Rudolf Carnap’s Logical Foundations of Probability is a work of technical philosophy, written during his years at the University of Chicago and published in the middle of the twentieth century. It sets out to rebuild inductive logic on exact foundations, using the machinery of symbolic logic and semantics that Carnap had spent the preceding decades developing. The book is not a survey; it is a construction, proceeding by definitions, theorems, and proofs, and it announces a second volume devoted to the system of inductive logic itself.
The governing conviction is stated plainly in the preface. All nondeductive reasoning is reasoning in terms of probability; inductive logic is therefore probability logic; and the probability at issue is a logical relation between two statements — the degree to which a hypothesis is confirmed by given evidence. The frequency conception of probability, though indispensable to science, cannot serve as the basic concept of inductive logic. From this it follows that the principles of inductive logic are analytic, and that inductive validity needs no synthetic presupposition such as a uniformity of nature. The book’s aim is to build a system on those terms, applicable to a deliberately simplified language.
From Explication to the Symmetrical c-Functions
The Program and Its Six Conceptions
Carnap opens with a preface that lays out a new approach to induction and probability, built on six basic conceptions. All inductive, nondeductive reasoning is reasoning in terms of probability; inductive logic is therefore identical with probability logic; the relevant probability concept is a logical relation between two statements—the degree of confirmation of a hypothesis on given evidence; the frequency concept, however scientifically important, is unsuitable as the basic concept of inductive logic; all principles and theorems of inductive logic are analytic; and inductive validity therefore does not rest on synthetic presuppositions such as the uniformity of the world. The major aim is to construct a system of inductive logic using symbolic logic and semantics, applicable to a simplified language system—lower functional logic with relations and identity. He maps the book’s structure: philosophical discussion in chapters i, ii, and iv; construction in chapters v–ix, proceeding “more geometrico” by definitions and proofs. Labeling conventions, summaries, small-type digressions, and reading guidance for philosophers, classical probability theorists, followers of von Mises or Reichenbach, mathematical statisticians of the Fisher, Neyman, Pearson, and Wald schools, and those concerned with practical decisions are all set out. Acknowledgments credit the University of Chicago and Rockefeller Foundation for a 1942–44 leave, and Feigl, Hempel, Quine, and Tintner for comments; the preface is signed at Chicago, March 1950.
Explication as the Method
Chapter i concerns explication: transforming an inexact prescientific concept, the explicandum, into an exact one, the explicatum, which must satisfy similarity, exactness, fruitfulness, and simplicity. Carnap distinguishes classificatory, comparative, and quantitative concepts, and formalization from interpretation. The book’s three tasks are clarifying degree of confirmation, induction, and probability. Because the explicandum is inexact, a proposed explication cannot be judged simply right or wrong, only more or less satisfactory.
Confirmation as a Semantical Relation
Confirmation is a complex procedure with components of different kinds; the book concerns only its logical aspect—relations between sentences. A scientist formulates observational results as a long report sentence e and asks whether and how strongly a hypothesis h is confirmed by it. Once h and e are given, the answer requires only logical analysis of their meanings, not factual knowledge. Since semantics is the theory of meaning, these are semantical concepts of confirmation, belonging to inductive rather than deductive logic. Carnap explains confirmation by analogy with L-implication: deciding whether e L-implies h needs no factual knowledge, only analysis of meanings; likewise we need not know whether e or h is true, or whether anyone believes e—unlike methodological problems of confirmation such as building apparatus or running experiments. Three semantical concepts correspond to three levels: the classificatory (“h is confirmed by e,” a relation between two sentences, symbolized ‘@’); the comparative (“h is more strongly confirmed by e than h′ by e′,” a tetradic relation, or triadic when restricted to same hypothesis or same evidence, with a “more or equal” form symbolized ‘𝔐’); and the quantitative degree of confirmation, ‘c(h,e) = q’, with q a real number between 0 and 1, to be defined as ‘c*’ in Volume II.
Two Concepts of Probability
Theories are commonly classed as classical, logical, or frequency, but these are not answers to the same problem. There are two principal explicanda: probability₁, degree of confirmation, and probability₂, relative frequency. The controversy between frequency theorists and others is futile, caused chiefly by authors not realizing the other side starts from a different explicandum. Nagel’s classification is cited—classical (Bernoulli, Laplace), objective logical relation (Keynes, Jeffreys), relative frequency (von Mises, Reichenbach, modern statistics)—and Carnap asks not what solutions authors offer but which problems they intend to solve. Phrases differ widely (“degree of belief,” “credibility,” “degree of reasonable expectation,” “partial truth”), so a unique explicandum is untenable; yet the number is chiefly two, each with classificatory, comparative, and quantitative forms. The distinction concerns explicanda, not explicata. Von Mises, whose frequency theory is the first complete one, holds probability₂ the only basis of the calculus and finds talk of the probability of an individual’s death meaningless; he misreads Keynes’s conception as thoroughly subjectivistic. Jeffreys demands that every theory of probability₁ meet requirements frequency theories fail, insisting no “objective” definition in terms of actual or possible observations is admissible, since observational results are initially unknown and we would have no starting point; in practice, he says, no statistician uses a frequency definition but all use degree of reasonable belief, usually without noticing.
Laying the Deductive Foundations
The chapter on deductive logic has three parts. The first gives the semantical foundations presupposed by what follows.
Q-Predicates and the Anatomy of Logical Space
Carnap opens by noting that counting structures would serve both deductive and inductive logic, and that the simplest case—one dyadic relation—has already been handled in graph theory, citing König’s 1936 treatise and its references to Jordan and Cayley. Graphs correspond to structures of symmetric relations, so those results must be generalized to nonsymmetric ones. He then introduces the Q-predicates: molecular predicates built by conjoining each primitive predicate or its negation, displayed for three predicates in a table of eight rows that mirrors a truth-table. Each Q-predicate abbreviates one such conjunction, and the definitions fix the Q-matrix, the Q-sentence, and κ, the number of Q-predicates, which equals 2^t. Every individual must possess exactly one Q-property, so the Q-properties form a division—the strongest partition expressible in the system, the infimae species of Aristotelian terminology. The theorems record that distinct Q-predicates are L-exclusive, that together they are L-disjunct, that each is factual, and that every molecular predicate expression is either L-empty or L-equivalent to a disjunction of some number of Q-predicates.
Logical Width
From this basis Carnap defines the width of a predicate expression: zero for L-empty expressions, otherwise the number of Q-predicates in its L-equivalent disjunction. With two independent properties such as Small and Black, their conjunction is narrower than either alone and their disjunction wider. Width is not extension—whether all black things happen to be small is a factual matter—but a logical relation: a wider property admits more possibilities. Comparing properties that do not stand in an L-implication relation requires extra conventions, which the primitive properties of any system supply. Q-properties receive width 1, the narrowest non-empty properties; relative width is width divided by κ. Carnap stresses that this concept matters greatly for inductive logic: the classical theory of probability fails to consider the width of the properties involved, and some of its principles, stated too generally, generate contradictions. A modification that makes degree of confirmation depend on widths will remove them.
The Logical Concept of Probability
Turning to probability itself, Carnap separates probability₁, the weight of evidence, from probability₂, relative frequency, and announces the explication of probability₁ as the book’s chief topic. It can be approached classificatorily (“h is confirmed by e”), comparatively, or quantitatively as a degree of confirmation c(h,e) = r. Three viewpoints explain it: as a measure of evidential support, as a fair betting quotient, and as an estimate of relative frequency. Support alone leaves numerical values unclear, since the assumptions that support is measured by nonnegative numbers and is additive over L-exclusive hypotheses are arbitrary. The betting analysis yields values in the interval from 0 to 1 and the complement rule for negations. The frequency analysis shows that a bettor’s fair quotient equals his estimate of the relative frequency, and Carnap concedes the circularity is admissible for clarifying an explicandum though not for a definitional system. The die example exposes confusions between the two probabilities, and a comparison with Reichenbach finds agreement on the explicanda but disagreement over identifying weight with frequency. Probability₁ does not supersede truth; inductive logic remains two-valued.
Inductive and Deductive Logic Compared
The final argument insists that the two logics share one structure. An effective procedure fixes each step and finishes in finitely many; a decision procedure determines truth or provability for a class of sentences. Deductive logic is the theory of L-implication, inductive logic the theory of degree of confirmation, and the comparison is hypothetical, since no adequate quantitative explicatum is yet known. In both branches the first problem—finding a conclusion—has no effective procedure: discovering theorems and framing hypotheses alike demand creative ability. The second problem—examining a result—also lacks a general method, since L-implication is not effective and Church showed no decision method can cover all sentences of lower functional logic; restricted methods exist in both fields, so an inductive machine is impossible in general but possible for limited domains, exactly as with deductive machines. The third problem, checking a given proof, is effectively decidable when the proof is fully laid out, though in practice only “practically effective,” since mathematicians abbreviate steps.
Insurance, Estimates, and the Limits of Simple Rules
Carnap tests his decision rules against a homely case: a man owns a $10,000 house, his evidence puts the chance of fire at one in a thousand, and a $5 premium would plainly make insurance worthwhile. Yet a rule that treats the most probable outcome as certain would have him act as though the house certainly would not burn, and so decline the policy. The rule of maximizing estimated gain fares no better in general, and Carnap moves to a rule of estimates: compute the probability-weighted value of the unknown magnitude and act as though it were near that figure. This handles the lottery, where a fair ticket costs exactly the estimated gain of one dollar, and the bookshop, where symmetry makes estimate and most probable value coincide. But applied literally it turns absurd—betting a thousand to one on exactly eighty copies sold—and it cannot say in which respects the estimate should govern conduct. Asymmetrical losses matter too: a bookseller who understocks loses profitable trade while overstocking merely costs storage, so he sensibly orders above his estimate.
Maximizing Estimated Gain and the Coin Bet
The refined rule compares actions by summing, over possible outcomes, the probability of each outcome given the action times the resulting gain, and selects the action with the highest estimate. It needs no pretence of knowledge and accommodates asymmetry. In simple bets it agrees with the cruder estimate rule, which may therefore be retained as a convenient approximation. A specialized version—accept favorable offers, reject unfavorable ones, treat neutral ones either way—looks like common sense until a coin flip exposes it. A man with $10,000 facing an even bet at even odds is told to accept, yet risking $8,000 on a toss strikes everyone as unreasonable; restricting the rule by the ratio of stake to fortune fails because other cases demand greater liberality, not less. The insurance case itself turns out to be an unfavorable offer that reason nonetheless endorses.
Reducing Confirmation to Finite Systems
Attention shifts to the technical machinery. Because the infinite system cannot be handled directly, Carnap defines confirmation in it as the limit of the values assigned in finite systems as the number of individuals grows. Since the tautological evidence carries no factual content, confirmation relative to it—null confirmation—becomes the basic notion, and the general value is recovered as a quotient of null confirmations. A sentence is equivalent to the disjunction of the state-descriptions in which it holds, so null confirmation for sentences reduces to a sum over state-descriptions, each of which must receive a positive value, the values summing to one. Any such assignment yields a regular confirmation function. Infinitely many such functions exist, much as many temperature scales suit the comparative “warmer,” and choosing them system by system independently will not do: they must fit together, since a sentence means the same thing in every system containing it.
What Axioms Alone Cannot Supply
Modern probability theories that lack an analogue of the principle of indifference state only theorems holding for every regular function, and so admit wildly inadequate distributions alongside sound ones. Such theories are correct but weak. Carnap expects to narrow the field in two large steps: first to the symmetrical functions, then by one further requirement to the function he will propose as explicatum. He lists theorems about confirmation values of zero and one, noting Keynes stated most of them but proved them only for finite systems because he conflated probability zero and one with deductive concepts.
Bayes’s Theorem and Relevance
Treating present knowledge, hypothesis, and prediction, Carnap derives the ratio by which an observation raises a hypothesis’s confirmation: posterior confirmation is proportional to prior confirmation and to likelihood, inversely proportional to expectedness—the more surprising the observation, the more it confirms. Bayes’s theorem extends this to competing hypotheses, and Carnap defends its validity once correctly formulated, while conceding that some applications via indifference are objectionable. “Prior” means prior to the new observation, not a priori. A medical example shows a chooser of tests preferring higher likelihood when expectedness is equal, lower expectedness when likelihood is equal.
The Relevance Quotient and Measure
Keynes’s notation is simplified, and the relevance quotient is defined as the factor by which evidence multiplies confirmation, exceeding one for positive relevance and falling below one for negative. It is commutative in hypothesis and evidence, though not always in the infinite system, and Johnson’s cumulative formula lets accumulated evidence be handled item by item. A separate relevance measure, built from m-values, is preferred to simple differences because it is symmetric and changes sign under negation, and it proves additive for exclusive disjunctions—a property m and c lack. Seventy-five sign-situations are tabulated, with the surprising result that two items may each be positively relevant to a hypothesis while their conjunction is negative.
Disjunctive and Conjunctive Analysis of Relevance
Carnap opens with two complementary ways of computing the relevance measure r of a sentence i to a hypothesis h on evidence e. The first, disjunctive analysis, breaks i into the state-descriptions s in its range and sums their r-values; since only state-descriptions inside the range of e·i contribute, the sum splits into a part where h holds (R₁, all nonnegative) and a part where it fails (R₂, all nonpositive). A table catalogues every deductive-relation case among e, h, and i, and the method rests on the additivity of r over exclusive disjunctions, terminating at state-descriptions that admit no further splitting. A zero value for i can mean either that every component is zero or that positive and negative contributions cancel.
The second, conjunctive method works from the opposite end: it analyzes i into its weakest conjunctive components, the negations of state-descriptions outside its range, which Carnap calls content-elements and whose class is the content of i. Here too the r-value of i is the sum over its content-elements, with only the ranges R₃ and R₄ of e·~i·h and e·~i·~h mattering. Content turns out to map neatly onto the L-concepts—universal when i is L-false, null when L-true, ordered by L-implication—so it captures assertive strength, a debt Carnap credits to Popper’s Logik while correcting an earlier misattribution to Wittgenstein in his own Semantics. Range nonetheless remains fundamental for inductive logic, since m– and c-functions add over disjunction but not conjunction; content serves chiefly as an auxiliary notion.
Extreme and Complete Relevance
Two sharper concepts follow. A sentence is extremely positive to h on e when it is positive and no stronger sentence is negative; in a finite system this amounts to raising the confirmation of h to 1, and the negative analogue lowers it to 0. Extreme irrelevance, by contrast, collapses into triviality: it holds only when e already implies h or its negation, or when e·i is L-false. Every state-description is always extremely positive, negative, or irrelevant, extreme relevance passes from a sentence to any stronger sentence not excluded by e, and—unlike simple relevance—these concepts do not depend on which regular c-function is chosen. Complete relevance mirrors this from the conjunctive side: i is completely positive when it is positive and no weaker sentence is negative, with complete irrelevance again confined to trivial cases, as with Keynes’s strict-sense irrelevance.
The Comparative Concept MC
How should an adequate comparative relation of confirmation be found? Carnap refuses to pick a c-function first and hunt for a matching relation, since infinitely many candidates compete. He begins instead from regular c-functions and demands a relation grasping logical connections that all of them share: if MC holds between two hypothesis-evidence pairs, every regular c must rank the first at least as high as the second, and conversely. Trivial cases—where e implies h or e′ implies ~h′—satisfy this but are too narrow, so he adds further conditions, including a nontrivial one where two confirming instances outweigh one. The worry that any list of cases looks arbitrary yields a second requirement: whatever every regular c-function ranks higher must fall under MC. Together the requirements fix MC uniquely as the most comprehensive relation in accord with all regular c-functions. Carnap then recasts the quantitative condition in purely comparative terms of L-implication between ranges, showing that a violation becomes possible only if certain ranges are non-null, which forces the condition that e′·h′ implies e·h while e implies h or e′. The definition of MC in the system combines non-L-false evidence with a disjunction of three cases, and a nine-term partition proof establishes that MC satisfies both requirements and is the only relation that does. The result transfers to infinite systems with a restriction to nongeneral sentences, and three worked examples illustrate the nontrivial case.
Hempel on Confirming Evidence
Carnap declines to construct a non-quantitative explicatum for classificatory confirmation; for his inductive logic, confirming evidence is handled by the general theory of relevance for regular c-functions and by the theory of c\*, and he sees the quantitative treatment as no disadvantage.
Gauging the Reliability of an Estimate
Carnap turns to the question of how much confidence an estimate deserves. An observer who has settled on a confirmation function and an estimate-function still needs a way to say how precise that estimate is—how likely the true frequency lies near the guessed value. A large sample on red hair in Chicago earns more trust than a small one, and the task is to give an exact explicatum for this inexact notion of reliability. Three methods are weighed. The first reads reliability off the confirmation of an interval around the estimate, taken as a density in the limit. The second fixes a confirmation value and measures the width of the interval that carries it—the older sense of “probable error,” with lower and upper bounds when the density is asymmetrical. The third, adopted as most adequate, abandons the question of whether the error is probably small and instead estimates the error itself. Estimating the signed error fails at once, since positive and negative errors cancel to zero; estimating its square succeeds, and this becomes the measure.
Estimated Standard Error
Following Gauss, Carnap treats square error as the fruitful quantity. Reliability is the estimate of square error, whose square root is the estimated standard error. A theorem shows this equals the estimate of the squared frequency minus the square of the estimated frequency—so the estimate of the square always exceeds the square of the estimate, the gap being the variance. Three worked examples with values 1, 2, 3 display the arithmetic: equal weights give a standard error of 0.82, a peaked distribution 0.63, a spread-out one 0.89. Carnap then warns against a disastrous fallacy: comparing two observers’ standard errors to judge one estimate more reliable than another, when both derive from the same confirmation function, proves nothing. The method compares estimates under different evidence within one system, not rival systems against each other.
Estimating Frequencies
The machinery extends to systems ℒ, where truth-frequency and relative truth-frequency are defined in the metalanguage. The estimate of a frequency is a weighted sum over k-sentences, and the estimate of relative frequency turns out to be the arithmetic mean of the c-values of the individual sentences. This justifies reading probability₁ as an estimate of relative truth-frequency and lets one explain probability to a subject through estimation—equal c-values of 0.3 across a hundred hypotheses meaning an estimate of thirty true. Property-frequencies follow analogously.
Predictive Inference and the Weight of Widths
The appendix to the outline of inductive logic opens with predictive inference, inference from one sample to another, called the most important kind of inductive inference. For a singular prediction, c* is pulled by two factors: the relative width of the property, purely logical, and the observed relative frequency, empirical. The value always lies between them, starting at the width before observation and drifting toward the frequency as the sample grows. This improves on Laplace’s rule of succession, which yields 1/2 for any property and contradicts itself across different properties, and on the straight rule, which gives probability 1 after three confirming instances—hardly acceptable.
Analogy, Inverse and Universal Inference
Inference by analogy raises a peculiar difficulty: probability seems to rise with shared properties, yet properties must be weighed, not counted—which the concept of width permits. The increase that analogy confers proves real but ordinarily small, matching the common view that analogical reasoning is valid yet weak. Inverse inference, from sample to population, depends on widths as well as frequencies. Universal inference tests a law against a sample free of counterexamples: for the infinite system confirmation falls to zero, and for finite systems it shrinks as the population grows, since a larger universe asserts more.
What Confirmation Actually Is
The central concept of the whole edifice is the degree of confirmation, written c(h,e): the strength of support a hypothesis h receives from evidence e. Carnap insists this is a semantical relation, not a psychological one. Once h and e are given, deciding how strongly e supports h requires only analysis of their meanings, in the way that deciding whether a set of premises logically implies a conclusion requires no factual investigation. Confirmation comes in three grades — classificatory, comparative, and quantitative — and the book pursues the quantitative grade while treating the other two as legitimate stopping points if no adequate numerical explicatum can be found.
Two features of this concept do most of the work. First, it is relative to a language: the degree of confirmation depends on the structure of the language system in which the sentences are formulated, which is why Carnap spends so long on state-descriptions, ranges, Q-predicates, and logical width before he defines a single confirmation function. He criticizes Keynes, Jeffreys, and others precisely for neglecting this dependence. Second, the concept is normative rather than descriptive. Carnap is not reporting how people actually weigh evidence; he is constructing an exact replacement for a vague prescientific notion, and he judges candidate replacements by the standards he sets out for explication — similarity to the original, exactness, fruitfulness, and simplicity.
The Two Probabilities and the Futile Quarrel
Much of the book’s polemical energy goes into separating probability₁, the degree of confirmation, from probability₂, the relative frequency. Carnap’s diagnosis of the long controversy between frequency theorists and their opponents is that it is largely a misunderstanding: the two sides are not offering rival answers to one question but answers to different questions. Von Mises speaks of the probability of an individual’s death as meaningless; Jeffreys declares that no objective definition in terms of actual or possible observations is admissible. Each is right about his own explicandum and wrong to think the other is doing the same job badly.
This distinction licenses a division of labor. Frequency statements remain scientifically indispensable, and Carnap treats the estimate of a relative frequency as one route into the quantitative concept — a fair betting quotient turns out to coincide with the estimate of relative frequency. But the logical concept cannot be reduced to it, and statements of degree of confirmation are L-determinate: true or false by virtue of meaning, not of how the world happens to be. Against Reichenbach, who identifies weight with relative frequency rather than its estimate, Carnap marks a clean parting of the ways.
Rules for Acting Under Uncertainty
A striking stretch of the argument turns from logic to decision. Carnap examines a series of rules for practical choice under uncertainty and finds each wanting in its literal form. Acting as though the most probable outcome were certain fails the householder deciding on fire insurance; acting on the estimate of gain works in the bookshop and the restaurant but breaks down when applied without qualification, since it would have a man bet a thousand to one on an exact figure. The rule of maximizing estimated gain survives these tests and handles asymmetrical losses, yet even it runs aground on the coin bet, where common sense balks at staking most of one’s fortune on an even wager. Carnap’s conclusion is not that rational decision is impossible but that a single simple rule cannot capture it; the discussion feeds directly into Daniel Bernoulli’s law of utility.
The Shape of the Whole
What holds the book together is a comparison. Deductive and inductive logic, Carnap argues, have the same fundamental architecture: in each, finding a conclusion is a matter of creative ingenuity for which no effective procedure exists; examining a result is likewise undecidable in general, as Church’s theorem shows for the predicate calculus; only the checking of a given proof is genuinely mechanical. An inductive machine is impossible for the first two tasks — but so is a deductive machine for their deductive counterparts. The difference between the two logics lies in their basic concepts, not in their degree of regulation by rules.
The constructive program proceeds in two large steps. Having reduced confirmation in an infinite language to confirmation in finite ones, and that in turn to the null confirmation of state-descriptions, Carnap is left with an infinite family of regular c-functions, all admissible. The symmetrical c-functions narrow the field by requiring that confirmation depend only on the structure of the evidence, not on which individuals happen to be named. One further requirement, reserved for the second volume, is to single out the function c* as the proposed explicatum.
A Logic of Science Without Foundations in Fact
The book’s significance lies in what it refuses as much as in what it builds. By making inductive validity analytic, Carnap removes the need for any metaphysical guarantee that the future will resemble the past; the uniformity of the world is not a premise of induction but a factual hypothesis among others. The cost is that inductive logic becomes a framework rather than a single doctrine — a family of functions constrained by adequacy conditions, with the choice among them settled by a further methodological decision. Later work on confirmation, including Hempel’s analysis of confirming evidence, would take up the problems Carnap left open. The result is a rigorous, unhurried attempt to give the logic of science the same exactness as its deductive counterpart, and to show that the attempt, while never complete, is not doomed.
