J . Fluid Mech. (1967), vol. 29, part 1, p p .

17-30 17
Printed in Gveat Britain

On the oscillatory instability of a differentially

heated fluid loop
By P I E R R E W E L A N D E R
University of Goteborg and
Woods Hole Oceanographic Institution

(Received 8 June 1966 and in revised form 24 October 1988)

A theoretical discussion is given of the motion of a fluid contained in a tube form-

ing a closed loop that is heated from below and cooled from above. The fluid is
assumed to have uniform temperature over each cross-section, and the heat
transfer is assumed proportional to the difference between the local temperatures
of the fluid and the tube. The latter temperature is prescribed. The system has
one steady solution with warm fluid rising in one branch and cold fluid sinking in
the other. This solution may, however, become unstable in an oscillatory manner.
A weak instability takes the form of pulsations, the motion being always of one
sign, while a strong instability takes the form of oscillations with zero mean
motion. These oscillations are irregular and do not repeat themselves even over
very long times.
These unstable motions are associated with thermal anomalies in the fluid that
are advected materially around the loop. The anomalies amplify through the
correlated variations in flow rate. A warm pocket of fluid creates maximum flow
rate going through the upper part and minimum flow rate going through the lower
part of the loop. Accordingly it passes quicker through the heat sink than through
the heat source, and the latter becomes more effective. Similarly, the heat sink
acts more effectively on a cold pocket of Auid.
The curve of neutral stability is worked out as a function of the two para-
meters of the problem, a non-dimensionalgravity and a non-dimensional friction
coefficient. The instability has also been studied by direct numerical time integra-
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

tion of the model equations.

It is suggested that the mechanism of instability found for this model operates
also in more complicated systems, and can explain the pulsative type of motions
observed recently in certain convection experiments.

1. Introduction
The problem of general thermal convection (fluid motion set up by heating and
cooling processes) is so far only incompletely understood. Most attention has been
paid to the case of BBnard convection (convection between two horizontal
boundaries kept at different temperatures). The stability problem and the range
of steady laminar convection is relatively well explored, while less is known about
the regimes of unsteady laminar convection and turbulent convection occurring
at high Rayleigh numbers. Only a few studies have been made of convection
2 Fluid Mech. 29
 18 Pierre Welander
created by more complicated distributions of heat sources and sinks, as occur in
many geophysical applications.
Restricting the attention to convection in a finite region, one finds that any
fluid particle carries out a periodic or quasi-periodic motion between the top and
the bottom, being heated a t the lower portion and cooled at the upper portion of
its orbit. It is suggested that insight into the mechanism of such a convection can
be obtained by considering a very simple model: a one-dimensional fluid moving
along a given closed loop and subjected to given heat sources and sinks along its
path. The model can be materialized as follows: take a narrow tube of uniform
cross-section and form it into a closed loop. The tube is filled with fluid that is kept
well mixed over the cross-section (the characteristic time for the mixing should be
small compared with the time required for the fluid to be advected an appreciable
distance along the loop). If the fluid is turbulent the transverse mixing is usually
strong enough but if the fluid is laminar one may have to introduce an artificial
‘diffusor’ in the tube. The tube walls are kept at a prescribed temperature that
varies along the loop. The heat transfer to the fluid can be assumed proportional
to the difference between the local temperature of the wall and the fluid.t
The fluid motion will be driven by the buoyancy force; it will also be retarded
by a frictional force. One can with good approximation introduce a frictional
force that is a function of the instantaneous flow rate; this is justified by the
assumption that momentum like the heat is diffused over the cross-section in a
time short compared with the ‘advection time ’.
The model described will have one solution of steady flow, whatever arrange-
ments of heat sources and sinksis made. This can be proved as follows. The buoy-
ancy force directed along the tube, integrated around the loop, is
B = Agp,a$Tdz,
where A is the cross-sectionalarea, g the acceleration of gravity,p, a meandensity,
LX the thermal expansion coefficient, T the temperature and z a, vertical co-
ordinate. I n a steady state, this force is balanced by a tangential friction force F ,
integrated along the tube. By assumption, F is a function of the flow rate q, F(q).
F(q) obviously increases with q, for any normal friction law. A typical curve
F = F(q)is drawn in figure 1 a. For any value of q one may further compute the
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

temperature distribution in the fluid using the assumed heating law and from
this obtain B(q).The relevant equation is (p/A) (8Tlas)= k(T,(s)- T), where s is
a co-ordinate along the tube, T,(s) the wall temperature and k a constant. It is
easy to show, and it is physically obvious, that when q becomes large the varia-
tions in T along the tube become small. Further, when q = 0 one has T = Toand
B is non-zero (except in the case when $Todz = 0, but then zero motion is one
solution). Obviously the two curves must cut at least in one point; this represents
a possible steady solution. The present proof assumes that the Boussinesq
approximation is valid, but it seems possible to generalize this.
t One may feel that the fluid should take on exactly the wall temperature if the trans-
verse mixing is instantaneous. However, in any experiment one finds a thermal resistance
in a thin boundary layer a t the wall, or in the wall itself. The heat flux can be assumed
proportional to the temperature drop over this region. The inner core of the fluid can still
be considered well mixed.
 Oscillatory instability of heated Jluid loop 19
Looking on the steady (non-zero)convection it seems likely, at the first glance,
that i t is stable. If the flow rate should increase above its balanced value,
the buoyancy forces would decrease and friction would increase, thereby counter-
acting the change. I n the case where zero motion is a steady solution, one can
obviously get instability. It is easy to show that, in this case, instability occurs
f

(4 (b)
FIGURE 1. Variation of total friction P and total buoyancy B with flow rate;
( a )n o equilibrium, ( b ) one unstable equilibrium. (Schematic picture.)

when B'(q)is larger than P'(q),at q = 0. There exists, however, another steady
solution (in fact, two solutions as seen from figure 1b). It looks,therefore, as if the
system would always end up in a stable situation. Some observations of a labor-
atory tube model suggested, however, that all steady solutions may eventually
become unstable when the buoyancy force is made large. To test this idea some
numerical experiments were run on the computer General Electric 225 at the
Woods Hole Oceanographic Institution. The model studied was the simplest
possible: a loop with two long vertical branches and a point heat source applied
at the lower end, a point heat sink at the upper end. The computations demon-
strated that the steady motion becomes unstable, and growing pulsative
motions start when a non-dimensional gravity parameter is large enough, for a
certain range of a frictional parameter. The instability was later demonstrated
analytically; it was found to be associated with infinitesimal perturbations. The
explanation in physical terms could also be given.
Instabilities of pulsative type have been reported for the BBnard convection
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

at Rayleigh numbers around lo6 by Thomas Rossby and for convection be-
tween rotating cylinders by Howard R. Snyder (private communications). In
studies of theoretical models of the oceanic circulation driven by a non-uniform
horizontal heating, there have also been reports on pulsative motions (Kirk
Bryan, private communication), but these cases seem more complex and there is
probably some more mechanism of instability involved. A further theoretical
study of the pulsative instabilities for a real two- or three-dimensional fluid
model would certainly be of interest and clarify the phenomenon further.

2. Derivation of the Model Equations

Consider the fluid in a portion of a tube with uniform cross-section area A and
length L (figure 2). The fluid is driven by the pressure difference between the end-
points and by a buoyancy force, and is retarded by a frictional force. The following
assumptions are made.
2-2
 20 Pierre Welander
(i) The Boussinesq approximation is valid.
(ii)The tangential friction force on the fluid is proportional to the instantaneous
flow rate q.
(iii) The temperature of the fluid is uniform over each cross-section.
(iv)The heat flux between the tube and the fluid is proportional to thedifference
between the wall temperature T, and the fluid temperature T. Tois prescribed
along the tube.

!i
FIGURE
2. Tube segment.

The equations of motion for the fluid are

where v is the fluid velocity, po a standard density, p the actual density, g the
acceleration of gravity, and F the friction force. The density and temperature
have a relation of the form
p = -p,aT+constant. (3)
Taking the tangential component of ( l ) ,one has
p,(dv/dt), = - (aP/as)-pg cos 4 + E7
and, integrating over the fluid volume,

Poat
al/v,dAds = -AZ(p’-p)-g
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

where p’ -p is the pressure difference between the end-points, and q5 is the angle

I
between the tube and the vertical.
We introduce the flow rate q = v,dA, which is constant along the tube in view
of ( 2 ) ,and replace p by T using (3).Under the assumptions (ii) and (iii) the above

1 s
equation can be written
poLq = -AZ(p‘-p)+Agpoa Tdzfconst. dz-p,LRq, (4)

where R is a frictional coefficient.Use has been made of the relation cos q5 ds = dx

where dx is a vertical increment.
A second equation is given by the heat flux condition
dT aT q a T
- = -+-- = Ic{To(~)-T}.
dt at A as
 Oscillatory instability of heatedsuid loop 21
We note that R and k both have the dimension of time -1; these times charac-
terize the viscous and thermal dissipation, respectively.
For a closed tube, equation (4)takes on the simple form

since p' = p and fdz = 0.

" I
@=zga Tdz-Rp (4a)

I n the following we are going to restrict ourselves to the simple model shown in
figure 3. The tube consists of two vertical branches of lengths &Lwith short con-
nexions a t the top and the bottom. The tubes are assumed insulated (k = 0 )
everywhere except at the top and bottom, where over short distances s the tem-
perature of the wall is kept a t -AT and + A T , respectively. We consider actually
the limiting case when As+O while k + m , in such a way that the heat flux
remains finite (point source and point sink).

-A

FIGURE
3. The simplified 4. Co-ordinate along the tube,
FIGURE
model. and antisymmetry condition.

The tube is symmetric with respect to the vertical. It is easy to see that by
starting from a temperature distribution that is antisymmetric with respect to the
centre of the loop, such a state of antisymmetry is always retained, see figure 4.
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

One can, in fact, prove (not shown here) that starting from any temperature
distribution the antisymmetric state is approached asymptotically at least in the
case when q has one sign, so that each part of the fluid passes repeatedly through
the heat source and heat sink.
It will, therefore, be natural to restrict the attention to this case. In the equa-
tions we let s vary in the interval from 0 to $L, and use the antisymmetry to
include contributions from the interval gL to L. The equation of motion is

'y
q. = - T ds - Rq.

The temperature of the fluid coming out from the heat source or heat sink is
determined by the temperature of the ingoing fluid and the flow rate only. It is
easily seen that, if a fluid particle passes the heat source during the short time At,
 22 Pierre Welunder
the difference in temperature between the outgoing and incoming fluid is, to the
first order,
Tout- Tin = (AT - Tin)(1- eckAt)
= (AT - Tin)(1- e-kaAs/lql). (7)
For small flow rates Tout= AT; for large flow rates Taut= Tin. For the heat sink
the same formula holds with AT replaced by - AT.
For the computation in the range 0 < s < &Lwe need to know Tin at the heat
source, when q > 0, and at the heat sink, when q < 0. By the antisymmetry we
have in the first case Tin = - q=&(L-As),
and in the second case Tin = - T,+,. I n
the limit when As+ 0 we denote these values simply - q=hL and - T,=,, respec-
tively.
Finally, the equations are non-dimensionalized by the transformation
L L
s+ -. s,
2
t+
2k As'
~ t, q+kAAs.q, T-+AT.T,

and take then the form

(8Tja.t)+q(aT/as)= 0,
T,=,+ T,=, = (1 + T,=,) (1- e-l'*) for q > 0,
T,=,+T,=,=(-l+T,=,)(l-el~*) for q < O .
The parameters of the problem are
a = gaATL/2(kh~)~
e ,= RL12kAs.

3. Steady motion
We assume that q > 0 ; hence the motion is upward in the branch 0 < s < 1.
I n steady state the temperature in this branch is uniform ;its value is denoted
by T.The other branch has the temperature - T.Denoting the flow rate q, (9) and
(11 a ) take on the form
eq = U P ,
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

2T = (1 + T )(1- e-lj?).
Eliminating ? one gets

It is easily seen that (15)has a single, positive solution. Values ij and T are given
in table 1 for some different values of uje.
Obviously the temperature becomes more uniform along the loop as the flow
rate increases. For very small flow rates the temperature of the rising and sinking
motion approaches the temperatures of the heat source and heat sink, respec-
tively. The case of no motion is degenerate, because of the assumption of insulated
branches. Any temperature distribution that satisfies the condition $ Tdx = 0
and that takes on the temperatures + 1 and - 1 a t the heat source and heat sink
will give an equilibrium solution. It is found that this equilibrium solution is
 Oscillatory instability of heated Jluid loop 23
unstable, at least when the initial temperatures lie between - 1 and 1. Physically,
this is seen by making a small displacement of the fluid from the equilibrium,
say, in the positive direction. The fluid coming out from the heat source (or the
heat sink) must necessarily have a higher (lower)]temperature than the fluid
that enters, and a positive buoyancy is created. When the branches are insulated
this added buoyancy is conserved and the system will accelerate in the positive
direction.

ale
-
q T
0.100 0.100 1.000
0.500 0.417 0.834
1.000 0.648 0.648
2.000 0.958 0.479
10~000 2.218 0.222
TABLE1. Values of non-dimensional flow rate and
temperature in steady motion

4. Stability of steady motion

We consider small deviations from the steady state, putting

q = ij+q’, T = T+T’, (16)

where q‘ is a function oft, and T‘ is a function o f t and s. ij and are assumed
positive.
The linearized forms of (9) and ( 10) are

+
q’ Eq’ =a s: T’ds,
(17)

while the boundary condition (1 1 a ) takes on the form

+ +
Ti=,, mTL==, nq’ = 0,
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

where

using the relation (14) between and F .

Introducing a time factor by putting
q’ = q*eTt, T’ = P(s)eri,

the equations for Q and ?(s) are

1
(r+s) 4=a/ Pas,
0

TP+ q(dP/ds) = 0,
with the boundary condition
P ( 0 )+ m P ( l ) +ng = 0. (24)
 24 Pierre Welunder
The solutions for $and $ are of the form
9 = c exp ( - rs/q),

and inserting these in the boundary condition (24) gives the characteristic equa-
tion
1+mexp( - r / ? j ) + +
naq
r(r e)
~(1 - exp( --r/q)}= 0.

FIGURE 5 . The function FIGURE

6. Tho function - u cot +w.
il(T) = (l-T-z)ln(l-T)/(l+T).

For r real and positive, all the terms in the left-hand side of (26) are positive;
hence no unstable solutions with exponential growth exist. There may still
exist solutions that are unstable in an oscillatory way. The neutral oscillations
are obtained by putting r = i w . In this case, after introducing the new parameters
a" = na/g, .? = e/q, 0 = WIG, (27)
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

the characteristic equation (26) can be written

Separating in real and imaginary parts, expressing a" in terms of T and 2: only,?
( + T ) ,the following two equations result:
and inserting m = (1 - T ) / 1
02+(E-&4)2 = IAY,
(z (29)
2. = - ( l / q o c o t g B , (30)
- 1-Tz 1-T
where A = A ( T )= --In----.
T2 1+T
f Note tthe relations
 Oscillatory instability of heated Jluid loop 25
We note that A > 0, since 0 < < 1. The behaviour of A ( T )is seenin figure 5.
In an (6,E)-plane (29) represents a circle with centre at &Aand diameter A . The
curve represented by (30) is shown in figure 6 (for = 1). When T is small,
which means large A , the two curves must obviously intersect, and two neutral

-.
0

Figure 7. Construction of points on the neutral curve.

https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

0 0 8 o Stable
0 0 0 0
0 g I
>
0 5 10

FIGURE 8. The neutral curve in a @, €)-plane,computed from the stability theory. The
cases run in the numerical integrations are indicated by an open circle (stable) or a solid
circle (unstable).
 26 Pierre Welander
solutions exist (figure 7). Thelimiting case when the two curves first make contact
occurs for T = 0.265. It is the largest for which the flow can become unstable,
and corresponds to a smallest possible flow rate ?j = 1.84. The angular frequency
of the oscillation is 3.60q. Since the angular frequency of a fluid particle is nij, the
thermal disturbance thus travels slightly faster than the fluid.
As the flow rate is increased above the critical value there is one upper and one
lower value of e for which neutral oscillations occur ; in the range between these
the solutions are amplified. The curve in the (4,€)-plane representing neutral
oscillations has been computed graphically, from the construction shown in
figure 7. The result is shown in figure 8. I n the asymptotic limits the upper and
lower values of B are 4?j2and an2,respectively. This result is easily found, noticing
that the two intersections between the circle and the - 6 cot +6 curve approach
0,= 277, Z, = A and 9, = 77, d, = n2/Awhen the radius of the circle gets large, and
that A itself has the asymptotic expression 4q2.
The neutral curve constructed corresponds to the first mode, in which the
temperature disturbance shows one wave along the loop. For larger ij values one
finds also higher modes with two, three, etc., waves along the loop. These come
from intersections of the circle with the different branches of the -6cotfrD
curve. It is easily shown that the corresponding neutral curves all are enclosed
by the curve for the first mode.
The present stability problem does not change qualitativelyin the more general
case where the frictional resistance is a non-linear function of q, as is expected in
turbulent flows. is, however, now a function of ?j, and this leads to a certain
distortion in the previous neutral curve.

5. A numerical experiment
Studies of the non-linear oscillations were made by numerical integration of
the equations (9), (10) and ( I l a ,b). These integrations were carried out on the
computer General Electric 225 at the Woods Hole Oceanographic Institution.
The flow rate q was directly obtained in curve form, by use of a machine plotter,
while the temperature field was printed only at certain times by command of the
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

operator. The computation over one cycle required about 1 see without printing,
and about 1min with printing at every time step.
For the integration the loop was divided in 16, or sometimes 32, equal segments,
each having a uniform temperature. Starting with zero temperature everywhere
except in the first segment next to the heat source, where the temperature was
assumed to be 1,the acceleration was computed by use of (9). The time required
to advect the fluid one segment forward along the loop was then found, keeping
the acceleration constant, and the temperature pattern was advanced one step.
The temperature of the fluid segment passing the heat source could be estimated
from a knowledge of the incoming fluid temperature and the mean flow rate
over the step, using (I 1 a, b).
The results of four different runs are shown in figure 9 a-d. I n the first case the
system is damped and the approach to the steady state is aperiodic. I n the second
case the system is still stable, but the approach to the steady state is oscillatory.
 Oscillatory instubility of heated Jluid loop 27
I n the third case we are very close to a neutral oscillation. A slight change in the
parameter values can change the solution to one that slowly amplifies. The
amplitude will be limited by non-linear effects and a steady pulsation develops.
The temperature field over one cycle in a pulsation is shown in figure 10,
while the variation of the flow rate and the total buoyancy $ T dz is shown in

{L
figure 11.

(4 , ,
t
0 5
P

2T

>t

FIGURE 9. Result of numerical integration in four cases: ( a )a = 0.4,E = 0.2;

( b ) a = 2.0, B = 1.0; ( c ) a = 20.0, E = 3.0; ( d ) a = 40.0, E = 6.0.
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

When the instability is made stronger a new phenomenon occurs that is

exemplified in figure 9d. The non-linear effects cannot limit the pulsations but
these grow until in one ‘back-oscillation’ the flow rate changes sign. The system
then flips over and pulsations build up around a reversed mean flow, etc. One
would perhaps expect to see a periodic behaviour develop, but in the numerical
experiment the oscillations never repeated themselves, even during integration
over several hundred cycles.
A similar ‘ergodic’ oscillation has been discussed in a report by Moore &
Spiegel (1966).I n their case the oscillator was even simpler, consisting of a single
particle elastically restrained in an unstable surrounding fluid. The oscillations,
which could be described by a third-order ordinary differential equation, became
irregular because of the existence of a number of (three-dimensional)limit cycles
that pass very close to each other at the equilibrium point. Even a small numerical
error near this point could shift the system into another limit-cycle and in the
 28 Pierre Welander
course of time a large departure would be found. I n the present case there is no
indication of such a phenomenon. Numerical integrations using different numbers
of segments have been carried out and compared, and it does not look as if
numerical errors cause the irregularity. Rather it seems that the explanation
lies in the large number of degrees of freedom of the system. With a continuous

1 2 3 4

/-‘. -,
,

\ ,“ I ’
5
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

1-1 - 1.0 to - 0.5 0 to + 0.5

- 0.5 to 0 + 0.5 to 1.0
FIGURE
10. Variation in the temperature field over one cycle in an almost
neutral oscillation (a = 20.0, E = 3.0).

distribution of temperature along the loop the number of degrees of freedom is, of
course, infinite. I n the actual integrations the number is finite, proportional to
division in segments, but this number is so large that any periodic or quasi-
periodic behaviour may come out only after a very long integration.
Using the numerical results it is possible to test the prediction of the linear
stability theory. I n figure 8 points are shown representing the numerical cases
 Oscillatory instability of heated Jluid loop 29
in a (4,€)-plane. One sees that the transition from stable to unstable solutions
occurs near the theoretical neutral curve, but that a certain increase of the
unstable region is present. This may be an effect of the difference approximation
and with a more accurate numerical method the agreement could probably be
improved.

FIGURE
11. Variation in the total buoyancy ((I) and flow rate ( b ) in the same case as re-
presented in figure 10.

6. Physical explanation of the instability

I n a steady motion one finds viscous and thermal dissipation that seem to
oppose any change in the flow rate. Any increase in flow rate, for example, would
cause an increase in friction and a decrease in total buoyancy, because the
heating/unit length of fluid column is diminished. One may therefore think that
a steady motion would always be stable. However, these two restraining effects
may not be in phase and an overshooting can then occur, eventually producing
growing oscillations.
I n the present case the mechanism of instability is made clear by figures 10
and 11. One sees in figure 10 a positive thermal anomaly, or ‘warm pocket ),that
emanates from the heat source and follows the fluid up along the right branch
(stages 2-5). It then passes the heat sink (stages 6-7) and can still be traced in the
fluid running down the left branch. Passing the heat source it again appears
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press

amplified. I n the opposite position one sees similarly a ‘cold pocket’ that is
regenerated each time it passes the heat sink. The variation of buoyancy and
flow rate shown in figure 11 explains this regeneration. As the ‘warm pocket ’
comes out from the heat source, positive buoyancy is built up and the flow
accelerates. When the pocket passes the heat sink the flow rate is maximum
(stage 6). Thus the effect of the heat sink is minimized. Half a cycle later when the
‘ warm pocket ’ passes the heat source, the flow rate is minimum and the heating
can regenerate the anomaly effectively. A similar reasoning holds for the ‘cold
pocket ’. One may see the mechanism most simply by considering the fluid as a
pendulum, with its mass centre towards the ‘ cold pocket ’. The heat source and
sink will be most effective when the pendulum is in its upper slower motion, and
less effective when the pendulum is in its lower more rapid motion.
One also wants to understand why, for a given mean flow rate 4 , oscillations
occur only for a certain range of the friction parameter E, excluding the very
 30 Pierre Welunder
small and very large values. We note that, for a fixed ?j, the limits of small and
large E represent also the limits of small and large values, respectively, of the
buoyancy parameter a (?j is a function of e/u alone).
In the limit of large e one finds that the acceleration term in (9) is small, and
buoyancy and friction balance. The lag between the buoyancy and flow shown in
figure 11is no longer present. When the warmest fluid reaches the cold source
and the buoyancy decreases (stages 5-6) the velocity immediately drops. The
effectiveness of the instability mechanism discussed earlier is therefore dimin-
ished. I n the limit of a small E , and thus a small a, (9)requires, on the other hand,
that ?j is almost constant. Thus the fluid has a large inertia. The flow can build up
slowly under the influence of a steady buoyancy force. However, it will not react
much to buoyancy variations in one cycle, and again the effectiveness of the
instability mechanism is diminished.

The author is indebted to Dr Joseph Keller for valuable discussions on the

problem. Dr Keller has looked further into the problem of the non-linear oscilla-
tions, in the limit of small inertia. His investigation is presented in another
article (Keller 1966).Miss Nancy Lockwood and Mrs Jacqueline Webster kindly
helped to set up the numerical problem for the computer. The work has been sup-
ported by the Office of Naval Research through Contract Nonr-2196(00) and
represents contribution number 1810 from the Woods Hole Oceanographic
Institution.
REFERENCES
KELLER,
J. 1966 J.Fluid Mech. 26, 599-606.
MOORE,
D. & SPIEGEL,E. A. 1966 A thermally excited non-linear oscillator. Astrophys. J .
143, 811-887.
https://doi.org/10.1017/S0022112067000606 Published online by Cambridge University Press