A Generalized Power Law Model of Citations

Ruheyan Nuermaimaiti1, Leonid V. Bogachev2 and Jochen Voss3

1
mmrn@leeds.ac.uk 2 L.V.Bogachevz@leeds.ac.uk 3 J.Voss@leeds.ac.uk
School of Mathematics, University of Leeds, Leeds LS2 9JT (United Kingdom)

Abstract
The classical power law model is widely used in informetrics to describe citations of scientific papers, although it
is not addressing variability across individual authors. We report our preliminary results for a novel model based
on a certain parametric form of the expected individual citation profile which generalizes the power law frequency
formula. The new model interpolates between large citation numbers, where the power law tail is reproduced, and
low citation numbers, which are usually truncated when fitting the power law model to the data. In addition, we
derive a deterministic limit shape of the citation profile, which can be used to make predictions about various
citation function such as the h-index.

Introduction
The classical power-law model was introduced by Lotka (1926) as an empirical match with
observed frequencies of citations in scientific publications. In a later development, Price (1965)
discovered an important connection with networks, whereby citations were interpreted as
nodes’ degrees. Examples of fitting the power law to the citation data can be found in Coile
(1977), Redner (1998), and Clauset, Shalizi & Newman (2009). In particular, it was found that
the power law frequencies do not necessarily fit well in the entire citation spectrum, so that a
suitable truncation of lower citation values may be needed.
Importantly, no assumptions are made in the power law model about the frequency distribution
of citations for an individual author randomly chosen from the population of authors. This
makes it difficult to project the model fitted to a pooled corpus of publications onto individual
authors, for example, for the purposes of evaluating their productivity.
The power-law model can be fitted to real-life data using standard statistical methods such as
the maximum likelihood or ordinary least squares estimation. As has been documented across
many use cases (Clauset, Shalizi & Newman, 2009), the power law usually fits quite well but
only in the tail region of the frequency range, which motivates the use of truncated power-law
models by excluding the lower values. This may decrease the utility of the model in estimation
of various functions of citations, such as the popular -index, introduced by Hirsch (2005) and
defined as the maximum number of an author's papers, each one cited at least times.
In an attempt to overcome this shortcoming, we propose a novel model by modifying the power
law setting. The new model interpolates between slow (almost flat) decay of the citation
frequencies at the bottom of the citation spectrum and then reproducing the power-law behavior
at the tail of the frequency distribution. As we will demonstrate below using a small real data
set, the model provides a very good fit across the entire citation spectrum. In addition, and in
contrast to the scale-free power law, our model possesses a deterministic limit shape of the
citation profile, which can be used, for example, to make meaningful estimation of the -index.
In particular, the estimation of the -index based on the modified model appears to be
significantly more accurate as compared to that in the standard power law model.

Power law frequencies

In its classical setting, the power law model states that the relative frequency of exactly
citations accumulated by a randomly sampled paper is proportional to the -th power of , with
some exponent (typically lying in the range ), that is (to include the case
and to normalize the sum of frequencies to unity),

843
where the normalizing constant is given by the reciprocal Riemann zeta function,

Here, “randomly sampled” means that a paper is sampled from a pooled corpus of papers
written by a (large) population of authors. For simplicity, interaction effects due to joint
authorship are not taken into account; such effects are complicated but have minor impact.
In terms of analysis of citation data, if there are papers with citations, out of the total
number of papers , then the power law model predicts that the relative frequencies are
approximately given by formula (1),

Note that the total number of papers and the total number of citations in this corpus are given,
respectively, by

Of course, in any real-life data set the numbers will reduce to zero for big enough (so that
the series in (4) are in fact finite sums), but this is reconciled with the prediction (1) simply by
the fact that the theoretical frequencies tend to zero as
As mentioned in the Introduction, the power law model does not address the frequency
distribution of citations for an individual author randomly chosen from the observed population
of authors (say, of size ) and featured by a collection of citation counts , that is, the numbers
of papers by this author that have citations ( ). Moreover, the number of observed
authors, , is often omitted in popular citation data sets (cf. Redner, 1998).
Having in mind statistically homogeneous populations of authors who produce their research
outputs according to the same probability distribution, it is natural to assume that these authors
are independent from one another due to lack of interaction. Equally reasonable is the
assumption of mutual independence of the counts for each individual author .
In this notation, the pooled numbers are given by

Then, according to the law of large numbers, for we have

where stands for expectation (statistical mean) and a random variable represents the counts
. Recalling (4), the mean number of papers per author is approximated as

provided that the series on the right is convergent. Thus, combining formulas (3), (6) and (7),
we obtain the link between the power-law frequencies and the expected counts ,

844
In practice, power law is often fitted in the tail of the frequency distribution, that is, for ,
with a suitably chosen truncation point . This leads to readjustment of the normalizing
constant in the frequency formula (1). Fitting such a model to the data requires optimization
over two parameters, and . Specifically, a natural heuristic tool to fit a truncated power-law
model is by looking at the frequency plots (e.g., histograms) with logarithmic scales on both
axes, whereby one seeks a straight-line fit, with the slope corresponding to (cf. Nicholls,
1987). An alternative approach (Clauset, Shalizi & Newman, 2009), which provides the helpful
smoothing of the discrete data, is via the complementary cumulative frequencies

Using again the log-log plots, a good fit corresponds to a straight line, with slope .

Generalized power-law model

Setting of the model
The generalized power-law (GPL) model introduced in this section is set out using two hyper-
parameters and , interpreted as the mean numbers per author of citations and papers,
respectively. These parameters can be estimated from the observed pooled corpus by

There are also two shape parameters, (akin to the power-law exponent) and .
Namely, the citation frequencies are now assumed to be of the form (cf. (1))

For small ( ) we have , while for larger we get a power-law dependence,

This may be viewed as an effective sewing of the formerly truncated lower values with the
power-law tail. Recalling the link (8) between the pooled frequencies and the individual
expected counts , we have

Hence, we can calibrate the model using (10) to make it consistent with the hyper-parameters,

and similarly

Considering the ratio (i.e., the mean number of citations per paper), we get

Hence, is expressed in terms of the hyper-parameters and the parameter ,

845
The shape parameter can be fitted to the data using either ordinary least squares or a suitable
version of the maximum likelihood estimation (cf. Nicholls, 1987).

Limit shape
It is useful to represent the citation profile of an individual author by ranking their papers
according to the citation scores (i.e., accumulated numbers of citations) ;
for example, is the score of the most cited paper. The citation profile is succinctly
visualized by the Young diagram formed by (left- and bottom-aligned) row blocks with
unit square cells; its upper boundary is the graph of the step function

where are the author’s citation counts. In particular, is their total number of papers.
A useful insight into the structure of the citations data may be available by looking at the shape
of suitably rescaled Young diagrams when both and are large (Vershik, 1996).
Specifically, set

and consider the rescaled (expected) shape

Substituting expressions (13) and approximating the sum by an integral like in (14) and (15), it
is easy to show that there is a limit shape given by

Data analysis

Goodness-of-fit
In this section, we fit the two models discussed above to real citation data, collected in January
2020 for a small population of authors identified as those who had published at least one
paper in the Electronic Journal of Probability in the first 10 issues (January–October) of volume
24 in 2019 (https://projecteuclid.org/euclid.ejp/1546571125) and who are also featured on
Google Scholar (https://scholar.google.com). The total counts for this data set are
(authors), (citations) and (papers), including papers
with zero citations. Noting that the observed frequency is quite high, it was
decided to omit the value at in the GPL model fitting (as seen in Figure 2, this value does
appear to be an outlier). According to (10), the hyper-parameters of the GPL are given by
and . In turn, the corresponding estimates for the shape
parameters are and , found numerically using the optim command in R.
The power-law model was fitted using a suitable truncation as explained after equation (8); the
fitted values, obtained using the poweRlaw package in R (https://cran.r-project.org/web/
packages/poweRlaw), are and . As we will see, the goodness-of-fit of the
power-law model is excellent, but a high value of is disappointing. Note that if we opted to
ignore truncation and tried to fit a power-law model in the entire range, the fitted value of the
exponent would change to .

846
First, let us report the results for the GPL model regarding the match of the theoretical limit
shape specified in equation (21). In Figure 1, this limit shape is plotted in scaled-
back coordinates for a better comparison with the data (represented by an empirical Young
diagram as explained above), that is, , with the scaling coefficients given by
formula (19) and estimated from the data as and .

Figure 1. Observed data in the Young diagram representation and the fitted limit shape, with
estimated parameters and . The left panel illustrates a match for lower values
of citations, while the right panel shows the tail comparison in the log-log coordinates.

We see from Figure 1 that the fit of the GPL is remarkably accurate, especially over a large
initial part of the citation spectrum. To inspect details of the tail behavior, we use the log-log
scale, revealing some minor discrepancies due to few extreme points.

Figure 2. Log-log plots comparing data and the fitted models. The left panel shows frequencies
, while the right panel features complementary cumulative frequencies . The left part of the
dashed lines indicates an extrapolation of the power law below the truncation point .
The observed frequency at (zero citations) is an outlier.

Next, we compare the fit of both models to the data using standard frequency plots (Figure 2).
Not surprisingly, the power law works extremely well at the tail but it is useless for smaller
values of . In contrast, the GPL strongly outperforms the power law over the initial range
(despite a visible outlier at ) but also works equally well at the tail.

Estimation of the -index

Let us now look at what our models can tell about the -index. According to Egghe & Rousseau
(2006), in the non-truncated power-law model the -index is estimated by the formula

847
where is the mean number of papers per author. In our case, ; using the
non-truncated estimate we get , while if we (formally) use the truncated fit
then the estimated value of would slightly change to .
In the GPL model, the -index geometrically corresponds to inscribing a biggest square inside
the empirical Young diagram of citations. Using the limit shape (21) and scaling back using the
coefficients and , we find that is the approximate solution of the equation

Solving this equation numerically yields . Comparing with the empirical (mean)
value , we see that the GPL estimation is superior to that of the power law, which
cannot capture the true -index lying deep below the truncation point .

Conclusion
In this paper, we have attempted to connect the pooled frequencies modeled via power law with
individual citations per author. We have also introduced the generalized power-law (GPL)
model which has been demonstrated to fit to the real data well in the entire range of citations,
unlike the power law which frequently needs to get truncated at the beginning. In a real data set
that we have studied, the GPL model reveals an inflated frequency of zero citations, which
draws attention to a possible lack of impact in scientific output. An additional novel feature of
the GPL model is that it possesses a deterministic limit shape which can be useful in estimating
informative function of the citation data such as the -index. Finally, it would be interesting to
compare our GPL model with alternative fitting approaches (Sichel, 1985).

Acknowledgments
Work by Ruheyan Nuermaimaiti is part of her PhD research at the School of Mathematics,
University of Leeds. The authors are grateful to Anna Senkevich for helpful insights concerning
data retrieval from the web and to an anonymous referee for suggested bibliographic references.

References
Clauset, A., Shalizi, C.R. & Newman, M.E.J. (2009). Power-law distributions in empirical data. SIAM
Review, 51, 661–703.
Coile, R.C. (1977). Lotka’s frequency distribution of scientific productivity. Journal of the American
Society for Information Science, 28, 366–370.
Egghe, L. & Rousseau, R. (2006). An informetric model for the Hirsch-index. Scientometrics, 69,
121–129.
Hirsch, J.E. (2005). An index to quantify an individual’s scientific research output. Proceedings of the
National Academy of Sciences of the United States of America, 102, 16569–16572.
Lotka, A.J. (1926). The frequency distribution of scientific productivity. Journal of the Washington
Academy of Sciences, 16, 317–323.
Nicholls, P.T. (1987). Estimation of Zipf parameters. Journal of the American Society for Information
Science, 38, 443–445.
Price, D.J.D.S. (1965). Networks of scientific papers. Science, 149, 510–515.
Redner, S. (1998). How popular is your paper? An empirical study of the citation distribution. The
European Physical Journal B – Condensed Matter and Complex Systems, 4, 131–134.
Sichel, H.C. (1985). A bibliometric distribution which really works. Journal of the American Society
for Information Science, 36, 314–321.
Vershik, A.M. (1996). Statistical mechanics of combinatorial partitions, and their limit shapes.
Functional Analysis and Its Applications, 30, 90–105.

848