You are on page 1of 25

Statistical Theory of Turbulence Author(s): G. I. Taylor Source: Proceedings of the Royal Society of London.

Series A, Mathematical and Physical Sciences, Vol. 151, No. 873 (Sep. 2, 1935), pp. 421-444 Published by: The Royal Society Stable URL: http://www.jstor.org/stable/96557 . Accessed: 04/12/2013 11:20
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The Royal Society is collaborating with JSTOR to digitize, preserve and extend access to Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences.

http://www.jstor.org

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

421

Statistical Theory of Turbulence


By G. I.
TAYLOR,

F.R.S.

(ReceivedJuly 4, 1935)
INTRODUCTIONAND SUMMARY OF PARTS I-IV

Sincethe time of OsborneReynoldsit has been knownthat turbulence to the coefficient producesvirtualmean stresseswhich are proportional of correlationbetweenthe componentsof turbulentvelocity at a fixed directions. The significanceof correlation point in two perpendicular betweenthe velocityof a particleat one time and that of the sameparticle at a later time, or between simultaneous velocitiesat two fixed points was discussedin 1921by the presentwriterin a theoryof " Diffusionby in the technique ContinuousMovements." The recent improvements have made it possibleactuallyto measuresome of measuring turbulence of the quantities envisagedin the theoryand thus to verifysome of the then put forward. relationships The theory has also been developedin severaldirectionswhich were not originallycontemplated. The theory, as originally put forward, when the motion provideda methodfor definingthe scale of turbulence is definedin the Lagrangian manner,and showedhow this scaleis related to diffusion. It is now shown that it can be applied either to the or to the Eulerianconceptionsof fluidflow. Lagrangian is producedin an air streamwith a definitescale by Whereturbulence means of a honeycombor regularscreen,eitherconceptioncan be used to definea length which is related to certain measurable propertiesof flow and is a definitefractionof the mesh-length,M, of the turbulenceproducingscreen. The Lagrangian conceptionleads to a length 11,which is analogousto the "Mischungsweg" of Prandtl. Experimentson diffusion behind screens,Part IV, show that 11= 0-1 M. The Eulerianconceptionleads to a definitelength 12which might be regarded as the averagesize of an with a hot wire, Part II, show that eddy. Correlationmeasurements 12 is about equal to 0-2 M. The theoryappliedin the Eulerian manner to thesecorrelation measurements also containsimplicitlya definitionof x, "the averagesize of the smallesteddies,"which are responsiblefor the dissipationof energyby viscosity.

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

422 It is provedthat

G. I. Taylor

W-== 15(u2/X2,), whereu2is the mean squarevariationin one componentof velocityand W is therateof dissipation of energy. This relationship is verified experimentally (PartII). The relationship betweenXand M is discussedand it is predicted that in an air streammovingwith velocityU will die down so that turbulence
IVu2

is determined that the scaleof turbulence provided by the mesh-length M where A is a universalconstant and B depends on the choice of the origin taken for x (the down-stream co-ordinate);u is a componentof is comparedwithresults turbulent velocity. This theoretical relationship in tunnels out in wind carried of experiments Englandand in America. The theory is applied in Part III, to determinethe distributionof dissipation acrossthe sectionof a parallelwall channel(two-dimensional pipe) and it is shownthat in the regionnearthe walls turbulent energyis it is than In the centralregion the producedmore rapidly dissipated. reverseis the case. In PartIV the resultsof diffusionexperiments madein Americaand at are discussedand it is shown that a the National PhysicalLaboratory can give Vv2, 11,and a length X, complete set of such measurements as a measureof the " smallestsize of eddy" in which may be regarded the Lagrangian theLagrangian system. ?n is connected, through equations of motion,with the averagespatialrate of changein pressure, namely

V(j_)2
by the formula,
:\/E@i ) A2 p -,;*.

that Xa is a Finallyit is shownthat the theoryleads to the prediction which exists at constant multiple of X. The only set of experiments 2? approximately. presentgives ?, that the "Reynolds All the above resultsare subjectto the restriction Number of Turbulence," namely l V/u2/v, is greaterthan some number which must be determined by experiment.

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theory of Turbulence PART I

423

At an early stage in the development of the theory of turbulence the idea arosethat turbulent motionconsistsof eddiesof more or less definite range of sizes. This conception combined with the already existing ideas of the KineticTheoryof Gases led Prandtland me independently " andis to introducethe lengthI whichis often calleda " Mischungsweg analogousto the " mean free path " of the KineticTheory. The length I could only be definedin relation to the definitebut quite erroneous conceptionthat lumps of air behavelike moleculesof a gas, preserving their identitytill some definitepoint in their path, when they mix with and attain the same velocity and other propertiesas their surroundings the mean value of the corresponding propertyin the neighbourhood. Such a conceptionmust evidentlybe regardedas a very rough representationof the truestateof affairs. If we considera numberof particles or small volumesof fluid startingfrom some definitelevel and carrying, say, heat in a directiontransverse to the mean streamlines, theiraverage distance from the level at which they started will go on increasing " in relation indefinitely so that we can only considera " Mischungsweg time of flight duringwhich we must considerthat the to some arbitrary particlespreservetheir individualpropertiesdistinctfrom those of their surroundings. Clearly this is an arbitraryconception and if pursued logicallyprobablyleads to a definitelywrong result. The only way in whicha smallvolumecan lose its heat is by conductivity to its surroundings. A decreasein molecularconductivity would thereforelead to an increasingtime during which the small volume would retain its heat distinct from its surroundingsand consequentlya decrease in conductivitywould necessarily lead to an increasein the " Mischungsweg." In all theorieswhich make use of I it is assumedthat 1 dependsonly on the dynamical conditionsof the fluid and is nearlyindependent of such physicalconstantsas thermalconductivity. " theories the length I is conIn all applicationsof " Mischungsweg sideredonly in relation to further,more or less arbitrary, assumptions the effectof turbulence concerning on the mean motion or of the mean motion on turbulence. It appearsas a fictitious length, the existence of whichis detected only by observationsof the distributionof mean etc. velocity, temperature, The difficultyof defininga " Mischungsweg," or scale of turbulence, without recourseto some definitehypotheticalphysical process which bears no relation to reality does not arise in such applications. The

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

424

G. 1. Taylor

difficulty,however, still exists and it led me, some years ago, to introduce the idea* that the scale of turbulence and its statistical properties in general can be given an exact interpretationby consideringthe correlation between the velocities at various points of the field at one instant of time or between the velocity of a particle at one instant of time and that of the same particle at some definite time, i, later. Some general relations applicable to either of these two aspects of the turbulent field were discussed, and the application of the definitions used in the second of them to diffusion in one dimension was worked out in detail. In this application of the theory the particles are conceived to move irregularly but with continuous velocity, v and v2is supposed to be independentof time. The diffusion of particles starting from a point (y- 0) is shown to depend on the correlation Rt between the velocity of a particle at any instant and that of the same particle after an interval of time i. In continuous turbulent movements Rt must be a function of i such that Rt =1 when i 0 and Rt -> 0 when i is large. If Y2 is the mean square of the distance through which the particles have diffused in time t it was proved that
(Y2)= dt

(1)

If the time of diffusion is small so that Rt has not departed appreciably from its initial value I 0, (1) becomes
( d(Y2)

=v2

so that

v't, VY2 (2) where v'= If the diffusion is taking place in a stream of air moving with velocity U and if the spread is observed at a small distance x down-streamfrom the source t x/U so that
_

(3)

UJ.

If the irregularmotion is of such a character that it is possible to define a time T such that Re = 0 for all values of i greaterthan T, so that there is no correlation between the velocities of a particle at the beginning and end of the time interval T, then Yv - v2
*

R Rt

(4)

'Proc. Lond. Math. Soc.,' vol. 20, p. 196 (1921).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theory of Turbulence

425

Yv is therefore constantfor all valuesof t > T in spiteof the fact that the value of Y2?is continually increasing and v2 is constant. Under thesecircumstances it is possibleto definea length11,such that _ T d = dt Vv2= d-v2JR 11 (5) It will be seen from (5) that the length11,definedas
11 v2

Rf

dt

(6)

bears the same relationshipto diffusionby turbulentmotion that the mean free path does to moleculardiffusion. In this sense it is very similar to the " Mischungsweg," 1, but with this importantdifference that the questionof mixturedoes not arisein definingit. As is pointedout above, theorieswhichdependessentially on the idea of mixtureby subdivisionand ultimatemoleculardiffusionlead to the " will dependvery greatlyon the expectationthat the " Mischungsweg moleculardiffusivepower of the fluid. In the theory of diffusionby continuousmovementsthe length 11bearsno relationto any processof mixture,indeed it is equally valid if mixturenever takes place. The diffusion would be to preventthe fluidfrom becoming effectof molecular
ever increasingly "spotty," i.e., it would tend to prevent a continual

increasein the deviationsof the measurable of the fluid froni properties their mean value in the neighbourhood. Mixturehas no effect in this theory on the diffusivepower of turbulentmotion.

CORRELATION IN THE TURBULENT FIELD WHEN DESCRIBEDIN THE EULERIAN MANNER

In a loose way it has been thoughtthat the "Mischungsweg"length I is relatedto, and even may be taken as a measureof the averagesize of the larger eddies in turbulentflow. It will be noticed that in the " theories,and also in the theory of diffusion original " Mischungsweg is definedin a Lagrangian by continuous movements, everything manner, i.e., by followingthe paths of particles. Whena field of eddyingflow is considered as an entityin itself, apartfrom its effectas a diffusiveagent, it is moreusualto thinkin termsof the Eulerian of fluidflow, conception i.e., a field of streamlines conceivedto exist in space at one instant of time. Any ideas we may have about " the size of an eddy" are likely in the Euleriansystem. For this reasonit would not' to be formulated
VOL. CLI.-A. 2F

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

426

G. 1. Taylor

be possible to connect directlythe size of an eddy, even if it could be with the value of I or of 11as definedby (6) in the accurately defined, Lagrangiansystem. At the same time it seems to be a matter of considerabletheoreticalinterestto investigatethe statisticalproperties with manner, in the Eulerian of a fieldof turbulent flow when described in some definite waythe a view to defininga lengthwhichmay represent "size of an eddy." The correlationtheory developedin my paper, " Diffusion by Conin this case and may be used is equallyapplicable tinuousMovements," to formulateanotherdefinitionof the scale of turbulence. It is clear of an eddya high degreeof we may meanby the diameter that whatever mustexist betweenthe velocitiesat two pointswhichare close correlation togetherwhen comparedwith this diameter. On the other hand, the correlationis likely to be small between the velocity at two points we imaginethat the apart. If, therefore, situatedmany eddy diameters correlation R. betweenthe values of u at two points distanty apart in for variousvalues has been determined the directionof the y co-ordinate of y we may plot a curveof R, againsty, and this curvewill represent, of u along the y axis. point of view, the distribution from the statistical If R. fallsto zero at, say,y Y, then a length'2 can be definedsuchthat
12

jR dy =0Ryddy.
o o

(7)

as the analoguein the Euleriansystem This length 12may be regarded system. It may be taken as a of 11,which is definedin the Lagrangian possibledefinitionof the " averagesize of the eddies."
METHODS FOR MEASURING 11 AND 12 EXPERIMENTAL

hot wire is capableof beingused to measureseveral The compensated in any statistical considered whichare necessarily of the quantities theory of turbulence. (1) u2can be measuredby means of a hot wire anemometer. If the are passedthrougha wire the heat producedcan disturbances amplified which will cause a deflection give rise to a currentin a thermojunction, to u2. in a galvanometer proportional to a (2) If two hot wires are set up at a distancey apart transverse streamof air and the currentsproducedby variationsin U at the two the points are sent throughthe two coils of an electricdynamometer, to u0U' will be proportional where u0 and u, are the resultingdeflection

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theory of Turbulence

-427

velocities at the two points. In this way Rz,u0ujzu can be measured. By repeating these measurements for a number of different distances of separation y between the two hot wires, Rv can be determined for all values;of y and hence by integration 12can be found. The (R., y) curve has already been obtained in certain cases by Messrs. Simmons and Salter at the National Physical Laboratory by this method (see fig. 1 of Part II). Another method is to arrange two equal hot wires on two arms of a Wheatstone bridge thus measuring (Ul - uO)2. If H2 and u12 are measured independently at the two stations, uoul can be found from the relationship (8) uo2+ 17- (uO-u)22uOu. Yet another method due to Prandtl* is to pass the currents from the two hot wires through coils which cause deflections of a spot of light in two directions at right angles to one another. If the two hot wires are identical and so close together that the correlation is nearly 1 0, the spot of light moves over a very elongated elliptic area, the long axis of which is at 450, to the deflections caused by either of the wires in the absence of disturbances from the other. By measuring the ratio of the principal axes of the elliptical blackened areas produced on a photographic plate by the moving spot of light during a prolonged exposure, it is possible to calculate RV. This method is specially suitable for measurements when the correlation is very high, i.e., 1 - RV is small. It is not so suitable for small correlations as the electric dynamometer ,method. Correlation measurements made in this way are shown in fig. 1 of Part III of this paper. (3) By introducing heat at a concentrated source or a line source in an air stream and measuring the spreading of the heat to leeward of thesource it should be possible to measure the quantity - Y which occurs in (1) and hence to find
rt~~~~~~~d

Rt dt for various values of t. If this reaches

-a constant value at some distance down-stream then 11 can be found. This method was suggested in my paper on " Diffusion by Continuous Movements." Up to the present, however, the theory has only been applied to cases like that Qf diffusion in the atmospheret where there is no a priorireason to suDDosethat any definite scale of turbulence can be
* Prandtl and Reichardt, " Einfluss von Wiirmeschichtung auf die Eigenschaften einer TurbulenterStromung." Deutsche Forschung, p. 110 (1934).
+ gV,ttrn 'Prnir Dnir Znc:
'

unl

11

1zA

(101IN

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

428

G. I. Taylor

defined. Indeed,Mr. 0. G. Suttonhas shownthat the best representation RSoc -n of diffusion in the air nearthe groundis obtainedby assuming so that Rt does not vanish howevergreat i may be. In fact j'R4 dt
o

in t so that 11, definedas in equation increased continuously with increase (6), would have no definitevalue. The turbulence whichoccursin wind tunnelsis producedor controlled size. In a windtunnel,therefore, by a honeycomb with cells of a definite to be of mightbe expected thereis an a priorireasonwhy the turbulence some definitescale. In fact, it might be expectedthat both 11and 12 would be some definitefractionof the mesh of the cells. Under these the diffusionequations(1) and (6) reduceto circumstances
dty2
= llv* (9)

x of thepointsat whichmeasureis validwhenthe distance Thisexpression ments of Y2 are made from the point or line source of diffusionis so x/U and U is the mean speed of the air great that Rt 0 where i stream.
APPLICATION OF DIFFUSION EQUATION WHEN TURBULENCE IS DECAYING

is not In the air streambehind a grid or honeycombthe turbulence constant. It decreases as the distance down-streamincreases. The precedingtheory cannot then be'applied without furtherinvestigation. as a functionof t the diffusionequationis If vl is considered
Ct
-t ~~~~d

21

s-di vt Jt_

(10)

di and t(Y) is the rate of increase in Y2 at time t after source. of the diffusionfrom a concentrated the beginning the velocityat time between of correlation the coefficient is If tR,_e t and that at time t-i, (10) may be written
for Y
vt-t i_-y2=t0
dt vt- t(tRt-t)dis

11

2 ~~~~ A/v where v'%, v't are written for V/v2t, motion is constantwith When the averageconditionof the turbulent respectto time tRt-t is the same as tRt+ or Rt and is a function of i

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theoryof Turbulence

429

only, so that (11) is identical with (1). When v' is not constant, it is not possible to proceed beyond (11), but the existing experimental evidence seems to show that turbulent diffusion is proportional to the speed, so that if matter from a concentrated source is diffused over an area downstream from the source, an increase in the speed of the whole system (i.e., proportional increases in turbulent and mean speed) leaves the distribution of matter in space unchanged (though the absolute concentration is reduced). The condition that this may be so is that tR,t is a function of n only where d- = v'd~ d (12) (v'/U) dx and x Ut is the distance down-stream from the source. The equation which represents the lateral spread of matter or heat from a concentrated source is therefore 1U dd V2) where
nz

HR,d-

(13) (14)

fV' dx,

and R, is the correlation between the velocities of a particle at times t, and t2 when
=,
B =

v' dt. If R. falls to zero at a finite value of -, say


0

and remains zero for all greater values of R,

is finite. XRd-m

If 11be written for


o

R,dn then (11) becomes


21Vx

1U d

(Y2)

-15

11.

(15)

This is the same expression as that found for turbulence which is not decaying.* It is worth noticing that (13) may be expressedin the form
i d (y2) w~(Y2)=
- inz e d-m.

x R~P,d

(16)

When x is small so that R= becomes


-

1 over the range from 0 to x, (16)-

(17)

The integral of (17) is


y2= * 2

or

Vy2

(18)

See equations (1) and (6).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

430

G. 1. Taylor

Whenthe turbulence is constantn xv'/U so that (18) reducesto the previous expression(3) for the spread of matter near a concentrated source. If the turbulence is not constantandif Y2 and v'/U aremeasured
at a number of values of x, then both N and i v' dx (Y2) can be found.
U
d-

Thus R,,dNcan be plottedagainst N and R),can be found graphically fromthis experimental curve.
MICRO-TURBULENCE AND DISSIPATION OF ENERGY

Besides the motions which are chiefly responsiblefor the diffusive the wholefieldmaybe in a stateof micro-turbulence, powerof turbulence i.e., there may exist very small-scaleeddies which, though they play a verysmallpartin diffusion, yet may be the principal agentsin the dissipation of energy. They may also be the principalcausesof the effects of turbulenceon the boundarylayer in wind tunnel work because the absolutemagnitude of the spaceratesof changein pressure may depend on them.
OF ENERGY DISSIPATION

The rate of dissipation of energyin a fluidat any instantdependsonly on the viscosity, t, and on the instantaneous distribution of velocity. of the essentialstatisticalproperties If, therefore,the representation of the velocityfieldcan be expressed the curve and similar by correlation R, curvesit must be possibleto deducefrom themthe rate of dissipation of energy. This would in general involvea complicated analysis,but the if the field of turbulent problemcan be much simplified flow is assumed to be isotropic.
ISOTROPIC TURBULENCE

In isotropicturbulence the average valueof anyfunctionof the velocity definedin relationto a given set of axes, is unaltered if the components, axes of referenceare rotated in any manner. That there is a strong motion has long been known. It has tendencyto isotropyin turbulent been shownby Fage and Townend,* for instance, that the average values of the three componentsof velocity in the centralregion of a pipe of squaresection are nearlyequal to one another. In the atmosphere the samephenomenon has been observed;though,as mightbe expected,the
* Townend, ' Proc. Roy. Soc.,' A, vol. 145 (1934) (see fig. 15, p. 203).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theoryof Turbulence

431

vertical components are smaller near the ground than the horizontal ones, this inequality decreases with height above the ground.* The assumption of isotropy immediately introduces many simplifications both into the statistical representation of turbulence and into the expression for the mean rate of dissipation of energy. The general expression for the rate of dissipation is
W

\\2 (au {2axax

av7,\2

\aa

aw

2+

27av

a9U,2

TX

ayi2 az axw
(19)

y + az

Making the assumption that the turbulence is statistically isotropic, the relations a) 2 at 2 aw 2 and

(auj2
and

au2

a2

/ aw av ay az 2

2=

'aw2

aw2

(20)

av au ax ay xxfa

auaw az ax

are immediately obtained so that p-6 (ax) +

aua) ay

av au

axay

(21)

Equation (21) contains three types of term. It will now be shown that these are all related to one another so that if the value of one is known the other two are known. That relationships can be found between the mean values of squares and products of au/ax, au/ay, av/ax, .., etc., is obvious. The simplest relationship is obtained as follows. The condition of continuity is

au+av+
so that

aaw

z a\ x U)+ (-v) ay, + ( -Wp


(aU
2

- 2 (ax ay + av aw+ a ay aZ az Tx)

(22

The conditions of statistical isotropy therefore lead to the relationship

2 au av

(23)

* Taylor, 'Q. J. R. Met. Soc.,' vol. 53, p. 210 (1927).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

432 or

G. 1. Taylor au av

axa=

(24)

In otherwordsthereis a definitecorrelation betweenau/ax coefficient and av/ay equal to 2.


MEAN VALUEOF GENERAL QUADRATIC FUNCTION OF au/aX, aViaX,
aU/ay, ..., ETC.

Considerthe most generalpossibleexpression for the mean value of any quadratic functionof the nine quantities

au av aw au av aw au av aw ay TX TX FXSTyS -y az z I TzX
In generalthere are 36 possiblecombinations of 9 things taken 2 at a time. Thus the most generalquadratic contains45 terms, expression namelythe 9 squaresof the quantities concerned andthe 36 combinations of 2. When the motion is statistically istotropicthe 45 terms fall into 10 groups,each of which contains 3 or 6 means which are equal to one another; for example,one group containing3 equal terms consists of
(aU)2 (\2

)2

and

aW 2

Anothercontaining6 equal termsconsistsof

axa: axayI ayaz' ayTax ay aza x


.

auav auaw avau avaw awau awav

The 10 possible independent mean values will be denotedby a,, a2,


a1o

according to the scheme laid out in Table I where the top row of

the tablegives the type termand all othertermsof the sametype can be obtainedby permuting the elementsof the type term. symmetrically The symbolwhich represents the mean value of any term of a type is given in the second row and the numberof independent termsin each groupis givenin the last row. In termsof thesesymbols(21) becomes W/
-

6a1+ 6a3+ 6a8.

(25)

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theoryof Turbulence

433

ZI-Ga 1'<, Ct

Hz
'!Ic c I( 1 U o

z
E

(~~~~~~1C ~ ~ ~ ' (Ict1 f e

I ,

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

434

G. 1. Taylor

I now propose to prove that the 10 values, a,, a2, ... al0, are interconnected, so that if the value of any one of them, which is not zero, is, known all the rest are known. For this purpose it is necessary to prove 9 linear relationships. One such relationship has already been proved (see equation (23)). Expressed in the symbols of Table I (23) may be, written
a, - - 2a6. (26),

Further relationships may be obtained as follows. Take any one of the 45 possible terms in the most general quadratic expression involving the 9 partial differentialsof (25). Transformu, v, w, x, y, z, by rotation of the axes to u' v', w', x', y', z'. The transformedexpression will still be, quadratic but will contain terms of other types than the original one. When the mean values of the terms in the transformed expression are considered it is a necessary consequence of the definition of isotropy that the value of each is equal to that of the type term in the group in which they are classed. A simple transformation is obtained by rotating the
axes through
450

about the axis of z so that x + y a2 = u+ u

2 IV/x'

V/2y'=-x+y z = z
Hence Au
aX

v/2v
w

U +v
w

(27>

1 '@uu _
aX,

.au'
au, aV X.,Z'\ aX' ayt~

v'
t
-j

_
a
-

'U ax -V X
2 au'

l axu' av a'x + -a aU

ey
y
1
(

-~X

- --

aU' au,

at

rt
a`j\
-

av y
aw

aut

aXt

ax

a/ x'
awf

ay
A u

T/

x'
wl

_
aw

-Fy7

aw I

az
_

az!+ az(28) + T-2;


By qa

Take forexamle( ;83a) = l. w' By squarinthetransforedexpressio Ajw


Take,forexp

au

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theoryof Turbulence

435

for au taking the mean value and substitutingthe symbol for the typetern fromTableI it will be foundthat corresponding

aui aavi2 =aut2 au,)2 a?x)/= al = 4; tAx'} a + ( aX )


=

avi
a+

ax'

I
(29)

Similarly
(7;)2=

(al+a3-a5-a

+a,+a8-a2-a5).

+ a,. =1 (aja = a3 =

a5a2, +

+ a6 + a8+ a2+ a5),

(30) (31)

( y)

(al + a3 -a5 + a2-a6-a8+a2-a5),

(aV) = a3-i

(a + a3 + a5-a2-a6-a.a8-a2+

a5).

(32)

Fromtheseequationsit will be found that


a2 -a5
O

(33)

and
a, -a
-

a,

Ia

C(

the type termscorrecan be derivedby transforming No furtherrelations to termsinvolvingw or z spondingwith a2, a5, a6, or a8. Proceeding

au=-a10
aya-a9-=
au av
2x8z=a7=

u _(a3- V

a3) _0, -ia7-a4+ a9-a2)-

(35) 0, (36)
(37)

auav

1 1

(a2-a9+

a4-a7

2,(a-

-4

,+7,-a4-

a9+a43*

hence since
a2

a9 =0

a7 (2--1)+a4=O
a

(38)

a =a4=
and hence

(a2H aID;-a4a4 (2-1)

a7

+ a4-

ag

a2),

combining (38) with (39)

+ a7-0

(39) (40)

a4= a7= O.

types up the resultsso far obtained6 of the 10 independent Summing

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

436

G. I. Taylor

of mean are zero, namely, a2, a4, a5,a7, a9, al0 and there are two independent relationships between the remaining 4 means, namely, a, =-2a6 and a-a3a6-a8 ?O.

Of these the first depends on incompressibility and isotropy. The second depends only on.isotropy. One further relationship can be obtained by volume integration of the general dissipation expression (19). This integration is well known*: it is

srr JJJ\ -Yax)ya:

r:

2 au\4- 2v) v2+ aw\+ law av za + azy


ax
2z

(auaW\2 02 ) + U+ aW ?( + a' V+ U 2dx dydz

d yd

=
where

)dxdydz-

an(q2)dS

+ 2

u v w dS, (41)

i=(wlay)

-- (av/lz), etc.

and the integrals are taken over the cloud surface S and through its volume. If the closed surface is large compared with the scale of the turbulence the surface integrals are small compared with the volume integrals which may therefore be neglected. Taking the mean value of all the quantitiesin (41) and expressingthe result for isotropic turbulence in terms of the symbols of Table I, (41) becomes Hence

W/ ,

6a,

+ + 6a+
a, + 2a8

6a8-6a3-6a8.

(42)

0,

(43)

solving (26), (34), and (43) it will be seen that al = 1a3 -2a6=--2a8. Three obvious corollaries to this result may be noticed: (1) The correlation coefficientbetween Au/ax and av/ay is - 2 (2) The correlation coefficientbetween au/ay and avlax is-4. (3) When the mean value of any one of the four possible types of quadratic terms which are not zero is known all the rest are known, so that the mean value of any quadraticfunction of the space rates of change
* See, for instance,the chapteron viscosity in Lamb's " Hydrodynamics."

(44)

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theory of Turbulence

437

of velocity is also known. In particularthe dissipation may be expressed in terms of


W/u
-

The correct expression is


3a3 + 6a3 -1 5a3 = 75 4

6a1 + 6a3 + 6a8 -

STATISTICAL REPRESENTATION OF MICROTURBULENCE

The value of ()

is clearly related to the way in which the value of

R, falls off from its initial value I 0 as y increases from zero. I have proved,* in fact, that 2 L 1Y4 2U2 l 2 (au ( .\ 2 (46)
The curvature of the Rv curve at y

0 O is therefore a measure of (au

so that

2u Lt
-y

2R

(47)

?/-*O \y2(7

The significance of the expression (47) can best be appreciated by defining a length X such that Lt ( 1 (48)
1-2 2 iS then a measure of the radius of curvature of the R, curve at y - 0. If the curve is drawn on such a scale that its height is H (corresponding with R -1 at y 0 is ?X2/2H. 0) the radius of curvature at y Another interpretation of X may be found by describing the parabola which touches the Rv curve at the origin. This parabola will cut the axis R, = 0 at the point y = X. X may roughly be regarded as a measure of the diameters of the smallest eddies which are responsible for the

dissipation of energy. CONNECTION BETWEEN DISSIPATION OF ENERGYAND CORRELATION FUNCTION R., Combining (45) with (47) and (48), the dissipation is related to the correlation function R, by the equation

W =15

Lt l
y->.

,Rv

(49)

* ' Proc. Lond. Math. Soc.,' vol. 20, p. 205, equation (14), (1921).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

438
or

G. 1. Taylor
W -15pu
/X2.

(50)

Since u2 and R. can be measured directly by means of the hot wire technique referred to earlier, the relationship (49) can be verified if W can be measured by other means. The way in which this can be done and the comparison between this statistical theory and the results of observation will be discussed later. In the meantime it may be noticed that if the Reynolds's stresses in geometrically similar fields of flow are proportional to u2 or u'2,W is proportionalto u'3, so that X is proportional to (u')-, and since ? is proportional to the curvatureof the RVcurve at 0 we are led to the prediction that the curvature of the R, curve at y its summit, y - 0, will be proportional to 1/u'. In the limit for veryhigh values of u' the RI, curve may be expected to have a pointed top.
TEST IN WIND TUNNEL OF SUGGESTIONFOR EXPERIMENTAL RELATIONS PREDICTED CORRELATION

It has been shown how measurements of correlation between the readings of two hot wires at points close together in a transverse section

of a pipe or wind tunnelcan give the valueof (u


(>)2

'measureIf similar

ments could be made in a line parallel to the main stream, values of could be obtained in the same way. Equation (44) shows that
(u 2
(au)2

and referringto equation (47) which is equally true when x is substituted for y, it will be seen that for the correlation to fall a given amount from its coincidence value 1 .0 the separation of the two hot wires must be 2 times as great when one lies up- or down-stream from the other as it is when they lie across the stream. This is a definite new theoretical prediction which could be tested. If difficultyis found in working with one hot wire down-stream from the other, measurements might be made with the two wires mounted at a fixed distance r apart on a rotating holder, and the variation in the correlation R as the holder is rotated might be found. The correlation between the values of u observed at two points situated at a short distance, r, aparttina line making an angle 0 to the wind direction
is*r2X,u

1Sv~~~~~~

2U 1 - -

) 2(

(51)

* Compare equation (47).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theoryof Turbulence


wherex' - x cos 0 + y sin 0. Since

439

aU= cos 0 au+ sin 0 LU ax' Ax ay


(51) becomes,in the notationof Table I,
R=-

1-2

r2

2 (acos20

+ a3sin20

+ 2a2 cOS

sin 0). O

(52)

is isotropicthis is When the turbulence


1 -R=

= 2 -2

2 (2au, oS2 0 -V+sin2 0). (C 2u2 aYi ~~~~~~~~~2 0


(COS2

hence from the definitiontof X


..

1-

R=

sin2 0).

(53)

that 1 - R should vary in the ratio 2: 1 as the It appears,therefore, holder is rotated for the maximum to the position to maximumto correlation. .minaimum
X AND, SCALEOF TURBULENCE RELATIONSHIP BETWEEN DIMENSIONAL

It has been shownby v. Karmanthat if the surfacestressin a pipe is expressedin the form T pVx2 then
U0 U
-

vx

f(a) ~a!

(54)

whereU0 is the maximum velocityin the middleof the pipe and U is the is associatedwith the conception velocityat radiusr. This relationship to thesquares stressesareproportional that the Reynolds's of theturbulent that the seems rate of velocity. It of dissipationof energy -components so far as changesin lineardimenin such a systemmustbe proportional,
sions, velocity, and density.are concerned, to pu'3/l, where 1 is some linear

4imension definingthe scale of the system. For turbulenceproduced therefore similarboundaries by geometrically W - constant For such systemstherefore /2 = 15 .
IV

(55)

t See equation(48).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

440

G. 1. Taylor

where C depends on the position relative to the solid boundaries of the point at which observationsaremade and on the elementused for defining1.

APPLICATION TO AIR STREAM BEHIND REGULAR GRIDS OR HONEYCOMBS

Formula (55) is specially well adapted for discussing the decay of turbulencein an air stream behind a grid or honeycomb, because it has been found that at a certain distance down-stream the stream becomes statistically uniform, i.e., the "wind shadow " of the grid disappears and the mean velocity becomes uniform. Under these circumstances it seems that the C of formula (55) must be a constant for any definite. form of grid. The researchesof Schlichting* have shown that at a short distance behind a cylindrical obstacle the wake assumes a definite form. The width of the wake and the velocity of the air in the middle of the wake depend on the drag coefficient of the obstacle so that obstacles, of very varied cross-sections produce identical wakes provided their drag coefficients are identical. For this reason it may be expected that if a regulargrid or honeycomb is constructedthe scale of the turbulentmotion produced by it at any distance down-stream beyond the point where the "' wind-shadow" has disappearedwill depend only on the form and mesh size of the grid, and not on the cross-section of the bars or sheets from which it is constructed. On the other hand, the velocities of the turbulent components will certainly depend on the drag coefficient of the bars, themselves as well as on the distance down-streamfrom the grid at which measurements are made. These considerations lead to the prediction that if only one-form of mesh is considered, say a square mesh, and if the length I in (55) is taken as M, the mesh length, i.e., the side of each square of the mesh, then the constant C in (55) will be an absolute constant independent of the form of the bars of the grid. We are thus led to a definite expression for x/M namely, (56) -=A \ where A is an absolute constant for all grids of a definite type, e.g., for
all s-iiiare-mesFh grids or honevcomhs.
* IIngen. Arch.,' vol. 1, p. 533 (1930).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theory of Turbulence


PREDICTION OF LAW OF DECAY OF TURBULENCE BEmND GRIDS AND HONEYCOMBS

441

We are now in a positionto predictthe way in which turbulence may be expectedto decaywhen a definitescale has been given to it as the air streampasses througha regulargrid or honeycomb. The rate of loss of kineticenergyof the turbulence per unit volumeis
ipU

d_(I2

+V2+W2),

whichin an isotropicfield of turbulence is


-

pU

(u2)

This must be equalto the rate of dissipation W, so that


3

U d (u - 15.S22

(57)

This equation is capable of experimentalverificationsindependently of the relationship (56) betweenXand M because,as has been shown,X is connectedwith R. through(48) and R, can be measuredinstrumentally. On the otherhand,if the relationship (56) betweenX and M is assumed to hold it is possibleto calculatethe law of decay of turbulence. Substitutingfor Xfrom (56), (57) becomes
u'3 dx

U d (u 2)

10
2' MA

(58)
58

and integrating (58) the followingvery simplelaw of decay is predicted,


u' =A2M -+ constant.

(59)

shouldbe applicable This expression to all cases wherethe turbulence is of a definitescale. The linearlaw of increase in U/u' shouldtherefore applyto all wind tunnelswherethe scale of turbulence is controlled by a honeycombor grid, and the value of the constantA determined experifor all squaregrids. Thus, the mentally,using (59), should be universal turbulencebehind a square-section honeycomb with long cells should obey the same law of decay as that producedby a square-mesh grid of flat slats or a square-mesh grid of roundbars,and the valuesof A should be identicalin all thesecases.
VOL. CLI.-A.

2 G

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

442

G. I. Taylor

For other types of grid or honeycomb, e.g., with hexagonal or triangular cells or a grid of parallel slats or plates, the constant A determined experimentally by applying"(59) to observed values of u' at different distances down the air stream might be expected to assume other values.
EXPECTED LIMITATIONSTO PREDICTED LINEAR LAW OF DECAY OF TURBULENCE

This is a very comprehensive prediction, but it is subject to certain limitations. In the first place it cannot be expected to apply when Mu'/v is small, for equations of the type (56) are not true when Mu'/v is small. In fact if (56) were supposed to hold when Mu'/v is small X would be greater than M, a condition which is clearly impossible at any rate near the grid. A second restriction is that the formula cannot be expected to apply in the region immediately behind the grid where the mean velocity is variable, i.e., where the " shadow " of the grid is still distinct. It is found experimentally that when the diameter of the bars of the grid is small compared with M the shadow may extend to as much as 20 M or 30 M behind the grid, but when the bars are as broad as W M the shadow disappearsa few mesh lengths down-streamfrom the grid. A third limitation may be expected to operate when the turbulence is not entirely due to the grid through which the stream passes. If, for instance, a very turbulent stream passes through a grid consisting of thin -wiresarranged in a large-scale mesh the scale of the turbulence in the stream might hardly be affected by its passage through the grid.
PREDICTIONS SUMMARY OF RESULTS AND THEORETICAL

(1) When the turbulence of a definite scale is produced or controlled in a stream of air by a honeycomb or grid of regularly spaced bars the scale of turbulence can be investigated in two ways. If the Lagrangian conception of fluid motion is adopted the scale of turbulence can be defined in referenceto the correlation R, between the velocity of a particle and that of the same particle at time . later. This conception is suited for `discussing experiments on diffusion of heat from a concentrated source. (2) If the diffusive spread of heat or matter from a line source is measured near the source it is proportional to the distance from the source and measures the transverse component of turbulent velocity independently of the scale.

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

Statistical Theory of Turbulence

443

(3) If the diffusivespreadis measuredat a number of positions extending far down-stream from the source a length 11analogous to the mean free path in kinetic theory of gases can be determined. It is anticipated that this will be some definite fraction of the mesh. size M of the honeycomb or grid. (4) Measurements of correlation between..simultaneous values of the velocity at points distributed along a line, can determine a length 12 which measuresthe scale of turbulencefrom the standpoint of the Eulerian representation of fields of flow. Both these lengths may be expected to be some definite fraction of the mesh length M, at any rate when the turbulence is not very small. (5) A third length X can. be defined in relation to the dissipation of energy by the equation

w -

15 U

This length may be taken to repre-

sent roughly the diameters of the smallest eddies into which the eddies defined by the scales 11or 12will break up. (6) If the rate of dissipation is proportional to the cube of the velocity, as it is where the Reynolds's stresses are proportional to the squares of the turbulent components of velocity, X is proportional to
V/u

In turbulence due to a square mesh honeycomb of mesh length M, I where A is a constant. This formula is inapplicable --= A /\/M when Mu'/v is small. (7) Using this value for X it is shown that the law of decay of turbulence is such that U/u' increases linearly with x in accordance with equation (59). (8) X is also directly connected with the correlation between simultaneous measurements of velocity at fixed points separated by a small distance. This correlation can be measured by suitable apparatus so that the theory can be verified experimentally. (9) In isotropic turbulence the mean value of any quadratic expression of the space rates of change in the velocity is known when the mean value of any one of the terms in it which is not zero is known. This leads to the prediction, which might be verified experimentally,that if the correlation between a component of velocity at a fixed point 0 and that at a neighbouring variable point P is measured, the surfaces of equal correlation are prolate spheroids with P as centre, the long axis is V2 times the equatorial axis and is directed in the direction in which the; velocity component is measured. This statement is identical,in substance, though not in form with that given on p. 439. 2 G2

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions

444

&GI. Taylor
CONCLUDING REMARKS

Of these results and predictions(1), (2) and (3) are substantiallyidentical with the conclusions put forward in 1921 in my paper, " Diffusion by Continuous Movements," where the suggestion that the diffusive power of turbulence should be used for the:purpose of measuring the scale of turbulence and the turbulent components was first made. Recently experimentsof this nature have been made by C. B. Schubauer* and (2) has been verified,as will be shown in Part IV of the present paper. Mr. Schubauer, however, worked quite independently of my previous work and indeed gives an empirical explanation of his experimental results. Conclusions (4) to (9) are, I believe, new. It will be shown in Part II that all these results, except (9), have now been verified experimentally and shown to be true. Experimentalwork is now in hand to test the truth of (9).

Statistical Theory of Turbulence-II


By G. I. TAYLOR, F.R.S. (ReceivedJuly 4, 1935)
MEASUREMENTSOF CORRELATIONIN THE EULERIAN REPRESENTATION OF TURBULENT FLOW

The methods describedin Part I have been used by Mr. L. F. G. Simmons, of the National Physical Laboratory,to find experimentallythe correlation between the turbulent components of velocity uo and u, at two points distant y apart in a direction transverse to the stream. The measurements were made at mean speed U 25 feet per second in a wind tunnel behind a honeycomb with 0 *9-inchsquare mesh. The results are shown in fig. 1 where the ordinates are Ru
-

Uo U2

and the abscissae are the

correspondingvalues of y. It will be seen that the RVcurve is apparently rounded at the top and that RVfalls to 0 08 at y = 0 38 inches. No measurementswere made beyond this point, but extrapolation seems to show that R =0 when y is about 0 5 inches, i.e., when y is slightly
ar.2ntEr th2nS.n

* ' Rep. Nat. Adv. Ctee. Aero., Wash.,' No. 524 (1935).

This content downloaded from 140.116.207.66 on Wed, 4 Dec 2013 11:20:08 AM All use subject to JSTOR Terms and Conditions