Geophysical Journal International

Geophys. J. Int. (2013) 192, 537556 doi: 10.1093/gji/ggs051

G
J
I
G
e
o
m
a
g
n
e
t
i
s
m
,
r
o
c
k
m
a
g
n
e
t
i
s
m
a
n
d
p
a
l
a
e
o
m
a
g
n
e
t
i
s
m
Flow throughout the Earths core inverted from geomagnetic
observations and numerical dynamo models
Julien Aubert
Dynamique des Fluides G eologiques, Institut de Physique du Globe de Paris, Sorbonne Paris Cit e, Universit e Paris-Diderot, UMR 7154 CNRS, 1 rue
Jussieu, F-75005 Paris, France. E-mail: [email protected]
Accepted 2012 October 30. Received 2012 October 22; in original form 2012 September 4
SUMMARY
This paper introduces inverse geodynamo modelling, a framework imaging ow throughout
the Earths core from observations of the geomagnetic eld and its secular variation. The
necessary prior information is provided by statistics from 3-D and self-consistent numerical
simulations of the geodynamo. The core method is a linear estimation (or Kalman ltering)
procedure, combined with standard frozen-ux core surface ow inversions in order to han-
dle the non-linearity of the problem. The inversion scheme is successfully validated using
synthetic test experiments. A set of four numerical dynamo models of increasing physical
complexity and similarity to the geomagnetic eld is then used to invert for ows at single
epochs within the period 19702010, using data from the geomagnetic eld models CM4 and
gufm-sat-Q3. The resulting core surface ows generally provide satisfactory ts to the secu-
lar variation within the level of modelled errors, and robustly reproduce the most commonly
observed patterns while additionally presenting a high degree of equatorial symmetry. The cor-
responding deep ows present a robust, highly columnar structure once rotational constraints
are enforced to a high level in the prior models, with patterns strikingly similar to the results
of quasi-geostrophic inversions. In particular, the presence of a persistent planetary scale,
eccentric westward columnar gyre circling around the inner core is conrmed. The strength of
the approach is to uniquely determine the trade-off between t to the data and complexity of
the solution by clearly connecting it to rst principle physics; statistical deviations observed
between the inverted ows and the standard model behaviour can then be used to quantitatively
assess the shortcomings of the physical modelling. Such deviations include the (i) westwards
and (ii) hemispherical character of the eccentric gyre. A prior model with angular momentum
conservation of the coremantle inner-core system, and gravitational coupling of reasonable
strength between the mantle and the inner core, is shown to produce enough westward drift
to resolve statistical deviation (i). Deviation (ii) is resolved by a prior with an hemispherical
buoyancy release at the inner-core boundary, with excess buoyancy below Asia. This latter re-
sult suggests that the recently proposed inner-core translational instability presently transports
the solid inner-core material westwards, opposite to the seismologically inferred long-term
trend but consistently with the eccentricity of the geomagnetic dipole in recent times.
Key words: Inverse theory; Dynamo: theories and simulations; Rapid time variations.
1 I NTRODUCTI ON
It has long been recognized that the variation of the Earths magnetic eld in time (known as the secular variation) provides a powerful probe
to infer the ow of electrically conducting uid at the top of the outer core, hence imaging part of the motion responsible for the geodynamo.
The rst decades of active research on the topic, summarized in the review of Bloxham & Jackson (1991), concentrated on laying down the
theoretical bases and numerical tools to handle the large amount of underdetermination of this core surface ow problem. At this point it was
acknowledged that, on account of the fact that the mantle is mostly insulating and the core highly conductive, magnetic eld measurements
can be continued downward from the Earth surface down to the core surface, but not further below, thus largely hindering our ability to image
ow throughout the core. Such an achievement would certainly be highly desirable to advance the general understanding of the geodynamo,
evaluate the state of its couplings with the adjacent layers (the inner core and the mantle), and properly initialize geomagnetic data assimilation
algorithms (see Fournier et al. 2010, for a review). Over the last two decades, the improved understanding of the magnetohydrodynamics
C
The Author 2012. Published by Oxford University Press on behalf of The Royal Astronomical Society. 537

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

538 J. Aubert
of rapidly rotating uids has brought a number of interesting dynamic assumptions which could help to circumvent the issue of downward
continuing the velocity eld from below the core surface. Recently, Jault (2008) observed that rapid, diffusionless dynamic phenomena
developing within a rapidly rotating uid on the Alfv en wave timescale

A
=
D()
1/2
B
(1)
are remarkably columnar (invariant in the direction of the Earths rotation axis) if
A
is much longer than the Earth inverse rotation rate

= 1/ (here D, and are, respectively, the outer-core thickness, magnetic permeability and density, and B is the typical magnetic
eld strength within the outer core). Recent independent estimates of B favouring a strength of several milliteslas (Aubert et al. 2009;
Buffett 2010), yield
A
on the order of a few years, possibly consistent with a 6-yr oscillation in the length-of-day (l.o.d.) variations (Gillet
et al. 2010) and the ratio =

/
A
(the Lehnert number) on the order of 10
4
. On that basis, it has been proposed (Pais &
Jault 2008; Gillet et al. 2009) that for ow on interannual to decadal timescales, a columnar downward continuation of the velocity eld
can be performed within the quasi-geostrophic framework formulated four decades ago (Busse 1970) and which has found increasingly
compelling numerical implementations from 1990 onwards (see a recent review by Finlay et al. 2010). These inversions account for the lateral
motion of coremantle boundary geomagnetic ux patches (e.g. Bloxham & Jackson 1991; Holme 2007; Finlay et al. 2010) by revealing a
planetary-scale, westwards, columnar eccentric gyre moving closer to the coremantle boundary in the Atlantic Hemisphere, and recessing
towards the inner-core boundary in the Pacic Hemisphere, where evidence of geomagnetic westward drift is more elusive. Such a gyre is
intriguing because it does not have an equivalent in the solutions so far produced by 3-D direct numerical dynamo modelling. The gyre is
robustly obtained over the last half century (Gillet et al. 2009), and may have been stable throughout the historical geomagnetic period, for as
long as the westward drift itself (Finlay et al. 2010). Such a persistence naturally raises the concern of the validity of the quasi-geostrophic
framework for timescales much longer than
A
, particularly the core overturn time
U
=D/U, where U is a typical core ow velocity. On these
timescales, thermal and magnetic anomalies have enough time to rearrange such as to respect thermal and magnetic wind equilibria which
can disrupt quasi-geostrophy (e.g. Aubert 2005). One of the most widely conjectured manifestations of this effect is for instance the existence
of polar vortices (Olson & Aurnou 1999; Sreenivasan & Jones 2005) within the tangent cylinder (the axial cylinder encircling the inner
core).
Interestingly, modern 3-D, self-consistent simulations of convective dynamos seem to generally have strengths and weaknesses nicely
complementing quasi-geostrophic models. Their large-scale magnetic eld output can be strikingly similar to the geomagnetic eld (Chris-
tensen et al. 2010) provided that three timescales are brought in reasonable proportion with respect to their Earth counterparts: the core
overturn time
U
, the Earth inverse rotation rate

and the magnetic dissipation time

= D
2
/, where is the magnetic diffusivity of the
outer-core liquid iron. Furthermore, numerical dynamos respecting these conditions also adequately render a surprisingly broad temporal
spectrum of geomagnetic variations (Christensen et al. 2012; Olson et al. 2012) around the overturn timescale
U
, thus suggesting that a
large part of the temporal geomagnetic power spectrum is affected by advective transport phenomena. Unfortunately, although the magnetic
Reynolds number Rm =

/
U
can be brought to the correct value pertaining to the Earths core of about 1000, the magnetic Ekman number
E

is currently limited to values of about 10

5
while the core value is on the order of 10
9
(see Christensen 2011, for a recent
review). As a consequence, numerical dynamos presently considered as Earth-like still lack the short timescale (interannual to decadal)
rotational dynamics precisely targeted by quasi-geostrophic models. Another connected question is whether the spatial and temporal similarity
of large-scale magnetic elds between models and Earth automatically implies that large-scale simulated ows are realistic. These ows are
indeed generated by convective instabilities at large scales, a mechanism which is very different from large-scale ows being maintained by
small-scale instabilities and trans-scale energy transfer in the Earths core.
The recent emergence of an inverse problem framework for numerical dynamos (Aubert & Fournier 2011; Fournier et al. 2011) and the
excellent match of selected models to the geomagnetic eld prompted the idea to use velocity statistics from Earth-like numerical dynamos
as a statistical prior in the underdetermined inversion for ow throughout the core. This forms goal (1) of the present work. Subsequent
goals are: (2) investigate the general ability of numerically simulated ows to quantitatively account for the geomagnetic secular variation,
(3) handle the underdetermination of the geomagnetic inverse problem in order to produce ow images in the entire volume of Earths core,
(4) assess the robustness of quasi-geostrophic ow images against this new approach, (5) highlight the parts of the ow signal requiring
additional modelling and nally (6) pave the way for time-dependent, prospective geomagnetic data assimilation algorithms.
In an initial approach of the subject (Aubert & Fournier 2011; Fournier et al. 2011), we have shown that good recovery of the internal
structure of numerical dynamos could be achieved by applying the theory of linear estimation in weakly non-linear cases. In the generally non-
linear context of the core owproblem, the linear estimation has a limited applicability and works only if we solve for small increments, starting
from initial guesses not too far from the nal solution. Although such a situation is expected to happen routinely in a properly implemented
data assimilation algorithm (Fournier et al. 2010), the application is difcult for a single-epoch inversion, or for the rst inversion of a data
assimilation sequence, since our best initial guesses are the time average elds of the dynamo simulation and the increments are expected to be
as large as these guesses (see for instance g. 3 in Fournier et al. 2011). Here, the approach is thus modied to resort to classical, frozen-ux
linearized inversions of the core surface ow, constrained by the numerical dynamo prior statistics, before resorting to linear estimation of
the deep ow. Models and methods are presented in Section 2. Section 3 presents the main results, which are discussed in Section 4.

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 539
2 MODELS AND METHODS
2.1 Numerical dynamos and geomagnetic eld models
The numerical dynamo model solves for Boussinesq convection and magnetic induction in the magnetohydrodynamic approximation in a
rapidly rotating spherical shell between radii r
i
and r
o
with the present Earths core aspect ratio r
i
/r
o
=0.35. A complete physical description
can be found in Aubert et al. (2009) and Aubert & Fournier (2011). The numerical implementation PARODY-JA is used (Dormy et al. 1998;
Aubert et al. 2008). Four cases have been calculated (Table 1). Models 1 and 2 originate from Aubert & Fournier (2011) and use rigid and
electrically insulating boundaries. The distribution of buoyancy sources aims at mimicking chemical convection (the dominant source of
buoyancy in the Earths core at present, see Aubert et al. 2009). The practical implementation comprises an imposed, spatially homogeneous
buoyancy ux at the inner boundary, zero buoyancy ux at the outer boundary, a sink term in the volume. Model 3 is similar to model 2,
but has a stress-free outer boundary, an axially rotating, conducting inner core (with same conductivity as the outer core) subject to viscous
and magnetic torques from the outer core, and to a gravitational torque from the mantle with expression (
ic

m
), where
ic
and

m
are, respectively, the inner-core and mantle rotation rate, and the coupling constant is as described in Aubert & Dumberry (2011).
The total (inner core, outer core and mantle) angular momentum is preserved (the mantle is also axially rotating under the inuence of the
opposite gravitational torque from the inner core). Model 4 is similar to model 2 but has an heterogeneous, longitudinally hemispherical
imposed buoyancy ux at the inner-core boundary, in addition to the homogeneous buoyancy ux. This conguration simulates the dynamic
consequences (Olson & Deguen 2012) of the inner-core translational instability (Alboussiere et al. 2010; Monnereau et al. 2010). Further
details about the numerical models are reported in Table 1.
Following our previous studies (Aubert & Fournier 2011; Fournier et al. 2011), the non-dimensional model output is scaled back to the
dimensional world using either canonical units, or units underlain by scaling principles known (or thought) to hold both in the numerical model
and in the Earths core. This is done in order to physically rationalize the parameter gap between the numerical model and the geodynamo.
The canonical length unit is used, the non-dimensional shell gap r
o
r
i
being assigned the value 2260 km. Owing to the homogeneity of
the magnetic induction equation (see Section 2.2.1), the magnetic eld unit is irrelevant here and only a time unit is needed in addition to
the length scale. Here the non-dimensional secular variation timescale of the model (Lhuillier et al. 2011) is assigned the Earth dimensional
value
SV
= 415 yr. The 50 yr uncertainty range reported by Lhuillier et al. (2011) only has a minor effect on the ow inversion results
presented in Section 3. Since the magnetic Reynolds number of models 24 is very close to the value of about 800 which is expected in the
Earths core (Christensen & Tilgner 2004), the choice of an advective timescale for rescaling the time axis is equivalent to the choice of a
timescale based on magnetic diffusion. Finally, for the sole illustrative purpose of comparing the simulated magnetic eld spectral properties
to the geomagnetic eld, the non-dimensional, convective-power based scaling prediction (Christensen & Aubert 2006) for the magnetic eld
amplitude in the models is adjusted to the dimensional value [B] = f
1/2
ohm
(
3
p
2
D
2
)
1/6
= 1.7 mT predicted by a high-power present core state
(Aubert et al. 2009). In the previous formula, p is the convective power density in the outer core and f
ohm
the ohmic dissipation fraction of
this power.
The output of the numerical models has been compared to magnetic eld and secular variation spectral coefcients obtained from the
geomagnetic eld models CM4 (Sabaka et al. 2004) and gufm-sat-Q3 (Finlay et al. 2012). For the satellite era, this latter model is chosen
because of its focus on extracting the geomagnetic signal of core origin. The magnetic eld output of models 2 and 3 is morphologically fairly
similar to the geomagnetic eld, as attested by the low
2
values reported in Table 1, and as illustrated in Fig. 1. The output also includes
equatorial magnetic patches of normal polarity, created by ux expulsion from radial equatorial upwellings. The patches do not have any
preferential drift direction in model 2. The choice of boundary conditions in model 3 produces a sizeable equatorial zonal ow below the
outer boundary (see Fig. 11 in Section 3.2.3), transporting equatorial magnetic ux patches in the westward direction at a speed comparable
Table 1. Properties of the numerical model cases. Models 1 and 2 correspond to the models fully described in Aubert
&Fournier (2011). Denitions for the left panel input parameters can also be found in that study: the Rayleigh number
Ra
Q
, and the viscous, magnetic and thermal Ekman numbers E, E

and E

. The right panel gives the values of the

output parameters, as dened in the introduction: the magnetic Reynolds number Rm, the Lehnert number , and the
measure
2
of morphological similarity to the geomagnetic eld (Christensen et al. 2010). Earth values are estimated
in Christensen &Aubert (2006); Aubert et al. (2009). Notes: (c) columnar ow, (am) angular momentumconservation
(see text), with an inner core to mantle gravitational coupling constant /D
3
= 5 10
4
as dened in Aubert
& Dumberry (2011) ( is the uid viscosity), (ep) presence of equatorial magnetic ux patches of normal polarity
at the outer boundary, (wdep) presence of westward-drifting equatorial ux patches of normal polarity at the outer
boundary, (ich) heterogeneous buoyancy release at the inner-core boundary with peak-to-peak magnitude at 80 per
cent of the homogeneous buoyancy ux, (eg) presence of an hemispherical, eccentric columnar gyre. The question
marks refer to features that are conjectured, but not directly observed in the Earths core.
Ra
Q
E E

Rm
2
Notes
Model 1 5.8 10
4
10
3
2.5 10
4
10
3
100 4 10
2
6
Model 2 2.7 10
5
3 10
5
1.2 10
5
3 10
5
858 1.4 10
2
1 c, ep
Model 3 2.7 10
5
3 10
5
1.2 10
5
3 10
5
982 1.3 10
2
0.8 c, am, wdep
Model 4 2.7 10
5
3 10
5
1.2 10
5
3 10
5
818 1.7 10
2
3 c, ich, eg
Earth O(10
13
) 3 10
15
3 10
9
O(10
15
) O(10
3
) O(10
4
) c?, eg?, wdep

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

540 J. Aubert
Figure 1. Coremantle boundary radial magnetic eld (a) and its rate of change (b) in a snapshot of model 2 and (c), (d) in the gufm-sat-Q3 geomagnetic eld
model (Finlay et al. 2012) for epoch 2001. Three levels of spherical harmonic truncation are presented for the model output. The geomagnetic eld models are
truncated at degree 13.
Figure 2. Energy spectra of the magnetic eld (a) and its secular variation (b) at the Earths surface, as function of the spherical harmonic degree. Two epochs
of the geomagnetic eld are represented (colour lines). The solid black line represents the average variance of the magnetic eld and secular variation of model
2. The dashed line corresponds to the snapshot shown in Fig. 1.
to the geomagnetic westward drift. Models 2 and 3 nally appear to provide a secular variation which is very similar to the geomagnetic
observations, including an important equatorial signal (with westward drift in the case of model 3). This good similarity is also attested by
comparing the magnetic eld and secular variation spectra of the models to the geomagnetic eld (Fig. 2), although the secular variation
spectra hint for model large scales (l = 1, 2) slightly underpowered with respect to the smaller scales when compared to the geomagnetic
data. In contrast, models 1 and 4 exhibit a less Earth-like magnetic eld morphology as attested by the larger
2
values in Table 1. Equatorial
magnetic eld dynamics is absent from model 1, and strongly reduced in model 4 with respect to its homogeneous counterpart model 2 (not
shown). It should nally be noted that none of the simulated magnetic elds present polarity reversals. This is not considered a problem given
the timescales on which the analysis is focused here.
While the velocity eld in model 1 signicantly deviates from columnar ow, snapshots from models 24 exhibit a fairly columnar
behaviour (see Fig. 6 in Section 3.1), with local deviations under the inuence of magnetic and thermal winds (most notably long-term polar
vortices in the tangent cylinder, see also Fig. 11 in Section 3.2.3). This observation and the Lehnert number values reported in Table 1 are
compatible with the condition 3 10
2
proposed by Jault (2008) for the presence of columnar ow in numerical dynamos. In contrast
with the other models, model 3 produces a westward zonal ow at equatorial position beneath the outer boundary (see Fig. 11). It should

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 541
nally be noted that none of the models 13 spontaneously generate eccentric columnar gyres, while model 4 does (see Fig. 14 in Section
3.2.4).
2.2 Inversion for core surface ows constrained by a numerical dynamo
The inversion proceeds in two stages: rst, the ow below the hydromagnetic boundary layer close to the coremantle boundary (the
near-surface ow) is obtained by inverting a linearized version of the non-linear frozen-ux induction equation. The second step is a linear
estimation (Aubert & Fournier 2011) of the deep ow from the near-surface ow, taking advantage of the strong linear couplings existing
between the two due to the inuence of the Coriolis force. This two-step formulation differs from the one-step approach undertaken in
Aubert & Fournier (2011) due to the need to handle the non-linearity of the core ow problem. It should also be noted that, in contrast to the
quasi-geostrophic approach, it does not necessarily involve a rigid enforcement of columnar ow at depth.
2.2.1 Surface ow
The magnetohydrodynamic induction equation is written as
B
t
= (u B) +
2
B, (2)
where t is time, B the magnetic eld and u the velocity eld. It is assumed (see Finlay et al. 2010, for justications) that magnetic diffusion
is of secondary importance owing to the rather large magnetic Reynolds number of the core, that the radial magnetic eld jump across the
thin viscous Ekman boundary layer close to the coremantle boundary is negligible, and that the radial velocity remains weak immediately
outside this boundary layer. The main source for the time variation of the outer boundary radial magnetic eld Br is then advection by the
lateral, near-surface velocity eld u
fs
immediately below the Ekman layer (Roberts & Scott 1965)
Br
t
=
H
(u
fs
Br) +. (3)
Here
H
is the horizontal divergence operator, and gathers all the secondary sources of secular variation which are neglected when adopting
the above assumptions, and when arbitrarily truncating the spectral representations of u
fs
and Br as
Br =

m
max
1
m m
max
1
|m| l l
max
1
Br
m
l
Y
m
l
(, ), (4)
u
fs
=
_
_
_
_
1
sin
T

+
S

+
1
sin
S

_
_
_
_
, (T, S) =

m
max
3
m m
max
3
|m| l l
max
3
(t
m
l
, s
m
l
)Y
m
l
(, ). (5)
Here l
max
1
= m
max
1
= 13 is the degree to which the magnetic eld and secular variation of core origin is assumed to be known, l
max
3
= m
max
3
= 30
is the degree to which core ow is expanded, Y
m
l
(, ) is the complex spherical harmonic function of degree l and order m on the unit sphere
mapped by the colatitude and the longitude . The lateral ow u
fs
is described by the toroidal and spheroidal potentials T and S. The
truncation level of the ow needs to be at least twice that of the data coefcients, in order to resolve the non-linear couplings present in (3).
It should however not greatly exceed that value because higher truncation levels result in increasingly costly inverse problem computations
(especially at depth) and because data provided up to degree and order 13 is not expected to constrain the additional ow coefcients.
Similarly to the approach presented in Whaler (1986), the direct problem (3) then writes, for m
max
1
m
1
m
max
1
, |m
1
| l
1
l
max
1
:
Br
m
1
l
1
t
=
1
r
o

m
max
1
m
2
m
max
1
|m
2
| l
2
l
max
1
Br
m
2
l
2

m
max
3
m
3
m
max
3
|m
3
| l
3
l
max
3
_
E
m
1
,m
2
,m
3
l
1
,l
2
,l
3
t
m
3
l
3
+ AG
m
1
,m
2
,m
3
l
1
,l
2
,l
3
s
m
3
l
3
_
+
m
1
l
1
, (6)
where r
o
is the coremantle boundary radius,
A =
1
2
[l
2
(l
2
+1) l
1
(l
1
+1) l
3
(l
3
+1)] , (7)

m
1
l
1
is the spherical harmonic expansion of the error , and E, G are the Elsasser and AdamsGaunt integrals
E
m
1
,m
2
,m
3
l
1
,l
2
,l
3
=
_
S
Y
m
1
l
1
_
Y
m
2
l
2

Y
m
3
l
3

Y
m
3
l
3

Y
m
2
l
2

_
dd, (8)

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

542 J. Aubert
G
m
1
,m
2
,m
3
l
1
,l
2
,l
3
=
_
S
Y
m
1
l
1
Y
m
2
l
2
Y
m
3
l
3
sin dd. (9)
These integrals relate to Wigner coefcients (James 1973) and are non-zero only for particular combinations of l
1, 2, 3
, m
1, 2, 3
respecting
triangular rules. They can be numerically evaluated in an efcient manner (here the algorithms and fortran routines detailed in Moon 1979,
are used). The direct problem (6) is non-linear, and linearized for each single epoch as
y = M(Br
m
l
)x
fs
+. (10)
Here y is the data vector containing the 196 complex coefcients Br
m
l
/t , m
max
1
m m
max
1
, |m| l l
max
1
. These reduce to N
d
= 195
real degrees of freedom, accounting for the facts that Br
m
l
and Br
m
l
are complex conjugates and that Br
0
0
= 0. The vector x
fs
is the state
vector containing the 1922 complex near-surface ow coefcients t
m
l
, s
m
l
, m
max
3
m m
max
3
, |m| l l
max
3
. Here again these reduce to
N
m
= 1920 real degrees of freedom. The matrix M contains the magnetic eld coefcients and coupling integrals. The inversion of Eq. (10)
usually implies additional assumptions to handle the underdetermination of this problem (since the number of unknowns N
m
is larger than
the number of equations N
d
), as well as regularizations to ensure that the solution is well-behaved from various point of views (see Finlay
et al. 2010, for a review). In the present approach, the statistics provided by the numerical dynamo model handle both the underdetermination
and regularization of the inversion, and provide a unique solution, clearly connected to the rst-principle physical content of the numerical
dynamo. This situation is arguably preferable to the use of weak regularization norms involving adjustable damping parameters used in
quasi-geostrophic core ow inversions (Pais & Jault 2008; Gillet et al. 2009). Furthermore, the mist norm associated to the statistical prior
can also be used to detect ow components which are compliant with or deviating from the prior assumptions. To these ends, the state vector
x
fs
is centered by removing a time average value x
fs
obtained fromthe numerical dynamo, such that x
fs
= x
fs
+ x
fs
. We then solve the problem
M x
fs
= y Mx
fs
. (11)
Introducing the statistical covariance matrix P
fs
for the components of the near-surface ow x
fs
, and the error covariance matrix R for the
components of the error , the stochastic inverse is (Aubert & Fournier 2011)
x
fs
= K
fs
(y Mx
fs
) , (12)
with the Kalman gain matrix
K
fs
= P
fs
M

_
MP
fs
M

+R
_
1
. (13)
Here the prime denotes the transpose complex conjugate. Similarly as in Fournier et al. (2011) and Aubert &Fournier (2011), the time average
state vector x
fs
and near-surface ow covariance matrix P
fs
are directly computed from a free run of the numerical dynamo model, by stacking
about 1000 decorrelated model snapshots and obtaining the rst and second statistical moments. In Aubert & Fournier (2011), it was shown
that the minimal temporal spacing between snapshots for obtaining a satisfactory decorrelation was the e-folding time
e
of the system. For
the priors computed here, the temporal spacing was set to 3
e
. The computation of P
fs
is performed up to spherical harmonic degree and
order 30, accordingly with the truncation level of the near-surface ow. The radius of the base of the hydromagnetic boundary layer is set at
200 km below the outer boundary in case 1, 80 km below the outer boundary in cases 2 and 4, and at the outer boundary in case 3 (since
this boundary is stress free). The structure of P
fs
is presented in Fig. 3. The block-diagonal structure is reminiscent of the results presented
in Aubert & Fournier (2011) and highlights the leading inuence of the Coriolis force, which linearly couples adjacent harmonic degrees l
within a block of constant harmonic order m, but does not couple distinct harmonic orders. Signicant linear couplings also exist between the
spheroidal and toroidal components of the ow, reecting the tight connection between surface ow rotation and upwelling usually observed
in numerical dynamos (see e.g. Aubert et al. 2008).
The error covariance matrix R is iteratively determined, following a procedure reminiscent from Pais & Jault (2008). The initial error
covariance matrix is assumed to be made up by a constant times the identity matrix. Then an initial velocity eld v is obtained, from which
the underparametrization error, or secular variation resulting from the inuence of the velocity eld with the unresolved magnetic eld, is
computed and added to the error arising from the neglect of magnetic diffusion and the observation error (these three sources are supposed
to be statistically independent and to overcome all the other sources)

m
l

2
=

[
H
(vBr
13<l30
))]
m
l

2
+

[e
r
B]
m
l

2
+
0
(l)
2
. (14)
Here Br
13 < l 30
is the small-scale part of the radial eld at the core-mantle boundary between degrees 14 and 30 (the contribution from higher
degrees has been checked to be negligible in all numerical dynamo models presented here), and e
r
B is the radial part of near-surface
magnetic diffusion. Both quantities are linearly estimated from the known part of the coremantle boundary magnetic eld and secular
variation, and the prior numerical model covariance properties, following the procedure outlined in Fournier et al. (2011) and Aubert &
Fournier (2011). Fig. 4 presents examples of such linear estimations. The data noise background
0
(l)
2
is set to correspond, at the Earth
surface, to a at energy spectrum per degree, at a level of 0.02 (nTyr
1
)
2
for gufm-sat-Q3 (Pais & Jault 2008) and 0.4 (nTyr
1
)
2
for CM4
(Gillet et al. 2009). From this single realization of the error
m
l
, an l-dependent statistical error model is constructed as
f (l) =
1
l +1
l

m=0
|
m
l
|
2
/(
lm
)
2
, (15)

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 543
Figure 3. Variance-normalized modulus of the covariance matrix P
fs
for the statistically centered toroidal and spheroidal components t
m
l
and s
m
l
of the
near-surface ow in model 2. Close-ups of selected regions are also enlarged.
Figure 4. (a) Radial magnetic eld at the core-mantle boundary in 2001 up to harmonic degree and order 30, linearly estimated from gufm-sat-Q3 provided up
to degree and order 13 and from the magnetic eld statistics of model 2, following the procedure outlined in Fournier et al. (2011); Aubert & Fournier (2011).
(b) Near-surface radial magnetic diffusion up to degree and order 13, linearly estimated from the same data and magnetic eld statistics as in (a).
where
lm
is the prior model variance for individual secular variation coefcients (This quantity is computed during the free dynamo model
run simultaneously with the ow statistics). The matrix R is nally updated as a diagonal matrix following this error model
R
m
l
m
l
= (
lm
)
2
f (l). (16)
The scheme is iterated until convergence (typically 10 iterations). The data t quality is measured through the normalized data mist

d
=
1
N
d
_
(y Mx
fs
)

R
1
(y Mx
fs
). (17)

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

544 J. Aubert
The energy of the inverse solution with respect to the statistical prior is evaluated through the normalized deviation to the prior model time
average

m
=
1
N
m
_
x

fs
P
1
fs
x
fs
. (18)
In the process of searching for the most probable solution given the data and the model statistics, the inversion (13) minimizes the functional
(Aubert & Fournier 2011)
J(x) = (y Mx
fs
)

R
1
(y Mx
fs
) + x

fs
P
1
fs
x
fs
. (19)
The inversion approximately balances each individual harmonic coefcient contribution in J. It is thus expected that for a correct t to the
data (
d
1), the deviation from the prior model time average will be
m

N
d
/N
m

d
< 1 since the spherical harmonic expansion of
the model exceeds that of the data by about a factor 10.
2.2.2 Deep ow
Once a near-surface ow inverse is obtained, a linear estimation of the corresponding deep ow is performed. The 3-D velocity vector
x =
_
t
m
l
(r
i
), . . . , t
m
l
(r
o
), s
m
l
(r
i
), . . . , s
m
l
(r
o
)
_
T
, (20)
where superscript T denotes the transpose, gathers the values of the statistically centred toroidal and spheroidal potentials of the velocity eld
u on the nodes of a numerical discrete grid (82 radial levels are used). The direct problem is written
H x = x
fs
, (21)
or, in an expanded form
_
0 . . . 0 1 0 . . . . . . . . . . . . 0
0 . . . . . . . . . . . . 0 1 0 . . . 0
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
t
lm
(r
i
)
.
.
.
t
lm
(r
E
)
.
.
.
t
lm
(r
o
)
s
lm
(r
i
)
.
.
.
s
lm
(r
E
)
.
.
.
s
lm
(r
o
)
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
_
t
lm
(r
E
)
s
lm
(r
E
)
_
. (22)
Here r
E
is the radius of the base of the hydromagnetic boundary layer and H is called the observation operator, containing ones in the cells
where the quantity is observed, and zeros otherwise. This vastly underdetermined problem is again solved under constraint of a numerical
dynamo prior model covariance matrix P cross-correlating all ow components (Aubert & Fournier 2011)
x = K x
fs
, (23)
with the Kalman gain matrix
K = PH

_
HPH

_
1
. (24)
The matrix P is computed using the same procedure as for P
fs
(which is a subset of P), and to the same spectral resolution. Its structure is
quite similar to the structures already described in detail in Fig. 3 of this study and g. 3 of Aubert & Fournier (2011), block-diagonal with
non-existent couplings between different m-values but signicant couplings between adjacent l-values within an m-block, and also exhibits
couplings between spheroidal and toroidal terms. Once the spheroidal and toroidal potentials are determined throughout the shell, the velocity
eld u can easily be determined after computing the poloidal potential p
lm
(r) through s
lm
(r) = [rp
lm
(r)]/rr.
3 RESULTS
3.1 Validation with synthetic experiments
In order to estimate the efciency of the whole core ow inversion procedure, we rst invert synthetic secular variation data, extracted and
realistically truncated from the model 2 snapshot presented in Fig. 1, and compare the recovered ow with the known reference velocity
eld of this snapshot. Table 2 summarizes the experiments performed. The recovery quality is quantitatively evaluated using the correlation

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 545
Table 2. Description of synthetic experiments and diagnostics for recovery quality. Correlation coefcients and pointwise
recovery factors are dened as in Amit et al. (2007), and evaluated between the recovery and reference truncated up to
degree 30 (rst number) and 13 (second number).
Exp. Correlation Pointwise
d

m
Notes
coefcient recovery factor
1 0.51, 0.70 0.4, 0.56 1.10 0.44 Twin experiment, actual small-scale Br and magnetic diffusion
are used for error computation
2 0.50, 0.69 0.4, 0.56 1.27 0.57 Same as (1), with linearly estimated errors, relevant
situation for the geomagnetic case
3 0.32, 0.38 0.25, 0.32 1.29 0.57 Same as (2), fraternal experiment using statistical
prior from model 1
Figure 5. Core surface owrecovery versus the reference for experiment 2. (a) Hammer projection of the near-surface reference ow(arrows, arbitrary scaling)
superimposed with a colour map of the near-surface toroidal scalar T (blue denotes a clockwise circulation when seen from the North pole). (b) same as (a),
for the recovered near-surface ow. The grey meridian locates the meridional cut drawn in Fig. 6. (c) Earth-surface energy spectrum of the secular variation
(black) as a function of the spherical harmonic degree. The residual left by the inversion (red), matches the sum (dashed red) of the modelled sources of error
(blue and green). (d) Energy spectrum of the near-surface reference and recovered ows, as functions of the spherical harmonic degree.
coefcient between the reference and the recovered ows, and the pointwise recovery factor already introduced in earlier similar tests (full
denitions in Amit et al. 2007). The synthetic data is considered noise-free, but subject to the already introduced underparametrization and
magnetic diffusion errors.
Experiments 1 and 2 are twin experiments, as they use as statistical prior the same model that served to produce the synthetic data
(but the snapshot under consideration is not part of the ensemble that produced the covariance matrices). Experiment 1 assumes perfect
knowledge of the small-scale magnetic eld and near-surface magnetic diffusion, while experiment 2 performs a linear estimation of these
hidden quantities from the known quantities (this is the realistic case). The near-surface ow recovery quality is fair in both cases (Table 2
and Fig. 5), especially at larger scales. The truncated geomagnetic data indeed do not constrain well the ow at harmonic degree larger than
15 (Fig. 5d). Data mists are close to 1, as expected from twin experiments, where data and model are compatible by construction. Deviations

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

546 J. Aubert
Figure 6. Deep ow recovery from the near-surface ow obtained in experiment 2, plotted against the reference ow. The grey lines locate the azimuth of the
meridional cut (same position also located in Fig. 5).
to the prior model time average are on the order of 0.5, consistently with Fig. 5(d) and N
m
> N
d
. The use of linearly estimated near-surface
magnetic diffusion and small-scale magnetic elds (as in Fig. 4) instead of the actual quantities does not lower the recovery quality, but results
in slightly larger data mists. Experiment 3 is a fraternal experiment using a different statistical prior (here model 1) instead of the model
that served to produce the synthetic data. This represents a situation where the prior presents an oversimplied description of the physics
underlying the data. The inversion yields almost the same data mist, a result which is surprising owing to the very distinct magnetic eld
morphologies of models 1 and 2 (see g. 3 in Fournier et al. 2011, and Fig. 1 a). The recovery quality is degraded, but again surprisingly good
considering that the velocity eld of model 1 (again g. 3 in Fournier et al. 2011) is morphologically very distinct from (and simpler than)
that of model 2. The recovery quality in experiment 3 is in line with the results of Rau et al. (2000), Amit et al. (2007), while experiment 2
generally yields better results, owing to the benets of error handling and appropriate prior statistical knowledge.
The quality of the deep ow recovery is evaluated in Fig. 6 from the surface results of experiment 2. Here the disparity between the
numbers of model and data degrees of freedom is acting twice (once at the stage of near-surface ow inversion, once at the stage of deep
ow inversion). The combination of the two effects attenuates the peak-to-peak magnitude of the recovery by about a factor two with respect
to the reference, while still maintaining the correct amplitude for the larger ow scales (the behaviour is qualitatively similar to that observed
in Fig. 5d). The most energetic features of the reference are well rendered in a morphological sense. However, there are obviously a number
of small-scale details missing in the recovery, which can thus be reasonably trustworthy if the focus is set on the large-scale, most energetic
features.
3.2 Geomagnetic inversions
Single-epoch inversions are performed using the geomagnetic data from models CM4 (in 1970, 1975, 1980, 1990, 1995, 2000) and gufm-
sat-Q3 (in 2000, 2001, 2002.5, 2004, 2010). It is emphasized that the present inversion scheme does not regularize the solutions in time.
Table 3 reports the quantities
d
and
m
for both geomagnetic eld models. With respect to the results obtained in the previous synthetic
Table 3. Normalized mist to the data
d
, and normalized deviation to the model time average

m
obtained when inverting the two geomagnetic eld models with the four priors described in
Section 2.1. The values reported are averages over the six epochs investigated in CM4, and the
ve epochs investigated in gufm-sat-Q3.

d
with prior 1 2 3 4
m
with prior 1 2 3 4
CM4 0.88 1.12 0.89 0.94 CM4 0.45 0.54 0.49 0.5
gufm-sat-Q3 1.34 1.32 1.18 1.22 gufm-sat-Q3 0.54 0.52 0.46 0.53

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 547
inversion experiments (Table 2), all inversions provide acceptable ts to the data, and the deviations of all inverted velocity elds from the
prior models time average are standard. The difference in the level of data mist
d
between CM4 and gufm-sat-Q3 originates in the different
values attributed to the noise background (see Section 2.2.1). In that respect, the noise background chosen for CM4 seems appropriate (
d
close to 1) while it is slightly optimistic for gufm-sat-Q3 (as was already mentioned in Pais & Jault 2008). In that latter case, the data mist
remains nevertheless reasonable as it is comparable to that obtained in synthetic experiments 2 and 3. As we shall see in the following, prior
models 3 and 4 have been elaborated to physically resolve deviations observed in inversions with prior 2, and yield consistently better (closer
to one) data mists with both geomagnetic eld models. Prior model 1 yields data mists comparable to prior model 2.
In the following, inversions performed with prior model 2 are rst investigated in detail. Model 2 is the central model to this study as
it is the most generic in terms of dynamo modelling [the model was part of a systematic parametric study in Christensen et al. (2010) and
investigated in Aubert & Fournier (2011)]. Model 1 is also generic but less physically realistic. Its purpose is an assessment of the general
robustness of the inversion. Inversions performed with models 3 and 4 are then investigated in order to attempt to explain and resolve the
deviations observed with model 2.
3.2.1 Inversions with prior model 2
The near-surface ow inversion for 2001 (Fig. 7a) recovers most previously obtained core surface ow features (Finlay et al. 2010), such as a
counter-clockwise gyre under the southern tip of Africa, a clockwise gyre east of Quebec, a westward equatorial owunder the Atlantic ocean.
The numerical dynamo prior imposes a high level of equator symmetry outside the axial cylinder tangent to the inner core, and each gyre gains
a symmetric counterpart (although the ow is not perfectly equator-symmetric as in quasi-geostrophic inversions). The ow also presents
a markedly hemispherical longitudinal pattern, with opposite rotations in the Atlantic and Pacic hemispheres. Deeper in the core (Figs 7b
and c), the highly equator-symmetric near-surface ow connects to a columnar ow (Fig. 7b), where ve large-scale convective cells can be
identied. Maps of the deep azimuthal velocity (Fig. 7c) also reveal a pair of nested columnar spirals anchored at the inner-core boundary,
inheriting their hemispherical character from the near-surface circulation. Together with the cylindrical radial ow (equatorial map in Fig. 7b),
they dene a planetary-scale circulation where upwelling from the eastern part of the inner-core boundary turns into eastward ow as it moves
upwards, and westward ow turns into downwelling in the Western Hemisphere as it approaches the inner-core boundary. The westward
(blue) part of this structure is stronger than its eastward counterpart and coincides with the quasi-geostrophic eccentric equatorial gyre (Pais
& Jault 2008), but touches the inner-core boundary instead of closing on itself. Similar images are found throughout the period 19702010
(Fig. 8), showing that the robustness and persistence of the gyre is supported by the present inverse geodynamo modelling approach. It should
be noted that quasi-geostrophic columnar motion outside the tangent cylinder is thus robustly obtained without being explicitly enforced.
Here it arises as a consequence of the physics contained in the prior. At the same time, the prior physics also permits ageostrophic deviations
such as thermal-wind driven polar vortices in the tangent cylinder (rotating westwards close to the coremantle boundary and eastwards close
to the inner-core boundary).
Fig. 9(a) shows that the inverted ow satises the data within the specied error level, as the residual left by the inversion (solid red)
is comparable to the sum (dashed red) of the modelled sources of error (the underparametrization error, blue, the magnetic diffusion error,
green, and the noise background, black). The near-surface ow is also generally compliant with the prior (Fig. 9b), as most individual ow
components remain within the 2 variance level when normalized with respect to the corresponding free run variance. This shows that
Earth-like, homogeneous numerical dynamos can provide a statistically compelling explanation of the geomagnetic secular variation. In the
present case however, only three spectral coefcients of the solution, t
0
1
, t
0
3
and t
1
2
, have variances exceeding the natural variability of the
prior by more than a factor 2. This underlines two important weaknesses of prior model 2, to which the inversion responds by stretching
the solution beyond the admissible statistical bounds provided by the prior. The two zonal coefcients t
0
1
and t
0
3
relate to the inability of the
prior to provide enough azimuthally averaged westward drift, while the coefcient t
1
2
relates to inability to produce a sizeable hemispherical
ow difference between the Atlantic and Pacic hemispheres (see insets in Fig. 9b).
3.2.2 Inversions with prior model 1 and robustness to a change of prior
Fig. 10 presents the near-surface inversion result for the same data as in Fig. 7, but now using prior model 1. Flows inverted using prior
models 1 and 2 are broadly similar at the largest (hemispherical) scales. Features well-identied in the literature such as the Quebec and
southern Atlantic vortices are consistently reproduced by both inversions. In contrast, the two inversions are markedly different at some other
locations (such as under India). Prior model 1 is less columnar than model 2, and this feature is also seen in the inversion result (compare the
magnitude of the northern and southern Pacic vortices). Flow inverted using prior model 1 is also slightly less energetic. Similarly to what
was observed with model 2, the inversion performed with model 1 needs to stretch the statistical deviation of the zonal and hemispherical
ow coefcients previously identied (not shown). The ow inversion result presented in Fig. 10 can be directly compared to the ow model
in g. 9 of Fournier et al. (2011), as it uses the same inversion prior and similar geomagnetic data. The two ows are morphologically
very different, highlighting the limits of a purely linear estimation of core surface ow from geomagnetic data, as discussed in Section 1.
In addition to rendering a pattern which is signicantly more compatible with the previous literature on core ow inversions, the present
inversion also renders a stronger ow (maximum 37 km yr
1
versus 22 km yr
1
in Fournier et al. 2011). The scheme indeed requests the

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

548 J. Aubert
Figure 7. Flow throughout the Earths core in 2001 from gufm-sat-Q3, inverted using prior model 2. (a) Atlantic-centered Hammer projection of the near-
surface ow (arrows, arbitrary scaling) superimposed with a colour map of the near-surface toroidal scalar T (blue denotes a clockwise circulation when seen
from the North pole). (b) Equatorial map (top) and isosurfaces (bottom) of the cylindrical radial velocity (red denotes an outward ow). The isosurface levels
are marked on the colour bar. (c) Equatorial map (top) and meridional cuts (bottom) of the azimuthal ow. Grey arrows in (b), (c) indicate the general ow
circulation. In (c), the westward eccentric columnar gyre appears in blue.
frozen-ux equation to be satised, thus setting an amplitude requirement on the ow. Such a requirement is not present in a linear estimation
of the ow based solely on its statistical correlations with magnetic eld and secular variation data, as done in Fournier et al. (2011).
3.2.3 Prior model 3, resolution of the t
0
1
t
0
3
deviation, and l.o.d. variations
The difculties encountered with the t
0
1
t
0
3
pair in Fig. 9(b) are indicative of a general inability of standard numerical dynamos to render a
westward-drifting zonal ow below the outer boundary at equatorial position. It has been suggested that an heterogeneous mantle control
could partially resolve this problem (Olson & Christensen 2002). Here an alternative approach is taken, attempting to explain westward drift
within the context of an homogeneous dynamo mechanism. The zonal owconguration in numerical dynamos with rigid boundaries is robust
to parameter variations, with shear created across the shell by the combined inuence of buoyancy and magnetic forces (Aubert 2005) and
resulting in an eastward ow near the inner-core boundary and almost zero rotation close to the equatorial outer boundary (Fig. 11a). It is now
assumed that the outer boundary is stress free, the inner boundary free to axially rotate under the inuence of viscous and magnetic torques,

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 549
Figure 8. Near-surface ow Hammer projections (a, same conventions as Fig. 7a) and equatorial maps of the azimuthal ow (b, same convention as Fig. 7c)
for epochs 1970, 1990 and 2010, using prior model 2.
that an additional gravitational torque couples the inner and outer boundaries, and that the total angular momentum of the solid core-uid
coremantle system is preserved. Being the only torque experienced by the outer boundary in that conguration, the gravitational torque
vanishes on the long term, thus maintaining alignment between the mantle and the inner core. The shear conguration then changes, with zero
average uid rotation with respect to the mantle immediately above the inner-core boundary and a sizeable westward rotation at equatorial
position close to the outer boundary (Fig. 11b). A similar process operates in the simulations of Buffett & Glatzmaier (2000). Implemented in
model 3, the conguration produces trains of westward-drifting equatorial magnetic ux patches of normal polarity, a feature which is very
reminiscent from the behaviour of the geodynamo. It additionally provides a physically consistent mechanism for creating l.o.d. variations.
Geomagnetic inversions performed with prior model 3 (Fig. 12a) shift the near-surface ow conguration observed in Fig. 7 towards a state
where the westward zonal ow is strengthened beneath the Atlantic, and the opposite rotation previously observed in the Pacic disappears.
The deep ow (Figs 12b and c) is still highly columnar, and directly reects these near-surface changes, with a strengthened retrograde
eccentric gyre and a strongly attenuated prograde gyre. The radial ow conguration remains similar, with upwelling being present mostly
in the Eastern Hemisphere and downwelling in the Western Hemisphere. These images are now strikingly similar to the quasi-geostrophic
inversion results (Pais & Jault 2008; Gillet et al. 2009). The normalized deviations of t
0
1
and t
0
3
are now inside the standard statistical deviation
of the prior, with respective values of 0.98 and 0.62, down from 3.98 and 2.25 in Fig. 9(b). The decisive inuence of the numerical prior for
determining the zonal ow, as well as other considerations regarding the resolution of the inversion are further discussed with the study of a
resolution matrix in Appendix A.
Having resolved the statistical deviation of the t
0
1
t
0
3
pair, it is interesting to compare the l.o.d. variations predicted by inversions with
priors 2 and 3 throughout the period 19702010. The outer-core angular momentum J is computed from the inverted azimuthal ow u

J =
_
V
r sin u

dV, (25)
where V is the core volume. The core ow-induced variations in the l.o.d. then follow from the formula (Jault et al. 1988; Jackson et al.
1993)
l.o.d. =
J
I
c
+ I
m
T
2
0
2
, (26)
where l.o.d. is the l.o.d. variation, J is the core angular momentum variation obtained from the inversions, I
c
+ I
m
= 8 10
37
kg m
2
is the moment of inertia of the coupled core-mantle system, T
0
= 24 hr is the reference l.o.d. The value used for the core density is

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

550 J. Aubert
Figure 9. Fit to the geomagnetic data from gufm-sat-Q3 in 2001 using prior model 2 and deviations from the prior model. (a) Earth-surface energy spectra
of the secular variation (black) as a function of the spherical harmonic degree, together with the t quality (red), and the various modelled errors (the sum
of which is represented in dashed red). (b) Variance spectrum of the individual toroidal near-surface ow spherical harmonic components. Each variance is
normalized with the variance obtained from a free run of the numerical dynamo prior model. Components with variance beyond the 2 level (dashed line)
have their corresponding azimuthal ow pattern represented in the insets (blue is westwards).
=11 000 kg m
3
. Fig. 13 presents the predicted and observed variations of the l.o.d. of core origin. As can be expected from the differences
in zonal ow morphologies (Fig. 11), these are quite larger when using prior model 3 than with prior model 2. The ratio between the two is
roughly consistent with the r.m.s. l.o.d. variations observed in free runs of the numerical priors (0.8 ms for model 2 and 3.2 ms for model 3).
Both priors roughly render the long-term decrease of the l.o.d. between 1970 and 2010 (this decrease being underestimated by prior model 2
and overestimated by prior model 3). Only prior model 3 is able to correctly render the l.o.d. variations over the last 10 yr. The results obtained
with prior model 3 are again very consistent with the results obtained in Gillet et al. (2009) with quasi-geostrophic core ow inversions.
3.2.4 Prior model 4 and resolution of the t
1
2
deviation
The statistical deviation of the t
1
2
near-surface ow coefcient displayed in Fig. 9(b) is now investigated. Here again the idea is to invoke an
additional feature previously not included in standard dynamo modelling. Heterogeneous boundary control is frequently invoked to explain
such planetary-scale, persistent non-zonal ow structures in the outer-core ow. The nested spirals in Fig. 7 could plausibly result from
coupling with a thermally heterogeneous mantle, as the resulting ows can have a large-scale spiralling structure (Sumita & Olson 1999),

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 551
Figure 10. Near-surface core ow in 2001 from gufm-sat-Q3, inverted using prior model 1 (same conventions as Fig. 7a).
Figure 11. Meridional cuts of the time averaged and azimuthally averaged ow (blue is westwards) in free runs (unconstrained by the data) of the numerical
prior models 2 (a) and 3 (b). In (b) velocities are measured with respect to the mantle.
do not drift in the mantle frame of reference, can locally account for the Atlantic westward drift (Olson & Christensen 2002) and exhibit a
weak hemispherical pattern in the secular variation amplitude (Christensen & Olson 2003). However, mantle-driven ows are generally not
fully longitudinally hemispherical as they inherit a harmonic order 2 structure from the seismically derived heat ow pattern imposed at the
coremantle boundary, and the orientation of their spirals tends to be mirror symmetric (Sumita & Olson 1999) to that observed here. To
test the other option of a control from below, a heterogeneous buoyancy release at the inner-core boundary is added to the numerical prior
model 2 to build model 4. In a rst step, direct numerical simulations are performed and visually compared to the images inverted with the
fully homogeneous, initial prior. A longitudinally hemispherical buoyancy pattern (Fig. 14a), with excess release below Asia yields a surface
circulation, cylindrical axial (Fig. 14b) and azimuthal (Fig. 14c) deep velocity patterns strikingly similar to the images inverted with prior
model 2 (Figs 7ac). In a second step, prior model 4 is used for geomagnetic inversion, the longitudinal orientation of the hemispherical
buoyancy release being varied (Fig. 14d). For epoch 2001, setting the center for excess buoyancy at about 70

E in the Asian Hemisphere

yields the best variance reduction for the t
1
2
outlier (at about 0.2), while also providing a slightly better overall t to the geomagnetic data.
4 DI SCUSSI ON
A theoretical framework, called inverse geodynamo modelling, has been built to produce inversions of the geomagnetic secular variation
for ows throughout the Earths core, the necessary prior information being supplied by a series of numerical models of the convective
geodynamo. The framework relies on classical data assimilation tools, but has been updated with respect to our previous work in order to
account for the non-linearity of the core ow problem. While the approach is not perfect, as it relies on rather detailed prior information,
the characteristics of which are subject to changes across the space of dynamo control parameters, it produces ows with features that are

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

552 J. Aubert
Figure 12. Flow throughout the Earths core in 2001 from gufm-sat-Q3, inverted using prior model 3 (same conventions as in Fig. 7). In (b), the uncapped
isosurfaces signal the use of a stress-free outer boundary. In (c), the strong equatorial westward zonal ow supported by the choice of boundary conditions and
angular momentum conservation appears clearly.
fairly robust against changes of the prior model. This fact suggests that ows produced even by the simplest numerical dynamo fairly well
describe the large-scale morphological characteristics of the much more turbulent ow generating the Earths magnetic eld, and provide a
satisfactory t to the secular variation within a realistically specied error level. The most important advantage that the method has over the
classical penalized least-squares approach is probably the fact that the solution is uniquely determined (absence of damping parameters) and
clearly connected to the rst principle physics supplied by the numerical dynamo prior model. Observed statistical deviations can then be
interpreted as shortcomings of the physical modelling. This possibility is rather novel among core ow inversion methods and has been used
here to attempt to physically describe the geomagnetic westward drift, and the global longitudinal hemisphericity of the geomagnetic secular
variation. Whether the proposed mechanisms are really the ones at work in the geodynamo or not is obviously still an open question, but the
present work opens interesting prospects by establishing a benchmark against which modelling ideas can be quantitatively tested.
The inverted deep ow has been shown to be strongly columnar if the prior Lehnert number is small enough, the limit of < 3 10
2
proposed by Jault (2008) being compatible with the present results. Owing to the smallness of that number in the Earth case, this produces
strong support to the hypothesis that ows inside Earths core are dominantly columnar, accordingly to the quasi-geostrophic assumption.
Furthermore, the results support the idea that the validity range of this assumption can be extended from short timescales (interannual to
decadal) towards timescales commensurate to core overturn and diffusive processes (centennial and more), at least for ow structures outside

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 553
Figure 13. Predicted l.o.d variations (blue, green) versus the observed variations attributed to core ow, computed as in g. 13 of Gillet et al. (2009).
the tangent cylinder. The existence of the giant eccentric westward columnar gyre proposed by Pais & Jault (2008) is thus conrmed within a
framework that allows this structure to persist over centuries, as seems to be the case in the Earth. In addition to presenting features strikingly
similar to quasi-geostrophic inversions, the inverted ows also have the additional strength of reconciling quasi-geostrophic, columnar ows
with longer-term ageostrophic features such as polar vortices inside the tangent cylinder.
Assuming that gravitational coupling of the mantle to the inner core is indeed the mechanism producing the geomagnetic westward
drift by deporting the thermal wind shear towards the vicinity of the core-mantle boundary, an additional constraint on the magnitude of
gravitational coupling can be derived from Fig. 13. In order to avoid excessive variations of the l.o.d. (as produced by prior model 3 and as
also seen in the results of quasi-geostrophic inversions, Gillet et al. 2009), a gravitational coupling constant on the order of /D
3
2
10
4
should probably be used in prior model 3 (although this remains to be checked in subsequent modelling). Scaling this value back to the
dimensional world using the scaling principles outlined in the introduction yields 2.3 10
20
N m yr. This is physically reasonable
(Aubert & Dumberry 2011) but too small to permit the explanation of the 6-year oscillation in the l.o.d. in terms of mantleinner-core
gravitational coupling (Mound & Buffett 2006), in which case an explanation in terms of rapid Alfv en waves (Gillet et al. 2010) would be
preferred.
Assuming that an heterogeneous buoyancy release at the inner core surface is at the origin of the hemispherical character of the secular
variation and its associated velocity eld structure, the giant eccentric westward columnar gyre, the present results predict that the center of the
excess buoyancy release should be in the Eastern Hemisphere. The inner-core possesses an instability mode (Monnereau et al. 2010) causing
its material to slowly translate along an axis parallel to the equatorial plane, with permanent melting in one hemisphere and crystallization in
the other. The direction of translation has been inferred to be west to east based on the comparison of a variable grain size scattering model with
inner-core surface seismic data. Should the light chemical elements at the origin of buoyancy be preferentially released in the crystallizing
hemisphere, this would indeed induce a hemispherical longitudinal buoyancy release pattern, but with the exact opposite distribution and
consequences (excess release and ow upwelling below America, decit and downwelling below Asia) to that deduced from the present
inverse geodynamo modelling results. There may be a few explanations to this discrepancy. First, the buildup of a dense layer (Alboussiere
et al. 2010) around the inner core as a consequence of its translational instability might reconcile the two views as this layer could act as
a buffer signicantly distorting the buoyancy pattern. Second, it is possible that the inner-core translation direction has changed through
time (Olson & Deguen 2012), in which case the results presented here would be supportive of a present translation direction opposite to the
long-term trend. It should be noted that this interpretation is compatible with the location of the recent-time geomagnetic dipole eccentricity
(Olson & Deguen 2012).
What should be the best numerical dynamo prior for geomagnetic inversion and data assimilation? Following the present study it is
obvious that a computationally tractable case underlain by physics closest to the real system and producing the most Earth-like output should
be chosen. However, it may be counterproductive to resolve all statistical deviations of a standard prior, as additional physical mechanisms
need to be invoked that are possibly not as geophysically robust than the basic formulation, and that can degrade the model compliance to the
geomagnetic eld (see for instance the higher
2
value of model 4 when compared to model 2 in Table 1). In the present case, the best model
would probably be model 3 as it invokes a simple and reasonable additional ingredient to clearly improve on the insufcient treatment of the
zonal ow in model 2, while providing a magnetic eld output which is even more compliant to the geomagnetic eld than that of model 2.
The hemispherical ow pattern can then be left as a perturbation in the inversion.

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

554 J. Aubert
Figure 14. Direct and inverse modelling of ow resulting from a hemispherical inner-core buoyancy release (prior model 4). (a) Near-surface ow Hammer
projection (same conventions as Fig. 7a), (b), (c) maps of the radial and azimuthal velocity eld in the equatorial plane (same conventions as Figs 7b and c),
from a direct numerical simulation (a free run unconstrained by the data) of model 4. In (a), (b), (c), the grey marks locate the positive and negative buoyancy
anomalies superimposed on the homogeneous buoyancy release. (d) Normalized mist to the data
d
and normalized variance of the t
1
2
spectral ow coefcient
obtained in inversions of gufm-sat-Q3 for 2001 using model 4, as a function of the central longitude of the positive buoyancy anomaly. Values previously
obtained in Fig. 9 with the homogeneous prior are reported as dashed lines.
The nal remark concerns the interesting prospects opened by the compatibility between 3-D inverse geodynamo modelling and quasi-
geostrophic core ow inversions. As mentioned in the introduction, the strengths and weaknesses of both approaches being complementary,
forthcoming efforts in the attempt of building a physically sound geomagnetic data assimilation scheme (this work being considered as one
of the steps) should include a combination of the two, in order to expand the time scale range over which the geomagnetic secular variation
can be accounted for. Future work also includes a regularization of the inversions in time (in addition to space as presently done), and the
buildup of a sequential assimilation scheme in time (as already undertaken in Aubert & Fournier 2011).
ACKNOWLEDGMENTS
JA wishes to thank Ingo Wardinski and an anonymous referee for insightful reviews. This work was supported by programs PNP/SEDI-TPS
of French Institut National des Sciences de lUnivers (INSU) of CNRS, and by the French Agence Nationale de la Recherche under the grant
ANR-2011-BS56-011. Numerical computations were performed at S-CAPAD, IPGP, France and using HPC resources from GENCI-IDRIS
(Grants 2011-042122 and 2012-042122). This is IPGP contribution 3345.

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Flow throughout the Earths core 555
REFERENCES
Alboussiere, T., Deguen, R. & Melzani, M., 2010. Melting-induced strati-
cation above the Earths inner core due to convective translation, Nature,
466(7307), 744747.
Amit, H., Olson, P. & Christensen, U., 2007. Tests of core ow imaging
methods with numerical dynamos, Geophys. J. Int., 168(1), 2739.
Aubert, J., 2005. Steady zonal ows in spherical shell dynamos, J. Fluid.
Mech., 542, 5367.
Aubert, J. & Dumberry, M., 2011. Steady and uctuating inner core rotation
in numerical geodynamo models, Geophys. J. Int., 184(1), 162170.
Aubert, J. & Fournier, A., 2011. Inferring internal properties of Earths core
dynamics and their evolution from surface observations and a numerical
geodynamo model, Nonlin. Process. Geophys., 18(5), 657674.
Aubert, J., Aurnou, J. & Wicht, J., 2008. The magnetic structure of
convection-driven numerical dynamos, Geophys. J. Int., 172, 945956.
Aubert, J., Labrosse, S. & Poitou, C., 2009. Modelling the palaeo-evolution
of the geodynamo, Geophys. J. Int., 179(3), 14141428.
Bloxham, J. & Jackson, A., 1991. Fluid ow near the surface of Earths
outer core, Rev. Geophys., 29(1), 97120.
Buffett, B.A., 2010. Tidal dissipation and the strength of the Earths internal
magnetic eld, Nature, 468(7326), 952955.
Buffett, B.A. & Glatzmaier, G.A., 2000. Gravitational braking of inner-core
rotation in geodynamo simulations, Geophys. Res. Lett., 27(19), 3125
3128.
Busse, F.H., 1970. Thermal instabilities in rapidly rotating systems, J. Fluid.
Mech., 44(3), 441460.
Christensen, U. & Aubert, J., 2006. Scaling properties of convection-driven
dynamos in rotating spherical shells and application to planetary magnetic
elds, Geophys. J. Int., 117, 97114.
Christensen, U. & Olson, P., 2003. Secular variation in numerical geody-
namo models with lateral variations of boundary heat ux, Phys. Earth
planet. Int., 138, 3954.
Christensen, U. & Tilgner, A., 2004. Power requirement of the geodynamo
from ohmic losses in numerical and laboratory dynamos, Nature, 429,
169171.
Christensen, U.R., 2011. Geodynamo models: Tools for understanding prop-
erties of earths magnetic eld, Phys. Earth planet. Int., 187, 157169.
Christensen, U.R., Wardinski, I. & Lesur, V., 2012. Timescales of geomag-
netic secular acceleration in satellite eld models and geodynamo models,
Geophys. J. Int., doi:10.1111/j.1365-246X.2012.05508.x.
Christensen, U.R., Aubert, J. & Hulot, G., 2010. Conditions for Earth-like
geodynamo models, Earth. planet. Sci. Lett., 296(3-4), 487496.
Dormy, E., Cardin, P. &Jault, D., 1998. MHDowin a slightly differentially
rotating spherical shell, with conducting inner core, in a dipolar magnetic
eld, Earth. planet. Sci. Lett., 160(1-2), 1530.
Finlay, C., Jackson, A., Gillet, N. & Olsen, N., 2012. Core surface magnetic
eld evolution 2000-2010, Geophys. J. Int..
Finlay, C.C., Dumberry, M., Chulliat, A. &Pais, M.A., 2010. Short timescale
core dynamics: theory and observations, Space. Sci. Rev., 155(1-4), 177
218.
Fournier, A. et al., 2010. An Introduction to data assimilation and pre-
dictability in geomagnetism, Space. Sci. Rev., 155(1-4), 247291.
Fournier, A., Aubert, J. & Th ebault, E., 2011. Inference on core surface ow
from observations and 3-D dynamo modelling, Geophys. J. Int., 186(1),
118136.
Gillet, N., Pais, M.A. & Jault, D., 2009. Ensemble inversion of time-
dependent core ow models, Geochem. Geophys. Geosys., 10(Q06004),
doi:10.1029/2008GC002290.
Gillet, N., Jault, D., Canet, E. & Fournier, A., 2010. Fast torsional waves
and strong magnetic eld within the Earths core, Nature, 465(7294),
7477.
Holme, R., 2007. Large-scale ow in the core, in Treatise on Geophysics,
Vol. 8: Core Dynamics, ed. Olson, P., p. 107130, Elsevier, Amsterdam.
Jackson, A., Bloxham, J. & Gubbins, D., 1993. Time-dependent ow at
the core surface and conservation of angular momentum in the coupled
core-mantle system, in Dynamics of the Earths Deep Interior and Earth
Rotation, Vol. 72, pp. 97107, eds Le Mou el, J.-L., Smylie, D.E. &
Herring, T., AGU Geophysical Monograph, Washington, DC.
James, R., 1973. Adams and Elsasser dynamo integrals, Proc. R. Soc. Lond.,
A, 331(1587), 469478.
Jault, D., 2008. Axial invariance of rapidly varying diffusionless motions in
the Earths core interior, Phys. Earth planet. Int., 166(1-2), 6776.
Jault, D., Gire, C. & Le Mou el, J.-L., 1988. Westward drift, core motions
and exchanges of angular momentum between core and mantle, Nature,
333, 353356.
Lhuillier, F., Fournier, A., Hulot, G. & Aubert, J., 2011. The geomagnetic
secular-variation timescale in observations and numerical dynamo mod-
els, Geophys. Res. Lett., 38(L09306), doi:10.1029/2011GL047356.
Monnereau, M., Calvet, M., Margerin, L. & Souriau, A., 2010. Lopsided
Growth of Earths Inner Core, Science, 328(5981), 10141017.
Moon, W., 1979. Numerical evaluation of geomagnetic dynamo integrals
(Elsasser and Adams-Gaunt integrals)., Comp. Phys. Comm., 16(2), 267
271.
Mound, J.E. & Buffett, B.A., 2006. Detection of a gravitational oscillation
in length-of-day, Earth planet. Sci. Lett., 243(3-4), 383389.
Olson, P. & Aurnou, J., 1999. A polar vortex in the Earths core, Nature,
402(6758), 170173.
Olson, P. & Christensen, U., 2002. The time averaged magnetic eld in nu-
merical dynamos with non-uniform boundary heat ow, Geophys. J. Int.,
151, 809823.
Olson, P. &Deguen, R., 2012. Eccentricity of the geomagnetic dipole caused
by lopsided inner core growth, Nat. Geosci., 5, 565569.
Olson, P., Christensen, U. & Driscoll, P., 2012. From superchrons to secular
variation: a broadband dynamo frequency spectrum for the geomagnetic
dipole, Earth. planet. Sci. Lett., 319-320, 7582.
Pais, M.A. & Jault, D., 2008. Quasi-geostrophic ows responsible for the
secular variation of the Earths magnetic eld, Geophys. J. Int., 173(2),
421443.
Rau, S., Christensen, U., Jackson, A. & Wicht, J., 2000. Core ow inversion
tested with numerical dynamo models, Geophys. J. Int., 141(2), 485497.
Roberts, P.H. & Scott, S., 1965. On the analysis of the secular variation. I:
a hydromagnetic constraint: theory, J. Geomag. Geoelectr., 17, 137151.
Sabaka, T., Olsen, N. &Purucker, M., 2004. Extending comprehensive mod-
els of the Earths magnetic eld with Orsted and CHAMP data, Geophys.
J. Int., 159(2), 521547.
Sreenivasan, B. & Jones, C., 2005. Structure and dynamics of the polar
vortex in the Earths core, Geophys. Res. Lett., 32(20).
Sumita, I. & Olson, P., 1999. A laboratory model for convection in earths
core driven by a thermally heterogeneous mantle, Science, 286(5444),
15471549.
Whaler, K.A., 1986. Geomagnetic evidence for uid upwelling at the core-
mantle boundary, Geophys. J. R. astr. Soc., 86(2), 563588.
APPENDI X A: A MODEL RESOLUTI ON MATRI X
The model resolution matrix K
fs
M (Fig. A1) indicates how much of a given near-surface ow is recovered after successive execution of the
direct and the inverse problems. Alternatively, this matrix also indicates whether the determination of a given ow coefcient relies entirely
on the data or also requires knowledge of the prior and of other ow coefcients. For a fully determined inverse problem, every coefcient is
perfectly resolved and K
fs
M is the identity matrix. Here, however, given the data resolution and the specied error, a perfect recovery of the
ow coefcients is not expected, especially beyond spherical harmonic degree 13.

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

556 J. Aubert
Figure A1. Model resolution matrix K
fs
M for the 2001 near-surface ow inversion of Fig. 12 (using prior model 3). (a) Full matrix, (b) diagonal coefcients.
In the case of the model 3 inversion presented in Fig. 12, the resolution (diagonal coefcients in Fig. A1a, also reported in Fig. A1b) is
above 0.5 for toroidal ow coefcients with spherical harmonic order between 1 and 6, or half the data truncation. This can be considered
satisfactory given the large underdetermination of the problem. The zonal ow (toroidal order 0) is a notable exception, with resolution
coefcients at or below 0.4. The inversions presented in Figs 7 and 12 have already underlined the decisive inuence of the prior in
determining the zonal ow. As their inverted columnar ows have signicant energies at harmonic orders 56, we expect these to be correctly
resolved.
These results contrast with the good recovery obtained up to degree 13 in the twin experiments presented in Fig. 5. For that synthetic
case, presumably, the perfect statistical compatibility between the prior and the searched owsolution is responsible for this extended recovery
range.
The weak diagonal coefcients in the lower right-hand side corner of the resolution matrix, together with the presence of off-diagonal
coefcients in the lower left-hand side corner, indicate that the determination of the spheroidal (upwelling) ow coefcients relies on the
prior and on their correlation with the toroidal ow coefcients.
For all coefcients, the determination of ow coefcients beyond degree 6 relies on the prior to an increasing extent as the spherical
harmonic degree increases.

b
y

g
u
e
s
t

o
n

J
a
n
u
a
r
y

2
4
,

2
0
1
3
h
t
t
p
:
/
/
g
j
i
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m