Statistical Fluid Mechanics, Volume II: Mechanics of Turbulence
By A. S. Monin and A. M. Yaglom
()
About this ebook
Written by two of Russia's most eminent and productive scientists in turbulence, oceanography, and atmospheric physics, this two-volume survey is renowned for its clarity as well as its comprehensive treatment. The first volume begins with an outline of laminar and turbulent flow. The remainder of the book treats a variety of aspects of turbulence: its statistical and Lagrangian descriptions, shear flows near surfaces and free turbulence, the behavior of thermally stratified media, and diffusion.
Volume Two continues and concludes the presentation. Topics include spectral functions, homogeneous fields, isotropic random fields, isotropic turbulence, self-preservation hypotheses, spectral energy transfer, the Millionshchikov hypothesis, acceleration fields, equations for higher moments and the closure problem, and turbulence in a compressible fluid. Additional subjects include general concepts of the local structure of turbulence at high Reynolds numbers, the theory of fully developed turbulence, the propagation of electromagnetic and acoustic waves through a turbulent medium, and the twinkling of stars. The book closes with a discussion of the functional formulation of the problem of turbulence, presenting the equations for the characteristic functional and methods for their solution.
Related to Statistical Fluid Mechanics, Volume II
Titles in the series (100)
Quantum Mechanics with Applications Rating: 2 out of 5 stars2/5Treatise on Physiological Optics, Volume III Rating: 0 out of 5 stars0 ratingsThe Theory of Heat Radiation Rating: 3 out of 5 stars3/5A First Look at Perturbation Theory Rating: 4 out of 5 stars4/5Electronic Structure and the Properties of Solids: The Physics of the Chemical Bond Rating: 3 out of 5 stars3/5Theory of Linear Physical Systems: Theory of physical systems from the viewpoint of classical dynamics, including Fourier methods Rating: 0 out of 5 stars0 ratingsQuantum Mechanics of One- and Two-Electron Atoms Rating: 0 out of 5 stars0 ratingsGroup Theory in Quantum Mechanics: An Introduction to Its Present Usage Rating: 0 out of 5 stars0 ratingsRational Mechanics: The Classic Notre Dame Course Rating: 5 out of 5 stars5/5Brownian Movement and Molecular Reality Rating: 0 out of 5 stars0 ratingsStatistical Fluid Mechanics, Volume II: Mechanics of Turbulence Rating: 0 out of 5 stars0 ratingsTheories of Figures of Celestial Bodies Rating: 0 out of 5 stars0 ratingsMathematics of Relativity Rating: 0 out of 5 stars0 ratingsPhysics of Waves Rating: 5 out of 5 stars5/5Equilibrium Statistical Mechanics Rating: 4 out of 5 stars4/5Neutrons, Nuclei and Matter: An Exploration of the Physics of Slow Neutrons Rating: 0 out of 5 stars0 ratingsReadable Relativity Rating: 4 out of 5 stars4/5Problems in Quantum Mechanics: Third Edition Rating: 3 out of 5 stars3/5Gravitational Curvature: An Introduction to Einstein's Theory Rating: 0 out of 5 stars0 ratingsLight Rating: 4 out of 5 stars4/5Dynamic Light Scattering: With Applications to Chemistry, Biology, and Physics Rating: 5 out of 5 stars5/5Thermoelectricity: An Introduction to the Principles Rating: 4 out of 5 stars4/5The Philosophy of Space and Time Rating: 5 out of 5 stars5/5Symmetry: An Introduction to Group Theory and Its Applications Rating: 4 out of 5 stars4/5An Introduction to Acoustics Rating: 1 out of 5 stars1/5Water Waves: The Mathematical Theory with Applications Rating: 0 out of 5 stars0 ratingsAn Elementary Survey of Celestial Mechanics Rating: 0 out of 5 stars0 ratingsQuantum Mechanics in Hilbert Space: Second Edition Rating: 4 out of 5 stars4/5An Elementary Treatise on Theoretical Mechanics Rating: 5 out of 5 stars5/5A History of Mechanics Rating: 4 out of 5 stars4/5
Related ebooks
Mathematical Aspects of Subsonic and Transonic Gas Dynamics Rating: 5 out of 5 stars5/5Group Theory in Quantum Mechanics: An Introduction to Its Present Usage Rating: 0 out of 5 stars0 ratingsClassical Electricity and Magnetism: Second Edition Rating: 4 out of 5 stars4/5Fundamental Principles of Modern Theoretical Physics: International Series of Monographs in Natural Philosophy Rating: 0 out of 5 stars0 ratingsThe Laminar Boundary Layer Equations Rating: 0 out of 5 stars0 ratingsThe Classical Theory of Fields Rating: 5 out of 5 stars5/5Methods of Quantum Field Theory in Statistical Physics Rating: 4 out of 5 stars4/5Mechanics and Electrodynamics Rating: 5 out of 5 stars5/5Course of Theoretical Physics Rating: 5 out of 5 stars5/5Quantum Theory of Collective Phenomena Rating: 0 out of 5 stars0 ratingsQuantum Theory of Many-Particle Systems Rating: 5 out of 5 stars5/5A Primer to the Theory of Critical Phenomena Rating: 0 out of 5 stars0 ratingsKinetic Boltzmann, Vlasov and Related Equations Rating: 0 out of 5 stars0 ratingsFoundations of Classical and Quantum Statistical Mechanics: International Series of Monographs in Natural Philosophy Rating: 0 out of 5 stars0 ratingsDistributions and Fourier Transforms Rating: 4 out of 5 stars4/5Fourier Series and Orthogonal Polynomials Rating: 0 out of 5 stars0 ratingsAsymptotic Theory of Supersonic Viscous Gas Flows Rating: 0 out of 5 stars0 ratingsA Mathematical Companion to Quantum Mechanics Rating: 0 out of 5 stars0 ratingsNonlinear Dynamics: Exploration Through Normal Forms Rating: 5 out of 5 stars5/5Periodic Differential Equations: An Introduction to Mathieu, Lamé, and Allied Functions Rating: 0 out of 5 stars0 ratingsSolved Problems in Classical Electromagnetism Rating: 0 out of 5 stars0 ratingsTheory of Linear Physical Systems: Theory of physical systems from the viewpoint of classical dynamics, including Fourier methods Rating: 0 out of 5 stars0 ratingsUnsteady Transonic Flow Rating: 0 out of 5 stars0 ratingsFunctional Analysis Rating: 4 out of 5 stars4/5Wave Propagation in Layered Anisotropic Media: with Application to Composites Rating: 0 out of 5 stars0 ratingsWater Waves: The Mathematical Theory with Applications Rating: 0 out of 5 stars0 ratingsAn Introduction to the Theory of the Boltzmann Equation Rating: 2 out of 5 stars2/5Low and High Frequency Asymptotics: Acoustic, Electromagnetic and Elastic Wave Scattering Rating: 0 out of 5 stars0 ratings
Physics For You
What If?: Serious Scientific Answers to Absurd Hypothetical Questions Rating: 5 out of 5 stars5/5Step By Step Mixing: How to Create Great Mixes Using Only 5 Plug-ins Rating: 5 out of 5 stars5/5Physics I For Dummies Rating: 4 out of 5 stars4/5Quantum Physics for Beginners Rating: 4 out of 5 stars4/5The Reality Revolution: The Mind-Blowing Movement to Hack Your Reality Rating: 4 out of 5 stars4/5Quantum Physics: A Beginners Guide to How Quantum Physics Affects Everything around Us Rating: 5 out of 5 stars5/5Quantum Computing For Dummies Rating: 0 out of 5 stars0 ratingsWelcome to the Universe: An Astrophysical Tour Rating: 4 out of 5 stars4/5String Theory For Dummies Rating: 4 out of 5 stars4/5Physics Essentials For Dummies Rating: 4 out of 5 stars4/5DIY Lithium Battery Rating: 3 out of 5 stars3/5How to Diagnose and Fix Everything Electronic, Second Edition Rating: 4 out of 5 stars4/5Quantum Spirituality: Science, Gnostic Mysticism, and Connecting with Source Consciousness Rating: 5 out of 5 stars5/5How to Teach Quantum Physics to Your Dog Rating: 4 out of 5 stars4/5Feynman Lectures Simplified 1A: Basics of Physics & Newton's Laws Rating: 5 out of 5 stars5/5The Grid: The Fraying Wires Between Americans and Our Energy Future Rating: 4 out of 5 stars4/5QED: The Strange Theory of Light and Matter Rating: 4 out of 5 stars4/5The Physics Book: From the Big Bang to Quantum Resurrection, 250 Milestones in the History of Physics Rating: 4 out of 5 stars4/5What the Bleep Do We Know!?™: Discovering the Endless Possibilities for Altering Your Everyday Reality Rating: 5 out of 5 stars5/5Beyond Weird: Why Everything You Thought You Knew about Quantum Physics Is Different Rating: 4 out of 5 stars4/5Unlocking Spanish with Paul Noble Rating: 5 out of 5 stars5/5The God Effect: Quantum Entanglement, Science's Strangest Phenomenon Rating: 4 out of 5 stars4/5Moving Through Parallel Worlds To Achieve Your Dreams Rating: 4 out of 5 stars4/5The Theory of Relativity: And Other Essays Rating: 4 out of 5 stars4/5The Dancing Wu Li Masters: An Overview of the New Physics Rating: 4 out of 5 stars4/5A Brief Welcome to the Universe: A Pocket-Sized Tour Rating: 5 out of 5 stars5/5Midnight in Chernobyl: The Untold Story of the World's Greatest Nuclear Disaster Rating: 4 out of 5 stars4/5
Reviews for Statistical Fluid Mechanics, Volume II
0 ratings0 reviews
Book preview
Statistical Fluid Mechanics, Volume II - A. S. Monin
STATISTICAL FLUID
MECHANICS:
Mechanics of Turbulence
Volume II
A. S. MONIN and A. M. YAGLOM
English edition updated, augmented and revised by the authors
Edited by John L. Lumley
DOVER PUBLICATIONS, INC.
Mineola, New York
Copyright
English translation copyright © 1975 by The Massachusetts Institute of Technology.
All rights reserved.
Bibliographical Note
This Dover edition, first published in 2007, is an unabridged republication of the work published in 1975 by The MIT Press, Cambridge, Massachusetts. That work was originally published in 1965 by Nauka Press, Moscow, under the title Statisicheskaya gridomekhanika-Mekhanika Turbulenosti, by Andrey S. Monin and Akiva M. Yaglom. Translated from the Russian by Scripta Graphica, Inc. This Dover edition is published by special arrangement with The MIT Press, 55 Hayward Street, Cambridge, MA 02142.
Library of Congress Cataloging-in-Publication Data
Monin, A. S.
Statistical fluid mechanics : mechanics of turbulence / A.S. Monin an A.M. Yaglom ; edited by John Lumley.
p. cm.
Originally published: Cambridge, Mass. : MIT Press, c 1971–1975.
Includes bibliographical references and index.
ISBN 0-486-45891-1 (pbk.)
1. Hydrodynamics. 2. Fluid mechanics. 3. Turbulence. I. Title.
QA911 .M6313 2007
532'.0527—dc22
2006053473
Manufactured in the United States by Courier Corporation
45891102
www.doverpublications.com
AUTHORS’ PREFACE
TO THE ENGLISH EDITION
We are very happy that this book, the preparation of which has taken so many years of our lives, is now available to English readers in its entirety. We have been very encouraged by the warm reception accorded to Volume 1 (judging from the reviews that we have seen), although it is not easy for readers to form an opinion on the basis of the first half, when the two volumes were prepared for use as a whole. As in Volume 1, we have done our best to make the book up-to-date by adding new material and figures (particularly in Chapter 8) amounting to nearly ten percent of the total, and more than 250 references to works which appeared after the publication of the Russian edition. We have also added supplementary remarks and errata to Volume 1. We understand that the incorporation of the additions in the main body of the book has made the work of our friendly editor, J. L. Lumley, much more difficult, and we are grateful to him for his attempts to make the book better.
A. S. Monin
A. M. Yaglom
EDITOR’S PREFACE
TO THE ENGLISH EDITION
As with Volume 1, the translated manuscript was sent to A. M. Yaglom, who revised both content and style extensively. The manuscript was then returned to me for editing. Afterthoughts, however, continued to arrive as postscripts to letters until at last one manuscript page had nearly doubled in length by five successive accretions. A halt was finally declared at the end of May, 1973. I have found the seven years’ interaction necessary to prepare these volumes for publication in English a thoroughly enjoyable and rewarding experience.
J. L. Lumley
CONTENTS
Authors’ Preface to the English Edition
Editor’s Preface to the English Edition
Chapter 6 Mathematical Description of Turbulence. Spectral Functions
11. Spectral Representations of Stationary Processes and Homogeneous Fields
11.1 Spectral Representation of Stationary Processes
11.2 Spectral Representation of Homogeneous Fields
11.3 Partial Derivatives of Homogeneous Fields. Divergence and Curl of a Vector Field
12. Isotropic Random Fields
12.1 Correlation Functions and Spectra of Scalar Isotropic Fields
12.2 Correlation Functions and Spectra of Isotropic Fields
12.3 Solenoidal and Potential Isotropic Vector Fields
12.4 One-Point and Two-Point Higher-Order Moments of Isotropic Fields
12.5 Three-Point Moments of Isotropic Fields
13. Locally Homogeneous and Locally Isotropic Random Fields
13.1 Processes with Stationary Increments
13.2 Locally Homogeneous Fields
13.3 Locally Isotropic Fields
Chapter 7 Isotropic Turbulence
14. Equations for the Correlation and Spectral Functions of Isotropic Turbulence
14.1 Definition of Isotropic Turbulence and the Possibilities of its Experimental Realization
14.2 Equations for the Velocity Correlations
14.3 Equations for the Velocity Spectra
14.4 Correlations and Spectra Containing Pressure
14.5 Correlations and Spectra Containing the Temperature
15. The Simplest Consequences of the Correlation and Spectral Equations
15.1 Balance Equations for Energy, Vorticity, and Temperature-Fluctuation Intensity
15.2 The Loitsyanskii and Corrsin Integrals
15.3 Final Period of Decay of Isotropic Turbulence
15.4 Experimental Data on the Final Period of Decay. The Decay of Homogeneous Turbulence
15.5 Asymptotic Behavior of the Correlations and Spectra of Homogeneous Turbulence in the Range of Large Length Scales (or Small Wave Numbers)
15.6 The Influence of the Spectrum Singularity on the Final Period Decay
16. Self-Preservation Hypotheses
16.1 The von Kármán Hypothesis on the Self-Preservation of the Velocity Correlation Functions
16.2 Less Stringent Forms of the von Kármán Hypothesis
16.3 Spectral Formulation of the Self-Preservation Hypotheses
16.4 Experimental Verification of the Self-Preservation Hypotheses
16.5 The Kolmogorov Hypotheses on Small-Scale Self-Preservation at High Enough Reynolds Numbers
16.6 Conditions for the Existence of Kolmogorov Self-Preservation in Grid Turbulence
16.7 The Meso-Scale Quasi-Equilibrium Hypothesis. Self-Preservation of Temperature Fluctuations
17. Spectral Energy-Transfer Hypotheses
17.1 Approximate Formulas for the Spectral Energy Transfer
17.2 Application of the Energy Transfer Hypotheses to the Study of the Shape of the Spectrum in the Quasi-Equilibrium Range
17.3 Application of the Energy-Transfer Hypotheses to Decaying Turbulence behind a Grid
17.4 Self-Preserving Solutions of the Approximate Equations for the Energy Spectrum
18. The Millionsfychikov Zero-Fourth-Cumulant Hypothesis and its Application to the Investigation of Pressure and Acceleration Fluctuations
18.1 The Zero-Fourth-Cumulant Hypothesis and the Data on Velocity Probability Distributions
18.2 Calculation of the Pressure Correlation and Spectra
18.3 Estimation of the Turbulent Acceleration Fluctuations
19. Dynamic Equations for the Higher-Order Moments and the Closure Problem
19.1 Equations for the Third-Order Moments of Flow Variables
19.2 Closure of the Moment Equations by the Moment Discard Assumption
19.3 Closure of the Second- and Third-Order Moment Equations Using the Millionshchikov Zero-Fourth-Cumulant Hypothesis
19.4 Zero-Fourth-Cumulant Approximation for Temperature Fluctuations in Isotropic Turbulence
19.5 Space-Time Correlation Functions. The Case of Stationary Isotropic Turbulence
19.6 Application of Perturbation Theory and the Diagram Technique
19.7 Equations for the Finite-Dimensional Probability Distributions of Velocities
20. Turbulence in Compressible Fluids
20.1 Invariants of Isotropic Compressible Turbulence
20.2 Linear Theory; Final Period of Decay of Compressible Turbulence
20.3 Quadratic Effects; Generation of Sound by Turbulence
Chapter 8 Locally Isotropic Turbulence
21. General Description of the Small-Scale Structure of Turbulence at Large Reynolds Numbers
21.1 A Qualitative Scheme for Developed Turbulence
21.2 Definition of Locally Isotropic Turbulence
21.3 The Kolmogorov Similarity Hypotheses
21.4 Local Structure of the Velocity Fluctuations
21.5 Statistical Characteristics of Accele ration, Vorticity, and Pressure Fields
21.6 Local Structure of the Temperature Field for High Reynolds and Peclet Numbers
21.7 Local Characteristics of Turbulence in the Presence of Buoyancy Forces and Chemical Reactions. Effect of Thermal Stratification
22. Dynamic Theory of the Local Structure of Developed Turbulence
22.1 Equations for the Structure and Spectral Functions of Velocity and Temperature
22.2 Closure of the Dynamic Equations
22.3 Behavior of the Turbulent Energy Spectrum in the Far Dissipation Range
22.4 Behavior of the Temperature Spectrum at Very Large Wave Numbers
23. Experimental Data on the Fine Scale Structure of Developed Turbulence
23.1 Methods of Measurement; Application of Taylor’s Frozen-Turbulence Hypothesis
23.2 Verification of the Local Isotropy Assumption
23.3 Verification of the Second Kolmogorov Similarity Hypothesis for the Velocity Fluctuations
23.4 Verification of the First Kolmogorov Similarity Hypothesis for the Velocity Field
23.5 Data on the Local Structure of the Temperature and other Scalar Fields Mixed by Turbulence
23.6 Data on Turbulence Spectra in the Atmosphere beyond the Low-Frequency Limit of the Inertial Subrange
24. Diffusion in an Isotropic Turbulence
24.1 Diffusion in an Isotropic Turbulence. Statistical Characteristics of the Motion of a Fluid Particle
24.2 Statistical Characteristics of the Motion of a Pair of Fluid Particles
24.3 Relative Diffusion and Richardson’s Four-Thirds Law
24.4 Hypotheses on the Probability Distributions of Local Diffusion Characteristics
24.5 Material Line and Surface Stretching in Turbulent Flows
25. Refined Treatment of the Local Structure of Turbulence, Taking into Account Fluctuations in Dissipation Rate
25.1 General Considerations and Model Examples
25.2 Refined Similarity Hypothesis
25.3 Statistical Characteristics of the Dissipation
25.4 Refined Expressions for the Statistical Characteristics of Small-Scale Turbulence
25.5 More General Form of the Refined Similarity Hypothesis
Chapter 9 Wave Propagation Through Turbulence
26. Propagation of Electromagnetic and Sound Waves in a Turbulent Medium
26.1 Foundations of the Theory of Electromagnetic Wave Propagation in a Turbulent Medium
26.2 Sound Propagation in a Turbulent Atmosphere
26.3 Turbulent Scattering of Electromagnetic and Sound Waves
26.4 Fluctuations in the Amplitude and Phase of Electromagnetic and Sound Waves in a Turbulent Atmosphere
26.5 Strong Fluctuations of Wave Amplitude
27. Stellar Scintillation
27.1 Fluctuations in the Amplitude and Phase of Star Light Observed on the Earth’s Surface
27.2 The Effect of Telescope Averaging and Scintillation of Stellar and Planetary Images
27.3 Time Spectra of Fluctuations in the Intensity of Stellar Images in Telescopes
27.4 Chromatic Stellar Scintillation
Chapter 10 Functional Formulation of the Turbulence Problem
28. Equations for the Characteristic Functional
28.1 Equations for the Spatial Characteristic Functional of the Velocity Field
28.2 Spectral Form of the Equations for the Spatial Characteristic Functional
28.3 Equations for the Space-Time Characteristic Functional
28.4 Equations for the Characteristic Functional in the Presence of External Forces
29. Methods of Solving the Equations for the Characteristic Functional
29.1 Use of a Functional Power Series
29.2 Zero-Order Approximation in the Reynolds Number
29.3 Expansion in Powers of the Reynolds Number
29.4 Other Expansion Schemes
29.5 Use of Functional Integrals
Bibliography
Supplementary Remarks to Volume 1
References
Errata to Volume 1
Author Index
Subject Index
6 MATHEMATICAL DESCRIPTION OF TURBULENCE. SPECTRAL FUNCTIONS
11. SPECTRAL REPRESENTATIONS OF STATIONARY PROCESSES AND HOMOGENEOUS FIELDS
11.1 Spectral Representation of Stationary Processes
The general concept of the random field was discussed in Chapt. 2 (see Vol. 1) where the main statistical characteristics of such fields, i.e., the mean values and the correlation functions of various orders, were introduced. This was sufficient for discussion of the simplest properties of turbulent flows in Chapts. 3–5. However, when we consider the finer properties of turbulence, we find that this requires a number of new mathematical ideas, and these will in fact be developed in the present chapter.
This chapter will be mainly concerned with the application of harmonic analysis to random functions of one or more variables, i.e., random processes and random fields. Harmonic analysis, i.e., the representation of functions by Fourier series or integrals, is very widely used in mathematical physics. It must be remembered, however, that representation by Fourier series is possible only for periodic functions, while representation by Fourier integrals is possible only for functions which vanish sufficiently rapidly at infinity. In applications, however, we frequently encounter nonperiodic functions which do not vanish at infinity and which, strictly speaking, cannot be represented by Fourier series or integrals. From this point of view, random functions turn out to have certain advantages as compared with ordinary, i.e., nonrandom, functions. The point is that a Fourier expansion (or, in other words, spectral representation) of a special form, and with a clear physical interpretation, is possible for any stationary random process and homogeneous random fields which, by definition, do not vanish at infinity.¹
Let us begin with a consideration of the case of a stationary process u(t). We shall suppose that u(t) = 0 [since, otherwise, we need only consider the process u′(t) = u(t) — u(t)) instead of u(t)]. The idea of a spectral expansion of the process u(t) is most readily understood by considering the following example. Let u(t) be a random function (in general complex)² of the form
where ω1, . . ., ωn are given numbers and Z1 . . ., Zn are complex random variables with the following properties:
(here and henceforth an asterisk will denote the complex conjugate). The correlation function of a complex process u(t) is conveniently defined by
(for real processes, this definition is of course the same as that given in Chapt. 2). From (11.1) and (11.2) we then have
We see that the correlation function depends only on t2 — t1, as should be the case for a stationary random process. Under certain additional conditions imposed on the variables Zk (which are automatically fulfilled when the multidimensional probability distributions for the variables Re Zk and Im Zk are Gaussian), all the higher-order moments and finite-dimensional probability distributions of the values of u(t) will also depend only on the time differences, i.e., the process u(t) will be stationary. Equation (11.1) will, in this case, define the spectral representation of this stationary process.³
The process defined by Eq. (11.1) will be real if, and only if, the number n = 2m is even, and when the 2m terms on the right-hand side of Eq. (11.1) split into m When this condition is satisfied, Eq. (11.1) may be rewritten in the form
where
and consequently
Hence it is clear that a real process of the form Eq. (11.1) is a superposition of uncorrelated harmonic oscillations with random amplitudes and phases. The correlation function for the process Eq. (11.5) is given by
It depends on the mean squares of the amplitudes Wk, but is independent of the statistical characteristics of the phases φk.
It can be shown that an arbitrary stationary random process u(t) has a spectral representation which is a direct generalization of Eq. (11.1) . However, in general, we must take the limiting form of Eq. (11.1) as n → ∞ by assuming that the frequencies ω1, . . ., ωn can approach each other without limit, and their number in a given interval along the ω axis can be infinite (although the sum of the complex amplitudes Zk is finite). Let the sum of the amplitudes Zk corresponding to frequencies ωk < ω be denoted by Z (ω). The function Z (ω) will be a random complex function such that Z (ω) ≡ 0 if u (t) = 0, and
[because of the properties of the amplitudes Zk defined by Eq. (11.2)]. Equation (11.8) can also be formally written down in the more convenient differential form
The fact that the process u(t) is obtained from Eq. (11.1) as n Zk → Z(ω) means that
where — Ω = ω0 < ω1 < . . . < ωn However, the right-hand side of Eq. (11.9) is equal to the improper Stieltjes integral (evaluated between — ∞ and ∞), so that we can write Eq. (11.9) in the symbolic form
This is, in fact, the general spectral representation of a stationary process u(t) first derived by Kolmogorov (1940a) and Cramer (1942) [see also Doob (1953), Yaglom (1962), and Rozanov (1967)]. In view of Eq. (11.9), the significance of this expansion lies in the possibility that any stationary random process u(t) can be replaced, to any required approximation, by the sum of uncorrelated harmonic oscillations with random amplitudes and phases.⁴
From Eq. (11.10) we can obtain the following inversion formula
which enables us to express Z (ω)in terms of u(t):
so that
It is clear from Eq. (11.11) that in the case of a Gaussian process u(t), the probability distributions of the values of Z (ω) will also be Gaussian. If the process u(t) is real (henceforth we shall consider only real processes), then it is clear that Z (— ω1)—Z(—ω2) = Z* (ω2)—Z* (ω1) for any ω2 > ω1 ≥ 0, i.e., dZ (— ω) = dZ* (ω). The spectral representation (11.10) can then be rewritten in a real form similar to Eq. (11.5).
The real physical significance of the spectral representation Eq. (11.10) can be seen from the fact that the spectral components of the process, corresponding in this representation to different parts of the frequency spectrum, can be isolated experimentally by means of suitably chosen filters. Filters are devices which transmit harmonic oscillations in a given frequency range, but reject oscillations corresponding to other frequencies. When a real stationary process described by Eq. (11.10) is acted upon by a filter with a bandwidth Δω = [ω1, ω2], the signal at the output of the filter is given by
where Re denotes the real part. This expression gives the spectral component of the process u(t) corresponding to the frequency interval Δω. The condition given by Eq. (11.8) [or Eq. (11.8′)] shows that spectral components corresponding to nonoverlapping frequency intervals are uncorrelated with each other. It is readily shown that, for a sufficiently narrow frequency range Δω, and a moderate time interval, the spectral component u(Δω, t) can be replaced with a high degree of accuracy by the harmonic oscillation 2ReZ (Δω) eiωt, although in the general case of a continuous spectrum, i.e., a process which cannot be represented by a finite or infinite sum of the form (11.1), the components u(Δω, t) will not be strictly periodic for any Δω.
The fact that the spectral components u (Δω, t) can be isolated experimentally gives a real significance to the frequency distribution of the (mean) energy of the process u(t). In physical applications, the energy of a process u(t) is usually proportional to [u (t)]² (for example, if u(t) is the velocity, then [u (t)]² is proportional to the kinetic energy). It follows that, for stationary random functions u(t), the quantity [u (t)]² = B (0) plays the role of the mean energy. Using Eqs. (11.12) and (11.8) we can readily show that the mean energy of a spectral component u (Δω, t) i.e., the mean energy corresponding to the harmonic oscillations with frequencies in the range Δω = [ω1, ω2] ), is given by
where
describes the distribution of the energy of the process u(tω < ∞.
The energy distribution over the spectrum for a turbulent flow is always continuous.is therefore represented by the integral of the spectral density function (or the spectrum) F (ω) of the process u(t):
Instead of the spectrum F (ω) defined for — ∞ < ω < ∞ (and even with respect to ω), it is frequently more convenient to consider the spectrum E (ω) = 2F ω < ∞. According to Eq. (11.14) we have
In view of Eqs. (11.15) and (11.8′), the evaluation of the mean values of double integrals with respect to Z*(ω) and Z(ω1) yields the correct result if we use the symbolic equation
where δ (ω) is the improper Dirac δ-function. This symbolic equation is equivalent to the rigorous relationship dZ*(ω)dZ(ω1) = χ(ω – ω1 )F(ω)dω, where χ(x) = 1 when x = 0, and χ(x) = 0 when x ≠ 0.
We can readily show from is the Fourier transform of the corresponding spectral density:
so that
The special case of Eq. (11.17) corresponding to τ = 0, i.e.,
has a particularly simple physical interpretation. It shows that the total energy of the process u(t) is the sum of the energies of the individual spectral components.
It is clear from Eqs. (11.14) and (11.18) that the Fourier transform of the correlation function of a stationary process should be a nonnegative function. This constitutes Khinchin’s theorem on the Fourier expansion of correlation functions [Khinchin (1934)]. Khinchin also showed that each function which has a nonnegative Fourier transform is the correlation function of some stationary random process. Therefore, to verify whether a given function is the correlation function of a stationary random process, we must find its Fourier transform and establish whether or not it is always nonnegative. For example, it follows that the functions
which correspond to the Fourier transforms
can all be correlation functions. In Eq. (11.20IV) Kv is the so-called Basset’s (or Macdonald’s) function (modified Bessel function of the third kind), and Γ is the Γ-function. At the same time the function equal to C (1 — α²τ²) for | τ 1/α and 0 for| τ | > 1/α cannot be a correlation function for any stationary random process, since it can be readily verified that its Fourier transform is not nonnegative for all ω.
The equations (11.17) and (11.18) can be verified directly by experiment, since the function B (τ) can usually be determined from a single measured realization of the process by means of time averaging (see Vol. 1, Sect. 4.7), and the function E (ω) can be measured independently with the aid of a set of narrow-band filters with different bandwidths [using Eqs. (11.13) and (11.15)]. Sup-pose, for example, that the process u(t) is realized in the form of a fluctuating voltage [if u(t) is a velocity or temperature fluctuation at a point in a turbulent flow, their transformation into voltage fluctuations is usually achieved automatically by the measuring instruments; see Vol. 1, Sect. 8.3]. Let us apply the signal u(tE (ω) dω (the time averaging in the formulas given by Eqs. (11.13)–(11.15) is, as a rule, performed directly by the watt meter which has a certain inertia). By varying the value of ω0 and differentiating the resulting empirical curve, we can find the function E (ω) which is necessary for the verification of the formulas given by Eqs. (11.17) and (11.18).
This type of verification of Eqs. (11.17) and (11.18) was first carried out by Taylor (1938b).⁶ For the function u(t), Taylor used the longitudinal velocity fluctuations at a fixed point in grid-generated turbulence in a wind tunnel. The corresponding spectral density E (ω) was measured by Simmons and Salter (1938) with the aid of a special set of electric filters (Fig. 1). Since no direct data were available at the time on the function B (τ), Taylor used measurements of the space correlation function B (x), i.e., the mean product of the simultaneous values of the longitudinal velocity components at two points separated by a distance x (in the direction of the wind-tunnel axis). Since the mean velocity u = U in the wind-tunnel was much greater than the fluctuating velocity u′ = u — U, Taylor assumed that the disturbances (eddies) were transported with the mean velocity without appreciable distortion. Hence, it follows that that
FIG. 1 Spectral density of the longitudinal velocity component behind a grid in a wind tunnel.
The assumption that this equation is valid has since been known as the Taylor hypothesis. In Fig. 2, which is taken from Taylor’s paper (1938b), the full points represent direct measurements of the function B (x), while crosses show the values of this function calculated from Eqs. (11.17) and (11.22) using the data of Fig. 1. As can be seen, the agreement is completely satisfactory.
Similar results have been frequently reported by many subsequent workers. Recent developments in electronics have permitted the direct determination of the time correlation functions B (τ)⁷ without the use of Taylor’s hypothesis. Such measurements can, of course, be used to verify directly the Taylor hypothesis [see, for example Favre et al. (1962), Frenkiel and Klebanoff (1966) or Comte-Bellot and Corrsin (1971)]. As an example, Fig. 3, which is taken from the paper by Favre et al. (1954) shows the results of filter measurements of the spectral density E (ω) of longitudinal velocity fluctuations in a wind tunnel, and also the values of E (ω) calculated from Eq. (11.18) using experimental data on the function B (τ). Here again the agreement is completely satisfactory.
FIG. 2 Comparison of measured values of the velocity correlation function (solid points) with calculations based on Eqs. (11.17) and (11.22) (circles with crosses).
FIG. 3 Comparison of the measured time velocity spectrum (solid points) with values calculated from Eq. (11.18) (open circles).
Let us now briefly consider the spectral representation of the derivatives of a stationary process u(t). If we define the derivative of u′(t) of a random function as the mean square limit of the corresponding finite differences, i.e.,
then from the spectral representation (11.10) we can readily show that
We note, however, that since the left-hand side of Eq. (11.23) contains the statistical mean, the derivative given by Eq. (11.24) characterizes the entire statistical ensemble of the realizations of the process u(t), and not each realization individually. It follows that, even if all the individual realizations of our process were smooth differentiable functions, but the statistical variance of the realization derivative at the point t were infinite, the derivative defined by Eq. (11.23) would not exist. On the other hand, if only the mean square derivative
(11.24) of the stationary process u(t) is finite, it can be shown that this implies that all the realizations of this process will be differentiable, and the derivative of the realization can then be regarded as coinciding with the realization of the random process (11.24) [see, for example, Doob (1953), Chapt. 11, Sect. 9, where a more rigorous formulation of this statement can be found]. Hence, processes for which the derivative (11.24) is finite can be referred to as differentiable processes.
Equation (11.24) defines the spectral representation of the derivative u′(t). It shows that the spectrum of the derivative is equal to ω²E (ω) [see Eq. (11.16)]. Of course, this spectrum is possible only if the function E (ω) vanishes at infinity sufficiently rapidly to ensure that
If this is not so, the process u(t) will not be differentiable in the above sense. In view of Eq. (11.17), the left-hand side of Eq. (11.25) is equal to B″ (0), and the condition (11.25) itself is equivalent to the condition for the existence of a finite second derivative of the function B (τ). The correlation function for the process (11.24) is given by
Similarly, for the nth derivative, we have
and
These equations are meaningful provided
or provided the function B (τ) is 2n-fold differentiable at the point τ = 0. In particular, the process u(t) will have derivatives of all orders if the integral in Eq. (11.29) converges for any n, i.e., the spectral density should fall to zero more rapidly than any finite negative power of | ω | as | ω | → ∞ (and the function B (τ) should have derivatives of all orders at the point τ = 0).
Since the correlation function and the spectrum are mutual Fourier transforms, all the formal relationships between these two functions can be inverted. In particular, the second equation in Eq. (11.28) corresponds to
which shows that the function E (ω) will have at least 2n derivatives if the correlation function B (τ) decreases at infinity more rapidly than | τ |–2n–1, so that
Hence, it is clear that the spectrum E (ω) will be infinitely differentiable provided only the correlation function B (τ) decreases at infinity more rapidly than any negative power of | τ |.
Suppose now that u(t) = {u1 (t), . . ., un (t)} is a multidimensional (vector) real random stationary process such that uk (t) = 0, k = 1, . . ., n. We shall assume that the correlation functions Bkl (τ) = uk (t) ul (t + τ) fall to zero at infinity sufficiently rapidly, so that they can be represented by the Fourier integrals
where
It is clear that the functions Fkk (ω), k = 1, . . ., n are the ordinary nonnegative and even (with respect to ω) spectra of the onedimensional processes uk (t), k = 1, . . ., n. As regards the functions Fkl (ω) with k ≠ l, i.e., the cross-spectral densities (or simply cross-spectra) of the processes uk (t) and ul (t), these functions can in general be complex.⁸ However, since the functions Bkl (τ) are real, we always have
Next, the obvious relationship
shows that
so that the matrix || Fkl (ω) || is Hermitian for all ω. Moreover, for any complex c1, . . ., cn and any ω we have
so that the matrix || Fkl (ω) || is nonnegative-definite for any ω.⁹ Cramer (1940) and Kolmogorov (1941e) have also shown that any matrix || Fkl (ω) || which satisfies the above conditions will be the spectral matrix, i.e., the matrix of the cross spectra, of a multidimensional stationary process.
The spectral representation of a multidimensional stationary process u (t) is of the form
where Z (ω)= {Z1 (ω), . . ., Zn (ω)} and the functions Z1 (ω), . . ., Zn (ω) are such that
[when k = l, Eq. (11.39) is obviously identical with Eq. (11.16)]. It is is clear that Eq. (11.39) leads directly to Eqs. (11.36) and (11.37).
11.2 Spectral Representation of Homogeneous Fields
The above discussion of the spectral representation of stationary random processes u (t) can be extended to homogeneous random fields u (x) (except for the experimental verification with the aid of filters). In this case, the harmonic oscillations eiωt are replaced by the plane waves eikx, and the random function Z (Δω) = Z ([ω1, ω2]) = Z (ω2) — Z (ω1) of the frequency interval Δω is replaced by the random function Z (Δk) of the multidimensional interval Δk = [k′, k″] in the space of the wave-number vectors k. Let Z (dk) = Z ([k, k + dk]) = dZ (k); then we can write down the spectral representation of a homogeneous random field u (x) in the form
where the integral is evaluated over the entire wave-number vector space and has the same meaning as the integral in Eq. (11.10). We shall suppose that the field u (x) is real, has a zero mean, and is such that
where B (r) = u (x) u (x + r) (the last condition ensures the continuity of the energy distribution of the field u (x) in the wave-number vector space). The random amplitudes dZ (k) of the representation (11.40) will then have the following properties:
where δ(k) is the multidimensional δ-function. When k = k1, the equation given by Eq. (and hence it is clear that F (—k) = F (k). The function F (k) will be called the three-dimensional spectral density (or simply the three-dimensional spectrum¹⁰) of the field u (x).¹¹ The formulas given by Eqs. (11.40) and (11.44) show that each homogeneous field can be approximated as closely as desired [in the sense explained in the discussion following Eq. (11.10)] by a finite sum of uncorrelated plane waves with different wavelengths and orientations and random amplitudes and phases. The complex amplitudes dZ (k) of these plane waves can be found from the field u (x) with the aid of the inversion formulas
where the integrals between infinite limits are to be understood as the limits (mean square limits, in fact) of the integrals between — A and + A as A → ∞ [see Eq. (11.11′)].
It follows from Eqs. (11.40) and (11.44) that
We note that the possibility of representation of the function B (r) by a Fourier integral of the form (must be nonnegative. The converse of this is also true: Any function B (r) which has a nonnegative Fourier transform, is the correlation function of some homogeneous random field [see, for example, Kampé de Fériet (1953) or Yaglom (1957, 1962)].
It follows from Eq. (11.46) that
Hence, by specifying the three-dimensional spectrum F (k) we simultaneously specify the correlation function B (r). Nevertheless, some workers use, instead of F (k), the less complete statistical characteristic
where dS (k) is an area element on the sphere | k |=k. This function depends on the single argument k (instead of the three arguments k1, k2, k3). The function E (k) will be referred to simply as the spectrum (without the adjective three-dimensional) of the field u (x). The spectrum E (k) does not uniquely define the function B (r) since the energy E (k) dk of plane waves with wave-number vectors in the range (k, k + dkis simply expressed in terms of E (k):
For any fixed value k = k0 we can split B (0) into two parts, namely,
This corresponds to the division of the field u (x) into two uncorrelated parts, namely, the macrocomponent u(k0, x) (set of disturbances with wavelengths greater than 2π/k0), and the microcomponent u(k0, x) (set of disturbances with wavelengths smaller than 2π/k0):
This is the analog of the usual division of an arbitrary homogeneous field into the mean value u (x) and a fluctuation u (x) — u (x) = u′ (x). The latter division can be regarded as a special case of the general relation (11.51) corresponding to k0 = 0.
In the case of a multidimensional (vector) homogeneous random field u (x) = {u1 (x), . . ., un(x)} with uj(x)=0, j= 1, . . ., n, the formula (11.40) will be valid for each component uj (x), so that
The quantities dZj (k), j = 1, . . ., will then satisfy Eqs. (11.42) and (11.43), and the condition (11.44) in the case of a field with sufficiently rapidly decreasing correlation functions Bjl (r) = uj (x) ul (x + r) is replaced by the following more general condition:
From equations (11.52) and (11.53) it follows that
and consequently
Since Bjl(r) are real functions, and Bjl (r) = Blj (–r), the three-dimensional cross spectra Fjl (k) of the fields uj (x) and ul (x) satisfy the conditions
Hence it follows that the spectral matrix
is Hermitian for any k. Moreover, it follows from equation (11.53) that the matrix (11.57) is nonnegative-definite for any k, i.e., it is such that
for any c1, . . ., cn. The converse statement is also true: Each matrix || Fjl (k) || which satisfies Eq. (11.56) and is nonnegative-definite for all k, is the spectral matrix of a multidimensional homogeneous random field [see, for example, Kampé de Fériet (1953) or Yaglom (1957, 1962)].
In the important special case when u (x) is the velocity field of a fluid flow, the quantity
is the mean kinetic energy per unit mass of the fluid. Therefore, in addition to the spectral densities Fjl (k), it is convenient to introduce the scalar spectral densities
which are such that
If we represent each component of the vector u (x) in the form of the sum (11.51), the vector field u (x) will be divided into a macrocomponent u (k0, x), with energy equal to
and a microcomponent u (k0, x), with energy equal to
which will be useful in our subsequent discussion:
differ from the corresponding fourth-order cumulants by terms which also tend to zero when | r | → ∞ or | r′ etc., on the right-hand sides of Eqs. (11.60)–(11.60VI) uniquely determine the corresponding moments (or combinations of moments) of which they are the Fourier transforms. We shall refer to them as the higher-order spectra (or spectral tensors) of the homogeneous fields u (x) the first four higher-order spectra satisfy the following relations:
Moreover, it is clear that the tensors Fij, Fij,l,mare symmetric in subscripts i, j, and that all the higher-order spectra transform to the complex-conjugate quantities when the signs of their arguments are reversed (the moments are real). They also satisfy the following simple relationships:
since
11.3 Partial Derivatives of Homogeneous Fields. Divergence and Curl of A Vector Field
As in the case of stationary random processes, the number of times one can differentiate the spectral density of a homogeneous random field F (k), increases with the rate at which the correlation function B (r) of this field falls to zero at infinity. Conversely, the rate at which the spectrum F (k) decreases to zero at infinity governs the smoothness of the correlation function B (r). Both these statements are simple consequences of the fact that the functions B (r) and F (k) are mutual Fourier transforms, so that
and
It follows from the last formula that if for | r | → ∞ the function B (r) falls to zero as r–N where N is an integer [or, in other words, if B (r) = O (r–N)], then the spectrum F (k) will be (N – 4)-fold differentiable with respect to its arguments for all values of k, but its partial derivatives of order N – 3 will exist only for k ≠ 0 but not at k = 0 [the function B (r) behaves in an analogous fashion if F (k) = O (k–N) as | k | = k → ∞)]. Since the integral on the right-hand side of Eq. (11.63) diverges logarithmically when B (r) = O (r–N), m1 + m2 + m3 = N – 3 and k = 0, and the function e–ikr (where C is a small constant), it is clear that, if B (r) = O (r–N), the partial derivates of order N – 3 of the function F (k) (which are continuous functions of k for k ≠ 0) will in general increase logarithmically as k → 0. The situation is precisely the same for the more general functions Bjl (r) and Fjl (k), which correspond to the multidimensional homogeneous field u (x) = {u1 (x), . . ., un (x)}. Therefore, for example, the general Taylor expansion of the function Fjl (k) in the neighborhood of k = 0, which is of the form
will for B (r) = O (r–N) contain only terms up to order N – 4 inclusive, and these will be followed by a residual term of the order of kN–4 ln k (when j = l, all the odd terms in Eq. (11.64) will, of course, be equal to zero). On the other hand, to insure that the spectrum Fjl(k) will be infinitely differentiable, the correlation function Bjl (k) should decrease to zero at infinity more rapidly than any finite negative power of | r |. Similarly, to ensure infinite differentiability of the function Bjl (r), we must have the appropriate rate of decrease to zero at infinity for the spectrum Fjl (k).
Let us suppose that
When this is so, the partial derivative
will exist in the sense analogous to Eq. (11.23), and so will all the lower-order partial derivatives. Moreover, in view of Eq. (11.66),
Similarly, if
(no summation over j and l), the derivatives
will exist, and
Of all the multidimensional random fields, the most important case for the theory of turbulence is the three-dimensional velocity field u (x) = {u1 (x), u2 (x), u3 (x)} . In view of Eq. (11.66),
and, hence,
and εjlm is a unit tensor of rank 3 which is antisymmetric in all its indices. Consequently,
It follows that the correlation function of the divergence of the field u (x) Similarly, the correlation tensor of the vorticity of the field u (x) The formulas for the correlation and spectral tensors of the vorticity field ω (x) = curl u (x) can be transformed with the aid of the following identity
which is readily verified. Moreover, it follows from Eq. (11.69) that
(no summation over j), and
We shall now restrict our attention to the case where the vector field u (x) is solenoidal, i.e., duj (x)/∂xj = 0. We then have
In view of Eq. (11.70) this means that
The four relationships in Eqs. (11.79) and (11.80) are not, of course, independent: From any one of them the other three follow automatically. Instead of Eqs. (11.79) and (11.80) we can also use the equivalent relationship kj dZj (k) = 0.
Substituting the identity given by Eq. (11.75) into Eq. (11.74), and using Eqs. (11.80) and (11.79), we can verify that for a solenoidal field u (x) the correlation and spectral tensors of the vorticity field are given by
and
The relationships given by Eq. (11.79) enable us to determine the general form of the tensor Fjl (k) [see Kampé de Fériet (1948)]. In view of Eq. (11.79), the vector k is the eigenvector of the Hermitian matrix || Fjl (k) || corresponding to its zero eigenvalue. Moreover, this matrix should have two further (in general, complex) eigenvectors, a(1) and b(1), which are orthogonal to each other and to the vector k and correspond to nonnegative eigenvalues λ1 (k) and λ2 (k). Let us normalize the vectors a(1) and b(1) and expand dZ (k) in terms of the vectors a(1), b(1), and k. Since k dZ (k) = 0, we have
where
and hence
by definition of a(1) and b(1)we have from Eqs. (11.84) and (11.85)
This is, in fact, the general form of the tensor Fjl (k) [since the tensor (This tensor depends on six complex functions of the argument k, Since, moreover, a/a, b/b, k/k [where a = | a |, b = | b | is an orthonormal triad of vectors, and consequently
from Eq. (11.86) and rewrite the expression for Fjl (k) in the form
The functions b² (k), a² (k), aj (k) are here related by
Using the conditions given by Eq. (11.79), we can establish another general representation of the tensor Fjl (k) which is occasionally quite convenient. Namely, let us consider the tensor
This tensor is symmetric and orthogonal to the vector k, i.e., it satisfies the condition kjΔjm (k) = 0. Moreover, if kmFm (k) = 0 then, clearly, Δjm (k) Fm (k) = Fj (k). Therefore, the transformation with the matrix || Δjm (k) || isolates the part of the vector F (k) which is orthogonal to the vector k. Therefore, for example, the tensor Fjl (k) which satisfies Eq. (11.79) can be always written in the form
where Φmn (k) is a symmetric tensor. We can take the original tensor Fmn (k) for Φmn (k). It is possible, however, to choose Φmn (k) to be an arbitrary Hermitian symmetric tensor with a nonnegative-definite matrix, in which case the tensor (11.89) will still satisfy all the requirements imposed on the spectral tensor of a solenoidal vector field.
If u (x) (x) is a homogeneous scalar field which is homogeneously correlated with u (x), then
i.e.,
and
According to is orthogonal to k, and consequently it can be represented in the form
The higher-order spectral tensors (or vectors) of a solenoidal field, [for example, Fij, l (k), Fijl (k, k′), can be represented in a similar way. In fact, if ∂uj/∂xj = 0, then
so that
Consequently,
The above results lead us to certain conclusions with regard to the behavior of the spectra of the field u (x) in the neighborhood of the origin of the wave vector space. Let us consider, for example, the tensor Fjl (k) and suppose that it can be expanded in the Taylor series Eq. (11.64) in the neighborhood of the point k = 0. In that case, the functions aj (k) and bj (k) in Eq. (11.86) can also be expanded in a Taylor series. Since, however, the vectors a (k) and b (k) should be orthogonal to all the vectors k/| k | for k = 0 (i.e., they should be equal to zero), it is clear that the expansion of Fjl (k) in a Taylor series should begin with second order terms:
begins with first order terms:
If we pass from the spectra to the correlation matrices, we have from Eq. (11.95),
and from Eq. (11.96) we have
12. ISOTROPIC RANDOM FIELDS
12.1 Correlation Functions and Spectra of Scalar Isotropic Fields
The spectra of homogeneous random fields will, in general, depend on the three variables k1, k2, k3 (while the higher-order spectra will depend on a still greater number of variables). In the subsequent analysis we shall be particularly concerned with special homogeneous and isotropic fields for which the main statistical characteristics depend only on one variable. Such homogeneous and isotropic fields will be discussed in this section.
A scalar random field u (x) corresponding to it are unaffected by all possible rotations of the set of points x1, x2, . . ., xN about axes passing through the origin, and by mirror reflections of this set of points in any plane passing through the origin. We shall confine our attention to fields which are both homogeneous and isotropic, and for the sake of brevity we shall refer to such fields simply as isotropic. In other words, by isotropic fields we shall understand random fields u (xare unaffected by any translations (i.e., shifts), rotations, and mirror reflections (i.e., space symmetries) of the set of points x1, x2, . . ., xN.
Since the field u (x) is homogeneous, its mean value u (x) should be a constant. Without loss of generality we can assume that u (x) = 0, having replaced, if necessary, the original field u (x) by the field u′ (x) = u (x) – u (x). We shall adopt this procedure henceforth. The correlation function B (x, x′) = u (x)u (x′) of an isotropic field u (x) must, of course, assume equal values for any two pairs of points (x, x′) In other words, if the distance between the points x and x′ is equal to the distance between x1 so that the function B (x, x′) can depend only on the distance r = | x′ – x | between the points x and x′ = x + r:
We thus see that in the case of an isotropic field u (x), the correlation function does, in fact, depend on the single variable r.
If we substitute Eq. (12.1) into the general formula (11.47) which defines the three-dimensional spectrum F (k) of a homogeneous field in terms of B (r) and transform to spherical coordinates, we find that, in the isotropic case, the spectrum will depend only on k = | k |:
Conversely, if
then the correlation function B (r) will depend only on r = | r |, and in view of Eq. (11.46) we have in this case
The function F (k) in equations (12.2) and (12.4) must be nonnegative. The set of correlation functions for all the possible isotropic random fields is therefore determined by the condition that the integral (12.2) should be nonnegative for all k ≥ 0. Hence, it follows that the functions
are the correlation functions of isotropic fields. In fact, it is readily verified that the three-dimensional spectra which correspond to these functions are, respectively,
i.e., they are nonnegative.
It is clear that, if we replace the variable r by | τ | in the functions (12.5), (12.5′), and (12.5″), we obtain examples of correlation functions of stationary processes. This is explained by the fact that the values of an isotropic field u (x) at points on the straight line x2 = x3 = 0 form a homogeneous field on the straight line (i.e., a stationary process in the variable x1). The one-dimensional Fourier transform of the function B (r) which is defined for r < 0 with the aid of the condition B (– r) = B (r) must be nonnegative throughout:
The function F1 (k1) defined by Eq. (12.7) or, what amounts to the same thing, by the inverse formulas
will be called the one-dimensional spectral density (or the onedimensional spectrum) of the field u (x). Comparison of Eq. (12.8) with the special case of Eq. (11.46) obtained for r2 = r3 = 0, will readily show that
Hence it also follows that for F (k) ≥ 0, the function F1 (k1) will always be nonnegative. On the other hand, it is readily seen that, from the fact that the function F1 (k1) of Eq. (12.9) is nonnegative, it does not immediately follow that the three-dimensional spectrum
is nonnegative. In fact, if we substitute, for example, the function given by Eq. (for 2F1(k) (with the argument ω replaced by k), we obtain a function F (k) which will also assume negative values.¹² We thus conclude the functionsB (r) = C max {1 —αr, 0} B (r) = Ce–αr cos βrcan be the correlation functions of homogeneous fields on a straight line (stationary processes), but they cannot be correlation functions of isotropic fields in three-dimensional space.
Instead of the three-dimensional spectrum F (k) we can also use the spectrum
Similarly, the function F1 (k1) can be replaced by E1 (k1) = 2F1(k1) which we shall again call the one-dimensional spectrum (or onedimensional spectral density). When applied to the functions E (k) and E1 (k1), the formulas given by Eqs. (12.2), (12.4), (12.9), and (12.10) assume the form
and
As already indicated, the rate at which the correlation function B (r) decreases to zero at infinity, determines the smoothness of the spectra F (k) and E (k), while the rate at which the spectrum F (k) or E (k) decreases, governs the degree of smoothness of the correlation function B (r). Moreover, the coefficients of the Taylor expansion for F (k) or E (k) (if the expansion is possible here) can be simply expressed in terms of the moments of the function B (r), while the Taylor expansion coefficients for B (r) can be expressed in terms of the moments of the function F (k) or E (k). In fact, by expanding, for example, the function sin kr/kr in Eqs. (12.2) and (12.4) in a Taylor series, we can verify that the Taylor expansion for the functions F (k), E (k) = 4πk²F (k) and B (r) at the point k = 0, and hence r = 0, are of the form
where
It is clear from Eq. (12.17) that, in particular, the derivatives of B (r) of order 4n at the point r = 0 are nonnegative, while the derivatives of order 4n + 2 at this point are nonpositive. Next, if F (k) ~ k⁴, i.e., E ~ k⁶), as k → 0 then in view of Eqs. (12.14) and (12.16), we have
We shall use this result below.
From the formulas given by Eqs. (12.4) and (12.2) or (12.12) we also have
Therefore,
where L is a characteristic length scale of the field u (x), which is called the integral length scale and is of the order of the distance over which there is an appreciable correlation between the values of the field at two points. Another and usually much smaller characteristic length scale of the field u (x) is the differential length scale which is given by the expression
This length scale is equal to the length of the segment cut on the horizontal axis by the parabola touching the graph of the correlation function B (r) at its apex (Fig. 4).
FIG. 4 Definition of the length scale λ.
12.2 Correlation Functions and Spectra of Isotropic Vector Fields
We shall now consider multidimensional isotropic random fields. At first sight it would appear natural to define the multidimensional field u (x) = {u1 (x), . . ., un (x) } as isotropic if all the probability densities for the values of any components of this field at a given set of points in space do not change under translations, rotations, and reflections of this set of points. We shall see later that another definition of the multidimensional isotropic field is more important for the theory of turbulence, but we shall begin our discussion with multidimensional fields which are isotropic in the sense indicated above. A multidimensional field of this kind is characterized by the correlation matrix
in which all the elements depend only on r = | r |. By analogy with the derivation of Eqs. (12.2) and (12.4), it can be shown that all the elements of this matrix can be written in the form
where
are the corresponding cross spectra which, in this case, depend only on k = | k |. Since by Eq. (12.24) the spectra Fjl (k) are real, they satisfy the conditions
for any k and any real c1, . . ., cn. Conversely, any functions which can be written in the form (12.24), where Fjl (k) satisfy the condition given by Eq. (12.25), are the elements of the spectral matrix of an isotropic random field.
Examples of multidimensional random fields in the theory of turbulence are the velocity field u (x) = {u1 (x), u2 (x), u3(x)}, a combination of the velocity field and some scalar fields (for example, the five-dimensional field {u (x), p (x), (x) } where p (x) (x) is the temperature) and, finally, a combination of scalar fields (for example, the two-dimensional field {p (x), (x) }). However, the above definition of isotropic multidimensional fields is completely satisfactory only for fields of the last type. The reason is that when the points x are rotated, the components of the velocity vector u (x) undergo a linear transformation (in accordance with the general transformation rule for the components of a vector). Hence, when applied to a vector field u (x) = {u1 (x), u2 (x), u3 (x) }, the isotropy condition must be formulated as follows: A random vector field u (x) is called isotropic if the probability densities for the values of its components at an arbitrary set of points x1, x2, . . ., xN are unaffected by any translations, rotations and/or reflections accompanied by simultaneous rotation and/or reflection of the coordinate system with respect to which the components of the vector are determined. Thus, for example, the two-dimensional probability density for the values of u1 (x1) and u2 (x2) (see axis (i.e., the irelative to the rotated
set of coordinates). For a field of the form {u (x), p (x), (x) }, which in addition to the vector u (x) contains some scalar components, the isotropy condition requires that the corresponding probability densities are unaffected by translations, rotations, and reflections, which are accompanied by a linear transformation of the components of the vector u (x).
FIG. 5 Properties of the isotrophic vector field.
Let us suppose now that u (x) is an isotropic random vector field. Since the field is homogeneous, its mean value u (x) = U is a constant vector. However, since the vector is isotropic, it must be unaffected by rotations. Consequently, it must be a zero vector. It follows that the mean value of an isotropic random vector field is necessarily equal to zero.
The situation is much more complicated in the case of the correlation tensor of an isotropic vector field
FIG. 6 associated with the points M and M′.
, the first axis of which lies along the vector r, are perpendicular to this vector as shown in Fig. 6. Let the values of the components of the tensor Bjl (r) will, clearly, depend only on r = | r |, will rotate together with the points (M, Maxes, we must have
, so that
[and, consequently, Bjl (r) , three components are zero and the two remaining components are equal. Therefore, there remain only two nonequal components which we shall denote as follows:
Here, and henceforth, uL and uN will denote the projections of the vector u on to the direction of r and, correspondingly, on to any perpendicular direction. The functions BLL (r) and BNN (r) are called the longitudinal and lateral correlation