 U ; with c0 D
x C ; 2 4a
c1 D
x C iy . C i/ Ci ; 4 8a
cQ0 D
y ; 2 4a c2 D
x iy . i/ i : 4 8a
Then we have iJ tanh.At=2/X D i tanh.a t/.c1 J U C c2 J U / D tanh.a t/W; where WD
ax ; 2
t
ay C ; 2
y C ; 2 4a
x : C 2 4a
410
15 The Pseudo-Differential Operator Technique
By computation we obtain hW; X i D p
1 f. C 2ay/2 C . 2ax/2 g; 4a
det cosh.At=2/ D cosh.a t/;
i hJ tanh.At=2/X; X i D
tanh.a t/ f. C 2ay/2 C . 2ax/2 g: 4a
Then the kernel of the fundamental solution is given by 1 3 Z 0 0 0 K.tI z; s; z ; s / D e i.ss /Ci.xx /Ci.yy / 2 R3 x C x0 y C y 0 ; ; ; ; d d d e tI 2 2 1 3 Z 0 0 0 0 0 e i fss C2a.x yxy /gCi.xx /Ci.yy / D 3 2 R ˛ t e tanh.a t/ exp .jj2 C jj2 / ddd cosh.a t/ 4a Z 1 a D exp i fs s 0 C 2a.x 0 y xy 0 /g 2 1 2 sinh.a t/ a jz z0 j2 C ˛ t d: tanh.a t/ 0
0
15.8 The Sub-Laplacian on Step-2 Nilpotent Lie Groups We apply the last construction of the fundamental solution to a case of the degenerate operators, that is, to the case of the sub-Laplacian on two-step free nilpotent Lie groups. So let FN CN.N 1/=2 Š RN ˚ RN.N 1/=2 be a connected and simply connected free two-step nilpotent Lie group with the Lie algebra fN CN.N 1/=2 (it is also identified with RN ˚ RN.N 1/=2 ). We fix a basis fXj ; Zj; k j 1 j; k N; j < kg of the Lie algebra fN CN.N 1/=2 . Their bracket relation is assumed to be ŒXj ; Xk D 2Zjk
15.8 The Sub-Laplacian on Step-2 Nilpotent Lie Groups
411
for 1 j < k N; and the group multiplication W FN CN.N 1/=2 FN CN.N 1/=2 ! FN CN.N 1/=2 is given by RN ˚ RN.N 1/=2 RN ˚ RN.N 1/=2 X X ˝ X X ˛ 3 xj X j ˚ zjk Zjk ; zQjk Zjk xQ j Xj ˚ X X X X xQ j Xj ˚ 7! xj Xj ˚ zjk Zjk zQjk Zjk X X xj C xQ j /Xj ˚ D zjk Ce zjk C xj xQ k xk xQ j Zjk : Let XQj be the left invariant vector fields on FN C.N 1/=2 : d f .g e tXj /jt D0 XQ j .f /g D dt X X @f @f @f D C xk xk ; @xj @zkj @zjk k<j
k>j
where g D .x; z/ 2 RN ˚ RN.N 1/=2 Š FN CN.N 1/=2 . Let .x; zI ; / D .xj ; zjk I j ; jk / 2 T .FN CN.N 1/=2 / Š RN ˚RN.N 1/=2 RN ˚ RN.N 1/=2 be the dual coordinates on the cotangent bundle; then we understand the symbol of vector fields XQ j and their Weyl symbol as 0 1 X X p w .XQ j / D .XQj / D 1 @j C xk kj xk jk A p D 1. ./x/j ;
k<j
k>j
where D ./ is an N N skew-symmetric matrix defined by ./
jk
D jk
.1 j < k N /:
Let P be the sub-Laplacian P D
N 1 X Q2 Xj I 2 j D1
then its Weyl symbol is given by w .P / D
N 1X w Q 2 .X j / 2 j D1
D
1 < X; HX >; X D t .x; /; 2
412
15 The Pseudo-Differential Operator Technique
with a 2N 2N matrix H defined by 0
. .//2 ./
H D@
./
1 A:
I
We consider the pseudo-differential operator p w .x; z; Dx ; Dz / with the Weyl symbol w .P / and construct the following fundamental solution E.t/ as a pseudodifferential operator of the Weyl symbol e.tI x; z; ; /: d w e .tI x; z; Dx ; Dz / C p w .x; z; Dx ; Dz /e w .tI x; z; Dx ; Dz / D 0 dt in .0; T / RN CN .N 1/=2 ; e w .0I x; z; Dx ; Dz / D I: Theorem 15.8.1. The Weyl symbol e.tI x; z; ; / is obtained as follows:
1 t tanh.i t 0 / e.tI x; z; ; / D p exp H X; X ; 2 i t 0 det cosh.it .// where
0 0 D @
./
0
0
./
1 A:
Then E.t/u.x; z/ D e w .tI x; z; Dx ; Dz /u.x; z/ Z Z N N.N 1/=2 D .2/ RN RN
RN.N 1/=2 RN.N 1/=2
Q e i <xx;>Ci
e.tI .x C x/=2; Q .z C zQ/=2; ; / u.x; Q zQ/ d dQz d dx: Q Corollary 15.8.2. The kernel function of the above fundamental solution K.tI x; z; x; Q zQ/ is given by Z Q Ci =t e i <x; . /x>=t K.tI x; z; x; Q zQ/ D .2 t/N=2N.N 1/=2 RN.N 1/=2 1 i ./ exp < x x; Q .x x/ Q > 2t tanh.i .// s h i ./ i det d ; sinh.i .// where < z; > D
P 1j
zjk jk .
15.8 The Sub-Laplacian on Step-2 Nilpotent Lie Groups
413
From the above expression of the kernel function K.t/, its value on the diagonal is given by Corollary 15.8.3. K.tI x; z; x; z/ D .2 t/
N=2N.N 1/=2
s
Z RN.N 1/=2
h
det
i ./ i d: sinh.i .//
We provide now the proof of Theorem 15.8.1. Let A D iJH be a 2N 2N matrix. Then the power of the matrix A is given by An D A.2i 0 /n1 D .2i 0 /n1 A: Since
0 ADi@
./ 2
./
I
(15.8.28)
1 A;
./
formula (15.8.28) is proved by the induction with respect to n if we note 0 A2 D 2 @
./2
./
./3
./2
1 A D 2iA 0 :
The next result will be used shortly. Lemma 15.8.4. If h0 .x/ is an entire function, then for the entire function h.x/ D xh0 .x/; we have h.tA=2/ D t A h0 .i 0 t/=2:
(15.8.29)
In the case f .x/ D cosh x, we have detfcosh.t A=2/g D detfcosh.i t/g:
(15.8.30)
Proof. Relation (15.8.29) is clear by the previous formula (15.8.28). Also, the formula cosh.t A=2/ D At h.i 0 t/=2 C I2N is derived from (15.8.28), where I2N D identity matrix of size .2N / .2N /
414
15 The Pseudo-Differential Operator Technique
and h.x/ D .cosh x 1/=x. We also have cosh.At=2/ D Ath.i 0 t/=2 C I2N 0 t ./h.it .// 2 Di@ 2t . .//2 h.it .// 0 it D@
2
1
2t h.it .//
A C I2N
t ./h.it .// 2
./h.it .// C I
it2 h.it .//
it2 . .//2 h.it .//
it ./h.it .// 2
1 A
CI
and 0 it det .cosh .tA=2// D det @ D det D det
2
./h.it .// C I
it2 h.it .//
it2 . .//2 h.it .// it2 ./h.it .// I it2 h .it .// it ./ 2 h .it .// C I
1 A
CI
I it2 h .it .// 0 it ./h.it .// C I
D det .it ./h.it .// C I / D det cosh .it .// : Theorem 15.8.1 is obtained by Lemma 15.8.4 and Theorem 15.4.8: The kernel K.tI x; z; x; Q zQ/ is given by the following formula: K.tI x; z; x; Q zQ/ D .2/N N.N 1/=2
Z
Z RN
RN.N 1/=2
Q e i <xx;>Ci
e.tI .x C x/=2; Q .z C zQ/=2; ; / d d: The following equation is clear for a symmetric nonsingular matrix M : Z RN
exp. < M ; > Ci < a; >/ d
D .det M /1=2 ./N=2 exp. < a; M 1a > =4/: Applying the above equation for M D
t tanh.it . // 2 it . /
and
a D .x x/ Q tanh.it .//.x C x/=2; Q we get the assertion of Corollary 15.8.2.
15.9 The Kolmogorov Operator
415
15.9 The Kolmogorov Operator Let P be an operator on R2 defined by P D
@ 1 @2 Cx : 2 2 @x @y
(15.9.31)
The next result deals with the exact form of a symbol of the fundamental solution of the Kolmogorov operator. Theorem 15.9.1. The fundamental solution for the operator (15.9.31) is E.t/ D e w .tI x; y; Dx ; Dy /, with the symbol t 2 t3 2 e.tI x; y; ; / D exp ixt : 2 24 The heat kernel of (15.9.31) is given by " 2# 1 3 3 0 2 0 1 0 exp .x x / 2.y y / .x C x / : K.tI x; y; x ; y / D t2 2t 2t t p
0
0
Proof. We have
In this case
H D
0 0
w .P / D
1 2 C ix: 2
0 ; 1
; p0 D 0
and X Dt .x; /: Note that A2 D 0: So we have tanh.At=2/ D At=2; 1 G.At=2/ D At 6 because G.x/ D
1 x C O.x 5 /: 3
bD0
416
15 The Pseudo-Differential Operator Technique
By Theorem 15.4.8, we have e.tI x; y; ; / n o t2 D exp i hJ.At=2/X; X i C thJX; Jp0 i C hJ.At=6/Jp0 ; Jp0 i 4 3 t t D exp hHX; X i i thX; p0 i hHJp0 ; Jp0 i 2 24 t 2 t3 2 D exp ixt : 2 24 The kernel of the fundamental solution is given by K.tI x; y; x 0 ; y 0 / 1 2 Z x C x0 0 0 e i.xx /Ci.yy / e tI ; ; d d D 2 2 R2 2 (r 3 )2 Z 1 2 0 3 0 2 t .x x i.x x t / / 5d d exp 4 2 C iˇ p D 2 2 24 2t 2t R2 3 1 2 p2 .xx0 /2 Z t 2 2t D p e exp C iˇ d; 2 24 t R with ˇ D y y0
x C x0 t: 2
So we have p n .x x 0 /2 6ˇ 2 o 2 p 24 exp K.tI x; y; x ; y / D 3 4 t 2 " 2t t p 2# 3 1 3 1 2.y y 0 / .x C x 0 / : D exp .x x 0 /2 t2 2t 2t t 0
0
To conclude, the method of pseudo-differential operators is feasible in several important cases, including degenerate operators (the Kolmogorov case), operators with potential (the Hermite case), and sub-elliptic Laplacians on two-step nilpotent Lie groups.
Appendix
Relationship Between the Heat and Wave Kernels This appendix presents the relationship between the wave kernel and the heat kernel of an operator. Since finding wave kernels is more difficult than finding heat kernels, this is not always an efficient method for computing heat kernels. On the other hand, the number of examples which can be worked out explicitly is limited. Here we provide just the example of the operator @2x . A similar method is used in [62] to provide the explicit formula of the heat kernel on the hyperbolic space. Let L be a second-order operator defined on C01 .Rn / with a self-adjoint nonpositive definite extension on L2 .Rn /. One defines the wave kernel of L to be the distribution u on Rn .1; 1/ which satisfies the following Cauchy problem: @2t u D Lu; ujt D0 D 0; @t ujt D0 D ı0 ; where ı0 is the Dirac distribution centered at x D 0. The solution u.x; t/ is a wave which is flat at t D 0 and has the initial velocity given by an impulse function centered at the origin. A formal computation shows that p 1 u.x; t/ D p (A.1) sin t L ; L where L is assumed to be positive definite. On the other side, the heat kernel of L is the distribution v on Rn .0; 1/ such that @t v D Lv; vjt D0 D ı0 : One may formally write v.x; t/ D e tL ;
t > 0:
(A.2)
417
418
Appendix
In order to establish relationship between formulas (A.1) and (A.2), we consider the following Fourier transform identity, where can be either a number or an operator: 2
e t D p D p Substituting
p
Z
1
4 t Z 1 4 t
e
2
s4t
s2
e 4t
1
Z
s2
ds D p e 4t cos.s/ ds 4 t 1 sin.s/ ds: @s
e
is
L for yields e
tL
1
D p 4 t
Z e R
2
s4t
@s
p
1 L
p sin.s L/
ds:
To conclude, we have Theorem 15.9.2. Let v.x; t/ be the heat kernel and u.x; t/ the wave kernel for L. Then Z 1 s2 t > 0: (A.3) e 4t @s u.x; s/ ds; v.x; t/ D p 4 t R As an immediate application of this formula, one may consider the operator L D @2x with the wave kernel u.x; t/ D H.t jxj/, where H.u/ D 1; if u > 0 is the Heaviside function. In this case the derivative in the dis0; if u 0 tribution sense @s u.x; s/ D @s H.s jxj/ D ı.s jxj/ yields the Dirac distribution. Then formula (A.3) provides the well-known expression for the heat kernel Z jxj2 1 1 s2 v.x; t/ D p e 4t ı.s jxj/ ds D p t > 0: e 4t ; 4 t R 4 t Other computations of heat kernels starting from the associated wave kernel constitute a launching ground for future research. A good reference for wave kernels is the paper [59].
Conclusions
At the end of this journey through several techniques of finding explicit formulas for heat kernels, we shall conclude with a few final remarks. Some of the methods presented in this monograph cover only certain operators in most of the cases and are not well suited for the others. We shall discuss in the following the experience gained from using these methods. More precisely, we shall be concerned with what method is suited to which operator. Elliptic and sub-elliptic operators via the geometric method. The heat kernel for a given elliptic operator has a determined physical meaning; i.e., it is the amount of heat transferred between two given points on a Riemannian manifold in a given time. It is physically reasonable to assume that the diffusion of heat occurs along the geodesics of the Riemannian geometry associated with the elliptic operator considered initially. There are a few classes of complete manifolds where the connectivity by geodesics holds; i.e., there is only one geodesic joining any two given points on the manifold. For instance, Hadamard manifolds (negatively curved Riemannian manifolds) and compact manifolds are just a couple of examples. However, the local connectivity by geodesics property always holds on any manifold (Whitehead’s theorem). For the elliptic operators associated with these types of Riemannian geometries, the expression of the heat kernel can be obtained by a closed-form formula that involves a product between the diffusion coefficient, called the volume function, and an exponential term of the action function along the geodesic. This action function, in the case of Riemannian geometry, is obtained by dividing the square of the Riemannian distance by 2t, where t denotes the time parameter. When the geodesics have conjugate points, the expression of the heat kernel is more complicated, as in the case of the circle S 1 . In other cases there might not be any explicit formulas, as the case of the sphere S 2 . We have extended this method to elliptic operators with potential. For these operators the geodesics are replaced by the solutions of the Euler–Lagrange equations associated with the Lagrangian obtained as a difference between the kinetic energy induced by the metric and the potential function. Explicit formulas exist only in the cases when the potential is linear or quadratic. Higher powers lead to unsolved problems, such as the exact solution for the quartic oscillator problem.
419
420
Conclusions
If the operator is sub-elliptic, the appropriate geometry is sub-Riemannian. Unfortunately, in this case the problem of connectivity by geodesics is a very complicated one, since the “local” and “global” behaviors in this case coincide. Since the sub-elliptic operators have always at least one missing direction, it was proved successful in several cases to integrate along the characteristic variety (the space where the principal symbol vanishes) a product between a volume function and the exponential of a modified complex action. The latter function has the important property that it contains information about the lengths of all the geodesics joining any two points. In this case the expression of the heat kernel has an integral representation that in general cannot be reduced to a function. However, in the case of the Kolmogorov operator this integral can be computed explicitly, and the result obtained is a function-type heat kernel. Despite its robustness, the geometric method has its own difficulties. It is not always easy to find the action function, since this is a solution of a nonlinear equation, called the Hamilton–Jacobi equation. In the Riemannian case one can use other equivalent formalisms to find the action function, such as the Lagrangian formalism and the Hamiltonian formalism. In either case, one ends up solving a system of ODEs with boundary conditions. In the case of sub-Riemannian geometry the aforementioned formalisms are not always equivalent, since they produce possibly distinct geodesics (regular and normal geodesics, respectively). It is believed that the normal geodesics are the ones useful in computing the heat kernel. Even if the action function has a reasonable expression, the transport equation, whose solution is the volume function, might not be easily solved explicitly, unless it is of a very particular type. The geometric method can be successfully applied in a number of cases when the Riemannian metric has a familiar expression and the transport equation has separable variables. The role of the Fourier transform. This method is used whenever one wants to eliminate a variable. One of the situations where the Fourier method was proved efficient was in computing the heat kernels for sub-elliptic operators with one missing direction. In this case the partial Fourier transform transforms the sub-elliptic operator into an elliptic one, for which the heat kernel can be obtained by any other method. Then an application of the inverse partial Fourier transform on the heat kernel of the elliptic operator provides the heat kernel for the sub-elliptic operator. The method of eigenfunction expansion. This method can be applied in the case when the operator has a discrete spectrum. For instance, the Laplace–Beltrami operator on a compact Riemannian manifold has this property. The formula for the heat kernel provided by this method involves infinite series, unless some generating formulas exist. Unfortunately, these are rare, and their occurrence deserves to be cherished. The limited number of cases when generating formulas can reduce the heat kernel series to a familiar function includes the generating formulas of Mehler, Hille–Hardy, and Poisson. In general, the method of eigenfunction expansion does not provide a closed-form expression for the kernel. In the case of sub-elliptic operators this method was not as successful as the geometric method.
Conclusions
421
The method of path integrals. In spite of its obscurity, this is the most remarkable of the methods, since it provides an integral formula over all the paths joining the endpoints, reminding us of the way radiation propagates between two given states, making the relationship with quantum mechanics. This path integral can be computed explicitly in some particular cases, by either van Vleck’s or Feynman– Kac’s formula. The first one provides the heat kernel as a function-type formula, reminding us of the geometric method, where the volume function is the square root of the van Vleck determinant. However, this formula is not successfully applicable to sub-elliptic operators. Explicitly working out heat kernels for sub-elliptic operators by path integrals, even for simple cases like the Grushin or Heisenberg operators, has yet to be accomplished at the moment. The stochastic method. Each of the differential operators considered is a generator for an associated Ito diffusion processes. The success of this method is based on the ability to compute the probability density function of the associated Ito diffusion. This method can be applied for both elliptic and sub-elliptic operators. The main difficulty of the method is solving the associated stochastic differential equation. This can be done easily for simple operators such as the Laplace, Grushin or Kolmogorov. In the last two of these cases the probability density is obtained by evaluating an expectation integral using Feynman–Kac’s formula. The stochastic methods have proved useful not only in the case of parabolic operators but also in solving Dirichlet problems. The method of Laguerre calculus. This is the symbolic tensor calculus induced by the Laguerre functions on the Heisenberg group. These functions have been used in the study of the twisted convolution, or equivalently, the Heisenberg convolution. The Laguerre calculus plays a role similar to the Fourier series for a reasonable function defined on the unit circle. In order to invert a sub-elliptic operator (not necessarily second-order) on the Heisenberg group, we need to invert the Laguerre tensor of the operator. The method was successful in computing the heat kernels for the Kohn Laplacian b D @Nb @Nb C @Nb @Nb and Paneitz operator. The advantage of this method over the others is its applicability to finding heat kernels to powers of sub-Laplacians, wave operators and higher-order operators as long as the Laguerre tensor is diagonalizable. However, the method is somewhat limited since it is heavily based on the group structure and orthogonality of the Laguerre functions. The method of pseudo-differential operators. Using the symbolic calculus of pseudo-differential operators, we retrieved the heat kernels of several operators, such as the Hermite, Grushin, Kolmogorov and step-2 sub-Laplacian N C i ˛T operators. The disadvantage of this method is its lim 12 .Z ZN C ZZ/ itation to strongly elliptic operators and some sub-elliptic operators defined on two-step Heisenberg manifolds. The technique cannot be applied to sub-Laplacians of step higher than 2. The sum-of-squares operators. One of the special classes of operators treated in this book is the sum of squares of n linear independent vector fields. If n is the dimension of the space, then the operator is elliptic. Otherwise, it is sub-elliptic. In the latter
422
Conclusions
case, if the vector fields and their iterated brackets generate the tangent space at each point, then we say the bracket generating property holds. If this holds, two things happen: the sum-of-squares operator is hypoelliptic (H¨ormander’s theorem), and the distribution generated by the vector fields has the global connectivity property (Chow’s theorem). If exactly one bracket is needed to generate the tangent space, then the operator is said to be of step 2. The prototype operator in this case is the Heisenberg operator. If exactly k 1 brackets are needed to generate the tangent space, then the operator is said to be of step k. Step k D 1 operators correspond to the elliptic case. The methods and techniques discussed in the present book apply for steps k D 1; 2. Even if most of the methods work theoretically for any step, we did not encounter any explicit example where the computation produces exact solutions. There are a few examples of superior step operators treated in the literature by some adhoc methods, which do not belong to any of the methods treated in this book. One of the further developments is to create techniques to approach these types of operators.
List of Frequently Used Notations and Symbols
Rn j˛j
jxj
The n-dimensional Euclidean space The length of the muti-index ˛1 C C ˛n , where ˛ D .˛1 ; : : : ; ˛n / Partial derivative with respect to xk @˛x11 @˛xnn P 1=2 n 2 if x D .x1 ; : : : ; xn / 2 Rn kD1 xk
L2 .I / Sn rx divX ijk .x/ ıx0 1 ; 2 ; 3 ; 4 Xt X.t/ E.Xt / Cov.Xt ; Xu / dX t F .x1 ; x2 I t1 ; t2 / P .AjB/ Wt W(t) pt .x0 ; x/ Gt .x/ Ik.x/ L x.t/; x.t/ P Scl S(x(t)) d.x0 ; x/ H(x, p) KO
Measurable and square integrable functions on I n-dimensional unit sphere in RnC1 The gradient in the x-variable The divergence of the vector field X The Christoffel symbols The Gamma function The Dirac distribution centered at x0 The theta-functions One-dimensional stochastic process n-dimensional stochastic process .Xt1 ; : : : ; Xtn / The expected value operator of the random variable Xt The covariance function The increment of the stochastic process Xt within time dt The joint distribution function of random variables X1 and X2 The conditional probability of A given B The one-dimensional Brownian motion The n-dimensional Brownian motion .Wt1 ; : : : ; Wtn / The probability density function Gaussian distribution of parameter t The Bessel function of first kind of order k Lagrangian function L W TM ! R The action associated with the classical Lagrangian The action evaluated along the path x.t/ The Riemannian distance between x0 and x Hamiltonian function H W T M ! R The integral kernel of the integral operator K
@xk @˛x
423
424
Px;yIt d m./ H2nC1 h2nC1 H G X K(x, y;t) Fy i [X, Y] hx; yi Hn .x/ L.a/ n TM H C 1 .M / H Id E4 e4 @ G F Wk.p/ P˛ b e.tI x; / MC .F / M.F / j ıi
List of Frequently Used Notations and Symbols
The space of continuous paths from x0 to x within time t The Weyl measure on the path space Px;yIt The Laplace operator 12 .@2x1 C C @2xn / The (2n C 1)-dimensional Heisenberg group The Lie algebra of the Heisenberg group H2nC1 P 2 The Heisenberg Laplacian 2n X kD1 k on the Heisenberg group H2nC1 The Grushin operator P 2 The sub-elliptic Laplacian m kD1 Xk on a manifold of dimension n>m Heat kernel—the amount of heat transferred from x0 to x within time t > 0 The partial Fourier transform with respect to y p The imaginary number 1 The bracket X Y P YX , with X , Y vector fields or operators The inner product nkD1 xk yk if x D .x1 ; : : : ; xn /; y D .y1 ; : : : ; yn / 2 Rn The Hermite polynomial of degree n The Laguerre polynomial of degree n and parameter a The tangent bundle of the manifold M A nonintegrable distribution—a sub-bundle of TM The set of smooth functions on the manifold M The set quaternion numbers The identity map The Engel group The Lie algebra of the Engel group The boundary of the set The noncommutative twisted product of G and F The Cauchy–Szeg¨o kernel The Paneitz operator The Kohn Laplacian The Weyl symbol The positive Laguerre matrix of F The Laguerre tensor of the operator F The Kronecker delta function
References
1. J. Aar˜ao. Fundamental solutions for some partial differential operators from fluid dynamics and statistical physics. SIAM Rev., 49, no. 2, pp. 303–314, 2007. 2. J. Aar˜ao. A transport equation of mixed type. J. Differ. Equ. 150, pp. 188–202, 1998. 3. A.A. Agrachev, U. Boscain, and M. Sigalotti. A Gauss-Bonnet-like formula on twodimensional almost – Riemannian manifolds. preprint, ArXiv:math/0609566v2. 4. M.Y. Antimirov, A.A. Kolyshkin, and R. Vaillancourt. Complex Variables. Academic Press, San Diego, CA, 1997. 5. M.F. Atiyah, R. Bott, and A. Shapiro. Clifford modules. Topology, 3, Suppl. 1, pp. 3–38, 1964. 6. V.I. Arnold. Mathematical Methods of Classical Mechanics. GTM 60, Springer-Verlag, Berlin, 1989. 7. R.O. Bauer. Analysis of the horizontal Laplacian for the Hopf fibration. Forum Math. 17, no. 6, pp. 903–920, 2005. 8. W. Bauer and K. Furutani. Quantization operators on quadrics. Kyushu J. Math. 62, no. 1, pp. 221–258, 2008. 9. W. Bauer and K. Furutani. Spectral Analysis and Geometry of a Sub-Riemannian Structure on S 3 and S 7 . J. Geom. Phys. 58, no. 12, pp. 1693–1738, 2008. 10. W. Bauer and K. Furutani. Zeta regularized determinant of the Laplacian for a class of spherical space forms. J. Geom. Phys. 58, pp. 64–88, 2008. 11. R. Beals. Analysis and geometry on the Heisenberg Group. Conference on Inverse Spectral Geometry, University of Kentucky, June 2002. 12. R. Beals. A note on fundamental solutions. Comm. PDE, 24, pp. 369–376, 1999. 13. R. Beals, B. Gaveau, and P.C. Greiner. On a geometric formula for the fundamental solution of subelliptic Laplacians. Math. Nachr. 181, pp. 81–163, 1996. 14. R. Beals, B. Gaveau, and P.C. Greiner. Complex Hamiltonian mechanics and parametrices for subelliptic Laplacians I, II, III. Bull. Sci. Math. 121, pp. 1–36, 97–149, 195–259, 1997. 15. R. Beals, B. Gaveau, and P.C. Greiner. The Green function of model step two hypoelliptic operators and the analysis of certain tangential cauchy riemannian complexes. Adv. Math. 121, 1996. 16. R. Beals, B. Gaveau, and P.C. Greiner. Hamilton–Jacobi theory and the heat kernel on Heisenberg groups. J. Math. Pures Appl. 79, pp. 633–689, 2000. 17. R. Beals, B. Gaveau, P. Greiner, and J. Vauthier. The Laguerre calculus on the Heisenberg group, II. Bull. Sci. Math. 110, pp. 255–288, 1986. 18. R. Beals and P.C. Greiner. Calculus on Heisenberg manifolds. Ann. Math. Studies, no. 119. Princeton University Press, Princeton, NJ, 1988. 19. C. Berenstein, D.C. Chang, and J. Tie. Laguerre Calculus and Its Application on the Heisenberg Group. AMS/IP Series in Advanced Mathematics, 22, International Press, Cambridge, MA, 2001. 20. A. Boggess and A. Raich. A simplified calculation for the fundamental solution to the heat equation on the Heisenberg group. Proc. Amer. Math. Soc. 137, pp. 937–944, 2009.
425
426
References
21. O. Calin and D.C. Chang. Geometric mechanics on a step 4 subRiemannian manifold. Taiwanese J. Math. 9, no. 2, pp. 261–280, 2005. 22. O. Calin and D.C. Chang. The geometry on a step 3 Grushin operator. Appl. Anal. 84, no. 2, pp. 111–129, 2005. 23. O. Calin and D.C. Chang. On a Class of Nilpotent Distributions, to appear in Taiwanese J. Math. 2010. 24. O. Calin and D.C. Chang. Geometric Mechanics on Riemannian Manifolds. Birkh¨auser, Boston, 2004. 25. O. Calin and D.C. Chang. Heat Kernels for Differential Operators with Radical Function Coefficients, to appear in Taiwanese J. Math. 2010. 26. O. Calin and D.C. Chang. Heat kernels for operators with gravitational potential operators in one and two variables. Pure Appl. Math. Quart. 6, no. 3, pp. 677–692, 2008. 27. O. Calin and D.C. Chang. Sub-Riemannian Geometry, General Theory and Examples. Cambridge University Press, Encyclopedia of Mathematics and Its Applications, 126, Cambridge, 2009. 28. O. Calin, D.C. Chang, and P.C. Greiner. Geometric Analysis on the Heisenberg Group and Its Generalizations. AMS/IP Series in Advanced Mathematics, 40, International Press, Cambridge, MA, 2007. 29. G. Campolieti and R. Makarov. Path integral pricing of Asian options on state-dependent volatility models. Quant. Fin. 8, pp. 147–161, 2008. 30. R. Camporesi. Harmonic analysis and propagators on homogeneous spaces. Phys. Rep. 196, no. 1, pp. 1–134, 1990. 31. C.H. Chang, D.C. Chang, B. Gaveau, P. Greiner, and H.P. Lee. Geometric analysis on a step 2 Grushin operator. Bull. Inst. Math. Academia Sinica (New Series), 4, no. 2, pp. 119–188, 2009. 32. D.C. Chang, S.C. Chang, and J. Tie. Laguerre calculus and Paneitz operator on the Heisenberg group. Sci. China Ser. A 2009. 33. D.C. Chang and I. Markina. Anisotropic quaternion Carnot groups: geometric analysis and Green’s function. Adv. Appl. Math. 39, pp. 345–394, 2007. 34. S.C. Chang, J. Tie, and C.T. Wu. Subgradient estimate and Liouville-type theorems for the CR heat equation on Heisenberg groups, to appear in Asian J. Math. 2011. ¨ 35. W.L. Chow. Uber Systeme von linearen partiellen Differentialgleichungen erster Ordnung. Math. Ann. 117, no. 1, pp. 98–105, 1939. 36. M. Christ and D. Geller. Singular integral characterizations of Hardy spaces on homogeneous groups. Duke Math. J. 51, pp. 547–598, 1984. 37. L.C. Cox, J.E. Ingersoll, and S.A. Ross. A theory of the term structure of interest rates. Econometrica, 53, pp. 385–407. 38. M. de Gosson. The Principles of Newtonian and Quantum Mechanics. Imperial College Press, London, 2001. 39. A. Debiard, B. Gaveau, and E. Mazet. Th´eor´emes de comparison in g´eom´etrie riemanniene. Publ. Kyoto University 12, pp. 391–425, 1976. 40. A.H. Dolley, C. Benson, and G. Ratcliff. Fundamental solutions for the power of the Heisenberg sub-Laplacian. Illinois J. Math. 37, pp. 455–476, 1993. 41. A. Erd´elyi. Higher Transcendental Functions, vols. I, II, III, McGraw-Hill, Inc., Springer, 1953. 42. R. Feynman. Space-time approach to nonrelativistic quantum mechanics. Rev. Mod. Phys. 20, pp. 367–387, 1948. 43. G.B. Folland. Harmonic analysis in phase space. Ann. Math. Studies, 122, Princeton University Press, Princeton, NJ, 1989. 44. G.B. Folland. Introduction to Partial Differential Equations, 2nd ed. Princeton University Press, Princeton, NJ, 1995. 45. G.B. Folland and E.M. Stein. Estimates for the @N b complex and analysis on the Heisenberg group. Comm. Pure Appl. Math. 27, pp. 429–522, 1974. 46. K. Furutani. The heat kernels and the spectrum of a class of nilmanifolds. Comm. Partial Differ. Equ. 21, nos. 3&4, pp. 423–438, 1996.
References
427
47. K. Furutani and C. Iwasaki. Grushin operator and heat kernel on milpotent Lie groups. RIMS Kˆokyˆuroku 1502, Developments of Cartan Geometry and Related Mathematical Problems, July 2006, Research Inst. Math. Sci., Kyoto University, Kyoto, Japan. 48. B. Gaveau. Principe de moindre action, propagation de la chaleur et estim´ees sous elliptiques sur certains groupes nilpotents. Acta Math. 139, pp. 95–153, 1977. 49. B. Gaveau. Syst´emes dynamiques associ´es a` certains op´erateurs hypoelliptiques. Bull. Sci. Math. 102, pp. 203–229, 1978. 50. I.M. Gelfand and G.E. Shilov. Generalized Functions I. Academic, New York, 1964. 51. D. Geller. Fourier analysis on the Heisenberg group. Proc. Natl. Acad. Sci. USA, 74, pp. 1328–1331, 1977. 52. M. Giaquinta and S. Hildebrand. Calculus of Variations, I, II, vol. 310. Springer, New York, 1977. 53. P.B. Gilkey. Invariance Theory, The Heat Equation, and the Atiyah-Singer Index Theorem. Publish or Perish, Inc., Wilmington, Delaware (USA) 1984, Studies in Adv. Math. CRC Press, Boca Raton, Ann Arbor, London, Tokyo 1994 (Second Edition). 54. H. Goldstein. Classical Mechanics. Addison-Wesley, Reading, MA, 1959. 55. C. Gordon and J. Dodziuk. Integral structures on H-type Lie algebras. J. Lie Theory, 12, no. 1, pp. 69–79, 2002. 56. C.R. Graham and J.M. Lee. Smooth solutions of degenerate Laplacians on strictly pseudoconvex Domains. Duke Math. J. 57, pp. 697–720, 1988. 57. P.C. Greiner. On the Laguerre calculus of left-invariant convolution operators on the Heisenberg group. Seminaire Goulaouic-Meyer-Schwartz, exp. XI, pp. 1–39, 1980–81. 58. P. Greiner and O. Calin. On subRiemannian geodesics. Analysis and Applications, 1, no. 3, pp. 289–350, 2003. 59. P.C. Greiner, D. Holcman, and Y. Kannai. Wave kernels related to second-order operators. Duke Math. J. 114, no. 2, pp. 329–387, 2002. 60. P.C. Greiner, J.J. Kohn, and E.M. Stein. Necessary and sufficient conditions for solvability of the Lewy equation. Proc. Natl. Acad. Sci. USA, 72, pp. 3287–3289, 1975. 61. P.C. Greiner and E.M. Stein. On the solvability of some differential operators of type b . Proc. International Conf., Cortona 1976–77, Ann. Scuola Norm. Sup. Pisa Cl. Sci. 4, pp. 106–165, 1978. 62. A. Grigor’yan and Masakazu Noguchi. The heat kernel on hyperbolic space. Bull. Lond. Math. Soc. 30, pp. 643–650, Cambridge University Press, 1998. 63. S.J. Gustavson and I.M. Sigal. Mathematical Concepts of Quantum Mechannics. Springer, Universitext, 2003. 64. K. Hirachi. Scalar pseudo-Hermitian Invariants and the Szeg¨o kernel on 3-dimensional CR manifolds. Lecture Notes in Pure and Appl. Math., Marcel-Dekker, 143, pp. 67–76, 1992. 65. L. H¨ormander. Hypoelliptic second order differential equations. Acta Math. 119, pp. 147–171, 1967. 66. L. H¨ormander. Pseudo-differential operators and hypoelliptic equations. Proc. Symposium on Singular Integrals, Am. Math. Soc. 10, pp. 138–183, 1967. 67. L. H¨ormander. The Weyl calculus of pseudo-differential operators. Comm. Pure Appl. Math. 32, pp. 359–443, 1979. 68. A. Hulanicki. The distribution of energy in the Brownian motion in the Gausssian field and analytic hypoellipticity of certain subelliptic operators on the Heisenberg group. Stud. Math. 56, pp. 165–173, 1976. 69. C. Iwasaki. The fundamental solution for pseudo-differential operators of parabolic type. Osaka J. Math. 14, pp. 569–592, 1977. 70. C. Iwasaki. A proof of the Gauss–Bonnet–Chern theorem by the symbol calculus of pseudodifferential operators. Jpn. J. Math. 21, pp. 235–285, 1995. 71. C. Iwasaki. Symbolic calculus for construction of the fundamental solution for a degenerate equation and a local version of Riemann-Roch theorem in Geometry, Analysis and Applications (R.S. Pathak, ed.). World Scientific, pp. 83–92, 2000.
428
References
72. C. Iwasaki and N. Iwasaki. Parametrix for a degenerate parabolic equation and its application to the asymptotic behavior of spectral functions for stationary problems. Publ. Res. Inst. Math. Sci. 17, pp. 557–655, 1981. 73. G. Johnson and M. Lapidus. The Feynman Integral and Feynman’s Operational Calculus. Oxford Science Publications, Oxford, 2000. 74. M. Kac. Integration in Function Spaces and Some of Its Applications. Accademia Nazionale dei Lincei, Scuola Normale Superioare, Pisa, 1980. 75. M. Kac. On distributions of certain Wiener functionals. Trans. Am. Math. Soc. 65, pp. 1–13, 1949. 76. R.P. Kanwal. Generalized Functions: Theory and Applications, 3rd ed. Birkh¨auser, Boston, 2004. 77. A. Kaplan. On the geometry of groups of Heisenberg type. Bull. Lond. Math. Soc. 15, no. 1, pp. 35–42, 1983. 78. A. Klinger. New derivation of the Heisenberg kernel. Comm. Partial Differ. Equ. 22, pp. 2051– 2060, 1997. 79. S. Kobayashi and K. Nomizu. Foundations of Differential Geometry, vols. I–II. Wiley Interscience, New York, 1996. 80. A.N. Kolmogorov. Uber die analytischen Methoden in der Wahrscheinlichkeitsrechnung. Math. Ann. 104, pp. 415–458, 1931. 81. A.N. Kolmogorov. Zuf¨allige Bewegungen (Zur Theorie der Brownschen Bewegung). Ann. Math. 35, no. 1, pp. 116–117, 1934. 82. H. Kumano-go. Pseudo-Differential Operators. MIT, Cambridge, MA, 1981. 83. H.H. Kuo. Introduction to Stochastic Integration. Springer, Universitext, New York, 2006. 84. J. Lafferty and G. Lebanon. Diffusion kernels on statistical manifolds. Jo. Mach. Learning Res. 2006 (http://www-2.cs.cmu.edu). 85. J. Lamperti. Probability. W.A. Benjamin, Inc. Reading, MA, 1966. 86. D.F. Lawden. Elliptic Functions and Applications, vol. 80. Springer, New York, 1989. 87. J.M. Lee. Pseudo-Einstein Structure on CR-Manifolds. Am. J. Math. 110, pp. 157–178, 1988. 88. J.F. Lingevitch and A.J. Bernoff. Advection of a passive scalar by a vortex couple in the small-diffusion limit. J. Fluid Mech. 270, pp. 219–249, 1994. 89. H.P. McKean. An upper bound to the spectrum of on a manifold of negative curvature. J. Diff. Geom. 4, pp. 359–366, 1970. 90. A. Melin. Lower bounds for pseudo-differential operators. Ark. Mat. 9, pp. 117–140, 1971. 91. S.G. Mikhlin. Compounding of double singular integrals. Doklady Akad. Nauk. USSR, 2, pp. 3–6, 1936. 92. R. Montgomery. A tour of subriemannian geometries, their geodesics and applications. AMS, Math. Surv. Monogr. 91, Providence, RI, 2002. 93. D. M¨uller and E.M. Stein. Lp -estimates for the wave equation on the Heisenberg group. J. Math. Pures Appl. 73, pp. 413–440, 1994. 94. S. Nishikawa. Variational Problems in Geometry. A.M.S. Translations of Mathematical Monographs, Iwanami Series in Modern Mathematics, 205, 2002. 95. B. Oksendal. Stochastic Differential Equations, An Introduction with Applications, 6th ed. Springer, Berlin, 2003. 96. J. Peetre. The Weyl transform and Laguerre polynomials. Le Matematiche, 27, pp. 301–323, 1972. 97. J.H. Rawnsley. A non-unitary pairing of polarization for the Kepler problem. Trans. Amer. Math. Soc. 250, pp. 167–180, 1979. 98. D.B. Ray and I.M. Singer. R-Torsion and the Laplacian on Riemannian manifolds. Adv. Math. 7, pp. 145–210, 1971. 99. S.M. Ross. Stochastic Processes. Wiley, New York, 1996. 100. L.P. Rothschild and E.M. Stein. Hypoelliptic differential operators and nilpotent Lie groups. Acta Math. 137, nos. 3/4, pp. 247–320, 1976. 101. L.S. Schulman. Ph.D. thesis. Phys. Rev. 176, p. 1558, 1968. 102. L.S. Schulman. Techniques and Applications of Path Integration. Dover Publications, Inc., Mineola, NY, 2005.
References
429
103. E.M. Stein. Harmonic Analysis – Real Variable Methods, Orthogonality, and Oscillatory Integrals. 43, Princeton Mathematical Series, Princeton University Press, Princeton, NJ, 1993. 104. R. Strichartz. Lp harmonic analysis and Radon transforms on the Heisenberg group. J. Funct. Anal. 96, pp. 350–406, 1991. 105. R. S. Strichartz. Sub-Riemannian geometry. J. Diff. Geom. 24, no. 2, pp. 221–263, 1986. 106. M. Taylor. Noncommutative Harmonic Analysis. Am. Math. Soc., Providence, RI, 1986. 107. S. Thangavelu. Lectures on Hermite and Laguerre Expansions. Mathematical Notes, no. 42, Princeton University Press, Princeton, NJ, 1993. 108. C. Tsutsumi. The fundamental solution for a degenerate parabolic pseudo-differential operator. Proc. Jpn. Acad. 50, pp. 11–15, 1974. 109. J.H. van Vleck. The correspondence principle in the statistical interpretation of quantum mechanics. Proc. Natl. Acad. Sci. USA, 14, no. 176, pp. 178–188, 1928. 110. N. Wallach. Symplectic Geometry and Fourier Analysis in Lie Groups: History, Frontiers and Applications. vol. V, Math. Sci. Press, 1977. 111. D.V. Widder. The Heat Equation. Academic, London, 1975. 112. S. Wolfgang. Wiener path integrals and the fundamental solution for the Heisenberg Laplacian. J. d’Analyse Math. 91, pp. 389–400, 2003.
Index
action, 24, 61 function, 231 integral, 13 adjoint operator, 166 area element, 180
degenerate operator, 407 diffusion coefficient, 183 distribution function, 177 drift, 173, 174, 183 Dynkin’s formula, 156
Bessel operator, 176 process, 176 bicharacteristic curve, 24 bicharacteristics system, 231 bilinear generating function, 103 boundary conditions, 24 bracket generating, 206 Brownian motion, 152 Brownian motion, 159, 176, 185, 188, 196 on circle, 181 with drift, 168
eigenfunction, 89, 99 eigenfunctions expansion, 403 eigenvalue, 89, 386 eigenvectors, 409 elastic force, 22 electric field, 23 Engel group, 205, 220 Euclidean Laplacian, 266 metric, 273 Euler’s constant, 317 Euler-Lagrange equations, 24 existence, 153, 194 expectation operator, 164 exponential map, 226 exponential martingale, 165
Campbell-Hausdorff formula, 207 Cauchy’s problem, 196, 372 Cauchy-Szeg¨o kernels, 295 characteristic set, 382, 387 Chow’s condition, 201 classical action, 24 Clifford algebra, 218 module, 218 compact support, 90 complete orthogonal system, 89, 99 complex Hamilton-Jacobi theory, 229 conditional expectation, 196 conditional distribution, 179 conservative system, 24 convolution operator, 289 CR-pluriharmonic functions, 310
Feynman-Kac formula, 195 fiber integration, 263 Fourier transform, 97, 307, 316, 355 fundamental solution, 407, 410, 415
Galileo’s equation, 22 Gaussian density, 168 Gaussian distribution, 175 Gaussian random variable, 163 generalized Brownian motion, 168 generalized transport equation, 237, 242
431
432 generating function, 92 generating function formula, 294 generator, 154, 170, 174, 184 of an Ito diffusion, 154 geometric Brownian motion, 171 gravitational field, 21 force, 22 potential, 97 Green functions, 206 Grushin operator, 263, 265, 267, 269, 335, 407 plane, 265 type operator, 263, 264
H¨ormander condition, 201, 238, 244, 263 Haar measure, 204, 209, 292 Hamilton-Jacobi equation, 24, 232, 246 Hamiltonian, 61 function, 23 mechanics, 23 Hamiltonian system, 232 harmonic oscillator, 220 polynomials, 275, 278 heat kernel, 403, 415 Heaviside function, 190 Heisenberg brush, 302 group, 87, 209, 256, 267, 289, 309, 333 Laplacian, 87 sub-Laplacian, 334 type algebra, 227 type group, 261 type Lie algebra, 216 vector fields, 87 Hermite function, 353 operator, 58, 333, 400, 403 polynomial, 91 Hessian matrix, 378 Hessian matrix, 382 Hille-Hardy’s formula, 94 homogeneous polynomial, 394 horizontal Laplacian, 277 hypoellipticity, 203, 238
integrated Brownian motion, 162 isotropic Heisenberg group, 328 Ito diffusion, 152, 167, 188, 196 Ito’s formula, 153, 155, 176, 181
Index Jacobi’s theta-function, 98 transformation, 98
K¨ahler form, 277 kinetic energy, 13 Kohn Laplacian, 291, 306, 349 Kolmogorov operator, 335, 342 Kolmogorov’s backward equation, 156, 196 forward equation, 166 operator, 188 Kolmogorov’s operator, 415 Kronecker delta function, 91, 297
Lagrangian, 13, 23 mechanics, 13 Laguerre calculus, 289, 307 functions, 296 matrix, 296 polynomials, 294, 300, 319 series expansion, 301 tensor forms, 298 Laguerre polynomial, 95 Langevin’s equation, 167 Laplace operator, 341 Laplacian, 99, 104, 196, 376 Legendre function, 104 polynomial, 102 Lewy operator, 291 Lie algebra, 208 Lie algebra, 290 lift, 24 linear noise, 170, 194 linear oscillator, 22 Lipschitz condition, 153 lognormal distributed, 159 distribution, 171 lognormal distributed, 172
mean, 168, 174 measurable functions, 152 measure space, 227 Mehler’s formula, 91, 231 modified Bessel equation, 178 modified Bessel function, 95 multi-index, 362
Index multi-product formula, 363 multiple solutions, 195
nilpotence class, 140 nilpotent Lie algebra, 214 Lie group, 219, 259, 410 Lie groups, 205, 207 nonholonomic sub-bundle, 204 normal distribution, 175
oscillatory integral, 363, 378
Paneitz operator, 292, 309, 320 parabolic equations, 397 partial Fourier transform, 228 path integrals, 230 pendulum string, 22 Poisson’s summation formula, 98 polar coordinates, 97, 180 Popp’s measure, 206 potential energy, 13 function, 196 power series, 93 principal bundle, 275 symbol, 232 principal value convolution, 293 probability density, 175 pseudo-differential operator, 362, 372
quadratic polynomial, 397 quaternion numbers, 273 quaternionic Heisenberg group, 213, 255
Riccati equation, 334, 339 Riemann-Stieltjes integral, 163 Riemannian manifold, 13 metric, 204, 211, 222 volume form, 274 Rodrigues’ formula, 94, 102
433 simple pendulum, 22 Sobolev space, 294 spectral decomposition, 225 spectrum, 227 spherical Grushin operator, 282, 285 spherical harmonics, 295 squared Bessel process, 192 state, 13 stochastic process, 152, 163 sub-elliptic estimation, 382 sub-Laplacian, 203, 232, 254, 282, 408, 410 sub-Riemannian manifold, 202, 263 metric, 204 structure, 201, 219, 263 submersion, 263 symbol, 363, 372, 395, 407 symbolic calculus, 382
tangent bundle, 13 Taylor expansion, 178, 379 Taylor’s formula, 92 thermodynamical systems, 202 theta-function, 7 trajectory, 13 transition density, 158, 171, 176, 184, 186, 194 transition distribution, 163 twisted convolution operator, 307, 318
uniqueness, 153, 194
Van Vleck determinant, 249 variance, 168, 174 vector bundle, 263 Volterra’s integral equation, 374 volume element, 204, 231, 242, 308 form, 263
Wald’s distribution, 177 Weyl symbol, 377, 381–383, 400, 401, 403, 411 Wiener integral, 161, 174
Applied and Numerical Harmonic Analysis J.M. Cooper: Introduction to Partial Differential Equations with MATLAB (ISBN 978-0-8176-3967-9) C.E. D’Attellis and E.M. Fernandez-Berdaguer: ´ Wavelet Theory and Harmonic Analysis in Applied Sciences (ISBN 978-0-8176-3953-2) H.G. Feichtinger and T. Strohmer: Gabor Analysis and Algorithms (ISBN 978-0-8176-3959-4) T.M. Peters, J.H.T. Bates, G.B. Pike, P. Munger, and J.C. Williams: The Fourier Transform in Biomedical Engineering (ISBN 978-0-8176-3941-9) A.I. Saichev and W.A. Woyczynski: ´ Distributions in the Physical and Engineering Sciences (ISBN 978-0-8176-3924-2) R. Tolimieri and M. An: Time-Frequency Representations (ISBN 978-0-8176-3918-1) G.T. Herman: Geometry of Digital Spaces (ISBN 978-0-8176-3897-9) A. Prochazka, ´ J. Uhli˜r, P.J.W. Rayner, and N.G. Kingsbury: Signal Analysis and Prediction (ISBN 978-0-8176-4042-2) J. Ramanathan: Methods of Applied Fourier Analysis (ISBN 978-0-8176-3963-1) A. Teolis: Computational Signal Processing with Wavelets (ISBN 978-0-8176-3909-9) W.O. Bray and C.V. Stanojevi´c: Analysis of Divergence (ISBN 978-0-8176-4058-3) G.T Herman and A. Kuba: Discrete Tomography (ISBN 978-0-8176-4101-6) J.J. Benedetto and P.J.S.G. Ferreira: Modern Sampling Theory (ISBN 978-0-8176-4023-1) A. Abbate, C.M. DeCusatis, and P.K. Das: Wavelets and Subbands (ISBN 978-0-8176-4136-8) L. Debnath: Wavelet Transforms and Time-Frequency Signal Analysis (ISBN 978-0-8176-4104-7) K. Grochenig: ¨ Foundations of Time-Frequency Analysis (ISBN 978-0-8176-4022-4) D.F. Walnut: An Introduction to Wavelet Analysis (ISBN 978-0-8176-3962-4) O. Bratteli and P. Jorgensen: Wavelets through a Looking Glass (ISBN 978-0-8176-4280-8) H.G. Feichtinger and T. Strohmer: Advances in Gabor Analysis (ISBN 978-0-8176-4239-6) O. Christensen: An Introduction to Frames and Riesz Bases (ISBN 978-0-8176-4295-2) L. Debnath: Wavelets and Signal Processing (ISBN 978-0-8176-4235-8) J. Davis: Methods of Applied Mathematics with a MATLAB Overview (ISBN 978-0-8176-4331-7) G. Bi and Y. Zeng: Transforms and Fast Algorithms for Signal Analysis and Representations (ISBN 978-0-8176-4279-2) J.J. Benedetto and A. Zayed: Sampling, Wavelets, and Tomography (ISBN 978-0-8176-4304-1) E. Prestini: The Evolution of Applied Harmonic Analysis (ISBN 978-0-8176-4125-2) O. Christensen and K.L. Christensen: Approximation Theory (ISBN 978-0-8176-3600-5) L. Brandolini, L. Colzani, A. Iosevich, and G. Travaglini: Fourier Analysis and Convexity (ISBN 978-0-8176-3263-2) W. Freeden and V. Michel: Multiscale Potential Theory (ISBN 978-0-8176-4105-4) O. Calin and D.-C. Chang: Geometric Mechanics on Riemannian Manifolds (ISBN 978-0-8176-4354-6)
Applied and Numerical Harmonic Analysis (Cont’d) J.A. Hogan and J.D. Lakey: Time-Frequency and Time-Scale Methods (ISBN 978-0-8176-4276-1) C. Heil: Harmonic Analysis and Applications (ISBN 978-0-8176-3778-1) K. Borre, D.M. Akos, N. Bertelsen, P. Rinder, and S.H. Jensen: A Software-Defined GPS and Galileo Receiver (ISBN 978-0-8176-4390-4) T. Qian, V. Mang I, and Y. Xu: Wavelet Analysis and Applications (ISBN 978-3-7643-7777-9) G.T. Herman and A. Kuba: Advances in Discrete Tomography and Its Applications (ISBN 978-0-8176-3614-2) M.C. Fu, R.A. Jarrow, J.-Y. J. Yen, and R.J. Elliott: Advances in Mathematical Finance (ISBN 978-0-8176-4544-1) O. Christensen: Frames and Bases (ISBN 978-0-8176-4677-6) P.E.T. Jorgensen, K.D. Merrill, and J.A. Packer: Representations, Wavelets, and Frames (ISBN 978-0-8176-4682-0) M. An, A.K. Brodzik, and R. Tolimieri: Ideal Sequence Design in Time-Frequency Space (ISBN 978-0-8176-4737-7) S.G. Krantz: Explorations in Harmonic Analysis (ISBN 978-0-8176-4668-4) G.S. Chirikjian: Stochastic Models, Information Theory, and Lie Groups, Volume 1 (ISBN 978-0-8176-4802-2) C. Cabrelli and J.L. Torrea: Recent Developments in Real and Harmonic Analysis (ISBN 978-0-8176-4531-1) B. Luong: Fourier Analysis on Finite Abelian Groups (ISBN 978-0-8176-4915-9) M.V. Wickerhauser: Mathematics for Multimedia (ISBN 978-0-8176-4879-4) P. Massopust and B. Forster: Four Short Courses on Harmonic Analysis (ISBN 978-0-8176-4890-9) O. Christensen: Functions, Spaces, and Expansions (ISBN 978-0-8176-4979-1) J. Barral and S. Seuret: Recent Developments in Fractals and Related Fields (ISBN 978-0-8176-4887-9) O. Calin, D.-C. Chang, K. Furutani, and C. Iwasaki: Heat Kernels for Elliptic and Sub-elliptic Operators (ISBN 978-0-8176-4994-4)