1

On the non-linear mechanics of hydrodynamic stability

By J. T. STUART
National Physical Laboratory, Teddington, Middlesex

(Received 11 November 1957)

SUMMARY
I n most work on the theory of stability of laminar flow,
infinitesimal disturbances only have been considered, so that
,only the initial growth of the disturbance has been determined.
I t is the object of the present paper to extend the theory to larger
amplitudes and to study the mechanics of disturbance growth with
the inherent non-linearity of the hydrodynamical system taken
into account.
The Reynolds stress (where averages are taken with respect
t o some suitable space coordinate) is the fundamental consequence
of the non-linearity, and its effects can be anticipated as follows.
Initially a disturbance grows exponentially with time according
to the linear theory, but eventually it reaches such a size that the
transport of momentum by the finite fluctuations is appreciable
and the associated mean stress (the Reynolds stress) then has an
appreciable effect on the mean flow. This distortion of the mean
flow modifies the rate of transfer of energy from the mean flow
to the disturbance and, since this energy transfer is the cause of
the growth of the disturbance, there is a modification of the rate
of growth of the latter.
It is suggested that, in many cases, an equilibrium state may
be possible in which the rate of transfer of energy from the
(distorted) mean flow to the disturbance balances precisely the
rate of viscous dissipation of the energy of disturbance. A theory
based on certain assumptions about the energy flow is given to
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

describe both the growth of the disturbance and the final

equilibrium state, and application is made to the cases of
Poiseuille flow between parallel planes and flow between rotating
cylinders. The distorted mean flow in the equilibrium state can
be calculated and from this, in the latter case, the torque required
t o maintain the cylinders in motion. Good agreement is obtained
with G. I. Taylor’s measurements of the torque for the case when
the inner cylinder rotates and the outer cylinder is at rest.

1. INTRODUCTION
The immediate objective of the theory of hydrodynamic stability is to
understand the mechanism of instability in laminar flow and to obtain a
F.M. A
 2 J. T. Stuart
criterion for its occurrence. A more fundamental objective is to understand
how, and under what circumstances, turbulence may arise from laminar
instability. The connection between laminar instability and turbulence
may not be a direct one, but under certain circumstances instability
of laminar flow will be a necessary prelude to transition to turbulence.
It is clear that the stability problem in its general form must be
considered to be non-linear, because the equations of motion are non-
linear.
The mathematical problem of hydrodynamic stability can be formulated
by taking the given steady-state solution of the equations of motion and
superimposing a disturbance of a suitable kind; this results in a set of
(non-linear) ' disturbance ' equations, which govern the behaviour of the
disturbance. If the solution of the equations shows that any disturbance
ultimately decays to zero, the flow is said to be stable; whereas if the
disturbance can be permanently different from zero, the flow is unstable.
It does not always happen that instability leads to turbulent motion, because
another (possibly more complex) form of laminar motion may be the result.
Indeed, it will be shown that this is often the case.
Naturally, the solution of the disturbance differential equations is
simplified considerably by linearization for small disturbances, and for
a description of theories based on this assumption the reader is referred
to the book by Lin (1955). On the basis of linear theory it is possible to
consider disturbances which contain an exponential time factor of the form
exp(kt), t being the time. The boundary conditions on the disturbance
equations require the vanishing at the boundaries of quantities like the
disturbance velocity components. Consequently, the boundary conditions
are homogeneous, and there is an eigenvalue problem for the determination
of the quantity k. I n this (linear) case, stability or instability is defined
as follows: if it is possible for k to have a positive real part, the flow is
unstable ; otherwise the flow is stable.
The prediction by linear theory of a disturbance which increases
exponentially with time is a feature which has occasionally given rise to
the suggestion that turbulence would necessarily ensue from the growth
of the disturbance to large amplitudes ; however, examples are known
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

in which this is not the case. On the other hand, it has been argued that
the non-linear terms will stabilize completely a flow which is unstable
according to linear theory, but such arguments can generally be refuted
(Stuart 1956a). I n this paper, certain features of the role played by the
non-linear terms of the equations of motion are discussed, and this leads
to a clarification of the connection between linear and non-linear instability
theories. Attention is restricted to flows which have constant local
Reynolds number.
I n cases of instability of fluid flow, the disturbance is usually periodic
in at least one direction of space. Thus it is convenient to take averages
with respect to one of the spatial dimensions, and to separate the flow into.
 On the non-linear mechanics of hydrodynamic stability 5
a mean part and a disturbance part, where the latter has zero mean*. It is
clear that the two parts of the flow are interdependent through the action
of the Reynolds stress (arising from the disturbance) on the mean flow.
(For a discussion of Reynolds stresses, the reader is referred to the books
by Goldstein (1938) and Townsend (1956).) On the basis of linear theory,
the disturbance is assumed to be so small that the effect of the Reynolds
stress on the mean motion can be neglected, in which case the mean flow
is the original laminar flow. However, in a non-linear theory the inter-
dependence of the mean and disturbance parts of the flow must be taken
into account. Let us now consider a flow whose local Reynolds number
does not vary, as in the case of flow between parallel planes or concentric
cylinders, and let the flow be perturbed by a small disturbance. Initially
the disturbance amplifies exponentially with time according to linear theory,
but eventually it reaches such a size that the mean transport of momentum
by the finite fluctuations is appreciable, and then the associated mean stress
(the Reynolds stress) has an appreciable effect on the mean flow. This
distortion of the mean flow clearly modifies the rate of transfer of energy
from the mean flow t o the disturbance and, since this energy transfer is
the cause of the growth of the disturbance, there is a modification of the
rate of growth of the latter. These processes, in which the disturbance
distorts the mean flow and the distortion of the mean flow modifies the rate
of growth of the disturbance, occur simultaneously.
It is natural t o enquire if an equilibrium state is possible, in which the
rate of transfer of energy from the mean flow to the disturbance balances
precisely the rate of viscous dissipation of energy of the disturbance. I n
such an equilibrium state, the disturbance will have a definite finite amplitude
and the mean flow will be distorted from its original laminar form.
Experimental evidence of an equilibrium state of this kind is afforded by
G. I. Taylor's observations on the instability of flow between rotating
cylinders, where the instability takes the form of cellular, toroidal vortices
spaced regularly along the axes of the cylinders. Taylor (1923, p. 342)
observed that " A moderate increase in the speed of the apparatus merely
increased the vigour of the circulation in the vortices without altering
appreciably their spacing or position ", and suggested that " The experiments
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

... indicate that the effect of the second-order [non-linear] terms is to prevent
* It should be mentioned that, an another formulation of the problem, the ' dis-
turbance ' is defined to be the whole of the deviation from the original laminar flow.
However, with this definition the ' disturbance ' must contain a ' mean ' part, a fact
which was overlooked by several authors in their studies of non-linear instability
theory (Stuart 1956 a). In the opinion of the writer, this approach does not yield
such a clear understanding of the physical processes involved in instability as does the
approach based on the concept of a mean flow and a disturbance flow (with zero mean)
interacting through the action of a Reynolds stress. Consequently, in this paper the
flow will always be separated into a mean part and a disturbance part (with zero mean).
The term ' disturbance ' will not be used to denote the deviation from the original
laminar flow.

A 2
 4 J . T. Stuart
the vortices from increasing indefinitely in activity ”. The combination of
the mean flow with a disturbance of definite amplitude may be referred to
as an equilibrium flow. Examples relating to Poiseuille flow between parallel
planes and to flow between rotating cylinders are described in $ 2 and $ 3 ,
and for the case of flow between cylinders good agreement with experiment
is obtained. For the case of Poiseuille flow, however, there is no experimental
evidence that an equilibrium flow of the kind described above does occur.
There is the possibility in this case, as in other cases, that a flow of this kind
does not occur, but rather that there is a continual generation of harmonics
of the basic disturbance and of other disturbances.
In the discussion above, attention has been paid to equilibrium flows
which may develop when the original laminar flow is unstable according to
linear theory. Some flows, however, such as Couette flow between parallel
planes and Poiseuille flow in a circular pipe, are completely stable against
infinitesimal disturbances. Even so, turbulence can occur at sufficiently
high Reynolds numbers. Furthermore, turbulence occurs in some flows
(for example, in Poiseuille flow between parallel planes) at a lower Reynolds
number than the critical according to linear theory, that is, it occurs when the
laminar flow is stable with respect to infinitesimal disturbances. A suggestion
which may lead to an explanation of such phenomena is that the appropriate
laminar flow may be unstable with respect to finite disturbances. When a
disturbance of suitable magnitude is present, the mean flow may be
distorted to such a form that the rate of transfer of energy to the disturbance
can balance exactly its rate of dissipation by viscosity. On the other hand,
a finite disturbance which is small enough will presumably decay to zero
amplitude, either because the rate of energy transfer to the disturbance is
insufficient to balance the rate of viscous dissipation of kinetic energy, or
because the energy transfer is actually from the disturbance t o the mean
flow. As an example of instability for finite disturbances, Meksyn & Stuart
(1951) considered the case of Poiseuille flow between parallel planes and
showed that the critical Reynolds number drops as the amplitude of the
disturbance rises. The reader is referred also to Lin (1955) and Stuart
(1956 b) for a discussion of these ideas. A discussion of the formulation of
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

the instability problem for plane Couette flow with finite disturbances was
given by Noether (1921).
It will be convenient to refer to non-linear disturbances as existing under
supercritical conditions if the Reynolds number is above the value which is
critical for linearized instability, and as existing under subcritical conditions
if the Reynolds number is such that the flow is stable with respect to
infinitesimal disturbances. A non-linear disturbance may clearly arise
spontaneously under supercritical conditions as a result of continued
amplification of a secondary disturbance ; the sequence of events which
would lead to a non-linear disturbance under subcritical conditions is less
evident and will not be discussed here.
If equilibrium flow is established consisting of a mean flow with a
.steady-amplitude finite disturbance, it does not follow that it is a stable flow
 On the non-linear mechanics of hydrodynamic stability 5
at all Reynolds numbers above the critical value for which it first occurs.
It is likely, however, to be stable for a certain range of Reynolds number
above the critical. Moreover, it is clear that it must become unstable at
some Reynolds number if turbulence is to occur at higher Reynolds numbers.
For example, in the case of flow between rotating cylinders mentioned
above, Taylor observed that “ a large increase [in the angular speed]
caused the symmetrical motion [of the cellular vortices] t o break down
into some kind of turbulent motion, which it was impossible to follow by
eye ”. Another interesting case is that of Poiseuille flow between parallel
planes, where the critical Reynolds iiumber according to linear theory is
a Reynolds number for which the flow is normally turbulent. Consequently,
equilibrium flows under supercritical conditions are, in this case, almost
certainly unstable. Similar considerations apply to equilibrium flows under
subcritical conditions, at least for the higher Reynolds numbers for which
they are valid.
T h e development of a non-linear instability theory for boundary-layer
flows, where the flow and local Reynolds number change in the stream
direction, presents additional difficulties. Whereas in the linear instability
theory it is permissible to regard a boundary-layer flow as nearly parallel
and to neglect boundary-layer growth, it does not seem obvious that such
an approximation is permissible in a non-linear theory. Because the local
Reynolds number increases in the downstream direction, any disturbance
is convected into regions of higher Reynolds number and the effect of this
continuous change of Reynolds number would have to be taken into account.
Consequently, the non-linear theory which will be described in this paper
for flows with constant local Reynolds number does not necessarily apply
quantitatively to the case of the boundary layer. T h e main features of the
theory are of wide applicability, but there may be additional factors
influencing them because of the growth of the thickness of the boundary
layer.
An interesting suggestion concerning the development of turbulence
from the growth of small disturbances has been made by Landau (1944)-
As noticed above, the occurrence of instability in a flow may lead to the
replacement of the original laminar flow by a new laminar flow, which consists
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

of a mean flow with a superimposed finite disturbance. This flow may

be expected to persist for a certain range of Reynolds number above the
critical value and then to become unstable at some Reynolds number against
a new (second) type of disturbance. A new equilibrium flow, consisting
of a mean flow with two superimposed modes of disturbance, is then
conceivable for a range of Reynolds number above the second critical value.
As the Reynolds number is raised still further, additional modes of
disturbance may appear successively until, at sufficiently large Reynolds
number, the flow is so highly disturbed as to be considered turbulent. In
the case of flow between rotating cylinders, experiments show that the
development of turbulence takes place fairly slowly as the Reynolds number
is raised ; this would correspond to the first two at least of the successive
 6 J . T.Stuart
critical Reynolds numbers of Landau’s theory being fairly widely spaced.
O n the other hand, there are cases of flow in which turbulence develops
rather suddenly as the Reynolds number is raised, and in these cases one
might infer that the critical Reynolds numbers are close together.
I n the discussion given above, attention has been entirely directed to
instability in fluids which are in motion. There is, however, another
important type of instability, namely that which occurs when a horizontal
layer of fluid is heated from below, causing an unstable density gradient
in the fluid. Instability takes the form of convection in cells of polygonal
planform, and occurs when a certain parameter (the Rayleigh number)
is above a critical value. T h e linear theory of instability is well understood
(see, for example, Pellew & Southwell 1940) and is known to be analogous,
under certain conditions, to the theory of instability of flow between
rotating cylinders. T h e fundamental physical process in the instability
is the conversion of potential energy associated with the gravitational field
into kinetic energy of the convective disturbance motion. When the
disturbance has a finite amplitude, the mean heat transport by the
convective motion causes a modification of the mean temperature distribu-
tion, where averages are taken over horizontal planes ; therefore, because
of the corresponding modification of the mean density distribution, the rate
of transfer of energy from the gravitational field to the disturbance is
modified. It appears that a steady (equilibrium) state is possible, in which
the rate of transfer of energy into the disturbance balances precisely the rate
at which energy is dissipated and diffused. T h e non-linear theory of thermal
convection has been studied by Sorokin (1954) and by Malkus & Veronis
(1958), both by methods related to those of the present paper and by
perturbation series expansions.
Some of the results obtained in this paper are similar to results obtained
from the turbulence models due to Burgers (1948) and Hopf (1956). For
a discussion of non-linear instability effects in the Burgers model, the reader
is referred to a paper by Stuart (1956 b).

2. DISTURBANCES
UNDER SUPERCRITICAL CONDITIONS IN POISEUILLE
FLOW
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

BETWEEN PARALLEL PLANES

In this section a simple treatment of the non-linear problem for
disturbances under supercritical conditions is described, and it is applied
in this section and in $ 3 to the cases of Poiseuille flow between parallel
planes and of flow between rotating cylinders. T h e method is based
essentially on a balance of energy, taking into account the mean-flow
distortion due to the Reynolds stress, and it is the ‘ integral ’ properties
of the flow which are treated rather than the spatial details. Consequently,
it is clear that the method has an obvious application to cases of instability,
such as that of flow between rotating cylinders, where it is the overall
properties of the flow (as opposed t o the details) which are important.
O n the other hand, in cases of instability which are governed by the
 On the non-linear mechanics of hydrodynamic stability 7
Orr-Sommerfeld equation (for example, the case of plane Poiseuille flow),
the instability depends very much upon precise details of the mean flow,
especially in the vicinity of the critical layer. Consequently, the method
is less valid for such cases, and the application to plane Poiseuille flow has
less validitythan the case of flow between rotating cylinders. Moreover, there
is less experimental evidence of quantitative value in the former case than
in the latter. In spite of these facts, the application to plane Poiseuille flow
is of considerable illustrative value, and is therefore presented in this section
before a treatment in § 3 of the case of flow between rotating cylinders, for
which a comparison with experiment will be made.
Consider a two-dimensional flow between parallel planes, and let x1
denote the coordinate parallel to the planes and x2 the coordinate normal
t o them, with ul, u2 denoting the corresponding components of velocity,
Let P denote pressure, p density, v kinematic viscosity and t the time.
Now let us introduce a reference length L, equal to half the distance between
the planes, and a reference speed Uo,the maximum speed in laminar flow.
'Thus we define
.XI = L x , x2 = Lx, u1 = uou, u2 = uow, P =ppu;, t = Lt/Uo. (2.1)
'Then the Navier-Stokes equations and equation of continuity are
au
at
au
ax
aw +u-aw +w-aw
-
au
- + u - +w-
ax
=

= -
ap + -
--
ax R
1
VZU,
1
at ax ax
-au+ - =aw
o,
ax ax J

where V2 = a2/ax2+ a2/ax2,and R = U,,L / v denotes the Reynolds number

for the problem.
The solution for undisturbed laminar Poiseuille motion is

z i = 1-x2, W = 0, = - 2 x / R fconst., (2.3)

https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

where a bar above a quantity denotes a mean value with respect to the
distance x. This is the basic laminar motion whose non-linear instability
characteristics we wish to examine.
Now let
u=G(x,t)+u'(x,x,t), w = W'(X,X,t), (2.4)
where u'and w' denote components of a finite disturbance (with zero mean).
Since the disturbance is finite, G is no longer given by (2.3) but is distorted
by the Reynolds stress. Furthermore, if the disturbance is growing or
decaying in amplitude, C depends on time. We suppose that u' and w'
.are harmonic in x with wavelength 2n/x and are given by Fourier series.
I f we substitute (2.4) into (2.2)) separate out the mean and disturbance
 8 J. T. Stuart
parts of the motion and then integrate the resulting disturbance equations
to obtain the energy balance relation, we have (without approximation)

= I l - ISIR, (2.5)
where the integrands are evaluated over a volume bounded by the planes
and by one wavelength. The mean velocity ii occurring in (2.5) is given
-aii+ - =au'w' ap 1 a 2 i i
by
aT ax
--
ax R- -9ax-*+ (2.6)
For a discussion of the derivation of the above equations, and for the
non-linear disturbance differential equations for u' and w', the reader is
referred to Stuart (1956a, b). Equations (2.5) and (2.6) will be used
here as an approximate basis for the solution of the non-linear problem
of the growth of the disturbance (u', w ' ) .
We note that in equation (2.5) the term on the left-hand side gives the
rate of growth of the disturbance energy within the volume considered.
On the right-hand side, the term Il is the integral of the product of the
Reynolds stress and the mean velocity gradient, and represents the rate
of energy transfer from the mean flow to the disturbance ; the term 12/R
is always positive and represents the rate of viscous dissipation of energy
of the disturbance. If Il is positive and greater than 12/R,the disturbance
energy is growing and the disturbance is increasing in amplitude.
Equation (2.6) shows how the distribution of mean motion is affected by
the viscous stress, pressure gradient, and Reynolds stress due to the
disturbance. An equilibrium flow is possible if ii can be so distorted by
the Reynolds stress that Il = I J R for a given Reynolds number.
It is worth considering at this point the boundary conditions on the
mean motion. For the velocity they are ii = 0 at z = +. 1. We shall also
adopt the condition that the mean pressure gradient shall remain unchanged ;
this is clearly physically realistic, since the pressure gradient is externally
applied*. As a consequence, the mean skin friction on the walls will be
the same in the equilibrium flow as in the laminar flow. However, the mean
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

velocity near the channel centre must drop when a disturbance is present,
because energy is required to maintain the disturbance. The unchanged
pressure gradient is given by equation (2.3) and the mean motion equation
thus takes the form
aa a - 2 1 a2ii
- + - (u'w') = - + - -
at ax R R ax2'
I n a state of equilibrium, ii is independent of time and equation (2.7) is
easily integrated to yield
ii(z)= 1 - x 2 + R 1: dz.

* Other boundary conditions, such as that of constancy of mass flux, yield qualita-
tively similar results for the amplitude of oscillation.
 On the non-linear mechanics of hydrodynamic stability 9
-
Since ufwf is - an odd function of z, C = 0 at x = 1. Moreover, since
Ilis positive, u'w' must be of opposite sign to aiildx in at least a dominant
range of z ; therefore, it must be dominantly positive when x is positive.
Thus 6 is less than unity at x = 0 and, in general, 6(x)is everywhere less
than its value for laminar flow. A corollary of the condition on the mean
pressure gradient is that the Reynolds number is based on a velocity U,
which is the maximum for laminar flow with the same mean pressure gradient.
Suppose we now define a stream function for the disturbed flow in the
form
+= +o(z,t ) + + l ( x , T ) e z a ( ~ - c r') + J 1 ( z , T ) e - z a ( s - c r 7 ) + +z(x, T)e2ia("-cr
7) +
...,
-
N

+&(z,~ (2.9)
) e - ~ l ~ ( ~ - ~ r ~ ) +

where the symbol represents a complex conjugate. represents a +

mean flow together with a periodic disturbance consisting of the first
harmonic with wavelength 2n/a, and higher harmonic components having
wave-numbers na ( n integral) but the same (real) wave velocity c,, which
is assumed to be independent of time. T h e amplification or damping of
the disturbance, and the consequent changes in the mean velocity (given by
21 = a#,/az), are accounted for by the dependence of all the +functions on 7.
If we substitute (2.9) into (2.2), utilize
= a+jax, = - a+lax, (2.10)
eliminate the pressure and separate out the Fourier components of the
disturbance, we obtain a set of equations for the functions ii, +1, &, and
so on. T h e equations are similar to (3.2), (3.3) and (3.4) of the paper by
Stuart (1956 b), except that allowance must be made for the dependence
of the functions on 7. Because of the non-linearity of the system, all of
the functions are mutually dependent and an infinite set of equations has
to be solved t o obtain an exact solution. I n particular, the mean-motion
equation is (2.6) with
z-7= ia{f&& - + 2(& J2 - C2) + ...I,
(2.1 1)'
where primes denote differentials with respect to z .
In the linear theory of instability, the Reynolds stress is neglected,
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

so that the mean velocity is the laminar Poiseuille flow. Furthermore,

the stream function (2.9) contains only the functions 4, and q$, the real
part of which determines the actual stream function. T h e problem then
is the solution of the Orr-Sommerfeld equation (see Lin 1955), namely
(ii-cr-ic2)(+; - ~ 2 + ~ ) - = - (i/aR)(~~-2aZ+;+a4~,), (2.12)
where ii = 1-z2 and the function +1 is proportional to exp(acj T ) . T h e
non-linear problem posed here is to follow the growth of a disturbance
at a given Reynolds number and, in particular, to find out if an
equilibrium flow is possible at that Reynolds number. T o this end,
the stream function (2.9) has been chosen to allow for changes in the basic
disturbance, partly through the generation of harmonics, due to the
non-linearity .
 10 J. T. Stuart
A simple approximate method of solution
T h e non-linear effects of a disturbance of finite amplitude appear t o
be of two kinds. On the one hand there is, through the Reynolds stress,
a modification of the energy transfer between the mean flow and the first
harmonic component, +1, of the disturbance. O n the other hand, higher
harmonics etc.) of the disturbance are generated and there is an
interchange of energy between them and both the mean flow and the first
harmonic component. T h e assumption which will be made here is that the
dominent non-linear interaction is that between the mean flow and the first
harmonic component of the disturbance. To put it in another way, the
effect of the Reynolds stress on the mean flow is taken into account, thus
causing modifications of the energy transfer between the mean flow and the
first-harmonic disturbance, but the generation of higher harmonics is
ignored. A similar assumption was made by Meksyn & Stuart (1951)
for disturbances under subcritical conditions. I n a formulation of this
kind there are two differential equations to determine C and +1, namely, (2.7)
and (2.12) with cireplaced by a - l a / a ~; in equation (Z.ll), +2 = +3 = ... = 0.
I n the present case of disturbances under supercritical conditions, we
shall make the additional assumption that the disturbance (u’, w’) is
similar in shape t o a solution given by linear theory, but that the solution
is multiplied by an amplitude factor, a(T), which is a function of time.
Then the approximate problem is to satisfy the disturbance energy-balance
equation (2.5) and the equation of mean motion (2.7), in which case the
amplitude is determined. If the amplitude of the disturbance in an
equilibrium flow is required, E is given in terms ofu’w’ by (2.8) and can
be substituted into (2.5). With the ‘ shape ’ assumption, (2.5) then gives
a relation between a2 and a4, so that a2 is determined in terms of integrals
involving functions given by linear theory. This is the most important result.
An equation illustrating the growth of the disturbance may be obtained
by noting that the term a t s / a ~in (2.7) is negligible at small times on the
basis of linear theory, and also at large times because it is to be expected
that a steady mean flow will be approached. If, on this basis, the term
a C / a ~is ignored at all times, (2.5) and (2.8) yield the equation
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

yl-aa2 = y2aa2-y3a2Ra4- Y4 -, a2 (2.13)

aT R
where a is the amplitude of the disturbance and

71 = j {Id‘12++21+121&
1
72 = 4 j 1
4+i+;-+?.+w,

1
0 0
(2.14)
4 j (+i+j-+T+;)2dx,
1 1
y3 = y4 = 2 j (1+”12+2~21+’12+~41+12}~~,
0 0
the suffixes r and i denoting real and imaginary parts. The function $ ( z )
is the amplitude distribution of a disturbance stream function according
-to linear theory. Equation (2.13) is of the form
dY (2.15)
= PlY - P2Y-P3Y27
 On the non-linear mechanics of hydrodynamic stability 11
where y is the square of the amplitude of the disturbance. Equation (2.15)
was given earlier by Landau (1944), although he does not state how, and
under what assumptions, it was derived.
The terms in (2.15) and the corresponding ones of (2.13) have the
following physical meanings. The term dy/d-r represents the rate of change
of the kinetic energy of the disturbance, the term p l y the rate of energy
transfer from the mean flow to the disturbance, the term P2y the rate of
viscous dissipation of the disturbance, and the non-linear term P3y2 the
restriction of the rate of energy transfer to the disturbance by the
Reynolds-stress distortion of the mean flow. For very small disturbances,
the non-linear term is negligible, and if P1 > P2, the disturbance amplifies
like C exp[(pl - P 2 ) 7 ] , where C is a constant. This solution corresponds
to linear stability. On the other hand, if the Reynolds stress is included
the general solution of equation (2.15) is
Ceii31-Psh
= 1 +(&)Ce".-@h ' (2.16)
P1- P2

When T + - co, y + C e ~ p [ ( ~ , - / 3 ~ ) ~which

], is the linearized solution
given above. Furthermore, when T + + co, y e (P1-P2)/up3, whatever
the value of C. This solution suggests that, whatever the initial size of
the disturbance, it always builds up to the same limiting amplitude. This
conclusion is physically reasonable, although it should be remembered
that the exact form of (2.16) is based on the assumption that an/& may
be neglected at all times.
Reverting now to (2.13), we note that the critical Reynolds number is
R, = Y 4 b Y 2 ,
the amplification factor of r,b2 on linear theory is

and the square of the equilibrium amplitude is

https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

Thus a, is proportional to the square root of the difference between the

actual Reynolds number and the critical Reynolds number. In order to
calculate specific values of the amplitude at a given Reynolds number, it is
necessary to approximate to the function 4. A fairly obvious choice would
be the function appropriate to the minimum critical Reynolds number. No
calculation of this function has been made, although Thomas (1953) has
calculated and tabulated 4 for tc = 1, R = lo4,c = c, + ici = 0.2375 + 0.0037i.
The function is illustrated in figure 1, 9? denoting the real part and 9 the
imaginary part. Numerical integration of (2.14) then gave
y1 = 2.05146, ye = 0.040192, y 3 = 0.002308, y4 = 247.62. (2.17)
 12 J . T.Stuart

Using these values, we find that uR,= 6150, which is in reasonable agreement
with Thomas’s critical value, namely, R, = 5780 at M = 1.02. Thus, even
though the function is not the correct stream function at the critical
Reynolds number, the energy-balance relation yields a fairly good approxi-
mation to the critical Reynolds number. It is suggested, therefore, that
the numbers (2.17) may be applicable over a wide range of R,and possibly
also for small variations of cc. Another characteristic of the linear instability
theory which can be calculated from (2.17) is the amplification consiant ci
at cc = 1, R = 104, and this is found to be ci = 0.00376. The closeness
of this to Thomas’s value, 0.0037, serves as a check on the accuracy of the
numerical work.

1 1 I . I I I
0 0.1 0’2 03 04 0.5 0.6 0.7 0.8 0’9 1.0
Channel 2 Wall
centre
Figure 1. Function 4 for Poiseuille flow at R = lo4, 01 = 1, c = 0.2375+0.00373.

For the non-linear properties, we consider a wave-number, cc = 1,

since this is close t o the value at the critical Reynolds number. The
equilibrium amplitude number, a,, rises from zero at the critical Reynolds
number (6150) through 0.00256 at R = lo4 to a maximum value of 0-00266
at R = 12300 ( = 2R,). T o obtain the actual (dimensionless) stream
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

function at a given value of R,the curves in figure 1 should be multiplied

by the appropriate value of a. The distribution of Reynolds stress is shown
in figure 2 (obtained from Thomas’s calculations), from which the actual
(dimensionless) Reynolds stress can be obtained by multiplying by 2.
Using the calculations described above, the mean velocity profile of
the equilibrium flow can be calculated from equation (2.8) ; for R = lo4,
it has the form shown in figure 3. The distorted profile is everywhere less
than the laminar profile because of the energy transfer to the disturbance,
but has the same gradient at the wall. It can readily be shown that the work
done by the pressure gradient to maintain the motion precisely balances the
total viscous dissipation of energy in the equilibrium flow. The root-mean-
squares of the disturbance velocity components, (>)I12 and are: (z2)1/2,
 O n the non-linear mechanics of hydrodynamic stability 13

I I I I

016 - -

-
--
012 -

ac
6-

_IL -

-
008-
8
a-
N

004 - -

0- 1 I I
0 02 0'4 06 0.8 1-0

centre

Figure 2. Reynolds stress function for Poiseuille flow at R = 104.

https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

I
0 02 0.4 0.6 0.8 1.0
Channel Wal
centre
Figure 3. Comparison of laminar and distorted mean flows.
 14 J. T. Stuart
shown in figure 4 for a = 1, R = lo4. The magnitudes are similar to those
applicable in turbulent motion at the same Reynolds number. A further
point worthy of note is that, at a Reynolds number of lo4, the maximum
value of the Reynolds stress is of the same order as the maximum value of
the viscous stress.
As mentioned in the Introduction, the equilibrium flow under super-
critical conditions in Poiseuille flow between parallel planes is almost
certainly unstable, because turbulence normally exists at those Reynolds
numbers for which the supercritical flow exists. However, it is of interest
that the magnitude (see figure 4) of disturbance which can be sustained
is similar to the magnitude appropriate to turbulent flow at the same
Reynolds number (Reichardt 1938; Laufer 1950).
0.051 I I I I I I I I 1

centre

Figure 4. Root-mean-square velocity fluctuations at R = lo4.

In connection with the establishment of the equilibrium flow, it has been

https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

suggested here that an infinitesimal disturbance at a given Reynolds number

can amplify into an equilibrium flow. Experimentally, however, it is the
equilibrium flow itself which is important, rather than the way in which
it is attained. I n practice the equilibrium flow may arise as follows. Suppose
the flow can be kept free of disturbances, so that disturbances and turbulence
do not occur under subcritical conditions. Then, as the Reynolds number
is raised, a weak disturbance will appear at the critical Reynolds number.
The disturbance will have a small but finite amplitude. Then, as the
Reynolds number is increased, the flow could (if it were not unstable)
retain the same wavelength as the disturbance which first occurred, but
with an amplitude a R-l(R - Rc)1’2. This idea is discussed again in $ 3
N

in connection with the flow between rotating cylinders.

 On the non-linear mechanics of hydrodynamic stability 15

3. DISTURBANCES
UNDER SUPERCRITICAL CONDITIONS IN FLOW BETWEEN
ROTATING CYLINDERS
The second application of the theory will be to the flow between two
concentric rotating cylinders, with Y , 8, 2: as the cylindrical coordinates
and u, v, w the corresponding velocity components. It is assumed that
the flow has rotational symmetry and is therefore independent of 8. Then
the Navier-Stokes and continuity equations are

av
- +u -
at
av
ay + w -av ( :> 1I
az + -ue’ = v vz-- , ‘z,

(3.1)
aw + u -aw +w- aw = - -1 -
- aP +vvzw,
at ay aZ P 22:

where
(3.2).

It is known from the work of Taylor (1923) and others that, when the
rotational speed of the inner cylinder (or, to be more specific, a parameter
sometimes called the Taylor number) is above a critical value, the steady
laminar flow is unstable. T h e disturbance which appears takes the form
of cellular, toroidal vortices spaced regularly along the axis of the cylinders.
The linear instability theory for this flow is well known (Taylor 1923)
and our purpose here is to study the non-linear theory, primarily in order
to obtain the amplitude of the equilibrium flow.
Since the flow which results from the instability is periodic with respect
to z, it is convenient to take averages with respect to z. I n order to allow
for the distortion of the disturbance by the non-linearity, we write
u = u’ = ul(r, t)e”z+u,(r, t)e%iaZ+... +u”,(r, t)e-iaZ+UU2(Y,t)e-2”Z+ ... (3.3)
together with exactly similar series for n’ = v - 6(r, t ) and w’ = w, where
the symbol denotes a complex conjugate and a bar above a quantity
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

denotes a mean value. If the above series are substituted into the equations
of motion (3.1) and the Fourier components are separated, there results
a set of equations for all the velocity functions involved in (3.3).
For the mean motion, we find

(3.4)’

(3.5).
similar equations have been derived and discussed by Townsend (1956)
in connection with turbulent flow. Equation (3.4) gives the radial pressure
gradient required to balance the centrifugal force and Reynolds stress,
 16 J. T. Stuart
while (3.5) gives the mean rotational velocity 5 as a function of the Reynolds
stress. The dependence of 5 on t is retained because, if the disturbance
is growing or decaying, the mean motion undergoes distortion in order
t o maintain the energy balance. A suitable boundary condition to apply
to the mean flow is that just enough external power is supplied to maintain
the angular velocities of the cylinders at constant values, even though the
skin friction is changing. If rl and r2 are the radii of the inner and outer
,cylinders and Ql and Qz their angular velocities, the boundary conditions
are
d = r1 Ql at r = yl; d = yZ Q2 at r = r2. (3.6)
I n a state of equilibrium, ad/& is zero and (3.5) can be integrated to give

and m = sZ2/sZl. With zero Reynolds stress, these relations become those
appropriate to laminar Couette flow between rotating cylinders.
I n addition to the mean motion equations, it is possible to write down
an infinite set of differential equations for the harmonic components of
the disturbance. These components are all mutually dependent because
of the non-linearity of the system. These equations will not be derived
or considered in detail here, because it is proposed to use an approximation
similar to that used in $2. The assumption is that the dominant non-linear
interaction is between the mean flow (a) and the first harmonic component
of the disturbance ( U ~ , ~ ~ ~ where
, W ~ ) ,the latter is assumed to be similar
in shape t o that given by linear theory but multiplied by an amplitude
factor. The mean flow is distorted by the Reynolds stress of the disturbance
and the consequent alteration of the energy transfer between mean flow and
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

disturbance determines the amplitude. As in $2, the disturbance energy-

balance relationship is required for this purpose.
If the equations of the mean motion are subtracted from the equations
(3.1), equations are obtained for the disturbance velocities u’, u’, w’ (as
defined by (3.3)). From these it is possible to obtain the following
disturbance energy equation, which is an exact consequence of the equations
for an axisymmetric disturbance :

-p JJ ( ~ 2 + q ‘ 2 + ~ ’ 2 ) r d (3.10)
~d~,
 On the non-linear mechanics of hydrodynamic stability 17
the vorticity components of the disturbance being given by

The three integral terms of (3.10) can be interpreted in a similar way to those
of (2.5). The net rate of increase of disturbance energy is equal to the
difference between the integral of the product of the Reynolds stress and
the flow shear (i35/ar- 5 / r ) , which represents the rate of transfer of kinetic
energy from the mean flow to the disturbance, and the rate of viscous
dissipation of kinetic energy.
The approximate non-linear problem of stability is that of solving the
equation of mean motion, (3.5), and the equation of energy balance for the
disturbance, (3.10), where, in the latter, the disturbance is assumed to be
specified in shape but not in amplitude. The pair of equations (3.5)and (3.10)
can then be used to determine the amplitude, as in $2. If the amplitude
of the disturbance in the equilibrium flow is required, G is given in terms
o f t l " by (3.7); then (3.10) involves both second and fourth powers of the
amplitude, and the latter can therefore be determined. Moreover, an
assumption similar to that of 3 2 concerning the growth of the disturbance
can be made. This assumption is that at7/at can be ignored at all times
on the grounds that it is certainly negligible at both small and large times.
By the procedure outlined above a differential equation of the form (2.15)
is obtained, and it has the solution (2.16). I n order to calculate the
coefficients in the equation, it is necessary to specify the shape of the flow
field so that several integrals can be evaluated. I n the detailed calculations
which follow, attention is restricted to the case Q2 = 0 so that comparison
can be made with experiments of Taylor (1936). Secondly, only the very
simplest case is considered, that in which the annular gap is small compared
with the radii of the cylinders. The linear stability problem (Chandrasekhar

'
1953) is then specified by
( 0 2 - h2) ( 0 2 - A2 - cr)2~)l+ X2 Tv, = 0 ,

vl = D2vl = D(D2-X2--)vl = 0 at = & i, J' (3.12)

where
1
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

d = YZ-rl, 2~0 = r1+r2, Y = r0+<d,

X =ud,
T = Q?rld3/u2,
cr = kd2/v,

R = Q2,rld/u.
D s d/d<,
J
r (3.13)

The disturbance is proportional to exp(kt). T h e symbol T denotes the

Taylor number and R denotes the Reynolds number.
T o the same approximation ( d < y o ) the mean flow is given by
d = $rlGl(l -25)-(azr~G2::d2/u2)~(t;), (3.14)
-112
Q(5) = - 2 r Zd5+2t;r2 Zd5+ J Zdt, (3.15)
-112 -112 -112
2 = - SP, P = (D2- X2)S, u1 = - ar, Ql P, v1 = (ar: Q; d / u ) S ,
(3.16)
P.4. B
 18 J, T . Stuart
u being the amplitude of the velocity u I . 'The flow shear is given to the
same approximation by
a@ G rl Q1 2a2r;,"2' d ( j l l 2 2 d i - - Z ( ( ) ) . (3.17)
ar r d -1/2
It has been assumed that the Reynolds-stress term in (3.14) is much larger
than the terms of order d/ro.
If the approximation d < r o is adopted and a few transformations are
applied, the largest terms of (3.10) yield
da2 y4a2
= y 2 a 2 - y Y s R2a 4 (3.18)
y1 d7 A2T '
where
T = 2ut/d2, (3.19)

This equation is of the form (2.15) and has the property that the amplitude
tends to a limiting value with the passage of time. If the Reynolds-stress
term ( - y 3 R2u4)is neglected and the disturbance is neutral, equation (3.18)
is simply Chandrasekhar's (1953) variational condition. The critical
Taylor number is given by h2Tc= y4/yz and the equilibrium amplitude by
(3.21)
Thus a p is proportional to the square-root of the difference between the
actual Taylor number and the critical Taylor number.
Equation (3.12) can be used to specify the function S ( l ) and, since we
are primarily concerned with the equilibrium flow, S(C) will be assumed
to be the function appropriate to cr = 0, h = hc = T , T = T, = 1708. This
is the case of a disturbance at the critical Taylor number, for which we
have, from the variational condition given by Chandrasekhar (1953),
, y = - - + -1+ - + - 1 C2 COST( A
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

8n2 7r4 2n2 477-3

cosh TC
x- (3.22)
cosh +T
to a good approximation, where A = -0.02686. Then for a range of
Taylor numbers above the critical value, and for the particular wave-number
X = T , the amplitude can be evaluated according to (3.20), (3.21) and (3.22).
Thus
(3.23)
This relation, together with (3.16) and (3.22), then gives the velocity
distribution of the disturbance.
 On the non-linear mechanics of hydrodynamic stability 19
Once the amplitude of the velocity distribution is known, the mean
skin friction can be calculated from (3.17) and the mean torque on a cylinder
of given length then follows. Taylor (1936) carried out experiments with
cylinders 84.4 cm long, the outer cylinder having radius r2 = 4-05 cm
while the radius of the inner cylinder varied from case to case. Here we
shall compare the measured and calculated torque for the case rl = 3-94cm.
The number of vortices spaced along the axis is 700 approximately and,
since alternate vortices rotate in opposite directions, the torque measured
is the mean torque, where the term ' mean ' is defined in the sense of this
paper. I n figure 5, G denotes the torque measured in units gmcm2sec-2
and N the angular speed measured in units rev.sec-l. It can be seen
that the agreement is good for Taylor numbers ( T being proportional
5.0
I I I I I I [ I I I I I
- -
Experiment ( G . 1 Taylor)
-
X
-
- 5-X\
-
- x'x

-
0
-
Theoretical
-
- Theoretical
f o r laminar flow
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press
 20 J. T. Stuart
will be valid for some range of Taylor iiumber above the critical value,
and figure 5 shows that this is indeed the case. For a non-linear instability
problem which has a certain similarity to the present one, that of thermal
convection in a horizontal layer of fluid, Malkus & Veronis (1958) have
devised a rigorous expansion procedure which gives the limiting steady
amplitude of convection for a certain range of Rayleigh number above the
critical value. Their method gives results very similar to those obtained
by an integral procedure analogous to the present one. Presumably such
an expansion would be valid for the cylinder problem also.
I I I

5 -0’4 -0.3 -0 2 -0-1 0 01 0.2 5

Figure 6. Mean flow between rotating cylinders ; log,,(N/~)= 2 . 5 , T = 20 98‘7.

With regard to the way in which an equilibrium state may be set up
at a Taylor number above the critical value, it has been suggested here
(though not proved rigorously) that an infinitesimal disturbance which
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press

increases exponentially with time will necessarily tend to an equilibrium

state. On the other hand, the experiments generally require a different
interpretation concerning growth. Taylor (1923, p. 331) states that
“ The speed of the motor [driving the cylinders] was ... gradually increased

till instability occurred’’ and (p. 342) that “ A moderate increase in the
speed of the apparatus merely increased the vigour of the circulation in the
vortices...”. It seems, therefore, that, as the Taylor number is raised,
instability sets in at the critical Taylor number with a definite wavelength.
Further increase of the Taylor number entails the flow retaining the same
wavelength, but with the amplitude of the disturbance motion given by
(3.23). Consequently, it seems that the non-linear stability problem of
growth at a given Reynolds or Taylor number, as treated herein, does not
correspond directly to the circumstances of Taylor’s experiments whereas
the treatment of the equilibrium flow presumably does.
 O n the non-linear mechanics of hydrodynamic stability 21

The work described in this paper was carried out partly in the
Aerodynamics Division of the National Physical Laboratory, and partly in
the Mathematics Department of the Massachusetts Institute of Technology
under contract with the U.S. Office of Naval Research. T h e author
particularly wishes to thank Professor C. C. Lin and D r W. V. R. Malkus
for many valuable discussions on the subject of this paper. Some of the
calculations of this paper were performed by Miss S. W. Skan, whom the
author also wishes to thank.

REFERENCES
BURGERS, J. M. 1948 Adv. Appl. Mech. 1, p . 17. New York : Academic Press.
CHANDRASEKHAR, S . 1953 Proc. Roy. Sac. A, 216, 293.
GOLDSTEIN, S . (Ed.) 1938 Modern Developments in Fluid Dynamics. Oxford :
Clarendon Press.
HOPI>,E. 1956 Proceedings of the Conference on Differential Equations, University
of Maryland (1955), p. 49.
LANDAU, L. 1944 C. R . (Doklady)Acad. Sci. U.R.S.S.44, 311.
LAUFER, J. 1950 Nut. Adv. Counc. Aero., Wash., Tech. Note no. 2123.
LIN, C. C. 1955 The. Theory of Hydrodynamic Stability. Cambridge University
Press.
MALKUS, W. V. R. & VERONIS, G . 1958 J. Fluid Mech. 4 (in the press).
MEKSYN, D. & STCJART, J . T. 1951 Pror. Roy. Sac. A, 208, 517.
NOETHER, F. 1921 Z . angew. Math. Mech. 1, 125.
PELLEW, A,& SOUTHWELL, R. V. 1940 Pro(. Roy. SVL.A, 176, 312.
KEICHARUT, H. 1938 Natuvwissenschafteti 26, 404.
SOROKIN, V . S . 1954 Prikl. Mat. iMekh. 18, 197.
STUART,J. T. 1956 a J. Aevo. Sci. 23, 86 & 894.
STUART,J. T. 1956 b Z . angew. Math. Mech., Sonderheft (1056), p. 932.
TAYLOR, G . I . 1923 Phil. Trans. A, 223, 289.
TAYLOR, G. I . 1936 Proc. Roy. Sac. A, 157, 546.
THOMAS, L. H . 1953 Phys. Rev. 91,780.
TOWNSEND, A . A. 1956 The Structure of Turbulent Shear Flow. Cambridge
University Press.
https://doi.org/10.1017/S0022112058000276 Published online by Cambridge University Press