Proo. Energy Combust. Sci. 1985, Vol. 11, pp. i 19 192. 0360-1285/85 $0.00 + .

50
Printed in Great Brilain. All rights reserved. Copyright © 1985 Pergamon Press Lid

P D F M E T H O D S F O R TURBULENT REACTIVE FLOWS

S. B. POPE
Sibley School o] Mechanical and Aerospace Engineering, Cornell University, Ithaca, N Y 14853, U.S.A.

Abstract --The aim of the methods described is to calculate the properties of turbulent reactive flow fields.
At each point in the flow field, a complete statistical description of the state of the fluid is provided by the
velocity-composition joint pdf. This is the joint probability density function (pdf) of the three components
of velocity and of the composition variables (species mass fractions and enthaipy). The principal method
described is to solve a modelled transport equation for the velocity-composition joint pdf. For a variable-
density flow with arbitrarily complex and nonlinear reactions, it is remarkable that in this equation the
effects of convection, reaction, body forces and the mean pressure gradient appear exactly and so do not
have to be modelled. Even though the joint pdf is a function of many independent variables, its transport
equation can be solved by a Monte Carlo method for the inhomogeneous flows of practical interest. A
second method that is described briefly is to solve a modelled transport equation for the composition joint
pdf.
The objective of the paper is to provide a comprehensive and understandable account of the theoretical
foundations of the pdf approach.

CONTENTS
1. Introduction 120
1.1. Introduction and objectives 120
1.2. Pdf methods 121
1.3. Outline 122
2. Probability Functions 122
2.1. Turbulent mixing layer 122
2.2. Random variables 123
2.3. Distribution function 123
2.4. Probability density function 124
2.5. Expectation 126
2.6. Measurements 128
2.7. Joint pdf's 129
2.8. Conditional probabilities 132
2.9. Limitations of one-point pdf's 133
3. Velocity-Composition Joint Pdf 134
3.1. Conservation equations 134
3.2. Velocity and composition spaces 136
3.3. Velocity-composition joint pdf 136
3.4. Discrete representation 138
3.5. Joint pdftransport equation 141
3.5.1. Evaluation of the integral I 142
4. Lagrangian Description 143
4.1. Equivalent systems 143
4.2. Lagrangian equations 144
4.3. Lagrangian pdf 145
4.4. Relationship between Lagrangian and Eulerian pdf's 146
4.5. Deterministic system 147
4.6. Stochastic systems 150
4.7. Normalization, mean continuity and pressure 153
4.8. Summary 156
5. Modelling 156
5.1. Guiding principles 157
5.2. Scalar dissipation 158
5.2.1. Exact equations 158
5.2.2. Deterministic model 159
5.2.3. Particle-interaction model 159
5.2.4. Langevin model 162
5.2.5. Conclusion 163
5.3. Energy decay and the return-to-isotropy 163
5.3.1. Analysis of the stochastic reorientation model 166
5.4. Rapid pressure 166
5.4.1. Solution of the Poisson equation 168

119
120
5.5. Langevin model
5.6. Joint pdf models
5.7. General flows
5.7.1. Inhomogeneity
5.7.2. Intermittency
5.7.3. Dissipation rates
5.7.4. Variable density
5.7.5. Laminar flamelets
6. Solution Algorithm
6.1. Introduction
6.2. Discrete representation
6.3. Fractional steps
6.4. First fractional step
6.5. Second fractional step: stochastic mixing model
6.6. Third fractional step: convection and mean pressure gradient
6.7. Determination of means
6.8. Variants
7. Composition Joint Pdf
7.1. Composition joint pdf equation
7.2. Solution algorithm
8. Conclusions
8.1. Summary
8.2. Discussion
8.2.1. Limitations of one-point closures
8.2.2. Advantages of the pdfapproach
8.2.3. Model improvements
8.2.4. Computational considerations
Acknowledgements
References
I. INTRODUCTION
1.1. Introduction and Objectives
This paper describes the use of evolution equations
for probability density functions (pdf's) to calculate
the properties of turbulent reactive flows.
Most flows in the environment, in process industry,
and in engineering equipment are turbulent. When
these flows involve heat transfer, mass transfer or
reaction, turbulence usually plays a dominant role in
determining the rates of these processes. The evident
importance of turbulent flows involving mixing and
reaction has stimulated a wide variety of theoretical
approaches, as revealed in a number of reviews, collec-
tive works, and specialist meetings. 1-8 The aim of
these approaches is to calculate the important proper-
ties of the flow field. For inert flows, the turbulence
closure problem 9 makes this a difficult task and, for
reactive flows, the difficulty is compounded by non-
linear reaction rates and large density variations. The
pdf methods described here are remarkably successful
in alleviating these difficulties: in a variable-density
flow, the effects of convection, reaction, body forces
and the mean pressure gradient can be treated without
approximation. The only remaining closure problems
are due to the fluctuating pressure gradient and to
molecular transport.
A cursory examination of the work on pdf methods
over the past fifteen years may leave the following
three impressions:
(i) pdf methods have great potential, since (as
mentioned above) they overcome the most
important closure problems;
S.B. POPE
169
171
173
173
174
175
176
176
178
178
178
178
179
180
180
182
185
186
187
187
188
188
190
190
190
190
190
191
191
(ii) because joint pdf's are functions of many (maybe
8) independent variables, numerical solutions to
their evolution equations would be extremely
difficult and costly, if not impossible; and
(iii) the mathematical formalism associated with the
pdf equations makes them inaccessible to all but
a handful of specialists, and has added little to
our understanding of the physical processes.
The objectives of this work are:
(i) to make pdf methods understandable from both
mathematical and physical viewpoints;
(ii) to show that the methods overcome the most
important closure problems;
(iii) to present models for the processes that cannot
be treated exactly;
(iv) to describe a computationally-efficient solution
algorithm for the modelled pdf equation that
makes it feasible to calculate the properties of
turbulent reactive flows of practical interest; and
(v) to show that considerable insight is gained by
describing physical processes in the multidimen-
sional state space in which the pdf's are defined.
In comparison to conventional turbulence
models,~°'.~~ pdf methods derive their advantage from
their more complete representation of the turbulent
flow field. Consider a flow involving three scalars
~b~(x_,t), q~2(A,t) and q~3~A,0 - - t h e s e could be fuel mass
fraction, oxidant mass fraction and enthalpy. With a
two-equation turbulence model (e.g. the k-e model),
at each point the turbulent reactive flow is repre-
sented by eight quantities--the means of the three
velocity components, the means of the three scalars, k
and ¢. In the pdf method, on the other hand, the flow is
represented by the joint pdf of the three velocities and
 Pdf methods for turbulent flows
the three scalars. At each point this is a function of six
independent variables. Clearly, the six-dimensional
joint pdf contains much more information than the
eight quantities considered in the conventional
turbulence-model approach.
Probability density functions have been used in the
study of turbulent reactive flows for at least 35
years.l 2 In the last ten years, several models have been
developed in which pdf's play a central role. 13"~4 In
these models, a parametric form for the pdf is
assumed 15-1s in terms of its first and second
moments, for which transport equations are solved.
The methods described here are distinctly different in
that the shape of the pdf is not assumed, but it is
calculated from a modelled evolution equation for the
pdf.
The pdf methods are applicable to a wide class of
flows. While we restrict our attention to single-phase,
low-Mach number flows, no restrictions are placed on
the complexity of the thermochemistry--there can be
large density variations, and there can be nonlinear
reactions involviag many species. Pdf methods are
well-suited to flows with complex thermochemistry,
and most of the work on pdf's in the last ten years has
been applied to reactive flows. This work, too, is
concerned with reactive flow: but this should not hide
the fact that (compared to conventional turbulence
models) pdf methods have many advantages for inert
and constant-density flows.
1.2. P d f M e t h o d s
The method of solving a modelled pdf equation
started in 1969. Lundgren 19 derived, modelled, and
solved a transport equation for the joint pdf of
velocity. In this equation, convective transport
appears in closed form. In mean-flow closures 1° a
model is needed for the Reynolds stresses (which
transport momentum), and in Reynolds-stress
closures* 1,2o a model is needed for the triple velocity
correlations (which transport the Reynolds stresses).
These transport processes are usually modelled by
gradient diffusion. But in the velocity pdf equation
these processes do not have to be modelled and so the
gradient-diffusion assumption is avoided. Lundgren
proposed simple relaxation models for the unknown
terms, and obtained analytic solutions to the modelled
pdf equation for simple flows.
Dopazo and O'Brien 21-23 and Pope 24 derived,
modelled, and solved the transport equation for the
composition joint pdf--that is, the joint pdf of a set of
scalars (e.g. mass fractions and enthalpy) that describe
the thermochemical state of the fluid. This equation is
remarkable in that arbitrarily complicated reactions
can be treated without approximation. This is in
marked contrast to conventional turbulence models in
which the mean reaction rate can be determined only
in special circumstances--when the reaction rate is
linear or when it is either very fast or very slow
compared with the turbulent time scales. For a
121
general, nonlinear, finite-rate reaction, conventional
turbulence models reach an impasse.
The modelled composition joint pdf equation has
been solved for a variety of flows, demonstrating its
ability to handle nonlinear reactions. Janicka et
al. 25'26 and Nguyen and Pope 2~ studied hydrogen-
air turbulent diffusion flames; Pope 2s and McNutt 29
studied turbulent premixed flames; and Sherikar and
Chevray a° and Givi et al. al studied the N O - O 3
reaction in turbulent shear flows. There have also
been several solutions of the transport equation for
the pdf of a conserved scalar, a2- a6
While the composition-pdf equation overcomes the
closure problem associated with nonlinear reaction
rates, it does not address the problem of determining
the turbulent flow field. In most of the composition pdf
calculations referenced above, the standard k - e
turbulence model was used to determine the mean
velocity and turbulence fields. The mean velocity and
the turbulent diffusivity and frequency were then used
as inputs to the modelled transport equation for the
composition joint pdf. In this equation, turbulent
transport is generally modelled by gradient diffusion.
For the inhomogeneous flows of practical interest,
analytical solutions to the pdf transport equations are
out of the question, and numerical methods, are
required. Standard techniques such as finite-differ-
ences are severely limited because of the large dimen-
sionality of joint pdf's. For example in a statistically
two-dimensional flow, the velocity joint pdf and the
joint pdf of three compositions are each a function of
five independent variables and time. It can be
estimated a2 that each time step in a finite-difference
solution to these joint pdf equations requires of the
order 1011 computer operations (additions or multi-
plications). With current computing rates of order 106
operations per second, such requirements are exces-
sive if not prohibitive. Further, the computer require-
ments rise exponentially with the dimensionality of
the pdf.
For problems with a large number of independent
variables, Monte Carlo methods aT,as usually provide
a feasible alternative means of obtaining numerical
solutions. In 1980, Pope a9 devised a Monte Carlo
method to solve the composition joint pdf equation.
This can be viewed as an extension to the inhomo-
geneous case of the stochastic mixing models of
Spielman and Levenspiel 4° and Flagan and
Appleton. 41 In this case the computer requirements
rise only linearly with the dimensionality of the pdf--
the best that can be achieved. The Monte Carlo
method has been used to solve the equation for the
joint pdf of three compositions in a one-dimensional
flOW 28 and that of two compositions in a two-dimen-
sional flow. a°,a 1
The ability of the Monte Carlo method to handle
joint pdf's of large dimensionality opened the way to
combining the velocity and composition pdf methods
to obtain the advantages of both. This is achieved by
considering the transport equation for the joint pdf of
the velocities and compositions. In this equation, con-
I22
vection appears in closed form (as it does in the
velocity-pdf equation) and reaction appears in closed
form (as it does in the composition-pdf equation).
Even though the velocity-composition joint pdf is a
function of many independent variables, its evolution
equation can be solved by a Monte Carlo method. 42
The joint pdf equation has been solved for a plane
jet, 't2''.3 for a thermal wake, 4~ for a premixed flame, 4s
and for a diffusion flame. 46
The pdf's considered here are Eulerian--that is,
they are the pdf's of fluid properties at a fixed spatial
location. But the pdf equations can be derived,
modelled and solved by either Eulerian or Lagrangian
methods. In the works cited above, the Eulerian view
is generally adopted. Here we make use of both
approaches.
Frost 47'4s and Chung 49'5° have obtained pdf equa-
tions from a Lagrangian formulation by using the
Langevin equation to model the fluid particle motion.
(The Langevin equation sl was originally devised to
describe Brownian motion.) Calculations of turbulent
diffusion flames using these pdf equations have been
performed by Frost *s and by Bywater. 52
1.3. Outline
This paper is primarily concerned with the
evolution equation for the velocity-composition joint
pdf--its derivation, interpretations, modelling, and
solution. The objective is to provide a comprehensive
and understandable account of the theoretical
foundations of the approach.
The next section contains a review of the definition
and properties of the probability functions used in
subsequent sections. In Section 3, the governing con-
servation equations of turbulent reactive flow are
presented: the properties of the velocity-composition
joint pdf are examined, and its evolution equation is
derived. In Section 4 this equation is studied from a
Lagrangian viewpoint. This Lagrangian description
affords a clear physical interpretation of the equation
and provides the framework for the modelling
(Section 5) and for the solution algorithm (Section 6).
Three types of models are described: linear deter-
ministic models; stochastic particle-interaction
models; and, models based on the Langevin equation.
In Section 7 the composition joint pdf equation is
examined. The final section contains a summary of the
most important points, and a discussion on the attri-
butes and limitations of the pdf approach.
2. P R O B A B I L I T Y F U N C T I O N S
Considering the stochastic nature of turbulence, it is
remarkable how little probability theory is used in
most theoretical approaches to the subject (see, for
example, HinzeS3). In this section various probability
functions are defined and their properties outlined in
order to provide the theoretical foundation on which
the rest of the paper is built. The texts on probability
theory by Drake s4 and Gnedenko 5s and those more
S.B. POPE
X~X.~ eC
v-XI
Ub'Tb ,.
splitterplate 'J
Uo,T;
FIG. 2.1. Sketch of a turbulent mixing layer showing
locations of (a) fully turbulent flow, (b) intermittent flow. (c)
non-turbulent flow.
closely related to turbulence by Lumley s6 and
Panchev 57 can be used as general references, but these
do not contain all the necessary results.
The reader familiar with probability theory may
wish just to examine the use of delta functions (most
importantly in Eqs (2.67) and (2.150)) before passing
on to Section 3.
In the following development, the concepts of prob-
ability are linked to turbulence by reference to
measurements in a hypothetical experiment involving
a turbulent flow.
2.1. Turbulent Mixing Layer
Figure 2.1 is a sketch of the turbulent mixing layer
formed between two parallel streams of air. The axial
and lateral coordinates are x 1 and x2, and the width of
the apparatus is assumed to be infinite in the spanwise,
Xa, direction. The lower air stream has an axial
velocity U° and absolute temperature Ta, while the
upper stream has a higher velocity U b and a higher
temperature Tb. The figure shows the flow a long time
after its initiation, when the free stream properties (Ua,
Ta, Ub and Tb) no longer vary with time. However, we
are also interested in the temporal development of the
flow from a uniform, quiescent initial condition. This
initial state exists before the time t = 0 when the flow
is started in a repeatable way.
It is assumed that the temperature T(x, t) is always
in the range,
T a<~ T<%Tb. (2.1)
This assumption is valid if the initial and boundary
conditions satisfy Eq. (2.1), and if the Mach number
Ma is small. Specifically, we require
Tb- Ta
Ma 2 << (2.2)
Tb+ T ~"
The normalized temperature ~b(x, t) is defined by
~b(x, t) - ( T ( x , t ) - T . ) / ( T b - T ~ ) , (2.3)
and, in view of Eq. (2.1), ~ adopts values between zero
and unity:
0 ~< ~b ~< 1. (2.4)
In referring to measurements of the normalized
temperature 0(x_,t) and of the velocity U(x,t), we
 Pdf methods for turbulent flows
assume that the measurements are free from error, and
that the measurement process does not affect the flow.
2.2. Random Variables
In the mixing-layer experiment, at a particular time
t = t o and at a particular location within the turbulent
flow x = x_o, the value of q~(x_0,to) is measured. The
experiment is then repeated and a second measured
value of ~b(x0, to) is obtained. (Recall that the time t is
measured from the initiation of the flow.) Almost
certainly, the two measured values of q~ will be
different. This is the nature of turbulent flows and the
result may not be surprising. But, the source of the
difference between the two realizations of the flow
needs careful consideration.
In both cases the flow is governed by the same
deterministic equations with the same initial and
boundary conditions. Why then is the flow not
uniquely determined as the solution of these
equations? The answer is that the initial and boun-
dary conditions, though nominally the same, are not
identical. Small air movements before the start of the
experiment lead to slightly different initial conditions.
Mechanical disturbances to the apparatus (such as
vibrations) lead to slightly different boundary con-
ditions. Even the governing equations may be slightly
different because impurities in the air lead to different
fluid properties. These differences, however small, are
amplified by the turbulence so that after some time the
temperature and velocity fields in the two experiments
are completely different.
We conclude, then, that in any turbulent flow
experiment, the boundary and initial conditions
cannot be controlled (or even known) sufficiently for
the evolution of the flow to be determined. Conse-
quently, it is appropriate to treat any flow property as
a random variable. The temperature tpx(~, to) and the
axial velocity U l(xo, to) are simple examples of flow
properties that can be treated as random variables.
A probabilistic theory makes no attempt to predict
the value of a random variable (q~(x_o, to), for example)
in a particular experiment, since this value is indeter-
minable. Instead, the aim of the theory is to determine
the probabilities of the random variables adopting
specified values. The relationship between these prob-
abilities and measured values is discussed in Section
2.6.
2.3. Distribution Function
We take as an example of a random variable the
normalized temperature at ( x , t ) = (Xo,to) which is
denoted simply by ~, rather than by q~(Xo,to). When
the experiment is performed and the temperature is
measured, the value of 0 can be plotted as a point on
h ~. E =
FIG. 2.2. Sample space (qs-spacel for the random variable 0,
showing the sample point ff = 4x
123
A
,o I .I
0 1/3 2/3 I '
FIG. 2.3. Sample space showing the region corresponding to
the event A - I/3 ~< ~ < 2/3: • sample point for which A
occurs; O sample point for which A does not occur.
the real line. Figure 2.2 shows the real line as the axis
of the independent variable ~, with the experimental
point (or sample point) plotted at ~, = q~. The ~, axis
(or ~b-space) is called the sample space of the random
variable ~b. Since ~ is a normalized temperature, ~, is
an independent temperature variable, and so q;-space
is also called temperature space.
The significance of the sample space is that any
event determined by the random variable corresponds
to a region of the sample space. F o r example, let the
event A be
A = (1/3 ~< th < 2/3). (2.5)
The region of ~k-space corresponding to this event is,
simply, 1/3 ~< ~b < 2/3, see Fig. 2.3. If the sample point
~, = ~b falls within this region, then the event A
occurred; if not, A did not occur.
The probability of the occurrence of the event A (or
simply, the probability of A) is denoted by P(A). The
probability of an event is a real number between zero
and unity. If an event cannot occur (i.e. it is impos-
sible) then its probability is zero. If an event must
occur (i.e. it is certain) then its probability is one. For
example, consider the events
B = ~b > 2, (2.6)
C -: sint~ >~ 0. (2.7)
Now, since ~b can only take values in the range
0 ~< ~ ~< 1 (Eq. 2.4), it is clear that B is impossible and
that C is certain:
P(B) = 0, (2.8)
P(C) = 1. (2.9)
These events are sketched on Fig. 2.4.
Since every event can be associated with a region of
the sample space, the probability of the event is equal
to the probability of the random variable being within
that region. The probability for any region can be
determined from the distribution function: for the
random variable ~b, the distribution function F~(~,) is
defined by
F~(~,) --- P(~ < ~k). (2.10)
Thus, for example, for the event
D =- ~b < 1/4, (2.11)
B
C
J i i i.//.
0 I 2 4
FIG. 2.4. Sample space showing the regions corresponding
to the events B --- q5 > 2 and C ~ sin O ~>0.
124 S.B. POPE

function F,($):
)
*<% d
f,(q/) - ~-~ F,(~). (2.19)
I I I I

By integrating this equation between ~ and

Flo. 2.5. Sample space showing the region ~t ~< @ < ~a as q/2 (~/~ > q/1 ) and comparing the result with Eq. (2.14).
the difference between the regions ~ < @2 and @ < @~. we find that

its probability is £ '+ f+(t~)d0 = F,(~, 2) - F+(qq)

1
P(D) = F#(I/4). (2.12)
= P(~t ~< ~b < qJ+). (2.20)
Consider the event E
That is, the probability of ~b in a given region is equal
E -- ¢q ~< ~b < ~k2, (2.13) to the integral of the pdf over that region. In
where ¢/2 > ~kx. The probability of E can be deter- particular, for an infinitesimal region,
mined by noting that the region ~1 ~< ~ < ~2 is equal P(+ ~< ~b < + + d e ) = f + ( + ) d ~ . (2.21)
to the (infinite) region ~ < I//2 minus the (infinite)
region ~, < ~kl, see Fig. 2.5. Thus, A probability density function has three fundamen-
tal properties. Since F,(+) is a non-decreasing
P(E) = P ( ~ ~< ~b < ~'2) function of q/, its derivative J'~(~0) cannot be negative:
= P(~ < ~/2)-P(~b < ~bl) f~(~0)/> 0. (2.22)
= F~(~k2)- F~(~kl). (2.14) The event - oo ~< ~b < ~ is certain, and so from Eq.
The fundamental properties of the distribution (2.20), we obtain
function are readily,~educed. Since the event
( - m < ~b < oo) is certain, we obtain f ° f~(~)d 0 1. (2.23)
~r,

P(~ < - o~) = F ~ ( - o0) = 0, (2.15)

and
P(qb < o o ) = F~(oo)= 1.
(These properties hold for any random variable and
do not depend upon our knowledge of q~, e.g. Eq.
(2.4)). The probability of the event E, Eqs (2.13)-
(2.14), is
P(E) = F~(~,2)-F¢(~bt) >/0,
and hence
F,(~2)/> F,(~I), for ~k2 > ~k1.
Thus F~(~O) is a non-decreasing function of ~ that
increases from zero to one as q~ varies from - oo to oo.
An example of F¢(~) for the mixing-layer experi-
ment is sketched on Fig. 2.6. It may be noted that in
this case F~(0) = 0 and F~(1) = 1, by virtue of Eq.
(2.4).
As I~,) tends to infinity, F,(~,) tends monotonically to
a constant (0 or 1), and so its derivative f , ( $ ) tends to
zero:
(2.16)
f ~ ( - m) = f , ( o c ) = 0. (2.24)
Figure 2.7 shows the pdffq,(~k) corresponding to the
distribution function of Fig. 2.6.
In some treatments of probability theory, the pdf is
defined only if the distribution function is continuous.
(2.17) In the present context, this restriction is unnecessary,
but clearly the definition of the pdf as the derivative of
a discontinuous function requires further comment. A
(2.18) simple example of a discontinuous distribution
function arises when the random variable ~b always
adopts the constant value ~p = 4',. Then,
P(q~ < ~,,) = 0,
P(~b = ~b,) = 1,
P(4~ > 4,,) = 0, (2.25)

2.4. Probability Density Function and the corresponding distribution function is

F¢(¢) = 0, ¢ ~< 4,,,
The probability density function (pdf)f~(q/) of the
random variable ~b is the derivative of the distribution = 1, ~, > ~b,. (2.26)

f~(~,)
T%(W) T
0.5 2.0
q,
I a,
O 0.5 I.O o 0.5 ,'o
FIG. 2.6. Distribution function F,(O)against 0- FIG. 2.7. Probability density function J~(~O)against ~,.
 Pdf methods for turbulent flows 125

Discontinuous distribution functions such as Eq.
(2.26) can be expressed as Heaviside functions, and
then the corresponding pdf's are composed of Dirac
delta functions. The definitions and properties of these
generalized functions are given briefly here: a more
detailed and rigorous treatment can be found in Refs
58 and 59.
The Heaviside function (or unit step function)
H(y) is given by
n(y) O, y <~O,
=
magnitude b, and its derivative bS(y-c) is a delta
function at y = c of magnitude b. (By "delta function
of magnitude b" we mean a delta function multiplied
by a constant of magnitude b.) From Eq. (2.31) we
obtain,
f :oob 3(y-c)g(y)dy = by(c). (2.32)
The discontinuous distribution function Eq. 12.26)
can now be written as

= 1, y > 0. (2.27) F,(qJ) = H ( q J - 4~*), {2.33)

and the corresponding pdf is obtained by differ-
A better definition of H(y) is that, for any function
entiating both sides,
g(y) (that is integrable and continuous at the origin)
f~(~O) = 5(~b-~*). {2.34)
f:H(y)g(y)dy=f:g(y)dy. (2.28)
In turbulent flows, distribution functions are
The Dirac delta function (or unit impulse function) encountered that are sometimes continuous, and
6(y) is the derivative of the Heaviside function: sometimes discontinuous. F o r example, Fig. 2.8(a)
shows the continuous distribution function and pdf of
dH(y) 4) when the measuring location x_0 is chosen in the
6(y) = - - (2.29)
dY center of the turbulent flow (see Fig. 2.1). If, on the
It is zero everywhere except at y = 0 where it is other hand, x_o had been located away from the
infinite. The area under the delta function is unity turbulent flow in the high temperature free stream
since for any positive number a, (where 4) = 1), then the distribution function would
have been,
f~ b(y)dj, H(a)-H(-a) 1. (2.30) F,(qJ) = H(~k- 1). (2.35)
a

The delta function is best defined by its fundamental This is shown on Fig. 2.8(c) with the corresponding
{sifting) property: pdf which is a delta function at ~O= 1. At the edge of
the mixing layer the flow is intermittent--sometimes
/: 5(y)g(y)dy = g(0). (2.31) turbulent, sometimes non-turbulent with ~ = 1. The
oc corresponding distribution, sketched on Fig. 2.8(b), is
The quantity bH(y-c) is a step at y = c of continuous up to qJ = 1, where there is a j u m p of
magnitude 0.35. The pdf is also continuous up to
= 1 where there is a delta function of magnitude
0.35.
,oyo /.,,/i. The probability of 4~ adopting the particular value
q~, can be deduced from Eq. (2.17):

0.5+2.0 II1!/\\\\ P(~b = 4),) = lim P(q~, ~< 4~ < q~,+AqJ)

A~0
/'°' . y , '.. = lim {F,(~b,+Aq,)-F,(q~,)].
0 0.5 z!o AO~0
(2.36)
%(q,)t f¢(~,)
1"014"0 If the distribution function F , ( ~ ) has a step of magni-
tude b at 0 = 4~, and, consequently, the pdff~(qJ) has
o.sTzo // a delta function of magnitude b at q~ = ~b,, then we
obtain
(b)

o 0.5 1.0 P(t~ = q~,) = b. (2.37)

1.0I If, on the other hand, F~(qJ) is continuous in the neigh-

borhood of 4~,, then P(q~ = 4),) is zero. The event
q~ = 4~, is not impossible, but it has probability zero
0.5T2.0 since it is just one of an infinite number of possible
(c) q,
events. Similarly, the event 4 ) ~ 4), has probability
one, but it is not certain.
0 01.5 1.0 |
We have taken as our example of a random variable
FIG. 2.8. Distribution functions and pdrs in a turbulent the normalized temperature q~. Consider now the
flow: (al fully turbulent region: (bt intermittent region; (c) temperature T measured in degrees Kelvin and the
non-turbulent region. same temperature 0 measured in degrees Rankine.
126 S.B. PoPE

With ? and 0 being the sample space variables corre- and

sponding to T and 0 we have ~F,(~, x_,t)
f , ( ¢ ; x_,t) -= c¢ (2.50)
9
0 = ~ T, (2.38)
A semi-colon is used to indicate that ;'~ is a density
and with respect to ~b (to the left of the semi-colon).
A flow is said to be statistically stationary if all
9
0 = ~ ?. (2.39) probabilities are independent of time. Thus for a
statistically stationary flow,
The event T < ? i s identical to the event 0 < 0 and so OF, (~, x, t)
(in an obvious notation) the distribution functions are = 0, (2.51)
t~t
equal:
and
Fr( ~ = Fo(O). (2.40)
= 0. (2.52)
Differentiating both sides with respect to ? we obtain Ot
dFr(? ) dO dF0(0 ) 9 If there is a coordinate system (~t, £2, xa) in which all
d? =fr(T) = d ? dO ='5 fe(O)" (2.41) probabilities are independent of x3, then the flow is
said to be statistically two-dimensional. For such a
Since 1 degree Kelvin corresponds to 9/5 degrees flOW,
Rankine, the probability per degree Kelvin, f r ( ~ , is
9/5 times the probability per degree Rankine, f0(0). OF,(~, xt, Y~2,x3, t)
= O, (2.53)
Note that f r and f0 have dimensions of (K)- 1 and
(R)-t respectively. (It is because of these observations and
that pdf's earn the designation "density".)
0f,(~; Yt,ff2,:~3, t) = 0. (2.54)
For the more general change of variables
0 = @(T) and 0 --- O(?), (2.42) Similarly, if all probabilities are independent of.'~2 and
2~, the flow is statistically one-dimensional. If all prob-
where O is a monotonically increasing function, we abilities are independent of position (yl,y2 and -~3)
obtain similar results: then the flow is statistically homogeneous.
Fr(?) = Fo(O), (2.43) By most definitions, all turbulent flows are three-
and dimensional and time dependent. But many flows of
interest are statistically stationary and many are
dO(?) 0 statistically one- or two-dimensional. The turbulent
Jr(?) = ~f0( )- (2.44)
mixing layer sketched in Fig. 2.1 is statistically two-
For O(T), being a monotonically decreasing function dimensional and, for large times, it can be expected to
of T, the event T < ? i s identical to the event 0/> 0 = be statistically stationary.
O(?). Thus
2.5. Expectation
P(T<?)=Pt01>0)=I-P(0<0), (2.45)
and so Most readers will be familiar with time averaging, a
procedure that is useful only in statistically-stationary
Fr(? ) = 1 - Fo(O). (2.46) flows. For the normalized temperature q~(x, t), the time
Differentiating both sides with respect to ? yields average (~b(x_,t))At is defined by

dO(?) r (0) (2.47)

1 I
(th(x_,t))at = ~-~ t+tu ~b(x~z)dz, (2.55)
fr(?)=- dT J® •
where At is a specified time interval. The time average
Since, in this case d O ( ? ) / d ? is negative, Eqs (2.44)
and (2.47) can be rewritten in the common form is itself a random variable. But if the time interval At is
sufficiently large, the measured value of (~b(x_,t))at
dO(?) may vary little in successive repetitions of the experi-
fT(?) = ~ fo(O). (2.48)
ment. In that case the measured time average is an
The random variable ~bthat we have been consider- approximation to a far more fundamental quantity--
the expectation of ~(x, t).
ing is the normalized temperature at the particular
In this sub-section the definition and properties of
location (x_,t) = (X_o,t o). The distribution function of q~
the expectation of a random variable are described. In
is F,(~b), and its pdf is f,~(0). Clearly, the normalized
the next sub-section the relationship between expecta-
"tempexature at any location is a random variable, and
consequently the distribution function and pdf are tions and measurable quantities (such as time
averages) is examined.
defined for all x and t:
The mathematical expectation of the random
F,(~O,x,t) - P(~b(x~r) < ~,), (2.49) variable ~b is denoted by E(~b) or ( 4 ) and is defined
 Pdf methods for turbulent flows
by,
E(4,) = <4,> - J_~ 0f~(0)d0. (2.56)
The terms mathematical expectation, expectation,
expected value, mean value and mean, are all synony-
mous. The mean <4,> is the center of the pdf, in that
the first moment about <4,> is zero,
A function of a random variable is itself a random
variable. For example, in the mixing layer between
two air streams of different temperatures, the density
p(4,) is a function of the normalized temperature. In
general, let Q(4') be a function* of the random variable
4': then the expectation of Q is defined by
(Q(4,)) = Q(0)f,(0)d0. (2.58)
oo
(It may be seen that the definition of (4,), Eq. (2.56), is
a particular case of this more general definition.) In
the infinitesimal region 0 ~< 4, < 0 + dO, the value of
Q (4,) is Q (0) and the probability is f~ (0) dO. Thus the
mean ( Q ) is the probability-weighted integral of all
possible values of Q.
After the mean, the second most important
numerical characteristic of a random variable is its
variance. The variance of 4, is denoted by D(4,) and is
defined by
D(4,) = E([4,- (4,)]2)
=
I_ ( 0 - (4,))2f,(0) d0.
The random variable 4, can be decomposed into its
(2.59)
mean (4,) and the fluctuation 4,',
4, = (4,)+4,', (2.60)
where,
4 , ' = 4 , - (4,). (2.61)
The variance is the mean of the square of the fluctua-
tion
D(4,) = (4,,2). (2.62)
The standard deviation, which is the square root of
the variance, is a measure of the width of the prob-
ability density function.
The variance is the second central moment of 4,.
The ruth central moment is defined by
p,, = ( 4 , " ) = t x (~k-(4,))mf~(0)d0. (2.63)
J - oc
It is evident from Eq. (2.63) that all the moments of 4,
can be determined from the pdf: but, in general, the
* Here and henceforth, we assume that functions such as
Q(q~) have whatever properties are required for the integrals
such as Eq. 12.58) to exist. In general these properties are
difficult to specify. If the pdf is continuous it is sufficient that
Q(~) be integrable. If the distribution function F,(~O) con-
tains discontinuities, then Q{q~)must be continuous at all
points of discontinuity of F~,(~ ).
127
pdf cannot be determined from a finite number of
moments.
The expectation ( Q ) of the random variable Q(4,) is
defined only if the integral
f ~ ,Q(O)If(O)d~
exists. For the pdf's considered here, it is certain that
the means and variances exist, and it is highly likely
that all the central moments exist. It is equally likely
that the expectation of rapidly increasing functions--
such as exp(exp(4,))--do not exist (for unbounded
random variables).
It proves to be very useful to identify F,{¢) and
f , ( 0 ) as the expectations of H(q, - 4,) and 6(q, - 4,). Let
G be the event
G -- 4, < 0. (2.64)
The probability of G is, simply,
P(G) = P(4, < 0) = F~(0). (2.65)
If the event G occurs, then the Heaviside function
H (0 - 4,) is unity, and if G does not occur, H (~b- 4, ) is
zero. Thus the expected value of H(~b -4,) is
<H(~O-4,)> = 1 . P ( G ) + 0 . (1 - P ( G ) )
= F~(0). (2.66)
Differentiating both sides with respect to O we obtain
( 6 ( 0 - 4,)) = f~(0). (2.67)
This property--that the pdf is the expected value of
the delta function--is of paramount importance in pdf
methods. Indeed in some treatments of the subject
(e.g. Ref. 60) it is taken as the starting point: b(qJ-4,)
is defined as the "fine grained" pdf (i.e. the pdf in one
realization of the flow), and then the pdf f,(~O) is
defined as ( 6 ( 0 - 4,)). Both approaches or viewpoints
lead to the same results. (It may also be noted that Eqs
(2.66) and (2.67) could also be obtained by substi-
tuting H ( 0 - 4 , ) and 6(~0-4,) for Q(0)in Eq. (2.58)--
provided that f~(0) is continuous. The above
development is to be preferred since it places no
restrictions on the pdf.)
An important result is that differentiation (with
respect to x_or t) and taking the mean, commute: that
is,
= ~x i , (2.68)
and
(04, / ~3(4,) (2.69)
These properties follow trivially from the definition of
a derivative:
= lim {(4,(t+~t))-(4,(t))} = ?(ep)
/its0 ~t
(2.70)
128 S.B. POPE

2.6. M e a s u r e m e n t s (iii) For large N, with probability 0.68, the error

From a theoretical viewpoint, the definition of the
mean <4~> as the mathematical expectation E(q~) is,
without doubt, superior to other possible choices.
Other possible definitions of the mean are: the
ensemble average (for flows that can be duplicated);
the realization average (for flows that can be re-
peated); the time average (for statistically-stationary
flows); the spatial average (for flows that are statisti-
cally homogeneous in one or more coordinate direc-
tions); and the phase average (for flows that are
statistically periodic). Of all the possible choices, only
the mathematical expectation is defined for all flows.
Consequently, only a theory based on this definition is
generally applicable. However, unlike the other
averages, the mathematical expectation cannot be
measured. We need, therefore, to consider the rela-
tionship between measurable quantities and the
mathematical expectation.
Consider the ensemble average of the random vari-
able q~ = tk(x_o,to) for the turbulent mixing-layer ex-
periment. A large number N of (nominally) identical
apparatus are constructed, andthe experiment is per-
formed on each one with (nominally) the same initial
and boundary conditions. For the nth apparatus, the
measured value of ~b is denoted by ~bt"). Then the
ensemble average <~b>Nof 4, is defined by
1 N
<4~>N -= ~- ~ ~b("). (2.71)
tl=[
The ensemble average <4~>N is clearly an estimate of
the mean <~b>since, for any N,
E(<~b>N) = <4,>. (2.72)
For large N, the Central Limit Theorem reveals the
relationship between <4~>N and <q~>. We define the
normalized error e by
= (<~b>u- <4~>)/~, (2.73)
where the standard error eu is defined by
e N =- t r l v / N , (2.74)
I<~b>.v- <qS>l is less than the standard error eN,
ex = (<qY2>AN)t:2. (2.77)
Thus, although <~> cannot be measured with
absolute certainty, for sufficiently large N, the
ensemble average <4~>Nprovides as good an approxi-
mation as may be required. It may be noted from (iii)
that the likely error is proportional to the standard
deviation ~ and decreases with the square root
of N. In order to halve the error, N must be increased
by a factor of four.
Similar results are obtained for the realization
average and, with some additional assumptions, for
the time, spatial and periodic averages. In practice,
then, the theory predicts values of the mathematical
expectation which can legitimately be compared with
experimentally determined averages. The ensemble
average of Q(~b), a function of 4', is defined by,
1 N
<Q(~b)>u - ~- ~ Q(4~"~). (2.78)
=1
The relationship between <Q>:¢ and <Q> is precisely
the same as that between <q~>Nand <q~>. In particular,
E(<Q> N) = (Q>, (2.79)
for any N >/1.
By analogy to Eqs (2.66) and (2.67) the ensemble-
averaged distribution function F,u(~O) and pdf f~u(~b)
are defined by
1 ~ H(~O_~b(.)),
(2.80)
and
l, N
f~N(~) = < 6 ( ~ - ¢ ) > N = ~ . ~ _ 6(~0-¢'"~).
(2.81)
Again, for any N/> 1, the expectations of these
ensemble averages are

and a is the standard deviation of tk, <qY2> t/2. Then as E{F~N(~) } = F,(~), (2.82)
N tends to infinity, according to the Central Limit and
Theorem, ss the pdf of e, f~ (~), tends to the standard-
ized normal distribution E{f~N(~b)} = f,(~k). (2.83)
~2 / 1 2\ These definitions are consistent since the ensemble
f~(*) = (2x)- : e x p ~ - ~ * ) . (2.75)
average of any function of 4, can be recovered without
error:
The following three results are readily deduced
from Eq. (2.75): 1 N
(i) Almost certainly (i.e. with probability one), as N
tends to infinity, <~b>Ntends to <q~>.
1 N
(ii) A finite N suffices to reduce the error = ~ ~ Q(~b("))= <Q(q~)>N.
I<~b>~- <~b>l below any specified positive limit n=|

with any required probability (less than unity). (2.84)

That is, for any a > 0 and p < 1, there is a finite
M such that To illustrate the difference between ensemble-
averaged distributions and their continuous counter-
P(l<4'>s-<4'>l < a) >I p, (2.76) parts, Fig. 2.9 shows distributions for the Gaussian or
for N > M. normal pdf,
 Pdf methods for turbulent flows
1.0 .....
(0)
0.5
0 ...... -'~ , q'
0.3 0.5 017
f~(~)
I
4.0 /,.--,,
(b) // "\\
o ..... 0.3
II I[l'"--;..J
0.5 0.7
Fro. 2.9. (a) Distribution functions against ~: dashed line
F,(~,): solid line FON(#). (b) PDF's against @: dashed line
f,($); vertical lines (representing delta functions of magni-
tude 0.05) .f,~ (¢).
f,(@) = (27~<~b'2>)- I/2exp(-½(~ -- <~>)2/<fb'2>),
with <4)> = 0.5 and <q5'2> = 0.01. On Fig. 2.9a, the
dashed line is the distribution function F,(~)
(obtained by integrating Eq. (2.85)), and the solid line
is F,~(0) for N = 20. (The 20 samples ~") are
normally-distributed random numbers with mean 0.5
and variance 0.01.) It may be seen that while F,(0) is
continuous, FoN(O ) is a piece-wise constant with N
points of discontinuity, On Fig. 2.9b, the dashed line is
the pdf f,(0), Eq. (2.85), and the delta functions (each
of magnitude 1/N= 1/20) form the ensemble-
averaged pdf f,N(0).
The definition of the ensemble-averaged pdf is saris-
factory since it possesses the required mathematical
properties, Eqs (2.83-84). But its relationship to the
pdf requires further comment: they are distinctly
different quantities in that f,u(O) is composed of delta
functions whereas f , ( 0 ) may be continuous. The rela-
tionship between f,N (0) and f , (0) can be understood
as follows.
Each value q~4,~is a sample of the random variable
~, and each location 0 = ~bt"~ is a sample point in
q/-space. Let m(0, A0) denote the number of sample
points in the region 0 ~ 4) < 0 + A 0 . The normalized
sample point number density is defined by
.. (,.(0.A0)~
ffN(O) -- um - -
and is the fraction of sample points per unit distance
along the O-axis. Clearly, f~N(O) is zero except at the
sample points where there is a delta function of magni-
tude 1/N. In other words, it is equal to the ensemble-
averaged pdf:
fgN(~) = f,u(O).
Thus the ensemble-averaged pdf is the normalized
sample point number density in 0-space.
As N tends to infinity, the ensemble-averaged pdf
.f,s, does not converge to f,, since f , s is still composed
129
of delta functions. The ensemble-averaged distribu-
,
tion function FN, is composed of steps and, even as N
tends to infinity, it does not become differentiable.
Nonetheless, for any continuous function Q(05), the
ensemble-average <Q(~b)>N almost certainly con-
verges to <Q($)>; and, for any N >~ 1, the expectation
o f f N, is f , .
:
It may be noted that the expectation of Eq. (2.87)
E {f**N(0) } = f,(0), (2.88)
shows that the pdf is the expectation of the normalized
sample point number density.
t
These considerations are important for the theory
developed below, where pdf's are represented in-
directly by ensembles of sample points. The principal
conclusions are that the pdf can be regarded as the
expectation of the normalized sample point density,
which is equal to the ensemble-averaged pdf. As the
size of the ensemble N tends to infinity, all ensemble
averages <Q>u almost certainly tend to the means
<Q>.
(2.85)
2.7. Joint Pdf's
For the mixing-layer experiment, the random
variable that we have been considering is the
normalized temperature at a particular location, 4~ =
$(x_o, to). We now consider a second random variable
U, which is the normalized axial velocity at the same
location:
U - (Ul(xo, to)--U,)/(Ub--Uo). (2.89)
When the steady state is reached, in the low-speed free
stream U and 4' are zero, and in the high-speed stream
U and $ are unity. The value of 4>is bounded by zero
and unity; but the same is not the case for the velocity,
since pressure fluctuations can cause U to exceed these
bounds. (Experimental data 61 show that U does
indeed exceed these bounds.)
A sample space (temperature space, or 0-space) has
been associated with the random variable q~, so that
any value of 4' can be plotted as a sample point 0 = $.
Similarly, velocity space of V-space is introduced so
that any value of U can be plotted as the sample point
V=U.
The distribution function F.(V) and the pdf f,(V)
are defined by,
F,(V) = P(U < V), (2.90)
(2.86)
and
dF~(V)
f~(V) = dV (2.91)
These definitions and the properties of the functions
are exactly analogous to those for 4', The mean
(mathematical expectation) of U is
(2.87)
<U> = [~ Vf.(V)dV, (2.92)
and the mean of a function* of U, Q(U) is
* See previousfootnote.
130 !ti.B. POPE

<Q(U)>=

The variance of U is
s m
-m
QU'l.L(Vv~ (2.93)
already been noted, the joint pdf cannot (in general)
be determined from the marginal pdf’s.
If Q(U,qb) is a function of U and 4, then its mean
(mathematical expectation) is
D(U) = (u’) = co (V-(U>)%(V)dV,
s -a (QWv4)> = m m QU'ttilf,,(KlL)dvW.
s-m I--co
(2.94)
(2.107)
where u is the fluctuation
If Q is a function only of U, then this formula yields
u f U-(U). (2.95)
The distribution functions Fo($) and F,(V) contain
all the information about 4 and U separately, but they
(QW>=
Q(V){f_miA&‘, ti)dl//
i_pm 1 dV
do not contain joint information. That is, if 4 and U = coQU'Vu(v)dv,
are measured, the probability of C$< $ is given by s-m (2.108)

FJJI) and the probability of U < V is given by F,(V). which is consistent with the earlier definition, Eq.
But F,($) and F,(V) do not contain sufficient in- (2.93).
formation to determine the probability of the joint The marginal pdf’s f,(V) and f,($) have means
(simultaneous) event 4 < $ and U c V. This informa- (U) and (4) and variances (u2> and (&‘). For the
tion is provided by the joint distribution function joint pdf f,,( V, $) another important characteristic is
F,( V, t,b)which is defined by the covariance (u@),
F,& $) = P(U < V, 4 < JI). (2.96)
(Uf$‘)E m m (V-(U>)(~-(~))f.s(V,~)dVd~.
It is emphasized that the event (U < V, 4 < $) refers s -0D s -co
to simultaneous values of U and 4. (2.109)
The following properties of the joint distribution The normalized covariance r is called the correlation
function are readily deduced: coeficient :
0 G F,,(V,$) g 1, (2.97) r = (f4qY)/{(u’) (@‘)}““. (2.110)
F,,(-oo,$) = F,,(V, -00) = 0, (2.98) If r is zero, then U and 4 are said to be uncorrelated. If
(2.99) It-1= 1, then U and 4 are perfectly correlated. In
F,,(% II) = F,($),
general, r lies between minus one and one,
F,,(V, a) = F,(V), (2.100)
-l,<r,<l. (2.111)
and, F,( V, $) is a non-decreasing function of both V
and I++. Perfect correlation occurs when U and C#J
are linearly
The joint pdf of U and 4,fN( V, IL),is defined by related,
4 = a+/?u, (2.112)
where a and ,9 are constants. If B is positive, then r = 1
Both F,(V, $) and &( V, $) are defined in the two- and U and C#J are perfectly positively correlated: if fl is
dimensional V-tj sample space. By integrating Eq. negative, r = - 1 and U and 4 are perfectly negatively
(2.101) over an infinitesimal area in this velocity- correlated.
temperature space, we obtain In nearly-homogeneous turbulence, experiment@’
show that fti( V, t,b)is a joint normal distribution:
f&+(V,l(/)dVd$ = P(V G U < V+dV,
fJV,$) = c-“‘exp(-i(V,$)), (2.113)
$6 4 < Jl+dJI). (2.102)
where the constant c is
Other properties of the joint pdf are:
c=4~~(~~)(~‘~)(1-r~), (2.114)
f”,#(VY#) 2 0, (2.103)
m P and A(V, $) is the positive definite quadratic function
_&(V,ti)dVdJ/ = 1, (2.104)
f -cc J‘-m A(v,lL) = f{(V-<U>)2/(a2>

O” .LJV> Jl)dV =.&W), (2.105) -2r(V-(U))(rL-(+>)/((u2>(&2>)“2

s -m (2.115)
+(rL-(#>)2/(&2>Ml-r2).
and
The marginal pdf’sf,(V) andf+(ll/) can be obtained by
L,(Kti)d$ =f.(V). (2.106) integrating Eq. (2.113) over $ and V respectively.
r -m Both are normal distributions,
The pdf’s &(JI) and fM(V) are called marginal pdf’s,
f.(V) = (2~(~~))-“~exp(-t(V-(U))‘/<u~)),
and the last two equations show that these marginal
pdf’s can be recovered from the joint pdf. But, as has (2.116)
 Pdf methods for turbulent flows
I~..0 '
0.5
0.0 I I
0.0 0.5 1.0
F~G. 2.10. Joint-normal pdf, and 100 sample points. The
probability of a sample point being inside each ellipse is
(from the inside outward) 0.2, 0.4, 0.6, 0.8 and 0.99.
and
f,(~O) = (2n<~b'2>)-1/2 exp(-½(~b - <~b>)2/<~b'2>).
The joint-normal pdf is maximum at (V, qJ)=
(<U>,<q~>) where 2, Eq. (2.115), is zero. Lines of
constant (positive) 2 are ellipses in the V - ¢ sample
space, and on these ellipses the probability density is
constant, Eq. (2.113). From Eqs (2.113-2.115) it can
be shown ~s that the probability P(2.) of a sample
point (V, ~O}= (U, ~b) lying within the ellipse given by
), = 2. is
P(),.) = 1 - e x p ( - ~ , . ) .
Figure 2.10 shows the joint-normal pdf correspond-
ing to ( U ) = 0.4, ( ~ ) = 0.5, (u ~) = 0.01, ((h '~) =
0.00696 and r =0.3. The maximum probability
density (at (V, q~) = ( ( U ) , (~b))) is 20. Five ellipses are
shown on which the probability densities are (from
the inside outward) 16, 12, 8, 4 and 0.2. The prob-
ability of a sample point being within each ellipse is
0.2, 0.4, 0.6, 0.8 and 0.99. Also shown are 100 sample
points which were obtained from an appropriate
q,
ffu~ Iv,#)
0.7
IO
0.5
o.3
do o15
FIG. 2.11. Joint-normal pdf: .L,o(V.q,) against V for @= 0.3.
0.4. 0.5.0.6 and 0.7.
131
random number generator. It may be seen that the
sample-point density approximates the probability
density.
Figure 2.10 shows f~(V,~O) as a contour plot.
Alternatively, f,,(V, ~) can be plotted against V for
different values of ~O.This representation is shown on
Fig. 2.11.
The joint normal distribution is defined by the
means, variances and eovariances of the random vari-
ables. It is emphasized, however, that in general a pdf
or a joint pdf cannot be determined from a finite
)
V number of moments. Because of their occurrence in
nearly-homogeneous turbulence, joint normal pdf's
have an important place in pdf methods. But in in-
homogeneous flows, especially turbulent reactive
flow, pdf's can be far from normal. 6a
At a given location in sample space, the distribution
function F.~(V,¢) can be continuous or it can be
discontinuous in either or both of V and ~b. For
example, if ~b adopts the constant value ~ = ~b,, then
(2.117) F,, is discontinuous in ~:
F.,(V,d/) = F , ( V ) n ( d / - ( p . ) . (2.119)
The corresponding pdf is
f.÷(V,~O) = fu(V)6(O- ~b.). (2.120)
If U also adopts a constant value U = U., then F,, is
discontinuous in both V and ¢:
F,,(V,¢) = H(V-U.)H(~O-4a.), (2.121)
(2.118) and the joint pdf is,
f~,(V,¢) = 6 ( V - U.)6(~O-t~.). (2.122)
The delta function product, or two-dimensional delta
function, in Eq. (2.122) represents a unit spike at
location (U,, ~ , ) in the V-~ sample space.
We now consider the transformation of the joint pdf
under a change of variables. (Equation 2.48 is the
transformation for a marginal pdf.) Let 9 and h be
random variables (with corresponding sample-space
variables # and/t) defined by
9 = G(U, th), # = G(V,~), (2.123)
and
h = H(U,(b), fi = H(V, qJ). (2.124)
We require that the functions G and H be such that
there is a one-to-one correspondence between points
in V-¢ space and those in #-/~ space. This require-
ment is met if the Jacobian
0(V, ~0) ---- 0/~ 0 ~ ' (2,125)
has the same sign everywhere.
A given event corresponds to a region of V-~ space
,io which, in turn, corresponds to a region of ~-Fi space.
As an example, the event
A - - ( V Q < ~ U < Vb,¢°~<~b<qjb), (2.126)
132 S.B. POPE

q, 2.8. Conditional Probabilities

q'b - q'* Conditional expectations play a central role in pdf
methods. Before examining their properties we need
to define conditional probability and conditional
o) b)'~/ ~v=v° pdf's.
Let A and B be two events with positive probabili-
I I - -
ties P(A) and P(B). The probability of the joint event
v, v b ~ i AB is P(AB). Since the joint event AB can only occur if
FIG. 2.12. The event A (Eq. 2.126) in: (a) V-qJspace. (b)~-ac both A and B occur, it is clear that,
space.
P(AB) ~< P(A), (2.132)
is sketched as a region of V-$ space on Fig. 2.12a and and
as a region of g-/i space in Fig. 2.12b. With fg,(~, ~) P(AB) ~< P(B). [2.133)
being the joint pdf of g and h, the probability of the
The conditional probability of A and B is defined by
event A is
P(AIB) =- P(AB)/P(B). {2.134)
P(A) = f~,(V,O)dVdO This is the probability of the event A, given that the
a
event B occurs. It is obvious from Eqs (2.132-2.134)
that P(AIB) conforms with the requirements on
probabilities, namely
0 ~< P(AIB) ~< 1. [2.135)
The meaning of conditional probabilities is
Now since there is a one-to-one correspondence illustrated in the following example. Let n denote the
between points in the two spaces, the last integral can outcome of the throw of a fair die (i.e. n = 1, 2, 3, 4, 5
•be rewritten 6'~ or 6 with equal probability), If A is the event "'n = 2",
and B is the event "n is even (2, 4 or 6)" then,
f_~ , (_O , [_i ) d _~ d [ i = ~ Jff , , ~ , ' [ i , ~O~,[i)dVd.',~,.
P(A) = 1/6 (2.136)
(2.128) P(B) = 1/2 {2.137)
Since the event A can be chosen arbitrarily, com- and
paring the last two equations we obtain the P(AB) = 1/6. (2.138)
transformation rule
From the definition of conditional probabilities, Eq.
,. ti, I 0t#. (2.134),
f"*(V' $ ) = fghl'q' ~[0(---~,
0)l" (2.129)
P ( A I B ) - (1/6)/(I/2)-- 1/3, (2.139)
The absolute value of the Jacobian represents the
volume ratio between the two spaces. The probability and
in an infinitesimal volume d V d 0 at (V,O) is P(BIA) = (1/6)/(1/6) = 1. (2.140)
f . , ( V , 0 ) d V d 0. In ~-~ space, the corresponding That is, if n is even (event B), then the probability of
volume is 10(~, h)/O(V, ~0)1d V d 0 and hence the corre- n = 2 (event A) is i/3, Eq. (2.139): if n = 2 , n is
sponding probability is certainly even, Eq. (2.140).
For the random variables U and ~b, consider the
events
A = U < I/, (2.141}
As Eq. (2.129) states, these two probabilities are equal.
The joint distribution function and pdf have been and
defined for the random variables U and ~bmeasured at B - ~k ~< ~b < qJ +A~O. [2.142)
the particular location Xo, t o. As with the marginal
pdf's, this definition can be generalized to make F~, The probabilities P(A), P(B) and P(AB) can be deter-
a n d J ~ functions ofx and t: mined from the joint pdf J.,(V, 0). Hence, using the
definition of conditional probabilities, we obtain,
F.4,(V,~O,x,t) - P(U(x_, t) < V, ~b(x_,t) < ~), P(AIB) = P(U < VI$ ~<4' < ~,+ A~O)
(2.130)
and
02
j=~(V, qJ;x_,t) ----~F,,(V,~O, x_,t). (2.131) (2.143)
As before, the semicolon is used to separate the The conditional distribution Junction F,,b~,(VIO) is
variables ( V, ¢) with respect to which f=~, is a density. defined as the limit of this conditional probability as
 Pdf methods for turbulent flows 133

Aq, tends to zero* : pdff, l~ is obtained by dividing f . , , Eq. (2.113), by f+,

Eq. (2.117):
F~,+(VIq,) = P(U < Vl+ = q,)
f~l,(Vl~k) = [2•<u2> (1 - r ) ] - uz
= I ~ J~,+(l?,~b)dl?/f+(qJ ). (2.144)
x e x p { - ~ ( V - < U I q ~ = q,>)2/((u2> (1-r2))},
The conditional Pdfful~ is the derivative (with respect
to V) of the conditional distribution function, where the conditional mean of U is
J~l~ (VI~b) -f~o(V,~O)/f~(O). (2.145) < u I4~ = q,> = < u> + <u¢'> ( ~ - <4,>)/<4,'2>.
It is simple to verify that F,** and f,l* have all the (2.151)
properties of a distribution function and a pdf, respec- Thus the conditional pdf f.l+ may be seen to be a
tively. The significance of the conditional pdf is that normal distribution with mean (Ul~b=~b> and
variance <u2>(l-r2). Similarly, f~l. is a normal
f~l,(Vl~h)dV
distribution with mean
is the probability that U is in the infinitesimal range
V ~ U < V+dV, given that (p is equal to ~. It is also <4,1u = v) = <4~>+<u#> (v-<u>)/<u2>
evident from Eq. (2.145) that the joint pdf can ix (2.1521
recovered from the conditional pdf by and variance (~,2> (1 -r2).
By definition, the random variables U and 4~ are
Ju+ = J'+f:l+ = J,,f¢,l,. (2.146)
independent if their joint pdf is the product of their
If Q(U, 4)) is a function of U and q), the conditional marginal pdf's:
expectation of Q, given q5 = ~k, is
f.,(V, ~O)=f.(V)f~(qJ). (2.153)
E(O(U,~b)l~b = ~ ) = <Q(U,6)Iq~ = O)
It then follows from Eq. (2.145) that the conditional
pdf's are equal to the unconditional pdf's,
=ff~/~,,(V,~h)Q(V,O)dV.
fm~ (VI¢,)= f~(V), (2.154)
(2.147) and
The unconditional mean of Q can then be obtained by f,l.(~'l V) = f , (~b). (2.155)
Equation (2.154) shows that, if U and 4) are
<Q(U,¢)>=ff~ f~ f.,(V, qOQ(V,O)dVdqJ independent, a knowledge of ~b provides no informa-
tion about U. In particular, the conditional
expectation of U is the same as the unconditional
expectation.
Independent random variables are uncorrelated
= i ,o+<++,+,+=+>+. since
d --- Qc

In particular, the unconditional expectation of U is <u4/>=~[+f?(v-<u))(~O-<¢>)f"o(v'W)dVdq',

(U> = [ ~ f~@)<UI6 = O>d~,. (2.149)
=
*c

In the subsequent derivation and examination of

f_?v- <U>)f,(V)dVf_ (¢J-+'+'+'+
do

pdf equations, terms of the form <US@-q,)> are =0. (2.156)

frequently encountered. Substituting U 6 ( 4 ) - ¢ ) f o r
Q(U, O) in Eq. (2.148) we obtain:
< U3(~b- 0)> = f/~.f*(q/)<U6(O-~)14)=~'>d~'
= f /~ f+(q/)<Ub(q/-O)l¢=q/>dq/
= ~ .I+(q/)6(q/-O)<UlO = q/> dq/
d-
=J~(0) (UI~ = ~'>-
This important result is used extensively below.
For the joint-normal distribution, the conditional
*If Fo(0) is discontinuous, F,I ~ and f.l, can still be
defined but not by Eqs (2.144) and (2.145).
But a lack of correlation does not, in general, imply
independence, For example, with ~ being a uniformly
distributed random variable between zero and one,
the random variables g =__cos(2n~) and h = sin(2n~)
are uncorrelated, but they are obviously not indepen-
dent.
2.9. Limitations of One-Point Pdf' s
All the pdf's considered in this paper are one-point
pdf's--pdf's of random variables at the same location
(2.150) and time. An example of a two-point pdf, is the joint
pdff2 + of the random variables O(x_,t) and 4~(x_,,t,).
The two-point distribution function F20 is defined by
F2,(~k,~O.,x_,t,x,,t.) - P(~b(.x,t) < qJ,~b(x_,,t,) < q,,),
(2.157)
134 S.B. POPE

and the two-point joint pdf is 3.4. In Section 3.5 the transport equation for
f(V,~O;x_,t) is derived from the transport equations
f24,(@,d/,; x_,t,x,,t,)
for U and 4~.
02
=_- F2c~(~,~.,x_,t,x_.,t.). (2.158) 3.1. Conservation Equations
The one-point pdf can be recovered by The state of a reacting gas mixture composed of s
species is fully described by the species mass fractions
f¢(~O; x, t)-- f?A,(C,,~,,;x_,t,x,,t,)d~,,, m,(x_, t) (~ = I, 2..... s), the specific enthalpy h(x, t), the
pressure p(x, t), and the velocity U(x_,t). An equation
(2.159) of state determines the density p as a function of m__.h
and p:
but, in general, the two-point pdf cannot be deter- p = p(~h,p), (3.1)
mined fromf,(~; x_,t) a n d f , ( ~ , ; x,, t,).
It is clear, then, that one-point pdf's do not provide where m denotes the set of species mass fractions.
a complete description of random fields. They provide The evolution of the reactive flow field is governed
information at each point separately, but no joint by the conservation equations of mass, momentum,
information at two or more points. As a consequence, chemical species and enthalpy 65.66:
one-point pdf's contain no information about the
frequency or length scale of the fluctuations. ~+~x,(pu,)= 0, (3.2)
To illustrate this observation, suppose that ~b(x, t) is
a travelling sine wave of frequency co, wave length L, DU~ _ Oz~j ?~P ÷pg/, (3.3)
and random phase, P Dt ~3x~ t?xj
~b(x, t) = ½+½sin {2n(cot+xJL+~)}, Dm, ~J~
(2.160) p. . . . . q-pSi, (3.4)
Dt c~xi
where ~ is a random variable, uniformly distributed and
between zero and one. The corresponding pdf is
Dh ~J~
f,(~0;x.t) = _1 { I - ( 2 ~ - 1)2} -1/2. (2.161)
P D----t= - ?x--~-+-pSh, (3.5).

where the material derivative (the rate of change

The pdf, it may be seen, is independent of both the
frequency co and the wave length L. Consequently the
one-point pdf contains insufficient information to
determine a quantity such as ((Od~/Ox~)2) which
clearly depends upon the wave length L. (It follows
from Eq. (2.160) that this quantity is given by,
04' 2
((-~xl) ) = ~ 2 / L 2 . ) (2.162,
3. VELOCITY-COMPOSITION JOINT PDF
In a low Mach number gaseous flow, the state of the
fluid at any location is fully described by the three
components of velocity (U = U t, U2, U3) and by a set
of a scalars (~b = ~b1, q~2,...,q~,). The set of scalars 4'
comprises the species mass fractions and the enthalpy.
In section 3.1 the transport equations for U and ~ are
presented.
A complete one-point statistical description of a
turbulent reactive flow is provided by f(V.~O;x,t)--
the joint pdf of U and ~. At a given location in
physical space and time (x_,t), the joint pdff(V_,~b) is
the probability density in the combined three-dimen-
sional velocity space (V-space) and the a-dimensional
composition space (~-space). The properties of the
velocity-composition space and of the pdf are out-
lined in Sections 3.2 and 3.3. The joint pdf can be
represented by an ensemble of delta functions (cf. Eq.
2.81). This discrete representation, which plays a
central role in the pdf approach, is described in section
following the fluid) is
D a
D--t = -~+ U' ~x~" (3.6)
In these equations, gj is the body force (per unit mass)
in the xfdirection; % is the sum of the viscous and
viscous-diffusive stress tensors66; S~ is the mass rate of
addition (per unit mass) of species ~ due to reaction;
J~ is the diffusive mass flux vector of species ~; Sh is the
source of specific enthalpy due to compressibility,
viscous dissipation and radiation; and J* is the specific
energy flux vector due to molecular transport. The
molecular transport quantities rij, J9 and J* are com-
plicated functions of the local properties and their
gradients. 66
The source of m~, S~, is a combination of reaction
rates and stoichiometfic coefficients. The reaction
rates depend upon the mass fractions, the pressure,
and the temperature T(x_,t). Since the temperature can
be determined from m, h and p,
T = T ( ~ h, p), (3.7)
so also can the source S~:
S~, = S~,(m__,h,p). {3.8)
The fluid mechanical equations (Eqs 3.2-3.3) are
coupled to the thermochemical equations (Eqs 3.1,
3.4, 3.5, 3.8) principally by the pressure and density. It
proves to be extremely useful to invoke the low Mach
number assumption to remove the coupling through
the pressure. Let x o be a reference location and let the
 Pdf methods for turbulent flows
reference pressure Po be
po(t) = (p(X_o,t)).
The pressure can then be written as
p(x, t) = po(t)+Ap(x~ t), (3.10)
where the pressure variation Ap is defined by
Ap(x, t) - p(x,t)-po(t ). (3.11)
The pressure-gradient term in the momentum equa-
tion can be rewritten as
g P - ~ (Ap(x_,t)+po(t))= ~A--~p (3.12)
?xj ~x s ~x s "
showing that the uniform reference pressure Po has no
effect.
The density and the reaction rates (and hence S,)
depend upon the magnitude of the pressure, not upon
the pressure gradient. The equation of state (Eq. 3.1)
can be expanded in a Taylor series in p around Po to
yield,
p(m_,h,p) = p(m,h, po){l q POp~PopAp .... }"
Now, since Ap/p o can be estimated to be of the order
of the Mach number squared,/7 only the first term in
the expansion need be retained. Thus, to a good
approximation,
p = p(m,h, po),
and similarly, the sources S, can be evaluated by
S~ = S~(_~h, po).
The source in the enthalpy equation is 6~
c~Ui Dp , , ,
pS h = zi.i-~xi+-~+(A--e),
where A and e are the rates of absorption and emission
of radiant energy (per unit volume). At low Math
number, the viscous dissipation term is negligible and
the pressure term is well approximated by
Dp dp0
D--}-= d-T'
In non-sooting gaseous flames, radiative heat transfer
can usually be neglected in the enthalpy equation. It is
assumed that the rate of absorption A is negligible, but
the rate of emission can be retained in the equation
since it is a function of m__,T and P0. Thus, with the
assumption of low Mach number and the neglect of
radiation absorption we obtain
S h = S,(m, h, Po, dPo/dt).
It may be seen that with the low Mach number
assumption, the pressure variations Ap do not affect
the thermochemical equations while the reference
pressure Po does not affect the momentum equations
directly. The equations are still coupled through the
density which depends upon Po. (Strahle~7 and others
135
have also determined that Ap does not affect the
thermoehemical equations at low Mach number.)
(3.9) In some applications (notably internal combustion
engines) the reference pressure po(t) varies signifi-
cantly with time. In other applications (open flames,
furnaces, and gas turbine combustors in the steady
state) the reference pressure is constant. For
simplicity, we consider the case of constant reference
pressure, but the extension to the variable case is
straightforward.
The mass fraction and enthalpy equations can be
written in a common form by defining the set of
a --- s + 1 scalars ~_(x~t) by:
dp~=m~, a = 1 , 2 ..... s = t r - 1 ,
and
~bo = h. (3.191
Then, with J~-= J~ and S, = Sh, the general scalar
transport equation is,
D~b~ ~J~
p .... ~-pS~, ~ = 1 , 2 ..... a. (3.20)
Dt c~xi
For a given reference pressure Po (assumed constant),
the density and sources depend only on the set of
(3.13) scalars ~:
p = p(~), (3.21)
and
S, = S,(~b). (3.22)
The set of a scalars ~ defined by Eq. (3.19) provides
(3.14)
a complete description of the thermochemical
properties of the gas mixture. Many combustion
problems involve a large number of species, and con-
(3.15)
sequently a is large. (For example, the combustion
of methanol may involve 26 species. 6s) In turbulent
combustion calculations, simplifying assumptions are
(3.16) usually made in order to reduce the number of scalars
needed to describe the gas composition. Extreme
examples are the idealized premixed and diffusion
flames, in each of which the gas composition is (by
assumption) determined by a single scalar. For the
idealized premixed flame 69 the single scalar is the
progress variable (or the normalized product concen-
(3.17) tration): for the idealized diffusion flame 7° the single
scalar is the mixture fraction (or the nozzle-fluid con-
centration).
In order to exploit such simplifying assumptions,
we redefine the set of scalars considered as follows:
~_ is a set of a scalars (~_=t~l,~ 2..... ~b,),
governed by the general transport equation Eq.
(3.20), which, in conjunction with the simplifying
assumptions (if any), is sufficient to determine the
(3.18) density p, the source terms S~ (ct = 1, 2,..., a) and
any other thermochemical property of interest.
The scalars ~bare called the composition variables since
they determine the composition of the gas mixture.
Their definition ensures that Eqs (3.20-3.22) hold, and
that (at least by assumption) the mass fractions and
enthalpy are known functions:
136 S.B. POPE
m, = m,(q~), (3.23)
and
h = h(¢_). (3.24)
In summary, the equations governing the evolution
of the turbulent reactive flow field are the mass con-
servation equation, Eq. (3.2), the momentum con-
servation equation, Eq. (3.3), the composition
equations, Eq. (3.20) and the equation of state, Eq.
(3.21). No assumptions have been made about the
reaction rates S t or the molecular transport terms r~,
J: and J~.
3.2. Velocity and Composition Spaces
At a given position and time (x_,t) = (x0, to) in a
turbulent flow, each component of the velocity
(Ui=U~(~t), i = 1 , 2 , 3 ) is a random variable.
Together, the three components form the random
vector U. Let (VI, 1/2, I/3) be the coordinates of the
point __Vin a three-dimensional Euclidean space that
we call the velocity space. Then the simultaneously-
measured values U1, U 2 and U 3 can be plotted as the
sample point V = U in the velocity space. The velocity
space (or V-space) is the sample space of the random
vector U.
Similarly, each of the composition variables (~b~=
~b~(x_0, to), ~ = 1,2 ..... a) is a random variable: taken
together they form a a-dimensional random vector ~.
Let ( ~ , ~2 ..... ~b¢)be the coordinates of the point ~ in
a a-dimensional Euclidean space that we call the
composition space. This is the sample space of the
composition variables, since a simultaneous measure-
ment of (~bt, ~b2..... ~b¢) can be plotted as the sample
point ff = ~b.
If the composition variable q~ is a mass fraction--
m~, say--then (from the definition of mass fractions)
~b~must lie between zero and unity:
0 ~< q~, ~ 1. (3.25)
A value of ~b, exceeding these bounds cannot occur. In
general, there is a region in composition space
(~_-space) called the allowed region within which any
point ~_corresponds to a possible composition ~b = ~.
Outside the allowed region, a point _~corresponds to a
composition ~b = ~, that cannot occur--possibly
because a negative mass fraction is implied.
Equations (3.21-3.24) show that properties such as
p, S,, m, and h are known in terms of the composition
variables ~. Since every point ~ in the allowed region
of the composition space corresponds to a possible
composition ~ = ~, the corresponding properties at
the point are p(~), S~(~), m~(~), h(~), and so on. In this
way, any property of ~b can be regarded as a function
in composition space. The density p is a scalar, while
quantities which have er components can be regarded
as vectors in ~-space. In particular, ~ = ( q ~ ,
~b2 ..... ~)er) and S = (S t, $2 ..... So) are vectors.
Processes, as well as properties, can be represented
in composition space. Consider a homogeneous,
~ th
FIG. 3.1. Reaction path in augmented composition space.
quiescent, reactive gas mixture, which at time t 0 has
the composition q~(to) = ~bo. According to Eq. (3.20),
as reaction takes place, the composition evolves by
dq~
dt = S~(~b), (3.26)
or, in terms of the vectors _~ and S,
d_~_= S(~b). (3.27)
dt --
(Because of the assumed homogeneity, ~ depends only
upon t, and the other terms in Eq. (3.20) are zero.) The
solution to Eq. (3.27) can be written
~(t) = O__(~b0, t), (3.28)
to show, explicitly the dependence of ~_(t) upon the
initial condition _~o.
At any time, the composition can be represented by
a point in ~-space. Initially the point is ~, -- ~(~b0, t).
Such a trajectory--or reaction path--is sketched on
Fig. 3.1 in the augmented composition space (~k-t
space). The composition is shown approaching the
value ~k = ~b, corresponding to chemical equilibrium.
It is evident from Eq. (3.27) that _S represents the
velocity of the point in composition space, and conse-
quently reaction paths are everywhere tangential to S.
The velocity vector U is a three-dimensional
random vector and q~ is a a-dimensional random
vector. Taken together, the (3 + a) quantities (U_.,¢_) =
(Ul, U2, U3, ~1, ~b2..... ~b,) form a (3 + a)-dimensional
random vector. The appropriate sample space for this
random vector is the (3 +a)-dimensional Euclidean
space in which the coordinates of the point (V, ~,) are
( Vt, I:2, V3, ~1, ~k2,..., ~). This is called the velocity-
composition space or V-_~ space. The experimental
techniques of laser-Doppler-anemometry and Raman
spectroscopy are approaching the stage where simul-
taneous measurements of U and ~_ are possiblefl t A
simultaneous measurement of these (3 + (r) quantities
could be plotted as a point in V-~_ space.
3.3. Velocity-Composition Joint Pdf
At the particular location and time (x, t) = (Xo,to),
the three components of velocity and the a corn-
 Pdf methods for turbulent flows 137

position variables, form the (3 + a)-dimensional infinitesimal volume dl/l at JI is f&(k) dk. Similarly, for
random vector (U_ 4). The (3+a)-dimensional the (3 +a)-dimensionakdom vector (IJ, p), the prob-
distribution function of(LJ, 8) is defined by ability of its being in the infinitesimal volume dy d& at
(V_L) is f(V_ &) dy d&. (These properties of the pdfs
F,,(Xi,)
- = PW, < V,, u, < v,, u, < V,, follow from the definition of the distribution function,
41 < $I,& < $~....r& < ti,). (3.29) Eq. (3.29).)
The sample point corresponding to a value of the
At every point in the velocity-composition space
random vector (U_&) is sure to lie somewhere in the
(v--k space), this distribution function has a value
Y-k sample space. Thus the probability of this event
between zero and one. The joint pdf of &J,$) is
(which is given by the integral off(V_t& over the whole
defined by, space) is one :

(3.39)

Similarly,
Before outlining the properties off* we introduce
some abbreviated notation. First, f (without sub-
scripts) is written forl$:

f(Y.!k) =f&,). (3.31) and

sf,(VM! = 1, (3.40)

An infinitesimal volume in y-space is written dy,

and an infinitesimal
dy = dV, dV, dV,,
volume in -$-space is d&,
(3.32) sf$&)dL = 1.
Let Q@,$) be a function* of U and I. Corre-
(3.41)

spondingly, Qw, I& is a function defined everywhere in

d&=d+,$,...d+,. (3.33) velocity-composition space. The mean, or expectation,
ofQis
Thus dydk is an infinitesimal (3+a)-dimensional
volume in velocity-composition space. Integration <Qcrr,&4> = QtX@fO!v~WYdk. (3.42)
over the whole of y-space or &-space is indicated by a 51
single integral sign: If Q is a function of U only, then its mean is

j( )dV= [;m {IK 1:x t )dV, dl/;dV,, <Q@D = Q&U-(X@L)dJ!d&

(3.34)
a2

(3.43)

Similarly, if Q is a function of & only,

Integration over the whole of v--k space is written as

( M!dk.
II
It may be recakd that & is always within the
The joint pdf of the three velocity components &a) allowed region of &-space. Physical properties such as
is obtained from f (Y, I& by integrating over k-space: the density p have no physical meaning outside the

f,(V) = f(X$L)dk.
s
And the marginal pdf of a single velocity compo-
(3.36)
allowed region. However, it aids both the notation
and the analysis to regard p&b) as a finite continuous
function defined everywhere in k-space. The definition
of p(k) outside the allowed region is arbitrary (except
nent--U,, say-is obtained by integrating fu over that p(&) should be finite and continuous) since
the other two velocities: p(&)f+ (I& is non-zero only in the allowed region. Thus,
the mean density can be determined from
f”,(V,) = a cc f,O!)dV, dV3. (3.37)
s I
Similarly the joint iif ifXthe compositions, f$ (&) is
obtained by integratingf(V_&) over velocity space
(PW =
sA$)f&)d&~ (3.45)

and is independent of the definition of p(k) outside the

allowed region. Similarly, all other functions (e.g. S@))
f,Clt) = (3.38)
are regarded as being defined everywhere in com-
position space.
The probability that J_Jis in the infinitesimal volume
dy at J! is f,(v)dy: the probability that $ is in the * See footnote in Section 2.5,
138 S.B. PoP~

In Section 2.5 it was shown that the pdf of a single where the density-weighted joint p d f f i s defined r° by
scalarf, (~) is equal to the expectation of a delta function
f(V, ~0) - f(V_, ~b)p(g,)/(p). (3.56)
( 6 ( ~ - $ ) ) , Eq. (2.67). Similarly, for the joint pdf we
have With this definition of f, Eq. (3.55) for the density-
weighted mean, ~ is directly analogous to Eq. (3.49) for
f(V__,g/) = (6(V-U_)6(~-~)), (3.46)
the mean (R).
where 6(_y_-U) is written for the three-dimensional delta The random vector (U, ~) has been defined as the
function at _V = U, velocity and composition at the particular location and
time ~ t) = (X_o,to). Clearly, the random vector (and
~ v(y_-u) = ~(v~ - u o6(v2 - u~)~(v3 - u3), hence the distribution function and pdf) can be defined
(3.47) at any ~ t). In general, then, the joint distribution
and, similarly, 6 ( ~ - ~ ) is the a-dimensional delta function is defined by,
function at ¢ = gt, F,#(_y_, ~,x_, t) = P(UI(x, t) < V1, U2(x_,t)

~ ( ¢ - ~ ) = ltI ~ ( 0 , - ~ , ) . (3.48)
< v2, U3(x_, t) < v3, ~,(x~ t)
< ,P~, 4,2(x_, t) < 02 ..... 4,,,(x_,t) < 0~,),
If R is any random variable, its mean can be deter-
mined by (3.57)
and the joint pdf is defined by,
(R>=ff<RIU=V.~_=¢>fV~t)dVd~, ~3+ct

(3.49)
where (RI__U = V_,q~ = ~ ) is the conditional expecta-
tion ofR, given that U = V and 4} -- ¢t (cf. Eq. 2.149). If
R is independent of U and q~, then the conditional
expectation is equal to the mean,
<RIU__ = V. 4~ = ~ ) = ( R ) , (3.50)
in which case Eq. (3.49) is not useful. At the opposite
extreme, if R is completely determined by U and q~ (i.e.
R = R(U__,~_)),then
( R ( _ ~ b ) I U -- V,~b = ~ ) = R(y_,~0), (3.51)
and Eq. (3.49) reduces to the definition of the
unconditional mean. Henceforth, conditional expecta-
tions are written (R IV, ~), it being implied that _Vis the
value of U_U_,and ~ is the value of ~p.
An important result embracing conditional expec-
tations and delta functions is
(R~(_V-U)6(~-q~)) = ( R I ~ g ~ ) f ( V . ~ ) . (3.52)
f v(y_~.;x_,t) -
OVI OV2OV30~IIO~I2 ...(~Ia
xF,2 V(y_,~O,x_,t). (3.58)
It is emphasized that f ~ ~; x_,t ) contains information
about every point ~ t) separately: it contains no two-
point information in physical space or time.
In applying the pdi" equation to inhomogeneous,
variable-density flows, the natural dependent variable
is neither f nor f b u t it is the mass density function (mdf)
~'(_y_,~O,x_; t) --- p(~)f(V__,~; x, t)
= (p)f(V__,~O;x~t). (3.59)
As is explained in the next subsection, the mass density
function ~" is the expected mass density in _ V - ~ - x
space.
The important properties of ~ follow immediately
from its definition and from the properties o f f :
f f : ' ( _ y . ¢, x_; t) dV dg, = (p), (3.60)

This result, which is used extensively below, can be

deduced from Eq. (3.42), (cf. Eq. 2.150). ff:v(y.¢,x;t)[p(g/)]-ldVd~= 1, (3.61)
For variable-density flows, ther.e can be an advan-
tage 17 in
. . . .
cons~denng densRy-weighted means (or Favre
(p)0,
averagesT°). For the random variable R, the density- d,} (3.62)
weighted mean/~ is defined by
and
R _= ( p R ) / ( p ) , (3.53)
and the fluctuation about this mean is f f ~(v__, q/,x; t)( Rl~ ~_) dV dy~-- (pR) --- <p)R.
R" - R - R. (3.54) (3.63)

F r o m Eqs (3.49) and (3.53) w e obtain In these equations, the means are functions of_x and t
and (RJV_, g/) is written for
(R(x, t)l__U(~, t) = V_, ~(x_, t) = 0 ) .

3.4. Discrete Representation

In Section 2.6 the ensemble-averaged pdf f~N was
defined by (Eq. 2.81)
 Pdf methods for turbulent flows 139

1 N produce 10 % accuracy, then 10,000 samples are needed

f,N(~o) -= ~ .~, ,~(¢- ~(')), (3.64) for 1% accuracy and 1,000,000 samples are needed for
0.1% accuracy. The over-riding strength of the
where 4~(') is the nth measured value ofq~(~o, to). The lxlf representation is its ability to handle pdf's of large
f¢,(¢ ) and the ensemble-averaged lxtff#n (¢) are related dimensionality. The discrete pdf is defined by N
in two ways. First, for any N(N >1 1), f,~ is the expected (3 + cr)-vectors, that is, by (3 + a)N numbers. Thus for
value off#n (Eq. 2.83): given N, the number of numbers required to represent
(f#N(~b)) = f#(~). (3.65) the pdf rises only linearly with the dimensionality of the
sample space. And Eq. (3.71) shows that the statistical
Second, for large N, the ensemble average of any
error is independent of the dimensionality of the sample
function Q(q~), space. Thus, if a three-dimensional pdf can be
(3.66) adequately represented by 1,000 samples (3,000 num-
bers) then a ten-dimensional pdf can be represented to
the same accuracy by just 10,000 numbers.
approximates the mean (Q(~)). The statisticalerror in
The ability to treat the problems of large dimen-
thisapproximation is of order N - i/2.
sionality and the slow convergence (8~2NaN-1/2)are
In thissection the discreterepresentationof pdCs as
characteristics of all Monte Carlo methods):
ensembles of delta functions is extended and general-
Each of the N vectors (_U("),~b(")) defines a sample
ized.This discreterepresentationiscentralto the whole
approach. First, for homogeneous constant-density
point in velocity-composition space. These points can
flows, a discrete representation for the joint p d f f ~ ~) be thought of as the locations (_V_,~) = (_U_("),~b(")) of N
is obtained: then, for the general case, a discrete notational particles in V_-g, space. The normalized
particle number density (in V_-¢ space) is
representation for the mass density function .a~(y_,~, x)
is given. 1
The discrete pdff~N (¢) (Eq. 3.64) is defined in terms of ~. 5(V-U("))f(~k-~b("))=f~(V__,~b), (3.72)
n=l
N sample values ~b("), n = 1,2 ..... N. In order to
represent the joint p d f f (V_,~), N samples are needed of and Eq. (3.68) shows that the pdf f (_y.,q/) is the expecta-
the (3 + tr)-dimensional random vector (_U_,~). Then the tion of this normalized particle number density.
This discrete representation o f f ( ~ ~,) (Eq. 3.67) is
discrete pdffN(V_, ~) is
used to study homogeneous flows. But for the general
1 N
variable-density inhomogeneous case, a discrete
f~(Y-"q/) --=N- n~l di(V-U("))6(~-~b("))' (3.67) representation of the mass density function 3~-(V__,~k,x_)is
required.
This is an ensemble of N (3+a)-dimensional delta
Consider the turbulent flow within a volume ¥ of
functions in velocity-composition space.
physical space {x-space). The expected mass of fluid
While it is natural to think of (_U_(")tq~(")) as simul-
within this volume is M,
taneously-measured values of the velocity and com-
position, the discrete representation (Eq. 3.67) is valid
M = f(p(3_) ) dx, (3.73)
for any random variables (_U_("),~b(")) with pdff(y_,~).
For then, the expectation of Eq. (3°67) is
where S d_xis written for the integral over ¥,
(fN(V, ~_)) = f v(y_,¢). (3.68)
The ensemble-average (Q(_U_,q~))N of any function f ( ) d x _ = f f f v ( )dx, dx2dx 3. (3.74,
Q(_U_,q~) can be obtained consistently from the discrete
Within the volume ¥, the mass density function
representation by
~(_V_,¢, x) is represented by N notional particles, each
<Q~)>~=ffQ(v.t)f,,V~t)dVdt representing a mass Am,
Am - M/N. (3.75)
1
= ~ - ~ Q(_U("),tb(")). (3.69) The nth particle has velocity U ("), composition tp(")
n=l
and position x ("). The (6+#)-dimensional vector
The ensemble average (Q>N approximates the mean (.U__('),~b('),x(')) defines the state of the particle and is
<Q> with a statistical error eQN: called the state vector. This state corresponds to the
point (y_,¢,x)= (_U_("),~b("),x(")) in the (6+a)-dimen-
<Q>~, = <Q> +eo~. (3.70)
sional state space--V(y_-~k-x) space. The discrete mass
The expected error E(e~N) is zero, and the standard density function ~-N(Y-,~, X) is defined by
deviation e~N is N

8~¢ = x/D(eQN) = N=':~ Dx//-D~. (3.71)

~N(Y_,~,x_) -= Am y ,~(y_-u("),~(C/-~(")),~-x'")).
n=l

This formula (Eq. 3.71 ) for the standard error reveals (3.76)
the strength and weakness of the discrete representa- The numbering of the particles is arbitrary. In
tion. The weakness is that the statistical error decreases other words the random vectors (U(n),~b("),x_(")),
slowly with increasing N: if 100 samples (N = 100) n = 1 , 2 , . . . , N , are identically distributed. Con-
140 S.B. POPE

sequently, the expectation of Eq. (3.76) can be written 3.76). In physical space, the expected particle number
N density is proportional to the mean fluid density (Eq.
<~-..(v_.~0.x)>= Am ~ <6(y_-u~"q6(~0-~"q 3.82). At a given location in physical space, the expected
particle number density in velocity-composition space
x 6 ~ - x~"~)>, (3.77) is proportional to the density-weighted joint p d f f (Eq.
3.84). In state space (V-~-x space) the expected particle
or, alternatively,
number density is proportional to the mass density
<~-N(Y_,¢,,x_)> function ~" (Eqs 3.78 and 3.79).
For constant-density homogeneous flows, the
= M<6(V_- U__.
t*~)6 ( ~ - ~ t'~) 6(~-x~"})), (3.78)
expected particle number density is uniform in physical
for any n (i ~< n ~< N). space and the statistics of U I"~and ~"~ are independent
The discrete mass density function ,a¢ has been of x ~"1.The volume averaRe of the discrete representa-
defined (Eq. 3.76) in terms of the state vectors of the N tion of ~(V,_C,x_) (Eq. 3.76) is then the same as the
notional particles, but the properties of these random discrete representation off(V_, ~) (Eq. 3.67), thus estab-
vectors have yet to be defined. Their properties are now lishing the consistency of the two representations.
deduced from the requirement that the expectation of We have obtained a prescription for determining the
ff~ is the mass density function ~': discrete representation ~-N(_~,~,x) from the mass
density function ~(y_, ~b,x). In the solution procedure
(~(.y_, ~O,x_)) = ~'(y_,~,x). (3.79) described in Section 6, the discrete representation is
Let h~) be the Ixtf of particle positions--that is, h(x) used, and consequently we need to consider the inverse
is the probability density of x t-~ = x. Then integrating problem--to determine (or, rather, to approximate)
Eq. (3.78) over velocity-composition space yields expectations from ~-N(-Y,~, x_).
To this end, let the volume ¥ be divided into K cells,
ff<.. >dVdq.=M<6~_-x'')> the kth being centered at ~k~ and having volume Yk.
These cells are sufficiently small that within each one
= Mh(x), (3.80) statistical homogeneity can be assumed, and the
total number of particles N is sufficiently large that the
and integrating Eq. (3.79) yields number of particles Nk in the kth cell is large.
The volume average of Eq. (3.82) over the kth cell is
f f <.,,>dv dq, = (3.81)
(Nk>/(NYk) = (p(xlk}))/M (3.85)
A comparison of these results shows that the particle where x_j~}is a location within the cell. Consequently, an
position pdf is approximation to the mean density is
h~) = (p(~))/M. (3.82)
(P(~k~)> ~ AmNk/¥k. (3.86)
Thus the number density of notional particles in
physical space is proportional to the mean fluid density. This is simply the mass of particles in the cell divided by
(It may be noted that the integral over ~ of each side of the cell volume. There are two approximations in Eq.
Eq. (3.82) is unity.) (3.86): first (Nk) has been replaced by Nk, resulting in a
Let f*(y_~ ~lx) denote the joint pdf of U ~'~ and ~"~ statistical error of relative magnitude N k 1/2; second,
conditional upon x t'~ = x. Thisis the joint pdfofU ~ and because of the assumed homogeneity with the cell, -~kl
qb~"~for a notional particle located at x. In terms of delta has been replaced by ~t~.
functions, this conditional pdf is A similar approximation is obtained for the density-
weighted mean of any function Q(U__,~b). Multiplying
Eq. (3.79) by Q(V_,~), integrating over v_-~k space, and
x 6~-xt"q)/<a~-xt"})). (3.83) taking the volume average over cell k we obtain

The numerator is <Scs)/M (Eq. 3.78) while the A m ( ~ (_U_c,}, ~,,)>,

(P(~'k~)Q(~'~)> = ~ Q ~ (3.87)
denominator is <9)/M (Eq. 3.82). Thus invoking Eqs
(3.79) and (3.59) we obtain
where ~o is a location within the cell, and summation is
f*(Y_, ~kl$) = .~(_~, qt. x)/(p(~)) over all the particles in the cell. Consequently an
approximation to ~ is
= f(_y_,¢,; x_). (3.84)
~. AmN k 1_ (U_~.~, {.}
This equation shows that, in order to satisfy the require- <p(x,.,)>~.,) { V. }N.~ Q ~ )'
ment <~'s) = ~,thepdfofnotionalparticleproperties
at a given location must be equal to the density- (3.88)
weighted pdf of the fluid properties.
or, recognizing that the term in braces is an approxima-
To summarize the properties of the discrete represen-
tion to <P(~k})) (Eq. 3.86),
tation for variable-density inhomogeneous flows: the
mass density function ~J(V. ~t, x) is represented by N 1
(~(x,k~) ~ ~ ~ Q(U_'"~,~b~"~). (3.89)
notional particles, each representing a mass Am (Eq. N~
 Pdf methods for turbulent flows 141

Thus, the ensemble average over the particles in expressed in terms of a conditional expectation,
the cell approximates the density-weighted mean.
jr c/ As _v,
3.5. Joint Pdf Transport Equation
F; ,,,OQ V~,)
A transport equation for the velocity-composition
joint pdfcan be derived from the transport equations for
U and qS, Eqs (3.3) and (3.20). These equations can be
xfdVd~. (3.98)
rewritten as The second step follows since, for U = V and ~b = #,
pOQ/OUs is a known function of_Vand ~0.Integration by
D U_____=2 As, (3.90) parts yields,
Dt
and V 0
D~=
-- = O:, (3.91)
Dt x [p(~0) (AslV__,~ ) f ] dVd~, (3.99)
where A is the force per unit mass, where

pAj(x_, t) = O~iJ bp ~-PgJ'

Oxi Oxj (3.92)
and O is the net source ofq~,
OJ7
pO:(x,t) =- --~x +PS,. (3.93)
There are several different methods of deriving pdf
transport equations. A useful method (originated by
Lundgren ~9 and subsequently used by several
researchers 2a.2L~7.6°) starts from the expression for
the pdf as the expectation of a delta function. Starting
from the expression
f(V, ¢; x, t) = (fi(V- U(x, t)) 6 ( ~ - ~(x_, t))>, (3.94)
Pope 72 used this method to derive the transport equa-
I - f f a-~rm(q')QVtY-,~_)(AslV-,~_}f3dV
d~_.
(3.100)
In Section 3.5.1 it is shown that for the wide class of
functions Q considered, the integral I is zero.
Following the same procedure for the final term in
Eq. (3.97) we obtain
aQ
x [pQ~)<o~IV~¢>f] dV dl/,, (3.101)
and combiningEqs (3.99)and (3.101)(withI = 0),the
second expressionfor(pDQIDt) isobtained,
- V

tion for f. A different method that does not involve delta

functions is used here. The method is to equate two
independent expressions for (pDQ/Dt), where Q(U, ~b)
is (almost) any function. (The choice of Q is discussed in (3.102)
Section 3.5.1.) With the aid of the continuity equation, By subtracting the second expression for (pDQ/Dt},
Eq. (3.2), we obtain Eq. (3.102), from the first expression, Eq. (3.95), we
obtain,
~ D Q effp(q,)Ov(Y_,q,)ydV__dq
(p__~_~l=_ '
e ffp(q,)VsQ(~q,)fdV_dq,
+ ~-~j [P(q~)<AslV--,q~}f]

bxj) --
(3.95)
This is the first expression for (pDQ/Dt>. The terms within the braces are independent of Q.
A second expression is obtained by relating changes Consequently, a sufficient condition for the satisfaction
in Q to changes in U and ~b, of Eq. (3.103) (for any Q) is that the terms within braces
sum to zero: but since Q can be chosen arbitrarily from a
DQ(U,~b) JQ DUs OQ D~b, wide class of functions, this is also a necessary condition.
(3.96)
Dt = OU~ Dt +t?~b,-- Dt Hence we obtain the transport equation for the
Note that summation is implied over repeated suffices (j velocity-scalar joint p d f f (V_,~0; ~ t):
and a). The material derivatives can be replaced by Aj
and ®, (Eqs 3.90 and 3.91) to yield p(¢/) +p(~.)V~ = Ob [p(d/)(AslV,~_>f ]
=( ~ s ){ .~: ,. (3.97,
---
d
[p(q,) < o , I ~ ~o>f].
oqJ,
The first term on the right-hand side can be re- (3.104)
142 S.B. PoP~

Equivalently, the transport equation for the mass The terms on the right-hand side of Eq. (3.109)
density function ( ~ = pf) is represent transport in velocity space by the viscous
stresses and by the fluctuating pressure gradient, and
a~ a [vj~] a transport in composition space by the molecular fluxes.
Before the equation can be solved for f, the conditional
a expectations appearing in these terms must be deter-
+ ~ - ~ [ < O , lV,~>~~r] = O. (3.105)
mined or approximated. In Section 5 models are
presented that approximate the conditional expecta-
It may be seen that the change of ~ with time is caused tions as functions of the joint pdf and of a specified
by transport in x-space (due to the velocity), transport turbulent time scale r(x_,t).
in V-space (due to A__),and transport in ~O-space (due to
O__).A complete discussion of this fundamental equation
is provided in the next section. 3.5.1. Evaluationof the integralI
Expressions for the conditional expectations of A and It is shown that the integral
are obtained from their defining equations. From Eq.
(3.93) we obtain I=-f f a-~j[P(~)O(~O)<A/YY,~>f]dV--d~,
=
I aJ~ \ (3.110)
is zero for a wide class of functions Q. It is simply shown
aJT that I is zero if Q has bounded support. In addition, with
-- - <~xi, V_,@>+ p(@)S,(@). very reasonable physical assumptions, it is shown
(3.106) that I is zero ifQ is monotonic as IvI tends to infinity and
the expectation <pAjQ> exists.
By decomposing the pressure p into its mean <p> and Let S(v) denote the surface of the sphere in V-space,
fluctuation p' centered at the origin, and with radius v,
p' -- p-<p>, (3.107) vz = V . V , (3.111)
fromEq. (3.92) we obtain and let g(v)be the probability density of[U_" U__]1/2 = v.
By first integrating over composition space, the integral
P(~-)<AjIV'~> = \ax i a<p> ap' +poAv,t)
Oxj axj I becomes

=
azu
--v,#L -
a<p> ' I= f J~j[<pA~QW>./~(V)]dV. (3.112)

+,o(@)g i. (3.108) This is the integral of a divergence in_V-space, and hence

using the divergence theorem, it can be rewritten as the
With these expressions for the conditional surface integral,
expectations, the joint pdf transport equation, Eq.
(3.104~can be written as, I = lira [ <pAjQlV)nJu_(V)dS(v) (3.113)
0i(
p(¢)~-+ p(~_)b~-7~
j p(_~)ej +
v~oo .IS(v)
= lira <pAjn~QIv>g(v), (3.114)

a<p>)~+~[p(~)
of a s ,(~)S] where n is the outward-pointing normal vector.
It is obvious from Eq. (3.114) that if Q has bounded
support then I is zero.
= ~[(_~. o~ We now consider less stringent conditions on Q. The
mean <pA~Q>can be expressed as

<pAjQ> = ~ <pAjQ[V>fu(V)dy, (3.115)

(3.109)
If the conditional expectations appearing on the right-
hand side were known, this equation could be solved for
f(# and S are known functions in ¢-spaee and <p> can
be determined from f, see Section 4.7). Thus, the
processes represented by the terms on the left-hand side
are accounted for exactly, without any approximation.
These processes are: transport in physical space; trans-
port in velocity space by gravity and by the mean
pressure gradient; and, transport in composition space
by reaction. It is remarkable that in a variable-density
flow with arbitrarily complicated reactions, all these
processes can be treated without approximation.
and its existence implies the absolute convergence of
this integral. That is, it implies the convergence of the
integral,
f I<PAjQIV>Ifu(V)dV
=fo{fs,~l(pA~QIV)l~(V)dS(v)}dv. (3.116,
Now if Q is monotonic as v--* oo, since p and Ai are
physical quantities it is reasonable to assume that
I<oA~QIv>l is also monotonic. Then, the convergence of
the integral (3.116) implies
 Pdf methods for turbulent flows
lim I(pAjQIv)lo(v ) = O. (3.117)
I , ~ oE
Now from Eq. (3.114),
I I I = lim I<pAjnjQIv)lg(v)
3
~< lira ~ I<pAjQIv>lo(v)=O. (3.118)
v ~ j= 1
Thus the monotonicity of Q at infinity and the existence
of <pAjQ>are sufficient to ensure that I is zero.
4. L A G R A N G I A N DESCRIPTION
The velocity-composition pdf equation derived in
the previous section is clearly useful, since the major
processes (convection and reaction) can be treated
without approximation. Before calculations can be
made using the equation, the conditional expectations
must be modelled, and a method of solving the equation
must be devised. With simple models and
for extremely simple flows, analytic solutions can
be obtained. But, in general, a numerical technique
is required. A standard technique--such as finite-
differences--is bound to be impracticable, -a2 because
f(V, ~,: x, t) is a function of (6 + ~r) variables and time.
The modelling and the numerical solution procedure
are strongly interconnected. The purpose of this section
is to show this interconnnection by considering both
topics in general terms. In the following two sections,
modelling (Section 5) and the solution procedure
(Section 6) are examined more specifically.
In Section 4.1 the observation is made that many
different systems evolve with the same pdf. In Sections
4.5 and 4.6 deterministic and stochastic systems
evolving with the same pdf are described. These systems
are used to interpret the pdf equation and to model it,
respectively. All the systems considered consist of
particles. Hence a Lagrangian viewpoint--developed
in Sections 4.2-4--is adopted. This Lagrangian
formulation provides a close link between the pdf
equation and the physical processes of turbulent reac-
tive flows.
There are subtle questions concerning the mean
pressure field, the mean continuity equation, and the
normalization of the mass density function. These
questions are addressed in Section 4.7.
This section may be the most difficult to comprehend:
several different pdf's are introduced, and (as usual with
Lagrangian formulations) the notation is involved. So
that the important conclusions are not obscured by
these difficulties, a summary is provided in Section 4.8.
The reader may find it useful to refer to this summary
after reading each sub-section.
4.1. EquivalentSystems
In Section 3, the joint pdf transport equation was
derived from the conservation equations,
--~-+-;-- (pU,) = 0, (4.1)
143
a
(-~+ U~-~x~)Ui= Aj(3_,t), (4.2)
and
~ - + U, ~b,, = e,(x, t), (4.3)
where A and O__are given by Eqs (3.92) and (3.93). In a
turbulent flow, these rates of change (A and O) are
random variables. Their conditional expectations are,
of course, not random variables, but are (unknown)
functions that are determined by the initial and
boundary conditions. Thus, for a given turbulent flow,
denoting these conditional expectations by a and _0,we
have
<Aj(.x,t)lV, qt) = ajV(y.,~_,x,t), (4.4)
and
<O,(x,t)lV,~> = O,(V,~_,x,t). (4.5)
In order to stress that_a and 0 are not random variables,
we note that
<a/(y_.~,x,t)> = aj V(y__,~_,x_,t), (4.6)
and that
<0,(_y, ¢~,x, t)l V__= V', ~ = ~'> = 0, (_V',~', x_,t).
(4.7)
The transport equations for the joint lxlff (Eq. 3.104)
and for the mass density function #- (Eq. 3.105) were
derived from the conservation equations Eqs (4.1-4.3).
In terms of a(y_,!/fi~t ) and _0(y_,~,x_,t), the transport
equation for arW__,~, x__;t) is,
O~ O O 0
at = 1, o2.
(4.8)
This is a deterministic equation. Although #- is the mass
density function of random variables, and ai and O, are
conditional expectations of random variables,
nevertheless, for a given flow, #'(y_, ~, x; t), al(_V_,~, x_,t)
and OfC~,~0,~ t) are functions determined by the initial
and boundary conditions. This equation (Eq. 4.8)
shows how these deterministic functions are related.
A second important observation is that the dynamics
of the velocity and composition fields enter the con-
servation equations (Eqs 4.1-4.3) through the random
variables A(.~ t) and O(~, t). But, in the equation for #"
(Eq. 4.8), the dynamics enter only through the con-
ditional expectations _aand 0. This is significant because
many different random variables can have the same
expectations, and consequently, the same mass-
density-function equation can describe the evolution
of many different stochastic systems. Two systems
that evolve with the same pdf or mass density function
are said to be stochasticallyequivalent.
Stochastically equivalent systems are used in two
ways: to interpret the equation for the mass density
function, and to provide a method for its solution. In
both cases, a Lagrangian view is adopted.
 144 S.B. POPE

4.2. Lagraru3ian Equations

In presenting transport equations, we have adopted d~tt )
the Eulerian view--we have considered the fluid
properties at a fixed location x_.The starting point for the ,_'t,_o.,)
alternative-- Lagrangian-- view is the concept ofafluid
particle. Let _x+ ~o, t) be the location (at time t) of the
fluid particle that at time to(t o ~< t) was located at
x = x o. The fluid particle has velocity U + (~o, t) and
composition ~b+ (~o, t): these are just the velocity and
composition at x_+ ~o, t),

U+(~.o,t) = UFx+~o,t),t], (4.9)

le
and
q~÷(X_o,t) = ~b[x+(xo, t),t]. (4.10) FIG. 4.1. Fluid particle paths and material element volume
change in x t-x2-t space.
By definition, the fluid particle moves with the fluid
velocity,
Ox ÷(xo, t) • + . 8(x +)
~-------~ = U_+(~o,t). (4.11) ptx ,t} ~(_-~o)= p(xo, to). (4.18)

Differentiating Eq. (4.9) with respect to time we Figure 4.1 is a sketch of fluid particle paths (or
obtain pathlines) in x_-t space for a two-dimensional flow. The
fluid particle which is initially at x o follows the fluid
particle path to reach x+(xo, t) at time t. The fluid
particles which initially occupy the (two-dimensional)
volume d¥(t o), at time t occupy the larger volume d'C(t ).
This volume expansion is due to a decrease in density,
= (DUj~ (4.12)
\-6T/x.x,' Eq. (4.18).
The state of a given fluid particle (i.e. given Xo)
since Ox~'/dt is just the velocity Us. This confirms the is defined by its position x_÷, its velocity U ÷ and
interpretation of the material derivative DfDt as the its composition ~b+. At any time t, the state of the
rate of change following the fluid. Thus, (from Eqs 4.2- particle can he represented by a point (V_.,~,x_)=
4.3) the Lagrangian equations for U + and ~+ are (_U_÷, q~+, x ÷ ) in the (6 + a)-dimensional state space or
V-~b-x_ space; and (U +,~+,x_ ÷ ) is the state vector.
-~U+(x_o,t) =Aj[x_+(Xo, t),tl, (4.1.3) As time progresses, the state of the particle changes
and the point defined by the state vector moves in state
and space. The Lagrangian equations show that the state
vector evolves by,
~-~ ~b~+(Xo, t) = O,[.x+~o, t), t'l, (4.14)
~ = . 4.191
where A__and O_Oare
_ given by Eqs (3.92-3.93).
x_+ U___+
Consider an infinitesimal material element of fluid
that at time to occupies the volume d¥(to) = dxo at X_o. Thus the vector (A, O__,U__÷ ) represents the rate of change
The mass of this element is of state, or the velocity in state space. The evolution of
the state of the particle can be represented by a path in
d m = P~o, to) d¥(to). (4.15)
the augmented state space or V-~b-x_-t space. For given
At a later time t > to, the material element occupies the x_o, th~ particle path is
volume d'C(t) at x+(xo, t), and its mass (which, by
definition of a material element, is unchanged) is,
= / ~+(t) / ' (4.20)
dm = p[A+~o,t),t]d'¢(t). (4.16)
L-x+(t) J
The volume d¥(t) is related to d¥(to) by the Jacobian
If at ~ t) = (X_o,to) the flow is turbulent, then the
a(xf,xLx~) _~. . initial state
d¥(t) = ~ - 7 - ~ , a [to/
U[XoI,Xo2,X03)

-
O(x_÷ )
d¥(to). (4.17)
/ to)/' (4.21)
Lx÷(to) J LX-o J
From these three equations we obtain the Lagrangian is a random vector. The rates of change A and 0 are also
form of the continuity equation random vectors. Consequently, different realizations of
 Pdf methods for turbulent flows
~v,,~
Fro. 4.2. Fluid particle paths in the augmented state space.
the same turbulent flow generate different particle paths
in the augmented state space.
Figure 4.2 is a sketch of particle paths for different
realizations of the same turbulent flow. This sketch
can only be suggestive of the true paths, since the
(7 +a)-dimensional augmented state space is repre-
sented as a three-dimensional space. The following
observations can be made:
(i) in different realizations, the initial vahie of x = Xo is
the same, but the initial values of U__+ and ~+ are
likely to be different;
(ii) the particle paths can cross, since the state vector
(_U_+,~b+,x+) does not uniquely determine the
rate of change vector (A_,0__,U + ); and
(iii) the state vector changes in time and physical space
on all scales, down to the smallest (Kolmogorov)
scales.
4.3. Lagranoian Pdf
The Lagrangian conditional joint pdf considered--
fL(V_,~,x; t[V__o,q,o,Xo)--is the joint pdf of the fluid
particle properties at time t, conditional upon the
particle properties at the earlier time to. More pre-
cisely, fL is the joint probability density of the event
St = {U+ ~o, t)=-V, ~b+ (X_o,t ) = qJ,_x+ ~o, t ) = x},
(4.22)
conditional upon the event
So - {__U+(Xo, to) = __U(xo,to) = __Vo,
~b+(X_o,to) = ~(Xo, to) = ~o}. (4.23)
This is a quantity of fundamental importance and
usefulness since, as is shown below, it is the transition
density for turbulent reactive flows. That is, the mass
density function at time t can be determined from its
value at to by
;f ffL(V,q,,x_;tlVo,q,o,Xo)
x .~ (Vo, q'o, Xo; to) dVo d~'o dxo.
(4.24)
145
The evolution equation for fL is derived here using
the delta-function method. First--mainly to simplify
the notation--severalfine-grained pdf's are introduced.
These are just delta-function products that can be
thought of as pdf's in a single realization of the flow. 6°
The Eulerian fine-grained joint pdf is
f ' ( V , ~ ; x, t) ~ 3(U_U_(x,t ) - V)f(~(x, t ) - ~),
(4.25)
and its expectation is fV(y_,~_;x,t) (cf. Eq. 3.46). The
Lagrangian fine-grained joint pdf is
f ; (y_,~b,_x;tlX_o)= fi(_U_+ (X_o,t) - v)f(~b + (x_o, t) - E)
x 6(x + (Xo, t ) - x). (4.26)
This pdfis conditional upon the initial particle location
x o, but not upon its initial properties. The conditional
Lagrangian PdffL can be expressed as
fL (Y, ~k,x; t IV_o,~o, X_o)
= (f~(_V,~,x, tlX_o)f (V_o,¢_o,X_o,to)>/f(Vo,~_o,Xo,to),
¢ . ¢ .
(4.27)
(cf. Eq. 2.145).
The evolution equation for fL is derived by
differentiating Eq. (4.27) with respect to time. In order
to do this we need to determine the derivatives of
the delta functions. For the temporal derivative of
6(-U-+ ~o, t ) - V_)we have
o ~ ÷-y_) ou? ~ + - v _ )
gt Ot OU+
- o u ? o ~ +-v_)
Ot Ovi
-
?v;
0 E [ c3t 6(~-*-V)
}
- ov, { A ~ ( x + ' t ) ' ~ + - V - ) } '
(4.28)
The second step follows since an infinitesimal change in
U + ofdU + has the same effect on the delta function as a
change in V of d V_= - d U + ; the third step follows
because OU+/dt is not a function of V; and the final
result is obtained by using Eq. (4.13).
Differentiating Eq. (4.27) with respect to time (noting
that only f ; is a function of t) we obtain
Of
OtL = l --~--~/{Ai(x-+'
~ t)f~f'} >// f
+(
+ ( - ~ / { U : ( x _ o, t)f;f' I ) / . f, (4.29,
wherefL,f~,f' a n d f h a v e the same arguments as in Eq.
(4.27), and Eqs (4.11) and (4.14) have been used to
eliminate Ox_+/Or and O~b+/~t. Now for the term in A~we
have (see Eq. 2.150)
! 46
(Adx +, t)f~f')/f = (A,(x_ +, t)l S,, So) (f~ f ' ) / f
--- (X~(x_,t)lS, So)fL,
where the conditions St and So are defined by Eqs
(4.22-23). In the final step, x + can be replaced by x since
x + = _xis one of the conditions in Sv By applying the
same treatment to the other terms, we obtain the
evolution equation forfL :
8'fa ~ . g [ f t . V J + a [ft.<A, IS,,So>]
~t cxi ov~
0
+~[A<odS, So>]=0.
The evolution equation for ft. has a similar form to
that for :~-, Eq. (4.8). There are two important differ-
ences: ft. is a pdf rather than a mass density function;
and in Eq. (4.31 ) A and O are conditioned on S Oas well
as on St.
(It is interesting to note that the form of Eq. (4.31)
does not change if additional fixed conditions are
applied to (or removed from)ft.. [By definition a fixed
condition is independent of t.] Let C denote a set of fixed
conditions and let f~(V,~O,x; tlC, xo) be the corre-
sponding Lagrangian conditional pdf. Then the evolu-
tion equation for f, is the same as Eq. (4.31) with f~
replacing ft. and C replacing So.)
4.4. Relationship between Lagrangian and
Eulerian Pdf' s
In this subsection fL is shown to be the transition
density (by establishing Eq. 4.24, and then further
relationships between ft. and .~- are demonstrated.
Consider the quantity
wherefp a n d f ' have the same arguments as in Eq. (4.27).
It is now shown that the expectation of.~" is equal to
each side of Eq. (4.24), thus establishing the equation's
validity.
Simply taking the expectation of Eq. (4.32) yields
(.~" (V_,~,x, t)) = (~ko)fL(V,~,x; tlVo,~ko,Xo)
x f (Vo, ~bo; x o, t o) dV o d~o dx_o
x ~(Vo, ~ko,X_o;to) dVo d~kodxo,
which is the same as the right-hand side of Eq. (4.24).
To establish that .~' is the fine-grained mass density
function (as is implied by the notation) is less simple. In
Eq. (4.32), the integral over the particle positions at time
to(j'dxo) can be replaced by the integral over their
positions at time t
S.B. POPE
f g(x°) dx-.
(4.30) g(x+) -
Now, the Jacobian is known from the Lagrangian
continuity equation Eq. (4.18) in terms of the densities.
Thus, Eq. (4.32) can be rewritten
"'(v__,C,,x;t)=fffp(~o)p(~_+)/p(~DXo,
d)
x ~ ( u + -v_)~(p_- -{z)
x ~(U[xo, to] - Vo)
(4.31) x ,~(4,[Xo, to] - ~ o )
x 6(x + -x_)dVod~odx +. (4.34)
The fight-hand side of this equation can be deter-
mined by integrating over V_o,~o and x + in turn. Only
6(_U_~o,t)-Vo) depends upon V o and its integral
over the whole of Vo-space is unity. Only P(~o) and
6(~b~o, to)-~ko) depend upon t#o, and in view of the
sifting property of the delta function (Eq. 2.32) we
have
fP(~o) {p(~rAo, t] } 6(~_[x o, t] - ~_o) d~o = 1.
(4.35)
Thus, Eq. (4.34) reduces to
~-'(v_, ~, x; t) = fp(~+)a(u + -_v)6(~ ÷ - ~)
x 6(x + - x ) d x +. (4.36)
Again in view of the sifting property, the only contri-
bution to this integral is from fluid particles located at
x + = x. And at this location, the Lagrangian properties
(e.g. U ÷ (xo, t)) can be written as Eulerian variables (i.e.
U(x, t)). We thus obtain
~r'(_y.,£, x_;t) = p(~)6(U(x, t)-V)6(~(x, t ) - ¢ ) ,
(4.32) (4.37)
and taking the mean
(,~"(v_y,~, x_; t)) = p(~)f (V__,£; x, t) = .T(V_, ~_,_x;t).
(4.38)
Equating the two different expressions for ( . T ' ) (Eqs
4.33 and 4.38) yields the required result
~ V~,,x;t)=f f ffL(V.~,x;tlVod,o,Xo)
x ff(Vo, ~bo,Xo; to)dVo d~odx o.
(4.39)
This identifies ft. as the relevant transition pdf: it is the
probability density of the fluid particle's transition from
(4.33)
the initial state (_Vo,~o,Xo) at time t o to the state
(V_,~k,x_) at time t. The mass density at time t is
(according to Eq. 4.39) the mass-probability-weighted
integral of this transition pdf over all initial states.
The Lagrangian pdf ft. has been conditioned on
Vo,~o and x o. In fact, the conditioning on V o is not
needed in order to obtain a relation similar to Eq. (4.39).
 Pdf methods for turbulent flows
The conditioning on ~o is needed in order to relate the
Jacobian d(x +)/0(xo) to densities. But in constant-
density flows (in which the Jacobian is unity) only the
conditioning on x o is needed. In this (constant-density)
case Eq. (4.39) reduces to
.f(V_,~ ;x_,t)= jlfL(V__,~h,x_;tlX_o)dxo.
It is useful to be able to derive Eulerian pdf evolution
equations from Lagrangian equations. This can be done
by differentiating Eq. (4.39) and substituting for OfL/dt
from Eq. (4.31):
a:(v__,~,x_;t)
?t
a [ft<@,lS.So)]}
x o~(Vo, ¢o, xo; to id_Vo d~o dx_o. (4.42)
It may be noted that, for each of the three terms, the
derivative (~/~x~,~/~V~or ~/0¢~) can be removed from
the integral, since ~(V_o,~o, Xo; to)dV o d~bo dx o does
not depend upon the state variables (V,~k,x). The
resulting integrals are then of the form
We now show that this integral is equal to
<Q(x_, t)lV__,q,> :(y_, ~,_x; t).
Multiplying Eq. (4.32)by an arbitrary function Q(~, t)
and taking the mean we obtain
<Q,~ (V_,q/,_x,t)> = p(~o)(Q.f~f >dVod~odx-o.
- fll
Having identified : ' as the fine-grained mass density
function, the left-hand side can be re-expressed as
<Q:~'> = <Qpf'>= <Q(x_,t)lv_,gL>p(g£).f(V,~_;x,t)
= <Q(x_,t)lV__,@_>~(V,~,x_;t ).
(The second step is analogous to Eq. 2.150.) Similarly,
the right-hand side of Eq. (4.43) is
(f f p(~o)<~Lr>dV_od~odxo
× ~(Vo, qJo, Xo; to)dVod~odx_ o.
Thus, combining the last three equations we obtain the
required result:
(f f <QlSt.So>f,:(V_o.~o,X_o:to)dY_od~_odx_o
= <Q(x_,t)]V,~>~(V__.~,x_: t).
147
With this result, Eq. (4.42) becomes
a~- ?
a-S- = - ax--~.[ ~ v , ] -o-v, [~<A'IV-'~>]
a
a0 [&~<®~lv_y,¢>]. (4.47)
(4.40)
This is the evolution equation for the Eulerian mass
density function ~- derived from the Lagrangian
equations for the evolution of fluid particle properties
(_U_+,q~+,x+). Inevitably the result is the same as Eq.
(4.8) which is derived from the evolution equations for
the Eulerian properties U and ~b.
4.5. DeterministicSystem
It has already been noted that the joint pdf transport
equation, Eq. (4.8), is deterministic. In this subsection
the equation is interpreted in terms of a deterministic
Lagrangian system.
In the joint pdf equation, the rates of change A~, t),
O(_x,t) and U(3.,t) appear as their conditionally
expected values:
l
-0v(-Y-'~'x' t)J w-L<U__~,/<--O('t)[V--'
t)lV_,
~' ~~k>J
>/" (4.48)
For a given flow, the conditional rate of change vector
(~0,V) is an unknown but nonetheless determined
vector function in state space.
Consider now the deterministic Lagrangian
equation
(4.49)
(4.43) In this case the state vector ~ _ ~ , ~) does not pertain to a
fluid panicle, since fluid particles evolve by a different
equation, Eq. (4.19). Equation (4.49) describes the
evolution of the state of notational quantities called
conditionally-expected particles, or conditional
(4.44) particles, for short.
From a given initial state (V_o,~bo,x_o)at time t -- to, a
conditional particle follows a path determined by Eq.
(4.49). In order to show exphcitly the dependence on the
initial state, we write this path as,
xo,,/ (4.50)
I_~(_Yo,g~o,X_o,t) d
In the augmented state space (_V_-~-x-t-space), this
equation defines the conditional path from the initial
state (V o, ~o, Xo).
(4.45) For the fluid particle with state (_U_+, q~+, x ÷ ), its rate
of change (A_,O_,U ÷ ) is a random variable (although the
x-component U + is, of course, known). Thus, in a
subsequent repetition of the flow, a fluid particle may be
found with the same state (_U__+,tb+,x+) but with a
different rate of change (A__,O, U ÷ ); see Fig. 4.2. For the
(4.46) conditional particle with state (O,~,~), its rate of
148 S.B. POPE

t The integration dV 0 dq/o dxo over the initial states can

be replaced by the integration dO__d~ d~ over the final
states, multiplied by the Jacobian:

ff
x {:(Vo, 6'o,X_o;to)
) /f
f
x ct(V°: ~°' x°)"(,':d O"d~ d~. (4.53)
atu~ ~_,x~ ) -
to I~ w,
The left-hand side of this equation can also be written as
IVo, ~.) IV,qJ) an integral over the final states, namely

FIG. 4.3. Conditional path from (Vo, ~o, Xo) at t = t o in the

augmented state space. ff
× {~,~(~,_~; t)}dOdq~di. (4.54)

change ($~__0) is not a random variable: it is the Now, since both these equations for ~ must hold for all
conditional expectation of the rate of change of fluid states (V, ~, x), the two terms in braces must be equal:
particles with the same state. (Yo, ~o, X_o)
The conditional path from the state (_Y_o,C/o,X_o)at :(0~_~, i_; t) = .~(_vo, ~o, X_o;to)
t = t o is sketched on Fig. 4.3. Again, this sketch is only (4.55)
suggestive since the (7 +a)-dimensional augmented
state space ~_-~-x-t-space) is shown as being 3- or, dividing by the Jacobian,
dimensional. The following observations can be made:
: .~(__V0,~o, Xo; to).
(i) the conditional path is uniquely defined by the
initial state; (4.56)
(ii) since the state vector uniquely determines the rate This formal solution, Eq. (4.56), has a simple
of change vector (.~ 0, O)conditional paths cannot geometric interpretation in terms of expected mass and
cross; and conditional tubes. Let So be the surface in state space that
(iii) the state vector changes in time and physical space encloses a (6 + a)-dimensional volume ¥o. At the time
on scales characteristic of the larger scale turbulent to, the quantity
motions.
While the behavior offluid particles and conditional
particles are quite different, their (one-point) statistics o

(4.57)
are identical. This is simply because in each case the
evolution for the mass density function ~ is Eq. (4.8). represents, the expectation of the mass within the
Thus, for a given flow, if initially (at time to) the particular region of state space ¥o: that is, m(to) is the
conditional particles are distributed in state space expected mass in ¥o at time t o.
identically to the fluid particles, then the conditional In the augmented state space (V-~k-_x-t), the condi-
and fluid particle distributions remain identical at tional tube (associated with ¥o) is defined to be the
future times. surface formed by all the conditional paths that at t = t o
For conditional particles, the transition density is
ft.(V, ¢/, x; ttV0,~o, Xo)= 3(_y_.-0")~(~- ~)5(x- i_),
(4.51)
since, by definition, (_0,0~,i_) is the state at time t of the v(t)
conditional particle that had the state (Y-o,~o, Xo) at
time t o. Thus, from Eq. (4.39), the mass density function
at time t is determined by its value at time t o and by the
conditional paths:

x_;t) = fff.(v_o. ¢,o.Xo;to)6(v-_o) "~~o

- -- path

× 6(~-~_),~(x-~_)dVodV, odxo. (4.52)

Because of the particular nature of the transition to
density, Eq. (4.51), this formal solution can be expressed v2,±
locally, rather than as an integral over all initial states. FtG. 4.4. Conditional tube in augmented state space.
 Pdf methods for turbulent flows 149

pass through the surface So. Since, by definition, the :(Y_o, 0o, Xo; to) ~ :N(Vo,~k0,X_o; to)
tube is everywhere tangential to conditional paths, a N
conditional particle cannot cross from inside to outside - A m ~ ~(V_o-U(O"')
n=|
the tube or vice versa. A conditional tube is sketched on
Fig. 4.4. Because x_and (V_,~k)are each represented by a x 5(~Lo-~b~n))&(Xo-_X~). (4.63)
single coordinate, the surface So appears as a closed Each particle represents a mass Am and (at time to ) (U~ ~,
curve, and the volume ¥o appears as a two-dimensional ~(O"),x~ )) is the state of the nth particle. Replacing ,~ by
area. ~-N on the right-hand side of Eq. (4.52) yields
At the later time t > to, the conditional tube occupies
the volume ¥(t) in state space, see Fig. 4.4. The expected :(y_,~,x_; t)
mass in ~(t) is

m,t,=ff;,~,,,.~(V,~,x_;t)dVd~dx_. (4.58,
Am ~,
n=l fff
t
~(-Y-o-- U---(o"))~(~Lo- ~b(o"))~fx-o-X-(on))

× ~(v_- O[Vo, ~o,_Xo, fl)~(~ - ~[V_o, ~_o,_xo,fl )

There is a orie-to-one correspondence between x 6(x_- x_[~o, ¢'o, X-o,t]) dV_o d~o dxo
the points in ¥o, (Vo,~o, Xo), and points in ¥(t), N

(O, ~, ~_). Consequently, the integral over ¥(t) can be = A m ~ 6(V-O~")[t])6(~_-~_(")[t])&(x_-~jn)[t]),

replaced by an integral over ¥o : (4.64)

m(t)=f f t_o, d/o, -o ) dV°

-- d ' ° dx°"
where
r O--'")(t)-~ r| n,l.n) ~., ,~--0
~-'L.V-O ,_Yf_O .~.,~*!.,.
(4.59)
I -- ¢
.,,., .,.,.,
t)
Now the formal solution to the pdf equation, Eq.
- L ~ - ~t-w-o ,~/£o , " o ,-/
(4.56), states that the integrals in Eq. (4.57) and (4.59)
(4.65)
are equal. Thus we obtain
re(t) = re(to). (4.60) The state FO~')(t),_~')(t), ~j')(t)] is the state at time t of
the conditional particle that at time to had the state
rU(o"1,~(o"),X(o")].Thus Eq. (4.64) shows that for all time the
Expected mass is constant along a conditional tube.
mass density function can be approximated by a large
The same analysis applies to an infinitesimal condi-
number of conditional particles.
tional tube. In an obvious notation we have
Through Eq. (4.64) we have reduced the problem of
dm(to) _= :(V_o,~o, Xo; to)d¥o, (4.01) determining Jr (_y,~k,_x; t) to that of determining the
din(t) - ff(V,~(,x; t)d¥(t) states of N conditional particles. Suppose that the
unknown conditional rates of change a(_V_,~,x_,t) and
_0(y_,~,x_,t) can be modelled as simply computed
= x_;,)atv_ , X-o)m 0 functions in state space. Then the particle states can be
= dm(to).

The Jacobian represents the volume ratiod¥(t)/d¥o for

the infinitesimal conditional tube. If the conditional
(4.62)
ro,.,)
determined by solving the equations

paths converge, d¥(t) decreases and the Jacobian isless /o_,', - J

than unity. In order that the expected mass remains (4.66)
constant, the mass density ~ must increase. Con-
from the initial conditions (~_~o"~,~b(0"),x_(o"~).For each
versely, if the conditional paths diverge, d¥(t) increases,
value of n, Eq. (4.66) is a set of ( 6 + a ) simultaneous
the Jacobian is greater than unity, and the mass density
ordinary differential equations. To solve these
decreases.
equations numerically is trivially simple compared to
There is a direct analogy between Eq. (4.56) for the
solving directly the equation for ~r (Eq. 4.8) which is a
mass den sity along a co nditional path, and Eq. (4.18 ) for
partial differential equation in (7 + a) dimensions.
the fluid densit y along a fluid particle path. This analogy
The conclusion that the pdf equation can be solved
may also be seen by comparing Figs 4.1 and 4.4. The
economically via Eq. (4.66) is based on the supposition
analogy arises because expected mass is conserved
that a and 0 can be modelled as simply computed
along conditional tubes, and mass is conserved
functions* in state space. In Section 5.2.2 such simple
following a material element: mass density is expected
deterministic models are examined and it is shown that,
mass per unit volume (in the state space), and the fluid
in general, they are unsatisfactory. Satisfactory deter-
density is mass per unit volume (in physical space).
ministic models can be constructed, but they are far
These considerations of conditional paths lead to a
computationally feasible method of solving the joint
* For the present purpose, a "simply computed function"
pdf equation. At the initial time t o, the mass density can be defined to be a function of V, ¢,, _x and t that can
function is approximated by a large number N of depend upon the first few moments of t~e pdf, but is other-
particles in state space (see Eq. 3.76): wise independent of the pdf.
150 S.B. POPE

from simple, and the computational advantages of Eq.

(4.66) are lost. Satisfactory modelling combined with
computational economy can be obtained through
stochastic models. These are discussed in the next U~

subsection and in Sections 5.2.3-4.

4.6. Stochastic Systems

Many different systems evolve with the same pdf.

Fluid particles and conditional particles behave quite
differently, but nevertheless their pdf's are the same. We
now consider systems of stochastic particles whose
behavior is physically and mathematically different , I , I , I ,
from that of fluid or conditional particles. The essence of 0 2 4 6
the whole approach is to construct a system of I/T

stochastic particles whose evolution is simply com-
puted and in which the pdf evolves in the same way
as the pdf of fluid particles.
We consider two types of stochastic processes.
Poisson processes form the basis of particle interaction
models (Section 5.3); and diffusion processes form the
basis for Lanoevin models (Section 5.5).
At time t, the state of a stochastic particle is denoted
by (_U_*(t), ~b*(t), x*(t)). This is a (6+a)-dimensional
state vector that corresponds to a point in state ~ace.
The state of the particle evolves according to a
stochastic prescription, and the corresponding point
moves--not necessarily continuously--in the aug-
mented state space (V, ~,, x, t).
In order to examine Poisson and diffusion processes
it is sufficient to consider a single (scalar) state property.
Let U*(t) be the state property at time t and let f*(V; t)
be the corresponding pdf. The stochastic process U*(t)
can be defined either in terms of the increment
A~,U*(t) -= U*(t+6t)- U*(t), (4.67)
or in terms of t he transition density f a ( V; t I Vo). This is
the probability density of the event U*(t+6t) = V
conditional upon U * (t) = Vo. In both cases we consider
the limit as the (positive) time interval, fit tends to zero.
(Note that A~, is the forward-difference operator.)
For a Poisson process, the increment is:
with probability 1 - fit~z: AotU*(t ) = 0,
with probability 6t/z : Af, U*(t) = U°(t)- U*(t),
(4.68)
where z(t) is a specified time scale and U°(t) is a random
variable with a specified conditional pdfo(V I Vo, t).
A sample of such a process is shown on Fig. 4.5. The
conditional pdf in this example is
g(V, Vo, t ) = fi ('V - 9 Vo) ., (4.69)
and the time scale z is taken to be constant. It may be
seen from the figure that U*(t) is constant except at a
finite number of points where it changes abruptly. In
this example, the value of U* is reduced by 10 ~o at each
such change. The time interval between abrupt changes
is a random variable: the probability density of the time
interval being At is r - 1e-a,/~ (see Ref. 55).
FZG.4.5. A sample of the Poisson.process U*(t).
An evolution equation for f * ( V ; 0 - - t h e pdf of
U*(t)--is now derived. Because U* is a discontinuous
function of time, the methods of derivation used in
Sections 3.5 and 4.3 cannot be employed. However the
evolution equation is readily derived from the
transition density through the relation
f*(V; t+fit) = ~f*(Vo; t)fo,(V; t I Vo) dVo,
w
(4.70)
(cf. Eq. 4.39). The transition density for the Poisson
process, corresponding to Eq. (4.68), is
J~,(v; tl Vo) = (1 -fit/z)fi(v- v o) + (fit/z)o(vI v o, t).
(4.71)
Substituting forJ~ t in Eq. (4.70) yields
f*(V; t +fit) = (1 -&/~)f*(V; t)+ (fit/~)
x ff*(Vo;t)g(VlVo, t)dVo. (4.72)
Hence in the limit as at tends to zero we obtain the
evolution equation
a/,(v; t)
= _I lf,(Vo; t)g(V[ Vo, t) dVo - f * ( V ; t)/r.
at z
(4.73)
It may be seen that although U*(t) evolves
stochastically and discontinuously, its pdf f*(V;t)
evolves deterministically and smoothly (provided that
neither f*(V; to) nor 9(V] Vo, t) is pathological).
The usefulness of the Poisson process is discussed
below, but now we turn our attention to diffusion
processes. A necessary preliminary is to describe the
Weiner process.
Let the time interval from t = 0 to t = T be divided
into N equal subintervals of duration at = T/N, and let
¢~,1 (n = 1, 2 ..... N) denote N independent standard-
ized normal random variables; thus
(~.1) = 0, (4.74)
and
(~.~,,) = 6.,.. (4.75)
 Pdf methods for turbulent flows 151

0.25 I

W(ln)
O
e

-0.25

_,.o

-0.50
I
0 0.5 1.0
; I I I
t
0 1.0
F1G. 4.6. A sample of the function W(t.). t

FIG. 4.7. Sample of the Weiner process W,.

Consider the quantity

ts < t., [W,~ - W~,] and [Wt. - W J are independent
Gaussian r a n d o m variables with zero means and
W(t.) - (6t) 1/2 ~ ~,), (4.76)
i=1
variances (t z - t 1) and (t 4 - ts) respectively.t
The increment
where t. = nAt. Alternatively, W(t.) can be defined by
the increment An, W, = Wz+6,- I41,, (4.80)

A~tW(t.- 1) = (6t)l/2~t.), n = 1,2 . . . . . N, is equal to fit 112 multiplied by a standardized Gaussian

random variable (el. Eq. 4.77). Consequently, as fit tends
(4.77) to zero, so does A~, I41,,showing that Wf is a continuous
with the initial condition W(to) = 0. Figure 4.6 shows function. On the other hand (A~, 14/,)/fit becomes infinite
W(t.) for T = 1 and N = 50. and hence Wt is not differerttiable.
The mean and variance of the difference The general diffusion process U*(t) considered has
W ( t . ) - W(t.,) (m < n) is readily deduced from Eqs the increment (as fit tends to zero)
(4.74-75): A6tU*(t) = D(U*[t], t)6t + [B( U*[t], t)] 1/2A~tWt,

( W ( t . ) - W(tm)) = (~t) ~/2 (~,~> = O, (4.81)

i=m+l
and
<[w(t.)- w(tm)Y>= 6t ~ ~ <¢,,,¢u)>
i=m+lj=m+l
i = m + l j = m + l
=6t(n-m) = t,-t..
Thus the difference W ( t . ) - IV(t.,)is a Gaussian random
variable with zero mean, and variance equal to the time
difference t . - t,~. This last result, it may be noted, is
independent of the choice of N. By a similar procedure it
is readily shown that (for m < n < p < q) the r a n d o m
variables [ W ( t . ) - W(tm)] and [ W ( t ~ ) - W(tp)] are un-
correlated and hence independent (in view of their
Gaussianity), with means of zero and variances ( t . - t . )
and t t q - tp) respectively.
The Weiner process, denoted by W,, can be thought of
as the discrete process described above in the limit as 6t
tends to zero (N tends to infinity). A sample is shown on
Fig. 4.7. In fact. it is better to define the Weiner process
by the properties deduced aboveTS: for 0 ~< t~ < t 2 ~<
where D and B are specified functions, B being non-
(4.78) negative. This increment is a Gaussian r a n d o m variable
with conditional mean
(A~,U*(t)IU*(t)= Vo)=D(Vo, t)t~t, (4.82)
and variance
<[A~,U*(t)- D(Vo, t)6t]2lU*(t) = V0> = B(Vo, t )6t.
(4.83)
(4.79)
Consequently the transition density f~t(V; tIVo) is a
Gaussian pdf with mean Vo + D(Vo, t)6t and variance
B(Vo, t)6t.
Since U*(t), Ffice IV,, is not differentiable, the evolu-
tion equation for its p d f f * ( V ; t) cannot be derived by
the methods of Sections 3.5 or 4.3. The standard
method ~4'7s involves g6,(AVI V0, t)--thepdfofA~,U*(t)
conditional upon U*(t) = Vo. In general, .q~ is simply
related to the transition density by
g~,(AVIVo, t) ----A,(Vo+AV; tl Vo), (4.84)
i" It follows from the central limit theorem that for a finite
time interval (t 2 - t 1), the increment [W,= - IV,,] is Gaussian
even if the infinitesimal increments ¢t,~ (Eqs 4.74-4.75) are
not.
152
and, in the case being considered, 9n, is Gaussian with
mean D( Vo, t )6t and variance B( Vo, t )6t.
Either from first principles or by combining Eqs
(4.70) and (4.84) (with Vo = V - A V ) we obtain
f*(V; t + 60 = f f * ( V - AV; t)O6,(AVl V - AV, t) d(AV).
(4.85)
The right-hand side can be re-expressed as a Taylor
series:
f*(V;t+6t)=f{f*(V;t)Ont(AV, V,t)
- AV O(f*g6')
OV
1 2 , )
tAV~20 (f*gn,) ~d(AV), (4.86)
+~, • OV2 ...
where f *96, is written for f * (V; t )06, (A V[ V, t ). Now the
derivatives and f * can be removed from the integral,
and the remaining terms can be identified as conditional
moments of An, U*(t):
f (A V)"an,(A V I V, t) d(A V) -- ( [An, U*(t)]"l V).
(4.87)
From Eqs (4.82-83) it may be seen that the first and
second moments are of order 6t while higher moments
are of smaller order.t Thus, to first order in 6t,
f*(V; t + 60 = f*(V; t ) - - ~ [f*(V; t)D(V, t)bt]
+~ ~ [f*(V; t)B(V, t)bt],
and hence in the limit as 6t tends to zero we obtain the
evolution equation for f * :
Of*(V;t) 0
. . . . [f*(V; t)O(V, t)]
0t 0V
1 02
+ 5 ~V5 [ f * ( v ; t)B(V,t)]. (4.88)
The evolution equation for the pdf in a diffusion
process, Eq. (4.88), is called the Fokker-Planck equa-
tion and has found wide application in many
fields. 73-76 The functions D(V, t) and B(V, t) are called
the coefficients of drift and diffusion. If D is uniform,
then its effect is to cause the pdfto movein V-space at the
speed D. If B is uniform (and D is zero) then Eq. (4.88)
reduces to the unsteady one-dimensional heat conduc-
tion equation:
Of* 1 B ozf*
Ot = 2 ~ " (4.89)
From this it may be seen tfiat the term in B causes the pdf
t These conditions on the moments are sufficient for Eq.
(4.88) to follow. The infinitesimal increment A6tU*(t) does
not have to be Gaussian.
S.B. POPE
to diffuse in V-space. (The diffusion process Eq. (4.81)
causes the pdf to diffuse in V-space. This should not be
confused with gradient-diffusion models of turbulent
transport which cause diffusion in physical space.)
For the particular case in which B is uniform and D is
linear in V,
D(V, t) = G(t)V, (4.90)
the general diffusion process (Eq. 4.81) reduces to
A6,U*(t ) = G(t)U*(t)6t + [B(t)] t/2A6,W,. (4.91)
This is the Langevin equation. Originally proposed s ~to
model the velocity of a particle undergoing Brownian
motion, the Langevin equation and its generalizations
have been widely studied.73- 7s
We now consider, in general terms, the modelling of a
physical process by a stochastic process. Let U(t) be a
physical quantity (e.g. one component of velocity) and
let A(t) (a random variable) be its rate of change. The pdf
of U(t)-f,(V; 0--evolves according to
Of,(V; t) _ O
Ot OI,; [a(V, t)f.(V; t)], (4.92)
(el. Eq. 4.8) where
a(V, t) - (A(t)l U(t) = V). (4.93)
For a Poisson process, U*(t) is a discontinuous
function of time, and for a diffusion process it is
continuous but not differentiable. Clearly then, neither
of these stochastic processes provides a good direct
model oftbe physical process U(t ). But the pdf's of U * (t)
and U(t) can evolve identically and hence f*(V; t) can
be a perfect model off, W; t). The conditions that have
to be satisfied for this to be the case can be deduced by
comparing Eqs (4.73), (4.88) and (4.92). For the Poisson
process, the time scale ~ and the pdf0 must be such that
xo(VlVo, t)d dV V,t)]. (4.94)
For the diffusion process, the functions D and B must be
such that
1
a(V,t)= D(V,t)-~[f,(V; t)]_lO_J_~[fu(V;t)B(V,t)].
(4.95)
And of course in both cases the initial conditions
f~(V; to) andf*(V; to) must be the same.
At the end of the last sub-section a computationally
feasible deterministic method of solving the joint pelf
equation was outlined. In the present context the
method is to approximate the initial pdf f,(V; t0) by a
large number of conditional particles in V-space:
1 N
f.(V; to) ~ -- ~ 6 ( V - O("~(to)). (4.96)
Nn=i
It was supposed that a(V,t) could be modelled as a
simply computed function, and hence the evolution of
 Pdf methods for turbulent flows 153

the conditional particles could be simply computed by The result is:

solving the N ordinary differential equations
Offt.= _vjOft Ms(x_,t) Offf.
d0q"}(t) = a(0~"l[t], t). (4.97) Ot Oxs OVj
dt
Although the selection of the initial conditions U(")(to)
--Gsk(X,t)o-O-~(f~.Vk)+~B(x,t)~i
may involve randomness, the subsequent evolution
according to Eq. (4.97) is deterministic. + ! { f A*(V_,~o,X_;t~(~l~o,X_,t )d~_o- f~}.
The pdfequations arising from the stochastic models,
Eqs (4.73 ) and (4.88), can be solved either by a determin- (4.103)
istic method or by a r a n d o m - - Monte Carlo--method. Both stochastic processes considered are Markov
The deterministic method is as described above, with processes. That is, given the current state U*(t), the
a(O(")[t], t) in Eq. (4.97) obtained from Eq. {4.94) for a future state U*(h) is independent of the past U*(to)
Poisson process, or from Eq. (4.95) for a diffusion (t o < t < t~). Consequently, 7~ the conditional pdf
process. But this method is not computationally feasible fz*(Y_, ~0,x; tlVo,~bo, Xo) evolves according to the same
since the integrals or differentials of f required by Eqs evolution equation as the unconditional pdf, Eq.
(4.94-95) are extremely difficult to compute from the (4.103). In fact, the conditional pdf can be determined
discrete representation. as the solution of this equation with the initial
The alternative--M onte Carlo--method is straight- condition
forward and computationally efficient. The initial pdf
f*(V;t0) is represented by a large number N of fL*(y_,~,X; to)=f(y_--Vo)t$(~--Ipo)f(_X--X_o). (4.104)
stochastic particles in V-space:
As before, the evolution equation for the (Eulerian)
1 N
mass density function ~ar*(V,~b,x_;t) can be obtained
f*(V; t 0) ~ ~-,~1= 6 ( V - U*t")(to)). (4.98)
from the conditional Lagrangian pdf via Eq. (4.39).
Differentiating this equation with respect to time and
The stochastic particles then evolve according to the
substituting the right-hand side of Eq. (4.103) for Of~./at
stochastic processes: Eq. (4.68) for a Poisson process,
yields:
Eq. (4.81) for a different process. (These equations are
simple to solve numerically.) Then, provided that the
modelling is accurate (i.e. Eq. (4.94) or (4.95) is satisfied), at = - ax--~.[ v s ~ * ] - [{Ms + Gsk V~}.~-*]
f*(V; t) is equal to f,(V; t) for all time.
1 a2~at* 1
This algorithm is called an indirect Monte Carlo
method 37-indirect because U* (t) does not model U(t)
directly.
We now revert to the general case in which a
stochastic particle has the (6+cr)-dimensional state
vector U* (t), ~b*(t), x* (t), with corresponding (uncondi- (4.105)
tional) Lagrangian pdff~V_, ~O,x; t) and Eulerian mass Although fff and ~'* have distinctly different
density function 3r*(y_,~/,x; t). In order to illustrate properties, it may be seen that their evolution equations
how the equations forf~and ~-* can be derived, we take (Eqs 4.103 and 4.105) are identical.
the following process as an example:
A~,x~(t) = U*(t)ft, (4.99)
4.7. Normalization, Mean Continuity and Pressure
A~fU*(t) = M j(x_*[t], t)St + Gjk(X_*[tl,t)U*(t)ft
The mass density function ~r(V__,~,x; t) satisfies the
+ [B(x_*[t], t)] 1/2A~t(Ws),, (4.100) following conditions:
with probability 1 - 6t/z: Ao,q~*(t) = 0, ~-(y_, ~O,x_;t) I> 0, (4.106)
with probability bur : A~¢~ (t) - dp~
, ~ 0
(t) - q~ (t).
.

ff~,~v(y_,~h,x_;t)dVd~=<p(x_,t)>,(4.107)
(4.101)
Here, M~(x, t), Gsk(X,t)and B(x, t) are specified func- and
tions; ~ is an isotropic Weiner process so that
<A$t(Wi)tA,~t(W j)t> = 6t611; (4.102) ff ~(V_, ~b,x; t) [p(~b)l- XdVd~ = 1, (4.108)

(see Section 3.3). Conversely, any function 3~-*(V, ~,, x;

and ~b° is a random a-vector with specified conditional
joint pdf O(~_l~_o,X_,t)--conditional, that is, on
~b*(t) = ~O0and x*(t) = x.
By the methods presented in this section, the
evolution equation for the Lagrangian joint pdf
fL*(V_,~,x;t) can be derived from Eqs (4.99-101).
t) satisfying these conditions is a valid mass density
function.t
1"~'* denotes any mass density function that approxi-
mates the fluid mass density function ~.
154 S.B. PoPE

We now ask: starting from a consistent initial con- conditions are satisfied if, and only if,
dition ~*(V_,~O,X;to), what conditions must be
<p(~[x_, t])) = q(x, t). (4.114)
satisfied by a modelled pdf equation in order that #'*
remains a valid mass density function? This ques- Hence proposition (ii).
tion is answered by establishing the following The terms to be modelled in the evolution equation
propositions. for ,~* represent transport in V - ~ space. Conse-
quently, when any modelled evolution equation for
(i) Any model that can be expressed in terms of the
,,at, is integrated over all V and ~b the modelled terms
evolution of the discrete mass density function
vanish leaving:
~-N satisfies the realizability condition, Eq.
(4.106).
(ii) The consistency condition, Eq. (4.107), and the
normalization condition, Eq. (4.108), are equi-
atLf f
valent (one implies the other). + ~-~x~ ~'*(V, ~0,x_; t)V~dV d~ = 0. (4.115)
(iii) The satisfaction of the mean continuity equation
a<p> a[<p>Oj Evaluating the integrals using Eqs (4.111-i 12) yields
-- ¢- - = O, (4.109)
at ax~ aq+ ~
-~- ~ x (qU~)=o, (4.116)
is a necessary and sufficient condition for the
satisfaction of the consistency condition. or
(iv) The mean continuity equation is satisfied for all 13 aO ,
time if, and only if, the mean pressure satisfies a 13--~tn q = - ax-----[' (4.117)
Poisson equation, Eq. (4.129).
where
Thus, if the model can be expressed in terms of the
evolution of the discrete mass density function, and if I3 a ÷0iaUx~ (4.118)
the mean pressure is correctly determined (by Eq. 13t - at
4.129), then the modelled mass density function # ' * is Rather than assuming the satisfaction of the mean
valid (i.e. it satisfies Eqs 4.106-108). continuity equation (Eq. 4.109), we define the function
The discrete representation of the mass density
X(~, t) by
function, Eq. (3.76), is a positive quantity (Am) multi-
plied by the sum of delta function locations moving in
<p>z(x_,
a<p>e [<P>O']
t) - -Ti- ax~ (4.119)
state space. But irrespective of the delta functions'
locations (i.e. irrespective of the modelling) the delta Then X -- 0 is a necessary and sufficient condition for
functions are non-negative and hence proposition (i) the continuity equation to be satisfied. Obtaining a 0 /
is established. Thus in pdf methods, realizability is ax~ from this equation and substituting into Eq.
guaranteed in a simple, natural manner. (4.117) yields:
Rather than assuming Eq. (4.107), we define
131nq 13In <p>
13~ = 13~ Z. (4.120)
q(x,t)-ff.*(V_,g,,x;t)dVd~, (4.110)
Since q and <p> are initially equal (for all x), then it is
and then define obvious from Eq. (4.120) that they remain equal if and
f(v,~;x_.t) - ,~*(v_,~,x_;t)/q(x,t). (4.111) only if ~ = 0. Thus proposition (iii) is established.
The satisfaction of the mean continuity equation for
Since f is non-negative and integrates to unity, it is a all time is equivalent to the satisfaction of
joint pdf. Density weighted means are defined (at any
x, t) by a_~_~= 0, (4.121)
~t
ff (4.112) with the initial condition Z(x_,to) = 0. Differentiating
Eq. (4.120) with respect to time we obtain
These are consistent definitions that assume Eq.
(4.106) but not Eqs (4.107-108). az a2 O~a__ a
Now with Q(_U_,~) = [p(q~)] - ~, Eq. (4.112) yields a--[=a-[i { < p > - q } + ax, ~{<P>-q}
+OOi a
~T ~ ~<P>-q) (4.122)
Suppose that at time t the consistency condition
=ff[PtrD]-t~*(v_Y.,qL)dVd~q. (<p> = q) and the continuity equation (X = 0) are
satisfied. Then the last term in Eq. (4.122) is clearly
(4.113)
zero, and so also is the next to last term. This follows
By comparing this equation with Eqs (4.107-108), we since (for the case considered) Eq. (4.120) yields a/at
clearly see that the consistency and normalization (<p> - q ) = o. Thus in order to satisfy Eq. (4.121) and
 Pdf methods for turbulent flows 155

hence the continuity equation we require To summarize this development: if, and only if, the
02 mean pressure satisfies Eq. (4.129) then, in turn, Eqs
at----~- { < p ) - q } = 0. (4.123) (4.123), (4.121) and the mean continuity equation, Eq.
(4.109) are satisfied, thus establishing proposition (iv).
We now show that this equation is satisfied if, and The starting point for this derivation of the Poisson
only if, the mean pressure satisfies a Poisson equation, equation for ( p ) is the general model for a stochastic
(Eq. 4.129). particle, Eq. (4.124). The result is valid even if U* is
For a fluid particle with state U+(0, ~b+(t), x+(t), not a differentiable function of time. For the more
the velocity evolves by (see Eq. 3.3) familiar but less general case of fluid particles, the
same equation is readily derived. From Eq. (3.3) we
aU; )_~ {ap &o~
at = -P@+ 7xj Ox,j r + a r have
apUj i_a__~ c3(p> FpAO ' (4.130)
Thus in a small time interval fit, the velocity increment
at ax i (pU~U j) = Ox~
A,,U ÷ due to the mean pressure gradient is
_ &p(~+ )-i [V<p>]x+. where

We consider, then, the general model for the PA o =_O"cij ap' +PgJ" (4.131)
velocity increment of a stochastic particle: ax~ axj

A~tU*(t) = -- 3tp(~*)- l[V(p>]xo +g_a,. (This redefinition of A ° is, for the case considered,
(4.124) consistent with definition (4.126).) The Poisson
equation, Eq. (4.129) is obtained by taking the mean
The first terra on the right-hand side represents the of Eq. (4.130), differentiating with respect to x j, and
effect of the mean pressure gradient. The random substituting
function ~ , can depend upon the state variables U*(t), a a 02<p)
~*(t), x_*(t) and time. The corresponding Eulerian axj at (pUj) = Or2 , (4.132)
mean momentum equation is obtained by deriving the
appropriate equation for F*(V_, ~k,x; t), multiplying by which is obtained from the continuity equation.
Vj, and integrating over all V and ~. The result is Finally, we note that the mean continuity equation,
Eq. (4.109), and the Poisson equation, Eq. (4.129), are
0
0-7(qC) = -Tx, ~
(qU,U)
q a(p)
<p> Oxj i_q,7to kinematic equations. That is, they hold at each time t
independent of the past or future. For a constant-
(4.125) density Newtonian fluid this is obvious because the
whereS" equations reduce to
lim {<p~,>/&}.
~°(x,t) - ( p ) - ' (4.126) V .<U> - 0, (4.133)
bt~O
and
From Eq. (4.116) we obtain
aq 2 0 0 V2(p) = - p
a~<U~U~) (4.134)
&2 .= axj dt (qO fl. (4.127) axi,gxj
But in variable-density flow, the occurrence of a< p )/d t
Adding 02(p)/at 2 t0 both sides and using Eq. (4.125)
and d2<p>/at 2 in Eqs (4.109) and (4.129) suggests the
to substitute for O/at(qOg) yields
contrary. However, the time derivatives of <p) can be
72 ~.2(p) 02(qU'-~.
Uj) re-expressed solely in terms of the fields at time t.
?t----~[ ( P ) - q } = Ot2 ~xiax j From Eqs (3.20) and (3.21) we obtain
ap a
0 ( q d(p)'~+ ~__~_.(q~O).
. (4.128) ~- = ~- p(~E~ t])
Oxj ~ axj / Ox~
In order to satisfy Eq. (4.123), the right-hand side of
Eq. (4.128) must be zero. For the case considered (in -- L=(~) ~ t ~
which [ ( p ) - q ] , but not necessarily its temporal
derivatives, is zero), this condition leads to the = L,@){-V,a~, 1 aJ_~7
-~ &, ~s,@)},
Poisson equation
a'<p> a2<p> a:<pU~U~> 0(<p>71 °) (4.135)
CXja-Y.j Ot2 ~XiOXj OXj where
(4.129)
L=(~k) = ~ p(~). (4.136)
f It may be seen that in order for the mean momentum to Each quantity on the right-hand-side of Eq. (4.135)
evolve continuously, the limit lim~t~0 (~,)/& must exist.
This does not imply that lim~r,o~t/fit exists: indeed this can be evaluated from the fields U(x_,t) and ~(x_,tt
limit exists for neither Poisson nor diffusion processes. independent of their rates of change. By the same
156 S.B. PoPE
method an expression for d2p/dt 2 can be derived, and
hence it may be seen that a(p)/dt and dZ(p)/dt 2 can
be determined from the dependent variables at a single
time.
(The conclusion that the pressure is determined
kinematically stems from the assumption that the
density is independent of pressure. Without this
assumption there would be an acoustic pressure field
that is not kinematically determined.)
4.8. Summary
We have considered several different systems of
particles. Theflu/d particle originating from x_o at time
t o has the state U+(x_o,t), q~+(x_o,t), x+ x(xo,t), which
evolves according to Eq. (4.19). Fluid particle paths
are sketched on Fig. 4.2.
The Lagrangian conditional joint pdf fLV(Y_.,~,X;
tlV_o,@o,X_o) is the joint pdf of the fluid particle
state (_U_+,@*,x+) at time t, conditional upon its
state being V(_Yo,¢o,~) at time t o. This is the tran-
sition density for turbulent reactive flows: the mass
density function at time t, ,~(y_,~,x_;t), is equal to
the mass-probability-weighted integral offL over all
initial states, Eq. (4.24).
Conditional particles evolve deterministically with
the same pdf as fluid particles. From the initial state
(_V_o,~o,X0) at time to, the state (O,0~.,~_) of a con-
ditional particle evolves according to Eq. (4.49). The
rate of change of state of a conditional particle is the
conditional expectation of the rate of change of state
of fluid particles with the same state. The solution to
the joint pdf equation can be formally written in terms
of conditional particle states, Eqs (4.52) and (4.56).
Many different systems of particles evolve with the
same pdf. The essence of the whole approach is to
construct a system of stochastic particles whose evolu-
tion is simply computed and in which the pdf evolves
in the same way as the pdf of fluid particles. The
stochastic particle originating from x_o at time t o has
the state (U*,@*,x*) with the corresponding pdf
fL*(v_, t, x; t).
In a Poisson process, the state of a stochastic
particle is constant except at a finite number of time-
points where it changes abruptly; see, for example,
Fig. 4.5. For a diffusion process, in the small time
interval fit, the change of state of a stochastic particle
has two components: a deterministic component of
order fit, and a random component with zero mean
and standard deviation of order (6t) t/2. The state
variables are continuous but non-differentiable func-
tions of time: see, for example, Fig. 4.7.
Equations (4.99-101) provide an example of a com-
bined Poisson and diffusion process. The correspond-
ing pdf equation for fL*, Eq. (4.103), shows that the
Poisson process gives rise to integrals over state space,
whereas the diffusion process leads to a term corre-
sponding to diffusion of the pdf in state space.
The task of modelling is to devise a stochastic
process such that the pdf fL*of the stochastic particles
evolves in the same way as the pdffL of fluid particles.
Then the corresponding mass density functions ~ *
and f f will also have the same evolution. To solve the
evolution equation for ~ * directly would be a pro-
hibitively difficult computational task. But it is
straightforward to compute the evolution of a large
number of stochastic particles which, through the
discrete representation Eq. (4.98), approximate the
mass density function if*.
The mass density function ~ satisfies realizability,
normalization and consistency conditions, Eqs
(4.106-108). The modelled mass density function if*
also satisfies these conditions provided that
(i) the model can be expressed in terms of the
evolution of the discrete mass density function:
and
(ii) the mean pressure is determined by the Poisson
equation, Eq. (4.129).
5. MODELLING
In a turbulent reactive flow, the evolution equation
for the joint p d f f ( V , ~ ; x, t) is (see Eq. 3.109):
,~f
+ Lp(~)gi--~xi). -~+-~ [p(¢)S,(~)f]
~u ~P' v, ,yj •
(5.1)
The mean pressure (p) is obtained from the Poisson
equation, (Eq. 4.129), and p(~k) and _S(~) are known
functions. Thus all the terms on the left-hand side of
Eq. (5.1) are in closed form. The three conditional
expectations containing zu, P' and J~ need to be
modelled while the remaining terms are known.
Before considering the terms to be modelled, we
emphasize that no modelling is needed for the terms
pertaining to reaction, the mean pressure gradient,
buoyancy, or transport in physical space.
In this section, models for the three conditional
expectations are presented. The emphasis is on the
form of the models, their qualitative performance, and
how they affect the solution procedure.
In Section 5.1, the general principles that guide the
modelling are outlined. In order to demonstrate the
qualitative features of different types of models, in
Section 5.2, deterministic, particle-interaction, and
Langevin models are presented for the effect of
molecular diffusion on the pdf f,(~) of a single com-
position. A combined deterministic and particle-inter-
action model for the velocity joint pdf f,(V) is
presented in Sections 5.3 and 5.4; and the generalized
Langevin model for f,(V) is described in Section 5.5.
In Section 5.6 the models for f~(¢) and fw(V) are
combined and extended to apply to the joint pdf
f(V,~). The models presented in Sections 5.2-5.6 are
for inert, constant-density, high-Reynolds-number,
 Pdf methods for turbulent flows
homogeneous turbulence. The development of more
accurate and general models is an active area of
current and future research which is discussed in
Section 5.7.
For the homogeneous flows considered in Sections
5.2-5.6, the joint pdff(V, ~,; x, t) is found to be a joint
normal distribution. 62 The aim of the modelling is,
therefore, to produce Gaussian pdf's whose first and
second moments evolve correctly.
The joint pdf f(V, ¢; x_,t) contains no information
about the time or length scales of the turbulent
fluctuations: such information must be supplied. All
the models presented involve a turbulent time scale
~(x_,t) defined by
"C = k / e . ,
where k(x_,t) is the turbulent kinetic energy
k = 5(uiu,>
1 = f f ~ ( V , - ( U , ) ) ( V , - ( U , ) )f dV dd/,
and e, is the rate of dissipation of k,
e=l~'i~dUJX)/p (5.4,
~xl I I "
(Here, and until Section 5.7, constant density is
assumed.) Since k is known as a function off, the time
scale T can be obtained by solving a modelled trans-
port equation for e,.~o Alternatively, in simple flows,
(or the length scale kX/2T) can be specified directly.
There is further discussion concerning the determina-
tion ofz in Section 5.7.3.
5.1. Guiding Principles
The modelling process consists of replacing exact
but unknown quantities in the joint pdf equation with
approximations in terms of known quantities. For
each of the exact quantities to be modelled (and for
the pdf equation itself) several quite general
properties can be deduced, usually by examining the
effects of various transformations. The modelling is
guided by the requirement that the modelled
quantities have the same properties as their exact
counterparts.
The six principles that have guided the modelling
presented here are:
(i) dimensional consistency,
(ii) coordinate system independence,
(iii) Gallilean invariance,
(iv) realizability,
(v) linearity and independence of conserved passive
scalars, and
(vi) boundedness of compositions.
The velocity U~ef (at a reference point and time) in a
flow can be expressed in various units (e.g. m/sec, ft/hr
etc.), and the reference point can be expressed in
various coordinate systems. But the physical process
(i.e. the flow), being oblivious of these human con-
157
structs, is invariant under a change of units and under
a change of coordinate system. Principles (i) and (ii)
guarantee that the model equations also have these
properties. Dimensional consistency simply requires
that the exact and modelled terms have the same
dimensions. Coordinate system independence is
guaranteed provided that the exact and modelled
terms can be written as Cartesian tensors with the
same suffices.
The principle of Gallilean invariance states that (in
Newtonian mechanics) the equations of motion are
invariant under a change of inertial frame. Conse-
quently, under such a change, exact and modelled
terms must transform in the same way.
Schumann 77 introduced the ¢ealizability principle
(5.2)
into turbulence modelling and it has subsequently
been used by several authors (e.g. Ref. 11). In fact the
ideas introduced by Schumann can be made more
precise by distinguishing between three different prin-
ciples: weak realizability; strong realizability ; and the
(5.3) accessibility of extreme states. These principles are
now described and illustrated.
The weak realizability principle states that if a
physical quantity (q, say) satisfies an inequality (e.g.
q > 0), then the value of the quantity obtaincdfrom the
model equations should satisfy the same inequality.
Examples of such inequalities are
~:(y_,~)/> 0, (5.5)
and
(ulu 1) >1 O, (5.6)
where u - U - ( U ) .
Equation (5.5) is the principal weak realizability
condition that the mass density function (or
equivalently the joint pdf) has to satisfy. If Eq. (5.5) is
satisfied, then all weak rcalizability conditions applic-
able to moments (e.g. Eq. 5.6) are also satisfied. As is
pointed out in Section 4.7, the discrete representation
of the mass density function satisfies Eq. (5.5), and
hence any model that can be expressed in terms of its
evolution satisfies the weak realizability condition. All
the models presented below satisfy this condition.
Another weak realizability condition to be satisfied by
the mass density function is the boundedness of com-
positions, principle (vi). This is discussed separately.
The strong re,alizability condition is concerned with
behavior of physical quantities in extreme states. If the
physical quantity q(t) satisfies the inequality
q >/0, (5.7)
then, by definition, q = 0 is an extreme state. Now if,
at some time t 1, q attains its extreme value (i.e.
q(q) = 0) then its rate of change at t I must also be
zero--for if not, either q will be negative at the next
instant of time, or, contrary to assumption, q is not a
differentiable function of time. Equation (5.7) then
implies
implies
(Oq) =0. (5.8)
- t #=0
158 S.B. Pope
In general, the stron9 realizability principle is that the
rate of change of a physical quantity in an extreme state
is zero.
Suppose that the evolution of q is modelled by the
equation
dq
d--t = R(q[t], rl't]), (5.9)
where r is some other physical quantity, and the
specific model is defined by the choice of the function
R. If q(t) = 0, then the strong realizability principle
requires
g(O,r[t]) = 0. (5.10)
But consider a model for which
R(O,r[t]) > 0. (5.11)
In this case the extreme state q(t)= 0 cannot be
reached (except as the initial condition) and so the
strong realizability principle Eq. (5.8) is not violated,
but it becomes mute. Thus we conclude that the strong
realizability condition is not violated by a model for
which
R(0, r[t]) I> 0. (5.12)
We introduce the principle of the accessibility of
extreme states, accordin 9 to which a model equation
should allow the extreme states to be reached. Equation
(5.10) is then a necessary condition for this principle to
be satisfied. But, unlike the realizability principles,
there is no fundamental reason why we should
demand that the accessibility principle be satisfied.
The models presented here conform to Eq. (5.12)
rather than (5.10). Thus the strong realizability condi-
tion is not violated but the extreme states may be
inaccessible. The extreme states generally considered
are two-dimensional turbulence)1.7a and the perfect
correlation of the velocity and scalar fields.t1"79 Since
neither of these states is likely to exist in the turbulent
reactive flows being considered, their accessibility is
most likely not an important consideration.
The fifth and sixth guiding principles are concerned
with the modelling of the scalars 05(_x,t). In a reactive
flow each scalar may have a different source S~(05) and
a different molecular transport term. But the modelled
pdf equation should be applicable to the simpler case
of 05 being a set of conserved passive scalars with the
same diffusivities: that is,
S(05) = 0, (5.13)
J~ = - F , (5.14)
with the diffusivity F and the density p being constant.
For this simple case, each scalar evolves indepen-
dently according to the same equation that is linear in
05
By considering linear transformations of the set of
scalars 05, Pope a° deduced constraints on the con-
struction of consistent closure approximations. For
scalar moments (e.g. for (ui05,) or (05,05#)), the evolu-
tion equations must be homogeneous (i.e. the same
scalar suffices must appear in each term, including any
repeated suffices).
In pdf models, the linearity and independence con-
ditions are simply stated in terms of the scalar incre-
ment. If the modelled mass density function
Jc*(V, ~, x; t) is represented by N stochastic particles
with the nth having the composition 05"~")(t), then the
increment
,(n)
A6,05*~")(t) -- 05~ (t+6t)-05~,(n) (t), (5.15)
must be a linear function of 05"~"~ (n = 1, 2..... N),
independent of any other scalar property (e.g.
(05~05#)). The models presented below satisfy this con-
dition.
As discussed in Section 3.2, composition variables
are usually bounded. For example, if 05~ is a species
mass fraction, then
0 ~< 05, ~< 1. (5.16)
Consequently there is an allowed region of com-
position space corresponding to possible values of 4>.
Outside the allowed region the joint p d f f ( ~ ~0; x, t) is
zero. Clearly a modelled pdf equation should preserve
this property.
5.2. Scalar Dissipation
In order to examine the qualitative performance of
linear deterministic models, particle-interaction
models, and Langevin models, we consider the effect
of molecular diffusion on the pdf of a conserved
passive scalar. This allows general conclusions to be
drawn (Section 5.2.5) about the use of these models in
pdf methods.
5.2.1. Exact equations
We consider the simple case of the evolution of the
pdf f~ (if; t) of a single conserved passive scalar 05(x, t)
in constant-density homogeneous turbulence. There
are no mean gradients (of ( U ) or (05)) and the
coordinate system is chosen so that the mean velocity
is zero.
Before examining the behavior of the pdf, we con-
sider the evolution of the mean (05) and the variance
(05,2). Assuming simple gradient diffusion, with F
being the constant diffusion coefficient, the molecular
flux of 05 is
J, = - F ~05 . (5.17)
0x i
Then, the transport equation for 05(x, t) (Eq. 3.20) is:
D05
p ~ = rvz05. (5.18)
For the case considered, the mean of this equation is
d(05)
= 0, (5.19)
dt
showing that the mean (qS) remains constant. An
equation for the variance (~b'z) is obtained by multi-
 Pdf methods for turbulent flows 159

plying Eq. (5.18) by 4} and taking the mean: The behavior of the model is best understood in terms
of the corresponding equation for a conditional
1 d<4}'2>
2 dt = - Q,' (5.20) particle:
d,~ 1
where the scalar dissipation is --dt* = -~C,(~-<4}>). (5.31)

0%
% = (F/p) \c3x i d x i / "
(5.21) In composition space (0-space) the conditional
particle moves towards the mean <4}> at a rate pro-
Since % is positive, it is evident that the variance (4},2) portional to its distance (in 0-space) from the mean
decays. <4}>. Thus, as time proceeds, the conditional paths
Just as z = k/e is the decay time scale of the velocity converge to the location ~ = <4}>. The constant of
fluctuations, r¢~ --- ½<4},2>/~ is the decay time scale proportionality ½C, is chosen to yield the correct
of the scalar fluctuations. The determination of e~ is decay rate of the variance.
discussed in Section 5.7.3. For the moment we adopt This model is attractive in its simplicity, and it is
Spalding's suggestion sl that z and z~ are propor- straightforward to verify that the modelled pdf
tional: equation admits the Gaussian solution Eq. (5.27) as it
z~ = z/C~, (5.22) should. But in general its qualitative behavior is incor-
rect. s2 From the arbitrary initial condition
where the empirical constant C, is taken to be 2.0.
f(l~; 0) = (¢'2>o 1/29([Ip - <4}>](4}'2>o ,/2),
Thus,
(5.32)
e~ = ½C~<4}'2>/z. (5.23)
(where # is an arbitrary standardized pdf) the solution
Defining the non-dimensional time t* by to the modelled Ixlf equation is
dt* = dt/z, (5.24) f(~b; t) = (4},2> -1/2g([-~ _ <4}>]<4},2>- 1/2).
the equation for the variance (Eq. 5.20) becomes (5.33)
d<¢'2> (The mean <4}> remains constant and the variance
dr* = -C~<{b'2>' (5.25) <4},2) decays according to Eq. 5.26.) As is readily seen,
the model predicts that the shape of the pdf remains
which, with the initial condition of (~'2)o at t* = 0, unchanged during decay, and consequently, from an
has the solution arbitrary initial pdf, there is no relaxation to a
(¢,2> = (4},2>o exp ( - C¢,t*). (5.26) Gaussian distribution. This incorrect behaviour might
be expected since the model contains no information
If the pdf of 4}, f(~b; t), is Gaussian then it is com- about the shape of the distribution--only the mean
pletely determined by the (constant) mean (4}> and <4}> appears in Eq. (5.30).
the variance (4},2> : In order for arbitrary pdf's to relax, the model must
f,(~) = (2r~<#2 >)- Inexp( _½[qj _ <4}>]2/<4},2>). contain information about the shape of the distribu-
tion. Such models have been proposed by Pope, 24
(5.27) Janicka et al. 2s and Dopazo. s3 In each case the model
A satisfactory modelled pdf equation should admit contains functions of the Ixlf integrated over com-
this Gaussian pdf as a solution. position space. The nonlinear and integral nature of
The evolution equation for f,(~b; t) is obtained from the resulting pdf equation, prohibits analytical solu-
the equation for f(V__,~k; x, t) (Eq. 3.109) by integrating tions. Numerical solutions can be obtained, but only
over velocity space and invoking the conditions of for one-dimensional (and possibly, two-dimensional)
constant density and statistical homogeneity: composition spaces. For flows with many species, and
hence a composition space of large dimension, the
computational work needed to perform the integra-
tion over ~,-space is prohibitive. 32 Fortunately, the
or same result can be obtained by computationally-
~fo
a---~= e [f*<(I'/p)V24}tO>]"
a6 (5.29)
efficient stochastic models.
5.2.3. Particle-interaction model
It may be seen that the evolution of the pdf de-
Consider the discrete representation in which the
pends entirely upon the conditional expectation
pdff~(¢,) is represented by N delta functions:
((F/p)V2O[~> which is to be modelled.
1 N
5.2.2. Deterministic model f~*Ntff; t) -= ~- ~ 6(~-4}*~"~(t)). (5.34)
nml
Dopazo 33 suggested the following deterministic
The discrete pdf evolves by the N stochastic particles
model:
4}*~'~moving in ~-space.
1 We now describe a stochastic mixing model (similar
({F/p)V2¢l~}>=-~C,(~-<4}>)/z. (5.30)
to a Poisson process) that simulates scalar dissipation.
160
In a small interval of time fit, there is a small prob-
ability that a pair of stochastic particles will "mix".
The time zN is defined by
%; --- z/(C,N).
Then in the time interval 6t, with probability 1 - 6t/r N,
the compositions of all the stochastic particles do not
change: but, with probability 6t/zs, the compositions
of a pair of particles change. These two particles
(denoted by p and q) are selected at random without
replacement and their values of 4" are replaced by
their common mean ½(~.(v}+ ~.(e). (Sampling with-
out replacement simply means that p and q may not
be the same particle.) This stochastic model can be
written
with probability 1 - 6tit s:
dp*t"l(t+St) = q~*t'}(t), n = 1,2 ..... N;
with probability 6t/Ts:
1 , ,
dp*~°)(t+6t)=dp*(q~(t+6t) = ~['~b (P}(t)+~b (q)(t)],
and
q~*t"~(t+ 6it) = ~*<"}(t), n = 1, 2..... N, n =~p, n 4: q.
It is important to realize that this stochastic process
is not a direct model of the physical process--a fluid
particle does not change its composition discon-
tinuously at discrete times. Rather, Eqs (5.36-37)
define a stochastic system whose expected state corre-
sponds to the expected state of the physical system.
The pdf for the stochastic system
f**(~h; t) --- E 6(0 - gb*('}(t)),
n=l
evolves continuously and deterministically, and
models the evolution of the Ixtff~(~). If the model is
accurate then f** is equal to f~ and the stochastic
system is equivalent to the physical system. Then,
consistent with Eq. (5.38), the mean of a function of ~b,
(Q(q~)), is equal to the expectation of the ensemble
average of Q(~b*~')):
1 s
(Q(q~)) = E ~ - , ~ Q(q~*("}).
To every stochastic model there is a corresponding
deterministic model. Using the techniques presented
in Section 4.6, or by other means, s4"ss the evolution
equation for f~*(~;t) corresponding to this model
(Eqs 5.35-37) can be shown to be
~- t) = 2C,1i( ~2~f~,(O+~b,)f~(~h -
Of~*(~; +')dgs'-f<.*(+)}.
(Note that the number of particles N does not appear
in Eq. (5.40). It is shown below that the speofication of
zN by Eq. (5.35) produces the correct decay rate of
S.B. POPE
(5.35) (a)
¢
(b)
(5.36)
(c)
(5.37)
FIG. 5.1. Particle trajectories: (a) fluid particles; (b)
stochastic particles; (c) conditional particles.
(q~,2) independent of N.) All the well-behaved deter-
ministic models proposed 24.26,83-85 contain an
integral over composition space such as that in Eq.
(5.40). In a numerical solution, these integrals are
computationally-expensive to perform, and con-
sequently it is advantageous to use the stochastic
(5.38) model instead.
The distinct behaviour of fluid particles, stochastic
particles and conditional particles is illustrated in Fig.
5.1. Fluid particle paths in B-space (Fig. 5.1a) are
erratic with variations on all time scales down to the
Kolmogorov scale. A stochastic particle path (Fig.
5.1b) is discontinuous since ~b*("l is a piecewise con-
stant function of t: the duration of the intervals of
constancy is of order z. Conditional paths (Fig. 5.1c)
are uniquely determined, smooth and continuous. In
(5.39) each case the particle paths converge since the
variance (~,2) decreases with time.
The preceding discussion is aimed at describing the
nature of stochastic models and at establishing their
connection to pdf equations. We now turn to examine
the performance of the stochastic model defined by
Eqs (5.35-37). This model, of which there are many
variants, is variously called Curl's model, the
coalescence/dispersal model, or the stochastic mixing
model. The evolution of the mean (~b) and variance
(5.40) (~b'2) can be determined either from the pdf equation,
Eq. (5.40), or by examining the model directly. It is
instructive to adopt the latter approach.
The modelled mixing process does not affect the
mean value of ~b. From Eq. (5.39) we obtain
 Pdf methods for turbulent flows 161

f l N
+ ft)> - = EJV_ + ft)

-- ~b**"~(t )] }. (5.41)

If in the time interval & there is no mixing then the O.4~

sum in Eq. (5.41) is zero (see Eq. 5.36). If a pair of
elements (p and q) mix, then the sum is again zero
since, from Eq. (5.37), we have
~*~P~(t + f t ) + ~ * ~ ( t + &) -- 4,*cP~(t) + ~*~'~(t). o,

(5.42)
Thus the model correctly leaves the mean <~b>
unaffected. 0.0 ~ ' ~I ~ 1 I ~
Since the mean is constant, a change in the variance 0.0 J.O 2.0 5.0
~0
<~b'2> is the same as a change in <~b2>, In the time
FIG. 5.2. Standardized pdf's f~*(¢) against ¢: . . . . . .
interval fit this is
Gaussian; standard stochastic mixing model;
improved stochastic mixing model. (f~*(~b) is sym-
metric about ¢, -- 0: it is shown for ~ >/0 only.)
=--1 E{ 2N1[(~*ln)(t-I-&)]z_ ~ [q~,~.~(t)]~}
N n= n=l ferent from a Gaussian pdf. In particular the flatness
(5.43) factor <~),4>/<~),2>2 for the mixing model's pdf is
infinite whereas it is equal to 3 for a Gaussian. Pope 8s
The only contribution to the right-hand side is from has developed an improved stochastic model that
the particles p and q in the event that they mix. The produces pdf's that are closer to Gaussian. Figure 5.2
probability of this event is &/%. Thus shows the pdf's for the standard mixing model and for
fit • the improved model compared with a Gaussian.
A~,<4~'2>=~-~ E{Eq~ ~>(t+ft)] +[¢b*~q~(t+ft)] ~ Stochastic mixing models are readily extended to
apply to a set of o scalars, ~ = Of1, ~b2..... ~,. The
- [~*~,~(t)] ~ _ [4,*~,~(t)] ~} joint pdf f,(~k;t) is represented by N stochastic
particles ( ~ ' ~ , n = 1, 2..... N), and the mixing occurs
= C , & E{2[½~b,~,~(t)+½~,~,~(t)]~
T by these particles moving in the a-dimensional com-
position space. For the simple stochastic model
-- [4,*(P)(t)] ~ -- [q~*(q)(t)] ~ } described, two particles (p and q) are selected at
random (in the same way as for a single scalar), and
= c , ft E{-½[~*~'~(t)]'-½[~*~,~(t)] ~
~r mixing occurs by:
+ 4~**'~(t)4~**~(t)}. (5.44) _~*~'}(t+ ft) = ½[q~*~(t)+ 4>*~q~It)],
The second step follows from the definitions of ~N and n = p or n = q, (5.48)
dp*~(t+&). Since 4~*¢~ is an element selected at
random from the ensemble, we have dp*{"}(t + f t ) = q~*{"~(t), n --k P and n ~ q. (5.49)

E [ ~ , ~ ] ~ = <~2> = <~b,~>+<q~>~, (5.45) Figure 5.3 illustrates the mixing process given by Eq.

and the same for E[~b*~] :. On the other hand, since

~*t~ and q~*~ are independent samples, we have
E { ~ b * ~ *~} = E{~b*~}E{~*~} = <~b>2. (5.46) Oq
Substituting these results in Eq. (5.44) and dividing by
fit we obtain p,q

A,~t<f'p'2>/ft -~ --C~b<~)'2>/'~. (5.47)

Hence, in the limit as fit tends to zero, the correct
decay rate of <~b':> is obtained (cf. Eq. 5,25).
(In fact, rather than showing that Eq. 5.47 yields the I
correct decay rate, this analysis is used to determine
the time z~ (Eq. 5.35) that produces the correct result.)
As well as having the correct effect on the mean and
variance, the stochastic mixing model causes arbitrary
initial pdf's to relax to a bell-shaped curve. Unfor- Fro. 5.3. Effect of mixing on the stochastic particles _~*~P~
tunately, this bell-shaped curve is significantly dif- and q~*~q~:O before mixing, • after mixing.
162 S.B. POPE

(5.48) for a two-dimensional ( 0 1 - ~ 2 ) composition , , , i

space. After mixing, both the stochastic particles ~*(P)

and ~*(~) lie half-way between their initial locations.
According to this simple mixing model the means
and variances evolve by,
I
d<4)~>
- - = 0, (5.50)
dt
ee"
and
d<4)'~r~'p>
dr* = - C,<~'_4)~>. (5.51) o

5.2.4. Langevin model

A model of scalar dissipation analogous to the
Langevin equation is (cf. Eqs 4.81 and 4.91): "1
0 I
A~,4)*(t) = G[4)*(t)- <4)(t)>]6t/~ tw

+[B<~'2>/~]t/2fls, W v (5.52)
Here G and B are constants to be determined (B > 0),
and 14:,is a Weiner process. Thus the first term on the
right-hand side of (5.52) represents a deterministic
change in the composition of the stochastic particle;
while the second term represents a random change of
mean zero and variance B< 4),2 > &/~.
Corresponding to this Langevi n model, the evolu-
tion for the pdff~*(~b; t) is (of. Eq. 4.88)
, ~32¢* 0
c3f~ d . .4_~B/,.h,2., v j ,
& . = -~--~ [f~ G ( ~ - < 4 ) > ) ] _ 2 "~" : a@2

(5.53) (b)

Multiplying this equation by (@-<4)>) and inte-

grating yields Eq. (5.19), showing that the mean <q~> -I 0 i i i i

does not change. Alternatively, the same conclusion t*

follows from the observation that the mean of Eq.
(5.52) is simply
A,t<4)*(t)> = 0.
The evolution of the variance <~b'2> can be deter-
mined either from Eq. (5.52) or by multiplying Eq.
(5.53) by (~k- <4)>)2 and integrating. The result is
d<4)'2>
- - = (2G+B)(4)'2>.
dt*
By comparing this equation with the known decay
rate, Eq. (5.25), we obtain a relationship between the
constants:
i value of B. (For B = 0, the Langevin equation, Eq.
G = -~(B+C,).
Figures 5.4a and 5.4b show (k*(t) according to the
Langevin model, Eq. (5.52), for two different values of
B--2.0 and 0.2 respectively. For both plots the same
initial condition 4)*(0) was used and also the same
sample of the Weiner process I,Y, namely that shown
on Fig. 4.7. A comparison of the figures illustrates the
role of the constant B in determining the variance of
the random fluctuations.
These plots of 4)*(t) according to the Langevin
model can also be compared to the plots of 4) ÷(t),
FIG. 5.4. Samples of (~*(t) given by the Langevin model with
(a) B = 2.0, (b) B = 0.2.
(5.54)
4)*(0 and q~(t) shown on Fig. 5.1. The behavior of
~b*(t) according to the Langevin model is similar to
that of fluid particles. However 4)*(t) has fluctuation
on all time scales whereas 4) + {t) has fluctuations only
(5.55) down to the Kolmogorov time scale.
The pdf. admits the desired Gaussian solution, Eq.
(5.27). Further it can be shown (by Fourier analysis)
that from any initial condition the pdf relaxes to a
Gaussian. The rate at which it relaxes depends on the
(5.56) (5.52), reduces to the deterministic model Eq. (5.30)
and there is no relaxation.)
The qualitative behavior of the Langevin model is
satisfactory in all respects except one--it violates the
principle of the boundedness of compositions.
Because of the diffusive nature of Eq. (5.53),.f~* (~; t) is
greater than zero for all values of tp. Thus if 4) were a
bounded scalar (e.g. 4)t> 0) then the boundedness
condition would be violated since the pdf would be
positive at values of ~ corresponding to impossible
compositions (e.g. ~ < 0). (Note that Fig. 5.4a shows
regions of negative 4)*.) For the simple one-dimen-
 Pdf methods for turbulent flows 163

sional case considered, the Langevin equation could the momentum equation (Eq. 3.3) becomes
be modified to guarantee boundedness. But for the
general case, such modifications would be prohibi- dui 6uj 0uj 0p' (5.61 )
tively complicated.

5.2.5. Conclusion (The mean pressure is uniform and, since the density is
constant, there is no buoyancy effect.) The Reynolds-
From the preceding examination of deterministic, stress equation is obtained by multiplying Eq. (5.61)
particle-interaction, and Langevin models, we can by Uk, adding the same equation with j and k com-
draw conclusions about their use in pdf methods. muted, and taking the mean. The result is
Because of their simplicity, linear deterministic
models (e.g. Eq. 5.30) are appealing, especially from ~(ujuk) ~(utujuk) = U(ukV2uj + ujV2uk )
a computational viewpoint. Since they do not cause P~ + P c3xl
the pdf shape to relax, they cannot be used alone. But / ap', ap'\
they can be used in conjunction with other models. A - \u, (5.62)
linear deterministic model is used to model part of the
pressure term (Section 5.4). Finally, making use of the fact that in homogeneous
For the case considered (scalar dissipation) the turbulence the spatial gradients of mean quantities
Langevin model is, in general, inapplicable because it ((u~ujuD, (uju~) and (u@')) are zero, we can write the
violates the principle of composition boundedness. Of Reynolds-stress equation as
the models described, only the particle interaction
model is satisfactory in all respects. (The simple model d
P "~ (UjUk) = --Pgjk + Pjk, (5.63)
described is unsatisfactory in that the pdf shapes
obtained are far from Gaussian, but the improved
where the dissipation tensor is
mixing models 85 are satisfactory.) In the next sub-
section particle-interaction models for the velocity t~ /Ouj Ou~\
joint pdf are described. ~jk=--2p t~xi-~X~), (5.64)
Although the Langevin model is inapplicable to
scalars, it is a good model for the velocities (which are and the pressure-rate-of-strain correlation is
not bounded). A generalized Langevin model for the / ,/auj Ou~\
joint pdf of velocity is presented in Section 5.5. PSk=--\p t~-~Xk+-~Xs)/. (5.65)

5.3. Energy Decay and the Return-to-lsotropy Of course, the Reynolds-stress equation can also be
derived from the joint pdf equation. This is achieved
We now turn our att,~tion to the joint pdf of the by multiplying the equation for f(_y,~;x_,t), Eq.
velocities fu(V;t) for constant-density homogeneous (3.109), by VjVk, integrating over V-q/ space and
turbulence with no mean velocity gradients. A good invoking the assumptions of homogeneity, etc.
approximation to this flow is the turbulence generated The decay of the turbulent kinetic energy
by a uniform stream passing through a grid. In a 1
coordinate system moving with the mean flow k = ~(uiui), (5.66)
velocity, the two principal observations are that the
turbulent kinetic energy decays, s6 and that the is due solely to the dissipation: contracting each side
Reynolds stresses tend towards isotropy, s~ of the Reynolds-stress equation yields
The evolution equation for the Reynolds stresses dk
(U~Uk) can be derived from the continuity and -- = -e, (5.67)
dt
momentum equations (Eqs 3.2-3.3). Since the density
is constant, the continuity equation (Eq. 3.2) becomes where the dissipation rate of turbulent kinetic energy
is
= O. (5.57) 1 ~_(Ou~Ouj;
g.x i
e= -~e~s= P \~xi ~3xil" (5.68)
By decomposing U__into its mean (in this case zero)
and fluctuation u There is no contribution from the pressure-rate-of-
strain because, by virtue of the continuity equation,
U(_x,t) = (U(x,t)) +u(_x.t), (5.58) Ps~ is zero. It may be noted that e is a positive quantity.
we obtain It is assumed that e is known, either from a
modelled transport equation or through the turbulent
?u~_= O, (5.59) time scale z - k/e. In terms of the normalized time
~'x i (dt* = dr~r), the kinetic energy equation is
With the assumption of a Newtonian fluid dk
- - = - k. (5.69)
dt*
TiJ = l ~\ / C.~j
w - - + ~('%i
- / /. {5.60)
The net rate of change of the Reynolds stresses is
164 S.B. POPE

often expressed as, not be linear in bjk as Rotta's model implies, Eq. (5.76).
d<ulu~> 2 (This is discussed more in Section 5.5.) Nevertheless,
dt = --'~.6jk'l-Rjk, (5.70) Rotta's model provides a reasonable representation of
the experimental results--at least with C 2 being a
where function of the anisotropy and Reynolds number.

Rjk E PflJp-
[2]
~,.ik---~l~6jk , (5.71)
For the flow considered, the evolution equation for
the joint pdf of velocity fu (V; t) is

(Lumley x~ and Launder et al.2°). The isotropic part P-h-: \ax, ax V

(--(2/3)e61k) causes the normal stresses (<ulul>,
<UzU2> and <u3u3>) to decay, while the anisotropic
part R~k serves to redistribute energy among the
Reynolds stresses: the trace Rjj is zero. The observed
(This equation is obtained by integrating the joint pdf
tendency of the Reynolds stresses to become isotropic
equation, Eq. (3.109), over composition space and
is due to the term R~k.
invoking the assumptions of constant density and
Rotta sa suggested the model
homogeneity.) The conditional expectations, which
2 are to be modelled, represent the (conditional) force
per unit volume due to viscous stresses and to the
fluctuating pressure gradient. A good model causes
where C2 is an empirical constant. The Reynolds- the pdfto relax to a Gaussian. The mean velocity <U>
stress equation then becomes does not change, while the second moments evolve
(approximately) according to Rotta's model, Eq.
dtd <uju,> = _~e6~ _C2 2 (5.73).
A particle interaction model--or, rather two
(5.73) models--produce the required result. The Reynolds-
Rotta's model is best understood in terms of the stress equation incorporating Rotta's model, Eq.
normalized anisotropy tensor (5.73), can be rewritten

1 1 d <UjUk> _<UiUk>_C,z <UlUk> kfjk .

bik =-- ~ <UlUk>/k-- ~ 6~. (5.74) dr*

This tensor has zero trace and each component is zero (5.80)
if the Reynolds stresses are isotropic. A scalar measure The first term on the right-hand side causes the kinetic
of the anisotropy is energy to decay without the anisotropy tensor being
b~ - b(~b~, (5.75) affected: the stochastic mixing model produces this
effect. The second term causes the anisotropy tensor
which is zero for isotropy and positive otherwise. to decay without the kinetic energy being affected: the
Equations (5.73-75) yield stochastic reorientation model produces this effect.
db~ The stochastic mixing model is directly analogous
dt* = -C2bjk' (5.76) to that used for the scalar dissipation. The joint pdf
f~_(V;t) is represented by the discrete pdf f*N
and which is composed of N delta functions:
db~ , 2 1 N
dr* = - 2 C 2 b " ' (5.77) L*u(V; t) --- ~- ~ ,~(V-U*~"I(t)). (5.81)
n=l

where The velocities U *(") of the stochastic particles remain

C~ -- C 2 - 1. (5.78)
It may be seen that for Cz < 1 (C~ < 0), the aniso-
tropy increases, contrary to experimental observa-
tion. 87 For C2 > 1 (C~ > Ok the anisotropy decays
exponentially with t*, and for C2 = 1 (C~ = 0) the
anisotropy is constant.
Launder et al. 2° used Rotta's model with the con-
stant value C 2 = 1.5. Lumley ~t suggested that Rotta's
model (and, hence, Eq. 5.76) accurately describes
existing experimental data, although C2 is not a
constant--it is a function of the anisotropy and
Reynolds number. Recent experiments and the re-
examination of existing data s7 confirm that C 2 is not
a constant and suggest that the return to isotropy may
constant except at discrete times. In the small interval
of time 6t, with probability 1-6t/z., the velocities
U *("1 of all the stochastic particles remain constant;
where
r~ =--~/(NC~), (5.82)
and C u is a constant to be determined. With prob-
ability 6t/zu, a pair of particles mix. The two particles
(denoted by p and q) are selected at random with-
out replacement and their velocities are replaced by
their common mean velocity:
1
U*(~(t + at) -- U*(~)(t + 5t) = ~ [_U*(p~(t)+ U*(~(t)].
(5.83)
 Pdf methods for turbulent flows 165

V3 d
- ~ (uju~) = - Cu(UlUk), (5.91 )

and hence, for compatibility with the first term in Eq.

(5.80), C, is unity.
The stochastic reorientation model differs from the
2 stochastic mixing model only in the definition of the
time scale, and in the operation performed on the pair
of particles selected. The time scale TR is defined by
ZR -- ¢/(C~N). (5.92)
p Vl
In the small time interval fit, with probability 6t/ZR,
tWO particles (p and q) are selected at random with-
out replacement. These particles are randomly re-
oil orientated in velocity space by

_U*(P)(t+f t) =__
U ( ~ )(t ) + 1-~11A U ( ~ ) (),
t
FIG. 5.5. Effect of mixing on the stochastic particles U *(p)
and U*(q): © before mixing, • after mixing. and (5.93)
|
U*(q)(t + ft ) -- ~_(~)(t) - 2 ~ZA U(Pq)(t),
Figure 5.5 shows, in velocity space, the effect of mixing
on the stochastic particles p and q, Eq. (5.83). After where ~/is a random vector uniformly distributed on
mixing, both particles lie half-way between their initial the unit sphere (that is, !/is a unit vector of random
locations. orientation). In this process, which is illustrated on
The momentum (per unit mass) M (p~) associated Fig. 5.6, neither the average velocity ~(~) nor the
with the particles p and q is, simply: velocity difference AU (n) is changed. Consequently,
neither the momentum nor the kinetic energy is
M (~) = U*(P)+U_ *(q) = 2 ~ (~), (5.84)
affected.
where the average velocity U (p~)is In Section 5.3.1 the stochastic reorientation model
1 ,(p) ,(q)
is analyzed to show that its effect on the Reynolds
D__(p~)_--~LU_ + U ]. (5.85) stresses is
d(ujuk) ,{ 2 }
Equation (5.83) shows that the average velocity is dt* = - C 2 (ujulc)-3 k f j k " (5.94)
unaffected by the mixing process
When the stochastic mixing model and the stochastic
O(P~)(t + fit) - _ lJ(P4)(t).
- (5.86)
reorientation model are used simultaneously, the net
The kinetic energy (per unit mass) associated with the effect is the sum of the two effects (Eqs 5.91 and 5.94).
particles is Thus, the stochastic models cause the Reynolds
stresses to evolve according to Rotta's model (Eq.
K (p~) -= 1 U * ( P ) ' U * ( p ) + I u *(q) "U *(~). (5.87) 5.80).
2~ -- 2-- --
In terms of the average velocity ~(P~) and the velocity
difference
A U (pq) = IU *(p)- U__*(~)[, (5.88)
the kinetic energy can be re-expressed as

= U (pq) • U(~)+~ [AU(~)] ~. (5.89)

Equation (5.83) shows that the change in kinetic

energy due to mixing is
Vl
1 i
K(Pq)(t + fit) - K(Pq)(t) = - ~ [AU(Pq)(t)] 2.

(5.90)
Thus, the reduction of the particle separation in
velocity space causes a decrease in kinetic energy.
The type of analysis used in Section 5.2.3. can be FIG. 5.6. Effect of stochastic reorientation showing that the
used to show that the effect of the stochastic mixing average velocity ~ and the velocity difference Am are un-
model on the Reynolds stresses is affected: © before reorientation, • after reorientation.
166 S.B. POPE

The reorientation process, Eq. (5.93), is directly Similarly, for j :z k, the conditional expectation
analogous to a collision between two Maxwellian E(q/lqk ) is zero since, for given qk (k ,==j} and z, the
molecules. Before the collision, the two molecules (p events qj = z and qj = - z are equally likely. Hence.
and q) have velocities U_*°'~ and U *~. They collide,
E(qfllk) = O. j ~= k. (5.101)
and their new velocities U*~)(t + fit) and U*~q)(t+ fit)
are such that both momentum and energy are con- By definition, the vector ~/has unit length:
served. In view of this analogy, it is not surprising that
qjqj "~- E(qjrlj) = E O h q l ~'- q 2 q 2 "b q3r]3) = I.
the stochastic reorientation model causes the pdf to
(5.102)
relax to a Gaussian (Maxwellian) distribution.
And, since by symmetry E071q1), E(v/2v/z) and E(v/zvt3)
5.3.1. Analysis of the stochastic reorientation model are equal, we obtain
The effect of the stochastic reorientation model 1
(Eqs 5.92-93) on the Reynolds stresses is determined E(qfl/k) = ~, j = k. (5.103)
in this subsection.
Equation (5.101) and (5.103) can be written in the
In the time interval 6t, with probability 6t/zR, the
common form
velocities of particles p and q change: otherwise the
velocities are constant. Thus the change in <UjUR) is 1
E(qjrlk ) = "~ 6jk, (5.104)
6t I E{U~n}(t +gt)U~P)(t +6t)
Returning to Eq. (5.96), since ~ is independent of the
velocities, the second term on the right-hand side is
+ U*~q~(t+ 6t)U~<~(t + 60
zero:
_ U.t~. U ' ~ - U.~ u * ~ } ,
(5.95)
where, here and below, U *~"~ is written for U*~")(t). (5.105)
From the definition of U*~P}(t+6t) and U*~}(t+6t),
Similarly the third term is zero. The final term is
Eq. (5.93), we obtain
E{ U*~m(t + 6t )U~P~(t + 60} E {~ r/jq,[AU°'q'] 2}

-- 17 ~ ~(Pq) ~(Pq) _t_ 1 tTt~),, A rT(P~) 1

= ~E(tljrlk)E{[AU(")] 2}

+~I _U~'}n~AU~'~+~I n~[zXU~~] 2 ~. 1

_ - ~ Vjk
x E/U*(p)tr*{p~_2U*{P}U*~q}~tr*{q)t;*tq)~
-- I i ~" i i i ~ wi ~i )"
(5.96)
1
Substituting for 10t ~ (Eq. 5.85), the first term becomes = -66j~{< v , v , > - < v,> <u,>}
E"f ~(pq) ~lpq) ~ - - 1 m,.~r r*~p)r r*<p}..t..r r *lP} U ~q) 1
=-~kl)jk. (5.106).
4-tT*(q) rf*(P) _t_/T*(q)//*(q}~
~'Jj ~'/k " ~ j ~k J All the terms in Eq. (5.95) have now been evaluated.
1 Using these results, the change in the Reynolds stress
= ~ (uju~) + (Uj> (U~). (5.97) is found to be

The last step follows from 6t 2

Aat<UjUk> = -- Nz-----Rs( < U j U k > - - 3 k6jkj" (5.107)
E {U *'m U~''~'} = E {U*°'~}E{ U *'~'} = < U j) < Uk>,
Dividing by 6t, substituting for rR (Eq. 5.92), and
(5.98) taking the limit as 6t tends to zero, we obtain the
(since U *(~) and U *(~) are independent) and from required result:
E~ tr*~t~*(~ = <U ~Uk) = <U ~) <U~) + <U~Uk>. d(UjUk>
-C" (ujuk)--:k~jk • (5.108)
dt*
(5.99)
The remaining terms in Eq. (5.90) depend upon q.
Since the orientation of ~/is random, each component 5.4. Rapid Pressure
(q~, k = 1,2,3) is symmetrically distributed about
While still considering constant-density homo-
zero--for any number z, the events qk = Z and
geneous turbulence, we now allow uniform mean
q~ = - z are equally likely. Consequently the expected
velocity gradients ~(Ui>/Ox j. The interaction of the
value of each component is zero:
mean velocity gradients with the turbulence gives rise
E(qk) = 0. (5.100) to "rapid" pressure fluctuations. In this section a
 Pdf methods for turbulent flows 167

deterministic model for these rapid pressure fluctua- where

tions is presented.
For constant-density flows, the continuity
equation, (Eq. 3.2), yields
t~Ui _ c3(Ui) = Ou~ =
An equation for the fluctuating velocity u_is obtained
by subtracting from Eq. (3.3) its mean. For a constant-
density Newtonian fluid, the result is,
~u
p ~ "~-p ~ [( Ui>Idj "~ ( Uj)IAi -~-uiuj (liildj>]
B~,j(v_) -- - ~1 f i l-rl & 0, O2 ~ <u,.(x_+_r)l_u_(x_)= X> dr.
(5.116)
0. (5.109)
A derivation of this equation is presented in Section
5.4.1 where several properties of the tensor B,,o(v) are
deduced.
The task of modelling the rapid pressure term has
now been reduced to that of modelling the tensor
B,,eflY.). Since B has dimensions of velocity and is
- -

independent of the scale of the turbulence, Pope ~2

suggested modelling it as a function v and the
=/tV2u~ Ox~"
Op' (5.110) Reynolds stresses. It is expedient to assume that B is a
linear function of v since this leads to a model that is
(For the homogeneous case, the gradients of the compatible with Reynolds stress models. Some
Reynolds stresses O(uluj)/Ox i are zero.) A Poisson support for the linearity assumption is provided by
equation for the fluctuating pressure is obtained by the exact result (Eq. 5.146 of Section 5.4.1 ):
differentiating Eq. (5.110) with respect to x j:
B,,e/ = v,,. (5.117)
V2P' -- -20 t3(Ui) c3uj t~2uiuj (5.111)
- - p - - The general form of the linear model is
Oxj ax~ Ox~gx/
This Poisson equation for p' shows that there are Bmo = vqCqm/j, (5.118)
two sources of pressure fluctuations; one due to the where C is a non-dimensional tensor function of the
interaction of turbulence with the mean velocity Reynolds stresses.
gradients, and the other solely due to turbulence. Before considering the determination of the tensor
These contributions to p' are denoted by pm and p(:) C, let us examine the form of the model, Combining
respectively. The contribution p(~) is termed the Eqs (5.118) and (5.115) yields
"'rapid" pressure since it changes rapidly in response
to changes in the mean velocity gradient. The c3Pm[u v ) = ~ O(Ui)~
-~x~ l_ = -Zp~x Cq.ejv q, (5.119)
separation o f f into two contributions allows its effect
on the joint velocity pdffu (V) to be separated (of. Eq. or alternatively
5.79):

. . . . . . . :_. v Oxj I- ex.

(5.120)
.... The effect of the rapid pressure on the joint pdf of
velocity (Eq. 5.112) is, then,
(5.112)
-Of,
g ....... _
2 cg(Ur) c3
The stochastic mixing and reorientation models have
been used to model the term in p(2): a deterministic
model for the term in pm is now developed. (5.121)
From Eq. (5.111 ) we have and the corresponding conditional Lagrangian
equation is
V,p(l) = _2p~(Ue.......~t3u,, (5.113)
c~x,. c3x:'
and differentiating with respect to xj (recalling that dt = "'" 20(Y
vx,.?) Cqmlj(Oq--(Uq)). t5.122)

the mean velocity gradient is uniform) we obtain Thus, the effect of the rapid pressure is to move
conditional particles in velocity space at a rate pro-
V 2 ?p(1) = ~(U:) ~2u~ . (5.114) portional to their distance from the mean velocity
~xj 2p Ox~OxrOx~
(_O~-(U)) and proportional to the mean velocity
This equation can be solved using Green's theorem to gradient. Since the mean velocity is unaffected, the net
yield an expression for the conditional rapid pressure motion in each direction (in V-space) is zero: and,
gradient in terms of the two-point conditional expec- since the kinetic energy is unaffected, the mean-square
tation of velocity: outward motion from V = ( U ) is balanced by the
(gff~)u } -2pC(~ U`) mean-square inward motion.
- - = v = B.~:(y_), As with the other linear deterministic models, this
\ (xj I- cx,,
model does not change the shape of the distribution,
(5.115) only its moments. Thus a joint normal distribution
168 S.B. Povw:

remains normal and, conversely, any other distribu-

tion does not relax to a normal distribution. For the
other processes considered, the lack of relaxation was
considered to be a defect in the model. But rapid
distortion theory 89 shows that the rapid pressure has
a deterministic effect on the turbulence, and hence,
does not cause the pdf to become Gaussian.
A completely satisfactory model for the non-dimen-
sional tensor C has yet to be developed. The best
available model is obtained by relating C to a fourth-
order tensor A that appears in Reynolds-stress Fla. 5.7. Sketch showing points x_and y in region R.
models. Multiplying Eq. (5.119) by (-f~vk) and inte-
grating we obtain separation vector r is
_/. Op,.\ /p,.Ou,,\ o<u, r = (r 'r) 1/2, (5.127)
/ = \ Ox l = 2p ~ x ~ Cq,,ej(uj,u~).
_n is the outward-pointing normal vector and O/On
(5.123) denotes the derivative in the direction ofn.
Now, the t e r m (pt!)Ouk/Oxj) appears as an unknown The Poisson equation
in the Reynolds-stress equation and is generally V2 tgp~t) 2p 0(U,,) O"u,,,
modelled by, (5.128)
Oxj = ~x~ OxeOxj'
~ x J = pk ~ x ~ A.ikt,,,, (5.124) (where O(Ut)/Ox,,, is uniform) has the solution

where there are several suggestions for the tensor A OPO':' - 2 p ~ B , ~ j + I ~ U +l~2', (5.129)
(Launder et al.,2° Lumley and Khajeh-Nouri 9° and
Lin and Wolfshteing!). Comparing these two where
equations, we see that A and C are related by
l f f f , R10:u"(~')dy,
r oy: oyj -
(5.130)
Ajkt,n = (2/k)Cq~j(ukuq), (5.125) B~ = -'~n

and
1 !
I~t) =-~1ffs°2P"tY)lds, Oy~On r (5.131)
Cq,,ej = -~k(UkU~)- Ajkem. (5.126)
and
From this last equation, Pope 72 evaluated the tensor
C corresponding to Launder et al.'s model for A. I~2~ = -4--~ 1 ffs ep",(y) a
aye--- 0n dS. (5.132)
For many flows the quantitative performance of
This solution follows directly from Green's second
Launder et al.'s model may be adequate. But their
identity (see, for example, Ref. 92).
model--and all the others mentioned--do not
We now take the expectation of Eq. (5.129) con-
guarantee (weak) realizability. In the pdf model, the
ditional on u(x_) = v, and then take the limit as R tends
(weak) realizability principle imposes no conditions
to infinity. In this process, both surface integrals
on C--except that it be finite. But, if the turbulence
vanish. For I}2~,for example, we have
becomes two-dimensional, the Reynolds stress tensor
becomes singular and then the right-hand side of Eq. = v>__ _1 f f ?p"'(y)lu(x)-- v; 0r-'
(5.126) becomes infinite, unless A satisfies a particular 4n J J s \ dyi ]-- -/an dS
condition. Current models for A do not satisfy this
condition and hence are badly behaved when the = - 4"-~ (Pt)(x-+r)]u(x-) = v)-ffn-n dS'
turbulence becomes two-dimensional. In view of this,
it would be preferable to model C directly. The corre- (5.133)
sponding model for A deduced from Eq. (5.125) would
then guarantee realizability. Now as R, and hence r, becomes infinite, pa~(x+r)
becomes independent of u(x_):
5.4.1. Solution of the Poisson equation lim (#U(x+r)lu(x) = v) = (p")) = 0.
r~ct2
An exact solution for the conditional rapid pressure (5.134)
gradient is obtained in terms of the two-point condi-
Consequently the derivative of (p"~(x+_r)lu(x)=_v)
tional expectation, Eqs (5.115-116). The solution
also tends to zero, yielding
follows from a straightforward application of Green's
theorem, but some care is needed to ensure proper lim (l}2~l_u(x) = v) = 0. (5.135)
R~oo
behavior at the origin and at infinity.
Let x and y = x + r be points within a regular region The same argument yields the same result for I~u.
R with bounding surface S, Fig. 5.7. The length of the The conditional expectation of the volume integral,
 Pdf methods for turbulent flows 169

Eq. (5.130), is to Poisson's equation 92 we obtain

(B~,r lu(x) = v ) -- (uAx)lu_(x) = _~)
(Br~OlU(~-)----"v-) -~ - 4 - 1~ f f IR 1r (¢~:um(-Y'
\ c y / c y j l u(x,
- -- = -v~)
/ dy
--

l ffs!~n(u,~+t)lu(x-)=v-)dS
4n
1 fffffR1 ~32
47r r t3rrt3rj
x (u,,(x_+_r)lu(x) = v)dr. (5.136) +~-n~ (u=(x +r)lu(x)=v.)--~-ndS.
In the limit as R tends to infinity, this equation defines (5.145)
the third-order tensor B,ej(v),
The first term on the right-hand side is just v,,, while,
B,,¢/(v_) = l,im (B,~,j [u(x) = v) as r tends to infinity, the other terms vanish. Hence we
R~
obtain
1 f l ~2
= -- 4--~ _J r ~ (urn(x-+ [)[-u(x-) = v) dr, B~f(v) = lim (B~flu(x) = v__)= vm. (5.146t
R~oo

(5.137)
where ~dr_ represents the volume integral over the 5.5. l_,angevin Model
whole physical space. The convergence of the integral In the previous two sections a model for the evolu-
is not in doubt since, as r tends to infinity, tion of the velocity joint pdf fu is described. The
(u~(x_+r)lu(x_) = v)) tends to zero more rapidly than dissipation and return-to-isotropy terms are modelled
i.-1
by the stochastic mixing and reorientation models
The conditional expectations of all the integrals in (Section 5.3), and a linear deterministic model is used
Eq. (5.129) have now been evaluated to yield the for the rapid pressure (Section 5.4). In this section an
required solution alternative model for the evolution offu is described.
This is the generalized Langevin model--a generaliza-
u_(x_)= v = - 2 p ~-~-~-_ B ~ q ~ ) , tion of Eq. (4.91) (see also Eq. 5.52). As in the previous
(5.138) sections, we consider constant-density homogeneous
turbulence with uniform mean velocity gradients.
where B,ej is given by Eq. (5.137). This solution is In terms of the velocity increment of a stochastic
useful because it shows explicitly the role of the mean particle,
velocity gradient and because several properties of
Bmq can be deduced. It is evident from Eq. (5.137) that An,U__*(t) = U*(t+ft)-U*(t), (5.147)
the result is unaffected by reversing the order of the generalized Langevin model is
differentiation: hence,
6t ~<p>
Bmej = Bmje. (5.139) AaU* = - - +-Go(U*-<Uj>)6t
p dx~
The continuity equation, Eq. (5.109), at the point y is +BI/2A~,(W~),, (5.148)
(~Um(Y__) CUm(X_"~r ) where W, is an isotropic Weiner process:
. . . . o, (5.14o)
~y',~ t3r~,
(Aa(Wi),As,(Wj),) = 6t~ u. (5.149)
and hence
The second-order tensor G and the positive scalar B
~r---~(u~,~-i-_r)l u(x) = v) = 0. (5.141) define the particular model, and their determination is
discussed below. Both G and B can depend upon time
Thus, in Eq. (5.137) the integrand vanishes i f / o r j is but are independent of U* and x (in homogeneous
contracted with m: turbulence). The first term on the right-hand side of
Eq. (5.148) defines the effect of the mean pressure
Bm,.j = 0, (5.142) gradient on the stochastic particle to be the same as
and that on a fluid particle.
With U* evolving according to the Langevin
B,e,~ = 0. (5.143) equation (Eq. 5.148), the Eulerian pdf f*v(y__;x_,t)
(Eq. 5.143 also follows from Eqs 5.139 and 5.142). evolves by (of. Eq. 4.105),
The final property of Bm/j is slightly more difficult to ~*+ VOW* = 1 0 ( p ) Ofu* 8
obtain. Contracting / andj in Eq. (5.136) yields
c3t c3xi p t3xi c~V/

(B~*//lu(x_) = v_) = -4nn r ~r/~,? 4"

I
:B ~C~2t*
.
Ju (5.15o)
x (u,~(x+r)Ju(x_) = v) d_r. (5.144)
Although the fluctuating velocity field is (by
Now the integrand is the Laplacian multiplied by the assumption) statistically homogeneous, the pdf
Green's function (I/r}. Hence. from the exact solution f*(V_, ~ t) depends upon x (because the mean velocity
170
(U(x,t)) is not uniform). It is useful, therefore, to
consider g(v; 0 - - t h e Eulerian joint pdf of the velocity
fluctuations:
g(v; t) - f * ( V ; x_,t),
where
v - V - (U(~, t)).
When the evolution equation for g is derived from Eqs
(5.150-152), the spatial derivative and mean pressure
terms vanish, leafing
Og ~O(Us) G,~ 0 [ g v ] ' 1 _ t~:g
0-7 = - 2
+ - ' - OvtOv~"
This equation is qualitatively correct in that it
admits joint normal solutions. And, as noted in
Section 5.2.4, the term in B causes arbitrary initial
distributions to relax to Gaussians (for B > 0).
The coefficient B is determined by requiring that (in
a limited sense) the Langevin equation be a direct
model of fluid particle behavior. Consider a time
interval s that is much less than ~, but much greater
than the Kolmogorov time scale %:
% a¢ s a< z.
The quantity
A~_UU+ft)= U+(t+s)-U+(t),
is the change in the Lagrangian velocity of a fluid
particle over the time interval s. According to
Kolmogorov's (1941) hypotheses, 93 the covariance of
A,U + (t) is (to first order in s):
(AsU+(t)A~U~(t)) = Coes6u,
where Co is a universal constant.
From the Langevin model (Eq. 5.148) we obtain (to
first order in s):
(A,U*(t)A,U*(t)) = Bs6 u,
for
0 ~< s ~ z.
By comparing Eqs (5.156) and (5.157), we see that (for
% << s << z) fluid particles and stochastic particles
have the same statistical properties provided that B is
chosen to be
B = Co e.
For time intervals s much less than %, the covariance
of A~U*(t) is of order s (Eq. 5.157), but for fluid
particles the covariance of A,U+(t) is only of order
s2. 93 Thus only in the limited range (5.154) does the
Langevin model provide a direct model of fluid
particle behavior.
For homogeneous isotropic turbulence (without
mean velocity gradients), the tensor Go can be
deduced without further modelling assumptions. The
tensor must be isotropic (Go oc 6o.), and the coefficient
of proportionality is determined by requirihg that the
S.B. PoPE
modelled equation yield the correct energy decay rate
(Eq. 5.67). Thus, for isotropic turbulence (with
(U)=0), the Langevin equation, (Eq. 5.148),
becomes
(5.151)
A,,U~' =-(~+~Co)(6t/z)U*+(Coe)'.'2A~,(Wi)v
(5.152) (5,160)
Anand and Pope .4 used this model equation to
calculate the diffusion of heat behind a line source in
grid turbulence. Excellent agreement with the avail-
able experimental data was obtained for C o ~ 2.1.
Since, according to Kolmogorov's hypotheses, Co is a
universal constant, we adopt this value for the general
(5.153) model.
We now consider modelling G~ifor the more general
case of homogeneous anisotropic turbulence with uni-
form mean velocity gradients. At this level of closure,
G can be a function of z, (ului) and O(Ui)/Oxj. The
most general model for G contains far more scalar
coefficients than could possibly be determined with
confidence. Haworth and Pope 94 considered a simpler
model for G that is linear in both the Reynolds
stresses and the mean velocity gradients:
(5.154) 0(Uk)
G ° = (°tt6iJ+°t2bij)/'c+Hqke Ox/ ' (5.161)
where
(5.155)
Hoke = fl160 6ke + f12 6ik 6je + f13 6i( 6jk
+ yi6obkr "~-Y26ikbjt, + ya6iebjk
+ y4bo6ke + Tsbik6je + y6bie6jk, (5.162)
and the eleven coefficients ~t,, tim and Ym are to be
(5.156) determined. (The anisotropy tensor b is defined by Eq.
5.74.)
In the alternative model presented in Sections
5.3--4, the stochastic reorientation model (that does
not depend upon 0(Uk)/0X<) simulates the effect of
(5.157) the "slow" pressure pt2), whereas a term linear in
O(Uk)/Oxe simulates the effect of the "rapid" pressure
p(t), Eq. (5.119). Similarly, the terms in ~tt and ~t2 in Eq.
(5.158) (5.161) could be identified with the slow pressure, and
the term in d(Uk)/COxe with the rapid pressure. Then,
it follows that H and C (Eq. 5.118) are related by
Hi)ke = 2Cjrki. (5.163)
However, the decomposition of Eq. (5.161) into slow
(5.159) and rapid parts contains the implicit assumption that
the slow pressure is independent of mean velocity
gradients. There being no strong basis for this
assumption, we follow Jones and Musonge 95 in re-
jecting it. As a consequence, H is not subject to many
of the constraints that otherwise follow from the
properties of B~c~(v_) = VqCq,,ej (see Section 5.4.1).
In the determination of the eleven coefficients in
Eqs (5.161-162), several subtleties arise. 9a Since the
coefficients/~ and ~'4 multiply
0(Uk) 0(Uk)
6ke. . . . . 0, (5.164)
Oxe Oxk
 Pdf methods for turbulent flows
their values are irrelevant and can be set arbitrarily.
The requirement that the model yields the correct
kinetic energy evolution leads to the condition:
0~1~ -(~"~-~Co)-~x2bijbji-T('~l-l-fl2 "~ f13 "J¢'~~l*b) kl
x O(Uk) -w*bubi/O(Uk)
Ox~ Oxr '
where
~* m. ~2.~-~3..1-~5.~6.
Two other requirements are imposed: that the model
produces the correct response for the rapid distortion
of isotropic turbulence; and, that the multiplier of
(0( Ut)/~x/- d( U/)/Ox k) has the correct form in the
limit of two-dimensional turbulence. 7s These two re-
quirements lead to the conditions
~=-1/5, fl2=4/5, ~3=-1/5,
and
2 1
f12--f13"~-~(~2--~3)'q-~(~6--~5) = 2. (5.167)
With the five conditions Eqs (5.165-167), six degrees
of freedom remain in the model.
The only other obvious source of information that
can be used to determine the coefficients is experi-
mental data on the evolution of Reynolds stresses in
homogeneous turbulence. But when the modelled
Reynolds stress equation is deduced from Eq. (5.161),
the coefficients appear in groups--only four of which
are linearly independent of Eqs (5.166-167). Thus two
degrees of indeterminacy remain. Haworth and
Pope 94 remove this indeterminacy by setting
~4 = 0,
and
3 1
y, = - ~ - ~ y * .
This last equation simplifies Eq. (5.165) to
1 3 ) , 0(Uk)
ctl = -- ~+~Co -ct2bijb~i-W bkibe ~
which is used to determine a~. The four remaining
coefficients (assumed constant) are determined by a
least-squares optimization against the available
experimental data. The values obtained are:
~2 = 3.7, y: = 3.01, Ya = -2.18 and Y5 = 4.29.
(5.171)
The values o f T 6 = - - 3.09 and y~ = - 1.28 then follow
from Eqs (5.167) and (5.169).
To summarize: for consistency with Kolmogorov's
scaling laws, the coefficient B in the Langevin
equation, (Eq. 5.148), is determined to be C0e, where
171
the universal constant Co has th~ value 2.1. A model
for the tensor G~j that is linear in both the anisotropy
tensor and the mean velocity gradients has been
proposed (Eqs 5.161-162). For homogeneous
turbulence, the resulting pdf equation yields a joint
normal distribution with the Reynolds stress
evolution closely matching experimental observa-
(5.165) tions.
5.6. Joint Pdf Models
In the previous sub-sections, models for the con-
ditional expectations
(FV2~b,I~),(/~V2UjlV--) and \Ox~ V_ , )
have been developed for homogeneous con-
stant-density turbulence, with no mean composition
gradients. With these models and a knowledge of the
(5.166) turbulent time scale T(t), the evolution of the com-
position pdff6_(.~;t), and of the velocity pdffa(V; t)
can be calculated. But, in order to calculate the evolu-
tion of the velocity-composition joint pdf f(V, ~; t) it
is necessary to model each quantity conditional on the
joint events U = V and ~b = ~. Thus, in this sub-
section, we re-examine the models to see how they
should be modified in order to represent the expecta-
tions
(rv2~lv,~), (.v2ujlv_.~,) and ~ V_,~ .
(5.1721
For the homogeneous flows considered, the joint
lxtf f ( _ ~ ; t) is found to be a joint normal distribu-
tion. ~2 Consequently, it is completely defined by the
means ( U ) and (4~), the Reynolds stresses (uiuj), the
(5o168) scalar variances (~b'_~b~),and the scalar fluxes (ui~b~).
These scalar fluxes contain all the joint information:
~(_y_) can be determined from ( U ) and (uiu~);f#_(.~_)
can be determined from (~b) and ( ~ ' ~ ) ; but the
(5.169) additional information contained in (u~q~) is needed
to determine the joint distribution f(y_,g,). (The
determination of the joint pdf from the first and
second moments is possible only because the pdf is
, known to be joint-normal.)
The models presented above are modified in such a
(5.170) way that the effect of the models on ~ and f~ is
unchanged. The effects of the modifications are
revealed in the behavior of the scalar fluxes.
We consider first the decay of the scalar flux for a
single scalar ~ in isotropic turbulence. For this case
the models described below yield the decay law
d(ui~)
dt* = -C:(ui~ >, (5.173)
where the positive coefficient Cy depends upon the
details of the model.
Since both (~b'2) and (u~ui) decay, it is inevitable
that (ui~) also decays. It is more informative to study
the evolution of the correlation coefficient r.
172 S.B. POPE

2 ,2 I N

q - A ~*(n)~-- r/'*(nbh*(n)~
This definition can be understood by noting that, if
the coordinate system is chosen such that the scalar 1 N
flux (which is a vector) is in the xt-direction, then Eq.
frT*(n)A ,4~*(n)._c.:h*(n)A TT*(nH
(5.174) becomes
(5.179)
r2 = (Ul4p')I/{(UtUt) (~b'2)} (5.175)
Now A6,q~*~"~is zero except, with probability 6t/z u, for
(cf. Eq. 2.110). If k, (~,z) and (u~q~) decay according the two elements p and q. If mixing takes place, then
to Eqs (5.69, 5.25 and 5.173), then it follows that r
evolves by A A*(p) = __1 ~,b,~p~_ d),(¢) ]
~t~', 2 ....... (5.180)
dr
-- = -C,r, (5.176)
dr* and similarly for ^ ,~*~) Thus, using Eq. (5.160) to
substitute for Ao,U *~"', Eq. (5.179) becomes
where
6t
A~,(ud,,) = - - - E [ U , *~P~(4, .* ~. - .~ *. ~ q
C, = C f - ~ (1C ~ + 1 ). (5.177) 2N¢ N
*(q} *(¢) *(p,
+u, (4~, -4~, )]
Sirivat and Warhaft 9~ provide experimental data
for the correlation coefficient r in decaying grid turbu- + ~._ 4,.*~, - ;+=Co or, -
lence. The data show that r decays, but a unique decay n=l T

constant C, is not evident. Values of C, in the range

0.35-0.85 are consistent with the data.
More recently, Shih and Lumley 9~ have proposed a
nonlinear model which agrees well with a second set of
data from Sirivat and Warhaft ~s obtained in grid
turbulence with a uniform mean temperature
gradient. If r * can be neglected in comparison to unity
(which is usually the case) this model reduces to the
decay law Eq. (5.176) with
C, = 0.55C~. (5.178)
For the flows of Sirivat and Warhaft 9~ the value of C~
is found to be approximately 1.6 which yields
C, ~ 1.4. Thus the experimental data do show that r
decays, but the value of the decay constant C, is
uncertain--if, indeed, it is a constant.
We now address the problem of combining models
for ~(V_) and f,_(~) to form a model for f(V,~). The
simplest assumptions are that the increment As~U* is
independent of q~*, and that As~* is independent of
U*. Since we are considering q~ to be a set of passive
scalars, the first of these assumptions is most reason-
able. Certainly for fluid particles, As~U+ is (deter-
ministically) independent of ~+. We adopt this
assumption without reservation or further discussion.
The second assumption--that As~* is independent of
U*--has no justification. However, we tentatively
adopt the assumption in order to explore its conse-
quence.
With the two independence assumptions, the
Langevin model for Aa:LI* (Eq. 5.160) can be com-
bined directly with the stochastic mixing model for
A~tq~* (Eqs 5.35, 5.48 and 5.49). The resulting joint
model is now analyzed to determine its effect on the
scalar flux (u~b,).
With f(V_,~k) being represented by N stochastic
particles U *t"~, q~*~"~, n = 1,2 ..... N, the change in
(u~4~,) in a small time interval 6t is (to first order in fit)
-~-(Co~,)l/2A,t(w(n))t]t
=
6t / 1 3
(5.181)
In the limit as 6t tends to zero we obtain the decay law
Eq. (5.173) with
1 3
C f = C~ +~+-~C o. (5.182)
This implies (Eq. 5.177) that the correlation coefficient
decay constant is
C , = ~3 C o+-~C~.
1
(5.183)
Taking Co = 2.1 and the conservative value C~ = 1.6,
this yields C, ~-2.4--a far higher value than is
suggested by the data.
The reason that the model yields too large a value of
C, lies in the assumption that A~,4~* is independent of
U*. Molecular mixing is a microscale process. When a
fluid particle's composition changes, it does so by
molecular exchange with neighboring fluid particles.
The velocities of neighboring fluid particles--or even
of particles separated by the Kolmogorov scale--are
strongly correlated. But in the stochastic mixing
model, the velocities of the two particles U *¢v~ and
U *l~ are independent.
The following modified mixing model suggests
itself. The first particle to be mixed (denoted by p) is
selected at random as before. Its velocity and com-
position are U *~ and q~*~). The choice of the second
particle (denoted by q) is biased towards those whose
velocity U *~ is close to U *~").The mixing operation is
then performed as before. Depending on the amount
of bias, this model yields decay constants in the range
 Pdf methods for turbulent flows 173

3 1 ~<3 1
~ C o - ~ C ~ <~C r ..~ 4 C o + ~ C , , (5.184) d(uluj) = -Cu,~(uiuj) , (5.189)
dr*

or (taking C o = 2.1, C~ = 2.0), and

0.575 ~< ('~ ~< 2.575. (5.185) d(~b',~}) = -C~(¢p'_tp~). (5.190)

Thus with biased selection, the modified mixing
model, in conjunction with the Langevin equation,
yields values of C, in the range of experimental
observations. (In the improved mixing model ss the
same effect can be achieved by biasing the amount of
mixing according to the velocity difference between
the two particles. This can be done in an efficient
algorithm, whereas an efficient algorithm for biased
selection is not evident.)
We now consider the use of the stochastic mixing
and reorientation models for A6,U* (Section 5.3)
instead of the Langevin equation. If the stochastic
models for A6,U* and A6t4~* are combined, and the
two independence assumptions are invoked (A6tU* is
independent of qS* and A~,4~* is independent of U*),
then the correlation coefficient r decays by Eq. (5.176)
with
1 1
c, = c i + 5 c , + 5 c ~ .
The three constants C~, C, and C, arise from the
three stochastic models--stochastic reorientation,
stochastic mixing in composition space, and
stochastic mixing in velocity space. Taking the values
C~ = 0.5 (C 2 = 1.5), C, = 2.0 and Cu - 1, we obtain
Cr = 2.0--again, a larger value than is suggested by
the data.
Rather than the second independence assumption,
an alternative hypothesis is that steep velocity
gradients cause steep composition gradients, and
hence mixing in g~-space is correlated with mixing in
V-space. The following stochastic model accounts for
this correlated mixing, with the constant C,~
(0 <~ C,o ~< 1)defining the degree of correlation.
The joint pdf f(V__,~k;t) is represented by N
stochastic particles U *("), ~*("), n = 1, 2 .... , N. In the
small time interval fit, the probability of the occur-
rence of correlated mixing is fit/r~,, where
Zu,~ = z/(Cuc~N).
When correlated mixing occurs, two of the N
stochastic particles (denoted by p and q) are selected
at random without replacement and their properties
are replaced by:
1 for these flows is less well established and far more
U*(P)(t + 6t) = U_*(q}(t+ 6t ) = ~ [U*(P)(t) + U*(~)(t)],
1 •
~*(p}(t + 6t) = ~*(~)tt + 6t) = 5 [4' (')(t)+ 4' (q)(t)].
The effect of this model alone on the Reynolds
stresses and scalar variances is:
dt*
Thus, in order to obtain the correct amount of mixing
(cf. Eqs 5.91 and 5.51), the stochastic mixing model in
composition space is used with the constant C,
modified to
C'~ = C , - C , , , (5.1913
and the stochastic mixing model in velocity space is
used with the constant C,( = 1) modified to
C'~ = Cu-C,~. (5.192)
In other words, correlated mixings occur with prob-
ability C,~N&/z, composition mixing with prob-
ability C'~N&/z, and velocity mixing with probability
C'..N ,~t/~.
This combination of the three stochastic mixing
models and the random reorientation model produces
the correct evolution of the Reynolds stresses and
scalar variances, independent of the value of C~. The
(5.186)
effect on the scalar-flux correlation coefficient r is to
cause decay according to Eq. (5.176) with
, 1 1
C, = C 2 + ~ C ~ + ~ C ~ - C , ~ . (5.193)
Thus a choice of Cu~ between unity and zero yields
values of C, in the range
1 _l<c,~<._ , 1 1
C2+~C
~ ' 2.~ C2+~C,+~, (5.194)
or (taking C[ = 0.5, C~ = 2.0)
1.0 ~ C, ~< 2.0, (5.195)
which is closer to the observed range than Eq. (5.186).
5.7. General Flows
The models that have been presented are for inert,
constant-density homogeneous flows. They are well-
understood and tested even though some components
(5.187)
have yet to be finalized. But we wish to apply the joint
pdf equation to more complicated flows--inhomo-
geneous, variable-density reactive flows. Indeed, the
virtue of the pdf equation is that it facilitates the
treatment of these complications. In this section the
extension to general flows is discussed. The modelling
difficult to test. But, it is argued, the exact terms in the
joint pdf equation and the processes already modelled
• account for the principal influences on tbejoint pdf.
5.7.1. lnhomogeneity
(5.188)
For a constant density Newtonian fluid with unit
Prandtl (or Schmidt) number (/z = F), the joint pdf
equation can be re-expressed as
174 S.B. Pore
.
? ( 0f'~ ~3~ f[- / OUjOU~I \]
1 ./,/'au~ aF~\
\]
O~ I- / ~UsO~<l \1
O
Ox Ol,~ [ f (p' l V. ~k)]. (5.196)
J )
The principal influence ofinhomogeneity is to cause
convective transport in physical space. This effect
appears in closed form in the joint pdf equation
through the term VjOf/Ox~. In contrast, in second-
order closures, this term gives rise to the correlations
<UiUjUk>,<UiUj~)'~>t and (uiqY,~b~) which are generally
modelled by gradient diffusion. It is emphasized that,
in the joint pdf equation, the convective term is in
closed form and therefore no .gradient diffusion
models are needed.
The first term on the right-hand side of Eq. (5.196)
represents transport in physical space due to
molecular diffusion, and (at high Reynolds number) is
usually negligible. Nevertheless, the term appears in
closed form and can be retained without additional
modelling being required.
The next three terms represent the effects of
dissipation and pressure fluctuations that have been
modelled in Sections 5.3-6. Since dissipation and (to a
lesser extent) the slow pressure are associated with the
smaller turbulent scales, it is reasonable to assume
that these processes are unaffected by inhomo-
geneities.
For flows remote from walls, it is assumed that the
rapid-pressure models still apply. An analysis of the
rapid pressure rate of strain 99 shows that, to first
order, departures from homogeneity have no effect.
Near walls, however, the velocity gradients and
Reynolds stresses vary rapidly. Further, the analysis of
Section 5.4.1. has to be modified to include surface
integrals at the wall. Consequently, the models (Eqs
5.119 and 5.148) (or at least the tensors C and G) have
to be modified: this is a subject for future work.
The final term in Eq. (5.196) represents transport
due to the fluctuating pressure. Pope 7~ suggested the
model
1
(p'IV, q~) = -~((utut,)-veve), (5.197)
which is consistent with Lumley's modeP ~ for (p'uj).
But this model has not been tested.
5.7.2. Intermittency
Near to the edge of boundary layers and free shear
layers the flow is intermittent--sometimes turbulent,
f~
(I-r)
o %
Fig. 5.8. Pdf of indicator ~br showing a delta function of
magnitude (1 -~,)at ~0;.= 0.
sometimes non-turbulent, t°° Consider a high
Reynolds number jet which transports a conserved
passive scalar 0r(x_, t), where 7 = cr + 1. At the jet exit
4~ is positive, while in the non-turbulent, irrotational
fluid remote from the jet q~r is zero. As the Reynolds
number tends to infinity, the condition ~b.~(x,t ) > 0
can be used as an indicator of turbulent fluid. That is,
if ~b~(~,t)> 0, the fluid is in turbulent motion, if
qbr(_x,t) = 0, the fluid is in irrotational, non-turbulent
motion. The intermittency factor ~,(~,t) is defined as
the probability of the fluid at (x_,t) being turbulent. In
terms of the indicator ~br, the intermittency factor is
y(x,t) = P(c~(x_,tl > 0). (5.198)
The pdf of 4)r, f , , (~07; x, t), is zero for ~O< 0, it has a
delta function of magnitude (1 -?(x~ t)) at ~O~= 0, and
it has a continuous distribution for ~b > 0, see Fig. 5.8.
Integration over the delta function and over the con-
tinuous distribution yields
f + f,~,(~by;x,t)d~b 7 = l-'~(x,t), (5.199)
and
f0 ~ f4, (~Or;x, t)dq/.r = ?(x_,t).
+
The joint pdf f ( ~ ~O;x, t) can be decomposed into
(5.200)
two parts: .the first f r ( ~ qJ; x, t)is the joint pdf of U
and q~ conditional upon the fluid being turbulent, the
second fNV(Y_,~_;x,t) is conditional on the fluid being
non-turbulent:
f = f i r + ( 1 -Y)fN- (5.201)
In terms of f,_~, (V, ~b,~ ; x_,t)--the j oint pdf of ~ q5
and ~br--these two conditional pdf's are
fr(V_, ~b;x_,t) = f (V_,~; x_,t l ~ > O)
y0 ~f,¢~ (V, 0 qJ~;x~ t) d~0JT(x_,t)
(5.202)
and
fs(V,~; x_,t ) = f ( V , ~ ; x_,tl~by=0)
= ;? f.~,(V__,~,O,;x,t)dO,/(1 -~,tx_,t)).
(5.203)
 Pdf methods for turbulent flows 175

The reason for wanting to discriminate between mining z. But there are two problems. First, the
turbulent and non-turbulent fluids is that their representation of turbulence by a single scale most
behavior is quite different. Consequently, a gain in likely limits the validity of the approach to reasonably
accuracy can be expected if different models are used simple flows? °2 Second, since measurements of e are
for the different states of the fluid. This can be done in scarce and difficult to perform, a modelled transport
two ways. The first way (that followed by Kollmann equation for e cannot be adequately tested. Thus the
and Janicka TM) is to derive and model the transport one-point joint pdf method (in common with k-e and
equations for f r , fN and ~,. The modelling that has been Reynolds-stress models) may suffer from the
discussed so far is appropriate to the turbulent part fr, inaccurate (or at least uncertain)determination of e.
and different modelling can be applied to fN. The rate Scale information can be incorporated by con-
at which non-turbulent fluid becomes turbulent is sidering multi-point pdf's, or the joint pdf of state
determined by the transport equation for 7. variables and their derivatives. Work on these pdf's
The second approach--that used by Pope + a - i s to has been performed by Lundgren, l°a Ievlev, 1°+
solve a single transport equation for Qian, l°s Meyers and O'Brien 1°~ and Pope ?°7 None
f~,(V,~,~Oy;x,t), but to use different modelling of these methods has been applied to inhomogeneous
depending on the value of Or- For+ ~ > 0 the flows. While multipoint pdf methods hold great
modelling appropriate to turbulent fluid is used, while promise for the future, they tie outside the scope of the
for ~,~ = 0 non-turbulent models are used. The rate of present work.
creation of turbulent fluid depends on the flux off~.~ Attention is now turned to the scalar dissipation
away from O r = 0. Thus, rather than solv'~ng and the model, Eq. (5.205). For a single, conserved,
equations for the conditional pdf's f r and fN, the passive scalar ~, C+ can be interpreted as the ratio of
equation for f~__~,is solved using conditional modelling. velocity-to-composition decay time scales:
Pope +3 presented such conditional models and used
C+ = U%, (5.206)
them to make calculations of a self-similar plane jet.
The calculated intermittency factor and conditional where
statistics agree well with the experimental data. But
since several of the conditional model constants were ~+ m ~1 ( ~ t2 )/~,. (5.207)
chosen by reference to these experimental data,
further calculations are needed to assess the univer- By measuring the decay of k and (~b'2) in grid turbu-
sality of the modelling. lence, C+ can be deduced. The measurements of Lin
and Lin 1os and those of Warhaft and Lumley 1o9 show
5.7.3. Dissipation rates that C, is not a universal constant. For given experi-
In the particle-interaction and Langevin models, mental conditions, C+ remains constant as the
rate information enters through the dissipation rate turbulence decays: but, depending upon the initial
and through the turbulent time scale t, conditions, values of C~ from 0.6 to 2.4 have been
observed.
z - kl~. Wartiaft 11o performed an experiment in which the
(Since the kinetic energy k can be determined from the scalar fluctuations were induced differently than in
joint pdf, a knowledge of either z or e is sufficient.) In Refs 108 and 109. Inthis case C+ is found in the range
addition, the scalar dissipation rate 1.2-3.1; and, for each of six configurations investi-
gated, C+ decreased by about 40 % over the length of
~ = (2F/p) ( ~ ~3~b~
~}~3x+/' (5.204) the wind tunnel.
In another experiment performed by Warhaft 111
is implicitly modelled by the stochastic mixing model grid turbulence with scalar fluctuations was passed
to be through an axisymmetric contraction. After the con-
traction C, was found to be in the range 1.9-3.1
e,o = C , ( (o',dp'¢)/z, (5.205)
(depending upon the configuration): C+ then in-
[cf. Eq. 5.51 ). This raises the questions: how can ~ (or creased significantly, nearly doubting by the end of the
e) be determined? Is Eq. (5.205) valid? and; if so, is C¢ tunnel.
a universal constant ? The purpose of this subsection is It is abundantly clear that, with measured values in
to outline the problems raised by these questions. the range 0.6-3.1, C+ is not a universal constant. It
For simple flows, ~(x_,t) can be specified. For would appear to be preferable therefore to abandon
example, Pope +:'+3 assumed z to be uniform across a Eq. (5.205) and instead solve a modelled transport
self-similar plane jet. But for most flows, a general equation for ~,p. Such equations have been
means of determining z(x~t) is required. Since k is proposed--by Newman et al., 112 for example. At this
known in terms off, an obvious method is to solve the level of closure, any modelled equation for ~,p implies
standard turbulence-model equation for e.l° Alterna- that (in decaying isotropic turbulence) the time scale
tively a modelled transport equation for T could be ratio C+ has a unique behavior, s° Thus, Newman et
solved. al.'s model causes C+ to remain at its initial value,
To solve a modelled transport equation for e (or z) consistent with the data of Lin and Lin 1°8 and
is. most likely, the best available method for deter- Warhaft and Lumley. 1°9 But the model is then incon-
176 S.B. POPE
sistent with Warhaft's data. 11°'1~: Worse, the prin-
ciple of linearity and independence of conserved
passive scalars constrains the form of the modelled
equation for e~p so that a consistent model is incapable
of allowing C, to remain constant.
The discussion above centers on scalar fluctuations
in decaying grid turbulence. But for shear flows in
which the scalar and velocity fields share a common
history and common boundary conditions, Eq. (5.205)
may be less unrealistic. B6guier et al. ~3 deduced the
value of C a from data of three shear flows: a boundary
layer, pipe flow and a plane wake. The value C# = 2.0
fits the data to within 20 % over nearly all of the flows.
Thus, at present, Eq. (5.205) may provide the best
model for e~B, but further investigations are clearly
needed.
Finally, we draw attention to the fact that none of
the models presented depends upon the molecular
transport properties. At high Reynolds number it can
be assumed H4 that the dissipation processes are
controlled by the large scales of the turbulence, inde-
pendent of the viscosity and diffusivity. However,
many flows of interest are at moderate Reynolds
numbers. It is possible to include a Reynolds number
dependence in the modelling (see, for example, Ref.
11), but this has yet to be done in pdf methods.
At moderate Reynolds number, the different
molecular diffusivities of different species are known
to have a significant effect in some reactive flows.~l ~
Also, when the Prandtl or Schmidt numbers are far
from unity--in liquids, for example--the modelling of
the mixing terms may not be realistic. Again, this is a
topic that has yet to be addressed in pdf methods.
5.7.4. Variable density
In the low Mach number flows considered, the
density p(~b) can vary significantly due to variations in
composition (and enthalpy). In both inert mixing
experiments and combustion experiments, density
variations of a factor of five or more are not
uncommon. It is clear, therefore, that the density
variations have a significant effect on the flow, and
that it would be inappropriate to treat them as small
fluctuations.
In the joint pdfequation (Eq. 3.109) it is remarkable
that the terms pertaining to convection, buoyancy, the
mean pressure gradient, and reaction appear in closed
form irrespective of variation of density in com-
position space. The conditional Lagrangian equation
(cf. Eq. 4.49) corresponding to these terms is
g = _4;) -
It may be seen that the conditional-particle
acceleration due to the mean pressure gradient is
inversely proportional to the density--light fluid is
accelerated more than heavy fluid. Libby and Bray t t6
have suggested that this effect is the cause of counter-
gradient diffusion in premixed flames. None of the
other terms in Eq. (5.208) contains the density
explicitly.
The effect of variable density on the conditional
expectations is largely unknown. The major effect is
likely to be on the fluctuating pressure p'. The Poisson
equation for the pressure becomes
=p~U i~U~ 8p fDU~
D (dU,~ t92% (5.209)
x,0x/
For a constant-density Newtonian fluid, only the first
term is non-zero, but, it may be seen, additional terms
arise due to density gradients, the velocity divergence
and the shear stresses. Not only are there additional
influences on p', but p' also affects the pdf in different
ways. For example, the term
(p,0v,;
~X i /
appears as a source in the turbulent kinetic energy
equation. (In constant-density flows the divergence of
velocity is zero and so pressure fluctuations do not
affect the kinetic energy.)
The effect of variable density on the modelled terms
is a subject f~r future work. It is emphasised, however,
that even in variable-density flows the terms per-
taining to convection, buoyancy, the mean pressure
gradient, and reaction appear in closed form in the
joint pdf equation.
5.7.5. Laminar flamelets
In the stochastic mixing model for f#_(~), it is
assumed that the rate of mixing is inversely propor-
tional to the turbulent time scale z. This assumption
(modified by the comments of Section 5.7.3) is valid
provided that the steepest gradients--which provide
the dominant contribution to (l'V2~b I ~ ) - - a r e caused
by turbulent straining. But rapid reactions can also
produce steep composition gradients. This effect is
studied in more detail for an idealised premixed
turbulent flame.
Let the single scalar q~ be the progress variable
(normally denoted by c 69) that is zero in the unburnt
p
0 0.5 l.o
FIG. 5.9. Density p and reaction rate S against ~ (arbitrary
scale).
 Pdf methods for turbulent flows 177

fuel-air mixture and unity in the fully burnt products.

Assuming simple gradient diffusion, the transport
equation for 4) is (Eq. 3.20)

p -~ = r +pS, (5.210)
h(@)

where the density p(O), the diffusivity F(~), and the

reaction rate S(@) are known functions of q~. Typical I
variations of p(~b), and S(q~) with 4) are shown in Fig. 0.0 0.5 I.O

5.9. Note that S(0) = S(1) = 0, and S(~b)/> 0.

Equation (5.210) admits a steady one-dimensional FIG. 5.11. Quantity h(O) against O, Eq. (5.216) (arbitrary
solution corresponding to a plane premixed laminar scale).
flame. At x~ = - ~ let the velocity be the laminar
flame speed S/ (U1 = S/, U2 = U3 = 0) and let the
mixture be unburnt, ~b = 0. At x~ = m there are pure quently, Eq. (5.215) holds for any plane premixed
products q~ = 1, and the velocity is U1 = Sep(O)/p(1), laminar flame governed by Eq. (5.210), irrespective of
U 2 = U 3 = 0. With these boundary conditions there its orientation or of a superimposed uniform fluid
is a steady one-dimensional solution to Eq. (5.210): velocity.
In the same way, the right-hand side of the con-
~b(x, t) = ~(x 1)' (5.211) servation equation for ~b (Eq. 5.210} is known:
This solution is sketched in Fig. 5.10. (The origin has
been arbitrarily chosen so that ~ ( 0 ) = 1/2.) The
laminar flame speed Se is the unique velocity
U~ (x~ = - ~ ) for which a stationary solution exists. where
From the laminar flame profile ~(xl), any function de
of the solution can be determined. For example, h(Cp) = hx(xl) = ~ F +oS = &p(O)--
dX 1 '

- - - - = 9x(xl ), (5.212) (5.217)

~xi dxi
where The function h(O) is sketched on Fig. 5.11. It is clear
from the last expression in Eq. (5.217) that
Ida(x1)]: h(0) = h(1) = 0, and h(O) >/0.
9x{x,) - i_d---d~--j. (5.213) We return now to high Reynolds number turbulent
Further, since ~(xl) increases monotonically with x 1, premixed flames and consider two extreme cases.
9x(x 1) is a unique function o f ~ : First, when the laminar flame thickness fir is much
larger than the Kolmogorov scale q (&/rl >> 1), the
9,,(xl ) = 0(4)). (5.214) steepest gradients are due to turbulent straining.
Thus, from the above three equations we obtain Then, mixing proceeds at a rate inversely proportional
to the turbulence time scale z in accord with the
stochastic mixing model. In the second extreme case
[~ ~x-~x~]~=¢ = 0(¢). (5.215)
(6e/rl << 1), the steepest gradients are found within
This result has been obtained for a stationary flame laminar flamelets. Burnt and unburnt mixture are
normal to the x~ coordinate direction with the origin separated by a reaction sheet (which may or may not
chosen so that O(0) = 1/2. But the left-hand side of be simply connected). Turbulence strains this sheet
Eq. (5.215) is a scalar that depends only on gradients and causes it to wrinkle. But, in the extreme limit
of ~b--it is unaffected by an arbitrary reorientation, considered, the strain rate is small (compared with
shift, or translation of the coordinate system. Conse- S r / f t ) and the radius of curvature is large (compared
with &). Thus, locally, the flame sheet behaves like a
plane premixed laminar flame.
In the transport equation for the joint pdf.f(V, @),
the conditional expectation of the right-hand side of
Eq. (5.210) appears. With the assumption of laminar
flamelet combustion this becomes

0.5 \ cx~/
where h is the known function given by Eq. (5.217).
We thus reach the remarkable conclusion that for
laminar flamelet combustion the reaction and mixing
terms appear in closed form.
FIG. 5.10. Premixed laminar flame profile e(x 0 against x~. Using this result (Eq. 5.218). Pope and Anand a5
178
have made joint pdf calculations to compare the
properties of premixed flames in flamelet and distri-
buted combustion.
6. S O L U T I O N A L G O R I T H M
6.1. Introduction
Throughout the development, the ability to solve
the modelled joint pdf equation has been a prime
consideration. The joint pdf can be represented com-
putationally by the discrete representation, and its
evolution can be computed via stochastic models. In
this section an algorithm is outlined to solve the
modelled joint pdf equation for an unsteady, variable-
density, three-dimensional turbulent reactive flow.
Variants Of the algorithm for simpler flows are
outlined in Section 6.8.
In the last section, alternative models are presented.
The solution algorithm is described for the velocities
being modelled by the Langevin equation, and the
scalar dissipation being modelled by the stochastic
mixing model. Solution algorithms for other modelled
pdf equations follow by analogy. It is assumed that the
time scale z(x_,t) is known.
The discrete representation, on which the solution
algorithm is based is reviewed briefly in Section 6.2.
From a given initial condition at t = t 0, the algorithm
marches in time with small time steps At. In the time
interval At, several different processes simultaneously
affect the evolution of the joint pdf. But the method of
fractional steps ~~7 (outlined in Section 6.3) is used so
that the effects of some of these processes can be
treated sequentially. In the first fractional step
(Section 6.4), the pdf evolves under the influence of all
processes except stochastic mixing, convection, and
the mean pressure gradient. Stochastic mixing takes
place during the second fractional step which is
described in Section 6.5. In the third fractional step
(Section 6.6), the mean pressure is determined by an
indirect algorithm. Then the joint pdf evolves under
the influence of convection and the mean pressure
gradient.
The algorithm described provides the solution in
terms of the discrete representation of the joint pdf.
But normally, the information sought is mean quanti-
ties, such as CI(.~ t), q~(x_,t) or g~(q~(x_,t)). Furthermore,
the mean velocity, the Reynolds stresses, and the mean
velocity gradients appear in the modelled Ixlf
equation and must, therefore, be determined from the
solution. An obvious way to determine means--that
described in Section 3.4--is to divide physical space
into cells. Then an ensemble average over the particles
in a cell approximates a density-weighted mean at the
cell location, Eq.(3.89). But this method--though con-
ceptually useful--has little merit as a computational
scheme. As described in Section 6.7, the method of
least-squares cubic splines can be used to determine
means from the discrete representation of the joint
pdf. This is a computationally-efficient method that
reduces the statistical error.
S.B. PoPE
6.2. Discrete Representation
We consider the turbulent flow within a three-
dimensional volume of physical space--the solution
domain. At time t, the mass of fluid within the solution
domain is M(t). The mass density function
3~'(y_,~, x; t) is represented by N(t) stochastic particles
in state space, each particle representing a fixed mass
Am. At time t, the state of the nth stochastic particle is
U__*(")(t),~*(")(t), x*("}(t), n = 1, 2..... N(t).
(6.1)
As time proceeds, the statesof the stochasticparticles
change--some particles may leave the solution
domain and some may enter as prescribed by the
boundary conditions. The discrete mass density
function, and the discrete pdf of the stochastic particle
states are given by
~¢~(V~b, x_; t) = pt~/)f~(V__, ~b;~ t)
N{t)
=am y. ~(v - u *("96 (t - ~ *(">)~x_-x
( *(")).
n=|
(6.2)
In the numerical solution, the states of the N
stochastic particles are known, but their expected
states are not. Nonetheless, the connection between
the discrete representation and the joint pdf is best
viewed in terms of expectations:
~(V, ~0,x_;t) = (~-*(V,~0,x; t)) (6.3)
or
f(V,~0; x_,t) = ( f ~ ( V , ~ ; x_,t)). (6.4)
It follows that Y~, must satisfy the consistency con-
dition
f f <~-~,(v__,~, x; t)> dV de
(N,, )
= Am.~l 6(x--x*(")[t]) = (p(x_,t)),
(6.5)
(cf. Eqs 3.82 and 4.107). Equivalently, ~'~ must satisfy
the normalization condition (Eq. 4.108)
ff<~,v(y_,~,x;t)>[p(~,)]-'dVde
( =
,,,,
Am ~ 6(x--x*(")[t p(q~*(")l-t = 1.
II=l
(6.6)
The pressure algorithm ensures that Eqs (6.5-6) are
satisfied.
6.3. Fractional Steps
The modelled equation for ~(V_, ~b,x; t) to be solved
 Pdf methods for turbulent flows 179

~'(t + At) - (I + AtPs)(l + AtP 2)(I + AtP l)~=(t),

~t ' Ox i p(q~) ~/ ~ ~ = (I + At[P 1 + P2 + Ps] + At2[paP2 + PsP1
~ 1 02~
= - 6,~ ~ , [(v~- v ~ ) ~ ] + ~ co~ a v~ o v~ + P2P1] + AtaPaP2P1)~(t)
= ~ ( t ) + At(PI + P2 + Ps)~-~'(t ) + ~9(At2).
(6.14)
A second expression for ~=(t + At) is obtained from the
Taylor expansion of 3~" about t:
~_t 02'~" At 2
The terms on the left-hand side are exact (cf. Eq. 3.109), ~-(t+At)=~.~(t)+ At4 0t 2 2[ ""
while the first two terms on the right-hand side are from
the Langevin model (cf. Eq. 5.150). The final term is due = ~,ar(t)+-~-At+O(At2). (6.15)
to the stochastic mixing model. It is a generalization of
the right-hand side of Eq. (5.40) to a set of ~ com- Thus, equating the right-hand sides of the last two
positions in variable-density flow. equations and dividing by At we obtain
Equation (6.7) can be written
0---[-= (PI + P 2 + P 3 ) ~ r +gP(At), (6.16)
(3t = (P~ + P2 + Pa)~'z;' (6.8)
which, except for the first-order truncation error, is the
same as Eq. (6.8). Consequently ~'(t) approximates
where P~, P: and P3 are operators corresponding to ~r(t) with an error of order At.
different processes. We define The three fractional steps (6.11-13) are qualitatively
the same. We examine the third step since P3 is the
simplest of the operators. In the limit as At tends to zero,
P3 - - Vi ~-~x-+ P(~) 8xg 0V~' (6.9)
and if the other two steps are omitted (or if P~ and P2 are
to correspond to the effects of convection and the mean set to the null operator), then Eq. (6.13) corresponds to
pressure gradient. Similarly, P2 corresponds to the the differential equation
stochastic mixing model and is defined so that P2 ~- is a~
equal to the final term in Eq. (6.7). The remaining
processes are contained in P~ :
c3~ 1 0(p) &~
-v~-~-~j~ p(~) ~xj ~vj" (6.17)
? 8 3S~
That is, it corresponds to ~ evolving under the
~ t3 1 0z influence of convection and the mean pressure gradient
vv~vvj alone. For finite At, Eq. (6.13) corresponds to the
approximate solution of this differential equation by the
(6.10) explicit Euler method, from the initial condition ~2(t):

where I is the identity operator. ~a~"(t+ At) = ~r2(t ) + AtP3,.~(t )

Equation (6.8) shows that all the processes affect the
evolution of ~ simultaneously. But the method of
fractional steps can be used to calculate the evolution of
by considering the sequential action of each of the
processes over sufficiently small time intervals At.
Starting from the given initial condition ~-(t0), the
method is used to calculate ~'(t )-- a first-order approxi-
mation to ~ ( t ) - - at successive time steps. For each time
step, the method is defined by the three fractional steps:
,~l(t) - (I + AtP1)~(t),
~ ( t ) = (l + AtP2)~l(t),
and
~(t+At) = (I+AtPs)~2(t).
a~,(t)
=~2(t)-AtVj
At a(p) ~¢2(t)
-t - - (6.18)
The details of the implementation of the three
fractional steps are described in the next three sub-
sections. While we have defined each step as an explicit
Euler step (Eqs 6.11-13), we note that any other scheme
(6.11) (that is at least first-order accurate) yields qualitatively
the same result (Eq. 6.16). Having established that ~ is a
(6.12) first-order approximation to ~r, it is no longer necessary
to distinguish between the exact and approximate
solution to the evolution equation: henceforth both are
(6.13) denoted by ~-.

Before discussing these steps, we first verify that :~(t) 6.4. First Fractional Step
is, as claimed, a first-order approximation to ~'(t).
Using Eqs (6.11-12) to eliminate ~ and ~a~z, Eq. (6.13) Although the first fractional step contains most of the
becomes processes, it is nevertheless the simplest to explain and
180 S.B. POPE
implement. The step corresponds to solving the
equation
g~r c3~r t~ ~ ~_
1 32~
+-~Coe ovjovj, (6.19)
for a time interval At, from the initial condition ~(t).
This is simply achieved since Eq. (6.19) is the evolution
equation corresponding to the stochastic system
A~,U* = g,ft + G,i[ U* - 0 ~]fit + (Co~)1/2A~,(I4I/).
a d ) ; = s~(~*)rt,
A~tx* = 0, (6.20)
(of. Eqs 4.99-100 ). Integrating these equafi ons for a time
At using the explicit Euler method yields (to first-order
in At):
U*(t + At) = U* + At{g, + Go[U* - 0 ~]}
+ (CoeAt)l/z~,,
~*[t + At) = G* + atsd4~*),
x~'(t ÷ At) = x*, (6.21)
where all quantifies on the right-hand side are at time t,
and ~ is a standardized joint normal random vector.
Thus in the first fraction step, the states of each of the
N(t) stochastic particles change independently
according to Eq. (6.21).
The computational work needed to evaluate Eqs
(6.21), depends crucially upon the reaction scheme. For
each particle in an inert constant-density flow, about
forty arithmetic operations ale needed to perform the
integration. But for a complicated reaction scheme
involving many species (for example Westbrook and
Dryer 6s) many thousands of operations may be
required to evaluate the sources S~(~*). A computation-
ally efficient algorithm is possible only if the sources _S
can be evaluated efficiently. If the number of com-
position variables tr can be reduced to three or less then
S can be tabulated (once) and then 4a operations per
particle are required to obtain the values of S from the
table. For combustion calculation, the simplified
reaction schemes of Westbrook and Dryer: ~9 are well-
suited to this purpose.
6.5. Second Fractional Step: Stochastic
Mixku3 Model
In the second fractional step the position x_* and
velocity U* of the stochastic particles are unaltered,
while the compositions may change by mixing.
The stochastic mixing model is described in Section
5.2.3 for the case of homogeneous turbulence. In that
case, in the small time interval 6t, the probability of a
pair of particles mixing is
C¢,N ft/z.
Consequently, the probability of the nth particle being
selected is
2C~rt/r,
since, if mixing takes place, two out of the N particles are
selected. Similarly, for the general inhomogeneous case,
in the small time interval At the probability of the nth
particle being selected is
1~n) - 2C~At/z(x*(')[t], t). (6.22)
(Clearly At should be chosen to be significantly smaller
than z.)
A simple computational algorithm to select particles
is to generate N independent random numbers (oCn~,
n = 1,2,...,N, each uniformly distributed between
zero and one. If ,
o(~)+P (n~ > 1, (6.23)
then the nth particle is selected for mixing. (A more
efficient, general, and complicated algorithm is given in
Ref. 85.)
Mixing takes place locally in physical space, and this
is reflected in the pairing of the selected particles.
Each particle selected is close (in physical space) to
its partner. A simple algorithm to accomplish this
local pairing, is to divide physical space into cells, and
to pair particles within each cell.t This introduces a
truncation error of the order of the cell size, but there
is little computational penalty in making this size very
small (compared to flow length scales).
The pairs of particles mix, as in the homogeneous
case, according to Eq. (5.48). This completes the second
fractional step.
6.6. Third Fractional Step: Convection and
Mean Pressure Gradient
In the third fraction step, the equation
c3~ 0,9v 1 t3(p) ~ "
at ~-Vj c~xj = p(~k) c3xj ~V2 ' (6.24)
is integrated for a time interval At from the initial
condition ¢~2(t) to yield ~ ( t + At). In this subsection we
describe both the numerical solution of Eq. (6.24) and
the determination of the mean pressure.
Corresponding to Eq. (6.24), the evolution equations
for stochastic particle properties are:
dU* - 1
dt = p(~b*[t])[V(P)]~*"" (6.25 k
d~* = O, (6.26)
dt
and
dx_*
= U*(t). (6.27)
dt
The initial conditions--denoted by x~2),U (2), ~(2)--are
the properties after the second fractional step. Equa-
tions (6.25-27) are integrated approximately so that the
"t"There are several solutions to the problem presented by
having an odd number of particles in a cell--but this is a
computational detail.
 Pdf methods for turbulent flows 181

properties after the third fractional step--~b TM, U (3), An elliptic partial differential equation for ( p ) is now
xt3)--are at least a first-order accurate approximation obtained by manipulating Eq. (6.33).
to the exact solution--U*(t+At), q~*(t+At), If there were no mean pressure gradient, then
x_*(t + At).t _x*(t + At) would assume the value
From Eq. (6.26) we obtain
X' -- x (2) +U(2)At. (6.34)
~(s~ = ~b~2)= q~*(t+At). (6.28)
And in general, from Eqs (6.31-32)it follows
Consequently, in this fractional step, the particle's
density remains constant. To aid the notation, we define
x * ( t + A t ) = x ,- ~ A1t 2 v*P(x (2 ))+tP(At a ). (6.35)
the particle's (constant) specific volume by
v* -- 1/p(t~(2)). (6.29) Substituting this expression for x*(t + At) in Eq. (6.33),
and expanding the delta function as a Taylor series in x
The pressure algorithm described below determines
about _x' yields
the mean pressure field to first order. The gradient of this
mean field is denoted by P(x_),
P(x) = V<p(x_,t)> + ¢(At), (6.30) \ z ox~ /
+ O(At s)
and, for the moment, is assumed to be known. Then,
Eq. (6.25) is integrated by 1 2 0
=h(x)+~At ~x<(~, )2P , ( x (2) )6(x--x)>
,
U(3) = U(z)_ Atv*P(x_(2))
= U*(t+At)+d)(At2). (6.31) + (~(At3), (6.36)
Having determined U*(t + At) to first order, x*(t + At) where
can be determined (from Eq. 6.27) to second order: h(x_) _ <v*6(~-x')>. (6.37)
Xt3) _ X_(2)+ 1 At(U(2 ) + U(3)), Since x_(z) approximates x_'with an error of order At, the
z term in _Pcan be re-expressed as
= x*(t + At) + t~(AtS). (6.32)
< (v*)2pl(x.(2))¢~_- x_')> _- <(~,*)2ei(_x(2))~(x-- x_(2))>
It may be seen from Eq. (6.32) that, as with all
+O(At)
Lagrangian methods, convection is treated without
= <(,~,)2~__x<2))>p,~)
reference to a grid, and spatial derivatives do not have to
be evaluated. Consequently the problems of false dif-
+ O(At)
fusion and artificial viscosity encountered in finite-
difference schemes 11s are avoided.
= ?(~) d O ,j) + ¢(At), (6.38)
To summarize, if the mean pressure is known, the
third fractional step consists of determining 4,(a), U TM
where
and x_(a) from Eqs (6.28), (6.31) and (6.32).
In principle, the mean pressure can be determined by 3'(X) -~= <(~*)26(_X--X(2))>. (6.39)
solving the Poisson equation, Eq. (4.129): But the
evaluation of the source terms--especially d2(p )~Or2 - The last step in Eq. (6.38) simply uses the definitions
is computationally impracticable. Instead, an indirect of _P (Eq. 6.30) and 3'(3_)-Substituting this result (Eq.
algorithm is used which results in an elliptic equation for 6.38) into Eq. (6.36) yields
(p). The algorithm is based on an observation that 1 2_~{ 8<P> 3
follows directly from the propositions established in M - ' = h(x)+~At 3'(_x)--~x+-x+
}+O(At ),
Section 4.7: tlae normalization condition (Eq. 4.108 or
6.6) is satisfied if, and only if, the mean pressure satisfies (6.40)
the Poisson equation. We deduce, then, the mean and hence the required elliptic equation for ( p ) is
pressure field that causes the consistency condition to
be satisfied at time t + At.
)~t~l-~x~; = 2[M -~ - h~_)]/At 2 + e(At).
For simplicity (and without loss of generality) we
consider a single particle (i.e. N = 1). Then, with a mass (6.41)
M of fluid in the solution domain, the consistency
condition at time t + At is The equations derived above are used in a five-step
scheme to determine the mean pressure and then to
M<,:*f(x-x*[t+~t])) = 1. (6.33)
perform the third fractional step:
(i) the function 3'~) (Eq. 6.39) is determined (or rather
+ The superscripts (2) and (3) refer to the states after the
second and third fraction steps, not to the second (n = 2) and approximated) from the initial positions x (2'") and
third 07 = 3) particles. Below. we write x_(2'') to denote the specific volumes v *(")of the N stochastic particles;
state of the nth particle after the second fraction step. (ii) for each particle, x'(") is determined from Eq. (6.34 ) ;
182
(iii) the function h(x) (Eq. 6.37) is determined from x_'(~)
and v *(");
(iv) Eq. (6,41) is solved to obtain a first-order approxi-
marion to (p(x, t)); and then
(v) Eqs (6.31-32) are solved to obtain x (3'") and
U(~.")--approximations to x*(")(t+At) and
__U*(")(t+ At).
6.7. Determination of Means
The algorithm described previously provides a
solution to the joint pdfequation in terms of the discrete
representation, Eq. (6.2). For this solution to be useful, a
method of determining mean quantities from the
discrete representation is required. In order to compare
theory with experiment, it is desirable to determine all
the mean quantities that are measured-- the means ( U )
and ( ~ ) ; the second, third and fourth moments; pdf's
f=, (I/1 ; x_,t) and f,, (qG; x_,t); and the joint pdf's of two
variables. (Joint pdf's of more than two variables
cannot readily be presented graphically.)
Not only are mean quantities required as outputs
from the solution, but they are also needed within the
solution algorithm. In the first fractional step, O,
ÙU/dxj and ul u~' must be determined in order to
evaluate the second-order tensor G O. And in the third
fractional step, the functions h~) and ~(~) (Eqs 6.37 and
6.39) must be determined as part of the pressure
algorithm.
In this section, the determination of means is
considered by reference to a one-dimensional example.
It is shown that the straight-forward method of
approximating means by ensemble averages over cells is
hopelessly inaccurate, especially for derivatives and
pdf's. But the method of least-squares cubic splines
provides a more accurate alternative.
The one-dimensional example considered is of a
normally-distributed random variable ~b(x) in the
interval 0 ~< x ~< 1. The mean is specified to be
(O(x)) = 3x z sin (nx), (6.42)
and the uniform standard deviation is
a m (~b'2) t/2 = I/4. (6.43)
,¢>
I.O
0.0 ,,~-
i I
o.0 0.5 I .o
X
FIG. 6.1. Quantity (q~(x)) against x: • ensemble average
from 1000samples and 20 cells; exact result.
S.B. POPE
The discrete representation of the p d f f , (¢; x) is
1 N
f~,~($;x) = ~- ~ 6 ( * - d p ( " ' ) 3 ( x - x ( " ) ) , (6:44)
ll=l
where x (") is the nth sample of a uniformly-distributed
random variable (0 ~< x (") ~< 1) and ~(") ts
qb(") = ( c~(xt"))) + a~j('), (6.45)
where ~(") is the nth sample of a standardized normal
random variable. The number of samples N is chosen to
be 1000.
The x-axis is divided into twenty equal-sized cells
(K = 20). Figure 6.1 shows the ensemble average of
for each cell compared with the specified mean, Eq.
(6,42). Some scatter is evident, but the overall agreement
is satisfactory: the r.m.s, error is 0.029. The second
derivative d2(~b)/dx 2 is calculated from the cell
averages by the three-point central difference. The
result, shown on Fig. 6.2, shows erratic behavior, the
r.m.s, error being 30.1. Not only is the inaccuracy
unacceptable, but also errors of this magnitude in d tT//
0x=, for example, when fed back into the solution
algorithm, would inevitably lead to numerical in-
stability.
An alternative method of determining means is least-
squares cubic splines (de Boor12°). A cubic spline is a
piecewise cubic polynomial which is continuous and
has continuous first and second derivatives. The x-axis
is divided into L equal intervals, and within each
interval the function (i.e. ( $ ) ) is approximated by a
cubic polynomial. Thus there are 4L polynomial coef-
ficients. At the (L - 1) points between the L intervals, the
function and its first two derivatives are continuous.
<~>/e
40
ZO
0 0.5~• • l.O
"ZO
-40
-60 Q--
FIG. 6.2. Term d2(ek(x))/dx z against ×: • from central
difference of cell averages: exact result.
 Pdf methods for turbulent flows
< ¢~ .> el
2°I
0 culties is to use smoothing cubic splines, x2o. 1,1 with the
-20 -
-40 ei
, I J 6.5 show the histograms for A~ ffi 0.25 and A~ ffi 0.1
0.0 0.5
X compared with the specified Gaussian pdf. It is immedi-
FIG. 6.3. Term d2~O(x))/dx" against x: • from least-
squares cubic spline; exact result.
This condition determines 3 ( L - 1) of the coefficients,
leaving L + 3 to be determined. These remaining coef-
ficients are determined by requiring that the mean-
square error between the spline and the data points
(~b('), x (')) be minimized. (An efficient algorithm using
B-splmes is described by de Boor. x2o)
With six intervals (L = 6), the least-squares cubic
spline approximates <~b(x)) with an r.m.s, error of
0.015 about half that of the cell averages. The approxi-
- -
mation to the second derivative shown on Fig. 6.3 shows
a more striking improvement. The r.m.s, error is 3.7, a
factor of 8 less than that for the cell average. To obtain
the same accuracy with cell averages, the number of
particles would have to be increased to N -- 64,000,
with an increase in storage and computer time of a factor
of 64. Clearly then, least-squares cubic splines provide
an efficient method of determining mean quantities.
The main reason for the improved accuracy of the
least-squares cubic splines is the reduction in the
number of degrees of freedom in the approximation
from 20 (K ffi 20) to 9 (L + 3 ffi 9). But the optimum
value of L is not known a priori--too large a value
f~(~).0t
1.0 ~
0.0 0.5
FIG. 6.4. Histogram (with interval A~ = 0.25) compared with the pdff~,(~; I/2).
183
increases the statistical error, while with too small a
value, the mean profile cannot be resolved. A more
sophisticated technique that overcomes these diffi-
smoothing parameter determined by cross-valida-
tion. 122"~'3 This method has proved successful in
several Monte Carlo solutions of the pdf
equations. 2~.43,4s
We consider now the determination of pdfs from the
discrete representation, taking f , ( ~ ; 1/2) as our
example. A straight-forward method is to form a
histogram with interval A~ from the ensemble of
particles in the cell centred at x ffi 1/2. Figures 6.4 and
1.0
ately apparent that this approximation is inadequate:
for large A~ (Fig. 6.4) the resolution is poor; for small
A~ (Fig. 6.5) the histogram is not smooth.
To estimate the pdffrom a finite number of samples is
a standard problem for which many algorithms have
been proposed, see for example Refs 124-128. (The
method of Crump and Seinf¢ld 12~'1zs is particularly
interesting in that it is directly analogous to cubic spline
smoothing using cross validation.)
All the methods mentioned must compromise
between statistical error and truncation error. If the cell
(in this case at x ffi I/2) is smell, then there are few
samples from which to estimate the pdf. But the
estimated pdf is an average over the cell, so that if the
cell is too large, significant truncation error results.
A method that does not require a compromise
between truncation and statistical error is to use least-
squares cubic splines (in x) to determine the distribu-
tion function F~ (~; x) for the discrete values of ~/
0f ffi ?A~, (~ integer). (6.46)
Then, at x ffi 1/2, a least-squares cubic spline (in ~) is
used to approximate the continuous distribution
function F~(~; I/2). The pdff~(~; 1/2) is then ob-
tained by differentiating F~(~; 1/2) with respect to ~,.
Figure 6.6 shows the pdf obtained using this
technique with AO -- 0.05, 6 splines in x, and 14 sphnes
1.0 I.~
184 S, B. POPE

3.0

2.0

1.0
f
o.o 0.5 t.O t .5

FIG. 6.5. Histogram (with interval A~, = 0.1) compared with the pdff4.(~,; 1/2).

f÷ ok)
2.0

1.0

i/ , , ~'--,.
v

o.o 0.5 hO t.5 q,

Fro. 6.6. Pdff÷($; 1/2) against ~b: . . . . . . Gaussian; from least-squares cubic splines with 1000
samples.

f¢(~k)
2.0

1.0

0.0 0.5 1.0 1.5

FIG. 6.7. Pdff,0k; 1/2) against t#: . . . . . . Gaussian; from least-squares cubic splines with 10,000
samples.
 Pdf methods for turbulent flows
in qJ between q, = - 1/2 and ~, = 2. It may be seen that
the agreement with the Gaussian is better than for the
histograms, but it is still not very good. In addition, the
approximated pdf contains small negative values, thus
violating a fundamental property of pdf's.
Since f , (~O;x) contains all the one-point statistical
information about 4~(x), it is not surprising that it is
difficult to obtain an accurate approximation to the pdf
from just 1000 data points. Figure 6.7 shows the
approximation to f,(q,; 1/2) obtained by the cubic-
spline techniques when the number of data points is
increased by a factor of 10 (i.e. N = 10,000). It may be
seen that the agreement is now satisfactory. Fortun-
ately, satisfactory estimates of simple means (e.g. (4~))
can be obtained with far fewer data points.
6.8. Variants
The solution algorithm has been described for a
general three-dimensional flow. We consider here the
simplifications possible for one and two-dimensional
flows, for self-similar flows, and for boundary-layer-
type flows.
If, in a given coordinate system, the joint pdfdoes not
vary in one coordinate direction, then the flow is
statistically two-dimensional. For example, in polar-
cylindrical (r, 0, z) coordinates, the flow is two-dimen-
sional (axisymmetric) if
~-0 f(V_, ~b; r,/9, z, t) = 0.
In the discrete representation of the joint pdf, the
location of the nth particle is r t~, 0~"~,z ~"~.But since (for
axisymmetric flows) no mean property depends on 0,
the value of 0 ~"~ is irrelevant and need not be calcu-
lated. Thus, the location of the particle is adequately
specified by r t"~ and z ~"~.
The simplification afforded by the two-dimension-
ality is not that 0 ~ need not be calculated, since the
integration of the stochastic equation for 0 ~'~is trivially
simple. Rather, the simplifications are that mean
quantities (e.g. ( U ) , V(p)) are determined as functions
of only two variables, and that (in the stochastic mixing
step) three-dimensional cells are replaced by two-
dimensional cells.
In the implementation of the stochastic models, the
solution domain is divided into K cells, the kth being
centered at -~k) and having volume "Ok"For axisym-
metric flows, the cells are two-dimensional areas on the
r-z plane, the kth centered at (rk,Zk) having an area Ak.
This cell represents the volume Yk of the body of
revolution formed by revolving Ak about the axis:
~k "" 27ZrkAk"
Now with NkIt ) particles in the kth cell, each one
representing a mass Am, the mean density is
( f l ( r k. .7k. t ) ) ~. A m N k / ~ k.
Hence, the particle number density on the r-z plane is
Nk/A k ~ ~ rk(Plrk, Zk, t)).
185
The accuracy with which means can be determined
from the particle properties depends upon the particle
number density. It is evident from Eq. (6.50) that near
the axis of symmetry (small rk) the particle number
density is relatively low and consequently means cannot
be determined accurately. It is desirable to control the
particle number density (and hence the accuracy) inde-
pendently of the coordinates and of the mean fluid
density. This car~ be achieved by a system of adjustable
weights: the nth particle with weight W ~"1represents a
mass W{"}Am. The corresponding discrete representa-
tion of the joint pdfis
,~-* (V__,~b,x; t) = p(~_)f*(V_,~_;x_,t)
N
=Am ~ W~")f(V-U *~"~)
n=l
× 6(q, - ~*~"~)6(x - x*~"l), (6.51)
where the mass of fluid in the solution domain is
N
M = Am ~ W ~").
n=l
From this it follows that for an axisymmetric flow, the
particle number density in the r-z plane is
Nk/A~'~ -~m rk(P(rk'zk't))/fl"k' 116.52)
where ~ is the average weight of particles in the kth cell.
It may be seen that a uniform particle number density
(6.47) can be achieved by choosing the weights to be propor-
tional to r(p).
(The implementation of the stochastic models has to
be modified to account for particles of different
weights.)
For one-dimensional flows, the solution algorithm
becomes still simpler. Only one coordinate is needed,
and all mean quantities are functions of a single spatial
variable.
Self-similar free shear flows--jets, wakes and mixing
layers--are two-dimensional. But a transformation of
the dependent variables (U and ~) can be used to make
the transport equations for statistical quantities
dependent upon a single spatial variable. Thus, for a
self-similar plane jet, Pope 42'43 solved the joint pdf
equation for fV(y__,~_;~, t), where r/ is the normalized
cross-stream variable.
A simplification is also possible for some free shear
flows that are not self-similar. The simplification--
which is analogous to the boundary-layer approxi-
m a t i o n - c a n be applied when the flow is predomin-
antly in one direction. More precisely, there must be
negligible probability of negative velocity in the
(6.48) dominant flow direction. Consider, for example, the
statistically two-dimensional turbulent mixing layer
formed between two parallel streams, Fig. 6.8. The joint
pdf equation is to be solved in the solution domain
(6.49) bounded by the flow surfaces a, b, c and d (that are
infinite in the spanwise x3-direction). A fluid particle
path (in x-space) is sketched on Fig. 6.8. Provided that
the velocity U 1 is always positive, all fluid particles
(6.50) within the solution domain entered through the
186
.....,.. b...... ~ - ~
Uo ~. . . . . ~ ~ 1 ~
U: ---.. --.. ~ --.....,,...__,~,._~. id particle
xlx2
LJ x3
FIG. 6.8. Sketch of a two-dimensional mixing layer showing a fluid particle path.
surfaces a, b and c: none entered through the down-
stream surface d. Consequently, in order to solve the
j pint pdfequation, boundary conditions are required on
the surfaces a, b and c, but not on the surface d. An
algorithm that exploits these observations is now out-
lined.
Rather than the mass density function, the dependent
variable considered is the axial momentum density
function
~¢t(V~k, xl,x2) =- Vl~,a;(V,~_,xl,x2).
From the modelled equation for ~ (Eq. 6.8) we readily
obtain
0Jg O~-
= (P~ +P2 + P ; ) ~
= (Pt + P2 + P'3)dt/Vt,
where the operators P t and P2 are defined as before
(Section 6.3) while the operator P; corresponds to
lateral convection and the mean pressure gradient:
10(p) 0
P3 =- - v~ ~x + p(O) Ox~ ovj"
Starting from the given boundary condition at
x ~ = 0, the method calculates ..g at successive axial
locations separated by a small distance Axe. We define
At(V~) to be the time taken for a particle with constant
axial velocity V, to pass from x t to x t +Ax~ :
At(V,) - A x t / V ,.
Then, from Eq. (6.54) we obtain
¢#(x t + A x , ) = [ I + A t ( V , ) { P t +Pz + P;}]./t'(x,)
+ ~Eat(v,) ~]
= [I + At(1/1)P;] [I + At(Vt)P2]
x El + At(V, )P ~].g (x ~) + CEAt(Vt )2],
where I is the identity operator.
It may be seen that this equation for ~¢(xt +z~x~) is
directly analogous to Eq. (6.14) f o r ff(t+At).
Consequently an algorithm for ~t' directly analogous to
that for .'T described in Sections 6.3-6 is possible. The
S.B. POPE
...- ~ ~1 instantaneousedge
of layer
j u
path
c ~ -,.---l
algorithm for ./t' differs in three respects. First, the
discrete representation is used to approximate ~g rather
than ~'. Consequently each stochastic particle
represents a fixed amount of axial momentum rather
than a fixed mass. Second, in each fractional step, the nth
particle evolves for a time
At {n) ~ / ~ I / U ~ (n), (6.58)
rather than for a fixed time At. Third, the axial pressure
gradient is assumed to be uniform across the flow and is
(6.53)
determined from the free-stream conditions. The
pressure algorithm described in Section 6.6 is used to
determine the lateral pressure gradient.
This algorithm is clearly inapplicable to flows in
which U, is likely to be zero or negative, since then At(")
would be infinite or negative, Eq. (6.58). In addition,
since the truncation error is of order Axl/U*, the
method is most easily applied to flows in which U* is
(6.54)
never very small (compared to a characteristic velocity
difference).
7. COMPOSITION JOINT PDF
(6.55) The velocity-composition joint pdf f(V_,~; x, t)
contains all the one-point statistical information about
the velocity and compositions. In order to solve the
modelled transport equation for f, the only information
required is the turbulent time scale ~(_x,t) (and, of course,
initial and boundary conditions). An alternative
approach that has been used by several authors 25,a7- 36
is to treat the velocity field with a conventional turbu-
(6.56)
lence model, and to solve a modelled transport equation
for the joint pdf of the compositions. This alternative
approach is described in this section.
Most of the composition pdf calculations have been
performed in conjunction with the k-e turbulence
model, t° Modelled transport equations are solved for
the turbulent kinetic energy k and its rate of dissipation
e. The turbulent time scale • is obtained from
(6.57) = k/~, (7.1)
and a turbulent viscosity/~T is defined by
#r = (P)C~,k2/e, (7.2)
where the constant C~, is ascribed the value 0.09. The
turbulent viscosity is used to determine the Reynolds
 Pdf methods for turbulent flows
stresses through the isotropic viscosity hypothesis:
.,, ,,. (dO, doff
+~6u((p) k dO/'~
+ l~r Oxl /"
This is the simplest possible consistent gradient-
diffusion model for momentum transport. The analo-
gous gradient-diffusion model for composition trans-
port is,
(p)ui'q~'~' = - Fr dxi
where the turbulent diffusion coefficient is
F r = ~r/ao,
and the turbulent Schmidt number % is taken to be 0.7.
If the mean density ( p ~ , t ) ) is known, then the
equations for k,e and the mean velocities O form a
closed set. In this approach, the mean density is deter-
mined from the solution to the modelled composition
joint pdf equation.
7.1. Composition Joint Pdf Equation
The joint pdf of the compositions q~is f~l~; x, t). The
transport equation forf¢ can be derived either from the
conservation equation for 4~, (following the procedure
of Section 3.5), or, more simply, by integrating the
equation for f(V, ~/,;x_,t) (Eq. 3.109) over velocity space.
By either method the result is:
• reactions without approximation. Consequently, it
+ ~-~ [o(~)s,(~)f~_] ~_
There is some advantage in considering the density-
weighted (Favre)joint pdf ~ , and the composition
mass density function o~ :
, ~ (~,_x; t) = (p)f~ (~; x, t) = p(~_)f, (~,; x_,t).
The second term in Eq. (7.6) can be re-expressed as
p(~) < Ui I~>f¢ = [ [~, + < u',' I~>] <p >f~
= rc,+<u;'l~,>]~,.
Then the joint pdf equation can be written (in terms of
the composition mass density function) as
( [-/laJ~[ \~ q ~ , finite-difference/Monte Carlo method in which the
The first two terms represent the rate of change of ~
following a particle moving with the mean velocity O;
the third term represents the effect of reaction. Both
these terms are in closed form whereas the terms on the
187
right-hand side of Eq. (7.9) contain conditional expecta-
tions that have to be modelled. The first term on the
right-hand side represents transport in composition
space due to molecular mixing. The stochastic mixing
model (Section 5.2.3) accounts for this process.
(7.3)
The final term in Eq. (7.9) represents transport in
physical space due to turbulent velocity fluctuations. In
the velocity-composition joint pdf equation this
process appears in closed form. But the composition pdf
J~ contains no information about the velocities and so
the conditional expectation (u" [~k) has to be modelled.
Most calculations have been performed using the
(7.4) gradient-diffusion model 24:
(p> (u',' I ~ ) ~ = - r r ~ff. (7.10)
(7.5)
The composition pdf approach--in which transport
equations are solved forf~ (or a~¢), k, ~ and 2 - - has two
disadvantages compared-to the-velocity-composition
pdfapproach--in which transport equations are solved
for f(y_, ~k; x_,t) (or a~-) and e. First, because of the use of
the isotropic viscosity hypothesis (Eq. 7.3), the k-e
model provides a far less accurate closure for the
velocity field. Second, the turbulent convection off¢ has
to be modelled, whereas that offappearsin closed form.
The assumption of gradient-diffusion transport Eq.
(7.10) is particularly questionable in variable-density
reactive flows. Nevertheless, for simple free shear flows,
the k-e model and the gradient-diffusion assumption
may be reasonably accurate. Then the composition pdf
approach has the advantage that the transport equation
for #'~ is simpler to solve than that for #'.
While the composition pdf approach is inferior to the
velocity-composition pdf approach, it still retains the
valuable attribute of being able to treat nonlinear
. (7.6) avoids the closure problem encountered when mean-
flow or second-order models are applied to reactive
flows.
7.2. Solution Aloorithm
(7.7) In this section we outline a solution algorithm for the
modelled composition joint pdf equation that is analo-
gous to the algorithm for fV(y_,~;x_,t) described in
Section 6. But first, we mention two different methods
that have been used previously. Janicka et al. 25 used a
(7.8) finite-difference technique to solve a modelled trans-
port equation for f¢ (~; x 1, x2) in a jet diffusion flame.
For one, or maybe two, composition variables finite-
difference solutions are possible, but, as the number of
variables increases, the computational requirements
become prohibitive. Pope 32 has devised a combined
computational effort increases only linearly with the
number of composition variables. Consequently it can
(7.9) be used 27-32 to solve the equation for the joint pdf of
many compositions.
The algorithm briefly described here (like that of
Section 6) does not require a finite-difference grid in
physical space, and so the truncation errors associated
188
with finite-difference approximations are avoided.
Consequently this algorithm is to be preferred over that
of P o p e . 32
The modelled equation for the composition mass
density function ~'~ (~, x; t) can be written
0
- 0 , [S~(~0)~',] + P 2 ~ , ,
where the operator P2 corresponds to the stochastic
mixing model. Using the method of fractional steps,
reaction and then mixing can be treated in the way
described in Sections 6.4-5. In the third and final
fractional step, the spatial transport terms are treated;
which amounts to solving Eq. (7.1 i) with the right-hand
side set to zero. As is now described, this is readily
achieved since the left-hand side of Eq. (7.11) corre-
sponds to a diffusion process in physical space.
With M being the mass in the solution domain, and
x_*(t) and ~b*(t) being the position and composition of a
stochastic particle, the composition mass density
function can be written
#-~,(q4 x; t) = M(b(~-f~*(t))b(x-x_*(t))).
We consider the diffusion process
a,,x_* (t) = D(x_*Et], t)rt + [B(x*[t], t)] t/2 aer _W,,
where the coefficients D
D_and B are to be determined and
_Wt is an isotropic Weiner process. This process is
directly analogous to Eq. (4.81). Hence from Eqs (4.88),
(4.39) and (7.12) we obtain the corresponding evolution
equation for ~-~:
gt 4-ox, [ D , ~ ] 2 Ox,gx, [B'~'4'] = 0.
In order for the diffusion process Eq. (7.13) to
correspond to the spatial transport in the modelled pdf
equation, Eq. (7.14) should be the same as Eq. (7.11)
with the right-hand side set to zero. This requirement
determines the coefficients to be
B = 2Fr/(p>,
and
D 0 + (p-~--~VFr.
Thus, in the solution procedure, spatial transport
(during the timelnterval At) is accounted for by moving
each stochastic particle independently by
x_*(t+ At)= x*(t)+ At[ O +~p) VFr]..,,
+[2Atrr/<p>]~f~tl~_
where ~ is a standardized joint normal random vector.
S.B. Pop~
8. CONCLUSIONS
The aim of this work has been to describe the
derivation, modelling, and solution of pdf transport
equations for turbulent reactive flows. The main con-
clusions are now summarized, and the attributes of the
velocity-composition joint pdfapproach are discussed
in Section 8.2.
(7.11) 8.1. Summary
At low Mach number, a turbulent flow of a single-
phase mixture of species can be described by the velocity
U(x_, t) and by a set of cr scalars q~(x,t). Any thermo-
chemical property of the mixture can be determined
from q~(x~t) and a reference pressure Po.
The joint pdff(V, ~O;x, t) is the probability density of
the simultaneous events U(x, t) = V_and ~(x, t) = tp. At
each point and time (x, t), the velocity-composition
joint p d f f contains a complete statistical description of
U__and ~, but it contains no two-point information. The
mean of any function of U and q~can be determined from
f, Eq. (3.42). In variable-density inhomogeneous flows,
rather than the joint pdf J~ it is most convenient to
consider the mass density function . f ( V , ~ , x ; t ) =
PilL)f( V, ~L; x, t).
(7.12) The three independent velocity variables V, the a
composition variables ~, and the three position
variables x_ form a (6+a)-dimensional sample space
called state space (or V-L-x_ space). Any random
(7.13) vector (U*,0b*,x*) has the same joint pdf as U(x, t)
and ~b(x,t) provided that:
(i) the joint pdf of x* is proportional to the mean
density <p(~ t)>, Eq. (3.82), and
(ii) the joint pdf of U_* and ~b* is equal to the density-
weighted joint pdf of U and ~b at x_*--
if(V_,~;x*, t)--Eq. (3.84).
Consequently, the mass density function can be
represented by N samples of such a random vector
(7.14) (U*~n),~b*~n',x_*l"~), n = 1,2 ..... N, Eq. (3.76). Each
sample of the random vector corresponds to a point in
state space and can be thought of as the state of a
notional particle. The expected particle number density
in state space is proportional to the mass density
function; the expected particle number density in
x-space is proportional to the mean fluid density; and, at
(7.15) given x_, the expected particle number density in V - ~
space is proportional to the density-weighted joint pdf
(7.16) A transport equation, Eq. (3.1091 for the velocity-
composition joint pdf can be derived without assump-
tion from the transport equations for U and ~b. For a
variable-density flow with arbitrarily complex and
nonlinear reactions, it is remarkable that in this
equation the effects of convection, reaction, body
forces and the mean pressure gradient appear exactly
and do not have to be modelled. The effects of
molecular processes and of the fluctuating pressure
(7.17)
gradient appear in the joint pdf equation as con-
ditional expectations that have to be modelled.
 Pdf methods for turbulent flows 189

The joint pdf equation can also be derived by a stochastic mixing model and the stochastic reorienta-
Lagrangian approach. The Lagrangian conditional tion model. In the stochastic mixing model, pairs of
joint pdf fL (_V,~, x; t[ V o, go, Xo~ is the joint pdf of the particles are randomly selected and their velocities
fluid particle state (U +, ~÷, x_+} at time t, conditional adopt their common mean velocity, Eq. (5.83). In the
upon its state being (Vo,~o, Xo) at time t o. This is the stochastic reorientation model, pairs of particles are
transition density for turbulent reactive flows: the randomly selected and are randomly reorientated in
mass density function at time t is equal to the mass- velocity space (Eq. 5.93). Both energy and momentum
probability-weighted integral of JL over all initial are conserved in this process. The effect of these two
states, Eq. (4.24). models on the Reynolds stresses (Eq. 5.80) is to cause
Many different systems of particles can evolve with decay and return to isotropy in accord with Rotta's
the same pdf as fluid particles. But there is only one model. Mean-velocity gradients give rise to rapid-
such deterministic system. In this system, the N pressure fluctuations. A deterministic model for their
notional particles are called conditional particles and effect (to be used in conjunction with the particle-
have the states (~t"}, ~1"), ~"~; n = 1,2 ..... N). The interaction models) is presented in Eq. (5.122) and
rate of change of state of conditional particles is equal shown to be consistent with Reynolds-stress models,
to the conditional expectation of the rate of change of Eq. (5.126).
state of fluid particles. The trajectories of conditional In the alternative--Langevin--model, the velocity
particles in the augmented state space ( V - ~ - x _ - t U*(t) of a stochastic particle evolves according to a
spacet are called conditional paths. These conditional diffusion process, Eq. (5.148). In the small time
paths define a formal solution to the joint pdf trans- interval fit, U*(t) changes deterministicaily by an
port equation, Eqs (4.52) and (4.56). A physical inter- amount ~tGoU ~, and randomly by amount (Btt)l/2~i
pretation of this solution is provided in Section 4 . 5 - (where ~ is a standardized joint normal random
loosely, probability is transported along conditional vector). For consistency with Kolmogorov's scaling
paths. laws, the coefficient B is determined to be Co~, where
The essence of the whole approach is to construct a Co = 2.1 is a universal constant. A model for the
system of stochastic particles whose evolution is tensor Go has been proposed (Eqs 5.161-162). For
simply computed and in which the pdf evolves in the homogeneous turbulence, the resulting pdf equation
same way as the pdf of fluid particles. Poisson yields a joint normal distribution with the Reynolds-
processes and diffusion processes (described in stress evolution closely matching experimental data.
Section 4.6~ provide the basis for such a stochastic The extension of these models to inhomogeneous,
system. variable-density reactive flows is discussed in Sections
The modelling is guided primarily by the known 5.6 and 5.7.
behavior of homogeneous turbulent flows. In these In Section 6, a solution algorithm is described to
simple flows, the joint pdf f ( V , ~ ; t) adopts "a"joint solve the modelled velocity-composition joint pdf
normal distribution. The aim of the modelling is there- equation for unsteady, three-dimensional, variable-
fore to produce a joint normal distribution whose first density flows. The velocities are modelled by the
and second moments evolve correctly. Langevin equation, and the scalar dissipation by the
For homogeneous turbulence, the joint pdf stochastic mixing model.
f(Y_.~;tl is represented by an ensemble of N The mass density function is represented by N
stochastic particles with states (U*~"~(t), ~*~"~(t); stochastic particles in state space, Eq. (6.2). Given the
n = 1,2 ..... N I. The modelling is performed, not states of the particles at time t, the states at time t + At
directly in the joint pdf equation, but indirectly by are calculated through a sequence of three fractional
prescribing the behavior of the N stochastic particles. steps. In the first fractional step, the velocities and
The primary effect of molecular diffusion on the compositions evolve due to reaction, the Langevin
composition variables ~b is to decrease the variance model and buoyancy, Eq. (6.21). In the second frac-
<~'~05~> while leaving the mean (qS~> unchanged, Eqs tional step, the particle compositions change
(5.50-511. The stochastic mixing model is used to according to the stochastic mixing model. And in the
produce this effect. Pairs of stochastic particles are third fractional step, the particle positions and
randomly selected from the ensemble (at a rate deter- velocities change due to convection and the mean
mined by the turbulent time scale ~) and their com- pressure gradient.
positions ~* adopt their common mean composition, The determination of the mean pressure field is an
Eq. ~5.481. integral part of the third fractional step. The indirect
Alternative models are presented for the effects of algorithm used is based on the observation that the
the fluctuating pressure gradient and viscous dissi- stochastic particle number density in physical space
pation. The first is based on stochastic particle-inter- should be proportional to the mean fluid density. The
action models, and the second on the Langevin mean pressure is determined--as the solution of an
equation. elliptic partial differential equation--by requiring that
With the particle-interaction models, in homo- this condition be satisfied. (Subtle connections
geneous turbulence with no mean-velocity gradients, between the mean pressure, the mean continuity
the decay of turbulence energy and the return to equation, and consistency conditions are discussed in
isotropy of the Reynolds stresses is simulated by the Section 4.7).
190 S.B. POPE
The algorithm provides a solution to the pdf equa-
tion in terms of stochastic particle properties. A con-
ceptually simple way to obtain mean quantities is to
divide physical space into cells, and then to form
ensemble averages over the particles in each cell. In
Section 6.7 it is shown that this straightforward
method is hopelessly inaccurate, and better methods
based on cubic splines are described.
For simpler flows--one and two-dimensional flows,
self-similar flows, and boundary-layer-type flows--
variants of the solution algorithm are described
in Section 6.8.
The composition joint pdf approach is discussed in
Section 7. In this approach, the mean velocity and
turbulence fields are determined using a standard
turbulence model and a modelled transport equation
is solved for the joint pdf of the compositions
f#-(~;x_, t). This approach is inferior to the velocity-
composition approach in that a turbulence model is
required and the turbulent transport off#_ has to be
modelled--usually by gradient diffusion, Neverthe-
less, nonlinear reactions can still be handled without
approximation, and the approach may be adequate
for simple flows. A solution algorithm for the
modelled composition transport equation is de-
scribed in Section 7.2.
8.2. Discussion
In the following four subsections we discuss: the
inherent limitations of one-point closures; the
advantages of the pdf approach compared to other
one-point closures; aspects of the modelling in need of
improvement; and the computational requirements of
the solution algorithm.
8.2.1. Limitations of one-point closures
The joint pdf method--in common with most other
turbulence modelsl°'lt--is a one-point closure. The
joint pdf contains no length-scale or time-scale in-
formation (see Section 2.9). In modelling the joint pdf
equation, the time scale z has to be specified either
directly, or indirectly through a modelled transport
equation for e. The direct specification of scale
information is only possible in simple flows, and the
validity of the modelled dissipation equation has often
been called into question (e.g. Ref. 102). In addition, as
discussed in Section 5.7.3, one-point closures are
incapable of cor.rectly determining the scalar dissi-
pation rate--at least in grid turbulence.
These inherent problems do not invalidate the
approach. For simple shear flows, for example, the use
of the modelled dissipation equation is justifiable, 1°2
and the scalar dissipation time scale appears to be
simply proportional to r . t l 3 For more complex
flows--with rapidly changing scales--the modelling is
less secure, but the magnitudes of the errors resulting
from the modelling inadequacies have yet to be
quantified, tSeveral multi-point pdf methods have
been suggested.) t°a- 1o7
8.2.2. Advantages of the pdf approach
Since the joint pdf provides a complete one-point
statistical description of the flow field, the pdf
approach has advantages over other one-point
closures. The principal advantages are that reaction
and convection transport can be treated without
approximation, even in variable density flows.
For constant-density homogeneous flows, the
models for the effects of the fluctuating pressure
gradient and for molecular mixing are compatible
with Reynolds-stress models and hence are equally
accurate. For inhomogeneous flows, the joint pdf
equation can be expected to be more accurate since
additional models are required in the Reynolds-stress
equations. Thus even for inert, constant-density flows.
joint pdf calculations can be expected to be more
accurate than Reynolds-stress calculations. But the
pdf approach has the additional advantage of being
able to handle, without approximation, nonlinear
reactions and the effect of variable density on con-
vection.
8.2.3. Model improvements
Several aspects of the modelling need to be refined,
but we identify two areas in particular need of further
study.
The treatment of scalar dissipation by the improved
stochastic mixing is only partially satisfactory.
Prandtl or Schmidt number effects are not accounted
for nor can the model be justified when reaction times
are of the order of the Kolmogorov time scale (or
smaller).
As discussed in Section 5.7.4, the large density
variations encountered in flows of interest undoubt-
edly have a significant influence on the modelled
terms. But little is known about the nature and mag-
nitude of variable-density effects.
In addressing these and other modelling questions,
we expect that direct numerical simulations of homo-
geneous turbulence 129-~31 will play a central role. In
the past, modelling has been guided by theory and
experimental data. But it is now possible to solve the
governing flow equations for homogeneous turbu-
lence at Reynolds numbers comparable to those of
wind-tunnel experiments. From the computed flow
fields, multi-point statistical information can be
extracted, and this provides an additional source of
guidance to modelling. In the joint pdf equation, the
modelling is most naturally viewed in Lagrangian
terms. While Lagrangian statistics are extremely diffi-
cult to obtain experimentally, it is relatively easy to
obtain them from direct numerical simulations.~ 31
8.2.4. Computationalconsiderations
The usefulness of the pdf approach depends on the
existence of a practicable means of solving the
modelled equations. Many Monte Carlo calculations
have been performed, both for the velocity-com-
position joint pdf 42 46 and for the composition joint
pdf. 27-32 Typical computational requirements are
 Pdf methods for turbulent flows
50,000 words of storage, and 5 min of CPU time on a
mainframe computer.
None of the calculations referenced above are for
the general three-dimensional case for which the
solution algorithm has been described (Section 6): the
most difficult case treated so far is a jet diffusion
flame. .6 Nevertheless, the modest computer require-
ments for the calculations performed give every
reason to suppose that the M o n t e Carlo method will
prove practicable for more difficult cases.
Acknowledgments--For comments and suggestions on the
first version of this paper, I am grateful to Professors W.
Kollmann, P. A, Libby and W. C. Reynolds as well as to the
editor and reviewers. I am grateful to D. C. Haworth for
providing suggestions and corrections on the revised version,
and for allowing me to use the unpubl'_'shed results of our
collaborative work (Ref. 94). And I thank Barbara Benedict
for her careful and patient typing of the paper.
This work was performed at the Massachusetts Institute of
Technology and at Cornell University. It was supported in
part by grant numbers CPE8000026 and CPE8212661 from
the National Science Foundation (Engineering Energetics
Program), and by grant numbers DKAG29-82-K-0017 and
DAAG29-84-K-0020 from the U.S. Army Research Office.
REFERENCES
1. MURTHY,S. N. B. (Ed.), Turbulent Mixina in Nonreactive
and Reactive Flows, Plenum Press (1975).
2. BRACCO,F. V. (Ed.), Combust. Sci. Technol. 13, 1-275
(1976).
3. BRODKEY,R. S. (Ed.), Turbulence in Mixino Operations,
Academic Press (1975).
4. LIBBY, P. A. and WmLIXMS, F. A. (Eds), Turbulent
Reactive Flows, Springer-Verlag (1980).
5. LmBY,P. A. and WILLIAMS,F. A., A. Rev. Fluid Mech. 8,
351-376 (1976).
6. KELLER, J. U. (Ed.), J. Non-equilib. Thermodyn. 6,
321-386 (1981).
7. KELLER,J. U. (Ed.), J. Non-equilib. Thermadyn. 7, 1-70
(1982).
8. JONES,W. P. and WHITELAW,J. H., Combust. Flame 48, 1
(1982).
9. LIEPMANN,H. W., Am. Sci. 67, 221-228 (1979).
10. LAUNDER,B. E. and SPALDING, D. B., Mathematical
Models of Turbulence, Academic Press (1972).
11. LUMLEY,J. L., Adv. appl. Mech. 18, 123 (1978).
12. HAWTHORNE,W. R., WEDDELL,D. S. and Ho~ lea,, H. C.,
Third Symposium on Combustion, Flame and Explosive
Phenomena, p. 266, Williams and Wilkins (1949).
13. BRAY, K. N. C. and Moss, J. B., University of
Southampton Report AASU335 (1974).
14. LOCKWOOD,F. C. and NAGUm,A. S., Combust. Flame 24,
109 (1975).
15. RHODES,R. P.,in Ref. 1, p. 235.
16. POPE,S. B., J. Non-Equilib. Thermodyn. 4, 309 (1979).
17. POPE,S. B., Phil. Trans. R. Soc. LoAd. A291, 529 (1979).
18. POPE,S. B., Turbulent Shear Flows 2 (L. J. S. Bradbury et
aL, Eds), 7-16, Springer-Verlag (1980).
19. LUNDGREN,T. S., Phys. Fluids 12, 485 (1969).
20. LAUNDER,B. E., REECE,G. J. and RODI,W., J. Fluid Mech.
68, 537 (1975).
21. DOPAZO.C. and O'BmEN, E. E., Acta astronaut. I, 1239
(1974).
22. DOPAZO.C. and O'BmEN, E. E., Phys. Fluids 17, 1968
(1975).
23. DoP Azo. C. and O'BmEN, E. E., C ombust.Sci. Technol.13,
99 (1976).
191
24. POPE,S. B., Combust. Flame 27, 299 (1976).
25. JANICKA,J., KOLBE,W. and KOLLMANN,W., Proc. Heat
Transf Fluid Mech. Inst., 296-312, Stanford University
Press (1978).
26. J^~CKA, J., KOLBE, W. and KOLLMANN,W., J. Non-
equilib. Thermodyn. 4, 47 (1978).
27. NGuY~, T. V. and POPE,S. B., Combust. Sci. Technol. 42,
13-45 (1984).
28. POPE,S. B., E~hteenth Symposium (lntl) on Combustion,
1001-1010, The Combustion Institute (1981).
29. McNuTT, D. G., M.S. Thesis, Massachusetts Institute of
Technology (1981).
30. SHI~KAR, S. V. and CHEVRAY,R., Third Symposium on
Turbulent Shear Flows, Davis, CA (1981).
31. GM, P., SImONANO,W. A. and POPE,S. B., Combust. Sci.
Technol. 37, 59-78 (1984).
32. PopE, S. B., Combust. Sci. Technol. 25, 159 (1981).
33. Doexzo, C., Phys. Fluids 18, 397 (1975).
34. KOLLMANN,W. and JANICKA,J., Phys. Fluids 25, 1755-
1769 (1982).
35. KOLLMANN, W. and SCHMITT, F., Seventh Biennial
Symposium on Turbulence, University of Missouri-Rolla
(1981).
36. WU, M.-Z. and O'BRIEN, E. E., Combust. Sci. Technol.
29, 53 (1982).
37. HANDSCOMa, D. C. and HAMMERSLEY,J. M., Monte
Carlo Methods, Methuen (1965).
38. DAHLQUIST,G. and BJORCK, A., Numerical Methods,
Prentice-Hall (1974).
39. POPE, S. B., MIT Report EL-80-012 (1980) (Revised
version appeared as Ref. 32).
40. SPmLMXN,L. A. and LEVENSPIEL,O., Chem. Engng Sci.
20, 247 (1965).
41. FLAGAN,R. C. and APPLETON,J. P., Combust. Flame 23,
249-267 (1974).
42. POPE,S. B., Turbulent Shear Flows 3 (L. J. S. Bradbury et
al., Eds), 113-123, Springer-Verlag (1982).
43. PopE, S. B., AIAA J. 22, 896 (1984).
44. ANAND,M. S. and POPE, S. B., Turbulent Shear Flows 4
(L. J. S. Bradbury et al., Eds), 46-61, Springer-Verlag
(1985).
45. PopE, S. B. and ANAND,M. S., Twentieth Syrup. (Into on
Combustion, The Combustion Institute (to be published)
(1985).
46. PopE, S. B. (in preparation) (1985).
47. FROST, V. A., Third All-Union ConJerence on Com-
bustion, Press. of Acad. Sci. USSR 1 (1960).
48. FROST,V. A., Fluid Mech., Soy. Res. 4, 124 (1975).
49. CHUNG,P. M. AIAA J. 7, 1982 (1969).
50. CHUNG,P. M., Combust. Sci. Technol. 13, 123 (1976).
51. [,ANGEVIN,P., C. R. Acad. Sci. Paris 146, 530 (1908).
52. BYWATER,R. J., AIAA J. 20, 824 (1982).
53. HINZE, J. O., Turbulence (2nd Edn), McGraw-Hill
(1975).
54. DRAKE, A. W., Fundamentals of Applied Probability
Theory, McGraw-Hill (1967).
55. GNEDENKO,B. V., The Theory of Probability, Chelsea
(1962).
56. LUMLEY, J. L., Stochastic Tools in Turbulence,
Academic Press (1970).
57. PANCttEV, S., Random Functions and Turbulence,
Pergamon Press (1971).
58. BOTKOV, E., Mathematical Physics, Addison-Wesley
(1968).
59. LIGHTHILL, M. J., Fourier Analysis and Generalized
Functions, Cambridge University Press (1970).
60. O'BRIEN,E. E., in Ref. 4, p. 185.
61. SPENCER, B. W., Ph.D. Thesis, University of Illinois
(1970).
62. TAVOULAmS,S. and CORRSlN, S., J. Fluid Mech., 104,
311-347 (1981).
63. DRAKE,M. C., PITZ, R. W. and LAPP, M.. AIAA paper
no. 84-0544 (1984).
64. JEFFREYS, H. and JEFFREYS, B., Methods o! Mathe-
192 S.B. POPE

matical Physics (3rd Edn), pp. 182-187, Cambridge 99. POPE, S. B., Ph.D. Thesis, University of London
University Press (1972). (1976).
65. BIRD, R. B., STEWARD, W. E. and LIGHTPOOT, E. N., 100. CORRSIN,S. and KISTLER, A. L., NACA Report 1244
Transport Phenomena, John Wiley and Sons (1960). (1955).
66. WILLIAMS,F. A., Combustion Theory, Addison-Wesley 101. KOLLMANN,W. and JANICKA.J., Phys. Fluids 25. 1755-
(1965). 1769 (1982).
67. STRAHLE,W. C., Combust. Sci. Technol. 29, 243 (1982). 102. POPE, S. B., The Limits of Applicability of Turbulent
68. WESTBROOK,C. K. and DRYER, F. L., Combust. Sci. Scale Equations, Report FS/77/7, Imperial College
Technol. 20, 125-140 (1979). (1977).
69. BRAY,K.N.C.,in Ref. 4, p. 115. 103. LUNDGREN,T. S., Statistical Models and Turbulence (M.
70. BILGER,R. W., in Ref. 4, p. 65. Rosenblatt and C. Van Atta, Eds), p. 70, Springer-
71. DRAKE, M. C., LAPP, M., PENNEY, C. M., WAP,SnAW, Verlag, Berlin (1971 ).
S. and GERHOLD,B. W., Eighteenth Symposium (Intl) on 104. IEVLEV,V. M., Dokl. Akad. Nauk. 208, 1044 (1973).
Combustion, p. 1521, The Combustion Institute (1981). 105. QIAN,J., Phys. Fluids 26, 2098-2104 (1983).
72. POPE,S. B., Phys. Fluids 24, 588-596 (1981). 106. MEYERS, R. E. and O'BRIEN, E. E.. Combust. Sci.
73. ARNOLD, L., Stochastic Differential Equations: Theory Technol. 26, 123-134 (1981).
and Applications, Wiley (1974). 107. POPE, S. B., Phys. Fluids 26, 3448-3450 (1983).
74. DURmN, P. A., NASA RP 1103 (1983). 108. LIN, S. C. and LIN, S. C., Phys. Fluids 16, 1587-1598
75. WAX,N., Noise and Stochastic Processes, Dover (1954). (1973).
76. VAN KAMPEN,N. G., Stochastic Processes in Physics and 109. WARHAFT,Z. and LUMLEY,J. L., J. Fluid Mech. 88, 659
Chemistry, North-Holland (1981). (1978).
77. SCHUMANN,U., Phys. Fluids 20, 721-725 (1977). 110. WARHAFT,Z., J. Fluid Mech. 104, 93 (1981 ).
78. SPEZtALE, C. G., Geophys. astrophys. Fluid Dyn. 23, 111. WARHAFT,Z., J. Fluid Mech. 99, 545 (1980).
62-84 (1983). 112. NEWMAN,G. R., LAUNDER,B. E. and LUMLEY,J. L., J.
79. SHIH,T.-H., Refined Modeling of Flows (J. P. Beneque Fluid Mech. 111,217 (1981).
et al., Eds), pp. 239-250, Paris: Presses Pont it 113. BEGUIER,C., DEKEYSER,I. and LAUNDER,B. E., Phys.
Chaussees (1983). Fluids 27, 307-310 (1978).
80. POPE,S. B., Phys. Fluids 26, 400-408 (1983). 114. TENNEKES,H. and LUMLEY, J. L.. A First Course in
81. SPALDING,D. B., Chem. Enang Sci. 26, 95 (1971). Turbulence, M IT Press, Cambridge (1972).
82. POPE, S. B., Combust. Flame 34, 103-105 (1979). 115. BILGER,R. W., AIAA J. 20, 962 (1982).
83. DOPAZO,C., Phys. Fluids 22, 20 (1979). 116. LmnY, P. A. and BRAY,K. N. C., Combust. Flame 39, 33
84. CURL,R. L., A.I.Ch.E.J. 9, 175 (1963). (1980).
85. POPE, S. B., Combust. Sci. Technol. 28, 131-145 (1982). 117. YANENKO, N. N., The Method of Fractional Steps,
86. COMTE-BELLOT,G. and CORRSlN,S., J. Fluid Mech. 18, Springer-Verlag (1971).
353-378 (1964). 118. ROACHE,P. J., Computational Fluid Dynamics, Hermosa
87. CHOI, K.-S. and LUMLEY,J. L., Proc. I U T A M Syrup. on (1972).
Turbulence and Chaotic Phenomena in Fluids, Kyoto, 119. WESTBROOK,C. K. and DRYER, F. L., Combast. Sci.
Japan (1983). Technol. 27, 31-43 (1981).
88. ROTTA,J. C., Z. Phys. 129, 547 (1951). 120. DE BOOR, C., A Practical Guide to Splines, Springer-
89. HUNT,J. R. C., Fluid Dyn. Trans. 9, 121-152 (1977). Verlag (1978).
90. LUMLEY,J. L. and KHAJEn-NouRI, B., Adv. Geophys. 121. REINSCH,C. H., Num. Math. 10, 177-183 (1967).
18A, 169 (1974). 122. CRAVEN,P. and WAHRA,G., Num. Math. 31, 377-403
91. LIN, A. and WOLFSHTEIN, M., First Symposium on (1979).
Turbulent Shear Flows, p. 437, Pennsylvania State Uni- 123. DIACO~S, P. and EFFROY, B., Sci. Am. 248, 116-130
versity (1977). (1983).
92. KELLOG, O. D., Foundations of Potential Theory, 124. BEAN, S. J. and TSOKOS, C. P., Int. statist. Rev. 48,
Springer-Verlag (1967). 267-287 (1980).
93. MONIN, A. S. and YAGLOM, A. M., Statistical Fluid 125. SCOTT,D. W. and FACTOR,L. E., d. Am. statist. Ass. 76,
Mechanics, Vol. 2, MIT Press, Cambridge (1975). 9-15 (1981).
94. HAWORTH D. C. and POPE, S. B., A generalized 126. GOOD, I. J. and GASKINS,R. A., J. Am. statist. Ass. 75,
Langevin model for turbulent flow. Submitted to Phys. 42-73 (1980).
Fluids (1985). 127. CRUMP, J. G. and SEINFELD, J. H., Aerosol Sci.
95. JONES, W. P. and MUSONGE,P., Fourth Symposium on Technol. 1, 15-34 (1982).
Turbulent Shear Flows, Karlsruhe (1983). 128. CRUMP,J. G. and SEINFELD,J. H., Aerosol Sci. Technol.
96. SIRIVAT, A. and WARI-IAET,Z., J. Fluid Mech. 120, 1,363-369 (1982).
475-501 (1982). 129. ROGALLO,R. S. and MO1N, P., A. Rev. Fluid Mech. 16,
97. SmH, T.-H. and LUMLEY, J. L., Report FDA-81-06, 99-137 (1984).
Cornell University (1981 ). 130. ROGALLO,R. S., NASA TM-81315 (1981).
98. SIRIV*AT,A. and WARHAFT, Z., J. Fluid Mech. 128, 131. RILEY,J. J. and PATTERSON,G. S., Phys. Fluids 17, 292
323-346 (1981). (1974).