Biblioranking Fundamental Physics (Updated To 2021/1/1) : CERN-TH-2018-066

CERN-TH-2018-066
Biblioranking fundamental physics

(updated to 2021/1/1)
Alessandro Strumiaa , Riccardo Torreb,c

arXiv:1803.10713v2 [cs.DL] 6 Apr 2021
a
Dipartimento di Fisica dell’Università di Pisa, Italy
b
CERN, Theory Division, Geneva, Switzerland
c
INFN, sezione di Genova, Italy
Abstract
Counting of number of papers, of citations and the h-index are the
simplest bibliometric indices of the impact of research. We discuss
some improvements. First, we replace citations with individual ci-
tations, fractionally shared among co-authors, to take into account
that different papers and different fields have largely different average
number of co-authors and of references. Next, we improve on citation
counting applying the PageRank algorithm to citations among papers.
Being time-ordered, this reduces to a weighted counting of citation de-
scendants that we call PaperRank. We compute a related AuthorRank
applying the PageRank algorithm to citations among authors. These
metrics quantify the impact of an author or paper taking into account
the impact of those authors that cite it. Finally, we show how self-
and circular- citations can be eliminated by defining a closed market of
Citation-coins. We apply these metrics to the InSpire database that
covers fundamental physics, presenting results for papers, authors,
journals, institutes, towns, countries, and continents, for all-time and
in recent time periods.
Contents
1 Introduction 2
2 Ranking papers 7
2.1 PaperRank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 PaperRank of papers: results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 PaperRank as the number of citations-of-citations . . . . . . . . . . . . . . . . . 10
2.4 Top-referred (recent) papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Paper metrics: correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Ranking authors 14
3.1 Sharing among co-authors: fractional counting . . . . . . . . . . . . . . . . . . . 16
3.2 Fractional counting and collaborations . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 PaperRank of authors: results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Author Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Removing self-citations and citation ‘cartels’: the Citation-coin . . . . . . . . . . 20
3.6 Author metrics: correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4 Rankings groups 24
4.1 Ranking institutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 Ranking towns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Ranking countries and continents . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.4 Ranking journals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5 Conclusions 28
A The InSpire and arXiv databases 34

A.1 Main trends in the fundamental physics literature . . . . . . . . . . . . . . . . . 34
A.2 Details about the dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1 Introduction
Bibliometrics can be a useful tool for evaluating research: it provides simple, quick, first ob-
jective measures of the impact of papers and authors and is increasingly being considered a
useful (although incomplete) evaluation criterion in postdoc/faculty recruitments, fundings,
and grant awards (Henneken and Kurtz, 2017, Kurtz, 2017, Kurtz and Henneken, 2017). In the
fundamental physics community that we consider in this paper, the most common measures
of the impact such as counting of number of papers, citations and Hirsch’s h-index (Hirsch,
2005), are inflating (Sinatra, Deville, Szell, Wang, and Barabási, 2015), making it harder to
identify the real impact of research, especially for the most recent literature. The more papers
one writes and the more citations these papers get, the bigger bibliometric estimators become:
it does not matter if these citations close in certain loops and/or remain confined in sub-fields,
2
or whether the paper has been written by a single author or in collaboration with thousands
of people.
We introduce new metrics and compare them with the existing ones, showing how they
address the issues mentioned above. We apply them to the InSpire 1 bibliographic database
(Holtkamp, Mele, Simko, and Smith, 2010, Ivanov and Raae, 2010, Klem and Iwaszkiewicz,
2011, Martin Montull, 2011),2 that covers fundamental physics literature after ≈ 1970.3 The
metrics, both the usual ones and the new ones that we introduce, can measure the impact of
papers, p, p0 , . . ., of authors, A, A0 , . . ., and of groups. They are defined as follows:
1. Number of papers
The most naive metric consists in counting the number of papers NApap = p∈A 1 written
P
by a given author A. This metric rewards the most prolific authors.
2. Number of citations
The most used metric consists in counting the number of citations Npcit received by a
paper p. An author A is then evaluated summing the number of citations NAcit received
by its papers. In formulæ:
X X
Npcit = 1, NAcit = Npcit , (1)
p0 →p p∈A
where the first sum runs over all papers p0 that cite p, and the second sum over all papers
p of author A.
3. h-index
The h-index is defined as the maximum h such that h papers have at least h citations
each. In formulæ, assuming that all papers of author A are sorted in decreasing order of
number of citations Npcit ≥ Np+1
cit
, it is given by
hA = max p | p ≤ Npcit .

(2)
p
This is proportional to NAcit , times a factor that penalises authors that write a small
number of highly cited papers (Hirsch, 2005).
1
https://inspirehep.net.
2
Other notable digital libraries and databases of research literature in various fields are, for instance,
ADS - The SAO/NASA Astrophysics Data System (http://www.adsabs.harvard.edu), CDS - CERN Docu-
ment Server (https://cds.cern.ch), arXiv.org (https://arxiv.org), Google Scholar (https://scholar.google.it), Mi-
crosoft Academic (https://academic.microsoft.com), DBLP - Computer Science Bibliography (https://dblp.uni-
trier.de), ACM Digital Library (https://dl.acm.org), PubMed (https://www.ncbi.nlm.nih.gov/pubmed), Math-
SciNet - Mathematical Reviews (https://mathscinet.ams.org), CiteSeerX (https://citeseerx.ist.psu.edu), Se-
mantic Scholar (https://www.semanticscholar.org), RePEc - Research Papers in Economics (http://repec.org),
IEEE Xplore Digital Library (https://ieeexplore.ieee.org), and zbMATH the first resource for mathematics
(https://zbmath.org).
3
The most relevant literature before ≈ 1970 has been added and is still being added on a request base to
InSpire.
3
As we will see, the average number of authors per paper and of references per paper increased,
in the last 20 years, by one to a few per-cent a year, and is significantly different in different
communities. Following basic common-sense, we propose an improved metric that renormalises
away such factors, and that cannot be artificially inflated adding more references and/or more
co-authors.
4. Number of individual citations

A citation from paper p0 to paper p is weighted as the inverse of the number of references
Npref
0 of paper p0 . Furthermore, the citations received by a paper p are equally shared
among its Npaut authors.4 In formulæ:
X 1 X Npicit
Npicit = , NAicit = . (3)
p0 →p
Npref
0
p∈A
Npaut
The definition of individual quantities is not new to the scientometrics literature and has
been extensively studied as an application of the so-called fractional counting. Fractional count-
ing has been considered both in the context of metrics and rankings (Aksnes, Schneider, and
Gunnarsson, 2012, Bouyssou and Marchant, 2016, Carbone, 2011, Egghe, 2008, Hooydonk,
1997, Leydesdorff and Bornmann, 2010, 2011, Leydesdorff and Opthof, 2010, Leydesdorff and
Shin, 2011, Rousseau, 2014), and in the context of constructing research networks (Leydesdorff
and Park, 2017, Perianes-Rodrı́guez, Waltman, and van Eck, 2016).
Furthermore, as well known in the literature, the division by Npref 0 factors out the different
publication intensity in the various fields, without the need of a field classification system.
Indeed, a paper has Npicit = 1 if it receives the mean number of citations in its field.
All the above metrics are defined “locally”, i.e. they can be computed for a paper/author
without knowing anything about all other papers but the ones that cite it. Therefore, they all
potentially suffer from the problem that even small sub-communities can inflate their indicators.
To overcome this problem one needs to define global measures, i.e. measures that know about
the whole community. The simplest such global measure of impact is given by the PageRank
algorithm, introduced in 1996 by Larry Page and Sergey Brin,5 the founders of Google (Brin,
Larry Page, and Winograd, 1999, Brin and Page, 1998).6 For a pedagogical introduction to
the PageRank see Rajaraman and Ullman (2009) and Tanase and Radu (2009). Applications
of the PageRank algorithm to citations network have already been considered, for instance, by
Chen, Xie, Maslov, and Redner (2007), Ding, Yan, Frazho, and Caverlee (2009), Ma, Guan, and
Zhao (2008), and Zhou, Orshanskiy, Zha, and Giles (2007). More advanced ranking algorithms,
based on integrated bibliographic information have also been proposed by Bini, del Corso, and
Romani (2010), Bini, del Corso, and Romani (2008), and del Corso and Romani (2009).
4
We assume that authors contributed equally because in fundamental physics authors are usually listed
alphabetically, with no information about who contributed more. The factor 1/Npaut , that is the one dictated
by conservation laws, will be further motivated in Section 3.2.
4
5. PaperRank
The PaperRank Rp of paper p and the PaperRank RA of author A are defined as
X Rp0 X Rp
Rp = ref
, RA = . (4)
0
p →p
Np 0
p∈A
Npaut
Namely, citations from papers p0 are weighted proportionally to their ranks Rp0 , that
get thereby determined trough a system of linear equations. The PaperRank provides a
metric which cannot be easily artificially inflated, because it is the bibliometric estimator
of a physical quantity: how many times each paper is read.
As we will see, the PaperRank singles out notable old papers which often do not have many ci-
tations. However, given that citations are time-ordered (newer papers cite older ones), the rank
reduces to a weighted sum over citation descendants (a combination of “citations of citations”)
which needs about 10-20 years to become a better indicator than the number of individual ci-
tations. In order to use information from the past, we define an alternative AuthorRank based
on citations among authors.
6. AuthorRank
We define the citation matrix which counts all individual citations from author A0 to A
X 1 1
NAicit
0 →A = , (5)
p →p
NpA NpA0 Npref
aut aut
A0
A0 A
where the sum runs over all papers pA0 of author A0 that cite papers pA of A. We then
define the AuthorRank RA as
X NAicit
0 →A
RA = R C
A0 A0 →A , C
A0 →A =P icit
, (6)
A0 A00 NA0 →A00
namely, as the principal eigenvector of the right stochastic matrix CA0 →A , which (thanks
to the normalization of each row provided by the sum in the denominator) tells the per-
centage of individual citations to A among all individual citations of A0 . The AuthorRank
gives more weight to citations coming from highly cited authors. We also use the Au-
thorRank of authors to define an improved ranking of papers, that we call AuthorRank
of papers, as
X X RA
Rp = . (7)
p0 →p A∈p0
Npaut ref
0 Np0
5
A similar idea, applied to bibliometrics, have been proposed long before the advent of the PageRank by
Pinski and Narin (1976).
6
Even if the word “Page” in PageRank may seem to refer to webpages, the name of the algorithm originates
from the name of one of its inventors, Larry Page.
5
Ideas similar to our AuthorRank have already been considered in the literature (Radicchi,
Fortunato, Markines, and Vespignani, 2009, West, Jensen, Dandrea, Gordon, and Bergstrom,
2013). Radicchi et al. (2009) applied an algorithm similar to our AuthorRank to the Physical
Review publication archive up to 2006. The implementation, called Science Author Rank
Algorithm (SARA) is different than ours in few respects: first, it is based on slices of the
full database, given by multiple graphs with equal number of citations, while our AuthorRank
includes information from a single graph constructed with weighted citations among all authors
after a given time (or for all times). Second, the SARA authors consider about half the number
of papers we consider and around one third of the citations we consider, all coming from the
single publisher Physical Review. While the analysis of Radicchi et al. (2009) is pioneering in
this direction, we believe that focusing on a single publisher introduces bias, since it does not
cover the whole scientific production of scientists: only a fraction of each author’s papers are
published on a Physical Review journal. We avoid this bias by considering the InSpire public
database including the scientific research in Fundamental Physics consisting of preprints and
articles published on about 2 · 103 journals. This should remove some of the bias induced by
considering a single publisher.
Third and most important difference between SARA and the AuthorRank is the contribution
of citations among papers to the weight of the links in the author-level graph: in the case of
SARA each paper contributes to the author-level graph with a weight 1/(m·n), with m and n the
number of authors in the citing and cited paper. In our case instead this weight is 1/NAicit
0 →A , with
this quantity defined in Eq. (5). As can be seen by this equation, this also contains a 1/Npref A0
.
In simpler words, the SARA author-level graph is determined using citations among papers, in
our notation Npcit , while our author-level graph is determined using individual citations among
papers, in our notation Npicit , as defined in Eq. (3). We find the latter more indicative, as we
find individual citations of papers, that are the building blocks of the paper-level graph, more
indicative than traditional citations. Being cited among many others is typically less relevant
than being cited with a few other references.
West et al. (2013) applied a similar algorithm to the Social Science Research Network
Community, calling it Author-Level Eigenfactor Metric. In this case some bias is introduced
by removing all self-citations from the graph, corresponding to the diagonal of the correlation
matrix NAicit
0 →A . As we discuss in Section 3.5, removing self-citations is an arbitrary procedure,
also admitting different implementations. The problem of making the effect of self citations less
relevant can be addressed in two different ways: by tuning the teleport probability parameter
in the PageRank algorithm, or by considering a more general quantity, like the Citation-coin
introduced below.
In this paper we stress that, while the AuthorRank is an interesting metric by itself, it
mainly identifies already well known authors, like Feynman and others (Radicchi et al., 2009),
and Jensen and others (West et al., 2013). We try to improve in extracting information from
the AuthorRank and use it to identify recent papers cited by highly ranked (with AuthorRank)
physicists. This allows us to define an AuthorRank for papers, that is a good candidate as an
early alert for potentially important papers.
As we stated before, removing self-citations carries some level of arbitrariness. One could
aim at resolving a better defined problem, consisting of removing all citation ‘cartels’, defined
6
as loops of citations among 3 authors, among 4 authors, etc. The mathematical problem of
removing all circular citations has a simple solution, inspired by economy: money.7 You don’t
get richer by giving money to yourself or by circulating money with friends. This leads us to
the definition of an additional metric:
7. Citation-coin
Author A ‘owes’ the number C //A of individual citations received minus the number of
individual citations given:8
X X 1
C
//A = (NAicit icit icit
0 →A − NA→A0 ) = NA − aut
. (8)
A0 p∈A
Np
This metric penalises authors who write many small papers which receive few citations
from others.
An important property of all above metrics is that they can be computed in practice. Our
metrics are intensive: so they can be used to rank groups, such as journals, institutes, countries,
etc. by simply summing over their members. Furthermore they can be restricted to any specific
time period, e.g. after year 2000.
The paper is structured as follows. In Section 2 we introduce the PaperRank R, discuss
its features and properties, and rank all papers in InSpire. In Section 3 we introduce the
number of individual citations Nicit , the AuthorRank R and the Citation-coin C //, discussing
their features and properties, and rank all authors in InSpire. In Section 4 we apply these
measures to rank groups: institutions, towns, countries, continents, and journals. Conclusions
are presented in Section 5. In Appendix A we describe the InSpire and arXiv databases and
their main features and trends, together with technical details.
Several of our results with complete tables are available at the PhysRank webpage9 .
2 Ranking papers
2.1 PaperRank
Given a citation network of Npap papers, the PaperRank Rp of each paper p is defined by
X Rp0
Rp = ℘ + α(1 − ℘) , (9)
p0 →p
Npref
0
where Npref
0 is the total number of references of each paper p0 that cites p. EquationEq. (9) is
linear, so its solution is unique, and can be computed efficiently iteratively (Rajaraman and
Ullman, 2009).
7
In mathematical language this consists in removing all closed loops from the authors’ graph, making it
acyclic.
8
A slight improvement will be added later, to avoid border effects due to intrinsic finite nature of the database.
9
http://rtorre.web.cern.ch/rtorre/PhysRank.
7
We elaborate on its meaning. Eq.Eq. (9) contains
P two arbitrary constants α and ℘. The
constant α just fixes the overall normalization p Rp = Rtot . When applied to internet, Rp
describes the probability that site p is visited and it is convenient
P cit to normalize it to one. We
choose Rtot equal to the total number of citations Rtot = p Np , in order to allow for an easier
comparison between the number of citations received by a paper and its PaperRank Rp . With
this normalisation Rp grows with time, as newer papers appear.
Viewing the rank as the probability that a paper is read, the parameter ℘ splits it into two
contributions: the first term is the probability that a reader reaches a paper by following a
reference to it; the second term, equal for all papers, simulates readers that randomly browse
the literature.
• In the limit ℘ = 0 the first contribution vanishes, and all papers have a common rank.
At first order in small ℘ 1, Rp starts discriminating the papers p:
Rtot 1 + ℘Npicit
℘1
Rp (℘) ' , (10)
Npap 1 + ℘
Npicit is the number of individual citations received by p defined in Eq. (3), which
where P
obeys p Npicit = Npap .
• In the limit ℘ = 1 the second contribution in Eq. (9) vanishes, and Rp only depends
on the structure of the network, provided that no closed sub-networks and dead-ends
exist (Rajaraman and Ullman, 2009). Recursive computations of Rp become slower as
℘ → 1.
Data about downloads of scientific articles would allow to extract the value of ℘ that better
fits the observed reading rate; however such data are not available in fundamental physics.10
We use a large ℘ = 0.99, such that the first contribution in Eq. (9) dominates for all relevant
authors.
2.2 PaperRank of papers: results

We compute the PaperRank by constructing a graph (and its transition matrix) having all
papers as nodes and citations as links. We consider the full InSpire database, as detailed in
Appendix A.2. Generally, a few hundred iterations of Eq. (9) are necessary for a percent level
convergence. The computation takes a few minutes on a laptop computer.
Table 1 (Table 2) shows the top-cited (top-ranked) papers in the InSpire database. Top-
ranked papers correspond to the papers with top PaperRank and tend to be old famous ones,
even with a relatively small number of citations. Top-cited papers, ranked with the usual
counting of the number of citations, tend to be modern, in the view of the inflation in the rate
of citations. The same effect was observed by Chen et al. (2007), who applied the PageRank
algorithm to the sub-set of papers published on Physical Review.
10
arXiv.org does not make public the number of downloads, to avoid the conversion of this information into
a relevant metric, and its consequent fate determined by Goodhart’s law.
8
Title 1st author Naut date Ncit Rp Rp
1 The Large N limit of superconformal fiel J.M.Maldacena 1 1998 16317 6725 22034
2 GEANT4–a simulation toolkit James.R.Allison 127 2003 13488 6790 2170
3 Measurements of Ω and Λ f S.Perlmutter 32 1999 12983 4124 7543
4 Observational evidence from supernovae f A.G.Riess 20 1998 12965 4267 7376
5 A Model of Leptons Steven.Weinberg 1 1968 12955 44908 41591
6 PYTHIA 6.4 Physics and Manual T.Sjostrand 3 2006 11823 3439 3798
7 Observation of a new particle in the sea ATLAS 2932 2012 11666 1652 4074
8 Observation of a New Boson at a Mass of CMS 2897 2012 11404 1638 4014
9 CP Violation in the Renormalizable Theor M.Kobayashi 2 1973 10695 10903 20359
10 Anti-de Sitter space and holography E.Witten 1 1998 10513 4471 13998
Table 1: Top-cited (highest number of citations) papers in the InSpire database.

2 Conservation of Isotopic Spin and Isotop C.N.Yang 2 1954 2852 41770 14876
3 Theory of Fermi interaction R.P.Feynman 2 1958 1748 39446 10778
4 Remarks on the Dirac theory of the posit W.Heisenberg 1 1934 122 38682 1223
5 On the Stopping of fast particles and on H.A.Bethe 2 1934 724 31585 1009
6 The S matrix in quantum electrodynamics F.J.Dyson 1 1949 771 31296 5044
7 Symmetries of baryons and mesons M.Gell.Mann 1 1962 1666 29947 8576
8 Field Theories with Superconductor Solut J.Goldstone 1 1961 2078 29244 5773
9 A Theory of the Fundamental Interactions J.S.Schwinger 1 1957 679 28331 11962
10 Space - time approach to quantum electro R.P.Feynman 1 1949 830 26963 3550
Table 2: Top-ranked (highest PaperRank) papers in the InSpire database.

2 Particle Creation by Black Holes S.W.Hawking 1 1974 8665 9142 28472
3 A Planar Diagram Theory for Strong Inter G.tHooft 1 1974 4989 8238 26994
4 Unity of All Elementary Particle Forces H.M.Georgi 2 1974 5090 10397 25424
5 Confinement of Quarks K.G.Wilson 1 1974 5376 23570 23192
6 Weak Interactions with Lepton-Hadron Sym S.L.Glashow 3 1970 6214 19199 22997
7 Pseudoparticle Solutions of the Yang-Mil A.A.Belavin 4 1975 2887 13352 22315
8 The Large N limit of superconformal fiel J.M.Maldacena 1 1998 16317 6725 22034
9 CP Violation in the Renormalizable Theor M.Kobayashi 2 1973 10695 10903 20359
10 Symmetry Breaking Through Bell-Jackiw An G.tHooft 1 1976 3769 8220 20081
Table 3: Top-referred (highest AuthorRank) papers in the InSpire database.
9
The difference between the two rankings is partly due to the fact that PaperRank penalises
recent papers. Papers tend to accumulate citations for about 10-20 years, while the rank
continues growing with time, and is highly suppressed for younger papers.
This also means that the PaperRank defined in Eq. (9) needs 10-20 years before providing a
better metrics than the number of citations. This is proven in the next section, where we show
that, for a time-ordered network (such as the network of citations), the PaperRank reduces to
the number of citations-of-citations.
2.3 PaperRank as the number of citations-of-citations

Internet allows for reciprocal links among pages, and the PageRank captures in a simple way
the self-interacting system. Citations among scientific papers are instead time-ordered, forming
an acyclic network. In the limit where citations of older papers to newer papers are ignored,11
no loops are possible within the network, and the implicit definition of the rank Rp of Eq. (9)
can be converted into the following explicit expression12
∞
X X 1 1
Rp ∝ ℘g ref
· · · ref
. (11)
g=0 pg →···→p
Np g
Np 1
Basically, Rp counts the number of citations-of-citations up to generation g. In the above

expression, the term with
- g = 0 contributes as unity, and accounts for the constant term in Eq. (9), which is
negligible for papers that receive citations from others;
- g = 1 contributes with the number of individual citations Npicit as in Eq. (10): the sum
runs over ‘first generation’ papers p1 that cite the paper p;
- g = 2 corresponds to ‘second generation’ papers p2 that cite the papers p1 that cite p;
- g = 3 corresponds to ‘third generation’ papers p3 that cite the papers p2 that cite the
papers p1 that cite p;
- for generic g the sum runs over all papers pg that cite paper p in g steps.
In other terms, for any given paper, we refer to papers that cite it as “first generation”, and
define as “second generation” those papers that cite at least a first generation paper, and so on.
A paper q can appear multiple times in different generations g, corresponding to all possible
citations paths from q to p. Eq.Eq. (11) shows that ℘ < 1 gives a cut-off on the number of
generations that one wants to consider, and that Rp (℘) grows with ℘, and with time.
11
We enforced time-ordering within the citations, deleting from the InSpire database a small number of
‘a-causal’ citations, where older papers cite newer papers (see Appendix A.2 for details). Since older papers
tend to accumulate large ranks, a-causal citations can artificially inflate the rank of a few recent papers.
12
This can be proven by substituting Eq. (11) into Eq. (9). A physicist can view in Eq. (11) a path-integral
within the network.
10
��
��
��
��

��
��
��
��
��
�
� � ��
��
��
Figure 1: Contributions of each generation to the citation chain of some notable papers.
• At one extremum, ℘ → 0, the rank Rp reduces to the “number of children” Npicit , without
checking if they are successful. Papers on hot topics can fast accumulate many citations,
even if later the hot topic becomes a dead topic. Too recent papers are penalised.
• At the other extremum, ℘ → 1, the rank Rp becomes the Adamo number: it counts
descendants. Seminal papers that open new successful fields are rated highly, and their
rank continues to grow. Too recent papers are highly penalised.
The PaperRank in Eq. (11) splits papers into two qualitatively different categories:
• Sub-critical papers that get some attention and get forgotten: this happens when, after
a long time (tens of years), the sum over g remains dominated by the first generation.
• Super-critical papers, that make history. If the citation rate is high enough, it can sustain
a ‘chain reaction’, such that late generations keep contributing significantly to the sum
over generations. At the same time, the original paper gets summarized in books and
ceases to be cited directly.
Figure 1 shows, for a few notable papers, how much different generations g contribute to
the sum in Eq. (11). We see that about 10 generations contribute significantly for old top-
ranked papers, while the 1st generation provides the dominant contribution to recent top-cited
papers.13
2.4 Top-referred (recent) papers

As explained in the previous section, the PaperRank can single out some notable papers with
few citations, provided that they are old. However, when applied to recent papers (less than 10-
13
We computed Rp (℘) analytically as function of ℘ for all papers p using Eq. (11) with the following ‘pruning’
algorithm. To start, one finds all papers with no citations and eliminates them from the database, after assigning
their contributions to the Rp of their references. The process is iterated. About 1000 iterations are needed to
prune all the InSpire citation tree to nothing, obtaining Rp (℘) as a power series in ℘.
11
Year Title 1st author Naut Ncit Rp Rp
1880 On the Relative Motion of the Earth and the A.A.Michelson 2 211 219 127
1890 Cathode rays J.J.Thomson 1 239 25 33
1900 On the electrodynamics of moving bodies Albert.Einstein 1 499 754 1272
1910 The Foundation of the General Theory of Rel Albert.Einstein 1 1104 1415 973
1910 Approximative Integration of the Field Equa Albert.Einstein 1 331 1552 6501
1910 Einstein’s theory of gravitation and its as de Sitter 1 219 3474 182
1920 About the Pauli exclusion principle E.P.Wigner 2 383 23416 2222
1920 Quantum Theory and Five-Dimensional Theory O.Klein 1 2481 3233 5741
1930 Quantised singularities in the electromagne P.A.M.Dirac 1 2269 20421 12586
1930 Remarks on the Dirac theory of the positron W.Heisenberg 1 122 38682 1223
1940 The Theory of magnetic poles P.A.M.Dirac 1 1071 6126 6821
1940 The S matrix in quantum electrodynamics F.J.Dyson 1 771 31296 5044
1940 Forms of Relativistic Dynamics P.A.M.Dirac 1 1828 1643 4095
1950 On gauge invariance and vacuum polarization J.S.Schwinger 1 5160 10263 11589
1950 Conservation of Isotopic Spin and Isotopic C.N.Yang 2 2852 41770 14876
1960 A Model of Leptons Steven.Weinberg 1 12955 44908 41591
1970 CP Violation in the Renormalizable Theory o M.Kobayashi 2 10695 10903 20359
1970 Confinement of Quarks K.G.Wilson 1 5376 23570 23192
1970 Particle Creation by Black Holes S.W.Hawking 1 8665 9142 28472
1980 The Inflationary Universe: A Possible Solut A.H.Guth 1 8171 9250 15347
1990 The Large N limit of superconformal field t J.M.Maldacena 1 16317 6725 22034
2000 GEANT4–a simulation toolkit GEANT 127 13488 6790 2170
2000 First year Wilkinson Microwave Anisotropy P WMAP 17 9040 2990 6507
2010 Observation of a new particle in the search ATLAS 2932 11666 1652 4074
Table 4: The top-cited, top-ranked, top-referred paper written in each decennium among those
listed in InSpire (which is highly incomplete before 1960).
20 years old), the PaperRank becomes highly correlated to the number of (individual) citations,
and therefore cannot perform better.
We thereby propose an early-alert indicator that recovers some information from the past.
First, we compute a rank among authors using all-time data. In particular, we adopt the
AuthorRank RA anticipated in Eq. (6) and better discussed in the next section. Next, we use
such rank to weight citations to papers, as in Eq. (7). This means that we give more weight
to citations from authors with higher RA , implementing a sort of representative democracy.
We dub papers with top AuthorRank of papers Rp as ‘top-referred papers’. Table 3 shows the
all-time list.
Table 4 shows the top-cited, top-ranked and top-referred papers published within each
decennium, based on all subsequent citations. For recent papers, top-cited and top-ranked
tend to be dominated by manuals of useful computer codes and by reviews.
We finally use Rp to find top-referred recent papers: Table 5 shows the top-referred papers
published after year 2010 and with less than 10 authors (because our goal is to identify no-
table recent papers less known than discoveries made by big experimental collaborations) and
more than 100 citations. Furthermore, we here removed self-citations, to avoid the list to be
12
1 Black Holes: Complementarity or Firewall A.Almheiri 4 2012 1129 230 4712
2 New Symmetries of QED D.Kapec 3 2015 137 36 1725
3 Conformal symmetry and its breaking in t J.M.Maldacena 3 2016 504 84 1092
4 On the Origin of Gravity and the Laws of E.P.Verlinde 1 2010 880 196 1082
5 Remarks on the Sachdev-Ye-Kitaev model J.M.Maldacena 2 2016 860 173 1059
6 Topological insulators and superconducto X.L.Qi 2 2011 1592 534 1026
7 Black holes and the butterfly effect S.H.Shenker 2 2013 733 198 939
8 Topological Insulators M.Zahid.Hasan 2 2010 1892 788 888
9 Shapiro Delay Measurement of A Two Solar P.Demorest 5 2011 2411 396 846
10 Investigating the near-criticality of th D.Buttazzo 7 2014 1086 127 845
Table 5: Top-referred papers written after 2010 with less than 10 authors and more than 100
citations.
� ��
� � ��
� �� ℛ� �� ℛ� �
// �
� ��
� ��
� ��
� ��
��
ℛ� ��
��
��
ℛ� ��
�
// �
Figure 2: Left(right) table: correlations between indices of papers(authors). For papers the cor-
relation is considered separately for the whole InSpire and for the eight main arXiv categories.
dominated by notable authors citing themselves.
2.5 Paper metrics: correlations

The left panel of Fig. 2 shows the correlations among the traditional counting of citations Npcit ,
the number of individual citations Npicit , the PaperRank Rp and the AuthorRank of papers
Rp , in the whole InSpire database. The number of individual citations is highly correlated
with the number of citations. Indeed, for papers, individual citations are just citations divided
by the number of references of the citing papers so that uncorrelation is proportional to the
variance around the average of the number of references per paper. The PaperRank and the
AuthorRank are less correlated to the number of (individual) citations and to each other and
represent fairly independent indices for ranking papers.
13
X X
Number of papers 1 Number of citations Ncit
papers papers
InSpires name All InSpires InSpires name All InSpires
1 G.Eigen.1 2508 1 R.V.Kowalewski.1 261176
2 S.L.Wu.1 2396 2 G.D.Cowan.1 247656
3 Kazuhiko.Hara.1 2359 3 Otmar.Biebel.1 236405
4 J.Brau.2 2310 4 J.Huston.1 241391
5 A.Seiden.1 2293 5 Christoph.Grab.1 225873
6 R.Kass.1 2276 6 Achim.Stahl.1 220954
7 David.M.Strom.1 2270 7 K.Moenig.1 222876
8 A.Bodek.2 2268 8 S.L.Wu.1 225586
9 W.T.Ford.1 2263 9 S.M.Spanier.1 215374
10 R.V.Kowalewski.1 2231 10 A.V.Gritsan.1 217487
X
h index 1 Average citations hNcit i
Ncit ≥h
InSpire name All InSpires InSpire name All InSpire
1 Kazuhiko.Hara.1 203 1 Y.Oohata.1 13488
2 J.Huston.1 200 2 N.Eiden.1 et al. 7673
3 H.H.Williams.1 198 3 J.L.Chuma.1 et al. 6744
4 S.L.Wu.1 197 4 S.Chowdhury.1 et al. 6389
5 A.G.Clark.1 196 5 P.Schaffner.1 6104
6 G.Eigen.1 196 6 S.B.Lugovsky.1 6043
7 P.K.Sinervo.1 195 7 K.S.Lugovsky.1 5844
8 M.J.Shochet.1 195 8 C.Roumenin.1 et al. 5746
9 J.Proudfoot.1 195 9 H.Yusupov.2 5732
10 S.M.Errede.1 195 10 V.S.Lugovsky.1 5195
Table 6: Authors listed according to traditional biblio-metric indices: total number of papers
(top left), of citations (top right), h-index (bottom left), average number of citations per paper
(bottom right).
3 Ranking authors
We start from the simplest and most naive metrics: in the top-left column of Table 6 we list
the authors with most papers. Within the InSpire database, they are all experimentalists that
participate in large collaborations with many co-authors. The extreme case are the ATLAS
and CMS collaborations with ∼ 103 papers and ∼ 103 authors.
In the top-right column of Table 6 we show the top-cited authors: again they are experi-
mentalists that participate in large collaborations. The citations of author A are counted in
the usual way: summing the citations received by all papers that include A as author, as in
Eq. (1). The bottom-left column of Table 6 shows the authors with highest h index, and the
bottom-right column the authors with the highest average number of citations per paper.
Next to each author we add symbols which show if they received the Nobel ( ), Dirac (
from ICTP and from IOP), Planck ( ), Sakurai (•), Wolf ( ) and Milner ( ) prizes. Small
inaccuracies are possible, as medalists have been identified from names. None of the top authors
14
InSpire name All InSpire name After 2000 InSpire name After 2010
1 E.Witten.1 3703 J.M.Maldacena.1 306.1 S.D.Odintsov.1 82.9
2 Steven.Weinberg.1 2487 S.D.Odintsov.1 276.1 J.M.Maldacena.1 82.3
3 G.tHooft.1 1592 E.Witten.1 272.0 N.Kidonakis.1 77.1
4 S.W.Hawking.1 1382 T.Padmanabhan.1 254.7 E.Witten.1 71.8
5 A.M.Polyakov.1 986.2 P.Z.Skands.1 254.4 U.G.Meissner.1 63.3
6 F.A.Wilczek.1 • 902.6 T.Sjostrand.1 • 252.2 A.Strominger.1 62.2
7 J.M.Maldacena.1 900.8 S.Nojiri.1 246.6 C.de.Rham.1 62.1
8 R.W.Jackiw.1 888.3 S.Mrenna.1 219.2 S.Tsujikawa.1 60.4
9 J.S.Schwinger.1 854.4 G.P.Salam.1 217.1 S.Nojiri.1 59.5
10 A.D.Linde.1 833.4 V.Springel.1 216.5 S.Capozziello.1 57.2
11 T.Sjostrand.1 • 829.7 D.T.Son.1 212.5 D.Stanford.1 57.2
12 L.Susskind.1 • 799.6 Ashoke.Sen.1 208.6 P.Z.Skands.1 57.1
13 S.L.Glashow.1 788.0 C.Vafa.1 186.7 L.Susskind.1 • 55.3
14 H.M.Georgi.1 • 775.0 M.Cacciari.1 180.7 S.Sachdev.1 50.0
15 N.Seiberg.1 719.0 U.G.Meissner.1 180.6 M.Czakon.1 49.4
16 P.A.M.Dirac.1 704.1 P.Nason.1 176.7 G.P.Salam.1 48.2
17 Sidney.R.Coleman.1 675.3 Ernest.Ma.1 170.2 R.Venugopalan.1 47.7
18 David.J.Gross.1 • 665.9 D.E.Kharzeev.1 160.3 M.Luscher.1 47.7
19 C.Vafa.1 643.0 S.Tsujikawa.1 158.9 R.E.Kallosh.1 47.1
20 S.J.Brodsky.1 • 626.4 A.Strominger.1 156.3 D.W.Hooper.1 46.4
21 J.R.Ellis.1 617.5 Martin.Bojowald.1 152.6 R.B.Mann.1 45.6
22 K.G.Wilson.1 613.5 V.A.Kostelecky.1 150.7 S.Hod.1 45.4
23 Abdus.Salam.1 607.3 S.S.Gubser.1 148.4 F.Maltoni.1 45.4
24 M.Luscher.1 589.9 S.Capozziello.1 145.4 R.C.Myers.1 45.2
25 J.D.Bjorken.1 586.6 J.Polchinski.1 144.7 A.D.Linde.1 44.9
26 R.L.Jaffe.1 579.0 Alan.D.Martin.1 144.3 A.De.Felice.1 43.2
27 A.Strominger.1 569.5 G.Amelino.Camelia.1 143.7 X.L.Qi.1 43.1
28 Ashoke.Sen.1 567.5 A.Loeb.1 143.1 S.F.King.1 42.9
29 G.Veneziano.1 563.2 D.W.Hooper.1 141.1 C.Bambi.1 42.5
30 C.N.Yang.1 548.7 A.Ashtekar.1 141.1 B.Schenke.1 42.3
31 J.Polchinski.1 546.2 F.Aharonian.1 139.6 T.Padmanabhan.1 42.1
32 Rabindra.N.Mohapatra.1 540.2 N.Arkani.Hamed.1 138.3 K.Hinterbichler.1 41.9
33 S.Deser.1 526.5 Wayne.Hu.1 137.1 H.T.Janka.1 41.1
34 Alexander.Vilenkin.1 525.5 A.C.Fabian.1 136.8 E.N.Saridakis.1 40.8
35 R.P.Feynman.1 524.9 Rong.Gen.Cai.1 136.5 P.Nason.1 40.2
36 L.Wolfenstein.1 • 524.6 G.R.Dvali.1 136.3 A.Mitov.1 40.1
37 Stephen.Louis.Adler.1 • 520.4 L.Susskind.1 • 136.1 Bing.Zhang.1 39.9
38 M.Gell.Mann.1 519.4 E.V.Shuryak.1 135.8 U.W.Heinz.1 39.3
39 B.Zumino.1 518.6 U.W.Heinz.1 134.5 Florian.R.A.Staub.1 39.3
40 John.H.Schwarz.1 518.4 F.Karsch.1 131.7 A.Strumia.1 39.2
41 M.A.Shifman.1 • 505.9 N.Kidonakis.1 131.5 J.Rojo.1 39.1
42 E.V.Shuryak.1 493.8 G.Soyez.1 131.1 T.Schwetz.1 39.0
43 M.B.Wise.1 • 485.9 A.D.Linde.1 128.9 A.J.Buras.1 39.0
44 G.W.Gibbons.1 484.8 A.A.Tseytlin.1 128.3 A.Loeb.1 38.8
45 J.D.Bekenstein.1 484.1 A.Strumia.1 124.7 G.Soyez.1 38.7
46 A.A.Tseytlin.1 483.9 M.Visser.1 123.9 E.Oset.1 38.1
47 L.N.Lipatov.1 474.9 S.D.M.White.1 122.6 V.Cardoso.1 37.8
48 H.Leutwyler.1 471.9 Nathan.J.Berkovits.1 122.2 E.P.Verlinde.1 37.2
49 T.D.Lee.1 459.0 S.Frixione.1 122.1 S.S.Ostapchenko.1 36.9
50 N.Isgur.1 • 456.9 J.R.Ellis.1 120.1 P.Bozek.1 36.8
Table 7: Authors sorted according to their number of individual citations NAicit for the whole
InSpire database (left), from year 2000 (middle), and from year 2010 (right).
15
according to traditional metrics received any of them.
Clearly, all the indices shown in Table 6 ceased to be relevant for experimentalists in view
of the large number of co-authors. This shows the need for an improved metrics that cor-
rects for the inflation in the number of co-authors and allows at least a naive comparison of
experimentalists with the rest of the community.
3.1 Sharing among co-authors: fractional counting

An improved metrics is obtained Pby attributing a fraction pA of any given paper to each author
A, and imposing the sum rule pA = 1. The fractions pA should tell how much each author
contributed to the paper. In the absence of this information, we assume that each co-author
contributed equally, so that pA = 1/Npaut .14 This is called ‘fractional counting’ in the bibliometric
literature.
Taking into account this factor, the total number of ‘individual papers’ of author A is given
by p∈A 1/Npaut . The same sharing among co-authors is applied to citations. The number of
P
‘individual citations’ received by author A is defined by summing over all its papers p taking
into account that citations are shared among co-authors, and weighted inversely to the number
of references:
X Npicit X 1 X 1
NAicit ≡ = .
p∈A
Npaut p∈A
(Number of authors of paper p) p0 →p (Number of references of p0 )
(12)
In the same way, we share the rank Rp of each paper equally among its authors. The rank
Rp of a paper approximates a physical quantity: how many times the paper is read. The rank
RA of an author inherits the same meaning: it tells the visibility of any author A, obtained by
summing the visibility of its papers p as in Eq. (4), i.e.15
X (PaperRank of paper p)
RA = (PaperRank of author A) = . (13)
p∈A
(Number of authors of paper p)
As discussed in Section 2.1 we consider ℘ = 0.99.
3.2 Fractional counting and collaborations

We share the number of citations received by a paper among its Naut authors as 1/Naut . Some
agencies adopt various less steep functions of Naut . This raises the question: how does the
total number of citations received scale, on average, with the number of co-authors?16 Figure 3
(described in more details in the caption) provides the answer: the total number of citations
14
Weighting authors proportionally to their AuthorRank gives a non-linear system of equations, with singular
solutions where a few notable authors collect all the weight of large collaborations. Restricting to single-author
papers would uniquely identify authors’ contributions, but at the price of discarding most literature.
15
For different methods see Ding et al. (2009) and Zhou et al. (2007).
16
We thank Paolo Rossi for raising this issue.
16
��
��

��
��
��
��-��
��-��
��
��-��
��-��
��
��-�� -�
��-��
��
��-��
��-��
� ��-�
� ��
�� -�� -��
Figure 3: Left: for each author in InSpire we compute its mean number of co-authors (per
paper, including the author itself ) and the total number of citations of all its papers. We do
not show the scatter plot, only its median, separately for each arXiv category. In both cases the
total number of citations grows roughly linearly with the number of authors. Right: As in the
middle panel, but showing on the vertical axis the number of individual citations, which roughly
does not depend on the number of authors.
received by an author grows roughly linearly with the mean number of co-authors. The right
panel of Fig. 3 shows that, instead, the total number of individual citations is scale-invariant,
namely roughy independent of the number of co-authors. This means that individual citations
do not reward nor penalise big collaborations, while citations reward big collaborations.
3.3 PaperRank of authors: results

We now apply the improved metrics described in the previous section to the InSpire database.
The left column of Table 7 lists the authors with the highest number of individual citations.
Two factors differentiate citations from individual citations. First, dividing by the number of
references (first factor in the denominator in Eq. (12)) counter-acts the inflation in the total
number of citations. This factors mildly penalises authors working in sectors (such as hep-ph)
where papers have a larger average number of references. Second, dividing by the number
of authors (second factor at the denominator in Eq. (12)) has a large impact: members of
huge collaborations no longer make the top positions of the list, which becomes dominated by
theorists. As discussed in Section 3.2, this happens because working in large collaborations does
not allow to recognise individual merit — not because working in large collaborations decreases
the average merit. Lists of bottom authors would similarly be dominated by theorists.
The left column of Table 8 shows the top-ranked authors in the InSpire database. The
PaperRank of authors identifies some older notable authors who received the prizes plotted in
17
InSpire name All InSpire InSpire name After 2000 InSpire name After 2010
1 Steven.Weinberg.1 235825 E.Witten.1 3670 N.Kidonakis.1 822.3
2 J.S.Schwinger.1 231194 J.M.Maldacena.1 3551 E.Witten.1 794.2
3 R.P.Feynman.1 162394 S.D.Odintsov.1 3027 J.M.Maldacena.1 786.5
4 M.Gell.Mann.1 156387 T.Sjostrand.1 • 3016 U.G.Meissner.1 785.3
5 C.N.Yang.1 114656 D.T.Son.1 2958 S.Sachdev.1 776.9
6 P.A.M.Dirac.1 111083 T.Padmanabhan.1 2824 S.Hod.1 752.2
7 Abdus.Salam.1 104766 P.Z.Skands.1 2804 D.Stanford.1 715.1
8 E.Witten.1 102753 S.Nojiri.1 2783 X.L.Qi.1 713.0
9 H.A.Bethe.1 102195 V.Springel.1 2701 S.D.Odintsov.1 704.1
10 G.tHooft.1 99645 A.Loeb.1 2594 Muhammad.Sharif.1 701.7
11 E.P.Wigner.1 90966 C.Vafa.1 2497 A.Loeb.1 700.3
12 T.D.Lee.1 86699 Ashoke.Sen.1 2468 S.Capozziello.1 674.2
13 W.Heisenberg.1 81238 G.P.Salam.1 2453 C.de.Rham.1 657.4
14 Stephen.Louis.Adler.1 • 76687 Martin.Bojowald.1 2373 L.Susskind.1 • 650.0
15 Yoichiro.Nambu.1 • 74084 S.Mrenna.1 2352 S.Tsujikawa.1 588.1
16 K.G.Wilson.1 69320 Ernest.Ma.1 2297 Bing.Zhang.1 577.6
17 S.L.Glashow.1 69308 A.C.Fabian.1 2285 D.W.Hooper.1 573.5
18 F.J.Dyson.1 68021 F.Aharonian.1 2189 C.Bambi.1 564.9
19 A.M.Polyakov.1 67138 Wayne.Hu.1 2162 R.B.Mann.1 564.0
20 J.D.Bjorken.1 63817 P.Nason.1 2132 A.Strominger.1 548.2
21 S.W.Hawking.1 62039 E.V.Shuryak.1 2121 E.Oset.1 534.9
22 B.Zumino.1 58680 G.R.Dvali.1 2100 P.Z.Skands.1 526.5
23 S.Mandelstam.1 57268 U.G.Meissner.1 2081 Xiao.Gang.Wen.1 520.4
24 G.Breit.2 52537 S.S.Gubser.1 2003 S.Nojiri.1 514.0
25 V.F.Weisskopf.1 52355 D.E.Kharzeev.1 1982 H.T.Fortune.1 512.1
26 Enrico.Fermi.1 52039 A.A.Tseytlin.1 1946 S.F.King.1 503.2
27 J.R.Oppenheimer.1 51472 M.Cacciari.1 1937 J.R.Ellis.1 500.6
28 David.J.Gross.1 • 50937 M.Zaldarriaga.1 1911 C.L.Kane.1 500.0
29 Peter.W.Higgs.1 • 50761 U.W.Heinz.1 1876 R.Myrzakulov.1 497.8
30 J.A.Wheeler.1 50238 Joseph.I.Silk.1 1873 A.Vishwanath.1 494.5
P
Table 8: Authors sorted according to the PaperRank of authors, RA = papers Rp /Naut for the
whole InSpire database (left), from year 2000 (middle), and from year 2010 (right).
front of their name, despite having less citations than modern authors, given the increase in
the rate of papers and of citations.
Anyhow, the main interest of our study is not re-discovering Feynman. We want to see if
our metrics do a better job than just citation counts in identifying modern authors with a high
impact. To achieve this, we set a lower cut-off on the publication year. We restrict the list
to papers published ‘From 2000’ (middle column of Table 8) and ‘From 2010’ (right column of
Table 8).
While switching from citations to individual citations is an obvious improvement, we find
that the rank does not improve over individual citations (they are strongly correlated) when
restricting to recent papers. As already discussed for papers, about 10-20 years are needed
before that the PaperRank becomes a better metrics. On shorter time-scales, the rank and
the number of individual citations are strongly correlated, and no significant differences arise;
a few authors have a rank significantly higher than their number of individual citations often
because they happen to be cited by reviews which fast received a large number of citations.
18
1 P.A.M.Dirac.1 292479 E.Witten.1 18748 N.Kidonakis.1 7248
2 Albert.Einstein.1 229323 T.Sjostrand.1 • 16669 E.Witten.1 5178
3 E.Witten.1 166374 J.M.Maldacena.1 16406 J.M.Maldacena.1 4799
4 Steven.Weinberg.1 147661 V.Springel.1 15267 P.Z.Skands.1 4335
5 G.tHooft.1 109341 P.Z.Skands.1 13705 A.A.Abdo.1 4294
6 J.S.Schwinger.1 98088 S.Mrenna.1 11931 M.Czakon.1 4160
7 Max.Born.1 81451 C.Vafa.1 11840 D.Stanford.1 4160
8 S.W.Hawking.1 79424 G.P.Salam.1 11091 X.L.Qi.1 3970
9 A.M.Polyakov.1 69293 D.T.Son.1 10511 M.Luscher.1 3929
10 R.P.Feynman.1 68094 P.Nason.1 10405 G.Harry.1 3666
11 M.Gell.Mann.1 66393 J.A.M.Vermaseren.1 9195 L.Susskind.1 • 3322
12 C.N.Yang.1 65056 Alan.D.Martin.1 8994 Alexander.Romanenko.1 3318
13 Enrico.Fermi.1 64907 S.D.M.White.1 8761 Xiao.Gang.Wen.1 3249
14 K.G.Wilson.1 54851 B.R.Webber.1 8673 S.Sachdev.1 3224
15 H.A.Bethe.1 53697 A.Loeb.1 8519 A.Strominger.1 3202
16 T.D.Lee.1 50413 Ashoke.Sen.1 8427 H.T.Janka.1 3108
17 L.Susskind.1 • 46873 L.E.Hernquist.1 8331 P.Nason.1 3089
18 Abdus.Salam.1 46278 J.Polchinski.1 8193 J.Rojo.1 3083
19 S.L.Glashow.1 45184 M.Cacciari.1 8146 A.Mitov.1 3077
20 R.W.Jackiw.1 44070 S.Frixione.1 8091 G.P.Salam.1 2934
21 F.A.Wilczek.1 • 43153 L.Susskind.1 • 8030 C.L.Kane.1 2831
22 H.M.Georgi.1 • 40178 A.C.Fabian.1 7809 F.Maltoni.1 2775
23 Stephen.Louis.Adler.1 • 39928 Wayne.Hu.1 7797 A.Vishwanath.1 2739
24 E.P.Wigner.1 39752 A.Strominger.1 7750 U.G.Meissner.1 2702
25 Sidney.R.Coleman.1 39247 M.Luscher.1 7710 J.Polchinski.1 2649
26 David.J.Gross.1 • 38777 Nathan.J.Berkovits.1 7679 Patrick.Huber.1 2606
27 A.D.Linde.1 38338 Xiao.Gang.Wen.1 7623 C.de.Rham.1 2577
28 J.D.Bjorken.1 38272 U.G.Meissner.1 7613 A.Loeb.1 2521
29 W.Heisenberg.1 36412 A.Y.Kitaev.1 7472 E.Berger.1 2492
30 S.Mandelstam.1 35261 F.Aharonian.1 7438 S.Forte.2 2463
Table 9: Authors sorted according to their Author Rank RA for the whole InSpire database
(left), from year 2000 (middle), and from year 2010 (right).
Restricting the sums to the N -th best papers of each author has little effect.
3.4 Author Rank

As outlined in the introduction, the citation matrix between authors (properly normalized)
CA0 →A defined in Eq. (6) allows to define an AuthorRank as
X
RA = ℘ RA0 CA0 →A + α(1 − ℘) , (14)
A0
where the second term gives a constant weight to each author, independently from the number
and quality of its papers. The network of citations among authors avoids time-directness, up to
time-scales comparable to the scientific ages of authors (which is enough for some goals, such as
studying how senior authors evaluate the work of younger authors). While formally analogous
to the ranking of papers, this ranking of authors is not a model of a physical process, because
one reads papers, not authors. The graph corresponding to the matrix CA0 →A contains cycles
19
1 E.Witten.1 3428 J.M.Maldacena.1 253.5 J.M.Maldacena.1 68.0
2 Steven.Weinberg.1 2261 T.Sjostrand.1 • 228.8 D.Stanford.1 50.2
3 G.tHooft.1 1366 P.Z.Skands.1 223.0 S.D.Odintsov.1 49.0
4 S.W.Hawking.1 1217 S.Mrenna.1 207.7 C.de.Rham.1 48.7
5 A.M.Polyakov.1 923.5 E.Witten.1 197.0 E.Witten.1 48.2
6 J.M.Maldacena.1 824.3 S.D.Odintsov.1 192.1 A.Strominger.1 47.9
7 J.S.Schwinger.1 739.7 G.P.Salam.1 181.2 N.Kidonakis.1 47.2
8 T.Sjostrand.1 • 736.1 S.Nojiri.1 173.1 P.Z.Skands.1 44.7
9 F.A.Wilczek.1 • 673.7 D.T.Son.1 162.8 M.Luscher.1 39.6
10 L.Susskind.1 • 658.6 V.Springel.1 157.4 G.P.Salam.1 38.5
11 R.W.Jackiw.1 655.8 T.Padmanabhan.1 156.6 M.Czakon.1 38.2
12 A.D.Linde.1 652.6 M.Cacciari.1 151.7 L.Susskind.1 • 36.9
13 S.L.Glashow.1 636.8 P.Nason.1 148.5 F.Maltoni.1 34.6
14 N.Seiberg.1 627.8 C.Vafa.1 134.7 S.Nojiri.1 34.0
15 P.A.M.Dirac.1 624.4 Ashoke.Sen.1 125.7 E.P.Verlinde.1 33.5
16 Sidney.R.Coleman.1 622.9 N.Arkani.Hamed.1 119.6 R.C.Myers.1 33.1
17 H.M.Georgi.1 • 614.4 A.Strominger.1 119.5 S.Tsujikawa.1 32.2
18 K.G.Wilson.1 552.5 V.A.Kostelecky.1 113.1 A.Mitov.1 32.0
19 David.J.Gross.1 • 543.2 J.Polchinski.1 108.0 G.Soyez.1 30.8
20 C.Vafa.1 521.4 G.Soyez.1 99.7 P.Nason.1 30.6
21 M.Luscher.1 508.0 D.E.Kharzeev.1 97.6 M.Cacciari.1 29.7
22 R.P.Feynman.1 459.9 P.Horava.1 97.3 A.D.Linde.1 28.4
23 A.Strominger.1 450.9 S.S.Gubser.1 96.5 A.A.Abdo.1 28.2
24 R.L.Jaffe.1 441.2 M.Luscher.1 94.6 Olivier.Mattelaer.1 28.0
25 J.Polchinski.1 439.7 S.Tsujikawa.1 93.8 A.De.Felice.1 27.9
26 J.D.Bjorken.1 422.5 L.Susskind.1 • 93.7 X.L.Qi.1 26.8
27 M.Gell.Mann.1 420.0 S.Frixione.1 92.7 D.Simmons.Duffin.1 26.7
28 C.N.Yang.1 415.9 N.Seiberg.1 90.3 K.Hinterbichler.1 26.4
29 B.Zumino.1 412.8 A.Ashtekar.1 89.4 R.Venugopalan.1 26.4
30 Abdus.Salam.1 410.0 M.A.Stephanov.1 87.5 N.Seiberg.1 26.2
Table 10: Authors sorted according to their Citation-coin C

//A of Eq. (15) for the whole InSpire
database (left), from year 2000 (middle), and from year 2010 (right)..
and also loops on the same node (self-citations): here we use ℘ = 0.9, such that self-citations
cannot boost the AuthorRank by more than one order of magnitude.
The left column of Table 9 shows the all-time AuthorRank. We see the emergence of old
authors such as Einstein, which were absent from previous top-rankings because poorly covered
and cited in the too recent InSpire database. Of course, the incompleteness of InSpire before
∼ 1960 makes results about older authors semi-quantitative.
This issue is avoided in the other columns of Table 9, where we show the AuthorRank
recomputed by restricting to recent papers only.
3.5 Removing self-citations and citation ‘cartels’: the Citation-coin

One of the unsatisfactory aspects of previous metrics is the effect of self-citations.
On short time-scales the PaperRank R and the number of (individual) citations are strongly
correlated, so they can be similarly inflated trough self-citations. Only on longer time-scales R
becomes a more robust measure: citations from paper p0 are weighted by its rank Rp0 , giving
20
relatively less weight to ‘below average’ papers that sometimes contain many self-citations.
Still, many below average papers can sum up to a significant total rank.
One can optionally count citations from published papers only, ignoring citations from
unpublished papers. However this choice discards information on good unpublished papers
(some well respected authors do not publish some of their papers).
Removing all self-citations is, by itself, an arbitrary choice. Furthermore it can be im-
plemented in different ways, for example removing citations from co-authors. Removing only
citations of an author to itself amounts to set to zero the diagonal elements of the citation
matrix NAicit
0 →A , reducing its N
2
entries to N (N − 1).
This does not protect from ‘citation cartels’. A second step in this direction consists in
removing citations exchanges A → A0 and A0 → A between all pairs of authors A, A0 . This
icit
amounts to subtract the symmetric part of the NA→A 0 matrix, reducing its entries to N (N −1)/2.
A third step is removing citations exchanges A → A0 , A0 → A00 and A00 → A among triplets
of authors A, A0 , A00 . A fourth step is removing all quadruplets etc.
A combinatorial computation shows that, after removing all possible ‘cartels’, only N − 1
entries remain. They can be described by N numbers C //A that sum up to 0. The meaning of C //A
0
can be intuitively understood by viewing NA0 →A as the total amount ‘paid’ by A to A, and N icit
icit
as a matrix of transactions. Then the Pphysical quantity unaffected by cartels is the net amount
icit icit
‘owned’ by each author A: C //A = A0 (NA0 →A − NA→A0 ). Citations are treated like money:
subtracting all possible citation ‘cartels’ is equivalent to count citations received as positive,
and citations given as negative. In doing this we proceed as described above: we actually count
individual citations (‘icit’): shared between co-authors, and divided by the number of references
of each paper. Then the price paid is the total number of papers written, independently of
their number of references. The Citation-coin of authors, C //A , can be written in terms of the
Citation-coin of their papers, C //p :
X C //p
C
//A = aut
, //p = Npicit − f (tend − tp )
C (15)
p∈A
N p
with f (tend − tp ) = 1. A paper has C //p < 0 when it receives a below-average number of
individual citations. This is the case for all recent papers, suggesting the introduction of a
factor f (tend − tp ) ≤ 1. All metrics penalise recent papers and authors (namely those active a
few years before tend , the end-date of the database), because citations accumulate over time.
This penalisation is stronger in the Citation-coin than in citation counting, if a paper p is paid
at the moment of its appearance tp . This boundary effect can be compensated by redefining
the Citation-coin in such a way that papers are ‘paid in instalments’, with the same time-scale
over which citations accumulate. One can conveniently choose f (t) ≈ 1 − e−t/τ , where τ ≈ 7 yr
is presently the time-scale over which citations are received by an average paper in the InSpire
database.17 In this way f (t) ' 1 for old papers, f (t) < 1 for young papers, f (t) ' 0 for very
recent papers. As a consequence recent authors do not necessarily have a negative Citation-
17
This number is obtained considering the whole InSpire and it varies by a factor of around two across the
main arXiv categories. As for fractional counting, in the case one wants to fine-tune comparisons within a given
sub-field, this number can be tuned to the behavior of that sub-field.
21
coin, and the lack of knowledge of future citations does not penalise their C
// more than their
number of citations.18
Table 10 lists authors according to their Citation-coin C //A . Authors that scored highly in
previous ranks by writing many papers with low impact have now disappeared, and some of
them actually got a negative C //A .
It is interesting to compare the Citation-coin to traditional metrics. Unlike the number of
(individual) citations, and unlike the number of papers, the C // does not reward authors that
publish many low impact papers. Like the h-index, the C // rewards both quality and quantity,
but in a different way. The h-index puts a threshold on the number of citations: as a result
papers below don’t contribute, and all papers above count equally. On the other hand, excellent
papers contribute significantly to the C //, while below-average papers contribute negatively. In
this respect, the Citation-coin can be considered an improvement of the h-index that is both
more indicative and less correlated (see Section 3.6) with citation counting. The C // also differs
from hNcit i, the number of (possibly individual) citations averaged over all papers of the author
under consideration. Unlike the C //, hNcit i only rewards quality: an author can maximise it
by writing only very few excellent papers. At the practical level, the C // identifies physicists
considered most esteemed (see Table 10), while traditional metrics fail (see Table 6).
Finally, from the database we can extract detailed reports about the metrics of each author:
Npap , Ncit , h, Nipap , Nicit , R, R, C //, their time evolution, the scientific age, the percentage of
given and received self-citations, the topics studied, the main collaborators, who the author
cites most, who cites the author most, etc. Similarly, these informations can be extracted and
compared for a group of authors. These detailed reports would be a good target for a future
improvement of the InSpire author profiles.
3.6 Author metrics: correlations

The right panel of Fig. 2 shows the correlations among the metrics for authors within the
whole InSpire database. The metrics are: number of papers NApap , number of citations NAcit ,
h-index squared h2A , number of individual papers NAipap , number of individual citations NAicit ,
PaperRank of authors RA , AuthorRank of authors RA , and Citation-coin C //A . We consider the
square of the h index since this is known to be almost fully correlated (0.99) with the number
of citations (Hirsch, 2005). From the table we see that our metrics for authors differ strongly
from the traditional ones, and also mildly differ between them. The main difference arises
because of the difference between experimentalists and theorists, so that the metrics become
more correlated if restricted within each group. However, the combined effect of dividing by
the number of references of the citing paper and the number of co-authors of the cited one
makes our proposed bibliometric indicators fairly uncorrelated with the existing ones.
18
Furthermore, one can define a C //+ metric which discards ‘negative’ papers with C
//p < 0 and sums the
contributions of ‘positive’ papers.
22
Institution All Institution After 2010 hep-ex hep-ph hep-th gr-qc astro-ph hep-lat nucl-th nucl-ex
1 CERN 25065 CERN 2817 413 652 184 10 49 66 8 84
2 Princeton U. 10774 Fermilab 1346 227 289 3 3 111 23 2 10
3 Harvard U. 10702 Brookhaven 791 77 202 7 0 15 103 54 73
4 SLAC 9913 Perimeter Inst. 751 0 81 434 133 35 2 0 0
5 Princeton, Inst 9699 DESY 705 147 274 70 1 38 10 1 9
6 Fermilab 9536 Princeton, Inst 686 0 41 522 6 90 0 0 0
7 Caltech 8470 Dubna, JINR 639 124 137 55 9 12 6 34 53
8 MIT, LNS 7658 KEK, Tsukuba 629 99 70 37 6 24 45 5 5
9 Brookhaven 7149 SLAC 603 71 177 47 2 72 0 1 5
10 LBL, Berkeley 6353 Caltech 584 40 53 159 93 166 1 2 4
11 Cambridge U. 5586 Beijing, Inst. 581 204 125 18 5 69 4 9 4
12 SUNY, Stony Bro 5371 Princeton U. 554 39 37 232 24 132 0 0 5
13 UC, Berkeley 4838 Munich, Max Pla 525 56 191 123 2 53 12 0 6
14 DESY 4580 LBL, Berkeley 519 57 84 27 1 80 4 28 37
15 Moscow, ITEP 4495 Valencia U., IF 492 52 304 7 14 36 7 17 1
16 Dubna, JINR 4431 Imperial Coll., 488 80 37 179 25 42 1 1 6
17 Washington U., 4089 Cambridge U., D 460 0 29 262 79 63 16 3 0
18 Munich, Max Pla 4024 Tokyo U., IPMU 416 27 116 165 15 67 2 1 0
19 Maryland U. 3801 Heidelberg, Max 400 33 221 1 0 68 0 3 9
20 Cornell U., LNS 3786 Stanford U., Ph 383 14 20 215 16 55 0 0 4
21 Los Alamos 3573 INFN, Rome 365 84 68 12 32 61 10 5 17
22 Oxford U. 3570 APC, Paris 359 22 9 57 72 170 0 0 0
23 Chicago U., EFI 3569 UC, Berkeley 358 15 79 70 7 103 4 3 5
24 Columbia U. 3560 Maryland U. 352 24 87 35 49 74 14 6 6
25 Imperial Coll., 3414 Heidelberg U. 352 63 85 44 4 37 10 18 23
26 Pennsylvania U. 3304 Potsdam, Max Pl 348 0 3 166 132 30 0 1 0
27 UCLA 3200 MIT 348 65 35 37 18 35 4 26 24
28 KEK, Tsukuba 3164 Munich, Tech. U 344 24 196 9 0 23 4 26 12
29 Wisconsin U., M 3101 Washington U., 338 21 57 48 6 22 67 41 12
30 Stanford U., Ph 3080 Harvard U., Phy 334 15 48 189 2 16 0 0 0
31 Saclay 3054 Jefferson Lab 333 10 119 0 0 0 48 22 41
32 Texas U. 2941 Wisconsin U., M 332 61 102 23 0 96 0 3 6
33 UC, Santa Barba 2920 Zurich, ETH 332 45 66 79 6 31 13 0 5
34 Santa Barbara, 2906 Zurich U. 324 40 190 2 12 56 0 0 4
35 Yale U. 2887 Frascati 324 81 46 13 3 3 6 0 22
36 Heidelberg U. 2847 Oxford U. 315 75 35 14 12 108 0 3 2
37 Harvard-Smithso 2633 Durham U., IPPP 314 2 267 14 0 18 0 0 0
38 Lebedev Inst. 2609 NIKHEF, Amsterd 312 61 114 18 21 23 0 0 18
39 Illinois U., Ur 2586 SUNY, Stony Bro 310 35 56 66 0 29 7 57 23
40 Argonne 2559 UC, Santa Barba 306 23 17 180 21 28 6 0 3
41 Kyoto U. 2445 Moscow, ITEP 305 55 72 66 2 6 13 7 33
42 Rutgers U., Pis 2441 Frankfurt U. 304 4 85 7 20 17 49 91 12
43 Tokyo U. 2430 Kyoto U., Yukaw 302 0 36 132 32 44 16 25 0
44 Durham U. 2397 UCLA 301 19 33 105 2 77 1 3 15
45 St. Petersburg, 2385 INFN, Pisa 296 63 45 25 16 44 15 13 3
46 Cambridge U., D 2343 Tokyo U. 294 19 78 43 9 25 23 22 15
47 Moscow, INR 2234 Ohio State U. 286 28 61 24 0 40 12 89 10
48 ICTP, Trieste 2183 Columbia U. 285 19 37 48 5 91 31 10 9
49 Bohr Inst. 2163 McGill U. 284 14 87 80 8 46 1 43 0
50 Minnesota U. 2139 Kyoto U. 282 20 36 77 23 51 12 23 6
51 Utrecht U. 2117 INFN, Turin 281 41 88 27 4 27 6 7 36
52 Frascati 2097 Bonn U. 279 55 93 27 0 16 11 10 8
53 Novosibirsk, IY 2096 Cracow, INP 278 47 134 0 0 6 0 57 12
54 Landau Inst. 2034 Harvard U. 278 5 35 177 6 33 0 0 0
55 Michigan State 2033 Madrid, IFT 274 1 105 96 9 38 8 3 0
56 Ohio State U. 2021 Los Alamos 271 20 78 3 3 26 18 33 22
57 Michigan U. 1963 UC, Irvine 267 38 100 6 0 83 0 0 0
58 Rutherford 1951 Orsay, LPT 265 1 179 30 26 3 13 0 0
59 Weizmann Inst. 1950 Tata Inst. 264 22 36 66 40 10 20 20 6
60 Rome U. 1897 Weizmann Inst. 258 16 55 80 3 29 1 1 6
61 MIT 1880 Southampton U. 253 1 134 48 26 13 25 0 0
62 Munich, Tech. U 1865 Stanford U., IT 253 1 45 180 1 14 0 0 0
63 Serpukhov, IHEP 1855 Pennsylvania U. 252 22 5 111 3 64 0 0 2
64 Penn State U. 1813 Bohr Inst. 252 18 52 103 6 33 14 3 12
65 Orsay, LPT 1804 Aachen, Tech. H 247 66 73 1 1 49 0 0 8
66 Ecole Normale S 1791 Yale U. 246 29 16 32 1 39 5 26 25
67 Hamburg U. 1777 Moscow, INR 241 39 43 23 4 47 0 3 18
68 Syracuse U. 1756 Darmstadt, GSI 240 7 36 1 0 13 5 42 26
69 Indiana U. 1755 Saclay, SPhT 237 0 102 75 3 19 0 27 2
70 UC, San Diego 1717 INFN, Padua 237 49 58 24 6 37 0 1 20
71 Rochester U. 1714 Argonne 236 33 83 2 0 12 5 15 11
72 Carnegie Mellon 1687 Manchester U. 235 65 73 12 0 50 0 6 2
73 Beijing, Inst. 1683 Johns Hopkins U 233 16 64 19 3 122 0 0 2
74 Garching, Max P 1665 INFN, Trieste 230 21 45 45 23 43 0 0 7
75 Chicago U. 1656 British Columbi 229 18 15 79 17 68 0 1 2
76 Frankfurt U. 1644 Michigan State 226 27 106 1 0 17 6 24 12
77 Tel Aviv U. 1604 Texas A-M 225 31 43 50 2 25 0 29 18
78 Texas A-M 1596 Humboldt U., Be 225 15 67 67 2 13 46 0 0
79 Zurich, ETH 1587 Rome U. 224 22 47 13 37 61 6 4 4
80 Bern U. 1584 MIT, Cambridge, 218 0 93 85 0 14 16 4 0
Table 11: Institutions listed according to their contribution to fundamental physics (as defined
by InSpire) quantified as the number of individual citations received by their affiliates as defined
in Eq. (16). Left: all time. Right: from 2010, and split within arXiv categories. The top ten
in each category are highlighted.
23
4 Rankings groups
Our metrics respect sum rules and thereby allow to define metrics for groups by simply summing
over their members. Furthermore, the main property of the Citation-coin holds not only for
authors of a set of papers, but for any group: it cannot be increased trough internal citations.
In order to show illustrative results we mostly use the number of individual citations as the
metric to rank groups, since individual citations have the benefit of being very fast to compute.
Of course, one could equally apply any other metric discussed before.
4.1 Ranking institutions

We share the number of citations received by each paper between the institutions I of its
authors with weights pI that sum up to 1. The weights are computed by first sharing equally
each paper between its Naut authors, and next between the affiliations of each author. Thereby
each author A contributes to institute I as 1/(NAaff Npaut ), where NAaff is the number of affiliations
of author A, that include institute I. When some affiliation is missed by InSpire, we renormalize
the pI such that they sum up to 1. Next, we sum over all papers p (optionally restricting to
recent papers, if one wants to evaluate the present situation, rather than the all-time record).
In formulæ, the number of individual citations received by institute I is:
X 1 X 1
NIicit = ppI Npicit , ppI = . (16)
p
Npaut A∈I,p NAaff
In the left column of Table 11 we list the institutes that most contributed to fundamental
physics, according to the whole InSpire database. In order to focus on the present situation,
in the second column we restrict to recent papers, written from 2010. The top positions are
occupied by research institutions rather than by teaching institutions. In the right columns
of Table 11 we show the contributions within each main arXiv category: the best institutions
strongly depend on the sub-field of interest. This means that generic rankings (e.g. at the
faculty level) are not much useful for authors interested in finding the most active institutions
within their specific sub-fields.
Concerning the time evolution, we compute the percentage of individual citations received
by papers written within any given year by authors affiliated to each institute, and fully fraction-
alized according to Eq. (16). Fig. 4a shows how the percentage impact of some main institutes
evolved with time. As papers published recently will accumulate most of their citations in the
future, in Fig. 4b we repeat the analysis in terms of fractionally counted papers. This indicator
provides a less significant proxy for scientific merit than citations, but its value is immediately
available, and helps to interpret the evolution in citations.
The black curve in Fig. 7 shows the time-dependence of the contribution of the top insti-
tution, CERN. It reached a maximum around 1965 (12% of world-wide individual citations
to 1965 papers have been given to CERN authors) and declined stabilising to ≈ 2%. All
main historical institutions show a similar trend, due to the fact that, until 1970, fundamental
physics was concentrated in a few institutions, and later became more distributed (especially
24
Institution Niaut Nicit PaperRank R AuthorRank R
1 Princeton, Inst. Adv 32 2.1 % 2.2 % 2.4 %
2 CERN 896 1.7 % 2.0 % 2.7 %
3 Fermilab 435 1.4 % 1.3 % 1.8 %
4 Texas U. 46 1.0 % 3.6 % 1.6 %
5 Caltech 90 0.89 % 3.6 % 1.6 %
6 Brookhaven 218 0.83 % 0.97 % 0.94 %
7 Cambridge U., DAMTP 49 0.78 % 0.96 % 0.79 %
8 SLAC 141 0.73 % 0.83 % 0.90 %
9 Dubna, JINR 312 0.62 % 0.54 % 0.51 %
10 DESY 276 0.59 % 0.46 % 0.57 %
11 Princeton U. 74 0.59 % 0.74 % 0.71 %
12 Harvard U., Phys. De 35 0.58 % 0.89 % 0.74 %
13 Maryland U. 72 0.55 % 0.45 % 0.46 %
14 Imperial Coll., Lond 121 0.54 % 0.56 % 0.55 %
15 Perimeter Inst. Theo 70 0.53 % 0.38 % 0.41 %
16 IPhT, Saclay 40 0.52 % 0.53 % 0.50 %
17 Stanford U., ITP 22 0.51 % 0.69 % 0.62 %
18 LBL, Berkeley 116 0.49 % 0.68 % 0.60 %
19 Munich, Max Planck I 124 0.47 % 0.36 % 0.44 %
20 Chicago U., EFI 41 0.46 % 0.60 % 0.54 %
Table 12: Institutes listed according to all-time bibliometric ranks of their last-year affiliates.
For each institute, we show the number of individual authors active in the last year, and the sum
of their biblio-metric indicators (percentage of the world total): number of individual citations,
PaperRank of authors, and AuthorRank of authors.
in theoretical physics). Half of the impact, quantified using individual citations, was produced
in the 12 top institutions in 1970, in 22 in 1980, 42 in 1990, 80 in 2000 and 160 now. As a
consequence the relative impact of the top institutions declined. Among the new institutions,
Perimeter and IPMU reached very high positions.
The list in Table 11 highlights the institutes with the most productive authors in recent
times (often young authors). The list in Table 12 highlights institutes with the most productive
affiliates, as instead quantified by their all-time biblio-metric rankings. Table 12 is produced
as follows. Since no list of present affiliates is available, we use the declared affiliations of
authors that wrote at least one paper in the last year, 2020. Authors with Naff affiliations are
assigned with fraction 1/Naff to each affiliation. When this number differs in recent papers,
we average over them respecting sum rules: each authors is affiliated to various institutions
with percentages that sum up to one. The average suppresses minor mistakes/missing data
in InSpire. For each institute I, we obtain a list of active affiliates with their percentages.
Summing over these percentages we determine the number of ‘individual authors’ Niaut affiliated
to each institution, shown in the 3rd column of Table 12. Next, summing over all affiliates using
the same weights, we compute the total biblio-metric ranking of all authors in each institute.
In column 4 we show the all-time number of individual citations, in column 5 the PaperRank
(Section 2.1) and in column the AuthorRank (Section 3.4). The latter 3 columns actually show
the world percentage of each metric in the various institutes: about 2% of researchers that most
25
��
��
��
��
��
��
��

��
��
��
��
��
��
��
��
��
��
� �
��
��
Figure 4: Percentage impact of some main institutes according to citations (left) and papers
(right).
contributed to fundamental physics can be found at IAS in Princeton, or at CERN.
4.2 Ranking towns

Sometimes multiple institutes are located nearby, and what matters is their total. We group
together institutes closer than about 30 km. In Fig. 5 we show a map with the places that
mostly contributed to fundamental physics: each contribution is plotted as a circle with area
proportional to the number of individual citations received by their papers written from year
2010, and color proportional to the contribution to experiment (green), theory (red), astro-
cosmology (blue) respectively. We focus on a relatively recent period, such that the map
photographs the present situation.
Similar maps can be computed for any given sub-topic or region. For instance, Fig. 6 shows
the same map separated according to papers published within the main arXiv categories, and
restricted to Europe.
4.3 Ranking countries and continents

We rank a country or continent C by summing the ranks of all institutes located in C. We
apply this to the number of individual citations:
X
NCicit = NIicit . (17)
I∈C
In the left panel of Fig. 7 we show the time evolution of the impact of papers written in main
representative countries within each year. The impact is quantified as the percentage of the
26
Figure 5: Institutes that contributed to fundamental physics plotted as dots with area propor-
tional to the number of individual citations received by their papers written from year 2010. We
group institutes closer than 30 km. The amount of green, red, blue in the color is proportional
to the contribution to experiment, theory, astro-cosmology respectively.
world total, in order to factor out the reduced number of citations of recent papers. USA is the
main country, but declining (from 70% around 1950 to 25% now); European countries are now
stable or slightly growing; China is growing. In the right panels of Fig. 7 the time evolution
of the percentage contribution of each country is shown separately, after the advent of arXiv,
within the main fields: experiment, theory, astro-cosmology.
Figure 8 shows the analogous plot for continents. We see that European physics suffered a
big decline after WW2, and returned to be the main actor only around 2000. The decline of
Asia around 1985 is due to the fall of Soviet Union (Mathematica geographic tools assign all
Russia to Asia); the present rise of Asia is mostly due to China.
Figure 9 shows the ratio between the number of individual citations and the population of
countries, while Fig. 10 shows the ratio between the number of individual citations and the
gross domestic product.
4.4 Ranking journals

When a paper is published on some journal (refereed or not), InSpire provides this information.
Table 13 lists journals according to the number of individual citations received by all papers
they published in fundamental physics, as included in InSpire. We separately show these data
for all InSpire, and restricting to articles published from 2000. Figure 11 shows the time
evolution of the percentage number of individual citations received by all papers on selected
journals. We see that internet brought a revolution around 2000: the decline of NPB and PLB
and the emergence of JHEP and Astrophysics J. as preferred journals.
27
Figure 6: As in Fig. 5, restricted to Europe, and showing separately the contributions within
main arXiv representative categories.
The 3rd column of Table 13 shows a direct measure of ‘quality’: the average number of
individual citations per paper, which roughly corresponds to what is known as ‘impact factor’.
According to this measure, the top-journals are those that publish reviews.
The 4th column shows a measure of both ‘quality’ and ‘quantity’: the Citation-coin C // of
the journal (difference between the number of individual citations and the number of published
papers). The top journal according to this measure is PRL. Journals that publish reviews
score well, but are limited by their restricted scope. The Citation-coin is negative for journals
that publish many papers that do not attract much interest, in particular those that publish
conference proceedings.
5 Conclusions
We applied improved bibliometric indicators to the InSpire database, which covers funda-
mental physics mostly after 1970. Figure 13 shows some main trends: growing rate of papers,
growing number of authors and of references per paper. Figure 15 shows the health status of
the main sub-fields.
The metrics that we explored are:
• The number of individual citations Nicit — defined as the number of citations divided
by the numbers of authors of the cited paper and of references of citing papers — that
28
�� -�� + ��-�� -�� + ��-�� + ��-�� + ��-�� -�� + ��-��
��
�
�
��
� �
��
� �
� ��
� � � �
� �
� � � �
� � ��
� � �
� � � �
� �
� � � � � �
� � �
� � � � � � � � � �
� � � �
��
� ��
� � � � �
��
� � �
��
��
��
��
� ��
��
�
�
� � �
��
��
�
��
��
� �
�
� �
�
�
�
�
�
�
� ��
�
� �
��
� ��
� ��
�
� � � �
� ��
� ��
� � � � � �
� � �
� � �
� ��
� �
� ��
�
� ��
� � ��
� � � � � � � � � � � � �
� � � � � �
� ��
� � � �
� � � � � � �
� � � � � � � � � � � � � � � �
��
� � � ��
� � � �
� �
��
��
� ��
� � � � � � �
�
� � � �
� �
� � � � � ��
�
��
� � �
�
� � � ��
� � � �
� ��
� � �
� � �
� � � � � �
� � �
��
��
�
� � �
� � �
� � � �
� � � � � � �
�
� ��
��
��
� �
� � � � � � �
� � ��
� � � � � � � � � � �
�
� � � � � �
� ��
� � � � ��
� � ��
� � ��
��
� � � � ��
� �
�
� ��
� �
��
� � � � �
� � � ��
�
�
��
��
� ��
��
��
��
��
��
� �
� �
�
�
�
�
� � � � �
�
�
�
�
� �
� ��
�
� � ��
�
�
�
� �
�
�
�
�
� �
�
� �
�
�
�
�
� �
��
�
�
�
�
�
� � � � � �
� �
� �
� � � � � � � � � � �
� � �
� � � � � � � � �
� ��
� � ��
��
� � ��
��
�
��
� � � � � �
� �
� � � �
� � � �
� � � � � � �
��
� � �
� � � ��
��
� � � � � � � � � � � � �
�
�
� � � ��
� �
� � � � � �
�
� � � � �
� � � � � � � � � � � � �
��
� � � ��
� � � � � �
� � �
��
�
� � ��
�
�
� � � � �
� � � � � ��
� � � � � � � �
� �
�
��
��
� � � � � � ��
�
� � � �
� � � � � � � � � � �
�
� ��
��
�
� � � �
� �
� � � �
� � � �
� �
� � � � �
� � �
��
��
��
� ��
� � � � �
� � �
� � �
� � �
� � �
��
��
� � � � � �
� � � � � � � � � �
� ��
� � � � � � ��
�
��
� � � �
� � � � � � � � � � � � �
��
� ��
� � � � � � � ��
� �
��
�
��
��
� � �
�
� �
� � � � � � �
� � � � �
� � � ��
� � �
��
� � � � ��
� � � � � � � � � � � �
� ��
��
� �
� �
� � � �
�
� � � � � � �
� ��
��
�
�
� ��
� �
�
� � � � � �
� � � � � �
� � ��
� ��
�
�
��
�
� �
��
� � ��
� ��
��
� ��
��
��
�
��
�
��
� � ��
��
��
��
��
��
��
��
��
��
��
��
�
��
� ��
��
��
� ��
�
�
� �
� � �
� � � � � � � � � � � � � � �
��
��
��/� �� * ��
Figure 7: Percentage impact of representative countries. For ease of visualisation, the USA
contribution has been multiplied by 1/5. The ∗ on France means that CERN is plotted separately.
�� -�� + ��-�� -�� + ��-�� + ��-�� + ��-�� -�� + ��-��

��
��
��
��
��
��
��
��
� � � �
��
��
��
Figure 8: Percentage impact by continent.
compensates for the recent inflationary trends towards more authors and references.
• The PaperRank R, which applies the PageRank algorithm to the citation network among
papers and that approximates a physical observable: how many times a paper is read.
• The AuthorRank R which applies the PageRank algorithm to the individual citation
network among authors.
• The Citation-coin C // equal to the difference between the number of received and given
individual citations. Since a naive definition of this indicator, that cannot be increased
29
Figure 9: Individual citations per country from 2010 divided by population. The top countries
are Switzerland, Germany, France, Italy, Slovenia.
Figure 10: Individual citations per country from 2010 divided by domestic gross product in
dollars. The top countries are Armenia, Estonia, Slovenia, Portugal, Italy.
by self-citations and circular citations, strongly penalizes young authors, we propose an

improvement that takes into account the average time in which citations are received
within the community.
An important feature of all these metrics is that they can be computed in practice.
In Section 2 we apply these metrics to papers. Table 1 shows the traditional list of all-time
top-cited (highest number of citations) papers. It can be compared with Table 2, which shows
top-ranked papers (papers with highest PaperRank, namely citations weighted proportionally
to the rank R of citing papers): the PaperRank retrieves old famous papers with relatively
few citations. When applied to the time-ordered citation network, the PageRank reduces to a
(weighted) counting of citations-of-citations described in Section 2.3. Thereby, when restricted
to recent papers, the PaperRank is dominated by the number of individual citations. Next,
Table 3 shows top-referred papers, where citations are weighted proportionally to the all-time
30
Journal, all InSpire Nicit Nicit /Npap C
// Journal, after 2010 Nicit Nicit /Npap C
//
1 Phys.Rev.D 113601 1.6 41080 Phys.Rev.D 19288 1.0 656
2 Phys.Lett.B 80214 1.5 25748 JHEP 13942 1.3 3179
3 Phys.Rev.lett. 67864 2.9 44359 Phys.Rev.lett. 9488 2.4 5517
4 Nucl.Phys.B 56150 2.3 31339 Phys.Lett.B 6008 1.3 1261
5 Astrophys.J. 38960 1.3 9920 Astrophys.J. 5929 0.6 −3700
6 Phys.Rev.C 30386 0.9 −4071 Phys.Rev.C 5666 0.9 −371
7 JHEP 30044 1.5 10024 Mon.Not.Roy.Astron.S 5386 0.5 −5091
8 Nucl.Phys.A 21518 0.7 −9310 Eur.Phys.J.C 4376 1.4 1216
9 Mon.Not.Roy.Astron.S 20594 0.9 −2525 JCAP 3174 1. −105
10 Nucl.Instrum.Meth.A 17592 1.0 380 Astron.Astrophys. 2647 0.5 −3068
11 Phys.Rev. 17351 2.4 10185 Nucl.Instrum.Meth.A 2291 0.7 −898
12 Astron.Astrophys. 14323 0.8 −4762 Class.Quant.Grav. 1896 0.8 −497
13 Astrophys.J.Lett. 10703 1.1 1351 JINST 1890 0.9 −268
14 Eur.Phys.J.C 9964 1.4 2801 Astrophys.J.Lett. 1661 0.7 −683
15 Phys.Rept. 9788 7.4 8462 Pos 1327 0.1 −9215
16 Class.Quant.Grav. 8822 0.9 −731 Nucl.Phys.B 1321 0.8 −327
17 Commun.Math.Phys. 7792 1.9 3658 J.Phys.Conf.Ser. 1183 0.2 −6450
18 Annals Phys. 7765 2.2 4169 J.Phys.G 1046 1. −44
19 Rev.Mod.Phys. 7235 4.8 5732 Phys.Rev.B 961 0.5 −967
20 Z.Phys.C 6939 1.4 1820 Astrophys.J.Suppl. 903 1.7 386
21 J.Math.Phys. 5717 0.7 −2289 Nucl.Phys.A 847 0.5 −1028
22 Astron.J. 5581 1.1 710 Nature 737 2.9 484
23 JCAP 5331 1.1 509 Eur.Phys.J.A 729 0.7 −252
24 Astrophys.J.Suppl. 4824 2.9 3148 Astropart.Phys. 717 1.4 196
25 Nature 4406 3.0 2945 Chin.Phys.C 706 0.7 −316
26 Prog.Theor.Phys. 4385 0.6 −2697 Phys.Rev.ST Accel.Be 672 0.9 −73
27 Int.J.Mod.Phys.A 4374 0.5 −4459 Comput.Phys.Commun. 630 2.0 314
28 Yad.Fiz. 4072 0.4 −5218 Int.J.Mod.Phys.A 622 0.4 −1152
29 Comput.Phys.Commun. 3864 2.6 2373 Phys.Rept. 584 6.2 490
30 JINST 3816 1.5 1330 Science 548 3.4 387
P
Table 13: Number of individual citations (Nicit = Ncit /Nref ), average number of individual
// = Nicit − Npap ) for some top journals.
citations per paper (Nicit /Npap ) and Citation-coin (C
The analysis is restricted to fundamental physics as included in the InSpire database. Left:
all time. Right: only papers published from year 2010.
AuthorRanks RA of citing authors. The AuthorRank identifies the recent papers that most
attract the attention of notable older authors. Finally the left panel of Fig. 2 shows the
correlations among paper metrics, showing how our proposed metrics are fairly independent
from the traditional ones and among each others.
In Section 3 we apply the new metrics to authors. Traditional metrics shown in Table 6 are
dominated by experimentalists who write more than 100 papers per year in collaborations with
more than 2000 authors. Considering instead the number of individual citations, the list in
Table 7 becomes dominated by theorists, especially those very active in relatively recent times.
Restricting to recent papers, the list includes some authors from fields that tend produce many
publications (tens per author per year). Less surprisingly, the list includes authors who produce
useful tools for collider experiments, which are presently very active. Ranking authors trough
their PaperRank, the all-time list in Table 8 is dominated by theorists (such as Weinberg,
31
��
�
� �
� �
� ��
��
� �
�
�
� �
�
� �
�
� �
� �
�
�
�
� ��
� �
�
�
� �
�
�
�
�
� �
� � � � �
� �
�
� ��
��
� �
� � �
�
�
�
�
�
�
� � �
� � �
� ��
� � �
� �
� � �
� � � � �
� � �
� �
� �
�
�
�
�
�
� � ��
� �
��
�
�
�
�
�
�
�
�
�
�
�
� �
�
� � � �
�
�
� �
�
� ��
� � � � � � � �
�
�
� � � � � � � � � � �
� � � � � � � � �
� � � � � �
�
�
�
� � � � � ��
� �
�
� �
�
�
� � ��
� � � �
� � � � �
� �
�
�
�
�
� � �
�
�
� � �
� � �
�
�
� �
� � �
� �
�
�
� �
�
�
� � � �
� �
� � �
� � � � � � � �
�
� � �
� � �
� �
�
� � �
�
� �
�
� �
� ��
� � � � � � � � � �
� � � � � � �
�
� � � � � � �
� �
� � � � � � � �
�
� � �
�
�
� �
� �
�
�
�
�
�
� � �
�
�
�
�
� �
� � �
� �
� � � �
��
� �
��
�
� � � � �
� �
� � ��
�
� � � � � � � � � � � � � � ��
� � � � � � � �
� � � � � � �
�
� �
� �
� � � � � � � � �
�
�
� �
��
� ��
�
� �
� � � � �
�
��
��
� � �
��
��
��
��
� �
� �
��
� � � �
��
��
� �
�
�
��
��
��
�
�
� ��
��
��
Figure 11: Time evolution of the fraction of individual citations received by papers published
on some notable journals.
Schwinger, Feynman, Gell Mann) that produced seminal papers after InSpire started, despite
the overall rate of papers and citations was a factor of few smaller than now (Fig. 14). On
the other hand, the recent-time list does not change significantly: the PaperRank is strongly
correlated to the number of individual citations, becoming a better metric only on longer time-
scales.
Due to this limitation, we developed the AuthorRank. Table 9 shows the result: authors
such as Dirac and Einstein now appear in the top of the list, despite having few papers with
few citations. The right columns of Table 9 shows recent authors listed weighting citations
according to the all-time AuthorRank of citing authors.
Table 10 lists authors according to their Citation-coin C
//: this metric rewards authors who
attract the interest of others by writing above-average papers, and penalising those that write
many below-average papers (or many recent papers, as papers need decades to be recognized
in terms of number of citations). The right columns of Table 10 again restrict the list to recent
times.
All our lists of authors also show the medals and prizes received by the various authors:
this shows that the non-traditional metrics agree much better than traditional metrics with
the opinions of the various panels (of course an unknown correlation between bibliometrics and
prize awards may still be present).
The right panel of Fig. 2 shows the correlations among indicators for authors. The metrics
we propose are fairly uncorrelated with traditional metrics and among each other.
Our metrics respect sum rules (their total is not inflated adding more authors or more ref-
erences) and are intensive: this means that groups can be ranked summing over their members.
In Section 4.1 we discussed the institutions that contain the authors that most contributed. In
Section 4.2 we grouped nearby institutes, providing maps of towns most active in fundamental
physics in different subfields. The same is done in Section 4.3 for countries and continents: in
view of the large statistics we also show the time evolution of their percentage impact. Finally,
32
in Section 4.4 we compute which journals publish the most impactful results in fundamental
physics, again showing the time evolution.
The different metrics that we propose give different informations on each author, providing
together a more complete view. Paperscape19 extracts information from arXiv and provides
a very useful visualisation of the citation graph among papers, and of the contribution of some
authors (those with unique names). It would be interesting to run the open-source Paperscape
code on the graph of individual citations among papers and authors extracted from the InSpire
database. Our indices could also be implemented in databases that index citations, such as
InSpire, in order to offer authors (institutes/journals/group) profiles with a larger variety of
information, able to give at one glance a much deeper and wider panoramic of each author
(institute/journal/group).
Several of our results with complete tables are available at the PhysRank webpage20 .
Disclaimer Technical details and limitations are described in the appendix. We repeat the
main caveats of our analysis: ‘fundamental physics’ here means ‘as included in the InSpire
database’; we do not correct for mistakes in InSpire (anyhow more accurate than commercial
databases). Omissions should be addressed to feedback@inspirehep.net or trough the on-line
forms on InSpire; we will update our results in some future. We just computed and showed
results, avoiding comments. We hope that authors of any field, journals boards, members of
institutes, towns, countries, and continents will understand that we cannot repeat all caveats
in all results.
Note added: 2021 update

All data have been updated to 2021/1/1. This is especially interesting, as many activities
switched to on-line mode in view of the covid-19 pandemic started during early 2020. Its effect
is in progress and will be better evaluated in the future. So far, Fig. 14 shows no noticeable
impact in the number of publications and related bibliometric indicators within the main sub-
fields. Similarly, Fig. 15 shows no noticeable impact on the number of authors who left the
various sub-fields, nor in the number of new authors. Fig. 7 indicates an accelerated growth
of publications from authors with Chinese affiliations relative to others. Fig. 8a shows that
Asia overcome NorthAmerica and that the European contribution, while still larger, declined
significantly. This is reflected in Fig. 4, that shows the time evolution of the percentage impact
of some main institutions.
Acknowledgments
We thank Roberto Franceschini, Christoffer Petersson, and Paolo Rossi for discussions that stimulated
this work. We thank Roberto Franceschini for participation in the first stage of this project. We are
19
http://paperscape.org.
20
http://rtorre.web.cern.ch/rtorre/PhysRank.
33
grateful to the InSpire team, and especially to Jacopo Notarstefano, for support and help with the
InSpire database.
A The InSpire and arXiv databases

The open-source InSpire bibliographic database21 covers fundamental physics world-wide. In-
Spire presently contains about 1.4 · 106 papers, 4 · 107 references, 105 authors, 104 institutions,
and 2·103 journals. InSpire started around 1965, but it also contains some notable older papers.
InSpire maps papers, authors, collaborations, institutes (affiliations), and journals to record
IDs (integer numbers) thereby addressing the problem of name disambiguation (Golosovsky
and Solomon, 2017, Martin Montull, 2011).
Starting from 1995, preprints for most of the papers contained in InSpire are available
through arXiv.org,22 which also covers fields beyond fundamental physics, so that not all of the
arXiv database is included into the InSpire one. The arXiv also provides a classification in
terms of categories, some of which contain sub-classes. The arXiv categories and the number
of papers in each of them are shown in the left histogram in Fig. 12.23 The right histogram
shows the fraction of papers in the various categories included in InSpire.
We also often show results for the main arXiv categories inside InSpire, defined as the arXiv
categories with more than 103 papers, and with a fraction included in InSpire larger than 50%.
These are: hep-ex (high-energy experiment), hep-ph (high energy theory/phenomenology), hep-
th (high energy theory), astro-ph (astrophysics and cosmology), hep-lat (lattice field theory),
nucl-ex (nuclear experiment), nucl-th (nuclear theory), gr-qc (general relativity and quantum
cosmology). Details of the dataset we consider and technical issues about the InSpire and
arXiv databases are discussed in Appendix A.2.
A.1 Main trends in the fundamental physics literature

The left panel of Fig. 13 shows the time evolution of some main factors: number of papers
per year (which increased by 5%/yr); average number of references per paper (increased by
3%/yr); number of citations per paper (roughly constant, taking into account that recent
papers, published in the past ≈ 15 years, necessarily received less citations); number of authors
per paper (increased from few to tens). We also see that most citations go to published papers.
In the right panel of Fig. 13 we see that the average age of references is increasing: before 1980
authors usually mostly cited recent papers. Contemporary papers cite references published ∆t
earlier with a distribution exp(−∆t/τ ) with τ ≈ 11 yr, see also Walker, Xie, Yan, and Maslov
(2007) and Wang, Song, and Barabási (2013).
Figure 14 shows the same trends within the main arXiv representative categories, showing
that papers with an increasingly large number of authors lie in experimental categories (hep-ex,
21
https://inspirehep.net.
22
https://arxiv.org.
23
The full list of sub-classes can be found in the arXiv API reference manual (https://arxiv.org/help/api/user-
manual).
34
Figure 12: Total number of papers in arXiv and fraction included in InSpire by gran cate-
gories. The number of papers in each category is computed considering only the main category
and ignoring cross-list, so that no paper can belong to more than one gran category.
nucl-ex, astro-ph), while papers in other fields keep having, on average, 2 − 3 authors.
Figure 15 shows the authors’ ‘birth’ and ‘death’ rates within the main arXiv categories as
function of time: in green the percentage of authors who published in year y but not in year
y − 1; in red the percentage of authors who published in year y − 1 but not in year y. The
balance is stable, with a significant growth of hep-ex when the Large Hadron Collider (LHC)
started, and of astro-ph until 2010.
Figure 16 shows the distributions of citations and individual citations for the whole InSpire
database. Citations do not follow a well behaved probability distribution due to the presence,
in the whole InSpire database, of communities with very different average number of authors.
This effect is clearer when going to the individual citations, that apply fractional counting
to the citations. Indeed, individual citations are reasonably well described by a log-normal
distribution, as already observed by Thelwall and Wilson (2014), with mean and standard
deviation both of order one. This shows how individual citations are well described by a
multiplicative stochastic process, allowing to combine and compare more heterogeneous sub-
fields.
In order to give an idea of how impact, as defined from citations and individual citations,
is distributed within the community, we computed the Gini index of these distributions. The
Gini index is used in economy as a measure of inequality of wealth. Just to give an idea of
its meaning, typical occidental countries presently have Gini coefficients between 0.2 and 0.4.
35
��
��
〈�� 〉 〈��
�� 〉
��
��

��
�� -��
��
��-��
��
��-��
� ��-��
��
�� 〉 ��-��
�
〈�
�� -��
�
〈�� 〉 ��-��
� � ��-��
��
��
Figure 13: Left: The vertical axis refers to different quantities: number of papers per year
(green), average number of references (red), of authors (magenta), of citations (blue), of ci-
tations among published papers (blue dashed). Figure 14 shows the same trends within arXiv
categories. Right: average age of references. Results before ∼ 1960 are affected by the smaller
size of the literature, and by the incomplete coverage in the InSpire database.
Even though this should not be interpreted analogously to the Gini coefficient in economy, it is
interesting to notice that the typical Gini coefficient of citations is between 0.7 (when looking
at single arXiv categories) and 0.8 (when considering the whole database). This means that
few papers get most of the citation impact: 4% of papers have more than 100 citations, and
they receive half of the total citations; half of the papers have less than 4 citations, and they
receive 2% of the total citations (see also Lehmann, Lautrup, and Jackson (2003)).
A.2 Details about the dataset

Any large database contains a small fraction of incomplete/inconsistent information, which may
affect any algorithmic study of the data in a variety of ways (del Corso and Romani, 2016).
The InSpire database is extremely curated. References are covered with an accuracy at the
% level, typically better than big private databases such as Scopus24 or WebOfScience25 , and
comparable to Google Scholar26 .
We obtained the InSpire database in the form of a ‘dump’ file.27 InSpire maps papers,
authors, collaborations, institutes (affiliations), and journals to record IDs (integer numbers)
thereby addressing the problem of name disambiguation (Golosovsky and Solomon, 2017, Mar-
24
https://www.scopus.com.
25
https://apps.webofknowledge.com.
26
https://scholar.google.com.
27
http://inspirehep.net/dumps/inspire-dump.html.
36
104 103 1.
Mean number of authors per paper
Fraction of published papers

0.8
Number of papers
102
0.6
103
0.4
101
0.2
102 100 0.
1995 2000 2005 2010 2015 2020 1995 2000 2005 2010 2015 2020 1995 2000 2005 2010 2015 2020
Year Year Year
80 80 3.
Mean number of references per paper
Mean number of citations per paper
Mean individual citations per paper

60 60
2.
40 40
1.
20 20
0 0 0.
1995 2000 2005 2010 2015 2020 1995 2000 2005 2010 2015 2020 1995 2000 2005 2010 2015 2020
Year Year Year
Figure 14: Trends within the main arXiv representative categories inside InSpire (papers
included in InSpire only).
tin Montull, 2011).

The extraction of an accurate date for each paper from the database suffers from some
uncertainty. There are several available dates: the date when the paper was added to the
database, a preprint date, often, but not always, corresponding to the arXiv preprint (when
available), a publication date (if the paper has been published on a journal), and an “earliest”
date, representing the first available date (not always present). Moreover, month information is
typically available only for arXiv papers and some published ones. In general we estimate our
uncertainty on the extracted dates at the percent level, in the sense that dates are accurately
extracted (at least the year) for about 99% of the papers. Given this uncertainty our sample
of papers consists of 1403881 papers from 1230 (de Sacrobosco, 1230) to 31 December 2020.
Another source of uncertainty comes from the author list of each paper. Some papers carry
an empty list of authors. This can have different reasons. For instance, only the name of the
collaboration is available for experimental papers (mainly conference notes) indexed from the
CDS database.28 All these papers are included in our analysis for what concerns papers, but
do not contribute to the metrics for authors. Similar problems extends to institutions and
28
https://cds.cern.ch.
37
��-�� -�� -�� -��
��
��

��
��
��
��

��
��
��

��
��
��-�� -�� -�� -��
��
��

��
��
��
��

��
��
��

��
��
Figure 15: Percentage of authors appeared or disappeared each year within the arXiv categories
fully covered by InSpire.
journals.
One more source of uncertainty is the extraction of references from papers, needed to produce
a citation map. This can be very simple when a bibtex or xml bibliography file is attached
to the paper, but can become an extremely complicated task for papers where only a pdf,
sometimes produced from a scanned paper, is present. InSpire uses state-of-the-art technology
for reference extraction (like Refextract29 and Grobid30 ), which is mainly automatic, with
human supervision only in case of errors and inconsistencies. Despite the advanced technology
for reference extraction, not all references are correctly extracted.
There are different kind of problems so that a reference can: simply be missed, be recognised
incompletely, be misidentified with another one, be assigned to an inexistent paper ID, point
out of the database (in which case it is counted in the number of references, but not indexed),
or point to a later paper (“a-causal” reference). All these effects are observed in the database.
While some are simple mistakes, or typos in the ID of the paper, the last could be a real
effect (below 1%), with the reference appearing in a subsequent version of the paper with no
available information on the dates of the different versions. Since “a-causal” citations can
generate anomalies in the computation of the rank (typically only when the a-causality is large,
i.e. in case of mistakes), we deleted them from our dataset. However, since dates are extracted
with an accuracy that is often of one year, we still consider causal all the references to papers
with the same date, regardless of the actual papers appearing order.
29
https://github.com/inspirehep/refextract.
30
https://github.com/kermitt2/grobid.
38
Figure 16: Distribution of citations (left) and individual citations (right) in the whole InSpire
database. The blue curve in the right panel is the fit to data with a log-normal distribution with
the parameters indicated in the caption. Both the distributions give a Gini index close to 0.8.
In any case, especially because of references pointing outside the database, the number of
references indexed for any given record is a better estimate of the actual number of references
than the one obtained by summing the indexed ones. Only when computing PaperRanks and
AuthorRanks, the number of references is defined equal to the number of indexed references,
which is needed to correctly normalize the transition matrix defining the citation graph. Given
that InSpire is complete only after 1970, this means that references of older papers are over-
attributed to those old notable papers that happen to be in InSpire.
In summary, the dataset consists of 1403881 papers, 78941 indexed authors, 8321 institutes,
2644 journals, 37172002 references (of which 26333765 indexed). We compute citations directly
from references and do not use citation information from the InSpire database.
Concerning the information on the arXiv database used in Fig. 12, we imported all records
using the arXiv API.31 All other information on arXiv papers and categories has been obtained
from the InSpire database. The full list of arXiv categories and subject-classes can be found
in the arXiv API reference manual.
All indices discussed in this paper can be computed in one hour on a laptop, apart for the
Author Rank, which involves a large, not very sparse matrix: ∼ 8 · 104 × 8 · 104 with about
5 × 108 non-vanishing entries.
References
Aksnes, D. W., Schneider, J. W., and Gunnarsson, 6 (1), 36–43. (Semantic Scholar) doi: [link]
M. (2012). Ranking national research systems by ci-
tation indicators. A comparative analysis using whole Bini, D., del Corso, G. M., and Romani, F. (2010).
and fractionalised counting methods. J. Informetrics, A combined approach for evaluating papers, authors
31
https://arxiv.org/help/api/user-manual.
39
and scientific journals. J. Comput. Appl. Math., 234 , Henneken, E. A. and Kurtz, M. J. (2017). Usage Bib-
3104-3121. (Semantic Scholar) doi: [link] liometrics as a Tool to Measure Research Activity.
Bini, D. A., del Corso, G. M., and Romani, F. (Semantic Scholar)
(2008). Evaluating scientific products by means of
Hirsch, J. E. (2005). An index to quantify an individ-
citation-based models: a first analysis and valida-
ual’s scientific research output. Proc. Nat. Acad. Sci.,
tion. ETNA, Electron. Trans. Numer. Anal., 33 , 1–
46 . (Semantic Scholar) doi: [link]
16. (zbMATH)
Holtkamp, A., Mele, S., Simko, T., and Smith, T.
Bouyssou, D. and Marchant, T. (2016). Ranking
(2010). Inspire: Realizing the dream of a global digi-
authors using fractional counting of citations: An ax-
tal library in high-energy physics.
iomatic approach. J. Informetrics, 10 (1), 183–199.
(Semantic Scholar) doi: [link] (InSpire)
Brin, S., Larry Page, R. M., and Winograd, T. (1999). Hooydonk, G. V. (1997). Fractional counting of mul-
The PageRank citation ranking: Bringing order to tiauthored publications: Consequences for the impact
the web. Stanford InfoLab, (1999). (Semantic of authors. JASIS , 48 (10), 944–945. (Semantic
Scholar) Scholar) doi: [link]
Brin, S. and Page, L. (1998). The anatomy of a large- Ivanov, R. and Raae, L. (2010). Inspire: A new scien-
scale hypertextual web search engine. Comput. Netw., tific information system for hep. J. Phys. Conf. Ser.,
30 (1-7), 107-117. (Semantic Scholar) doi: [link] 219 . (InSpire) doi: [link]
Carbone, V. (2011). Fractional counting of author- Klem, J. and Iwaszkiewicz, J. (2011). Physicists
ship to quantify scientific research output. CoRR, get inspired: Inspire project and grid applications.
abs/1106.0114 . (Semantic Scholar) J. Phys. Conf. Ser., 331 . (InSpire) doi: [link]
Chen, P., Xie, H., Maslov, S., and Redner, S. (2007).
Kurtz, M. J. (2017). Comparing People with Biblio-
Finding scientific gems with google’s pagerank al-
metrics.
gorithm. J. Informetrics, 1 , 8-15. (Semantic
Scholar) doi: [link] (Semantic Scholar)
del Corso, G. M. and Romani, F. (2009). Versatile Kurtz, M. J. and Henneken, E. A. (2017). Measur-
weighting strategies for a citation-based research eval- ing metrics - a 40-year longitudinal cross-validation
uation model. Bull. Belg. Math. Soc. - Simon Stevin, of citations, downloads, and peer review in astro-
16 (4), 723–743. (zbMATH) physics. Journal of the American Society for Infor-
mation Science and Technology, 68 , 695-708. (Se-
del Corso, G. M. and Romani, F. (2016). A multi-
mantic Scholar) doi: [link]
class approach for ranking graph nodes: models and
experiments with incomplete data. Inf. Sci., 329 , 619- Lehmann, S., Lautrup, B., and Jackson, A.
637. (Semantic Scholar) (2003). Citation distributions in high-energy physics.
de Sacrobosco, J. (1230). Tractatus de sphaera. Phys. Rev. E , 68 . (InSpire) doi: [link]
(InSpire) Leydesdorff, L. and Bornmann, L. (2010). How
Ding, Y., Yan, E., Frazho, A. R., and Caverlee, J. fractional counting affects the impact factor: Steps
(2009). Pagerank for ranking authors in co-citation towards field-independent classifications of scholarly
networks. JASIST , 60 (11), 2229–2243. (Semantic journals and literature. CoRR, abs/1007.4749 . (Se-
Scholar) doi: [link] mantic Scholar)
Egghe, L. (2008). Mathematical theory of the h- and Leydesdorff, L. and Bornmann, L. (2011). How frac-
g-index in case of fractional counting of authorship. tional counting of citations affects the impact factor:
JASIST , 59 (10), 1608–1616. (Semantic Scholar) Normalization in terms of differences in citation po-
doi: [link] tentials among fields of science. JASIST , 62 (2), 217–
229. (Semantic Scholar) doi: [link]
Golosovsky, M. and Solomon, S. (2017). Grow-
ing complex network of citations of scientific pa- Leydesdorff, L. and Opthof, T. (2010). Normaliza-
pers - measurements and modeling. Phys. Rev. E , tion at the field level: fractional counting of citations.
95 (012324). (Semantic Scholar) doi: [link] J. Informetrics, 4 , 644-646. (Semantic Scholar)
40
Leydesdorff, L. and Park, H. W. (2017). Full and Rousseau, R. (2014). A note on the interpolated
fractional counting in bibliometric networks. J. In- or real-valued h-index with a generalization for frac-
formetrics, 11 (1), 117–120. (Semantic Scholar) tional counting. Aslib J. Inf. Manag., 66 (1), 2–12.
doi: [link] (Semantic Scholar) doi: [link]
Leydesdorff, L. and Shin, J. C. (2011). How to evalu- Sinatra, R., Deville, P., Szell, M., Wang, D., and
ate universities in terms of their relative citation im- Barabási, A.-L. (2015). A century of physics. Nature
pacts: Fractional counting of citations and the nor- Physics, 11 (10), 791–796. (Semantic Scholar) doi:
malization of differences among disciplines. JASIST , [link]
62 (6), 1146–1155. (Semantic Scholar) doi: [link]
Tanase, R. and Radu, R. (2009). The Mathematica
Ma, N., Guan, J., and Zhao, Y. (2008). Bringing of Web Search.
pagerank to the citation analysis. Inf. Process. Man-
age., 44 (2), 800–810. (Semantic Scholar) doi: Thelwall, M. and Wilson, P. (2014). Regression
[link] for citation data: An evaluation of different meth-
ods. J. Informetrics, 8 (4), 963 - 971. (Semantic
Martin Montull, J. (2011). Inspire: Managing meta-
Scholar) doi: [link]
data in a global digital library for high-energy physics.
Commun. Comput. Info. Sci., 240 , 269-274. (In- Walker, D., Xie, H., Yan, K.-K., and Maslov, S.
Spire) doi: [link] (2007). Ranking scientific publications using a sim-
ple model of network traffic. J. Stat. Mech., 0706 .
Perianes-Rodrı́guez, A., Waltman, L., and van Eck,
(Semantic Scholar) doi: [link]
N. J. (2016). Constructing bibliometric networks:
A comparison between full and fractional count- Wang, D., Song, C., and Barabási, A.-L. (2013).
ing. J. Informetrics, 10 (4), 1178–1195. (Semantic Quantifying long-term scientific impact. Science,
Scholar) doi: [link] 342 (6154), 127–132. (Semantic Scholar) doi:
Pinski, G. and Narin, F. (1976). Citation influ- [link]
ence for journal aggregates of scientific publications: West, J. D., Jensen, M. C., Dandrea, R. J., Gor-
Theory, with application to the literature of physics. don, G. J., and Bergstrom, C. T. (2013). Author-
Inf. Process. Manage., 12 (5), 297–312. (Semantic level eigenfactor metrics: Evaluating the influence of
Scholar) doi: [link] authors, institutions, and countries within the social
Radicchi, F., Fortunato, S., Markines, B., and Vespig- science research network community. JASIST , 64 (4),
nani, A. (2009). Diffusion of scientific credits and the 787–801. (Semantic Scholar) doi: [link]
ranking of scientists. Phys. Rev. E , 80 , 056103. (Se-
Zhou, D., Orshanskiy, S. A., Zha, H., and Giles, C. L.
mantic Scholar) doi: [link]
(2007). Co-ranking authors and documents in a het-
Rajaraman, A. and Ullman, J. D. (2009). Mining of erogeneous network. Proceedings of the 7th IEEE
massive datasets. Cambridge University Press. (Se- International Conference on Data Mining (ICDM
mantic Scholar) doi: [link] 2007), 739–744. (Semantic Scholar) doi: [link]
41

Biblioranking Fundamental Physics (Updated To 2021/1/1) : CERN-TH-2018-066

Uploaded by

Copyright:

Available Formats

Biblioranking Fundamental Physics (Updated To 2021/1/1) : CERN-TH-2018-066

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Biblioranking Fundamental Physics (Updated To 2021/1/1) : CERN-TH-2018-066

Uploaded by

Copyright:

Available Formats

CERN-TH-2018-066

Biblioranking fundamental physics

Alessandro Strumiaa , Riccardo Torreb,c

A The InSpire and arXiv databases 34

4. Number of individual citations

2.2 PaperRank of papers: results

Table 1: Top-cited (highest number of citations) papers in the InSpire database.

Title 1st author Naut date Ncit Rp Rp

Table 2: Top-ranked (highest PaperRank) papers in the InSpire database.

Title 1st author Naut date Ncit Rp Rp

Table 3: Top-referred (highest AuthorRank) papers in the InSpire database.

2.3 PaperRank as the number of citations-of-citations

Basically, Rp counts the number of citations-of-citations up to generation g. In the above

��� ����������� ��� ���������� ������� ��� �

2.4 Top-referred (recent) papers

�� ���� ��� ���� ���� ���� ���� ����

ℛ� ������ ���� ���� ���� ����

������ ���� ���� ����

dominated by notable authors citing themselves.

2.5 Paper metrics: correlations

3.1 Sharing among co-authors: fractional counting

As discussed in Section 2.1 we consider ℘ = 0.99.

3.2 Fractional counting and collaborations

������ ����� ������ �� ���������� ���������

3.3 PaperRank of authors: results

3.4 Author Rank

Table 10: Authors sorted according to their Citation-coin C

3.5 Removing self-citations and citation ‘cartels’: the Citation-coin

3.6 Author metrics: correlations

4.1 Ranking institutions

����� ���������� �� ������

contributed to fundamental physics can be found at IAS in Princeton, or at CERN.

4.2 Ranking towns

4.3 Ranking countries and continents

4.4 Ranking journals

��� ���-�� + ����-�� ���-�� + ���-�� + ���-��� + ����-�� �����-�� + ��-��

������ ������������ ���� ������������ ������� ������

Figure 8: Percentage impact by continent.

by self-citations and circular citations, strongly penalizes young authors, we propose an

Note added: 2021 update

A The InSpire and arXiv databases

A.1 Main trends in the fundamental physics literature

������� ��� �� ���������� �� �����

A.2 Details about the dataset

Mean number of authors per paper

Fraction of published papers

Mean number of citations per paper

Mean individual citations per paper

tin Montull, 2011).

��� ��� ��� ���

��� ��� ��� ���

����� ��� ����� ����

����� ��� ����� ����

����� ��� ����� ����

��� ��� ��� ���

��� ��� ��� ���

����-�� ���-��� ����-�� �����-��

��� ��� ��� ���

��� ��� ��� ���

��

��

ℛ� ��

��

��

��

�� -�� + ��-�� -�� + ��-�� + ��-�� + ��-�� -�� + ��-��

��

��

��

��

��

��

��

��

��

��-�� -�� -�� -��

��

��

��

��

��

��

��