Professional Documents
Culture Documents
by
Haralabos Marmanis
M. Sc., University of Illinois, 1996
Diploma, Aristotle University of Thessaloniki, 1994
Thesis
Submitted in partial fulllment of the requirements for
the Degree of Doctor of Philosophy
in the Division of Applied Mathematics at Brown University
May 2000
Abstract of \Analogy between the Electromagnetic and Hydrodynamic Equations: Appli-
cation to Turbulence," by Haralabos Marmanis, Ph.D., Brown University, May 2000
In this thesis we have accomplished the following tasks:
1. A new theory of turbulence is initiated, for the purpose of describing the dynamical
behavior of averaged
ow quantities in incompressible
uid
ows of high Reynolds
numbers. It is based on a new analogy between electromagnetism and turbulent
hydrodynamics. The central idea of the theory is the presentation of turbulent dy-
namics as an interplay between the vorticity eld (w = r u) and the Lamb vector
(l = w u). Two important concepts are introduced: a) the turbulent charge and
b) the turbulent current. This leads to a closed linear system of equations for the
averaged vorticity and the averaged Lamb vector. The averaged distributions of the
turbulent charge and the turbulent current appear as source terms in this system.
The premise of the theory is that the averaged distributions of the sources are the
most natural choice of turbulent quantities to be modeled. The general framework of
the theory is presented.
2. The case of homogeneous turbulence is studied next. Direct numerical simulations
of decaying homogeneous turbulence in a periodic box are conducted. The Fourier
coecients of the vorticity eld are recorded as functions of time, while turbulence
decays. These functions are projected onto the temporal eigenfunctions provided by
the theory. Three eigenfunctions suce to reproduce the time evolution of a Fourier
coecient, for nearly all the period of decay. The eigenvalues of the vorticity are
discrete, which suggests that the phenomenon of intermittency arises naturaly in this
context. The statistical properties of turbulent waves are inquired and shown to result
in a Kolmogorov type of spectrum.
3. Two cases of inhomogeneous turbulence are studied: Flow in a channel, and
ow in a
circular pipe. Direct numerical simulations are used to calculate the distribution of the
turbulent sources. The partial universality of the averaged velocity prole, from the
center of the duct to the wall, is inferred by the partial universality of the turbulent
current. This kind of universality gives excellent agreement with the experimental
data available for circular pipes, in a regime of Reynolds numbers that spans four
orders of magnitude.
We suggest that the proposed theory is most suitable for the study of
ows that are
unsteady and involve complex geometries. The implementation of the theory in a numerical
code will result in a powerful tool for the design of industrial applications. Aside from the
simplicity of the equations to be solved, the linearity of the equations suggests that a library
of cases can be built so that knowledge about one kind of geometry can be readily used for
another (more complex) kind of geometry, under the same
ow conditions.
c Copyright
by
Haralabos Marmanis
2000
This dissertation by Haralabos Marmanis is accepted in its present form by
the Division of Applied Mathematics as satisfying the
dissertation requirement for the degree of
Doctor of Philosophy
Date
GEORGE KARNIADAKIS, Director
Date
MARTIN MAXEY, Reader
Date
KATEPALLI R. SREENIVASAN, Reader
Date
ROBERT H. KRAICHNAN, Reader
Date
Peder J. Estrup
Dean of the Graduate School and Research
ii
The Vita of Haralabos Marmanis
Haralabos Marmanis was born in Athens (Greece) on January 19, 1970. He obtained a
diploma in civil engineering from the Aristotle University of Thessaloniki, and a M.Sc.
degree in theoretical and applied mechanics from the University of Illinois. A list of his
publications reads as follows:
2. Marmanis, H. 1998
The kinetic theory of point vortices
Proc. Roy. Soc. A , 454, 587-606
3. Marmanis, H. 1998
Analogy between the Navier-Stokes and Maxwell's equations:
Application to Turbulence.
Phys. Fluids 10 (6), 1428-1437
4. Marmanis H. 1999
Turbulence, electromagnetism and quantum mechanics: A common perspective.
In Photon: Old problems in light of new ideas, Ed. V.V. Dvoeglazov
Nova Science Publications
v
Acknowledgments
This work has started, in the summer of 1993, from the encouragement that the author
received by Dr. Robert H. Kraichnan. Since that time, Dr. Kraichnan has helped the
author with his perspicacious comments and his moral support. Hence, it is an honor for
the author to dedicate this work to him as a cordial devoir.
Although the initiation of this work was due to Dr. Kraichnan, the completion of it is
due to Professor George Karniadakis. Starting with the dawn of 1996, he has continuously
provided support and guidance that were necessary in order to deal with the labyrinth of
our modern academic life. Thus, in a state of gratication, the author wishes to express his
gratitude and to thank him wholeheartedly.
The vivid discussions and the congenial academic atmosphere in the division of Applied
Mathematics has played a catalytic role in the present work. Especially, Professor David
Gottlieb, Professor Martin Maxey, Professor Chi-Wang Shu, and Dr. Andrew Poje have
helped or advised the author, several times and in manifold ways. It is a pleasure for the
author to avow this support. Finally, the author would like to acknowledge and thank Mike
Kirby and the rest of our CFD group for many hours of fruitful and enjoyable confabulations.
Financial support, by the National Science Foundation under the grants CTS-9417520
and CTS-9619232 and by the Air Force grant AFOSR F49620 94-1-0313, is gratefully ac-
knowledged.
vi
Contents
1 Introduction 1
1.1 The Denition of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The basic eld variables: Vorticity and Lamb's vector . . . . . . . . . . . . 9
1.3 An outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Homogeneous Turbulence 30
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Direct numerical simulation and validation . . . . . . . . . . . . . . . . . . . 35
3.3 Signal analysis of the vorticity . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.4 Kolmogorov spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.5 Concluding remarks for homogeneous turbulence . . . . . . . . . . . . . . . 54
4 Inhomogeneous Turbulence 64
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 Turbulent duct
ows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
vii
4.2.1 The scaling of the turbulent current in channel
ows . . . . . . . . . 71
4.2.2 A candidate universality for circular pipes . . . . . . . . . . . . . . . 75
5 Concluding remarks 83
5.1 The motive and the rationale . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2 A summary of the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
viii
List of Tables
2.1 The correspondence between microscopic electromagnetism and unaveraged
turbulent hydrodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1 The above values describe the initial congurations and give the range of Re
for the four runs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 The coecients of the power laws and the coecients of the polynomial
ttings for the four runs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
ix
List of Figures
2.1 The distribution of the turbulent charge for a Burgers' vortex as a function of
the radial distance. We show four dierent distributions that correspond to
four cases with dierent kinematic viscosities but they are otherwise identical.
The square symbols (2) correspond to = 0:01, the triangles (4) correspond
to = 0:001, the circles () correspond to = 0:005, and the diamonds (3)
correspond to = 0:0001. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Comparison of the elds < u2r w > and < u2 > r W in a channel
ow
(Re = 5 000). The averaging is done along the streamwise direction. . . . . 25
3.1 For steady state
ows, the magnitudes of the turbulent sources in the central
region of ducts essentially vanish. Thus these regions are characterized as
homogeneous. Top: Channel
ow data 42], the center is located at y = 0
and the wall is located at y = 1. These data are averaged in the streamwise
and spanwise directions. Bottom: Circular pipe
ow data (present DNS), the
wall is at r = 1. These data are not averaged in the homogeneous directions. 59
x
3.2 The product of the maximum wavenumber times the Kolmogorov length scale
is shown as a function of time for all four runs. Values larger than one are
considered to re
ect an adequate spatial resolution. See Table 1, for reference
to the symbols used. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3 Three-dimensional structure of the turbulent charge at an intermediate stage
of the decay (Re = 20). The gure on the top represents the unaltered DNS
data, whereas the gure on the bottom represents the same data after a spa-
tial averaging has been performed. The lter length scale was approximately
equal to two Kolmogorov length scale units. The data were taken from run 4. 61
3.4 Kinetic energy versus t;1 . The symbols are data taken from run 4. The two
solid lines are two dierent ts to these data (a power law t and a second
order polynomial in powers of t;1 ) as indicated. . . . . . . . . . . . . . . . . 62
3.5 The vorticity Fourier coecient, with k = (1 1 0). The DNS signal (solid
line), the signal constructed by the rst eigenmode (2), by the rst two
eigenmodes (), by the rst three eigenmodes (). . . . . . . . . . . . . . . . 62
3.6 Comparison of the Fourier vorticity signals. The Fourier coecients for the
two wavevectors k = (1 1 0) and k = (1 0 1) have been set equal at the
beginning of the calculation. Their time evolution, according to M98, should
be the same. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
xi
4.2 The non-dimensional turbulent current J~z , as calculated by (4.6), versus y + .
The square symbols (2) denote Re = 180, the crosses (+) denote Re = 395,
the diamonds () denote Re = 590, and the solid line denotes the universal
curve given by equation (4.12). . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3 Averaged turbulent current versus distance from the wall. Comparison be-
tween DNS data of run 3 (symbols) and the model suggested by (4.17). . . 76
4.4 Averaged streamwise velocity (scaled with its value at the centerline) versus
distance from the wall (scaled with the radius of the pipe). The square
symbols (2) are data from Lekakis (Re=30912), the diamonds (3) are data
from Zagarola (Re=25817), and the () circles are the data from Nikuradze's
experiments (Re=26092). The solid line is the theoretical prediction. . . . . 77
4.5 Averaged streamwise velocity (scaled with its value at the centerline) versus
distance from the wall (scaled with the radius of the pipe). The diamonds
(3) are data from Zagarola (Re=439207), and the () circles are the data
from Nikuradze's experiments (Re=419421). The solid line is the theoretical
prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.6 Averaged streamwise velocity (scaled with its value at the centerline) versus
distance from the wall (scaled with the radius of the pipe). The diamonds
(3) are data from Zagarola (Re=1783953), and the () circles are the data
from Nikuradze's experiments (Re=1844000). The solid line is the theoretical
prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
xii
4.7 Averaged streamwise velocity (scaled with its value at the centerline) versus
distance from the wall (scaled with the radius of the pipe). The diamonds
(3) are data from Zagarola (Re=19903343), and the () circles are data from
Zagarola (Re=10300141). The solid line is the theoretical prediction. . . . . 81
4.8 This is the absolute value of the error in the calculated value of the wall
shear stress for ten Reynolds numbers that span four orders of magnitude.
The values are scattered but all of them are more than 1% accurate. . . . . 82
xiii
Chapter 1
Introduction
The turbulent motion of
uids is an eminent problem of science, not only for its numerous
engineering applications but also for its role as an archetype of non-equilibrium statistical
mechanics. It is considered the last unsolved problem of 20th century physics, and there is
a good reason for that. Fluid and plasma turbulence are omnipresent in nature, from the
very small to the very large scales of the universe, and naturally one may ponder on why is
turbulence such a prevalent phenomenon ? Does it contain elements that are deeply rooted
in the fundamental natural laws ? Is it possible that turbulent-like mechanisms are hidden
in other phenomena but rarely recognized as such ? These questions are very serious and
important but we will not deal with them herein. Instead, we will focus on the problem of
modeling turbulence in incompressible, viscous
ows.
The governing equations for these
ows are the Navier-Stokes equations, and the prop-
erty of incompressibility is expressed by the divergence-free character of the velocity eld.
The validity of these equations will not be questioned here. Although doubts have been
articulated, neither strongly persuasive arguments have been given nor progress emerged
from such hypotheses. The cited reviews, textbooks and treatises cover these issues satisfac-
1
2
torily. Our choice of incompressible viscous
ows at high Reynolds numbers is based on the
belief that this type of
ows is suciently broad to be of practical interest and adequately
general to include all the fundamental dynamics of motion. For example, the subsonic
ow
around a
ying object, and generally the
ows around vehicles moving in a gaseous or liquid
substance, can be conceived as incompressible, high Reynolds number
ows. The range of
scales that it covers is vast, and extends from the mixing of cream in a coee cup to oceanic,
atmospheric and even interstellar
ows. Nevertheless, it is true that a great number of im-
portant processes involved in applications, such as combustion, compressibility or chemical
reactions are excluded. This choice has been made for the following two reasons. Firstly,
the great challenge of high Reynolds number
ows is not the possible degree of perplexity of
a particular problem, but the inherent nonlinearity included in the Navier-Stokes equations.
Secondly, since we introduce new concepts and a novel point of view, it is reasonable to
apply the theory at its simplest form and examine whether it can be useful or not. If it is,
then an extension towards more elaborated cases can be undertaken.
Finally, we should emphasize that our main objective is the study of three-dimensional
turbulence, although the theory can be readily extended to the idealized concept of two-
dimensional turbulence. In contrast, the dynamics ensued by Burger's equation { often
referred to as one-dimensional turbulence or \Burgerlence" { can not be considered as
turbulence because the vorticity eld and the Lamb vector are fundamental variables in the
theory of meta
uid dynamics. Since there is not vorticity in one dimension, this kind of
dynamics will be excluded from farther consideration.
3
1.1 The Denition of the Problem
Let us consider an incompressible viscous
uid that occupies a given region of space. This
region may be either a domain enclosed by a prescribed boundary or an open domain that
is simply- or multiply-connected. Under the in
uence of any forcing, the
uid will be set in
motion. Energy will be drawn from the work done by the forcing, it will be redistributed
among dierent parts of the
uid, and eventually it will be dissipated to molecular heat.
Our problem is to describe these
uid motions that require a signicant agitation of the
uid
constituents, in order to redistribute and eventually dispose o the energy that is provided
from the forcing.
We consider our
uid as a continuum medium, rather than an enormous assembly of
molecules. This means that the beautiful kinetic theory of gases or other atomistic points of
view will not concern us. The ratio of the mean free path (between molecular collisions) to
the smallest length scales (subsumed in the
ow) is an extremely small number, and similar
arguments hold for the relevant time scales as well. Hence, the continuum hypothesis is
well justied 1], 2], 3], 4]. Subsequently, in order to describe the motion of the
uid
we need to establish a reference frame. There are merely two ways of describing the
uid
motion. The rst is the material description, which chronicles the history of each particle.
The second is the spatial description, which presents the state of motion at each spatial
location for all times. The rst description is called the \Lagrangian" description whereas
the second is called the \Eulerian" description (for a brief account of a relevant historical
anecdote see 5]). We will use the Eulerian description of
uid motion, unless otherwise
noted. This means that we are examining what happens at every point of our domain, at
any instant of time. Each point is associated with a three-dimensional vector x (given in
terms of our reference frame), and each instant of time is associated with the value of a real
4
variable t. The motion is completely determined if we know the velocity vector u(x t) and
the pressure p(x t), for all points in space and time.
The governing equations that these elds obey have an interesting history. In 1736, Euler
had indicated by inferences that his Mechanica was the rst part of a planned treatise in
six parts, the sixth of which was to be concerned with the motion of
uids. Although he
never completed his planned treatise, Euler gradually built his great memoirs (e.g. 6], 7],
8]) that described the motion of ideal
uids. Several of the basic ideas of
uid dynamics
can be found in the works of Euler's antecedents, except that their exposition is incomplete
and sometimes even obscure. Euler's work has certainly been the rst perspicuous and
perspicacious exposition of the foundations of
uid mechanics. In Fourier's words1 9]:
\On est parvenu a exprimer par des equations generales a dierences partielles
les conditions du mouvement des
uides. Cette decouverte, qui est un des plus
beaux resultats de la Geometrie moderne, est due a d'Alembert et a Euler : : :
Euler : : : donne ces equations sous une forme simple et distincte qui embrasse
tous les cas possibles, et il les demontre avec cette clarite admirable qui est le
caractere principal de tous ses ecrits."
ru = 0: (1.2)
Euler's work paved the way for Navier 12], and Stokes 13], who presented the dynamic
equation of motion for a viscous, or non-ideal,
uid. The modied equations (1.1), which
are called the Navier-Stokes equations, can be written as follows
!
@ u = ; (w u) ; r p + u2 + r2u (1.3)
@t 2
where is the coecient of kinematic viscosity. The vector product of the vorticity with
the velocity is the Lamb vector, and it will be denoted as
l(x t) w u :
The quantity in the parenthesis of the second term in the r.h.s. of (1.3) is the Bernoulli
energy function, or total head, and will be denoted as
!(x t) p + u2 :
2
@ w = ; r l + r2 w : (2.1)
@t
In addition, since vorticity is the curl of the velocity, we have that
rw = 0 : (2.2)
@w = ; r l ; r r w : (2.3)
@t
The above equations are valid locally, whether the
uid motion is turbulent or not, and
averaged forms of these equations should also be valid# as long as the averaging operator
commutes with spatial and temporal dierentiation. A nice property of equation (2.1) is
14
15
that it contains only the vorticity and the Lamb vector. So, if one knew the Lamb vector
then the solution of (2.1), with appropriate boundary and initial conditions, would be a
matter of solving a linear partial dierential equation. Nevertheless, this is again a closure
problem and the question of how to obtain the Lamb vector remains. One of the possibilities
would be to model the curl of the Lamb vector by some other quantity involving only the
vorticity. This is not very promising though, since by the denition of the Lamb vector it
is obvious that velocity gradients dierent from the vorticity emerge and no progress will
be made. However, notice that we can consider the Lamb vector as another dynamical
variable, besides vorticity, and envision equations (2.1) and (2.2) as an incomplete system
of four equations, where the evolution and the divergence of the vorticity are given, whereas
the evolution and the divergence of the Lamb vector are missing. Thus our goal will be
to obtain these relationships and arrive at a complete system where the evolution of one
eld involves the spatial changes of the other, while extra nonlinear terms are conceived
as their sources. Whenever the theory is applied in a particular case, these sources will be
considered as having been determined not from the values of the elds but from the external
conditions of the
ow, i.e. the geometry and the gross energetics.
Now, let us examine the divergence of the Lamb vector. By applying the divergence
operator on both sides of the Navier-Stokes, we get
ergo the divergence of the Lamb vector is equal to the \lumpiness" of the Bernoulli energy
function (the term lumpiness has been introduced by Morse & Feshbach 32]). It follows that
the same quantity is endowed with a clear physical meaning: it represents the tendency of
the energy to clump at some regions where its value is positive, and to evacuate some other
regions where its value is negative# the stronger the tendency, the greater the magnitude of
16
the divergence. In the case where the Laplacian is zero, the energy has no curvature at all
and its density arranges itself so as to average out the dierences imposed by the boundary
conditions. Notice that for an incompressible Newtonian
uid, the divergence of the Lamb
vector is the same for both the inviscid and the viscous case# this is evident from direct
calculation of the equations.
The above considerations lead naturally to the introduction of the turbulent charge
density n(x t), whose mathematical denition is given by
This new quantity can be written in terms of the velocity and the vorticity, by using a
familiar vector identity, i.e.
r l(x t) = u r w ; w w : (2.6)
This equation explicitly states that the turbulent charge density is identically zero for
irrotational
ows. Hence, as a property of rotational
ows only, it ts genuinely in the
category of desired
ow quantities set forth in the introduction. Nevertheless, its merit is
not simply its connection with rotational
ows, but rather the specic way in which the
vorticity appears in (2.6). For instance, think of the turbulent
ow in a pipe or in a channel
and use any proper averaging operator in (2.6). It can be shown that the average of the rst
part is equal to the average of the second part whenever an isotropic situation occurs, thus
the average turbulent charge density is zero in these regions. This is the case for the center
region of a pipe or a channel and consequently the distribution of the turbulent charge
in these cases should be concentrated on the boundaries. We will elaborate more on this
matter in the next chapters.
We will proceed now with the derivation of the equation that describes the evolution of
17
the Lamb vector. The starting point is the Navier-Stokes equations written in the form
@ l = ; @ 2 u ; r @ ! + r2 @ u : (2.8)
@t @t2 @t @t
The rst two parts on the r.h.s. of (2.8) can be characterized as inviscid contributions to the
evolution of the Lamb vector, in the sense that they are the only parts present in the inviscid
case. Whereas the third part can be considered as the viscous correction to it. Therefore,
we will rst write down the evolution of the Lamb vector as if the
uid was inviscid (i.e.
governed by the Euler equations), and then require that viscous eects will occur only
through the last part of (2.8). This is not an exact derivation based on the Navier-Stokes,
but it is physically plausible and reduces to the exact result in the case of zero viscosity.
Moreover, the exact result from the Navier-Stokes demands the modeling of a nonlinear
term who's physical meaning is not clear. For a theory of high Reynolds numbers, the
viscous corrections are not the dominant terms and their eects are of primary importance
in the very small scales which, as we shall see, can eventually be ltered out by an averaging
procedure.
The calculation of the inviscid parts is given in Appendix A, and the result is
@ l = u2 r w ; un ; r (u w)u
@t
; w r(! + u2) ; 2(l r)u : (2.9)
The rst part on the r.h.s. of the above equation resembles, apart from the u2 multiplicative
factor, the rst part that appears in the evolution equation of the vorticity. The rest of
the terms have no equivalent in the vorticity equation and can not be written only as
18
combinations of w or l. Thus we arrive at the concept of the second turbulent source,
namely, the turbulent current vector j(x t) whose mathematical denition is given by
If we add the viscous part, which results from the substitution of the Navier-Stokes in the
last part of (2.8), to the above result then we will get the following governing equation
@ l = u2 r w ; j + r r l (2.13)
@t
In conclusion, we have found a system of four equations (i.e. (2.3), (2.2), (2.5) and
(2.13)) that are valid locally at any part of the
uid to a certain degree of approximation,
as far as viscous corrections are concerned. These equations should be correct whether the
uid is turbulent or laminar, although in the latter cases the equations (2.5) and (2.13)
are nugatory. The derivation of the latter equations involved the introduction of two new
concepts: the turbulent charge density n(x t) and the turbulent current vector j(x t). These
quantities should be taken as an input and can not be determined by the theory, thus our
19
ability to obtain them by observation is of fundamental importance. Finally, we remark
that this system of equations is linear, therefore it does not produce new unknowns after
averaging.
rb = 0 (2.14)
@b = ;r e (2.15)
@t
r e = 4 (2.16)
@ e = c2 r b ; 4 i (2.17)
@t
where b(x t) is the microscopic magnetic eld, e(x t) is the microscopic electric eld, (x t)
is the microscopic charge density, and i(x t) is the microscopic current. A more economical
description of the elds is obtained by introducing the vector potential a(x t) and the scalar
potential (x t). It is easy to see that the substitutions
b = r a (2.18)
e = ; @@ta ; r (2.19)
yield (2.14) and (2.15) as identities. Table (2.2) gives the complete correspondence between
the variables of the two systems, when = 0.
20
The correspondence between the equations is given, in a synoptic way, as follows (recall
that = 0)
rw = 0 $ rb = 0
@w = ; r l $ @b = ;r e
@t @t
rl = n $ r e = 4
@ l = u2 r w ; j $ @@te = c2 r b ; 4 i :
@t
The two systems share the common feature of having their primary elds varying ex-
tremely rapidly in space and in time. One may argue that in electromagnetism the charge
and the current are given, whereas in our case they are part of the problem. In fact, this
is not so. The reason is that the spatial variations of the microscopic electromagnetic elds
occur over distances of the order of 10;8 cm or less, and their temporal
uctuations oc-
cur over periods ranging from 10;13 sec for nuclear vibrations to 10;17 sec for electronic
orbital motion. Macroscopic measuring devices average over intervals much larger than
these. Therefore, all the microscopic
uctuations are averaged out, giving relatively smooth
and slowly varying macroscopic quantities 33]. These macroscopic quantities are deter-
mined by experiments and that's why they are considered as an input in the macroscopic
21
Maxwell equations. The importance of this distinction between knowns and unknowns in
electromagnetism is worth of a small digression.
If we insist in making precise the electromagnetic sources, then we need to invoke con-
cepts and methods used in quantum electrodynamics. The resulting qualitative picture for
the electromagnetic sources in the latter representation is akin to the picture of turbulent
sources in the frame of the proposed theory, but with a very important dierence: The
turbulent sources can not have a point-like structure. Point particles have been adopted in
quantum electrodynamics from the classical theory, but they give in both cases inadmissible
singularities. Thus, the idea of extended particles (i.e. by delineating a compact support
for certain properties of the elds) is especially welcome.
Now, the plan for new turbulence theory is exactly the same as it is for electromag-
netism 33]. Since we have obtained the fundamental \microscopic" equations (2.3), (2.2),
(2.5), and (2.13), we will rst average all quantities over the space variables, by applying
a spatial ltering method as proposed by Russako 34]. The resulting average sources are
to be determined by experiment, actual or numerical, and be considered as an input in the
equations of the meta
uid dynamics. Let us start with the ltering method. The spatial
average of a function A(x t) with respect to a test function f (x) will be dened in terms
of a convolutional integral given by
Z
hA(x t)i = f (x0)A(x ; x0 t) d3x0
0 0
-0.02 -0.02
-0.06 -0.06
-0.08 -0.08
-0.1 -0.1
0 0.1 0.2 0.3 0.4 0.5
X
Figure 2.1: The distribution of the turbulent charge for a Burgers' vortex as a function
of the radial distance. We show four dierent distributions that correspond to four cases
with dierent kinematic viscosities but they are otherwise identical. The square symbols
(2) correspond to = 0:01, the triangles (4) correspond to = 0:001, the circles ()
correspond to = 0:005, and the diamonds (3) correspond to = 0:0001.
in the same gure, that for higher Reynolds numbers the width of the turbulent charge
distribution decreases, according to the decrease of the vortex radius. Thus,
ows that
consist of an ensemble of these vortices will have a vanishing turbulent charge distribution
after averaging over distances on the order of the Kolmogorov scale, and they will form an
appropriate physical setting of homogeneous turbulence.
We can now proceed with the averaging of the equations (2.3), (2.2), (2.5), and (2.13).
The large-scale Lamb and vorticity eld quantities L and W will be dened as the averages
of the small-scale elds l and w, i.e. L hl(x t)i W hw(x t)i. Hereafter, we will
make the convention of denoting the average of a quantity by capitalizing the variable's
name. Then, the two homogeneous equations (2.2) and (2.3), after ltering, become
rW = 0 (2.20)
24
@ W = ;r L ; r r W : (2.21)
@t
On the other hand, the ltered inhomogeneous equations (2.5) and (2.13), after averaging,
become
On the basis of classical arguments, this approximation is reasonable since the energy
length scales dier drastically from the dissipation scales, for suciently high Reynolds
25
I
1.0 10
8.57143
7.14286
5.71429
4.28572
0.5
2.85714
1.42857
0
-1.42857
-2.85714
0.0
-4.28571
-5.71429
-7.14286
-8.57143
-0.5 -10
-1.0
0.0 0.5 1.0 1.5 2.0
I1
1.0 10
8.57143
7.14286
5.71429
4.28572
0.5
2.85714
1.42857
0
-1.42857
-2.85714
0.0
-4.28571
-5.71429
-7.14286
-8.57143
-0.5 -10
-1.0
0.0 0.5 1.0 1.5 2.0
Figure 2.2: Comparison of the elds < u2 r w > and < u2 > r W in a channel
ow
(Re = 5 000). The averaging is done along the streamwise direction.
number
ows. Although a mathematical proof can not be given, evidence supporting the
validity of the above relation is provided by DNS results of homogeneous turbulence and of
turbulent channel
ows# see for example gure (2.2). This approximation implies that the
kinetic energy is uncorrelated to the Laplacian of the velocity but it does not require that
the kinetic energy is constant (neither in space nor in time). Once this approximation has
been made and introduced in the equations, all the
ows can be classied into two distinct
cases. The rst is the \homogeneous" case, where the turbulent sources vanish, and the
26
second is the \inhomogeneous" case, where the turbulent sources do not vanish.
The new equations in the homogeneous case read
rW = 0
@ W = ;r L ; r r W
@t
rL = 0
@ L = hu2ir W + r r L : (2.25)
@t
The rst of these equations merely states that the vorticity is the curl of a vector (i.e. the
velocity vector) whereas the third is a denition. These equations can be reduced in the
usual fashion into wave equations for the elds. In the homogeneous case, we take hu2i to be
equal to the square root of the total kinetic energy per unit mass and ltered volume. Then,
these equations can be solved both analytically and numerically for appropriate boundary
conditions# for instance, in a periodic cubic box.
In the case of inhomogeneous turbulence, the turbulent sources do not vanish, and our
ltered model equations have the following form
rW = 0
@ W = ;r L + r r W
@t
r L~ = N (x t)
@ L~ = hu2ir W
~ ; J(x t) + r r L : (2.26)
@t
The two new eld quantities L~ and W
~ are introduced for the sake of generality. They
represent the eect that the ltering of the turbulent sources introduces to the basic elds
L and W. It can be shown that their relation is given, in tensor notation, as follows
0 1
X @Q
L~ = L + 4 @P ; @x + : : :A
27
W~ = W ; 4 (M + : : :) :
the quantities P, M, Q (and similar higher order objects) represent the macroscopically
averaged turbulent charge dipole, vortex dipole, and turbulent charge quadrapole, and
higher moment densities of the particular turbulent
ow. These moments are zero in the
case of the homogeneous system resulting in the simplication of the equations, as discussed
above. The relations used to give the elds L~ and W
~ in terms of the basic elds L and
W will be referred to as the constitutive relations of the particular
ow conformation. The
simplest of all is a linear relation of the following type
L~ = L W
~ = W:
The constants and can be interpreted as the hydrodynamic analogues of the dielectric
and magnetic permeability constants, respectively. In the simplest cases they can be consid-
ered as scalars, whereas in general they will be second-order tensors re
ecting the present
anisotropies. The determination of constitutive relations between the associated tilde-eld
quantities and the elds L and W can be either a subject of data analysis or it can be
asserted a priori and tested a posteriori.
The eld quantities L~ and W
~ are not necessary for simple cases, such as
ows of a
homogeneous
uid in ducts or over a vehicle. However, they can be important in cases
where the
uid is made up of two or more constituents (e.g. a liquid lled with solid
particles or gas bubbles), cases where external body forces have many localized maxima in
an otherwise quiescent forcing background, or nally in cases where non-Newtonian
uids
with large macromolecules are considered. For, in these cases, local elds and sources
will be associated with the additive particulates or external forcing distribution and their
contribution will alter the eect of averaging# especially when the size of the particulates
28
(whether solid particles, gas bubbles or polymer macromolecules) or the ne-structure size
of the forcing is comparable or larger then the length scale associated with the spatial
averaging. It is possible that, for simple geometrical congurations, these eects will be
amenable to an explicit calculation. Although we will not pursue such a task herein, we
should remark that research in this area (that spans a very wide spectrum of engineering
applications) can be advantageous since the general framework of the theory will remain
unchanged# unlike the situations that would result from such modications for any current
turbulent model available.
The boundary conditions (B.C.) associated with the above PDE system can be deduced
from the familiar integral equivalents of the proposed equations. For example, the B.C. on
the wall of a channel, for the normal components of L and W, will be
W n^ = 0
L~ n^ = 4
where n^ is a unit normal to the surface of the wall. In words, we ensure that the normal
component of vorticity W is continuous and the discontinuity of the normal component of
L~ at any point is equal to 4 times the surface charge density
at that point.
For the tangential components, the B.C. will be
n^ L = 0
~ = p42 K
n^ W hu i
where K the surface current. It is understood that the surface current has only components
parallel to the surface at every point. The tangential component of L across the surface of
~ is discontinuous by an amount
the wall is continuous, while the tangential component of W
whose magnitude is equal to p4hu2 i times the magnitude of the surface current density and
29
whose direction is parallel to K n^ .
The model system (2.26) is a set of eight equations involving the components of the four
elds L, W, L~ and W
~ . When the constitutive relationships are dened, the description of
the turbulent sources renders the model equations a closed system of PDEs and essentially
achieves the transformation of the turbulence closure problem into a study of turbulent
sources for dierent geometries of interest. The most welcome features of the new equa-
tions, (2.25) and (2.26), are their linear (in terms of the vorticity and the Lamb vector)
structure and their resemblance to the equations that describe the electromagnetic eld in
vacuum. The former feature calls for an application of the superposition principle whereas
the latter can provide physical insight for many complicated cases of turbulent
ows (if
an already examined electromagnetic problem exists). The superposition principle will be
most appropriate for the construction of solutions in complicated geometries by combining
solutions that are obtained in simpler geometries. For example, let us suppose that a DNS
study can give us the turbulent charge distribution for certain geometries, say over a
at
or corrugated wall, and also the dependence of its magnitude on Reynolds number. Then
we can solve, only once, for these standardized cases and construct the solution of a more
general problem by making fair use of the superposition principle.
Chapter 3
Homogeneous Turbulence
Homogeneous turbulence is an intensively studied topic in turbulence theory, because the
theoretical analysis is facilitated signicantly by the assumptions that are associated with
this kind of
ows. In this Chapter, we dene homogeneous turbulence in a way that encom-
passes the traditional denition but it is based entirely on the nature of the newly dened
turbulent sources. We show that the conventional DNS of homogeneous turbulence are,
indeed, homogeneous in the new sense, and that for channel and circular pipe
ows the
region away from the walls can also be considered as being homogeneous. The new dual
symmetry of the proposed theory is stated explicitly. The DNS of homogeneous turbulence
are presented and their results are exploited for the analysis of the vorticity equation. The
eigenvalues of the latter are found to form a discrete set, which suggests that the phe-
nomenon of intermittency could be viewed as the manifestation of \jumps" between various
eigenstates. Lastly, the predicted waves are taken to form an ensemble of weakly interacting
waves. The theory of wave turbulence is invoked and it is shown that Kolmogorov solutions
can be obtained as universal solutions of the kinetic equation for the velocity correlator.
We will see taht this result presupposes a non-vanishing value of the kinematic viscosity.
30
31
3.1 Introduction
A turbulent velocity
ow eld is typically considered as a continuous random function of
position and time 37]. This means that, although the detailed properties of turbulent
velocity or pressure signals do not appear to be predictable, their statistical properties can
be analyzed and reproduced. Since the early investigations of G. I. Taylor, it was clear
that an assumption regarding the averaged quantities at one point would greatly simplify
the mathematical analysis. The use of this assumption gave birth to the classical denition
of homogeneous turbulence: A turbulent motion whose average properties are invariant
under spatial translation. This type of homogeneous turbulence is an idealistic situation,
merely a mathematical construction. It requires that the space is innite { if it weren't, a
translation by a distance equivalent to the size of the spatial extension of the
ow would
give dierent statistics { and thereby the kinetic energy is innite, as well. Nevertheless,
the concept is useful under the premise that
ows which occur in nite volumes do exhibit
invariance of their statistics for lengths that are smaller than a characteristic length of the
ow domain. Several theoretical and numerical aspects of this problem have been studied
extensively in the past (4], 3]). These works serve as a point of reference and validation
for many approaches on turbulence. In the last decades, the focus of the studies made in
homogeneous turbulence was primarily the investigation of the small-scale structure and the
exploitation of the ideas rst introduced by Kolmogorov 38], 39] and Obukhov 40]. As far
as applications are concerned, the long-term scope of the modern studies is the employment
of this knowledge to large eddy simulation (LES) schemes, and the improvement of our
understanding about \macroscopic" properties of turbulent
ows that are not homogeneous.
Today, there is not a generally accepted theory of turbulence but there is a common
belief that a complete description of the turbulent motion should somehow rely, at least
32
partially, on experimental input. Thus, the main issue appears to be the choice of the
quantities that should be given in a turbulence theory. In the previous chapter we provided
a general framework for the turbulent motion of incompressible
uids (hereafter referred to
as M98), in which the quantities to be prescribed as input have been dictated by the theory
itself and not through empirical observation. The basic equations of M98 predicate that
the averaged vorticity (W) and Lamb vector (L) satisfy the following system of equations
rW = 0
@ W = ;r L ; r r W
@t
rL = N
@ L = c2 r W ; J + r r L (3.1)
@t
where is the kinematic viscosity of the
uid, and c2 =< u u >.
In the previous chapter, we have shown that the system of equations (3.1) is derived
from the Navier-Stokes equations under two assumptions. The
rst assumption is that
the following relationship holds
The physical interpretation of this assumption is that the most energetic scales are well
separated from the dissipative scales, and consequently the two terms in the product are
uncorrelated. It does not require that the kinetic energy is constant# neither in space nor in
time. The assumption has been shown to be valid for a turbulent channel ow (see gure
(2.2)). The validity of (3.2) depends, of course, on the Reynolds number but the dependence
is favorable for our results. For, it is an experimental fact that as the Reynolds number
increases the energetic scales become well separated from the dissipative scales 41].
33
The second assumption is related to the viscous terms in the evolution equation of
the Lamb vector. It is essentially an extended dual symmetry that is imposed in the
fundamental equations. This dual symmetry presents the dynamical evolution of the Lamb
vector as the dual equation of the vorticity evolution equation. This can best be illustrated
in the homogeneous case, by setting c2 = 1 and rewriting the second and fourth equation
of (3.1) as follows
@ X = ;r (Y + r X) (3.3)
@t
where if X = W then Y = L and = , whereas if X = L then Y = ;W and =
; . Therefore, as a conclusion we can state the following: The proposed equations are
approximate in the viscous case, but they present a dual symmetry that coincides with the
usual electromagnetic duality principle for = 0
the latter is very important since in the
inviscid case the equations are exact.
We will adopt a dierent denition of homogeneity that stems quite naturally from the
theory itself. That is, the adjective `homogeneous' will refer to the vanishing of the averaged
turbulent sources. Mathematically speaking this means
and
J hun + r (u w)u + w r(! + u2) + 2(l r)u i = 0 : (3.5)
Despite the complex nonlinearities that occur in the expressions above, these conditions
have a clear physical meaning. They require that the averaged Bernoulli energy function
(i.e. h!i) arranges itself so as to average out the dierences imposed by the boundary
conditions. All signicant structures that are related to the dissipation of energy need to
be smaller than the lter length-scale, but there is no need for innite extent of the domain
34
under consideration, and therefore the kinetic energy is nite.
For instance, in the prototype case of shear turbulent
ow in a duct (e.g. channel
ow and pipe
ow), the central region of the duct can be considered homogeneous by
both denitions. This is well known in the classical case, where the adaptation of Taylor
hypothesis of `frozen convection' allows measurements that are taken from ducts (with a
grid of cylinders further upstream) to be used in the study of homogeneous and isotropic
turbulence. Nevertheless, under the new denition and from a purely mathematical point
of view this may not be obvious at all. The non-linearities in (2.10) are hard to interpret
outside the M98-frame. Thus, the claim that the two denitions should be compatible (in
such standard cases as the case of
ow in the central region of ducts) needs some evidence
of support. This evidence is presented in Figure 3.1, where we show the turbulent sources
for the case of a channel
ow 42] and the case of a pipe
ow at Re = 3280 (for more
details on this DNS, see the next chapter) and Re = 3750, respectively. The characteristic
length for the case of the channel is half the height of the channel, whereas for the case
of the pipe the characteristic length is equal to its radius. For both channel and pipe the
characteristic velocity is the centerline velocity. It should be clear from this gure that the
central region is practically of zero magnitude. Incidentally, we should emphasize that the
essential structure of the turbulent sources does not depend on the averaging
procedure. In order to prove this statement, we have averaged the turbulent sources in
the plot on the top of Figure 3.1, i.e. the channel data, along the homogeneous directions
(i.e. streamwise and spanwise) but we have not averaged the results of the pipe simulation
which are shown on the bottom of Figure 3.1. The striking similarity of their distributions
shows that the averaging merely subtracts the noise.
In turn, this suggests that the averaged velocity eld can, and in general will, have a
35
spatial variation while the turbulence is homogeneous# this is not possible in the classical
sense because that would destroy the invariance of the statistics for translations with length
comparable or larger than the lter length-scale. A turbulent
ow that is homogeneous in
the classical sense it is also considered homogeneous in the new sense, as long as the lter
macroscale is suciently large to encompass a large number of turbulent charges of both
signs. Therefore, the vanishing of the turbulent sources allows a wider class of ows to be
considered as homogeneous. For instance, consider the
ow motion that would result inside
a box, made of xed rigid walls, after a plane grid (that is initially parallel to one of the
sides) has been rapidly moved from one side to another, for a number of times. This
uid
motion will denitely be turbulent, for a suciently violent motion of the grid that will
stir the
uid and transmit to it a large portion of its energy. However, this motion will
not be homogeneous according to the old denition (due to the presence of the boundaries)
whereas it will be homogeneous according to the new denition. The same holds for the
uid motion that results in an otherwise ambient
uid due to a rapid oscillation of a grid
somewhere inside the
uid.
where N is the number of grid points per direction and i j m = 1 2 : : : N . The wavenum-
bers of the Fourier space are given by
ki =
ni 2L (3.7)
and
X
u(x t) = uF (k t) exp(j k x) : (3.9)
k
Since the velocity eld is real, it follows that the complex conjugates of its Fourier compo-
nents uF (k t) are equal to uF (;k t) and therefore only one half of the Fourier components
37
need to be calculated.
The time advancement algorithm is the sti&y stable scheme introduced by Karniadakis
et al. 46]. This semi-implicit scheme allows third-order accuracy in time, and it has been
proved to work better for long-time integrations. The initial conditions chosen for the
velocity eld were of two types. The rst, type (I), refers to elds that are created by a
random number generator and are endowed with a specic energy spectrum. The latter is
determined by four coecients (c1, c2, c3 , and c4) in the following manner
The second, type (II), refers to elds that consist of a collection of Burgers vortices with
their axes distributed uniformly along all three coordinate directions. The dierent type of
initial conditions allows us to compare the initial transients related to the arbitrariness of
the initial elds. This is important in determining the early evolution of the decay, since
this requires an unbiased origin of time. The time domain for our results starts when two
simulations with dierent type of initial conditions begin to exhibit the same behavior and
concomitantly both the energy and the enstrophy decay. These criteria are sucient to
ensure that the eects of the initial transients disappear. The author did not nd this issue
addressed in the literature, despite the numerous DNS studies that are related to freely
decaying homogeneous turbulence (see, for example, 47]).
We have conducted several DNS studies over a wide range of the parameter space.
Herein, we present four representative cases. Table 3.2 provides information relevant to the
conditions under which the four DNS have been made.
We note that the new code was tested, for the accuracy of its results, by comparison
with the code used by Ruetsch & Maxey 48]. Both codes gave identical results for the
same initial conditions and long-time integrations. The spatial resolution, measured by
38
Run 1 Run 2 Run 3 Run 4
Symbol 2 4 3
N 64 64 96 128
Type of IC I II I I
c1 0.8 | 0.8 0.02
c2 0.5 | 0.5 4.0
c3 -1.2 | -1.2 -1.5
c4 1.0 | 1.0 2.0
0.01 0.005 0.01 0.01
Re 46 { 13 60 { 17 67 { 16 61 { 11
Table 3.1: The above values describe the initial congurations and give the range of Re
for the four runs.
@ 2U = ;c2r W + J ; r r L (3.11)
@t2
;r @@t! ; r (;r L ; r r W) :
Hence, due to the cancelation of the third and fth term, we obtain the following equation
for the averaged velocity eld U
@ 2U = c2r2U + 2 r4 U + J (3.13)
@t2 t
which is a wave equation with a forcing term exclusively expressed in terms of the turbulent
sources. Now, according to our denition of homogeneity, we have N = 0 and J = 0, thus
we nally get
@ 2U = c2r2 U + 2r4U : (3.14)
@t2
It follows that the equation for the averaged vorticity eld is given by a similar equation
@ 2W = c2r2W + 2 r4W : (3.15)
@t2
Due to the linear form of these equations, each direction is uncoupled from the others and
we can therefore drop the vectorial notation.
Since we can take the Fourier transform of the above equations and obtain a single ODE
in time, for each Fourier coecient of vorticity, the solutions of this ODE can be directly
compared with the signals, i.e. time-histories, obtained by DNS. The Fourier transform of
equation (3.15) is
@ 2 Wk + (c2k2 ; 2k4 ) W = 0 (3.16)
@t2 k
where Wk is the Fourier image of any of the vorticity components with wavenumber k.
The kinematic viscosity can be considered as a constant, and thus the temporal behavior
42
of vorticity is essentially determined by the dependence of c on time. Equation (3.16),
2
c2 (t) / t;1
c2 (t) / t;2
c2 (t) / t2
The list is not exhaustive but rather indicative of what sort of functions give an exact
analytical solution.
On the other hand, the DNS results show that the decay of c2 , as a function of t given
by
c2 (t) = a0 + t + t2 (3.17)
represents well all our DNS data and it can be thought of as another tting form than the
power law models ( i.e. c2 (t) / t; ) that are found in the literature (see, for example, 47]
and references therein). This is shown clearly in Figure (3.4), where we plot the energy
versus t;1 , for run 4, and two tting curves to it. The rst is a power law type of decay,
whereas the second is a polynomial of second order (in powers of inverse time). It is clear
that the polynomial of second order is a better representation of the recorded values than
the power law. Table 3.3 reports the values of the coecients for each tting curve and
for all runs. As the polynomial order increases the tting becomes better but an analytical
43
Run 1 Run 2 Run 3 Run 4
Symbol 2 4 3
1.16 1.36 1.59 1.46
a0 -0.0375 -0.0671 -0.2511 -0.0754
1.03659 1.4010 3.0992 2.0748
-0.0241 -0.0098 -0.0468 0.1193
Table 3.2: The coecients of the power laws and the coecients of the polynomial ttings
for the four runs.
solution is not feasible anymore. This is to be expected, of course, since a polynomial tting
in inverse powers of time is essentially a Taylor series expansion around zero. We neglect
the coecient a0 from the subsequent analysis, since its eect can always be calculated
later, but we do report its value for completeness. The presence and the non-zero value of
a0 simply re
ects the fact that our calculations have a nite length# if they did not, then
c2(t) goes to zero as t goes to innity, and thereby the value of a0 becomes zero.
The exact values of and depend on the particular time interval that we examine. If
we x the time interval the values of these parameters are not the same for dierent runs#
a situation similar to the absence of an unequivocally accepted value of . Our simulations
are unique, in the sense that we include two dierent types of initial conditions. This allows
us to examine whether dierent initial conditions can aect the decay of c2 . In fact, we
found that the initial period of decay has an entirely dierent behavior for the two dierent
types of initial conditions. Here, we will not elaborate further on this point since we are
merely seeking a good representation of how c2 decays within a desirable interval of time.
The polynomial model of the energy decay is valid for a time interval that is broader than
the domain where the power law is valid.
We will now write equation (3.16) in the more abstract form
D Wk = 0 (3.18)
44
where the D operator is given by
@ 2 + (c2(t)k2 ; 2 k4) :
D = @t 2
If Wk is given at two time instances, then the D operator describes how the turbulent
ow quantities will decay in time. One initial condition is the initial value of each Fourier
coecient. The other condition involved simply requires that everything should decay to
zero for t ! 1. The D operator is a Schr%odinger operator, and it is well known that every
Schr%odinger problem is associated with a Sturm-Liouville problem 32]. This means that
the theory naturally provides a basis of temporal eigenfunctions, which can be used in the
description of the decay process. Thus, it is reasonable to express an arbitrary experimental
signal Wkexp in terms of the eigenfunctions of D and write
X
Wkexp = cnWk(n) (3.19)
n
where the Wk(n) are given as the solutions of
Without loss of generality, we can choose the following conditions for our problem
i i
Wk(n) (t=0)
= 0 Wk
(n)
(t!1)
=0:
where c2(t) has been assumed to take the form (3.17). Thus, the values of k and k are
related to the and , respectively, by the following formulas
k = k2 k = k2 :
45
Another special case of time dependency for c that has been observed experimentally (50]#
2
51]), is the so-called linear decay law. The linear law of decay appears as a special case of
our model, by setting = 0 in (3.17). We adopt this decay law only because, in the context
of our analysis, the eigenfunctions take on very simple expressions and provide the ground
for an exact analytical treatment.
A last simplication in (3.21) is the lumping of the eigenvalue with the viscous term,
i.e.
^ (kn) = (kn) + 2 k4 : (3.22)
The rst step towards the solution of (3.21) is a transformation of the dependent variable.
In particular, the idiosyncrasy of (3.21) is that we demand the solution to vanish at the end
of an interval where the points are singular. This problem is usually solved by the removal
of the asymptotic behavior of the solution at innity. If we let time to approach innity,
then (3.21) reads
# i
d2Wk(n) ; ^
(n) (n)
dt2 t!1 k k t!1 = 0 :
W (3.23)
Equation (3.23) has solutions that vanish at innity only when ^ (kn) > 0, so we will set (kn) =
(k(n))2 > 0, which according to (3.22) ensures that ^ (kn) > 0. On physical grounds, this is
quite reasonable. D is the operator that describes the decay process, and its eigenvalues
should have the same eect on the system as does the viscosity, at least for the large scale
motions. In that sense, the eigenvalues of D are a type of generalized viscosity for the
system, and the k 's dened above are the dynamic equivalent of the term k2 .
We come next to the already mentioned transformation. By keeping a consistent nota-
tion, we set
^k(n) = (^k(n))2 (3.24)