An Axiomatic Theory of Inductive Inference∗

Luciano Pomatto † Alvaro Sandroni ‡

April 17, 2017

Abstract

This paper develops an axiomatic theory of induction that speaks to the

recent debate on Bayesian orgulity. It shows the exact principles associated

with the belief that data can corroborate universal laws. We identify two

types of disbelief about induction: skepticism that the existence of universal

laws of Nature can be determined empirically, and skepticism that the true

law of Nature, if it exists, can be successfully identified. We formalize and

characterize these two dispositions towards induction by introducing novel

axioms for subjective probabilities. We also relate these dispositions to the

(controversial) axiom of sigma-additivity.

∗We are grateful to Nabil Al-Najjar, Frederick Eberhardt and Alvaro Riascos. All remaining
errors are our own.
†Division of the Humanities and Social Sciences, Caltech, Pasadena, CA, 91125. (e-mail:

luciano@caltech.edu)
‡Department of Managerial Economics and Decision Sciences, Kellogg School of Management,

Northwestern University, Evanston, IL 60208 (e-mail: sandroni@kellogg.northwestern.edu).

1



1 Introduction

We seek an axiomatic understanding of specific problems of induction. Informally,

induction is taken to mean the process of using empirical evidence to validate

general claims, and, for our purposes, it is critical to differentiate between two

types of epistemic skepticism about induction.

One may doubt it is possible to know whether Nature abides by any law.1 Any

empirical regularity may be a temporary fluke. Hence, patterns can suggest, but

not prove, the existence of universal laws. So, one may ascribe non-vanishing odds

to the idea that Nature does not follow any law, no matter how numerous and

consistent the data may grow to be. We refer to this disposition as Humean skep-

ticism, with the caveat that we do not claim to provide a complete representation

of Hume’s (and other authors) actual statements.

In addition, even if it is taken for granted that Nature abides by a law, one

may be skeptical that such law can be inferred with arbitrarily high precision,

even when the data grows without bounds. Let’s say that in each period either

0 or 1 must occur and that 1 has been observed every period, over a long time,

say t periods. The data is consistent with the law “Nature produces only 1,”

and with the law “Nature produces 1 until period t and 0 afterwards,” among

(infinitely) many other laws. So, one may mantain a non-vanishing doubt that

empirical evidence can validate a specific law, even under the assumption that

the data follow one. We refer to this form of skepticism as Goodman’s skepticism,

with the same caveat as above.

We consider a probabilistic framework in which an agent, named Bob, is en-

dowed with a coherent view of world (i.e., a finitely additive probability measure)

over paths (i.e., infinite binary sequences). As the data unfolds, Bob updates

his view of the world through Bayes’ rule.2 No restrictions are placed on which

1The definition of “law” is subjective as we make clear in the main text.
2This framework follows de Finetti’s (1970) viewpoint that inference involves personal judge-

ments of likelihood that must be formalized in a coherent way. See de Finetti (1970) for a
connection between coherent views of the world and Dutch books. We make no original attempt
to justify Bayesianism and subjectivism.

2



paths may be produced. So, no relationship between past and future are, a priori,

required (apart from the idea that either 0 or 1 occurs each period). Bob is not

dogmatic about induction either. He believes that the data may or may not follow

eternal laws.

Under the lenses of this formal framework, we formalize Hume’s and Good-

man’s skepticisms by introducing two novel axioms for subjective probabilities.

These axioms refer to Bob’s belief as the data unfolds and becomes arbitrarily

numerous. Bob’s view of the world is inductive in the sense of Hume, as we define

it, if, under data compatible with laws, Bob expects to become almost convinced

that Nature indeed follows laws. This axiom rules out Humean skepticism about

induction. Bob’s view of the world is inductive in the sense of Goodman if he

expects to successfully identify Nature’s law up to a vanishing degree of error,

conditional on Nature abiding by one. This axiom rules out Goodman’s skepti-

cism about induction.

A natural starting line of inquiry is the extent of the connection between the

two problems of induction. We start by asking whether Goodman’s skepticism

implies Hume’s skepticism and the converse implication. Neither is true. Some

coherent views exhibit Goodman’s skepticism, but not Hume’s skepticism and,

conversely, some coherent views exhibit Hume’s skepticism, but not Goodman’s

skepticism. Thus, these two types of skepticism are not logically nested.

Of particular interest are the coherent views that express skepticism in the

sense of Goodman, but not in the sense of Hume. If, say, confronted with the

question of whether or not the data is generated by a Turing machine, such views

of the world express conviction that with enough data it is possible to make this

determination with near certainty. In spite of this remarkable confidence on the ca-

pacity of Bayes’ rule to address this apparently insurmountable inference problem,

the same view of the world, if confronted with the (arguably simpler) question

of which Turing machine generates the data, assuming that one does, remains

skeptical that this determination can be made with arbitrarily high precision.

The celebrated theorems of Levy (1937), Doob (1949) and Blackwell and Du-

bins (1962) make clear that under σ-additivity a Bayesian must believe that his

opinion about a given hypothesis will converge to the truth. In particular, σ-

additivity excludes both Hume’s and Goodman’s skepticism, and therefore it im-

3



plies a form of “Bayesian orgulity” (Belot, 2013). Different results were obtained

by Elga (2015), Juhl and Kelly (1994), and Kelly (1996), among others, who have

shown that there exist non σ-additive coherent views of the world which allow for

Humean skepticism. Hence, in the absence of σ-additivity, epistemic skepticism

is allowed.

Our results reveal a complex relationship between skepticism and subjective

probability. There are non σ-additive coherent views that rule out Humean skep-

ticism and others yet that rule out Goodman’s skepticism. Thus, the spectrum

of coherent views is rich enough to allow, at the same time, both orgulity and

skepticism about induction. In particular, in the absence of σ-additivity, orgulity

and skepticism are allowed. Orgulity is not an exclusive property of σ-additivity

and may hold with or without it. This is a difficulty for a clear-cut theory of

induction that seeks the root causes of orgulity and skepticism about induction.

We show that while lack of σ-additivity does not assure Hume’s skepticism

and Goodman’s skepticism, it always assures skepticism in at least one of these

two ways. This is demonstrated by the structure theorem for coherent views of

the world. It shows that a coherent view is inductive in the sense of Hume and

in the sense of Goodman if and only if it is σ-additive. Thus, σ-additivity is the

definitive condition that assumes away both Hume’s and Goodman’s skepticism

about induction. It is not necessary to rule out either type of skepticism, but it

is required to rule out both types simultaneously.

The interpretation of the structure theorem requires considerable care. The

equivalence between induction and σ-additivity may suggest that the problems of

how to conceptually justify either induction or σ-additivity are, in fact, one and

the same problem and that σ-additivity is the root and only cause of the conviction

in the ultimate success of induction. This reading of the structure theorem may

prove incomplete. Consider an alternative approach, where the focus is not on

using data to ultimately, i.e. in the limit as data grows, uncover eternal laws

of Nature, but on making predictions within a practical (i.e., bounded) future.

Consider the case where a long sequence of 1’s has been observed. One may wonder

if “Nature produces only 1’s.” One may also wonder whether or not “Nature will

produce only 1’s for the next 1000 periods.” Our last result concerns the latter

case, where Bob remains agnostic about the validity of universal claims, but asks

4



whether regularities in the past can be used to make sharp predictions about

a bounded future. This result shows that any coherent view of the world, no

matter how it is formed, must be confident that multiple repetitions of Bayes’

rule transform pattern data into a near infallible guide to a bounded future.

Moreover, after numerous enough data there must be high confidence on lim-

ited, but correct, inductive inferences. This holds even if, a priori, no assumption

is made on the relationship between past and future in the sense that the data

may unfold according to any path, including those without patterns. It also holds

even if Bob is a skeptic in regards to the use of data to ultimately validate spe-

cific or general laws. Eventually there must be high confidence that the past is

a limited, but successful, guide to the future. This conclusion follows from con-

ditional probability alone, and holds for any coherent view of the world. Thus,

some confidence in inductive inference follows from coherence.

1.1 Literature on Bayesian Orgulity

The paper speaks to the recent debate on “Bayesian orgulity,” originated with

Belot (2013, 2015). Central to Belot’s thesis is the argument that the convergence

results of Levy, Doob and Blackwell and Dubins are proof that Bayesianism implies

epistemic arrogance. The debate has spurred different views. Huttegger (2015)

argued that the issue of convergence to the truth should be put in the context of

a long, but finite horizon. Weatherson (2015) revisited Belot’s argument from the

perspective of Bayesian imprecise probability. The work closest to this paper is

Elga (2015), who showed the existence of non σ-additive subjective probabilities

expressing epistemic humility.

This paper is also connected to the work of Kelly (1996), who formalized the

connection between inductive inference and finitely additive probabilities, to the

work of Gilboa and Samuelson (2012), who analyzed how subjectivity can enhance

inductive inference, and to Al-Najjar, Pomatto and Sandroni (2014), who study

how different dispositions towards induction can affect incentive problems.

5



2 Basic Concepts and Results

2.1 Patterns and Coherence

An agent, named Bob, observes, in every period, one of two possible outcomes, 0

or 1. The set Ω = {0, 1}∞ is the set of all paths or infinite histories of outcomes.
Given a path ω and a time t, we denote by ωt the set of paths that share with

ω the same first t outcomes. We call ωt a finite history. We fix an algebra Σ

of subsets of Ω (subsets of Ω mentioned in the text belong to Σ, even when not

stated explicitly). The agent is endowed with a finitely additive probability P on

Σ.3 The measure P captures Bob’s subjective viewpoint on how the outcomes

will evolve. We refer to P as a coherent view of the world.

Some paths are governed by a law or pattern and some are not. For instance,

the path 1∞ = (1, 1, 1, ...) follows the law “Nature produces only the outcome 1.” A

classic example of a pattern is given by periodic paths, defined by repeated cycles

as in (1, 0, 1, 0, ...) or (1, 1, 0, 0, 1, 1, ...) or, more generally, eventually periodic path,

i.e., sequences that are periodic after some point in time. Both examples are

subsumed by the class of computable paths, which consists of all sequences that

can be generated by a Turing machine (i.e., all paths that are the output of some

finite program running on a computer with unlimited storage).

In order to speak of induction it is critical to demarcate between paths gov-

erned by a law from paths that do not follow any discernible pattern. This dis-

tinction can be made in many different ways and the precise way in which this

determination should be made is orthogonal to the central questions in this pa-

per. So, we need not take a definitive stance of this matter. Instead, we assume

that the final determination of what constitutes a law is subjective. That is, Bob

determines which set of paths A ⊆ Ω are the ones that abide by a law. The
complement of A are the set of paths that according to Bob do not follow any

pattern. For simplicity, we often refer to paths in A as laws and to paths not in

A as non-laws.

We make the following assumptions on A and P .

3That is, a function P : Σ → [0, 1] such that P (Ω) = 1 and for every pair of disjoint sets E1
and E2 in Σ it satisfies P (E1 ∪E2) = P (E1) + P (E2).

6



Assumption 1. A is countable.

While flexible enough to capture many formal definitions of pattern, including

the set of periodic, eventually periodic or computable paths, the assumption is

not, however, without loss of generality. It greatly simplifies the analysis because

it rules out both conceptual and technical difficulties that are outside the scope

of this paper. The main implication of Assumption 1 is that it allows a view

of the world to assign strictly positive probability to each lawlike path. If, for

example, A was uncountable, then Bob would have to assign zero probability to

most individual laws. Formally:

Remark 1 For any coherent view of the world, there can be, at most, countably

many paths with strictly positive probability.

The result applies to Bob’s view of the world both before and, by Bayes’ rule,

after the data is observed.

An alternative approach, which allows to capture more complex inference prob-

lems, is to consider non-deterministic laws. In Section 7 we discuss this alternative

approach and, in particular, the difficulties it involves.

Assumption 2. P({ω}) > 0 for every ω ∈ A, and P(Ac) > 0.

Bob believes that any law in his set A is, a priori, possible. Bob also does not

rule out the possibility that Nature does not follow any pattern. This assump-

tion enables Bayesian inferences about universal laws.4 Assumption 2 simplifies

the notation and the statement of some of the results, but can be substantially

weakened. Formally, all results in the paper continue to hold if their statements

are modified by replacing the condition “for every ω ∈ A” with “for every ω ∈ A
such that P ({ω}) > 0”.

Assumption 3. Given any finite history ωt, A∩ωt 6= ∅ and Ac ∩ωt 6= ∅.

Given any finite history ωt, no matter how complex or simple it may be, there

are infinitely many laws that are compatible with it (i.e., there are infinitely many

4As is well known, Bayesian inference about an hypothesis requires the latter to have initial
positive probability. See, for example, Broad (1918), Wrinch and Jeffreys (1919) and Edgeworth
(1922), among others. See also Zabell (2011) for these and other references.

7



laws ω ∈ A such that the first t outcomes are equal to ωt) as well as uncountably
many non-laws that are also compatible with it. So, for any data, Bob can never

rule the hypothesis that Nature abides by laws nor the hypothesis that it does

not. This captures the idea that there are many different ways in which past and

future can relate to each other. The history 1, 1, 1, 1, 1 is equally compatible with

the law “always 1” and with the law “1 in the first 5 periods and 0 afterwards.”

In sum, the assumption ensures that it is not possible to deduce conclusively,

from any finite data, whether or not Nature abides by laws, nor if so, to which

law. Hence, it makes clear that induction, in this paper, refers to probabilistic

inferences that can approach certainty, but never reach it in finite time. This

assumption is also satisfied by all canonical definitions of patterns, and so is useful

for the interpretation of the results. However, our results remain unchanged under

the weaker condition that there exists (at least) one law ω̄ with the property that

for every t there is a law ω ∈ A distinct from ω̄ such that ωt = ωt. So, upon
observing t outcomes matching the path ω̄, Bob cannot conclude with certainty

that the law, if it exists, must be ω̄.

Finally, we emphasize that while our main examples of laws and patterns refer

to celebrated ideas such as Turing machines and periodicity, our results would

continue to hold even if Bob had an eccentric understanding of what is a law

or pattern. That is, none of our results depend on the labels given to laws and

non-laws, neither do they depend on the nature of the paths that are categorized

as laws and non-laws (provided that Assumption 1 on the existence of at most

countably many laws hold). The key point is that whatever Bob’s understanding

of what constitutes laws and patterns might be, he privileges paths in the set A by

assigning strictly positive probability to each of them. This is a non-judgemental,

but meaningful, differentiation of laws and non-laws because, as we discussed,

only countably many paths can have strictly positive probability.

We fix for the remainder of the paper a set of paths A satisfying Assumptions 1

and 3. We also restrict the attention to views of the world that satisfy Assumption

2.

8



2.2 Induction and the Separation Theorem

We now formalize specific forms of induction.

Definition 1 A coherent view of the world P is inductive in the sense of Hume

if for every path ω ∈ A,

P
(
A | ωt

)
→ 1 as t →∞. (1)

From sufficient data with a pattern, Bob ultimately concludes, with proba-

bility approaching certainty, that Nature must follow some law. A view of the

world that violates (1) is such that the probability of the set A of lawlike paths

remains bounded away from 1, regardless of the number of realizations. Any such

worldview captures what we refer to as Humean skepticism: Bob maintains a

non-vanishing doubt that perhaps Nature does not work through eternal laws, no

matter how consistent and numerous the data he observes.

At each point in time, the observed finite history ωt is consistent with a path

following a pattern as well as with a path that does not follow a pattern. So, by

Bayes’ rule, even a view of the world that is inductive in the sense of Hume will

always attach non-zero probability to the event that Nature does not abide by laws

(under Assumption 2). What distinguishes between skepticism and inductivity in

the sense of Hume is whether or not Bob’s doubt on the regularity of Nature

vanishes, as the number of observations that exhibit a pattern goes to infinity.

Definition 2 A coherent view of the world P is inductive in the sense of Good-

man if for every path ω ∈ A,

P
(
{ω} | A∩ωt

)
→ 1 as t →∞. (2)

If it is granted that Nature abides by some law and sufficient data with a

pattern is observed, Bob infers Nature’s true law with increasing precision, and

ultimately concludes it is eternal. A view of the world that violates (2) captures

what we refer to as Goodman’s skepticism: even assuming that an underlying

law exists and that extensive evidence is available, Bob remains skeptical he will

9



ever be able to perfectly single out the data generating law with arbitrarily high

confidence.

As in the case of induction in the sense of Hume, at no point in time Bob’s

inference is solved perfectly. He will always attach nonzero odds to multiple paths.

However, a view of the world that is inductive in the sense of Goodman is confident

he must ultimately put almost all mass on the law generating the data.

The distinction we make here need not be seen as the formal counterpart of the

classic and the new riddle of induction (see Goodman (1955) and Stalker (1994),

for a discussion) and the above terminology is used mostly as a mnemonic device.

Fundamentally, we ask two direct inference questions: Within the present proba-

bilistic framework can one tell, from sufficient data and with arbitrary precision,

(1) whether Nature must abide by a law and (2) if so, which law?

We now examine the logical connection between these two questions.

The Separation Theorem There exist coherent views of the world that are in-

ductive in the sense of Hume but not in the sense of Goodman, and views

that are inductive in the sense of Goodman but not in the sense of Hume.

The Separation Theorem shows that Hume’s skepticism and Goodman’s skep-

ticism are not logically nested. One does not imply the other. In the Appendix

we provide simple examples of views satisfying only one of the properties. Given

the separation theorem, it is meaningful to consider those coherent views of the

world that express both types of faith in induction.

Definition 3 A coherent view of the world P is inductive if it is inductive in the

sense of Hume and is inductive in the sense of Goodman.

Under an inductive view of the world, skepticism about induction vanishes.

Bob interprets evidence consistent with a pattern as a sign of the existence of an

underlying law of Nature, and expects further evidence to allow him to single out

the correct law with virtually exact precision. So, inductive views express great

confidence in the power of empirical evidence to predict the future. This can be

expressed as follows:

10



Definition 4 A coherent view of the world P is confident that enough pattern

data transforms the past into a near infallible guide to the future if for every path

ω ∈ A,
P
(
{ω} | ωt

)
→ 1 as t →∞. (3)

So, conditional on a sufficient long pattern data ωt, the future is forecasted

with arbitrarily high degree of certainty.

Remark 2 A coherent view of the world P is inductive if and only if it is confident

that enough pattern data transforms the past into a near infallible guide to the

future.

So, partial induction is the necessary and sufficient condition for confidence

that sufficient pattern data is a near perfect guide to the future. Remark 2 delivers

an initial characterization of induction that will prove useful.

3 Orgulity and σ-additive Coherent Views

This section examines inductive properties of σ-additive coherent views. These

results are known and adapted to our framework. We refer to known results as

“propositions” and to novel ones as “theorems.”

Proposition 1 If a coherent view of the world is σ-additive then it is inductive.

The proof of this result can be found in Kelly (1995). Under σ-additivity, after

multiple observations consistent with a pattern, Bob infers Nature’s underlying

law with arbitrary accuracy and concludes with almost certainty that Nature

cannot follow a different law. However, σ-additivity entails even stronger forms

of faith in induction.

Definition 5 A coherent view of the world P is completely inductive in the sense

of Hume if

P
(
A | ωt

)
→ 1 as t →∞, for every path ω ∈ A

and

P
(
Ac | ωt

)
→ 1 as t →∞, for P-almost every path ω in Ac (4)

11



A coherent view of the world that is completely inductive in the sense of

Hume expresses full confidence that, with sufficient data, laws and non-laws can

be distinguished empirically and with near certainty. So, complete induction in the

sense of Hume is an expression of confidence that a remarkably difficult inference

problem can be resolved with arbitrarily high precision.

Proposition 2 Any σ-additive coherent view of the world is completely induc-

tive in the sense of Hume.

This result has led Belot (2013) to speak of “Bayesian orgulity.” The basic

inference problem is difficult. Yet, σ-additive coherent views are confident that

finite, but long enough, data suffices to determine with arbitrarily high precision

whether or not Nature is governed by a law.

In addition,

Definition 6 A coherent view of the world P is completely inductive if it is

completely inductive in the sense of Hume and inductive in the sense of Goodman.

Combining propositions 1 and 2 yields:

Corollary 1 If a coherent view of the world P is σ-additive then it is completely

inductive.

Under σ-additivity, Bob must express the following viewpoint on induction:

“I do not know whether Nature works through laws or not, but given sufficient

data I will find out with an arbitrarily high degree of certainty. If Nature generates

the data based on a law, I will ultimately conclude that Nature works through laws

and uncover the law Nature abides by, up to a vanishing error. If the data is not

governed by a law, then, in the long run, I will become near certain that Nature

does not follow laws. This is true even though any finite data is simultaneously

consistent with countably many laws and uncountably many non-laws.”

So, under σ-additivity, Bob believes that Bayes’ rule resolves these essential

problems of induction. With sufficient data, Nature’s law is eventually uncovered.

A false inference of laws, when Nature follows none, is unlikely. The intuition

behind these results is as follows: First, let’s assume, for simplicity, that Nature

12



abides either by the law “always 1” or to a law “1 until period t and 0 thereafter,”

for some t > 0. No sequence of 1’s, either large or small, suffices to infer Nature’s

law conclusively, but there is a crucial difference between a short and a long

sequence. Ex-ante, the odds of the law “always 1” are fixed and strictly positive.

The odds of the laws “1 until some period t ≥ m and 0 thereafter” are arbitrarily
small if m is sufficiently high. It is here that the assumption of σ-additivity is

used. Under σ-additivity, such tail events must be unlikely. It now follows, by

Bayes’ rule, that conditional on a sufficiently long sequence of 1’s, the likelihood of

the law “always 1” eventually dominates the likelihood of any competing standing

theories. Thus, under σ-additivity, Bob cannot express Goodman’s skepticism.

The intuition regarding Hume’s skepticism is related, but not identical. As-

sume, for simplicity, that Nature either abides by the law “always 1” or does not

abide by any law. Once again, no sequence of 1’s, either large or small, suffices for

conclusive inference. For any sequence of 1’s, no matter how long, there are still

many non-laws that are consistent with it. However, the set of non-laws that are

consistent with consecutive 1’s until period t, shrinks monotonically to the empty

set as t goes to infinity. This follows because no non-law is consistent with an infi-

nite sequence of 1’s. So, under σ-additivity, the ex-ante odds of the set of standing

non-laws (i.e., those consistent with data of consecutive 1’s until period t) goes

to zero as t goes to infinity. Hence, by Bayes’ rule, conditional on a sufficiently

long sequence of 1’s, the relative likelihood of the law “always 1” is much higher

than the competing and still standing non-laws. Thus, under σ-additivity, Bob

cannot express Hume’s skepticism. Finally, the intuition regarding property (4) is

also similar. The set of laws consistent with non-pattern data of length t shrinks

monotonically to the empty set as t goes to infinity (because no law is consistent

with an infinite sequence of non-pattern data). Thus, under σ-additivity, it is

unlikely that laws are consistent with long non-pattern data. Hence, property (4)

holds and so does complete induction in the sense of Hume.

4 Orgulity and General Coherent Views

As shown, σ-additive coherent views rule out skepticism about induction. We

now consider Bob’s conclusions about the ultimate fate of multiple repetitions of

13



Bayes’ rule for general, not necessarily σ-additive, coherent views of the world.

We start with an important result, a corollary of Elga (2015) (related results can

also be found in Juhl and Kelly (1994) and Kelly (1996)): 5

Proposition 3 Let ε > 0. There exists a coherent view of the world P such that

P
(
A | ωt

)
≤ ε for every t and every ω ∈ A.

The view P displays a complete failure of induction in the sense of Hume. Un-

der P , no evidence can overturn Bob’s initial pessimistic belief on the existence

of laws. Hence, σ-additivity suffices to rule out Hume’s skepticism about induc-

tion, and this condition cannot be completely disposed of. Elga (2015) shows

that not all coherent views are inductive in the sense of Hume. On the other

hand, the separation theorem shows that some non σ-additive coherent views are

inductive in the sense of Hume. Moreover, there are also coherent views that are

not σ-additive, but nevertheless are inductive in the sense of Goodman. Lack of

σ-additivity does not assure skepticism in the sense of Hume and does not assure

skepticism in the sense of Goodman either. Other strong forms of induction can

also be obtained without σ-additivity.

The Complete Humean Induction Theorem There exists a coherent view

of the world P that is not σ-additive but is completely inductive in the

sense of Hume.

Insomuch as confidence about Humean induction must be granted under σ-

additivity, the same confidence must also be granted without σ-additivity, for

some coherent views of the world. An example of such a view can be found in the

Appendix.

Consider, for instance, the case where A is the set of computable paths. The

Complete Humean Induction Theorem shows that some, but not all, coherent

views of the world express the belief that even a fundamental problem such as

whether or not Nature can be reduced to a Turing machine can be solved (up to

5The construction in Elga (2015) does not immediately apply to our framework (where As-
sumptions 1-3 hold). For completeness, we provide an alternative construction in the Appendix.

14



a vanishing error) empirically, even in the absence of σ-additivity. In this sense,

Bayesian orgulity is not restricted to σ-additivity. It extends to other coherent

views of the world as well.

5 The Axiomatization of Induction

The Separation and the Complete Humean Induction theorems present a difficulty

for the development of a crisp theory of inductive inference. The difficulty is that

confidence on solving induction problems is not only a product of well understood

conditions such as σ-additivity, but also of properties coherent views might have,

which are less understood and intuitively less clear. The Complete Humean In-

duction Theorem is particularly challenging because it shows that confidence on

empirical solutions to strong forms of inference problems can be obtained under

conditions other than σ-additivity. However, let Σ̄ be the smallest algebra that

contains all finite histories and all singletons {ω} for ω ∈ A. This is the small-
est algebra which allows to express property (3), which is equivalent to a view

P being inductive. The key point of this algebra is as follows: it is possible to

obtain property (1) and also property (2) without σ-additivity. It is even possi-

ble to combine properties (1) and (4) without σ-additivity (and, hence, produce

complete Humean induction). However, on Σ̄, it is not possible for property (3)

without σ-additivity. This makes σ-additivity not only sufficient, but necessary

for partial induction (and, hence, for complete induction as well). Thus,

The Structure Theorem A coherent view of the world P is inductive if and

only if is σ-additive on Σ̄.

The Structure Theorem is a full characterization result that delivers an ax-

iomatic understanding of induction. The key result is the demonstration that

while lack of σ-additivity does not assure skepticism in the sense of Goodman and

it does not ensure skepticism in the sense of Hume either, it always assures skep-

ticism in at least one of these two senses. So, on Σ̄, any result that holds without

σ-additivity holds under some skepticism over induction. Conversely, results that

require σ-additivity, also require induction.

15



The collection Σ̄ is smaller than the σ-algebras commonly used in probability

theory. While σ-algebras are mathematically convenient under σ-additivity, they

do not play a particular role under finite additivity. What makes Σ̄ appealing

in the context of induction is that Σ̄ is the simplest (i.e. the smallest) algebra

that allows to distinguish between inductive and non-inductive views of the world.

Small algebras such as Σ̄ have an additional advantage. Because Σ̄ is countable,

finitely additive measures can be defined on P using only elementary mathematics,

and without invoking the (uncountable) Axiom of Choice.

6 Pragmatism, Induction and de Finetti

So far, we have focused on induction in the sense of the empirical validation of

eternal laws of Nature. There are, however, other perspectives on induction, such

as the one in which Bob is concerned with making accurate predictions about

the practical future, rather than uncovering universal laws of nature, or even

questioning their existence.6

If a law or theory makes predictions that are accurate within some finite hori-

zon then the theory predicts as if it were correct. Thus, the argument goes, data

need not uncover the actual data generating process. Nor does it need to reveal

whether or not a law exists. It only needs to allow for accurate predictions for

the practical future. To fix ideas, we refer to this perspective as pragmatism, with

no claim that our narrow use of this terminology comprehends most associations

with this word.

We now revisit the different problems on induction, but from a more pragmatic

perspective. In doing so we take a shortcut in the conceptual development. We

define pragmatic inductive views as requiring that enough pattern data leads to a

near infallible guide to a bounded future, instead of first making a distinction be-

tween induction in the sense of Hume and Goodman and then obtaining accurate

predictions as a result of both conditions as we did in Remark 2.

Definition 7 A coherent view of the world P is pragmatically inductive if, for

6See Russell (1912, Ch. VI) for a discussion of induction which clearly distinguishes between
the two perspectives.

16



every path ω ∈ A and every natural number k,

P
(
ωt+k | ωt

)
→ 1 as t →∞. (5)

So, with enough pattern data, Bob is convinced that the next outcomes can be

predicted with near certainty. This follows, in Bob’s belief, even if Nature abides

by no laws or if it abides by a law that cannot be inferred from the data. The

only claim is that after enough pattern data Nature behaves as if it abides by a

(data-inferred) law for a bounded, but arbitrarily long future.

We now turn to the concept of complete induction in the sense of Hume, from

the pragmatic perspective. Let U be the set of unions of finite histories. So, a set
U ∈ U is a union of finite histories such as ωt, where ω ∈ Ω and t is a natural
number. Any arbitrarily complex set E ⊆ Ω can be approximated in terms of
finite histories by choosing a set U ∈U such that E ⊆ U. 7

Definition 8 A coherent view of the world P is pragmatically completely inductive

in the sense of Hume if for any set U ∈U such that A ⊆ U,

P
(
U | ωt

)
→ 1 as t →∞ on every ω in A

and for any set V ∈U such that Ac ⊆ V,

P
(
V | ωt

)
→ 1 as t →∞ on P -almost every ω in Ac.

Given the requirement for any set in U that contains A or Ac, there is, in
particular, the same requirement for sets arbitrarily close to A or Ac. Sufficient

pattern data leads to near certainty of finite histories associated with laws and

sufficient non-pattern data leads to near certainty of finite histories associated

with non-laws. Combining the two definition yields,

Definition 9 A coherent view of the world P is pragmatically completely inductive

if it is pragmatically inductive and pragmatically completely inductive in the sense

of Hume.

7For instance, the set Ac of paths not following a pattern can be written as Ac =
⋂

n Un,
where (Un) is a decreasing sequence in U.

17



So, in particular, enough pattern data leads to a near infallible guide to a

bounded future and enough non-pattern data leads to near certainty of future

finite histories associated with non-laws.

The Pragmatic Induction Theorem Every coherent view of the world is prag-

matically completely inductive.

Unlike the previous results, the Pragmatic Induction Theorem holds for all

coherent views of the world. No matter how coherent beliefs are formed, they must

express confidence that mechanical repetitions of Bayes’ rule transform sufficiently

numerous pattern data into a near infallible guide to a bounded future. In the case

of non-pattern data then, provided that the data is sufficiently long, there must

be confidence, approaching certainty, of an observable future associated with non-

laws. This holds without any other assumption such as σ-additivity. Therefore,

any coherent view of the world contains a seed of orgulity.

The concerns one may have about the orgulity of Bayesians, may not go away,

at least completely, by abandoning σ-additivity. The Pragmatic Induction Theo-

rem relies on multiple repetitions of Bayes’ rule alone, hence it holds with or with-

out σ-additivity. However, the extent to which this remaining form of orgulity is a

difficulty for the Bayesian paradigm is a question beyond the scope of this paper.

According to one viewpoint, the cases of successful inference that follow from the

repetition of Bayes’ rule can be seen as a desideratum that provide support to

the Bayesian approach. According to a different viewpoint, the Pragmatic Induc-

tion Theorem can be seen as an expression of excessive confidence of the same

paradigm. This paper does not resolve this fundamental tension but it helps to

make precise the conditions under which orgulity holds.

While the Pragmatic Induction Theorem relies only on coherence and Bayes’

rule, it is embedded in a standpoint that can be traced back to de Finetti. The

key conceptual point advanced by de Finetti is that the Bayesian perspective on

inference effectively solves the problem of induction. As he wrote in de Finetti

(1970):8

8See chapters 11.1.5 and 11.2.1. For de Finetti’s (1970) perspective on induction, see also de
Finetti (1970b, 1972).

18



In the philosophical arena, the problem of induction, its meaning, use

and justification, has given rise to endless controversy, which, in the

absence of an appropriate probabilistic framework, has inevitably been

fruitless, leaving the major issues unresolved. It seems to me that the

question was correctly formulated by Hume [...]

In our formulation, the problem of induction is, in fact, no longer a

problem: we have, in effect, solved it without mentioning it explicitly.

Everything reduces to the notion of conditional probability [...]

In this sense, the Pragmatic Induction Theorem can be seen as formalization

of de Finetti’s viewpoint on induction. However, to the best our knowledge, de

Finetti never made a distinction between the two basic inference problems (i.e.,

does Nature abides by laws, and if so which one?) and never examined these

problems in a formal model. While pragmatism is the additional element necessary

for the formalization of this viewpoint, there is a yet more basic contribution. de

Finetti mostly wrote about induction in the context, as in de Finetti (1969), of

exchangeable beliefs (i.e. beliefs such that the order in which different outcomes

occur over time is irrelevant). Exchangeability not only rules out elementary laws

such as “1 until period t and 0 afterwards,” it is also a critical assumption on

the data and, hence, an assumption on how past and future must relate to each

other. In contrast, in the Pragmatic Induction Theorem, confidence on limited,

but successful, inductive inference about the future holds without assumptions on

how the past and the future must relate to each other. The conclusions about the

future depends on the data, but there is no restriction on the data generating

process itself.

7 Extensions

This paper dealt with some inductive inference problems, but left others un-

examined. Perhaps the most basic limitation in this paper is that the data-

generating processes are deterministic. A natural extension could go as follows:

The Blackwell-Dubins theorem extends Proposition 1 to stochastic data generat-

ing processes. Let’s say that there are countably many (possibly stochastic) data

19



generating processes P1,P2,P3, ... and Bob’s belief (a prior over {P1,P2,P3, ...})
assigns, ex-ante, strictly positive probability to each of them. If all probabilities

are σ-additive then Bob’s predictions will become eventually indistinguishable

from the data generating process, no matter which one.

In spite of the power of the Blackwell-Dubins theorem, new difficulties arise

in the case of stochastic data generating processes. For example, if two processes

are identical in all but the first period, then it may be impossible to empirically

determine which process runs the data. This determination is not relevant for

predicting the future after period 1 (see Lehrer and Smorodinski (1996) and Ace-

moglu, Cherzonukov and Yildiz (2016) on this problem). Other difficulties may

prove currently intractable. The Blackwell-Dubins theorem relies heavily on σ-

additivity. For general coherent views, there are some conceptual advances and

some analytical methods for Bayesian learning were developed in Pomatto, Al-

Najjar, and Sandroni (2014). With some effort, these techniques can be applied to

prove a version of the Pragmatic Induction Theorem for stochastic data-generating

processes. The Complete Humean Induction and the Separation Theorems are

existence results and so still hold when the set of data generating processes is

expanded. The main hurdle is the Structure Theorem. For a counterpart of that

result, one must find an algebra on which induction is equivalent to σ-additivity

when the data generating processes can be stochastic. This is a (very) difficult

problem.

8 Appendix

8.1 Proof of the Separation Theorem

We now provide examples of views that are inductive in the sense of Hume, but

not in the sense of Goodman, or are inductive in the sense of Goodman but not

in the sense of Hume.

Fix a σ-additive measure Pσ =
∑

ω∈A βωδω, where each δω is the measure

putting probability 1 on a path ω and (βω) are strictly positive weights such that∑
ω∈A βω = 1. Being σ-additive, it is inductive by Proposition 1.

We start with the following result.

20



Lemma 1 There exists a finitely additive probability S satisfying the following

two properties:

• S (ωt) = Pσ (ωt) for every ωt;

• S (A) = 0.

So, any finite history has the same probability under S as under P . However,

under S almost every path will eventually cease to follow a pattern.

Proof of Lemma 1. Let F be the algebra generated by all finite histories.
Consider the algebra A generated by F and the set A. As proved in  Loś and
Marczewski (1949), a set E ⊆ Ω belongs to A if and only if it is of the form
E = (F1 ∩A) ∪ (F2 ∩Ac) where F1,F2 belong to F. Let M be defined as

M ((F1 ∩A) ∪ (F2 ∩Ac)) = Pσ (F2)

for every set (F1 ∩A) ∪ (F2 ∩Ac) in A. It can be easily verified that M is a well
defined probability measure on A. Let S be any measure extending M from A to
Σ (see, for example,  Loś and Marczewski (1949) for a proof that such an extension

exists). By construction, S satisfies the desired properties.

The mixture Q = 1
2
Pσ +

1
2
S satisfies assumptions 1 and 2. It is inductive in

the sense of Goodman but not in the sense of Hume. The intuition for why S

is inductive in the sense of Goodman is as follows: when conditioning on A the

measure Q reduces to the σ-additive measure Pσ, which is inductive. Formally,

because S (A) = 0 then for every ω ∈ A we have

Q
(
{ω}|A∩ωt

)
=

1
2
Pσ ({ω}∩A)

1
2
Pσ (A∩ωt) + 12S (A∩ω

t)
= Pσ

(
{ω}|ωt

)
for each t. The measure Pσ is σ-additive hence inductive, so Pσ ({ω}|ωt) converges
to 1 for every ω ∈ A. Hence, Q is inductive in the sense of Goodman. To see that
it is not inductive in the sense of Hume, notice that for every ω ∈ A, we have

Q
(
A|ωt

)
=
Pσ (A∩ωt) + S (A∩ωt)

Pσ (ωt) + S (ωt)
=
Pσ (A∩ωt)

2Pσ (ωt)
=

1

2

21



Hence, Q (A|ωt) remains equal to 1
2

no matter how large t is. So, Q is not inductive

in the sense of Hume.

We now construct an example of a measure inductive in the sense of Hume

but not in the sense of Goodman. As implied by assumption 3, we can fix a path

ω̄ ∈ A with the property that for every t we can find another path ω̄t ∈ A distinct
from ω̄ such that ω̄tt = ω̄

t (so ω̄t and ω̄ coincide on the first t outcomes but differ

on some later outcome). As is well known, there exist finitely additive probability

measures that assign probability 0 to each single path but probability 1 to the

whole set {ω̄1, ω̄2, ...}(see, for example, Rao and Rao 1983). Let R be such a a
measure. We consider the mixture

P =
1

2
Pσ +

1

2
R

It satisfies assumptions 1 and 2. In addition,

P
(
A|ωt

)
=
Pσ (A∩ωt) + R (A∩ωt)

Pσ (ωt) + R (ωt)
= 1

since Pσ (A) = R (A) = 1. To see that P is not inductive in the sense of Goodman

consider the finite history ω̄t. Bayes’ rule implies

P
(
{ω̄}|A∩ ω̄t

)
=

Pσ ({ω̄})
Pσ (ω̄

t) + R (ω̄t)
.

By definition the measure R assigns probability 0 to every finite set of paths.

Hence R ({ω̄k : k ≥ 1, ω̄k ∈ ω̄t}) = R ({ω̄k : k ≥ 1}) for every t, so that R (ω̄t) = 1.
Therefore

P
(
{ω̄}|A∩ ω̄t

)
=

Pσ ({ω̄})
Pσ (ω̄

t) + 1

As t →∞, σ-additivity implies that Pσ (ω̄t) converges to Pσ ({ω̄}), so P ({ω̄}|A∩ ω̄t)
converges to 1

2
. Hence, P is inductive in the sense of Hume but not in the sense

of Goodman.

22



8.2 Proof of the Complete Humean Induction Theorem

The proof follows the same argument in the second part of the proof of the Separa-

tion Theorem. Let Pσ and R be defined as in the above proof, and let ω̃ be a path

such that ω̃ /∈ A and ω̃1 6= ω̄1 (since A is countable, such a path exists). Consider
the mixture P = 1

3
Pσ +

1
3
R + 1

3
δω̃. As shown above, we have P (A|ωt) → 1 as

t →∞ for every path ω ∈ A. Given the path ω̃, we have that for every t > 1,

P
(
Ac|ω̃t

)
=
Pσ
(
Ac ∩ ω̃t

)
+ R

(
Ac ∩ ω̃t

)
+ 1

Pσ
(
ω̃t
)

+ R
(
ω̃t
)

+ 1

since ω̃t 6= ω̄t then R
(
ω̃t
)

= R
({
ω̄k : ω̄k ∈ ω̃t

})
= 0. Therefore

P
(
Ac|ω̃t

)
=
Pσ
(
Ac ∩ ω̃t

)
+ 1

Pσ
(
ω̃t
)

+ 1

since ω̃ /∈ A, then Pσ
(
ω̃t
)
→ 0, so P

(
Ac|ω̃t

)
→ 1. Therefore, P is completely

inductive in the sense of Hume. To see that P is not σ-additive, notice that for

every n, we have

P ({ω̄k : k ≥ n}) =
1

3

∑
k≥n

Pσ ({ω̄k}) +
1

3
R ({ω̄k : k ≥ n}) .

Because R assigns probability 0 to every finite set of paths, we have R ({ω̄k : k ≥ n}) =
1 for every n. Hence, P ({ω̄k : k ≥ n}) ≥ 13 for every n, even if ∩n{ω̄k : k ≥ n} =
∅. Hence P is not σ-additive.

8.3 Proof of the Structure Theorem

We denote by F the algebra generated by all finite histories. Hence F ⊆ Σ̄ ⊆ Σ.
A result related to the next lemma appears in Al-Najjar, Pomatto and Sandroni

(2014).

Lemma 2 A set E belongs to Σ̄ if and only if there exists a set F belonging to

F such that the symmetric difference E4F is finite and included in A.

Proof. Let E be the collection of sets E for which there exists a set F ∈ F

23



such that the symmetric difference E4F is finite and included in A. We prove
that E ⊆ Σ̄. Let E and F ∈ F be such that E4F is finite and included in
A. Because E\F is finite and included in A and Σ̄ is an algebra containing each
singleton {ω} for paths in A, then F ∪ (E\F) ∈ Σ̄. Similarly, F\E ∈ Σ̄ and
so E = (F ∪ (E\F))\(F\E) ∈ Σ̄. We now show that Σ̄ ⊆ E. It follows from
the definition that E satisfies F ⊆ E and {ω} ∈ E for each ω ∈ A. We now
prove that E is an algebra. Let E ∈E be such that E 4F is finite and included
in A for some F ∈ F. Because Ec 4 Fc = E 4 F and Fc ∈ F , it follows
that Ec ∈ E. Now let E1,E2 ∈ E, and fix F1,F2 ∈ F such that E1 4 F1 and
E2 4F2 are finite and included in A. Let E = E1 ∪E2 and F = F1 ∪F2. Then
E 4F ⊆ (E1 4F1) ∪ (E2 4F2). Hence E 4F is finite and satisfies E 4F ⊆ A.
Thus, E is closed under union and complementation. Therefore, E is an algebra.
So, Σ̄ ⊆E. Thus Σ̄ = E.

We can now proceed with the proof. Let P be σ-additive. As shown in,

for instance, Shiryaev (1996) (page 134), σ-additivity implies that P must sat-

isfy P (ωt) → P ({ω}) as t → ∞, for every ω ∈ A. Therefore, P ({ω}|ωt) =
P ({ω}) /P (ωt) → 1 whenever P ({ω}) > 0. So, by Remark 2, P is inductive
in the sense of Hume and in the sense of Goodman. Conversely, suppose P is

inductive in both sense. We now show it is σ-additive on Σ̄. Let µ be the re-

striction of P on F. The measure µ is σ-additive on F (see the discussion in
Example 10.4.2. in Rao and Rao (1983)). So, by Carateodory theorem it admits

a σ-additive extension Pµ on the σ-algebra generated by F. In order to show that
P is σ-additive (on Σ̄) we prove that Pµ (E) = P (E) for every E ∈ Σ̄.

Let E ∈ Σ̄ and choose a set F ∈ F such that E4F is finite and included in
A. By additivity, any measure Q satisfies

Q (E) = Q (F) +
∑

ω∈E−F

Q ({ω}) −
∑

ω∈F−E

Q ({ω}) (6)

By construction, Pµ and P coincide on F. Hence P (F) = Pµ (F). Since P is in-
ductive, for every ω ∈ A, by Remark 2 it satisfies P ({ω}|ωt) = P ({ω}) /P (ωt) →
1, i.e. P ({ω}) = limt P (ωt). The σ-additivity of Pµ and the fact P and Pµ coin-

24



cide on F imply

Pµ ({ω}) = lim
t
Pµ
(
ωt
)

= lim
t
P
(
ωt
)

= P ({ω})

for every ω ∈ A. In particular, this holds for every ω ∈ E4F . We can therefore
conclude from (6) that

Pµ (E) = Pµ (F) +
∑

ω∈E−F

Pµ ({ω}) −
∑

ω∈F−E

Pµ ({ω})

= P (F) +
∑

ω∈E−F

P ({ω}) −
∑

ω∈F−E

P ({ω})

= P (E) .

Because E is arbitrary, it then follows that P and Pµ coincide on Σ̄. Hence P is

σ-additive on Σ̄.

8.4 Proof of the Pragmatic Induction Theorem

Endow Ω with the product topology, and let B be the Borel σ-algebra generated.
Let F be, as before, the algebra generated by all finite histories. Given any
coherent view of the world P (satisfying, as usual, assumptions 1 and 2) consider

the restriction µ of P on F. Following the proof of the Structure Theorem, the
measure µ admits a σ-additive extension Pσ on B.

We now show that P is pragmatically inductive. For each ω ∈ A we have
Pσ ({ω}) > 0. To see this, notice that σ-additivity implies Pσ ({ω}) = limt Pσ (ωt).
For each t, we have Pσ (ω

t) = P (ωt) ≥ P ({ω}) > 0. Hence Pσ ({ω}) > 0.
Therefore, by σ-additivity, Pσ ({ω}|ωt) → 1 as t → ∞. Since Pσ

(
ωt+K|ωt

)
≥

Pσ ({ω}|ωt), we conclude that Pσ
(
ωt+K|ωt

)
→ 1 as t →∞. Because Pσ

(
ωt+K|ωt

)
=

P
(
ωt+K|ωt

)
for every t, we conclude that P is pragmatically inductive.

The result that P is pragmatically completely inductive in the sense of Hume

can be proved as a consequence of the following general principle: for every set

U ∈U and every history ωt, we have

P
(
U | ωt

)
≥ Pσ

(
U | ωt

)
.

25



We now prove this claim. The collection U of unions finite histories forms a base
for the topology. Since the product topology is separable, each U ∈ U can be
written as U =

⋃∞
n=1 hn where each hn is a finite history. For each m, we have

that
⋃m
n=1 hn belongs to F, hence

P (U) ≥ P

(
m⋃
n=1

hn

)
= Pσ

(
m⋃
n=1

hn

)

Since
⋃m
n=1 hn ↑ U as m → ∞, σ-additivity implies Pσ (

⋃m
n=1 hn) ↑ Pσ (U) as

m →∞. Therefore P (U) ≥ Pσ (U). For each t and path ω, the set U∩ωt is open,
and the same argument as above implies that P (U ∩ωt) ≥ Pσ (U ∩ωt). Because
Pσ and P coincide on F, we also have P (ωt) = Pσ (ωt). Hence P (U | ωt) ≥
Pσ (U | ωt), as claimed.

Because Pσ is σ-additive, it is completely inductive in the sense of Hume. So, if

A ⊆ U and Ac ⊆ V then Pσ (U|ωt) → 1 for every ω ∈ A and P (V |ωt) → 1 for P-
almost every path ω ∈ Ac. Since P (U|ωt) ≥ Pσ (U|ωt) and P (V |ωt) ≥ Pσ (V |ωt),
it then follows that P is pragmatically completely inductive in the sense of Hume.

8.5 Proof of other results in the text

Proof of Remark 1. The proof of this result is standard, and included only

for the sake of completeness. Let D = {ω : P ({ω}) > 0} be the set of paths to
which P attaches strictly positive probability. The additivity of P implies that for

each positive integer k, the set Dk = {ω : P ({ω}) > k−1} must be finite. Hence
D = ∪∞k=1Dk is countable.

Proof of Remark 2. Assumptions 1, 2 and 3 imply that for each ω and t, the

conditional probabilities P (·|ωt) and P (·|ωt ∩A) are well defined. In addition,
by the law of total probability, for each ω ∈ A we have

P
(
{ω}|ωt

)
= P

(
{ω}|ωt ∩A

)
P
(
A|ωt

)
for each ω ∈ A. Hence, as t → ∞, it follows that P ({ω}|ωt) → 1 if and only
if P ({ω}|ωt ∩A) P (A|ωt) → 1. That is, if and only if P ({ω}|ωt ∩A) → 1 and
P (A|ωt) → 1.

26



Proof of Proposition 3. Let ε ∈ (0, 1) and let Pσ be a σ-additive measure that
satisfies assumptions 1-3. Using Lemma 1, let S be a probability measure that

satisfies S (ωt) = Pσ (ω
t) for every history, but S (A) = 0. Let P = εPσ+(1 −ε) S.

Then, for every ω ∈ A and every t, we have

P
(
A|ωt

)
=
εPσ (A∩ωt) + (1 −ε) S (A∩ωt)

Pσ (ωt)
=
εPσ (A∩ωt)
Pσ (ωt)

≤ ε.

References

[1] Acemoglu, D., Cherzonukov, V. and M. Yildiz (2016): “Fragility of Asymp-

totic Agreement under Bayesian Learning.” Theoretical Economics, 11, 187-

227.

[2] Al-Najjar, N., Pomatto, L., and A. Sandroni (2014). “Claim Validation.” The

American Economic Review, 104(11), 3725-3736.

[3] Belot, G. (2015): “Bayesian orgulity.” Philosophy of Science. 80(4). 483-503.

[4] Belot, G. (2015): “Objectivity and Bias.” Mind, forthcoming.

[5] Blackwell, D. and L. Dubins (1962): “Merging of opinions with increasing

information.” The Annals of Mathematical Statistics, 33(3), 882-886.

[6] Broad, C. D. (1918): ”On the relation between induction and probability -

(Part I.).” Mind 27-4: 389-404.

[7] Doob, J. L. (1949): “Application of the theory of martingales.” Le calcul des

probabilites et ses applications, 23-27.

[8] de Finetti, B. (1969). “Initial probabilities: A prerequisite for any valid in-

duction”. Synthese, 20(1), 2-16.

[9] de Finetti, B. (1970): Theory of Probability, vol. 2. Wiley, New York.

[10] de Finetti, B. (1972). Probability, Induction, and Statistics. Wiley, New York.

27



[11] Edgeworth, F. Y. (1922): “The philosophy of chance.” Mind, 31-123, 257-283.

[12] Elga, A. (2015): “Bayesian humility.” Philosophy of Science. Forthcoming.

[13] Gilboa, I., and L. Samuelson (2012): “Subjectivity in inductive inference.”

Theoretical Economics, 7(2), 183-215.

[14] Goodman, N. (1955): Fact, Fiction and Forecast. Harvard University Press,

Cambridge.

[15] Howson, C. and P. Urbach. (2006). Scientific reasoning: The Bayesian ap-

proach.

[16] Huttegger, S. (2015): “Bayesian convergence to the truth and the meta-

physics of possible worlds.” Philosophy of Science. 82(4), 587-601.

[17] Kelly, K.T., and C. Juhl. (1994). “Reliability, convergence, and additivity.”

PSA: Proceedings of the Biennial Meeting of the Philosophy of Science As-

sociation, 181-189.

[18] Kelly, K. T. (1996): Logic of Reliable Inquiry. Oxford University Press.

[19] Lehrer, E., and R. Smorodinsky (1996): “Merging and learning.” Statistics,

Probability and Game Theory, 147–168.

[20] Lévy, P. (1937): Theorie de l’Addition des Variables Aléatoires. Gauthier-

Villars, Paris.

[21]  Loś, J. and E. Marczewski (1949): “Extensions of Measures.” Fundamenta

Mathematicae, 1(36), 267-276.

[22] Pomatto, L., N. Al-Najjar, and A. Sandroni (2014): “Merging and testing

opinions.” The Annals of Statistics 42(3), 1003-1028.

[23] Rao, K.P.S. and M. Rao (1983): Theory of Charges, Academic Press, New

York.

[24] Russell, B. (1912). The Problems of Philosophy. Oxford.

28



[25] Shiryaev, A. N. (1996): Probability. Springer-Verlag, New York.

[26] Stalker, D. F. (1994): Grue!: the New Riddle of induction. Open Court

Publishing Company.

[27] Weatherson, B. (2015): “For Bayesians, rational modesty requires impreci-

sion.” Ergo, an Open Access Journal of Philosophy. 2.

[28] Wrinch, D., and H. Jeffreys. “On some aspects of the theory of probability.”

Philosophical Magazine. 38-228. 715-731.

[29] Zabell, S. L. (2011): “Carnap and the logic of inductive inference.” Handbook

of the history of logic. 265-309.

29