A&A 587, A150 (2016) Astronomy

DOI: 10.1051/0004-6361/201527295 &


c ESO 2016 Astrophysics

Solar activity during the Holocene: the Hallstatt cycle

and its consequence for grand minima and maxima
I. G. Usoskin1 , Y. Gallet2 , F. Lopes2 , G. A. Kovaltsov3,4 , and G. Hulot2

1
Sodankylä Geophysical Observatory (Oulu unit) and ReSoLVE Center of Excellence, University of Oulu, 90014 Oulou, Finland
e-mail: [email protected]
2
Institut de Physique du Globe de Paris, Sorbonne Paris Cité, Université Paris Diderot, UMR 7154 CNRS, 75005 Paris, France
3
Ioffe Physical-Technical Institute, 194021 St. Petersburg, Russia
4
Institute of Physics of the Earth, Russian Academy of Sciences, 119991 Moscow, Russia
Received 2 September 2015 / Accepted 14 January 2016

ABSTRACT

Aims. Cosmogenic isotopes provide the only quantitative proxy for analyzing the long-term solar variability over a centennial
timescale. While essential progress has been achieved in both measurements and modeling of the cosmogenic proxy, uncertain-
ties still remain in the determination of the geomagnetic dipole moment evolution. Here we aim at improving the reconstruction of
solar activity over the past nine millennia using a multi-proxy approach.
Methods. We used records of the 14 C and 10 Be cosmogenic isotopes, current numerical models of the isotope production and trans-
port in Earth’s atmosphere, and available geomagnetic field reconstructions, including a new reconstruction relying on an updated
archeo- and paleointensity database. The obtained series were analyzed using the singular spectrum analysis (SSA) method to study
the millennial-scale trends.
Results. A new reconstruction of the geomagnetic dipole field moment, referred to as GMAG.9k, is built for the last nine millennia.
New reconstructions of solar activity covering the last nine millennia, quantified in terms of sunspot numbers, are presented and
analyzed. A conservative list of grand minima and maxima is also provided.
Conclusions. The primary components of the reconstructed solar activity, as determined using the SSA method, are different for
the series that are based on 14 C and 10 Be. This shows that these primary components can only be ascribed to long-term changes in the
terrestrial system and not to the Sun. These components have therefore been removed from the reconstructed series. In contrast, the
secondary SSA components of the reconstructed solar activity are found to be dominated by a common ≈2400-year quasi-periodicity,
the so-called Hallstatt cycle, in both the 14 C and 10 Be based series. This Hallstatt cycle thus appears to be related to solar activity.
Finally, we show that the grand minima and maxima occurred intermittently over the studied period, with clustering near lows and
highs of the Hallstatt cycle, respectively.
Key words. Sun: activity – solar-terrestrial relations

1. Introduction levels. In particular, although the existence of grand minima and

maxima in solar activity has been known for a long time (see
Solar activity, which generically includes different manifesta- the review of Usoskin 2013), it has remained a matter of debate
tions of fluctuating processes in the solar convection zone on whether grand minima and maxima are separate activity modes
the surface and in the corona, varies on timescales from seconds of the solar dynamo or simply non-Gaussian tails of its variabil-
to millennia. Its long-term variability, with timescales longer ity (e.g., Moss et al. 2008; Passos et al. 2014; Karak et al. 2015).
than several centuries, can only be studied using indirect prox-
ies, such as cosmogenic radionuclides (e.g., Beer et al. 2012; To settle this question, a new approach was recently devel-
Usoskin 2013). For this purpose, the most useful cosmogenic oped (Usoskin et al. 2014), which takes into account the full
isotopes are radiocarbon 14 C and beryllium 10 Be. These isotopes range of uncertainties associated with a modern reconstruction
allow reconstructing the past solar activity over the Holocene of the 14 C global production rate (Roth & Joos 2013), an accu-
period, which spans the past 11 millennia, that is, since the rate millennial-scale archeomagnetic field reconstruction (Licht
last deglaciation. Recently, several long-term solar activity re- et al. 2013), and a detailed 14 C production model (Kovaltsov
constructions have been published (e.g., Solanki et al. 2004; et al. 2012). Based on an analysis of the probability density func-
Vonmoos et al. 2006; Muscheler et al. 2007; Usoskin et al. 2007; tion, the resulting reconstruction, limited to the past 3000 years
Steinhilber et al. 2012). They revealed variability in the solar (Licht et al. 2013), made it possible for the first time to show
activity on centennial and millennial scales, ranging from very that grand minima of solar activity correspond to a distinct oper-
low (almost spotless Sun) to high activity levels. However, since ational mode of the solar activity (Usoskin et al. 2014). Because
these reconstructions rely on slightly different basic models and of poor statistics, however, the result was inconclusive for the
different datasets, they sometimes differ in details and overall nature of the grand maxima. Constraining the solar activity over
longer period is more problematic, in particular because of un-
 certainties acknowledged before 500–1000 BC in paleo- and
Tables C.1 and C.2 are only available at the CDS via anonymous
ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via archeomagnetic field reconstructions (Snowball & Muscheler
http://cdsarc.u-strasbg.fr/viz-bin/qcat?J/A+A/587/A150 2007). However, a large set of new archeo- and paleointensity
Article published by EDP Sciences A150, page 1 of 10
 A&A 587, A150 (2016)

Q( C) (atoms/cm /sec)
data was acquired in the past few years (see Appendix A), which,
14
as we argue, can improve our knowledge of the geomagnetic 2.5 A) C

2
dipole moment evolution over most of the Holocene. Here we
take advantage of the new data to better constrain the long-term
solar activity, as revealed from the use of a synthetic index of rel- 2.0
ative sunspot numbers. We first extend the approach of Usoskin
et al. (2014) to the last 9 millennia using the reconstruction of
the 14 C global production rate (Roth & Joos 2013), the 10 Be

14
1.5
GRIP dataset (Yiou et al. 1997), a new dipole moment recon-
struction hereafter referred to as GMAG.9k (see Appendix A),
-6 -5 -4 -3 -2 -1 0 1 2
and 14 C and 10 Be production models (Kovaltsov et al. 2012;

Be flux (atoms/cm /sec)

10
Kovaltsov & Usoskin 2010). The recovered reconstructions aim B) Be
0.015
at scrutinizing the variability in the solar activity over millen-

2
nial and centennial timescales. We also discuss the robustness
of our new solar activity reconstructions using the results de-
rived from other recent archeo- and paleomagnetic Holocene
field models. Finally, we present new results providing impor- 0.010
tant observational constraints on the solar dynamo.

-6 -5 -4 -3 -2 -1 0 1 2
10
2. Data kyears -BC/AD
2.1. Cosmogenic radionuclide records Fig. 1. Time series of cosmogenic radionuclide data. Panel A): decadal
We used two sets of cosmogenic radionuclide data (14 C in tree radiocarbon 14 C global production rate (Roth & Joos 2013) with the
trunks and 10 Be in polar ice; panels A and B in Fig. 1, respec- 95% confidence interval plotted in gray. Panel B): quasi-decadal vari-
ability of 10 Be flux in the GRIP ice core (Yiou et al. 1997). The formal
tively) as tracers of solar activity (e.g., Beer et al. 2012; Usoskin 1σ error of 7% (relative to the given value) is indicated by the error bar
2013). next to the legend.
Radiocarbon 14 C is produced in the terrestrial atmosphere by
cosmic rays and then takes part in the global carbon cycle (e.g.,
Bard et al. 1997; Beer et al. 2012; Roth & Joos 2013). The mea- depositional flux measured in central Greenland in the frame-
sured quantity, the relative concentration Δ14 C of radiocarbon in work of the Greenland Ice Core Project (GRIP) for the period
tree rings, needs to be corrected for the apparent decay and for before 1645 AD (Yiou et al. 1997). We considered the mean data
the carbon cycle effect to reconstruct the 14 C production rate. set reduced to quasi-decadal time resolution. The corresponding
Here we used the 14 C production rate, Q(14 C), as reconstructed rate is shown in Fig. 1B, where we also plot a formal 1σ un-
by Roth & Joos (2013) for the Holocene, using the globally aver- certainty error (estimated to be 7% in relative terms, Yiou et al.
aged IntCal09 (Reimer et al. 2009) radiocarbon database and the (1997)).
dynamical BERN3D-LPJ carbon cycle model, which is a new-
generation carbon-cycle climate model, featuring a 3D dynamic 2.2. Axial dipole evolution over the past 9000 years
ocean, reactive ocean sediments, and a 2D atmosphere compo-
nent coupled to the Lund-Potsdam-Jena dynamic global vegeta- Two approaches can be used to constrain the axial dipole mo-
tion model. The data were reduced to the decadal temporal res- ment evolution over the past few millennia. The first consists
olution. For the decades around years 775 AD and 994 AD, the in constructing global geomagnetic field models in the form of
production rate was corrected to remove the modeled contribu- time-varying series of Gauss coefficients by taking advantage of
tion that is due to the occurrence of two extreme solar particle all available archeo- and paleomagnetic data (see, e.g., Korte
events (Usoskin et al. 2013; see also the discussion in Miyake & Constable 2005, 2011; Korte et al. 2011; Licht et al. 2013;
et al. 2012; Miyake et al. 2013; and Bazilevskaya et al. 2014). Pavón-Carrasco et al. 2014; Nilsson et al. 2014) and using the
Finally, we only consider data before 1900 AD because of the corresponding axial dipole component |g10 |. Differences among
Suess effect, which is related to extensive burning of fossil fuel, these models mainly come from the treatment applied to the
which dilutes radiocarbon in the natural reservoirs and makes data, in particular the way experimental and dating uncertain-
the use of the 14 C data after 1900 more uncertain. ties are being handled (see discussion and details in the refer-
Following Usoskin et al. (2014), we used 1000 individual re- ences above). In the present study, we considered three recent
alizations of the Q(14 C) ensemble to describe the consequences models, referred to as A_FM (Licht et al. 2013), SHA.DIF.14k
of uncertainties in the data and carbon cycle modeling (Roth (Pavón-Carrasco et al. 2014), and pfm9k.1a/b (Nilsson et al.
& Joos 2013). The corresponding production rate is shown in 2014). We note that A_FM and SHA.DIF.14k were built using
Fig. 1a (mean (black) and 95% range (gray shaded area) of the archeomagnetic and volcanic data sets, whereas pfm9k.1a/b also
1000 realizations). took sedimentary data into account. We assume that pfm9k.1a/b
The radionuclide 10 Be is produced by cosmic rays in the at- supersedes the slightly older CALS10k.1b field model con-
mosphere through spallation reactions (Beer et al. 2012). It be- structed by Korte et al. (2011) using practically the same dataset.
comes attached to aerosols and is relatively quickly precipitated The second approach is based on archeo- and paleointen-
to the ground. Because of this fast precipitation, it is not com- sity data collected worldwide, using archeological artifacts and
pletely mixed in the atmosphere and is subject to some com- volcanic rocks. The data are transformed into virtual axial
plicated transport. We relied on the parameterization of the at- dipole moments (VADM), which are then carefully weighted
mosphere transport and deposition of beryllium proposed by to produce a worldwide average VADM. To be valid, this
Heikkilä et al. (2009). Here we used a long series of 10 Be “paleomagnetic” approach requires a dual averaging of the data,
A150, page 2 of 10
 I. G. Usoskin et al.: Solar activity during the Holocene: the Hallstatt cycle and its consequence for grand minima and maxima

both in time and space, to best smooth out non-dipole field com-
A)

VADM (10 A m )
12

2
ponents (for a discussion, see, e.g., Korte & Constable 2005;
Genevey et al. 2008). Here we consider the two most recent

22
mean VADM curves built in this way, one by Genevey et al. 10
(2008), which encompasses the past 3000 years, and a second
one by Knudsen et al. (2008), which covers the entire Holocene. 8
In addition, and because quite a large number of additional inten-
sity data have recently been collected, an updated mean VADM
6
curve was also produced for the purpose of the present study.
For this we used the GEOMAGIA50.v3 data base (Brown et al. -7 -6 -5 -4 -3 -2 -1 0 1 2
2015), to which we added or modified about 390 individual in- kyears -BC/AD
tensity data points (see more details in Appendix A). The new
data compilation contains 4764 intensity values dated to between B)

VADM (10 A m )
2
7000 BC and 2000 AD. 12
To build this new VADM curve, we first carried out a se-

22
ries of computations to explore the effects of changing the
width of the temporal averaging sliding windows (200, 500, and 10
1000 years) and the size of the region of spatial weighting (over
regions of 10◦ , 20◦ , and 30◦ width). We also used a bootstrap GMAG AF_M G08
technique to simulate the effect of noise in the intensity data 8 pfm9k.1a pfm9k.1b
within their age uncertainties and within their 2σ experimen- Kn08 SHA-DIF.14k
tal error bars (see, e.g., Korte et al. 2009; Thébault & Gallet -1 0 1 2
2010). For each set of parameters, an ensemble of 1000 individ- kyears -BC/AD
ual curves was computed, allowing us to obtain at each epoch
the mean VADM and its standard deviation, together with the Fig. 2. Time series of the axial dipole moment reconstructions spanning
maximum and minimum VADM from the 1000 possible values. the past 9000 years (panel A)), with a zoom for the last 3200 years
Results from different computations are shown in Appendix A. (panel B)). The black solid line depicts GMAG.9k (the reconstruction
presented and used in this work) with ±1σ and the full range vari-
This analysis revealed that VADMs derived using sliding win- ability presented by the gray shading and the hatching, respectively.
dows with widths of 500 years and 1000 years are very simi- Other reconstructions shown are (Licht et al. 2013) denoted as AF_M,
lar. VADM values also appear to be relatively insensitive to the (Genevey et al. 2008) denoted as G08, (Nilsson et al. 2014) denoted as
size of the area chosen for the regional averaging. Some differ- pfm9k.1b and pfm9k.1a, (Knudsen et al. 2008) denoted as Kn08, and
ences, but still quite limited, are observed when the width of (Pavón-Carrasco et al. 2014) denoted as SHA-DIF.14k. For better read-
the sliding window is reduced to 200 years, which reveals en- ability, error bars were omitted for these curves, but this does not affect
hanced variations compared to that obtained when using sliding the discussion of the results (see text).
windows of larger widths, as expected. Averaging over such a
narrow window, however, is reasonable only for the most recent
time interval (here, the past 3500 years), which is documented offset between ≈500 AD and ≈1500 AD. This offset also per-
by a rather large number of data points (3756 among the 4764 sists when a sliding-window duration of 500 years is used for
available data) with a relatively wide but still uneven geograph- the VADM computations. However, the A_FM, SHA-DIF.14k
ical distribution (see for instance Fig. 4 in Knudsen et al. 2008; and pfm9k.1a/b curves generally lie within the envelope of pos-
Genevey et al. 2008). Because of this, we finally decided to build sible VADM values (Fig. 2). The same observation holds for all
a composite VADM variation curve, which we hereafter refer to the periods before 1000 BC (Fig. 2A). This encouragingly sug-
as GMAG.9k. This curve was computed for every 10 years, us- gests that relying on the GMAG.9k ensemble of 1000 individ-
ing sliding windows of widths 200 years between 1500 BC and ual VADM curves to reconstruct the solar activity as done in
2000 AD and 500 years between 7000 BC and 1500 BC, with a the present study, can be considered as a conservative procedure
spatial weighting over regions of 10◦ in size for both time inter- from a geomagnetic point of view.
vals (numerical values for this composite curve are provided in
the CDS, Tables C.1 and C.2).
Figure 2 shows this new GMAG.9k curve. This updated 3. Reconstructing the solar activity
VADM curve does not markedly differ from previous dipole mo-
ment curves. Its behavior over the past 3000 years is very similar Since details of the reconstruction of solar activity from cosmo-
to that of the VADM curve obtained by Genevey et al. (2008), genic nuclides are described elsewhere (e.g., Beer et al. 2012;
who also used a spatial weighting, but with a smaller number of Usoskin 2013), we only briefly describe this reconstruction and
different and distant regions (30◦ in size) and a smaller dataset recall important relevant information. Cosmogenic isotopes are
selected based on specific quality criteria. We note, however, produced by cosmic rays in the terrestrial atmosphere. Since cos-
that the new VADM curve tends to lie slightly below that of mic rays are modulated by solar magnetic activity, the variability
Genevey et al. (2008). Differences with the VADM variation of cosmogenic isotope production reflects the latter. However,
curve of Knudsen et al. (2008) are larger. However, the latter two terrestrial processes may disturb this relation. One process
was computed using sliding windows of width 500 years with is additional shielding of Earth from cosmic rays by the geomag-
no geographical weighting (these authors concluded that they netic field, whose changes must be known independently. For
had no significant bias despite the poor spatial data distribution). this purpose, we relied on the GMAG.9k axial dipole evolution
Comparison with the A_FM, SHA-DIF.14k and pfm9k.1a/b constructed as described in Sect. 2.2. As discussed in Sect. 2.1,
curves (Licht et al. 2013; Pavón-Carrasco et al. 2014; Nilsson another important process is transport and deposition of the
et al. 2014) reveals a fairly similar evolution except for a small nuclides in the terrestrial system. Because of the poorly known
A150, page 3 of 10
 A&A 587, A150 (2016)

details of climate variability in the past, the related transport A) SN-14C SN-10Be Us14
80
models are commonly adjusted to modern conditions, which
may lead to some uncertainties in the older part of the time 60
interval.
40
We converted the cosmogenic isotope production rate to the 20
Galactic cosmic ray (GCR) flux variability using recent produc-
tion models. The global production of 14 C was modeled using 0
the model of Kovaltsov et al. (2012), while the production of 20 B)

Sunspot number
10
Be was modeled using an updated version of the model of
Kovaltsov & Usoskin (2010). Cosmic ray variability was calcu- 10
lated in terms of the heliospheric modulation potential (see def- 0
initions and formalism in Usoskin et al. 2005), also considering
α-particles and heavier species of cosmic rays (Webber & Higbie -10
2009). This modulation potential was furthermore converted into
-20
decadal (solar-cycle averaged) sunspot number through the open
solar magnetic flux model (Solanki et al. 2000; Krivova et al. 10 C)
2007), which relates the solar surface magnetic cycle to the
5
emerging magnetic flux (Cameron & Schüssler 2015). We note
that the overall reconstruction method used here is similar to that 0
previously used by Usoskin et al. (2014).
-5
Uncertainties were assessed straightforwardly by computing
-10
a large ensemble of individual reconstructions. We used the set
of 1000 time-varying individual archeomagnetic reconstructions -7 -6 -5 -4 -3 -2 -1 0 1 2
of GMAG.9k (see Sect. 2.2), which account in particular for kyears -BC/AD
experimental and age uncertainties. This ensemble was cross-
used with a similar ensemble of 1000 production rates of 14 C Fig. 3. Panel A): raw reconstructions of the sunspot numbers (mean
to account for measurement and compilation uncertainties in the curves) SN-14C (blue) and SN-10Be (red), compared to the recent 3 kyr
reconstruction (Usoskin et al. 2014, – green curve). Panel B): first com-
IntCal09 and SHCal04 data, in the air-sea gas exchange rate, in
ponent of the singular spectrum analysis (SSA – see Appendix B) for
the terrestrial primary production, and in the closure of the at- the SN-14C (blue) and SN-10Be (red) series. The shaded areas depict
mospheric CO2 budget (Roth & Joos 2013). GMAG.9k was also the uncertainties related to the parameter L of the SSA. Panel C): same
cross-used in the same way with a set of 1000 10 Be series. In as in panel B), but for the second SSA components of the SN-14C
that case, however, 10 Be series of decadal values were generated (blue) and SN-10Be (red) series. The large dots and red stars denote
around the mean provided by GRIP using normally distributed times of the grand minima (see Table 1) and grand maxima (Table 2),
random numbers (with a standard deviation equal to 7% of the respectively.
mean value, Yiou et al. 1997) to reflect known errors. In both
cases, all possible combinations of the ensembles yielded 106 se- GMAG pfm9k.1a
ries of the reconstructed heliospheric modulation potential, next 100 Kn08 SHA-DIF.14k
Sunspot number

converted into 106 series of sunspot numbers. These series re-

flect the error propagation through all the intermediate steps. An
additional random error with σ = 0.5 was finally added to each
computed sunspot number to account for the small possible error 50
related to the conversion between the modulation potential and
the solar open magnetic flux (Solanki et al. 2004).
Decadal sunspot numbers reconstructed in this way from the
14
C and 10 Be data are henceforth denoted as SN-14C and SN- 0
10Be, respectively. Ensemble means of these SN-14C and SN-
-7 -6 -5 -4 -3 -2 -1 0 1 2
10Be are shown in Fig. 3A. These reconstructions agree well
with the earlier reconstruction (Usoskin et al. 2014) that cover kyears -BC/AD
the last 3000 years.
Fig. 4. Comparison of alternative SN-14 sunspot number
We checked the influence of the choice of axial dipole mo- reconstructions when relying on different axial dipole reconstruc-
ment reconstruction on the SN-14C reconstruction by also con- tions, (same notations as in Fig. 2). Only the mean curves of the
sidering alternative geomagnetic field models. Results are shown corresponding ensembles are shown.
in Fig. 4. This figure clearly shows that all 14 C-based SN re-
constructions lie close to each other and reveal a common gen- 4. Long-term behavior
eral pattern. In particular, Figs. 3A and 4 both display a long-
term trend in the 14 C-based SN reconstructions over the past To investigate the long-term behavior of the reconstructed so-
9000 years. This trend, however, is different from that of the lar activity, we relied on the singular spectral analysis (SSA)
10
Be-based SN reconstruction. What causes these long-term method (Vautard & Ghil 1989; Vautard et al. 1992) as described
trends is unclear: they may reflect various combinations of cli- in Appendix B. This SSA was applied to both the SN-14C and
mate effects or improperly corrected geomagnetic field effects. SN-10Be series, and the corresponding two first SSA compo-
We discuss this important point in the next section. nents are shown in Figs. 3B and C. The robustness of these SSA
A150, page 4 of 10
 I. G. Usoskin et al.: Solar activity during the Holocene: the Hallstatt cycle and its consequence for grand minima and maxima

Table 1. Grand minima with their centers, approximate duration, and

comments (1 – listed in Usoskin et al. 2007; 2 – listed in Inceoglu et al. 100 A) SN-14C-C Us14
2015).

Sunspot number
Center Duration Comment
(-BC/AD) (years)
50
1680 80 Maunder†
1470 160 Spörer
1310 80 Wolf
1030 80 Oort
690 80 1, 2 0
−360 80 1, 2
−750 120 1, 2
−1385 70 1, 2 -7 -6 -5 -4 -3 -2 -1 0 1 2
−2450 40 2
−2855 90 1, 2 100 B) SN-10Be-C Us14
−3325 90 1, 2
−3495 50 1, 2
−3620

Sunspot number
50 1, 2
−4220 30 1, 2
−4315 50 1, 2 50
−5195 50 2
−5300 50 1, 2
−5460 40 1, 2
−5610 40 1, 2
−6385 130 1, 2
0
(†)
Notes. Independently known.
-7 -6 -5 -4 -3 -2 -1 0 1 2

Table 2. Grand maxima with their centers, approximate duration, and

kyears -BC/AD
comments (1 – listed in Usoskin et al. 2007; 2 – listed in Inceoglu et al. Fig. 5. Corrected sunspot number reconstructions SN-14C-C (panel A))
2015). and SN-10Be-C (panel B)), after removing long-term trends (see
Sect. 4.1). The black curves and the gray shading depict the mean and
Center Duration Comment the 95% range (over 106 ensemble members) of the reconstructions, re-
(-BC/AD) (years) spectively, and the green curve represents the 3 kyr reconstruction by
Usoskin et al. (2014). Stars and circles denote grand maxima and min-
1970 80 Modern† ima, respectively, as in Fig. 3. Tables for this plot are available at the
505 50 2 CDS, including the mean values and the uncertainties of the sunspot
305 30 2 numbers reconstructed here as shown in Fig. 5.
−245 70 2
−435 50 1, 2
−2065 50 1, 2
−2955 30 2 Just as clearly, it also appears that these primary compo-
−3170 100 1, 2 nents are different for the two series: SN-14C yields a single,
−3405 50 2 nearly symmetric wave along the entire time interval of 9 mil-
−3860 50 1, 2
lennia with a range of about 20 in sunspot number, while SN-
−6120 40 1, 2
−6280 40 2 10Be yields a nearly monotonous trend within the same range
−6515 70 1 of about 20 in sunspot number. The fact that these trends differ
−6710 40 1 so much implies that they can hardly be related to a common
process. This makes terrestrial processes, in particular transport
Notes. (†) Independently known. and deposition, a much more likely cause. Differences in the
very long term Holocene trends between the two isotopes have
previously been noted and ascribed to such terrestrial processes
(Vonmoos et al. 2006; Usoskin et al. 2009; Steinhilber et al.
results has also been assessed by considering a wide range of 2012; Inceoglu et al. 2015). Climate change, in particular, is a
values for the embedding dimension L. The resulting uncertain- likely cause because it affects the two isotopes in very different
ties are indicated by means of shaded areas in these figures. ways (Beer et al. 2012), with 14 C being sensitive to long-term
changes in the ocean circulation (e.g., Hua et al. 2015), while
10
4.1. Multimillennial trend: possible climate influence Be is mainly sensitive to large-scale atmospheric dynamics. In
any case, it is quite clear that these long-term trends are unlikely
Here, we first consider the long-term primary SSA compo- to be of solar origin. For this reason, we decided to remove them
nents of the SN-14C and SN-10Be series. These are shown in from our original SN-14C and SN-10Be series to produce what
Fig. 3B. The shaded areas show the full range of computa- we hereafter refer to as the SN-14C-C and SN-10Be-C series,
tions for L-values ranging between 150 and 200 for 14 C and be- where the last “C” stands for “corrected”. The corresponding
tween 120 and 170 for 10 Be. These primary components are well solar activity reconstructions are shown in Fig. 5. (We note that
identified. since these reconstructions were corrected for long-term trends,
A150, page 5 of 10
 A&A 587, A150 (2016)
100 32

This work
80 L14 64
Sunspot number

H98
SS15
60 C14 128

Period (Years)
40 256

20 512

0 1024

1600 1700 1800 1900

Years
2048
Fig. 6. Decadally averaged sunspot numbers for the period 1600–
1900 AD. The thick gray curve represents this work (uncertainties are
not shown). Other curves correspond to sunspot number series: L14 4096
-6000 -5000 -4000 -3000 -2000 -1000 0 1000
(Lockwood et al. 2014), H98 (Hoyt & Schatten 1998), SS15 (Svalgaard Years -BC/AD
& Schatten 2015), and the C14 international sunspot number (v.2)
scaled with a factor 0.6 (Clette et al. 2014). Fig. 7. Wavelet coherence between the SN-14C and SN-10Be series.
The color code gives the value of the coherence from 0 (blue) to 1
(red). The arrows denote the relative phase between the series so that
the right-pointing arrows correspond to an exact in-phase and the left-
they only reflect relative changes within the solar activity, strictly pointing arrows to an exact anti-phase relation. Black contours delimit
speaking). the areas of high coherence (95% confidence level). White curves de-
We also show in Fig. 6 the comparison of the sunspot num- limit the cone of influence where results can be influenced by the edges
ber series for the period 1600–1900 AD reconstructed here from of the time series (beyond which the analysis is possibly biased).
14
C with other series based on sunspot counts and drawings
(Hoyt & Schatten 1998; Lockwood et al. 2014; Clette et al.
2014; Svalgaard & Schatten 2015). All series exhibit the same principal component analysis applied by Steinhilber et al. (2012)
centennial-scale evolution, while at shorter timescales the varia- to a composite solar activity reconstruction likewise revealed a
tions differ from one series to the next. However, the reconstruc- Hallstatt cycle in the heliospheric modulation potential that is
tion clearly lies within the range of the other sunspot series. synchronous with the reconstruction shown in Fig. 3C, although
it was not explicitly characterized.
This Hallstatt cycle has so far either been ascribed to cli-
4.2. Common signature of the Hallstatt cycle mate variability (Vasiliev & Dergachev 2002) or to geomagnetic
We now consider the second SSA components as shown in fluctuations, particularly geomagnetic pole migration (Vasiliev
Fig. 3C. These components are dominated by a ≈2400 yr period- et al. 2012). However, the fact that the signal we found is in
icity that is remarkably coherent between the SN-14C and SN- phase and of the same magnitude in the two cosmogenic isotope
10Be series. The formal Pearson correlation coefficient between reconstruction implies that it can hardly be of climatic origin.
the two curves shown in Fig. 3C is 0.77 ± 0.01, which is highly As already pointed out, 14 C and 10 Be respond differently to cli-
significant with p < 10−5 estimated using the non-parametric mate changes. In particular, 14 C is mostly affected by the ocean
random phase method (Ebisuzaki 1997; Usoskin et al. 2006). ventilation and mixing, while 10 Be (in particular its deposition
on central Greenland) is mainly affected by the large-scale at-
To confirm the significance of this observation, we also
mospheric circulation, particularily in the North Atlantic region
carried out a wavelet coherence analysis of the SN-14C and
(Field et al. 2006; Heikkilä et al. 2009). It can also hardly be of
SM-10Be series. The wavelet coherence is a normalized cross-
geomagnetic origin and related to geomagnetic pole migration,
spectrum of the two series and provides a measure of their co-
since 14 C is globally (hemispherically) mixed in the terrestrial
variance in time-frequency domain. It was calculated using the
system and insensitive to the migration of these poles. To repro-
Morlet basis and a code originally provided by Grinsted et al.
duce the observed Hallstatt cycle, the dipole moment (VADM)
(2004), but modified to adopt the non-parametric random phase
would have to vary, with the corresponding period, in a range
method for assessing confidence level (Ebisuzaki 1997; Usoskin
of ≈2 × 1022 A m2 (i.e., by about 20%). This is not supported
et al. 2006). The corresponding wavelet coherence is displayed
by any geomagnetic field reconstruction (Fig. 2). The SSA anal-
in Fig. 7. It shows that while the coherence is strong but in-
ysis of the geomagnetic series (not shown) does not yield the
termittent at shorter timescales and nearly absent at the longest
Hallstatt cycle.
timescales (cf. Usoskin et al. 2009), there is a wide band of very
We thus conclude that the ≈2400-yr Hallstatt cycle is most
high coherence that is in phase along the entire time interval for
likely a property of the long-term solar activity.
the periods of 2000–3000 years, consistent with the periods seen
in the second SSA component plotted in Fig. 3C.
It is therefore clear that the ≈2400 yr quasi-periodicity 5. New constraints on the temporal distribution
is common to both series and that it dominates their super- of grand minimum and grand maximum events
millennial timescale variability. It is related to the so-called
Hallstatt cycle that is known in Δ14 C (e.g., Damon & Sonett Using the reconstructed SN-14C-C and SN-10Be-C time series
1991; Vasiliev & Dergachev 2002; Ma 2007), but has been shown in Fig. 5, we provide a list of grand minima and max-
poorly documented until now in the 10 Be data (McCracken ima (Tables 1 and 2, respectively). In contrast to earlier work
et al. 2011; Hanslmeier et al. 2013). We also note that the (Usoskin et al. 2007; Steinhilber et al. 2008), we propose a
A150, page 6 of 10
 I. G. Usoskin et al.: Solar activity during the Holocene: the Hallstatt cycle and its consequence for grand minima and maxima

0.5 (see Fig. 3C). This feature was only hinted at in passing by
0.4 A) Min Steinhilber et al. (2012). Figure 8 shows the pdf (built using the
superposed epoch analysis) of the grand minima and maxima
0.3 times of occurrence relative to the time of occurrence of the near-
PDF

0.2
est Hallstatt cycle low and high. A tendency to cluster is clearly
observed. We checked that a similar tendency also appears when
0.1 considering the lists of grand maxima and minima provided by
Usoskin et al. (2007) and Inceoglu et al. (2015) over the same
0.0 time period. We speculate that this clustering might mean that
B) Max the probability of a switch of the solar dynamo from the nor-
0.3
mal mode to the grand minimum mode (resp. grand maximum
mode), according to Usoskin et al. (2014), is modulated by the
0.2
PDF

Hallstatt cycle. Although this interpretation relies on (necessar-

ily) arbitrary choices for defining grand minimum and maximum
0.1
events, it clearly deserves more investigation to better constrain
the behavior of the solar dynamo on centennial and millennial
0.0
timescales.
-1500 -1000 -500 0 500 1000 1500
Epoch time (years)
6. Conclusions
Fig. 8. Probability density function (pdf) of the time of occurrence of
grand minima (panel A)) and grand maxima (panel B)) relative to the Here we presented new reconstructions of solar activity (quanti-
time of occurrence of the nearest low and high, respectively, of the fied in terms of sunspot numbers) spanning the past 9000 yr (ta-
Hallstatt cycle, using the superposed epoch analysis. The times of oc- bles are available at the CDS) and assessed their accuracy using
currence of highs and lows of the Hallstatt cycle are defined by con- different geomagnetic field reconstructions and current cosmo-
sidering the average of the two curves (second SSA components of the genic isotope production models.
SN-14C and SN-10Be series) shown in Fig. 3C (leading to 5530 BC, We found that the primary SSA components of the recon-
3650 BC, 810 BC, and 1510 AD for the lows and to 6670 BC, 4600 BC, structions that are based on 14 C and 10 Be are significantly dif-
1970 BC, and 160 AD for the highs). ferent. These primary components probably reflect long-term
changes in the terrestrial system that affect the 14 C and 10 Be
isotopes in different ways (ocean circulation and/or large-scale
conservative list based on both the SN-14C-C and SN-10Be-C atmospheric transport). These components were therefore re-
reconstructions, meaning that we only list events that are simul- moved to produce meaningful corrected sunspot number se-
taneously seen in both reconstructions (a time adjustment of the ries. In contrast, the secondary SSA components of the 14 C and
10
Be-based series for ±40 years was allowed owing to the dating 10
Be based reconstructions revealed a common remarkably syn-
uncertainties (Muscheler et al. 2014)). To identify grand min- chronous ≈2400-year quasi-periodicity. We therefore concluded
ima, the following criterion was used (with one exception, see that this so-called Hallstatt periodicity most likely reflects some
below): the event in both reconstructions (using the mean of periodicity in the solar activity.
the ensemble) must correspond to a SN value below a thresh- From the two cosmogenic isotope records, we finally defined
old value of S N = 20 for at least 30 years. Although the event a conservative list of grand minima and grand maxima covering
ca. 6385 BC lies slightly above this threshold, we also consid- the past 9 millennia. An important finding is that the grand min-
ered it as a grand minimum because it clearly has a Spörer-type ima and maxima occurred intermittently over the studied period,
(i.e., prolonged) shape and occurs in both series. It is possible with clustering near minima and maxima of the Hallstatt cycle,
that the level of activity was slightly overestimated for this event respectively. The Hallstatt cycle thus appears to be a long-term
because of uncertainties in dipole moment evolution during the feature of solar activity that needs to be taken into account in
older part of the time interval. To identify grand maxima, we models of solar dynamo.
similarly requested the events to have a SN value exceeding the
threshold of S N = 55 for at least 30 years in both reconstruc-
Acknowledgements. We are thankful to Raphael Roth and Fortunate Joos for
tions. We thus defined 20 grand minima with a total duration providing data on 14 C production rates and for useful discussions on the carbon
of 1460 years (≈17% of time) and 14 grand maxima with a to- cycle. Support by the Academy of Finland to the ReSoLVE Center of Excellence
tal duration of 750 years (≈8%). These numbers are similar to (project no. 272157) and by IPGP through its invitation program is acknowl-
those estimated earlier by Usoskin et al. (2007), although we edged. Y.G. and G.K. were partly financed by grant N 14.Z50.31.0017 of the
note that more grand maxima are now identified. In contrast, Russian Ministry of Science and Education. This is IPGP contribution No. 3699.
these numbers are significantly lower than those recently esti-
mated by Inceoglu et al. (2015), who relied on a different type
References
of analysis and set of criteria (somewhat less restrictive, leading
to the identification of a larger set of events, which is essentially Bard, E., Raisbeck, G., Yiou, F., & Jouzel, J. 1997, Earth Planet. Sci. Lett., 150,
inclusive of the set of events we identified here, see Tables 1 453
Bazilevskaya, G. A., Cliver, E. W., Kovaltsov, G. A., et al. 2014, Space Sci. Rev.,
and 2). 186, 409
Times of grand minima and maxima identified in this way Beer, J., McCracken, K., & von Steiger, R. 2012, Cosmogenic Radionuclides:
are shown in Figs. 3C and 5 as circles and stars, respectively. An Theory and Applications in the Terrestrial and Space Environments (Berlin:
important observation from this figure is that the occurrence of Springer)
Brown, M. C., Donadini, F., Korte, M., et al. 2015, Earth Plan. Space, 67, 83
grand minima and maxima appears intermittent in time. Grand Cai, S., Tauxe, L., Deng, C., et al. 2014, Earth Planet. Sci. Lett., 392, 217
minima seem to be closely related to the Hallstatt cycle, being Cai, S., Chen, W., Tauxe, L., et al. 2015, J. Geophys. Res., 120, 2056
clearly more numerous during the lows of the Hallstatt cycle Cameron, R., & Schüssler, M. 2015, Science, 347, 1333

A150, page 7 of 10
 A&A 587, A150 (2016)

Clette, F., Svalgaard, L., Vaquero, J., & Cliver, E. 2014, Space Sci. Rev., 186, 35 Krivova, N., Balmaceda, L., & Solanki, S. 2007, A&A, 467, 335
Cromwell, G., Tauxe, L., Staudigel, H., & Ron, H. 2015, Phys. Earth Planet. Licht, A., Hulot, G., Gallet, Y., & Thébault, E. 2013, Phys. Earth Planet. Inter.,
Inter., 241, 44 224, 38
Damon, P., & Sonett, C. 1991, in The Sun in Time, eds. C. Sonett, M. Giampapa, Lockwood, M., Owens, M. J., & Barnard, L. 2014, J. Geophys. Res., Space
& M. Matthews (Tucson, USA: University of Arizona Press), 360 Phys., 119, 5172
de Groot, L. V., Béguin, A., Kosters, M. E., et al. 2015, Earth Planet. Sci. Lett., Ma, L. H. 2007, Sol. Phys., 245, 411
419, 154 Mane, R. 1981, in Dynamical Systems and Turbulence, eds. D. Rand, &
Di Chiara, A., Tauxe, L., & Speranza, F. 2014, Phys. Earth Planet. Inter., 234, 1 L. Young (Springer-Verlag), Lect. Notes Math., 898, 230
Donadini, F., Korhonen, K., Riisager, P., & Pesonen, L. J. 2006, EOS McCracken, K., Beer, J., Steinhilber, F., & Abreu, J. 2011, Space Sci. Rev., 324
Transactions, 87, 137 Miyake, F., Nagaya, K., Masuda, K., & Nakamura, T. 2012, Nature, 486, 240
Ebisuzaki, W. 1997, J. Climate, 10, 2147 Miyake, F., Masuda, K., & Nakamura, T. 2013, Nat. Comm., 4, 1748
Field, C., Schmidt, G., Koch, D., & Salyk, C. 2006, J. Geophys. Res., 111, Moss, D., Sokoloff, D., Usoskin, I., & Tutubalin, V. 2008, Sol. Phys., 250, 221
D15107 Muscheler, R., Joos, F., Beer, J., et al. 2007, Quat. Sci. Rev., 26, 82
Gallet, Y., & Al Maqdissi, M. 2010, Akkadica, 131, 29 Muscheler, R., Adolphi, F., & Knudsen, M. F. 2014, Quat. Sci. Rev., 106, 81
Gallet, Y., & Butterlin, P. 2015, Archaeometry, 57, S1, 263 Nachasova, I., & Burakov, K. 1995, Geomagn. Aeron., 35
Gallet, Y., Le Goff, M., Genevey, A., Margueron, J., & Matthiae, P. 2008, Nachasova, I., & Burakov, K. 1998, Geomagn. Aeron., 38
Geophys. Res. Lett., 35, 2307 Nilsson, A., Holme, R., Korte, M., Suttie, N., & Hill, M. 2014, Geophys. J. Int.,
Gallet, Y., Genevey, A., Le Goff, M., et al. 2009, Comptes Rendus Physique, 10, 198, 229
630 Osete, M. L., Catanzariti, G., Chauvin, A., et al. 2015, Phys. Earth Planet. Inter.,
Gallet, Y., Molist Montaña, M., Genevey, A., et al. 2015, Phys. Earth Planet. 242, 24
Inter., 238, 89 Passos, D., Nandy, D., Hazra, S., & Lopes, I. 2014, A&A, 563, A18
Genevey, A., Gallet, Y., Constable, C. G., Korte, M., & Hulot, G. 2008, Pavón-Carrasco, F. J., Osete, M. L., Torta, J. M., & De Santis, A. 2014, Earth
Geochem., Geophys., Geosyst., 9, 04038 Planet. Sci. Lett., 388, 98
Genevey, A., Gallet, Y., Thébault, E., Jesset, S., & Le Goff, M. 2013, Geochem., Reimer, P. J., Baillie, M. G. L., Bard, E., et al. 2009, Radiocarbon, 51, 1111
Geophys., Geosys., 14, 2858 Roth, R., & Joos, F. 2013, Clim. Past, 9, 1879
Golub, G., & Kahan, W. 1965, J. Soc. Indust. Appl. Math. B: Num. Anal., 2, 205 Shaar, R., Tauxe, L., Ben-Yosef, E., et al. 2015, Geochem., Geophys., Geosyst.,
Golyandina, N., Nekrutkin, V., & Zhigljavsky, A. 2001, Analysis of Time Series 16, 195
Structure: SSA and related techniques (Boca Raton, London, New York, Snowball, I., & Muscheler, R. 2007, Holocene, 17, 851
Washington DC: Chapman and Hall/CRC) Solanki, S., Schüssler, M., & Fligge, M. 2000, Nature, 408, 445
Grinsted, A., Moore, J. C., & Jevrejeva, S. 2004, Nonlin. Processes Geophys., Solanki, S., Usoskin, I., Kromer, B., Schüssler, M., & Beer, J. 2004, Nature, 431,
11, 561 1084
Hanslmeier, A., Brajša, R., Čalogović, J., et al. 2013, A&A, 550, A6 Steinhilber, F., Abreu, J. A., & Beer, J. 2008, Astrophys. Space Sci. Trans., 4, 1
Hassani, H. 2007, Journal of Data Science, 5, 239 Steinhilber, F., Abreu, J., Beer, J., et al. 2012, Proc. Nat. Acad. Sci. USA, 109,
Heikkilä, U., Beer, J., & Feichter, J. 2009, Atmos. Chem. Phys., 9, 515 5967
Hong, H., Yu, Y., Lee, C. H., et al. 2013, Earth Planet. Sci. Lett., 383, 142 Stillinger, M. D., Feinberg, J. M., & Frahm, E. 2015, J. Archaeol. Sci., 53, 345
Hoyt, D. V., & Schatten, K. H. 1998, Sol. Phys., 179, 189 Svalgaard, L., & Schatten, K. H. 2015, ArXiv e-prints [arXiv:1506.00755]
Hua, Q., Webb, G. E., Zhao, J., et al. 2015, Earth Planet. Sci. Lett., 422, 33 Takens, F. 1981, in Dynamical Systems and Turbulence, eds. D. Rand, &
Inceoglu, F., Simoniello, R., Knudsen, V. F., et al. 2015, A&A, 577, A20 L. Young (Springer-Verlag), Lect. Notes Math., 898, 366
Kapper, K. L., Donadini, F., & Hirt, A. M. 2015, Phys. Earth Planet. Inter., 241, Thébault, E., & Gallet, Y. 2010, Geophys. Res. Lett., 37, L22303
21 Usoskin, I. G. 2013, Liv. Rev. Sol. Phys., 10, 1
Karak, B. B., Kitchatinov, L. L., & Brandenburg, A. 2015, ApJ, 803, 95 Usoskin, I. G., Alanko-Huotari, K., Kovaltsov, G. A., & Mursula, K. 2005,
Kissel, C., Laj, C., Rodriguez-Gonzalez, A., et al. 2015, Earth Planet. Sci. Lett., J. Geophys. Res., 110, A12108
430, 178 Usoskin, I. G., Voiculescu, M., Kovaltsov, G. A., & Mursula, K. 2006,
Kittler, J., & Young, P. C. 1973, Pattern Recognition, 5, 335 J. Atmosph. Sol.-Terr. Phys., 68, 2164
Knudsen, M. F., Riisager, P., Donadini, F., et al. 2008, Earth Planet. Sci. Lett., Usoskin, I. G., Solanki, S. K., & Kovaltsov, G. A. 2007, A&A, 471, 301
272, 319 Usoskin, I. G., Horiuchi, K., Solanki, S., Kovaltsov, G. A., & Bard, E. 2009,
Korhonen, K., Donadini, F., Riisager, P., & Pesonen, L. J. 2008, Geochem., J. Geophys. Res., 114, A03112
Geophys., Geosys., 9, Q04029 Usoskin, I. G., Kromer, B., Ludlow, F., et al. 2013, A&A, 552, L3
Korte, M., & Constable, C. 2005, Earth Planet. Sci. Lett., 236, 348 Usoskin, I. G., Hulot, G., Gallet, Y., et al. 2014, A&A, 562, L10
Korte, M., & Constable, C. 2011, Phys. Earth Planet. Inter., 188, 247 Vasiliev, S., & Dergachev, V. 2002, Annales Geophys., 20, 115
Korte, M., Donadini, F., & Constable, C. G. 2009, Geochem. Geophys. Geosys., Vasiliev, S. S., Dergachev, V. A., Raspopov, O. M., & Jungner, H. 2012,
10, Q06008 Geomagn. Aeron., 52, 121
Korte, M., Constable, C., Donadini, F., & Holme, R. 2011, Earth Planet. Sci. Vautard, R., & Ghil, M. 1989, Physica D Nonlin. Phenom., 35, 395
Lett., 312, 497 Vautard, R., Yiou, P., & Ghil, M. 1992, Physica D Nonlin. Phen., 58, 95
Kovaltsov, G., & Usoskin, I. 2010, Earth Planet. Sci. Lett., 291, 182 Vonmoos, M., Beer, J., & Muscheler, R. 2006, J. Geophys. Res., 111, A10105
Kovaltsov, G., Mishev, A., & Usoskin, I. 2012, Earth Planet. Sci. Lett., 337, Webber, W. R., & Higbie, P. R. 2009, J. Geophys. Res., 114, A02103
114 Yiou, F., Raisbeck, G., Baumgartner, S., et al. 1997, J. Geophys. Res., 102, 26783

A150, page 8 of 10
 I. G. Usoskin et al.: Solar activity during the Holocene: the Hallstatt cycle and its consequence for grand minima and maxima
13 13
Appendix A: GMAG.9k axial dipole evolution A) ΔT=200 yr, ΔS=10 x10
o o
B) ΔT=500 yr, ΔS=10 x10
o o

VADM (10 A m )
2
12 12

Here we used the GEOMAGIA50.v3 database (Donadini et al.

11 11

22
2006; Korhonen et al. 2008; Brown et al. 2015) to which we
added recent archeo- and paleointensity results (Cai et al. 2014, 10 10

2015; Cromwell et al. 2015; de Groot et al. 2015; Di Chiara 9 9

et al. 2014; Gallet et al. 2008, 2009, 2015; Gallet & Al Maqdissi
2010; Hong et al. 2013; Kapper et al. 2015; Kissel et al. 2015; 8 8

Osete et al. 2015; Shaar et al. 2015; Stillinger et al. 2015). 7 7

-1500 -1000 -500 0 500 1000 1500 2000 -1500 -1000 -500 0 500 1000 1500 2000
Concerning the data compiled in this version of GEOMAGIA, 13 13
o o o o
the Mesopotamian data from Nachasova & Burakov (1995, C) ΔT=200 yr, ΔS=30 x30 D) ΔT=500 yr, ΔS=30 x30

VADM (10 A m )
2
12 12
1998) were modified according to Gallet et al. (2015). Further
revisions were also discussed in Genevey et al. (2013; also A.

22
11 11

Genevey, pers. comm.) and Gallet & Butterlin (2015). 10 10

Axial dipole evolution over the past 9000 years, referred to
9 9
as GMAG.9k, was constrained by virtual axial dipole moments
(VADM) averaged over sliding windows of 200 years between 8 8

1500 BC and 2000 AD and of 500 years between 7000 BC

7 7
and 1500 BC shifted by 10 years and using a regional weight- -1500 -1000 -500 0 500 1000 1500 2000 -1500 -1000 -500 0 500 1000 1500 2000

ing scheme over regions of 10◦ width. Weights of one third or Epoch Epoch
two thirds were assigned to the regions that contained at a given Fig. A.1. Comparison between VADM curves computed between
time interval only one or two individual intensity (VADM) data 1500 BC and 2000 AD using sliding windows of ΔT = 200 years (pan-
points, respectively. For those intensity data with no age uncer- els A), C)) and 500 years (B), D)) shifted by 50 years, and using a
tainties provided in the GEOMAGIA50.v3 database, we used weighting over regions of ΔS = 10◦ × 10◦ (A), B)) and ΔS = 30◦ × 30◦
the same approach as in Licht et al. (2013). We computed the (C), D) width. The thick black line exhibits the averaged VADM com-
means of the known age uncertainties over 500 yr long time in- puted using a bootstrap scheme (see main text and legend of Tables C.1
tervals between 1000 BC and 2000 AD, over 2000 yr interval and C.2 available at CDS), with its 1σ uncertainties (dotted lines) and
between 3000 BC and 1000 BC and over 4000 yr interval be- the envelope of possible VADM values (gray lines).
tween 7000 BC and 3000 BC. After multiplication by a factor
of 1.5, the corresponding values were assigned to the intensity 13
o o
13
o o
data with unknown age uncertainties within the periods of con- 12 A) ΔT=500 yr, ΔS=10 x10 12 B) ΔT=1000 yr, ΔS=10 x10
VADM (10 A m )
2

cern. Following Knudsen et al. (2008), when no experimental 11 11

22

errors were provided on the data, we assigned errors amount- 10 10

ing 25% of the corresponding intensity values. We also note that 9 9

for the archeomagnetic data for which no attempt was made to 8 8

take the cooling rate effect on thermoremanent magnetization 7 7

acquisition into account, we systematically implemented a cool- 6 6
ing rate correction of 5% decrease (see for instance in Genevey 5 5
et al. 2008). Mean VADM estimates were then derived using 13
-6000 -4000 -2000 0 2000
13
-6000 -4000 -2000 0 2000

a bootstrap technique to account for the noise in the available C) ΔT=500 yr, ΔS=30 x30
o o
D) ΔT=1000 yr, ΔS=30 x30
o o
VADM (10 A m )

12 12
2

paleo- and archeointensity data within their age uncertainties 11 11

and within their 2σ experimental uncertainties. 1000 VADM
22

10 10
curves, also derived using different randomly attributed locations
9 9
of the weighting regions, were hence determined, whose statis-
8 8
tics are summarized in Tables C.1 and C.2 (available in CDS)
7 7
for the periods 1500 BC-2000 AD and 7000 BC-1500 BC, re-
6 6
spectively. These tables provide for each epoch (first column) a
5 5
mean VADM (second column), a standard deviation (third col- -6000 -4000 -2000 0 2000 -6000 -4000 -2000 0 2000
umn), and the maximum and minimum values defining the enve- Epoch Epoch
lope of possible VADM results (fourth and fifth columns). The
variability of VADM is shown in Figs. A.1 (between 1500 BC Fig. A.2. Comparison between VADM curves computed between
7000 BC and 2000 AD using sliding windows of ΔT = 500 years (pan-
and 2000 AD) and A.2 (between 7000 BC and 2000 AD).
els A), C)) and 1000 years (B), D)) shifted by 50 years, and using a
weighting over regions of ΔS = 10◦ × 10◦ (A), B)) and ΔS = 30◦ × 30◦
(C), D) width. The thick black line exhibits the averaged VADM com-
Appendix B: Identification of long-term trends puted using a bootstrap scheme (see main text and legend of Tables C.1
and C.2 available at CDS), with its 1σ uncertainties (dotted lines) and
by singular spectrum analysis the envelope of possible VADM values (gray lines).
The non-parametric singular spectrum analysis (SSA) of time
series is based on the Karhunen-Loeve spectral decomposition
theorem (Kittler & Young 1973) and the Mané-Takens embed-
ded theorem (Mane 1981; Takens 1981). It allows a time series The basic version of SSA that we used consists of four
to be decomposed into several components with distinct tempo- straightforward steps (see, e.g., Golyandina et al. 2001; Hassani
ral behaviors and is very convenient to identify long-term trends 2007): embedding, singular value decomposition, grouping, and
and quasi-periodic oscillations. reconstructions.
A150, page 9 of 10
 A&A 587, A150 (2016)

When considering a real-value time series x (x1 , x2 , . . . , xN ), “empirical orthogonal functions”) of the matrix D = X XT . If we
the first step of this SSA consists of embedding this series into denote with d the number
√ of nonzero eigenvalues, we may next
an L-dimensional vector space, using lagged copies of x to form define Vi = XT Ui / λi (i = 1, . . . , d). Then, the trajectory matrix
the so-called trajectory (Hankel) matrix (where K = N − L + 1
√ as a sum of elementary matrices X = X1 +· · ·+Xd ,
can be written
and L is a parameter to be chosen), where Xi = λi Ui ViT .
  Once this decomposition has been completed, the third step
 x1 x2 . . . xK  consists of the construction of groups of components by rear-
 x2 x3 . . . xK+1 
ranging X into X = XG1 + XG2 + . . . , where each XG is the
X =  .
 . . . . . . . . . . . . 
(B.1)
sum (group) of a number of Xi . The choice of the components
 x x xN  to be considered
L L+1 . . . √ in each group is made empirically by grouping
eigentriples ( λ, U, V) with similar eigenvalues. Finally, a diag-
The second step consists of performing a singular value decom- onal averaging is applied to each XG to make it take the form of
position (Golub & Kahan 1965) of the trajectory matrix. This a trajectory matrix, from which the associated time series com-
provides a set of L eigenvalues λi (arranged in decreasing order ponent x̃G of length N can be recovered (for details, see, e.g.,
λ1 ≥ λ2 ≥ · · · ≥ λL ≥ 0) and eigenvectors Ui (often called Golyandina et al. 2001; Hassani 2007).

A150, page 10 of 10