Download as pdf or txt
Download as pdf or txt
You are on page 1of 64

Reports on Progress in Physics

REVIEW ARTICLE You may also like


- Log filtering method based on user
Magnetoconvection behaviors
Nan Wu, Xueming Tang and Ying Pan

To cite this article: M R E Proctor and N O Weiss 1982 Rep. Prog. Phys. 45 1317 - Behavior modeling method and its
application in combat simulation
Wang Shaoqin, Xu Bingli, Niu Shulai et al.

- Design and Implementation of Trusted


Plug-in Based on Kylin Operating System
View the article online for updates and enhancements. Platform
Yong Gao and Xuejun Yu

This content was downloaded from IP address 155.198.90.122 on 24/02/2023 at 10:21


Rep. Prog. Phys., Vol. 45,1982. Printed in Great Britain

Magnetoconvection

M R E Proctor and N 0 Weiss


Department of Applied Mathematics and Theoretical Physics, University of Cambridge,
Cambridge CB3 9EW, UK

Abstract

In the outer layers of the Sun and other late-type stars thermal convection is affected
by the presence of magnetic fields. This review deals principally with the interaction
between convection and an externally imposed magnetic field in a Boussinesq fluid.
Further effects of compressibility are considered briefly at the end. Magnetoconvec-
tion exhibits a particularly rich variety of behaviour when the ratio of the magnetic
to the thermal diffusivity is small; the governing equations then allow both steady and
oscillatory solutions. Our account provides a unified description of linear theory and
of the interesting new results that are found in the non-linear regime. Recent work
has classified the behaviour of isolated flux tubes and the relationship between
time-dependent and steady solutions. Convection in a magnetic field is now the
best-studied example of double-diffusive convection and serves as a guide to the
behaviour of related systems.

This review was received in October 1981.

0034-4885/82/111317 +58$08.00 @ 1982 The Institute of Physics 1317


1318 M R E Proctor and N 0 Weiss

Contents
Page
1. Introduction 1319
2. Boussinesq magnetoconvection 1320
2.1. The governing equations 1320
2.2. The boundary conditions 1323
2.3. Special geometries 1323
3. Kinematic theory 1324
3.1, The magnetic Reynolds number 1324
3.2. Flux ropes and flux expulsion 1325
3.3. Hexagonal cells 1327
4. Oberbeck convection 1328
5 . Rayleigh-BCnard convection: linear stability theory 1331
5.1. The eigenvalue problem 1331
5.2. Bifurcations from the static solution 1332
5.3. The influence of the horizontal wavenumber and the boundary
conditions on stability 1334
6. Two-dimensional convection: modified perturbation theory 1335
6.1, Steady convection 1335
6.2. Oscillatory convection 1337
6.3, Behaviour of oscillatory solutions when r(') is close to r ( e ) 1337
6.4. Subcritical steady convection 1341
7. The transition from oscillatory to steady convection in a truncated
model problem 1344
7.1. The modal equations 1344
7.2. The stability of steady solutions 1345
7.3. Oscillatory solutions 1346
8. Two-dimensional numerical experiments 1350
8.1. The development of non-linear magnetoconvection 1350
8.2. Spatially asymmetrical solutions 1357
8.3, Preferred length scales for convection 1358
9. Axisymmetric convection 1361
9.1. Steady solutions 1361
9.2. Numerical experiments 1362
10. The transition from a kinematic to a dynamical regime 1365
10.1. The peak field strength 1365
10.2. A boundary layer model 1367
11. Further effects 1369
11.1. Horizontal fields 1369
11.2. Modified boundary conditions 1372
11.3. Rotation 1372
11.4. Magnetic buoyancy 1372
11.5, Compressibility 1373
12. Astrophysical implications 1375
References 1377
Magnetoconvection 1319

1. Introduction

Cosmical magnetic fields are generally associated with turbulent motion, which may
be driven by convection. Late-type stars like the Sun have deep outer convective
zones, where their magnetic fields are generated, and much of the theoretical work
on turbulent magnetic fields has been related to dynamo theories of their origin. In
this review we shall be concerned with the more restricted problem of convection in
an externally imposed magnetic field. There are two aspects of the interaction between
magnetic fields and convection: on the one hand, the motion sweeps magnetic flux
aside and concentrates it into isolated tubes or sheets; on the other, the Lorentz force
affects, and may suppress, the pattern of convection. At the surface of the Sun,
magnetic fields have a highly intermittent structure. Most of the flux is confined to
tubes in which the fields are intense; in larger features, such as sunspots, heat transport
is partially inhibited by the magnetic field.
The nature of magnetoconvection depends crucially on the ratio f = T / K of the
magnetic diffusivity T to the thermal diffusivity K . When f is large, as it is in laboratory
experiments or in the Earth's core, convection arises first as steady motion and the
main effect of the field is to reduce the heat transport. In the astrophysical context
5 is usually very small (owing to radiative heat transport) and that regime can only
be studied theoretically. In this review we concentrate principally on the idealised
problem of convection in a Boussinesq fluid permeated by an initially uniform vertical
magnetic field, with f < 1; linear stability theory (Thompson 1951, Chandrasekhar
1952) shows that in this case convection first sets in as growing oscillations (closely
related to trapped slow magnetoacoustic waves) which carry relatively little heat. In
what follows, we shall discuss at some length the development of non-linear periodic
solutions and the transition from time-dependent to steady finite-amplitude convec-
tion. Apart from its relevance to astrophysics, magnetoconvection is probably the
most thoroughly studied example of double-diffusive convection, and many of the
results described here have analogues in problems such as thermosolutal convection
or convection in a rotating system.
Since linear theory was extensively discussed by Chandrasekhar (1961) this review
is slanted towards more recent, non-linear results. Developments in bifurcation theory
and in the application of asymptotic techniques have made it possible to describe the
development of finite-amplitude behaviour in the neighbourhood of bifurcations from
the static conducting state. At the same time the availability of larger and faster
computers has facilitated exploration of the non-linear regime by a systematic series
of numerical experiments. These advances have led to a fairly complete understanding
of non-linear Boussinesq magnetoconvection in simplified (two-dimensional or axisym-
metric) geometries. The unfolding of the bifurcation pattern, the formation of isolated
sheets or tubes of flux, and their effect on heat transport can all be described.
Magnetoconvection offers a paradigm of the combined use of analytical and numerical
techniques in order to elucidate complex non-linear behaviour. The subject now
stands at a threshold, with the more relevant but much more complicated three-
dimensional, compressible problem awaiting thorough investigation. We shall attempt
to provide a full description of Boussinesq theory in simplified geometries, while
1320 M R E Proctor and N 0 Weiss

pointing the way to treatments of the more difficult compressible and three-dimensional
configurations.
The plan of the review is as follows. In § 2 we derive the equations that govern
Boussinesq magnetoconvection and discuss boundary conditions and simplified
geometries. Then, in 8 3, we treat the kinematic effects of prescribed velocity fields
on magnetic fields, including flux expulsion and the formation of isolated sheets or
tubes of flux. Dynamical effects are introduced in § 4 by considering the simpler
Oberbeck problem (where convection is driven by a prescribed temperature gradient)
and demonstrating the exclusion of motion from the flux sheets. Linear stability theory
for the Rayleigh-B6nard problem is summarised in § 5 and the next five sections are
devoted to non-linear behaviour. Two-dimensional magnetoconvection is discussed
in detail: results obtained by perturbation methods are described in § 6; in 8 7 these
results are extended into the non-linear regime by adopting a truncated model system,
before presenting numerical results for the full problem in § 8. Axisymmetric mag-
netoconvection is described in § 9, and § 10 discusses the transition from the kinematic
regime to one in which the field is dynamically active. In 9 11 we briefly review
extensions of the theory to include more exotic effects, emphasising magnetic buoyancy
and developments in the theory of compressible magnetoconvection. Finally, we
comment on the behaviour of three-dimensional, turbulent magnetic fields and revert
to our original, astrophysical motivation.

2. Boussinesq magnetoconvection

2.1 The governing equations


I

For almost all of this review (until 011.5) we shall adopt the Boussinesq approximation
(e.g. Spiegel and Veronis 1960), so it is important to specify the circumstances in
which this approximation is valid.
Let us consider a horizontally stratified fluid layer of depth d , referred to a Cartesian
coordinate system with the z axis pointing vertically upwards, and let F ( z ) , p ( z ) ,
F ( z ) be suitable horizontally averaged values of temperature T, density p and pressure
p , respectively. Then we can define the corresponding scale heights HT, H,, H p such
that HT = [ - (d/dz) In T(z)]-', etc. The Boussinesq approximation can be justified
if d << min [HT,H,, H,] and the fluctuations in thermodynamic quantities are small
compared to their mean values (Malkus 1964, Gough 1969, Velarde and Perez Cordon
1976). It follows that the Mach number must be small; if, in addition, the AlfvCn
speed is much smaller than the sound speed the continuity equation reduces to

v*u=o. (2.1)
Fluctuations in density affect only the buoyancy term in the equation of motion, which
then takes the form
Du
po-=-Vp+pg+jABfpovV2u
Dt

where the subscript zero denotes a reference value, U is the fluid velocity, g = - g i
and the viscosity pov is assumed uniform. The magnetic field B satisfies
V*B=O (2.3)
Magnetoconvection 1321

and, in the magnetohydrodynamic approximation, the current density j = p i l V A B


(Roberts 1967, Moffatt 1978). Then the evolution of the field is governed by the
induction equation
aB/at=Vr\(uA B ) + ~ V ' B (2.4)
where the electrical conductivity (poq)-' is assumed uniform.
The linearised equation of state for a perfect gas takes the form
Sp - S p ST
(2.5)
PO PO To'
In order to derive the Boussinesq equation of state it is convenient to consider
deviations from a state of hydrostatic equilibrium with
dp/dz = -gpo (2.6)
and a uniform field Bo, and write p = p + Sp, p m = pm+ Spm, etc, where p m = / BI2/2po.
Then we can rewrite equation (2.2) in the form
Du
PO-= -VP'+p'g + p i 1 (B * V)B+ p o u V 2 ~ (2.7)
Dt
where the total pressure P = p +pm. From (2.7) it follows that the fluctuations in the
total pressure gradient are comparable in magnitude with the fluctuations in the
buoyancy force. Hence

Thus the fluctuations in total pressure are small, so that at leading order in (2.5) we
may write

(2.9)

where 6 is the ratio of gas to magnetic pressure in the hydrostatic state. Since we
are considering thermally driven motion, with temperature fluctuations of the order
of AT, we must require that lSp/pol s AT/To. This can be achieved either if p = 0 ( 1 )
and lapmI << pmo (a situation of little interest in the solar context) or if ISp, 1 -pmo and
PAT b T o . Then (2.5) can be rewritten

(2.10)

(Roberts and Stewartson 1977, Spiegel and Weiss 1982). The two terms on the
right-hand side of (2.10) lead, respectively, to thermal and magnetic buoyancy forces,
when this equation is substituted into (2.7). We shall discuss magnetic buoyancy later
(§ 11.4) but for the present we suppose that T o / p h T < <1, so that (2.10) simplifies to
Sp = -p&iST (2.11)
where the coefficient of thermal expansion E = 7';' for a perfect gas (Chandrasekhar

(z)
1961). Thus we can rewrite the equation of motion (2.2) as

- - v - _ aSTg
Du -
_ - + - j1 A B+ vV2u (2.12)
Dt Po
where po and E may be regarded as constant.
1322 M R E Proctor and N 0 Weiss

Finally we have the energy equation. As in the non-magnetic case (Spiegel and
Veronis 1960) this reduces to
DT/Dt=u , V T a d + ~ V 2 T (2.3 3)
where the adiabatic temperature gradient VT,, = g5To/cp,and the thermal conduc-
tivity (c,po) is assumed uniform, with cp the specific heat at constant pressure. For a
layer with top and bottom boundaries at fixed temperatures T I , T2 the potential
temperature difference across the layer is given by
AT =(T2-T1)-gGTod/cp. (2.14)
For model calculations it is convenient to assume that g5Tod << c,AT, so that (2.13)
becomes the familiar heat flow equation:
DT/Dt = K V ~ T , (2.15)
In order to render the equations dimensionless, we scale lengths with the layer
depth d , times with the thermal relaxation time d 2 / K , velocities with K / d and magnetic
fields with IBol. We define T* = ( T - Tl)/AT. Then equations (2.12), (2.15) and (2.4)
become

= -vp * + R T * +~ ~ Q ( AVB ) A B + v 2 U (2.16)

aT"
-+(U * V)T*= V2T* (2.17)
at

_aB_ VA(U AB)=(V'B (2.18)


at
with V U = V * B = 0 , where the modified pressure p * is proportional to S p ' / p o +
a

g5.z ( T o - Tl)/AT. The dimensionless parameters that appear in (2.16)-(2.18) are the
Rayleigh number
H =g5ATd3/~u (2.19)
the Chandrasekhar number
Q = /BoI2d2/popoTV (2.20)
and the Prandtl numbers
U =U/K
(2.21)
s= 77/K

representing the ratios of the viscous diffusivity U and the magnetic diffusivity 77 to
the thermal diffusivity K .
The pressure may be eliminated from the equation of motion by taking its curl,
yielding the vorticity equation

iao 1
- --VA(U
at
ACO) = ~ v T * A ~ + ( Q v A ( J A B ) + v ~ (2.22)
~

where the vorticity o = V A U and J = pd. We note that thermal convection is gener-
ated dnly by horizontal gradients of temperature.
Magnetoconvection 1323

2.2. The boundary conditions


In most of what follows, we shall confine our attention to the following standard
problem: the fluid is contained between horizontal planes held at fixed temperatures,
and Bo = Bo2. The lateral boundaries are vertical, and there is no transfer of mass,
heat or magnetic flux across them. Thus
at z = 0, -
1;T* = 1 , 0 respectively; U 2 = 0
(2.23)
at lateral boundaries: U n* = 0 , n* * VT* = 0, B 6 = 0
where n^ is the unit normal, and 6 2 = 0.
We need additional boundary conditions on U and B . In the absence of a magnetic
field, one might assume either rigid boundaries, on which U = 0, or free boundaries,
on which tangential viscous stresses vanish (Chandrasekhar 1961). When a magnetic
field is present, it is natural to assume that rigid boundaries are perfect thermal and
electrical conductors, so B 2 = B o , n^ A E = 0, where E is the electric field; if, on the
other hand, the boundaries are free, one could match B to an external vacuum
(potential) field tending to Bo as 121+CO (Chandrasekhar 1961). The latter boundary
condition is non-local, however, and almost all work has foilowed
Chandrasekhar in using the idealised boundary condition

~ A B = O 2 =o, 1. (2.24)

This choice of boundary conditions has the appealing consequence that the eigenfunc-
tions of the linear stability problem remain proportional to e x p ( i n m ) with n an
integer (Rayleigh 1916). Even in the non-linear regime the idealised free boundary
conditions lead to solutions that are periodic in z . These dynamically passive conditions
seem appropriate in modelling stellar convection, while the exterior potential field
condition is more suitable for laboratory experiments (Roberts 1967).

2.3. Special geometries


Equations (2.16)-(2.18) can be greatly simplified for highly symmetrical configurations.
In two-dimensional convection BT y^ = U * y^ = 0 and all physical quantities are
independent of y in a Cartesian coordinate system ( x , y , 2). Then the solenoidal fields
U and B may be expressed in terms of a stream function + ( x , z ) and a flux function
( y component of the vector potential) A(x, z ) such that

(2.25)

Then the vorticity o = (0,w , 0) where w = -V2$ and we define J = V A B * y^ = -V2A.


The vorticity equation (2.22) now has only a y component, which is

(2.26)

while the heat equation (2.17) and the induction equation (2.18) become

(2.27)
1324 M R E Proctor and N 0 Weiss

and

(2.28)

where it has been assumed that there is no y component of the electric field. The
idealised boundary conditions on U and B may be rewritten in terms of 4, w and A
as
4 = w = a A / a z = 0 on all boundaries, while A ( 0 , z ) = 0, A ( h ,z ) = 1. (2.29)
In axisymmetric convection B * 4 = U 4 = 0 ; dT*/acp = 0 and all components of U
and B, referred to cylindrical polar coordinates (r, Q , z ) , are independent of Q , where
4 is the unit vector in the azimuthal direction. Then U and B may be expressed in
terms of a Stokes stream function Y(r,z ) and a corresponding flux function x(r,z )
such that

(2.30)

The vorticity o = ( 0 ,w , 0) where w = - ( l / r ) D 2 Y and we define J = V A B . @ =


- ( l / r ) D 2 X . Then (2.22), (2.17) and (2.18) become

(2.31)

(2.32)

(2.33)

where the Stokes operator D 2= a 2 / a r 2 - ( l / r ) a / a r+ a2/dz ’. Once again, the boundary


conditions become
Y=w =ax/& = o on all boundaries
(2.34)
X ( 0 , Z )= 0 x(A,z)=$A’.

3. Kinematic theory

3.1. The magnetic Reynolds number


Before tackling the full dynamical problem, it is essential to establish the kinematic
effects of motion on the magnetic field in a highly conducting fluid. In the induction
equation (2.18), 5 = 0 corresponds to the case of a perfect electrical conductor; field
lines are then frozen in to the fluid (Roberts 1967, Cowling 1976a, Moffatt 1978).
If f # O advection competes with diffusion of magnetic flux, and the outcome is
determined by the value of the magnetic Reynolds number
R m=(V/f) (3.1)
Magnetoconvection 1325

where U is a measure of the (dimensionless) speed of the flow. (In dimensional terms,
R, = LJd/q.) If R, is small, as in most experiments, the field differs only slightly from
B o ; if R, is large, as in most astrophysical and geophysical situations, the field is
substantially distorted by the motion. In the rest of this section we shall discuss the
effect of a layer of persistent eddies on an initially uniform magnetic field.

3.2. Flux ropes and flux expulsion


Within a convecting layer we can isolate regions of converging flow. When R, is
large, magnetic flux is swept into these regions until advection is eventually balanced
by diffusion. Solutions have been obtained for a number of simple patterns of motion
(Parker 1963, 1979, Clark 1964, 1965, 1966, Weiss 1966, Clark and Johnson 1967,
Busse 1975, Moffatt 1978, Proctor and Weiss 1978). The formation of flux ropes is
best illustrated by considering the simple two-dimensional flow given by
*(x, 2 ) = u x z . (3.2)
The field then remains vertical (since dU,/az = 0) and, from (2.28), the flux function
A = A(x, t ) satisfies
aA aA a2A
-= U x - + + 7 . (3.3)
at ax ax
If the field B = dA/ax is initially uniform in the neighbourhood of x = 0, then for
small times (3.3) shows that B increases exponentially with t. At large times (3.3)
admits a steady solution
~ ~( - u x 2 / 2 < ) .
B ( X ) = B , exp (3.4)
If the total flux per unit length in the y direction is finite and equal to 2B&,
Bmax=BoL(2Rm/.rr)’/*.Thus the flux is confined to a sheet of semi-width S =
O ( R , ’ / 2 ) ,and with a maximum field aR;”. The development of this steady state
may be described in terms of the Gaussian similarity solution

(3.5)

where g(t) = (1+ C e-2Lit)-1’2and C is an arbitrary positive constant. We note that


the time taken for the steady state to be established does not depend on 5, but only
on the ‘turnover time’ T~ = U-’.
Within an individual eddy, there are closed streamlines that do not pass through
the magnetic boundary layers where the field is concentrated. In a steady state, for
R, >> 1, it follows from (2.28) that at leading order

(3.6)

so that B is parallel to U. However, integration of the steady form of (2.28) over the
area S bounded by any such closed streamline C yields

(3.7)
1326 M R E Proctor and N 0 Weiss

so that dA/d$ = 0; hence B = 0 everywhere in the region of these closed streamlines


(cf Batchelor 1956). Thus magnetic flux is expelled from the interior of the eddy and
concentrated in regions of converging flow (in the neighbourhood of heteroclinic orbits
of the motion (Childress 1979)). The expulsion of an initially uniform field can be
followed by numerical integration of the time-dependent induction equation. Figure
1 shows the process occurring with a velocity defined by the stream function 4 =
-1
7~ U sin T X sin TTZ, for R , = 250. This value shows the typical behaviour for large
R,; for R,s 50, the field relaxes to its steady value without reconnection of the lines
of force (Parker 1966). The time scale for flux expulsion depends on 6, and is larger
than that for flux concentration: within the eddy, diffusion becomes important after
a time of the order of ( T ~ R (Parker
~ ~ 1963,
) Weiss 1966, Moffatt and Kamkar 1982).
Analogous results hold for an axisymmetric configuration. For a stagnation point
flow given by the Stokes stream function
(3.8)

I'
0
1

6 7 8

9 10 20

Figure 1. Flux expulsion in two dimensions. Successive stages in the distortion of an initially uniform
magnetic field at times r/Tl (values given for each diagram), where T~ = 5 ~ ~ (from
/ 8 Galloway and Weiss
1981).
Magnetoconvection 1327

the field remains vertical and in a steady state


BZ(y,Z ) = Bmax exp (-iRmr2). (3.9)
If the total flux is equal to rB0L2,B,,, = $BoL2R,; the greater amplification is due
to the fact that the flux is concentrated into a tube rather thafl a sheet. Flux expulsion
occurs as for the two-dimensional case (Moffatt 1978). In a closed cylindrical domain
(0 s z s 1,O s r s ro} bounded by a surface S, flux may be concentrated either on the
axis or at the periphery. In the limit R,+ 0;) it can be shown that almost all the flux
is confined to the central tube (Galloway et a1 1978). The steady version of equation
(2.33) may be rewritten
v + (ux) = b~ (r2v(X/r2)).
+ (3.10)
Integrating over the volume of the domain and applying the divergence theorem, we
find that

(3.11)

from (2.34). Now the average field on the boundary r = ro is

(3.12)

since the total flux through the domain can be shown from (2.30) to be 2 7 ~ ~ l ~ = ~ ~
Thus B is unaffected by the motion, and the flux contained in an annular region of
thickness -R,”2r0 near r = ro tends to zero as R,+ W. Thus all the flux must be in
the neighbourhood of the axis. Numerical integration of the time-dependent equation
(2.33) (Galloway and Moore 1979) confirms this result: figure 2 shows an example
taken from a non-linear computation. (Analogous arguments applied to the two-
dimensional problem yield the rather uninteresting result that the average values of
B must be the same on either lateral boundary.)

3.3. Hexagonal cells


Convection in an axisymmetric cell introduces some three-dimensional effects but the
cells cannot be combined to form a tesselated pattern of convection. The kinematic
effect of motion in a layer of hexagonal cells has been modelled by Clark and Johnson
(1967) and Galloway and Proctor (1982). They selected the velocity field proposed
by Christopherson (1940) which takes the form
U 2 7Tx TX
(3.13)

The streamlines of the horizontal flow at z = 1 are shown in figure 3(a), while contours
of the vertical velocity are displayed in figure 3(b). Unlike the cylindrical cell, where
the horizontal velocity is everywhere radial and converges or diverges only at the axis,
the hexagonal cell has streamlines that converge or diverge at the cell corners as well.
Thus one might expect flux tubes to develop both at the centre and at the corners
of the cell. Figures 3 ( c ) , ( d ) and ( e ) show perspective plots of the vertical field at
z = 0, $ and 1, respectively, for R , = 200. It can be seen that a significant proportion
of the flux remains at the centre of the cell even at z = 1, where the flow is radially
1328 M R E Proctor and N 0 Weiss

,/----
----_

Figure 2. Flux expulsion in cylindrical geometry. Contours of ( a ) y, (lines of force) and ( b )4' ' (streamlines)
in a steady state with R, >> 1. Note that almost all the flux is concentrated near the axis (from Galloway
et al 1977).

outward from the centre. Calculations suggest that no more than about 35% of the
flux is located at the corners of the ceil. This ratio obviously depends on the particular
choice of U in (3.13); preliminary calculations suggest that as the downward motion
becomes relatively more vigorous the proportion of flux in the central region increases.

4. Oberbeck convection

We saw in the previous section that convective eddies tend to concentrate magnetic
flux into isolated sheets or ropes. Within these regions the Lorentz force is relatively
strong; unless Bo is very small, flux concentration will therefore be impeded by the
dynamical effects of the magnetic field. To proceed further we have to solve the
equation of motion. This is simplified by considering a variant of the standard problem
in which the temperature distribution is prescribed. For the Oberbeck problem we
suppose that U, 5 and are all very small; in the limit of high thermal diffusion,
(2.17)reduces to

V2T*= 0 . (4.1)
From ( 2 . 2 2 ) , the buoyancy force will generate convective eddies provided that T"
varies horizontally at z = 0 or z = 1.
Magnetoconvection 1329

80
'r LO

(C) (d) (e)

Figure 3. Flux expulsion in a hexagonal convection cell. ( a )Streamlines of the horizontal flow at the top
of the cell (from Clark and Johnson 1967), ( b )contours of the vertical velocity (from Chandrasekhar 1961).
Profiles of the vertical field strength in a steady state with R, = 200 at (c) z = 0 , ( d )z = i, ( e ) t = 1 (from
Galloway and Proctor 1981).

The two-dimensional problem with

T" = 0 at z = 0 T*=cos~xatz=l
(4.2)
T*(x, 2 ) =cos TX sinh m/sinh T

was studied by Peckover and Weiss (1978). The corresponding isotherms are shown
in figure 4(a): the horizontal temperature gradient is a maximum at (i,1) and motion
is driven for all values of the modified Rayleigh number R / a (which is independent
of K). When (R/a)<c 1 and Q is small (the kinematic limit) contours of o and @ for
the resulting flow are as depicted in figures 4(b) and (c). Provided J / u is sufficiently
small (7) << v ) magnetic flux is expelled from the eddy and concentrated into sheets,
as described in 0 3. In the kinematic limit the lines of force are similar to those in
figure 1.
As Q increases the flux sheets become dynamically active and generate vorticity
with the opposite sign to that produced by the buoyancy force. Thus motion is excluded
from the regions of strong magnetic field. Figures 4(e), ( f ) and (g) show contours of
A , o and 1+4for a steady state in this dynamical regime. If Q is further increased the
magnetic forces quench the motion throughout the region and the magnetic Reynolds
number tends monotonically to zero. The effect of increasing Q for a fixed value of
R / a is shown in figure 5 , which shows a sequence of steady-state magnetic fields.
The actual peak field in the flux sheet at (1,l) is proportional to B" = Q1'2Bz(l,l).
-
In the kinematic regime B* RY2Q1'*(cf equation (3.5)) but as Q is increased the
1330 M R E Proctor and N 0 Weiss

la)

(e)

Figure 4. Oberbeck convection with R / u = lo4. Contours of ( a ) T*, ( b ) w , (c) $ in the absence of a
magnetic field. Contours of ( d ) A , ( e ) w , (f) $ for Q = 1600, = &: note the shaded regions where the
vorticity is reversed (from Peckover and Weiss 1978).

10 1 25

Figure 5. Suppression of convection by a magnetic field. Steady field configurations for the Oberbeck
problem with R/u = lo4, [ = 1 and Q values given for each diagram (from Peckover and Weiss 1978).
Magnetoconvection 1331

effectivemagnetic Reynolds number can be approximated by


R,=R,,-CQ (4.3)
where R,, and C are constants. It follows that B*(Q) has a maximum: for the case
illustrated in figure 5 this maximum occurs at Q = 150.
Similar results can be obtained in an axisymmetric geometry, which actually proves
more favourable for asymptotic analysis. We shall return to this question in § 10 in
order to elucidate the transition from kinematic to dynamical behaviour for the more
important Rayleigh-BCnard problem.

5. Rayleigh-Benard convection: linear stability theory

5.1,The eigenvalue problem


The governing equations (2.16)-(2.18) admit the obvious static solution

U =o T"z1-z
B =i i
p=R z--
f, +constant. (5.1)

In order to investigate its stability we consider small perturbations such that lu 1 << 1 and
T * = 1- t + 8 (x, t ) B = i +6(x, t ) (5.2)
(Chandrasekhar 1961). To leading order the vorticity equation (2.22) becomes
1 act, aJ
(+ at - R V @A Z * + L Q - az
+V~~ (5.3)

since J =V A 6. Similarly equations (2.17) and (2.18) reduce to


a6
-=u 'i+V20 (5.4)
at
and

Now the z component of (5.3) is


-
1 -aw,
-=Q-+V
a J2
z ~,
0- at az
while the z component of the curl of ( 5 . 5 ) is
J J , aw,
-=-
at az
+ CV2Jz. (5.7)

These equations describe damped torsional Alfv6n waves: since there are no source
terms, the solutions decay exponentially for any reasonable choice of boundary
conditions. Hence w z and J, are both zero for any unstable perturbation.
From (2.1) and (2.3) it follows that U and 6 are purely poloidal, and we can write
U = v A v A (F(X, t ) ; )
6 = V A V A (fT(x,t ) i ) .
1332 M R E Proctor and N 0 Weiss

Then we may recast equations (5.3)-(5.5) in the form

_
a@-- -V;F + V2@
at
(5.10)

(5.11)

where 57; = a2/dx + a2/ay ', and seek normal modes with
(F,0, H ) = (&z), fi(z))f(.x, y ) exp b * t ) (5.12)
where 0:f = -a2f. (Note that the horizontal planform is degenerate in linear theory;
for any a there are solutions for f that correspond to rolls, axisymmetric rings, square
cells, hexagons, etc.) Then (5.9)-(5.11) become
[(T-~s * - ( d l - a ' ) I ( d - a ')F = -Re^ + 5Q (dz - a ' ) d , f i
[s * - ( d l - a ')]e^ = a 2R (5.13)
[s*-[(d:-a2)]fi=dz#
where dzfi = &/dz, etc.
To proceed further we must specify the boundary conditions. For the idealised
free boundary conditions (2.23) and (2.24) we may set fi, & a s i nn m , f i a c o s n m .
Then the eigenvalues s * satisfy the cubic characteristic equation
P 2 ( s * + p 2 ) ( s * + ~ p 2 ) ( s * + 5 P 2 ) + ~ I Q2 nP 27~ 2 (s * +P2)-Rua2(s*+l'p2)=0 (5.14)
where p 2 = a + n 2 v 2 .

5.2. Bifurcations from the static solution


We first seek the condition for a simple bifurcation: s * = 0. This is traditionally known
as an 'exchange of stabilities'. From (5.14) s * = 0 when R = R ( e )where
,
n2 p 2 ) / a 2 .
R ( e ) = ( p 6 + Q2 7~ (5.15)
It is obvious that for a given value of a', R'" is least when n = 1; this holds for other
bifurcations too, since increasing n is equivalent to decreasing the depth of the layer.
We therefore set n = 1 henceforth. When Q = 0, R ( e = ) R o ,where
R , ( ~=)p 6 l a 2 = ( r r 2 + a 2 ) 3 / a 2 . (5.16)
R o is least when a = rr/JZ; this minimum value, 2 7 r 4 / 4 = 657.51, is the critical
Rayleigh number for the onset of convection in the absence of a magnetic field
(Rayleigh 1916).
It will prove convenient, especially when we come to consider non-linear solutions,
to introduce a more compact notation by defining a normalised Rayleigh number
r = R/Ro. (5.17)
At the same time we define modified parameters r, s and q by setting
r = p 2t s =p-Zs* q = (.rr2/P4)Q. (5.18)
Magnetoconvection 1333

These scalings take up all the geometrical factors so that the characteristic equation
(5.14)simplifies to
s3 + (1+ U + l ) s 2+ [u(l- r + l q )+ I(1+ u)]s + a l ( l + q - r ) = 0. (5.19)
The condition (5.15) for a simple bifurcation is now r = r ( e )where
,.(e) = 1+q, (5.20)
Of the three roots of (5.19), one is always real and negative. The other pair may
be real, or else complex conjugates. In the latter case there is a Hopf bifurcation
from the static solution (when Re (s) = 0, s = hi&) at r = r(') where

(5.21)
provided that

(5.22)

is positive. Hence, if a Hopf bifurcation is possible it always occurs at a lower value


of r than the simple bifurcation. In traditional terms, overstability sets in at r = <r ( e )
(Thompson 1951, Chandrasekhar 1952). It is clear from (5.22) that a necessary
condition for the existence of a Hopf bifurcation is

(5.23)

Now 5 >> 1 for liquid metals (e.g. Cowling 1976a); hence oscillatory convection will
not arise in laboratory experiments or geophysical applications. Within a star, however,
f is typically very small, owing to the high radiative conductivity. In astrophysical
applications, therefore, convection generally sets in as growing oscillatory solutions.
For 4' > 1 or q <go, the only bifurcation from the static solution occurs at r = r(e'
and the static solution is unstable for all r > r ( e ' , If, on the other hand, (5.23) is
satisfied, so that 6 ;> 0, the static solution is unstable for all r > r"). The behaviour
of the eigenvalues in the complex s plane is illustrated in figure 6; as r passes through

Im Is1

I
Figure 6. Behaviour of the eigenvalues when convection sets in via a Hopf bifurcation at R = R'"'.The
arrows indicate the evolution of the eigenvalues s with increasing R in the complex plane.
1334 M R E Proctor and N 0 Weiss

r"), the real part of the complex conjugate pair changes sign. As r is further increased,
Re (s) increases while /Im (s)/decreases. At r = r"), say, Im (s) = 0 and the two roots
coalesce. Thereafter both eigenvalues remain on the real axis; one increases with
increasing r, but the other decreases and passes through zero at r = r(e). In linear
theory, therefore, the transition from oscillatory to direct instability occurs at r = r")
(Danielson 1961, Weiss 1964). The condition for (5.19) to have two equal roots is that
4a12(r- 5 q ) 3 - a 2 M + 2 a ( l - 5 ) L + (1- 5)'(1 -(+)'(a -f)2 =0 (5.24)
where
L = r [ ( ( l + 2 5 - 26') - a ( l + 66 -46') + d 4 - 5 ) -a3]
- 5q[(2-25 -J2) - a ( 4 - 65 - 5') + a 2 ( 1-46) + a 3 ] (5.25)
and
M = r2[-(1 + 8f - 8 f 2 )+ 2 a ( 5 - 45) - a 2 ]- 2r5q[(10- 195 + l o t 2 )-a(l +I) + a 2 ]
+12q2[(8- 85 - f 2 ) -2 4 4 - 5 5 ) -a2]. (5.26)

Values of r ' l ) , if needed, must in general be found numerically. We note, however,


that r'i) depends on [q and that in the limit (q + IX, at fixed a and f
-
r ( ' ) lq. (5.27)
For 6 ;< 0, the only real root of (5.24) is less than r(') and therefore uninteresting.
For a perfect fluid ( K = v = 7 = 0) the dimensional frequency of oscillation R
satisfies

(5.28)

(WalCn 1949, Cowling 1953, 1976a). This shows clearly that the oscillations should
be thought of as trapped hydromagnetic waves. From (5.28), the transition from
neutral oscillations to exponentially growing instabilities occurs when
gEATd=P
2
T
2 2
Bo/kopoa
2
. (5.29)
This is the same result as (5.27). When the diffusivities are small, but their ratios
remain of the order of unity, the real part of the complex eigenvalues remains small.
Only when s is real ( r > r ( i ) )does the instability grow rapidly.
These results may be summarised in slightly different terms as follows: when (5.23)
is satisfied, so that 6 ;>0, overstable convection first sets in at r'"), and there is a
transition (in linear theory) to overturning convection at r ( i ) ;as far as linear solutions
are concerned, r ( e )appears irrelevant. If, on the other hand, 6: < 0, a 'leak' instability
appears first at I " ) , and r ( i )has no physical significance.

5.3. The influence of the horizontal wavenumber and the boundary conditions on
stability
The parameters defined in (5.17) and (5.18) depend on the horizontal wavenumber
a . In practice we wish to determine the values of A T at which bifurcations occur as
functions of a . In order to discuss the influence of horizontal wavelength on the onset
of instability, we therefore have to work in terms of R and Q.
Magnetoconvection 1335

From (5.15), is a minimum when a = a0 where


2a: + 3 r 2 a : = r 4 ( v+2Q) (5.30)

(Chandrasekhar 1961). Thus a C increases with Q ; as Q+m, a , - ( ~ ~ Q / 2 ) ~and ’~,


R (e) = R mln
(e)
-
r2Q. A similar effect occurs for the oscillatory solutions, with the same
power law dependence on Q of both R(” and the critical wavenumber (though the
multiplicative constants are different). Thus, for large Q the Lorentz forces favour
vertically elongated cells with motion predominantly along the field lines. Owing to
the large value of a c ,lateral diffusion becomes extremely effective.
The values of the Rayleigh number for which bifurcations occur depend on the
particular choice of boundary conditions, though the bifurcation pattern is qualitatively
unchanged. Chandrasekhar (1961) obtained values of R“’ for the cases when both
upper and lower boundaries are rigid, and when one is rigid and one free. The former
can be realised in the laboratory, and the experimental results obtained by Nakagawa
(1955, 1957) are in remarkably good agreement with theoretical predictions. When
there is a Hopf bifurcation, treatment of realistic boundary conditions is more compli-
cated. Gibson (1960) investigated the behaviour of R(” at large Q, and found that
it differs only slightly from that found with the idealised boundary conditions (2.24).

6. Two-dimensional convection: modified perturbation theory

6.1. Steady convection


It is known that in the neighbourhood of the simple bifurcation at r ( e )there exist
non-trivial steady solutions corresponding to finite-amplitude convection (see, for
example, Iooss and Joseph 1980). The linear theory presented in the previous section
does not, however, tell us whether these solutions will be stable. We therefore
introduce a small parameter E such that

4=E$1+E242+ ...
8= 4 + e 2 e 2 + ... (6.1)
A = x + E A ~ + E ’ A .~. +
.
and
r =r(ej+Ery) +E2rf’ + .. , (6.2)
(Malkus and Veronis 1958, Veronis 1959). At each stage in the expansion we may
define a column vector

(6.3)

where J,,= -V2A,. If we substitute from (6.1) and (6.2) into (2.26)-(2.28) and set the
time derivatives to zero then at leading order in E the equations are linear and can
be written in the compact form
LY\Ir,=O (6.4)
1336 M R E Proctor and N 0 Weiss

where the self-adjoint differential operator

0
and A = a 2 / p 2 r 2 .This is just the two-dimensional form of the eigenvalue problem
discussed in § 5 : thus r'eJ= 1+ q and the eigenfunction

L a sin a x cos r z J
where the normalisation is chosen for subsequent convenience. (Note that $l CC
- a F / d x , where F is the potential defined in ( 5 . 8 ) . ) Since the operator L is self-adjoint
we have the identity
(yr?LWn)= (yrTLW1) = 0 (6.7)
for all n, where the angle brackets denote averages over the domain (0 S x d 2 r l a ;0d
z s I}.
To second order in E we obtain

LW2 = 1 r(eJa($l, e l ) / a ( x , z )
-Alq'sa(h A l ) / a i x , 2 )
(6.8)

where on= -V2$,. It follows from (6.6) that the two Jacobians in the first row on the
right-hand side of (6.8) are identically zero. The solvability condition on this
inhomogeneous equation is obtained by applying (6.7) to (6.8): it then follows that
r ? ) = 0 (as expected from the symmetry of the problem). If we impose the orthogonality
condition
(yr?Wfi) = 0 (a f 1) (6.9)
then the solution of (6.7) is

1
0
sin
- (l/r) 2rz . (6.10)
( r 2 p 2 / 1 2 a sin
3) 2 a x
At third order we have

Lyr3 = a(41,e 2 ) / a ( x ,2 )

The solvability condition yields, from (6.7) and (6.11),


i' (6.11)

2 r 2 ( a 2- r 2 )
r:"'=l+q+ a2p212 9 (6.12)
Magnetoconvection 1337

a result obtained by Veronis (1959, in a footnote). We note that for f sufficiently


small, rt) is positive or negative according as to whether the cells are tall and thin
( a > T ) or short and squat ( a < T I , respectively.
We are now able to discuss the stability of finite-amplitude solutions when w i < 0.
Then one real eigenvalue of the linear problem increases through zero as r increases
through r ( e ) . It follows that the finite-amplitude solution is stable if r f ) > 0 and unstable
if r f ) < 0 (PoincarC 1885, Joseph 1981). In this case we may describe the corresponding
bifurcations as supercritical or subcritical, respectively. We should note, however,
that when U', > 0 and r(') < r ( e )the relation between r!) and the stability of the steady
solution is quite different.

6.2. Oscillatory convection


If f < 1 and q > qo then there exist finite-amplitude periodic solutions in the neighbour-
hood of the Hopf bifurcation at r(') (Hopf 1942). We assume that the motion has a
period 2 ~ / 6and expand W and r as power series in E so that

(6.13)

(Veronis 1959). From (6.1) and (6.13) it follows that $,, e,,, A,, have the same spatial
dependence as before (though they are now time-dependent) and that 01= r?) = 0.
The evaluation of W 2 and r e ' is inevitably more complicated than that of r f ) : it is,
however, possible to find, as a quadratic equation for Wi,the condition that r?) = 0
(Knobloch et al 1981). The regions where r?) is positive can therefore be delimited
and it turns out that r y ) > O for values of a in the neighbourhood of T (square cells)
though, with small enough or large enough values of a and suitable choices of U and
f , r y ) may be negative.
Hopf's (1942) theorem determines the stability of these oscillatory solutions. If
r?) > 0 the bifurcation is supercritical and the solutions are stable in its neighbourhood;
if r y ' < O the bifurcation is subcritical and the solutions are unstable. To proceed
further one might continue the expansion to higher order in E , though the radius,of
convergence may be very small (min ( 1 , f , U ) ) .
The presence of a Hopf bifurcation at r ( O ) affects the stability of the steady branch
that bifurcates from the static solution at r(e). The steady branch is always unstable
in the neighbourhood of r ' e ) but the behaviour of non-linear solutions, to be discussed
below, depends on the sign of r f ) . We consider perturbations to the (trivial and
finite-amplitude) steady solutions, varying as exp ( S T ) . Only two eigenvalues are of
interest and figure 7 shows the behaviour of their real parts as r passes through r ( e ) .
Note that as r is decreased through r ( e ) one eigenvalue for the static solution increases
through zero, while the other remains positive. It follows that on the finite-amplitude
steady branch one eigenvalue is always large and positive, while the other is small
and has the same sign as r f ) .

6.3. Behaviour of oscillatory solutions when r'"' is close to r ( e )


In general, non-linear time-dependent behaviour has to be followed using computa-
tional techniques. There is, however, one limit when many properties of the oscillatory
branch are accessible to analysis. If q = q o , r ( O ) = r(j)= r ( e ) and there are two zero
1338 A4 R E Proctor and N 0 Weiss

t
E7 lbl

/ \

Figure 7. Growth rates of small disturbances to static and finite-amplitude solutions for ( a )rt) > O and
( b ) r t ) < O . The signs of the real parts of the two relevant eigenvalues are shown (in brackets if the
eigenvalues are complex).

eigenvalues. When
q =qO+E2 (E << 1) (6.14)

say, 6,is small and r ( O ) is close to r ( e ) ,so that

(6.15)

and

(6.16)

Following Knobloch and Proctor (1981) we write

(6.17)

and, instead of (6,1), expand $, e and A as follows:


4 =Ea(T*)$l +o ( ~ ~ )
0= E w ) e +l E 2 C ( 7 * ) e 2 + - ~ ( E 3 ) (6.18)
A = x + E ~ ( *T) A + E ’e ( T * ) A + O ( E’)
where we have introduced a new (slow) time T* = €7 and $1, 81, etc, are defined by
Magnetoconvection 1339

(6.6) and (6.10). On substituting into the governing equations (2.26)-(2.28) we obtain

Eb’ = -b + ~ ( 1E-’ C ) + O ( E ~ )
EC’= a ( - c+ab)+O(E2) (6.19)
= -ld + a - E2ae + o ( E ~ )
E e l = -(4 - w ) l e + w u d + o ( E ~ )

where the geometrical parameter


w = 4.rr2/p2 (6.20)
and the primes denote derivatives. Thus O < w < 4 and w increases as cells grow
flatter, By back-substitution and differentiation, all other variables can be eliminated
in favour of a and (6.19) reduces to
a”-Ma3+MNu =EF(u)+O(E’) (6.21)
with
M = - r f ) u l ( 1+a + N = -(I -- p ) / r Y ) (6.22)
and
F ( a )= a (1 + a + l)-1[Ca2a’
+ Da”‘+ ( 1 - /A[)a’] (6.23)
where
2 l b +t )
C = a(1-l) ( 2 + 3 - a l ( l - l ) 4 - w
-(
(l+a) 2 w
(w - 1)+-
4 - -w2 )
(6.24)
D=l+l/(T+(l+(T+[)/a[.
In the limit E + 0 equation (6.21) reduces to Duffing’s equation, which Kas a
single-parameter family of periodic solutions that depend on the value of the
Hamiltonian
g.4u”- &a + 4MNu ’, (6.25)
For small E , 8 evolves on yet another, slower time scale T’ = E T * = 6’7, according to
the equation
d8/dTt = (a’F(a)) (6.26)
where the average is taken over a period, The critical points of (6.26) give the values
of 8 for which periodic solutions of (6.2) are possible, and the equation can also be
used to determine the stability of these solutions. The periodic solutions of Duffing’s
equation can be expressed in terms of Jacobian elliptic functions. For M, N > 0,

(6.27)

where the parameter m (0 s m < 1 ) is given by


8 = m M N 2 ( 1+ m ) - 2 . (6.28)
1340 M R E Proctor a n d N 0 Weiss

ForM<Oand(m-$NsO
a=(-) 2mN ' I 2 cn((-) MN ' I 2T * / m )
(6.29)
2m-1 1-2m
and
MN'm(1 - m )
g=- > 0. (6.30)
(2m - 1)2
Finally, for M < 0 and N > 0 there is a second solution
~ = ( S ) ~ " d n ( ( - -MN
) ' I 2.;*/m)
(6.31)
m-2
and

(6.32)

For given r?), the parameter m is determined as a function of F by the solvability


condition (6.26). Knowing r?), I.L and m, we can decide which of the three different
types of solution, in (6.27), (6.29) and (6.31), is appropriate.
When r $ ) < O and r ( ' ) s r < r ( ' ) both M and N are positive, from (6.22), and the
oscillations have the form (6.27). The Hopf bifurcation is supercritical and both the
amplitude and the period of the motion become larger as r increases from r"), while
m increases from zero to unity. For small m the solutions of (6.27) are approximately
sinusoidal but as m + 1 they acquire a characteristically flat-topped snoidal form (cf
the finite-amplitude motion of a simple pendulum). Finally, for r = rmax< r ( e ) ,m = 1,
the period becomes infinite and the trajectory degenerates into a heteroclinic orbit
which connects the saddle points that correspond to the two unstable steady solutions
a = *N'/'. The relationship between the oscillatory and steady solution branches is
illustrated in figure 8 ( a ) ; sketches of the solution trajectories are shown in figure 14.
The oscillatory solution branch is quite different when r?) > 0, as can be seen from
figure 8(6) (which shows the typical case with r?) >O). In the neighbourhood of r")
the oscillations are given by (6.29) and remain roughly sinusoidal over the range AS
with 0 s m <$. At S , m = i, F = 1 and so N changes sign; the solution continues
smoothly through S and on to B ( m ~ 0 . 9 3 r, = rmax)where it loses stability. At B
the solution has a characteristically spiky cnoidal form. As m is further increased, r
decreases along the oscillatory branch until, at C, m = 1 and the period becomes
infinite. At this point there is a transition from a degenerate heteroclinic orbit to a
symmetrical pair of homoclinic orbits: that is to say, for larger values of r there is a
unique unstable limit cycle about the origin while for smaller values of r there are
two unstable limit cycles about the steady solutions a = *N ' I 2 . These latter are given
by (6.31) and are of dnoidal form. As r decreases from C to D along the oscillatory
branch m decreases and the limit cycles collapse on to the steady branch at r = rH< rmax.
The steady branch is stable for r > r H and there is a Hopf bifurcation at r H . (The
bifurcation appears to be subcritical, though this cannot be confirmed without proceed-
ing to higher order in the expansion scheme.) For r slightly less than rmin the steady
solution is unstable with two complex conjugate eigenvalues. For r slightly greater
than r ( e ) ,on the other hand, there are two real positive eigenvalues (cf figure 7(a)).
Thus there is some r, r ( e )< r < r H , at which the two real eigenvalues coalesce and then
Magnetoconvection 1341

-
/

0 Y
I I I I I I I
03 05 07 09
P P
Figure 8. Relationships between the branches of steady and oscillatory solutions when r ( O J is close to r ( e J .
( a ) The maximum and the mean of a* plotted as a function of when rf’ < O ; the broken line denotes
the unstable steady branch. ( b ) The same functions plotted for a case where rf’, r?’ > O . The segment
ED of the steady branch EDF is unstable; the unstable portion of the oscillatory branch is indicated by a
broken line (from Knobloch and Proctor 1981).

become complex: proceeding from r ( e )to rH along the finite-amplitude steady branch
gives a succession of eigenvalues like those obtained proceeding from d e )to r(’) along
the curve in figure 6 .
Figure 9 shows a sequence of trajectories projected onto the a‘a plane for five
different values of r. For r < r ( e ) there is an unstable node at the origin which is
transformed into a saddle point for r > r ( e ) ;the two symmetrical steady solutions are
unstable nodes for r close to r ( e ) but become unstable foci before r reaches rH and
stable foci for r > r H . In the range rH < r < rmax there is a pair of unstable (dnoidal)
limit cycles about the steady solutions, or a single unstable (cnoidal) limit cycle about
the origin, as well as the stable cnoidal limit cycle. These sketches summarise the
rather complicated behaviour that has been described above.
We have investigated behaviour when r(’) and r ( e ) are both close to r ( i ) by means
of an expansion in powers of E . An alternative approach is to describe the ‘unfolding’
of the bifurcation by using centre manifold theory (Guckenheimer and Knobloch
1982), which allows one to enumerate the different possible types of behaviour when
q is close to qo. The range of behaviour that has been described is characteristic of
bifurcations of co-dimension two (Guckenheimer 1981) which can be reduced to
normal forms (Coullet 1981). As we shall see in 08 7 and 8 below, the qualitative
picture in figure 9 can be extrapolated into the fully non-linear domain.

6.4. Subcritical steady convection


As we saw in 8 6.1 the steady branch bifurcates from the static solution at r ( e ) and its
slope there may be either positive or negative. It is important to establish the value
of rmin, the lowest value of r for which stable steady convection can occur; in particular,
we should determine under what conditions subcritical steady convection, with rmin<
r @ ) ,is possible. It is clear that if r f ) < O then rmin<r(e)but the converse is not true:
we shall demonstrate that rminmay be less than r ( e )even if r f ) > 0.
1342 M R E Proctor and N 0 Weiss

A B c D
ibl

Figure 9. Sketches showing the behaviour of solutions when r y ) >O. ( a ) U * as a function of r (cf figure
8(6)). ( 6 ) Phase portraits in the a’-a plane for values of r as indicated. The origin 0 is always unstable,
and is enclosed by a stable (cnoidal) limit cycle L. Points S correspond to the steady branch, which is
unstable in B but subsequently stable. The limit cycles L’ are always unstable; they are initially of dnoidal
form but merge to form a cnoidal trajectory at D (from Knobloch and Proctor 1981).

The rates of working of the different forces that appear in the equation of motion
can be compared by forming power integrals. We form the scalar product with the
velocity of the equation of motion (2.16) and average over the domain (for time-
dependent motion this spatial average must be combined with a time or an ensemble
average) to give
R ( W e ) = (lWI2) - [Q( U * J A B ) (6.33)
where w = U ‘2. The magnetic term can be rewritten, by using the scalar product
with B of the induction equation (2.18), so that (6.33) becomes
R ( w 6 )= ( ~ w ~ ~ ) + L ~ Q ( ~ J I ’ ) . (6.34)
This power integral states that the rate of working of the buoyancy force is balanced
by the rates of viscous and ohmic dissipation in the Boussinesq approximation.
(Strictly, the global rate of working of the buoyancy force is zero in a steady
compressible flow and only the pressure gradient contributes to the power integral.
Magnetoconvection 1343

It can be shown, however, that in the Boussinesq approximation this contribution is


equal to the left-hand side of (6.34) (Malkus 1964).) Finally, taking the product with
8 of the temperature equation (2.17),we obtain
( w e )= (lv@). (6.35)
Consider first the case when Q = 0 and the flow is two-dimensional. If we assume
that both U and 8 are small and that they can be expanded as power series in E , as
in (6,1),then we have, to second order in E , from (6.34) and (6.35)

and

Recalling that u 1 and satisfy the steady versions of (5.3) and (5.4),we may derive
the further relations
R ~ ( w= (W
~ ~ m3)
~ ) e (Wl83)=(v61 ' v83). (6.38)
From (6.36)-(6.38)it follows that
R:")(wlOl)
= R0(lV821~). (6.39)
Now suppose that Q is finite but that the magnetic Reynolds number ~ /is lsmall.
Then an extension of the above argument yields
R:"'(Wl81)= R ' " ( I V ~ Z- f)Q~[ ()U 1 ' J2 A B1) +2(U1 J1 A B2)] (6.40)
whence
(6.41)

which is equivalent to (6.12).


Alternatively, we might suppose that U = 0 ( 1 ) , t is very small and that there exist
steady solutions with R close to R o (Busse 1975). Then Q cannot be too large (in a
sense to be defined precisely later), and the PCclet number E and the Reynolds number
E / U remain small though the magnetic Reynolds number R , = ~ /may l be large. In
these circumstances, the magnetic field can be greatly distorted by the motion (cf § 3
above) so J can no longer be expanded in powers of E . Thus we put
R=Ro+SR (6.42)
and combine (6.39) with (6.34) to give
SR(wi8i)= E ~ R O ( ~ V( ~~/ EZ )~~ Q
~ ()J+' ) . (6.43)
In order to proceed further, we assume that Q is sufficiently small that the form of
the velocity field u 1 is not affected. Thus J depends only on R , and the kinematic
distortion of the magnetic field resembles that discussed in § 3. In particular, for large
R , magnetic flux is expelled from the centre of the region and confined to sheets of
-
thickness R ; 1 / 2 . It follows that ( J 2 ) R? and so (6.43) becomes
SR = E 2~o+E-1/2~t1/2~ (6.44)
approximately, where C is a constant that depends only on cy ; for the case a = 7~ (the
only one computed directly by Busse) /7 = 17.25. It is now clear that the analysis is
valid if 1112Qs E ' I 2 << 1.
1344 M R E Proctor and N 0 Weiss

The lowest value of R for which steady convection can occur in this regime is
obtained by minimising SR with respect to E . Then

min (SR)= $ ( 1 7 . 2 5 ~ l ~ ' ~ Q ) ~ / ' for E = a(17.2511/2Q/~4)2/5. (6.45)

Hence there is subcritical convection, with Rmin<R@',if Q > 28112. Moreover,


a similar result (but with different numerical constants) holds for all values of CY. In
particular, subcritical convection still occurs when R ?' > 0, in which case the steady
branch must have at least two turning points. Moreover, when Q is very large, so
that R'" >> Ro,steady convection is possible arbitrarily close to Roprovided that 1112Q
is sufficiently small: in dimensional terms, this requires that

(6.46)

When 5 is small and Rmin << R (e' it is likely that convection sets in via a Hopf bifurcation
at R(')<< R'": it can be shown that for Qo = (p4/.r2)qo < Q < O(l3l4)steady convection
can occur before the Hopf bifurcation but, for larger Q, Rmin>R(") wherever (6.45)
is valid. Strictly, (6.43) holds only if l1l2Q<< 1, though one might expect the expansion
to be rather more robust (Busse 1975); in fact, the numerical computations in 8
below show that the analysis remains reliable for SR =sRo.

7. The transition from oscillatory to steady convection in a truncated model


problem

7.1.The modal equations


In the last section we considered the behaviour of non-linear solutions in the neighbour-
hood of the bifurcations from static equilibrium. To proceed further, we should
integrate the partial differential equations (2.26)-(2.28) numerically on a computer;
this has been done and the results will be described in § 8. However, only stable
steady solutions are generated in the numerical experiments, and it is difficult to link
the steady branch to the bifurcation at r(e' when r m i n # r ( e )Hence
. it is instructive,
and also convenient, to introduce a truncated model problem with the following
properties (Knobloch et a1 1981):
(i) the bifurcations at r") and are identical to those for the full problem;
(ii) to second order in E the finite-amplitude solutions are identical to those for
the full two-dimensional problem;
(iii) the solutions are bounded as functions of time for each value of r ; and
(iv) the form of the stable parts of the branches of oscillatory and steady solutions
is qualitatively similar to that for the full two-dimensional problem (this must be
verified a posteriori).
These aims can be achieved by expressing 4, 0, A in terms of a restricted set of
spatial modes and obtaining a system of ordinary differential equations that governs
their evolution in time (cf Veronis 1965, 1966). Thus we write
$ =a( T ) ~ I

= b ( T ) 6 1+ C ( T ) 6 2 (7.1)
A = x +d(~)Al+e(.r)A,
Magnetoconvection 1345

where iJ1,el,etc, are given by (6.6) and (6.10). (Note that (7.1) can be obtained by
setting E = 1 in (6.18).) We then substitute these expressions into (2.26)-(2.28) and
consistently neglect all terms generated that involve higher harmonics so as to obtain
the system
U =a[-a+rb-&d((3-w)e-l}]
d =-b + a ( l - c )
c =w(-c +ab) (7.2)
d = -5d + a (1 - e )
6 = -(4 - w ) l e + w a d
(cf equation (6,19)),where dots denote differentiation with respect to T .
This model system has been constructed so as to satisfy the first two conditions
enumerated above. It can be shown that the solutions of (7.2) are uniformly bounded
in time and, as we shall see, they mimic many properties of the full problem. Moreover,
the divergence of the flow in phase space
ad ad ac ad a i
-+-+-+-+-=
aa ab ac ad ae
- [ u + ( l + w ) + 5 ( 5 w)]
- (7.3)

is always negative, so trajectories are attracted to a set of measure zero in the phase
space; in particular, they may be attracted to a fixed point, a limit cycle or, perhaps,
a strange attractor. Finally, the equations (7.2) possess an important symmetry, for
they are invariant under the transformation
(a, b, c, d , e ) += (-a, -6, c, -d, e ) . (7.4)
When q = 0 the first three equations decouple from the rest and can be transformed
to the system described by Lorenz (1963), which has been extensively investigated.
Although the system (7.2) can display all the behaviour associated with the Lorenz
equations we shall be concerned with solutions in a different regime, for values of r
sufficiently small that the Lorenz attractor has not yet appeared.

7.2. The stability of steady solutions


The system (7.2) is more tractable than the full problem. Steady finite-amplitude
solutions are readily obtained and their stability can be explicitly investigated.
Moreover, time-dependent solutions can be computed with adequate precision. In
the steady state

where ii = (4 - w)12/w, and


r =1 + a’ + 6 (1+ a 2 ) ( i i+ a2)-2[p+ (4 - w ) a 2 ] q . (7.6)
Note that this solution does not depend on u. Equation (7.6) is a cubic in a 2 , which
-
has zero, one or two turning points for a 2 > 0, and r a 2 as a 2 += CO. Thus the steady
solution branch has one of the three forms shown in figure 10.
When overstability does not occur (G:< 0) the stability properties of the steady
solutions are as indicated in figure 10. The two eigenvalues that are of interest are
1346 M R E Proctor and N 0 Weiss

Figure 10. Sketches of the amplitude-Rayleigh number diagram for the truncated model problem when
overstability does not occur. The signs of the two relevant eigenvalues are shown for the cases with ( a )
zero, ( b ) one and (c) two turning points along the steady branch (from Knobloch et a1 1981).

both real and they can only change sign at r L e )or at a turning point on the steady
branch. We observe that rmin< r ( e ) when r ( e )< 0 but that rminmay still be less than
r ( e )when r @ )> 0, if there are two turning points.
These stability properties become more complicated with the addition of an
oscillatory branch. When r ( e )> 0 and there are no turning points on the steady branch
the eigenvalues are real and positive in the neighbourhood of r ( e )(cf figure 7 ( a ) )but
then become complex so that there is a Hopf bifurcation from the steady branch
at r = r m i n > r ( e ) . This behaviour, which is sketched in figure l l ( a ) , is identical to
that discussed in 0 6.3. When T ( ~ ) < O and there is a single turning point, as in
figure 11(6), rmin<r(e), When r:"'>O and there are two turning points there may
be a Hopf bifurcation before the first turning point, as in figure l l ( c ) . Usually,
however, the eigenvalues remain real, as in figure l l ( d ) .

Figure 11. As for figure 10, but for the case when convection sets in via a Hopf bifurcation. Broken lines
imply instability.

7.3. Oscillatory solutions


The stable part of the oscillatory branch can be investigated by integrating the equations
(7.2) numerically as an initial-value problem. A simple example, resembling figure
l l ( d ) , is illustrated in figure 12. At small amplitudes the oscillations are almost
Magnetoconvection 1347

Figure 12. Oscillatory solutions in the truncated model problem. The variables a, d and e are plotted as
functions of time T together with the limit cycle projected on to the a-a' plane for q = 2.5, m = 2 , 5 = 0.4,
(T = 1 and r = 3.4 (from Knobloch er al 1981).

sinusoidal but as r increases the profiles grow more jagged and eventually develop a
characteristic structure with a small peak followed by a long plateau (cf figure 8 ( a ) ) .
At the same time the period grows larger and eventually tends to infinity as r + rmax.
Simultaneously, the RMS value of a approaches the value on the unstable steady
branch, as shown in figure 13. (Because of the peak, however, the maximum value

r
Figure 13. The end of the oscillatory branch. (a) The maximum and RMS values of a and (6) the period
of the oscillation as functions of r. As in figure 8 ( a ) the broken line shows the unstable steady branch.
Parameters are as for figure 12 (from Knobloch er a / 1981).

of a remains above the steady branch.) At r = rmax there must therefore be a hetero-
clinic orbit and the oscillatory branch terminates on the unstable steady branch, as
described in Q 6 . 3 . For r marginally greater than rmax there are no periodic solutions
1348 M R E Proctor and N 0 Weiss

and all trajectories are eventually attracted to the stable part of the steady solution
branch. Figure 14 shows sketches of trajectories in the ua plane for values of r
slightly less than, equal to, and slightly greater than rmax. The results obtained for
the case when r ( O ) is very close to r @ )are generic, not only for the truncated model
problem but also for the full two-dimensional equations, as we shall see below.

~ 5 U

Figure 14. Phase portraits when r y ) < 0. Sketches of trajectories in the a-a' plane when r is ( a ) just less
than, (6) equal to and (c) just greater than rmax. U and S denote unstable and stable steady solutions,
respectively (from D a Costa et al 1981).

The behaviour of the oscillatory branch for the case in figure l l ( b ) is similar to
that described above. When there is a Hopf bifurcation from the steady branch the
oscillations develop differently. Figure 15 shows solutions for an example like that

ai

Figure 15. As figure 12, but with q = 5, m = 2, [ = 0.67, c7 = 10 and r = 6.265.


Magnetoconvection 1349

in figure l l ( a ) . The form of the oscillatory branch is like that in figure 8(6): the
oscillations have a clearly cnoidal character but as r increases a plateau emerges and
the period grows longer. The oscillatory branch loses stability at rmax>rmin; the
unstable part has not been followed numerically but results are consistent with a
transition from cnoidal to dnoidal types of motion and the presence of a subcritical
Hopf bifurcation at rmin.
There is another means of capturing oscillations at a Hopf bifurcation, which was
not present in the parameter range studied in § 6 . Figure 16 shows limit cycles
projected onto the ad plane for an example resembling figure l l ( c ) . In this case the
Hopf bifurcation is actually at rmax. The oscillation seems quite normal until r
approaches rmax. Then the trajectory passes increasingly close to the unstable critical
points, which are spirals. The solution ‘feels’ the complex eigenvalues and describes
a loop around each point before escaping from its neighbourhood. As r + rmaxmore
and more loops are made until finally the critical points become stable and the
trajectory spirals into one of them. Such behaviour cannot be understood in terms
of a two-dimensional phase diagram.

l
i id)
i

Figure 16. Looping the loop. Limit cycles in the a-d plane for q = 5, w = 2, [ = 0.598 and (r = 10. ( a )
r = 6.2 (one loop); (6) r = 6.25 (two loops); (c) detail of ( 6 ) ; ( d ) r = 6.26 (eight loops) (from Knobloch e l
al 1981).
1350 M R E Proctor and N 0 Weiss

There may also be bifurcations from the oscillatory branch, leading to further,
more complicated types of oscillation. In the cases discussed so far, the oscillations
were symmetrical in that the trajectories transform into themselves under the symmetry
operation (7.4). In a certain parameter range there is a bifurcation from symmetrical
to asymmetrical oscillations. Figure 17 shows limit cycles before and after such a

lbl 4

Figure 17. Period doubling. Limit cycles in the a-a’plane for q = 5, w = 2, [ = 0.4, U = I O . (a)r = 5.20
(symmetrical); ( b ) I = 5.25 (asymmetrical); ( c ) r = 5.280 (detail showing period doubling); ( d ) r = 5.288
(quadrupled period) (from Knobloch et a / 1981).

bifurcation; there is, of course, a pair of asymmetrical solutions which transform into
each other under the symmetry operation. This bifurcation may be followed by a
sequence of period-doubling bifurcations such that after the nth bifurcation the solution
repeats itself after 2” cycles around the origin. Figure 17 shows details of solutions
with n = 1 and 2. This phenomenon appears to be an example of a Feigenbaum (1978,
1979) sequence, with an accumulation point at which the period is infinite. For the
case illustrated this sequence is followed by an inverse sequence of bifurcations at
which the period is halved, followed by a bifurcation back to symmetry. It seems
likely that there exist ranges of parameters for which the period-doubling cascade is
followed by chaotic solutions, suggesting the appearance of a strange attractor, as in
a related problem (Da Costa et a1 1981, Knobloch and Weiss 1981). Although
bifurcations to asymmetry have been found for the full problem, chaotic behaviour
has not yet been found in magnetoconvection, in contrast to thermosolutal convection
(Huppert and Moore 1976). These exotic bifurcations offer a fascinating, and tantalis-
ing, topic for research. At the moment it is not clear whether they will persist when
more terms are added to the expansions in (7.1) and whether they are relevant to the
full problem. On the other hand, it seems likely that if there are aperiodic oscillations
in the full problem then they are preceded by a Feigenbaum sequence. This would,
however, be extremely difficult to detect in observations or numerical experiments.

8. Two-dimensional numerical experiments


8.1. The development of non-linear magnetoconvection
The standard two-dimensional problem, defined in 8 2.3, has been extensively investi-
gated by solving the full set of non-linear partial differential equations on a computer
Magnetoconvection 1351

(Weiss 1981a, b, c). In this section we shall summarise these results and show how
they relate to the solutions obtained by expansion methods in the last two sections.
When the amplitude of convection is small the two complementary approaches are
in excellent agreement but qualitatively new effects are present in the fully non-linear
regime. The kinetic energy is a useful measure of the vigour of convection. Let

U* = ( / U1)' = A JO JO 4w dx dz.

Figure 18 shows the RMS velocity U as a function of R for two typical cases. Only
stable parts of the steady and oscillatory branches have been obtained numerically
and the unstable portions are conjectural. The first case (with A =;) resembles that
in figure l l ( a ) , with no turning points on the steady branch; the second (with A = 1)
resembles figure 1l(d). The general behaviour of the non-linear oscillatory solutions
is very similar to that described in the preceding section. In particular, the period
increases towards the end of the oscillatory branch.

.
N
0 e

0 e
R/R,
Figure 18. Subcritical and supercritical two-dimensional convection. Scaled RMS velocity plotted against
Rayleigh number, for q = 2.5, U = [ = 3, (a) A = 1, (6) A =f. Only the maximum values of the spatially
averaged velocity are shown on the oscillatory branches; the unstable segments of the steady branches are
conjectural (from Weiss 1 9 8 1 ~ ) .

We shall, for the moment, concentrate on results obtained for A = 1, when RY)=
Ro+27r2Q,from (6.41), and is independent of i.The thermal efficacy of convection
is measured by the Nusselt number

N =A-* jo*
(1- M / d z ) dx (8.2)
1352 M R E Proctor and N 0 Weiss

evaluated at z = 0 or z = 1. The Nusselt number is the ratio of the total heat flux to
the flux that there would be in the absence of convection: for a purely conducting
solution N = 1 and for small amplitude convection N 9 1+ 2c, where c is defined in
(7.1). When Q = 0, N increases monotonically with R along the steady branch, which
bifurcates from R o as shown in figure 19; for Q = 100, f = O . l there is a Hopf
bifurcation from the static solution and the steady branch has (at least) two turning
points. Stable steady solutions can be found for R > Rmi,= 1780 which is very close
to the value of 1746 predicted by equation (6.45). Thus Busse’s (1975) formula
remains accurate even with [‘”Q i= 30. Inspection of the solutions confirms that
a$/az varies linearly across the flux sheets, within which the field has a Gaussian
profile, as described in 8 3.

R, R’” B R ”

Figure 19. Convection with Gaussian flux sheets. Nusselt number as a function of Rayleigh number with
Q = 0 and with Q = 100, [ = 0.1, A = cr = 1. The full curves connect stable steady solutions, while vertical
bars indicate the maximum amplitudes of oscillatory solutions. The value of Rmincalculated from (6.45)
is indicated by a letter B (from Weiss 1981a).

Subcritical convection with Rmingiven approximately by (6.45) occurs provided


<
[1’2Q5100 and is sufficiently small. For larger values of f”2Q the form of the
solutions changes and the expression (6.45) ceases to be accurate. This regime can
be investigated by varying [ while holding [Q fixed. (In dimensional terms this
corresponds to varying 7 while Bo is held fixed.) Figure 20 shows the results of three
series of numerical experiments with Q = 100. For f = 1 there are no oscillations and
N increases monotonically with R. Overstability is possible for [ < 0.6581; for = 0.5, <
rmln is slightly greater than r ( e ) and the steady branch resembles that in figure 18(a).
Results obtained for f = 0.2, 0.1, 0.05, 0.025 are also similar to those shown in figure
2 0 ( b ) : subcritical convection takes place on the steady branch and R,,, apparently
tends to a limit as [ + 0. In this limit, R,,, 6500, while R‘” = 6962. Thus steady
i=

convection first appears at a value of R that is determined by the field strength and
is independent of the magnetic diffusivity. Furthermore, this value is close to that at
which the overturning instability occurs on linear theory in the limit as all diffusivities
go to zero (cf 0 5.2).
Equation (6.45) leads one to expect that Rmln--f-2’5 as [ + O for fixed [Q, in
contradiction to what is found. To understand why, we must look in detail at the
Magnetoconvection 1353
-
io)

4 -

3-

2-

1, '
3-

2-

1 1 I 1 1 I

solutions. Figure 21 shows contours of 4 (streamlines), U , T" (isotherms) and A (lines


of force) for a typical solution with 5 = 1. The streamlines fill the entire cell and are
qualitatively similar to those of the linear eigenfunction. The vorticity field is slightly
distorted but always has the same sign. The temperature is distorted by the motion
to form a hot rising and a cold sinking plume and the flux function looks similar when

Figure 21. Steady non-linear convection. Contours of $, U, T* and A with [ = 1, R = 6000; [Q = 100,
U = 1, A = 1 (from Weiss 1981a).

rotated through 90". When 5 = 0.025 things are very different: in figure 22 magnetic
flux is almost entirely concentrated into isolated sheets and the Lorentz force generates
countervorticity at their edges so that motion is excluded from the regions where the
field is strong (cf figure 4). Figure 23 shows profiles of the horizontal velocity (-a*/&)
and the vertical field (aA/ax)across the flux sheets for the cases described above.
When 5 = 1 the velocity varies linearly and the field is Gaussian (cf § 3) but when f
is small the field is roughly constant and there is scarcely any transverse motion in
the flux sheet (apart from a slow, thermally driven countercell). The width of the flux
sheet is independent of i though the current sheath becomes thinner as 5 + 0.
1354 M R E Proctor and N 0 Weiss

Figure 22. As figure 21, but with [ = 0.025, R = 6500. Motion is excluded from the flux sheets.

Figure 23. Stagnant flux sheets. Variation of U, and B, across the flux sheets for the parameter values of
( a ) figure 21 and ( b ) figure 22 (from Weiss 1981a).

Figures 21 and 22 help to explain the occurrence of subcritical steady convection


when t is small. If the magnetic Reynolds number is sufficiently large, flux is expelled
from the centre of a convection cell and concentrated at its edges, allowing convection
to proceed in the field-free central region. Although the rate of ohmic dissipation
increases as flux sheets are formed there is a compensating increase in the rate of
working of the buoyancy force, as indicated by equation (6.44). It follows that
subcritical steady solutions cannot be described by models that do not allow for
variations in the horizontal structure of the field. Thus the mean field approximation
(Van der Borght et a1 1972) is inadequate when ( < 1. Fortunately, the second-order
magnetic term (represented by e in the truncated model of 9 7) does provide a limited
description of flux expulsion (just as c describes the formation of thermal boundary
layers). Hence that model yields a reasonable qualitative representation of subcritical
behaviour, though it overestimates the ease with which subcritical convection can
occur. Furthermore, to describe the development of stagnant flux sheets would require
more modes than were included in § § 6 or 7.
On the oscillatory branch solutions grow in amplitude as t is decreased. As can
be seen from figure 20(b), both the maximum amplitude and the time-averaged value
of N increase initially along the oscillatory branch, reach maxima and then decrease
Magnetoconvection 1355

as R approaches R,,,, which is slightly greater than Rmin.When R is near R ( O ) the


non-linear oscillations are similar to the eigenfunctions of the linear problem. Figure
24 shows the time-dependent fields at successive intervals during an oscillation: the

Figure 24. Small-amplitude oscillatory solutions. Contours of $, U , T* and A at six equal time intervals
during a half-cycle for Q = 100, [ = 0.1, R = 3000 (from Weiss 1981a).

vorticity shows some structure but has no significant subsidiary maxima. The time-
dependent fields in figure 25, for R close to R,,,, are completely different. As the
field is concentrated, zones of strong reversed vorticity are created, which eventually
engulf the central zone of thermally generated vorticity, leading to an interval of
stasis. Magnetic flux is expelled from the centre and the field lines are almost straight
during the period of stasis, when they gradually diffuse towards the centre. The
behaviour of N , U’ and the excess magnetic energy density (measured as the square
of the AlfvCn speed to facilitate comparison with U’)

A V’ = a f ~IB (l2 - 1) (8.3)


is shown in figure 26(a). Note (i) the relatively long periods of inactivity in the kinetic
energy and, to a lesser extent, in N ; (ii) the disparity in magnitude between A V 2 and
U’, showing that the magnetic energy is much greater than the kinetic energy in this
non-linear regime (for linear solutions, A V 2 / U 2= (1+ a ) / (-l f ) = 2.2); and (iii) the
asymmetry in the profile of A V2.
1356 M R E Proctor and N 0 Weiss

Figure 25. Larger amplitude oscillations. As figure 24, but with R = 6500.

Figure 26. Symmetrical and asymmetrical oscillations. Nusselt number as a function of time ( a ) for the
parameters of figure 2 5 , (6) for [ = 0.2, R = 6000 (from Weiss 1981a).
Magnetoconvection 1357

The period P of the oscillations varies along the oscillatory branch, dropping at
first and then rising steeply as R approaches R,,,. Moreover, the RMS value of N at
the end of the oscillatory branch is less than the value when R = Rmin. The oscillatory
branch apparently terminates on the unstable part of the steady branch as the period
becomes infinite, in the manner described for small-amplitude solutions in § 6.3 and
for the truncated model problem in § 7.3. Furthermore, as R approaches R,,, there
may be a bifurcation leading to solutions that are slightly asymmetrical in time. In
figure 26(a) successive peaks, separated by an interval $P,are identical but the solutions
in figure 26(b), obtained for 4' = 0.05, have peaks that repeat regularly but are slightly
different, so that both N and U vary with period P. No further bifurcations have so
far been discovered.

8.2. Spatially asymmetrical solutions


The solutions that have just been described have point symmetry about the midpoint
(BA, B). They can be expanded in Fourier series of the form

E E $mn ( t ) sin m a x sin n m


m n
(8.4)

where Gmn = 0 if ( m + n ) is odd, and it can be shown that this property is preserved
as the solutions evolve in time, The numerical experiments show, however, that in
certain circumstances small disturbances with ( m + n ) even, introduced by rounding
errors, will grow, leading to a manifest lack of symmetry. Figure 27 shows a solution
obtained with 5 = 0.1, Q = 1210, R = lo4 in which 90% of the magnetic flux is on
one side of the cell. The region occupied by the magnetic field occupies about one-third
of the cell and is almost stagnant.

Figure 27. Spatially asymmetrical solutions. As figure 21, but with R = lo4, Q = 1210, 5 = 0.1 (from Weiss
1981b).

Although no formal theory is available, the development of asymmetry may be


explained as follows. When the flux sheets are dynamically active the tangential as
well as the normal component of velocity must vanish at their edges (cf figure 23(6)).
Thus they change the effective lateral boundary conditions. If a little flux is transferred
from the right-hand to the left-hand side of the cell, the boundary condition is relaxed
on the right, allowing the flow to accelerate and so to pump more flux across the cell.
In a steady state the right-hand flux sheet has a Gaussian profile and the average
values of B are the same at either edge of the cell (cf § 3.2). Convection is more
vigorous within the field-free region when the configuration becomes asymmetrical
but, since the field suppresses motion throughout a large part of the cell, the Nusselt
number is slightly less than it would be for a symmetrical solution.
1358 M R E Proctor and N 0 Weiss

For steady solutions this asymmetry is most marked when R is near Rmin(provided
Q is sufficiently large). As the configuration is varied from a state with stagnant flux
ropes to one in which the field is dynamically passive, the degree of asymmetry
diminishes and in the kinematic regime symmetrical solutions are apparently stable.
The transition from the steady to the oscillatory branch is more complicated than in
the cases described in § 8.1. As R approaches Rminthere are oscillations about a
steady state which do not result in a reversal of the main flow. Figure 28 shows the
fields during a complete cycle. The magnetic and field-free regions may be considered
separately. In the former the strong field is unstable to oscillatory modes and non-linear
oscillations modulate the main flow without reversing it (Galloway 1978). As R is
further decreased these oscillatory solutions give way to ones in which the flow reverses.
Figure 29 shows the fields for a half-cycle of such an asymmetrical oscillation. The
transition from oscillations of the type in figure 29 must occur through a degenerate
periodic solution similar to that discussed in 8 6.3.

Figure 28. Asymmetrical vacillation. As figure 24, but now for a full cycle with R = lo4, Q = 1960, [ = 0.1
(from Weiss 1981b).

8.3. Preferred length scales for convection


Linear theory predicts that for Q sufficiently large convection first appears with narrow
cells of width A a Q-'I6, as shown in § 5.3. (Note that A is not necessarily the same
as A, the width of the periodic box in which calculations are carried out.) Non-linear
Magnetoconvection 1359

Figure 29. Asymmetrical oscillations. As figure 28, but for a half-cycle only with Q = 2890.

studies of convection in the absence of a magnetic field (Moore and Weiss 1973)
indicate that the preferred horizontal scales lies in the range 1 s A ' S 1.5. Numerical
experiments with A = 2, 3, 4, R = lo4 and Q = 800 show, however, that for finite-
amplitude convection in the presence of a strong magnetic field cells with A ' = 2 are
preferred. Figure 30 shows a steady solution, obtained by integrating the equations
from an irregular initial state, for A = 4. The two cells are of roughly equal width and
are separated by a slab of flux. We observe that the vorticity has the same sign in
each cell and that the effective widths of the eddies themselves are approximately
1.3. In this regime the field is confined to isolated regions and the flow has much the
same horizontal scale as it would have had in the absence of any field.
The behaviour of cells with A ' = 2 is most conveniently investigated by conducting
a series of experiments in which R is fixed while Q is varied. (Note that increasing
Q for fixed R is broadly equivalent to decreasing R for fixed 0, and Q ' e ) , Q'"' etc, 9

may be defined by obvious analogy.) For R = lo4, 10' the steady branch continues
until Q = Q,,, (equivalent to Rmin!)and Q >
,,, Q(')> Q ( e ) . As Q approaches Q,,
the eddy is squeezed out of existence by the expanding slabs of flux.
Figure 31 shows N plotted against Q for R = lo5 and A = A ' = i, 2. For A = $ both
Q(') and Q@) are close to their maximum values as functions of A. As Q is decreased
from a large value overstable convection first appears, with narrow cells, around
Q = 65 000. For A = the steady branch, which bifurcates from the static solution at
Q'"==8000, corresponds to that shown in figure l l ( a ) . In fact, 'steady' solutions do
1360 M R E Proctor and N 0 Weiss

Figure 30. Asymmetrical convection with A'2.2. Streamlines and lines of force with A =4, R = lo4,
Q = 1000, [ = 0.1.

- e -

0 1 2 3 L S
Li x 10."
Figure 31. Preferred length scales in linear and non-linear regimes. N as a function of Q for R = lo5,
4,
J = 0.2, cr = 1 and A = 2.

not appear until Q = 3000 and even then the flow vacillates, like the solutions in
figure 28. For A = 2, Q(e)= 2000 but 'steady' solutions persist until Q = 45 000.
The numerical results described in this section indicate that the time-averaged
heat flux for oscillatory convection remains comparatively small. In many circum-
stances we wish to ascertain the smallest value of R (or, alternatively, the largest
value of Q) for which convection can contribute significantly to the heat transport.
Thus we must find the minimum value of Rmin(or the maximum of a,,,) as a function
of A'. As we have seen, in the regime of interest this occurs for A ' = 2 and non-linear
computations imply that Rmin-@ (or Qmax-R/l). In dimensional terms, we find
that convection becomes effective when
gE ATd -B ; / k o p o . (8.5)
Magnetoconvection 1361

This non-linear result agrees with that derived from linear theory in equation (5.29).
Indeed, it is the only simple relationship that is dimensionally consistent and does
not involve the diff usivities.

9. Axisymmetric convection

9.1. Steady solutions


In this section we shall consider axisymmetric cells, with the idealised boundary
conditions (2.34). The velocity and magnetic fields are described by the Stokes stream
function V and the flux function x,respectively, which evolve according to equations
(2.31) and (2.33). As Rayleigh (1916) observed: 'We may regard the hexagon as
deviating comparatively little from the circular form'. Axisymmetric convection is
geometrically three-dimensional and allows us to investigate the behaviour of isolated
tubes of flux (cf § 3.2).
The linear eigenvaiue problem, as described in § 5, is independent of the horizontal
planform. Now 9 and x are related to the potentials F and H defined in (5.8) by

(9.1)

Clearly F, H and 8 satisfy Bessel's equation of order zero, while


(D2+a2)V=(D2+(y2)G(-~r2)=0. (9.2)
Thus the spatial dependence of the eigenfunction is given by
V a rJl(ar)sin TZ
8 =Jo(ar) sin T Z (9.3)
(x- t r 2 )CC r ~ ~ ( acosr )m.
The radius of the cell is A = j l / a , where jl = 3.831 71 . . . is the first non-trivial zero
of J l ( x ) . In the absence of a magnetic field there is a simple bifurcation at R = R o ( a )
which is a minimum when a = 1r/J2, so that A = 1.725. Throughout the rest of this
subsection we shall confine our attention to solutions obtained with this value of A.
In the presence of a magnetic field there is a simple bifurcation at and a Hopf
bifurcation at R'"', provided that (5.23) is satisfied. Modified perturbation theory is
significantly more complicated in axisymmetric geometry, and the results can no longer
be expressed in closed form. Moreover, even for steady solutions R:"'depends on
the Prandtl number U . When Q = 0, R:") K U - * for the astrophysically interesting
case u << 1, but the form of the solution changes when the Reynolds number €U-' is
large, leading to a 'flywheel' regime in which the steady solutions are independent of
U (Jones et a1 1976, Proctor 1977). We shall only develop modified perturbation
theory in the case (+ + CO when, for Q << 1, R f' tends to a constant value (Liang et a1
1969). Setting
V = ErJl(ar)sin T Z +. .. (9.4)
we have
Ry' = 12.56+ 1.32Q[-2+O(Q). (9.5)
For this value of a , R:"'> 0 if Q is sufficiently small.
1362 M R E Proctor and N 0 Weiss

If ( K 1, SO that R , = E / ( is large while E is small, the flux will be concentrated


into a central rope and the field will have a Gaussian profile provided that Q is not
too large (see § 3.2). Then the form of the steady branch can be obtained by an
analysis similar to that in § 6.4. We need to calculate the contribution from the ohmic
dissipation term in the analogue of (6.43). The field is confined to a region of radius
O(R,"2) and its magnitude is increased by a factor O(R,). Thus ( J 2 ) - R $ and so
SR = 12.56e2+ 153Q (9.6)
(cf (6.44)) where the constant is obtained from a detailed calculation (Proctor and
Galloway 1979). Since 37r2< 153, R remains greater than R ( e )along the steady
solution branch in this regime. The contrast with the two-dimensional result (6.44)
is caused by the enhanced dissipation within an axial flux rope. Flux expulsion on its
own is not sufficient to produce subcritical convection.
Subcritical convection might, however, be achieved if the Lorentz forces are strong
enough to affect the form of the flow in the neighbourhood of the flux rope. Then
flux concentration could be halted, so reducing ohmic dissipation, without spoiling
the beneficial effects of flux expulsion. Numerical experiments (§ 9.2) do indeed
confirm that R,,,<R"' for sufficiently large Q and sufficiently small [. In the limit
when [ is extremely small, so that In R , >> 1, the dynamical effects of the flux rope
can be determined analytically. Using methods to be described in § 10 below, solutions
can be obtained for all Q << R,/ln R , (Proctor and Galloway 1979). For R, >>
Q In R , >> 1 the expression (9.6) is replaced by
SR = 12.56c2+ 123/ln ( E / [ ) (9.7)
which is independent of Q! Then R,,,=Ro+ 123/11n ( 1 , and this is clearly less than
R'". For yet larger values of Q the field profile becomes flatter than Gaussian and
the theory ceases to apply. We must turn to the computer to find out more.

9.2. Numerical experiments


Galloway (1976, 1978) and Galloway and Moore (1979) carried out a thorough
investigation of axisymmetric magnetoconvection. In the following discussion we shall
emphasise the differences between their three-dimensional geometry and the two-
dimensional geometry of § 8. In their calculations, they fixed the value of R and
varied Q. Figure 32 shows the Nusselt number as afunction of Q for R = 20R, = 1 3 5 ~ ~

t
10 10: io3 10"
Q
Figure 32. Axisymmetric magnetoconvection. N as a function of Q for R = 20R,, A = 4, [ = 0.1, U = 1.
Note the occurrence of hysteresis (from Galloway and Moore 1979).
Magnetoconvection 1363

and A =$. In the kinematic limit, when Q is small, there is a steady solution with a
thin central flux rope (cf § 3.2). The field has a Gaussian profile as shown in figure
3 3 ( a ) but the Lorentz force has no significant effect on the vorticity. For larger values
of Q the field becomes dynamically important and motion is excluded from the central
flux tube. Note that the transition from a kinematic to a dynamical regime occurs for
much smaller values of Q than in the two-dimensional geometry. However, the
Lorentz force has little effect outside the flux tube and therefore the velocity and

Figure 33. Field profiles in an axial flux rope. The parameters are as for figure 32 with ( a ) Q = 2 and ( b )
Q = 100 (from Galloway and Moore 1979).

hence the Nusselt number are only marginally reduced. Figure 34 shows a steady
solution for Q = 100. The profiles of vorticity reveal a prominent ridge due to the
current sheet at the periphery of the tube, whose field no longer has a Gaussian form:
in the interior the field strength is almost constant, rising to a maximum at the edge
(owing to the presence of a weak counterflow) as shown in figure 33(b). As Q is
further increased the Nusselt number begins to fall along the upper branch of figure
32, though steady solutions persist until Q = 4000. In this regime motion outside the
flux rope is progressively inhibited, and the rope expands until it fills about half the

Figure 34. Steady axisymmetric convection with a dynamically limited flux rope. Contours of P,T* and
x and profiles of w for the case in figure 3 3 ( b ) (from Galloway and Moore 1979). ( a ) Streamlines, ( b )
isotherms, ( c ) vorticity perspective, ( d ) magnetic lines of force.
1364 M R E Proctor and N 0 Weiss

volume of the cell. Ohmic dissipation takes over from viscous dissipation in balancing
the work done by the buoyancy force.
For 5 5 0 0 s Q d Q(’) = 10 300 there are oscillatory solutions in which the whole
flow changes sign. Figure 35 shows the behaviour of the fields over a complete cycle
for Q = 7000. A series of runs made as Q was progressively decreased from Q(”
shows striking hysteresis. Oscillations appear again but with a different, greater
amplitude, as can be seen in figure 32. Regular oscillations persist until Q = 2500 <
Q ( i ) =2565. For 2 0 0 0 3 Q B 1000 the oscillations have a different form: figure 36
shows the behaviour of the fields over two complete cycles with Q = 1500. Notice
that the oscillation is confined to the region of strong magnetic field, and that the
sense of motion in the main eddy does not change. These solutions resemble the
two-dimensional results in figure 28, and the periods can be estimated by applying
linear theory within the flux rope (Galloway 1978). For lower values of Q the motion
is steady, but the Nusselt number is lower than in the sequence for ascending Q, at
least for Q b 5 0 . Apparently there are two distinct families of solutions in the
dynamical regime with different proportions of flux in the central rope. These families

lbi

Figure 35. Axisymmetric oscillations. As in figure 34 at eight equally spaced intervals over an entire cycle,
with Q = 7000 (from Galloway and Moore 1979).
Magnetoconvection 1365

Figure 36. Axisymmetric vacillation. As in figure 35 but now for two complete cycles, with Q = 1500.

presumably correspond to the symmetrical and asymmetrical solutions in the two-


dimensional geometry, though there is no obvious symmetry to be broken in an
axisymmetric configuration.
In general, the axisymmetric results described above parallel those obtained for
two-dimensional rolls. The latter allow a more precise description of the bifurcation
structure and the development of non-linear convection. However, the transition
from the dynamical to the kinematic regime is more readily described in the axisym-
metric geometry as we shall see in the following section.

10. The transition from a kinematic to a dynamical regime


10.1, The peak field strength
For a given Rayleigh number, the peak field in the flux tube varies with Q. It is
convenient to express this field as a dimensionless Alfvtn speed V", such that
V" = VOB" Vo = (Q[m)"* B" = max {~B,(o,
z)1}. (10.1)
z
1366 M R E Proctor and N 0 Weiss

Figure 37 shows the behaviour of V * as a function of V oin axisymmetric geometry


(Galloway and Moore 1979). In the kinematic regime V*KR,Vo; as Lorentz forces
become significant the slope decreases and V * reaches a maximum V,. Immediately
after the maximum V * decreases slowly as V ois increased and the field profile remains
approximately Gaussian. In the fully dynamical regime when the flux tube is stagnant
V * decreases more rapidly. Eventually there is a transition to oscillatory convection
and for Q > Q‘O) the layer is stable and V * = Vo. It is of particular interest to establish
the maximum value V , of the field in an isolated flux tube (Parker 1981).

1 10 100
v,
Figure 37. Field amplification. V* as a function of V, (log-log plot; arbitrary units) for R = lOR,, [ = 0.5,
U = 1, A = 4 (from Galloway and Moore 1979).

Let U be a characteristic speed of convection in the kinematic regime. For the


whole of this section we shall assume that U is sufficiently large that the Reynolds
number U/u is very small: the numerical experiments suggest that the form of
asymptotic results obtained for small U/u may hold for high Reynolds numbers too.
Then we can estimate the dependence of V , on U by the following argument. In
axisymmetric geometry the vorticity equation (2.31)simplifies to

(10.2)

Thus the vorticity w can be written as the sum of two contributions WO and w1 with
corresponding velocities uo and u l , such that

(10.3)

-
Clearly luo/ U ; the transition from a kinematic to a dynamical regime occurs when
the effective magnetic Reynolds number R, is reduced from its original value U / l
due to the counter flow generated by the Lorentz forces, i.e. when lull is comparable
to U. The current density J vanishes on the axis and outside the flux tube; we may
model this simply by supposing J to be concentrated in a sheath of radius E about
-
the axis, where E Rm1” << 1; then (10.3) is dominated by radial derivatives so that
w1 = -dul/dr, say, where u 1 is the induced vertical velocity. Then

(10.4)
Magnetoconvection 1367

For simplicity we suppose A = 1; then the boundary conditions on u1 are

dul/dr = 0 r=0,1 lo
1
ruldr=O. (10.5)

It follows that
Odrse
(10.6)

where
U1 -- - 2 L~ 2 [I~~E/-$$(~-E~)]Q~B*~

= -$E2/ln E/Q[B*~ (10.7)


(Galloway et a1 1978). The induced velocity distribution extends across the whole
cell but, provided /In E / is large, it is only significant in the neighbourhood of the flux
rope.
-
Kinematic amplification is halted when U1 U, and so E* reaches a maximum
B, such that
B2-R:/Q1nRm (10.8)
whence

(10.9)

This result is confirmed by a more respectable calculation which will be described in


the next section.
A similar treatment can be applied to the two-dimensional geometry (Galloway
et a1 1978). No logarithmic terms appear in the solutions, and the induced velocity
-
is of the order of U1 R,'/2QlB*2. Thus the maximum field is given by

(10.10)
-
The prediction that V , [-'I4can be checked by comparison with the results of
numerical experiments (Peckover and Weiss 1978, Weiss 1981b). These computations
suggest that V,-l-o.2 which is in reasonable agreement with (10.10). It follows from
this equation, or from (10.9), that V , may be arbitrarily large if the ratio v / q is
sufficiently big.

10.2. A boundary layer model


In the axisymmetric geometry, it is convenient to scale the dimensionless velocity with
U, which is based on lu0/ and thus known, so that for steady fields the induction
equation (2.33) can be written
U vx = R ; ~ D ~ ~U = Uv = UVA(9@/r). (10.11)
As above, v can be written as v o + vl, where v 1 is the induced velocity. For large R,,
as previously shown, all the flux is confined to a tube of thickness Rm1" at the axis.
If we suppose that, near r = 0, v is dominated by terms with a radial length scale
1368 M R E Proctor and N 0 Weiss

>>RA1”,we may write


W - ir2f(z)+ O(r2) r+O (10.12)
where f ( z ) =u(O, z ) 2. Substituting (10.12) into (10.11) and ignoring higher-order
terms, we find that
x = x (5,z ) = A: ’[ 1- exp (-5’f/4z )] (10.13)
where the boundary layer coordinate 5 = Rm1/’r, The axial magnetic field is then
B,(O, z ) = R,A2f(z)/4z. (10.14)
The solution (10.13) breaks down at z = 1, but this can be shown not to affect the
determination of f ( z ) at leading order.
To calculate the induced velocity we need the vorticity equation for ql,which can
be written

(10.15)

The solution of this boundary layer equation has to be matched to a Stokes flow in
the rest of the cell. Because the axis is singular it turns out that the boundary conditions
at r = A are not necessary to determine the leading-order form of TInear r = 0. After
some a!gebra it emerges that
QA
q1-- (10.16)
32
so f ( z ) can be written

(10.17)

If f o ( z )is known, this equation can be solved to yield

(10.18)

where A = i A 4 Q In R, (Galloway et a1 1978). It follows from (10.17) that for each


z , Q’/’B,(O, z ) attains its maximum when A is of the order of unity.
Figure 38 shows Q1/’f(z)/2z (proportional to the axial field) as a function of Q1’2
(proportional to the mean dimensional field) for various values of z, when u o is
generated by Oberbeck convection in a deep cylindrical container. For this model
B* = B,(O, 0), and as Vo is varied V* reaches a maximum V, such that
U2
Vk = 0.017 -( d i ) (10.19)
In R,
in precise agreement with (10.9).
In the Rayleigh-BCnard problem the simple separation of f into f o and f l is not
possible, since the temperature field is itself determined by the form of the total
velocity field. Nevertheless, some analytical progress is possible when the PCclet
Magnetoconvection 1369

I 1_
0 5 10
c1"2

Figure 38. Field amplification in axisymmetric Oberbeck convection. Axial field as a function of mean
field (linear plot; arbitrary units) with z = 0, A, 2A (in descending order) in a deep cylinder (A << 1) (from
Galloway et al 1978).

number E is small and Q is not too large. In the flux rope regime (R, = ~ / l 1)
>the
>
relation (9.6) is replaced by

SR = 1 2 . 5 6 ~ ' +27.520 (10.20)

(in the case A = 1.725 . . .) where f ( z ) is given by (10.17) and f o ( z )= (T/&) sin m.
These coupled equations can then be solved numerically to give E as a function of
SR, 5 and Q (Proctor and Galloway 1979). Thus it is possible to calculate Rminas a
function of Q for fixed f < c 1; as was shown in 8 9.1, Rmintends to a constant value
for large Q (provided Q << R,/ln (R,) so that the analysis remains valid). It can then
be shown that for R less than this limiting value there is a maximum value for Q for
which steady convection can occur while, for larger R , E tends to a constant at large
Q. Figure 39 shows V" as a function of Vo in the two respective cases. Note that
once again there is an initial kinematic phase followed by a dynamical regime and
that V" reaches a maximum V, (cf figures 37 and 38).
The above approach is less successful when applied to two-dimensional geometry.
Because there is no singularity at an axis, the induced velocity fills the entire cell.
Hence (10.17) must be replaced by a non-linear integro-differential equation, which
cannot be solved by any standard method.

11. Further effects

11.1. Horizontal fields


So far this review has concentrated on the standard problem defined in Q 2. Moreover,
non-linear results are available only for special geometries (two-dimensional or axisym-
metric). Such solutions are probably unstable to three-dimensional disturbances,
though no quantitative stability analysis is available. In this section we shall consider
1370 M R E Proctor and N 0 Weiss

vo

Figure 39. Field amplification for the flux rope model. V* as a function of V , for I<<1, R = 666 (-)
and 663 (- - -). Note that in the former case the magnetic field is unable to suppress convection while the
flux rope is thin (from Proctor and Galloway 1979).

the consequences of modifying the standard problem and including additional effects
in the governing equations. The survey will be comparatively superficial, with an
emphasis on physical ideas rather than on detailed calculations. Many of the topics
to be covered have in any case been reviewed elsewhere, in the context of hydromag-
netic dynamo theory (Moff att 1978, Parker 1979).
The standard problem may be changed by imposing a horizontal rather than a
vertical magnetic field. If the motion is confined to rolls with axes parallel to the field,
then the latter has no dynamical effect. An inclined magnetic field eliminates the
degeneracy of the linear stability problem, favouring rolls whose axes lie in the same
plane as the magnetic field (Chandrasekhar 1961). Nevertheless the alternative
configuration, with the roll axes transverse to the magnetic field, has some relevance
in astrophysics (see below). Suppose then that Bo =I?& ; in two-dimensional geometry
the boundary conditions on the flux function A become
aA
-= 0 on all boundaries A(x,0) = 1, A(x, 1)= 0 (11.1)
ax
instead of (2.29). Otherwise the problem is unchanged.
In the absence of any dissipation the transition from neutral oscillations to exponen-
tially growing instabilities occurs when

(11.2)
FOP0

(cf equation (5.29)). Thus instability first occurs when a is very small and the cells
are elongated in the direction of the field; with a vertical field, on the other hand, it
Magnetoconvection 1371

follows from (5.29) that large values of a are preferred. Aspects of the development
of non-linear magnetoconvection in a perfect fluid have been discussed by Pedlosky
(1978). When diffusion is included, stability criteria can readily be obtained by
replacing Q by Q/A2 in the expressions given in § 5.2. Results analogous to those
obtained in $ 6 by modified perturbation theory can also be found (Arter 1982).
In the fully non-linear regime there are, however, important differences. Numerical
experiments (Arter 1982) confirm that flux is expelled into sheets at the upper and
lower boundaries; when 6 < 1 these sheets are contained within the thermal boundary
layers but the magnetic regions cannot be isolated from the convective eddies as they
were, for instance, in figure 23(b). As a consequence horizontally elongated cells are
preferred for steady motion when Q is large. Figure 40 shows convection in a single
cell with A = 6 and 6 = 0.2. The horizontal flux sheets are dynamically active, and
produce a greater variety of time-dependent behaviour than when the field is vertical.

Figure 40. Convection in a horizontal field. Contours of @, w , T*, A for A = 6 and R =3x l o 4 , Q = 1350,
[ = 0.2, (T = 1 (from Arter 1982).

For 6 < 1 and Q sufficiently large a branch of oscillatory solutions bifurcates from
the static solution as before. Initially these oscillations are symmetrical but they
become asymmetrical as R is increased. There is a transition to steady convection
with Rmin<R'", but as R is increased past Rminthere is a Hopf bifurcation from the
steady branch leading to oscillations about the steady solution without reversal of the
flow. These vacillations develop in amplitude as R is increased and may grow until
the sense of motion is eventually reversed. It is uncertain whether the steady branch
regains stability as R +CO.
1372 M R E Proctor and N 0 Weiss

11.2. Modified boundary conditions


All non-linear investigations have so far been restricted to the standard boundary
conditions (2.23) and (2.24) or (11.1). In particular, it has been assumed throughout
that the temperature is fixed on the upper and lower boundaries. This is a special
case of the more general boundary condition
dT
-+WT =constant (11.3)
az

where the Biot number 28 can vary from zero (fixed flux) to infinity (fixed temperature).
In the absence of a magnetic field, fixed flux boundary conditions favour instabilities
on long horizontal scales (see Hurle et a1 (1967) and references therein). The
non-linear problem has been attacked by asymptotic methods based on shallow water
theory (Chapman et al 1980, Chapman and Proctor 1980): the mode of maximum
growth rate on linearised theory is found to lose stability to disturbances with larger
scales until the entire region is occupied by a single cell. It should, however, be noted
that as A becomes greater than (HJd)”’, where H, is the pressure scale height, the
Boussinesq approximation ceases to be valid and subcritical instabilities appear
(Depassier and Spiegel 1981). The development of flattened cells has also been
demonstrated in numerical experiments (Hewitt et a1 1980). With small but finite
values of W the preferred length scale is of the order of B3-1’4 for weakly non-linear
theory but B3-’I3 in the non-linear regime (Busse and Riahi 1980, Proctor 1981).
With an imposed horizontal magnetic field, the preference for horizontally
elongated cells would be enhanced by fixed flux boundary conditions. With a vertical
field it seems likely that broader cells will form between the flux sheets and that the
steady solution branch will become yet more subcritical. Exclusion of motion from
the flux sheets leads to an increased vertical temperature contrast; hence, the top of
a flux sheet will be colder than its surroundings. The corresponding density gradient
should enhance the field strength within the sheet.

11.3. Rotation
The effect of rotation on convection has many similarities to that of a magnetic field.
When both effects are simultaneously present, the critical Rayleigh number may be
lower than if either constraint had been acting separately. Linear theory is inevitably
more complicated than that described in Q 5 and only a few intrepid spirits have
ventured into the non-linear domain. This work has been reviewed by Acheson
(1978a) and Roberts (1978); it is of particular relevance to the hydromagnetic dynamo
problem, but lies beyond the scope of this review.

11.4. Magnetic buoyancy


We saw in § 2 that the density is, in general, affected by fluctuations in both temperature
-
and magnetic pressure. When To/pAT 1 the appropriate Boussinesq equation of
state is (2.10) rather than (2.11) and magnetic buoyancy becomes significant. As a
result, fluid elements with strong magnetic fields tend to float upwards, an effect that
is particularly important for isolated flux ropes (Parker 1979). Instabilities driven by
magnetic buoyancy are important when the scale of variations in the direction of the
imposed horizontal field is much larger than the layer depth. Then the energy equation
Magnetoconvection 1373

has to be modified and (2.13) is replaced by

(11.4)

(Spiegel and Weiss 1982). However, magnetic buoyancy does not significantly affect
the solutions for magnetoconvection that have been described in previous sections.
The stability of a stratified horizontal field has been*studied in considerable detail
(Acheson 1978b, 1979 and references therein). The field is stably stratified if
g V / BI < 0 but instabilities driven by magnetic buoyancy may occur when the field
strength decreases upwards. Since we are concerned with thermally driven motion
we should discover how a stratified magnetic field abets or inhibits the onset of
convection. The full three-dimensional problem is complicated: hydromagnetic waves
travelling along the field may be destabilised by thermal or magnetic buoyancy.
However, if the field is sufficiently strong we might expect convection to occur in
rolls parallel to the undisturbed magnetic field. In the absence of motion let there be
a field of the (dimensional) form
B =B ( z ) j B = Bo(1+ z/HB) He >> d (11.5)
and let us suppose that the velocity is confined to the xz plane and independent of
y . Then the field lines remain straight but are advected with the motion and a
non-dissipative gaseous layer is unstable to interchanges if

(11.6)

where y is the ratio of the principal specific heats (Acheson 1979, Spiegel and Weiss
1982); this condition can alternatively be expressed as

(11.7)

where H, = -(d In p/dz)-'.


When diffusion is present it can readily be shown that the governing equations
are identical in form to those describing two-dimensional thermosolutal convection
(Spiegel and Weiss 1982). Magnetic buoyancy leads to double-diff usive behaviour of
the kind that produces layering in experiments with a stabilising solute gradient (Turner
1973). There is a simple bifurcation from the static solution when

(11.8)

but for [ < 1 and Bi/HBsufficiently large convection first sets in via a Hopf bifurcation
when

(11.9)
ugd
Thus the bifurcation structure for this problem is reminiscent of that described in § 7.

11.5. Compressibility
The double-diffusive aspects of Boussinesq magnetoconvection are of most interest
when [ < 1. In astrophysical situations, where 4' is small, compressibility cannot be
1374 M R E Proctor and N 0 Weiss

ignored. In the absence of a magnetic field a non-dissipative perfect gas is stably


stratified provided that the temperature gradient is subadiabatic. A necessary and
sufficient condition for stability is provided by the Schwarzschild criterion

(I. 1.10)

(Spiegel 1972). This criterion is modified by the imposition of a magnetic field Bo$;
a sufficient condition for stability is (in dimensional form)

(11.11)

(Gough and Tayler 1966). In order to proceed to the Boussinesq limit, three para-
meters must become small; these are d/H,, the Mach number M = U / V s ,where V s
is the sound speed, and the AlfvCnic Mach number

MA E VA/ Vs = ( 2 / ~ p ) ~ ” VA = B o ( P Ob)-1’2. (11.12)


Setting M i << 1in (11.6) yields as a necessary condition for instability that the potential
temperature gradient

(11.13)

If, in addition, the layer depth is much smaller than a scale height a sufficient condition
for instability is that

(11.14)

from (5.29). As one might expect, it is much more difficult to produce convection in
a very shallow layer.
Almost all direct investigations of compressible magnetoconvection are confined
to linear theory (so that M<c 1). It is convenient to discuss separately the effects of
stratification and of increasing the AlfvCn speed. We shall first consider the case when
d << H p and M A is allowed to vary. A uniform medium permeated by a vertical
magnetic field can support three different types of wave: these are the pure AlfvCn
wave (with a purely toroidal velocity field) and the fast and slow magnetoacoustic
waves (Cowling 1976a). In the limit MA<<1 the fast magnetoacoustic wave is longi-
tudinal and travels isotropically with speed Vs,while the slow wave travels along the
field lines with speed VA. The latter may be destabilised by a superadiabatic tem-
perature gradient (Kato 1966, Cowling 1976b). These unstable modes correspond to
overstable oscillations in the Boussinesq approximation. When M A >> 1, on the other
hand, the fast mode travels isotropically with the AlfvCn speed and the slow mode
travels along the field lines with speed V s (these oscillations are essentially sound
waves that use the flux tubes as waveguides). Once again, the slow mode is destabilised
by superadiabatic stratification.
Overstability is caused by thermal diffusion (Cowling 1976a), and we can simplify
the discussion by supposing that v = 77 = O((T = t = 0). Then in terms of the modified
parameters
?=or 4 = C+tq (11.15)
Magnetoconvection 1375

(which do not depend on v or 7 )we find that in the Boussinesq limit


f'"' -- 0 f(Ij -4 -+ CO (11.16)
so that there is instability for all positive f (Thompson 1951, Danielson 1961). Note
that thermal diffusion allows overstable oscillations where, from (11.8), non-dissipative
layers would be stable. Kat0 (1966) studied the effect of increasing MA. For finite
M A overstability is only possible if P exceeds a certain critical value that increases
linearly with 4, and 7") is increased. Thus overstability is inhibited when M A is
increased. (Note that Kato (1966) applies a local analysis; Saito and Kat0 (1968)
discuss the effect of imposing boundary conditions.)
When d = O(H,) the effects of stratification become significant. The onset of
convection in a field-free polytropic atmosphere has been studied in some detail
(Spiegel 1965, Gough et a1 1976, Graham and Moore 1978). In the presence of a
magnetic field the slow magnetoacoustic modes are modified by stratification. The
classification and behaviour of these modes have been explored by Antia and Chitre
(1979). The corresponding problem with a horizontal field is treated by Antia et a1
(1978a). When MA>> 1 the slow magnetoacoustic mode behaves like a vertically
propagating sound wave in a polytropic atmosphere (Lamb 1932). Such waves can
be thermally destabilised (Spiegel 1964, Jones 1976, Antia et a1 1978b). The associ-
ated magnetic problem is discussed by Syrovatskii and Zhugzhda (1967) and Zhugzhda
(1970).
Little is known about non-linear compressible magnetoconvection, when the Mach
number M becomes finite. The treatment of Oberbeck convection described in § 4
can be extended to the compressible case (Cattaneo, in preparation). It is no longer
clear what limits the field strength in isolated flux sheets. Thin tubes, unaffected by
external convection, have received considerable attention (e.g. Parker 1979, Spruit
1981). These slender tubes are liable to an instability which leads to partial evacuation
and enhancement of the field strength (Roberts and Webb 1978, Spruit 1979, Spruit
and Zweibel 1979, Unno and Ando 1979). We conjecture that a proper numerical
treatment of the non-linear problem would show local concentrations of flux between
convection cells with peak fields B * approaching the pressure balancing limit ( 2 ~ ~ p ) " ~ ,
for M of the order of unity.

12. Astrophysical implications

This review has naturally been slanted towards those problems for which specific
results have been obtained, with an emphasis on the available non-linear solutions.
This has led us to focus on idealised geometries, where branches of oscillatory and
steady solutions can be followed from their initial bifurcations well into the non-linear
domain. In a real, three-dimensional world the two-dimensional and axisymmetric
solutions are almost certainly unstable to three-dimensional perturbations (Galloway
and Weiss 1981) though no proper analysis has yet been carried out. In this concluding
section we shall describe what is conjectured about three-dimensional, turbulent
magnetoconvection and its relationship to fields that are observed upon the surface
of the Sun. Inevitably, the discussion is more qualitative.
The statistical properties of turbulent magnetic fields have been investigated in
considerable detail (Moffatt 1978, Parker 1979). More recently, the development of
intermittent field structures has been studied in extensive numerical calculations
1376 M R E Proctor and N 0 Weiss

(Pouquet and Patterson 1978, Orszag and Tang 1979, LCorat et al 1981, Meneguzzi
et al 1981). The computations of Meneguzzi et a1 (1981) show that in an incompressible
turbulent flow driven by random (non-thermal) forces the field becomes extremely
intermittent at high magnetic Reynolds numbers. Figure 41 shows the regions where
/ B (is within 5 % of its.maximum value for R,= 100; the detailed structure changes
with time but remains more intermittent than the corresponding vorticity. In a
turbulent convecting layer one would therefore expcct magnetic flux to be confined
to ropes that wander between convection ‘cells’ that last for approximately one
turnover time. If these ropes are dynamically active, motion will be excluded from
them and they will affect the pattern of convection in such a way that they remain at
the interfaces between adjacent cells, where flux can be transferred from one rope to
another (Galloway and Weiss 1981). This situation can to some extent be modelled
by supposing that there is a separation of scales between the main convective eddies
and the small-scale (non-helical) turbulence, so that the effects of the latter can be
parametrised by introducing an eddy diffusivity i j into the induction equation (2.4),
with a corresponding magnetic Reynolds number d,==200. Then most of the flux

Figure 41. Intermittency in turbulent magnetic fields. Regions for which IBI has more than 95% of its
maximum value are shaded, in a numerical simulation of three-dimensional time-dependent magneto-
hydrodynamic turbulence with R, = 100 and v = 7) (from Meneguzzi et a / 1981).
Magnetoconvection 1377

will end up in isolated ropes that are separated from the weaker small-scale fields;
the properties of the latter have been explored by Knobloch and Rosner (1981).
Such speculations cannot be checked by comparison with experiments. Fortu-
nately, the Sun provides an excellent laboratory where turbulent magnetic fields can
be observed. At the photosphere, convection occurs predominantly in cells with
diameters of about 1000 km (the solar granulation). Recent observations have
confirmed that almost all magnetic flux is confined to isolated tubes. Within these
tubes the fields are so intense that the magnetic pressure approaches the external gas
pressure and greatly exceeds the dynamic pressure of convection (Stenflo 1977). The
problems raised by such filamentary magnetic fields go well beyond the range of the
theoretical results in earlier sections of this review (Parker 1979, 1981, Spruit 1981)
but it is clear that any successful description must include the effects of the convective
cells in which the fields are embedded (Weiss 1982).
The small-scale intermittent fields are apparently a by-pr-duct of larger-scale fields
that are responsible for the solar cycle. The large-scale fields also emerge through
the photosphere in isolated ropes, which may assemble to form sunspots, with
diameters of up to 40 000 km. Sunspots are cool because normal (overturning)
convection is suppressed within them; the energy that is radiated from their umbrae
must be carried by relatively inefficient oscillatory convection. Thus the determination
of Rminor Q,, in § 7 may be regarded as an attempt to answer the question: what
critical field strength is needed to form a pore or sunspot? The results are crudely
consistent with the field strengths observed. There is strong but indirect evidence
from observations that magnetic activity is a typical feature of stars like the Sun, with
deep outer convective zones (Noyes 1981). The fields that emerge in active regions
are generated by a dynamo process, operating deep in the convective zone. Simplified
dynamo models can easily be constructed but the structure of the large-scale field is
largely determined by its interaction with convection. Any detailed theory must
include the effects of rotation, magnetic buoyancy and compressibility, with mag-
netoconvectioa as the most significant ingredient.

References

Acheson D J 1978a Rotating Fluids in Geophysics ed P H Roberts and A M Soward (New York: Academic)
p 315
-1978b Phil. Trans. R. Soc. 289 459
-1979 Solar Phys. 62 23
Antia H K and Chitre S M 1979 Solar Phys. 63 67
Antia H M, Chitre S M and Gokhale M H 1978a Solar Phys. 60 31
Antia H M, Chitre S M and Kale D M 1978b Solar Phys. 56 275
Arter W 1982 Geophys. Astrophys. Fluid Dyn. to be submitted
Batchelor G K 1956 J. Fluid Mech. 1 177
Busse F H 1975 J. Fluid Mech. 71 193
Busse F H and Riahi N 1980 J. Fluid Mech. 96 243
Chandrasekhar S 1952 Phil. Mag 43 501
-1961 Hydrodynamic and Hydromagnetic Stability (Oxford: Clarendon)
Chapman C J, Childress S and Proctor M R E 1980 Earth Planet. Sci. Lett. 51 362
Chapman C J and Proctor M R E 1980 J. Fluid Mech. 101 759
Childress S 1979 Phys. Earth Planet. Interiors 20 172
Clark A 1964 Phys. Fluids 7 1299
-1965 Phys. Fluids 8 644
__ 1966 Phys. Fluids 9 485
1378 M R E Proctor and N 0 Weiss

Clark A and Johnson A C 1967 Solar Phys. 2 433


Christopherson D G 1940 Q. J. Math. 11 63
Coullet P 1981 Woods Hole Oceanographic Institution Rep. 81-102, p 94
Cowling T G 1953 The Sun ed G P Kuiper (Chicago: University of Chicago Press) p 532
-1976a Magnetohydrodynamics (Bristol: Adam Hilger Ltd)
-1976b Mon. Not. R. Astron. Soc. 177 409
Da Costa K, Knobloch E and Weiss N 0 1981 J. Fluid Mech. 109 25
Danielson R E 1961 Astrophys. J. 134 289
Depassier M C and Spiegel E A 1981 Astron. J. 86 496
Feigenbaum M J 1978 J. Star. Phys. 19 25
-1979 J. Stat. Phys. 21 669
Galloway D J 1976 Problems of Stellar Convection ed E A Spiegel and J-P Zahn (Berlin: Springer-Verlag)
p 188
__ 1978 Mon. Not. R. Astron. Soc. 184 49P
Galloway D J and Moore D R 1979 Geophys. Astrophys. Fluid Dyn. 12 73
Galloway D J and Proctor M R E 1982 Planetary and Stellar Magnetism ed A M Soward (New York:
Gordon and Breach)
Galloway D J, Proctor M R E and Weiss N 0 1977 Nature 266 686
-1978 J. Fluid Mech. 87 243
Galloway D J and Weiss N 0 198 1 Astrophys. J. 243 945
Gibson R D 1966 Proc. Camb. Phil. Soc. 62 287
Gough D 0 1969 J. Atmos. Sci. 26 448
Gough D 0, Moore D R, Spiegel E A and Weiss N 0 1976 Astrophys. J. 206 536
Gough D 0 and Tayler R J 1966 Mon. Not. R. Astron. Soc. 133 85
Graham E and Moore D R 1978 Mon. Not. H.Astron. Soc. 183 617
Guckenheimer J 1981 Dynamical Systems and Turbulence ed D Rand and L-S Young (Berlin: Springer-
Verlag)
Guckenheimer J and Knobloch E 1982 to be published
Hewitt J M, McKenzie D P and Weiss N 0 1980 Earth Planet. Sci. Lett. 51 370
Hopf E 1942 Ber. Math. Phys. Sachs. A k a d . Wiss. 94 1
Huppert H E and Moore D R 1976 J. FluidMech. 78 821
Hurle D T J, Jakeman E and Pike E R 1967 Proc. R. Soc. A 296 469
Iooss G and Joseph D D 1980 Elementary Stability and Bifurcation Theory (Berlin: Springer-Verlag)
Jones C A 1976 Mon. Nor. R. Astron. Soc. 176 145
Jones C A, Moore D R and Weiss N 0 1976 J. Fluid Mech. 73 353
Joseph D D 1981 Hydrodynamic Instabilities and the Transition to Turbulence ed H L Swinney and J P
Gollub (Berlin: Springer-Verlag) p 27
Kat0 S 1966 Publ. Astron. Soc. Japan 18 201
Knobloch E and Proctor M R E 1981 J. Fluid Mech. 108 291
Knobloch E and Rosner R 1981 Astrophys. J. 247 300
Knob1och.E and Weiss N 0 1981 Phys. Lett. 85A 127
Knobloch E, Weiss N 0 and Da Costa L 1981 J. Fluid Mech. 113 153
Lamb Sir H 1932 Hydrodynamics (Cambridge: Cambridge University Press)
LBorat J, Pouquet A and Frisch U 1981 J. Fluid Mech. 104 419
Liang S F, Vidal A and Acrivos A 1969 J. Fluid Mech. 36 239
Lorenz E N 1963 J. Atmos. Sci. 20 130
Malkus W V R 1964 Geophys. Fluid Dyn. vol 1, Woods Hole Oceanographic Institution Rep. 64-46, p 1
Malkus W V R and Veronis G 1958 J. Fluid Mech. 4 225
Meneguzzi M, Frisch U and Pouquet A 1981 Phys. Rev. Lett. 47 1060
Moffatt H K 1978 Magnetic Field Generation in Electrically Conducting Fluids (Cambridge: Cambridge
University Press)
Moffatt H K and Kamkar S 1982 Planetary and Stellar Magnetism ed A M Soward (New York: Gordon
and Breach)
Moore D R and Weiss N 0 1973 J. Fluid Mech. 58 289
Nakagawa Y 1955 Nature 175 417
__ 1957 Proc. R. Soc. A 240 108
Noyes R S 1981 Solar Phenomena in Stars and Stellar Systems ed R M Bonnet and A K Dupree (Dordrecht:
D Reidel) p 1
Orszag S and Tang C-H 1979 J. Fluid Mech. 90 129
Magnetoconvection 1379

Parker E N 1963 Astrophys. J. 138 552


-1979 Cosmical Magnetic Fields-their origin and their activity (Oxford: Clarendon)
-1981 Solar Phenomena in Stars and Stellar Systems ed R M Bonnet and A K Dupree (Dordrecht: D
Reidel) p 33
Parker R L 1966 Proc. R. Soc. A 291 60
Peckover R S and Weiss N 0 1978 Mon. Not. R. Astron. Soc. 182 189
Pedlosky J 1978 Geophys. Fluid Dyn. vol 1 Woods Hole Oceanographic Institution Rep. 78-67, p 110
Poincar6 H 1885 Acta Math. 7 259
Pouquet A and Patterson G S 1978 J. Fluid Mech. 85 305
Proctor M R E 1977 J. Fluid A4ech. 82 97
_- 1981 J. Fluid Mech. 113 469
Proctor M R E and Galloway D J 1979 J. Fluid Mech. 90 273
Proctor M R E and Weiss N 0 1978 Rotating Fluids in Geophysics ed P H Roberts and A M Soward
(New York: Academic) p 389
Rayleigh (Lord) 1916 Phil. Mug. 32 529
Roberts B and Webb A R 1978 Solar Phys. 56 5
Roberts P H 1967 A n Introduction to Magnetohydrodynamics (London: Longmans)
-1978 Rotating Fluids in Geophysics ed P H Robert and A H Soward (New York: Academic) p 421
Roberts P H and Stewartson K 1977 Astron. Nachr. 298 311
Saito M and Kat0 S 1968 Solar Phys. 3 531
Spiegel E A 1964 Astrophys. J. 141 1068
-1965 Astrophys. J. 141 1068
__ 1972 Ann. Rev. Astron. Astrophys. 10 261
Spiegel E A and Veronis G 1960 Astrophys. J. 131 442
Spiegel E A and Weiss N 0 1982 Geophys. Astrophys. Fluid Dyn. in press.
Spruit H C 1979 Solar Phys. 61 373
-1981 The Sun as a Star ed S D Jordan (NASA/CNRS)
Spruit H C and Zweibel E H 1979 Solar Phys. 62 15
Stenflo J 0 1977 Basic Mechanisms of solar activity ed V Bumba and J Kleczek (Dordrecht: D Reidel) p
69
Syrovatskii J and Zhugzhda Y D 1967 Astron. Z h . 44 1180 (Engl. transl. 1968 Sou. Astron. 11 945)
Thompson W B 1951 Phil. Mag. 42 1417
Turner J S 1973 Buoyancy Efects in Fluids (Cambridge: Cambridge University Press)
Unno W and Ando H 1979 Geophys. Astrophys. Fluid Dyn. 12 107
Van der Borght R, Murphy J S and Spiegel E A 1972 Austr. J . Phys. 25 703
Velarde M G and Perez Cordon R 1976 J. Physique 37 177
Veronis G 1959 J. Fluid Mech. 5 401
-1965 J. Marine Res. 23 1
-1966 J. Fluid Mech. 4 545
Wal6n C 1949 On the Vibratory Rotation of the Sun (Henrik Lindstihls Bokhandel)
Weiss N 0 1964 Phil. Trans. R. Soc. A 256 99
-1966 Proc. R. Soc. A 293 310
-1981a J. Fluid Mech. 108 247
-1981b J. Fluid Mech. 108 273
-1981c J. Geophys. Res. 86b 11689
-1982 Solar activity ed C Jordon (Oxford: EPS) p 35
Zhugzhda Y D 1970 Astron. Z h . 47 340 (Engl. transl. 1971 Sou. Astron. A J 14 274)

You might also like