Asymptotic Analysis of Random Walks Light Tailed Distributions 1st Edition A.A. Borovkov All Chapter Instant Download

Full download ebooks at ebookmeta.
com
Asymptotic Analysis of Random Walks Light Tailed

Distributions 1st Edition A.A. Borovkov
https://ebookmeta.com/product/asymptotic-analysis-of-random-
walks-light-tailed-distributions-1st-edition-a-a-borovkov-2/
OR CLICK BUTTON
DOWLOAD NOW
Download more ebook from https://ebookmeta.com

More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Asymptotic Analysis of Random Walks Light Tailed

Distributions 1st Edition A.A. Borovkov
https://ebookmeta.com/product/asymptotic-analysis-of-random-
walks-light-tailed-distributions-1st-edition-a-a-borovkov/
Random Walks on Infinite Groups 1st Edition Steven P.

Lalley
https://ebookmeta.com/product/random-walks-on-infinite-
groups-1st-edition-steven-p-lalley/
Continuous Time Random Walks for the Numerical Solution

of Stochastic Differential Equations 1st Edition Nawaf
Bou-Rabee
https://ebookmeta.com/product/continuous-time-random-walks-for-
the-numerical-solution-of-stochastic-differential-equations-1st-
edition-nawaf-bou-rabee/
Asymptotic Geometric Analysis Part II 1st Edition Shiri

Artstein-Avidan
https://ebookmeta.com/product/asymptotic-geometric-analysis-part-
ii-1st-edition-shiri-artstein-avidan/
Statistical Analysis of Graph Structures in Random
Variable Networks V. A. Kalyagin
https://ebookmeta.com/product/statistical-analysis-of-graph-
structures-in-random-variable-networks-v-a-kalyagin/
Distributions 1st Edition Jacques Simon
https://ebookmeta.com/product/distributions-1st-edition-jacques-
simon/
Theory of Distributions 2nd Edition Svetlin G. Georgiev
https://ebookmeta.com/product/theory-of-distributions-2nd-
edition-svetlin-g-georgiev/
Distributions, Volume 3 1st Edition Jacques Simon
https://ebookmeta.com/product/distributions-volume-3-1st-edition-
jacques-simon/
Philosophers Walks 1st Edition Bruce Baugh
https://ebookmeta.com/product/philosophers-walks-1st-edition-
bruce-baugh/
Asymptotic Analysis of Random Walks
This is a companion book to Asymptotic Analysis of Random Walks: Heavy-Tailed

Distributions by A. A. Borovkov and K. A. Borovkov. Its self-contained systematic
exposition provides a highly useful resource for academic researchers and professionals
interested in applications of probability in statistics, ruin theory, and queuing theory. The
large deviation principle for random walks was first established by the author in 1967,
under the restrictive condition that the distribution tails decay faster than exponentially.
(A close assertion was proved by S. R. S. Varadhan in 1966, but only in a rather special
case.) Since then, the principle has been treated in the literature only under this condition.
Recently, the author, jointly with A. A. Mogul’skii, removed this restriction, finding a
natural metric for which the large deviation principle for random walks holds without any
conditions. This new version is presented in the book, as well as a new approach to
studying large deviations in boundary crossing problems. Many results presented in the
book, obtained by the author himself or jointly with co-authors, are appearing in a
monograph for the first time.
A . A . B o r ov k ov is Principal Research Fellow and a Russian Academy of Sciences

Adviser at the Sobolev Institute of Mathematics, where he also served as the founding
head of the Department of Probability Theory and Mathematical Statistics. He was the
founding Chair in Probability Theory and Mathematical Statistics at the Novosibirsk State
University in 1965. He has authored and co-authored several influential research
monographs. In 1979 he was awarded (together with V. V. Sazonov and V. Statulevičius)
the State Prize of the USSR for outstanding research results on asymptotic methods of
probability theory, and was elected a full member of the Academy of Sciences of the
USSR in 1990. He received the 2002 Russian Government Prize in Education, the 2003
Russian Academy of Sciences A. A. Markov Prize, and in 2015 the Russian Academy of
Sciences A. A. Kolmogorov Prize for his joint work with A. A. Mogul’skii on the
extended large deviation principle (2015).
Encyclopedia of Mathematics and Its Applications
This series is devoted to significant topics or themes that have wide application in
mathematics or mathematical science and for which a detailed development of the
abstract theory is less important than a thorough and concrete exploration of the
implications and applications.
Books in the Encyclopedia of Mathematics and Its Applications cover their subjects
comprehensively. Less important results may be summarized as exercises at the ends of
chapters. For technicalities, readers can be referred to the bibliography, which is expected
to be comprehensive. As a result, volumes are encyclopedic references or manageable
guides to major subjects.
ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS
All the titles listed below can be obtained from good booksellers or from Cambridge University Press. For a
complete series listing, visit www.cambridge.org/mathematics
126 O. Calin and D.-C. Chang Sub-Riemannian Geometry

127 M. Grabisch et al. Aggregation Functions
128 L. W. Beineke and R. J. Wilson (eds.) with J. L. Gross and T. W. Tucker Topics in Topological Graph Theory
129 J. Berstel, D. Perrin and C. Reutenauer Codes and Automata
130 T. G. Faticoni Modules over Endomorphism Rings
131 H. Morimoto Stochastic Control and Mathematical Modeling
132 G. Schmidt Relational Mathematics
133 P. Kornerup and D. W. Matula Finite Precision Number Systems and Arithmetic
134 Y. Crama and P. L. Hammer (eds.) Boolean Models and Methods in Mathematics, Computer Science, and
Engineering
135 V. Berthé and M. Rigo (eds.) Combinatorics, Automata and Number Theory
136 A. Kristály, V. D. Rădulescu and C. Varga Variational Principles in Mathematical Physics, Geometry, and
Economics
137 J. Berstel and C. Reutenauer Noncommutative Rational Series with Applications
138 B. Courcelle and J. Engelfriet Graph Structure and Monadic Second-Order Logic
139 M. Fiedler Matrices and Graphs in Geometry
140 N. Vakil Real Analysis through Modern Infinitesimals
141 R. B. Paris Hadamard Expansions and Hyperasymptotic Evaluation
142 Y. Crama and P. L. Hammer Boolean Functions
143 A. Arapostathis, V. S. Borkar and M. K. Ghosh Ergodic Control of Diffusion Processes
144 N. Caspard, B. Leclerc and B. Monjardet Finite Ordered Sets
145 D. Z. Arov and H. Dym Bitangential Direct and Inverse Problems for Systems of Integral and Differential
Equations
146 G. Dassios Ellipsoidal Harmonics
147 L. W. Beineke and R. J. Wilson (eds.) with O. R. Oellermann Topics in Structural Graph Theory
148 L. Berlyand, A. G. Kolpakov and A. Novikov Introduction to the Network Approximation Method for
Materials Modeling
149 M. Baake and U. Grimm Aperiodic Order I: A Mathematical Invitation
150 J. Borwein et al. Lattice Sums Then and Now
151 R. Schneider Convex Bodies: The Brunn–Minkowski Theory (Second Edition)
152 G. Da Prato and J. Zabczyk Stochastic Equations in Infinite Dimensions (Second Edition)
153 D. Hofmann, G. J. Seal and W. Tholen (eds.) Monoidal Topology
154 M. Cabrera García and Á. Rodríguez Palacios Non-Associative Normed Algebras I: The Vidav–Palmer and
Gelfand–Naimark Theorems
155 C. F. Dunkl and Y. Xu Orthogonal Polynomials of Several Variables (Second Edition)
156 L. W. Beineke and R. J. Wilson (eds.) with B. Toft Topics in Chromatic Graph Theory
157 T. Mora Solving Polynomial Equation Systems III: Algebraic Solving
158 T. Mora Solving Polynomial Equation Systems IV: Buchberger Theory and Beyond
159 V. Berthé and M. Rigo (eds.) Combinatorics, Words and Symbolic Dynamics
160. B. Rubin Introduction to Radon Transforms: With Elements of Fractional Calculus and Harmonic Analysis
161 M. Ghergu and S. D. Taliaferro Isolated Singularities in Partial Differential Inequalities
162 G. Molica Bisci, V. D. Radulescu and R. Servadei Variational Methods for Nonlocal Fractional Problems
163 S. Wagon The Banach–Tarski Paradox (Second Edition)
164 K. Broughan Equivalents of the Riemann Hypothesis I: Arithmetic Equivalents
165 K. Broughan Equivalents of the Riemann Hypothesis II: Analytic Equivalents
166 M. Baake and U. Grimm (eds.) Aperiodic Order II: Crystallography and Almost Periodicity
167 M. Cabrera García and Á. Rodríguez Palacios Non-Associative Normed Algebras II: Representation Theory
and the Zel’manov Approach
168 A. Yu. Khrennikov, S. V. Kozyrev and W. A. Zúñiga-Galindo Ultrametric Pseudodifferential Equations
and Applications
169 S. R. Finch Mathematical Constants II
170 J. Krajíček Proof Complexity
171 D. Bulacu, S. Caenepeel, F. Panaite and F. Van Oystaeyen Quasi-Hopf Algebras
172 P. McMullen Geometric Regular Polytopes
173 M. Aguiar and S. Mahajan Bimonoids for Hyperplane Arrangements
174 M. Barski and J. Zabczyk Mathematics of the Bond Market: A Lévy Processes Approach
175 T. R. Bielecki, J. Jakubowski and M. Niewȩgłowski Fundamentals of the Theory of Structured Dependence
between Stochastic Processes
176 A. A. Borovkov Asymptotic Analysis of Random Walks: Light-Tailed Distributions
Asymptotic Analysis of Random Walks
Light-Tailed Distributions
A . A . B O ROV KOV
Sobolev Institute of Mathematics, Novosibirsk
Translated by
V. V. U LYA N OV
Lomonosov Moscow State University
and HSE University, Moscow
M . V. Z H I T L U K H I N
Steklov Institute of Mathematics, Moscow
University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906
Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit
of education, learning, and research at the highest international levels of excellence.
www.cambridge.org
Information on this title: www.cambridge.org/9781107074682
DOI: 10.1017/9781139871303
© A. A. Borovkov 2020
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2020
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Borovkov, A. A. (Aleksandr Alekseevich), 1931– author.
Title: Asymptotic analysis of random walks : light-tailed distributions /
A.A. Borovkov, Sobolev Institute of Mathematics, Novosibirsk ;
translated by V.V. Ulyanov, Higher School of Economics, Mikhail
Zhitlukhin, Steklov Institute of Mathematics, Moscow.
Description: Cambridge, United Kingdom ; New York, NY : Cambridge
University Press, 2020. | Series: Encyclopedia of mathematics and its
applications | Includes bibliographical references and index.
Identifiers: LCCN 2020022776 (print) | LCCN 2020022777 (ebook) |
ISBN 9781107074682 (hardback) | ISBN 9781139871303 (epub)
Subjects: LCSH: Random walks (Mathematics) | Asymptotic expansions. |
Asymptotic distribution (Probability theory)
Classification: LCC QA274.73 .B6813 2020 (print) | LCC QA274.73 (ebook) |
DDC 519.2/82–dc23
LC record available at https://lccn.loc.gov/2020022776
LC ebook record available at https://lccn.loc.gov/2020022777
ISBN 978-1-107-07468-2 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
Contents
Introduction page ix
1 Preliminary results 1
1.1 Deviation function and its properties in the one-
dimensional case 1
1.2 Deviation function and its properties in the multidimen-
sional case 13
1.3 Chebyshev-type exponential inequalities for sums of
random vectors 18
1.4 Properties of the random variable γ = (ξ ) and its
deviation function 25
1.5 The integro-local theorems of Stone and Shepp
and Gnedenko’s local theorem 32
2 Approximation of distributions of sums of random variables 39
2.1 The Cramér transform. The reduction formula 39
2.2 Limit theorems for sums of random variables in the Cramér
deviation zone. The asymptotic density 43
2.3 Supplement to section 2.2 50
2.4 Integro-local theorems on the boundary of the Cramér zone 65
2.5 Integro-local theorems outside the Cramér zone 70
2.6 Supplement to section 2.5. The multidimensional case.
The class of distributions ER 75
2.7 Large deviation principles 81
2.8 Limit theorems for sums of random variables with
non-homogeneous terms 90
2.9 Asymptotics of the renewal function and related problems.
The second deviation function 101
2.10 Sums of non-identically distributed random variables in
the triangular array scheme 116
v
vi Contents
3 Boundary crossing problems for random walks 136

3.1 Limit theorems for the distribution of jumps when the end
of a trajectory is fixed. A probabilistic interpretation of the
Cramér transform 137
3.2 The conditional invariance principle and the law
of the iterated logarithm 139
3.3 The boundary crossing problem 150
3.4 The first passage time of a trajectory over a high level and
the magnitude of overshoot 155
3.5 Asymptotics of the distribution of the first passage time
through a fixed horizontal boundary 169
3.6 Asymptotically linear boundaries 186
3.7 Crossing of a curvilinear boundary by a normalised
trajectory of a random walk 188
3.8 Supplement. Boundary crossing problems in the
multidimensional case 201
3.9 Supplement. Analytic methods for boundary crossing
problems with linear boundaries 216
3.10 Finding the numerical values of large deviation probabilities 228
4 Large deviation principles for random walk trajectories 234
4.1 On large deviation principles in metric spaces 234
4.2 Deviation functional (or integral) for random walk
trajectories and its properties 246
4.3 Chebyshev-type exponential inequalities for trajectories of
random walks 265
4.4 Large deviation principles for continuous random walk
trajectories. Strong versions 270
4.5 An extended problem setup 282
4.6 Large deviation principles in the space of functions without
discontinuities of the second kind 290
4.7 Supplement. Large deviation principles in the space (V, ρV ) 300
4.8 Conditional large deviation principles in the space (D, ρ) 305
4.9 Extension of results to processes with independent increments 314
4.10 On large deviation principles for compound renewal
processes 329
4.11 On large deviation principles for sums of random variables
defined on a finite Markov chain 340
Contents vii
5 Moderately large deviation principles for the trajectories of

random walks and processes with independent increments 343
5.1 Moderately large deviation principles for sums Sn 343
5.2 Moderately large deviation principles for trajectories sn 350
5.3 Moderately large deviation principles for processes with
independent increments 358
5.4 Moderately large deviation principle as an extension of the
invariance principle to the large deviation zone 361
5.5 Conditional moderately large deviation principles for the
trajectories of random walks 363
6 Some applications to problems in mathematical statistics 375
6.1 Tests for two simple hypotheses. Parameters
of the most powerful tests 375
6.2 Sequential analysis 380
6.3 Asymptotically optimal non-parametric goodness of fit tests 386
6.4 Appendix. On testing two composite parametric hypotheses 395
6.5 Appendix. The change point problem 397
Basic notation 407
References 410
Index 419
Introduction
This book is devoted mainly to the study of the asymptotic behaviour of the
probabilities of rare events (large deviations) for trajectories of random walks. By
random walks we mean the sequential sums of independent random variables or
vectors, and also processes with independent increments. It is assumed that those
random variables or vectors (jumps of random walks or increments of random
processes) have distributions which are ‘rapidly decreasing at infinity’. The last
term means distributions which satisfy Cramér’s moment condition (see below).
The book, in some sense, continues the monograph [42], where more or less
the same scope of problems was considered but it was assumed that the jumps of
random walks have distributions which are ‘slowly decreasing at infinity’, i.e. do
not satisfy Cramér’s condition. Such a division of the objects of study according to
the speed of decrease of the distributions of jumps arises because for rapidly and
slowly decreasing distributions those objects form two classes of problems which
essentially differ, both in the methods of study that are required and also in the
nature of the results obtained.
Each of these two classes of problems requires its own approach, to be devel-
oped, and these approaches have little in common. So, the present monograph,
being a necessary addition to the book [42], hardly intersects with the latter in
its methods and results. In essence, this is the second volume (after [42]) of a
monograph with the single title Asymptotic Analysis of Random Walks.
The asymptotic analysis of random walks for rapidly decreasing distributions
and, in particular, the study of the probabilities of large deviations have become
one of the main subjects in modern probability theory. This can be explained as
follows.
• A random walk is a classical object of probability theory, which presents huge
theoretical interest and is a mathematical model for many important applications
in mathematical statistics (sequential analysis), insurance theory (risk theory),
queuing theory and many other fields.
• Asymptotic analysis and limit theorems (under the unbounded growth of some
parameters, for example the number of random terms in a sum) form the chief
ix
x Introduction
method of research in probability theory. This is due to the nature of the main
laws in probability theory (they have the form of limit theorems), as well as
the fact that explicit formulas or numerical values for the characteristics under
investigation in particular problems, generally, do not exist and one has to find
approximations for them.
• Probabilities of large deviations present a considerable interest from the math-
ematical point of view as well as in many applied problems. Finding the
probabilities of large deviations allows one, for example, to find the small
error probabilities in statistical hypothesis testing (error probabilities should
be small), the small probabilities of the bankruptcy of insurance companies
(they should be small as well), the small probabilities of the overflow of
bunkers in queuing systems and so on. The so-called ‘rough’ theorems about the
probabilities of large deviations (i.e. about the asymptotics of the logarithms of
those probabilities; see Chapters 4 and 5) have found application in a series of
related fields such as statistical mechanics (see, for example, [83], [89], [178]).
• Rapidly decreasing distributions deserve attention because the first classical
results about the probabilities of large deviations of sums of random variables
were obtained for rapidly decreasing distributions (i.e. distributions satisfying
Cramér’s condition). In many problems in mathematical statistics (especially
those related to the likelihood principle), the condition of rapid decrease turns
out to be automatically satisfied (see, for example, sections 6.1 and 6.2 below).
A rapid (in particular, exponential) decrease in distributions often arises in
queuing theory problems (for example, the Poisson order flow is widely used),
in risk theory problems and in other areas. Therefore, the study of problems with
rapidly decreasing jump distributions undoubtedly presents both theoretical and
applied interest. Let us add that almost all the commonly used distributions in
theory and applications, such as the normal distribution, the Poisson distribu-
tion, the -distribution, the distribution in the Bernoulli scheme, the uniform
distribution, etc. are rapidly decreasing at infinity.
Let ξ , ξ1 , ξ2 , . . . be a sequence of independent identically distributed random
variables or vectors. Put S0 = 0 and

n
Sn := ξk , n = 1, 2, . . .
k=1
The sequence {Sn ; n 0} is called a random walk. As has been noted, a random
walk is a classical object of probability theory. Let us mention the following
fundamental results related to random walks.
• The strong law of large numbers, on the convergence Sn /n → Eξ almost surely
a.s.
as n → ∞.
• The functional central limit theorem, on the convergence in distribution of the
process ζn (t), t ∈ [0, 1], with values (Sk − Ak )/Bn at the points t = k/n,
k = 0, 1, . . . , n, to a stable process, where Ak , Bn are appropriate normalising
constants. For example, in the case Eξ = 0, Eξ 2 = σ 2 < ∞, the polygonal
Introduction xi
line ζn (t) for Ak = 0, B2n = nσ 2 converges in distribution to a standard
Wiener process (the invariance principle; for details, see e.g. [11], [39] or the
introduction in [42]).
• The law of the iterated logarithm, which establishes upper and lower bounds for
the trajectories of {Sk }.
None of the above results describes the asymptotic behaviour of the probabil-
ities of large deviations of trajectories of {Sk }. We mention the following main
classes of problems.
(a) The study of the probabilities of large deviations of sums of random variables
(or vectors); for example, the study of the asymptotics (for Eξ = 0, Eξ 2 < ∞
in the one-dimensional case) of probabilities
√
P(Sn x) for x n, n → ∞. (0.0.1)
(b) The study of the probabilities of large deviations in boundary crossing prob-
lems. For example, in this class of problems belongs a problem concerning
the asymptotics of the probabilities

x
P max ζn (t) − √ g(t) 0 (0.0.2)
t∈[0,1] n
√
for an arbitrary function g(t) on [0, 1] in the case x n as n → ∞ (the
process ζn (t) is defined above).
(c) The study of the more general problem about the asymptotics of the probabil-
ities

x √
P ζn (·) ∈ √ B , x n, (0.0.3)
n
where B is an arbitrary measurable set in one or another space of the functions
on [0, 1].
This monograph is largely devoted to the study of problems concerning prob-
abilities (0.0.1)–(0.0.3) and other closely related problems, under the assumption
that Cramér’s moment condition
[C] ψ(λ) := Eeλξ < ∞
is satisfied for some λ = 0 (its statement here is given for a scalar ξ ). This condition
means (see subsection 1.1.1) that at least one of the ‘tails’ P(ξ x) or P(ξ < −x)
of the distribution of the variable ξ decreases as x → ∞ faster than some exponent.
In the monograph [42], it was assumed that condition [C] is not satisfied but the
tails of the distribution of ξ behave in a sufficiently regular manner.
As we noted above, the first general results for problem (a) about the asymp-
totics of the probabilities of large deviations of sums of random variables go
back to the paper of Cramér. Essential contributions to the development of this
xii Introduction
direction were also made in the papers of V.V. Petrov [149], R.R. Bahadur and
R. Ranga Rao [5], C. Stone [175] and others. One should also mention the papers
of B.V. Gnedenko [95], E.A. Rvacheva [163], C. Stone [174] and L.A. Shepp [166]
on integro-local limit theorems for sums of random variables in the zone of normal
deviations, which played an important role in extending such theorems to the zone
of large deviations and forming the most adequate approach (in our opinion) to
problems concerning large deviations for sums Sn . This approach is presented in
Chapter 2 (see also the papers of A.A. Borovkov and A.A. Mogul’skii [56], [57],
[58], [59]).
The first general results about the joint distribution of Sn and Sn := maxkn Sk
(this is a particular case of the boundary crossing problem (b); see (0.0.2)) in
the zone of large deviations were obtained by A. A. Borovkov in the papers
[15] and [16], using analytical methods based on solving generalised Wiener–
Hopf equations (in terms of Stieltjes-type integrals) for the generating function
of the joint distribution of Sn and Sn . The solution was obtained in the form
of double transforms of the required joint distribution, expressed in terms of
factorisation components of the function 1 − zψ(λ). The double transforms, as
functions of the variables λ and z, can be inverted if one knows the poles of those
transforms as a function of the variable λ and applies modifications of the steepest
descent methods in the variable z. Those results allowed author A.A. Borovkov in
[17]–[19] to find a solution to a more general problem about the asymptotics of the
probabilities (0.0.2). Later in the work by A. A. Borovkov and A. A. Mogul’skii
[53], the asymptotics of the probability (0.0.2) were found for some cases
using direct probability methods without the factorisation technique (see
Chapter 3).
The first general results on the rough asymptotics in problem (c), i.e. on
the logarithmic asymptotics of the probability of the general form (0.0.3) for
arbitrary sets B in the one-dimensional case (these results constitute the large
deviation principle) were obtained in the papers of S.R.S. Varadhan [177] (for
a special case and x n) and A.A. Borovkov [19] (for x = O(n)). In the
paper of A.A. Mogul’skii [131], the results of the paper [19] were transferred
to the multidimensional case. The large deviation principle was later extended
to a number of other objects (see, for example, [83], [178], [179]). However,
the large deviation principle in the papers [177], [19], [131] was established
under a very restrictive version of Cramér’s condition, that ψ(λ) < ∞ for all
λ, and only for continuous trajectories ζn (t) with Ak = 0, Bn = n. In a more
recent series of papers by A.A. Borovkov and A.A. Mogul’skii [60]–[68] (see also
Chapter 4) substantial progress in the study of the large deviation principle was
made: Cramér’s condition, mentioned above in its strong form, was weakened
or totally removed and the space of trajectories was extended up to the space of
functions without discontinuities of the second kind. At the same time the form of
the results changed, so it became necessary to introduce the notion of the ‘extended
large deviation principle’ (see Chapter 4).
Introduction xiii
Let us mention that in the paper of A.A. Borovkov [19] the principle
of moderately large deviations was established as well for ζn (t) as x = o(n),
n → ∞. After that A.A. Mogul’skii [131] extended those results to the multi-
dimensional case. In a recent paper [67] those results were strengthened (see
Chapter 5 for details).
Also, note that a sizable literature has been devoted to the large deviation
principle for a wide class of random processes, mainly Markov processes and
processes arising in statistical mechanics (see, for example, [83], [89], [93], [178],
[179]). Those publications have little in common with the present monograph in
their methodology and the nature of the results obtained, so we will not touch upon
them.
The above overview of the results does not in any way pretend to be complete.
Our goal here is simply to identify the main milestones in the development of the
limit theorems in probability theory that are presented in this monograph. A more
detailed bibliography will be provided in the course of the presentation.
Let us now provide a brief description of the contents of the book.

This monograph contains main and supplemental sections. The presentation
of the main sections is, generally, self-sufficient; they contain full proofs. The
supplemental sections contain results which are close to the main results, but
the presentation in those sections is concise and, typically, proofs are not given.
We have included in these supplements results whose proofs are either too
cumbersome or use methods beyond the scope of this monograph. In those cases,
we supply references to the omitted proofs.
This book consists of six chapters. The first chapter contains preliminary
results which will be used in what follows. In sections 1.1 and 1.2 we discuss
Cramér’s condition and state the properties of the deviation function, which plays
an important role throughout the book. We obtain inequalities for the distributions
of the random variables Sn and their maxima Sn . In section 1.3 we establish
exponential Chebyshev-type inequalities for sums of random vectors in terms of
deviation functions. Section 1.5 is devoted to the integro-local limit theorems of
Gnedenko and Stone.
In the second chapter we study the asymptotics of the distributions of the sums
Sn of random variables and vectors (problem (a)). In section 2.1, we introduce
the notion of the Cramér transform and establish the so-called reduction formula,
which reduces a problem concerning integro-local theorems in the zone of large
deviations to the same problem in the zone of normal deviations. In section 2.2,
integro-local limit theorems in the so-called Cramér deviation zone are obtained.
Section 2.3 contains supplements to the results of section 2.2, which are provided
without proofs. They include local theorems for densities in the homogeneous case
and integro-local limit theorems in the multidimensional case. In section 2.4, we
obtain integro-local theorems on the border of the Cramér zone of deviations and
in section 2.5 theorems that apply outside the Cramér zone. In the latter case, it
xiv Introduction
is possible to obtain substantial results only for special classes of distributions
which vary sufficiently regularly at infinity. Section 2.6 contains supplements
to the results of sections 2.4 and 2.5 relating to the multidimensional case. In
section 2.7, we present the large deviation principle for the sums Sn . The influence
of one or several ‘non-homogeneous’ terms in the sum Sn on the form of the results
obtained in sections 2.2–2.7 is studied in section 2.8. Section 2.9 contains additions
concerning the large deviation principle for renewal functions, the probability that
the sequence {Sn }∞n=1 reaches a distant set and other related problems.
The third chapter is devoted to boundary crossing problems (problem (b); see
(0.0.2)) for random walks. In section 3.1, we investigate the limiting behaviour
of the conditional distribution of the jumps of a random walk when the end of
a trajectory (the sum Sn ) is fixed. This allows us to understand the probabilistic
meaning of the Cramér transform. In section 3.2, conditional invariance principles
and the law of the iterated logarithm are established (again, when the end of a
trajectory is fixed). The problem of the crossing of an arbitrary boundary by a
trajectory {Sk } is considered in section 3.3. In section 3.4, we study the joint
distribution of the first time that a random walk crosses a high level and the
overshoot over this level.
The first passage time over a fixed level (in particular, the zero level) is studied
in section 3.5. In section 3.6, the distribution of the first passage time over a
curvilinear boundary for a class of asymptotically linear boundaries is considered.
In section 3.7, we consider the same problem as in section 3.6 but for arbitrary
boundaries and normalised trajectories on the segment [0, 1]. A generalisation
of those results to the multidimensional case is provided without proof in sec-
tion 3.8. In section 3.9, also without proof, we give an account of the analyt-
ical approach to the study of the joint distribution of the variables Sn and Sn .
The probability
uyx,n = P(Sn−1 < x, Sn x + y)
y
satisfies an integro-difference equation, so the generating function ux (z) =
∞ n y
n=1 z ux,n satisfies the generalised (in terms of the Stieltjes integral) Wiener–
Hopf integral transform on the half-line. For the Laplace transform uy (λ, z) of
y
ux (z) (in x), we find an explicit form in terms of the so-called V-factorisation of
the function 1−zψ(λ). It turns out that this double transform can be asymptotically
inverted, since it is possible to find an explicit form for the pole of the function
uy (λ, z) in the plane λ. This allows us to find the asymptotics in x of the
y
function ux (z), and then, using a modification of the steepest descent method, the
y
asymptotics of ux,n .
In the last section of Chapter 3 (section 3.10) we consider numerical methods
for finding values of parameters in terms of which the investigated probabilities
can be described.
Chapter 4 is devoted to the large deviation principles (l.d.p.) for trajectories
{Sk }nk=1 (problem (c); see (0.0.3)). The first three sections are of a preliminary
Introduction xv
nature. In the first section, we introduce the notions of a local and an extended l.d.p.
for random elements in an arbitrary metric space. We find conditions when a local
l.d.p. implies an extended l.d.p. In the second section, we study the functional (i.e.
the integral) of deviations, in terms of which all the main results of Chapters 4 and 5
are obtained. In section 4.3, we obtain exponential Chebyshev-type inequalities
for trajectories of a random walk (extensions of the inequalities in section 1.3
to the case of trajectories); using these inequalities, upper bounds on the l.d.p.
will be obtained. In section 4.4, we establish strong (compared to the already
known) versions of the l.d.p. for continuous normalised trajectories ζn (t) of a
random walk as x ∼ n, n → ∞. The strengthening consists in eliminating
some conditions or substantially weakening them. In section 4.5, we consider
an extended setting of the problem, which arises when the Cramér condition is
not required to hold over the the whole axis or under simultaneous extension of
the space where the trajectories are defined to the space D of functions without
discontinuities of the second type. We introduce a metric in this space which is
more relevant to the problems we consider than the Skorokhod metric, as it allows
the convergence of continuous processes to discontinuous ones. In section 4.6, we
establish the local l.d.p. in the space D and also the so-called extended l.d.p. In
section 4.7, l.d.p.’s in the space of functions of bounded variations are presented
without proof. Conditional l.d.p.’s in the space D of trajectories of random walks
with the end that is localised (in some form) are considered in section 4.8. Some
results obtained earlier in Chapter 4 are extended in section 4.9 to processes
with independent increments. As a corollary, we obtain Sanov’s theorem about
large deviations of empirical distribution functions. In section 4.10, we briefly
discuss approaches to how one can obtain l.d.p.’s for compound renewal processes
and, in section 4.11, for sums of random variables defined on a finite Markov
chain.
Chapter 5 is devoted to the moderately large deviation principle (m.l.d.p.) for
trajectories of random walks ζn (t) and processes with independent increments,
√
when x = o(n), x n as n → ∞. Section 5.1 contains a presentation of
the m.l.d.p. for sums Sn of random variables. The moderately large deviation
principle for trajectories is formulated and proved in section 5.2. Similar results
for processes with independent increments are established in section 5.3. As
a corollary, we obtain a counterpart of Sanov’s theorem for moderately large
deviations of empirical distribution functions. The connection of the m.l.d.p. to
the invariance principle is considered in section 5.4. The end of section 5.5 is
devoted to conditional m.l.d.p. with a localised end.
Some applications of the results obtained in Chapters 2–5 to problems of
mathematical statistics are provided in Chapter 6. The following problems are
considered: the finding of parameters (the small probabilities of errors of the
first and second types) for the most powerful test for two simple hypotheses
(section 6.1); the finding of parameters for optimal tests in sequential analysis
(also with small probabilities of errors; section 6.2).
xvi Introduction
In section 6.3, we construct asymptotically optimal non-parametric goodness of
fit tests. Some asymptotic results concerning the testing of two complex parametric
hypotheses are provided in section 6.4.
In section 6.5 we find asymptotics for the main characteristics in some change
point problems.
Let us now mention the main distinguishing features of the book.

(1) The traditional circle of problems about limit theorems for sums Sn in the
book is considerably extended. It includes the so-called boundary crossing
problems related to the crossing of given boundaries by trajectories of random
walks. To this circle belong, in particular, problems about probabilities of large
deviations of the maxima Sn = maxkn Sk of sums of random variables, which
are widely used in applications. For the first time, a systematic presentation
of a unified approach to solve the above-mentioned problems is provided. In
particular, we present a direct probabilistic approach to the study of boundary
crossing problems for random walks.
(2) For the first time in the monographic literature, the so-called extended large
deviation principle, which is valid under considerably wider assumptions than
before and under a wider problem setup, is presented.
(3) For the first time in the monographic literature, the conditional l.d.p. and the
moderately large deviation principle are presented.
(4) Results concerning the large deviation principle are extended to multidi-
mensional random walks and in some cases to processes with independent
increments.
(5) The book contains a considerable number of applications of the results we
obtain to certain problems in mathematical statistics.
The author is grateful to A.A. Mogul’skii and A.I. Sakhanenko for useful
remarks, and also to T.V. Belyaeva for the help with preparing the manuscript
for printing.
1
Preliminary results
1.1 Deviation function and its properties in the one-dimensional case

1.1.1 Cramér’s conditions
Let ξ , ξ1 , ξ2 , . . . be a sequence of independent identically distributed random
variables,

n
S0 = 0, Sn = ξk for n 1, Sn = max Sk .
0kn
k=1
An important role in describing the distribution asymptotics of the values Sn and

Sn , as well as the whole trajectory of {Sk }nk=0 for large n, is played by the so-called
deviation function (rate function). The deviation function is most informative if at
least one of Cramér’s moment conditions is met:
[C± ] There exists λ ≷ 0 such that

ψ(λ) := eλt P (ξ ∈ dt) < ∞.
A condition [C] will be used to mark the fulfilment of at least one of these
conditions:

[C] = [C+ ] [C− ].
We denote an intersection of the conditions [C± ] as

[C0 ] = [C+ ] [C− ].
Condition [C0 ] means, evidently, that
ψ(λ) < ∞ for sufficiently small |λ|.
If
λ+ := sup λ : ψ(λ) < ∞ , λ− = inf λ : ψ(λ) < ∞ ,
then correspondingly conditions [C± ], [C], [C0 ] can be written in the forms
λ± ≷ 0, λ+ − λ− > 0, |λ± | > 0.
1
2 Preliminary results
These conditions, which are called Cramér’s conditions, characterise the decay
rate of the ‘tails’ F± (t) for the distribution of a random variable ξ . When the
condition [C+ ] is met, by virtue of Chebyshev’s exponential inequality we have
F+ (t) := P (ξ t) e−λt ψ(λ) for λ ∈ (0, λ+ ), t > 0,
and, therefore, F+ (t) decreases exponentially as t → ∞. Conversely, if F+ (t) <
ce−μt for some c < ∞, μ > 0 and for all t > 0, then for λ ∈ (0, μ) we have
0
eλt P(ξ ∈ dt) 1 − F+ (0),
−∞
∞ ∞ ∞
e P (ξ ∈ dt) = −
λt
e dF+ (t) = F+ (0) + λ
λt
eλt F+ (t) dt
0 0 0
∞
cλ
F+ (0) + cλ e(λ−μ)t dt = F+ (0) + < ∞,
0 μ −λ
cλ
ψ(λ) 1 + < ∞.
μ−λ
There is a similar connection between the decay rate of F− (t) = P (ξ − t) as
t → ∞ and the finiteness of ψ(λ) under condition [C− ].
It is clear that condition [C0 ] implies the exponential decay of F+ (t) + F− (t) =
P |ξ | t , and vice versa.
Hereafter we also use the conditions
[C∞± ] = {λ± = ±∞}
and the condition

[C∞ ] = |λ± | = ∞ = [C∞+ ] [C∞− ].
It follows from the above that the condition [C∞+ ] ([C∞− ]) is equivalent to the
fact that the tail F+ (t) (F− (t)) diminishes faster than any exponent, as t increases.
It is clear that, for instance, an exponential distribution meets condition [C+ ]

[C∞− ] while a normal distribution meets condition [C∞ ].
Along with Cramér’s conditions we will also assume that the random variable
ξ is not degenerate, i.e. ξ ≡ const. (or D ξ > 0, which is the same).
The properties of the Laplace transform ψ(λ) of a distribution of random
variable ξ are set forth in various textbooks; see e.g. [39]. Let us mention the
following three properties, which we are going to use further on.
ψ (λ)
( 1) The functions ψ(λ) and ln ψ(λ) are strictly convex; the ratio strictly
ψ(λ)
increases on (λ− , λ+ ).
The analyticity property of ψ(λ) in a strip Re λ ∈ (λ− , λ+ ) can be supplemented
with the following ‘extended’ continuity property on a segment [λ− , λ+ ] (on the
strip Re λ ∈ [λ− , λ+ ]).
( 2) The function ψ(λ) is continuous ‘from within’ a segment [λ− , λ+ ]; i.e.
ψ(λ± ∓ 0) = ψ(λ± ) (the cases ψ(λ± ) = ∞ are not excluded).
1.1 Deviation function and its properties in the one-dimensional case 3
The continuity on the whole line might fail as, for instance, λ+ < ∞, ψ(λ+ ) <
∞, ψ(λ+ + 0) = ∞, which is the case for the distribution of a random variable ξ
with density f (x) = cx−3 e−λ+ x for x 1, c = const.
( 3) If E |ξ |k < ∞ and the right-hand side of Cramér’s condition [C+ ] is met,
then the function ψ is k times right-differentiable at the point λ = 0,
ψ (k) (0) = E ξ k =: ak ,
and, as λ ↓ 0,

k
λj
ψξ (λ) = 1 + aj + o(λk ).
j!
j=1
It also follows that the next representation takes place as λ ↓ 0:

k
γj λ j
ln ψξ (λ) = + o(λk ), (1.1.1)
j!
j=1
where the γj are so-called semi-invariants (or cumulants) of order j of a random

variable ξ . It is not difficult to check that
γ1 = a1 , γ2 = a02 = σ 2 , γ3 = a03 , . . . , (1.1.2)
where a0k = E (ξ − a1 )k is a central moment of kth order.
1.1.2 Deviation function

Under the condition [C], a pivotal role in describing the asymptotics of probabil-
ities P(Sn x) is played by the deviation function.
Definition 1.1.1. The deviation function of a random variable ξ is the function 1
(α) := sup αλ − ln ψ(λ) . (1.1.3)
λ
The meaning of the name will become clear later. In classical convex analysis
the right-hand side of (1.1.3) is known as the Legendre transform of the function
A(λ) := ln ψ(λ).
Consider a function A(α, λ) := αλ − A(λ) presented under the sup sign
in (1.1.3). The function −A(λ) is strictly concave (see property ( 1)), so the
function A(α, λ) is the same (note also that A(α, λ) = − ln ψα (λ), where ψα (λ) =
e−λα ψ(λ) is the Laplace transform of the distribution of the random variable
ξ − α and, therefore, from the ‘qualitative’ point of view, A(α, λ) possesses all the
properties of the function −A(λ)). It follows from what has been said that there
always exists a unique point λ = λ(α) on the ‘extended’ real line [−∞, ∞]),
1 This function is often referred to as a rate function.

where the sup in (1.1.3) is attained. When α increases, the values of A(α, λ) for
λ > 0 will increase (in proportion to λ), and for λ < 0 they will decrease.
Therefore, a graph of A(α, λ) as a function of λ, will, roughly speaking, ‘roll over’
to the right as α increases. It means that the maximum point λ(α) will also shift
to the right (or will stay still, if λ(α) = λ+ ).
Let us move on to an exact formulation. The following three sets play an
important role in studying the properties of the function (α):
A = λ : A(λ) < ∞ = λ : ψ(λ) < ∞ , A = A (λ) : λ ∈ A ,
and the convex envelope S of the support of the distribution of ξ . It is clear that
the values λ± are bounds for the set A. The right and left bounds α± , s± for the
sets A , S, are evidently given by
ψ (λ± ∓ 0)
α± = A (λ± ∓ 0) = ;
ψ(λ± ∓ 0)
s+ = sup t : P(ξ t) < 1 ,
s− = inf t : P(ξ t) > 0 ,
where A (λ+ − 0) = lim A (λ) and A (λ− + 0) is defined analogously.

λ↑λ+
If s+ < ∞ (the variable ξ is bounded above), then asymptotically the function
A(λ), as λ → ∞, will increase linearly, so that (α) = ∞ for α > s+ . In a
similar way (α) = ∞ for s− > −∞, α < s− . Thus, we may confine ourselves
to considering the properties of the function on [s− , s+ ].
The value of α+ determines the angle at which a curve A(λ) = ln ψ(λ) meets
a point λ+ , A(λ+ ) . The value of α− has an analogous meaning. If α ∈ [α− , α+ ]
then the equation Aλ (α, λ) = 0 or, which is the same, the equation
ψ (λ)
= α, (1.1.4)
ψ(λ)
always has a unique solution λ(α) on the segment [λ− , λ+ ] (the values of λ± can
be infinite). This solution λ(α), as the inverse function to the function ψ (λ)/ψ(λ)
(see (1.1.4)), which is analytic and strictly increasing on (λ− , λ+ ), is also analytic
and strictly increasing on (α− , α+ ):
λ(α) ↑ λ+ as α ↑ α+ ; λ(α) ↓ λ− as α ↓ α− . (1.1.5)
From the equations

ψ (λ(α))
(α) = αλ(α) − A λ(α) , =α (1.1.6)
ψ(λ(α))
we obtain
ψ (λ(α))
(α) = λ(α) + αλ (α) − λ (α) = λ(α).
ψ(λ(α))
Taking into account that ψ (0)/ψ(0) = a1 = Eξ , 0 ∈ [λ− , λ+ ], a1 ∈ [α− , α+ ],
we get the following representation for the function :
(1) If α0 ∈ [α− , α+ ], α ∈ [α− , α+ ], then
α
(α) = (α0 ) + λ(v)dv. (1.1.7)
α0
Since λ(a1 ) = (a1 ) = 0 (which follows from (1.1.4) and (1.1.6)), in particular
for α0 = a1 we have
α
(α) = λ(v)dv. (1.1.8)
a1
The functions λ(α), (α) are analytic on (α− , α+ ).

Now let us consider what is happening outside the segment [α− , α+ ]. For
definiteness, let λ+ > 0. We are going to study the behaviour of the functions
λ(α), (α) for α α+ . Similar considerations can be made in the case λ− < 0,
α α− .
First let λ+ = ∞, i.e. let the function ln ψ(λ) be analytic on the whole semiaxis
λ > 0, so that the tail F+ (t) decreases, as t → ∞, faster than any exponent. We
will assume, without loss of generality, that
s+ > 0, s− < 0. (1.1.9)
This can always be achieved using a shift transformation of a random variable;

further, we assume, without loss of generality, as in many limit theorems about the
distribution of Sn , that E ξ = 0, using the fact that the problem of examining the
distribution of Sn is ‘translation-invariant’. It can also be noted that ξ −a (α −a) =
ξ (α) (see property (4) below) and that (1.1.9) always holds, if E ξ = 0.
(2) (i) If λ+ = ∞ then α+ = s+ .
Hence, if λ+ = ∞, s+ = ∞, then we always have α+ = ∞ and for any α α−
both (1.1.7) and (1.1.8) hold.
(ii) If s+ < ∞ then λ+ = ∞, α+ = s+ ,
(α+ ) = − ln P(ξ = s+ ), (α) = ∞ for α > α+ .
Similar statements are true for s− , α− , λ− .

Proof. (i) First let s+ < ∞. Then the asymptotics of ψ(λ) and ψ (λ), as λ → ∞,
are determined by the corresponding integrals in the vicinity of s+ :
ψ(λ) ∼ E (eλξ ; ξ > s+ − ε), ψ (λ) ∼ E (ξ eλξ ; ξ > s+ − ε)
as λ → ∞ and for any fixed ε > 0. It follows that

ψ (λ) E (eλξ ; ξ > s+ − ε)
α+ = lim = lim = s+ .
λ→∞ ψ(λ) λ→∞ E (ξ eλξ ; ξ > s+ − ε)
If s+ = ∞, then ln ψ(λ) increases faster than any linear function as λ → ∞,
and, therefore, the derivative ln ψ(λ) increases without limit, α+ = ∞.
(ii) The first two statements are evident. Let p+ = P(ξ = s+ ) > 0. Then
ψ(λ) ∼ p+ eλs+ ,
αλ − ln ψ(λ) = αλ − ln p+ − λs+ + o(1) = (α − α+ )λ − ln p+ + o(1)
as λ → λ+ = ∞. It follows from this and (1.1.6) that

− ln p+ for α = α+ ,
(α) =
∞ for α > α+ .
If p+ = 0, then from the relation ψ(λ) = o(eλs+ ) as λ → ∞, we obtain in a

similar way that (α+ ) = ∞. The property (2) is proved.
Now let 0 < λ+ < ∞. Then s+ = ∞. If α+ < ∞, then it is necessary that
ψ(λ+ ) < ∞, ψ(λ+ + 0) = ∞, ψ (λ+ ) < ∞. The left derivative is meant. If
ψ(λ+ ) = ∞, then ln ψ(λ+ ) = ∞ and ln ψ(λ) → ∞ as λ ↑ λ+ , α+ = ∞,
which contradicts the assumption α+ < ∞. Since ψ(λ) = ∞ for λ > λ+ , it
follows that λ(α), having reached λ+ with increasing α, stops at this point, so that
for α α+ we have
λ(α) = λ+ , (α) = (α+ ) + λ+ (α − α+ ) = αλ+ − A(λ+ ). (1.1.10)
Thus, for α α+ the function λ(α) is a constant, and (α) grows linearly.
Moreover, the relations (1.1.7), (1.1.8) remain valid.
If α+ = ∞, then α < α+ for all finite α α− , and again we are dealing
with the ‘regular’ situation considered before (see (1.1.7), (1.1.8)). Since λ(α) is
non-decreasing, those relations imply the convexity of (α).
In sum, we can formulate the next property.
(3) The functions λ(α), (α) may have discontinuities only at the points s± in
the case P(ξ = s± ) > 0. These points separate the domain (s− , s+ ) of finiteness
and continuity (in the extended sense) of the function from the domain α ∈
[s− , s+ ], where (α) = ∞. On [s− , s+ ] the function is convex. (If one defines
convexity in the ‘extended’ sense, i.e. allowing infinite values, then is convex on
the whole line.) The function is analytic on the interval (α− , α+ ) ⊂ (s− , s+ ). If
λ+ < ∞, α+ < ∞, then the function (α) is linear on (α+ , ∞) with a slope angle
of λ+ ; at the boundary point α+ the continuity of the first derivatives persists. If
λ+ = ∞, then (α) = ∞ on (α+ , ∞). An analogous property is valid for the
function (α) on (−∞, α− ).
If λ− = 0, then α− = a1 and λ(α) = (α) = 0 with α a1 .
In fact, since λ(a1 ) = 0 and ψ(λ) = ∞ for λ < λ− = 0 = λ(a1 ), with
the decrease of α down to α− = a1 , the point λ(α), having reached 0, stops, and
λ(α) = 0 for α α− = a1 . It follows from this and from the first identity in
(1.1.6) that (α) = 0 for α a1 .
If λ− = λ+ = 0 (the condition [C ] is not satisfied), then λ(α) = (α) ≡ 0
for all α. This is evident, since the value under the sup sign in (1.1.3) is equal to
−∞ for all λ = 0. In this case the limit theorems stated in the subsequent sections
would not be informative.
By summarising the properties of , we can conclude, in particular, that on the
whole line the function is
(a) convex: for α, β ∈ R, p ∈ [0, 1]
pα + (1 − p)β p(α) + (1 − p)(β); (1.1.11)
(b) lower semicontinuous:
lim (α) (α0 ), α0 ∈ R. (1.1.12)
α→α0
Properties (a), (b) are known as general properties of the Legendre transform of
a convex lower-semicontinuous function A(λ) (see e.g. [159]).
We will also need the following properties of function .
(4) Under trivial conventions about notation, for independent random variables
ξ and η we have
(ξ +η) (α) = sup αλ − A(ξ ) (λ) − A(η) (λ) = inf (ξ ) (γ ) + (η) (α − γ ) ,
λ γ

α−b
(cξ +b) (α) = sup αλ − λb − A(ξ ) (λc) = (ξ ) .
λ c
It is clear that the infimum infγ in the first relation is reached at a point γ , such
that λ(ξ ) (γ ) = λ(η) (α − γ ). If ξ and η are identically distributed, then γ = α/2
and, therefore,
α α α
(ξ +η) (α) = (ξ ) + (η) = 2(ξ ) .
2 2 2
It is also evident that for all n 2
α
αλ
(α) = sup αλ − nA (λ) = n sup
(Sn ) (ξ )
− A (λ) = n(ξ )
(ξ )
.
n n
(5) The function (α) attains its minimal value, which is equal to 0, at the point
α = E ξ = a1 . For definiteness, let α+ > 0. If a1 = 0, E |ξ k | < ∞, then
1 γ3
λ(0) = (0) = (0) = 0, (0) = , (0) = − , ... (1.1.13)
γ2 γ22
(in the case α− = 0, right derivatives are meant ). As α ↓ 0, the next representa-
tion takes place:

k
(j) (0)
(α) = α j + o(α k ). (1.1.14)
j!
j=2
The semi-invariants γj are defined in (1.1.1) and (1.1.2).

If the double-sided Cramér’s condition [C0 ] is met, then the expansion of (α)
into a series (1.1.14) for k = ∞ holds and is called a Cramér’s series.
The proof of property (5) should not cause any trouble, and we leave it to the
reader.
(6) If condition [C+ ] is met, then there exist constants c1 > 0 and c2 such that,
for all α,
(α) c1 α − c2 .
If λ+ = ∞, then (α) α as α → ∞, i.e. there exists a function v+ (α) → ∞,
as α → ∞, such that (α) αv+ (α).
Proof. By virtue of (1.1.10), λ(α) ↑ λ+ as α ↑ α+ , so λ(α) = λ+ for α α+ .
Since the function is convex, (α) = λ(α) ↑ as α ↑; then, having taken a point
α0 > 0, such that λ(α0 ) > 0, we obtain
(α) (α0 ) + λ(α0 )(α − α0 ) for all α.
If λ+ = ∞, then λ(α) ↑ ∞ as α ↑ ∞ and
α
1 1
v+ (α) := (α) = (0) + λ(t)dt → ∞ as α → ∞.
α α 0
The property (6) is proved.
If follows from this property that under condition [C0 ] there exist constants
c1 > 0, c2 > 0, such that
(α) c1 |α| − c2 for all α.
If |λ± | = ∞, then (α) |α| as |α| → ∞.
(7) An inversion formula holds: for λ ∈ (λ− , λ+ )
ln ψ(λ) = sup αλ − (α) . (1.1.15)
α
This means that when condition [C] is met, the deviation function uniquely
determines the Laplace transform of ψ(λ), and, hence, the distribution of a random
variable ξ . The formula (1.1.15) also indicates that the iterated Legendre transform
of the convex function ln ψ(λ) leads to the same original function.
Proof. Let us denote the right-hand side of (1.1.15) by T(λ) and show that T(λ) =
ln ψ(λ) for λ ∈ (λ− , λ+ ). If, in order to find the sup in (1.1.15), we set the
derivative with respect to α of the function under the sup sign equal to zero, then
we obtain the equation
λ = (α) = λ(α). (1.1.16)
Since λ(α) on (α− , α+ ) is the function inverse to ln ψ(λ) (see (1.1.4)), then for
λ ∈ (λ− , λ+ ) the equation (1.1.16) has an evident solution
α = a(λ) := ln ψ(λ) . (1.1.17)
Taking into account that λ a(λ) ≡ λ, we obtain
T(λ) = λa(λ) − a(λ) ,
T (λ) = a(λ) + λa (λ) − λ a(λ) a (λ) = a(λ).
Since a(0) = a1 and T(0) = −(a1 ) = 0,
λ
T(λ) = a(u)du = ln ψ(λ). (1.1.18)
0
The statement is proved, as well as another inversion formula (the last equality in
(1.1.18); this expresses ln ψ(λ) in terms of the integral of the function a(λ), which
is inverse to λ(α)).
By virtue of Chebyshev’s exponential inequality, for all n 1, λ 0, x 0,
we have
P (Sn x) e−λx ψ(λ) = exp −λx + n ln ψ(λ) . (1.1.19)
Since λ(α) 0 for α a1 , by setting α = x/n and by substituting λ = λ(α) 0
in (1.1.19) we obtain the property
(8) For all n 1 and α = x/n a1 ,
P(Sn x) e−n(α) .
The next property can be named as an exponential modification of the
Kolmogorov–Doob inequality.
(9) Theorem 1.1.2. (i) For all n 1, x 0 and λ 0, one has
P(Sn x) e−λx max 1, ψ n (λ) . (1.1.20)
(ii) Let Eξ < 0, λ1 := max λ : ψ(λ) 1 . Then, for all n 1 and x 0, one
has
P (Sn x) e−λ1 x . (1.1.21)
If λ+ > λ1 then ψ(λ1 ) = 1, (α) λ1 α for all α and (α1 ) = λ1 α1 , where
ψ (λ1 )
α1 := arg λ(α) = λ1 = , (1.1.22)
ψ(λ1 )
so that a line y = λ1 α is tangent to the convex function y = (α) at the
point (α1 , λ1 α1 ). In addition, along with (1.1.21), the next inequality holds
for α := x/n:
P (Sn x) e−n1 (α) , (1.1.23)
where

λ1 α for α α1 ,
1 (α) =
(α) for α > α1 .
If α α1 then the inequality (1.1.23) coincides with (1.1.21); for α > α1 it
is stronger than (1.1.21).
(iii) Let Eξ 0, α = x/n Eξ . Then, for all n 1,
P (Sn x) e−n(α) . (1.1.24)
Theorem 1.1.2 distinguishes three non-overlapping possibilities,

(a) Eξ < 0, λ+ = λ1 ,
(b) Eξ < 0, λ+ > λ1 ,
(c) Eξ 0,
for which P (Sn x) is bounded by the right-hand sides of inequalities (1.1.21),
(1.1.23), (1.1.24) correspondingly. However, by accepting some natural conven-
tions, one can express all three stated inequalities in the unique form of (1.1.23).
Indeed, let us turn to the definition (1.1.22) of α1 . As already noted (see (1.1.4)),
λ(α) is a solution of the equation ψ (λ)/ψ(λ) = α, which is unique for
ψ (λ) ψ (λ)
α ∈ [α− , α+ ], α+ := lim , α− := lim .
λ↑λ+ ψ(λ) λ↓λ− ψ(λ)
For α α+ the function λ(α) is defined as a constant λ+ (see (1.1.10)). This
means that for λ1 = λ+ the value of α1 is not uniquely defined and may take any
value from α+ to ∞, so that, by setting α1 = max α : λ(α) = λ1 = λ+ = ∞,
we turn the inequality (1.1.23) in the case λ1 = λ+ (i.e. in case (a)) into the
inequality
P (Sn x) e−nλ1 α = e−λ1 x ,
i.e. into the inequality (1.1.21).
If Eξ 0 then λ1 = 0. If λ+ = 0 then λ+ = λ1 and we have the same situation
as before, but now P (Sn x) allows only the trivial bound of 1. If λ+ > 0 then
α1 = Eξ ; for α < α1 the bound (1.1.23) is trivial again, and for α > α1 it coincides
with (1.1.24).
Corollary 1.1.3. If we set
⎧
⎨ ψ (λ1 ) , for λ > λ ,
+ 1
α1 := max α : λ(α) = λ1 = ψ(λ1 )
⎩
∞, for λ+ = λ1 ,
then the inequality (1.1.23) holds without any additional conditions on Eξ and λ+
and comprises inequalities (1.1.21), (1.1.24).
From the results of [16] (where the asymptotics of the distributions of the
maximum of sequential sums of random variables are studied; see also Chapter 3)
it is not difficult to extract the information that the exponents in the inequality
(1.1.23) are asymptotically unimprovable:
1 x
lim ln P (Sn x) = −1 (α) as = α;
n→∞ n n
the same can be deduced using the large deviation principle for trajectories of {Sk }
(see Chapter 4).
Proof of Theorem 1.1.2. (i) The random variable
η(x) := inf{k > 0 : Sk x}
is a stopping time. Thus, the event {η(t) = k} and the random variable Sn − Sk are
independent and

n
n
ψ n (λ) = EeλSn E(eλSn ; η(x) = k) E eλ(x+Sn −Sk ) ; η(x) = k
k=1 k=1

n
= eλx ψ n−k (λ)P η(x) = k eλx min 1, ψ n (λ) P η(x) n .
k=1
Hence we obtain statement (i).
(ii) Inequality (1.1.21) directly follows from (1.1.20), if one takes λ = λ1 . Now
let λ+ > λ1 . Then, obviously, ψ(λ1 ) = 1 and from the definition of the function
(α) it follows that
(α) λ1 α − ln ψ λ1 = λ1 α.
Furthermore,
(α1 ) = λ1 α1 − ln ψ λ1 = λ1 α1 ,
so that the curves y = λ1 α and y = (α) are tangent to each other at the point
(α1 , λ1 α1 ).
Next, it is clear that ψ λ(α) 1 for α α1 . For α = x/n, the optimal choice
of λ in (1.1.20) would be λ = λ(α). For such an λ(α), i.e. α = x/n, we obtain
P (Sn x) e−n(α) .
Together with (1.1.21) this proves (1.1.23). It is also clear that (α) > λ1 α for
α > λ1 , which proves the last statement of (ii).
(iii) Since λ(a1 ) = 0 and λ(α) is non-decreasing, λ(α) 0 for α a1 . In the
case a1 0 one has ψ(λ) 1 for λ 0. Thus ψ λ(α) 1 for α a1 . By
substituting λ = λ(α) in (1.1.20) for α a1 , one obtains (1.1.24). Theorem 1.1.2
is proved.
The probabilistic sense of the deviation function is clarified by the following
statement. Let (α)ε = (α − ε, α + ε) be an ε-neighbourhood of α. For any set
B ⊂ R denote
(B) = inf (α).
α∈B
Theorem 1.1.4. For all α ∈ R and ε > 0,

1 Sn
lim ln P ∈ (α)ε = − (α)ε , (1.1.25)
n→∞ n n

1 Sn
lim lim ln P ∈ (α)ε = −(α). (1.1.26)
ε→0 n→∞ n n
The statement (1.1.26) in Theorem 1.1.4, referred to as the local large deviation
principle, allows another forms. It will be shown in section 4.1 that the statement
(1.1.26) is equivalent to the following: given any α, for any sequence εn that
converges to zero sufficiently slowly, one has

1 Sn
(α) = − lim ln P ∈ (α)εn . (1.1.27)
n→∞ n n
This relation can be expressed in the form

Sn
ln P ∈ (α)εn = −n(α) + o(n) as n → ∞.
n
Thus the dependence on n and α of the left-hand side of this equality is asymptot-
ically factorisable: it is presented as a product of the factors n and (α), where
the former depends only on n, while the latter depends only on α.
Theorem 1.1.4 will be proved in section 2.7. Large deviation principles will be
discussed in detail in Chapters 4 and 5 in a more general setting. Note that for
a sequence of sums Sn , the so-called moderately large deviation principle holds;
this is formulated and proved in Chapter 5.
Now let us consider a few examples in which the values of λ± and α± and the
functions ψ(λ), λ(α), (α) can be calculated in an explicit form.
Example 1.1.5. If ξ is normally distributed with parameters (0, 1) then
2 /2 α2
ψ(λ) = eλ , |λ± | = |α± | = ∞ λ(α) = α, (α) =
.
2
Example 1.1.6. For the Bernoulli scheme with parameter p ∈ (0, 1) we have
ψ(λ) = peλ + q, |λ± | = ∞, α+ = 1, α− = 0, a1 = E ξ = p,

α(1 − p) α 1−α
λ(α) = ln , (α) = α ln + (1 − α) ln for α ∈ (0, 1),
p(1 − α) p 1−p
(0) = − ln(1 − p), (1) = − ln p, (α) = ∞ for α ∈ [0, 1].
Example 1.1.7. For an exponential distribution with rate β we have
β 1
ψ(λ) = , λ+ = β, λ− = −∞, α+ = ∞, α− = 0, a1 = ,
β −λ β
1
λ(α) = β − , (α) = αβ − 1 − ln αβ for α > 0.
α
Example 1.1.8. For a centred Poisson distribution with rate β we have
ψ(λ) = exp β[eλ − 1 − λ] , |λ± | = ∞, α− = −β, α+ = ∞, a1 = 0,

β +α α+β
λ(α) = ln , (α) = (α + β) ln −α for α > −β.
β β
1.2 Deviation function and its properties in the multidimensional case 13
1.2 Deviation function and its properties in the multidimensional case

1.2.1 Cramér’s conditions
Now let ξ , ξ1 , ξ2 , . . . be a sequence of independent identically distributed random
vectors in a d-dimensional Euclidean space Rd . For vectors α = (α(1) , . . . , α(d) )
and β = (β(1) , . . . , β(d) ) from Rd we denote the dot product and the norm
respectively by
α, β = α(1) β(1) + · · · + α(d) β(d) , |α| = α, α1/2 .
As before, let S0 = 0, Sn = ξ1 + · · · + ξn for n 1. We assume that the
distribution of vector ξ , as in section 1.1, is not degenerate, i.e. there is no b ∈ Rd ,
b = 0, such that ξ , b = const. with unit probability.
The Laplace transform of the distribution of a random vector ξ is denoted, as
before, by
ψ(λ) := Eeλ,ξ , λ ∈ Rd .
In the multidimensional case we use Cramér’s conditions in the following form:
[C] There exists λ ∈ Rd such that ψ(λ) < ∞ in some vicinity of λ.
[C0 ] Condition [C] is met for λ = 0.
[C∞ ] Condition [C] is met for all λ ∈ Rd .
It is evident that [C∞ ] ⊂ [C0 ] ⊂ [C].
1.2.2 Deviation function

The deviation function of a random vector ξ , which is also known as the Legendre
transform of the function
A(λ) := ln ψ(λ)
or as the conjugate function to A(λ), in convex analysis (cf. [159]), is defined by
(α) = (ξ ) (α) := sup α, λ − A(λ) , α ∈ Rd . (1.2.1)
λ
Let us clarify the main properties of the deviation function.
−
→
( 1) Young’s inequality: For all α and λ from Rd ,
A(λ) + (α) α, λ. (1.2.2)
This inequality follows immediately from the definition. By putting λ = 0
in (1.2.2), we get (α) 0 for all α ∈ Rd .
An important role in the multidimensional case is played (just as before) by
the sets
A = λ : A(λ) < ∞ ,

∂A(λ) ∂A(λ)
A = A (λ), λ ∈ A , where A (λ) = ,..., = grad A(λ),
∂λ(1) ∂λ(d)
and the convex envelope S of the support of the distribution of ξ .
The following properties of the function (α) hold.
−
→
( 2) (α) = ∞ for α ∈ S; A ⊂ S. If the support S is bounded then A = Rd ,
A = S. For all α ∈ Rd there exists a unique point λ(α) ∈ Rd such that
(α) = λ(α), α − A λ(α) .
The functions λ(α) and (α) are analytic on A ,

2
∂ (α)
(a1 ) = 0, λ(a1 ) = grad (α)|α=a1 = 0, (a1 ) :=
∂α ∂α
= M −1 ,
(i) (j) α=a1
where a1 = Eξ and M −1 is the inverse covariance matrix
M = Eξ(i)
0 0
ξ(j) , 0
ξ(i) = ξ(i) − Eξ(i) , i = 1, . . . , d.
If the vector ξ has a normal distribution with mean a1 and covariance matrix M
then it is easy to verify that
1
(α) = (α − a1 )M −1 (α − a1 )T .
2
−
→
( 3) The function (α) is convex: for α, β ∈ Rd , p ∈ [0, 1] one has (1.1.11).
The sets A and v = α : (α) v for all v 0 are convex.
−
→
( 4) The function (α) is lower semicontinuous everywhere:
lim (α) (α0 ), α0 ∈ Rd . (1.2.3)

α→α0
−
→
( 5)
(αt)
∞ (α) := lim = sup λ, α. (1.2.4)
t→∞ t λ∈A
−
→
( 6) An inversion formula holds: for all λ ∈ Rd
ln ψ(λ) = sup α, λ − (α) .

α
Thus, as in the one-dimensional case, the deviation function (α) under [C]
uniquely defines the distribution of ξ .
The next property can be named as a consistency property. Under natural
conventions about notation one has
−
→
( 7)
(ξ(1) ) (α(1) ) = inf (α)

α(2) ,...,α(d)
and, more generally, for all b ∈ Rd ,
(ξ ,b) (β) = inf (α) : α, b = β .

−
→ −
→
Proofs of the properties ( 2)–( 7) can be found in [159], where general
properties of the Legendre transform of convex lower-semicontinuous functions
are studied. In the same book and in [48] the consistency property can be found in
a more general form, i.e. in the form of the deviation function of a random vector
ξ H + b, where H is a matrix of order (d, r).
Directly from the definition of the function (α) we obtain the property
−
→
( 8) (1) Let b ∈ Rd and let H be an invertible square matrix of order d. Then
(ξ H+b) (α) = (α − b)H −1 .

α
(2) (Sn ) (α) = n .
n
The proof is the same as for property (4) in the one-dimensional case.
−
→
Property ( 5) allows one to obtain the following property:
−
→
( 9) If [C0 ] is met then there exist constants c1 > 0 and c2 such that
(α) c1 |α| − c2 .
If [C∞ ] is met then
(α) |α| as |α| → ∞.

−
→
The first statement follows from ( 5) and the fact that under [C0 ] the point
λ = 0 is an inner point of A and, therefore, that the right-hand side of (1.2.4) is
uniformly positive over α. Under [C∞ ] one has A = Rd , and the right-hand side
of (1.2.4) is infinite.
The probabilistic sense of the deviation function remains the same (see Theo-
rem 1.1.4):
Theorem 1.2.1. The statements of Theorem 1.1.4 remain valid in the case of a
d-dimensional random vector ξ , d > 1.
Our commentaries on Theorem 1.1.4 also remain valid, including the statement
of the equivalence of relations (1.1.25) and (1.1.27). Theorem 1.2.1 will be proved
in section 2.7.
−
→
In the next section we will need a strengthening of the property ( 4). Denote
by <∞ = α : (α) < ∞ the set for which is finite. The function (α)
is continuous in the interior (<∞ ) of <∞ (this property is inherent to any
convex function on Rd ), and lower semicontinuous, if the boundary is included.
It turns out that the latter property can be strengthened up to lower continuity
and continuity inside <∞ on the boundary ∂<∞ (see Lemma 1.2.2 below).
Hereafter such an extension will be required for convex lower-semicontinuous
functions J = J(y) 0 defined on an arbitrary linear metric space (Y, ρ).
Hence we will formulate and prove Lemma 1.2.2 in the general case, for which the
functions have the aforementioned properties in an arbitrary metric space (Y, ρ).
Let the function J = J(y) : Y → [0, ∞] be convex and lower semicontinuous.
Since the function J is convex, the set J<∞ := {y ∈ Y : J(y) < ∞} and its closure
[J<∞ ] are convex. For any set B ⊂ Y denote
J(B) := inf J(y).
y∈B
Lemma 1.2.2. Let (Y, ρ) be an arbitrary linear metric space, and let a function
J : Y → [0, ∞] be lower semicontinuous and convex, y ∈ [J<∞ ], y0 ∈ J<∞ . Then
the following properties hold.
(i) (Lower continuity)
lim J((y)ε ) = J(y). (1.2.5)
ε→0
(ii) (Continuity inside J<∞ along rays)

lim J((1 − p)y0 + py) = J(y). (1.2.6)
p↑1
For a set B ⊂ Rd we denote by (B) and [B] the interior and the closure of B
respectively.
Corollary 1.2.3. Let B ⊂ Rd be a convex set and let the function J be lower
semicontinuous and convex, J((B)) < ∞. Then
J([B]) = J(B) = J((B)). (1.2.7)
Proof of Lemma 1.2.2. (i) Property (1.2.5) is a consequence of the lower semi-
continuity of J (cf. (1.2.3)) and the relation (1.2.6). Indeed, by virtue of the lower
semicontinuity, we have on the one hand
lim J((y)ε ) J(y).
ε→0
On the other hand, choose y0 ∈ J<∞ and let yp := (1 − p)y0 + py, so that
y = y1 . According to the properties of a linear metric space (see [112], p. 23)
we have yp → y1 as p → 1. Hence for every ε > 0 there exists an arbitrarily large
p = p(ε) < 1 such that yp(ε) ∈ (y)ε and p(ε) → 1 as ε → 0. Therefore, by virtue
of (1.2.6),
lim J((y)ε ) lim J(yp(ε) ) = J(y).
ε→0 ε→0
Property (1.2.5) is proved.

(ii) Let us prove property (1.2.6). If J(y) = ∞ then (1.2.6) follows from (1.2.3)
for J. If J(y) < ∞ then consider the function
g(p) := J(yp ), p ∈ [0, 1].
The function J is convex and therefore on the one hand
g(p) pJ(y1 ) + (1 − p)J(y0 ), lim g(p) J(y1 ).
p→1
On the other hand, by virtue of (1.2.3) and the convergence yp → y1 as p → 1,
we have
lim g(p) J(y1 ).

p→0
This proves (1.2.6). Lemma 1.2.2 is proved.

Proof of Corollary 1.2.3. (i) Note the following in advance. Since B and J<∞ are
convex sets, the sets [B], [B] ∩ J<∞ are also convex. According to the condition,
J((B)) < ∞ and, therefore, the set (B) ∩ J<∞ is not empty. Let us take points
y0 ∈ (B) ∩ J<∞ , y1 ∈ [B] ∩ J<∞ and show that a half-open interval
[y0 , y1 ) := {yp := py1 + (1 − p)y0 ; p ∈ [0, 1)} is a subset of (B). (1.2.8)
Both the points y0 and y1 belong to the convex set [B] ∩ J<∞ . Hence, yp also
belongs to that set for all p ∈ [0, 1] and, in order to prove (1.2.8), we have to
exclude the possibility yp ∈ ∂B for p ∈ [0, 1). But if yp ∈ ∂B then there exists a
sequence of points v → yp such that v ∈ [B]. In this case,
v − py1
w := → y0 as v → yp .
1−p
Since y0 ∈ (B), w ∈ (B) as v is close enough to yp . Since v = py1 + (1 − p)w and
y1 ∈ [B], v ∈ [B]. The derived contradiction proves (1.2.8).
(ii) Now let us turn to the direct proof of (1.2.7). Statement (1.2.8) implies the
inequality
J((B)) J(py1 + (1 − p)y0 ), 0 p < 1. (1.2.9)
Since y0 , y1 belong to J<∞ , by virtue of statement (ii) of Lemma 1.2.2, the right-
hand side of (1.2.9) converges to J(y1 ) as p → 1. Therefore,
J((B)) J(y1 ).
Hence, due to the arbitrariness of y1 ∈ [B] ∩ J<∞ , we obtain the inequality
J((B)) J([B]),
which, along with the evident relations
J((B)) J(B) J([B]),
proves (1.2.7). The corollary is proved.

Remark 1.2.4. In Corollary 1.2.3, the condition J((B)) < ∞ cannot be replaced
with the condition J(B) < ∞, as evidenced (in the case Y = Rd , J = ) by the
following example.
Example 1.2.5. A random vector ξ ∈ R2 takes three values, β (1) = (0, 2), β (2) =
(1, 1), β (3) = (1, −1) with the probabilities p(1) = 1/2, p(2) = 1/4, p(3) = 1/4,
correspondingly. By virtue of the local large deviation principle (1.1.27) (see
Theorem 1.2.1), one has
(β (i) ) = − ln p(i) , i = 1, 2, 3.
Now consider a set B = B1 ∪ {β (2) }, where B1 is an open triangle with vertices
β (1) , β (2) , β (4) = (1, 2). Since (α) = ∞, if a point α lies outside the closed
triangle with vertices β (1) , β (2) , β (3) then (B) = − ln p(2) = ln 4. In addition,
([B]) (β (1) ) = ln 2. Therefore, we have
([B]) < (B) < ∞.
The equalities (1.2.7) fail.
1.3 Chebyshev-type exponential inequalities for sums of random vectors

1.3.1 Basic inequalities for random vectors
Consider an arbitrary measurable set B ⊂ Rd and, as before, put
(B) := inf (α).
α∈B
If the set B is empty, we put (B) = ∞. For an arbitrary set B ⊂ Rd we will

denote by (B) the interior of B, i.e. the collection of all points α that belong to
B along with a neighbourhood, and by [B] the closure of the set B. For a v 0
consider the set
v := {α : (α) v},
of all points at which the value of the deviation function does not exceed v. We
will denote the set on which the deviation function is finite by

<∞ := v = {α : (α) < ∞}.
v0
Note that in the one-dimensional case d = 1 Chebyshev’s exponential inequal-

ity (see property (8) in section 1.1) for convex sets B can be written in the form
P(ξ ∈ B) e−(B) . (1.3.1)
Indeed, if (B) = 0 then the inequality is trivial. If (B) > 0 then either b+ :=
inf{α : α ∈ B} > Eξ or b− := sup{α : α ∈ B} < Eξ . If b+ > Eξ , b+ ∈ B, then
(b+ ) = (B) > 0 and the relation (1.3.1) follows from Chebyshev’s inequality
P(ξ ∈ B) P(ξ b+ ) e−(b+ ) . (1.3.2)
If (b+ ) < (B) (which is possible only if b+ ∈ B, P(ξ b+ ) = 1 and has
a discontinuity at point b+ ) then (B) = ∞, P(ξ ∈ B) = 0 and inequality (1.3.1)
holds as before. The case b− < Eξ is considered analogously. It is clear that ‘one-
sided’ sets B (i.e. those lying wholly on one side of Eξ ) can be embedded in convex
sets with the same bounds b± , and the inequality (1.3.1) holds for them as well.
1.3 Exponential inequalities for sums of random vectors 19
We return to the general case d 1. The main assertions of the present section
are stated below.
Theorem 1.3.1. If B is an arbitrary open convex set then
P(ξ ∈ B) e−(B) . (1.3.3)

Theorem 1.3.2. For an arbitrary set B, one has
P(ξ ∈ B) e−([B
con ])
, (1.3.4)
where Bcon is the convex envelope of B.

It follows from Theorems 1.3.1 and 1.3.2 that inequality (1.3.3) holds for both
open convex and closed convex sets. Thus, selecting a class of convex sets and
using the deviation function allows one to find a unified simple form (1.3.3) of
the inequality, which it is still natural to call Chebyshev’s inequality and which
holds in spaces of any dimension, including the infinite-dimensional case (see
section 4.3 below).
Theorems 1.3.1 and 1.3.2 are consequences of a more general assertion
(Theorem 1.3.4) that employs broader conditions ensuring that inequalities of
the form (1.3.3) hold true. Verifying these conditions in spaces of high dimension
can be quite difficult. For this reason, we have stated simpler versions of
Theorems 1.3.1, 1.3.2 as our main results. These theorems will suffice for the
purposes of forthcoming sections and for some applications as well.
To state Theorem 1.3.4, we will need several concepts.
If (B) = 0 then also ([B]) = 0, and all the assertions we have already stated
before and will state hereafter would be trivial. Therefore, it would be assumed
throughout the section that (B) > 0. This means that we also assume that
Cramér’s condition [C] is met:
Eeλ,ξ < ∞ for λ taking values in a body (i.e. a compact set with non-empty
interior).
In this case the set <∞ will also be a body.
To simplify the preliminaries, we will assume for the present that the stronger
condition [C0 ] is met.
In this case one can assume without loss of generality that Eξ = 0, ‘shifting’ the
set B if necessary. Then (α) → ∞ as |α| → ∞, and the sets v form a family of
increasing sets (as v increases) that eventually fill the set <∞ , which coincides
(up to its boundary) with the convex envelope S of the support of the distribution
of ξ . As the boundary of v approaches that of <∞ (as v increases), the growth
of v in that direction slows down or stops altogether, so that the boundaries
of the sets v and <∞ may coincide in some locations. The boundaries ∂v
of the sets v are called level surfaces. One should note that the level surfaces
corresponding to large enough v may have ‘cavities’; i.e. for some directions e,
the equation (te) = v can have no solutions.
Let us state the conditions for Theorem 1.3.4. First let (B) < ∞, i.e. there
exist points α ∈ B for which (α) < ∞. In this case, under the condition [C0 ],
there exists a v = vB at which, by virtue of the continuity properties of the function
(noted above in Lemma 1.2.2) and the monotone growth of the compacts v ,
these compacts will touch the set [B] for the first time. In other words, there exists
a point αB , at which the minimum
min (α) = ([B]) = (αB ) =: vB

α∈[B]
is attained. It is clear that the point αB lies on the boundaries of the sets vB and
[B], and that the sets touch each other: αB ∈ vB ∩ [B] = ∅.
If only condition [C] is met then, in the case where the set B is unbounded, the
sets vB and [B] may ‘touch’ each other ‘at infinity’.
If (B) = ∞ then the set B does not intersect the convex set <∞ .
The symbol e will be used for unit vectors.
Definition 1.3.3. Let (B) < ∞. In this case, the set B will be called -separable
if there exists a hyperplane = := {α : e, α = b} (going through the point αB
if the latter exists) that separates the sets B and vB = {α : (α) vB } in the
following sense:
B ⊂ > := {α : e, α > b}, vB ⊂ := {α : e, α b}. (1.3.5)
If (B) = ∞ then the set B is called -separable if there exists a hyperplane

= that separates the sets B and <∞ :
B ⊂ > , <∞ ⊂ .
Theorem 1.3.4. If B is a -separable set then
P(ξ ∈ B) e−([B]) . (1.3.6)

Clearly, a convex set B with (B) < ∞ cannot be -separable.
Note that -separability is essential for the assertion of Theorem 1.3.4 to hold
(without that assumption, inequality (1.3.6) is, generally speaking, wrong), but
this property is quite restrictive. What bounds can one obtain for P(ξ ∈ B) if the
set B is not -separable?
If B ⊂ ∪Bk is a subset of the union of an at most countable collection of convex
sets Bk then Theorem 1.3.2 implies that

P(ξ ∈ B) e−([Bk ]) . (1.3.7)
Theorems 1.3.2 and 1.3.4 do not provide any further meaningful bounds. Using
a somewhat different approach, we can consider a random variable γ := (ξ )
and an ‘iterate’ deviation function (γ ) , i.e. the deviation function for a random
variable γ which is equal to the value of the deviation function = (ξ ) (α) at
the point ξ . Properties of the random variable γ and the deviation function (γ )
are studied in section 1.4.
Theorem 1.3.5. If (B) Eγ then
P(ξ ∈ B) e−
(γ ) ((B))
.
1.3.2 Proofs of Theorems 1.3.1, 1.3.2, 1.3.4 and 1.3.5

−
→
Proof of Theorem 1.3.4. Put β = e, ξ . The property ( 7) can be written as
(β) (v) = inf (α). (1.3.8)

e,α=v
(i) Let the set B be -separable and (B) = ∞. Then > ∩ <∞ = ∅ and so
(> ) = ∞. In other words, (α) = ∞ for all α ∈ > . It follows from this and
(1.3.8) that (β) (v) = ∞ for all v > b and therefore (β) ((b, ∞)) = ∞.
Further, by virtue of (1.3.5), one has
P(ξ ∈ B) P(ξ ∈ > ) = P(e, ξ > b) = P(β > b). (1.3.9)
Since P(β > b) = limk→∞ P(β b + 1/k), Chebyshev’s exponential inequality

(1.3.2) yields
P(β > b) lim e− = e−

(β) (b+1/k) (β) (b+0)
, (1.3.10)
k→∞
where (β) (b + 0) = (β) ((b, ∞)) = ∞. Therefore, P(ξ ∈ B) = 0, and so

inequality (1.3.6) is proved.
(ii) Now let (B) < ∞. It follows from (1.3.8) that
(β) ((b, ∞)) = inf (α) = (> ).

>
Therefore, once again using inequalities (1.3.9), (1.3.10), we get
P(ξ ∈ B) e− = e−(> ) .

(β) ((b,∞))
But the sets vB and > are disjoint. Hence (α) > vB for any α ∈ > . This
means that (> ) vB . Inequality (1.3.6) is established and Theorem 1.3 is
proved.
Note that if the set B is not -separable then, generally speaking, the equality
([B]) = ∞ does not imply that P(ξ ∈ B) = 0 (recall that the set <∞ coincides
up to its boundary with the convex envelope of the support of the distribution of
ξ ). This is demonstrated by the following example.
Example 1.3.6. Let be a unit sphere in Rd , d 2, and let a random vector ξ
have a uniform distribution on . Also let B be the closure of the exterior. Then
⊂ B, ([B]) = ∞ and P(ξ ∈ B) = 1.
Proof of Theorem 1.3.1. Since the sets v , <∞ are convex, according to the
Hahn–Banach theorem (see e.g. [110], p. 137) the open convex set B is -
separable and so by Theorem 1.3.4 one has (1.3.6) for that set. It remains to make
use of Corollary 1.2.3, by virtue of which the right-hand side of (1.3.6) coincides
with the right-hand side of (1.3.3). Theorem 1.3.1 is proved.
Proof of Theorem 1.3.2. Since B ⊂ Bcon , it suffices to verify that (1.3.4) holds for
any convex set B = Bcon .
The set BN := {α ∈ B : |α| N} is convex, along with B. The ε-neighbourhood
(BN )ε of the set BN is also convex. Therefore, by Theorem 1.3.1,
P(ξ ∈ B) P(ξ ∈ (BN )ε ) + PN e−((BN )ε ) + PN , PN := P(|ξ | > N).
Now let ε = εk → 0 as k → ∞, even out the spaces on this line and let αk be
a sequence of points from (BN )εk ⊂ [(BN )εk ] such that (αk ) ((BN )εk ) +
1/k. Assuming without loss of generality that αk converge to α0 ∈ [BN ] as k →
∞, we obtain from the lower semicontinuity of the function that
P(ξ ∈ B) lim e−((BN )εk ) + PN lim e−(αk )+1/k
k→∞ k→∞
−(α0 ) −([BN ])
+ PN e + PN e + PN . (1.3.11)
Since ([BN ]) ↓ ([B]) as N ↑ ∞, passing to the limit in (1.3.11) as N ↑ ∞ we
obtain (1.3.6). Theorem 1.3.2 is proved.
Proof of Theorem 1.3.5. Since B ⊂ {α : (α) (B)}, one has, for (B) Eγ ,
that
P(ξ ∈ B) P((ξ ) (B)) = P(γ (B)) e−
(γ ) ((B))
.
The theorem is proved.
1.3.3 Inequalities for sums of random vectors

As before, set

n
Sn := ξi for n 1.
i=1
Put

Sn λ
ζn := , A(ζn ) (λ) := ln Eeλ,ζn = nA .
n n
−
→
Then the deviation function (ζn ) (α) for ζn , by virtue of property ( 8) is equal to
(ζn ) (α) = n(α). (1.3.12)
Thus, the following assertion immediately follows from Theorems 1.3.1 and 1.3.2.
Corollary 1.3.7. (i) For an arbitrary open convex set B one has the inequality
P(ζn ∈ B) e−n(B) . (1.3.13)
(ii) For an arbitrary measurable set B, one has the inequality
P(ζn ∈ B) e−n([B
con ])
,
where Bcon is the convex envelope of B.
From the results contained in this monograph (see below), it follows that the
bounds in Corollary 1.3.7(i) are ‘exponentially’ unimprovable. Such exact bounds
for the probability P(ζn ∈ B) for arbitrary sets B cannot be found. However, the
following assertion holds true. As before, we let γ = (ξ ).
Corollary 1.3.8. If (B) Eγ then
P(ζn ∈ B) e−n
(γ ) ((B))
. (1.3.14)
Inequality (1.3.14) is not ‘exponentially’ unimprovable owing to the ‘losses’ in
the first inequality in (1.3.15) (see below). This is also indicated by the inequality
(γ ) (v) < v − Eγ (see (1.4.8) below), which holds for random variables ξ that
are unbounded from above. However, for large v one has (γ ) (v) ∼ v, and so
inequalities (1.3.13), (1.3.14), in a certain sense, converge (i.e. (γ ) ((B)) ∼
(B) for large (B)).
Proof of Corollary 1.3.8. By virtue of the convexity of the deviation function
(α), one has
n
n(ζn ) γk , γk := (ξk ).
k=1
Since B ⊂ {α : (α) (B)}, we have the implications

n
{ζn ∈ B} ⊂ {(ζn ) (B)} ⊂ γk n(B) .
k=1
From this relation, in the case where (B) Eγ , using Chebyshev’s exponential
inequality we obtain
n
γk n(B) e−n ((B)) .
(γ )
P(ζn ∈ B) P (1.3.15)
k=1
Corollary 1.3.8 is proved.
1.3.4 Appendix1. Strengthening of exponential inequalities

for non-convex sets
As was noted before, the inequality (1.3.14) in Corollary 1.3.8 is not ‘exponen-
tially’ unimprovable. In [20] a strengthening of that inequality was obtained where
the ‘exponential’ component has the form e−n(B) , the same as for convex sets B
but with a multiplier which is exponential with respect to n. Let Bu = α : (α)
u}, so that (Bu ) ≡ u, and let ∂Bu be a boundary of the set Bu , Vu be a ‘volume’
of ∂Bu (the area for d = 3, the length for d = 2). The set Bu is the complement
to a convex set and is the ‘least convenient’ for obtaining upper bounds for the
probability P(ζn ∈ Bu ).
Theorem 1.3.9. [20] Let the condition [C0 ] be met. Then for every ε > 0 and for
all sufficiently large un one has

eλu n d−1/2 −un
P(ζn ∈ Bu ) (1 + ε)Vu e , (1.3.16)
4ru

where ru = supα∈∂Bu |α|, λu = supα∈∂Bu λ(α),
If u → 0, un 1, as n → ∞ then for all sufficiently large n the inequality
(1.3.16) can be written in the form
√
2 π πe (d−1)/2 μ1 3(d−1)/2
P(ζn ∈ Bu ) (1 + ε) (un)(d−1)/2 e−un ,
(d/2) 2 μd
where μ21 and μ2d are respectively the largest and the smallest eigenvalues of the
matrix
2
∂ (α)
M −1 = ∂α ∂α
,
(i) (j) α=Eξ
which is the inverse to the matrix M = Eξ(i)

0 ξ 0 , where ξ 0 = ξ − Eξ , of
(j) (i) (i) (i)
central second moments.
It is clear that for any B ⊂ Rd the probability P(ζn ∈ B) is bounded by the
right-hand side of (1.3.16) with u = (B). For small u the boundary ∂Bu is close
to the ellipse αM −1 α T = u.
Theorem 1.3.9 is proved in [20] by means of majorisation of the left-hand side of
(1.3.16) with P(ζn ∈ Bu ), where
Bu is a complement to the polyhedron ∂ Bu located
between ∂Bu and ∂Bu− , > 0. The boundary ∂ Bu has, as n grows, a growing
number of faces (in order to make → 0), and the probability P(ζn ∈ Bu ) is
bounded by the sum of the probabilities of reaching the corresponding subspaces
(see (1.3.7)), which allows simple exponential bounds.
Note that the inequalities in Theorem 1.3.9, in contrast with the exact inequali-
ties met in previous sections, are asymptotic.
From the integro-local theorem 2.3.2, obtained below in section 2.3 for sums Sn
(with some additional conditions), it is not difficult to find by means of integration
the exact asymptotic behaviour of the probability P(ζn ∈ Bu ) for the ‘least
convenient’ sets Bu . It has the form
cu nd/2−1 e−un .
Comparison with (1.3.16) shows that the error of inequality (1.3.16) is of order
√
n as n → ∞ – the same as for the exact inequalities in Theorems 1.3.1–
1.3.4.
1.4 The random variable γ = (ξ ) and its deviation function 25
1.4 Properties of the random variable γ = (ξ )

and its deviation function
1.4.1 Invariance of γ under linear transformations of ξ
−
→
By virtue of property ( 8) from subsection 1.2.2, for a fixed vector b ∈ Rd and
square non-singular matrix H the following relations hold true:
(ξ +b) (α) = (ξ ) (α − b), (ξ H) (α) = (ξ ) (αH −1 ),
where H −1 is the inverse to the matrix H. Hence
(ξ +b) (ξ + b) = (ξ ) (ξ + b − b) = γ ,
i.e. the value of γ does not change under the shift ξ . Similarly, we get
(ξ H) (ξ H) = (ξ ) (ξ HH −1 ) = γ ,
i.e. the value of γ does not change under ‘rotation and contraction’ of the vector
ξ . Thus, a linear transformation of the vector ξ does not affect the value of γ .
1.4.2 Properties of the random variable γ and the function (γ )

in the one-dimensional case, d = 1
Let s± be the boundaries of the set <∞ or, equivalently, the boundaries of the
convex envelope S of the support for the distribution of ξ , ± = (s± ), ∗ =
max{+ , − }, ∗ = min{+ , − }.
Theorem 1.4.1. (i) The distribution of a random variable γ = (ξ ) satisfies
the inequalities
⎧ −v
⎨ 2e , if v ∗ ,
P(γ v) e−v , if v ∈ (∗ , ∗ ], (1.4.1)
⎩ ∗
0, if v > .
(ii) The value Eγ satisfies the inequalities
∗
Eγ 2 − e−∗ − e− 2.
(iii) The following dichotomy holds: either ∗ = ∞ and then
(γ ) (v) ∼ v as v → ∞
(or, equivalently, λ+ := sup{t : Eetγ < ∞} = 1), or ∗ < ∞, and

(γ )
then, in the case where ∗ = + , one has s+ < ∞, γ + (and hence,

λ+ = ∞), P(γ = + ) = P(ξ = s+ ) = e−+ . Similar relations hold true
(γ )
in the case where ∗ = − .

Proof. (i) Consider the equation
(α) = v > 0.
If v ∗ then, by virtue of the convexity of the function (α) and its continuity
inside [s− , s+ ], there exist two solutions α± (v) of that equation; α+ (v) > Eξ
and α− (v) < Eξ . If v ∈ (∗ , ∗ ] then there exists only one solution: α+ (v) (if
∗ = − ) or α− (v) (if ∗ = + ). In that case, we will introduce the second
‘solution’, which does not exist, by setting it equal to ∓∞. If v > ∗ then there
are no solutions, and we set α± (v) = ±∞. It follows that {(ξ ) v} is the union
of two disjoint events {ξ α− (v)}, {ξ α+ (v)} (these events may be empty).
Therefore, for γ = (ξ ) we obtain
P(γ v) = P(ξ α− (v)) + P(ξ α+ (v)),
so that, by Chebyshev’s inequality,

⎧ −v
⎨ 2e , if v ∗ ,
P(γ v) e−(α− (v)) + e−(α+ (v)) e−v , if v ∈ (∗ , ∗ ],
⎩
0, if v > ∗ .
(ii) The second assertion of the theorem follows from the first and the equality
∞
Eγ = P(γ v)dv.
0
(iii) Let ∗ = ∞. It follows from (i) that λ+ 1. We have to prove the

(γ )

converse inequality. Put θn := 1n nk=1 γk , where the γk are independent copies
of γ . By inequality (1.3.1),
1
lim ln P(θn v) −(γ ) (v). (1.4.2)
n→∞ n
We will obtain a lower bound for the left-hand side of this inequality. By virtue of

the convexity of the function , for ζn := 1n nk=1 ξk one has
1
n
(ζn ) (ξk ) = θn , (1.4.3)
n
k=1
and therefore
P(θn v) P((ζn ) v) P(ζn α+ (v)).
For the sake of definiteness, let + = ∞. Then the function (α) continuously
increases from 0 to ∞ on (Eξ , ∞) and therefore (α+ (v) + 0) = (α+ (v)) = v.
By the large deviation principle for ζn (see Theorem 1.1.4) it is not hard to obtain
(see also Theorem 2.2.2)
1 1
lim ln P(θn v) lim ln P(θn v)
n→∞ n n→∞ n
1
lim ln P(ζn α+ (v)) −(α+ (v) + 0) = −v. (1.4.4)
n→∞ n
1.4 The random variable γ = (ξ ) and its deviation function 27
It follows from (1.4.2) and (1.4.4) that
(γ ) (v) v. (1.4.5)
Further, by the property (1) (see (1.1.8)), the deviation function (γ ) (v)
for the random variable γ (and for any other random variable as well) allows a
representation of the form
v
(v) =
(γ )
λ(γ ) (u)du, (1.4.6)
Eγ
where λ(γ ) (u) is the value of λ at which the supremum is attained in the definition
(γ )
(γ ) (u) = supλ {λu−ln Eeλγ }. Moreover, λ(γ ) (u) ↑ λ+ := sup{λ : Eeλγ < ∞}
as u ↑ ∞. From that it follows that there exists the limit
(γ ) (v) (γ )
lim = λ+ . (1.4.7)
v→∞ v
(γ )
Now the required inequality λ+ 1 follows from (1.4.5) and (1.4.7), which
(γ )
proves that λ+ = 1.
If ∗ = + < ∞ then the assertion of the theorem follows in an obvious way
from the equality
+ = − ln P(ξ = s+ )
(see (2ii)). Theorem 1.4.1 is proved.
(γ )
Relations (1.4.7) for λ+ = 1 mean that, under broad assumptions,
− ln P(γ v) ∼ v as v→∞
(see also subsection 1.4.3). This is, in a certain sense, an analogue of the relation
P F(ξ ) < t = P ξ < F (−1) (t) = F F (−1) (t) ≡ t,
which is true under some assumptions on the distribution function F(t) = P(ξ <
(γ )
t). Also observe that, by virtue of the relation (1.4.6) and the equality λ+ = 1,
one has the inequality
(γ ) (v) < v − Eγ for v > Eγ . (1.4.8)
Example 1.4.2. Suppose that ξ has a normal distribution. By virtue of subsec-
tion 1.4.1, one can assume that Eξ = 0, Eξ 2 = 1. Then (α) = α 2 /2 (see
Example 1.1.5), so that γ = ξ 2 /2, Eγ = 1/2,
√ ∞
2 −u2 /2 1
P(γ v) = 2P(ξ 2v) = √ √ e du ∼ √ e−v as v → ∞.
2π 2v πv
Further,
∞
2 /2 1 2 /2−x2 /2 1
Eeλγ = Eeλξ =√ eλx dx = √ for λ 1.
2π −∞ 1−λ
Therefore the equation for the point λ(γ ) (v) takes the form of a partial derivative
with respect to λ

1 1
λv + ln(1 − λ) = v − =0
2 λ 2(1 − λ)
and it has the unique solution λ(γ ) (v) = 1 − 1/(2v). Hence we find that
v
1 1
(γ ) (v) = 1− du = v − (ln v + ln 2 + 1).
Eγ 2u 2
It is also clear that, in this example, one has s± = ±∞, ± = ∞.
Example 1.4.3. Suppose that ξ follows the exponential distribution. By virtue of
subsection 1.4.1, one can assume
P(ξ t) = e−t for t 0.
Then (see Example 1.1.7)
(α) = α − 1 − ln α for α 0,
so that γ = ξ − 1 − ln ξ . The equation (α) = v has the solution α+ (v) =
v + ln v + 1 + O(ln v/v) as v → ∞. Therefore,

ln v
P(γ v) = P ξ v + ln v + 1 + O
v

−(v+ln v+1+O(ln v/v)) 1 −v−1 ln v
=e = e 1+O as v → ∞.
v v
In this example, s− = 0, s+ = ∞, ± = ∞.
Example 1.4.4. For a Bernoulli random variable ξ (P(ξ = 1) = p = 1 − P(ξ =
0)) one has (see Example 1.1.6)
ξ 1− ξ
γ = (ξ ) = ξ ln + (1− ξ ) ln , Eγ = −p ln p − (1− p) ln(1− p) ln 2.
p 1− p
Hence
− ln p with probability p,
γ =
− ln(1 − p) with probability 1 − p.
If p = 1/2 then the variable γ degenerates to the constant ln 2. For p < 1/2, the
variable γ is given by an affine transformation of ξ :
ξ +a
γ = with a = ln(1 − p), b = ln(1 − p) − ln p,
b−a
so that
(γ ) (v) = (v(b − a) − a).
The case p > 1/2 is dealt with in a similar way. In this example we have s− = 0,
s+ = 1, − = − ln(1 − p), + = − ln p.
Another Random Document on
Scribd Without Any Related Topics
»Sire», kuiskasi Gilbert, »teillä ei ole enää mitään pelättävää.
Sallikaa minun palata kuningattaren luo.»
»Menkää!» kehoitti kuningas ja puristi hänen kättänsä.
Gilbertin poistuessa huoneesta astuivat Isnard ja Vergniaud sisälle.
He olivat lähteneet kansalliskokouksesta ja tulivat nyt

suojelemaan kuningasta kansansuosiollaan ja tarpeen mukaan
myöskin ruumiillaan.
»Missä kuningas on?» kysyivät he.
Gilbert viittasi kädellään, ja molemmat kansanedustajat kiiruhtivat

kuninkaan luo.
Päästäkseen kuningattaren luo Gilbertin täytyi kulkea monen

huoneen
kautta ja muiden ohella myöskin kuninkaan makuusuojan läpi.
Kaikkialla oli tungokseen asti kansaa.
»Kas vain», sanoivat miehet koetellessaan kuninkaan vuoteen

pehmeyttä, »paksulla Vetolla on vuode, joka totisesti on parempi
kuin meikäläisten!»
Tämä meno ja puuha ei ollut enää huolestuttavaa. Ensimmäinen

kiihtymys oli häipynyt.
Gilbert jatkoi matkaansa rauhoittuneena.
Astuessaan saliin, minne oli kuningattaren jättänyt, hän silmäili

pikaisesti ympärilleen ja hengitti kevyemmin.
Kuningatar oli entisellä paikallaan. Pikku kruununprinssin päässä
oli
punainen myssy kuten hänen isällänsäkin.
Viereisestä huoneesta kuului raskasta astuntaa ja Gilbert suuntasi

katseensa ovelle.
Santerre oli tulossa. Tämä jättiläinen astui saliin.
»Täälläkö Itävallatar on tavattavissa?» kysyi hän.
Gilbert astui hänen luokseen salin poikki.
»Herra Santerre», sanoi hän.
Santerre pyörähti häneen päin.
»No mutta, tekö se olettekin, tohtori Gilbert!» huudahti hän

iloisesti.
»Joka ei ole unohtanut», virkkoi Gilbert, »että te olette niitä,

jotka avasitte hänelle Bastiljin portit… Sallikaa minun esitellä teidät
kuningattarelle, herra Santerre.»
»Kuningattarelle? Minutko kuningattarelle?» ihmetteli oluenpanija.
»Niin juuri, kuningattarelle. Kieltäydyttekö?»
»En maar kieltäydykään!» vastasi Santerre. »Aioin esittäytyä

omin päin, mutta koska te olette saapuvilla…»
»Minä tunnen herra Santerren», virkkoi kuningatar, »ja tiedän,

että nälkävuonna hän yksin ravitsi puolet Saint-Antoinen
esikaupungin asukkaista».
Santerre seisoi mykkänä. Sitten hänen hämmentynyt katseensa
osui kruununprinssiin, ja nähdessään, että suuret hikikarpalot
vierivät pitkin lapsi-rukan poskia, hän virkkoi miehilleen:
»Ottakaa pois punainen myssy lapsen päästä! Näettehän, että

hän on tukehtumaisillaan!»
Kuningatar kiitti häntä katseellaan.
Silloin kunnon flaamilainen kumartui kuningattareen päin, nojaten

pöydän reunaan, ja virkkoi puoliääneen:
»Teillä on taitamattomia ystäviä, madame. Minä tunnen sellaisia,

jotka
palvelisivat teitä paremmin!»
Tunnin kuluttua väkijoukko oli poistunut, ja kuningas palasi

sisarensa
kanssa huoneeseen, missä kuningatar ja lapset odottelivat.
Kuningatar riensi häntä vastaan ja heittäytyi hänen jalkoihinsa.

Molemmat lapset tarttuivat hänen käsiinsä. He syleilivät toisiaan
kuin
haaksirikosta pelastuneet.
Silloin vasta kuningas huomasi, että punainen myssy oli yhä

hänen
päässään.
»Ah, olin unohtanut sen!» huudahti hän.
Hän riisti sen päästään ja viskasi kauas lattialle.

Muuan nuori tykistöupseeri, tuskin kahdenkolmattakaan ikäinen,
oli katsellut koko tätä kohtausta nojatessaan erääseen
rantapengermän puuhun. Ikkunasta hän oli nähnyt kuningasta
uhanneet vaarat ja hänen osakseen tulleet nöyryytykset. Kun sitten
lisäksi ilmestyi punainen myssy kuninkaan päähän, ei hän voinut
hillitä itseään, vaan mutisi:
»Ah, olisipa minulla edes tuhatkaksisataa miestä ja pari tykkiä,

niin vapauttaisin kuninkaan tuota pikaa koko tuosta roskaväestä!»
Mutta koska hänellä ei ollut tuhattakahtasataa miestä eikä paria

tykkiä ja kun hän ei voinut kauempaa katsella inhoittavaa
näytelmää, kääntyi hän poispäin.
Tämä nuori upseeri oli Napoleon Bonaparte.

XVII
Vastavaikutus
Tuileriein tyhjennys oli käynyt yhtä alakuloisesti ja mykästi kuin

rynnäkkö oli ollut meluava ja pelottava.
Päivän vähäisiä tuloksia kummeksien joukko päättelikin: — Emme

ole
saavuttaneet mitään. Sinne on mentävä uudelleen!
Se oli tosiaankin liian paljon, ollakseen uhkaus, mutta liian vähän

merkitäkseen tosihyökkäystä.
Ne, jotka olivat tehneet tapahtumista johtopäätöksiä, olivat

arvostelleet Ludvig XVI:tta sen mukaan, mitä huhu tiesi hänestä
kertoa. He muistelivat, kuinka kuningas oli paennut Varennesiin
lakeijaksi pukeutuneena, ja puhelivat:
»Kuullessaan ensimmäisen melun Ludvig piiloutuu johonkin

kaappiin, pöydän alle, verhon taakse. Tarvitsee vain sohaista
miekalla, ja hänestä selvitään, ja sitten voi sanoa kuin Hamlet, joka
luullen surmanneensa Tanskan tyrannin huudahti: 'Rotta!'»
Mutta nyt olikin käynyt ihan toisin. Milloinkaan kuningas ei ollut
esiintynyt näin tyynesti, tekisi mieli sanoa, näin suurpiirteisesti.
Solvaus oli ollut tavaton, mutta se ei ollut kohonnut hänen

alistumisensa tasalle. Hänen arkailevaa lujuuttaan, jos sellaisesta
voi tässä puhua, oli tarvinnut ärsyttää, ja tässä ärtymystilassa se oli
jäykistynyt kovaksi kuin teräs. Äärimmäisen pulan kannustamana
hän oli kalpenematta viisi tuntia katsellut kuinka kirveet välähtivät
hänen päänsä päällä, ja piikit, sapelit ja pistimet uhkasivat lävistää
hänen rintansa. Kymmenessä mitä verisimmässä taistelussa ei
ketään kenraalia voisi uhata sellainen vaara, jota kuningas uhitteli
tämän kapina-aallon vyöryessä hitaasti hänen editseen! Théroignet,
Saint-Huruget, Lazouskit, Fournierit, Verrierèt, kaikki nuo tottuneet
murhaajat olivat lähteneet liikkeelle vakain aikein surmata hänet, ja
silloin tämä odottamaton hengen ylevyys, joka oli kohonnut
myrskyn yläpuolelle, oli kirvoittanut puukon heidän kädestään.
Ludvig XVI oli kestänyt kilvoituksen. Kuninkaallinen Ecce homo oli
näyttänyt otsaa, jota punainen myssy peitti, kuten Jeesuksen otsaa
orjantappurakruunu. Ja kuten Jeesus kesken häväistystä ja
pahoinpitelyä oli sanonut: »Minä olen teidän vapahtajanne!»
samoin Ludvig XVI oli kesken solvauksia ja herjauksia lausunut
yhtenään: »Minä olen teidän kuninkaanne!»
Niin oli käynyt. Murtaessaan auki Tuileriein portit

vallankumouksen ajatus oli luullut tapaavansa kuninkuuden
voimattoman ja vapisevan varjon, ja suureksi kummakseen se
kohtasikin keskiaikaisen uskon, uhmaavana ja elävänä! Ja hetken
seisoi vastakkain kaksi maailmankatsomusta, toinen syrjään
painumaisillaan, toinen sijalle nousemassa; se oli jotakin hirveää,
ikäänkuin taivaalla nähtäisiin yhtaikaa kaksi aurinkoa, toinen
nousemassa ennenkuin toinen on ehtinyt laskea! Mutta molemmat
loistivat yhtä mahtavasti ja kirkkaasti, kansan vaatimuksessa oli
yhtä lujaa uskoa kuin kuninkuuden vastustuksessa.
Rojalistit olivat haltioissaan. Voittohan oli heidän.
Kuningas, jota yhä pakotettiin tottelemaan kansalliskokousta ja

joka oli ollut valmis vahvistamaan toisen kahdesta asetuksesta,
sovellutti kielto-oikeutensa molempiin, sillä hän tiesi, ettei vaara
olisi suurempi, vaikka hän kumoaisi molemmatkin.
Tänä kohtalokkaana kesäkuun 20 päivänä kuninkuus oli muuten

vaipunut niin syvälle, että se tuntui tavanneen jo kuilun pohjan ja
että sen täytyi siitä lähtien nousta.
Ja tilanne näytti tosiaankin kehittyvän siihen suuntaan.
Kesäkuun 21 päivänä kansalliskokous päätti, ettei

kansankokouksia enää sallittaisi tulliporttien sisäpuolella. Siten
paheksuttiin edellisen päivän mellakkaa, tai enemmänkin, se
tuomittiin.
Kesäkuun 20 päivän iltana Pétion oli saapunut Tuilerieihin, kun

mellakka oli päättymässä.
»Sire», sanoi hän kuninkaalle, »nyt vasta olen saanut kuulla,

mihin
tilanteeseen teidän majesteettinne on joutunut».
»Sepä kummallista», vastasi kuningas. »Ja tätä on kumminkin

kestänyt
jo melko kauan.»
Seuraavana päivänä perustuslailliset, rojalistit ja feuillantit
ehdottivat kansalliskokoukselle, että julistettaisiin sotalaki
pantavaksi voimaan.
Lukija muistaa, mihin edellinen sotalaki oli johtanut lähes vuotta

varemmin Mars-kentällä.
Pétion riensi kansalliskokoukseen.
Ehdotusta perusteltiin väitteillä, että uusia väenkokouksia kuului

olevan tekeillä. Pétion vakuutti, ettei uusista väenkokouksista ollut
puhettakaan. Hän vastasi Pariisin turvallisuudesta. Ehdotus
hylättiin.
Istunnosta Pétion lähti kahdeksan aikaan illalla Tuilerieihin

vakuuttamaan kuninkaalle pääkaupungin olevan rauhallisen. Hänen
mukanaan oli Sergent, vaskipiirrosten kaivertama, Marceaun vävy,
valtuustonjäsen ja poliisilaitoksen johtomiehiä. Heidän seuraansa oli
liittynyt pari kolme valtuustonjäsentä.
Carrousel-aukiolla jotkut Ludvig pyhän järjestön ritarit,

perustuslailliset kaartilaiset ja kansalliskaartilaiset herjasivat heitä.
Pétionia hätyytettiin, Sergentia lyötiin rintaan ja kasvoihin, vaikka
hänen yllänsä oli valtuuston vyöhyt, kaatoipa muuan isku hänet
katuun.
Perille päästyään Pétion huomasi, että häntä odotti taistelu.
Marie-Antoinette loi häneen silmäyksen, jommoisia vain Maria

Teresian tyttären silmät osasivat sinkauttaa: kaksi vihan ja
ylenkatseen leimahdusta, kaksi pelottavaa, säkenöivää salamaa.
Kuningas tiesi jo, mitä kansalliskokouksessa oli tapahtunut.

»Vai niin, hyvä herra», sanoi hän, »tekö siis väitätte, että rauha
on
palannut pääkaupunkiin?»
»Niin, sire», vastasi Pétion, »kansa on esittäytynyt teille, se on

nyt
tyyni ja tyytyväinen».
»Myöntäkää, herra», jatkoi kuningas hyökkäävästi, »myöntäkää,

että eilispäivä oli julkean häväistyksen päivä ja ettei valtuusto ole
tehnyt, mitä sen olisi pitänyt ja mitä se olisi voinut tehdä».
»Sire», vastasi Pétion, »valtuusto on tehnyt velvollisuutensa.

Yleinen mielipide arvostelkoon sen tekoja.»
»Sanokaa: kansa kokonaisuudessaan, herra.»
»Valtuusto ei pelkää kansankaan arvostelua.»
»Kuinka on Pariisin laita tällä hetkellä?»
»Se on rauhallinen, sire.»
»Se ei ole totta!»
»Sire…»
»Olkaa vaiti!»
»Kansan hallintomiehen ei tarvitse olla vaiti, sire, kun hän täyttää

velvollisuutensa ja lausuu totuuden.»
»Hyvä on, voitte poistua.»

Pétion kumarsi ja lähti.
Kuningas oli esiintynyt niin kiihkeästi, hänen piirteensä ilmaisivat

niin syvää kiukkua, että kuningatarkin, tuittupäinen nainen,
tulisieluinen amatsooni, kauhistui.
»Hyvä Jumala», sanoi hän Rödererille, kun Pétion oli kadonnut,

»eikö teidänkin mielestänne kuningas esiintynyt hyvin kiihkeästi ja
ettekö pelkää, että hänen tulisuutensa tuottaa hänelle haittaa
pariisilaisten taholta?»
»Madame», vastasi Röderer, »kukaan ei kummastele, että

kuningas käskee alamaisen vaieta, joka unohtaa esiintyä
kunnioittavasti».
Seuraavana päivänä kuningas kirjoitti kansalliskokoukselle

valittaen, mikä häväistys oli kohdannut linnaa, kuninkuutta ja
kuningasta. Sitten hän sepitti julistuksen kansalle.
Oli siis kaksi kansaa: kansa, joka oli järjestänyt kesäkuun 20

päivän, ja kansa, jolle kuningas esitti valittelunsa.
Kesäkuun 24 päivänä kuningas ja kuningatar olivat mukana

kansalliskaartin katselmuksessa, ja heitä tervehdittiin haltioitunein
suosionosoituksin.
Samana päivänä Pariisin direktorio erotti pormestarin. Mistä se

sai moisen rohkeuden? Kolmea päivää myöhemmin asia selvisi.
Lafayette oli lähtenyt leiristään yhden ainoan upseerinsa kanssa,

saapunut Pariisiin kesäkuun 27 päivänä ja mennyt ystävänsä de la
Rochefaucauldin luo.
Yön kuluessa lähetettiin tieto perustuslaillisille, feuillanteille ja
rojalisteille ja huolehdittiin siitä, että parvekkeet olisivat täynnä
seuraavana päivänä.
Seuraavana päivänä kenraali tuli kansalliskokoukseen.

Kolminkertainen eläköönhuuto tervehti häntä, mutta jokainen niistä
hukkui girondelaisten murinaan. Kaikki tunsivat, että tulossa oli
tuima istunto.
Kenraali Lafayette oli kelpo mies sanan puhtaimmassa

merkityksessä, mutta se ei vielä merkitse rohkeutta. Harvoin on
todella kelpo mies samalla rohkea.
Lafayette vaistosi vaaran. Yksin kaikkia vastassa hän joutuisi

pelaamaan kansansuosionsa rippeillä. Jos hän menettäisi ne,
tuhoutuisi hän mukana; jos hän voittaisi, voisi hän pelastaa
kuninkaan.
Hänen yrityksensä oli sitäkin kauniimpi, kun hän tiesi, että

kuningas kaihtoi ja kuningatar vihasi häntä. Olihan Marie-Antoinette
sanonut: »Mieluummin tuhoudun Pétionin toimesta kuin pelastun
Lafayetten avulla!»
Ehkä hän tulikin suorittamaan vain aliluutnantin uhkamielisen

teon, vastaamaan haasteeseen.
Paria viikkoa aikaisemmin hän oli kirjoittanut yhtaikaa sekä

kuninkaalle että kansalliskokoukselle: kuninkaalle kehoittaen tätä
vastarintaan ja kansalliskokoukselle uhaten sitä, jos se jatkaisi
hyökkäilyjään.
»Hän uskaltaa uhitella armeijansa keskellä», oli muuan edustaja
sanonut. »Saammepa nähdä, puhuuko hän samalla lailla
joutuessaan seisomaan yksin meidän parissamme.»
Nämä sanat oli kerrottu Lafayettelle, kun hän majaili leirissään

Maubeugessa, ja kenties se oli syynä hänen Pariisin-matkaansa.
Hän nousi puhujalavalle toisten osoittaessa suosiotaan, toisten

uhkaavasti muristessa.
»Hyvät herrat», sanoi hän, »minua on soimattu, että sepitin

tämän kuun kuudentenatoista päivätyn kirjeeni leirissä.
Velvollisuuteni oli torjua syytös, että arkailen lähteä siitä
kunniakkaasta rintasuojuksesta, jonka joukkojen kiintymys on
luonut ympärilleni, ja tulla yksinäni teidän eteenne. Minua kutsui
lisäksi vieläkin voimakkaampi syy. Kesäkuun
kahdennenkymmenennen päivän ilkiteko on herättänyt kaikkien
kunnon kansalaisten närkästystä, etenkin armeijassa. Upseerit,
aliupseerit ja sotamiehet ovat kuin yksi mies. Olen saanut kaikilta
armeijaosastoilta kirjelmiä, joista huokuu rakkaus perustuslakiin ja
viha vehkeilijöitä vastaan. Olen lopettanut nämä mielenilmaukset.
Olen katsonut tehtäväkseni ilmaista yksin kaikkien tunteet. Nyt
puhun teille kansalaisena. On jo aika turvata perustuslaki ja taata
kansalliskokouksen vapaus ja kuninkuuden arvo. Rukoilen
kansalliskokousta päättämään, että kesäkuun
kahdennenkymmenennen päivän ilkiteko rangaistaisiin kuin
majesteettirikos ja että ryhdytään tehokkaihin keinoihin, jotta
kaikkia perustuslain säätämiä arvoasteita kunnioitetaan ja etenkin
teidän ja kuninkaan arvovaltaa. Armeijalle on taattava varmuus
siitä, ettei perustuslakia loukata rahdunkaan vertaa maan rajojen
sisällä, kun kelpo ranskalaiset vuodattavat vertansa puolustaakseen
sitä rajalla.»
Sitä mukaa kuin Lafayette lähestyi loppulausettaan Guadet oli

noussut hitaasti paikaltaan. Suosionosoitusten raikuessa tämä
terävä girondelaispuhuja kohotti kätensä merkiksi, että hän halusi
vastata kenraalin puheeseen. Kun girondelaispuolue tahtoi
sinkauttaa ivan vasaman, antoi se jousen Guadetille, ja Guadetin
tarvitsi vain umpimähkään siepata nuoli viinestään.
Tuskin olivat viimeiset suosionosoitukset vaienneet, kun hänen

värähtelevän äänensä kaiku jo täytti salin.
»Heti kun näin Lafayetten», huudahti hän, »heräsi minussa

muuan ylen lohdullinen ajatus. Meitä ei siis enää uhkaakaan
ulkonainen vihollinen, päättelin itsekseni. Herra Lafayette on tullut
ilmoittamaan meille voitostaan ja heidän tuhostaan! Se oli
lyhytaikainen harhaluulo. Vihollisemme ovat yhä elossa ja samat
kuin ennenkin. Ulkonaiset vaaramme eivät ole hälvenneet. Ja silti
on herra Lafayette Pariisissa. Hän esittäytyy kelpo miesten ja
armeijan puhetorveksi. Keitä ovat nämä kelpo miehet? Kuinka tämä
armeija on voinut keskustella? Mutta herra Lafayette näyttäköön
meille ensin lomalupansa.»
Nämä sanat vihjasivat girondelaisille, että tuuli alkoi puhaltaa

myötäisesti. Ja tuskin oli Guadet lopettanut, kun koko sali jymähti
suosionosoitusten ukkosesta.
Muuan edustaja huusi paikaltaan:
»Hyvät herrat, unohdatte, kenelle ja kenestä puhutaan!

Unohdatte, mikä mies Lafayette on! Lafayette on Ranskan
vapauden vanhin poika. Lafayette on uhrannut vallankumoukselle
omaisuutensa, aatelisarvonsa, elämänsä!»
»Mutta tehän pidätte hänen hautajaispuhettaan!» huusi muuan

ääni.
»Hyvät herrat», sanoi Ducos, »keskustelun vapautta on loukattu,

kun joukossamme esiintyy kansalliskokoukselle vieras kenraali».
»Siinä ei ole kaikki!» huudahti Vergniaud. »Tämä kenraali on

poistunut vartioasemaltaan vihollisen edestä. Hänelle eikä suinkaan
sille marsalkalle, jonka hän on jättänyt omalle tilalleen, on uskottu
se armeijaosasto, jonka hän on hylännyt. Tutkikaamme, onko hän
saanut laillisen loman. Jos hän on poistunut leiristään lomatta,
vangittakoon hänet ja tuomittakoon karkulaisena.»
»Se oli minunkin kysymykseni ydin», huomautti Guadet, »ja minä

kannatan Vergniaudin ehdotusta».
»Kannatetaan, kannatetaan!» huusi koko Gironde. »Äänestys!»

sanoi
Gensonné.
Äänestyksessä saivat Lafayetten ystävät kymmenen äänen

enemmistön.
Kuten rahvas kesäkuun 20 päivänä, hänkin oli uskaltanut joko

liian paljon tai liian vähän. Se oli sellainen voitto, josta Pyrrhos
vaikeroi, menetettyään puolet armeijastaan: »Vielä yksi sellainen
voitto, ja minä olen hukassa!»
Kuten Pétion, niin Lafayettekin lähti kansalliskokouksesta

kuninkaan luo.
Hänet otettiin vastaan leppeämmin, mutta ei yhtään suopeammin.
Lafayette oli uhrannut kuninkaalle ja kuningattarelle enemmän

kuin elämänsä: hän oli uhrannut kunniansa.
Jo kolmannen kerran hän teki tämän lahjoituksen,

kallisarvoisemman kuin yksikään niistä, joita kuninkaat voivat
tehdä. Ensimmäisen kerran Versaillesissa lokakuun 6 päivänä,
toisen kerran Mars-kentällä heinäkuun 17 päivänä ja kolmannen
kerran tänään.
Hän toivoi vielä. Siitä toivonkipinästä hän nyt tuli puhumaan

hallitsijalleen. Seuraavana päivänä hän pitäisi kuninkaan kanssa
kansalliskaartin katselmuksen. Ei olisi epäilystäkään, ettei
kuninkaan ja entisen ylipäällikön osanotto herättäisi yleistä
hurmiota. Lafayette käyttäisi hyväkseen tätä vaikutusvaltaa,
marssisi kansalliskokoukseen, pidättäisi girondelaisedustajat, ja
näin syntyvän hälinän aikana kuningas lähtisi Pariisista Maubeugen
leiriin ja olisi siten täysin turvassa.
Se olisi uhkarohkea kaappaus, mutta silloisissa oloissa se tuntui

miltei varmalta.
Ikävä vain, että kello kolmen aikaan aamulla Danton tuli Pétionin
luo ja paljasti salahankkeen. Päivän koitteessa Pétion peruutti
katselmuksen.
Kuka oli pettänyt kuninkaan ja Lafayetten? Kuningatar!
Eikö hän ollut sanonut, että hän mieluummin tuhoutuisi jonkun

toisen toimesta kuin pelastuisi Lafayetten avulla?
Hänellä olikin tarkka käsi: hän oli menossa tuhoon Dantonin
toimesta!
Hetkellä, jolloin katselmuksen olisi pitänyt alkaa, Lafayette lähti

Pariisista ja palasi leiriin. Eikä hän silti ollut kadottanut kaikkea
toivoaan kuninkaan pelastamisesta.
XVIII
Vergniaud valmistautuu puhumaan
Lafayetten voitolla, epäilyttävällä voitolla, joka johti

peräytymiseen, oli omituiset seuraukset.
Se oli lyönyt rojalistit. Girondelaiset sensijaan olivat näennäisestä

tappiosta vahvistuneet, sillä se oli innostanut heitä osoittamalla
kuilun, johon he olivat olleet suistumaisillaan.
Jos Marie-Antoinetten sydän olisi vihannut vähemmän, olisi

Gironde sillä hetkellä kai ollut lopussa.
Hoville ei saisi nyt suoda aikaa korjata erehdystä, jonka se oli

tehnyt. Jokainen haki ja itsekukin luuli löytäneensä sopivan keinon.
Kun keinot sitten esitettiin, havaittiin ne kaikki riittämättömiksi ja
hylättiin.
Rouva Roland, puolueen sielu, halusi panna kansalliskokouksen

järähtämään. Kuka voisi sitä järäyttää? Kuka pystyisi iskun
sivaltamaan? — Vergniaud!
Mutta mitä teki tämä Akilles teltassaan tai oikeammin tämä
Reinhold
Armidan puutarhoissa? — Hän lempi.
Kovin vaikeata on rakastaessaan vihata.
Hän rakasti kaunista rouva Simon Candeillea, näyttelijätärtä,

runoilijatarta, soittotaiteilijatarta. Ystävät hakivat häntä toisinaan
pari kolme päivää, löytämättä häntä kotoaan. Vihdoin he löysivät
hänet kauniin naisen jalkojen juuresta. Hänen toinen kätensä lepäsi
naisen polvilla, ja toinen näppäili haaveksivasti harpun kieliä.
Ja joka ilta hän meni teatterin orkesteriin paukuttamaan käsiään

sille, jota hän palvoi kaiket päivät.
Eräänä iltana kansalliskokouksesta lähti kaksi epätoivoista

edustajaa. Vergniaudin toimettomuus kauhistutti heitä Ranskan
takia. Nämä miehet olivat Grangeneuve ja Chabot.
Edellinen oli bordeauxilainen asianajaja, Vergniaudin ystävä ja

kilpailija ja, kuten hänkin, Gironden edustaja; jälkimäinen oli
entinen kapusiinimunkki, laatinut yksin tai muiden avulla
Sanskulottien katkismuksen, joka vuodatti kuninkuuden ja
uskonnon ylitse luostarikammiossa kertynyttä sappea.
Synkkänä ja mietteissään Grangeneuve asteli Chabotin rinnalla.

Tämä silmäili toveriaan ja oli näkevillään hänen otsallaan ajatusten
luoman varjon.
»Mitä ajattelet?» kysyi Chabot.
»Ajattelen», vastasi toinen, »että tämä hitaus heikentää

isänmaata ja surmaa vallankumouksen».
»Ah, niinkö arvelet?» naurahti Chabot ominaisella katkeralla
tavallaan.
»Ajattelen», pitkitti Grangeneuve, »että jos kansa suo

kuninkuudelle aikaa, kansa on tuhon oma!»
Chabot nauroi kimakasti.
»Ajattelen», jatkoi Grangeneuve, »että vallankumouksilla on vain

yksi otollinen hetki. Jotka lyövät sen laimin, eivät löydä sitä
toistamiseen, vaan saavat tehdä siitä tilin Jumalalle ja
jälkimaailmalle.»
»Ja sinä arvelet, että Jumala ja jälkimaailma vaativat meidät tilille

laiskuudestamme ja toimettomuudestamme?»
»Pahoin pelkään.»
Lyhyen vaitiolon jälkeen Grangeneuve jatkoi:
»Kuulehan, Chabot, olen varma, että kansa on väsynyt

viimeisestä tappiostaan. Se ei nouse enää ilman voimakasta
kannustusta, ilman veristä kiihoketta. Se tarvitsee raivon tai kauhun
kuohuntaa, mistä se ammentaa verestä tarmoa.»
»Kuinka sille järjestetään se raivon tai kauhun tilaisuus?» kysyi

Chabot.
»Sitä juuri mietinkin», vastasi Grangeneuve, »ja luulen jo

päässeenikin salaisuuden perille».
Chabot siirtyi lähemmäksi. Toverin äänensävy ilmaisi, että hän

saisi
kuulla jotakin hirveää.
»Mutta», jatkoi Grangeneuve, »löydänkö miehen, joka olisi kyllin

päättäväinen sellaiseen tekoon?»
»Puhu!» virkkoi Chabot äänellä, jonka tiukkuus oli omiaan

hälventämään hänen toverinsa epäilykset. »Minä olen valmis
kaikkeen, tuhotakseni sen, mitä vihaan, ja minä vihaan kuninkaita
ja pappeja!»
»No niin», sanoi Grangeneuve, muistellen menneisyyttä, »olen

nähnyt kaikkien vallankumousten kehdossa puhdasta verta,
Lukretiuksen ajoista Sidneyn päiviin asti. Valtiomiehille
vallankumoukset ovat teorioita, mutta kansoille ne ovat kostoa. Jos
siis tahdotaan työntää laumaa kostoon, on sille esitettävä uhri. Hovi
epää meiltä sellaisen uhrin; tehkäämme se siis omasta
keskuudestamme hyvän asiamme eduksi!»
»En ymmärrä», sanoi Chabot.
»Jonkun meistä — tunnetuimpien, kiihkeimpien,

moitteettomimpien joukkoon kuuluvan — tulee sortua aateliston
uhrina.»
»Jatka.»
»Tuon uhrin täytyy olla kansalliskokouksen jäsen, jotta

kansalliskokous tarttuisi kostonkalpaan. Lyhyesti sanoen minun
pitää ruveta siksi uhriksi!»
»Mutta ylimykset eivät iske sinuun, Grangeneuve. He varovat sitä

tekemästä!»
»Tiedän sen. Siksipä sanoinkin, että olisi löydettävä joku
päättäväinen mies…»
»Mitä varten?»
»Iskemään minua.»
Chabot väistyi askeleen syrjään, mutta Grangeneuve tarttui hänen

käsivarteensa.
»Chabot», virkkoi hän, »sanoit äsken olevasi valmis tuhoamaan

kaikki,
mitä vihaat. Oletko valmis surmaamaan minut?»
Munkki oli vaiti. Grangeneuve jatkoi:
»Puheeni on turhaa, elämäni tarpeeton vapaudelle, mutta

kuolemani hyödyttäisi sitä. Ruumiini kohoaisi kapinan lipuksi, ja
minä sanon sinulle…»
Grangeneuve viittasi uhkaavin elein kädellään Tuilereihin päin.
»Tuon linnan ja sen asukkien pitää tuhoutua siinä myrskyssä!»
Chabot silmäili Grangeneuviä puistattavan ihailun vallassa.
»No?» kärtti Grangeneuve.
»Ah, ylevä Diogenes», sanoi Chabot, »sammuta lyhtysi! Se mies

on löydetty!»
»Sopikaamme siis kaikesta jo tänä iltana. Ensi yönä kävelen

yksinäni täällä» (he olivat tällöin Louvren ristikkoportin kohdalla)
»autioimmassa ja pimeimmässä kolkassa… Jos pelkäät kätesi
hervahtavan, ota mukaasi pari isänmaanystävää. Teen tämän
liikkeen, jotta he tuntisivat minut.»
Grangeneuve kohotti molemmat kätensä.
»He surmaavat minut, ja lupaan kaatua ääntäkään päästämättä.»
Chabot pyyhki otsaansa nenäliinalla.
»Huomisaamuna», jatkoi Grangeneuve, »löydetään ruumiini.

Sinä syytät hovia. Kansan kostava käsi tekee loput.»
»Hyvä on», sanoi Chabot. »Ensi yönä siis!»
Ja nämä eriskummaiset salaliittolaiset puristivat toistensa kättä ja

erosivat.
Grangeneuve meni asuntoonsa, teki testamenttinsa ja päiväsi sen

ikäänkuin se olisi Bordeauxissa laadittu vuotta aikaisemmin.
Chabot käväisi Palais-Royalissa aterioimassa, poikkesi sitten

rautakauppaan ja osti puukon.
Sieltä lähdettyään hän näki sattumalta teatterien

ohjelmailmoituksia.
Neiti Candeille esiintyi illalla. Munkki tiesi nyt mistä löytäisi

Vergniaudin.
Hän meni Comédie Françaiseen, nousi kauniin näyttelijättären

aitioon ja tapasi siellä tavalliset ihailijat: Vergniaudin, Talman,
Chénierin ja Dugazonin.
Candeille esiintyi kahdessa näytelmässä, ja Chabot viipyi
teatterissa loppuun asti.
Kun näytäntö oli päättynyt, kaunis näyttelijätär pukeutunut ja

Vergniaud tarjoutunut saattamaan hänet Richelieu-kadulle, jonka
varrella hän asui, kapusi Chabot myös vaunuihin.
»Teillä on kai jotakin sanottavaa minulle, Chabot?» kysyi

Vergniaud, joka oivalsi, että kapusiinimunkilla oli hänelle asiaa.
»Niin on… mutta olkaa huoleti, se ei vie paljon aikaa.»
»Sanokaa se heti.»
Chabot otti esille kellonsa.
»Ei ole vielä aika tullut», virkkoi hän.
»Milloin siis?»
»Sydänyöllä.»
Kaunista Candeillea puistatti tämä salaperäinen keskustelu.
»Voi, hyvä herra!» mutisi hän.
»Rauhoittukaa», lohdutti Chabot, »Vergniaudilla ei ole mitään

pelättävää, mutta isänmaa tarvitsee häntä».
Ajoneuvot vierivät näyttelijättären asunnolle päin.
Candeille ja molemmat miehet pysyivät äänettöminä.

Näyttelijättären oven edessä Vergniaud kysyi:
»Tuletteko mukaan, Chabot?»
»En, te lähdette minun mukanani.»
»Mutta minne te viette hänet, hyvä Jumala?» vaikeroi Candeille.
»Parinsadan askeleen päähän. Neljännestunnin perästä hän on

vapaa,
lupaan sen.»
Vergniaud puristi kauniin lemmittynsä kättä, viittasi häntä

rauhoittumaan ja lähti Chabotin kanssa Traversière-kadulle.
He menivät Saint-Honoré-kadun poikki ja sitten Echelle-kadulle.
Tämän kadun kulmassa munkki laski kätensä Vergniaudin

olkapäälle ja osoitti toisella kädellä miestä, joka käveli Louvren
autioitten muurien kupeella.
»Näetkö?» kysyi hän.
»Mitä?»
»Tuon miehen?»
»Näen», vastasi girondelainen.
»Hän on ystävämme Grangeneuve.»
»Mitä hän tuolla tekee?»
»Odottaa.»
»Mitä hän odottaa?»

»Että hänet tapettaisiin.»
»Ettäkö hänet tapettaisiin?»
»Niin.»
»Kenen pitäisi surmata hänet?»
»Minun!»
Vergniaud silmäili Chabotia kuin mielenvikaista.
»Muista Spartaa, muista Roomaa», sanoi Chabot, »ja kuuntele».
Hän kertoi Vergniaudille koko jutun.
Munkin puhuessa Vergniaudin pää painui kumaraan. Hän tunsi,

kuinka kaukana hän, veltostunut kansanjohtaja, rakastunut leijona,
oli hirveästä tasavaltalaisesta, joka Deciuksen lailla etsi kuilua,
syöksyäkseen siihen, jotta hänen kuolemansa pelastaisi isänmaan.
»Hyvä on», sanoi hän, »pyydän aikaa kolme päivää puheeni

valmistukseen».
»Jo kolmen päivän perästä…?»
»Ole huoleti, kolmen päivän perästä murskaannun epäjumalan

edessä tai kaadan sen!»
»Annatko sanasi, Vergniaud?»
»Annan.»
»Miehen sanan?»
»Tasavaltalaisen sanan.»
»Nyt en tarvitse sinua. Mene rauhoittamaan lemmittyäsi.»
Vergniaud palasi Richelieu-kadulle, ja Chabot riensi

Grangeneuven luokse.
Nähdessään miehen lähestyvän Grangeneuve vetäytyi kaikkein

pimeimpään soppeen. Chabot seurasi häntä. Grangeneuve pysähtyi
muurin juurelle voimatta mennä pitemmälle. Chabot läheni yhä.
Silloin Grangeneuve teki sovitun liikkeen, kohottaen molemmat
kätensä. Ja kun Chabot pysyi liikkumatta, sanoi hän:
»No, mikä sinua pidättää? Iske toki!»
»Se on tarpeetonta», vastasi Chabot. »Vergniaud aikoo puhua.»
»Olkoon niin!» huoahti Grangeneuve. »Mutta luulen silti tätä

keinoa tehokkaammaksi.»
Mitä arvelette kuninkuuden voivan tällaisia miehiä vastaan?

XIX
Vergniaud puhuu
Jo oli aikakin Vergniaudin tehdä päätös. Vaara kasvoi sekä ulkona

että
sisällä.
Ulkona, Regensburgissa, lähettiläsneuvosto oli yksimielisesti

kieltäytynyt ottamasta vastaan Ranskan ministeriä.
Englanti, joka nimitti itseään Ranskan ystäväksi, varusteli

suunnatonta armeijaa.
Keisarikunnan ruhtinaat, jotka kerskuivat äänekkäästi

puolueettomuudellaan, opastelivat salaisesti vihollista sopiviin
asemiin.
Badenin herttua oli päästänyt itävaltalaiset Kehliin, penikulman

päähän Strassburgista.
Flandriassa olivat asiat vieläkin hullummin. Luckner, iäkäs,

itsepäinen sotakarhu, vastusteli kaikkia Dumouriezin suunnitelmia,
ja tämä oli kuitenkin ainoa mies, joka Ranskalla oli panna vihollista
vastaan. Jollei hän ollutkaan nerokas, oli hän kuitenkin älykäs.
Lafayette oli hovin puolella, ja hänen viimeinen tekosensa oli

osoittanut selvästi, ettei kansalliskokouksen, toisin sanoin Ranskan,
sopinut luottaa häneen.
Biron, urhea ja vakaa mies, ensimmäisten vastoinkäymisten

lannistamana, ajatteli vain puolustussotaa.
Näin oli laita ulkona.
Sisällä Elsassi pyysi äänekkäästi aseita, mutta sotaministeri, joka

oli kokonaan hovin puolella, varoi niitä lähettämästä.
Etelässä muuan prinssien kenraaliluutnantti, Languedocin ja

Cévennien
kuvernööri, jätti valtuutensa ylimystön vahvistettaviksi.
Lännessä eräs yksinkertainen talonpoika, Allan Redeler, kuulutti

messun aikana, että kuninkaan ystävät kokoontuisivat asestettuina
naapurikappelin edustalle.
Viisisataa talonpoikaa kertyi paikalle. Vendéessa ja Bretagnessa

oli
lietsottu chouanilaisuutta: tarvitsi vain sysätä se liikkeelle.
Ja miltei kaikkien departementtien johtoelimet lähettelivät

vallankumousvastaisia julistuksia.
Vaara oli suuri, uhkaava, hirveä, niin suuri, ettei se uhannut enää
yksilöitä, vaan koko isänmaata.
Sanat: »Isänmaa on vaarassa!» kulkivat kuiskauksena suusta
suuhun,
vaikkei niitä vielä lausuttukaan ääneen.
Kansalliskokous odotti.
Chabat ja Grangeneuve olivat sanoneet: »Kolmen päivän perästä

Vergniaud puhuu.»
Laskettiin tunnit, jotka olivat kulumassa.
Vergniaudia ei näkynyt kansalliskokouksessa ensimmäisenä eikä

toisena
päivänä. Kolmantena kaikki saapuivat paikalle jännityksestä
väristen.
Ainoakaan edustaja ei ollut jäänyt pois. Parvekkeet olivat

täpötäynnä.
Viimeisenä astui sisälle Vergniaud.
Tyytyväisyyttä ilmaiseva kohina täytti salin. Parvekkeet

paukuttivat
käsiään kuin teatterin permanto, kun suosittu näyttelijä astuu esille.
Vergniaud kohotti päätänsä nähdäkseen, kenelle taputettiin. Kun

suosionosoitukset yltyivät, huomasi hän, että juuri hänelle
taputettiin.
Hän oli näihin aikoihin lähes kolmenneljättä vuoden ikäinen.

Luonteeltaan hän oli mietiskelevä ja laiska. Hänen uinaileva älynsä
rakasti huolettomuutta. Vain nautinnoissa hän osoitti tulisuutta.
Olisi voinut sanoa, että hänellä oli kiire poimia täysin käsin
nuoruuden kukkia, koska elämän kevät jäisi hänen osaltaan perin
lyhyeksi. Hän meni myöhään makuulle ja nousi vasta vähää ennen
puoltapäivää. Kun hänen piti puhua, valmisteli hän puhettaan
kolme neljä päivää, siloitteli, kiilloitti ja hioi sitä, kuten soturi
taistelun aattona aseitaan. Hän oli puhujana sentapainen kuin
miekkailukoulussa taitava aseen käyttelijä. Isku tuntui hänestä
onnistuneelta vain jos se oli sivallettu loistavasti ja sai osakseen
voimakkaan kättenpaukkeen. Hänen puheensa oli säästettävä
vaaran hetkiksi, äärimmäisen tilanteen turvaksi.
Hän ei ollut tavallisten hetkien mies, on sanonut muuan runoilija;

hän oli suurten päivien mies.
Vergniaud oli rakenteeltaan pikemminkin pieni kuin kookas,

mutta hän oli tanakka, eräänlainen voimailijatyyppi. Hänen
tukkansa oli pitkä ja aaltoileva. Puhuessaan hän ravisti sitä kuin
leijona harjaansa. Leveän otsan alla säihkyivät tuuheitten
kulmakarvojen suojassa mustat silmät, milloin lempeinä, milloin
palavina. Nenä oli lyhyt, hieman leveä, ja sen pielet ylpeästi
koholla. Huulet olivat turpeat, ja kuten lähteestä vesi pulppuaa
vuolaana ja solisten, samoin hänen suustansa sanat syöksyivät
valtavana ryöppynä, vaahtoisina pärskyten ja helisten. Hänen
rokonarpinen ihonsa näytti rosoiselta marmorilta, jota taiteilijan
taltta ei ole vielä sijoitellut, vaan jossa yhä näkyvät käsityöläisen
vasaran jäljet. Se hohti purppuranpunaisena tai valahti kalpeaksi,
kuinka kulloinkin veri ahtautui päähän tai pakeni sydämeen. Omissa
oloissaan tai väkijoukossa liikkuessaan hän näytti tavalliselta
kansalaiselta, johon teräväsilmäisinkään historioitsija ei olisi
kohdistanut huomiotaan, mutta kun intohimon lieska kiehutti hänen
vertansa, kun hänen kasvojensa lihakset nytkähtelivät, kun hänen
kohotettu kätensä vaati hiljaisuutta ja vangitsi kuulijakunnan, tuli

Asymptotic Analysis of Random Walks Light Tailed Distributions 1st Edition A.A. Borovkov All Chapter Instant Download

Uploaded by

Copyright:

Available Formats

Asymptotic Analysis of Random Walks Light Tailed Distributions 1st Edition A.A. Borovkov All Chapter Instant Download

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Asymptotic Analysis of Random Walks Light Tailed Distributions 1st Edition A.A. Borovkov All Chapter Instant Download

Uploaded by

Copyright:

Available Formats

Full download ebooks at ebookmeta.

Asymptotic Analysis of Random Walks Light Tailed

Download more ebook from https://ebookmeta.com

Asymptotic Analysis of Random Walks Light Tailed

Random Walks on Infinite Groups 1st Edition Steven P.

Continuous Time Random Walks for the Numerical Solution

Asymptotic Geometric Analysis Part II 1st Edition Shiri

Distributions 1st Edition Jacques Simon

Theory of Distributions 2nd Edition Svetlin G. Georgiev

Distributions, Volume 3 1st Edition Jacques Simon

Philosophers Walks 1st Edition Bruce Baugh

This is a companion book to Asymptotic Analysis of Random Walks: Heavy-Tailed

A . A . B o r ov k ov is Principal Research Fellow and a Russian Academy of Sciences

Encyclopedia of Mathematics and Its Applications

126 O. Calin and D.-C. Chang Sub-Riemannian Geometry

Cambridge University Press is part of the University of Cambridge.

3 Boundary crossing problems for random walks 136

5 Moderately large deviation principles for the trajectories of

Let us now provide a brief description of the contents of the book.

Let us now mention the main distinguishing features of the book.

1.1 Deviation function and its properties in the one-dimensional case

An important role in describing the distribution asymptotics of the values Sn and

It also follows that the next representation takes place as λ ↓ 0:

where the γj are so-called semi-invariants (or cumulants) of order j of a random

1.1.2 Deviation function

1 This function is often referred to as a rate function.

A = λ : A(λ) < ∞ = λ : ψ(λ) < ∞ , A = A (λ) : λ ∈ A ,

where A (λ+ − 0) = lim A (λ) and A (λ− + 0) is defined analogously.

λ(α) ↑ λ+ as α ↑ α+ ; λ(α) ↓ λ− as α ↓ α− . (1.1.5)

From the equations

The functions λ(α), (α) are analytic on (α− , α+ ).

s+ > 0, s− < 0. (1.1.9)

This can always be achieved using a shift transformation of a random variable;

(α+ ) = − ln P(ξ = s+ ), (α) = ∞ for α > α+ .

Similar statements are true for s− , α− , λ− .

ψ(λ) ∼ E (eλξ ; ξ > s+ − ε), ψ (λ) ∼ E (ξ eλξ ; ξ > s+ − ε)

as λ → ∞ and for any fixed ε > 0. It follows that

as λ → λ+ = ∞. It follows from this and (1.1.6) that

If p+ = 0, then from the relation ψ(λ) = o(eλs+ ) as λ → ∞, we obtain in a

λ(α) = λ+ , (α) = (α+ ) + λ+ (α − α+ ) = αλ+ − A(λ+ ). (1.1.10)

The semi-invariants γj are defined in (1.1.1) and (1.1.2).

Theorem 1.1.2 distinguishes three non-overlapping possibilities,

Theorem 1.1.4. For all α ∈ R and ε > 0,

ψ(λ) = peλ + q, |λ± | = ∞, α+ = 1, α− = 0, a1 = E ξ = p,

ψ(λ) = exp β[eλ − 1 − λ] , |λ± | = ∞, α− = −β, α+ = ∞, a1 = 0,

1.2 Deviation function and its properties in the multidimensional case

1.2.2 Deviation function

(α) = λ(α), α − A λ(α) .

The functions λ(α) and (α) are analytic on A ,

where a1 = Eξ and M −1 is the inverse covariance matrix

lim (α) (α0 ), α0 ∈ Rd . (1.2.3)

ln ψ(λ) = sup α, λ − (α) .

(ξ(1) ) (α(1) ) = inf (α)

and, more generally, for all b ∈ Rd ,

(ξ ,b) (β) = inf (α) : α, b = β .

(ξ H+b) (α) = (α − b)H −1 .

If [C∞ ] is met then

(α)  |α| as |α| → ∞.

(ii) (Continuity inside J<∞ along rays)

(α) = λ(α), α − A λ(α) .

ln ψ(λ) = sup α, λ − (α) .

(ξ ,b) (β) = inf (α) : α, b = β .

(α) |α| as |α| → ∞.

P(ξ ∈ B) P(ξ ∈ > ) = P(e, ξ > b) = P(β > b). (1.3.9)

(β) ((b, ∞)) = inf (α) = (> ).

P(ξ ∈ B) e− = e−(> ) .

which is the inverse to the matrix M = Eξ(i)