You are on page 1of 6

MNRAS 468, 5014–5019 (2017) doi:10.

1093/mnras/stx627

Most frequent value statistics and distribution of 7 Li


abundance observations

J. Zhang1,2‹
1 Department of Mathematics and Physics, Hebei GEO University, Shijiazhuang 050016, China
2 National Astronomical Observatories/Yunnan Observatory, Chinese Academy of Sciences, Kunming, 650011, China

Accepted 2017 March 10. Received 2017 March 1; in original form 2017 January 23

ABSTRACT
The problem of primordial lithium is one of the key issues in many astrophysical fields, espe-
cially nuclear astrophysics and cosmology. We apply a novel modern statistical method – the
most frequent value (MFV) procedure – to study the distribution modelling for 7 Li abundance
observations, constructing a robust and effective astrostatistical algorithm. Considering prior
non-Gaussian error distributions of measurements of astrophysical quantities, the robust MFV
method is essential as a standard statistical algorithm, according to the minimization for the in-
formation loss principle. Associated with global optimization, the MFV procedure has already
been shown to be a more powerful tool than some classic statistical algorithms originating
from the least-squares or maximum likelihood principle for dealing with the unknown distri-
bution problems of observed astrophysical quantities. Applying the MFV algorithm, we obtain
the MFV M = 2.205 and the dihesion ε = 0.067 of A(Li) measurements, which demonstrate
certain advantages to solving problems with a non-Gaussian error distribution.
Key words: nuclear reactions, nucleosynthesis, abundances – methods: statistical – stars:
abundances – star: statistics.

lithium problems imply the possibilities of other unknown mecha-


1 I N T RO D U C T I O N
nisms or non-standard physics solutions.
It is common knowledge that, based on the standard big bang model, From another perspective of stellar evolution, modern astrophys-
we can understand the nucleosynthesis of the lightest nuclei (mainly ical theory encounters an important unsolved problem of lithium
including isotopes of H, He and Li, such as deuterium, 3 He, 4 He, chemical evolution in the galaxy (Guiglion et al. 2016). It is well
and so on) from the first ∼1 s to ∼3 min at the beginning of the Uni- known that lithium can be synthesized via nuclear reactions in the
verse (Fields 2011; Bertulani & Kajino 2016; Cyburt et al. 2016; stars. However, lithium is an easily destroyed element in a more
Gustavino et al. 2016). The calculated results of big bang nucle- moderate temperature case (T ∼ 2.5 × 106 K; e.g. stellar interiors
osynthesis (BBN) give us a useful probe to demonstrate the diver- via the 7 Li(p, α)4 He reaction) and a relatively lower temperature
sity of the chemical evolution in the early cosmos. There is no doubt (T ∼ 2 × 106 K) for 6 Li. So, the observed isotopic abundances of
that modern cosmology has made tremendous progress, due to the lithium have provided a critical constraint not only for standard BBN
concordance of theoretical predictions and Wilkinson Microwave but also for stellar evolution (Casey et al. 2016; Cyburt et al. 2016).
Anisotropy Probe (WMAP) observations of the cosmic microwave Therefore, measurements of lithium abundance have played a vital
background, with the exception of the primordial lithium problem role in the astrophysics and astronomy of the early Universe. By
(Fields 2011). There is a large non-negligible mismatch between observing the spectra of absorption lines of atmospheres in stars,
the observed 7 Li abundances and the BBN+WMAP predictions. Be- astronomers have obtained the lithium abundance. In view of the
sides, the second lithium problem involves 6 Li formed from the difficulties of measuring 6 Li (which needs higher resolution R ≈
2
H(α, γ )6 Li reaction. Applying the first fruitful measurements of 100 000 and high signal-to-noise ≈500), we generally investigate
the 2 H(α, γ )6 Li reaction, Anders et al. (2014) obtained the big bang the observed abundance of 7 Li.
lithium isotopic ratio, which cannot be explained by the standard There are many astrophysical channels to produce lithium, in-
big bang model (see also Mukhamedzhanov & Shubhchintak 2016). cluding Galactic cosmic rays (Reeves 1970), core-collapse (Hart-
Using comparisons of observational data and theory, the primordial mann et al. 1999), novae (Arnould & Norgaard 1975; Tajitsu
et al. 2015), cool bottom-burning in low-mass giants (Sackmann
& Boothroyd 1999), hot bottom-burning in asymptotic giant branch
(AGB) stars (Abia, Pavlenko & de Laverny 1999), etc. However,
 E-mail: zhangphysics@126.com the problem of observed Li-rich giants is still not well understood

C 2017 The Author

Published by Oxford University Press on behalf of the Royal Astronomical Society


MFV statistics and distribution of 7 Li 5015
(Casey et al. 2016). According to classical stellar structure and duce the MFV procedure and the numerical calculation algorithm.
evolution, the lithium abundance should present a tendency to de- The final section contains results and a discussion.
cline along with the giant phase, especially after first dreg up. On
the contrary, relatively younger Population I stars illustrate unex-
pectedly higher abundance than that of metal-poor Population II 2 O B S E RVAT I O N S A N D DATA S E T S E L E C T I O N
stars. Most puzzlingly, there exist more than a dozen red giants that
Because metal-poor Population II stars probably keep the primor-
show very high lithium abundances. Some abundances A(Li) even
dial chemical abundance of their predecessor molecular cloud in
exceed 3, largely beyond the prediction of standard BBN, where
the early galaxy (Sneden, Cowan & Gallino 2008; Zhang, Cui
A(Li) = 12 + lg[n(Li)/n(H)] and n denotes the atom number density
& Zhang 2010a,b), the observational abundances of lithium in
(Kirby et al. 2012). This means that other unknown mechanisms
metal-poor dwarfs or early subgiants are the critical observed con-
of forming lithium cannot be ruled out, probably involving mixing
straints, independent of that of standard BBN+WMAP for the cos-
mechanisms.
mic lithium problem. Spite & Spite (1982a,b) first showed that
In addition, the meteoritic and solar photospheric abundances
shallow convection zones in metal-deficient warm dwarfs should
of lithium are 3.26 and 1.05 dex, respectively (Grevesse, Asplund
preserve the fossil abundances of lithium, and that they demon-
& Sauval 2007; Lodders & Palme 2009), which demonstrate the
strated approximately the same observational lithium abundance
lithium chemical evolution history. Obviously, the observed abun-
(i.e. the Spite plateau), no matter how temperature and metallicity
dance of lithium is a powerful probe for understanding the pri-
vary within the scope of −2.4 ≤ [Fe/H] ≤ −1.4 and 5700  Teff 
mordial lithium problem in the early cosmos. For a few decades,
6250 K (Bonifacio et al. 2007; Spite, Spite & Bonifacio 2012; Fu
researchers have measured many samples of primordial abundances
et al. 2015). Therefore, the identification of lithium abundance on
in standard BBN and photospheric lithium abundances, observed in
the Spite plateau is pivotal in order to constrain the standard the-
metal-poor stars including dwarfs and giants. However, the mea-
ory, and this has inspired us to evaluate the more accurate lithium
surements of lithium unexpectedly disagree with the predictions of
abundances on the Spite plateau according to the new statistic
standard theory. There is a vital gap between the primordial lithium
method.
abundance, A(Li) ∼ 2.6 dex (Spergel et al. 2003, 2007) and that of
In particular, we do not take the non-local thermodynamic equi-
the Spite plateau, A(Li) ∼ 2.2 dex (Spite & Spite 1982a,b). Clearly,
librium (NLTE) into consideration because of the small comprehen-
it is very important to determine a reliable value of lithium abun-
sive effects of three-dimensional (3D) hydrodynamic characteristics
dance on the Spite plateau, and this has motivated us to investigate
and the NLTE of stellar atmospheres (Korn et al. 2006). Following
a new reliable statistical method.
the selection of Crandall et al. (2015) and Spite et al. (2012), we
Tremendous advances have been made in astrostatistics studies
choose metal-poor stars with −2.8 ≤ [Fe/H] ≤ −2.0 and 5900  Teff
related to many astrophysical fields (Podariu et al. 2001; Chen &
in order to reflect the primordial lithium abundance. Consequently,
Ratra 2003, 2011; Wall & Jenkins 2012; Crandall & Ratra 2015;
stars beyond these ranges of metallicity and effective temperature
Chen, Kumar & Ratra 2017). There are various statistical methods
will be ruled out. Ultimately, we simply analyse the 66 measure-
to deal with the problems of astrophysical measurements (e.g. big
ments of lithium abundance similar to Crandall et al. (2015), which
data of abundances). One of the most common means of astrostatis-
are shown in Fig. 1.
tics is the conventional χ 2 approach, which assumes a statistically
independent Gaussian error distribution and known error bars with-
out systematic effects (Gott et al. 2001). Gott et al. argued that
3 THEORETICAL MODEL
median statistics was another useful method in comparison with the
χ 2 approach, due to fewer assumptions about the data set. Moreover, In astrophysical fields, the parameter estimation of the observed
Avelino, Martins & Pinto (2002) further promoted previous work measurements is a vital task. For years, many algorithms have been
and applied modified median statistics to investigate Type Ia super- constructed. The most common methods are the weight mean and
novae data. The more widely applied median statistics, discussed least-squares method (LSM), which require that the observed mea-
in Chen & Ratra (2003), constrains the mass density of the cosmos. surements obey the Gaussian distribution. However, the observa-
Furthermore, Chen & Ratra (2011) used a median statistical method tions of astrophysical quantities show that the LSM is often more
to assess the Hubble constant (H0 ) and they estimated systematic vulnerable to the effects of outliers in measurements than other ro-
errors for different statistical techniques. Based on the weighted bust methods. In particular, if the error distribution of the observed
mean and median statistics, Crandall & Ratra (2014) found good data is non-Gaussian, then the application of the LSM is impossibly
consistency by estimating cosmological parameter values. Recently, optimal. Unfortunately, some statistical methods based on the LSM
Crandall et al. have investigated the error distribution of the distance often are used to deal with astrophysical measurements without
moduli measurements of the Large Magellanic Cloud (Crandall & considering the non-Gaussian error distribution.
Ratra 2015) and 7 Li abundance observations (Crandall, Houston & The error distribution is of great significance in the analysis of the
Ratra 2015), which have illustrated the non-Gaussianity of error measurements. Obviously, it is necessary to apply two procedures to
distribution. It can be seen that the uncertainties and distributions assess the characteristics of Li abundances. One method is to apply
of scientific measurements are attracting more widespread interest the Gaussian approximation, which is dependent on the maximum
in physical and astrophysical fields (Bailey 2017). likelihood principle or the LSM. The other method is to use the
In this paper, we use a novel method – the most frequent value MFV procedure via the iteration algorithm, which relies on the
(MFV) procedure (Steiner 1988, 1991, 1997; Steiner & Haja- minimization for the information loss principle, almost regardless
gos 2001; Kemp 2006; Szucs, Civan & Virag 2006; Szegedi 2013; of the given priori limiting distribution. The MFV procedure is
Szegedi & Dobroka 2014) – to evaluate lithium abundances on the a new efficient robust method, proposed by Steiner (1991, 1997)
Spite plateau. The MFV technique has been applied in some sub- and Szucs et al. (2006) for geophysics. In order to remove the
jects, but mainly in geosciences at present. In the next section, we weakness involving the low effectiveness and high sensitivity of
describe the observation and data selection. In Section 3, we intro- sparsely distributed big data, Steiner argued for the adoption of the

MNRAS 468, 5014–5019 (2017)


5016 J. Zhang

Figure 1. Count, probability density (red line) and box-plot of Li observations. [A colour version of this figure is available in the online version.]

maximum reciprocals principle, 4 R E S U LT S A N D D I S C U S S I O N


 1 We have fitted Weibull, gamma and lognormal distributions to A(Li)
= max, (1)
i
Xi2 + S 2 data using maximum likelihood estimation, as shown in Fig. 2.
Four classical goodness-of-fit plots are shown, including a den-
where Xi denotes the residuals and deviations and S denotes the sity plot, the empirical cumulative distribution function (CDF)
measurement error. Based on the minimization of the information plot, Q–Q plot and P–P plot in statistics. With more observa-
divergence (relative entropy) denoting the measure of information tions of Li abundances, we can obviously demonstrate the presence
loss (Huber 1981; Steiner 1991, 1997), Steiner proposed the MFV and magnitude of discrepancies among different distributions and
procedure and the scaling factor ε called dihesion, in order to evalu- measurements.
ate the parameter of scale in such a way as to reduce the information By fitting the lithium abundance measurements to common dis-
loss. Furthermore, the MFV and the dihesion can be obtained via tributions including Cauchy and Student’s t, Crandall et al. (2015)
iterations. In the (j + 1)th step of the MFV procedure, the corre- gave A(Li) = 2.21 ± 0.065 and found the characteristics of the
sponding formula of iterations for the MFV M are non-Gaussian error distribution of lithium observations in the Spite
 n  2   2  2  plateau. The non-Gaussian data might not be uncommon in astro-
i=1 εj x i / εj + xi − M j physics and other fields (Chen, Gott & Ratra 2003; Bailey 2017). It
Mj +1 =     2  . (2) is very interesting and important to distinguish between the Gaus-
n
i=1 εj / εj + xi − Mj
2 2
sianity and non-Gaussianity when observers deal with big data anal-
ysis. Because of the central limit theorem, Gaussian statistics have
Here, xi is a series of observations and the dihesion εj can be calcu-
been applied to nearly every field. However, if errors do not match
lated by
the Gaussian assumption, researchers have to be cautious when


n  2   2  2 2 choosing the appropriate statistical method. Even assuming that the
3 i=1 εk xi − Mj
4
/ εk + xi − Mj observed distribution curve seems to be bell-shaped, we still can-
εk+1
2
=
 .
2 2
(3) not determine whether it is Gaussian. For example, considering a
n 
i=1 εk / εk + xi − Mj
4 2
Student’s t-distibution with one degree of freedom (i.e. the stan-
dard Cauchy distribution or Breit–Wigner distribution), we cannot
In the first step, the initial value M0 for the iteration is given as obtain the finite mean and variance of the Cauchy distribution via
the average value of the measurements, and the initial value of ε is integration (Walck 2007). This implies an exception of the central
defined as limit theorem.
√ From another perspective, if we suppose two independent
3
ε0 = (xmax − xmin ). (4) stochastic variables A and B with standard Gaussian distributions,
2 the distribution of the ratio A/B is, surprisingly, the Cauchy form.
We can set the threshold criterion that constrains the accuracy in More amazingly, the distribution of the superposition of many inde-
iterations. After a defined number of iterations, the calculation can pendent Cauchy-distributed variables still remains a Cauchy form,
be stopped when the dihesion is smaller than the threshold value regardless of the sample size. Obviously, this demonstrates the lim-
(e.g. 10−5 ). Finally, the MFV M and the dihesion ε are given as itation of applying standard Gaussian statistics to some physical
a result. In particular, unlike the standard deviation in the LSM, cases, such as the ratio form problem, even though both numerator
the dihesion ε is an important parameter and is not sensitive to the and denominator obey the Gaussian distribution (Walck 2007). Fur-
outliers. thermore, it is essential that not all observed symmetrical unimodal

MNRAS 468, 5014–5019 (2017)


MFV statistics and distribution of 7 Li 5017

Figure 2. Four goodness-of-fit plots for various distributions fitted to observed abundances of Li, including Weibull, gamma and lognormal distributions fitted
to serving sizes from the measurements. [A colour version of this figure is available in the online version.]

distributions are Gaussian. Thus, we have to investigate a robust most, and is not affected by the extremal observed points. Nonethe-
statistical approach that can be applied to various distributions in- less, the mode is not fit when the occurring numbers are the same.
cluding Gaussian and non-Gaussian distributions. Also, there is no unique mode when the distribution of observations
Here, using the MFV procedure, we determine the Li abundance is multi-peak.
of the Spite plateau and we attain the MFV M = 2.205 and the In order to overcome the above shortcoming of the classic statis-
dihesion ε = 0.067, which illustrate certain advantages when deal- tical method, the MFV applies a new weight function
ing with problems with the non-Gaussian error distribution. These  2
n
results are very similar to the estimated value 2.21 ± 0.065 of Cran- 1
F (ε) = ε3 , (5)
dall et al. (2015), but might present a larger difference between the
i=1
ε + (xi − M)2
2

approaches of the two different principles along with the increase of


observations. This is expected as a result of asymptotically unimodal which controls well the accuracy of the measurements for reducing
distributions of the present observed data, as shown in top-left panel the effect of the dihesion ε from data fluctuation. Furthermore, the
of Fig. 2 . In this case, the MFV should be a close approximation to statistical efficiency of the MFV method largely exceeds that of the
the median value used in Crandall et al. (2015). Because the error LSM (Steiner 1991, 1997).
distribution of Li is just not normal (Crandall et al. 2015), the results In addition, it is very important that the descriptive statistics de-
of the MFV method are more credible than other statistical methods termine the symmetry of observational data and heavy-tailed or
based on the Gaussian distribution (Steiner 1991, 1997). light-tailed characteristics of distributions. This characterization of
There are several improvements over the LSM or other traditional measurements is generally represented by skewness and kurtosis.
statistical algorithms in astrophysical fields. For example, the me- The skewness–kurtosis graph proposed by Cullen & Frey (1999) is
dian statistical method only requires independent observations and shown in Fig. 3, which helps us to understand the relations between
is free of systematic errors, without having to consider the Gaussian the observed true distribution and common various distributions
feature (Gott et al. 2001). However, the MFV is different from the (Gaussian, uniform, logistic, exponential). Meanwhile, we also ap-
median, although some features are similar. In general, the median ply the non-parametric bootstrap procedure to assess the uncertainty
is robust to parameters and not sensitive to outliers or mistakes, of the evaluated results of skewness and kurtosis originating from
but it is not representative enough relative to the MFV procedure the measurements (Efron & Tibshirani 1994). The estimated skew-
(Steiner 1991, 1997). For instance, when the sample space of the ness and kurtosis are −0.146 and 2.554, respectively. From Figs 2
measurements is not large, there is a relatively great difference be- and 3, we can find the difference between the distributions of sim-
tween the median and the average value. Moreover, because of the ulations and the observed distribution of Li abundances.
lack of stability, when the sample space of the measurements is To sum up, accurately determining the A(Li) value in the Spite
large, the median is often affected by the middle term of the se- plateau serves a key role in many astrophysical fields, especially nu-
quence of observations. Besides, the MFV is not the mode of the clear astrophysics and cosmology. Because there is a non-Gaussian
generic statistics. Generally, the mode is the number that occurs error distribution of A(Li), we adopt the MFV procedure according

MNRAS 468, 5014–5019 (2017)


5018 J. Zhang

Observation Theoretical distributions


bootstrapped values normal
uniform
exponential
logistic
beta
lognormal
gamma
(Weibull is close to gamma and lognormal)
kurtosis

10

square of skewness
Figure 3. Cullen and Frey graph. [A colour version of this figure is available in the online version.]

to the minimization of the information loss, rather than the weighted Bonifacio P. et al., 2007, A&A, 462, 851
mean method and the median statistics. The results of the MFV Casey A. R. et al., 2016, MNRAS, 461, 3336
method are more robust and efficient, but still – as the calculated Chen G., Ratra B., 2003, PASP, 115, 1143
results present – almost as good as previous classic works (Crandall Chen G., Ratra B., 2011, PASP, 123, 1127
Chen G., Gott J. R. III, Ratra B., 2003, PASP, 115, 1269
et al. 2015). Obviously, this algorithm has enormous potential for
Chen Y., Kumar S., Ratra B., 2017, ApJ, 835, 86
observed astrophysical fields, such as the observed data of Type Ia
Crandall S., Ratra B., 2014, PhLB, 732, 330
supernova (Han & Podsiadlowski 2004; Wang & Han 2012; Riess Crandall S., Ratra B., 2015, ApJ, 815, 87
et al. 1998; Perlmutter et al. 1999). Crandall S., Houston S., Ratra B., 2015, Mod. Phys. Lett. A, 30, 1550123
Cullen A., Frey H., 1999, Probabilistic Techniques in Exposure Assessment.
Plenum, New York
AC K N OW L E D G E M E N T S Cyburt R. H., Fields B. D., Olive K. A., Yeh T.-H., 2016, Rev. Mod. Phys.,
88, 015004
I would like to thank the anonymous referee for useful comments
Efron B., Tibshirani R., 1994, An Introduction to the Bootstrap. Chapman
and suggestions. I acknowledge B. Zhang and Z-W. Han for their
and Hall, London
kind help with this work, as described in the paper. I also ac- Fields B. D., 2011, Annual Review of Nuclear and Particle Science, 61, 47
knowledge support from the National Natural Science Foundation Fu X., Bressan A., Molaro P., Marigo P., 2015, MNRAS, 452, 3256
of China (Grant Nos 11547041, 11403007, 11673007, 11643007, Gott J. R. III, Vogeley M. S., Podariu S., Ratra B., 2001, ApJ, 549, 1
11333004, U1531130, 11673059, 11390374 and 11521303) and Grevesse N., Asplund M., Sauval A. J., 2007, Space Sci. Rev., 130,
the Chinese Academy of Sciences (Nos KJZD-EW-M06-01 and 105
QYZDB-SSW-SYS00). Guiglion G., de Laverny P., Recio-Blanco A., Worley C. C., De Pascale M.,
Masseron T., Prantzos N., Mikolaitis Š., 2016, A&A, 595, A18
Gustavino C., Anders M., Bemmerer D., Elekes Z., Trezzi D., 2016, Euro-
REFERENCES pean Phys. J. A, 52, 74
Han Z., Podsiadlowski P., 2004, MNRAS, 350, 1301
Abia C., Pavlenko Y., de Laverny P., 1999, A&A, 351, 273 Hartmann D., Myers J., Woosley S., Hoffman R., Haxton W., 1999, in
Anders M. et al., 2014, Phys. Rev. Lett., 113, 042501 Ramaty R., Vangioni-Flam E., Cassé M., Olive K., eds, ASP Conf.
Arnould M., Norgaard H., 1975, A&A, 42, 55 Ser. Vol. 171, LiBeB Cosmic Rays, and Related X- and Gamma-Rays.
Avelino P. P., Martins C. J. A. P., Pinto P., 2002, ApJ, 575, 989 Astron. Soc. Pac., San Francisco, p. 235
Bailey D. C., 2017, Royal Society Open Science, 4, 160600 preprint Huber P., 1981, Robust Statistics. Wiley, New York
(arXiv:1612.00778) Kemp A. W., 2006, Steiner’s Most Frequent Value. Encyclopedia of Statis-
Bertulani C. A., Kajino T., 2016, Progress in Particle and Nuclear Physics, tical Sciences, Vol. 12. Wiley, New York
89, 56 Kirby E. N., Fu X., Guhathakurta P., Deng L., 2012, ApJ, 752, L16

MNRAS 468, 5014–5019 (2017)


MFV statistics and distribution of 7 Li 5019
Korn A. J., Grundahl F., Richard O., Barklem P. S., Mashonkina L., Collet Steiner F., ed., 1997, Optimum Methods in Statistics. Akademia Kiado,
R., Piskunov N., Gustafsson B., 2006, Nature, 442, 657 Budapest
Lodders K., Palme H., 2009, Meteoritics and Planetary Science Supplement, Steiner F., Hajagos B., 2001, Acta Geod. Geophys. Hungarica, 36, 327
72, 5154 Szegedi H., 2013, Geosciences and Engineering, 2, 102
Mukhamedzhanov A. M., Shubhchintak B. C. A., 2016, Phys. Rev. C, 93, Szegedi H., Dobroka M., 2014, Acta Geod. Geophys. Hungarica, 49, 95
045805 Szucs P., Civan F., Virag M., 2006, Hydrogeology Journal, 14, 31
Perlmutter S. et al., 1999, ApJ, 517, 565 Tajitsu A., Sadakane K., Naito H., Arai A., Aoki W., 2015, Nature, 518, 381
Podariu S., Souradeep T., Gott J. R. III, Ratra B., Vogeley M. S., 2001, ApJ, Walck C., ed., 2007, Handbook on Statistical Distributions for Experimen-
559, 9 talist, International Report SUF-PFY/96-01. University of Stockholm,
Reeves H., 1970, Nature, 226, 727 Stockholm
Riess A. G. et al., 1998, AJ, 116, 1009 Wall J. V., Jenkins C. R., 2012, Practical Statistics for Astronomers. Cam-
Sackmann I.-J., Boothroyd A. I., 1999, ApJ, 510, 217 bridge Univ. Press, Cambridge
Sneden C., Cowan J. J., Gallino R., 2008, ARA&A, 46, 241 Wang B., Han Z., 2012, New Astron. Rev., 56, 122
Spergel D. N. et al., 2003, ApJS, 148, 175 Zhang J., Cui W., Zhang B., 2010a, MNRAS, 409, 1068
Spergel D. N. et al., 2007, ApJS, 170, 377 Zhang J., Cui W., Zhang B., 2010b, MNRAS, 402, 956
Spite F., Spite M., 1982a, A&A, 115, 357
Spite M., Spite F., 1982b, Nature, 297, 483
Spite M., Spite F., Bonifacio P., 2012, Mem. Soc. Astron. Ital. Suppl., 22, 9
Steiner F., 1988, Geophys. Trans., 34, 226
Steiner F., ed., 1991, The Most Frequent Value: Introduction to Modern
Conception Statistics. Akademia Kiado, Budapest This paper has been typeset from a TEX/LATEX file prepared by the author.

MNRAS 468, 5014–5019 (2017)

You might also like