Oscar Sheynin

Theory of Probability. A Historical Essay

Revised and Enlarged Edition

Berlin
2017

1
Contents
Preface
0. Introduction
0.1. The Stages
0.2. Mathematical Statistics
0.3. The Theory of Errors
0.4. The Statistical Method
0.5. Some nasty facts
1. The Antenatal Period
1.1. Randomness, Probability Expectation
1.2. Mathematical Treatment of Observations
2. The Early History
2.1. Stochastic Ideas in Science and Society
2.2. Mathematical Investigations
3. Jakob Bernoulli and the Law of Large Numbers
3.1. Bernoulli’s Works
3.2. The Art of Conjecturing (1713),
Part 4: Its Main Propositions
3.3. Bernoulli’s Contemporaries
4. De Moivre
and the De Moivre – Laplace Limit Theorem
4.1. The Measurement of Chance (1712)
4.2. Life Insurance
4.3. The Doctrine of Chances (1718, 1738, 1756)
4.4. The De Moivre – Laplace Theorem
5. Bayes
5.1. The Bayes Formula and Induction
5.2. The Limit Theorem
5.3. Additional Remark
6. Other Investigations before Laplace
6.1 Stochastic Investigations
6.2. Statistical Investigations
6.3. Mathematical Treatment of Observations
7. Laplace
7.1. Theory of Probability
7.2. Theory of Errors
7.3. Philosophical Views
7.4. Conclusions
8. Poisson
8.1. Subjective Probability
8.2. Two New Notions
8.3. The De Moivre – Laplace Limit Theorem
8.4. Sampling without Replacement
8.5. Limit Theorems for Poisson Trials
8.6. The Central Limit Theorem
8.7. The Law of Large Numbers
8.8. The Theory of Errors and Artillery
8.9. Statistics
9. Gauss, Helmert, Bessel
9A. Gauss
9A.1. The Method of Least Squares before 1809
9A.2. Theoria Motus (1809)
9A.3. Determining the Precision of Observations (1816)
9A.4. The Theory of Combinations (1823 – 1828)
9A.5. Additional Considerations
9A.6. More about the Method of Least Squares
9B. Helmert
9C. Bessel

2
10. The Second Half of the 19th Century
10.1. Cauchy
10.2. Bienaymé
10.3. Cournot
10.4. Buniakovsky
10.5. Quetelet
10.6. Galton
10.7. Statistics
10.8. Statistics and Natural Sciences
10.9. Natural Scientists
11. Bertrand and Poincaré
11.1.Bertrand
11.2. Poincaré
12. Geometric Probability
13. Chebyshev
13.1. His Contributions
13.2. His Lectures
13.3. Some General Considerations
14. Markov, Liapunov, Nekrasov
14.1. Markov: General Scientific Issues
14.2. Markov: His Main Investigations
14.3. Markov: His Personal Traits
14.4. Liapunov
14.5. Nekrasov
15. The Birth of Mathematical Statistics
15.1. The Stability of Statistical Series
15.2. The Biometric School
15.3. The Merging of the Continental Direction
and the Biometric School?
Supplement
References
Index of Names

3
3
 Annotation
This book covers the history of probability up to Kolmogorov with
essential additional coverage of statistics up to Fisher. Based on my
work of ca. 50 years, it is the only suchlike book. Gorrochurn (2016)
is similar but his study of events preceding Laplace is absolutely
unsatisfactory. Hald (1990; 1998) are worthy indeed but the
Continental direction of statistics (Russian and German statisticians) is
omitted, it is impossible to find out what was contained in any
particular memoir of Laplace and the explanation does not always
explain the path from, say, Poisson to a modern interpretation of his
results. Finally, the reader ought to master modern math. statistics.
I included many barely known facts and conclusions, e. g., Gauss’
justification of least squares (yes!), the merits of Bayes (again, yes!),
the unforgivable mistake of Laplace, the work of Chebyshev and his
students (merits and failures) etc., etc.
The book covers an extremely wide field, and is targeted at the
same readers as any other book on history of science. Mathematical
treatment is not as difficult as it is for readers of Hald.

4
4
 Preface

To disregard bygone years and only

kow-tow to the present is mean,
barbarous and ignorant
A. S. Pushkin

I do feel how wrongful it was to work for so

many years at statistics and neglect its history
K. Pearson (1978, p. 1)

I have attempted to include everything essentially interesting. A

historical essay such as this one can (and hopefully will) be also used
for methodological purposes, so that the narrative should not be dry,
but it is not for me to determine whether my book is now suitable
enough for a comparatively broad circle of readers.
The book is intended for those interested in the history of
mathematics or statistics and more or less acquainted with the latter. It
will also be useful for statisticians. My exposition is based, in the first
place, on my own investigations published over some 50 years. True, I
am not satisfied with a few of them anymore. Note also, that I was
unable to check the proofs of some of my papers which are therefore
corrupted by misprints. I bear in mind a manuscript that was smuggled
out of the Soviet Union (Sheynin 1989a) as well as my Russian
articles in the Istoriko-Matematicheskie Issledovania from 1993
onward whose proofs I never saw. A good few years ago this journal,
the only Russian outlet for papers on history of mathematics, was
killed by the transformers of the Academy of Sciences.
I describe the origin of the notions of randomness and subjective or
logical probability in antiquity, discuss how laymen comprehended
the main notions of the theory of probability, dwell on the birth of
political arithmetic and study the history of the theory of probability
proper. I also trace the development of statistics and its penetration
into natural sciences as well as the history of the mathematical
treatment of observations (Ptolemy, Al-Biruni, Kepler, the classical
error theory). I stop at the axiomatization of probability and at the
birth of the real mathematical statistics, i.e., at Kolmogorov and
Fisher.
From adjoining general sources1 written from a modern point of
view, I mention Stigler (1986), Hald (1990; 1998) and Farebrother
(1999). The first of these, in spite of its title, only dwells on separate
chapters of the history of statistics and is utterly corrupted by
slandering the memory of Euler and Gauss. And since it was
universally hailed with no one defending those giants, it showed that
the scientific community was (and is) seriously ill.
The next two books are specimens of an exposition of a
mathematical subject, but they are intended for really qualified
readers; then, some topics in Hald (1998), especially the description
of the work of Russian mathematicians, are omitted and the exposition
is not always connected with specific contributions; thus, it is difficult
to see what had Laplace to say in any of his early memoirs. Finally,

5
5
Farebrother’s book dwells on the treatment of observations. My own
booklet (2006/2009) is a collection of 832 short dictums, – of
pronouncements made by great many scholars over the centuries on
probability, statistics, theory of errors and randomness. I see it as a
very important scientific and methodological supplement to
“ordinary” treatises.
During the last years, quite a few worthless or mediocre
contributions to my subject have appeared which was apparently made
possible by unsatisfactory preliminary reviewing (and then justified by
subsequent superficial abstracting). I do not mention such literature
and I also note that in 1915 the Petersburg Academy of Sciences
awarded a gold medal to Chuprov for a review written on its request
(Sheynin 1990c/2011, p. 50). Then, I quote Truesdell (1984, p. 292):

By definition, now, there is no learning, because truth is dismissed

as an old-fashioned superstition. Instead […] there is perpetual
‘research’ on anything and everything. In virtue of the Parkinson’s
law, the professional historian must keep on publishing. Whiteside’s
monument to Newton, like Wren’s masterpiece for St. Paul, will soon
be hidden by towering concrete hives of new bureaus and new slums.

The general situation is bad. For example, it is difficult to publish

an honest critical review of a new book. Periodicals receive free
copies of them for reviewing from the publishers, and Editors are
therefore obliged only to approve sweet nothings. As an example, I
advise readers to compare my reviews of two books (Sheynin 2006c;
2006d) with any other of their published reviews2. See also § 0.5.
With sincere gratitude I recall the late Professors Youshkevitch,
who was always favourably disposed towards me, and Truesdell, the
Editor of the Archive for History of Exact Sciences, who had to busy
himself with my English and compelled me to pay due attention to
style. In 1991, after moving to Germany, I became able to continue
my work largely because of Professor Pfanzagl’s warm support. In
particular, he secured a grant for me (which regrettably dried up long
ago) from Axel-Springer Verlag. In my papers, I had acknowledged
the help of many colleagues including the late Doctors Chirikov (an
able mathematician whose bad health thwarted his scientific career)
and Eisenhart.
The reader should bear in mind that even Markov did not always
distinguish between strict and non-strict inequalities. A second similar
fact is that the distinction between a magnitude sought (for example,
an unknown constant) and its statistical estimate had not been
explicitly indicated until perhaps the end of the 19th century (and still
later in the Biometric school). Then, the expression such as P (x = m),
used for example by Laplace, should be understood as P(m < x < m +
dm) with one or both inequalities being possibly non-strict.
I am using the following abbreviations: CLT – central limit
theorem; LLN – law of large numbers; and MLSq – method of least
squares. W-i means Gauss, Werke, Bd. i (reprint: Hildesheim, 1973 –
1981) and W/Erg-i is Gauss, Werke, Ergänzungsreihe, Bd. i, the
reprint of the correspondence of Gauss (Hildesheim, 1975 – 1987).

6
6
Bände 1, 3, 4 and 5 are his correspondences with Bessel, Gerling,
Olbers and Schumacher respectively. Finally notation S, G, i means
that the source in question is available as a downloadable file i on my
website www.sheynin.de which is being copied by Google (Google,
Oscar Sheynin, Home). I attach this notation to a source if I have
provided its English translation from other languages or if that source
is rare.
When describing the contributions of previous years and centuries I
sometimes use modern terms but indicate them in square brackets.
Thus, [probability] implies that the appropriate author had not applied
that expression.
I have gathered the Notes at the end of the chapter in question. I am
mentioning many Russian sources; some of them, translated by myself
into English. A double date provided in a reference indicates the dates
of both the original and the later edition to which I refer.
I have managed to publish abroad a substantial number of papers
while still living under a dictatorial regime in Moscow, and the
difficulties which I had to overcome have been unknown to the
external world. Without Youshkevitch they would have been
insurmountable and in any case in addition to English texts (even of
the reviews for Zentralblatt MATH) I had to submit their Russian
versions for the censors. Add to this all the humiliation meted out to a
Jew, and you will begin to understand my former situation. I described
it in my Russian autobiography, see my site.

Notes
1. Since I also dwell on population statistics, I ought to mention J. & M.
Dupâquier (1985). Among other issues, they describe the history of national and
international societies and institutions.
2. Here is a typical case. In 1998 Desrosières stated that Poisson had formulated
the strong LLN, and Gauss had derived the normal distribution as the limit of the
binomial (see my review in Isis, vol. 92, 2001, pp. 184 – 185) whereas Stigler (1999,
p. 52) called him a first-class scholar.

7
7
 0. Introduction
0.1. The Stages
Kolmogorov (1947, p. 54) “tentatively” separated the history of
probability into four stages: the creation of its “elements” (from
Pascal and Fermat to Jakob Bernoulli); the 18th, and the
commencement of the 19th century (from De Moivre to Poisson); the
second half of the 19th century (Chebyshev, Markov; Liapunov and
the origin of mathematical statistics); and the beginning of the 20th
century. Gnedenko (1958) and Prokhorov & Sevastianov (1999)
offered, roughly speaking, the same pattern and connected the fourth
period with the introduction of the ideas and methods of the set theory
and the theory of functions of a real variable.
I stress two points. First, I think that there existed an initial version
of the theory of probability whose acme were the LLN, the De
Moivre – Laplace theorem (in essence proved by the former), and the
inverse Bayes theorem (§ 5.2). Second, the modern stage of the
theory, considered up to Kolmogorov, began with Chebyshev, and
this fact should be here more clearly reflected. And so, my pattern of
the history of probability is as follows.
1. Its antenatal period (from Aristotle to the mid-17th century).
2. The creation of its initial version (finally achieved by Jakob
Bernoulli, De Moivre and Bayes).
3. Its development as an applied mathematical discipline (from
Bayes to Laplace and Poisson to Chebyshev).
4. A rigorous proof of its limit theorems (Chebyshev, Markov,
Liapunov) and its gradual transition to the realm of pure mathematics.
5. Axiomatization.
Laplace reasonably transferred the theory of probability from pure
(as understood by his predecessors) to applied mathematics and
Poisson and Poincaré followed suit. During the third, and partly the
fourth stage, mathematicians barely regarded the theory of probability
as a serious scientific discipline. And even now they barely recognize
the theory of errors.
0.2. Mathematical Statistics
Its separation from probability or from statistics in general is
difficult. It originated in the early years of the 20th century as the result
of the work of the Biometric school and the Continental direction of
statistics. Its aim is the systematization, processing and utilizing
statistical data (Kolmogorov & Prokhorov 1988/1990, p. 138). They
added a definition of statistical data:

“Information on the number of objects in some more or less

extensive collection that have some specific properties”.

They apparently excluded the theory of errors and it is unclear

whether they meant raw or corrected (by exploratory data analysis)
information.
Theoretical statistics is wider since it additionally studies the
collection and that same analysis of data.

8
8
 For Pearson, statistics remained an applied mathematical discipline
whereas Fisher managed to create its theory, mathematical statistics.
0.3. The Theory of Errors
From its origin in the mid-18th century and until the 1920s the
stochastic theory of errors had been a most important chapter of
probability theory. Not without reason had P. Lévy (1925, p. vii)
maintained that without it his main work on stable laws of distribution
would have no raison d’être1. Actually, for the theory of errors that
book was meaningless (Sheynin 1995c) and it is incomprehensible
why he had not noticed it.
In turn, mathematical statistics borrowed its principles of maximum
likelihood and minimal variance from the error theory. Today, the
stochastic theory of errors is the application of the statistical method
to the treatment of observations2.
The history of the theory of errors has its own stages. In ancient
times, astronomers were dealing with observations as they saw fit. At
the second stage, beginning perhaps with Tycho Brahe, observations
ceased to be “private property”, but their treatment was not yet
corroborated by quantitative considerations. This happened during the
third stage (Simpson, Lambert), and the final, fourth stage was the
completion of the classical theory of errors (Laplace and especially
Gauss) although later Helmert fruitfully continued the relevant
investigations.
The main peculiarity of the error theory is the usage of the notion of
real (true) value of the constant sought, see § 6.3.3 and Sheynin
(2007a), and I emphasize that it is really needed in experimental
science rather than being outdated and on its way out. Fourier
(1826/1890, p. 534) defined it as the limit of the arithmetic mean,
which, incidentally, provides a new dimension to the Gaussian
postulate of the mean [an expression due to Bertrand (1888a, p.
176)], see § 9A.2-2, and to the attempts to justify its usage by the
notion of consistency, cf. §§ 9A.4-7, 11.2-8, 13.2-7 and 14.4-2. See
also § 6.3.3. It is a peculiar fact that mathematicians neglected (and
still neglect) the reasonable basis of Gauss’ attitude and preferred
(prefer) the hardly useful Laplacean approach.
0.4. The Statistical Method
It might be thought that statistics and statistical method are
equivalent notions; it is usual, nevertheless, to apply the former term
when studying population and to use the latter in all other instances
and especially when applying statistics to natural sciences. However,
there also exist such expressions as medical and stellar statistics, and,
to recall, theory of errors (§ 0.3).
I understand that, respectively, they are applications of the
statistical method to medicine, stellar astronomy and treatment of
observations. This explanation is in line with Pearson’s statement
(1892, p. 15): The unity of all science consists alone in its method.
And to my mind the statistical method is mathematical (or theoretical)
statistics. It is tempting to suggest that mathematics is the application
of the mathematical method, i. e., of the introduction and study of
systems which possibly bear no relation to reality; a simple example:
complex numbers.

9
9
 Three stages might be distinguished in the history of the statistical
method. At first, conclusions were being based on (statistically)
noticed qualitative regularities, a practice which conformed to the
qualitative essence of ancient science. Here, for example, is the
statement of the Roman scholar Celsus (1935, p. 19):

Careful men noted what generally answered the better, and then
began to prescribe the same for their patients. Thus sprang up the Art
of medicine.

The second stage (Tycho in astronomy, Graunt in demography and

medical statistics) was distinguished by the availability of statistical
data. Scientists had then been arriving at important conclusions either
by means of simple stochastic ideas and methods or even directly, as
before. During the present stage, which dates back to the end of the
19th century, inferences are being checked by quantitative stochastic
rules.
0.5. Some Nasty Facts
I briefly discuss some facts concerning my subject, but I will hardly
err when connecting them with history of mathematics or perhaps with
history of science as well.
1. Disdainful attitude towards information. This is contrary to the
proper statement of Shaw & Austin (§ 10.8.3) and mostly due to the
influence from beyond. The scientologists (no connection with
religion) who vainly attempt to estimate the worth of scientists by
numerical measurements apparently do not consider reviewing or
abstracting as real scientific work. The scientific rat race (publish or
perish!) therefore prevents careful and honest reviewing. The situation
with respect to reviewing is horrible and the more so since for most
researchers the main abstracting journals are now out of reach.
2. The same rat race hinders honest work in general.
3. Standardization, or the sledgehammer law engenders robots and
is the curse on science. No one ever compelled Chekhov to write just
like Tolstoy did or required Agatha Christie to imitate Edgar Poe.
Science is however standardized by a sledgehammer. The format of
the presentation of papers must be standardized however different are
their aims, styles or lengths. And why? To justify the very existence of
those scientologists? A most simple example: In a manuscript, I
referred both to K. and E. S. Pearson, but a worker of the editorial
staff deleted the initials! Spelling of names is required to disregard the
rights of authors or their publications abroad. An eminent Russian
author, Bernstein published many notes in France, signed his name as
I mentioned. At the very least, it had thus become his pen name, but
present authors are required to mention its ugly version, Bernshtein.
And now other aspects of the sledgehammer law: periodicals do not
publish translations, period! In the West, Russian literature is badly
known. If a journal has a thousand readers (which is more than
generous) only one or two will be able to come across a reference to
an important Russian source, get hold of it and understand something.
One or two out of a thousand! Crass robotic stupidity.

10
10
 Finally, no one will ever know how many worthy contributions have
not been published since their authors were unable to overcome the
sledgehammer law! See also my Preface.
The late professor Truesdell edited 49 volumes of his prestigious
periodical, Archive for History of Exact Science without standardizing
anything, and nothing bad had happened, but after his death that
journal was promptly driven into the general fold.
I have described all this in detail (Sheynin 2017) and insistently
recommend the readers to see it. My epigraph was a phrase from
Shakespeare: Something is rotten in the state of Denmark.

Notes
1. In 1887 Chebyshev (§ 13.1-4) indicated that the CLT can substantiate the
MLSq which had nothing to do with the Gaussian approach and Poincaré (§ 11.2-
8), in the last years of his life, stated that the theory of errors had naturellement been
his main aim in probability. However, his contribution to that theory is barely
significant.
2. I especially note its application to metrology (Ku 1969). There also existed a
determinate branch of the error theory, unconnected with stochastic considerations.
Here is one of its problems: Compile such a program for measuring angles in the
field that the unavoidable errors, both systematic and random, will least corrupt the
final results (the coordinates of the stations of the trigonometric network). Another
problem: bearing in mind the same goal, determine the optimal form of a triangle of
triangulation. It is opportune to mention Cotes (1722) who solved 28 pertinent
problems concerning plane and spherical triangles with various sets of measured
elements. I leave such problems aside although, in principle, they can be included in
the province of the design of experiments. The determinate error theory is also
related to the exploratory data analysis that aims at uncovering the underlying
structures (e.g., the systematic errors). Or, rather, that branch of the error theory was
swallowed by both of those new disciplines.

11
11
 1. The Antenatal Stage
1.1. Randomness, Probability, Expectation
Is an infinite (much more difficult: finite) sequence random or not?
This is a fundamental problem which I leave aside. Chaitin (1975) is a
popular discussion of the efforts to answer this question. The role of
randomness (for example, in studying the evolution of species or in
the kinetic theory of gases) is obvious. In statistics, a random variable
ought to be stable (best of all, possess an invariable law of
distribution) which is not necessary in natural science1. Thus,
Kolmogorov (1983/1992, p. 515) stated:

We should distinguish randomness in the wider sense (absence of

any regularity) and stochastic ransom events (which constitute the
subject of probability theory).

1.1.1. Aristotle. Ancient scholars repeatedly mentioned randomness

and (logical or subjective) probability2. The first notion implied lack
of aim or law, such as a discovery of a buried treasure (Aristotle,
Metaphys. 1025a) or a sudden meeting of two persons known to each
other (Aristotle, Phys. 196b 30). In the second example randomness
might be interpreted as an intersection of two chains of determinate
events3 and in both cases a small change in the action(s) of those
involved (in these chains) would have led to an essential change of the
result. Thus, the treasure would have remained hidden, the meeting
would not have taken place, cf. § 11.2-9 for a link with modernity;
true, randomness is not anymore connected with lack of intention.
In each of these illustrations the sudden event could have been
(although was not) aimed at; Aristotle would not have called random a
meeting with a stranger, or a discovery of a rusty nail. I would have
asked: Suppose that a venomous snake bites a man in his own kitchen,
should have Aristotle called such an event random?
The examples above also mean that randomness is a possibility, and
Aristotle (Methaphys. 1064b – 1065a) indeed said so. It was Hegel
(1812/1978, pp. 383 – 384) who formulated the converse statement.
Suppose that a discrete random variable takes values xi, i = 1, 2, …, n,
with certain probabilities. Then, according to Hegel, any xi itself is
random.
Aristotle’s special example (Phys. 199b 1; also see De generatione
animalium 767b5) considered deviations from law, monstrosities. The
first departure of nature from the type

12
12
 Is that the offspring should become female instead of male; … as it
is possible for the male sometimes not to prevail over the female …

His was the first, not really convincing example of a dialectical

conflict between randomness and necessity4.
Aristotle (De Caelo 283b 1, also see Phys. 196b and Rhetorica
1369a) stated that “The products of chance and fortune are opposed to
what is, or comes to be, always or usually”. Similar assertions are
made about accidents (Metaphys. 1025a, 1065a), chance conjunctions
(Anal. Post. 87b) and coincidences (Parva naturalia 463b). Finally,
Aristotle (Phys. 197b 0, 197b 14 and 197a 5) distinguishes between
chance (which is “in the sphere of moral actions”) and spontaneity
(“in lower animals and in many inanimate objects”).
Junkersfeld (1945) was and possibly remains the author of the best
account of Aristotle’s concept of chance, and she (p. 22) stated that by
chance he meant something that takes place

Occasionally; has the character of an end; is such that it might

have been the object of a natural or of a rational appetite; was not in
fact the object of any appetite but came into being by accident.

The circularity of this definition is actually due to Aristotle.

That, according to Aristotle, chance takes place rarely, is also
indirectly seen in his statement (Metaphys. 1064b 15, 1026b, 1027a;
Ethica Eudemia 1247b) that “none of the traditional sciences busies
itself about the accidental […] but only sophistic […]”. He (Ethica
Eud. 1247a) notes, however, that “chance largely enters” navigation
and strategy, and even that navigation “is as in throwing dice”.
Regrettably, he did not say anything about sophistic. I also note that
neither does the theory of probability consider the accidental, but
rather studies the laws of randomness. In the words of Schiller
(Spaziergang, 1795),

“Der Weise […] sucht das vertraute Gesetz in des Zufalls

grausenden Wundern, sucht den ruhenden Pol in der Erscheinungen
Flucht”.

Plato (Cioffari 1935, p. 30) has a better opinion about navigation:

although

Chance is almost everything in the arts of the […] pilot, and the
physician, and the general, […] yet in a storm there must surely be a
great advantage in having the aid of the pilot’s art.

And here is another example connected with astronomy (De caelo

292a 30 and 289b 22) that shows, incidentally, that games of chance
even then provided examples of stochastic considerations:

Ten thousand Coan throws [whatever that meant] in succession

with the dice are impossible and it is therefore difficult to conceive

13
13
that the pace of each star should be exactly proportioned to the size of
its circle5.

Cicero (Franklin 2001, p. 164) came to a similar conclusion

unconnected, however, with natural sciences.
Aristotle (Anal. Priora 70a 0) also stated that probability

Is a generally approved proposition: what men know to happen or

not to happen, to be or not to be, for the most part thus and thus […],
e. g., “the envious hate”6.

He (Rhetorica 1402a 5; De Poetica 1461b) believed that “it is possible

that improbable things” will happen. Yes, but his was of course only a
qualitative formula. And, while discussing poetry, he (De Poetica
1460a 25) introduced a rudimentary scale of subjective probabilities:
“a likely impossibility is always preferable to an unconvincing
possibility”. Understandably, he (Rhetorica 1376a 19) recommended
the use of probabilities in law courts. He (Problemata 951b 0) also
thought that it was better to acquit a “wrong-doer” than to condemn an
innocent person7 so that the statistical idea about errors of the two
kinds is seen here.
Aristotle several times mentions luck (Metaphys. 1065a; Rhetorica
1361b) and fortune (Magna Moralia 1206b, 1270a) as notions
expressing deviations from reasonable qualitative expectation.
Finally, Aristotle (for example, Ethica Nicomachea 1104a 24)
believed that mean behaviour, moderation possessed optimal
properties. Analogous statements had appeared even earlier in ancient
China; the doctrine of means is attributed to a student of Confucius
(Burov et al 1973, pp. 119 – 140). Again, a similar teaching existed in
the Pythagorean school (Makovelsky 1914, p. 63), and Nicomachus
of Gerasa (1952, p. 820) stated that a perfect number was in the realm
of equality, was a mean between numbers the sum of whose divisors
was less, and greater that the number itself; was between excess and
deficiency. In medicine (§ 6.2) the mean was considered as the ideal
state (of health), while in games of chance (§ 5) the (arithmetic) mean
was believed to possess certain stochastic properties.
In the new time, the arithmetic mean became the main estimator of
the constants sought (§ 1.2.4) and has been applied in civil life (§
2.1.2). In addition, it is obviously connected with the appropriate
expectation.
Both Plato and Aristotle, as witnessed by Simplicius (Sambursky
1956, p. 37; he only provided an exact reference on p. 3 of the reprint
of this paper), called natural sciences “the science of the probable
[eikotologia]”. A later scholar, Levi ben Gerson, thought that the
determinism of natural laws was only approximate and probable
(Rabinovitch 1973, p. 77, with a reference to Levi’s work), and,
similarly (Ibidem, p. 166), Maimonides held that natural philosophy
only offered probable theories.
1.1.2. The Bible and the Talmud (Sheynin 1998b). The earlier part
of the Talmud is an interpretation of the first five books of the Old
Testament. Called Mishna, it is subdivided into more than 60 treatises.

14
14
The other part of the Talmud is made up of later commentaries on the
Mishna. The Jerusalem Talmud was essentially completed in the
fourth century and precedes the more influential Babylonian Talmud
by about a century. I have seen the English edition of the Babylonian
version (six volumes; London, 1951 – 1955) and I also refer to the
German edition of the Talmud (12 vols, Berlin, 1930 – 1936) and the
French edition of the Jerusalem Talmud (six volumes; Paris, 1960).
When referring to the Talmud, I abbreviate it; thus, T/Avoth means
treatise Avoth from the Talmud.
My reasoning is not linked with religious belief, but it might be
emphasized that the sacred texts represented the general feelings of
ancient communities. Therefore, any statements there, which
complement and sometimes precede Aristotle’s teaching, characterize
the contemporaneous knowledge stored by mankind.
Ancient commentators of the Mishna hardly set high store by
natural science or mathematics. Thus, Rabbi Elieser ben Chisma
(T/Avoth318) argued that the laws concerning

Bird offerings and the onset of menstruation are essential

traditional ordinances, but the calculation of the seasons and
geometry are but the after-course of wisdom.

Maimonides (1977, pp. 118 – 129), however, was a later exception:

Our sages confirmed [that mathematical astronomy was] the true

wisdom in the sight of the people. But the theories of the astrologists
are devoid of any value.

It is generally known that theology had hindered the advancement

of science as well as the introduction of preventive measures against
deadly small-pox (§ 6.2.3). On the other hand, the most celebrated
scholars including Kepler and Newton had been inspired by wishing
to discover Divine laws of nature. Scientists believed that those laws
directly showed the acts of God whereas the Bible was, after all,
written by mortals.
In any case theology demanded logical thought. Thus, the Old
Testament (Proverbs 14:28) contains a direct and a contrary
proposition: “In a multitude of people is the glory of a king, but
without people a prince is ruined”. See also Matthew 13:34, but the
example above is interesting also in that it could have well been
formulated in the mid-17th century by John Graunt or William Petty,
the fathers of political arithmetic, or by Johann Süssmilch in the next
century. Consider also the questions listed by Moses (Numbers 13: 17
– 20) which also provide a link with political arithmetic. He sent spies
to the land of Canaan to find out

Whether the people who dwell in it are strong or weak, whether

they are few or many, […] whether the land is rich or poor […].

Several authors preceded me. Hasofer (1967) studied the use of lots
as described in the Talmud and Rabinovitch (1973), drawing on his

15
15
earlier papers, dwelt on stochastic considerations in the same source.
He aptly refuted an author of the Encyclpaedia Hebraica (Tel Aviv,
1962, XIV, pp. 920 – 921) who had maintained that

Ancient Jewish thought did not know [did not make use of] the
concept of probability.

I am greatly indebted to Rabinovich, but am not satisfied either with

his selection of cases or explanations. He discusses marginal
problems, finds non-existing subjects (axioms of probability theory)
or makes too much of some other topics (law of large numbers,
sampling).
I am adding other important facts and, unless stated otherwise,
offering my own comments.
Randomness is mentioned several times in the Old Testament:

By chance I happened to be on mount Gilbo’a (2 Samuel 1:6)

Now there happened to be there a worthless fellow (2 Samuel 20:1)
A certain man drew his bow at a venture and struck the King of
Israel (1 Kings 22:34, 2 Chronicles 18:33)

Here, and in the other cases, randomness implicitly meant lack of

purpose, cf. § 1.1.1. Ecclesiastes 9:11, however, states much more:
Time and chance determine the fate of man.
The Talmud makes indirect use of probability (Rabinovitch 1973,
Chapter 4). Thus, in certain cases prohibited fruit could not have
exceeded 1/101 of its general quantity. Under less rigid demands it
was apparently held that only the sign of the deviation from 1/2 was
essential (Ibidem, p. 45).
Rabinovitch (1973, p. 44) noted that in the absence of a “clear
majority” the Talmud regards doubts as half and half which is
certainly only valid for subjective probabilities8. He refers to two
tracts but in one of them (Makhshirin 23–11) the statement is rather that
the majority is equivalent to the whole. And at least in one case that
principle seems to have been misapplied (Rabinovitch p. 45):

If nine shops sell ritually slaughtered meat, and one sells meat that
is not ritually slaughtered [obviously, meat of cattle, twice] and he
[and someone] bought in one of them and does not know which one –
it is prohibited because of the doubt; but if meat was found [in the
street] one goes after the majority.

Another example (Rabinovitch p. 40) obviously shows that

subjective probability can lead to sophisms, see § 8.1 and Sheynin
(2002b). I bear in mind the opinion of Rabbi Shlomo ben Adret (1235
– 1310), whom Rabinovitch mentions time and time again. There are
several pieces of meat, all of them kosher except one. Eating the first
one is allowed, since it is likely kosher; the same with the second one
etc., and when only two pieces of meat are left, the forbidden piece
was likely already eaten and they are also allowed.

16
16
 Both the Bible and the Talmud provide examples of attempts to
distinguish between randomness and causality and to act
accordingly. Genesis 41:1 – 6 discusses cows and ears of corn as seen
by the Pharaoh in two consecutive dreams. The dreams, essentially the
same, differed in form. Both described an event with an extremely low
probability (more precisely, a miracle), and they were thus divine
rather than random. Then, Job (9:24 and 21:17 – 18) decided that the
world was “given over to the wicked” [this being the cause] since the
alternative had a low pseudo-statistical probability: “their lamp was
put out rarely”. This statement possibly brought forward a
commentary (T/Avoth 415) to the effect that we are unable to explain
his conclusion.
A random event of sorts with a rather high probability seems
nevertheless to have been presented in Exodus 21:29: an attack by an
ox will be likely if, and only if, he “has been accustomed to gore”.
And here are several examples from the Talmud. If in three
consecutive days (not all at once, or in four days) three (nine) persons
died in a town “bringing forth” 500 (1500) soldiers, the deaths should
be attributed to a plague, and a state of emergence must be declared
(Taanid 34)9. The probability of death of an inhabitant during three
days was apparently considered equal to 1/2, see above, so that a
random and disregarded death of three people (in the smaller town)
had probability 1/8. The case of all three dying at once was for some
reason left aside. An early commentator, Rabbi Meir, lamely
explained the situation by mentioning the goring ox (the German
edition of the Talmud, Bd. 3, p. 707)10. A similar and simpler example
concerned an amulet (Sabbath 62): for being approved, it should have
healed three patients consecutively. No follow-up checks had been
demanded.
The second example (Leviticus 6:3 – 10) concerned the annual Day
of Atonement when the High Priest brought up two lots, for the Lord
and “for Azazel”, one in each hand, from an urn. During 40 years, the
first lot invariably came up in the Priest’s right hand and that was
regarded as a miracle and ascribed to his special merit [to a cause]. A
special example concerned the redemption of the first born by lot
(Jerus. Talmud/Sangedrin 14). Moses wrote “Levite” on 22, 273
ballots and added 273 more demanding five shekels each. The
interesting point here is that only 22,000 “Levite” ballots were needed
so that Moses ran the risk of losing some of the required money.
Nevertheless, the losing ballots turned up at regular intervals, which
was regarded as a miracle. The existence of the superfluous ballots
was not explained; I believe that the Israelites were afraid that the last
273 of them to draw the lots will be the losers. Such misgivings are,
however, unfounded, see § 8.4 for an explanation based on subjective
probabilities and Tutubalin (1972, p. 12) for a real demonstration.
Rabinovitch (1977, p. 335) provided a similar example, again from
the Talmud. I cite finally the ruling about abandoned infants
(Makhshirin 27). A child, found in a town whose population was
mostly gentile, was supposed to be gentile, and an Israelite otherwise
(also when the two groups were equally numerous).

17
17
 1.1.3. Medicine (Sheynin 1974, pp. 117 – 121). Hippocrates
described a large number of [case histories] containing qualitative
stochastic considerations in the spirit of the later Aristotle (and
ancient science in general). Thus (1952a, pp. 54 – 55): “It is probable
that, by means of […], this patient was cured”; or (1952b, p. 90),

To speak in general terms, all cases of fractured bones are less

dangerous than those in which …

He (see below) also understood that healing depended on causes

randomly changing from one person to another. Then, Hippocrates
formulated qualitative [correlational] considerations, as for example
(1952c, No. 44):

Persons who are naturally very fat are apt to die earlier than those
who are slender.

He (1952c, § 69 p. 116; § 71 p. 117) also states that much depends

on the constitution and general condition of the patient. Thus, in the
second case:

Men’s constitutions differ much from one another as to the facility

or difficulty with which dislocations are reduced […].

Aristotle left similar reasoning, for example (Problemata 892a 0):

“Why is it that fair men and white horses usually have grey eyes?” Or,
noting some dependence between the climate or weather and human
health (Ibidem 859b 5, 860a 5).
Galen also made use of stochastic reasoning. Most interesting is his
remark (1951, p. 202) which I shall recall in § 11.2-9:

… in those who are healthy […] the body does not alter even from
extreme causes; but in old men even the smallest causes produce the
greatest change.

One of his pronouncements (Ibidem, p. 11) might be interpreted as

stating that randomness was irregular: The body has “two sources of
deterioration, one intrinsic and spontaneous, the other [which is]
extrinsic and accidental”, affects the body occasional[ly], irregular[ly]
and not inevitabl[y]. In principle, one of his conclusions (1946, p.
113) is connected with clinical trials:

What is to prevent the medicine which is being tested from having a

given effect on two [on three] hundred people and the reverse effect on
twenty others, and that of the first six people who were seen at first
and on whom the remedy took effect, three belong to the three
hundred and three to the twenty without your being able to know
which three belong to the three hundred, and which to the twenty […]
you must needs wait until you see the seventh and the eighth, or, to put
it shortly, very many people in succession.

18
18
 Galen (1951, Book 1, Chapter 5, p. 13) also thought that a mean
state, a mean constitution were best, cf. § 1.1.1:

Health is a sort of harmony […] all harmony is accomplished and

manifested in a two-fold fashion; first, in coming to perfection […]
and second, in deviating slightly from this absolute perfection […].

There also (pp. 20 – 21 of Book 1, Chapter 6) he says, for example,

that “A good constitution [is] a mean between extremes”.
1.1.4. Astronomy. Astronomers understood that their observations
were imperfect; accordingly, they attempted to determine some
bounds for the constants sought. Thus (Toomer 1974, p. 139), the
establishment of bounds “became a well-known technique […]
practised for instance by Aristarchus, Archimedes and
Erathosthenes”. Here, for example, is Aristarchus(1959, p. 403):

The diameter of the sun has to the diameter of the earth a ratio
greater than [… 19:3] but less than [… 43:6], –

greater than 6.33 and less than 7.17.

This did not however exclude the need to assign point estimates
which was done by taking into account previous data (including
bounds), qualitative considerations and convenience of further
calculations. Modern mathematical statistics teaches that in case of
large errors any observation can be chosen as the final result. Then,
concerning the last-mentioned circumstance Neugebauer (1950, p.
252) remarks:

The ‘doctoring’ of numbers for the sake of easier computation is

evident in innumerable examples of Greek and Babylonian astronomy.
Rounding-off in partial results as well as in important parameters can
be observed frequently often depriving us of any hope of
reconstructing the original data accurately.

And, again (Neugebauer 1975, p. 107):

In all ancient astronomy, direct measurements and theoretical

considerations are […] inextricably intertwined […] ever present
numerical inaccuracies and arbitrary rounding […] repeatedly have
the same order of magnitude as the effects under consideration11.

Babbage (1874) seems to have been the only author in the new
times who paid special attention to various methods of doctoring. He
distinguished hoaxing or forging (downright deceiving), trimming
(leaving the mean intact but “improving” the precision of
measurement) and cooking (selecting observations at will). I return to
errors of observation below.
Theoretical considerations partly replaced measurements in the
Chinese meridian arc of 723 – 726 (Beer et al 1961, p. 26; Needham
1962, p. 42). At the end of their paper the authors suggested that it
was considered more suitable to present harmonious results.

19
19
 Special attention was being paid to the selection of the optimal
conditions for observation; for example, to the determination of time
intervals during which an unavoidable error least influenced the final
result12. Thus (Ptolemy 1984, IX, 2, p. 421, this being the Almagest,
his main contribution): on certain occasions (during stations) the
motion of planets is too small to be observable. No wonder that he
(Ibidem, III 1, p. 137) “abandoned” observations “conducted rather
crudely”. Al-Biruni (1967, pp. 46 – 51), the only Arab scholar to
surpass Ptolemy and to be a worthy forerunner of Galileo and Kepler,
to whom I return below, rejected four indirect observations of the
latitude of a certain town in favour of its single and simple direct
measurement.
Studying this aspect, Aaboe & De Solla Price (1964, pp. 2 and 3)
argued that

In the pre-telescopic era there is […] a curious paradox that even a

well-graduated device [their estimate: its error was 5′] for measuring
celestial angles […] is hardly a match for the naked and unaided eye
judiciously applied […].

The function of smaller antique instruments, as they believe, was to

serve as a means for avoiding calculations while the

Characteristic type of measurement depended not on instrumental

perfection but on the correct choice of crucial phenomena.

They even, and I would say, mistakenly, called antique observations

qualitative.
Hipparchus was aware that, under favourable conditions, a given
error of observation can comparatively little influence the unknown
sought (Toomer 1974, p. 131). And (Neugebauer 1950, p. 250)
Babylonian astronomers of the Seleucid period possessed some
understanding of such phenomena: their lunar and planetary
computations “were based on an exceedingly small number of
observations” and a “very high accuracy” of these observations was
not needed. He continued:

It seems to me one of the most admirable features of ancient

astronomy that all efforts were concentrated upon reducing to a
minimum the influence of the inaccuracy of individual observations
with crude instruments by developing to the farthest possible limits the
mathematical consequences of a very few basic elements.

Elsewhere he (1948, p. 101) argued that in antiquity

Observations were more qualitative than quantitative; when angles

are equal may be decided fairly well on an instrument but not how
large are the angles says Ptolemy with respect to the lunar and solar
diameter[s].

20
20
His reference is not exactly correct; it should have been Almagest (V
14, p. 252; H 417). The wording in the edition of 1984 is somewhat
different.
Regular observations constituted the third main feature of ancient
astronomy. Neugebauer (1975, p. 659) credited Archimedes and
Hipparchus with systematic observation of the apparent diameters of
the sun and the moon. Moreover, otherwise Hipparchus could have
hardly been able to compile his star catalogue. Ptolemy (1984, III, 1,
pp. 132 and 136; IV, 9, p. 206) apparently made regular observations
as well. For example, in the second instance Ptolemy mentions his
“series of observations” of the sun.
Al-Biruni (1967) repeatedly tells us about his own regular
observations, in particular (p. 32) for predicting dangerous landslides
(which was hardly possible; even latitude was determined too crude).
Then, Levi ben Gerson (Goldstein 1985, pp. 29, 93 and 109)
indirectly but strongly recommended them. In the first two cases he
maintained that his regular observations proved that the declination of
the stars and the lunar parallax respectively were poorly known. Thus,
already in those times some contrast between the principle of regular
(and therefore numerous) observations and the selection of the best of
these began to take shape, also see § 1.2.2.
Astronomers undoubtedly knew that some errors, for example those
caused by refraction, acted one-sidedly (e. g., Ptolemy 1984, IX 2); I
also refer to Lloyd (1982, p. 158, n 66) who argues that Ptolemy
“implicitly recognized what we should describe as systematic errors”
and notes that he “has a [special] term for significant or noteworthy
differences”. Nevertheless, the separation into random and systematic
errors occurred only at the end of the 18th century (§ 6.3.1). But, for
example, the end of one of Ptolemy’s statements (1956, III, 2, p. 231)
hints at such a separation:

Practically all other horoscopic instruments […] are frequently

capable of error, the solar instruments by the occasional shifting of
their positions or of their gnomons, and the water clocks by stoppages
and irregularities in the flow of the water from different causes and by
mere chance.

Al-Biruni (1967, pp. 155 – 156) formulated a similar statement

about water clocks.
Here are my conclusions derived from the Almagest:

Ptolemy knew that errors of observation were unavoidable or

almost so.
He mentioned a number of sources of error.
He knew how to minimize the influence of some of them.
He gave thought to the choice of methods of observation.
He knew that some errors acted systematically.

It is possible that, when selecting a point estimate for the constants

sought, ancient astronomers were reasonably choosing almost any
number within the appropriate bounds (see above). Indeed, modern

21
21
notions about treating observations, whose errors possess a “bad”
distribution, justify such an attitude, which, moreover, corresponds
with the qualitative nature of ancient science. Ptolemy’s cartographic
work corroborates my conclusion: he was mainly concerned with
semblance of truth [I would say: with general correctness] rather than
with mathematical consistency (Berggren 1991). A related fact
pertains even to the Middle Ages (De Solla Price 1955, p. 6):

Many medieval maps may well have been made from general
knowledge of the countryside without any sort of measurement or
estimation of the land by the ‘surveyor’.

Many authors maintained that Ptolemy had borrowed observations

from Hipparchus, and, in general, doctored them while R. R. Newton
(1977, p. 379) called him “the most successful fraud in the history of
science”. Yes, he likely borrowed from Hipparchus, but in good faith,
in accordance with the day’s custom. No, he had not doctored any
observations, but rejected, adjusted or incorporated them “as he saw
fit” (Gingerich 1983, p. 151; also see Gingerich 2002), he was an
opportunist ready “to simplify and to fudge” (Wilson 1984, p. 43).
I adduce now three noteworthy statements which partly shaped my
opinion above.
1) Kepler (1609/1992, p. 642/324):

We have hardly anything from Ptolemy that we could not with good
reason call into question prior to its becoming of use to us in arriving
at the requisite degree of accuracy.

2) Laplace (1796/1884, p. 413):

Hipparchus among all ancient astronomers deserves the gratitude

of astronomy for the large number and precision of his observations,
for important conclusions which he had ben able to make by
comparing them with each other and with earlier observations, and
for the witty methods by which he guided himself in his research.
Ptolemy, to whom we are mostly indebted for acquainting us with his
work, had invariably based himself on Hipparchus’ observations and
theories. He justly appraised his predecessor…

Indeed, Ptolemy (1984, IX 2, p. 421; H 210) also see (1984 III 1, p.

136; H 200) called him “a great lover of truth”. And Laplace on the
next page:

His Tables of the Sun, in spite of their imperfection, are a durable

monument to his genius and Ptolemy respected them so much that he
subordinated his own observations to them.

3) Newcomb (1878, p. 20): “… all of Ptolemy’s Almagest seems to

me to breathe an air of perfect sincerity”13.

22
22
 Al-Biruni (1967, p. 152) was the first to consider, although only
qualitatively, the propagation of computational errors and the
combined effect of observational and computational errors:

The use of sines engenders errors which become appreciable if they

are added to errors caused by the use of small instruments, and errors
made by human observers.

One of his statements (Ibidem, p. 155) on the observation of lunar

eclipses for determining the longitudinal difference between two cities
testified to his attempt to exclude systematic influences from final
results: Observers of an eclipse should

Obtain all its times [phases] so that every one of these, in one of the
two towns, can be related to the corresponding time in the other. Also,
from every pair of opposite times, that of the middle of the eclipse
must be obtained.

Such a procedure would have ensured some understanding of the

systematic influences involved, cf. Boscovich’ calculation of a
latitudinal difference in § 6.3.2.
For Al-Biruni, see Al-Khazini (1983, pp. 60 – 62), the arithmetic
mean was not yet the universal estimator of the constants sought;
when measuring the density of metals he made use of the [mode], the
[midrange] as well as of some values situated within the extreme
measurements14, also see Sheynin (1992; 1996a, pp. 21 – 23).
Levi ben Gerson (Goldstein 1985, p. 28) stated that (at least) “all
Ptolemy’s predecessors preferred” to use no instruments at all. He also
distrusted the precision of Ptolemy’s invention, “the instrument of the
rings” as well as astrolabes and quadrants in general. He knew that
small errors of observation can produce large errors in the resulting
position of the stars. Regrettably he (Ibidem, p. 29) attributed this fact
to the “third kind” of errors thus spoiling their classification. And he
even claimed to have constructed and tested an instrument which
allowed to make error-free observations.
Levi (Ibidem, pp. 29, 93 and 109) several times mentions his
regular observations and maintains (Rabinovitch 1974, p. 358) that
repeat observations ought to be made “as often as is required”. This is
interesting but not definite enough. Moreover, the same author (1973,
p. 77) quoted Levi’s belief in the universal existence of uncertainty,
and thus in the impossibility of such an instrument:

Perfect knowledge of a thing is to know it as it is […], that is, to

know that aspect that is determined and bounded and to know also
that indeterminacy which is in it.

1.1.5. Maimonides and Thomas Aquinas. In accordance with the

Talmud, the consumption of some foods was allowed only for priests
and in many other cases the part of the forbidden food should not have
exceeded certain limits, cf. §1.1.2. In this connection Maimonides

23
23
(Rabinovitch 1973, p. 41) listed seven relevant ratios, i.e., seven
different probabilities of eating the forbidden food.
The Talmud also qualitatively discussed the estimation of prices for
quantities depending on chance (Franklin 2001, p. 261). It is
opportune to recall here that the Roman lawyer Ulpianus (170 – 228)
compiled a table of expectations of life, common for men and women
(Sheynin 1977b, pp. 209 – 210), although neither the method of its
compilation, nor his understanding of expectation are known. His
table was being used for determining the duration of some allowances
(Kohli & van der Waerden 1975, p. 515). At least methodologically
his table constituted the highest achievement of demographic statistics
until the 17th century.
Maimonides (Rabinovitch 1973, p. 164) mentioned expectation on
a layman’s level. He noted that a marriage settlement (providing for a
widow or a divorced wife) of 1000 zuz “can be sold for 100 [of such
monetary units], but a settlement of 100 can be sold only for less than
10”. It follows that there existed a more or less fixed expected value of
a future possible gain15.
A marriage settlement is a particular case of insurance; the latter
possibly existed in an elementary form even in the 20th century BC
(Raikher 1947, p. 40). Another statement of Maimonides (Rabinovitch
1973, p. 138) can also be linked with jurisprudence and might be
considered as an embryo of Jakob Bernoulli’s thoughts about
arguments (cf. §3.2.1):

One should not take into account the number of doubts, but rather
consider how great is their incongruity and what is their disagreement
with what exists. Sometimes a single doubt is more powerful than a
thousand other doubts.

Incongruity and disagreement, however, rather have to do with

opinions.
We also find an embryo of a random variable in Maimonides’
works (Rabinovitch 1973, p. 74):

Among contingent things some are very likely, other possibilities [of
other things?] are very remote, and yet others are intermediate.

In the new time one of the first to follow suit in natural science (not
in lotteries) was Maupertuis (1745, vol. 2, pp. 120 – 121) who
effectively explained instances when a child resembled one of his
remote ancestors, as well as [mutations] by non-uniform randomness.
At the same time, however, while discussing the origin of eyes and
ears in animals, he (1751, p. 146) only compared “une attraction
uniforme & aveugle” with some “principe d’intelligence” and came
out in favour of design.
An important digression from statistics is not out of order. While
considering a case of uncertain paternity, the Talmud (Makkot 315–16;
Rabinovitch 1973, p. 120) declared: “What we see we may presume,
but we presume not what we see not”, cf. the examples of direct and

24
24
contrary propositions at the beginning of this subsection. Compare
Newton (1729/1960, p. 547):

I have not been able to discover the cause of these properties of

gravity from phenomena, and I frame no hypotheses [hypotheses non
fingo].

Maimonides (1975, p. 123) advised physicians and judges to test, so

to say, simpler hypotheses, and then, only when necessary, to go on to
more complex assumptions. Cure ought to be attempted “through
food”, then by “gentle” medicine whereas “strong drugs” should be
considered as the last means. Just the same, a judge should try to
“effect a settlement” between litigants, then judge “in a pleasant
manner” and only then become “more firm”. Compare Newton’s Rule
No. 1 of reasoning in philosophy (Ibidem, p. 398):

We are to admit no more causes of natural things than such as are

both true and sufficient to explain their appearances16.

Moreover, Occam’s razor (William Occam, 1287 – 1347) comes to

mind: when two explanations are possible, the simpler is usually
better.
Thomas (1952) was the main commentator of Aristotle and he
strove to adapt the pagan Philosopher to Christianity. Just as his hero,
he believed that random events occurred in the minority of cases and
were due to some hindering causes (Sheynin 1974, p. 103):

Casual and chance events are such as proceed from their causes in
the minority of cases and are quite unknown:

Some causes are so ordered to their effects as to produce them not

of necessity but in the majority of cases, and in the minority to fail in
producing them … [which] is due to some hindering cause.

Again following Aristotle, Thomas illustrated this statement by the

“production of woman” which was nevertheless “included in nature’s
intention”17. He (Ibidem, p. 108) also maintained that law courts
should guide themselves by stochastic considerations.18

In the business affairs of men […] we must be content with a

certain conjectural probability.

On the introduction of moral certainty see §§ 2.1.2, 2.2.2 and 3.2.2.

Finally, Thomas (Ibidem p. 107) attributed ranks and degrees to
miracles; on this point see also Kruskal (1988). At a stretch, this
meant an introduction of qualitative probabilities. Byrne (1968)
studied the work of Thomas and argued (pp. 202 – 208) that he had
used a rudimentary frequency theory of probability but I fail to see
this. Anyway, Thomas provided a link between medieval and modern
science.

25
25
 1.2. Mathematical Treatment of Observations
1.2.1. Theory of Errors: General Information. I introduce some
notions and definitions which will be needed in § 1.4. Denote the
observations of a constant sought by

x1, x2, …, xn, x1 ≤ x2 ≤ … ≤ xn. (1)

It is required to determine its value, optimal in some sense, and

estimate the residual error. The classical theory of errors considers
independent observations (see § 9A.4-4) and, without loss of
generality, they might be also regarded as of equal weight. This
problem is called adjustment of direct observations.
Suppose now that k unknown magnitudes x, y, z, … are connected
by a redundant system of n physically independent equations (k < n)

ai x + bi y + ci z + … + si = 0 (2)

whose coefficients are given by the appropriate theory and the free
terms are measured. The approximate values of x, y, z, … were usually
known, or they could have been calculated by solving any subsystem
of k equations from (1) hence the linearity of (2). The equations are
linearly independent (a later notion), so that the system is inconsistent
(which was perfectly well understood). Nevertheless, a solution had to
be chosen which should have left small enough residual free terms
(call them vi). More precisely, those vi’s should have obeyed the
properties of usual random errors (approximately equal number of
positive and negative errors and more errors small in absolute value
than large errors).
The values of the unknowns thus obtained are their estimates (x0, y0,
…) and this problem, again with an evaluation of their errors, is called
adjustment of indirect measurements.
Since the early 19th century the usual condition for solving (2) was
that of least squares
W = vi2 = [vv] = v12 + v22 + … + vn2 = min19, (3)
so that
 W/  x =  W/  y = … = 0. (4)

Conditions (4) easily lead to a system of normal equations

[aa]x0 + [ab]y0 + … + [as] = 0, [ab]x0 + [bb]y0 + … + [bs] = 0, …, (5)

having a positive definite and symmetric matrix. For direct

measurements the same condition (3) leads to the arithmetic mean.
Another no less important and barely known to statisticians pattern of
adjusting indirect observations is described in § 9A.4-9.
1.2.2. Regular Observations. I have mentioned them in § 1.1.4.
They are necessary for excluding systematic, and compensating the
action of random errors. Kepler is known to have derived the laws of
planetary motion by issuing from Tycho’s regular observations, who

26
26
had thought that they provided a means for averaging out “random,
instrumental and human error” (Wesley 1978, pp. 51 – 52). Note,
however, that both instrumental and human errors can well be partly
random which testifies to Wesley poor knowledge of the theory of
errors. Wesley also states that Tycho (somehow) combined
measurements made by different instruments and that exactly that
approach might have produced most favourable results20.
Nevertheless, it seems that, when compiling his star catalogues,
Flamsteed, the founder of the Greenwich observatory, made use of
only a part of his observations (Baily 1835, p. 376):

Where more than one observation of a star has been reduced, he

has generally assumed that result which seemed to him most
satisfactory at the time, without any regard to the rest. Neither […]
did he reduce the whole (or anything like the whole) of his
observations. […] And, moreover, many of the results, which have
been actually computed, […] have not been inserted in any of his MS
catalogues.

Reduction, however, was a tiresome procedure and, anyway,

Flamsteed likely considered his results as preliminary which is
indirectly testified by the last lines of the passage just above and by
his own pronouncements (Sheynin 1973c, pp. 109 – 110), for example
by his letter of 1672, see Rigaud (1841, pp. 129 – 131):

I […] give you the sun’s diameters, of which I esteem the first, third
and fourth too large […] the rest I esteem very accurate, yet will not
build upon them till I have made some further trials with an exacter
micrometer.

Bradley’s principle of treating observations remains somewhat

unexplained (Sheynin 1973c, p. 110; Rigaud 1832). In one case
(Rigaud, p. 78) he derived the arithmetic mean of 120 observations,
and he (Ibidem, p. 17) supplemented his discovery of nutation of the
Earth’s axis by stating that

This points out to us the great advantage of cultivating [astronomy]

as well as every other branch of natural knowledge by a regular series
of observations and experiments.

He also discovered the aberration of light. At the same time he

(Rigaud 1832, p. 29) reported that

When several observations have been taken of the same star within
a few days of each other, I have either set down the mean result, or
that observation which best agrees with it.

And Boyle, the cofounder of scientific chemistry and co-author of

the Boyle – Mariotte law, kept to his own rule (Boyle 1772, p. 376):

27
27
 Experiments ought to be estimated by their value, not their number;
[…] a single experiment […] may as well deserve an entire treatise
[…]. As one of those large and orient pearls may outvalue a very
great number of those little […] pearls, that are to be bought by the
ounce […].

So are series of observations needed? All depends on the order of

the random errors, their law of distribution, on the magnitude of
systematic influences, the precision and accuracy required (the first
term concerns random errors, the second one describes systematic
corruption) and on the cost of observation. In any case, it is hardly
advisable to dissolve a sound observation in a multitude of worse
measurements. In geodesy, observations ought to be carried out under
differing (but good enough) meteorological conditions to lessen the
influence of systematic errors. Mendeleev (§ 10.9.3) likely held the
same opinion. In general, that same aim requires a prior knowledge of
the number of observations which prevents the application of
sequential analysis.
1.2.3. Galileo. When treating discordant observations of the
parallax of the New Star of 1572 made by several astronomers,
Galileo (1632, Day Third) formulated some propositions of the not yet
existing error theory21 and, first of all, indicated the properties of
usual random errors (also known to Kepler, see § 1.2.4). The method
of observation was of course worthless: in those days, even annual star
parallaxes remained unyielding to measurement. Astronomers were
only interested in placing the New Star either “beneath” the Moon or
“among” the fixed stars (but even that problem was really insoluble).
In essence, Galileo compared two natural scientific hypotheses with
each other and chose the latter. His test, later applied by Boscovich (§
6.3.2), was the minimal sum of absolute values of the parallaxes.
Because of computational difficulties, however, Galileo only took into
account some of the pairs of observations.
Buniakovsky (1846, Chapter on history of probability) mentioned
his investigation in a few lines but did not provide a reference;
Maistrov (1967/1974, pp. 30 – 34) described Galileo’s reasoning, but
see Hald (1990, pp. 149 – 160) for a detailed and rigorous discussion.
Apparently during 1613 – 1623 Galileo wrote a note about the
game of dice first published in 1718 (F. N. David 1962, pp. 65 – 66;
English translation on pp. 192 – 195). He calculated the number of all
the possible outcomes (and therefore, indirectly, the appropriate
probabilities) and testified that gamblers were believing that 10 or 11
points turned out more often than 9 or 12. If only these events are
considered (call them A and B respectively), then the difference
between their probabilities

P(A) = 27/52, P(B) = 25/52, P = 1/26 = 0.0385

can apparently be revealed. On determination of such small

differences see also Note 10 to Chapter 2.
In 1610 – 1612 several astronomers discovered sunspots. Daxecker
(2004; 2006) singled out the contribution of Christopher Scheiner and

28
28
discussed his book Rosa ursina sive Sol of 1626 – 1630. Much more
known is Galileo’s achievement (1613) who managed to separate the
regular rotation of the spots with the Sun itself from their random
proper movement relative to the Sun’s disk and estimated the period
of the Sun’s rotation as one lunar month; the present-day estimate is
24.5 – 26.5 days.
Humboldt (1845 – 1862, 1858, p. 64n) thought that sunspots could
have well been observed much earlier, for example “On the coast of
Peru, during the garua [?] […] even with the naked eye”, but he was
unable to substantiate his opinion.
Here, however, is the great traveller Marco Polo (Jennings 1985, p.
648). He described his conversation with the astronomer Jamal-ud-
Din, a Persian, and his team of Chinese astronomers. They discovered
the sunspots (and apparently observed them repeatedly) when “the
desert dust veiled the Sun”. Marco Polo narrative appeared in 1319
and his conversation with astronomers took place in the last quarter of
the 13th century somewhere near the present-day Chinese city Tianjin.
Marco Polo offered no comment, and it seems that he had not become
interested at all.
The subsequent history of sunspots is connected, first and foremost,
with their diligent observer, Schwabe, but it was Humboldt’s
description of his work that turned the attention of astronomers to
them (Clerke 1885/1893, p. 156).
While denying any possibility of an irregular motion of a heavenly
body, Galileo (1623/1960, p. 197) apparently denied randomness in
general:

Those lines are called regular which, having a fixed and definite
description, have been susceptible of definition and of having their
qualities and properties demonstrated. […] But irregular lines are
those which have no determinacy whatever […] and hence
indefinable. […] The introduction of such lines is in no way superior
to the sympathy, antipathy, occult properties, influences and other
terms employed by some philosophers as a cloak for the correct reply,
which would be I do not know.

This was possibly directed against some of Kepler’s utterances and

then, even if to a small extent, explains Galileo’s still mysterious
failure to recognize the Keplerian laws of planetary motion.
1.2.4. Kepler. Randomness and the Treatment of Observations.
Randomness played a certain part in Kepler’s astronomical
constructions. True, he (1606/2006, p. 163) denied it:

But what is randomness? Nothing but an idol, and the most

detestable of idols; nothing but contempt of God sovereign and
almighty as well as of the most perfect world that came out of His
hands.

Nevertheless, his laws of planetary motion were unable to justify

the values of the eccentricity of their orbits. He finally had to consider
them random, caused by disturbances, which was quite in the

29
29
Aristotelian spirit (§ 1.1.1)22, occasioned by deviation from (Divine)
laws of nature. In this connection I quote Poincaré (1896/1912, p. 1)
who most clearly formulated the dialectical link between randomness
and necessity in natural sciences:

There exists no domain where precise laws decide everything, they

only outline the boundaries within which randomness may move. In
accordance with this understanding, the word randomness has a
precise and objective sense.

Kepler (1596) first encountered those eccentricities when

attempting to construct a model of the solar system by inserting the
five regular solids between the spheres of the then six known planets:
they, the eccentricities, and, for that matter, unequal one to another,
much worried him. At that time, Kepler understood eccentricity as the
preordained eccentric position of the Sun as measured from the centre
of the circular orbit of a given planet.
He then changed his (actually, ancient) definition of eccentricity:

In the motion of the planets, the variables, if any (such as, in the
motion of the planets, the varying distance from the sun, or the
eccentricity), arise from the concurrence of extrinsic causes (Kepler
1609/1992, Chapter 38, p. 405).

On the same page he illustrated his opinion by obstacles which

prevent rivers from descending “towards the centre of the earth”, and
finally, on the next page, he concluded that “other causes are
conjoined with the motive power from the sun” [affect their motion].
Kepler (1620 – 1621/1952, p. 932) voiced his main statement in a
later contribution:

If the celestial movements were the work of mind, as the ancients

believed, then the conclusion that the routes of the planets are
perfectly circular would be plausible. […]
But the celestial movements are […] the work of […] nature […]
and this is not proved by anything more validly than by observation of
the astronomers, who […] find that the elliptical figure of revolution
is left in the real and very true movement of the planet. […]
Because in addition to mind there was then need of natural and
animal faculties [which] followed their own bent […].

Later, however, Kepler returned to his still cherished explanation of

the construction of the Solar system by means of the regular solids and
argued that

The origin of the eccentricities of the individual planets [is] in the

arranging of the harmonies between their motions (Kepler 1619/1997,
title of Chapter 9 of Book 5 on p. 451).

On that very page he explained that God had combined the

planetary motions with the five regular solids and thus created the
only most perfect prototype of the heaven.
30
30
 Again in the same chapter, in Proposition 5, on p. 454, he indirectly
mentioned in this connection his second law of planetary motion; for
that matter, he could have referred to it in his previous source, see
above. So it was this law according to which the preordained
eccentricities regulated the planetary motions. Had Kepler realized
that his contrived theory of the construction of the Solar system was
patently wrong, he would have probably returned to his previous
opinion about the “faculties” which “followed their own bent”. True,
it is difficult to see how, even acknowledging his theory, could have
eccentricities regulated planetary motions.
On the contrary, Kepler repeated his statement about the just
mentioned preordained eccentricities in the second edition of 1621 of
his theory (Kepler 1596/1963, pp. 97 and 104). Note, however, that he
had not abandoned his opinion about the possibility of external forces
corrupting Divine intention.
Without referring to anyone, Kant (1755/1910, p. 337) repeated
Kepler’s pronouncement about external forces:

The multitude of circumstances that participate in creating each

natural situation, does not allow the preordained regularity to occur.
Why are their [the planets’] paths not perfectly circular? Is it not seen
clearly enough, that the cause that established the paths of celestial
bodies […] had been unable to achieve completely its goal? […] Do
we not perceive here the usual method of nature, the invariable
deflection of events from the preordained aim by various additional
causes?

In § 3.2.4 I mention Kepler in connection with his ideas about the

end of the world.
Kepler (1604/1977, p. 337) also decided that a possible (that is, an
aimless) appearance of the New star in a definite place and at a
definite moment (both the place and the moment he, in addition,
considered remarkable) was so unlikely that it should have been called
forth by a cause [it had an aim], cf. § 1.1.1.
Kepler (Sheynin 1974, § 7) considered himself the founder of
scientific astrology, of a science of [correlational] rather than strict
influence of heaven on men and states. Thus (Kepler 1619/1952, book
4, pp. 377 – 378), his heavenly bodies were not Mercury, but
Copernicus and Tycho Brahe, and the constellations at his birth had
only awakened rather than hardened his spirit and the abilities of his
soul. And (1610/1941, p. 200), “heaven and earth are not coupled as
cog-wheels in a clockwork”. Before him Tycho likely held the same
view (Hellman 1970, p. 410). As an astrologer, Ptolemy (1956, I, 2
and I, 3) also believed that the influence of the heaven was a tendency
rather than a fatal drive, that astrology was to a large extent a science
of qualitative correlation, and Al-Biruni (1934, p. 232) likely thought
the same way: “The influence of Venus is towards …”, “The moon
tends …”23
They both thus anticipated Tycho and Kepler. For the latter, the
main goal of astrology was not the compilation of horoscopes
concerning individuals, but the determination of tendencies in the

31
31
development of states for which such circumstances as geographical
position, climate, etc., although not statistical data, should also be
taken into account. Cf. the approach of political arithmeticians (§
2.1.4).
A few decades ago physicists and mechanicians began to recognize
randomness as an essentially more important agent, a fact which I am
leaving aside, as well as chaotic phenomena, that is, exponentially
deviating paths of, for example, a ball on a non-elliptical billiard table.
Owing to unavoidable small uncertainty of its initial conditions, the
path becomes a cloud which fills a certain region. This chaos defies
simple quantitative definition24.
I qualitatively explain the difference between ordinary randomness
and chaotic motion: however long and complicated is the fall of a
tossed coin, the outcomes of this toss do not change. However, the
instability of chaotic motion rapidly increases with time and an
uncountable infinity of possible paths (a cloud) inevitably appears.
Kepler had to carry out enormous calculations and, in particular, to
adjust both direct and indirect measurements. The most interesting
example in the first case (Kepler 1609/1992, p. 200/63) was the
adjustment of the four following observations (I omit the degrees)

x1 = 23′39″, x2 = 27′37″, x3 = 23′18″, x4 = 29′48″.

Without any explanation, Kepler selected x = 24′ 33″ as the “medium

ex aequo et bono” (in fairness and justice). A plausible reconstruction
(Filliben, see Eisenhart 1976) assumes that x was a generalized
arithmetic mean with weights of observations being 2, 1, 1, and 0 (the
fourth observation was rejected). But the most important circumstance
here is that the Latin expression above occurred in Cicero’s Pro A.
Caecina oratio and carried an implication Rather than according to
the letter of the law. Rosental & Sokolov (1956, p. 126), in their Latin
textbook intended for students of law, included that expression in a list
of legal phrases and adduced Cicero’s text (p. 113; German
translation, see Sheynin 1993c, p. 186). In other words, Kepler, who
likely read Cicero, called the ordinary arithmetic mean the letter of the
law, i.e., the universal estimator [of the parameter of location].
It might be thought that such a promotion was caused by increased
precision of observations; their subjective treatment (§ 1.1.4) became
anachronistic. In addition, astronomers possibly began perceiving the
mean as an optimal estimator by analogy with the ancient idea of the
expediency of “mean” behaviour (§ 1.1.1)25.
Overcoming agonizing difficulties, Kepler repeatedly adjusted
indirect measurements. I dwell here on two points only. And, first of
all: How had he convinced himself that Tycho’s observations were in
conflict with the Ptolemaic system of the world? I believe that Kepler
applied the minimax principle (§ 6.3.2) demanding that the residual
free term of the given system of equations, maximal in absolute value,
be the least from among all of its possible “solutions”. His equations
were not even algebraic, but, after some necessary successive
approximations, they might have been considered linear. He
apparently determined such a minimum, although only from among

32
32
some possibilities, and found out that that residual was equal to 8′
which was inadmissible, see his appropriate statement (1609/1992, p.
286/113):

The divine benevolence had vouchsafed us Tycho Brahe, a most

diligent observer, from whose observations the 8′ error in this
Ptolemaic computation is shown […] [and, after a few lines] because
they could not have been ignored, these eight minutes alone will have
led the way to the reformation of all of astronomy, and have
constituted the material for a great part of the present work.

Then, when actually adjusting observations, he (Ibidem, p. 334/143)

corrupted them by small arbitrary corrections. He likely applied
elements of what is now called statistical simulation, but in any case
he must have taken into account the properties of “usual” random
errors, i.e., must have chosen a larger number of small positive and
negative corrections and about the same number of the corrections of
each sign. Otherwise, Kepler would have hardly achieved success.
I stress finally that Kepler (1609/1992, p. 523/256), although not
from the very beginning, came to understand that each observation
had to be taken into consideration:

Since the first and the third position […] agree rather closely, some
less thoughtful person will think that it [the constant sought] should be
established using these, the others being somehow reconciled. And I
myself tried to do this for rather a long time.

Here is an overview (Pannekoek 1961/1989, pp. 339 – 340):

In former centuries the astronomer selected from among his

observations those that seemed the best; this made him liable to bias
or inclined to select such data as showed a possibly unreal agreement.
[…] A new attitude was brought into being, typical of the nineteenth-
century scientist towards his material; it was no longer a mass of data
from which he selected what he wanted, but it was the protocol of an
examination of nature, a document of facts to which he had to defer.

Notes
1. See Note 22 (examples of randomness in the general sense).
2. I leave aside the views of Democritus, Epicurus and Lucretius since I think
that their works are not sufficiently understandable. Russell (1962, p. 83) considered
them “strict determinists”, but many other scholars were of the opposite opinion.
Thus, Kant (1755/1910, p. 344) remarked that the random mutual movement of the
Lucretius’ atoms had not created the world. Many ancient scientists reasoned on
randomness, see Note 3. And here is a strange statement of Strabo (1969, 2.3.7), a
geographer and historian:

Such a distribution of animals, plants and climates as exists, is not the result of
design – just as the difference of race, or of language, is not, either – but rather of
accident and chance.

Chrysippus (Sambursky 1956/1977, p. 6) held that chance was only the result of
ignorance and St. Augustinus, and, much later, Spinoza and D’Alembert expressed
similar thoughts (M. G. Kendall 1956/1970, p. 31, without an exact reference).

33
33
Kendall also maintains that Thomas (Sheynin 1974, p. 104) stated something
similar: a thing is fortuitous with respect to a certain, but not to a universal cause but
I think that that proposition is rather vague.
3. Bru (Cournot 1843/1984, p. 306) noted that a number of ancient scholars
expressly formulated such an explanation of chance. An example taken from ancient
Indian philosophy (Belvalkar & Ranade 1927, p. 458) admits of the same
interpretation:

The crow had no idea that its perch would cause the palm-branch to break, and
the palm-branch had no idea that it would be broken by the crow’s perch; but it all
happened by pure Chance.

Lack of aim or intersection of chains of events may be seen in Hobbes’ remark

(1646/1840, p. 259):

When a traveller meets with a shower, the journey had a cause, and the rain had
a cause […]; but because the journey caused not the rain, nor the rain the journey,
we say they were contingent one to another.

He added, however, that the rain was a random event since its cause was
unknown, cf. above.
4. More interesting is a by-law pronounced in ancient India, between the 2nd
century BC and 2nd century of our era (Bühler 1886/1967, p. 267):

The witness [in law-suits pertaining to loans], to whom, within seven days after he
has given evidence, happens [a misfortune through] sickness, a fire or the death of a
relative, shall be made to pay the debt and a fine.

This was an attempt to isolate necessity (a speedy divine punishment) from

chance. Another example (Hoyrup 1983) describes the death of 20 murderers in an
accident with only one person (the one who had unsuccessfully tried to prevent the
murders) surviving. This story, concerning the year 590 or thereabouts, was possibly
invented, but in any case it illustrated the same attempt.
5. Apparently: an invariable mutual arrangement of the stars cannot be random.
Levi ben Gerson (1999, p. 48) left a similar but less direct statement, but, strictly
speaking, such arguments are not convincing. It is impossible to say beforehand
which outcomes of ten (say) throws of a coin exhibit regularity, and which are a
result of chance: all of them are equally probable. It is opportune to recall the
D’Alembert – Laplace problem: The word Constantinople is composed of printer’s
letters; was the composition random? D’Alembert (1768a, pp. 254 – 255) stated that
all the arrangements of the letters were equally probable only in the mathematical
sense, but not “physically” so that the word was not random. Laplace (1776/1891, p.
152 and 1814/1995, p. 9) more correctly decided that, since the word had a definite
meaning [had an aim], its random [aimless] composition was unlikely. He thus
reasonably refused to solve this problem strictly. Poisson (1837a, p. 114) provided
an equivalent example and made a similar conclusion. Matthiesen (1867), however,
reported an extremely rare event: in a game of whist it occurred that each of the four
gamblers received cards of only one suit. It is of course impossible to check this
story and, anyway, it is reasonable to follow Laplace and Poisson.
6. Cicero understood the probable just as Aristotle did (Franklin 2001, p. 116).
Much later, in the Digest (the Roman code of civil laws, 533), the same
interpretation was indirectly repeated (Ibidem, p. 8). I supplement this information
about Cicero by another of his statements (1997, Book 1, § 12, p. 7):

Many things are probable and […] though these are not demonstrably true, they
guide the life of the wise man because they are so significant and clear-cut.

7. Later authors repeatedly expressed the same idea; I name Maimonides

(Rabinovitch 1973, p. 111), Thomas Aquinas (Byrne 1968, pp. 223 and 226) and
even Peter the Great (in 1716, his Kriegs-Reglement, see Sheynin 1978b, p. 286,
Note 39). I do not think, however, that practice followed such statements.

34
34
 8. Half-proofs were mentioned in law courts apparently in the 1190s (Franklin
2001, p. 18).
9. These numbers indirectly indicated the population of the towns. Deaths of
infants hardly counted here.
10. And here is Markov’s opinion of 1915 about graduates of Russian Orthodox
seminaries (Sheynin 1993a, p. 200):

They are getting accustomed by their schooling to a special kind of reasoning.

They must subordinate their minds to the indications of the Holy fathers and replace
their minds by the texts from the Scripture.

11. To compare all this with the practice of the 19th and partly 20th century: Gauss
calculated his measured angles to within 0.001 (actual precision about 700 hundred
times lower); Karl Pearson kept to the same habit, even if not to the same extent,
and at least one similar example pertained to Fisher, see a discussion of this subject
in Science, vol. 84, 1936, pp. 289 – 290, 437, 483 – 484 and 574 – 575.
12. A good example about the determinate error theory (Chapter 0, Note 2).
13. And here is a general estimate of Ptolemy written by a Soviet dinosaur during
Russia’s brighter (!) years (Chebotarev 1958, p. 579): his system “held mankind in
spiritual bondage for fourteen centuries”.
14. In accordance with the Talmud (Kelim 176), the volume of a “standard” hen’s
egg, which served as the unit of volume, was defined as the mean of the “largest”
and the “smallest” eggs [from a large batch]. The Talmud also provided elementary
considerations about linear measurements and some stipulations regarding their
admissible errors (Sheynin 1998b, p. 196).
15. Large payments were thus valued comparatively higher and this subjective
attitude can also be traced in later lotteries up to our days: although large winnings
are unrealistic, gamblers are apt to hope for them. The expectations of the various
winnings in the Genoese lottery, which had been carried out from the mid-15th
century, confirm the conclusion made above: they decreased with the increase in the
theoretically possible gain (N. Bernoulli 1709/1975, p. 321; Biermann 1957;
Bellhouse 1981). Recent experiments (Cohen et al 1970, 1971) seem to suggest that
psychological subjective probabilities differ from objective statistical probabilities.
An embryo of the notion of expectation might be seen in the administration of
justice in 11th-century India (Al-Biruni 1887, vol. 2, pp. 158 – 160). Thus,

If the suitor is not able to prove his claim, the defendant must swear. […] There
are many kinds of oath, in accordance with the value of the object o the claim.

The oath apparently became ever more earnest as that value increased; the
probability of lying with impunity multiplied by the value in question was the
expectation of fraudulent gain. However, when

the object of claim was of some importance, the accused man was invited to drink
some kind of a liquid which in case he spoke the truth would do him no harm.

16. At least once Ptolemy (1984, III 4, p. 153; H 232) stated that a simpler
hypothesis “would seem more reasonable”.
17. Hardly anyone later recalled that doubtful example; Lamarck (1815, p. 133),
however, believed that there existed deviations from divine design in the tree of
animal life and explained them by “une … cause accidentelle et par consequent
variable”. In 1629, William Ames, a theologian, stated that random events might
occur even with probability p ≥ 1/2, see Bellhouse (1988, p. 71) who does not
elaborate and provides no exact reference.
18. The Laws of Manu (Ibidem) and the ancient Chinese literature (Burov et al
1972, p. 108) contain examples of decisions based on elementary stochastic
considerations, e.g., accept as true the statement of the majority.
19. I am using the Gauss apt notation

[ab] = a1b1 + a2b2 + … + anbn.

35
35
 20. But how did Tycho treat his observations when one or even more of his
instruments had to be temporarily taken out of service? Its (their) removal from the
pool could have led to a systematic shift in the mean measurement.
21. Possibly somewhat exaggerating, Rabinovitch (1974, p. 355), who described
the legal problems and rituals of the Judaic religion, concluded that these
propositions (not formulated by Ptolemy) were known even in antiquity.
22. Chance also began to be recognized in biology (and even Aristotle thought
that monstrosities were random, see § 1.1.1). Harvey (1651/1952, p. 338) stated that
spontaneous generation (then generally believed in) occurred accidentally, as though
aimlessly, again see § 1.1.1:

Creatures that arise spontaneously are called automatic […] because they have
their origin from accident, the spontaneous act of nature.

I would say that Harvey considered randomness an intrinsic feature of nature.

Lamarck (1809, p. 62), also see Sheynin (1980, p. 338), kept to the same
opinion. Harvey (1651/1952, p. 462) also believed that the form of hen’s eggs was
“a mere accident” and thus indicated an example of intraspecific variation and
Lamarck (1817, p. 450) explained such variations by accidental causes.
23. Ptolemy’s source was devoted to astrology. Thorndike (1923, p. 115) testifies
that in the Middle Ages the Tetrabiblos exerted “great influence”.
24. Ekeland (2006) provides pictures of such clouds. Chaos is one of the main
subjects of his book, but its description is not good enough, and he (p. 125) even
compares it with a game of chance. Randomness, which is another of his subjects, is
treated hardly better (p. 86): reality “lies somewhere between” order and dependence
of everything on everything else. There are other defects as well, and one of them is
an extremely bad documentation.
25. Astronomers certainly applied the arithmetic mean even before Kepler did.
Tycho (Plackett 1958/1970, pp. 122 – 123) combined 24 of his observations into
(12) pairs and calculated the (generalized arithmetic) mean of these pairs, and of
three separate observations assigning equal weight to each of the 15 values thus
obtained. He chose the pairs in a manner allowing the elimination of the main
systematic errors and, apparently, so as to estimate, even if qualitatively, the
influence of random errors in 12 cases out of the 15. The separate observations
could have been somehow corrected previous to the adjustment. I shall describe a
similar case in § 6.3.2.
The mean occurs in early approximate calculations of areas of figures and
volumes of bodies so as to compensate the errors of approximate formulas and/or
the deviations of the real figures and bodies from their accepted models (Colebrooke
1817, p. 97).

36
36
 2. The Early History
2.1. Stochastic Ideas in Science and Society
2.1.1. Games of Chance. They fostered the understanding of the
part of chance in life and, even in antiquity, illustrated practically
impossible events (§ 1.1.1) whereas mathematicians discovered that
such games provided formulations of essentially new problems.
Furthermore, although Pascal did not apply his relevant studies to any
other domain, he (1654b/1998, p. 172) had time to suggest a
remarkable term for the nascent theory, – Aleae geometria, La
Géométrie du hazard, – and to indicate his desire to compile a
pertinent tract. Later Huygens (1657/1920) introduced his study of
games of chance by prophetically remarking that it was not a simple
“jeu d’esprit” and that it laid the foundation “d’une spéculation fort
intéressante et profonde”. Leibniz (1704/1996, p. 506) noted that he
had repeatedly advocated the creation of a “new type of logic” so as to
study “the degrees of probability” and recommended, in this
connection, to examine all kinds of games of chance1. In 1703 he
wrote to Jakob Bernoulli (Kohli 1975b, p. 509):

I would like that someone mathematically studies different games

(in which excellent examples of [the doctrine of estimating
probabilities] occur). This would be both pleasant and useful and not
unworthy either of you or of another respected mathematician.

Indeed, games of chance, and possibly only they, could have at the
time provided models for posing natural and properly formulated
stochastic problems. And, in addition, they were in the social order of
the day.
Rényi (1969) attempted to conjecture the essence of Pascal’s
proposed tract. He could have been right in suggesting its subject-
matter but not with regard to the year, – 1654, – when Pascal or rather
Rényi described it. Another shortcoming of Rényi’s attempt is that in
spite of treating philosophical issues, Pascal (again, Rényi) had not
mentioned Aristotle.
The theory of probability had originated in the mid-17th century
rather than earlier; indeed, exactly then influential scientific societies
came into being and scientific correspondence became usual. In
addition, during many centuries games of chance had not been
sufficiently conducive to the development of stochastic ideas (M. G.
Kendall 1956/1970, p. 30). The main obstacles were the absence of

37
37
“combinatorial ideas” and of the notion of chance events, superstition,
and moral or religious “barriers” to the development of stochastic
ideas, see also § 6.2.3. In essence, combinatorial analysis dates back to
the 16th century although already Levi ben Gerson (Rabinovitch
1973, pp. 147 – 148) had created its elements.
Montmort (1713/1980, p. 6) had testified to the superstition of
gamblers; Laplace (1814/1995, pp. 92 – 93) and Poisson (1837a, pp.
69 – 70) repeated his statement (and adduced new examples). When a
number has not been drawn for a long time in the French lottery,
Laplace says, “the mob is eager to bet on it” and he adds that an
opposite trend is also noticeable. The same illusions exist in our time
although Bertrand (1888a, p. XXII) had convincingly remarked that
the roulette wheel had “ni conscience ni mémoire”. Even a just game
(with a zero expectation of loss) is ruinous (§ 6.1.1) and is therefore
based on superstition while lotteries are much more harmful. Already
Petty (1662/1899, vol. 1, p. 64) stated that they were “properly a Tax
upon unfortunate self-conceited fools” and Arnauld & Nicole
(1662/1992, p. 332) indicated that an expectation of a large winning in
a lottery was illusory. In essence, they came out against hoping for
unlikely favourable events, cf. § 1.1.5, Note 15.
2.1.2. Jurisprudence. I mentioned it in §§ 1.1.1 and 1.1.5 and, in
particular, I noted that one of the first tests for separating chance from
necessity was provided for the administration of justice. It seems,
however, that the importance of civil suits and stochastic ideas in law
courts increased exactly in the mid-17th century2. Descartes
(1644/1978, pt. 4, § 205, p. 323) put moral certainty into scientific
circulation, above all apparently bearing in mind jurisprudence3.
Arnauld & Nicole (1662/1992) mentioned it; Jakob Bernoulli
(1713a/1999, pt. 4, Chapter 3) later repeated one of their examples
(from their pt. 4, Chapter 15). True, Leibniz (§3.1.2) doubted that
observations might lead to it.
In the beginning of the 18th century, Niklaus Bernoulli (§ 3.3.2)
devoted his dissertation to the application of the “art of conjecturing”
to jurisprudence. Elements of numerical measures in that science are
known even in the Roman canon law: it “had an elaborate system of
full proofs, half proofs, and quarter proofs” (Garber & Zabell 1979, p.
51, Note 23). On the Roman code of civil law see also Note 6 in
Chapter 1 and Franklin (2001, p. 211).
Leibniz (1704/1996, pp. 504 – 505) mentioned degrees of proofs
and doubts in law and in medicine and indicated that

Our peasants have since long ago been assuming that the value of a
plot is the arithmetic mean of its estimates made by three groups of
appraisers4.

2.1.3. Insurance of Property and Life Insurance. Marine

insurance was the first essential type of insurance of property but it
lacked stochastic ideas or methods. In particular, there existed an
immoral and repeatedly prohibited practice of betting on the safe
arrivals of ships. Anyway, marine insurance had been apparently
based on rude and subjective estimates. Chaufton (1884, p. 349)

38
38
maintained that in the Middle Ages definite values were assigned to
risks in marine operations but he possibly meant just such estimates.
And here is a quote from the first English Statute on assurance
(Publicke Acte No. 12, 1601; Statutes of the Realm, vol. 4, pt. 2, pp.
978 – 979):

And whereas it hathe bene tyme out of mynde an usage amongste

merchants, both of this realme and of forraine nacyons, when they
make any great adventure, […] to give some consideracion of money
to other persons […] to have from them assurance made of their
goodes, merchandizes, ships, and things adventured, […] whiche
course of dealinge is commonly termed a policie of assurance […].

Now, a lucid (regrettably, unsubstantiated) statement about the

indirect importance of marine insurance (O’Donnell (1936, p. 78):

Wide research concedes that Life Insurance came into its own not
by a front-door entrance, but by marine insurance porthole.

Life insurance exists in two main forms. Either the insurer pays the
policy-holder or his heirs the stipulated sum on the occurrence of an
event dependent on human life; or, the latter enjoys a life annuity. Life
insurance in the form of annuities was known in Europe from the 13th
century onward although later it was prohibited for about a century
until 1423 when a Papal bull officially allowed it (Du Pasquier 1910,
pp. 484 – 485). The annuitant’s age was not usually taken into
consideration either in the mid-17th century (Hendriks 1853, p. 112),
or even, in England, during the reign of William III [1689 – 1702] (K.
Pearson 1978, p. 134). Otherwise, as it seems, the ages had been
allowed for only in a very generalized way (Sheynin 1977b, pp. 206 –
212; Kohli & van der Waerden 1975, pp. 515 – 517; Hald 1990, p.
119). It is therefore hardly appropriate to mention expectation here,
but at the end of the 17th century the situation began to change.
It is important to note that in the 18th, and even in the mid-19th
century, life insurance therefore hardly essentially depended on
stochastic considerations5; moreover, the statistical data collected by
the insurance societies as well as their methods of calculations
remained secret. A special point is that more or less honest business
based on statistics of mortality hardly superseded downright cheating
before the second half of the 19th century. Nevertheless, beginning at
least from the 18th century, the institute of life insurance strongly
influenced the theory of probability, see §§ 4.2 and 6.1.1c.
I single out the Memorandum of De Witt (1671). He separated four
age groups (5 – 53; 53 – 63; 63 – 73; and 73 – 80 years) and assumed
that the chances of death increased in a definite way from one group
to the next one but remained constant within each of them. According
to his calculations, the cost of an annuity for “young” men should
have been 16 times higher than the yearly premium (not 14, as it was
thought).
Eneström (1896/1897, p. 66) called De Witt’s exposition unclear.
Thus, the risk of dying always concerned a three year old infant,

39
39
which was not only not explained, but expressed in a misleading way;
and the proposed chances of death at various ages were contrary to
what was actually calculated.
An appendix to the main text (Hendriks 1853, pp. 117 – 118)
contains an interesting observation. Examining “considerably more
than a hundred different classes, each class consisting of about one
hundred persons”, De Witt found that

For young lives each of these classes always produced to the

annuitants […] a value of more than sixteen florins of capital arising
from one florin of annual rent. […] Thus […] in practice, when the
purchaser of several life annuities comes to divide his capital […]
upon several young lives – upon ten, twenty, or more – this annuitant
may be assured, without hazard or risk of the enjoyment of an
equivalent, in more than sixteen times the rent which he purchases.

This statement belongs to the prehistory of the LLN and it also

shows a peculiar aspect of business in those times.
In the same year, in a letter to another mathematician, Hudde, De
Witt (Hendriks 1853, p. 109) in an elementary way calculated the cost
of annuity on several lives (an annuity that should be paid out until the
death of the last person of the group; usually, of a married couple). In
the process, he determined the distribution of the maximal term of a
series of observations [obeying a uniform law]. Kohli & van der
Waerden (1975) described in detail the history of the institution of
life insurance including the work of De Witt and Huygens (§ 2.2.2),
and I only note that the former had not justified his assumed law of
mortality. A likely corollary of De Witt’s work was that the price of
annuities sold in Holland in 1672 – 1673 depended on the age of the
annuitants (Commelin 1693, p. 1205).
The first estimation of the present worth of life annuities, based on
a table of expectations of life, was made by the Praetorian Prefect
Ulpianus, see § 1.1.5. Leibniz (1986, pp. 421 – 432), also see Leibniz
(2000), in a manuscript written in 1680, described his considerations
about state insurance, see Sofonea (1957a). He had not studied
insurance as such, but maintained that the “princes” should care about
the poor, remarked that the society ought to be anxious for each
individual etc. Much later Süssmilch (§ 6.2.2) formulated similar
ideas.
Tontines constituted a special form of mutual insurance. Named
after the Italian banker Laurens Tonti (Hendriks 1863), they, acting as
a single body, distributed the total sums of annuities among their
members still alive, so that those, who lived longer, received
considerable moneys. Tontines were neither socially accepted nor
widespread “on the assumed rationale that they are too selfish and
speculative” (Hendriks 1853, p. 116). Nevertheless, they did exist in
the 17th century, see Sheynin (1977b, pp. 210 – 211).
2.1.4. Population Statistics. The Old Testament (Numbers,
Chapter 1) reports on a general census, or, more precisely, on a census
of those able to bear arms. To recall (§ 1.1.2), the Talmud estimated
the population of towns only by the number of soldiers “brought

40
40
forth” [when needed]. In China, in 2238 BC or thereabouts, an
estimation of the population was attempted and the first census of the
warrior caste in Egypt occurred not later than in the 16th century BC
(Fedorovitch 1894, pp. 7 – 21). In Europe, even in 15th century Italy,
for all its achievements in accountancy and mathematics (M. G.
Kendall 1960),

counting was by complete enumeration and still tended to be a

record of a situation rather than a basis for estimation or prediction
in an expanding economy.

Only Graunt (1662) and, to a lesser extent, Petty (1690) can be

called the fathers of population statistics. They studied population,
economics, and commerce and discussed the appropriate causes and
connections by means of elementary stochastic considerations, also
see Urlanis (1963) and K. Pearson (1978, pp. 30 – 49). It was Petty
who called the new discipline political arithmetic. He (Petty
1690/1899, vol. 2, p. 244) plainly formulated his denial of
“comparative and superlative Words” and attempted to express
himself in “Terms of Number, Weight, or Measure…”; Graunt
undoubtedly did, if not said the same.
Petty (1927, vol. 1, pp. 171 – 172) even proposed to establish a
“register generall of people, plantations & trade of England”. In
particular, he thought of collecting the accounts of all the

Births, Mariages, Burialls […] of the Herths, and Houses […] as

also of the People, by their Age, Sex, Trade, Titles, and Office.

As noticed by Greenwood (1941 – 1943/1970, p. 61), the scope of

this Register was to be “wider than that of our existing General
Register Office”.
Strictly speaking, neither Petty, nor, as it seems, his followers ever
introduced any definition of political arithmetic. However, without
violating Petty’s thoughts quoted above, it is possible to say that the
aims of this new scientific discipline were to study from a socio-
economic point of view states and separate cities (or regions) by
means of (rather unreliable) statistical data on population, industry,
agriculture, commerce etc.
At least 30 from among Petty’s manuscripts (1927) pertained to
political arithmetic. This source shows him as a philosopher of science
congenial in some respects with Leibniz, his younger contemporary. I
adduce one passage (Ibidem, pp. 39 – 40); also see Sheynin (1977b,
pp. 218 – 220):

What is a common measure of Time, Space, Weight, & motion?

What number of Elementall sounds or letters, will […] make a speech
or language? How to give names to names, and how to adde and
subtract sensata, & to ballance the weight and power of words; which
is Logick & reason.

41
41
 Graunt (1662) studied the weekly bills of mortality in London
which began to appear in the 16th century and had been regularly
published since the beginning of the 17th century. For a long time his
contribution had been attributed to Petty. However, according to Hull
(Petty 1899, vol. 1, p. lii), Petty

Perhaps suggested the subject of inquiry, […] probably assisted

with comments upon medical and other questions here and there […],
procured [some] figures […] and may have revised, or even written
the Conclusion…

If so, Petty still perhaps qualifies as co-author, but I shall not

mention him anymore. And I also quote his Discourse (1674):

I have also (like the author of those Observations [like Graunt])

Dedicated this Discourse to […] the Duke of Newcastle.

And so, Graunt was able to use the existing fragmentary statistical
data and estimated the population of London and England as well as
the influence of various diseases on mortality. His main merit
consisted in that he attempted to find definite regularities in the
movement of the population. I only indicate that he established that
both sexes were approximately equally numerous (which contradicted
the then established views) and that out of 27 newly born about 14
were boys. When dealing with large numbers, Graunt did not doubt
that his conclusions reflected objective reality which might be seen as
a fact belonging to the prehistory of the LLN. Thus, the statistical ratio
14:13 was, in his opinion, an estimate of the ratio of the respective
[probabilities].
Nevertheless, he uncritically made conclusions based on a small
number of observations as well and thought that the population
increased in an arithmetical progression. Many authors had since
replaced that law by the geometrical progression definitely introduced
by Süssmilch and Euler (§ 6.2.2).
In spite of the meagre and sometimes wrong information provided
in the bills about the age of those dying, Graunt was able to compile
the first mortality table (common for both sexes). He calculated the
relative number of people dying within the first six years and within
the next decades up to age 86. According to his table, only one person
out of a hundred survived until that age. Now, how exactly had Graunt
calculated his table? Opinions vary, but, in any case, the very
invention of the mortality table was the main point here. The indicated
causes of death were also incomplete and doubtful, but Graunt
formulated some important relevant conclusions as well (although not
without serious errors)6. His general methodological (but not factual)
mistake consisted in that he assumed, without due justification, that
statistical ratios during usual years (for example, the per cent of yearly
deaths) were stable.
It is generally known that Graunt had essentially influenced later
scholars and here are a few pertinent opinions (Huygens, letter of

42
42
1662/1888 – 1950, 1891, p. 149; Süssmilch, see K. Pearson 1978, pp.
317 – 318; Willcox, see Graunt 1662/1939, p. x; Hald 1990, p. 86).

1. Grant’s [!] discourse really deserves to be considered and I like

it very much. He reasons sensibly and clearly and I admire how he
was able to elicit all his conclusions from these simple observations
which formerly seemed useless.

2. The discovery of statistical regularities was even as feasible as

that as America, it wanted only its Columbus…

3. Graunt is memorable mainly because he discovered […] the

uniformity and predictability of many biological phenomena taken in
the mass […] thus he, more than any [other] man, was the founder of
statistics.

4. Graunt, in his own words, reduced the data from “several great
confused Volumes into a few perspicuous Tables” and analysed them
in “ a few succinct Paragraphs” which is exactly the aim of statistics.
Hald could have referred here to Kolmogorov & Prokhorov, see §
0.2.
Halley (1694a; 1694b), a versatile scholar and an astronomer in the
first place, compiled the next mortality table. He made use of
statistical data collected in Breslau7, a city with a closed population.
Halley applied his table for elementary stochastic calculations
connected with life insurance and he was also able to find out the
general relative population of the city. Thus, for each thousand infants
aged less than a year, there were 855 children from one to two years
of age, …, and, finally, 107 persons aged 84 – 100. After summing up
all these numbers, Halley obtained 34 thousand (exactly) so that the
ratio of the population to the newly born occurred to be 34. Until 1750
his table remained the best one (K. Pearson 1978, p. 206).
For a long time, the explanation of the calculations in his main
(first) paper had not been properly understood. Then, the yearly rate of
mortality in Breslau was 1/30, the same as in London, and yet Halley
considered that city as a statistical standard. If only such a notion is
appropriate, standards of several levels ought to be introduced.
Finally, Halley thought that the irregularities in his data

Would rectify themselves, were the number of years [of observation]

much more considerable.

Such irregularities could have well been produced by systematic

influences, but, anyway, Halley’s opinion shows the apparently wide-
spread belief in an embryo of the LLN.
Halley’s second note is indeed interesting as a reasoning on the
welfare of the population.
In spite of the deficiencies mentioned, great success came
immediately. K. Pearson (1978, p. 78) indicated that Halley had made
“all the use that a modern actuary could” of his data and that he had
computed his life-table “as we should do it today”. Sofonea (1957b, p.

43
43
31*) called Halley’s contribution “the beginning of the entire
development of modern methods of life insurance”, and Hald (1990,
p. 141) stated that it “became of great importance to actuarial
science”.
Later scholars applied Halley’s contribution; the most important
example is De Moivre’s introduction (1725/1756/1967) of the uniform
law of mortality for ages beginning at 12 years.
In 1701 Halley (Chapman 1941, p. 5) compiled a chart of Northern
Atlantic showing the lines of equal magnetic declinations so that he
(and of course Graunt) might be called the founders of exploratory
data analysis, a most important, even if elementary stage of statistical
investigations, cf. § 10.8.3.
In 1680 – 1683 Leibniz wrote several manuscripts mostly
pertaining to the so-called statecraft (§ 6.2.1) and first published in
1866 (Leibniz 1986, pp. 340 – 349, 370 – 381, and 487 – 491), see
also Sheynin (1977b, pp. 224 – 227). He recommended the
publication of “state tables” (numerical or not?) of remarkable facts
and their comparison, year with year, or state with state. Their
compilation, as he suggested, should have been the duty of special
recording offices and, as it seems, for such offices Leibniz (disorderly)
listed 56 questions from which I mention the number of inhabitants of
a state and the comparison of the birth rate and mortality. Then, he
thought it advisable to collect information about scientific
achievements, “clever ideas” and medical and meteorological
observations, and to establish “sanitary boards” for compiling data on
a wide range of subjects (meteorology, medicine, agriculture).
One of Leibniz’ manuscripts (Leibniz 1986, pp. 456 – 467, or, with
a German translation, 2000, pp. 428 – 445) was devoted to political
arithmetic. There, he introduced the moyenne longueur de la vie
humaine8, necessary, as he remarked, for calculating the cost of
annuities; assumed without substantiation several regularities, for
example, that the ratio of mortality to population was equal to 1:40;
and wrongly stated that the mortality law for each age group including
infants was uniform. Following Arnauld & Nicole (1662/1992, pp.
331 and 332), he discussed apparence or degré de la probabilité and
apparence moyenne [expectation]. See §6.2.1 about the history of
university statistics.
Leibniz began by stating that in a throw of two dice the occurrence
of seven points is three times (actually six times) more probable than
of twelve points. Todhunter (1865, p. 48) indicated his second mistake
of the same kind. And Leibniz ended his essay by arguing that, since
hardly 1/9 or 1/10 of fertile women conceive yearly, the birth-rate can
be nine or ten times higher than it actually is. I simply do not
understand this argument.
Population statistics owed its later development to the general
problem of isolating randomness from Divine design. Kepler and
Newton achieved this aim with regard to inanimate nature, and
scientists were quick to begin searching for the laws governing the
movement of population, cf. K. Pearson’s appropriate remark in §
2.2.3.
2.2. Mathematical Investigations

44
44
 2.2.1. Pascal and Fermat. In 1654 Pascal and Fermat exchanged
several letters (Pascal 1654a) which heralded the beginning of the
formal history of probability. They discussed several problems; the
most important of them was known even at the end of the 14th century.
Here it is: Two or three gamblers agree to continue playing until one
of them scores n points; for some reason the game is, however,
ínterrupted on score a:b or a:b:c (a, b, c < n) and it is required to
divide the stakes in a reasonable way9. Both scholars solved this
problem (the problem of points; see Takácz 1994) by issuing from one
and the same rule: the winnings of the gamblers should be in the same
ratio(s) as existed between the expectations of their scoring the n
points, see for example Sheynin (1977b, pp. 231 – 239). The actual
introduction of that notion, expectation, was their main achievement.
They also effectively applied the addition and the multiplication
theorems10.
The methods used by Pascal and Fermat differed. In particular,
Pascal solved the above problem by means of the arithmetic triangle
composed, as is well known, of binomial coefficients of the
development (1 + 1)n for increasing values of n. Pascal’s relevant
contribution (1665) was published posthumously, but Fermat was at
least partly familiar with it. Both there, and in his letters to Fermat,
Pascal in actual fact made use of partial difference equations (Hald
1990, pp. 49 and 57).
The celebrated Pascal wager (1669/2000, pp. 676 – 681), also
published posthumously, was in essence a discussion about choosing a
hypothesis. Does God exist, rhetorically asked the devoutly religious
author and answered: you should bet. If He does not exist, you may
live calmly [and sin]; otherwise, however, you can lose eternity. In the
mathematical sense, Pascal’s reasoning11 is vague; perhaps he had no
time to edit his fragment. Its meaning is, however, clear: if God exists
with a fixed and however low probability, the expectation of the
benefit accrued by believing in Him is infinite.
2.2.2. Huygens. Huygens was the author of the first treatise on
probability (1657). Being acquainted only with the general contents of
the Pascal – Fermat correspondence, he independently introduced the
notion of expected random winning and, like those scholars, selected
it as the test for solving stochastic problems. Note that he went on to
prove that the “value of expectation”, as he called it, of a gambler who
gets a in p cases and b in q cases was

pa  qb
. (1)
pq

Jakob Bernoulli (1713a/1999, p. 9) justified the expression (1)

much simpler than Huygens: if each of the p gamblers gets a, and each
of the q others receives b, and the gains of all of them are the same,
then the expectation of each is equal to (1). After Bernoulli, however,
expectation began to be introduced formally: expressions of the type
of (1) followed by definition.
Huygens solved the problem of points under various initial
conditions and listed five additional problems two of which were due

45
45
to Fermat, and one, to Pascal. He solved them later, either in his
correspondence, or in manuscripts published posthumously. In
essence, they demanded the use of the addition and multiplication
theorems, the actual introduction of conditional probabilities and the
formula (in modern notation)

P(B) = ΣP(Ai)P(B/Ai), i = 1, 2, …, n.

I describe two of the five additional problems. Problem No. 4 was

about sampling without replacement. An urn contained 8 black balls
and 4 white ones and it was required to determine the ratio of chances
that in a sample of 7 balls 3 were, or were not white. Huygens
determined the expectation of the former event by means of a partial
difference equation (Hald 1990, p. 76), cf. Korteweg’s remark about
Huygens’ analytical methods below. Nowadays such problems leading
to the hypergeometric distribution (J. Bernoulli 1713a/1999, pp. 167
– 168; De Moivre 1712/1984, Problem 14 and 1718/1756, reprint
1967, Problem 20) appear in connection with statistical inspection of
mass production.
Pascal’s elementary Problem No. 5 was the first to discuss the
gambler’s ruin. Gamblers A and B undertake to score 14 and 11 points
respectively in a throw of 3 dice. They have 12 counters each and it is
required to determine the ratio of the chances that they be ruined. The
stipulated numbers of points occur in 15 and 27 cases and the ratio
sought is therefore (5/9)12.
In 1669, in a correspondence with his brother Lodewijk, Huygens
(1895), see Kohli & van der Waerden (1975), discussed stochastic
problems connected with mortality and, to be sure, life insurance. So it
happened that the not yet formed theory of probability spread over
new grounds. Issuing from Graunt’s mortality table (§ 2.1.4),
Huygens (pp. 531 – 532) introduced the probable duration of life (but
not the term itself) and explained that it did not coincide with expected
life. On p. 537 he specified that the latter ought to be used in
calculations of annuities and the former for betting on human lives.
Indeed, both he (pp. 524 – 526) and Lodewijk (pp. 484 – 485)
mentioned such betting. Christiaan also showed that the probable
duration of life could be determined by means of the graph (a
continuous curve passing through empirical points given by Graunt’s
table of mortality; plate between pp. 530 and 531) of the function

y = 1 – F(x),

where, in modern notation, F(x) was a remaining unknown integral

distribution function with admissible values of the argument being
0 ≤ x ≤ 100.
Also in the same correspondence Huygens (p. 528) examined the
expected period of time during which 40 persons aged 46 will die out;
and 2 persons aged 16 will both die. The first problem proved too
difficult, but Huygens might have remarked that the period sought was
40 years (according to Graunt, 86 years was the highest possible
age). True, he solved a similar problem but made a mistake. He

46
46
assumed there that the law of mortality was uniform and that the
number of deaths will decrease with time, but for a distribution,
continuous and uniform in some interval, n order statistics will divide
it into (n + 1) approximately equal parts and the annual deaths will
remain about constant. In the second problem Huygens applied
conditional expectation when assuming that one of the two persons
will die first. Huygens never mentioned De Witt (§ 2.1.3) whose work
(an official and classified document) had possibly been remaining
unknown12.
When solving problems on games of chance, Huygens issued from
expectations which varied from set to set rather than from constant
probabilities. He was therefore compelled to compose and solve
difference equations (Korteweg, see Huygens 1888 – 1950, 1920, p.
135) and he (like Pascal, see § 2.2.1) should be recalled in connection
with their history. See also Shoesmith (1986).
While developing the ideas of Descartes and other scholars about
moral certainty (§ 2.1.2), Huygens maintained that proofs in physics
were only probable and should be checked by appropriate corollaries
and that common sense should determine the required degree of
certainty of judgements in civil life. In a letter of 1691 Huygens (1888
– 1950, t. 10, p. 739) had indeed mentioned Descartes and, without
justification, dismissed probabilities of the order of p = 10–11 although
he hardly applied this, or any other number as a criterion. Note that
Borel (1943/1962, p. 27) proposed p = 10–6 and 10–15 as insignificant
on the human and the terrestrial scale respectively. However, a single
value is hardly suitable for any event. Also see Sheynin (1977b, pp.
251 – 252).
2.2.3. Newton. Newton left interesting ideas and findings pertaining
to probability (Sheynin 1971a), but much more important were his
philosophical views. Here is the opinion of K. Pearson (1926):

Newton’s idea of an omnipresent activating deity, who maintains

mean statistical values, formed the foundation of statistical
development through Derham [a religious philosopher], Süssmilch [§
6.2.2], Niewentyt [a statistician], Price [Chapt. 5] to Quetelet [§ 10.5]
and Florence Nightingale13…

Newton had not stated such an idea (although he thought that God
regularly delivered the system of the world from accumulating
corruptions, see below). In 1971 E. S. Pearson answered my question
on this point:

From reading [the manuscript of K. Pearson (1978)] I think I

understand what K. P. meant […]. He has stepped ahead of where
Newton had got to, by stating that the laws which give evidence of
Design, appear in the stability of the mean values of observations …

I have since found that K. Pearson (1978, pp. 161 and 653) had
attributed to De Moivre (1733/1756, reprint 1967, pp. 251 – 252) the
Divine stability of statistical ratios, that is, the original determination

47
47
or original design and referred to Laplace who (1814/1995, p. 37)
had indeed formulated a related idea:

In an indefinitely continued sequence of events, the action of

regular and constant causes ought, in the long run, to outweigh that of
irregular causes.

However, as I also note in § 7.1-3, Laplace never mentioned Divine

design.
K. Pearson (1926) then went over to De Moivre (§ 4.4) and Bayes
(Chapter 5) and maintained that their work was motivated by
theological and sociological causes rather than by mathematics itself.
And here is Newton’s most interesting pronouncement (1704/1782,
Query 31):

Blind fate could never make all the planets move one and the same
way in orbs concentrick, some inconsiderable irregularities excepted,
which may have risen from the mutual actions of comets and planets
upon one another, and which will be apt to increase, till this system
wants a reformation. Such a wonderful uniformity in the planetary
system must be allowed the effect of choice. And so must the
uniformity in the bodies of animals.

I have indicated that such considerations are logically imperfect but

practically certain (Note 5 in Chapter 1). The idea of a divine
reformation of the system of the world was later abandoned, but
Newton’s recognition of the existence and role of its random
disturbances is very important. Random, I specify, in the same sense
as the outcome of coin-tossing is. But at the same time Newton (1958,
pp. 316 – 318), just like Kepler (§ 1.2.4), denied randomness and
explained it by ignorance of causes. It was the future theologian
Bentley, who, in 1693, expressed his thoughts after discussing them
with Newton. The texts of two of his sermons, of Newton’s letters to
him, and an article on Newton and Bentley are in Newton (1958).
Finally, Newton’s remark (Schell 1960), that the notion of chance
might be applied to a single trial, has a philosophical side14.
When studying the chronology of ancient kingdoms, Newton
(1728, p. 52) left an interesting statement:

The Greek Chronologers […] have made the kings of their several
Cities […] to reign about 35 or 40 years a-piece, one with another;
which is a length so much beyond the course of nature, as is not to be
credited. For by the ordinary course of nature Kings Reign, one with
another, about 18 or 20 years a-piece; and if in some instances they
Reign, one with another, five or six years longer, in others they reign
as much shorter: 18 or 20 years is a medium.

Newton derived his own estimate from other chronological data and
his rejection of the twice longer period was reasonable. Nevertheless,
a formalized reconstruction of his decision is difficult: within one and
the same dynasty the period of reign of a given king directly depends

48
48
on that of his predecessor. Furthermore, it is impossible to determine
the probability of a large deviation of the value of a random variable
from its expectation without knowing the appropriate variance (which
Newton estimated only indirectly and in a generalized way). K.
Pearson (1928a) described Newton’s later indication of the sources of
his estimate and dwelt on Voltaire’s adjoining remarks, and,
especially, on the relevant work of Condorcet.
I am now mentioning Newton’s manuscript (1967, pp. 58 – 61)
written sometime between 1664 and 1666. This is what he wrote:

If the Proportion of the chances […] bee irrational, the interest

[expectation] may bee found after ye same manner.

He thought of a ball falling upon the centre of a circle divided into

sectors whose areas were in “such proportion as 2 to √5”. If the ball
“tumbles” into the first sector, a person gets a, otherwise he receives
b, and his “hopes is worth”

(2a + b√5) ÷ (2 + √5).

This was a generalization of expectation as defined by Huygens (§

2.2.2) and the first occurrence of geometric probability (§ 6.1.6).
Newton’s second example was a throw of an irregular die. He
remarked that [nevertheless] “it may bee found how much one cast is
more easily got than another”. He hardly thought about analytic
calculations, he likely bore in mind statistical probabilities. I can only
add that Newton may well have seen Graunt’s contribution (§ 2.1.4).
In 1693, when answering a question, Newton (Gani 1982)
determined the [probability] of throwing not less than one, two, and
three sixes with six, 12 and 18 dice respectively (cf. the De Méré
problem in Note 10). In the last-mentioned case, for example, his
calculations can be described by the formula

P = 1 – C182 (5 / 6)16 (1 / 6) 2  C181 (5 / 6)17 (1 / 6)  (5 / 6)18 .

2.2.4. Arbuthnot. Arbuthnot (1712) collected the data on births

(more precisely, on baptisms) in London during 1629 – 1710. He
noted that during those 82 years more boys (m) were invariably born
than girls (f) and declared that that fact was not the Effect of Chance
but Divine Providence, working for a good End. Indeed, as he added,
boys and men were subject to greater dangers and their mortality was
higher than that of the females, “as Experience convinces us”. Even
disregarding such [hardly exhibited] regularities as the “constant
Proportion” m:f and “fix’d limits” of the difference (m – f), the “Value
of Expectation” of a random occurrence of the observed inequality
was less than (1/2)82, he stated, see however Note 5 in Chapter 1.
Arbuthnot could have concluded that the births of both sexes
obeyed [the binomial distribution], which, rather than the inequality
m > f, manifested Divine design; and could have attempted to estimate
its parameter (approximately equal to 14:13, see § 2.1.4). Then, he had
not remarked that baptisms were not identical with births; that

49
49
Christians perhaps somehow differed from other people and, again,
that London was perhaps an exception. Graunt (1662, end of Chapter
3) had indeed stated that during 1650 – 1660 less than half of the
general population had believed that baptism was necessary.
And Arbuthnot had not known the comparative mortality of the
sexes. Nevertheless, later authors took note of his paper, continued to
study the same ratio m:f and, by following this subject, made
important stochastic findings (see especially § 4.4). Freudenthal
(1961, p. xi) even called Arbuthnot the author of the first publication
on mathematical statistics. From among many other recent
commentators I name Shoesmith (1987) and H. A. David & Edwards
(2001, pp. 9 – 11) and I note that Arbuthnot was the first to publish a
trick equivalent to the application of a generating function of the
binomial distribution although only for its particular case. Jakob
Bernoulli (§ 3.1.2) applied a generating function before Arbuthnot,
but his book only appeared in 1713.
In 1715, ‘sGravesande (K. Pearson 1978, pp. 301 – 303; Hald
1990, pp. 279 – 280) improved on Arbuthnot’s reasoning and
discussed it with Niklaus Bernoulli, cf. § 3.3.4.
Bellhouse (1989) described Arbuthnot’s manuscript written likely
in 1694. There, the author examined the game of dice, attempted to
study chronology (two examples, cf. § 2.2.3) and to a certain extent
anticipated his published note of 1712.
The study of the sex ratio at birth by most eminent scholars was
based on Arbuthnot’s data.

Notes
1. He himself began studying games of chance in 1675 (Biermann 1955). See
Gini (1946), Kohli (1975b) and Sylla (1998) for the main body of his
correspondence with Jakob Bernoulli (with comments). Also published (in the
original Latin) is the entire correspondence (Leibniz 1971, pp. 10 – 110; Weil et al
1993).
2. The volte-face of the public mood can be perceived when considering the
problem of absentees about whom nothing is known. So as not to violate God’s
commandment [which one?], Kepler (1610/1941, p. 238) as an astrologer refused to
state whether a particular absentee was alive or not. Jakob Bernoulli (1713/1999, p.
235), however, suggested to study, in such cases, the pertinent stochastic
considerations as also did Niklaus Bernoulli (§ 3.3.2).
3. This notion was introduced about 1400 for the solution of ethical problems
(Franklin 2001, p. 69). Recall also Thomas’ “conjectural probability” in § 1.1.5.
4. In 1705 he repeated his statement about the appraisal of lots (end of § 2.1.2) in
a letter to Jakob Bernoulli (Kohli 1975b, p. 512). Much earlier, he included it in a
Latin manuscript of 1680 – 1683 (Leibniz 2000, pp. 428 – 445, with a German
translation).
5. Several authors mentioned the practice of insuring a number of healthy infants
possibly substantiated by intuitive stochastic reasoning, see § 3.2.3.
6. It might be thought that Graunt attempted to allow for systematic corruptions
of the data. Thus, he reasonably supposed that the number of deaths from syphilis
was essentially understated [not only because of the difficulty in diagnosing but
also] out of ethical considerations.
7. He (as well as Leibniz) obtained them from Caspar Neumann, a Magister der
Philosophie and Member of the Societät der Wissenschaften in Berlin. In a letter of
1692 Leibniz (1970, p. 279) stated that the data were interesting. On Halley see
Böckh (1893), Chapman (1941) and Cook (1998).
8. Perhaps he was not acquainted with the correspondence of Huygens (§ 2.2.2).

50
50
 9. About 1400 an anonymous Italian author (Franklin 2001, pp. 294 – 296)
correctly solved this problem for the case of two gamblers, but had not sufficiently
justified his solution and made a mistake in a more complicated instance. It is much
more important to note, however, that he had not introduced the notion of
expectation (cf. below).
10. The term probability does not appear in the extant part of the correspondence,
only chance is applied. De Méré, a man of the world, had unintentionally initiated
that correspondence by asking Pascal why the chances of two apparently equivalent
events were different. An elementary calculation shows that either the gamblers
were able to reveal a difference of probabilities equal to 0.0264, cf. § 1.2.3, or, as
Ore (1960, pp. 411 – 412) and van der Waerden (1976) believed, De Méré was
able to calculate the appropriate probabilities, – but still thought that achieving a six
in four throws of a die and two sixes in 24 throws of two dice should have been
equally probable since 24/36 = 4/6. Actually,

P1 = 1 – (5/6)4 ≈ 0.5177, P2 = 1 – (35/36)24 ≈ 0.4913.

Later Jakob Bernoulli (1713a/1999, p. 32) considered the same problem.

A queer episode concerning De Méré occurred in the 19th century. Georg Cantor
mistakenly thought that by his conclusion the man of the world had wished to
destroy science. Accordingly, he privately called Kronecker (who had denied the
emerging set theory) “Herr von Méré” (Fraenkel 1930, p. 199).
11. For a similar reasoning see Arnauld & Nicole (1662/1992, p. 334).
12. During the last years of his life Jakob Bernoulli vainly asked Leibniz to
procure for him a copy of De Witt’s work.
13. I would have excluded Niewentyt from the Pearsonian chain; Derham,
however, had been very influential.
14. Here is a similar statement formulated in the 4th century BC (Burov et al 1972,
p. 203):

Who even before battle gains victory by military estimations has many chances
[…] who has many chances gains victory; who has few chances does
not gain victory; all the less he who has no chances at all.

51
51
 3. Jakob Bernoulli and the Law of Large Numbers
I consider Bernoulli’s main work, the Ars conjectandi (AC)1,
published posthumously in Latin and touch on his Diary
(Meditationes) for 1684 – 1690. Only the stochastic part of the latter
was published together with the AC (both in their original Latin), with
other materials and comments (J. Bernoulli 1975). I also discuss
related topics and dwell on Bernoulli’s contemporaries. The AC was
translated into German and its separate parts have also appeared in
other living languages. In 1913 Uspensky translated Part 4 into
Russian (reprint: J. Bernoulli 1986). My own translation, or rather
rendition of the same part has also appeared (J. Bernoulli 2005). An
English translation of the entire Ars has been published; I reviewed it
(Sheynin 2006a) and concluded that the Cobbler should stick to his
last. In the sequel, I refer to the German translation of Bernoulli
(1899, reprinted in 1999).
3.1. Bernoulli’s Works
3.1.1. The Diary. There, Bernoulli studied games of chance and the
stochastic side of civil law. He (1975, p. 47) noted that the
probability2 of a visitation of the plague in a given year was equal to
the ratio of the number of these visitations during a long period of
time to the number of years in that period. I stress that Bernoulli thus
applied the definition of probability of an event (of statistical
probability!) rather than making use of chances. An interesting point
in this connection is that he (p. 46, marginal note) wrote out the
imprint of a review published in 1666 of Graunt’s book (§ 2.1.4)
which Bernoulli possibly had not seen; he had not referred to it either
in the Meditationes itself or in the AC. But the most important part of
his Diary is a (fragmentary) proof of the LLN. This fact means that
Bernoulli proved it not later than in 1690.
3.1.2. The Art of Conjecturing (1713a, b). Its Contents. The last
part of this book is entitled The use and application of the previous
doctrine to civil, moral and economic affairs (J. Bernoulli
1713a/1999, p. 229) but nothing of the sort had appeared there3.
Interesting problems are solved in parts 1 and 3 (the study of random
sums for the uniform and the binomial distributions, a similar
investigation of the sum of a random number of terms for a particular
discrete distribution, a derivation of the distribution of the first order
statistic for the discrete uniform distribution and the calculation of
probabilities appearing in sampling without replacement). The
author’s analytical methods included combinatorial analysis and
calculation of expectations of winning in each set of finite and infinite
games and their subsequent summing.
Part 1 is a reprint of Huygens’ tract (§ 2.2.2) including the solution
of his five additional problems, one of which Bernoulli (1713a/1999,
p. 167) had, however, carried over to Part 3, together with vast and
valuable commentaries. Nevertheless, this form again testifies that he
was unable to complete his contribution. Also in Part 1 Bernoulli (pp.
22 – 28), while considering a game of dice, compiled a table which
enabled him to calculate the coefficients of xm in the development of
(x + x2 + … + x6)n for small natural values of n.

52
52
 Part 2 did not bear on probability. It dealt with combinatorial
analysis and it was there that the author introduced the Bernoulli
numbers.
Part 4 contained the LLN. There also we find a not really formal
“classical” definition of probability (a notion which he had not applied
when formulating that law), a reasoning, in Chapter 2, on the aims of
the art of conjecturing (determination, as precise as possible, of
probabilities for choosing the best solutions of problems, apparently in
civil life) and elements of stochastic logic4.
Bernoulli likely considered the art of conjecturing as a
mathematical discipline based on probability as a measure of certainty
and on expectation and including (the not yet formally introduced)
addition and multiplication theorems and crowned by the LLN.
Bernoulli informed Leibniz about the progress in his work in a
letter of 3 Oct. 1703 (Kohli 1975b, p. 509). He was compiling it for
many years with repeated interruptions caused by his “innate laziness”
and worsening of health; the book still lacked its “most important
part”, the application of the art of conjecturing to civil life;
nevertheless, he, Bernoulli, had already shown his brother [Johann]
the solution of a “difficult problem, special in its own way” [§ 3.2.3],
that justified the applications of the art of conjecturing.
Most important both in that letter and in the following
correspondence of 1703 – 17055 (Ibidem, pp. 510 – 512) was the
subject of statistical probabilities, also see §§ 3.2.2 – 3.2.3. Leibniz
never agreed that observations could secure moral certainty but his
arguments were hardly convincing. Thus, he in essence repeated the
statement of Arnauld & Nicole (1662/1992, pp. 304 and 317) that the
finite (the mind; therefore, observations) cannot always grasp the
infinite (for example, God, but also, as Leibniz stated, any
phenomenon depending on innumerable circumstances).
These views were possibly caused by his understanding of
randomness as something “whose complete proof exceeds any human
mind” (Leibniz 1686/1960, p. 288). His heuristic statement does not
contradict a modern approach to randomness founded on complexity
and he was also right in the sense that statistical determinations cannot
definitively corroborate a hypothesis.
In his letter of 3 Dec. 1703 Leibniz (Gini 1946, p. 405) had also
maintained that the allowance for all the circumstances was more
important than subtle calculations, and Bortkiewicz (1923, p. 12) put
on record Keynes’ favourable attitude towards this point of view and
indicated the appropriate opinion of Mill (1843/1886, p. 353) who had
sharply contrasted the consideration of circumstances with “elaborate
application” of probability. Mill could have mentioned applied
mathematics in general; the circumstances there are the extents to
which its models are compatible with reality. Calculations (not
necessarily stochastic) are hardly more important than such
circumstances, cf. Gauss’ opinion in §§ 9A and 9A.5-2. But
circumstances should not be contrasted with calculations, the less so
because they can initially remain insufficiently known.
Bernoulli paid due attention to Leibniz’ criticism: more than a half
of Chapter 4 of Part 4 of the AC in essence coincided with the

53
53
respective passages from his letters to Leibniz; in that chapter,
Bernoulli (1713a/1999, p. 250), in particular, discussed the objections
made by “scientists”, that is, by Leibniz6.
3.2. The Art of Conjecturing, Part 4: Its Main Propositions
3.2.1. Stochastic Assumptions and Arguments. Bernoulli
examined these in Chapters 2 and 3 but did not return to them
anymore; he possibly thought of applying them in the unwritten pages
of his book. The mathematical aspect of his considerations consisted
in the use of the addition and the multiplication theorems for
combining various arguments.
Unusual was the non-additivity of the [probabilities]7. Here is one
of his examples (p. 244): “something” possesses 2/3 of certainty but
its opposite has 3/4 of certainty; both possibilities are probable and
their probabilities are as 8:9. Koopman (1940) resumed, in our time,
the study of non-additive probabilities whose sources can be found in
the medieval doctrine of probabilism that considered the opinion of
each theologian as probable. Franklin (2001, p. 74) attributed the
origin of probabilism to the year 1577, or, in any case (p. 83), to 1611.
Nevertheless, similar pronouncements on probabilities of opinion go
back to John of Salisbury (the 12th century) and even to Cicero
(Garber & Zabell 1979, p. 46).
I note a “general rule or axiom” concerning the application of
arguments (pp. 234 and 236): out of two possibilities, the safer, the
more reliable, or at least the more probable should be chosen8;
gamblers, however, always acted that same way (§ 1.2.3) if only they
did not follow superstitious beliefs (§ 2.1.1).
3.2.2. Statistical Probability and Moral Certainty. Before going
on to prove his LLN, Bernoulli (p. 246) explained that the theoretical
“number of cases” was often unknown, but what was impossible to
obtain beforehand, might at least be determined afterwards, i.e., by
numerous observations. The application of statistical [probabilities],
he maintained, was not new at all and referred to the “celebrated”
Arnauld, the co-author of Arnauld & Nicole (1662/1992)9. In his
Diary, Bernoulli indirectly mentioned Graunt (§ 3.1.1) and,
furthermore, quite to the point, in connection with the impossibility of
determining, without applying statistical data, how much more
probable was the case of a youth outliving an old man than the
opposite instance10.
He informed Leibniz about his opinion (cf. § 3.1.2) and added that
exactly that consideration led him to the idea of replacing, when
necessary, prior knowledge by posterior. Recall also Bernoulli’s
reasoning on the statistical probability of a plague epidemic (§ 3.1.1).
I discussed moral certainty in §§ 2.1.2 and 2.2.2. Bernoulli (p. 238)
maintained that it ought to be admitted on a par with absolute certainty
and that judges must have firm instructions about what exactly (for
example, 0.99 or 0.999 of certainty) constituted moral certainty. The
latter idea was hardly ever put into practice; furthermore, the
probability of a just sentence must be the higher the more important it
is. On p. 249 Bernoulli mentioned moral certainty once more. His
theorem will show, he declared, that statistical [probability] was a

54
54
morally certain [a consistent, in modern terms] estimator of the
theoretical [probability]11.
3.2.3. The Law of Large Numbers. Bernoulli proved a proposition
that, beginning with Poisson, is called the LLN (§ 8.7). Let r and s be
natural numbers, t = r + s, n, a large natural number, ν = nt, the
number of independent12 trials in each of which the studied event
occurs with [probability] r/t, µ – the number of the occurrences of the
event (of the successes). Then Bernoulli proved without applying
mathematical analysis that

  r 1 1
P |  |    1  (1)
  t t  1  c

and estimated the value of ν necessary for achieving a given c > 0. In

a weaker form Bernoulli’s finding meant that

  r 
limP = |  |    = 1, ν → ∞, (2)
  t 

where, as also in (1), r/t was the theoretical, and µ/ν, the statistical
probability.
Markov (Treatise, 1924, pp. 44 – 52) improved Bernoulli’s
estimate mainly by specifying his intermediate inequalities and K.
Pearson (1925), by applying the Stirling formula, achieved a
practically complete coincidence of the Bernoulli result with the
estimate that makes use of the normal distribution as the limiting case
of the binomial law13. In addition, Pearson (p. 202) considered
Bernoulli’s estimate of the necessary number of trials in formula (1)
“crude” and leading to the ruin of those who would have applied it. He
also inadmissibly compared Bernoulli’s law with the wrong
Ptolemaic system of the world (and De Moivre with Kepler and
Newton):

Bernoulli saw the importance of a certain problem; so did Ptolemy,

but it would be rather absurd to call Kepler’s or Newton’s solution of
planetary motion by Ptolemy’s name!

The very fact described by formulas (1) and (2) was, however,
extremely important for the development of probability and
statistics14; and, anyway, should we deny the importance of existence
theorems?
And so, the LLN established a correspondence between the two
probabilities15. Bernoulli (p. 249) had indeed attempted to ascertain
whether or not the statistical probability had its “asymptote”; whether
there existed such a degree of certainty, which observations, no matter
how numerous, were unable to ensure. Or, in my own words, whether
there existed such positive numbers ε and δ < 1, that

  r 
lim P |  |    ≤ 1 – δ, ν → ∞.
  t 

55
55
 He answered his question in the negative: no, such numbers did not
exist. He thus established, within the boundaries of stochastic
knowledge, a relation between deductive and inductive methods and
combined statistics with the art of conjecturing. Strangely enough,
statisticians for a long time had not recognized this fact. Haushofer
(1872, pp. 107 – 108) declared that statistics, since it was based on
induction, had no “intrinsic connections” with mathematics based on
deduction (consequently, neither with probability). A most noted
German statistician, Knapp (1872, pp. 116 – 117), expressed a
strange idea: the LLN was hardly useful since statisticians always
made only one observation, as when counting the inhabitants of a city.
And even later, Maciejewski (1911, p. 96) introduced a “statistical law
of large numbers” instead of the Bernoulli proposition that allegedly
impeded the development of statistics. His own law qualitatively
asserted that statistical indicators exhibited ever lesser fluctuations as
the number of observations increased.
All such statements definitely concerned the Poisson law as well
(European statisticians then hardly knew about the Chebyshev form
of the LLN) and Maciejewski’s opinion likely represented the
prevailing attitude of statisticians. Here, indeed, is what Bortkiewicz
(1917, pp. 56 – 57) thought: the expression law of large numbers
ought to be used only for denoting a “quite general” fact, unconnected
with any definite stochastic pattern, of a higher or lower degree of
stability of statistical indicators under constant or slightly changing
conditions and given a large number of trials. Even Romanovsky
(1912, p. 22; 1924, pt. 1, p. 15; 1961, p. 127) kept to a similar view.
Thus, in the last-mentioned contribution he stressed the natural-
scientific essence of the law and called it physical.
The LLN has its prehistory. It was thought, long before Bernoulli,
that the number of successes in n “Bernoulli” trials with probability p
was approximately equal to

µ = np. (3)

Cardano (Ore 1963, pp. 152 – 154 and 196), for example, applied
this formula in calculations connected with games of dice. When
compiling his mortality table, Halley (§ 2.1.4) assumed that
“irregularities” in his data would have disappeared had he much more
observations at his disposal. His idea can be interpreted as a statement
on the increase in precision of formula (3) with n, see however §
2.1.4. Also see Graunt’s reasoning (same subsection).
A second approach to the LLN took shape in astronomy when the
arithmetic mean became the universal estimator of the constant sought
(§ 1.2.4), cf. § 6.3.3. If the expectation of each of the magnitudes (1.1)
is equal to that constant, i.e., if systematic errors are absent, and if (as
always was the case) their variances are bounded, it could be thought
that a was approximately equal to the arithmetic mean of the
observations.
Similar but less justified statements concerning sums of magnitudes
corrupted by random errors had also appeared. Thus, Kepler (Sheynin

56
56
1973c, p. 120) remarked that the total weight of a large number of
metal money (the mean weight!) of the same coinage did not depend
on the inaccuracy in the weight of the separate coins. Then (§ 2.1.3),
when buying annuities “upon several young lives”, the expectation of
a gain Exi from each such transaction was obviously positive; if
constant, the buyer could expect a total gain of n∑Exi. Much later
Condorcet (1785a, p. 226) testified that those engaged in such
“commerce” (and apparently ignorant of the LLN) had regarded it as
“sûre”.
There also likely existed a practice of an indirect participation of
(petty?) punters in many games at once. At any rate, both De Moivre
(1718/1756, reprint 1967, Problem 70) and Montmort (1708/1980, p.
169) mentioned in passing that some persons bet on the outcomes of
games16. The LLN has then been known but that practice could have
existed from much earlier times. And, finally, Gower (1993, p. 272)
noted that Boscovich (1758, § 481) had [somewhat vaguely]
maintained that the sum of random magnitudes decreased with an
increase in the number of terms.
The attitude of Kepler and Boscovich is understandable: a similar
wrong opinion about random sums had persisted at least until the 19th
century, so that Helmert (1905, p. 604) thought advisable to refute it
in his own field.
3.2.4. Randomness and Necessity. Apparently not wishing to
encroach upon theology, Bernoulli (beginning of Chapter 1) refused to
discuss the notion of randomness. Then, in the same chapter, he
offered a subjective description of the “contingent” but corrected
himself at the beginning of Chapter 4 where he explained randomness
by the action of numerous complicated causes. Finally, the last lines
of his book contained a statement to the effect that some kind of
necessity was present even in random things. He referred to Plato who
had taught that after a countless number of centuries everything
returned to its initial state. Bernoulli likely thought about the archaic
notion of the Great Year whose end will cause the end of the world
with the planets and stars returning to their positions at the moment of
creation. Without justification, he widened the boundaries of
applicability of his law and his example was, furthermore, too
complicated. It is noteworthy that Kepler (1596) believed that the end
of the world was unlikely. In this, the first edition of this book, his
reasoning was difficult to understand but later (1621) he substantiated
his conclusion by stating, in essence, like Oresme (1966, p. 247) did
before him, that two [randomly chosen] numbers were “probably”
incommensurable17. Bernoulli (end of Chapter 1) also borrowed
Aristotle’s example (§ 1.1.1) of finding a buried treasure, but, unlike
him, had not connected it with randomness.
3.3. Bernoulli’s Contemporaries
I dwell somewhat on the ideas and findings of some of Bernoulli’s
contemporaries but I postpone the discussion of De Moivre, whose
first publication had appeared before the AC did, until Chapter 4.
3.3.1. Arnauld. Arnauld & Nicole anonymously put out their book,
Art of Reasoning (1662/1992)18. Arnauld was its main author and I
mentioned him in Note 11 of Chapter 2 (in connection with the Pascal

57
57
wager), in § 2.1.2 (moral certainty) and § 2.1.1 (an advice to neglect
unlikely favourable events). Then (§§ 3.2.2 and 3.2.1), I noted that
Bernoulli mentioned him when justifying the use of statistical
probabilities and borrowed his principle of behaviour. Finally,
Arnauld repeatedly, although without a formal definition, applied the
term probabilité (for example, on pp. 331 and 332), and degrez de
probabilité. Recall that Leibniz (§§ 3.1.2 and 2.1.4), in turn, borrowed
from him a reasoning and a term.
3.3.2. Niklaus Bernoulli. He published a dissertation on the
application of the art of conjecturing to jurisprudence (1709/1975). An
utterly unsuitable English translation is available in the Internet: Latin
phrases up to 40 lines long are rendered in even longer sentences and
in general very much is incomprehensible.
The dissertation contained
a) The calculation of the mean duration of life for persons of
different ages.
b) A recommendation of its use for ascertaining the value of
annuities and estimating the probability of death of absentees about
whom nothing is known. He proposed to consider them dead when the
probability of their death is twice higher than that of the alternative.
c) Methodical calculations of expected losses in marine insurance.
d) The calculation of expected gains (more precisely, of expected
losses) in the Genoese lottery.
e) Calculation of the probability of truth of testimonies.
f) The determination of the life expectancy of the last survivor of a
group of men (pp. 296 – 297; Todhunter 1865, pp. 195 – 196).
Assuming a continuous uniform law19 of mortality, he calculated the
expectation of the appropriate [order statistic]. He was the first to use,
in a published work, both this distribution and an order statistic.
g) A comment on the introduction of expectation by Huygens (p.
291; Kohli 1975c, p. 542), see expression (2.1). Bernoulli interpreted
it as a generalized arithmetic mean and the centre of gravity “of all
probabilities” (this is rather loose).
Apparently in accordance with his subject he had not discussed the
treatment of observations, cf. § 2.1.4. Bernoulli’s work undoubtedly
fostered the spread of stochastic notions in society (cf. § 2.1.2), but I
ought to add that not only did he pick up some hints included in the
manuscript of the Ars conjectandi, he borrowed separate passages
both from it and even from the Meditationes (Kohli 1975c, p. 541),
never intended for publication. His numerous general references to
Jakob do not excuse his plagiarism.
3.3.3. Montmort. He is the author of an anonymous book
(1708/1980), important in itself and because of its obvious influence
upon De Moivre as well as on Niklaus Bernoulli, the correspondence
with whom Montmort included in 1713 in the second edition of his
work. In the Introduction (1708/1713, p. iii) he indicated that in
practical activities and considerations it was desirable to be guided by
“geometry” rather than by superstition, cf. § 2.1.1. However, he (p.
xii) added that, since he was unable to formulate appropriate
“hypotheses”, he was not studying the applications of [stochastic]
methods to civil life20.

58
58
 Henny (1975) and Hald (1990) examined Montmort’s findings. The
latter, on his p. 290, listed Montmort’s main methods: combinatorial
analysis, recurrent formulas and infinite series; and on p. 209 Hald
added the method (the formula) of inclusion and exclusion

P(ΣAi) = ΣP(Ai) – ΣP(Ai·Aj) + ΣP(Ai·Aj·Ak ) – …, (4)

where A1, A2, …, An were events and i < j < k < …This formula is an
obvious stochastic corollary of a general proposition about arbitrarily
arranged sets.
Here are some problems solved by Montmort (1708/1980, pp. 244 –
246/NNo. 46 – 50 and 203 – 205/NNo. 200 – 202; 130 – 143), see
Hald (1990, pp. 196 – 198; 206 – 213; 292 – 297; and 328 – 336)
respectively:
1) The problem of points. Montmort arrived at the negative
binomial distribution. He returned to this problem in his
correspondence with Niklaus Bernoulli (Hald 1990, pp. 312 – 314).
2) A study of throwing s points with n dice, each having f faces.
Montmort applied the combinatorial method and formula (4).
3) A study of arrangements and, again, of a game of dice. Montmort
arrived at the multivariate hypergeometric, and the multinomial
distributions.
4) A study of occupancies. Tickets numbered 1, 2, …, n, are
extracted from an urn one by one without replacement. Determine the
probability that at least one ticket with number k, 1 ≤ k ≤ n, will occur
at the k-th extraction. Montmort derived the appropriate formulas

Pn = 1 – 1/2! + 1/3! – … + (–1)n – 1/n!, limPn = 1 – 1/e, n → ∞.

Niklaus Bernoulli and De Moivre returned to this problem, see H. A.

David & Edwards (2001, pp. 19 – 29).
3.3.4. Montmort and Niklaus Bernoulli: Their Correspondence.
I outline their correspondence of 1710 – 1713 (Montmort 1708/1980,
pp. 283 – 414).
1) The strategic game Her (Hald 1990, pp. 314 – 322). The modern
theory of games studies it by means of the minimax principle.
Nevertheless, already Bernoulli indicated that the gamblers ought to
keep to [mixed strategies]. See also Bellhouse et al (2015).
2) The gambler’s ruin. Montmort wrote out the results of his
calculations for some definite initial conditions whereas Bernoulli
indicated, without derivation, the appropriate formula (an infinite
series). Hald believes that he obtained it by means of the method of
inclusion and exclusion. On this point and on the appropriate findings
of Montmort and De Moivre see also Thatcher (1957), Takácz (1969)
and Kohli (1975a).
3) The sex ratio at birth (Montmort 1708/1713, pp. 280 – 285;
Shoesmith 1985a). I only dwell on Bernoulli’s indirect derivation of
the normal distribution (Sheynin 196821; 1970a, pp. 201 – 203). Let
the sex ratio be m/f, n, the total yearly number of births, and µ and (n –
µ), the numbers of male and female births in a year. Denote

59
59
 n/(m + f) = r, m/(m + f) = p, f/(m + f) = q, p + q = 1,

and let s = 0(√n). Then Bernoulli’s derivation (Montmort 1708/1980,

pp. 388 – 394) can be presented as follows:

P(|µ– rm| ≤ s) ≈ (t – 1)/t,

t ≈ [1 + s(m + f)/mfr]s/2 ≈ exp[s2(m + f)2/2mfn],
P (|µ – rm| ≤ s) ≈ 1 – exp(s2/2pqn),

P[|µ – np|/ npq ≤ s] ≈ 1 – exp(–s2/2).

This result does not however lead to an integral theorem since s is

restricted (see above) and neither is it a local theorem; for one thing, it
lacks the factor 2 /  22.
4) The Petersburg game. In a letter to Montmort, Bernoulli (Ibidem,
p. 402) described his invented game. B throws a die; if a six arrives at
once, he receives an écu from A, and he obtains 2, 4, 8, … écus if a six
only occurs at the second, the third, the fourth, … throw. Determine
the expectation of B’s gain. Gabriel Cramer insignificantly changed
the conditions of the game; a coin appeared instead of the die, and the
occurrence of heads (or tails) has been discussed ever since. The
expectation of gain became

Eξ = 1∙1/2 + 2∙1/4 + 4∙1/8 + … = ∞, (5)

although a reasonable man will never pay any considerable sum in

exchange for it.
This paradox is still being examined. Additional conditions were
being introduced; for example, suggestions were made to neglect
unlikely gains, i.e., to truncate the series (5); to restrict beforehand the
possible payoff; and, the most interesting, to replace expectation by
moral expectation23. In addition, Condorcet (1784, p. 714) noted that
the possibly infinite game nevertheless provided only one trial and
that only some mean indicators describing many such games could
lead to an expedient solution. Actually issuing from the same idea,
Freudenthal (1951) proposed to consider a number of games with the
role of the gamblers in each of them to be decided by lot. Finally, the
Petersburg game caused Buffon (1777, § 18) to carry out the
apparently first statistical experiment. He conducted a series of 2048
games; the mean payoff was 4.9 units, and the longest duration of play
(in six cases), nine throws24. From a theoretical point of view, the
game was interesting because it introduced a random variable with an
infinite expectation.

Notes
1. Bernoulli (1713a/1999, p. 233) had additionally explained this expression by
the Greek word stochastice which Bortkiewicz (1917, p. X), with a reference to
him, put into scientific circulation. Already Wallis (1685, p. 254) had applied the
expression stochastic (iterative) process and Prevost & Lhuilier (1799, p. 3)
mentioned stochastics, or “l’art de conjecturer avec rigueur sur l’utilité et l’étendue
[and the extensiveness] du principe par lequel on estime la probabilité des causes”.
Hagstroem (1940) indicated that Plato and Socrates had applied the term

60
60
stochastics and that the Oxford English Dictionary had included it with a reference
to a source published in 1662.
2. Bernoulli had not invariably applied this term, see § 3.1.2.
3. Bearing in mind the published subject of that part, it would have been
expedient to isolate the mentioned applications of the art of conjecturing.
4. Concerning the use of probable inferences in civil life see also the opinion of
Cicero in Note 6 to Chapter 1. And, in connection with Leibniz’ understanding of
randomness (below), I recall Cicero (1991, Buch 2, § 17, p. 149) once more:
“Nothing is more opposed to calculation and regularity than chance”.
The publishers appended to the AC the author’s French contribution Lettre à un Amy
sur les Parties du Jeu de Paume (J. Bernoulli 1975) in which he calculated the
expectations of winnings in a time-honoured variety of tennis.
5. I mentioned it in §§ 2.1.1 and 2.1.2.
6. In 1714, in a letter to one of his correspondents, Leibniz (Kohli 1975b, p. 512)
softened his doubts about the application of statistical probabilities and for some
reason added that the late Jakob Bernoulli had “cultivated” [the theory of
probability] in accordance with his, Leibniz’, “exhortations”.
7. On Bernoulli’s non-additive probabilities see Shafer (1978) and Halperin
(1988).
8. See Arnauld & Nicole (1662/1992, p. 327): we should choose the more
probable.
9. I have found there only one appropriate but not really convincing example, see
§ 2.1.2. On their p. 281 these authors mention the possibility of posterior reasoning.
10. Cf. his Example 4 (Bernoulli 1713a/1999, p. 236).
11. Leibniz (1668 – 1669 (?)/1971, pp. 494 – 500 reasoned on the application of
moral certainty in theology. One of its chapters should have included the expression
“infinite probability or moral certainty”. In a later manuscript of 1693 he (Couturat
1901, p. 232), unfortunately, as it seems, isolated logical certainty, physical certainty
or “logical probability”, and physical probability. His example of the last-mentioned
term, the southern wind is rainy, apparently described a positive correlative
dependence.
12. De Moivre (§ 4.1) was the first to mention independence.
13. Markov had not applied that formula apparently because Bernoulli did not
yet know it.
14. Strengthened by the prolonged oblivion of De Moivre’s finding (§ 4.3).
15. Throughout Part 4, Bernoulli considered the derivation of the statistical
probability of an event given its theoretical probability and this is most definitely
seen in the formulation of his Main Proposition (the LLN) in Chapter 5. However,
both in the last lines of that chapter, and in Chapter 4 he mentioned the inverse
problem and actually alleged that he solved it as well. I return to this point in my
Chapter 5.
16. Cournot (1843, §11) also mentions them vaguely.
17. For us, Oresme’s understanding of incommensurability is unusual, but I do
not dwell on this point. Before him, Levi ben Gerson (1999, p. 166) stated that the
heavenly bodies will be unable to return to their initial position if their velocities
were incommensurable. He had not, however, mentioned the end of the world.
18. Bernoulli possibly thought about that expression when choosing a title for his
book (and for the new discipline of the same name, the predecessor of the theory of
probability).
19. Huygens’ appropriate reasoning (§ 2.2.2) appeared in print much later.
20. Cf. Daniel Bernoulli’s moral expectation (§ 6.1.1).
21. Only in its reprint of 1970 (p. 232).
22. Nevertheless, A. P. Youshkevich (1986) reported that at his request three
mathematicians, issuing from the description offered by Hald, had concluded that
Bernoulli had come close to the local theorem. Neither had Hald (1998, p. 17)
mentioned that lacking factor.
23. See Daniel Bernoulli’s memoir of 1738 in § 6.1.1. He published it in
Petersburg, hence the name of the game.
24. O. Spieß (1975) dwelt on the early history of the Petersburg game and Jorland
(1987) and Dutka (1988) described later developments. Dutka also adduced the
results of its examination by means of statistical simulation.

61
61
 4. De Moivre and the De Moivre – Laplace Limit Theorem
4.1. The Measurement of Chance (1712)
In his first probability-theoretic work De Moivre (1712/1984)
justified the notion of expected random gain by common sense rather
than defining it formally as has been done later, cf. § 2.2.2; introduced
the multiplication theorem for chances (mentioning independence of
the events) and applied the addition theorem, again for chances; and,
in solving one of his problems (No. 26), applied the formula (3.4) of
inclusion and exclusion. I describe some of his problems; I have
mentioned Problem 14 (repeated in De Moivre’s Doctrine of chances)
in § 2.2.2.
1) Problem No. 2. Determine the chances of winning in a series of
games for two gamblers if the number of remaining games is not
larger than n, and the odds of winning each game are a/b. De Moivre
notes that the chances of winning are as the sums of the respective
terms of the development of (a +b)n.
2) Problem No. 5. The occurrence of an event has a chances out of
(a + b). Determine the number of trials (x) after which it will happen,
or not happen, with equal probabilities1. After determining x from the
equation

(a + b)x – bx = bx,

De Moivre assumes that a/b = 1/q, q → ∞, and obtains

1 + x/q + x2/2q2 + x3/6q3 + … = 2, x = qln2, (1)

whose left part resembles the Poisson distribution.

3) A lemma. Determine the number of chances for the occurrence
of k points in a throw of f dice each having n faces. Later De Moivre
(1730, pp. 191 – 197; 1718 and 1756, Problem No. 3, Lemma) solved
this problem by means of a generating function of a sequence of
possible outcomes of a throw of one die.
4) Problem No. 9 (cf. Pascal’s problem from § 2.2.2). Gamblers A
and B have p and q counters, and their chances of winning each game
are a and b, respectively. Determine the odds of their ruining. By a
clever trick that can be connected with the notion of martingale
(Seneta 1983, pp. 78 – 79) De Moivre obtained the sought ratio:

PA/PB = aq(ap – bp) ÷ bp(aq – bq). (2)

He left aside the elementary case of a = b.

5) Problem No. 25. Ruining of a gambler during a finite number of
games played against a person with an infinite capital. De Moivre
described the solution in a generalized way; its reconstruction is due
to Hald (1990, pp. 358 – 360). The conditions of the problem had not
been formulated explicitly.
4.2. Life Insurance
De Moivre first examined life insurance in the beginning of the
1720s and became the most influential author of his time in that field.
Issuing from Halley’s table (§ 2.1.4), he (1725/1756, pp. 262 – 263)

63
62
assumed a continuous uniform law of mortality for all ages beginning
with 12 years and a maximal duration of life equal to 86 years. I
describe some of those of his numerous findings which demanded the
application of the integral calculus.
1) Determine the expected duration of life for a man of a given age
if the maximal duration, or complement of life is n (n = 86 – age). The
answer is n/2 (p. 288). Reconstruction:
n

 xdx / n = n/2.
0

2) Determine the probability of one person outliving another one if

the complements of their lives are n and p, n > p (p. 324). Here, in
essence, is De Moivre’s solution. Let the random durations of the lives
of A and B be ξ and η. Then, since at some moment x the complement
of A’s life is (n – x),

P(ξ ≥ x, η = x) = [(n – z)/n]dz/p, P(ξ > η) =

p


0
[(n – z)/n]dz/p = 1 – p/2n.

3) Determine the expected time Eζ during which both men with the
same complements of life as in the previous problem do not die (p.
288). De Moivre only provided the answer; a reconstruction (Czuber,
Note 22 to the German translation of 1906 of De Moivre’s work) is as
follows.

P(x ≤ ξ ≤ x + dx or x ≤ η ≤ x + dx) =
[(n – x)/n]dx/p + [(p – x)p]dx/n,
p

Eζ =  {[(n – x)/np] + [p – x)p/n]}dx = p/2 – p2/6n.

0

Note that probabilities of the type of P (ξ ≥ x) easily lead to integral

distribution functions.
Hald (1990, pp. 515 – 546) described in detail the work of De
Moivre and of his main rival, Simpson (1775), in life insurance.
Simpson improved on, and in a few cases corrected the former’s
findings. After discussing one of the versions of mutual insurance,
Hald (p. 546) concluded that Simpson’s relevant results represented
“an essential step forward”. However, the relations between De
Moivre and Simpson became horrible and K. Pearson (1978) justly
called Simpson a most disreputable character (p. 145), an unblushing
liar and a thorough knave at heart (p. 184).
4.3. The Doctrine of Chances (1718, 1738, 1756)
This work published in three editions, in 1718, 1738, and,
posthumously, in 1756 (reprinted in 1967), was De Moivre’s main
achievement. He developed it from his previous memoir (§ 4.1) and
intended it for gamblers so that many results were provided there
without proof. This fact together with other circumstances2 caused this
extremely important book, whose translation into French

64
63
contemplated both Lagrange and Laplace3, to remain barely known
for many decades. I refer to the reprint of its last edition.
In his Introduction, De Moivre listed his main methods:
combinatorial analysis, recurrent sequences (whose theory he himself
developed) and infinite series; in particular, he applied appropriately
truncated divergent series. Also in the Introduction, on pp. 1 – 2 he
provided the “classical” definition of probability, usually attributed to
Laplace, but kept to the previous reasoning on expectation (§ 4.1) and
even introduced the value of expectation (p. 3); formulated the
multiplication theorem for probabilities (not for chances, as
previously) and, in this connection, once more mentioned
independence. Two events, A and B, were independent, if, as he
stated,

P(B) = P(B/A), P(A) = P(A/B)

(modern notation here and below). For dependent events (p. 6), three
in number (say),

P(A·B·C) = P(A) P(B/A)P(C/A·B). (3)

I list now some of the problems from the Doctrine mentioned by

Hald (1990, pp. 409 – 413) without repeating those described in § 4.1
and, for the time being, leaving aside the normal distribution.
1) The Huygens additional Problem No. 4 (§ 2.2.2) including the
multivariate case. The appearance of the hypergeometric distribution:
Problems NNo. 20 and 26.
2) Runs of successes in n Bernoulli trials including the case of
n → ∞: Problems NNo. 34 and 74.
3) Coincidences. A generalization of Montmort’s findings (§ 3.3.3)
by the method of inclusion and exclusion: Problems 35 and 36.
4) The gambler’s ruin: Problems 58 – 71.
5) Duration of game: Problems 58 – 64, 68 – 71.
For the general reader the main merit of the Doctrine was the study
of many widely known games whereas De Moivre himself, in
dedicating its first edition to Newton (reprinted in 1756 on p. 329),
perceived his main goal in working out

A Method of calculating the Effects of Chance […] and thereby

fixing certain rules, for estimating how far some sort of Events may
rather be owing to Design than Chance […] [so as to learn] from your
Philosophy how to collect, by a just Calculation, the Evidences of
exquisite Wisdom and Design, which appear in the Phenomena of
Nature throughout the Universe.

I stress that De Moivre wrote this dedication before proving his

limit theorem (§ 4.4). See Pearson’s statement on Newton’s influence
in § 2.2.3. The aim of De Moivre’s theory of probability was therefore
the separation of necessity and randomness of any kind.

65
64
 4.4. The De Moivre – Laplace Theorem
In 1730 De Moivre published his Miscellanea analytica. Later he
appended two supplements; I am interested in the second one (1733)4,
which he printed in a small number of copies and sent out to his
colleagues. In 1738 De Moivre translated it into English and included
in the second, and then, in an extended form, in the third edition of the
Doctrine (pp. 243 – 254 in 1756). Its title includes the words binomial
(a + b)n which means that, although studying the particular case of the
symmetric binomial, De Moivre thought about the general case. He (p.
250) also expressly and justly stated that the transition to the general
case was not difficult. Strangely enough, even recently some authors
(Schneider 1988, p. 118) maintained that De Moivre had only
considered the particular case.
The date of the compilation of this supplement is known: on the
first page of its Latin original De Moivre stated that he had concluded
(at least its mathematical part) about 12 years earlier, i.e., much earlier
than he published his Misc. anal., but we can only follow that
contribution (Sheynin 1970a).
1) In Book 5 of the Misc. anal. De Moivre determined the ratio of
the middle term of the symmetric binomial to the sum of all of its
terms, and in the first supplement to that work he derived,
independently from, and simultaneously with Stirling, the so-called
Stirling formula. Only the value of the constant, 2π , the latter
communicated to him5.
2) In the same Book, De Moivre calculated the logarithm of the
ratio of the middle term of the binomial (1 + 1)n to the term removed
by l from it:

(m + l – 1/2)ln(m + l – 1) + (m – l + 1/2)ln(m – l + 1)
– 2mlnm + ln[(m +l)/m], m = n/2.

However, only in the second supplement De Moivre transformed this

expression, obtaining, as n → ∞, – 2l2/n. The ratio itself thus became
equivalent to

1 – 2l2/n + 4l4/2n2 – ... (4)

Actually, as corroborated by his further calculations, De Moivre

thought about the inverse ratio.
3) Also in the same supplement, after integrating (4), De Moivre
calculated the ratio of the sum of the terms between the middlemost
and the one removed from it by l to the sum of all the terms. It was
equal to

[2/ 2πn ] (l – 2l3/1·3n + 4l5/2·5n2 – …). (5)

He then calculated this sum either by numerical integration, or, for

l < √n/2, by leaving only a few of its first terms. For n → ∞ his main
result can be written in modern notation as

66
65
   np
b
1
lim P[a ≤
npq
≤ b] =
2 
a
exp(–z2/2) dz. (6)

This is the integral De Moivre – Laplace theorem (see § 7.1-3), as

Markov (1924, p. 53) called it, – a particular case of the CLT, a term
introduced by Polya (1920). Note that neither De Moivre, nor Laplace
knew about uniform convergence that takes place here.
Todhunter (1865, pp. 192 – 193) inadequately described the
essence of De Moivre’s finding. He failed to note that De Moivre had
actually considered the general case of p ≠ q and only stated that “by
the aid of Stirling’s Theorem the value of Bernoulli’s Theorem is
[was] largely increased”6. De Morgan (1864) and then Eggenberger
(1894) were the first to note that De Moivre had arrived at the [normal
distribution]. Concerning De Morgan see Note 1 to Chapter 8.
De Moivre (1718/1756, p. 252) mentioned the study of the sex ratio
at birth (§ 2.2.4) and illustrated it by imagined throws of 14 thousand
dice each having 35 faces and painted in two colours, 18 faces white
and 17 black7. His reasoning (and his general considerations on p.
251) meant that, for him, the binomial distribution was a divine law of
nature, stochastic only because of possible deviations from it. De
Moivre thus actually recognized the mutual action of necessity and
randomness, cf. § 3.2.4.

Notes
1. De Moivre thus made use both of chances and probability.
2. De Moivre’s symbolism soon became dated; the English language had been
little known on the Continent; Todhunter, the most influential historian of
probability of the 19th century, inadequately described De Moivre’s main finding (§
4.4); and, last but not least, Laplace (1814/1995, p. 119) did not sufficiently explain
it.
3. Lagrange’s letter to Laplace of 30 Dec. 1776 in t. 14 of his Oeuvres (1892), p.
66.
4. I call this memoir a supplement only for the sake of tradition: its extant copies
in large libraries were bound to the Misc. anal.
5. In the same supplement De Moivre included a table of lg n! for n = 10 (10) 900
with 14 decimals; reprinted: (1718/1756, p. 333). Eleven or twelve decimals were
correct; a misprint occurred in the value of lg 380!.
6. In 1740, Simpson directly dwelt on the general case (Hald 1990, pp. 21 – 23).
7. Regular 35-hedrons do not exist, but this is not here important. De Moivre
thought about 14 thousand yearly births with m:f = 18:17.

67
66
 5. Bayes
5.1. The Bayes Formula and Induction
I dwell on the posthumous memoir (Bayes 1764 – 1765) complete
with the commentaries by Price. In its first part Bayes introduced his
main definitions and proved a few theorems; note that he defined
probability through expectation. There was no hint of the so-called
Bayes theorem

P(B /Ai )P(Ai )

P(Ai/B) = n
. (1)
 P(B /A )P(A )
j 1
j j

Lubbock et al (1830, p. 48) first applied the term itself, as noted by

H. A. David, see H. A. David & Edwards (2001, p. 215). He referred
to an oral communication from Hald. I return to formula (1) in §§ 7.1-
1 and 9.2-2. Here I indicate that Bayes had in essence introduced
induction into probability and that his approach that assumed the
existence of prior probabilities or distributions (see below) greatly
influenced the development of mathematical statistics1.
Bayes then studied an imaginary experiment, a ball falling on point
r situated in a unit square ABCD, “to the left” or “to the right” of some
straight line MN parallel to, and situated between AB and CD. If, after
(p + q) trials, the point r occurred p times to the right of MN and q
times, to the left of it, then

c 1

P(b  r  c) =  u p (1  u ) q du ÷ v
p
(1  v )q dv (2)
b 0

where bc is a segment within AD. Bayes derived the denominator of

(2) obtaining the value of the [beta-function] B(p + 1; q + 1) and
spared no effort in estimating its numerator. The right side of (2) is
now known to be equal to the difference of two values of the
incomplete beta-function

Ic(p + 1; q + 1) – Ib(p + 1; q + 1).

Thus, given the results of the experiment, and assuming a uniform

prior distribution2 of the location of MN and r, he determined the
appropriate theoretical probability considered as a random variable.
In his covering letter, Price provided purely methodical
illustrations; the most interesting of them required the [probability] of
the next sunrise observed 106 times in succession. Formula (2)
indirectly answers his question if b = 1/2 and c = 1 are chosen; it also
provides the probability of the contrary event if b = 0 and c = 1/2.
Price (Bayes 1764/1970, pp. 149 and 150 – 151) also solved the same
question for p = 1 and q = 0 and obtained P = 3/4 which is doubtful:
knowing nothing about the essence of a phenomenon we should have
got P = 1/2 (cf. Poisson’s reasoning in § 8.1). In this case, formula (2)
is wrong. Note also that Chebyshev (1879 – 1880/1936, p. 158)
formulated the same problem on an everyday level: To determine the

68
67
probability of a student’s successful answer to the next question after
his previous successes.
I dwell somewhat on the above. The probability (this time, the
actual probability) of the next sunrise is

p 1
1 1

P =  x p 1dx ÷ x
p
dx =
0 0
p2

(cf. §7.1-5). Now, P(p = 1) = 2/3, not good enough either.

Polyá (1954, p. 135) remarked that each consecutive success
(sunrise) provided ever less justification for the next one.
Cournot (1843, § 93) considered a similar problem: A woman gave
birth to a boy; determine the probability that her next child will also
be a boy. Without justification, he stated that “perhaps” the odds were
2:1 but that it was impossible to solve that problem. See the opinions
of Laplace (§§ 7.1-1 and 7.1-5), Gauss (§ 9A.2-2) and Chebyshev (§
13.2-7) about the Bayesian approach. Another point concerned the
Bayesian treatment of an unknown constant r in formula (2) as a
random variable (Neyman 1938/1967, p. 337).
Beginning with the 1930s and perhaps for three decades English
and American statisticians had been denying Bayes. I am, however,
leaving aside that period and I only note that the first and the main
critic of the Bayes “theorem” or formula was Fisher (1922, pp. 311
and 326) but that he had not specified what exactly did he refuse to
comply with. It seems that he disagreed with the introduction of
hardly known prior probabilities and/or with the assumption that they
were equal to one another, cf., however, Laplace’s general statement
about rectifying hypotheses (§ 7.2-1). The papers of Cornfield (1967)
and Barnard (1967) were also insufficiently definite; the former
figuratively remarked, on his p. 41, that Bayes had returned from the
cemetery.
It is methodologically important to note that the inverse probability
defined by formula (1) is tantamount to conditional probability given
that the stipulated condition has indeed been fulfilled.
5.2. The Limit Theorem
I dwell now on the case of n = (p + q) → ∞ which Bayes had not
expressly discussed. Price, however, remarked that, for a finite n, De
Moivre’s results were not precise. I indicate in addition that another
posthumous note written by Bayes was published at the same time,
1764, as his main memoir, and in the same periodical, and there he
warned mathematicians about the danger of applying divergent series.
He had not named De Moivre, but he apparently had in mind the
derivation of the De Moivre – Laplace theorem (4.6) as well.
Timerding, the Editor of the German translation of the Bayes
memoir, nevertheless went on to consider the limiting case. He issued
from Bayes’ calculations made for large but finite values of p and q.
Applying a clever trick, he proved that, as n → ∞, the probability α of
the ball falling to the right of MN obeyed the proposition
|αa|
z
1
2 0
limP{ ≤ z} = exp(–w2/2)dw, (3)
pq /n3

69
68
where (not indicated by Timerding) a = p/n = Eα, pq/n3 = varα.
The functions in the left sides of formulas (4.6) and (3) are random
variables, centred and normed in the same way, and it is remarkable
that Bayes, without knowing the notion of variance, apparently
understood that (4.6) was not sufficiently precise for describing the
problem inverse to that studied by De Moivre. Anyway, Price (Bayes
1764/1970, p. 135) stated that he knew

of no person who has shewn how to deduce the solution of the

converse problem […]. What Mr De Moivre has done therefore
cannot be thought sufficient …

Jakob Bernoulli (Note 15 in my Chapter 3) maintained that his

formulas were also fit for solving the inverse problem – but how
precisely? True, De Moivre (1718/1756, p. 251) also mentioned the
inverse problem:

Conversely, if from numberless Observations we find the Ratio of

the Events to converge to a determinate quantity […], then we
conclude [how precisely?] that this Ratio expresses the determinate
Law according to which the Event is to happen.

In my opinion, this insufficiently known proposition is very

important. Together with the integral De Moivre – Laplace theorem it
completed the creation of the first version of the theory of probability.
Bayes was actually the main predecessor of Mises who had not,
however, mentioned him.
5.3. Additional Remark
Stigler (1983) quoted a curious statement (Hartley 1749, pp. 338 –
339) and interpreted it as a testimony against Bayes’ priority. After
referring to De Moivre, Hartley wrote, in part:

An ingenious friend has communicated to me a solution of the

inverse problem of determining the probability of an event given the
number of times it happened and failed.

Later Stigler (1986, pp. 98, 132) recalled Hartley and his own
earlier paper of 1983, but did not definitively repeat his previous
inference. Then, however, he (1999, pp. 291 – 301) reprinted that
paper and added a tiny footnote brushing aside all the criticism
published by that time.
Stigler inferred that the author of the Bayes’ theorem was
Saunderson (1682 – 1739), and by applying formula (1), he even
found that his conclusion was three times more probable than the
former opinion. However, he based himself on subjective estimates
and in particular assumed that the prior probabilities of the authorship
of Bayes and Saunderson were the same. This means that the extra-
mathematical arguments (for example, the evidence of Price, a close
friend of Bayes) are not considered at all. And it is opportune to recall
the opinion of Gauss (§ 3.1.2): applications of the theory of

70
69
probability can be greatly mistaken if the essence of the studied object
is disregarded.
In addition, not only a honest personality like Saunderson, but
almost any pretender will be able to claim equal prior rights with an
established author (or a politician of the past). For my part, I have
argued that it was Bayes himself who communicated to Hartley the
solution “of the inverse problem”. Referring to later sources, Zabell
(1989, p. 316) concluded that Stigler’s opinion was not serious.

Notes
1. A modern encyclopaedia (Prokhorov 1999b) contains 14 items mentioning
him; for example, Bayesian estimator, Bayesian approach, etc. There also, on p. 37,
the author of the appropriate entry mistakenly attributes formula (1) to Bayes.
2. Bayes himself had not stated that this distribution was uniform, but it is
nevertheless necessary to make this assumption (K. Pearson 1978, p. 364). Without
any explanation provided, Mises (1919, § 9.2) remarked that Bayes had considered
the general case as well. Following Czuber, to whom he referred, Mises proved that
the influence of non-uniformity of the prior distribution weakens with the increase in
the number of observations.

71
70
 6. Other Investigations before Laplace
6.1. Stochastic Investigations
6.1.1. Daniel Bernoulli. He published a number of memoirs
pertaining to probability and statistics, and, before that, he (1735)
provided a stochastic reasoning on the structure of the Solar system.
The inclinations of the orbits of the five (excepting the Earth) then
known planets with respect to the Earth (considered as random
variables with a continuous uniform distribution) were small, and the
probability of a “random” origin of that circumstance, as he
concluded, was negligible. I dwelt on the logic of such considerations
in § 1.1.1; here, however, enters a new dimension (see § 10.8.4): it
was possible to study, instead of the inclinations, the arrangement of
the poles of the orbits (Todhunter 1865, p. 223).
In this subsection, I consider only some of Bernoulli’s memoirs and
I postpone the study of his other work until §§ 6.2.3 and 6.3, but my
general conclusion is that he, together with De Moivre, was the main
predecessor of Laplace.
1) Moral expectation. While attempting to explain the paradoxical
nature of the Petersburg game (§ 3.3.4), Bernoulli (1738) suggested
that the gain y of a gambler was determined by his winnings x in
accord with the differential equation (the first such equation in
probability)

dy = cdx/x, c > 0, so that y = f(x) = cln(x/a)

where a was the gambler’s initial capital.

Bernoulli also proposed that the expected winnings
(p1x1 + p2x2 + … + pnxn)/Σpi where pi were the appropriate
probabilities be replaced by their “moral expectation”

Σpif(xi)/Σpi.

He indicated but had not proved (see § 7.1-9) that even a “just”
game with a zero expected loss for each participant became
disadvantageous because the moral expectation of winnings, again for
each, was negative, and that the infinite expected gain in the
Petersburg game (3.5) can be replaced by a finite moral expectation.
Then, applying his innovation to a study of marine shipping of freight,
he maintained (again, without proof, see same subsection below) that
the freight should be evenly distributed among several vessels.
Bernoulli appended the text of a letter of 1732 from Gabriel Cramer
to Nikolaus Bernoulli which contained his (not Daniel’s) term moral
expectation. Cramer also indirectly suggested the choice of

f(x) = min (x; 224) or f(x) = x.

In a letter of 1742 Bernoulli left a curious statement concerning

politics (P. N. Fuss 1843/1968, t. 2, p. 496): “I believe that
mathematics can also be rightfully applied in politics”. After having
referred to a positive opinion of Maupertuis, he continued:

72
71
 An entirely new science will emerge provided that as many
observations are made in politics as in physics.

With all respect, this is at least unclear.

Moral expectation had become popular and Laplace (1812/1886, p.
189) therefore proposed a new term for the previous “usual”
expectation calling it mathematical; his expression persists to no
purpose at least in the French and Russian literature and I refuse to
apply it. Modern statistical damage functions are akin to moral
expectation and moreover, at the end of the 19th century, issuing from
Bernoulli’s idea, economists began to develop the theory of marginal
utility thus refuting Bertrand’s opinion (1888a, p. 66) that moral
expectation was useless:

The theory of moral expectation became classical, and never was a

word used more exactly. It was studied and taught, it was developed in
books really celebrated. With that, the success came to a stop; no
application was made, or could be made, of it.

2) A limit theorem. While studying the same problem concerning

the sex ratio at birth (§§ 2.2.4, 3.3.4, 4.4), Bernoulli (1770 – 1771), in
the first part of his memoir, assumed that male and female births were
equally probable. It followed that the probability that the former
constituted a half of 2N births will be

P = [1·3·5·…·(2N – 1)] ÷ [2·4·6·…·2N] = q(N).

He calculated this fraction not by the Wallis formula but by means

of differential equations. After deriving q(N – 1) and q(N + 1) and the
two appropriate values of Δq, he arrived at

dq/dN = – q/(2N + 2), dq/dN = – q/(2N – 1)

and, “in the mean”, dq/dN = – q/(2N + 1/2). Assuming that the
solution of this equation passed through point N = 12 and q(12) as
defined above, he obtained

q = 1.12826/ 4 N  1 .

Application of differential equations was Bernoulli’s usual method

in probability, also see § 6.1.1-1.
Bernoulli also determined the probability of the birth of
approximately m boys (see below):

P(m = N ± µ) = qexp(– µ 2/N) with µ = 0(√N). (1)

In the second part of his memoir Bernoulli assumed that the

probabilities of the birth of both sexes were in the ratio of a:b.
Equating the probabilities of m and (m + 1) boys being born, once
more for 2N births, he thus obtained the [expected] number of male
births

73
72
 2 Na  b 2 Na
Em = M = 
ab ab

which was of course evident. More interesting was Bernoulli’s

subsequent reasoning for determining the probability of an arbitrary m
(for µ of the order of √N):

P(m = M + µ + 1) – P(m = M + µ) ≡ dπ =
2N  M  
π – (a/b)π dµ,
M   1
μ  1  μa /b
– dπ/π = dµ.
m  μ 1

The subsequent transformations included the expansion of

ln[(M + 1 + µ)/(M + 1)] into a power series. Bernoulli’s answer was

( a  b) 2
P(m = M  µ) = π = P(m = M) exp [– ],
2bM

hence (1). Note that Bernoulli had not applied the local De Moivre
(– Laplace) theorem.
Issuing from some statistical data, Bernoulli compared two possible
ratios a/b but had not made a final choice in favour of either of them.
He also determined such a value of µ that the sum of probabilities (1),
beginning from µ = 0, equalled one half. Applying summation rather
than integration, he had not therefore arrived at an integral limit
theorem and (also see above) he did not refer to, and apparently had
not known about De Moivre’s findings. This shows, once again (cf. §
4.4), that they had for a long time been forgotten.
3) Urn problems. I consider two of these. An urn contains n pairs of
white and black stripes. Determine the number (here and below,
actually, the expected number) of paired stripes left after (2n – r)
extractions without replacement. By the combinatorial method
Bernoulli (1768a) obtained

x = r(r – 1)/(4n – 2); and x = r2/4n if n = ∞.

He derived the same result otherwise: when r decreases by dr the

corresponding dx is either zero [(r – 2x) cases] or dr (2x cases) so that

dx = [(r – 2x)∙0 + 2x∙dr]/r, x = r2/4n since r = 2n if x = n.

Bernoulli then generalized his problem by considering unequal

probabilities of extracting the stripes of different colours and he
(1768b) applied his findings to study the duration of marriages, a
subject which was directly linked with insurance of joint lives.
Suppose now that each of two urns contains an equal number n of
balls, white and black, respectively. Determine the number of white
balls in the first urn after r cyclic interchanges of one ball. Bernoulli

74
73
(1770) solved this problem by the same two methods. Thus, he issued
from the differential equation

dx = – xdr/n + [(n – x)/n]dr so that x ≈ (1/2)n[1 + e– 2r/n].

Bernoulli then combinatorially considered the case of three urns

with balls of three different colours. He noted that the number of
white balls in the first urn was equal to the sum of the first, the fourth,
the seventh, … terms of the development of [(n – 1) + 1]r divided by
nr–1. For the other urns he calculated, respectively, the sums of the
second, the fifth, the eighth, …, and the third, the sixth, the ninth, …
terms. For the first urn he obtained

1
A= [(n – 1)r + Cr3 (n – 1)r–3 + Cr6 (n – 1)r–6 + ...] ≈ ne– r/n S. (2)
n r 1

The expression designated by S obeyed the differential equation

Sdr3/n3 = d3S

and was therefore equal to

S = aer/n + be– r/2n sin(r√3/2n) + ce– r/2n cos(r√3/2n)

where, on the strength of the initial conditions, a = 1/3, b = 0, c = 2/3.

Bernoulli derived similar expressions for the other urns, calculated
the number of extractions leading to the maximal number of white
balls in the first urn, and, what is extremely interesting, he also noted
the existence of a limiting state, of an equal number of balls of each
colour in each urn. This can be easily verified by referring to the
theorem on the limiting transition matrix in homogeneous Markov
chains. Physicists could have noticed here a stochastic model of the
previously received notion of the thermal death of (a finite) universe.
Bernoulli obtained formula (2) by issuing from differential
equations

dx = – xdr/n + [n – (x + y)]dr/n, dy = – ydr/n + xdr/n

where x, y, and [n – (x + y)] were the numbers of white balls in the

urns after r interchanges1. I return to this problem in §7.1-3; here, I
note that Todhunter (1865, pp. 231 – 234) simplified Bernoulli’s
solution and made it more elegant. He wrote the differential equations
as

dx = (dr/n) (z – x), dy = (dr/n) (x – y), dz = (dr/n) (y – z)

and noted that the sum S was equal to

S = (1/3)[eαr/n + eβr/n + eγr/n]

3
with α, β, γ being the values of 1.

75
74
 Bernoulli’s x in his first problem, and his S and A from (2) depend
on discrete “time” r/n, which is characteristic of stochastic processes
with non-homogeneous time. The same is true about the ratio s/ in §
6.2.3 below.
6.1.2. D’Alembert. In the theory of probability, he is mostly known
as the author of patently wrong statements2. Thus, he (1754)
maintained that the probability of heads appearing twice in succession
was equal to 1/3 rather than to 1/4. Then, he (1768a) reasoned on the
difference between “mathematical” and “physical” probabilities3,
stating without justification that, for example, after one of two
contrary events had occurred several times in succession, the
appearance of the other one becomes physically more probable. He
was thus ridden by prejudices which Montmort had already
mentioned and which Bertrand later refuted by a few words (§ 2.1.1).
At the same time, D’Alembert recommended to determine
probabilities experimentally but had not followed his own advice
(which saved him from revealing his mistakes). Finally, he (1768b)
denied the difference (perfectly well understood by Huygens, § 2.2.2)
between the mean, and the probable durations of life and even
considered its existence as an (additional) argument against the theory
of probability itself.
It is opportune to recall Euler’s opinion as formulated in one of his
private letters of 1763 (Juskevic et al 1959, p. 221): D’Alembert tries
“most shamelessly to defend all his mistakes”. Anyway, D’Alembert
(1768d, pp. 309 – 310) did not ascribe the theory of probability to “a
precise and true calculus with respect either to its principles or
results”4.
On the other hand, D’Alembert thought that, in a single trial, rare
events should be considered unrealizable (Todhunter 1865, § 473)
and that absolute certainty was qualitatively different from “the
highest probability”. It followed from the latter statement that, given a
large number of observations, an unlikely event might happen (cf. the
strong law of large numbers), and, taken together, his considerations
meant that the theory of probability ought to be applied cautiously.
D’Alembert (1768c) also reasonably objected to Daniel Bernoulli’s
work on prevention of smallpox and formulated his own pertinent
ideas (§ 6.2.3). I ought to add that D’Alembert was indeed
praiseworthy for his work in other branches of mathematics (and in
mechanics); note also that Euler had not elaborated his likely correct
remark. On D’Alembert’s work see also Yamazaki (1971).
6.1.3. Lambert. He was the first follower of Leibniz in attempting
to create a doctrine of probability as a component of a general
teaching of logic. Like D’Alembert (Note 3 to this Chapter), Lambert
explained randomness by ignorance of causes, but he also stated that
all digits in infinite decimal developments of irrational numbers were
equally probable, which was an heuristic approach to the notion of
normal numbers, and he formulated a modern-sounding idea about the
connection of randomness and disorder (Lambert 1771, § 324; 1772 –
1775), also see Sheynin (1971a, pp. 238 – 239; 1971b, p. 246; 1974,
pp. 136 – 137).

76
75
 His considerations were forgotten, only Cournot (1851/1975, § 33
note) and Chuprov (1909/1959, p. 188) mentioned them.
Lambert did not go out of the confines of uniform randomness. To
put his ideas in perspective, I ought to add that the philosophical
treatises of the 18th century testify to the great difficulties experienced
in generalizing the notion of randomness, also see § 2.2.4. One
example: even in the 19th century, many scientists, imagining that
randomness was only uniform, refused to recognize the evolution of
species, and two authors (Baer 1873, p. 6; Danilevsky 1885, pt. 1, p.
194) independently mentioned the philosopher depicted in Gulliver’s
Travels (but borrowed by Swift from Raymond Lully, 13th – 14th
centuries). That “inventor”, hoping to get to know all the truths, was
putting on record each sensible chain of words that appeared from
among their uniformly random arrangements.
6.1.4. Buffon. He is mostly remembered for his definitive
introduction of geometric probabilities (§ 6.1.6). Then, he reasonably
suggested that the value of winnings in a game of chance diminished
with the increase of the gambler’s capital (cf. § 6.1.1) and
experimentally studied the Petersburg game (§ 3.3.4), proposed the
value 1/10,000 as a universally negligible probability, attempted to
determine the probability of the next sunrise (see Chapter 5)5, cf. §
7.1-5, and compiled tables of mortality which became popular.
Negligible, as he thought, was the probability of death of a healthy
man aged 56 during the next 24 hours, but his figure was apparently
too low; K. Pearson (1978, p. 193) thought that 1/1,000 would have
been more appropriate. In addition, negligibility ought to be only
chosen for a particular event rather than assigned universally. All the
above is contained in Buffon’s main work (1777). There also (§ 8,
Note) he published the text of his letter of 1762 to Daniel Bernoulli
which contained an embryo of Quelelet’s celebrated Average man
(see my § 10.5):

Mortality tables are always concerned with the average man, that
is, with people in general, feeling themselves quite well or ill, healthy
or infirm, robust or feeble.

6.1.5. Condorcet. He attempted to apply the theory of probability

to jurisprudence in the ideal and tacitly assumed case of independent
judgements made by jurors or judges. He also estimated the
trustworthiness of testimonies and critically considered electoral
problems. His main method was the application of difference
equations. Todhunter (1865, pp. 351 – 410) described the work of
Condorcet in detail and concluded (p. 352) that in many cases it was
“almost impossible to discover” what he had meant6: “The obscurity
and self contradiction are without any parallel …” He, Todhunter, will
provide some illustrations, “but no amount of examples can convey an
adequate impression of the extent of the evils”. At the very least,
however, Laplace and Poisson continued to apply probability to
jurisprudence and certainly profited to some extent from the work of
Condorcet. Poisson (1837a, p. 2) mentioned his ideas quite
favourably.

77
76
 I note however that, while discussing games of chance, Condorcet
(1785b/1847, p. 561) expressed himself rather unfortunately, and
stated on the next page without any justification that Daniel Bernoulli
had not removed all the objections to the “rule” of expectation and
that this was allegedly achieved by D’Alembert. In 1772, in a letter to
Turgot, he (Henry 1883/1970, pp. 97 – 98) told his correspondent that
he was “amusing himself” by calculating probabilities, had compiled
“a booklet on that subject” and was keeping to the opinions of
D’Alembert. That booklet is unknown. On Condorcet see also Zabell
(1988b) and Yamazaki (1971).
6.1.6. Geometric Probabilities. These were decisively introduced
in the 18th century although the definition of the notion itself, and, for
that matter, only on a heuristic level, occurred in the mid-19th century
(§ 10.3). Newton (§ 2.2.3) was the first to think about geometric
probability; Daniel Bernoulli (§ 6.1.1) tacitly applied it in 1735 as did
somewhat later (possibly even earlier) De Moivre (1725/1756, p.
323), and then T. Simpson (1757) (§ 6.3.1) and Bayes (§ 5.1).
Dealing with the continuous uniform distribution, De Moivre
assumed, for example, that if 0 < ξ < b and 0 < a < b,
then P (0 < ξ < a) = [0; a] ÷ [0; b].
Simpson noted that in his case (a continuous triangular distribution)
probabilities were proportional to the areas of the appropriate figures.
Bayes assumed that, for a continuous uniform distribution, the
probabilities of a ball falling on different equal segments were equal
to one another.
The Michell (1767) problem became classical: Determine the
probability that two stars from all of them, uniformly distributed over
the celestial sphere, were situated not farther than 1° from each other.
Choose an arbitrary point (A) on a sphere with centre O and imagine a
circle perpendicular to OA having distance 1° from A. The probability
sought is the ratio of the surface of the spherical segment thus
obtained to that of the sphere. Newcomb and Fisher calculated the
expected number of closely situated stars (§ 10.8-4) and general issues
were also debated by others. Thus, Proctor (1874, p. 99) wished to
determine

What peculiarities of distribution might be expected to appear

among a number of points spread over a plane surface perfectly at
random.

His was a question which concerned the deviations of an empirical

density curve from its theoretical counterpart and now belonged to
mathematical statistics. And Bertrand (1888a, pp. 170 – 171)
remarked that without studying other features of the sidereal system it
was impossible to decide whether stars were arranged randomly.
Buffon (§ 6.1.4) expressly studied geometric probability; the first
report on his work (Anonymous 1735) had appeared long before his
contribution. Here is his main problem: A needle of length 2r falls
“randomly” on a set of parallel lines. Determine the probability P that
it intersects one of them. It is easily seen that

78
77
 P = 4r/πa (3)

where a > 2r is the distance between adjacent lines. Buffon himself

had, however, only determined the ratio r/a for P = 1/2. His main aim
was (Buffon 1777/1954, p. 471) to “put geometry in possession of its
rights in the science of the accidental [du hasard]”. Many
commentators described and generalized the problem above. The first
of them was Laplace (§7.1-4) who noted that formula (3) enabled to
determine [with a low precision] the number π. I treat the further
history of geometric probability in my Chapter 12.
6.2. Statistical Investigations
6.2.1. Staatswissenschaft (Statecraft, University Statistics). In
mid-18th century Achenwall (Sheynin 1997b) created the Göttingen
school of Staatswissenschaft which described the climate,
geographical situation, political structure and economics of separate
states and estimated their population by issuing from data on births
and mortality but did not study relations between quantitative
variables. Achenwall referred to Süssmilch (§ 6.2.2), advised state
measures fostering the multiplication of the population and
recommended censuses without which (1763/1779, p. 187) a
“probable estimate” of the population could be still got, see above. He
(1752/1756, Intro.) also left an indirect definition of statistics:

In any case, statistics is not a subject that can be understood at

once by an empty pate. It belongs to a well digested philosophy, it
demands a thorough knowledge of European state and natural history
taken together with a multitude of concepts and principles, and an
ability to comprehend fairly well very different articles of the
constitutions of present-day kingdoms [Reiche].

Achenwall’s student Schlözer (1804, p. 86) figuratively stated that

“History is statistics flowing, and statistics is history standing still”. In
spite of its elegance, this phrase is meaningless: Schlözer himself
(§ 14-3) stated that it was necessary to compare one state with another
and the changes in a state in time. And even Leibniz recommended
such comparisons in his manuscripts of the 1680s (Sheynin 1977b, p.
224).
However, for those keeping to Staatswissenschaft this pithy saying
became the definition of statistics which was thus not compelled to
study causal connections in society or discuss possible consequences
of innovations; which thus failed to adhere to the goals of political
arithmetic (§ 2.1.4). The second distinction between the two
disciplines consisted in that only political arithmetic was seriously
interested in studying population. Finally, the methods of investigation
were also different: not numbers, but wordy descriptions of
remarkable features lay at the heart of the works of the Göttingen
school.
Knies (1850, p. 24) quoted unnamed German authors who had
believed, in 1806 and 1807, that the issues of statistics ought to be the
national spirit, love of freedom, the talent and the characteristics of
great personalities and ordinary people of a given state. This criticism,

79
78
of course, had to do with the limitations of mathematics in general.
Recall (§ 1.1.2) that Moses wished to find out not only whether the
population of Canaan was numerous or not, but also whether the
people there were “strong or weak”.
Tabular statistics which had originated with Anchersen (1741)
could have served as an intermediate link between words and
numbers, but Achenwall was apparently opposed to it. Anyway, he
(1752, Intro.) stated that he had “experienced a public attack” against
the first edition of that book (published in 1749 under a previous title)
by Anchersen. “Tabular” statisticians continued to be scorned, they
were even called Tabellenfabrikanten and Tabellenknechte (slaves of
tables) (Knies 1850, p. 23). When describing Russia in 1734, I. K.
Kirillov used many tables, but his work was only published in 1831
(Ploshko & Eliseeva 1990, pp. 65 – 66). Another contribution which I
have not seen is Golitzin (1807).
By the end of the 19th century, owing to the heterogeneity of its
subject, Staatswissenschaft disintegrated. K. Pearson (1978, p. 125)
remarked that political economy (Adam Smith) was the first
discipline to break off from it and that the “evolution of Political
Philosophers” had further curtailed the Staatswissenschaft. All this
means that statistics, in its modern sense, owes its origin to political
arithmetic. Consequently, I dwell below on contributions which had
not belonged to the former subject, but, to the contrary, were
mathematical or, in any case, issued from statistical data.
Apparently in the mid-19th century the Staatswissenschaft had
regenerated and partly transferred to political arithmetic. It exists
nowadays, at least in Germany, it is taught there in some universities
and it can be considered as the application of the statistical method to
the various aspects of the life of states.
K. Pearson ( p. 29) named Edward Chamberlayne (1616 – 1703)
as the “English Achenwall” but he also noted that Chamberlayne had
“copied” his book from a French work of 1661 (which he did not see).
6.2.2. Population Statistics. Süssmilch (1741) adhered to the
tradition of political arithmetic. He collected vast statistical data on
the movement of population and attempted (as Arbuthnot did, see §
2.2.4) to reveal in it divine providence but he treated his materials
rather loosely. Thus, when taking the mean of the data pertaining to
towns and rural districts, he tacitly assumed that their populations
were equally numerous; in his studies of mortality, he had not
attempted to allow for the differences in the age structure of the
populations of the various regions etc. Nevertheless, it is possible to
believe that his works paved the way for Quetelet (§ 10.5); in
particular, he studied issues which later came under the province of
moral statistics (e.g., illegitimate births, crime, suicides). And his
tables of mortality had been in use even in the beginning of the 19th
century. Like Graunt, Süssmilch discussed pertinent causes and
offered conclusions. Thus, he (1758) thought of examining the
dependence of mortality on climate and geographical position and he
knew that poverty and ignorance were conducive to the spread of
epidemics.

80
79
 Süssmilch’s main contribution, the Göttliche Ordnung, marked the
origin of demography. Its second edition of 1765, included a chapter
“On the rate of increase and the period of doubling [of the
population]” written jointly with Euler. Partly reprinted in the latter’s
Opera omnia (in t. 7 of ser. 1, 1923), it served as the basis of one of
Euler’s memoirs (Euler 1767). Süssmilch naturally thought that the
multiplication of mankind was a divine commandment and that,
therefore, rulers must take care of their subjects. Quite consistently, he
condemned wars and luxury and indicated that the welfare of the poor
was to the advantage of both the state, and the rich. His collaboration
with Euler and frequent references to the latter in his Göttliche
Ordnung certainly mean that Euler had shared his general social
views.
Both as a statistician and as a deeply religious person Süssmilch
opposed polygamy. In this, he followed Graunt and perhaps other
statisticians. And it seems that Daniel Bernoulli, the bachelor, was
the only partisan of that practice. He (1768c, p. 103) stated, without
any substantiation, that polygamy “certainly” fostered the
multiplication of population. His was a letter to Euler, another deeply
religious person, who likely did not comment on that point. The
biographers of Bernoulli were not at all sure that he was really
religious. On Süssmilch see Birg (1986) and Pfanzagl & Sheynin
(1997).
Malthus (1798) picked up one of the conclusions made in the
Göttliche Ordnung, viz., that the population increased in a geometric
progression (which is still a more or less received statement).
Statisticians have been hard put to reconcile the Biblical command
(Genesis 1:28) Be fruitful and increase in number … with the
registered growth of the population. Pearson (1978, p. 337)
commented on this difficulty experienced by an early British
statistician:

Apparently his view is that while the Creator would not approve of
starvation for thinning humanity, he would have no objections to
plague or war! It reminds me of the farmer’s wife. She saved an
earwig from her hot iron and threw it in the kitchen fire saying I
nearly killed ye with the hot iron!

Euler is known to have left hardly any serious contribution to the

theory of probability (see also § 6.3.1 devoted to the theory of errors),
but he published a few memoirs on population statistics collected in
the same volume 7 of his works. When treating statistical data, he did
not introduce any stochastic laws (for example, laws of mortality), but
such concepts as increase in population and the period of its doubling
are due to him, and his reasoning was always elegant and
methodically interesting, in particular for life insurance (Paevsky
1935). See Sheynin (2007b).
Lambert published a mainly methodical study in population
statistics (1772). Without due justification, he proposed there several
laws of mortality (§ 9). One of them was represented by a sum of two
terms and he explained that they described physical processes; now,

81
80
we also see that they belonged to types IX and X of the Pearson
curves. Then, he formulated the problem concerning the duration of
marriages, statistically studied children’s mortality from smallpox and
the number of children in families (§ 108). See Sheynin (1971b) and
Daw (1980) who also appended a translation of Lambert’s discussion
of the smallpox issue.
When considering the last-mentioned subject, Lambert issued from
data on 612 families having up to 14 children and, once more without
substantiation, somehow adjusted his materials. It is remarkable that
he arbitrarily increased the total number of children by one half and
that the new data, as he maintained, were “smoother”. It might be
thought that Lambert attempted to allow for stillbirths and the death of
children. Elsewhere in his work he (§ 68) indicated that statistical
investigations should reveal [and explain] irregularities.
6.2.3. Medical Statistics. Population statistics was not closely
linked with any embryo of medical statistics, Graunt and Süssmilch (§
6.2.2) notwithstanding. Schlözer passed medical problems over in
silence. Statisticians have been ignorant of probability theory,
recognized only Bernoulli trials (§ 3.2.3). Interesting enough was
Poisson’s introduction of the (soon forgotten) term social arithmetic
(§ 8), i. e., population and medical statistics.
Medical statistics originated in the 19th century, partly because of
the need to combat the devastating visitations of cholera. Interestingly
enough, the expression medical probability appeared not later than in
the mid-18th century (Mendelsohn 1761, p. 204). At the end of that
century Condorcet (1795/1988, p. 542) advocated collection of
medical observations7 and Black (1788, p. 65) even compiled a
possibly forgotten “Medical catalogue of all the principle diseases and
casualties by which the Human Species are destroyed or annoyed” that
reminded of Leibniz’ thoughts (§ 2.1.4). Descriptions belonging to
other branches of natural sciences as well have actively been compiled
(mostly later) and such work certainly demanded preliminary
statistical efforts8. Some authors mistakenly stated that their
compilations ruled out the need for theories (cf. D’Alembert’s
opinion in Note 7). Until the beginning of the 20th century, the
partisans of complete descriptions continued to deny sampling in
statistics proper (§ 10.8-2).
Especially important was the study of inoculation, of the method of
preventing smallpox by communicating a mild form of smallpox from
one person to another. The history of smallpox epidemics and
inoculation is described in various sources (Condamine 1759, 1763,
1773; Karn 1931). In his first memoir, Condamine listed the
objections against inoculation, both medical and religious. He ended
his contribution thus:

Since the inoculation of the Royal Family in England, this practice

became generally accepted in France.

His statement did not however square with Daniel Bernoulli’s

appeals in its favour. In his second memoir Condamine (p. 464)
mentioned the Bernoulli family:

82
81
 In Basel, Messrs Bernoulli, whose name can by itself authorize
doubtful opinions on many points, have not rested content to declare
themselves publicly in favour of inoculation. They obtained approval
for the first experimentation from the faculties of medicine and
theology in Basel. The younger of the two brothers [Johann II, 1710 –
1790] and the only one of them married, decided to participate, and in
1756 he inoculated his two younger sons, and, a year ago, their elder
brother.

Johann II had indeed many sons, not all of them famous, whereas
his brother was Daniel Bernoulli. An approval from theologians was
really needed. White (1896/1898) described the “warfare of science
with theology” including, in vol. 2, pp. 55 – 59, examples of fierce
opposition to inoculation (and, up to 1803, to vaccination of
smallpox). Many thousands of Canadians perished in the mid-19th
century only because, stating their religious belief, they had refused to
be inoculated. White clearly distinguished between theology, the
opposing force, and “practical” religion.
Condamine (1773) includes the text of a third memoir on the same
subject (pp. 221 – 282) read in 1764 and the author’s correspondence,
in particular with Daniel Bernoulli. He had given Bernoulli data on
smallpox epidemics which the latter used in his research (below).
Karn stated at the very beginning of her article that

The method used in this paper for determining the influence of the
death-rates from some particular diseases on the duration of life is
based on suggestions which were made in the first place by D.
Bernoulli.

Bernoulli (1766) justified inoculation. That procedure, however,

spread infection, was therefore somewhat dangerous for the
neighbourhood and prohibited for some time, first in England, then in
France. Referring to statistical data, but not publishing it, Bernoulli
suggested that 1/n was the yearly rate of the occurrence of smallpox in
those who have not had it before; that 1/m was the corresponding
mortality; that m = n = 8 and that the inoculation itself proved fatal in
0.5% of cases.
He formed the appropriate differential equation whose solution

mξ
s=
1  (m  1)e x / n

showed the relation between age x (in years) and the number of people
of the same age, ξ, of which s had not contacted smallpox. Also by
means of a differential equation he derived a similar formula for a
population undergoing inoculation, that is, for its 99.5% which safely
endured it and were not anymore susceptible to the disease. It
occurred that inoculation lengthened the mean duration of life by 3
years and 2 months and that it was therefore, in his opinion, extremely
useful. The Jennerian vaccination, –

83
82
 The inestimable discovery by Jenner, who has thereby become one
of the greatest benefactors of mankind (Laplace 1814/1995, p. 83), –

was introduced at the end of the 18th century. Its magnificent success
had not however ruled out statistical studies. Thus, Simon (1887, vol.
1, p. 230) formulated a question about the impermanence of protection
against post-vaccinal smallpox and concluded that only
comprehensive national statistics could provide an answer.
D’Alembert (1761b; 1768c) criticized Daniel Bernoulli9. Not
everyone will agree, he argued, to lengthen his mean duration of life at
the expense of even a low risk of dying at once of inoculation; then,
moral considerations were also involved, as when inoculating
children. Without denying the benefits of that procedure, D’Alembert
concluded that statistical data on smallpox should be collected,
additional studies made and that the families of those dying of
inoculation should be indemnified or given memorial medals.
He also expressed his own thoughts, methodologically less evident
but applicable to studies of even unpreventable diseases. Dietz &
Heesterbeek (2002) described Bernoulli’s and D’Alembert’s
investigations on the level of modern mathematical epidemiology and
mentioned sources on the history of inoculation. For his part, K.
Pearson (1978, p. 543) stated that inoculation was

Said to have been a custom in Greece in the 17th century and was
advocated […] in the Phil. Trans. of the Royal Society in 1713.

Also see Sheynin (1972a/1977, pp. 114 – 116; 1982, pp. 270 –
272).
6.2.4. Meteorology. In § 2.1.4 I noted that Leibniz recommended
regular meteorological observations. And, indeed (Wolf 1935, p. 312),

Observations of barometric pressure and weather conditions were

made at Hanover, in 1678, and at Kiel, from 1679 to 1714, at the
instigation of Leibniz.

The Societas meteorologica Palatina in Pfalz (a principality in

Germany) was established in 1780, and, for the first time in the history
of experimental science, it organized cooperation on an international
scale (Sheynin 1984b, § 3.1). At about the same time the Société
Royale de Médecine (Paris) conducted observations in several
European countries (Kington 1974). And even in the 1730s – 1740s
they were carried out in several towns in Siberia in accordance with
directions drawn up by Daniel Bernoulli in 1733 (Tikhomirov 1932).
In the second half of the 18th century several scholars (the
meteorologist Cotte, Lambert and Condorcet) proposed plans for
comprehensive international meteorological studies.
Lambert (1773) studied the influence of the Moon on the air
pressure. Daniel Bernoulli, in a letter of 1759 (Radelet de Grave et al
1979, p. 62), encouraged him and noted that since the Moon’s distance
was not constant, its influence on the air ought to vary in time, cf. §

84
83
7.2-8. Then, Toaldo (1775; 1777) statistically studied the connections
between phenomena concerning meteorology and stated that the
weather depended on the configurations of the Moon. His opinion was
not abandoned until the mid-19th century (Muncke 1837, pp. 2052 –
2076), but either then, or later, in the second half of that century, for
example when the connection between cyclones and solar activity had
been studied (Sheynin 1984a, § 4.2), no embryo of correlation theory
was established, see § 10.7.
Lamarck, the most eminent biologist of his time, seriously
occupied himself with physics, chemistry and meteorology. In
meteorology, his merits had for a long time been ignored (Muncke
1837), but he is now remembered for his “pioneer work in the study of
weather” (Shaw & Austin 1942, p. 130) and I (Sheynin 1984b, § 6)
quoted several of his important pronouncements. He repeatedly
applied the term météorologie statistique (e.g., 1800 – 1811, t. 4, p. 1)
whose aim (Ibidem, t. 11, p. 9 – 10) was the study of climate, or, as he
(Ibidem, t. 4, pp. 153 – 154) maintained elsewhere, the study of the
climate, of regularities in the changes of the weather and of the
influence of various meteorological phenomena on animals, plants and
soil.
6.3. Mathematical Treatment of Observations
In modernity, mathematical treatment of observations became
necessary after regular astronomical observations (Tycho Brahe,
§ 1.2.2) had begun. A new problem of natural sciences, the
determination of the figure and the size of the Earth (of the Earth’s
ellipsoid of rotation), presented itself in the second half of the 18th
century. By means of meridian arc measurements the lengths of those
arcs were calculated (indirectly, by triangulation). After determining
the length of one degree of the meridian in two different and observed
latitudes it becomes possible to calculate both parameters of the
ellipsoid whereas redundant measurements lead to equations of the
type of (1.2) in these unknowns which can then be derived more
precisely10.
The term “Theory of errors” (Theorie der Fehler) is due to
Lambert (1765a, Vorberichte and § 321) who defined it as the study
of the relations between errors, their consequences, circumstances of
observation and the quality of the instruments. He isolated the aim of
the “Theory of consequences” as the study of functions of observed
(and error-ridden) quantities. In other words, he introduced the
determinate error theory (Note 2 to § 0.3) and devoted to it §§ 340 –
426 of his contribution. Neither Gauss, nor Laplace ever used the
new terminology, but Bessel (1820, p. 166; 1838b, § 9) applied the
expression “theory of errors” without mentioning anyone and by the
mid-19th century it became generally known. As far as that theory is
concerned, Lambert was Gauss’ main predecessor (see § 6.3.1).
I shall separately consider the adjustment of direct and indirect
measurements; note, however, that scientists of the 18th century
recognized the common character of these problems. Thus, in both
cases the unknowns were called by the same term, “Mittel” (Lambert
1765b, § 6) or “milieu” (Maire & Boscovich 1770, pp. 484 and 501),
also see the method of averages (§ 6.3.2).

85
84
 6.3.1. Direct Measurements. The first to touch on this case was
Cotes (1722, see Gowing 1983, p. 107).Without any justification he
advised to regard the weighted arithmetic mean, which he compared
with the centre of gravity of the system of points, – of the
observations,– as the “most probable” estimator of the constant
sought:

Let p be the place of some object defined by observation, q, r, s the

places of the same object from subsequent observations. Let there also
be weights P, Q, R, S reciprocally proportional to the displacements
arising from the errors in the single observations, and which are given
by the limits of the given errors; and the weights P, Q, R, S are
conceived as being placed at p, q, r, s, and their centre of gravity Z is
found; I say the point Z is the most probable place of the object.

Cotes appended a figure (perhaps representing a three-dimensional

picture) showing nothing except these four points. He had not
explained what he meant by most probable, nor did he exemplify his
rule. Nevertheless, his authority apparently gave support to the
existing common feeling (§ 1.2.4). Without mentioning Cotes and
putting forth qualitative considerations, Condamine (1751, p. 223)
recommended to apply it. Then, Laplace (1814/1995, p. 121) stated
that “all calculators” followed the Cotes rule. Elsewhere Laplace
(1812/1886, pp. 351 – 353) remarked that astronomers had begun to
follow Cotes after Euler (1749), but even before Cotes Picard
(1693/1729, pp. 330, 335, 343) called the arithmetic mean véritable.
T. Simpson (1756) applied, for the first time ever, stochastic
considerations to the adjustment of measurements; for that matter, he
made use of generating functions. He aimed to refute some authors
(left unnamed) who had maintained that one good observation was as
plausible as the mean of many of them, cf. § 1.2.2. Simpson assumed
that the chances of observational errors

– v, – v + 1, …, – 2, – 1, 0, 1, 2, …, v – 1, v

were equal [proportional] either to

r–v, r–v+1, …, r–2, r–1, 1, r, r2, …, rv–1, rv

or to

r– v, 2r–v+1, …, (v – 1)r–2, vr–1, (v + 1), vr, (v – 1)r2, …, 2rv–1, rv.

He assumed that the observational errors obeyed some density law

(taking r = 1 he thus introduced the uniform and the triangular discrete
distributions) and his was the first actual introduction of random
errors.
Denote the observational errors by εi, and by N, the number of some
chances. Then, as Simpson noted,

N(ε1 + ε2 + … + εn = m) was the coefficient of rm in the expansions of

86
85
 (r–v + … + r0 + … + rv)n = r–vn(1 – r)–n(1 – r2v+1)n,

(r–v + 2r–v+1 + … + (v + 1)r0 + … + 2rv–1 + rv)n =

r–vn(1 – r)–2n(1 – rv+1)2n.

The left sides of these two equalities were generating functions with
unit coefficients in the first case, and coefficients

1, 2, …, v + 1, … 2, 1

in the second instance.

For both these cases Simpson determined the probability that the
absolute value of the error of the arithmetic mean of n observations
was less than some magnitude, or equal to it11. Consequently, he
decided that the mean was always [stochastically] preferable to a
separate observation and he thus arbitrarily and wrongly generalized
his proof. Simpson also indicated that his first case was identical with
the determination of the probability of throwing a given number of
points with n dice each having (v + 1) faces. He himself (1740,
Problem No. 22), and earlier Montmort (§ 3.3.3), although without
introducing generating functions, and De Moivre (1730, pp. 191 –
197) had studied the game of dice.
Soon Simpson (1757) reprinted his memoir adding to it an
investigation of the continuous triangular distribution. He passed over
to the continuous case by assuming that |v| → ∞ leaving the magnitude
(m/n)/v constant. Here, the fraction in the numerator was the
admissible error of the mean and n, as before, the number of
observations. Simpson’s graph however represented a finite v and a
continuous argument (the observational errors) and the curve of the
error of the mean did not possess the distinctive form of the normal
distribution.
Simpson naturally had no knowledge of the variance and the
calculation of the probability that the error of the mean exceeded the
error of a single observation occurred to be difficult (Shoesmith
1985b).
Without mentioning Simpson, Lagrange (1776a) studied the error
of the mean for several other and purely academic distributions, also
by applying generating functions (even for continuous laws, thus
anticipating the introduction of characteristic functions). A possible
though inadequate reason for leaving out Simpson was the heated
dispute over priority between De Moivre and him. Lagrange
apparently had not wanted to be even indirectly involved in it. De
Moivre was a scholar of a much higher calibre (a fact clearly
recognized by Simpson) and 43 years the senior. At least on several
important occasions Simpson did not refer to De Moivre and, after
being accused by the latter (1725; only in edition of 1743, p. xii) of

Mak[ing] a Shew of new Rules, and works of mine, appeal[ed] to all

mankind, whether in his treatment of me [of Simpson], he has [not]

87
86
discovered an air of self-sufficiency, ill-nature, and inveteracy,
unbecoming a gentleman (Simpson, posth. publ. 1775, p. 144).

See Pearson’s opinion about him at the end of § 4.2.

Lagrange’s memoir contained other findings of general
mathematical interest. He was the first to use integral transformations,
and, in Problem 6, he derived the equation of the multivariate normal
distribution (K. Pearson 1978, p. 599). In his § 18 he introduced the
term courbe de la facilité des erreurs. Also see Sheynin (1973a, § 2).
Lambert (1760, §§ 271 – 306) described the properties of “usual”
random errors, classified them in accordance with their origin (§ 282),
unconvincingly proved that deviating observations should be rejected
(§§ 287 – 291) and estimated the precision of observations (§ 294),
again lamely but for the first time ever. He then formulated an
indefinite problem of determining a [statistic] that with maximal
probability least deviated from the real value of the constant sought
(§ 295) and introduced the principle of maximal likelihood, but not the
term itself, for a continuous density (§ 303), maintaining, however (§
306), that in most cases it will provide estimates little deviating from
the arithmetic mean. The translator of Lambert’s contribution into
German left out all this material claiming that it was dated12.
Lambert introduced the principle of maximum likelihood for an
unspecified, more or less symmetric and [unimodal] curve, as shown
on his figure, denote this curve by φ(x – xo) where xo was the sought
[parameter of location]. Let the observations be x1, x2, …, xn, and,
somewhat simplifying his reasoning, write his [likelihood function] as

φ(x1 – xo) φ(x2 – xo) … φ(xn – xo).

When differentiating this function, Lambert had not indicated that the
argument here was the parameter xo, etc.
In a few years Lambert (1765a) returned to the treatment of
observations. He attempted to estimate the precision of the arithmetic
mean, but did not introduce any density and was unable to formulate a
definite conclusion. He also partly repeated his previous
considerations and offered a derivation of a density law of errors
occurring in pointing an instrument (§§ 429 – 430) in accordance with
the principle of insufficient reason: it was a semi-circumference (with
an unknown radius) simply because there were no reasons for its
“angularity”.
Johann III Bernoulli (1785) published a passage from a
manuscript of Daniel Bernoulli (1769/1997) which he had received in
1769 but which was written, as its author had told him, much earlier.
There, Daniel assumed the density law of observational errors as a
semi-ellipse or semi-circumference of some radius r ascertained by
assigning a reasonable maximal error of observation and the [location
parameter] equal to the weighted arithmetic mean with posterior
weights

pi = r2 – ( x – xi)2. (4)

88
87
Here, xi were the observations and x , the usual mean. If required,
successive approximations could have been made.
The first to apply weighted or generalized arithmetic means was
Short (1763). This estimator demanded a subjective selection of
weights and, moreover, it only provided a correction to the ordinary
arithmetic mean which tended to vanish for even density functions.
In his published memoir Daniel Bernoulli (1778) objected to the
application of the arithmetic mean which (§ 5) only conformed with
an equal probability of all possible errors and was tantamount to
shooting blindly13. Instead, he suggested [the maximum likelihood
estimator of the location parameter] and supported his idea (§ 9) by
indicating that, when one out of several possible and incompatible
events had occurred, it should be thought that it was the event that
possessed the highest probability.
Listing a few reasonable restrictions for the density curve (but
adding to these the condition of its cutting the abscissa axis almost
perpendicularly), he selected a semi-circumference with radius equal
to the greatest possible, for the given observer, error. He then (§ 11)
wrote out the [likelihood function] as

{[r2 – (x – x1)2] [r2 – (x – x2)2] [r2 – (x – x3)2] …}1/2,

where, in somewhat different notation, x was the unknown abscissa of

the centre of the semi-circumference, and x1, x2, x3, …, were the
observations. Preferring, however, to calculate the maximum of the
square of that function, Bernoulli thus left the semi-circumference for
an arc of a parabola. He certainly had not known that the variance of
the result obtained will therefore change.
For three observations his [likelihood equation], as it occurred, was
of the fifth degree. Bernoulli numerically solved it in a few particular
instances with some values of x1, x2 and x3 chosen arbitrarily (which
was admissible for such a small number of them). In turn, I present his
equation as

x  x1 x  x2
+ 2 +…=0
r 2  ( x  x1 ) 2 r  ( x  x2 ) 2

so that the maximum likelihood estimate is

[ px] 1
x0 = , pi = 2 (5; 6)
 ip r  ( x 0  xi )
2

with unavoidable use of successive approximations. These formulas

are lacking in Bernoulli’s memoir although the posterior weights (6)
were the inverse of the weights (4) from his manuscript. This fact
heuristically contradicted his own preliminary statement about
shooting skilfully. Neither would have astronomers of his time
approved weights increasing towards the tails of a distribution. It is
now known, however, that these are expedient in case of some

89
88
densities. I also note that, according to Bernoulli, the properly normed
density was

y = (3/4r3) [r2 – (x – xo)2], xo – r ≤ x ≤ xo + r

and that the weights (6) should be corrected accordingly.

Daniel Bernoulli is invariably described as a mathematician and
natural scientist who was able to achieve great success with
comparatively little mathematics, and I fail to understand why he did
not indicate formulas (5) and (6).
Euler (1778) commented on Bernoulli’s memoir. He (§ 6) objected
to the [principle of maximum likelihood] but his reasoning was not
good enough. He argued that, on the one hand, the result of an
adjustment should barely change whether or not a deviating
observation was adopted, but, on the other hand, that the value of the
[likelihood function] essentially depended on that decision.
Euler then (§ 7) remarked that, in general, there was no need

To have recourse to the principle of the maximum, since the

undoubted precepts of the theory of probability are quite sufficient to
resolve all questions of this kind.

He had not noticed that his requirement led to the choice of the
median.
In the positive part of his commentary, Euler recommended, instead
of the arithmetic mean, the estimate (5) with posterior weights (4) and
he mistakenly assumed that Bernoulli had actually chosen these same
weights. While developing his thoughts, and denoting the n
observations by П + а, П + b, П + с, …, where

a + b + c + … = 0, (7)

he formed the equation

nx3 – nr2x + 3Bx – C = 0,

B = a2 + b2 + c2 + …, C = a3 + b3 + c3 + …,

from which the estimate П + x should have been calculated with x

equal to its root least in absolute value. Condition (7) meant that the
estimate sought was the closest possible to the arithmetic mean; Euler
himself (§ 9) justified his choice of the root by noting that x = 0 as
r → ∞, that is, as n → ∞, also see below.
Euler (§ 11) also remarked that estimate (5) with weights (4) can be
obtained from the condition

[r2 – (xo – a)2]2 + [r2 – (xo – b)2]2 + [r2 – (xo – c)2]2 + … = max. (8)

The magnitudes in parentheses are the deviations of observations from

the estimate sought and their fourth powers are negligible so that
condition (8) is equivalent to the requirement

90
89
 (xo – a)2 + (xo – b)2 + (xo – c)2 + … = min, (9)

whence, in accordance with condition (7), follows the arithmetic

mean. Condition (9) is heuristically similar to the principle of least
squares (which in case of one unknown indeed leads to the arithmetic
mean) and condition (8) with weights (4) resembles the Gaussian
principle of maximum weight (of least variance). True, if the density
of the observational errors is known (which was the unrealistic
assumption of both Bernoulli and Euler), then other estimates can be
better than the arithmetic mean, cf. the opinion of Poincaré (§ 11.2).
A small deviation from condition (9) does exist and it is easy to see
that it is occasioned by inevitable deviations of the observations from
the proposed (or tacitly assumed) symmetrical law. Bernoulli himself
noted this fact when (see above) numerically adjusting several sets of
three observations. So, I say once more that in actual fact Bernoulli
proposed the general arithmetical mean.
In his last memoir Daniel Bernoulli (1780) separated, for the first
time ever, observational errors into random (momentanearum) and
systematic (chronicarum), although not for observations in general.
Thus, he assumed an equal number of slow and quick swings of the
pendulum and said nothing about the possible interdependence of
consecutive swings and his systematic error was the simplest possible.
But Daniel was the first to apply the normal distribution for
calculations in a field belonging to the theory of errors, to use the
probable error as a measure of precision in the same field and to
isolate the two kinds of errors.
As I remarked in § 1.1.4, even ancient astronomers undoubtedly
knew that some errors were systematic. And here is the opinion of D.
T. Whiteside (private communication, 1972):

Newton in fact (but not in explicit statement) had a precise

understanding of the difference between random and structurally
‘inbuilt’ errors. He was certainly, himself, absorbed by the second
type of ‘inbuilt’ error, and many theoretical models of differing types
of physical, optical and astronomical phenomena were all consciously
contrived so that these structural errors should be minimized. At the
same time, he did, in his astronomical practice, also make suitable
adjustment for ‘random’ errors in observation …

I return to Bernoulli. Since he considered pendulums14, he indicated

that these errors acted proportionally to the square root of, and to the
time itself respectively. Making use of his previous findings (§ 6.1.1,
formula (1)), Bernoulli justified his investigation by the [normal
distribution] which thus first occurred in the theory of errors, although
only as a limiting law.
The number of vibrations of a seconds pendulum during a day is 2N
= 86,400; Bernoulli assumed that (N + µ) of them were slower, and
(N – µ) faster than stipulated, with periods of (1 + α) and (1 – α)
respectively. His simple pattern meant that the number of positive
(say) errors possessed a symmetric binomial distribution and that the

91
90
error of the pendulum accumulated after a large number of vibrations
had a normal distribution.
In his previous work Bernoulli (1770 – 1771) noted that, for
N = 10,000,


1
[2/ πN ]  exp( x 2 N )dx 
0
2

if µ = 47.25. Now, having N = 43,200, he obtained, for the same

probability of 1/2,

µ = 47.25 4.32 ≈ 100.

It was this calculation that caused his conclusion (above) about the
behaviour of random errors. Already in the 19th century, however, it
became known that such errors can possess other laws of distribution
(e.g., § 10.9.4).
Note also that Bernoulli came close to introducing the probable
error; to recall (§ 2.2.2), Huygens discussed the probable duration of
life. Bernoulli was also the first to introduce elementary errors. I do
not however set high store by this fact; indeed, this notion is not
necessary for proving the CLT. I conclude by remarking that
Bernoulli had not investigated the more general pattern of an unequal
number of the slower and the faster vibrations although it
corresponded to the case of unequal probabilities of male and female
births, also studied by him. Neither had he said anything about the
possible dependence between the periods of successive vibrations.
6.3.2. Indirect measurements. Here, I consider the adjustment of
redundant systems

aix + biy + … + si = vi, i = 1, 2, …, n (10)

in k unknowns (k < n) and residual free terms vi (see § 1.2.1).

1) In case of two unknowns (cf. beginning of § 6.3) astronomers
usually separated systems (10) into all possible groups of two
equations each and averaged the solutions of these groups. In other
words, if (xij; yij) is the solution of group (i; j), i, j = 1, 2, …, n, i < j,
then, in accordance with this method of combinations, the final
estimates of the unknowns were

xo = (1/ Cn2 )Σxij, yo = (1/ Cn2 )Σyij.

The residual free terms were thus neglected.

In 1757 and later Boscovich (Cubranic 1961, pp. 90 – 91; Maire &
Boscovich 1770, pp. 483 – 484) applied this method but it did not
satisfy him, see below. Interestingly enough, in the first case he
(Cubranic 1961, p. 46) derived the arithmetic mean of four latitudinal
differences in an unusual way: he first calculated the half-sums of all
six pairwise differences and then took their mean. He apparently
attempted to exclude, without changing the final result, the

92
91
unavoidable systematic errors and thus to ensure a (qualitative)
estimation of the order of random errors15.
In the 19th century, it was discovered that the MLSq led to the same
result as the method of combinations, although only if the particular
solutions were appropriately weighted (Whittaker & Robinson 1949,
p. 251).
2) For the case of three unknowns the method of combinations
becomes unwieldy. In an astronomical context, Mayer (1750) had to
solve 27 equations in three unknowns. He separated them into three
groups of nine equations each, calculated three particular solutions
(see below), and, finally, averaged them. The plausibility of the results
thus obtained depended on the expediency of the separation and it
seems (Stigler 1986, pp. 21 – 25) that Mayer had indeed made a
reasonable choice. Being mostly interested in only one unknown, he
included the equations with their greatest and smallest in absolute
value coefficients in the first, and the second group respectively. Note
also that Mayer believed that the precision of results increased as the
number of observations, but in his time this mistake was
understandable.
Mayer solved each group of equations under an additional condition

Σvi = 0, (11)

where i indicates the number of an equation; if the first group includes

the first nine of them, then i = 1, 2, …, 9. Biot (1811, pp. 202 – 203)
testified that before the advent of the MLSq astronomers had always
applied the method of Mayer and Laplace (1812/1886, pp. 352 – 353)
stated almost the same.
In a letter of 1850 Gauss (W/Erg-5, No. 3/6, p. 90) remarked that
Mayer had only calculated by means of primitive combinations. He
referred to Mayer’s manuscripts, but it is likely that Mayer’s trick was
almost the same in both cases. And Gauss himself, in an earlier letter
of the same year (Ibidem, pp. 66 – 67), recommended a similar
procedure for calibrating an aneroid.
Condition (11) determines the method of averages and Lambert’s
recommendation (1765b, § 20) about fitting an empirical straight line
might be interpreted as its application. Lambert separated the points
(the observations) into two groups, with smaller and larger abscissas,
and drew the line through their centres of gravity. He employed a
similar procedure when fitting curves by separating the points into
several groups.
The method of averages was intuitively considered as conforming
to the equal possibility of errors of each sign (Maire & Boscovich
1770, p. 501), and, apparently, as leading in case of one unknown to
the arithmetic mean. See § 10.1 for its further history.
3) The Boscovich method. He (Maire & Boscovich, Ibidem)
adjusted systems (10) under additional conditions

v1 + v2 + … + vn = 0, |v1| + |v2| + ... + |vn| = min, (12; 13)

93
92
the first of which determined the method of averages. It can be
allowed for by summing all the equations and eliminating one of the
unknowns from the expression thus obtained. The mean (milieu), as
Boscovich remarked, should be connected

Par une certaine loi aux règles des combinaisons fortuites et du

calcul des probabilités16.

He was unable, however, to explain how his conditions conformed to

his aim.
Boscovich’s second condition (13)17 linked his method with the
[median]. Indeed, his geometric adjustment of systems (10) consisted
in constructing a straight line whose slope occurred to be equal to the
median of some fractions. In other words: for meridian arc
measurements systems (10) are

ai x + y + si = vi. (14)

After allowing for condition (12), we have

[ai – (1/n)Σai]x + [si – (1/n)Σsi] = 0.

Calculate the n values of x and choose as the estimate their median.

Laplace (§ 7.2-6) also made use of the Boscovich method.
4) The minimax method. According to it, systems (10) are solved
under the additional condition

|vmax| = min,

with the minimum being determined from all possible and expedient
solutions18. In § 2.1.4 I indicated that Kepler had apparently made use
of some elements of this method (true, not even for algebraic
equations). It does not ensure optimal, in any sense, results, but allows
to check whether the theory, underlying the given system (10), is
correct. Indeed, any other method of its solution will lead to a greater
value of |vmax|, the gap between theory and observation will be wider,
and the correctness of the former might mistakenly be questioned.
Gusak (1961) described the history of the minimax method from
1778, when Euler had applied it to an important but, in my context,
irrelevant study, to Chebyshev. However, Euler (1749) made use of
the rudiments of that method much earlier. When solving systems of
the type of (10), he compared only a few “solutions” with each
other19. Then, Lambert (1765a, § 420) recommended the same
method but owned that he did not know how to apply it “in a general
manner and without many roundabout ways”. Laplace (1789/1895,
pp. 493, 496 and 506 and elsewhere) applied the minimax method for
preliminary investigations, – for checking whether or not the results of
meridian arc measurements and pendulum observations contradicted
the theory according to which the Earth was an oblate ellipsoid of
rotation. Since the method of minimax has no stochastic underpinning,
I am not describing the appropriate algorithms introduced by Laplace;

94
93
I note, however, that it is applied in the theory of statistical decision-
making (Lehmann 1959, Chapter 9).
5) Euler (1749, 1755, 1770) had to treat indirect measurements as
well. At least in the first two instances his goal was much more
difficult than that outlined in § 1.2.1 where the underlying theory was
supposed to be known for sure. Concerning the first of his
contributions, Wilson (1980, p. 262n) remarked that Euler was

Stymied by the finding that, for certain of the variables, the

equations led to wildly different values, so that averages [values
adjusted in accordance with some definite rule] would seem
meaningless.

Euler did not attempt to build a more or less general theory; he

rather restricted his efforts to achieving practical results. In
accordance with the considerations above, he turned in some cases to
the minimax principle. On another important occasion Euler’s attitude
(1770) is difficult to explain: he did not keep to any definite method
and, moreover, combined equations in a doubtful manner. So as to
eliminate one unknown, he subtracted each equation from (say) the
first one, thus tacitly regarding it as possessing much more weight.
See Sheynin (2007b, § 3.5). This procedure seems to have been
necessary. Ivory (§ 10.9.1) acted the same way.
6.3.3. The True Value of a Measured Constant (Sheynin 2007a).
Astronomers, geodesists, metrologists and other specialists making
measurements have always been using this expression. Mathematical
statistics has done away with true values and introduced instead
parameters of densities (or distribution functions), and this was a step
in the right direction: the more abstract was mathematics becoming,
the more useful it proved to be.
Fisher was mainly responsible for that change; indeed, he (1922,
pp. 309 – 310) defined the notions of consistency, efficiency and
sufficiency of statistical estimators without any reference to true
values. But then, on p. 311, he accused the Biometric school of
applying the same names to

The true value which we should like to know […] and to the particular
value at which we happen to arrive.

So the true value was then still alive and even applied, as in the
lines above, to objects having no existence in the real world.
Incidentally, the same can be said about Gauss (1816, §§ 3 and 4)
who repeatedly considered the true value of a measure of precision of
observations. And Hald (1998) mentioned the true value many times
in Chapters 5 and 6; on p. 91 he says: “the estimation of the true
value, the location parameter…”
So what is a true value? Markov (1924, p. 323) was the only
mathematician who cautiously, as was his wont, remarked:

It is necessary in the first place to presume the existence of the

numbers whose approximate values are provided by observations.

95
94
 This phrase first appeared in the 1908 edition of his Treatise (and
perhaps in its first edition of 1900). He did not attempt to define true
value, but this is exactly what Fourier (1826/1890, p. 534) had done
more than a century before him. He determined the véritable objet de
la recherche (the constant sought, or its “true” value) as the limit of
the arithmetic mean of n appropriate observations as n → ∞.
Many authors, beginning perhaps with Timerding (1915, p. 83) [and
including Mises (1919/1964a, pp. 40 and 46)], without mentioning
Fourier and independently from each other, introduced the same
definition. One of them (Eisenhart 1963/1969, p. 31) formulated the
unavoidable corollary: the mean residual systematic error had to be
included in that “true” value:

The mass of a mass standard is […] specified […] to be the mass of

the metallic substance of the standard plus the mass of the average
volume of air adsorbed upon its surface under standard conditions.

However, even leaving systematic influences aside, the precision of

observations is always restricted (§ 11.2-8) so that the term “limit” in
the Fourier definition (which is in harmony with the Mises definition
of probability) must not be understood literally. I indicate also that
Gauss (W-9, pp. 278 – 281), see also Schreiber (1879, p. 141)
measured each angle in the field until becoming convinced that further
work was meaningless.
The concept of true value is not, however, universally recognized.
Chatterjee (2003, p. 264) mentioned the true value syndrome which
was ultimately left behind.

Notes
1. Lagrange (1777) solved a similar problem for a finite number of urns and balls
of two colours as well as some other stochastic problems by means of partial
difference equations.
2. He published many memoirs and papers on the theory of probability and its
applications (§ 6.2.3) and it is difficult to organize them bibliographically; on this
point see Paty (1988). Todhunter (1865) devoted an entire chapter to D’Alembert.
3. Cf. the D’Alembert – Laplace problem (Note 4 in Chapter 1). In 1750
D’Alembert declared that randomness was only caused by ignorance (Note 2 in
Chapter 1). The denial of randomness, also upheld by Kepler (§ 1.2.4) and Laplace
(§ 7.3), although only by mere words, proved fruitless.
4. Regarding his really strange attitude towards medicine see Note 7.
5. His unsubstantiated conclusion was absolutely wrong. Loveland (2001)
attempted to reconstruct Buffon’s reasoning.
6. Recall however (§ 3.3.4) that Condorcet reasonably remarked on the
Petersburg game.
7. D’Alembert (1759/1821, p. 163) should also be mentioned. The first edition of
this contribution published in 1759 apparently had not contained any such statement.
Note, however, that he died in 1783 so that he formulated his similar desire in the
18th century. D’Alembert even stated that a physician was a blind man who can
strike either the disease or the patient by his club and added, on p. 167, that the best
doctor was the one who least believed in medicine. See also § 10.8.
8. The same author, Black, appended a “Chart of all the fatal diseases and
casualties in London during […] 1701 – 1776” to his book. It will warn us to make
the best disposition and preparation for defence. However, in his previous book
Black (1782) stated contradictory ideas.

96
95
 9. In the first case he discussed Bernoulli’s report; I stress that the latter’s memoir
appeared only in 1766. Later D’Alembert rewrote his memoirs. See Todhunter
(1865, pp. 265 – 271, 277 – 278 and 282 – 286) for a detailed description of his
proposals.
10. Not the semiaxes of the ellipsoid, a and b (a > b), were determined, but rather
a and the flattening (a – b)/a. The flattening had also been derived from pendulum
observations; see § 10.9.1 where I describe the pertinent work of Ivory.
11. The distributions introduced by Simpson, if considered continuous, can be
directly compared with each other in the sense that the respective variances are v2/3
and v2/6.
12. In a letter of 1971 E. S. Pearson informed me that “curiously” his father’s
Lectures (1978), – then not yet published, – omitted Lambert. He explained:

It was not because [Lambert’s] writings were in German of which my father was
an excellent scholar. I suppose […] that he selected the names of the personalities
he would study from a limited number of sources, e.g., Todhunter, and that these did
not include Lambert’s name. [Todhunter did refer to Lambert but had not described
his work.] Of course, K. P. was over 70 by the time his history lectures passed the
year 1750, and no doubt his exploration was limiting itself to the four Frenchmen,
Condorcet D’Alembert, La Grange and Laplace.

13. Here, however, is K. Pearson’s reasonable qualitative statement (1978, p.

268): small errors are more frequent and have their due weight in the mean.
14. For this reason his memoir was attributed to practical mechanics and until my
publication (Sheynin 1972a) its stochastic nature had not been noticed.
15. Tycho’s example (Note 25 in Chapter 1) is more convincing.
16. The last term deserves attention: it was hardly used before Boscovich.
17. Galileo (§ 1.2.3) and Daniel Bernoulli (1735/1987, pp. 321 – 322) applied
this condition in the case in which the magnitudes such as vi were positive by
definition. The latter derived the plane of the solar equator in such a way that the
sum of the inclinations of the planetary orbits, considered positive, relative to the
equator, was minimal. W. Herschel (1805) determined the movement of the Sun by
issuing from the apparent motion of the stars. The sum of these motions depends on
the former and its minimal value, as he assumed, provided an expedient estimation
of that movement, cf. Kepler’s similar method at the end of § 1.2.4.
Note that in those times the motion of a star could have been discovered only in
the plane perpendicular to the line of vision. Here is W. Herschel’s earlier reasoning
(1783/1912, vol. 1, p. 120):

We ought […] to resolve that which is common to all the stars […] into a single
real motion of the Solar system, as far as that will answer the known facts, and only
to attribute to the proper motions of each particular star the deviations from the
general law the stars seem to follow …

Such, he added, were “the rules of philosophizing”. Compare now Newton’s Rule
No. 1 of reasoning in philosophy (§ 1.1.5).
When treating direct measurements W. Herschel (1806) preferred the [median]
rather than the arithmetic mean (Sheynin 1984a, pp. 172 – 173).
18. It is remarkable that the minimax method corresponds, as Gauss (1809b,
§ 186) noted, to the condition

lim(v12k + v22k + ... + vn2k) = min, k → ∞.

19. Stigler (1986, pp. 27 – 28) called Euler’s memoir (1749) a “statistical failure”
and, in his opinion, Euler was a mathematician who “distrusted” the combination of
equations. Without perceiving the main goal of the method of minimax, and
mentioning a classic in a free and easy manner, Stigler got into a mess. See Wilson’s
statement at the end of § 6.3.2. In his second book Stigler (1999, pp. 317 – 318)
unblushingly called Euler a great statistician but did not notice his lame argument
against the [method of maximum likelihood] (§ 6.3.1). Moreover, as stated there, in
the particular instance under discussion Euler should have opted for the not yet

97
96
named median (which he did not mention) rather than the mean (which he
preferred).
For that matter, in the 18th century practitioners experienced difficulties when
deciding how to adjust their observations (Bru 1988, pp. 225 – 226); and at the turn
of that century Laplace and Legendre simply refused to adjust a triangulation chain
laid out between two baselines. Instead, likely fearing the propagation of large
errors, they decided to calculate each half of the chain starting from its own baseline
(Sheynin 1993b, p. 50). Much later Laplace (ca. 1819/1886, pp. 590 – 591) defended
their decision by the previous ignorance of the “vraie théorie” of adjustment and
added that his justification of the MLSq had changed the situation.
I supplement Bru’s description by indicating that Maupertuis (1738, p. 160;
1756b, pp. 311 – 319) calculated his triangulation twelve times (each time taking
into account differing sets of measured angles), selected two of his results and
adopted their mean value.
It is instructive to note that, before the adjustment proper of the Soviet primary
triangulation, each of its chains situated between baselines and astronomically
determined azimuths was replaced by an appropriate geodetic line (cf. beginning of
§ 9B). Only these lines were then adjusted after which each chain was finally dealt
with independently from one another. One of the benefits of this procedure was that
it prevented the systematic errors from “freely walking” over the entire network, as
Izotov, the leading assistant of Krasovsky, the calculator of the Krasovsky spheroid,
explained ca. 1950 in one of his lectures at the Moscow Geodetic Institute which I
attended.

7. Laplace
7.1. Theory of probability
Laplace devoted a number of memoirs to the theory of probability
and later combined them in his Théorie analytique des probabilités
(abbreviation: TAP) (1812). When referring to it, I often indicate only
the page numbers. I describe its second Livre; in the first one he
studied the calculus of generating functions with application to the
solution of ordinary and partial difference equations and the
approximate calculation of integrals.
1) In Chapter 1 Laplace provided the “classical” definition of
probability (introduced by De Moivre, see § 4.3), formulated the
addition and multiplication theorems for independent events as well as
theorems concerning conditional probabilities. He described the same
material in his Essai philosophique …1 where he (1814/1995, p. 10),
in addition, included the so-called Bayes theorem, see formula (5.1),
calling it a principle. Much earlier he (1774/1891, p. 29) introduced a
“fundamental principle”, – the same theorem for the case of constant
prior probabilities P(Ai):

P(Ai/B)/P(Aj/B) = P(B/Ai)/P(B/Aj).

2) In Chapter 2 Laplace solved a number of problems by means of

difference, and partial difference equations. I consider three other
problems.

98
97
 a) In an astronomical context Laplace studied sampling with
replacement. Tickets numbered from 0 to n are extracted from an urn.
Determine the probability that the sum of k numbers thus extracted
will be equal to s (p. 257). Let these numbers be t1, t2, …, tk, then

t1 + t2 + ... + tk = s. (1)

Laplace calculated the number of combinations leading to equality (1)

allowing for the condition ti ≤ n, i = 1, 2, ..., k by assigning to these ti
probabilities

(1 – ln+1)/(n + 1) (2)

with l = 0 for ti ≤ n and l = 1 otherwise. Earlier, I (Sheynin 1973a, pp.

291 – 298) discussed Laplace’s use of discontinuity factors in
somewhat more detail. I also described his similar method which he
applied in 1810 and which dates back to De Moivre and Simpson
(Ibidem, pp. 278 – 279). If, for example, two (three) of the t’s exceed
n, that factor in (2) is raised to the second (to the third) power etc.
Laplace calculated the probability sought and considered the case of
s, n → ∞ and his formula on p. 260 for the distribution of the sum of
independent, continuous variables obeying the uniform law on interval
[0; 1] corresponds with modern literature (Wilks 1962, § 8.3.1) which
does not, however, demand large values of s and n.
Also in an astronomical context, already in 1776, Laplace solved a
problem concerning such distributions by very complicated recursion
relations (Sheynin 1973a, pp. 287 – 290). Note, however, that even
Simpson and Lagrange (§ 6.3.1) obtained similar findings in the
theory of errors.
Laplace treated the two other problems alluded to above in the same
way as he did earlier in 1781.
b) Non-negative [random variables] t1, t2, …, tk with differing laws
of distribution φi(t) are mutually independent and their sum is s.
Determine the integral

 ψ t ; t ;
1 2 ; tk  φ1  t  φ 2  t  φ  t  dt1dt2  dtk

over all possible values of the variables; ψ is not yet chosen. Laplace
then generalizes his very general problem still more by assuming that
each function φi(t) can be determined by different formulas on
different intervals of its domain.
When solving this problem, Laplace made use of the same
discontinuity factor as above and derived the Dirichlet formula (an
expression for the n-tuple integral of the product of n power functions
over the area in which the sum of the arguments, i. e., of the bases of
those functions, is restricted by the interval [0, 1]), even in a more
general version. The case of ψ ≡ 1 enabled him to determine the
probability of equality (1) (which interested Laplace here also). He
then once more specified his problem by assuming that

φi(t) = a + bt + ct2.
99
98
 When solving that problem, Laplace derived a multiple integral of
u, u1, u2, … over the area

0 ≤ u + u1 + u2 + … ≤ s

and differentiated it with respect to s, – with respect to that area! He

had not mentioned that he calculated that derivative rather than the
integral and only provided the final answer. I note that a simple
transformation u = sx, u1 = sx1, u2 = sx2, … saves us from that
unusual differentiation (Sheynin 1973a, p. 292).
c) An interval OA is divided into equal or unequal parts and
perpendiculars are erected to the interval at their ends. The number of
perpendiculars is n, their lengths (moving from O to A) form a non-
increasing sequence and the sum of these lengths is given. Suppose
now that the sequence is chosen repeatedly; what, Laplace asks, would
be the mean broken line connecting the ends of the perpendiculars?
The mean value of a current perpendicular? Or, in the continuous
case, the mean curve? Each curve might be considered as a realization
of a stochastic process and the mean curve sought, its expectation.
Laplace was able to determine this mean curve (Sheynin 1973a, p.
297) by issuing from his previous problem2 and, in 1781, he attempted
to apply this finding in the theory of errors (§ 7.2) and for studying
expert opinions. Suppose that some event can occur because of n
mutually exclusive causes. Each expert arranges these in an increasing
(or decreasing) order of their [subjective] probabilities, which, as it
occurs, depend only on n and the number of the cause, r, and are
proportional to

1 1 1
  ...  .
n n 1 n  r 1

The comparison of the sums of these probabilities for each cause

allows to show the mean opinion about its importance. To be sure,
different experts will attribute differing perpendiculars to one and the
same cause.
3)The third Chapter is devoted to the integral “De Moivre –
Laplace” theorem and to several interesting problems connected with
the transition to the limit. In proving that theorem (§ 4.4) Laplace
applied the Euler – MacLaurin summation formula, and, a second
innovation, calculated the remainder term to allow for the case of
large but finite number of trials. His formula was (α = n /2 xx ) :

P(|µ – np – z| ≤ l) =

α
(2/√π)  exp(– t2)dt + α 1/π exp(–l2n/2xx′). (3)
0

Here p was the probability of success in a single Bernoulli trial, µ,

the total number of successes in n trials, q = 1 – p, z was unknown but

100
99
|z| < 1, x = np + z, and x′ = nq – z.
Laplace indicated that his theorem was applicable for estimating the
theoretical probability given statistical data, cf. the Bayes theorem in
§ 5.2, but his explanation was not clear, cf. Todhunter (1865, pp. 554
– 556). Insufficiently clear is also Hald’s description (1990, § 24.6).
Already Daniel Bernoulli (§ 6.1.1) solved one of Laplace’s
problem: There are two urns, each containing n balls, some white and
the rest black; on the whole, there are as many white balls as black
ones. Determine the probability u that the first urn will have x white
balls after r cyclic interchanges of one ball. The same problem was
solved by Lagrange (1777/1869, pp. 249 – 251), Malfatti
(Todhunter 1865, pp. 434 – 438) and Laplace (1811; and in the same
way in the TAP).
Laplace worked out a partial difference equation, “mutilated it most
unsparingly” (Todhunter 1865, p. 558), and obtained a partial
differential equation with argument x

u′r/n = 2u + 2µu′µ + u″µµ , for x = (n + µ√n)/2

and expressed its solution in terms of functions related to the

[Chebyshev –] Hermite polynomials (Molina 1930, p. 385). Hald
(1998, p. 339) showed, however, that Todhunter’s criticism was
unfounded.
Later Markov (1915b) somewhat generalized this problem by
considering the cases of n → ∞ for r/n → ∞ and n → ∞ and r/n =
const and Steklov (1915) proved the existence and uniqueness of the
solution of Laplace’s differential equation with appropriate initial
conditions added whereas Hald (2002) described the history of those
polynomials. Hostinský (1932, p. 50) connected Laplace’s equation
with the Brownian motion and thus with the appearance of a random
process (Molina 1936).
Like Bernoulli, Laplace discovered that in the limit, and even in the
case of several urns, the expected (as he specified on p. 306) numbers
of white balls became approximately equal to one another in each of
them. He also remarked that this conclusion did not depend on the
initial distribution of the balls. Finally, in his Essai (1814/1995, p. 42),
Laplace noted that nothing changed if new urns, again with arbitrary
distributions of balls, were added to the original urns. He declared,
apparently too optimistically, that

These results may be extended to all naturally occurring

combinations in which the constant forces animating their elements
establish regular patterns of action suitable to disclose, in the very
mist of chaos, systems governed by these admirable laws.

Divine design was absent, cf. De Moivre’s dedication of his book

to Newton in § 4.3.
The Daniel Bernoulli – Laplace problem coincides with the
celebrated Ehrenfests’ model (1907) which is usually considered as
the beginning of the history of stochastic processes so that its
formulation is really unnecessary. As I noted in § 6.1.1c, the existence

101
100
of the limiting state in this problem can be justified by the Markov
ergodic theorem for Markov chains. See also Sheynin (1972a/1977,
pp. 127 – 128).
4) I touch on Chapter 4 in § 7.2-4. Laplace devoted Chapter 5 to
the detection of constant causes (forces) in nature. Thus, he attempted
to estimate the significance of the daily variation of the atmospheric
pressure. K. Pearson (1978, p. 723) noted that nowadays the Student
distribution could be applied in such investigations, that some of the
assumptions which Laplace made proved wrong, etc. and, in addition,
that Laplace had unjustifiably rejected those days during which the
variation exceeded 4 mm.
Laplace remarked that the calcul des probabilités can be applied to
medicine and economics. It may be argued that he thought about
stochastic analysis of statistical data, see his Essai (1814/1995, p. 61).
He even inserted a section on the application of the probability
calculus to moral sciences.
Concerning geometric probability, Laplace only discussed the
Buffon problem and stated (p. 365) that geometric probability can be
applied to rectify curves “ou carrer leurs surfaces” (nowadays this
carrer certainly looks strange).To repeat (§ 6.1.6), a needle of length
2r falls from above on a set of parallel lines. The distance between
adjacent lines is a  2r and the probability p that the needle intersects
a line is

p = 4r/πa.

Without proof Laplace mistakenly stated that, for a = 1, 2r = π/4

was the optimal length of the needle for statistically determining π
although he provided the correct answer, 2r = 1, in the first edition of
the TAP.
Both Todhunter (1865, pp. 590 – 591) and Gridgeman (1960)
examined this problem by applying stochastic considerations, and
both proved that Laplace’s previous answer was correct.
5) In Chapter 6 Laplace solved some problems by means of the
Bayes approach (see § 5.1) although without referring to him; true, he
mentioned Bayes elsewhere (1814/1995, p. 120). Here is one of them.
Denote the unknown probability that a newly born baby is a boy by
x and suppose that during some time p boys and q girls were born.
Then the probability of that “compound” event will be proportional to

y = xp(1 – x)q. (4)

If z(x) is the prior distribution of x, then

b 1

P(a ≤ x ≤ b) =  yz dx ÷  yz dx, 0 < a < b < 1. (5)

a 0

If, as Laplace nevertheless assumed, z was constant, and if p and q

were large, the probability sought will be expressed by an integral of
an exponential function of a negative square.

102
101
 And so, Laplace actually estimated the probability x. For the curve
(4) the point of its maximum

α = p/(p + q) (6)

seems to be its natural estimator, but Ex, or, more precisely, the
expectation of a random variable ξ with distribution

xp(1 – x)q ÷  xp(1 – x)q dx,

0

does not coincide with (6): the latter is only an asymptotically

unbiased estimator of x. This expectation is evidently

p 1
Eξ= . (7)
pq2

The introduction of functions z(x) allowed to assume an equal

probability for each value of x, but the choice of such functions
remained undecided. Of course, nothing more could have been
achieved.
Laplace went on to discuss the bivariate case and then solved
another problem. Suppose that the inequality p > q persisted during a
number of years. Determine the probability that the same will happen
for the next hundred years. There is no doubt that Laplace understood
that his problem only made sense under invariable social and
economic conditions. Here is his answer:

1 1

P =  xp(1 – x)qz100dx ÷  xp(1 – x)qdx

0 0

where z is the sum of the first n terms of the development

[x + (1 – x)]2n and 2n = p + q.

A similar problem in which q = 0, p = m and z = xn led Laplace to

the probability of such z:

P = (m + 1) ÷ (m + n + 1).

In the Essai Laplace (1814/1995, p. 11) applied this formula,

slightly different from his previous formula (7), for solving Price’s
problem about the next sunrise (§ 5.1) but he only mentioned Buffon
(§ 6.1.4), and, as expected, did not agree with his solution.
I consider this problem in more detail. Recall (§ 5.1) that one of
Price’s results was doubtful. Fries (1842/1974, p. 7 and 188/140)
indicated that P tends to disappear as n → ∞ and that therefore the
phenomenon described by the Laplace formula (he did not mention
the sunrise) cannot be a law of nature. He concluded that, when

103
102
issuing from repeated observations and posterior probability, it was
impossible to guess the prior probability and he thus hardly
recognized the Bernoulli theorem.
Fries decided that the solution of the Price problem just as the
Laplacean solution of the D’Alembert – Laplace problem (Note 5 to
Chapter 1) is based on philosophical (not mathematical) probabilities
and philosophical induction. Finally and without any special study
Fries (p. 157/139) declared that the MLSq was completely subjective.
All this he briefly stated already in his Introduction. It is exactly the
Bernoulli theorem that answers Fries. Laplace (1814/1995, p. 10)
actually remarked that his formula was necessarily justified by the
principle of insufficient reason but with the increase of the number of
trials the accepted assumption should have been gradually specified.
He himself (§ 7.2-1) recommended a similar procedure.
Zabell (1989) comprehensively studied the Price problem but all
but ignored Fries. I note that it was Fries who had effectively
introduced philosophical probabilities before Cournot (1843)
considered them in detail.
Back to my main subject. Laplace determined the population of
France given sampling data, and, for the first time ever, estimated the
precision of (his version of) sampling. Suppose that N and n are the
known numbers of yearly births in France as a whole and in some of
its regions and m is the population of those regions. Laplace naturally
assumed that

M = (m/n)N.

He then had to estimate its error, the fraction (Hald 1998, p. 288)

1 1


0
xN+n (1 – x)m–n+M–N dx ÷ 
0
xn(1 – x)m–n dx.

K. Pearson (1928a) noted some imperfections in Laplace’s

reasoning and achieved a reduction of the variance of his result: it
should have been multiplied by [(N – n) ÷ (N + n)]1/2. Here are his two
main remarks. First, Laplace considered (m, n) and (M, n) as
independent samples from the same infinite population whereas they
were not independent and the very existence of such a population was
doubtful. Second, Laplace chose for the magnitude sought an
absolutely inappropriate uniform prior distribution. Pearson also
negatively described Laplace’s calculation of the incomplete beta-
function. However, he (1934, Intro.) also owned that that problem
remained very difficult which thus actually exonerated Laplace.
Pearson’s first remark had to do with Laplace’s supplementary urn
problem. Suppose that an urn contains infinitely many white and black
balls. After n drawings without replacement m white balls were
extracted; a second sample of an unknown volume provided r white
balls. Denoting

k = nr/m + z,

104
103
Laplace derived a limit theorem

m3
P(|k – nr/m| < z) = 1 – 2  exp(– m3z2/S)dz,
S

S = 2nr(n – m)(m + r).

The limits of integration, as Laplace formally assumed, were z and

∞.
Later Markov (1900b) proved that, for an unknown m,

m r t
P [|  |< (1 / k )  (1/n) ] > 1 – 1/t2, t > 0.
n k 2

He (1914a) then specified that all prior probabilities of the appearance

of a white ball were equal to one another and proved, in addition, that
the same inequality of the Bienaymé – Chebyshev type held also for
“indefinite” [random] fractions m/n and r/k. As though the last of the
Mogicanes, Markov consistently refused to use the then new term,
random variable, see § 14.2-1.
6) In Chapter 7 Laplace studied the influence of a possible
inequality of probabilities assumed equal to each other. For example,
when tossing a coin the probability of heads can be (1 ± a)/2 with an
unknown a. Supposing that both signs were equally probable, Laplace
derived the probability of throwing n heads in succession

P = (1/2) [(1 + a )n + (1 – a)n] ÷ 2n

which for n > 1 was greater than 1/2n.

Suppose now (the general case) that the probability is not p, as
assumed, but (p + z), |z| ≤ a, with density φ(z). Then the probability of
a “compound” event y will be

a a

P =  y (p + z) φ(z) dz ÷  φ(z) dz
a a

(cf. formula (5) above). In case of an unknown density φ(z) it should

be replaced by the density of z (which he called the probability of z),
Laplace adds. The appearance of denominators in such formulas
seems to be unnecessary.
The special reasoning in this chapter, as also in one of the examples
in Chapter 3, can be justified by considering Markov chains. Actually
it is tantamount to stating that an infinite shuffling of a deck of cards
results in an equal probability of all of their possible arrangements
(Feller 1950, § 9 of Chapter 15).
Laplace then considers tickets put into an urn. Suppose, he says,
that the probabilities of their extraction are not equal to one another.
However, the inequalities will be reduced had the tickets been put into

105
104
the urn not in an assigned order, but according to their random
extraction from an auxiliary urn, and still more reduced in case of
additional auxiliary urns. Laplace had not proved this statement, he
only justified it by a general principle: randomness diminished when
subjected to more randomness. This is perhaps too general, but
Laplace’s example was tantamount to reshuffling a deck of cards.
7) Chapter 8 was devoted to population statistics, to the mean
durations of life and marriages. Laplace did not apply there any new
ideas or methods. However, he studied anew the Daniel Bernoulli
model of smallpox (§ 6.2.3), adopted more general assumptions and
arrived at a more general differential equation (Todhunter 1865, pp.
601 – 602).
8) In Chapter 9 Laplace considered calculations made in
connection with annuities and introduced the “Poisson” generalization
of the Jakob Bernoulli theorem (as noted by Molina 1930, p. 372).
Suppose that two contrary events can occur in each independent (as he
clearly indicated) trial i with probabilities qi and pi respectively, qi + pi
= 1, i = 1, 2, …, s, and that these events signify a gain ν and a loss µ,
respectively. For constant probabilities q and p the expected gain after
all these trials will be s(qν – pµ), as Laplace for some reason
concluded in a complicated way. He then estimated this magnitude for
the case of a large s by means of his limit theorem (3). Then,
generalizing the result obtained to variable probabilities, he introduced
the characteristic function of the final gain

[p1 + q1exp(ν1ωi)] [p2 + q2exp(ν2ωi)] … [ps + qsexp(νsωi)],

applied the inversion formula and obtained the normal distribution, all
this similar to the derivation of the law of distribution of a linear
function of observational errors (§ 7.2-4, see also his earlier memoirs).
9) In Chapter 10 Laplace described his thoughts about moral
expectation (§ 6.1.1). If the physical capital of a gambler is x, his
moral capital will be

y = klnx + lnh, h, x > 0.

Let Δx take values a, b, c, … with probabilities p, q, r, … Then

Ey = k[pln(x + a) + qln(x + b) + …] + lnh,

EΔy < EΔx. (8)

In other words, even a just game (EΔx = 0) is disadvantageous.

Todhunter (1865, p. 215) proved inequality (8) simpler than Laplace
did. However, a more general expression Ef(x) ≥ f (Ex) holds for
convex functions (Rao 1965, § 1e5) so that, if x > 0,

E(– lnx) ≥ – lnEx, Elnx ≤ lnEx < Ex.

Laplace then proved that the freight in marine shipping should be

evenly distributed among several vessels. I provide my own proof

106
105
(Sheynin 1972a/1977, pp. 111 – 113). Suppose that the capital of the
freightowner is a, the value of the freight, A, the probability of a safe
arrival of a vessel, p, and q = 1 – p. Then
a) If the freight is thus distributed on n vessels, the moral
expectation of the freightowner’s capital is (here and below 0 ≤ k ≤ n
and k is the number of the lost ships)

y(n) = Σ Cnk pn–k qk ln{[A(n – k)/n] + a}. (9)

b) Independently from n the corresponding expectation is equal to

the right side of (9) where, however, the logarithm is replaced by its
argument, so that obviously

a + AΣ[(n – k)/n]pn–k qk = a + Ap.

c) For any increasing function f (x) the moral expectation (9) is

restricted:

y(n) = Σ Cnk pn – kqkf{[A(n – k)/n] + a} < f (A + a) (p + q)n =

f (A + a).

d) Let f (x) be continuous and increasing and have a decreasing

derivative. Then y(n) increases monotonically but is restricted by the
moral expectation (9). The proof is here rather long and I refer readers
to my paper (1972b/1977, pp. 112 – 113).
Many authors after Laplace dwelt on moral expectation (cf. §
6.1.1); Fourier (1819) and Ostrogradsky (N. I. Fuss 1836, pp. 24 –
25), also see Ostrogradsky (1961, pp. 293 – 294), attempted to
develop it. The former thought that it should be individually specified
for each person whereas Fuss reported that Ostrogradsky

Did not at all admit Daniel Bernoulli’s hypothesis; he expressed the

“moral fortune” by an arbitrary function of the physical fortune and
he was able to solve the main problems connected with the moral
fortune with such breadth and as precisely as could only be desired.

Note, however, that the logarithmic function also appears in the

celebrated Weber – Fechner psychophysical law and is applied in the
theory of information. Nothing more is known about either of these
cases and since Fourier had not published anything more on this
subject, and Ostrogradsky himself never published anything, they both
possibly had second thoughts.
The connection of marine insurance with moral expectation
provided an occasion for Laplace (1814/1995, p. 89) to express
himself in favour of insurance of life and even to compare a nation
with an association

Whose members mutually protect their property by proportionally

supporting the costs of this protection.

107
106
 10) In the eleventh, the last, Chapter, and, in part, in Supplement 1
to the TAP, Laplace examined the probability of testimonies. Suppose
that an urn contains 1,000 numbered tickets. One of them is extracted,
and a witness states that that was ticket number i, 1 ≤ i ≤ 1,000. He
may tell the truth and be deceived or not; or lie, again being deceived
or not. Laplace calculated the probability of the fact testified by the
witness given the probabilities of all the four alternatives. In
accordance with one of his corollaries, the witness’s mistake or lie
becomes ever more probable the less likely is the fact considered in
itself (p. 460).
Laplace next introduced the prior probability of a studied event
confirmed by m witnesses and denied by n others. If it is 1/2 and the
probability of the truthfulness of each witness is p, then the probability
of the event was

p m n
P= .
p m  n  (1  p )m  n

Suppose now that the probabilities of truthfulness are pi > 1/2 and
the prior probability of the event is 1/n. If the event is reported by a
chain of r witnesses, then (p. 466)

( np1  1) ( np2  1) ... ( npr  1)

P = (1/n) + [(n – 1)/n]
( n  1) r

so that for n = 2 and n → ∞

P = (1/2) +(1/2) (2p1 – 1) (2p2 – 1) … (2pr – 1) and P = p1 p2 … pr

respectively.
Laplace next examines verdicts brought in by s independent judges
(jurors) assuming that each of them decides justly with probability
p > 1/2. The probability of a unanimous verdict is

ps + (1 – p)s = i/n.

Here, the right side is known (n is the total number of verdicts of

which i were brought in unanimously). For s = 3 (p. 470)

p = 1/2 + [(4i – n)/12n]1/2.

If 4i < n, it should have been concluded that Laplace’s (very

restrictive) assumptions were wrong; he did not however make this
remark.
If, in different notation, the probability of a just verdict reached by
each judge (juror) was unknown, and p judges condemned, and q of
them acquitted the defendant, the probability of a just final verdict was
(p. 527)

108
107
 1 1


1/ 2
up(1 – u)qdu ÷  up(1 – v)qdv
0

(cf. formulas from Laplace’s Chapter 6). Laplace stated that the
verdicts were independent, but only in passing (1816, p. 523). Poisson
(1837a, p. 4) indicated that Laplace had considered the defendant
innocent unless and until pronounced guilty: his formulas had not
included any prior probability of guilt. In Poisson’s opinion, this
should be assumed to exceed 1/2. I note that his remark had nothing to
do with any individual case.
In § 8.9.1 I return to the application of probability in jurisprudence;
here, I additionally refer to Zabell (1988a).
7.2. Theory of Errors
Laplace’s work on the theory of errors can be easily separated into
two stages. While treating it in the 18th century, he was applying the
comparatively new tool, the density3, and trying out several rules for
the selection of estimators of the true values of the constants sought
(cf. § 6.3.3). His equations proved too complicated and he had to
restrict his attention to the case of three observations. Later Laplace
proved (not rigorously) several versions of the CLT and was able to
drop his restriction, but he had to adopt other conditions. Here is
Bienaymé’s precise conclusion (1853/1867, p. 161), also noticed by
Idelson (1947, p. 11):

For almost forty years Laplace had been presenting […] memoirs
on probabilities, but […] had not wanted to combine them into a
general theory.

However, Bienaymé continued, the CLT [non-rigorously proved by

Laplace] enabled him to compile his TAP.
1) The Year 1774. Without substantiation, Laplace assumed that,
for any x1 and x2, the sought density ψ(x) of observational errors
satisfied the equation

ψ′(x2)/ψ′(x1) = ψ(x2)/ψ(x1)

and obtained

ψ(x) = (m/2)e–m|х|. (10)

Later, while discussing suchlike decisions, Laplace (1798 –

1825/1878 – 1882, t. 3, p. xi) argued that the adopted hypotheses
ought to be “incessantly rectified by new observations” until
“veritable causes or at least the laws of the phenomena” be
discovered. A similar passage occurred in his Essai (1814/1995, p.
116). Cf. Double et al (1835, pp. 176 – 177): the main means for
revealing the “vérité” were induction, analogy and hypotheses
founded on facts and “incessantly verified and rectified by new
observations”.
Suppose that the observations are a, b, and c, and p = b – a,

109
108
q = c – b. Issuing from the [likelihood function]

f(x) = ψ(x) ψ(p – x) ψ(p + q – x) (11)

rather than from the density, Laplace determined the parameter

sought, e [the median], with respect to curve (11); alternatively, he
applied the condition

∫|х – е| f(x)dx = min, |х| < + ∞

whence it followed that the integrals of f(x) over (– ∞; e] and [e; + ∞)

were equal to each other so that e, just the same, was the [median].
Neither function (10) nor (11) contained a location parameter.
Eisenhart (1964) noted that Pitman (1939) had made similar use of a
function of the type of (11).
For small values of m the magnitude x = e – a ≈ (2p + q)/3 and
therefore e coincided with the arithmetic mean and function (10)
became

ψ(x) = (m/2) (1 – m|х|) ≈ m/2 = Const.

Laplace was not satisfied with these corollaries and had thus rejected
the [median]. Note that for a [random variable] ξ with density (10)
varξ = 2/m2 so that a small m really invites trouble.
He then studied the case of an unknown parameter m by applying
the principle of inverse probability, that is, by the so-called Bayes
formula (5.1) with equal prior probabilities, but made a mistake in his
calculations (Sheynin 1977a, p. 7). Stigler (1986, pp. 115 – 116)
explained its cause but wrongly indicated that, since Laplace had not
then read the Bayes memoir, he was unable to borrow the “Bayes
formula”. Yes, indeed unable, but simply because that formula was
lacking in the work of his predecessor. Neither had he mentioned my
remark.
2) The Year 1781. Laplace again issued from the [likelihood
function] of the type of (11) and put forward four possible conditions
for determining the real value of the constant sought: the integrals of
f(x) or xf(x) over [– N; 0] and [N; 0], where N was the greatest possible
error, should be equal to each other; or, the value of the second
integral over [– N; N] should be minimal; and his final condition was
the application of the [maximum likelihood principle]. Recall,
however, that the curve (11) did not include a location parameter so
that it should have been somehow inserted. Anyway, Laplace decided
in favour of his third condition (which coincided with the first one).
So as to select a density, Laplace examined a special problem
(§ 7.1-2), and, not really convincingly, obtained a “mean” law of error

y = (1/2a) lna/|х|, |х| ≤ a. (12)

He referred to the principle of insufficient reason and noted that

function (12) was even and decreased with |х|, – that is, conformed to

110
109
the properties of “usual” errors; the restriction х ≠ 0 hardly bothered
him.
Next Laplace studied what might be called the multidimensional
Bayes method. Suppose that observational errors εi, i = 1, 2, …, n,
having “facility” xi have occurred. Then the probability of the
observed system of errors is proportional to

x1 x2 ...xn
P=
 ... x1x2 ...xn dx1dx2 ...dxn
where the integrals are taken over all possible values of each variable;
actually, Laplace considered a more general case in which each εi
occurred ki times. Multiplying the obtained expression by the product
of all the differentials, Laplace arrived at the probability element (of
the differential) of an n-dimensional random vector. It was now
possible for him to determine the law of distribution of observational
errors provided that the prior information stated above was available.
In connection with density (12) Laplace carried out a special study
introducing the Dirac delta-function which had already appeared in
Euler’s works, see Truesdell (1984, p. 447, Note 4, without an exact
reference). One of Laplace’s conditions for determining an estimator
xo of the real value of the constant sought given its observations x1, x2,
…, xn was (§ 7.2-1) that the integrals

 ψ(x – x1) ψ(x – x2) … ψ(x – xn)

over [– a; xo] and [xo; a] should be equal to each other. Laplace

indicated, without proof, that in case of an infinite a the arithmetic
mean can be obtained from the density law (12). He apparently
thought that the function (12) then became constant, cf. his similar
derivation in § 7.2-1.
Laplace then went over to a “much more general proposition” for
density

y = φ(αx) = φ(– αx) = q, if αx = 0, and y = 0 otherwise; α → 0.

In actual fact, he considered a sequence of functions φ(αx) such that

φ(αx) = q(α), α = {α1; α2; … αn; …} → 0.

If αx = t, then

φ(t) = q if t = 0, |х| < + ∞; φ(t) = 0 otherwise (when t ≠ 0, |х| = + ∞),

and, obviously,



φ(t)dt = C ( = 1).

111
110
 Laplace had not written these last equalities, but I think that he had
actually introduced the Dirac delta-function

φ(t) = lim(λ/√π)exp(– λ2t2), λ → ∞.

Laplace could have regarded the equalities above as representing a

uniform distribution of observational errors having an arbitrarily wide,
rather than assigned beforehand domain. His proposition consisted in
that the unknown constant xo was equal to the appropriate arithmetic
mean, but it can hardly be proved in the context of generalized
functions: Laplace had to consider the integral

 φ[α(x – x1)] φ[α(x – x2)] … φ[α(x – xn)],

which does not exist in their language.

3) The Years 1810 – 1811. Laplace (1810a) considered n
[independent] discrete random errors (or magnitudes) uniformly
distributed on interval [– h; h ]. After applying a particular case of
characteristic functions and the inversion formula, he proved, very
carelessly and non-rigorously, that, in modern notation, as n → ∞,

|  i | 3
s

lim P (
n
≤ s) =
 2 
0
exp(– x2/2σ2)dx, i = 1, 2, …, n, (13)

where σ2 = h2/3 was the variance of each ξi. He then generalized his
derivation to identically but arbitrarily distributed variables possessing
variance. When proving the [CLT] he (1810a/1898, p. 304) made use
of an integral of a complex-valued function and remarked that he
hoped to interest “géométres” in that innovation and thus separated
himself from (pure) mathematicians, see also similar reservations
elsewhere (Laplace 1774/1891, p. 62; 1812/1886, p. 365).
In a supplement to this memoir Laplace (1810b), apparently
following Gauss, turned his attention to the MLSq, and derived it
without making any assumptions about the arithmetic mean (cf.
§ 9A.2-2), but he had to consider the case of a large number of
observations and to suppose that the means of their separate groups
were also normally distributed.
Soon enough Laplace (1811) returned to least squares. This time he
multiplied the observational equations in one unknown

aix + si = εi, i = 1, 2, …, n,

where the right sides were errors rather than residuals, by indefinite
multipliers qi and summed the obtained expressions:

[aq]x + [sq] = [εq].

The estimator sought was

xo = – [sq]/[aq] + [εq]/[aq]  – [sq]/[aq] + m.

112
111
Tacitly assuming that all the multipliers qi were of the same order,
Laplace non-rigorously proved another version of the CLT for n → ∞:


1 [qq]
P(m = α) = exp(– α2/2 σ 2m ), σ 2m = k″ , k″ =  x2ψ(x)dx
σ m 2π [aq]2 

where ψ(x) was an even density of observational errors possessing

variance.
Then Laplace determined the multipliers by introducing the
condition



|z|P(z)dz = min (14)

which led him to equalities q = µai, and then to the principle of least
squares (in the case of one unknown)

x = [as] ÷ [aa].

Finally, Laplace generalized his account to the case of two

unknowns. He multiplied the observational equations (in two
unknowns) by two sets of indefinite multipliers {mi} and {ni}4 and
obtained a bivariate normal distribution for independent components
and, once more applying the condition of least absolute expectation,
arrived at the principle of least squares. Todhunter (1865, pp. 578 –
588), who referred to Ellis (1849), studied the case of three or more
unknowns.
And so, the derived principle essentially depended on the existence
of the normal distribution. Indeed, the CLT (and therefore a large
number of observations) was necessary and the use of conditions of
the type of (14) would have otherwise been extremely difficult. No
wonder that Laplace’s theory had not been enjoying practical success.
I adduce a wrong statement formulated on this point by Tsinger (1862,
p. 1) who compared the importance of the Gaussian and the
Laplacean approaches:

Laplace provided a rigorous [?] and impartial investigation […]; it

can be seen from his analysis that the results of the method of least
squares receive a more or less significant probability only on the
condition of a large number of observations; […] Gauss endeavoured,
on the basis of extraneous considerations, to attach to this method an
absolute significance [wrong]. With a restricted number of
observations we have no possibility at all to expect a mutual
cancellation of errors and […] any combination of observations can
[…] equally lead to an increase of errors as to their diminution.

With regard to Gauss see § 9A. Here, I note that Tsinger lumped
together both justifications of the MLSq due to Gauss and that
practice demanded the treatment of a finite (and sometimes a small)

113
112
number of observations rather than limit theorems. Tsinger’s high-
handed attitude towards Gauss (and his blind respect for Laplace) was
not an isolated occurrence, see §§ 10.8.5 and 13.2-7. This was partly
occasioned by Gauss’ disrespect of Legendre (§ 9A.1.2) and,
consequently, by the (antiscientific) ignorance of Gauss by French
mathematicians including Poisson, and partly because they had been
unquestionably following Laplace’s approach (Tsinger was not the
only one so affected).
Even not so long ago Eisenhart (1964, p. 24) noted that the
existence of the second formulation of the MLSq

Seems to be virtually unknown to almost all American users of least

squares except students of advanced mathematical statistics.

4) Chapter 4 of the TAP. Laplace non-rigorously proved the CLT

for sums and sums of absolute values of independent, identically
distributed errors restricted in value as well as for the sums of their
squares and for their linear functions. All, or almost all of this material
had already been contained in his previous memoirs although in 1811
he only proved the local theorem for linear functions of errors.
In § 23 Laplace formulated his aim: to study the mean result of
“observations nombreuses et non faites encore …” This was
apparently the first explicit statement concerning general populations;
see § 14.2-1 for the appropriate opinion of Chebyshev and Markov
and § 10.8.5 for similar statements in physics.
5) In Supplement 1 to the TAP Laplace (1816) considered
observational equations in (let us say) two unknowns

ai x + bi y + li = vi, i = 1, 2, …, s.

Suppose that Δx and Δy are the errors of the least-squares estimators

of the unknowns, denote the even density of the observational errors
by φ(u/n) with |u|  n, the moments by the letter k with appropriate
subscripts, ξ = Δx√s, η = Δy√s,

k s
[vv ] 2k 2 n 2 s
β2 =
kk 4  2k 2
2
, Q2
= 
i 1
(ai ξ + bi η)2
and t =
s

k
.

Laplace calculated

P(ξ; η) ~ exp{– Q2(2[vv] – 2t√s)},

P(t) ~ exp{– (β2/4n4) [t + (Q2/s√s)]2}.

It thus occurred that P(ξ; η; t) which he also obtained showed that t

was independent of ξ; η; or, that the sample variance was independent
from the estimators of the unknowns, cf. §9A-5; to repeat, the
observational errors were assumed to possess an even distribution, –
and a normal distribution in the limit. For a proof of Laplace’s result
see Meadowcroft (1920).

114
113
 Laplace also considered non-even distributions and recommended,
in such a case, to demand that the sum of vi be zero. Since [av] = 0 is
the first normal equation written down in another form, this demand is
fulfilled for ai = const (or bi = const); otherwise, it is an additional
normal equation corresponding to a fictitious unknown, the mean
systematic error of observations.
Finally, Laplace derived a formula for estimating the precision.
Without explanation (which appeared on p. 571 of his Supplement 2)
he approximated the squared sum of the real errors by the same sum
of the residuals and arrived at an estimator of the variance

[vv]
m= .
s

Without naming anyone Gauss (1823b, §§ 37 – 38) remarked that

that formula was not good enough, see § 9A.4-6. Interestingly,
Laplace (1814/1995, p. 45) stated that

The weight of the mean result increases like the number of

observations divided [the French word was indeed divisé] by the
number of parameters.

6) In Supplement 2 to the TAP Laplace (1818) adopted the normal

law as the distribution of observational errors themselves and not only
as the law for their means. Indeed, the new “repeating” theodolites
substantially reduced the error of reading and equated its order with
that of the second main error of the measurement of angles in
triangulation, the error of sighting, which was not yet sufficient. The
error in the sum of the three angles of a triangle (the appropriate
discrepancy, or its closing) could therefore be also regarded as
normally distributed with density

φ(x) = h /3π exp(– hx2/3),

where h = 1/2σ2 was the measure of precision of an angle.

Tacitly assuming that h was a [random variable], Laplace proved
that

Eh = 3n/2θ2, ψ(x) = hn/2exp[(– h/3)θ2]

were its expectation and density, θ2, the sum of n squares of the
triangular discrepancies. He computed the probability of the joint
realization of errors obeying the normal law in a triangle and
concluded that an equal distribution of the closing of the triangle
among its angles was advantageous. The MLSq leads to the same
conclusion, and for that matter, irrespective of the normal law. Then,
when adjusting a chain of triangles rather than a separate triangle, two
additional conditions (never mentioned by Laplace) have to be
allowed for, – those corresponding to the existence of two baselines

115
114
and, possibly, two astronomical azimuths, – and a preliminary
distribution of the closings is of course possible but not necessary.
And so, let the observational errors have density

φ(x) = h /π exp(– hx2).

Denote the closing of triangle i by Ti and suppose that the errors of the
angles αi, βi and γi already obey the condition

αi + βi + γi = Ti.

Laplace derived the relations

P(αi; Ti) ~ h /3π exp [– (h/3)Ti2],

P(T1; T2; …; Tn) ~ ( h /3π )n/2 exp {– (h/3)[TT]},


h n / 2 e  ( h / 3)[TT ] 3n  2 3n
P(h) ~ 
, Eh =  hP(h)dh =  .
2 [TT ] 2 [TT ]
h
n / 2  ( h / 3)[TT ] 0
e dh
0

The error involved in the approximation just above can be easily

estimated: in his Supplement 3 (ca. 1819) Laplace took n1 = 26 and
n2 = 107. Finally, supposing that h = Eh,

σ = 1/ 2h = [TT ] / 3n ,

not a bad result (improved by the approximation above!).

Laplace next investigates the adjustment of equations in one
unknown by the MLSq for normally distributed errors. The interesting
point here is that he had not indicated that the distribution of the
residuals was also normal; in other words, that that distribution was
[stable].
In the same Supplement Laplace discussed the Boscovich method
of adjusting meridian arc measurements (§ 6.3.2-3). Write the initial
equations in a different form,

pi y – ai + xi = 0, i = 1, 2, …, n, pi > 0, a1/p1 > a2/p2 > … > an/pn.

The second unknown is presumed to be eliminated, and xi are the

residual free terms. The Boscovich conditions, or, rather, his second
condition, leads to

y = ar/pr

with error – xr/pr, i.e., to the calculation of this second unknown from
one equation only. This latter is determined by inequalities

p1 + p2 + … + pr–1 < pr+ pr+1 + … + pn,

p1 + p2 + … + pr > pr+1 + pr+2 + … + pn.

116
115
 And so, these inequalities determine the sample median of the
fractions ai/pi. Suppose now that the observational errors have an even
density φ(x) and

k″ =  x2φ(x)dx.
0

Then, as Laplace showed, basing his derivation on variances5 rather

than on absolute expectations as before, the Boscovich method was
preferable to the MLSq if, and only if,

4φ2(0) > 1/(2k″).

According to Kolmogorov (1931), the median is preferable to the

arithmetic mean if

1/[2σφ(m)] < 1, σ2 = 2k″,

and m is the population median.

While translating Laplace’s Mécanique Céleste into English,
Bowditch (Laplace 1798 – 1825/1832, vol. 2, § 40, Note) stated:

The method of least squares, when applied to a system of

observations, in which one of the extreme errors is very great, does
not generally give so correct a result as the method proposed by
Boscovich […]; the reason is, that in the former method, this extreme
error [like any other] affects the result in proportion to the second
power of the error; but in the other method, it is as the first power.

In other words, the robustness of the Boscovich method is

occasioned by its connection with the median.
7) In Supplement 3 to the TAP Laplace (ca. 1819) begins by
evaluating a chain of 26 triangles (Perpignan – Formentera) which
was a part of a much longer independent chain of 107 triangles. For
the same normal distribution φ(x) he has

 

ε =  | x |  ( x )dx , ε′ =  x 2 ( x )dx , ε′ = πε2/2.

 

The empirical value of ε for the longer chain was

(1/107) (|T1| + |T2| + … + |T107|) = 1.62 so that (1.622/2)π = 4.13.

With subscripts 1 and 2 denoting the shorter and the longer chains
respectively, Laplace has

[TT]1 = 4.13·26 = 107.8; empirical value, (26/107)[TT]2 = 108.8.

117
116
 This calculation shows that, first, Laplace preferred to evaluate
[TT]1 by [TT]2 rather than use its actual value which was hardly
correct since the pertinent conditions of observations could well have
been different. Second, Laplace has thus qualitatively checked the
realization of the normal law.
Next Laplace considers the adjustment of equations

pi x = ai + miαi + niβi, i = 1, 2, …, n

in one unknown, x, and independent errors αi and βi both distributed

normally with differing measures of precision; he only mentioned
independence later (1827/1904, p. 349). Laplace explained his
calculation by referring to his pp. 601 – 603 of the same Supplement
which does not help but at least he concluded that the error of x was
distributed normally so that after all he knew that the normal law was
[stable], cf. § 9A.2-6. However, the variance of the emerged law
depended on the application of the MLSq which meant that the result
just formulated was not sufficiently general.
Also in 1827 Laplace (1904, p. 343) stated that the MLSq was a
particular case of the “most advantageous” method of adjustment
(based on the minimal value of the expected absolute error and the
presumed normal law, see end of § 7.2-3). Before 1823, he would
have been partly in the right, but not afterwards, not since Gauss’
second justification of the MLSq had appeared.
8) In Note 4 to this Chapter I indicated that Laplace had
successfully treated the case of dependence between random
variables. Elsewhere, however, he (1827) somehow erred when
investigating the atmospheric pressure. Its mean daily variation in
Paris during 11 years was 0.763 mm, or 0.940 mm, if, during the same
years, only three months, from February to April, were taken into
consideration. When attempting to find out whether the difference
between the two values was significant, Laplace had not indicated that
they were not independent6. He made one more mistake: when solving
his equations in two unknowns, the action of the Moon and the time of
the maximal pressure, he had not stated that, again, the appropriate
free terms were not independent. Without justifying his remark, K.
Pearson (1914 – 1930, vol. 3A, p. 1) stated that

Condorcet often and Laplace occasionally failed because [the] idea

of correlation was not in their mind.

Elsewhere, he (1978, p. 658) left a similar remark, again without

substantiation; there also, on p. 671, he added that Laplace was “rarely
a good collector, or a safe guide in handling [the data]”. Pearson
exaggerated: on the then possible scientific level, and issuing from
observations, Laplace proved that the Solar system will remain stable
for a long time and completed the explanation of the motion of its
bodies in accordance with the law of universal gravitation.
7.3. Philosophical Views
Laplace (1814/1995, p. 2) stated that, for a mind, able to
“comprehend” all the natural forces, and to “submit these data to

118
117
analysis”, there would exist no randomness “and the future, like the
past, would be open” to it. Nowadays, this opinion cannot be upheld
because of the recently discovered phenomenon of chaos (§ 1.2.4);
however, other remarks are also in order.
a) Such a mind does not exist and neither is there any
comprehensive theory of insignificant phenomena, a fact which
Laplace undoubtedly knew. He therefore actually recognized
randomness (Dorfman 1974, p. 265).
b) In addition, there exist unstable movements, sensitive to small
changes of initial conditions, cf. § 11.2-9.
c) Already previous scholars, for example, Maupertuis (1756a, p.
300) and Boscovich (1758, §385), kept to the “Laplacean
determinism”. Both mentioned calculations of past and future (“to
infinity on either side”, as Boscovich maintained) but, owing to
obvious obstacles, see above Item a, both disclaimed any such
possibility.
In his Essai Laplace (1814/1995, p. 37) additionally provided
examples of “statistical determinism”, – of the stability of the number
of dead letters and of the profits made by those who ran lotteries. He
explained all this by the action of the LLN (more precisely, by its,
then barely known, Poisson form, see § 7.1-8). Participation in
lotteries only depends on free will, cf. Quetelet’s similar statement in
§ 10.5 and Petty’s opinion (§ 2.1.1)7.
In his early memoirs, Laplace (e.g., 1776/1891, pp. 144 – 145), like
Newton (§ 2.2.3), had not recognized randomness and explained it by
ignorance of the appropriate causes, or by the complexity of the
studied phenomenon. He even declared that the theory of probability,
that estimated the degrees of likelihood of phenomena, was indebted
for its origin to the weakness of the mind and a similar statement
occurred in his Essai (1814/1995, p. 3). Thus, probability became for
him an applied mathematical discipline servicing natural sciences8 (cf.
§ 0.1) and, even for this reason alone, he had not separated
mathematical statistics from it, although he (1774/1891, p. 56) noted
the appearance of “un nouveau genre de problème sur les hasards”,
and even (1781/1893, p. 383) of “une nouvelle branche de la théorie
des probabilités”9.
A curious statement deserves to be included (Laplace 1796/1884, p.
504):

Had the Solar system been formed perfectly orderly, the orbits of
the bodies composing it would have been circles whose planes
coincided with the plane of the Solar equator. We can perceive
however that the countless variations, that should have existed in the
temperatures and densities of the diverse parts of these grand masses,
gave rise to the eccentricities of their orbits and the deviations of their
movement from the plane of that equator.

The causes mentioned by Laplace could have hardly be called

external, and the main relevant explanation of randomness, deviation
from the laws of nature, persisted. Leaving aside the planes of the

119
118
planetary orbits, I deny his (and Kant’s, and Kepler’s, see § 1.2.4)
opinion concerning eccentricities.
Newton theoretically proved that the Keplerian laws of planetary
motion resulted from his law of universal gravitation. In my context, it
is necessary to stress: he also proved that the eccentricity of the orbit
of a given planet is determined by the planet’s initial velocity. For
some greater values of that velocity the orbit will become parabolic
(with its eccentricity equal to unity, rather than less than unity as in
the case of ellipses), for other still greater values, hyperbolic (with
eccentricities greater than unity). And for a certain value of that
velocity an elliptic orbit will become circular. But is it really
necessary to refer to a rigorous proof? Indeed, it is difficult to imagine
that such changes do not occur gradually, that, consequently, the
eccentricity does not vary continuously with the velocity.
All these findings, as Newton proved, persisted for planets (not
material points) having uniform density. So what should we think
about Laplace’s explanation? I believe that the variations of densities
(but hardly temperatures) peculiar to a given planet could have
somewhat corrupted the eccentricity caused by its initial velocity. I am
unable to say whether they could have also caused a corruption of
some other kind, but in any case I need not discuss this problem.
So it really seems that Laplace (and Kant) were mistaken (Kepler
was obviously ignorant of the law of universal gravitation). I am not
sure that Kant had studied Newton attentively enough, but Laplace
certainly did, although a bit later, in t. 1 of his Traité de Méc. Cél.
(1798/1878, Livre 2, chapters 3 and 4).
Witness finally Fourier’s comment (1829, p. 379) on Laplace’s
Exposition: it “is an ingenious epitome of the principal discoveries”.
And on the same page, discussing Laplace’s “historical works” (to
whose province the Exposition belonged):

If he writes the history of great astronomical discoveries, he

becomes a model of elegance and precision. No leading fact ever
escapes him. […] Whatever he omits does not deserve to be cited.

7.4. Conclusions
Laplace collected his earlier memoirs in one contribution which
cannot, however, be regarded as a single whole. He never thought
about solving similar problems in a similar way (and his Essai was not
a masterpiece of scientific-popular literature, see Note 1 to this
Chapter). Then, many authors complained that Laplace had described
his reasoning too concisely. Here, for example, is what Bowditch
(Todhunter 1865, p. 478), the translator of Laplace’s Traité de
mécanique céleste into English, sorrowfully remarked:

Whenever I meet in La Place with the words ‘Thus it plainly

appears’ I am sure that hours, and perhaps days of hard study will
alone enable me to discover how it plainly appears.

This can also be said about the TAP.

120
119
 The Laplacean definition of probability (to repeat: first introduced
by De Moivre, see § 4.3) was of course unsatisfactory, but nothing
better had appeared until the advent of the axiomatic theory (or, the
Mises debatable formula). Here is the testimony of Kamke (1933, p.
14): In 1910, it was said at Göttingen University that probability was a
number situated between 0 and 1 about which nothing more was
known. Similar statements were due to Mises in 1919, to Keynes in
1921, and to P. Lévy (who was born in 1886) in his earlier life
(Cramér 1976, § 2.1) as well as to Poincaré (§ 11.2-1) and Markov
(§ 14.1-5).
But the opinion of Doob (1989) was even more interesting. In 1946

To most mathematicians mathematical probability was to

mathematics as black marketing to marketing; […] the confusion
between probability and the phenomena to which it is applied […]
still plagues the subject; [the significance of the Kolmogorov
monograph] was not appreciated for years, and some mathematicians
sneered that […] perhaps probability needed rigor, but surely not
rigor mortis; […] the role of measure theory in probability […] still
embarrasses some who like to think that mathematical probability is
not a part of analysis.

All this means that Laplace is here exonerated. However, he had not
even heuristically introduced the notion of random variable and was
therefore unable to study densities or characteristic functions as
mathematical objects. His theory of probability quite properly
remained an applied mathematical discipline since he made
outstanding discoveries in mathematics, astronomy and physics.
However, it did not yield to development which necessitated its
construction anew. It is opportune to note that Maxwell referred to
Laplace only twice (§ 10.9.5), and Boltzmann did not mention him at
all.
Laplace had not regarded himself as a pure mathematician (§ 7.2-3)
and I quote the opinion of Fourier (1829, pp. 375 – 376):

We cannot affirm that it was his destiny to create a science entirely

new, like Galileo and Archimedes; to give to mathematical doctrines
principles original and of immense extent, like Descartes, Newton and
Leibniz; or, like Newton, to be the first to transport himself into the
heavens, and to extend to all the universe the terrestrial dynamics of
Galileo: but Laplace was born to perfect everything, to exhaust
everything, and to drive back every limit, in order to solve what might
have appeared incapable of solution. He would have completed the
science of the heavens, if that science could have been completed.

Laplace introduced partial differential equations and, effectively,

stochastic processes into probability, and non-rigorously proved
several versions of the CLT by applying characteristic functions and
the inversion formula.
On that basis, he constructed his version of the theory of errors,
which essentially depended on the existence of a large number of

121
120
normally distributed observational errors and was therefore
unsuccessful. In the not yet existing mathematical statistics Laplace
investigated the statistical significance of the results of observation,
introduced the method of statistical simulation, studied his version of
sampling and extended the applicability of the Bayesian approach to
statistical problems.
He knew the Dirichlet formula (even in a generalized version),
introduced the Dirac delta-function and integrals of complex-valued
functions. He had also indicated (long before the strong law of large
numbers became known) that in probability theory limit was
understood in a special way. Molina (1930, p. 386) quoted his memoir
(1786/1894, p. 308) where Laplace had contrasted (although not
clearly enough) the “approximations” admitted in the theory of
probability with certainty provided in analysis.
It is my belief that a balanced opinion of a classic is needed, and I
am now discussing Laplace’s mistakes and shortcomings. That his
contributions are extremely difficult to understand is generally known,
see for example Bowditch’s opinion above. Then, Laplace was
“extremely careless in his reasoning and in carrying out formal
transformations” (Gnedenko & Sheynin 1978/2001, p. 224 with
examples attached). And here are some of my present comments.
1) See my comments just above the quotation from Doob.
2) Laplace made a mistake when studying the Buffon problem
(§ 7.1-4).
3) He applied an unsuitable model when calculating the population
of France (Ibidem). And he presented his final estimate (1812, pp. 399
and 401) in a hardly understandable way; in any case, Poisson (1812)
wrongly reported his result.
4) There also, Laplace bravely, and without any reservation,
calculated the probability of a certain demographic relation persisting
for a hundred years to come.
5) He made a mistake while discussing the treatment of
observations (§ 7.2-1).
6) Here is Laplace’s opinion (1814/1995, p. 81) about mortality
tables: “There is a very simple way of constructing” [them] from the
registers of births and deaths”. But the main point is to study the
plausibility of these registers, to single out possible corruptions and
exceptional circumstances etc. Then, the boundaries of the constructed
mortality table have to be determined both in time and territory.
7) Poisson (§ 8.9.1) noted that in applying probability theory to
jurisprudence, Laplace had not allowed for a prior probability of
defendants’ guilt. Such a probability has nothing in common with the
presumption of innocence in each individual case, cf. Quetelet’s
inclinations to crime (§ 10.5).
8) The method of least squares. The proper attitude for Laplace
would have been to acknowledge the Gaussian demand (§ 9A.2) for
studying the treatment of a small number of observations and to
restrict therefore the importance of his own results. Instead, he insisted
on his own approach and virtually neglected Gauss. Later French
scientists including Poisson followed suit and even the most eminent
mathematicians (or at least those of them who had not studied

122
121
attentively the treatment of observations) became confused. When
proceeding to prove the CLT, Chebyshev (§ 13.1-4) remarked that it
leads to the MLSq!
9) Laplace did not explain, for example in his Essai (1814), the
dialectic of randomness and necessity.
10) See my comments on the eccentricities of the planetary orbits in
§ 7.3. The last edition of his book of 1796 during his lifetime appeared
in 1813 without any corrections of that subject.

Notes
1. The Essai ran through a number of editions and was translated into many
languages. It attracted the public to probability, but the complete lack of formulas
there hindered its understanding. The appearance of Quetelet’s superficial
contributions written in good style (§ 10.5) had a negative effect on the fate of the
Essai.
2. For a simpler derivation of its equation see Todhunter (1865, pp. 545 – 546).
3. Laplace applied several pertinent terms. In his TAP, he finally chose loi de
probabilité or loi des erreurs.
4. The quantities [εm] and [εn] which appeared here were not independent.
Without indicating this, Laplace correctly solved his problem.
5. In his Supplement 3 Laplace once more applied the variance as the main
measure of precision of observations.
6. Cf. Note 4. Retaining excessive decimals (see above) was of course traditional,
see Note 11 to Chapter 1.
7. Kant (1763/1912, p. 111) indicated that the relative number of marriages
(which obviously depended on free will) was constant.
8. The subjects discussed by Laplace in his Exposition (1796/1884) had not
demanded stochastic reasoning (see however end of this subsection), but he
undoubtedly applied them, for example, in the Traité (1798 – 1825), to say nothing
about the treatment of observations, and his determinism had not hindered him at all.
Thus, Laplace (1812/1886, p. 361) stated that a certain magnitude, although having
been indicated by [numerous] observation[s], was neglected by most astronomers,
but that he had proved its high probability and then ascertained its reality (although
did not provide his calculations). In general, unavoidable ignorance concerning a
single random event becomes a cognizable regularity.
9. Lagrange, in a letter to Laplace of 13.1.1775, see t. 14 of his Oeuvres, 1892, p.
58, used this latter expression. Inductive stochastic conclusions occurred in the
Talmud (§ 1.1.2) and Arbuthnot’s memoir (§ 2.2.4) and the work of many other
authors, especially Bayes, which had appeared before Laplace, might be today
attributed, at least in part, to mathematical statistics.

8. Poisson
Poisson’s publications concerning the theory of probability began to
appear in 1811 when he published abstracts of two of Laplace’s early
memoirs, and in 1812 he continued this work by an abstract of
Laplace’s Théorie analytique. These abstracts, published in Nouv.
Bull. des Sciences Soc. Philomatique de Paris, were not really
interesting, but who could have done better? Only Fourier. See Bru

123
122
(1981; 2013) for a description of the French mathematical community
during Poisson’s lifetime and his general and for many years dominant
role there. Bru also took care to explain much of Poisson’s
mathematics.
Like Laplace, Poisson had published a number of memoirs on the
theory of probability, then combined them in his monograph (1837a)
whose juridical title did not reflect its contents; only its subtitle
promised to discuss, as a preliminary, the general principles of the
calculus of probability. I describe both this contribution (referring
only to its page numbers) and his other works. First, however, I quote
Poisson’s statement (p. 1) about the place of probability in
mathematics and then describe the scope of the Elements of the
Calculus of Probability and Social Arithmetic as formulated by him
(Poisson 1837c, p. 26).
And so, probability became

Une des principales branches des mathématiques, soit par le

nombre et l’utilité de ses applications, soit par le genre d’analyse
auquel il a donne naissance [to which it gave birth].

The Elements (1837c) listed: 1) Topics of probability itself (general

principles, the Bernoulli theorem, probabilities of future events
derived from the probabilities of similar previous events). 2) Tables of
mortality, mean duration of life, smallpox, inoculation and
vaccination. Here also, expectation, cf. § 13.3. 3) Institutions
depending on probabilities of events (annuities, insurance, loans).
4) Mean values of a large number of observations.
The soon forgotten term social arithmetic (appearing also below in
§ 8.9) thus designated population and medical statistics. Now, we
would rather say social statistics.
8.1. Subjective Probability
The aim of the calculus of probability, as Poisson (pp. 35 – 36)
maintained, was the determination, in any doubtful “questions”, of the
ratio of the cases favourable for the occurrence of an event to all
possible cases, and its principles should be regarded as “un
supplément nécessaire de la logique”. He (pp. 30 and 31) remarked
that probability changed with experience, and was subjective, but that
the chance of an event remained constant. Already Leibniz (§ 2.1.1)
and then De Morgan (1847)1 and Boole (1952) attempted to justify
probability by elements of mathematical logic, see also Halperin
(1988).
The stressed difference between chance and probability (also
recognized by Cournot, see § 10.3) is now forgotten, although
Poisson attempted to adhere to it. Thus (p. 47), he showed that the
subjective probability of extracting a white ball from an urn
containing white and black balls in an unknown proportion was equal
to 1/2 “as it should have been”. This conforms to the principles of the
theory of information and he himself was satisfied with the result
obtained since it corresponded to “la perfaite perplexite de notre
esprit”.

124
123
 Davidov (1854, p. 66) who was well acquainted with foreign
literature, noted:

Vague ideas on probability and an inexact distinction between

subjective and objective probabilities are among the main obstacles
against a speedy development of practical medicine.

8.2. Two New Notions

Poisson (1829, § 1) defined the distribution function of a discrete
random variable ξ as

F(x) = P(ξ < x)

and (Ibidem) introduced the density as the derivative of F(x). Later he

(1837b, pp. 63 and 80) similarly treated the continuous case. Davidov
(1885) and Liapunov (1900) had noted his innovation, but
distribution functions only became generally used in the 20th century.
Poisson (pp. 140 – 141) was also the first to introduce the notion of
a discrete random variable although he named it by an obviously
provisional term, chose A2. He then (p. 254) considered a random
variable with values being multiples of some ω, assumed that ω → 0
and thus went over, in accordance with the tradition of his day, to a
continuous variable3. As compared with Simpson, who studied
random observational errors (§ 6.3.1), Poisson’s innovation here was a
formal heuristic definition of a random variable and its more general
(not necessarily connected with the theory of errors) understanding.
8.3. The De Moivre – Laplace Limit Theorem
Poisson (1837a, p. 189) provided his own derivation of that theorem
by issuing from the probability4 of the occurrence of contrary events A
and B not less than m times (not more than n times) in µ = m + n
Bernoulli trials

m ( m  1) 2 m ( m  1) ... ( m  n  1) n
P = pm{1 + mq + q +…+ q } = (1)
2! n!

 
xn
=  Xdx ÷  Xdx, X= , (2)
a 0
(1  x ) 1

where p and q were the probabilities of the occurrence of these events

in a single trial, p + q = 1.
His results were, however, tantamount to formula (7.3), see Sheynin
(1978b, pp. 253 – 254). Montmort (1708/1713, p. 244), also see
Todhunter (1865, p. 9), knew formula (1) and formula (2) occurred in
Laplace’s TAP, Chapter 6.
For small values of q Poisson (p. 205) derived the approximation

P ≈ e-ω(1 + ω + ω2/2! + … + ωn/n!), (3)

where mq ≈ µq = ω. He had not provided the expression

125
124
 P(ξ = m) = e-ωωm/m!.

8.4. Sampling Without Replacement

Poisson (pp. 231 – 234) examined sampling without replacement
from an urn containing a white balls and b black ones (a + b = c) and
applied the result obtained for appraising a model of France’s electoral
system. Suppose that the sample contained m white, and n black balls
(m + n = s). Its probability, as Poisson indicated, was represented by
the [hypergeometric] distribution. For large a and b as compared with
the sample, Poisson determined an approximate expression for that
probability under an additional condition

n > m. (4)

If a series of k such samples are made, then

s1 + s2 + … + sk = c.

After calculating the probability of the condition (4) being fulfilled j

times out of k, Poisson concluded that, even if b only somewhat
exceeded a, j will apparently be too large. For k = 459, which was the
number of electoral districts in France, c = 200,000, equal to the
number of the voters (less than 1% of the population!). Suppose also
that each voter is a member of one of the two existing parties; that the
voters are randomly distributed over the districts; and that the
proportion of party memberships is 90.5:100. Then, as Poisson
concluded, remarking, however, that his model was too simplified, the
probability of electing a deputy belonging to the less numerous party
was very low.
Poisson (1825 – 1826) studied sampling without replacement also
in connection with a generally known game. Cards are extracted one
by one from six decks shuffled together as a single whole until the
sum of the points in the sample obtained was in the interval [31; 40].
The sample is not returned and a second sample of the same kind is
made. It is required to determine the probability that the sums of the
points are equal. Here is his solution, see Sheynin (1978b, pp. 290 –
292), Gnedenko & Sheynin (1978/2001, pp. 236 – 237) and § 8.5
below.
An urn contains x1 balls numbered one, x2 balls numbered two, …,
xi balls numbered i (x1 + x2 + … + xi = s). What is the probability of
extracting without replacement a1 balls numbered one, a2 balls
numbered two, …, ai balls numbered i (a1 + a2 + … + ai = n) if

a1 + 2a2 + … + iai = x? (5)

Without taking into account this condition Poisson got

n !( s  n)! x1 ! x2 ! xi !
P= ... 
s! a1 !( x1  a1 )! a2 !( x2  a2 )! ai !( xi  ai )!

126
125
 1
(s + 1)  (1  y ) s Ydy,
0

Y=
a1 a2 ai
x1 !  y  x2 !  y  xi !  y 
   
a1 !( x1  a1 )! 1  y  a2 !( x2  a2 )! 1  y 
...
ai !( xi  ai )! 1  y 
.

For i = 2 the system of probabilities thus calculated defines a

hypergeometric distribution, as it is now called. So as to allow for
condition (5), Poisson replaced Y by the sum of its values
corresponding to such sets of {a1; a2; …; ai} for which that restriction
is fulfilled.
Considering a generating function

yt x1 yt 2 x2 yt i xi
(1  ) (1  ) ...(1  ) ,
1 y 1 y 1 y

Poisson noted that the probability sought was equal to the coefficient
of t in

1
(s + 1)  (1  y  yt ) x1 (1  y  yt 2 ) x2 ...(1  y  yt i ) xi dy.
0

A second sample of extracted balls {b1; b2; …; bi} had to be

introduced and the probability of the joint occurrence of condition (5)
and of restriction

b1 + 2b2 + … + ibi = z

had to be determined by means of a bivariate generating function, also

see § 8.5 below. Only later, when solving the electoral problem
(above), Poisson remarked that the result of the second sampling
might be (subjectively!) considered independent from the first one if
only it, the first one, remained unknown. Poisson’s solution was
indeed ingenious.
Suppose (Poisson 1837a, pp. 231 – 234) that an urn contains a
white balls and b black ones. Two samples without replacement are
made, one after the other, and g and m white balls and h and n black
ones are extracted respectively, g + h = r. The probability of the
second sample is

P(a; b; m; n) =  [P(a – g; b – h; m; n) P(a; b; g; h)]

where the sum is extended over g, h = 0, 1, 2, …; g + h = r and the

letters in parentheses are the appropriate arguments. The right side of
the formula does not depend on r which might be therefore assumed to
be zero. This remark indeed proves Poisson’s statement as well as the
finding of another author, Mondesir (1837). This episode and its

127
126
further history is described in Sheynin (2002b). Here, I only mention
Chuprov. In a letter of 1921 he (Sheynin 1990/2011, p. 145) stated:

Not knowing the prior data it is impossible to distinguish a series of

numbers obtained when extracting the tickets without replacement
from a series obtained according to the usual way of replacing […]
the ticket.

Also see Chuprov (1923, pp. 666 – 667; 1924, p. 490). The first to
consider sampling without replacement was Huygens (§ 2.2.2).
8.5. Limit Theorems for the Poisson Trials
Suppose that contrary events A and B occur in trial j with
probabilities pj and qj (pj + qj = 1). Poisson (p. 248) determined the
probability that in s trials event A occurred m times, and event B, n
times (m + n = s). He wrote out the generating function of the random
variable m (or, the bivariate generating function of m and n) as

X = (up1 + vq1) (up2 + vq2) … (ups + vqs)

so that the probability sought was the coefficient of umvn in the

development of X. His further calculations (lacking in Chapter 9 of
Laplace’s TAP) included transformations

u = eix, v = e–ix, upj + vqj = cosx + i(pj – qj)sinx = ρi exp(irj),

ρj = {cos2x + [(pj – qj)sinx]2}1/2, rj = arctan[(pj – qj)tanx].

Excluding the case of pj or qj decreasing with an increasing s, and

without estimating the effect of simplifications made, Poisson (pp.
252 – 253) derived the appropriate local and integral limit theorems
for large values of s. They were, however, complicated and their
importance apparently consisted in extending the class of studied
random variables.
8.6. The Central Limit Theorem
Poisson (p. 254) introduced a [lattice] random variable whose
values were multiples of some ω on a finite interval and depended on
the number of the trial. Applying the appropriate characteristic
function and the inversion formula, he determined the probability that
the sum of these values s was obeying certain inequalities aω ≤ s ≤
bω. He then went over to the sum of continuous variables by assuming
that ω → 0, a, b → ∞ with finite aω and bω and (p. 268) derived the
[CLT] for s under a single (not adequately explained) condition, again
without estimating the effect of simplifications made5. In accordance
with the context, it seems, however, that he supposed that the
variances of the terms of s were finite and bounded away from zero.
He (p. 258) also made use of the Dirichlet discontinuity factor which
he considered known. Dirichlet introduced it in two papers, both
published in 1839, see his Werke, Bd. 1, 1899, pp. 377 – 410.
Poisson (1824; 1829) earlier proved several versions of the CLT in
the same way. He (1824, §§ 4 and 6) introduced then the so-called
Cauchy distribution and found out that it was [stable]. For a modern
exposition see Hald (1998, pp. 317 – 327) who took into

128
127
consideration all the appropriate deliberations of Poisson. Those
proofs were methodically bad since the conditions of the theorems
were not provided whereas Hald remarked that that defect had then by
usual. Laplace (and Poisson) treated probability as a branch of applied
mathematics (§ 0.2 and elsewhere),but such omissions are still
unforgivable6. The proofs themselves were not really studied: Hald
somehow thought that they were rigorous.
Poisson (1824, §§ 8 – 10) also considered a linear function

E = a1ε1 + a2ε2 + … + anεn

of discrete and continuous independent random variables εi. In the

second instance he (1824, p. 288) obtained the appropriate CLT and
noted that that theorem did not hold for variables with density

φ(x) = e–2|x|, |x| < + ∞

and either ai = 1/(i + 1) or 1/(2i – 1). Markov (1899c, p. 42)

mentioned these exceptional cases in his debates with Nekrasov about
the CLT; in the translation of his note, I have inadvertently omitted his
exact reference to Poisson.
Poisson also applied the CLT for estimating the significance of
discrepancies between empirical indicators obtained from different
series of observations. For the Bernoulli trials he studied the
discrepancies between probabilities of events (p. 224) and between the
appropriate frequencies (p. 294), and, for his own pattern (§ 8.7),
between the mean values of a random variable (p. 288). Cournot
(1843, Chapters 7 and 8) borrowed his findings without mentioning
him.
8.7. The Law of Large Numbers
Here is how he defined this law in his Préambule (p. 7):

Things of every kind obey a universal law that we may call the law
of large numbers. Its essence is that if we observe a very large number
of events of the same nature, which depend on constant causes and on
causes that vary irregularly, sometimes in one manner sometimes in
another, i.e., not progressively in any determined sense, then almost
constant proportions will be found among these numbers.

He went on to state qualitatively that the deviations from his law

became ever smaller as the number of observations increased.
Bortkiewicz (1904, p. 826, Note 13) remarked that the Préambule was
largely contained in Poisson’s previous work (1835). Poisson (1837a,
pp. 8 – 11) illustrated his vague definition by various examples,
which, however, did not adequately explain the essence of the law but
were interesting indeed. Thus (pp. 9 and 10), the LLN explains the
stability of the mean sea level and the existence of a mean interval
between molecules. Beginning with 1829, Poisson’s contributions had
been containing many direct or indirect pronouncements on molecular
conditions of substance, local parameters of molecular interactions,

129
128
etc. sometimes connected with the LLN (Sheynin 1978b, p. 271, note
25) but they remained unnoticed.
Poisson then (pp. 138 – 142) formulated but did not prove three
propositions characterizing the LLN. These were based on the
standard formula (which Poisson had not written out)

P(B) = ΣP(Ai) P(B/Ai).

In actual fact, he studied the stability of statistical indicators by

means of the CLT, see Hald (1998, pp. 576 – 582).
Poisson described his law in a very complicated way and its
sufficiently detailed exposition is still lacking. No wonder that
Bortkiewicz (1894 – 1896, Bd. 8, p. 654) declared that

There hardly exists such a theorem that had met with so many
objections as the law of large numbers.

Here, in addition, is a passage from Bortkiewicz’ letter to Chuprov

of 1897 (Sheynin 1990c/2011, p. 60):

Or take […] my last three-hour talk with Markov about the law of
sm. [small] numbers [§ 15.1.2]. It caused me nothing but irritation. He
again demanded that I change the title. With respect to this topic we
got into conversation about the law of l. nn. It happens that Markov
(like Chebyshev) attributes this term to the case when all the
probabilities following one another in n trials are known beforehand.
[…] In concluding, Markov admitted that perhaps there did exist
‘some kind of ambiguity’ in Poisson’s reasoning, but he believed that
it was necessary to take into account the later authors’ understanding
of the term ‘law of l. nn.’ …

It is indeed difficult to examine Poisson’s considerations on that

point, but at least one of his examples (p. 148ff) is clear. It deals with
a throw of many coins of the same denomination and “mode de
fabrication”. And, although Poisson (p. 147) argued that the
probability of (say) heads could be established statistically, it seems
that his example had to do with unknown probabilities. Other
examples mentioned above (sea level and interval between molecules)
certainly included unknown probabilities.
The LLN was not recognized for a long time. In 1855 Bienaymé
declared that it contained nothing new (§ 10.2) which apparently
compelled Cournot (1843) to pass it over in silence; his views were
certainly known even in 1842 (Heyde & Seneta 1977, pp. 46 – 47).
Even much later Bertrand (1888a, pp. XXXII and 94) considered it
unimportant and lacking in rigour and precision. However, already
Bessel (1838a, especially § 9) guardedly called the Poisson law a
“principle” of large numbers, Buniakovsky (1846, p. 35) mentioned it
and Davidov (1854; 1857, p. 11) thought it important. It is
nevertheless possible (§ 3.2.3) that statisticians had recognized the
Bernoulli, and the Poisson (and the Chebyshev) laws of large
numbers only in the qualitative sense.

130
129
 8.8. The Theory of Errors and Artillery
In the theory of errors Poisson offered his proof of the CLT (§ 8.6)
and a distribution-free test for the evenness of the density of
observational errors (1829, § 10). He (1837b) also applied the theory
of probability and the error theory to artillery firing, although mostly
in a methodical sense7. He recommended the variance as the main
estimator of scattering which conformed to Laplace’s later idea, see
§ 7.2-6. In one of his problems Poisson (1837b, § 7) determined the
distribution of the square of the distance of some point from the origin
given the normal distributions of the point’s distances from the two
coordinate axes. He thus was perhaps the first to treat clearly the
densities as purely mathematical objects.
Poisson had not referred to the Gaussian theory of errors, had
therefore essentially lessened the importance of his contributions to
that discipline. Here, for example, is his opinion (1833, p. 36) about
the merits of Legendre:

It is to him that the sciences of observation owe the rule of

calculation which [?] is named the MLSq. Laplace showed its entire
probable benefit with respect to the precision of the results.

Gauss does not exist! And Poisson’s (1837a) layman’s, as I insist,

considerations about measurements and observations are almost
useless. Unlike Poisson and other French mathematicians, see end of
§ 7.2-3, Laplace (1812/1886, p. 353) objectively described the
discovery of the MLSq:

Legendre conceived a simple idea to consider the sum of the

squares of observational errors and to render it minimal which
immediately leads to as many final equations as there are elements to
be corrected. That learned geometer was the first to publish the
indicated method, but we ought to acknowledge that Gauss had been
invariably applying the same idea for many years before Legendre’s
publication and communicated it to many astronomers.
8.9. Statistics
In § 6.2 I described the development of statistics in the 18th century
and I return to this subject in Chapter 10. Here, I discuss the
appropriate pronouncements of Poisson and some other scholars.
Recall first of all (§ 8.6) that Poisson investigated the significance of
empirical discrepancies. Quetelet (1869, t. 1, p. 103), who
corresponded with Poisson, testified that the latter had mentioned
statisticians, who

Pretended to substitute their fantasies for the veritable principles of

[their] science, with derisive severity.

In a few other cases (and twice in joint papers) Poisson expressed

himself more definitely. Thus (Libri-Carrucci et al 1834, p. 535):

The most sublime problems of the arithmétique sociale can only be

resolved with the help of the theory of probability.

131
130
A year later Double et al (1835, p. 174) stated that

Statistics carried into effect always is, after all, the functioning
mechanism of the calculus of probability, necessarily concerning
infinite [?] masses, an unrestricted number of facts; and (p. 176) [with
respect to the applicability of mathematics] the state of the medical
sciences is not worse than, not different from the situation with all the
physical and natural sciences, jurisprudence, moral and political
sciences etc.

This opinion was, however, questioned. Poinsot (Poisson 1836, p.

380) declared that the application of the calculus of probability to
“moral things”, such as the verdicts of law courts and elections, was a
“dangerous illusion”, also see § 8.9.18. Double (1837, pp. 362 – 363)
sharply objected to the application of statistics in medicine and stated
that “each case appear[ed] to me [to him] a new and a separate
problem”. However, he mistakenly identified statistics with the
numerical method (see § 10.8). Cauchy (1821/1897, p. V) cautiously
pronounced a similar opinion: The only method of natural sciences
consisted in subjecting observations to calculus, but the mathematical
sciences should not “exceed their bounds”. Later he (1845/1896, p.
242), however, expressed himself quite differently: statistics, as he
maintained, provided the means, infallible in a sense, for judging
doctrines and institutions, and should be applied “with full rigour”.
8.9.1. Jurisprudence. Poisson (1837a, pp. 1 – 2) thought that the
study of the probabilities of verdicts and, in general, of majority
decisions, was a most important application of the calculus of
probability. He (p. 17) perceived his main goals in that field as an
examination of the stability of the rate of conviction, a derivation of
the number of jurors and of the majority necessary for conviction so as
to lower the probability of miscarriage of justice, a comparison of
judicial statistics of different countries and (p. 7) the proof of the
applicability of mathematical analysis to “matters that are called
moral”. On moral sciences see Note 8.
Poisson was mainly interested in studying criminal offences. Unlike
Laplace, he (pp. 4 and 318) introduced a positive probability (k) of
the defendant’s guilt. One of his formulas (p. 333) determined the
probability that the defendant, convicted by (n – i) jurors out of n, was
really guilty:

Pi = ktm/[ktm + (1 – k)], t = u/(1 – u).

Here, m = n – 2i, and u was the probability of a correct verdict

reached by each juror (judge). Poisson noted that the right side did not
depend on n; however, supposing that n was odd (say), i could have
varied up to its greatest value, (n – 1)/2, – and what mattered was the
sum of all of the values of Pi. Poisson derived a similar formula for a
continuous random u by introducing its unknown prior density.
One of Poisson’s statements (pp. 375 – 376) is debatable: he
thought that the rate of conviction should increase with crime.

132
131
 The application of the theory of probability to jurisprudence
continued to be denied. Here are the two most vivid pertinent
statements (Mill 1843/1886, p. 353; Poincaré 1896/1912, p. 20):

1) Misapplications of the calculus of probability […] made it the

real opprobrium of mathematics. It is sufficient to refer to the
applications made of it to the credibility of witnesses, and to the
correctness of the verdicts of juries.

2) People “influence each other” and act like the “moutons de

Panurge”.

Nevertheless, the pertinent work of Laplace and Poisson (and of

their predecessor, Condorcet, § 6.1.5) had undoubtedly attracted the
public to the problems of the administration of justice and showed
what could be hoped for in the ideal case. I return to Poincaré in
§ 11.2.
8.9.2. Medical Statistics. I mentioned this discipline in § 6.2.3.
Now I say that Poisson had certainly contributed to its development.
Here is a statement of Gavarret (1840, p. xiii), his former student who
later took to medicine:

Only after long reflection on the lectures and writings of the

illustrious geometer, we grasped all the extensiveness of the
systematic application of the experimental method in the art of
healing.

In his book, that became very well known, Gavarret explained the
normal approximation to the binomial law and the calculation of
admissible discrepancies of frequencies in series of Poisson trials and
(p. 194) stressed the importance of checking null hypotheses. True,
their introduction was a logical consequence of Poisson’s
deliberations.
In Russia, in the 1850s, Davidov (Ondar 1971), who was well
acquainted with the work of Poisson and Cournot (§ 10.3),
popularized the application of the statistical method to medicine, see
Ondar (1971). I mention him again in §§ 8.1 and 10.4-8.
Poisson (1837a, p. VI) and Gavarret after him insisted on the need
to collect many observations. Here is Poisson:

Medicine will not become either a science or an art if not based on

numerous observations, on the tact and proper experience of the
physicians who judge the similarity of cases and take into account
exceptional circumstances.

And so, medicine was then not a science or an art!

Nevertheless, it was difficult to collect many observations. Here is
the opinion of a noted physician (Liebermeister ca. 1876, pp. 935 –
940):

133
132
 Theoreticians rather often categorically tell us, practical
physicians, that all our inferences about the advantages or
shortcomings of some methods of treatment, so far as they are based
on results which have really taken place, simply remain up in the air if
only we do not apply rigorous rules of the theory of probability. […]
Until now physicians have applied that theory so seldom not so much
because they sometimes did not attach proper significance to it, but
mainly since its analytical arsenal was too imperfect and awkward.
[…] Mathematicians say: If you, physicians, wish to arrive at
plausible conclusions, you must invariably work with large numbers;
you ought to collect thousands and hundred thousands observations.
[…] This, however, is impossible for statistics of a general
practitioner. And, nevertheless, if this condition is fulfilled, it will
often be doubtful whether the theory of probability will be necessary
in the same pressing manner. […] Gavarret somewhat arbitrarily
presumed, as Poisson also did in several problems, that 0.9953 or
212:213 […] is a sufficient measure of probability. […] Suppose that
the successes of two methods of treatment are only as 10:1, would not
that be sufficient for preferring the first one?

He exaggerated, but his criticism was still valid. Then, beginning

with 1863 (Chauvenet) and even earlier astronomers and geodesists
had begun to offer tests for rejecting outliers quite in vein with
Liebermeister’s reasoning. Previous practitioners had also made
plausible inferences on the strength of scarce data (Bull 1959) whereas
Niklaus Bernoulli (§ 3.3.2) thought that an absentee ought to be
declared dead once his death becomes only twice as probable as the
alternative.
Modern statistics including the theory of errors cannot restrict it
activity to the case of a large number of observations. Therefore,
Liebermeister rather than Gavarret was the pioneer of medical
statistics. Freudenthal & Steiner (1966, pp. 181 – 182) mistakenly
attached to Gavarret, not to Liebermeister, the transition from
unconditional certitude to reasonable degrees of probability. In 1889,
Liebermeister’s collected medical works had appeared in Leipzig after
which he still published a few books. It is worthwhile to mention that
Liebermeister in a medical context, studied the possibility of
distinguishing between equality and inequality of success probabilities
in two (small) series of binomial trials. Starting from a Laplacean
formula based on the existence of uniform prior distribution, and
assuming that the two probabilities coincided, he considered the size
of the tail probability (of the hypergeometric distribution). His main
formula had hardly ever reappeared. See Seneta (1994).
I treat the further history of medical statistics in § 10.8.1, but it is
opportune to say now that in the 19th century that discipline had still
been rather primitive and that the demand advocated by Gavarret (and
Poisson) was indeed unrealistic.

Notes
1. When describing his attempt to generalize the normal law, De Morgan (1864,
p. 421) declared that if the probability of a certain event was 2.5, it meant that the
event “must happen twice with an even chance of happening a third time”. In 1842,

134
133
in a letter to John Herschel (Sophia De Morgan 1882, p. 147), he stated that
“undoubtedly”

sin∞ = 0, cos∞ = 0, tan∞ =  1, cotan∞ = tan∞.

I did not find Herschel’s answer.

2. Earlier Poisson (1830, pp. 141 and 146) used the same letter A for designating
an observed constant, – of “some thing”. Consequently, it hardly stood later for
“aléatoire”.
3. Poisson (1837a, p. 274, and earlier (1833, p. 637) corroborated the transition
from discrete to continuous by a trick that can be described by Dirac’s delta-
function. When considering density φ(x) equal to zero everywhere excepting a finite
number of points ci, i = 1, 2, …, n, and such that

ci  


ci  
φ(x) dx = gi, Σgi = 1, ε → 0,

Poisson had thus introduced that function of the type of

φ(x) = Σgi δ(x – ci).

4. Poisson was unable to keep to his announced distinction between chance and
probability and I am therefore making use of the modern term.
5. Poisson referred to his p. 155 and to his memoir (1829, § 8), but, as I see it, the
situation remained unclear. Later he (1837a, pp. 312 – 313) repeated the formula of
the CLT for the mean value of a random variable without introducing any conditions
and even without demanding that its domain was restricted to a finite interval.
6. From Laplace to the 1930s the theory of probability had been developing as a
branch of applied mathematics (§ 0.1), see also Note 5. And here is Poisson (1837a,
§ 84):

There exists a very high probability that these unknown chances little differ from
the ratio …

7. From 1812 (and until?) Poisson was “examinateur de l’arme de l’artillerie”

(Arago 1850, p. 602).
8. Laplace (1814/1995, p. 61) urged to

Apply to the political and moral sciences the method based on observation and
the calculus, a method that has served us so successfully in the natural sciences.

It is difficult to say what exactly is included into moral sciences; see however
Poinsot’s statement below. Beginning at least with Quetelet, the study of
phenomena depending on free will (although only crimes, suicides, and marriages)
was considered to constitute the subject of moral statistics. Then, however, the
domain of that branch of statistics essentially broadened and includes now, for
example, philanthropy and professional and geographical mobility of the population.

135
134
 9. Gauss, Helmert, Bessel
9A. Gauss

This subchapter is mostly devoted to the MLSq1.

I (1979) have somewhat dwelt on Gauss’ investigations in
probability proper. He was a tireless collector of statistical data, even
of non-essential nature (Biermann 1991), and successfully managed
the widows’ fund of the Göttingen University. His correspondence
and scientific legacy includes a study of the mortality of newly-born
and of the members of tontines (of closed societies of mutually
insured persons, see § 2.1.3). In the theory of probability, he left the
inversion formula for the Fourier transform of the density function.
Gauss also solved the first problem in the metric theory of numbers.
He considered the expansion of a number M (0 < M < 1) into a
continued fraction with unit numerators and investigated the
probability P(n; x) that, beginning with its (n + 1)-st convergent, the
“tail” of this fraction was less than x. If all the permissible values of M
were equally probable or more or less so, then, as he (W-10/1, pp. 371
– 372) explained his problem in a letter of 1812 to Laplace, P(0; x) =
x and

ln (1  x )
lim P(n; x) = , n → ∞.
ln 2

Nevertheless, he was not quite satisfied with this solution and asked
Laplace to have a look at the problem. He, Gauss, was sure that
Laplace will find a plus complete solution, – a pre-limiting expression.
A phrase from Gauss’ Mathematisches Tagebuch written in 1800 (p.
552 of the Werke) testifies that Gauss had already derived the equality
above – and had then been satisfied with his work.
Stäckel (Gauss, W-10/1, pp. 554 – 556) and then Kuzmin (1928)
proved this equality and the latter also derived an asymptotic
expansion for P(n; x).
Here, I also repeat in a few words Gauss’ general opinion (W-12,
pp. 201 – 204) about the applications of the theory of probability as
described by W. E. Weber in one of his letters of 1841. If only based
on numbers, Gauss reasoned, such applications could be greatly
mistaken; the nature of the studied subject ought also to be taken into
account. However, probability provides clues when nothing except
numbers is known, as for example when dealing with annuities; and in
jurisprudence and it can determine the desired number of witnesses
and jurors [but hardly without allowing for “the nature” of law courts].
See also Gauss’ opinion about the rejection of outlying observations in
§ 9A.5-2. As shown in the sequel, quite a few authors shared Gauss’
viewpoint.
9A.1. The Method of Least Squares before 1809

136
135
 It had been indirectly and inaccurately applied from the mid-18th
century (§ 6.3.2) and its peculiar version was possibly known even
earlier. When some point P was graphically intersected from three or
more given stations, a triangle, or a polygon of errors appeared on the
surveyor’s table sheet and it was apparently natural to select the
position of P by eye in such a manner that the sum of the squares of its
distances from the sides of the triangle (of the polygon) was minimal.
To a certain extent I can justify my opinion by mentioning an
experimental smoothing of a broken line by eye (Tutubalin 1973, p.
27): on the whole, the curves thus drawn were as accurate as if having
been determined by the MLSq.
Euler (1778) can be considered as Gauss’ predecessor in the
heuristic sense (§ 6.3.1), but Gauss possibly did not see that memoir (a
commentary on Daniel Bernoulli’s memoir of the same year). The
regrettably incomplete list of books which Gauss had borrowed from
the library of the Göttingen university (Dunnington 1955, pp. 398 –
404) does not include the volume of the contributions of the
Petersburg Academy in which those two papers were published.
However, in his letters Gauss expressed his surprise at the fact that the
principle of least squares was not discovered earlier.
9A.1.1. Huber. Many authors, for example Merian (1830, p. 148),
stated that somewhat before 1802 the Swiss mathematician and
astronomer Huber had discovered the principle of least squares, but
that, living far from scientific centres, he had not reported his finding
to anyone. However, Dutka (1990), who referred to a forgotten paper
(W. Spieß 1939), concluded otherwise. It occurs that Spieß quoted
Huber himself who had mentioned “Legendre’s criterion [Maßstab]
of least squares”.
9A.1.2. Legendre (1805, pp. 72 – 73) introduced the principle of
least squares:

Among all the principles that can be proposed [for solving

redundant systems of linear equations], I think there is no one more
general, more exact, and easier to apply, than that which we have
made use of in the preceding researches and which consists in making
the sum of the squares of the errors [of the residuals] a minimum. In
this way there is established a sort of equilibrium among the errors,
which prevents the extremes to prevail and is well suited to make us
know the state of the system most near to the truth.

Translation by Hald (1998, p. 119). Legendre also indicated that the

absolute values of the extremes [again: of the residuals] should be
confined within the shortest possible interval. He had not added that it
was the minimax principle (§ 6.3.2) rather than his innovation that
ensured his desire.
9A.1.3. Adrain. The American mathematician Adrain (1809)
justified the principle of least squares and the [normal distribution] 2 at
about the same time as Gauss did and applied it to the solution of
several problems, see below (Dutka 1990). He also indicated that the
lack of space prevented him to discuss the adjustment of pendulum
observations. About ten years later he (1818a) published that study in

137
136
which he revealed two mistakes in Laplace’s pertinent calculations
(1798 – 1825, t. 2, § 42 of Livre 3). The same year his derivation of
the length of the major semi-axis of the Earth’s ellipsoid of rotation
(1818b) appeared. Incidentally, that length (6378.629 km) was
sufficiently close to a determination of 1940 by F. N. Krasovsky
(6378.245 km).
Adrain’s main paper was first mentioned much later (C. Abbe
1871) but his second article had become known to Olbers (W/Erg-4,
No. 2, p. 711) who informed Gauss about it. An American author,
wrote Olbers to Gauss, had mentioned his previous paper and
“ascribed” the MLSq to himself. Gauss hardly made any comment;
the priority strife with Legendre was apparently enough for him.
Here are Adrain’s derivations of the normal distribution.
a) Lines a and b are measured in the field with errors x and y
respectively and

x/a = y/b (1)

and the total error is fixed:

x + y = c. (2)

Introducing the density of the observational errors φ and tacitly

assuming their independence, Adrain applied the principle of
[maximum likelihood]

φ(x; a) φ(y; b) = max

so that, after allowing for conditions (1) and (2),

[φ′(x; a)/φ(x; a)]dx + [φ′(y; b)/φ(y; b)]dy = 0, φ′(x; a)/φ(x; a) = mxa, etc.

b) Suppose that for linear measurements

x2 + y2 = r2,

then

W = φ(x) φ(y) – λ(x2 + y2) = max, φ′(x)φ(y) – 2λx = 0,

φ(x)φ′(y) – 2λy = 0,
φ′(x)/xφ(x) = φ′(y)/yφ(y) = c, etc.

Adrain then wrote out the joint distribution of both these errors and
indicated that the appropriate contour lines of equal probabilities were
ellipses (ellipses of errors, as they were later called in the theory of
errors).
Conditions (1) and (2) hardly conform to reality; thus, the former
describes the action of systematic errors. Also arbitrary is the
condition applied in the second justification. Nevertheless, John
Herschel (1850), Maxwell (1860), Thomson & Tait (1867, p. 314)
and Krylov (1950, Chapt. 8) repeated that demonstration without any

138
137
references (Sheynin 1965). Later on Kac (1939) and Linnik (1952)
weakened the condition of independence.
Adrain was now able to prove quite simply that the arithmetic mean
of direct measurements was optimal; this, of course, conformed to the
principle of least squares in case of several unknowns. Finally, Adrain
showed how to adjust a traverse (a polygon with measured sides and
bearings) by the principle of least squares and, what is also
remarkable, he calculated corrections to directly measured magnitudes
rather than to their functions which were not independent from each
other, see Sheynin (1965), a paper which in general I am now
dissatisfied.
9A.1.4. Gauss. He (1809a; 1809b, § 186) applied the principle of
least squares from 1794 or 1795. In the second instance, he called it
“our principle”: Unser Princip, dessen wir uns seit dem Jahre 1795
bedient haben, and in both cases he mentioned Legendre. Much later
Gauss (1823a) once more mentioned Legendre, although not in his
main memoir of 1823.
Legendre (letter to Gauss 31.5.1809, see Gauss, W-9, p. 380) was
badly offended. He rightfully stated that priority is only established by
publication. Gauss did not answer him, and Legendre (1820, pp. 79 –
80) charged him with appropriating the MLSq.
Many authors commented on this episode, and I quote May (1972,
p. 309) and Biermann (1966, p. 18):

Gauss cared a great deal for priority. […] But to him this meant
being first to discover, not first to publish; and he was satisfied to
establish his dates by private records, correspondence, cryptic
remarks in publications. […] Whether he intended it so or not, in this
way he maintained the advantage of secrecy without losing his
priority in the eyes of later generations.

What is forbidden for usual authors, ought to be allowed for

Gausses and in any case we must respect his [Gauss’] initial
considerations.

It seems that Legendre could have stated in 1820 that he was the
inventor of the MLSq and that, in spite of Gauss’ claim, everyone will
agree with him. Regrettably, this did not happen. Laplace (end of
§ 8.8) objectively described the discovery of the MLSq but did not
add that Legendre had not really substantiated it. Then, he offered his
own version of the theory of errors only suitable for a large number of
observations and for other conditions necessary for the CLT, and other
French mathematicians including Poisson followed him and Legendre
and did not even mention Gauss. This seriously diminished the
importance of their work.
In his letter to Gauss, Legendre also noticed that Euler rather than
Laplace, as Gauss (1809, § 177) stated, calculated the integral of the
exponential function of the negative square. Later Gauss (Monatliche
Correspondenz, Bd. 21, p. 280) explained that he noticed his mistake
when the printing of his book was almost completed and that it was
Laplace who provided the final formulation of the result. The Editors

139
138
of Gauss (1887), who noticed that explanation did not supply the year
of that Bd. 21.
On 30 Jan. 1912 Gauss (W-10/1, p. 373) only answered Laplace: he
had applied the MLSq long before 1805 but had no desire to publish a
fragment. Gauss’ preliminary reports had appeared in the Göttingische
gelehrte Anzeigen and Legendre hardly saw the report (1809a). Later
it was reprinted (W-6, pp. 59 – 60), and here is an excerpt:

The author has been applying the main principles which are here
considered for fourteen years now and long ago had communicated to
his astronomical friends. They lead to the same method which
Legendre […] published a few years ago.

I (Sheynin 1999a; 1999c) described the possible cases in which

Gauss could have applied the MLSq before 1805 and named many of
his colleagues and friends to whom he had communicated his
discovery. Unexpectedly, it occurred that von Zach, who allegedly
refused to testify to Gauss’ priority, had not until 1805 known the
formulation of the principle of least squares, and, furthermore, that he
(1813, p. 98n) indirectly agreed with the latter’s statements by
repeating them without any qualification remarks:

The celebrated Dr Gauss was in possession of that method since

1795 and he advantageously applied it when determining the elements
of the elliptical orbits of the four new [minor] planets as it can be seen
in his excellent work [Theoria motus].

Regrettably, “it” is not seen there.

This passage is even more important than Zach’s editorial
acceptance of Gauss’ priority (noticed by Dutka 1996, p. 357). In
1809, Zach’s periodical, Monatliche Correspondenz, carried an
anonymous review of Gauss’ Theoria motus, and there, on p. 191,
Gauss’ pertinent claim was repeated.
Gauss’ early use of the MLSq is not generally accepted, see for
example Marsden (1995, p. 185) who nevertheless had not mentioned
the opposite opinion of Brendel (1924) and Galle (1924, p. 9) or of
Gauss’ contemporaries3. In any case, Gerardy (1977), drawing on
archival sources, discovered that Gauss, in 1802 – 1807, had
participated in land surveying (in part, for his own satisfaction) and
concluded, on p. 19 (note 16) that Gauss started using the method not
later than in 1803. Regrettably, Gerardy concentrated on describing
Gauss’ simple calculations and his statement mentioned just above
was not quite definite.
There are many other instances including that mentioned by von
Zach (above) in which Gauss could have well applied his invention at
least for preliminary, trial calculations, or short cuts. For him, the
MLSq was not a cut and dry procedure, see § 9A.5-3. Then, weighing
the observations as well as the presence of possible mistakes in the
data could have hampered justification.
As to the communication of his discovery, I proved that among
those whom Gauss had informed before 1805 were Bessel and

140
139
Wolfgang Bolyai (the father of the cofounder of the non-Euclidean
geometry, Janos or Johann Bolyai), and Olbers about whom it was
known long ago. In 1812, Olbers promised Gauss to state publicly that
he came to know about the MLSq from Gauss a few years before
Legendre’s publication, but he only did so in 1816. In 1812 − 1815,
Olbers just did not publish anything suitable for inserting such a
remark, see the appropriate volume of the Catalogue of Scientific
Literature of the Royal Society.
9A.2. Theoria Motus (1809b)
In accordance with the publisher’s demand, this book appeared in
Latin. Its German original is lost and Gauss’ correspondence (letter
from Olbers of 27.6.1809, see W/Erg-4, No. 1, p. 436) proves that,
while translating, he essentially changed its text. The treatment of
observations occupies only a small part of the book.
1) The Boscovich method (see § 6.3.2-3). Suppose that n equations
(1.2) in m unknowns (n > m) are adjusted by that method. Then, as
Gauss (§ 186) remarked, equation (6.13) meant that exactly m residual
free terms will be zero. Somewhat below, in the same § 186, Gauss
qualified his statement by taking into account the other Boscovich
equation (6.12) but mistakenly attributed it to Laplace. In § 174 he
stated that the formulated corollary was undesirable although in §§
188 – 189 he apparently agreed that the Boscovich method (§ 6.3.2
might ensure a first approximation. His remark, that can be easily
proved, means that he knew an important theorem in linear
programming.
2) The [normal distribution] (§§ 175 – 177). Gauss (§ 177) assumed
“as an axiom” that the arithmetic mean of many observations was the
most probable value of the measured constant “if not absolutely
precisely, then very close to it”. He (§ 175) derived the density φ of
observational errors (although not introducing any new term)
believing that it was [unimodal] and “in most cases” even; this, then,
was his understanding of the properties of random errors. Finally, in
order to justify the principle of [maximal likelihood], Gauss (§ 176)
proved the “fundamental principle” of inverse probability, see my
§ 7.1-1, for the case of equal probabilities of the various hypotheses.
However, the principle of the arithmetic mean (above) already implied
this restriction (Whittaker & Robinson 1924/1949, p. 219).
And so, if the observations are denoted by xi, i = 1, 2, …, n, then,
according to the principle of maximal likelihood,

[φ′(x1 – a)/φ(x1 – a)] + [φ′(x2 – a)/φ(x2 – a )] + ... +

[φ′(xn – a)/φ(xn – a )] = 0,

where a is the estimator sought, coinciding, as stipulated, with the

arithmetic mean xo. If

xi = x1 – nN, i = 2, 3, …, n,

then

x1 + (x2 + x3 + … + xn) = x1 + (n – 1)x1 – n(n – 1)N,

141
140
 N = (x1 – xo)/(n – 1),
φ′(x1 – xo)/φ(x1 – xo) = (1 – n)φ′(– N)/φ(– N) = – (1 – n)φ′(N)/φ(N),
φ′[N(n – 1)]/{(1 – n)φ[N(n – 1)]} = – φ′(N)/φ(N),
φ′(x)/xφ(x) = Const,

φ(x) = (h/√π)exp(– h2x2), h > 0. (3)

Gauss (§ 178) called h the “measure of precision” (gradus

praecisionis). It might be supposed that, from the very beginning, he
was not satisfied with his derivation. His wording of the principle of
the arithmetic mean and of the properties of the density of
observational errors contained qualification remarks whereas the
obtained principle of least squares (see below § 9A.2-3) occurred to be
an axiom. Again, it is difficult to believe that Gauss was pleased with
the appearance of a universal law of error. Later he (1821/1887, pp.
193 and 194; 1823a/1887, p. 196) remarked that his derivation had
depended on a hypothetically assumed distribution. And here is
Bertrand’s opinion (1888a, p. XXXIV): Gauss had not claimed to
establish the “vérité”, he attempted to search for it. Bertrand (pp. 180
– 181) also remarked that the mean of the values of some function did
not coincide with the mean value of its arguments, which, in his
opinion, testified against the principle of arithmetic mean. Gauss,
however, considered direct measurements. Note also that he (his letter
to Encke of 1831; W-8, pp. 145 – 146) “not without interest”
acquainted himself with the attempt of his correspondent to justify the
arithmetic mean by deterministic analytical axioms. Many authors
made similar efforts and Zoch (1935) concluded that, although they
were unsuccessful, the postulate of the arithmetic mean can
nevertheless be established without stochastic considerations. His
finding was unrelated to the theory of errors, but the pertinent
investigations apparently served as the point of departure for the
theory of invariant statistical hypotheses and estimators (Lehmann
1959, Chapter 6). Encke (1832, p. 74), as it occurred, was not satisfied
with either substantiations of the MLSq.
Gauss (1845/1873, p. 143) left a lesser known statement about the
arithmetic mean. He remarked that the random variations corrupting
observations mostly compensate one another so that the mean
becomes ever more reliable as the number of observations increases.
This is “generally absolutely right”, and often led to “splendid results”
in natural sciences. However, Gauss continued, an important
condition, often overlooked and difficult to check, was that the
disordered variations ought to be entirely independent from each
other, cf. § 9A.4-4.
3) The principle of least squares (§ 179) followed immediately.
Gauss, however, added that, similar to the principle of the arithmetic
mean, it should be considered an axiom [considered as a corollary of
an axiom?]. A special point here is that, instead of the real errors the
principle of least squares was formulated with regard to residual free
terms. Helmert (1872, p. 75) indicated this fact but paid scant
attention to it and had not mentioned Gauss. Apparently he had not
realized that the normal law was [stable], cf. §§ 7.2.7.

142
141
 4) The precision of the arithmetic mean. Gauss, naturally, restricted
his attention to the case of the [normal distribution]. Later he (§ 9A.4)
certainly abandoned this restriction.
5) The precision of a random sum (marginal note to § 183, included
not in the German translation, but in Gauss’ W-7). Suppose that

x = a + b + c + ...,

then

hx = 1 ÷ [(1/ha2) + (1/hb2) + (1/hc2) + ... ]1/2.

Gauss did not explain his note; it might be supposed that the terms
above were normally distributed since he only introduced h for that
law. However, he may well have derived this formula in the general
case. Incidentally, it shows that at least in the case of the normal
distribution the random error of the sum increases as the square root of
the number of its terms.
6) The precision of the [estimators of the] unknowns (§ 182; 1811,
§ 13). Suppose that these estimators are determined by solving a
system of normal equations in accordance with the Gauss method of
successive eliminations. Then, assuming that the precision of a direct
measurement is unity, the precision of the estimator of the last
unknown is equal to the root of its coefficient in the last reduced
equation. Also see my § 9A.4-5.
9A.3. Determining the Precision of Observations (1816)
1) The precision of the measure of precision h in formula (3).
Suppose that the errors of m [independent] observations are α, β, γ, …
Then the most probable value of that magnitude is determined by the
condition

hmexp[– h2(α2 + β2 + γ2 + …)] = max

and is therefore equal to

ho = {m/[2(α2 + β2 + γ2 + …]}1/2 = 1/σ√2.

In the last expression, which is my own, σ is the mean square error

of an observation. Gauss also indicated that

P(ho – λ ≤ h ≤ ho + λ) = θ(λ√m/ho), θ(t) = (2/√π)  exp (– z2)dz,

0
so that, for P = 1/2, λ = ρho/√m, ρ ≈ 0.477. In addition, for distribution
(3),

P(|х| ≤ ρ√h) = 1/2, and r = ρ/h

is the probable error formally introduced by Bessel (1816, pp. 141 –

142).
Let

143
142
 

Sn = |α|n + |β|n + |γ|n + ..., Kn = 


xnφ(x)dx,

then, for large values of m,

P (– λ ≤ Sn – mKn ≤ λ) = θ{λ/[2m(K2n – Kn2)]1/2}, (4)

where mKn is the most probable [the mean] value of Sn. In actual fact,
Gauss treated absolute moments and the formula for Kn should be
corrected. Formula (4) was proved by Helmert (1876b) and then by
Lipschitz (1890), but Cramér (1946, § 28.2) noted that it was a
particular case of the CLT.
Finally, Gauss derived a formula for the absolute moments of the
normal law

mKn = Sn0 = mП[(n – 1)/2]/hn√π, П(x) = Г(x + 1),

so that h (and therefore r) could have been estimated by S n , the mean

value of Sn. Comparing the probable intervals of r for different n,
Gauss concluded that n = 2 secured its best estimator.
In one of his letters of 1825 Gauss (W-8, p. 143) objected to the
probable error as “depending on a hypothesis” [on the law of
distribution]. Still, again in his correspondence (Sheynin 1994a, p.
261), and even in a paper (1828b), he applied it quite a few times.
Natural scientists, for example Mendeleev (§ 10.9.3) and Newcomb
(§ 10.9.4), followed suit and Bomford (1971, pp. 610 – 611)
“reluctantly” changed from probable to mean square error in that
edition of his book. However, L. O. Struve (1887, last, unnumbered,
page) proposed to abandon the probable error.
2) Denote 1/h√2 = α and let n = 2. Then

[m(K4 – K22)]1/2 = α2 2m

and, in accordance with formula (4), the sum of squares S2 is

distributed normally N[mα2; α2 2m ]. This is the asymptotic chi-
squared distribution, cf. Cramér (1946, § 20.2).
9A.4. The Theory of Combinations (1823 – 1828)
I consider the main part of this memoir in which Gauss provided his
definitive justification of the MLSq by the principle of maximum
weight [of minimal variance], and I add a few words about its
supplement (1828a).
1) Random errors and the density of observational errors. Gauss (§§
1 – 3) distinguished between random and systematic errors but had not
provided their formal definition. He (§ 4) then repeated (see my
§ 9A.2-2) the definition of density and listed its properties. The mean
value of the errors (§ 5) was equal to zero; otherwise, as Gauss
additionally remarked, it determined the action of constant errors.
2) The measure of precision. Gauss (§ 6) introduced a measure of
precision [the variance]

144
143
 

m2 = 

x2φ(x)dx

calling it the mean error to be feared, – des mittleren zu befürchtenden

Fehler, errorum medium metuendum (1821/1887, p. 194; 1823b, § 7).
In his letters to Encke of 23 Aug. 1831 (W-8, pp. 145 – 146), to
Bessel of 28.2.1839 (Ibidem, pp. 146 – 147), to Schumacher of 25
Nov. 1844 (reported by Helmert in his Introduction to Gauss 1887)
and in § 7 of the Theor. motus Gauss stressed that an integral measure
of precision was preferable to a local measure. He (1823b, § 6) also
indicated that the quadratic function was the simplest [from integral
measures], and in 1821 he (1887, p. 192) dwelt on his choice in more
detail: it was also connected with

Some other, extremely important advantages which no other

function possesses. However, any other even degree could have been
selected as well …

Could have been chosen in spite of the advantages of the variance?

Bienaymé (1853/1867, pp. 167 – 169) proved that a formula of the
type of (5), see below, was not valid for any other even exponent; a
clear exposition of this proof is due to Idelson (1947, pp. 269 – 271).
Therefore, Bienaymé continued, the choice of the variance was
unavoidable. I doubt, however, that, as he believed (p. 169), Gauss
was here mistaken. The sample variance (see § 9A.4-6) is distribution-
free.
3) An inequality of the Bienaymé – Chebyshev type. Gauss (§ 9)
examined the probability

m

µ = P(|ξ| ≤ λm) = 
 m
φ(x)dx

for a [unimodal] density of observational errors ξ having variance m2

and proved (§ 10) that

2
λ ≤ µ√3 for µ ≤ 2/3 and λ ≤ for ≤ µ ≤ 1.
3 1 μ

Cramér (1946, § 15.7 and Example 4 to Chapters 15 – 20) more

easily proved this “remarkable” theorem, as Gauss called it, whereas
Seal (1967/1970, p. 210) indicated, that Gauss had wished to abandon
the universality of the normal distribution since it occurred that,
anyway, P(|ξ| ≤ 2m) ≥ 0.89. But should we forget his own, although
indirect arguments and doubts?
4) Independence. In § 18 Gauss offered his definition, although not
quite formal, of independent functions of observations: they should
not contain common observations. In § 19 he specified that those
functions were linear; otherwise his statement would have

145
144
contradicted the Student – Fisher theorem on the independence of the
sample variance and the arithmetic mean.
Therefore, if some observation was common for two functions of
observational results, the errors of these functions will not be
independent from one another and the mean value of their product will
not therefore vanish4. In one of his examples, Gauss calculated the
variance of a linear form of independent random variables.
Gauss (1809b, § 175; 1823b, § 15) mentioned independence even
earlier but without explanation, and, later he (1826/1887, p. 200;
1828, § 3) described the mutual dependence of magnitudes known
from observation by the existence of functional connections between
them. This meant, for example, that the adjusted angles of a triangle,
since their sum was equal to 180° plus the spheroidal excess, were
dependent on one another. See also end of § 9A.2-2.
In mathematical statistics the definition of independence is
different. An orthogonal transformation of independent and normally
distributed magnitudes leads to their as though “adjusted” values, – to
their linear forms of a certain type, which are nevertheless
independent (the Fisher lemma; Cramér (1946, § 29.2)). Here is K.
Pearson’s appropriate statement (1920/1970, p. 187) which I do not
however understand: for Gauss

The observed variables are independent, for us [they] are

associated or correlated. For him the non-observed variables are
correlated owing to their known geometrical relations with observed
variables; for us, [they] may be supposed to be uncorrelated causes,
and to be connected by unknown functional relations with the
correlated variables.

According to Krengel (2011), the modern notion of independence

of events is due to Bohlmann whom Kolmogorov in 1933 had not
mentioned. Kolmogorov introduced independence of events and
random variables.
5) The principle of maximum weight for [unbiased] estimators.
Gauss described this subject ponderously. For that matter, Helmert
(1872) and Idelson (1947) are in general much better understood.
Suppose that, without loss of generality, the initial equations are

ai x + bi y = Gi = gi + εi, i = 1, 2, …, n

where εi is the error of the free term gi. The estimators of the
unknowns might be represented by linear forms, for example by x =
[αG] with unknown coefficients αi so that

mx2 = [αα]m2 (5)

where m2 is the variance of an observation.

It is easy to prove that [aα] = 1, [bα] = 0 and the condition of
maximal weight will be

W = [αα] – 2Q11[aα] – 2Q12[bα] = max

146
145
where Q11 and Q12 are the Lagrange multipliers. Similar
considerations, and, in particular, an estimation of precision
resembling formula (5), are also possible for the other unknowns. It
occurs that the estimators of the unknowns are determined from the
normal equations and their weights are calculated by means of the
Lagrange multipliers of the type of Qii which, like the other
multipliers Qij, are determined from the same normal equations with
partly unit and partly zero free terms. Thus, in formula (5) [αα] = Q11.
According to the above, it follows that such formulas can be made
use of even before observation; the general layout of the geodetic
network and the crude values of its angles obtained during
reconnaissance make it possible to calculate the Qij. And (what Gauss
had not known) these multipliers are connected with covariations;
thus, Q12 = E(xy).
6) The estimator of the sample [variance]. Gauss (§§ 37 – 38)
proved that, for n observations and k unknowns, the unbiased sample
variance and its estimator were, respectively,

m2 = E[vv]/(n – k), mo2 = [vv]/(n – k) (6a, b)

where vi were the residual free terms of the initial equations. Instead
of the mean value, the sum of squares [vv] itself has to be applied.
Coupled with the principle of maximal weight (of least variance),
formulas (6) provide effective estimators, as they are now called.
Gauss (1823a/1887, p. 199) remarked that the acceptance of his
formula (6b) instead of the previous expression (§ 7.2-5), whose
denominator was equal to n, was demanded by the “dignity of
science”.
Gauss stressed that the estimator (6) was unbiased; however, the
practically applied estimator is not m2, but the biased m. Furthermore,
unbiased estimators do not exist in every case and some bias is
allowed (Sprott 1978, p. 194). Finally, I note that Czuber (1891, p.
460) testified that Helmert had thought that varmo2/mo2 was more
important than varmo2 by itself and Eddington (1933, p. 280)
expressed the same opinion. Czuber also proved that, for the normal
distribution, that relative error was minimal for the estimator (6b).
7) The precision of the estimator of the sample variance. Gauss (§
40) directly calculated the boundaries of the var mo2 by means of the
fourth moment of the errors and indicated that for the normal
distribution

varmo2 = 2m4/(n – k). (6c)

He somehow erred in calculating the abovementioned boundaries,

see (15); in addition, his formulas should have included the unknown
magnitude Eεi2 (εi were the observational errors) rather than m2.
Formula (6c) shows that mo2 is a consistent estimator of the sample
variance; this property persists in the general case, see formulas (15).
Many years later Bertrand (1888a) criticized the Gauss formula
(6c). Tacitly assuming the normal distribution, he provided an

147
146
example in which his own estimate of σ2 was less than that provided
by Gauss. He forgot, however, that formula (6c) provided an unbiased
estimate whereas his own estimate was biased. Then, he calculated σ2
forgetting the Gauss formula for the case of normal distribution. It was
this episode that led Czuber to the discussion described in § 9A.4-6.
8) Other topics. Gauss also determined the variance of a linear
function of the estimators of the unknowns (which are not
independent) and provided expedient procedures for further
calculations after additional data become known or after the weights
of some observations have to be changed.
9) Another manner of adjusting observations. In the supplement
(1828a) to his memoir Gauss described the adjustment of observations
by the MLSq according to the pattern of conditional observations. In
geodetic practice, it is often expedient to issue from the directly
measured magnitudes and conditional equations rather than from
observational equations (1.2). Sometimes both kinds of equations are
made use of at the same time, but I leave this case aside and consider
now a (later) typical chain of, say, 10 triangles of triangulation. Each
angle is measured as are the lengths of two extreme sides (baselines)
whose directions (azimuths) are determined by astronomical
observations. The observational errors are such that both the baselines
and the azimuths might be considered exact; only the angles are
adjusted. Each measured angle qi provides an equation

xi – qi = vi (7)

where the first term is the true value of the angle and the right side is
the sought correction. Now, the condition of closing the first triangle
(I disregard its excess) is

x1 + x2 + x3 – 180° = 0. (8)

Extremely simple is also the condition that demands that the

azimuth of the first baseline plus the algebraic sum of the appropriate
angles be equal to the azimuth of the second baseline. The sine
theorem is however needed for the transition from the first baseline to
the second one, but a first approximation is achieved by introducing
the measured angles so that the required trigonometric equation is
linearized. It follows that all the conditions can be written as

[av] + w1 = 0, [bv] + w2 = 0, etc. (9)

Formed by means of equations (7), they should be exactly fulfilled

and the number of the terms in the square brackets is either three, as in
equations of the type of (8), or more, depending on the number of the
triangles in the chain. The adjustment proper consists in determining
the conditional minimum of [vv] with the usual application of the
Lagrange multipliers and the corrections vi are determined through
these multipliers. Strangely enough, only Helmert (1872, p. 197) was
the first to provide such an explanation.

148
147
 10) A new exposition of the memoir (Sheynin 2012a; 2014).
Gauss could have derived formulas (6) in the very beginning of the
memoir, just after he introduced the variance (§ 9A.4-2) since the
required conditions (linearity of the initial equations, (physical)
independence of their free terms and unbiasedness of the estimators of
the unknowns) were not connected with the further exposition. And
the MLSq directly followed from those formulas (6). Hundreds of
textbooks had been describing the MLSq as justified by Gauss in
1809: such an approach was incomparably easier. Now we see that it
is quite possible to follow the memoir of 1823. Its very existence had
been barely known, see Eisenhart (1964, p. 24) in § 7.1-3.
The most eminent scientists (Boltzmann 1896/1909, p. 570;
Chebyshev, see § 13.2-7) had been barely acquainted with the work of
Gauss.
Many authors beginning with Gauss himself had derived the
formula (6) which was not difficult. The main point, however, is that
the proof does not depend on the condition of least squares. On the
contrary, this condition can now be introduced at once since it means
minimum variance. The formulas derived by Gauss for constructing
and solving the normal equations and calculation of the weights of
xˆ , yˆ ,... and of their linear functions will still be useful.
Gauss had actually provided two justifications of least squares (of
which I only left the second one), but why did not he even hint at this
fact? I can only quote Kronecker (1901, p. 42) and Stewart (Gauss
1823b – 1828/1995, p. 235):

The method of exposition in the “Disquisitiones [Arithmeticae”,

1801] as in his works in general is Euclidean. He formulates and
proves theorems and diligently gets rid of all the traces of his train of
thoughts which led him to his results. This dogmatic form was
certainly the reason for his works remaining for so long
incomprehensible.

Gauss can be as enigmatic to us as he was to his contemporaries.

Gauss himself actually said so. His eminent biographer, Sartorius

von Waltershausen (1856/1965, p. 82) testified: He had used to say
that, after constructing a good building, the scaffolding should not be
seen. And he had often remarked that his method of description
strongly hindered readers less experienced in mathematics.
Finally, I note Gauss’ words (letter to W. Olbers 30.7.1806): Meine
Wahlspruch [motto] ist aut Caesar, aut nihil.
The second substantiation of the MLSq can be accomplished by
applying the notions of multidimensional geometry (Kolmogorov
1946; Hald 1998, pp. 473 – 474). Nevertheless, the new exposition of
the memoir of 1823 is essential, and it appeared more than 200 years
after its publication!
Kolmogorov (p. 64) also believed that the formula for m2 (6a)
should, after all, be considered as its definition. Much earlier Tsinger
(1862, § 33) stated that it already “concealed” the MLSq which,

149
148
however, was only a hint at the real possibility of understanding
Gauss. Harter (1977, p. 28) stated almost the same.
9A.5. Additional Considerations
Having substantiated the MLSq, Gauss nevertheless deviated from
rigid rules; one pertinent example is in § 6.3.2. Here, I have more to
say.
1) The number of observations. In his time, methods of geodetic
observations were not yet perfected. Gauss himself was successfully
developing them and he understood that a formal estimation of
precision can only describe the real situation after all the conditions (§
9A.4-9) were allowed for, i.e., only after all the field work was done.
It is no wonder, then, that Gauss continued to observe each angle at
each station until being satisfied that further work was useless, see the
very end of Chapter 6.
2) Rejection of outliers. This delicate operation does not yield to
formal investigation since observations are corrupted by systematic
errors, and, in general, since it is difficult to distinguish between a
blunder and a “legitimate” large error. Statistical tests, that had
appeared in the mid-19th century, have not been widely used in the
theory of errors. Here is the opinion of the authors (Barnett & Lewis
1978, p. 360) of a book on this subject:

When all is said and done, the major problem in outlier study
remains the one that faced the very earliest workers […] – what is an
outlier and how should we deal with it?

I still refer to Laplace (1918, p. 534) and Gauss (letter to Olbers of

1827, W-8, pp. 152 – 153):

For a successful application of the calculus of probability to

geodetic observations one should honestly report about all of his own
observations which he admitted and not reject any of them solely
because they are somewhat (!) remote from the rest.

When the number of observations was not very large, and a sound
knowledge of the subject was lacking, rejection was always doubtful,
and in any case nothing should be concealed so that others will be
able to decide otherwise.

3) Calculations. Without even an arithmometer, Gauss was able to

carry out difficult calculations; once he solved a system of 55 normal
equations (letter to Olbers of 1826; W-9, p. 320). For other examples
see Sheynin (1979, p. 53). His preparatory work (station adjustment;
compilation of the initial equations, see § 9A.4-9, and of the normals
themselves) had to be very considerable as well.
Sometimes Gauss applied iterative calculations (letter to Gerling of
1823; W-9, pp. 278 – 281), also see Forsythe (1951) and Sheynin
(1963). The first to put on record this fact, in 1843, was Gerling
himself. Then, Gauss (1809b, § 185) left an interesting qualitative
remark stating that “it is often sufficient” to calculate approximately
the coefficients of the normal equations. The American astronomer

150
149
Bond (1857) had applied Gauss’ advice and Newcomb (1897a, p. 31)
followed suit.
As a calculator of the highest calibre (Maennchen 1918/1930, p. 3),

Gauss was often led to his discoveries by means of mentally

agonizing precise calculations […]; we find [in his works] substantial
tables whose compilation would in itself have occupied the whole
working life of some calculators of the usual stamp.

I ought to add that Gauss made some mistakes in his computations

possibly because, first, he had not invariably checked them, see for
example Gerardy (1977) or his own methodological note (1823c)
where the signs of dx and dy were wrong. Second, Gauss calculated
“unusually fast” (Maennchen 1918/1930, p. 65ff) which caused
mistakes and additional difficulties in proving that he had applied the
MLSq before 1805.
Maennchen did not study Gauss’ geodetic calculations possibly
because in his time the solution of systems of linear equations had not
yet attracted the attention of mathematicians.
For my part, I note that, when compiling a certain table of
mortality, Gauss (W-8, pp. 155 – 156) somehow calculated the values
of exponential functions bn and cn for n = 3 and 7(5)97 with lgb =
0.039097 and lgc = – 0.0042225.
Here, now, is Subbotin’s conclusion (1956, p. 297) about the
determination of the orbits of celestial bodies but applicable to my
subject as well: Lagrange and Laplace

Restricted their attention to the purely mathematical aspect [of the

problem] whereas Gauss had thoroughly worked out his solution from
the point of view of computations taking into account all the
conditions of the work of astronomers and [even] their habits.

4) Estimation of precision (Sheynin 1994a, pp. 265 – 266). In his

letters to Bessel (in 1821) and Gerling (in 1844 and 1847) Gauss
stated that the estimation of precision based on a small number of
observations was unreliable. In 1844 he combined observations made
at several stations and treated them as a single whole, cf. Laplace’s
attitude (§ 9A.2-7). And in 1847 Gauss maintained that, lacking
sufficient data, it was better to draw on the general knowledge of the
situation.
9A.6. More about the Method of Least Squares
1) In spite of Gauss’ opinion, his first justification of the MLSq
became generally accepted (§ 9A4-10 and Sheynin 1995c, § 3.4), in
particular because the observational errors were (and are)
approximately normal whereas his mature contribution (1823b) was
extremely uninviting; and the work of Quetelet (§ 10.5) and Maxwell
(§ 10.8.5) did much to spread the idea of normality. Examples of
deviation from the normal law were however accumulating both in
astronomy and in other branches of natural sciences as well as in
statistics (Sheynin 1995c, § 3.5; again Quetelet and Newcomb, see §

151
150
10.8.4). See also § 9C in which I describe the surprising attitude of
Bessel who had actually concealed such deviations.
And, independently from that fact, several authors came out against
the first substantiation. Markov (1899a), who referred to Gauss
himself (to his letter to Bessel, see my § 9A.4-2), is well known in this
respect but his first predecessor was Ivory (§ 10.9-1).
The second justification was sometimes denied as well. Thus,
Bienaymé (1852, p. 37) declared that, unlike Laplace, Gauss had
provided considerations rather than proofs (?); see also Poincaré’s
opinion in § 11.2-7.
2) When justifying the MLSq in 1823 in an essentially different
way, Gauss called the obtained estimators most plausible (maxime
plausibiles, or, in his preliminary report (1821), sicherste, rather than
as before, maxime probabile, wahrscheinlichste. For the case of the
normal distribution, these are jointly effective among unbiased regular
estimators5.
3) Mathematicians had not paid due attention to Gauss’ work on the
MLSq (§§ 9A.2-3 and 13.2-7), and neither did statisticians, see the
Epigraph to this book which apparently complements the following
passage (Eisenhart 1978, p. 382):

When Karl Pearson and G. Udny Yule began to develop the

mathematical theory of correlation in the 1890s, they found that much
of the mathematical machinery that Gauss devised […] was
immediately applicable. […] Gauss’ contributions to the method of
least squares embody mathematics essential to statistical theory and
its applications in almost every field of science today.

I really think that K. P. and Yule only discovered Gauss at a late stage
of their work.
9B. Helmert
It was Helmert who mainly completed the development of the
classical Gaussian theory of errors; furthermore, some of his findings
were interesting for mathematical statistics. With good reason
Schumann (1917, p. 97) called him “Master of both the niedere
[surveying and applied] geodesy and higher [triangulation etc.,
gravimetry, figure of the Earth] geodesy”. Until the 1930s, Helmert’s
treatise (1872) remained the best source for studying the error theory
and the adjustment of triangulation.
Indeed, its third, posthumous edition of 1924 carried a few lines
signed by a person (H. Hohenner) who explained that, upon having
been asked by the publishers, he had stated that the treatise still
remained the best of its kind. His opinion, he added, convinced the
publishers.
Helmert (1886, pp. 1 and 86) was the first to consider appropriate
geodetic lines rather than chains of triangulation, and this innovation,
developed by Krasovsky, became the essence of the method of
adjustment of the Soviet primary triangulation (see Note 19 to Chapter
6 and Sakatow 1950, § 91). Another of his lesser known contributions
(Helmert 1868) was a study of various configurations of geodetic
systems. Quite in accordance with the not yet existing linear

152
151
programming, he investigated how to achieve necessary precision
with least possible effort, or, to achieve highest possible precision
with a given amount of work. Some equations originating in the
adjustment of geodetic networks are not linear, not even algebraic;
true, they can be linearized (§ 9A.4-9), and perhaps some elements of
linear programming could have emerged then, in 1868, but this had
not happened. Nevertheless, Helmert noted that it was expedient to
leave some angles of a particular geodetic system unmeasured, but
this remark was only academic: all angles have always been measured
at least for securing a check upon the work as a whole.
I describe now Helmert’s stochastic findings.
1) The chi-square distribution. I (1966) noted that E. Abbe (1863),
see also M. G. Kendall (1971), derived it as the distribution of the
sum of the squares of normally distributed errors. He wished to offer a
test for revealing systematic errors, and he required, in particular, the
distribution of the abovementioned function of the errors since it was
indeed corrupted by those errors. Exactly his test rather than the
distribution obtained was repeatedly described in the geodetic
literature whereas Linnik (1958/1961, pp. 109 – 113) introduced a
modified version of the Abbe test.
Helmert (1876b) provided his own derivation of the χ2 distribution
which he first published without justification (1875a). Neither then
nor much later (see Item 2) did he mention Abbe. Actually, he
continued after Gauss (1816), see § 9A.3, by considering
observational errors ε1, ε2, …, εn and the sum of their powers Σεin for
the uniform and the [normal] distributions and for an arbitrary
distribution as n → ∞. In the last instance, he proved the Gauss
formula (4) and then specified it for the abovementioned distributions.
He derived the χ2 distribution by induction beginning with n = 1 and
2; Hald (1952, pp. 258 – 261) provided a modernized derivation.
2) Much later Helmert (1905) offered a few tests for revealing
systematic influences in a series of errors which he wrote down as

v1ε1 + v2ε2 + … + vnεn

with vi = 1 or – 1 and εi > 0. He issued from the formula

P(|ξ – Еξ| ≤ m) ≈ 0.68 (10)

where m was the mean square error of ξ (and thus restricted his
attention to the normal law): if the inequality in the left side of (10)
did not hold, then, as he thought, systematic influences were present.
When deriving his tests, Helmert considered Σvi, Σ|vi|, runs of signs of
the vi and functions of the errors εi themselves and in this last-
mentioned case he provided a somewhat modified version of the Abbe
test.
3) The Peters formula (1856) for the mean absolute error. For n
normally distributed errors it was

θ = Σ|vi|/ n( n  1) , 1 ≤ i ≤ n (11)

153
152
with vi being the deviations of the observations from their arithmetic
mean. Helmert (1875a) derived this formula anew because Peters had
tacitly and mistakenly assumed that these deviations were mutually
independent. Passing over to the errors εi, Helmert calculated the
appropriate integral applying for that purpose the Dirichlet
discontinuity factor. However, since the normal distribution is stable,
it is possible to say now at once (H. A. David 1957) that formula (11)
is correct because

EΣ|vi| = n( n  1) /h√π

where h is the appropriate parameter [measure of precision] of the

initial normal distribution and, as it should be, θ = 1/h√π.
Helmert also attempted to generalize the Peters formula by
considering indirect measurements with k unknowns (k > 1). He was
unable to derive the appropriate formula but proved that a simple
replacement of (n – 1) in formula (10) by (n – k) resulted in
underestimating the absolute error.
4) Helmert (1876b) calculated the variance of the estimator (11).
His main difficulty here was the derivation of E|vivj|, i < j, but he was
able to overcome it and obtained

{π/2 + arcsin[1/(n – 1)] – n + n( n  2) }/πnh2.

Fisher (1920, p. 761) derived this formula independently.

5) In the same paper Helmert investigated the precision of the
Gauss formula (6b). For direct measurements it can be replaced by the
expression for the mean square error

[vv ]
m= .
n 1

Helmert derived it for the normal distribution by the principle of

maximum likelihood, but had not remarked that the estimator obtained
(which, however, directly followed from (6a) and was always applied
in practice in geodesy) was, unlike the Gauss formula, biased.
Denote the observational errors by εi and their mean by ε, then

vi = εi – ε

and the probability that these errors had occurred, as Helmert

indicated in the context of his proof, was equal to

P = n(h/√π)nexp[– h2([vv] + nε2)] dv1 dv2 … dvn–1 dε. (12)

This formula shows that, for the normal distribution, [vv], – and,
therefore, the variance as well,– and the arithmetic mean are
independent. Helmert had thus proved the important Student – Fisher
theorem although without paying any attention to it.

154
153
 A special feature in Helmert’s reasoning was that, allowing for (6c),
he wrote down the Gauss formula 6b) for the case of direct
measurements (and, to repeat, for the normal distribution) as6

[vv ] 2
mo2 = [1 ± ]; (13)
n 1 n 1

that is, he considered the variance together with its mean square error.
Gauss (1816, §§ 6 and 8) sometimes, but not always, acted the same
way (although without applying that term, mean …, which I found in
the mid-19th century in the works of Chebyshev and Russian
artillerists).
In addition, Helmert noted that for small values of n the varmo2 did
not estimate the precision of formula (6b) good enough and derived
the following formula

[vv ] 2 (n / 2)
E[m – ] = (1/h2){1 – √2 }. (14)
n 1  [(n  1) / 2] n  1

He issued from the probability of the values of vi, i = 1, 2, …, (n – 1),

P = √n(h/√π)n–1 exp(– h2[vv]) dv1 dv2 … dvn–1

that follows from formula (12), noted that the probability P(ε ≤ [vv] ≤
ε + dε) was equal to the appropriate integral, and introduced new
variables

t1 = √2(v1 + 1/2v2 + 1/2v3 + 1/2v4 + … + 1/2vn–1),

t2 = 3 / 2 (v2 + 1/3v3 + 1/3v4 + … + 1/3vn–1),
t3 = 4 / 3 (v3 + 1/4v4 + … + 1/4vn–1), …,
tn – 1 = n /( n  1) vn–1.

Note that [vv] = [tt] where, however, the first sum consisted of n
terms and the second one, of (n – 1) terms, and the Jacobian of the
transformation was √n. The derivation of formula (14) now followed
immediately since Helmert knew the χ2 distribution. Taken together,
the transformations from {ε} to {v} and from {v} to {t} are called
after him.
Kruskal (1946) transformed formula (12) by introducing a
bivariate “Helmert distribution” with variables

s= [vv ] / n , u = x – µ,

where x was the arithmetic mean of n normally distributed

observations N(µ; σ), and replaced h by σ. He mentioned several
authors who had derived that new distribution by different methods,
determined it himself by induction and indicated that the Student
distribution followed from it, see Hald (1998, p. 424).

155
154
 Finally, Helmert corrected the boundaries of the estimator (6b). As
indicated by Gauss they were

2(ν4 – 2s4)/(n – k); [1/(n – k)](ν4 – s4) + (k/n)(3s4 – ν4)

where ν4 was the fourth moment of the errors and s2 = Em2. Helmert
had discovered that the lower boundary was wrong and Kolmogorov
et al (1947) independently repeated his finding. Here is the final
result; Maltzev (1947) proved that the lower bound was attainable. For
non-negative and then non-positive (v4 – 3s4), the product
(n – k)var mo2 as it occurred, was contained within, respectively,
boundaries

[(ν4 – s4) – (k/n) (ν4 – 3s4); (ν4 – s4)], (15a)

[(ν4 – s4); (ν4 – s4) + (k/n)(3s4 – ν4)]. (15b)

9C. Bessel
Friedrich Wilhelm Bessel (1784 – 1846) was an outstanding
astronomer and an eminent mathematician. He and Gauss were the
originators of a new direction in practical astronomy and geodesy
which demanded a thorough examination of the instruments and
investigation of the plausibility of observational methods. Each
instrument was accused of every possible defect and was only
exonerated after proving itself irreproachable. See an appropriate
quotation from Newcomb in Schmeidler (1984, pp. 32 – 34).
I mentioned Bessel in §§ 6.3 and 8.7. His achievements in
astronomy and geodesy are well known; I name the determination of
astronomical constants; the first determination of a star’s parallax; the
discovery of the personal equation; the development of a method of
adjusting triangulation (see however below); design and use of
metallic bars for measuring baselines, and the derivation of the
parameters of the Earth’s ellipsoid of rotation which enjoyed
international recognition (Strasser 1957, p. 39).
The personal equation is the systematic difference of the moments
of the passage of a star through the cross-hairs of an astronomical
instrument as recorded by two observers. When studying this
phenomenon, it is necessary to compare the moments fixed by two
astronomers at different times and, consequently, to take into account
the correction of the clock (if they use only one). Bessel (1823) had
indeed acted appropriately, but in one case he failed to do so, and his
pertinent observations proved useless. He made no such comment;
Bessel (1838a, §§ 1 and 2) determined the densities of two
functions of a continuously and uniformly distributed [random
variable], and, unlike Laplace, he clearly formulated this problem.
Nevertheless, he erred in his computations of the pertinent variances
and probable errors. He also determined the density of the total
observational error made up of many heterogeneous components but a
rigorous solution of such problems became only possible mush later (§
13.1-4)7.
I have discovered 33 mistakes in arithmetic and elementary algebra
(except those noticed by the Editor) in his Abhandlungen (1876). They

156
155
did not influence his conclusions but they throw doubt on his more
serious calculations. Here is just one of them (1876, Bd. 2, p. 376):
√4:√5 = 1/1.409; actually, however, 1/1.118.
One more example, this time concerning Bessel’s reasoning (1818;
1838a). He presented three series of Bradley’s observations, 300, 300
and 470 in number, and stated that their errors almost precisely
obeyed normal distributions. Actually, he was wrong and it is difficult
to believe that he was mistaken (especially see below). Moreover, he
thus missed the opportunity to discover an example of long series not
quite normally distributed errors of precise observations. Later,
scientists gradually discovered such series, in the first place see
Newcomb (1886).
Bessel’s contribution (1838a) included a proof of a version of the
CLT (rigorously proved only by Liapunov and Markov). Bessel stated
that, given more observations, the deviation from normality will
disappear. Did not he notice that he thus undermined the essence of
that theorem? Did not he formulate his wrong conclusion to save that
proof?
It became customary to measure each angle of a chain of
triangulation an equal number of times and, which was more
important, to secure their mutual independence so as to facilitate the
treatment of the observations, – to separate the station adjustment
from the adjustment of the chain as a whole. Bessel, however, did not
keep to the abovementioned condition (and had to adjust all the
observations at once). There are indications that the actual rejection of
his method annoyed him8.
I have since discovered other examples of Bessel’s misleading
statements in his popular writings. True, at least one of them pertains
to the time of his fatal illness, but I venture to suppose that a very ill
person should all the more try to avoid mistakes.
1. Bessel (1843). This is his report of the same year read out to the
physical section of the Königsberg physical-economic society in
which he had been very active. Schumacher published the texts of
these reports (1848b), and Bessel (1848a), about which I say a few
words below, is included in that collection.
And so, Bessel (1843) described the life and work of William
Herschel. Among other things, he properly discussed Herschel’s hunt
for double stars and his attempt at counting the stars in the Milky
Way, but he did not explain that there are two types of double stars
nor did he say that the Milky Way is only one of the countless
galaxies.
Herschel came to understand that his telescope did not penetrate to
the boundaries of the Milky Way (F. G. W. Struve 1847, p. 34; Hoskin
1959) whereas Bessel (p. 474, left column) stated quite the opposite.
Another mistake concerned the discovery of the planet Uranus.
Contrary to Bessel’s statement (p. 469, left column), Herschel
discovered a moving body and decided that it was a comet. Finally,
Bessel (p. 470, right column) mentioned Caroline, the sister of
William, and remarked that she was still alive and assisted her brother.
Actually, Caroline died several decades later than he.
2. Bessel (1845). This is a newspaper article which had nothing to

157
156
do with astronomy. Bessel stated that under such parameters as
territory, climate etc. (political system not mentioned) only mental
development of the population determined its acceptable maximal
number. However, a territory becomes more or less populated when
people turn from hunting to farming (Bessel’s own example), but are
farmers more mentally developed than hunters?
Then Bessel turned his attention to the United States and provided
his own data about the population of Native Americans taken out of
thin air and damnably wrong.
3. Bessel (1848a). The date of the report is unknown. Bessel
mentioned Delambre’s Astronomy which was not quite definite, but
sufficient for stating that the report was read in 1821 or later.
The significance of Jakob Bernoulli’s law of large numbers was not
discussed, Lambert’s preference of maximum-likelihood estimators
over the arithmetic mean (p. 401) was mostly imagined and Laplace’s
Essai philosophique of 1816 was not even mentioned. Population
statistics studied, for example, by De Moivre, Nicolaus and Daniel
Bernoulli, was completely left out. It is difficult to conclude that
Bessel’s quite elementary exposition had satisfied his listeners.
In his correspondence, Gauss several times indicated Bessel’s
shortcomings.
1. Gauss – Olbers, 2 Aug. 1817. Bessel had overestimated the
precision of some of his measurements. On 2 Nov. 1817 Olbers
confidentially informed Bessel about Gauss’ opinion.
2. Gauss – Schumacher, between 14 July and 8 Sept. 1826. He
stated the same about Bessel’s investigation of the precision of the
graduation of a limb.
3. Gauss – Schumacher 27 Dec. 1846. He negatively described
some of Bessel’s posthumous manuscripts. In one case he was
shocked by Bessel’s carelessness8.
I am at a loss: how was it possible to pass these statements over?
And, again, how was it possible for Bessel to be at once a great
scholar and a happy-go-lucky scribbler? Cf. Goethe (Faust, pt. 1, Sc.
2): Two souls are living in his breast.

Notes
1. This term should only be applied to the method as substantiated by Gauss in
1823; until then, strictly speaking, the principle of least squares ought to be thought
of.
2. Adrain included his work in a periodical published by himself for the year
1808; however, its pertinent issue appeared only in 1809 (Hogan 1977). Adrain’s
library included a copy of Legendre’s memoir (Coolidge 1926) in which, however,
the normal distribution was lacking; furthermore, it is unknown when had Adrain
obtained that memoir. The term normal distribution appeared in 1873 (Kruskal
1978) and was definitively introduced by K. Pearson (1894).
3. Their opinion should not be forgotten. Here is another example. Encke (1851,
p. 2) believed that Gauss had applied the MLSq when determining the orbit of
Ceres, the first-discovered minor planet (Gauss did not comment). In Note 19 to
Chapter 6 I mentioned an inadmissible free and easy manner adopted by a certain
author (Stigler 1986) with respect to Euler. His attitude towards Gauss was not
better. Here are his statements: Legendre “immediately realized the method’s
potential” (p. 57), but “there is no indication that [Gauss] saw its great potential
before he learned of Legendre’s work” (p. 146); then (p. 143), only Laplace saved
Gauss’s argument [his first justification of the MLSq] from joining “an

158
157
accumulating pile of essentially ad hoc constructions”; and, finally (p. 145), Gauss
“solicited reluctant testimony from friends that he had told them of the method
before 1805”. I (Sheynin 1999a; 1999c) had refuted these astonishing declarations
(see also the very end of § 9A.1 about Olbers) which Stigler (1999), the first ever
slanderer of the memory of the great man, repeated slightly less impudently.
Regrettably, no one supported me; on the contrary, Stigler’s first book met with
universal approval although he, in addition, left aside the ancient history as well as
such scholars as Kepler, Lambert and Helmert. Hald (1998, p. xvi), whose
outstanding contribution deserves highest respect, called Stigler’s book “epochal”. I
am unable to understand suchlike opinions. The attitude of the scientific comunity
towards Stigler proves that it is seriously ill.
4. It is not amiss to add that the primary triangulation of the Soviet Union
consisted of chains independent one from another in the Gauss’ sense. This, together
with other conditions, enabled the geodesists to estimate realistically the precision of
the whole great net (Sakatow 1950/1957, pp. 438 – 440). And in general, geodesists,
not necessarily mentioning Gauss, were keeping to his opinion. I also note that
Kapteyn (1912), who had not cited Gauss and was unsatisfied with the then
originating correlation theory, proposed to estimate quantitatively the dependence
between series or functions of observations by issuing from the same notion of
independence, see Sheynin (1984a, § 9.2.1). His article went unnoticed.
5. Concerning this rarely mentioned concept see Cramér (1946, § 32.6).
6. In the theory of errors, the application of the mean square error with a double
sign became standard (§ 9B-4).
7. In 1839 Gauss informed Bessel (W-8, pp. 146 – 147) that he had read the
latter’s memoir with interest although the essence of the problem had been known to
him for many years.
8. In 1825, Gauss had a quarrel with Bessel but no details are known (Sheynin
2001d, p. 168). Even in 1817 Olbers (Erman 1852, Bd. 2, p. 69) regretted that the
relations between Bessel and Gauss were bad. In 1812, in a letter to Olbers, Bessel
(Ibidem, Bd. 1, p. 345) had called Gauss “nevertheless” the inventor of the MLSq,
but in 1844, in a letter to Humboldt (Sheynin 2001d, p. 168), he stressed
Legendre’s priority.

159
158
 10. The Second Half of the 19th Century
Here, I consider the work of several scholars (§§ 10.1 – 10.6),
statistics (§ 10.7), and its application to various branches of natural
sciences (§ 10.8). The findings of some natural scientists are discussed
in § 10.9 since it proved difficult to describe them elsewhere.
10.1. Cauchy
Cauchy published not less than 10 memoirs devoted to the
treatment of observations and the theory of probability. Eight of them
(including those of 1853 mentioned below) were reprinted in t. 12, sér.
1, of his Oeuvres complètes (1900). In particular, he studied the
solution of systems of equations by the principle of minimax (§ 6.3.2)
and proved the theorem in linear programming known to Gauss
(§ 9A.2-1). He had also applied the method of averages (§ 6.3.2) and
Linnik (1958/1961, § 14.5), who cited his student L. S. Bartenieva,
found out that the pertinent estimators were unbiased and calculated
their effectiveness for the cases of one and two unknown(s). I briefly
describe some of Cauchy’s findings.
Cauchy (1853b) derived the even density of observational errors
demanding that the probability for the error of one of the unknowns,
included in equations of the type of (1.2), to remain within a given
interval, was maximal. Or, rather, he derived the appropriate
characteristic function

φ(θ) = exp(– θµ+1), c, θ > 0, µ real (1)

and noted that the cases µ = 1 and 0 led to the [normal law] and to the
“Cauchy distribution”, see § 8.6. The function (1) is characteristic
only when 1 < µ ≤ 1 and the appropriate distributions are [stable].
In two memoirs Cauchy (1853c; 1853d) proved the [CLT] for the
linear function

A = [mε] (2)

of [independent] errors εi having an even density on a finite interval.

In both cases he introduced characteristic functions of the errors and
of the function (2), obtained for the latter

Ф(θ) = exp(– sθ2)

where 2s was close to σ2, the variance of (2), and, finally, arrived at

α
2
P(|А| ≤ α) ≈
σ π

0
exp(– x2/2σ2)dx.

It is important that he had also estimated the errors due to

assumptions made and Freudenthal (1971, p. 142) even declared that
his proof was rigorous by modern standards; see, however Heyde &
Seneta (1977, pp. 95 – 96).
Cauchy devoted much thought to interpolation of functions, and, in
this connection, to the MLSq, but, like Poisson, he never cited Gauss.
In one case he (1853a/1900, pp. 78 – 79) indicated that the MLSq
160
159
provided most probable results only in accordance with the
Laplacean approach [that is, only for the normal distribution] and
apparently considered this fact as an essential shortcoming of the
method.
10.2. Bienaymé
Heyde & Seneta (1977) described his main findings; I follow their
account and abbreviate their work as HS. Bru et al (1997) published
two of Bienaymé’s manuscripts and other relevant archival sources.
1) A limit theorem (Bienaymé 1838; HS, pp. 98 – 103). Bienaymé
had “essentially” proved the theorem rigorously substantiated by
Mises (1919; 1964a). The abovementioned adverb appeared in Mises
(1964b, p. 352). Suppose that n trials are made with some event Ai
from m mutually exclusive events (i = 1, 2, …, m) occurring in each
trial with probability θi and that xi is the number of times that Ai
happened, Σxi = n. Treating the probabilities θi as random variables,
Bienaymé studied the distribution of their linear function in the
limiting case xi, n → ∞, xi/n = Ci. As a preliminary, he had to derive
the posterior distribution of the θi given xi tacitly assuming that the
first (m – 1) of these probabilities were random variables with a
uniform prior distribution. Actually, Bienaymé proved that the
assumption about the prior distributions becomes insignificant as the
number of the multinomial trials increases.
Note that Nekrasov (1890) had forestalled Czuber, whom Mises
named as his predecessor. Assuming some natural restrictions, he
proved a similar proposition concerning the Bernoulli trials.
2) The Liapunov inequalities (Bienaymé 1840b; HS, pp. 111 –
112). Without proof, Bienaymé1 indicated that the absolute initial
moments of a discrete [random variable] obeyed inequalities which
could be written as

(E|ξ|m)1/m ≤ (E|ξ|n)1/n, 0 ≤ m ≤ n.

Much later Liapunov (1901a, § 1) proved that

(E|ξ|m)s–n < (E|ξ|n)s–m < E(|ξ|s)m–n, s > m > n ≥ 0.

He applied these inequalities when proving the [CLT].

3) The law of large numbers. Bienaymé (1839) noted that the
fluctuation of the mean statistical indicators was often greater than in
accordance with the Bernoulli law, and suggested a possible reason:
some causes acting on the studied events, as he thought, remained
constant within a given series of trials but essentially changed from
one series to the next one. Cournot, Lexis and other “Continental”
statisticians took up this idea without citing Bienaymé (Chapter 15)
but it was also known in the theory of errors since systematic errors
can behave in a similar way. Bienaymé, in addition, somehow
interpreted the Bernoulli theorem as an attempt to study suchlike
patterns of the action of causes. He (1855/1876) repeated this
statement and, on p. 202, he mistakenly reduced the Poisson LLN to
the case of variable probabilities whose mean value simply replaced
the constant probability of the Bernoulli trials, also see HS, § 3.3.

161
160
 4) The Bienaymé – Chebyshev inequality (Bienaymé 1853; HS, pp.
121 – 124; Gnedenko & Sheynin 1978/2001, pp. 258 – 262). This is
the name of the celebrated inequality

P(|ξ – Eξ| < β) > 1 – varξ/β2, β > 0. (3)

Differing opinions were pronounced with regard to its name and to

the related method of moments. Markov touched on this issue four
times. In 1912, in the Introduction to the German edition of his
Treatise (1900a/1908), he mentioned “the remarkable Bienaymé –
Chebyshev method”. At about the same time he (1912b, p. 218)
argued that

Nekrasov’s statement [that Bienaymé’s idea was exhausted in

Chebyshev’s works] is refuted by indicating a number of my papers
which contain the extension of Bienaymé’s method [to the study of
dependent random variables].

Then, Markov (1914b/1981, p. 162) added that the “starting point”

of Chebyshev’s second proof of Poisson’s LLN “had been […]
indicated by […] Bienaymé” and that in 1874 Chebyshev himself
called this proof “a consequence of the new method that Bienaymé
gave”. Nevertheless, Markov considered it “more correct” to call the
method of moments after both Bienaymé and Chebyshev, and
“sometimes” only after the latter, since “it only acquires significance
through Chebyshev’s work” [especially through his work on the
CLT]. Finally, Markov (Treatise, 1924, p. 92) stated that Bienaymé
had indicated the main idea of the proof of the inequality (3), although
restricted by some conditions, whereas Chebyshev was the first to
formulate it clearly and to justify it.
Bienaymé (1853/1867, pp. 171 – 172) considered a random sum,
apparently (conforming to the text of his memoir as a whole)
consisting of identically distributed terms, rather than an arbitrary
magnitude ξ, as in formula (3). This is what Markov possibly thought
of when he mentioned some conditions. HS, pp. 122 – 123, regarded
his proof, unlike Chebyshev’s substantiation [§ 13.1-3], “short,
simple, and […] frequently used in modern courses …” Yes, Hald
(1998, p. 510) repeated it in a few lines and then got rid of the sum by
assuming that it contained only one term. Gnedenko (1954/1973, p.
198) offered roughly the same proof but without citing Bienaymé.
Bienaymé hardly thought that his inequality was important
(Gnedenko & Sheynin 1978/2001, p. 262; Seneta 1998, p. 296). His
main goal was to prove that only the variance was an acceptable
estimator of precision in the theory of errors (see § 9A.4-2) and,
accordingly, he compared it with the fourth moment of the sums of
random [and independent] errors. Consequently, and the more so since
he never used integrals directly, I believe that Chebyshev (1874), see
also Gnedenko & Sheynin 1978/2001, p. 262) overestimated the part
of his predecessor in the creation of the method of moments. Here are
his words:

162
161
 The celebrated scientist presented a method that deserves special
attention. It consists in determining the limiting value of the integral
[…] given the values of the integrals…

The integrand in the first integral mentioned by Chebyshev was f(x)

and the limits of integration were [0; a]; in the other integrals, xf(x),
x2f(x), … and the limits of integration, [0; A], f(x) > 0 and A > a.
5) Runs up and down (Bienaymé 1874; 1875; HS, pp. 124 – 128).
Suppose that n observations of a continuous random variable are
given. Without proof Bienaymé indicated that the number of intervals
between the points of extremum (almost equal to the number of these
points) is distributed approximately normally with parameters

mean …(2n – 1)/3, variance ... (16n – 29)/90. (4)

He maintained that he knew this already 15 or 20 years previously.

HS states that these findings were discovered anew; nevertheless, the
authors derive formulas (4), the first of them by following Bertrand,
also see Moore (1978, p. 659). Bienaymé checked the agreement
between several series of observations and his findings. Some of the
data did not conform to his theory and he concluded that that
happened owing to unrevealed systematic errors. I return to his test in
§ 10.3.
6) The method of least squares (Bienaymé 1852; HS, pp. 66 – 71).
Bienaymé correctly remarked that least variance for each estimator
separately was not as important as the minimal simultaneous
confidence interval for all the estimators. Keeping to the Laplacean
approach to the MLSq (see his remark in my § 9A.6-1), he restricted
his attention to the case of a large number of observations. Bienaymé
also assumed that the distribution of the observational errors was
known and made use of its first moments and even introduced the first
four cumulants and the multivariate Gram – Charlier series (Bru
1991, p. 13; Hald 2002, pp. 8 – 9). He solved his problem by applying
the principle of maximum likelihood, introducing the characteristic
function of the [vector of the] errors and making use of the inversion
formula. True, he restricted his choice of the [confidence] region; on
the other hand, he derived here the χ2 distribution. Bienaymé’s
findings were interesting indeed, but they had no direct bearing on the
theory of errors. Furthermore, his statement (pp. 68 – 69) that both the
absolute expectation and variance were unreliable estimators of
precision was certainly caused by his adoption of the method of
maximum likelihood and is nowadays forgotten.
7) A branching process (Bienaymé 1845; HS, pp. 117 – 120).
Bienaymé had formulated the properties of criticality of a branching
process while examining the same problem of the extinction of noble
families that became attributed to Galton. D. G. Kendall (1975)
reconstructed Bienaymé’s proof and reprinted his note and Bru (1991)
quoted a passage from Cournot’s contribution of 1847 who had
solved a stochastic problem in an elementary algebraic way and had
indicated that it was tantamount to determining the probability of the
duration of the male posterity of a family,– to a problem in which

163
162
“Bienaymé is engaged”. Bru thought it highly probable that Cournot
had borrowed his study from Bienaymé.
8) An approach to the notion of sufficient estimator (Bienaymé
1840a; HS, pp. 108 – 110). When investigating the stability of
statistical frequencies (see also Item 3), Bienaymé expressed ideas that
underlie the notion of sufficient estimators. For m and n Bernoulli
trials with probability of success p the number of successes has
probability

P(ξi = k) = Csk pk(1 – p)s–k

with s = m and s = n respectively and i = 1 and 2 denoting these series

of trials. It is easy to ascertain that the probability P(ξ1 = r,
ξ2 = a – r|ξ1 + ξ2 = a) does not depend on p. Bienaymé thought that
this property, that takes place when the totality of the trials is
separated into series, could prove the constancy of the laws of nature.
However, statisticians (Fourier, whom he mentioned; Quetelet 1846,
p. 199 ) pragmatically considered such a separation as the best method
for revealing variable causes. HS additionally noted that Bienaymé
should have understood that all the information about the unknown
probability p [if it were constant] was provided by the totality of the
trials, that Bienaymé had wrongly calculated the variance of the
hypergeometric distribution, and that he made use of a particular case
of the CLT for checking a null hypothesis.
10.3. Cournot
Cournot intended his main contribution (1843) for a broader circle
of readers. However, lacking a good style and almost completely
declining the use of formulas, he hardly achieved his goal. Recall also
(the end of § 8.7) that Cournot passed over in silence the LLN. I
describe his work as a whole; when referring to his main book, I
mention only the appropriate sections.
1) The aim of the theory of probability. According to Cournot
(1875, p. 181), it was “The creation of methods for assigning
quantitative values to probabilities”. He thus moved away from
Laplace (§ 7.3) who had seen the theory as a means for revealing the
laws of nature. Cf. Chebyshev’s opinion in § 13.2-1.
2) The probability of an event (§ 18): this is the ratio of the extent
(étendue) of the favourable chances to the complete extent of all the
chances2. The modern definition replaced “extent” by a clear
mathematical term, “measure”. I stress that Cournot’s definition
included geometric probability, which until him had been lacking any
formula, and thus combined it with the classical case. Cournot (§ 113)
also introduced probabilities unyielding to measurement and (§§ 233
and 240-8) called them philosophical. They might be related to expert
estimates whose treatment is now included in the province of
mathematical statistics. Cf. Fries in § 7.1-5.
3) The term médiane. This is due to Cournot (§ 34).
4) The notion of randomness. In a book devoted to exonerating
games of chance La Placette (1714) explained (not clearly enough)
randomness as an intersection of independent chains of determinate
events, thus repeating the statements of many ancient scholars (see §

164
163
1.1.1). Cournot (§ 40) expressed the same idea, and, in § 43, indirectly
connected randomness with unstable equilibrium by remarking that a
right circular cone, when stood on its vertex, fells in a “random”
direction. This was a step towards Poincaré’s viewpoint (§ 11.2-9).
Cournot (1851, § 33, Note 38; 1861, § 61, pp. 65 – 66) also recalled
Lambert’s attempt to study randomness (see my § 6.1.3)3, and (1875,
pp. 177 – 179) applied Bienaymé’s test (§ 10.2-3) for investigating
whether the digits of the number π were random. He replaced its first
36 digits by signs plus and minus (for example, 3; 1; 4; 1 became – ;
+; –) and counted 21 changes in the new sequence. Comparing the
fractions 21/36 = 0.583 and [(2·36 + 1)/3·36] = 0.667, see the first of
the formulas (3), Cournot decided that the accordance was good
enough (?), but reasonably abstained from a final conclusion.
5) A mixture of distributions. Given the densities of separate groups
of ni, i = 1, 2, …, m, observations, Cournot (§ 81) proposed the
weighted mean density as their distribution. He had not specified the
differences between the densities, but in § 132 he indicated that they
might describe observations of different precision and in § 135 he
added, in a Note, that observational errors approximately followed the
[normal law]. I describe the attempts to modify the normal law made
by astronomers in § 10.8.4.
6) Dependence between the decisions of judges and/or jurors.
Cournot (1838; 1843, §§ 193 – 196 and 206 – 225) gave thought to
this issue. Suppose (§ 193) that in case of two judges the probabilities
of a correct verdict are s1 and s2. Then the probability that they agree
is

p = s1s2 + (1 – s1) (1 – s2) (5)

1 1
so that, if s1 = s2 = s > 1/2, s =  .
2 2 2 p 1

Statistical data provided the value of p; it should have obviously

exceeded 1/2. If the data made it also possible to ascertain s1 and s2,
and equations of the type of (5) will not be satisfied, it will be
necessary to conclude that the verdicts were not independent.
Cournot’s study was hardly successful in the practical sense, but it at
least testified to an attempt to investigate dependence between some
events.
7) A critical attitude towards statistics; a description of its aims and
applications. Cournot (§ 103) declared that statistics had blossomed
exuberantly and that [the society] should be on guard against its
premature and wrong applications which might discredit it for some
time and delay the time when it will underpin all the theories
concerning the “organization sociale”. Statistics, he (§ 105) continued,
should have its theory, rules, and principles, it ought to be applied to
natural sciences, to physical, social and political phenomena; its main
goal was (§ 106) to ascertain “the knowledge of the essence of
things”, to study the causes governing the phenomena of the physical
world and social life (§ 120). The theory of probability was applicable

165
164
to statistics (§ 113) and the “principe de Bernoulli” was its only
pertinent sound foundation (§ 115).
These statements were not at all unquestionable (§§ 6.2.1 and 9B).
Cournot, however, went further: the theory of probability might be
successfully applied in astronomy, and the “statistique des astres, if
such an association of words be permitted, will become a model for
every other statistics” (§ 145). He himself, however, statistically
studied the parameters of planetary and cometary orbits, but not the
starry heaven, and his statement was inaccurate: first (§ 10.8), by that
time statistics had begun to be applied in a number of branches of
natural sciences; second (Sheynin 1984a, § 6), stellar statistics had by
then already originated.
8) Explanation of known notions and issues. Cournot methodically
explained the notion of density (§§ 64 – 65) and the method of
calculating the density of a function of a (of two) [random variable(s)]
(§§ 73 – 74). He also described how is or should statistics be applied
in natural sciences and demography, discussed the treatment of data
when the probabilities of the studied events were variable, etc.
Taken as a whole, Cournot made a serious contribution to
theoretical statistics. Chuprov (1905/1960, p. 60), bearing in mind
mathematics as well as philosophy and economics, called him a man
of genius. Later he (1909/1959, p. 30) stated that Cournot was

One of the most profound thinkers of the 19th century, whom his
contemporaries failed to appreciate, and who rates ever higher in the
eyes of posterity.

Lastly, Chuprov (1925a/1926, p. 227) characterized the French

scientist as “the real founder of the modern philosophy of statistics”.
All this seems to be somewhat exaggerated and in any case I do not
agree with Chuprov’s opinion about Cournot’s “real substantiation”
and “canonical” proof of the LLN (1905/1960, p. 60; 1909/1959, pp.
166 – 168). Up to 1910, when he began corresponding with Markov,
Chuprov was rather far from mathematical statistics. He had not
remarked that Cournot did not even formulate that law, and that his
“Lemma”, as Chuprov called it [rare events do not happen, see
Cournot (1843, § 43) interpreted by Chuprov as “do not happen
often”], was not new at all, see my §§ 2.1.2, 2.2.2 and 3.2.2
concerning moral certainty and § 6.1.2 with regard to D’Alembert
who formulated the same proposition.
Cournot obviously never had anything in common with precise
measurements (observations) and his considerations about them are
hardly useful. He had not, but should have known that in 1817
Humboldt introduced isotherms (§ 10.8.3). He had not mentioned
Daniel Bernoulli’s study of 1766 of smallpox epidemics and his
description of the tontines was wrong.
10.4. Buniakovsky
Several European mathematicians had attempted to explicate the
theory of probability simpler than Laplace did. Lacroix (1816), the
mathematical level of whose book was not high, Cournot (§ 10.3),
whose main contribution was translated into German in 1849, and De

166
165
Morgan (1845) might be named here; concerning the last-metioned
author see however Note 1 to Chapter 8. In Russia, Buniakovsky
(1846) achieved the same aim; his treatise was the first comprehensive
Russian contribution so that P. B. Struve (1918, p. 1318) called him
“a Russian student of the French mathematical school”. I discuss the
main issues considered by him both in his main treatise and elsewhere,
see also Sheynin (1991b), this being a general description of
Buniakovsky’s work. A practically complete list of his contributions is
in Materialy (1917).
1) The theory of probability. In accordance with its state in those
times, Buniakovsky (1846, p. I) attributed it to applied mathematics.
He (Ibidem) also maintained that

The analysis of probabilities considers and quantitatively estimates

even such phenomena […] which, due to our ignorance, are not
subject to any suppositions.

This mistaken statement remained, however, useless: Buniakovsky

never attempted to apply it; furthermore, he (p. 364; 1866a, p. 24)
went back on his opinion.
2) Moral expectation (see § 6.1.1). Independently from Laplace,
Buniakovsky (1846, pp. 103 – 122) proved Daniel Bernoulli’s
conclusion that an equal distribution of a cargo on two ships increased
the moral expectation of the freightowner’s capital as compared with
transportation on a single ship. Later he (1880) considered the case of
unequal probabilities of the loss of each ship. Then, he (1866a, p. 154)
mentioned moral expectation when stating that the statistical studies
of the productive population and the children should be separated, and
he concluded with a general remark:

Anyone, who does not examine the meaning of the numbers, with
which he performs particular calculations, is not a mathematician.

For a long time statisticians shunned mathematics (end of § 10.7)

since they wrongly understood the essence of that science. And even
in the first half of the 20th century Soviet statisticians shunned it since
they correctly understood the ensuing danger to Marxism (Note 7 to
Chapter 15).
Buniakovsky passed over in silence Ostrogradsky’s attempt to
generalize the concept of moral expectation (§ 7.1-9).
3) Geometric probabilities (§ 6.1.4). Buniakovsky (1846, pp. 137 –
143) generalized the Buffon problem by considering the fall of the
needle on a system of congruent equilateral triangles. His geometric
reasoning was, however, complicated and his final answer, as Markov
(Treatise, 1900/1924, p. 270) maintained, was wrong. Markov himself
had been solving an even more generalized problem concerning a
system of congruent scalene triangles, but his own graph was no less
involved and, as it seems, no one has checked his solution.
Buniakovsky also investigated similar problems earlier (1837) and
remarked then that their solution, together with [statistical simulation],

167
166
might help to determine the values of special transcendental functions.
In the same connection, Laplace only mentioned the number π.
4) “Numerical” probabilities. Buniakovsky (1836; 1846, pp. 132 –
137) solved an elementary problem on the probability that a quadratic
equation with coefficients “randomly” taking different integral values
had real roots. Much more interesting are similar problems of later
origin, for example on the reducibility of fractions (Chebyshev, see
§ 13.2-8) or those concerning the set of real numbers.
5) A random walk. Buniakovsky (1846, pp. 143 – 147) calculated
the probability that a castle, standing on square A of a chessboard,
reached square B (possibly coinciding with A) in exactly x moves if its
movement was “uniformly” random. Before that, random walks (here,
it was a generalized random walk) had occurred only indirectly, when
studying a series of games of chance.
Buniakovsky’s problem was, however, elementary. The castle can
only be in two states,– it can reach B either in one move, or in two
moves; the case A ≡ B belongs to the latter, but might be isolated for
the sake of expediency. Buniakovsky divided the squares in three
groups: square A itself; 14 squares lying within reach of the very first
move from A; and the rest 49 squares and he formed and solved a
system of three pertinent difference equations for the number of cases
leading to success. It turned out that the mean probability (of its three
possible values) was equal to 1/64 and did not depend on x. He had
not interpreted his result, but properly indicated that it was also
possible to solve the problem in an elementary way, by direct
calculation. Note that the first n moves (n ≥ 1), if unsuccessful, do not
change anything, and this circumstance apparently explains the
situation.
6) Statistical control of quality. Buniakovsky (1846, Addendum;
1850) proposed to estimate probable military losses (actually, mean
losses, cf. beginning of § 11.2) in battle by sample data in each arm of
the engaged forces, – to use stratified sampling, as it is now called.
His study was hardly useful, the more so since he applied the
Bayesian approach assuming an equal prior probability of all possible
losses, but he (1846, pp. 468 – 469) also indicated that his findings
might facilitate the acceptance “of a very large number of articles and
supplies” only a fraction of which was actually examined.
Statistical control of quality was then still unknown although even
Huygens (§ 2.2.2) had solved a pertinent urn problem. Ostrogradsky
(1848), possibly following Buniakovsky4, picked up the same issue.
He stated, on p. 322, that the known solutions [of this problem] “sont
peu exactes et peu conformes aux principes de l’analyse des hasards”.
He did not elaborate, and his own main formula (p. 342) was
extremely involved (and hardly checked by anyone since).
Two other authors had preceded Buniakovsky who considered an
equivalent problem: Simpson (1740, Problem 6) and Öttinger (1843,
p. 231). Here is the former:

There is a given Number of each of several Sorts of Things […] as

(a) of the first Sort, (b) of the second, etc. put promiscuously together;
out of which a given Number (m) is to be taken, as it happens; To find

168
167
the probability that there shall come out precisely a given Number of
each Sort…

7) The history of the theory of probability. Buniakovsky was one of

the first (again, after Laplace) to consider this subject and a few of his
factual mistakes might well be overlooked. His predecessors were
Montucla (1802) and Lubbok & Drinkwater (1830). In his popular
writings Buniakovsky showed interest in history of mathematics in
general and in this field he possibly influenced to some extent both
Markov, see Sheynin (1989а, § 3), and the eminent historian of
mathematics, V. V. Bobynin (1849 – 1919).
8) Population statistics. Buniakovsky (1846, pp. 173 – 213)
described various methods of compiling mortality tables, studied the
statistical effect of a weakening or disappearance of some cause of
death (cf. § 6.2.3), calculated the mean and the probable durations of
marriages and associations and, following Laplace, solved several
other problems.
After 1846, Buniakovsky actively continued these investigations.
He compiled mortality tables for Russia’s Orthodox believers and
tables of their distribution by age (1866a; 1866b; 1874) and estimated
the number of Russian conscripts ten years in advance (1875b). No
one ever verified his forecast and the comments upon his tables
considerably varied. Bortkiewicz (1889; 1898b) sharply criticized
them, whereas Davidov (Ondar 1971), who, in 1886, published his
own study of mortality in Russia, noted a serious methodical mistake
in their compilation but expressed an opposite opinion. Finally,
Novoselsky (1916, pp. 54 – 55) mainly repeated Davidov’s criticism,
but indicated that Buniakovsy’s data were inaccurate and incomplete
(as Buniakovsky himself had repeatedly stressed) and called his tables
“a great step forward”:

A new period in the study of mortality in Russia started with […]

the demographic investigations […] made by Buniakovsky […] and
especially […] with his Essay (1866a). His contributions on
population statistics represent an outstanding and remarkable
phenomenon not only in our extremely poor demographic literature
but in the rich realm of foreign writings as well, and particularly of
his time. […] Due to the clearness, depth and nicety of his analysis,
Buniakovsky’s works fully retain their importance for the present day
…
His tables constitute a great step forward but do not represent
sufficiently well the picture of Russian mortality. This is explained by
the inaccuracy of the main initial materials, lack […] of many
necessary data and a defect in the very method of compiling these
tables.

On Buniakovsky’s method of compiling mortality tables see articles

on his method and on Mortality tables in Demografichesky (1985).
In 1848 Buniakovsky published a long newspaper article devoted to
a most important subject, to the dread of cholera. However, he likely
had not paid due attention to this work. Much later Enko (1889)

169
168
provided the first mathematical model of an epidemic (of measles). It
is now highly appreciated (Dietz 1988; Gani 2001) and it might be
regretted that Buniakovsky did not become interested in such issues.
9) Among other studies, I mention Buniakovsky’s solution of a
problem in the theory of random arrangements (1871) that, however,
hardly found application, and an interesting urn problem (1875a)
connected with partition of numbers. An urn contains n balls
numbered from 1 through n. All at once, m balls (m < n) are extracted;
determine the probability that the sum of the numbers drawn was
equal to s. This problem is due to Laurent (1873) who referred to
Euler (1748, Chapter 16).
The problem demanded from Buniakovsky the calculation of the
coefficient of tmxs in the development of

(1 + tx) (1 + tx2) … (1 + txn).

He solved this problem by means of an involved partial difference

equation and only for small values of m; he then provided a formula
for the transition from m to (m + 1). Laplace (§ 7.1-2) solved the same
problem in a different way.
10) Approximate stochastic summing and an automatic abacus.
Buniakovsky (1867) solved a few problems on the approximate
summing of the values of functions (e. g., square and cubic roots for
consecutive natural numbers), or air pressure during six months. He
had not stated the aim of this latter summing, but it is evident: for
deriving mean values.
Buniakovsky reported about his abacus at a sitting of the
mathematical and physical class of the Petersburg Academy, then
published an appropriate memoir (1867). There, he described the
application of such abacuses for treating observations of
meteorological elements. Prudnikov (1954, p. 81) suggested that
Chebyshev had arrived at the idea about the structure of his
arithmometer under the influence of that memoir.
11) From other subjects, apart from participation in compiling
explanatory dictionaries of Russian language, I name the solution of a
(barely useful) problem on the theory of random arrangements,
linguistics (see below), the LLN, treatment of observations, and
testimonies, elections and verdicts. The last-mentioned item deserves
notice because of later events. I (Sheynin 1989а, p. 340) noted that in
1912 Markov applied to the Most Holy Synod for excommunication
from the Russian orthodox Church. Explaining his request, he
mentioned, in part, his disagreement with Buniakovsky who had
stated that some events ought to be simply believed. Markov,
however, had not known the most clear statement made by the latter
(1866b, p. 4) to the effect that “truths cognized by revelation” ought to
be separated from everything else. See also § 14.1.
In a popular article Buniakovsky (1847) dealt in particular with
linguistics. He mentioned his previous unpublished work on this topic
and explained the aims of statistically studying linguistics.
Regrettably, nothing more is known about him here. On early
statistical work in this field see Knauer (1955).

170
169
 For several decades Buniakovsky’s treatise (1846) continued to
influence strongly the teaching of probability theory in Russia; in spite
of the work of previous Russian authors, he it was who originated the
real study of the theory beyond Western Europe.
Several authors expressed their opinion about Buniakovsky (1846).
Vasiliev (1921, p. 36) decided that

This thorough and clearly written book is one of the best from the
European mathematical literature on the theory of probability. It
much assisted the dissemination of the interest in this science among
Russian mathematicians and increased the significance of teaching
probability here as compared with the universities of other countries.

Indeed, Buniakovsky was vice-president of the Petersburg

Academy! See also in § 10.5 a statement about the instruction in
probability in Belgium.
Markov (1914b/1981, p. 162) considered Buniakovsky’s treatise
(1846) a beautiful work and Steklov (1924, p. 177), president of the
Russian Academy of Sciences, believed that for his time it was
complete and outstanding.
I am duty bound, however, to remark that Buniakovsky did not pay
attention to the work of Chebyshev; after 1846, he actually left
probability for statistics. Then, Buniakovsky devoted more than 60
pages to the treatment of observations, but had not thrown proper light
on the achievements of Gauss and his exposition was old-fashioned.
This is all the more regrettable since even Shiyanov (1836) did more
justice to Gauss (but still had not described the second justification of
the MLSq).
10.5. Quetelet
At the beginning of his scientific career Quetelet visited Paris and
met leading French scientists. Some authors indicated that he was
much indebted to Laplace but I think that the main inspiration to him
was Fourier (1821 – 1829).
Quetelet tirelessly treated statistical data and attempted to
standardize statistics on an international scale. He was co-author of
the first statistical reference book (Quetelet & Heuschling 1865) on
the population of Europe (including Russia) and the USA that
contained a critical study of the initial data; in 1853, he (1974, pp. 56
– 57) served as chairman of the Conférence maritime pour l’adoption
d’un système uniforme d’observation météorologiques à la mer and
the same year he organized the first International Statistical Congress.
K. Pearson (1914 – 1930, 1924, vol. 2, p. 420) praised Quetelet for
“organizing official statistics in Belgium and […] unifying
international statistics”. The latter (1846, p. 364) complained that
different states are apparently pleased to prevent any rapprochement
[of statistical materials].
Mouat (1885, p. 15) testified that about 1831 – 1833 Quetelet had
suggested “the formation of a Statistical Society in London”. It was
indeed established in 1834 and now called Royal Stat. Soc.
Quetelet popularized the theory of probability. His own writings on
this subject intended for the general reader are not really interesting,

171
170
but Mansion (1904, p. 3) claimed that there were “peu de pays” where
instruction in that discipline had been on a par with Belgian practice
which fact he attributed to the lasting influence of Quetelet.
Quetelet’s writings (1869; 1871) contain many dozens of pages
devoted to various measurements of the human body, of pulse and
respiration, to comparisons of weight and stature with age, etc. and he
extended the applicability of the [normal law] to this field. Following
Humboldt’s advice he (1870), introduced the term anthropometry and
thus curtailed the boundaries of anthropology. He was possibly
influenced by Babbage (1857), an avid collector of biological data. In
turn, Quetelet impressed Galton (1869, p. 26) who called him “the
greatest authority on vital and social statistics”. While discussing
Galton (1869), K. Pearson (1914 – 1930, vol. 2, 1924, p. 89)
declared:

We have here Galton’s first direct appeal to statistical method and

the text itself shows that [the English translation of Quetelet (1846)]
was Galton’s first introduction to the […] normal curve.

In those days the preliminary analysis of statistical materials was

extremely important first and foremost because of large systematic
corruptions, forgeries and incompleteness of data. Quetelet came to
understand that statistical documents were only probable and that, in
general, tout l’utilité of statistical calculations consisted in estimating
their trustworthiness (Quetelet & Heuschling 1865, p. LXV).
Quetelet (1846) left recommendations concerning the compilation
of questionnaires and the preliminary checking of the data; maintained
(p. 278) that too many subdivisions of the data was a charlatanisme
scientifique, and, what was then understandable, opposed sampling (p.
293). This contribution contained many reasonable statements. In
1850, apparently bearing in mind its English translation, Darwin
(1887, vol. 1, p. 341) noted:

How true is a remark […] by Quetelet, […] that no one knows in

disease what is the simple result of nothing being done, as a standard
with which to compare homeopathy, and all other such things.

Almost a null hypothesis! Quetelet (1846, p. 259) also declared that

The plants and the animals have remained as they were when they
left the hands of the Creator.

Lamarck was the first who attempted to construct a theory of

evolution, and Quetelet’s statement possibly testifies that his thoughts
had been more or less discussed. But Quetelet never mentioned
Lamarck or Wallace or Darwin. His attitude partly explains why the
statistical study of the evolution of species had begun comparatively
late (and only in the Biometric school). I note that Knapp (1872b),
while discussing Darwin’s ideas, had not mentioned randomness and
said nothing about a statistical study of biological problems.

172
171
 Quetelet collected and systematized meteorological observations5
and described the tendency of the weather to persist by elements of the
theory of runs, see § 10.8.3. Keeping to the tradition of political
arithmetic (§ 2.1.4), he discussed the level of postal charges (1869, t.
1, pp. 173 and 422) and rail fares (1846, p. 353) and recommended to
study statistically the changes brought about by the construction of
telegraph lines and railways (1869, t. 1, p. 419). His special
investigation (1836, t. 2, p. 313; first noted in 1832) was a quantitative
description of the changes in the probabilities of conviction of the
defendants depending on their personality (sex, age, education) and
Yule (1900/1971, pp. 30 – 32) favourably, on the whole, commented
upon his work as the first attempt to measure dependence
(association).
Quetelet is best remembered for the introduction of the Average
man (1832a, p. 4; 1832b, p. 1; 1848a, p. 38), inclinations to crime
(1832b, p. 17; 1836, t. 2, p. 171 and elsewhere) and marriage (1848a,
p. 77; 1848b, p. 38), – actually, the appropriate probabilities, – and
statements about the constancy of crime (1829, pp. 28 and 35 and
many other sources). In spite of his shortcomings (below), the two
last-mentioned items characterized Quetelet as the originator of moral
statistics with Süssmilch as his predecessor.
The Average man, as he (1848a, p. 38) thought, had mean physical
and moral and intellectual features, was the alleged type of the nation
and even of the entire mankind. And he (1848a, pp. 91 – 92) properly
related the mean inclinations to the Average man. Reasonable
objections were levelled against this concept to which I add that
Quetelet had not specified the notion of average as applied here.
Sometimes he had in mind the arithmetic mean, in other cases (1848a,
p. 45), however, it was the median, and he (1846, p. 216) only
mentioned the Poisson LLN in connection with the mean human
stature. Cournot (1843, p. 143) stated that the Average man was
physiologically impossible (the averages of the various parts of the
human body [and of weight and height] were inconsistent one with
another), and Bertrand (1888a, p. XLIII) ridiculed Quetelet:

In the body of the average man, the Belgian author placed an

average soul. [The average man] has no passions or vices [wrong], he
is neither insane nor wise, neither ignorant nor learned. […] [He is]
mediocre in every sense. After having eaten for thirty-eight years an
average ration of a healthy soldier, he has to die not of old age, but of
an average disease that statistics discovers in him.

Nevertheless, the Average man is useful even now at least as an

average producer and consumer; Fréchet (1949) replaced him by a
closely related “typical” man.
Quetelet (1848a, p. 82; 1869, t. 2, p. 327) indicated that the real
inclination to crime of a given person might well differ considerably
from the apparent mean tendency. It seems, however, that he had not
sufficiently stressed these points; a noted statistician (Rümelin 1867,
p. 25) forcibly denied any criminal tendency in himself.

173
172
 Chuprov (1909/1959, p. 23) neatly summed up the arguments of
Quetelet’s followers and opponents:

Their [his worshippers’ zealous beyond reasoning] naïve

admiration for “statistical laws”; their idolizing of stable statistical
figures; and their absurd teaching that regarded everyone as
possessing the same “mean inclinations” to crime, suicide and
marriage, undoubtedly provoked protests. Regrettably, however, the
protests were hardly made in a scientific manner.

Quetelet (1836, t. 1, p. 10) declared that the [relative] number of

crimes was constant, and that

Each social regime presupposes […] a certain number and a

certain order of crimes, these being merely the necessary
consequences of its organization.

However, he had not justified his statement by statistical data. The

alleged constancy did not take place (Rehnisch 1876): Quetelet had
not studied criminal statistics attentively enough. Then, constancy of
crime could only happen under constant social conditions, but this
consideration had only indirectly followed from his statements.
A special point concerns laws of distribution. Quetelet (1848a, p.
80; 1869, t. 2, pp. 304 and 347) noticed that the curves of the
inclinations to crime and to marriage plotted against ages were
exceedingly asymmetric. He (1846, pp. 168 and 412 – 424) also knew
that asymmetric densities occurred in meteorology and he (1848a, p.
viii) introduced a mysterious “loi des causes accidentelles” whose
curve could be asymmetric (1853, p. 57)! And still he (1853)
explained this fact by special causes and anomalies.
In short, he had revealed here and elsewhere, see above and my
paper (1986a), his general attitude which Knapp (1872a, p. 124)
explained by his “spirit, rich in ideas, but unmethodical and therefore
unphilosophical”. His was a polite appraisal indeed; in actual fact,
Quetelet often acted in a happy-go-lucky manner.
Nevertheless, Quetelet had been the central figure of statistics in the
mid-19th century. Freudenthal (1966, p. 7) correctly concluded that
before Quetelet there were statistical bureaux and statisticians, but no
statistics [as a discipline].

10.6. Galton
Being influenced by his cousin, Darwin, Galton began to study the
heredity of talent and published an important treatise (1869) on that
subject; incidentally, he introduced the term eugenics. In a letter of
1861 Darwin (1903, p. 181) favourably mentioned it. He (1876, p. 15)
also asked Galton to examine his investigation of the advantages of
cross-fertilization as compared with spontaneous pollination. Galton
(Ibidem) compared the two processes with regard to their
characteristics and, in particular, to the ordered heights of the
seedlings. In the latter instance, it occurred that the signs of almost all

174
173
the differences between the corresponding heights coincided. See a
similar study by Seidel in § 10.8.1.
Galton (1863) devised an expedient system of symbols for weather
charts and immediately discovered the existence of previously
unknown anticyclones. From the point of view of statistics, he had
thus reasonably studied his initial data. Galton (K. Pearson 1914 –
1930, vol. 2, Chapter 12) also invented composite photographs of
kindred persons (of people of a certain nationality or occupation, or
criminals), all of them taken on the same film with an appropriately
shorter exposure. In any case, his innovation was much more justified
than Quetelet’s Average man.
Galton(1892) became the main inventor of fingerprinting. Because
of its reliability, it did not demand statistical analysis and superseded
the previous system of identification developed by Alph. Bertillon
(1893). This latter procedure was partially based on anthropometry
and made use of from the 1890s to the beginning of the 20th century.
Another of Galton’s invention (1877) was the so-called quincunx, a
device for visually demonstrating the appearance of the normal
distribution as the limiting case of the binomial law (Stigler 1986, pp.
275 –281). A special feature of that device was that it showed that the
normal law was stable. Galton’s main statistical merit consisted,
however, in the introduction of the notions of regression and
correlation. The development of correlation theory became one of the
aims of the Biometric school (§ 14.2), and Galton’s close relations
with Pearson were an important cause of its successes.
Recall (§§ 1.1.1 and 1.1.3) that reasoning in the spirit of qualitative
correlation was not foreign to ancient scholars which was in
conformity with the qualitative nature of the science of those days.
And what about modernity? In the 1870s, several scientists (C.
Meldrum, in 1872 and 1875; J. N. Lockyer, in 1873; H. F. Blanford, in
1880, see Sheynin (1984a, p. 160)) took notice of the dependence
between solar activity and elements of terrestrial magnetism and on
meteorological phenomena but not a word did they say about
developing a pertinent quantitative theory. And even though Seidel, in
1865 – 1866 (§ 10.8.1), quantitatively studied the dependence between
two, and then three factors, he did not hint at generalizing his findings.
Galton was meritorious indeed! For the sake of comprehensiveness I
repeat (Note 4 to Chapter 9) that in 1912 Kapteyn provided an
“astronomical” version of the correlation coefficient.
10.7. Statistics
Here, I discuss the situation in statistics in the 19th century. Related
material is in §§ 6.2 and 10.5.
The Staatswissenschaft held its ground for many decades. In
France, Delambre (1819, p. LXVII) argued that statistics was hardly
ever engaged in discussions or conjectures and did not aim at
perfecting theories, and that political arithmetic ought to be
“distinguished” from it. Under statistics he understood geodetic,
meteorological and medical data, mineralogical descriptions and even
art expositions. I believe however that the two last-mentioned items
were rather soon excluded from such lists6.

175
174
 The newly established London Statistical Society declared that
statistics “does not discuss causes, nor reason upon probable effects”
(Anonymous 1839, p. 1). True, they denied that “the statist [!] rejects
all deductions, or that statistics consists merely of columns of figures”
and stated that “all conclusions shall be drawn from well-attested data
and shall admit of mathematical demonstration”. This announcement
was thus ambiguous; the Society attempted to adhere to its former
statement, but in vain. Anyway, Woolhouse (1873, p. 39) testified that

These absurd restrictions have been necessarily disregarded in […]

numerous papers.

Indeed, that statistics should explain the present state of a nation by

considering its previous states was declared a century before that
(Gatterer 1775, p. 15). And the very title of Dufau (1840) called
statistics “The theory of studying the laws according to which the
social events are developing”7.
During the 19th century the importance of statistics had been
considerably increasing. Graunt (1662/1939, p. 79) was not sure
whether his work would be “necessary to many, or fit for others, than
the Sovereign, and his chief Ministers …” and the classical
investigations of the sex ratio at birth (§§ 2.2.4, 3.3.4, 4.4, 6.1.1) had
not found direct applications. However, by the mid-19th century it
became important to foresee how various transformations will
influence society and Quetelet (§ 10.5) repeatedly stressed this point.
Then, at the end of the 19th century censuses of population, answering
an ever widening range of questions, began to be carried out in
various countries. It is nevertheless instructive to compare the
situation at that time with what is happening nowadays8.
1) Public opinion was not yet studied, nor was the quality of mass
production checked by statistical methods, cf. § 10.4-6.
2) Sampling had been considered doubtful9. Cournot (1843) passed
it over in silence and Laplace’s determination of the population of
France based on sampling (§ 7.1-5) was largely forgotten even in spite
of the inexorable increase in statistical materials. Indeed, already the
beginning of the century witnessed “legions” of new data (Lueder
1812, p. 9) and the tendency to amass sometimes useless or unreliable
data revealed itself in various branches of natural sciences (§ 10.8).
Note however that the need for observations (in natural science?)
increases with knowledge (Descartes 1637/2012, p. 63).
Quetelet (§ 10.5) opposed sampling. Even much later Bortkiewicz
(1904, p. 825) and Czuber (1921, p. 13) called sampling “conjectural
calculation” and Chuprov (1912) had to defend that procedure
vigorously.
3) The development of the correlation theory began at the end of the
19th century (§§ 10.6, 15.2), but even much later Kaufman (1922, p.
152) declared that

The so-called method of correlation adds nothing essential to the

results of elementary analysis.

176
175
See, however, § 14.1-4.
4) Apart from the error theory the variance began to be applied in
statistics only after Lexis (§ 15.1), but even later Bortkiewicz (1894 –
1896, Bd. 10, pp. 353 – 354) stated that the study of precision was an
accessory goal, a luxury, and that the statistical flair was much more
important, cf. the opinion of Gauss in § 9A.5-1. This point of view
had perhaps been caused by the presence of large systematic
corruptions in the initial materials.
5) Not just a flair, but a preliminary data analysis (which, however,
does not call off the definitive estimation of the plausibility of the
final results and which received general recognition only a few
decades ago) is necessary, and should be the beginning of the
statistician’s work. Splendid examples of such analysis had occurred
much earlier and to these I attribute the introduction of contour lines
(Halley, in 1701, see § 2.1.4, drew lines of equal magnetic
declinations over North Atlantic, also see § 10.8.3, the discoveries
made by Humboldt in § 10.8.3 and by Galton in § 10.6).
6) Econometrics originated only in the 1930s.
I list now the difficulties, real and imaginary, of applying the theory
of probability to statistics.
7) The absence of “equally possible” cases whose existence is
necessary for understanding the classical notion of probability.
Statisticians repeatedly mentioned this cause, also see § 3.2.3. True,
Cournot (§ 10.3-7 and -8) explained that equipossibility was not
necessary (and, in the first place, mentioned the “Bernoulli
principle”), but his advice was hardly heard. Lexis (report of
1874/1903, pp. 241 – 242; 1886, pp. 436 – 437; 1913, p. 2091) also
cited equipossibility. In the second case, in a paper devoted to the
application of probability theory to statistics, he even added that the
introduction of equipossibility led to the subjectivity of the theory of
probability. Elsewhere, however, Lexis reasoned differently; he had
no integral viewpoint. Thus, statistics is mainly based on the theory of
probability (1877, p. 5); if the statistical probability tends to its
theoretical counterpart, equally possible cases might be assumed
(Ibidem, p. 17); and the “pattern” of the theory of probability is the
highest scientific form in which statistics might be expressed
(1874/1903, p. 241).
8) Disturbance of the constancy of the probability of the studied
event and/or of the independence of trials. I repeat (§ 3.2-3) that
before Lexis statisticians had only recognized the Bernoulli trials; and
even much later Kaufman (1922/1928, pp. 103 – 104) declared that
the theory of probability was applicable only to these trials, and, for
that matter, only in the presence of equally possible cases. He
mentioned several allegedly likeminded authors including Markov
and Yule, but did not supply the exact references and I am inclined to
believe that these authors wished to investigate whether or not the
given statistical trials were Bernoullian. As to the equally possible
cases, see Item 7 above. After all, Kaufman’s opinion about
correlation theory (Item 3) can be understood, but this time he proved
himself hopelessly lagging behind life.

177
176
 9) The abstract nature of the (not yet axiomatized) theory of
probability. As I noted in § 6.3.3, the history of mathematics testifies
that the more abstract it became, the wider had been the range of its
applicability. Nevertheless, statisticians had not expected any help
from the theory of probability. Block (1878/1886, p. 134) thought that
it was too abstract and should not be applied “too often”, and Knapp
(1872, p. 115) called it difficult and hardly useful beyond the sphere
of games of chance and insurance.
There also, on pp. 116 – 117, he added a sound statement:

Placing coloured balls in Laplacean urns is not enough for shaking

scientific statistics out of them.

For his part, Chuprov (1922/1960, pp. 415 – 416) formulated a

related opinion:

Only statisticians armed with mathematics can defeat

mathematicians playing with statistics.

Statisticians had not been armed and did not trust mathematicians.
Even in 1911 G. von Mayr, an eminent statistician of the old school
stated that mathematical formulas were unnecessary for statistics and
then privately confessed that he cannot endure mathematics
(Bortkevich & Chuprov 2005, Letter 109). And mathematicians, if not
playing with statistics, had not always understood statistics (§ 8.9-2).
It is not amiss to mention here the pioneer attempt to create
mathematical statistics (Wittstein 1867). He compared the situation in
statistics with the childhood of astronomy and stressed that statistics
(and especially population statistics) needed a Tycho and a Kepler to
proceed from reliable observations to regularities. Specifically, he
noted that statisticians did not understand the essence of probability
theory and never estimated the precision of the results obtained. The
term mathematical statistics is apparently due to him.
10.8. Statistics and Natural Sciences
In the 19th century, the statistical method gave rise to a number of
disciplines and I discuss the relevant situation in several branches of
natural sciences. First, I cite the opponents of that method; as
described in § 10.5, their stance can be explained by ignorance, by
attaching mean indicators to individuals. Thus, Comte (1830 – 1842, t.
3/1893, No. 40, p. 329): the use of statistics in medicine is a “profonde
dégénération directe de l’art médicale”. Also see similar statements in
§ 8.9.
To begin in earnest, however, I note the existence of the so-called
numerical method usually attributed to the French physician Louis
(1825) who introduced it by calculating the frequencies of the
symptoms of various diseases so as to facilitate diagnosing. No
stochastic considerations or estimations of the reliability of the
conclusions were involved. Louis (pp. xvii – xviii) even thought that,
given observations, any physician ought to make the same conclusion.
His method remained in vogue for a few decades.

178
177
 He and his adherents attempted to replace qualitative descriptions
by directly obtained statistical data (cf. Petty’s statement in § 2.1.4).
Bouillaud (1836), who inserted numerous passages from Laplace’s
Essai philosophique (1814) in his book, favourably described the
numerical method and called it a supplement to other methods (pp.
190 – 191), and (p. 187) added only a few words about the “calcul
approximatif ou des probabilités”. That method, as he stated, was
almost always the only means for generalizing the results obtained;
and the advantages of this “kind of calculus” are such that its
discussion was not necessary. Also see Proctor’s statement in § 10.8.4.
Bouillaud (pp. 186 – 187) thought that the calculus of probability was
approximate, but almost always the only means for generalizing the
obtained results, whereas medicine, if only based on probability, will
remain a game of chance of sorts since physicians ought to consider
the personality of their patients. And he also stated that medical
statistics was yet in its cradle, but is still applied with some success
and will considerably develop, cf. § 10.8.1.
And here is D’Alembert (1759/1821, p. 163):

Systematic medicine seems to me […] a real scourge of mankind.

Observations, numerous and detailed, sensibly corresponding one to
another, – this is […] to what reasoning in medicine ought to reduce
itself.

On D’Alembert see also Note 7 to Chapter 6. Another author who

advocated compilation of medical observations is mentioned at the
beginning of § 6.2.3.
Unlike Bouillaud, Gavarret did not sidestep this issue (§ 8.9.2), and
he (1840, p. x) reasonably remarked that the numerical method was
not in itself scientific and was not based on “general philosophy”.
There also existed a wrong opinion (D’Amador 1837, p. 12)
attributing the numerical method to probability theory. Guy (1852, p.
803), who was an eminent physician and a statistician of note, applied
the term “numerical or statistical method”, but only to prevent
statistics from being confused with Staatswissenschaft (§ 6.2.1). There
also (pp. 801 – 802) he claimed that medicine had a special relation
with statistics:

There is no science which has not sooner or later discovered the

absolute necessity of resorting to figures as measures and standards
of comparison; nor is there any reason why physiology and medicine
should claim exemption. […] On the contrary, they […] may hope to
derive the greatest benefit from the use of numbers. […] Without
statistics a science bears to true science the same sort of relation
which tradition bears to history.

However, the numerical method (not necessarily in medicine) can

be traced back to the 18th century (see below and § 6.2.3) and my
description (§§ 10.8.1 – 10.8.4) shows that it continued in existence
for many decades. Furthermore, empiricism had been a feature of the

179
178
Biometric school (§ 15.2). Greenwood (1936, p. 139) indirectly
praised it, perhaps excessively:

Some heart-breaking therapeutic disappointments in the history of

tuberculosis and cancer would have been avoided if the method of
Louis had been not merely praised but generally used during the last
fifty years.

In statistics proper, Fourier’s fundamental Recherches (1821 –

1829) concerning Paris and the Département de la Seine might be here
mentioned. This contribution almost exclusively consisted of
statistical tables with data on demography, industry, commerce,
agriculture and meteorology. True, empiricism was not sufficient even
for compiling tables. Then, the abundance of materials led to the
wrong idea that a mass of heterogeneous data was better than a small
amount of reliable observations (§ 10.8.1).
In actual fact, the numerical method originated with Anchersen
when statisticians have begun to describe states in a tabular form (and
thus facilitated the use of numbers), see § 6.2.1. Recall (§ 2.1.4),
moreover, that Leibniz recommended compilation of Staatstafeln.
10.8.1. Medicine. See also § 8.9.2. In 1835, Double et al (§ 8.9)
indicated that statistics might be applied in medicine. Surgery
occurred to be the first branch of medicine to justify their opinion.
Already in 1839 there appeared a (not really convincing) statistical
study of the amputation of limbs. Soon afterwards physicians learned
that the new procedure, anaesthesia, could cause complications, and
began to compare statistically the results of amputation made with and
without using it. The first such investigation (J. Y. Simpson 1847 –
1848/1871, p. 102) was, however, unfortunate: its author had
attempted to obtain reliable results by issuing from materials
pertaining to several English hospitals during 1794 – 1839:

The data I have adduced […] have been objected to on the ground
that they are collected from too many different hospitals, and too
many sources. But, […] I believe all our highest statistical authorities
will hold that this very circumstance renders them more, instead of
less, trustworthy.

I ought to add, however, that Simpson (Ibidem, p. 93) stated that

only a statistical investigation could estimate the ensuing danger.
He (1869 – 1870/1871, title of contribution) also coined the term
Hospitalism which is still in vogue. He compared mortality from
amputations made in various hospitals and reasonably concluded, on
the strength of its monotonous behaviour, that mortality increases with
the number of beds; actually (p. 399), because of worsening of
ventilation and decrease of air space per patient. Virchow (1868 –
1869/1879, Bd. 2, p. 21) noted that the pernicious influence of bad air
had been known about six hundred years previously.
Justification of conclusions based on monotonic behaviour of some
indications was not restricted to medicine; Quetelet’s table of

180
179
probabilities of conviction of defendants depending on their
personality showed a monotonous increase of those probabilities (cf.
§ 10.5).
At about the same time Pirogov introduced anaesthesia in military
surgery and began to compare the merits of the conservative treatment
of the wounded versus amputation. Much later he (1864, p. 690)
called his time “transitional”:

Statistics shook the sacred principles of the old school, whose views
had prevailed during the first decades of this century, – and we ought
to recognize it,– but it had not established its own principles.

Pirogov (1849, p. 6) reasonably believed that the application of

statistics in surgery was in “complete agreement” with the latter
because surgical diseases depended incomparably less on individual
influences. However, he repeatedly indicated that medical statistics
was unreliable. Thus (1864/1865 – 1866/1920, p. 20):

Even a slightest oversight, inaccuracy or arbitrariness makes [the

data] far less reliable than the figures founded only on a general
impression with which one is left after a mere but sensible observation
of cases.

Later he (1879/1882, p. 40) singled out an important pertinent

cause:

Extremely different circumstances separate the entire mass of

information in too insignificant and very dissimilar groups which does
not allow any correct conclusion about the worth of a certain
amputation. [In 1849] I […] had not yet known all the false ways to
which the number sometimes leads us.

In essence, he advocated attentive allowance for all circumstances

and minimal statistical technique which was in accordance with his
time and especially so with the originating military surgery (of which
he was the founder).
Pirogov was convinced in the existence of regularities in mass
phenomena. Thus (1850 – 1855/1961, p. 382), each epidemic disease
as well as each “considerable” operation had a constant mortality rate,
whereas war was a “traumatic epidemic” (1879/1882, p. 295). This
latter statement apparently meant that under war conditions the
sickness rate and mortality from wounds obeyed statistical laws. Then
(1854, p. 2), the skill of the physicians [but not of witch doctors]
hardly influenced the total result of the treatment of many patients.
Here is his highly relevant opinion (1871, pp. 48 – 49):

On what does the success of treatment or the decrease of mortality

in the army depend? Surely not on therapy and surgery by themselves.
Without an efficient administration [of medicine] little can the masses
expect from therapy and surgery even in time of peace, much less
during such a catastrophe as war.

181
180
 Note finally Pirogov’s possibly correct statement (1864, pp. 5 – 6):
without the not yet existing doctrine of individuality, real progress in
medical statistics is impossible.
Pirogov participated in the Crimean war, in which Florence
Nightingale, on the other side, showed her worth both as a medical
nurse and a statistician, cf. Pearson’s relevant statement in § 2.2.3.
She would have wholeheartedly approved of Pirogov’s conclusion
(above) concerning the success of treatment.
Such new disciplines as epidemiology and public hygiene, the
forerunner of ecology, appeared within medicine in the 19th century. I
discussed the inoculation of smallpox in § 6.2.3 and mentioned
Enko’s essential finding at the end of § 10.4. In 1866, Farr
(Brownlee 1915) preceded Enko; his study of cattle plague only
methodically belonged to epidemiology, and, interestingly enough,
Brownlee published his note in a medical journal. Farr indicated that
he had also investigated the visitations of cholera and diphtheria of
1849 and 1857 – 1859 respectively.
Here is his reasoning. Denote the number of attacks of the plague
during a period of four weeks by s. He noted that the third differences
of lns were constant, so that

s = Cexp{δt[t + m)2+ n]}, C > 0, δ < 0.

It was Brownlee who supplied this formula because Farr was unable
to insert it in his newspaper letter. Farr’s calculated values of s did not
agree with actual figures, but at least he correctly predicted a rapid
decline of the epidemic.
It seems that epidemiology was properly born when cholera
epidemics had been ravaging Europe. The English physician Snow
(1855) compared mortality from cholera for two groups of the
population of London, – for those whose drinking water was either
purified or not. He ascertained that purification decreased mortality by
eight times, and he thus discovered how did cholera epidemics spread,
and proved the essential applicability of the first stage of the statistical
method (§ 0.4). Pettenkofer (1886 – 1887) published a monstrous
collection of statistical materials pertaining to cholera, but he was
unable to process them. He (1865, p. 329) stressed that no cholera
epidemic was possible at a certain moment without a local
“disposition” to it and he attached special importance to the level of
subsoil water. His view does not contradict modern ideas about the
necessary threshold values. However, Pettenkofer did not believe in
contemporary bacteriological studies and opposed Snow. For an
estimate of his views see Winslow (1943/1967, p. 335).
Seidel (1865) arranged the years 1856 – 1864 in the decreasing
order of the first, and then of the second series of numbers describing
two phenomena. The conformity between the two series was, in his
opinion, striking so that Seidel thus, like Galton later on, applied rank
correlation.
He (1865 – 1866) investigated the dependence of the monthly cases
of typhoid fever on the level of subsoil water, and then on both that

182
181
level and the rainfall. It occurred that the signs of the deviations of
these figures from their mean yearly values coincided twice more
often than not and Seidel quantitatively (although indirectly and with
loss of information) estimated the significance of the studied
connections. His work remained, however, completely forgotten and
Weiling (1975) was likely the first to recall it.
Already Leibniz (§ 2.1.4) recommended to collect and apply
information concerning a wide range of issues, which, as I add now,
pertained to public hygiene. Condorcet (1795/1988, pp. 316 and 320)
described the aims of “mathématique sociale” [social statistics] and
mentioned the study of the influence of temperature, climate,
properties of soil, food and general habits on the ratio of men and
women, birth-rate, mortality and number of marriages. Much later, M.
Lévy (1844) considered the influence of atmosphere, water and
climate as well as of the suitable type of clothes and appropriate food
on man.
From its origin in the mid-19th century, public hygiene began
statistically studying a large number of problems, especially those
caused by the Industrial Revolution in England and, in particular, by
the great infant mortality. Thus, in Liverpool only 2/3 of the children
of gentry and professional persons lived to the age of five years
(Chadwick 1842/1965, p. 228).
Pettenkofer (1873) estimated the financial loss of the population of
Munich ensuing from such diseases as typhoid fever and his booklet
can be attributed to this discipline. In Russia his student Erismann
(1887) published a contribution on sanitary statistics.
At the turn of the 19th century Jenner introduced smallpox
vaccination instead of inoculation. It had not, however, preclude the
need for solving statistical problems about the estimation of various
possible versions of vaccination and inoculation was not ruled out at
once. See also § 6.2.3.
10.8.2. Biology. The attempts to connect the appearance of leaves,
flowers and fruits on plants of a given species with the sums of mean
daily temperatures began in the 18th century (Réaumur 1738) and
Quetelet (1846, p. 242) proposed to replace those sums by the sums
of squares, but he was still unable to compare both procedures
quantitatively. Also in the 19th century, vast statistical materials
describing the life of plants were published (Aug. P. DeCandolle
1832), and Babbage (1857) compiled a statistical questionnaire for
the class of mammalia. I mentioned his similar work in § 10.5. In
Russia, Baer (1860 – 1875) with associates conducted a large-scale
statistical investigation of fishing.
Humboldt created the geography of plants (Humboldt & Bonpland
1815; Humboldt 1816) which was based on collection and estimation
of statistical data. Alph. DeCandolle (1855, t. 1, p. vi) however also
mentioned Linné, his own father Aug. DeCandolle and Brown;
Darwin (1903, vol. 2, p. 26, letter of 1881) only mentioned Humboldt.
Darwin had to study various statistical problems, for example on
cross-fertilization of plants (§ 10.6), the life of earthworms (§ 12-2)
and on the inheritance of a rare deformity in humans (1868/1885, vol.
1, p. 449). In the last-mentioned case Stokes provided the solution

183
182
(apparently by applying the Poisson distribution) at his request.
Statistical tables and summaries with qualitative commentaries occur
in a number of Darwin’s writings and he also collected statistical data.
Being the main author of the hypothesis of the origin of species, he
made use of such terms as variation and natural selection without
defining any of them. And, when reasoning about randomness, he
understood it in differing ways. In the problem concerning the
deformity Darwin decided that it was not random (not merely
possible, cf. Kolmogorov’s statement in § 1.1). In two other cases in
which he discussed the hypothesis of evolution he understood
randomness as ignorance of causes (1859/1958, p. 128), cf. Laplace
(§ 7.3), and, in 1881, as lack of purpose (1903, p. 395), cf. Aristotle
(§ 1.1.1). It is also remarkable that Darwin (1859/1958, p. 77) actually
described randomness as the effect of complicated causes, cf.
Poincaré (§ 11.2-9):

Throw up a handful of feathers, and all fall to the ground according

to definite laws; but how simple is the problem where each shall fall
compared with problems in the evolution of species.

The stochastic essence of the evolution hypothesis was evident both

for its partisans and the opponents; Boltzmann, however, was an
exception (§ 10.8.5-4).
I reconstruct now Darwin’s model of evolution. Introduce an n-
dimensional (possibly with n = ∞) system of coordinates, the body
parameters of individuals belonging to a given species (males and
females should, however, be treated separately), and the appropriate
Euclidean space with the usual definition of distances between its
points. At moment tm each individual is some point of that space and
the same takes place at moment tm+1 for the individuals of the next
generation. Because of the “vertical” variation, these, however, will
occupy somewhat different positions. Introduce in addition point (or
subspace) V, corresponding to the optimal conditions for the existence
of the species, then its evolution will be represented by a discrete
stochastic process of the approximation of the individuals to V (which,
however, moves in accordance with the changes in the external world)
and the set of individuals of a given generation constitutes the
appropriate realization of the process. Probabilities describing the
process (as well as estimates of the influence of habits, instincts, etc.)
are required for the sake of definiteness, but they are of course
lacking.
The main mathematical argument against Darwin’s hypothesis was
that a purposeful evolution under “uniform” randomness was
impossible; see end of § 6.1.3 with regard to the difficulties of
generalizing the notion of randomness. Only Mendel’s contributions
(1866; 1866 – 1873, publ. 1905), forgotten until the beginning of the
20th century, allowed to answer such criticisms. True, great many
objections and problems still remain, but at the very least Darwin had
transformed biology as a science. In addition, his work was
responsible for the appearance of the Biometric school (§ 15.2).

184
183
 An essentially new stage in the development of the Darwinian ideas
had occurred at the end of the 18th century (De Vries 1905) and
somewhat later (Johannsen 1922/1929). De Vries stressed the
importance of sports (although did not explain their relation to
mutations, a term which he himself introduced somewhat earlier)
which considerably strengthened the theory of evolution.
Andersson (1929), who briefly discussed the work of Johannsen,
quoted him, regrettably without indicating the source:

The science of evolution has turned into an Augeas stables which

really ought to be mucked.

And here is Johannsen himself.

p. 355. Galton’s statistics of heredity were quite erroneous – a

combination between collective measurements of unsorted rough
material and biological analysis of the real units of certain
populations.
p. 356. These statistical researches in heredity are naturally of
importance from the sociological point of view and of practical
interest in insurance calculations and so on. But they do not reach the
biological problems of heredity.
p. 357. The [genotypic] differences are discontinuous […] rather
contrary to Darwinism.
p. 359. We cannot do without statistics!

From the mathematical point of view, Mendel did nothing except

for an elementary application of the binomial distribution, but his
memoir marked the origin of a new direction in biology, of genetics,
and provided an example of a fundamental finding achieved by
elementary means. Mendel had based his conclusions on experiments,
and these became the object of many discussions with regard to his
initial data and to his subjective and objective honesty. Such scholars
as Fisher (1936) and van der Waerden (1968) participated in the
debates, and finally all doubts have possibly blown over the more so
since Mendel’s life and his meteorological observations and
investigations unquestionably testify in his favour. It is thought that
Mendel was born in a mixed Czech-German family; actually,
however, he was German, and in 1945 – 1946 the descendants of his
relatives were driven out of the then Czechoslovakia10.
10.8.3. Meteorology. The material pertaining to the 18th century is
in § 6.2.4. Humboldt (1818, p. 190) maintained that

To discover the laws of nature [in meteorology] we ought to

determine the mean state of the atmosphere and the constant type[s] of
its variations before examining the causes of the local perturbations11.

He (1845 – 1862, Bd. 1, pp. 18 and 72; Bd. 3, p. 288) conditioned

the investigation of natural phenomena by examination of mean states.
In the latter case he mentioned “the sole decisive method [in natural
sciences], that of the mean numbers” which (1845, Bd. 1, p. 82)

185
184
“show us the constancy in the changes”. He himself (1817, p. 466)
introduced isotherms and climatic belts (known to ancient scholars
who had only possessed qualitative knowledge of temperature). He
thus separated climatology from meteorology and he also noted the
existence of local corruptions of the temperature.
Much later he (1845 – 1862, Bd. 4, p. 59) added that he had
borrowed the idea of contour lines from Halley (§ 2.1.4) [and had
therefore also applied a splendid particular instance of exploratory
data analysis].
Humboldt (1817, p. 532) also recommended the application of
contour lines for winter and summer. It is somewhat strange that,
when offering a definition of climate, he (1831, p. 404) had not
directly linked it with mean states but later scholars have formulated
this tie ever more explicitly (Körber 1959, p. 296). Chuprov
(1922b/1960, p. 151), for example, had identified climate with a
system of certain mean values.
Humboldt (1843, t. 1, p. 83) also formulated a statement which
could have shown that the time for passing from mean values to
distributions had not yet come:

Exact sciences only developed to the extent that they considered

physical phenomena in their ensemble and gradually ceased to attach
too much importance either to culminating points isolated among a
line of facts, or to the extreme temperatures during some days of the
year.

He (Ibidem, p. 405) provided a splendid example: large nuggets

cannot be a sure indication of the mean content of gold in their
vicinity.
Köppen (1874, p. 3) dwelt on the same subject. He believed that
the “introduction of the arithmetical mean in meteorology was the
most important step”, but that it was not sufficient all by itself.
Dove (1837, p. 122) came out against “the domination” of mean
values; largely following Humboldt (see above), he (1839, p. 285)
formulated the aims of meteorology as the “determination of mean
values [of temperature], derivation of the laws of [its] periodic
changes and indication of rules for [determining its] irregular
changes”, and he attached no less importance to the spatial scatter of
the temperature. Later Dove (1850, p. 198) introduced monthly
isotherms.
Buys Ballot (1850, p. 629) was even more straightforward: the
study of deviations from mean values (mean states) constituted the
second stage in the development of meteorology. He (1847, p. 108)
noted that a similar process was going on in astronomy (planetary
orbits had been assumed elliptical, then perturbations began to be
studied). The same, he added, was the situation concerning all
sciences that did not admit experimentation. He could have cited
geodesy and the ever more precise determination of the figure of the
Earth.
In the 18th and apparently in the early 19th century the mean
monthly and even the mean yearly temperatures had been considered

186
185
as the means of their extreme values (Cotte 1788, p. 9) so that the
occurred introduction of the arithmetic mean of all of their values was
an essential step forward.
Meteorological observations multiplied, and they had been
published without being of use to the general readership of scientific
periodicals. Biot (1855, pp. 1179 – 1180), for example, had opposed
that practice and Mendeleev (1876/1946, p. 267) remarked that the
prevailing “collecting” school of meteorologists needed nothing but
“numbers and numbers”. Later he (1885/1952, p. 527) optimistically
decided that a new meteorology was being born and that “little by
little” it had begun, basing its work on statistical data, to “master,
synthesize, forecast”.
Lamont (1867, p. 247) maintained that the irregular temporal
changes of the atmosphere were not random “in the sense of the
calculus of probability” and (p. 245) recommended his own method of
studying, instead, simultaneous observations made at different
localities. Quetelet (1849, t. 1, Chapter 4, p. 53) remarked that the
differences of such observations conformed to accidental errors, but
he did not elaborate. Much earlier Lamont (ca. 1839, p. 263) indicated
that the air pressure and temperature are very changeable and their
mean values are barely reliable. Again, he did not elaborate and stated
without proof that a year of simultaneous observations is tantamount
to 30 years of usual observations.
Quetelet (1846, p. 275) resolutely contended that meteorology was
alien to statistics: unlike the “physicist”, the statistician wishes to
know, first of all, everything that can influence man and contribute to
his welfare. In addition to meteorology, he cited other “alien”
sciences, such as physical geography, mineralogy, botany. His
statement is correct only insofar as statistical meteorology, stellar
statistics etc. belong to the appropriate sciences.
The study of densities of the distributions of meteorological
elements began in the mid-19th century; Quetelet, for example, knew
that these densities were asymmetric (§ 10.5). At the end of the
century Meyer (1891, p. 32), when mentioning that fact, stated that the
theory of errors was not applicable to meteorology. However,
mathematical statistics does not leave aside the treatment of
asymmetric series of observations, and already K. Pearson (1898)
made use of Meyer’s material for illustrating his theory of asymmetric
curves.
Lamarck was one of the first scholars to note the dependence of the
weather on its previous state, see for example t. 5, pp. 5 and 8 and t.
11, p. 143 of his Annuaries (1800 – 1811). In t. 11, p. 122 he
essentially repeated his first pronouncement and formulated it as an
aphorism:

The entire state of things in the atmosphere […] results not only
from an ensemble of causes which tend to operate, but also from the
influence of the previous state.

Quetelet (1852; 1853, p. 68; 1849 – 1857/1857, pt. 5, pp. 29 and

83) repeatedly mentioned lasting periods of fair or foul weather. He

187
186
(1852; 1857) analysed that phenomenon applying elementary
stochastic considerations and concluded that the chances of the
weather persisting (or changing) were not independent. Köppen’s
analysis (1872) was more mathematically oriented, see Sheynin
(1984b, p. 80).
Quetelet was also praiseworthy for compiling and systematizing
meteorological observations. Köppen (1875, p. 256) noted that
observations made since the early 1840s at the “entire network of
stations” in Belgium “proved to be the most lasting [in Europe] and
extremely valuable”. And Faraday praised Quetelet’s observations of
atmospheric electricity (Note 5).
10.8.4. Astronomy. Here, I only discuss the events of the 19th
century. Already Daniel Bernoulli (§ 6.1.1) and Laplace (§ 7.1-2)
stochastically studied regularities in the Solar system. In actual fact,
they considered the planets as elements of a single population, and this
approach was vividly revealed in the later investigations of the
asteroids. Newcomb (1861a and elsewhere) repeatedly compared the
theoretical (calculated in accordance with the uniform distribution)
and the real parameters of their orbits; true, he was yet unable to
appraise quantitatively his results. Poincaré (§ 11.2-5) stochastically
although worthlessly estimated the total number of the small planets.
Of special interest are Newcomb’s considerations (1862) on the
distribution of the asteroids, likely based on his later published and
even more interesting statement (1881). His former contribution
makes difficult reading mostly because of its loose style. As I
understand him, Newcomb intuitively arrived at the following
proposition: a large number of independent points A1 = (B1 + b1t),
A2 = (B2 + b2t), … where t denoted time, and the other magnitudes
were constant, will become almost uniformly distributed over a
circumference. In 1881 Newcomb remarked that the first pages of
logarithmic tables wore out “much faster” than the last ones and set
out to derive the probability that the first significant digits of
empirically obtained numbers will be n1, n2, … Without any proof he
indicated that, if numbers s1, s2, …, sn were selected “at random”, the
positive fractional parts of the differences (s1 – s2), (s2 – s3), … will
tend, as n → ∞, to a uniform distribution over a circumference, and
that the empirical magnitudes, to which these differences conform,
will have equally probable mantissas of their logarithms. Newcomb’s
reasoning heuristically resembles the Weyl celebrated theorem that
states that the terms of the sequence {nx}, where x is irrational, n = 1,
2, …, and the braces mean “drop the integral part”, are uniformly
distributed on a unit interval. In the sense of the information theory,
Newcomb’s statement means that each empirical number tends to
provide one and the same information. Several authors, independently
one from another, proved that Newcomb was right. One of them
(Raimi 1976, p. 536) called his statement an “inspired guess” and
reasonably noted that it was not, however, universally valid.
By the mid-century, after processing observations made over about
a century, a rough periodicity of the number of sunspots was
established (cf. § 1.2.3). Newcomb (1901), who studied their
observations from 1610 onward, arrived at T = 11.13 years. The

188
187
present-day figure is T ≈ 11 years but a strict periodicity is denied. In
any case, it might be thought that the numbers of sunspots constitute a
time series, an object for stochastic studies. I note that Newcomb
considered the maxima and the minima of that phenomenon as well as
half the sums of the numbers of the sunspots “corresponding to the
year of minimum and the following maximum, or vice versa” (p. 4).
He determined the four appropriate values of T and their mean without
commenting on the possible dependence between them.
The variation of the terrestrial latitudes is known to be caused by
the movement of the pole about some point along a curve resembling
a circumference with period 1.2 years. Newcomb (1892) checked the
then proposed hypothesis that the movement was periodic with T =
1.17 years and he assumed that the pole moved uniformly along a
circumference. Some of his calculations are doubtful (and not
sufficiently detailed, a feature peculiar to many of his works), but he
correctly concluded that the hypothesis was [apparently] valid.
In 1767 Michell (§ 6.1.6) attempted to determine the probability
that two stars were close to each other. By applying the Poisson
distribution, Newcomb (1859 – 1861, 1860, pp. 437 – 439) calculated
the probability that some surface with a diameter of 1° contained s
stars out of N scattered “at random” over the celestial sphere and
much later Fisher (Hald 1998, pp. 73 – 74) turned his attention to that
problem. Boole (1851/1952, p. 256) reasoned on the distinction
between a uniform and any other random distribution:

A ‘random distribution’ meaning thereby a distribution according

to some law or manner, of the consequences of which we should be
totally ignorant; so that it would appear to us as likely that a star
should occupy one spot of the sky as another. Let us term any other
principle of distribution an indicative one.

His terminology is now unsatisfactory, but his statement shows that

Michell’s problem had indeed led to deliberations of a general kind.
See also Newcomb (1904a) who thought about the subjective
difference between those distributions.
Newcomb (1861b) also solved a related problem in which he
determined the probability of the distance between the poles of two
great circles randomly situated on a sphere. Issuing from other initial
considerations, Laplace (1812, p. 261) and Cournot (1843, § 148)
earlier provided solutions differing both from each other and from
Newcomb’s answer (Sheynin 1984a, pp. 166 – 167).
F. G. W. Struve (1827, pp. xxxvii – xxxix) determined the
probability that two or three stars are situated near to each other.
Bertrand (1888, pp. 170 – 171) noted however, that likely/unlikely
relative situations of two stars can be considered in different ways (not
only by distances between them) and concluded (pp. 4 – 7) that the
Michell problem was unsolvable (that there are different ways to
describe randomness, see § 11.1-1.
About 1784 William Herschel started counting the number of stars
situated in different regions of the sky. He thought that his telescope
was able to penetrate right up to the boundaries of the Milky Way and

189
188
hoped to determine its configuration but later understood his mistake
(§ 9C).
In one of its sections he (1784/1912, p. 158) counted the stars in six
fields selected “promiscuously” and assumed the mean number of
them as an estimate for the entire section. Much later Herschel (1817)
proposed a model of a uniform spatial distribution of the stars. He
fixed the boundaries for the distances of the stars of each magnitude
but allowed the stars to be randomly distributed within these
boundaries.
Herschel thus provided an example of randomness appearing
alongside necessity; cf. Poincaré’s statement in § 1.2.4. When
estimating the precision of his model for the stars of the first seven
magnitudes, Herschel calculated the sum of the deviations of his
model from reality. For the first four magnitudes that sum was small
although the separate deviations were large. Recall (§ 6.3.2) that,
when adjusting observations, Boscovich applied a similar test with
respect to absolute deviations. Herschel himself (1805) made use of it
when determining the direction of the Sun’s movement (cf. Note 17 to
Chapter 6).
Herschel (1817/1912, p. 579) indicated that

Any star promiscuously chosen […] out of [14,000 stars of the first
seven magnitudes] is not likely to differ much from a certain mean size
of them all.

He certainly did not know that, with regard to size, the stars are
incredibly different; its mean value is a worthless quantity, and, in
general, stochastic statements, made in the absence of data, are hardly
useful. A formal check in accordance with the Bienaymé –
Chebyshev inequality would have revealed Herschel’s mistake. But in
any case it occurred that the stars, even earlier than the asteroids, had
been considered as elements of a single population (in the last-
mentioned instance, wrongly).
Stellar statistics really originated in the mid-19th century with the
study of the proper motions of hundreds of stars (until 1842, when
astronomers started to use the Doppler’s invention, only in the
directions perpendicular to the appropriate lines of sight).
Argelander (1837, p. 581) considered 560 stars with perceptible
proper motions and determined the Sun’s motion more reliably than it
was achieved previously. Otto Struve (1842; 1844), then F. G. W.
Struve (1852, pp. clxxxii – clxxxv) made the next steps. For the first
of his studies the Royal Astronomical Society awarded O. Struve its
gold medal (Airy 1842).
When studying the Sun’s motion, astronomers beginning with
Herschel thought that the peculiar motions of the stars (their motions
relative to the Sun) were [random variables] and Kapteyn (1906a, p.
400) called the random distribution of the direction of the peculiar
motions a fundamental hypothesis.
The calculated mean proper motions for stars of a given magnitude
proved, however, almost meaningless since the magnitudes depended
on distances. Beginning with W. Herschel, astronomers thought that

190
189
the proper motions were random, but they understood randomness in
different ways. Newcomb (1902a) assumed that their projections on
an arbitrary axis were normally distributed. He derived, although
without providing his calculations, the density laws of their
projections on an arbitrary plane and their own distribution. Both
these laws were connected with the χ2 distribution.
The general statistical study of the starry heaven became more
important than a precise determination of the parameters of some star
(Hill & Elkin 1884, p. 191):

The great Cosmical questions to be answered are not so much what

is the precise parallax of this or that particular star, but – What are
the average parallaxes of those of the first, second, third and fourth
magnitude respectively, compared with those of lesser magnitude?
[And] What connection does there subsist between the parallax of a
star and the amount and direction of its proper motion or can it be
proved that there is no such connection or relation?

Then, Kapteyn (1906b; 1909) described a stochastic picture of the

stellar universe by the laws of distribution of the (random!)
parameters, parallaxes and peculiar motions, of the stars. He (1906a)
also initiated the study of the starry heaven by [stratified] sampling;
here is a passage from a letter that he received in 1904 on this subject
from one of his colleagues, Edward Pickering, and inserted on his p.
67:

As in making a contour map, we might take the height of points at

the corners of squares a hundred meters on a side, but we should also
take the top of each hill, the bottom of each lake, […], and other
distinctive points.

In statistics, sampling became recognized at about the same time,

although not without serious resistance (You Poh Seng 1951) and its
most active partisan was Kiaer, also see § 10.7-2.
Newcomb (1902b, pp. 302 and 303) offered a correct estimate of
Kapteyn’s work:

In recent times what we may regard as a new branch of

astronomical science is being developed. […] This is what we now
call the science of stellar statistics. The statistics of the stars may be
said to have commenced with Herschel’s gauges of the heavens […].
The outcome of Kapteyn’s conclusions is that we are able to describe
the universe as a single object …

The compilation of vast numerical materials (catalogues,

yearbooks) was also of a statistical nature. Moreover, sometimes this
direction of work had been contrasted to theoretical constructions.
Thus, Proctor (1872) plotted 324 thousand stars on his charts
attempting to leave aside any theories on the structure of the stellar
system, but the development of astronomy proved him wrong.

191
190
 Calculation and adjustment of observations, their reasonable
comparison has always been important for astronomy. Here, I again
ought to mention, in the first place, Newcomb. Benjamin (1910) and
many other commentators stated that he had to process more than 62
thousand observations of the Sun and the planets and that his work
included a complete revision of the constants of astronomy. He
necessarily discussed and compared observations obtained at the main
observatories of the world but he hardly had any aids except for
logarithmic tables. In addition he published some pertinent theoretical
studies. He was of course unable to avoid the perennial problem of the
deviating observations. At first he regarded them with suspicion, then
(1895, p. 186), however, became more tolerant. If a series of
observations did not obey the normal law, Newcomb (1896, p. 43)
preferred to assign a smaller weight to the “remote” observations, or,
in case of asymmetric series, to choose the median instead of the
arithmetic mean. He had not mentioned Cournot (§ 10.3-3), and, in
two memoirs published at the same time, he (1897a; 1897b) called the
median by two (!) other, nowadays forgotten, terms.
Mendeleev (§ 10.9.3) objected to combining different summaries of
observations; Newcomb, however, had to do it repeatedly, and in such
cases he (1872) hardly managed without subjective considerations and
assigned weights to individual astronomical catalogues depending on
their systematic errors. Interestingly enough, he then repeated such
adjustments with weights, depending on random errors.
After determining that the normal law cannot describe some
astronomical observations necessarily made under unfavourable
conditions, Newcomb (1886) proposed for them (and, mistakenly, for
all astronomical observations altogether) a generalized law, a mixture
of normal laws with differing measures of precision occurring with
certain probabilities. The measure of precision thus became a discrete
random variable, and the parameters of the proposed density had to be
selected subjectively. Newcomb noted that his density led to the
choice of a generalized arithmetic mean with weights decreasing
towards the “tails” of the variational series. I (§ 6.3.1) have remarked,
however, that that mean was hardly better than the ordinary arithmetic
mean.
He had also introduced some simplifications, and Hulme & Symms
(1939, p. 644) noted that they led to the choice of the location
parameter by the principle of maximum likelihood. Newcomb hardly
knew that his mixture of normal laws was not normal (Eddington
1933, p. 277). In turn, two authors generalized Newcomb’s law
(Lehmann – Filhès 1887; K. F. Ogorodnikov, three English papers in
1928 – 1929), see Sheynin (1995c, pp. 179 – 182), but their work was
of little practical importance.
Like Mendeleev (§ 10.9.3), Newcomb (1897b, p. 165) thought that
the discrepancy between two empirical magnitudes was essential if it
exceeded the sum of the two appropriate probable errors, and it seems
that this rigid test had been widely accepted in natural sciences. Here
is Markov’s relevant pronouncement from a rare source (Sheynin
1990b; pp. 453 – 454): he

192
191
 Like[d] very much Bredikhin’s rule according to which ‘in order to
admit the reality of a computed quantity, it should at least twice
numerically exceed its probable error’. I do [he does] not know,
however, who established this rule or whether all experienced
calculators recognized it.

In other words, the difference between zero and a “real” non-zero

magnitude must twice exceed its probable error, a statement that
conformed to Mendeleev’s and Newcomb’s opinion. But still,
Newcomb several times indicated that some magnitude a determined
by him had mean square error b even when the latter much exceeded
the former including the case (1901, p. 9) of a = 0.05 and b = 0.92!
Dorsey & Eisenhart (1969, p. 50) recommended another rule in which
the probable errors characterized separate observations.
Repeatedly applying the MLSq, Newcomb sometimes deviated
from strict rules; see one such example in § 9A.5-3. In another case he
(1895, p. 52) thought that small coefficients in a system of normal
equations might be neglected, but he had not provided any
quantitative test. Newcomb realized that, when forming normal
equations, the propagation of round-off errors could result in their
interdependence, and he reasonably concluded that in such cases the
calculations should be made with twice as many significant digits.
This is what he (1867) did when studying the calculations of the
Kazan astronomer Kowalski, who had noted that, out of the four
normal equations which he formed, only two were independent. It is
now known that ill-conditioned observational equations should rather
be processed without forming normal equations, – for example, by
successive approximations.
Newcomb’s calculation (1874, p. 167) presents a special case.
Having 89 observational equations in five unknowns, he formed and
solved the normal equations. Then, however, he calculated the
residual free terms of the initial equations and somehow solved these
equations anew (providing only the results of both solutions). He
apparently wished to exclude systematic influences as much as
possible, but how?
Newcomb (1895, p. 82; 1897b, p. 161) mistakenly stated, although
mentioning earlier the definitive Gaussian justification of the MLSq,
that the method was inseparable from the normal law. I note also his
unfortunate reasoning (Newcomb & Holden 1874, p. 270) similar to
the one made by Clausius (§ 10.8.5): for systematic error s and
random errors r1 and r2, as he proved, and only for the normal law by
considering the appropriate double integral, that

E[(s + r1) (s + r2)] = s2.

It might be concluded that Newcomb necessarily remained more or

less within the boundaries of the classical theory of errors and simple
stochastic patterns. At the same time, the extant correspondence
between him and K. Pearson during 1903 – 1907 (Sheynin 2002a, §
7.1) testifies that he wished to master the then originating

193
192
mathematical statistics. Here is a passage from his archival letter of
1903 to Pearson (I have supplied the reference):

You are the one living author whose production I nearly always
read when I have time and can get at them, and with whom I hold
imaginary interviews while I am reading.

I mention finally Newcomb’s statistical contribution (1904b) in

which he examined the classical problem of the sex ratio at birth (see
§§ 2.2.4, 3.3.4, 4.4 and 6.1.1). He assumed that there existed three
kinds of families numbered, say, m, n, and n, for whom the
probabilities of the birth of a boy were p, p + α and p – α respectively
and he studied, in the first place, the births of twins. The sex of the
embryo, as he thought, became established only after the action of a
number of successive causes made it ever more probable in either
sense.
10.8.5. Physics. 1)The kinetic theory of gases originated in mid-19th
century as the result of the penetration of the statistical method into
physics. Truesdell (1975) discussed its early history; thus (p. 28), it
was Waterson who, in 1843, introduced the mean free path of a
molecule, cf. § 8.7, but his innovation was not published. Clausius
likely published the first memoir (1849) which belonged to physics
(but did not deal with the molecular hypothesis) and contained ideas
and methods of the theory of probability.
After Poisson’s death that theory sank into oblivion (§§ 0.1 and
7.4). No wonder that Clausius (1889 – 1891, p. 71) made a point to
prove the equality E(ξ/Eξ) = 1 for the velocity ξ of a molecule. See a
similar case in § 10.8.4.
Maxwell twice mentioned Laplace (Sheynin 1985, pp. 364 and
366n), although without providing any definite references, whereas
Boltzmann, who cited many scholars and philosophers in his popular
writings, never recalled him. Khinchin (1943/1949, p. 2) maintained
that Maxwell and Boltzmann applied

Fairly vague and somewhat timid probabilistic arguments that do

not pretend here to be the fundamental basis, and play approximately
the same role as purely mechanical considerations. […] Far reaching
hypotheses are made concerning the structure and the laws of
interaction between the particles […]. The notions of the theory of
probability do not appear in a precise form and are not free from a
certain amount of confusion which often discredits the mathematical
arguments by making them either devoid of any content or even
definitely incorrect. The limit theorems […] do not find any
applications […]. The mathematical level of al these investigations is
quite low, and the most important mathematical problems which are
encountered in this new domain of application do not yet appear in a
precise form.

His statement seems too harsh, written from the standpoint of

statistical mechanics of the mid-20th century. Then, I believe that it
was partly occasioned by Boltzmann’s verbose style of writing. Third,

194
193
physicists certainly applied the LLN indirectly. Fourth, Khinchin said
nothing about positive results achieved in physics (formulation of the
ergodic hypothesis, use of infinite general populations, Maxwell’s
indirect reasoning about randomness). My first remark is indeed
essential; here is an extract from Maxwell’s letter of 1873 (Knott
1911, p. 114):

By the study of Boltzmann I have been unable to understand him.

He could not understand me on account of my shortness, and his
length was and is an equal stumbling block to me.

And Boltzmann (1868/1909, p. 49) indeed owned that it was

difficult to understand Maxwell’s “Deduktion” (1867) “because of its
extreme brevity”. I emphasize that statistical mechanics could not
have appeared unless and until the kinetic theory with its
mathematical shortcomings had been established (Truesdell 1975;
Brush 1976).
2) Clausius. He (1857/1867, pp. 238 and 248) asserted that
molecules moved with essentially differing velocities. Even
Boscovich (1758, § 481) stated something similar but perhaps
presumed that the differences between these velocities were not large:
The “points” [atoms] of “a particle” [of light, as in § 477, or of any
body, as in § 478] move “together with practically the same velocity”,
and the entire particle will “move as a whole with the single motion
that is induced by the sum [the mean] of the inequalities pertaining to
all its points”.
Clausius used a single mean velocity such as to make the entire
kinetic energy of a gas equal to its actual value. Later he (1862/1867,
p. 320) maintained that the velocities of molecules randomly differed
one from another.
And he (1858/1867, p. 268) studied the length of the free path of a
molecule. Denote the probability of a unit free path by a, then

W = ax = (e–x),  > 0

will be the probability of its being equal to x; here,  is derived from

the molecular constants of the substance. Similar considerations are in
other works of Clausius (1862/1867, § 29; 1889 – 1891, pp. 70 – 71
and 119).
He (1889 – 1891, pp. 70 – 71) also calculated the mean free path of
a molecule. Actually, without writing it out, he considered free paths
of random length  and calculated the expected free path as an integral
over all of its possible values from 0 to ∞.
Suppose now that

 = 1 + 2 + … + m

where m is an arbitrary natural number. Then, according to Clausius’

assumptions, k, k = 1, 2, …, m, will not depend on
(1 + 2 + … + k–1) and the characteristic function for k will be equal
to the product of these functions for the previous ’s. In this instance,

195
194
all these functions are identical, and F(s), the integral distribution
function of , is therefore infinitely divisible.
Distribution functions had first appeared (or were easy to be derived
from the correspondence of) Huygens (§ 2.2.2), from De Moivre (§
4.2) and Davidov (1885) and directly in Poisson (§ 8.2).
Clausius’ achievements were interesting, but he did not attempt to
construct the kinetic theory of gases on a stochastic basis.
Nevertheless, his role in this direction, at least from the viewpoint of
probability theory, had not yet ben studied. See Schneider (1974) who
reviewed his concrete physical research by stochastic considerations.
His great merit is seen in Maxwell’s statement (1875/1890, p. 427):

Clausius opened up a new field […] by showing how to deal

mathematically with moving systems of innumerable molecules.

3) Maxwell (1860) established his celebrated distribution of the

velocities of monatomic molecules

1
φ(x) = exp( x 2 /α 2 ).
α π

He tacitly assumed that the components of the velocity were

independent; later this restriction was weakened (§ 9A.1.3). He then
maintained that the average number of particles with velocities within
the interval [v; v + dv] was proportional to

4
f(x) = 3
v 2 exp(v 2 /α 2 )dv.
α π

This can be justified by noting that the probability of such velocities

can also be represented as

2  v  dv

 dφ  sin θdθ
0 0

v
t 2 exp( t 2 /α 2 ) dt.

It is presumed here that the components of the velocity in each of the

three dimensions have the same distribution. Maxwell left interesting
statements about the statistical method in general, and here is one of
them (1873b/1890, p. 374):

We meet with a new kind of regularity, the regularity of averages,

which we can depend upon quite sufficiently for all practical
purposes, but which can make no claim to that character of absolute
precision which belongs to the laws of abstract dynamics.

The drafts of the source just mentioned (Maxwell 1990 – 2002,

1995, pp. 922 – 933) include a previously unpublished and very
interesting statement (p. 930): abandoning the “strict dynamical
method” and adopting instead the statistical method “is a step the
philosophical importance of which cannot be overestimated”.
196
195
And here is his definition (not quite formal) of the statistical method
which heuristically resembles the formulation provided by
Kolmogorov & Prokhorov (§ 0.2): it consisted in “estimating the
average condition of a group of atoms” (1871/1890, p. 253), in
studying “the probable [not the average!] number of bodies in each
group” under investigation (1877, p. 242).
Maxwell gave indirect thought to randomness. Here is his first
pronouncement (Maxwell 1859/1890, vol. 1, pp. 295 – 296) which
was contained in his manuscript of 1856 (1990 – 2002, 1990, p. 445),
and it certainly describes his opinion about that phenomenon:

There is a very general and very important problem in Dynamics

[…]. It is this – Having found a particular solution of the equations of
motion of any material system, to determine whether a slight
disturbance of the motion indicated by the solution would cause a
small periodic variation, or a derangement of the motion […].

Maxwell (1873a, p. 13) later noted that in some cases “a small

initial variation may produce a very great change […]”. Elsewhere he
(report read 1873, see Campbell & Garnett 1882/1969, p. 440)
explained that in such instances the condition of the system was
unstable and prediction of future events becomes impossible. He
(Ibidem, p. 442) provided an example of instability of a ray within a
biaxial crystal and prophetically stated (p. 444) that in future
physicists will study “singularities and instabilities”.
In a manuscript of the same year, 1873 (p. 360), Maxwell remarked
that

The form and dimensions of the orbits of the planets […] are not
determined by any law of nature, but depend upon a particular
collocation of matter. The same is the case with respect to the size of
the earth.

This was an example illustrating Poincaré’s statement concerning

randomness and necessity (§ 1.2.4), but I ought to add that it was not
sufficiently specific; the eccentricities of planetary orbits depend on
the velocities of the planets, cf. end of § 7.3.
And here is Maxwell’s position (1875/1890, p. 436) concerning
randomness in the atomic world:

The peculiarity of the motion of heat is that it is perfectly irregular;

[…] the direction and magnitude of the velocity of a molecule at a
given time cannot be expressed as depending on the present position
of the molecule and the time.

At the very end of his life Maxwell (1879/1890, pp. 715 and 721)
introduced a definition for the probability of a certain state of a system
of material particles:

I have found it convenient, instead of considering one system of […]

particles, to consider a large number of systems similar to each other

197
196
[…]. In the statistical investigation of the motion, we confine our
attention to the number of these systems which at a given time are in a
phase such that the variables which define it lie within given limits.

Boltzmann (1868, § 3) defines the probability of the system being in

a phase […] as the ratio of the aggregate time during which it is in
that phase to the whole time of the motion […].

4) If the classical definition of probability is included here, we can

say that Boltzmann used three formulations. Maxwell (§ 10.8.5-2)
mentioned one of them, and another reference can be added:
Boltzmann (1868/1909, Bd. 1, p. 50). Another one was that applied by
Maxwell (see same subsection) although sometimes Boltzmann
(1878/1909, p. 252) did not indicate which one he was employing. He
(1872/1909, p. 317) apparently thought that these posterior
probabilities were equivalent.
In other words, with respect to separate molecules Boltzmann
introduced the time average probability, – and maintained that it was
equivalent to the “usual” phase average probability, also see § 12-2.
When studying polyatomic gases, Boltzmann (1871) defined the
probability of its state as a product such as fdω where f was some
function, varying in time, of the coordinates and velocities of the
separate molecules and dω, the product of the differentials of those
parameters. For stochastic processes, such functions determine the
distribution of a system of random variables at the appropriate
moment. Zermelo (1900, p. 318) and then Langevin (1913/1914, p. 3)
independently stressed the demand to provide a “definition correcte et
claire de la probabilité” (Langevin).
Like Maxwell, Boltzmann (1887/1909, p. 264; 1899, Bd. 2, p. 144)
used the concepts of fictitious physical systems and infinite general
population.
Boltzmann (1896/1909, p. 570) stated that the [normal law]
followed from equal probabilities of positive and negative elementary
errors of the same absolute value. His was of course an unworthy
formulation of the CLT.
I ought to add that Boltzmann respected the theory of probability.
Thus (1872/1909, p. 317)

An incompletely proved theorem whose correctness is questionable

should not be confused with completely proved propositions of the
theory of probability. Like the results of any other calculus, the latter
show necessary inferences made from some premises.

And again (1895/1909, p. 540): the theory of probability “is as exact

as any other mathematical theory” if, however, “the concept of equal
probabilities, which cannot be determined from the other fundamental
notions, is assumed”.
From 1871 onward Boltzmann had been connecting the proof of
the second law of thermodynamics with stochastic considerations.
Thus, he (1872/1909, pp. 316 – 317) declared that the problems of the
mechanical theory of heat are also problems of probability theory.

198
197
Then, however, he (1886/1905, p. 28) indicated that the 19th century
will be the age of “mechanical perception of nature, the age of
Darwin”, and (1904a/1905, p. 368) that the theory of evolution was
understandable in mechanical terms, that (1904b, p. 136) it will
perhaps become possible to describe electricity and heat mechanically.
The possible reason for his viewpoint was that he did not recognize
objective randomness. Another reason valid for any scholar was of
course the wish to keep to “abstract dynamics”, see Maxwell’s
statement on the “new kind of regularity” (§ 10.8.5-3) and the opinion
of Hertz (1894, Vorwort): “Physicists are unanimous in that the aim of
physics is to reduce the phenomena of nature to the simple laws of
mechanics”. And here is a lucid description of this point as far as
Boltzmann was considered (Rubanovsky 1934, p. 6): in his works

Randomness […] struggles with mechanics. Mechanical philosophy

is still able […] to overcome randomness and wins a Pyrrhic victory
over it but recedes undergoing a complete ideological retreat.

10.9. Natural scientists

10.9.1. Ivory. In a letter to Olbers of 1827, Gauss (W/Erg-4, No. 2,
pp. 475 – 476) called Ivory an “acute” mathematician, but indicated
that the “spirit” of the MLSq was alien to him. In 1825 – 1830 Ivory
published 11 papers in one and the same periodical [the last of these
was Ivory (1830)] devoted to the derivation of the flattening of the
Earth’s ellipsoid of rotation by means of pendulum observations. It is
not amiss to add that his main contributions pertained to the theory of
the figure of the Earth (attraction of a material point by an ellipse).
In accordance with the Clairaut theorem, the Earth’s flattening (see
Note 10 to Chapter 6) is determined by two observations of [the
acceleration of] gravity at different latitudes; however, unavoidable
errors and local irregularities in the figure of the Earth necessitate the
use of redundant observations. See the same remark in § 6.3. To
strengthen Gauss’ remark, I state that Ivory was simply ignorant of
the MLSq and without justification called it not good enough. He
denied it in words but applied the MLSq, perhaps not even realizing it
at once. Thus, starting from equations of the type (1.1) with ai = 1 he
(1826b, pp. 244 – 245) stated that the condition ∑vi = 0, unlike the
requirement of the MLSq [av] = [bv] = … = 0, see equations (1.5),
was expedient. He failed to notice that in his case the expedient
condition coincided with the demand that [av] = 0.
Then, having at his disposal 5 – 7 observations, only one of which
was made at a southern station, he (1826a, p. 9) combined it with each
of the others (so as to have pairs with a large latitudinal difference
between stations) and calculated the flattening from the thus obtained
pairs. The weight of the equatorial observation became absurdly great
and its error corrupted all the pairs in the same way. An utterly
unworthy manner of treatment, as Gauss stated but it seems that the
answer to this problem depends on the unknown magnitude of the
systematic errors.
Only later did Ivory remark that the local anomalies of gravity can
essentially influence the end result, – and rejected a large part of the

199
198
available observations, – up to 31%, see Ivory (1826b, p. 242), – and
even began doubting whether it was possible to derive a single
flattening. Local anomalies are indeed extremely troublesome (also
see § 10.8.3 where I indirectly mentioned local perturbations of
temperature) but Ivory attempted to get rid of them too radically.
Finally, when estimating the precision of his results, he had not
applied the variance.
Before adjusting pendulum observations it is possible to replace
stations having almost the same latitude by one fictitious mean station
(which Ivory had not done). If, however, the longitudes of such
stations are also almost the same, all of them could have been
corrupted by the same local gravimetric anomaly so that the weight of
the mean station should not be increased as compared with either of
them.
I ought to add two remarks. First, his final result (1828, p. 242) was
sufficiently close to the flattening of the Krasovsky ellipsoid of 1940
(Sakatov 1950, p. 364): e = 0.00333 – 0.00338 and 0.00335,
respectively. In addition, Ivory (1825, p. 7), without, however,
mentioning Gauss, maintained that the MLSq should be substantiated
by the principle of maximum weight. Second, Ivory actually wished to
solve two problems at once: to find out whether the observations were
consistent with an ellipsoidal Earth, and to adjust them. It is the
minimax method (§ 6.3.2) rather than the MLSq that is best for
solving the first problem.
10.9.2 Fechner. He (1860) was the founder of psychophysics and
therefore became one of the first to introduce the statistical method
into physics, although not in the crucial direction. He (1860, Bd. 1, p.
8, see also 1877, p. 213) defined it as an “exact doctrine on the
functional correspondence or interdependence of body and soul”.
And here is a modern down to earth definition (New Enc. Brit., 15th
ed., vol. 9, 1997, p. 766): psychophysics is a

Study of quantitative relations between psychological events and

physical events, or, more specifically, between sensations and the
stimuli that produce them.

There also Fechner’s book just mentioned is called a classic.

Fechner (1855 and 1864) missed the opportunity to comment on the
developing kinetic theory of gases. Moreover, he (1874b, pp. 7 and 9;
1897, p. 15) repeatedly treated physics on a par with (practical)
astronomy by stating that both these branches of natural sciences had
to do with symmetric distributions and true values of magnitudes
sought. His mathematical tools and the very approach were primitive
and almost everything he achieved had to be repeated at a much
higher level. Ebbinghaus (1908, p. 11) possibly having in mind
Fechner’s non-scientific writings called him “a philosopher full of
fantasies” and, at the same time, “a most strict physicist” who had
“put […] together psychophysics as a new branch of knowledge”. As
chance would have it, Fechner published those writings under the pen
name Dr Mises.

200
199
 Fechner himself (1877, p. 215) figuratively estimated his own work
(and made known the existence of opposition to it):

The Tower of Babel was not completed because the workers were
unable to explain to each other how should they build it. My
psychophysical structure will probably survive because the workers
cannot see how they might demolish it.

Being the co-author of the logarithmic Weber – Fechner law

connecting stimuli with sensations, Fechner extended the range of its
application after making a great number of experiments (1860; 1887).
He studied the methods of experimentation, and made reasonable
pertinent statements. The modern method of paired comparisons (H.
A. David 1963) owes much to him.
In the theory of errors Fechner had been mentioning Gauss, but he
also attempted, sometimes unsuccessfully, to introduce his own
innovations, or to repeat unknown to him previous findings but he
nevertheless somewhat furthered that theory. Thus, issuing from
elementary but apparently non-rigorous considerations, he (1874a, p.
74) provided a formula for estimating the precision of observations
which coincided with the Peters formula (9B.11) but was applicable to
any distributions. Then, proceeding from the Gaussian formulas (§
9A.3), he compared two competing expressions connecting the
magnitudes of the stars with their brightness, solved redundant
systems of equations by the method of pairwise combinations (§
6.3.2), and remarked, without substantiation (and hardly correctly),
that that method asymptotically tended to the MLSq (1887, p. 217).
Fechner’s main innovation was, however, the collective, – actually,
the set of observed values of a random variable. He (1897) proposed
to study collectives by applying several mean values, their mutual
arrangement, and their deviations (both absolute and normed) from the
observations. He mostly paid attention to asymmetric collectives and
even attempted to discover a universal asymmetric distribution for
errors in natural sciences (cf. § 10.5). Fechner especially examined the
double normal law (two different normal laws for the smaller and the
larger values of observations in the variational series respectively,
turning into one another at the point of maximal probability, i.e., at the
mode), and the double lognormal law. Fechner also attempted,
although not very successfully, to separate the real and the apparent
(caused by an insufficient number of observations) asymmetry.
Finally, Fechner (1897, pp. 365 – 366) studied the dependence of
the successive daily air temperatures by comparing their course with
the arrangement of winning (numbered) tickets of a reputed German
lottery. When examining the results of the lottery, he achieved an
interesting result pertaining to the runs up and down (cf. § 10.2-5).
Furthermore, Fechner even introduced a measure of dependence
varying from 0 to 1, but describing only “positive” dependences. His
contribution appeared posthumously, after the Galton correlation
theory had emerged.
Mises (1928/1972, pp. 26 and 99) highly appraised Fechner’s
efforts and owned (p. 99) that Fechner’s “constructions prompted, at

201
200
least me [Mises], to adopt a new viewpoint”. Two more passages by
K. Pearson (1905, p. 189) and Freud (1925/1963, p. 86) are in order:

All the leading statisticians from Poisson to Quetelet, Galton,

Edgeworth and Fechner […] have realized that asymmetry must be in
some way described before we can advance in our theory of variation
[in biology].

I was always open to the ideas of Fechner and have followed that
thinker upon many important points.

10.9.3. Mendeleev. From 1893 to 1907 Mendeleev was Director of

Russia’s Main Board of Measures and Weights which was established
on his initiative on the basis of its modest predecessor (where, from
1892, Mendeleev was scientific curator). Actually, he was an
extremely versatile scientist, studied statistics of population and
industry and thought that the range of statistics is unrestricted.
Mendeleev processed observations both as a chemist and a
metrologist. He (1872b/1951, p. 101) distrusted data “obtained under
differing conditions, by different methods and observers” as compared
with those “achieved by precise methods and experienced persons”.
And (1887/1934, p. 82) “disadvantageous” data ought to be rejected
“by a clear critical appraisal”, otherwise “a realistic result” is
impossible to get.
No wonder that Mendeleev (1872a/1939, p. 144) preferred “to
make a few but precise and repeated measurements” and objected to
amassing observations; true, this attitude was partly due to his wish to
avoid calculations, cf. Boyle’s statement in § 1.2.2. Mendeleev
effectively repeated this reasoning (1875a/1939, p. 256) and, for
example in the first case (an attempt at refining the Boyle – Mariotte
law), he added that the measurements ought to be made at
“significantly different pressures”. He was apparently concerned with
systematic errors which could have been otherwise almost the same.
Mendeleev (1875b/1950, p. 209) thought that an observational
series should be “harmonious”, that is, that its median should coincide
with its arithmetic mean, or (his second definition) that the mean of its
middlemost third should coincide with the mean of the means of its
extreme thirds. In the first case, he mistakenly added that the
coincidence meant that the appropriate distribution was normal. He
had not said how to treat observations which did not obey his wish; cf.
end of § 10.8.3.
The deviation of the arithmetic mean from the median, normed in a
certain way, is nowadays recognized as a measure of asymmetry of
the appropriate distribution (Yule & Kendall 1958, p. 161). For
Mendeleev, the probable error was the main estimator of precision and
he (1860/1947, p. 46) assumed that the admissible deviation between
two means was the sum of their probable errors (cf. § 10.8.4). Suppose
that these errors are equal to each other; then, for the normal
distribution, their sum is 1.35σ where σ is the standard deviation (or
the mean square error) of each mean. On the other hand, the standard
deviation of the difference between the means is σ√2 and it thus

202
201
occurs that the studied difference is essential when it is equal to its
standard deviation. Mendeleev’s (or Bredikhin’s, or Newcomb’s)
rule seems to be too rigid. A different rule was recommended more
recently (Dorsey & Eisenhart 1969, p. 50) in which the probable
errors of individual measurements were involved instead.
Mendeleev had not mentioned the second Gaussian justification of
the MLSq and made a few mistakes in his theoretical considerations.
One of them was an excessive belief in the arithmetic mean
(1856/1937, p. 181; 1877/1949, p. 156; 1895/1950, p. 159): he thought
that that estimator ought to be chosen if nothing was known about the
precision of the individual results; cf. § 7.2-6.

Notes
1. He published many very short notes and insufficiently described his findings,
sometimes, like in this case, without proof.
2. Cournot only explained his understanding of the distinction between chance
and probability in § 48 and not clearly enough, see his §§ 12 and 240/3. True, in his
Préface he published Poisson’s letter of 1836 where its author had indicated that
with regard to that point they were unanimous, cf. § 8.1.
3. No one apparently recalled it before Cournot; and only Chuprov (1909/1959,
p. 188) mentioned it afterwards.
4. He read his work in 1846, but the existing materials testify that already then he
could have known Buniakovsky’s book, and exactly him did Ostrogradsky
apparently criticize (see next sentence). He barely busied himself with probability,
but he (1858) made the calculations necessary for the work of a society of mutual
insurance. In § 7.1-9 I mentioned his attempt to generalize the notion of moral
expectation. On Ostrogradsky see Gnedenko (1951).
5. Over the years, Faraday (1991 – 2008) several times expressed his high
opinion about Quetelet’s measurements of atmospheric electricity, and I especially
note two of his letters (to Quetelet, No. 1367 of 1841 in vol. 3, 1996, p. 42, and to
Richard Taylor, the then Editor of the Lond., Eduinb. and Dublin Phil. Mag.,
No. 2263 of 1850, 1999, p. 270):

1) You are indeed a worthy example in activity & power to all workers in science
and if I cannot imitate your example I can at least appreciate & value it.
2) Faraday approvingly remarked on the absence of “imagination or hypothesis”
in Quetelet’s work and added: Such was the true method by which advances in
science in this very difficult part could be really made.

See also the opinion of Köppen about the meteorological observations in

Belgium at the very end of § 10.8.3.
6. Achenwall (Schlözer 1804, § 3) declared that statistics of a given state is a
collection of its remarkable features and it is unclear how Delambre’s differing
understanding of statistics related to political arithmetic. One point is however clear:
the refusal to study causes and effects was quite foreign to Petty and Graunt.
7. The Paris statistical society was established a bit earlier than 1804 (Schlözer
1804, § 7). It attempted to collect and treat data on many diverse matters, in the first
place concerning France.
8. The Handbook of social indicators published in 1989 by the UN, see W. F. M.
De Vries (2001), listed several hundred indicators separated into 13 groups. They
help to trace the range of problems of modern statistics. Several papers on the
newest goals of statistics in the “information society” are collected in the
International Statistical Review, vol. 71, No. 1, 2003. It is opportune to add that
Leibniz (Sheynin 1977, pp. 222 – 227) had interesting thoughts about the aim of
statistics.
9. It had been applied in England from the 12th century onward for checking the
quality of batches of new coins (Stigler 1977). Ptukha (1961) described its usage in
Russia from the 17th century. In the beginning of the 18th century Marshall Vauban

203
202
estimated France’s agricultural production by sampling, but apparently
unsuccessfully (Moreau de Jonnès 1847, pp. 53 – 54).
Lagrange (1796), see Pearson (1978, pp. 628 – 635), published an essay on
political arithmetic in which he had to use sampling.
10. Private communication (2003) by Prof. Walter Mann, a grandson of Mendel’s
nephew, Alois Schindler, and a typographic text of the latter’s manuscript (of his
report of 1902). That manuscript was, however, published (Krizenecky 1965, pp. 77
– 100).
11. Still earlier the problem of allowing for local anomalies presented itself when
pendulum observations began to be used for determining the flattening of the Earth’s
ellipsoid of rotation, see § 10.9.1.

204
203
 11. Bertrand and Poincaré
Passing now to Bertrand, I disturb the chronology of description,
but not its logic: he was not interested in the work of Chebyshev. This
is also true with regard to Poincaré who never mentioned
Chebyshev’s followers in probability (Markov and Liapunov) either.
11.1. Bertrand
In 1855 Bertrand had translated Gauss’ works on the MLSq into
French1, but his own work on probability began in essence in 1887 –
1888 when he published 25 notes in one and the same periodical as
well as his main treatise (1888a), written in great haste and carelessly,
but in a very good literary style. I take up its main issues and state
right now that it lacks a systematic description of its subject.
1) “Uniform” randomness. By several examples Bertrand proved
that the expression “at random”, or even “uniformly” random, was not
definite enough. Thus, he maintained that the Michell problem
(§ 6.1.6) should have been generalized: remarkable was not only a
small distance between stars, but some other features of their mutual
arrangement as well. One of his examples (p. 4) became classical.
Determine the probability, Bertrand asked, that a randomly drawn
chord of a given circle was longer than the side of an equilateral
triangle inscribed in the circle. He listed three possible answers:
a) One endpoint of the chord is fixed; p = 1/3.
b) The chord’s direction is fixed; p = 1/2.
c)The location of the centre of the chord in any point of the circle is
equally probable; p = 1/4.
A curious statement about this problem is due to Darboux
(1902/1912, p. 50):

In accord with considerations which seem equally plausible, he

[Bertrand] derived two different values for the probability sought, 1/2
and 1/3. He investigated this question and found its solution, but left
its discovery to the readers.

In failing to mention the third solution he possibly followed Poincaré

(§ 12-4). I return to this problem in Chapter 12.
2) Statistical probability and the Bayesian approach. Heads
appeared m = 500,391 times in n = 106 tosses of a coin (p. 276). The
statistical probability of that event is p = 0.500391; it is unreliable, not
a single of its digits merits confidence. After making this astonishing
declaration, Bertrand compared the probabilities of two hypotheses,
namely, that the probability was either p1 = 0.500391, or p2 =
0.499609. However, instead of calculating

[p1mp2n] ÷ [p2mp1n],

he applied the De Moivre – Laplace theorem and only indicated that

the first probability was 3.4 times higher than the second one. So what
should have the reader thought? And his choice of the two
probabilities was unfortunate.
As I understand him, Bertrand (p. 161) “condemned” the Bayes
“principle” only because the probability of the repetition of the

205
204
occurrence of an event after it had happened once was too high (cf. the
problem about the sunrise in § 5.1). This conclusion was too hasty,
and the reader was again left in suspense: what might be proposed
instead? Note that Bertrand (p. 151) mistakenly thought that the De
Moivre – Laplace theorem precisely described the inverse problem,
the estimation of the theoretical probability given the statistical data.
3) Statistics of population. Bertrand indicated that there existed a
dependence between trials (or their series) and that the probabilities of
the studied events could change. He referred only to Dormoy and had
not provided any concrete examples, but he (p. 312) noted that, when
studying the sex ratio at birth, both Laplace and Poisson had assumed
without justification that the probability of a male birth was constant
in time and space. Yes, but their mistake was only methodological
since they could not have failed to understand this circumstance (as I
mentioned in § 7.1-5).
Bortkiewicz (1930, p. 53) concluded that Dormoy was much less
important than Lexis; see however end of § 15.1.1.
4) Mathematical treatment of observations. Bertrand paid much
attention to this issue, but his reasoning was amateurish and
sometimes wrong. Even if, when translating Gauss (see above), he
had grasped the essence of the MLSq, he barely remembered that
subject after more than 30 years. Thus, he (pp. 281 – 282) attempted
to prove that the sample variance (9.6b) might be replaced by another
estimator of precision having a smaller variance. He failed to notice,
however, that, unlike the Gauss statistic, his new estimator was biased.
Furthermore, when providing an example, Bertrand calculated the
variance of (9.6b) for the case of the normal distribution instead of
applying the Gauss formula (9.6c).
At the same time Bertrand formulated some sensible remarks. He
(p. 248) expressed a favourable opinion about the second Gauss
justification of the MLSq but indicated (p. 267) that, for small errors,
the even distribution

φ(x) = a + bx2

could be approximately represented by an exponential function of a

negative square, – that the first substantiation of the method was also
approximately valid. Finally, Bertrand provided an argument against
the postulate of the arithmetic mean, see § 9A.2-2.
5) Several interesting problems in addition to that described in
§ 11.1-1 dwell on a random composition of balls in an urn; on
sampling without replacement; on the ballot problem; and on the
gambler’s ruin.
a) White and black balls are placed in the urn with equal
probabilities and there are N balls in all. A sample made with
replacement contained m white balls and n black ones. Determine the
most probable composition of the urn (pp. 152 – 153). Bertrand
calculated the maximal value of the product of the probabilities of the
sample and of the hypotheses on the composition of the urn.
b) An urn has sp white balls and sq black ones, p + q = 1.
Determine the probability that after n drawings without replacement

206
205
the sample will contain (np – k) white balls (p. 94). Bertrand solved
this problem applying the [hypergeometric distribution] and obtained,
for large values of s and n, an elegant formula

1
P= s /( s  n ) exp[– k2s/2pqn(s – n)].
2πpqn

He (1887b) published this formula earlier without justification and

noted that the variable probability of extracting the balls of either
colour was “en quelque sorte un régulateur”.
c) Candidates A and B scored m and n votes respectively, m > n and
all the possible chronologically differing voting records were equally
probable. Determine the probability P that, during the balloting, A was
always ahead of B (p. 18). Following André (1887), who provided a
simple demonstration, Bertrand proved that

P = (m – n)/(m + n), (1)

see also Feller (1950/1968, § 1 of Chapter 3). Actually, Bertrand

(1887a) was the first to derive formula (1) by a partial difference
equation. This ballot problem has many applications (Feller, Ibidem).
Takácz (1982) traced its history back to De Moivre (§ 4.1-5). He
indicated that it was extended to include the case of m ≥ µn for
positive integral values of µ and that he himself, in 1960, had further
generalized that extended version.
d) I select one out of the few problems on the gambler’s ruin
discussed by Bertrand (pp. 122 – 123). Gambler A has m counters and
plays with an infinitely rich partner. His probability of winning any
given game is p. Determine the probability that he will be ruined in
exactly n games (n > m). Bertrand was able to solve this problem by
applying formula (1). Calculate the probability that A loses (n + m)/2
games and wins (n – m)/2 games; then, multiply it by the probability
that during that time A will never have more than m counters, that is,
by m/n.
In a brief chapter he largely denied almost everything done in the
“moral applications” of probability by Condorcet (but did not refer to
Laplace or Poisson).
In two of his notes Bertrand (1888b; 1888c) came close to proving
that for a sample from a normal population the mean and the variance
were independent. Heyde & Seneta (1977, p. 67n) indicated this fact
with respect to Bertrand’s second note; see §§ 7.2-5 and 9B-5 for the
previous findings of Laplace and Helmert.
Taken as a whole, Bertrand’s treatise is impregnated with its non-
constructive negative (and often unjustified) attitude towards the
theory of probability and treatment of observations. And at least once
he (pp. 325 – 326) wrongly alleged that Cournot had supposed that
judges decided their cases independently one from another, see § 10.3-
6. I ought to add, however, that Bertrand exerted a strong (perhaps too
strong) influence upon Poincaré, and, its spirit and inattention to
Laplace and Bienaymé notwithstanding, on the revival of the interest
of French scientists in probability (Bru & Jongmans 2001).

207
206
 11.2 Poincaré
In the theory of probability, Poincaré is known for his treatise
(1896); I refer to its extended edition of 1912. I note first of all that he
had passed over in silence not only Russian mathematicians, but even
Laplace and Poisson, and that his exposition was imperfect.
Commenting on the first edition of his treatise, Bortkiewicz
(Bortkevich & Chuprov 2005, Letter 19 of 1897) noted:

The excessively respectful attitude towards […] Bertrand is

surprising. No traces of a special acquaintance with the literature on
probability are seen. The course is written in such a way as though
Laplace and Poisson, especially the latter, never lived.

Following Bertrand, Poincaré (p. 62) called the expectation of a

random variable its probable value; denoted the measure of precision
of the normal law either by h or by √h; made use of loose expressions
such as “z lies between z and z + dz” (p. 252).
Several times Poincaré applied the formula

  ( x )  ( x )dx
n
 ( x0 )
lim = ,n→∞ (2)
  ( x )  ( x )dx  ( x0 )
n

where Ф(x) was a restricted positive function, xo, the only point of its
maximum, and the limits of integration could have been infinite
(although only as the result of a formal application of the Bayesian
approach). Poincaré (p. 178) only traced the proof of (2) and, for
being true, some restrictions should perhaps be added. To place
Poincarè’s trick in the proper perspective, see Erdélyi (1956, pp. 56 –
57). I discuss now some separate issues mostly from Poincaré’s
treatise.
1) The theory of probability. Poincaré (p. 24) reasonably stated that
a satisfactory definition of prior probability was impossible, cf. § 7.4.
Strangely enough, he (1902/1923, p. 217) declared that “all the
sciences” were nothing but an “unconscious application” of the
calculus of probability, that the theory of errors and the kinetic theory
of gases were based on the LLN and that the calculus of probability
“will evidently ruin them” (les entrainerait évidemment dans sa
ruine). Therefore, as he concluded, the calculus was only of practical
importance2. Another strange pronouncement is in his treatise (1896,
p. 34). As I understand him, he maintained that a mathematician is
unable to understand why forecasts concerning mortality figures come
true.
In a letter of ca. 1899 partly read out at the hearing of the notorious
Dreyfus case (Le procès 1900, t. 3, p. 325; Sheynin 1991a, pp. 166 –
167) Poincaré followed Mill (§ 8.9.1), even generalized him to
include all “moral sciences” and declared that the appropriate findings
made by Condorcet and Laplace were senseless. And he objected to
a stochastic study of handwriting for identifying the author of a certain
document.

208
207
 The interest in application of probability to jurisprudence is now
revived. Heyde & Seneta (1977, p. 34) had cited several pertinent
sources published up to 1975; to these I am adding Zabell (1988a),
Gastwirth (2000) and Dawid (2005) who emphasized the utmost
importance of interpreting background information concerning
stochastic reasoning, cf. § 3.1.2.
2) Poincaré (1892a) had published a treatise on thermodynamics
which Tait (1892) criticized for his failure to indicate the statistical
nature of this discipline. A discussion followed in which Poincaré
(1892b) stated that the statistical basis of thermodynamics did not
satisfy him since he wished to remain “entirely beyond all the
molecular hypotheses however ingenious they might be”; in
particular, he therefore passed the kinetic theory of gases over in
silence. Soon he (1894/1954, p. 246) made known his doubts: he was
not sure that that theory could account for all the known facts. In a
later popular booklet Poincaré (1905/1970, pp. 210 and 251) softened
his attitude: physical laws will acquire an “entirely new aspect” and
differential equations will become statistical laws; laws, however, will
be shown to be imperfect and provisional.
3) Geometric probability. On its previous history see § 6.1.6; its
further development is described in Chapter 12. Here, I only indicate
that Poincaré explained the paradoxical nature of the Bertrand
problem (§ 11.1-1).
4) The binomial distribution. Suppose that m Bernoulli trials with
probability of success p are made and the number of successes is α.
Poincaré (pp. 79 – 84), in a roundabout and difficult way, derived (in
modern notation) E(α – mp)2 and E|α – mp|. In the first case he could
have calculated Eα2; in the second instance he obtained

E|α – mp| ≈ 2mpq Cmmp pmpqmq, q = 1 – p.

5) The Bayesian approach: estimating the total number (N) of the

asteroids. Poincaré (pp. 163 – 168) assumed that only M of them were
known and that, during a certain year, n minor planets were observed,
m of which were known before. Introducing a constant probability p =
n/N of observing an asteroid during a year and applying the Bayesian
approach, he obtained

EN ≈ n/p.

He was not satisfied with this pseudo-answer and assumed now that
p was unknown. Again applying the Bayesian approach and supposing
that p took with equal probability all values within the interval [0; 1],
he derived instead

EN = (M/m)n.

He could have written this formula at once; in addition, it was

possible to recall the Laplace problem of estimating the population of
France by sample data (§ 7.1-5). It is nevertheless interesting that

209
208
Poincaré considered the unknown number of the minor planets as a
random variable.
6) Without mentioning Gauss (1816, § 5), Poincaré (pp. 192 – 194)
derived the moments of the [normal] distribution

φ(y) = h /π exp(– hy2) (3)

obtaining

(2 p )!
Ey2p = (4)
h p p! 2 2 p

and proved, by issuing from formula (2), that the density function
whose moments coincided with the respective moments of the
[normal] law was [normal]. This proposition was, however, due to
Chebyshev (1887a), see also Bernstein (1945/1964, p. 420).
Then Poincaré (pp. 195 – 201) applied his investigation to the
theory of errors. He first approximately calculated E y 2p for the mean
y of a large number n of observations having Eyi = 0 and Eyi2 =
Const, equated these moments to the moments (4) and thus expressed
h through Eyi2. This was a mistake: y , being a mean, had a measure
of precision nh rather than h. Poincaré (p. 195) also stated that Gauss
had calculated E y 2; actually, Gauss (1823b, §15) considered the
mean value of ∑yi2/n.
The main point here and on pp. 201 – 206, where Poincaré
considered the mean values of (y1 + y2 + … + yn)2p with identical and
then non-identical distributions of the terms and Eyi = 0, was a non-
rigorous proof of the CLT: for errors of sensiblement the same order
and constituting une faible part of the total error, the resulting error
follows sensiblement the Gauss law (p. 206)3.
Also for proving the normality of the sum of errors Poincaré (pp.
206 – 208, only in 1912) introduced characteristic functions which did
not conform to their modern definition. Nevertheless, he was able to
apply the Fourier formulas for passing from them to densities and
back. These functions were

f(α) = Σpxeαx, f(α) =  φ(x)eαxdx (5)

and he noted that

f(α) = 1 + αEx/1! + α2Ex2/2! + … (6)

Markov (1898/1951, p. 269) referred, but had not commented on

Poincaré (1896, pp. 169 – 186 = 1912, pp. 189 – 206). I repeat that
there, on p. 173/194, Poincaré had applied his formula (2).
7) Homogeneous [Markov chains]. Poincaré provided interesting
examples that might be interpreted in the language of these chains.
a) He (p. 150) assumed that all the asteroids moved along one and
the same circular orbit, the ecliptic, and explained why they were

210
209
uniformly scattered across it. Denote the longitude of a certain minor
planet by l = at + b where a and b are random and t is the time, and,
by φ(a; b), the continuous joint density function of a and b. Issuing
from the expectation

Eeiml =  φ(a; b)eim(at + b)dadb

(which is the appropriate characteristic function in the modern sense),

Poincaré not very clearly proved his proposition that resembled the
celebrated Weyl theorem (beginning of § 10.8.4). The place of a
planet in space is only known with a certain error, and the number of
all possible arrangements of the asteroids on the ecliptic might
therefore be assumed finite whereas the probabilities of the changes of
these arrangements during time period [t; t + 1] do not depend on t.
The uniform distribution of the asteroids might therefore be justified
by the ergodic property of homogeneous Markov chains having a
finite number of possible states.
b) The game of roulette. A circle is alternately divided into a large
number of congruent red and black sectors. A needle is whirled with
force along the circumference of the circle, and, after a great number
of rotations, stops in one of the sectors. Experience proves that the
probabilities of red and black coincide and Poincaré (p. 148)
attempted to justify that fact. Suppose that the needle stops after
travelling a distance s (2π < s < A). Denote the corresponding density
by φ(x), a function continuous on [2π; A] and having a bounded
derivative on the same interval. Then, as Poincaré demonstrated, the
difference between the probabilities of red and black tended to zero as
the length of each red (and black) arc became infinitesimal (or, which
is the same, as s became infinitely large). He based his proof on the
method of arbitrary functions (Khinchin 1961, No. 2, pp. 88 –
89/2004, pp. 421 – 422; von Plato 1983) and himself sketched its
essence. Poincaré also indicated that the rotation of the needle was
unstable: a slight change in the initial thrust led to an essential change
in the travelled distance (and, possibly, to a change from red to black
or vice versa).
c) Shuffling a deck of cards (p. 301). In an extremely involved
manner, by applying hypercomplex numbers, Poincaré proved that
after many shuffling all the possible arrangements of the cards tended
to become equally probable. See end of § 7.1-6.
8) Mathematical treatment of observations. In a posthumously
published Résumé of his work, Poincaré (1921/1983, p. 343) indicated
that the theory of errors “naturally” was his main aim in the theory of
probability and that statement reflected the situation in those times. In
his treatise he (pp. 169 – 173) derived the normal distribution of
observational errors mainly following Gauss; then, like Bertrand,
changed the derivation by assuming that not the most probable value
of the estimator of the [location parameter] coincided with the
arithmetic mean, but its mean value. He (pp. 186 – 187) also noted
that, for small absolute errors x1, x2, …, xn, the equality of some f(z) to
the mean value of f(xi), led to the equality of z, the estimate of the true
value of the constant sought, and the arithmetic mean of xi. It seemed
211
210
to him that he thus corroborated the Gauss postulate4. Finally,
Poincaré (p. 188) indicated that the [variance] of the arithmetic mean
tended to zero with the increase in the number of observations and
referred to Gauss (who nevertheless had not stated anything at all
about the case of n → ∞, cf. § 9A.4-7). “Nothing”, however, followed
since other linear means had the same property, as Markov
(1899a/1951, p. 250) stated mentioning a wrong remark made by
Maievsky. Poincaré himself (1912, pp. 196 – 201 and 217) twice
proved the [consistency] of the arithmetic mean. In the second case he
issued from a characteristic function of the type of (5) and (6) and
passed on to the characteristic function of the arithmetic mean. He
noted that, if that function could not be represented as (6), the
consistency of the arithmetic mean was questionable, and he
illustrated that fact by the Cauchy distribution. Perhaps because of all
this reasoning on the mean Poincaré (p. 188) declared that Gauss’
rejection of his first substantiation of the MLSq was “assez étrange”
and corroborated this conclusion by remarking that the choice of the
[parameter of location] should not be made independently from the
distribution (and thus directly contradicting Gauss’ mature approach
which presumed unknown distributions). This statement corresponds
to modern thoughts, but the distributions are hardly known.
Poincaré (pp. 217 – 218) also stated that very small errors made it
impossible to obtain absolute precision even as n → ∞. If so, these
errors originate from the non-evenness of the law of distribution
(Bayes; see Stigler (1986, pp. 94 – 95) and Cournot (1843, § 137)),
the variability of that law (again Cournot) and, I would add, some
interdependence of the observations.
9) Randomness. Poincaré discussed randomness both in his treatise
and in his popular scientific booklets, but his various interpretations of
chance were not definitely enough compared one with another.
a) Instability of equilibrium or movement. Some of the statements
made by Aristotle (§ 1.1.1) and Galen (§ 1.1.3) meant that small
causes might lead to considerable consequences, and Maxwell
(§ 10.8.5-3), who described this phenomenon, apparently thought
about randomness but did not mention it.
Poincaré (p. 4) was the first to state directly that randomness was
instability of equilibrium or movement and he (pp. 4 – 5) provided a
few examples: the instability of a cone stood on its vertex formerly
mentioned by Cournot (§ 10.3-4); the roulette; the scattering of the
asteroids; unstable states of the atmosphere. His third example, just
like Newton’s reasoning on the irregularities in the Solar system
(§ 2.2.3), was nevertheless connected with great intervals of time.
Poincaré also argued that Laplace (§ 7.3), whom he did not name,
was wrong: forecasts of the future were impossible because of the
instability of motion. I have not found any connections between the
just described explanation of randomness and Poincaré’s study of
stability in mathematics or astronomy.
b) Complicated causes. Already Leibniz (§ 3.1.2) heuristically
explained randomness by the complexity of causes. Laplace (end of
§ 7.4) wrongly explained the existence of the eccentricities of
planetary orbits by the action of great many complicated causes.

212
211
Maxwell (§ 10.8.5-3) assumed that the distribution of the velocities of
molecules set in after a great number of collisions among a great
number of particles, but he did not mention randomness. And once
more Poincaré was the first to do so. He (pp. 7 – 8) maintained that
the molecular motion was random because of the combined action of
instability and complexity of causes, but he then mentioned the
shuffling of cards, the mixing of liquids and powders and (p. 15)
“even” of molecules in the kinetic theory of gases.
c) Small causes leading to small consequences. Poincaré (p. 10)
provided only one example, that, furthermore, did not belong to
natural sciences: small causes led to small errors of measurement; he
also indicated that these errors were considered random because their
causes were too complicated.
d) Intersection of chains of determinate events. I mentioned this
explanation in §§ 1.1.1 and 10.3-4. Poincaré (p. 10) allowed it, but his
first two explanations were his main ones; and he apparently forgot
here about the third one.
e) Randomness and necessity. Poincaré (see § 1.2.4) formulated a
highly proper idea on the combined action of randomness and
necessity. Regrettably, he had not mentioned the appearance of
necessity in mass random phenomena.
For Poincaré, the theory of probability remained an accessory
subject, and his almost total failure to refer to his predecessors except
Bertrand testifies, as mentioned by Bortkievicz, that he was not duly
acquainted with their work. Furthermore: in 1912 he was already able
to, but did not apply the arsenal of Markov chains. At the same time,
however, he became the author of a treatise that for about 20 years had
remained the main writing on probability in Europe. Le Cam’s
declaration (1986, p. 81) that neither Bertrand, nor Poincaré
“appeared to know“ the theory was unjust: he should have added that,
at the time, Markov was apparently the only one who did master
probability. On Bertrand see end of § 11.1.

Notes
1. The title-page of the French translation carried a phrase “Translated and
published avec l’autorisation de l’auteur”, but Bertrand himself (C. r. Acad. Sci.
Paris, t. 40, 1855, p. 1190) indicated that Gauss, who had died the same year, was
only able to send him “quelques observations de détail”.
2. Poincaré always applied the term “calcul” rather than “théorie” of probability.
It is hardly amiss to note that in 1882 – 1891 Markov had published five
mimeographed editions of his lectures called Theory of probability, but that he
called his treatise (1900 and later editions) Calculus of probability (in both cases, in
Russian). Another point: at least in 1892 Poincaré was not prepared to believe in the
statistical nature of the second law of thermodynamics; in addition to § 11.2-2
above, see Sheynin (1991a, p. 141).
3. For Poincaré the theory of probability remained a branch of applied
mathematics (§ 0.1 and Note 6 to Chapter 8).
4. In the same context Poincaré (p. 171) argued that everyone believed that the
normal law was universal: experimentalists thought that that was a mathematical
fact and mathematicians believed that it was experimental. Poincaré referred to the
oral statement of Lippmann, an author of a treatise on thermodynamics.

213
212
 12. Geometric Probability
On the development of the notion of geometric probability in the
18th century and earlier see § 6.1.6, and on its definition by Cournot
see § 10.3-4; I described the Bertrand problem on the length of a
random chord in § 11.1-1. Here, I discuss the further history of the
same notion.
1) Cournot (1843, § 74) applied geometric probability for deriving
the distribution of a function of several random arguments. Here is
one of his examples. The arguments of the function u = |x – y| are
uniformly distributed on segment [0; 1]. After calculating the areas of
the appropriate figures, he concluded that

P(u ≥ a) = (1 – a2), 0 ≤ a ≤ 1.

The determination of the probability of the contrary event would

have led Cournot to the once popular encounter problem (Laurent
1873, pp. 67 – 69): two persons are to meet at a definite spot during a
specified time interval, their arrivals are independent and occur “at
random”. The first one to arrive waits for a certain time and then
leaves. Determine the probability of the encounter.
2) Most eminent natural scientists of the 19th century tacitly applied
geometric probability. Boltzmann (§ 10.8.5-1) defined the time
average probability that the velocity of a molecule was contained in an
interval [c; c + dc] as the ratio of the time during which that event
took place to the total time of observation. I do not dwell on an earlier
definition of probability in physics or the further considerations
concerning the ergodic hypothesis. Maxwell (§ 10.8.5-2) applied
geometric probability while deriving his celebrated law.
When studying the life of earthworms, Darwin (1881/1945, pp. 52
– 55) strewed paper triangles over some ground. They were dragged
away by the worms but he recovered most of them and found out that
the worms had not seized “indifferently by chance any part” of the
triangles. He thought about several possibilities of [uniform]
“chance”, and, in particular, he decided, in actual fact, that the number
of times a worm would have seized “by chance” any side of a triangle
was proportional to its length1.
3) Seneta et al (2001) described the pertinent investigations of
Sylvester, Crofton and Barbier which led to the appearance of
integral geometry. I only mention Sylvester’s remarkable problem: To
determine the probability that four points taken “at random” within a
finite convex domain will form a convex quadrilateral. See Czuber
(1908/1968, pp. 99 – 102) for a few particular cases of that problem.
4) Czuber (1884, p. 11): Two points, M and N, are randomly
situated on segment AB = a. Determine the probability of MN > NA.
We have

P(x ≤ MA ≤ x +dx) = dx/a, P(MA > NA|x ≤ MA ≤ x +dx) = x/a.

 xdx /a  1/ 2.
2
P(MN > NA) =
0

214
213
 Indeed, suppose that the points A, M, N, B are situated in that order
from left to right. (The other possible case can be considered
likewise.) Then, if MN > NA, move M towards N until MN < NA,
which is always possible, QED.
5) Poincaré (1896, p. 97; 1912, p. 118) noted that the probability
that a point (x; y) was situated within some figure was equal to the
appropriate integral

 φ(x; y) dx dy

where φ should be somehow specified. He then went over to the

Bertrand problem (§ 11.1-1) but mentioned only two of its solutions
and provided his own reasoning tacitly assuming that φ ≡ 1. The chord
can be fixed with respect to the centre of the circle O and the polar
axis passing through, and beginning in O, by two parameters, ω and
α,– the polar angles of A, an endpoint of the chord, and of P, its
centre; or, by two other parameters, θ and ρ,– the polar coordinates of
P. Now, the integrals over the given circle

 dωdα ≠  dρdθ

and this, as Poincaré stated, explained the paradoxical nature of the

problem.
He also studied the probability that rotated figures satisfy certain
conditions but he did not state that this investigation was connected
with the Bertrand paradox, cf. § 12-7 below.
6) Czuber (1908/1968, pp. 107 – 108) discovered three more natural
solutions of the Bertrand problem.
a) One endpoint of the chord is fixed, and the chord passes
through any point of the circle; p = 1/3 + √3/2π ≈ 0.609.
b) Both endpoints of the chord are chosen randomly; this case
coincided with Bertrand’s first version.
c) Two points of the chord situated inside the circle are chosen
randomly; p = 1/3 + 3√3/4π ≈ 0.746.
7) It turned out (De Montessus 1903) that the Bertrand problem had
an uncountable set of answers. Suppose that Ox is the x-axis and mark
points D and C on its positive half,– its intersections with concentric
circumferences with common centre in point O and radii OD = 1/2
and OC = 1. Arbitrary

215
214
 Fig. 1. De Montessus (1903).
A point moves along the axis
from D to infinity, and,
correspondingly, the probability
sought in the Bertrand problem is
seen to have an uncountable set of
values. OD = 1/2, OC = 1.

points M2(x) and M3(x) are situated on the same semiaxis, between the
two circles and beyond the larger of them respectively. Tangents A2B2
and A3B3 to the smaller circumference pass through M2 and M3
respectively, and M3T is the tangent to the larger circumference with
point of contact T. Finally, M1(x) is an arbitrary point on the same
semiaxis inside the smaller circle.
For points M2 and M3 the probability sought is, respectively,

P2 = angle A2M2O/π = [2arcsin(1/2x)]/π,

P3 = angle A3M3O/angle TM3O = [arc sin(1/2x)]/[arcsin(1/x)],

with 1/2 ≤ x ≤ 1 and x ≥ 1 respectively.

When moving from point O in the positive direction (say), the
probability P2 decreases from 1 at point D to 1/3, and, from point C to
infinity, probability P3 increases from 1/3 to 1/2. It is rather difficult to
prove that P3 increases monotonically (and De Montessus had not
done it), but already for x = 1.01 and 1.1 it is 0.36 and 0.41
respectively and it reaches value (1/2 – 1/1,600 ) at x = 10.
Note that the coincidence of points M2 or M3 with D leads to
Bertrand’s first solution and the movement M3 → ∞ provides his
second case. His third solution concerned a point rather than a straight
line and was thus different.
De Montessus calculated the general mean probability of the
studied event. However, it was hardly proper to include in the
calculation, as he did, points such as M1 for which the stipulated
condition was certainly satisfied. More important, while calculating
the mean probability for the continuous case, De Montessus first
determined a finite sum, and, when adding together the appropriate
fractions, he added separately their numerators and their
denominators. Nevertheless, his mean probability (P = 1/2) was
correct and could have been established by noting that the studied
interval beyond the circle was infinite.
8) Borel (1909/1950, p. 132) solved the encounter problem (§ 12-1)
and on pp. 148 – 149 he indicated, without referring to anyone that
most of the natural methods of solving the Bertrand problem led to
P = 1/2.
9) Schmidt (1926) issued from Poincaré’s considerations and
indicated that the probability sought should persist under translation
and rotation of the coordinate system (invariance under reflection is
now also included). Accordingly, he proved that this condition is only
fulfilled for the (ρ; θ) coordinate system, see Item 5, and when

216
215
transforming that system into another one (with the appropriate
Jacobian being of course allowed for)2.
And so, commentators have finally agreed that the probability
sought was 1/2 which was tantamount to ignorance and could have
been stated from the very beginning.
For a modern viewpoint on geometric probability see M. G.
Kendall & Moran (1963); in particular, following authors of the 19th
century (e.g., Crofton 1869, p. 188), they noted that it might
essentially simplify the calculation of integrals. Then, Ambartzumian
(1999) indicated that geometric probability and integral geometry are
connected with stochastic geometry.

Notes
1. Darwin thus considered several possibilities of a “random” dragging of
triangles and in that sense his study forestalled the Bertrand problem on the length
of a random chord. Darwin attempted to ascertain whether or not the worms acted
somewhat intelligently, and he concluded that they had not seized the triangles
indifferently.
2. Prokhorov (1999b) believed that, from the geometrical point of view, the most
natural assumption in the Bertrand problem was that θ and ρ were independent and
uniformly distributed, 0 ≤ θ ≤ 2π, 0 ≤ ρ ≤ 1.

217
216
 13. Chebyshev
13.1. His Contributions
1) His Master’s dissertation (1845). It was intended as a manual for
students of the Demidov Lyceum in Yaroslavl and Chebyshev gave
there an account of the theory of probability barely applying
mathematical analysis; for example, he replaced integration by
summing. Already then, however, he consistently estimated the errors
of pre-limiting relations. The dissertation apparently had an addendum
published somewhat later, see § 13.1-2.
Incidentally, I doubt that this manual was a proper textbook. A
general survey of the theory of probability and its applications would
have been more useful the more so since Chebyshev’s reasoning was
necessarily burdensome.
2) The Poisson LLN (Chebyshev 1846); see Prokhorov (1986) for
a detailed exposition. Chebyshev solved the following problem. In n
[independent] trials the probability of success was p1, p2, …, pn.
Determine the probability that the total number of successes was µ.
By clever reasoning he obtained the formula

n  m 1
m( n  m)  ns   n(1  s ) 
m
1
P(µ ≥ m) ≤    
2 n m  ns  m  nm 

where m > ns +1 and s was the mean probability of success.

This result was interesting in itself and, in addition, it enabled
Chebyshev to prove the Poisson theorem, cf. § 8.7. He did not fail to
indicate the necessary number of trials for achieving a stipulated
probability of the approximation of the frequency of success µ/n to s.
As stated in the title of the memoir, Chebyshev had indeed not applied
any involved mathematical tools, but his transformations were
burdensome. His proof was rigorous (although he had not indicated
that the trials were independent) and he (1846) had the right to
reproach Poisson whose method of derivation did not provide the
limits of the error of his approximate analysis. See however the very
end of § 13.3. Later Chebyshev (1879 – 1880/1936, pp. 162 – 163,
translation p. 152) explicated one of his intermediate transformations
more clearly, also see Bernstein (1945/1964, p. 412).
3) The Bienaymé – Chebyshev inequality (cf. § 10.2-4).
Chebyshev (1867) considered discrete [random variables] with a finite
number of possible values; without loss of generality I simplify his
derivation by assuming that each of the n variables takes an equal
number of values. Chebyshev showed that

P{|Σ(ξi – Eξi)| < α  [Eξ 2

i  (Eξ i ) 2 ] } > 1 – 1/α2, α > 0. (1)

Unlike Heyde & Seneta (§ 10.2-4) I believe that Chebyshev derived

this inequality in about the same way as Bienaymé did, only in much
more detail. True, he restricted his attention to discrete variables
whereas Bienaymé, without elaborating, apparently had in mind the
continuous case; his memoir was devoted to the mathematical
treatment of observations. Modern authors, whom I mentioned in

218
217
§ 10.2-4, repeat the derivation for the latter instance; actually, already
Sleshinsky (1893) had done it.
Chebyshev immediately derived a corollary, which, in somewhat
different notation, was

lim P(Σ[|ξi – Eξi|/n] < ε) = 1, n → ∞

and he (1879 – 1880/1936, pp. 166 – 167, translation pp. 155 – 156)
specified this formula for the case in which the [random variables]
coincided one with another. Chebyshev thus obtained a most
important and very simple corollary: the arithmetic mean was a
[consistent] estimator of the expectation of a random variable. Both
corollaries assume that the expectations and variances of the
appropriate variables are uniformly restricted and Chebyshev had
indeed indicated this restriction (in another language). In the last-
mentioned source, and even earlier in another context, he (1867, p.
183) introduced indicator variables (taking values 0 and 1 with
respective probabilities) but not the term itself.
4) [The central limit theorem]. The title of the appropriate memoir
(1887b) mentions two theorems the first of which was the proposition
on the arithmetic mean (see Item 3) and Chebyshev only repeated its
formula. He then went on to the CLT noting that it “leads” to the
MLSq, – leads, as I comment, in accordance with the Laplacean
approach.
Chebyshev first of all referred to his inequalities for an integral of a
non-negative function whose moments up to some order coincided
with the same moments of the appropriate, in a definite sense, normal
distribution. He (1874) had published these inequalities without proof
and Markov (1884) and then Stieltjes substantiated them. Chebyshev
himself also justified them afterwards but without mentioning his
predecessors. A detailed history of these inequalities is due to Krein
(1951). Later Stieltjes (1885) expressed regret that he had overlooked
that Markov paper.
Chebyshev considered [random variables] u1, u2, …, un having
densities φi(x) and moments

Eui = 0, |Eui2| < C, |Eui3| < C, …

These conditions are not sufficient. The random variables ought to be

independent, and Chebyshev certainly thought so, but he had not
indicated the restriction

lim
 Eu 2
i
≠ 0, i = 1, 2, …, n, n → ∞. (2)
n

On the other hand, it was not necessary to demand that the moments
were uniformly bounded and Chebyshev possibly did not express such
a restriction. Here is Liapunov’s indirect testimony (1901b, p. 57): it
occurs that Chebyshev sometimes used the singular form instead of
the plural. Liapunov provided a few examples and, in particular,

219
218
quoted Chebyshev’s expression “the absolute value of the
mathematical expectations” from his formula of the CLT.
Chebyshev noted that the density f(x) of the fraction

x = Σui/√n (3)

can be determined by means of the multiple integral

f(x) =   ... φ1(u1)φ2(u2) …φn (un)du1du2 … dun (4)

extended over the values of the variables at which the fraction above
is situated within the interval [x; x + dx]. He multiplied both parts of
(4) by esx where s was some constant and integrated them over
(– ∞; + ∞) so that the right side became separated into a product of n
integrals with the same limits of integration. Chebyshev then
developed both parts in powers of s (the right side, after taking its
logarithm) and equated the coefficients of the same powers of that
magnitude to each other. Thus the integrals

 f(x)dx,  xf(x)dx,  x2f(x)dx, …

or, in other words, the moments of magnitude (3), were determined up

to some order (2m – 1). It occurred that, as n → ∞, again with the
same limits of integration,

 esxf(x)dx = exp(s2/2q2) (5)

where 1/q2 was the arithmetic mean of the second moments of ui and
it is here that the condition (2) was needed. Applying his previously
mentioned estimates of the integral of a non-negative function,
Chebyshev now completed his proof:

lim P(α ≤
u i
β

≤ β) = (1/√π)  exp(– x2)dx, n → ∞. (6)

2  Eu
2
i α

For finite values of n the same probability, as Chebyshev indicated

without a rigorous demonstration, was determined by a development
in polynomials now called after him and Hermite.
Bernstein (1945/1964, pp. 423 – 425) indicated that the
abovementioned expansion in powers of s diverged at |s| ≠ 0 and that
Markov (1898/1951, p. 268), when proving the Chebyshev theorem
anew, without explaining the situation had therefore introduced an
additional restriction,– not (2), but

lim Eun2 ≠ 0, n → ∞.

220
219
 In addition, Markov wrote out the expression that Chebyshev had
actually applied in his investigation:

m
 
lim 
 ui  1


 
  tmexp(– t2)dt, n → ∞. (7)
 2 Eui 2
 

Issuing from the Chebyshev inequalities, Markov also proved that

these expressions meant that the appropriate density tended to the
normal density whereas Chebyshev had apparently thought it evident.
For a detailed discussion see Kolmogorov’s commentary in
Chebyshev (1944 – 1951, vol. 3, 1948, pp. 404 – 409).
Sleshinsky (1892) turned his attention to the Chebyshev
demonstration of the CLT even before Markov did. He issued from
Cauchy’s findings (§ 10.1) and, as he stated on his p. 204, aimed at
simplifying (not specifying) his predecessor. In spite of
Freudenthal’s opinion and following Heyde & Seneta (1977, pp. 95 –
96), I think that the Cauchy investigation was nevertheless imperfect,
see end of § 13.1. And, once more repeating the last-mentioned
commentators, I note that Sleshinsky apparently proved the CLT
rigorously although only for a linear function of observational errors
having an even density. Liapunov (1900, p. 360) remarked that these
conditions were too restrictive. Like Chebyshev (§ 0.3, Note 1),
Sleshinsky maintained that his findings justified the MLSq [once
more: only in the Laplacean sense].
13.2. His Lectures
Chebyshev delivered lectures on the theory of probability at
Petersburg University from 1860 to 1882. In 1936, A. N. Krylov
published those read in 1879/1880 as recorded by Liapunov and I
refer to his publication by mentioning only the page numbers of this
source and of its translation of 2004.
In his Foreword Krylov declared that Liapunov had reproduced the
lectures “exactly as they were read, including all the fine points of his
accompanying remarks”. Prudnikov (1964, p. 183), however, thought
differently:
It was hardly possible to write down Chebyshev’s lectures minutely
and it is natural that their extant record is fragmentary.

This seems to be at least partly true. Krylov also indicated that he

had rewritten the Liapunov manuscript

In accordance with the new system of spelling at the same time

checking all the derivations …

I translated this book correcting perhaps a hundred (I repeat: a

hundred) mathematical misprints but I do not claim that I revealed all
of them. Ermolaeva (1987) briefly described a more detailed record of
Chebyshev’s lectures read during September 1876 – March 1878,
discovered by herself but still unpublished. She had not indicated
whether the newly found text essentially differed from the published
version and she is behaving like a dog in the manger.
221
220
 The lectures were devoted to definite integrals, the theory of finite
differences and the theory of probability. I discuss only their last
section but I begin by several general comments. Chebyshev
attempted to apply the simplest methods; for example, he used
summing, and, if necessary, went on to integration only at the last
moment; he introduced characteristic functions only in the discrete
case; as I mentioned above; he did not specify that he considered
independent events or variables; he was not interested in the
philosophical aspect of probability1; and, among the applications of
the theory of probability, he almost exclusively discussed the
mathematical treatment of observations.
1) The main notions. Chebyshev (p. 148, translation p. 141)
declared that the aim of the theory of probability was

To determine the chances of the occurrence of a certain event, and

he continued: the word ‘event’ means anything whose probability is
being determined, and probability serves to denote some magnitude
that is to be measured.

The use of “chance” and “probability” in the same sentence was

perhaps an elegant variation; in essence, however, Chebyshev made a
small heuristic step towards an axiomatic theory2. I also adduce a
modern formula (Prokhorov & Sevastianov 1999, p. 77): the theory
of probability studies mathematical models of random events and

Given the probabilities of some random events, makes it possible to

determine the probabilities of other random events somehow
connected with the first ones.

Chebyshev (p. 160/150) introduced an unusual and hardly useful

generalized definition of expectation, – of the expectation of the
occurrence of one out of several incompatible events. The sum of the
products of the type piai, as he stated, described these events by their
probabilities and the magnitudes “measuring” them. Note that he
mainly discussed discrete variables.
Tacitly following Laplace (§ 7.4), Chebyshev (p. 165/155)
indicated that the concept of limit in probability theory differed from
that in analysis. Still, I am unable to agree with such equalities (or
misprints?) as (pp. 167, 183, 204/156, 171, 190)

lim m/n = p. (8)

2) The limit theorem for Poisson trials (p. 167 and 201ff/156,
187ff). Determine the probability Pn, m that in n trials an event having
probabilities pi, i = 1, 2, …, n, respectively, occurred m times.
Applying a little known formula from the first section of his lectures
(p. 59/63)

π
Am = (1/2π) 
π
f[eφi]e–mφidφ, (9)

222
221
for the coefficients of the series

f(x) = Ao + A1x + A2x2 + … + Amxm + …,

Chebyshev obtained (qi = 1 – pi)


Pn, m = (1/2π) 

[p1eφi+ q1] [p2eφi + q2] …[pneφi + qn]e–mφidφ.

After some transformations when considering only small values of φ it

occurred that

π
Pn, m = (1/π)  exp(– nQφ2/2)cos[(np – m)φ]dφ
0

where p was the mean probability of success and Q = Σpiqi/n.

Assuming for large values of n an infinite upper limit in the obtained
integral, Chebyshev finally got

t
P[|m/n – p| < t 2Q /n ] = (2/√π)  exp(– z2)dz
0

(without the sign of limit!) and noted that formula (8), or, as he
concluded, the Poisson LLN, followed from it. He naturally did not
here admonish his predecessor.
3) The Bernoulli pattern (pp. 168 – 175/157 – 163). Chebyshev
wrote out the generating function of the binomial distribution (in usual
modern notation)

ΣPn, m tm = (pt + q)n, m = 0, 1, 2, …, n (10)

and calculated the appropriate expectation and variance by the modern

method (by differentiating this function once and twice etc.) although
without indicating its generality. He then repeated this derivation
otherwise. Assuming that in equation (10) t = eα, Chebyshev
multiplied both its parts by e–αpn, developed the exponential functions
in powers of α and equated the coefficients of α and then of α2. In
concluding, he (pp. 179 – 183/167 – 171) derived the local and the
integral De Moivre – Laplace limit theorems and (pp. 183 – 186/171
– 175) paid attention to the calculation of the integral of the
exponential function of a negative square. I note his unusual manner
which, in this case, becomes evident when he stated that the
abovementioned integral with the limits of integration being [u; + ∞)
was equal to the value of the integrand at the lower limit multiplied by
some proper fraction, – rather than by a real number situated in the
interval (0; 1).
4) A limit theorem for the multinomial distribution (pp. 205 – 207
and 214 – 218/190 – 193, 198 – 203). Chebyshev considered n trials
in each of which occurred one and only one event out of A1, A2, …, Ak
with event Ai meaning that some function took the value i. All the

223
222
events were equally probable so that each had probability 1/k.
Suppose that event Ai happened mi times, and

m1 + m2 + … + mk = n, P(m1 + 2m2 + ... + kmk = s) = Ps,

then

ΣPsts = tn(tk – 1)n/kn(t – 1)n (11)

after which Chebyshev determined Ps.

When considering the limiting case he expressed the right side of
(11) as

f (t) = Ao + A1t + A2t2 + … + Asts + …

and made use of the expression (9) so as to obtain

π
knPs = (1/2π) 
π
eφi(n–s){[ekφi – 1] ÷ [eφi – 1]}ndφ,

where the n-th power of the fraction was equal to

en(k–1)φi/2 [sin(kφ/2) ÷ sin(φ/2)]n

so that

π
1 φ
Ps =
π 
0
cos{([n(k – 1) – 2s]
2
}{[sinkφ/2] ÷ [ksinφ/2]}ndφ,

and, for large values of k, again without the sign of limit,

u
P(|s – kn/2| < ku n /6 ) = (2/√π)  exp (– t2)dt.
0

5) [The central limit theorem] (pp. 219 – 223/203 – 206). At the

time, Chebyshev had not yet known its rigorous proof. I only note his
pronouncement (p. 224/206): the formula that he obtained was not
derived

In a rigorous way […]. We have made various assumptions but did

not determine the boundary of the ensuing error. In its present state,
mathematical analysis cannot derive this boundary in any satisfactory
fashion.

6) Statistical inferences. Chebyshev solved two problems which,

however, were considered before him. In the first of these he (pp. 187
– 192/175 – 180) derived the Bayes limit theorem (§ 5.2). He
ascertained that, given the length 2w of the segment [b, c] and large

224
223
numbers p and q (notation as in formula (5.2)), probability P takes its
maximal value at

b = p/(p + q) – w, c = p/(p + q) + w

which is natural. After that Chebyshev (p. 142) had indeed derived
formula (5.3), but in his Lectures he had not mentioned Bayes at all.
In the second problem he (pp. 193 – 201/181 – 187) studied the
probability of a subsequent result in Bernoulli trials. An event
occurred m times in n trials; determine the probability that it will
happen r times in k new trials. Only guiding himself mostly by the
Stirling theorem, Chebyshev non-rigorously derived an integral limit
theorem similar to that obtained by Laplace (§ 7.1-5). His formula
(again without the sign of limit) was

 r m m 1 1 
t
m
P |  |  t 2 (1  ) (  )  = (2/√π)  exp(– z2)dz. (12)
 k n n n n k  0

Later Markov (1914a) indicated the same formula, although

correctly written down as a limit theorem. He hardly remembered its
occurrence in Chebyshev’s lectures.
7) Mathematical treatment of observations (pp. 224 – 252/207 –
231). Chebyshev (p. 227/209) proved that the arithmetic mean was a
[consistent] estimator of the unknown constant. Unlike Poincaré
(§ 11.2-7), he (pp. 228 – 231/209 – 212) justified its optimality by
noting that, among linear estimators, the mean ensured the shortest
probable intervals for the ensuing error. The variance of the arithmetic
mean was also minimal (Ibidem); although Chebyshev had not paid
special attention to that estimator of precision. In principle, he based
his reasoning on the definitive Gaussian substantiation of the MLSq
(§ 9A.4).
But at the same time Chebyshev (pp. 231 – 236/212 – 216) derived
the normal distribution as the universal law of error in about the same
way as Gauss did in 1809 (§ 9A.2). “The Gauss method”, Chebyshev
(p. 250/229) maintained, bearing in mind exactly that attempt later
abandoned by Gauss, was based on the doubtful “law of hypotheses”,
– on the “Bayes theorem” with equal prior probabilities. Chebyshev
several times censured that “law” when discussing the Bayesian
approach in his lectures. In this case, it is opportune to recall
Whittaker & Robinson’s remark in § 9A.2-2. I also note that
Chebyshev (p. 249/228) wrongly thought that the Gauss formula (9.6)
for the sample variance had only appeared “recently” and that it
assumed a large number of observations. Chebyshev indicated that he
considered only random errors having zero expectations, but he did
not mention that the Gauss formula provided an unbiased estimation.
It might be concluded that the treatment of observations hardly
interested him.
8) Cancellation of a fraction (pp. 152 – 154/144 – 146). Determine
the probability P that a “random” fraction A/B cannot be cancelled.

225
224
Denote the probability that a prime number m cannot be cancelled out
of A/B by pm. Then

P = p2 p3 p5 … pm.

Since the probability that A or B is divisible by m is 1/m (this was

an essential assumption, see comment below!),

pm = 1 – 1/m2,
P = (1 – 1/22) (1 – 1/32) (1 – 1/52) … (1 – 1/m2) …, (13)
1/P = 1 + 1/22 + 1/32 + 1/42 + … = π2/6, (14)

P = 6/π2.

Chebyshev did not explain the transition from product to series, but
it was known to Euler (1748, Chapter 15, §§ 275 – 277). Chebyshev
determined the sum (14) by two different methods. One of them
consisted in equating the coefficients of x2 in two different expansions
of ln(sinx/x) of which at least the second one was again known to
Euler (Ibidem, Chapter 9, § 158):

ln(1 – x2/6 + x4/120 – ...) = ln[(1 – x2/π2) (1 – x2/4π2) (1 – x2/9π2) ...].

Chebyshev also remarked that if a fraction cannot be reduced by 2,

3, or 5, then 1/19 < 1 – P < 1/20, which testifies once again that he
paid due attention to practical considerations3. Markov4 remarked that
Kronecker (1894, Lecture 24) solved the same problem but indicated
Dirichlet’s priority. Kronecker had not supplied an exact reference
and I was unable to check his statement; he added that Dirichlet had
determined the probability sought “if it existed at all”.
Bernstein (1928/1964, p. 219) refuted Chebyshev’s solution by
noting that his assumption led to contradiction. He also adduced
further considerations, and, in particular, indicated, on p. 220, that the
theory of numbers dealt with regular number sequences whose
limiting or asymptotic frequencies of numbers of some class, unlike
probabilities, “we will never determine experimentally”, might be
studied. See Postnikov (1974) on the same problem and on the
stochastic theory of numbers.
13.3. Some General Considerations
And so, Chebyshev argued that the propositions of the theory of
probability ought to be rigorously demonstrated and its limit theorems
should be supplemented by estimation of the errors of pre-limiting
relations (Kolmogorov 1947, p. 56). He himself essentially developed
the LLN and, somewhat imperfectly, proved for the first time the
CLT; on the study of these two issues depended the “destiny” of the
theory of probability (Bernstein 1945/1964, p. 411). His students,
Markov and Liapunov in the first place, also contributed to the
theory (§§ 14.1, 14.2, 14.4).
Kolmogorov continued:

226
225
 Chebyshev was the first to appreciate clearly and use the full power
of the concepts of random variable and [its] expectation.

I take issue with the full power. Indeed, Chebyshev had not made use
of Poisson’s heuristic definition of random variable (§ 8.2), had not
applied this term5 and did not study densities or generating functions
as mathematical objects. Then, the entire development of the theory of
probability might be described as an ever fuller use of the power of
the abovementioned concepts; thus, it had since began to study
dependent random variables, their systems and chains.
Here also is Bernstein’s conclusion (1945/1964, p. 432):

The genius of Chebyshev and his associates, who, in this field

[theory of probability], have left mathematicians of Western Europe
far behind, have surmounted the crisis of the theory of probability that
had brought its development to a stop a hundred years ago.

“Crisis” may be understood as a dangerous and unstable state; in

this case, as the theory’s extremely unfavourable state as compared
with the main branches of mathematics then rapidly developing in
Europe. Two circumstances ought to be mentioned here. First,

In spite of his splendid analytical talent, Chebyshev was a

pathological conservative.

This is the opinion of Novikov (2002, p. 330) who corroborated it

by referring to V. F. Kagan (1869 – 1953), an eminent geometrician.
The latter,

When being a young Privat-Docent, had listened to Chebyshev’s

scornful statement on the trendy disciplines like the Riemann
geometry and complex-variable analysis.

Even Liapunov (1895/1946, pp. 19 – 20), who (Bernstein

1945/1964, p. 427)

Understood and was able to appreciate the achievements of the

West European mathematicians, made in the second half of the [19th]
century, better than the other representatives of the [Chebyshev]
Petersburg school,

called Riemann’s ideas “extremely abstract”; his investigations,

“pseudo-geometric” and sometimes, again, too abstract and having
nothing in common with Lobachevsky’s “deep geometric studies”.
Liapunov obliquely recalled Klein, but disregarded him. Klein had in
1871 presented a unified picture of the non-Euclidean geometry in
which the findings of Lobachevsky and Riemann appeared as
particular cases.
On the other hand, Tikhomandritsky (1898, p. IV) testified that in
1887 he had shown Chebyshev his “course” and that the latter “stated
that […] it is necessary to transform the entire theory of probability”.

227
226
It is difficult to say what exactly did he mean. His words must have
been known; I found later references to them (Maciejewski 1911, p.
87; Gnedenko & Gikhman 1956, p. 487). On the Petersburg school
of the theory of probability see also Bernstein (1940).
But why neither Poisson nor Poincaré had not attempted to prove
rigorously their statements? With respect to the CLT Poisson
apparently was unable to act differently, but I have indicated the main
reason in Notes 5 and 6 to Chapter 8 and in § 7.4. Neither densities
nor characteristic functions had hardly been treated as mathematical
objects.
And the destiny of probability theory not less depended on the
proper handling of those two objects which had only occurred in the
1920s (P. Lévy). True, Poisson (1837a, p. 1) indicated that in the 18th
century did the calculus (not the theory) of probability

Wholly extend and became one of the main branches of

mathematics owing both to the number and utility of its applications
and the type of the analysis which it beget.

Still, this does not contradict the statements above. And here is a
suitable statement of Markov from his report of 1921 (Sheynin 2006a,
p. 152):

It was usual to consider the theory of probability as an applied

science in which mathematical rigour was superfluous.

A special comment on Chebyshev is warranted. For many yedars

had had been associated with problems in artillery (Prudnikov 1950)
and certainly with, and essentially influenced Maievsky, the most
eminent specialist in ballistics, and the founder of modern external
ballistics, see his book of 1872 (Mandryka 1954, p. 162).
Chebyshev published a note (1870) in the memoirs of the Belgian
academy which was favourably discussed by mostly French authors. It
apparently had not contained any essentially new material as
compared with his previous publications but nevertheless it should
have been, but was not mentioned in his collected works (1899 –
1907; 1944 – 1951). I have almost completely reprinted it in one of
my collections (S, G, 78) as Mayevski (1872) also did long before me.
However, I have also described the comments mentioned above. In
somewhat more detail see the same in my Russian collection S, G, 82.
As a sideline I add that both Chebyshev and Mayevski had been
applying the term mean square error so that it seems that it had thus at
least found its way into Russian geodesy.
And here are a few lines from Maievsky (1872, p. XII):

Chebyshev provides formulas for interpolating by the method of

least squares. We have applied them for determining the projection of
the path [of the shell] on the vertical plane of the firing by issuing
from its results.

228
227
 Notes

1. Prudnikov (1964, p. 91) quoted a paper of V. A. Latyshev, an educationalist

and a student of Chebyshev, published in 1893:

One of the most distinguished [Russian] mathematicians […] had the habit of
expressly telling his students that he did not advise [them] to engage in the
philosophical aspect of mathematics since this was not very helpful for acquiring the
knowledge of mathematics, and even rather harmful.

Prudnikov added that Latyshev had certainly meant Chebyshev. Recall (§ 5.1)
that Chebyshev formulated the problem on the next sunrise in everyday language.
2. Chebyshev (1845/1951, p. 29) provided a similar definition of the aims of the
theory of probability much earlier. It is hardly amiss to remark that for Laplace the
theory served for discovering the laws of nature. Boole (1851/1952, p. 251)
expressed ideas similar to those formulated by Chebyshev:

The object of the theory of probabilities may be thus stated: Given the separate
probabilities of any propositions to find the probability of another proposition.

He (1854, p. 288) was also the first to argue that the theory should be
axiomatized, see Supplement. On Boole’s probability see Hailperin (1976) who
does not, however, dwell on axiomatization.
3. Note however Chebyshev`s unqualified statement (1879 – 1880/1936, p. 214,
translation p. 198): Different lotteries are equally fair if the expected gains are the
same and equal to the [equal] stakes. This contradicts the reasonable opinion of both
D’Alembert and Buffon (§§ 6.1.2 and 6.1.4) that a low probability of a single
[favourable] event be disregarded.
4. I refer to the German translation of his Treatise (1912, p. 148) from the Russian
edition of 1908 and to p. 241 of its last edition of 1924.
5. The term “random quantity” appeared at the end of the 19 th century (Vasiliev
1885, pp. 127 – 131; Nekrasov 1888, p. 77) whereas the English expression
“random magnitude” was possibly introduced later (Whitworth 1901, p. 207). I had
not, however, seen the previous editions of that book. The present English term is of
course random variable.

229
228
 14. Markov, Liapunov, Nekrasov
I consider here the work of three outstanding scholars; with regard
to Markov and Nekrasov, however, qualification remarks will
follow.
14.1. Markov: General Scientific Issues
I consider his main results in § 14.2; here, I briefly treat some
additional matters, but see § 15.1.3 for his study of statistical series.
Finally, in § 14.3 I describe Markov’s personality.
1) History of the theory of probability. Markov undoubtedly paid
attention to it. He investigated the Bernoulli LLN (§ 3.2.3); in 1913
he initiated a jubilee meeting of the Petersburg Academy of Sciences
celebrating the bicentenary of that law, as well as the publication of a
Russian translation of pt. 4 of the Ars Conjectandi (see § 3). Markov
left several statements about the history of the Bienaymé –
Chebyshev inequality (§ 10.2-4), somehow argued for the second
Gauss justification of the MLSq (as mentioned in § 9A.6-1),
introduced an apt term, “De Moivre – Laplace limit theorem”
(Treatise 1924, p. 53) and stressed De Moivre’s part in establishing
the “Stirling formula”. This last edition of his Treatise includes many
interesting historical remarks. As far as his number-theoretic papers
collected in his Selected works (1951) are concerned, I can at least add
that they contain many references to his predecessors.
2) Insurance of life. In his Treatise (1900a), Markov described the
pertinent theory but did not add any new findings. However, he
actively collaborated with pension funds scrupulously considering all
practical details of their work (Sheynin 1997d), and in 1906 he
published two newspaper articles destructively criticizing a proposed
scheme for insuring children (reprinted in the same later translated
article).
3) Calculations. “Markov liked calculating and was good at it”
(Linnik et al 1951, p. 615). In the theory of probability, most
important is his table of the normal distribution (1888) giving it to 11
digits for the argument x = 0 (0.001) 3 (0.01) 4.8 and adduced the
differences of all the necessary orders (for example, with the first
three differences for x ≤ 2.649). According to a reputed reference book
(Fletcher et al 1962), two tables of the normal distribution, one of
them Markov’s, and the other, published ten years later, remained
beyond compare up to the 1940s. In an indirect way, Markov (1899b,
p. 30) made known his attitude toward calculation:

Many mathematicians apparently believe that going beyond the

field of abstract reasoning into the sphere of effective calculations is
humiliating.

4) Correlation theory. In § 10.6-3 I indicated that statisticians had

doubted it. The same was true with regard to Markov. Slutsky (1912a)
had collected and generalized the relevant findings of the Biometric
school, and even a few decades later Kolmogorov (1948) called his
book important and interesting. Markov, however, did not duly
estimate it. He mentioned it in three letters to Chuprov, all written in
1912 (Ondar 1977/1981, pp. 53 – 58), and he (p. 53) stated that it

230
229
interested, but did not “attract” him, and (p. 58) did not “like it very
much”.
Also in 1912, Slutsky exchanged a few letters with Markov and, in
particular (Sheynin 1990c/2011, p. 64) stated:

I believe that the shortcomings of Pearson’s exposition are

temporary and of the same kind as the known shortcomings of
mathematics in the 17th and 18th centuries. A rigorous basis for the
work of the geniuses was built only post factum, and the same will
happen with Pearson. I took upon myself to describe what was done.
Sometime A. A. Chuprov will set forth the subject of correlation from
the philosophical and logical point of view, and describe it as a
method of research. An opportunity will present itself to a ripe
mathematical mind to develop the mathematical basis of the theory.

In a few years Markov (1916a/1951, p. 533) critically mentioned

the correlation theory:

Its positive side is not significant enough and consists in a simple

usage of the method of least squares to discover linear dependences.
However, not being satisfied with approximately determining various
coefficients, the theory also indicates their probable errors and enters
here the region of fantasy, hypnosis and faith in such mathematical
formulas that, in actual fact, have no sound scientific justification.

Now, discovering dependences, even if only linear, is indeed

important; and the estimation of plausibility of the results obtained is
an essential part of any investigation. True, at the time such estimation
had not been done properly. Considering a paper of a contemporary
Russian author, Markov (Ibidem pp. 534 – 535) pointed out an
obvious senselessness: the calculated correlation coefficient was 0.09
with probable error 0.14. In addition, these figures greatly changed
when Markov left aside some of the observations made use of.
However (Linnik; see his comment on that paper (Markov 1951, p.
670), without knowledge of the distribution of the population, the
sample correlation coefficient cannot properly estimate the general
coefficient.
Nevertheless, I quote Bernstein (1928/1964, p. 231):

Excluding biological applications, most of its [of the correlation

theory] practical usage is based on misunderstanding.

He had not regrettably elaborated.

5) Principles of the theory of probability. In essence, Markov left
that issue aside. Thus, in the German edition of his Treatise (1912,
translated from the Russian edition of 1908, p. iii) he declared that he
did not discuss it in detail. At about the same time, he (1911c/1981,
pp. 149 – 150) pessimistically estimated suchlike efforts:

I shall not defend these basic theorems connected to the basic

notions of […] equal probability, of independence of events, and so

231
230
on, since I know that one can argue endlessly on the basic principles
even of a precise science such as geometry.

Markov (Treatise, 1908, p. and 1924, c. 2) also stated, somewhat

indefinitely, that

Various concepts are defined not by words, each of which can in

turn demand definition, but rather by [our] attitude towards them
ascertained little by little.

Apparently, some (not various) concepts must be admitted without

definition; there are no modern definition either of point or straight
line.
It ought to be added, however, that, except for the axiomatic
approach, only Mises was able to abandon the classical definition of
probability (see also § 7.4). Note, however, that Markov, like a student
of Chebyshev, underrated the then originating axiomatic direction of
probability as well as the theory of functions of a complex variable
(A. A. Youshkevich 1974, p. 125).
On p. 10 of his Treatise (1924) Markov formulated the following
axiom: If there are several equally possible events, some of them
favourable, the others not, with regard to event A, then, after A occurs,
the unfavourable events “fall through” whereas the others remain
equally possible. I do not see how can it be otherwise. Then, on pp. 13
– 19 Markov proved the addition and the multiplication theorems (in a
rather complicated way and mentioning his axiom) and on p. 24
concluded that these theorems along with his axiom serve

As an unshakeable base for the calculus of probability as a chapter

of pure mathematics.

So here we are! His axiom (many pages apart, not displayed in the
text and difficult to find!), never mentioned by any later author,
allegedly transformed the theory of probability1 …
6) Mathematical statistics. In 1910 Markov (Ondar 1977/1981, p. 5)
had denied Pearson, but by the end of his life he somewhat softened
his attitude. Here is a passage from Chuprov’s letter, written
apparently in 1924, to Isserlis (Sheynin 1990c/2011, p. 76):

Markov regarded Pearson, I may say, with contempt. Markov’s

temper was no better than Pearson’s, he could not stand even slightest
contradictions either [§ 14.3]. You can imagine how he took my
persistent indications to the considerable scientific importance of
Pearson’s works. My efforts thus directed were not to no avail as
proved by [Markov 1924]. After all, something [Pearsonian] was
included in the field of Markov’s scientific interests.

Chuprov (1925b/1981, p. 155) also published a review of the

mentioned edition of Markov’s Treatise. Here, I only cite his
reasonable criticism of Markov’s treatment of correlation theory:

232
231
 The choice of questions on which attention is concentrated is
fortuitous, their treatment within the bounds of the chapter on the
method of least squares is incomplete, the connection made between
the theory of correlation and the theory of probability is inadequate…

Now, what statistical innovations had Markov included in this last

edition of his Treatise? A study of statistical series and of the
Pearsonian correlation theory (§ 14.2-1). He considered linear
correlation and applied the MLSq for determining the parameters of
the lines of regression and discussed the case of [random variables]
with densities of their distributions being quadratic forms. Markov
also included a general reference to Slutsky (1912a) and certainly did
not repeat his earlier harsh words about imagination, hypnotism, etc.
Below (§ 14.2-1) I shall add that Markov paid no attention either to
the chi-squared test or to the Pearsonian curves.
7) Teaching probability theory in school. In 1914 Nekrasov made
an attempt to introduce probability into the school curriculum.
Markov, who could not stand him at all, either as a man, or as a
mathematician, was not invited to the pertinent discussion by
correspondence, but he voiced his opinion in an ad hoc paper (1915a).
He sharply protested against the concrete school programme proposed
by Nekrasov, but, as I understand him, did not object to the very
principle. In 1914, he published a relevant newspaper article
(reprinted in Sheynin 1993a, p. 200), and in 1916 he was member of
the Commission established by the Academy of Sciences to study
Nekrasov’s proposal. Its review was extremely negative (Report 1916)
both with respect to Nekrasov’s programme and to his understanding
of the main issues of mathematical analysis, cf. Nekrasov’s statement
about the concept of limit in § 14.5.
8) Methodological issues. Many authors praised the methodological
value of Markov’s contributions. Bernstein (1945/1964, p. 425) stated
that Markov’s Treatise and memoirs were “specimens of preciseness
and lucidity of exposition”. Linnik et al (1951, p. 615) maintained
that Markov’s language was distinct and clear, and that he thoroughly
trimmed the details. A striking example proving the opposite is
Markov’s failure to discuss the adjustment of direct conditional
observations (§ 9A.4-9) in his Treatise. And I do not trust Chuprov
(1925b/1981, p. 154) who thought that the exposition in Markov’s
Treatise was “transparently clear”.
Except Markov himself (§ 14.2-1) the only author with whom I
agree is Idelson (1947, p. 101). He remarked that the chapter on the
MLSq in the Treatise (1924) was ponderously written. Indeed,
Markov’s general rule was to rewrite his formulas rather than to
number, and then to refer to them. Thus, on pp. 328 – 330 of the
Treatise a long equality appeared five times in succession! Then, he
disregarded demonstrative pronouns. On p. 328, for example, he
wrote: “The choice of coefficients [a displayed line of them followed]
is at our disposal. We shall subject the coefficients [the same line was
repeated] to two conditions …”
Then, Markov refused to apply the term random magnitude (as it
has been called in Russia), see § 14.2-1, and the expressions normal

233
232
law and coefficient of correlation were likewise absent in his works.
As the Russian saying goes: “The whole company broke step, the
lieutenant alone is in step”. Markov’s literary style was pedestrian and
sometimes hardly understandable (1906/1951, p. 341) and, from one
edition to another, the structure of his Treatise became ever more
complicated. A few more remarks are in §14.2-1.
14.2. Markov: His Main Investigations
1) Mathematical treatment of observations. Linnik et al (1951, p.
637) believed that, when substantiating the MLSq, Markov had “in
essence” introduced concepts “equivalent” to the modern concepts of
unbiased and effective statistics for estimating parameters of the laws
of distribution. Markov, however, only indirectly estimated
parameters (he never used such an expression), and, which is more
important, it is just as possible to attribute those concepts to Gauss.
Nor do I agree with Idelson (1947, p. 14) who mentioned the Gauss
method, developed by Markov “up to the highest logical and
mathematical perfection”. In § 9A.6-1 I mentioned Markov’s resolute
stand for the second Gauss substantiation of the MLSq; this (and his
remark on the [consistency] of the arithmetic mean being inadequate,
see § 11.2-8) is all that he really accomplished here. Neyman (1934,
p. 595) erroneously attributed that justification to Markov and F. N.
David & Neyman (1938) repeated that statement, but then Neyman
(1938/1952, p. 228) recognized his mistake. Nevertheless, the
mysterious Gauss – Markov theorem is still alive and kicking. It was
Scheffé (1959, p. 14) who introduced that term (H. A. David, see H.
A. David & Edwards 2001, p. 218) although Plackett (1949) had noted
that mistake.
And I add now that Markov (1899a/1951, p. 246) had at the same
time denied any optimality of the MLSq, so why did he feel necessary
to substantiate it at all?
In his Treatise (1900) Markov in essence combined the treatment of
observations with the study of correlation (§ 14.1-4), statistical series
and interpolation; this, perhaps, reflected his attempt to include the
MLSq into the then originating mathematical statistics, but his
innovation was methodically doubtful.
The discussion of statistical series was rather involved for an
educational aid and did not mention Chuprov’s relevant papers
(1916; 1918 – 1919) the first of which Markov himself had
communicated to the Izvestia of the Petersburg Academy of Sciences.
Note that Chuprov (1925b/1981, pp. 154 and 155) politely remarked
that Markov had left out the works of other authors not belonging to
the “stream” of his own contributions. I would say that this criticism
was too mild.
In connection with statistical series Markov (Treatise 1924, pp. 349
– 353) considered Weldon’s experiment with 26,306 throws of 12 dice
(K. Pearson 1900) and decided, after applying the CLT and the Bayes
theorem with transition to the normal law, that the probability of a 5 or
a 6 was higher than 1/3. Unlike Pearson, he had not used the chi-
squared test and he could have left an impression that (although
suitable for a small number of trials as well) it was not needed at all.

234
233
 As to interpolation, the only point of contact with the MLSq was
the calculation of the empirical coefficients according to the principle
of maximal weight.
Markov passed over in silence the Pearsonian curves perhaps owing
to their insufficient substantiation. However, again in 1924, he
reprinted the Introduction to the edition of 1913 where he had stated
that the use of approximate methods in applied mathematics was
unavoidable even when an estimation of their error was impossible,
and (1915a, p. 32) maintained that “empirical” formulas did not
demand theoretical proof. I think that Markov followed here (as he did
with his chains which he left without any applications to natural
sciences) his own rigid principle hardly worthy of exact imitation
(Ondar 1977/1981, Letter 44 to Chuprov of 1910): “I shall not go a
step out of that region where my competence is beyond any doubt”.
The explication of the MLSq proper was involved, and Markov
himself knew it. In a letter of 1910 to Chuprov he (Ibidem, p. 21)
wrote: “I have often heard that my presentation is not sufficiently
clear”. In 1893, his former student, Koialovitch (Sheynin 2006b, p.
85), writing to Markov, formulated some puzzling questions:

As far as I understand you, you consider each separate observation

as a value of a possible result. Thus, a series of results […] is possible
for each measurement, and one of them is realized. I am prepared to
understand all this concerning one observation. However, if there are,
for example, two observations, then I cannot understand the difference
between the series of all the possible results of the first observation
[…] and the similar series for the second measurement […] The
problem will certainly be solved at once if you say that the
probabilities of the same error in these two series are different, but
you will hardly want to introduce the notion of probability of error in
your exposition.

See also Ermolaeva (2009) and my Russian contribution on

Koialovich in S, G, 85. The situation had not improved with time.
Contrary to what he himself (Treatise 1924, pp. 323 and 373) stated
while following Chebyshev (1879 – 1880/1936, p. 227, translation p.
208), Markov (1924, pp. 327 and 374) maintained that only one
possible observation corresponded to each actually made. He never
clearly explained that observational errors were [random variables]
and that a series of observations was a [random] sample and had a
density function. I note that Laplace, in the very beginning of § 23 of
his Analytic Theory of probability, mentioned numerous and not yet
made observations…
Wherever possible, Markov (Ondar 1977/1981, Letter 53 to Chuprov
of 1912) excluded “the completely undefined expressions random and
at random”. Instead, he added, he introduced an appropriate
explanation in each particular case. However, at least sometimes he
simply wrote indefinite which was much worse; incidentally, the
translators of Ondar (1977) modernized Markov’s letters by rendering
indefinite as random. Cf. Note 5 to Chapter 13.

235
234
 The chapter on the MLSq in Markov’s Treatise was hardly inviting
either for mathematicians or geodesists. Both would have been
disappointed by the lack of discussion of Pearson’s work whereas the
latter, in addition, had not needed interpolation or investigation of
statistical series but would have wished to see much more about
correlation. And the absence of the Gauss brackets (Note 19 in my
Chapter 1) as well as the appearance of the long-ago dated term
practical geometry instead of geodesy (p. 462) would have annoyed
them.
I also mention that Markov destructively criticized a paper
(Galitzin 1902) devoted to the study of the solidity of glass tubes. His
review was extant as a manuscript and I (Sheynin 1990b) published it.
Markov had not applied any new method, but he thoroughly treated
Galitzin’s data and allowed for every possible circumstance. It was in
connection with the discussion of Galitzin’s paper that Markov stated
his opinion about the “Bredikhin rule” (§ 10.8.4).
2) The LLN. Markov (1906/1951, p. 341) noted that the condition

1
limE [(ξ1 + ξ2 + … + ξn) – (Eξ1 + Eξ2 + … + Eξn)]2 = 0, n → ∞ (1)
n2

was sufficient for the sequence ξ1, ξ2, …, ξn, … of [random variables]
to obey the LLN; or, in accordance with his formula, to comply with
the condition

limP{|(ξ1 + ξ2 + … + ξn ) – (Eξ1 + Eξ2 + … + Eξn)| < ε} = 1, n → ∞.

Then Markov (Ibidem, pp. 342 – 344; Treatise, 1913, pp. 116 –
129) derived a few relevant sufficient conditions for sequences of
independent, and, especially, dependent random variables and
(1906/1951, p. 351; Treatise 1913, p. 119; Treatise, 1924, p. 174)
provided examples of sequences not obeying the law, and, in addition
(Treatise, 1913, p. 129), proved that apart from condition (1),
independent variables obeyed the LLN if, for every i, there existed the
moments

Eξi = ai, E|ξi – ai|1+δ < C, 0 < δ < 1.

In connection with his investigations of the LLN Markov (Treatise,

1900; p. 86 in the edition of 1924) had proved that, for a positive
random variable ξ,

P (ξ ≤ t2Eξ) > 1 – 1/t2

and Bortkiewicz (1917, p. 36) and Romanovsky (1925a; 1925b)

called this inequality after Markov.
3) [The CLT]. As I mentioned at the end of § 13.1, Markov
specified the conditions of theorem (13.6) proved by Chebyshev. He
(1898/1951, p. 268) considered independent [random variables] ui
with zero expectations2 and, following Chebyshev, supposed that, for
finite or3 infinite values of k,

236
235
 lim|Eunk| < + ∞, n → ∞. (2)

In addition, Markov, however, demanded that

limEun2 ≠ 0, n → ∞. (3)

He several times returned to the CLT.

a) For equalities (13.7) to hold, he (1899a/1951, p. 234) assumed
condition (3) and, for the transition to the new theorem he (p. 240)
additionally introduced two restrictions: as n → ∞,

limE[(u1 + u2 + … + un)2] = ∞,
lim[E(u1 + u2 + … + un)2/n] ≠ ε. (4; 5)

b) Later Markov (1907, p. 708) again proved formula (13.7).

Referring to his papers (1898; 1899a), he now introduced conditions
(2) (for finite values of k) and (5) but did not restrict the values of ui.
On his next page Markov abandoned condition (5) “if only”

limEun2 = ∞, n → ∞ (6)

and the values of ui remained finite. Restrictions (4) and (6) certainly
coincided.
Finally, Markov (1908a) essentially extended the applicability of
the method of moments by replacing his conditions by Liapunov’s
single restriction (1901a/1954, p. 159)

E | u | i
2

 var u 
lim 2 1  / 2
= 0, δ > 0, n → ∞. (7)
i

In 1913 Markov included a modified version of his last-mentioned

study in his Treatise (Markov 1924; reprint: 1951, pp. 319 – 338).
In connection with condition (3) Markov (1899c, p. 42) mentioned
the example provided by Poisson (1824, § 10). The latter proved that
the limiting distribution of the linear form

L = ε1 + 1/3ε2 + 1/5ε3 + …

of [random variables] εi with density e–2|х| was

lim P(|L| ≤ с) = 1 – (4/π)arctane–2c, n → ∞.

In this example

lim var[εn/(2n – 1)] = 0, n → ∞.

Markov himself (1899a/1951, pp. 242 – 246) also provided an

example in which the condition (3) had not held and the CLT did not

237
236
take place and he mentioned Poisson without adducing the exact
reference.
The appearance of condition (5) remains, however, unclear.
Nekrasov (1900 – 1902, 1902, pp. 292 and 293) introduced it for
independent variables instead of restriction (3). Liapunov
(1901a/1954, p. 175) maintained that it was not sufficient (but was he
then acquainted with the third part of Nekrasov’s contribution?) and
mentioned Markov’s examples. Seneta (1984, p. 39) indicated,
however, that Markov’s published papers had not contained such
examples and that condition (3) was necessary and sufficient for the
CLT in the case of uniformly restricted variables.
4) Markov chains. This term is due to Bernstein (1926, §16);
Markov himself (1906/1951, p. 354) called them simply chains. He
issued from a paper by Bruns of the same year, but the prehistory of
Markov chains is much richer. Here are the main relevant issues.
a) The Daniel Bernoulli – Laplace urn problem, the predecessor of
the Ehrenfests’ model (§ 7.1-3).
b) The study of the Brownian motion (Brush 1968).
c) The problem of the extinction of families (§ 10.2-7).
d) The problem of random walks (Dutka 1985).
e) Some of Poincaré’s findings (§ 11.2-6).
f) The work of Bachelier, for example, Bachelier (1900) on
financial speculations, also see Courtault et al (2000) and Taqqi
(2001).
Markov (1906/1951, pp. 345 and 354) considered simple
homogeneous chains of random events and discrete [random
variables] and proved that the LLN was applicable both to the number
of successes and to the sequences of these variables. Later he
(1910/1951, p. 476) extended the first of these findings to simple
nonhomogeneous chains.
Markov proved the CLT for his chains by means of condition
(13.7). He considered simple homogeneous chains of events (1906)
and of [random variables] (1908b); simple nonhomogeneous (1910)
and complex homogeneous (1911a; 1911b) chains of [random
variables]; simple homogeneous chains of indirectly observed events
(1912a). While studying the chains, Markov established important
ergodic theorems but had not paid them any special attention; in this
connection, I mentioned one of his solved problems in § 7.1-3.
It is difficult to imagine that Markov had not grasped the essential
importance of the chains for various applications, but he did not say
anything about that, and his only relevant and mostly methodical
example (Treatise, 1913) was a study of the alternation of consonants
and vowels in the Russian language, see Petruszewycz (1983). In 1910
Markov himself, in his letters to Chuprov, remarked more than once
that he was restricting his field of work by what was well known to
him, see § 14.2-1. I note also that at the time physics was not yet duly
studied in Russia (Kolmogorov 1947, p. 59).
5) In concluding, I ought to add that Markov widely applied the
method of moments, and only he who repeats some of his
investigations (for example, of his study of the limiting behaviour of
the terms obtained by decomposing algebraic fractions), will be able

238
237
to estimate the obstacles which he overcame. Bernstein (1945/1964,
p. 427) contrasted Markov and Liapunov. The latter had applied the
classical transcendental analysis as developed by that time whereas
the method of moments, Bernstein maintained,

Did not facilitate the problem [of proving the CLT] but rather
transferred all its difficulties elsewhere.

Soon Bernstein (1947, p. 44) remarked that Chebyshev had

expressed a negative opinion about the method of moments for
characteristic functions which in those times had not yet answered the
requests of mathematical rigour.
It might be imagined, however, that Markov wished to ascertain
how powerful was the method of moments; he himself (Treatise,
1913, p. 322) indicated that Liapunov had “shaken” the importance of
the method of moments and that he, Markov, therefore decided to
prove the CLT anew (see above).
Liapunov (1900/1954, p. 125) called the Markov proof of the CLT
too complicated and unwieldy because of its connection with a special
theory. However, Krein (1951, pp. 8 – 9) noted that the method of
moments had not yet become useless. It is still being applied (A. D.
Soloviev 1997/2008, p. 355):

Considerably improved since Markov’s time, [it] is still used […]

for solving such problems in which the moments of random variables
are derived much easier than their distributions.

14.3. Markov: His Personal Traits

The sequel will substantiate my decision to discuss Markov’s
personality at the end of this section. For his biography see Markov
Jr (1951), a noted mathematician in his own right, and Grodzensky
(1987). These authors describe many cases of his principled stand on
burning social and political issues whereas Grodzensky published
many pertinent newspaper letters4. In particular, Markov struggled
against anti-Semitism (a subject which Markov Jr was unable to
describe in 1951!) and denounced the Russian Orthodox Church, see
also Sheynin (1989a, pp. 340 – 341; 2007c). The Press used to call
him Militant academician (Nekrasov 1916a, p. 9) and Andrew the
Furious (Neyman 1978). Other sources include Gnedenko & Sheynin
(1978), and Seneta (2001). See Markov’s correspondence in S, G, 16.
I have mentioned one of his letters in §14.1-7.
In 1901 Tolstoy was excommunicated from the Church. Then,
during his last days, the Most Holy Synod discussed whether he
should be “admitted to the bosom of the Church” and decided against
it (Anonymous 1910; archival sources of the Synod). This goes to
show that in 1912 Tolstoy’s excommunication was likely well
remembered.
Yes, in 1912 Markov submitted a request to the Synod for
excommunication. He quoted his Treatise to the effect that

239
238
 We should regard stories about incredible events allegedly having
occurred in bygone times with extreme doubt

and added that he did not

Sympathise with religions which, like Orthodoxy, are supported by,

and in turn lend their support to fire and sword.

His request was not granted; the Synod resolved that Markov “had
seceded from God’s church” (Emeliakh 1954, pp. 400 – 401 and 408).
This paper provides, in particular, the text of an internal letter of the
Synod which stated that excommunication would have been too
honourable for Markov. I think that it was the notorious Beilis case (a
blood libel) which prompted Markov’s request.
Markov’s last and certainly useless protest took place in 1921
(Grodzensky 1987, p. 137) when 15 professors of the Petrograd
University declared that applicants ought to be chosen according to
their knowledge rather than to class or political considerations;
Markov was the first to sign their statement.
Markov’s attitude towards other scholars had not rarely been
wrong. The excessive sharpness of his statements is generally known
and here is a passage from a letter of Zhukovsky, the then President
of the Moscow Mathematical Society, of 23.11.1912 to Markov (S,
G, 16, section 5.1, Letter 47):

I cannot fail to reproach you for the expressions concerning the

honourable Sergei Alekseevich Chaplygin in your letter. They can
hardly be called proper.

Chaplygin (1869 – 1942) was cofounder of aero hydrodynamics

(and an active member of the Society). Markov’s letter reflected the
polemic between him and Nekrasov in the Society’s periodical,
Matematichesky Sbornik.
The second and last example (K. A. Andreev’s letter of 1915 to
Nekrasov; Sheynin 1994f, p. 132): Markov

Remains to this day an old and hardened sinner in provoking

debate. I had understood this long ago, and I believe that the only way
to save myself from the trouble of swallowing the provocateur’s bait is
a refusal to respond to any of his attacks …

Andreev had published a posthumous manuscript of V. G.

Imshenetsky and Markov severely criticized its incompleteness.
Nevertheless, Markov himself, soon before his death, agreed to
publish his last, and also incomplete manuscript (Besikovitch 1924, p.
XIV).
Markov had been sharply and even provocatively behaving towards
Nekrasov whom he overpowered by many rude postcards (Nekrasov’s
complaint of 1915 to the permanent secretary of the Academy of
Sciences (Nekrasov 1916a, pp. 56 – 62; S, G, 16).

240
239
 In his own scientific work, Markov had been too rigid, see §§ 14.1-
4, 14.1-5 (the statement of A. A. Youshkevich), 14.2-1 and 14.2-4,
which negatively influenced his work. During his last years, in spite of
extremely difficult conditions of life in Russia and his worsened
health, he completed (?) the last edition of his Treatise (published
posthumously) but hardly sufficiently described the new findings of
the Biometric school. Such scholars as Yule and Student (Gosset)
were not mentioned, Fisher was also ignored.
14.4 Liapunov
The theory of probability remained an episode in his scientific
work. He (1900; 1901a) proved the [CLT] assuming a single condition
(7). I briefly repeat (Bernstein 1945/1964, pp. 427ff) that a
characteristic function determines the sought law of distribution
independently from the existence of the relevant moments and that the
expansion in powers of s which Chebyshev (§ 13.1-4) made use of
did not anymore lead to difficulties after replacing that argument by is.
Liapunov proved that under his condition the characteristic function of
a centred and normed sum of random variables tended to the
characteristic function of a normed normal law. I also mention
Lindeberg (1922b, p. 211) whose proof of the CLT was simpler and
became better known5. He referred to his previous paper (1922a) and
continued:

I see now that already Liapunov had explicated general findings

which not only surpass the results achieved by Mises [a reference to
his article of 1919 followed] but which make it possible to derive most
of what I have established. […] The study of Liapunov’s work
prompted me to check anew the method that I have applied.

A special point is connected here with the CLT for large deviations.
Chebyshev thought that the limits of integration, α and β in formula
(13.6) describing that theorem, were “any”. Nekrasov (1911, p. 449)
arbitrarily interpreted that expression as “variable”. I discuss
Nekrasov in § 14.5; here, I say that he could have well indicated that,
on the contrary, he had generalized the Chebyshev theorem. In his
previous polemic paper Liapunov (1901b, p. 61) declared that he had
assumed that these limits were given beforehand and that otherwise
the probability, written down in the left side of formula (13.6), could
have no limit at all, – but nevertheless be asymptotically expressed by
the normal law of distribution6.
14.5. Nekrasov
Nekrasov’s life and work are clearly separated into two stages.
From 1885 and until about 1900 he had time to publish remarkable
memoirs both in Russia and Germany and to become Professor and
Rector of Moscow University; I mentioned him in § 10.2. In 1898 he
sketched the proof of the [CLT] for sums of [lattice random variables].
Then, however, his personality changed. His writings (only on
probability and statistics) became unimaginably verbose, sometimes
obscure and confusing, and inseparably linked with ethical, political
and religious considerations. Here is a comparatively mild example
(1906, p. 9): mathematics accumulated

241
240
 Psychological discipline as well as political and social arithmetic
or the mathematical law of the political and social development of
forces which depend on mental and physiological principles.

Furthermore, Nekrasov’s work began to abound with elementary

mathematical mistakes and senseless statements. Thus (Nekrasov
1901, p. 237): it is possible to assume roughly that x n, n > 0, is the
limit of sin x as |х| → 0, and

The conclusions made by [Chebyshev, Markov and Liapunov]

never differ from such an understanding of limit.

I provide a second and last out of many possible illustrations from

Nekrasov’s letter of 20.12.1913 to Markov (Archive, Russian Acad.
Sci., Fond 173, inventory 1, 55, No. 5; S, G, 16, section 5.1, Letter
17):

I distinguish the viewpoints of Gauss and Laplace [on the MLSq]

by the moment with regard to the experiment. The first one is
posterior and the second one is prior. It is more opportune to judge à
posteriori because more data are available, but this approach is
delaying, it lags behind, drags after the event.

The attendant reasons for such a change were Nekrasov’s religious

upbringing (before entering Moscow University he graduated from a
Russian Orthodox seminary), his work from 1898 onward as a high
official at the Ministry of People’s Education7, and his reactionary
views. At least once Nekrasov (A. V. Andreev 1999, p. 103)
mentioned the Integral Knowledge of the religious philosopher V. S.
Soloviev (1853 – 1900) and it is opportune to quote Soloviev’s
pronouncement (Radlov 1900, p. 787) with which, in actual fact,
Nekrasov became absorbed: “veritable knowledge is a synthesis of
theology, rational philosophy and positive science”. Andreev indeed
maintains that Nekrasov became split between mathematics and such
philosophy. Bortkiewicz (1903) notes that Nekrasov “especially often
mentioned Soloviev in vain”, – and sometimes justifiably, as I am
inclined to believe.
And here is Liapunov’s suitable remark (1901b, p. 63):

All of Nekrasov’s objections are based on misunderstandings […].

Some of them are simply unjustified statements […], others either
have no relations to the papers which he criticizes or are extremely
indefinite.

Concerning Nekrasov’s social and political views I turn to his letter

of 1916 to P. A. Florensky (Sheynin 1993a, p. 196): “the German –
Jewish culture and literature” pushes “us” to the crossroads. World
War I was then going on, but that fact only to some extent exonerates
Nekrasov. Florernsky was ab eminent theologian, a philosopher of
mathematics and an out-and-out anti-Semite. I have seen in the

242
241
internet a story about his statement to the effect that, had he been a
Jew, he would himself killed the Christian boy. He spoke about the
notorious blood libel of Beilis (who was later acquitted).
I shall now dwell on some concrete issues.
1) Teaching the theory of probability. In § 14.1-7 I mentioned
Nekrasov’s proposal for teaching probability in school and the
rejection of the curriculum drawn up by him. I add now that already in
1898 Nekrasov made a similar proposal concerning the Law Faculty
of Moscow University, also rejected or at least forgotten; however
(Sheynin 1995a, p. 166), during 1902 – 1904 the theory of probability
was not taught there even at the Physical and Mathematical Faculty,
and hardly taught during 1912 – 1917.
2) The MLSq. Nekrasov (1912 – 1914) mistakenly attributed to
Legendre an interpolation-like application of the method and (1914)
acknowledged his failure to notice, in 1912, the relevant work of
Yarochenko (1893a; 1893b), but still alleged (wrongly) to have
considered the issue in a more general manner. Yarochenko justified
the arithmetic mean and the MLSq in general by a reference to
Chebyshev’s memoir (1867), – that is, by the Bienaymé – Chebyshev
inequality (§ 9A.4-3). Note that the first such statement appeared
simultaneously with the Chebyshev memoir (Usov 1867). Recall also
Nekrasov’s strange pronouncement about Laplace and Gauss quoted
above.
3) [The CLT]. It was Nekrasov who had considered the CLT for
large deviations, – for the case that began to be studied only 50 years
later. Suppose that independent [lattice] random variables (linear
functions of integral variables) ξi, i = 1, 2, …, n, have finite mean
values ai and variances σi2 and

m = ξ1 + ξ2 + … + ξn.

Denote

|x(m)| = |m – Σai|/(Σσi2)1/2.

Nekrasov restricted his attention to the case in which |x| < np,
0 < p < 1/6 and stated that, for all values of m1 and m2 which
conformed to that condition,

P(m1 < ξ1 + ξ2 + … + ξn < m2) ~ [1/ 2π ]  exp (– t2/2)dt.

The limits of integration were x(m1) and x(m2) respectively.

In all, Nekrasov (1898) formulated six theorems and proved them
later (1900 – 1902). Neither Markov, nor Liapunov had sufficiently
studied them; indeed, it was hardly possible to understand him and A.
D. Soloviev (1997/2008, p. 359) reasonably stated:

I am firmly convinced that no contemporaneous mathematician, or

later historian of mathematics had [has] ever studied it [the memoir
(1900 – 1902)] in any detail.

243
242
 He himself was only able to suggest that Nekrasov had indeed
proved his theorems and he reminded his readers that Markov had
indicated some mistakes made by Nekrasov. Furthermore, Soloviev
(pp. 356 – 357) remarked that Nekrasov had wrongly understood the
notion of lattice variables (not like I described above). In his general
conclusion Soloviev (p. 362) stated that Nekrasov had imposed on the
studied variables an excessively strict condition (the analyticity of the
generating functions in some ring, which was much stronger than
presuming the existence of all of the moments) and that it was
generally impossible to check his other restrictions. Both Soloviev,
and the first of the modern commentators, Seneta (1984, §6), agree in
that Nekrasov’s findings had not influenced the development of the
theory of probability8. This regrettable outcome was certainly caused
both by Nekrasov’s inability to express himself intelligibly and by the
unwieldiness of his purely analytical rather than stochastic approach
(Soloviev, p. 363).
Chuprov (Bortkevich & Chuprov 2005, letter to Bortkiewicz of 22
Nov. 1896) actually proved that Nekrasov did not study mathematical
statistics at all or even sufficiently mastered the theory of probability:

Noting the word dispersia [variance] in my [student] composition,

he timorously asked me: So do you really apply the theory of
probability to the dispersion of light?

Notes
1. Still, Markov had a forerunner: Donkin (1851) expressed a quite similar
statement apparently based on the principle of insufficient reason which Keynes
(1921/1973, p. 44) later renamed the principle of indifference. Boole (1854b/2003,
p. 163) thought that perhaps [Donkin’s principle] might […] be regarded as
axiomatic. Neither Donkin, nor Keynes mentioned that that principle or axiom
transforms the theory of probability.
2. Until he began to study his chains, Markov had always introduced these two
conditions. In one case (1899a/1951, p. 240) he apparently had not repeated them
from his p. 234.
3. A misprint occurred in the Russian translation of the French original.
4. Grodzensky regrettably had not adduced an index of the letters he discovered
and did not indicate which of them had indeed been published at once.
5. Thus, Gnedenko (1954/1973, pp. 254 – 259) proves the theorem under the
Lindeberg condition and then explains that the Liapunov restriction leads to the
former.
6. Liapunov’s correspondence with K. A. Andreev in 1901 (Sheynin 1989b)
testifies that he had initially wished to publish his note in the Matematichesky
Sbornik, that the leadership of the Moscow Mathematical Society (Bugaev,
Nekrasov (!)) opposed his desire, and that he essentially expanded his first draft on
Andreev’s advice.
7. Here is K. A. Andreev’s opinion (letter of 1901 to Liapunov; Gordevsky
1955, p. 40): Nekrasov

Reasons perhaps deeply, but not clearly, and he expresses his thoughts still more
obscurely. I am only surprised that he is so self-confident. In his situation, with the
administrative burden weighing heavily upon him, it is even impossible, as I
imagine, to have enough time for calmly considering deep scientific problems, so
that it would have been better not to study them at all.

244
243
 Here is an example of Nekrasov’s deep thought (1916b, p. 23): he mentioned
almost all the problems of the yet not existing theory of catastrophes and introduced
that very term, catastrophe.
8. It might be added, however, that Markov (1912b, p. 215) sometimes
considered the refutation of Nekrasov’s mistaken statements as one of the aims of
his work. A similar explanation is contained in one of his letters of 1910 to
Chuprov (Ondar 1977/1981, p. 5).

245
244
 15. The Birth of Mathematical Statistics
15.1. The Stability of Statistical Series
By the end of the 19th, and in the beginning of the 20th century,
statistical investigations on the Continent were mostly restricted to the
study of population. In England, on the contrary, the main field of
application for statistical studies at the time had been biology. It is
possible to state more definitely that the so-called Continental
direction of statistics originated as the result of the work of Lexis
whose predecessors had been Poisson, Bienaymé, Cournot and
Quetelet. Poisson and Cournot (§ 8.6) examined the significance of
statistical discrepancies “in general” and assuming a large number of
observations, – without providing concrete examples. Cournot (§
10.3-6) also attempted to reveal dependence between the decisions
reached by judges (or jurors). Bienaymé (§ 10.2-3) was interested in
the change in statistical indicators from one series of trials to the next
one and Quetelet (§ 10.5) investigated the connections between causes
and effects in society, attempted to standardize statistical data
worldwide and created moral statistics.
All this had been occurring against the background of statements
that the theory of probability was only applicable to statistics if, for a
given totality of observations, “equally possible cases” were in
existence, and the appropriate probability remained constant (§ 10.7).
15.1.1. Lexis. He (1879) proposed a distribution-free test for the
equality of probabilities in different series of observations; or, in other
words, a test for the stability of statistical series. Suppose that there
are m series of ni observations, i = 1, 2, …, m, and that the probability
of success was constant throughout and equal to p. If the number of
successes in series i was ai, the variance of these magnitudes can be
calculated by two independent formulas (Lexis 1879, § 6)

σ12 = pqn, σ22 = [vv]/(m – 1) (1; 2)

where n was the mean of ni, vi, the deviations of ai from their mean,
and q = 1 – p. Formula (2) was due to Gauss, see (9.6b); he also knew
formula (1), see a posthumously published note: W-8, 1900, p. 133.
The frequencies of success could also be calculated twice. Note
however that Lexis applied the probable error rather than the variance.
Lexis (§ 11) called the ratio

Q = σ2/σ1 (3)

the coefficient of dispersion perhaps choosing the letter Q in honour of

Quetelet. In accordance with his terminology, the case Q = 1
corresponded to normal dispersion (with some random deviations
from unity nevertheless considered admissible); he called the
dispersion supernormal, and the stability of the totality of observations
subnormal if Q > 1 (and indicated that the probability p was not then
constant); and, finally, Lexis explained the case Q < 1 by dependence
between the observations, called the appropriate variance subnormal,
and the stability, supernormal. He did not, however, pay attention to
this last-mentioned case.

246
245
 But how exactly, in Lexis’ opinion, could the probability vary? No
universal answer was of course possible. He (1876, pp. 220 – 221 and
238) thought that the variations followed a normal law, but then he
(1877, § 23) admitted less restrictive conditions (evenness of the
appropriate [density function]) and noted that more specific
restrictions were impossible. Moreover, I am not sure that Lexis had
firmly broken off with previous traditions, see § 10.7-7.
Then, Lexis (1879) discussed this issue once more, and even
mentioned “irregular waves” (§ 22), but it is very difficult to follow
him. Time and time again he interrupted himself by providing
statistical examples and never gave precise formulations. Recall (my §
10.8.4) that Newcomb introduced a mixture of normal laws instead of
one single distribution and that at the very least his exposition was
quite definite. At the same time Lexis made a common mistake by
believing that the relation between the mean square error and the
probable error remained constant irrespective of the appropriate
distribution.
He had not calculated either the expectation, or the variance of his
coefficient (which was indeed difficult), neither did he say that that
was necessary. Recall (§ 9A.4) that Gauss, after introducing the
sample variance, indicated that it was [unbiased] and determined its
variance. Lexis’ main achievement was perhaps his attempt to check
statistically some stochastic model; it is apparently in this sense that
Chuprov’s remark on the need to unite him and Pearson (§ 15.2)
should be understood.
I note finally that Lexis (1879, § 1) qualitatively separated
statistical series into several types and made a forgotten attempt to
define stationarity and trend.
A French actuary Dormoy (1874; 1878) preceded Lexis, but at the
time even French statisticians (who barely participated in the
development of the Continental direction of statistics) had not noticed
his theory. It was Lexis who first discovered Dormoy (Chuprov
1909/1959, p. 236) and Chuprov (1926, p. 198, Russian translation
1960, p. 228, English translation 2004, p. 78) argued that the Lexian
theory of dispersion ought to be called after Dormoy and Lexis. He
thus opposed a later opinion of Bortkiewicz (1930, p. 53) who ranked
Dormoy far below Lexis. Be that as it may, later statisticians had only
paid attention to Lexis.
15.1.2. Bortkiewicz. I mentioned him in § 8.7 in connection with
the LLN and, in § 10.7-4, I dwelt on his statement about the
estimation of precision of statistical inferences. Of Polish descent,
Vladislav Iosifovich Bortkevich was a lawyer by education. He was
born and studied in Petersburg, but at the end of the 19th century he
continued his education in Germany (he was Lexis’ student) and in
1901 secured a professorship in Berlin and remained there all his life
as Ladislaus von Bortkiewicz. In 1912 the Russian statistician P. D.
Azarevich (Fortunatov 1914, p. 237) mentioned him thus: “Each time
I see him, I feel sorry that he was lost to Russia. There’s a genuine
man of science”. In a letter of 1905 to Chuprov (Sheynin 1990c/2011,
p. 56) Bortkiewicz indicated that in Germany he felt himself
“perfectly well”, whereas a cataclysm was possible in Russia.

247
246
Bortkiewicz had indeed published most of his contributions in
German (which he knew hardly worse than Russian), but he did not
lose his ties with Russia. He (1903) sharply criticized Nekrasov for
the latter’s statements that the theory of probability can soften “the
cruel relations” between capital and labour (p. 215) and (p. 219)
exonerate the principles of firm rule and autocracy as well as for
Nekrasov’s “sickening oily tone” (p. 215) and “reactionary longings”
(p. 216)1. Then, Slutsky (1922) referred to a letter received from
Bortkiewicz, see also Sheynin (2007f), and, finally, at least during his
last years he was connected with the then existing in Berlin Russian
Scientific Institute and Russian Scientific Society (Sheynin 2001f, p.
228; Bortkevich & Chuprov, 2005, pp. 9 – 12).
Bortkiewicz achieved interesting findings and his example is
extremely instructive since he was not initially acquainted with
mathematics. In 1896, in a letter to Chuprov (Sheynin 1990c/2011, p.
58), he declared that the differentiation of an integral with respect to
its (lower) limit was impossible. Many authors deservedly praised him
for his scientific work. Thus (Woytinsky 1961, pp. 452 – 453), he was
called “the statistical Pope” whereas H. Schumacher (1931, p. 573)
explained Bortkiewicz’ attitude towards science by a quotation from
the Bible (Exodus 20:3): “You shall have no other gods before me”.
Here is how Chuprov’s student and the last representative of the
Continental direction, Anderson2 (1932, p. 243/1963, Bd. 2, p. 531)
described his achievement in studying statistical series:

Our (younger) generation of statisticians is hardly able to imagine

that mire in which the statistical theory had got into after the collapse
of the Queteletian system, or the way out of it which only Lexis and
Bortkiewicz have managed to discover.

Yes, but neither the former (§ 15.1.1) nor the latter (see below) had
successfully overcome the occurring difficulties. Later Anderson
included Chuprov.
Bortkiewicz’ work is insufficiently known mostly because of his
pedestrian style and excessive attention to details but also since
German statisticians and economists of the time (Bortkiewicz was also
a celebrated economist) had been avoiding mathematics. He did not
pay attention to improving his style. Chuprov (Bortkevich & Chuprov
2005, Letter 35 of 1898) criticized his friend for a methodically
unfortunate paper on Pareto but Bortkiewicz refused to mend his
ways. And Winkler (1931, p. 1030) quoted a letter from Bortkiewicz
(date not given) who was glad to find in him one of the five expected
readers of his work (which one?)! Here is what Anderson (1932, p.
245/1963, Bd. 2, p. 533) had to say:

Bortkiewicz did not write for a wide circle of readers […] and was
not at all a good exponent of his own ideas. In addition, he made very
high demands on the readers’ schooling and intellect. With
stubbornness partly caused by his reclusive life, […] he refused to
follow the advice of […] Chuprov …

248
247
 His Italian papers of 1908 – 1909 written as a defence of his law of
small numbers (see below) are virtually unknown. The initial German
text of one of them is kept at Uppsala and copied in S, G, 25.
Bortkiewicz had determined EQ and EQ2. Chuprov several times
mentioned this fact (Sheynin 1990c/2011, pp. 87, 93, 139) and in 1916
Markov (Ondar 1977/1981, p. 93) stated that Bortkiewicz’ “research
[…] while not fully accurate, is significant” and (Markov 1911c/1981,
p. 153) that “some” of his relevant studies “deserve greater attention”.
It is most interesting that Bortkiewicz introduced his law of small
numbers (1898a) for studying the stability of statistical series3. He
argued that a series consisting of independent observations with
differing probabilities of the occurrence of a rare event might be
considered as a sample from a single totality. This fact, or, more
precisely, the decrease of the pertinent coefficient of dispersion to
unity with the decrease of the number of observations he had indeed
called the law of small numbers.
From the very beginning his publication aroused debates (Sheynin
1990c/2011, pp. 59 – 62). Chuprov (Bortkevich & Chuprov 2005,
Letter 2 of 1896) advised Bortkiewicz to refer to Poisson, but in 1909
– 1911, in his letters to Chuprov, Bortkiewicz stressed the distinction
between his law and the Poisson formula. The low value of
probability, as he argued, was not his main assumption; the rarity of
the event might have been occasioned by a small number of
observations. Incidentally, this explanation raises doubts about the
applicability here of the Poisson distribution, and, for that matter,
Bortkiewicz had never comprehensively explained his law. Here is
what Chuprov wrote to Markov in 1916 (Sheynin 1990c/2011, pp. 91
– 92):

It is difficult to say to what extent the law of small numbers enjoys

the recognition of statisticians since it is not known what, strictly
speaking, should be called the law of small numbers. Bortkiewicz did
not answer my questions formulated in the note on p. 398 of the
second edition of the Essays [Chuprov 1909/1959, p. 285] either in
publications or in written form; I did not question him orally at all
since he regards criticisms of the law of sm. numb. very painfully.

Mathematicians now simply dismiss the law of small numbers as

another term for the Poisson limit theorem (Kolmogorov 1954,
without explaining the situation). However, the first to deny that law
was L. Whittaker (1914) who showed her standpoint in the very title
of her paper. Bortkiewicz sharply objected to it in his polemic paper
(1915a).
Markov repeatedly discussed that law in his letters of 1916 to
Chuprov (Ondar 1977/1981); he indicated that Bortkiewicz had
wrongly combined his data and (p. 108) “chose material that was
pleasing to him”4 and that (pp. 81 and 108) for small numbers the
coefficient of dispersion cannot be large. He also publicly repeated his
last-mentioned statement (1916b, p. 55). In 1916, in an answer to
Markov, Chuprov (Sheynin 1990c/2011, p. 91) apparently disagreed

249
248
that Bortkiewicz had wrongly combined his materials and reported
that Yastremsky (1913) had also proved Markov’s main statement.
Then, Quine & Seneta (1987) described Bortkiewicz’ law and
indicated more definitely that for small independent and integral
random variables a large value of Q was unlikely.
No one ever noted that the Bortkiewicz coefficient Q1 differed from
Q (so that some comments had been wrong) and that it was the ratio of
two dependent random variables, call tem ξ and η: Q1 = Eξ/Eη, and,
consequently, no one noted that Bortkiewicz wrongly assumed that
Eξ/Eη = E(ξ/η). Chuprov was an exception, but he said nothing in
public whereas Bortkiewicz arbitrarily answered him that that wrong
equality holds approximately.
I ought to add that after 60 years of its neglect Bortkiewicz was the
author who picked up Poisson’s law (also Newcomb, see § 10.8.4,
and Chebyshev) and that for a long time his contribution (1898a) had
remained the talk of the town. Thus, Romanovsky (1924, book 17, p.
15) called Bortkiewicz’ innovation “the main statistical law”.
Bortkiewicz invariably defended his law but it is difficult to study it
thoroughly; commentators whom I mentioned above did not examine
it comprehensively (and neither did other authors), and I refer readers
to my paper (2008). I have minutely studied his contribution and
decided that he did nothing except (as Kolmogorov stated) recalling
Poisson. Here, I only say that his coefficient of dispersion differed
from the Lexian Q which meant that some conclusions of the
commentators mentioned above became meaningless.
Another point here is that the decrease of the coefficient of
dispersion with the decrease of the number of observations tells us
nothing about the underlying probability (probabilities) of the studied
event. Chuprov (1909/1959, p. 277) noted this fact but only mentioned
Lexis, not Bortkiewicz, and did not repeat his remark in any later
Russian or German paper.
15.1.3. Markov and Chuprov. In his letters of 1910 to Chuprov,
Markov (Ondar 1977) proved that Lexis’ considerations were wrong.
Thus, it occurred that the dispersion could also be normal when the
observations were dependent. In addition, he constructed an example
of independent observations which, when being combined into series
in different ways, were characterized either by super- or subnormal
dispersions. However, later Chuprov, in a letter of 1923 to his former
student Chetverikov (Sheynin 1990c/2011, p. 139), remarked that
stability was only determined for concrete series.
Also in 1910, Chuprov, in a letter to Markov, provided examples of
dependences leading to super- and sub-normality of dispersion; in
1914 he even decided that the coefficient of dispersion should be
“shelved” to which Bortkiewicz strongly objected (Sheynin
1990c/2011, p. 140). Then, in 1916 both Markov and Chuprov proved
that EQ2 = 1 (see details Ibidem, pp. 140 – 141). Finally, Chuprov
(1918 – 1919; see Ibidem, pp. 113 – 114) definitively refuted the
applicability of the coefficient of dispersion.
I cannot understand why, even in 1921 Chuprov (2009, p. 88) wrote
to Gulkevich (a noted Russian diplomatist who refused to return to the

250
249
Soviet Union and became the assistant of F. Nansen at the League of
Nations):

One of the most important doctrines of the theory of statistics,

which I had until now completely acknowledged and professed, the
Lexian theory of the stability of statistical numbers, is to a large
extent, as it occurs, based on a mathematical misunderstanding. This
knocks out one of the foundations of the theory whose central part is
now hanging on air. I do not want to resign myself to this fact without
providing a replacement, but I cannot manage. My attempts encounter
that same objection and I have almost concluded that the obstacle is
essentially insurmountable. […] That’s how it is turning out for the
time being.

All this was almost forgotten. Thus, Särndal (1971, pp. 376 – 377),
who briefly described the work of Lexis and noted that it prompted
Charlier “to look […] into questions of non-normality of data”, did
not mention either it or any later developments.
Here is what Anderson (1926/1963, Bd. 1, p. 31) stated, and what
significantly complements his own opinion cited in § 15.1.2: Chuprov

Essentially changed the traditional Lexis’ doctrine […] from which

actually pretty little is now left.

In the same contribution Chuprov (1918 – 1919, p. 205) proved, in

a most elementary way, a general formula for the variance:

 n n

(1/n)E   ( xi  Ei ) 2  =
 i 1 i 1 
n n
(1/n2)  E(ξi – Eξi)2 + (1/n2)   [E(xixj)– Eξi Eξj].
i 1 i 1 j i

Included here were n random variables ξi anyhow dependent on each

other and the results of a single observation xi of each of them. I note
that Chuprov partly issued from his manuscript (1916 or 1917) There,
he determined EQ2 anew and provided qualitative considerations
concerning the distribution of the coefficient of dispersion. Chuprov
sent his manuscript to Markov, and it is mentioned or implied in their
correspondence of 1917 (Sheynin 1990c/2011, pp. 94 – 96) and,
possibly, Ondar 1977/1981, p. 116ff).
Romanovsky (1923) published a very favourable review of
Chuprov’s work but did not indicate that the latter’s notation was too
involved and impeded understanding. A notorious case is presented in
his paper (Chuprov 1923). There, for example on p. 472, he applied
two-storeyed superscripts and two-storeyed subscripts in the same
formula. Hardly has any other author (certainly including
Bortkiewicz) allowed himself to take such liberties, to expect his
readers to understand suchlike gibberish.
A few additional words are here in order. While studying the
stability of statistical series, Chuprov achieved really interesting

251
250
results, see Seneta (1987). On the other hand, since he considered
problems of the most general nature, he inevitably derived awkward
formulas, and the same Romanovsky (1930, p. 216) noted that
Chuprov’s formulas, although “being of considerable theoretical
interest”, were “almost useless” due to complicated calculations
involved. And, on the next page: the estimation of the empirical
coefficient of correlation for samples from arbitrary populations was
possible almost exclusively by means of Chuprov’s formulas, which
however were “extremely unwieldy […], incomplete and hardly
investigated”.
15.2. The Biometric School
That name itself recalls the periodical Biometrika whose first issue
appeared in 1902 with a subtitle Journal for the Statistical Study of
Biological Problems. Its first editors were Weldon (a widely educated
biologist who died in 1906), Pearson and Davenport5 “in
consultation” with Galton. The editorial there contained the following
passage:

The problem of evolution is a problem in statistics […] [Darwin

established] the theory of descent without mathematical conceptions6
[…] [but] every idea of Darwin – variation, natural selection […] –
seems at once to fit itself to mathematical definition and to demand
statistical analysis. […] The biologist, the mathematician and the
statistician have hitherto had widely differentiated fields of work. […]
The day will come […] when we shall find mathematicians who are
competent biologists, and biologists who are competent
mathematicians …

Much later Pearson (1923, p. 23) once more mentioned Darwin:

We looked upon Darwin as our deliverer, the man who had given
new meaning to our life and the world we inhabited.

Here is a passage from a note which Pearson had compiled (and

apparently sent around) in 1920 and which his son, E. S. Pearson
(1936 – 1937, vol. 29, p. 164), quoted: The aim of the Biometric
school was

To make statistics a branch of applied mathematics, […] to extend,

discard or justify the meagre processes of the older school of political
and social statisticians, and, in general, to convert statistics in this
country from being the playing field of dilettanti and controversialists
into a serious branch of Science. […] Inadequate and even erroneous
processes in medicine, in anthropology [anthropometry], in
craniometry, in psychology, in criminology, in biology, in sociology,
had to be criticized […] with the aim of providing those sciences with
a new and stronger technique.

Note that almost all the disciplines mentioned above were included
in Pearson’s main field of interests and that he had not found a single
kind word for Continental statisticians.

252
251
 Here is a peculiar passage (Pearson 1907, p. 613):

I have learned from experience with biologists, craniologists,

meteorologists, and medical men (who now occasionally visit the
biometricians by night!) that the first introduction of modern
statistical methods into an old science by the layman is met with
characteristic scorn; but I have lived to see many of them tacitly
adopting the very processes they began by condemning.

The immediate cause for establishing Biometrika seems to have

been scientific friction and personal disagreement between Pearson
and Weldon on the one hand, and biologists, especially Bateson, on
the other hand, who exactly at that time had discovered the unnoticed
Mendel. It was very difficult to correlate Mendelism and biometry: the
former studied discrete magnitudes while the latter investigated
continuous quantitative variations.

Between the Biometric and Mendelian schools there was nothing

fundamentally incompatible; in pursuit of the same objective they
followed different lines of approach which were essentially
complementary rather than antagonistic. E. S. Pearson (1936, p. 227).

On pp. 169 – 170 E. S. P. noted that the study of social problems by

the Mendelian approach had been superficial.
The rapid success of the new school was certainly caused by the
hard work of its creators, but also by the efforts of their predecessor,
Edgeworth. Chuprov (1909/1959, p. 27 – 28) provided his correct
characteristic. A talented statistician (and economist), he was
excessively original and had an odd style; he was therefore unable to
influence strongly his contemporaries. However, in his native country
he at least paved the way for the perception of mathematical-statistical
ideas and methods. His works have appeared recently in three
volumes (1996). See also Schumpeter (1954/1955, p. 831) and M. G.
Kendall (1968/1970, pp. 262 – 263). I have quoted them as well as
Chuprov elsewhere (Sheynin 2006/2009, NNo. 537 – 539).
Pearson, perhaps at once, became the main editor of Biometrika,
and among his authors were Chuprov and Romanovsky7.
A list of more than 600 of Pearson’s publications is in Morant et al
(1939) and Merrington et al (1983), and his son, Egon Pearson (1936
– 1937) described his life and work. Many of his earlier papers are
reprinted (K. Pearson 1948) and his manuscripts are held in University
College London.
Hardly known are his thoughts about physics which he studied until
1893. Thus (1891, p. 313) negative matter exists in the universe, and
physical variation effects were perhaps due to the general
construction of our space, see Clifford (1885/1886, p. 202). He did
not however mention Riemannian spaces whereas it is nowadays
thought that the curvature of space–time is caused by forces operating
in it. Lastly (Pearson 1887, p. 114), all atoms in the universe of
whatever kind appear to have begun pulsating at the same instant.

253
252
 Unexpectedly, at least for me, Tee (2003, 15 May) noted that
Pearson’s Grammar of Science (1892) denied the existence of atoms
and only repented in 1911, in the third edition of that book. Tee (from
University of Auckland) is an extremely careful reviewer; I had not
checked his statement which nevertheless seems at least partly
doubtful (all atoms […] have begun etc., see a bit above).
Pearson (1857 – 1936) was an applied mathematician and a
philosopher, but in the first place a co-founder of biometry. The
beginning of his scientific work can be connected with his Grammar
of science which earned him the brand of a

Conscientious and honest enemy of materialism and one of the most

consistent and lucid Machians.

That was Lenin’s conclusion (1909/1961, pp. 190 and 274). Note
that the latter term is tantamount to Mach’s philosophy, i.e., to a
variety of subjective idealism. It is, however, difficult to imagine that
Pearson evaded reality. But at the same time Mach’s followers define
the aim of science as description rather than study of phenomena and
Pearson separated experience (statistical data) from theory (from the
appropriate stochastic patterns), although he did not at all keep to the
tradition of the Staatswissenschaft (§ 6.2.1).
Pearson’s Grammar became widely known8 and he was elected to
the Royal Society. Newcomb, as President of the forthcoming
International Congress of Arts and Sciences (St. Louis, 1904), invited
him to read a report on methodology of science9. Such scholars as
Boltzmann and Kapteyn had participated there. Newcomb’s attitude
towards Pearson was also reflected in one of his pronouncements of
1903, see § 10.8-4.
S. L. Zabell somewhere noticed that Mach (1897, Introduction) had
mentioned Pearson’s Grammar…:

The publication [of the Grammar] acquainted me with a researcher

whose erkenntnisskritischen [Kantian] ideas on every important issue
coincide with my own notions and who knows how to oppose, candidly
and courageously, extra-scientific tendencies in science.

I mention two more facts concerning Pearson. In 1921 – 1933 he

had read a special course of lectures at University College and in 1978
his son, E. S. Pearson, published them making use of the extant notes
and likely providing the title itself. After Todhunter (1865), this
contribution was apparently the first considerable work in its field and
on its first page the author expressed his regret that he did not study
the history of statistics earlier (see one of the Epigraphs to this book).
E. S. Pearson supplied a Preface where he illustrated his father’s
interest in general history. But in my context it is more important to
mention K. Pearson’s fundamental biography of Galton (1914 –
1930), perhaps the most immense book from all works of such kind,
wherever and whenever published.
Pearson also devoted several papers to the history of probability and
statistics; I mentioned three of them (§§ 2.2.3, 3.2.3 and 7.1-5) and

254
253
resolutely disagreed with the main conclusion of the second one. In
two more articles Pearson (1920; 1928b) studied the history of
correlation and maintained that some authors including Gauss could
have applied the ideas and methods of correlation theory, but that it
would be nevertheless wrong to attribute to them its beginnings. He
often successfully attempted to introduce the statistical method and
especially the theory of correlation into many branches of science. See
however Bernstein’s statement about that theory in § 14.1-4.
The work of Pearson and his followers [Student (real name,
Gosset), Yule and others] is partly beyond the boundaries of my
investigation and I shall only sketch the main directions of Pearson’s
subsequent (after about 1894) studies, of the person who (Hald 1998,
p. 651)

Between 1892 and 1911 […] created his own kingdom of

mathematical statistics [!] and biometry in which he reigned
supremely, defending its ever expanding frontiers against attacks.

And, finally (Fisher 1937, p. 306), soon after Pearson’s death:

K. Pearson’s plea of comparability [between the methods of

moments and maximum likelihood] is […] only an excuse for
falsifying the comparison.

Egon Pearson kept silent.

Positive opinions had been certainly held as well (Mahalanobis
1936; Eisenhart 1974), see also Newcomb’s letter to Pearson in
§ 10.8-4 and the statements of Bernstein and Kolmogorov below.
Pearson’s main merits include the development of the principles of
the correlation theory and contingency, the introduction of the
“Pearsonian curves” for describing empirical distributions, rather than
for replacing the normal law by another universal density, which was
what Newcomb (§ 10.8.4) had attempted to accomplish, and the χ2 test
as well as the compilation of numerous statistical tables. Pearson
(1896 with additions in 1901 and 1916) constructed the system of his
curves in accordance with practical considerations but had not
sufficiently justified it by appropriate stochastic patterns. That system
was defined as the solution of the differential equation

xk
y′ = y (4)
a  bx  cx 2

with four parameters. The case b = c = 0 naturally led to the normal

distribution; otherwise, 12 types of curves appeared of which at least
some were practically useful. Pearson determined the parameters by
the method of moments, – by four sample moments of the appropriate
distribution10. Recall (§ 10.2-4) that the same term is used in the
theory of probability in a quite another way. The “statistical” method
of moments is opportune, but the estimates thus calculated often have
an asymptotic efficiency much less than unity (Cramér 1946, § 33.1).

255
254
 Bernstein (1946, pp. 448 – 457) indicated a stochastic pattern
(sampling with balls being added) that led to the most important
Pearsonian curves. He referred to Markov and, on p. 337, to Polya
(1931) as his predecessors. Markov (1917) had indeed considered the
abovementioned pattern and mentioned the Pearsonian curves on his
very first page; Bernstein, however, mistakenly indicated another of
his papers.
Pearson (E. S. Pearson 1936 – 1937, vol. 29, p. 208 with reference
to his record of the lectures of his father) paid special attention to the
notion of correlation and stated that

The purpose of the mathematical theory of statistics is to deal with

the relationship between 2 or more variable quantities, without
assuming that one is a single-valued mathematical function of the rest.

Abbe (§ 9B-1) and then Helmert derived the χ2 distribution for

revealing systematic influences in the theory of errors (§ 9B-1)
whereas Pearson (1900) introduced the chi-squared test in the context
of mathematical statistics. True, not at once, he began applying it for
checking the goodness of fit, independence in contingency tables and
homogeneity. In spite of its importance, the chi-squared test hardly
“clears” empiricism of its dangers as Fisher, in 1922, claimed (Hald
1998, p. 714).
I (§ 14.1-4) noted that the Continental statisticians were not
recognizing Pearson, also see Ondar (1977/1981, p. 142) who quoted
Chuprov’s similar statement. Many of his colleagues, Chuprov wrote,
“like Markov, shelve the English investigations without reading
them”. The cause of that attitude was the empiricism of the Biometric
school (Chuprov 1918 – 1919, Bd. 2, pp. 132 – 133):

The disinclination of English researchers for the concepts of

mathematical probability and mathematical expectation caused much
trouble. The refusal to use these basic notions obscured the stochastic
statement of problems; on occasion, it even directed the attempts to
solve them on a wrong track. If, however, this attire, so uninviting to
the Continental eye, is shed and the discarded is picked up, then it will
be distinctly seen that Pearson and Lexis often offer different in form
but basically kindred methods for solving essentially similar
problems.
It seems that nowadays one of the most important problems of the
flourishing stochastic theory of statistics is exactly to work out a
synthesis eliminating the contradictions between the two currents
which are directed to common goals. Not Lexis against Pearson, but
Pearson cleansed by Lexis and Lexis enriched by Pearson should be
the slogan of those who are not satisfied by the soulless empiricism of
the post-Queteletian statistics and strive for constructing its rational
theory.

Fisher (1922, pp. 311 and 329n) also indicated that Pearson had
been confusing theoretical and empirical indicators. Similar
pronouncements were due to Anderson (Sheynin 1990c/2011, p.

256
255
149), but I shall only quote Bernstein and Kolmogorov. Bernstein
(1928/1964, p. 228), when discussing a new cycle of problems in the
theory of probability which comprises the theories of distribution and
of the general non-normal correlation, wrote:

From the practical viewpoint the Pearsonian English school is

occupying the most considerable place in this field. Pearson fulfilled
an enormous work in managing statistics; he also has great
theoretical merits, especially since he introduced a large number of
new concepts and opened up practically important paths of scientific
research. The justification and criticism of his ideas is one of the
central problems of current mathematical statistics. Charlier and
Chuprov, for example, achieved considerable success here whereas
many other statisticians are continuing Pearson’s practical work,
definitely losing touch with probability theory …

The modern period in the development of mathematical statistics

began with the fundamental works of […] K. Pearson, Student, Fisher.
… Only in the contributions of the British school did the application of
probability theory to statistics cease to be a collection of separate
isolated problems and became a general theory of statistical testing of
stochastic hypotheses … Kolmogorov (1947, p. 63).

The investigations made by Fisher, the founder of the modern

British mathematical statistics, were not irreproachable from the
standpoint of logic. The ensuing vagueness in his concepts was so
considerable, that their just criticism led many scientists (in the Soviet
Union, Bernstein) to deny entirely the very direction of his research.
Ibidem, p. 64.

Enumerating the “main weaknesses” of the Pearsonian school,

Kolmogorov (1948/2002, pp. 68 – 69) indicated that

Rigorous results concerning the proximity of empirical sample

characteristics to theoretical related only to the case of independent
trials. Notions held by the English statistical school about the logical
structure of the theory of probability which underlies all the methods
of mathematical statistics remained on the level of the eighteenth
century. In spite of the immense […] work done […] the auxiliary
tables used in statistical studies proved highly imperfect in respect of
cases intermediate between “small” and “large” samples.

On the statements of other scientists about Pearson see Sheynin

(2010a). On Yule see M. G. Kendall (1951) and on Gosset (Student)
see E. S. Pearson (1990). His collected papers appeared in 1992. And
just a few words about him from Irwin (1978, p. 409):

He was one of the pioneers in the development of modern statistical

method and its application to the design and analysis of experiments.

257
256
 Irwin (p. 410) also quoted Fisher: Gosset was the Faraday of
statistics. Note that the German statistician Lüroth was the first to
introduce the t-distribution (Pfanzagl & Sheynin 1996).
15.3. The Merging
of the Continental Direction and the Biometric School?
So, did the two statistical streams merge, as Chuprov would have
it? In 1923 he had become Honorary Fellow of the Royal Statistical
Society and in 1926, after his death, the Society passed a resolution of
condolence (Sheynin 1990c/2011, p. 156) which stated that his

Contributions to science were admired by all […]. They did much

to harmonise the methods of statistical research developed by
continental and British workers.

In § 14.1-4 I mentioned the unilateral and, for that matter, only

partly successful attempts made by Chuprov, and the vain efforts of
Slutsky to reconcile Markov with Pearson’s works. And Bauer
(1955, p. 26) reported that he had investigated, on Anderson’s
initiative, how both schools had been applying analysis of variance
and concluded (p. 40) that their work was going on side by side but
did not tend to unification. More details about Bauer`s study are
contained in Heyde & Seneta (1977, pp. 57 – 58) where it also
correctly indicated that, unlike the Biometric school, the Continental
direction had concentrated on nonparametric statistics.
I myself (Gnedenko & Sheynin 1978/2001, p. 275) suggested that
mathematical statistics properly originated as the coming together of
the two streams; even now I think that that statement was not original
(but am unable to mention anyone). However, now I correct myself.
At least until the 1920s, say, British statisticians had continued to
work all by themselves. E. S. Pearson (1936 – 1937), in his study of
the work of his father, had not commented on Continental statisticians
and the same is true about other such essays (Mahalanobis 1936;
Eisenhart 1974). We only know that K. Pearson regretted his
previous neglect of the history of statistics (see one of my Epigraphs).
I believe that English, and then American statisticians for the most
part only accidentally discovered the findings already made by the
Continental school. Furthermore, the same seems to happen nowadays
as well. Even Hald (1998) called his book History of Mathematical
Statistics, but barely studied the work of that school. In 2001,
Biometrika (vol. 88) published five essays devoted to its centenary but
not a word was said in any of them about the Continent, not once was
Chuprov mentioned. It is opportune to add that Cramér (1946,
Preface) aimed to unite, in his monograph, English and American
statistical investigations (and, in the first place, the work of Fisher)
with the new, purely mathematical theory of probability created
“largely owing to the work of” French and Russian mathematicians.
Later he (1981, p. 315) stated that

In the years before the war [before WWI] it had seemed to me that
continental mathematicians and the Anglo-Saxon statisticians were

258
257
working without sufficient mutual contact, and that it might be useful
to try to join both these lines of research.

In 1919 there appeared in Biometrika an editorial remarkably

entitled Peccavimus! (we were guilty). Its author, undoubtedly
Pearson, corrected his mathematical and methodological mistakes
made during several years and revealed mostly by Chuprov (Sheynin
1990c/2011, p. 75) but he had not taken the occasion to come closer to
the Continental statisticians.

Notes
1. Bortkiewicz’ paper appeared in a rare Russian political periodical published
abroad. I discovered that journal in the Rare books section of the (former) Lenin
State Library in Moscow. A few other copies of the periodical’s same issue, which I
since found in Germany, do not, however, contain the paper in question; perhaps it
was only included in a part of the edition.
2. Oskar Nikolaevich Anderson (1887 – 1960), a Russian German, emigrated in
1920. In 1924 – 1942 he lived and worked in Bulgaria, then in Germany (in West
Germany), was the leading statistician in both these countries and a founder-member
of the international Econometric Society (Anderson 1946). Also see Sheynin
(1990c/2011, pp. 80 – 82), H. & R. Strecker (2001) and his collected works
(Anderson 1963). Anderson was the last representative of the Continental direction.
3. In 1897 Bortkiewicz also unsuccessfully attempted to publish his work in
Russian, in a periodical of the Petersburg Academy of Sciences. His request was
refused since the contribution was to appear elsewhere (although only in German),
see Sheynin (1990c/2011, p. 61).
4. This charge was not proved; furthermore, it contradicts our perception of his
personality.
5. An author of a paper published in 1896, of a book devoted to biometry which
appeared in 1899, and of two subsequent notes (M. G. Kendall & Doig 1968).
6. Already in Darwin’s times, a theory was supposed to be quantitatively
corroborated. Darwin, however, provided nothing of the sort and it would be more
proper to say, as in § 10.8.2, “hypothesis of the origin of species”.
7. In 1912, Slutsky had submitted two manuscripts to Pearson who rejected both.
Three letters from Slutsky to Pearson (but no replies) are extant (Univ. College
London, Pearson Papers 856/4 and 856/7; Sheynin 1999c). In this connection
Slutsky had corresponded with Chuprov (Sheynin 1990c/2011, pp. 65 – 67) and
soon published elsewhere (1914) one of his manuscripts, – the one, whose refusal by
Pearson he called a misunderstanding. It was Chuprov who recommended him a
proper outlet.
8. Here is Neyman’s remarkable recollection (E. S. Pearson 1936 – 1937, vol.
28, p. 213): in 1916, he read the Grammar of Science on advice of his teacher at
Kharkov University, S. N. Bernstein, and the book greatly impressed “us”. I note
that, in turn, see below, Pearson (1978, p. 243) had mentioned Lenin: Petersburg

Has now for some inscrutable reason been given the name of the man who has
practically ruined it.

It is not difficult to imagine that Pearson was given a hostile reception in the
Soviet Union. This issue is beyond my chronological boundaries and I only mention
two episodes (Sheynin 1998c, pp. 536 and 538, note 16).
a) Maria Smit, the future Corresponding Member of the Academy of Sciences (!),
1930: the Pearsonian curves are based

On a fetishism of numbers, their classification is only mathematical.

Although he does not want to subdue the real world as ferociously as it was
attempted by […] Gaus [Gauss], his system nevertheless rests only on a
mathematical foundation and the real world cannot be studied on this basis at all.

259
258
 b) A. Ya. Boiarsky, L. Zyrlin, 1947: they blasphemously charged Pearson with
advocating racist ideas that “forestalled the Göbbels department”.
Only somewhat more reserved was the anonymous author in the Great Sov. Enc.,
2nd ed., vol. 33, 1955, p. 85.
Soviet statistics and statisticians endured real suffering. The same Smit, in 1931
(Sheynin 1998c, p. 533, literal translation): “the crowds of arrested saboteurs are full
of statisticians”. She herself had probably helped to assure the success of that
process.
And let us recall Schlözer (1804, § 15): Statistics and despotism do not get along
together.
9. Pearson refused to come because of his financial problems and unwillingness
to leave his Department at University College London under “less complete
supervision” (Sheynin 2002b, pp. 143 and 163, Note 8).
10. Here is a passage from the extant part of an unsigned and undated letter
certainly written by Slutsky to Markov, likely in 1912 (Sheynin 1999c, p. 132):

…are not independent in magnitude from the sum of the already accumulated
deviations or that the probabilities of equal deviations are not constant, we shall
indeed arrive at the formula [Slutsky wrote down formula (4) with k = 0 and F(x)
instead of the trinomial in the denominator]. […] Much material [already shows that
the Pearsonian curves are useful but] […] it seems desirable also for the
asymmetric Pearson curves […] to provide a theoretical derivation which would put
[them] in the same line as the Gauss curve on the basis of the theory of probability
(hypergeometric series).

260
259
 Supplement: Axiomatization
I present a bibliographic survey of some important points.
The main essays are Barone & Novikoff (1978) and Hochkirchen
(1999) and among the lesser known authors is Bernstein (1917). After
Hilbert (1901), Kolmogorov (1933) made the decisive step and
Freudenthal & Steiner (1966, p. 190) commented: he came with the
Columbus’ egg. As the legend goes, Columbus cracked an egg which
enabled it to stand firmly on his table. Among the new sources I list
Hausdorff (2006) who left an important unpublished contribution, see
Girlich (1996), Shafer & Vovk (2001) and Krengel (2011) who
stressed the role of Bohlmann. Vovk & Shafer (2003, p. 27)
characterized their book:

We show how the classical core of probability theory can be based

directly on game-theoretic martingales, with no appeal to measure
theory. Probability again becomes [a] secondary concept but is now
defined in terms of martingales.

In concluding, I quote Boole (1854/1952, p. 288):

The claim to rank among the pure sciences must rest upon the
degree in which it [the theory of probability] satisfies the following
conditions: 1° That the principles upon which its methods are founded
should be of an axiomatic nature.

Boole formulated two more conditions of a general scientific essence.

Bibliography
Barone J., Novikoff A. (1978), History of the axiomatic formulation of
probability from Borel to Kolmogorov. Arch. Hist. Ex. Sci., vol. 18, pp. 123 – 190.
Bernstein S. N. (1917, in Russian), An essay on the axiomatic justification of the
theory of probability. Sobranie Sochineniy (Works), vol. 4. N. p., 1964, pp. 10 – 60.
Boole G. (1854), On the conditions by which the solution of questions in the
theory of probability are limited. In author’s Studies in Logic and Probability, vol. 1.
London, 1952, pp. 280 – 288.
Freudenthal H., Steiner H.-G. (1966), Die Anfänge der
Wahrscheinlichkeitsrechnung. In Grundzüge der Mathematik, Bd. 4. Hrsg H.
Behnke et al. Göttingen, pp. 149 – 195.
Girlich H.-J. (1996), Hausdorffs Beiträge zur Wahrscheinlichkeitstheorie. In F.
Hausdorff zum Gedächtnis, Bd. 1. Hrsg E. Brieskorn. Braunschweig – Wiesbaden,
pp. 31 – 70.
Hausdorff F. (2006), Ges. Werke, Bd. 5. Berlin.
Hilbert D. (1901), Mathematische Probleme. Ges. Abh., Bd. 3. Berlin, 1970, pp.
290 – 329.
Hochkirchen T. (1999), Die Axiomatisierung der Wahrscheinlichkeitsrechnung.
Göttingen.
Kolmogorov A. N. (1933, in German), Foundations of the Theory of Probability.
New York, 1950, 1956.
Krengel U. (June 2011), On the contributions of G. Bohlmann to probability
theory. J. Electron. Hist. Prob. Stat., www.jehps.net
Shafer G., Vovk V. (2001), Probability and Finance. New York.
Vovk V. G., Shafer G. R. (2003), Kolmogorov’s contributions to the foundations
of probability. Problems of Information Transmission, vol. 39, pp. 21 – 31.

261
260
 References
Abbreviation
AHES = Arch. Hist. Ex. Sci.
Hist. Scient. = Historia Scientiarum (Tokyo)
IMI = Istoriko-Matematicheskie Issledovania
ISI = Intern. Stat. Inst.
ISR = Intern. Stat. Rev.
JNÖS = Jahrbücher f. Nationalökonomie u. Stat.
MS = Matematich. Sbornik
OC = Oeuvr. Compl.
PT = Phil. Trans.
UMN = Uspekhi Matematich. Nauk

Bernoulli, J. (1975a; 1986)

David, H. A. & Edwards (2001)
Farebrother (1999)
Freudenthal & Steiner (1966)
Gauss (1887)
Gnedenko & Sheynin (1978)
Hald (1990; 1998; 2003; 2004)
Heyde & Seneta (2001)
Johnson & Kotz (1997)
Kendall M. G. & Doig (1962 – 1968)
Kendall M. G. & Plackett (1977)
Kotz & Johnson (1982 – 1989)
Kotz et al (1997 – 1999)
Kruskal & Tanur (1978)
Pearson E. S. & Kendall (1970)
Pearson K. (1978)
Sheynin et al (1970b)
Schneider (1988)
Todhunter (1865)
2. My Publications Devoted to Various Scientists
and Barely Mentioned in the Main Text
1970b and 1972a (Daniel Bernoulli)
1971b (Lambert)
1973b (Boscovich)
1976, 1977a (Laplace)
1986a (Quetelet)
1990/2011c (Chuprov)
1991b (Buniakovsky)
1994b (Bertrand)
1995 (Helmert)
1998c (statistics in the USSR)
1998c (Bible, Talmud)
2003b (Nekrasov)
2004c (Fechner)
2006/2009 (collected quotes)
2008a, 2012c (Bortkiewicz)
2010b (Bayes)
2010a (K. Pearson)
2012a (Gauss)
2016a (essence of statistics)

262
261
 3. General Literature

Aaboe A., De Solla Price D. J. (1964), Qualitative measurements in antiquity. In

L’aventure de la science (Mélanges A. Koyré, t. 1). Paris, pp. 1 – 20.
Abbe C. (1871), Historical note on the method of least squares. Amer. J. Sci. Arts,
ser. 3, vol. 1, pp. 411 – 415.
Abbe E. (1863), Über die Gesetzmäßigkeit in der Verteilung der Fehler bei
Beobachtungsreihen. Ges. Abh., Bd. 2, 1989, pp. 55 – 81.
Achenwall G. (1752), Staatsverfassung der europäischen Reiche im Grundrisse.
Göttingen, this being the second edition of Abriß der neuesten Staatswissenschaft,
etc. Göttingen, 1749. Many later editions up to 1798 but in 1768 the title changed
once more.
--- (1763), Staatsklugheit und ihren Grundsätzen. Göttingen. Fourth edition, 1779.
Adrain R. (1809), Research concerning the probabilities of the errors which happen
in making observations. All three of his papers (see below) are reprinted in Stigler
(1980, vol. 1).
--- (1818a), Investigation of the figure of the Earth and of the gravity in different
latitudes. Trans. Amer. Phil. Soc., new ser., vol. 1.
--- (1818b), Research concerning the mean diameter of the Earth. Ibidem.
Aira G. B. (1842), Address. Mem. Roy. Astron. Soc., vol. 19, pp. 271 – 283.
Alberuni (Al-Biruni) (1887), India, vols. 1 – 2. Delhi, 1964. The second quotation
from this source in my text is omitted from the abridged edition of 1983.
--- (1934), The Book of Instruction in the Elements of the Art of Astrology. London.
--- (1967), Determination of Coordinates of Cities. Beirut.
Al-Khazini (1983), Kniga Vesov Mudrosti ( The Book of the Balance of Wisdom).
Nauchnoe Nasledstvo, vol. 6, pp. 15 – 140.
Ambartzumian R. V. (1999, Russian), Stochastic geometry. In Prokhorov (1999b,
p. 682).
Anchersen J. P. (1741), Descriptio statuum cultiorum in tabulis. Copenhagen –
Leipzig.
Anderson O. (1926 Bulg.), Zum Gedächtnis an Professor A. A. Tschuprow
(Junior). In author’s Works (1963, Bd. 1, pp. 28 – 38).
--- (1932), Ladislaus von Bortkiewicz. Z. f. Nationalökonomie, Bd. 3, pp. 242 – 250.
Ibidem, Bd. 2, pp. 530 – 538.
--- (1946), [Autobiography]. Archiv Ludwig-Maximilians-Univ. München, Signatur
II-734.
--- (1963), Ausgewählte Schriften, Bde 1 – 2. Tübingen.
André D. (1887), Solution directe du problème résolu par Bertrand. C. r. Acad. Sci.
Paris, t. 105, pp. 436 – 437.
Andreev A. V. (1999, Russian), Theoretical basis of confidence (a sketch of
Nekrasov’s portrait). IMI, vol. 4 (39), pp. 98 – 113.
Anonymous (1735), Géométrie. Hist. Acad. Roy. Sci. avec les Mém. Math. et Phys.,
pp. 43 – 45 of the Histoire. Certainly written by Buffon.
Anonymous (1839), Introduction. J. Stat. Soc. London, vol. 1, pp. 1 – 5.
Anonymous (1910, Russian), The Holy Synod and Tolstoy. Newspaper Rech, 8
Nov. 1910, p. 3.
Arago F. (1850), Poisson. OC, t. 2, 1854, pp. 593 – 671.
Arbuthnot J. (1712), An argument for divine Providence taken from the constant
regularity observed in the birth of both sexes. In M. G. Kendall & Plackett (1977,
pp. 30 – 34).
Argelander F. (1837), Über eigene Bewegung des Sonnensystems. In the Memoirs
of the Petersburg Academy and in Astron. Nachr., Bd. 16, No. 363 – 364, pp. 43 –
55; Bd. 17, No. 398, pp. 209 – 215.
Aristarchus (1959), On the sizes and distances of the sun and the moon. In Heath
Sir Thomas, Aristarchus of Samos. Oxford, pp. 353 – 414.
Aristotle (1908 – 1930, 1954), Works, vols. 1 – 12. London. I am referring to many
treatises from that source although the authorship of two of them (Problemata and
Magna Moralia) is doubtful. There is also a new edition of Aristotle (Princeton,
1984, in two volumes).
Arnauld A., Nicole P. (1662), L’art de penser. Published anonymously. Paris,
1992. English translation: Edinburgh – London, 1850.

264
262
Babbage C. (1857), On tables of the constants of nature and art. Annual Rept.
Smithsonian Instn for 1856, pp. 289 – 302. Abstract published in 1834.
--- (1874), Of observations. Ibidem for 1873, pp. 187 – 197.
Bachelier L. (1900), Théorie de la spéculation. Paris, 1995. Engl. transl.: Princeton,
2006.
Baer K. (1860 – 1875), Issledovania o Sostoianii Rybolovstva v Rossii
(Investigations on the State of Fishing in Russia), vols. 1 –9. Petersburg.
--- (1873), Zum Streit über den Darwinismus. Dorpat [Tartu].
Baily Fr. (1835), An Account of the Revd John Flamsteed. London.
Barnard G. A. (1967), The Bayesian controversy in statistical inference. J. Inst.
Actuaries, vol. 93, pt. 1, pp. 229 – 269.
Barone J., Novikoff A. (1978), A history of the axiomatic formulation of
probability from Borel to Kolmogorov. AHES, vol. 18, pp. 123 – 190.
Bauer R. K. (1955), Die Lexische Dispersionstheorie in ihren Beziehungen zur
modernen statistischen Methodenlehre etc. Mitteilungsbl. f. math. Statistik u. ihre
Anwendungsgebiete, Bd. 7, pp. 25 – 45.
Bayes T. (1764 – 1765), An essay towards solving a problem in the doctrine of
chances. PS Roy. Soc., vols. 53 – 54 for 1763 – 1764, pp. 360 – 418 and 296 – 325.
German transl.: Leipzig, 1908. Reprint of pt. 1: Biometrika, vol. 45, 1958, pp. 293 –
315, and E. S. Pearson & Kendall (1970, pp. 131 – 153).
Beer A. et al (1961), An 8th century meridian line. Vistas Astron., vol. 4, pp. 3 – 28.
Bellhouse D. R. (1981), The Genoese lottery. Stat. Sci., vol. 6, pp. 141 – 148.
--- (1988), Probability in the 16th and 17th centuries. ISR, vol. 56, pp. 63 – 74.
--- (1989), A manuscript on chance written by J. Arbuthnot. ISR, vol. 57, pp. 249 –
259.
Bellhouse D. R., Fillion N. (2015), Le her and other problems in probability etc.
Stat. Sci., vol. 30, pp. 26 – 39.
Belvalkar S. K., Ranade R. D. (1927), History of Indian Philosophy, vol. 2. Poona.
Benjamin M. (1910), Newcomb. In Leading American Men of Science. Ed., D. S.
Jordan. New York, pp. 363 – 389.
Berggren J. L. (1991), Ptolemy’s map of earth and the heavens: a new
interpretation. AHES, vol. 43, pp. 133 – 144.
Bernoulli Daniel (1735), Recherches physiques et astronomiques … Quelle est la
cause physique de l’inclinaison des plans des planètes… In Bernoulli, D. (1987, pp.
303 – 326). [All the other memoirs are reprinted in Bernoulli, D. (1982). When
possible, I provide their titles in translation from the Latin.]
--- (1738; 1982, pp. 223 – 234, in Latin), Exposition of a new theory on the
measurement of risk. Econometrica, vol. 22, 1954, pp. 23 – 36.
--- (1766; 1982, pp. 235 – 267), Essai d’une nouvelle analyse de la mortalité causée
par la petite vérole, et des avantages de l’inoculation pour la prévenir.
--- (1768a; 1982, pp. 276 – 287), De usu algorithmi infinitesimalis in arte
coniectandi specimen.
--- (1768b; 1982, pp. 290 – 303), De duratione media matrimoniorum, etc. [On the
mean duration of marriages for any ages of husband and wife and on other bordering
issues]. Russian transl. in Ptukha (1955, pp. 453 – 464). S, G, 5.
--- (1768c), Letter to Euler. Priroda, No. 5, 1982, pp. 103 – 104. Only in Russian
translation (by A. P. Youshkevich).
--- (1769, Latin), Manuscript, same title as in 1778. In English, in Festschrift for
Lucien Le Cam. New York, 1997, pp. 358 – 367.
--- (1770; 1982, pp. 306 – 324), Disquisitiones analyticae de nouo problemata
coniecturale.
--- (1770 – 1771; 1982, pp. 326 – 338, 341 – 360), Mensura sortis ad fortuitam
successionem rerum naturaliter contingentium applicata.
--- (1778; 1982, pp. 361 – 375, in Latin), The most probable choice between several
discrepant observations and the formation therefrom of the most likely induction.
Biometrika, vol. 48, 1961, pp. 3 – 13, with translation of Euler (1768). Reprint: E. S.
Pearson & Kendall (1970, pp. 155 – 172).
--- (1780; 1982, pp. 376 – 390), Specimen philosophicum de compensationibus
horologicis, et veriori mensura temporis.
--- (1982; 1987), Werke, Bde. 2 – 3. Basel.
Bernoulli Jakob (manuscript; 1975, partial publ.), Meditationes. In Bernoulli, J.
(1975, pp. 21 – 90).

265
263
--- (1713a), Ars conjectandi. Reprint: Bernoulli J. (1975, pp. 107 – 259). German
transl.: Wahrscheinlichkeitsrechnung. Leipzig, 1899. Its reprint: Frankfurt/Main,
1999. Russian transl. of pt. 4: 1913 reprinted in Bernoulli, J. (1986, pp. 23 – 59).
References in text to German translation. English translation by Edith Dudley Sylla
of 2006 (Sheynin 2006a) is rotten and dangerous.
--- (1713b), Lettre à un amy sur les parties du jeu de paume. Reprint: J. Bernoulli
(1975, pp. 260 – 286).
--- (1975), Werke, Bd. 3. Basel. Includes reprints of several memoirs of other
authors and commentaries.
--- (1986), O Zakone Bolshikh Chisel (On the Law of Large Numbers). Moscow.
Ed., Yu. V. Prokhorov. Contains reprint of Markov (1914b) and commentaries.
--- (2005), On the Law of Large Numbers. Berlin. This is my translation of pt. 4 of
the Ars Conjectandi.
Bernoulli Johann III (1785), Milieu à prendre entre les observations. Enc.
Méthodique. Mathématiques, t. 2. Paris, pp. 404 – 409.
Bernoulli Niklaus (1709), De Usu Artis Conjectandi in Jure. Reprint: Bernoulli J.
(1975, pp. 289 – 326). English translation is available in the Internet, but it is hardly
useful.
Bernstein S. N. (1926), Sur l’extension du théorème limite du calcul des
probabilités aux sommes de quantités dépendantes. Math. Annalen, Bd. 97, pp. 1 –
59.
--- (1928, Russian), The present state of the theory of probability and its
applications. In Bernstein (1964, pp. 217 – 232). S, G, 7.
--- (1940, Russian), The Petersburg school of the theory of probability. Uch. Zapiski
Leningradsk. Gos. Univ. No. 55 (Ser. math. sci. No. 10), pp. 3 – 11.
--- (1945, Russian), On Chebyshev’s work on the theory of probability. In Bernstein
(1964, pp. 409 – 433).
--- (1946), Teoria Veroiatnostei (Theory of Prob.). Moscow – Leningrad. 4th edition.
--- (1964), Sobranie Sochineniy (Coll. Works), vol. 4. Moscow.
Bertillon Alph. (1893), Instructions signalétiques. Melun.
Bertrand J. (1887a), Solution d’une problème. C. r. Acad. Sci. Paris, t. 105, p. 369.
--- (1887b), Sur les épreuves répétées. Ibidem, pp. 1201 – 1203.
--- (1888a), Calcul des probabilités. 2nd ed., 1907. Reprints: New York, 1970, 1972.
Second ed. practically coincides with the first one.
--- (1888b), Sur l’évaluation a posteriori de la confiance méritée par la moyenne
d’une série de mesures. C. r. Acad. Sci. Paris, t. 106, pp. 887 – 891.
--- (1888c), Sur l’erreur à craindre dans l’évaluation des trois angles d’un triangle.
Ibidem, pp. 967 – 970.
Besikovitch A. S. (1924, Russian), Biographical essay [of A.A. Markov]. In Markov
(1924, pp. iii – xiv).
Bessel F. W. (1816), Untersuchungen über die Bahn des Olbersschen Kometen.
Abh. Preuss. Akad. Berlin, math. Kl. 1812 – 1813, pp. 119 – 160. Bessel (1876) only
contains a passage from this contribution.
--- (1818), Fundamenta astronomiae. Königsberg. See a fragment in a German
translation (Schneider 1988, pp. 277 – 279).
--- (1820), Beschreibung des auf des Königsberger Sternwarte. Astron. Jahrb.
(Berlin) für 1823, pp. 161 – 168.
--- (1823), Persönliche Gleichung bei Durchgangsbeobachtungen. In Bessel (1876,
Bd. 3, pp. 300 – 304).
--- (1838а), Untersuchung über die Wahrscheinlichkeit der Beobachtungsfehler. In
Bessel (1876, Bd. 2, pp. 372 – 391).
--- (1838b), Gradmessung in Ostpreussen. Berlin. Partly reprinted in Bessel (1876,
No. 135).
--- (1843), Sir William Herschel. Ibidem, Bd. 3, pp. 468 – 478.
--- (1845), Übervölkerung. Ibidem, pp. 387 – 407.
--- (1848a), Über Wahrscheinlichkeitsrechnung. In Bessel (1848b, pp. 387 – 407).
--- (1848b), Populäre Vorlesungen über wissenschaftliche Gegenstände. Hamburg.
Editor, H. C. Schumacher.
--- (1876), Abhandlungen, Bde 1 – 3. Leipzig. Editor, R. Engelmann.
Bienaymé I. J. (1838), Sur la probabilité des résultats moyens des observations.
Mém. рres. Acad. Roy. Sci. Inst. France, t. 5, pp. 513 – 558.

266
264
--- (1839), Théorème sur la probabilité des résultats moyens des observations. Soc.
Philomat. Paris, Extraits, sér. 5, pp. 42 – 49. Also: L’Institut, t. 7, No. 286, pp. 187
– 189.
--- (1840a), Principe nouveau du calcul des probabilités avec ses applications aux
sciences d’observation. Soc. Philomat. Paris, Extraits, sér. 5, pp. 37 – 43. Also:
L’Institut, t. 8, No. 333, pp. 167 – 169.
--- (1840b), Quelques propriétés des moyens arithmétiques de puissances de
quantités positives. Ibidem, pp. 67 – 68. Also: L’Institut, t. 8, No. 342, pp. 216 –
217.
--- (1845), De la loi de multiplication et de la durée des familles. Reprint: D. G.
Kendall (1975, pp. 251 – 253).
--- (1852), Sur la probabilité des erreurs d’après la méthode des moindres carrés. J.
Math. Pures Appl., sér. 1, t. 17, pp. 33 – 78. Also: Mém. рres. Acad. Sci. Inst.
France, sér. 2, t. 15, 1858, pp. 615 – 663.
--- (1853), Considérations à l’appui de la découverte de Laplace sur la loi de
probabilité dans la méthode des moindres carrés. C. r. Acad. Sci. Paris, t. 37, pp.
309 – 324. Also: J. Math. Pures Appl. , sér. 2, t. 12, 1867, pp. 158 – 176.
--- (1855), Sur un principe que M. Poisson avait cru découvrir et qu’il avait appelé
Loi des grands nombres. C. r. Acad. Sci. Morales et Politiques, sér. 3, t. 11, pp. 379
– 389. Also: J. Soc. Stat. Paris, 1876, pp. 199 – 204.
--- (1874), Sur une question de probabilités. Bull. Soc. Math. France, t. 2, pp. 153 –
154.
--- (1875), Application d’un théorème nouveau du calcul des probabilités. C. r.
Acad. Sci. Paris, t. 81, pp. 417 – 423. Also: Bull. Math. Astr., t. 9, pp. 219 – 225.
Biermann K.-R. (1955), Über eine Studie von Leibniz zu Fragen der
Wahrscheinlichkeitsrechnung. Forschungen und Fortschritte, Bd. 29, No. 4, pp. 110
– 113.
--- (1957, Russian), The problems of the Genoese lottery in the works of classics of
the theory of probability. IMI, vol. 10, pp. 649 – 670. S, G, 36.
--- (1966), Über die Beziehungen zwischen Gauss und Bessel. Mitt. Gauss-Ges.
Göttingen, Bd. 3, pp. 7 – 20. S, G, 72
Biermann K.-R., Faak, M. (1957), Leibniz’ “De incerti aestimatione”.
Forschungen und Fortschritte, Bd. 31, No. 2, pp. 45 – 50.
Biot J. B. (1811), Traité élémentaire d’astronomie physique, t. 2. Paris – St.
Pétersbourg. 2nd ed.
--- (1855), Sur les observatoires météorologiques permanents que l’on propose
d’établir en divers points de l’Algérie. C. r. Acad. Sci. Paris, t. 41, pp. 1177 – 1190.
Birg S., Editor (1986), Ursprunge der Demographie in Deutschland. Leben und
Werke J. P. Süssmilch’s. [Coll. Papers.] Frankfurt/Main.
Black W. (1788), Comparative View of the Mortality of the Human Species.
London.
Block M. (1878), Traité théorique et pratique de statistique. Paris, 1886.
Böckh R. (1893), Halley als Statistiker. Bull. ISI, t. 7, No. 1, pp. 1 – 24.
Boltzmann L. (1868), Studien über das Gleichgewicht der lebenden Kraft. In
Boltzmann (1909, Bd. 1, pp. 49 – 96).
--- (1871), Analytischer Beweis des zweiten Hauptsatzes. Ibidem, pp. 288 – 308.
--- (1872), Weitere Studien über das Wärmegleichgewicht. Ibidem, pp. 316 – 402.
--- (1878), Weitere Bemerkungen über einige Probleme. Ibidem, Bd. 2, pp. 250 –
288.
--- (1886), Der zweite Hauptsatz der mechanischen Wärmetheorie. In Boltzmann
(1905, pp. 25 – 50).
--- (1887), Über die mechanischen Analogien des zweiten Hauptsatzes. In
Boltzmann (1909, Bd. 3, pp. 258 – 271).
--- (1895), On certain questions of the theory of gases. Ibidem, pp. 535 – 544.
--- (1895 – 1899), Vorlesungen über Gastheorie, Bde 1 – 2. Leipzig.
--- (1896), Entgegnung auf die […] Betrachtungen des Hrn. Zermelo. In Boltzmann
(1909, Bd. 3, pp. 567 – 578).
--- (1904a), Entgegnung auf einen von [...] Ostwald [...] gehaltenen Vortrag. In
Boltzmann (1905, pp. 364 – 378).
--- (1904b), Vorlesungen über die Prinzipe der Mechanik, Tl. 2. Leipzig.
--- (Lectures 1900 – 1902), Über die Principien der Mechanik. In Boltzmann (1905,
pp. 308 – 337).

267
265
--- (1905), Populäre Schriften. Leipzig, 1925. Later ed., 1979.
--- (1909), Wissenschaftliche Abhandlungen, Bde 1 – 3. Leipzig.
Bomford G. (1971), Geodesy. Oxford. First two eds: 1952, 1962. [Oxford, 1980.]
Bond G. P. (1857), On the use of equivalent factors in the method of least squares.
Mem. Amer. Acad. Arts Sciences, new ser., vol. 6, pt. 1, pp. 179 – 212.
Boole G. (1851), On the theory of probabilities. In author’s book (1952, pp. 247 –
259).
--- (1854a), On the conditions by which the solution of questions in the theory of
probabilities are limited. Ibidem, pp. 280 – 288.
--- (1854b), Laws of Thought. New York, 2003.
--- (1952), Studies in Logic and Probability, vol. 1. London.
Borel E. (1943), Les probabilités et la vie. Engl. transl.: New York, 1962.
Bortkevich V. I. (1889, Russian), On Russian mortality. Vrach, vol. 10, pp. 1053 –
1056.
--- (1903, Russian), The theory of probability and the struggle against sedition.
Osvobozhdenie. Stuttgart, book 1, pp. 212 – 219. Signed “B”. S, G, 4.
Bortkevich V. I., Chuprov A. A. (2005), Perepiska (Correspondence) 1895 – 1926.
Berlin. S, G, 9.
Bortkiewicz L. von (1894 – 1896), Kritische Betrachtungen zur theoretischen
Statistik. JNÖS, 3. Folge, Bde 8, 10, 11, pp. 641 – 680, 321 – 360, 701 – 705.
--- (1898a), Das Gesetz der kleinen Zahlen. Leipzig.
--- (1898b), Das Problem der Russischen Sterblichkeit. Allg. stat. Archiv, Bd. 5, pp.
175 – 190, 381 – 382.
--- (1904), Anwendungen der Wahrscheinlichkeitsrechnung auf Statistik. Enc. math.
Wiss., Bd. 1, pp. 821 – 851.
--- (1915a), Realismus und Formalismus in der mathematischen Statistik. Allg. stat.
Archiv, Bd. 9, pp. 225 – 256.
--- (1915b), Lexis zum Gedächtnis. Z. f. d. gesamte Versicherungs-Wiss., Bd. 15,
pp. 117 – 123.
--- (1915c), Lexis. Nekrolog. Bull. ISI, t. 20, No. 1, pp. 328 – 332.
--- (1917), Die Iterationen. Berlin.
--- (1923), Wahrscheinlichkeit und statistische Forschung nach Keynes. Nord. Stat.
Tidskrift, Bd. 2, pp. 1 – 23.
--- (1930), Lexis und Dormoy. Nordic Stat. J., vol. 2, pp. 37 – 54.
Boscovich R. (1758). Philosophiae Naturalis Theoria. Latin – English edition:
Chicago – London, 1922. English translation from the edition of 1763: Theory of
Natural Philosophy. Cambridge, Mass., 1966.
Bouillaud J. (1836), Essai sur la philosophie médicale. Paris.
Boyle R. (posthumous, 1772), A physico-chymical essay. Works, vol. 1. Sterling,
Virginia, 1999, pp. 359 – 376.
Brendel M. (1924), Über die astronomische Arbeiten von Gauss. In Gauss, W-11,
Tl. 2, Abt. 3. Separate paging.
Brownlee J. (1915), Historical note on Farr’s theory of epidemic. Brit. Med. J., vol.
2, pp. 250 – 252.
Bru B. (1981), Poisson, le calcul des probabilités et l’instruction publique. In
Métivier et al (1981, pp. 51 – 94). English translation: J. Electron. Hist. Probab.
Stat., 2005.
--- (1988), Laplace et la critique probabiliste des mesures géodésiques. In Lacombe,
H., Costabel, P. (1988, pp. 223 – 244).
--- (1991), A la recherche de la démonstration perdue de Bienaymé. Math. Inf. Sci.
Hum., 29e année, No. 114, pp. 5 – 17.
--- (2013), Poisson et le calcul des probabilités. In Poisson (2013, pp. 333 – 355).
Bru B., Bru M.-F. Bienaymé O. (1997), La statistique critiquée par le calcul des
probabilités. Rev. Hist. Math., t. 3, pp. 137 – 239.
Bru B., Jongmans F. (2001), Bertrand. In Heyde et al (2001, pp. 185 – 189).
Brush S. G. (1968), Brownian movement from Brown to Perrin. AHES, vol. 5, pp.
1 – 36. Reprint: M. G. Kendall et al (1977, pp. 347 – 382).
--- (1976), The Kind of Motion We Call Heat, Bks 1 – 2. Amsterdam.
Bühler G., Editor (1886), Laws of Manu. Oxford, 1967.
Buffon G. L. L. (1777), Essai d’arithmétique morale. In Buffon (1954, pp. 456 –
488).

268
266
--- (1954), Œuvres philosophiques. Paris. Editors, J. Piveteau, M. Fréchet, C.
Bruneau.
Bull J. P. (1959), Historical development of clinical therapeutic trials. J. Chronic
Diseases, vol. 10, pp. 218 – 248.
Buniakovsky V. Ya. (1836, Russian), Determination of the probability that a
randomly chosen quadratic equation with integral coefficients has real roots. Mém.
Imp. Acad. Sci. St. Pétersb., 6me sér., Sci. math., phys. et natur., t. 3 (Sci. math. et
phys., t. 1), No. 4, pp. 341 – 351.
--- (1836 – 1837, Russian), On the application of the analysis of probabilities to
determining the approximate values of transcendental numbers. Ibidem, pp. 457 –
467 and No. 5, pp. 517 – 526.
--- (1846), Osnovania Matematicheskoi Teorii Veroiatnostei (Principles of the Math.
Theory of Probability). Petersburg. Abbreviated “Detailed Contents” and chapter
“History of Probability” reprinted in Prokhorov (1999b, pp. 867 – 869 and 863 –
866).
--- (1847, Russian), On the possibility of introducing definite measures of
confidence in the results of some observational sciences and in the first place of
statistics. Sovremennik, vol. 3, pp. 36 – 49.
--- (1850), Sur une application curieuse de l’analyse des probabilités. Mém. Imp.
Acad. Sci. St. Pétersb., 6me sér., Sci. math., phys. et natur., t. 6 (Sci. math. et phys., t.
4), No. 3, pp. 233 – 258.
--- (1866a, Russian), Essay on the law of mortality in Russia and on the distribution
of the Orthodox believers by ages. Zapiski Imp. Akad. Nauk St. Petersb., vol. 8, No.
6. Separate paging.
--- (1866b, Russian), Tables of mortality and of population for Russia. Mesiatseslov
(Calendar) for 1867. Petersburg. Supplement, pp. 3 – 53.
--- (1868, Russian), A few remarks on the laws of the movement of the population in
Russia. Russk. Vestnik, vol. 73, pp. 5 – 20.
--- (1871, Russian), On a special kind of combinations occurring in problems
connected with defects. Zapiski Imp. Akad. Nauk St. Petersb., vol. 20, No. 2.
Separate paging.
--- (1874, Russian), Anthropological [Anthropometric] studies. Ibidem, vol. 23, No.
5. Separate paging.
--- (1875a, Russian), On a problem concerning the partition of numbers. Ibidem, vol.
25, No. 1. Separate paging.
--- (1875b, Russian), On the probable number of men in the contingents of the
Russian army in 1883, 1884 and 1885. Ibidem, No. 7. Separate paging.
--- (1880, Russian), On maximal quantities in issues concerning moral benefit.
Ibidem, vol. 36, No. 1. Separate paging.
Burov V. G., Viatkin R. V., Titarenko M. A., Editors (1972 – 1973),
Drevnekitaiskaia Filosofia (Ancient Chinese Philosophy), vols. 1 – 2. Moscow.
Bursill-Hall P., Еditor (1993), R. J. Boscovich. His Life and Scientific Work. Rome.
Buys Ballot C. H. D. (1847), Les changements périodiques de temperature. Utrecht.
--- (1850), Die periodischen […] Änderungen der Temperatur. Fortschritte Phys.,
Bd. 3 für 1847, pp. 623 – 629.
Byrne E. F. (1968), Probability and Opinion. The Hague.
Campbell L., Garnett W. (1882), Life of Maxwell. London, 1884. New York –
London, 1969.
Сauchy A. L. (1821), Course d’analyse de l’Ecole royale polytechnique. OC, sér. 2,
t. 3. Paris, 1897.
--- (1845), Sur les secours que les sciences du calcul peuvent fournir aux sciences
physiques ou même aux sciences morales. OC, sér. 1, t. 9. Paris, 1896, pp. 240 –
252.
--- (1853a), Sur la nouvelle méthode d’interpolation comparée à la méthode des
moindres carrés. OC, t. 12. Paris, 1900, pp. 68 – 79.
--- (1853b), Sur les résultats moyens d’observations de même nature, et sur les
résultats les plus probables. Ibidem, pp. 94 – 104.
--- (1853c), Sur la probabilité des erreurs qui affectent des résultats moyens
d’observations de même nature. Ibidem, pp. 104 – 114.
--- (1853d), Sur les résultats moyens d’un très grand nombre d’observations. Ibidem,
pp. 125 – 130.
Celsus A. C. (1935, English), De Medicina, vol. 1. London. Written in 1st century.

269
267
Chadwick E. (1842), Report on the Sanitary Condition of the Labouring
Population. Edinburgh, 1965.
Chapman S. (1941), Halley As a Physical Geographer. London.
Chatterjee S. K. (2003), Statistical Thought: a Perspective and History. Osford.
Chaufton A. (1884), Les assurances, t. 1. Paris.
Chebotarev A. S. (1958), Sposob Naimenshikh Kvadratov etc. (Method of Least
Squares). Moscow.
Chebyshev P. L. (1845, Russian), An essay on an elementary analysis of the theory
of probability. In Chebyshev (1944 – 1951, vol. 5, pp. 26 – 87).
--- (1846), Démonstration élémentaire d’une proposition générale de la théorie des
probabilités. J. reine u. angew. Math., Bd. 33, pp. 259 – 267.
--- (1867), Des valeurs moyennes. J. Math. Pures et Appl., t. 12, pp. 177 – 184.
--- (1870), Appendice to Mayevski (1870), Formules d’interpolation par la méthode
des moindres carrés. Mémoires couronnés et mémoires des savants étrangers. Acad.
Roy. des Sciences, des Lettres et des Beau-Arts de Belg., t. 21, pp. 25 – 33.
--- (1874), Sur les valeurs limites des intégrales. Ibidem, t. 19, pp. 157 – 160.
--- (Lectures 1879/1880), Teoria Veroiatnostei (Theory of Probability). Moscow –
Leningrad, 1936. S, G, 3.
--- (1887a, Russian), Sur les résidus intégraux qui donnent des valeurs approchées
des intégrales. Acta Math., t. 12, 1888 – 1889, pp. 287 – 322.
--- (1887b, Russian), Sur deux théorèmes relatifs aux probabilités. Ibidem, t. 14,
1890 – 1891, pp. 305 – 315.
--- (1899 – 1907), Oeuvres, tt. 1 – 2. Pétersbourg. Reprint: New York, 1962.
--- (1944 – 1951), Polnoe Sobranie Sochineniy (Complete Works), vols. 1 – 5.
Moscow – Leningrad.
Chetverikov N. S., Editor (1968), O Teorii Dispersii (On the Theory of
Dispersion). Moscow.
Chirikov M. V., Sheynin O. B. (1994), see Sheynin (1994f).
Chuprov A. A. (1905), Die Aufgaben der Theorie der Statistik. Schmollers Jahrb. f.
Gesetzgebung, Verwaltung u. Volkwirtschaft in Dtsch. Reiche, Bd. 29, pp. 421 –
480. Quoted from its Russian translation (Chuprov 1960, pp. 43 – 90).
--- (1909), Ocherki po Teorii Statistiki (Essays on the Theory of Statistics).
Moscow, 1959. Second ed., 1910.
--- (1912, Russian), Sampling investigation. Translated in Sheynin, O., Chuprov’s
early paper on sampling. JNÖS, Bd. 216, 1997, pp. 658 – 671.
--- (1916, Russian), On the expectation of the coefficient of dispersion. Izvestia Imp.
Akad. Nauk, vol. 10, pp. 1789 – 1798. S, G, 35.
--- (1916 or very beginning of 1917, Russian), On the mean square error of the
coefficient of dispersion. Manuscript, Archive, Russian Acad. Sci., Fond 173, Inv. 1,
No. 50. S, G, 35.
--- (1918 – 1919), Zur Theorie der Stabilität statistischer Reihen. Skand.
Aktuarietidskr., t. 1 – 2, pp. 199 – 256, 80 – 133.
--- (1922a); Lehrbücher der Statistik. Nordisk Statistisk Tidskrift, Bd. 1, pp. 139 –
160, 329 – 340. Reprinted in Chuprov (1960, pp. 413 – 429).
--- (1922b), Das Gesetz der großen Zahlen und der stochastisch-statistische
Standpunkt […]. Nordisk Statistisk Tidskrift, Bd. 1, pp. 39 – 67. Reprinted in
Chuprov (1960, pp. 141 – 162).
--- (1923), On the mathematical expectation of the moments of frequency
distributions in the case of correlated observations. Metron, t. 2, pp. 461 – 493, 646
– 683.
--- (1924), Ziele und Wege der stochastischen Grundlagen der statistischen Theorie.
Nord. Stat. Tidskrift, t. 3, pp. 433 – 493.
--- (1925a), Grundbegriffe und Grundprobleme der Korrelationstheorie. Leipzig –
Berlin. Quoted from its Russian version of 1926. English translation: London, 1939.
--- (1925b), Review of Markov (1924). In Ondar (1977/1981, pp. 154 – 157).
--- (1926), Teorien för statistiska räckors stabilitet. Nord. Stat. Tidskrift, Bd. 5, pp.
195 – 212. Russian translation: Chuprov (1960, pp. 224 – 239). S, G, 35.
--- (1960), Voprosy Statistiki (Issues in Statistics). Reprints and/or translations of
papers. Moscow.
--- (2009), Pis’ma (Letters to) K. N. Gulkevich, 1919 – 1921. Co-compilers C.
Wittich, G. Kratz. Berlin. S, G, 28.
Cicero(n) M. T. (1991), Über die Wahrsagung. München – Zürich.

270
268
--- (1997), Nature of the Gods. Oxford.
Cioffari V. (1935), Fortune and Fate from Democritus to St. Thomas Aquinas. New
York.
Clausius R. (1849), Über die Natur derjenigen Bestandteile der Erdatmosphäre.
Annalen Phys. Chem., Bd. 76 (152), pp. 161 – 188.
--- (1857), Über die Art der Bewegung welche wir Wärme nennen. In Clausius
(1867, pp. 229 – 259).
--- (1858), Über die mittlere Länge der Wege ... bei der Molecularbewegung.
Ibidem, pp. 260 – 276.
--- (1862), Über die Wärmeleitung gasförmiger Körper. Ibidem, pp. 277 – 326.
--- (1867), Abhandlungen über die mechanische Wärmetheorie, Abt. 2.
Braunschweig.
--- (1889 – 1891), Die kinetische Theorie der Gase. Braunschweig.
Clerke Agnes M. (ca. 1885), Popular History of Astronomy. London, 1893.
[London, 1902.]
Clifford W. K. (1885), Common Sense of the Exact Sciences. New York 1946. K.
Pearson essentially supplemented the posthumous edition of 1886.
Cohen J., Chesnick E. I. (1970), The doctrine of psychological chances. Brit. J.
Psychol., vol. 61, pp. 323 – 334.
Cohen J., Chesnick E. I., Haran D. (1971), Evaluation of compound probabilities
in sequential choice. Nature, vol. 232, pp. 414 – 416.
Colebrooke H. T. (1817), Algebra and Mensuration from the Sanscrit of
Brahmegupta and Bhascara. London. [Wiesbaden, 1973.]
Commelin C. (1693), Beschryvinge der Stadt Amsterdam. Amsterdam.
Comte A. (1830 – 1842), Cours de philosophie positive, tt. 1 – 6. Paris, 1893, t. 3.
Condamine C. M. de la (1751), Mesure des trois premièrs dégrés du méridien.
Paris.
--- (1759), Sur l’inoculation de la petite vérole. Hist. Acad. Roy. Sci. Paris 1754
avec Mém. math. et phys., pp. 615 – 670 of the Mémoires.
--- (1763), Second mémoire sur l’inoculation […]. Ibidem, pp. 439 – 482 of the
Mémoires.
--- (1773), Histoire de l’inoculation. Amsterdam.
Condorcet M. J. A. N. (1784), Sur le calcul des probabilités. Hist. Acad. Roy. Sci.
Paris 1781 avec Mém. Math. et Phys. pour la même année. Paris, 1784, pp. 707 –
728.
--- (1785a), Essai sur l’application de l’analyse à la probabilité des décisions
rendues à la pluralité des voix. New York, 1972.
--- (1785b), [An anti-scientific] Eloge de M. [D.] Bernoulli. Hist. Acad. Roy. Sci.
Paris 1782 avec Mém. Math. et Phys., pp. 82 – 107 of the Histoire. Reprint in
author’s Œuvres, t. 2. Paris, 1847, pp. 545 – 580. S, G, 39.
--- (1795), Esquisse d’un tableau historique des progrès de l’esprit humain suivi de
Fragment sur l’Atlantide. Paris, 1988.
Cook A. H. (1998), Halley: Charting the Heavens and the Seas. Oxford.
Coolidge J. L. (1926), Adrain and the beginnings of American mathematics. Amer.
Math. Monthly, vol. 33, No. 2, pp. 61 – 76.
Cornfield J. (1967), The Bayes theorem. Rev. ISI, vol. 35, pp. 34 – 49.
Cotes R. (1722), Aestimatio errorum in mixta mathesis per variationes partium
trianguli plani et sphaerici. Opera Misc. London, 1768, pp. 10 – 58.
Cotte L. (1788), Sur l’utilité des observations météorologiques. In author’s
Mémoires sur la météorologie, tt. 1 – 2. Paris (t. 1, pp. 1 – 32).
Cournot A. A. (1838), Sur l’applications du calcul des chances à la statistique
judiciaire. J. Math. Pures et Appl., sér. 1, t. 3, pp. 257 – 334.
--- (1843), Exposition de la théorie des chances et des probabilités. Paris, 1984.
Editor B. Bru. S, G, 54.
--- (1851), Essai sur les fondements de nos connaissances et sur les caractères de la
critique philosophique. Paris, 1975.
--- (1861), Traité de l’enchainement des idées fondamentales dans les sciences et
dans l’histoire. Paris, 1982.
--- (1875), Matérialisme, vitalisme, rationalisme. Paris, 1979.
Courtault J.-M. et al (2000), Bachelier. On the centenary of “Théorie de la
spéculation”. Math. Finance, vol. 10, pp. 341 – 353.
Couturat L. (1901), La logique de Leibniz. Paris.

271
269
Cramér H. (1946), Mathematical Methods of Statistics. Princeton. 13th printing,
1974.
--- (1976), Half a century with probability theory: some personal recollections.
Annals Prob., vol. 4, pp. 509 – 546.
--- (1981), Mathematical probability and statistical inference. ISR, vol. 49, pp.
309 – 317.
Crofton M. W. (1869), On the theory of local probability applied to straight lines
drawn at random in a plane. PT Roy. Soc., vol. 158 for 1868, pp. 181 – 199.
Cubranic N. (1961), Geodetski rad R. Boskovica. Zagreb.
Czuber E. (1891), Zur Kritik einer Gauss’schen Formel. Monatshefte Math. Phys.,
Bd. 2, pp. 459 – 464.
--- (1908), Wahrscheinlichkeitsrechnung und ihre Anwendung, Bd. 1. New York,
1968. First edition, 1903.
--- (1921), Die statistischen Forschungsmethoden. Wien.
D’Alembert J. Le Rond (1754), Croix ou pile. Enc. ou dict. raisonné des sciences,
des arts et des métiers, t. 4. Stuttgart, 1966, pp. 512 – 513.
--- (1759), Essai sur les élémens de philosophie. OC, t. 1, pt. 1. Paris, 1821, pp. 116
– 348.
--- (1761a), Sur le calcul des probabilités. Opusc. Math., t. 2. Paris, pp. 1 – 25.
--- (1761b), Sur l’application du calcul des probabilités à l’inoculation de la petite
vérole. Ibidem, pp. 26 – 95.
--- (1768a), Doutes et questions sur le calcul des probabilités. In author’s book
Mélanges de litterature, d’histoire et de philosophie, t. 5. Amsterdam, pp. 239 –
264.
--- (1768b), Sur la durée de la vie. Opusc. Math., t. 4. Paris, pp. 92 – 98.
--- (1768c), Sur un mémoire de M. Bernoulli concertant l’inoculation. Ibidem, pp.
98 – 105.
--- (1768d), Sur le calcul des probabilités. Ibidem, pp. 283 – 310.
D’Amador R. (1837), Mémoire sur le calcul des probabilités appliqué à la
médecine. Paris.
Danilevsky N. Ya. (1885, Russian), Darvinism, vol. 1, pts 1 – 2. Petersburg.
Darboux G. (read 1901; 1902), Eloge historique de Bertrand. In author’s Eloges
académiques et discours. Paris, pp. 1 – 60.
Darwin C. (1859), Origin of Species. London – New York, 1958. [Manchester,
1995.]
--- (1868), The Variation of Animals and Plants under Domestication, vols. 1 – 2.
London, 1885. [London, 1988.]
--- (1876), The Effects of Cross and Self-Fertilisation in the Vegetable Kingdom.
London, 1878. [London, 1989.]
--- (1881), The Formation of Vegetable Mould. London, 1945. [London, 1989.]
--- (1887), Life and Letters. New York – London, 1897, vols. 1 – 2. [New York,
1969.]
--- (1903), More Letters, vol. 1. London.
David F. N. (1962), Games, Gods and Gambling. London.
David F. N., Neyman J. (1938), Extension of the Markoff theorem on least squares.
Stat. Res. Memoirs, vol. 2, pp. 105 – 117.
David H. A. (1957), Some notes on the statistical papers of Helmert. Bull. Stat. Soc.
New South Wales, vol. 19, pp. 25 – 28. First publ. Ibidem, 1954.
--- (1963), The Method of Paired Comparisons. London – New York, 1988.
--- (1998), Early sample measures of variability. Stat. Sci., vol. 13, pp. 368 – 377.
David H. A., Edwards A. W. F. (2001), Annotated Readings in the History of
Statistics. New York.
Davidov A. Yu. [1854], Lektsii Matematicheskoi Teorii Veroiatnostei (Lectures on
Math. Theory of Probability). N. p., n. d.
--- (1857), Theory of mean quantities with its application to compilation of mortality
tables. In Rechi i Otchet, Proiznesennye v Torzhestvennom Sobranii Moskovskogo
Universiteta (Orations and Report at the Grand Meeting of Moscow Univ.).
Moscow, first paging.
--- [1885], Teoria Veroiatnostei, 1884 – 1885 (Theory of Probability). N. p., n. d.
--- (1886, Russian), On mortality in Russia. Izvestia Imp. Obshchestva Liubitelei
Estestvoznania, Antropologii i Etnografii, vol. 49, No. 1, pp. 46 – 66.

272
270
Daw R. H. (1980), J. H. Lambert, 1727 – 1777. J. Inst. Actuaries, vol. 107, pp. 345
– 350.
Dawid Ph. (2005), Statistics on trial. Significance, vol. 2, No. 1, pp. 6 – 8.
Daxecker F. (2004), The Physicist and Astronomer Christopher Scheiner.
Innsbruck.
--- (2006), Die Hauptwerk des Astronomen P. Christoph Scheiner JJ. Innsbruck.
DeCandolle Alph. (1855), Géographie botanique raisonnée, tt. 1 – 2. Paris.
DeCandolle Aug. P. (1832), Physiologie végétale, tt. 1 – 3. Paris.
Delambre J. B. J. (1819), Analyse des travaux de l’Académie … pendant l’année
1817, partie math. Mém. Acad. Roy. Sci. Inst. de France, t. 2 pour 1817, pp. I –
LXXII of the Histoire.
Demografichesky (1985), Demografichesky Enziklopedichesky Slovar
(Demographic Enc. Dict.). Moscow.
De Moivre A. (1712, in Latin). English transl.: De mensura sortis or the
measurement of chance. ISR, vol. 52, 1984, pp. 236 – 262. Commentary (A. Hald):
Ibidem, pp. 229 – 236.
--- (1718), Doctrine of Chances. Later editions: 1738, 1756. References in text to
reprint of last edition: New York, 1967.
--- (1725), Treatise on Annuities on lives. London. Later edition of 1743
incorporated in the Doctrine (1756, pp. 261 – 328) in a somewhat improved form, as
its anonymous editor stated on p. xi. German transl.: Wien, 1906.
--- (1730), Miscellanea Analytica de Seriebus et Quadraturis. London. French
translation: Paris, 2009.
--- (1733, Latin), Transl. by author: A method of approximating the sum of the terms
of the binomial (a + b)n expanded into a series from whence are deduced some
practical rules to estimate the degree of assent which is to be given to experiments.
Incorporated in subsequent editions of the Doctrine (in 1756, in an extended version,
on pp. 243 – 254).
--- (1756), This being the last edition of the Doctrine.
De Montessus R. (1903), Un paradoxe du calcul des probabilités. Nouv. Annales
Math., sér. 4, t. 3, pp. 21 – 31.
De Morgan A. (1845), Theory of probabilities. Enc. Metropolitana, Pure sciences,
vol. 2. London, pp. 393 – 490.
--- (1847), Formal Logic, or the Calculus of Inference, Necessary and Probable.
Second ed.: Chicago, 1926.
--- (1864), On the theory of errors of observation. Trans. Cambr. Phil. Soc., vol. 10,
pp. 409 – 427.
De Morgan Sophia Elizabeth (1882), Memoir of Augustus De Morgan. London.
Descartes R. (1644), Les principes de la philosophie. Œuvres, t. 9, pt. 2 (the whole
issue). Paris, 1978. Reprint of the edition of 1647.
DeVries W. F. M. (2001), Meaningful measures: indicators on progress, progress
on indicators. ISR, vol. 69, pp. 313 – 331.
De Witt J. (1671, in vernacular), Value of life annuities in proportion to redeemable
annuities. In Hendriks (1852, pp. 232 – 249).
Dietz K. (1988), The first epidemic model: historical note on P. D. Enko. Austr. J.
Stat., vol. 30A, pp. 56 – 65.
Dietz K., Heesterbeek J. A. P. (2000), D. Bernoulli was ahead of modern
epidemiology. Nature, vol. 408, pp. 513 – 514.
--- (2002), D. Bernoulli’s epidemiological model revisited. Math. Biosciences, vol.
180, pp. 1 – 21.
O’Donnell T. (1936), History of Life Insurance. Chicago.
Doob J. L. (1989), Commentary on probability. In Centenary of Math. in America,
pt. 2. Providence, Rhode Island, 1989, pp. 353 – 354. Editors P. Duren et al.
Dorfman Ya. G. (1974), Vsemirnaia Istoria Fiziki (Intern. History of Physics).
Moscow.
Dormoy E. (1874), Théorie mathématique des assurances sur la vie. J. des actuaries
française, t. 3, pp. 283 – 299, 432 – 461.
--- (1878), Same title, t. 1. Paris. Incorporates his paper of 1874.
Dorsey N. E., Eisenhart C. (1969), On absolute measurement. In Ku (1969, pp. 49
– 55).

273
271
Double F. J., Dulong P. L., Larrey F. H., Poisson S.-D. (1835), Review of Civiale,
Recherches de statistique sur l’affection calculeuse. C. r. Acad. Sci. Paris, t. 1, pp.
167 – 177. Translation: Intern. J. Epidemiology, vol. 30, 2001, pp. 1246 – 1249.
Double F. J. (1837, in French), Inapplicability of statistics to the practice of
medicine. Lond. Medical Gaz., vol. 20, No. 2, pp. 361 – 364. Transl. from the Gaz.
Médicale.
Dove H. W. (1837), Meteorologische Untersuchungen. Berlin.
--- (1839), Über die nicht periodischen Änderungen der Temperaturvertheilung auf
der Oberfläche der Erde. Abh. Preuss. Akad. Wiss. Berlin, Phys. Abh. 1838, pp. 285
– 415.
--- (1850), Über Linien gleicher Monatswärme. Ibidem, Phys. Abh. 1848, pp. 197 –
228.
Dufau P. A. (1840), Traité de statistique ou théorie de l’étude des lois, d’après
lesquelles se développent des faits sociaux. Paris.
Dunnington W. (1955), Gauss: Titan of Science. New York.
Dupâquier J. et M. (1985), Histoire de la démographie. Paris.
Du Pasquier L. G. (1910), Die Entwicklung der Tontinen bis auf die Gegenwart. Z.
schweiz. Stat., 46. Jg, pp. 484 – 513.
Dutka J. (1985), On the problem of random flights. AHES, vol. 32, pp. 351 – 375.
--- (1988), On the St. Petersburg paradox. AHES, vol. 39, pp. 13 – 39.
--- (1990), R. Adrain and the method of least squares. AHES, vol. 41, pp. 171 – 184.
--- (1996), On Gauss’ priority in the discovery of the method of least squares.
AHES, vol. 49, pp. 355 – 370.
Ebbinghaus H. (1908), Abriss der Psychologie. Leipzig.
Eddington A. S. (1933), Notes on the method of least squares. Proc. Phys. Soc.,
vol. 45, pp. 271 – 287.
Edgeworth F. Y. (1996), Writings in Probability, Statistics and Economics, vols 1 –
3. Cheltenham. Editor, C. R. McCann, Jr.
Edwards A. W. F. (1987), Pascal’s Arithmetic Triangle. Baltimore, 2002.
--- (1994), R. A. Fisher on K. Pearson. Notes and Records Roy. Soc. London, vol.
48, pp. 97 – 106.
Eggenberger J. (1894), Darstellung des Bernoullischen Theorems etc. Mitt.
Naturforsch. Ges. Bern No. 1305 – 1334 für 1893, pp. 110 – 182. Also published
separately: Berlin, 1906.
Ehrenfest P. & T. (1907), Über zwei bekannte Einwände gegen das Boltzmannsche
H-Theorem. In Ehrenfest, P. (1959), Coll. Scient. Papers. Amsterdam, pp. 146 –
149.
Eisenhart C. (1963), Realistic evaluation of the precision and accuracy of
instrument calibration. In Ku (1969, pp. 21 – 47).
--- (1964), The meaning of “least” in least squares. J. Wash. Acad. Sci., vol. 54, pp.
24 – 33. Reprint: Ku (1969, pp. 265 – 274).
--- (1974), Pearson. Dict. Scient. Biogr., vol. 10, pp. 447 – 473.
--- (1976), [Discussion of invited papers.] Bull. ISI, vol. 46, pp. 355 – 357.
--- (1978), Gauss. In Kruskal & Tanur (1978, vol. 1, pp. 378 – 386).
Ekeland I. (2006), The Best of All Possible Worlds. Chicago.
Emeliakh L. I. (1954, Russian), The case of A. A. Markov’s excommunication
from the Church. Voprosy Religii i Ateisma, vol. 2, pp. 397 – 411. S, G, 85.
Encke J. F. (1851), Über die Bestimmung der elliptischen Elemente bei
Planetenbahnen. Abh. Kgl. Akad. Wiss. Berlin for 1849, pp. 1 – 68 of second paging.
Eneström G. (1896, Swedish), Sur la méthode de J. de Witt (1671) pour le calcul de
rentes viagères. Archief voor de verzekerings-wetenschap, t. 3, 1897, pp. 62 – 68.
Enko P. D. (1889, Russian), On the course of epidemics of some infectious
diseases. Intern. J. Epidemiology, vol. 18, 1989, pp. 749 – 755. Partial transl. by K.
Dietz.
Erdélyi A. (1956), Asymptotic Expansions. New York.
Erisman F. F. (1887), Kurs Gigieny (A Course in Hygiene), vols. 1 – 2. Moscow.
Vol. 2 contains a supplement on sanitary statistics.
Erman A., Editor (1852), Briefwechsel zwischen W. Olbers und F. W. Bessel, Bde
1 – 2. Leipzig.
Ermolaeva N. S. (1987, Russian), On an unpublished course on the theory of
probability by Chebyshev. Voprosy Istorii Estestvozn. i Tekhniki No. 4, pp. 106 –
112.

274
272
--- (1990, Russian), The method of least squares in Markov’s letter to B. M.
Koialovich. IMI, vol. 13 (48), pp. 89 – 110.
Euler L. (1748, Latin), Introduction to Analysis of the Infinite, Book 1. New York,
1988.
--- (1749), Recherches sur la question des inégalités du mouvement de Saturne et de
Jupiter. Opera Omnia, ser. 2, t. 25. Zürich, 1960, pp. 45 – 157.
--- (1755), Elémens de la trigonométrie sphéroidique tirés de la méthode des plus
grands et plus petits. Ibidem, ser. 1, t. 27. Zürich, 1954, pp. 309 – 339.
--- (1767), Recherches générales sur la mortalité et la multiplication du genre
humain. Ibidem, ser. 1, t. 7. Leipzig, 1923, pp. 79 – 100. His manuscript Sur
multiplication du genre humaine is also there, pp. 545 – 552.
--- (1770), Expositio methodorum, cum pro determinanda parallaxi solis […].
Ibidem, ser. 2, t. 30. Zürich, 1964, pp. 153 – 231.
--- (1778, Latin). [Commentary to D. Bernoulli (1778).] English transl.: 1961, and
its reprint 1970, are together with the transl. and reprint of D. Bernoulli.
Faraday M. (1991 – 2008), Correspondence, vols. 1 – 5. London.
Farebrother R. W. (1993), Boscovich’s method for correcting discordant
observations. In Bursill-Hall (1993, pp. 255 – 261).
--- (1999), Fitting Linear Relationships. History of the Calculus of Observations
1750 – 1900. New York.
Fechner G. T. (1855), Über die physikalische und philosophische Atomenlehre.
Leipzig, 1864.
--- (1860), Elemente der Psychophysik, Bde 1 – 2. Leipzig. Third ed., 1907, its
reprint, 1964.
--- (1874a), Über die Bestimmung des wahrscheinlichen Fehlers eines
Beobachtungsmittels. Annalen Phys. Chem., Jubelbd., pp. 66 – 81.
--- (1874b), Über den Ausgangswert der kleinsten Abweichungssumme […]. Abh.
Kgl. Sächs. Ges. Wiss., Bd. 18 (Bd. 11 of the Math.-Phys. Kl.), No. 1, pp. 3 – 76.
--- (1877), In Sachen der Psychophysik. Leipzig.
--- (1882), Revision der Hauptpunkte der Psychophysik. Leipzig.
--- (1887), Über die Methode der richtigen und falschen Fälle. Abh. Kgl. Sächsische
Ges. Wiss., Bd. 13 (22), pp. 109 – 312.
--- (1897), Kollektivmasslehre. Leipzig. Editor (and actual co-author) G. F. Lipps.
Fedorovitch L. V. (1894), Istoria i Teoria Statistiki (History and Theory of
Statistics). Odessa.
Feller W. (1950), Introduction to Probability Theory and Its Applications, vol. 1.
New York – London. Third ed., 1968.
Fisher R. A. (1920), Mathematical examination of the methods of determining the
accuracy of an observation. Monthly Notices Roy. Astron. Soc., vol. 80, pp. 758 –
770.
--- (1922), On the mathematical foundations of theoretical statistics. PT Roy. Soc.,
vol. A222, pp. 309 – 368.
--- (1936), Has Mendel’s work been rediscovered? Annals of Science, vol. 1, pp. 115
– 137.
--- (1937), Professor K. Pearson and the method of moments. Ibidem, vol. 7, pp. 303
– 318.
--- (1956), Statistical Methods and Statistical Inference. Edition of 1973 reprinted
with separate paging in Fisher (1990).
--- (1990), Statistical Methods, Experimental Design and Scientific Inference.
Oxford.
Fletcher A., Miller J. C. P. et al (1946), Index of Mathematical Tables, vol. 1.
Oxford, 1962.
Forsythe G. E. (1951), Gauss to Gerling on relaxation. Math. Tables and Other
Aids to Computers, vol. 5, pp. 255 – 258.
Fortunatov A. (1914, Russian), To the memory of P. D. Azarevitch. Statistich.
Vestnik No. 1 – 2, pp. 233 – 238.
Fourier J. B. J., (1819), Extrait d’un mémoire sur la théorie analytique des
assurances. Annales chim. phys., t. 10, pp. 177 – 189.
--- Editor (1821 – 1829), Recherches statistiques sur la ville de Paris et de
Départament de la Seine, tt. 1 – 4. Paris.
--- (1826), Sur les résultats moyens déduits d’un grand nombre d’observations.
Œuvres, t. 2. Paris, 1890, pp. 525 – 545.

275
273
--- (1829), Historical Eloge of the Marquis De Laplace. London, Edinburgh and
Dublin Phil. Mag., 2nd ser., vol. 6, pp. 370 – 381. The original French Eloge was
published in 1831: Mém. Acad. Roy. Sci. Inst. de France, t. 10, pp. LXXX – CII.
Fraenkel A. (1930), G. Cantor. Jahresber. Deutsch. Mathematiker-Vereinigung, Bd.
39, pp. 189 – 266.
Franklin J. (2001), The Science of Conjecture. Baltimore.
Fréchet M. (1949), Réhabilitation de la notion statistique de l’homme moyen. In
author’s book Les mathématiques et le concret. Paris, 1955, pp. 317 – 341.
Freud S. (1925), Selbstdarstellung. Werke, Bd. 14. Frankfurt/Main, 1963, pp. 31 –
96.
Freudenthal H. (1951), Das Peterburger Problem in Hinblick auf Grenzwertsätze
der Wahrscheinlichkeitsrechnung. Math. Nachr., Bd. 4, pp. 184 – 192.
--- (1961), 250 years of mathematical statistics. In Quantitative Methods in
Pharmacology. Amsterdam, 1961, pp. xi – xx. Editor H. De Jonge.
--- (1966), De eerste ontmoeting tussen de wiskunde en de sociale wetenschappen.
Verh. Knkl. Vlaamse Acad. Wetenschappen, Letteren en Schone Kunsten Belg., Kl.
Wetenschappen, Jg. 28, No. 88. Separate paging.
--- (1971), Cauchy. Dict. Scient. Biogr., vol. 3, pp. 131 – 148.
Freudenthal H., Steiner H.-G. (1966), Aus der Geschichte der
Wahrscheinlichkeitstheorie und der mathematischen Statistik. In Grundzüge der
Mathematik, Bd. 4. Göttingen, 1966, pp. 149 – 195. Editors H. Behnke et al.
Fuss N. I. (1836), Compte rendu de l’Académie pour 1835. In Recueil des Actes de
la Séance Publique Acad. Imp. Sci. St. Pétersbourg 29 Sept. 1835. St. Pétersbourg,
1836. Separate paging.
Fuss P. N. (1843), Correspondance mathématique et physique de quelques célèbres
géomètres du XVIII siècle, tt. 1 – 2. New York – London, 1968.
Galen C. (1946), On Medical Experience. London. Written in 2nd century.
--- (1951), Hygiene. Springfield, Illinois. Written in 2nd century.
Galilei G. (1613, Italian), History and demonstrations concerning sunspots and their
phenomena. In author’s book Discoveries and Opinions of Galileo. Garden City,
New York, 1957, pp. 88 – 144.
--- (1632, Italian), Dialogue concerning the Two Chief World Systems. Berkeley –
Los Angeles, 1967.
Galitzin B. (1902), Über die Festigkeit des Glasses. Izv. Imp. Akad. Nauk, ser. 5,
vol. 16, No. 1, pp. 1 – 29.
Galle E. (1924), Über die geodätischen Arbeiten von Gauss. In Gauss, W-11/2, Abt.
1. Separate paging.
Galton F. (1863), Meteorographica. London – Cambridge.
--- (1869), Hereditary Genius. London – New York, 1978.
--- (1877), Typical laws of heredity. Nature, vol. 15, pp. 492 – 495, 512 – 514, 532 –
533.
--- (1892), Finger Prints. London.
Gani J. (1982), Newton on a “Question touching ye different Odds upon certain
given Chances upon Dice”. Math. Scientist, vol. 7, pp. 61 – 66.
--- (2001), P. D. Enko. In Heyde & Seneta (2001, pp. 223 – 226).
Garber D., Zabell S. (1979), On the emergence of probability. AHES, vol. 21, pp.
33 – 53.
Gastwirth J. L., Editor (2000), Statistical Science in the Courtroom. New York.
Gatterer J. C. (1775), Ideal einer allgemeinen Weltstatistik. Göttingen.
Gauss C. F. (1809a, German), Preliminary author’s report about Gauss (1809b). In
Gauss (1887, pp. 204 – 205).
--- (1809b, Latin), Theorie der Bewegung, Book 2, Section 3. German transl.
Ibidem, pp. 92 – 117.
--- (1811, Latin), Aus der Untersuchung über die elliptischen Elemente der Pallas.
Ibidem, pp. 118 – 128.
--- (1816), Bestimmung der Genauigkeit der Beobachtungen. Ibidem, pp. 129 – 138.
--- (1821, German), Preliminary author’s report about Gauss (1823b, pt. 1). Ibidem,
pp. 190 – 195.
--- (1823a, German), Preliminary author’s report about Gauss (1823b, pt. 2). Ibidem,
pp. 195 – 199.
--- (1823b, Latin), Theorie der den kleinsten Fehlern unterworfenen Combination
der Beobachtungen, pts. 1 – 2. Ibidem, pp. 1 – 53.

276
274
--- (1823c), Anwendung der Wahrscheinlichkeitsrechnung auf eine Aufgabe der
practischen Geometrie. W-9. Göttingen – Leipzig, 1903, pp. 231 – 237.
--- (1826, German), Preliminary author’s report about Gauss (1828). Ibidem, pp. 200
– 204.
--- (1828a, Latin), Supplement to Gauss (1823b). German transl.: Ibidem, pp. 54 –
91.
--- (1828b), Bestimmung des Breitenunterschiedes zwischen den Sternwarten von
Göttingen und Altona etc. W-9, pp. 5 – 64. S, G, 72.
--- (1845; Nachlass), Anwendung der Wahrscheinlichkeitsrechnung auf die
Bestimmung der Bilanz für Witwenkassen. W-4, 1873, pp. 119 – 183.
--- (1855), Méthode des moindres carrés. Paris.
--- (1863 – 1930), Werke, Bde 1 – 12. Göttingen a. o. Reprint: Hildesheim, 1973 –
1981.
--- (1887), Abhandlungen zur Methode der kleinsten Quadrate. Hrsg, A. Börsch &
P. Simon. Latest ed.: Vaduz, 1998.
Gavarret J. (1840), Principes généraux de statistique médicale. Paris.
Gerardy T. (1977), Die Anfänge von Gauss’ geodätische Tätigkeit. Z. f.
Vermessungswesen, Bd. 102, pp. 1 – 20.
Gingerich O. (1983), Ptolemy, Copernicus, and Kepler. In Adler M. G. & van
Doren J., Еditors (1983), Great Ideas Today. Chicago, pp. 137 – 180.
--- (2002), The trouble with Ptolemy. Isis, vol. 93, pp. 70 – 74.
Gini C. (1946), Gedanken von Theorem von Bernoulli. Z. für Volkswirtschaft u.
Statistik, 82. Jg., pp. 401 – 413.
Gnedenko B. V. (1951, Russian), On Ostrogradsky’s work in the theory of
probability. IMI, vol. 4, pp. 99 – 123. S, G, 5.
--- (1954, Russian). Theory of probability. Moscow, 1973. [Providence, RI, 2005.]
First Russian ed., 1950.
--- (1958), Main stages in the history of the theory of probability. Actes VIIIe
Congrès Hist. Sci. 1956. N. p., 1958, vol. 1, pp. 128 – 131.
--- (1959, Russian), On Liapunov’s work on the theory of probability. IMI, vol. 12,
pp. 135 – 160.
Gnedenko B. V., Gikhman I. I. (1956, Russian), Development of the theory of
probability in the Ukraine. IMI, vol. 9, pp. 477 – 536.
Gnedenko B. V., Sheynin O. B. (1978). See Sheynin (1978a).
Goldstein B. R. (1985), The ‘Astronomy’ of Levi ben Gerson (1288 – 1344). New
York.
Golitzin Iv. (1807), Statisticheskie Tablitzy Rossiyskoi Imperii (Stat. Tables of the
Russian Empire). Moscow.
Gordevsky D. Z. (1955, Russian), K. A. Andreev. Kharkov.
Gower B. (1993), Boscovich on probabilistic reasoning and the combination of
observations. In Bursill-Hall (1993, pp. 263 – 279).
Gowing R. (1983), Roger Cotes – Natural Philosopher. Cambridge.
Graunt J. (1662), Natural and Political Observations Made upon the Bills of
Mortality. Baltimore, 1939. Editor, W. F. Willcox.
Great Books (1952), Great Books of the Western World, vols 1 – 54. Chicago.
Greenwood M. (1936), Louis and the numerical method. In author’s Medical
Dictator. London, pp. 123 – 142.
--- (1940), A statistical mare’s nest? J. Roy. Stat. Soc., vol. 103, pp. 246 – 248.
--- (1941 – 1943), Medical statistics from Graunt to Farr. Biometrika. Reprint:
Pearson & Kendall (1970, pp. 47 – 120).
Gridgeman N. T. (1960), Geometric probability and the number π. Scripta Math., t.
25, pp. 183 – 195.
Grodzensky S. Ya. (1987, Russian), A. A. Markov. Moscow.
Gusak A. A. (1961, Russian), La préhistoire et les débuts de la théorie de la
représentation approximative des fonctions. IMI, vol. 14, pp. 289 – 348.
Guy W. A. (1852), Statistics, medical. In Cyclopaedia of Anatomy and Physiology,
vol. 4. London, pp. 801 – 814.
Hagstroem K.-G. (1940), Stochastik, ein neues – und doch ein altes Wort. Skand.
Aktuarietidskr., t. 23, pp. 54 – 57.
Hailperin (Halperin ?) T. (1976), Boole’s Logic and Probability. Amsterdam, 1986.
Hald A. (1952), Statistical Theory with Engineering Applications. New York, 1960.

277
275
--- (1990), History of Probability and Statistics and Their Applications before 1750.
New York.
--- (1998), History of Mathematical Statistics from 1750 to 1930. New York.
--- (2002), On the History of Series Expansions of Frequency Functions and
Sampling Distributions, 1873 – 1944. Copenhagen, this being Roy. Danish Acad.
Sci. Letters, Mat.-fys. Meddelelser No. 49.
--- (2003), The History of the Law of Large Numbers and Consistency. Univ.
Copenhagen, Dept. applied math. & statistics, Preprint No. 2.
--- (2004), History of Parametric Statistical Inference from Bernoulli to Fisher,
1713 to 1935. Copenhagen.
Halley E. (1694a), An Estimate of the Degree of Mortality of Mankind. Baltimore,
1942.
--- (1694b), Some further considerations on the Breslaw bills of mortality. Phil. PT
Soc., vol. 17 for 1693, No. 198, pp. 654 – 656. Reprint of volume: New York, 1963.
Halperin T. (1988), The development of probability logic from Leibniz to MacColl.
Hist. and Phil. of Logic, vol. 9, pp. 131 – 191.
Hartley D. (1749), Observations on Man. London. Reprint of edition of 1791:
Poole, 1998.
Harvey W. (1651, Latin), Anatomical Exercises on the Generation of Animals. In
Great Books (1952), vol. 28, pp. 329 – 496.
Hasofer A. M. (1967), Random mechanisms in Talmudic literature. Biometrika, vol.
54, pp. 316 – 321. Enlarged version (1977): Some aspects of Talmudic probabilistic
thought. In Proc. Assoc. Orthodox Jewish Scientists, vol. 2. Jerusalem, pp. 63 – 80.
Haushofer D. M. (1872), Lehr- und Handbuch der Statistik. Wien.
Hegel G. W. F. (1812), Wissenschaft der Logik, Tl. 1. Hamburg, 1978.
Hellman C. D. (1970), Brahe. Dict. Scient. Biogr., vol. 2, pp. 401 – 416.
Helmert F. R. (1868), Studien über rationelle Vermessungen im Gebiete der
höheren Geodäsie. Z. Math. Phys., Bd. 13, pp. 73 – 120, 163 – 186.
--- (1872), Ausgleichungsrechnung nach der Methode der kleinsten Quadrate.
Leipzig. Subsequent eds: 1907 and 1924.
--- (1875a), Über die Berechnung des wahrscheinlichen Fehlers aus einer endlichen
Anzahl wahrer Beobachtungsfehler. Z. Math. Phys., Bd. 20, pp. 300 – 303.
--- (1875b), Über die Formeln für den Durchschnittsfehler. Astron. Nachr., Bd. 85,
pp. 353 – 366.
--- (1876a), Genauigkeit der Formel von Peters zur Berechnung des
wahrscheinlichen Beobachtungsfehlers directer Beobachtungen. Ibidem, Bd. 88, pp.
113 – 132.
--- (1876b), Über die Wahrscheinlichkeit der Potenzsummen der
Beobachtungsfehler. Z. Math. Phys., Bd. 21, pp. 192 – 218.
--- (1886), Lotabweichungen, Heft 1. Berlin.
--- (1905), Über die Genauigkeit der Kriterien des Zufalls bei Beobachtungsreihen.
Sitz. Ber. Kgl. Preuss. Akad. Wiss., Phys.-Math. Cl., Halbbd. 1, pp. 594 – 612. Also
in Helmert (1993, pp. 189 – 208).
--- (1993), Akademie-Vorträge. Frankfurt am Main. Reprints of author’s reports.
Hendriks F. (1852 – 1853), Contributions to the history of insurance. Assurance
Mag., vol. 2, pp. 121 – 150, 222 – 258; vol. 3, pp. 93 – 120.
--- (1863), Notes on the early history of tontines. J. Inst. Actuaries, vol. 10, pp. 205
– 219.
Henny J. (1975), Niklaus und Johann Bernoullis Forschungen auf dem Gebiet der
Wahrscheinlichkeitsrechnung in ihrem Briefwechsel mit Montmort. In J. Bernoulli
(1975, pp. 457 – 507).
Henry M. Ch. (1883), Correspondance inédite de Condorcet et de Turgot. Genève,
1970.
Herschel J. (1850), [Review of] Quetelet (1846). Edinb. Rev., or, Critical J., vol.
92, No. 185, pp. 1 – 57. Publ. anonymously.
Herschel W. (1783), On the proper motion of the Sun. In Herschel (1912, vol. 1, pp.
108 – 130).
--- (1784), Account of some observations. Ibidem, pp. 157 – 166.
--- (1805), On the direction and motion of the Sun. Ibidem, vol. 2, pp. 317 – 331.
--- (1806), On the quantity and velocity of the solar motion. Ibidem, pp. 338 – 359.
--- (1817), Astronomical observations and experiments tending to investigate the
local arrangement of celestial bodies in space. Ibidem, pp. 575 – 591.

278
276
--- (1912), Scientific papers, vols. 1 – 2. London. [London, 2003.]
Hertz H. (1894), Die Principien der Mechanik, this being the author’s Ges. Werke,
Bd. 3. Leipzig.
Heyde C. C., Seneta E. (1977), Bienaymé. New York.
---, Editors (2001), Statisticians of the Centuries. New York.
Hill D., Elkin W. L. (1884), Heliometer-determination of stellar parallax. Mem.
Roy. Astron. Soc., vol. 48, pt. 1 (the whole issue).
Hippocrates (1952a), On the epidemics. In Great Books (1952, vol. 10, pp. 44 –
63).
--- (1952b), On fractures. Ibidem, pp. 74 – 91.
--- (1952c), Aphorisms. Ibidem, pp. 131 – 144.
Hobbes T. (1646), Of liberty and necessity. English Works, vol. 4. London, 1840,
pp. 229 – 278.
Hochkirchen C. (1999), Die Axiomatisierung der Wahrscheinlichkeitsrechnung.
Göttingen.
Hogan E. R. (1977), R. Adrain: American mathematician. Hist. Math., vol. 4, pp.
157 – 172.
Hoskin M. A. (1959), William Herschel. New York.
Hostinský B. (1932), Application du calcul des probabilités à la théorie du
mouvement Brownien. Annales Inst. H. Poincaré, t. 3, pp. 1 – 72.
Hoyrup J. (1983), Sixth-century intuitive probability: the statistical significance of
a miracle. Hist. Math., vol. 10, pp. 80 – 84.
Humboldt A. (1816), Sur les lois que l’on observe dans la distribution des formes
végétales. Annales Chim. Phys., t. 1, pp. 225 – 239.
--- (1817), Des lignes isothermes. Mém. Phys. Chim. Soc. d’Arcueil, t. 3, pp. 462 –
602.
--- (1818), De l’influence de la déclinaison du Soleil sur le commencement des
pluies équatoriales. Annales Chim. Phys., t. 8, pp. 179 – 190.
--- (1831), Fragmens de géologie et de climatologie asiatiques, t. 2. Paris.
--- (1845 – 1862), Kosmos, Bde. 1 – 5 (1845, 1847, 1850, 1858, 1862). Stuttgart.
English transl. of vol. 4: New York, 1858.
Humboldt A., Bonpland A. J. A. (1815 – 1825), Nova genera et species
plantarum, tt. 1 – 7. Paris.
Huygens C. (1657), De calcul dans les jeux de hasard. In Huygens (1888 – 1950, t.
14, pp. 49 – 91). Calculus of gambling, 1998.
--- (1888 – 1950), Oeuvres complètes, tt. 1 – 22. La Haye. Volumes 1, 6, 10 and 14
appeared in 1888, 1895, 1905 and 1920 respectively. Huygens’ correspondence is in
vols. 1 and 6, S, G, 85.
Idelson N. I. (1947), Sposob Naimenshikh Kvadratov etc (Method of Least Squares
etc). Moscow.
Irwin J. O. (1978), Gosset W. S. In Kruskal & Tanur (1978, vol. 1, pp. 409 – 413).
Ivory J. (1825), On the method of least squares. London, Edinburgh and Dublin
Phil. Mag., vol. 65, pp. 1 – 10, 81 – 88, 161 – 168.
--- (1826a), On the ellipticity of the Earth as deduced from experiments with the
pendulum. Ibidem, vol. 68, pp. 3 – 10, 92 – 101.
--- (1826b), On the methods proper to be used for deducing a general formula for the
length of the seconds pendulum. Ibidem, pp. 241 – 245.
--- (1828), Letter to the Editor relating to the ellipticity of the Earth as deduced from
experiments with the pendulum. Ibidem, New ser., vol. 3, pp. 241 – 243.
--- (1830), On the figure of the Earth. Ibidem, New ser., vol. 7, pp. 241 – 243.
Jennings J. (1985), The Journeyer. London. This is a translation of Marco Polo’s
book of 1298.
Johnson N. L., Kotz S., Editors (1997), Leading Personalities in Statistical
Sciences. New York. Collection of biographies partly reprinted from Kotz &
Johnson (1982 – 1989), incorporated in Kotz & Johnson (1982 – 1989/2006).
Johansenn W. (1922), Biology and statistics. Nordic Stat. J., vol. 1, 1929, pp. 351 –
361.
Jorland G. (1987), The Saint Petersburg paradox 1713 – 1937. In The Probabilistic
Revolution, vols 1 – 2. Cambridge (Mass.), 1987, vol. 1, pp. 157– 190. Editors, L.
Kruger, G. Gigerenzer, M. S. Morgan.
Junkersfeld J. (1945), The Aristotelian – Thomistic Concept of Chance. Notre
Dame, Indiana.

279
277
Juskevic [Youshkevitch] A. P., Winter E., Hoffmann P., Editors (1959), Die
Berliner und die Petersburger Akademie der Wissenschaften in Briefwechsels
Eulers, Bd. 1. Berlin.
Kac M. (1939), On a characterization of the normal distribution. In Kac,
Probability, Number Theory and Statistical Physics. Cambridge (Mass.), 1979, pp.
77 – 79.
Kamke E. (1933), Über neuere Begründungen der Wahrscheinlichkeitsrechnung.
Jahresber. Deutschen Mathematiker-Vereinigung, Bd. 42, pp. 14 – 27.
Kant I. (1755), Allgemeine Naturgeschichte und Theorie des Himmels. Ges.
Schriften, Bd. 1. Berlin, 1910, pp. 215 – 368.
--- (1763), Der einzig mögliche Beweisgrund zu einer Demonstration des Daseins
Gottes. Ibidem, Bd. 2, 1912, pp. 63 – 163.
Kapteyn J. C. (1906a), Plan of Selected Areas. Groningen.
--- (1906b), Statistical methods in stellar astronomy. [Reports] Intern. Congr. Arts &
Sci. St. Louis 1904. N. p., vol. 4, 1906, pp. 396 – 425.
--- (1909), Recent researches in the structure of the universe. Annual Rept
Smithsonian Instn for 1908, pp. 301 – 319.
--- (1912), Definition of the correlation-coefficient. Monthly Notices Roy. Astron.
Soc., vol. 72, pp. 518 – 525.
Karn M. Noel (1931), An inquiry into various death-rates and the comparative
influence of certain diseases on the duration of life. Annals of Eugenics, vol. 4, pp.
279 – 326.
Kaufman A. A. (1922), Teoria i Metody Statistiki (Theory and Methods of
Statistics). Moscow. Fourth ed. Fifth, posthumous edition, Moscow, 1928. German
edition: Theorie und Methoden der Statistik. Tübingen, 1913.
Kendall D. G. (1975), The genealogy of genealogy: branching processes before
(and after) 1873. Bull. London Math. Soc., vol. 7, pp. 225 – 253.
Kendall M. G. (Sir Maurice) (1951), George Udny Yule, 1871 – 1951. J. Roy. Stat.
Soc., vol. 115A, pp. 156 – 161.
--- (1956), The beginnings of a probability calculus. Biometrika, vol. 43, pp. 1 – 14.
Reprint: E. S. Pearson & Kendall (1970, pp. 19 – 34).
--- (1960), Where shall the history of statistics begin? Biometrika, vol. 47, pp. 447 –
449. Reprint: Ibidem, pp. 45 – 46.
--- (1968), F. Y. Edgeworth. Biometrika, vol. 55, pp. 269 – 275. Reprint: Ibidem, pp.
257 – 263.
--- (1971), The work of Ernst Abbe. Biometrika, vol. 58, pp. 369 – 373. Reprint: M.
G. Kendall & Plackett (1977, pp. 331 – 335).
Kendall M. G., Doig A. G. (1962 – 1968), Bibliography of Statistical Literature,
vols. 1 – 3. Vol. 1, 1962, covers the period 1950 – 1958, vol. 2, 1965, represents
period 1940 – 1949, and vol. 3, 1968, lists literature up to 1940 with supplements to
vols. 1 and 2.
Kendall M. G., Moran P. A. P. (1963), Geometrical Probabilities. London.
Kendall M. G., Plackett R. L., Editors (1977), Studies in the History of Statistics
and Probability, vol. 2. London. Collected reprints of papers.
Kepler J. (1596). Mysterium cosmographicum. Ges. Werke, Bd. 8. Munich, 1963,
pp. 7 – 128, this being a reprint of second edition (1621, with additions to many
chapters). English transl: New York, 1981.
--- (1604, German), Thorough description of an extraordinary star. Vistas in
Astronomy, vol. 20, 1977, pp. 333 – 339.
--- (1606, Latin), Über den neuen Stern im Fuß des Schlangenträger. Würzburg,
2006.
--- (1609, Latin), New Astronomy. Cambridge, 1992, 2015. Double paging.
--- (1610), Tertius interveniens. Das ist Warnung an etliche Theologos, Medicos und
Philosophos. Ges. Werke, Bd. 4. München, 1941, pp. 149 – 258.
--- (1618 – 1621, Latin), Epitome of Copernican Astronomy, Books 4 and 5, 1620 –
1621. Great Books (1952, vol. 16, pp. 845 – 1004).
--- (1619, Latin), Harmony of the World. Philadelphia, 1997.
Keynes J. M. (1921), Treatise on Probability. Coll. Writings, vol. 8. London, 1973.
Khinchin A. Ya. (1943, Russian), Mathematical Foundations of Statistical
Mechanics. New York, 1949.
--- (1961, Russian), The Mises frequency theory and modern ideas of the theory of
probability. Science in Context, vol. 17, 2004, pp. 391 – 422.

280
278
Kington J. A. (1974), The Societas meteorologica Palatina. Weather, vol. 29, No.
11, pp. 416 – 426.
Knapp G. F. (1872), Quetelet als Theoretiker. JNÖS, Bd. 18, pp. 89 – 124.
Knauer K. (1955), Grundfragen einer mathematischen Stilistik. Forschungen u.
Fortschritte, Bd. 29, pp. 140 – 149.
Knies C. G. A. (1850), Die Statistik als selbstständige Wissenschaft. Kassel.
[Frankfurt, 1969.]
Knott C. G. (1911), Life and Work of P. G. Tait. Cambridge.
Kohli K. (1975a), Spieldauer: von J. Bernoullis Lösung der fünfte Aufgabe von
Huygens bis zu den Arbeiten von de Moivre. In J. Bernoulli (1975, pp. 403 – 455).
--- (1975b), Aus dem Briefwechsel zwischen Leibniz und Jakob Bernoulli. Ibidem,
pp. 509 – 513.
--- (1975c), Kommentar zur Dissertation von N. Bernoulli. Ibidem, pp. 541 – 556.
Kohli K., van der Waerden B. L. (1975), Bewertung von Leibrenten. In J.
Bernoulli (1975, pp. 515 – 539).
Koialovitch B. M. (1893 – 1909, Russian), Four letters to A. A. Markov. Archive,
Russian Acad. Sci., Fond 173, Inventory 1, 10. S, G, 85.
Kolmogorov A. N. (1931, Russian), The method of median in the theory of errors.
In Kolmogorov (1992, pp. 115 – 117).
--- (1946, Russian), Justification of the method of least squares. Sel. Works, vol. 2,
pp. 285 – 302.
--- (1947, Russian), The role of Russian science in the development of the theory of
probability. Uch. Zapiski Moskovsk. Gos. Univ. No. 91, pp. 53 – 64. S, G, 7.
--- (1948, Russian), Slutsky. Math. Scientist, vol. 27, 2002, pp. 67 – 74.
--- (1954, Russian), Law of small numbers. Bolshaia Sov. Enz. (Great Sov. Enc.), 2nd
edition, vol. 26, p. 169. Published anonymously.
--- (1983), On the logical foundations of probability theory. Sel. Works, vol. 2, pp.
515 – 519. (See next item.)
--- (1985 – 1986, Russian), Selected Works, vols. 1 – 2. Dordrecht, 1991 – 1992.
Kolmogorov A. N., Petrov A. A., Smirnov Yu. M. (1947, Russian), A formula of
Gauss in the method of least squares. Izvestia Akad. Nauk SSSR, ser. math., vol. 11,
pp. 561 – 566. English transl. in Kolmogorov (1992, vol. 2, pp. 303 – 308).
Kolmogorov A. N., Prokhorov Yu. V. (1988, Russian), Mathematical statistics.
Enc. Math., vol. 6, 1990, pp. 138 – 142.
Koopman B. O. (1940), The bases of probability. Bull. Amer. Math. Soc., vol. 46,
pp. 763 – 774.
Köppen W. (1872), Die Aufeinanderfolge der unperiodischen
Witterungserscheinungen. Repert. Met., Bd. 2. Petersburg, pp. 187 – 238.
--- (1874), Über die Abhängigkeit des klimatischen Characters der Winde von ihrem
Ursprunge. Ibidem, Bd. 4, No. 4, pp. 1 – 15.
--- (1875, in Russian), On the observation of periodic phenomena in nature. Zapiski
Russk. Geografich. Obshchestvo po Obshchei Geografii, vol. 6, No. 1, pp. 255 –
276.
Körber H.-G. (1959), Über Humboldts Arbeiten zur Meteorologie und
Klimatologie. Humboldt Gedenkschrift. Berlin, pp. 289 – 335.
Kotz S., Johnson N. L., Editors (1982 – 1989), Enc. of Statistical Sciences, second
edition, vols. 1 – 16 with single paging. Hobokan, New Jersey, 2006.
Kotz S., Read C. D., Banks D. L., Editors (1997 – 1999), Update Volume, vol. 1 –
3 to Kotz & Johnson (1982 – 1989). New York.
Krein M. G. (1951, Russian), Chebyshev’s and Markov’s ideas in the theory of the
limiting values of integrals and their further development. UMN, vol. 6, No. 4, pp. 3
– 120.
Krengel, U. (2011), On the contributions of Georg Bohlmann. J. Electronique
d’hist. du probabilités et de la statistique.
Kritzenwecky J., Editor (1965), G. J. Mendel. Leipzig.
Kronecker L. (1894), Vorlesungen über Mathematik, Bd. 1. Leipzig.
--- (1901), Vorlesungen über Zahlentheorie, Bd. 1. Leipzig.
Kruskal W. (1946), Helmert’s distribution. Amer. Math. Monthly, vol. 53, pp. 435 –
438.
--- (1978), Formulas, numbers, words: statistics in prose. In: New Directions for
Methodology of Social and Behavioral Science. San Francisco, 1981. Editor, D.
Fiske, pp. 93 – 102.

281
279
--- (1988), Miracles and statistics: the casual assumption of independence. J. Amer.
Stat. Assoc., vol. 83, pp. 929 – 940.
Kruskal W., Tanur J. M., Editors (1978), International Encyclopedia of Statistics,
vols. 1 – 2. New York.
Krylov A. N. (1950), Lektsii o Priblizhennykh Vychisleniyakh (Lectures on
Approximate Calculations). Moscow.
Ku H. H., Editor (1969), Precision Measurement and Calibration. Nat. Bureau
Standards Sp. Publ. 300, vol. 1. Washington.
Kuzmin R. (1928, Russian), Sur un problème de Gauss. Atti Congr. Intern. Matem.
Bologna 1928. Bologna, 1932, t. 6, pp. 83 – 89.
Lacombe H., Costabel P., Editors (1988), La figure de la Terre du XVIIIe siècle à
l’ère spatiale. Paris.
Lacroix S.-F. (1816), Traité élémentaire du calcul des probabilités. Paris.
Subsequent editions: 1828, 1833, 1864; German transl., 1818.
Lagrange J. L. (1776a), Sur l’utilité de la méthode de prendre le milieu entre les
résultats de plusieurs observations. Oeuvr., t. 2. Paris, 1868, pp. 173 – 236.
--- (1776b), Letter to Laplace of 30 Dec. Oeuvr., t. 14. Paris, 1892, p. 66.
--- (1777), Recherches sur les suites récurrentes. Oeuvr., t. 4. Paris, 1869, pp. 151 –
251.
Lamarck J. B. (1800 – 1811), Annuaire météorologique [, tt. 1 – 11]. Paris.
Extremely rare.
--- (1809), Philosophie zoologique, t. 2. Paris, 1873.
--- (1815), Histoire naturelle des animaux sans vertèbres, t. 1. Paris.
--- (1817), Espèce. Nouv. Dict. Hist. Natur., t. 10, pp. 441 – 451.
Lambert J. H. (1760), Photometria. Augsburg. Incomplete transl. into German:
Ostwald Klassiker, NNo. 31 – 33, 1892.
--- (1765a), Anmerkungen und Zusätze zur practischen Geometrie. In Lambert,
Beyträge zum Gebrauche der Mathematik und deren Anwendung, Tl. 1. Berlin,
1765, pp. 1 – 313.
--- (1765b), Theorie der Zuverlässigkeit der Beobachtungen und Versuche. Ibidem,
pp. 424 – 488.
--- (1771), Anlage zur Architectonic, Bd. 1. Also Phil. Schriften, Bd. 3. Hildesheim,
1965.
--- (1772), Anmerkungen über die Sterblichkeit, Todtenlisten, Geburthen und Ehen.
Beyträge, Tl. 3. Berlin, 1772, pp. 476 – 569.
--- (1772 – 1775), Essai de taxéométrie ou sur la mesure de l’ordre. Nouv. Mém.
Acad. Roy. Sci. et Belles-Lettres Berlin for 1770, pp. 327 – 342; for 1773, pp. 347 –
368. Also Phil. Schriften, Bd. 8/1. Hildesheim, 2007, pp. 423 – 460.
--- (1773), Observations sur l’influence de la Lune dans le poids de l’atmosphère.
Nouv. Mém. Acad. Roy. Sci. Berlin for 1771, pp. 66 – 73.
Lamont J. (1867), Über die Bedeutung arithmetischer Mittelwerthe in der
Meteorologie. Z. Öster. Ges. Met., Bd. 2, No. 11, pp. 241 – 247.
Langevin P. (1913), La physique du discontinu. In collected articles Les progrès de
la physique moléculaire. Paris, 1914, pp. 1 – 46.
Laplace P. S. (1774), Sur la probabilité des causes par les événements. OC, t. 8.
Paris, 1891, pp. 27 – 65.
--- (1776), Recherches sur l’intégration des équations différentielles aux différences
finies. Ibidem, pp. 69 – 197.
--- (1781), Sur les probabilités. OC, t. 9. Paris, 1893, pp. 383 – 485.
--- (1786), Sur les approximations des formules qui sont fonctions de très-grands
nombres, Suite. OC, t. 10. Paris, 1894, pp. 295 – 338.
--- (1789), Sur quelques points du système du monde. OC, t. 11. Paris, 1895, pp. 477
– 558.
--- (1796), Exposition du système du monde. OC, t. 6. Paris, 1884 (the whole
volume, this being a reprint of the edition of 1835).
--- (1798 – 1825), Traité de mécanique céleste. OC, tt. 1 – 5. Paris, 1878 – 1882.
English transl. of vols. 1 – 4 by N. Bowditch: Celestial Mechanics (1829 – 1839).
New York, 1966.
--- (1810a), Sur les approximations des formules qui sont fonctions de très grands
nombres et sur leur application aux probabilités. OC, t. 12. Paris, 1898, pp. 301 –
345.
--- (1810b), Same title, supplement. Ibidem, pp. 349 – 353.

282
280
--- (1811), Sur les intégrales définies. Ibidem, pp. 357 – 412.
--- (1812), Théorie analytique des probabilités. OC, t. 7, No. 1 – 2. Paris, 1886.
Consists of two parts, an Introduction (1814) and supplements, see below. Theory of
probability proper is treated in pt. 2. Sections 21 – 24 of Chapter 4: S, G, 58.
--- (1814), Essai philosophique sur les probabilités. In Laplace (1812/1886, No. 1,
separate paging). English transl: Philosophical Essay on Probabilities. New York,
1995. Translator and editor A. I. Dale.
--- (1816), Théor. anal. prob., Supplément 1. OC, t. 7, No. 2, pp. 497 – 530.
--- (1818), Théor. anal. prob., Supplément 2. Ibidem, pp. 531 – 580.
--- (ca. 1819), Théor. anal. prob., Supplément 3. Ibidem, pp. 581 – 616.
--- (1827), Sur le flux et reflux lunaire atmosphérique. OC, t. 13. Paris, 1904, pp.
342 – 358.
La Placette J. (1714), Traité des jeux de hasard. La Haye.
Laurent H. (1873), Traité du calcul des probabilités. Paris.
Le Cam L. (1986), The central limit theorem around 1935. Stat. Sci., vol. 1, pp. 78
– 96.
Legendre A. M. (1805), Nouvelles méthodes pour la détermination des orbites des
comètes. Paris.
--- (1814), Méthode des moindres carrés. Mém. Acad. Sci. Paris, Cl. Sci., Math. et
Phys. t. 11, pt. 2, année 1810, pp. 149 – 154.
--- (1820), Nouvelles méthodes etc. Suppl. 2.
Lehmann E. L. (1959), Testing Statistical Hypotheses. [New York, 1997.]
Lehmann-Filhès R. (1887), Über abnorme Fehlervertheilung und Verwerfung
zweifelhafter Beobachtungen. Astron. Nachr., Bd. 117, pp. 121 – 132.
Leibniz, G. W. (1668 – 1669 (?), manuscript), Demonstrationum Catholicarum
conspectus. Sämmtl. Schriften und Briefe, 6. Reihe, Bd. 1. Berlin, 1971, pp. 494 –
500.
--- (1680?, manuscript, publ. 1872), Öffentliche Assecuranzen. In author’s book
(1986, pp. 421 – 432).
--- (1680 – 1683, manuscript, published 1866), Essay de quelques raisonnemens
nouveau sur la vie humaine. In Leibniz (2000, pp. 428 – 445 with German
translation). S, G, 58.
--- (1683, manuscript), Meditatio iuridico-mathematica de interusurio simplice. With
a German translation in Leibniz (2000, pp. 273 – 293).
--- (1686, manuscript), Allgemeine Untersuchungen über die Analyse der Begriffe
und wahren Sätze. In author’s book Fragmente zur Logik. Berlin, 1960, pp. 241 –
303.
--- (1704), Neue Abhandlungen über menschlichen Verstand. Hamburg, 1996.
--- (1970), Sämmtl. Schriften und Briefe, 8. Reihe, Bd. 1. Berlin.
--- (1971), Math. Schriften, 1. Reihe, Bd. 3. Hildesheim.
--- (1986), Sämmtl. Schriften und Briefe, Reihe 4, Bd. 3. Berlin.
--- (2000), Hauptschriften zur Versicherungs- und Finanzmathematik. Editors, E.
Knobloch et al. Berlin.
Lenin V. I. (1909), Materialism i Empiriokritizism. Polnoe Sobranie Sochineniy
(Complete Works), 5th edition, entire vol. 18, 1961.
Le procès (1900), Le procès Dreyfus devant le Conceil de guerre de Rennes, tt. 1 –
3. Paris.
Levi ben Gerson (14th century, publ. 1999), The Wars of the Lord, vol. 3. New
York.
Lévy M. (1844), Traité d’hygiène. Paris, 1862.
Lévy P. (1925), Calcul des probabilités. Paris.
Lexis W. (1874, report), Naturwissenschaft und Sozialwissenschaft. In Lexis (1903,
pp. 233 – 251).
--- (1876), Das Geschlechtsverhältnis der Geborenen und die
Wahrscheinlichkeitsrechnung. JNÖS, Bd. 27, pp. 209 – 245. Reprinted in Lexis
(1903, pp. 130 – 169).
--- (1877), Zur Theorie der Massenerscheinungen in der menschlichen Gesellschaft.
Freiburg i/Bayern.
--- (1879), Über die Theorie der Stabilität statistischer Reihen. JNÖS, Bd. 32, pp. 60
– 98. Reprinted in Lexis (1903, pp. 170 – 212).
--- (1886), Über die Wahrscheinlichkeitsrechnung und deren Anwendung auf die
Statistik. JNÖS, Bd. 13 (47), pp. 433 – 450.

283
281
--- (1903), Abhandlungen zur Theorie der Bevölkerungs- und Moralstatistik. Jena.
--- (1913), Review of A. A. Kaufmann (1913). Schmollers Jahrbuch f.
Gesetzgebung, Verwaltung u. Volkswirtschaft in Deutschen Reiche, Bd. 37, pp. 2089
– 2092.
Liapunov A. M. (1895, Russian), Chebyshev. In Chebyshev, P. L. Izbrannye
Matematich. Trudy (Sel. Math. Works). Moscow – Leningrad, 1946, pp. 9 – 21. S,
G, 36.
--- (1900), Sur une proposition de la théorie des probabilités. Izvestia Imp. Acad. Sci.
St. Pétersb., sér. 5, t. 13, pp. 359 – 386.
--- (1901a), Nouvelle forme du théorème sur la limite de probabilité. Mém. Imp.
Acad. Sci. St. Pétersb., sér. 8, Cl. phys.-math., t. 12, No. 5, separate paging. Also in
Liapunov (1954, pp. 157 – 178).
--- (1901b, Russian), An answer to Nekrasov. Zapiski Kharkov Univ., vol. 3, pp. 51
– 63.
--- (1954), Sobranie Sochineniy (Coll. Works), vol. 1. Moscow.
Libri – Carrucci G. B. I. T., Lacroix S. F., Poisson S. D. (1834), Report on
Bienaymé´s manuscript [Bienaymé (1838)]. Procès verbaux des séances, Acad. Sci.
Paris, t. 10, pp. 533 – 535.
Liebermeister C. (ca. 1877), Über Wahrscheinlichkeitsrechnung in Anwendung auf
therapeutische Statistik. Sammlung klinischer Vorträge No. 110 (Innere Medizin No.
39). Leipzig, pp. 935 – 961.
Lindeberg J. W. (1922a), Über das Exponentialgesetz in der
Wahrscheinlichkeitsrechnung. Annales Acad. Scient. Fennicae, t. A16 für 1920, No.
1, pp. 1 – 23.
--- (1922b), Eine neue Herleitung des Exponentialgesetzes in der
Wahrscheinlichkeitsrechnung. Math. Z., Bd. 15, pp. 211 – 225.
Linnik Yu. V. (1951, Russian), Commentary on Markov (1916a). In Markov (1951,
pp. 668 – 670).
--- (1952, Russian), Remarks concerning the classical derivation of the Maxwell
law. Doklady Akad. Nauk SSSR, vol. 85, pp. 1251 – 1254.
--- (1958, Russian), Method of Least Squares etc. Oxford, 1968. Reference in text to
English edition of 1961.
Linnik Yu. V., Sapogov N. A., Timofeev V. N. (1951, Russian), Essay on the
work of Markov in number theory and theory of probability. In Markov (1951, pp.
614 – 640). S, G, 5.
Lipschitz R. (1890), Sur la combinaison des observations. C. r. Acad. Sci. Paris, t.
111, pp. 163 – 166.
Lloyd G. E. R. (1982), Observational error in later Greek science. In Science and
Speculation. Editors, J. Barnes et al. Cambridge, pp. 128 – 164.
Louis P. C. A. (1825), Recherches anatomico-pathologiques sur la phtisie. Paris.
Loveland R. (2001), Buffon, the certainty of sunrise, and the probabilistic reductio
ad absurdum. AHES, vol. 55, pp. 465 – 477.
Lubbock W., Drinkwater Bethune J. E. (1830), Treatise on Probability. Appended
to Jones D. (1844), On the Values of Annuities, vol. 2. London.
Lueder A. F. (1812), Kritik der Statistik und Politik. Göttingen.
Mach E. (1897), Die Mechanik in ihrer Entwicklung. Leipzig. Third edition, 1911.
Maciejewski C. (1911), Nouveaux fondements de la théorie de la statistique. Paris.
Maennchen Ph. (1918), Gauss als Zahlenrechner. Gauss W-10/2, Abt. 6, 1930.
Separate paging.
Mahalanobis P. C. (1936), Note on the statistical and biometric writings of K.
Pearson. Sankhya, vol. 2, pp. 411 – 422.
Maimonides M. (12th century; 1975), Introduction to the Talmud. New York.
--- (12th century; 1977), Letters. New York.
Maire [C.], Boscovich [R. J.] (1770), Voyage astronomique et géographique dans
l’État de l’Église. Paris.
Maistrov L. E. (1967, Russian), Probability Theory. Historical Sketch. New York –
London, 1974.
Makovelsky [O.] (1914), Dosokratiki (Presocratics), pt. 1. Kazan.
Malthus T. R. (1798), Essay on the Principle of Population. Works, vol. 1. London,
1986.
Maltsev A. I. (1947, Russian), Remark on the work of Kolmogorov et al (1947).
Izvestia Akad. Nauk SSSR, Ser. math., vol. 11, pp. 567 – 578.

284
282
Mandryka A. P. (1984, Russian), Nikolai Vladimirovich Mayevski. Moscow.
Mansion P. (1904), Sur la portée objective du calcul des probabilités. Mathesis, sér.
3, t. 4. Suppl. 2. Separate paging.
Markov A. A. [Jr.] (1951 Russian), Biography of A. A. Markov [Sr.]. In Markov
(1951, pp. 599 – 613). S, G, 5.
Markov A. A. [Sr] (1884, Russian), Proof of some of Chebyshev’s inequalities. In
author’s book Izbrannye Trudy po Teorii Nepreryvnykh Drobei (Sel. Works on
Theory of Continued Fractions). Moscow – Leningrad, 1948, pp. 15 – 24.
--- (1888), Table des valeurs de l’intégrale … St. Pétersbourg.
--- (1898, Russian), Sur les racines de l’équation … In Markov (1951, pp. 255 –
269).
--- (1899a, Russian), The law of large numbers and the method of least squares.
Ibidem, pp. 231 – 251.
--- (1899b, Russian), Application of continued fractions to calculation of
probabilities. Izvestia Fiz.-Mat. Obshchestvo Kazan Univ., ser. 2, vol. 9, No. 2, pp.
29 – 34.
--- (1899c, Russian), Answer [to Nekrasov]. Ibidem, No. 3, pp. 41 – 43.
--- (1900a), [Treatise] Ischislenie Veroiatnostei (Calculus of Probabilities).
Subsequent editions: 1908, 1913, and (posthumous) 1924. German transl.: Leipzig –
Berlin, 1912.
--- (1900b, Russian), On probability a posteriori. Soobshcheniya Kharkov
Matematich. Obshchestvo, ser. 2, vol. 7, No. 1, pp. 707 – 714.
--- (1906, Russian), Extension of the law of large numbers to magnitudes dependent
on one another. In Markov (1951, pp. 339 – 361). S, G, 5.
--- (1907, Russian), On some cases of the theorems on the limit of expectation and
on the limit of probability. Izvestia Imp. Akad. Nauk. St. Petersb., ser. 6, t. 1, pp. 707
– 714.
--- (1908a, Russian), On some cases of the theorem on the limit of probability.
Ibidem, t. 2, pp. 483 – 496.
--- (1908b, Russian), Extension of the limit theorems of the calculus of probability
to a sum of magnitudes connected into a chain. In Markov (1951, pp. 363 – 397). S,
G, 5.
--- (1910, Russian), Investigation of the general case of trials connected into a chain.
Ibidem, pp. 465 – 507. S, G, 5.
--- (1911a, Russian), On connected magnitudes not forming a real chain. Ibidem, pp.
399 – 416.
--- (1911b, Russian), On a case of trials connected into a complex chain. Ibidem, pp.
417 – 436.
--- (1911c, Russian), On the basic principles of the calculus of probability and on the
law of large numbers. In Ondar (1977/1981, pp. 149 – 153).
--- (1912a, Russian), On trials connected into a chain by unobserved events. In
Markov (1951, pp. 437 – 463).
--- (1912b, Russian), A rebuke to Nekrasov. MS, vol. 28, pp. 215 – 227.
--- (1913, Russian), Example of a statistical investigation of the text of [Pushkin’s]
Evgeniy Onegin illustrating the connection of trials into a chain. Izvestia Imp. Akad.
Nauk St. Pétersb., ser. 6, t. 7, pp. 153 – 162.
--- (1914a, Russian), On probability a posteriori. Second note. Soobshcheniya
Kharkov Matematich. Obshchestvo, ser. 2, vol. 14, No. 3, pp. 105 – 112.
--- (1914b, in Russian), Bicentennial of the law of large numbers. In Ondar
(1977a/1981, pp. 158 – 163).
--- (1915a, Russian), On Florov’s and Nekrasov’s scheme of teaching the theory of
probability in school. Zhurnal Ministerstva Narodn. Prosv., New ser., vol. 57, May,
pp. 26 – 34 of section on Modern chronicle.
--- (1915b, Russian), On a problem by Laplace. In Markov (1951, pp. 549 – 571).
--- (1916a, Russian), On the coefficient of dispersion. Ibidem, pp. 523 – 535. S, G,
5.
--- (1916b, Russian), On the coefficient of dispersion for small numbers. Strakhovoe
Obozrenie, No. 2, pp. 55 – 59. S, G, 5.
--- (1917, Russian), On some limit formulas of the calculus of probability. In
Markov (1951, pp. 573 – 585).
--- (1951), Izbrannye Trudy (Sel. Works). N. p.

285
283
--- (1990, Russian), On solidity of glass. Manuscript of ca. 1903. Incorporated in
Sheynin (1990b). S, G, 85.
Marsden B. G. (1995), 18th and 19th century developments in the theory and
practice of orbit determination. In Taton & Wilson (1995, pp. 181 – 190).
Materialy (1917), Materialy dlia Biograficheskogo Slovaria Deistvitelnykh Chlenov
Akademii Nauk (Materials for a Biographical Dictionary of the Full Members of the
Acad. Sci.). Petrograd.
Matthiesen L. (1867), Vermischtes aus dem Gebiete der
Wahrscheinlichkeitsrechnung. Arch. Math. Phys., Bd. 47, pp. 457 – 460.
Maupertuis P. L. M. (1738), Relation du voyage fait par ordre du Roi au cercle
polaire etc. Œuvres, t. 3, pp. 68 – 175.
--- (1745), Venus physique. Oeuvres, t. 2, pp. 1 – 133.
--- (1751), Systéme de la nature. Ibidem, pp. 135 – 184.
--- (1756a), Sur le divination. Ibidem, pp. 298 – 306.
--- (1756b), Opérations pour déterminer la figure de la Terre et les variations de la
pesanteur. Oeuvres, t. 4, pp. 285 – 346.
--- (1756c), Ouvres, tt. 1 – 4. Lyon.
Maxwell J. C. (1859), On the stability of the motion of Saturn’s rings. In Maxwell
(1890, vol. 1, pp. 288 – 376).
--- (1860), Illustrations of the dynamical theory of gases. Ibidem, pp. 377 – 410.
--- (1867), On the dynamical theory of gases. Ibidem, vol. 2, pp. 26 – 78.
--- (1871), Introductory lecture on experimental physics. Ibidem, pp. 241 – 255.
--- (1873a), Matter and Motion. London.
--- (1873b), Molecules. In Maxwell (1890, vol. 2, pp. 361 – 378).
--- (1873, report), Does the progress of physical science tend to give any advantage
to the opinion of necessity over that of the contingency of events. In Campbell &
Garnett (1884/1969, pp. 434 – 444).
--- (manuscript 1873), Discourse on molecules. Ibidem, pp. 358 – 361. Possibly an
excerpt.
--- (1875), On the dynamical evidence of the molecular constitution of bodies.
In Maxwell (1890, vol. 2, pp. 418 – 438).
--- (1877), Review of H. W. Watson (1876), Treatise on the Kinetic Theory of
Gases. Oxford. Nature, vol. 16, pp. 242 – 246.
--- (1879), On Boltzmann’s theorem. In Maxwell (1890, vol. 2, pp. 713 – 741).
--- (1890), Scientific Papers, vols. 1 – 2. Cambridge. Reprints: Paris, 1927, New
York, 1965.
--- (1990 – 2002), Scientific Letters and Papers, vols. 1 – 3. Cambridge.
May K. O. (1972). Gauss. Dict. Scient. Biogr., vol. 5, pp. 298 – 315.
Mayer T. (1750), Abhandlung über die Umwälzung des Mondes um seine Axe.
Kosmograph. Nachr. u. Samml. für 1748, pp. 52 – 183.
Mayevsky N. V. (1870), Mémoire sur les experiences faits […] au mois de
Novembre 1867, pout déterminer les pressions des gaz de la poudre dans l’ame des
bouches à feu. Acad. Roy. des Sciences, des Lettres et des Beau-Arts de Belg., t. 21,
pp. 3 – 24.
--- (1872), Traité de balistique extérieure. Paris. Russian edition: 1870. Russian
lithographic edition:1859.
Meadowcroft L. V. (1920), On Laplace’s theorem on simultaneous errors.
Messenger Math., vol. 48, pp. 40 – 48.
Mendel J. G. (1866, German), Experiments in plant hybridization. In Bateson, W.
(1909), Mendel’s Principles of Heredity. Cambridge, 1913, pp. 317 – 361.
--- (1905, German), Letters to C. Naegeli, 1866 – 1873. Genetics, vol. 35, No. 5, pt.
2, pp. 1 – 28.
Mendeleev D. I., all publications in Russian (1856), Specific volumes. In
Mendeleev (1934 – 1952, vol. 1, 1937, pp. 139 – 311).
--- (1860), On the cohesion of some liquids. Ibidem, vol. 5, 1947, pp. 40 – 55.
--- (1872a), On the compressibility of gases. Ibidem, vol. 6, 1939, pp. 128 – 171.
--- (1872b), Report on the experiments of 1867 and 1869. Ibidem, vol. 16, 1951, pp.
99 – 113.
--- (1875a), On the elasticity of gases, pt. 1. Ibidem, vol. 6, pp. 221 – 589.
--- (1875b), Progress of work on the restoration of the prototypes of measures of
length and weight. Ibidem, vol. 22, 1950, pp. 175 – 213.

286
284
--- (1876), On the temperatures of the atmospheric layers. Ibidem, vol. 7, 1946, pp.
241 – 269.
--- (1877), The oil industry of Pennsylvania and in the Caucasus. Ibidem, vol. 10,
1949, pp. 17 – 244.
--- (1885), Note on the scientific work of A. I. Voeikov. Ibidem, vol. 25, 1952, pp.
526 – 531.
--- (1887), Investigation of the specific weight of aqueous solutions. Ibidem, vol. 3,
1934, pp. 3 – 468.
--- (1895), On the weight of a definite volume of water. Ibidem, vol. 22, 1950, pp.
105 – 171.
--- (1934 – 1952), Sochinenia (Works). Moscow – Leningrad.
Mendelsohn M. (1761), Über die Wahrschscheinlichkeit. Phil. Schriften, Tl. 2.
Berlin, pp. 189 – 228.
Merian P. (1830), D. Huber. In Verh. Allg. schweiz. Ges. ges. Naturwiss. in ihrer
16. Jahresversamml. 1830. St. Gallen, pp. 145 – 152.
Merrington M. et al (1983), List of the Papers and Correspondence of Karl
Pearson. London.
Métivier M., Costabel P., Dugac P., Editors (1981), Poisson et la science de son
temps. Paris.
Meusnier N. (2009), Fermat et les prémices d’une mathématisation du hasard.
Annales Fac. Sciences Toulouse, No. 18, pp. 87 – 118.
Meyer Hugo (1891), Anleitung zur Bearbeitung meteorologischer Beobachtungen.
Berlin.
Michell J. (1767), An inquiry into the probable parallax and magnitude of the fixed
stars. PT Roy. Soc. Abridged, vol. 12, 1809, pp. 423 – 438.
Mill J. S. (1843), System of Logic. London, 1886. Many more editions, e. g. Coll.
Works, vol. 8. Toronto, 1974
--- (1964a), Selected Papers, vol. 2. Providence, Rhode Island.
Mises R. von (1919), Fundamentalsätze der Wahrscheinlichkeitsrechnung. Math. Z.,
Bd. 4, pp. 1 – 97. Partly reprinted in Mises (1964a, pp. 35 – 56).
--- (1928), Wahrscheinlichkeit, Statistik und Wahrheit. Wien. Subsequent editions,
for example, Wien, 1972. English transl.: New York, 1981.
--- (1964a), Selected Papers, vol. 2. Providence, Rhode Island.
--- (1964b), Mathematical Theory of Probability and Statistics. New York. Editor,
H. Geiringer.
Molina E. C. (1930), The theory of probability: some comments on Laplace’s
Théorie analytique. Bull. Amer. Math. Soc., vol. 36, pp. 369 – 392.
--- (1936), A Laplacean expansion for Hermitian – Laplace functions of high order.
Bell Syst. Techn. J., vol. 15, pp. 355 – 362.
Mondesir E. (1837), Solution d’une question qui se présente dans le calcul des
probabilités. J. Math. Pures et Apppl., t. 2, pp. 3 – 10.
Montmort P. R. (1708), Essay d’analyse sur les jeux de hazard. Paris, 1713.
Published anonymously. References in text to reprint: New York, 1980.
Montucla J. E. (an X, 1802), Histoire des mathématiques, t. 3. Paris.
Moore P. G. (1978), Runs. In Kruskal & Tanur (1978, vol. 1, pp. 655 – 661).
Morant G. M. et al (1939), Bibliography of the Statistical and Other Writings of
Karl Pearson. London.
Mouat F. J. (1885), History of the Statistical Society of London. Jubilee Volume of
the Stat. Soc. London, pp. 14 – 71.
Muncke G. W. (1837), Meteorologie. Gehler’s Phys. Wörterbuch, Bd. 6/3. Leipzig,
pp. 1817 – 2083.
Needham J. (1962), Science and Civilization in China, vol. 4, pt. 1. Cambridge.
Nekrasov P. A., all publications in Russian
--- [1888], Teoria Veroiatnostei. Lectures of 1887 – 1888. N. p., n. d., lithograph.
Subsequent editions: 1894 (lithograph), 1896 and 1912.
--- (1890), The proposition inverse with respect to the Jakob Bernoulli theorem.
Trudy Obshchestvo Liubitelei Estestvoznania, Antropologii i Etnografii, Otdelenie
fizich. nauk, vol. 3, No. 1, pp. 45 – 47.
--- (1898), General properties of mass independent phenomena in connection with
approximate calculation of functions of very large numbers. MS, vol. 20, pp. 431 –
442.

287
285
--- (1900 – 1902), New principles of the doctrine of probabilities of sums and
means. MS, vols 21 – 23, pp. 579 – 763; 1 – 142, 323 – 498; 41 – 455.
(1901), Concerning a simplest theorem on probabilities of sums and means. MS,
vol. 22, pp. 225 – 238.
--- (1902), Filosofia i Logika Nauki etc. (Philosophy and Logic of Science).
Moscow.
--- (1906), Osnovy Obshchestvennykh i Estestvennykh Nauk v Srednei Shkole.
(Principles of Social and Natural Sciences in High School). Petersburg.
--- (1911), On the principles of the law of large numbers, the method of least squares
and statistics. MS, vol. 27, pp. 433 – 451.
--- (1912 – 1914), The Laplacean theory of the method of least squares simplified by
a theorem of Chebyshev. MS, vols. 28 – 29, pp. 228 – 234 and 190 – 191.
--- (1916a), Srednia Shkola […] (Secondary School […]. Petrograd.
--- (1916b), Prinzip Ekvivalentnosti Velichin etc. (Principle of the Equivalence of
Magnitdes). Petrograd.
Neugebauer O. (1948), Mathematical methods in ancient astronomy. In author’s
book (1983, pp. 99 – 127).
--- (1950), The alleged Babylonian discovery of the precession of the equinoxes.
Ibidem, pp. 247 – 254.
--- (1975), History of Ancient Mathematical Astronomy, pts 1 – 3. Berlin. [Berlin,
2004.]
--- (1983), Astronomy and History. Sel. Essays. New York
Newcomb S. (1859 – 1861), Notes on the theory of probability. Math. Monthly, vol.
2, 1860, pp. 134 – 140, 272 – 275.
--- (1860), [Discussion of the principles of probability theory.] Proc. Amer. Acad.
Arts and Sciences, vol. 4 for 1857 – 1860, pp. 433 – 440.
--- (1861a), On the secular variations and mutual relations of the orbits of the
asteroids. Mem. Amer. Acad. Arts and Sciences, New ser., vol. 8, pt. 1, pp. 123 –
152.
--- (1861b), Solution of problem. Math. Monthly, vol. 3, pp. 68 – 69.
--- (1862), Determination of the law of distribution of the nodes and perihelia of the
small planets. Astron. Nachr., Bd. 58, pp. 210 – 220.
--- (1867), Investigation of the orbit of Neptune. Smithsonian Contr. to Knowledge,
vol. 15. Separate paging.
--- (1872), On the Right Ascensions of the Equatorial Fundamental Stars.
Washington.
--- (1874), Investigation of the orbit of Uranus. Smithsonian Contr. to Knowledge,
vol. 19. Separate paging.
--- (1878), Researches of the motion of the Moon, pt. 1. Wash. Observations for
1875, Appendix 2.
--- (1881), Note on the frequency of use of the different digits in natural numbers.
Amer. J. Math., vol. 4, pp. 39 – 40.
--- (1886), A generalized theory of the combinations of observations. Ibidem, vol. 8,
pp. 343 – 366.
--- (1892), On the law and the period of the variation of terrestrial latitudes. Astron.
Nachr., Bd. 130, pp. 1 – 6.
--- (1895), The Elements of the Four Inner Planets and the Fundamental Constants
of Astronomy. Washington.
--- (1896), On the solar motion as a gauge of stellar distances. Astron. J., vol. 17, pp.
41 – 44.
--- (1897a), A new determination of the precessional constant. Astron. Papers, vol.
8, pp. 1– 76.
--- (1897b), A new determination of the precessional motion. Astron. J., vol. 17, pp.
161 – 167.
--- (1900), On the distribution of the mean motions of the minor planets. Astron. J.,
vol. 20, pp. 165 – 166.
--- (1901), On the period of the Solar spots. Astrophys. J., vol. 13, pp. 1 – 14.
--- (1902a), On the statistical relations among the parallaxes and the proper motions
of the stars. Astron. J., vol. 22, pp. 165 – 169.
--- (1902b), The Universe as an organism. In author’s Sidelights on Astronomy. New
York – London, 1906, pp. 300 – 311.
--- (1904a), On the Position of the Galactic. Carnegie Instn of Wash., Publ. No. 10.

288
286
--- (1904b), Statistical Inquiry into the Probability of Causes of Sex in Human
Offspring. Ibidem, Publ. No. 11.
Newcomb S., Holden E. S. (1874), On the possible periodic changes of the Sun’s
apparent diameter. Amer. J. Sci., ser. 3, vol. 8 (108), pp. 268 – 277.
Newton I. (1669 – 1671), Lectiones opticae. Opera quae Extant Omnia, vol. 3.
London, 1782, pp. 250 – 437. [London, 1931.]
--- (1704), Optics. Ibidem, vol. 4, pp. 1 – 264. [London, 1931.]
--- (1728), Chronology of Ancient Kingdoms Amended. London. [London, 1770.]
--- (1729), Mathematical Principles of Natural Philosophy. Third edition. Berkeley,
1960.
--- (1958), Papers and Letters on Natural Philosophy. Cambridge.
--- (1967), Mathematical Papers, vol. 1. Cambridge.
Neyman J. (1934), On two different aspects of the representative method. J. Roy.
Stat. Soc., vol. 97, pp. 558 – 625. Reprinted in Neyman (1967, pp. 98 – 141).
--- (1938), L’estimation statistique traitée comme un problème classique de
probabilité. Reprinted Ibidem, pp. 332 – 353.
--- (1938), Lectures and Conferences on Mathematical Statistics and Probability.
Washington, 1952.
--- (1967), Selection of Early Statistical Papers. Berkeley.
--- (1978), Review of Ondar (1977). Hist. Math., vol. 5, pp. 485 – 486.
Nicomachus of Gerasa (1st century, 1952), Introduction to Arithmetic. In Great
Books (1952, vol. 11, pp. 811 – 848).
Novikov S. P. (2002, Russian), The second half of the 20th century and its result:
the crisis of the physical and mathematical association in Russia and in the West.
IMI, vol. 7 (42), pp. 326 – 356.
Novoselsky S. A. (1916), Smertnost i Prodolzhitelnost Zhizni v Rossii (Mortality
and Duration of Life in Russia). Petrograd
Ondar Kh. O. (1971, Russian), On the work of A. Yu. Davidov in the theory of
probability and on his methodological views. Istoriya i Metodologiya Estestven.
Nauk, vol. 11, pp. 98 – 109. S, G, 5.
---, Editor (1977, Russian), Correspondence between Markov and Chuprov on the
Theory of Probability and Mathematical Statistics. New York, 1981.
Ore O. (1960), Pascal and the invention of probability theory. Amer. Math. Monthly,
vol. 67, pp. 409 – 419.
--- (1963), Cardano, the Gambling Scholar. Princeton, 1963; New York, 1965.
Oresme N. (14th century, publ. 1966, Latin and English), De Proportionibus
Proportionum and Ad Pauca Respicientes. Editor E. Grant. Madison..
Ostrogradsky M. V. (1848), Sur une question des probabilités. Bull. phys.-math.,
Imp. Acad. Sci. St. Pétersb., t. 6, No. 21 – 22, pp. 321 – 246. Russian translation:
Ostrogradsky (1961, pp. 215 – 237).
--- (1858, Russian), Note on the fund for honourably discharged persons. In
Ostrogradsky (1961, pp. 297 – 300).
--- (1961), Pedagogicheskoe Nasledie (Pedagogic Heritage). Editors I. B.
Pogrebyssky, A. P. Youshkevich. Moscow.
Paevsky V. V. (1935, Russian), Euler’s work in population statistics. In Euler.
Memorial volume. Moscow – Leningrad, pp. 103 – 110. S, G, 5.
Pannekouk A. (1961), History of Astronomy. New York, 1989.
Pascal B. (1654a, French), Correspondence with P. Fermat. OC, t. 1, pp. 145 – 166.
--- (1654b), A la très illustre Académie Parisienne de Mathématique (Latin and
French). Ibidem, pp. 169 – 173.
--- (1665), Traité du triangle arithmétique. Ibidem, pp. 282 – 327.
--- (1669), Pensées. Collection of fragments. OC, t. 2, pp. 543 – 1046.
--- (1998 – 2000), Oeuvres complètes, tt. 1 – 2. Paris.
Paty M. (1988), D’Alembert et les probabilités. In Rashed R., Editor, Sciences à
l’époque de la Révolution française. Paris, pp. 203 – 265.
Pearson E. S. (1936 – 1937), K. Pearson: an appreciation of some aspects of his life
and work. Biometrika, vol. 28, pp. 193 – 257; vol. 29, pp. 161 – 248.
--- (1990), ‘Student’. A Statistical Biography of W. S. Gosset. Editor, R. L. Plackett,
assisted by G. A. Barnard. Oxford.
Pearson E. S., Kendall M. G., Editors (1970), Studies in the History of Statistics
and Probability [vol. 1]. London. Collected reprints of papers.
Pearson E. S., Wishart J., Editors (1943), Student’s Collected Papers. London.

289
287
Pearson K. (1887), On a certain atomic hypothesis. Trans. Cambr. Phil. Soc., vol.
14, pp. 71 – 120.
--- (1891), Atom squirts. Amer. J. Math., vol. 13, pp. 309 – 362.
--- (1892), Grammar of Science. [Bristol, 1991; Tokyo, 1991.]
--- (1894), On the dissection of asymmetrical frequency curves. Phil. Trans. Roy.
Soc., vol. A185, pp. 71 – 110.
--- (1896), Skew variation in homogeneous material. Ibidem, vol. A186, pp. 343 –
414.
--- (1898), Cloudiness. Proc. Roy. Soc., vol. 62, pp. 287 – 290.
--- (1900), On a criterion that a given system of deviations … can be reasonably
supposed to have arisen from random sampling. London, Edinburgh and Dublin
Phil. Mag., ser. 5, vol. 50, pp. 157 – 175.
--- (1904), On a generalized theory of alternative inheritance. PT Roy. Soc., vol.
A203, pp. 53 – 86.
--- (1905), Das Fehlergesetz und seine Verallgemeinerung durch Fechner und
Pearson: a rejoinder. Biometrika, vol. 4, pp. 169 – 212.
--- (1907), On correlation and methods of modern statistics. Nature, vol. 76, pp. 577
– 578, 613 – 615, 662.
--- (1914 – 1930), Life, Letters and Labours of Fr. Galton, vols 1, 2, 3A, 3B.
Cambridge.
--- (1919), Peccavimus! Biometrika, vol. 12, pp. 259 – 281.
--- (1920), Notes on the history of correlation. Biometrika, vol. 55, pp. 459 – 467.
Reprinted in E. S. Pearson & Kendall (1970, pp. 185 – 205).
--- (1923), Darwin. London.
--- (1925), James Bernoulli’s theorem. Biometrika, vol. 17, pp. 201 – 210.
--- (1926), A. De Moivre. Nature, vol. 117, pp. 551 – 552.
--- (1928a), On a method of ascertaining limits to the actual number of marked
individuals […] from a sample. Biometrika, vol. 20A, pp. 149 – 174.
--- (1928b), The contribution of G. Plana to the normal bivariate frequency surface.
Ibidem, pp. 295 – 298.
--- (1934), Tables of the Incomplete Beta-Function. Cambridge, 1968.
--- (1948), Early Statistical Papers. Cambridge.
--- (1978), History of Statistics in the 17th and 18th Centuries against the Changing
Background of Intellectual, Scientific and Religious Thought. Lectures 1921 – 1933.
Editor E. S. Pearson. London.
Peters [C. A. E.] (1856), Über die Bestimmung der wahrscheinlichen Fehlers eines
Beobachtung. Astron. Nachr., Bd. 44, pp. 29 – 32.
Petrov V. V. (1954, Russian), The method of least squares and its extreme
properties. UMN, vol. 1, pp. 41 – 62.
Petruszewycz M. (1983), Description statistique de textes littéraires russes par les
méthodes de Markov. Rev. Etudes Slaves, t. 55, pp. 105 – 113.
Pettenkofer, M. (1865), Über die Verbreitungsart der Cholera. Z. Biologie, Bd. 1,
pp. 322 – 374.
--- (1873), Über den Werth der Gesundheit für eine Stadt. Braunschweig. English
transl.: Bull. Hist. Med., vol. 10, No. 3 and 4, 1941.
--- (1886 – 1887), Zum gegenwärtigen Stand der Cholerafrage. Arch. f. Hyg., Bd. 4,
pp. 249 – 354, 397 – 546; Bd. 5, pp. 353 – 445; Bd. 6, pp. 1 – 84, 129 – 233, 303 –
358, 373 – 441; Bd. 7, pp. 1 – 81.
Petty W. (1662), A Treatise of Taxes and Contributions. In Petty (1899, vol. 1, pp. 1
– 97).
--- (1674), Discourse Read before the Royal Society. London.
--- (1690). Political Arithmetic. In Petty (1899, vol. 2, pp. 239 – 313).
--- (1899), Economic Writings, vols. 1 – 2. Ed., C. H. Hull. Cambridge. Reprint:
London, 1997.
--- (1927), Papers, vols. 1 – 2. London. Reprint: London, 1997.
Pfanzagl J., Sheynin O. (1997), Süssmilch. Incorporated in second edition (2006)
of Kotz & Johnson, vol. 13, pp. 8489 – 8491 but somehow appeared there
anonymously.
Picard, J. (1693), Observations astronomiques faites en diverse endroits du royame.
Mém. Acad. Roy. Sci. 1666 – 1699, t. 7. Paris, 1729, pp. 329 – 347.

290
288
Pirogov N. I. (1849, Russian), On the achievements of surgery during the last five
years. Zapiski po Chasti Vrachebnykh Nauk Med.-Khirurgich. Akad., Year 7, pt. 4,
sect. 1, pp. 1 – 27.
--- (1850 – 1855, Russian), Letters from Sevastopol. In Pirogov (1957 – 1962, vol.
8, 1961, pp. 313 – 403).
--- (1854), Statistischer Bericht über alle meine im Verlauf eines Jahres ...
vorgenommenen oder beobachteten Oprationsfälle, this being the author’s booklet
Klinische Chirurgie, No. 3. Leipzig. Reprinted in vols. 5 and 6 of his Sobranie
Sochineniy (see vol. 5, p. 20).
--- (1864), Grundzüge der allgemeinen Kriegschirurgie. Leipzig. Russian version:
1865 – 1866.
--- (1871), Bericht über die Besichtigung der Militär-Sanitäts-Anstalten in
Deutschland, Lothringen und Elsass im Jahre 1870. Leipzig.
--- (1879, Russian), Das Kriegs-Sanitäts-Wesen und die Privat-Hilfe auf dem
Kriegsschauplätze etc. Leipzig, 1882.
--- (1957 – 1962), Sobranie Sochineniy (Coll. Works), vols. 1 – 8. Moscow.
Pitman E. J. G. (1939), Estimation of the location and scale parameters.
Biometrika, vol. 30, pp. 391 – 421.
Plackett R. L. (1958), The principle of the arithmetic mean. Biometrika, vol. 45, pp.
130 – 135. Reprinted in E. S. Pearson & Kendall (1970, pp. 121 – 126).
--- (1972), The discovery of the method of least squares. Biometrika, vol. 59, pp.
239 – 251. Reprinted in Kendall & Plackett (1977, pp. 279 – 291).
--- (1988), Data analysis before 1750. ISR, vol. 56, pp. 181 – 195.
Plato J. von (1983), The method of arbitrary functions. Brit. J. Phil. Sci., vol. 34,
pp. 37 – 47.
Poincaré H. (1892a), Thermodynamique. Paris.
--- (1892b), Réponse à P. G. Tait. Œuvres, t. 10. Paris, 1954, pp. 236 – 237.
--- (1894), Sur la théorie cinétique des gaz. Ibidem, pp. 246 – 263.
--- (1896), Calcul des probabilités. Paris, 1912, reprinted 1923 and 1987.
--- (1902), La science et l’hypothèse. Paris, 1923.
--- (1905), La valeur de la science. Paris, 1970.
--- (1921), Résumé analytique [of own works]. In Mathematical Heritage of H.
Poincaré. Providence, Rhode Island, 1983. Editor F. E. Browder, pp. 257 – 357.
Poisson S.-D. (1812), Review of Laplace (1812). Nouv. Bull. des Sciences. Soc.
Philomatique de Paris, t. 3, pp. 160 – 163. S, G, 58.
--- (1824 – 1829), Sur la probabilité des résultats moyens des observations. Conn.
des Temps, pour 1827, pp. 273 – 302; pour 1832, pp. 3 – 22.
--- (1825 – 1826), Sur l’avantage du banquier au jeu de trente-et-quarante. Annales
Math. Pures et Appl., t. 16, pp. 173 – 208.
--- (1830), Formules des probabilités […] qui peuvent être utiles dans l’artillerie.
Mémorial de l’Artillerie, No. 3, pp. 141 – 156. This periodical was also published in
Bruxelles with a different numbering of pages.
--- (1833a), Traité de mécanique, t. 1. Paris. Second edition. First edition, 1811.
--- (1833b), Discours prononcé aux funérailles de Legendre. J. f. d. reine u. angew.
Math., Bd. 10, pp. 360 – 363. S, G, 58.
--- (1835), Recherches sur la probabilité des jugements etc. C. r. Acad. Sci. Paris, t.
1, pp. 473 – 494.
--- (1836), Note sur la loi des grands nombres. Ibidem, t. 2, pp. 377 – 382.
--- (1837a), Recherches sur la probabilité des jugements en matière criminelle et en
matière civile. Paris. [Paris, 2003.] S, G, 53.
--- (1837b), Sur la probabilité du tir a la cible. Mémorial de l’Artillerie, No. 4, pp. 59
– 94.
--- (1837c), Programmes de l’enseignement de l’Ecole Polytechnique […] pour
l’année scolaire 1836 – 1837. Paris.
--- (2013), Les mathématiques au service de la science. Paris. Editor, Yvette
Kossmann-Schwarzbach.
Polya G. (1920), Über den zentralen Grenzwertsatz der
Wahrscheinlichkeitsrechnung und das Momentenproblem. Math. Z., Bd. 8, pp. 171 –
181.
--- (1931), Sur quelques points de la théorie des probabilités. Annalen Inst. H.
Poincaré, t. 1, pp. 117 – 161.
--- (1954), Mathematics and Plausible Reasoning. Princeton.

291
289
--- (1984), Collected Papers, vol. 4. Cambridge (Mass.).
Postnikov A. G. (1974), Veroiatnostnaia Teoria Chisel (Stochastic Number
Theory). Moscow.
Prevost P., Lhuillier S. A. J. (1799), Sur l’art d’estimer la probabilité des causes
par les effets. Mém. Acad. Roy. Sci. et Belles-Lettres Berlin avec l’Histoire 1796, pp.
3 – 24 of the second paging.
Price D. J. (1955), Medieval land surveying and topographical maps. Geogr. J., vol.
121, pp. 1 – 10.
Le procès (1900), Le procès Dreyfus devant le Conceil de guerre de Rennes, tt. 1 –
3. Paris.
Proctor R. A. (1872), On star-grouping. Proc. Roy. Instn Gr. Brit., vol. 6, pp. 143 –
152.
--- (1874), The Universe. London.
Prokhorov Yu. V. (1986, Russian), The law of large numbers and the estimation of
the probabilities of large deviations. In Bernoulli, J. (1986, pp. 116 – 150).
--- (1999, Russian), The Bertrand paradox. In Prokhorov (1999b, p. 46).
---, Editor (1999b), Veroiatnost i Matematicheskaia Statistika. Enziklopedia
(Probability and Math. Statistics. Enc.). Moscow.
Prokhorov Yu. V., Sevastianov B. A. (1999, Russian), Probability theory. In
Prokhorov (1999b, pp. 77 – 81).
Prudnikov V. E. (1964, Russian), Chebyshev. [Leningrad, 1976.]
Ptolemy (2nd century; 1956), Tetrabiblos. In Greek and English. London.
--- (2nd century; 1984), Almagest. London.
Ptukha M. V. (1955), Ocherki po Istorii Statistiki v SSSR (Essays on the History of
Statistics in the Soviet Union), vol. 1. Moscow.
--- (1961, Russian), Sample inspection of agriculture in Russia in the 17th and 18th
centuries. Uch. Zap. po Statistike, vol. 6, pp. 94 – 100. S, G, 36.
Quetelet A. (1829), Recherches statistiques sur le Royaume des Pays-Bas, Mém.
Acad. Roy. Sci., Lettres et Beaux-Arts Belg., t. 5, separate paging.
--- (1832a), Recherches sur la loi de la croissance de l’homme. Ibidem, t. 7. Separate
paging.
--- (1832b), Recherches sur le penchant au crime. Ibidem. Separate paging.
--- (1836), Sur l’homme et le développement de ses facultés, ou essai de physique
sociale, tt. 1 – 2. Bruxelles.
--- (1846), Lettres ... sur la théorie des probabilités. Bruxelles.
--- (1848a), Du système social. Paris.
--- (1848b), Sur la statistique morale. Mém. Acad. Roy. Sci., Lettres et Beaux-Arts
Belg., t. 21. Separate paging.
---, publie par (1849 – 1857), Sur le climat de la Belgique, tt. 1 – 2. Bruxelles.
--- (1852), Sur quelques propriétés curieuses qui présentent les résultats d’une série
d’observations. Bull. Acad. Roy. Sci., Lettres, Beaux-Arts Belg., t. 19, pt. 2, pp. 303
– 317.
--- (1853), Théorie des probabilités. Bruxelles.
--- (1869), Physique sociale, tt. 1 – 2. Bruxelles, this being a new edition of Quetelet
(1836).
--- (1870), Des lois concernant le développement de l’homme. Bull. Acad. Roy. Sci.,
Lettr., Beaux-Arts Belg., 39e année, t. 29, pp. 669 – 680.
--- (1871), Anthropométrie. Bruxelles.
--- (1974), Mémorial. Bruxelles.
Quetelet A., Heuschling X. (1865), Statistique internationale (population).
Bruxelles.
Quine M. P., Seneta E. (1987), Bortkiewicz’ data and the law of small numbers.
ISR, vol. 55, pp. 173 – 181.
Rabinovitch N. L. (1973), Probability and Statistical Inference in Ancient and
Medieval Jewish Literature. Toronto.
--- (1974), Early antecedents of error theory. AHES, vol. 13, pp. 348 – 358.
--- (1977), The one and the many: stochastic reasoning in philosophy. Annals of sci.,
vol. 13, pp. 348 – 358.
Radelet de Grave P., Scheuber V. (1979), Correspondance entre D. Bernoulli et
J.-H. Lambert. Paris.
Radlov E. L. (1900, Russian), V. S. Soloviev: his religious and philosophical views.
Enz. Slovar Brfockhaus & Efron, halfvol. 60, pp. 785 – 792.

292
290
Raikher V. K. (1947), Obshchestvenno-Istoricheskie Tipy Strakhovania (Social
Types of Insurance in History). Moscow – Leningrad.
Raimi R. A. (1976), The first digit problem. Amer. Math. Monthly, vol. 83, pp. 521
– 538.
Raju C. K. (2010), Probability in ancient India. Handbook of the Philosophy of
Science, vol. 7. N. p., pp. 1 – 24.
Rao C. R. (1965), Linear Statistical Inference and Its Applications. New York.
Réaumur R. A. (1738), Observations du thermomètre etc. Hist. Acad. Roy. Sci.
Paris avec Mém. math.-phys. 1735, pp. 545 – 576.
Rehnisch E. (1876), Zur Orientierung über die Untersuchungen und Ergebnisse der
Moralstatistik. Z. Philos. u. philos. Kritik, Bd. 69, pp. 43 – 115, this being pt 2 of his
article. Part 3 had not appeared.
Renyi A. (1969), Briefe über Wahrscheinlichkeitsrechnung. Budapest.
Report (1916, Russian), Report of the [ad hoc] Commission to discuss some issues
concerning the teaching of mathematics in high school. Izvestia Imp. Akad. Nauk,
ser. 6, vol. 10, No. 2, pp. 66 – 80.
Rigaud S. P. (1832), Miscellaneous Works and Correspondence of J. Bradley.
Oxford. [Oxford, 1972.]
--- (1841), Correspondence of Scientific Men of the 17th Century, vol. 2. Oxford.
Romanovsky, V. I. (1912), Zakon Bolshikh Chisel i Teorema Bernoulli (The Law of
Large Numbers and the Bernoulli Theorem). Warsaw. Also in Protokoly Zasedaniy
Obshchestvo Estestvoispytatelei Univ. Varshava for 1911, No. 4, pp. 39 – 63.
--- (1923, Russian), Review of Chuprov (1918 – 1919), Vestnik Statistiki, Book 13,
No. 1 – 3, pp. 255 – 260.
--- (1924, Russian), Theory of probability and statistics according to some newest
works of Western scholars. Vestnik Statistiki, No. 4 – 6; No. 7 – 9, pp. 1 – 38 and 5 –
34.
--- (1925a, Russian), Generalization of the Markov inequality and its application in
correlation theory. Bull. Sredneaziatsk. Gos. Univ., vol. 8, pp. 107 – 111.
--- (1925b), Généralisation d’une inégalité de Markoff. C. r. Acad. Sci. Paris, t. 180,
pp. 1468 – 1470.
--- (1930), Matematicheskaia Statistika (Math. Statistics). Moscow – Leningrad.
--- (1961), Same title, book 1. Tashkent. Editor, T. A. Sarymsakov.
Rosental I. S., Sokolov V. S. (1956), Uchebnik Latinskogo Yazyka (Textbook of
Latin). Moscow.
Rubanovsky L. M. (1934), Metody Fizicheskoi Statistiki (Methods of Physical
Statistics). Leningrad – Moscow.
Rümelin G. (1867), Über den Begriff eines socialen Gesetzes. In author’s Reden
und Aufsätze. Freiburg i/B – Tübingen, 1875, pp. 1 – 31.
Russell B. (1962), History of Western Philosophy. London.
Sakatov P. S. (1950, Russian), Lehrbuch der höheren Geodäsie. Berlin, 1957. Other
Russian editions, 1953, 1964.
Sambursky S. (1956), On the possible and probable in ancient Greece. Osiris, vol.
12, pp. 35 – 48. Reprinted in M. G. Kendall & Plackett (1977, pp. 1 – 14).
Särndal C.-E. (1971), The hypothesis of elementary errors and the Scandinavian
school in statistical theory. Biometrika, vol. 58, pp. 375 – 392. Reprinted in Kendall
M. G. & Plackett (1977, pp. 1 – 14).
Sartorius von Waltershausen, W. (1856), Gauss zum Gedächtnis. Wiesbaden,
1965.
Scheffé H. (1959), Analysis of Variance. New York.
Schell E. D. (1960), Pepys, Newton and probability. Amer. Statistician, vol. 14, No.
4, pp. 27 – 30.
Schlözer A. L. (1804), Theorie der Statistik. Göttingen.
Schmeidler F. (1984), Leben und Werk des Königsberger Astronomen Bessel.
Kelkheim/T.
Schmidt O. Yu. (1926, Russian), On the Bertrand paradox in the theory of
probability. MS, vol. 33, pp. 33 – 40.
Schneider I. (1968), Der Mathematiker A. De Moivre. AHES, vol. 5, pp. 177 – 317
(the entire No. 3 – 4).
--- Editor (1988), Die Entwicklung der Wahrscheinlichkeitstheorie von den
Anfängen bis 1933. Darmstadt. Collection of fragments almost exclusively in
German with some comment.

293
291
Schreiber [O.] (1879), Richtungsbeobachtungen und Winkelbeobachtungen. Z. f.
Vermessungswesen, Bd. 8, pp. 97 – 149.
Schumacher H. (1931), Ladislaus von Bortkiewicz. Allg. stat. Arch., Bd. 21, pp.
573 – 576.
Schumann R. (1917), F. R. Helmert. Österr. Z. f. Vermessungswesen, Jg. 15, No.
718, pp. 97 – 100.
Schumpeter (1954), History of Economic Analysis. New York, 1955.
Seal H. L. (1967), Historical development of the Gauss linear model. Biometrika,
vol. 54, pp. 1 – 24. Reprinted in E. S. Pearson & Kendall (1970, pp. 207 – 230).
Seidel L. (1865), Über den … Zusammenhang … zwischen der Häufigkeit der
Typhus-Erkrankungen und dem Stande des Grundwassers. Z. Biol., Bd. 1, pp. 221 –
236.
--- (1866), Vergleichung der Schwankungen der Regenmengen mit der
Schwankungen in der Häufigkeit des Typhus. Ibidem, Bd. 2, pp. 145 – 177.
Seneta E. (1983), Modern probabilistic concepts in the work of E. Abbe and A. De
Moivre. Math. Scientist, vol. 8, pp. 75 – 80.
--- (1984), The central limit theorem and linear least squares in pre-revolutionary
Russia: the background. Ibidem, vol. 9, pp. 37 – 77.
--- (1987), Chuprov on finite exchangeability, expectation of ratios and measures of
association. Hist. Math., vol. 14, pp. 243 – 257.
--- (1994), Carl Liebermeister’s hypergeometric tails. Ibidem, vol. 21, pp. 453 – 462.
--- (1998), Bienaymé (1798 – 1878): criticality, inequality and internationalization.
ISR, vol. 66, pp. 291 – 301.
Seneta E., Parshall K. H., Jongmans F. (2001), Nineteenth-century developments
in geometric probability. AHES, vol. 55, pp. 501 – 524.
Shafer G. (1978), Non-additive probabilities in the work of Bernoulli and Lambert.
AHES, vol. 19, pp. 309 – 370.
Shaw N., Austin E. (1942), Manual of Meteorology, vol. 1. Cambridge.
Sheynin O. B. (1963), Adjustment of a triangulation figure by frame structure
analogue. Survey Rev., vol. 17, pp. 55 – 56.
--- (1965, Russian), On Adrain’s work in the theory of errors. IMI, vol. 16, pp. 325 –
336. S, G, 1.
--- (1966), Origin of the theory of errors. Nature, vol. 211, pp. 1003 – 1004.
--- (1968), On the early history of the law of large numbers. Biometrika, vol. 55, pp.
459 – 467. Reprinted in E. S. Plackett & Kendall (1970, pp. 231 – 239).
--- (1970а, Russian), On the history of the De Moivre – Laplace limit theorems.
Istoria i Metodologia Estestven. Nauk, vol. 9, pp. 199 – 211.
--- (1970b), D. Bernoulli on the normal law. Biometrika, vol. 57, pp. 199 – 202.
Reprinted in M. G. Kendall & Plackett (1977, pp. 101 – 104).
--- (1970c, Russian), Combinatorial analysis and the theory of probability. Co-
authors L. E. Maistrov, B. A. Rosenfeld. Chapter in Istoria Matematiki s
Drevneishikh Vremen do Nachala 19-go Stoletia (Hist. of Math. from Most Ancient
Times to Beginning of the 19th Century), vol. 2, 18th century. Editor, A. P.
Youshkevich. Moscow, 1970, pp. 81 – 97.
--- (1971a), Newton and the classical theory of probability. AHES, vol. 7, pp. 217 –
243.
--- (1971b), Lambert’s work in probability. Ibidem, pp. 244 – 256.
--- (1971c), On the history of some statistical laws of distribution. Biometrika, vol.
58, pp. 234 – 236. Reprinted in M. G. Kendall & Plackett (1977, pp. 328 – 330).
--- (1972a), D. Bernoulli’s work on probability. RETE. Strukturgeschichte der
Naturwissenschaften, Bd. 1, pp. 273 – 300. Reprinted in M. G. Kendall & Plackett
(1977, pp. 105 – 132).
--- (1972b, Russian), Theory of probability. Co-author L. E. Maistrov. Chapter in
same source as in (1970c), vol. 3, 19th century. Editor, A. P. Youshkevich. Moscow,
1972, pp. 126 – 152.
--- (1973a), Finite random sums. Historical essay. AHES, vol. 9, pp. 275 – 305.
--- (1973b), Boscovich’s work on probability. Ibidem, pp. 306 – 324.
--- (1973c), Mathematical treatment of astronomical observations. Historical essay.
AHES, vol. 11, pp. 97 – 126.
--- (1974), On the prehistory of the theory of probability. AHES, vol. 12, pp. 97 –
141.
--- (1975), Kepler as a statistician. Bull. ISI, t. 46, pp. 341 – 354.

294
292
--- (1976), Laplace’s work on probability. AHES, vol. 16, pp. 137 – 187.
--- (1977a), Laplace’s theory of errors. AHES, vol. 17, pp. 1 – 61.
--- (1977b), Early history of the theory of probability. Ibidem, pp. 201 – 259.
--- (1978а, Russian), Theory of probability. Co-author, B. V. Gnedenko. Chapter in
Mathematics of the 19th century [vol. 1]. Editors, A. N. Kolmogorov, A. P.
Youshkevitch. Basel, 1992 and 2001, pp. 211 – 288.
--- (1978b), Poisson’s work in probability. AHES, vol. 18, pp. 245 – 300.
--- (1979), Gauss and the theory ofberrors. AHES, vol. 20, pp. 21 – 72.
--- (1980), On the history of the statistical method in biology. AHES, vol. 22, pp.
323 – 371.
--- (1982), On the history of medical statistics. AHES, vol. 26, pp. 241 – 286.
--- (1983), Corrections and short notes on my papers. AHES, vol. 28, pp. 171 – 195.
--- (1984a), On the history of the statistical method in astronomy. AHES, vol. 29,
pp. 151 – 199.
--- (1984b), On the history of the statistical method in meteorology. AHES, vol. 31,
pp. 53 – 93.
--- (1985), On the history of the statistical method in physics. AHES, vol. 33, pp.
351 – 382.
--- (1986a), Quetelet as a statistician. AHES, vol. 36, pp. 281 – 325.
--- (1986b, Russian), J. Bernoulli and the beginnings of probability theory. In
Bernoulli J. (1986, pp. 83 – 115).
--- (1989a), Markov’s work on probability. AHES, vol. 39, pp. 337 – 377; vol. 40, p.
387.
--- (1989b, Russian), A. M. Liapunov’s letters to K. A. Andreev. IMI, vol. 31, pp.
306 – 313. S, G, 1.
--- (1990a, Russian), On the history of the statistical method in natural science. IMI,
vol. 32/33, pp. 384 – 408. S, G, 1
--- (1990b), Markov’s review of Galitzin (1902). Ibidem, pp. 451 – 467. S, G, 85.
--- (1990c, Russian), Chuprov: Life, Work, Correspondence. Göttingen, 1996 and
2011.
--- (1991a), Poincaré’s work in probability. AHES, vol. 42, pp. 137 – 172.
--- (1991b), On the works of Buniakovsky in the theory of probability. AHES, vol.
43, pp. 199 – 223.
--- (1992), Al-Biruni and the mathematical treatment of observations. Arabic
Sciences and Phil., vol. 2, pp. 299 – 306.
--- (1993а, Russian), Markov’s letters in newspaper Den in 1914 – 1915. IMI, vol.
34, pp. 194 – 206. S, G, 85.
--- (1993b), On the history of the principle of least squares. AHES, vol. 46, pp. 39 –
54.
--- (1993c), Treatment of observations in early astronomy. Ibidem, pp. 153 – 192.
--- (1993d), Chuprov, Slutsky and Chetverikov: some comments. Hist. Math., vol.
20, pp. 247 – 254.
--- (1994a), Gauss and geodetic observations. AHES, vol. 46, pp. 253 – 283.
--- (1994b), Chebyshev’s lectures on the theory of probability. Ibidem, pp. 321 –
340.
--- (1994c), Bertrand’s work on probability. AHES, vol. 48, pp. 155 – 199.
--- (1994d), Ivory’s treatment of pendulum observations. Hist. Math., vol. 21, pp.
174 – 184.
--- (1994e), Theory of errors. In Companion Enc. Hist. and Phil. of Math. Sciences,
vol. 2, pp. 1315 – 1324. London, Editor I. Grattan-Guinness. [Johns Hopkins Univ.
Publ., 2005.]
--- (1994f, Russian), Correspondence between P. A. Nekrasov and K. A. Andreev.
IMI, vol. 35, pp. 124 – 147. Co-author, M. V. Chirikov. S, G, 1.
--- (1995a, Russian), Nekrasov’s letters to Chuprov. IMI, vol. 1 (36), No. 1, pp. 159
– 167.
--- (1995b), Helmert’s work in the theory of errors. AHES, vol. 49, pp. 73 – 104.
--- (1995c), Density curves in the theory of errors. Ibidem, pp. 163 – 196.
--- (1995d), Introduction of statistical reasoning into astronomy. In Taton & Wilson
(1995, pp. 191 – 197).
--- (1996a), History of the Theory of Errors. Egelsbach.
--- (1996b), A forerunner of the t-distribution. Biometrika, vol. 83, pp. 891 – 898.
Co-author, J. Pfanzagl.

295
293
--- (1996c), Selection and treatment of observations by Mendeleev. Hist. Math. , vol.
23, pp. 54 – 67.
--- (1997a), Lüroth. Co-author J. Pfanzagl. In Johnson & Kotz (1997, pp. 203 –
204). Incorporated in second edition (2006) of Kotz & Johnson.
--- (1997b), Achenwall. In second edition of Kotz & Johnson (2006, vol. 1, pp. 26 –
27).
--- (1997c), Süssmilch. See Pfanzagl & Sheynin.
--- (1997d, Russian), Markov and insurance of life. Math. Scientist, vol. 30, 2005,
pp. 5 – 12.
--- (1998a), Theory of probability: definition and relation to statistics. AHES, vol.
52, pp. 99 – 108.
--- (1998b), Statistical thinking in the Bible and the Talmud. Annals Sci., vol. 55, pp.
185 – 198.
--- (1998c), Statistics in the Soviet epoch. JNÖS, Bd. 217, pp. 529 – 549.
--- (1999a), Discovery of the principle of least squares. Hist. Scient., vol. 8, pp. 249
– 264.
--- (1999b, Russian), Slutsky: fifty years after his death. IMI, vol. 3 (38), pp. 128 –
137.
--- (1999c), Gauss and the method of least squares. JNÖS, Bd. 219, pp. 458 – 467.
--- (2000), Bessel: some remarks on his work. Hist. Scient., vol. 10, pp. 77 – 83.
--- (2001a), Pirogov as a statistician. Ibidem, pp. 213 – 225.
--- (2001b), Gauss. In Heyde & Seneta (2001, pp. 119 – 122).
--- (2001c), Mendel. In Heyde & Seneta (2001, pp. 190 – 193.
--- (2001d), Gauss, Bessel and the adjustment of triangulation. Hist. Scient., vol. 11,
pp. 168 – 175.
--- (2001e), Social statistics and probability theory in the 19th century. Ibidem, pp.
86 – 111. Shorter version: JNÖS, Bd. 223, 2003, pp. 91 – 112.
--- (2001f), Anderson’s forgotten obituary of Bortkiewicz. JNÖS, Bd. 221, pp. 226
– 236.
--- (2002a), Newcomb as a statistician. Hist. Scient., vol. 12, pp. 142 – 167.
--- (2002b), Sampling without replacement: history and applications. NTM, Intern.
Z. f. Geschichte, u. Ethik Naturwiss., Technik u. Med., Bd. 10, pp. 181 – 187.
--- (2003a), Russian and English), Gumbel, Einstein and Russia. Moscow. S, G, 12.
--- (2003b), Nekrasov’s work on probability: the background. AHES, vol. 57, pp.
337 – 353.
--- (2003c), On the history of Bayes’s theorem. Math. Scientist, vol. 28, pp. 37 – 42.
--- (2003d), Geometric probability and the Bertrand paradox. Hist. Scient., vol. 13,
pp. 42 – 53.
--- (2004), Fechner as a statistician. Brit. J. Math., Stat. Psychology, vol. 57, pp. 53 –
72.
--- (2006a, Russian), On the relations between Chebyshev and Markov. IMI, vol. 11
(46), pp. 148 – 157.
--- (2006b), Markov’s work on the treatment of observations. Hist. Scient., vol. 16,
pp. 80 – 95.
--- (2006c), Review of T. M. Porter’s Karl Pearson (Princeton, 2006). Ibidem, pp.
206 – 209.
--- (2006d), Review of translation of Jakob Bernoulli’s Ars Conjectandi by E. D.
Sylla (Baltimore, 2006). Hist. Scient., vol. 16, pp. 212 – 214.
--- (2006/2009), Theory of Probability and Statistics As Exemplified in Short
Dictums. Berlin, 2009.
--- (2007a), The true value of a measured constant and the theory of errors. Hist.
Scient., vol. 17, pp. 38 – 48.
--- (2007b), Euler’s work in probability and statistics. In Euler Reconsidered.
Tercentenary Essays. Ed., R. Baker. Heber City, Uta, pp. 281 – 316.
--- (2007c), Markov: integrity is just as important as scientific merits. NTM, Intern.
Z. f. Geschichte u. Ethik Naturwiss., Technik, Med., Bd. 15, pp. 289 – 294.
--- (2007d, Russian), The correspondence between Slutsky and Bortkevich. Finansy
i Biznes, No. 4, pp. 139 – 154. Co-authors G. Rauscher, C. Wittich.
--- (2008a), Bortkiewicz’ alleged discovery: the law of small numbers. Hist. Scient.,
vol. 18, pp. 36 – 48.
--- (2008b), Romanovsky’s correspondence with K. Pearson and F. A. Fisher.
Archives d’histoire des sciences, t. 58, No. 160 – 161, pp. 365 – 384.

296
294
--- (2010a) Karl Pearson a century and a half after his birth. Math. Scientist, vol. 35,
pp. 1 – 9.
--- (2010b), The inverse law of large numbers. Math. Scientist, vol. 35, pp. 132 –
133.
--- (2011), Statistics, History of. Intern. Enc. of Stat. Sci. Göttingen, pp. 1493 –
1504. Author of § 1: M. Lovrich.
--- (2012а), New exposition of Gauss’ final justification of least squares. Math.
Scientist, vol. 37, pp. 147 – 148.
--- (2012b), Poisson and statistics. Ibidem, pp. 149 – 150.
--- (2012с), Ladislaus von Bortkiewicz: a scientific biography. In W. Wieslaw,
Editor, Dzieje Matematyki Polskiej. Wroclaw, pp. 249 – 266.
--- (2013), Poisson et la statistique. In Poisson (2013, pp. 357 – 366).
--- (2014a), Gauss and the method of least squares. Silesian Stat. Rev., No. 12 (18),
pp. 9 – 36.
--- (2014b), Elementary exposition of Gauss (1823). Ibidem, pp. 39 – 47.
--- (2014c), Antistigler. Ibidem, pp. 48 – 52. S, G, 31.
--- (2014d), Theory of errors and statistics. Ibidem, pp. 53 – 54.
--- (2014e), Randomness and determinism. Ibidem, pp. 57 – 74.
--- (2014, 2015), Cluster of anniversaries. Math. Scientist, vol. 39, pp. 11 – 16; vol.
40, pp. 1 – 6.
--- (2016a), The history and principle of statistics. Wiley, Stats Ref: Statistics
reference online
--- (2016b), The other Bessel, 2 pages in S, G, 72.
--- (2016c), On the history of university statistics. Silesian Stat. Rev., No. 14/20, pp.
7 – 25.
--- (2017), Black Book. S, G, 80.
Shiyanov A. N. (1836), On the method of least squares. In Bolotov A. P.
Geodeziya, pt. 1. Petersburg, pp. 321 – 360.
Shoesmith E. (1985a), N. Bernoulli and the argument for Divine Providence. ISR,
vol. 53, pp. 255 – 259.
--- (1985b), T. Simpson and the arithmetic mean. Hist. Math., vol. 12, pp. 352 – 355.
--- (1986), Huygens’ solution to the gambler’s ruin problem. Ibidem, vol. 13, pp.
157 – 164.
--- (1987), The Continental controversy over Arbuthnot’s argument for Divine
Providence. Ibidem, vol. 14, pp. 133 – 146.
Short J. (1763), Second paper concerning the parallax of the Sun. PT Roy. Soc., vol.
53, pp. 300 – 342.
Simon J. (1887), Public Health Reports, vols 1 – 2. London.
Simpson J. Y. (1847 – 1848), Anaesthesia. Works, vol. 2. Edinburgh, 1871, pp. 1 –
288.
--- (1869 – 1870), Hospitalism. Ibidem, pp. 289 – 405.
Simpson T. (1740), Nature and Laws of Chance. London.
--- (1756), On the advantage of taking the mean of a number of observations. PT
Roy. Soc., vol. 49, pp. 82 – 93.
--- (1757), Extended version of same: in author’s book Miscellaneous Tracts on
Some Curious […] Subjects […]. London, pp. 64 – 75.
--- (1775), Doctrine of Annuities and Reversions. London.
Sleshinsky I. V. (1892, Russian), On the theory of the method of least squares.
Zapiski Novoross. Obshchestva Estestvoispytatelei, vol. 14, pp. 201 – 264.
--- (1893, Russian), On Chebyshev’s theorem. Zapiski Novoross. Univ., vol. 59, pp.
503 – 506.
Slutsky E. E. (1912a), Teoria Korreliatsii (Correlation theory). Kiev. S, G, 38.
--- (1912b, English), Three letters to K. Pearson. Univ. College London. Pearson
Papers 856/4 and 856/7. In Sheynin (1999e, pp. 78 – 84).
--- (1914), On the criterion of goodness of fit of the regression lines and on the best
method of fitting them to data. J. Roy. Stat. Soc., vol. 77, pp. 78 – 84.
--- (1922, Russian), On then logical foundation of the calculus of probability. In
Slutsky (1960, pp. 18 – 24).
--- (1960), Izbrannye Trudy (Selected Works). Moscow.
Snow J. (1855), On the mode of communication of cholera. In Snow on Cholera.
New York, 1965, pp. 1 – 139.

297
295
Sofonea T. (1957a), Leibniz und sein Projekt zur Errichtung staatlicher
Versicherungsanstalten. Schweiz. Versicherungs-Zeitschrift, Jg. 25, pp. 144 – 148.
--- (1957b), E. Halley (1656 – 1742) und seine Sterbetafel 300 Jahre nach seiner
Geburt. Het Verzerkerings Archief, t. 34, pp. 31* – 42*.
Soloviev A. D. (1997, Russian), Nekrasov and the central limit theorem of the
theory of probability. Archives Intern. d’Hist. Sci., vol. 58, 2008, No. 160 – 161, pp.
353 – 364.
Spieß O. (1975), Zur Vorgeschichte des Petersburger Problems. In J. Bernoulli
(1975, pp. 557 – 567).
Spieß W. (1939), Kann man für D. Huber Ansprüche als Erfinder der Methode der
kleinsten Quadrate geltend machen? Schweiz. Z. Vermessungswesen u.
Kulturtechnik, Bd. 37, pp. 11 – 17, 21 – 23.
Steklov V. A. (1915, Russian), On a problem of Laplace. Izvestia Imp. Akad. Nauk,
Ser. 6, t. 9, pp. 1515 – 1537.
--- (1924, in Russian), A. A. Markov. Izvestia Ross. Akad. Nauk, ser. 6, vol. 16 for
1922, pp. 169 – 184. S, G, 85.
Stigler S. M. (1977), Eight centuries of sampling inspection: the trial of the pyx. J.
Amer. Stat. Assoc., vol. 72, pp. 493 – 500.
---, Editor (1980), American Contributions to Mathematical Statistics in the 19th
Century, vols. 1 – 2. New York. Reprints of papers of early American authors. No
single paging.
--- (1986), History of Statistics. Cambridge (Mass.).
--- (1999), Statistics on the Table. Cambridge (Mass.). Collected (perhaps revised)
papers.
Stiltjes T. J. (1885), Note a l’occasion de la reclamation de M. Markoff. OC, t. 2.
Berlin, 1993, pp. 430 – 431/518 – 519.
Strabo (1969), Geography, vol. 1. Written about the year 7.
Strecker H., Strecker R. (2001), Anderson. In Heyde & Seneta (2001, pp. 377 –
381).
Struve F. G. W. (1847), Etude d’astronomie stellaire. Petersburg.
--- (1852), Stellarum fixarum […] positiones mediae etc. Petersburg.
Struve O. (1842), Bestimmung der Constante des Präcession. Petersburg.
--- (1844), Same title. Mém. Acad. Imp. St. Pétersbourg, 6ème sér., Sci. math., phys.
et natur., t. 3 (5), pp. 17 – 24.
Struve P. B. (1918, in Russian), Who was the first to indicate that statistics can be
applied to philological studies? Izvestia Ross. Akad. Nauk, Ser. 6, t. 12, pp. 1317 –
1318.
Subbotin M. F. (1956, Russian), Astronomical and geodetic work of Gauss. In
Gauss. Memorial volume. Moscow, p. 243 – 310.
Süssmilch J. P. (1741), Die Göttliche Ordnung in den Veränderungen des
menschlichen Geschlechts, aus der Geburt, dem Tode und der Fortpflanzung
desselben. Berlin, 1765. Several subsequent editions.
--- (1758), Gedancken von dem epidemischen Krankheiten, etc. In Wilke, J., Editor,
Die Königliche Residenz Berlin und die Mark Brandenburg im 18. Jahrhundert.
Berlin, 1994, pp. 69 – 116.
Sylla E. D. (1998), The emergence of mathematical probability from the perspective
of the Leibniz – Jakob Bernoulli correspondence. Perspective on Sci., vol. 6, pp. 41
– 76.
Tait P. G. (1892), Poincaré’s Thermodynamics. Nature, vol. 45, pp. 245 – 246.
Takácz (Takács) L. (1967), Combinatorial Methods in the Theory of Stochastic
Process. New York.
--- (1969), On the classical ruin problems. Ibidem, vol. 64, pp. 889 – 906.
--- (1982), Ballot problems. In second edition of Kotz & Johnson (2006, vol. 1, pp.
183 – 188).
--- (1994), The problem of points. Math. Scientist, vol. 19, pp. 119 – 139.
Taqqi M. S. (2001), Bachelier and his times; a conversation with B. Bru. Finance
Stoch., vol. 5, pp. 3 – 32.
Tassi P. (1988), De l’exhaustif au partiel; un peu d’histoire sur le développement
des sondages. J. Soc. Stat. Paris, t. 129, No. 1 – 2, pp. 116 – 132.
Taton R., Wilson C., Editors (1995), General History of Astronomy, vol. 2B.
Cambridge.
Tee G. J. (2003), K. Pearson on atoms. Human Nature Rev., 3, p.302.

298
296
Thatcher A. R. (1957), A note on the early solutions of the problem of the duration
of play. Biometrika, vol. 44, pp. 515 – 518. Reprinted in E. S. Pearson & Kendall
(1970, pp. 127 – 130).
Thomas Aquinas (1952, English), Summa Theologica. Great Books, vols. 19 and
20. Subdivided in many tracts.
Thomson W., Tait P. G. (1867), Treatise on Natural Philosophy, vol. 1. Oxford.
[New York, 2002.]
Thorndike L. (1923), History of Magic and Experimental Science, vol. 1. New
York.
Tikhomandritsky M. A. (1898), Kurs Teorii Veroiatnostei (Course in Theory of
Probability). Kharkov.
Tikhomirov E. I. (1932, Russian), Directions to Russian meteorological stations in
the 18th century. Izvestia Glavnoi Geofisich. Obs., No. 1 – 2, pp. 3 – 12.
Timerding H. E. (1915), Analyse des Zufalls. Braunschweig.
Toaldo J. (1775, Italian), Witterungslehre für den Feldbau. Berlin, 1777.
--- (1777), Essai de météorologie. J. Phys., t. 10, pp. 249 – 279, 333 – 367.
Todhunter I. (1865), History of the Mathematical Theory of Probability from the
Time of Pascal to That of Laplace. New York, 1949, 1965.
Toomer G. J. (1974), Hipparchus on the distances of the sun and moon. AHES, vol.
14, pp. 126 – 142.
Truesdell C. (1975), Early kinetic theories of gases. AHES, vol. 15, pp. 1 – 66.
--- (1984), An Idiot’s Fugitive Essays on Science. New York. Collected reprints of
author’s introductions and reviews on history and philosophy of natural sciences.
Tsinger V. Ya. (1862), Sposob Naimenshikh Kvadratov (Method of Least Squares).
Dissertation. Moscow.
Tutubalin V. N. (1972), Teoria Veroiatnostei v Estestvoznanii (Theory of
Probability in Natural Sciences). Moscow.
--- (1973), Statisticheskaia Obrabotka Riadov Nabliudeniy (Stat. Treatment of
Series of Observations).Moscow.
Urlanis B. C. (1963, Russian), The tercentenary of population statistics. Uch. Zap.
po Statistike, vol. 7, pp. 150 – 160.
Usov N. A. (1867, Russian), Remark about the Chebyshev theorem. MS, vol. 2, pp.
93 – 95 of second paging. S, G, 5.
Vasiliev A. V. (1885), Teoria Veroiatnostei. Kazan.
Waerden van der B. L. (1968), Mendel’s experiments. Centaurus, vol. 12, pp. 275
– 288.
--- (1976, Russian), Die Korrespondenz zwischen Pascal und Fermat über
Wahrscheinlichkeitsprobleme. IMI, vol. 21, pp. 228 – 232.
Wallis J. (1685), Treatise of Algebra. London.
Weil A., Truesdell C., Nagel F. (1993), Der Briefwechsel von J. Bernoulli. Basel.
Weiling F. (1975), Mendel sowie die von Pettenkoffer angeregten Untersuchungen
von des Zusammenhanges Cholera- und Typhus- Massenerkrankungen mit dem
Grundwasserstand. Sudhoffs Arch., Bd. 59, pp. 1 – 19.
Wesley W. G. (1978), The accuracy of Tycho Brahe’s instruments. J. Hist. Astron.,
vol. 9, pp. 42 – 53.
White A. D. (1896), History of the Warfare of Science with Theology in
Christendom, vols. 1 – 2. London – New York, 1898. Many later editions up to
1955.
Whittaker E. T., Robinson G. (1924), Calculus of Observations. London, 1949.
Whittaker L. (1914). On the Poisson law of small numbers. Biometrika, vol. 10, pp.
36 – 71.
Whitworth W. A. (1867), Choice and Chance. New York, 1959, this being a reprint
of ed. of 1901.
Wilks S. S. (1962), Mathematical Statistics. New York.
Wilson C. (1980), Perturbations and solar tables from Lacaille to Delambre. AHES,
vol. 22, pp. 53 – 304.
--- (1984), The sources of Ptolemy’s parameters. J. Hist. Astron., vol. 15, pp. 37 –
47.
Winkler W. (1931), Ladislaus von Bortkiewicz. Schmollers Jahrbuch f.
Gesetzgebung, Verwaltung u. Volkswirtschaft im Deutschen Reich, 55. Jg, pp. 1025
– 1033.

299
297
Winslow C.-E. A. (1943), The Conquest of Epidemic Disease. New York – London,
1967.
Wishart J. (1927), On the approximate quadrature of certain skew curves with an
account of the researches of Bayes. Biometrika, vol. 19, pp. 1 – 38, 442.
Wittstein Ph. (1867), Mathematische Statistik. Hannover.
Wolf Abr. (1935), History of Science, Technology and Philosophy in the 16th and
17th Centuries. London, 1950.
Woolhouse W. S. B. (1873), On the philosophy of statistics. J. Inst. Actuaries, vol.
17, pp. 37 – 56.
Woytinsky W. S. (1961), Stormy Passage. New York.
Yamazaki E. (1971), D’Alembert et Condorcet: quelques aspects de l’histoire du
calcul des probabilités. Jap. Studies Hist. Sci., vol. 10, pp. 60 – 93.
Yarochenko S. P. (1893a, Russian), On the theory of the method of least squares.
Zapiski Novoross. Univ., vol. 58, pp. 193 – 208 of second paging.
--- (1893b), Sur la méthode des moindres carrés. Bull. Sciences Math., sér. 2, t. 17,
pp. 113 – 125.
Yastremsky B. S. (1913, Russian), The law of sufficiently large numbers applied to
estimate the stability of statistical series. Izbrannye Trudy (Sel. Works). Moscow,
1964, pp. 13 – 32.
You Poh Seng (1951), Historical survey of the development of sampling theories
and practice. J. Roy. Stat. Soc., vol. A114, pp. 214 – 231. Reprinted: M. G. Kendall
& Plackett (1977, pp. 440 – 458).
Youshkevitch A. A. (1974), Markov. Dict. Scient. Biogr., vol. 9, pp. 124 – 130.
Youshkevitch A. P. (1986, Russian), N. Bernoulli and the publication of J.
Bernoulli’s Ars Conjectandi. Theory of Probability and Its Applications, vol. 31,
1987, pp. 286 – 303.
Yule G. U. (1900), On the association of attributes in statistics. PT Roy. Soc, vol.
A194, pp. 257 – 319. Also in Yule, Statistical Papers. London, 1971, pp. 7 – 69.
Yule U., Kendall M. G. (1937), Introduction to Theory of Statistics. London, 1958.
Zabell S. L. (1988a), The probabilistic analysis of testimony. J. Stat. Planning and
Inference, vol. 20, pp. 327 – 354.
--- (1988b), Buffon, Price and Laplace: scientific attribution in the 18th century.
AHES, vol. 39, pp. 173 – 181.
--- (1989), The rule of succession. Erkenntnis, Bd. 31, pp. 283 – 321.
Zach F. X. von (1813), Sur le degré du méridien. Mém. Acad. Imp. Sci., Littérature,
Beaux-Arts Turin pour 1811 – 1812. Sci. math. et phys., pp. 81 – 216.
Zermelo E. (1900), Über die Anwendung der Wahrscheinlichkeitsrechnung auf
dynamische Systeme. Phys. Z., Bd. 1, pp. 317 – 320.
Zoch R. T. (1935), On the postulate of the arithmetic mean. Annals Math. Stat., vol.
6, pp. 171 – 182.

Three sources, additionally recalled/found:

Bellhouse D. R. (2011), Abraham De Moivre. Boca Raton, FL.
Cardano G. (1961), Book on Games of Chance. Mineola, NY, 2015.
Lehmann E. L. (2011), Fisher, Neyman and the Creation of Classical Statistics.
New York.

300
298
 Index of Names
This Index does not cover the General Literature and my own name is also absent.
The numbers refer to subsections rather than to pages.

Aaboe, A. 1.1.4
Abbe, C. 9A.1.3
Abbe, E. 9B, 15.2
Achenwall, G. 6.2.1, Notes to chapter 10
Adrain, R. 9A.1.3, Notes to chapter 9
Afanasieva – Ehrenfest, T. A. 7.1, 14.2
Airy G. B. 10.8.4
Al-Biruni (973 – 1048) Preface, 1.1.4, 1.2.4, Notes to chapter 1
Al-Khazini, 1.1.4
Ambartsumian, R. V. 12
Ames, W. Notes to chapter 1
Anchersen, J. P. 6.2.1, 10.7
Anderson, O. 15.1.2, 15.1.3, 15.2, Notes to chapter 15
Andersson T. 10.8.2
André, D. 11.1
Andreev, A. V. 14.5,
Andreev, K. A. 14.3, Notes to chapter 14
Arago, F. Notes to chapter 8
Arbuthnot, J. 2.2.4, 6.2.2, Notes to chapter 7
Archimedes (ca. 287 – 212 BC), 1.1.4, 7.4
Argelander F., 10.8.4
Aristarchus (end of 4th c. – first half of 3rd c. BC), 1.1.4
Aristotle (384 – 322 BC) 0.1, 1.1.1, 1.1.2, 1.1.3, 1.1.5, 1.2.4, Notes to chapter 1,
2.1.1, 3.2.4, 10.8.2, 11.2
Arnauld, A. 2.1.1, 2.1.2, 2.1.4, Notes to chapter 2, 3.1.2, 3.2.2, 3.3.1, Notes to
chapter 3
Asarevich, P. D. 15.1.2
Augustinus, St. (354 – 430) Notes to chapter 1
Austin, E. 0.5, 6.2.4, 10.8.3
Azarevich P. D. 15.1.2
Babbage, C. 1.1.4, 10.5, 10.8.2
Bachelier, L. 14.2
Baer, K. 6.1.3, 10.8.2
Baily, F. 1.2.2
Barbier, E. 12
Barnard, G. A. 5.1
Barnett N. 9A5
Bartenieva, L. S. 10.1
Bateson, W. 15.2
Bauer, R. K. 15.3
Bayes, T. 0.1, 2.2.3, chapter 5, 6.1.6, 7.1, 7.2, Notes to chapter 7, 10.4, 11.1, 11.2,
13.2
Beer, A. 1.1.4
Beilis, M. 14.3, 14.5
Bellhouse, D. R. Notes to chapter 1, 2.2.4, 3.3.4
Belvalkar, S. K. Notes to chapter 1
Benjamin, M. 10.8.4
Bentley, R. 2.2.3
Berggren, J. L. 1.1.4
Bernoulli, Daniel Notes to chapter 3, 6.1.1, 6.1.2, 6.1.4 – 6.1.6, 6.2.2 – 6.2.4, 6.3.1,
Notes to chapter 6, 7.1, 9C, 10.3, 10.4, 10.8.4, 14.2
Bernoulli, Jakob 0.1, 1.1.5, 2.1.1, 2.1.2, 2.2.2, 2.2.4, Notes to chapter 2, 3, 4.3, 4.4,
5.2, 7.1, 8, 8.6, 8.7, 9C, 10.2, 10.3, 10.7, 11.2, 13.2, 14.1
Bernoulli, Johann I 3.1.2
Bernoulli, Johann II 6.2.3
Bernoulli, Johann III 6.3.1

301
299
Bernoulli, Niklaus Notes to chapter 1, 2.1.2, 2.2.4, Notes to chapter 2, 3.3.2, 3.3.3,
3.3.4, 6.1.1, 8.9.2, 9C
Bernstein, S. N. 0.5, 11.2, 13.1 – 13.3, 14.1, 14.2, 14.4, 15.2, Notes to chapter 15
Bertillon, Alph. 10.6
Bertrand, J. 0.3, 2.1.1, 6.1.1, 6.1.2, 6.1.6, 8.7, 9A.2, 9A4, 10.2, 10.5, 10.8.4, 11,
11.1, 11.2, Notes to chapter 11, 12, Notes to chapter 12
Besikovitch, A. S. 14.3
Bessel, F. W. Preface, 6.3, 8.7, 9A.1.4, 9A.3 – 9A.6, 9C, Notes to chapter 9
Bienaymé, I. J. 7.1, 7.2, 8.7, 9A.4, 9A.6, 10.2, 10.3, 10.8.4, 11.1, 13.1, 14.1, 14.5,
15.1
Biermann, K.-R. Notes to chapter 1, Notes to chapter 2, 9A, 9A.1.4
Biot, J. B. 6.3.2, 10.8.3
Birg, S. 6.2.2
Black, W. 6.2.3, Notes to chapter 6
Blanford, H. F. 10.6
Block, M. 10.7
Bobynin, V. V. 10.4
Böckh, R. Notes to chapter 2
Bohlmann G. 9A.4
Boiarsky, A. Ya. Notes to chapter 15
Boltzmann, L. 7.4, 9A.4, 10.8.2, 10.8.5, 12, 15.2
Bolyai, J. 9A.1.4
Bolyai, W. 9A.1.4
Bomford, G. 9A.3
Bond, G. P. 9A.5
Bonpland, A. J. A., 10.8.2
Boole, G. 8.1, 10.8.4, Notes to chapter 13, Notes to chapter 14
Borel, E. 2.2.2, 12
Bortkiewicz, L. von, 3.1.2, 3.2.3, Notes to chapter 3, 8.7, 10.4, 10.7, 11.1, 11.2,
14.2, 14.5, 15.1.1, 15.1.2, 15.1.3, Notes to chapter 15
Boscovich, R. J. 1.1.4, 1.2.3, 3.2.3, 6.3, 6.3.2, Notes to chapter 6, 7.2, 7.3, 9A.2,
10.8.4, 10.8.5
Bouillaud, J. 10.8
Bowditch, N. 7.2, 7.4
Boyle, R. 1.2.2, 10.9.3
Bradley, J. 1.2.2, 9C
Brahe, T. 0.3, 0.4, 1.2.2, 1.2.4, Note to chapter 1, 6.3, Notes to chapter 6, 10.7
Bredikhin, F. A. 10.8.4, 10.9.3, 14.2
Brendel, M. 9A.1.4
Brown, R. 7.1, 14.2
Brownlee, J. 10.8.1
Bru, B. Notes to Chapt. 1, Notes to chapter 6, 8, 10.2, 11.1
Bruns, H. 14.2
Brush, S. G. 10.8.5, 14.2
Bühler, G. Notes to chapter 1
Buffon, G. L. L. 3.3.4, 6.1.4, 6.1.6, Notes to chapter 6, 7.1, 7.4, 10.4, Notes to
chapter 13
Bugaev, N. V. Notes to chapter 14
Bull, J. P. 8.9.2
Buniakovsky, V. Ya. 1.2.3, 8.7, 10.4, Notes to chapter 10
Burov, V. G. 1.1.1, Notes to chapter 1, Notes to chapter 2
Buys Ballot, C. H. D. 10.8.3
Byrne, E. F. 1.1.5, Notes to chapter 1
Caesar 9A.4
Campbell, L. 10.8.5
Cantor, G. Notes to chapter 2
Cardano, G. 3.2.3
Cauchy, A. L. 8.6, 8.9, 10.1, 11.2, 13.1
Celsus, A. C. (ca. 25 BC – ca. 50 AD) 0.4
Chadwick E. 10.8.1
Chaitin, G. J. 1.1

302
300
Chamberlayne, E. 6.2.1
Chaplygin, S. A. 14.3
Chapman, S. 2.1.4, Notes to chapter 2
Charlier, C. V. L. 10.2, 15.1.3, 15.2
Chatterjee S. K. 6.3.3
Chaufton, A. 2.1.3
Chauvenet, W. 8.9.2
Chebotarev, A. S. Notes to chapter 1
Chebyshev, P. L. 0.1, Notes to chapter 0, 3.2.3, 5.1, 6.3.2, 7.1, 7.2, 7.4, 8.7, 9A.4,
9B, 10.2 – 10.4, 10.8.4, 11, 11.2, chapter 13, 14.1, 14.2, 14.4, 14.5
Chekhov, A. P. 0.5
Chetverikov, N. S. 15.1.3
Chirikov, M. V. Preface, Notes to chapter 14
Chrysippus (208 or 205 – 280) Notes to chapter 1
Christie, A. 0.5
Chuprov, A. A. Preface, 6.1.3, 8.4, 8.7, 10.3, 10.5, 10.7, 10.8.3, Notes to chapter 10,
10.5, 10.7, 11.2, 14.1, 14.2, Notes to chapter 14, 14.5, 15.1.1 – 15.1.3, 15.2, 15.3,
Notes to chapter 15
Cicero M. T. (106 – 43 BC) 1.1.1, 1.2.4, Notes to chapter 1, 3.2.1, Notes to chapter
3
Cioffari, V. 1.1.1
Clairaut, A. C. 10.9.1
Clausius, R. 10.8.4
Clerke, Agnes M. 1.2.3
Clifford, W. K. 15.2
Cohen, J. Notes to chapter 1
Colebrooke, H. T. Notes to chapter 1
Columbus Chr. 2.1.4
Commelin, C. 2.1.3
Comte, A. 10.8
Condamine, C. M. de la 6.2.3, 6.3.1
Condorcet, M. J. A. N. 2.2.3, 3.2.3, 3.3.4, 6.1.5, 6.2.3, 6.2.4, Notes to chapter 6, 7.2,
8.9.1, 10.8.1, 11.1, 11.2
Confucius (ca. 551 – 479 BC) 1.1.1
Cook, A. H. Notes to chapter 2
Coolidge, J. L. Notes to chapter 9
Copernicus, N. 1.2.4
Cornfield, J. 5.1
Cotes, R. Notes to chapter 0, 6.3.1
Cotte, L. 6.2.4, 10.8.3
Cournot, A. A. Notes to chapter 1, Notes to chapter 3, 5.1, 6.1.3, 7.1, 8.1, 8.6, 8.7,
8.9.2, 10.2, 10.3, 10.4, 10.5, 10.7, 10.8.4, Notes to chapter 10, 11.1, 11.2, 12, 15.1
Courtault, J.-M. 14.2
Couturat, L. Notes to chapter 3
Cramer, G. 3.3.4, 6.1.1
Cramér, H. 7.4, 9A.3, 9A.4, Notes to chapter 9, 15.2, 15.3
Crofton, M. W. 12
Cubranic, N. 6.3.2
Czuber, E. 4.2, Notes to chapter 5, 9A4, 10.2, 10.7, 12
D’Alembert, J. Le Rond, Notes to chapter 1, 6.1.2, 6.1.3, 6.1.5, 6.2.3, Notes to
chapter 6, 7.1, 10.3, 10.8, Notes to chapter 13
D’Amador, R. 10.8
Danilevsky, N. Ya. 6.1.3
Darboux G. 11.1
Darwin, C. 10.5, 10.6, 10.8.2, 10.8.5, 12, Notes to chapter 12, 15.2, Notes to chapter
15
Davenport, C. B. 15.2
David, F. N. 1.2.3, 14.2
David, H. A. 2.2.4, 3.3.3, 5.1, 9B, 10.9.2, 14.2
Davidov, A. Yu. 8.1, 8.2, 8.7, 8.9.2, 10.4, 10.8.5
Daw, R. H. 6.2.2

303
301
Dawid, Ph. 11.2
Daxecker, F. 1.2.3
DeCandolle Alph. 10.8.2
DeCandolle, Aug. P. 10.8.2
Delambre, J. B. J. 9C, 10.7, Notes to chapter 10
De Mére, C. 2.2.3, Notes to chapter 2
Demidov, P. G. 13.1
Democritus (ca. 460 – ca. 370 BC) Notes to chapter 1
De Moivre, A. 0.1, 2.1.4, 2.2.2, 2.2.3, 3.2.3, 3.3, 3.3.3, 3.3.4, Notes to chapter 3,
chapter 4, 5.2, 5.3, 6.1.1, 6.1.6, 6.3.1, 7.1, 7.4, 9C, 10.8.5, 11.1, 13.2, 14.1
De Montessus, R. 12
De Morgan, A. 4.4, 8.1, Notes to chapter 8, 10.4
De Morgan, Sophia Notes to chapter 8
Derham, W. 2.2.3, Notes to chapter 2
Descartes, R. 2.1.2, 2.2.2, 7.4
De Solla Price, D. J. 1.1.4
Desrosières, A. Notes to Preface
De Vries H. 10.8.2
De Vries, W. F. M. Notes to chapter 10
De Witt, J. 2.1.3, 2.2.2, Notes to chapter 2
Dietz, K. 6.2.3, 10.4
Dirac, P. A. M. 7.2, 7.4, Notes to chapter 8
Dirichlet, P. G. L. 7.1, 7.4, 8.6, 9B, 13.2
Doig, A. G. Notes to chapter 15
Donkin, W. F. Notes to chapter 14
Doob, J. L. 7.4
Doppler, C. 10.8.4
Dorfman, Ya. G. 7.3
Dormoy, E. 11.1, 15.1.1
Dorsey, N. E. 10.8.4, 10.9.3
Double, F. J. 7.2, 8.9, 10.8.1
Dove, H. W. 10.8.3
Dreyfus, A. 11.2
Drinkwater Bethune, J. E. 10.4
Dufau, P. A. 10.7
Dunnington G. W. 9A.1
Dupâquier, J. Note to Preface
Dupâquier, M. Note to Preface
Du Pasquier, L. G. 2.1.3
Dutka, J. Notes to chapter 3, 9A.1.1, 9A.1.3, 9A.1.4, 14.2
Ebbinghaus, H. 10.9.2
Eddington, A. S. 9A4, 10.5, 10.8.4
Edgeworth, F. Y. 10.9.2, 15.2
Edwards, A. W. F. 2.2.4, 3.3.3, 5.1, 14.2
Eggenberger, J. 4.4
Ehrenfest, P. 7.1, 14.2
Ehrenfest, T., see Afanasieva-Ehrenfest T.
Eisenhart, C. Preface, 1.2.4, 6.3.3, 7.2, 9A4, 9A.6, 10.8.4, 10.9.3, 15.2, 15.3
Ekeland, I. Notes to chapter 1
Eliseeva, I. I. 6.2.1
Elkin, W. L. 10.8.4
Ellis R. L. 7.2
Emeliakh, L. I. 14.3
Encke, J. F. 9A.2, 9A4, Notes to chapter 9
Eneström G. 2.1.3
Enko, P. D. 10.4, 10.8.1
Epicurus (342 or 341 – 271 or 270 BC) Note to chapter 1
Erathosthenes (ca. 276 – 194 BC), 1.1.4
Erdélyi, A. 11.2
Erismann, F. F. 10.8.1
Erman, A. Notes to chapter 9

304
302
Ermolaeva, N. S. 13.2, 14.2
Euler, L. Preface, 2.1.4, 6.1.2, 6.2.2, 6.3.1, 6.3.2, Notes to chapter 6, 7.1, 7.2, 9A1,
9A1.1, 9A.1.4, Notes to chapter 9, 10.4, 13.2
Faraday, M. 10.8.3, Notes to chapter 10, 15.2
Farebrother, R. W. Preface
Farr, W. 10.8.1
Fechner, G. T. 6.1.1, 7.1, 10.9.2
Fedorovitch, L. V. 2.1.4
Feller, W. 7.1, 11.1
Fermat, P. 0.1, 2.2.1, 2.2.2
Filliben 1.2.4
Fisher, R. A. Preface, 0.2, Notes to chapter 1, 5.1, 6.1.6, 6.3.3, 9A.4, 9B, 10.7.2,
10.8.2, 10.8.4, 10.7.4, 14.3, 15.2, 15.3
Flamsteed, J. 1.2.2
Fletcher, A. 14.1
Florensky, P. A. 14.5
Forsythe, G. E. 9A.5
Fortunatov, A. 15.1.2
Fourier, J. B. J. 0.3, 6.3.3, 7.1, 7.3, 7.4, 8, 9, 9A, 10.2, 10.5, 10.8, 11.2
Franklin, J. 1.1.1, 1.1.5, Notes to chapter 1, 2.1.2, Notes to chapter 2, 3.2.1
Fraenkel, A. Notes to chapter 2
Fréchet, M. 10.5
Freud, S. 10.9.2
Freudenthal, H. 2.2.4, 3.3.4, 8.9.2, 10.1, 10.5, 13.1
Fries, J. F. 7.1, 10.3
Fuss, N. I. 7.1
Fuss, P. N. 6.1.1
Galen, C. (129 – 201?) 1.1.3, 11.2
Galilei, G. 1.1.4, 1.2.3, Notes to chapter 6, 7.4
Galitzin, B. B. 14.2
Galle, A. 9A.1.4
Galton, F. 10.2, 10.5, 10.6, 10.7, 10.8.1, 10.8.2, 10.9.2, 15.2
Gani, G. 2.2.3, 10.4
Garber, D. 2.1.2, 3.2.1
Garnett, W. 10.8.5
Gastwirth, J. L. 11.2
Gatterer, J. C. 10.7
Gauss, C. F. Preface, Notes to Preface, 0.3, Notes to chapter 1, 3.1.2, 5.1, 5.3, 6.3,
6.3.1 – 6.3.3, Notes to chapter 6, 7.2, 7.4, 8.8, chapter 9, 10.1, 10.7, 10.8.4, 10.9.1,
10.9.2, Notes to chapter 10, 11.1, 11.2, Notes to chapter 11, 13.2, 14.1, 14.2, 14.4,
14.5, 15.1.1, 15.2, Notes to chapter 15
Gavarret, J. 8.9.2, 10.8
Gerardy, T. 9A.1.4, 9A.5
Gerling, C. L., Preface, 9A.5
Gikhman, I. I. 13.3
Gingerich, O. 1.1.4
Gini, C. Notes to chapter 2, 3.1.2
Gnedenko, B. V. 0.1, 7.4, 8.4, 10.2, Notes to chapter 10, 13.3, 14.3, Notes to chapter
14, 15.3
Goebbels, J. P. Notes to chapter 15
Goethe, I. W. 9C
Goldstein, B. R., 1.1.4
Golitsin I. 6.2.1
Gordevsky, D. Z. Notes to chapter 14
Gower, B. 3.2.3
Gowing, R. 6.3.1
Gram, J. P. 10.2
Graunt, J. 0.4, 1.1.2, 2.1.4, 2.2.2 – 2.2.4, Notes to chapter 2, 3.1.1, 3.2.2, 3.2.3, 6.2.2,
6.2.3, 10.7, Notes to chapter 10
‘sGravesande, W. J. (1688 – 1742), 2.2.4
Greenwood, M. 2.1.4, 10.8
Gridgeman, N. T. 7.1

305
303
Grodsensky, S. Ya. 14.3, Notes to chapter 14
Gul’kevich,K. N. 15.1.3
Gusak, A. A. 6.3.2
Guy, W. A. 10.8
Hagstroem, K.-G. Notes to chapter 3
Hald, A. Preface, 1.2.3, 2.1.3, 2.1.4, 2.2.1, 2.2.2, 2.2.4, 3.3.3, 3.3.4, Notes to chapter
3, 4, 4.1 – 4.3, Notes to chapter 4, 5.1, 6.3.3, 7.1, 8.6, 8.7, 9A1.2, 9A4, 9B, Notes to
chapter 9, 10.2, 10.8.4, 15.2, 15.3
Halley, E. 2.1.4, Notes to chapter 2, 3.2.3, 4.2, 10.7, 10.8.3
Halperin, T. Notes to chapter 3, 8.1, Notes to chapter 13
Harter, H. L. 9A4
Hartley, D. 5.3
Harvey, W. Notes to chapter 1
Hasofer, A. M. 1.1.2
Haushofer, D. M. 3.2.3
Heesterbeek, J. A. P. 6.2.3
Hegel, G. W. F. 1.1.1
Hellman, C. D. 1.2.4
Helmert, F. R. 0.3, 3.2.3, 9A.2 – 9A.4, 9B, Notes to chapter 9, 11.1, 15.2
Hendriks, F. 2.1.3
Henny, J. 3.3.3
Henry, M. Ch. 6.1.5
Hermite, C. 7.1, 13.1
Herschel, Caroline 9C
Herschel, J. Notes to chapter 8, 9A.1.3
Herschel, W. Notes to chapter 6, 9C, 10.8.4
Hertz, H. 10.8.5
Heuschling, X. 10.5
Heyde, C. C. 8.7, 10.1, 10.2, 11.1, 11.2, 13.1, 15.3
Hill, D. 10.8.4
Hipparchus (180 or 190 – 125 BC) 1.1.4
Hippocrates (460 – 377 or 356 BC) 1.1.3
Hobbes, T. Notes to chapter1
Hogan, E. R. Notes to chapter 9
Hohenner, H. 9B
Holden, E. S. 10.8.4
Hoskin, M. A. 9C
Hostinský, B. 7.1
Hoyrup, J. Notes to chapter 1
Huber, D. 9A.1.1
Hudde, J. H. 2.1.3
Hulme H. R. 10.8.4
Humboldt, A. 1.2.3, Notes to chapter 9, 10.3, 10.5, 10.7, 10.8.2, 10.8.3
Huygens, C. 2.1.1, 2.1.3, 2.1.4, 2.2.2, 2.2.3, Notes to chapter 2, 3.1.2, 3.3.2, Notes to
chapter 3, 4.3, 6.1.2, 6.3.1, 8.4, 10.4, 10.8.5
Huygens, L. 2.2.2
Idelson, N. I. 7.2, 9A.4, 14.1, 14.2
Imshenetsky, V. G. 14.3
Irwin, J. O. 15.2
Isotov, A. A. Notes to chapter 6
Isserlis, L. 14.1
Ivory, J. Notes to chapter 6, 9A.6, 10.9.1
Jamal-ud-Din, 12th century, 1.2.3
Jenner E. 6.2.3, 10.8.1
Jennings, J. 1.2.3
John of Salisbury (1115 or 1120 – 1180) 3.2.1
Johannsen, W. 10.8.2
Jongmans F. J. 11.1
Jorland, G. Notes to chapter 3
Junkersfeld, J., 1.1.1
Juskevic, A. P., see Youshkevitch, A. P.

306
304
Kac, M. 9A.1.3
Kagan, V. F. 13.3
Kamke, E. 7.4
Kant, I. 1.2.4, Note to chapter 1, 7.3, Notes to chapter 7, 15.2
Kapteyn, J. C. Notes to chapter 9, 10.6, 10.8.4, 15.2
Karn, M. Noel, 6.2.3
Kaufman, A. A. 10.7
Kendall, D. G. 10.2
Kendall, M. G. (Sir Maurice) Note to chapter 1, 2.1.1, 2.1.4, 9B, 10.9.3, 12,
15.2, Notes to chapter 15
Kepler, J. Preface, 1.1.2, 1.1.4, 1.2.2, 1.2.3, 1.2.4, Notes to chapter 1, 2.1.4, 2.2.3,
Notes to chapter 2, 3.2.3, 3.2.4, 6.3.2, Notes to chapter 6, 7.3, Notes to chapter 9,
10.7
Keynes, J. M. 3.1.2, 7.4, Notes to chapter 14
Khinchin, A. Ya. 10.8.5, 11.2
Kiaer, A. N. 10.8.4
Kington, J. A. 6.2.4
Kirillov I. K., 6.2.1
Klein, F. 13.3
Knapp, G. F. 3.2.3, 10.5, 10.7
Knauer, K. 10.4
Knies, C. G. A. 6.2.1
Knott, C. G. 10.8.5
Kohli, K. 1.1.5, 2.1.1, 2.1.3, 2.2.2, Notes to chapter 2, 3.1.2, 3.3.2, 3.3.4, Notes to
chapter 3
Koialovitch, B. M. 14.2
Kolmogorov, A. N. Preface, 0.1, 0.2, 1.1, 2.1.4, 7.2, 7.4, 9A4, 9B, 10.8.2, 10.8.5,
13.1, 13.3, 14.1, 14.2, 15.1.2, 15.2
Koopman, B. O. 3.2.1
Köppen, W. 10.8.3, Notes to chapter 10
Körber, H.-G. 10.8.3
Korteweg, D. J. 2.2.2
Kowalski, M. A. 10.8.4
Krasovsky, F. N. Notes to chapter 6, 9A.1.3, 9B, 10.9.1
Krein, M. G. 13.1, 14.2
Krengel, U. 9A.4
Krizenecky, J. Notes to chapter 10
Kronecker, L. Notes to chapter 2, 9A4, 13.2
Kruskal, W. 1.1.5, 9B, Notes to chapter 9
Krylov, A. N. 9A.1.3, 13.2
Ku, H. H. Note to Chapt. 0
Kuzmin, R. 9A
Lacroix, S. F. 10.4
Lagrange, J. L. 4.3, Notes to chapter 4, 6.3.1, Notes to chapter 6, 7.1, Notes to
chapter 7, 9A.4, 9A.5, Notes to chapter 10
Lamarck, J. B. Notes to chapter 1, 6.2.4, 10,5, 10.8.3
Lambert, J. H. 0.3, 6.1.3, 6.2.2, 6.2.4, 6.3, 6.3.1, 6.3.2, Notes to chapter 6, Notes to
chapter 9, 9C, 10.3
Lamont, J. 10.8.3
Langevin, P. 10.8.5
Laplace, P. S. Preface, 0.1, 0.3, 1.1.4, Note to chapter 1, 2.1.1, 2.2.3, 4.3, 4.4, Notes
to chapter 4, 5.1, 5.2, 6.1.1, 6.1.5, 6.1.6, 6.2.3, 6.3, 6.3.1, 6.3.2, Notes to chapter 6,
7, 8, 8.3, 8.5, 8.6, 8.8, 8.9.1, Notes to chapter 8, 9A, 9A.1.3, 9A.1.4, 9A.6, 9A.2,
9A.5, Notes to chapter 9, 9C, 10.1, 10.3 – 10.5, 10.7, 10.8, 10.8.2, 10.8.4, 10.8.5,
11.1, 11.2, 13.1, 13.2, Notes to chapter 13, 14.1, 14.2, 14.4, 14.5
La Placette, J. 10.3
Latyshev, V. A. Notes to chapter 13
Laurent, H. 10.4, 12
Le Cam, L. 11.2
Legendre, A. M. Notes to chapter 6, 7.2, 8.8, 9A.1.1 – 9A.1.4, Notes to chapter 9,
14.5

307
305
Lehmann, E. L. 6.3.2, 9A.2
Lehmann-Filhès, R. 10.8.4
Leibniz, G. W. 2.1.1 – 2.1.4, Notes to chapter 2, 3.1.2, 3.2.2, 3.3.1, Notes to chapter
3, 6.1.3, 6.2.1, 6.2.3, 6.2.4, 7.4, 8.1, 10.8, 10.8.1, Notes to chapter 10, 11.2
Lenin, V. I. 15.2, Notes to chapter 15
Levi ben Gerson (1288 – 1344) 1.1.1, 1.1.4, Note to chapter 1, 2.1.1, Notes to
chapter 3
Lévy, M. 10.8.1
Lévy, P. 0.3, 7.4, 13.3
Lewis, T. 9A5
Lexis, W. 10.2, 10.7, 11.1, 15.1, 15.1.1, 15.1.2, 15.1.3, 15.2
Liapunov, A. M. 0.1, 8.2, 9C, 10.2, 11, 13.1 – 13.3, 14, 14.1, 14. 2, 14.4, 14.5, Notes
to chapter 14
Libri-Carrucci, G. B. I. T. 8.9
Liebermeister, C. 8.9.2
Lindeberg, J. W. 14.4, Notes to chapter 14
Linnik, Yu. V. 9A.1.3, 9B, 10.1, 14.1, 14.2
Lippmann, G. Notes to chapter 11
Lipschitz, R. 9A.3
Lloyd, G. E. R. 1.1.4
Lluilier, S. A. Notes to chapter 3
Lobachevsky, N. I. 13.3
Lockyer, J. N. 10.6
Louis, P. C. A. 10.8
Loveland, J. Notes to chapter 6
Lubbock, W. 5.1, 10.4
Lucretius (I century BC) Note to chapter 1
Lueder, A. F. 10.7
Lully, R. (ca. 1235 – ca. 1315) 6.1.3
Lüroth J. 15.2
Mach, E. 15.2
Maciejewski, C. 3.2.3, 13.3
MacLaurin, C. 7.1
Maennchen, Ph. 9A.5
Mahalanobis, P. C. 15.2, 15.3
Maievsky, N. V. 11.2, 13.3
Maimonides M. (1135 – 1204) 1.1.1, 1.1.2, 1.1.5, Note to chapter 1
Maire, C. 6.3, 6.3.2
Maistrov L. E. 1.2.3
Makovelsky, A. O. 1.1.1
Malfatti, F. 7.1
Malthus, T. R. 6.2.2
Maltsev, A. I. 9B
Mandryka A. P. 13.3
Mann, W. Notes to chapter 10
Mansion, P. 10.5
Mariotte, E. 1.2.2, 10.3
Markov, A. A., Sr Preface, 0.1, Notes to chapter 1, 3.2.3, Notes to chapter 3, 4.4,
6.1.1, 6.3.3, 7.1, 7.2, 7.4, 8.6, 8.7, 9A.6, 9C, 10.2 – 10.4, 10.7, 10.8.4, 11, 11.2,
Notes to chapter 11, 13.1 – 13.3, 14, 14.1, 14.5, Notes to chapter 14, 15.1.2, 15.1.3,
15.2, 15.3, Notes to chapter 15
Markov, A. A., Jr 14.3
Marsden, B. G. 9A.1.4
Matthiesen, L. Notes to chapter 1
Maupertuis, P. L. M. 1.1.5, 6.1.1, Notes to chapter 6, 7.3
Maxwell, J. C. 7.4, 9A.1.3, 9A.6, 10.8.5, 11.2, 12
May, K. O. 9A.1.4
Mayer, T. 6.3.2
Mayr, G. von 10.7
Meadowcroft, L. V. 7.2
Meldrum, C. 10.6
Mendel, J. G. 10.8.2, Notes to chapter 10, 15.2

308
306
Mendeleev, D. I. 1.2.2, 9A.3, 10.8.3, 10.8.4, 10.9.3
Mendelsohn, M. 6.2.3
Merian, P. 9A.1.1
Merrington, M. 15.2
Meyer, H. 10.8.3
Michell, J. 6.1.6, 10.8.4, 11.1
Mill, J. S. 3.1.2, 8.9.1, 11.2
Mises, R. von Notes to chapter 5, 5.2, 6.3.3, 7.4, 10.2, 10.9.2, 14.1, 14.4
Molina, E. C. 7.1, 7.4
Mondesir, E. 8.4
Montmort, P. R. 2.1.1, 3.2.3, 3.3.3, 3.3.4, 4.3, 6.1.2, 6.3.1, 8.3
Montucla, J. E. 10.4
Moore, P. G. 10.2
Moran, P. A. P. 12
Morant, G. M. 15.2
Moreau de Jonnès, Notes to chapter 10
Mouat, F. J. 10.5
Muncke, G. W. 6.2.4, 10.8.3
Nansen, F. 15.1.3
Needham, J. 1.1.4
Nekrasov, P. A. 8.6, 10.2, Notes to chapter 13, 14, 14.1 – 14.5, Notes to chapter 14,
14.5, 15.1.2
Neugebauer, O. 1.1.4
Neumann, C. Notes to chapter 2
Newcomb, S. 1.1.4, 6.1.6, 9A.5, 9A3, 9A.6, 9C, 10.8.4, 10.9.3, 15.1.1, 15.2
Newton, I. Preface, 1.1.2, 1.1.5, 2.1.4, 2.2.3, 3.2.3, 4.3, 6.1.6, 6.3.1, Notes to chapter
6, 7.1, 7.3, 7.4, 11.2
Newton, R. R. 1.1.4
Neyman, J. 5.1, 14.2, 14.3, Notes to chapter 15
Nicole, P. 2.1.1, 2.1.2, 2.1.4, Notes to chapter 2, 3.1.2, 3.2.2, 3.3.1, Notes to chapter
3
Nicomachus of Gerasa (ca. 100 BC) 1.1.1
Nieuwentit, B. (1654 – 1718) 2.2.3, Notes to chapter 2
Nightingale, F. 2.2.3, 10.8.1
Novikov, S. P. 13.3
Novoselsky, S. A. 10.4
Occam W. 1287 – 1347, 1.1.5
O’Donnell 2.1.3
Ogorodnikov, K. F. 10.8.4
Olbers, H. W. Preface, 9A.1.3, 9A.1.4, 9A.2, 9A4, 9A.5, 9C Notes to chapter 9,
10.9.1
Ondar, Kh. O. 8.9.2, 10.4, 14.1, 14.2, Notes to chapter 14, 15.1.2, 15.1.3, 15.2
Ore, O. Notes to chapter 2, 3.2.3
Oresme, N. (ca. 1323 – 1382) 3.2.4, Notes to chapter 3
Ostrogradsky, M. V. 7.1, 10.4, Notes to chapter 10
Öttinger, L. 10.4
Paevsky, V. V. 6.2.2
Pannekouk, A. 1.2.4
Pareto, W. 15.1.2
Pascal, B. 0.1, 2.1.1, 2.2.1, 2.2.2, Notes to chapter2, 3.3.1, 4.1
Paty, M. Notes to chapter 6
Pearson, E. S. 0.5, 2.2.3, Notes to chapter 6, 15.2, 15.3, Notes to chapter 15
Pearson, K. 0.2, 0.4, 0.5, Note to chapter 1, 2.1.3, 2.1.4, 2.2.3, 2.2.4, Notes to
chapter 2, 3.2.3, 4.2, 4.3, Notes to chapter 5, 6.1.4, 6.2.1 – 6.2.3, 6.3.1, Notes to
chapter 6, 7.1, 7.2, 9A.4, 9A.6, Notes to chapter 9, 10.5, 10.6, 10.8.1, 10.8.3, 10.8.4,
10.9.2, Notes to chapter 10, 14.1, 14.2, 15.1.1, 15.2, 15.3, Notes to chapter 15
Peter the Great Note to chapter 1
Peters, C. A. F. 6.3.2, 9B, 10.9.2
Petruszewycz, M. 14.2
Pettenkofer, M. 10.8.1
Petty, W. 1.1.2, 2.1.1, 2.1.4, 7.3, 10.8, Notes to chapter 10
Pfanzagl, J. Preface, 6.2.2, 15.2

309
307
Picard, J. 6.3.1
Pickering, E. 10.8.4
Pirogov, N. I. 10.8.1
Pitman, E. J. G. 7.2
Plackett, R. L. Notes to chapter 1, 14.2
Plato (429 – 348 BC) 1.1.1, 3.2.4, Notes to chapter 3
Plato, J. von 11.2
Ploshko V. G., 6.2.1
Poe, E. 0.5
Poincaré, H. Notes to chapter 0, 1.2.4, 6.1.3, 6.3.1, 7.4, 8.9.1, 9A.6, 9B, 10.3, 10.8.2,
10.8.4, 10.8.5, 11, 11.1, 11.2, 12, 13.2, 13.3, 14.2
Poinsot, L. 8.9, Notes to chapter 8
Poisson, S.-D. Note to Preface, 0.1, Note to chapter 1, 2.1.1, 3.2.3, 4.1, 5.1, 6.1.5,
7.1 – 7.4, 8, Notes to chapter 8, 10.1, 10.2, 10.5, 10.8.2, 10.8.4, 10.8.5, 10.9.2,
Notes to chapter 10, 11.1, 11.2, 12, 13.1 – 13.3, 14.2, 15.1, 15.1.2
Polo, Marco 1.2.3
Polya, G. 4.4, 5.1, 15.2
Postnikov, A. G. 13.2
Prevost, P. Notes to chapter 3
Price, R. 2.2.3, 5.1 – 5.3, 7.1
Proctor, R. A. 6.1.6, 10.8, 10.8.4
Prokhorov, Yu. V. 0.1, 0.2, 2.1.3, 2.1.4, Notes to chapter 5, 10.8.5, Notes to chapter
12, 13.1, 13.2
Prudnikov, V. E. 10.4, 13.2, 13.3, Notes to chapter 13
Ptolemy, C. (IIc.) Preface, 1.1.4, 1.2.4, Note to chapter 1, 3.2.3
Ptukha, M. V. Notes to chapter 10
Quetelet, A. 2.2.3, 6.1.4, 6.2.2, 7.3, 7.4, Notes to chapter 7, 8.9, Notes to chapter
8, 9A.6, 10.2, 10.5, 10.7, 10.8.1 – 10.8.3, 10.9.2, Notes to chapter 10, 15.1, 15.1.1,
15.1.2, 15.2
Quine, M. P. 15.1.2
Rabbi Elieser ben Chisma, 1.1.2
Rabbi Meir, 1.1.2
Rabbi Schlomo ben Adret, 1235 – 1310, 1.1.2
Rabinovitch, N. L. 1.1.2, 1.1.4, Note to chapter 1, 2.1.1
Radelet de Grave P. 6.2.4
Radlov, E. L. 14.5
Raikher, V. K. 1.1.5
Raimi, R. A. 10.8.4
Ranade R. D. Note to chapter 1
Rao, C. R. 7.1
Réaumur, R. A. 10.8.2
Rehnisch, E. 10.5
Rényi, A. 2.1.1
Riemann, B. 13.3
Rigaud, S. P. 1.2.2
Robinson, G. 6.3.2, 9A.2, 13.2
Romanovsky, V. I. 3.2.3, 14.2, 15.1.2, 15.1.3, 15.2
Rosental, I. S. 1.2.4
Rubanovsky L. M. 10.8.5
Rümelin, G. 10.5
Russell, B. Note to chapter 1
Sakatov, P. S. 9A5, 9B, Notes to chapter 9, 10.9.1
Sambursky, S. 1.1.1, Note to chapter 1,
Särndal, C.-E. 15.1.3
Sartorius von Waltershausen, W. de 9A.4
Saunderson, N. 5.3
Scheffé, H. 14.2
Scheiner, Chr. 1.2.3
Schell, E. D. 2.2.3
Schiller, F. 1.1.1
Schindler, A. Notes to chapter 10

310
308
Schlözer, A. L. 6.2.1, 6.2.3, Notes to chapters 10 and 15
Schmeidler, F. 9C
Schmidt, O. Ju. 12
Schneider, I. 4.4, 10.8.5
Schreiber, O. 6.3.3
Schumacher, H. Preface, 9A4, 9C, 15.1.2
Schumann R. 9B
Schumpeter J. A. 15.2
Schwabe, H. 1.2.3
Seal, H. L. 9A.4
Seidel, L. 10.6, 10.8.1
Seneta, E. 4.1, 8.7, 8.9.2, 10.1, 10.2, Notes to chapter 10, 11.1, 11.2, 12, 13.1, 14.2,
14.3, 15.1.2, 15.1.3, 15.3
Sevastianov, B. A. 0.1, 13.2
Shafer, G. Notes to chapter 3
Shakespeare W. 0.5
Shaw, N. 0.5, 6.2.4, 10.8.3
Shiyanov, A. N. 10.4
Shoesmith, E. 2.2.2, 2.2.4, 3.3.4, 6.3.1
Short, J. 6.3.1
Simon, J. 6.2.3
Simplicius (died 549) 1.1.1
Simpson, J. Y. 10.8.1
Simpson, T. 0.3, 4.2, Notes to chapter 4, 6.1.6, 6.3.1, Notes to chapter 6, 7.1, 8.2,
10.4
Sleshinsky, I. V. 13.1
Slutsky, E. E. 14.1, 15.1.2, 15.3, Notes to chapter 15
Smit, M. Notes to chapter 15
Smith, A. 6.2.1
Snow, J. 10.8.1
Socrates (470 or 469 – 399 or 390 BC) Notes to chapter 3
Sofonea, T. 2.1.3
Sokolov, V. S. 1.2.4
Soloviev, A. D. 14,2, 14.5
Soloviev, V. S. 14.5
Spieß, O. Notes to chapter 3
Spieß, W. 9A.1.1
Spinoza, B. Note to chapter 1
Sprott D. A. 9A4
Stäckel, P. 9
Steiner, H.-G 8.9.2
Steklov, V. A. 7.1, 10.4
Stewart G. W. 9A4
Stieltjes, T. J. 13.1
Stigler, S. M. Preface, Note to Preface, 5.3, 6.3.2, Notes to chapter 6, 7.2, Notes to
chapter 9, 10.6, Notes to chapter 10
Stirling, J. 3.2.3, 4.4, 13.2, 14.1
Stokes, G. G. 10.8.2
Strabo (64 or 63 BC – 23 or 24 AD) Note to chapter 1
Strasser G. 9C
Strecker, H. Notes to chapter 15
Strecker, R. Notes to chapter 15
Struve, L O. 9A3
Struve F. G. W. 9C, 10.8.4
Struve O. 10.8.4
Struve, P. B. 9C, 10.4
Student (Gosset, W.S.) 7.1, 9A4, 9B, 14.3, 15.2
Subbotin, M. F. 9A.5
Süssmilch, J. P. 1.1.2, 2.1.3, 2.1.4, 2.2.3, 6.2.1 – 6.2.3, 10.5
Swift, J. 6.1.3
Sylla, E. D. Notes to chapter 2
Sylvester, J. J. 12

311
309
Symms S. L. T. 10.8.4
Tait, P. G. 9A.1.3, 11.2
Takácz (Takács), L. 2.2.1, 3.3.4, 11.1
Taqqi, M. S. 14.2
Taylor, R. Notes to chapter 10
Tee, J. O. 15.2
Thatcher, A. R. 3.3.4
Thomas Aquinas (1225 or 1226 – 1274) 1.1.5, Notes to chapters 1 and 2
Thomson, W. 9A.1.3
Thorndike, L. Note to chapter 1
Tikhomandritsky, M. A. 13.3
Tikhomirov, E. I. 6.2.4
Timerding, H. E. 5.2, 6.3.3
Toaldo, J. 6.2.4
Todhunter, I. 2.1.3, 2.1.4, 3.3.2, 4.4, Notes to chapter 4, 6.1.1, 6.1.2, 6.1.5, Notes to
chapter 6, 7.1, 7.2, 7.4, Notes to chapter 7, 8.3, 15.2
Tolstoy, L. N. 0.5, 14.3
Tonti, L. 2.1.3
Toomer, G. J. 1.1.4
Truesdell, C. Preface, 0.5, 7.2, 10.8.5
Tsinger, V. Ya. 7.2, 9A4
Turgot, A. R. J. 6.1.5
Tutubalin, V. N. 1.1.2, 9A.1
Ulpianus D. (ca. 170 – 228), 1.1.5, 2.1.3
Urlanis, B. Ts. 2.1.4
Usov, N. A. 14.5
Uspensky, J. V. 3
Vasiliev, A. V. 10.4, 10.5, Notes to chapter 13
Vauban, Sebastein le pretre de, Notes to chapter 10
Virchow R. von 10.8.1
Voltaire (Arouet, M. E.) 2.2.3
Waerden, B. L. van der, 1.1.5, 2.1.3, 2.2.2, Notes to chapter 2, 10.8.2
Wallace, A. R. 10.5
Wallis, J. Notes to chapter 3, 6.1.1
Waterston J. J. 10.8.5
Weber, E. H. 6.1.1, 7.1, 10.9.2
Weber, W. E. 9A
Weil, A. Notes to chapter 2
Weiling, F. 10.8.1
Weldon, W. F. R. 14.2, 15.2
Wesley, W. G. 1.2.2
Weyl, H. 10.8.4, 11.2
White, A. D. 6.2.3
Whiteside, D. T. Preface, 6.3.1
Whittaker, E. T. 6.3.2, 9A.2, 13.2
Whittaker, L. 15.1.2
Whitworth, W. A. Notes to chapter 13
Willcox W. F. 2.1.4
Wilks, S. S. 7.1
William III 2.1.3
Wilson, C. 1.1.4, 6.3.2, Notes to chapter 6
Winkler, W. 15.1.2
Winslow, C.-E. A. 10.8.1
Wittstein, Th. 10.7
Wolf, Abr. 6.2.4
Woolhouse, W. S. B. 10.7
Woytinsky W. S. 15.1.2
Wren Ch. Preface
Yamazaki, E. 6.1.2, 6.1.5
Yarochenko, S. P. 14.5
Yastremsky, B. S. 15.1.2
You Poh Seng 10.8.4

312
310
Youshkevitch, A. A. 14.1, 14.3
Youshkevitch, A. P. Preface, Notes to chapter 3, 6.1.2
Yule, G. U. 9A.5, 9A.6, 10.5, 10.7, 10.9.3, 14.3, 15.2
Zabell, S. L. 2.1.2, 3.2.1, 5.3, 6.1.5, 7.1, 11.1, 11.2, 15.2
Zach, F. X. von 9A.1.4
Zermelo, E. 10.8.5
Zhukovsky, N. E. 14.3
Zoch, R. T. 9A.2
Zyrlin, L. Notes to chapter 15

313
311