Professional Documents
Culture Documents
Textbook Lectures On Inductive Logic 1St Edition Jon Williamson Ebook All Chapter PDF
Textbook Lectures On Inductive Logic 1St Edition Jon Williamson Ebook All Chapter PDF
Jon Williamson
Visit to download the full and correct content document:
https://textbookfull.com/product/lectures-on-inductive-logic-1st-edition-jon-williamson/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
https://textbookfull.com/product/inductive-logic-
programming-29th-international-conference-ilp-2019-plovdiv-
bulgaria-september-3-5-2019-proceedings-dimitar-kazakov/
https://textbookfull.com/product/lectures-on-astrophysics-1st-
edition-steven-weinberg/
https://textbookfull.com/product/lectures-on-imagination-1st-
edition-paul-ricoeur/
https://textbookfull.com/product/lectures-on-geometry-woodhouse/
Lectures on New Testament Theology 1st Edition Baur
https://textbookfull.com/product/lectures-on-new-testament-
theology-1st-edition-baur/
https://textbookfull.com/product/lectures-on-lyapunov-
exponents-1st-edition-marcelo-viana/
https://textbookfull.com/product/network-flow-algorithms-1st-
edition-williamson/
https://textbookfull.com/product/lectures-on-convex-optimization-
yurii-nesterov/
https://textbookfull.com/product/twenty-lectures-on-algorithmic-
game-theory-1st-edition-tim-roughgarden/
LECTURES ON INDUCTIVE LOGIC
Lectures on Inductive Logic
JON WILLIAMSON
Professor of Reasoning, Inference and Scientific Method, University of Kent
3
3
Great Clarendon Street, Oxford, OX2 6DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© Jon Williamson 2017
The moral rights of the author have been asserted
First Edition published in 2017
Impression: 1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2016953026
ISBN 978–0–19–966647–8
Printed and bound by
CPI Group (UK) Ltd, Croydon, CR0 4YY
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
‘Tell me, what drove thee from the Table Round,
My brother? was it earthly passion crost?’
‘Nay,’ said the knight;‘for no such passion mine.
But the sweet vision of the Holy Grail
Drove me from all vainglories, rivalries,
And earthly heats that spring and sparkle out
Among us in the jousts’
( from The Holy Grail by Alfred Lord Tennyson)
PREFACE
Inductive logic seeks to determine the extent to which the premisses of an argument entail
its conclusion. This book offers an introduction to the topic and develops a new account of
inductive logic which appeals to developments in Bayesian epistemology (a theory about
how strongly one should believe the various propositions that one can express).
Chapter 1 introduces perhaps the simplest and most natural account of inductive lo-
gic, classical inductive logic, which is attributable to Ludwig Wittgenstein. An inductive logic
needs to capture two kinds of partial entailment: the first kind, logical entailment, quantifies
the extent to which premisses entail a conclusion in virtue of logical connections between
them; the second kind, inductive entailment, quantifies the extent to which a sample of past
observations entails the next outcome to be sampled. Unfortunately, as we shall see, clas-
sical inductive logic fails to adequately capture inductive entailment. Thus there is a need
to develop more sophisticated inductive logics.
Chapter 2 presents enough logic and probability theory for the reader to begin to study
inductive logic, while Chapter 3 introduces the ways in which logic and probability can be
combined in an inductive logic. Then, in Chapter 4, we encounter the most influential ap-
proach to inductive logic, due to W.E. Johnson and Rudolf Carnap. We see that while this
approach can account for inductive entailment, it does so at the expense of logical entail-
ment. Thus we remain in need of an inductive logic that captures both logical and inductive
entailment.
In Chapter 5 we see how an alternative approach to inductive logic follows rather nat-
urally from the philosophical theory of objective Bayesian epistemology. This approach
captures logical entailment because it extends classical inductive logic (Chapter 6). It cap-
tures inductive entailment because of its connections with statistical theory (Chapter 7).
Chapter 8 defends the approach by tackling several key criticisms that are often levelled at
inductive logic. Chapter 9 presents a formal justification of the objective Bayesian approach.
We take stock in Chapter 10.
This book is aimed at researchers and students who have some knowledge of elemen-
tary logic and probability, and who are familiar with mathematical reasoning and notation.
The material presented in this book is well suited to being delivered as a series of lec-
tures. An undergraduate course to philosophy students might omit the starred sections,
which are less central to the main thrust of the argument and can be more technical, and
might skim over some other mathematical details. For graduate students in philosophy,
these lectures could be given in conjunction with a more detailed consideration of Bayesian
epistemology (see, e.g., Williamson, 2010b) and other philosophical notions of probability
(Gillies, 2000). Alternatively, the material in this book could be treated alongside that of
Howson (2000) to provide a thorough introduction to induction from both the subjective
Bayesian perspective (Howson’s view) and the objective Bayesian perspective developed
viii | PREFACE
here. For mathematics students, these lectures could be given in conjunction with material
from Paris and Vencovská (2015), which is a work in the Johnson–Carnap tradition, and
the lectures could be extended by filling in the material sketched in §9.4.2, which can be
found in Landes and Williamson (2013, 2015). For students in computing, these lectures
could be augmented by a thorough discussion of inference in inductive logic using Bayesian
networks (Haenni et al., 2011; Korb and Nicholson, 2003; Neapolitan, 2004; Williamson,
2005).
ACKNOWLEDGEMENTS
I am grateful to the UK Arts and Humanities Research Council for supporting the research
project From objective Bayesian epistemology to inductive logic; that research, conducted in
collaboration with Teddy Groves and Jürgen Landes, informs much of this book. Lectures
based on the material in this book have been given at the fifth Indian School on Logic and
its Applications, the 2015 Spring School on Combining Probability and Logic, and to phil-
osophy students at the University of Kent. I am grateful to the audiences for their feedback.
The book has also benefited immensely from discussions with and comments from David
Corfield, Teddy Groves, Jürgen Landes, Jeff Paris, Soroush Rad, Christian Wallmann and
Michael Wilde.
I would like to thank Dan Taber, Keith Mansfield and all the team at OUP for ensuring
that the process of producing this book ran smoothly.
This book is dedicated to Kika, Charlie and Anouk, with thanks for everything.
CONTENTS
∗
The starred sections contain material which is less central to the main thrust of the argument and which can be
more technical.
xii | CONTENTS
4 Carnap’s Programme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.1 Conditionalizing on a Blank Slate 59
4.2 Pure and Applied Inductive Logic 61
4.3 Conditionalization 63
4.4 The Permutation Postulate 65
4.5 The Principle of Indifference 68
4.6 Which Value in the Continuum? 71
4.7 Which Continuum of Inductive Methods? 72
4.8 Capturing Logical Entailment 72
4.9 Summary 74
5 From Objective Bayesian Epistemology to Inductive Logic . . . . . . . . . 75
5.1 Objective Bayesian Epistemology 75
5.2 * Objective versus Subjective Bayesian Epistemology 77
5.3 Objective Bayesian Inductive Logic 81
5.4 * Language Invariance 85
5.5 * Finitely Generated Evidence Sets 91
5.6 Updating, Expansion and Revision 94
5.6.1 Maxent and Conditionalization 96
5.6.2 Maxent and KL-updating 101
5.7 Summary 103
6 Logical Entailment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.1 Truth Tables with Probabilities 105
6.2 Logical Irrelevance Revisited 108
6.3 Context and Chance Constraints 111
6.4 Constraints on Conditional Probabilities 115
6.5 Revision Under Constraints 118
6.6 Lottery and Preface Paradoxes Revisited 120
6.7 The Fundamental Inductive Pattern Revisited 122
6.7.1 A Surprising Consequence 122
6.7.2 An Otherwise Surprising Consequence 123
6.7.3 A Plausible Consequence 124
6.8 * Inferences in Predicate Inductive Logic 126
7 Inductive Entailment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.1 Syntactic Relevance 134
7.2 The Calibration Norm 135
7.3 Extended Example 138
7.4 Is this Application of Confidence Intervals Legitimate? 142
7.5 Uniqueness of the Interval 145
7.6 Loss of Information 146
7.7 Generalization 147
∗
The starred sections contain material which is less central to the main thrust of the argument and which can be
more technical.
CONTENTS | xiii
References 193
Index 199
∗
The starred sections contain material which is less central to the main thrust of the argument and which can be
more technical.
1
In this chapter we shall examine a rather natural and intuitive inductive logic, which we shall
call classical inductive logic, as a means to acquaint ourselves with the goals of inductive logic.
Although it will turn out (in §1.4) that classical inductive logic is not altogether viable, this
logic will provide a useful platform upon which to develop a more successful inductive logic
later on in the book.
a → b, b | a?
Another way of putting this question is: do the premisses a → b, b deductively entail (i.e.
force the truth of) the conclusion a?
We know the argument is invalid by considering its truth table:
a b a→b b a
T T T T T
T F F F T
F T T T F
F F T F F
The two left-most columns go through all the truth assignments to the atomic sentences,
while the next two columns give the corresponding truth values of the premisses and the
right-most column gives the corresponding truth value of the conclusion. The argument
is invalid because the third truth assignment makes the premisses true but the conclusion
false.
Hence this argument, which is often called affirming the consequent, is deemed fallacious,
and we write:
a → b, b | a.
While the argument seems a poor one from a deductive point of view, we can ask: to what
extent is the conclusion plausible, given the premisses? This can be viewed as a question of
partial entailment: to what extent do the premisses entail the conclusion? If, in the presence
of the premisses, some level of plausibility y attaches to the conclusion, then we can write:
a → b, b |≈ ay .
If Tr is the number of the truth-grounds of a proposition ‘r’, and if Trs the number of
the truth-grounds of a proposition ‘s’ that are at the same time truth-grounds of ‘r’,
then we call the ratio Trs : Tr the degree of probability that the proposition ‘r’ gives to
the proposition ‘s ’.
(Wittgenstein, 1922, §5.15)
Wittgenstein also noted that this proportion can be read off a standard truth table
(Wittgenstein, 1922, §5.151). Consider our example:
a b a→b b a
T T T T T
T F F F T
F T T T F
F F T F F
PATTERNS OF PARTIAL ENTAILMENT AND SUPPORT | 3
Here, one out of the two truth assignments that make the premisses true also makes the
conclusion true, so
a → b, b |≈ a1/2 .
We shall call the inductive logic with this notion of partial entailment classical inductive
logic, or CIL for short. We use this ‘classical’ for three reasons. First, it is perhaps the most
natural and basic inductive logic. Second, it was put forward by Wittgenstein at the same
time that he put forward the truth-table method for deciding propositional validity in what
is now known as classical deductive logic. Third, this inductive logic follows very naturally
from what is generally known as the classical interpretation of probability:
The theory of chance consists in reducing all the events of the same kind to a certain
number of cases equally possible, that is to say, to such as we may be equally undecided
about in regard to their existence, and in determining the number of cases favorable
to the event whose probability is sought. The ratio of this number to that of all the
cases possible is the measure of this probability, which is thus simply a fraction whose
numerator is the number of favorable cases and whose denominator is the number of
all the cases possible.
(Laplace, 1814, pp. 6–7)
a implies b
b true
a more credible
4 | CLASSICAL INDUCTIVE LOGIC
The key point to note about this inference is that the conclusion says that a becomes
more credible; it is thus asserting that there is a change in degree of partial entailment, that
is, a relation of support.
We saw above that the premisses entail the conclusion to degree 12 :
a → b, b |≈ a1/2 .
a b a
T T T
T F T
F T F
F F F
Two out of the four truth assignments render the conclusion true. Hence,
|≈ a1/2 .
Thus the premisses neither support nor undermine the conclusion—they fail to move its
probability away from 12 . The premisses are neutral with respect to the conclusion, written:
a → b, b ⊥
⊥ a.
It might appear, then, that classical inductive logic fails to validate the fundamental in-
ductive pattern. But Polya actually interprets the fundamental inductive pattern in a slightly
different way. He says, ‘the verification of a consequence renders a conjecture more cred-
ible’ (Polya, 1954, p. 5). This can be interpreted as: given a → b, learning b makes a more
credible. A very similar inference scheme was also advocated by Jaynes (2003, p. 4), who
gives much the same interpretation as Polya. Hempel (1945, p. 103) called this inference
pattern the entailment condition, and gives a similar interpretation: ‘Any sentence which is
entailed by an observation report is confirmed by it.’ Hempel argues that any reasonable
concept of partial entailment should satisfy this condition.
Thus the question is whether b changes the degree to which a is entailed, when a → b is
held fixed. If so, then b does support a relative to a → b and we can write:
[a → b], b a.
Here the square brackets specify what is held fixed, while denotes positive support.
Now,
a → b |≈ a1/3 .
This can be seen by the fact that one out of the three truth assignments that satisfy a → b
also satisfies a:
PATTERNS OF PARTIAL ENTAILMENT AND SUPPORT | 5
a b a→b a
T T T T
T F F T
F T T F
F F T F
On the other hand, we saw that a → b, b |≈ a1/2 . Hence b raises the degree of partial
entailment of a from 1/3 to 1/2, and we do indeed have that:
[a → b], b a.
Note that the maximum amount that the degree of partial entailment could be raised is
1 – 1/3 = 2/3. So b raises it by 1/2–1/3
2/3
= 25% of the maximum amount. This allows us to
quantify the degree to which b supports a, relative to a → b, and we can write:
[a → b], b 25% a.
In sum, then, classical inductive logic does indeed validate the fundamental inductive
pattern.
(a ∧ c) → b
b
a∧c
a b c (a ∧ c) → b b a∧c
T T T T T T
T T F T T F
T F T F F T
T F F T F F
F T T T T F
F T F T T F
F F T T F F
F F F T F F
The truth table shows several things. First, this is an invalid argument (lines 2,5,6).
Second, lines 1,2,5,6 show that:
(a ∧ c) → b, b |≈ a ∧ c1/4 .
6 | CLASSICAL INDUCTIVE LOGIC
(a ∧ c) → b |≈ a ∧ c1/7 .
So b supports a ∧ c, but not by as much as in the case of the fundamental inductive pattern
considered in §1.2.1:
This may be explained by the fact that classical inductive logic considers a ∧ c inherently
less plausible than a: we have that |≈ a ∧ c1/4 while |≈ a1/2 .
a implied by b
b false
a less credible
Polya expressed the inference thus: ‘our confidence in a conjecture can only diminish
when a possible ground for the conjecture is exploded’ (Polya, 1954, p. 20). In the inference,
a is the conjecture and b is the ground for the conjecture.
Here is the truth table for this inference:
a b b→a ¬b a
T T T F T
T F T T T
F T F F F
F F T T F
b → a, ¬b |≈ a1/2 .
b → a |≈ a2/3 .
[b → a], ¬b 25% a.
Thus classical inductive logic does indeed validate this inductive pattern of Polya’s.
PATTERNS OF PARTIAL ENTAILMENT AND SUPPORT | 7
1.2.4 Analogy
Polya (1954, §13.9) proposes:
a analogous to b
b true
a more credible
Here ‘a is analogous to b’ is understood as there being some common ground g, such that
g → a and g → b. That is,
g→a
g→b
b true
a more credible
Classical inductive logic validates such an inference, as we shall now see. First,
g → a, g → b, b |≈ a2/3 .
g a b g→a g→b b a
T T T T T T T
T T F T F F T
T F T F T T F
T F F F F F F
F T T T T T T
F T F T T F T
F F T T T T F
F F F T T F F
g → a, g → b |≈ a3/5 .
Hence,
[g → a, g → b], b 16.6̇% a.
Note that classical inductive logic also validates the inference if the first two premisses
are not held fixed:
|≈ a1/2 ,
8 | CLASSICAL INDUCTIVE LOGIC
so
g → a, g → b, b 33.3̇% a.
Exercise
(i) To what extent does a ∨ b entail a?
(ii) To what extent do the premisses of the following argument entail its conclusion? To
what extent do they support the conclusion?
(a ∨ b) → c
(a ↔ c) → b
c
r ¬r
chemotherapy +6 –2
radiotherapy +4 –1
cancer will otherwise recur. Bayesian decision theory, which is arguably our best guide to
decision making in such a scenario, says that the consultant should choose the act which
has maximum expected utility. Thus they should choose chemotherapy if
where P(r) and P(¬r) are the probabilities of r and ¬r respectively. According to the rules
of probability, P(¬r) = 1 – P(r), so they should choose recurrence if
that is, if
The key task that remains, then, is to determine P(r), the probability of recurrence, from
available evidence.
Suppose, for example, that the consultant’s evidence consists of the facts that the cancer
metastasized (m) and that metastasis together with the presence of a particular biological
marker (b) ensure recurrence. (In reality, of course, a consultant’s evidence will be much
more complex, but this simplified example will suffice to illustrate the role, if not the extent,
of the available evidence. A more realistic version will be explored at the end of §6.4.) The
task is to determine what degree of plausibility the evidence gives to r. This is really just a
question of inductive logic:
m, m ∧ b → r |≈ r? .
According to classical inductive logic, we need to build a truth table to answer this
question:
m b r (m ∧ b) → r m r
T T T T T T
T T F F T F
T F T T T T
T F F T T F
F T T T F T
F T F T F F
F F T T F T
F F F T F F
Two of the three truth assignments that satisfy the premisses also satisfy the
conclusion, so
m, m ∧ b → r |≈ r2/3 .
10 | CLASSICAL INDUCTIVE LOGIC
Now 2/3 > 1/3, so giving chemotherapy maximizes expected utility and the consultant
should choose to give the patient chemotherapy.
In sum, inductive logic is essential to decision making. Decision making requires ascer-
taining the probabilities of the various possible outcomes that make a difference to the
utilities of the available acts. These probabilities are not normally a part of one’s evidence,
but need to be derived from the available evidence. It is inductive logic that tells us how to
derive these probabilities: inductive logic tells us the extent to which evidence entails the
various outcomes relevant to the decision problem. Thus we need to investigate whether
there is some inductive logic—such as classical inductive logic—that stands out as being
most appropriate.
Something like Carnap’s theory [of inductive logic] would be required if an electronic
reasoning machine is ever built.
(Good, 1950, p. 48)
This is because the notion of partial entailment has such wide applicability in artificial
intelligence. Consider a partial entailment relation of the form:
ϕ1 , . . . , ϕk |≈ ψ y .
In natural language processing, ψ might be the hypothesis that a word has a certain
meaning, while ϕ1 , . . . , ϕk might contain information about the language as well as con-
textual information about the words preceding the word in question (Berger et al., 1996).
There are countless other potential applications of inductive logic in artificial intelli-
gence. Consequently, a viable inductive logic would be of enormous interest to researchers
in these fields.
|≈ Br101
1/2
.
12 | CLASSICAL INDUCTIVE LOGIC
This is because half the truth assignments make this proposition true.
Br101
T
F
This degree of partial entailment seems reasonable enough, in the absence of any evi-
dence that indicates otherwise. But classical inductive logic will say that, in the presence of
premisses which assert just that 100 ravens have all been found to be black, the degree of
partial entailment remains 12 :
This is because one out of the two truth assignments that make the premisses true also
makes the conclusion true:
Thus, according to classical inductive logic the premisses are neutral with respect to the
conclusion, that is, they neither support nor undermine the conclusion:
Br1 , . . . , Br100 ⊥
⊥ Br101 .
This seems totally unreasonable. Observing 100 ravens and finding them all to be black
should lend some support to the conclusion that the next raven will be found to be black.
This inability to capture learning from experience is a clear deficiency of classical inductive
logic.
Although Carnap highlighted the significance of this problem for inductive logic, a simi-
lar point had previously been made by George Boole, in the context of drawing balls from
an urn with replacement:
It follows, therefore, that if the number of balls be infinite, and all constitutions of the
system be equally probable, the probability of drawing m white balls in succession will
be 21m , and the probability of drawing m + 1 white balls in succession 2m+1 1
; whence
the probability that after m white balls have been drawn, the next drawing will furnish
a white one, will be 12 . In other words, past experience does not in this case affect future
expectation.
(Boole, 1854, pp. 371–2)
INDUCTIVE ENTAILMENT AND LOGICAL ENTAILMENT | 13
Wittgenstein pointed out that this phenomenon is due to the propositions being logically
independent, that is, having no propositional variables in common, when viewed from the
perspective of propositional logic:
Classical inductive logic appeals to truth-table methods, which are only suitable for de-
ciding entailment in propositional logic. Thus classical inductive logic treats the Bri as
distinct propositional variables. Since these propositional variables have no logical struc-
ture from the point of view of propositional logic, classical inductive logic cannot capture
any connection between them. Because the premisses are logically independent of the con-
clusion, classical inductive logic cannot but deem them to be neutral with respect to the
conclusion. This deficiency, then, points to an inadequacy of propositional inductive logic,
and motivates a move to predicate logic to try to capture the connection between the Bri .
This was essentially what motivated Carnap’s programme for inductive logic.
That inductive logic needs to capture these two kinds of partial entailment motivates the
following slogan:
(1945) abandoned the classical notion of partial entailment to try to capture learning from
experience.
We see then that the field of inductive logic underwent a schism, with some researchers
focussing on logical entailment and others focussing on inductive entailment. In fact,
Wesley Salmon argued that it is not possible to capture both notions of partial entailment
in a single inductive logic. Here, Salmon uses ‘partial entailment’ to refer to what we call
‘logical entailment’, and ‘degree of confirmation’ for what we call ‘partial entailment’:
The dilemma seems to be as follows. In order for an inductive logic to capture logical
entailment, the following principle must hold:
Logical Irrelevance. If ϕ and ψ are logically independent then they are neutral with
respect to each other (i.e., neither changes the degree of partial entailment of
the other).
Kohta tuli äitikin ulos, alkoi torua häntä, tempasi nuken hänen
kädestään, löi sen kappaleiksi ja lähetti Dinan jotakin toimittamaan.
Koko kylä oli kohta siinä koolla: poikia, tyttöjä, vanhoja eukkoja ja
ikämiehiäkin. Kaikki solkkasivat vierasta kieltään.
— Ai urus, ai Ivan!
Tästä lähtien Žilinin kätevyydestä kulki maine, että hän oli oikea
mestari. Alkoi tulla väkeä hänen luokseen kaukaisista kylistä; ken toi
pyssyn tai pistoolinlukon, ken kellorähjän korjattavaksi. Toipa isäntä
hänelle työkalujakin, hohtimet, poran ja viilan.
Žilin alkoi kysellä isännältä lähemmin ukosta, mikä hän oikein oli.
Isäntä selitti:
— Siinä vasta mies oli! Ei olekaan hänen veroistaan džigittiä näillä
mailla, monta venäläistä on ottanut hengiltä ja rikas oli aikoinaan.
Kolme eukkoa on ukolla ollut ja kahdeksan poikaa. Kylässä asuivat
kaikki yhdessä. Saapuivat sitten venäläiset, hävittivät kylän ja
tappoivat seitsemän poikaa. Yksi poika vain jäi henkiin ja antautui
venäläisille. Ukko matkusti perästä ja heittäytyi itsekin venäläisten
armoille. Eleli heidän parissaan kolmisen kuukautta, löysi sieltä
poikansa, surmasi hänet omin käsin ja sitten pakeni. Sen päivän
jälkeen ei ole enää sotaa käynyt, vaan Mekkaan matkusti Allahia
rukoilemaan. Siksi nyt turbaania kantaa. Mekassa käynyttä sanotaan
Hadziksi, ja sellaisilla on turbaani päässä. Ukko ei suosi teikäläisiä.
Käskee tappamaan sinut, mutta minun ei käy sitä tekeminen, kun
olen maksanut sinusta rahat. Pidän sinusta, Ivan, en hennoisi
tappaa, en edes luotani laskea, ellen olisi antanut sanaani. — Isäntä
nauraa virnisteli virkkaessaan venäjäksi: »Sinu, Ivan, hyvä, minu,
Abdul, hyvä.»
IV
Žilin eleli näin jälleen kuukauden päivät. Hän käveli päivisin ympäri
aulia (kylää) tai askarteli jotakin. Mutta kun yö saapui ja aulissa
kaikki äänet vaikenivat, hän vajassaan tonki maata kaivellen. Kivet
tekivät kaivaessa vastusta, mutta hän kuopi kivet irti viilan tyngällä ja
sai lopulta kaivetuksi seinän alle aukon, josta sopivana hetkenä voi
ryömiä ulos. »Kunpa vain tuntisin, seudut ja tietäisin, mihin päin
lähteä», hän mietiskeli. »Tataritpa eivät virka mitään.»
— Älä, koskee!
Žilinin oli raskasta kulkea, hänenkin jalkansa olivat verillä ja hän oli
ihan uupumaisillaan. Hän kumartui, kohotteli Kostylinin paremmin
selkäänsä ja kantaa retuutti häntä pitkin tietä.