HAL Id: jpa-00247956
https://hal.archives-ouvertes.fr/jpa-00247956
Submitted on 1 Jan 1994
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Velocity intermittency in turbulence : how to objectively characterize it ?
Antoine Naert, Laurent Puech, Benoît Chabaud, Joachim Peinke, Bernard Castaing, Bernard Hebral
To cite this version:
Antoine Naert, Laurent Puech, Benoît Chabaud, Joachim Peinke, Bernard Castaing, et al.. Veloc-
ity intermittency in turbulence : how to objectively characterize it ?. Journal de Physique II, EDP
Sciences, 1994, 4 (2), pp.215-224. �10.1051/jp2:1994124�. �jpa-00247956�
J. Phys II France 4 j1994) 215-224 FEBRUARY 1994, PAGE 215
Classification Physics Abstracts
47.25
Velocity intermittency in turbulence
:how to objectively
characterize it ?
Antoine Naert, Laurent Puech, Benoit Chabaud, Joachim Peinke, Bemard
Castaing
andBernard Hebral
Centre de Recherches sur (es Trds Basses Tempdratures (*), CNRS, B-P- 166, 3804? Grenoble Cedex 9, France
(Received 22 Marc-h J993, re>,ised 2J september J993, accepted 25 October J993j
R4sum4. L'dvolution d'un certain paramdtre A~ ddfini dans l'article a dtd utilisde pour tenter de trancher entre diffdrents moddles de l'intermittence en turbulence. Le but de cet article est de valider une mdthode
e~pdrimentale
de ddterrnination de ce paramdtre,inddpendamment
de toutmeddle, h partir des densitds de
probabilitd
de diffdrences de vitesse (pdo. Les conditionsexpdrimentales
quiperrnettraient
de pousser plus loin la caractdrisation de ces pdf sent dgalement discutdes.Abstract. The evolution of A~, a parameter whose definition is recalled in the paper, has been used in order to test various interrnittency models in turbulence. This paper aims to validate an experimental, model independent, method for
extracting
this parameter from thevelocity
difference
probability density
functions(pdo.
We also discuss the charactenstics of an experiment which could push further the analysis of these pdf.1. Introduction.
Statistics of
velocity
difference between twopoints
is the main tool forcharacterizing developed
turbulent flows[1, 21. Recently
theinterejt
shifted from the moments of thedistribution of these
velocity
differences to thestudy
of the wholeprobability density
function(pdf~ [3-51.
Models wereproposed [4-91,
based on variousphysical grounds,
each onegiving
agood
agreement with the observedpdfs.
Thus theproblem
is less in this agreement than in theinsight
thecorresponding analysis gives
into theunderlying physics.
Most of the
proposed
models can bepresented
in a unified wayconnecting
ther to the
large
scale one[10, 1].
If &u is the considered component of thevelocity
difference&u
=
u,(x
+r) u~(x)
(*) Laboratoire associd h l'Universitd Joseph Fourier.
and
P~
~~ itsprobability density
function(m)
=
(&u~) ),
these models assume that"r "r
)Pr~~) "iGr~~)PL()j$ (I)
r r r
where L is the
large integral
scale and G a distribution whichdepends
on the model.In this paper we show the interest of
using
alog
normal ansatz for G~:~
ln~ «/«~
G~ cc exp
~
(2)
"r 2 A
(in general
«~, whichgives
the maximum ofG,
is different from«~).
This interest does notonly
lie in that this should be thelimiting
form for G~ at a small rlO].
Themajor
interest is that the parameter A~ obtained infitting
theexperimental P~
withequations (1, 2),
measures themean
squared
deviation((&
In«)~),
even if G~significantly
differs from alog
normal distribution. This isimportant
for thefollowing
reasons. Theexisting
theories onintermittency
in
developed
turbulencepredict
both differentshapes
for G and different behaviours for((&
In«)~)
ve;sus r,Physically (
(& In«)~)
measures theprogression
ofintermittency along
the scales
(number
of steps in the cascademodels).
For instance theKolmogorov
Obukhovtheory [2],
which can be seen as oneparticular
multifractal model,predicts
alogarithmic dependence
1(&
in«)~)
=
(
in ~On the other
hand,
the variational model [71gives
a power lawdependence
(
(b In «)~)
cc r~ POur message is that
discriminating
between these modelsby looking
at the behaviour of((&
In «)~)
versus r is much easier thanby determining
theshape
ofG,
to which the3u is
poorly
sensitive,We demonstrate this
point
in sections 2 and 3by constructing
an artificial&u
using equation (I)
where G isdeliberately
different from alog
normal distribution, In section 3 we use a binomial distributionrecently proposed
ascorresponding
to the random beta model, a multifractal type one, In each case we show that the behaviour of A~ well reflects that of(
In « )~)
In section 4 we address the
question
of theshape
of G. We discuss the characteristic of anexperiment
which couldadequately
discriminate between differentproposed shapes.
We show that alarge
number ofsignificant
measurements are necessary, of the order of 10?points.
2. A first
example.
As discussed in the introduction we now construct an artificial
pdf using equation ii)
P~~(x)
=
~~
~~~
i~~
exp~~~
(l + b
In~
«
~
exp
~~
~
(3)
(2
ar « 2 A 2«
where x stands for ~~ and « for
I
andA~ is the normalisation factor. We denote it
«~ «~
N° 2 MEASURING INTERMITTENCY IN TURBULENCE 217
P~~
because we shall treat it as anexperimental
distribution and search for the parameter A which makes the « theoretical » distribution~~~~~
(2
A$
~~~)~A~
~~P
~
~~~
having
ashape
the closestpossible
to theP~~
one. We then compare A ~ with :n ~r
~~~
~~lAe (in
« )2 ~~~in2
fi
« ~~ ~ ~ ~~~ "~
d in « ~~~
We have thus first to
adequately
determine the difference inshape
betweenP~~
and~ih
We
begin by normalizing
them to the same meansquared
value,namely1, Pe(X')
" "e
Pex("eX~) (6)
p,(x<)
= «~
p~(«~x<) (7)
with
+JJ
+JJx'~ P~(x')d~<'=
j(a~x')~P~~(a~x') d(a~,<')
=
-« £Ye
~-«
and
+~J
+~X~~
~t(X~)
i~~~ j
(£Yt~')~ ~th(£YtX~) d(£YtX~)
~ l-oJ £Yt -~
In a real
experiment,
theby accumulating
the events inhistograms
where x' has agiven
valuexl
within A. Letn~ be the number of events such as
X,' ~,t~
~ X,' + d.
If N is the total number of measurements, P
(x,')
= 2A
The difference between two models for the
compared
to the difference weexpect between two realisations of the same
experiment,
which isgiven by
(n~
h,)~
= h~and thus
(P (x,') > (x;'))~
~
~
In a first
approach,
a measure of the difference betweenP~~
and P~~ could thus be theintegral
~+"
(Pe(x') P,(x'))~
~,
-m
Pt(x')
P~(P
~) is the normalised version ofP~,(P~~).
Seeequations
(6,7).
However such anexpression
could not converge ifP~
andP~
has very differentasymptotic
behaviour when x'~ ± oo
(see
Sect.
4).
We thusprefer
to define~
+°~
~
(Pe(x'l-Pt(x'l)~
~, ~
~
_~
P~(.<')+P~(x')
Let us now compare
P~~ given by equation (3)
withA~
=
0.49 to P
~~
(Eq. (4)).
Theintegrals giving P~(x')
andP,(x')
have been calculatedby
thetrapezoidal
rule with an absoluteprecision
of 10-~°((3
In«)~)
has been evaluated within 10-5 of relativeprecision
andX~(A
~) was calculated within 10-2 of relativeprecision [12].
The minimum of
X~(A
~) was estimatedby
the method of Brent[121.
Three A~ values are chosen in theneighbourhood
of theexpected optimum
: A),
Aj,
and aparabola
is fitted withthe three
corresponding X~' X~(A)),
The minimum of thisparabola gives Al,
The process isthen iterated with
Al
and the two closest of thepreceding Al,
It isstopped
when two successive minima differby
less than 10-3 in absoluteprecision,
6 0" ~
x2
4
0'~
2 0" ~
b
0 0.2 0.4 0.6 0.8
Fig.
I. The difference X~ versus the deformation parameter b.In
figure
I we haveplotted
the smallestX~ (Eq, (8))
versus b which measures the difference between the«
experimental
» G distribution and alog
normal distribution,The
X~
values arerising
ratherrapidly
with b, Thenthey
saturate at arelatively
small b value. This can be understood. When b islarge
the width of the distribution G becomes very small. G is then similar to a Diracdistribution,
whatever itsshape
and thecorresponding pdf
goes to a Gaussian
shape (see Eqs,
(Ii and(4)),
ThusX~
should go to zero whenb goes to
infinity,
For b= loo,
X~
isequal
to 1,7 x10~~.
The small values of
X~
are coherent with the small difference observed betweenA~ and
(
(3 In «)~) (Eq. (5)),
This is shown infigure
2, The decrease in A~corresponds
to theabove mentioned
sharpening
of G when b grows,Looking
at differentA~
values(A~
=
0.49,
0.7 andI)
we have been able to summarize theX~
behaviourby
anapproximate scaling (Fig. 3).
Within a few percent,X~/A~
behaves as a universal function of bA. Thisscaling
isprobably
fortuitous and not even true in theneighbourhood
of b=
0. However the values we have
spanned
are theexperimentally interesting
ones[7, 1II
and thisapproximate scaling
couldhelp
inestimating
an effective b value from the measuredX~
and A 2 values.The conclusion of this section is thus clear : the parameter ~ obtained from the
experimental
method we discuss
always gives
agood
estimate of((31n«)~)
even if the realG function is
significantly
different from alog-normal
distribution. As this result has beenN° 2 MEASURING INTERMITTENCY IN TURBULENCE 219
o.5
,-
0.45
0. 4
0.35
0 3
<ln20~
0.25
~2 0.2
0 0.2 0.4 0,6 O-S I
b Fig. ?. Comparison between and
((6
In rr)~).
0 0 0 6 , ~ -. .i T --1 .-~-r i~
xz/~4
0.004
0.002
b~
0
0 0.1 0.2 0.3 0.4 0.5
Fig. 3. The appro~imate sc311ng ~ummanzing the X~ behaviour 1,e13us and b.
obtained
through
a verypeculiar
and ad hoc class of G function. we shallstrengthen
the argumentby using
a more credible class of function,recently propo~ed.
This will inparticular
introduce the effect of the asymmetry of G iersiis In r,, Another conclusion is that the value of
X~
if;tati,tically ~ignificant,
couldhelp
inrevealing
a difference between the real G and alog-normal
distribution. This lastpoint
will bedeveloped
in the fourth section, 3. The binomial distribution.Recently
Benzi et al.[81
derived a formula for thevelocity
differences within the framework of the random beta model[131.
Their result for theP~(&v
cc~ ~j
~ a "~(
l a)~
B~~~~ii
"~exp[- CB~
~'~ii ~'~(3v
)~(9)
~~
K~
where
f,,
=
'
=
2~" C
=
(2 Vi j~' Vo being
the characteristicvelocity
difference atlarge
L
~cale. a
=
7/8 and B
=
1/2 are chosen such i to
reproduce
the(~
values obtainedby fitting
the
experimental p-order
moment((3uy)
with a power law in I((&CY)
CC i~~I<fi ~'ii Iii i'ili'i'jt it -T ' lL3W'Rh1>»~ '
With the notation of our
equation (1)
we have" "
V0 ~i~ B~~~
" "n
B~~~ (lo)
The distribution G is
given by
G(I
cc
( [~j a~~~(I
-a)~B~3(In ~i +~lnB), (11)
~n K=0
~ ~n 3
Thus
~3
In "
~)
=
~~~
j~ ((3K)~)
~n 3
=
n(
~~~)~ "~~
"~~(12)
3 (~x +
B(1
a))~
Following
the same process as in section 2 weidentify
thepdf given by equation (9)
withP~~(3u).
We then normalize itthrough equation (6),
and we compare this normalizedpdf P~(x')
with the normalized«
log
normal» oneP~, searching
for the minimum ofX~. This minimum value and the
corresponding
A~ value aregiven
infigures
4 and 5, for alarge
range of values of n,~ , ~. 4 X2
4
10"~
2
0'~
n
0 20 40 60 80 100
Fig, 4, The difference X~ versus n for the binomial distribution,
0 3
~~ ~,
,
;
;
~" <ln2~
0 .2 '
o. i
n
0
0 20 40 60 80 100
Fig.
5.Comparison
between A~ and((6
In«)~)
for the binomial distribution.N° 2 MEASURING INTERMITTENCY IN TURBULENCE 221
The difference between A~ and
( (3
In «)~),
and the minimum values ofX~
arelarger
here than in theprevious
case. This isprobably
due to thelarge dissymmetry
in In « of the« binomial » distribution,
This
dissymmetry
appears whencomparing
atypical
« binomial »
«
log
normal » onegiving
the lowestX~
(seeFig. 6).
Our choice ofX~ (Eq. (8)) clearly
emphasizes
the center of thedistribution,
where theexperimental precision
is the best one, Thediscrepancy
appears in thewings.
lo(~P
3
5 P
/ P
(n=15)
',,' b~
7
-lo .5 0 5 lo
x
Fig,
6.Comparison
between the « binomial» pdf (n = 15) (full line) and the best fitted « log normal »
one (dashed line), The experimental points are given to
figure
atypical uncertainty.
Here the number ofindependent
measurements is 2 x 10~. A significantcomparison
needs to take account of the skewnessthrough
the asymmetry of Pi (Eq. (I)). See reference [71.We note however that the difference between
A~
and((3
In«)~)
remains within 15 fl.More
important,
A~ appears asnearly
linear in n in the present range, which is asignificant
characteristic of( (3
In «)~)
in this type of models. The behaviour of A~ thus well reflects the behaviour of((3
In«)~),
even in this extreme case.In
particular
the recent evidence[7, 121
for a power law behaviour versus r forA~ shows that
((3
In«)~)
presents a similar behaviour. In the cascade models like therandom-beta one discussed in this
section,
the number of steps n should not be taken as linear in In r butproportional
tor~P
4. The statistical error.
The minimum
x~
values obtained in the twoprevious
sections wererelatively
small. So thequestion
raises of theexperimental
accuracy needed todistinguish
between alog
normal distribution of « and a deformed one for instance, In this section we estimate the minimumx~
to beexpected
if one compares anexperimental pdf (with
N measurementpoints)
with itsexact
shape, asymptotically
obtained with an infinite number of measurements.Let us call J the size of a class, like in section 2. We have to estimate
X~=2 ( ~~~~~'~ ~'~~'~~~A= ( E, (13)
,=-«
Pe(x,)+Pt(.«)
,=_«
where
P~
i~ now theexperimental pdf
andPj
itsasymptotic
limit when N~ oo. The number of
points
in the class is n, N~P~(,i,
) for ourparticular experiment
and it~ averageexpectation
value is
h,
= NAPj(.i,).
Twolimiting
ca~es have to be considered.2
k,
~~ ~~'~ ~' ~~~~
~~~~'~
~~~~"'~
NJ
(F~(v, P~(
i,I
=
~ (ii~ fi,
I
=
~'
(14)N- J- N-
J~
as~uming
a Poisson distribution for ii,. We get2 (P
~(.t',
Pi (,ii
)~~ l
~~ ~ P
~
(.t,
+ P (-I,) Niii
h,
« I. If fi~ i~ zero we haveE,
k~=
2 ~P i-r, 2
If ii, is or
larger,
we canneglect P~(,i,)
andE,
n,=
2
AP~(,1,
=
2
Calling ar(fi,)
the Nprobability
to get n, we haveE, =2r(0)j+2zar(fi,)(=2(1+ 7r(01)(=/=4AP~(_r,).
(16) Note that ifx~
were defined with
only P~
(and notP~
+Pj)
in the denominator, the sum ofequation
(13) woulddiverge.
We shallinterpolate
between these twolimiting
case~taking E,
= 4
k,/N
for fi, w andE,
=
for
k,
m Let us call 2 i~ thelength
of the interval onwhich k~ m For a
large
N themajor
contribution to X~ comes fmm this interval and4
~
2,ij,
X~=- (17)
We can go a step further
assuming
aparticular shape
forP,.
Anexponential shape
has often been used to fit suchpdf.
Itgives
P,(.1)
=
~-exp(-
i-11
,'2)
, 2
(the
variance ofP~(,i)
is I). ~i~ isgiven by
~~
exp(-.ij,
,. 2)=
, 2 4
or .i~, =
~-
ln (2, 2 NJ1
,, 2 The contribution of j-v ~_<~ is
m
,/j
2 4
P,(~<)
d~< = 4 exp (- iii , 2)=
~
'i,
N° 2 MEASURING INTERMITTENCY IN TURBULENCE 223
Then
X~"~(l
+ In
(2,2NA)), (18)
The result
depends
on the size of the class as too small a classgives large
statistical errors in thecounting,
On the other hand toolarge
classes introduce errors inestimating
theintegrals,
We shall take J
= 1/30 as a
compromise,
that it classes 30 times smaller than the root meansquare of 3u,
From the
preceding
section we know that an order ofmagnitude
forX~
which allows some distinction between thepossible shapes
is lo ~ This ask for N mlo?
uncorrelated measure-ments. This is close to the
large~t samples
used yet[71.
For
distinguishing
between twoproposed shapes
the strategy is thus clear. First the numberof measurements must be sufficient for the
« statistical
»
X~ being
smaller than thatgiving
the difference between the twoshapes.
Second theX~
must be estimatedusing alternatively
the twoproposed shapes
as the« theoretical » one, The smaller X~ is the be~t and must be of order of the
« statistical
» value to be the
good
one.5. Conclusion.
The conclusion of this
study
is that theshape
of theshape.
The parameter ~ obtainedby using
alog
normalansatz for G is
always
a reasonable measure of((3
In jr)2),
Physically
this parameter is anobjective
measure of the evolution of the difference ofvelocity pdfs,
In all modelsreferring
to an energy cascade, as for instance the random beta model discussed here in ~ection 3,(
In «j~)
isproportional
to the number of cascade steps.Our
study
thusfully
validate~ anexperimental
way ofmeasuring
thisimportant
parameter,This method ha~
already
shown that((&
In(,)~)
does not behave as In r, as isgenerally
assumed, but more as a power law in ;.The second
question
addressed in this paper concerned theshape
of the distribution ofi, ; G. We have shown that a
significant
check of thisshape
asks for verylarge
statistical~ample,,
however not out of the presentexperimental possibilitie~.
Inparticular, distinguishing
between a binomial di~tribution as
proposed
in reference(8]
and alog
normal one asproposed
in[41
i~certainly
feasible.Acknowledgments.
Thi~ work has been
partly ,upported by
a DRET contract (N° 9?/105) and theRdgion
Rhbne-Alpe,,
References
II Kolmogorov A. N., Dr)i/ At-ad Nauk. ssSR 30 (19411 301 31 (1941) 338 and 32 (1941) 16.
[21 Monin A. S.~ Yaglom A, M., Statistical Fluid Mechanic~ (the MIT Press, 1975).
[3] Gagne Y., Thesis (19871 (Institut National Polytechnique de Grenoble, unpublished) Advances in Turbulence 3
(Springer Verlag
1991).[4j Castaing B., C-R At-ad. SC (Pa;is) 309 ~erie II (19891503.
(5] Kraichnan R., Pfi,vs Rei Lett. 65 j1990j 575 She Z. S., F/ilid Din. Research 8 j1991j 143.
[61 Andrews L. C.,
Phillips
R. L,,Shivamoggi
B. K,, Beck J. K., Joshi M. L., Phys. Fluids Al (1989j 999.[7]
Castaing
B., Gagne Y.,Hopfinger
E,,Physica
D 46(1990j
177.[81 Benzi R,, Biferale L., Paladin G., Vulpiani A.,
Vergassola
M., Phys. Re». Lett. 67 (1991) 2299.[91 Eggers J., Grossmann S., Phys. Rev. A 45 (1992) 2360,
[10]
Castaing
B,, Gagne Y., Turbulence inSpatially
ExtendedSystems,
R. Benzi Ed, jLes Houches 19931.[tl]
Castaing
B,, Gagne Y., Marchand M., to appear in Physica D.[12] Press W. H.,
Flannery
B. P,,Teukolsky
S. A.,Vetterling
W, T., NumericalRecipes (Cambridge University
Press, 1986),[13j Benzi R,, Paladin G,, Parisi G., Vulpiani A., J, Phys. A17 (1984) 3521,