Bill George Notes

Lectures in Turbulence for the 21st Century
William K. George Professor of Turbulence Emeritus Department of Applied Mechanics Chalmers University of Technology Gothenburg, Sweden
Contents
1 The 1.1 1.2 1.3 1.4 1.5 1.6 1.7 Nature of Turbulence The turbulent world around us . What is turbulence? . . . . . . . Why study turbulence? . . . . . . The cost of our ignorance . . . . What do we really know for sure? Our personal adventure . . . . . . A brief outline . . . . . . . . . . . 9 9 11 13 14 15 16 17 19 19 20 20 21 23 24 24 26 28 28 29 29 32 33 34 34 35 38 41 41 43 45 46
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
2 The Elements of Statistical Analysis 2.1 Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 The Ensemble and Ensemble Averages . . . . . . . . . 2.2.1 The mean (or ensemble) average . . . . . . . . 2.2.2 Fluctuations about the mean . . . . . . . . . . 2.2.3 Higher moments . . . . . . . . . . . . . . . . . 2.3 Probability . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 The histogram and probability density function 2.3.2 The probability distribution . . . . . . . . . . . 2.3.3 Gaussian (or normal) distributions . . . . . . . 2.3.4 Skewness and kurtosis . . . . . . . . . . . . . . 2.4 Multivariate Random Variables . . . . . . . . . . . . . 2.4.1 Joint pdfs and joint moments . . . . . . . . . . 2.4.2 The bi-variate normal (or Gaussian) distribution 2.4.3 Statistical independence and lack of correlation 2.5 Estimation from a Finite Number of Realizations . . . 2.5.1 Estimators for averaged quantities . . . . . . . . 2.5.2 Bias and convergence of estimators . . . . . . . 2.6 Generalization to the estimator of any quantity . . . . 3 Reynolds Averaged Equations 3.1 The Equations Governing the Instantaneous Fluid 3.2 Equations for the Average Velocity . . . . . . . . 3.3 The Turbulence Problem . . . . . . . . . . . . . . 3.4 The Origins of Turbulence . . . . . . . . . . . . . 3
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
Motions . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . .
4 3.5 3.6 3.7 4 The 4.1 4.2 4.3 4.4 4.5
CONTENTS The importance of non-linearity . . . . . . . . . . . . . . . . . . . The Eddy Viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . The Reynolds Stress Equations . . . . . . . . . . . . . . . . . . . Turbulence Kinetic Energy The Kinetic Energy of the Fluctuations The Dissipation Rate . . . . . . . . . . . The Production . . . . . . . . . . . . . . The Transport (or Divergence) Terms . The Intercomponent Transfer of Energy 48 50 56 61 61 64 67 69 72 77 77 78 80 88 93 93 97 97 99 102 105 106 107 109 113 116 119 119 121 123 125 128 128 130 132 135 135 136 137 142
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
5 A First Look at Homogeneous Turbulence: 5.1 Introduction . . . . . . . . . . . . . . . . . . . 5.2 Why are homogeneous ows important? . . . 5.3 Decaying turbulence: a brief overview . . . . 5.4 A second look at simple shear ow turbulence
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
6 Turbulent Free Shear Flows 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 6.2 The averaged equations . . . . . . . . . . . . . . . . 6.2.1 The shear layer quations . . . . . . . . . . . . 6.2.2 Order of magnitude estimates . . . . . . . . . 6.2.3 The streamwise momentum equation . . . . . 6.2.4 The transverse momentum equation . . . . . 6.2.5 The free shear layer equations . . . . . . . . . 6.3 Two-dimensional Turbulent Jets . . . . . . . . . . . 6.3.1 Similarity analysis of the plane jet . . . . . . 6.3.2 Implications of the Reynolds stress equations 6.4 Other Free Shear Flows . . . . . . . . . . . . . . . . . 7 WALL-BOUNDED TURBULENT FLOWS 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 7.2 Review of laminar boundary layers . . . . . . . . . 7.3 The outer turbulent boundary layer . . . . . . . 7.4 The inner turbulent boundary layer . . . . . . . . 7.5 The viscous sublayer . . . . . . . . . . . . . . . . . 7.5.1 The linear sublayer . . . . . . . . . . . . . . 7.5.2 The sublayers of the constant stress region . 7.6 Pressure gradient boundary layers and channel ow 7.7 The inertial sublayer . . . . . . . . . . . . . . . . . 7.7.1 Some history . . . . . . . . . . . . . . . . . 7.7.2 Channel and Pipe Flows . . . . . . . . . . . 7.7.3 Boundary Layers . . . . . . . . . . . . . . . 7.7.4 Summary and conclusions . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
CONTENTS 8 Stationarity Random Processes 8.1 Processes statistically stationary in time 8.2 The autocorrelation . . . . . . . . . . . 8.3 The autocorrelation coecient . . . . . . 8.4 The integral scale . . . . . . . . . . . . 8.5 The temporal Taylor microscale . . . . . 8.6 Time averages of stationary processes . . 8.7 Bias and variability of time estimators .
5 149 149 150 152 152 152 155 157 161 161 162 163 165 167 169 170 170 171 172 173 174 175 175 176 177 179 180 181 185 185 187 190 191 193 194 199 199 203 205 206 209
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
9 Homogeneous Random Processes 9.1 Random elds of space and time . . . . . . . . . . . . . . . . 9.2 Multi-point correlations . . . . . . . . . . . . . . . . . . . . . 9.3 Spatial integral and Taylor microscales . . . . . . . . . . . . . 9.3.1 Integral scales . . . . . . . . . . . . . . . . . . . . . . 9.3.2 Taylor microscales . . . . . . . . . . . . . . . . . . . . 9.4 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5 Implications of Continuity . . . . . . . . . . . . . . . . . . . . 9.5.1 Reduction of Bi,j () to three independent components . r 9.6 Relations among the derivative moments . . . . . . . . . . . . 9.7 Elimination of the cross-derivative moments . . . . . . . . . . 9.7.1 The homogeneous dissipation . . . . . . . . . . . . . 9.7.2 Pressure uctuations in homogeneous turbulence . . . 9.8 Isotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.1 An example from uid mechanics: viscous stress . . . . 9.8.2 Isotropic single-point correlations . . . . . . . . . . . . 9.8.3 Isotropic two-point statistics . . . . . . . . . . . . . . 9.8.4 Derivative moments in isotropic turbulence . . . . . . 9.8.5 Isotropic integral scale and Taylor microscales . . . . . 9.9 Axisymmetric Turbulence . . . . . . . . . . . . . . . . . . . . 10 Decomposing turbulence 10.1 Scales of turbulence . . . . . . . . . . . . . . 10.2 The anatomy of turbulent motions . . . . . 10.3 Fields of Finite Extent . . . . . . . . . . . . 10.4 Homogeneous Fields . . . . . . . . . . . . . 10.5 Are Homogeneous Fields and Periodic Fields 10.6 Inhomogeneous elds of Innite Extent . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . the . .
. . . . . . . . . . . . . . . . Same? . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
11 Decomposing Homogeneous Turbulence 11.1 Generalized functions . . . . . . . . . . . . . . . . . 11.2 Fourier transforms of homogeneous turbulence . . 11.3 The Three-dimensional Energy Spectrum Function 11.4 DNS turbulence . . . . . . . . . . . . . . . . . . . . 11.5 One-dimensional spectra . . . . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
CONTENTS 11.6 Spectral symmetries . . . . . . . . . . . . . . . . . . . . . . . . . 211 11.7 Consequences of incompressibility . . . . . . . . . . . . . . . . . 212 11.8 Implications of Isotropy on Spectra . . . . . . . . . . . . . . . . . 213
12 Measuring spectra from experimental and DNS data 215 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 12.2 Determining the wave-number spectrum from the spatial auto-correlation216 12.3 The nite Fourier transform . . . . . . . . . . . . . . . . . . . . . 218 12.4 Taylors Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . 218 12.4.1 The Frozen Field Hypothesis . . . . . . . . . . . . . . . . . 218 12.4.2 The Eect of a Fluctuating Convection Velocity . . . . . . 218 12.5 Resolution and Spatial Filtering . . . . . . . . . . . . . . . . . . . 218 13 Dynamics of Homogeneous Turbulence 13.1 The Fourier transformed instantaneous equations . . . . 13.2 The spectral equations . . . . . . . . . . . . . . . . . . . 13.3 The eect of Reynolds number on the spectral equations 13.4 Conicting views of spectral transfer . . . . . . . . . . . 219 219 221 224 225 227 227 230 232 234 237 238 240 240 300 301 301 303 305
. . . .
. . . .
. . . .
. . . .
. . . .
14 Kolmogorov Turbulence 14.1 The universal equibrium range . . . . . . . . . . . . . . . 14.2 The inertial subrange . . . . . . . . . . . . . . . . . . . . 14.3 Scaling the energy and dissipation ranges . . . . . . . . . . 14.4 The k 5/3 and r4/3 laws . . . . . . . . . . . . . . . . . . . . 14.4.1 Dimensional and physical analysis . . . . . . . . . . 14.4.2 A useful empirical low wavenumber spectral model 14.4.3 Models for the spectrum in the univeral equilibrium 14.4.4 Deduction of k 5/3 -range from asymptotic analysis 14.4.5 The inertial range at nite Reynolds numbers . . .
. . . . . . . . . . . . . . . . . . . . . . . . range . . . . . . . .
A Signal Processing A.1 Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.2 The measurement chain . . . . . . . . . . . . . . . . . . . . . . . A.3 Analog-to-digital conversion . . . . . . . . . . . . . . . . . . . . .
B Random Processes 309 B.1 Time-averaged statistics . . . . . . . . . . . . . . . . . . . . . . . 311 B.1.1 Time mean value . . . . . . . . . . . . . . . . . . . . . . . 312 B.1.2 Higher moments . . . . . . . . . . . . . . . . . . . . . . . . 314 C Fourier analysis of time varying signals C.1 Fourier series . . . . . . . . . . . . . . C.2 Fourier transform . . . . . . . . . . . . C.3 Convolution . . . . . . . . . . . . . . . C.4 The nite Fourier transform . . . . . . 317 318 320 320 321
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
CONTENTS
C.5 The shift theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 D Digital Fourier transforms 325 D.1 Aliasing of periodically sampled data . . . . . . . . . . . . . . . . 325 D.2 The Discrete Fourier transform . . . . . . . . . . . . . . . . . . . 328 D.3 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 E Generalized functions F Spectral analysis of random signals F.1 The Fourier transform of a random signal . . . . . . . . . . . . F.2 Proof of Wiener-Khinchin Theorem using generalized functions F.3 The nite Fourier transform . . . . . . . . . . . . . . . . . . . F.3.1 An indirect method . . . . . . . . . . . . . . . . . . . F.3.2 A direct method . . . . . . . . . . . . . . . . . . . . . F.4 An example: digital spectral analysis of random data . . . . . 333 339 339 341 343 344 345 346
. . . . . .
. . . . . .
G Windows and Filters 349 G.1 Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 G.2 Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
CONTENTS
Chapter 1 The Nature of Turbulence

1.1 The turbulent world around us
The turbulent motion of uids has captured the fancy of observers of nature for most of recorded history. From howling winds to swollen oodwaters, the omnipresence of turbulence paralyzes continents and challenges our quest for authority over the world around us. But it also delights us with its unending variety of artistic forms. Subconsciously we nd ourselves observing exhaust jets on a frosty day; we are willingly hypnotized by licking ames in an open hearth. Babbling brooks and billowing clouds fascinate adult and child alike. From falling leaves to the swirls of cream in steaming coee, turbulence constantly competes for our attention. Turbulence by its handiwork immeasurably enriches the lives of even those who cannot comprehend its mysteries. Art museums are lled with artists attempts to depict turbulence in the world around us. The classic sketch of Italian renaissance artist and engineer, Leonardo da Vinci, shown in Figure 1.1 represents both art and early science. And as the tongue-in-cheek poem below by Corrsin (one of the turbulence greats of the past century) shows, even for those who try, the distinction between art and research is often dicult to make.
SONNET TO TURBULENCE
by
S. Corrsin1
(For Hans Liepmann 2 on the occasion of his 70th birthday, with apologies to Bill S. and Liz B.B.) Shall we compare you to a laminar ow? You are more lovely and more sinuous.
Stan Corrsin was a famous and much beloved turbulence researcher and professor at the Johns Hopkins University. 2 Hans Liepmann was another famous turbulence researcher and professor at Cal Tech, who was Corrsins Ph.D. dissertation advisor.
1
10
CHAPTER 1. THE NATURE OF TURBULENCE Rough winter winds shake branches free of snow, And summers plumes churn up in cumulus. How do we perceive you? Let me count the ways. A random vortex eld with strain entwined. Fractal? Big and small swirls in the maze May give us paradigms of ows to nd. Orthonormal forms non-linearly renew Intricate ows with many free degrees Or, in the latest fashion, merely few As strange attractor. In fact, we need Cray 3s3 . Experiment and theory, unforgiving; For serious searcher, fun ... and its a living!
Figure 1.1: Leonardo da Vincis observation of turbulent ow: Drawing of a free water jet issuing from a square hole into a pool (courtesy of eFluids.com). These lectures will mostly deal with the equations used to describe the mechanics of turbulence. It is only equations which can give us the hope of predicting turbulence. But your study of this subject will be missing a great deal if this is all you learn. The advantage of studying turbulence is that you truly can see it almost everywhere as it mixes and diuses, disrupts and dissipates the world around us. So teach yourself to observe the natural and manmade processes around you. Not only will your life become more interesting, but your learning will be enhanced as well. Be vigilant. Whenever possible relate what you are learning to what you see. Especially note what you do not understand, and celebrate when and if you do. Then you will nd that the study of turbulence really is fun.
3
At the time this poem was written, the Cray 2 was the worlds most powerful computer.
1.2. WHAT IS TURBULENCE?
11
1.2
What is turbulence?
Turbulence is that state of uid motion which is characterized by apparently random and chaotic three-dimensional vorticity. When turbulence is present, it usually dominates all other ow phenomena and results in increased energy dissipation, mixing, heat transfer, and drag. If there is no three-dimensional vorticity, there is no real turbulence. The reasons for this will become clear later; but briey, it is ability to generate new vorticity from old vorticity that is essential to turbulence. And only in a three-dimensional ow is the necessary stretching and turning of vorticity by the ow itself possible. For a long time scientists were not really sure in which sense turbulence is random, but they were pretty sure it was. Like anyone who is trained in physics, we believe the ows we see around us must be the solution to some set of equations which govern. (This is after all what mechanics is about writing equations to describe and predict the world around us.) But because of the nature of the turbulence, it wasnt clear whether the equations themselves had some hidden randomness, or just the solutions. And if the latter, was it something the equations did to them, or a consequence of the initial conditions? All of this began to come into focus as we learned about the behavior of strongly non-linear dynamical systems in the past few decades. Even simple nonlinear equations with deterministic solutions and prescribed initial conditions were found to exhibit chaotic and apparently random behavior. In fact, the whole new eld of chaos was born in the 1980s4 , complete with its new language of strange attractors, fractals, and Lyapunov exponents. Such studies now play a major role in analyzing dynamical systems and control, and in engineering practice as well.
Figure 1.2: Turbulence in a water jet. Photo from Dimotakis, Miake-Lye and Papantoniou, Phys. Flds., 26 (11), 3185 3192. Turbulence is not really chaos, at least in the sense of the word that the dynamical systems people use, since turbulent ows are not only time-dependent but space dependent as well. But as even the photos of simple turbulent jets
The delightful book by James Gleik Chaos: the making of a new science provides both interesting reading and a mostly factual account.
4
12
CHAPTER 1. THE NATURE OF TURBULENCE

( = smoke-wire location)
40
85
x/
185
Figure 1.3: Axisymmetric wakes from four dierent generators. Photo from S.C. Cannon, Ph.D. Dissertation., U.of Ariz, 1991. and wakes shown in Figures 1.2 and 1.3 make clear, turbulence has many features that closely resemble chaos. Obvious ones include spatial and temporal intermittency, dissipation, coherent structures, sensitive dependence of the instantaneous motions on the initial and upstream conditions, and even the near-fractal distribution of scales. In fact, the ows we see themselves bear an uncanny resemblance to the phase plane plots of strange attractors. No one would ever confuse a jet with a wake, but no two wakes seem to be quite alike either. Because of the way chaos has changed our world view, most turbulence researchers now believe the solutions of the uid mechanical equations to be deterministic. Just like the solutions of non-linear dynamical systems, we believe turbulent solutions to be determined (perhaps uniquely) by their boundary and
1.3. WHY STUDY TURBULENCE?
13
initial conditions5 . And like non-linear dynamical systems, these deterministic solutions of the non-linear uid mechanics equations exhibit behavior that appears for all intents and purposes to be random. We call such solutions turbulent, and the phenomenon turbulence. Because of this chaotic-like and apparently random behavior of turbulence, we will need statistical techniques for most of our study of turbulence. This is a course about the mechanical mysteries of turbulence. It will attempt to provide a perspective on our quest to understand it. The lack of a satisfactory understanding of turbulence presents one of the great remaining fundamental challenges to scientists and to engineers as well, since most technologically important ows are turbulent. The advances in understanding over the past few decades, together with the advent of large scale computational and experimental capabilities, present the scientist and engineer with the rst real capabilities for understanding and managing turbulent ows. As a result, this is a really wonderful time to study this subject.
1.3
Why study turbulence?
There really are the TWO reasons for studying turbulence engineering and physics! And they are not necessarily complementary, at least in the short run. Certainly a case can be made that we dont know enough about turbulence to even start to consider engineering problems. To begin with (as we shall see very quickly over the next few lectures), we always have fewer equations than unknowns in any attempt to predict anything other than the instantaneous motions. This is the famous turbulence closure problem. Of course, closure is not a problem with the so-called DNS (Direct Numerical Simulations) in which we numerically produce the instantaneous motions in a computer using the exact equations governing the uid. Unfortunately we wont be able to perform such simulations for real engineering problems until at least a few hundred generations of computers have come and gone. And this wont really help us too much, since even when we now perform a DNS simulation of a really simple ow, we are already overwhelmed by the amount of data and its apparently random behavior. This is because without some kind of theory, we have no criteria for selecting from it in a single lifetime what is important. The engineers counter argument to the scientists lament above is: airplanes must y, weather must be forecast, sewage and water management systems must be built,
If it comes as a surprise to you that we dont even know this for sure, you might be even more surprised to learn that there is a million dollar prize for the person who proves it.
5
14
CHAPTER 1. THE NATURE OF TURBULENCE society needs ever more energy-ecient hardware and gadgets.
Thus, the engineer argues, no matter the inadequate state of our knowledge, we have the responsibility as engineers to do the best we can with what we have. Who, considering the needs, could seriously argue with this? Almost incredibly some physicists do! The same argument happens in reverse as well. Engineers can become so focused on their immediate problems they too lose the big picture. The famous British aerodynamicist M. Jones captured this well when he said, A successful research enables problems which once seemed hopelessly complicated to be expressed so simply that we soon forget that they ever were problems. Thus the more successful a research, the more dicult does it become for those who use the result to appreciate the labour which has been put into it. This perhaps is why the very people who live on the results of past researches are so often the most critical of the labour and eort which, in their time, is being expended to simplify the problems of the future6 It seems evident then that there must be at least two levels of assault on turbulence. At one level, the very nature of turbulence must be explored. At the other level, our current state of knowledge however inadequate it might be must be stretched to provide engineering solutions to real problems. The great danger we face is of being deceived by the successes and good fortune of our engineering solutions into thinking we really understand the physics. But the real world has a way of shocking us back to reality when our tried and tested engineering model fails miserably on a completely new problem for which we have not calibrated it. This is what happens when we really dont understand the physics behind what we are doing. Hopefully this course will get you excited about both the physics and the applications, so you wont fall into this trap.
1.4
The cost of our ignorance
It is dicult to place a price tag on the cost of our limited understanding of turbulence, but it requires no imagination at all to realize that it must be enormous. Try to estimate, for example, the aggregate cost to society of our limited turbulence prediction abilities which result in inadequate weather-forecasts alone. Or try to place a value on the increased cost to the consumer of the need of the designer of virtually every uid-thermal system from heat exchangers to hypersonic planes to depend on empiricism and experimentation, with the resulting need for abundant safety factors and non-optimal performance by all but the crudest measures. Or consider the frustration to engineers and cost to management of the never-ending need for code-validation experiments every time a new class
6
Jones, B.M. (1934) Stalling, The Wilbur Wright Memorial Lecture. J. Aeron. Sci
1.5. WHAT DO WE REALLY KNOW FOR SURE?
15
of ows is encountered or major design change contemplated. The whole of idea of codes in the rst place was to be able to evaluate designs without having to do experiments or build prototypes. Some argue that our quest for knowledge about turbulence should be driven solely by the insatiable scientic curiosity of the researcher, and not by the applications. Whatever the intellectual merits of this argument, it is impossible to consider the vastness and importance of the applications and not recognize a purely nancial imperative for fundamental turbulence research. The problem is, of course, that the cost of our ignorance is not conned to a single large need or to one segment of society, but is spread across the entire economic spectrum of human existence. If this were not the case, it would be easy to imagine federal involvement at the scale of Americas successful moon venture or the international space station, or at very least a linear accelerator or a Galileo telescope. Such a commitment of resources would certainly advance more rapidly our understanding. But the turbulence community those who study and those who use the results have failed ourselves to recognize clearly the need and nature of what we really do. Thus in turbulence, we have been forced to settle for far, far less than required to move us forward very fast, or maybe at all. Hopefully you will live to see this change. Or even better, perhaps you will be among the ones who change it.
1.5
What do we really know for sure?
Now even from these brief remarks, you have probably already gured out that the study of turbulence might be a little dierent than most of the subjects you have studied. This is a subject we are still studying. Now not everyone who teaches courses on this subject (and especially those who write books about it) will tell you this, but the truth is: we really dont know a whole lot for sure about turbulence. And worse, we even disagree about what we think we know! Now, as you will learn in this course (or maybe heard somewhere before), there are indeed some things some researchers think we understand pretty well like for example the Kolmogorov similarity theory for the dissipative scales and the Law of the Wall for wall-bounded ows, ideas you will soon encounter. These are based on assumptions and logical constructions about how we believe turbulence behaves in the limit of innite Reynolds number. But even these ideas have never really been tested in controlled laboratory experiments in the limits of high Reynolds number, because no one has ever had the large scale facilities required to do so.7 It seems to be a characteristic of humans (and contrary to popular belief, scientists and engineers are indeed human) that we tend to accept ideas which
The proposal to build the Nordic Wind Tunnel at Chalmers back at the beginning of this millenium was an attempt to ll this gap.
7
16
have been around a while as fact, instead of just working hypotheses that are still waiting to be tested. One can reasonably argue that the acceptance of most ideas in turbulence is perhaps more due to the time lapsed since they were proposed and found to be in reasonable agreement with a limited data base, than that they have been subjected to experimental tests over the range of their assumed validity.8 Thus it might be wise to view most established laws and theories of turbulence as more like religious creeds than matters of fact. The whole situation is a bit analogous to the old idea that the sun and stars revolved around the earth it was a ne idea, and even good today for navigational purposes. The only problem was that one day someone (Copernicus, Brahe and Galileo among them) looked up and realized it wasnt true. So it may be with a lot of what we believe today to be true about turbulence some day you may be the one to look at evidence in a new way and decide that things we thought to be true are wrong.
1.6
Our personal adventure
This is a turbulence course. You are enthused I hope, at least for the moment, to learn about turbulence. But since no two people could be in complete agreement about something like turbulence about which we know so little, this will be perhaps a pretty unusual course. I will try really hard to be honest in what I tell you. Even so, you should not trust me entirely, nor anyone else for that matter. It will really be up to you to distinguish among what you wish to consider as fact, working hypothesis, or to dismiss as fantasy. It is also very important that you try to keep track of which is which in your mind, and be willing to let ideas move from one category to the other as your understanding and information grows. Like dierent artists painting the same scene, the pictures you and I paint will as much reect our own personalities and histories, as the facts. But, like real works of art, both my picture of turbulence and yours might enable others to see things that they would have otherwise missed. This does not imply, however, that there are not real truths to be found only that we at this point can not say with condence what they are. Above all, we must not forget that we seek truth and understanding, the rst step toward which is learning and admitting what we do not know. Of course we will try to never completely forget that there are real problems to be solved. Throughout these lectures I will try to use many illustrations from my own experience. But the real goal is to help you develop enough fundamental understanding that you can sort through the many options available to you for the particular problems you will encounter in the real world. And maybe, with a little luck, you will even be able to make your own contribution to the state of our knowledge about turbulence. But at very least I hope the result of this course
This point was made rather forcefully by Robert R. Long, (Professor Emeritus, Johns Hopkins University) in his famous footnoted Journal of Fluid Mechanics paper in 1982.
8
1.7. A BRIEF OUTLINE
17
will be to make you open to new ideas, however uncomfortable they make make you feel initially. I encourage you to not be lazy. Too many study turbulence hoping for easy and quick answers, general formulas, and word pictures. The fact is the study of turbulence is quite dicult, and demands serious commitment on the part of the student. The notations are sometimes complex, and they must be this way to succinctly express the real physics. The equations themselves are extremely dicult, yet only by using them to express ideas can we say we understand the physics. Word pictures and sketches can help us, but they cannot be extrapolated to real problems. Of course we must resort to simplications and at times even heuristic reasoning to understand what our equations are telling us. But be careful to never confuse these simplications and pedagogical tools with the real ows you are likely to encounter. Sometimes they are useful in understanding, yet sometimes they can be misleading. There is no substitute for actually looking at a ow and analyzing exactly which terms in the governing equations are responsible for what you see. If this all seems a bit discouraging, look at it this way. If the turbulence problem were easy, it would have been solved years ago. Like applying Newtons law (or even relativity) to point masses with known forces, every engineer could do turbulence on his laptop.9 The turbulence problem has been worked on for over a century by many very smart people. There has certainly been progress, some would even say great progress. But not enough to make the study of turbulence easy. This problem is dicult. Even so, the equations require no more skills than undergraduate mathematics just a lot of it. So be of brave heart and persevere. Do not quit before the end of an analysis. Actually carrying things out yourself is the only road to complete understanding. The dierence between the success and failure of your eort will be almost entirely measured by your willingness to spend time and think dicult and complex thoughts. In other words, you cant be lazy and learn turbulence.
1.7
A brief outline
Now for some specics: this book will provide an introduction to the fundamentals of turbulent ow. The focus will be on understanding the averaged equations of motion and the underlying physics they contain. The goal will be to provide you with the tools necessary to continue the study of turbulence, whether in the university or industrial setting. Topics covered include: what is turbulence; the Reynolds-averaged equations; instability and transition; simple closure models; the Reynolds stress equations; simple decaying turbulence; homogeneous shear ow turbulence; free turbulent shear ows; wall-bounded turbulent ows; multipoint and spectral considerations; and multi-point similarity in turbulence.
9
Some indeed might now think this is possible, and for some very simple problems, it is.
18
Study questions for Chapter 1

1. Observe your surroundings carefully and identify at least ten dierent turbulent phenomena for which you can actually see ow patterns. Write down what you nd particularly interesting about each. 2. Talk to people (especially engineers) you know (or even dont know particularly well) about what they think the turbulence problem is. Decide for yourself whether they have fallen into the trap that Professor Jones talks about in the quotation used in this text. 3. In 1990 I wrote an ASME paper entitled The nature of turbulence. (You can download a copy for yourself from the TRL website.) In this paper I suggested that most turbulence researchers wouldnt recognize a solution to the turbulence problem, even if they stumbled across it. My idea was that if you dont know what you are looking for, you arent likely to know when you nd it. What do you think about this, especially in light of your interviews above? 4. Some believe that computers have already (or at least soon will) make experiments in turbulence unnecessary. The simplest ow one can imagine of suciently high Reynolds number to really test any of the theoretical ideas about turbulence will require a computational box of approximately (105 )3 , because of the large range of scales needed. The largest simulation to-date uses a computational box of (103 )3 , and takes several thousand hours of processor time. Assuming computer capacity continues to double every 1.5 years, calculate how many years it will be before even this simple experiment can be done in a computer. 5. The famous aerodynamicist Theordore von Karman once said: A scientist studies what is; an engineer creates what has never been. Think about this in the context of the comments in Chapter 1, and about the diering goals of the scientist and the engineer. Then try to gure out how you can plot a life course that will not trap you into thinking your own little corner of the world is all there is. 6. The instructor has essentially told you that you really should believe nothing he says (or anyone else says, for that matter), just because he (or they) said it. Think about what the scientic method really is, and how you will apply it to your study of the material in this course. 7. Think about the comments that ideas become accepted simply because they have been around awhile without being disproved. Can you think of examples from history, or from your own personal experience? Why do you think this happens? And how can we avoid it, at least in our work as scientists and engineers?
Chapter 2 The Elements of Statistical Analysis

Original version February 26, 1987, last revised 11 Feb 2010.
2.1
Foreword
Much of the study of turbulence requires statistics and stochastic processes, simply because the instantaneous motions are too complicated to understand. This should not be taken to mean that the governing equations (usually the NavierStokes equations) are stochastic. Even simple non-linear equations can have deterministic solutions that look random. In other words, even though the solutions for a given set of initial and boundary conditions can be perfectly repeatable and predictable at a given time and point in space, it may be impossible to guess from the information at one point or time how it will behave at another (at least without solving the equations). Moreover, a slight change in the initial or boundary conditions may cause large changes in the solution at a given time and location; in particular, changes that we could not have anticipated. In this chapter we shall introduce the simple idea of the ensemble average. Most of the statistical analyses of turbulent ows are based on the idea of an ensemble average in one form or another. In some ways this is rather inconvenient, since it will be obvious from the denitions that it is impossible to ever really measure such a quantity. Therefore we will spend the last part of this chapter talking about how the kinds of averages we can compute from data correspond to the hypothetical ensemble average we wish we could have measured. In later chapters we shall introduce more statistical concepts as we require them. But the concepts of this chapter will be all we need to begin a discussion of the averaged equations of motion in Chapter 3. 19
20
CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS
2.2
2.2.1
The Ensemble and Ensemble Averages

The mean (or ensemble) average
The concept of an ensemble average is based upon the existence of independent statistical events. For example, consider a number of individuals who are simultaneously ipping unbiased coins. If a value of one is assigned to a head and the value of zero to a tail, then the arithmetic average of the numbers generated is dened as: 1 XN = xn (2.1) N where our nth ip is denoted as xn and N is the total number of ips. Now if all the coins are the same, it doesnt really matter whether we ip one coin N times, or N coins a single time. The key is that they must all be independent events meaning the probability of achieving a head or tail in a given ip must be completely independent of what happens in all the other ips. Obviously we cant just ip one coin once and count it N times; these clearly would not be independent events. Exercise Carry out an experiment where you ip a coin 100 times in groups of 10 ips each. Compare the values you get for X10 for each of the 10 groups, and note how they dier from the value of X100 . Unless you had a very unusual experimental result, you probably noticed that the value of the X10 s was also a random variable and diered from ensemble to ensemble. Also the greater the number of ips in the ensemble, the closer you got to XN = 1/2. Obviously the bigger N , the less uctuation there is in XN . Now imagine that we are trying to establish the nature of a random variable, x. The nth realization of x is denoted as xn . The ensemble average of x is denoted as X (or x), and is dened as X = x lim
N 1 xn N N n=1
(2.2)
Obviously it is impossible to obtain the ensemble average experimentally, since we can never have an innite number of independent realizations. The most we can ever obtain is the arithmetic mean for the number of realizations we have. For this reason the arithmetic mean can also referred to as the estimator for the true mean or ensemble average. Even though the true mean (or ensemble average) is unobtainable, nonetheless, the idea is still very useful. Most importantly, we can almost always be sure the ensemble average exists, even if we can only estimate what it really is. Note that in some particularly dicult cases this may require imagining that we can sum our random variable at a given instant and time across an innite number of universes which are governed by the same statistical rules. So the fact of the existence of the ensemble average does not always mean that it is easy to obtain in practice.
2.2. THE ENSEMBLE AND ENSEMBLE AVERAGES
21
Nonetheless, unless stated otherwise, all of the theoretical deductions in this book will use this ensemble average; and therefore are completely general. Obviously this will mean we have to account for these statistical dierences between true means and estimates of means when comparing our theoretical results to actual measurements or computations. In general, the xn could be realizations of any random variable. The X dened by equation 2.2 represents the ensemble average of it. The quantity X is sometimes referred to as the expected value of the random variable x, or even simply its mean. For example, the velocity vector at a given point in space and time, , t, in a x given turbulent ow can be considered to be a random variable, say ui ( , t). If x (n) there were a large number of identical experiments so that the ui ( , t) in each of x (n) them were identically distributed, then the ensemble average of ui ( , t) would x be given by N 1 (n) ui ( , t) = Ui ( , t) lim x x ui ( , t) x (2.3) N N n=1 Note that this ensemble average, Ui ( , t), will, in general, vary with the index pendent variables and t. It will be seen later that under certain conditions x the ensemble average is the same as the average which would be generated by averaging in time, or even space. But even when a time (or space) average is not meaningful, however, the ensemble average can still be dened; e.g., as in an non-stationary or periodic ow. Only ensemble averages will be used in the development of the turbulence equations in this book unless otherwise stated. Thus the equations derived will be completely general, and quite independent of the particular nature of the ow, or even its statistical character.
2.2.2
Fluctuations about the mean
It is often important to know how a random variable is distributed about the mean. For example, Figure 2.1 illustrates portions of two random functions of time which have identical means, but are obviously members of dierent ensembles since the amplitudes of their uctuations are not distributed the same. It is possible to distinguish between them by examining the statistical properties of the uctuations about the mean (or simply the uctuations) dened by: x = x X It is easy to see that the average of the uctuation is zero, i.e., x = 0 (2.5) (2.4)
On the other hand, the ensemble average of the square of the uctuation is not zero. In fact, it is such an important statistical measure we give it a special name, the variance, and represent it symbolically by either var[x] or (x )2 . The
22
Figure 2.1: A typical random function of time with non-zero mean value.
variance is dened as: var[x] (x )2 = [x X]2 =

N
(2.6) (2.7)
lim
1 N
[xn X]2
n=1
Note that the variance, like the ensemble average itself, can never really be measured, since it would require an innite number of members of the ensemble. It is straightforward to show from equation 2.2 that the variance in equation 2.6 can be written as: var[x] = x2 X 2 (2.8) Thus the variance is the second-moment minus the square of the rst-moment (or mean). In this naming convention, the ensemble mean is the rst moment. Exercise Use the denitions of equations 2.2 and 2.7 to derive equation 2.8. The variance can also referred to as the second central moment of x. The word central implies that the mean has been subtracted o before squaring and averaging. The reasons for this will be clear below. If two random variables are identically distributed, then they must have the same mean and variance. The variance is closely related to another statistical quantity called the standard deviation or root mean square (rms) value of the random variable x,which is
2.2. THE ENSEMBLE AND ENSEMBLE AVERAGES
23
Figure 2.2: Two random functions of time having the same mean and variance, but very dierent higher moments.
denoted by the symbol, x . Thus, x (var[x])1/2

2 or x = var[x].
(2.9)
2.2.3
Higher moments
Figure 2.2 illustrates two random variables of time which have the same mean and also the same variances, but clearly they are still quite dierent. It is useful, therefore, to dene higher moments of the distribution to assist in distinguishing these dierences. The m-th moment of the random variable is dened as:
N 1 m x = lim xn N N n=1 m
(2.10)
It is usually more convenient to work with the central moments dened by: (x )m = (x X)m = lim
N 1 [xn X]m N N n=1
(2.11)
24
The central moments give direct information on the distribution of the values of the random variable about the mean. It is easy to see that the variance is the second central moment (i.e., m = 2).
2.3
2.3.1
Probability
The histogram and probability density function
The frequency of occurrence of a given amplitude (or value) from a nite number of realizations of a random variable can be displayed by dividing the range of possible values of the random variables into a number of slots (or windows). Since all possible values are covered, each realization ts into only one window. For every realization a count is entered into the appropriate window. When all the realizations have been considered, the number of counts in each window is divided by the total number of realizations. The result is called the histogram (or frequency of occurrence diagram). From the denition it follows immediately that the sum of the values of all the windows is exactly one. The shape of a histogram depends on the statistical distribution of the random variable, but it also depends on the total number of realizations, N , and the size of the slots, c. The histogram can be represented symbolically by the function Hx (c, c, N ) where c x < c + c, c is the slot width, and N is the number of realizations of the random variable. Thus the histogram shows the relative frequency of occurrence of a given value range in a given ensemble. Figure 2.3 illustrates a typical histogram. If the size of the sample is increased so that the number of realizations in each window increases, the diagram will become less erratic and will be more representative of the actual probability of occurrence of the amplitudes of the signal itself, as long as the window size is suciently small. If the number of realizations, N , increases without bound as the window size, c, goes to zero, the histogram divided by the window size goes to a limiting curve called the probability density function, Bx (c). That is, Bx (c) lim H(c, c, N )/c N c 0 (2.12)
Note that as the window width goes to zero, so does the number of realizations which fall into it, N H. Thus it is only when this number (or relative number) is divided by the slot width that a meaningful limit is achieved. The probability density function (or pdf) has the following properties: Property 1: Bx (c) > 0 always. (2.13)
2.3. PROBABILITY
25
Figure 2.3: Histogram, together with its limiting probability density function.
Property 2: P rob{c < x < c + dc} = Bx (c)dc where P rob{ } is read the probability that. Property 3: P rob{c < x} = Property 4:
(2.14)
Bx (c)dc
(2.15)
Bx (x)dx = 1
(2.16)
The condition imposed by property (1) simply states that negative probabilities are impossible, while property (4) assures that the probability is unity that a realization takes on some value. Property (2) gives the probability of nding the realization in a interval around a certain value, while property (3) provides the probability that the realization is less than a prescribed value. Note the necessity of distinguishing between the running variable, x, and the integration variable, c, in equations 2.14 and 2.15. Since Bx (c)dc gives the probability of the random variable x assuming a value between c and c + dc, any moment of the distribution can be computed by integrating the appropriate power of x over all possible values. Thus the n-th moment is given by: xn =

cn Bx (c)dc
(2.17)
26
Exercise: Show (by returning to the denitions) that the value of the moment determined in this manner is exactly equal to the ensemble average dened earlier in equation 2.10. (Hint: use the denition of an integral as a limiting sum.) If the probability density is given, the moments of all orders can be determined. For example, the variance can be determined by: var{x} = (x X)2 =
(c X)2 Bx (c)dc
(2.18)
The central moments give information about the shape of the probability density function, and vice versa. Figure 2.4 shows three distributions which have the same mean and standard deviation, but are clearly quite dierent. Beneath them are shown random functions of time which might have generated them. Distribution (b) has a higher value of the fourth central moment than does distribution (a). This can be easily seen from the denition (x X) =
4
(c X)4 Bx (c)dc
(2.19)
since the fourth power emphasizes the fact that distribution (b) has more weight in the tails than does distribution (a). It is also easy to see that because of the symmetry of pdfs in (a) and (b), all the odd central moments will be zero. Distributions (c) and (d), on the other hand, have non-zero values for the odd moments, because of their asymmetry. For example, (x X)3 =

(c X)3 Bx (c)dc
(2.20)
is equal to zero if Bx is an even function.
2.3.2
The probability distribution
Sometimes it is convenient to work with the probability distribution instead of with the probability density function. The probability distribution is dened as the probability that the random variable has a value less than or equal to a given value. Thus from equation 2.15, the probability distribution is given by
Fx (c) = P rob{x < c} =
Bx (c )dc
(2.21)
Note that we had to introduce the integration variable, c , since c occurred in the limits. Equation 2.21 can be inverted by dierentiating by c to obtain Bx (c) = dFx dc (2.22)
2.3. PROBABILITY
27
Figure 2.4: Relation of skewness to the shape of the pdf and nature of the signal.
28
2.3.3
Gaussian (or normal) distributions
One of the most important pdfs in turbulence is the Gaussian or Normal distribution dened by 1 2 2 BxG (c) = e(cX) /2x (2.23) 2x where X is the mean and x is the standard derivation. The factor of 1/ 2x insures that the integral of the pdf over all values is unity as required. It is easy to prove that this is the case by completing the squares in the integration of the exponential (see problem 2.2). The Gaussian distribution is unusual in that it is completely determined by its rst two moments, X and . This is not typical of most turbulence distributions. Nonetheless, it is sometimes useful to approximate turbulence as being Gaussian, often because of the absence of simple alternatives. It is straightforward to show by integrating by parts that all the even central moments above the second are given by the following recursive relationship,
n (x X)n = (n 1)(n 3)...3.1x 4 6 Thus the fourth central moment is 3x , the sixth is 15x , and so forth.
(2.24)
Exercise: Prove this. The probability distribution corresponding to the Gaussian distribution can be obtained by integrating the Gaussian pdf from to x = c; i.e., 1 c (c X)2 /2x 2 FxG (c) = e dc 2x (2.25)
The integral is related to the erf-function tabulated in many standard tables and function subroutines, but usually with the independent variable normalized by x ; i.e., c = c/x . Alternatively we can subtract FxG (c) from unity to obtain the probability that x c as 1 FxG (c). This is related to the complementary error function, erfc(c), also usually easily available.
2.3.4
Skewness and kurtosis
Because of their importance in characterizing the shape of the pdf, it is useful to dene scaled (or normalized) versions of third and fourth central moments: the skewness and kurtosis respectively. The skewness is dened as third central moment divided by the three-halves power of the second; i.e., S= (x X)3 (x X)2 3/2 (2.26)
The kurtosis is dened as the fourth central moment divided by the square of the second; i.e., (x X)4 K= (2.27) (x X)2 2
2.4. MULTIVARIATE RANDOM VARIABLES
29
Both these are easy to remember if you note the S and K must be dimensionless. The pdfs in Figure 2.4 can be distinguished by means of their skewness and kurtosis. The random variable shown in (b) has a higher kurtosis than that in (a). Thus the kurtosis can be used as an indication of the tails of a pdf, a higher kurtosis indicating that relatively larger excursions from the mean are more probable. The skewnesses of (a) and (b) are zero, whereas those for (c) and (d) are non-zero. Thus, as its name implies, a non-zero skewness indicates a skewed or asymmetric pdf, which in turn means that larger excursions in one direction are more probable than in the other. For a Gaussian pdf, the skewness is zero and the kurtosis is equal to three (see problem 2.4). The atness factor, dened as (K 3), is sometimes used to indicate deviations from Gaussian behavior. Exercise: Prove that the skewness and kurtosis of a Gaussian distributed random variable are 0 and 3 respectively.
2.4
2.4.1
Multivariate Random Variables

Joint pdfs and joint moments
Often it is important to consider more than one random variable at a time. For example, in turbulence the three components of the velocity vector are interrelated and must be considered together. In addition to the marginal (or single variable) statistical moments already considered, it is necessary to consider the joint statistical moments. For example if u and v are two random variables, there are three second-order moments which can be dened u2 , v 2 , and uv. The product moment uv is called the cross-correlation or cross-covariance. The moments u2 and v 2 are referred to as the covariances, or just simply the variances. Sometimes uv is also referred to as the correlation. In a manner similar to that used to build-up the probability density function from its measurable counterpart, the histogram, a joint probability density function (or jpdf), Buv , can be built-up from the joint histogram. Figure 2.5 illustrates several examples of jpdfs which have dierent cross-correlations. For convenience the uctuating variables u and v can be dened as u = u U v = v V (2.28) (2.29)
where as before capital letters are used to represent the mean values. Clearly the uctuating quantities u and v are random variables with zero mean. A positive value of u v indicates that u and v tend to vary together. A negative value indicates that when one variable is increasing the other tends to be decreasing. A zero value of u v indicates that there is no correlation between
30
Figure 2.5: Contours of constant probability for four dierent joint probability density functions. Try to gure out what the moments would be for each and how they would dier.
2.4. MULTIVARIATE RANDOM VARIABLES
31
u and v . As will be seen below, it does not mean that they are statistically independent. It is sometimes more convenient to deal with values of the cross-variances which have been normalized by the appropriate variances. Thus the correlation coecient is dened as: u v uv (2.30) [u 2 v 2 ]1/2 The correlation coecient is bounded by plus or minus one, the former representing perfect correlation and the latter perfect anti-correlation. As with the single-variable pdf, there are certain conditions the joint probability density function must satisfy. If Buv (c1 , c2 ) indicates the jpdf of the random variables u and v, then: Property 1: Buv (c1 , c2 ) > 0 always. Property 2: (2.31)
P rob{c1 < u < c1 + dc1 , c2 < v < c2 + dc2 } = Buv (c1 , c2 )dc1 , dc2 Property 3:
(2.32)
Buv (c1 , c2 )dc1 dc2 = 1
(2.33)
Property 4:
Buv (c1 , c2 )dc2 = Bu (c1 )
(2.34)
where Bu is a function of c1 only. Property 5:
Buv (c1 , c2 )dc1 = Bv (c2 )
(2.35)
where Bv is a function of c2 only. The functions Bu and Bv are called the marginal probability density functions, and they are simply the single variable pdfs dened earlier. The subscript is used to indicate which variable is left after the others are integrated out. Note that Bu (c1 ) is not the same as Buv (c1 , 0). The latter is only a slice through the c2 -axis, while the marginal distribution is weighted by the integral of the distribution of the other variable. Figure 2.6 illustrates these dierences.
32
Figure 2.6: Surface representation of a joint probability density function.
If the joint probability density function is known, the joint moments of all orders can be determined. Thus the m, n-th joint moment is u v =
m n
cm cn Buv (c1 , c2 )dc1 dc2 1 2
(2.36)
where m and n can take any value. The corresponding central-moment is: (u U ) (v V ) =
m n
(c1 U )m (c2 V )n Buv (c1 , c2 )dc1 dc2
(2.37)
In the preceding discussions, only two random variables have been considered. The denitions, however, can easily be generalized to accommodate any number of random variables. In addition, the joint statistics of a single random variable at dierent times or at dierent points in space could be considered. This will be discussed later when stationary and homogeneous random processes are considered.
2.4.2
The bi-variate normal (or Gaussian) distribution
If u and v are normally distributed random variables with standard deviations given by u and v , respectively, with correlation coecient uv , then their joint
2.4. MULTIVARIATE RANDOM VARIABLES probability density function is given by 1 (c1 U )2 (c2 V )2 c1 c2 BuvG (c1 , c2 ) = exp + uv 2 2 2u v 2u 2v u v
[ ]
33
(2.38)
This distribution is plotted in Figure 2.7 for several values of uv where u and v are assumed to be identically distributed (i.e., u2 = v 2 ). It is straightforward to show (by completing the square and integrating) that this yields the single variable Gaussian distribution for the marginal distributions (see problem 2.5). It is also possible to write a multivariate Gaussian probability density function for any number of random variables. Exercise: Prove that equation 2.23 results from integrating out the dependence of either variable using equations 2.34 or 2.35.
2.4.3
Statistical independence and lack of correlation
Denition: Statistical Independence Two random variables are said to be statistically independent if their joint probability density is equal to the product of their marginal probability density functions. That is, Buv (c1 , c2 ) = Bu (c1 )Bv (c2 ) (2.39)
It is easy to see that statistical independence implies a complete lack of correlation; i.e., uv 0. From the denition of the cross-correlation, (u U )(v V ) = = =
(c1 U )(c2 V )Buv (c1 , c2 )dc1 dc2 (c1 U )(c2 V )Bu (c1 )Bv (c2 )dc1 dc2
(c1 U )Bu (c1 )dc1
(c2 V )Bv (c2 )dc2 (2.40)
= 0
where we have used equation 2.39 since the rst central moments are zero by denition. It is important to note that the inverse is not true lack of correlation does not imply statistical independence! To see this consider two identically distributed random variables, u and v , which have zero means and a non-zero correlation u v . From these two correlated random variables two other random variables, x and y, can be formed as: x = u + v y = u v (2.41) (2.42)
34
Figure 2.7: Jpdf on left (symmetry around slanted lines through the origin) shows correlation while the jpdf on the right (symmetry about both axes) formed by summing and dierencing the two variables does not.
Clearly x and y are not statistically independent since the quantities from which they were formed are not statistically independent. They are, however, uncorrelated because: xy = (u + v )(u v ) 2 2 = u + u v u v v = 0
(2.43)
since u and v are identically distributed (and as a consequence u 2 = v 2 ). Figure 2.7 illustrates the change of variables carried out above. The jpdf resulting from the transformation is symmetric about both axes, thereby eliminating the correlation. Transformation, however, does not insure that the distribution is separable, i.e., we did not insure that Bx,y (a1 , a2 ) = Bx (a1 )By (a2 ), as required for statistical independence. Exercise: Apply the transformation above, equation 2.41, to the jointly Gaussian pdf given by equation 2.38 with uv = 0. Then use it to determine whether in fact x and y are statistically independent.
2.5
2.5.1
Estimation from a Finite Number of Realizations

Estimators for averaged quantities
Since there can neverbe innite number of realizations from which ensemble an averages (and probability densities) can be computed, it is essential to ask: How
2.5. ESTIMATION FROM A FINITE NUMBER OF REALIZATIONS
35
many realizations are enough? The answer to this question must be sought by looking at the statistical properties of estimators based on a nite number of realizations. There are two questions which must be answered. The rst one is: Is the expected value (or mean value) of the estimator equal to the true ensemble mean? Or in other words, is the estimator unbiased? The second question is: Does the dierence between the value of the estimator and that of the true mean decrease as the number of realizations increases? Or in other words, does the estimator converge in a statistical sense (or converge in probability). Figure 2.8 illustrates the problems which can arise.
2.5.2
Bias and convergence of estimators
A procedure for answering these questions will be illustrated by considering a simple estimator for the mean, the arithmetic mean considered above, XN . For N independent realizations, xn , n = 1, 2, , N where N is nite, XN is given by: XN =
N 1 xn N n=1
(2.44)
Now, as we observed in our simple coin-ipping experiment, since the xn are random, so must be the value of the estimator XN . For the estimator to be unbiased, the mean value of XN must be the true ensemble mean, X; i.e., limN XN = X (2.45)
It is easy to see that since the operations of averaging and adding commute, XN = =
N 1 xn N n=1
(2.46) (2.47) (2.48)
N 1 xn N n=1 1 = NX = X N
(Note that the expected value of each xn is just X since the xn are assumed identically distributed). Thus xN is, in fact, an unbiased estimator for the mean. The question of convergence of the estimator can be addressed by dening the square of variability of the estimator, say 2 N , to be: X 2 N X (XN X)2 var{XN } = X2 X2 (2.49)
36
Figure 2.8: Three dierent estimators for the mean: The rst of which converges to the mean with increasing number of samples, the middle converges to the wrong mean, and the bottom does not converge at all.
2.5. ESTIMATION FROM A FINITE NUMBER OF REALIZATIONS
37
Now we want to examine what happens to XN as the number of realizations increases. For the estimator to converge it is clear that x should decrease as the number of samples increases. Obviously, we need to examine the variance of XN rst. It is given by: var{XN } = (XN X)2
[
N 1 xn X) = N n=1
]2
(2.50)
]2
N N 1 1 = xn X) N n=1 N n=1 N 1 = (xn X) N n=1
(2.51) (2.52)
]2
since XN = X from equation 2.46. Using the fact that the operations of averaging and summation commute, the squared summation can be expanded as follows:
[
]2
(xn X) = =
n=1
N N 1 (xn X)(xm X) N 2 n=1 m=1
N 1 (xn X)2 N 2 n=1 1 var{x}, (2.53) = N where the next to last step follows from the fact that the xn are assumed to be statistically independent samples (and hence uncorrelated), and the last step from the denition of the variance. It follows immediately by substitution into equation 2.49 that the square of the variability of the estimator, XN , is given by:
1 var{x} N X2 [ ] 1 x 2 = (2.54) N X Thus the variability of the estimator depends inversely on the number of independent realizations, N , and linearly on the relative uctuation level of the random variable itself, x /X. Obviously if the relative uctuation level is zero (either because there the quantity being measured is constant and there are no measurement errors), then a single measurement will suce. On the other hand, as soon as there is any uctuation in the x itself, the greater the uctuation (relative to the mean of x, x = X), then the more independent samples it will take to achieve a specied accuracy. 2 N = X Example: In a given ensemble the relative uctuation level is 12% (i.e., x /X = 0.12). What is the fewest number of independent samples that must be acquired to measure the mean value to within 1%?
38
CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS Answer Using equation 2.54, and taking XN = 0.01, it follows that: (0.01)2 = 1 (0.12)2 N (2.55)
or N 144. The variability is of great value in estimating statistical error. But it is, of course, of no value when the mean value of the process itself if zero, or nearly so. In such cases we have to satisfy ourselves with simply the variance (or standard deviation) of XN . This has proven to be quite terrifying to many investigators who have been afraid to show data because of what they believed to be its large relative errors. In fact were being fooled by the only apparently large scatter, since they were trying to measure a quantity whose average value was zero. This can happen quite easily since most the plotting routine are self-scaling. In most cases, a careful look at the actual numbers labelling the abscissa can put things into the proper perspective.
2.6
Generalization to the estimator of any quantity
Similar relations can be formed for the estimator of any function of the random variable, say f (x). For example, an estimator for the average of f based on N realizations is given by: N 1 FN fn (2.56) N n=1 where fn f (xn ). It is straightforward to show that this estimator is unbiased, and its variability (squared) is given by: 2 N = F 1 var{FN (x)} N f (x)2 (2.57)
Example: Suppose it is desired to estimate the variability of an estimator for the variance based on a nite number of samples as: varN {x}
N 1 (xn X)2 N n=1
(2.58)
(Note that this estimator is not really the best that we could do since it presumes that the mean value, X, is known, whereas in fact usually only XN is obtainable as in Problem 2.6 below. The errors this induces can be quite serious if the number of independent samples is quite small. The so-called Student-T distribution is an attempt to deal with this problem. In modern turbulence research the number of independent samples in a properly designed experiment is usually very large, so the dierences are slight. )
2.6. GENERALIZATION TO THE ESTIMATOR OF ANY QUANTITY
39
Answer Let f = (x X)2 in equation 2.57 so that FN = varN {x}, f = var{x} and var{f } = var{(x X)2 } = {(x X)2 var[x X])}2 . Then: 2 N var 1 var{(x X)2 } = N (var{x})2 (2.59)
This is easiest to understand if we rst expand only the numerator to obtain: var{(x X)2 } = {(x X)2 var[x]}2 = (x X)4 [var{x}]2 Thus 2 N = var (2.60)
1 (x X)4 [var{x}]2 (2.61) N [var{x}]2 Obviously to proceed further we need to know how the fourth central moment relates to the second central moment. As noted earlier, in general this is not known. If, however, it is reasonable to assume that x is a Gaussian distributed random variable, we know from section 2.3.4 that the kurtosis is 3. Then for Gaussian distributed random variables, 2 N = var 2 N (2.62)
Thus the number of independent data required to produce the same level of convergence for an estimate of the variance of a Gaussian distributed random variable is 2 times that of the mean. It is easy to show that the higher the moment, the more the amount of data required (see Problem 2.7). As noted earlier, turbulence problems are not usually Gaussian, and in fact values of the kurtosis substantially greater than 3 are commonly encountered, especially for the moments of dierentiated quantities. Clearly the non-Gaussian nature of random variables can aect the planning of experiments, since substantially greater amounts of data can be required to achieved the necessary statistical accuracy.
Problems for Chapter 2

1. By using the denition of the probability density function as the limit of the histogram of a random variable as the internal size goes to zero and as the number of realizations becomes innite (equation 2.12), show that the probability average dened by equation 2.17 and the ensemble average dened by equation 2.2 are the same. 2. By completing the square in the exponential, prove that the pdf for the normal distribution given by equation 2.23 integrates to unity (equation 2.16) 3. Prove equation 2.24. 4. Prove for a normal distribution that the skewness is equal to zero and that the kurtosis is equal to three.
40
CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS 5. Show by integrating over one of the variables that the Gaussian jpdf given by equation 2.38 integrates to the marginal distribution pdf given by equation 2.23, regardless of the value of the correlation coecient. 6. Find the variability of an estimator for the variance using equation 2.60, but with the sample mean, XN , substituted for the true mean, X. 7. Create a simple estimator for the fourth central moment assuming the second to be known exactly. Then nd its variability for a Gaussian distributed random variable. 8. You are attempting to measure a Gaussian distributed random variable with 12 bit A/D converter which can only accept voltage inputs between 0 and 10. Assume the mean voltage is +4, and the rms voltage is 4. Show what a histogram of your measured signal would look like assuming any voltage which is clipped goes into the rst or last bins. Also compute the rst three moments (central) of the measured signal.
Chapter 3 The Reynolds Averaged Equations and the Turbulence Closure Problem
3.1 The Equations Governing the Instantaneous Fluid Motions
All uid motions, whether turbulent or not, are governed by the dynamical equations for a uid. These can be written using Cartesian tensor notation as: ui ui p Tij + uj = + t xj xi xj { } uj + uj =0 + t xj xj
[ ]
(v)
(3.1) (3.2)
where ui ( , t) represents the i-th component of the uid velocity at a point in x (v) x space, [ ]i = xi , and time, t. Also p( , t) represents the static pressure, Tij ( , t), x x the viscous (or deviatoric) stresses, and the uid density. The tilde over the symbol indicates that an instantaneous quantity is being considered. Also the Einstein summation convention has been employed.1 In equation 3.1, the subscript i is a free index which can take on the values 1, 2, and 3. Thus equation 3.1 is in reality three separate equations. These three equations are just Newtons second law written for a continuum in a spatial (or Eulerian) reference frame. Together they relate the rate of change of momentum per unit mass (ui ), a vector quantity, to the contact and body forces. Equation 3.2 is the equation for mass conservation in the absence of sources (or sinks) of mass. Almost all ows considered in this book will be incompressible, which implies that the derivative of the density following the uid material (the
1
Einstein summation convention: repeated indices in a single term are summed over 1,2, and
3.
41
42
CHAPTER 3. REYNOLDS AVERAGED EQUATIONS
term in brackets) is zero. Thus for incompressible ows, the mass conservation equation reduces to: D = + uj =0 (3.3) Dt t xj From equation 3.2 it follows that for incompressible ows, uj =0 xj (3.4)
The viscous stresses (the stress minus the mean normal stress) are represented (v) (v) by the tensor Tij . From its denition, Tkk = 0. In many ows of interest, the uid behaves as a Newtonian uid in which the viscous stress can be related to the uid motion by a constitutive relation of the form 1 (v) Tij = 2 sij skk ij 3
[ ]
(3.5)
The viscosity, , is a property of the uid that can be measured in an independent experiment. sij is the instantaneous strain rate tensor dened by 1 ui uj + sij 2 xj xi
[ ]
(3.6)
From its denition, skk = uk /xk . If the ow is incompressible, skk = 0 and the Newtonian constitutive equation reduces to (v) Tij = 2ij s (3.7)
Throughout this text, unless explicity stated otherwise, the density, = and the viscosity will be assumed constant. With these assumptions, the instantaneous momentum equations for a Newtonian uid reduce to:
[
ui ui 1 p 2 ui + uj = + 2 t xj xi xj
(3.8)
where the kinematic viscosity, , has been dened as:
(3.9)
Note that since the density is assumed constant, the tilde is no longer necessary. Sometimes it will be more instructive and convenient to not explicitly include incompressibility in the stress term, but to refer to the incompressible momentum equation in the following form:
] (v) ui p Tij ui + uj = + t xj xi xj [
(3.10)
This form has the advantage that it is easier to keep track of the exact role of the viscous stresses.
3.2. EQUATIONS FOR THE AVERAGE VELOCITY
43
3.2
Equations for the Average Velocity
Turbulence is that chaotic state of motion characteristic of solutions to the equations of motion at high Reynolds number. Although laminar solutions to the equations often exist that are consistent with the boundary conditions, perturbations to these solutions (sometimes even innitesimal) can cause them to become turbulent. To see how this can happen, it is convenient to analyze the ow in two parts, a mean (or average) component and a uctuating component. Thus the instantaneous velocity and stresses can be written as: ui = Ui + ui p = P +p (v) (v) (v) Tij = Tij + ij
(v)
(3.11)
where Ui , p, and Tij represent the mean motion, and ui , p, and ij the uctuating motions. This technique for decomposing the instantaneous motion is referred to as the Reynolds decomposition. Note that if the averages are dened as ensemble means, they are, in general, time-dependent. For the remainder of this book, unless otherwise stated, the density will be assumed constant so and its uctuation is zero. Substitution of equations 3.11 into equations 3.10 yields (Ui + ui ) (Ui + ui ) (P + p) (Tij + ij ) + (Uj + uj ) + = t xj xi xj
[ ]
(v) (v)
(3.12)
This equation can now be averaged to yield an equation expressing momentum conservation for the averaged motion. Note that the operations of averaging and dierentiation commute; i.e., the average of a derivative is the same as the derivative of the average. Also, the average of a uctuating quantity is zero.2 Thus the equation for the averaged motion reduces to: T Ui Ui P ui + Uj = + ij uj t xj xi xj xj
[ ]
(v)
(3.13)
where the remaining uctuating product term has been moved to the right-hand side of the equation. Whether or not this last term is zero like the other uctuating terms depends on the correlation of terms in the product. In general, these correlations are not zero. The mass conservation equation can be similarly decomposed. In incompressible form, substitution of equations 3.11 into equation 3.4 yields: (Uj + uj ) =0 xj
2
(3.14)
These are easily proven from the denitions of both.
44 of which the average is:
Uj =0 xj
(3.15)
It is clear from equation 3.15 that the averaged motion satises the same form of the mass conservation equation as does the instantaneous motion, at least for incompressible ows. How much simpler the turbulence problem would be if the same were true for the momentum! Unfortunately, as is easily seen from equation 3.13, such is not the case. Equation 3.15 can be subtracted from equation 3.14 to yield an equation for the instantaneous motion alone; i.e., uj =0 xj (3.16)
Again, like the mean, the form of the original instantaneous equation is seen to be preserved. The reason, of course, is obvious: the continuity equation is linear. The momentum equation, on the other hand, is not; hence the dierence. Equation 3.16 can be used to rewrite the last term in equation 3.13 for the mean momentum. Multiplying equation 3.16 by ui and averaging yields: ui uj =0 xj (3.17)
This can be added to uj ui /xj to obtain: uj ui ui uj + 0 = uj + ui = ui uj xj xj xj xj (3.18)
where again the fact that arithmetic and averaging operations commute has been used. The equation for the averaged momentum, equation 3.13 can now be rewritten as: [ ] (v) Tij Ui Ui P + Uj + ui uj (3.19) = t xj xi xj xj The last two terms on the right-hand side are both divergence terms and can be combined; the result is:
] Ui P [ (v) Ui + Uj = + Tij ui uj t xj xi xj [ ]
(3.20)
Now the terms in square brackets on the right have the dimensions of stress. The rst term is, in fact, the viscous stress. The second term, on the other hand, is not a stress at all, but simply a re-worked version of the uctuating contribution to the non-linear acceleration terms. The fact that it can be written this way, however, indicates that at least as far as the mean motion is concerned, it acts as though it were a stress hence its name, the Reynolds stress. In the succeeding sections the consequences of this dierence will be examined.
3.3. THE TURBULENCE PROBLEM
45
3.3
The Turbulence Problem
It is the appearance of the Reynolds stress which makes the turbulence problem so dicult at least from the engineers perspective. Even though we can pretend it is a stress, the physics which give rise to it are very dierent from the viscous stress. The viscous stress can be related directly to the other ow properties by constitutive equations, which in turn depend only on the properties of the uid (as in equation 3.5 for a Newtonian uid). The reason this works is that when we make such closure approximations for a uid, we are averaging over characteristic length and time scales much smaller than those of the ows we are interested in. Yet at the same time, these scales are much larger than the molecular length and time scales which characterize the molecular interactions that are actually causing the momentum transfer. (This is what the continuum approximation is all about.) The Reynolds stress, on the other hand, arises directly from the ow itself! Worse, the scales of the uctuating motion which give rise to it are the scales we are interested in. This means that the closure ideas which worked so well for the viscous stress, should not be expected to work too well for the Reynolds stress. And as we shall see, they do not. This leaves us in a terrible position. Physics and engineering are all about writing equations (and boundary conditions) so we can solve them to make predictions. We dont want to have to build prototype airplanes rst to see if they will fall out of the sky. Instead we want to be able to analyze our designs before building the prototype, both to save the cost in money and in lives if our ideas are wrong. The same is true for dams and bridges and tunnels and automobiles. If we had condence in our turbulence models, we could even build huge one-os and expect them to work the rst time. Unfortunately, even though turbulence models have improved to the point where we can use them in design, we still cannot trust them enough to eliminate expensive wind tunnel and model studies. And recent history is full of examples to prove this. The turbulence problem (from the engineers perspective) is then three-fold: The averaged equations are not closed. Count the unknowns in equation 3.20 above. Then count the number of equations. Even with the continuity equation we have at least six equations too few. The simple ideas to provide the extra equations usually do not work. And even when we can x them up for a particular class of ows (like the ow in a pipe, for example), they will most likely not be able to predict what happens in even a slightly dierent environment (like a bend). Even the last resort of compiling engineering tables for design handbooks carries substantial risk. This is the last resort for the engineer who lacks equations or cannot trust them. Even when based on a wealth of experience, they require expensive model testing to see if they can
46
CHAPTER 3. REYNOLDS AVERAGED EQUATIONS be extrapolated to a particular situation. Unfortunately so innitely clever is Mother Nature in creating turbulence that is unique to a particular set of boundary conditions that often they cannot. Turbulent ows are indeed ows!. And that is the problem.
3.4
The Origins of Turbulence
Turbulent ows can often be observed to arise from laminar ows as the Reynolds number, (or some other relevant parameter) is increased. This happens because small distubances to the ow are no longer damped by the ow, but begin to grow by taking energy from the original laminar ow. This natural process is easily visualized by watching the simple stream of water from a faucet (or even a pitcher). Turn the ow on very slowly (or pour) so the stream is very smooth initially, at least near the outlet. Now slowly open the faucet (or pour faster) and observe what happens, rst far away, then closer to the spout. The surface begins to exhibit waves or ripples which appear to grow downstream. In fact, they are growing by extracting energy from the primary ow. Eventually they grow enough that the ow breaks into drops. These are capillary instabilities arising from surface tension. But regardless of the type of instability, the idea is the same: small (or even innitesimal) disturbances have grown to disrupt the serenity (and simplicity) of laminar ow. The manner in which instabilities grow naturally in a ow can be examined using the equations we have already developed above. We derived them by decomposing the motion into a mean and a uctuating part. But suppose instead we had decomposed the motion into a base ow part (the initially laminar part) and into a disturbance which represents a uctuating part superimposed on the base ow. The result of substituting such a decomposition into the full Navier-Stokes equations and averaging is precisely that given by equations 3.13 and 3.15. But the very important dierence is the additional restriction that what was previously identied as the mean (or averaged) motion is now also the base or laminar ow. Now if the base ow is really a laminar ow (which it must be by our original hypothesis), then our averaged equations governing the base ow must yield the same mean ow solution as the original laminar ow on which the disturbance was superimposed. But this can happen only if these new averaged equations reduce to exactly the same laminar ow equations without any evidence of a disturbance. Clearly from equations 3.13 and 3.15, this can happen only if all the Reynolds stress terms vanish identically! Obviously this requires that the disturbances be innitesimal so the extra terms can be neglected hence our interest in innitesimal disturbances. So we hypothesized a base ow which was laminar and showed that it is unchanged even with the imposition of innitesimal disturbances on it but only as long as the disturbances remain innitesimal! What happens if the disturbance
3.4. THE ORIGINS OF TURBULENCE
47
starts to grow? Obviously before we conclude that all laminar ows are laminar forever we better investigate whether or not these innitesimal disturbances can grow to nite size. To do this we need an equation for the uctuation itself. An equation for the uctuation (which might be an imposed disturbance) can be obtained by subtracting the equation for the mean (or base) ow from that for the instantaneous motion. We already did this for the continuity equation. Now we will do it for the momentum equation. Subtracting equation 3.13 from equation 3.11 yields an equation for the uctuation as: ui p Ui ui ui ui = + ij uj uj uj + Uj t xj xi xj xj xj xj
[ ]
(v)
(3.21)
It is very important to note the type and character of the terms in this equation. First note that the left-hand side is the derivative of the uctuating velocity following the mean motion. This is exactly like the term which appears on the left-hand side of the equation for the mean velocity, equation 3.13. The rst two terms on the right-hand side are also like those in the mean motion, and represent the uctuating pressure gradient and the uctuating viscous stresses. The third term on the right-hand side is new, and will be seen later to represent the primary means by which uctuations (and turbulence as well!) extract energy from the mean ow, the so-called production terms. The last term is quadratic in the uctuating velocity, unlike all the others which are linear. Note that all of the terms vanish identically if the equation is averaged, the last because its mean is subtracted from it. Now we want to examine what happens if the disturbance is small. In the limit as the amplitude of the disturbance (or uctuation) is innitesimal, the bracketed term in the equation for the uctuation vanishes (since it involves products of innitesimals), and the remaining equation is linear in the disturbance. The study of whether or not such innitesimal disturbances can grow is called Linear Fluid Dynamic Stability Theory. These linearized equations are very dierent from those governing turbulence. Unlike the equations for disturbances of nite amplitude, the linearized equations are well-posed (or closed) since the Reynolds stress terms are gone. Therefore, in principle, they can be solved exactly with no need for closure approximations. The absence of the non-linear terms, however, constrains the validity of the linear analysis to only the initial stage of disturbance growth. This is because as soon as the uctuations begin to grow, their amplitudes can no longer be assumed innitesimal and the Reynolds stress (or more properly, the non-linear uctuating terms), become important. As a result the base ow equations begin to be modied so that the solution to them can no longer be identical to the laminar ow (or base ow) from which it arose. Thus while linear stability theory can predict when many ows become unstable, it can say very little about transition to turbulence since this process is highly non-linear. It is also clear from the above why the process of transition to turbulence is so dependent on the state of the background ow. If the disturbances present
48
in the base ow are small enough, then Linear Stability Theory will govern their evolution. On the other hand if the disturbances to the base ow are not small enough, Linear Stability Theory can never apply since the non-linear terms will never be negligible. This is so-called by-pass transition. It is not uncommon to encounter situations like this in engineering environments where the incoming ow has a modest turbulence level super-imposed upon it. In such cases, the nature of the disturbances present is as important as their intensities, with the consequence that a general transition criterion may not exist, and perhaps should not even be expected.
3.5
The importance of non-linearity
We saw in the preceding section that non-linearity was one of the essential features of turbulence. When small disturbances grow large enough to interact with each other, we enter a whole new world of complex behavior. Most of the rules we learned for linear systems do not apply. Since most of your mathematical training has been for linear equations, most of your mathematical intuition therefore will not apply either. On the other hand, you may surprise yourself by discovering how much your non-mathematical intuition already recognizes non-linear behavior and accounts for it. Consider the following simple example. Take a long stick with one person holding each end and stand at the corner of a building. Now place the middle of the stick against the building and let each person apply pressure in the same direction so as to bend the stick. If the applied force is small, the stick deects (or bends) a small amount. Double the force, and the deection is approximately doubled. Quadruple the force and the deection is quadrupled. Now you dont need a Ph.D. in Engineering to know what is going to happen if you continue this process. The stick is going to break! But where in the equations for the deection of the stick is there anything that predicts this can happen? Now if you are thinking only like an engineer, you are probably thinking: hes asking a stupid question. Of course you cant continue to increase the force because you will exceed rst the yield stress, then the breaking limit, and of course the stick will break. But pretend I am the company president with nothing more than an MBA.3 I dont know much about these things, but you have told me in the past that your computers have equations to predict everything. So I repeat: Where in the equations for the deection of this stick does it tell me this is going to happen?
For some reason the famous O-ring disaster of the the Challenger space shuttle in 1983 comes to mind here. The decision by the manufacturer, Morton Thiokol, to launch at temperatures below that at which the O-ring seals had been tested was made entirely by MBAs and lawyers, over the objections of the scientists and engineers present. Aside from the tragedy of the lives lost, including Gregory Jarvis whom my former oce building at University at Bualo is named after, they blew up a billion dollar spacecraft.
3
3.5. THE IMPORTANCE OF NON-LINEARITY
49
The answer is very simple: There is nothing in the equations that will predict this. And the reason is also quite simple: You lost the ability to predict catastrophes like breaking when you linearized the fundamental equations which started out as Newtons Law too. In fact, before linearization, they were exactly the same as those for a uid, only the constitutive equation was dierent. If we had NOT linearized these equations and had constitutive equations that were more general, then we possibly could apply these equation right to and past the limit. The point of fracture would be a bifurcation point for the solution. Now the good news is that for things like reasonable deections of beams, linearization works wonderfully since we hope most things we build dont deect too much especially if you are sitting on a geological fault as I am at the moment of this writing.4 Unfortunately, as we noted above, for uids the disturbances tend to quickly become dominated by the non-linear terms. This, of course, means our linear analytical techniques are pretty useless for uid mechanics, and especially turbulence. But all is not lost. Just as we have learned to train ourselves to anticipate when sticks break, we have to train ourselves to anticipate how non-linear uid phenomena behave. Toward that end we will consider two simple examples: one from algebra the logistic map, and one from uid mechanics simple vortex stretching. Example 1: An experiment with the logistic map. Consider the behavior of the simple equation: yn+1 = ryn (1 yn ) (3.22)
where n = 1, 2, , 0 < y < 1 and r > 0. The idea is that you pick any value for y1 , use the equation to nd y2 , then insert that value on the right-hand side to nd y3 , and just continue the process as long as you like. Make sure you note any dependence of the nal result on the initial value for y. First notice what happens if you linearize this equation by disregarding the term in parentheses; i.e., consider the simpler equation yn+1 = ryn . My guess is that you wont nd this too exciting unless, of course, you are one of those individuals who likes watching grass grow (or golf on TV). Now consider the full equation and note what happens for r < 3, and especially what happens for very small values of r. Run as many iterations as necessary to make sure your answer has converged. Do NOT try to take short-cuts by programming all the steps at once. Do them one at a time so
I am sitting at this moment of this writing at the Institute for Theoretical Physics at the University of California/Santa Barbara. If you are reading this, it is likely that an earthquake did not happen during the writing session.
4
50
CHAPTER 3. REYNOLDS AVERAGED EQUATIONS you can see what is happening. Believe me, it will be much easier this way in the long run. Now research carefully what happens when r = 3.1, 3.5, and 3.8. Can you recognize any patterns. Vary r between 3 and 4 to see if you can nd the boundaries for what you are observing. Now try values of r > 4. How do you explain this?
Example 2: Stretching of a simple vortex. Imagine a simple vortex lament that looks about like a strand of spaghetti. Now suppose it is in an otherwise steady inviscid incompressible ow. Use the vorticity equation to examine the following: Examine rst what happens to it in two-dimensional velocity eld. Note particularly whether any new vorticity can be produced; i.e., can the material derivative of the vorticity ever be greater than zero? (Hint: look at the j ui /xj -term.) Now consider the same vortex lament in a three-dimensional ow. Note particularly the various ways new vorticity can be produced if you have some to start with! Does all this have anything to do with non-linearities? Now you are ready for a real ow. A Simple Experiment: The Starbucks5 problem Go to the nearest coee pot (or your favorite coee shop) and get a cup of coee. (Note that you are not required to drink it, just play with it.) Then slowly and carefully pour a little cream (or half and half, skim milk probably wont work) into it. Now ever so gently, give it a simple single stir with a stick or a spoon and observe the complex display that you see. Assuming that the cream and coee move together, and that the vorticity (at least for a while) moves like uid material, explain what you see in the light of Example 2 above.
3.6
The Turbulence Closure Problem and the Eddy Viscosity
From the point of view of the averaged motion, at least, the problem with the non-linearity of the instantaneous equations is that they introduce new unknowns,
Starbucks is a very popular chain of coee shops in the USA and many other countries who have only recently discovered what good coee tastes like.
5
3.6. THE EDDY VISCOSITY
51
the so-called Reynolds stress, into the averaged equations. There are six new individual Reynolds stress components we must deal with to be exact: u2 , u2 , 1 2 u2 , u1 u2 , u1 u3 , and u2 u3 . These have to be related to the mean motion itself 3 before the equations can be solved, since the number of unknowns and number of equations must be equal. The absence of these additional equations is often referred to as the Turbulence Closure Problem. A similar problem arose when the instantaneous equations were written (equations 3.1 and 3.2), since relations had to be introduced to relate the stresses (in particular, the viscous stresses) to the motion itself. These relations (or constitutive equations) depended only on the properties of the uid material, and not on the ow itself. Because of this fact, it is possible to carry out independent experiments, called viscometric experiments, in which these uid properties can be determined once and for all. Equation 3.5 provides an example of just such a constitutive relation, the viscosity, , depending only in the choice of uid. For example, once the viscosity of water at given temperature is determined, this value can be used in all ows at that temperature, not just the one in which the evaluation was made. Or for another example, if we are working on a problem of air ow, we only need to go to reference book somewhere and we can nd a complete specication of how the viscosity of air depends on temperature and pressure. Someone somewhere else has already compiled this information from independent experiments. It is tempting to try such an approach for the turbulence Reynolds stresses (even though we know the underlying requirements of scale separation are not satised). For example, a Newtonian type closure for the Reynolds stresses, often referred to as an eddy or turbulent viscosity model, looks like: 1 1 ui uj + ui ui = t Sij Skk ij 3 3
[ ]
(3.23)
where t is the turbulence viscosity (also called the eddy viscosity), and Sij is the mean strain rate dened by: 1 Ui Uj + Sij = 2 xj xi
[ ]
(3.24)
The second term vanishes identically for incompressible ow. For the simple case of a two-dimensional shear ow, equation 3.23 for the Reynolds shear stress reduces to U1 (3.25) u1 u2 = t x2 Note this model is the direct analogy to the Newtonian model for viscous (v) stress in a uid. The Reynolds stresses, ui uj replaces the viscous stress, ij . The counterpart to the mechanical pressure is the mean normal Reynolds stress, ui ui /3. And like its uid counterpart it, the Reynolds stress can depend only on the mean strain rate at a single instant and single location in the ow, so has
52
Figure 3.1: Schematic of axisymmetric plume no history or non-local dependence. This absence will turn out to be fatal in most applications. Moreover, unlike like the viscosity, , which depends only on the uid and not the motion itself, the turbulence viscosity, t , depends entirely on the motion. That such a simple model can adequately describe the mean motion in at least one ow is illustrated by the axisymmetric buoyant plume sketched in Figure 3.1. Figures 3.2 and 3.3 show the calculation of the mean velocity and temperature proles respectively. Obviously the mean velocity and temperature proles are reasonably accurately computed, as are the Reynolds shear stress and lateral turbulent heat ux shown in Figures 3.4 and 3.5. The success of the eddy viscosity in the preceding example is more apparent than real, however, since the value of the eddy viscosity and eddy diusivity (for the turbulent heat ux) have been chosen to give the best possible agreement with the data. This, in itself, would not be a problem if that chosen values could have been obtained in advance of the computation, or even if they could be used to successfully predict other ows. In fact, the values used work only for this ow, thus the computation is not a prediction at all, but a postdiction or hindcast from which no extrapolation to the future can be made. In other words, our turbulence model is about as useful as having a program to predict yesterdays weather. Thus the closure problem still very much remains. Another problem with the eddy viscosity in the example above is that it fails to calculate the vertical components of the Reynolds stress and turbulent heat ux. An attempt at such a computation is shown in Figure 3.6 where the vertical turbulent heat ux is shown to be severely underestimated. Clearly the value of the eddy viscosity in the vertical direction must be dierent than in the radial
53
Figure 3.2: Mean velocity proles for axisymmetric plume
Figure 3.3: Mean temperature proles for axisymmetric plume
54
Figure 3.4: Reynolds shear stress proles for axisymmetric plume
Figure 3.5: Radial turbulent heat ux for axisymmetric plume
55
Figure 3.6: Vertical turbulent heat ux for axisymmetric plume
56
direction. In other words, the turbulence for which a constitutive equation is being written is not an isotropic medium. In fact, in this specic example the problem is that the vertical component of the heat ux is produced more by the interaction of buoyancy and the turbulence, than it is by the working of turbulence against mean gradients in the ow. We will discuss this in more detail in the next chapter when we consider the turbulence energy balances, but note for now that simple gradient closure models never work unless gradient production dominates. This rules out many ows involving buoyancy, and also many involving recirculations or separation where the local turbulence is convected in from somewhere else. A more general form of constitutive equation which would allow for the nonisotropic nature of the medium (in this case the turbulence itself) would be 1 1 ui uj + uk uk ij = ijkl Skl Smm kl 3 3
[ ]
(3.26)
This closure relation allows each component of the Reynolds stress to have its own unique value of the eddy viscosity. It is easy to see that it is unlikely this will solve the closure problem since the original six unknowns, the ui uj , have been traded for eighty-one new ones, ijkl . Even if some can be removed by symmetries, the remaining number is still formidable. More important than the number of unknowns, however, is that there is no independent or general means for selecting them without considering a particular ow. This is because turbulence is indeed a property of the ow, not of the uid.
3.7
The Reynolds Stress Equations
It is clear from the preceding section that the simple idea of an eddy viscosity might not be the best way to approach the problem of relating the Reynolds stress to the mean motion. An alternative approach is to try to derive dynamical equations for the Reynolds stresses from the equations governing the uctuations themselves. Such an approach recognizes that the Reynolds stress is really a functional6 of the velocity; that is, the stress at a point depends on the velocity everywhere and for all past times, not just at the point in question and at a particular instant in time. The analysis begins with the equation for the instantaneous uctuating velocity, equation 3.21. This can be rewritten for a Newtonian uid with constant viscosity as: ui p Ui ui ui ui + Uj = + ij uj uj uj t xj xi xj xj xj xj
[ ]
(v)
(3.27)
Note that the free index in this equation is i. Also, since we are now talking about turbulence again, the capital letters represent mean or averaged quantities.
6
A functional is a function of a function
3.7. THE REYNOLDS STRESS EQUATIONS Multiplying equation 3.27 by uk and averaging yields: ui ui uk + Uj uk = t xj
[ ]
57
p + uk ij (3.28) xi xj [ ] { } Ui ui uk uj uk uj xj xj uk
(v)
Now since both i and k are free indices they can be interchanged to yield a second equation given by7 : uk uk p = ui + ui kj (3.29) ui + Uj ui t xj xk xj [ ] { } uk Uk ui uj ui uj xj xj Equations 3.28 and 3.29 can be added together to yield an equation for the Reynolds stress, ui uk ui uk 1 p p + Uj = ui + uk t xj xk xi [ ] uk ui ui uj + uk uj xj xj
(v) (v)
(v)
(3.30)
1 kj + ui + uk ij xj xj
[
Uk Ui ui uj + uk uj xj xj
It is customary to rearrange the rst term on the right hand side in the following way:
[
p p ui + uk = xk xi
ui uk p + xk xi [pui kj + puk ij ] + xj
(3.31)
The rst term on the right is generally referred to as the pressure strain-rate term. The second term is written as a divergence term, and is generally referred to as the pressure diusion term. We shall see later that divergence terms can never create nor destroy anything; they can simply move it around from one place to another.
Alternatively equation 3.21 can be rewritten with free index k, then multiplied by ui and averaged
7
58
The third term on the right-hand side of equation 3.30 can similarly be rewritten as:
(v) ui kj + uk ij = xj xj (v)
(v) uk ij xj
(v) ui kj xj
(3.32)
(v) (v) [ui kj + uk ij ] xj
The rst of these is also a divergence term. For a Newtonian uid, the last is the so-called dissipation of Reynolds stress by the turbulence viscous stresses. This is easily seen by substituting the Newtonian constitutive relation to obtain: 1 uk ui (v) uk (v) ui ij + kj = 2 sij + skj xj xj xj xj
[ ] [ ]
(3.33)
It is not at all obvious what this has to do with dissipation, but it will become clear later on when we consider the trace of the Reynolds stress equation, which is the kinetic energy equation for the turbulence. Now if we use the same trick from before using the continuity equation, we can rewrite the second term on the right-hand side of equation 3.30 to obtain:
[
ui uk + uk uj = ui uk uj ui uj xj xj xj
(3.34)
This is also a divergence term. We can use all of the pieces we have developed above to rewrite equation 3.30 as:
3.7. THE REYNOLDS STRESS EQUATIONS
59
p ui uk ui uk + Uj ui uk = + t xj xk xi { 1 + [puk ij + pui kj ] ui uk uj xj +2[sij uk + skj ui ]} [ ] Uk Ui ui uj + uk uj xj xj [ ] uk ui 2 sij + skj xj xj
(3.35)
This is the so-called Reynolds Stress Equation which has been the primary vehicle for much of the turbulence modeling eorts of the past few decades. The left hand side of the Reynolds Stress Equation can easily be recognized as the rate of change of Reynolds stress following the mean motion. It seems to provide exactly what we need: nine new equations for the nine unknowns we cannot account for. The problems are all on the right-hand side. These terms are referred to respectively as 1. the pressure-strain rate term 2. the turbulence transport (or divergence) term 3. the production term, and 4. the dissipation term. Obviously these equations do not involve only Ui and ui uj , but depend on many more new unknowns. It is clear that, contrary to our hopes, we have not derived a single equation relating the Reynolds stress to the mean motion. Instead, our Reynolds stress transport equation is exceedingly complex. Whereas the process of averaging the equation for the mean motion introduced only six new independent unknowns, the Reynolds stress, ui uj , the search for a transport equation which will relate these to the mean motion has produced many more unknowns. They are: pui ui sjk sij sjk ui uk uj ui p xj T OT AL 3 unknowns 27 9 27 (3.36) (3.37) (3.38) (3.39) (3.40) (3.41)
9 75
60
Not all of these are independent, since some can be derived from the others. Even so, our goal of reducing the number of unknowns has clearly not been met. Equations governing each of these new quantities can be derived from the original dynamical equations, just as we did for the Reynolds stress. Unfortunately new quantities continue to be introduced with each new equation, and at a faster rate than the increase in the number of equations. Now the full implications of the closure problem introduced by the Reynolds decomposition and averaging has become apparent. No matter how many new equations are derived, the number of new unknown quantities introduced will always increase more rapidly. Our attempt to solve the turbulence problem by considering averages illustrates a general principle. Any time we try to fool Mother Nature by averaging out her details, she gets her revenge by leaving us with a closure problem more equations than unknowns. In thermodynamics, we tried to simplify the consideration of molecules by averaging over them, and were left with the need for an equation of state. In heat transfer, we tried to simplify considerations by which molecules transfer their kinetic energy, and found we were lacking a relation between the heat ux and the temperature eld. And in uid mechanics, we tried to simplify consideration of the mean motion of molecules and ended up with viscous stress. In all of these cases we were able to make simple physical models which worked at least some of the time; e.g., ideal gas, Fourier-Newtonian uid. And these models all worked because we were able to make assumptions about the underlying molecular processes and assume them to be independent of the macroscopic ows of interest. Unfortunately such assumptions are rarely satised in turbulence. It should be obvious by now that the turbulence closure problem will not be solved by the straight-forward derivation of new equations, nor by direct analogy with viscous stresses. Rather, closure attempts will have to depend on an intimate knowledge of the dynamics of the turbulence itself. Only by understanding how the turbulence behaves can one hope to guess an appropriate set of constitutive equations AND understand the limits of them. This is, of course, another consequence of the fact that the turbulence is a property of the ow itself, and not of the uid!
Chapter 4 The Turbulence Kinetic Energy

4.1 The Kinetic Energy of the Fluctuations
It is clear from the previous chapter that the straightforward application of ideas that worked well for viscous stresses do not work too well for turbulence Reynolds stresses. Moreover, even the attempt to directly derive equations for the Reynolds stresses using the Navier-Stokes equations as a starting point has left us with far more equations than unknowns. Unfortunately this means that the turbulence problem for engineers is not going to have a simple solution: we simply cannot produce a set of reasonably universal equations. Obviously we are going to have to study the turbulence uctuations in more detail and learn how they get their energy (usually from the mean ow somehow), and what they ultimately do with it. Our hope is that by understanding more about turbulence itself, we will gain insight into how we might make closure approximations that will work, at least sometimes. Hopefully, we will also gain an understanding of when and why they will not work. An equation for the uctuating kinetic energy for constant density ow can be obtained directly from the Reynolds stress equation derived earlier, equation 3.35, by contracting the free indices. The result is:
[
ui ui + Uj ui ui t xj { } 2 = pui ij q 2 uj + 4sij ui xj ui Ui 4sij 2ui uj xj xj
(4.1)
where the incompressibility condition (uj /xj = 0) has been used to eliminate the pressure-strain rate term, and q 2 ui ui . The last term can be simplied by recalling that the velocity deformation rate tensor, ui /xj , can be decomposed into symmetric and anti-symmetric parts; 61
62 i.e.,
CHAPTER 4. THE TURBULENCE KINETIC ENERGY
ui = sij + ij xj
(4.2)
where the symmetric part is the strain-rate tensor, sij , and the anti-symmetric part is the rotation-rate tensor, ij , dened by: uj 1 ui ij = 2 xj xi
[ ]
(4.3)
Since the double contraction of a symmetric tensor with an anti-symmetric tensor is identically zero, it follows immediately that: sij ui = sij sij + sij ij xj = sij sij
(4.4)
Now it is customary to dene a new variable k, the average uctuating kinetic energy per unit mass, by: 1 1 1 k ui ui = q 2 = [u2 + u2 + u2 ] 2 3 2 2 2 1 (4.5)
By dividing equation 4.1 by 2 and inserting this denition, the equation for the average kinetic energy per unit mass of the uctuating motion can be re-written as:
[
+ Uj k = t xj xj
1 1 pui ij q 2 uj + 2sij ui 2 Ui ui uj 2sij sij xj
(4.6)
The role of each of these terms will be examined in detail later. First note that an alternative form of this equation can be derived by leaving the viscous stress in terms of the strain rate. We can obtain the appropriate form of the equation for the uctuating momentum from equation 3.21 by substituting the incompressible Newtonian constitutive equation into it to obtain:
[
1 p 2 ui Ui ui ui + Uj ui = + 2 uj uj uj t xj xi xj xj xj xj
(4.7)
If we take the scalar product of this with the uctuating velocity itself and average, it follows (after some rearrangement) that:
[
+ Uj k t xj
1 1 = pui ij q 2 uj + k 2 xj ui ui Ui ui uj xj xj xj
xj
(4.8)
4.1. THE KINETIC ENERGY OF THE FLUCTUATIONS
63
Both equations 4.6 and 4.8 play an important role in the study of turbulence. The rst form given by equation 4.6 will provide the framework for understanding the dynamics of turbulent motion. The second form, equation 4.8 forms the basis for most of the second-order closure attempts at turbulence modelling; e.g., the socalled k- models (usually referred to as the k-epsilon models). This because it has fewer unknowns to be modelled, although this comes at the expense of some extra assumptions about the last term. It is only the last term in equation 4.6 that can be identied as the true rate of dissipation of turbulence kinetic energy, unlike the last term in equation 4.8 which is only the dissipation when the ow is homogeneous. We will talk about homogeniety below, but suce it to say now that it never occurs in nature. Nonetheless, many ows can be assumed to be homogeneous at the scales of turbulence which are important to this term, so-called local homogeniety. Each term in the equation for the kinetic energy of the turbulence has a distinct role to play in the overall kinetic energy balance. Briey these are: Rate of change of kinetic energy per unit mass due to non-stationarity; i.e., time dependence of the mean: k (4.9) t Rate of change of kinetic energy per unit mass due to convection (or advection) by the mean ow through an inhomogenous eld : Uj k xj (4.10)
Transport of kinetic energy in an inhomogeneous eld due respectively to the pressure uctuations, the turbulence itself, and the viscous stresses: xj
{
1 1 pui ij q 2 uj + 2sij ui 2
(4.11)
Rate of production of turbulence kinetic energy from the mean ow (gradient): Ui (4.12) ui uj xj Rate of dissipation of turbulence kinetic energy per unit mass due to viscous stresses: 2sij sij (4.13) These terms will be discussed in detail in the succeeding sections, and the role of each examined carefully.
64
4.2
The Rate of Dissipation of the Turbulence Kinetic Energy.
The last term in the equation for the kinetic energy of the turbulence has been identied as the rate of dissipation of the turbulence energy per unit mass; i.e., ui ui ui uj = 2sij sij = + xj xj xj xi
{ }
(4.14)
It is easy to see that 0 always, since it is a sum of the average of squared quantities only (i.e., sij sij 0). Also, since it occurs on the right hand side of the kinetic energy equation for the uctuating motions preceded by a minus sign, it is clear that it can act only to reduce the kinetic energy of the ow. Therefore it causes a negative rate of change of kinetic energy; hence the name dissipation. Physically, energy is dissipated because of the work done by the uctuating viscous stresses in resisting deformation of the uid material by the uctuating strain rates; i.e., (v) = ij sij (4.15) This reduces to equation 4.14 only for a Newtonian uid. In non-Newtonian uids, portions of this product may not be negative implying that it may not all represent an irrecoverable loss of uctuating kinetic energy. It will be shown in Chapter 5 that the dissipation of turbulence energy mostly takes place at the smallest turbulence scales, and that those scales can be characterized by the so-called Kolmogorov microscale dened by:
(
)1/4
(4.16)
In atmospheric motions where the length scale for those eddies having the most turbulence energy (and most responsible for the Reynolds stress) can be measured in kilometers, typical values of the Kolmogorov microscale range from 0.110 millimeters. In laboratory ows where the overall scale of the ow is greatly reduced, much smaller values of K are not uncommon. The small size of these dissipative scales greatly complicates measurement of energy balances, since the largest measuring dimension must be about equal to twice the Kolmogorov microscale. And it is the range of scales, L/, which makes direct numerical simulation of most interesting ows impossible, since the required number of computational cells is several orders of magnitude greater than (L/)3 . This same limitation also aects experiments as well, which must often be quite large to be useful. One of the consequences of this great separation of scales between those containing the bulk of the turbulence energy and those dissipating it is that the dissipation rate is primarily determined by the large scales and not the small. This is because the viscous scales (which operate on a time scale of tK = (/)1/2 ) dissipate rapidly any energy sent down to them by the non-linear processes of
4.2. THE DISSIPATION RATE
65
scale to scale energy transfer. Thus the overall rate of dissipation is controlled by the rate of energy transfer from the energetic scales, primarily by the non-linear scale-to-scale transfer. This will be discussed later when we consider the energy spectrum. But for now it is important only note that a consequence of this is that the dissipation rate is given approximately as: u3 L (4.17)
where u2 q 2 /3 and L is something like an integral length scale. It is easy to remember this relation if you note that the time scale of the energetic turbulent eddies can be estimated as L/u. Thus dk/dt = (3/2)du2 /dt can estimated as (3u2 /2)/(L/u). Sometimes it is convenient to just dene the length scale of the energycontaining eddies (or the pseudo-integral scale) as: l u3 (4.18)
Almost always, l L, but the relation is at most only exact theoretically in the limit of innite Reynolds number since the constant of proportionality is Reynolds number dependent. Some just assume ratio to be constant (and even universal), and even refer to l as though it were the real integral scale. Others believe that the experimental scatter observed in the constant is because of the diering upstream conditions and that the ratio may not be constant at all. It is really hard to tell who is right in the absence of facilities or simulations in which the Reynolds number can vary very much for xed initial conditions. Nonetheless, there is really overwhelming evidence (experimental and theoretical) that this ratio depends very much on the type of ow being considered. This all may leave you feeling a bit confused, but thats the way turbulence is right now. Its a lot easier to teach if we just tell you one view, but thats not very good preparation for the future. Here is what we can say for sure. Only the integral scale, L, is a physical length scale, meaning that it can be directly observed in the ow by spectral or correlation measurements (as shown later). The pseudo-integral scale, l, on the other hand is simply a denition; and it is only at innite turbulence Reynolds number that it has any real physical signicance. But it is certainly a useful approximation at large, but nite, Reynolds numbers. We will talk about these subtle but important distinctions later when we consider homogeneous ows, but it is especially important when considering similarity theories of turbulence. For now simply le away in your memory a note of caution about using equation 4.17 too freely. And do not be fooled by the cute description this provides. It is just that, a description, and not really an explanation of why all this happens sort of like the weather man describing the weather. Using equation 4.18, the Reynolds number dependence of the ratio of the
66
Kolmorgorov microscale, K , to the pseudo-integral scale, l, can be obtained as: K 3/4 = Rl (4.19) l where the turbulence Reynolds number, Rl , is dened by: ul u4 Rl = (4.20) Example Estimate the Kolmogorov microscale for u = 1 m/s and L = 0.1m for air and water. air For air, Rl = 1 (0.1)/15 106 7 103 . Therefore l/K 8 102 , so K 1.2 104 m or 0.12mm. water For water, Rl = 1 (0.1)/106 105 . Therefore l/K 5 103 , so K 2 105 m or 0.02mm.
Exercise: Find the dependence on Rl of the time-scale ration between the Kolmorogov microtime and the time scale of the energy-containing eddies. Thus the dissipative scales are all much smaller than those characterizing the energy of the turbulent uctuations, and their relative size decreases with increasing Reynolds number. Note that in spite of this, the Kolmogorov scales all increase with increasing energy containing scales for xed values of the Reynolds number. This fact is very important in designing laboratory experiments at high turbulence Reynolds number where the nite probe size limits spatial resolution. The rather imposing size of some experiments is an attempt to cope with this problem by increasing the size of the smallest scales, thus making them larger than the resolution limits of the probes being used. Exercise: Suppose the smallest probe you can build can only resolve 0.1 mm. Also to do an experiment which is a reasonable model of a real engineering ow (like a hydropower plant), you need (for reason that will be clear later) a scale separation of at least L/K = 104 . If your facility has to be at least a factor of ten larger than L (which you estimate as l), what is its smallest dimension? It will also be argued later that these small dissipative scales of motion at very high Reynolds number tend to be statistically nearly isotropic; i.e., their statistical character is independent of direction. We will discuss some of the implications of isotropy and local isotropy later, but note for now that it makes possible a huge reduction in the number of unknowns, particularly those determined primarily by the dissipative scales of motion.
4.3. THE PRODUCTION
67
4.3
The Kinetic Energy of the Mean Motion and the Production of Turbulence
An equation for the kinetic energy of the mean motion can be derived by a procedure exactly analogous to that applied to the uctuating motion. The mean motion was shown in equation 3.19 to be given by: T Ui P Ui + Uj = + ij (ui uj ) t xj xi xj xj
[ ]
(v)
(4.21)
By taking the scalar product of this equation with the mean velocity, Ui , we can obtain an equation for the kinetic energy of the mean motion as:
[
Ui
Ui Tij ui uj Ui P + Uj Ui = + Ui t xj xi xj xj
(v)
(4.22)
Unlike the uctuating equations, there is no need to average here, since all the terms are already averages. In exactly the same manner that we rearranged the terms in the equation for the kinetic energy of the uctuations, we can rearrange the equation for the kinetic energy of the mean ow to obtain:
[
+ Uj K= xj { } 1 1 P Ui ij ui uj Ui + 2Sij Ui xj 2 Ui 2Sij Sij +ui uj xj
(4.23)
1 1 K Q2 = Ui Ui (4.24) 2 2 The role of all of the terms can immediately be recognized since each term has its counterpart in the equation for the average uctuating kinetic energy. Comparison of equations 4.23 and 4.6 reveals that the term ui uj Ui /xj appears in the equations for the kinetic energy of BOTH the mean and the uctuations. There is, however, one VERY important dierence. This production term has the opposite sign in the equation for the mean kinetic energy than in that for the mean uctuating kinetic energy! Therefore, whatever its eect on the kinetic energy of the mean, its eect on the kinetic energy of the uctuations will be the opposite. Thus kinetic energy can be interchanged between the mean and uctuating motions. In fact, the only other term involving uctuations in the equation for the kinetic energy of the mean motion is a divergence term; therefore it can only move the kinetic energy of the mean ow from one place to another.
where
68
Therefore this production term provides the only means by which energy can be interchanged between the mean ow and the uctuations. Understanding the manner in which this energy exchange between mean and uctuating motions is accomplished represents one of the most challenging problems in turbulence. The overall exchange can be understood by exploiting the analogy which treats ui uj as a stress, the Reynolds stress. The term: ui uj Ui /xj (4.25)
can be thought of as the working of the Reynolds stress against the mean velocity gradient of the ow, exactly as the viscous stresses resist deformation by the instantaneous velocity gradients. This energy expended against the Reynolds stress during deformation by the mean motion ends up in the uctuating motions, however, while that expended against viscous stresses goes directly to internal energy. As we have already seen, the viscous deformation work from the uctuating motions (or dissipation) will eventually send this uctuating kinetic energy on to internal energy as well. Now, just in case you are not all that clear exactly how the dissipation terms really accomplish this for the instantaneous motion, it might be useful to examine exactly how the above works. We begin by decomposing the mean deformation rate tensor Ui /xj into its symmetric and antisymmetric parts, exactly as we did for the instantaneous deformation rate tensor in Chapter 3; i.e., Ui = Sij + ij xj where the mean strain rate Sij is dened by 1 Ui Uj Sij = + 2 xj xi and the mean rotation rate is dened by 1 Ui Uj ij = 2 xj xi
[ ] [ ]
(4.26)
(4.27)
(4.28)
Since ij is antisymmetric and ui uj is symmetric, their contraction is zero so it follows that: Ui = ui uj Sij (4.29) ui uj xj Equation 4.29 is an analog to the mean viscous dissipation term given for incompressible ow by: Tij
(v) Ui
xj
= Tij Sij = 2Sij Sij
(v)
(4.30)
It is easy to show that this term transfers (or dissipates) the mean kinetic energy directly to internal energy, since exactly the same term appears with the opposite
4.4. THE TRANSPORT (OR DIVERGENCE) TERMS
69
sign in the internal energy equations. Moreover, since Sij Sij 0 always, this is a one-way process and kinetic energy is decreased while internal energy is increased. Hence it can be referred to either as dissipation of kinetic energy, or as production of internal energy. As surprising as it may seem, this direct dissipation of energy by the mean ow is usually negligible compared to the energy lost to the turbulence through the Reynolds stress terms. (Remember, there is a term exactly like this in the kinetic energy equation for the uctuating motion, but involving only uctuating quantities; namely, 2sij sij .) We shall show later that for almost always in turbulent ow, sij sij >> Sij Sij . What this means is that the energy dissipation in a turbulent ow is almost entirely due to the turbulence. There is a very important dierence between equations 4.29 and 4.30. Whereas the eect of the viscous stress working against the deformation (in a Newtonian uid) is always to remove energy from the ow (since Sij Sij > 0 always), the eect of the Reynolds stress working against the mean gradient can be of either sign, at least in principle. That is, it can either transfer energy from the mean motion to the uctuating motion, or vice versa. Almost always (and especially in situations of engineering importance), ui uj and Sij have the opposite sign. Therefore, ui uj Sij > 0 almost always, so kinetic energy is removed from the mean motion and added to the uctuations. Since the term ui uj Ui /xj usually acts to increase the turbulence kinetic energy, it is usually referred to as the rate of turbulence energy production, or simply the production. Now that we have identied how the averaged equations account for the production of turbulence energy from the mean motion, it is tempting to think we have understood the problem. In fact, labeling phenomena is not the same as understanding them. The manner in which the turbulence motions cause this exchange of kinetic energy between the mean and uctuating motions varies from ow to ow, and is really very poorly understood. Saying that it is the Reynolds stress working against the mean velocity gradient is true, but like saying that money comes from a bank. If we want to examine the energy transfer mechanism in detail we must look beyond the single point statistics, so this will have to be a story for another time. Example: Consider how the production term looks if the Reynolds stress is modelled by an turbulent viscosity.
4.4
The Transport (or Divergence) Terms
The overall role of the transport terms is best understood by considering a turbulent ow which is completely conned by rigid walls as in Figure 4.1. First consider only the turbulence transport term. If the volume within the connement is denoted by Vo and its bounding surface is So , then rst term on the right-hand
70
Figure 4.1: Turbulence conned by rigid walls.
side of equation 4.6 for the uctuating kinetic energy can be integrated over the volume to yield: 1 1 pui ij q 2 uj + sij ui dV 2 Vo xj [ ] 1 1 2 = pui ij q uj + sij ui nj dS 2 So
[ ]
(4.31)
where we have used the divergence theorem again! We assumed our enclosure to have rigid walls; therefore the normal component of the mean velocity (un = uj nj ) must be zero on the surface since there can be no ow through it (the kinematic boundary condition). This immediately eliminates the contributions to the surface integral from the puj nj and q 2 uj nj terms. But the last term is zero on the surface also. This can be seen in two ways: either by invoking the no-slip condition which together with the kinematic boundary condition insures that ui is zero on the boundary, or by noting from Cauchys theorem that sij nj is the viscous contribution to the normal contact force per unit area on the surface (i.e., t(v) ) whose scalar product with ui must be identically n zero since un is zero. Therefore the entire integral is identically zero and its net contribution to the rate of change of kinetic energy is zero. Thus the only eect of the turbulence transport terms (in a xed volume at least) can be to move energy from one place to another, neither creating nor destroying it in the process. This is, of course, why they are collectively called the transport terms. This spatial transport of kinetic energy is accomplished by the acceleration of adjacent uid due to pressure and viscous stresses (the rst
4.4. THE TRANSPORT (OR DIVERGENCE) TERMS
71
and last terms respectively), and by the physical transport of uctuating kinetic energy by the turbulence itself (the middle term). This role of these turbulence transport terms in moving kinetic energy around is often exploited by turbulence modellers. It is argued, that on the average, these terms will only act to move energy from regions of higher kinetic energy to lower. Thus a plausible rst-order hypothesis is that this diusion of kinetic energy should be proportioned to gradients of the kinetic energy itself. That is, 1 1 k puj q 2 uj + sij ui = ke 2 xj (4.32)
where ke is an eective diusivity like the eddy viscosity discussed earlier. If we use the alternative form of the kinetic energy equation (equation 4.8), there is no need to model the viscous term (since it involves only k itself). Therefore our model might be: 1 1 k puj q 2 uj = kealt (4.33) 2 xj These, of course, look much more complicated in a real model because of the need to insure proper tensorial invariance, etc., but the physics is basically the same. If you think about it, that such a simple closure is worth mentioning at all is pretty amazing. We took 9 unknowns, lumped them together, and replaced their net eect by the simple gradient of something we did know (or at least wanted to calculate), k. And surprisingly, this simple idea works pretty well in many ows, especially if the value of the turbulent viscosity is itself related to other quantities like k and . In fact this simple gradient hypothesis for the turbulence transport terms is at the root of all engineering turbulence models. There are a couple of things to note about such simple closures though, before getting too enthused about them. First such an assumption rules out a countergradient diusion of kinetic energy which is known to exist in some ows. In such situations the energy appears to ow up the gradient. While this may seem unphysical, remember we only assumed it owed down the gradient in the rst place. This is the whole problem with a plausibility argument. Typically energy does tend to be transported from regions of high kinetic energy to low kinetic energy, but there is really no reason for it always to do so, especially if there are other mechanisms at work. And certainly there is no reason for it to always be true locally, and the gradient of anything is a local quantity. Let me illustrate this by a simple example. Lets apply a gradient hypothesis to the economy a plausibility hypothesis if you will. By this simple model, money would always ow from the rich who have the most, to the poor who have the least. In fact, as history has shown, in the absence of other forces (like revolutions, beheadings, and taxes) this almost never happens. The rich will always get richer, and the poor poorer. And the reason is quite simple, the poor are usually borrowing (and paying interest), while the rich are doing the loaning (and collecting interest). Naturally there are individual exceptions and
72
great success stories among the poor. And there are wealthy people who give everything away. But mostly in a completely free economy, the money ows in a counter-gradient manner. So society (and the rich in particular) have a choice risk beheading and revolution, or nd a peaceful means to redistribute the wealth like taxes. While the general need for the latter is recognized (especially among those who have and pay the least), there is, of course, considerable disagreement of how much tax is reasonable to counter the natural gradient. Just as the simple eddy viscosity closure for the mean ow can be more generally written as a tensor, so can it be here. In fact the more sophisticated models write it as second or fourth-order tensors. More importantly, they include other gradients in the model so that the gradient of one quantity can inuence the gradient of another. Such models can sometimes even account for counter-gradient behavior. If your study of turbulence takes you into the study of turbulence models, watch for these subtle dierences among them. And dont let yourself be annoyed or intimidated by their complexity. Instead marvel at the physics behind them, and try to appreciate the wonderful manner in which mathematics has been used to make them properly invariant so you dont have to worry about whether they work in any particular coordinate system. It is all these extra terms that give you reason to hope that it might work at all.
4.5
The Intercomponent Transfer of Energy
The objective of this section is to examine how kinetic energy produced in one velocity component of the turbulence can be transferred to the other velocity components of the uctuating motion. This is very important since often energy is transferred from the mean ow to a only a single component of the uctuating motion. Yet somehow all three components of the kinetic energy end up being about the same order of magnitude. The most common exception to this is very close to surfaces where the normal component is suppressed by the kinematic boundary condition. To understand what is going on, it is necessary to develop even a few more equations; in particular, equations for each component of the kinetic energy. The procedure is almost identical to that used to derive the kinetic energy equation itself. Consider rst the equation for the 1-component of the uctuating momentum. We can do this by simply setting i = 1 and k = 1 in the equation 3.35, or derive it from scratch by setting the free index in equation 3.27 equal to unity (i.e. i=1); i.e.,
[
u1 1 p 1 1j U1 u1 u1 u1 + Uj = + uj uj uj t xj x1 xj xj xj xj
(v)
(4.34)
Multiplying this equation by u1 , averaging, and rearranging the pressure-velocity gradient term using the chain rule for products yields:
4.5. THE INTERCOMPONENT TRANSFER OF ENERGY 1- component

[
73
= + +
1 2 Uj u xj 2 1 u1 p x1 { } 1 1 2 pu1 1j u1 uj + 2s1j u1 xj 2 U1 2s1j s1j u1 uj xj
(4.35)
All of the terms except one look exactly like the their counterparts in equation 4.6 for the average of the total uctuating kinetic energy. The single exception is the rst term on the right-hand side which is the contribution from the pressure-strain rate. This will be seen to be exactly the term we are looking for to move energy among the three components. Similar equations can be derived for the other uctuating components with the result that 2- component
[ ]
= + + and
1 2 Uj u xj 2 2 u2 p x2 { } 1 1 2 pu2 2j u2 uj + 2s2j u2 xj 2 U2 u2 uj 2s2j s2j xj
(4.36)
74 3- component
[
= + +
1 2 Uj u xj 2 3 u3 p x3 { } 1 1 2 pu3 3j u3 uj + 2s3j u3 xj 2 U3 u3 uj 2s3j s3j xj
(4.37)
Note that in each equation a new term involving a pressure-strain rate has appeared as the rst term on the right-hand side. It is straightforward to show that these three equations sum to the kinetic energy equation given by equation 4.6, the extra pressure terms vanishing for the incompressible ow assumed here. In fact, the vanishing of the pressure-strain rate terms when the three equations are added together gives a clue as to their role. Obviously they can neither create nor destroy kinetic energy, only move it from one component of the kinetic energy to another. The precise role of the pressure terms can be seen by noting that incompressibility implies that: uj p =0 (4.38) xj It follows immediately that: u2 u1 u3 p = p + p x1 x2 x3 Thus equation 4.35 can be written as:
[ ] [ ]
(4.39)
= +
1 2 Uj u xj 2 1 [ ] u2 u3 p + p x2 x3 { } 1 1 2 pu1 1j u1 uj + 2s1j u1 xj 2 U1 u1 uj 2s1j s1j xj
(4.40)
Comparison of equation 4.40 with equations 4.36 and 4.37 makes it immediately apparent that the pressure strain rate terms act to exchange energy between components of the turbulence. If pu2 /x2 and pu3 /x3 are both positive, then energy is removed from the 1-equation and put into the 2- and 3-equations since the same terms occur with opposite sign. Or vice versa.
4.5. THE INTERCOMPONENT TRANSFER OF ENERGY
75
The role of the pressure strain rate terms can best be illustrated by looking at a simple example. Consider a simple homogeneous shear ow in which Ui = U (x2 )1i and in which the turbulence is homogeneous. For this ow, the assumption of homogeneity insures that all terms involving gradients of average quantities vanish (except for dU1 /dx2 ). This leaves only the pressure-strain rate, production and dissipation terms; therefore equations 4.36, 4.37 and 4.40 reduce to: 1-component: u2 u3 U1 u2 1 = p + p u1 u2 1 t x2 x3 x2 2-component: u2 2 = t 3-component: u2 3 = t where 1 2s1j s1j 2 2s2j s2j 3 2s3j s3j (4.44) (4.45) (4.46) +p u2 x2 u3 x3 2 (4.42)
[ ]
(4.41)
+p
(4.43)
It is immediately apparent that only u2 can directly receive energy from the 1 mean ow because only the rst equation has a non-zero production term. Now lets further assume that the smallest scales of the turbulence can be assumed to be locally isotropic. While not always true, this is a pretty good approximation for high Reynolds number ows. (Note that it might be exactly true in many ows in the limit of innite Reynolds number, at least away from walls.) Local isotropy implies that the component dissipation rates are equal; i.e., 1 = 2 = 3 . But where does the energy in the 2 and 3-components come from? Obviously the pressure-strain-rate terms must act to remove energy from the 1-component and redistribute it to the others. As the preceding example makes clear, the role of the pressure-strain-rate terms is to attempt to distribute the energy among the various components of the turbulence. An easy way to remember this is to think of the pressure strain rate terms as the Robin Hood terms: they steal from the rich and give to the poor. In the absence of other inuences, they are so successful that the dissipation by each component is almost equal, at least at high turbulence Reynolds numbers. In fact, because of the energy re-distribution by the pressure strain rate terms, it is uncommon to nd a turbulent shear ow away from boundaries where the kinetic energy of the turbulence components dier by more than 30-40%, no matter which component gets the energy from the mean ow.
76
Example: In simple turbulent free shear ows like wakes or jets where the energy is primarily produced in a single component (as in the example above), typically u2 u2 + u2 where u2 is the kinetic of the component produced directly 1 2 3 1 by the action of Reynolds stresses against the mean velocity gradient. Moreover, u2 u2 . This, of course, makes some sense in light of the above, since both 2 3 o-axis components get most of their energy from the pressure-strain rate terms. It is possible to show that the pressure-strain rate terms vanish in isotropic turbulence. This suggests (at least to some) that the natural state for turbulence in the absence of other inuences is the isotropic state. This has also been exploited by the turbulence modelers. One of the most common assumptions involves setting these pressure-strain rate terms (as they occur in the Reynolds shear equation) proportional to the anisotropy of the ow dened by: aij = ui uj q 2 ij /3 (4.47)
Models accounting for this are said to include areturn-to-isotropy term. An additional term must also be included to account for the direct eect of the mean shear on the pressure-strain rate correlation, and this is referred to as the rapid term. The reasons for this latter term are not easy to see from single point equations, but fall out rather naturally from the two-point Reynolds stress equations we shall discuss later.
Chapter 5 A First Look at Homogeneous Turbulence:

5.1 Introduction
The main purpose of this chapter is to examine the single point equations developed in the preceding chapters as they apply to homogeneous turbulence. Since we really havent dened yet what we mean by homogenous lets do that now. By homogeneous we mean the statistics of the turbulence are independent of the physical origin in space. In other words, no matter where we put the coordinate system, we get the same numbers. For single point statistical quantities, this means they must be constant in the homogeneous directions (however many there are). For example, in a homogeneous ow which is homogeneous in all directions, u2 ( ) = u2 ( ) for all and x , so we could just drop the -argument comx x 1 x 1 x 2 pletely and call it u1 . An obvious consequence of this which we have already noted in previous chapters is that all spatial gradients of a homogeneous process must be zero. We shall see later that another important consequence of this is that we can approximate true ensemble averages by averages over homogeneous directions in space instead. A closely related concept to the concept of homogeneity is that of stationarity. A stationary random process is one in which the statistics are independent of origin in time. And this of course means all single-time averages must be time independent; i.e., u(t) = U , [(t) U ]2 = u2 , etc. Many experiments are u approximately stationary random processes, or at least we try to perform them that way. The reason we shall see later in Chapter 8 is that it greatly simplies averaging by allowing us to time-average instead of performing the experiment many times to build up an ensemble over which to average at an instant. This works for stationary process since in principle the statistics are time-independent. We only need to insure that we have enough independent pieces of information, since the variable at one instant can still be correlated with that at another. This is because stationarity does not imply that two-time statistics like u(t)u(t ) are time-independent, only that they are independent of origin in time. Hence 77
78
CHAPTER 5. A FIRST LOOK AT HOMOGENEOUS TURBULENCE:
they could be a function of the time dierence t t. Analogously, homogeneous statistics involving two locations (e..g. two-point statistics like u( )u( )) can x x depend on the separation vector . We will have much more to say about x x this in later chapters, since two-point statistics are the key to understanding the dierent scales of turbulence motions and how they interact. Homogeneous turbulence is an important subject in its own right but can only be properly discussed with much more powerful statistical tools than we have discussed so far. These will receive considerable attention later. Nonetheless it important to discuss homogeneous turbulencet here briey since so many of the ideas commonly believed about turbulence come from our study of it. And another reason is that in fact you should be impressed by how little we actually know for sure about it. Most books and discussions on this subject present all of this believed stu as fact. Hopefully you will not only be convinced that little is really fact, but that there are plenty of opportunities for you to contribute to this eld.1 If nothing else, it should be obvious after reading this chapter that most of the simple turbulence models should not be expected to work too well, since they really cant predict much of the observed behavior, at least not without assuming the answer at the outset. And of course, this begs the bigger question: if the models dont work too well for simple homogeneous ows, can they ever be trusted for more complicated ones? The answer is: perhaps (in the absence of any alternatives) they are useful as engineering tools, but they certainly do not represent a scientic understanding of the observed phenomena.
5.2
Why are homogeneous ows important?
We have been able to gain considerable insight into how turbulent ows behave by considering in the preceding chapters the single point energy and Reynolds stress equations. For example, we now know how to recognize the dierence between the unsteady (or non-stationary) contribution to the derivative following the mean motion and that due to convection through an inhomogeneous eld by the mean ow (section 4.1. Obviously if the eld is stationary or homogeneous, one or the other of these terms vanishes. We also learned that the so-called transport terms just move energy around, and that they vanish in homogeneous ows (section 4.4). Then we saw that there are some terms which can remain, even if the ow is homogeneous (e.g., sections 4.2 and 4.3). We, in fact, already used a simple unsteady homogeneous example to illustrate how the turbulence obtains energy from the mean ow via the Reynolds stress working against the mean ow gradient (section 4.5). And we used the same examI remember my own experience as a student learning turbulence for the rst time lamenting that there was nothing about this beautiful subject left for me to do, everything was known. So I turned my attention to free shear ows in the hopes I would nd some small thing I could contribute. Imagine my surprise years later when I discovered the truth! I vowed that I would never leave students thiniking this way, at least until I was positive something was true.
1
5.2. WHY ARE HOMOGENEOUS FLOWS IMPORTANT?
79
ple to illustrate how the pressure strain-rate terms describe the process by which all the components are supplied with energy, even if all the energy is produced in only one of them. Finally, we saw saw that the energy is dissipated primarily by the turbulence at the very smallest scales of motion, which themselves have the least energy but are often nearly locally homogeneous (section 4.2). Clearly homogeneous ows must be of considerable interest since they allow us to examine the behavior of the terms which do not vanish without additional complications of those which do. Unfortunately, like many of the ows we consider in this course, homogeneous turbulence is an ideal state which can at best be approximated in a wind tunnel or in the computer. Nonetheless it is still of considerable interest to the engineer and the physicist, to the turbulence modeler and to the theoretician. It is of interest to engineers and turbulence modelers since it oers a simple way to test whether some terms have been modeled correctly without the complications of the others, or more importantly, without the complications of boundary conditions. But, by contrast, such ows are of interest to physicists and mathematicians because they appear to be anomalous, meaning the results do not seem to be completely consistent with our ideas about how they should behave. These anomalies of homogeneous turbulence present a dilemma for the modelers and engineers: if they are truly anomalous, should they even try to account for the results at all with their models? Many think not, and many in fact do not. Even the ones who believe the results of homogeneous ows are important to modeling are very selective about which experimental results they use, and assume all the other experiments to be wrong or simply ignore the ones they dont like. But this creates a problem for those actually trying to nd the truth, since new ndings and insights can be treated quite harshly because these new ideas can be quite threatening to those who have invested so much time building models based the old ones. These anomalies of homogeneous ows also present a real dilemma for the physicists and mathematicians. Anomalies in science sometimes mean there is something big is about to be discovered. One possibility is that we have been doing something very wrong in our experiments and computations, and we hopefully are about to learn what. Or even more exciting, there may be something very wrong with our ideas, and we might be about to have some new ones which may completely change our understanding of how the world is. The best example of the latter possibility is illustrated by the general assessment of physics near the end of the 19th century. Scientists then thought they had learned all the fundamental ideas of physics, with nothing important left to be done but engineering applications of well-established principles. All that was left, some said, was to account for two anomalous experiments the RayleighJeans experiment2 and the Mickelson-Morley experiment3 and many thought
The Rayleigh-Jeans experiment measured thermal radiation from a black box The Mickelson-Morley experiment measured the speed of light and showed it to be apparently independent of the speed of the source.
3 2
80
these experiments to be in error. The resolution of these two anomalies gave us quantum mechanics and relativity, both of which completely changed our picture of the world around us and had dramatic inuences on engineering and life on this planet. Will such be the case with our view of turbulence when we resolve the remaining dilemmas of homogeneous ows in turbulence? Or will we simply nd out that most of the conicting experiments have been wrong. The search for answers will be exhilarating for some, and frustrating for others, but necessary for all especially for those like you entering the eld now. If we nd the answers, they may completely change how we do things, like quantum mechanics or relativity did. Or we may nd they have little eect at all and our old way of doing things is the best we can do. A good example of this was the recognition by Copernicus that the earth was not the center of the universe. The new world view completely changed the way we think about the universe, but had no eect on navigation. Sailors still nd their way across great oceans using the principles of Ptolemaic navigation which is based on the idea that the earth is at the center.4 My personal suspicion is that things will be changed a lot as our understanding of these anomalies grows, simply because our present turbulence models really dont extrapolate too well to new problems. Unlike the example of Ptolemaic navigation which works perfectly well on the surface of the earth, our turbulence models are really very reliable, especially at predicting new things we havent built into them. This probably means we still have much to learn about turbulence, and as we learn our ideas will change and our models improve. So we might as well begin this learning process with the problem which is at least in principle the easiest: homogeneous turbulent ows.
5.3
Decaying turbulence: a brief overview
Look, for example, at the decay of turbulence which has already been generated. If this turbulence is homogeneous and there is no mean velocity gradient to generate new turbulence, the kinetic energy equation reduces to simply: d k = dt (5.1)
Note that the time derivative is just an ordinary derivative, since there is no dependence of any single point quantity on position. This is often written (especially for isotropic turbulence) as: ] [ d 3 2 (5.2) u = dt 2 where 3 k u2 (5.3) 2
I can personally testify that this works, since I arrived in Kinsale, Ireland after an arduous voyage across the Atlantic from America in my 42 foot sailboat Wings by exactly this method.
4
5.3.
DECAYING TURBULENCE: A BRIEF OVERVIEW
81
Now you cant get any simpler than this. Yet unbelievably we still dont have enough information to solve it. Lets try. Suppose we use the extended ideas of Kolmogorov we introduced in Chapter 4 to relate the dissipation to the turbulence energy, say: u3 = f (Re) (5.4) l Already you can see we have two problems, what is f (Re), and what is the time dependence of l? Now there is practically a dierent answer to these questions for every investigator in turbulence most of whom will assure you their choice is the only reasonable one. Figure 5.1 shows two attempts to correlate some of the grid turbulence data (1) using the longitudinal integral scale for l, i.e., l = L11 , or simply L. The rst thing you notice is the problem at low Reynolds number. The second is probably the possible asymptote at the higher Reynolds numbers. And the third is probably the scatter in the data, which is characteristic of most turbulence experiments, especially if you try to compare the results of one experiment to the other. Lets try to use the apparent asymptote at high Reynolds number to our advantage by arguing that f (Re) A, where A is a constant. Note that this limit is consistent with the Kolmogorov argument we made back when we were talking about the dissipation earlier, so we might feel on pretty rm ground here, at least at high turbulent Reynolds numbers. But before we feel too comfortable about this, we should note that a careful examination of the data suggests that the asymptote depends on the details of how the experiment was forced at the large scales of motion (e.g. which grid was used, etc.) . This is not good, since if true it means that the answer depends on the particular ow exactly what we wanted to avoid by modelling in the rst place. Nonetheless, lets proceed by assuming in spite of the evidence that A 1 and L is the integral scale. Now how does L vary with time? Figure 5.2 shows the ratio of the integral scale to the Taylor microscale from the famous ComteBellot/Corrsin (1971) experiment. One might assume, with some theoretical justication, that L/ const. This would be nice since you will be able to show that if the turbulence decays as a power law in time, say u2 tn , then t1/2 . But as shown in Figure 5.3 from Wang et al (2000), this is not a very good assumption for the DNS data available at that time. Now I believe this is because of problems in the simulations, mostly having to do with the fact that turbulence in a box is not a very good approximation for truly homogeneous turbulence unless the size of the box is much larger than the energetic scales. Figure 5.4 shows what happens if you try to correct for the nite box size, and now the results look pretty good. You can see immediately that if I am right and L t1/2 then u2 t1 . Now any careful study of the data will convince you that the energy indeed decays as a power law in time, but there is no question that n = 1, but n < 1, at least for most of the experiments. Most people have tried to x this problem changing p. But I say the problem is in f (Re) and the assumption that u3 /L
82
Figure 5.1: Collection of experimental data showing relation of the longitudianl 1) (1) (1) integral scale, L1,1 , to l = u3 /; i.e., = L1,1 /l = L1,1 /u3 (from Gamard and George 2000).
5.3.
83
Figure 5.2: Plots of longitudinal integral scale (L1,1 = Lf ) and lateral integral (1) scale (L2,2 = Lg ) to the Taylor microscale, g (data of Comte-Bellot and Corrsin 1971).
(1)
84
constant exponent region |< >|
4 L/ 3
00000000 L/ corrected Wray ............... Lm/m uncorrected Wray
L/=4.397
4 time
Figure 5.3: Ratio of longitudinal integral scale to the Taylor microscale, g from recent DNS of isotropic decaying turbulence by Wray (1998), from Wang and George 2003). Note how dierent the results are when the large scales missing due to the nite computational box are accounted for.
5.3.
85
4.5
|<constant exponent region > |
3.5
2.5 L/ 2 1.5 oooooooo L/ corrected deBruyn Kops/Riley 1 ................. Lm/m corrected deBruyn Kops/Riley L/=3.400 0.5 0 0
0.1
0.2
0.3
0.4
0.5 time
0.6
0.7
0.8
0.9
Figure 5.4: Ratio of longitudinal integral scale to the Taylor microscale, g from recent DNS of isotropic decaying turbulence by DeBruyn Kops and Riley (1999), from Wang and George 2003). Note how dierent the results are when the large scales missing due to the nite computational box are accounted for.
86
at nite Reynolds numbers. I would argue that n 1 only in the limit of innite Reynolds number. To see why I believe this, try doing the problem another way. We will show in a later chapter that if the turbulence decays as a power law, then the Taylor microscale, g , must be proportional exactly to t1/2 . Thus we must have (assuming isotropy): k dk k = 10 2 dt g t It is easy to show that k tn where n is given by: d2 10 g = dt n (5.6) (5.5)
and any value of n 1 is acceptable. Obviously the dierence lies in the use of the relation u3 /L at nite Reynolds numbers. Figure 5.5 shows data taken downstream of a grid by Genevieve Comte-Bellot and Stanley Corrsin in one of the most famous turbulence experiments ever. The experiments were carried out at the Johns Hopkins University while I was a student there, and I can still remember the huge hand-drawn graphs spread out over the long coee table. Genevieve was using them to determine whether data satised a power law decay, which in the days before modern computer graphics was the way you had to do things. You can judge from yourself from the gure, but clearly a power law of n = 1.28 works pretty well. Believe it or not, this whole subject is one of the really hot debates of the last decade, and may well be for the next as well. The reason, at least in part, is that the value of the decay exponent, n, seems to a function of how the ow started (e.g., which grid, what Reynolds number, etc.). Until quite recently actually (George 1992 Phys. Fluids was the turning point) people believed (hoped is perhaps a more accurate description) that such turbulence had universal decay characteristics. There is pretty convincing evidence that it does not. As gure ?? makes clear, there is even convincing evidence from recent experiments at Imperial College of London by Christos Vassilicos and his co-workers that some kinds of grids (a particular class of fractal grids) like the one shown in gure ?? dont decay as power-laws at all, but in fact decay exponentially. If you examine the equations above, you can see that this can happen only if the Taylor microscale remains constant during decay, so that the time derivative of the kinetic energy is proportional to the kinetic energy. Figure ?? show that this is exactly what happens. Interestingly, as shown in George 1992 (Physics of Fluids) and George and Wang (2009), both kinds of decay, power-law and exponential, are consistent with an equilibrium similarity analysis of the spectral energy equations. And the theory even predicts the observed dependence on initial (or upstream) conditions. But while this shows that what we observe is consistent with the equations and how the turbulence generally evolves, we really dont have clue yet why it behaves one
5.3.
87
Figure 5.5: Turbulence intensity decay downstream of a square-rod grid with M = 25.4mm and U = 20m/s The open circles represent the mean square streamwise uctuations and the plus signs the cross-stream (from Comte-Bellot and Corrsin 1971).
88
tr 17 1.2 1 u2 (m/s)2 0.8 0.6 0.4 0.2 0 0 exponential theory
2 x (m)
Figure 5.6: Variation of turbulence intensity with downstream position for one space-lling fractal grid. Solid line shows exponential t (from Hurst and Vassilicos (2007)). way or the other. Who knows, maybe some of you will be involved in resolving it, since it really is one of the most fundamental questions in turbulence.
5.4
A second look at simple shear ow turbulence
Lets consider another homogeneous ow that seems pretty simple at rst sight, homogeneous shear ow turbulence with constant mean shear. We already considered this ow when we were talking about the role of the pressure-strain rate terms. Now we will only worry, for the moment, about the kinetic energy which reduces to: [ ] dU 1 k = uv (5.7) t 2 dy Now turbulence modelers (and most experimentalists as well) would love for the left-hand side to be exactly zero so that the production and dissipation exactly balance. Unfortunately Mother Nature, to this point at least, has not allowed such a ow to be generated. In every experiment to-date after some initial adjustments, the energy increases with time (or equivalently, increases with increasing
5.4. A SECOND LOOK AT SIMPLE SHEAR FLOW TURBULENCE
89
Figure 5.7: Diagram of space-lling square fractal grid, from gure 33 of Hurst and Vassilicos (2007).
90
8 7 6 5 (mm) 4 3 2 1 0 0 0.5 1 1.5 x/xpeak 2 tr = 2.5 tr = 5.0 t = 8.5 r tr = 13.0 tr = 17.0 2.5 3
Figure 5.8: Downstream variation of Taylor microscale, , for dierent fractal grids showing approach to constant level as ow evolves downstream (from Hurst and Vassilicos (2007).)
5.4. A SECOND LOOK AT SIMPLE SHEAR FLOW TURBULENCE
91
downsteam distance as the turbulence is convected down the tunnel). Lets make a few simple assumptions and see if we can gure out what is going on. Suppose we assume that the correlation coecient uv/u2 = C is a constant. Now, we could again assume that u3 /l, at least in the very high Reynolds number limit. But for reasons that will be obvious below, lets assume something else we know for sure about the dissipation; namely that: =D u2 2 (5.8)
where is the Taylor microscale and D 15 (exact for isotropic turbulence). Finally lets assume that the mean shear is constant so dU/dy = K is constant also. Then our problem simplies to: d3 2 u2 u = KCu2 D 2 dt 2 (5.9)
Even with all these simplications and assumptions the problem still comes down to What is = (t)?. Now the one thing that all the experiments agree on is that = o is approximately constant. (I actually have a theory about this, together with M. Gibson, and it even predicts this result, George and Gibson 1992 Experiments in Fluids) Now you have all you need to nish the problem, and I will leave it for you. But when you do you will nd that the turbulence grows (or decays) exponentially. How fast it grows (or decays) depends on the ratio of the production to dissipation; i.e., P uvdU/dy (5.10)
It is clear from the experiments that P/ depends on the upstream or initial conditions, as the theory suggests it should. But it is not at all clear how. One possibiliy is that the higher Reynolds number characterising these initial conditions, the closer P/ is to unity. If so, then you will only get P/ 1 as an innite Reynolds number limit. Which in turn implies you can never really achieve the ideal ow many people would like where the production and dissipation exactly balance. But here we are again, with lots of questions and no denitive highl Reynolds number experiments to guide us, nor little hope for help from the low Reynolds number numerical simulations. So here is one more set of questions to add to the list of things about turbulence we would like to know and to which you might contribute.
92
Chapter 6 Turbulent Free Shear Flows

6.1 Introduction
Free shear ows are inhomogeneous ows with mean velocity gradients that develop in the absence of boundaries. As illustrated in Figures 6.1 and 6.2, turbulent free shear ows are commonly found in natural and in engineering environments. The jet of air issuing from ones nostrils or mouth upon exhaling, the turbulent plume from a smoldering cigarette, and the buoyant jet issuing from an erupting volcano all illustrate both the omnipresence of free turbulent shear ows and the range of scales of such ows in the natural environment. Examples of the multitude of engineering free shear ows are the wakes behind moving bodies and the exhausts from jet engines. Most combustion processes and many mixing processes involve turbulent free shear ows. Free shear ows in the real world are most often turbulent. Even if generated as laminar ows, they tend to become turbulent much more rapidly than the wallbounded ows which we will discuss later. This is because the three-dimensional vorticity necessary for the transition to turbulence can develop much more rapidly in the absence of walls that inhibit the growth velocity components normal to them. The tendency of free shear ows to become and remain turbulent can be greatly modied by the presence of density gradients in the ow, especially if gravitational eects are also important. Why this is the case can easily be seen by examining the vorticity equation for such ows in the absence of viscosity,
[
i ui p i + uj = j + ijk t xj xj xj xk
(6.1)
The cross-product of the density and pressure gradients of last term vanishes identically in constant density ows or barotropic1 ows, but can act quite dramatically to either increase or decrease vorticity production. For example, in the
A barotropic ow is one in which the gradients of density and pressure are co-linear, because the density is a function of the pressure only.
1
93
94
CHAPTER 6. TURBULENT FREE SHEAR FLOWS
Figure 6.1: Left: Exhaust from smokestack. Right: Exhaust from rocket engine
vertically-oriented buoyant plume generated by exhausting a lighter uid into a heavier one, the principal density gradient is across the ow and thus perpendicular to the gravitational force which is the principal contributor to the pressure gradient. As a consequence the turbulent buoyant plume develops much more quickly than its uniform density counterpart, the jet. On the other hand, horizontal vorticity (vertical motions) in a horizontal free shear ows in a stably stratied environment (uid density decreases with height) can be quickly suppressed since the density and pressure gradients are in opposite directions. Free turbulent shear ows are distinctly dierent from the homogeneous shear ows. In a free turbulent shear ow, the vortical uid is spatially conned and is separated from the surrounding uid by an interface, the turbulent-nonturbulent interface (also known as the Corrsin superlayer after itself discoverer). The turbulent/non-turbulent interface has a thickness which is characterized by the Kolmogorov microscale, thus its characterization as an interface is appropriate. The actual shape of the interface is random and it is severely distorted by the energetic turbulent processes which take place below it, with the result that at any given location the turbulence can be highly intermittent. This means that at a given location, it is sometimes turbulent, sometimes not. It should not be inferred from the above that the non-turbulent uid outside
6.1. INTRODUCTION
95
Figure 6.2: Left: Steam engine. Right: Red Arrow Jets.
the superlayer is quiescent. Quite the opposite is true since the motion of the uid at the interface produces motions in the surrounding stream just as would the motions of a solid wall. Alternately, the ow outside the interface can be viewed as being induced by the vortical motions beneath it. It is easy to show that these induced motions are irrotational. Thus since these random motions of the outer ow have no vorticity, they can not be considered turbulent. Figure 6.3 shows records of the velocity versus time at two locations in the mixing layer of a round jet. When turbulent uid passes the probes, the velocity signals are characterized by bursts of activity. The smooth undulations between the bursts in the lower gure are the irrotational uctuations induced by the turbulent vorticity on the other side of the interface. Note that near the center of the mixing layer where the shear is a maximum, the ow is nearly always turbulent, while it becomes increasingly intermittent as one proceeds away from the region of maximum production of turbulence energy. This increasing intermittency toward the outer edge is a characteristic of all free shear ows, and is an indication of the fact that the turbulent/non-turbulent interface is constantly changing its position. One of the most important features of free shear ows is that the amount of uid which is turbulent is continuously increased by a process known as entrainment. No matter how little uid is in the ow initially, the turbulent part of the ow will continue to capture new uid by entrainment as it evolves. The photograph of an air jet in Figure 1.2 illustrates this phenomenon dramatically. The mass ow of the jet increases at each cross-section due to entrainment. Entrainment is not unique to turbulent ows, but is also an important characteristic of laminar ow, even though the actual mechanism of entrainment is quite dierent. There are several consequences of entrainment. The rst and most obvious is that free shear ows continue to spread throughout their lifetime. (That this is the case is obvious from the pictures of Figures 6.1 and 6.2. A second consequence of entrainment is that the uid in the ow is being continuously diluted by the addition of uid from outside it. This is the basis of many mixing processes, and without such entrainment our lives would be quite dierent. A third consequence
96
12
Maximum shear Outer edge 10
U (m/s)
2 0
t (s)
Figure 6.3: Velocity versus time, axisymmetric jet. is that it will never be possible to neglect the turbulent transport terms in the dynamical equations, at least in the directions in which the ow is spreading. This is because the dilution process has ensured that the ow can never reach homogeneity since it will continue to entrain and spread through its lifetime (Recall that the transport terms were identically zero in homogeneous ows). Thus in dealing with free shear ows, all of the types of terms encountered in the turbulence kinetic energy equation of Chapter ?? must be dealt with advection, dissipation, production, and turbulent transport. Turbulent free shear ows have another distinctive feature in that they very often give rise to easily recognizable large scale structures or eddies. Figures 1.2 and 1.3 also illustrate this phenomenon, and coherent patterns of a scale equal to the lateral extent of the ow are clearly visible. The same features are are more dicult to see at higher Reynolds numbers because of the increasing energy of the smaller scales, but none-the-less they can still be discerned in the high Reynolds number jet of Figure 6.4. These large eddies appear to control the shape of the turbulent/non-turbulent interface and play an important role in the entrainment process. They stretch the small scale vorticity on the interface (the so-called Corrsin superlayer) so these are amplied, with the result that the vorticity diuses rapidly into the non-turbulent uid which has been engulfed by the large eddies. These large eddies are also important for the processes by which the turbulence gains and distributes energy from the mean ow. A feature which free shear ows have in common with the homogeneous ows
6.2. THE AVERAGED EQUATIONS
97
Figure 6.4: Smoke visualization of jet mixing layer at Reynolds number of 65,000.
discussed in Chapter 5 is that their scales continue to grow as long as the ow remains turbulent. The dynamical equations and boundary conditions for many free shear ows can be shown to admit to similarity solutions in which the number of independent variables is reduced by one. According to the equilibrium similarity principle set forth in this chapter, such ows might be expected to asymptotically achieve such a state; and this, in fact, occurs. In the limit of innite Reynolds number, some such ows can even be characterized by a single time and length scale, thus satisfying the conditions under which the simplest closure models might be expected to work. Care must be taken not to infer too much from the ability of a given closure model to predict such a ow, since any model which has the proper scale relations should work. Finally there is the important question of whether free shear ows become asymptotically independent of their initial conditions (or source conditions). The conventional wisdom until very recently (and that presented in all texts prior to this one) has been that they do. If correct, this means that there is nothing that could be done to alter the far downstream ow. As we shall see later in this chapter, there is considerable theoretical and experimental evidence, however, that this traditional view is wrong. This opens up previously unimagined possibilities for ow control at the source. In the remainder of this chapter, the averaged equations of motion will be simplied, and similarity solutions for several ideal shear ows will be derived and discussed in detail. The role of the large eddies will be discussed, and mechanisms for turbulent entrainment will be examined. The energy balance of several turbulent free shear ows will be studied in some detail. Finally, the eects of connement and non-homogeneous external boundary conditions will be considered.
6.2
6.2.1
The averaged equations

The shear layer quations
One of the most important ideas in the history of uid mechanics is that of the boundary layer approximation. These approximations to the Navier-Stokes equations were originally proposed by Prandtl in his famous theory for wall boundary
98
layers. By introducing a dierent length scale for changes perpendicular to the wall than for changes in the ow direction, he was able to explain how viscous stresses could survive near the wall at high Reynolds number. These allowed the no-slip condition at a surface to be satised, and resolved the paradox of how there could be drag in the limit of zero viscosity. It may seem strange to be talking about Prandtls boundary layer idea in a section about free shear ows, but as we shall see below, the basic approximations can be applied to all thin (or slowly growing) shear ows with or without a surface. In this section, we shall show that free shear ows, for the most part, satisfy the conditions for these boundary layer approximations. Hence they belong to the general class of ows referred as boundary layer ows. One important dierence will lie in whether momentum is being added to the ow at the source (as in jets) or taken out (by drag, as in wakes). A related inuence is the presence (or absence) of a free stream in which our free shear ow is imbedded. We shall see that stationary free shear ows fall into two general classes, those with external ow and those without. One easy way to see why this makes a dierence is to remember that these ows all spread by sharing their momentum with the surrounding ow, or by stealing momentum from it, almost always entraining mass from the surrounding uid at the same time. (You dont have to think very hard to see that as mass is entrained, it is carries its own momentum with it into the shear ow.) You should expect (and nd) that even a small free stream velocity (or pressure gradient or even free stream turbulence) can make a signicant dierence, since the momentum carried in is mixed in with that of the uid particles which are already part of the turbulence. The longer the ow develops (or the farther downstream one looks), the more these simple dierences can make a dierence in how the ow spreads. In view of this, it should be no surprise that the presence or absence of an external stream plays a major role in determining which mean convection terms which must be retained in the governing equations. And also in determining whether an experiment or simulation is a valid approximation to a ow in an innite environment, since both most be performed in a nite box or windtunnel. We will consider here only ows which are plane (or two-dimensional) in the mean (although similar considerations can be applied to ows that are axisymmetric in the mean, see the exercises). In eect, this is exactly the same as assuming the ow is homogeneous in the third direction. Also we shall restrict our attention to ows which are statistically stationary, so that time derivatives of averaged quantities can be neglected. And, of course, we have already agreed to conne our attention to Newtonian ows at constant density. It will be easier to abandon tensor notation for the moment, and use the symbols x, U, u for the streamwise direction, mean and uctuating velocities respectively, and y, V, v for the cross-stream. Given all this, the mean momentum equations reduce to:
6.2. THE AVERAGED EQUATIONS x-component: U U U 1 P u2 uv 2U 2U +V = + 2 + 2 x y x x y x y
99
(6.2)
y-component: U V 2V V 1 P uv v 2 2V +V = + 2 + 2 x y y x y x y (6.3)
In addition, we have the two-dimensional mean continuity equation which reduces to: U V + =0 (6.4) x y
6.2.2
Order of magnitude estimates
Now lets make an order of magnitude estimate for each of the terms. This procedure may seem trivial to some, or hopeless hand-waving to others. The fact is that if you fall into either of these groups, you have missed something important. Learning to make good order-of-magnitude arguments and knowing when to use them (and when not to use them) are two of the most important skills in uid mechanics, and especially turbulence. To do this right we will have to be very careful to make sure our estimates accurately characterize the terms we are making them for. Naturally we should not expect changes of anything in the x-direction to scale the same as changes in the y-direction, especially in view of the above. (This, after all, is the whole idea of thin shear ow.) So lets agree that we will pick a length scale, say L, characteristic of changes or mean quantities in the x-direction; i.e. 1 (6.5) x L where for now means of the order of magnitude of. And we can do the same thing for changes of mean quantities in the y-direction by dening a second length scale, say , to mean: 1 (6.6) y Note that both these scales will vary with the streamwise position where we evaluate them. A good choice for might be proportional to the local lateral extent of the ow (or its width), while L is related to the distance from the source. Consider the mean velocity in equation 7.10. It occurs in ve dierent terms: U alone; twice with x-derivatives, U/x and 2 U/x2 ; and twice with y-derivatives, U/y and 2 U/y 2 . Now it would be tempting to simply pick a scale velocity for U , say Us , and use it to estimate all ve terms, say as: Us , Us /L, Us /L2 , Us /, and Us / 2 . But this is much too na and fails to appreciate the true role of the ve, terms we are evaluating.
100
Look at the example shown in Figure 6.4. Our simple approach would provide an appropriate estimate for a jet if we took our velocity scale equal to the mean centerline velocity at a given streamwise location; i.e., Us (x) = Ucl (x). This is because both the mean velocity and the changes in the mean velocity across the ow are characterized by the centerline velocity. But by contrast, look at the wakes shown in Figure 1.3. If we denote the free stream velocity (i.e. far away from the wake) by Uo and the velocity at the wake centerline by Ucl , we can dene a wake decit by Ucl = Uo Ucl (6.7) Even relatively close to the wake generators, the wake decit is small compared to the free stream velocity (i.e., Ucl << Uo ). So the obvious choice to scale U (x, y) would be Uo . On the other hand, an estimate for the velocity gradient across the ow of Uo / would be much too big, again because the decit is so small. Obviously a better choice would be to use the centerline mean velocity decit, Ucl ; i.e., U Ucl (6.8) y In the order of magnitude analysis below, we shall try to keep the discussion as general as possible by using Us to characterize the mean velocity when it appears by itself, and Us to represent changes in the mean velocity. For the jet example of the preceding paragraph, both Us and Us are the same; i.e., Us = Ucl and Us = Ucl . But for the wake they are dierent because of the external stream; i.e., Us = Uo and Us = Ucl . If you can keep in mind why these dierences exist among the various ows, it will be a lot easier to both understand the results and not confuse them. Now we could distinguish changes of velocity in the x-direction from those in the y-direction. But this level of complexity is not necessary (at least for the examples considered here), especially since we have left the precise denition of L rather nebulous. What we can do is to use the same estimate for changes in the velocity as for the y-direction, and dene our length scale L to make the estimate based on both correct; i.e., U/x Us /L. To see why this makes sense physically and can be reasoned (as opposed to guessed), lets look at the wake. Pick a spot outside, near the edge of the wake fairly close the generator (say point A). Now proceed at constant y far downstream in x. Eventually the wake will have spread past you and you will be close enough to the centerline so the local mean velocity will be closer to Ucl than Uo . Obviously we have simply traveled far enough at constant y to insure that Ucl is the proper scale for the changes in velocity in the x-direction. If the distance downstream over which this change occurred is taken as L, then the proper estimate is easily seen to be: Ucl U (6.9) x L But this is exactly what we would have gotten by taking Us = Ucl as we agreed above. We simply have absorbed any dierences into our choice of L. When
101
considering specic problems and applying similarity techniques, the seemingly arbitrary choices here become quite precise constraints (as we shall see). We still havent talked about how to estimate the velocity scale for V , the cross-stream mean velocity component. From the continuity equation we know that: V U = (6.10) y x From our considerations above, we know that: U Us x L (6.11)
If there is no mean cross ow in the external stream, then the scale for V is the same as the scale for changes in V . Therefore, V Vs y (6.12)
It follows immediately that the order of magnitude of the cross-stream velocity is: Vs Us L (6.13)
We might have expected something like this if we had thought about it. If the V velocity were of the same order as the U -velocity, how could the ow in any sense be a thin shear ow. On the other hand, it also makes sense that Vs /Us /L, since both are some measure of how the ow spreads. Note that equation 6.13 would not be the correct estimate for Vs if there were an imposed cross-ow, since then we would have to consider V and changes in V separately (exactly as for U ). The mean pressure gradient term is always a problem to estimate at the outset. Therefore it is better to simply leave this term alone, and see what is left at the end. In the estimates below you will see a question mark, which simply means we are postponing judgement until we have more information. Sometimes we will have to keep the term simply because we dont know enough to throw it away. Other times it will be obvious that it must remain because there is only one term left that must be balanced by something. Now we have gured out how to estimate the order of magnitude of all the terms except the turbulence terms. For most problems this turns out to be pretty straightforward if you remember our discussion of the pressure strain-rate terms. They so eectively distribute the energy that even that the three components of velocity are usually about the same order of magnitude. So if we chose a turbulence scale as simply u, then u2 u2 , v 2 u2 , and w2 u2 . But what about the Reynolds shear stress components like uv? When acting against the mean shear to produce energy, they tend to be well-correlated and the maximum value of uv/urms vrms < 0.5. Obviously the right choice for the order of magnitude is: uv u2 . This is not always the right choice though since some cross-stress like
102
uw are identically zero in plane ows because of homogeniety. Sometimes the Reynolds shear stress will be identically zero at some points in the ow (like when it is changing sign from one side to the other of a symmetry plane). When this happens terms you neglected can sneak back into the problem. The important point to remember is that you only estimating which terms might be the most important most of the time, and must look again after you have analyzed or solved the equations to see if your estimates were correct
6.2.3
The streamwise momentum equation
Lets look now at the x-component of the mean momentum equation and write below each term its order of magnitude. U x Us Us L U + U y ( ) Us Us L V 1 P u2 uv 2U 2U + 2 + 2 x x y x y 2 2 u u Us Us ? 2 2 L L
Now the survival of at least one of the terms on the left-hand side of the equation is the essence of free shear ow, since the ow is either being speeded up or slowed down by the external ow or surroundings. Since we have chosen the primary ow direction to be x, then the largest of these acceleration (or deceleration) terms is the rst. Therefore to see the relative importance of the remaining terms, we need to re-scale the others by dividing all the estimates by Us Us /L. Doing this we have: U U x 1 = + V U y Us Us uv 2U 2U + 2 + 2 y x y ( ) ( ) 2 u L L Us Us Us L Us
1 P u2 x x 2 u ? Us Us
So what do these mean? And how do we decide whether they are of the same order as our leading term, or much less? (Note that if any are bigger than our rst term, it either means we have scaled it wrong, or that we guessed wrong about which term was the largest.) The beginning of the answer lies in remembering
103
that this is a book about turbulence. Therefore almost all the interesting ows are at high Reynolds number. But what does this mean? Now we can say. It could mean the viscous terms in the mean momentum equation are negligible, which in turn means that: Us L >> 1 and Us L >> Obviously the second criterion is much more stringent, since L/ is typically of order 10 or less for our thin shear layers. When Us / > 1000, the contributions of the viscous stresses are certainly negligible, at least as far as the x-component of the mean momentum equation is concerned. But if Us / 100 only, this is a pretty marginal assumption, and you might want to retain the last term in your analysis. Such is unfortunately the case in many experiments that are often at quite low Reynolds numbers. So, what then, you ask, do we do about the turbulence terms? To repeat: this is a book about TURBULENCE! Which means there is no way we are going to throw away all the turbulence terms on the right-hand side of the equation. There is no magic or hocus-pocus about this; you simply cant have turbulence without at least one turbulence term. And if there is no turbulence, we really arent too interested. So, what does this tell us? It tells us about ! Surprised? I bet you thought it would tell us about u, right? Not so. Look at the right-hand side of the second equation on the previous page and the orders of magnitude below it. Almost always, u < Us , sometimes much less and almost never larger. Then the biggest turbulence term is the one involving the Reynolds shear stress, uv/y, which we have estimated as [u2 /(Us Us )](L/). Hence there can be no turbulence terms at all unless: u2 L Us Us Wow! Look what we have learned about turbulence without ever solving a single equation. We know how the growth of our free shear ow relates to the turbulence intensity at least in terms of order of magnitude. The similarity theories described below will even be able to actually deduce the x-dependence of , again without actually solving the equations. Does the argument above mean the other turbulence term is negligible since it is only of order u2 /Us Us ? Well, no question we should expect it to be smaller, typically less than 5%. But unlike the viscous terms, this term often does not get smaller the farther we go in the streamwise direction. So the answer depends on the question we are asking and how accurate we want our answer to be. If
104
we are asking only rst order questions and are willing to accept errors in our solutions (or experiments) to within a 510 percent, then this normal stress term is certainly negligible. Before we go on, consider the following: since /L < 0.1 typically, to rst order free shear ows do not grow at all! So before you place your bets about whether d/dx for a particular shear ow, for example, is 0.095 instead of 0.09 (turbulence wars have been waged over less), then you better make sure you are dealing with a full deck meaning these second order (and often neglected) terms had better be retained in any calculation or experimental verication (like momentum or energy balances). Much of the confusion in the literature about free shear ows comes from the failure to use second order accuracy to make second order statements. And even more unbelievably, most attempts to do so-called code validation of second-order turbulence models are based on measurements that are only rst order accurate. Now I know you think I am kidding but try to do a simple momentum balance on the data to within order u2 /Us Us . Did I hear you say it was ridiculous to use second-order equations (like the Reynolds stress equations or even the kinetic energy equation) and determine constants to three decimal places using rst order accurate data? Hopefully you have followed all this. If not, go forward and faith will come to you. Ill summarize how we do this: You have to look very carefully at the particular ow you wish to analyze. Then, based on physical reasoning and data (if you have it), you make estimates you think are appropriate. Then you use these estimates to decide which terms you have to keep, and which you can ignore. And when you are completely nished, you carefully look back to make sure that all the terms you kept really are important and remain important as the ow develops. Finally, if you are really good at this, you look carefully to make sure that all the terms you kept do not all vanish at the same place (by going from positive to negative, for example). If this happens, then the terms you neglected may in fact be the only terms left in the equation that are not exactly zero oops! This can make for some very interesting situations. (The famous critical layer of boundary layer stability theory is an example of this.) Example: Turbulence through a contraction Consider the kinetic energy balance of a turbulence ow, homogeneous in lateral planes that is being accelerated through a contraction so that U/x > 0. First use the component Reynolds stress equations to show that the u2 is decreased by the production term (and the pressure strain-rate term as well), and that v 2 is increased. Then
105
using the continuity equation together with lHpitals rule show that to leading o order at the centerline (assume symmetry around this line) the turbulence kinetic energy equation reduces to: U k U = [u2 v 2 ] x x (6.14)
Assuming that upstream u2 > v 2 , re-examine your assessment of which terms are important near the points where k is maximal and where u2 = v 2 . (Note that these points occur nearly, but not quite, in the same place.)
6.2.4
The transverse momentum equation
Now we must also consider the transverse or y-momentum equation. The most important thing we have to remember is that the y-equation can NOT be considered independently from the x-momentum equation. We would never consider trying to simplify a vector like (a, b, c) by dividing only one component by S say to produce (Sa, b, c), since this would destroy the whole idea of a vector as having a direction. Instead you would re-scale as (Sa, Sb, Sc) to preserve the direction. We must do the same for a vector equation. We have already decided that the rst term on the left-hand side of the x-momentum equation was the term we had to keep, and we divided by its order of magnitude, Us Us /L, to make sure it was of order one. Thus we have already decided how we are going to scale all the components of the vector equation, and so we must do exactly the same thing here. But rst we must estimate the order of magnitude of each term, exactly as before. Using our previous results, heres what we get: V x Us /L Us L U V y ( ) Us /L Us L + V = 1 P uv v 2 2V 2V + 2 + 2 y x y x y 2 2 u Us /L Us /L u ? L L2 2
Now dividing each term by Us Us /L, exactly as before, we obtain: U V x L + V V y

(
Us Us
106
uv 1 P = y x 2 u ? Us Us
v 2 2V 2V + 2 + 2 y x y ( ) ( ) 2 u L Us Us Us L L Us
Unless you have seen this all before, the left-hand side is probably a surprise: none of the mean convection terms are of order one! In fact the only estimated term that is of order one in the whole equation is v 2 /y, and only because we have already agreed that (u2 /Us Us )(L/) had to be of order one to keep a turbulence term in the x-momentum equation. Of course, there cannot be an equation with only a single term equal to zero, unless we have badly over-estimated its order of magnitude. Fortunately there is the pressure term left to balance it, so to rst order in /L u2 /Us Us , the y-momentum equation reduces to simply: 0 1 P v 2 y y (6.15)
This simple equation has an equally simple interpretation. It says that the change in the mean pressure is only due to the radial gradient of the transverse component of the Reynolds normal stress, v 2 . We can integrate equation 6.15 across the shear layer from a given value of y to innity (or anywhere else for that matter) to obtain: P (x, y) = P (x, ) v 2 , (6.16)
assuming of course the free stream value of v 2 to be zero. If the free stream is at constant (or zero) mean velocity, then P (x, ) = P = constant, so we can write simply: P (x, y) = P v 2 (6.17)
Either of these can be substituted into the x-momentum equation to eliminate the pressure entirely, as we shall show below. 2 2 Now since v 2 << Us typically, P P , at least to rst order in u2 /Us . Therefore one might be tempted to conclude that the small pressure changes across the ow are not important. And they are not, of course, if only the streamwise momentum is considered. But without this small mean pressure gradient across the ow, there would be no entrainment and no growth of the shear layer. In other words, the ow would remain parallel. Clearly whether an eect is negligible or not depends on which question is being asked.
6.2.5
The free shear layer equations
If we assume the Reynolds number is always large enough that the viscous terms can be neglected, then as noted above, the pressure term in the x-momentum
6.3. TWO-DIMENSIONAL TURBULENT JETS
107
equation can be evaluated in terms of P and v 2 . Thus, to second-order in u2 /Us Us or /L, the momentum equations for a free shear ow reduce to a single equation: U U U + V x y
{ } ] dP [ 2 = uv u v 2 dx y x { }
(6.18)
As we have seen above, the second term on the left-hand side (in curly brackets) may or may not be important, depending on whether Us = Us or Us /Us << 1. Clearly this depends on the velocity decit (or excess) relative to the free stream. The second term in brackets on the right-hand side is also second-order (in u2 /Us Us /L) compared to the others, and so could rightly be neglected. It has been retained for now since in some ows it does not vanish with increasing distance from the source. Moreover, it can be quite important when considering integrals of the momentum equation, since the proles of u2 and v 2 do not vanish as rapidly with increasing y as do the other terms.
6.3
Two-dimensional Turbulent Jets
Turbulent jets are generated by a concentrated source of momentum issuing into an ambient environment. The state of the environment and the external boundary conditions are very important in determining how the turbulent ow evolves. The simplest case to consider is that in which the environment is at rest and unbounded, so all the boundary conditions are homogeneous. The jet itself can be assumed to issue from a line source or a slot as shown in Figure 6.5. The jet then evolves and spreads downstream by entraining mass from the surroundings which are at most in irrotational motion induced by the vortical uid within the jet. But no new momentum is added to the ow downstream of the source, and it is this fact that distinguishes the jet from all other ows. As we shall see below, however, that the rate at which momentum crosses any x-plane is not quite constant at the source value due to the small streamwise pressure gradient arising from the turbulence normal stresses. The averaged continuity equation can be integrated from the centerline (where it is zero by symmetry) to obtain V , i.e., V =
0 y
U d y x
(6.19)
It immediately follows that the V velocity at is given by: V = V d = U (x, y)dy dx 0 (6.20)
Twice the integral represents the total volume ow rate crossing a given x-plane. Therefore it makes sense that the rate of increase of this integral is the entrainment velocity, V . Obviously, V cannot be zero if the jet is to spread, at least in a
108
Figure 6.5: Sketch of plane jet showing jet source, coordinate system and typical mean velocity prole.
two-dimensional ow. It is easy to see that the entrainment velocity calculated from the integral is consistent with our order of magnitude estimate above (i.e., V Us /L). Note that the integral makes no sense if the mean velocity U does not go to zero with increasing |y|. If the free stream is assumed to have zero streamwise velocity, then the scales for the velocity and gradients of it are the same, so the equations reduce to simply equation 7.16 and the mean continuity equation. The latter can be multiplied by U to yield: { } U V U + =0 (6.21) x y When this is added to equation 7.16 the terms can be combined to obtain:
] 2 [ 2 U + UV + uv + u v 2 = 0 x y y x
(6.22)
This can be integrated across the ow for any given value of x to obtain: d 2 [U + (u2 v 2 )]dy = 0 dx (6.23)
where we have assumed that U , u2 , and v 2 vanish as |y| . (Remember these assumptions the next time someone tries to tell you that a small co-owing stream and modest background turbulence level in the external ow are not important.) Equation 6.23 can in turn be integrated from the source, say x = 0, to obtain:
Mo =
[U 2 + (u2 v 2 )]dy
(6.24)
where Mo is the rate at which momentum per unit mass per unit length is added at the source. The rst two terms of the integrand are the ux of streamwise momentum due to the mean ow and the turbulence. The last term is due to the streamwise pressure gradient obtained by integrating the y-momentum equation.
109
Equation 6.24 is a very powerful constraint, and together with the reduced governing equations allows determination of a similarity solution for this problem. It also is of great value to experimentalists, since it can be used to conrm whether their measurements are valid and whether the ow indeed approximates a plane jet in an innite environment at rest. Be forewarned, most do not! And the explanations as to why range from hilarious to pathetic. Consider this, if you measure the velocity moments to within say 1%, then you should be able to estimate the integral even more accurately. The reason is that the random errors are reduced by the integration in proportion to the inverse of the square root of the number of points you used in the integration if the errors are really random and statistically independent of each other. (Note that this is exactly like adding the statistically independent random numbers in Chapter 2.) So be very wary of measurements which have not been tested against the momentum integral for this and all ows. Unfortunately there is usually a reason why the authors fail to mention it: they would rather you not know.
6.3.1
Similarity analysis of the plane jet
Foreword The results of this section are VERY MUCH at odds with those in all texts, including the most recent. There is an increasing body of literature to support the position taken however, especially the multiplicity of solutions and the role of upstream or initial conditions. Since these ideas have a profound eect on how we think about turbulence, it is important that you consider both the old and the new, and be aware of why the conclusions are dierent.2 The message you should take from the previous section is that regardless of how the plane jet begins, the momentum integral should be constant at Mo . Note that Schneider (1985) has shown that the nature of the entrained ow in the neighborhood of the source can modify this value, especially for plane jets. Also the term u2 v 2 , while important for accounting for the momentum balance in experiments, can be neglected in the analysis below with no loss in generality. The search for similarity solutions to the averaged equations of motion is exactly like the approach utilized in laminar ows (e.g., Batchelor Fluid Dynamics 1967) except that here there are more unknowns than equations, i.e., the averaged equations are not closed. Thus solutions are sought which are of the form: U = Us (x)f (y, ) uv = Rs (x)g(y, )
2
(6.25) (6.26) (6.27)
This section has been taken in part from George 1995 Some new ideas for similarity of turbulent ows, Turbulence, Heat and Mass Transfer, Lisbon 1994, (Hanjalic and Pereira, eds.), 24 - 49, Elsevier, Amsterdam. Also of interest might be the paper that started this line of thinking: George 1989 The Self-Preservation of Turbulent Flows and Its Relation to Initial Conditions and Coherent Structures, in Advances in Turbulence, (George and Arndt, eds.), 39 73, Hemisphere (now Bacon and Francis), NY.
110 where
y = y/(x)
(6.28)
and where represents any dependence on upstream (or source) conditions we might not have thought of yet. The scale functions are functions of x only, i.e., Us = Us (x) Rs = Rs (x) and = (x) only. Note that the point of departure from the traditional turbulence analyses was passed when it was not arbitrarily decided 2 that Rs = Us , etc. (c.f., Tennekes and Lumley 1972, Townsend 1976, Pope 2000). Now lest you think there is something weird about the way we have written the form of the solutions we say we are looking for, consider this. Suppose you were looking for some x-dependent quantities with which you could normalize the proles to make them all collapse together like any experimentalist almost always does. Such a search is called the search for scales, meaning scales which collapse the data. Now precisely what does collapse the data really mean? Exactly this: collapse the data means that the form of the solution must be an x-independent solution to the governing equations. But this is exactly what we are looking for in equations 6.25 and 6.26. Obviously we need to plug them into the averaged equations and see if such a solution is possible. We already know which terms in these equations are important. What we want to do now is to see if these governing equations admit to solutions for which these terms remain of equal importance throughout the jet development. These are called equilibrium similarity solutions. Note that for any other form of solution, one or more terms could die o, so that others dominate. Dierentiating equations 6.25 and 6.26 using the chain-rule, substituting into the averaged momentum equation, and clearing terms yields:
[
dUs 2 dUs d y Rs f + f f (y )dy = g 2 Us dx Us dx dx Us 0
(6.29)
Now, all of the x-dependence is in the square bracketed terms, so equilibrium similarity solutions are possible only if all the bracketed terms have the same x-dependence, i.e., dUs Rs d dx Us dx Us (6.30)
We can dene the coecients of proportionality to be n and B respectively; i.e., d dUs = n dx Us dx d Rs = B dx Us (6.31) (6.32)
Note that n and B can at most depend on the details of how the ow began; i.e., the mystery argument .
6.3. TWO-DIMENSIONAL TURBULENT JETS Substitution into the momentum integral yields to rst order,
2 AUs = Mo
111
(6.33)
where A
f 2 (y, )dy
(6.34)
A also can at most depend only on how the ow began (i.e., ), since every other dependence has been eliminated. Now we learn somthing quite remarkable. Since Mo is independent of x, then the only possible value of n is n = 1/2. It follows immediately that:
1/2 Us Mo 1/2
(6.35)
It is easy to see that all of the relations involving Us are proportional to d/dx, leaving only
y 1 Rs 2 f f f (y )dy = g 2 d/dx 2 Us 0 { } [ ]
(6.36)
Thus the only remaining necessary condition for similarity is

2 Rs Us
d dx
(6.37)
Note that, unlike the similarity solutions encountered in laminar ows, it is possible to have a jet which is similar without having some form of power law behavior. In fact, the x-dependence of the ow may not be known at all because of the closure problem. Nonetheless, the proles will collapse with the local length scale. In fact, it is easy to show that the proles for all plane jets will be alike, if normalized properly, even if the growth rates are quite dierent and they began with quite dierent source conditions. To show this, dene the scale velocity Us to be the centerline velocity Ucl by absorbing an appropriate factor into the prole function f (y). (In fact, it will now have the value unity at y = 0.) Also, the entire factor in brackets on the right-hand side of equation 6.36 can be absorbed into the Reynolds stress prole,g(y); i.e., by dening g (y, ) to be: Rs g(y, ) g (y, ) = 2 Us d/dx
[ ]
(6.38)
Finally, if the length scale is chosen the same for all ows under consideration (e.g., the half-width, say 1/2 , dened as the distance between the center and where U = Ucl /2), then the similarity equation governing all jets reduces to: f2 + f
0 y
f (y )dy = 2 g
(6.39)
112
Figure 6.6: Sketch of net mass ow crossing any downstream plane for round jet as function of distance from the source.
Thus all jets, regardless of source conditions, will produce the same similarity velocity prole, when scaled with Ucl and 1/2 . This will not be true for the Reynolds stress, however, unless the coecient of proportionality between Rs and 2 Us d/dx is the same for all plane jets. In general, it is not; so here is where the dependence on source conditions shows up. It is immediately obvious from equation 6.36 that the usual arbitrary choice of 2 Rs = Us considerably restricts the possible solutions to those for which d/dx = constant, or plane jets which grow linearly. However, even if the growth were linear under some conditions, there is nothing in the theory to this point which demands that the constant be universal and independent of how the jet began. The idea that jets might all be described by the same asymptotic growth rate stems from the idea of a jet formed by a point source of momentum only, say Mo . Such a source must be innitesimal in size, since any nite size will also provide a source of mass ow. In the absence of scaling parameters other than distance from the source, x, the only possibilities are Us (Mo /x)1/2 , Rs Mo /x, and 2 x. Obviously Rs = Us , and the constants of proportionality can be assumed universal since there is only one way to create a point source jet. The problem with the point source idea is not in the idea itself, but that it has been accepted for so long by so many without justication as the asymptotic state of all nite source jets. The usual line of argument (e.g., Monin and Yaglom 1972) (if one is presented at all) is the just a plausiblity argument that goes like this. The jet entrains mass, and the mass ow of the jet increases with distance downstream like that shown in Figure 6.6. Since the mass that has been entrained ultimately overwhelms that added at the source, it is argued that the latter can be neglected. Hence the far jet is indistinguishable from a point source jet. And what is the experimental proof oered to substantiate this plausible sounding argument? Precisely that the proles of U/Ucl collapse to a single prole for all types of initial conditions. But as was demonstrated above, the mean momentum equation tells us that the mean velocity proles will collapse to a single prole no matter what! If there is a dependence on the initial conditions it
113
can only show up in the other turbulence moments, in proles of the Reynolds 2 shear stress normalized by Ucl (instead of d/dx), and, of course, in d/dx itself. And guess what quantities no two experiments ever seem to agree on? All of the above! Could all those experiments plus this nice theory with virtually no assumptions be wrong? Many still think so. It is truly amazing how much good science people will throw away in order to hang-on to a couple of ideas that were 2 never derived in the rst place but only assumed: in this case, Rs = Us and the asymptotic independence of upstream conditions. But then there are still many people who believe the world is at. George (1989) demonstrated for an axisymmetric jet that even simple dimensional analysis suggests that the classical theory is wrong, and the same arguments apply here. Suppose that in addition to momentum, mass is added at the source at a rate of mo . Now there is an additional parameter to be considered, and as a consequence, an additional length scale given by L = m2 /Mo . Thus the most that 0 1/2 can be inferred from dimensional analysis is that /x, Us x1/2 /Mo and Rs x/Mo are functions of x/L, with no insight oered into what the functions might be.
6.3.2
Implications of the Reynolds stress equations
George 1989 further argued that some insight into the growth rate of the jet can be obtained by considering the conditions for similarity solutions of the higher moment equations, in particular the kinetic energy equation. The choice of the kinetic energy equation for further analysis was unfortunate since it implicitly assumed that the three components of kinetic energy all scaled in the same manner. This is, in fact, true only if d/dx = const, which is certainly not true a priori for turbulent shear ows. Therefore, here the individual component equations of the Reynolds stress will be considered (as should have been done in George 1989). For the plane jet the equation for u2 can be written to rst order (Tennekes and Lumley 1972) as U u2 u2 +V x y } 2 u { U = p + u2 v 2uv 2u x y y (6.40)
where u is the energy dissipation rate for u2 . By considering similarity forms for the new moments like 1 2 u 2 u p x 1 2 u v 2 u = Ku (x)k() = Pu (x)pu () = Tu2 v (x)t() = Du (x)d() (6.41) (6.42) (6.43) (6.44)
114
2 and using Rs = Us d/dx, it is easy to show that similarity of the u2 -equation is possible only if 2 Ku Us 3 Us d Pu dx 3 d Tu2 v Us dx 3 Us d Du dx
(6.45) (6.46) (6.47) (6.48)
All of these are somewhat surprising: The rst (even though a second moment like the Reynolds stress) because the factor of d/dx is absent; the second, third and fourth because it is present. Similar equations can be written for the v 2 , w2 , and uv-equations; i.e. v 2 v 2 +V x y } v { = 2p + v 3 2pv 2v y y w2 w2 +V x y } w { = 2p + w2 v 2w z y
(6.49)
(6.50)
uv uv +V x y ( ) } u v { U + = p + uv 2 v 2 y x y y
(6.51)
When each of the terms in these equations is expressed in similarity variables, the resulting similarity conditions are: Us Kv d dx Us Kw d Dw Pw dx Dv P v Tv3 Tw2 v Us Kv d dx Us Kw d dx (6.52) (6.53)
(6.54) (6.55)
6.3. TWO-DIMENSIONAL TURBULENT JETS and the real surprise,

(
115
Kv
2 Us
d dx
)2
(6.56)
There is an additional equation which must be accounted for: namely that the sum of the pressure strain-rate terms in the component energy equations be zero (from continuity). Thus, p or in similarity variables, Pu (x)pu () + Pv (x)pv () + Pw (x)pw () = 0 This can be true for all only if Pu Pv Pw An immediate consequence is that Du Dv Dw (6.60) (6.59) (6.58) u v w + p + p = 0 x y z (6.57)
From equations 6.46, 6.52 and 6.53 it also follows that the constraint imposed by 6.59 can be satised only if Ku Kv Kw But from equation 6.56, this can be true only if d = constant dx (6.62) (6.61)
The relations given by equations 6.61 and 6.60 were assumed without proof in the George 1989 analysis. The additional constraint imposed by equation 6.62 was not derived, however, and arises from the additional information provided by the pressure strain-rate terms. Hence, similarity solutions of the Reynolds stress equations are possible only if Ds (x)
3 Us
(6.63)
It is an immediate consequence of the earlier discussion on the nature of the dissipation that there are only two possibilities for this to occur: i) Either the local Reynolds number of the ow is constant so that the eect of the dissipation on the energy containing eddies (and those producing the Reynolds stress as well) does not vary with downstream distance; or
116
ii) The local turbulence Reynolds number is high enough so that the relation q 3 /L is approximately valid (for a physical length L !). Unlike some ows (like the axisymmetric jet or plane wake) where the local Reynolds number is constant, for the plane jet it varies with downstream distance. Therefore the only possibility for similarity at the level of the Reynolds stresses is (ii). This can occur only when the turbulence Reynolds number is large enough, typically 104 . Since the local Reynolds number for the plane jet continues to increase with increasing downstream distance, this state will eventually be reached. The higher the source Reynolds number, the closer to the exit plane the similarity of the moments will be realized.
6.4
Other Free Shear Flows
The plane jet is but one of the ows which can be analyzed in the manner described above. A few of the possibilities which have already been analyzed are the axisymmetric jet, plane and axisymmetric wakes (George 1989, 1995 Hussein et al. 1994, Moser et al. 1997, George and Davidson 2004, Johansson et al. 2005, Johansson and George 2006, Ewing et al. 2007). Other possibilities include free shear layers, thermal plumes and the self-propelled wake to mention but a few. All of these fall into the two categories described above: Flows which evolve at constant Reynolds number, and ows which do not. The axisymmetric jet and the plane wake are of the former type, and hence when properly scaled (using the techniques described above) will yield Reynolds number and source dependent solutions. These have been already discussed in detail in the cited papers and will not be discussed further here. The second type of ows (those for which the local Reynolds number varies with streamwise distances) also fall into two types: Those for which the local Reynolds number increases downstream (like the plane jet, plume or the shear layer), and those for which it decreases (like the axisymmetric wake). When the Reynolds number is increasing with x, the ow will eventually reach the state of full similarity where all of the mean and turbulence quantities collapse. This state will be characterized by the innite Reynolds number dissipation relation q 3 / which will be manifested in the approach of d/dx to its asymptotic value. (This has been shown above to be constant for the plane jet, but will be dierent for wakes, for example.) Generally this approach will coincide with a turbulent Reynolds number of q 4 / 104 and the emergence in the spectra of the k 5/3 range. Before this, the lack of collapse will be most evident in those quantities which depend directly on d/dx, like uv, v 2 , etc. Other quantities like the mean ow will collapse much earlier, and as noted above will collapse to proles independent of source conditions. The latter will not be the case for the second moment quantities since a dependence in the asymptotic value of d/dx will result in dierences in the Reynolds stress equations themselves. Perhaps the most troubling (and for that reason the most interesting) ows
6.4. OTHER FREE SHEAR FLOWS
117
are those where the local Reynolds number is decreasing like the axisymmetric wake. In these cases, the mean velocity proles will collapse, assuming the Reynolds number to be large enough that the viscous terms in the mean momentum equation are negligible. The asymptotic growth rate (corresponding to q 3 / or in the case of the axisymmetric wake, x1/3 ) will only be achieved as long as the local Reynolds number is high enough (again q 4 / 104 ). As soon as it drops below this value, the growth rate and scale parameters will begin to deviate (perhaps substantially) from the asymptotic power law forms. The turbulence quantities will begin to reect this in the lack of collapse again rst noticeable in quantities with v 2 , etc. which have a direct dependence on d/dx. The mean velocity prole, however, will continue to collapse when scaled in local variables. The same will be true for ows in which the source Reynolds number is not high enough for the ow to ever achieve the requisite turbulence Reynolds number the mean velocity will collapse even though the x-dependence will be all wrong at least if asymptotic behavior is expected. To further confuse matters, the axisymmetric wake eventually nds a new low Reynolds number similarity solution state, in which it stays for ever.
118
Chapter 7 WALL-BOUNDED TURBULENT FLOWS

7.1 Introduction
Without the presence of walls or surfaces, turbulence in the absence of density uctuations could not exist. This is because it is only at surfaces that vorticity can actually be generated by an on-coming ow is suddenly brought to rest to satisfy the no-slip condition. The vorticity generated at the leading edge can then be diused, transported and amplied. But it can only be generated at the wall, and then only at the leading edge at that. Once the vorticity has been generated, some ows go on to develop in the absence of walls, like the free shear ows we considered earlier. Other ows remained attached to the surface and evolve entirely under the inuence of it. These are generally referred to as wall-bounded ows or boundary layer ows. The most obvious causes for the eects of the wall on the ow arise from the wall-boundary conditions. In particular, The kinematic boundary condition demands that the normal velocity of the uid on the surface be equal to the normal velocity of the surface. This means there can be no ow through the surface. Since the velocity normal to the surface cannot just suddenly vanish, the kinematic boundary condition ensures that the normal velocity components in wall-bounded ows are usually much less than in free shear ows. Thus the presence of the wall reduces the entrainment rate. Note that viscosity is not necessary in the equations to satisfy this condition, and it can be met even by solutions to to the inviscid Eulers equations. The no-slip boundary condition demands that the velocity component tangential to the wall be the same as the tangential velocity of the wall. If the wall is at rest relative, then the no-slip condition demands the tangential ow velocity be identically zero at the surface. 119
120
CHAPTER 7. WALL-BOUNDED TURBULENT FLOWS
Figure 7.1: Flow around a simple airfoil without separation.
It is the no-slip condition, of course, that led Ludwig Prandtl1 to the whole idea of a boundary layer in the rst place. Professor Prandtl literally saved uid mechanics from dAlemberts paradox: the fact that there seemed to be no drag in an inviscid uid (not counting form drag). Prior to Prandtl, everyone thought that as the Reynolds number increased, the ow should behave more and more like an inviscid uid. But when there were surfaces, it clearly didnt. Instead of behaving like those nice potential ow solutions (like around cylinders, for example), the ow not only produced drag, but often separated and produced wakes and other free shear ows. Clearly something was very wrong, and as a result uid mechanics didnt get much respect from engineers in the 19th century. And with good reason: how useful could a bunch of equations be if they couldnt nd viscous drag, much less predict how much? But Prandtls idea of the boundary layer saved everything. Prandtls great idea was the recognition that the viscous no-slip condition could not be met without somehow retaining at least one viscous stress term in the equations. As we shall see below, this implies that there must be at least two length scales in the ow, unlike the free shear ows we considered in the previous chapter for which the mean ow could be characterized by only a single length scale. The second length scale characterizes changes normal to the wall, and make it clear precisely which viscous term in the instantaneous equations is important.
Prandtl was a famous German professor at Gttingen in the early 19th century, and founder o of the famous institute there which so dominated much of 20th century uid dynamics thinking, well into the second half of the century. Among his most famous students were T. Von Karman, and H. Schlichting.
7.2. REVIEW OF LAMINAR BOUNDARY LAYERS
121
7.2
Review of laminar boundary layers
Lets work this all out for ourselves by considering what happens if we try to apply the kinematic and no-slip boundary conditions to obtain solutions of the Navier-Stokes equations in the innite Reynolds number limit. Lets restrict our attention for the moment to the laminar ow of a uniform stream of speed, Uo , around a body of characteristic dimension, D, as shown in Figure 7.1. It is easy to see the problem if we non-dimensionalize our equations using the free stream boundary condition and body dimension. The result is: p Dui 1 2 ui = + j xi Re x2 D t (7.1)
2 t where ui ui /Uo , xi xi /D, Uo t/D and p p/(Uo ). The kinematic viscosity, has disappeared entirely, and is included in the Reynolds number dened by:
Re
Uo D
(7.2)
Now consider what happens as the Reynolds number increases, due to the increase of Uo or L, or even a decrease in the viscosity. Obviously the viscous terms become relatively less important. In fact, if the Reynolds number is large enough it is hard to see at rst glance why any viscous term should be retained at all. Certainly in the limit as Re , our equations must reduce to Eulers equations which have no viscous terms at all; i.e., in dimensionless form, Di u p = xi Dt (7.3)
Now if we replace the Navier-Stokes equations by Eulers equations, this presents no problem at all in satisfying the kinematic boundary condition on the bodys surface. We simply solve for the inviscid ow by replacing the boundary of the body by a streamline. This automatically satises the kinematic boundary condition. If the ow can be assumed irrotational, then the problem reduces to a solution of Laplaces equation, and powerful potential ow methods can be used. In fact, for potential ow, it is possible to show that the ow is entirely determined by the normal velocity at the surface. And this is, of course, the source of our problem. There is no role left for the viscous no-slip boundary condition. And indeed, the potential ow has a tangential velocity along the surface streamline that is not zero. The problem, of course, is the absence of viscous terms in the Euler equations we used. Without viscous stresses acting near the wall to retard the ow, the solution cannot adjust itself to zero velocity at the wall. But how can viscosity enter the equations when our order-of-magnitude analysis says they are negligible at large Reynolds number, and exactly zero in the innite Reynolds
122
number limit? At the end of the nineteenth century, this was arguably the most serious problem confronting uid mechanics. Prandtl was the rst2 to realize that there must be at least one viscous term in the problem to satisfy the no-slip condition. Therefore he postulated that the strain rate very near the surface would become as large as necessary to compensate for the vanishing eect of viscosity, so that at least one viscous term remained. This very thin region near the wall became known as Prandtls boundary layer, and the length scale characterizing the necessary gradient in velocity became known as the boundary layer thickness. Prandtls argument for a laminar boundary layer can be quickly summarized using the same kind of order-of-magnitude analysis we used in Chapter 6. For the leading viscous term: 2u Us 2 2 (7.4) y where is the new length scale characterizing changes normal to the plate near the wall and we have assumed Us = Us . In fact, for a boundary layer next to walls driven by an external stream these can both be taken to be the free stream speed U . For the leading convection term: U U U2 s x L (7.5)
The viscous term can survive only if it is the same order of magnitude as the convection term. Hence it follows that we must have: Us U2 s 2 L (7.6)
This in turn requires that the new length scale must satisfy: L Us or
[ [ ]1/2
(7.7)
]1/2
L Us L
(7.8)
Thus, for a laminar ow, grows like L1/2 . Now if you go back to your uid mechanics texts and look at the similarity solution for a Blasius boundary layer, you will see this is exactly right if you take L x, which is what we might have guessed anyway.
Actually about the same time a Swedish meteorologist named Ekman realized the same thing must be true for rotating ows with a horizontal surface, like the earths boundary layer, for example. He invented what we now call the Ekman layer, which had the really strange characteristic that the ow direction changed with height. You can actually observe this on some days simply by looking up and noting the clouds at dierent heights are moving in dierent directions.
2
7.3. THE OUTER TURBULENT BOUNDARY LAYER
123
It is very important to remember that the momentum equation is a vector equation, and therefore we have to scale all components of this vector equation the same way to preserve its direction. Therefore we must carry out the same kind of estimates for the cross-stream momentum equations as well. For a laminar boundary layer this can be easily be shown to reduce to:
[
]{
u u v +u +v x y t
p 1 = + Re L y
]{
2v 1 + 2 Re y
2v 2 x
(7.9)
Note that a only single term survives in the limit as the Reynolds number goes to innity, the cross-stream pressure gradient. Hence, for very large Reynolds number, the pressure gradient across the boundary layer equation is very small. Thus the pressure gradient in the boundary layer is imposed on it by the ow outside the boundary layer. And this ow outside the boundary layer is governed to rst order by Eulers equation. The fact that the pressure is imposed on the boundary layer provides us an easy way to calculate such a ow. First calculate the inviscid ow along the surface using Eulers equation. Then use the pressure gradient along the surface from this inviscid solution together with the boundary layer equation to calculate the boundary layer ow. If you wish, you can even use an iterative procedure where you recalculate the outside ow over a streamline which was been displaced from the body by the boundary layer displacement thickness, and then re-calculate the boundary layer, etc. Before modern computers, this was the only way to calculate the ow around an airfoil, for example. And even now it is still used for most calculations around aerodynamic and hydrodynamic bodies.
7.3
The outer turbulent boundary layer
The understanding of turbulent boundary layers begins with exactly the same averaged equations we used for the free shear layers of Chapter 7; namely, x-component: U U U 1 P u2 uv 2U 2U +V = + 2 + 2 x y x x y x y (7.10)
y-component: V V 1 P uv u2 2V 2V U +V = + 2 + 2 x y y x y x y two-dimensional mean continuity U V + =0 x y (7.12) (7.11)
In fact, the order of magnitude analysis of the terms in this equation proceeds exactly the same as for free shear ows. If we take Us = Us = U , then the
124
ultimate problem again reduces to how to keep a turbulence term. And, as before this requires: u2 L Us Us No problem, you say, we expected this. But the problem is that by requiring this be true, we also end up concluding that even the leading viscous term is also negligible! Recall that the leading viscous term is of order: L Us (7.13)
compared to the unity. In fact to leading order, there are no viscous terms in either component of the momentum equation. In the limit as U / , they are exactly the same as for the free shear ows we considered earlier; namely, U U U + V x y and 0= 1 P 2 v y y (7.15)
{ }
1 P 2 = uv u x y x
(7.14)
And like the free shear ows these can be integrated from the free stream to a given value of y to obtain a single equation:
] [ 2 U U dP U +V = uv u v 2 x y dx y x { }
(7.16)
The last term in brackets is the gradient of the dierence in the normal Reynolds stresses, and is of order u2 /U 2 compared to the others, so is usually just ignored. The bottom line here is that even though we have attempted to carry out an order of magnitude analysis for a boundary layer, we have ended up with exactly the equations for a free shear layer. Only the boundary conditions are dierent most notably the kinematic and no-slip conditions at the wall. Obviously, even though we have equations that describe a turbulent boundary layer, we cannot satisfy the no-slip condition without a viscous term. In other words, we are right back where we were before Prandtl invented the boundary layer for laminar ow! We need a boundary layer within the boundary layer to satisfy the no-slip condition. In the next section we shall in fact show that such an inner boundary layer exists. And that everything we analyzed in this section applies only to the outer boundary layer which is NOT to be confused with the outer ow which is non-turbulent and still governed by Eulers equation. Note that the presence of P in our outer boundary equations means that (to rst order in the turbulence intensity), the pressure gradient is still imposed on the boundary layer by the ow outside it, exactly as for laminar boundary layers (and all free shear ows, for that matter).
7.4. THE INNER TURBULENT BOUNDARY LAYER
125
7.4
The inner turbulent boundary layer
We know that we cannot satisfy the no-slip condition unless we can gure out how to keep a viscous term in the governing equations. And we know there can be such a term only if the mean velocity near the wall changes rapidly enough so that it remains, no matter how small the viscosity becomes. In other words, we need a length scale for changes in the y-direction very near the wall which enables us keep a viscous term in our equations. This new length scale, lets call it . (Note that this choice of symbol is probably a bit confusing since it is very close to the symbol, K that we used for the Kolmogorov microscale. But we must make some compromises for the sake of history, since almost the whole world calls it . In fact most use exactly the same symbol for the Kolmogorov microscale.) Obviously we expect that is going to be much smaller than , the boundary layer thickness. But how much smaller? Obviously we need to go back and look at the full equations again, and rescale them for the near wall region. To do this, we need to rst decide how the mean and turbulence velocities scale near the wall scale. We are clearly so close to the wall and the velocity has dropped so much (because of the no-slip condition) that it makes no sense to characterize anything by U . But we dont have anyway of knowing yet what this scale should be, so lets just call it uw (assuming uw << U ) and dene it later. Also, we do know from experiment that the turbulence intensity near the wall is relatively high (30% or more). So there is no point in distinguishing between a turbulence scale and the mean velocity, we can just use uw for both. Finally we will still use L to characterize changes in the x-direction, since these will vary even less rapidly than in the outer boundary layer above this very near wall region we are interested in (due to the wall itself). For the complete x-momentum equation we estimate: U U x uw uw L + U y ( ) uw uw L V 1 P u2 uv 2U 2U + 2 + 2 x x y x y 2 2 uw uw uw uw ? 2 2 L L
where we have used the continuity equation to estimate V uw /L near the wall. As always, we have to decide which terms we have to keep so we know what to compare the others with. But that is easy here: we MUST insist that at least one viscous term survive. Since the largest is of order uw / 2 , we can divide by it to obtain:
126
U x ( ) uw L U
U y ( ) uw L V
2U 1 P u2 uv 2U = + 2 + 2 x x y x y ( ) 2 uw uw ? 1 L L2 Now we have an interesting problem. We have the power to decide whether the Reynolds shear stress term survives or not by our choice of ; i.e., we can pick uw / 1 or uw / 0. Note that we can not choose it so this term blows up, or else our viscous term will not be at least equal to the leading term. The most general choice is to pick /uw so the Reynolds shear stress remains too. (This is called the distinguished limit in asymptotic analysis.) By making this choice we eliminate the necessity of having to go back and nd there is another layer in which only the Reynolds stress survives as we shall see below. Obviously if we choose /uw , then all the other terms vanish, except for the viscous one. In fact, if we apply the same kind of analysis to the y-mean momentum equation, we can show that the pressure in our near wall layer is also imposed from the outside. Moreover, even the streamwise pressure gradient disappears in the limit as uw / . These are relatively easy to show and left as exercises. So to rst order in /L /(uw L), our mean momentum equation for the near wall region reduces to: U 0 uv + y y
[ ]
(7.17)
In fact, this equation is exact in the limit as uw / , but only for the very near wall region! Equation 7.17 can be integrated from the wall to location y to obtain: 0 = uv uv|y=0 + U U y y (7.18)
y=0
From the kinematic and no-slip boundary conditions at the wall we immediately know that uv|y=0 0. We also know that the wall shear stress is given by: w U y (7.19)
y=0
Substituting this we obtain our equation for the very near wall (in the limit of innite Reynolds number) as: U w = uv + y (7.20)
7.4. THE INNER TURBULENT BOUNDARY LAYER
127
We immediately recognize one of the most important ideas in the history of boundary layer theory; namely that in the limit of innite Reynolds number the total stress in the wall layer is constant. Not surprisingly, the wall layer is referred to quite properly as the Constant Stress Layer. It is important not to forget (as many who work in this eld do) that this result is valid only in the limit of innite Reynolds number. At nite Reynolds numbers the total stress is almost constant, but never quite so because of the terms we have neglected. This dierence may seem slight, but it can make all the dierence in the world if you are trying to build an asymptotically correct theory, or even just understand your experiments. Before leaving this section we need to resolve the important questions of what is uw , our inner velocity scale. It is customary to dene something called the friction velocity, usually denoted as u , by: u2 w (7.21)
Now using this, equation 7.20 can be rewritten as: u2 = uv + U y (7.22)
It should be immediately obvious that the choice is uw = u ; in other words, the friction velocity is the appropriate scale velocity for the wall region. It follows immediately from our considerations above that the inner length scale is = /u . An interesting consequence of these choices is that the inner Reynolds number is unity; i.e., u / = 1, meaning that viscous and inertial terms are about the same. But then this is precisely why we dened as we did in the rst place to ensure that viscosity was important. It is important not to read too much into the appearance of the wall shear stress in our scale velocity. In particular, it is wrong to think that it is the shear stress that determines the boundary layer. In fact, it is just the opposite: the outer boundary layer determines the shear stress. If you have trouble understanding this, just think about what happens if turn o the outer ow: the boundary layer disappears. The wall shear stress appears in the scale velocity only because of the constant stress layer in which the Reynolds stress imposed by the outer ow is transformed into the viscous stress on the wall. Finally we can use our new length scale to dene where we are in this near wall layer. In fact, we can introduce a whole new dimensionless coordinate called y + dened as: yu y (7.23) y+ = When moments of the velocity eld, for example, are normalized by u (to the appropriate power) and plotted as functions of y + , they are said to be plotted in inner variables. Similarly, we can non-dimensionalize equation 7.20 using inner
128
variables and rewrite it as simply: 1 = ri + dfi dy + (7.24)
where if we suppress for the moment a possible x-dependence we can write: fi (y + ) and ri (y + ) U (x, y) u uv u2 y (7.25)
(7.26)
By contrast an outer dimensionless coordinate, y, is dened by: y= (7.27)
In terms of these coordinates the outer equations (for the mean ow) are generally considered valid outside of y + = 30 or so. And the inner equations are needed inside of y = 0.1 if we take = 0.99 , where 0.99 is dened to be equal to the value of y at which the mean velocity in the boundary layer is 0.99% of its free stream value, U . It is easy to see that the ratio of y + to y is the local Reynolds number + where u + = (7.28) Obviously the higher the value of + , the closer the inner layer will be to the wall relative to . On the other hand, since the wall friction (and hence u ) drops with increasing distance downstream, both the outer boundary layer and inner boundary grow in physical variables (i.e., the value of y marking their outer edge increases). Sorting all these things out can be very confusing, so it is very important to keep straight whether you are talking about inner, outer or physical variables.
7.5
7.5.1
The viscous sublayer

The linear sublayer
It should be clear from the above that the viscous stress and Reynolds stress cannot both be important all the way across the constant stress layer. In fact, inside y + = 3, the Reynolds stress term is negligible (to within a few percent), so very near the wall equation 7.20 reduces to: u2 U y (7.29)
7.5. THE VISCOUS SUBLAYER or in inner variables: 1 dfi dy +
129
(7.30)
This equation is exact right at the wall, for any Reynolds number. It can immediately be integrated to obtain the leading term in an expansion of the velocity about the wall as: fi (y + ) = y + (7.31) or in physical variables U (x, y) = u2 y (7.32)
Note that some authors consider the extent of the linear sublayer to be y + = 5, but by y + = 3 the Reynolds stress has already begun to evolve, making the approximations above invalid. Or said another way, the linear approximation is only good to within about 10% at y + = 5; and the linear approximation deteriorates rapidly outside this. It should be obvious why this subregion very near the wall is usually referred to as the linear sublayer. It should not be confused with the term viscous sublayer, which extends until the mean ow is dominated by the Reynolds stress alone at about y + = 30. This linear sublayer is one of the very few EXACT solutions in all of turbulence. There are NO adjustable constants! Obviously it can be used with great advantage as a boundary condition in numerical solutions if the resolution is enough to resolve this part of the ow at all. And it has great advantages for experimentalists too. They need only resolve the velocity to y + = 3 to be able to get an excellent estimate of the wall shear stress. Unfortunately, few experiments can resolve this region. And unbelievably, even some who are able to measure all the way to the wall sometimes choose to ignore the necessity of satisfying equation 7.31 or 7.32. Look at it this way, if these equations are not satised by the data, then either the data is wrong, or the Navier-Stokes equations are wrong. If you think the latter and can prove it, you might win a Nobel prize. The important point is that the mean velocity prole is linear at the wall! It is easy to show using the continuity equation and the wall boundary conditions on the instantaneous velocity at the wall that the Reynolds stress is cubic at the 3 wall; i.e., uv/u2 = d3 y + . See if you can now use equation 7.20 to show that this implies that a fourth order expansion of the mean velocity at the wall yields: u+ = y + + c4 y +
4
(7.33)
where c4 = d3 /4. It is not clear at this point whether c4 (and d3 ) are Reynolds number dependent or not. Probably not, but they are usually taken as constant anyway by turbulence modelers seeking easy boundary conditions. Note that this fourth order expansion is a good description of the mean velocity only out to y + 7, beyond which the cubic expansion of the Reynolds stress (on which it is based) starts to be invalid.
130
Exercise: Show that the presence of a streamwise pressure gradient in the viscous sublayer equations introduces a quadratic term into the near wall mean 2 velocity prole; i.e., + y + /2 where + = (/u3 )dP/dx. See if you can use your ability to scale the y-momentum equation to show that it is imposed from the outer ow onto the near wall layer.
7.5.2
The sublayers of the constant stress region
As we move out of the linear region very close to the wall the Reynolds stress rapidly develops until it overwhelms the viscous stress. Also, as we move outward, the mean velocity gradient slowly drops until the viscous stress is negligible compared to the Reynolds shear stress. We call this region of adjustment where both the viscous and Reynolds stresses are important in the mean momentum equation, the buer layer or buer region. It roughly extends from about y + = 3 to y + = 30. On the inside viscous stresses dominate. By y + = 30, however, the viscous shear stress is less than 1 % percent or so of the Reynolds shear stress and therefore nearly negligible. Outside approximately y + = 30, we have only: u2 uv (7.34)
In other words, the Reynolds shear stress is itself nearly constant. Viscous eects on the mean ow are gone, and only the inertial terms from the uctuating motion remain. This remains true for increasing distance from the wall until the mean convection terms begin to be important, which is about y/99 0.1 for boundary layers. Obviously a necessary condition for this to be true is that + = 99 / >> 300, since otherwise there will be no region where both the viscous stress and mean convection terms are negligible. When this condition is satised, the region of approximately constant Reynolds stress is referred to as the constant Reynolds stress layer or inertial sublayer. We will have much more to say about the inertial sublayer later, after we consider boundary layers with pressure gradient and channel ows. But note for now that there are many in the turbulence community who ignore this necessary condition for its existence. This is especially true when interpreting the results of experiments and DNS, most of which are at low Reynolds number and thus never achieve a real inertial sublayer. In particular, you should be quite skeptical when you nd conceptual and computational models based on them, as many are. High Reynolds number boundary layers with a true inertial sublayers quite possibly behave very dierently; but at this writing we are only beginning to nd out. But we have learned enough to be quite suspicious. Figure 7.2 summarizes the important regions of a turbulent boundary layer. The constant stress layer has two parts: a viscous sublayer and an inertial sublayer. The viscous sublayer itself has two identiable regions: the linear sublayer where
7.5. THE VISCOUS SUBLAYER
131
Free Stream U
1111 0000
8
Outer Boundary Layer
1 0 1 0 1 0 1 0 1 0 1 0 1 0 11 00 11 00
y+
1 0 1 0 1 0 1 0 1 0 1 0 11 00 1 0
y
Outer Equations
1111111111111111111 1 0000000000000000000 0 1 0 1 0 1 0 1 0 11 11 00 00 11 11 00 00 1 0 1 0 11 00 11 1 00 0 1 0 Inertial Sublayer 1 0 1 0 1 1 0 0 Overlap Region 1 1 0 0 1 1 0 0 1 0 1 0 1 0 11111111111111111 1 00000000000000000 0 111 1 000 0 1 0 1 1 0 0 1 0 1 0 1 0 1 0 111111111111111111 1 000000000000000000 0 111 1 000 0 1 0 1 0 Viscous 11111111111111111 1 00000000000000000 0 111 1 000 0 1 0 Sublayer 1 1 0 0 1 0 1 1 0 0 1 0 1 1 0 0 1 0 11111111111111111111111111111 00000000000000000000000000000 11111111111111111111111111111 00000000000000000000000000000
0.1+ 0.1 300 Inner Equations Meso Layer 30 Buffer Layer 3 Linear Sublayer
Figure 7.2: Sketch showing the various regions of the turbulent boundary layer in inner and outer variables. Note that in + is less than approximately 3000, then the inertial layer cannot be present, and for + less than about 300, the mesolayers extend into the outer layer.
132
only viscous shear stresses are important, and a buer layer in which both viscous and Reynolds shear stresses are important. And as we shall see later, some (me among them) believe that the lower part of the inertial sublayer (30 < y + < 300 approximately) is really a mesolayer in which the Reynolds shear stress is constant at u2 , but the energetics of the turbulence (the multi-point equations) are still inuenced by viscosity. Or said another way, the mean ow equations are indeed inviscid, but the multi-point equations for the turbulence are not. This close to the wall there simply is not the scale separation between the energy and dissipative scales required for viscosity to not directly aect the Reynolds stress producing motions.
7.6
Pressure gradient boundary layers and channel ow
We will consider the near wall regions of pressure gradient boundary layers and channel ows together, in part because many have been led to believe that the near wall regions of both are identical; and even that they are identical to the near wall region of the zero pressure gradient turbulent boundary layer we considered above. From an engineering perspective, the problem would be greatly simplied if this were true. You can decide for yourself whether this is reasonable or not.3 The averaged equations for a turbulent boundary layer (including pressure gradient) look like this: U U 1 dP U U +V = + uv + (7.35) x y dx y y where the y-momentum equation for the boundary layer has been integrated to replace the local pressure by that imposed from outside the boundary layer, P , which is in turn assumed to be a function of the streamwise coordinate, x, only. By contrast fully developed turbulent channel ows are homogeneous in the streamwise direction (only the pressure varies with x), so the convection terms (left hand side) are identically zero and the mean momentum equation reduces to: 1 dP d dU 0= + uv + (7.36) dx dy dy The cross-stream momentum equation can be used to argue that P is independent of y, to at least second order in the turbulence intensity. Clearly equations 7.35 and 7.36 reduce to equation 7.17 only if the extra terms in each vanish; i.e., the mean convection terms on the left-hand side of equation 7.35 and the pressure gradient term in both. This is presumed to happen at
This chapter was taken from a more extensive discussion of the issues raised in this section and the next by George, W.K. (2007) Is there really a universal log law for turbulent wallbounded ows?, Phil. Trans. Roy. Soc. A, 365, pp. 789 - 806.
3
7.6. PRESSURE GRADIENT BOUNDARY LAYERS AND CHANNEL FLOW133

0.0 -0.1 -0.2 -0.3 -0.4
+
4
Re = 2000 Re = 950 Re = 550 Re = 180
Re = 2000 Re = 950 Re = 550 Re = 180
y dU /dy
<uv>
+
-0.5 -0.6 -0.7 -0.8 -0.9 -1.0 0 50 100
+ +
0 1.00E+00
1.00E+01
y+
150
200
250
1.00E+02 +
1.00E+03
1.00E+04
Figure 7.3: Plots of uv+ and y + /R+ 1 for DNS of channel ow, R+ = 180, 550, 950, and 2000.
Figure 7.4: Plots of y + dU + /dy + for DNS of channel ow, R+ = 180, 550, 950, and 2000.
suciently high Reynolds number. But there is a problem which is seldom (if ever) addressed. It is obvious if equation 7.36 is integrated to obtain: u2 U y dP = uv + , y dx
[ ]
(7.37)
or in so-called inner variables, y + = yu / and U + = U/u : 1 = uv+ dU + + y + dy + (7.38)
where + is a dimensionless pressure gradient dened as: + = dP u3 dx (7.39)
For a channel ow, the force exerted on the overall ow due to the streamwise pressure gradient is exactly balanced by the wall shear stress, so that: w R dP = u2 = dx (7.40)
where R is the half-height of the channel. Therefore for a channel, + = 1/R+ , where R+ = u R/. Thus equation 7.38 becomes simply: 1=
+ + dU uv dy +
y+ + + R
(7.41)
Since the viscous stress term is less than about 1% by y + = 30, this means that Reynolds shear stress drops linearly by 10% over the region of interest (y/R 0.1), independent of the Reynolds number. Therefore only in the innermost part of the constant stress layer can the pressure gradient be assumed negligible, and nowhere if the Reynolds number is not extremely high.
134
The interplay of the Reynolds shear stress and the mean pressure gradient is illustrated in gure 7.3 using the recent DNS channel ow data from Jimenez and his co-workers (e.g., del Alamo et al. 2006). The values of R+ are 180, 500, 950 and 2000. The linear drop of the total shear stress (viscous plus Reynolds) insures that there is no constant stress region, since its contribution is more than 1% of the total beyond y + = 2, 5, 10 and 20 respectively. And, contrary to popular belief, the situation will not improve with Reynolds number, since by y + = 0.1R+ 0.2R+ (the approximate outer limit of the inertial layer) the pressure gradient will always have reduced the Reynolds shear stress by 10 % - 20 %.
Many assume (mostly because they have been told) that the inertial region is described by a logarithmic prole, and that is built into many turbulence models. Figure 7.4 uses the same DNS data to illustrate that there is certainly not a logarithmic region for these data. If the velocity prole were really logarithmic, the quantity plotted, y + dU + /dy + would be constant over some region. Clearly it is not, at least for Reynolds number range of DNS data currently available.
By contrast, the boundary layer at zero pressure gradient does not have this pressure gradient problem, since the pressure gradient term is identically zero. If the Reynolds number is high enough for the convection (advection) and viscous terms to be negligible over some region (e.g., + >> 300), it truly does have a y-independent stress layer (even though it continues to vary slowly with x). Therefore it can (at least in principle) behave like the equation 7.17, even if pipe or channel ow cannot. Boundary layers with pressure gradient, however, do not behave like either a zero pressure gradient boundary layer or a channel (or pipe) ow, since over the range logarithmic behavior is expected the role of the pressure gradient depends on the value of + . Therefore the eect of the pressure gradient over the overlap region will in principle be dierent for each imposed + (with presumably dierent log parameters for each).
Thus there is no reason a priori to believe boundary layers and pipes/channels to have identical inertial layers (or even mesolayers). The most that can be expected is that they might be identical only for the part of the ow which satises y + << 0.1R+ or y + << 0.1/+ (and even then only if the residual x-dependence of the boundary layer is ignored). For most boundary layer experiments, this is a very small region indeed. Therefore, while pipe/channel experiments (or DNS) may be of considerable interest in their own right, they can not be a substitute for high Reynolds number boundary layer experiments. This is especially true if the goal is to evaluate or substantiate theories to within 10% (since their underlying equations dier over the overlap region by this amount).
7.7. THE INERTIAL SUBLAYER
135
7.7
7.7.1
The inertial sublayer

Some history
The realization that the Reynolds stress might be nearly constant over some region of the ow led von Krmn (1930) and Prandtl (1932) to postulate that the a a velocity prole might be logarithmic there. The easiest derivation begins with using an eddy viscosity argument. Simply model the Reynolds shear stress with an eddy viscosity, say uv = e U/y u2 (since the viscous term is negligible in the inertial region). Then on dimensional grounds choose e = u y, where is a constant of proportionality. Integration yields immediately: 1 u+ = ln y + + B (7.42) where and B are assumed constant This was originally believed to apply outside of y + = 30 out to about y/99 0.1 0.2 or y/R 0.2 for a channel or pipe ow. The results looked pretty good, especially given the data at the time. So not only was the idea of the log law born, it came (on the basis of quite limited evidence) to be considered universal. The subsequent renements by Millikan (1938), Isakson (1937), Clauser (1954) and Coles (1956) were so strongly embraced by the community that they not only appear in virtually all texts, it is only in the past decade or so that serious challenges to the arguments are even publishable. At this time (and over the past decade or so), the log sublayer has been (and is) one of the most discussed and debated subjects in Fluid Mechanics. There is one whole school of thought that says the mean prole is logarithmic quite independent of whether the external ow is a boundary layer, a pipe, a channel, or for that matter about anything you can think of. Moreover, it is usually argued that the three paramenters in this universal log prole are themselves universal constants. This whole idea of a universal log law has pretty much been elevated to the level of the religion and why not? Life becomes a lot simpler if some things can be taken as known to be true. And even the word, universality has a nice ring to it. For example, some go so far as to say that we can be so sure of these things that we dont even need to bother measure down below the log layer, since we already know the answer we need only nd the shear stress which ts the universal log prole onto our data, and thats the shear stress. And if the measurements down in the linear region dont collapse when plotted as U + = U/u versus y + = yu / and the inferred value of u , why then clearly the measurements of U or y must be wrong down there. The possibility that the log law might be wrong, or even that the parameters might not be constant is just too painful to contemplate. So no matter how much evidence piles up to the contrary, the universalists persist in their beliefs a lot like the people who still believe that the earth is at4 . It just makes the earth so much easier to draw.
4
There really are people who still believe the earth is at, and they even have explanations
136
Now you probably suspect from the tone of the above that I am not a member of the universalist school. And you are certainly right and I dont believe the earth is at either. But I am not one of the other group either the so-called power-law school who believe that the mean velocity prole in this region is given by a power law. The power law people are at somewhat of a disadvantage since their parameters are Reynolds number dependent. The fact that their power laws may t the data better is scoed at, of course, by the universalists who naturally believe any data for which their universal log law does not t perfectly must be wrong, or certainly in need of careful manipulation. So where to these ideas come from? And why do people hold them so zealously? Well, the rst is easy and we shall see below. The answer to the second question is probably the same as for every religious quarrel that has ever been the less indisputable evidence there is for an idea, the more people argue that there is.
7.7.2
Channel and Pipe Flows
So back to question one. Here in a nutshell is the answer: Lets consider two simple changes to the closure solution presented above; i.e., u2 = t Consider the following: We are in a ow region where the viscous stress is negligible. The only parameters we have to work with are the distance from the wall, y, and the friction velocity, u . It follows immediately on dimensional grounds alone that: t = Cu [y + a] (7.44) U y (7.43)
where the coecient C must be exactly constant in the limit of innite Reynolds number as must the additive oset a. Note that an oset is necessary since we really dont have any idea how big an eddy is relative to the wall, nor what the eect of the viscous sublayer beneath us is. For the ows homogeneous in x, we can take C(Re) and a = a(Re) to be Reynolds number dependent, but independent of y. This implies we can integrate equation 7.43 using equation 7.44 to obtain: homogeneous ows u+ = Ci (Re) ln[y + + a(RE)+ ] + Bi (Re) (7.45)
for all our observations that suggest it is not. Turbulence experts are capable of these kinds of arguments too.

25
40 35 30 25
137
20
Re_D = 7.43E4a = 1.44E5a = 4.11E5 = 1.35E6 = 4.46E6
20 15 10 5 0 1
10
Re_D = 7.43Ea = 1.44E5a = 4.11E5 = 1.35E6 = 4.46e6
(Uc -U)/u *
15
U+
10
100
(R-r)+
1000
10000
0 1.E-03
1.E-02
1.E-01
1.E+00
(R-r)/R
Figure 7.5: Five velocity proles in inner Figure 7.6: Five velocity proles in outer variables. Superpipe data of McKeon et variables. Superpipe data of McKeon et al. 2004a. al. 2004a. where B(Re) is the integration constant. Since in the limit of innite Reynolds number, our original momentum equation was independent of Reynolds number, then so must our solution be in the same limit; i.e. Bi (Re) Bi Ci (Re) Ci a(Re)+ a+ (7.46) (7.47) (7.48)
This is, of course, the usual log law but with an oset. And Ci = 1/ is the inverse of the usual von Karman constant. A much more formal derivation is included in the appendix to this chapter, which avoids both the problems raised earlier about the pressure gradient and the need to assume a particular turbulence model. There really arent many high Reynolds number channel simulations, but the same considerations can be applied to turbulent pipe ows. And there are some really impressively high Reynolds number pipe ow experiments that were conducted by Beverly McKeon, Lex Smits and their co-workers using the superpipe facility at Princeton. As can clearly be seen in Figures 7.5 and 7.6, the superpipe data really does show the expected logarithmic region, and its extent increases with increasing Reynolds number. It can also be argued theoretically for channel and pipe ows that there should also be a logarithmic friction law. Figure 7.7 shows that this seems to be true also.
7.7.3
Boundary Layers
The evidence above for the logarithmic prole and friction laws would seem to be rather convincing, at least for a pipe or channel. So whats the problem?
138
40 38 36 34 32
Uc / u*
30 28 26 24 22 20 1.E+03
Superpipe data (all) Theory (Re dep.) Theory (Re indep.)
1.E+04
R+
1.E+05
1.E+06
Figure 7.7: Plot of Uc /u versus R+ using superpipe data of McKeon et al. (2004b). The rst problem we have to confront is what do we mean by innite Reynolds number which is of course the only limit in which our inner equations are valid anyway? For a parallel ow like a pipe or channel where all the x-derivatives of mean quantities are zero (i.e., homogeneous in x), this is no problem, since at least the local Reynolds number would not change with downstream distance. But for a developing ow like a boundary layer or wall jet it is a BIG problem since the further you go, the bigger the local Reynolds number gets, so you never stop evolving. So we need to ask ourselves: how might things be dierent in a developing boundary layer. Many want to assume that nothing is dierent. But then the whole idea of a boundary layer is that it spreads, so it is the departures from a parallel ow that make it a boundary layer in the rst place. The most obvious way (maybe the only way) to account for the variation of the local Reynolds number as the ow develops is to include a factor y + in our eddy viscosity; i.e., t = C(Re)[y + ] u y (7.49)
where can depend on the local Reynolds number. Since we know the ow very near the wall is almost parallel, it is obvious that is going to be have to be quite small. But it is easy to show that no matter how small is, unless it is identically zero our equation will never integrate to a logarithm. In fact, we obtain: u+ = Ci (Re)[y + + a+ ] + Bi (Re)
(7.50)
where as before all the parameters must be asymptotically constant. As you can see this is a power law. It can become a log law only if is asymptotically zero, which of course would mean that the boundary layer has stopped growing. This seems unlikely, at least to me.

10000
1100
139
900
1000
Osterlund, Rtheta= 20,562 Stanislas et al., Rtheta=21,000 700
Y+
Y+
500
100
10 Osterlund, Rtheta= 20,562 Stanislas et al., Rtheta=21,000 1 0 5 10 15 20 25 30 35

100 0 5 10 15 U+ 20 25 30 35 300
U+
Figure 7.8: Comparison of mean velocity pro- Figure 7.9: Expanded linear-linear plot les from Osterlund (2000) at R = 20, 562 of data in gure ?? from 100 < y + < and Stanislas et al. (2006) at R = 21, 000. 1, 100. On the other hand, suppose the ow is exactly parallel, like a channel or a pipe. Then our would be exactly zero, and we get a logarithmic prole. This is to me one of the great and beautiful mysteries of calculus: how a tiny innitesimal can change a logarithm to a power law. In fact, other arguments based on functional analysis and quite independent of closure models can produce the same results, and show that the additive parameter for the power law is identically zero. Moreover, it is possible using a powerful new method called Near-asymptotics to actually deduce the Reynolds number dependence of most of the parameters in both the log and the power laws. This is an area still undergoing development at the moment. My advice is: reject any simple argument which dismisses these new developments. They may not be as simple as the old log law, but they might have the advantage of being correct. Boundary layer developments over the past decade have been discussed in detail in a recent paper (George 2006 AIAA Journal) from the perspective of the ideas presented here, so the very brief summary below will not suce for a careful reading of that paper. The primary concerns are twofold: rst the apparent lack of a consistent theory for the log law in boundary layers; and second, the validity of the experiments. Both these are discussed below. First, there is good reason to believe that the underlying log theory utilized above so convincingly for pipe/channel ow does not apply to boundary layers. The theory for pipe and channel ows depends crucially on the existence of the Reynolds number independent limits of the scaled proles. If the inner and outer proles do not both scale with u , then there is no possibility of a logarithmic prole in the overlap region. Because the boundary layer is not homogeneous in the streamwise direction, there is no theoretical argument that can be used to justify an outer decit law for boundary layers using u (as described in detail in George 2006 and George and Castillo 1997). Thus in the absence of supporting theory, any further inferences based on this decit scaling are at most built on a
140
foundation of empiricism, no matter how good the empirical collapse over some range of the data. In spite of the theoretical objections, there is still evidence that a log friction law with = 0.38 is an accurate description of at least the friction law for the boundary layer (e.g., Osterlund 2000, Nagib et al. 2004), and perhaps even the velocity proles. In an eort to resolve the apparent paradox, George (2006) suggested these might represent the leading terms in a logarithmic expansion of the power law solutions, and showed that a value near 0.38 was consistent with the power law coecients. In fact, the dependence of the skin friction (or equivalently, u /U ) on Reynolds number (or + ) that results from a power law theory (which is theoretically defensible from rst principles, George and Castillo 1997), are virtually indistinguishable from the log law ts to the same power law theoretical curves. Whatever the reason for the apparent success of the log law in zero pressure gradient boundary layers, in the absence of a consistent log theory for the boundary layer (or any developing wall-bounded ow), there is no reason to believe that logarithmic friction and velocity laws for boundary layers should be linked to those for pipes and channels, no matter how good the empirical curve ts. The consequences of this are quite important, since it means that boundary layers could be quite well described by logarithms with = 0.38, independent of all other considerations. Second, there are reasons to believe there are signicant problems with at least some of the boundary layer mean velocity measurements that have been used to argue for the log law with = 0.38. George (2006) pointed out that the recent results from Nagib et al. (2004) are not consistent with the momentum integral, diering by as much as 30-40%.5 Thus either the ow is not a two-dimensional incompressible smooth wall turbulent boundary layer, or the skin friction and/or velocity measurements are in error. George (2006) also considered in detail the other extensive and relatively re cent set of mean velocity measurements by Osterlund 2000. Contrary to the claims made by Osterlund et al. (2000), these data were shown to be equally consistent with either log or power law curve ts, and in fact the curve ts were indistinguishable. It was also pointed out that in the absence of Reynolds stress measurements, there was no way to conrm that any of the measured proles were consistent with the mean equations of motion. This was of considerable concern, since unlike earlier boundary data (e.g., Smith & Walker (1959)), the Osterlund data showed virtually no Reynolds number dependence in the overlap region, but did in the outer region of the ow (where one would least expect to nd it). The afore-mentioned concern about the Osterlund experiment was considerably heightened by recent results from on-going experiments at the Lille boundary layer facility (mentioned above) by Stanislas and co-workers (see Carlier and
This problem seems to have disappeared in more recently reported versions of this data, but with no explanation oered. The data have still not been made available to the general public as of this writing.
5

10000
1100
141
900
1000
Osterlund, Rtheta= 20,562 Stanislas et al., Rtheta=21,000 700
Y+
Y+
500
100
10 Osterlund, Rtheta= 20,562 Stanislas et al., Rtheta=21,000 1 0 5 10 15 20 25 30 35

100 0 5 10 15 U+ 20 25 30 35 300
U+
Figure 7.10: Comparison of mean velocity Figure 7.11: Expanded linear-linear plot proles from Osterlund (2000) at R = 20, 562 of data in gure ?? from 100 < y + < and Stanislas et al. (2006) at R = 21, 000. 1, 100. Stanislas et al. 2004), Stanislas et al. 2006), which became available to me in the course of preparing this paper. Their mean velocity prole obtained at R = 21, 000 is plotted in Figure 7.10 along with the corresponding prole from Osterlund (2000). The friction coecients for the two experiments were almost exactly the same (meaning the normalized shear stresses were in agreement), as is evident from the near overlay of the curves at the largest distances from the wall (eectively U /u ). This is encouraging, since the Osterlund shear stress was estimated using an oil lm method, and the Lille result obtained by micro-PIV. Incidentally, the latter value diers by about 3.5% from the shear stress estimated using the Clauser chart on the Lille data, a substantial dierence in view of the questions being asked. By contrast to the data at large distances from the wall, the mean velocity proles near the wall (inside 0.1599 or y + < 1200) dier substantially until they come together again inside y + = 10. Figure 7.11 shows a linear-linear plot of both sets of data showing only the region from 100 < y + < 1, 100. The two proles appear virtually identical over this range, but shifted in both velocity and position. This is dicult to understand. Both sets of measurements were obtained using hot-wire anemometry, so there is no obvious reason for the dierence. The Lille proles, however, were conrmed by more sparsely spaced PIV measurements. Moreover the Lille proles are consistent with the measured Reynolds stress proles (from PIV) and the dierential equations of motion. Both the log prole and power law prole, U + = Ci (y + + a+ ) , can be optimized to t the Lille velocity proles to within 0.2% for 50 < y + < 0.1 + . This was no surprise, since as pointed out by George (2006), the functional forms are indistinguishable, at least over the range of data available. The results for , Co , Ci and a+ were 0.119, 0.97, 9.87 and -8.3 respectively, dierent from the earlier estimates of George and Castillo (1997) as expected since the data were dierent than that previously considered. The values for the logarithmic t, on the other
142
hand, were quite surprising given the dierence in the measured proles, since the optimal values were 0.384, 4.86, -2.03, and 4 for , Bi , Bo and a+ respectively. By comparison, the log values for the Osterlund prole were = 0.384, Bi = 4.16, + and a = 0. In other words, the values of from the log curve ts to the Stanislas et al. and Osterlund experiments were identical, even though the proles and other constants diered substantially. Thus in spite of the dierences and shortcomings of the various experiments (which remain to be explained and reconciled), there would appear to be increasing evidence for = 0.38 (or even 0.384) for boundary layers.
7.7.4
Summary and conclusions
So in summary, there is no justication, theoretical or experimental, for a universal log law for all wall-bounded ows, no matter how aesthetically appealing or potentially useful an idea. At very least, boundary layers and pipe/channel ows are fundamentally dierent. Or viewed another way, the log law represents the inertial region of pipes, channels and boundary layers to about the same degree that their underlying equations have the same terms, which is to within about 10%. Thus the historical value of = 0.41 is probably best seen as a compromise for dierent ows, accurate to within these limits. The log theory does apply quite rigorously to pipe ows with = 0.43, and perhaps to other wall-bounded ows homogeneous in horizontal planes (e.g., channels, Couette ow, the neutral planetary boundary layer, etc.). But it is a power law theory for the boundary layer that can be derived from rst principles using equilibrium similarity analysis and near-asymptotics. This theory predicts (without additional assumptions) a number of things that have also been observed, but which require additional hypotheses with a log theory. Among them are: diering 2 outer scales for the normal and shear stress components (Uo and u2 respectively), the consequent dependence of the turbulence properties of boundary layers in the overlap region on mixed scales, and the dependence of pressure uctuations on the ratio Uo /u . Moreover, the same principles can be used to predict dierent results for dierent ows (like wall-jets and boundary layers with pressure gradient), again as observed. Nonetheless, theoretical arguments notwithstanding, the log law also appears to apply to developing boundary layers. If not the leading term in a logarithmic expansion of the power law solution, it is at least a local and empirical description And to the degree that developing boundary layers can be described this way, the value of for them appears (at least at this time) to be about 0.38.
143
Appendix: Formal derivation of log proles for pipe ow

The derivation of the log law for pipe and channel ows can be carried out much more rigorously than presented above, and quite independent of any eddy viscosity model. This derivation also avoids any problems one might infer from the arguments above (and the DNS data) about the importance of the pressure gradient. The problem is that for pipe and channel ows is that there is really no constant Reynolds stress region nor any region where the pressure gradient can be ignored. We owe to Isakson (1937) and Millikan (1938) the original arguments that the contrary is true. The basic arguments have been presented in detail in many texts (e.g., Tennekes and Lumley 1972, Panton 1996 and in slightly more general form in Wosnik et al. 2000), and will only be summarized here. We will present them for pipe ow only, since experiments in pipes are much easier to realize at high Reynolds numbers than in channels, in part due to the diculty of maintaining a two-dimensional mean ow in the latter. The superpipe data of Zagorola and Smits (1998), for example, go as high as R+ = 500, 000. The basic theoretical arguments for channels, however, are the same. The counterpart to equation 7.36 for fully-developed ow in an axisymmetric pipe with smooth walls is: 1 P 1 U 0= + r uv + x r r r
[ ]
(7.51)
where r is measured from the pipe centerline, and U, u and v are to be interpreted as corresponding to the streamwise and radial velocity components respectively. Since P is nearly independent of r, we can multiply by r and integrate from the wall, R, to the running coordinate, r, to obtain the counterpart to equation 7.37 as: u2 r = R
( )[
U 1 (R2 r2 ) 1 dP uv + r 2 R dx
(7.52)
Integration all the way to the centerline (r = 0) yields the relation between the pipe radius, the wall shear stress and the imposed pressure gradient as: u2 = R dP 2 dx (7.53)
Thus of the four parameters in the equation, R, , (1/)dP/dx and u2 , only three are independent. It follows immediately from dimensional analysis that the mean velocity for the entire ow can be written in either of two ways: U = fi (r+ , R+ ) u (7.54)
144 and
U Uc (7.55) = fo (r, R+ ) u where Uc is the mean velocity at the centerline, R+ = u R/ is the ratio of outer to inner (or viscous) length scales, r+ = ru / is an inner normalization of r, and r = r/R is an outer normalization of r. Note that the velocity dierence from the centerline, or velocity decit, is used in the last expression to avoid having to take account of viscous eects as R+ . Figures 7.5 and 7.6 show some of the recent superpipe data of McKeon et al. (2004b) plotted in inner and outer variables respectively. (Note that the conventional labels inner and outer may appear opposite to the what they should be, since the outer is really the core region for the pipe and inner is a thin region closest to the pipe walls.) Clearly the inner scaled proles appear to collapse near the wall, nearly collapse over a large intermediate range, and diverge when (R r)/R > 0.1 or so. This means that the extent of the region of nearcollapse in inner variables increases indenitely as the Reynolds number, R+ , increases. One might easily infer that in this region of near-collapse, the collapse will also improve to some asymptotic limiting prole (in inner variables). Similarly the outer scaled proles appear to nearly collapse as long as R+ r+ > 300 500 approximately, and again one could infer that the region of collapse might improve and continue all the way to the wall if the Reynolds number increased without bound. We can dene hypothetical inner and outer limiting proles as fi (r+ ) and fo (r) respectively; i.e., limR+ fi (r+ , R+ ) = fi (r+ ) limR+ fo (r, R+ ) = fo (r) (7.56) (7.57)
For nite values of R+ , both equations 7.56 and 7.57 describe functionally the entire prole, and R+ acts as a parameter to distinguish the curves when they diverge. To see how they dier, consider the limit as R+ . Clearly viscosity has disappeared entirely from fo (r), so it can at most represent the mean velocity prole away from the near wall region where viscous eects are not important. By contrast, fi (r+ ) can only describe the very near wall region, since it has retained no information about R. Now it is possible that the two limiting proles, fi and fo , dont link up; i.e., neither describes the ow far enough toward the pipe center in the rst case and toward the wall in the latter that they both describe a common region. But suppose they do (Millikans great idea), so that both the inner and outer scalings have a common (or overlap) region. Or thought of another way: can we stretch the region over which the inner region collapses the data so that it overlaps a similar stretch in the other direction of the outer scaled version? In fact there are a variety of ways to show that the answer is yes. The traditional way is to set the inner limit of fo equal to the outer limit of fi and ask
145
whether there can be a solution in the limit of innite Reynolds number. Similar results can be obtained by matching derivatives in the limit (c.f., Tennekes and Lumley 1972), or using matched asymptotic expansions (e.g., Panton 1996). Alternatively, Wosnik et al. (2000) used the methodology of near asymptotics to seek not an overlap region, but instead a common region which survives at nite Reynolds number as the limits are approached. Regardless, all of the methodologies conclude that the mean velocity prole in the common (or overlap) region should be logarithmic and given by the following equations: U Uc 1 = ln (1 r + a) + Bo u U 1 = ln (R+ r+ + a+ ) + Bi u
(7.58) (7.59)
where a = a/R, a+ = au / and a is a spatial oset which is a necessary consequence of the need for invariance (c.f., Oberlack 2001, Wosnik et al. 2000). In addition, the friction velocity and centerline velocity must be related by the following relationship (or friction law): Uc 1 = ln R+ + C u where C = Bi Bo (7.61) (7.60)
Thus it is not enough to simply draw a logarithmic curve on a friction plot or an inner velocity plot and conclude anything more than that an empirical t is possible. In fact empirical log ts always always seem to work, at least over some limited range, for just about any curve. Therefore it is only when ts to all the three plots (friction, inner and outer mean velocity) can be linked together with common parameters using equations 7.58 to 7.61 that it can truly be concluded that pipe/channel ows are logarithmic and that theory and experiment agree. The asymptotic theories conclude that , Bi , and Bo must be constant, but only because the matching is done in the limit as R+ . Near-asymptotics, by contrast tells how these limits are approached (inversely with powers of ln R+ ) and also how the dierent parameters are linked together; i.e., they must either be independent of R+ or satisfy: d d (1/) = (Bi Bo ) (7.62) + d ln R d ln R+ But regardless of whether , Bi or Bo are constants (i.e., independent of Reynolds number) or only asymptotically constant, only two of them can be chosen independently. So how well does this work? Quite well actually. Figure 7.7 shows data for Uc /u from the superpipe experiments of McKeon et al. (2004a), along with ln R+
146
several logarithmic ts to the data, both with average rms errors of about 0.2%. One curve uses constant values of and C, and another the variable Reynolds number version proposed by Wosnik et al. 2000. The values of and C = Bi Bo were 0.427 and 7.19 for the constant parameter (Reynolds number independent) analysis, while the limiting values for and C = Bi Bo for the Reynolds number dependent analysis were 0.429 and 7.96 respectively. The reason for the dierence between the two values of C can be seen by examining the Reynolds number dependence in the Wosnik et al. theory for which: (1 + )A (ln R+ ) 1 1 A = (ln R+ )1+
Ci = Ci +
(7.63) (7.64)
Substituting these into equation 7.60 yields the rened friction law as: Uc 1 A = ln R+ + C + , u (ln R+ ) (7.65)
All of the extra Reynolds number dependence is in the last term of equation 7.65, and in fact it is this term which adjusts C from its asymptotic value of 7.96 to 7.19 over the range of the experiments. For this data set the optimal values of and A were given by -0.932 and 0.145 respectively, so the variation in over the entire range of the data was only from 0.426 to 0.427. The corresponding variation of the last term in the friction law, however, was from 0.690 to 0.635, enough to account for the slight lack of collapse of the mean velocity proles in outer variables noted in gure 7.6. Note that the constants determined by the Wosnik et al. 2000 used an earlier (and uncorrected) form of the superpipe data, which showed a slightly dierent Reynolds number dependence. The dierences are due to the static hole corrections in the new data set. Unlike the conclusions from earlier versions of the superpipe data, there would appear to be little reason to consider the Reynolds number dependent version superior. Table 7.1 summarizes the parameters from individual ts to ve of the McKeon et al. (2004a) proles selected to cover the entire range of the data. The value of determined from the friction data was taken as given, and the values of Bi and a were determined from a regression t of each inner prole between 50 < (R r)+ < 0.1R+ . The average rms errors are approximately 0.2% for all inner proles. These same values, together with C = Bi Bo , were then used to determine Bo , by optimizing the t to the same prole in outer variables over the same range. The values of are remarkably constant, as are those of Bi . There might be a slight Reynolds number trend in the values of Bo . In view of the closest distance to the wall which can be measured (relative to a+ ), the variation in a+ is probably random positioning errors.
7.7. THE INERTIAL SUBLAYER ReD 105 R+ 103 Bi Bo a+ % errin % errout 0.743 1.44 4.11 1.82 3.32 8.51 0.426 0.426 0.427 5.62 5.50 5.64 -1.65 1.77 -1.65 -1.33 -1.34 -5.10 0.169 0.269 0.427 0.657 1.13 1.71 13.5 44.6 25.2 76.4 0.427 0.427 5.87 5.85 -1.43 -1.46 -11.9 -1.5 3.26e-04 0.188 1.37 1.55
147
Table 7.1: Parameters for ts of log law to inner and outer proles of McKeon et al. (2004a) using friction law values for Reynolds number dependent parameters. If the Reynolds number dependence is truly negligible, then the inner and outer mean velocity proles should collapse when dierent Reynolds numbers are plotted together as in Figures 7.5 and 7.6. The collapse of the proles in inner variables is excellent, consistent with the observations that is nearly constant, and Bi and a+ are nearly so. The outer variable plot does not collapse quite so well, especially over the range for which the proles are logarithmic. This lends support for the variable Reynolds number approach, which shows that the only signicantly Reynolds number dependent parameter is Bo . This has implications for the asymptotic friction law, however, since the asymptotic values of C = Bi Bo are dierent, 7.19 versus 7.96. So where does this leave us? These experiments (and most other pipe experiments as well) show an almost perfect agreement with the theoretical predictions, both the asymptotic and near-asymptotic versions. Not only does there appear to be a region of logarithmic behavior in the mean velocity proles where we expected to nd it (30 50 < y + < 0.1 0.2R+ ), the parameters determined from ts to these and the logarithmic friction law satisfy the constraints among them. This is about the strongest experimental conrmation for a theory that can possibly be imagined. So the analysis presented here (of part of the most recent version of the superpipe experiments) suggests strongly that the value of is about 0.43, a bit lower than the value of 0.44-0.45 suggested from the earlier uncorrected data and slightly higher than the estimate of McKeon et al. (2004a) of 0.42 using a larger set of the same data used herein. But all are higher than the earlier accepted value of 0.41, however, and most certainly not lower. The asymptotic value of the additive constant for the outer velocity prole (and friction law) can still be debated, but this debate in no way detracts from the overall conclusion. In spite of the absence of a constant total stress region (and hence the lack of validity of the early arguments for it), the logarithmic theory for a pipe ow can be taken as fact. One can infer this is probably also true for the channel, once data at suciently high Reynolds number becomes available to test it.
148
Chapter 8 Stationarity Random Processes

This chapter on stationary random processes does not really t here if we were following the logical order of presentation of the material, since it has absolutely nothing to do with the chapters which immediately follow it. And it is a actually a bit dangerous to put it here, since it leads many people to the conclusion that the Reynolds-averaged equations of the next few chapters are valid only for stationary random processes. This is absolutely false. The RANS equations are completely general, whether the under-lying statistical processes are stationary or not, since they have been averaged in the sense of the true (or ensemble) average of the preceding chapter. Nonetheless there is an advantage pedagogically to putting this chapter here since many students at this point are trying to gure out how the results the preceding chapter apply to the time-varying (and often statistically) stationary processes they are seeing in the lab, or have seen in their experience. It is not hard to think of the nite time estimator for the time average, say UT , as being related to the the nite sample estimator, XN , of the preceding chapter, especially if the process is sampled digitally. It is not at all obvious, however, What constitutes an independent realization is the process is time-varying? This chapter both introduces the idea of a stationary random process, and answers this question. Much more detail about this important class of random process is included in the appendices, including a complete discussion of spectral analysis.
8.1
Processes statistically stationary in time
Many random processes have the characteristic that their statistical properties do not appear to depend directly on time, even though the random variables themselves are time-dependent. For example, consider the signals shown in Figures 2.2 and 2.5. When the statistical properties of a random process are independent of time, the random process is said to be stationary. For such a process all the moments are time-independent, e.g., (t) = U , etc. Note that we have used a tilde u 149
150
CHAPTER 8. STATIONARITY RANDOM PROCESSES
to represent the instantaneous value (i.e., u(t)), a capital letter to represent its average (i.e., U = (t)), and we dene the uctuation by a lower case letter (i.e., u u = u(t)U ). Also since the process is assumed stationary, the mean velocity and velocity moments are time-independent. In fact, the probability density itself is time-independent, as should be obvious from the fact that the moments are time independent. An alternative way of looking at stationarity is to note that the statistics of the process are independent of the origin in time. It is obvious from the above, for example, that if the statistics of a process are time independent, then un (t) = un (t + T ), etc., where T is some arbitrary translation of the origin in time. Less obvious, but equally true, is that the product u(t)u(t ) depends only on the time dierence t t and not on t (or t ) directly. This consequence of stationarity can be extended to any product moment. For example, u(t)v(t ) can depend only on the time dierence t t. And u(t)v(t )w(t ) can depend only on the two time dierences t t and t t (or t t ) and not t, t or t directly.
8.2
The autocorrelation
One of the most useful statistical moments in the study of stationary random processes (and turbulence, in particular) is the autocorrelation dened as the average of the product of the random variable evaluated at two times, i.e. u(t)u(t ). Since the process is assumed stationary, this product can depend only on the time dierence = t t. Therefore the autocorrelation can be written as: C( ) u(t)u(t + ) (8.1)
The importance of the autocorrelation lies in the fact that it indicates the memory of the process; that is, the time over which a process is correlated with itself. Contrast the two autocorrelations shown in Figure 8.1. The autocorrelation of a deterministic sine wave is simply a cosine as can be easily proven. Note that there is no time beyond which it can be guaranteed to be arbitrarily small since it always remembers when it began, and thus always remains correlated with itself. By contrast, a stationary random process like the one illustrated in the gure will eventually lose all correlation and go to zero. In other words it has a nite memory and forgets how it was. Note that one must be careful to make sure that a correlation really both goes to zero and stays down before drawing conclusions, since even the sine wave was zero at some points. Stationary random processes always have two-time correlation functions which eventually go to zero and stay there. Example 1. Consider the motion of an automobile responding to the movement of the wheels over a rough surface. In the usual case where the road roughness is ran-
8.2. THE AUTOCORRELATION
151
Figure 8.1: Autocorrelations for two random processes and a periodic one.
domly distributed, the motion of the car will be a weighted history of the roads roughness with the most recent bumps having the most inuence and with distant bumps eventually forgotten. On the other hand if the car is traveling down a railroad track, the periodic crossing of the railroad ties represents a deterministic input and the motion will remain correlated with itself indenitely. This can be a very bad thing if the tie crossing rate corresponds to a natural resonance of the suspension system of the vehicle. Since a random process can never be more than perfectly correlated, it can never achieve a correlation greater than is value at the origin. Thus |C( )| C(0) (8.2)
An important consequence of stationarity is that the autocorrelation is symmetric in the time dierence, = t t. To see this simply shift the origin in time backwards by an amount and note that independence of origin implies: u(t)u(t + ) = u(t )u(t) = u(t)u(t ) Since the right hand side is simply C( ), it follows immediately that: C( ) = C( ) (8.4) (8.3)
152
8.3
The autocorrelation coecient

u(t)u(t + ) C( ) = C(0) u2
It is convenient to dene the autocorrelation coecient as: ( ) where u2 = u(t)u(t) = C(0) = var[u] Since the autocorrelation is symmetric, so is its coecient, i.e., ( ) = ( ) (8.7) (8.6) (8.5)
It is also obvious from the fact that the autocorrelation is maximal at the origin that the autocorrelation coecient must also be maximal there. In fact from the denition it follows that (0) = 1 (8.8) and ( ) 1 for all values of . (8.9)
8.4
The integral scale
One of the most useful measures of the length of time a process is correlated with itself is the integral scale dened by Tint
( )d
(8.10)
It is easy to see why this works by looking at Figure 8.2. In eect we have replaced the area under the correlation coecient by a rectangle of height unity and width Tint .
8.5
The temporal Taylor microscale

dC d 1 d2 C + 2 2 d 2 =0 1 3 d3 C 3! dt
The autocorrelation can be expanded about the origin in a MacClaurin series; i.e., C( ) = C(0) + +
=0
(8.11)
=0
But we know the autocorrelation is symmetric in , hence the odd terms in must be identically zero (i.e., dC/d | =0 = 0, d3 /d 3 | =0 , etc.). Therefore the expansion of the autocorrelation near the origin reduces to:
8.5. THE TEMPORAL TAYLOR MICROSCALE
153
Figure 8.2: The autocorrelation coecient showing relation of the integral scale to the area under the autocorrelation coecient curve.
1 2 d2 C C( ) = C(0) + 2 d 2
+
=0
(8.12)
Similarly, the autocorrelation coecient near the origin can be expanded as: ( ) = 1 + 1 d2 2 d 2 2 +
=0
(8.13)
where we have used the fact that (0) = 1. If we dene = d/d we can write this compactly as: 1 (8.14) ( ) = 1 + (0) 2 + 2 Since ( ) has its maximum at the origin, obviously (0) must be negative. We can use the correlation and its second derivative at the origin to dene a special time scale, (called the Taylor microscale 1 ) by: 2 2 (0) (8.15)
Using this in equation 8.14 yields the expansion for the correlation coecient near the origin as: 2 (8.16) ( ) = 1 2 +
The Taylor microscale is named after the famous English scientist G.I. Taylor who invented it in the 1930s. Among his many other accomplishments he designed the CQR anchor which is still found on many boats today.
1
154
Figure 8.3: The autocorrelation coecient for positive time lags together with its oscullating parabola showing the Taylor microscale.
Thus very near the origin the correlation coecient (and the autocorrelation as well) simply rolls o parabolically; i.e., ( ) 1 2 2 (8.17)
This parabolic curve is shown in Figure 8.3 as the osculating (or kissing) parabola which approaches zero exactly as the autocorrelation coecient does. The intercept of this osculating parabola with the -axis is the Taylor microscale, . The Taylor microscale is signicant for a number of reasons. First, for many random processes (e.g., Gaussian), the Taylor microscale can be proven to be the average distance between zero-crossing of a random variable in time. This is approximately true for turbulence as well. Thus one can quickly estimate the Taylor microscale by simply observing the zero-crossings using an oscilloscope trace. The Taylor microscale also has a special relationship to the mean square time derivative of the signal, [du/dt]2 . This is easiest to derive if we consider two stationary random signals, say u and u , we obtain by evaluating the same signal at two dierent times, say u = u(t) and u = u(t ). The rst is only a function of t, and the second is only a function of t . The derivative of the rst signal is du/dt and the second du /dt . Now lets multiply these together and rewrite them as: d2 du du = u(t)u (t ) dt dt dtdt (8.18)
where the right-hand side follows from our assumption that u is not a function of t nor u a function of t.
8.6. TIME AVERAGES OF STATIONARY PROCESSES
155
Now if we average and interchange the operations of dierentiation and averaging we obtain: du du d2 = u u (8.19) dt dt dtdt Here comes the rst trick: u u is the same as u(t)u(t ), so its average is just the autocorrelation, C( ). Thus we are left with: du du d2 = C(t t) dt dt dtdt (8.20)
Now we simply need to use the chain-rule. We have already dened = t t. Lets also dene = t + t and transform the derivatives involving t and t to derivatives involving and . The result is: d2 d2 d2 = 2 2 dtdt d d So equation 8.23 becomes: du du d2 d2 = 2 C( ) 2 C( ) dt dt d d (8.22) (8.21)
But since C is a function only of , the derivative of it with respect to is identically zero. Thus we are left with: du du d2 = 2 C( ) (8.23) dt dt d And nally we need the second trick. Lets evaluate both sides at t = t (or = 0 to obtain the mean square derivative as: du dt
( )2
d2 C( ) d 2 =0
(8.24)
But from our denition of the Taylor microscale and the facts that C(0) = u2 and C( ) = u2 ( ), this is exactly the same as: du dt
( )2
=2
u2 2
(8.25)
This amazingly simple result is very important in the study of turbulence, especially after we extend it to spatial derivatives.
8.6
Time averages of stationary processes
It is common practice in many scientic disciplines to dene a time average by integrating the random variable over a xed time interval, i.e.,
156
Figure 8.4: Integration of random time signal from 0 to T is area under the signal
1 T2 UT u(t)dt T T1
(8.26)
For the stationary random processes we are considering here, we can dene T1 to be the origin in time and simply write: UT 1 T
0 T
u(t)dt
(8.27)
where T = T2 T1 is the integration time. Figure 8.4 shows a portion of a stationary random signal over which such an integration might be performed. The time integral of u(t) over the interval (O, T ) corresponds to the shaded area under the curve. Now since u(t) is random and since it forms the upper boundary of the shaded area, it is clear that our estimator for the time average over the interval (0, T ), UT , is itself random and will depend on which particular section of the signal is being integrated. Thus, UT is a lot like the estimator for the mean based on a nite number of independent realizations, XN we encountered earlier in Section 2.5. It will be shown in the analysis presented below that if the signal is stationary, the time average dened by equation 8.27 is an unbiased estimator of the true average U . Moreover, the estimator converges to U as the time becomes innite; i.e., for stationary random processes U = lim 1T u(t)dt T T 0 (8.28)
8.7. BIAS AND VARIABILITY OF TIME ESTIMATORS
157
Thus the time and ensemble averages are equivalent in the limit as T , but only for a stationary random process.
8.7
Bias and variability of time estimators
It is easy to show that the estimator, UT , is unbiased by taking its ensemble average; i.e., 1T 1T UT = u(t)dt = (t)dt u (8.29) T 0 T 0 Since the process has been assumed stationary, u(t) is independent of time. It follows that: 1 UT = (t)T = U u (8.30) T To see whether the estimate improves as T increases, the variability of UT must be examined, exactly as we did for XN earlier in Section 2.5.2. To do this we need the variance of UT given by: var[UT ] = [UT UT ]2 = [UT U ]2
{ }2
1 T = [(t) U ]dt u T2 0 1 TT = [(t) U ][(t ) U ]dtdt u u T2 0 0 1 TT = u(t)u(t )dtdt T2 0 0
(8.31) (8.32)
But since the process is assumed stationary u(t)u(t ) = C(t t) where C(t t) = u2 (t t) is the correlation function dened earlier and (t t) is the correlation coecient. Therefore the integral can be rewritten as: var[UT ] = 1 TT C(t t)dtdt 2 0 T 0 u2 T T = (t t)dtdt T2 0 0 (8.33) (8.34)
Now we need to apply some fancy calculus. If new variables = t t and = t + t are dened, the double integral can be transformed to (see Figure 8.5):
[ ] 2T 0 2T + var[u] T d d( ) + d d( ) var[UT ] = 2T 2 0 T
(8.35)
where the factor of 1/2 arises from the Jacobian of the transformation. The integrals over d can be evaluated directly to yield:
{ } 0 var[u] T [( )[T ]d + ( )[T + ]d var[UT ] = 2T 2 0 T
(8.36)
158
Figure 8.5: Integration domain for transformed variables.
By noting that the autocorrelation is symmetric, the second integral can be transformed and added to the rst to yield at last the result we seek as: var[u] T | | d var[UT ] = ( ) 1 T T T
[ ]
(8.37)
Now if our averaging time, T , is chosen so large that | |/T < 1 over the range for which ( ) is non-zero, the integral reduces: var[UT ] 2var[u] T ( )d T 0 2Tint = var[u] T
(8.38)
where Tint is the integral scale dened by equation 8.10. Thus the variability of our estimator is given by: 2Tint var[u] 2 T = (8.39) U T U2 Therefore the estimator does, in fact, converge (in mean square) to the correct result as the averaging time, T increases relative to the integral scale, Tint . There is a direct relationship between equation 8.39 and equation 2.54 which gave the mean square variability for the ensemble estimate from a nite number of statistically independent realizations, XN . Obviously the eective number of independent realizations for the nite time estimator is: Nef f = T 2Tint (8.40)
8.7. BIAS AND VARIABILITY OF TIME ESTIMATORS
159
Figure 8.6: The integral for two integral scales in time eectively acts as an independent realization, at least for the time integral of a random signal.
so that the two expressions are equivalent. Thus, in eect, portions of the record separated by two integral scales behave as though they were statistically independent, at least as far as convergence of nite time estimators is concerned. Thus what is required for convergence is again, many independent pieces of information. This is illustrated in Figure 8.6. That the length of the record should be measured in terms of the integral scale should really be no surprise since it is a measure of the rate at which a process forgets its past. Example It is desired to measure the mean velocity in a turbulent ow to within an rms error of 1% (i.e., = 0.01). The expected uctuation level of the signal is 25% and the integral scale is estimated as 100 ms. What is the required averaging time? From equation 8.39 T = 2Tint var[u] 2 U2 = 2 0.1 (0.25)2 /(0.01)2 = 125sec
(8.41)
Similar considerations apply to any other nite time estimator and equation 2.57 can be applied directly as long as equation 8.40 is used for the number of independent samples. It is common experimental practice to not actually carry out an analog integration. Rather the signal is sampled at xed intervals in time by digital means and the averages are computed as for an ensemble with a nite number of realizations. Regardless of the manner in which the signal is processed, only a nite portion of
160
a stationary time series can be analyzed and the preceding considerations always apply. It is important to note that data sampled more rapidly than once every two integral scales do not contribute to the convergence of the estimator since they can not be considered independent. If N is the actual number of samples acquired and t is the time between samples, then the eective number of independent realizations is { N t/Tint if t < 2Tint Nef f = (8.42) N if t 2Tint It should be clear that if you sample faster than t = 2Tint you are processing unnecessary data which does not help your statistics converge. You may wonder why one would ever take data faster than absolutely necessary, since it simply lls up your computer memory with lots of statistically redundant data. When we talk about measuring spectra you will learn that for spectral measurements it is necessary to sample much faster to avoid spectral aliasing. Many wrongly infer that they must sample at these higher rates even when measuring just moments. Obviously this is not the case if you are not measuring spectra.
Chapter 9 Homogeneous Random Processes

Acknowledgement: The gures in this chapter were prepared by Abolfazl Shiri.
9.1
Random elds of space and time
To this point only temporally varying random elds have been discussed. For turbulence however, random elds can be functions of both space and time. For example, the temperature could be a random scalar function of time t and position , i.e., x = ( , t) x (9.1) The velocity is another example of a random vector function of position and time, i.e., = ( , t) u ux (9.2) or in tensor notation, ui = ui ( , t) x (9.3) In the general case, the ensemble averages of these quantities are functions of both position and time; i.e., ( , t) ( , t) x x ui ( , t) Ui ( , t) x x (9.4) (9.5)
If only stationary random processes are considered, then the averages do not depend on time and are functions of only; i.e., x ( , t) ( ) x x ui ( , t) Ui ( ) x x (9.6) (9.7)
Now the averages may not be position dependent either. For example, if the averages are independent of the origin in position, then the eld is said to be homogeneous. Homogeneity (the noun corresponding to the adjective homogeneous) is exactly analogous to stationarity except that position is now the variable, and not time. 161
162
CHAPTER 9. HOMOGENEOUS RANDOM PROCESSES
It is, of course, possible (at least in concept) to have homogeneous elds which are either stationary or non-stationary. Since position, unlike time, is a vector quantity it is also possible to have only partial homogeneity. For example, a eld might be homogeneous in the x1 - and x3 -directions, but not in the x2 -direction. In fact, it appears to be dynamically impossible to have ows in nature which are homogeneous in all variables and stationary as well, but the concept is useful, nonetheless. Such so-called forced turbulence, however, can be generated in a computer, and is both stationary and approximately homogeneous. Homogeneity will be seen to have powerful consequences for the equations governing the averaged motion, since the spatial derivative of any averaged quantity must be identically zero. Thus even homogeneity in only one direction can considerably simplify the problem. For example, in the Reynolds stress transport equation, the entire turbulence transport is exactly zero if the eld is homogeneous.
9.2
Multi-point correlations
The concept of homogeneity can also be extended to multi-point statistics. Consider for example, the correlation between the velocity at one point and that at another as illustrated in Figure 9.1. If the time dependence is suppressed and the eld is assumed statistically homogeneous, this correlation is a function only of the separation of the two points, i.e., ui ( , t)uj (x , t) Bi,j () x r where is the separation vector dened by r = x r x or ri = xi xi (9.9) (9.10) (9.8)
Note that the convention we shall follow for vector quantities is that the rst subscript on Bi,j is the component of velocity at the rst position, , and the second subscript is the component of velocity at the x . For scalar quantities we shall simply put a symbol for the quantity to second, x hold the place. For example, we would write the two-point temperature correlation in a homogeneous eld as: ( , t)(x , t) B, () x r (9.11)
Figure 9.2 shows a typical example. Note how the correlation peaks at zero separation and dies o as the separation vector increases. A mixed vector/scalar correlation like the two-point temperature velocity correlation would be written as: ui ( , t)(x , t) Bi, () x r (9.12)
9.3. SPATIAL INTEGRAL AND TAYLOR MICROSCALES
163
Figure 9.1: Velocity at two points separated by = x . r x
On the other hand, if we meant for the temperature to be evaluated at and the x velocity at x we would have to write: ( , t)ui (x , t) B,i () x r (9.13)
Most books dont bother with this subscript notation, and simply give each new correlation a new symbol. At rst this seems much simpler, and as long as you are only dealing with one or two dierent correlations, it is. But after we introduce a few more correlations and you read about a half-dozen pages, you nd that you have completely forgotten which symbol stands for which correlation. Then because it is usually very important to know exactly what the forgotten symbol means, you thumb madly through the book trying to nd where they were dened in the rst place. Since we will use many dierent correlations and would like to avoid this useless waste of time thumbing through the book, we will use this comma system to help us remember1 .
9.3
Spatial integral and Taylor microscales
Just as for a stationary random process, correlations between spatially varying, but statistically homogeneous, random quantities ultimately go to zero; i.e., they
For this system we can thank Professors Monin and Yaglom who wrote the famous two volume compendium translated into English by Professor John Lumley [?]
1
164
Figure 9.2: Two-point correlation of a scalar eld with vector separation, = r (r1 , r2 , 0); i.e., in the x1 -x2 -plane.
165
become uncorrelated as their locations become widely separated. Statistical phenomena that behave this way are sometimes referred as ergodic processes. Because position (or relative position) is a vector quantity, however, the correlation may die o at dierent rates in dierent directions. Thus direction must be an important part of any denition of scales from correlation functions, both the directions associated with the quantities and the direction of the separation between them. Consider for example the one-dimensional spatial correlation which is obtained by measuring the correlation between the temperature at two points along a line in the x-direction, say, B, (r) (x1 , x2 , x3 , t)(x1 + r, x2 , x3 , t) = B, (r, 0, 0)
(1)
(9.14)
where B, is dened by equation 9.11. The superscript (1) denotes the coordinate direction in which the scalar separation, r, has been chosen. This distinguishes it from the vector separation, of B, (). Also, note that the correlation r r at zero separation is just the variance; i.e., B, (0) = 2
(1)
(9.15)
Figure 9.3 shows a typical one-dimensional correlation. For separations along the 2- and 3-axes we could dene other correlations given by: B, (r) (x1 , x2 , x3 , t)(x1 , x2 + r, x3 , t) = B, (0, r, 0) B, (r) (x1 , x2 , x3 , t)(x1 , x2 , x3 + r, t) = B, (0, 0, r)
(3) (2)
(9.16) (9.17) (9.18)
In fact, we could dene a one-dimensional correlation in an arbitrary direction, say given by the unit vector /r, but this would just be the original correlation r (/r) r function; i.e., B, (r) = B, (). r
9.3.1
Integral scales
1 (1) B, (r)dr 2 0
The integral scale of ( , t) in the x1 -direction can be dened as: x

(1) L
(9.19)
Since the integral dened to be the area under positive values of the two-point correlation normalized by its value at zero separation, it has the simple geometrical interpretation shown in Figure 9.3. In particular, it is just equal to the intercept of the ordinate (separation axis) of a rectangle with the same area and abscissa intercept as the correlation function. It is easy to see from the gure that the integral scale can be a useful measure of how large a separation is required in
166
Figure 9.3: One-dimensional correlation of a scalar eld, B, (r) = B, (r, 0, 0). (1) The intercept of the rectangle with the r-axis illustrates the integral scale, L .
(1)
order for the most of the process to become uncorrelated with itself, typically a few integral scales. It is clear that there are at least two more integral scales which could be dened by considering separations in the 1- and 2 directions; i.e., 1 (2) B, (r)dr 2 0
L and
(3) L
(2)
(9.20)
1 (3) 2 B, (r)dr 0
(9.21)
In fact, an integral scale could be dened for any direction, say /r; i.e., r
(/r) r L,
1 B, ()dr r = 2 0
(9.22)
where the integration is in the direction of the separation vector . Note that just r because we can dene an integral scale does not necessarily mean it exists, since the integral of the correlation function can be zero (e.g., if the correlation function changes sign). Well see examples of this later. The situation is even more complicated when correlations of vector quantities are considered. For example, consider the correlation of the velocity vectors at two points, Bi,j (). Clearly Bi,j () is not a single correlation, but rather nine r r separate correlations: B1,1 (), B1,2 (), B1,3 (), B2,1 (), B2,2 (), etc. For each of r r r r r these an integral scale can be dened once a direction for the separation vector
167
is chosen. For example, the integral scales associated with B1,1 for the principal directions are 1 (1) L1,1 B1,1 (r, 0, 0)dr (9.23) u2 0 1 1 (2) L1,1 B1,1 (0, r, 0)dr (9.24) u2 0 1 1 (3) B1,1 (0, 0, r)dr (9.25) L1,1 u2 0 1 Integral scales can similarly be dened for the other components of the correlation tensor. Two of particular importance in the historical development of turbulence theory are 1 (1) L1,1 B1,1 (r, 0, 0)dr (9.26) u2 0 1 1 (1) L2,2 B2,2 (r, 0, 0)dr (9.27) u2 0 2 In general, each integral scale will be dierent, unless restrictions beyond simple homogeneity are placed on the process (e.g., like isotropy discussed below). Thus, it is important to specify precisely which integral scale is being referred to, which components of the vector quantities are being used, and in which direction the integration is being performed. Finally note that the integral scales dened in this section are all physical integral scales, meaning that they are directly obtainable from the two-point statistics. They should not confused with the pseudo-integral scale, u3 /, dened by equation 4.18 in Chapter 4. As noted there, the pseudo-integral scale is very much associated with the idea of a large gap in size between the energy-containing scales and those dissipating most of the energy. In the limit of innite Reynolds numbers, it can be argued that the physical and pseudo-integral scales should be proportional, but the constants of proportionality can vary greatly from ow to ow. So MIND THE GAP when using or making arguments from either type. The biggest problem in determining integral scales in the laboratory or from experimental data is most often with the experiment or simulation itself. In particular, the correlation at separations signicantly greater than integral scale can aect the integration. This also means that unless the purpose is to examine the eect of the boundaries on the ow, the windtunnel or computational box must also much greater in size than the integral scale for the experiment or DNS to be valid at all. This is a serious problem with many experiments and especially DNS which purport to be simulations of homogeneous turbulence (and free shear ows like jets or wakes), but in fact are not.
9.3.2
Taylor microscales
Similar considerations apply to the Taylor microscales, regardless of whether they are being determined from the correlations at small separations, or from the mean
168
Figure 9.4: Sketch showing relation of Taylor microscale, f to oscullating (1) parabola of correlation, f (r) = B1,1 (r).
square uctuating gradients. The two most commonly used Taylor microscales are often referred to as f and g and are dened by: 2 2 f and 2 2 g u2 2 [u2 /x1 ]2 (9.29) u2 1 [u1 /x1 ]2 (9.28)
The subscripts f and g are historical, and refer to the autocorrelation coecients dened by: f (r) and g(r) u1 (x1 , x2 , x3 )u1 (x1 + r, x2 , x3 ) B1,1 (r, 0, 0) = 2 u1 B1,1 (0, 0, 0) u2 (x1 , x2 , x3 )u1 (x2 + r, x2 , x3 ) B2,2 (r, 0, 0) = 2 u1 B2,2 (0, 0, 0) (9.30)
(9.31)
It is straightforward to show from the denitions that f and g are related to the curvature of the f and g correlation functions at r = 0. Specically, 2 = f and 2 = g 2 d2 f /dr2 |r=0 2 d2 g/dr2 |r=0 (9.32)
(9.33)
9.4. SYMMETRIES
169
Since both f and g are symmetrical functions of r, df /dr and dg/dr must be zero at r = 0. It follows immediately that the leading r-dependent term in the McClaurin series expansion about the origin of each autocorrelations is of parabolic form; i.e., r2 f (r) = 1 2 + (9.34) f and g(r) = 1 r2 + 2 g (9.35)
This is illustrated in Figure 9.4 which shows that the Taylor microscales are the intersection with the r-axis of a parabola tted to the appropriate correlation function at the origin. Fitting a parabola is a common way to determine the Taylor microscale, but to do so you must make sure you resolve accurately to scales much smaller than it (typically an order of magnitude smaller is required). Otherwise the Taylor microscale you determine is more related to the spatial ltering of the probe or numerical algorithm than to the actual ow.
9.4
Symmetries
It is easy to see that the consideration of vector quantities raises special considerations with regard to symmetries of two-point statistical quantities. These can be very dierent, depending on whether the quantities involved are scalars or vectors, or some combination of each. These symmetry conditions can be used to great advantage to reduce the quantity of measurements, and as well to conrm that the elds are truly homogenous and the the measured quantities correct. For example, as illustrated in Figures 9.2 and 9.3, the correlation between a scalar function of position at two points is symmetrical in , i.e., r B, () = B, () r r (9.36)
This is easy to show from the denition of B, and the fact that the eld is homogeneous. Simply shift each of the position vectors by the same amount r to obtain: B, () ( , t)(x , t) r x x r r = ( , t)(x , t) x = ( , t)( , t) x r = B, () r (9.37)
r x since x = . Clearly the positions are reversed and the separation vector is pointing the opposite way.
170
Such is not the case, in general, for vector functions of position. For example, see if you can prove to yourself the following for a scalar velocity correlation: B,i () = Bi, () r r Similarly, the velocity-velocity correlation must satisfy: Bi,j () = Bj,i () r r Clearly the latter is symmetrical in the variable only when i = j. r Exercise: Prove equations 9.38 and 9.39. (Hint: Note that ( )uj ( , t) = x r x uj ( , t)( ) = Bj, ().) x x r r These properties of the two-point correlation function will be seen to play an important role in determining the interrelations among the dierent two-point statistical quantities. They will be especially important when we talk about spectral quantities. (9.39) (9.38)
9.5
Implications of Continuity
The equations of mass conservation can be used to relate the derivatives of the correlations. These can be used to great advantage in reducing the number of correlations which must be measured to completely describe a turbulent eld, and even to verify whether the measured correlations are internally consistent with each other. Also, continuity together with homogeniety has important implications for the dissipation, which we have already encountered in Section 4.1. We shall see another important application in the section below when we consider isotropy, where it reduces the number of correlations (or spectra) which must be measured to a single one.
9.5.1
Reduction of Bi,j () to three independent compor nents
To see how this works, rst write the incompressible continuity equation at point (i.e., ui ( , t)/xi = 0), then multiply it by the velocity at second point, say x x , t), and average to obtain: uj (x uj (x , t) ui ( , t) x =0 xi (9.40)
But since uj (x , t) does not depend on , it can be pulled inside the derivative, as x can the average, yielding: ui ( , t)uj (x , t) x xi = Bi,j (, t) = 0, r (9.41) xi
9.6. RELATIONS AMONG THE DERIVATIVE MOMENTS
171
the last line following from the fact that homogeneity implies that ui ( , t)uj (x , t) = x Bi,j ) only. r For convenience lets dene i = xi + xi together with ri = xi xi , which in turn implies that xi = (i + ri )/2 and xi = (i ri )/2. Application of the chain-rule implies immediately that: m rm = + xi xi m xi rm = mi mi m rm = i ri Using the chain-rule result in equation 9.43 yields immediately: Bi,j () = 0 r ri (9.43)
(9.42)
since Bi,j is a function of only. r By writing the continuity equation at the point, x , and multiplying by the velocity at the point, , another three equations can be similarly derived. The x result is: Bi,j () = 0 r (9.44) rj It is clear that with both equations 9.43 and 9.44 together, we have six equations in nine unknowns. Since our equations are dierential equations, this means that all correlations can be derived to within a constant if any of three are known. But since all the correlations go to zero with increasing separation, this means that only three of the components of the two-point Reynolds stress tensor are independent. Said another way, we only need to measure three of them to nd all the others but only if the ow is incompressible. These relations among the various components of the two-point Reynolds stress equations will especially useful when we consider spectra in the next section, since the dierential equations will be replaced by algebraic ones.
9.6
Relations among the derivative moments
Homogeniety also considerably reduces the number of independent derivative moments. To see how, consider the average of the product of the velocity derivatives at two dierent points. By using the fact that ui ( , t) is not a function of x nor x , t) a function of , it follows that: is uj (x x ui ( , t) uj (x , t) x 2 x = Bi,j (x , t) xm xn xm xn 2 Bi,j (, t) r = rm rn (9.45) (9.46)
172
where we have used equation 9.41 twice in succession. But since the order of dierentiation is irrelevant in the last expression on the right-hand side, the same must be also true on the left-hand side. In particular it must be true for = 0 (i.e., x = ). Thus the derivative moments must be r x related by: ui uj ui uj = (9.47) xm xn xn xm where we have reversed the indices with respect to which we are dierentiating.
9.7
Elimination of the cross-derivative moments
The incompressible continuity equation can be used together with homogeneity to deduce another important result for homogeneous ows: namely that sij sij = i i . And immediate consequence is that for homogeneous turbulence: = 2sij sij = ui ui xj xj (9.48)
To see this, multiply the continuity equation by u1 /x1 , u2 /x2 , and u3 /x3 respectively, then average to obtain: u1 x1
[ ]2
u1 u2 u1 u3 + = 0 x1 x2 x1 x3
[ ]2
(9.49) (9.50) (9.51)
u2 u2 u1 + x2 x1 x2
u2 u3 = 0 x2 x3
[ ]2
u3 u1 u3 u3 u2 + + x3 x1 x3 x2 x3
= 0
Now by using equation 9.47 for the derivative moments, we can reverse the indices in the crossed moments and rewrite this as: u1 x1
[ ]2
u1 u2 u1 u3 + = 0 x2 x1 x3 x1
[ ]2
(9.52) (9.53) (9.54)
u1 u2 u2 + x2 x1 x2
u2 u3 = 0 x3 x2
[ ]2
u1 u3 u2 u3 u3 + + x3 x1 x3 x2 x3
= 0
Since there are three independent equations, they can be solved to obtain the crossed moments in terms of the squared derivatives as:
9.7. ELIMINATION OF THE CROSS-DERIVATIVE MOMENTS

]2
173
u1 u2 1 u1 = + x2 x1 2 x1 u1 u3 1 u1 = + x3 x1 2 x1 1 u2 u3 u1 = x3 x2 2 x1
[ [
]2
u2 + x2 u2 x2 u2 + x2
[ [
]2
u3 x3 u3 + x3 u3 + x3
[ [
(9.55) (9.56) (9.57)
]2
]2
]2 ]2
]2
]2
We will use them immediately in the section immediately following to simplify the dissipation in homogeneous ows. These relations will also be found to be particularly useful when the additional constraints of axisymmetry or isotropy are imposed toward the end of this chapter.
9.7.1
The homogeneous dissipation

= 2sij sij { } ui ui ui uj = + xj xj xj xi =
The dissipation is given by = 2sij sij , which can be expanded to obtain:
(9.58)
[ ] [ ]2 [ ]2 u 2 u1 u1 1 + + x1 x2 x3
u2 + x1 u3 + x1
[
]2 ]2
u2 + x2 u3 + x2
[
]2 ]2
u2 + x3 u3 + x3
[
]2
]2
[ ] u 2 u1 u2 u1 u3 1 + + x1 x2 x1 x3 x1
u2 u1 u2 + + x1 x2 x2
]2
u2 u3 x3 x2
[
u3 u1 u3 u2 u3 + + + x1 x3 x2 x3 x3
]2
(9.59)
From equations 9.52 to 9.54 it is immediately obvious the second bracketed term is identically zero for homogenous incompressible ow (since each line adds to zero). But the rst bracketed term is just [ui /xj ]2 . Thus for homogeneous incompressible ow the dissipation is given by: = ui ui , xj xj (9.60)
174
exactly as we surmised must be true from the various forms of the turbulence kinetic energy equation of section 4.1. There are several other similar relations that appear in turbulent ow analyses and experiments. For example, application of the derivative symmetry conditions lead immediately to the conclusion that in homogeneous turbulence the mean square strain rate is equal the mean square rotation rate which is in turn equal to half the mean square vorticity (or enstrophy); i.e., 1 sij sij = ij ij = i i 2 (9.61)
Execise: Prove that for homogeneous incompressible ow the mean square strain rate equals the mean square rotation rate and half the mean square vorticity for homogeneous incompressible ow; i.e., equations 9.61 above.
9.7.2
Pressure uctuations in homogeneous turbulence
Equations 9.61 have an interesting application to the pressure uctuatuations in homogeneous turbulence. By taking the divergence of the constant desnsity Navier-Stokes equations, it follows immediately that the pressure uctuations are described by the solution to the following Poissons equation: 1 ui uj 2 p = xj xi (9.62)
Decomposing the velocity and pressure into mean and uctuating parts (e.g., p = P + p, etc), and subtracting the mean equation leaves an equation for the pressure uctuations, p( , t), as: x 1 Ui uj Ui uj ui uj ui uj 2 p = + + xj xi xj xi xj xi xj xi
[ ] [ ]
(9.63)
The rst two terms are usually referred to as turbulence-mean shear interaction, and the latter as the turbulence-turbulence interaction terms. But we can further decompose the last uctuating term into the dierence between the strain-rate tensor squared and the rotation-rate tensor squared; i.e., ui uj = sij sij ij ij xj xi Thus equation 9.63 can be rewritten as: 1 Ui uj Ui uj 2 p = + + [(sij sij (sij sij ) (ij ij ij ij )] xj xi xj xi (9.65)
[ ]
(9.64)
9.8. ISOTROPY
175
Thus the turbulence turbulence interaction terms consist of the dierence of two quantities (the square of the uctuating strain and rotation rates) which individually can be quite large, but their dierence quite small. And in fact, as we have seen above, in homogenous turbulence the dierence of the averages of the squared strain-rate and rotation-rate tensors is exactly zero! Exercise: Show that the square of the rotation rate is equal to half the square of the vorticity (i.e., ij ij = i i /2); and rewrite equation 9.65 using it.
9.8
Isotropy
It should be clear even at this point that two point statistical quantities can be quite complex, even to simply dene which ones we are talking about. Fortunately there are symmetries which can both help us reduce the number, and also help us decide the degree to which real ows or numerically generated ones approximate homogenous ones. In addition to these basic symmetries imposed by homogeneity, it is sometimes convenient to make even more stringent assumptions about the statistical character of the turbulence; e.g., axisymmetry or isotropy. Axisymmetry in this context does not mean the same thing as the axisymmetric shear ows discussed earlier in which the statistical properties were constant in circular contours about a centerline. What axisymmetry means here is that the statistics at every point have a preferred direction about which the properties are symmetrical. And similarly isotropic means is that there is no preferred direction at all in fact, the turbulence statistical properties at a point even have reectional symmetry. In this section we shall consider only implications of isotropy.
9.8.1
An example from uid mechanics: viscous stress
Before looking at turbulence, lets make a point of contact with what you should have already from your study of Fluid Mechanics. (If you havent seen this before, pick up any good uid mechanics book.) In order to relate the viscous (or (v) deviatoric) stress tensor, say ij ( , t), to the velocity eld one begins by noting x that the stress tensor must be a functional of the velocity at all points in the ow and at all times up to the present. Then by a serious of arguments that relate to how it must appear in dierent coordinate systems one decides that it can really only be a function of the strain-rate. Then by another series of assumptions we argue that there might be some ows in which history played no role, and in fact further, the stress might depend on only the rst power of the strain-rate locally. (v) With all of these assumptions we can write: ij = Cijkl skl , meaning that we still have 81 undetermined material coecients. But we know immediately that some of these must be equal since both the stress tensor and the strain-rate
176

(v)
tensor are symmetrical; i.e., ij = ji and sij = sji . This reduces the number of unknowns from 81 to 36. Now we begin to make assumptions about the material, in particular that it is an isotropic material. This means that it should be be independent of how we view it, or even independent of whether we view it in a mirror. So step by step we consider what happens if we rotate about the 1-axis, then the 2-axis, etc., and insist these remain invariant during the transformation. And then we examine what happens if we reect about the axes; i.e,, the 1-axis becomes the negative 1-axis, etc. When have exhausted all the possibilities we end up with only two undetermined constant, our familiar Newtonian uid model; i.e., (v) ij = 2[sij skk ij /3] + 2 skk . (9.66) Amazingly, by simply exploring the implications of symmetries and isotropy, we have reduced the number of unknowns to two simple material properties. And the second of these, the second viscosity, 2 , can often be assumed to be zero. The implications of isotropy for turbulence are at least as profound. Even though ows are never really quite isotropic, we can learn a great deal by exploring what they would be like if they were. And of course we will not be able to reduce our result to simple material constants, since it is the statistics of the ow that will be assumed to be isotropic, not the underlying material.
9.8.2
Isotropic single-point correlations
There are a number of statistical quantities we have seen already for which we stated the eects of isotropy. For example, consider the single point Reynolds stress tensor, ui uj . First rotate the coordinate system 90 around the x1 -axis so the old x3 -axis becomes the new x2 axis and the old negative x2 -axis becomes the new x3 -axis. It is easy to see u2 u3 in the new coordinate system must be equal to u2 u3 in the old. But isotropy requires that the form of ui uj be independent of coordinate system. This clearly is possible only if u2 u3 = 0. Rotations and reections about the other axis lead to similar conclusions for all the o-diagonal terms. Moreover they also imply that all the diagonal terms must be equal. Therefore isotropy implies that u2 = u2 = u2 and ui uj = 0 3 2 1 for i = j. Similar considerations apply to the other second order tensors we encountered as well. We also encountered the scalar-vector correlations, like the pressure-velocity correlation, puj . Consider a mirror image transformation in which the old x1 axis becomes the negative of the new x1 axis. It follows immediately that p u1 = pu1 since the direction of the velocity is reversed. But isotropy requires that the correlation be independent of rotations and reections of the coordinate system. Hence the only possible value of the single point pressure-velocity correlation is zero. Similar considerations apply to all vector statistical properties, including the
9.8. ISOTROPY
177
mean ow. So, for example, the only possible value of the mean ow in isotropic turbulence is zero; i.e., no mean ow at all.
9.8.3
Isotropic two-point statistics
If a turbulent eld is assumed to be isotropic, this very much restricts the form of the two-point correlations (and their Fourier transforms as well). And as we shall see below, the consequences of isotropy for two-point statistical properties are considerably more interesting than for the single-point quantities. The presence of the separation vector = x makes the problem a bit more complicated than r x for the single point quantities considered above. First it introduces a whole new scalar invariant, r = || = (ri ri )1/2 . Second, if some of the statistical properties r under consideration are vectors, it introduces new angles into consideration from the inner products that can be formed with it; e.g., = ui ri = | ||| cos , where u r u r is the angle between them. It is clear that we should expect very dierent results depending on whether the correlations under consideration are scalars, vectors or tensors, and indeed this is the case. Consider rst the two-point correlation of an isotropic scalar eld like B, () r above. We already know that B, () = B, () from homogeneity alone. But if r r the eld is isotropic this correlation must be invariant no matter how the coordinate system is rotated or reected about itself. Since the components of change r with coordinate system and only its length, r = || is invariant, B, () can be r r invariant only if depends only on r; i.e., B, () = B (r) r (9.67)
Now consider a two-point correlation comprised of a vector and a scalar quantity, like Bp,i () = p( ui ( +). Again homogeneity alone implies that Bp,i () = r x x r r Bi,p (), and vice versa, so this must also be true for an isotropic eld as well. r But as we rotate and reect both the magnitude, r, and the angle between and r must be invariant. It is not immediately obvious, but straightforward to show, u that the most general form of a two-point scalar-vector correlation in an isotopic eld is given by: ri Bp,i () = BpL (r) r (9.68) r rj rj (9.69) Bj,p () = BLp (r) = BpL (r) r r r x r Two-point correlations involving two vectors, say ui = ui ( ) and uj = uj ( +), x in addition to the separation vector, are even more interesting. The reasons r can be seen in Figure 9.5. In addition to the symmetry requirements imposed by homogeneity, Bi,j () = Bj,i (), all of the angles between the vectors must r r be maintained as the coordinate system is rotated and reected. Again it is complicated but straightforward to show that the most general form of a twopoint vector-vector correlation in an isotropic eld is given by: Bi,j () = A(r)ri rj + B(r)ij r (9.70)
178
Figure 9.5: Sketches showing which velocity components are used to make up BLL , BLL , BN N , BLM ,BLN and BN M respectively, and their relation to the separation vector.
9.8. ISOTROPY
179
where A(r) and B(r) are two functions which depend on r only, and ij is the isotropic tensor (or Kronecker delta tensor). We will nd considerable use for the spectral equivalent of this result in the next chapter. It is customary (and convenient) to express the functions A and B in terms of the two-point velocity correlations we commonly measure (see Figure 9.5). The rst of these is denoted as BLL (r) and called the longitudinal correlation because the two-components of velocity are aligned with the separation vector. And the second is denoted as BN N (r) and called the normal or transverse correlation, since the two components of velocity are chosen to be in the same plane but perpendicular to the separation variable. To see how these determine A and B simply follow the directions above. For example BLL (r) would have to be the same as B1,1 (r, 0, 0), but B2,2 (0, r, 0) would also suce (as would any choice for which the vectors are collinear). Substituting into equation 9.70 yields BLL (r) as: BLL (r) = B1,1 (r, 0, 0) = A(r)r2 + B(r) Similarly, BN N (r) can be determined from: BN N (r) = B2,2 (r, 0, 0) = B(r) (9.72) (9.71)
since r2 r2 = 0 if = (r, 0, 0). We now have two equations in two unknowns, so we r can solve for A(r) and B(r) to obtain the general form: Bi,j () = [BLL (r) BN N (r)] r ri rj + BN N (r)ij r2 (9.73)
The number of unknown functions can be reduced to one (either BLL or BN N ) by employing the continuity equations derived in above in Section 9.5 Exercise: Find BLL in terms of BN N Use the isotropic relation of equation 9.73 together with the incompressible continuity equations 9.43 and 9.44 to obtain: BLL = BN N + r dBLL dr (9.74)
(Hint: Note that BLL /ri = dBLL /dr (r/ri ) and that r/rj = rj /r. Show 2 2 2 this by dierentiating r2 = r1 + r2 + r3 ) term by term.)
9.8.4
Derivative moments in isotropic turbulence
Isotropy has a powerful inuence on the velocity derivative moments. In fact, only one of them is independent, so all of the others can be expressed in terms of it. For example, if we choose to express the others in term of [u1 /x1 ]2 , the results are: u1 x1
[ ]2
u2 = x2
]2
u3 = x3
]2
(9.75)
180
u1 x1
]2
1 u1 = 2 x2 1 u1 = 2 x3 1 u2 = 2 x3
[ [
]2 ]2 ]2
1 u2 = 2 x1 1 u3 = 2 x1 1 u3 = 2 x2
[ [
]2
]2
]2
(9.76)
and u1 x1
[ ]2
u1 u2 x2 x1 u2 u3 = x3 x1 u2 u3 = x3 x2
(9.77)
It follows immediately the dissipation for isotropic turbulence can be quite simply expressed as: u1 = 15 x1
[ [ ]2
]2
(9.78)
But any of the other derivatives could have been used as well; e.g., 15 u2 = 2 x1 . (9.79)
9.8.5
Isotropic integral scale and Taylor microscales
It is straightforward to show from the form of the isotropic correlation function that the integral scales based on the longitudinal correlation, BLL (r), is double that of the normal correlation, BN N (r). Exercise: Prove that L2,2 and L1,1 are twice L1.1 . (Hint: Use equation 9.74 and integrate by parts.) It is similarly quite easy to show from the derivative relations and the de nitions of the Taylor microscales that f = 2g . It follows immediately that the dissipation can be expressed in terms of the Taylor microscales for isotropic turbulence as: = 30 u2 u2 = 15 2 2 g f (9.80)
(1) (2) (1)
9.9. AXISYMMETRIC TURBULENCE since u2 = u2 = u2 = u2 . 1 2 3 Exercise: Prove that 2 = 22 . f g
181
9.9
Axisymmetric Turbulence
A less restrictive hypothesis about turbulence than isotropy is that all statistical measures have rotational symmetry about a given axis. This is called axisymmetric turbulence.2 In fact very few ows satisfy the conditions for isotropic turbulence, but many seem to satisfy the requirements and constraints for axisymmetric turbulence. For example, in turbulent shear ows, the produced component of the turbulence energy is larger than the other two components which are approximately equal. This is especially true for the derivative moments, where many ows appear to satisfy the axisymmetric relations but not the isotropic ones.3 In spite of this axisymmetric turbulence has not been much exploited in modeling, perhaps because there is virtually nothing about it in texts. We shall to remedy that in this section. We focus here only on the velocity derivatives. If we imagine the ow to have a preferred direction, say prescribed by the unit vector, , then a Taylor expansion of the most general form of the two-point axisymmetric tensor around the point = 0 is given by: r Bi,j () = 2(02 22 + 02 )ri rj (200 + 00 ) + r2 {[402 + 222 302 ) r +2 (202 22 822 )
}]
ij + 00 + r2 (302 222 + 22 2 ) i j (9.81)
+2r(i rj + ri j )(200 02 )
where is a unit vector in the preferred direction and = /r is the inner r product of with a unit vector in the direction of the separation vector, . The r two invariants, 00 and 00 can be related to the components of the turbulence energy by: 1 00 = u2 2 1 00 = u2 u2 2 1 = u2 u2 3 1
(9.82) (9.83)
where it is presumed that u2 = u2 . These will not enter the axisymmetric 3 2 derivative relations.
The theory of axisymmetric turbulence was rst developed by Chrandresekhar in 1950. The idea of locally axisymmetric turbulence appears to have originated with George and Hussein (1991), J. Fluid Mechanics, vol. 233, pp 1 -23.
3 2
182
CHAPTER 9. HOMOGENEOUS RANDOM PROCESSES It is relatively easy to show that:

2 2 2 B1,1 () = 200 [202 + 2022 ]r1 402 (r2 + r3 ) r 2 B2,2 () = (200 + 00 ) [402 + 622 + 02 + 22 ]r1 r 2 2 [202 + 02 ]r2 [402 222 + 302 ]r3 2 B2,2 () = (200 + 00 ) [402 + 622 + 02 + 22 ]r1 r 2 2 [402 222 + 302 ]r2 [202 + 02 ]r3
(9.84) (9.85) (9.86)
Note that the last two just have r2 and r3 reversed, consistent with the symmetry about the 1-axis. The o-diagonal terms of Bi,j reduce to: B1,2 () = 2(02 + 22 )r1 r2 r B1,3 () = 2(02 + 22 )r1 r3 r B2,3 () = 2(02 00 + 00 )r2 r3 r (9.87) (9.88) (9.89)
After some work the general derivative moments can be shown to be given by: ui uj = (202 + 222 202 )(in jm + im jn ) (9.90) xm xn + {(802 422 + 602 )mn (402 222 1622 )1m 1n } ij (402 202 )[i (1m jn + jm 1n ) + j (1m in + im 1n )]
It follows that of the 45 possible derivative moment combinations, all but 15 are zero in axisymmetric turbulence. The non-zero moments depend only on 02 , 22 , 02 , and 22 . If the 1-axis is chosen as the axis of symmetry,then the non-zero moments are: u1 x1 u3 = x3 u1 = x3 u3 = x1 u3 = x2
[ [ [ [ [ ]2
= 402 + 422
]2
(9.91) (9.92) (9.93) (9.94) (9.95)
u2 x2 u1 x2 u2 x1 u2 x3
[ [ [
]2 ]2 ]2 ]2
= 402 + 402
]2
= 802
]2
= 802 + 1222 + 202 + 222

]2
= 802 + 422 + 602
The cross-derivative moments are:
9.9. AXISYMMETRIC TURBULENCE
183
u1 u2 u1 u3 = = 2(02 + 22 ) x2 x1 x3 x1 u2 u3 = 2(02 22 + 02 ) x3 x2
(9.96) (9.97)
It follows immediately by substitution into their denitions that the dissipation, , and enstrophy (mean square vorticity), i i , are given by: / = i i = [6002 + 2022 + 2002 + 422 ] (9.98)
There are a variety of ways to solve for the four invariants in terms of four independent derivative moments. These choice is generally made from what is easiest to obtain experimentally. For example, with hot-wire anemometers a particularly convenient choice of derivatives to measure are u1 /x1 , u1 /x3 , u2 /x1 and u3 /x1 . It is easy to show the the relations above yield: 1 u1 = 3 x3
[ [ [ [ ]2
02 22 02 22
]2
(9.99) 1 u1 2 x3 u1 + x1 u1 + x3
[ [ [ ]2
1 u1 = 4 x3 1 u2 = 6 x3 1 u2 = 2 x1
(9.100)
]2
]2
]2
3 u1 2 x3
[
(9.101)
[ ]2
]2
]2
10 u1 3 x1
]2
1 u2 3 x3
(9.102)
It follows that the dissipation and enstropy are given for this choice of independent derivatives as:
[ [ [ ] ]2 ]2 [ ]2 5 u 2 u1 u2 8 u2 1 = + 2 + 2 + 3 x1 x3 x1 3 x3
(9.103)
Exercise: Derive equations 9.102 and 9.103. Alternatively one might choose derivatives depending only on measurements in the 1- and 2-directions; e.g., when using planar particle image velocimetry (PIV). The dissipation can be shown in this case to be given by: u1 = x1
[ ]2
u1 + 2 x2
]2
u2 + 2 x1
]2
u2 + 8 x2
]2
(9.104)
184
Exercise: Derive equation 9.104. Hint: rst solve for the invariants in terms of these derivative moments.
Chapter 10 How do we decompose turbulence into scales of motion ?

10.1 Scales of turbulence
In the preceding chapter we have come to understand which terms do what in the incompressible Reynolds-averaged Navier-Stokes equations. We know now that it is the Reynolds stress working against the mean velocity gradient that accounts for taking energy from the mean ow and putting it into the turbulence (at least in most situations). And we know it is the dissipation (of turbulence energy) which takes this kinetic energy out from the turbulence and sends it along on an irreversible path to internal energy. So it seems there are no mysteries left, right? Actually there is quite a bit left to think about, especially if we want to know how these processes actually happen. And given all of the moments about which we need to make closure hypothesis to use the RANS e equations at all, we really all the how we possibly can. Let me point something out again, something you probably just let slide by the rst time. The Reynolds stresses are mostly associated with the large and energetic scales of motion, the so-called energy containing scales. But the dissipation takes place at the very smallest scales of motion, those characterized in high Reynolds number turbulence by the Kolmogorov microscale. If this picture is really true, how does the energy get from the large scales to the small scales? Or putting it another way: where in the equations is there a term (or terms) that accounts for this apparently necessary transfer of energy from large to small scales? Obviously this whole breakdown of the energetic scales to smaller scales is the key to the entire role of turbulence as the primary mechanism for energy dissipation. If this didnt happen, then there would be no reason to expect < sij sij > >> Sij Sij . And thus no way the turbulence dissipation could dominate direct viscous dissipation by the mean ow, which it most certainly does in most instances where turbulence occurs. But why does this really matter? Consider the following examples. If you 185
186
CHAPTER 10. DECOMPOSING TURBULENCE
wanted to increase dissipation (e.g., to slow down an airplane) then you would want to increase the rate at which energy is transferred to the smaller scales. If you wanted to decrease dissipation (e.g., to save fuel in ight), then obviously you would like to interfere with and reduce this downward cascade. Regardless of its practical interest, the manner in which dierent scales of motion exchange energy is clearly VERY important. But suppose you nd you are not able to signicantly alter the rate at which energy is sent from large to small scales? What can you do now? The obvious alternative is to prevent the energy from getting into the turbulence in the rst place. But how do you do this? You know it is the Reynolds stress working against the mean velocity gradient that produces turbulence energy from the mean ow. But this is like saying that money comes from a bank. Where does the Reynolds stress come from? How does it accomplish this energy transfer from mean to uctuation? You can guess that it must have something to do with the large scales of motion, since thats where the most energy is. But how and why did it get there? Clearly some kind of ow instability must be at work here. If so, is it the same kind of instabilities that transfer energy the from the large scales at which it is produced to the much smaller scales at which it is dissipated? Or is there something entirely dierent at work here? An even more fundamental question is: what do we mean when we talk about scales of motion at all? When we look up at the clouds or the exhaust from a smoke stack, we can certainly imagine that we see dierent scales of motion. The same is true when we observe the breakdown of the smoke rising from a smoldering cigarette or match, or as we pour cream into a cup of coee. But to carry out a physical analysis, we must somehow translate our visualization of scales into mathematics of scale. And that is much more dicult. There are many ways we can imagine spatially averaging, ltering (or windowing) the information about turbulence that we see or measure or simulate. For example, we can easily invent a volume-averaged turbulent velocity eld, (v) ui ( , t) dened by x 1 ui ( x , t)dx x (10.1) v v This is just a convolution integral which produces a new velocity, uv , from the i original instantaneous eld by averaging it around some neighborhood v of the point . And of course in doing so we will smooth out (or average out) uctuations x that are smaller than the characteristic dimension of the volume over which we have averaged. This would seem to be exactly what we are looking for, since simply by changing the volume, or looking at the dierences between volumes of dierent sizes, we can look at any scale of motion we please. But now try to do the same volume-averaging to the Navier-Stokes equations. Immediately you will discover that you cannot compute anything without facing the same kind of closure problems that always result from averaging. No matter how we average, or over what scale, we are always left with trying to account for ui ( , t) = x
(v)
10.2. THE ANATOMY OF TURBULENT MOTIONS
187
what is aecting our averages by things that you thought we averaged out. When we made the continuum hypothesis, we ended up with stresses. When we used the Reynolds decomposition to produce the RANS equations, we ended up with Reynolds stresses. And no matter how we spatially average, we will always end up with sub-volume stresses. Such volume averages are the basis most so-called large eddy simulations (or LES), and nding suitable closure models for the sub-volume (or sub-grid) scale stresses represents one of the greatest challenges. Unfortunately, in turbulence (like the stresses in many non-Newtonian uids as well), often the eects of the sub-grid stresses are not local (in time or space), so simple gradient hypotheses do not work very well. Obviously, just as we must look at molecular structure to understand non-Newtonian eects in uids, or look at grain structure to understand some solid materials, we need to gure out a way to literally tear the turbulence apart to understand what is really going on. This chapter and the ones that follow it are all how to do this, and do it in a way that leads us to dynamical equations. Only by analyzing such equations can we truly say we understand the physics.
10.2
The anatomy of turbulent motions
Clearly we need a way to describe mathematically the dierent scales of motion. Moreover, it can only be useful to us (wearing our engineering hats) if it leads us to a set of equations that describe these dierent scales. And these equations must tell us how the dierent scales transfer energy to each other, and how they get energy from the mean ow in the rst place. These are among the most dicult questions in turbulence for the physicists and mathematicians, and the successful resolution will be of enormous importance for engineers. In this chapter we will look at one very powerful approach to this problem. Now it is really simple to see the evolution from large scales to small scales in nature. All you have to do is look into a coee cup after you have added cream and stirred. But it is really dicult to try to quantify exactly how all this relates to simple vortical structures. Even when we succeed, it is not at all clear how these structures (or vortices) at dierent scales interact with (or evolve from) each other. Worse, though, it is almost impossible to write down equations to describe what we see. Think of the following analogy with biology. It is one thing to talk about the anatomy of the human body, another to talk about its physiology. Anatomy describes what is there (like arms and legs and internal organs), physiology explains how things work. Both are essential, but anatomy alone would be pretty useless, and physiology alone impossible. But at the same time, imagine trying to even talk about physiology without having anatomical terms so we know which part of the body is being analyzed. So it is with turbulence: the anatomy must allow us to do physiology. Words that do not lead to equations are of very limited value, no matter how vivid and accurate their description of the ow.
188
So we need a mathematical way to break turbulent velocities down into various scales of motion. Now it is here that most introductory books simply give up, and either try to convince you (usually with lots of hand-waving and verbal descriptions) that it is all obvious. Or else they go immediately into spectral or structure function analysis, leaving you mostly on your own to gure out why these are important. The former approach cannot possibly succeed since you dont have to go very far to realize something is missing; namely any way to connect all these mental pictures with equations. And the latter leaves you shaking your head wondering what happened, why we switched to Fourier space, and how these powerful mathematical techniques could possibly have any relevance to real ows. I am going to try something totally dierent in the history of introductory turbulence, so you have a chance here to either make me look like an educational pioneer or a complete fool. It is my view that most students are smart enough and know enough simple math that we can provide a satisfactory and correct answer to the question of scales in turbulence motion. This will eventually lead us to a subject that strikes terror into the hearts of even many senior turbulence investigators: the POD. But dont worry. We only need some rather simple calculus. Luckily for us, we will nd it leads us to a gold mine of well-known mathematical results accumulated over the past 200 years. Now the method we shall discuss is surely not the only way to decompose a turbulent ow. And it may not even ultimately prove to be the best way. But it certainly will be seen to be a sure way to both break the ow down into things which can be sometimes be interpreted as scales of motion. From it, we shall see that spectral analysis falls out rather naturally as the ideal way to decompose a homogeneous ow. Most importantly, whether we can interpret the results as scales or not, we can always end up with equations which describe how each piece evolves and interacts with the others. The problem was originally posed for turbulence by John Lumley1 in the following manner: suppose we have a random velocity eld, ui (), where represents , t or some subset of them. We seek to nd a deterministic vector eld, say i (), x which has the maximum projection on our random vector eld, ui , in a mean square sense. In other words, we would like to nd a whole new deterministic eld represented by i () for which < |(ui (), ())|2 > is maximized. i If we think of both ui and i as functions belonging to Hilbert space (I bet this sends most of you scurrying for those dusty old math books), then this inner product can be written as : (ui (), ()) = i

V
ui () ()d() i
(10.2)
where the integral is over the entire domain dened by the physical extent of the turbulent eld.
Professor John Lumley, now at Cornell, was (and still is) one of the leading turbulence experts in the last half of the 20th century. And he is rather special to me since he was my Ph.d dissertation advisor.
1
10.2. THE ANATOMY OF TURBULENT MOTIONS
189
It is pretty easy to show by the calculus of variations2 that the appropriate choice of i () to maximize its projection onto the velocity eld is the solution to the following integral equation:
region
Rij (, )j ( )d( ) = i ()
(10.3)
where Ri,j (, ) is the two point correlation function given by: Ri,j (, ) < ui ()uj ( ) > and =< |2 | > . where is dened by:
(10.4)
(10.5)
=
region
ui ()i ()d()
(10.6)
Exercise: Prove that equations 10.3 to 10.5 result from maximizing the projection of () onto ui (): in a mean square sense. Note that you must rule out simply making as large as possible by dividing by the square root of its projection on itself, (i (), i ()); i.e., maximize:
[
region region
ui ()i ()d()
i ()i ()d()
]1/2
(10.7)
Alternatively, you can maximize (ui (), ()) subject to the constraint that (i (), ()) = i i 1. Either way the result should be equation 10.3. Equation 10.3 is an integral equation since the unknown function to be solved for, i (), occurs on both sides, only one of them inside the integral. Also note that the value of is itself unknown, and must be determined as a part of the solution. This appears quite complicated, but luckily for us, integral equations are well-known to mathematicians who have spent a few hundred years learning and writing books about them. This particular integral equation is a member of the family of Fredholm integral equations.3 Usually it does not have a single solution set i , , but many. In general, the number, nature and character of these solutions depends on both the nature of the kernel and the type of region over which the integral is taken. Thus, nding the very best choice of our special deterministic function, i (), to represent the random velocity eld, ui (), has reduced to nding a solution to an integral equation for i () in which the kernel is given by the two-point correlation function, Ri,j (, ). Now this might not have come as a surprise to you; but if not, you may well be the rst person ever for which that is true.
2 3
It only took me about 20 years, but it really is trivial Fredhom was a Swedish mathematician of the 19th century.
190
The two-point correlation is, of course, itself deterministic since it is obtained from averages of the velocity eld. So all of the information about our particular solution to the Navier-Stokes equations (our ow of interest) is contained in this kernel. As we shall see, the big problem in applying equation 10.3 is in nding out sucient information about Ri,j (, ) to make the integration possible (if it is possible at all).
10.3
Fields of Finite Extent
The most familiar application of equation 10.3 is to ows in which the region is of nite extent in one or more directions (or time), either naturally or because of articially imposed boundaries. If the eld is of nite total energy, then the classical Hilbert-Schmidt theory applies. (Another quick trip to the math books might help here too). According to the Hilbert-Schmidt theory of integral equations, there is not just a single solution to equation 10.3, but there are denumerably innitely (n) many solutions (or eigenfunctions), i (), n = 1, 2, . For each eigenfunction, there is a corresponding eigenvalue, n . Moreover, these solutions are orthogonal. (I sure this brings back lots of fond memories from your math courses, right?) All this means is that: (p) (n) i ()i ()d() = pn (10.8)
region
Now since we have all of these solutions and they are orthogonal, we can reconstruct the original velocity eld from them as: ui () =
n=1
an i ()
(n)
(10.9)
The random coecients an are functions of the variables not used in the integral, and must be determined by projection back onto the velocity eld; ie
an =
region
ui () i ()d()
(n)
(10.10)
It is easy to show using the fact that the (n) () are orthonormal that: n =< an am > mn (10.11)
Thus the random coecients are uncorrelated. And, since we can reconstruct the velocity, of course we can reconstruct the two point Reynolds stress tensor, Ri,j (, ). The result after self-multiplication and averaging of equation 10.9 is: Ri,j (, ) =
n=1
n i ()j ( )
(n) (n)
(10.12)
The eigenvalues are ordered (meaning that the lowest order eigenvalue is bigger that the next, and so on); i.e, 1 > 2 > 3 . Thus the representation is optimal
10.4. HOMOGENEOUS FIELDS
191
(or proper) in the sense that the fewest number of terms is required to capture the energy. This is VERY important, and frankly it is the only reason any one cares at all about the POD. It is truly the most ecient way to break-down a eld of nite total energy into pieces, at least from the perspective of the energy. Thus the POD has provided several insights and possibilities: First, because of the nite total energy it has produced a denumerably innite set of orthogonal functions which optimally (in a mean square sense) describe the ow. This should make you feel VERY comfortable if you have had an applied math course, since you know lots of examples of orthogonal functions and can appreciate well what their advantages are. Dont be too bothered by the fact that we really dont know in general what our particular orthogonal functions are, and may not even be able to nd them analytically. But we do know how to nd them numerically and empirically if we have enough information about the two point velocity correlation. We simply put our measured two-point correlation function into the integral of equation 10.18, and use some appropriate method method to calculate them. Usually this is by discretizing the integral rst, so it becomes a matrix equation. Then with an appropriate subroutinue, out they pop. Second a nite subset of these functions can be used to produce a nite number of equations for analysis. This is accomplished by using them in a Galerkin projection on the governing equations (in our case the instantaneous Navier-Stokes equations). By truncating after a specied number of modes, the innitely dimensional governing equations are reduced to a nite set. We are not going to talk about this much in this course, but if you are into using the POD for control, or even if you are carrying out DNS (Direct Numerical Simulations) or LES (Large Eddy Simulations), you probably will nd someone using them, even just to clean up numerical data. POD (Proper Orthogonal Decomposition) techniques are currently in vogue to generate appropriate bases for dynamical systems models of turbulence (v Holmes et al 1996, Glauser et al 1993), but they have been used for more than 30 years to investigate coherent structures in turbulence (eg Lumley 1967, George 1989b, Moin and Moser 1989).
10.4
Homogeneous Fields
Really interesting things happen to our optimal projection integral, equation 10.3, if the ow is homogeneous in . Recall that homogeniety means the statistics are x independent of origin. In particular, the two point correlation with separation x = x reduces to Ri,j ( , x ) = Bi,j (). Note that by denition homogeneous r x r ows are not of nite total energy since they are of innite extent, so the HilbertSchmidt theory cannot apply to them. For elds homogeneous in , equation 10.3 can easily be shown to transform x to: x Bi,j (, t)j ( + )d = |2 |i ( , t) r x r r (10.13)
Since the (x) on the right hand side is a function of only, it can be included in x
192
the integral on the left so there is no x-dependence left on the right hand side; ie

x r j ( + ) Bi,j (, t) r d = |2 | r i ( , t) x
(10.14)
It is immediately obvious that solution itself (the term in brackets) must eliminate the x-dependence on the left hand side. There is really only one function which can accomplish this miracle, the exponential function. Therefore the eigenfunctions must be of the form ( ) x ] where is a wavenumber vector and all values of it are possible; ie exp[ik x k < . The coecients of the exponential, say ui ( t), can be shown to < k k, be given by 1 x ui ( t) = k, ui ( , t)eik d x x (10.15) 2 and the velocity eld can be reconstructed from them by

ui ( , t) = x
x ui ( , t)eik d x k
(10.16)
Thus, unlike the nite total energy case considered above, the solutions to equation 10.3 for homogeneous elds do not need to be empirically determined. They are ordinary analytical functions and we know what they are; they are the familiar Fourier transform which depends on the continuous vector variable k. This is indeed a wonderful surprise, mostly because over the past few hundred years we have learned more about the mathematics of Fourier transforms than just about any other function. Even though we are dealing with particularly nasty functions these functions are not only random, but may not even go to zero at innity (since the eld is homogeneous) there is a whole body of mathematics, the theory of generalized functions, to assure us that we are still able to proceed. Exercise: Consider a random eld of nite total energy which is a function of one variable only, say u(r). Now the integral equation of equation 10.18 becomes:
R(r, r )(r )dr = (r)
(10.17)
where the kernel is given by the cross-correlation, R(r, r ) = u(r)u(r ). Imagine that you only have data for a xed number of points, say only r, r = nr. Show how you can discretize this into a standard matrix eigenvalue problem which can be solved numerically. Exercise: Now consider the same problem as above, but this time show how you might develop a solution by successive approximations, where you guess a solution, substitute it into the left-hand side and integrate to obtain the right hand side. Then use the new right-hand side to substitute into the the left-hand side again, etc. etc. For a simple example, let R(r, r ) = exp[A(r + r )].
10.5. ARE HOMOGENEOUS FIELDS AND PERIODIC FIELDS THE SAME?193
10.5
Are Homogeneous Fields and Periodic Fields the Same?
Contrary to popular assumption (especially in the DNS and LES communities), these are not the same thing. The velocity eld is said to be periodic in the variable x if u(x) = u(x + L) where L is the period and the dependence on the other variables has been suppressed for now, as has the fact that the eld is a vector. Homogeniety, on the other hand, means the statistics are independent of origin. If a ow is homogeneous in a single variable, say x, then the two point correlation with separations in x reduces to R(x, x ) = R(r) where r = x x is the separation. By denition, homogeneous ows are not of nite total energy since they are of innite extent, so the Hilbert-Schmidt theory cannot apply to them. By oontrast, periodic elds are of nite total energy only if a single period is considered, since otherwise they repeat to innity. Now if periodicity and homogeniety are so dierent, why does the confusion arise? The POD provides the answer. For elds homogeneous in x, equation 10.3 can be shown to transform to
R(r)(x + r)dr = (x)
(10.18)
Since the (x) on the right hand side is a function of x only, it can be included in the integral on the left. Since there is now no x-dependence left on the right hand side, it is immediately obvious that solution itself must eliminate the xdependence on the left hand side. Therefore the eigenfunctions must be of the form (x) exp (ikx) where k is a wavenumber and all values of k are possible; ie < k < . The coecients, u(k), can be shown to be given by u(k) = 1 u(x)eikx dx 2 (10.19)
and the velocity eld can be reconstructed from them by
u(x) =
u(k)eikx dk
(10.20)
Thus, as noted earlier, the optimized projection integral for homogeneous elds reduces to the familiar Fourier transform which depends on the continuous variable k, so the number of eigenfunctions is non-denumerable. The situation for periodic elds is almost the same, but not quite and that little dierence is at the root of the confusion. Any periodic eld, even a random one, can be represented by a Fourier series; ie u(x) =
n=
an ei2nx/L
(10.21)
194
where the an are random and are determined in the usual manner. Using the orthogonality, the two-point correlation function can be written as R(x, x ) =
n=
< |an |2 > ei2n(x x)/L
(10.22)
Thus the two-point correlation for periodic ows, like homogeneous ows, depends only on the dierence variable r = x x. Hence the eigenvalue problem of the POD reduces to exactly the form of equation 10.18, except now the limits of integration are (L/2, L/2). It is easy to see that the POD modes must also be harmonic functions, like those for homogeneous ows. But there is a very important dience which is obvious from the integral: for periodic ows the wavenumber must be given by k = 2n/L and n can only take integer values! Moreover, the number of POD modes is now denumerably innite instead of being non-denumerable (ie continuous in k). Therefore the POD modes and the Fourier modes are identical. Thus the use of Fourier series to represent periodic elds is indeed optimal, at least in a mean square sense. Now the relation between a boxed homogeneous eld and a periodic eld can be readily determined by noting that because the energy is made nite by the box, the Hilbert-Schmidt theory again applies; hence the number of eigenfunctions becomes denumerable. If the kernel of boxed eld is now in addition assumed to be periodic, the Fourier series representation above follows immediately. Thus the periodic elds usually assumed for calculation are dynamically equivalent to a boxed homogeneous eld with the additional assumption of periodicity of the instantaneous elds. The assumption of periodicity has not only made the eigenfunctions denumerable, but it has forced the phase relations of all the scales, and this must also be of particular concern for the largest ones. Such calculations of bounded elds, like their experimental counterparts, can only be representative of homogeneous elds for scales of motion much smaller than the computational box (or lowest wavenumber) and for limited times. Whether current computations are acceptable is open to debate. My own view is that the best test of whether the eld is a good model of a truly homogeneous ow is best measured by its departures (or lack of them) from similarity theory. We will talk about this in detail in the next chapter, but recognize that the whole subject is being debated right now, sometimes rather hotly. Who knows, you might be lucky enough nd yourself in the middle of this debate quite soon.
10.6
Inhomogeneous elds of Innite Extent
None of the approaches above applies to ows which are inhomogeneous, but of innite extent (like most shear ows in the streamwise direction). In fact, it has not been at all clear until recently whether the integral of equation 10.3 even exists in such cases. All attempts to-date to apply the POD to the ow in these inhomogeneous directions have ended up applying the Hilbert-Schmidt
10.6. INHOMOGENEOUS FIELDS OF INFINITE EXTENT
195
theory to articially truncated, nite regions of the ow. And as a result, the eigenfunctions and eigenvalues found are dependent on the particular domain included in the decomposition. Clearly this is because it is the nite domain itself which is making the energy nite. Recently, however, one of my Ph.d. students, Dan Ewing 1995 (see also Ewing and George 1995) was able to show that if similarity solutions of the two-point Reynolds stress equations were possible, then the POD could be applied in similarity coordinates and the eigenfunctions were harmonic functions in it. By using a logarithmic coordinate transformation he was able to identify a number of ows for which two-point similarity was possible, thus for these ows the eigenfunctions are known analytically. Most importantly, the eigenfunctions were independent of the domain, at least in principle. For the far axisymmetric jet, the appropriate modes were u(, x) x1 exp(i) (10.23) where ln x/Lo (10.24) and Lo is prescribed by the initial conditions.4 Thus two-point similarity and equation 10.3 have yielded an optimal set of eigenfunctions into which the ow can be decomposed. The two point correlations, Rij (x, x ) =< ui (x)uj (x ) >, could all be expressed in the form, Rij (x, x ) = Q(x, x )exp[i( )] = Q(x, x )exp[i ln x /x] (10.25)
where Q(x, x ) = Us (x)Us (x )d/dx and for this ow Us (x) 1/x and d(x)/dx = constant. Note the dependence of the correlation in similarity variables on , an obvious counterpart to the x x dependence of homogeneous ows. Now these functional forms are interesting for a couple of reasons. First, because they settle the question of whether equation 10.3 can be applied to a ow of innite extent that is not homogeneous: It can! Second, for similarity ows of innite extent, the optimal basis functions are analytical functions, and they are harmonic functions in the similarity variable = ln x/Lo . Third, there is a continuum of eigenfunctions since all values of the reduced wavenumber, , are possible; ie < < . This last fact is the most interesting of all since it represents the counterpart of the homogeneous analysis above. Hence the denumerable POD modes of the Hilbert- Schmidt theory for an inhomogeneous nite energy ow have given way to the non-denumerable modes of Ewing. Thus, once again, the POD suggests that conning a ow changes the fundamental nature of it, consistent with observation. There is at least one more interesting aspect of the these inhomogeneous eigenfunctions. It is easy to show by expanding the logarithm of equation 10.25 that the
Interestingly, no length scale can be formed for a point source jet from the two parameters available, the kinematic viscosity and the rate at which momentum is added per unit mass. 2 Hence Lo must depend on nite source eects, like perhaps (Bo /Mo )1/2 where Bo is the rate of mass addition per unit mass (v George 1989a).
4
196
limiting forms of at least these inhomogeneous eigenfunctions are ordinary Fourier modes. From its Taylor expansion about x = x , ln x /x = (x x)/x + . It follows for small values of (x x)/x that Rij exp [ik(x x] where k is the ordinary, but local, wavenumber dened by k = x. Thus the usual assumptions of local homogeniety and the use of spectral analysis for the small scale motions are justied, at least in this case. Whether this is a general property of the POD is still very much the subject of debate (cf Holmes et al 1996).
Bibliography
[1] Ewing D (1995) On Multi-point Simlarity Solutions in Turbulent FreeShear Flows. PhD diss., Dept Mech Engr, SUNY/Bualo, Bualo, NY.a [2] Ewing D and George WK (1995) Similarity Analysis of the Two-Point Velocity Correlation Tensor in the Turbulent Axisymmetry Jet. Turbulence, Heat and Mass Transfer, Lisbon 1994 (Hanjalic and Pereira, eds.), Begell House Inc., NY, 49 56. [3] Holmes P, Lumley JL and Berkooz G (1996) Turbulence, Coherent Structures, Dynamical Systems and Symmetry CUP, Cambridge, UK. [4] George WK (1989b) Insight into the Dynamics of Coherent Structures from a Proper Orthogonal Decomposition. in Zorin Zaric Symp on Nearwall Turbulence, Dubrovnik, Yug (S.Kline ed), Hemisphere, NY. [5] Lumley JL (1967) The Structure of Inhomogeneous Turbulent Flows. in Atm Turb and Radio Wave Propag, Nauka, Moscow.
197
198
BIBLIOGRAPHY
Chapter 11 Decomposing Homogeneous Turbulence

Homogeneous turbulence provides us the only opportunity to study turbulence in the absence of turbulence transport and even more important, the absence of boundary conditions. Our study is facilitated by the results of section 10.4 which showed that homogenous turbulence is optimally decomposed into functions of exponential type. Thus the most natural way to analyze homogeneous ows is to use Fourier analysis. But because homogeneous ows are of necessity of innite extent, clearly Fourier integrals in the ordinary sense will not work since the integrals involving instantaneous quantities will not converge. Clearly we will need to generalize the usual denitions in order to make them useful. The classic texts have mostly approached this problem by using so-called Fourier-Stieljes integration, which can be quite forbidding to the new student to the subject (and many older students as well). Fortunately there is another much simpler way to approach the subject using generalized functions. (Note that the reader who is unfamiliar with this subject might rst want to study the appendices C, E, and F, which deal with temporal Fourier analysis only without the complications of a vector argument.)
11.1
Generalized functions
For homogeneous ows, we found that the appropriate choices of eigenfunctions to solve equation 10.3 were exponential functions. The coecients of these complex exponential eigenfunctions, say ui ( t), were given by: k, ui ( t) = k, 1 x ui ( , t)eik d , x x 3 (2)

(11.1)
and the velocity eld can be reconstructed from them by: ui ( , t) = x

x ui ( t)eik d k, k
(11.2)
199
200
CHAPTER 11. DECOMPOSING HOMOGENEOUS TURBULENCE
Now if you have studied Fourier transforms in an applied math course, you have probably already spotted one potential problem: the integrals of equations E.1 and E.2 may not even exist at least in the ordinary sense. A truly homogeneous eld has no spatial bounds and must be truly innite. Moreover, since its statistical properties are independent of origin, the velocity uctuations simply go on forever. Thus our random velocity eld is really rather nasty, mathematically speaking, and most certainly the integrals in the ordinary sense become unbounded. So we have a dilemma. Our attempt to nd the optimal way to decompose this ow has led us to Fourier transforms, but they do not seem to apply to the very problem which gave rise to them turbulence which is homogeneous. The answer lies in a major mathematical development of the 20th century the theory of generalized functions. There are numerous references which one can consult for a more proper mathematical treatment than the rather cursory and intuitive treatment here. (Lumley 1970, Lighthill 1955 are two of my favorites). In brief the basic idea is to replace functions whose integrals do not converge, with functions which do. Great idea, Im sure you are thinking, but doesnt this require magic? In truth it is almost magic, since in the end we almost never worry about what we have done, and almost always just go on doing regular mathematics like nothing ever happened. Impossible, you say. Lets consider a simple example in one dimension. Suppose I want to take the integral of the function, f (x) = 1, from (, ). Obviously this integral does not exist. Nor, in fact does its Fourier transform exist (in the ordinary sense). Now consider a second function, say: gL (x) = ex
2 /2L2
(11.3)
Now since the tails of this function roll-o exponentially, it certainly is integralable; in particular, dx = 2L (11.4) (You know this from Chapter 2, since (1/ 2L)exp(x2 /(2L2 ) is exactly the Gaussian which integrates to unity.) Our integrable function gL (x) also has a wonderful Fourier transform, wonderful in the sense that not only does it exist, all its derivatives exist also; i.e.,
ex
2 /2L2
1 ikx x2 /2L2 2 2 e e dx = 2Lek L /2 (11.5) FT{e }= 2 This is easy to compute by completing the square. So we have one nasty function, f (x) = 1, and one wonderful function, gL (x); the former has no integral, and hence no transform (in the ordinary sense), but the latter has both. Now note something interesting. The limit of gL (x) 1 as L , which is exactly the value of our nasty function, f (x). In fact, we could just dene a new function by the product fL (x) = f (x)gL (x) and note that:
x2 /2L2
11.1. GENERALIZED FUNCTIONS
201
lim fL (x) = lim f (x)gL (x) = f (x)

L
(11.6)
In fact, even more interestingly, the Fourier transform of our new function, fL (x), also exists in the ordinary sense. In this case, its just the Fourier transform of gL itself. Here is where one of the really good ideas of the last century appears1 , the magic if you will. Lets just dene the Fourier transform of our nasty function, f (x), in the sense of generalized functions to simply be the limit of the Fourier transform of fL (x) as L ; i.e, 1 ikx e f (x)gL (x)dx L 2
FTgf {f (x)} = lim FT{fL (x)} = lim

L
(11.7)
The Fourier transform of 1 in the sense of generalized functions is so useful, we have given it a special name, the delta-function; i.e., (y) lim GL (y)
L
(11.8)
where GL (y) can be any function whose integral is unity and which becomes undened at y = 0 and zero everywhere else in the limit as L . Im sure you have seen before, but you may not have realized that it was a generalized function. In general, the generalized functions are not uniquely dened. For example, all the functions below are suitable for dening (y): 1 2 2 GL (y) = ey /2L 2 y|/L G2L (y) = e sin(y/L) G3L (y) = y/L
(11.9) (11.10) (11.11)
The rst and last have continuous derivative everywhere, the second has a singularity at the origin. When working with Fourier transforms, it is generally best to dene them in terms of functions which both go to zero exponentially fast, and which have all derivatives continuous. There is nothing in this course which needs anything more than GL (y), the Gaussian version, or 2L times it. We can generalize this whole procedure to almost any arbitrary function, whether deterministic or random. For example, suppose we have a random homogenous function (in one variable), say v(x). Then we can dene its Fourier transform in the sense of generalized functions to be:
One of the rst to see this was the electrical engineer named Heaviside and he invented the step function which bears his name.
1
202
v (k) FTgf {v(x)} = =
lim FT{v(x)gL (x)} lim
(11.12) (11.13)
1 ikx e v(x)gL (x)dx L 2
where gL (x) can be any function for which the product v(x)gL (x) is integrable and for which:
L
lim v(x)gL (x) = v(x)
(11.14)
Obviously a suitable choice is the Gaussian function we started o with; i.e., gL (x) = ex
2 /2L2
(11.15)
Exercise: Show that the Fourier transforms in the sense of generalized functions of eikxo , coskxo and sinkxo are (xo ), [(xo ) + (xo )]/2 and i[(xo ) + (xo )]/2 respectively using the Gaussian version of gL (x) dened above. Exercise: Compute the inverse transforms from the above example. Do NOT use the short-cut version where you assume the properties of a delta-function, but instead work with the actual transformed version of f (x)gL (x) under the limit sign, then take the limits. For the rest of this course, we will simply agree that whenever there is any doubt, we always mean the Fourier transform in the sense of generalized functions. For example, when we take the three dimensional spatial Fourier transform of the velocity eld, ui ( , t), we really mean the Fourier transform in the sense of x generalized functions dened by: ui ( t) FTgf {ui ( , t)} k, x 1 i x = lim e k [ui ( , t)gL3 ( )]d x x x L (2)3 (11.16) (11.17)
where gL3 ( ) is some suitably dened function which makes the integral exist. An x excellent choice for gL3 ( ) would be: x gL3 ( ) = e[x1 +x2 +x3 ]/2L x
2 2 2 2
(11.18)
whose Fourier transform (in the ordinary sense) is given by: L3 2 2 2 2 e[k1 +k2 +k3 ]L /2 (11.19) 3/2 (2) We will use exactly this denition in Section 11.23 to show that Fourier coecients in non-overlapping wavenumber bands are uncorrelated. GL3 ( = k)
11.2. FOURIER TRANSFORMS OF HOMOGENEOUS TURBULENCE
203
Exercise: Find the Fourier transform of 1 in three-dimensions using generalized functions, then show how you might represented it symbolically as a three dimensional delta-function, ( k). Exercise: If the Fourier transform can be represented in the sense of generalized functions as (| o |), nd the inverse Fourier transform in the sense of k k generalized functions.
11.2
Fourier transforms of homogeneous turbulence
We have agreed already that we will always, when necessary, interpret our Fourier transforms in the sense of generalized functions. So if we agree to only transform over the space variables, we are left with the following Fourier transform vector pair: ui ( t) = k, ui ( ) = x 1 x d eik ui ( , t) x x 3 (2)
x d e+ik ui ( t) k k,
(11.20) (11.21)
Note that we have represented a triple integral by a single integral sign and moved the dierential, d or d next to the integral sign, so it will be obvious which x k, variables are being integrated. Also it is understood that everything to the right of the dierential is to be integrated over those variables. There is still one little problem. We set out to nd one deterministic vector function which best described our random eld, and we have ended up nding not just an innity of them (like for the inhomogeneous elds above), but in fact a continuum of them: the number of eigenfunctions is non-denumerable. As inconvenient as this might appear, that is the way things are. It is an inescapable consequence of the fact that we let the boundaries (and the energy) in the eld go to innity. But who cares, it is a small price to pay since we have this wonderful Fourier analysis tool to work with. Now Im sure you are asking: What is he so excited about? Why is the applicability of Fourier analysis such a BIG THING? There are two big reasons (among many). The rst has to do with what happens when you take the inverse transform of equation F.9 at the point x , multiply it by the complex conjugate of the inverse transform at point , and average to get the two-point correlation, x Ri,j ; i.e., x Ri,j (x , , t) = ui ( , t)uj ( , t) x x
dk
(11.22)
+i(km x kp xp ) m
d e k
i (k , t) ( t) u uj k,
204
But we have assumed the eld to be homogeneous so the two-point correlation can depend at most on the vector separation, = x ; i.e., r x x Ri,j (x , , t) = Bi,j (, t) r Therefore equation F.3 is simply:

(11.23)
Bi,j (, t) = r
dk
d e+i(km xm kp xp ) i (k , t) ( t) k u uj k,
(11.24)
and the left-hand side has no separate dependence on either or x separately, x but is only a function of . Now look carefully at the right-hand side. Clearly, r unless a miracle occurs in the integration, the right-hand side is going to always depend on x and . x Guess what? You probably guessed it. A miracle DOES occur well, not really a miracle, but even better than a miracle. This miracle can be proven to be true. The miracle is that since both sides of equation F.3 MUST depend only on = x , it follows immediately that the Fourier components in nonr x overlapping wavenumber bands must be uncorrelated. Say what, you say? Exactly this:
{
k i (k , t) ( t)dk d u uj k, or equivalently:
Fi,j ( t)d , k = k, k k = 0 ,k k
(11.25)
k) i (k , t) ( t) = Fi,j ( t)(k u uj k, k,
(11.26)
where ( ) is the familiar delta-function (not to be confused with the Kronecker delta tensor) and Fi,j ( t) is a deterministic function called the velocity crossk, spectrum tensor. It is easy to see by substitution that our two-point velocity correlation function is the three-dimensional inverse Fourier transform (in the ordinary sense) of the velocity cross-spectrum tensor; i.e.,

Bi,j (, t) = r
eikm rm Fi,j ( t)d k, k
(11.27)
It is a bit more dicult to show that the cross-spectrum is the three-dimensional Fourier transform (in the ordinary sense) of the two-point velocity correlation function; i.e., t) = 1 eikm rm Bi,j (, t)d r r (11.28) Fi,j (k, (2)3 Thus the cross-spectrum and the two-point correlation form a Fourier transform pair. Exercise: Use the denition of gL (x) in the preceding chapter and prove that (k) (k ) = F (k)(k k) if u(x) is a homogeneous random variable and u(k) u u
11.3. THE THREE-DIMENSIONAL ENERGY SPECTRUM FUNCTION 205 is dened in the sense of generalized functions by: u(k) = 1 2
dx eikx u(x)
(11.29)
Exercise: Carry out the same exercise and derive equation F.8 using gL3 ( ) x as dened in the preceding chapter. The implications of what we have accomplished become immediately obvious if we evaluate the inverse transform of equation F.9 at = 0 to regain the singler point cross-correlation; i.e.,
Bi,j (0, t) =
d Fi,j ( t) k k,
(11.30)
Clearly Fi.j ( t) is telling us how the single-point Reynolds stress, ui uj = k, Bi,j (0, t), is distributed over the various wavenumbers (or scales of the turbulence). This is even more obvious if we contract the two indices by letting i = j, sum and divide by two to get the energy; i.e., 1 2 1 1 q = Bi,i (0, t) = 2 2 2
d Fi,i ( t) k k,
(11.31)
The contracted cross-spectrum is usually called simply the energy spectrum, and it tells us exactly how the turbulence energy is distributed with wavenumber. But this is almost what we wanted in the rst place a way to tell the energy associated with one scale from the energy associated with another.
11.3
The Three-dimensional Energy Spectrum Function
Now it really is a nuisance to have to deal with the energy spectrum, Fi,i ( t), k, dened above. Even though it is a scalar, it is still a function of four independent variables, and t. This is, of course, a lot less nuisance than the full crossk spectrum tensor, Fi,j ( t)). Nonetheless, it is still rather dicult to draw pictures k, of things that depend on even three-variables, much less show their time dependence. So theoreticians have dened another kind of spectrum, E(k, t) which is dened as the integral of Fi,i ( t) over spherical shells of radius k | i.e., k, k|; 1 dS(k) Fi,i ( t) k, E(k, t) 2 k=| k|
(11.32)
where means a surface integral and dS(k) is a spherical area element. This has the advantage of having only the scalar argument, k, dened by:
2 2 2 k 2 = k1 + k2 + k3
(11.33)
206
This new kind of spectrum is properly called the three-dimensional spectrum function (to distinguish it from the contracted cross-spectrum we used above). But most people simply refer to it as just the spectrum or the energy spectrum, which is ne as long as there is no confusion. (Be careful in skipping from book to book and paper to paper since some books do not include the factor of 1/2 in the denition.) Most of the time the surface integral of equation 11.32 is carried out in spherical coordinates, (k, , ), so equation 11.32 becomes:
/2
E(k, t) =
=0 =/2
Fi,i ( t)k 2 sindd k,
(11.34)
where Fi,i ( t) must, of course, be written in spherical coordinates too. This has k, particular advantages for isotropic ow, as will be seen later. It is easy to see that the integral of E(k, t) over all wavenumbers yields the turbulence kinetic energy (per unit mass); i.e.,
1 ui ui = E(k, t)dk 2 k=0
(11.35)
This follows immediately from the fact that the integral of Fi,i ( t)/2 over all k, wavenumbers yields the kinetic energy (per unit mass) and the fact that we have already integrated over two of the variables in the denition of E(k, t). This is, of course, the reason why most people simply refer to E(k, t) as the energy spectrum. Unfortunately there are a lot of other denitions of spectra which integrate to the energy which will shall learn about below, so be sure both you and the person you are talking to agree on your denitions. Figure 13.1 shows an example of the energy spectrum function determined from an attempt using DNS to simulate isotropic turbulence in a computer. Figure ?? shows the energy spectrum function inferred from grid turbulence measurements of one-dimensional spectra (see below) at dierent positions downstream in a wind tunnel assuming isotropy (as described in a the next chapter). The problem with E(k, t) is that it is not simply related to the correlation functions, unless you are willing to make other assumptions like isotropy. This is because the value of E(k, t) at a given wavenumber really has information from all three spectral components in it, so information as been smeared out, or averaged. But this is not averaging in the sense that we have been using it at least not without other hypotheses as noted below. It is really only exactly what it says it is: an integration of the energy spectrum over a surface of radius k.
11.4
DNS turbulence
In fact, researchers who do DNS with spectral codes take this average over spherical shells of radius k quite literally and use it to their advantage in a very interesting manner. Since even a modest simulation of decaying turbulence takes
11.4. DNS TURBULENCE
207
0.03 time=1.326 time=1.700 time=2.219 time=3.687 time=4.050 time=4.711 time=6.000 time=7.340
0.025
0.02
0.015
0.01
0.005
10
15 k
20
25
30
Figure 11.1: DNS data of Wray [?] for decaying isotropic turbulence showing how E(k, t) decays with time. Note limited resolution at low wavenumbers.
208
hundreds of processor hours, it costs far too much to do more than a few runs. This means it is impossible to ever have enough data to do ensemble averages. One way to beat this problem is to compute so-called forced turbulence, in which the turbulence is forced continuously so the energy input exactly balances the dissipation rate. Such turbulence is then stationary, so the usual time averages can be performed. Unfortunately, many interesting ows, like decaying turbulence or homogeneous shear ow turbulence, are not stationary. So what the DNS people do is to pick the starting spectrum, E(k, 0), then take the amplitude of their Fourier coecients equal to the square root, so that the starting Fourier components are: E(k, 0) ui ( 0) = k, (k)3
1/2
ei(k)
(11.36)
where (k) is initial phase and k is the resolution between the selected wavenumbers. (You will understand this better after we consider nite domain transforms in the next chapter, but for now simply think of 1/k 3 as an approximations to the delta function ( we encountered above.) They then use a random k k) number generator to randomize the phases at each value of k. (Note that it is also important to make sure these initial values satisfy the incompressible continuity equation, which transformed becomes simply ki ui = 0.) Once the calculation begins, the Fourier-transformed Navier-Stokes equations govern the evolution of both the amplitudes and the phases. It is easy to see why these particular DNS techniques are still somewhat controversial, even after twenty years. You are working with only a single member of an ensemble, and a very special one at that. And you are interpreting the integration over the spherical shells of radius k as an average in the ensemble sense. Now this is really a very clever trick that makes the computations possible. But it assumes the turbulence evolves to a state independent of its initial conditions, consistent with the long held view about turbulence. Unfortunately even computations done in this way suggest quite strongly that the decay rate of the turbulence energy depends on how you begin; i.e., on the initial conditions. This is a very serious problem indeed, unless the only conditions that matter are the shape of the starting spectrum. At this moment, most people believe (some vehemently) in the asymptotic independence, which is the traditional view of turbulence. I am of the opposite point of view, but suspect strongly that all the necessary information about the initial conditions is in the starting spectrum. You will have a chance to judge for yourself whose views are more reasonable from the examples in the following chapters. In spite of those who would like to pretend there is no problem, the next decade may prove very interesting as computers get more powerful and tougher questions are asked of the simulations. And who knows, maybe you will be among those who will provide the answers to questions that many now are afraid to ask.
11.5. ONE-DIMENSIONAL SPECTRA
209
11.5
One-dimensional spectra
There is another type of spectrum, the so-called one-dimensional spectrum, we need to consider before we press on to the applications. In truth, for the theory we really dont need this now; but I know you will all be thinking about how what we talk about relates to things we can measure, right? So we might as well put all the cards on the table at once. The rst thing you need to know is that NONE of the spectra we talked about above can actually be measured in a turbulent ow or least havent yet. Think about what would be required. You would have to make two-point correlation measurements of all three components of velocity in all three directions, then Fourier transform them in all three directions. And you would have to have enough of such measurements to satisfy statistical convergence criteria, and over a large enough spatial extent to not have a window problem, and with enough resolution to not alias well you get the idea. The last two problems are tractable in most cases if special care is taken, but its the three components of velocity with separations in three directions that is the killer (at least for now maybe when holographic PIV techniques nally mature). Now you might conclude from the above that spectral analysis is impossible. Not quite. If we cant do exactly what we want, we invent something else. It turns out that for turbulent ows in which there is a substantial mean velocity, say U = (U1 , 0, 0) and |U1 | is a lot bigger than the rms turbulence velocity, the result of our measurement of a frequency spectrum with a xed probe, is pretty close to the one-dimensional wavenumber spectrum dened by: Fi,j (k1 )
(1)
Fi,j ( t)dk2 dk3 k,
(11.37)
In other words, the one-dimensional cross-spectrum is just the original crossspectrum with two of the wavenumber directions integrated out. The superscript 1 is necessary to name the function so we can tell it apart from the original. (Remember: you cant name a function by its argument.) The superscript could be 1, 2, or 3, and is determined by the one variable that is NOT integrated out. The way this happens is that the uctuations we see at a frequency f are mostly due to spatial disturbances of wavelength = U1 /f being convected past us. Thus what shows up at frequency f is really a disturbance at wavenumber k1 = 2f /U1 . Really it is all the same as simply saying that /t U1 /x1 . This interpretation is called Taylors Hypothesis and works pretty well as long as the local turbulence intensity is less than 10 20%. Above this the interpretation becomes very dicult. Now one-dimensional spectra have some very funny and strange characteristics, which really drive non-turbulence people crazy. The most annoying feature is that they always end up with a nite value of the spectrum at zero wavenumber. This can be very misleading, since in fact there is no energy at zero wavenumber ever! It has simply been aliased there by waves traveling sideways. This is illustrated
210
in Figure ?? which shows a probe traveling through a eld of xed waves, all of wavelength o = 2/ko , but with their wavefronts aligned in dierent directions. Only those with wavefronts exactly perpendicular to the direction of travel of the probe get measured at the right wavenumber, ko . All the rest are seen at ko cos where is the angle of incidence. Example: Consider a turbulent eld to have all its energy concentrated on one shell at wavenumber, | = ko . This can be represented as Fi,i ( = Fo (| ko ). k| k) k| (1) 2 2 Apply the denition of equation 11.37 to show that Fi,i = Fo [1 k /ko ]. Note that all of the energy in the eld is at | = ko , there is none at ko (or above) in k| the one-dimensional spectrum. The one-dimensional spectrum does have several nice features, however. The rst is that each one-dimensional spectrum is the one-dimensional Fourier transform of the corresponding two point correlation functions, Bi,j (r, 0, 0), Bi,j (0, r, 0), and Bi,j (0, 0, r); i.e., 1 ikr e Bi,j (r, 0, 0)dr 2 1 ikr (2) Fi,j (k) = e Bi,j (0, r, 0)dr 2 1 ikr (3) Fi,j (k) = e Bi,j (0, 0, r)dr 2 Fi,j (k) =
(1)
(11.38) (11.39) (11.40)
And in fact, these correlations can be shown to be the inverse Fourier transforms of the one-dimensional spectra; i.e.,
Bi,j (r, 0, 0) = Bi,j (0, r, 0) = Bi,j (0, 0, r) =
e+ikr Fi,j (k)dk e+ikr Fi,j (k)dk e+ikr Fi,j (k)dk

(3) (2)
(1)
(11.41) (11.42) (11.43)
Thus like there three-dimensional counterparts, they form a Fourier transform pair. Proof of equation 11.39 Start with equation F.9 and set = (r, 0, 0) to r obtain:

Bi,j (r, 0, 0) =
eik1 r Fi,j ( t)d k, k

[
ik1 r
dk1 e
Fi,j ( t)dk2 dk3 k,
(11.44)
But the term in brackets is just Fi,j (k1 ).
(1)
11.6. SPECTRAL SYMMETRIES
211
Proof of equation 11.39 Start with the denition of equation F.10 and substitute from equation F.9 to obtain: Fi,j (k1 ) =

(1)
Fi,j ( t)dk2 dk3 k,

[ ] 1 ikm rm e Bi,j (, t)d dk2 dk3 r r (2)3 {

(11.45)
}
1 dr1 eik1 r1 = (2)
dr2 dr3 Bi,j (, t)(r2 )(r3 ) r
where the delta-functions result from the following double Fourier transform in the sense of generalized functions of 1; i.e., 1 ik2 r2 k3 r3 e 1 dk2 dk3 = (r2 )(r3 ) (2)2 Integration over r2 and r3 yields immediately equation 11.39. An interesting feature of the one-dimensional spectrum is that its value at the origin (k = 0) is proportional to an integral scale. For example, to obtain the (1) longitudinal integral scale you would obtain from the two-point correlation, B1,1 , (1) start with the corresponding one-dimensional spectrum. F1,1 and set k = 0 to obtain:
(1) F1,1 (0)
(11.46)
1 = B1,1 (r, 0, 0)dr 0 (1) u2 L1,1 1 =
(11.47)
where we have used the symmetry of B1,1 (r, 0, 0) about r = 0. In fact, this is usually the best way to determine an integral scale, since you can use Fast Fourier Transform algorithms to speed up the computations (as shown in the appendices) and completely by-pass the computation of the correlation all together.
11.6
Spectral symmetries
The symmetries encountered in Section 9.4 all have their counterpart in spectral space. The Fourier representation, however, is generally complex; so what might be simply a symmetry about the origin shows up as a Hermitian property in the cross-spectrum. For example, consider the cross spectrum which is the three-dimensional Fourier transform of the cross-correlation Bi,j () and given by: r Fi,j ( = k) 1 (2)3

eikm rm Bi,j ()d r r
212
CHAPTER 11. DECOMPOSING HOMOGENEOUS TURBULENCE 1 (2)3 1 = (2)3 = Fj,i ( k) =

eikm rm Bj,i ()d r r e+ikm rm Bj,i ()d r r (11.48)
since Bi,j is real. But also it follows that:
Fi,j ( = k)
1 (2)3 1 = (2)3 = Fi,j ( k)
eikm rm Bi,j ()d r r e+ikm rm Bi,j ()d r r (11.49)
This last property is called Hermitian and corresponds to symmetry of the crossspectra.
11.7
Consequences of incompressibility
In section 9.5 we explored the consequences of the incompressible continuity equation on the two-point correlation tensor. To see the counterparts for Fourier analysis, consider rst the continuity equation for just the instantaneous uctuating velocity given by uj /xj = 0. Fourier transforming this in the sense of generalized functions yields: 1 (2)3 1 (2)3

ikm xm
uj ( , t) x d = 0 x xj

(11.50)
This can be integrated by parts to obtain:

[
ikm xm
] uj ( , t) x 1 d = [ikj ] x xj (2)3 = [ikj ]( t) u k,
eikm xm uj ( , t)d x x (11.51)
Thus the counterpart to the instantaneous continuity equation in Fourier space is: [ikj ]j ( t) = 0 u k, (11.52) Equations 9.43 and 9.44 can be Fourier transformed in a similar manner to obtain their counterparts in spectral space as: kj Fi,j ( = 0 k) ki Fi,j ( = 0 k) Exercise Prove these. (11.53) (11.54)
11.8. IMPLICATIONS OF ISOTROPY ON SPECTRA
213
11.8
Implications of Isotropy on Spectra
By direct analogy with the two-point correlations it follows that the most general forms of the two-point cross-spectral spectra in an isotropic eld are given by: F, ( = F (k) k) ki Fi, ( = FL (k) k) k = C(k)ki kj + D(k)i,j Fi,j (k) (scalar) (vector) (tensor) (11.55) (11.56) (11.57)
where F , FL , C and C are all functions of k = | only. k| It is straightforward to show (using continuity and the denitions) that C(k) and D(k) are simply related to the energy spectrum function, E(k), so the most general form of the two-point vector-vector spectrum in an isotropic eld is given by: E(k) 2 [k ij ki kj ] Fi,j ( = k) 4k 4 (11.58)
Exercise: Prove equation 11.58 Hint: use the denition of E(k) and change the surface integral to spherical coordinates using dS(k) = k 2 sin dd. This isotropic spectral equation also has important implications for the onedimensional spectra as well. In particular, it is straightforward to show that:
(1) F1,1 (k1 )
1 E(k) 2 2 = [k k1 ]dk 3 2 k1 k 1 E(k) 2 (1) 2 F2,2 (k1 ) = [k + k1 ]dk 4 k1 k 3
(11.59) (11.60)
Exercises: Prove equations 11.59, and 11.60. Hint: substitute the general isotropic expression into the denition of the one-dimensional spectra and change 2 2 to polar coordinates, noting that d = kdk where 2 = k2 + k3 . It follows by repeated dierentiation (and application of Leibnitz rule for differentiating under the integral sign) that: dF k 1,1 E(k) = dk (1) d 1 dF1,1 = k3 dk k dk Exercises: Prove equations 11.61 and 11.62.
2 (1) 2 d F2,2 k dk 2 (1)
(11.61) (11.62)
214
Chapter 12 Measuring spectra from experimental and DNS data

12.1 Introduction
The theoretical considerations of the next chapters will make clear the importance of turbulence spectra. It is important both to be able to measure them experimentally, or alternatively, calculate them from DNS data. Unfortunately, there are several types of problems which one encounters, some of which are common to both, others unique to each. All are extremely important, and failure to properly understand the limitations can render the data useless. The primary common problems for both DNS and experiments are the statistical errors one encounters from having only a limited number of samples, and the problems related to the nite spatial extent of the simulation or experimental facility. The sampling problems are the exact counterpart to those we encountered in Chapter 2 when making statistical estimates of single point quantities. This statistical uncertainty complicates all spectral data interpretation, especially the identication of spectral peaks and roll-os. The nite domain problems arise from the fact that no experiment or simulation can be truly homogeneous, hence can at best be a model of homogeneous ows over some limited range of scales or times. It may seem that these dierences are unimportant, but failure to recognize them can result in models for turbulence which depend more on the boundary conditions of the experiments (or DNS) than on the dynamics of the turbulence. Also, the nite size of measuring arrays or windows can itself create spectral leakage This chapter will focus on how to obtain a velocity spectrum from time-varying or spatially vary data from one or more probes. We will rst examine how one processes data from time-varying signals, then we will discuss how such data can be interpreted as a space-varying using Taylors hypothesis. Finally the ideas will be generalized to making spatial measurements directly. 215
216CHAPTER 12. MEASURING SPECTRA FROM EXPERIMENTAL AND DNS DATA
12.2
Determining the wave-number spectrum from the spatial auto-correlation
The most obvious way to obtain a frequency spectrum is to measure the two-point correlation to as large a separation as possible, then Fourier transform it to directly obtain the corresponding one-dimensional spectrum. For example, suppose you have two probes, one xed, the other movable, each measuring the streamwise component of the velocity. You would like to use them to measure B1,1 (r, 0, 0) where: B1,1 (r, 0, 0) < u1 (x1 , x2 , x3 )u1 (x1 + r, x2 , x3 ) > (12.1)
And you would then like to Fourier transform B1,1 (r, 0, 0) to obtain the one(1) dimensional spectrum F1,1 (k1 ) where:
(1) F1,1 (k1 )
1 = FT{B1,1 (r, 0, 0)} = B1,1 (r, 0, 0)ek1 r dr 2
(12.2)
But there is a limit to how far apart you can put your probes because of the nite extent of your facility. If your xed probe is at the center, the most you can measure is L/2 < r < L/2; i.e, you can only obtain:
{
B1,1L (r1 )
B1,1 (r1 , 0, 0) 0
, L/2 < r < L/2 , otherwise
(12.3)
Therefore you really cant perform the integration above, even though your estimate of B1,1 is perfect for the separations you can measure (and of course it is never is, due to the statistical error we will discuss later). The most you can possibly compute is the nite space transform given by: F1,1L (k1 ) =
(1)
1 L/2 B1,1L (r1 )ek1 r1 dr1 2 L/2
(12.4)
We can re-write this as the product of the true correlation, B1,1 (r1 , 0, 0), and a window function as:
(1) F1,1L (k1 )
1 B1,1L (r1 )wL (r1 )ek1 r1 dr1 = 2

{
(12.5)
where wL (r) is the top-hat window function dened as: wL (r) = 1 0 , |r| L/2 , |r| > L/2 (12.6)
Parsevals Theorem tells us that the Fourier transform of the product of two functions is the convolution of their Fourier transforms. Thus our measured spectrum is given by:
12.2. DETERMINING THE WAVE-NUMBER SPECTRUM FROM THE SPATIAL AUTO-CORR
F1,1L (k1 ) =
(1)
F1,1 (k1 k)wL (k)dk =
(1)
F1,1 (k)wL (k1 k)dk
(1)
(12.7)
where w(k) is the Fourier transform of the top-hat window function given by: L wL (k) FT{wL (r)} = 2
( )
sin(kL/2) (kL/2)
(12.8)
It is clear that in spite of our best eorts, we have not measured the true spectrum at all. Instead all we have obtained is a garbled version of it which ltered through the window function, w(k). The entire process is very much like looking through a glass-window at something on the other side. If the glass is not of high quality, then the image we see is distorted, perhaps even to the point where we cannot even recognize it. This can happen with spectra as well, so it is very important understand what the window has done. Figure ?? shows both the top-hat and its Fourier transform. Note that most the area under w is between the zero-crossings at k = /L. But notice also the other peaks at higher and lower wavenumbers, and especially the rather strong negative peaks. These side-lobes roll o as |2/kL|; but even so they can cause considerable leakage from high spectral values to low ones, and also cause false peaks due to the negative values. Obviously, the larger the domain (L/2, L/2), the closer the narrower the lter, and the closer the window approaches a delta function, (k). In the limit of innite L, the true spectrum is recovered, since convolution of any function with a delta function simply reproduces the original function. True spectra can never be negative, of course. But measured spectra can be, so they must be interpreted carefully. Usually it is desirable to reduce these unphysical and spurious values by introducing additional window functions into the data processing before Fourier transformation. (Afterwards is too late!) Popular choices are the hanning window, the hamming window, and the Parzen window. All of these choice pre-multiply the measured correlation by a function which rolls o less abruptly than the top-hat. This reduces the side-lobes and their bad inuences, usually by making them roll-o more rapidly. But it also reduces the resolution in wavenumber space by making the eective span-wise extent less and the band-width greater. As a result, there is no simple best way to process data. Each experiment must be considered separately, and each result carefully analyzed to see which peaks are real, which are false, which are reduced by leakage, which roll-os are physical and which are not. Usually this is accomplished by analyzing the same data with several dierent windows. Then, with luck, it will be possible to infer what the real spectrum might be. But the entire process is really an art. Like any art skill, the more practice you have, the better you get at making the right choices. And like any newly acquired skill, beginners be especially careful.
218CHAPTER 12. MEASURING SPECTRA FROM EXPERIMENTAL AND DNS DATA
12.3
The nite Fourier transform
There is another way to obtain the spectrum of a signal without rst computing the autocorrelation direct transformation of the incoming signals. The basic idea is quite simple and follows directly from the denition of the Fourier transform of the signal. For example, consider the same velocity component, u1 ( , considered x above with x2 and x3 held constant. The one-dimensional transform in the x1 direction of this velocity component is given by: Now again in the real world, our information is limited to a nite domain, say (L/2, L/2) as before. Therefore the most we can really compute from our data is:
12.4
12.4.1 12.4.2
Taylors Hypothesis
The Frozen Field Hypothesis The Eect of a Fluctuating Convection Velocity
12.5
Resolution and Spatial Filtering
Chapter 13 Dynamics of Homogeneous Turbulence

13.1 The Fourier transformed instantaneous equations
We are about ready to apply our results from Fourier analysis to the dynamical equations for homogeneous turbulence. Now with all the dierent types of spectra ying around it would be really easy to forget that this was, after all, the whole point of decomposing the instantaneous velocity in the rst place. But our original purpose was to investigate the dierent roles that the dierent scales of motion play in the dynamics of the motion. And dynamics means the interplay of accelerations and forces; so back to the instantaneous equations we must go. Now there are two ways we can proceed: We could Fourier transform the instantaneous equations for the uctuations in a homogeneous ow. Then we could examine these, and from them even build equations for the energy spectra by using the Wiener-Kinchine relation. Alternatively we could build a set of equations for the two-point velocity correlations, and transform them to get the spectral equations. Either way, we end up in the same place with a set of equations for the velocity cross-spectra, which we can then integrate over spherical shells of radius k to get an equation for E(k, t). Lets try both, since each has a unique piece of information. Lets begin by simply using equation 3.27 for the instantaneous uctuating velocity. For now lets just assume there is no mean velocity at all and that the ow is homogeneous. If we substitute equation 11.21 for the instantaneous velocity and dene a similar transform for the uctuating pressure, the integrand of our transformed equation reduces to:
ui ( t) k, + (ikj )j (k )i (k t)dk u u k, t 1 k, k, = ki p( t) 2k 2 ui ( t)
(13.1)
219
220
CHAPTER 13. DYNAMICS OF HOMOGENEOUS TURBULENCE
The second term on the left-hand side is a convolution over the wavenumber vector and results from the fact that multiplication in space corresponds to convolution in wavenumber, and vice versa. Since we are assuming incompressible ow, the pressure term can also be expressed in terms of the velocity using the continuity equation. For an exercise see if you can show this. Note that like the velocity uctuation itself, the Fourier velocity coecients, t), are themselves random and of zero mean. Thus each realization of them ui (k, will be dierent, and it is only the spectra formed from them which will be deterministic. This means averaging in some way is probably required at some point. But even without any kind of averaging, equation 13.1 is a very important equation because even a cursory examination tells us a lot about turbulence. First note that the viscous terms are weighted by a factor of k 2 , compared to the the temporal decay term. Obviously viscous eects are concentrated at much smaller scales (higher wavenumbers) than are inertial eects. Second, note that the non-linear terms show up in a convolution involving three k. wavenumbers: k , and also k This is the only place the non-linearity shows k, up. Since we know non-linearity is the essence of turbulence, then we can say for sure that turbulent non-linear interactions involve only triads of wavenumber vectors. Some possibilities are shown in Figure ??. The study of which triads dominate the energy transfer has been important to turbulence for a long time. Although all triads can be important, it is generally believed that the local interactions are the most important; i.e., those for which all three wavenumbers have about the same magnitude. This can be very important in making models for the turbulence like LES, for example. So what have we learned so far? That the energetic scales decay primarily by the non-linear triad interactions which move the energy to higher and higher wavenumbers. Eventually the viscosity comes to dominate these high wavenumbers and the energy is dissipated at the smallest scales of motion. Isnt it amazing how we can learn all this just by examining an equation without ever solving it? The neat thing is that once you learn to do this for turbulence you can do it for just about every eld of science or engineering. The study of turbulence really has changed your life. The discretized version of equation 13.1 is the basis for most spectral method DNS solutions (and pseudo-spectral) method solutions using very large parallel computers. Modern Fast Fourier Transform (FFT) algorithms allow huge arrays to be rapidly manipulated much more quickly than nite dierence techniques. The emergence of these techniques, especially over the past decade, has been one of the most exciting aspects of turbulence research. And it gets even more interesting with each increase in computer capacity. Yet as we noted above, though, there are some fundamental questions which remain, some of which we will mention below.
13.2. THE SPECTRAL EQUATIONS
221
13.2
The spectral equations
It is straight-forward to proceed from equation 13.1 to a set of spectral equations. It is worthwhile though to back up a step and arrive at the same point another way via the two-point correlation equations. The procedure for deriving these equations are almost the same as those used to derive the single-point Reynolds stress equations, with only one dierence: the instantaneous equations at one point are multiplied by the velocity component at another. Thus the equation for the instantaneous uctuating velocity at , say ui = x ui ( , t), is multiplied by the uctuating velocity at x , say ui ui (x , t); i.e., x
[
ui And vice versa, ui

[
ui + = t ui + = t
(13.2)
]
(13.3)
These are averaged and added to obtain an equation for the two-point velocity correlation as: < ui uj > < ui uk uj > < ui uj uk > + + (13.4) t xk xk ( ) ( 2 ) < ui uj > 2 < ui uj > 1 < puj > < p ui > = + + + xi xj xk xk xk xk Note that we used a trick to pull the derivatives outside the product: ui is not a function of x , nor is ui a function of . Also we have assumed no mean velocity. x Now lets consider the implications of homogeneity. We know that this means that the two-point moments can only depend on = x . For convenience lets r x = x + , and then change variables from x , to , . r x x dene another variable, The chain-rule immediately implies that: j rj = + xi j xi rj xi = + i ri Similarly, it is easy to show that: = xi i ri (13.6)
(13.5)
Now this may seem like so much hocus-pocus until you realize that we are only dierentiating functions of only, so all the derivatives involving are identically r zero.
222
CHAPTER 13. DYNAMICS OF HOMOGENEOUS TURBULENCE Thus homogeneity reduces our two-point Reynolds stress equation to: Bi,j (, t) r = [Bik,j (, t) Bi,jk (, t)] r r t rk [ ] 1 Bp,j (, t) Bi,p (, t) r r 2 Bi,j (, t) r + + 2 ri rj rk rk (13.7) (13.8)
Dont lose heart yet, we are almost there. Now we could Fourier transform this directly, but since we are only going to work with the trace, lets contract the indices rst, then Fourier transform. As usual, setting i = j causes the pressure term to drop out, and we are left with: Bi,i (, t) r 2 Bi,i (, t) r [Bik,i (, t) Bi,ik (, t)] + 2 r r = t rk rk rk (13.9)
We can immediately take the three-dimensional Fourier transform to obtain: Fi,i ( t) k, = Gi,i ( t) 2k 2 Fi,i ( t) k, k, t (13.10)
where Gi,i contains the non-linear interactions and is related to the transform of the triple moment terms by: Gi,i ( t) = ikl [Fil,i ( t) Fil,i ( t)] k, k, k, (13.11)
What is arguably the most famous equation in turbulence theory results immediately by integrating over spherical shells of radius k and invoking the denition of the three-dimensional energy spectrum function, E(k, t). The result is: E(k, t) = T (k, t) 2k 2 E(k, t) t (13.12)
where T (k, t) is non-linear spectral energy transfer dened as the integral over spherical shells of radius k of Gi,i ( t)/2. Note that like all other averaged equak, tions in turbulence, this equation is not closed since there are two unknowns but only one equation. Nonetheless, there will be much we can learn from it. It is easy to show by integration that the left-hand side integrates to the lefthand side of equation 5.2, while the last term on the right-hand side integrates to the dissipation; i.e.,
1 2 <q > = E(k, t)dk 2 0 = 2 k 2 E(k, t)dk
(13.13) (13.14)
It follows that the integral of the spectral transfer term must be identically zero; i.e., T (k, t)dk = 0 (13.15)
0
13.2. THE SPECTRAL EQUATIONS
223
0.03 time=1.326 time=1.700 time=2.219 time=3.687 time=4.050 time=4.711 time=6.000 time=7.340
0.025
0.02
0.015
0.01
0.005
10
15 k
20
25
30
Figure 13.1: DNS data of Wray [?] for decaying isotropic turbulence showing how E(k, t) decays with time. Note limited resolution at low wavenumbers.
224

6 4 2 0
T/ u 2
-2 -4 -6 -8 -10 -12 -14 0 2 4 6 8 10 12 14 16
TIME=1.710
Wray 98 data T from eq. 6 using E (G92) T from eq. 3 using E (K41)
]]
Figure 13.2: DNS data of Wray [?] for decaying isotropic turbulence showing T (k, t) and 2E(k, t) at a single time. (This can also be proven directly from the denition of T .) Thus the net eect of the spectral energy transfer term is zero. It can only act to move energy from one scale to another. In other words, it is exactly the term we were looking for, since it alone can move energy from the scales where they are generated to those where it can be dissipated.
13.3
The eect of Reynolds number on the spectral equations
We are now in a position to examine how turbulence changes during decay. Figure ?? depicts linear-linear plots of typical energy, dissipation and energy transfer spectra. The energy spectrum, E(k, t), rises rather rapidly for small wavenumbers, probably as k m where 1 < m < 5/2 and is determined by the power law decay exponent of the energy. Most believe E(k, t) peaks near k L1 where L is the longitudinal integral scale. After this is rolls-o for a while as k 5/3+ where is very close to zero for large Reynolds numbers. Then nally it rolls o exponentially for wavenumbers above about one fth the inverse of the Kolmogorov microscale. 1/. Since the kinetic energy of the turbulence is the integral under the energy spectrum, it is easy to see that most of the energy comes from the spectral region near the peak. Nonetheless, the contribution from the slow roll-o can not be neglected and must be considered in designing experiments (or simulations) to avoid underestimating the energy.
13.4. CONFLICTING VIEWS OF SPECTRAL TRANSFER
225
The dissipation spectrum is, of course, very closely related to the energy spectrum since it is the same spectrum multiplied by 2k 2 . The eect of the pre-factor is weight the dissipation to the highest wavenumbers so most of the dissipation occurs at wavenumbers higher than the peak which is near k 0.2/K . This means that experiments or simulations must carefully resolve wavenumbers to at least K (or higher) if the dissipation is to be accurately determined. Resolution requirements are much more severe for higher spectral moments. The spectral transfer term does exactly what we thought it should do. The transfer term is negative for low wavenumbers and nearly equal to E/t. Thus on the average, it takes energy from the energy-containing scales where there is very little dissipation and moves it somewhere. But where? The answer lies in the high wavenumbers, where the spectral transfer is positive and nearly equal to the dissipation.
13.4
Conicting views of spectral transfer
Since Kolmogorov it has been believed that it is the ratio of the integral scale, L, to the Kolmogorov microscale, K , which determines the shape of the the spectra, and the separation between the energy and dissipation ranges. But there are recent indications that this may not always be true, if at all, as discussed below. For now, lets assume it is true. Consider what happens as R L/K becomes very large. The Kolmogorovian view of the energy, dissipation and transfer spectra for three dierent values of R are illustrated in Figure ??. For R 100 (which is typical of many simulations and laboratory experiments), the energy and dissipative ranges overlap, and there is really no region at all where the spectral transfer is near zero. By contrast, for R > 104 , the dissipative ranges and energy-containing ranges are widely separated. At low wavenumbers, the spectral decay is almost entirely due to the spectral transfer, and this region is virtually independent of viscosity. At high wavenumbers, there is virtually no energy, and the balance is almost entirely between spectral transfer and dissipation. And in-between the energy and dissipative ranges, there is an extensive spectral region where there is virtually no energy, no dissipation, and the spectral transfer is zero. This is the so-called inertial subrange of the spectrum which only emerges at very high wavenumbers, and will seen below to be the region characterized by the near k 5/3 behavior noted above. So where it the problem with this picture. The problem is that this picture has never been conrmed in any ow for which dE/dt was not identically zero due to stationarity. Figure ?? is taken from a paper by Helland et al. in 1977 and shows data taken in several wind tunnels, one the Colorado State University wind tunnel which is among the largest in the world. Figure ?? shows similar data in DNS of decaying turbulence taken from George and Wang (2002). These DNS data are at values of R between 30 and 60, which is about the same as most of the wind tunnel data. The CSU data are at R 240. Clearly neither the DNS
226
nor the experiments in decaying turbulence show the expected region in which T (k, t) 0. By contrast, the forced DNS of Kaneda et al. (2002) does show an extensive range for which T 0. So we are left with the following dilemma: does the Kolmogorov picture appear to be correct for the forced turbulence only because of the absence of alternatives. In particular, forced turbulence is stationary, so above the wavenumber of the forcing, no other value of T is possible if the dissipative wavenumbers are large enough. Or would the results for decaying turbulence give the same results if the Reynolds number were high enough? If not, could it be that the whole idea of Kolmogorov is wrong and there is another idea which better describes the world as we nd it? The next chapter explores such a possibility.
Chapter 14 Kolmogorov Turbulence

There are several dierent approaches to the spectral energy equation, each representing very dierent kinds of turbulence. This chapter presents and explores the classical theory, mostly due to Kolmogorov (1941). It has been developed and rened by many. It was rst presented in the form presented here by George K. Batchelor and his co-workers, and most eloquently in his now classic book Homogeneous turbulence (Batchelor 1953). Most of the things we thought we knew about turbulence in the 20th century (including all turbulence models) are based at some level on these ideas. So even if they turn out not to be completely correct, it is important to learn about them if for no other reason than to understand how and why the engineering models have developed the way they did. Ive already given some hints in the previous chapters that I dont believe this so-called Kolmogorov view of turbulence to be the whole story. In fact, at this writing I have come to believe that there probably are three dierent kinds of turbulence, only one of which behaves like Kolmogorov. This is not a question of Kolmogorovs ideas being wrong, but instead my belief that the underlying assumptions do not apply to all ows. Therefore in preparation for the new ideas in subsequent chapters, I shall try to make it quite clear what what the underlying assumptions are.
14.1
The universal equibrium range
We shall start our discussion with the equation 13.12 for decaying homogenous turbulence, since it has all the features we need; i.e., E(k, t) = T (k, t) 2k 2 E(k, t) t (14.1)
where E(k, t) is the three-dimensional energy spectrum function, and T (k, t) is the non-linear spectral energy transfer, also dened as the integral over spherical shells of of radius k the non-linear spectral transfer. It has already been noted that this equation is not closed since there are two unknowns but only one equation. 227
228
CHAPTER 14. KOLMOGOROV TURBULENCE
Now the rst thing we need to remind ourselves is what the basic terms in this equation are. The left-hand-side is just the rate at which energy changes (decays in this case) at any given wavenumber, so its integral over all wavenumbers is the rate of change of kinetic energy with time. The last term on the right-hand-side is the rate at which kinetic energy is converted to internal energy by viscosity and the deformation of uid elements; i.e., the rate-of-dissipation, . Clearly this must be must equal the integral of the left-hand-side (since there is no production or transport in our simplied homogenous problem): i.e.,
d k 2 E(k, t)dk = E(k, t)dk = 2 dt 0 0
(14.2)
All of this, of course means that the integral of the non-linear transfer, T (k, t) over all wavenumbers is exactly zero; i.e.,
T (k, t)dk = 0.
(14.3)
In words, there is no net transfer of energy due to the non-linear (or triadic) interactions; whatever is taken out at one wavenumber must be put back somewhere else (in wavenumber space). The basis of Kolmogorovs ideas (or hypotheses) was that there might be situations in which the T (k, t) takes energy out at low wavenumbers and puts it in at high wavenumbers. In fact this seems to be correct, even for turbulent shear ows since the energy is usually put into the ow at the large scales (the energetic scales corresponding roughly to the peak in the energy spectrum). This is because in ows with mean shear the turbulence energy is produced by the working of the Reynolds shear stresses agains the mean ow gradient. And the Reynolds stress spectrum peaks about where the energy spectrum does, but it falls o with increasing wavenumber much more rapidly. So the bottom line is that even if we dont have a simple homogenous decaying turbulence, the turbulence at suciently high-wavenumbers behaves as though we do at least we hypothesize that it might do so. In fact many ows actually do behave this way, even at modest Reynolds numbers. So here is our picture: energetic scales of motion corresponding wavenumbers of the same order of magnitude as the energy spectrum peak, say kpeak and dissipation mostly at much higher wavenumbers. Now we take this picture one step further. If most of the energy is at the smaller wavenumbers (the energetic scales), then these energetic scales will dominate the non-linear transfer (the triadic-interactions there). Moveover, since the uctuating velocity gradients are relatively small at these low wavenumbers, compared to those at much higher wavenumbers, then there is really very little dissipation down here at these wavenumbers where the energy mostly is. This can be easily seen by looking at the last term in equation 14.1 where the multiplication of E(k, t) by k 2 skews its integral to much higher wavenumbers, and makes it nearly zero for low wavenumbers. Now we are ready for Kolmogorovs fundamental hypothesis, and on which
14.1. THE UNIVERSAL EQUIBRIUM RANGE
229
all the other consequences of this theory are based: The Universal Equilbrium Range Hypothesis. In brief the hypothesis is that for wavenumbers suciently higher than the those characterizing the energy-containing range (say k >> kpeak ), E(k, t)/t 0. And most importantly, it is negligible relative to the non-linear transfer and dissipation. The basic argument supporting this hypothesis is that the times scales of the smallest scales or eddies (or highest wavenumbers) are so much smaller than the time scales of the energy-containing eddies (or energetic wavenumbers) that the smallest scales are eectively in statistical equilibrium. We can see why the equilibrium range hypothesis might be true by the following arguments. (Note that we will need to use information that we later derive from the consequences of Kolmogorovs hypothesis, so if you think we are assuming things we have not proven yet, we are. But hang in there for now just dont forget your questions.) We know from experience that most of the dissipation occurs for wavenumbers below k Kol = 1 (Typically about 99%.) And we have guessed earlier that we can dene a dissipative time scale from just the rate of dissipation of turbulence energy, , and the kinematic viscosity, , and it is given on dimensional grounds alone by Kol = ()1/4 . So lets compare Kol to a time-scale characteristic of the energy-containing eddies, say l l/u kpeak u/2 where 3u2 /2 is the kinetic energy. Obviously we need an estimate for , so we will use another consequence of the theory we are deriving; namely we dene l = u3 / and hope that l is proportional to the size of the energetic scales (roughly the integral scale) or l 2/kpeak . Now we can substitute for u3 /l to obtain: Kol ul = l
[ ]1/4
= Rl
1/4
(14.4)
where we have dened a turbulence Reynolds number as Rl = ul/. Clearly the bigger the turbulence Reynolds number the smaller the ratio of time scales and the more the dissipative scales seem to be in equilibrium, at least relative to the energetic scales. But what about the wavenumbers that are much larger than kpeak , but much smaller that k = 1/Kol ? How do we know whether these are in near statistical equilibrium? We can estimate a time-scale for an arbitrary wavenumber by simply using the energy spectrum itself and the local wavenumber. On dimensional grounds alone we dene the local time scale to be: k = [k 3 E(k, t)]1/2 (14.5)
Now we need to use another result of the theory we are about to derive, namely the form of the spectrum in the so-called inertial subrange, or more commonly, the k 5/3 range. We will derive this later, but for now note that in it, E(k, t) = Kol 2/3 k 5/3 , where Kol is the so-called Kolmogorov constant which is usually assumed to be between 1.5 and 1.7. (The real miracle here is that the dissipation, , should make its way into an expression about the behaviour of the spectrum in a range in which there is assumed to be no dissipation, but save your questions about
230
this till later there is is reason.) Substituing this result into our local time scale 1/2 leads to the time scale of eddies in the inertial subrange as k = K k 2/3 1/3 . Using = u3 /l yields immediately the ratio of the time scale at given value of k in the inertial subrange to that of the energetic scales as: 1 k = (kl)2/3 l (14.6)
This can be much less than unity only if kl >> 1, to which we are of course willing to restrict ourselves. So the fundamental hypothesis of the equilibrium subrange can at most be true for wavenumbers large compared to those characterizing the energy-containing range (kl >> 1) , and for high turbulence Reynolds numbers (Rl >> 1 or l/Kol >>>> 1). It has been generally assumed, at least since the 1960s that this hypothesis of Kolmogorovs in fact describes all turbulence for scales much smaller than those at which the energy is put into the ow, no matter how generated. Moreover, in spite of our inability to pin down exactly the value of Kol or even to measure the dissipation, this whole part of the spectrum has been widely assumed to be universal, hence the phase Universal Equilibrium Range. And there have been a number of experiments and simulations (using forced DNS) which purport to support the idea. Unfortunately it does not seem to have been noticed that there is one other very important circumstance under which Kolmogorovs hypothesis is satised exactly; namely ows which are statistically stationary so the left-hand-side of equation 14.1 is identically zero. Clearly you cant prove Kolmogorovs theory by using data from statistically stationary ows. And in fact it has been the non-stationary ows which have been the problem, decaying turbulence in particular. Nontheless, before examining them (in the next chapter), lets try to consider the full consequences of Kolmogorovs ideas.
14.2
The inertial subrange
Now lets assume Kolmogorovs ideas to be correct and that they indeed represent a spectral range which can be considered to be in statistical equilibrium. This of course presumes the Reynolds number to be suciently high, or more particularly l/Kol >> 1. If so then we can split the energy equation at some arbitrary wavenumber, say km , where km is some intermediate wavenumber simultaneously satisfying k >> kpeak and simultaneously km >> 1/Kol . Then if kpeak <<<< 1/Kol , we can safely assume there to be very little dissipation below k < km so that the energy equation reduces simply to: E T, t k < km (14.7)
For convenience we can dene a spectral ux, say k (k, t), so that:
14.2. THE INERTIAL SUBRANGE
231
k (k, t) =
T (k , t)dk
(14.8)
where k is just a dummy integration variable (which we need since we have used k as the upper limit of integration). Note that many confuse this spectral ux, k with the rate of dissipation of kinetic energy, . It is not! In fact it varies with wavenumber, k, and represents the net eect of the non-linear terms in removing energy from wavenumbers smaller than k. So why you ask must we confuse things by calling it k . The answer will come soon, and represents one of the most beautiful results in turbulence theory and also one of the most poorly understood. Now since T (k, t) is the energy added (or removed if the sign is negative) at each wavenumber, it is easy to see that the spectral ux is the net energy crossing each wavenumber (from low to high since we have included the negative sign in the denition). Dierentiation yields immediately the non-linear spectral transfer as the negative of the gradient of the spectral ux; i.e., dk (k, t) dk We can integrate equation 14.7 from 0 k km to obtain: T (k, t) = (14.9)
km d km E(k, t)dk = T (k, t)dk = k (km , t) (14.10) dt 0 0 But since we have assume almost all of the dissipation to occur at much higher wavenumbers than km and almost all of the energy to be at wavenumbers below km , it follows immediately that: km d km dk (k, t) d 1 E(k, t)dk ui ui = dk = k (km , t) dt 0 dt 2 dk 0 [ ]
(14.11)
since by denition k (0, t) = 0. In words, the spectral ux crossing our intermediate wavenumber, km , is almost equal to the time derivative of the kinetic energy. Now lets examine what is happening at the higher wavenumbers. For k > km the turbulence is assumed to be in in near statistical equilibrium so we can throw away the time derivative on the left-hand-side and write approximately: 0 T 2k 2 E, Integration from km to innity yields: 0
km
k > km
km
(14.12)
T (k, t)dk 2
E(k, t)dk
(14.13)
But we already have observed that the integral of the non-linear spectral transfer, T (k, t) over all wavenumbers is zero. Therefore
232
km
T (k, t)dk =
km
T (k, t)dk = k (km , t)
(14.14)
Also since by hypothesis we have assumed the Reynolds number to be so high that there is relatively little dissipation for wavenumbers below km , it follows immediately that the last integral equation 14.13 is just approximately the dissipation; i.e.,
km
E(k, t)dk
(14.15)
Thus the integral of our spectral energy equation for k > km is just: k (km , t) = (14.16)
Thus the net spectral ux at our intermediate wavenumber km is approximately equatl to the real dissipation of turbulence kinetic energy. This is of course a consequence of the fact that we have assumed no dissipation at the low wavenumbers just transfer out of the energy-containing eddies by the non-linear interactions to ever smaller scales. It is but a small extension of our arguments to realize that if our underlying assumptions are correct, then these arguments are exact at innite Reynolds number; i.e. when l/Kol . In this limit, there can be no dissipation at the energy-containing scales, and no energy at the dissipative scales. If the latter seems counter-intuitive, then dont think of the limit, just the limiting process where less and less energy is available at the higher wavenumbers due to the vicous dissipation, and less and less dissipation at the smaller wavenumbers. And of course it is the non-linear terms (represented here by the spectral ux) that move energy from the low wavenumbers to the high ones. Exactly how they do this is not completely understood, but clearly it is through the non-linear triadic interactions.
14.3
Scaling the energy and dissipation ranges
From the consderations above we can infer two dierent scaling laws, one for high wavenumbers (or the dissipative scales of motion) and one for low wavenumbers (or the energetic scales of motion). First for the so-called energy-contaning range of scales we can see that the only important parameters (at least in the limit of innite l/Kol are u2 and l = u3 /. (Recall that the latter is really l = u3 /k (km ), but since we are at assuming very high values of l/Kol we use instead.) Thus on dimensional grounds alone we would expect spectra to collapse at low wavenumbers when plotted as: E(k) E(k) = 2 ul (14.17)
14.3. SCALING THE ENERGY AND DISSIPATION RANGES
233
Figure 14.1: Plots of streamwise and lateral one-dimensional spectra measured downstream of grid, plotted in energy-containing range variables (data of Mydlarski and Warhaft 1996, from Gamard and George 2000).
234
where we have dropped for now the dependence on t. E(k) is a non-dimensional spectrum and k = kl is a non-dimensional wavenumber. This is the so-called energy-range form of the spectrum and u2 and l are referred to as energycontaining range variables. Exactly the same scaling can be applied to the one-dimensional spectra as well. Figure 14.1 shows spectra for dierent wavenumbers plotted in this manner. The (1) particular plots shown actually use the measured spatial integral scales, L1,1 and (1) L2,2 instead of the pseudo-integral scale l, so the collapse is actually a bit better than it would have been with l.1 (More will be said about this later.) The collapse at low wavenumbers is quite impressive, and departure is gradual with increasing wavenumber until the near exponential roll-o with increasing Reynolds number as the ratio of l/Kol increases. The collapse is, of course, not perfect since we are not really at innite Reynolds number (i.e., l/Kol ) and there are still small (but nite) viscous eects felt at all wavenumbers. Now consider what happens for high wavenumbers. For this region of the spectrum dissipation dominates, and the entire equilibrium range is pretty much described by only and . We can use these to write another non-dimsionalized form of the spectrum as: E + (k + ) = 1 5/4 1/4 E(k) (14.18)
Now E + (k + ) is a non-dimensional spectrum of the dimensionless wavenumber, k + = k Kol . This is the so-called Kolmogorov non-dimensional spectrum and the spectrum is said to be normalized in Kolmogorov variables. Figure 14.2 shows the same one-dimensional spectral data as shown in Figure 14.1, but this time plotted in Kolmogorov variables. Note the near collapse at high wavenumbers, but the clear variation with l/Kol at the low wavenumbers. Also clear from both gures is that there seems to be range of wavenumbers between both extremes (say 1/l << k << 1/Kol ) where both the high and low wavenumber scalings work pretty well. This, as we shall see below, is the inertial subrange where the famous Kolmogorov k 5/3 law wil be seen to live.
14.4
The k 5/3 and r4/3 laws
One of the main reasons for the acceptance2 of Kolmogorovs ideas about the local equilibrium (and even universality) of the small scale (or high wavenumber)
The original Mylarski/Warhaft paper uses l = u3 / instead. This didnt happen immediately, and it really wasnt until the tidal channel measurements of Grant et al. reported in a famous 1963 Marseille turbulence meeting that people began to generally accept them. Batchelor and others, of course, believed them much earlier, probably because of their theoretical elegance. But even Kolmogorov himself was beginning to doubt and in the same meeting produced another version that included intermittency. This latter has come to be known as Kolmogorov 1963 or just K63, while his earlier theory is usually just referred to as K41. It is K41 that is the focus of this chapter.
2 1
14.4. THE K 5/3 AND R4/3 LAWS
235
Figure 14.2: Plots of streamwise and lateral one-dimensional spectra measured downstream of a grid, plotted in Kolmogorov variables (data of Mydlarski and Warhaft 1996, from Gamard and George 2000).
236
turbulence was the success of the the inertial subrange prediction. As we shall see below, a slight extension of the arguments above leads immediately to prediction that the spectrum in the inertial subrange should vary as E(k) = Kol 2/3 k 5/3 . Similar considerations for the velocity structure functions yields [ui ( + ) x r (n) n/3 n/3 n ui ( ] = i x r . If it is assumed that the turbulence in this equilibrium range is also universal (meaning it is the same in all ows and the Reynolds number is high enough), then the coecients are also presumed to be universal constants. The third-order structure function is the most interesting of all, since it can be shown directly from the dynamic equation for the second-order structure function that the coecient is exactly 4/5; i.e [u(x + r) u(x)]3 = (4/5) r. Now the truth is that in spite of the enthusiasm for these ideas (actually its more like religious fervor3 ), they seem to work really well only some of the time, and not so well other times. One reason for this not-so-great behavior is completely obvious. The Reynolds number (i.e., the ratio of integral scale to Kolmogorov microscale) of many ows, especially those we generate in our laboratories and computers, is not nearly at high enough for the theory to even begin to apply. Thus there really is no inertial subrange in which k , nor sometimes is the Reynolds number even high enough for there to be an equilibrium range at all. Even so, people still draw k 5/3 lines on their measured spectra, and seem to take comfort in the fact there is always a point of tangency. But there are other reasons for the failure of the theory as well. First there seem to be ows which march to a dierent drummer entirely, especially non-equilibrium ows. We shall discuss some of these in the next chapter. And also there are ows in which energy is being put into or taken out of the inertial subrange directly by the mean ow or from other sources like buoyancey. Thus the the spectral ux is not constant but increases or decreases with wavenumber. But sometimes even when this happens we can make a dierent spectral model, still in part based on Kolmogorovs ideas. We shall include an example later. And there seem to also be eects of internal intermittency; i.e., eects arising from the fact that the dissipation is not uniformly distributed throughout the ow but is sometimes highly concentrated into compact regions (like spaghetti or lasagna). Actually it was Kolmogorov himself in 1963 who rst built a model for this based on the assumption that the dissipation was log-normally distributed. There has been much work on this over the past four decades; and although it is important, we shall not discuss it here since it will lead us too far from our story. The bottom line though is that internal intermittency will cause a slight Reynolds number dependence in the inertial subrange exponents (e.g., from 5/3 to 5/3 + (Re) where (Re) > 0. Note that this is the opposite of the low Reynolds number eects considered later.
3
I just saw the results of recent study that suggests the more poorly people actually understand something and the less condent they are that the data actually support it, the more intensely they will defend it. This happens in turbulence too.
237
14.4.1
Dimensional and physical analysis
In section 14.2 we argued for the existence of a spectral range, the inertial subrange, in which for any wavenumber, km , the spectral ux, i (km ) was constant and approximately equal to the actual rate of dissipation, . Moreover we argued that in the limit of innte Reynolds number that k (km ) = exactly. This was an immediate consequence of our assumption that there was no dissipation at the lower wavenumbers or in our inertial range, and therefore any energy dissipated had to be passed through to the dissipative range. Now look at the consequence for the spectrum E(k) in this inertial subrange. The only parameter surviving in the equations is k = . There is eectively no energy here, so u2 and l are not important here; nor are there any viscous eects so we can forget about also. This means that E(k) must be entirely determined by k and alone, the latter ONLY because it is equal to the spectral ux in the inertial range, k . To emphasize this point we shall write k below in our derived result, and set it equal to only at the very end. (This may seem rather silly to you now, but it will save you much agonizing later when you try to gure out whether these theoretical results should apply to the ows you encounter which dont seem to behave this way like in turbomachinery or in geophysical uid dynamics.) Now we apply the Buckingham- theorem which in its simplest form says the number of dimensionless ratios possible is equal to the number of parameters minus the number of dimensions. We have three parameters (E(k), k and k ) and two dimensions (length and time). Therefore we know immediately that E a b k c = const and must be dimensionless. Since the dimensions of E are k length3 /time2 , are length2 /time3 and k is 1/length, clearly the constant can be dimensionless only if E(k) 2/3 k 5/3 . Moreover if the spectrum itself is assumed to be universal, then so is the coecient, say Kol . Exercise Prove using dimensional analysis that if there is a range of scales in physical space for which and r are the only parameters, then the nth -order (n) structure function must be given by [ui ( + ) ui ( ]n = i rn/3 . x r x The results for the structure function and spectrum are generally assumed to be equivalent, but I must confess I have a certain uneasiness about the latter. My reason is that in spectral space we know that the spectral coecients in nonoverlapping bands are uncorrelated. So when we make arguments about spectral ux we can pretty much be sure that we are talking about something that corresponds to the mathematical decomposition of the equations, with energy being transferred from the eigenfuntion at one wavenumber to that at another. It is not so obviously true (at least to me) in structure function space, since we for sure can not say that turbulence at one physical scale is uncorrelated with that at another. But maybe Im just being a bit pedantic here, since Im not aware that anyone else seems too worried about this. It does make a dierence, however, when one
238
thinks about closure in LES (Large Eddy Simulations). In particular, can one apply Kolmogorovs ideas to small volumes of uid below some cuto size with the same degree of condence one could make a truncation in wavenumber space above some wavenumber? My personal opinion is: probably not.
14.4.2
A useful empirical low wavenumber spectral model
There really are no known analytical solutions to the spectral energy equation, so when one is needed we have to resort to empirical ones. For the low wavenumber spectral models, these are only a little more than just ts to data, but they should try to get the physics more or less right. For example, one popular choice for an spectrum in the energy-containing range (originally proposed by von-Karman and Howard 1938) is given by: E(k) = [u2 l] C4 k 4 [1 + (k/ke )2 ]17/6 (14.19)
For small k this will start o as k 4 , which is consistent with the controversial Loitsianskii invariant. And for k >> ke it rolls o as k 5/3 , which is consistent with the upper limit of the energy scaled spectrum in the limit of innite Reynolds + number, E (k + ). The coecient C4 and parameter, ke are usually chosen so the the spectrum integrates to the all the kinetic energy, 3u2 /2 and gives the right value of Kol in the limit of large k; i.e., 3 2 u = E (k, t)dk 2 0 (14.20)
17/3 and u2 lC4 ke = Kol . Obviously you will need to use the innite Reynolds number relationship = u3 /l.
Exercise: Find the value of C4 and ke l which satises these constraints. (Hint: non-dimensionalize the wavenumber by ke , then you will discover the integral to be one of the standard tabulated integral (a gamma-function actually). Then solve for the values of ke l and C4 that give you the right kinetic energy, 3u2 /2, and Kol . Try two values of Kol : 1.5 and 1.7.) Substitute equation 14.19 into the isotropic spectral relations and show that the corresponding one-dimensional spectrum is given by: F1,1 (k1 ) =
(1)
C [1 + (k1 /ke )2 ]5/6
(14.21)
What are the values of C and ke l for the two cases above? (Remember the onedimensional spectrum should only integrate to u2 , and from the isotropic relation between the three-dimensional spectrum function and the one-dimensional spectrum, the one-dimensional Kolmogorov constant is 9Kol /55.)
14.4. THE K 5/3 AND R4/3 LAWS p kpeak L ke L kpeak /ke Cp 1 1.3630 1.7597 0.7746 0.3229 1.34 1.2943 1.4437 0.8965 0.4535 2 1.2272 1.1203 1.0954 0.8455 3 1.1804 0.8798 1.3416 2.2249 4 1.1570 0.7468 1.5492 6.2528
239
Table 14.1: Spectral parameters as function of p It is possible to generalize the result above into more variations. One used by Wang and George, JFM 2003 (see their appendix) writes: E(k) = [u2 L] Cp (kL)p [1 + (kL/ke L)2 ]5/6+p/2 ] (14.22)
Here L is the physical (or real integral scale) and the values of Cp and ke L are chosen so that the energy and integral scale spectral integrals give the correct values. The same trick of non-dimensionalizing the wavenumber by ke allows both integrals to be expressed exactly as beta-functions. Values of kpeak L, ke L, kpeak /ke and Cp are tabulated in Table 1 for p = 1, 1.34, 2, 3 and 4. Note that this method of choosing the coecients (i.e., insisting the integral scale be correct as well as the energy) does not allow an independent determination of the von Karman constant. Finally, Gamard and George (2000 J. Flow, Turbulence and Combustion, see the appendix) used a slightly modied version with near asympototics which actually allows for nite Reynolds number departures from the innite Reynolds number solutions; e.g., E(k, L/Kol ) = [u2 L] Cp (kL)p [1 + (kL/ke L)2 ]5/6+p/2/2 (14.23)
where is a function of 1/ ln L/Kol . The whole point is that even an empirical expression like this can be of enormous value in trying to understand some of the ner points of turbulence and turbulence measurement. For example, Reynolds (Ann. Rev. Fluid Mech. 1976) uses Equation 14.22 to examine how turbulence modeling constants depend on the energy spectrum assumed. George et al. 1984 use a slightly dierent version to obtain an excellent t to the one-dimensional velocity and pressure spectra in a jet mixing layer. Wnstrm et al. (2007 ASME paper and in her Ph.D. dissertaa o tion) even used equation 14.19 to evaluate how the nite spatial dimensions of a PIV interrogation volume aected the mean square value of dierent components of the velocity. This worked in her case because the spatial ltering was cutting o in the inertial subrange, so the viscous range was not important. The bottom line is: dont be too analytically lazy to be creative. It is amazing how much you can learn from a simple empirical model thoughtfully applied.
240
14.4.3
ll in
Models for the spectrum in the univeral equilibrium range
14.4.4
Deduction of k 5/3 -range from asymptotic analysis
There is another way to deduce the existence of the k 5/3 -range by working with the scaling laws for low and high wavenumber spectra (equations 14.17 and 14.18). It has the advantage that it really is useful in highlighting that our deductions are really valid only at very high Reynolds number, in fact innite to be precise. This was to the best of my knowledge presented by Tennekes and Lumley in their pioneering book An First Course in Turbulence, and which even though outdated is still better than most newer books on turbulence. What they did is to recognize that both the inner and outer scaled forms of the spectrum are valid if you keep the Reynolds number dependence, in their case l/Kol where they used l = u3 / for their low wavenumber scaling. Thus both the following expressions for E(k) are completely valid at all wavenumbers: E(k) = u2 lE + (k + , R) E(k) = 5/4 1/4 E(k, R) (14.24) (14.25)
where k + = k Kol and k = k l and R = l/Kol . The dependence on R inside the functional relationship is what makes the energy variables scaled spectrum start deviating at high wavenumbers, and vice versa for the high wavenumber scaled spectrum. For example, in Figures 14.1 and 14.2, R (but dened using L instead of l) is the parameter that labels the indidual curves. Note how they depart from collapse as R changes. Now if the spectra have to be equal, so must their derivatives with respect to wavenumber; i.e. dE dE = u2 l dk dk dE = 5/4 1/4 dk dk dk dE + dk + dk + dk
(14.26) (14.27)
But dk/dk = l and dk + /dk = Kol . Hence the derivatives can be equal only if: dE dE + (14.28) = 5/4 1/4 Kol dk + dk We can eliminate u2 by substituting using = u3 /l, which must be true, at least at innite Reynolds number. Doing this, using the denition of Kol = ( 3 /)1/4 , and multiplication of both sides by k 8/3 yields after some rearrangement: u2 l2
241
+ + dE(k, R 8/3 dE (k , R) k 8/3 = k+ (14.29) dk + dk Now comes the trick part, which is the same trick used to derive the log or power law solutions to the boundary layer. We argue that in the limit as R = l/Kol , the ratio of k/k + becomes undened. Therefore in this limit, both sides of equation 14.29 must equal the same constant, which for convenience we dene to be (3/5) Kol ; i.e.,
3 dE(k, R = Kol lim k 8/3 R 5 dk dE + (k + , R) 3 8/3 lim k + = Kol + R dk 5
(14.30) (14.31)
But since the right-hand side of both equations is a constant we can integrate immediately to obtain: E (k) = Kol k 5/3
+ E (k + ) = Kol k + 5/3
(14.32) (14.33)
Thus we have recovered the inertial subrange we obtained by dimensional analysis. But we now recognize that it is not just in-between the energy-containing range and the dissipative range, it is an extension of both of them. It is truly a matched layer that comes from stretching and matching two solutions in the limit of innite Reynolds number. Clearly, even if all theoretical assumptions are correct, we should expect to nd Kol strictly constant and universal in the real world only at very, very high Reynolds numbers. Also we should expect spectra at low wavenumbers to collapse with u2 and l to only collapse the low wavenumber range perfectly at very, very high Reynolds numbers. And the same for the Kolmogorov scaled spectra at high wavenumbers. In the next section we shall examine what happens if we relax this requirement on the Reynolds number a bit.But before we do, lets see what happens if we try putting the high wavenumber and low wavenumber spectra together to form a composite spectrum. The basic idea is to write one of them in the variables of the other, multiply them together and divide by the common part (a trick I learned from Chuck Van Atta, a highly respected turbulence experimentatlist at UC San Diego, now deceased). For example, we can write the high wavenumber spectrum using k + = k/R. The common part to both high and low wavenumber spectra (in k variables would be just be equation 14.32 above. Thus our composite spectrum in energy variables would just be:
+ Ecomposite (k, R) = E (k) E (k/R) [Kol k 5/3 ]1
(14.34)
300
Note the substitution of k/R for k + inside the function E + . As the value of k increases, the increasing viscous eects included in E + will cut o the composite spectrum at very high wavenumbers, exactly like happens in nature. And the higher the value of R, the higher the value of k before this happens.
14.4.5
The inertial range at nite Reynolds numbers
ll in from Gamard and George 2000
Appendix A Signal Processing

Foreword: This appendix was largely taken from the notes from courses on turbulence measurement at the Danish Technical University and Chalmers developed by Knud Erik Meyer, Peter B.V. Johansson and William K. George. Most things that we measure are what we called random processes. In measurements of uid ows, turbulence is one example of very important random component. To get useful results from measurements from a random process we need to use statistical tools since we do not have much use for a collection of instantaneous values alone. This is not always simple to do in practice, but the underlying ideas are quite straightforward. This chapter is about the process of getting from a physical quantity to a numerical estimate of the quantity. For more reading on data from random processes, we recommend the very thorough book by Bendat and Piersol [?]. Several special details related to ow measurements are covered in [?].
A.1
Signals
We dene a signal as the variation in time and space of any relevant physical quantity occurring in a useful system or device. The signal itself is formed by a series of measurements of the quantity. We can divide signals from random processes into three dierent types: Stationary: a signal whose statistical properties do not vary with time. This means that it does not matter when you choose to do your measurement. This is a fundamental assumption in most experiments and also an assumption that we will do throughout this chapter. Periodic: a signal that repeats itself after a nite amount of time. Common examples occur in rotating machinery. Sometimes such a signal can be treated as though it were locally stationary statistically by dividing each period into small time-slots. Data from the same time-slot from many periods 301
302
35
APPENDIX A. SIGNAL PROCESSING
30
25
u [m/s]
20
15
10
0 0
8 t [ms]
10
12
14
Figure A.1: Hot-wire measurement of the main ow velocity component in a turbulent ow behind a cylinder sampled at 100 kHz. Data from [?]. can then be treated as a stationary signal. Usually though, when doing statistical analysis, it is better to treat such signals by using only information from exactly the same phase of the signal (called phase-averaging). Transient: a signal that only exists for a nite amount of time, usually short. An example could be a sudden close of a valve in a ow system. If the process creating the signal is repeatable, the signal can sometimes be treated as a periodic signal. There can of course be signals that do not fall within these three types; e.g., signals that are not repeatable. For such a signal, many of the statistical tools discussed in the present chapter are not applicable and it is only possible to average between dierent realizations of the same experiment; i.e., the experiment must be repeated over and over. An example of a signal is shown in gure A.1. Even though the signal does seem to have some repeating patterns, the signal has a strong random component
A.2. THE MEASUREMENT CHAIN

physical quantity transducer
303
numerical value
signal conditioning (filter, amplification)
analog to digital conversion
Figure A.2: The measurement chain due to ow turbulence. There are many questions we could try to answer using the data shown in gure A.1. A few of them are: What is the mean velocity? Is the data record long enough for an accurate estimate? Can we estimate the turbulence intensity? How accurate will our estimate be? Is there a dominant frequency in the signal due to vortex shedding from the cylinder? And if so, how can we estimate it? What would the optimal sampling frequency be for the questions above? These questions and several more can be answered using the statistical tools that we present in these appendices.
A.2
The measurement chain
Before we look at the statistical tools, we think it is instructive to understand the basic elements of a measurement. Any measurement can be described by the measurement chain shown in gure A.2. Here the measurement is divided into three steps. The rst step is a transducer (based on some physical principle) that converts the physical quantity into another physical quantity suitable for further signal processing. This is very often a voltage, but there are other possibilities as well; e.g., a displacement or an optical signal. Then there is nearly always some sort of signal conditioning. The converted signal might be amplied. It is also both unavoidable and desirable to have some ltering of the signal. Sometimes this is inherent to the transducer. Finally, before a measurement is usable, it should be converted into a numerical value. This is accomplished by an analog-to-digital (A/D) converter. For a simple example: consider the measurement of the temperature in room by a glass thermometer. The transducer is an enclosed amount of uid whose volume changes with temperature (e.g., ethanol). When the uid (mostly in the bulb) changes its volume to adjust to the temperature around it, this change is amplied by letting the change in volume cause a thin column of the uid to rise or fall in a very thin pipe. The temperature can then read be comparing
304

analog to digital conversion card
tranducer 1
amplifier
anti-aliasing filter anti-aliasing filter
sample & hold sample & hold
tranducer 2
amplifier
tranducer n
amplifier
anti-aliasing filter
sample & hold
M U L T I P L E X E R
analog to digital converter
computer
Figure A.3: Schematics of an analog-to-digital card the height of the column to a calibrated scale placed along the thin pipe. The conversion to a numerical value is done by you when you compare the column height to the scale. If the thermometer is moved into a cold air stream from a an open window, it will not instantly show the new temperature. You will need to wait at least a few minutes to read the new temperature. It is inherent to the construction of the thermometer that it lters out fast (or high frequency) variations of the temperature, and only follows slower variations. Thus in signal processing terminology, the thermometer is acting like a low-pass lter on the temperature uctuations, the net eect of which is to integrate out the faster time-dependent variations. The amplication (sensitivity of column movement) and the ltering properties have been xed by the design of the thermometer, and by your choice of the type of thermometer to use. You are also making some decisions about the conversion to numerical values. You can choose to report the temperature as 21 C, as 21.4 C or as 21.43 C depending on number of divisions on the scale and on your ability to interpolate between divisions. In the example with the glass thermometer of the preceding section, the conversion to a numerical value was the last step before the nal result. A few decades ago, a large part of the data processing was done by analog systems using specially built electrical circuits to evaluate quantities like mean values or frequency content. Today, the strategy is usually to convert measurements to a digital form as early as possible and perform data processing digitally. This is partly because digital processing has become very cheap, but in many cases, it is also because digital processing is much more exible. In the following section, we will therefore only consider digital data processing.
A.3. ANALOG-TO-DIGITAL CONVERSION
305
When a measurement variable is converted into digital form, it is discretized in three dierent ways: The value of the variable is discretized into a number of bit levels The variable is discretized in time by sampling The variable is discretized in space by selection of measurement points. The consequences of this discretization will be discussed in the subsequent sections. In any measurement, you will need to make some decisions regarding all three elements in the measurements chain. We will cover the properties of dierent transducers in the discussion of the dierent measurements techniques and we will cover the use of lters in section G.2. In the next section we will examine how an analog-to-digital converter works and what options are available.
A.3
Analog-to-digital conversion
If you are sampling data with a computer, most investigators today use an analogto-digital card (A/D card). A schematic of typical components is shown in gure A.3. An A/D card will typically have several data channels so that it can be connected to several transducers. Each channel has its own signal conditioning. The conversion to a digital number will be only for a xed voltage range; e.g., from 0.0 to 10.0V or from -5.0V to 5.0V. The signal from the transducer must therefore be rst amplied (or attenuated) to match this voltage range. Often an oset value should be added to place the mean value near the middle of the range, thereby minimizing the possibility of instantaneous values outside it. Depending on what will be done with the data, an important element can be an anti-aliasing lter. This is just a low-pass lter that removes high frequency parts of the signal before digitization. This lter is essential for spectral analysis of digital time-series data,and will be discussed in detail in section G.2. The basic problem it prevents is called aliasing, which means that frequencies of the digitized data show up at dierent frequencies than they should. Once data has been sampled incorrectly and is aliased, it is nearly impossible to de-alias or correct it; so proper choice (and use) of anti-aliasing lters can be quite crucial to the success or failure of the experiment. Three considerations are primary: First, do the lters roll-o (or cut-o) rapidly enough to really remove the parts of the signal that might otherwise be aliased? Second, are the lters for dierent channels phase-matched so their eect on all channels is the same? And nally, are the phase changes at dierent frequencies that are introduced by the lters acceptable. Since the more severe the cut-o, the more the phase changes with frequency, generally some trade-os must be made. A popular compromise choice is the Bessel lter whose phase variation with frequency is linear, which corresponds to a simple time delay.
306

input signal
Vs
+
comparator
t
digital-to-analog converter with counter
Vdac
t stop
Figure A.4: Principle of an analog-to-digital converter
Most A/D cards only have one analog-to-digital converter (A/D converter). The A/D converter is therefore connected (or multiplexed) to the data channels through a multiplexer which only connects the A/D to one channel at a time. This means that the channel signals are not converted at the same time, which can create phase problems in the data analysis if they are to used together. It is often desirable to take samples from all channels at exactly the same time. Therefore some A/D cards have a sample-and-hold component for each channel just before the multiplexer. At a trigger pulse or at a xed frequency, each channel voltage is stored, and then the stored channel voltages can be read one by one by using the multiplexer and the A/D converter. But even sample-and-hold will not x the problem if the anti-aliasing lters are not properly phase-matched to each other. The principles of an A/D converter are illustrated in gure A.4. The A/D converter determines the current signal voltage Vs by comparing the voltage to voltage generated by a digital-to-analog converter (DAC). A common method is to increment the D/A using a digital counter count from zero until the corresponding voltage from DAC matches the signal voltage. A comparator will then send a stop signal to the counter. The digital number in the counter will then be the digital representation of the signal voltage. It is obvious that the conversion takes some time and also that the more bits in the converted value (resolution of voltage), the longer the time needed for the conversion. Some A/D converters use more sophisticated and thus faster algorithms. The basic specication of an A/D converter is the number of conversions per second and the number of bits in the
A.3. ANALOG-TO-DIGITAL CONVERSION bits 8 12 16 levels resolution 28 = 256 0.4% 12 2 = 4096 0.02% 16 2 = 65536 0.002%
307
Table A.1: Common resolution of AD cards converted digital value. A number of properties should be evaluated when an A/D card or an A/D converter is selected for measuring task: Resolution: The resolution of the full range is determined by the number of bits in the converted digital value. Common values are shown in table A.1. Converters with as low as 10 bits or as high as 22 bits are also available. Even conversion to a 8 bit number has a resolution better than 1% of full range, and this is sometimes sucient with proper pre-conditioning of the signal (e.g., oset, amplication, etc.). A/D cards or converters with 16 bits are quite aordable today and can be very convenient; e.g., for signal that has small uctuations around a large mean value. Sometimes a 16 bit card can measure directly without as much signal conditioning as one with fewer bits. Sampling rate: General purpose 12 bit A/D cards are available with sampling rates up to 1000 kHz; i.e., up to 1.000.000 data samples per second can be obtained. If such an A/D card is connected to 5 channels, this means that each channel with be sampled with 200 kHz. Dedicated 8 bit converters can have sample rates up 1000 MHz. Cards may have local memory buers and can then only sample at the highest sample rates until the buer is full. Channels: As mentioned above, general purpose A/D cards multiplex several channels into a single A/D converter. Channels can be single ended or dierential. Single ended channels use a common ground and therefore only one wire is connected for each channel. This always introduces ground loops in the system, which can substantially increase the electronic noise and greatly complicate analysis. Much to be preferred are dierential channels which connect each pair of wires separately, so that all channels can be electrically isolated from each other. A general purpose card often comes with either 16 single ended channels or 8 dierential channels. The sampleand-hold option illustrated in gure A.3 will increase the cost of the card. Sensitivity: The sensitivity is limited by the number of bits and the range over which conversion is performed. When electronic noise is negligible, the effective noise on the system is set by the volts/bit. This is the so-called quantization noise. Since it appears on the digitized data as broadband
308
APPENDIX A. SIGNAL PROCESSING noise, it must be insured to be well below any signal of interest. The important control factors here are the signal conditioning (oset, amplication and ltering), and the general noise specications of the A/D card.
Appendix B Random Processes

Foreword: This section overlaps considerably (at this point at least) the material in Chapter 8. It was largely taken from the notes from courses on turbulence measurement at the Danish Technical University and Chalmers developed by Knud Erik Meyer, Peter B.V. Johansson and William K. George.
Many processes we wish to measure are in fact random processes, meaning that one can not guess exactly future values by looking at it past values. Some processes really are deterministic (meaning that for xed initial conditions and boundary values the same answer will be obtained every time in the process), but they behave chaotically and appear to be random. Turbulence is a good example of this. But even processes that are not really random, appear to us as random when we try to measure them because of our measurement (or computational) errors. Thus just about everything we encounter can be treated as random processes. We will call a variable measured from a random process a random variable. In the following, we will denote the random variable as u, since we often measure a velocity or velocity component. The variable could just as well be anything else, like local pressure, temperature, etc. And it could be a function of time and space, and most certainly which (of possibly many) experiments we performed. But to make statistics we need to be able to deal with independent events. The easiest way to understand this is to imagine doing the same experiment quite independently in many dierent universes simultaneously. Then we could compare what happened at exactly the same place at the same time in each universe, or average the results together. When we can compute such an average, the true average, we denote it by u. Even though we have still not dened the true average, we will presume it to exist. In fact we shall see that the only way to get a true average is to have an innite number of universes in which we are simultaneously performing an identical experiment in each. We will call our ith attempt to measure our random variable as the ith -realization, and denote it as u(i) . In the paragraphs below we shall assume each realization to be statistically independent from all others. This is discussed in detail in sec309
310
APPENDIX B. RANDOM PROCESSES
tion 2.4.3 of chapter 2, which provides a denition in terms of the joint probability density function. But for our purposes the most important consequence is that if two random variables, u(i) and u(j) , are statistically independent, then are uncorrelated. This means that u(i) u(j) = 0 if i = j. (Note that the inverse is not necessarily true; i.e., lack of correlation does not imply statistical independence.) Usually the rst thing we want to know about a random variable, u, is its mean value, u. This is really not a simple as it might seem, since the best we can ever do is estimate the true mean value (which we have presumed to exist for any random process). The reason is obvious if we look at the denition of the ensemble mean (or true) mean (or average):
N 1 (i) u = lim u N N i=1
(B.1)
Clearly one can never have an innite number of independent universes (other than in our imaginations). And in fact we can never even have an innite number in the one universe we have. Your rst thought is probably: Do we really need an innite number? Whats wrong with just taking a nite number, say UN given by:
N 1 (i) UN = u N i=1
(B.2)
Our problem is that this estimator (or estimate of the mean) is itself a random variable. This means that our every attempt to measure the mean will most likely produce a dierent result. Obviously we need to know a couple of things for measurements to make any sense. First, is there some reason to believe that the mean of our estimates is close to the true mean of the underlying random process? And second, is there hope that by making more estimates or for longer times that we have a better chance of getting closer to the correct answer? These two questions are usually asked this way: Is the estimator biased? And does the estimator converge with increasing time or number of samples? The rst question is generally addressed by examining the bias, the second by examining its variability. We shall examine both of these in the following subsections. The problem is that in a real measurement, we never have the true mean, say u, but can only use an estimator for the ensemble average based on a nite number of realizations, say N ; i.e., UN =
N 1 (i) u N i=1
(B.3)
Since u(i) is a random variable then UN is also a random variable. This means that if we repeat the experiment, we should not expect to get two values of UN that are exactly the same. There are two important questions to investigate when an estimator is used.
B.1. TIME-AVERAGED STATISTICS
311
1. Does the average of our estimator give the right average; i.e., does U N u? If this is the case, the estimator is an unbiased estimator. 2. Does the estimator converge towards the true value as we take more and more samples? We talk about convergence. It is straightforward to see that UN is an unbiased estimator by comparing eqs. (B.10) and (B.3). We can analyze the convergence by looking at the variability dened as (UN u)2 2 = . (B.4) N u2 If we insert the denition of our estimator from eq. (B.3), we get that 2 N
2 1 var(u) 1 u = = N u2 N u2
(B.5)
where the variance is estimated as var(u) =

N 1 (i) (u u)2 N i=1
(B.6)
and the standard deviation, u is the square root of the variance. Sometimes the variability is expressed using the turbulence intensity Tu = u /u (here u is a velocity) as 1 u Tu N = = . N u N (B.7)
Looking eq. (B.5) or (B.7), we see that N 0 as N . We also see that at N 1/ N ; i.e., if we increase the number of samples N four times, we should expect the condence interval between our estimated mean value and the true mean value to get 50% smaller. As discussed in detail in section 2.5 of chapter 2, N is an estimator of the standard deviation of UN (but only if the N samples are uncorrelated ). Thus N is a measure of the uncertainty caused by the limited number of samples, and it is therefore an important parameter for the design of an experiment.
B.1
Time-averaged statistics
As discussed in section A.1, most experiments are random processes designed to be stationary.
312
B.1.1
Time mean value

1T u(t)dt. T T 0
The time-averaged mean value for stationary signal is dened as u lim (B.8)
We have used the same symbol for the average as for the true average dened above, since if the process is stationary, both time and ensemble averages are the same. Like the true average though, this expression can not be evaluated in a real experiment since we can not measure for an innite time. So instead we are forced to deal with an estimator for the time average, say UT , dened by: 1T UT u(t)dt. T 0 (B.9)
Also usually we will not represent the signal as analog (or continuous), but instead use a discrete representation. The latter problem can be easily solved by rewriting equation (B.8) into a discrete form using values u1 , u2 . . . uN of u sampled with a constant time between samples t, UT
N N 1 (i) 1 (i) u t = u N t i=1 N i=1
(B.10)
since T = N t. Thus our discrete time average estimator looks exactly like our estimator for independent realizations from before. Therefore we can approximate the time average value this way, and it does not depend on the time between samples, t, as long as they are statistically independent. In a real measurement, of course, we can only take a limited number of samples, N . Moreover, our random process is a continuous process with its own memory of how it was at an earlier time (since the ow at one time is linked to other times by the governing equations). How then can any two samples be truly statistically independent? Given this question, we obviously need to worry about the same things we did before: Does our time average estimator converge? And does it converge to the right value? By comparing eqs. (B.8) and (B.9), it is easily seen that the estimator is unbiased; i.e., U T u as T . The question of convergence is addressed as before by studying the variability, which is given in this case by: 2 = T var{UT } (UT u)2 = u2 u2 (B.11)
(Note that in practice we use UT instead of u in the integral above.) After some eort involving multiple integrations, it can be shown that the relative error in estimating the mean is given by:

T =
2I T
var{u} u
2I u T u
(B.12)

35 1 0.8 u [m/s] 0.6 30 R 0.4 0.2 0 25 0 2 4 t [ms] 6 8 0 0.5 1 [ms] 1.5
313
(a) sample of velocity data
(b) autocorrelation function
Figure B.1: Hot-wire measurement downstream of a grid. where I is the integral time scale for the process dened by: [u u]2 I =
[u(t) u][u(t + ) u]d
(B.13)
Note that the integrand is the two-time correlation (or autocorrelation function) for the process dened by: R( ) = u(t)u(t + ). (B.14)
Also the variance in the continuous formulation can be estimated using its nite time estimator as: 1T varT {u} = [u(t) u]2 dt. (B.15) T 0 Comparing equation B.12 to equation (B.7) it can be seen that T /2I corresponds to the N of our independent samples above. This means that To have uncorrelated measurements, the time between samples should be at least two times the integral time scale (T /N 2I). If you have measured with a time between samples that is shorter than two times the integral time scale, you can estimate the standard deviation of the mean value using eq. (B.12). Said in another way, you will not get better convergence for the estimate of the mean value by sampling faster than one sample every 2I. Example: Autocorrelation from hot-wire measurement Figure B.1 shows a hot-wire measurement made downstream of a grid in a wind tunnel. The data is taken with a single wire probe with a sampling frequency of
314
50 kHz. The total data record is 2.6 seconds long (131072 samples). The rst 10 ms of the record is shown in gure B.1(a). The mean value of the total record is u = 30.23 m/s (here assumed to be the true mean value) and this is indicated as a horizontal dotted line in the gure. Figure B.1(b) shows the autocorrelation found by evaluating eq. (B.14) for the total data record. By integrating the autocorrelation, the integral time scale is found to be I = 0.20 ms. This is indicated in gure B.1(b) as a dashed rectangle with the same area as the area below the autocorrelation curve. In gure B.1(a) samples taken with a distance of 2I = 0.4 ms is shown with the symbol . The 10 ms record of the ow therefore corresponds to 25 statistically independent samples. The standard deviation of u can be calculated from the data to be u = 1.24 m/s and relative value of the standard deviation of the mean value can therefore be estimated using eq. (B.12) as UT 2I u 2 0.20 ms 1.24 m/s = T = = = 0.8% u T u 10 ms 30.23 m/s or a value of UT = 0.25 m/s. The actual mean value of the sample shown in gure B.1(a) is UT = 30.06 m/s or a deviation from u of 0.6%. The mean value of the rst 25 samples marked with the symbol gure B.1(a) is UN = 29.98 m/s or a deviation from u of 0.8%. Both of these deviations are clearly with in the statistical variation indicated by the value of T .
B.1.2
Higher moments
We often want to characterize our data using higher order moments. A higher moment is dened as 1T 2 n u = lim (u (t)) dt (B.16) T T 0 where we have dened the uctuation, u (t), to be; u (t) = u(t) u (B.17)
Using our previous results, we can get an error estimate for an arbitrary moment by substituting u with un and N with T /2I in eq. (B.5) to obtain 2 n u 2I var(un ) 2I u2n un 2 = = T un 2 T un 2
(
(B.18)
For the second moment (the variance of u) we get 2 2 u 2I = T u4 1 (u2 )2

)
(B.19)
Thus in order to estimate the error of the second moment, we need to know the fourth! For the third moment (the skewness) we get 2 3 u 2I = T
(
u6 1 (u3 )2
(B.20)
315
For the fourth moment (the atness or sometimes called the kurtosis) the error estimate becomes ( ) 2I u8 2 u 4 = 1 (B.21) T u4 2 Equations (B.19B.21) are very hard to compute from measured data. Therefore it is customary to estimate these using the approximation that the random process can take an analytical form. The most common form used in turbulence is to assume the process to be Gaussian. A random process is Gaussian if the probability density function p(u ) can be expressed as 1 u2 p(u ) = exp 2 2 2
(B.22)
where is the standard deviation, = u2 . Using the probability density function, the moments can be alternatively dened as un = un p(u )du (B.23)
This can be used together with the denition of the Gaussian to obtain a simple analytical form for the error of each moment. For a Gaussian process, the odd moments like the skewness is identical zero, so one can only obtain numbers for the second and forth moments 2 2 ,G = u 2 4 ,G = u Example: Deciding sampling times We want to measure in a turbulent ow and preliminary measurements have estimated a turbulence intensity of Tu = 20% and an integral time scale of I = 1.5 ms. We want to measure the mean velocity with a an error not larger than 1%. If we assume the errors to be distributed according to a Gaussian random process, a 1% error corresponds to two times the standard deviation and is sometimes expressed as a 95% condence interval. We therefore want to estimate the necessary sample times for the standard deviation of the estimated mean value to be T = 0.5%. Using equations (B.7) and (B.12) we nd the sample time to be T = 2 1.5 103 s 0.22 2I(Tu)2 = = 4.8 s T 0.0052 4I T 22I T (B.24) (B.25)
corresponding to an eective number of samples of N= T 4.8 s = = 1600 2I 2 1.5 103 s
316
Statistically it does not matter if we take more samples during the 4.8 seconds of sampling. The optimal sampling rate fs can therefore be considered to be fs = N 1600 = = 333 Hz T 4.8 s
We can use eqs. (B.24) and (B.25) to estimate the uncertainty for the higher moments for a sample time of T = 4.8 s. Using a factor of two on the standard deviations corresponding to a 95% condence interval, we nd the uncertainty on the second moment (the variance) to be 7% and for the fourth moment to be 17%. To obtain one percent uncertainty on the fourth moment, we must have a total sampling time equal to at least 1320 seconds (22 minutes!) corresponding to 1,750,000 uncorrelated samples!
Appendix C Fourier analysis of time varying signals

Foreword: This appendix was largely taken from the notes from courses on turbulence measurement at the Danish Technical University and Chalmers developed by Knud Erik Meyer, Peter B.V. Johansson and William K. George.
There are two ways of looking at how a signal evolves in time. One way is to look at the signal as a sequence of events (e.g., opening and closing a valve). Another way is to characterize the signal in terms of frequencies being present. An example is how we perceive sound. If we are listening to somebody talking, we understand words by listening to the sequences of dierent sounds in the words. However, the single sounds in the words (corresponding to letters in written words) are characterized by the combinations of frequencies in the sound which is a result of how we form the sound with our mouth. It is also interesting that we are able to recognize people we know only by a small sample of their voice. We simply recognize the combination of frequencies that characterize their voice. The human ear is a sophisticated analyzer of frequencies in a sound signal. The sound sensing organ has about 15 000 hair cells that each are tuned to a specic frequency. The brain therefore receives a measurement of the content of each frequency present in a sound. When we want to do a similar analysis of a measured signal we usually use Fourier analysis. This is an important tool to characterize data from a measurement from a turbulent ow, but it is also an important part of many optical measurement systems where the raw transducer signal is a frequency. The frequency is often determined by a Fourier analysis. In this section we focus on Fourier analysis in time, but it can just as well be done in space (e.g., if instantaneous data along a line is available) . 317
318
APPENDIX C. FOURIER ANALYSIS OF TIME VARYING SIGNALS
C.1
Fourier series
If we have a periodic signal (i.e., a signal that repeats itself every time interval T ) it can be developed in a Fourier series as: t u(t) = A0 + An cos 2n T n=1

t + Bn sin 2n T n=1
(C.1)
The frequencies present in this decomposition, fn = n/T , are harmonics (or integer multiples) of the fundamental frequency 1/T . The Fourier coecients, An and Bn , are given by: An =
( ) t 1 T /2 u(t) cos 2n dt T T /2 T
(C.2) (C.3)
( ) t 1 T /2 u(t) sin 2n dt Bn = T T /2 T
Example: Square wave Consider the square wave signal shown in gure C.1. Show that it can be reconstructed using the following expression: 4 u(t) = 1 t sin 2n T n=1,3,5... n

(C.4)
Figure C.1 illustrates that for n = 1 (dashed curve) the result is just a sine curve. As the Fourier components for increasing values of n are added, however, the resulting curve approaches the square wave signal. It is sometimes convenient to use a complex notation in the formulation of Fourier series. We can dene a complex coecient as Cn = An iBn and rewrite equations (C.2C.3) as: 1 T /2 Cn = u(t)ei2nt/T dt T T /2 (C.5)
The ratio between the real and imaginary values can provide phase information about the signal. It is also convenient to introduce negative values of n. This corresponds to negative frequencies (which can be thought of as waves going backwards in time). Then the values are symmetric in the sense that A(n) = A(n) and B(n) = B(n). The use of negative values of n means that the reconstruction formula now can be written very compactly as u(t) =
n=
Cn e+i2nt/T
(C.6)
C.1. FOURIER SERIES
319
n=1 and n=3 1 0.5 0 0.5 1 0 0.2 0.4 0.6 0.8 1 1 0.5 0 0.5 1 0 0.2 0.4
n=5
0.6
0.8
n=7 1 0.5 0 0.5 1 0 0.2 0.4 0.6 0.8 1 1 0.5 0 0.5 1 0 0.2 0.4
n=9
0.6
0.8
Figure C.1: Reconstruction of square wave signal using Fourier series
320
C.2
Fourier transform
The Fourier series discussed in the previous section only applies for a periodic deterministic signal; i.e., a signal that is exactly the same from period to period. But sometimes we must deal with a single pulse or even a random processes. We therefore need to be able to decompose a signal that is not repeatable. We can overcome this by using the Fourier transform. This can be viewed as a Fourier series in the limit for which the period, T becomes innite, and the corresponding values of n/T become a continuous range of frequencies f , meaning that all frequencies are now possible. We dene then, the Fourier transform of the function u(t) as:
u(f ) =
ei2f t u(t)dt
(C.7)
These are really the continuous counterpart to the Fourier series coecients of a periodic signal, and can similarly be used to reconstruct the original signal. We call this reconstruction the inverse Fourier transform and dene it as:
u(t) =
e+i2f t u(f )df
(C.8)
As in the complex Fourier series, we use negative values of the frequency f . An implicit assumption is that the integrals converge. This is, of course, never true for a stationary random process, so we need to use the idea of generalized functions to insure that they do. These are discussed in detail in section E, and will be important when we consider stationary random processes. In practice, however, these Fourier integrals exist for all experimental signals, sincethe time domain over which the integral can be performed is nite. Some of the consequences of this truncation in time are discussed under the sections about ltering and windows below.
C.3
Convolution
f g =
A convolution is an operation on two functions f and g dened as: f ( )g(t )d (C.9)
A convolution is an integral that expresses the amount of overlap of one function g as it is shifted over another function f . It therefore blends one function with another. We have already seen one example of a convolution in the autocorrelation function dened in section 8.2. The use of windows discussed in the next section can also conveniently be expressed in terms of convolutions. For the Fourier transform, some important relations for convolutions exist. Here, we will let F denote the Fourier transform and F 1 denote the inverse
C.4. THE FINITE FOURIER TRANSFORM Fourier transform. The most important results are: F[f g] F(f g) F 1 [F(f )F(g)] F 1 [F(f ) F(g)] = = = = F[f ]F[g] F[f ] F(g) f g fg
321
(C.10) (C.11) (C.12) (C.13)
In words, this means that the convolution of two functions in the time space corresponds to the product of the transformed functions in the frequency space and vice versa.
C.4
In any application of Fourier analysis we are always limited by the length of the time record, T . This means that the most we can expect to be able to transform is the nite time transform given by:
T /2 T /2
uiT (f ) =
ei2f t u(t)dt
(C.14)
where for convenience we have written it over the symmetric interval in time (T /2, T /2). Now with a little thought, it is clear that we are actually taking the Fourier transform of the product of two functions, the correlation (the part we want) plus the window function; i.e., uiT (f ) = F[u(t)wT (t)] = where wT ( ) is dened by: 1, wT ( ) = 0, | | > T /2 T /2 T /2 (C.16)
ei2f t u(t)wT (t)dt
(C.15)
From the results of the preceding section we immediately recognize that the Fourier transform we seek is the convolution of the true Fourier transform with the Fourier transform of the window function; i.e,, uiT (f ) = ui (f ) wT (f )
= =
u(f f )wT (f )df u(f )wT (f f )df
(C.17)
322
Obviously the Fourier transform at a given frequency is contaminated by the Fourier transform at all other frequencies, the exact amount depending on the window function, the Fourier transform of which is given by: sin(f T ) WT (f ) = T f T (C.18)
(See if you can show this.) The so-called side-lobes have their zeros at frequency which are multiples of 1/T and roll-o only as f 1 ; moreover, every other side-lobe is negative. We shall see later that windows play a very important role in spectral analysis. Unfortunately their importance is poorly understood by many (even texts) which erroneously confuse Fourier transforms and Fourier series. For example, if you simply imagine a nite piece of record to be repeated periodically, you will completely overlook the fact that your spectral analysis has been corrupted by the nite time window.
C.5
The shift theorem
Consider a signal, u(t), which is shifted in time by amount to obtain u(t + ). Sometimes this is unavoidable; e.g., in a multichannel A/D converter without simultaneous sample-and-hold where the channels are scanned sequentially. And sometimes we introduce time lags deliberately; for example, to maximize the correlation with another signal. Or as another example, in the next section it will be seen to be convenient to consider the Fourier transform over the interval (0, T ) instead of (T /2, T /2). It is straightforward to show that time-lags introduce a phase-shift on the Fourier transform which is linear with frequency. From the denition of the Fourier transform, it follows immediately that: F[u(t + )] =
ei2f t u(t + )dt
(C.19)
But we can dene x = t + and transform the integral to obtain: F[u(t + )] =
ei2f (x )u(x)dx
= e+i2
ei2f x u(x)dx (C.20)
= e+i2 F[u(x)] = e+i2 u(f )
Note that F [u(x)] = F[u(t)] since both t and x are just dummy integration variables in the integrals. If we look at the phase of the Fourier transform, we see immediately that our time displacement corresponds to a phase shift which is linear with frequency; that is: [ ] Im shif ted (f ) = tan1 = unshif ted + 2f (C.21) Re
C.5. THE SHIFT THEOREM
323
It is easy to show the the inverse is also true: a linear phase shift corresponds to a simple time delay in the signal. This can sometimes be used to great advantage in designing measurement systems.
324
Appendix D Digital Fourier transforms

D.1
Aliasing of periodically sampled data
Before considering what happens if we digitally sample an signal for a nite length time interval, lets consider rst what happens if we digitally sample an innitely long record. The easiest way to think about this is to make a simple model for the analog-to-digital conversion process in which the digitally sampled signal is treated as the limit of a continuously varying one. Although this involves generalized functions (which are treated in detail in section E), the only result we need to use here is the familiar delta-function, (t) which is innite at t = 0, zero everwhere else, and in the limit as 0 has the following properties:

+ + +
(t)dt = 1
(D.1) (D.2) (D.3)
f (t)(t)dt = f (0)
f (t)(t to )dt = f (to )
Using these we can dene a sampling function, say g(t) as: g(t) = t
n=
(t nt)
(D.4)
where t is the distance between samples. (Note that multiplication by t makes g(t) dimensionless, since the delta-function of time has dimensions of inverse time.) Now we can represent our sampled signal as: us (t) = u(t)g(t) 325 (D.5)
326
APPENDIX D. DIGITAL FOURIER TRANSFORMS
It is easy to see that this has all of the properties of a sampled time series. For example, the time average over the interval from 0 to T is given by: 1 UT = T
0 T
1 us (t)dt = T
u(t) t
n=
(t nt) dt (D.6)
N 1 = u(nt) N n=1
since T = N t is the record length. But this is exactly the digital approximation given by equation B.10. Similarly we can examine the Fourier transform of our digitally sampled signal which is given by: F[us (t)] = =
ei2f t us (t)dt ei2f t u(t)g(t)dt (D.7)
Since the right-hand side is the Fourier transform of the product of two functions, we know immediately that the result is the convolution of the Fourier transform of each function individually; i.e., F[us (t)] = {F[u(t)] F[g(t)]} = where as shown below, g (f ) is given by: g (f ) =
n=
u(f f )(f )df g
(D.8)
(f n/t) :
(D.9)
Proof of equation D.9 First note that g(t) is periodic, so it can be represented by a Fourier series. If we center the delta function over one period in the interval (t/2, t/2), the Fourier series coecients are given by; 1 t/2 i2mt/t Cm = e t(t)dt = 1 t t/2 Thus the Fourier series representation of g(t)is given by: g(t) =
m=
(D.10)
e+i2mt/t
(D.11)
Now take the Fourier transform, term by term, of this Fourier series representation which is given by:
[
m=
+i2mt/t
= =
m= m=
ei2f t e+i2mt/t dt
(f m/t) (D.12)
D.1. ALIASING OF PERIODICALLY SAMPLED DATA

(a) Fourier transform of true signal |( f )|
327
-2fs
-fs
fs
2fs
(b) Fourier transform of properly sampled signal |( f )|
-2fs
-fs
fs
2fs
(c) Fourier transform of aliased signal |( f )|
-4fs
-3fs
-2fs
-fs
fs
2fs
3fs
4fs
Figure D.1: Relation between aliasing and sample frequency fs = 1/t
It follows immediately that the Fourier transform of our sampled signal is given by: F[us (t)] = = =
u(f f )(f )df g
m= m=
u(f f )(f m/t)df (D.13)
u(f m/t)
Thus the Fourier transform of our sampled signal is an innitely repeated version of the Fourier transform of the original signal. This is illustrated in gure D.1. We will assume that gure D.1(a) shows the real Fourier transform of our signal. Note that the signal is symmetric around f = 0 as we have just discussed. The result of a discrete Fourier transform with data sampled with a sample frequency fs that is three times the maximum frequency present in the signal is shown in gure D.1(b). We see that the result
328
is repeated with a period of fs . If we instead use a sample frequency that is only 1.5 times the maximum frequency in the signal, we get the result shown in gure D.1(c). The result from the true distribution is still repeated with a period of fs , but there now is an overlap between the periods. The result (shown as a thick line) is an incorrect (or aliased) distribution in the regions of overlap. This is an important problem called aliasing. If we sample too slowly, the results get corrupted lower frequencies get a contribution from higher frequencies and vice versa. This is also illustrated in gure D.3 of section D.3. The sine signal has period frequency of 0.7 the sampling frequency. However, two other sine signals with period time of 0.3 and 1.3 of sampling frequency, respectively, also match the data samples. After sampling, there is no way of determining what the real signal really was. The key to avoiding aliasing is to satisfy the so-called Nyquist criterion: A signal must be sampled with a sample rate that is at least twice the maximum frequency present in the signal. In gure D.3 this means that only the slowest signal (period frequency of 0.3 time the sampling rate) would be correctly sampled with the illustrated sampling; i.e., there is no sine function with a frequency lower than 0.3 that matches the sampled data. You may think that you are only interested in the lower frequencies and therefore do not have to worry about higher frequencies. This is wrong. If you proceed this way it is quite likely that the lower frequencies will be corrupted by the higher frequencies! Worse, you will have little chance of detecting the problem and you will have no way of xing the problem after the data is taken. You may argue that you dont know what the highest frequencies are in your signal. There is always some noise that we cannot control. Well, it is your task to ensure this is not a problem! This is the reason why most A/D cards have a low pass lter before the sampling and the A/D conversion. To be safe, the lter frequency should be 35 times lower than the sampling frequency. We will have a closer look on lters in section G.2.
D.2
The Discrete Fourier transform
As in section B.1, the problem is both that the signal is only known at discrete points and that we only will have data for a limited time period. Now we will look at how a discrete formulation can be made. First we have to change eqs. (C.7) to a nite time estimate,
uT (f ) =
ei2f t u(t)dt
(D.14)
Note that we have deliberately shifted the time axis by +T /2. This introduces a linear phase shift equivalent to multiplying the Fourier coecients by e+if T compared to the symmetric nite transform of the previous chapter. Generally this will not present a problem, but should not be ignored. This is not exactly what we want, since we have sampled the velocity at discrete
D.2. THE DISCRETE FOURIER TRANSFORM times. We really have un = u(nt) n = 0, 1, 2, . . . , N 1
329
(D.15)
Using the basic denitions of integral calculus, we can discretize the integral of equation (D.14) as: uT (f ) =
N 1 n=0
ei2f nt un t
(D.16)
N is the total number of samples and T = N t is the total sample time. The time between samples, t, is given by the sampling frequency, t = 1/fs = T /N . This is, of course, an approximation which becomes exact in the limit as the number of points, N , becomes innite and as the interval between them, t, goes to zero. Now since we only have N data points, we can only calculate N independent Fourier coecients. In fact, since we are in the complex domain, we can only calculate N/2, since the real and imaginary parts are independent of each other. So we might as well pick the frequencies for which we will evaluate the sum of equation D.16 for maximum convenience. For almost all applications this turns out to be integer multiples of the inverse record length; i.e., fm = m m = T N t m = 0, 1, 2, . . . , N 1 (D.17)
Substituting this into equation D.16 yields our discretized Fourier transform as:
{
uT (fm ) = T
1 N 1 i2mn/N e un N n=0
m = 0, 1, 2, . . . , N 1
(D.18)
This equation can be evaluated numerically for each of the frequencies fm dened in eq. (D.17). The Fourier coecients, uT (fm ), found from eq. (D.18) are complex numbers. For future reference note they negative frequencies are mapped into fm = N 1, N 2, N 3, N m instead of at negative values of m, as illustrated in Figure D.1. It is easy to show the original time series data points can be recovered by using the inverse discrete Fourier transform: 1 un = T
{N 1
m=0
+i2mn/N
uT (fm )
(D.19)
The real and imaginary value of at frequency fm corresponds to the coecients of the cosine and sine parts respectively, in a reconstruction of the signal. In fact if we divide uT (fm ) by T, then it is exactly the Fourier series coecient that would represent a periodic signal of period T , but equal to our piece of the record over the interval 0, T . Thus the algorithms for treating a discretized periodic signal and a discretized piece of an innitely long signal are the same. Because of this the
330
terms in brackets of equations D.18 and D.19 are so important they are usually written as a pair like this: um = un =
1 N 1 i2mn/N e un N n=0
N 1 m=0
(D.20) (D.21)
e+i2mn/N uk
Equations D.20 and D.21 are computationally very demanding since they require N 2 operations. Fortunately if N is an integer power of 2, 3 or 5, there is an algorithm called the Fast Fourier Transform (or simply the FFT) that is computationally very ecient and requires only N ln N operations.1 The FFT algorithm is available in many software packages for numerical calculations. Hardware solutions (dedicated electronic devices) are also available and are used in several measurement instruments. Example: Look up the FFT implementation in MATLAB. Note that MATLAB uses a slightly dierent denition with the indices n and m running from 1 to N instead of from 0 to N 1, with corresponding dierent arguments inside the summations. A change of variables m = m + 1 and n = n + 1 in eq. (D.18) yields exactly the same denition as used in the MATLAB implementation.
D.3
An Example
Let us try to do an FFT analysis on a very simple signal: we will sample a cosine function with 10 samples per period and we will use N = 32 samples and put them into a vector u. The samples are illustrated with the symbol in gure D.2(a). Assuming a unit time between samples, the frequency in the signal is f = 0.1. The theoretical result in the frequency domain should therefore be a single peak at f = 0.1 and zero value of f everywhere else. The result of running an FFT on these 32 samples is shown in gure D.2(b) as the magnitude of u(f ). The MATLAB command used for the calculation is simply uf=abs(fft(u)). The values of the frequency are found using equatiob (D.17) by dividing the vector index with N . The result is probably a bit surprising. We do get a clear peak at the signal frequency f = 0.1. However, the result also show some frequency content at all other frequencies. Furthermore we see that the result is perfectly symmetric around f = 0.5 and that we therefore also have a clear peak at f = 0.9. To understand the symmetry, remember that we introduced the concept of negative frequencies in the denition of the Fourier Transform. If we instead calculated the frequencies using an index range from N/2 to N/2, we would have gotten a result that was symmetric around f = 0. Actually, we can evaluate eq. (D.18) for any value of m. However, since cosine and sine give the same values
1
See [?] for more details on the FFT.
D.3. AN EXAMPLE
331
(a) cos(x) sampled at 10 samples per period 20 1 0.5 0 0.5 1 0 5 10 t 15 20 0 0 || u 15
(b) spectrum of sampled signal (N=32)
10
0.2
0.4 f
0.6
0.8
(c) sinc(x) and its absolute value (dots) 1 0.8 0.6 || 0.4 0.2 0 0.2 0.4 10 5 0 x 5 10 0 0 5 15 20
(d) spectrum of sampled signal (N=256)
10
0.2
0.4 f
0.6
0.8
Figure D.2: FFT on sampled cosine signal
332
1
0.5
0.5
3 t
Figure D.3: Sine signal (solid curve) with frequency 0.7 sampled at unit time between samples sine curves with frequencies 0.3 and 1.3 also matches sampled data. for arguments that have a dierence of 2, the result will repeat itself every time m is increased with N . Returning to gure D.2(b), we see that the reason why we get a symmetry around f = 0.5 is that we see the real, positive frequencies in the region from f = 0 to f = 0.5 and then the negative frequencies from the next period of the result (repeated with fs = 1) in the region from f = 0.5 to f = 1. The symmetric part is only there because it is a technical convenient way of doing the calculations. When we present the results, we therefore only need to present the rst half of the result corresponding to positive frequencies.
Appendix E Generalized functions

In Appendix C we dened the Fourier transform and Inverse Fourier Transform by equations C.7 and C.8 respectively. The Fourier coecients u(t), of the signal, u(t) were given by u(t) =

ei2f t u(t)dt
(E.1)
From them the original signal can be reconstructed using inverse Fourier transform by: u(t) =

e+i2f t u(t)df
(E.2)
Now if you have studied Fourier transforms in an applied math course, you have probably already spotted one potential problem: the integrals of equations E.1 and E.2 may not even exist at least in the ordinary sense. Figure E.1 illustrates the problem. For example, a stationary random process has no bounds (unless we arbitrarily truncate it over some interval). Moreover, since its statistical properties are independent of origin, the uctuations simply go on forever. Thus our random signal is really rather nasty, mathematically speaking, and most certainly the integrals in the ordinary sense become unbounded. There are, of course, many other kinds of signals for which the integrals do not converge either, some of which are of great interest to us. Examples include: u(t) = 1, u(t) = cos(2fo t), u(t) = sin(2fo t), to cite but a few. So we have a dilemma. Our decomposition of signals using Fourier transforms does not seem possible, since the integrals do not converge. The resolution to our dilemma lies in a major mathematical development of the 20th century the theory of generalized functions. There are numerous references which one can consult for a more proper mathematical treatment than the rather cursory and intuitive treatment here. (Lumley 1970, Lighthill 1955 are two of my favorites). In brief the basic idea is to replace functions whose integrals do not converge with functions which do. Great idea, Im sure you are thinking, but doesnt this require magic? In truth it is almost magic, since in the end we almost never worry about what we have done, and almost always just go on doing regular mathematics like nothing ever happened. 333
334
APPENDIX E. GENERALIZED FUNCTIONS
35
30
25
u [m/s]
20
15
10
0 0
8 t [ms]
10
12
14
Figure E.1: Hot-wire measurement of the main ow velocity component in a turbulent ow behind a cylinder sampled at 100 kHz. Data from [?].
335 Only rarely will we have to actually consider that we are working with generalized functions. Impossible, you say. Lets consider the following simple example. Suppose I want to take the integral of the function, f (t) = 1, from (, ). Obviously this integral does not exist. Nor, in fact does its Fourier transform exist, at least (in the ordinary sense). Now consider a second function, say: gT (t) = et
2 /2T 2
(E.3)
Now since the tails of this function roll-o exponentially, it certainly is integrable. And in fact the integral is given by: dt = 2T (E.4) (You know this from Chapter 2, since (1/ 2T )exp(t2 /(2T 2 ) is exactly the Gaussian which integrates to unity.) Our integrable function gT (t) also has a wonderful Fourier transform, wonderful in the sense that not only does it exist, all its derivatives exist also; i.e.,
et
2 /2T 2
FT{et
2 /2T 2
}=
ei2f t et
2 /2T 2
dt =
2T e(2f T )
2 /2
(E.5)
This is easy to compute by completing the square. So we have one nasty function, f (t) = 1, and one wonderful function, gT (t); the former has no integral, and hence no transform in the ordinary sense, but the latter has both. Now note something interesting. The limit of gT (t) 1 as T , which is exactly the value of our nasty function, f (t). In fact, we could just dene a new function by the product fT (t) = f (t)gT (t) and note that:
T
lim fT (t) = lim f (t)gT (t) = f (t)

T
(E.6)
In fact, even more interestingly, the Fourier transform of our new function, fT (t), also exists in the ordinary sense. In this case, its just the Fourier transform of gT (t) itself. Here is where one of the really good ideas of the last century appears1 , the magic if you will. Lets just dene the Fourier transform of our nasty function, f (t), in the sense of generalized functions to simply be the limit of the Fourier transform of fT (t) as T ; i.e, FTgf {f (t)} = lim FT{fT (t)} = lim
T
ei2f t f (t)gT (t)dx
(E.7)
The Fourier transform of 1 in the sense of generalized functions is so useful, we have given it a special name, the delta-function; i.e.,
One of the rst to see this was the electrical engineer named Heaviside and he invented the step function which bears his name.
1
336
(t) lim GT (t)

T
(E.8)
where GT (t) can be any function whose integral is unity and which becomes undened at t = 0 and zero everywhere else in the limit as T . Im sure you have seen before, but you may not have realized that it was a generalized function. In general, the generalized functions are not uniquely dened. For example, all the functions below are suitable for dening (t): GT (t) = et /2T G2L (t) = e|t|/T sin(t/T ) G3L (t) = t/T
2 2
(E.9) (E.10) (E.11)
The rst and last have continuous derivative everywhere, the second has no derivative at the origin. When working with Fourier transforms, it is generally best to dene them in terms of functions which both go to zero exponentially fast, and which have all derivatives continuous. There is nothing in this course which needs anything more than GT (t), the Gaussian version, or 2T times it. We can generalize this whole procedure to almost any arbitrary function, whether deterministic or random. For example, suppose we have a stationary random process of time, say v(t). Then we can dene its Fourier transform in the sense of generalized functions to be: v (f ) FTgf {v(t)} = = lim FT{v(t)gT (t)}
(E.12) (E.13)
lim
ei2f t v(t)gT (t)dt
where gT (t) can be any function for which the product v(t)gT (t) is integrable and for which:
T
lim v(t)gT (t) = v(t)
(E.14)
Obviously a suitable choice is the Gaussian function we started o with; i.e., gT (t) = et
2 /2T 2
(E.15)
Exercise: Show that the Fourier transforms in the sense of generalized functions of ei2fo t , cos(2fo t) and sin(2fo t) are (fo ), [(fo )+(fo )]/2 and i[(fo )+ (fo )]/2 respectively using the Gaussian version of gT (t) dened above. Exercise: Compute the inverse transforms from the above example. Do NOT use the short-cut version where you assume the properties of a delta-function, but
337 instead work with the actual transformed version of f (t)gT (t) under the limit sign, then take the limits. For the rest of this course, we will simply agree that whenever there is any doubt, we always mean the Fourier transform in the sense of generalized functions. For example, when we consider homogeneous random processes in threedimensions, we consider the three dimensional spatial Fourier transform of the velocity eld, ui ( , t), dened by: x ui ( t) = k, 1 (2)3

x eik ui ( , t)d x x
(E.16)
(Note the factors of 2 appear in a spatial transform because we are transforming over the wavenumber vector, instead of f .) Since we are transforming functions k, for which the integrals do not converge, we really mean the Fourier transform in the sense of generalized functions dened by: ui ( t) FTgf {ui ( , t)} k, x 1 i x = lim e k [ui ( , t)gL3 ( )]d x x x T (2)3
(E.17) (E.18)
where gL3 ( ) is some suitably dened function which makes the integral exist. An x excellent choice for gL3 ( ) would be: x gL3 ( ) = e[x1 +x2 +x3 ]/2L x
2 2 2 2
(E.19)
whose Fourier transform (in the ordinary sense) is given by: GL3 ( = k) L3 2 2 2 2 e[k1 +k2 +k3 ]L /2 (2)3/2 (E.20)
We have used exactly this denition to show that Fourier coecients in nonoverlapping wavenumber bands are uncorrelated. Exercise: Find the Fourier transform of 1 in three-dimensions using generalized functions, then show how you might represented it symbolically as a three-dimensional delta-function, ( k). Exercise: If the Fourier transform can be represented in the sense of generalized functions as (| o |), nd the inverse Fourier transform in the sense of k k generalized functions.
338
Appendix F Spectral analysis of random signals

Stationary random processes commonly arise in the study of random signals, and especially in the study of turbulence. As we shall see below, Fourier transforming a random signal leads to a Fourier transform that is itself random; i.e., the Fourier coecients associated with any frequency are random and will be dierent for each realization. Thus we must use statistical analysis in conjunction with Fourier transformation if we wish to examine the frequency content of our signals. This could be quite complicated were it not for the implications of stationarity. As introduced in Chapter 8, a stationary random process is one for which all the statistics are independent of origin in time. Of particular interest to us in this appendix will be the autocorrelation given by u(t)u(t + ) and its Fourier transform, the spectrum.
F.1
The Fourier transform of a random signal
The Fourier transform and the inverse Fourier of a vector function of time, say ui (t), form a transform pair given by:
ui (f ) = ui (t) =
dt ei2f t ui ( , t) x df e+i2f t ui (f )
(F.1) (F.2)
Note that we have moved the dierentials, dt and df , next to the integral sign, so it will be obvious which variable is being integrated. Also it is understood that everything to the right of the dierential is to be integrated over those variables. Finally, note that although we have chosen in this section to talk about vector functions of time, everything can be directly applied to scalar functions of time as well just eliminate the subscripts. 339
340
APPENDIX F. SPECTRAL ANALYSIS OF RANDOM SIGNALS
Now Im sure you are asking: Why is the applicability of Fourier analysis such a BIG THING? There are two big reasons (among many). The rst has to do with what happens when you take the inverse transform at the time t, multiply it by the complex conjugate of the inverse transform at time t , and average to get the two-point correlation, Ri,j (t, t ); i.e., Ri,j (t, t ) = ui (t)uj (t )
(F.3)
+i2(f t f t)
df
df e
(f )j (f ) ui u
(F.4)
But we have assumed the eld to be stationary, so the two-point correlation can depend at most on the time separation, = t t; i.e., Ri,j (t, t ) = Bi,j ( ) Therefore equation F.3 is simply:
(F.5)
Bi,j ( ) =
df
df e+i2(f t f t) (f )j (f ) ui u
(F.6)
and the left-hand side has no dependence on either t or t separately, but is only a function of = t t. Now look carefully at the right-hand side. Clearly, unless a miracle occurs in the integration, the right-hand side is going to always depend on t and t. Guess what? You probably guessed it. A miracle DOES occur well, not really a miracle, but even better than a miracle. This miracle can be proven to be true by using the generalized functions of the previous section. The miracle is that since both sides of equation F.3 MUST depend only on = t t, it follows immediately that the Fourier components in non-overlapping frequency bands must be uncorrelated. Say what, you say? Exactly this:
{
(f ) (f )df df ui uj or equivalently:
Si,j (f )df , f = f 0 , f = f
(F.7)
u (f )j (f ) = Si,j (f )(f f ) ui
(F.8)
where ( ) is the familiar delta-function (not to be confused with the Kronecker delta tensor) and Si,j (f ) is a deterministic function called the velocity crossspectrum tensor. It is easy to see by substitution that our two-point velocity correlation function is the inverse Fourier transform (in the ordinary sense) of the velocity crossspectrum tensor; i.e., Bi,j ( ) =

ei2f Si,j (f )df
(F.9)
F.2. PROOF OF WIENER-KHINCHIN THEOREM USING GENERALIZED FUNCTIONS341 It is a bit more dicult to show that the cross-spectrum is the Fourier transform (in the ordinary sense) of the two-point velocity correlation function; i.e.,
Si,j (f ) =
ei2f Bi,j ( )d
(F.10)
Thus the cross-spectrum and the two-point correlation form a Fourier transform pair. u (f )j (f ) ui Exercise: Use the denition of gT (t) in the preceding chapter and prove that = Sij (f )(f f ) if ui (t) is a stationary random variable and ui (f ) is dened in the sense of generalized functions by:
ui (f ) = limT
dt ei2f t ui (t)gT (t)
(F.11)
(Hint: The solution is in the next section. But see if you can work it out before looking at it.) The implications of what we have accomplished become immediately obvious if we evaluate the inverse transform of equation F.9 at = 0 to regain the singlepoint cross-correlation; i.e.,
Bi,j ( ) =
df Si,j (f )
(F.12)
Si.j (f ) is telling us how the single-point Reynolds stress, ui uj = Bi,j ( ), is distributed over the various frequencies. This is even more obvious if we contract the two indices by letting i = j, sum and divide by two to get the energy; i.e., 1 2 1 1 q = Bi,i (0, t) = df Fi,i (f ) 2 2 2 (F.13)
Clearly the contracted cross-spectrum (usually called simply the energy spectrum) is telling us exactly how the turbulence energy is distributed with frequency.
F.2
Proof of Wiener-Khinchin Theorem using generalized functions
Start with the denition of the Fourier transform in the sense of generalized functions:
ui (f ) =gf =
ei
2f t
ui (t)dt ei
2f t
limT
ui (t) gT (t)dt
(F.14)
342
where we choose for convenience: gT (t) = e[t

2 /2
T 2]
(F.15)
Now examine the product (f ) uj (f ). Substitution from equation F.14 and ui changing the dummy variable under the integral sign yields: (f ) uj (f ) = limT (f ) uj (f )T ui ui

(F.16)
where (f ) uj (f )T is dened to be the integral before taking the limit; i.e., ui (f ) ui uj (f )T =

ei
2 [f t f t]
ui (t)uj (t )et
2 /2T 2
et
2 /2T 2
dtdt (F.17)
t
L/2 -L/2 L/2
-L
-L/2
-L
Figure F.1: Leftmost: original domain. Rightmost: transformed domain. If ui (t) is a stationary random process, then the two-time correlation is a function only of = t t only; i.e. ui (t)uj (t ) = Bi,j ( ). So we need to transform the integral to reect this. For convenience we also dene a second new independent variable as = t + t. The original and transformed domains can be viewed as the limits as L of the nite domains shown in Figure F.1. For 0, L + L and 0, L L + . Dening f = f f and transforming the integral yields: (f ) uj (f )T ui { 1 L d ei = limL 2 0
[2f +f ]
Bi,j ( ) e
2 /4T 2
(F.18)
L
d e
i [2f +f ]
Bi,j ( ) e
2 /4T 2
L+ L+ L
d ei
i f
2 /4L2
d e
2 /4T 2
F.3. THE FINITE FOURIER TRANSFORM

343
Taking the limit as L and recombining the integrals yields immediately: (f ) ui uj (f )T =

d e
i [2f +f ]
Bi,j ( ) e
2 /4T 2
d ei
2 /4T 2
(F.19)
The factor of 1/2 comes from the Jacobian since we have mapped into a domain with twice the area. It is easy to see that the inner integral is just the Fourier transform of a Gaussian with standard deviation 2T and transform variable f / 2, so it reduces to: 2 2 2 d ei f e /4T = 2 T e[2f T ] (F.20)
which when integrated over all f has integral unity. Therefore in the limit as T this will become our delta-function, (f ) = (f f ). Since it is independent of we can move it in front of the integral over to obtain: (f ) uj (f ) = limT (f ) uj (f )T ui ui [2f T ]2 = limT 2 T e (F.21)
d ei
[2f +f ]
Bi,j ( ) e
2 /4T 2
By taking the limit as T it is easy to see that this reduces to: (f ) uj (f ) = Si,j ([f + f ]/2)(f f ) ui
(F.22)
Since the delta function kills of anything for which f = f , this is equivalent to: (f ) uj (f ) = Si,j (f )(f f ) ui (F.23)
where Si,j (f ) is the spectrum given by the Fourier transform of the autocorrelation function; i.e., Si,j (f ) = ei2f Bi,j ( )d (F.24)
F.3
We saw earlier in section D.2 that in reality we are always limited to a nite domain. This means that the most we can expect to be able to transform is the nite time transform given by equation D.14; i.e., for a vector eld:
T /2 T /2
uiT (f ) =
ei2f t ui (t)dt
(F.25)
where for later convenience we have written it over the symmetric interval in time. Note that this introduces a linear phase shift equivalent to multiplying the Fourier coecients by e+if T . Clearly if ui (t) is random, then so is uiT (f ), just as was ui (f ) above for the innite domain. What is not so obvious is the relation between the spectrum associated with ui (f ) and that associated with uiT (f ). Note that equation F.8 is not particularly useful to us in actually computing the spectrum, since it contains the generalized function, (f f ), and thus has meaning only in the limit.
344
F.3.1
An indirect method
Of course we could take the data and compute from it u(t)u(t + ) for all the time-lags available to us on the interval 0, T . For example if we have a stationary random process we could dene an estimator for the correlation from a time average like this: 1 T | | Bi,jT ( ) = ui (t)uj (t + )dt (F.26) T 0 Note that it may seem strange to divide by T when the integration is over the interval T | | since this produces a biased estimator, but it actually has a lower variability this way. (See if you can show this.) Also, one disadvantage of this method it is considerably more computationally intensive than the method outlined in the following section, the number of multiplications being proportional to N 2 for N point records. Once we compute Bi,jT ( ), we can Fourier transform it to get a spectral estimator, i.e.,
T /2 T /2
Si,jT (f ) =
ei2f Bi,jT ( )d
(F.27)
where we have made the interval symmetrical since we need both positive and negative time lags. Note that we have completely avoided the need for generalized functions, since our correlation estimator is an ordinary function. The problem with this method can be seen by re-writing the spectral estimator using a window function like this;
Si,jT (f ) = where WT ( ) is dened by: 1, WT ( ) = 0,
ei2f Bi,jT ( )WT ( )d
(F.28)
T /2 T /2 (F.29) | | > T /2
Now it is clear that we are actually taking the Fourier transform of the product of two functions, the correlation (the part we want) plus the window function. Thus the spectrum we obtain is not the spectrum we seek, but the convolution of the true spectrum with the window function; i.e,, Si,jT (f ) = Sij (f ) WT (f ) =
Si,j (f f )WT (f )df
(F.30)
In other words, the spectrum at a given frequency is contaminated by the spectrum at all other frequencies, the exact amount depending on the window function. This window function is given by: sin(f T ) WT (f ) = T f T (F.31)
F.3. THE FINITE FOURIER TRANSFORM
345
The so-called side-lobes have their zeros at frequency which are multiples of 1/T , but roll-o only as f 1 which is much slower than most spectra. Moreover, every other one is negative, which means you can actually compute negative spectral values for some frequencies. This is of course, quite unphysical. Although other windows (e.g., Hanning, Hamming and Parzen) can be introduced to improve things, the best approach is to avoid this method if at all possible.
F.3.2
A direct method
Another better way to proceed is to work directly with the nite transform of equation F.25, and create the spectral estimator dened by: u (f )jT (f ) iT u (F.32) T The spectrum generated by a single record of data will of course be random, since the ui s are random. But the trick is to sub-divide a very long record into blocks of data which are transformed separately, then averaged the spectral estimators for each block. The more independent blocks that are averaged together, the smoother the resulting spectral estimates (i.e., the more blocks, the lower the variability). We can show directly that this makes sense by substituting from equation D.14 into the denition of equation F.32 to obtain: Si,jT (f ) = 1 Si,jT (f ) = T
{
T /2
} {
T /2
e+i2f t ui (t)dt
T /2
T /2
i2f t
uj (t )dt
(F.33)
where we have changed the dummy integration variable in the second expression to t since we want to put the integrals together. Combining the integrals and averaging yields: 1 T /2 T /2 i2(f t f t) Si,jT (f ) = e ui (t)uj (t )dtdt T T /2 T /2 (F.34)
But because the process is statistically stationary, the cross-correlation depends only on the time dierence, t t = ; i.e., ui (t)uj (t ) = Bij ( ) only. Thus, we can use the same trick we used in section 8.7 to transform the integrals, then integrate one of them out to obtain:
Si,jT (f ) =
i2f
| | d Bi,j ( ) 1 T
(F.35)
It is easy to see that as T this estimator reduces to exactly equation F.10, so this estimator (unlike the one of the previous section) is unbiased. But like it, it is also clearly the convolution of the spectrum we want with a window. This time, however, the window is given by the Fourier transform of:
346
1 WT ( ) = 0,
| | , T
T T (F.36) | | > T
It is straightforward to show that in this case WT (f ) is given by: sin(f T ) WT (f ) = T f T

[ ]2
(F.37)
So the sidelobes have their zeros at multiples of f = 1/T just like the top-hat window of section F.3.1, but they roll o as f 2 . Moreover, they are always positive, meaning that it is impossible to produce negative spectral values.
F.4
An example: digital spectral analysis of random data
We will now use Fourier analysis on some real data. First, we will use the rst 1024 samples data shown in gure A.1 (rst 10 ms of data). The data are taken 4 diameters downstream of a cylinder in a turbulent cross ow. Flow settings were matched to a vortex shedding frequency of 1000 Hz. Putting the velocity data into a vector u, we use to following MATLAB commands: N=1024; dt=0.00001; uf=fft(u(1:N)); f=[0:N-1]/(N*dt); uf2=abs(uf).^2/(N*dt); loglog(f(2:N/2),uf2(2:N/2)); % % % % % % number of samples time between samples do fft make frequency axis calculate spectrum loglog plot of result
The result is shown in gure F.2 (marked rst record). The six MATLAB script lines make the following operations: First we dene number of samples to be 1024 and the time between samples to be 0.01 ms (i.e. sample frequency of 100 kHz). We do the fast Fourier transport on the rst 1024 samples and store them in vector uf. The frequency axis is then computed using eq. (D.17)), and the power spectrum is calculated using eq. (F.32)). Finally, the plot is made. Note that the plot only is produced for indices from 2 to N/2. The rst index correspond to f = 0 or m = 0 in eq. (D.18)), and (as can easily been seen from the FFT algorithm), it corresponds to the mean value of u and not a real frequency. It is therefore not included in the plot. If we wish to obtain the real zero frequency we must estimate it by extrapolation from the lowest frequencies we have. As discussed in section D.3, the result is symmetric around N/2 and we therefore
F.4. AN EXAMPLE: DIGITAL SPECTRAL ANALYSIS OF RANDOM DATA347
10
First record 128 records
10
10 S(f)
10
10
10
10
10
10
10 f [Hz]
Figure F.2: Power spectrum of the data presented in gure A.1.
348
only present the rst half of the result. The results are usually plotted in a double logarithmic plot since both axis cover several decades. The result of this analysis is shown as rst record in Figure F.2. The result is not very satisfying. We do see a peak near 1000 Hz, but also other large peaks. Basically the the result seem very noisy. What we have done is to estimate 1024 frequencies using 1024 data samples. Our input data is a random data set, and therefore the result is just another random data set. To get a better estimate we clearly need to use much more input data. But using a longer data record will not by itself solve the problem, since we will also get more frequencies in the result (increase the resolution of the frequency axis). We there need to do some sort of averaging. One methods is to average adjacent frequencies, which is reduces the variability just like averaging any random numbers. It also decreases the frequency resolution by smoothing the data. Another approach is to divide the original data record into a number of shorter records of the same length and then take the average of the results. The full data series of the data shown in Figure A.1 consists of 131 072 samples (1.3 seconds). This can be split into 128 records of each 1024 samples. The average of the results of FFT analysis of each of these records is also shown in Figure F.2. The result is much smoother and now has a single dominant peak at 1000 Hz. But no matter which way we do it, the resulting spectrum has approximately the same resolution.
Appendix G Windows and Filters

G.1
Windows
The other problem in gure D.2(b) was the frequency content outside the signal frequency. It turns out that this is related to the limited number of samples that we have. The Fourier transform is really an integral from to , not a limited time. We can describe what we do by using the concept of a window. We will use a top hat window. This can be dened as
{
wT (t) =
1; T /2 < t < T /2 0; otherwise
(G.1)
We can now describe the nite time sampling of the real signal u(t) by multiplying with this window function, uT (t) = u(t)wT (t) (G.2) and in the frequency domain we nd uT (f ) = F[uT (t)] = F[u(t)wT (t)] Using eq. (C.11), we nd that

(G.3)
uT (f ) = u(f ) wT (f ) =
u(f f1 )wT (f1 )df1 =
u(f1 )wT (f f1 )df1 (G.4)
We see that our transformed signal gets contaminated with all other frequencies coming from the convolution with the transformed window function wT . This can be seen in the example with the cosine function in gure D.2. The t axis limits in eq. (G.1)) can be changed to suit the actual sampling. This is 349
350
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.5
APPENDIX G. WINDOWS AND FILTERS
Hanning Hamming Bartlett
0 t
0.5
Figure G.1: Examples of common window functions shown in gure D.2(a). The Fourier Transform of the top-hat window function dened in eq. (G.1) is the function sinc(x) = sin(x)/x. This function is shown in gure D.2(c) where the absolute value of the function is also shown as a dotted curve. We can see how the sinc function is convoluted by the original cosine frequency by doing a new FFT calculation with so-called zero-padding. This means that samples with the value zero after the original 32 sampled data. This will increase the resolution of the frequency. In MATLAB, zero-padding to a total record of 256 is simply done by the command uf=abs(fft(u,256)). The result of this command is shown in gure D.2(d). This result is very close the the analytic result of our sampling procedure. We see how the original frequency at 0.1 has been convolved with the sinc function. The points from gure D.2(b) are shown in gure D.2(d) as small green rings. We see that the courser calculation are samples of the more detailed curve. The lobes of the sinc function falls of as 1/f and therefore in the example of gure D.2 moves spectral energy from the real frequency into a large range of the total spectrum. There are several ways to x the window problem. The best way is to make the record length (the window length) longer. Of course, this requires planning before you do the experiment. Another way is to multiply the sampled data with a window function with no sharp corners. Examples of some common window functions are shown in gure G.1. It may seem strange to take perfectly good data and weight it. One eect of windows is to smooth out real peaks (reduce
G.2. FILTERS
351
resolution). There might also be bias eects, i.e. energy is put into the wrong frequencies. Even though weighting windows can solve problems, they can also create them, so they should therefore be used with care.
G.2
Filters
We can describe lters in the frequency domain in way similar to how we used windows in the time domain. We could dene a very simple lter (low-pass sharp cut-o lter) as { 1; |f | < fL wLP (f ) = (G.5) 0; otherwise We nd the new frequency distribution as uLP (f ) = u(f )wLP (f ) (G.6)
The ltered signal in the time-domain is then found using the inverse Fourier Transform on uLP (f ). If we apply such a lter to, for example, a square wave signal, we would get a result similar to what we see in gure C.1. If the cut-o frequency fL is a slightly above a frequency of 9 times the square wave period time, we would get the result shown in the last plot of gure C.1 (marked n = 9). We get some distortion of the signal in terms of ringing. Often this is not desirable and a sharp cut-o lter is therefore not a practical solution. Instead, a much smoother lter function should be used. Filters and other electronic component (e.g. ampliers) are often described by the relative change on a logarithmic scale. Traditionally the unit deciBel (dB) is used. This unit was originally used in acoustics, where it is a measure for the ratio between two sound intensity levels (measured in W/m2 ). If I2 is a reference level (e.g. the smallest sound intensity sensed by the human ear), the level of an intensity I1 is found in dB as In = 10 log I1 I2 (G.7)
From this denition, we see that 1 dB corresponds to a 25.9% increase in intensity level. If I1 is two times I2 we get In = 3.01 dB. In electronics, this correspond to dierences in power levels. Often we want instead to describe dierences in voltage levels. We can nd the corresponding ratio N by relating two voltages V1 and V2 to the ratio power levels the voltages will make in a resistor R, N = 10 log V12 /R V1 = 20log 2 V2 V2 (G.8)
For voltages we nd, that if V1 is two times V2 then N = 6.02 dB. Simple lters can be made with combinations of condensers (capitors), coils (inductors) and resistors. A lter that removes high frequencies is called a lowpass lter, and a lter that removes low frequencies is called a high-pass lter. A
352
APPENDIX G. WINDOWS AND FILTERS
Figure G.2: Simple rst order lters
G.2. FILTERS
353
Figure G.3: Amplitude characteristics (left) and phase characteristics (right) high-pass and a low-pass lter can be combined into a band-pass lter. The lters that can be described by rst order dierential equations are therefore called rstorder lters. A rst order low-pass lter rolls o at -6 dB/octave, a second order at -12 dB/oct, and so on. As illustrated in gure G.3, a lter can be described by characteristics of amplitude and phase of the output signal compare to input signal as a function of the frequency. A lter is characterized by a cut-o frequency (frequency where attenuation is 3 dB) and then the slope of the characteristic in dB/decade. As it can be seen, lters change not only the amplitude but also the phase of the output signal. Usually this is not a problem, but if two signals are to be compared after passing two dierent lters, care should be taken. More advanced lters with dierent slopes in dB/decade can be built. Filters that roll-o at -48 dB/octave or more are not uncommon, and can be very useful in preventing aliasing. Details are not covered here, but can be found in signal processing books.

Bill George Notes

Uploaded by

Copyright:

Available Formats

Bill George Notes

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bill George Notes

Uploaded by

Copyright:

Available Formats

Lectures in Turbulence for the 21st Century

4 3.5 3.6 3.7 4 The 4.1 4.2 4.3 4.4 4.5

Chapter 1 The Nature of Turbulence

1.2. WHAT IS TURBULENCE?

CHAPTER 1. THE NATURE OF TURBULENCE

1.3. WHY STUDY TURBULENCE?

Why study turbulence?

The cost of our ignorance

1.5. WHAT DO WE REALLY KNOW FOR SURE?

What do we really know for sure?

CHAPTER 1. THE NATURE OF TURBULENCE

Our personal adventure

1.7. A BRIEF OUTLINE

CHAPTER 1. THE NATURE OF TURBULENCE

Study questions for Chapter 1

Chapter 2 The Elements of Statistical Analysis

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

The Ensemble and Ensemble Averages

2.2. THE ENSEMBLE AND ENSEMBLE AVERAGES

Fluctuations about the mean

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

variance is dened as: var[x] (x )2 = [x X]2 =

2.2. THE ENSEMBLE AND ENSEMBLE AVERAGES

denoted by the symbol, x . Thus, x (var[x])1/2

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

is equal to zero if Bx is an even function.

The probability distribution

Fx (c) = P rob{x < c} =

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

Gaussian (or normal) distributions

Skewness and kurtosis

2.4. MULTIVARIATE RANDOM VARIABLES

Multivariate Random Variables

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

2.4. MULTIVARIATE RANDOM VARIABLES

Buv (c1 , c2 )dc1 dc2 = 1

Buv (c1 , c2 )dc2 = Bu (c1 )

where Bu is a function of c1 only. Property 5:

Buv (c1 , c2 )dc1 = Bv (c2 )

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

Figure 2.6: Surface representation of a joint probability density function.

cm cn Buv (c1 , c2 )dc1 dc2 1 2

(c1 U )m (c2 V )n Buv (c1 , c2 )dc1 dc2

The bi-variate normal (or Gaussian) distribution

Statistical independence and lack of correlation

(c1 U )Bu (c1 )dc1

(c2 V )Bv (c2 )dc2 (2.40)

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

Estimation from a Finite Number of Realizations

2.5. ESTIMATION FROM A FINITE NUMBER OF REALIZATIONS

Bias and convergence of estimators

(2.46) (2.47) (2.48)

CHAPTER 2. THE ELEMENTS OF STATISTICAL ANALYSIS

2.5. ESTIMATION FROM A FINITE NUMBER OF REALIZATIONS

N N 1 1 = xn X) N n=1 N n=1 N 1 = (xn X) N n=1

N N 1 (xn X)(xm X) N 2 n=1 m=1

Generalization to the estimator of any quantity

2.6. GENERALIZATION TO THE ESTIMATOR OF ANY QUANTITY

Problems for Chapter 2